diff --git a/.gitattributes b/.gitattributes index 5133731a49d8c59d0f823515e563dc2b53e8be2d..de00aa56e4a8f5aecb0f68497fd5c01666e4c886 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1959,3 +1959,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Robust[[:space:]]Multi-Object[[:space:]]4D[[:space:]]Generation[[:space:]]for[[:space:]]In-the-wild[[:space:]]Videos/5608bb67-dff7-4952-a0b9-ba92a1ee0b94_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Robust[[:space:]]Multimodal[[:space:]]Survival[[:space:]]Prediction[[:space:]]with[[:space:]]Conditional[[:space:]]Latent[[:space:]]Differentiation[[:space:]]Variational[[:space:]]AutoEncoder/67d06a41-5db8-4d3b-a154-67ea148145c7_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Robust-MVTON_[[:space:]]Learning[[:space:]]Cross-Pose[[:space:]]Feature[[:space:]]Alignment[[:space:]]and[[:space:]]Fusion[[:space:]]for[[:space:]]Robust[[:space:]]Multi-View[[:space:]]Virtual[[:space:]]Try-On/e4a7a6ed-e1ee-46d8-ba87-bfe4f3f1f820_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/RoomPainter_[[:space:]]View-Integrated[[:space:]]Diffusion[[:space:]]for[[:space:]]Consistent[[:space:]]Indoor[[:space:]]Scene[[:space:]]Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/RoomTour3D_[[:space:]]Geometry-Aware[[:space:]]Video-Instruction[[:space:]]Tuning[[:space:]]for[[:space:]]Embodied[[:space:]]Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Rotation-Equivariant[[:space:]]Self-Supervised[[:space:]]Method[[:space:]]in[[:space:]]Image[[:space:]]Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/S2D-LFE_[[:space:]]Sparse-to-Dense[[:space:]]Light[[:space:]]Field[[:space:]]Event[[:space:]]Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/S2Gaussian_[[:space:]]Sparse-View[[:space:]]Super-Resolution[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/S4-Driver_[[:space:]]Scalable[[:space:]]Self-Supervised[[:space:]]Driving[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]with[[:space:]]Spatio-Temporal[[:space:]]Visual[[:space:]]Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SACB-Net_[[:space:]]Spatial-awareness[[:space:]]Convolutions[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAIST_[[:space:]]Segment[[:space:]]Any[[:space:]]Infrared[[:space:]]Small[[:space:]]Target[[:space:]]Model[[:space:]]Guided[[:space:]]by[[:space:]]Contrastive[[:space:]]Language-Image[[:space:]]Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SALAD_[[:space:]]Skeleton-aware[[:space:]]Latent[[:space:]]Diffusion[[:space:]]for[[:space:]]Text-driven[[:space:]]Motion[[:space:]]Generation[[:space:]]and[[:space:]]Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SALOVA_[[:space:]]Segment-Augmented[[:space:]]Long[[:space:]]Video[[:space:]]Assistant[[:space:]]for[[:space:]]Targeted[[:space:]]Retrieval[[:space:]]and[[:space:]]Routing[[:space:]]in[[:space:]]Long-Form[[:space:]]Video[[:space:]]Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAM-I2V_[[:space:]]Upgrading[[:space:]]SAM[[:space:]]to[[:space:]]Support[[:space:]]Promptable[[:space:]]Video[[:space:]]Segmentation[[:space:]]with[[:space:]]Less[[:space:]]than[[:space:]]0.2%[[:space:]]Training[[:space:]]Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAM-REF_[[:space:]]Introducing[[:space:]]Image-Prompt[[:space:]]Synergy[[:space:]]during[[:space:]]Interaction[[:space:]]for[[:space:]]Detail[[:space:]]Enhancement[[:space:]]in[[:space:]]the[[:space:]]Segment[[:space:]]Anything[[:space:]]Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAM2-LOVE_[[:space:]]Segment[[:space:]]Anything[[:space:]]Model[[:space:]]2[[:space:]]in[[:space:]]Language-aided[[:space:]]Audio-Visual[[:space:]]Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAM2Object_[[:space:]]Consolidating[[:space:]]View[[:space:]]Consistency[[:space:]]via[[:space:]]SAM2[[:space:]]for[[:space:]]Zero-Shot[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAMBLE_[[:space:]]Shape-Specific[[:space:]]Point[[:space:]]Cloud[[:space:]]Sampling[[:space:]]for[[:space:]]an[[:space:]]Optimal[[:space:]]Trade-Off[[:space:]]Between[[:space:]]Local[[:space:]]Detail[[:space:]]and[[:space:]]Global[[:space:]]Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAMWISE_[[:space:]]Infusing[[:space:]]Wisdom[[:space:]]in[[:space:]]SAM2[[:space:]]for[[:space:]]Text-Driven[[:space:]]Video[[:space:]]Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAR3D_[[:space:]]Autoregressive[[:space:]]3D[[:space:]]Object[[:space:]]Generation[[:space:]]and[[:space:]]Understanding[[:space:]]via[[:space:]]Multi-scale[[:space:]]3D[[:space:]]VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SASep_[[:space:]]Saliency-Aware[[:space:]]Structured[[:space:]]Separation[[:space:]]of[[:space:]]Geometry[[:space:]]and[[:space:]]Feature[[:space:]]for[[:space:]]Open[[:space:]]Set[[:space:]]Learning[[:space:]]on[[:space:]]Point[[:space:]]Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SAT-HMR_[[:space:]]Real-Time[[:space:]]Multi-Person[[:space:]]3D[[:space:]]Mesh[[:space:]]Estimation[[:space:]]via[[:space:]]Scale-Adaptive[[:space:]]Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SATA_[[:space:]]Spatial[[:space:]]Autocorrelation[[:space:]]Token[[:space:]]Analysis[[:space:]]for[[:space:]]Enhancing[[:space:]]the[[:space:]]Robustness[[:space:]]of[[:space:]]Vision[[:space:]]Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SCAP_[[:space:]]Transductive[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]via[[:space:]]Supportive[[:space:]]Clique-based[[:space:]]Attribute[[:space:]]Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SCFlow2_[[:space:]]Plug-and-Play[[:space:]]Object[[:space:]]Pose[[:space:]]Refiner[[:space:]]with[[:space:]]Shape-Constraint[[:space:]]Scene[[:space:]]Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SCSA_[[:space:]]A[[:space:]]Plug-and-Play[[:space:]]Semantic[[:space:]]Continuous-Sparse[[:space:]]Attention[[:space:]]for[[:space:]]Arbitrary[[:space:]]Semantic[[:space:]]Style[[:space:]]Transfer/fae97f07-7418-476a-be91-97b643e54a78_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SCSegamba_[[:space:]]Lightweight[[:space:]]Structure-Aware[[:space:]]Vision[[:space:]]Mamba[[:space:]]for[[:space:]]Crack[[:space:]]Segmentation[[:space:]]in[[:space:]]Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SDBF_[[:space:]]Steep-Decision-Boundary[[:space:]]Fingerprinting[[:space:]]for[[:space:]]Hard-Label[[:space:]]Tampering[[:space:]]Detection[[:space:]]of[[:space:]]DNN[[:space:]]Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SDGOCC_[[:space:]]Semantic[[:space:]]and[[:space:]]Depth-Guided[[:space:]]Bird's-Eye[[:space:]]View[[:space:]]Transformation[[:space:]]for[[:space:]]3D[[:space:]]Multimodal[[:space:]]Occupancy[[:space:]]Prediction/14986d14-2105-4389-9a81-6b64509e2379_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SEAL_[[:space:]]Semantic[[:space:]]Attention[[:space:]]Learning[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Representation/382a607d-24da-4509-8433-748d58e2f990_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SEC-Prompt_SEmantic[[:space:]]Complementary[[:space:]]Prompting[[:space:]]for[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SEEN-DA_[[:space:]]SEmantic[[:space:]]ENtropy[[:space:]]guided[[:space:]]Domain-aware[[:space:]]Attention[[:space:]]for[[:space:]]Domain[[:space:]]Adaptive[[:space:]]Object[[:space:]]Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SET_[[:space:]]Spectral[[:space:]]Enhancement[[:space:]]for[[:space:]]Tiny[[:space:]]Object[[:space:]]Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SF2T_[[:space:]]Self-supervised[[:space:]]Fragment[[:space:]]Finetuning[[:space:]]of[[:space:]]Video-LLMs[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SF3D_[[:space:]]Stable[[:space:]]Fast[[:space:]]3D[[:space:]]Mesh[[:space:]]Reconstruction[[:space:]]with[[:space:]]UV-unwrapping[[:space:]]and[[:space:]]Illumination[[:space:]]Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SFDM_[[:space:]]Robust[[:space:]]Decomposition[[:space:]]of[[:space:]]Geometry[[:space:]]and[[:space:]]Reflectance[[:space:]]for[[:space:]]Realistic[[:space:]]Face[[:space:]]Rendering[[:space:]]from[[:space:]]Sparse-view[[:space:]]Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SGC-Net_[[:space:]]Stratified[[:space:]]Granular[[:space:]]Comparison[[:space:]]Network[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]HOI[[:space:]]Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SGCR_[[:space:]]Spherical[[:space:]]Gaussians[[:space:]]for[[:space:]]Efficient[[:space:]]3D[[:space:]]Curve[[:space:]]Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SGFormer_[[:space:]]Satellite-Ground[[:space:]]Fusion[[:space:]]for[[:space:]]3D[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SGSST_[[:space:]]Scaling[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Style[[:space:]]Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SIDA_[[:space:]]Social[[:space:]]Media[[:space:]]Image[[:space:]]Deepfake[[:space:]]Detection,[[:space:]]Localization[[:space:]]and[[:space:]]Explanation[[:space:]]with[[:space:]]Large[[:space:]]Multimodal[[:space:]]Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SILMM_[[:space:]]Self-Improving[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]for[[:space:]]Compositional[[:space:]]Text-to-Image[[:space:]]Generation/d6a00039-382c-4d20-89db-a164b6033fad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SINR_[[:space:]]Sparsity[[:space:]]Driven[[:space:]]Compressed[[:space:]]Implicit[[:space:]]Neural[[:space:]]Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SIR-DIFF_[[:space:]]Sparse[[:space:]]Image[[:space:]]Sets[[:space:]]Restoration[[:space:]]with[[:space:]]Multi-View[[:space:]]Diffusion[[:space:]]Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SKDream_[[:space:]]Controllable[[:space:]]Multi-view[[:space:]]and[[:space:]]3D[[:space:]]Generation[[:space:]]with[[:space:]]Arbitrary[[:space:]]Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SKE-Layout_[[:space:]]Spatial[[:space:]]Knowledge[[:space:]]Enhanced[[:space:]]Layout[[:space:]]Generation[[:space:]]with[[:space:]]LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SLADE_[[:space:]]Shielding[[:space:]]against[[:space:]]Dual[[:space:]]Exploits[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SLAM3R_[[:space:]]Real-Time[[:space:]]Dense[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]from[[:space:]]Monocular[[:space:]]RGB[[:space:]]Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SLVR_[[:space:]]Super-Light[[:space:]]Visual[[:space:]]Reconstruction[[:space:]]via[[:space:]]Blueprint[[:space:]]Controllable[[:space:]]Convolutions[[:space:]]and[[:space:]]Exploring[[:space:]]Feature[[:space:]]Diversity[[:space:]]Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SMILE_[[:space:]]Infusing[[:space:]]Spatial[[:space:]]and[[:space:]]Motion[[:space:]]Semantics[[:space:]]in[[:space:]]Masked[[:space:]]Video[[:space:]]Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SMTPD_[[:space:]]A[[:space:]]New[[:space:]]Benchmark[[:space:]]for[[:space:]]Temporal[[:space:]]Prediction[[:space:]]of[[:space:]]Social[[:space:]]Media[[:space:]]Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SOAP_[[:space:]]Vision-Centric[[:space:]]3D[[:space:]]Semantic[[:space:]]Scene[[:space:]]Completion[[:space:]]with[[:space:]]Scene-Adaptive[[:space:]]Decoder[[:space:]]and[[:space:]]Occluded[[:space:]]Region-Aware[[:space:]]View[[:space:]]Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SOGS_[[:space:]]Second-Order[[:space:]]Anchor[[:space:]]for[[:space:]]Advanced[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/c50dd783-fbc4-44df-8116-14786787e34e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SOLAMI_[[:space:]]Social[[:space:]]Vision-Language-Action[[:space:]]Modeling[[:space:]]for[[:space:]]Immersive[[:space:]]Interaction[[:space:]]with[[:space:]]3D[[:space:]]Autonomous[[:space:]]Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SOLVE_[[:space:]]Synergy[[:space:]]of[[:space:]]Language-Vision[[:space:]]and[[:space:]]End-to-End[[:space:]]Networks[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SP3D_[[:space:]]Boosting[[:space:]]Sparsely-Supervised[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]via[[:space:]]Accurate[[:space:]]Cross-Modal[[:space:]]Semantic[[:space:]]Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SPA-VL_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Safety[[:space:]]Preference[[:space:]]Alignment[[:space:]]Dataset[[:space:]]for[[:space:]]Vision[[:space:]]Language[[:space:]]Models/22cd485b-929e-4951-969d-47a8232e9e43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SPAR3D_[[:space:]]Stable[[:space:]]Point-Aware[[:space:]]Reconstruction[[:space:]]of[[:space:]]3D[[:space:]]Objects[[:space:]]from[[:space:]]Single[[:space:]]Images/3000205b-bdbd-4656-8ffc-c648c9a58288_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SPARC_[[:space:]]Score[[:space:]]Prompting[[:space:]]and[[:space:]]Adaptive[[:space:]]Fusion[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Multi-Label[[:space:]]Recognition[[:space:]]in[[:space:]]Vision-Language[[:space:]]Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SPARS3R_[[:space:]]Semantic[[:space:]]Prior[[:space:]]Alignment[[:space:]]and[[:space:]]Regularization[[:space:]]for[[:space:]]Sparse[[:space:]]3D[[:space:]]Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SPC-GS_[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]with[[:space:]]Semantic-Prompt[[:space:]]Consistency[[:space:]]for[[:space:]]Indoor[[:space:]]Open-World[[:space:]]Free-view[[:space:]]Synthesis[[:space:]]from[[:space:]]Sparse[[:space:]]Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SPMTrack_[[:space:]]Spatio-Temporal[[:space:]]Parameter-Efficient[[:space:]]Fine-Tuning[[:space:]]with[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]for[[:space:]]Scalable[[:space:]]Visual[[:space:]]Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SSHNet_[[:space:]]Unsupervised[[:space:]]Cross-modal[[:space:]]Homography[[:space:]]Estimation[[:space:]]via[[:space:]]Problem[[:space:]]Reformulation[[:space:]]and[[:space:]]Split[[:space:]]Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/STAA-SNN_[[:space:]]Spatial-Temporal[[:space:]]Attention[[:space:]]Aggregator[[:space:]]for[[:space:]]Spiking[[:space:]]Neural[[:space:]]Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/STAR-Edge_[[:space:]]Structure-aware[[:space:]]Local[[:space:]]Spherical[[:space:]]Curve[[:space:]]Representation[[:space:]]for[[:space:]]Thin-walled[[:space:]]Edge[[:space:]]Extraction[[:space:]]from[[:space:]]Unstructured[[:space:]]Point[[:space:]]Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/STCOcc_[[:space:]]Sparse[[:space:]]Spatial-Temporal[[:space:]]Cascade[[:space:]]Renovation[[:space:]]for[[:space:]]3D[[:space:]]Occupancy[[:space:]]and[[:space:]]Scene[[:space:]]Flow[[:space:]]Prediction/8219c456-c593-40fc-a736-867ed4131468_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/STDD_[[:space:]]Spatio-Temporal[[:space:]]Dual[[:space:]]Diffusion[[:space:]]for[[:space:]]Video[[:space:]]Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_content_list.json b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..790827114049ef870b138080349e8ae24df14fd8 --- /dev/null +++ b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_content_list.json @@ -0,0 +1,1280 @@ +[ + { + "type": "text", + "text": "RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing", + "text_level": 1, + "bbox": [ + 96, + 130, + 903, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhipeng Huang $^{1}$ Wangbo Yu $^{1,2}$ Xinhua Cheng $^{1}$ Chengshu Zhao $^{1}$ Yunyang Ge $^{1}$ Mingyi Guo $^{1}$ Li Yuan $^{1,2\\dagger}$ Yonghong Tian $^{1,2\\dagger}$", + "bbox": [ + 99, + 178, + 921, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Peking University \n $^{2}$ Pengcheng Laboratory", + "bbox": [ + 415, + 232, + 609, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03d79a1da8db38566108b54f9b459eed0d8275ffc8ab738c8fd2f160b1e6ab1a.jpg", + "image_caption": [ + "Figure 1. RoomPainter is capable of generating high-fidelity and consistent texture for a given room mesh." + ], + "image_footnote": [], + "bbox": [ + 94, + 323, + 903, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 549, + 326, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Indoor scene texture synthesis has garnered significant interest due to its important potential applications in virtual reality, digital media and creative arts. Existing diffusion-model-based researches either rely on per-view inpainting techniques, which are plagued by severe cross-view inconsistencies and conspicuous seams, or adopt optimization-based approaches that involve substantial computational overhead. In this work, we present RoomPainter, a framework that seamlessly integrates efficiency and consistency to achieve high-fidelity texturing of indoor scenes. The core of RoomPainter features a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a neighbor-integrated attention mechanism for zero-shot texture map generation. Using the MVIS, we firstly generate texture map for the entire room to ensure global consistency, then adopt its variant, namely Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) to repaint individual instances", + "bbox": [ + 88, + 584, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "within the room, thereby further enhancing local consistency and addressing the occlusion problem. Experiments demonstrate that RoomPainter achieves superior performance for indoor scene texture synthesis in visual quality, global consistency and generation efficiency.", + "bbox": [ + 511, + 551, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 662, + 645, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality 3D content synthesis is in great demand, owing to its diverse applications across the entertainment industry, robotic simulation and mixed-reality environments. The design of indoor scene typically requires professionals to spend dozens of hours to complete. Generating texture for indoor scene through automated methods can significantly reduce labor costs and save time. With the advancement of text-to-image diffusion models [28, 30, 33, 34], large-scale indoor scene texture synthesis has made significant progress. Although existing indoor texture synthesis methods produce high-quality results, they face several significant challenges. Specifically, these methods can be categorized into two groups: inpainting-based methods [6, 32] and optimization-based methods [7, 24, 48].", + "bbox": [ + 509, + 690, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "574", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inpainting-based methods utilize depth-to-image diffusion models [33, 55] to progressively generate texture for neighboring views, with the guidance of reference geometry and previously generated views. However, due to their per-view inpainting strategy, these approaches often result in inconsistencies in texture style across instances within the room and fail to effectively address the loop-closure problem. Optimization-based methods mitigate these inconsistencies by employing score distillation loss [29] or its variations [43] in global texture maps. However, the optimization process incurs significant time costs for indoor scene texture synthesis, due to the introduction of a large number of learnable parameters and also prone to issue of training instability. To address these challenges, we propose an efficient and high-fidelity indoor scene texturing framework, dubbed RoomPainter. The core of RoomPainter lies in a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a related view-based attention mechanism for zero-shot texture map generation. Using the MVIS, we generate a holistic texture map for the entire room, achieving better global consistency than per-view inpainting-based methods. Subsequently, to paint the textured areas raised by occlusions between objects while preserving the global consistency established in the first stage, we adopt a refine stage to repaint each instance in the room. In this stage, we apply an Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) technique at the instance scale to perform texture inpainting and refinement for each individual instance.", + "bbox": [ + 93, + 90, + 480, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are summarized below:", + "bbox": [ + 93, + 585, + 480, + 613 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose RoomPainter, an effective framework for indoor scene texture synthesis that excels in producing consistent, high-quality texture with remarkable time efficiency.", + "- We introduce Multi-View Integrated Sampling (MVIS), a zero-shot technique that leverages view-weighted texture rectification and related view-based attention to generate global consistent texture map using 2D diffusion model.", + "- We further present Multi-View Integrated Repaint Sampling (MVRS), which simultaneously refines the textured areas and inpaints the occlusion areas based on their context, thereby facilitates fine-detailed texture synthesis with local consistency.", + "- We conduct extensive experiments and compare our method with strong baselines. The results demonstrate that our method not only achieves superior performance in visual quality for indoor scene texture synthesis but also excels in generation efficiency." + ], + "bbox": [ + 89, + 626, + 478, + 893 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 517, + 90, + 661, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Generation with Diffusion Model", + "text_level": 1, + "bbox": [ + 517, + 114, + 828, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rapid advancements in image [26, 28, 33, 34, 55] and video generation models [3, 4, 44, 45, 49, 53] in recent years, the field of 3D generation has experienced significant progress. Various methods now generate objects [11, 29, 41, 43, 50-52] or scenes [10, 18, 27, 53, 54, 56] based on textual descriptions via leveraging SDS Loss and its variants [29, 43] to optimize implicit representation such as NeRF [25] and 3D Gaussian Splatting [20]. In the realm of scene generation, some approaches utilize text-guided inpainting models and depth prediction models, which incrementally construct scenes via a warping process [9, 12, 18, 27, 54]. Other methods focus on fine-tuning image generation models using 3D data, allowing them to generate multi-view consistent images or panorama for 3D tasks with the guidance of semantic images or depth maps. Moreover, some methods predict depth from a given image to initialize 3D Gaussian Splatting, followed by further optimization of the 3D Gaussian Splatting using video generation models.", + "bbox": [ + 514, + 137, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.3D Object Texturing", + "text_level": 1, + "bbox": [ + 516, + 435, + 707, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of 3D object texturing methods is to create high-quality texture for the given 3D object that align with user-provided textual descriptions. Early methods, such as [6, 32, 39, 42], utilized depth-aware Diffusion Models [26, 33, 55] to iteratively apply inpainting for each viewpoint. At each step, the texture is generated based on the texture from the previous viewpoint, gradually rotating the object to produce a complete texture. While these methods can generate texture in a relatively short time, they fail to address issues such as loop-closure and visible seams. Other works, such as [8, 24, 48], leverage differentiable rendering combined with SDS [29] to texture the given object through optimization techniques. Additionally, approaches like [5, 14, 19, 22] use UV maps as intermediaries and apply different fusion methods to integrate multiple of view independent UV maps into a consistent UV map. During the Diffusion model sampling process, this consistent UV map is used to adjust the Diffusion model's input, ensuring that the resulting texture is more consistent, with fewer visible seams and improved multi-view consistency.", + "bbox": [ + 514, + 459, + 903, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. 3D Indoor Scene Texturing", + "text_level": 1, + "bbox": [ + 516, + 771, + 756, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the task of 3D indoor scene texture generation, the objective is to create texture for a given indoor scene based on a user-provided textual description, which presents challenges such as object occlusion and multi-view consistency. SceneTex [7] defines an implicit texture field and optimizes it using VSD [43]. By employing an optimization-based approach to refine the texture field, SceneTex naturally re", + "bbox": [ + 514, + 795, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "575", + "bbox": [ + 486, + 946, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d112579350850e288373a3e791576afbb2987acfbd57e82f2cd992a0a3a9957.jpg", + "image_caption": [ + "Figure 2. Illustration of Multi-view Integrated Sampling For $N$ surrounding viewpoints in the room, centered around the room's center point and guided by the corresponding depth maps, we use a diffusion model to generate the denoised observation $\\mathcal{I}_t^n$ at timestep $t$ . This observation is then projected into UV space to obtain texture maps that corresponding to respective viewpoints. The texture maps from multiple viewpoints are dynamically merged to produce the texture map for the current timestep, which subsequently guides the sampling process for the next timestep." + ], + "image_footnote": [], + "bbox": [ + 101, + 92, + 901, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "solves the issue of multi-view consistency in 3D scene texture. Furthermore, SceneTex incorporates a cross-attention mechanism to transfer texture from visible regions to occluded areas. However, this process can result in blurriness in the occluded regions, and the optimization procedure itself is computationally intensive. When methods like Text2Tex and TEXTure [6, 32], typically used for object texture generation, are applied to scene texture generation, they can only enforce style consistency across multiple objects based on textual control. However, these methods do not ensure global style consistency for the entire scene. Alternatively, if the entire room is treated as a single object and methods such as TEXTure or Text2Tex are employed with the camera positioned at the center of the room, generating the scene texture by rotating the camera, global style consistency can be achieved. However, this approach still faces challenges such as occlusion and visible seams between regions.", + "bbox": [ + 88, + 409, + 480, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other approaches, such as [46], use panoramic images to generate a rough texture for the entire room. The camera position is then changed to generate texture for occluded regions. Subsequently, an MLP is employed to predict the texture of unseen areas based on the color and texture coordinates from the UV map. However, this method requires the scene to have an initial texture and relies on texture style transfer, meaning it cannot generate texture from scratch for a completely untextured scene.", + "bbox": [ + 89, + 681, + 482, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 829, + 191, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a 3D indoor scene mesh $\\mathcal{M}$ , our objective is to efficiently generate texture that not only align with the provided textual descriptions but also maintain global consistency.", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since our framework employs pretrained text-to-image diffusion model to generate texture for the given indoor scene mesh, we will start with an overview of the text-to-image Diffusion model and mesh rendering in Sec. 3.1. Following this, we introduce Multi-View Integrated Sampling (abbreviated as MVIS) in Sec. 3.2, which addresses the challenge of maintaining multi-view consistency while preserving high-frequency details in the diffusion model sampling process. Subsequently, In Sec. 3.3, we introduce a variant of MVIS, namely Multi-View Integrated Repaint Sampling (MVRS), to inpaint the occlusion region and simultaneously refine existing texture for each instance within the indoor scene. Finally, in Sec. 3.4, we introduce a Related View-based Attention mechanism used in the MVIS and MVRS processes, which enhances multi-view consistency during texture synthesis.", + "bbox": [ + 511, + 409, + 906, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 659, + 653, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Model Sampling. Diffusion models [17, 38] are latent variable models that comprises a forward process $q(x_{t}|x_{0})$ and a learned sampling process $p_{\\theta}(x_{t - 1}|x_t)$ . Given an initial data point $x_0$ , its noisy counterpart at time $t$ is denoted as $x_{t}$ , and the forward process $q$ is expressed as:", + "bbox": [ + 511, + 681, + 906, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {(1 - \\bar {\\alpha} _ {t})} \\epsilon , \\epsilon \\sim N (0, I), \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 767, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\in [0, \\infty)$ and the values of $\\alpha_{t}$ are defined by a scheduler to controls the amount of noise added to $x_0$ at time step $t$ . Additionally, it has $\\beta_{t} = 1 - \\alpha_{t}$ and $\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}$ . In the sampling process $p_{\\theta}$ , with the DDPM sampler [17], the diffusion model progressively denoises an initial noise image $x_{T} \\sim \\mathcal{N}(0, I)$ to clean image. Specifically, at any intermediate time step $t$ , the sampling process", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "576", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$p_{\\theta}$ can be formed as:", + "bbox": [ + 89, + 90, + 230, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} \\epsilon , \\epsilon \\sim N (0, I), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 119, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the value of $\\sigma_{t}$ and $\\mu_{t - 1}$ is given by:", + "bbox": [ + 89, + 148, + 380, + 164 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t - 1} = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} x _ {0, t} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} x _ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 175, + 482, + 209 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 224, + 482, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the estimation of the original image $x_{0,t}$ at time step $t$ is calculated as:", + "bbox": [ + 89, + 263, + 483, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {0, t} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(x _ {t} , t\\right)}{\\sqrt {\\alpha_ {t}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 303, + 482, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(x_t,t)$ is the noise that estimated by the diffusion model. As $t$ decreases from $T$ to 0, the sampling process $p_{\\theta}$ ultimately samples a clean data point $x_0$ . Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads depth-aware text-to-image diffusion models [1, 55], where the output of the model $\\epsilon_{\\theta}(x_t,y,d,t)$ is conditioned on a text prompt $y$ and a depth map $d$ .", + "bbox": [ + 89, + 349, + 483, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we utilize Stable Diffusion [28] as our base model. It performs denoising in the latent space and employs an autoencoder $\\mathcal{D}(\\mathcal{E}(\\cdot))$ for the conversion between image and latent representations. Therefore, latent $x_0$ generated by its diffusion process will be decoded to the image $\\mathcal{D}(x_0)$ through its decoder.", + "bbox": [ + 89, + 470, + 483, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mesh Rendering. Given a mesh $\\mathcal{M}$ , a texture map $\\mathcal{T}$ and a viewpoint $C$ , the rendering function $\\mathcal{R}$ can be employed to produce the rendered image as $I = \\mathcal{R}(\\mathcal{T},\\mathcal{M},C)$ . Conversely, the inverse rendering function: $\\mathcal{R}^{-1}$ can reconstruct the texture map from the image, yielding $\\mathcal{T}' = \\mathcal{R}^{-1}(I,\\mathcal{M},C)$ .", + "bbox": [ + 89, + 561, + 483, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mult-View Integrated Sampling", + "text_level": 1, + "bbox": [ + 89, + 664, + 372, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from object-level multi-view diffusion models [21, 36, 37], diffusion models that capable of ensuring multi-view consistency in scene level image generation [40] are extremely limited.", + "bbox": [ + 89, + 686, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we propose Multi-View Integrated Sampling (MVIS), grounded with a Related View-based Attention technique, to adapt a standard text-to-image model to generate view-consistent images in a training-free manner. Based on MVIS, we first implement a global texturing stage to generate consistent texture for the scene mesh.", + "bbox": [ + 89, + 750, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, given a set of predefined cameras $\\{C\\} = C^n, n = 1, \\ldots, N$ , we first render their corresponding depth maps $d_n$ and similarity masks $S_n$ , where each", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pixel in the similarity mask represents the inverted normalized value of the cosine similarity between the normal vectors of the visible faces and the viewing direction, with values ranging from [0, 1] [6, 14, 32]. Based on the similarity masks, we construct the texture weighting maps $W_{n}$ for each camera, defined as $W_{n} = \\mathcal{R}^{-1}(S_{n},\\mathcal{M},C^{n})$ , which are then used to dynamically merge per-view texture maps.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Algorithm 1, in the sampling process, we firstly sample $\\{x_T^n\\}_{n=1}^N \\sim \\{\\mathcal{N}(0,I)\\}$ for $N$ camera views. Subsequently, at each denoising step $t$ , we decode all $N$ estimated $x_{0,t}^n$ into $\\mathcal{I}_t^n$ and back project them into independent per-view texture map $\\mathcal{T}_t^n$ . We then merge these per-view texture maps to form a global texture map $\\mathcal{T}_t$ using dynamic merge [2, 22], which can be formulated as:", + "bbox": [ + 511, + 198, + 905, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {t} = \\frac {\\sum_ {n = 1} ^ {N} \\left(W _ {n} ^ {e x p (t)} \\cdot \\mathcal {T} _ {t} ^ {n}\\right)}{\\sum_ {n = 1} ^ {N} W _ {n} + \\gamma}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 315, + 903, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\exp(t)$ is used to dynamically regulate variations in consistency during the sampling process at timestep $t$ . As $t$ decreases, $\\exp(t)$ increases linearly, leading to a sharper merged texture. With the merged texture map, we again render per view images from $\\mathcal{T}_t$ , using foreground mask $M^n$ for viewpoint $C^n$ , and encode them into latent space to obtain $\\tilde{x}_{0,t}$ , which is then used to replace the previous obtained less $x_{0,t}^n$ in Eq. 3 for consistent view generation. After repeating these denoising steps, we are able to generate a consistent texture map, denote as $\\mathcal{T}_{MVIS}$ .", + "bbox": [ + 511, + 366, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-View Integrated Repaint Sampling", + "text_level": 1, + "bbox": [ + 511, + 529, + 864, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the texture map $\\mathcal{T}_{MVIS}$ maintains stylistic and consistent throughout the room, there are still areas that remain untextured due to occlusions. To address this problem, we adopt a refine stage to inpaint the occlusion areas and further enhance the texture fidelity, while preserving the global consistency established in the first stage.", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we decouple each instance in the room, resulting a total of $K$ instances $\\mathcal{M}_1,\\ldots ,\\mathcal{M}_K$ . We then design a variant of MVIS, namely multi-view integrated repaint sampling (MVRS), to perform texture inpainting and refinement for each individual instance. As shown in Fig. 3, the core of MVRS lies in incorporating diffusion-based inpainting [23] into the MVIS process. For each individual instance, given $N$ predefined camera views, we firstly render $N$ images $\\mathcal{I}^n$ along with their inpainting masks $P^n$ that indicates untextured region of them. Subsequently, for any time step $t$ in the MVIS process, we encode $\\mathcal{I}^n$ into the latent space and add noise on it to produce a noisy latent $\\tilde{x}_t^n$ , followed by sampling $x_{t}^{n}$ from MVIS, then combine them into a new latent using the inpainting mask $P^n$ , where the painted area keeps unchanged, finally passing it into the U-Net for the next denoising step. This process introduces information from the painted area into the MVIS process,", + "bbox": [ + 511, + 643, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "577", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8a3867a78cc9c482a20b45f15b4b17e79815f9232a2ae2c8a32ca8c9885d90b8.jpg", + "image_caption": [ + "Figure 3. Illustration of Multi-View Integrated Repaint Sampling(MVRS) Due to occlusion between instances, certain areas remain untextured after the first stage of texture generation. To address this issue, we perform texture generation for each instance within the room based on the painted areas. For $N$ different viewpoints of each instance, guided by the corresponding depth maps, at sampling step $t$ of MVRS, we combine the painted areas $(x_{MVIS})$ with the sampling results of MVIS at step $t + 1$ $(x_{0,t + 1})$ using a mask $P$ to form $x_{t}$ , which serves as the input for the sampling process of timestep t. Noise corresponding to the current timestep was added before mask combine. Upon completing the MVRS process, the texture map for a specific instance in the room is fully generated." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 268 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Multi-View Integrated Sampling" + ], + "code_body": "Input: mesh $\\mathcal{M}$ , text $y$ , cameras $\\{C^1,\\dots ,C^N\\}$ \nParameters: DDPM noise schedule $\\{\\sigma_t\\}_{t = T}^0$ \nInitialization: $\\{x_T^n\\}_{n = 1}^N\\sim \\{\\mathcal{N}(0,I)\\}$ \nfor $t\\in \\{T\\dots 0\\}$ do \nfor $n\\in \\{1\\dots N\\}$ do \n $\\epsilon_t^n\\gets \\epsilon_\\theta (x_t^n,y,d_n,t)$ $x_{0,t}^{n}\\gets \\frac{x_{t}^{n} - \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}^{n}}{\\sqrt{\\alpha_{t}}}$ $\\mathcal{I}_t^n\\gets \\mathcal{D}(x_{0,t}^n)$ $\\mathcal{T}_t^n\\gets \\mathcal{R}^{-1}(\\mathcal{I}_t^n,\\mathcal{M},C^n)$ \nend for \n $\\mathcal{T}_t =$ dynamic_merge( $\\{\\mathcal{T}_t^n\\}_{n = 1}^N$ ) \nif $t > 0$ then \nfor $n\\in \\{1\\dots N\\}$ do \n $\\epsilon^n\\sim \\mathcal{N}(0,I)$ $\\tilde{x}_{0,t}^{n}\\gets \\mathcal{E}(M^{n}\\odot \\mathcal{R}(\\mathcal{T}_{t},\\mathcal{M},C_{n}) + (1 - M^{n})\\odot \\mathcal{I}_{t}^{n})$ $\\mu_{t - 1}^{n}\\gets \\frac{\\sqrt{\\alpha_{t - 1}}\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\tilde{x}_{0,t}^{n} + \\frac{\\sqrt{\\alpha_{t}}(1 - \\bar{\\alpha}_{t - 1})}{1 - \\bar{\\alpha}_{t}} x_{t}^{n}$ $x_{t - 1}^{n}\\gets \\mu_{t - 1}^{n} + \\sigma_{t}\\epsilon^{n}$ \nend for \nend if \nend for \nTMVIS $=$ T0 \nreturn Texture map TMVIS", + "bbox": [ + 106, + 407, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "effectively transferring the texture of the painted area to the regions with holes while ensuring consistency across multiple viewpoints. After iterating through all instances within the room, we obtain the final output scene texture $\\mathcal{T}_{MVRS}$ .", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Related View-based Attention", + "text_level": 1, + "bbox": [ + 513, + 387, + 777, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the MVIS and MVRS process, we employ a 2D Diffusion Model for zero-shot multi-view sampling. A key challenge in maintaining view consistency is sharing information across multiple images from different viewpoints. Inspired by [15], we modify the self-attention mechanism in the diffusion model to ensure that each view incorporates sampling information from related views during the sampling process. During the sampling process, denote the queries, keys and values derived from the deep feature of the denoising U-Net for viewpoint $n$ as $Q_{n}$ , $K_{n}$ , and $V_{n}$ , respectively. For each viewpoint, assume that it has a total of $R$ associated viewpoints (view that has overlap). Subsequently, the Related View-based Attention for viewpoint $n$ is computed as:", + "bbox": [ + 511, + 410, + 906, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s o f t m a x} \\left(\\frac {Q _ {n} \\tilde {K} _ {n} ^ {T}}{\\sqrt {d}} \\tilde {V} _ {n}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 633, + 906, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{K}_n = \\left[K_1K_2\\ldots K_R\\right]^T$ and $\\tilde{V}_n = \\left[V_1V_2\\ldots V_R\\right]^T$ . In the first room-scale texturing stage, we consider the $R$ related viewpoints for view $n$ as its adjacent views, i.e., its left and right view. When sampling at instance scale, we define the related views as all the $N$ predefined camera view.", + "bbox": [ + 511, + 686, + 906, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 799, + 645, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the general implementation of the baseline methods used for comparison. Next, we provide an overview of the implementation details of our proposed method. Finally, we present our quantitative experiments, qualitative results, and ablation studies.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "578", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bf78b4c1a0cb9f69752dc94637fd32dd25d3383370c47a6a3be5ccfeeeeb724d.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons. Text2Tex-H [6] suffers from occlusion and visible seams. Text2Tex-C [6] struggles to maintain style consistency across all instances. SceneTex [7] produces unrealistic texture and results in blurry regions. In contrast, our method generates high-quality texture while preserving overall style consistency across instances in the scene. Ceilings and back-facing walls are excluded for improved visualizations." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 903, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Baseline Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 731, + 375, + 747 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Text2Tex-H [6]: This variant of Text2Tex is implemented using a holistic text prompt provided by user and treats the entire room as a cohesive entity. The camera is positioned at the center of the room, with a field of view of $60^{\\circ}$ , and the elevation angle is set to $0^{\\circ}$ , and azimuth angles uniformly sampled from the range $[0^{\\circ}, 360^{\\circ}]$ , which results in a total of twelve perspectives.", + "- Text2Tex-C [6]: For comparative purposes, we employ the original configuration of Text2Tex, wherein each" + ], + "bbox": [ + 89, + 753, + 482, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "room instance, including the interior walls and furniture, is textured using an instance-specific text prompt. And for interior walls texturing, we utilize the text prompt such as: \"A Chinese style living room, no furniture\" and sampled 12 perspectives same as Text2Tex-H. All the textured instances are subsequently combined, forming the output textured room.", + "bbox": [ + 524, + 731, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- SceneTex [7]: We compare against SceneTex using its original setup, while employing the same holistic prompt as utilized in our approach.", + "bbox": [ + 513, + 837, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "579", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4766f501f5344637745ee3097aeae72152506d41bddea829d503674777935f77.jpg", + "image_caption": [ + "\"A Cream style master bedroom\"", + "Ours (W/O Attn)", + "Figure 5. Ablation studies on the multiview-consistency module. Synthesizing texture without the Related View-based attention leads to inconsistencies across views, as shown in the leftmost column. When synthesizing texture without MVIS, noticeable inconsistencies appear across different viewpoints (middle column). In contrast, our full method samples multi-view images with significantly stronger consistency. The areas of major inconsistency in the images are highlighted with red and blue boxes. In the same column, boxes of the same color indicate that they are close to each other in 3D space. Zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 148, + 114, + 370, + 616 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc70eb03cd093f87aa7e09a55473049c93de1e773bf5fb8e032200e289c41abc.jpg", + "image_caption": [ + "Ours (W/O MVIS)" + ], + "image_footnote": [], + "bbox": [ + 405, + 114, + 624, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f98974ec98f7fd41d39d9e2178e7749a2571b4d4331d8c9102884943fba2207.jpg", + "image_caption": [ + "Ours (Full)" + ], + "image_footnote": [], + "bbox": [ + 663, + 114, + 880, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 736, + 307, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we used the SDXL model [28] as the text-to-image generator and incorporated depth information through ControlNet [1, 55]. For the text prompt, we implemented view-specific prompt combining the holistic prompt with the names of the types of objects present within the field of view. For sampling, we applied the DDPM [17] as the sampler, setting the sampling steps to 50. We used PyTorch3D [31] for rendering and texture projection. Additionally, UV atlas generation for the mesh was handled", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with Xatlas [47]. All experiments were performed on an NVIDIA RTX A6000 GPU. Further implementation details are available in the supplementary materials.", + "bbox": [ + 511, + 737, + 906, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 799, + 717, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our quantitative experiment were conducted across 10 scenes including 5 bedroom and 5 living room from the 3D-FRONT dataset [13], using five different text styles totally. To evaluate the generated texture, we calculated both the CLIP Score(CLIP) [16], which assesses the alignment", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "580", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4429644c7a34f76f38f37ecf8093098ecc3d13e5297ff6ac1c578af60c0d8c20.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparisons. We report the Generation time and 2D metrics result for quantitative comparisons. The generation time is shown in the format minutes. The 2D metrics include: CLIP Score (CLIP) [16], Aesthetic Score(AS) [35]. We show that our method produces high quality textures efficiently." + ], + "table_body": "
MethodGeneration Time(mins) ↓CLIP ↑AS ↑
Text2Tex-H [6]8.5021.584.34
Text2Tex-C [6]70.7521.934.85
SceneTex [7]2614.5021.874.75
(Ours) Full46.0023.475.03
(Ours) w/o Attn-23.325.01
(Ours) w/o MVIS-23.275.01
(Ours) w/o MVRS-22.394.40
", + "bbox": [ + 94, + 88, + 478, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between generated texture and the textual description, and the CLIP Aesthetic Score(AS) [35], which measures texture quality. We also compared the average processing time of each texture generation method in minutes. The results are summarized in Table.1. We render 10 images from novel views for each room to calculate the average of CLIP and AS. Due to the presence of numerous distinct objects within the room, the methods in baselines do not adjust the text prompts for texture generation based on the variation of objects from different viewpoints. Therefore, we employ a holistic prompt to compute the CLIP score. Since our evaluation metric involves image sampling from new viewpoints, Text2Tex-H, which fails to address the occlusion problem, generates texture containing large occluded regions, resulting in the lowest performance in terms of the metric. However, due to the limited number of sampled viewpoints, it exhibits the fastest generation speed. We configure SceneTex according to its original setup, sampling from a sphere centered at the room's center. As a result, regions not observed by SceneTex often exhibit blurriness. Meanwhile, in addition to optimizing the texture field using VSD, SceneTex also requires training LoRA, which results in the longest processing time. Text2Tex-C employs an instance-specific approach for texture generation, utilizing different prompts for each object, which results in higher performance metrics. Additionally, it incorporates viewpoint selection and texture optimization processes, leading to a longer computation time.", + "bbox": [ + 93, + 321, + 480, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 94, + 756, + 279, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide a visual comparison in fig.4. Specifically, Text2Tex-H [6] shows that simply applying object texture generation method to indoor scene texturing task cannot resolve the occlusion problem (see the large black areas in the first column of fig.4) and contain obvious seams. Text2Tex-C [6] texturing the room completely but suffering from obvious seams and global style consistency across objects. Although SceneTex introduces the use of cross-attention", + "bbox": [ + 94, + 781, + 480, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to generate texture for occluded regions, noticeable blurriness still persists in these areas. Furthermore, due to its optimization-based approach, SceneTex often produces unrealistic texture. In contrast, our method generates globally consistent texture with a unified style in the first stage. In the second stage, through per-instance inpainting and refinement, we successfully maintain global style consistency while addressing issues of occlusion and unrealistic texture.", + "bbox": [ + 516, + 92, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 516, + 220, + 674, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig.5, all images are raw samples from the final step, thus revealing inherent multi-view inconsistencies.", + "bbox": [ + 516, + 244, + 903, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the MVIS. In the first stage of texturing, we introduce MVIS to enhance the global consistency of the texture. Qualitative results from this stage are shown in the second column of Fig.5. The figure demonstrates that without MVIS, texture consistency across multiple viewpoints can be significantly impact. We further conduct quantitative ablation on dataset adopted in previous experiments. Results are shown in the seventh row of Tab.1. The quantitative experimental results clearly demonstrate that MVIS improves the quality of the generated texture.", + "bbox": [ + 516, + 273, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Related View-based Attention. We incorporate the Related View-based Attention module in both stages of texture generation to enhance consistency. Both qualitative(the first column of Fig.5) and quantitative experiments(the fifth row of the Tab.1) demonstrate that using only MVIS or MVRS is insufficient to generate consistent and high-quality texture. The Related View-based Attention module effectively enhances consistency while ensuring the quality of the generated texture.", + "bbox": [ + 516, + 425, + 903, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the MVRS. In the second stage of texturing, we use MVRS to generate texture for the occluded regions of each instance and refine the texture generated in the first stage. Quantitative experimental results (seventh row in Tab.1) show that without MVRS, the quality of the texture significantly deteriorates. This is because, without MVRS, the texture generated in the first stage suffers from low-quality results due to occlusion. We will provide additional qualitative materials demonstrating the effectiveness of MVRS in the supplementary materials.", + "bbox": [ + 516, + 561, + 903, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 516, + 724, + 630, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose an indoor scene texturing framework named RoomPainter that generates high-quality texture with multiview consistency and time efficiency. By leveraging the carefully designed modules including view-weighted texture rectification and related view-based attention, proposed view-integrated diffusion preserves the cross-view consistency during the diffusion process. Qualitative and quantitative results demonstrate the effectiveness of RoomPainter in indoor scene texturing, especially in global consistency and generation efficiency.", + "bbox": [ + 516, + 750, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "581", + "bbox": [ + 486, + 946, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] xinsir/controlnet-depth-sdx1-1.0. https://huggingface.co/xinsir/controlnet-depth-sdx1-1.0, 2023.4.7", + "[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 4", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563-22575, 2023. 2", + "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2", + "[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 1, 2, 3, 4, 6, 8", + "[7] Dave Zhenyu Chen, Haoxuan Li, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Scenetex: High-quality texture synthesis for indoor scenes via diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21081-21091, 2024. 1, 2, 6, 8", + "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 2", + "[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2", + "[10] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2920–2929, 2023. 2", + "[11] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene, 2025. 2", + "[12] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 7", + "[14] Chenjian Gao, Boyan Jiang, Xinghui Li, Yingpeng Zhang, and Qian Yu. Genesistex: Adapting image denoising diffusion to texture space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4620-4629, 2024. 2, 4", + "[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 5", + "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7, 8", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7", + "[18] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2", + "[19] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Texgen: Text-guided 3d texture generation with multi-view sampling and resampling. arXiv preprint arXiv:2408.01291, 2024. 2", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2", + "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 4", + "[22] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 2, 4", + "[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 4", + "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 2", + "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "582", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2", + "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2", + "[27] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2", + "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 4, 7", + "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1", + "[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 7", + "[32] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 1, 2, 3, 4", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2", + "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 8", + "[36] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 4", + "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4", + "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + ], + "bbox": [ + 91, + 92, + 482, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[39] Jiaxiang Tang, Ruijie Lu, Xiaokang Chen, Xiang Wen, Gang Zeng, and Ziwei Liu. Intex: Interactive text-to-texture synthesis via unified depth-aware inpainting. arXiv preprint arXiv:2403.11878, 2024. 2", + "[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 4", + "[41] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2", + "[42] Qi Wang, Ruijie Lu, Xudong Xu, Jingbo Wang, Michael Yu Wang, Bo Dai, Gang Zeng, and Dan Xu. Roomtex: Texturing compositional indoor scenes via iterative inpainting. arXiv preprint arXiv:2406.02461, 2024. 2", + "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[44] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2", + "[45] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2", + "[46] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 3", + "[47] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 7", + "[48] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 1, 2", + "[49] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 2", + "[50] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. 2023. 2", + "[51] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian." + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "583", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744, 2023.", + "[52] Wangbo Yu, Chaoran Feng, Jiye Tang, Jiashu Yang, Zhenyu Tang, Xu Jia, Yuchao Yang, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2", + "[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2", + "[54] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 30(12):7749-7762, 2024. 2", + "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 7", + "[56] Haiyang Zhou, Xinhua Cheng, Wangbo Yu, Yonghong Tian, and Li Yuan. Holodreamer: Holistic 3d panoramic world generation from text descriptions. arXiv preprint arXiv:2407.15187, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "584", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_model.json b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd897aaf7f7c021408ec5f87491f9ea2e7b53276 --- /dev/null +++ b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_model.json @@ -0,0 +1,1993 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.904, + 0.154 + ], + "angle": 0, + "content": "RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.179, + 0.923, + 0.219 + ], + "angle": 0, + "content": "Zhipeng Huang\\(^{1}\\) Wangbo Yu\\(^{1,2}\\) Xinhua Cheng\\(^{1}\\) Chengshu Zhao\\(^{1}\\) Yunyang Ge\\(^{1}\\) Mingyi Guo\\(^{1}\\) Li Yuan\\(^{1,2\\dagger}\\) Yonghong Tian\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.233, + 0.61, + 0.272 + ], + "angle": 0, + "content": "\\(^{1}\\)Peking University \n\\(^{2}\\)Pengcheng Laboratory" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.324, + 0.905, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.515, + 0.816, + 0.53 + ], + "angle": 0, + "content": "Figure 1. RoomPainter is capable of generating high-fidelity and consistent texture for a given room mesh." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.55, + 0.327, + 0.566 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.585, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Indoor scene texture synthesis has garnered significant interest due to its important potential applications in virtual reality, digital media and creative arts. Existing diffusion-model-based researches either rely on per-view inpainting techniques, which are plagued by severe cross-view inconsistencies and conspicuous seams, or adopt optimization-based approaches that involve substantial computational overhead. In this work, we present RoomPainter, a framework that seamlessly integrates efficiency and consistency to achieve high-fidelity texturing of indoor scenes. The core of RoomPainter features a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a neighbor-integrated attention mechanism for zero-shot texture map generation. Using the MVIS, we firstly generate texture map for the entire room to ensure global consistency, then adopt its variant, namely Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) to repaint individual instances" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.629 + ], + "angle": 0, + "content": "within the room, thereby further enhancing local consistency and addressing the occlusion problem. Experiments demonstrate that RoomPainter achieves superior performance for indoor scene texture synthesis in visual quality, global consistency and generation efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.663, + 0.646, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.691, + 0.907, + 0.903 + ], + "angle": 0, + "content": "High-quality 3D content synthesis is in great demand, owing to its diverse applications across the entertainment industry, robotic simulation and mixed-reality environments. The design of indoor scene typically requires professionals to spend dozens of hours to complete. Generating texture for indoor scene through automated methods can significantly reduce labor costs and save time. With the advancement of text-to-image diffusion models [28, 30, 33, 34], large-scale indoor scene texture synthesis has made significant progress. Although existing indoor texture synthesis methods produce high-quality results, they face several significant challenges. Specifically, these methods can be categorized into two groups: inpainting-based methods [6, 32] and optimization-based methods [7, 24, 48]." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "574" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.573 + ], + "angle": 0, + "content": "Inpainting-based methods utilize depth-to-image diffusion models [33, 55] to progressively generate texture for neighboring views, with the guidance of reference geometry and previously generated views. However, due to their per-view inpainting strategy, these approaches often result in inconsistencies in texture style across instances within the room and fail to effectively address the loop-closure problem. Optimization-based methods mitigate these inconsistencies by employing score distillation loss [29] or its variations [43] in global texture maps. However, the optimization process incurs significant time costs for indoor scene texture synthesis, due to the introduction of a large number of learnable parameters and also prone to issue of training instability. To address these challenges, we propose an efficient and high-fidelity indoor scene texturing framework, dubbed RoomPainter. The core of RoomPainter lies in a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a related view-based attention mechanism for zero-shot texture map generation. Using the MVIS, we generate a holistic texture map for the entire room, achieving better global consistency than per-view inpainting-based methods. Subsequently, to paint the textured areas raised by occlusions between objects while preserving the global consistency established in the first stage, we adopt a refine stage to repaint each instance in the room. In this stage, we apply an Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) technique at the instance scale to perform texture inpainting and refinement for each individual instance." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.586, + 0.482, + 0.614 + ], + "angle": 0, + "content": "The main contributions of our work are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.627, + 0.48, + 0.683 + ], + "angle": 0, + "content": "- We propose RoomPainter, an effective framework for indoor scene texture synthesis that excels in producing consistent, high-quality texture with remarkable time efficiency." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.687, + 0.48, + 0.744 + ], + "angle": 0, + "content": "- We introduce Multi-View Integrated Sampling (MVIS), a zero-shot technique that leverages view-weighted texture rectification and related view-based attention to generate global consistent texture map using 2D diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.748, + 0.48, + 0.819 + ], + "angle": 0, + "content": "- We further present Multi-View Integrated Repaint Sampling (MVRS), which simultaneously refines the textured areas and inpaints the occlusion areas based on their context, thereby facilitates fine-detailed texture synthesis with local consistency." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.823, + 0.48, + 0.895 + ], + "angle": 0, + "content": "- We conduct extensive experiments and compare our method with strong baselines. The results demonstrate that our method not only achieves superior performance in visual quality for indoor scene texture synthesis but also excels in generation efficiency." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.627, + 0.48, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.091, + 0.662, + 0.105 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.116, + 0.829, + 0.129 + ], + "angle": 0, + "content": "2.1. 3D Generation with Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.138, + 0.905, + 0.424 + ], + "angle": 0, + "content": "With the rapid advancements in image [26, 28, 33, 34, 55] and video generation models [3, 4, 44, 45, 49, 53] in recent years, the field of 3D generation has experienced significant progress. Various methods now generate objects [11, 29, 41, 43, 50-52] or scenes [10, 18, 27, 53, 54, 56] based on textual descriptions via leveraging SDS Loss and its variants [29, 43] to optimize implicit representation such as NeRF [25] and 3D Gaussian Splatting [20]. In the realm of scene generation, some approaches utilize text-guided inpainting models and depth prediction models, which incrementally construct scenes via a warping process [9, 12, 18, 27, 54]. Other methods focus on fine-tuning image generation models using 3D data, allowing them to generate multi-view consistent images or panorama for 3D tasks with the guidance of semantic images or depth maps. Moreover, some methods predict depth from a given image to initialize 3D Gaussian Splatting, followed by further optimization of the 3D Gaussian Splatting using video generation models." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.436, + 0.708, + 0.451 + ], + "angle": 0, + "content": "2.2.3D Object Texturing" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.46, + 0.905, + 0.761 + ], + "angle": 0, + "content": "The goal of 3D object texturing methods is to create high-quality texture for the given 3D object that align with user-provided textual descriptions. Early methods, such as [6, 32, 39, 42], utilized depth-aware Diffusion Models [26, 33, 55] to iteratively apply inpainting for each viewpoint. At each step, the texture is generated based on the texture from the previous viewpoint, gradually rotating the object to produce a complete texture. While these methods can generate texture in a relatively short time, they fail to address issues such as loop-closure and visible seams. Other works, such as [8, 24, 48], leverage differentiable rendering combined with SDS [29] to texture the given object through optimization techniques. Additionally, approaches like [5, 14, 19, 22] use UV maps as intermediaries and apply different fusion methods to integrate multiple of view independent UV maps into a consistent UV map. During the Diffusion model sampling process, this consistent UV map is used to adjust the Diffusion model's input, ensuring that the resulting texture is more consistent, with fewer visible seams and improved multi-view consistency." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.772, + 0.757, + 0.787 + ], + "angle": 0, + "content": "2.3. 3D Indoor Scene Texturing" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.796, + 0.904, + 0.9 + ], + "angle": 0, + "content": "In the task of 3D indoor scene texture generation, the objective is to create texture for a given indoor scene based on a user-provided textual description, which presents challenges such as object occlusion and multi-view consistency. SceneTex [7] defines an implicit texture field and optimizes it using VSD [43]. By employing an optimization-based approach to refine the texture field, SceneTex naturally re" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.947, + 0.51, + 0.956 + ], + "angle": 0, + "content": "575" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.093, + 0.902, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.313, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Illustration of Multi-view Integrated Sampling For \\(N\\) surrounding viewpoints in the room, centered around the room's center point and guided by the corresponding depth maps, we use a diffusion model to generate the denoised observation \\(\\mathcal{I}_t^n\\) at timestep \\(t\\). This observation is then projected into UV space to obtain texture maps that corresponding to respective viewpoints. The texture maps from multiple viewpoints are dynamically merged to produce the texture map for the current timestep, which subsequently guides the sampling process for the next timestep." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.41, + 0.482, + 0.681 + ], + "angle": 0, + "content": "solves the issue of multi-view consistency in 3D scene texture. Furthermore, SceneTex incorporates a cross-attention mechanism to transfer texture from visible regions to occluded areas. However, this process can result in blurriness in the occluded regions, and the optimization procedure itself is computationally intensive. When methods like Text2Tex and TEXTure [6, 32], typically used for object texture generation, are applied to scene texture generation, they can only enforce style consistency across multiple objects based on textual control. However, these methods do not ensure global style consistency for the entire scene. Alternatively, if the entire room is treated as a single object and methods such as TEXTure or Text2Tex are employed with the camera positioned at the center of the room, generating the scene texture by rotating the camera, global style consistency can be achieved. However, this approach still faces challenges such as occlusion and visible seams between regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.818 + ], + "angle": 0, + "content": "Other approaches, such as [46], use panoramic images to generate a rough texture for the entire room. The camera position is then changed to generate texture for occluded regions. Subsequently, an MLP is employed to predict the texture of unseen areas based on the color and texture coordinates from the UV map. However, this method requires the scene to have an initial texture and relies on texture style transfer, meaning it cannot generate texture from scratch for a completely untextured scene." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.192, + 0.847 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Given a 3D indoor scene mesh \\(\\mathcal{M}\\), our objective is to efficiently generate texture that not only align with the provided textual descriptions but also maintain global consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Since our framework employs pretrained text-to-image diffusion model to generate texture for the given indoor scene mesh, we will start with an overview of the text-to-image Diffusion model and mesh rendering in Sec. 3.1. Following this, we introduce Multi-View Integrated Sampling (abbreviated as MVIS) in Sec. 3.2, which addresses the challenge of maintaining multi-view consistency while preserving high-frequency details in the diffusion model sampling process. Subsequently, In Sec. 3.3, we introduce a variant of MVIS, namely Multi-View Integrated Repaint Sampling (MVRS), to inpaint the occlusion region and simultaneously refine existing texture for each instance within the indoor scene. Finally, in Sec. 3.4, we introduce a Related View-based Attention mechanism used in the MVIS and MVRS processes, which enhances multi-view consistency during texture synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.655, + 0.675 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Diffusion Model Sampling. Diffusion models [17, 38] are latent variable models that comprises a forward process \\( q(x_{t}|x_{0}) \\) and a learned sampling process \\( p_{\\theta}(x_{t - 1}|x_t) \\). Given an initial data point \\( x_0 \\), its noisy counterpart at time \\( t \\) is denoted as \\( x_{t} \\), and the forward process \\( q \\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.768, + 0.907, + 0.788 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {(1 - \\bar {\\alpha} _ {t})} \\epsilon , \\epsilon \\sim N (0, I), \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(t \\in [0, \\infty)\\) and the values of \\(\\alpha_{t}\\) are defined by a scheduler to controls the amount of noise added to \\(x_0\\) at time step \\(t\\). Additionally, it has \\(\\beta_{t} = 1 - \\alpha_{t}\\) and \\(\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}\\). In the sampling process \\(p_{\\theta}\\), with the DDPM sampler [17], the diffusion model progressively denoises an initial noise image \\(x_{T} \\sim \\mathcal{N}(0, I)\\) to clean image. Specifically, at any intermediate time step \\(t\\), the sampling process" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.232, + 0.106 + ], + "angle": 0, + "content": "\\(p_{\\theta}\\) can be formed as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.121, + 0.483, + 0.137 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} \\epsilon , \\epsilon \\sim N (0, I), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.15, + 0.381, + 0.165 + ], + "angle": 0, + "content": "where the value of \\(\\sigma_{t}\\) and \\(\\mu_{t - 1}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.176, + 0.483, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t - 1} = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} x _ {0, t} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} x _ {t}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.226, + 0.483, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.294 + ], + "angle": 0, + "content": "and the estimation of the original image \\( x_{0,t} \\) at time step \\( t \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.304, + 0.483, + 0.339 + ], + "angle": 0, + "content": "\\[\nx _ {0, t} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(x _ {t} , t\\right)}{\\sqrt {\\alpha_ {t}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.472 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(x_t,t)\\) is the noise that estimated by the diffusion model. As \\(t\\) decreases from \\(T\\) to 0, the sampling process \\(p_{\\theta}\\) ultimately samples a clean data point \\(x_0\\). Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads depth-aware text-to-image diffusion models [1, 55], where the output of the model \\(\\epsilon_{\\theta}(x_t,y,d,t)\\) is conditioned on a text prompt \\(y\\) and a depth map \\(d\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.563 + ], + "angle": 0, + "content": "In this paper, we utilize Stable Diffusion [28] as our base model. It performs denoising in the latent space and employs an autoencoder \\(\\mathcal{D}(\\mathcal{E}(\\cdot))\\) for the conversion between image and latent representations. Therefore, latent \\(x_0\\) generated by its diffusion process will be decoded to the image \\(\\mathcal{D}(x_0)\\) through its decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Mesh Rendering. Given a mesh \\(\\mathcal{M}\\), a texture map \\(\\mathcal{T}\\) and a viewpoint \\(C\\), the rendering function \\(\\mathcal{R}\\) can be employed to produce the rendered image as \\(I = \\mathcal{R}(\\mathcal{T},\\mathcal{M},C)\\). Conversely, the inverse rendering function: \\(\\mathcal{R}^{-1}\\) can reconstruct the texture map from the image, yielding \\(\\mathcal{T}' = \\mathcal{R}^{-1}(I,\\mathcal{M},C)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.373, + 0.682 + ], + "angle": 0, + "content": "3.2. Mult-View Integrated Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Different from object-level multi-view diffusion models [21, 36, 37], diffusion models that capable of ensuring multi-view consistency in scene level image generation [40] are extremely limited." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.854 + ], + "angle": 0, + "content": "To address this challenge, we propose Multi-View Integrated Sampling (MVIS), grounded with a Related View-based Attention technique, to adapt a standard text-to-image model to generate view-consistent images in a training-free manner. Based on MVIS, we first implement a global texturing stage to generate consistent texture for the scene mesh." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 2, given a set of predefined cameras \\(\\{C\\} = C^n, n = 1, \\ldots, N\\), we first render their corresponding depth maps \\(d_n\\) and similarity masks \\(S_n\\), where each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "pixel in the similarity mask represents the inverted normalized value of the cosine similarity between the normal vectors of the visible faces and the viewing direction, with values ranging from [0, 1] [6, 14, 32]. Based on the similarity masks, we construct the texture weighting maps \\( W_{n} \\) for each camera, defined as \\( W_{n} = \\mathcal{R}^{-1}(S_{n},\\mathcal{M},C^{n}) \\), which are then used to dynamically merge per-view texture maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.906, + 0.304 + ], + "angle": 0, + "content": "As shown in Algorithm 1, in the sampling process, we firstly sample \\(\\{x_T^n\\}_{n=1}^N \\sim \\{\\mathcal{N}(0,I)\\}\\) for \\(N\\) camera views. Subsequently, at each denoising step \\(t\\), we decode all \\(N\\) estimated \\(x_{0,t}^n\\) into \\(\\mathcal{I}_t^n\\) and back project them into independent per-view texture map \\(\\mathcal{T}_t^n\\). We then merge these per-view texture maps to form a global texture map \\(\\mathcal{T}_t\\) using dynamic merge [2, 22], which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.316, + 0.905, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {t} = \\frac {\\sum_ {n = 1} ^ {N} \\left(W _ {n} ^ {e x p (t)} \\cdot \\mathcal {T} _ {t} ^ {n}\\right)}{\\sum_ {n = 1} ^ {N} W _ {n} + \\gamma}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.907, + 0.519 + ], + "angle": 0, + "content": "Here, \\( \\exp(t) \\) is used to dynamically regulate variations in consistency during the sampling process at timestep \\( t \\). As \\( t \\) decreases, \\( \\exp(t) \\) increases linearly, leading to a sharper merged texture. With the merged texture map, we again render per view images from \\( \\mathcal{T}_t \\), using foreground mask \\( M^n \\) for viewpoint \\( C^n \\), and encode them into latent space to obtain \\( \\tilde{x}_{0,t} \\), which is then used to replace the previous obtained less \\( x_{0,t}^n \\) in Eq. 3 for consistent view generation. After repeating these denoising steps, we are able to generate a consistent texture map, denote as \\( \\mathcal{T}_{MVIS} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.866, + 0.546 + ], + "angle": 0, + "content": "3.3. Multi-View Integrated Repaint Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Although the texture map \\(\\mathcal{T}_{MVIS}\\) maintains stylistic and consistent throughout the room, there are still areas that remain untextured due to occlusions. To address this problem, we adopt a refine stage to inpaint the occlusion areas and further enhance the texture fidelity, while preserving the global consistency established in the first stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Specifically, we decouple each instance in the room, resulting a total of \\(K\\) instances \\(\\mathcal{M}_1,\\ldots ,\\mathcal{M}_K\\). We then design a variant of MVIS, namely multi-view integrated repaint sampling (MVRS), to perform texture inpainting and refinement for each individual instance. As shown in Fig. 3, the core of MVRS lies in incorporating diffusion-based inpainting [23] into the MVIS process. For each individual instance, given \\(N\\) predefined camera views, we firstly render \\(N\\) images \\(\\mathcal{I}^n\\) along with their inpainting masks \\(P^n\\) that indicates untextured region of them. Subsequently, for any time step \\(t\\) in the MVIS process, we encode \\(\\mathcal{I}^n\\) into the latent space and add noise on it to produce a noisy latent \\(\\tilde{x}_t^n\\), followed by sampling \\(x_{t}^{n}\\) from MVIS, then combine them into a new latent using the inpainting mask \\(P^n\\), where the painted area keeps unchanged, finally passing it into the U-Net for the next denoising step. This process introduces information from the painted area into the MVIS process," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "577" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Figure 3. Illustration of Multi-View Integrated Repaint Sampling(MVRS) Due to occlusion between instances, certain areas remain untextured after the first stage of texture generation. To address this issue, we perform texture generation for each instance within the room based on the painted areas. For \\(N\\) different viewpoints of each instance, guided by the corresponding depth maps, at sampling step \\(t\\) of MVRS, we combine the painted areas \\((x_{MVIS})\\) with the sampling results of MVIS at step \\(t + 1\\) \\((x_{0,t + 1})\\) using a mask \\(P\\) to form \\(x_{t}\\), which serves as the input for the sampling process of timestep t. Noise corresponding to the current timestep was added before mask combine. Upon completing the MVRS process, the texture map for a specific instance in the room is fully generated." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.389, + 0.398, + 0.404 + ], + "angle": 0, + "content": "Algorithm 1 Multi-View Integrated Sampling" + }, + { + "type": "algorithm", + "bbox": [ + 0.107, + 0.408, + 0.484, + 0.781 + ], + "angle": 0, + "content": "Input: mesh \\(\\mathcal{M}\\) , text \\(y\\) , cameras \\(\\{C^1,\\dots ,C^N\\}\\) \nParameters: DDPM noise schedule \\(\\{\\sigma_t\\}_{t = T}^0\\) \nInitialization: \\(\\{x_T^n\\}_{n = 1}^N\\sim \\{\\mathcal{N}(0,I)\\}\\) \nfor \\(t\\in \\{T\\dots 0\\}\\) do \nfor \\(n\\in \\{1\\dots N\\}\\) do \n\\(\\epsilon_t^n\\gets \\epsilon_\\theta (x_t^n,y,d_n,t)\\) \n\\(x_{0,t}^{n}\\gets \\frac{x_{t}^{n} - \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}^{n}}{\\sqrt{\\alpha_{t}}}\\) \n\\(\\mathcal{I}_t^n\\gets \\mathcal{D}(x_{0,t}^n)\\) \n\\(\\mathcal{T}_t^n\\gets \\mathcal{R}^{-1}(\\mathcal{I}_t^n,\\mathcal{M},C^n)\\) \nend for \n\\(\\mathcal{T}_t =\\) dynamic_merge(\\(\\{\\mathcal{T}_t^n\\}_{n = 1}^N\\) ) \nif \\(t > 0\\) then \nfor \\(n\\in \\{1\\dots N\\}\\) do \n\\(\\epsilon^n\\sim \\mathcal{N}(0,I)\\) \n\\(\\tilde{x}_{0,t}^{n}\\gets \\mathcal{E}(M^{n}\\odot \\mathcal{R}(\\mathcal{T}_{t},\\mathcal{M},C_{n}) + (1 - M^{n})\\odot \\mathcal{I}_{t}^{n})\\) \n\\(\\mu_{t - 1}^{n}\\gets \\frac{\\sqrt{\\alpha_{t - 1}}\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\tilde{x}_{0,t}^{n} + \\frac{\\sqrt{\\alpha_{t}}(1 - \\bar{\\alpha}_{t - 1})}{1 - \\bar{\\alpha}_{t}} x_{t}^{n}\\) \n\\(x_{t - 1}^{n}\\gets \\mu_{t - 1}^{n} + \\sigma_{t}\\epsilon^{n}\\) \nend for \nend if \nend for \nTMVIS \\(=\\) T0 \nreturn Texture map TMVIS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "effectively transferring the texture of the painted area to the regions with holes while ensuring consistency across multiple viewpoints. After iterating through all instances within the room, we obtain the final output scene texture \\(\\mathcal{T}_{MVRS}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.388, + 0.779, + 0.403 + ], + "angle": 0, + "content": "3.4. Related View-based Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.623 + ], + "angle": 0, + "content": "In the MVIS and MVRS process, we employ a 2D Diffusion Model for zero-shot multi-view sampling. A key challenge in maintaining view consistency is sharing information across multiple images from different viewpoints. Inspired by [15], we modify the self-attention mechanism in the diffusion model to ensure that each view incorporates sampling information from related views during the sampling process. During the sampling process, denote the queries, keys and values derived from the deep feature of the denoising U-Net for viewpoint \\( n \\) as \\( Q_{n} \\), \\( K_{n} \\), and \\( V_{n} \\), respectively. For each viewpoint, assume that it has a total of \\( R \\) associated viewpoints (view that has overlap). Subsequently, the Related View-based Attention for viewpoint \\( n \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.634, + 0.907, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s o f t m a x} \\left(\\frac {Q _ {n} \\tilde {K} _ {n} ^ {T}}{\\sqrt {d}} \\tilde {V} _ {n}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.907, + 0.786 + ], + "angle": 0, + "content": "where \\(\\tilde{K}_n = \\left[K_1K_2\\ldots K_R\\right]^T\\) and \\(\\tilde{V}_n = \\left[V_1V_2\\ldots V_R\\right]^T\\). In the first room-scale texturing stage, we consider the \\(R\\) related viewpoints for view \\(n\\) as its adjacent views, i.e., its left and right view. When sampling at instance scale, we define the related views as all the \\(N\\) predefined camera view." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.646, + 0.817 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we first describe the general implementation of the baseline methods used for comparison. Next, we provide an overview of the implementation details of our proposed method. Finally, we present our quantitative experiments, qualitative results, and ablation studies." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.905, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.65, + 0.908, + 0.707 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons. Text2Tex-H [6] suffers from occlusion and visible seams. Text2Tex-C [6] struggles to maintain style consistency across all instances. SceneTex [7] produces unrealistic texture and results in blurry regions. In contrast, our method generates high-quality texture while preserving overall style consistency across instances in the scene. Ceilings and back-facing walls are excluded for improved visualizations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.376, + 0.748 + ], + "angle": 0, + "content": "4.1. Baseline Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.755, + 0.483, + 0.86 + ], + "angle": 0, + "content": "- Text2Tex-H [6]: This variant of Text2Tex is implemented using a holistic text prompt provided by user and treats the entire room as a cohesive entity. The camera is positioned at the center of the room, with a field of view of \\(60^{\\circ}\\), and the elevation angle is set to \\(0^{\\circ}\\), and azimuth angles uniformly sampled from the range \\([0^{\\circ}, 360^{\\circ}]\\), which results in a total of twelve perspectives." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.861, + 0.483, + 0.891 + ], + "angle": 0, + "content": "- Text2Tex-C [6]: For comparative purposes, we employ the original configuration of Text2Tex, wherein each" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.755, + 0.483, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.732, + 0.907, + 0.837 + ], + "angle": 0, + "content": "room instance, including the interior walls and furniture, is textured using an instance-specific text prompt. And for interior walls texturing, we utilize the text prompt such as: \"A Chinese style living room, no furniture\" and sampled 12 perspectives same as Text2Tex-H. All the textured instances are subsequently combined, forming the output textured room." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.838, + 0.907, + 0.884 + ], + "angle": 0, + "content": "- SceneTex [7]: We compare against SceneTex using its original setup, while employing the same holistic prompt as utilized in our approach." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "579" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.226, + 0.129, + 0.489 + ], + "angle": 270, + "content": "\"A Cream style master bedroom\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.101, + 0.314, + 0.114 + ], + "angle": 0, + "content": "Ours (W/O Attn)" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.116, + 0.371, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.1, + 0.575, + 0.114 + ], + "angle": 0, + "content": "Ours (W/O MVIS)" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.116, + 0.625, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.101, + 0.808, + 0.115 + ], + "angle": 0, + "content": "Ours (Full)" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.116, + 0.882, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.641, + 0.907, + 0.711 + ], + "angle": 0, + "content": "Figure 5. Ablation studies on the multiview-consistency module. Synthesizing texture without the Related View-based attention leads to inconsistencies across views, as shown in the leftmost column. When synthesizing texture without MVIS, noticeable inconsistencies appear across different viewpoints (middle column). In contrast, our full method samples multi-view images with significantly stronger consistency. The areas of major inconsistency in the images are highlighted with red and blue boxes. In the same column, boxes of the same color indicate that they are close to each other in 3D space. Zoom in for the best view." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.737, + 0.308, + 0.753 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this experiment, we used the SDXL model [28] as the text-to-image generator and incorporated depth information through ControlNet [1, 55]. For the text prompt, we implemented view-specific prompt combining the holistic prompt with the names of the types of objects present within the field of view. For sampling, we applied the DDPM [17] as the sampler, setting the sampling steps to 50. We used PyTorch3D [31] for rendering and texture projection. Additionally, UV atlas generation for the mesh was handled" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.738, + 0.907, + 0.784 + ], + "angle": 0, + "content": "with Xatlas [47]. All experiments were performed on an NVIDIA RTX A6000 GPU. Further implementation details are available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.718, + 0.816 + ], + "angle": 0, + "content": "4.3. Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our quantitative experiment were conducted across 10 scenes including 5 bedroom and 5 living room from the 3D-FRONT dataset [13], using five different text styles totally. To evaluate the generated texture, we calculated both the CLIP Score(CLIP) [16], which assesses the alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "580" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.212 + ], + "angle": 0, + "content": "
MethodGeneration Time(mins) ↓CLIP ↑AS ↑
Text2Tex-H [6]8.5021.584.34
Text2Tex-C [6]70.7521.934.85
SceneTex [7]2614.5021.874.75
(Ours) Full46.0023.475.03
(Ours) w/o Attn-23.325.01
(Ours) w/o MVIS-23.275.01
(Ours) w/o MVRS-22.394.40
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.223, + 0.482, + 0.292 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons. We report the Generation time and 2D metrics result for quantitative comparisons. The generation time is shown in the format minutes. The 2D metrics include: CLIP Score (CLIP) [16], Aesthetic Score(AS) [35]. We show that our method produces high quality textures efficiently." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.322, + 0.482, + 0.743 + ], + "angle": 0, + "content": "between generated texture and the textual description, and the CLIP Aesthetic Score(AS) [35], which measures texture quality. We also compared the average processing time of each texture generation method in minutes. The results are summarized in Table.1. We render 10 images from novel views for each room to calculate the average of CLIP and AS. Due to the presence of numerous distinct objects within the room, the methods in baselines do not adjust the text prompts for texture generation based on the variation of objects from different viewpoints. Therefore, we employ a holistic prompt to compute the CLIP score. Since our evaluation metric involves image sampling from new viewpoints, Text2Tex-H, which fails to address the occlusion problem, generates texture containing large occluded regions, resulting in the lowest performance in terms of the metric. However, due to the limited number of sampled viewpoints, it exhibits the fastest generation speed. We configure SceneTex according to its original setup, sampling from a sphere centered at the room's center. As a result, regions not observed by SceneTex often exhibit blurriness. Meanwhile, in addition to optimizing the texture field using VSD, SceneTex also requires training LoRA, which results in the longest processing time. Text2Tex-C employs an instance-specific approach for texture generation, utilizing different prompts for each object, which results in higher performance metrics. Additionally, it incorporates viewpoint selection and texture optimization processes, leading to a longer computation time." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.757, + 0.281, + 0.772 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.782, + 0.481, + 0.9 + ], + "angle": 0, + "content": "We provide a visual comparison in fig.4. Specifically, Text2Tex-H [6] shows that simply applying object texture generation method to indoor scene texturing task cannot resolve the occlusion problem (see the large black areas in the first column of fig.4) and contain obvious seams. Text2Tex-C [6] texturing the room completely but suffering from obvious seams and global style consistency across objects. Although SceneTex introduces the use of cross-attention" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.093, + 0.904, + 0.212 + ], + "angle": 0, + "content": "to generate texture for occluded regions, noticeable blurriness still persists in these areas. Furthermore, due to its optimization-based approach, SceneTex often produces unrealistic texture. In contrast, our method generates globally consistent texture with a unified style in the first stage. In the second stage, through per-instance inpainting and refinement, we successfully maintain global style consistency while addressing issues of occlusion and unrealistic texture." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.222, + 0.676, + 0.235 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.245, + 0.904, + 0.273 + ], + "angle": 0, + "content": "As shown in Fig.5, all images are raw samples from the final step, thus revealing inherent multi-view inconsistencies." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.424 + ], + "angle": 0, + "content": "Effectiveness of the MVIS. In the first stage of texturing, we introduce MVIS to enhance the global consistency of the texture. Qualitative results from this stage are shown in the second column of Fig.5. The figure demonstrates that without MVIS, texture consistency across multiple viewpoints can be significantly impact. We further conduct quantitative ablation on dataset adopted in previous experiments. Results are shown in the seventh row of Tab.1. The quantitative experimental results clearly demonstrate that MVIS improves the quality of the generated texture." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.426, + 0.904, + 0.56 + ], + "angle": 0, + "content": "Effectiveness of the Related View-based Attention. We incorporate the Related View-based Attention module in both stages of texture generation to enhance consistency. Both qualitative(the first column of Fig.5) and quantitative experiments(the fifth row of the Tab.1) demonstrate that using only MVIS or MVRS is insufficient to generate consistent and high-quality texture. The Related View-based Attention module effectively enhances consistency while ensuring the quality of the generated texture." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.562, + 0.904, + 0.711 + ], + "angle": 0, + "content": "Effectiveness of the MVRS. In the second stage of texturing, we use MVRS to generate texture for the occluded regions of each instance and refine the texture generated in the first stage. Quantitative experimental results (seventh row in Tab.1) show that without MVRS, the quality of the texture significantly deteriorates. This is because, without MVRS, the texture generated in the first stage suffers from low-quality results due to occlusion. We will provide additional qualitative materials demonstrating the effectiveness of MVRS in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.726, + 0.631, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.751, + 0.904, + 0.9 + ], + "angle": 0, + "content": "We propose an indoor scene texturing framework named RoomPainter that generates high-quality texture with multiview consistency and time efficiency. By leveraging the carefully designed modules including view-weighted texture rectification and related view-based attention, proposed view-integrated diffusion preserves the cross-view consistency during the diffusion process. Qualitative and quantitative results demonstrate the effectiveness of RoomPainter in indoor scene texturing, especially in global consistency and generation efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.947, + 0.51, + 0.957 + ], + "angle": 0, + "content": "581" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] xinsir/controlnet-depth-sdx1-1.0. https://huggingface.co/xinsir/controlnet-depth-sdx1-1.0, 2023.4.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.591 + ], + "angle": 0, + "content": "[7] Dave Zhenyu Chen, Haoxuan Li, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Scenetex: High-quality texture synthesis for indoor scenes via diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21081-21091, 2024. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[10] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2920–2929, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[12] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[13] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Chenjian Gao, Boyan Jiang, Xinghui Li, Yingpeng Zhang, and Qian Yu. Genesistex: Adapting image denoising diffusion to texture space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4620-4629, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[18] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[19] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Texgen: Text-guided 3d texture generation with multi-view sampling and resampling. arXiv preprint arXiv:2408.01291, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[22] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "582" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[32] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[36] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Jiaxiang Tang, Ruijie Lu, Xiaokang Chen, Xiang Wen, Gang Zeng, and Ziwei Liu. Intex: Interactive text-to-texture synthesis via unified depth-aware inpainting. arXiv preprint arXiv:2403.11878, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[41] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[42] Qi Wang, Ruijie Lu, Xudong Xu, Jingbo Wang, Michael Yu Wang, Bo Dai, Gang Zeng, and Dan Xu. Roomtex: Texturing compositional indoor scenes via iterative inpainting. arXiv preprint arXiv:2406.02461, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[44] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[45] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[46] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.901, + 0.688 + ], + "angle": 0, + "content": "[47] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[48] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[49] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[51] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "583" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[52] Wangbo Yu, Chaoran Feng, Jiye Tang, Jiashu Yang, Zhenyu Tang, Xu Jia, Yuchao Yang, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[54] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 30(12):7749-7762, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[56] Haiyang Zhou, Xinhua Cheng, Wangbo Yu, Yonghong Tian, and Li Yuan. Holodreamer: Holistic 3d panoramic world generation from text descriptions. arXiv preprint arXiv:2407.15187, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "584" + } + ] +] \ No newline at end of file diff --git a/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_origin.pdf b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..08905f4d483630a3cb6cdaf46828daaf82e5a8ed --- /dev/null +++ b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/cdff11d5-3d3b-4439-92c3-e436eb05d496_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5bf0dabecc71014d7f983ca8328016913416d9ede4af73ce259363cdcf39f7b +size 4773626 diff --git a/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/full.md b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ef847b57a7a9e77ddcf708efa32a380b53b4999 --- /dev/null +++ b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/full.md @@ -0,0 +1,279 @@ +# RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing + +Zhipeng Huang $^{1}$ Wangbo Yu $^{1,2}$ Xinhua Cheng $^{1}$ Chengshu Zhao $^{1}$ Yunyang Ge $^{1}$ Mingyi Guo $^{1}$ Li Yuan $^{1,2\dagger}$ Yonghong Tian $^{1,2\dagger}$ + +$^{1}$ Peking University + $^{2}$ Pengcheng Laboratory + +![](images/03d79a1da8db38566108b54f9b459eed0d8275ffc8ab738c8fd2f160b1e6ab1a.jpg) +Figure 1. RoomPainter is capable of generating high-fidelity and consistent texture for a given room mesh. + +# Abstract + +Indoor scene texture synthesis has garnered significant interest due to its important potential applications in virtual reality, digital media and creative arts. Existing diffusion-model-based researches either rely on per-view inpainting techniques, which are plagued by severe cross-view inconsistencies and conspicuous seams, or adopt optimization-based approaches that involve substantial computational overhead. In this work, we present RoomPainter, a framework that seamlessly integrates efficiency and consistency to achieve high-fidelity texturing of indoor scenes. The core of RoomPainter features a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a neighbor-integrated attention mechanism for zero-shot texture map generation. Using the MVIS, we firstly generate texture map for the entire room to ensure global consistency, then adopt its variant, namely Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) to repaint individual instances + +within the room, thereby further enhancing local consistency and addressing the occlusion problem. Experiments demonstrate that RoomPainter achieves superior performance for indoor scene texture synthesis in visual quality, global consistency and generation efficiency. + +# 1. Introduction + +High-quality 3D content synthesis is in great demand, owing to its diverse applications across the entertainment industry, robotic simulation and mixed-reality environments. The design of indoor scene typically requires professionals to spend dozens of hours to complete. Generating texture for indoor scene through automated methods can significantly reduce labor costs and save time. With the advancement of text-to-image diffusion models [28, 30, 33, 34], large-scale indoor scene texture synthesis has made significant progress. Although existing indoor texture synthesis methods produce high-quality results, they face several significant challenges. Specifically, these methods can be categorized into two groups: inpainting-based methods [6, 32] and optimization-based methods [7, 24, 48]. + +Inpainting-based methods utilize depth-to-image diffusion models [33, 55] to progressively generate texture for neighboring views, with the guidance of reference geometry and previously generated views. However, due to their per-view inpainting strategy, these approaches often result in inconsistencies in texture style across instances within the room and fail to effectively address the loop-closure problem. Optimization-based methods mitigate these inconsistencies by employing score distillation loss [29] or its variations [43] in global texture maps. However, the optimization process incurs significant time costs for indoor scene texture synthesis, due to the introduction of a large number of learnable parameters and also prone to issue of training instability. To address these challenges, we propose an efficient and high-fidelity indoor scene texturing framework, dubbed RoomPainter. The core of RoomPainter lies in a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a related view-based attention mechanism for zero-shot texture map generation. Using the MVIS, we generate a holistic texture map for the entire room, achieving better global consistency than per-view inpainting-based methods. Subsequently, to paint the textured areas raised by occlusions between objects while preserving the global consistency established in the first stage, we adopt a refine stage to repaint each instance in the room. In this stage, we apply an Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) technique at the instance scale to perform texture inpainting and refinement for each individual instance. + +The main contributions of our work are summarized below: + +- We propose RoomPainter, an effective framework for indoor scene texture synthesis that excels in producing consistent, high-quality texture with remarkable time efficiency. +- We introduce Multi-View Integrated Sampling (MVIS), a zero-shot technique that leverages view-weighted texture rectification and related view-based attention to generate global consistent texture map using 2D diffusion model. +- We further present Multi-View Integrated Repaint Sampling (MVRS), which simultaneously refines the textured areas and inpaints the occlusion areas based on their context, thereby facilitates fine-detailed texture synthesis with local consistency. +- We conduct extensive experiments and compare our method with strong baselines. The results demonstrate that our method not only achieves superior performance in visual quality for indoor scene texture synthesis but also excels in generation efficiency. + +# 2. Related Works + +# 2.1. 3D Generation with Diffusion Model + +With the rapid advancements in image [26, 28, 33, 34, 55] and video generation models [3, 4, 44, 45, 49, 53] in recent years, the field of 3D generation has experienced significant progress. Various methods now generate objects [11, 29, 41, 43, 50-52] or scenes [10, 18, 27, 53, 54, 56] based on textual descriptions via leveraging SDS Loss and its variants [29, 43] to optimize implicit representation such as NeRF [25] and 3D Gaussian Splatting [20]. In the realm of scene generation, some approaches utilize text-guided inpainting models and depth prediction models, which incrementally construct scenes via a warping process [9, 12, 18, 27, 54]. Other methods focus on fine-tuning image generation models using 3D data, allowing them to generate multi-view consistent images or panorama for 3D tasks with the guidance of semantic images or depth maps. Moreover, some methods predict depth from a given image to initialize 3D Gaussian Splatting, followed by further optimization of the 3D Gaussian Splatting using video generation models. + +# 2.2.3D Object Texturing + +The goal of 3D object texturing methods is to create high-quality texture for the given 3D object that align with user-provided textual descriptions. Early methods, such as [6, 32, 39, 42], utilized depth-aware Diffusion Models [26, 33, 55] to iteratively apply inpainting for each viewpoint. At each step, the texture is generated based on the texture from the previous viewpoint, gradually rotating the object to produce a complete texture. While these methods can generate texture in a relatively short time, they fail to address issues such as loop-closure and visible seams. Other works, such as [8, 24, 48], leverage differentiable rendering combined with SDS [29] to texture the given object through optimization techniques. Additionally, approaches like [5, 14, 19, 22] use UV maps as intermediaries and apply different fusion methods to integrate multiple of view independent UV maps into a consistent UV map. During the Diffusion model sampling process, this consistent UV map is used to adjust the Diffusion model's input, ensuring that the resulting texture is more consistent, with fewer visible seams and improved multi-view consistency. + +# 2.3. 3D Indoor Scene Texturing + +In the task of 3D indoor scene texture generation, the objective is to create texture for a given indoor scene based on a user-provided textual description, which presents challenges such as object occlusion and multi-view consistency. SceneTex [7] defines an implicit texture field and optimizes it using VSD [43]. By employing an optimization-based approach to refine the texture field, SceneTex naturally re + +![](images/1d112579350850e288373a3e791576afbb2987acfbd57e82f2cd992a0a3a9957.jpg) +Figure 2. Illustration of Multi-view Integrated Sampling For $N$ surrounding viewpoints in the room, centered around the room's center point and guided by the corresponding depth maps, we use a diffusion model to generate the denoised observation $\mathcal{I}_t^n$ at timestep $t$ . This observation is then projected into UV space to obtain texture maps that corresponding to respective viewpoints. The texture maps from multiple viewpoints are dynamically merged to produce the texture map for the current timestep, which subsequently guides the sampling process for the next timestep. + +solves the issue of multi-view consistency in 3D scene texture. Furthermore, SceneTex incorporates a cross-attention mechanism to transfer texture from visible regions to occluded areas. However, this process can result in blurriness in the occluded regions, and the optimization procedure itself is computationally intensive. When methods like Text2Tex and TEXTure [6, 32], typically used for object texture generation, are applied to scene texture generation, they can only enforce style consistency across multiple objects based on textual control. However, these methods do not ensure global style consistency for the entire scene. Alternatively, if the entire room is treated as a single object and methods such as TEXTure or Text2Tex are employed with the camera positioned at the center of the room, generating the scene texture by rotating the camera, global style consistency can be achieved. However, this approach still faces challenges such as occlusion and visible seams between regions. + +Other approaches, such as [46], use panoramic images to generate a rough texture for the entire room. The camera position is then changed to generate texture for occluded regions. Subsequently, an MLP is employed to predict the texture of unseen areas based on the color and texture coordinates from the UV map. However, this method requires the scene to have an initial texture and relies on texture style transfer, meaning it cannot generate texture from scratch for a completely untextured scene. + +# 3. Methods + +Given a 3D indoor scene mesh $\mathcal{M}$ , our objective is to efficiently generate texture that not only align with the provided textual descriptions but also maintain global consistency. + +Since our framework employs pretrained text-to-image diffusion model to generate texture for the given indoor scene mesh, we will start with an overview of the text-to-image Diffusion model and mesh rendering in Sec. 3.1. Following this, we introduce Multi-View Integrated Sampling (abbreviated as MVIS) in Sec. 3.2, which addresses the challenge of maintaining multi-view consistency while preserving high-frequency details in the diffusion model sampling process. Subsequently, In Sec. 3.3, we introduce a variant of MVIS, namely Multi-View Integrated Repaint Sampling (MVRS), to inpaint the occlusion region and simultaneously refine existing texture for each instance within the indoor scene. Finally, in Sec. 3.4, we introduce a Related View-based Attention mechanism used in the MVIS and MVRS processes, which enhances multi-view consistency during texture synthesis. + +# 3.1. Preliminaries + +Diffusion Model Sampling. Diffusion models [17, 38] are latent variable models that comprises a forward process $q(x_{t}|x_{0})$ and a learned sampling process $p_{\theta}(x_{t - 1}|x_t)$ . Given an initial data point $x_0$ , its noisy counterpart at time $t$ is denoted as $x_{t}$ , and the forward process $q$ is expressed as: + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {(1 - \bar {\alpha} _ {t})} \epsilon , \epsilon \sim N (0, I), \qquad (1) +$$ + +where $t \in [0, \infty)$ and the values of $\alpha_{t}$ are defined by a scheduler to controls the amount of noise added to $x_0$ at time step $t$ . Additionally, it has $\beta_{t} = 1 - \alpha_{t}$ and $\bar{\alpha}_{t} = \prod_{s=1}^{t} \alpha_{s}$ . In the sampling process $p_{\theta}$ , with the DDPM sampler [17], the diffusion model progressively denoises an initial noise image $x_{T} \sim \mathcal{N}(0, I)$ to clean image. Specifically, at any intermediate time step $t$ , the sampling process + +$p_{\theta}$ can be formed as: + +$$ +x _ {t - 1} = \mu_ {t - 1} + \sigma_ {t} \epsilon , \epsilon \sim N (0, I), \tag {2} +$$ + +where the value of $\sigma_{t}$ and $\mu_{t - 1}$ is given by: + +$$ +\mu_ {t - 1} = \frac {\sqrt {\bar {\alpha} _ {t - 1}} \beta_ {t}}{1 - \bar {\alpha} _ {t}} x _ {0, t} + \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1})}{1 - \bar {\alpha} _ {t}} x _ {t}, \tag {3} +$$ + +$$ +\sigma_ {t} = \frac {1 - \bar {\alpha} _ {t - 1}}{1 - \bar {\alpha} _ {t}} \beta_ {t}, \tag {4} +$$ + +and the estimation of the original image $x_{0,t}$ at time step $t$ is calculated as: + +$$ +x _ {0, t} = \frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} \left(x _ {t} , t\right)}{\sqrt {\alpha_ {t}}}, \tag {5} +$$ + +where $\epsilon_{\theta}(x_t,t)$ is the noise that estimated by the diffusion model. As $t$ decreases from $T$ to 0, the sampling process $p_{\theta}$ ultimately samples a clean data point $x_0$ . Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads depth-aware text-to-image diffusion models [1, 55], where the output of the model $\epsilon_{\theta}(x_t,y,d,t)$ is conditioned on a text prompt $y$ and a depth map $d$ . + +In this paper, we utilize Stable Diffusion [28] as our base model. It performs denoising in the latent space and employs an autoencoder $\mathcal{D}(\mathcal{E}(\cdot))$ for the conversion between image and latent representations. Therefore, latent $x_0$ generated by its diffusion process will be decoded to the image $\mathcal{D}(x_0)$ through its decoder. + +Mesh Rendering. Given a mesh $\mathcal{M}$ , a texture map $\mathcal{T}$ and a viewpoint $C$ , the rendering function $\mathcal{R}$ can be employed to produce the rendered image as $I = \mathcal{R}(\mathcal{T},\mathcal{M},C)$ . Conversely, the inverse rendering function: $\mathcal{R}^{-1}$ can reconstruct the texture map from the image, yielding $\mathcal{T}' = \mathcal{R}^{-1}(I,\mathcal{M},C)$ . + +# 3.2. Mult-View Integrated Sampling + +Different from object-level multi-view diffusion models [21, 36, 37], diffusion models that capable of ensuring multi-view consistency in scene level image generation [40] are extremely limited. + +To address this challenge, we propose Multi-View Integrated Sampling (MVIS), grounded with a Related View-based Attention technique, to adapt a standard text-to-image model to generate view-consistent images in a training-free manner. Based on MVIS, we first implement a global texturing stage to generate consistent texture for the scene mesh. + +As shown in Fig. 2, given a set of predefined cameras $\{C\} = C^n, n = 1, \ldots, N$ , we first render their corresponding depth maps $d_n$ and similarity masks $S_n$ , where each + +pixel in the similarity mask represents the inverted normalized value of the cosine similarity between the normal vectors of the visible faces and the viewing direction, with values ranging from [0, 1] [6, 14, 32]. Based on the similarity masks, we construct the texture weighting maps $W_{n}$ for each camera, defined as $W_{n} = \mathcal{R}^{-1}(S_{n},\mathcal{M},C^{n})$ , which are then used to dynamically merge per-view texture maps. + +As shown in Algorithm 1, in the sampling process, we firstly sample $\{x_T^n\}_{n=1}^N \sim \{\mathcal{N}(0,I)\}$ for $N$ camera views. Subsequently, at each denoising step $t$ , we decode all $N$ estimated $x_{0,t}^n$ into $\mathcal{I}_t^n$ and back project them into independent per-view texture map $\mathcal{T}_t^n$ . We then merge these per-view texture maps to form a global texture map $\mathcal{T}_t$ using dynamic merge [2, 22], which can be formulated as: + +$$ +\mathcal {T} _ {t} = \frac {\sum_ {n = 1} ^ {N} \left(W _ {n} ^ {e x p (t)} \cdot \mathcal {T} _ {t} ^ {n}\right)}{\sum_ {n = 1} ^ {N} W _ {n} + \gamma}. \tag {6} +$$ + +Here, $\exp(t)$ is used to dynamically regulate variations in consistency during the sampling process at timestep $t$ . As $t$ decreases, $\exp(t)$ increases linearly, leading to a sharper merged texture. With the merged texture map, we again render per view images from $\mathcal{T}_t$ , using foreground mask $M^n$ for viewpoint $C^n$ , and encode them into latent space to obtain $\tilde{x}_{0,t}$ , which is then used to replace the previous obtained less $x_{0,t}^n$ in Eq. 3 for consistent view generation. After repeating these denoising steps, we are able to generate a consistent texture map, denote as $\mathcal{T}_{MVIS}$ . + +# 3.3. Multi-View Integrated Repaint Sampling + +Although the texture map $\mathcal{T}_{MVIS}$ maintains stylistic and consistent throughout the room, there are still areas that remain untextured due to occlusions. To address this problem, we adopt a refine stage to inpaint the occlusion areas and further enhance the texture fidelity, while preserving the global consistency established in the first stage. + +Specifically, we decouple each instance in the room, resulting a total of $K$ instances $\mathcal{M}_1,\ldots ,\mathcal{M}_K$ . We then design a variant of MVIS, namely multi-view integrated repaint sampling (MVRS), to perform texture inpainting and refinement for each individual instance. As shown in Fig. 3, the core of MVRS lies in incorporating diffusion-based inpainting [23] into the MVIS process. For each individual instance, given $N$ predefined camera views, we firstly render $N$ images $\mathcal{I}^n$ along with their inpainting masks $P^n$ that indicates untextured region of them. Subsequently, for any time step $t$ in the MVIS process, we encode $\mathcal{I}^n$ into the latent space and add noise on it to produce a noisy latent $\tilde{x}_t^n$ , followed by sampling $x_{t}^{n}$ from MVIS, then combine them into a new latent using the inpainting mask $P^n$ , where the painted area keeps unchanged, finally passing it into the U-Net for the next denoising step. This process introduces information from the painted area into the MVIS process, + +![](images/8a3867a78cc9c482a20b45f15b4b17e79815f9232a2ae2c8a32ca8c9885d90b8.jpg) +Figure 3. Illustration of Multi-View Integrated Repaint Sampling(MVRS) Due to occlusion between instances, certain areas remain untextured after the first stage of texture generation. To address this issue, we perform texture generation for each instance within the room based on the painted areas. For $N$ different viewpoints of each instance, guided by the corresponding depth maps, at sampling step $t$ of MVRS, we combine the painted areas $(x_{MVIS})$ with the sampling results of MVIS at step $t + 1$ $(x_{0,t + 1})$ using a mask $P$ to form $x_{t}$ , which serves as the input for the sampling process of timestep t. Noise corresponding to the current timestep was added before mask combine. Upon completing the MVRS process, the texture map for a specific instance in the room is fully generated. + +Algorithm 1 Multi-View Integrated Sampling +Input: mesh $\mathcal{M}$ , text $y$ , cameras $\{C^1,\dots ,C^N\}$ +Parameters: DDPM noise schedule $\{\sigma_t\}_{t = T}^0$ +Initialization: $\{x_T^n\}_{n = 1}^N\sim \{\mathcal{N}(0,I)\}$ +for $t\in \{T\dots 0\}$ do +for $n\in \{1\dots N\}$ do + $\epsilon_t^n\gets \epsilon_\theta (x_t^n,y,d_n,t)$ $x_{0,t}^{n}\gets \frac{x_{t}^{n} - \sqrt{1 - \bar{\alpha}_{t}}\epsilon_{t}^{n}}{\sqrt{\alpha_{t}}}$ $\mathcal{I}_t^n\gets \mathcal{D}(x_{0,t}^n)$ $\mathcal{T}_t^n\gets \mathcal{R}^{-1}(\mathcal{I}_t^n,\mathcal{M},C^n)$ +end for + $\mathcal{T}_t =$ dynamic_merge( $\{\mathcal{T}_t^n\}_{n = 1}^N$ ) +if $t > 0$ then +for $n\in \{1\dots N\}$ do + $\epsilon^n\sim \mathcal{N}(0,I)$ $\tilde{x}_{0,t}^{n}\gets \mathcal{E}(M^{n}\odot \mathcal{R}(\mathcal{T}_{t},\mathcal{M},C_{n}) + (1 - M^{n})\odot \mathcal{I}_{t}^{n})$ $\mu_{t - 1}^{n}\gets \frac{\sqrt{\alpha_{t - 1}}\beta_{t}}{1 - \bar{\alpha}_{t}}\tilde{x}_{0,t}^{n} + \frac{\sqrt{\alpha_{t}}(1 - \bar{\alpha}_{t - 1})}{1 - \bar{\alpha}_{t}} x_{t}^{n}$ $x_{t - 1}^{n}\gets \mu_{t - 1}^{n} + \sigma_{t}\epsilon^{n}$ +end for +end if +end for +TMVIS $=$ T0 +return Texture map TMVIS + +effectively transferring the texture of the painted area to the regions with holes while ensuring consistency across multiple viewpoints. After iterating through all instances within the room, we obtain the final output scene texture $\mathcal{T}_{MVRS}$ . + +# 3.4. Related View-based Attention + +In the MVIS and MVRS process, we employ a 2D Diffusion Model for zero-shot multi-view sampling. A key challenge in maintaining view consistency is sharing information across multiple images from different viewpoints. Inspired by [15], we modify the self-attention mechanism in the diffusion model to ensure that each view incorporates sampling information from related views during the sampling process. During the sampling process, denote the queries, keys and values derived from the deep feature of the denoising U-Net for viewpoint $n$ as $Q_{n}$ , $K_{n}$ , and $V_{n}$ , respectively. For each viewpoint, assume that it has a total of $R$ associated viewpoints (view that has overlap). Subsequently, the Related View-based Attention for viewpoint $n$ is computed as: + +$$ +\operatorname {s o f t m a x} \left(\frac {Q _ {n} \tilde {K} _ {n} ^ {T}}{\sqrt {d}} \tilde {V} _ {n}\right), \tag {7} +$$ + +where $\tilde{K}_n = \left[K_1K_2\ldots K_R\right]^T$ and $\tilde{V}_n = \left[V_1V_2\ldots V_R\right]^T$ . In the first room-scale texturing stage, we consider the $R$ related viewpoints for view $n$ as its adjacent views, i.e., its left and right view. When sampling at instance scale, we define the related views as all the $N$ predefined camera view. + +# 4. Experiments + +In this section, we first describe the general implementation of the baseline methods used for comparison. Next, we provide an overview of the implementation details of our proposed method. Finally, we present our quantitative experiments, qualitative results, and ablation studies. + +![](images/bf78b4c1a0cb9f69752dc94637fd32dd25d3383370c47a6a3be5ccfeeeeb724d.jpg) +Figure 4. Qualitative comparisons. Text2Tex-H [6] suffers from occlusion and visible seams. Text2Tex-C [6] struggles to maintain style consistency across all instances. SceneTex [7] produces unrealistic texture and results in blurry regions. In contrast, our method generates high-quality texture while preserving overall style consistency across instances in the scene. Ceilings and back-facing walls are excluded for improved visualizations. + +# 4.1. Baseline Implementation Details + +- Text2Tex-H [6]: This variant of Text2Tex is implemented using a holistic text prompt provided by user and treats the entire room as a cohesive entity. The camera is positioned at the center of the room, with a field of view of $60^{\circ}$ , and the elevation angle is set to $0^{\circ}$ , and azimuth angles uniformly sampled from the range $[0^{\circ}, 360^{\circ}]$ , which results in a total of twelve perspectives. +- Text2Tex-C [6]: For comparative purposes, we employ the original configuration of Text2Tex, wherein each + +room instance, including the interior walls and furniture, is textured using an instance-specific text prompt. And for interior walls texturing, we utilize the text prompt such as: "A Chinese style living room, no furniture" and sampled 12 perspectives same as Text2Tex-H. All the textured instances are subsequently combined, forming the output textured room. + +- SceneTex [7]: We compare against SceneTex using its original setup, while employing the same holistic prompt as utilized in our approach. + +![](images/4766f501f5344637745ee3097aeae72152506d41bddea829d503674777935f77.jpg) +"A Cream style master bedroom" +Ours (W/O Attn) +Figure 5. Ablation studies on the multiview-consistency module. Synthesizing texture without the Related View-based attention leads to inconsistencies across views, as shown in the leftmost column. When synthesizing texture without MVIS, noticeable inconsistencies appear across different viewpoints (middle column). In contrast, our full method samples multi-view images with significantly stronger consistency. The areas of major inconsistency in the images are highlighted with red and blue boxes. In the same column, boxes of the same color indicate that they are close to each other in 3D space. Zoom in for the best view. + +![](images/dc70eb03cd093f87aa7e09a55473049c93de1e773bf5fb8e032200e289c41abc.jpg) +Ours (W/O MVIS) + +![](images/4f98974ec98f7fd41d39d9e2178e7749a2571b4d4331d8c9102884943fba2207.jpg) +Ours (Full) + +# 4.2. Implementation Details + +In this experiment, we used the SDXL model [28] as the text-to-image generator and incorporated depth information through ControlNet [1, 55]. For the text prompt, we implemented view-specific prompt combining the holistic prompt with the names of the types of objects present within the field of view. For sampling, we applied the DDPM [17] as the sampler, setting the sampling steps to 50. We used PyTorch3D [31] for rendering and texture projection. Additionally, UV atlas generation for the mesh was handled + +with Xatlas [47]. All experiments were performed on an NVIDIA RTX A6000 GPU. Further implementation details are available in the supplementary materials. + +# 4.3. Quantitative Analysis + +Our quantitative experiment were conducted across 10 scenes including 5 bedroom and 5 living room from the 3D-FRONT dataset [13], using five different text styles totally. To evaluate the generated texture, we calculated both the CLIP Score(CLIP) [16], which assesses the alignment + +
MethodGeneration Time(mins) ↓CLIP ↑AS ↑
Text2Tex-H [6]8.5021.584.34
Text2Tex-C [6]70.7521.934.85
SceneTex [7]2614.5021.874.75
(Ours) Full46.0023.475.03
(Ours) w/o Attn-23.325.01
(Ours) w/o MVIS-23.275.01
(Ours) w/o MVRS-22.394.40
+ +Table 1. Quantitative comparisons. We report the Generation time and 2D metrics result for quantitative comparisons. The generation time is shown in the format minutes. The 2D metrics include: CLIP Score (CLIP) [16], Aesthetic Score(AS) [35]. We show that our method produces high quality textures efficiently. + +between generated texture and the textual description, and the CLIP Aesthetic Score(AS) [35], which measures texture quality. We also compared the average processing time of each texture generation method in minutes. The results are summarized in Table.1. We render 10 images from novel views for each room to calculate the average of CLIP and AS. Due to the presence of numerous distinct objects within the room, the methods in baselines do not adjust the text prompts for texture generation based on the variation of objects from different viewpoints. Therefore, we employ a holistic prompt to compute the CLIP score. Since our evaluation metric involves image sampling from new viewpoints, Text2Tex-H, which fails to address the occlusion problem, generates texture containing large occluded regions, resulting in the lowest performance in terms of the metric. However, due to the limited number of sampled viewpoints, it exhibits the fastest generation speed. We configure SceneTex according to its original setup, sampling from a sphere centered at the room's center. As a result, regions not observed by SceneTex often exhibit blurriness. Meanwhile, in addition to optimizing the texture field using VSD, SceneTex also requires training LoRA, which results in the longest processing time. Text2Tex-C employs an instance-specific approach for texture generation, utilizing different prompts for each object, which results in higher performance metrics. Additionally, it incorporates viewpoint selection and texture optimization processes, leading to a longer computation time. + +# 4.4. Qualitative Analysis + +We provide a visual comparison in fig.4. Specifically, Text2Tex-H [6] shows that simply applying object texture generation method to indoor scene texturing task cannot resolve the occlusion problem (see the large black areas in the first column of fig.4) and contain obvious seams. Text2Tex-C [6] texturing the room completely but suffering from obvious seams and global style consistency across objects. Although SceneTex introduces the use of cross-attention + +to generate texture for occluded regions, noticeable blurriness still persists in these areas. Furthermore, due to its optimization-based approach, SceneTex often produces unrealistic texture. In contrast, our method generates globally consistent texture with a unified style in the first stage. In the second stage, through per-instance inpainting and refinement, we successfully maintain global style consistency while addressing issues of occlusion and unrealistic texture. + +# 4.5. Ablation Studies + +As shown in Fig.5, all images are raw samples from the final step, thus revealing inherent multi-view inconsistencies. + +Effectiveness of the MVIS. In the first stage of texturing, we introduce MVIS to enhance the global consistency of the texture. Qualitative results from this stage are shown in the second column of Fig.5. The figure demonstrates that without MVIS, texture consistency across multiple viewpoints can be significantly impact. We further conduct quantitative ablation on dataset adopted in previous experiments. Results are shown in the seventh row of Tab.1. The quantitative experimental results clearly demonstrate that MVIS improves the quality of the generated texture. + +Effectiveness of the Related View-based Attention. We incorporate the Related View-based Attention module in both stages of texture generation to enhance consistency. Both qualitative(the first column of Fig.5) and quantitative experiments(the fifth row of the Tab.1) demonstrate that using only MVIS or MVRS is insufficient to generate consistent and high-quality texture. The Related View-based Attention module effectively enhances consistency while ensuring the quality of the generated texture. + +Effectiveness of the MVRS. In the second stage of texturing, we use MVRS to generate texture for the occluded regions of each instance and refine the texture generated in the first stage. Quantitative experimental results (seventh row in Tab.1) show that without MVRS, the quality of the texture significantly deteriorates. This is because, without MVRS, the texture generated in the first stage suffers from low-quality results due to occlusion. We will provide additional qualitative materials demonstrating the effectiveness of MVRS in the supplementary materials. + +# 5. Conclusion + +We propose an indoor scene texturing framework named RoomPainter that generates high-quality texture with multiview consistency and time efficiency. By leveraging the carefully designed modules including view-weighted texture rectification and related view-based attention, proposed view-integrated diffusion preserves the cross-view consistency during the diffusion process. Qualitative and quantitative results demonstrate the effectiveness of RoomPainter in indoor scene texturing, especially in global consistency and generation efficiency. + +# References + +[1] xinsir/controlnet-depth-sdx1-1.0. https://huggingface.co/xinsir/controlnet-depth-sdx1-1.0, 2023.4.7 +[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 4 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563-22575, 2023. 2 +[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2 +[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 1, 2, 3, 4, 6, 8 +[7] Dave Zhenyu Chen, Haoxuan Li, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Scenetex: High-quality texture synthesis for indoor scenes via diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21081-21091, 2024. 1, 2, 6, 8 +[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 2 +[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2 +[10] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2920–2929, 2023. 2 +[11] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene, 2025. 2 +[12] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2 + +[13] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 7 +[14] Chenjian Gao, Boyan Jiang, Xinghui Li, Yingpeng Zhang, and Qian Yu. Genesistex: Adapting image denoising diffusion to texture space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4620-4629, 2024. 2, 4 +[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 5 +[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7, 8 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7 +[18] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2 +[19] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Texgen: Text-guided 3d texture generation with multi-view sampling and resampling. arXiv preprint arXiv:2408.01291, 2024. 2 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2 +[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 4 +[22] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 2, 4 +[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 4 +[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 2 +[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2 +[27] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 4, 7 +[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1 +[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 7 +[32] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 1, 2, 3, 4 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 8 +[36] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 4 +[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4 +[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using + +nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[39] Jiaxiang Tang, Ruijie Lu, Xiaokang Chen, Xiang Wen, Gang Zeng, and Ziwei Liu. Intex: Interactive text-to-texture synthesis via unified depth-aware inpainting. arXiv preprint arXiv:2403.11878, 2024. 2 +[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 4 +[41] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2 +[42] Qi Wang, Ruijie Lu, Xudong Xu, Jingbo Wang, Michael Yu Wang, Bo Dai, Gang Zeng, and Dan Xu. Roomtex: Texturing compositional indoor scenes via iterative inpainting. arXiv preprint arXiv:2406.02461, 2024. 2 +[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[44] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2 +[45] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2 +[46] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 3 +[47] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 7 +[48] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 1, 2 +[49] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 2 +[50] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. 2023. 2 +[51] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. + +Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744, 2023. +[52] Wangbo Yu, Chaoran Feng, Jiye Tang, Jiashu Yang, Zhenyu Tang, Xu Jia, Yuchao Yang, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2 +[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2 +[54] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 30(12):7749-7762, 2024. 2 +[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 7 +[56] Haiyang Zhou, Xinhua Cheng, Wangbo Yu, Yonghong Tian, and Li Yuan. Holodreamer: Holistic 3d panoramic world generation from text descriptions. arXiv preprint arXiv:2407.15187, 2024. 2 \ No newline at end of file diff --git a/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/images.zip b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db53931485cc6ca22af95c6d46afc5c2542af209 --- /dev/null +++ b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d361e7f0c23701e6fd6cf5c6d75368da808553042c2465766f0c3560c47e7508 +size 599616 diff --git a/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/layout.json b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dab82164f55f954cc02dc5ba7d4d55ce4f7b26a1 --- /dev/null +++ b/2025/RoomPainter_ View-Integrated Diffusion for Consistent Indoor Scene Texturing/layout.json @@ -0,0 +1,8498 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 553, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 553, + 121 + ], + "type": "text", + "content": "RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": "Zhipeng Huang" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Wangbo Yu" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Xinhua Cheng" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Chengshu Zhao" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Yunyang Ge" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Mingyi Guo" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Li Yuan" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "text", + "content": " Yonghong Tian" + }, + { + "bbox": [ + 61, + 141, + 564, + 173 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 254, + 184, + 373, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 184, + 373, + 215 + ], + "spans": [ + { + "bbox": [ + 254, + 184, + 373, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 254, + 184, + 373, + 215 + ], + "type": "text", + "content": "Peking University \n" + }, + { + "bbox": [ + 254, + 184, + 373, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 254, + 184, + 373, + 215 + ], + "type": "text", + "content": "Pengcheng Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 256, + 553, + 399 + ], + "blocks": [ + { + "bbox": [ + 58, + 256, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 58, + 256, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 58, + 256, + 553, + 399 + ], + "type": "image", + "image_path": "03d79a1da8db38566108b54f9b459eed0d8275ffc8ab738c8fd2f160b1e6ab1a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 407, + 499, + 419 + ], + "lines": [ + { + "bbox": [ + 108, + 407, + 499, + 419 + ], + "spans": [ + { + "bbox": [ + 108, + 407, + 499, + 419 + ], + "type": "text", + "content": "Figure 1. RoomPainter is capable of generating high-fidelity and consistent texture for a given room mesh." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "spans": [ + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 463, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 463, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 463, + 297, + 715 + ], + "type": "text", + "content": "Indoor scene texture synthesis has garnered significant interest due to its important potential applications in virtual reality, digital media and creative arts. Existing diffusion-model-based researches either rely on per-view inpainting techniques, which are plagued by severe cross-view inconsistencies and conspicuous seams, or adopt optimization-based approaches that involve substantial computational overhead. In this work, we present RoomPainter, a framework that seamlessly integrates efficiency and consistency to achieve high-fidelity texturing of indoor scenes. The core of RoomPainter features a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a neighbor-integrated attention mechanism for zero-shot texture map generation. Using the MVIS, we firstly generate texture map for the entire room to ensure global consistency, then adopt its variant, namely Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) to repaint individual instances" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 498 + ], + "type": "text", + "content": "within the room, thereby further enhancing local consistency and addressing the occlusion problem. Experiments demonstrate that RoomPainter achieves superior performance for indoor scene texture synthesis in visual quality, global consistency and generation efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 525, + 395, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 395, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 395, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 547, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 547, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 547, + 555, + 715 + ], + "type": "text", + "content": "High-quality 3D content synthesis is in great demand, owing to its diverse applications across the entertainment industry, robotic simulation and mixed-reality environments. The design of indoor scene typically requires professionals to spend dozens of hours to complete. Generating texture for indoor scene through automated methods can significantly reduce labor costs and save time. With the advancement of text-to-image diffusion models [28, 30, 33, 34], large-scale indoor scene texture synthesis has made significant progress. Although existing indoor texture synthesis methods produce high-quality results, they face several significant challenges. Specifically, these methods can be categorized into two groups: inpainting-based methods [6, 32] and optimization-based methods [7, 24, 48]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "574" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 453 + ], + "type": "text", + "content": "Inpainting-based methods utilize depth-to-image diffusion models [33, 55] to progressively generate texture for neighboring views, with the guidance of reference geometry and previously generated views. However, due to their per-view inpainting strategy, these approaches often result in inconsistencies in texture style across instances within the room and fail to effectively address the loop-closure problem. Optimization-based methods mitigate these inconsistencies by employing score distillation loss [29] or its variations [43] in global texture maps. However, the optimization process incurs significant time costs for indoor scene texture synthesis, due to the introduction of a large number of learnable parameters and also prone to issue of training instability. To address these challenges, we propose an efficient and high-fidelity indoor scene texturing framework, dubbed RoomPainter. The core of RoomPainter lies in a zero-shot technique that effectively adapts a 2D diffusion model for 3D-consistent texture synthesis, along with a two-stage generation strategy that ensures both global and local consistency. Specifically, we introduce Attention-Guided Multi-View Integrated Sampling (MVIS) combined with a related view-based attention mechanism for zero-shot texture map generation. Using the MVIS, we generate a holistic texture map for the entire room, achieving better global consistency than per-view inpainting-based methods. Subsequently, to paint the textured areas raised by occlusions between objects while preserving the global consistency established in the first stage, we adopt a refine stage to repaint each instance in the room. In this stage, we apply an Attention-Guided Multi-View Integrated Repaint Sampling (MVRS) technique at the instance scale to perform texture inpainting and refinement for each individual instance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 464, + 294, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 294, + 486 + ], + "type": "text", + "content": "The main contributions of our work are summarized below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 496, + 293, + 708 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 496, + 293, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 293, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 293, + 540 + ], + "type": "text", + "content": "- We propose RoomPainter, an effective framework for indoor scene texture synthesis that excels in producing consistent, high-quality texture with remarkable time efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 544, + 293, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 293, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 293, + 589 + ], + "type": "text", + "content": "- We introduce Multi-View Integrated Sampling (MVIS), a zero-shot technique that leverages view-weighted texture rectification and related view-based attention to generate global consistent texture map using 2D diffusion model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 592, + 293, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 293, + 648 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 293, + 648 + ], + "type": "text", + "content": "- We further present Multi-View Integrated Repaint Sampling (MVRS), which simultaneously refines the textured areas and inpaints the occlusion areas based on their context, thereby facilitates fine-detailed texture synthesis with local consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 651, + 293, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 293, + 708 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 293, + 708 + ], + "type": "text", + "content": "- We conduct extensive experiments and compare our method with strong baselines. The results demonstrate that our method not only achieves superior performance in visual quality for indoor scene texture synthesis but also excels in generation efficiency." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 405, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 405, + 83 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 405, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 91, + 507, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 507, + 102 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 507, + 102 + ], + "type": "text", + "content": "2.1. 3D Generation with Diffusion Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 109, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 109, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 109, + 553, + 335 + ], + "type": "text", + "content": "With the rapid advancements in image [26, 28, 33, 34, 55] and video generation models [3, 4, 44, 45, 49, 53] in recent years, the field of 3D generation has experienced significant progress. Various methods now generate objects [11, 29, 41, 43, 50-52] or scenes [10, 18, 27, 53, 54, 56] based on textual descriptions via leveraging SDS Loss and its variants [29, 43] to optimize implicit representation such as NeRF [25] and 3D Gaussian Splatting [20]. In the realm of scene generation, some approaches utilize text-guided inpainting models and depth prediction models, which incrementally construct scenes via a warping process [9, 12, 18, 27, 54]. Other methods focus on fine-tuning image generation models using 3D data, allowing them to generate multi-view consistent images or panorama for 3D tasks with the guidance of semantic images or depth maps. Moreover, some methods predict depth from a given image to initialize 3D Gaussian Splatting, followed by further optimization of the 3D Gaussian Splatting using video generation models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 345, + 433, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 345, + 433, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 345, + 433, + 357 + ], + "type": "text", + "content": "2.2.3D Object Texturing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 364, + 553, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 364, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 364, + 553, + 602 + ], + "type": "text", + "content": "The goal of 3D object texturing methods is to create high-quality texture for the given 3D object that align with user-provided textual descriptions. Early methods, such as [6, 32, 39, 42], utilized depth-aware Diffusion Models [26, 33, 55] to iteratively apply inpainting for each viewpoint. At each step, the texture is generated based on the texture from the previous viewpoint, gradually rotating the object to produce a complete texture. While these methods can generate texture in a relatively short time, they fail to address issues such as loop-closure and visible seams. Other works, such as [8, 24, 48], leverage differentiable rendering combined with SDS [29] to texture the given object through optimization techniques. Additionally, approaches like [5, 14, 19, 22] use UV maps as intermediaries and apply different fusion methods to integrate multiple of view independent UV maps into a consistent UV map. During the Diffusion model sampling process, this consistent UV map is used to adjust the Diffusion model's input, ensuring that the resulting texture is more consistent, with fewer visible seams and improved multi-view consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 611, + 463, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 463, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 463, + 623 + ], + "type": "text", + "content": "2.3. 3D Indoor Scene Texturing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 630, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 630, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 630, + 553, + 712 + ], + "type": "text", + "content": "In the task of 3D indoor scene texture generation, the objective is to create texture for a given indoor scene based on a user-provided textual description, which presents challenges such as object occlusion and multi-view consistency. SceneTex [7] defines an implicit texture field and optimizes it using VSD [43]. By employing an optimization-based approach to refine the texture field, SceneTex naturally re" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "text", + "content": "575" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 73, + 552, + 238 + ], + "blocks": [ + { + "bbox": [ + 62, + 73, + 552, + 238 + ], + "lines": [ + { + "bbox": [ + 62, + 73, + 552, + 238 + ], + "spans": [ + { + "bbox": [ + 62, + 73, + 552, + 238 + ], + "type": "image", + "image_path": "1d112579350850e288373a3e791576afbb2987acfbd57e82f2cd992a0a3a9957.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Illustration of Multi-view Integrated Sampling For " + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "text", + "content": " surrounding viewpoints in the room, centered around the room's center point and guided by the corresponding depth maps, we use a diffusion model to generate the denoised observation " + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t^n" + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 247, + 555, + 304 + ], + "type": "text", + "content": ". This observation is then projected into UV space to obtain texture maps that corresponding to respective viewpoints. The texture maps from multiple viewpoints are dynamically merged to produce the texture map for the current timestep, which subsequently guides the sampling process for the next timestep." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 324, + 294, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 324, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 294, + 539 + ], + "type": "text", + "content": "solves the issue of multi-view consistency in 3D scene texture. Furthermore, SceneTex incorporates a cross-attention mechanism to transfer texture from visible regions to occluded areas. However, this process can result in blurriness in the occluded regions, and the optimization procedure itself is computationally intensive. When methods like Text2Tex and TEXTure [6, 32], typically used for object texture generation, are applied to scene texture generation, they can only enforce style consistency across multiple objects based on textual control. However, these methods do not ensure global style consistency for the entire scene. Alternatively, if the entire room is treated as a single object and methods such as TEXTure or Text2Tex are employed with the camera positioned at the center of the room, generating the scene texture by rotating the camera, global style consistency can be achieved. However, this approach still faces challenges such as occlusion and visible seams between regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 540, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 647 + ], + "type": "text", + "content": "Other approaches, such as [46], use panoramic images to generate a rough texture for the entire room. The camera position is then changed to generate texture for occluded regions. Subsequently, an MLP is employed to predict the texture of unseen areas based on the color and texture coordinates from the UV map. However, this method requires the scene to have an initial texture and relies on texture style transfer, meaning it cannot generate texture from scratch for a completely untextured scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 657, + 117, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 117, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 117, + 670 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Given a 3D indoor scene mesh " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", our objective is to efficiently generate texture that not only align with the provided textual descriptions but also maintain global consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 324, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 517 + ], + "type": "text", + "content": "Since our framework employs pretrained text-to-image diffusion model to generate texture for the given indoor scene mesh, we will start with an overview of the text-to-image Diffusion model and mesh rendering in Sec. 3.1. Following this, we introduce Multi-View Integrated Sampling (abbreviated as MVIS) in Sec. 3.2, which addresses the challenge of maintaining multi-view consistency while preserving high-frequency details in the diffusion model sampling process. Subsequently, In Sec. 3.3, we introduce a variant of MVIS, namely Multi-View Integrated Repaint Sampling (MVRS), to inpaint the occlusion region and simultaneously refine existing texture for each instance within the indoor scene. Finally, in Sec. 3.4, we introduce a Related View-based Attention mechanism used in the MVIS and MVRS processes, which enhances multi-view consistency during texture synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 522, + 400, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 400, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 400, + 534 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": "Diffusion Model Sampling. Diffusion models [17, 38] are latent variable models that comprises a forward process " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{0})" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": " and a learned sampling process " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t)" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": ". Given an initial data point " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": ", its noisy counterpart at time " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": ", and the forward process " + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 608, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 608, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 345, + 608, + 555, + 624 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {(1 - \\bar {\\alpha} _ {t})} \\epsilon , \\epsilon \\sim N (0, I), \\qquad (1)", + "image_path": "9ef7eacbba5002dd19116b234d4cf83c8e2a24706d98239f2802f4f217f88344.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "t \\in [0, \\infty)" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " and the values of " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " are defined by a scheduler to controls the amount of noise added to " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ". Additionally, it has " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{t} = 1 - \\alpha_{t}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ". In the sampling process " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", with the DDPM sampler [17], the diffusion model progressively denoises an initial noise image " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " to clean image. Specifically, at any intermediate time step " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", the sampling process" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "576" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 141, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 141, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 141, + 83 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 141, + 83 + ], + "type": "text", + "content": " can be formed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 95, + 295, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 95, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 104, + 95, + 295, + 108 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\mu_ {t - 1} + \\sigma_ {t} \\epsilon , \\epsilon \\sim N (0, I), \\tag {2}", + "image_path": "60fc95a09d350867e2a1415d4f24857f8a21b15f4192242e7cc9bd7d37c33521.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "type": "text", + "content": "where the value of " + }, + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "type": "inline_equation", + "content": "\\mu_{t - 1}" + }, + { + "bbox": [ + 55, + 118, + 233, + 130 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 139, + 295, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 139, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 89, + 139, + 295, + 166 + ], + "type": "interline_equation", + "content": "\\mu_ {t - 1} = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} x _ {0, t} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} x _ {t}, \\tag {3}", + "image_path": "3ab32085394a842a1d09ce75929ba901b99192356722891811d32d985107d612.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 178, + 295, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 178, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 178, + 295, + 204 + ], + "type": "interline_equation", + "content": "\\sigma_ {t} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} \\beta_ {t}, \\tag {4}", + "image_path": "6f0259beb09f7f4e3754d8501a5102cac611687760f022d3614146962da017cf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "text", + "content": "and the estimation of the original image " + }, + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "inline_equation", + "content": "x_{0,t}" + }, + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 240, + 295, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 240, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 111, + 240, + 295, + 268 + ], + "type": "interline_equation", + "content": "x _ {0, t} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(x _ {t} , t\\right)}{\\sqrt {\\alpha_ {t}}}, \\tag {5}", + "image_path": "0ec6edde649c050e0defa6d7a13c5c0184110d306337b885d1d324ee52940dcc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " is the noise that estimated by the diffusion model. As " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " decreases from " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " to 0, the sampling process " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " ultimately samples a clean data point " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": ". Moreover, this process can be generalized for learning a marginal distribution using an additional input condition. That leads depth-aware text-to-image diffusion models [1, 55], where the output of the model " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,y,d,t)" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " is conditioned on a text prompt " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " and a depth map " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": "In this paper, we utilize Stable Diffusion [28] as our base model. It performs denoising in the latent space and employs an autoencoder " + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\cdot))" + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": " for the conversion between image and latent representations. Therefore, latent " + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": " generated by its diffusion process will be decoded to the image " + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(x_0)" + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": " through its decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": "Mesh Rendering. Given a mesh " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": ", a texture map " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": " and a viewpoint " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": ", the rendering function " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": " can be employed to produce the rendered image as " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "I = \\mathcal{R}(\\mathcal{T},\\mathcal{M},C)" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": ". Conversely, the inverse rendering function: " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": " can reconstruct the texture map from the image, yielding " + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{T}' = \\mathcal{R}^{-1}(I,\\mathcal{M},C)" + }, + { + "bbox": [ + 55, + 445, + 296, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 526, + 228, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 228, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 228, + 540 + ], + "type": "text", + "content": "3.2. Mult-View Integrated Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "content": "Different from object-level multi-view diffusion models [21, 36, 37], diffusion models that capable of ensuring multi-view consistency in scene level image generation [40] are extremely limited." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": "To address this challenge, we propose Multi-View Integrated Sampling (MVIS), grounded with a Related View-based Attention technique, to adapt a standard text-to-image model to generate view-consistent images in a training-free manner. Based on MVIS, we first implement a global texturing stage to generate consistent texture for the scene mesh." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "As shown in Fig. 2, given a set of predefined cameras " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{C\\} = C^n, n = 1, \\ldots, N" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we first render their corresponding depth maps " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "d_n" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and similarity masks " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_n" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", where each" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "pixel in the similarity mask represents the inverted normalized value of the cosine similarity between the normal vectors of the visible faces and the viewing direction, with values ranging from [0, 1] [6, 14, 32]. Based on the similarity masks, we construct the texture weighting maps " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "W_{n}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " for each camera, defined as " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "W_{n} = \\mathcal{R}^{-1}(S_{n},\\mathcal{M},C^{n})" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", which are then used to dynamically merge per-view texture maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": "As shown in Algorithm 1, in the sampling process, we firstly sample " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "\\{x_T^n\\}_{n=1}^N \\sim \\{\\mathcal{N}(0,I)\\}" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": " camera views. Subsequently, at each denoising step " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": ", we decode all " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": " estimated " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "x_{0,t}^n" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t^n" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": " and back project them into independent per-view texture map " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t^n" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": ". We then merge these per-view texture maps to form a global texture map " + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 157, + 554, + 240 + ], + "type": "text", + "content": " using dynamic merge [2, 22], which can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 250, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 250, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 375, + 250, + 553, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {t} = \\frac {\\sum_ {n = 1} ^ {N} \\left(W _ {n} ^ {e x p (t)} \\cdot \\mathcal {T} _ {t} ^ {n}\\right)}{\\sum_ {n = 1} ^ {N} W _ {n} + \\gamma}. \\tag {6}", + "image_path": "d6ef763bef0c35a76984f964421dbc9bdab14232333795907b682476f8d8d1cc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\exp(t)" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": " is used to dynamically regulate variations in consistency during the sampling process at timestep " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": " decreases, " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\exp(t)" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": " increases linearly, leading to a sharper merged texture. With the merged texture map, we again render per view images from " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": ", using foreground mask " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "M^n" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": " for viewpoint " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "C^n" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": ", and encode them into latent space to obtain " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{0,t}" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": ", which is then used to replace the previous obtained less " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "x_{0,t}^n" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": " in Eq. 3 for consistent view generation. After repeating these denoising steps, we are able to generate a consistent texture map, denote as " + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{MVIS}" + }, + { + "bbox": [ + 313, + 290, + 555, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 419, + 529, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 529, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 529, + 432 + ], + "type": "text", + "content": "3.3. Multi-View Integrated Repaint Sampling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "Although the texture map " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{MVIS}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " maintains stylistic and consistent throughout the room, there are still areas that remain untextured due to occlusions. To address this problem, we adopt a refine stage to inpaint the occlusion areas and further enhance the texture fidelity, while preserving the global consistency established in the first stage." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": "Specifically, we decouple each instance in the room, resulting a total of " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " instances " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1,\\ldots ,\\mathcal{M}_K" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ". We then design a variant of MVIS, namely multi-view integrated repaint sampling (MVRS), to perform texture inpainting and refinement for each individual instance. As shown in Fig. 3, the core of MVRS lies in incorporating diffusion-based inpainting [23] into the MVIS process. For each individual instance, given " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " predefined camera views, we firstly render " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^n" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " along with their inpainting masks " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "P^n" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " that indicates untextured region of them. Subsequently, for any time step " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " in the MVIS process, we encode " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^n" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " into the latent space and add noise on it to produce a noisy latent " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{x}_t^n" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ", followed by sampling " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{t}^{n}" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " from MVIS, then combine them into a new latent using the inpainting mask " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "P^n" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ", where the painted area keeps unchanged, finally passing it into the U-Net for the next denoising step. This process introduces information from the painted area into the MVIS process," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "577" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 213 + ], + "type": "image", + "image_path": "8a3867a78cc9c482a20b45f15b4b17e79815f9232a2ae2c8a32ca8c9885d90b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": "Figure 3. Illustration of Multi-View Integrated Repaint Sampling(MVRS) Due to occlusion between instances, certain areas remain untextured after the first stage of texture generation. To address this issue, we perform texture generation for each instance within the room based on the painted areas. For " + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": " different viewpoints of each instance, guided by the corresponding depth maps, at sampling step " + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": " of MVRS, we combine the painted areas " + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "(x_{MVIS})" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": " with the sampling results of MVIS at step " + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "(x_{0,t + 1})" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": " using a mask " + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 220, + 555, + 287 + ], + "type": "text", + "content": ", which serves as the input for the sampling process of timestep t. Noise corresponding to the current timestep was added before mask combine. Upon completing the MVRS process, the texture map for a specific instance in the room is fully generated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 65, + 323, + 296, + 618 + ], + "blocks": [ + { + "bbox": [ + 56, + 308, + 243, + 319 + ], + "lines": [ + { + "bbox": [ + 56, + 308, + 243, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 243, + 319 + ], + "type": "text", + "content": "Algorithm 1 Multi-View Integrated Sampling" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "lines": [ + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": "Input: mesh " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " , text " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " , cameras " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{C^1,\\dots ,C^N\\}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " \nParameters: DDPM noise schedule " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{\\sigma_t\\}_{t = T}^0" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " \nInitialization: " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{x_T^n\\}_{n = 1}^N\\sim \\{\\mathcal{N}(0,I)\\}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "t\\in \\{T\\dots 0\\}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " do \nfor " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "n\\in \\{1\\dots N\\}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\epsilon_t^n\\gets \\epsilon_\\theta (x_t^n,y,d_n,t)" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "x_{0,t}^{n}\\gets \\frac{x_{t}^{n} - \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}^{n}}{\\sqrt{\\alpha_{t}}}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t^n\\gets \\mathcal{D}(x_{0,t}^n)" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t^n\\gets \\mathcal{R}^{-1}(\\mathcal{I}_t^n,\\mathcal{M},C^n)" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " \nend for \n" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t =" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " dynamic_merge(" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_t^n\\}_{n = 1}^N" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " ) \nif " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " then \nfor " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "n\\in \\{1\\dots N\\}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\epsilon^n\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{0,t}^{n}\\gets \\mathcal{E}(M^{n}\\odot \\mathcal{R}(\\mathcal{T}_{t},\\mathcal{M},C_{n}) + (1 - M^{n})\\odot \\mathcal{I}_{t}^{n})" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\mu_{t - 1}^{n}\\gets \\frac{\\sqrt{\\alpha_{t - 1}}\\beta_{t}}{1 - \\bar{\\alpha}_{t}}\\tilde{x}_{0,t}^{n} + \\frac{\\sqrt{\\alpha_{t}}(1 - \\bar{\\alpha}_{t - 1})}{1 - \\bar{\\alpha}_{t}} x_{t}^{n}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "x_{t - 1}^{n}\\gets \\mu_{t - 1}^{n} + \\sigma_{t}\\epsilon^{n}" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " \nend for \nend if \nend for \nTMVIS " + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 65, + 323, + 296, + 618 + ], + "type": "text", + "content": " T0 \nreturn Texture map TMVIS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "effectively transferring the texture of the painted area to the regions with holes while ensuring consistency across multiple viewpoints. After iterating through all instances within the room, we obtain the final output scene texture " + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{MVRS}" + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 307, + 476, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 307, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 476, + 319 + ], + "type": "text", + "content": "3.4. Related View-based Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": "In the MVIS and MVRS process, we employ a 2D Diffusion Model for zero-shot multi-view sampling. A key challenge in maintaining view consistency is sharing information across multiple images from different viewpoints. Inspired by [15], we modify the self-attention mechanism in the diffusion model to ensure that each view incorporates sampling information from related views during the sampling process. During the sampling process, denote the queries, keys and values derived from the deep feature of the denoising U-Net for viewpoint " + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "inline_equation", + "content": "Q_{n}" + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "inline_equation", + "content": "K_{n}" + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "inline_equation", + "content": "V_{n}" + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": ", respectively. For each viewpoint, assume that it has a total of " + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": " associated viewpoints (view that has overlap). Subsequently, the Related View-based Attention for viewpoint " + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 325, + 555, + 493 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 381, + 502, + 555, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 502, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 381, + 502, + 555, + 534 + ], + "type": "interline_equation", + "content": "\\operatorname {s o f t m a x} \\left(\\frac {Q _ {n} \\tilde {K} _ {n} ^ {T}}{\\sqrt {d}} \\tilde {V} _ {n}\\right), \\tag {7}", + "image_path": "316c1b0ad82808e36679ce65c138431b1572262d240b9140cf514c00576d9c50.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\tilde{K}_n = \\left[K_1K_2\\ldots K_R\\right]^T" + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\tilde{V}_n = \\left[V_1V_2\\ldots V_R\\right]^T" + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "content": ". In the first room-scale texturing stage, we consider the " + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "content": " related viewpoints for view " + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "content": " as its adjacent views, i.e., its left and right view. When sampling at instance scale, we define the related views as all the " + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 544, + 555, + 622 + ], + "type": "text", + "content": " predefined camera view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 633, + 395, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 395, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 395, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "In this section, we first describe the general implementation of the baseline methods used for comparison. Next, we provide an overview of the implementation details of our proposed method. Finally, we present our quantitative experiments, qualitative results, and ablation studies." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "578" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 553, + 507 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 553, + 507 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 553, + 507 + ], + "type": "image", + "image_path": "bf78b4c1a0cb9f69752dc94637fd32dd25d3383370c47a6a3be5ccfeeeeb724d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 514, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 555, + 559 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons. Text2Tex-H [6] suffers from occlusion and visible seams. Text2Tex-C [6] struggles to maintain style consistency across all instances. SceneTex [7] produces unrealistic texture and results in blurry regions. In contrast, our method generates high-quality texture while preserving overall style consistency across instances in the scene. Ceilings and back-facing walls are excluded for improved visualizations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 579, + 230, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 230, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 230, + 592 + ], + "type": "text", + "content": "4.1. Baseline Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 597, + 295, + 705 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "text", + "content": "- Text2Tex-H [6]: This variant of Text2Tex is implemented using a holistic text prompt provided by user and treats the entire room as a cohesive entity. The camera is positioned at the center of the room, with a field of view of " + }, + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "text", + "content": ", and the elevation angle is set to " + }, + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "text", + "content": ", and azimuth angles uniformly sampled from the range " + }, + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "inline_equation", + "content": "[0^{\\circ}, 360^{\\circ}]" + }, + { + "bbox": [ + 55, + 597, + 295, + 681 + ], + "type": "text", + "content": ", which results in a total of twelve perspectives." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 681, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 681, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 55, + 681, + 295, + 705 + ], + "type": "text", + "content": "- Text2Tex-C [6]: For comparative purposes, we employ the original configuration of Text2Tex, wherein each" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 579, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 579, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 321, + 579, + 555, + 662 + ], + "type": "text", + "content": "room instance, including the interior walls and furniture, is textured using an instance-specific text prompt. And for interior walls texturing, we utilize the text prompt such as: \"A Chinese style living room, no furniture\" and sampled 12 perspectives same as Text2Tex-H. All the textured instances are subsequently combined, forming the output textured room." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 663, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 663, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 314, + 663, + 555, + 700 + ], + "type": "text", + "content": "- SceneTex [7]: We compare against SceneTex using its original setup, while employing the same holistic prompt as utilized in our approach." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "579" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 91, + 227, + 488 + ], + "blocks": [ + { + "bbox": [ + 63, + 178, + 78, + 387 + ], + "lines": [ + { + "bbox": [ + 63, + 178, + 78, + 387 + ], + "spans": [ + { + "bbox": [ + 63, + 178, + 78, + 387 + ], + "type": "text", + "content": "\"A Cream style master bedroom\"" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 79, + 192, + 90 + ], + "lines": [ + { + "bbox": [ + 127, + 79, + 192, + 90 + ], + "spans": [ + { + "bbox": [ + 127, + 79, + 192, + 90 + ], + "type": "text", + "content": "Ours (W/O Attn)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 91, + 227, + 488 + ], + "lines": [ + { + "bbox": [ + 91, + 91, + 227, + 488 + ], + "spans": [ + { + "bbox": [ + 91, + 91, + 227, + 488 + ], + "type": "image", + "image_path": "4766f501f5344637745ee3097aeae72152506d41bddea829d503674777935f77.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 507, + 555, + 563 + ], + "lines": [ + { + "bbox": [ + 55, + 507, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 555, + 563 + ], + "type": "text", + "content": "Figure 5. Ablation studies on the multiview-consistency module. Synthesizing texture without the Related View-based attention leads to inconsistencies across views, as shown in the leftmost column. When synthesizing texture without MVIS, noticeable inconsistencies appear across different viewpoints (middle column). In contrast, our full method samples multi-view images with significantly stronger consistency. The areas of major inconsistency in the images are highlighted with red and blue boxes. In the same column, boxes of the same color indicate that they are close to each other in 3D space. Zoom in for the best view." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 248, + 91, + 382, + 487 + ], + "blocks": [ + { + "bbox": [ + 279, + 79, + 351, + 90 + ], + "lines": [ + { + "bbox": [ + 279, + 79, + 351, + 90 + ], + "spans": [ + { + "bbox": [ + 279, + 79, + 351, + 90 + ], + "type": "text", + "content": "Ours (W/O MVIS)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 248, + 91, + 382, + 487 + ], + "lines": [ + { + "bbox": [ + 248, + 91, + 382, + 487 + ], + "spans": [ + { + "bbox": [ + 248, + 91, + 382, + 487 + ], + "type": "image", + "image_path": "dc70eb03cd093f87aa7e09a55473049c93de1e773bf5fb8e032200e289c41abc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 406, + 91, + 539, + 487 + ], + "blocks": [ + { + "bbox": [ + 451, + 79, + 494, + 91 + ], + "lines": [ + { + "bbox": [ + 451, + 79, + 494, + 91 + ], + "spans": [ + { + "bbox": [ + 451, + 79, + 494, + 91 + ], + "type": "text", + "content": "Ours (Full)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 91, + 539, + 487 + ], + "lines": [ + { + "bbox": [ + 406, + 91, + 539, + 487 + ], + "spans": [ + { + "bbox": [ + 406, + 91, + 539, + 487 + ], + "type": "image", + "image_path": "4f98974ec98f7fd41d39d9e2178e7749a2571b4d4331d8c9102884943fba2207.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 583, + 188, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 188, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 188, + 596 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "In this experiment, we used the SDXL model [28] as the text-to-image generator and incorporated depth information through ControlNet [1, 55]. For the text prompt, we implemented view-specific prompt combining the holistic prompt with the names of the types of objects present within the field of view. For sampling, we applied the DDPM [17] as the sampler, setting the sampling steps to 50. We used PyTorch3D [31] for rendering and texture projection. Additionally, UV atlas generation for the mesh was handled" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 584, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 555, + 620 + ], + "type": "text", + "content": "with Xatlas [47]. All experiments were performed on an NVIDIA RTX A6000 GPU. Further implementation details are available in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 633, + 439, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 439, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 439, + 646 + ], + "type": "text", + "content": "4.3. Quantitative Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Our quantitative experiment were conducted across 10 scenes including 5 bedroom and 5 living room from the 3D-FRONT dataset [13], using five different text styles totally. To evaluate the generated texture, we calculated both the CLIP Score(CLIP) [16], which assesses the alignment" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "580" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 167 + ], + "type": "table", + "html": "
MethodGeneration Time(mins) ↓CLIP ↑AS ↑
Text2Tex-H [6]8.5021.584.34
Text2Tex-C [6]70.7521.934.85
SceneTex [7]2614.5021.874.75
(Ours) Full46.0023.475.03
(Ours) w/o Attn-23.325.01
(Ours) w/o MVIS-23.275.01
(Ours) w/o MVRS-22.394.40
", + "image_path": "4429644c7a34f76f38f37ecf8093098ecc3d13e5297ff6ac1c578af60c0d8c20.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 57, + 176, + 294, + 231 + ], + "lines": [ + { + "bbox": [ + 57, + 176, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 57, + 176, + 294, + 231 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons. We report the Generation time and 2D metrics result for quantitative comparisons. The generation time is shown in the format minutes. The 2D metrics include: CLIP Score (CLIP) [16], Aesthetic Score(AS) [35]. We show that our method produces high quality textures efficiently." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 255, + 294, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 255, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 294, + 588 + ], + "type": "text", + "content": "between generated texture and the textual description, and the CLIP Aesthetic Score(AS) [35], which measures texture quality. We also compared the average processing time of each texture generation method in minutes. The results are summarized in Table.1. We render 10 images from novel views for each room to calculate the average of CLIP and AS. Due to the presence of numerous distinct objects within the room, the methods in baselines do not adjust the text prompts for texture generation based on the variation of objects from different viewpoints. Therefore, we employ a holistic prompt to compute the CLIP score. Since our evaluation metric involves image sampling from new viewpoints, Text2Tex-H, which fails to address the occlusion problem, generates texture containing large occluded regions, resulting in the lowest performance in terms of the metric. However, due to the limited number of sampled viewpoints, it exhibits the fastest generation speed. We configure SceneTex according to its original setup, sampling from a sphere centered at the room's center. As a result, regions not observed by SceneTex often exhibit blurriness. Meanwhile, in addition to optimizing the texture field using VSD, SceneTex also requires training LoRA, which results in the longest processing time. Text2Tex-C employs an instance-specific approach for texture generation, utilizing different prompts for each object, which results in higher performance metrics. Additionally, it incorporates viewpoint selection and texture optimization processes, leading to a longer computation time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 599, + 171, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 599, + 171, + 611 + ], + "spans": [ + { + "bbox": [ + 58, + 599, + 171, + 611 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 619, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 619, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 619, + 294, + 712 + ], + "type": "text", + "content": "We provide a visual comparison in fig.4. Specifically, Text2Tex-H [6] shows that simply applying object texture generation method to indoor scene texturing task cannot resolve the occlusion problem (see the large black areas in the first column of fig.4) and contain obvious seams. Text2Tex-C [6] texturing the room completely but suffering from obvious seams and global style consistency across objects. Although SceneTex introduces the use of cross-attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 73, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 167 + ], + "type": "text", + "content": "to generate texture for occluded regions, noticeable blurriness still persists in these areas. Furthermore, due to its optimization-based approach, SceneTex often produces unrealistic texture. In contrast, our method generates globally consistent texture with a unified style in the first stage. In the second stage, through per-instance inpainting and refinement, we successfully maintain global style consistency while addressing issues of occlusion and unrealistic texture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 175, + 413, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 413, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 413, + 186 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 194, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 216 + ], + "type": "text", + "content": "As shown in Fig.5, all images are raw samples from the final step, thus revealing inherent multi-view inconsistencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 217, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 335 + ], + "type": "text", + "content": "Effectiveness of the MVIS. In the first stage of texturing, we introduce MVIS to enhance the global consistency of the texture. Qualitative results from this stage are shown in the second column of Fig.5. The figure demonstrates that without MVIS, texture consistency across multiple viewpoints can be significantly impact. We further conduct quantitative ablation on dataset adopted in previous experiments. Results are shown in the seventh row of Tab.1. The quantitative experimental results clearly demonstrate that MVIS improves the quality of the generated texture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 337, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 443 + ], + "type": "text", + "content": "Effectiveness of the Related View-based Attention. We incorporate the Related View-based Attention module in both stages of texture generation to enhance consistency. Both qualitative(the first column of Fig.5) and quantitative experiments(the fifth row of the Tab.1) demonstrate that using only MVIS or MVRS is insufficient to generate consistent and high-quality texture. The Related View-based Attention module effectively enhances consistency while ensuring the quality of the generated texture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 445, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 563 + ], + "type": "text", + "content": "Effectiveness of the MVRS. In the second stage of texturing, we use MVRS to generate texture for the occluded regions of each instance and refine the texture generated in the first stage. Quantitative experimental results (seventh row in Tab.1) show that without MVRS, the quality of the texture significantly deteriorates. This is because, without MVRS, the texture generated in the first stage suffers from low-quality results due to occlusion. We will provide additional qualitative materials demonstrating the effectiveness of MVRS in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 574, + 386, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 574, + 386, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 574, + 386, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 594, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 594, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 594, + 553, + 712 + ], + "type": "text", + "content": "We propose an indoor scene texturing framework named RoomPainter that generates high-quality texture with multiview consistency and time efficiency. By leveraging the carefully designed modules including view-weighted texture rectification and related view-based attention, proposed view-integrated diffusion preserves the cross-view consistency during the diffusion process. Qualitative and quantitative results demonstrate the effectiveness of RoomPainter in indoor scene texturing, especially in global consistency and generation efficiency." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "text", + "content": "581" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] xinsir/controlnet-depth-sdx1-1.0. https://huggingface.co/xinsir/controlnet-depth-sdx1-1.0, 2023.4.7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "text", + "content": "[2] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects. arXiv preprint arXiv:2407.02430, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 294, + 290 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 292, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 294, + 346 + ], + "type": "text", + "content": "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4169-4181, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 402 + ], + "type": "text", + "content": "[6] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 468 + ], + "type": "text", + "content": "[7] Dave Zhenyu Chen, Haoxuan Li, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Scenetex: High-quality texture synthesis for indoor scenes via diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21081-21091, 2024. 1, 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 524 + ], + "type": "text", + "content": "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 568 + ], + "type": "text", + "content": "[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "text", + "content": "[10] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2920–2929, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[11] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[12] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[13] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[14] Chenjian Gao, Boyan Jiang, Xinghui Li, Yingpeng Zhang, and Qian Yu. Genesistex: Adapting image denoising diffusion to texture space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4620-4629, 2024. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "text", + "content": "[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "type": "text", + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 555, + 386 + ], + "type": "text", + "content": "[18] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "text", + "content": "[19] Dong Huo, Zixin Guo, Xinxin Zuo, Zhihao Shi, Juwei Lu, Peng Dai, Songcen Xu, Li Cheng, and Yee-Hong Yang. Texgen: Text-guided 3d texture generation with multi-view sampling and resampling. arXiv preprint arXiv:2408.01291, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 555, + 544 + ], + "type": "text", + "content": "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "text", + "content": "[22] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion. arXiv preprint arXiv:2311.12891, 2023. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 580, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 634 + ], + "type": "text", + "content": "[23] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "type": "text", + "content": "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "582" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "text", + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[27] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 249 + ], + "type": "text", + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "text", + "content": "[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[31] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[32] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 482 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 548 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "text", + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 656 + ], + "type": "text", + "content": "[36] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[37] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[38] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[39] Jiaxiang Tang, Ruijie Lu, Xiaokang Chen, Xiang Wen, Gang Zeng, and Ziwei Liu. Intex: Interactive text-to-texture synthesis via unified depth-aware inpainting. arXiv preprint arXiv:2403.11878, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[40] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[41] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[42] Qi Wang, Ruijie Lu, Xudong Xu, Jingbo Wang, Michael Yu Wang, Bo Dai, Gang Zeng, and Dan Xu. Roomtex: Texturing compositional indoor scenes via iterative inpainting. arXiv preprint arXiv:2406.02461, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "text", + "content": "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 421 + ], + "type": "text", + "content": "[44] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "text", + "content": "[45] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 476, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 533 + ], + "type": "text", + "content": "[46] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 551, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 551, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 551, + 544 + ], + "type": "text", + "content": "[47] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "text", + "content": "[48] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4347-4356, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[49] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[50] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, et al. Nofa: Nerf-based one-shot facial avatar reconstruction. 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "text", + "content": "[51] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "583" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 329 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Hifi-123: Towards high-fidelity one image to 3d content generation. arXiv preprint arXiv:2310.06744, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[52] Wangbo Yu, Chaoran Feng, Jiye Tang, Jiashu Yang, Zhenyu Tang, Xu Jia, Yuchao Yang, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[54] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 30(12):7749-7762, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "text", + "content": "[56] Haiyang Zhou, Xinhua Cheng, Wangbo Yu, Yonghong Tian, and Li Yuan. Holodreamer: Holistic 3d panoramic world generation from text descriptions. arXiv preprint arXiv:2407.15187, 2024. 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "584" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_content_list.json b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6f4ab24c4775b7a2af43dd73826e2fa36461cb2 --- /dev/null +++ b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_content_list.json @@ -0,0 +1,1574 @@ +[ + { + "type": "text", + "text": "RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation", + "text_level": 1, + "bbox": [ + 204, + 130, + 794, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingfei Han\\*1,3, Liang Ma\\*1, Kamila Zhumakhanova\\*1, Ekaterina Radionova\\*1, Jingyi Zhang\\*2 \nXiaojun Chang\\*1,4, Xiaodan Liang\\*1,2, Ivan Laptev\\*1 \n $^{1}$ Department of Computer Vision, MBZUAI $^{2}$ Shenzhen Campus of Sun Yat-Sen University \n $^{3}$ ReLER Lab, AAII, UTS $^{4}$ University of Science and Technology of China \nhttps://roomtour3d.github.io", + "bbox": [ + 127, + 202, + 870, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 340, + 326, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-and-Language Navigation (VLN) suffers from the limited diversity and scale of training data, primarily constrained by the manual curation of existing simulators. To address this, we introduce RoomTour3D, a video-instruction dataset derived from web-based room tour videos that capture real-world indoor spaces and human walking demonstrations. Unlike existing VLN datasets, RoomTour3D leverages the scale and diversity of online videos to generate open-ended human walking trajectories and open-world navigable instructions. To compensate for the lack of navigation data in online videos, we perform 3D reconstruction and obtain 3D trajectories of walking paths augmented with additional information on the room types, object locations and 3D shape of surrounding scenes. Our dataset includes $\\sim 100K$ open-ended description-enriched trajectories with $\\sim 200K$ instructions, and $17K$ action-enriched trajectories from 1847 room tour environments. We demonstrate experimentally that RoomTour3D enables significant improvements across multiple VLN tasks including CVDN, SOON, R2R, and REVERIE. Moreover, RoomTour3D facilitates the development of trainable zero-shot VLN agents, showcasing the potential and challenges of advancing towards open-world navigation.", + "bbox": [ + 89, + 375, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past years, Vision-and-Language Navigation [2, 21, 24, 32, 42, 54] has largely relied on human-designed simulators and annotated trajectories. R2R [2] initiated a benchmark for language-guided navigation in simulated indoor settings, while CVDN [51], REVERIE [42], and SOON [66] expanded VLN to dialogue-based and object-focused tasks. However, these curated simulations lack scene diversity and fail to capture real-world complexity.", + "bbox": [ + 89, + 758, + 482, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address limited diversity, recent methods propose the use of richer and more varied training data. AirBERT [15] combines discrete Airbnb images for panoramic views, which lack consistency and naturalistic context of an indoor scene. ScaleVLN [55] utilizes laboriously curated 3D scenes [45, 58], but suffers from reconstruction quality and scalability. More recently, YTB-VLN [34] attempts to use video frames to compose panoramic views and organize instructions with predefined templates, yet overlooks object variety and geometry structure. NaVid [60] constructs sequential single-view trajectories from MatterPort3D [4] and R2R [2] annotations, paired with general video data to train a sim-to-real agent. None of these approaches simultaneously achieves scalability in scene diversity, openness in object variety, or comprehensive geo-perception in spatial representations, each of which is critical to training effective and open-world navigation agents.", + "bbox": [ + 511, + 327, + 906, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the challenge, we introduce RoomTour3D, a novel dataset that provides a geometry-aware, spatially enriched training environment for VLN agents. Built upon easily accessible room tour videos from the Internet, RoomTour3D captures continuous movement through real estates with a hand-held camera from a first-person perspective. Each frame presents a realistic, agent-centric view and showcases a rich array of indoor items. The continuous flow of these frames captures multiple views of the environment, presenting diverse room layouts and inherently embedding the geometric properties of the spaces. To unleash the power of these videos, we propose an automatic and extendable pipeline to obtain open-ended geometry-aware human walking trajectories, spatially contextualized textual instructions using open vocabularies.", + "bbox": [ + 511, + 590, + 908, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better model the navigation scenario, we take advantage of the continuous walk-through trajectories and densely sample frames from room tour videos. Then, we use COLMAP [46, 47] to reconstruct 3D scenes of real-estates to obtain the geometric information. With known", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Co-first authors.", + "bbox": [ + 114, + 887, + 215, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27586", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "camera locations and orientations, we sample \"decision-making\" frames at points of maximal yaw rotation and sequence further frames every $\\sim 1.5$ meters to finalize trajectories. Additionally, our pipeline uses extensive annotations via RAM [61] for object tagging, Grounding-DINO [38] for precise localization, and Depth-Anything [59] for estimating relative object-camera distances. To integrate knowledge of object variety, geometric awareness, and human walking preferences into model training, we employ GPT-4 [40] to generate navigation instructions for both summarization and task-specific navigation tasks.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our RoomTour3D is an effort to create a comprehensive database derived from room tour videos and enriched by human-living knowledge. The dataset includes $\\sim 100\\mathrm{K}$ open-ended trajectories with $\\sim 200\\mathrm{K}$ descriptions, and $\\sim 17\\mathrm{K}$ geometry-aware trajectories with navigable actions from 1847 homes. Moreover, we are releasing intermediate products such as object tags, bounding boxes, depth maps, room locations, and the necessary code and prompts used to generate instructions. To validate the robustness, we conducted experiments with NaviLLM [63], a generalist model based on Large Language Model (LLM), to train a unified multi-task navigation agent. Integrating our data improved baseline performances on CVDN, SOON, R2R, and REVERIE by over $6\\%$ , notably achieving a $9.8\\%$ boost on SOON and setting new state-of-the-art results. Furthermore, our enriched action-instruction data facilitates training an end-to-end zero-shot navigation agent, advancing towards open-world embodied navigation.", + "bbox": [ + 91, + 258, + 483, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we make the following key contributions:", + "bbox": [ + 109, + 531, + 472, + 546 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Video Collection for Complex Environments: We curate a novel dataset of diverse videos tailored for navigation tasks, distinguishing it from existing datasets such as YTB-VLN [34]. Our dataset features longer videos, representing more complex environments, and has fewer shot changes, ensuring continuity and contextual consistency.", + "- Automatic Video Processing: We propose a pipeline to extract geometry-aware navigation instructions, aligning spatial understanding with navigation goals. Additionally, we generate open-vocabulary instructions for open-ended trajectories to enhance real-world applicability.", + "- Demonstrating Data Effectiveness: Through extensive experiments and ablation studies, we demonstrate that our dataset significantly improves the performance of state-of-the-art models." + ], + "bbox": [ + 89, + 547, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 791, + 232, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-and-Language Navigation", + "text_level": 1, + "bbox": [ + 89, + 816, + 380, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Navigating unseen indoor environments with natural language instructions is vital for enabling embodied agents to assist humans. Various scenarios have been explored, including fine-grained instruction following (R2R [2]) and", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dialogue-based navigation (CVDN [51]), object localization from instructions (SOON [66], REVERIE [42]), and embodied question answering through active 3D exploration [10, 56]. While substantial work focuses on task-specific models [1, 6, 7, 14, 16-20, 25, 30, 37, 39, 43, 44, 52, 62, 66], they often lack generalization across tasks. In this context, NaviLLM [63] introduces an embodied generalist model that simultaneously addresses multiple tasks through a single framework, demonstrating strong generalization ability.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Data-Centric Methods for VLN", + "text_level": 1, + "bbox": [ + 511, + 250, + 792, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The scarcity of VLN training data remains a critical issue and results in poor generalization of VLN agents to unseen environments. Most of existing VLN datasets such as R2R [2], RxR [24], CVDN [51] and SOON [66] are produced in simulators, which constrains data scalability due to the high labor costs involved. To tackle the problem, data augmentation [12, 13, 23, 27, 28, 35, 49] and self-exploration in simulator environments [31, 53] have been investigated.", + "bbox": [ + 511, + 268, + 903, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLN-BERT [17] and AirBERT [15] attempt to use web-based image-caption pairs for pre-training, however, resulting trajectories often fail to mimic realistic navigation. Similarly, automatic dataset generation pipelines [8, 22], including ScaleVLN [55], rely on manually curated 3D scenes or synthetic environments, which are costly to produce and lack the photorealism needed for robust real-world generalization. PanoGen [26] enhances VLN training by generating diverse text-conditioned panoramic environments using text-to-image diffusion models and recursive outpainting. While this approach addresses the scarcity of training environments, it relies on synthetic panoramas and may not generalize well to real environments. YTB-VLN [34] advances scalability by leveraging YouTube room tour videos to generate path-instruction pairs but omits explicit path geometry, essential for robotic navigation.", + "bbox": [ + 511, + 405, + 905, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our work, we address the limitations by designing RoomTour3D with properties: (i) free-form and open-vocabulary path annotations instead of template-based instructions, (ii) extraction of open-ended trajectories from sequential video clips, and (iii) inclusion of turning points and spatially close frames as navigable candidate actions, moving beyond panoramic nodes. Furthermore, our approach integrates 3D reconstruction of indoor videos to retrieve trajectory geometry and employs an LLM to generate detailed, object-aware instructions with enhanced spatial understanding.", + "bbox": [ + 511, + 647, + 905, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Zero-shot Navigation", + "text_level": 1, + "bbox": [ + 511, + 821, + 712, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the substantial semantic variations in complex real-world scenarios, fully-supervised VLN models often struggle to generalize across diverse navigation scenes. Zero-shot VLN has thus gained attention as it eliminates prior", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8af1a3f33c14c429498d63dcb7b01f77acca153579fa8dc4ceb4ca5320fa13e2.jpg", + "image_caption": [ + "Figure 1. Overview of our RoomTour3D data generation. Starting from a room tour video, we first apply BLIP-2 [29] on frame sequence to predict the room locations. Next, we use RAM [61] and Grounding-DINO [38] to identify objects within the frames and employ Depth-Anything [59] for depth prediction. Subsequently, COLMAP is used to reconstruct the 3D scene with complete geometry information, and we sample human walking trajectories from the continuous frames. The trajectory captures open-world objects, their positions, and depths relative to the camera. Finally, we use advanced LLM, i.e., GPT-4 to generate the free-form descriptions for pretraining, namely description-enriched trajectories. Specifically, for the trajectory shown in the figure, which involves instant turning points, we specially treat $<0>$ to $<6>$ as walking trajectory, $<\\mathrm{A}> <\\mathrm{B}>$ and $<\\mathrm{C}>$ as side-watching points and use them as negative candidates for navigation finetuning task, namely action-enriched trajectories. For more details, please refer to Section 3." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 883, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "knowledge of environments and instructions, effectively mitigating environmental biases.", + "bbox": [ + 88, + 542, + 482, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Commercial model based methods utilize advanced LLMs and robust frameworks for seamless solutions. MapGPT [5] incorporates a map-based prompting system with global spatial reasoning and adaptive path planning. DiscussNav [39] employs a multi-expert framework where LLMs specialize in subtasks like instruction analysis and vision perception. NavGPT [64] focuses on explicit reasoning by combining commonsense reasoning with visual observations. As to non-commercial methods, LangNav [41] uses language as the primary perceptual space, while NavCoT [33] introduces parameter-efficient training to allow LLMs to autonomously reason and act.", + "bbox": [ + 88, + 571, + 482, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show that usage of our action-enriched data for navigation tuning results in superior zero-shot performance over all non-commercial methods and reaches comparable results to commercial approaches based on GPT-3.5.", + "bbox": [ + 88, + 753, + 482, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. RoomTour3D", + "text_level": 1, + "bbox": [ + 89, + 818, + 232, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the automatic data curation pipeline of RoomTour3D. We detail the process of annotations, from sampling open-ended human walking trajectories to generating corresponding descriptions with open-", + "bbox": [ + 88, + 840, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "world object variety and spatial awareness. Enabled by reconstructed 3D scenes, we further sample navigable trajectories with actions. The overall pipeline of our data generation is illustrated in Figure 1. Please refer to Appendix 7 for details about video collection.", + "bbox": [ + 511, + 542, + 906, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Description-Enriched Trajectories", + "text_level": 1, + "bbox": [ + 511, + 625, + 813, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this subsection, we detail the process of generating controllable descriptions for open-ended trajectories. We start by generating human-walking trajectories by uniformly sampling frames at a rate of one frame every two seconds, which aligns with the average human walking speed of 1.42 meters per second [57], typically slower in indoor environments. Subsequently, to annotate these trajectories, as shown in Figure 2, we employ expert models such as BLIP-2 [29], RAM [61], Grounding-DINO [38], and Depth-Anything [59] to gather extensive information on object variety, spatial positions, and depth measurements. Finally, we integrate this information into GPT4 [40] to generate detailed and coherent traje.", + "bbox": [ + 509, + 643, + 906, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Variety and Spatial Awareness. In order to harness object variety and enable spatial awareness, we compose three expert models and design a textual template, i.e., \"There is a object tag to the spatial position of current spot", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27588", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/911aaf5c0c530d215601da82152c47e13079298286803f71626c19a04793eb77.jpg", + "image_caption": [ + "Figure 2. Instruction generation in a controllable way. (a) Using open-sourced expert models, we identify what objects are in the frames, and assess how far an object is and determine where an object is located. The information is then textualized to create richly detailed frame captions. (b) BLIP-2 is adopted to predict and smooth room location across sequential frames. (c) Combining room locations and object information, we use GPT-4 for controllable and open-vocabulary instruction generation. The prompt consists of a task instruction that defines the generation task, and in-context examples that constrain the output style." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 901, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in relative distance\", to organize the multi-source information to ease GPT generation. Firstly, we used RAM [61] to annotate the object categories within the frames. Based on these category tags, we employed Grounding DINO [38] to locate the objects in the frames. Subsequently, we used Depth-Anything [59] to predict the depth maps corresponding to the frames.", + "bbox": [ + 88, + 438, + 480, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using this data, we identify the spatial locations and distances of objects relative to the current camera position. By analyzing object bounding box center positions and depth map locations, we can generate frame captions, as illustrated in Figure 2. Finally, objects in different frames can be easily correlated and capture the progression across different frames. More details in the spatial awareness data generation are provided in Appendix 9.", + "bbox": [ + 88, + 547, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Room Location Annotation in Videos. To determine the camera of each frame, w.r.t. the room category, we used BLIP2 [29] in visual question-answering mode, posing the question, \"Which room am I in?\" A predefined list of 16 common room types (i.e., bathroom, kitchen, bedroom, living room, hallway, office, dining room, foyer, laundry room, outside, porch, garage, front, patio, driveway, and backyard) was used as possible answers. This list was curated by analyzing 10 randomly sampled long videos, using BLIP2 in generative mode to identify and rank the most frequently mentioned room types. For frame-level predictions, we switched to BLIP2's discriminative mode and applied temporal smoothing to denoise outputs.", + "bbox": [ + 88, + 670, + 482, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We validated this approach by manually annotating 50 video clips, achieving an accuracy of $85\\%$ . The use of", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BLIP2 leverages its open-world knowledge, while limiting room types to 16 categories for discriminative selection simplifies outputs and reduces ambiguity. Any loss in open-vocabulary flexibility is addressed during GPT-based trajectory summarization.", + "bbox": [ + 511, + 438, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Controllable Instruction Generation. To generate descriptions that accurately capture human-walking trajectories and reflect the environment, we integrate frame-level room locations with frame captions composed of object descriptions. We then employ GPT-4-Turbo [40] for controllable instruction generation, leveraging the multi-source information contained in the composed captions. As depicted in Figure 2, we organize the prompt using the \"Task instruction - In-context examples - Prediction\" scheme. This approach defines our instruction generation task as describing object progression along the moving trajectories, and includes two examples to ensure GPT produces instruction-style texts only. As shown in Figure 2 (c), captions of frames along the trajectory are embedded into the prompt and input into GPT.", + "bbox": [ + 511, + 515, + 908, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Action-Enriched Trajectories", + "text_level": 1, + "bbox": [ + 511, + 744, + 774, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Environment Reconstruction. To obtain the geometric information of trajectories within RoomTour3D, we employ COLMAP [46, 47] for 3D reconstruction. This process allows us to infer the 3D layout of environments in the videos, providing a detailed geometric context for navigation tasks. Specifically, we sample the videos at 3 frames per second to balance accuracy and execution time. To further improve time efficiency, we split the videos into 100-second video clips with 10-second overlaps between adjacent clips and", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27589", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e19f7087947122c0d3fc40f131c599bd7334c845cd723cacb4e816e05627cfa3.jpg", + "image_caption": [ + "(a) Pretraining with Summarization Task" + ], + "image_footnote": [], + "bbox": [ + 94, + 109, + 501, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output:", + "text_level": 1, + "bbox": [ + 101, + 207, + 155, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Move through the corridor and step forwards.Pass by a paint on the wall. Then advance through the bedroom and into the bathroom.Start with a view of the bed,curtain..", + "bbox": [ + 101, + 220, + 250, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1838f762e8d572fe34dfde7ee2d1506cdd8c4333250bf16271252c30888f763f.jpg", + "image_caption": [ + "Embodied Agent" + ], + "image_footnote": [], + "bbox": [ + 290, + 224, + 325, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output:", + "bbox": [ + 292, + 275, + 344, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 295, + 291, + 354, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summarization", + "bbox": [ + 403, + 282, + 503, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Navigation", + "bbox": [ + 403, + 297, + 480, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": ":", + "bbox": [ + 509, + 135, + 555, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5a07fc8369152ba31dab3d9612dc1d8aaab2dfe3a1a4560859ee1169b74c1159.jpg", + "image_caption": [ + "(b) Finetuning with Navigation Task" + ], + "image_footnote": [], + "bbox": [ + 558, + 112, + 630, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9538f1c804d37c0a76cb42c6b789daf2d8c94bd50f6e3fed3d316fede151c528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 112, + 712, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e118a104bf57f46bbeeb47c8f83214436bab7676019c225eac5490a517df05d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 112, + 797, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7132859559d05b920b5f5f44cdd5d42af2e1e90601e05ae4b3886b993d73f3ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 831, + 113, + 893, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Move through the corridor and step forwards... advance through the bedroom ... Start with a view of the bed, curtains... Move closer, noting the headboard... Pass the dresser and ...", + "bbox": [ + 558, + 159, + 895, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task:", + "bbox": [ + 517, + 229, + 553, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compare the History and Instruction to infer your current progress, and then select the correct direction from the candidates to go to the target location.", + "bbox": [ + 558, + 205, + 874, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "< Cand>:", + "bbox": [ + 509, + 270, + 558, + 281 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eceadb299b02e2ebd900bf6fec1c6e85e7623a658b5fad72f2ef98223754423a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 250, + 630, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9cc547a5a38a2057ed7f3e72fb00ffe50e1f8789d9af34313daccdbd6fef4f1d.jpg", + "image_caption": [ + "Figure 3. Model training diagram with RoomTour3D. We design two tasks for our RoomTour3D to boost NaviLLM. (a) Pretraining: Sampled frames on the trajectory are treated as candidate observations. Model is optimized to summarize object progression along the path. (b) Finetuning: Each frame is considered as a navigable step. Given historical observation $<0>$ to $<2>$ and navigation instruction, the model is prompted to predict the next action by selecting from candidate observations View A, View B, View C and View D." + ], + "image_footnote": [], + "bbox": [ + 648, + 250, + 717, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1660967e4f96da8597f609b3aaa90d22a702bb2340d61c876ba5e7262db9e310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 248, + 803, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbd8c90c1b6c5ed22a2425cb4c1f7b969152dc941be7c90b9e02bde90453f9d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 248, + 888, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "perform reconstruction on the clips simultaneously.", + "bbox": [ + 89, + 393, + 429, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following this, we merge the resulting sub-models reconstructed from the video clips. For every two adjacent clips, if the reconstructed models have more than three overlapping frames, the models are merged and readjusted into one. However, due to varying reconstruction quality, a single video clip can produce more than one model. To manage this, we construct a graph for merging these sub-models. Each model is treated as a node, and we connect two models if they have more than three overlapping frames. We then apply Depth-First Search [50] to merge any two connected models and replace the original model nodes with the new merged one, continuing this process until no connected nodes remain.", + "bbox": [ + 89, + 412, + 483, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Navigable Action Sampling. We enhance navigation action diversity by using frames from room tour videos as navigable actions by sampling at significant view-change points within a small radius. These points are identified by reconstructing 3D scenes to measure camera orientation differences and distances between frames, capturing varied views from revisited locations or turning points. Frames with substantial view changes are retained using cosine similarity thresholding, followed by non-maximum suppression to isolate major shifts. DBSCAN clustering [11] groups spatially close frames with different views to ensure diverse navigable actions. These measures ensure the robustness to the misidentification of significant view change points that may arise due to inaccuracies in 3D reconstruction. Finally, we identify distinct walking paths within each cluster. For each path, we select the most recent frame as a positive candidate and the frame with the highest angular difference as a negative candidate, enhancing the diversity of navigable actions. More detailed description of our navigable points", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sampling approach are provided in Appendix 8.", + "bbox": [ + 513, + 393, + 828, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Vision-and-Language Navigation Model", + "text_level": 1, + "bbox": [ + 511, + 421, + 872, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce a practice to use our data to train a generalist embodied agent. To start with, we first provide a concise introduction to the state-of-the-art VLN model, NaviLLM [63], which is an LLM-based navigation agent. Then, we introduce two tasks that are adapted for our RoomTour3D data, i.e., vision-instruction summarization task for pretraining and action-instruction navigation task for finetuning.", + "bbox": [ + 511, + 446, + 906, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Revisiting NaviLLM", + "text_level": 1, + "bbox": [ + 511, + 575, + 707, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NaviLLM is a SOTA LLM-based model for embodied navigation, excelling on benchmarks like CVDN and SOON. It processes panoramic inputs by encoding environmental views and integrating them with navigation instructions. Specific tokens are defined for different types of inputs: for historical observations and for candidate views at each navigational step.", + "bbox": [ + 511, + 598, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, NaviLLM receives instructions and a sequence of candidate views. At each step, the candidate observations are input into the model along with the instructions. The model predicts the next action by selecting the appropriate view from the candidates, and this selected view is then cached as a token, updating the model's internal state for future decisions. At the last step of the navigation task, the model summarizes all tokens as a separate training task to ensure comprehensive understanding and retention of the navigational history. For testing, the model similarly uses historical observations, i.e., tokens, accumulated during navigation, and candidate views, i.e., , at each step to decide the next", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27590", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "action. This process ensures that the model's actions align with the given instructions and the observed environment.", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Summarization Task for Pretraining", + "text_level": 1, + "bbox": [ + 89, + 130, + 406, + 147 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To leverage the rich, sequential nature of videos and enhance future planning capabilities, we adapt NaviLLM to use the description-enriched trajectories from our RoomTour3D dataset for a summarization task. Each frame is treated as a candidate view and wrapped with $\\langle \\text{cand} \\rangle$ tokens. Similar to selected panoramic views at each navigation step, these frames are considered as selected actions when executing navigation instructions. As shown in Figure 3(a), the frame tokens and summarization task instruction are organized into a unified prompt and input into the LLM. The model is expected to output a trajectory summary containing object progression and room locations, as specified by the task instruction. The model is trained using next-token prediction loss, consistent with the original language model training methodology.", + "bbox": [ + 89, + 152, + 482, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Navigation Task for Finetuning", + "text_level": 1, + "bbox": [ + 89, + 387, + 366, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to enable learning navigation decision-making from scalable scenes, we adapt NaviLLM to action-enriched trajectories from our RoomTour3D. Unlike panoramic views capturing observations from a single location, our data provides candidate views from frames at different locations and orientations, with only one frame directed toward the destination. Each frame in the video sequence is treated as a potential navigable action and wrapped with $\\langle \\text{cand} \\rangle$ tokens. These candidate views are presented to the model and processed in the same way as panoramic views. As shown in Figure 3(b), the model processes the inputs to predict the next action, selecting the appropriate frame from the candidate views. The selected action is then cached as a $\\langle \\text{hist} \\rangle$ token for subsequent decision-making steps.", + "bbox": [ + 89, + 409, + 482, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During fine-tuning, each frame is treated as a navigable step, with the next trajectory frame as the target action and $\\langle \\text{STOP} \\rangle$ as an alternative. The model uses historical observations and navigation instructions to iteratively predict the next action, building a detailed understanding of the path. At the final step, the model summarizes the navigation path, incorporating object progression and room locations. This summarization task enhances its ability to recall navigational history and improves performance.", + "bbox": [ + 89, + 636, + 482, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Tasks and Experiments", + "text_level": 1, + "bbox": [ + 89, + 784, + 310, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section outlines our experimental setup and presents the results. Detailed implementation information can be found in Appendix 11.", + "bbox": [ + 89, + 810, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. During pretraining, we follow practice from NaviLLM [63] and perform teacher-forcing training on the combined dataset from our video-instruction data from", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RoomTour3D, together with CVDN [51], SOON [66], R2R [2], REVERIE [42] and ScanQA [3], and augmented data from R2R and REVERIE. In the multi-task fine-tuning stage, we alternate between teacher forcing and student forcing on the combined data from our action-instruction data from RoomTour3D, together with CVDN, SOON, R2R, REVERIE, ScanQA and LLaVA-23k [36].", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the impact of our data on navigation agent training, we test on CVDN, SOON, R2R, and REVERIE. CVDN requires navigating towards a target by understanding dialog history, linking dialogue comprehension to actions. SOON tasks the agent with locating objects without bounding boxes, emphasizing semantic-visual alignment. R2R involves following step-by-step instructions, requiring dynamic progress tracking and precise alignment with navigational history. REVERIE focuses on localizing distant objects based on concise instructions, aided by ground truth bounding boxes at waypoints.", + "bbox": [ + 511, + 198, + 903, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For the navigation tasks, we follow the evaluation methodology from [2] using the following navigation metrics: Success Rate (SR), which measures whether the agent reaches the target location within a set distance threshold; Success Rate Weighted by Path Length (SPL), which is the SR adjusted by the ratio of the ground truth path length to the actual path traveled; Goal Progress (GP), the advancement in meters towards the goal. GP is utilized for the CVDN dataset, whereas SR and SPL are the metrics for other datasets.", + "bbox": [ + 511, + 364, + 903, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison on Supervised Tasks", + "text_level": 1, + "bbox": [ + 511, + 529, + 805, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, we performed a one-time fine-tuning on the four tasks in a fully supervised manner. To begin, our experiments reiterate the superiority of multitask training over single-task training. Also, incorporating our RoomTour3D data into the pre-training process led to consistent improvements across all metrics on Val-U, achieving state-of-the-art results in the GP metric in the CVDN dataset.", + "bbox": [ + 511, + 551, + 903, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, finetuning with our action-enriched data results in state-of-the-art performance on both Val-U and Test sets across SOON, R2R and REVERIE tasks. While the improvement on the CVDN and SOON datasets is modest, the most significant boost compared to the reproduced baseline is observed in R2R Val-U and REVERIE Val-U, with gains of approximately $5.7\\%$ and $6\\%$ , respectively. The improvement in R2R is largely driven by enhanced spatial awareness, stemming from the inclusion of proximity data, which helps the model better understand object distance and position. Similarly, gains in REVERIE are attributed to a combination of open-vocabulary tags, spatial awareness, and the addition of room type data, which encourages the model to infer the layout of environments, thereby boosting its spatial reasoning capabilities. Moreover, our use of open-ended instructions allows the model to adapt flexibly to diverse", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f043ae34705e56b2392b05c6876deea666c3c6f6d6fc36fb6b537f5fbf05a2b2.jpg", + "table_caption": [ + "Table 1. Overall comparison with the baseline methods. Our RoomTour3D data can boost NaviLLM by a margin on SOON, R2R and REVERIE on SPL metric and on CVDN GP metric. *denotes reproduced results. RT3DDesc and RT3DAction stand for description-enriched trajectories only and action-enriched trajectories." + ], + "table_footnote": [], + "table_body": "
MethodsCVDNSOONR2RREVERIE
Val-UTestVal-UTestVal-UTestVal-UTest
Models Focusing on Single Task
PREVALENT [16]3.152.44--5351--
HOP [43]4.413.24--575926.124.3
HAMT [6]5.135.58--616030.226.7
DUET [7]--22.621.4605833.736.0
VLN-SIG [25]5.525.83--6260--
VLN-PETL [44]5.696.13--605827.726.7
NavGPT2 [65]----6160--
BEV-BERT [1]----646036.436.4
Unified Model For All Tasks
NaviLLM(w. Pretrain) [63]6.167.9029.226.3596035.732.3
NaviLLM(w. Pretrain)*6.09-28.0-56.7-31.4-
NaviLLM+RT3DDesc(Ours)6.967.5530.226.562.361.837.135.1
NaviLLM+RT3DAction(Ours)6.337.2231.727.862.462.237.436.4
", + "bbox": [ + 133, + 135, + 862, + 391 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e06ffe75a18e4dc9dcc5d65649ca41ab29924edaf7693e3a407c83df405029a.jpg", + "table_caption": [ + "Table 2. Ablation study on the input modalities for trajectory summarization task." + ], + "table_footnote": [], + "table_body": "
Object tagsDepth & Bounding BoxRoom typeCVDN GP↑SOONR2RREVERIE
SR↑SPL↑SR↑SPL↑SR↑SPL↑
×××6.0933.6428.0165.5256.6738.3231.35
✓××5.4132.5226.5163.6155.7642.5234.37
✓✓×6.4937.6230.4068.3761.7041.7236.04
✓✓✓6.9638.8030.2169.3762.2843.2537.10
", + "bbox": [ + 125, + 417, + 867, + 518 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65ceaa449853a6aa73f6f8d14f01bb60e56ff7331609c2e450b02b7c5805a808.jpg", + "table_caption": [ + "Table 3. Overall comparison with SOTA zero-shot methods on R2R. ${}^{ \\dagger }$ denotes training exclusive navigable actions. ${}^{ * }$ denotes using 36 views setting. RT3D stands for our RoomTour3D." + ], + "table_footnote": [], + "table_body": "
MethodsVal Unseen
SR↑SPL↑
Random Walk [41]32
Commercial Model
NavGPT(GPT-3.5) [64]*13.899.12
NavGPT(GPT-4) [64]3429
MapGPT(GPT-4) [5]38.825.8
MapGPT(GPT-4V) [5]43.734.8
DiscussNav(GPT-4) [39]4340
Open-source Model
LangNav(LLaMA2-7B) [41]00
NavCoT(LLaMA2-7B) [33]7.786.50
DuET (Init. LXMERT [48])10
NaviLLM [63]†00
NaviLLM+RT3D(Ours)14.3310.86
", + "bbox": [ + 122, + 583, + 444, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scenarios, fostering more robust and generalizable performance and better contextual understanding.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison on Zero-shot Task", + "text_level": 1, + "bbox": [ + 511, + 532, + 787, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the substantial indoor knowledge contained in our data and its effectiveness for embodied action and language instructions, we conduct zero-shot experiments on embodied action prediction, as shown in Table 3.", + "bbox": [ + 511, + 556, + 905, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We removed all action and geometric data from the training datasets and retrained NaviLLM with and without our RoomTour3D dataset. Without action prediction data, NaviLLM lacked the ability to learn effective navigable action selection. However, with the inclusion of our action-enriched trajectories, NaviLLM achieved a $14.33\\%$ SR and a $10.86\\%$ SPL, outperforming open-source models built on LLaMA-7B and reaching results comparable to NavGPT [64], which leverages GPT-3.5. These improvements validate the effectiveness of our 3D trajectories mined from room tour video reconstructions and emphasize the value of our action-enriched trajectories. This highlights the significant contribution of our dataset to advancing open-world navigation.", + "bbox": [ + 511, + 619, + 906, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 845, + 663, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of open-world semantics and spatial awareness. As shown in Table 2, we analyzed the impact of various in", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27592", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruction: Exit the sewing room. Turn right. Go toward the glass cabinet with the dolls in it. Turn into the doorway on the left. Pass the bed and go through the next doorway on the left into the bathroom. Wait by the sink. (Instruction_id : 4676_0)", + "bbox": [ + 130, + 90, + 866, + 113 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9bf6007d195df537182c3350ca85990d0a7ef379298454e9619c57b7b63bba6.jpg", + "image_caption": [ + "Figure 4. Paths of NaviLLM [63] and ours on R2R-unseen. Purple and green circles denote the start and target locations, respectively, and the red circle represents incorrect endpoint. According to the instruction, the agent should turn left at the waypoint marked with yellow. Our method makes the correct decision, while the baseline is confused by similar entrance at the waypoint, thus mistakenly turns right." + ], + "image_footnote": [], + "bbox": [ + 133, + 114, + 549, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b08e8dfcb8e39667396b9957e364d71711d36f4edfc1075952c3cd28f3deea88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 114, + 864, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formation types on instruction generation. Adding object variety significantly improved performance on REVERIE with increase SPL from $31.35\\%$ to $34.37\\%$ , as this dataset relies on object grounding. However, it had no direct impact on SOON, possibly because SOON relies solely on detailed textual descriptions without explicit bounding box annotations. After introducing depth estimation, which helps determine the relative distances of objects, the performance on SOON, R2R and REVERIE achieve marginal boosts. This demonstrates that enhancing spatial awareness significantly contributes to indoor navigation tasks. Furthermore, incorporating room locations, which capture the scene semantics along the trajectory, provided a moderate boost across all four VLN tasks. This further highlights the critical role of object variety and spatial awareness in improving navigation performance.", + "bbox": [ + 88, + 388, + 485, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of action-instruction data. As shown in Table 1, we test the effect of incorporating video-action-instruction data into the training dataset. It is evident that this approach improves the SPL metric across the test scenes of SOON, R2R, and REVERIE. We believe that incorporating geometric information and movement-inclusive instructions helps the model better align the relationship between action and observation changes, thereby further enhancing the model's embodied capabilities.", + "bbox": [ + 89, + 628, + 483, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Data correctness verification", + "text_level": 1, + "bbox": [ + 89, + 773, + 346, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluated the correctness of our automated data-generation pipeline by manually rating 100 randomly sampled trajectory descriptions on a 4-point relevance scale: 1 for \"totally irrelevant\", 2 for \"partially relevant\", 3 for \"mostly relevant\" and 4 for a \"perfect match\". The evaluation yielded an average score of 3.08, with $74\\%$ of descriptions rated as \"mostly relevant\" or \"perfect match,\" demon-", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "strating the method's effectiveness in generating meaningful, visually aligned descriptions.", + "bbox": [ + 511, + 388, + 906, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Navigation Case Visualization", + "text_level": 1, + "bbox": [ + 511, + 430, + 779, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 4, selecting the correct action is critical at specific decision points, such as when a left turn is required to follow the instruction accurately. In the example, at step $②$ , both the rooms to the left and right could satisfy the latter part of the instruction, \"pass the bed and go into the bathroom.\" However, the baseline method incorrectly chooses a right turn at the designated left-turn point, causing it to deviate from the intended path. Once this error occurs, even with scene graph history, the model struggles to realign with the correct trajectory. This challenge is particularly common in household environments, where bedroom layouts often appear similar. It further demonstrates the effectiveness of our data alignment in improving adherence to action-based instructions.", + "bbox": [ + 511, + 453, + 908, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 681, + 633, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present RoomTour3D, a novel dataset automatically curated from room tour videos for VLN tasks. By leveraging the rich, sequential nature of video data and incorporating object variety and spatial awareness, we generate 200K navigation instructions and 17K action-enriched trajectories from 1847 room tour scenes. Additionally, we produce navigable trajectories from video frames and reconstructed 3D scenes, which significantly boost the performance and set new state-of-the-art results on the SOON and REVERIE benchmarks. This approach also enables the development of a trainable zero-shot navigation agent, demonstrating the effectiveness and scalability of RoomTour3D in advancing VLN research.", + "bbox": [ + 511, + 705, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. BEVbert: Multimodal map pre-training for language-guided navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2737-2748, 2023. 2, 7", + "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments, 2018. 1, 2, 6", + "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 5", + "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 5", + "[5] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee K Wong. Mapgpt: Map-guided prompting for unified vision-and-language navigation. arXiv preprint arXiv:2401.07314, 2024. 3, 7", + "[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 7", + "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation, 2022. 2, 7", + "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation, 2022. 2", + "[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 5", + "[10] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering, 2017. 2", + "[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, page 226-231. AAAI Press, 1996. 5, 1", + "[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation, 2018. 2", + "[13] Tsu-Jui Fu, Xin Eric Wang, Matthew Peterson, Scott Grafton, Miguel Eckstein, and William Yang Wang. Coun" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "terfactual vision-and-language navigation via adversarial path sampling, 2020. 2", + "[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14911-14920, 2023. 2", + "[15] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation, 2021. 1, 2", + "[16] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13137-13146, 2020. 2, 7", + "[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language BERT for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2", + "[18] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3d world into large language models, 2023.", + "[19] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6683-6693, 2023.", + "[20] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding, 2023. 2", + "[21] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. arXiv preprint arXiv:1905.12255, 2019. 1", + "[22] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning, 2023. 2", + "[23] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 2", + "[24] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 1, 2", + "[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 2, 7", + "[26] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27594", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "navigation. Advances in Neural Information Processing Systems, 36:21878-21894, 2023. 2", + "[27] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information, 2021. 2", + "[28] Jialu Li, Hao Tan, and Mohit Bansal. EnvEdit: Environment editing for vision-and-language navigation, 2022. 2", + "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 3, 4", + "[30] Xiangyang Li, Zihan Wang, Jiahao Yang, Yaowei Wang, and Shuqiang Jiang. KERM: Knowledge enhanced reasoning for vision-and-language navigation, 2023. 2", + "[31] Xiwen Liang, Fengda Zhu, Lingling Li, Hang Xu, and Xiaodan Liang. Visual-language navigation pretraining via prompt-based environmental self-exploration, 2022. 2", + "[32] Xiwen Liang, Liang Ma, Shanshan Guo, Jianhua Han, Hang Xu, Shikui Ma, and Xiaodan Liang. CorNav: Autonomous agent with self-corrected planning for zero-shot vision-and-language navigation, 2024. 1", + "[33] Bingqian Lin, Yunshuang Nie, Ziming Wei, Jiaqi Chen, Shikui Ma, Jianhua Han, Hang Xu, Xiaojun Chang, and Xiaodan Liang. NavCoT: Boosting llm-based vision-and-language navigation via learning disentangled reasoning. arXiv preprint arXiv:2403.07376, 2024. 3, 7", + "[34] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H. Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos, 2023. 1, 2", + "[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup, 2021. 2", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 6", + "[37] Rui Liu, Wenguan Wang, and Yi Yang. Volumetric environment representation for vision-language navigation. In CVPR, pages 16317-16328, 2024. 2", + "[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding Dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 4", + "[39] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. arXiv preprint arXiv:2309.11382, 2023. 2, 3, 7", + "[40] OpenAI. GPT-4: Generative pre-trained transformer 4. OpenAI API, 2022. 2, 3, 4, 1", + "[41] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. arXiv preprint arXiv:2310.07889, 2023. 3, 7", + "[42] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: Remote embodied visual referring expression in real indoor environments, 2020. 1, 2, 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: History-and-order aware pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15418-15427, 2022. 2, 7", + "[44] Yanyuan Qiao, Zheng Yu, and Qi Wu. VLN-PETL: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 2, 7", + "[45] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-Matterport 3D dataset (HM3D): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1", + "[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 4, 2, 3", + "[47] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1, 4, 2, 3", + "[48] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 7", + "[49] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout, 2019. 2", + "[50] Robert Tarjan. Depth-first search and linear graph algorithms. In 12th Annual Symposium on Switching and Automata Theory (swat 1971), pages 114-121, 1971. 5", + "[51] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation, 2019. 1, 2, 6", + "[52] Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu, and Qijun Chen. Vision-and-language navigation via causal learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[53] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks, 2022. 2", + "[54] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIV 16, pages 413–430. Springer, 2020. 1", + "[55] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 1, 2", + "[56] Erik Wijmans, Samyak Datta, Oleksandr Maksymets, Abhishek Das, Georgia Gkioxari, Stefan Lee, Irfan Essa, Devi Parikh, and Dhruv Batra. Embodied question answering" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in photorealistic environments with point cloud perception, 2019. 2", + "[57] Wikipedia. Preferred walking speed. https://en.wikipedia.org/wiki/Preferred_walking_speed.3", + "[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1", + "[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4", + "[60] Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and Wang He. NaVid: Video-based vlm plans the next step for vision-and-language navigation. arXiv preprint arXiv:2402.15852, 2024. 1", + "[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize Anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 2, 3, 4", + "[62] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation, 2024. 2", + "[63] Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. Towards learning a generalist model for embodied navigation, 2023. 2, 5, 6, 7, 8", + "[64] Gengze Zhou, Yicong Hong, and Qi Wu. NavGPT: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 3, 7", + "[65] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 7", + "[66] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. SOON: Scenario oriented object navigation with graph-based exploration, 2021. 1, 2, 6" + ], + "bbox": [ + 91, + 90, + 483, + 700 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27596", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_model.json b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_model.json new file mode 100644 index 0000000000000000000000000000000000000000..659ed55ea133bc33f6d67c1860918a10b068bd31 --- /dev/null +++ b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.795, + 0.176 + ], + "angle": 0, + "content": "RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.203, + 0.871, + 0.294 + ], + "angle": 0, + "content": "Mingfei Han\\*1,3, Liang Ma\\*1, Kamila Zhumakhanova\\*1, Ekaterina Radionova\\*1, Jingyi Zhang\\*2 \nXiaojun Chang\\*1,4, Xiaodan Liang\\*1,2, Ivan Laptev\\*1 \n\\(^{1}\\)Department of Computer Vision, MBZUAI \\(^{2}\\)Shenzhen Campus of Sun Yat-Sen University \n\\(^{3}\\)ReLER Lab, AAII, UTS \\(^{4}\\)University of Science and Technology of China \nhttps://roomtour3d.github.io" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.342, + 0.327, + 0.359 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.486, + 0.722 + ], + "angle": 0, + "content": "Vision-and-Language Navigation (VLN) suffers from the limited diversity and scale of training data, primarily constrained by the manual curation of existing simulators. To address this, we introduce RoomTour3D, a video-instruction dataset derived from web-based room tour videos that capture real-world indoor spaces and human walking demonstrations. Unlike existing VLN datasets, RoomTour3D leverages the scale and diversity of online videos to generate open-ended human walking trajectories and open-world navigable instructions. To compensate for the lack of navigation data in online videos, we perform 3D reconstruction and obtain 3D trajectories of walking paths augmented with additional information on the room types, object locations and 3D shape of surrounding scenes. Our dataset includes \\(\\sim 100K\\) open-ended description-enriched trajectories with \\(\\sim 200K\\) instructions, and \\(17K\\) action-enriched trajectories from 1847 room tour environments. We demonstrate experimentally that RoomTour3D enables significant improvements across multiple VLN tasks including CVDN, SOON, R2R, and REVERIE. Moreover, RoomTour3D facilitates the development of trainable zero-shot VLN agents, showcasing the potential and challenges of advancing towards open-world navigation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Over the past years, Vision-and-Language Navigation [2, 21, 24, 32, 42, 54] has largely relied on human-designed simulators and annotated trajectories. R2R [2] initiated a benchmark for language-guided navigation in simulated indoor settings, while CVDN [51], REVERIE [42], and SOON [66] expanded VLN to dialogue-based and object-focused tasks. However, these curated simulations lack scene diversity and fail to capture real-world complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.586 + ], + "angle": 0, + "content": "To address limited diversity, recent methods propose the use of richer and more varied training data. AirBERT [15] combines discrete Airbnb images for panoramic views, which lack consistency and naturalistic context of an indoor scene. ScaleVLN [55] utilizes laboriously curated 3D scenes [45, 58], but suffers from reconstruction quality and scalability. More recently, YTB-VLN [34] attempts to use video frames to compose panoramic views and organize instructions with predefined templates, yet overlooks object variety and geometry structure. NaVid [60] constructs sequential single-view trajectories from MatterPort3D [4] and R2R [2] annotations, paired with general video data to train a sim-to-real agent. None of these approaches simultaneously achieves scalability in scene diversity, openness in object variety, or comprehensive geo-perception in spatial representations, each of which is critical to training effective and open-world navigation agents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.821 + ], + "angle": 0, + "content": "To address the challenge, we introduce RoomTour3D, a novel dataset that provides a geometry-aware, spatially enriched training environment for VLN agents. Built upon easily accessible room tour videos from the Internet, RoomTour3D captures continuous movement through real estates with a hand-held camera from a first-person perspective. Each frame presents a realistic, agent-centric view and showcases a rich array of indoor items. The continuous flow of these frames captures multiple views of the environment, presenting diverse room layouts and inherently embedding the geometric properties of the spaces. To unleash the power of these videos, we propose an automatic and extendable pipeline to obtain open-ended geometry-aware human walking trajectories, spatially contextualized textual instructions using open vocabularies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To better model the navigation scenario, we take advantage of the continuous walk-through trajectories and densely sample frames from room tour videos. Then, we use COLMAP [46, 47] to reconstruct 3D scenes of real-estates to obtain the geometric information. With known" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.216, + 0.9 + ], + "angle": 0, + "content": "* Co-first authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27586" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "camera locations and orientations, we sample \"decision-making\" frames at points of maximal yaw rotation and sequence further frames every \\(\\sim 1.5\\) meters to finalize trajectories. Additionally, our pipeline uses extensive annotations via RAM [61] for object tagging, Grounding-DINO [38] for precise localization, and Depth-Anything [59] for estimating relative object-camera distances. To integrate knowledge of object variety, geometric awareness, and human walking preferences into model training, we employ GPT-4 [40] to generate navigation instructions for both summarization and task-specific navigation tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.26, + 0.485, + 0.53 + ], + "angle": 0, + "content": "Our RoomTour3D is an effort to create a comprehensive database derived from room tour videos and enriched by human-living knowledge. The dataset includes \\(\\sim 100\\mathrm{K}\\) open-ended trajectories with \\(\\sim 200\\mathrm{K}\\) descriptions, and \\(\\sim 17\\mathrm{K}\\) geometry-aware trajectories with navigable actions from 1847 homes. Moreover, we are releasing intermediate products such as object tags, bounding boxes, depth maps, room locations, and the necessary code and prompts used to generate instructions. To validate the robustness, we conducted experiments with NaviLLM [63], a generalist model based on Large Language Model (LLM), to train a unified multi-task navigation agent. Integrating our data improved baseline performances on CVDN, SOON, R2R, and REVERIE by over \\(6\\%\\), notably achieving a \\(9.8\\%\\) boost on SOON and setting new state-of-the-art results. Furthermore, our enriched action-instruction data facilitates training an end-to-end zero-shot navigation agent, advancing towards open-world embodied navigation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.532, + 0.473, + 0.547 + ], + "angle": 0, + "content": "In this work, we make the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.548, + 0.483, + 0.64 + ], + "angle": 0, + "content": "- Video Collection for Complex Environments: We curate a novel dataset of diverse videos tailored for navigation tasks, distinguishing it from existing datasets such as YTB-VLN [34]. Our dataset features longer videos, representing more complex environments, and has fewer shot changes, ensuring continuity and contextual consistency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.64, + 0.483, + 0.715 + ], + "angle": 0, + "content": "- Automatic Video Processing: We propose a pipeline to extract geometry-aware navigation instructions, aligning spatial understanding with navigation goals. Additionally, we generate open-vocabulary instructions for open-ended trajectories to enhance real-world applicability." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.715, + 0.483, + 0.775 + ], + "angle": 0, + "content": "- Demonstrating Data Effectiveness: Through extensive experiments and ablation studies, we demonstrate that our dataset significantly improves the performance of state-of-the-art models." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.548, + 0.483, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.233, + 0.807 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.381, + 0.833 + ], + "angle": 0, + "content": "2.1. Vision-and-Language Navigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Navigating unseen indoor environments with natural language instructions is vital for enabling embodied agents to assist humans. Various scenarios have been explored, including fine-grained instruction following (R2R [2]) and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "dialogue-based navigation (CVDN [51]), object localization from instructions (SOON [66], REVERIE [42]), and embodied question answering through active 3D exploration [10, 56]. While substantial work focuses on task-specific models [1, 6, 7, 14, 16-20, 25, 30, 37, 39, 43, 44, 52, 62, 66], they often lack generalization across tasks. In this context, NaviLLM [63] introduces an embodied generalist model that simultaneously addresses multiple tasks through a single framework, demonstrating strong generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.25, + 0.793, + 0.265 + ], + "angle": 0, + "content": "2.2. Data-Centric Methods for VLN" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.905, + 0.405 + ], + "angle": 0, + "content": "The scarcity of VLN training data remains a critical issue and results in poor generalization of VLN agents to unseen environments. Most of existing VLN datasets such as R2R [2], RxR [24], CVDN [51] and SOON [66] are produced in simulators, which constrains data scalability due to the high labor costs involved. To tackle the problem, data augmentation [12, 13, 23, 27, 28, 35, 49] and self-exploration in simulator environments [31, 53] have been investigated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.906, + 0.648 + ], + "angle": 0, + "content": "VLN-BERT [17] and AirBERT [15] attempt to use web-based image-caption pairs for pre-training, however, resulting trajectories often fail to mimic realistic navigation. Similarly, automatic dataset generation pipelines [8, 22], including ScaleVLN [55], rely on manually curated 3D scenes or synthetic environments, which are costly to produce and lack the photorealism needed for robust real-world generalization. PanoGen [26] enhances VLN training by generating diverse text-conditioned panoramic environments using text-to-image diffusion models and recursive outpainting. While this approach addresses the scarcity of training environments, it relies on synthetic panoramas and may not generalize well to real environments. YTB-VLN [34] advances scalability by leveraging YouTube room tour videos to generate path-instruction pairs but omits explicit path geometry, essential for robotic navigation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.906, + 0.814 + ], + "angle": 0, + "content": "In our work, we address the limitations by designing RoomTour3D with properties: (i) free-form and open-vocabulary path annotations instead of template-based instructions, (ii) extraction of open-ended trajectories from sequential video clips, and (iii) inclusion of turning points and spatially close frames as navigable candidate actions, moving beyond panoramic nodes. Furthermore, our approach integrates 3D reconstruction of indoor videos to retrieve trajectory geometry and employs an LLM to generate detailed, object-aware instructions with enhanced spatial understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.714, + 0.837 + ], + "angle": 0, + "content": "2.3. Zero-shot Navigation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Given the substantial semantic variations in complex real-world scenarios, fully-supervised VLN models often struggle to generalize across diverse navigation scenes. Zero-shot VLN has thus gained attention as it eliminates prior" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27587" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.094, + 0.885, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.415, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Figure 1. Overview of our RoomTour3D data generation. Starting from a room tour video, we first apply BLIP-2 [29] on frame sequence to predict the room locations. Next, we use RAM [61] and Grounding-DINO [38] to identify objects within the frames and employ Depth-Anything [59] for depth prediction. Subsequently, COLMAP is used to reconstruct the 3D scene with complete geometry information, and we sample human walking trajectories from the continuous frames. The trajectory captures open-world objects, their positions, and depths relative to the camera. Finally, we use advanced LLM, i.e., GPT-4 to generate the free-form descriptions for pretraining, namely description-enriched trajectories. Specifically, for the trajectory shown in the figure, which involves instant turning points, we specially treat \\(<0>\\) to \\(<6>\\) as walking trajectory, \\(<\\mathrm{A}> <\\mathrm{B}>\\) and \\(<\\mathrm{C}>\\) as side-watching points and use them as negative candidates for navigation finetuning task, namely action-enriched trajectories. For more details, please refer to Section 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.543, + 0.483, + 0.572 + ], + "angle": 0, + "content": "knowledge of environments and instructions, effectively mitigating environmental biases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.572, + 0.483, + 0.753 + ], + "angle": 0, + "content": "Commercial model based methods utilize advanced LLMs and robust frameworks for seamless solutions. MapGPT [5] incorporates a map-based prompting system with global spatial reasoning and adaptive path planning. DiscussNav [39] employs a multi-expert framework where LLMs specialize in subtasks like instruction analysis and vision perception. NavGPT [64] focuses on explicit reasoning by combining commonsense reasoning with visual observations. As to non-commercial methods, LangNav [41] uses language as the primary perceptual space, while NavCoT [33] introduces parameter-efficient training to allow LLMs to autonomously reason and act." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.754, + 0.483, + 0.815 + ], + "angle": 0, + "content": "We show that usage of our action-enriched data for navigation tuning results in superior zero-shot performance over all non-commercial methods and reaches comparable results to commercial approaches based on GPT-3.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.233, + 0.835 + ], + "angle": 0, + "content": "3. RoomTour3D" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we present the automatic data curation pipeline of RoomTour3D. We detail the process of annotations, from sampling open-ended human walking trajectories to generating corresponding descriptions with open-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.618 + ], + "angle": 0, + "content": "world object variety and spatial awareness. Enabled by reconstructed 3D scenes, we further sample navigable trajectories with actions. The overall pipeline of our data generation is illustrated in Figure 1. Please refer to Appendix 7 for details about video collection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.815, + 0.642 + ], + "angle": 0, + "content": "3.1. Description-Enriched Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.839 + ], + "angle": 0, + "content": "In this subsection, we detail the process of generating controllable descriptions for open-ended trajectories. We start by generating human-walking trajectories by uniformly sampling frames at a rate of one frame every two seconds, which aligns with the average human walking speed of 1.42 meters per second [57], typically slower in indoor environments. Subsequently, to annotate these trajectories, as shown in Figure 2, we employ expert models such as BLIP-2 [29], RAM [61], Grounding-DINO [38], and Depth-Anything [59] to gather extensive information on object variety, spatial positions, and depth measurements. Finally, we integrate this information into GPT4 [40] to generate detailed and coherent traje." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Object Variety and Spatial Awareness. In order to harness object variety and enable spatial awareness, we compose three expert models and design a textual template, i.e., \"There is a object tag to the spatial position of current spot" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27588" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.903, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.353, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Instruction generation in a controllable way. (a) Using open-sourced expert models, we identify what objects are in the frames, and assess how far an object is and determine where an object is located. The information is then textualized to create richly detailed frame captions. (b) BLIP-2 is adopted to predict and smooth room location across sequential frames. (c) Combining room locations and object information, we use GPT-4 for controllable and open-vocabulary instruction generation. The prompt consists of a task instruction that defines the generation task, and in-context examples that constrain the output style." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.482, + 0.544 + ], + "angle": 0, + "content": "in relative distance\", to organize the multi-source information to ease GPT generation. Firstly, we used RAM [61] to annotate the object categories within the frames. Based on these category tags, we employed Grounding DINO [38] to locate the objects in the frames. Subsequently, we used Depth-Anything [59] to predict the depth maps corresponding to the frames." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.548, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Using this data, we identify the spatial locations and distances of objects relative to the current camera position. By analyzing object bounding box center positions and depth map locations, we can generate frame captions, as illustrated in Figure 2. Finally, objects in different frames can be easily correlated and capture the progression across different frames. More details in the spatial awareness data generation are provided in Appendix 9." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Room Location Annotation in Videos. To determine the camera of each frame, w.r.t. the room category, we used BLIP2 [29] in visual question-answering mode, posing the question, \"Which room am I in?\" A predefined list of 16 common room types (i.e., bathroom, kitchen, bedroom, living room, hallway, office, dining room, foyer, laundry room, outside, porch, garage, front, patio, driveway, and backyard) was used as possible answers. This list was curated by analyzing 10 randomly sampled long videos, using BLIP2 in generative mode to identify and rank the most frequently mentioned room types. For frame-level predictions, we switched to BLIP2's discriminative mode and applied temporal smoothing to denoise outputs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We validated this approach by manually annotating 50 video clips, achieving an accuracy of \\(85\\%\\). The use of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.514 + ], + "angle": 0, + "content": "BLIP2 leverages its open-world knowledge, while limiting room types to 16 categories for discriminative selection simplifies outputs and reduces ambiguity. Any loss in open-vocabulary flexibility is addressed during GPT-based trajectory summarization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.742 + ], + "angle": 0, + "content": "Controllable Instruction Generation. To generate descriptions that accurately capture human-walking trajectories and reflect the environment, we integrate frame-level room locations with frame captions composed of object descriptions. We then employ GPT-4-Turbo [40] for controllable instruction generation, leveraging the multi-source information contained in the composed captions. As depicted in Figure 2, we organize the prompt using the \"Task instruction - In-context examples - Prediction\" scheme. This approach defines our instruction generation task as describing object progression along the moving trajectories, and includes two examples to ensure GPT produces instruction-style texts only. As shown in Figure 2 (c), captions of frames along the trajectory are embedded into the prompt and input into GPT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.776, + 0.763 + ], + "angle": 0, + "content": "3.2. Action-Enriched Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D Environment Reconstruction. To obtain the geometric information of trajectories within RoomTour3D, we employ COLMAP [46, 47] for 3D reconstruction. This process allows us to infer the 3D layout of environments in the videos, providing a detailed geometric context for navigation tasks. Specifically, we sample the videos at 3 frames per second to balance accuracy and execution time. To further improve time efficiency, we split the videos into 100-second video clips with 10-second overlaps between adjacent clips and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27589" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.093, + 0.418, + 0.105 + ], + "angle": 0, + "content": "(a) Pretraining with Summarization Task" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.11, + 0.502, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.208, + 0.156, + 0.22 + ], + "angle": 0, + "content": "Output:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.222, + 0.25, + 0.302 + ], + "angle": 0, + "content": "Move through the corridor and step forwards.Pass by a paint on the wall. Then advance through the bedroom and into the bathroom.Start with a view of the bed,curtain.." + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.226, + 0.326, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.234, + 0.435, + 0.246 + ], + "angle": 0, + "content": "Embodied Agent" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.276, + 0.346, + 0.288 + ], + "angle": 0, + "content": "Output:" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.292, + 0.356, + 0.303 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.284, + 0.504, + 0.293 + ], + "angle": 0, + "content": "Summarization" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.298, + 0.482, + 0.307 + ], + "angle": 0, + "content": "Navigation" + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.093, + 0.81, + 0.105 + ], + "angle": 0, + "content": "(b) Finetuning with Navigation Task" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.136, + 0.557, + 0.147 + ], + "angle": 0, + "content": ":" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.113, + 0.631, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.113, + 0.714, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.113, + 0.798, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.833, + 0.114, + 0.894, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.16, + 0.896, + 0.195 + ], + "angle": 0, + "content": "Move through the corridor and step forwards... advance through the bedroom ... Start with a view of the bed, curtains... Move closer, noting the headboard... Pass the dresser and ..." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.23, + 0.555, + 0.241 + ], + "angle": 0, + "content": "Task:" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.207, + 0.875, + 0.242 + ], + "angle": 0, + "content": "Compare the History and Instruction to infer your current progress, and then select the correct direction from the candidates to go to the target location." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.271, + 0.56, + 0.282 + ], + "angle": 0, + "content": "< Cand>:" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.251, + 0.632, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.25, + 0.718, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.249, + 0.805, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.249, + 0.889, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 3. Model training diagram with RoomTour3D. We design two tasks for our RoomTour3D to boost NaviLLM. (a) Pretraining: Sampled frames on the trajectory are treated as candidate observations. Model is optimized to summarize object progression along the path. (b) Finetuning: Each frame is considered as a navigable step. Given historical observation \\(<0>\\) to \\(<2>\\) and navigation instruction, the model is prompted to predict the next action by selecting from candidate observations View A, View B, View C and View D." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.394, + 0.431, + 0.409 + ], + "angle": 0, + "content": "perform reconstruction on the clips simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Following this, we merge the resulting sub-models reconstructed from the video clips. For every two adjacent clips, if the reconstructed models have more than three overlapping frames, the models are merged and readjusted into one. However, due to varying reconstruction quality, a single video clip can produce more than one model. To manage this, we construct a graph for merging these sub-models. Each model is treated as a node, and we connect two models if they have more than three overlapping frames. We then apply Depth-First Search [50] to merge any two connected models and replace the original model nodes with the new merged one, continuing this process until no connected nodes remain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Navigable Action Sampling. We enhance navigation action diversity by using frames from room tour videos as navigable actions by sampling at significant view-change points within a small radius. These points are identified by reconstructing 3D scenes to measure camera orientation differences and distances between frames, capturing varied views from revisited locations or turning points. Frames with substantial view changes are retained using cosine similarity thresholding, followed by non-maximum suppression to isolate major shifts. DBSCAN clustering [11] groups spatially close frames with different views to ensure diverse navigable actions. These measures ensure the robustness to the misidentification of significant view change points that may arise due to inaccuracies in 3D reconstruction. Finally, we identify distinct walking paths within each cluster. For each path, we select the most recent frame as a positive candidate and the frame with the highest angular difference as a negative candidate, enhancing the diversity of navigable actions. More detailed description of our navigable points" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.394, + 0.829, + 0.41 + ], + "angle": 0, + "content": "sampling approach are provided in Appendix 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.873, + 0.44 + ], + "angle": 0, + "content": "4. Vision-and-Language Navigation Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.569 + ], + "angle": 0, + "content": "In this section, we introduce a practice to use our data to train a generalist embodied agent. To start with, we first provide a concise introduction to the state-of-the-art VLN model, NaviLLM [63], which is an LLM-based navigation agent. Then, we introduce two tasks that are adapted for our RoomTour3D data, i.e., vision-instruction summarization task for pretraining and action-instruction navigation task for finetuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.708, + 0.594 + ], + "angle": 0, + "content": "4.1. Revisiting NaviLLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.704 + ], + "angle": 0, + "content": "NaviLLM is a SOTA LLM-based model for embodied navigation, excelling on benchmarks like CVDN and SOON. It processes panoramic inputs by encoding environmental views and integrating them with navigation instructions. Specific tokens are defined for different types of inputs: for historical observations and for candidate views at each navigational step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "During training, NaviLLM receives instructions and a sequence of candidate views. At each step, the candidate observations are input into the model along with the instructions. The model predicts the next action by selecting the appropriate view from the candidates, and this selected view is then cached as a token, updating the model's internal state for future decisions. At the last step of the navigation task, the model summarizes all tokens as a separate training task to ensure comprehensive understanding and retention of the navigational history. For testing, the model similarly uses historical observations, i.e., tokens, accumulated during navigation, and candidate views, i.e., , at each step to decide the next" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27590" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.123 + ], + "angle": 0, + "content": "action. This process ensures that the model's actions align with the given instructions and the observed environment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.408, + 0.148 + ], + "angle": 0, + "content": "4.2. Summarization Task for Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.381 + ], + "angle": 0, + "content": "To leverage the rich, sequential nature of videos and enhance future planning capabilities, we adapt NaviLLM to use the description-enriched trajectories from our RoomTour3D dataset for a summarization task. Each frame is treated as a candidate view and wrapped with \\( \\langle \\text{cand} \\rangle \\) tokens. Similar to selected panoramic views at each navigation step, these frames are considered as selected actions when executing navigation instructions. As shown in Figure 3(a), the frame tokens and summarization task instruction are organized into a unified prompt and input into the LLM. The model is expected to output a trajectory summary containing object progression and room locations, as specified by the task instruction. The model is trained using next-token prediction loss, consistent with the original language model training methodology." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.388, + 0.367, + 0.405 + ], + "angle": 0, + "content": "4.3. Navigation Task for Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.636 + ], + "angle": 0, + "content": "In order to enable learning navigation decision-making from scalable scenes, we adapt NaviLLM to action-enriched trajectories from our RoomTour3D. Unlike panoramic views capturing observations from a single location, our data provides candidate views from frames at different locations and orientations, with only one frame directed toward the destination. Each frame in the video sequence is treated as a potential navigable action and wrapped with \\( \\langle \\text{cand} \\rangle \\) tokens. These candidate views are presented to the model and processed in the same way as panoramic views. As shown in Figure 3(b), the model processes the inputs to predict the next action, selecting the appropriate frame from the candidate views. The selected action is then cached as a \\( \\langle \\text{hist} \\rangle \\) token for subsequent decision-making steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.773 + ], + "angle": 0, + "content": "During fine-tuning, each frame is treated as a navigable step, with the next trajectory frame as the target action and \\( \\langle \\text{STOP} \\rangle \\) as an alternative. The model uses historical observations and navigation instructions to iteratively predict the next action, building a detailed understanding of the path. At the final step, the model summarizes the navigation path, incorporating object progression and room locations. This summarization task enhances its ability to recall navigational history and improves performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.312, + 0.803 + ], + "angle": 0, + "content": "5. Tasks and Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.856 + ], + "angle": 0, + "content": "This section outlines our experimental setup and presents the results. Detailed implementation information can be found in Appendix 11." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. During pretraining, we follow practice from NaviLLM [63] and perform teacher-forcing training on the combined dataset from our video-instruction data from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "RoomTour3D, together with CVDN [51], SOON [66], R2R [2], REVERIE [42] and ScanQA [3], and augmented data from R2R and REVERIE. In the multi-task fine-tuning stage, we alternate between teacher forcing and student forcing on the combined data from our action-instruction data from RoomTour3D, together with CVDN, SOON, R2R, REVERIE, ScanQA and LLaVA-23k [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.365 + ], + "angle": 0, + "content": "To evaluate the impact of our data on navigation agent training, we test on CVDN, SOON, R2R, and REVERIE. CVDN requires navigating towards a target by understanding dialog history, linking dialogue comprehension to actions. SOON tasks the agent with locating objects without bounding boxes, emphasizing semantic-visual alignment. R2R involves following step-by-step instructions, requiring dynamic progress tracking and precise alignment with navigational history. REVERIE focuses on localizing distant objects based on concise instructions, aided by ground truth bounding boxes at waypoints." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.905, + 0.516 + ], + "angle": 0, + "content": "Evaluation Metrics. For the navigation tasks, we follow the evaluation methodology from [2] using the following navigation metrics: Success Rate (SR), which measures whether the agent reaches the target location within a set distance threshold; Success Rate Weighted by Path Length (SPL), which is the SR adjusted by the ratio of the ground truth path length to the actual path traveled; Goal Progress (GP), the advancement in meters towards the goal. GP is utilized for the CVDN dataset, whereas SR and SPL are the metrics for other datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.806, + 0.546 + ], + "angle": 0, + "content": "5.1. Comparison on Supervised Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.905, + 0.657 + ], + "angle": 0, + "content": "As shown in Table 1, we performed a one-time fine-tuning on the four tasks in a fully supervised manner. To begin, our experiments reiterate the superiority of multitask training over single-task training. Also, incorporating our RoomTour3D data into the pre-training process led to consistent improvements across all metrics on Val-U, achieving state-of-the-art results in the GP metric in the CVDN dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Notably, finetuning with our action-enriched data results in state-of-the-art performance on both Val-U and Test sets across SOON, R2R and REVERIE tasks. While the improvement on the CVDN and SOON datasets is modest, the most significant boost compared to the reproduced baseline is observed in R2R Val-U and REVERIE Val-U, with gains of approximately \\(5.7\\%\\) and \\(6\\%\\), respectively. The improvement in R2R is largely driven by enhanced spatial awareness, stemming from the inclusion of proximity data, which helps the model better understand object distance and position. Similarly, gains in REVERIE are attributed to a combination of open-vocabulary tags, spatial awareness, and the addition of room type data, which encourages the model to infer the layout of environments, thereby boosting its spatial reasoning capabilities. Moreover, our use of open-ended instructions allows the model to adapt flexibly to diverse" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27591" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Overall comparison with the baseline methods. Our RoomTour3D data can boost NaviLLM by a margin on SOON, R2R and REVERIE on SPL metric and on CVDN GP metric. *denotes reproduced results. RT3DDesc and RT3DAction stand for description-enriched trajectories only and action-enriched trajectories." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.136, + 0.864, + 0.392 + ], + "angle": 0, + "content": "
MethodsCVDNSOONR2RREVERIE
Val-UTestVal-UTestVal-UTestVal-UTest
Models Focusing on Single Task
PREVALENT [16]3.152.44--5351--
HOP [43]4.413.24--575926.124.3
HAMT [6]5.135.58--616030.226.7
DUET [7]--22.621.4605833.736.0
VLN-SIG [25]5.525.83--6260--
VLN-PETL [44]5.696.13--605827.726.7
NavGPT2 [65]----6160--
BEV-BERT [1]----646036.436.4
Unified Model For All Tasks
NaviLLM(w. Pretrain) [63]6.167.9029.226.3596035.732.3
NaviLLM(w. Pretrain)*6.09-28.0-56.7-31.4-
NaviLLM+RT3DDesc(Ours)6.967.5530.226.562.361.837.135.1
NaviLLM+RT3DAction(Ours)6.337.2231.727.862.462.237.436.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.401, + 0.74, + 0.415 + ], + "angle": 0, + "content": "Table 2. Ablation study on the input modalities for trajectory summarization task." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.419, + 0.868, + 0.519 + ], + "angle": 0, + "content": "
Object tagsDepth & Bounding BoxRoom typeCVDN GP↑SOONR2RREVERIE
SR↑SPL↑SR↑SPL↑SR↑SPL↑
×××6.0933.6428.0165.5256.6738.3231.35
✓××5.4132.5226.5163.6155.7642.5234.37
✓✓×6.4937.6230.4068.3761.7041.7236.04
✓✓✓6.9638.8030.2169.3762.2843.2537.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Table 3. Overall comparison with SOTA zero-shot methods on R2R. \\( {}^{ \\dagger } \\) denotes training exclusive navigable actions. \\( {}^{ * } \\) denotes using 36 views setting. RT3D stands for our RoomTour3D." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.584, + 0.445, + 0.833 + ], + "angle": 0, + "content": "
MethodsVal Unseen
SR↑SPL↑
Random Walk [41]32
Commercial Model
NavGPT(GPT-3.5) [64]*13.899.12
NavGPT(GPT-4) [64]3429
MapGPT(GPT-4) [5]38.825.8
MapGPT(GPT-4V) [5]43.734.8
DiscussNav(GPT-4) [39]4340
Open-source Model
LangNav(LLaMA2-7B) [41]00
NavCoT(LLaMA2-7B) [33]7.786.50
DuET (Init. LXMERT [48])10
NaviLLM [63]†00
NaviLLM+RT3D(Ours)14.3310.86
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "scenarios, fostering more robust and generalizable performance and better contextual understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.788, + 0.55 + ], + "angle": 0, + "content": "5.2. Comparison on Zero-shot Task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.906, + 0.618 + ], + "angle": 0, + "content": "To further demonstrate the substantial indoor knowledge contained in our data and its effectiveness for embodied action and language instructions, we conduct zero-shot experiments on embodied action prediction, as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.832 + ], + "angle": 0, + "content": "We removed all action and geometric data from the training datasets and retrained NaviLLM with and without our RoomTour3D dataset. Without action prediction data, NaviLLM lacked the ability to learn effective navigable action selection. However, with the inclusion of our action-enriched trajectories, NaviLLM achieved a \\(14.33\\%\\) SR and a \\(10.86\\%\\) SPL, outperforming open-source models built on LLaMA-7B and reaching results comparable to NavGPT [64], which leverages GPT-3.5. These improvements validate the effectiveness of our 3D trajectories mined from room tour video reconstructions and emphasize the value of our action-enriched trajectories. This highlights the significant contribution of our dataset to advancing open-world navigation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.665, + 0.863 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Effect of open-world semantics and spatial awareness. As shown in Table 2, we analyzed the impact of various in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.132, + 0.092, + 0.867, + 0.114 + ], + "angle": 0, + "content": "Instruction: Exit the sewing room. Turn right. Go toward the glass cabinet with the dolls in it. Turn into the doorway on the left. Pass the bed and go through the next doorway on the left into the bathroom. Wait by the sink. (Instruction_id : 4676_0)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.115, + 0.55, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.116, + 0.865, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 4. Paths of NaviLLM [63] and ours on R2R-unseen. Purple and green circles denote the start and target locations, respectively, and the red circle represents incorrect endpoint. According to the instruction, the agent should turn left at the waypoint marked with yellow. Our method makes the correct decision, while the baseline is confused by similar entrance at the waypoint, thus mistakenly turns right." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.389, + 0.486, + 0.629 + ], + "angle": 0, + "content": "formation types on instruction generation. Adding object variety significantly improved performance on REVERIE with increase SPL from \\(31.35\\%\\) to \\(34.37\\%\\), as this dataset relies on object grounding. However, it had no direct impact on SOON, possibly because SOON relies solely on detailed textual descriptions without explicit bounding box annotations. After introducing depth estimation, which helps determine the relative distances of objects, the performance on SOON, R2R and REVERIE achieve marginal boosts. This demonstrates that enhancing spatial awareness significantly contributes to indoor navigation tasks. Furthermore, incorporating room locations, which capture the scene semantics along the trajectory, provided a moderate boost across all four VLN tasks. This further highlights the critical role of object variety and spatial awareness in improving navigation performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.485, + 0.766 + ], + "angle": 0, + "content": "Effect of action-instruction data. As shown in Table 1, we test the effect of incorporating video-action-instruction data into the training dataset. It is evident that this approach improves the SPL metric across the test scenes of SOON, R2R, and REVERIE. We believe that incorporating geometric information and movement-inclusive instructions helps the model better align the relationship between action and observation changes, thereby further enhancing the model's embodied capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.774, + 0.347, + 0.789 + ], + "angle": 0, + "content": "5.4. Data correctness verification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluated the correctness of our automated data-generation pipeline by manually rating 100 randomly sampled trajectory descriptions on a 4-point relevance scale: 1 for \"totally irrelevant\", 2 for \"partially relevant\", 3 for \"mostly relevant\" and 4 for a \"perfect match\". The evaluation yielded an average score of 3.08, with \\(74\\%\\) of descriptions rated as \"mostly relevant\" or \"perfect match,\" demon-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.389, + 0.907, + 0.419 + ], + "angle": 0, + "content": "strating the method's effectiveness in generating meaningful, visually aligned descriptions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.431, + 0.78, + 0.447 + ], + "angle": 0, + "content": "5.5. Navigation Case Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.909, + 0.664 + ], + "angle": 0, + "content": "As shown in Figure 4, selecting the correct action is critical at specific decision points, such as when a left turn is required to follow the instruction accurately. In the example, at step \\(②\\), both the rooms to the left and right could satisfy the latter part of the instruction, \"pass the bed and go into the bathroom.\" However, the baseline method incorrectly chooses a right turn at the designated left-turn point, causing it to deviate from the intended path. Once this error occurs, even with scene graph history, the model struggles to realign with the correct trajectory. This challenge is particularly common in household environments, where bedroom layouts often appear similar. It further demonstrates the effectiveness of our data alignment in improving adherence to action-based instructions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.634, + 0.697 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.909, + 0.889 + ], + "angle": 0, + "content": "In this paper, we present RoomTour3D, a novel dataset automatically curated from room tour videos for VLN tasks. By leveraging the rich, sequential nature of video data and incorporating object variety and spatial awareness, we generate 200K navigation instructions and 17K action-enriched trajectories from 1847 room tour scenes. Additionally, we produce navigable trajectories from video frames and reconstructed 3D scenes, which significantly boost the performance and set new state-of-the-art results on the SOON and REVERIE benchmarks. This approach also enables the development of a trainable zero-shot navigation agent, demonstrating the effectiveness and scalability of RoomTour3D in advancing VLN research." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27593" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. BEVbert: Multimodal map pre-training for language-guided navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2737-2748, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[5] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee K Wong. Mapgpt: Map-guided prompting for unified vision-and-language navigation. arXiv preprint arXiv:2401.07314, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.611 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[10] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, page 226-231. AAAI Press, 1996. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Tsu-Jui Fu, Xin Eric Wang, Matthew Peterson, Scott Grafton, Miguel Eckstein, and William Yang Wang. Coun" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "terfactual vision-and-language navigation via adversarial path sampling, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14911-14920, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[16] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13137-13146, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language BERT for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3d world into large language models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[19] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6683-6693, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[20] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[21] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. arXiv preprint arXiv:1905.12255, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[22] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[23] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[24] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27594" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "navigation. Advances in Neural Information Processing Systems, 36:21878-21894, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[27] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.166, + 0.482, + 0.194 + ], + "angle": 0, + "content": "[28] Jialu Li, Hao Tan, and Mohit Bansal. EnvEdit: Environment editing for vision-and-language navigation, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.196, + 0.482, + 0.249 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.253, + 0.482, + 0.295 + ], + "angle": 0, + "content": "[30] Xiangyang Li, Zihan Wang, Jiahao Yang, Yaowei Wang, and Shuqiang Jiang. KERM: Knowledge enhanced reasoning for vision-and-language navigation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.297, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[31] Xiwen Liang, Fengda Zhu, Lingling Li, Hang Xu, and Xiaodan Liang. Visual-language navigation pretraining via prompt-based environmental self-exploration, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.341, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[32] Xiwen Liang, Liang Ma, Shanshan Guo, Jianhua Han, Hang Xu, Shikui Ma, and Xiaodan Liang. CorNav: Autonomous agent with self-corrected planning for zero-shot vision-and-language navigation, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[33] Bingqian Lin, Yunshuang Nie, Ziming Wei, Jiaqi Chen, Shikui Ma, Jianhua Han, Hang Xu, Xiaojun Chang, and Xiaodan Liang. NavCoT: Boosting llm-based vision-and-language navigation via learning disentangled reasoning. arXiv preprint arXiv:2403.07376, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[34] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H. Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.554 + ], + "angle": 0, + "content": "[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[37] Rui Liu, Wenguan Wang, and Yi Yang. Volumetric environment representation for vision-language navigation. In CVPR, pages 16317-16328, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding Dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[39] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. arXiv preprint arXiv:2309.11382, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[40] OpenAI. GPT-4: Generative pre-trained transformer 4. OpenAI API, 2022. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[41] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. arXiv preprint arXiv:2310.07889, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: Remote embodied visual referring expression in real indoor environments, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[43] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: History-and-order aware pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15418-15427, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[44] Yanyuan Qiao, Zheng Yu, and Qi Wu. VLN-PETL: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[45] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-Matterport 3D dataset (HM3D): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 4, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[47] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1, 4, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[48] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[49] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[50] Robert Tarjan. Depth-first search and linear graph algorithms. In 12th Annual Symposium on Switching and Automata Theory (swat 1971), pages 114-121, 1971. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[51] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[52] Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu, and Qijun Chen. Vision-and-language navigation via causal learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[53] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[54] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIV 16, pages 413–430. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[55] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Erik Wijmans, Samyak Datta, Oleksandr Maksymets, Abhishek Das, Georgia Gkioxari, Stefan Lee, Irfan Essa, Devi Parikh, and Dhruv Batra. Embodied question answering" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27595" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "in photorealistic environments with point cloud perception, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[57] Wikipedia. Preferred walking speed. https://en.wikipedia.org/wiki/Preferred_walking_speed.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[60] Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and Wang He. NaVid: Video-based vlm plans the next step for vision-and-language navigation. arXiv preprint arXiv:2402.15852, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize Anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[62] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[63] Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. Towards learning a generalist model for embodied navigation, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[64] Gengze Zhou, Yicong Hong, and Qi Wu. NavGPT: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[65] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[66] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. SOON: Scenario oriented object navigation with graph-based exploration, 2021. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27596" + } + ] +] \ No newline at end of file diff --git a/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_origin.pdf b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3af7d40e8fe9ed7c500028d7407230c47d1ec4f2 --- /dev/null +++ b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/0dcfb182-7bce-42ad-a99b-7e9e0d669261_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfb16df592a5396c586cc456d9f47439e2bcd3f1c3101b4524fcd446c5b6b2f3 +size 2925029 diff --git a/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/full.md b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..43fa5b6379cb90ac5a39da00bcd981e88db0dd27 --- /dev/null +++ b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/full.md @@ -0,0 +1,310 @@ +# RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation + +Mingfei Han\*1,3, Liang Ma\*1, Kamila Zhumakhanova\*1, Ekaterina Radionova\*1, Jingyi Zhang\*2 +Xiaojun Chang\*1,4, Xiaodan Liang\*1,2, Ivan Laptev\*1 + $^{1}$ Department of Computer Vision, MBZUAI $^{2}$ Shenzhen Campus of Sun Yat-Sen University + $^{3}$ ReLER Lab, AAII, UTS $^{4}$ University of Science and Technology of China +https://roomtour3d.github.io + +# Abstract + +Vision-and-Language Navigation (VLN) suffers from the limited diversity and scale of training data, primarily constrained by the manual curation of existing simulators. To address this, we introduce RoomTour3D, a video-instruction dataset derived from web-based room tour videos that capture real-world indoor spaces and human walking demonstrations. Unlike existing VLN datasets, RoomTour3D leverages the scale and diversity of online videos to generate open-ended human walking trajectories and open-world navigable instructions. To compensate for the lack of navigation data in online videos, we perform 3D reconstruction and obtain 3D trajectories of walking paths augmented with additional information on the room types, object locations and 3D shape of surrounding scenes. Our dataset includes $\sim 100K$ open-ended description-enriched trajectories with $\sim 200K$ instructions, and $17K$ action-enriched trajectories from 1847 room tour environments. We demonstrate experimentally that RoomTour3D enables significant improvements across multiple VLN tasks including CVDN, SOON, R2R, and REVERIE. Moreover, RoomTour3D facilitates the development of trainable zero-shot VLN agents, showcasing the potential and challenges of advancing towards open-world navigation. + +# 1. Introduction + +Over the past years, Vision-and-Language Navigation [2, 21, 24, 32, 42, 54] has largely relied on human-designed simulators and annotated trajectories. R2R [2] initiated a benchmark for language-guided navigation in simulated indoor settings, while CVDN [51], REVERIE [42], and SOON [66] expanded VLN to dialogue-based and object-focused tasks. However, these curated simulations lack scene diversity and fail to capture real-world complexity. + +To address limited diversity, recent methods propose the use of richer and more varied training data. AirBERT [15] combines discrete Airbnb images for panoramic views, which lack consistency and naturalistic context of an indoor scene. ScaleVLN [55] utilizes laboriously curated 3D scenes [45, 58], but suffers from reconstruction quality and scalability. More recently, YTB-VLN [34] attempts to use video frames to compose panoramic views and organize instructions with predefined templates, yet overlooks object variety and geometry structure. NaVid [60] constructs sequential single-view trajectories from MatterPort3D [4] and R2R [2] annotations, paired with general video data to train a sim-to-real agent. None of these approaches simultaneously achieves scalability in scene diversity, openness in object variety, or comprehensive geo-perception in spatial representations, each of which is critical to training effective and open-world navigation agents. + +To address the challenge, we introduce RoomTour3D, a novel dataset that provides a geometry-aware, spatially enriched training environment for VLN agents. Built upon easily accessible room tour videos from the Internet, RoomTour3D captures continuous movement through real estates with a hand-held camera from a first-person perspective. Each frame presents a realistic, agent-centric view and showcases a rich array of indoor items. The continuous flow of these frames captures multiple views of the environment, presenting diverse room layouts and inherently embedding the geometric properties of the spaces. To unleash the power of these videos, we propose an automatic and extendable pipeline to obtain open-ended geometry-aware human walking trajectories, spatially contextualized textual instructions using open vocabularies. + +To better model the navigation scenario, we take advantage of the continuous walk-through trajectories and densely sample frames from room tour videos. Then, we use COLMAP [46, 47] to reconstruct 3D scenes of real-estates to obtain the geometric information. With known + +camera locations and orientations, we sample "decision-making" frames at points of maximal yaw rotation and sequence further frames every $\sim 1.5$ meters to finalize trajectories. Additionally, our pipeline uses extensive annotations via RAM [61] for object tagging, Grounding-DINO [38] for precise localization, and Depth-Anything [59] for estimating relative object-camera distances. To integrate knowledge of object variety, geometric awareness, and human walking preferences into model training, we employ GPT-4 [40] to generate navigation instructions for both summarization and task-specific navigation tasks. + +Our RoomTour3D is an effort to create a comprehensive database derived from room tour videos and enriched by human-living knowledge. The dataset includes $\sim 100\mathrm{K}$ open-ended trajectories with $\sim 200\mathrm{K}$ descriptions, and $\sim 17\mathrm{K}$ geometry-aware trajectories with navigable actions from 1847 homes. Moreover, we are releasing intermediate products such as object tags, bounding boxes, depth maps, room locations, and the necessary code and prompts used to generate instructions. To validate the robustness, we conducted experiments with NaviLLM [63], a generalist model based on Large Language Model (LLM), to train a unified multi-task navigation agent. Integrating our data improved baseline performances on CVDN, SOON, R2R, and REVERIE by over $6\%$ , notably achieving a $9.8\%$ boost on SOON and setting new state-of-the-art results. Furthermore, our enriched action-instruction data facilitates training an end-to-end zero-shot navigation agent, advancing towards open-world embodied navigation. + +In this work, we make the following key contributions: + +- Video Collection for Complex Environments: We curate a novel dataset of diverse videos tailored for navigation tasks, distinguishing it from existing datasets such as YTB-VLN [34]. Our dataset features longer videos, representing more complex environments, and has fewer shot changes, ensuring continuity and contextual consistency. +- Automatic Video Processing: We propose a pipeline to extract geometry-aware navigation instructions, aligning spatial understanding with navigation goals. Additionally, we generate open-vocabulary instructions for open-ended trajectories to enhance real-world applicability. +- Demonstrating Data Effectiveness: Through extensive experiments and ablation studies, we demonstrate that our dataset significantly improves the performance of state-of-the-art models. + +# 2. Related Work + +# 2.1. Vision-and-Language Navigation + +Navigating unseen indoor environments with natural language instructions is vital for enabling embodied agents to assist humans. Various scenarios have been explored, including fine-grained instruction following (R2R [2]) and + +dialogue-based navigation (CVDN [51]), object localization from instructions (SOON [66], REVERIE [42]), and embodied question answering through active 3D exploration [10, 56]. While substantial work focuses on task-specific models [1, 6, 7, 14, 16-20, 25, 30, 37, 39, 43, 44, 52, 62, 66], they often lack generalization across tasks. In this context, NaviLLM [63] introduces an embodied generalist model that simultaneously addresses multiple tasks through a single framework, demonstrating strong generalization ability. + +# 2.2. Data-Centric Methods for VLN + +The scarcity of VLN training data remains a critical issue and results in poor generalization of VLN agents to unseen environments. Most of existing VLN datasets such as R2R [2], RxR [24], CVDN [51] and SOON [66] are produced in simulators, which constrains data scalability due to the high labor costs involved. To tackle the problem, data augmentation [12, 13, 23, 27, 28, 35, 49] and self-exploration in simulator environments [31, 53] have been investigated. + +VLN-BERT [17] and AirBERT [15] attempt to use web-based image-caption pairs for pre-training, however, resulting trajectories often fail to mimic realistic navigation. Similarly, automatic dataset generation pipelines [8, 22], including ScaleVLN [55], rely on manually curated 3D scenes or synthetic environments, which are costly to produce and lack the photorealism needed for robust real-world generalization. PanoGen [26] enhances VLN training by generating diverse text-conditioned panoramic environments using text-to-image diffusion models and recursive outpainting. While this approach addresses the scarcity of training environments, it relies on synthetic panoramas and may not generalize well to real environments. YTB-VLN [34] advances scalability by leveraging YouTube room tour videos to generate path-instruction pairs but omits explicit path geometry, essential for robotic navigation. + +In our work, we address the limitations by designing RoomTour3D with properties: (i) free-form and open-vocabulary path annotations instead of template-based instructions, (ii) extraction of open-ended trajectories from sequential video clips, and (iii) inclusion of turning points and spatially close frames as navigable candidate actions, moving beyond panoramic nodes. Furthermore, our approach integrates 3D reconstruction of indoor videos to retrieve trajectory geometry and employs an LLM to generate detailed, object-aware instructions with enhanced spatial understanding. + +# 2.3. Zero-shot Navigation + +Given the substantial semantic variations in complex real-world scenarios, fully-supervised VLN models often struggle to generalize across diverse navigation scenes. Zero-shot VLN has thus gained attention as it eliminates prior + +![](images/8af1a3f33c14c429498d63dcb7b01f77acca153579fa8dc4ceb4ca5320fa13e2.jpg) +Figure 1. Overview of our RoomTour3D data generation. Starting from a room tour video, we first apply BLIP-2 [29] on frame sequence to predict the room locations. Next, we use RAM [61] and Grounding-DINO [38] to identify objects within the frames and employ Depth-Anything [59] for depth prediction. Subsequently, COLMAP is used to reconstruct the 3D scene with complete geometry information, and we sample human walking trajectories from the continuous frames. The trajectory captures open-world objects, their positions, and depths relative to the camera. Finally, we use advanced LLM, i.e., GPT-4 to generate the free-form descriptions for pretraining, namely description-enriched trajectories. Specifically, for the trajectory shown in the figure, which involves instant turning points, we specially treat $<0>$ to $<6>$ as walking trajectory, $<\mathrm{A}> <\mathrm{B}>$ and $<\mathrm{C}>$ as side-watching points and use them as negative candidates for navigation finetuning task, namely action-enriched trajectories. For more details, please refer to Section 3. + +knowledge of environments and instructions, effectively mitigating environmental biases. + +Commercial model based methods utilize advanced LLMs and robust frameworks for seamless solutions. MapGPT [5] incorporates a map-based prompting system with global spatial reasoning and adaptive path planning. DiscussNav [39] employs a multi-expert framework where LLMs specialize in subtasks like instruction analysis and vision perception. NavGPT [64] focuses on explicit reasoning by combining commonsense reasoning with visual observations. As to non-commercial methods, LangNav [41] uses language as the primary perceptual space, while NavCoT [33] introduces parameter-efficient training to allow LLMs to autonomously reason and act. + +We show that usage of our action-enriched data for navigation tuning results in superior zero-shot performance over all non-commercial methods and reaches comparable results to commercial approaches based on GPT-3.5. + +# 3. RoomTour3D + +In this section, we present the automatic data curation pipeline of RoomTour3D. We detail the process of annotations, from sampling open-ended human walking trajectories to generating corresponding descriptions with open- + +world object variety and spatial awareness. Enabled by reconstructed 3D scenes, we further sample navigable trajectories with actions. The overall pipeline of our data generation is illustrated in Figure 1. Please refer to Appendix 7 for details about video collection. + +# 3.1. Description-Enriched Trajectories + +In this subsection, we detail the process of generating controllable descriptions for open-ended trajectories. We start by generating human-walking trajectories by uniformly sampling frames at a rate of one frame every two seconds, which aligns with the average human walking speed of 1.42 meters per second [57], typically slower in indoor environments. Subsequently, to annotate these trajectories, as shown in Figure 2, we employ expert models such as BLIP-2 [29], RAM [61], Grounding-DINO [38], and Depth-Anything [59] to gather extensive information on object variety, spatial positions, and depth measurements. Finally, we integrate this information into GPT4 [40] to generate detailed and coherent traje. + +Object Variety and Spatial Awareness. In order to harness object variety and enable spatial awareness, we compose three expert models and design a textual template, i.e., "There is a object tag to the spatial position of current spot + +![](images/911aaf5c0c530d215601da82152c47e13079298286803f71626c19a04793eb77.jpg) +Figure 2. Instruction generation in a controllable way. (a) Using open-sourced expert models, we identify what objects are in the frames, and assess how far an object is and determine where an object is located. The information is then textualized to create richly detailed frame captions. (b) BLIP-2 is adopted to predict and smooth room location across sequential frames. (c) Combining room locations and object information, we use GPT-4 for controllable and open-vocabulary instruction generation. The prompt consists of a task instruction that defines the generation task, and in-context examples that constrain the output style. + +in relative distance", to organize the multi-source information to ease GPT generation. Firstly, we used RAM [61] to annotate the object categories within the frames. Based on these category tags, we employed Grounding DINO [38] to locate the objects in the frames. Subsequently, we used Depth-Anything [59] to predict the depth maps corresponding to the frames. + +Using this data, we identify the spatial locations and distances of objects relative to the current camera position. By analyzing object bounding box center positions and depth map locations, we can generate frame captions, as illustrated in Figure 2. Finally, objects in different frames can be easily correlated and capture the progression across different frames. More details in the spatial awareness data generation are provided in Appendix 9. + +Room Location Annotation in Videos. To determine the camera of each frame, w.r.t. the room category, we used BLIP2 [29] in visual question-answering mode, posing the question, "Which room am I in?" A predefined list of 16 common room types (i.e., bathroom, kitchen, bedroom, living room, hallway, office, dining room, foyer, laundry room, outside, porch, garage, front, patio, driveway, and backyard) was used as possible answers. This list was curated by analyzing 10 randomly sampled long videos, using BLIP2 in generative mode to identify and rank the most frequently mentioned room types. For frame-level predictions, we switched to BLIP2's discriminative mode and applied temporal smoothing to denoise outputs. + +We validated this approach by manually annotating 50 video clips, achieving an accuracy of $85\%$ . The use of + +BLIP2 leverages its open-world knowledge, while limiting room types to 16 categories for discriminative selection simplifies outputs and reduces ambiguity. Any loss in open-vocabulary flexibility is addressed during GPT-based trajectory summarization. + +Controllable Instruction Generation. To generate descriptions that accurately capture human-walking trajectories and reflect the environment, we integrate frame-level room locations with frame captions composed of object descriptions. We then employ GPT-4-Turbo [40] for controllable instruction generation, leveraging the multi-source information contained in the composed captions. As depicted in Figure 2, we organize the prompt using the "Task instruction - In-context examples - Prediction" scheme. This approach defines our instruction generation task as describing object progression along the moving trajectories, and includes two examples to ensure GPT produces instruction-style texts only. As shown in Figure 2 (c), captions of frames along the trajectory are embedded into the prompt and input into GPT. + +# 3.2. Action-Enriched Trajectories + +3D Environment Reconstruction. To obtain the geometric information of trajectories within RoomTour3D, we employ COLMAP [46, 47] for 3D reconstruction. This process allows us to infer the 3D layout of environments in the videos, providing a detailed geometric context for navigation tasks. Specifically, we sample the videos at 3 frames per second to balance accuracy and execution time. To further improve time efficiency, we split the videos into 100-second video clips with 10-second overlaps between adjacent clips and + +![](images/e19f7087947122c0d3fc40f131c599bd7334c845cd723cacb4e816e05627cfa3.jpg) +(a) Pretraining with Summarization Task + +# Output: + +Move through the corridor and step forwards.Pass by a paint on the wall. Then advance through the bedroom and into the bathroom.Start with a view of the bed,curtain.. + +![](images/1838f762e8d572fe34dfde7ee2d1506cdd8c4333250bf16271252c30888f763f.jpg) +Embodied Agent + +Output: + + + +Summarization + +Navigation + +: + +![](images/5a07fc8369152ba31dab3d9612dc1d8aaab2dfe3a1a4560859ee1169b74c1159.jpg) +(b) Finetuning with Navigation Task + +![](images/9538f1c804d37c0a76cb42c6b789daf2d8c94bd50f6e3fed3d316fede151c528.jpg) + +![](images/8e118a104bf57f46bbeeb47c8f83214436bab7676019c225eac5490a517df05d.jpg) + +![](images/7132859559d05b920b5f5f44cdd5d42af2e1e90601e05ae4b3886b993d73f3ef.jpg) + +Move through the corridor and step forwards... advance through the bedroom ... Start with a view of the bed, curtains... Move closer, noting the headboard... Pass the dresser and ... + +Task: + +Compare the History and Instruction to infer your current progress, and then select the correct direction from the candidates to go to the target location. + +< Cand>: + +![](images/eceadb299b02e2ebd900bf6fec1c6e85e7623a658b5fad72f2ef98223754423a.jpg) + +![](images/9cc547a5a38a2057ed7f3e72fb00ffe50e1f8789d9af34313daccdbd6fef4f1d.jpg) +Figure 3. Model training diagram with RoomTour3D. We design two tasks for our RoomTour3D to boost NaviLLM. (a) Pretraining: Sampled frames on the trajectory are treated as candidate observations. Model is optimized to summarize object progression along the path. (b) Finetuning: Each frame is considered as a navigable step. Given historical observation $<0>$ to $<2>$ and navigation instruction, the model is prompted to predict the next action by selecting from candidate observations View A, View B, View C and View D. + +![](images/1660967e4f96da8597f609b3aaa90d22a702bb2340d61c876ba5e7262db9e310.jpg) + +![](images/cbd8c90c1b6c5ed22a2425cb4c1f7b969152dc941be7c90b9e02bde90453f9d8.jpg) + +perform reconstruction on the clips simultaneously. + +Following this, we merge the resulting sub-models reconstructed from the video clips. For every two adjacent clips, if the reconstructed models have more than three overlapping frames, the models are merged and readjusted into one. However, due to varying reconstruction quality, a single video clip can produce more than one model. To manage this, we construct a graph for merging these sub-models. Each model is treated as a node, and we connect two models if they have more than three overlapping frames. We then apply Depth-First Search [50] to merge any two connected models and replace the original model nodes with the new merged one, continuing this process until no connected nodes remain. + +Navigable Action Sampling. We enhance navigation action diversity by using frames from room tour videos as navigable actions by sampling at significant view-change points within a small radius. These points are identified by reconstructing 3D scenes to measure camera orientation differences and distances between frames, capturing varied views from revisited locations or turning points. Frames with substantial view changes are retained using cosine similarity thresholding, followed by non-maximum suppression to isolate major shifts. DBSCAN clustering [11] groups spatially close frames with different views to ensure diverse navigable actions. These measures ensure the robustness to the misidentification of significant view change points that may arise due to inaccuracies in 3D reconstruction. Finally, we identify distinct walking paths within each cluster. For each path, we select the most recent frame as a positive candidate and the frame with the highest angular difference as a negative candidate, enhancing the diversity of navigable actions. More detailed description of our navigable points + +sampling approach are provided in Appendix 8. + +# 4. Vision-and-Language Navigation Model + +In this section, we introduce a practice to use our data to train a generalist embodied agent. To start with, we first provide a concise introduction to the state-of-the-art VLN model, NaviLLM [63], which is an LLM-based navigation agent. Then, we introduce two tasks that are adapted for our RoomTour3D data, i.e., vision-instruction summarization task for pretraining and action-instruction navigation task for finetuning. + +# 4.1. Revisiting NaviLLM + +NaviLLM is a SOTA LLM-based model for embodied navigation, excelling on benchmarks like CVDN and SOON. It processes panoramic inputs by encoding environmental views and integrating them with navigation instructions. Specific tokens are defined for different types of inputs: for historical observations and for candidate views at each navigational step. + +During training, NaviLLM receives instructions and a sequence of candidate views. At each step, the candidate observations are input into the model along with the instructions. The model predicts the next action by selecting the appropriate view from the candidates, and this selected view is then cached as a token, updating the model's internal state for future decisions. At the last step of the navigation task, the model summarizes all tokens as a separate training task to ensure comprehensive understanding and retention of the navigational history. For testing, the model similarly uses historical observations, i.e., tokens, accumulated during navigation, and candidate views, i.e., , at each step to decide the next + +action. This process ensures that the model's actions align with the given instructions and the observed environment. + +# 4.2. Summarization Task for Pretraining + +To leverage the rich, sequential nature of videos and enhance future planning capabilities, we adapt NaviLLM to use the description-enriched trajectories from our RoomTour3D dataset for a summarization task. Each frame is treated as a candidate view and wrapped with $\langle \text{cand} \rangle$ tokens. Similar to selected panoramic views at each navigation step, these frames are considered as selected actions when executing navigation instructions. As shown in Figure 3(a), the frame tokens and summarization task instruction are organized into a unified prompt and input into the LLM. The model is expected to output a trajectory summary containing object progression and room locations, as specified by the task instruction. The model is trained using next-token prediction loss, consistent with the original language model training methodology. + +# 4.3. Navigation Task for Finetuning + +In order to enable learning navigation decision-making from scalable scenes, we adapt NaviLLM to action-enriched trajectories from our RoomTour3D. Unlike panoramic views capturing observations from a single location, our data provides candidate views from frames at different locations and orientations, with only one frame directed toward the destination. Each frame in the video sequence is treated as a potential navigable action and wrapped with $\langle \text{cand} \rangle$ tokens. These candidate views are presented to the model and processed in the same way as panoramic views. As shown in Figure 3(b), the model processes the inputs to predict the next action, selecting the appropriate frame from the candidate views. The selected action is then cached as a $\langle \text{hist} \rangle$ token for subsequent decision-making steps. + +During fine-tuning, each frame is treated as a navigable step, with the next trajectory frame as the target action and $\langle \text{STOP} \rangle$ as an alternative. The model uses historical observations and navigation instructions to iteratively predict the next action, building a detailed understanding of the path. At the final step, the model summarizes the navigation path, incorporating object progression and room locations. This summarization task enhances its ability to recall navigational history and improves performance. + +# 5. Tasks and Experiments + +This section outlines our experimental setup and presents the results. Detailed implementation information can be found in Appendix 11. + +Datasets. During pretraining, we follow practice from NaviLLM [63] and perform teacher-forcing training on the combined dataset from our video-instruction data from + +RoomTour3D, together with CVDN [51], SOON [66], R2R [2], REVERIE [42] and ScanQA [3], and augmented data from R2R and REVERIE. In the multi-task fine-tuning stage, we alternate between teacher forcing and student forcing on the combined data from our action-instruction data from RoomTour3D, together with CVDN, SOON, R2R, REVERIE, ScanQA and LLaVA-23k [36]. + +To evaluate the impact of our data on navigation agent training, we test on CVDN, SOON, R2R, and REVERIE. CVDN requires navigating towards a target by understanding dialog history, linking dialogue comprehension to actions. SOON tasks the agent with locating objects without bounding boxes, emphasizing semantic-visual alignment. R2R involves following step-by-step instructions, requiring dynamic progress tracking and precise alignment with navigational history. REVERIE focuses on localizing distant objects based on concise instructions, aided by ground truth bounding boxes at waypoints. + +Evaluation Metrics. For the navigation tasks, we follow the evaluation methodology from [2] using the following navigation metrics: Success Rate (SR), which measures whether the agent reaches the target location within a set distance threshold; Success Rate Weighted by Path Length (SPL), which is the SR adjusted by the ratio of the ground truth path length to the actual path traveled; Goal Progress (GP), the advancement in meters towards the goal. GP is utilized for the CVDN dataset, whereas SR and SPL are the metrics for other datasets. + +# 5.1. Comparison on Supervised Tasks + +As shown in Table 1, we performed a one-time fine-tuning on the four tasks in a fully supervised manner. To begin, our experiments reiterate the superiority of multitask training over single-task training. Also, incorporating our RoomTour3D data into the pre-training process led to consistent improvements across all metrics on Val-U, achieving state-of-the-art results in the GP metric in the CVDN dataset. + +Notably, finetuning with our action-enriched data results in state-of-the-art performance on both Val-U and Test sets across SOON, R2R and REVERIE tasks. While the improvement on the CVDN and SOON datasets is modest, the most significant boost compared to the reproduced baseline is observed in R2R Val-U and REVERIE Val-U, with gains of approximately $5.7\%$ and $6\%$ , respectively. The improvement in R2R is largely driven by enhanced spatial awareness, stemming from the inclusion of proximity data, which helps the model better understand object distance and position. Similarly, gains in REVERIE are attributed to a combination of open-vocabulary tags, spatial awareness, and the addition of room type data, which encourages the model to infer the layout of environments, thereby boosting its spatial reasoning capabilities. Moreover, our use of open-ended instructions allows the model to adapt flexibly to diverse + +Table 1. Overall comparison with the baseline methods. Our RoomTour3D data can boost NaviLLM by a margin on SOON, R2R and REVERIE on SPL metric and on CVDN GP metric. *denotes reproduced results. RT3DDesc and RT3DAction stand for description-enriched trajectories only and action-enriched trajectories. + +
MethodsCVDNSOONR2RREVERIE
Val-UTestVal-UTestVal-UTestVal-UTest
Models Focusing on Single Task
PREVALENT [16]3.152.44--5351--
HOP [43]4.413.24--575926.124.3
HAMT [6]5.135.58--616030.226.7
DUET [7]--22.621.4605833.736.0
VLN-SIG [25]5.525.83--6260--
VLN-PETL [44]5.696.13--605827.726.7
NavGPT2 [65]----6160--
BEV-BERT [1]----646036.436.4
Unified Model For All Tasks
NaviLLM(w. Pretrain) [63]6.167.9029.226.3596035.732.3
NaviLLM(w. Pretrain)*6.09-28.0-56.7-31.4-
NaviLLM+RT3DDesc(Ours)6.967.5530.226.562.361.837.135.1
NaviLLM+RT3DAction(Ours)6.337.2231.727.862.462.237.436.4
+ +Table 2. Ablation study on the input modalities for trajectory summarization task. + +
Object tagsDepth & Bounding BoxRoom typeCVDN GP↑SOONR2RREVERIE
SR↑SPL↑SR↑SPL↑SR↑SPL↑
×××6.0933.6428.0165.5256.6738.3231.35
✓××5.4132.5226.5163.6155.7642.5234.37
✓✓×6.4937.6230.4068.3761.7041.7236.04
✓✓✓6.9638.8030.2169.3762.2843.2537.10
+ +Table 3. Overall comparison with SOTA zero-shot methods on R2R. ${}^{ \dagger }$ denotes training exclusive navigable actions. ${}^{ * }$ denotes using 36 views setting. RT3D stands for our RoomTour3D. + +
MethodsVal Unseen
SR↑SPL↑
Random Walk [41]32
Commercial Model
NavGPT(GPT-3.5) [64]*13.899.12
NavGPT(GPT-4) [64]3429
MapGPT(GPT-4) [5]38.825.8
MapGPT(GPT-4V) [5]43.734.8
DiscussNav(GPT-4) [39]4340
Open-source Model
LangNav(LLaMA2-7B) [41]00
NavCoT(LLaMA2-7B) [33]7.786.50
DuET (Init. LXMERT [48])10
NaviLLM [63]†00
NaviLLM+RT3D(Ours)14.3310.86
+ +scenarios, fostering more robust and generalizable performance and better contextual understanding. + +# 5.2. Comparison on Zero-shot Task + +To further demonstrate the substantial indoor knowledge contained in our data and its effectiveness for embodied action and language instructions, we conduct zero-shot experiments on embodied action prediction, as shown in Table 3. + +We removed all action and geometric data from the training datasets and retrained NaviLLM with and without our RoomTour3D dataset. Without action prediction data, NaviLLM lacked the ability to learn effective navigable action selection. However, with the inclusion of our action-enriched trajectories, NaviLLM achieved a $14.33\%$ SR and a $10.86\%$ SPL, outperforming open-source models built on LLaMA-7B and reaching results comparable to NavGPT [64], which leverages GPT-3.5. These improvements validate the effectiveness of our 3D trajectories mined from room tour video reconstructions and emphasize the value of our action-enriched trajectories. This highlights the significant contribution of our dataset to advancing open-world navigation. + +# 5.3. Ablation study + +Effect of open-world semantics and spatial awareness. As shown in Table 2, we analyzed the impact of various in + +Instruction: Exit the sewing room. Turn right. Go toward the glass cabinet with the dolls in it. Turn into the doorway on the left. Pass the bed and go through the next doorway on the left into the bathroom. Wait by the sink. (Instruction_id : 4676_0) + +![](images/f9bf6007d195df537182c3350ca85990d0a7ef379298454e9619c57b7b63bba6.jpg) +Figure 4. Paths of NaviLLM [63] and ours on R2R-unseen. Purple and green circles denote the start and target locations, respectively, and the red circle represents incorrect endpoint. According to the instruction, the agent should turn left at the waypoint marked with yellow. Our method makes the correct decision, while the baseline is confused by similar entrance at the waypoint, thus mistakenly turns right. + +![](images/b08e8dfcb8e39667396b9957e364d71711d36f4edfc1075952c3cd28f3deea88.jpg) + +formation types on instruction generation. Adding object variety significantly improved performance on REVERIE with increase SPL from $31.35\%$ to $34.37\%$ , as this dataset relies on object grounding. However, it had no direct impact on SOON, possibly because SOON relies solely on detailed textual descriptions without explicit bounding box annotations. After introducing depth estimation, which helps determine the relative distances of objects, the performance on SOON, R2R and REVERIE achieve marginal boosts. This demonstrates that enhancing spatial awareness significantly contributes to indoor navigation tasks. Furthermore, incorporating room locations, which capture the scene semantics along the trajectory, provided a moderate boost across all four VLN tasks. This further highlights the critical role of object variety and spatial awareness in improving navigation performance. + +Effect of action-instruction data. As shown in Table 1, we test the effect of incorporating video-action-instruction data into the training dataset. It is evident that this approach improves the SPL metric across the test scenes of SOON, R2R, and REVERIE. We believe that incorporating geometric information and movement-inclusive instructions helps the model better align the relationship between action and observation changes, thereby further enhancing the model's embodied capabilities. + +# 5.4. Data correctness verification + +We evaluated the correctness of our automated data-generation pipeline by manually rating 100 randomly sampled trajectory descriptions on a 4-point relevance scale: 1 for "totally irrelevant", 2 for "partially relevant", 3 for "mostly relevant" and 4 for a "perfect match". The evaluation yielded an average score of 3.08, with $74\%$ of descriptions rated as "mostly relevant" or "perfect match," demon- + +strating the method's effectiveness in generating meaningful, visually aligned descriptions. + +# 5.5. Navigation Case Visualization + +As shown in Figure 4, selecting the correct action is critical at specific decision points, such as when a left turn is required to follow the instruction accurately. In the example, at step $②$ , both the rooms to the left and right could satisfy the latter part of the instruction, "pass the bed and go into the bathroom." However, the baseline method incorrectly chooses a right turn at the designated left-turn point, causing it to deviate from the intended path. Once this error occurs, even with scene graph history, the model struggles to realign with the correct trajectory. This challenge is particularly common in household environments, where bedroom layouts often appear similar. It further demonstrates the effectiveness of our data alignment in improving adherence to action-based instructions. + +# 6. Conclusion + +In this paper, we present RoomTour3D, a novel dataset automatically curated from room tour videos for VLN tasks. By leveraging the rich, sequential nature of video data and incorporating object variety and spatial awareness, we generate 200K navigation instructions and 17K action-enriched trajectories from 1847 room tour scenes. Additionally, we produce navigable trajectories from video frames and reconstructed 3D scenes, which significantly boost the performance and set new state-of-the-art results on the SOON and REVERIE benchmarks. This approach also enables the development of a trainable zero-shot navigation agent, demonstrating the effectiveness and scalability of RoomTour3D in advancing VLN research. + +# References + +[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. BEVbert: Multimodal map pre-training for language-guided navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2737-2748, 2023. 2, 7 +[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments, 2018. 1, 2, 6 +[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 5 +[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 5 +[5] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee K Wong. Mapgpt: Map-guided prompting for unified vision-and-language navigation. arXiv preprint arXiv:2401.07314, 2024. 3, 7 +[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 7 +[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation, 2022. 2, 7 +[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation, 2022. 2 +[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 5 +[10] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering, 2017. 2 +[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, page 226-231. AAAI Press, 1996. 5, 1 +[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation, 2018. 2 +[13] Tsu-Jui Fu, Xin Eric Wang, Matthew Peterson, Scott Grafton, Miguel Eckstein, and William Yang Wang. Coun + +terfactual vision-and-language navigation via adversarial path sampling, 2020. 2 +[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14911-14920, 2023. 2 +[15] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation, 2021. 1, 2 +[16] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13137-13146, 2020. 2, 7 +[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language BERT for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2 +[18] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3d world into large language models, 2023. +[19] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6683-6693, 2023. +[20] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding, 2023. 2 +[21] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. arXiv preprint arXiv:1905.12255, 2019. 1 +[22] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning, 2023. 2 +[23] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 2 +[24] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 1, 2 +[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 2, 7 +[26] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language + +navigation. Advances in Neural Information Processing Systems, 36:21878-21894, 2023. 2 +[27] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information, 2021. 2 +[28] Jialu Li, Hao Tan, and Mohit Bansal. EnvEdit: Environment editing for vision-and-language navigation, 2022. 2 +[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 3, 4 +[30] Xiangyang Li, Zihan Wang, Jiahao Yang, Yaowei Wang, and Shuqiang Jiang. KERM: Knowledge enhanced reasoning for vision-and-language navigation, 2023. 2 +[31] Xiwen Liang, Fengda Zhu, Lingling Li, Hang Xu, and Xiaodan Liang. Visual-language navigation pretraining via prompt-based environmental self-exploration, 2022. 2 +[32] Xiwen Liang, Liang Ma, Shanshan Guo, Jianhua Han, Hang Xu, Shikui Ma, and Xiaodan Liang. CorNav: Autonomous agent with self-corrected planning for zero-shot vision-and-language navigation, 2024. 1 +[33] Bingqian Lin, Yunshuang Nie, Ziming Wei, Jiaqi Chen, Shikui Ma, Jianhua Han, Hang Xu, Xiaojun Chang, and Xiaodan Liang. NavCoT: Boosting llm-based vision-and-language navigation via learning disentangled reasoning. arXiv preprint arXiv:2403.07376, 2024. 3, 7 +[34] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H. Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos, 2023. 1, 2 +[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup, 2021. 2 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 6 +[37] Rui Liu, Wenguan Wang, and Yi Yang. Volumetric environment representation for vision-language navigation. In CVPR, pages 16317-16328, 2024. 2 +[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding Dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 4 +[39] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. arXiv preprint arXiv:2309.11382, 2023. 2, 3, 7 +[40] OpenAI. GPT-4: Generative pre-trained transformer 4. OpenAI API, 2022. 2, 3, 4, 1 +[41] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. arXiv preprint arXiv:2310.07889, 2023. 3, 7 +[42] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: Remote embodied visual referring expression in real indoor environments, 2020. 1, 2, 6 + +[43] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: History-and-order aware pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15418-15427, 2022. 2, 7 +[44] Yanyuan Qiao, Zheng Yu, and Qi Wu. VLN-PETL: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 2, 7 +[45] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-Matterport 3D dataset (HM3D): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1 +[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 4, 2, 3 +[47] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1, 4, 2, 3 +[48] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 7 +[49] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout, 2019. 2 +[50] Robert Tarjan. Depth-first search and linear graph algorithms. In 12th Annual Symposium on Switching and Automata Theory (swat 1971), pages 114-121, 1971. 5 +[51] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation, 2019. 1, 2, 6 +[52] Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu, and Qijun Chen. Vision-and-language navigation via causal learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[53] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks, 2022. 2 +[54] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIV 16, pages 413–430. Springer, 2020. 1 +[55] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 1, 2 +[56] Erik Wijmans, Samyak Datta, Oleksandr Maksymets, Abhishek Das, Georgia Gkioxari, Stefan Lee, Irfan Essa, Devi Parikh, and Dhruv Batra. Embodied question answering + +in photorealistic environments with point cloud perception, 2019. 2 +[57] Wikipedia. Preferred walking speed. https://en.wikipedia.org/wiki/Preferred_walking_speed.3 +[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1 +[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4 +[60] Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and Wang He. NaVid: Video-based vlm plans the next step for vision-and-language navigation. arXiv preprint arXiv:2402.15852, 2024. 1 +[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize Anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 2, 3, 4 +[62] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation, 2024. 2 +[63] Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. Towards learning a generalist model for embodied navigation, 2023. 2, 5, 6, 7, 8 +[64] Gengze Zhou, Yicong Hong, and Qi Wu. NavGPT: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 3, 7 +[65] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 7 +[66] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. SOON: Scenario oriented object navigation with graph-based exploration, 2021. 1, 2, 6 \ No newline at end of file diff --git a/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/images.zip b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..220916a90a33579f740a05f37d05ce6c8c2aa47f --- /dev/null +++ b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:640c31ca540a834daf30d4ab3fd3616305f4a32e0ef1c4e6f018fd9c0c5e176f +size 625295 diff --git a/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/layout.json b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e2c48427e335e77a0fd90280f2b1028ff6376a64 --- /dev/null +++ b/2025/RoomTour3D_ Geometry-Aware Video-Instruction Tuning for Embodied Navigation/layout.json @@ -0,0 +1,7850 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 486, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 486, + 139 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 486, + 139 + ], + "type": "text", + "content": "RoomTour3D: Geometry-Aware Video-Instruction Tuning for Embodied Navigation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "text", + "content": "Mingfei Han\\*1,3, Liang Ma\\*1, Kamila Zhumakhanova\\*1, Ekaterina Radionova\\*1, Jingyi Zhang\\*2 \nXiaojun Chang\\*1,4, Xiaodan Liang\\*1,2, Ivan Laptev\\*1 \n" + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "text", + "content": "Department of Computer Vision, MBZUAI " + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "text", + "content": "Shenzhen Campus of Sun Yat-Sen University \n" + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "text", + "content": "ReLER Lab, AAII, UTS " + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 160, + 533, + 232 + ], + "type": "text", + "content": "University of Science and Technology of China \nhttps://roomtour3d.github.io" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 270, + 200, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 270, + 200, + 284 + ], + "spans": [ + { + "bbox": [ + 151, + 270, + 200, + 284 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "text", + "content": "Vision-and-Language Navigation (VLN) suffers from the limited diversity and scale of training data, primarily constrained by the manual curation of existing simulators. To address this, we introduce RoomTour3D, a video-instruction dataset derived from web-based room tour videos that capture real-world indoor spaces and human walking demonstrations. Unlike existing VLN datasets, RoomTour3D leverages the scale and diversity of online videos to generate open-ended human walking trajectories and open-world navigable instructions. To compensate for the lack of navigation data in online videos, we perform 3D reconstruction and obtain 3D trajectories of walking paths augmented with additional information on the room types, object locations and 3D shape of surrounding scenes. Our dataset includes " + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "inline_equation", + "content": "\\sim 100K" + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "text", + "content": " open-ended description-enriched trajectories with " + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "inline_equation", + "content": "\\sim 200K" + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "text", + "content": " instructions, and " + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "inline_equation", + "content": "17K" + }, + { + "bbox": [ + 55, + 297, + 297, + 571 + ], + "type": "text", + "content": " action-enriched trajectories from 1847 room tour environments. We demonstrate experimentally that RoomTour3D enables significant improvements across multiple VLN tasks including CVDN, SOON, R2R, and REVERIE. Moreover, RoomTour3D facilitates the development of trainable zero-shot VLN agents, showcasing the potential and challenges of advancing towards open-world navigation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 698 + ], + "type": "text", + "content": "Over the past years, Vision-and-Language Navigation [2, 21, 24, 32, 42, 54] has largely relied on human-designed simulators and annotated trajectories. R2R [2] initiated a benchmark for language-guided navigation in simulated indoor settings, while CVDN [51], REVERIE [42], and SOON [66] expanded VLN to dialogue-based and object-focused tasks. However, these curated simulations lack scene diversity and fail to capture real-world complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 464 + ], + "type": "text", + "content": "To address limited diversity, recent methods propose the use of richer and more varied training data. AirBERT [15] combines discrete Airbnb images for panoramic views, which lack consistency and naturalistic context of an indoor scene. ScaleVLN [55] utilizes laboriously curated 3D scenes [45, 58], but suffers from reconstruction quality and scalability. More recently, YTB-VLN [34] attempts to use video frames to compose panoramic views and organize instructions with predefined templates, yet overlooks object variety and geometry structure. NaVid [60] constructs sequential single-view trajectories from MatterPort3D [4] and R2R [2] annotations, paired with general video data to train a sim-to-real agent. None of these approaches simultaneously achieves scalability in scene diversity, openness in object variety, or comprehensive geo-perception in spatial representations, each of which is critical to training effective and open-world navigation agents." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 650 + ], + "type": "text", + "content": "To address the challenge, we introduce RoomTour3D, a novel dataset that provides a geometry-aware, spatially enriched training environment for VLN agents. Built upon easily accessible room tour videos from the Internet, RoomTour3D captures continuous movement through real estates with a hand-held camera from a first-person perspective. Each frame presents a realistic, agent-centric view and showcases a rich array of indoor items. The continuous flow of these frames captures multiple views of the environment, presenting diverse room layouts and inherently embedding the geometric properties of the spaces. To unleash the power of these videos, we propose an automatic and extendable pipeline to obtain open-ended geometry-aware human walking trajectories, spatially contextualized textual instructions using open vocabularies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "To better model the navigation scenario, we take advantage of the continuous walk-through trajectories and densely sample frames from room tour videos. Then, we use COLMAP [46, 47] to reconstruct 3D scenes of real-estates to obtain the geometric information. With known" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 132, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 132, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 132, + 712 + ], + "type": "text", + "content": "* Co-first authors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27586" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "camera locations and orientations, we sample \"decision-making\" frames at points of maximal yaw rotation and sequence further frames every " + }, + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "inline_equation", + "content": "\\sim 1.5" + }, + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": " meters to finalize trajectories. Additionally, our pipeline uses extensive annotations via RAM [61] for object tagging, Grounding-DINO [38] for precise localization, and Depth-Anything [59] for estimating relative object-camera distances. To integrate knowledge of object variety, geometric awareness, and human walking preferences into model training, we employ GPT-4 [40] to generate navigation instructions for both summarization and task-specific navigation tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "content": "Our RoomTour3D is an effort to create a comprehensive database derived from room tour videos and enriched by human-living knowledge. The dataset includes " + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{K}" + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "content": " open-ended trajectories with " + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\sim 200\\mathrm{K}" + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "content": " descriptions, and " + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\sim 17\\mathrm{K}" + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "content": " geometry-aware trajectories with navigable actions from 1847 homes. Moreover, we are releasing intermediate products such as object tags, bounding boxes, depth maps, room locations, and the necessary code and prompts used to generate instructions. To validate the robustness, we conducted experiments with NaviLLM [63], a generalist model based on Large Language Model (LLM), to train a unified multi-task navigation agent. Integrating our data improved baseline performances on CVDN, SOON, R2R, and REVERIE by over " + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "content": ", notably achieving a " + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 56, + 205, + 296, + 419 + ], + "type": "text", + "content": " boost on SOON and setting new state-of-the-art results. Furthermore, our enriched action-instruction data facilitates training an end-to-end zero-shot navigation agent, advancing towards open-world embodied navigation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 421, + 289, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 289, + 433 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 289, + 433 + ], + "type": "text", + "content": "In this work, we make the following key contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 613 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": "- Video Collection for Complex Environments: We curate a novel dataset of diverse videos tailored for navigation tasks, distinguishing it from existing datasets such as YTB-VLN [34]. Our dataset features longer videos, representing more complex environments, and has fewer shot changes, ensuring continuity and contextual consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 506, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 566 + ], + "type": "text", + "content": "- Automatic Video Processing: We propose a pipeline to extract geometry-aware navigation instructions, aligning spatial understanding with navigation goals. Additionally, we generate open-vocabulary instructions for open-ended trajectories to enhance real-world applicability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": "- Demonstrating Data Effectiveness: Through extensive experiments and ablation studies, we demonstrate that our dataset significantly improves the performance of state-of-the-art models." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 627, + 142, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 142, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 142, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 233, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 233, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 233, + 659 + ], + "type": "text", + "content": "2.1. Vision-and-Language Navigation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Navigating unseen indoor environments with natural language instructions is vital for enabling embodied agents to assist humans. Various scenarios have been explored, including fine-grained instruction following (R2R [2]) and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "dialogue-based navigation (CVDN [51]), object localization from instructions (SOON [66], REVERIE [42]), and embodied question answering through active 3D exploration [10, 56]. While substantial work focuses on task-specific models [1, 6, 7, 14, 16-20, 25, 30, 37, 39, 43, 44, 52, 62, 66], they often lack generalization across tasks. In this context, NaviLLM [63] introduces an embodied generalist model that simultaneously addresses multiple tasks through a single framework, demonstrating strong generalization ability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 198, + 485, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 485, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 485, + 209 + ], + "type": "text", + "content": "2.2. Data-Centric Methods for VLN" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 553, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 320 + ], + "type": "text", + "content": "The scarcity of VLN training data remains a critical issue and results in poor generalization of VLN agents to unseen environments. Most of existing VLN datasets such as R2R [2], RxR [24], CVDN [51] and SOON [66] are produced in simulators, which constrains data scalability due to the high labor costs involved. To tackle the problem, data augmentation [12, 13, 23, 27, 28, 35, 49] and self-exploration in simulator environments [31, 53] have been investigated." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 321, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 554, + 513 + ], + "type": "text", + "content": "VLN-BERT [17] and AirBERT [15] attempt to use web-based image-caption pairs for pre-training, however, resulting trajectories often fail to mimic realistic navigation. Similarly, automatic dataset generation pipelines [8, 22], including ScaleVLN [55], rely on manually curated 3D scenes or synthetic environments, which are costly to produce and lack the photorealism needed for robust real-world generalization. PanoGen [26] enhances VLN training by generating diverse text-conditioned panoramic environments using text-to-image diffusion models and recursive outpainting. While this approach addresses the scarcity of training environments, it relies on synthetic panoramas and may not generalize well to real environments. YTB-VLN [34] advances scalability by leveraging YouTube room tour videos to generate path-instruction pairs but omits explicit path geometry, essential for robotic navigation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 513, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 554, + 644 + ], + "type": "text", + "content": "In our work, we address the limitations by designing RoomTour3D with properties: (i) free-form and open-vocabulary path annotations instead of template-based instructions, (ii) extraction of open-ended trajectories from sequential video clips, and (iii) inclusion of turning points and spatially close frames as navigable candidate actions, moving beyond panoramic nodes. Furthermore, our approach integrates 3D reconstruction of indoor videos to retrieve trajectory geometry and employs an LLM to generate detailed, object-aware instructions with enhanced spatial understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 651, + 436, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 436, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 436, + 662 + ], + "type": "text", + "content": "2.3. Zero-shot Navigation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Given the substantial semantic variations in complex real-world scenarios, fully-supervised VLN models often struggle to generalize across diverse navigation scenes. Zero-shot VLN has thus gained attention as it eliminates prior" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27587" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 541, + 322 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 541, + 322 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 541, + 322 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 541, + 322 + ], + "type": "image", + "image_path": "8af1a3f33c14c429498d63dcb7b01f77acca153579fa8dc4ceb4ca5320fa13e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "text", + "content": "Figure 1. Overview of our RoomTour3D data generation. Starting from a room tour video, we first apply BLIP-2 [29] on frame sequence to predict the room locations. Next, we use RAM [61] and Grounding-DINO [38] to identify objects within the frames and employ Depth-Anything [59] for depth prediction. Subsequently, COLMAP is used to reconstruct the 3D scene with complete geometry information, and we sample human walking trajectories from the continuous frames. The trajectory captures open-world objects, their positions, and depths relative to the camera. Finally, we use advanced LLM, i.e., GPT-4 to generate the free-form descriptions for pretraining, namely description-enriched trajectories. Specifically, for the trajectory shown in the figure, which involves instant turning points, we specially treat " + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "inline_equation", + "content": "<0>" + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "inline_equation", + "content": "<6>" + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "text", + "content": " as walking trajectory, " + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "inline_equation", + "content": "<\\mathrm{A}> <\\mathrm{B}>" + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "inline_equation", + "content": "<\\mathrm{C}>" + }, + { + "bbox": [ + 54, + 328, + 555, + 417 + ], + "type": "text", + "content": " as side-watching points and use them as negative candidates for navigation finetuning task, namely action-enriched trajectories. For more details, please refer to Section 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 430, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 430, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 295, + 453 + ], + "type": "text", + "content": "knowledge of environments and instructions, effectively mitigating environmental biases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 453, + 295, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 453, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 54, + 453, + 295, + 596 + ], + "type": "text", + "content": "Commercial model based methods utilize advanced LLMs and robust frameworks for seamless solutions. MapGPT [5] incorporates a map-based prompting system with global spatial reasoning and adaptive path planning. DiscussNav [39] employs a multi-expert framework where LLMs specialize in subtasks like instruction analysis and vision perception. NavGPT [64] focuses on explicit reasoning by combining commonsense reasoning with visual observations. As to non-commercial methods, LangNav [41] uses language as the primary perceptual space, while NavCoT [33] introduces parameter-efficient training to allow LLMs to autonomously reason and act." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 597, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 597, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 54, + 597, + 295, + 645 + ], + "type": "text", + "content": "We show that usage of our action-enriched data for navigation tuning results in superior zero-shot performance over all non-commercial methods and reaches comparable results to commercial approaches based on GPT-3.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 142, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 142, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 142, + 661 + ], + "type": "text", + "content": "3. RoomTour3D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 666, + 295, + 714 + ], + "type": "text", + "content": "In this section, we present the automatic data curation pipeline of RoomTour3D. We detail the process of annotations, from sampling open-ended human walking trajectories to generating corresponding descriptions with open-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": "world object variety and spatial awareness. Enabled by reconstructed 3D scenes, we further sample navigable trajectories with actions. The overall pipeline of our data generation is illustrated in Figure 1. Please refer to Appendix 7 for details about video collection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 495, + 498, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 498, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 498, + 508 + ], + "type": "text", + "content": "3.1. Description-Enriched Trajectories" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 510, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 664 + ], + "type": "text", + "content": "In this subsection, we detail the process of generating controllable descriptions for open-ended trajectories. We start by generating human-walking trajectories by uniformly sampling frames at a rate of one frame every two seconds, which aligns with the average human walking speed of 1.42 meters per second [57], typically slower in indoor environments. Subsequently, to annotate these trajectories, as shown in Figure 2, we employ expert models such as BLIP-2 [29], RAM [61], Grounding-DINO [38], and Depth-Anything [59] to gather extensive information on object variety, spatial positions, and depth measurements. Finally, we integrate this information into GPT4 [40] to generate detailed and coherent traje." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Object Variety and Spatial Awareness. In order to harness object variety and enable spatial awareness, we compose three expert models and design a textual template, i.e., \"There is a object tag to the spatial position of current spot" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27588" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 552, + 273 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 552, + 273 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 552, + 273 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 552, + 273 + ], + "type": "image", + "image_path": "911aaf5c0c530d215601da82152c47e13079298286803f71626c19a04793eb77.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 279, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 279, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 555, + 335 + ], + "type": "text", + "content": "Figure 2. Instruction generation in a controllable way. (a) Using open-sourced expert models, we identify what objects are in the frames, and assess how far an object is and determine where an object is located. The information is then textualized to create richly detailed frame captions. (b) BLIP-2 is adopted to predict and smooth room location across sequential frames. (c) Combining room locations and object information, we use GPT-4 for controllable and open-vocabulary instruction generation. The prompt consists of a task instruction that defines the generation task, and in-context examples that constrain the output style." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 294, + 430 + ], + "type": "text", + "content": "in relative distance\", to organize the multi-source information to ease GPT generation. Firstly, we used RAM [61] to annotate the object categories within the frames. Based on these category tags, we employed Grounding DINO [38] to locate the objects in the frames. Subsequently, we used Depth-Anything [59] to predict the depth maps corresponding to the frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "text", + "content": "Using this data, we identify the spatial locations and distances of objects relative to the current camera position. By analyzing object bounding box center positions and depth map locations, we can generate frame captions, as illustrated in Figure 2. Finally, objects in different frames can be easily correlated and capture the progression across different frames. More details in the spatial awareness data generation are provided in Appendix 9." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 531, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 686 + ], + "type": "text", + "content": "Room Location Annotation in Videos. To determine the camera of each frame, w.r.t. the room category, we used BLIP2 [29] in visual question-answering mode, posing the question, \"Which room am I in?\" A predefined list of 16 common room types (i.e., bathroom, kitchen, bedroom, living room, hallway, office, dining room, foyer, laundry room, outside, porch, garage, front, patio, driveway, and backyard) was used as possible answers. This list was curated by analyzing 10 randomly sampled long videos, using BLIP2 in generative mode to identify and rank the most frequently mentioned room types. For frame-level predictions, we switched to BLIP2's discriminative mode and applied temporal smoothing to denoise outputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We validated this approach by manually annotating 50 video clips, achieving an accuracy of " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". The use of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "type": "text", + "content": "BLIP2 leverages its open-world knowledge, while limiting room types to 16 categories for discriminative selection simplifies outputs and reduces ambiguity. Any loss in open-vocabulary flexibility is addressed during GPT-based trajectory summarization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 408, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 587 + ], + "type": "text", + "content": "Controllable Instruction Generation. To generate descriptions that accurately capture human-walking trajectories and reflect the environment, we integrate frame-level room locations with frame captions composed of object descriptions. We then employ GPT-4-Turbo [40] for controllable instruction generation, leveraging the multi-source information contained in the composed captions. As depicted in Figure 2, we organize the prompt using the \"Task instruction - In-context examples - Prediction\" scheme. This approach defines our instruction generation task as describing object progression along the moving trajectories, and includes two examples to ensure GPT produces instruction-style texts only. As shown in Figure 2 (c), captions of frames along the trajectory are embedded into the prompt and input into GPT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 590, + 474, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 474, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 474, + 604 + ], + "type": "text", + "content": "3.2. Action-Enriched Trajectories" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "3D Environment Reconstruction. To obtain the geometric information of trajectories within RoomTour3D, we employ COLMAP [46, 47] for 3D reconstruction. This process allows us to infer the 3D layout of environments in the videos, providing a detailed geometric context for navigation tasks. Specifically, we sample the videos at 3 frames per second to balance accuracy and execution time. To further improve time efficiency, we split the videos into 100-second video clips with 10-second overlaps between adjacent clips and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27589" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 87, + 307, + 148 + ], + "blocks": [ + { + "bbox": [ + 110, + 73, + 255, + 83 + ], + "lines": [ + { + "bbox": [ + 110, + 73, + 255, + 83 + ], + "spans": [ + { + "bbox": [ + 110, + 73, + 255, + 83 + ], + "type": "text", + "content": "(a) Pretraining with Summarization Task" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 87, + 307, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 87, + 307, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 87, + 307, + 148 + ], + "type": "image", + "image_path": "e19f7087947122c0d3fc40f131c599bd7334c845cd723cacb4e816e05627cfa3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 164, + 95, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 164, + 95, + 174 + ], + "spans": [ + { + "bbox": [ + 62, + 164, + 95, + 174 + ], + "type": "text", + "content": "Output:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 175, + 153, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 153, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 153, + 239 + ], + "type": "text", + "content": "Move through the corridor and step forwards.Pass by a paint on the wall. Then advance through the bedroom and into the bathroom.Start with a view of the bed,curtain.." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 178, + 178, + 199, + 200 + ], + "blocks": [ + { + "bbox": [ + 178, + 178, + 199, + 200 + ], + "lines": [ + { + "bbox": [ + 178, + 178, + 199, + 200 + ], + "spans": [ + { + "bbox": [ + 178, + 178, + 199, + 200 + ], + "type": "image", + "image_path": "1838f762e8d572fe34dfde7ee2d1506cdd8c4333250bf16271252c30888f763f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 185, + 266, + 194 + ], + "lines": [ + { + "bbox": [ + 207, + 185, + 266, + 194 + ], + "spans": [ + { + "bbox": [ + 207, + 185, + 266, + 194 + ], + "type": "text", + "content": "Embodied Agent" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 218, + 211, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 218, + 211, + 228 + ], + "spans": [ + { + "bbox": [ + 179, + 218, + 211, + 228 + ], + "type": "text", + "content": "Output:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 181, + 231, + 217, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 231, + 217, + 239 + ], + "spans": [ + { + "bbox": [ + 181, + 231, + 217, + 239 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 224, + 308, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 224, + 308, + 232 + ], + "spans": [ + { + "bbox": [ + 247, + 224, + 308, + 232 + ], + "type": "text", + "content": "Summarization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 247, + 236, + 294, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 236, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 247, + 236, + 294, + 243 + ], + "type": "text", + "content": "Navigation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 107, + 340, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 340, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 340, + 116 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 342, + 89, + 386, + 121 + ], + "blocks": [ + { + "bbox": [ + 367, + 73, + 495, + 83 + ], + "lines": [ + { + "bbox": [ + 367, + 73, + 495, + 83 + ], + "spans": [ + { + "bbox": [ + 367, + 73, + 495, + 83 + ], + "type": "text", + "content": "(b) Finetuning with Navigation Task" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 89, + 386, + 121 + ], + "lines": [ + { + "bbox": [ + 342, + 89, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 342, + 89, + 386, + 121 + ], + "type": "image", + "image_path": "5a07fc8369152ba31dab3d9612dc1d8aaab2dfe3a1a4560859ee1169b74c1159.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 394, + 89, + 436, + 120 + ], + "blocks": [ + { + "bbox": [ + 394, + 89, + 436, + 120 + ], + "lines": [ + { + "bbox": [ + 394, + 89, + 436, + 120 + ], + "spans": [ + { + "bbox": [ + 394, + 89, + 436, + 120 + ], + "type": "image", + "image_path": "9538f1c804d37c0a76cb42c6b789daf2d8c94bd50f6e3fed3d316fede151c528.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 444, + 89, + 488, + 120 + ], + "blocks": [ + { + "bbox": [ + 444, + 89, + 488, + 120 + ], + "lines": [ + { + "bbox": [ + 444, + 89, + 488, + 120 + ], + "spans": [ + { + "bbox": [ + 444, + 89, + 488, + 120 + ], + "type": "image", + "image_path": "8e118a104bf57f46bbeeb47c8f83214436bab7676019c225eac5490a517df05d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 509, + 90, + 547, + 118 + ], + "blocks": [ + { + "bbox": [ + 509, + 90, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 509, + 90, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 509, + 90, + 547, + 118 + ], + "type": "image", + "image_path": "7132859559d05b920b5f5f44cdd5d42af2e1e90601e05ae4b3886b993d73f3ef.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 126, + 548, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 126, + 548, + 154 + ], + "spans": [ + { + "bbox": [ + 342, + 126, + 548, + 154 + ], + "type": "text", + "content": "Move through the corridor and step forwards... advance through the bedroom ... Start with a view of the bed, curtains... Move closer, noting the headboard... Pass the dresser and ..." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 182, + 339, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 182, + 339, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 339, + 190 + ], + "type": "text", + "content": "Task:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 163, + 535, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 163, + 535, + 191 + ], + "spans": [ + { + "bbox": [ + 342, + 163, + 535, + 191 + ], + "type": "text", + "content": "Compare the History and Instruction to infer your current progress, and then select the correct direction from the candidates to go to the target location." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 214, + 342, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 214, + 342, + 223 + ], + "spans": [ + { + "bbox": [ + 312, + 214, + 342, + 223 + ], + "type": "text", + "content": "< Cand>:" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 354, + 198, + 386, + 238 + ], + "blocks": [ + { + "bbox": [ + 354, + 198, + 386, + 238 + ], + "lines": [ + { + "bbox": [ + 354, + 198, + 386, + 238 + ], + "spans": [ + { + "bbox": [ + 354, + 198, + 386, + 238 + ], + "type": "image", + "image_path": "eceadb299b02e2ebd900bf6fec1c6e85e7623a658b5fad72f2ef98223754423a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 397, + 198, + 439, + 239 + ], + "blocks": [ + { + "bbox": [ + 397, + 198, + 439, + 239 + ], + "lines": [ + { + "bbox": [ + 397, + 198, + 439, + 239 + ], + "spans": [ + { + "bbox": [ + 397, + 198, + 439, + 239 + ], + "type": "image", + "image_path": "9cc547a5a38a2057ed7f3e72fb00ffe50e1f8789d9af34313daccdbd6fef4f1d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "type": "text", + "content": "Figure 3. Model training diagram with RoomTour3D. We design two tasks for our RoomTour3D to boost NaviLLM. (a) Pretraining: Sampled frames on the trajectory are treated as candidate observations. Model is optimized to summarize object progression along the path. (b) Finetuning: Each frame is considered as a navigable step. Given historical observation " + }, + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "type": "inline_equation", + "content": "<0>" + }, + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "type": "inline_equation", + "content": "<2>" + }, + { + "bbox": [ + 55, + 255, + 555, + 299 + ], + "type": "text", + "content": " and navigation instruction, the model is prompted to predict the next action by selecting from candidate observations View A, View B, View C and View D." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 449, + 197, + 492, + 239 + ], + "blocks": [ + { + "bbox": [ + 449, + 197, + 492, + 239 + ], + "lines": [ + { + "bbox": [ + 449, + 197, + 492, + 239 + ], + "spans": [ + { + "bbox": [ + 449, + 197, + 492, + 239 + ], + "type": "image", + "image_path": "1660967e4f96da8597f609b3aaa90d22a702bb2340d61c876ba5e7262db9e310.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 500, + 197, + 544, + 239 + ], + "blocks": [ + { + "bbox": [ + 500, + 197, + 544, + 239 + ], + "lines": [ + { + "bbox": [ + 500, + 197, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 500, + 197, + 544, + 239 + ], + "type": "image", + "image_path": "cbd8c90c1b6c5ed22a2425cb4c1f7b969152dc941be7c90b9e02bde90453f9d8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 312, + 263, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 263, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 263, + 323 + ], + "type": "text", + "content": "perform reconstruction on the clips simultaneously." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 327, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 482 + ], + "type": "text", + "content": "Following this, we merge the resulting sub-models reconstructed from the video clips. For every two adjacent clips, if the reconstructed models have more than three overlapping frames, the models are merged and readjusted into one. However, due to varying reconstruction quality, a single video clip can produce more than one model. To manage this, we construct a graph for merging these sub-models. Each model is treated as a node, and we connect two models if they have more than three overlapping frames. We then apply Depth-First Search [50] to merge any two connected models and replace the original model nodes with the new merged one, continuing this process until no connected nodes remain." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "Navigable Action Sampling. We enhance navigation action diversity by using frames from room tour videos as navigable actions by sampling at significant view-change points within a small radius. These points are identified by reconstructing 3D scenes to measure camera orientation differences and distances between frames, capturing varied views from revisited locations or turning points. Frames with substantial view changes are retained using cosine similarity thresholding, followed by non-maximum suppression to isolate major shifts. DBSCAN clustering [11] groups spatially close frames with different views to ensure diverse navigable actions. These measures ensure the robustness to the misidentification of significant view change points that may arise due to inaccuracies in 3D reconstruction. Finally, we identify distinct walking paths within each cluster. For each path, we select the most recent frame as a positive candidate and the frame with the highest angular difference as a negative candidate, enhancing the diversity of navigable actions. More detailed description of our navigable points" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 312, + 507, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 507, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 507, + 324 + ], + "type": "text", + "content": "sampling approach are provided in Appendix 8." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 334, + 534, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 534, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 534, + 348 + ], + "type": "text", + "content": "4. Vision-and-Language Navigation Model" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "content": "In this section, we introduce a practice to use our data to train a generalist embodied agent. To start with, we first provide a concise introduction to the state-of-the-art VLN model, NaviLLM [63], which is an LLM-based navigation agent. Then, we introduce two tasks that are adapted for our RoomTour3D data, i.e., vision-instruction summarization task for pretraining and action-instruction navigation task for finetuning." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 456, + 433, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 433, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 433, + 470 + ], + "type": "text", + "content": "4.1. Revisiting NaviLLM" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": "NaviLLM is a SOTA LLM-based model for embodied navigation, excelling on benchmarks like CVDN and SOON. It processes panoramic inputs by encoding environmental views and integrating them with navigation instructions. Specific tokens are defined for different types of inputs: for historical observations and for candidate views at each navigational step." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "During training, NaviLLM receives instructions and a sequence of candidate views. At each step, the candidate observations are input into the model along with the instructions. The model predicts the next action by selecting the appropriate view from the candidates, and this selected view is then cached as a token, updating the model's internal state for future decisions. At the last step of the navigation task, the model summarizes all tokens as a separate training task to ensure comprehensive understanding and retention of the navigational history. For testing, the model similarly uses historical observations, i.e., tokens, accumulated during navigation, and candidate views, i.e., , at each step to decide the next" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27590" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "action. This process ensures that the model's actions align with the given instructions and the observed environment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 249, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 249, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 249, + 117 + ], + "type": "text", + "content": "4.2. Summarization Task for Pretraining" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "type": "text", + "content": "To leverage the rich, sequential nature of videos and enhance future planning capabilities, we adapt NaviLLM to use the description-enriched trajectories from our RoomTour3D dataset for a summarization task. Each frame is treated as a candidate view and wrapped with " + }, + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\langle \\text{cand} \\rangle" + }, + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "type": "text", + "content": " tokens. Similar to selected panoramic views at each navigation step, these frames are considered as selected actions when executing navigation instructions. As shown in Figure 3(a), the frame tokens and summarization task instruction are organized into a unified prompt and input into the LLM. The model is expected to output a trajectory summary containing object progression and room locations, as specified by the task instruction. The model is trained using next-token prediction loss, consistent with the original language model training methodology." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 224, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 224, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 224, + 320 + ], + "type": "text", + "content": "4.3. Navigation Task for Finetuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "text", + "content": "In order to enable learning navigation decision-making from scalable scenes, we adapt NaviLLM to action-enriched trajectories from our RoomTour3D. Unlike panoramic views capturing observations from a single location, our data provides candidate views from frames at different locations and orientations, with only one frame directed toward the destination. Each frame in the video sequence is treated as a potential navigable action and wrapped with " + }, + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "inline_equation", + "content": "\\langle \\text{cand} \\rangle" + }, + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "text", + "content": " tokens. These candidate views are presented to the model and processed in the same way as panoramic views. As shown in Figure 3(b), the model processes the inputs to predict the next action, selecting the appropriate frame from the candidate views. The selected action is then cached as a " + }, + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "inline_equation", + "content": "\\langle \\text{hist} \\rangle" + }, + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "text", + "content": " token for subsequent decision-making steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 612 + ], + "type": "text", + "content": "During fine-tuning, each frame is treated as a navigable step, with the next trajectory frame as the target action and " + }, + { + "bbox": [ + 55, + 504, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\langle \\text{STOP} \\rangle" + }, + { + "bbox": [ + 55, + 504, + 295, + 612 + ], + "type": "text", + "content": " as an alternative. The model uses historical observations and navigation instructions to iteratively predict the next action, building a detailed understanding of the path. At the final step, the model summarizes the navigation path, incorporating object progression and room locations. This summarization task enhances its ability to recall navigational history and improves performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 621, + 190, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 190, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 190, + 635 + ], + "type": "text", + "content": "5. Tasks and Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "content": "This section outlines our experimental setup and presents the results. Detailed implementation information can be found in Appendix 11." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Datasets. During pretraining, we follow practice from NaviLLM [63] and perform teacher-forcing training on the combined dataset from our video-instruction data from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "RoomTour3D, together with CVDN [51], SOON [66], R2R [2], REVERIE [42] and ScanQA [3], and augmented data from R2R and REVERIE. In the multi-task fine-tuning stage, we alternate between teacher forcing and student forcing on the combined data from our action-instruction data from RoomTour3D, together with CVDN, SOON, R2R, REVERIE, ScanQA and LLaVA-23k [36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 289 + ], + "type": "text", + "content": "To evaluate the impact of our data on navigation agent training, we test on CVDN, SOON, R2R, and REVERIE. CVDN requires navigating towards a target by understanding dialog history, linking dialogue comprehension to actions. SOON tasks the agent with locating objects without bounding boxes, emphasizing semantic-visual alignment. R2R involves following step-by-step instructions, requiring dynamic progress tracking and precise alignment with navigational history. REVERIE focuses on localizing distant objects based on concise instructions, aided by ground truth bounding boxes at waypoints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 408 + ], + "type": "text", + "content": "Evaluation Metrics. For the navigation tasks, we follow the evaluation methodology from [2] using the following navigation metrics: Success Rate (SR), which measures whether the agent reaches the target location within a set distance threshold; Success Rate Weighted by Path Length (SPL), which is the SR adjusted by the ratio of the ground truth path length to the actual path traveled; Goal Progress (GP), the advancement in meters towards the goal. GP is utilized for the CVDN dataset, whereas SR and SPL are the metrics for other datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 493, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 493, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 493, + 432 + ], + "type": "text", + "content": "5.1. Comparison on Supervised Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 520 + ], + "type": "text", + "content": "As shown in Table 1, we performed a one-time fine-tuning on the four tasks in a fully supervised manner. To begin, our experiments reiterate the superiority of multitask training over single-task training. Also, incorporating our RoomTour3D data into the pre-training process led to consistent improvements across all metrics on Val-U, achieving state-of-the-art results in the GP metric in the CVDN dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "Notably, finetuning with our action-enriched data results in state-of-the-art performance on both Val-U and Test sets across SOON, R2R and REVERIE tasks. While the improvement on the CVDN and SOON datasets is modest, the most significant boost compared to the reproduced baseline is observed in R2R Val-U and REVERIE Val-U, with gains of approximately " + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": ", respectively. The improvement in R2R is largely driven by enhanced spatial awareness, stemming from the inclusion of proximity data, which helps the model better understand object distance and position. Similarly, gains in REVERIE are attributed to a combination of open-vocabulary tags, spatial awareness, and the addition of room type data, which encourages the model to infer the layout of environments, thereby boosting its spatial reasoning capabilities. Moreover, our use of open-ended instructions allows the model to adapt flexibly to diverse" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27591" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 107, + 528, + 310 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Overall comparison with the baseline methods. Our RoomTour3D data can boost NaviLLM by a margin on SOON, R2R and REVERIE on SPL metric and on CVDN GP metric. *denotes reproduced results. RT3DDesc and RT3DAction stand for description-enriched trajectories only and action-enriched trajectories." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 107, + 528, + 310 + ], + "lines": [ + { + "bbox": [ + 82, + 107, + 528, + 310 + ], + "spans": [ + { + "bbox": [ + 82, + 107, + 528, + 310 + ], + "type": "table", + "html": "
MethodsCVDNSOONR2RREVERIE
Val-UTestVal-UTestVal-UTestVal-UTest
Models Focusing on Single Task
PREVALENT [16]3.152.44--5351--
HOP [43]4.413.24--575926.124.3
HAMT [6]5.135.58--616030.226.7
DUET [7]--22.621.4605833.736.0
VLN-SIG [25]5.525.83--6260--
VLN-PETL [44]5.696.13--605827.726.7
NavGPT2 [65]----6160--
BEV-BERT [1]----646036.436.4
Unified Model For All Tasks
NaviLLM(w. Pretrain) [63]6.167.9029.226.3596035.732.3
NaviLLM(w. Pretrain)*6.09-28.0-56.7-31.4-
NaviLLM+RT3DDesc(Ours)6.967.5530.226.562.361.837.135.1
NaviLLM+RT3DAction(Ours)6.337.2231.727.862.462.237.436.4
", + "image_path": "f043ae34705e56b2392b05c6876deea666c3c6f6d6fc36fb6b537f5fbf05a2b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 77, + 331, + 531, + 411 + ], + "blocks": [ + { + "bbox": [ + 156, + 317, + 452, + 328 + ], + "lines": [ + { + "bbox": [ + 156, + 317, + 452, + 328 + ], + "spans": [ + { + "bbox": [ + 156, + 317, + 452, + 328 + ], + "type": "text", + "content": "Table 2. Ablation study on the input modalities for trajectory summarization task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 331, + 531, + 411 + ], + "lines": [ + { + "bbox": [ + 77, + 331, + 531, + 411 + ], + "spans": [ + { + "bbox": [ + 77, + 331, + 531, + 411 + ], + "type": "table", + "html": "
Object tagsDepth & Bounding BoxRoom typeCVDN GP↑SOONR2RREVERIE
SR↑SPL↑SR↑SPL↑SR↑SPL↑
×××6.0933.6428.0165.5256.6738.3231.35
✓××5.4132.5226.5163.6155.7642.5234.37
✓✓×6.4937.6230.4068.3761.7041.7236.04
✓✓✓6.9638.8030.2169.3762.2843.2537.10
", + "image_path": "9e06ffe75a18e4dc9dcc5d65649ca41ab29924edaf7693e3a407c83df405029a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 75, + 462, + 272, + 659 + ], + "blocks": [ + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "type": "text", + "content": "Table 3. Overall comparison with SOTA zero-shot methods on R2R. " + }, + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "type": "text", + "content": " denotes training exclusive navigable actions. " + }, + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "type": "inline_equation", + "content": "{}^{ * }" + }, + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "type": "text", + "content": " denotes using 36 views setting. RT3D stands for our RoomTour3D." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 462, + 272, + 659 + ], + "lines": [ + { + "bbox": [ + 75, + 462, + 272, + 659 + ], + "spans": [ + { + "bbox": [ + 75, + 462, + 272, + 659 + ], + "type": "table", + "html": "
MethodsVal Unseen
SR↑SPL↑
Random Walk [41]32
Commercial Model
NavGPT(GPT-3.5) [64]*13.899.12
NavGPT(GPT-4) [64]3429
MapGPT(GPT-4) [5]38.825.8
MapGPT(GPT-4V) [5]43.734.8
DiscussNav(GPT-4) [39]4340
Open-source Model
LangNav(LLaMA2-7B) [41]00
NavCoT(LLaMA2-7B) [33]7.786.50
DuET (Init. LXMERT [48])10
NaviLLM [63]†00
NaviLLM+RT3D(Ours)14.3310.86
", + "image_path": "65ceaa449853a6aa73f6f8d14f01bb60e56ff7331609c2e450b02b7c5805a808.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "scenarios, fostering more robust and generalizable performance and better contextual understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 422, + 482, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 482, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 482, + 435 + ], + "type": "text", + "content": "5.2. Comparison on Zero-shot Task" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 489 + ], + "type": "text", + "content": "To further demonstrate the substantial indoor knowledge contained in our data and its effectiveness for embodied action and language instructions, we conduct zero-shot experiments on embodied action prediction, as shown in Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "content": "We removed all action and geometric data from the training datasets and retrained NaviLLM with and without our RoomTour3D dataset. Without action prediction data, NaviLLM lacked the ability to learn effective navigable action selection. However, with the inclusion of our action-enriched trajectories, NaviLLM achieved a " + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "inline_equation", + "content": "14.33\\%" + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "content": " SR and a " + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "inline_equation", + "content": "10.86\\%" + }, + { + "bbox": [ + 313, + 491, + 555, + 658 + ], + "type": "text", + "content": " SPL, outperforming open-source models built on LLaMA-7B and reaching results comparable to NavGPT [64], which leverages GPT-3.5. These improvements validate the effectiveness of our 3D trajectories mined from room tour video reconstructions and emphasize the value of our action-enriched trajectories. This highlights the significant contribution of our dataset to advancing open-world navigation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 670, + 406, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 406, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 406, + 683 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Effect of open-world semantics and spatial awareness. As shown in Table 2, we analyzed the impact of various in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27592" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 80, + 72, + 530, + 90 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 530, + 90 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 530, + 90 + ], + "type": "text", + "content": "Instruction: Exit the sewing room. Turn right. Go toward the glass cabinet with the dolls in it. Turn into the doorway on the left. Pass the bed and go through the next doorway on the left into the bathroom. Wait by the sink. (Instruction_id : 4676_0)" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 91, + 336, + 256 + ], + "blocks": [ + { + "bbox": [ + 82, + 91, + 336, + 256 + ], + "lines": [ + { + "bbox": [ + 82, + 91, + 336, + 256 + ], + "spans": [ + { + "bbox": [ + 82, + 91, + 336, + 256 + ], + "type": "image", + "image_path": "f9bf6007d195df537182c3350ca85990d0a7ef379298454e9619c57b7b63bba6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "type": "text", + "content": "Figure 4. Paths of NaviLLM [63] and ours on R2R-unseen. Purple and green circles denote the start and target locations, respectively, and the red circle represents incorrect endpoint. According to the instruction, the agent should turn left at the waypoint marked with yellow. Our method makes the correct decision, while the baseline is confused by similar entrance at the waypoint, thus mistakenly turns right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 341, + 91, + 529, + 257 + ], + "blocks": [ + { + "bbox": [ + 341, + 91, + 529, + 257 + ], + "lines": [ + { + "bbox": [ + 341, + 91, + 529, + 257 + ], + "spans": [ + { + "bbox": [ + 341, + 91, + 529, + 257 + ], + "type": "image", + "image_path": "b08e8dfcb8e39667396b9957e364d71711d36f4edfc1075952c3cd28f3deea88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "text", + "content": "formation types on instruction generation. Adding object variety significantly improved performance on REVERIE with increase SPL from " + }, + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "inline_equation", + "content": "31.35\\%" + }, + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "inline_equation", + "content": "34.37\\%" + }, + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "text", + "content": ", as this dataset relies on object grounding. However, it had no direct impact on SOON, possibly because SOON relies solely on detailed textual descriptions without explicit bounding box annotations. After introducing depth estimation, which helps determine the relative distances of objects, the performance on SOON, R2R and REVERIE achieve marginal boosts. This demonstrates that enhancing spatial awareness significantly contributes to indoor navigation tasks. Furthermore, incorporating room locations, which capture the scene semantics along the trajectory, provided a moderate boost across all four VLN tasks. This further highlights the critical role of object variety and spatial awareness in improving navigation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 606 + ], + "type": "text", + "content": "Effect of action-instruction data. As shown in Table 1, we test the effect of incorporating video-action-instruction data into the training dataset. It is evident that this approach improves the SPL metric across the test scenes of SOON, R2R, and REVERIE. We believe that incorporating geometric information and movement-inclusive instructions helps the model better align the relationship between action and observation changes, thereby further enhancing the model's embodied capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 613, + 212, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 212, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 212, + 624 + ], + "type": "text", + "content": "5.4. Data correctness verification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "We evaluated the correctness of our automated data-generation pipeline by manually rating 100 randomly sampled trajectory descriptions on a 4-point relevance scale: 1 for \"totally irrelevant\", 2 for \"partially relevant\", 3 for \"mostly relevant\" and 4 for a \"perfect match\". The evaluation yielded an average score of 3.08, with " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "74\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " of descriptions rated as \"mostly relevant\" or \"perfect match,\" demon-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "text", + "content": "strating the method's effectiveness in generating meaningful, visually aligned descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 341, + 477, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 477, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 477, + 354 + ], + "type": "text", + "content": "5.5. Navigation Case Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 359, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 556, + 525 + ], + "type": "text", + "content": "As shown in Figure 4, selecting the correct action is critical at specific decision points, such as when a left turn is required to follow the instruction accurately. In the example, at step " + }, + { + "bbox": [ + 313, + 359, + 556, + 525 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 313, + 359, + 556, + 525 + ], + "type": "text", + "content": ", both the rooms to the left and right could satisfy the latter part of the instruction, \"pass the bed and go into the bathroom.\" However, the baseline method incorrectly chooses a right turn at the designated left-turn point, causing it to deviate from the intended path. Once this error occurs, even with scene graph history, the model struggles to realign with the correct trajectory. This challenge is particularly common in household environments, where bedroom layouts often appear similar. It further demonstrates the effectiveness of our data alignment in improving adherence to action-based instructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 540, + 388, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 388, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 388, + 552 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 559, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 556, + 704 + ], + "type": "text", + "content": "In this paper, we present RoomTour3D, a novel dataset automatically curated from room tour videos for VLN tasks. By leveraging the rich, sequential nature of video data and incorporating object variety and spatial awareness, we generate 200K navigation instructions and 17K action-enriched trajectories from 1847 room tour scenes. Additionally, we produce navigable trajectories from video frames and reconstructed 3D scenes, which significantly boost the performance and set new state-of-the-art results on the SOON and REVERIE benchmarks. This approach also enables the development of a trainable zero-shot navigation agent, demonstrating the effectiveness and scalability of RoomTour3D in advancing VLN research." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27593" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Dong An, Yuankai Qi, Yangguang Li, Yan Huang, Liang Wang, Tieniu Tan, and Jing Shao. BEVbert: Multimodal map pre-training for language-guided navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2737-2748, 2023. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "type": "text", + "content": "[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sunderhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments, 2018. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 258 + ], + "type": "text", + "content": "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. ScanQA: 3D question answering for spatial scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 295, + 315 + ], + "type": "text", + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 295, + 360 + ], + "type": "text", + "content": "[5] Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee K Wong. Mapgpt: Map-guided prompting for unified vision-and-language navigation. arXiv preprint arXiv:2401.07314, 2024. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 295, + 395 + ], + "type": "text", + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. In NeurIPS, 2021. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "type": "text", + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation, 2022. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 483 + ], + "type": "text", + "content": "[8] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Learning from unlabeled 3d environments for vision-and-language navigation, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "text", + "content": "[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 576 + ], + "type": "text", + "content": "[10] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "text", + "content": "[11] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, page 226-231. AAAI Press, 1996. 5, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Daniel Fried, Ronghang Hu, Volkan Cirik, Anna Rohrbach, Jacob Andreas, Louis-Philippe Morency, Taylor Berg-Kirkpatrick, Kate Saenko, Dan Klein, and Trevor Darrell. Speaker-follower models for vision-and-language navigation, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Tsu-Jui Fu, Xin Eric Wang, Matthew Peterson, Scott Grafton, Miguel Eckstein, and William Yang Wang. Coun" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "terfactual vision-and-language navigation via adversarial path sampling, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 160 + ], + "type": "text", + "content": "[14] Chen Gao, Xingyu Peng, Mi Yan, He Wang, Lirong Yang, Haibing Ren, Hongsheng Li, and Si Liu. Adaptive zone-aware hierarchical planner for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14911-14920, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 196 + ], + "type": "text", + "content": "[15] Pierre-Louis Guhur, Makarand Tapaswi, Shizhe Chen, Ivan Laptev, and Cordelia Schmid. Airbert: In-domain pretraining for vision-and-language navigation, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 555, + 252 + ], + "type": "text", + "content": "[16] Weituo Hao, Chunyuan Li, Xiujun Li, Lawrence Carin, and Jianfeng Gao. Towards learning a generic agent for vision-and-language navigation via pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13137-13146, 2020. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "text", + "content": "[17] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. A recurrent vision-and-language BERT for navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1643–1653, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 342 + ], + "type": "text", + "content": "[18] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3D-LLM: Injecting the 3d world into large language models, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "text", + "content": "[19] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6683-6693, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "text", + "content": "[20] Minyoung Hwang, Jaeyeon Jeong, Minsoo Kim, Yoonseon Oh, and Songhwai Oh. Meta-exlore: Exploratory hierarchical vision-and-language navigation using scene object spectrum grounding, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "text", + "content": "[21] Vihan Jain, Gabriel Magalhaes, Alexander Ku, Ashish Vaswani, Eugene Ie, and Jason Baldridge. Stay on the path: Instruction fidelity in vision-and-language navigation. arXiv preprint arXiv:1905.12255, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[22] Aishwarya Kamath, Peter Anderson, Su Wang, Jing Yu Koh, Alexander Ku, Austin Waters, Yinfei Yang, Jason Baldridge, and Zarana Parekh. A new path: Scaling vision-and-language navigation with synthetic instructions and imitation learning, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 556, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 589 + ], + "type": "text", + "content": "[23] Jing Yu Koh, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Pathdreamer: A world model for indoor navigation, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "type": "text", + "content": "[24] Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-across-room: Multilingual vision-and-language navigation with dense spatiotemporal grounding. arXiv preprint arXiv:2010.07954, 2020. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[25] Jialu Li and Mohit Bansal. Improving vision-and-language navigation by generating future-view image semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10803-10812, 2023. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27594" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "navigation. Advances in Neural Information Processing Systems, 36:21878-21894, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[27] Jialu Li, Hao Tan, and Mohit Bansal. Improving cross-modal alignment in vision language navigation via syntactic information, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 131, + 294, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 294, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 294, + 153 + ], + "type": "text", + "content": "[28] Jialu Li, Hao Tan, and Mohit Bansal. EnvEdit: Environment editing for vision-and-language navigation, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 155, + 294, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 155, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 294, + 197 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 200, + 294, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 294, + 233 + ], + "type": "text", + "content": "[30] Xiangyang Li, Zihan Wang, Jiahao Yang, Yaowei Wang, and Shuqiang Jiang. KERM: Knowledge enhanced reasoning for vision-and-language navigation, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 235, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 235, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 294, + 268 + ], + "type": "text", + "content": "[31] Xiwen Liang, Fengda Zhu, Lingling Li, Hang Xu, and Xiaodan Liang. Visual-language navigation pretraining via prompt-based environmental self-exploration, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 270, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 294, + 313 + ], + "type": "text", + "content": "[32] Xiwen Liang, Liang Ma, Shanshan Guo, Jianhua Han, Hang Xu, Shikui Ma, and Xiaodan Liang. CorNav: Autonomous agent with self-corrected planning for zero-shot vision-and-language navigation, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 315, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 369 + ], + "type": "text", + "content": "[33] Bingqian Lin, Yunshuang Nie, Ziming Wei, Jiaqi Chen, Shikui Ma, Jianhua Han, Hang Xu, Xiaojun Chang, and Xiaodan Liang. NavCoT: Boosting llm-based vision-and-language navigation via learning disentangled reasoning. arXiv preprint arXiv:2403.07376, 2024. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "text", + "content": "[34] Kunyang Lin, Peihao Chen, Diwei Huang, Thomas H. Li, Mingkui Tan, and Chuang Gan. Learning vision-and-language navigation from youtube videos, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 406, + 294, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 438 + ], + "type": "text", + "content": "[35] Chong Liu, Fengda Zhu, Xiaojun Chang, Xiaodan Liang, Zongyuan Ge, and Yi-Dong Shen. Vision-language navigation with random environmental mixup, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 440, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 294, + 462 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "type": "text", + "content": "[37] Rui Liu, Wenguan Wang, and Yi Yang. Volumetric environment representation for vision-language navigation. In CVPR, pages 16317-16328, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "text", + "content": "[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding Dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "text", + "content": "[39] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. arXiv preprint arXiv:2309.11382, 2023. 2, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "type": "text", + "content": "[40] OpenAI. GPT-4: Generative pre-trained transformer 4. OpenAI API, 2022. 2, 3, 4, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[41] Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, and Yoon Kim. LangNav: Language as a perceptual representation for navigation. arXiv preprint arXiv:2310.07889, 2023. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[42] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. REVERIE: Remote embodied visual referring expression in real indoor environments, 2020. 1, 2, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[43] Yanyuan Qiao, Yuankai Qi, Yicong Hong, Zheng Yu, Peng Wang, and Qi Wu. HOP: History-and-order aware pretraining for vision-and-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15418-15427, 2022. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[44] Yanyuan Qiao, Zheng Yu, and Qi Wu. VLN-PETL: Parameter-efficient transfer learning for vision-and-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15443-15452, 2023. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 185, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 250 + ], + "type": "text", + "content": "[45] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-Matterport 3D dataset (HM3D): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 294 + ], + "type": "text", + "content": "[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 4, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[47] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1, 4, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[48] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[49] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[50] Robert Tarjan. Depth-first search and linear graph algorithms. In 12th Annual Symposium on Switching and Automata Theory (swat 1971), pages 114-121, 1971. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "text", + "content": "[51] Jesse Thomason, Michael Murray, Maya Cakmak, and Luke Zettlemoyer. Vision-and-dialog navigation, 2019. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 511 + ], + "type": "text", + "content": "[52] Liuyi Wang, Zongtao He, Ronghao Dang, Mengjiao Shen, Chengju Liu, and Qijun Chen. Vision-and-language navigation via causal learning. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 567 + ], + "type": "text", + "content": "[53] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 568, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 634 + ], + "type": "text", + "content": "[54] Xin Eric Wang, Vihan Jain, Eugene Ie, William Yang Wang, Zornitsa Kozareva, and Sujith Ravi. Environment-agnostic multitask learning for natural language grounded navigation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIV 16, pages 413–430. Springer, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[55] Zun Wang, Jialu Li, Yicong Hong, Yi Wang, Qi Wu, Mohit Bansal, Stephen Gould, Hao Tan, and Yu Qiao. Scaling data generation in vision-and-language navigation. In ICCV 2023, 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[56] Erik Wijmans, Samyak Datta, Oleksandr Maksymets, Abhishek Das, Georgia Gkioxari, Stefan Lee, Irfan Essa, Devi Parikh, and Dhruv Batra. Embodied question answering" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27595" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 555 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "in photorealistic environments with point cloud perception, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[57] Wikipedia. Preferred walking speed. https://en.wikipedia.org/wiki/Preferred_walking_speed.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 184 + ], + "type": "text", + "content": "[58] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "type": "text", + "content": "[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "text", + "content": "[60] Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and Wang He. NaVid: Video-based vlm plans the next step for vision-and-language navigation. arXiv preprint arXiv:2402.15852, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "text", + "content": "[61] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize Anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 385 + ], + "type": "text", + "content": "[62] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 419 + ], + "type": "text", + "content": "[63] Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. Towards learning a generalist model for embodied navigation, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 295, + 464 + ], + "type": "text", + "content": "[64] Gengze Zhou, Yicong Hong, and Qi Wu. NavGPT: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 518 + ], + "type": "text", + "content": "[65] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. Navigpt-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 295, + 555 + ], + "type": "text", + "content": "[66] Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. SOON: Scenario oriented object navigation with graph-based exploration, 2021. 1, 2, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27596" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_content_list.json b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3fba5d6c28b11202a7636a0532a9b596a39a3eb7 --- /dev/null +++ b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_content_list.json @@ -0,0 +1,2074 @@ +[ + { + "type": "text", + "text": "Rotation-Equivariant Self-Supervised Method in Image Denoising", + "text_level": 1, + "bbox": [ + 166, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanze Liu1 Jiahong Fu1 Qi Xie1,† Deyu Meng1,2,3", + "bbox": [ + 233, + 178, + 763, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong University, Xi'an, China", + "bbox": [ + 333, + 198, + 661, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Pengcheng Laboratory, Shenzhen, China", + "bbox": [ + 331, + 215, + 665, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Macau University of Science and Technology, Taipa, Macao", + "bbox": [ + 254, + 233, + 741, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised image denoising methods have garnered significant research attention in recent years, for this kind of method reduces the requirement of large training datasets. Compared to supervised methods, self-supervised methods rely more on the prior embedded in deep networks themselves. As a result, most of the self-supervised methods are designed with Convolution Neural Networks (CNNs) architectures, which well capture one of the most important image prior, translation equivariant prior. Inspired by the great success achieved by the introduction of translational equivariance, in this paper, we explore the way to further incorporate another important image prior. Specifically, we first apply high-accuracy rotation equivariant convolution to self-supervised image denoising. Through rigorous theoretical analysis, we have proved that simply replacing all the convolution layers with rotation equivariant convolution layers would modify the network into its rotation equivariant version. To the best of our knowledge, this is the first time that rotation equivariant image prior is introduced to self-supervised image denoising at the network architecture level with a comprehensive theoretical analysis of equivariance errors, which offers a new perspective to the field of self-supervised image denoising. Moreover, to further improve the performance, we design a new mask mechanism to fusion the output of rotation equivariant network and vanilla CNN-based network, and construct an adaptive rotation equivariant framework. Through extensive experiments on three typical methods, we have demonstrated the effectiveness of the proposed method. The code is available at: https://github.com/liuhanze623/AdaReNet.", + "bbox": [ + 89, + 316, + 483, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "During the process of capturing and transmitting images, unexpected noises are frequently introduced [16, 28]. Such noise can severely degrade image quality and disrupt subsequent image processing tasks. Image denoising is a tech-", + "bbox": [ + 89, + 814, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/031d15c4d8f19aae03236bff2d1623d9e7b08992e58ce746422a9b091da14bed.jpg", + "image_caption": [ + "(a) Input image" + ], + "image_footnote": [], + "bbox": [ + 519, + 287, + 645, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82dee4da83fe713d603bbe8467bb6e8460ead0480fdf9a923f01e053748543e7.jpg", + "image_caption": [ + "(b) CNN output" + ], + "image_footnote": [], + "bbox": [ + 647, + 289, + 772, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a6e72eba01528aa8564a56f09f52d5397e01e56f6fc3a12ca51c691363bb8afd.jpg", + "image_caption": [ + "(c) Equ. conv. output", + "Figure 1. Illustration of the output feature map of a typical image obtained by standard CNN and our used rotation equivariant convolution neural network. Both networks are initialized randomly." + ], + "image_footnote": [], + "bbox": [ + 774, + 289, + 900, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nique designed to address the standard inverse problem in image processing and is widely utilized across various fields. With the rapid advancement of deep learning (DL), denoisers based on different deep networks have achieved outstanding results.", + "bbox": [ + 511, + 463, + 906, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early, it was believed that the success of deep learning is primarily due to the exhaustive utilization of training data. Thus most early DL-based image denoising are in supervised manner [2, 7, 14, 20, 34, 42, 56, 57], where models are trained on extensive datasets that include pairs of clean and noisy images, thereby learning the transformation from noisy to clean images. However, in reality, compiling such comprehensive datasets is both costly and time-intensive, posing substantial challenges [1, 35, 50].", + "bbox": [ + 511, + 546, + 908, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised approaches thus achieved significant research attention in recent years. These approaches rely less on extensive supervised datasets, but pay more attention to the full utilization of the prior information inherent in deep networks [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Notably, Lehtinen et al. introduced the innovative self-supervised learning approach called Noise2Noise [26], which enables training on pairs of noisy images depicting the identical scene. [26] was the first self-supervised algorithm that achieved performance comparable to supervised image denoising. Other methods involve training denoising models on single noisy images by designing blind-spot networks [3, 21, 22, 48], while some approaches devise strategies to generate training image pairs from noisy im", + "bbox": [ + 511, + 689, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 886, + 245, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12720", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ages [6, 18, 32, 33, 51] and have achieved excellent results.", + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nowadays, it has become common sense that the prior information inherent in deep networks plays an important role in improving the performance of DL-based methods in image processing tasks [11, 39, 46, 49]. The most classic example should be the CNN, which well captures the translation-equivariant prior of natural images. In CNNs, shifting an input image of CNN is equivalent to shifting all of its intermediate feature maps and the output image. Compared with fully-connected neural networks, this translational equivariance property brings in rational weight sharing, makes the network parameters being used more efficiently, and thus leads to substantially better performance. Inspired by the significant success brought about by the introduction of translational equivariance, another essential image prior (As illustrated in Fig. 1), rotation-equivariant prior has to be taken into consideration very recently. Rotation-equivariant CNNs have been applied and achieved notable performance improvement in multiple supervised image processing tasks [11, 49].", + "bbox": [ + 88, + 108, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared with supervised learning, self-supervised learning approaches indeed depend more on the prior knowledge embedded within the network, for the information that can be learned from the dataset is less. As a result, most current self-supervised image denoising methods are all based on the CNN architectures, which is largely due to the reliance on the prior of translational equivariance. Therefore, it holds even greater significance to incorporate rotational equivariant design into the deep networks for self-supervised methods.", + "bbox": [ + 89, + 397, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there are two critical issues that need to be addressed when introducing equivariant priors into self-supervised image denoising. On the one hand, although, Xie et al. have recently constructed a rotation equivariant CNN suitable for image processing [49] and Fu et al. further analyzed the global rotation equivariance error of the entire network [11], current rotation equivariant designs for image processing are usually in ResNet[15] structure without upsampling or downsampling. For example, for the superresolution network in [49] where upsampling module is inevitable, only the portion of the network before upsampling is rotation equivariant and the equivariance for the entire network can not be guaranteed. Since self-supervised networks often utilize U-Net structures, which include multiple upsampling and downsampling modules, it is imperative to first explore the impact of these modules on rotation equivariance of the network.", + "bbox": [ + 89, + 550, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, rotation equivariant design for deep networks inevitably introduces parameter sharing and convolution kernel parameterization, which usually degrades the representation accuracy of the network. However, image denoising tasks often have high requirements for the network's representation accuracy, since important high", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "frequency components need to be reconstructed. Besides, not all areas of natural images strictly comply with rigid rotation equivariance, indiscriminately adopting a rotation equivariant network for the entire image can often be detrimental to the reconstruction performance of images.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This study primarily focuses on integrating rotation equivariant prior to existing self-supervised denoising techniques while solving the aforementioned issues. The key contributions can be summarized as follows:", + "bbox": [ + 511, + 167, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We explore the way for introducing rotation equivariant prior into self-supervised image denoising frameworks at the network architecture level. Particularly, for the first time, we rigorously analyze the impact of upsampling and downsampling on equivariant networks theoretically. The equivariant errors of upsampling and downsampling layer indeed approach zero when the resolution of the input image increases, though the approach rate $(O(h), h$ denoted the mesh size of image) is slower than equivariant convolutional layer (whose equivariant error approaches zero at a rate of $O(h^{2})$ ). Then, by taking U-Net as an example, we further analyze the rotation equivariant error of the entire network, showing that by simply replacing all the convolution layers with rotation equivariant convolutions [49], we can indeed achieve a reliable rotation equivariant network for self-supervised image denoising.", + "bbox": [ + 511, + 229, + 906, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We have further developed an adaptive rotation equivariant network to enhance the representation accuracy. Specifically, we design a fusion module for integrating the advantages of both rotation-equivariant and nonequivariant networks. The module can automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. This design provides greater flexibility and yields improved denoising results.", + "- We conducted comprehensive experiments across various self-supervised denoising methods, demonstrating the effectiveness of integrating rotation-equivariant image priors into neural networks for self-supervised techniques. Our approach provides a novel perspective in the field of self-supervised image denoising." + ], + "bbox": [ + 511, + 472, + 905, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work and Prior Knowledge", + "text_level": 1, + "bbox": [ + 511, + 715, + 838, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Denoising", + "text_level": 1, + "bbox": [ + 511, + 741, + 679, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-learning Denoising Methods: Conventional denoising methods predominantly depend on the statistical characteristics of images and mathematical modeling, often employing image priors instead of learned denoisers. Notably, NLM [4, 5] and BM3D [9, 29] have been proposed that can effectively remove noise based on the exploitation of image self-similarity. WNNM [12, 13] treated image denoising as a low-rank matrix approximation problem and achieved great performance.", + "bbox": [ + 511, + 763, + 905, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12721", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Denoising Methods: The majority of DL-based denoising algorithms are supervised [2, 7, 14, 52, 56, 57]. DnCNN [56] employed an end-to-end training strategy to learn the mapping between noisy images and the residual components, markedly improving denoising efficacy. FFDNet [57] proposed a versatile denoising architecture by incorporating the noise level as a network parameter. Nevertheless, to train a proficient model, an extensive collection of paired noisy and clean images is necessary to comprehensively capture the range of image contents and noise variations [2, 14, 19, 25, 37, 53]. The acquisition of such training datasets can be prohibitively costly and challenging, and in the case of certain medical images, it may be virtually unattainable.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised Denoising Methods: Self-supervised deep learning methods have garnered significant interest due to their independence from clean reference images [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Noise2Noise [26], proposed by Lehtinen et al., was the first self-supervised method that achieved performance comparable to supervised methods only using paired noisy images. Noise2Void [21] and Noise2Self [3] used a single noisy image and employed a blind-spot network to predict the clean pixel values based on neighboring pixels, thereby avoiding collapse to an identity mapping. [6, 18, 32, 33, 51] devise strategies to generate training image pairs from noisy images to train a network. R2R [33] has demonstrated that the cost function defined on the noisy/noisy image pairs generated by this method is statistically equivalent to the supervised method. Single-image denoising methods [36, 41] capitalize on the statistical properties of the image itself, showing powerful denoising capabilities without extensive training datasets. These methods lack clean images for supervision and rely more on the prior knowledge embedded within the network, which is the reason for the success of these CNN-based self-supervised image denoising methods. Inspired by this, we explore the way to further incorporate rotation equivariant image prior to these methods with rigorous theoretical analysis.", + "bbox": [ + 91, + 303, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Rotation Equivariance", + "text_level": 1, + "bbox": [ + 89, + 679, + 302, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equivariant Convolution Neural Networks (ECNNs) have drawn substantial interest within the field of computer vision these years [8, 17, 30, 39, 40, 45-47, 58]. Their key strength stems from their ability to effectively handle image rotations through architectural design, which significantly enhances the model's generalization and robustness.", + "bbox": [ + 89, + 702, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cohen et al. introduced Group Equivariant Convolutional Networks (G-CNNs) [8] and first integrated $\\frac{\\pi}{2}$ degree rotation equivariance into the neural network. Recently, the filter parametrization technique has been widely employed and Weiler et al. used harmonics as steerable filters to achieve exact equivariance [45, 46]. However, a notable limitation of these filter parameterization techniques", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is the inaccuracy in their representation, which significantly affects low-level vision tasks. Fortunately, Xie et al. addressed this issue by proposing Fourier series expansion-based filter parametrization, which has relatively high expression accuracy [49]. The proposed Fconv exhibits precise equivariance in the continuous domain, degrading to approximation only after discretization.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equivariant convolutions, distinct from data augmentation techniques, inherently integrate rotational symmetry image priors into the network architecture, thereby guaranteeing the network's inherent equivariance and offering superior interpretability and generalizability. However, in the U-Net structure, there is no theoretical guarantee for the design of rotation equivariance. To the best of our knowledge, this study represents the first application of rotation equivariant convolution in the field of self-supervised image denoising and offers theoretical assurance of equivariance for networks based on the U-Net architecture.", + "bbox": [ + 511, + 198, + 906, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Prior Knowledge about Equivariance", + "text_level": 1, + "bbox": [ + 511, + 369, + 836, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equivalence to a transformation indicates that the application of a transformation to the input results in a corresponding, predictable transformation of the output [39, 45, 49]. Concretely, consider a mapping $\\Psi$ that transforms the input feature space to the output feature space, and let $G$ denote a set of transformations. For any $g\\in G$ , the following relationship holds:", + "bbox": [ + 511, + 392, + 905, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left[ \\pi_ {g} [ f ] \\right] = \\pi_ {g} ^ {\\prime} [ \\Psi [ f ] ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 500, + 903, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ represents any feature map within the input feature space, and $\\pi_g$ and $\\pi_g'$ describe the action of the transformation $g$ on the input and output features, respectively.", + "bbox": [ + 511, + 523, + 905, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 582, + 687, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. ECNNs for Self-supervised Image Denoising", + "text_level": 1, + "bbox": [ + 511, + 606, + 888, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised image denoising methods lack clean images for supervision and rely more on the prior knowledge embedded within the network. Therefore, introducing rotation equivariant image prior to this field is reasonable, which can be achieved by ECNNs [49]. The U-Net can effectively restore high-frequency details in images with an encoder-decoder architecture, making it widely used in this task. Consequently, to construct an equivariant self-supervised denoising network, we need to discuss the impact of the essential upsampling and downsampling operators in the U-Net network on equivariance.", + "bbox": [ + 511, + 628, + 905, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the subsequent section, we first present some notations and concepts, and then theoretically analyze the impact of upsampling and downsampling operators on equivariant networks. Furthermore, we deduce the equivariant error for the complete U-Net network and provide theoretical guarantees for the implementation of equivariance in U-Net architectures.", + "bbox": [ + 511, + 796, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12722", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0bb10b52cb4f3715d20db0c3873d303d685ca87942634912483bed24003a532f.jpg", + "image_caption": [ + "Figure 2. The network architecture of the equivariant N2N method. The network can be divided into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1. Notations and Concepts", + "text_level": 1, + "bbox": [ + 89, + 311, + 303, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first introduce some necessary notations and preliminaries as follows.", + "bbox": [ + 89, + 330, + 482, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider the equivariance on the orthogonal group $O(2)$ . Formally, $O(2) = \\{A \\in \\mathbb{R}^{2 \\times 2} | A^T A = I_{2 \\times 2}\\}$ , which contains all rotation and reflection matrices. Without ambiguity, we use $A$ to parameterize $O(2)$ . The Euclidean group $E(2) = \\mathbb{R}^2 \\rtimes O(2)$ ( $\\rtimes$ is a semidirect-product), whose element is represented as $(x, A)$ . Restricting the domain of $A$ and $x$ , we can also use this representation to parameterize any subgroup of $E(2)$ . In practice, the subgroup is usually assumed to contain $t$ rotations with $\\frac{2\\pi}{t}$ degree for an integer $t \\in \\mathbb{N}_+$ .", + "bbox": [ + 89, + 363, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An image $I \\in R^{n \\times n}$ is viewed as a two-dimensional discretization of a smooth function $r: \\mathbb{R}^2 \\to \\mathbb{R}$ , at the cell-center of a regular grid with $n \\times n$ cells, i.e., for $i, j = 1, 2, \\dots, n$ ,", + "bbox": [ + 89, + 513, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i j} = r \\left(x _ {i j}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 574, + 480, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{ij} = \\left(\\left(i - \\frac{n + 1}{2}\\right)h, \\left(j - \\frac{n + 1}{2}\\right)h\\right)^T$ , $h$ is the mesh size.", + "bbox": [ + 89, + 598, + 482, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An intermediate feature map $F \\in \\mathbb{R}^{n \\times n \\times t}$ in equivariant networks is a multi-channel tensor, which can be viewed as the discretization of a continuous function defined on $\\tilde{E} = \\mathbb{R}^2 \\rtimes S$ , where $S$ is a subgroup of $O(2)$ and $t$ is the number of elements in $S$ . Formally, $F$ can be represented as a three-dimensional grid tensor sampled from a smooth function $e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}$ , i.e., for $i, j = 1, 2, \\dots, n$ ,", + "bbox": [ + 89, + 618, + 483, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i j} ^ {\\backslash A} = e (x _ {i j}, A), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 731, + 482, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{ij}$ is defined in (2) and $A \\in S$ .", + "bbox": [ + 94, + 762, + 349, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With above notations, the transformations on the input and feature maps can be mathematically formulated. Specifically, in the continuous domain, for an input $r \\in C^{\\infty}(\\mathbb{R}^{2})$ and feature map $e \\in C^{\\infty}(E(2))$ , the transformation $\\tilde{A} \\in O(2)$ acts on $r$ and $e$ respectively by:", + "bbox": [ + 89, + 776, + 482, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\pi_ {\\tilde {A}} ^ {R} [ r ] (x) = r (\\tilde {A} ^ {- 1} x), \\forall x \\in \\mathbb {R} ^ {2}, \\\\ \\pi_ {\\tilde {A}} ^ {E} [ e ] (x, A) = e \\left(\\tilde {A} ^ {- 1} x, \\tilde {A} ^ {- 1} A\\right), \\forall (x, A) \\in E (2). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 863, + 480, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, if $A_{\\theta} \\in O(2)$ is the rotation matrix $\\left[ \\begin{array}{c}\\cos \\theta ,\\sin \\theta \\\\ -\\sin \\theta ,\\cos \\theta \\end{array} \\right]$ , then the corresponding rotation operators can be expressed by $\\pi_{\\theta}^{R}$ and $\\pi_{\\theta}^{E}$ .", + "bbox": [ + 513, + 90, + 903, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, in the discrete domain, we can also define the transformation $\\tilde{A} \\in S$ on the input image and feature map as followings:", + "bbox": [ + 511, + 154, + 905, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\tilde {\\pi} _ {\\bar {A}} ^ {R} (I)\\right) _ {i j} = \\pi_ {\\bar {A}} ^ {R} [ r ] (x _ {i j}), \\\\ \\left(\\tilde {\\pi} _ {\\tilde {A}} ^ {E} (F)\\right) _ {i j} ^ {A} = \\pi_ {\\tilde {A}} ^ {E} [ e ] \\left(x _ {i j}, A\\right), \\tag {5} \\\\ \\forall i, j = 1, 2, \\dots , n, A \\in S. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 210, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, rotation operators can be denoted as $\\tilde{\\pi}_{\\theta}^{R}$ and $\\tilde{\\pi}_{\\theta}^{E}$ .", + "bbox": [ + 511, + 292, + 901, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Equivalence of Downsampling and Upsampling", + "text_level": 1, + "bbox": [ + 511, + 319, + 897, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the U-Net architecture commonly used in self-supervised image denoising typically integrates upsampling and downsampling operators. In specific applications, these layers will inevitably affect the equivariance of the network. Recent studies [11, 49] have analyzed the equivariance properties of the group convolution layer. However, the effects of upsampling and downsampling on network equivariance remain unexplored.", + "bbox": [ + 511, + 338, + 905, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First of all, We provide the definitions of commonly used downsampling methods in the continuous domain.", + "bbox": [ + 511, + 460, + 905, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Maxpooling Downsampling. Maxpooling is a commonly used downsampling method in CNNs, which reduces the spatial dimensions of feature maps by sliding a fixed-size window over the feature map and selecting the maximum value within each region as the output [23]. In the continuous domain, we can define maxpooling operator $M(\\cdot)$ as follows,", + "bbox": [ + 511, + 491, + 905, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ M (F) ] (x, A) = \\max _ {\\Omega_ {x}} F _ {i j} ^ {A}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 598, + 903, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x = [x_{1},x_{2}]^{T}\\in \\mathbb{R}^{2}$ denotes the spatial coordinates, and $x_{1}\\in [i,i + 1]$ , $x_{2}\\in [j,j + 1]$ , $\\Omega_{x} = \\{(i,j),(i + 1,j),(i,j + 1),(i + 1,j + 1)\\}$ .", + "bbox": [ + 511, + 626, + 903, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stride Downsampling. Stride Downsampling is also a widely used downsampling operator which reduces the size of the feature map by adjusting the stride of the convolution operation [24]. In the continuous domain, we can define stride downsampling operator $S(\\cdot)$ as follows,", + "bbox": [ + 511, + 672, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ S (F) ] (x, A) = F _ {i, j + 1} ^ {A}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 761, + 903, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x = [x_{1},x_{2}]^{T}\\in \\mathbb{R}^{2}$ denotes the spatial coordinates, and $x_{1}\\in [i,i + 1],x_{2}\\in [j,j + 1]$ $\\Omega_{x} = \\{(i,j),(i+$ $1,j),(i,j + 1),(i + 1,j + 1)\\} .$", + "bbox": [ + 511, + 792, + 903, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the above two common downsampling operators, it is of great significance to analyze their impact on the equivariant network. Therefore, we constructed their equivariant errors under the rotational equivariant structure.", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12723", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 Assume that a feature map $F \\in \\mathbb{R}^{n \\times n \\times t}$ is discretized from the smooth function $e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}$ , $|S| = t$ , the mesh size is $h$ , $D(\\cdot)$ is the downsampling operator. If for any $A, B \\in S, x \\in \\mathbb{R}^2$ , the following conditions are satisfied:", + "bbox": [ + 88, + 90, + 483, + 165 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\| \\nabla e (x, A) \\| \\leq G, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 166, + 482, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "then the following results are satisfied:", + "bbox": [ + 89, + 186, + 349, + 203 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| D \\left[ \\tilde {\\pi} _ {B} ^ {E} \\right] (F) (x, A) - \\pi_ {B} ^ {E} [ D (F) ] (x, A) \\right| \\leq 2 \\sqrt {2} G h. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 209, + 482, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The downsampling operator $D(\\cdot)$ can be either $M(\\cdot)$ or $S(\\cdot)$ . Theorem 1 reveals that the equivariant error of the downsample operator is primarily influenced by the mesh size $h$ , and it indeed approach zero when the resolution of the input image increases with the approach rate $O(h)$ .", + "bbox": [ + 89, + 236, + 483, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we provide the definitions of commonly used upsampling methods in the continuous domain.", + "bbox": [ + 89, + 311, + 483, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Nearest Neighbor Upsampling. Nearest neighbor interpolation is an image scaling method that fills the pixels of the interpolated image by selecting the original pixel value closest to the target pixel position. In the continuous domain, we can define the nearest neighbor operator $N(\\cdot)$ as follows,", + "bbox": [ + 89, + 342, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ N (F) ] (x, A) = F _ {i ^ {*} j ^ {*}} ^ {A}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 431, + 480, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(i^{*},j^{*}) = arg\\min_{ij}||x_{ij} - x||_2^2$", + "bbox": [ + 89, + 453, + 357, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bilinear Upsampling. Bilinear interpolation calculates the new pixel value by taking the weighted average of the four surrounding known pixel values. In the continuous domain, we can define the bilinear interpolation operator $B(\\cdot)$ as follows,", + "bbox": [ + 89, + 469, + 483, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ B (F) ] (x, A) = \\sum_ {i = 1} ^ {2} \\sum_ {j = 1} ^ {2} \\lambda_ {i j} f \\left(Q _ {i j}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 542, + 482, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{ij}$ are the coefficients of bilinear interpolation and $f(Q_{ij})$ represent the grid points, $x = [x_1,x_2]^T\\in \\mathbb{R}^2$ denotes the 2D spatial coordinates, $x_{1}\\in [i,i + 1]$ , $x_{2}\\in [j,j + 1]$ .", + "bbox": [ + 89, + 588, + 483, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both of the aforementioned upsampling operators are widely utilized across various network architectures, making it essential to analyze their mathematical properties within rotational equivariant networks. Accordingly, we evaluated their equivariant errors under a rotational equivariant framework.", + "bbox": [ + 89, + 648, + 483, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2 Assume that a feature map $F \\in \\mathbb{R}^{n \\times n \\times t}$ is discretized from the smooth function $e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}, |S| = t$ , the mesh size is $h$ , $U(\\cdot)$ is the upsampling operator. If for any $A, B \\in S, x \\in \\mathbb{R}^2$ , the following conditions are satisfied:", + "bbox": [ + 88, + 748, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\| \\nabla e (x, A) \\| \\leq G, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 824, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "then the following results are satisfied:", + "bbox": [ + 89, + 845, + 349, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| U \\left[ \\tilde {\\pi} _ {B} ^ {E} \\right] (F) (x, A) - \\pi_ {B} ^ {E} [ U (F) ] (x, A) \\right| \\leq 2 (\\sqrt {2} + 1) G h. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 868, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The upsampling operator $U(\\cdot)$ can be either $N(\\cdot)$ or $B(\\cdot)$ . It is worth noting that the above conclusions indicate that the error introduced by the upsampling operator in the rotational equivariant network is related to $h$ , aligning with established understanding.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3. Analysis of Complete U-Net Network", + "text_level": 1, + "bbox": [ + 511, + 172, + 818, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We take the network design of the N2N method [26] as an example to give the derivation of the rotation equivariant error for the entire U-Net architecture. As shown in Fig. 2, we decompose the network into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer (Equivariant Convolution) and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers, we provide the equivariant error for each block and subsequently derive the equivariant error bounds for the complete network.", + "bbox": [ + 511, + 190, + 906, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3 demonstrates the equivariance error of the complete U-Net network under discrete angles, and the corollary further provides the equivariance error of the complete U-Net network under any rotation angle. Note that the mesh size of upsampling and downsampling are different, we define the mesh size of the original picture to be $h$ , the mesh size after a $\\times 2$ downsampling is $2h$ , and so on.", + "bbox": [ + 511, + 357, + 905, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 3 For an image $X$ with size $H\\times W\\times n_0$ , and a $N$ -layer rotation equivariant $U$ -Net network $\\mathrm{UNet}_{eq}(\\cdot)$ whose channel number of the $l^{th}$ layer is $n_l$ , rotation equivariant subgroup is $S\\leqslant O(2),|S| = t$ , and activation function is set as ReLU. If the latent continuous function of the $c^{th}$ channel of $X$ denoted as $r_c:\\mathbb{R}^2\\to \\mathbb{R}$ , and the latent continuous function of any convolution filters in the $l^{th}$ layer denoted as $\\phi^l:\\mathbb{R}^2\\to \\mathbb{R}$ , $DB_{i}$ and $UB_{i}$ represent the downsampling block and the upsampling block, respectively. $\\hat{\\Psi}$ , $\\hat{\\Phi}$ , and $\\hat{\\Upsilon}$ represent the convolutional layers in the input, middle, and output stages, respectively. We define:", + "bbox": [ + 511, + 476, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {U N e t} _ {e q} (\\cdot) = \\hat {\\Upsilon} \\left[ \\hat {U} B _ {m} \\dots \\hat {U} B _ {1} \\left[ \\hat {\\Phi} \\left[ \\hat {D} B _ {m} \\dots \\hat {D} B _ {1} \\left[ \\hat {\\Psi} \\right] \\right] \\dots \\right] (\\cdot), \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 647, + 905, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the following conditions are satisfied:", + "bbox": [ + 511, + 684, + 764, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| r _ {c} (x) \\right| \\leq F _ {0}, \\| \\nabla_ {x} r _ {c} (x) \\| \\leq G _ {0}, \\| \\nabla_ {x} ^ {2} r _ {c} (x) \\| \\leq H _ {0}, \\\\ \\left| \\phi^ {l} (x) \\right| \\leq F _ {l}, \\| \\nabla_ {x} \\phi^ {l} (x) \\| \\leq G _ {l}, \\| \\nabla_ {x} ^ {2} \\phi^ {l} (x) \\| \\leq H _ {l}, \\tag {15} \\\\ \\forall \\| x \\| \\geq (p + 1) h / 2, \\phi_ {l} (x) = 0, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 704, + 905, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ is the filter size, $h$ is the mesh size, $\\theta_{k} = \\frac{2k\\pi}{t}$ , $k = 1,2,\\dots ,t$ . $\\nabla_{x}$ and $\\nabla_{x}^{2}$ denote the operators of gradient and Hessian matrix, respectively. We have", + "bbox": [ + 511, + 770, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathrm {U N e t} _ {e q} \\left[ \\tilde {\\pi} _ {\\theta_ {k}} ^ {R} \\right] (X) - \\tilde {\\pi} _ {\\theta_ {k}} ^ {R} \\left[ \\mathrm {U N e t} _ {e q} \\right] (X) \\right| \\leq R _ {1} h + R _ {2} h ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 821, + 903, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{1}, R_{2}$ are two constants with respect to $N, n_{l}$ and the upper bound in (15), their specific values can be found in the supplementary materials.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12724", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca00c297445f151b65f928c399921d23eb8f3c23b76c42722455854486b3af48.jpg", + "image_caption": [ + "Figure 3. Illustrations of our proposed adaptive network AdaReNet. Specifically, $I \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times \\mathrm{C}}$ represents a noisy image, where $H$ and $W$ represent the spatial dimensions, and $C$ denotes the channel dimension. The Vanilla Module and EQ Module each produce their respective preliminary denoising results, denoted as $f_c$ and $f_e$ . The Fusion Module $Mask(\\cdot)$ automatically decides which areas of the image to use more EQ Module would gain more benefit. After adaptive fusion by $Mask(\\cdot)$ and correction by the Self-correcting Module $S_c(\\cdot)$ , the final denoised image $\\bar{I}$ is output." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 908, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Corollary 1 Under the same condition as Theorem 3, for an arbitrary $\\theta \\in [0,2\\pi ]$ , let $\\pi_{\\theta}$ denote the rotation transformation, then $\\forall \\theta$ we have", + "bbox": [ + 89, + 426, + 483, + 472 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathrm {U N e t} _ {e q} \\left[ \\tilde {\\pi} _ {\\theta} ^ {R} \\right] (X) - \\tilde {\\pi} _ {\\theta} ^ {R} \\left[ \\mathrm {U N e t} _ {e q} \\right] (X) \\right| \\leq R _ {1} h + R _ {2} h ^ {2} + R _ {3} t ^ {- 1} h, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 484, + 483, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $R_{1}, R_{2}, R_{3}$ are constants that can be found in the supplementary materials.", + "bbox": [ + 89, + 517, + 483, + 547 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f2c85e6f5a7481384f6cadf7273fa12fe9f52168c394457b3782ee1e9b10cde.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 575, + 191, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/311276b789292fa4a963cd43fc3403acd7f9ba656973b27f4455f5b711df2367.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 191, + 575, + 287, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d197b7e75b11fd15a10152a02f592a8b63f043f80544922bf958c0da424558e8.jpg", + "image_caption": [ + "(c)", + "Figure 4. (a) An image from the Kodak dataset, (b) the heatmap of the low-frequency component, (c) the heatmap of the high-frequency component, (d) the output of our proposed MaskNetwork (the brighter area indicates the use of more Vanilla Module)." + ], + "image_footnote": [], + "bbox": [ + 287, + 575, + 383, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08f1bfb0ff3129c1b59f28fb1aafee7177e32e0251253f1ea3cdc541044854a2.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 383, + 575, + 478, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Proposed Adaptive Network AdaReNet", + "text_level": 1, + "bbox": [ + 89, + 756, + 429, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rotation equivariant design usually degrades the representation accuracy of the network because of parameter sharing and filter parameterization and not all areas of natural images strictly comply with rigid rotation equivariance. In order to solve the above problems, the proposed network predominantly comprises four primary modules, as depicted in Fig. 3, which are the Vanilla Module, EQ Module, Fusion Module, and Self-correcting Module.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vanilla Module. This module utilizes a conventional CNN architecture, maintaining the original design where only translation equivariance is incorporated into the network. We define the result of the input image $I$ after passing through the Vanilla Module as $f_{c}$ .", + "bbox": [ + 511, + 426, + 906, + 503 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {c} = \\operatorname {V M} (I). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 516, + 903, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EQ Module. This module employs the same architecture as the previous one, with the conventional convolution replaced by the rotation equivariant network to incorporate rotation equivariance prior into the architecture. We define the result of input image $I$ after passing through the EQ Module as $f_{e}$ ,", + "bbox": [ + 511, + 544, + 906, + 635 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {e} = \\operatorname {E Q} (I). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 637, + 903, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fusion Module. This module introduces a MaskNetwork $Mask(\\cdot)$ , a specialized network designed to merge the outputs from the Vanilla Module and the EQ Module. It employs several layers of standard convolutions to automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. The output of $Mask(\\cdot)$ is shown in Fig. 4(d). We define the result of input image $I$ after passing through MaskNetwork as $M_{f}$ ,", + "bbox": [ + 511, + 661, + 906, + 797 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM _ {f} = \\operatorname {M a s k} (I). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 810, + 903, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Self-correcting Module. Following the Fusion Module, a self-correcting module is applied to refine the fused result. This is achieved through the use of simple ResNet Blocks. We define the Self-correcting Module as $S_{c}(\\cdot)$ .", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the network presented in Fig. 3, we can derive the following mathematical representation to encapsulate the training and inference processes of the entire network:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} = M _ {f} \\odot f _ {c} + (1 - M _ {f}) \\odot f _ {e}, \\bar {I} = S _ {c} (\\hat {I}), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 165, + 482, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat{I}$ denotes the results of adaptive fusion, and $\\bar{I}$ represents the final restoration result after going through the Self-correcting Module, $\\odot$ denote element-wise multiplication.", + "bbox": [ + 89, + 199, + 483, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Function. We incorporate the loss of the two subnetworks as regularization terms into the main loss. The loss function is defined as follows:", + "bbox": [ + 89, + 262, + 483, + 308 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} = \\| \\bar {I} - \\text {t a r g e t} \\| _ {2} + \\alpha_ {1} \\| f _ {c} - \\text {t a r g e t} \\| _ {2} + \\alpha_ {2} \\| f _ {e} - \\text {t a r g e t} \\| _ {2}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 321, + 483, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha_{1},\\alpha_{2}$ are hyperparameters and we empirically set $\\alpha_{1} = \\alpha_{2} = 0.1$", + "bbox": [ + 89, + 354, + 483, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remark. Embedding rotational equivariance into the network is highly effective for self-supervised image denoising task. Since natural images do not adhere to strict rotational equivariance, our proposed adaptive rotational equivariant network, named AdaReNet, can automatically decide which regions of the image to apply the rotation-equivariant network, thereby further enhancing performance. By observing the output of the Fusion Module in Fig. 4(d), it is noted that the mask values are larger at the edge details of the pattern, indicating that the adaptive equivariant network tends to use the outputs from the Vanilla Module more frequently in these areas. Conversely, for the restoration of the majority of low-frequency components in the image, the network predominantly uses the outputs from the EQ Module. This aligns with the common sense that convolutions are more adept at fitting high-frequency information[43].", + "bbox": [ + 89, + 386, + 483, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 647, + 223, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conducted experiments based on the existing setup and validated our method in three classic approaches. For the Noise2Noise [26] and Noise2Void [21] methods, we showed experiments on U-Net [38], which can significantly speed up the training process while maintaining acceptable performance, and demonstrated the effectiveness of our method. Besides, we also conducted the R2R experiments based on DnCNN [33], further validating the superiority of the proposed method. Due to space limitations, we only present partial results. Implementation details and further experiments concerning different models, various datasets, model parameter counts, as well as experiments in the field of self-supervised fluorescence microscopy denoising[27] can be found in supplementary materials.", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experiments on Compared Methods", + "text_level": 1, + "bbox": [ + 511, + 90, + 830, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rotation Equivariant N2N: We selected the most classical Gaussian noise for our experiments, and randomized the standard deviation $\\sigma \\in [0,50]$ of noise for each training example individually. N2N-EQ denotes the rotation equivariant network, while N2N-EQ $^+$ serves as the adaptive rotation equivariant network. The notation in other experiments is similar. The results are shown in Tab. 1. The superiority of our method can also be observed from Fig. 5.", + "bbox": [ + 511, + 112, + 906, + 233 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ffef4fbc996fce9776ca9a9c3814b4d88e958f1c5c3ef29972d54537cbeb587e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetGaussian25Gaussian50
N2N [26]N2N-EQN2N-EQ+N2N [26]N2N-EQN2N-EQ+
Kodak [10]31.47/0.87431.60/0.87831.72/0.88028.29/0.77828.58/0.79028.69/0.791
BSD300 [31]30.18/0.86930.28/0.87230.36/0.87327.02/0.76227.24/0.77127.31/0.772
Set14 [54]30.02/0.85130.06/0.85430.19/0.85527.16/0.76827.32/0.77527.44/0.777
", + "bbox": [ + 514, + 244, + 903, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rotation Equivariant N2V: We conducted experiments with U-Net architectures. The results are presented in Tab. 2 and Fig. 6. The equivariant error1 of networks N2V, N2VEQ, and $\\mathrm{N2V - EQ^{+}}$ are 0.233, 0.068, and 0.076, respectively. This verifies that the improvements are achieved by reducing equivariant errors. Besides, Tab. 3 further performed diverse scenarios trained on the BSD500 dataset. Our method consistently achieved superior results.", + "bbox": [ + 511, + 340, + 906, + 462 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d71bb2efadd643e5809733f4b6c5e951579ff0cb20db2ae92f720d128f8cef5f.jpg", + "table_caption": [ + "Table 1. N2N: three networks with U-Net architecture were tested under conditions of Gaussian noise at levels 25 and 50." + ], + "table_footnote": [], + "table_body": "
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]28.17/0.82029.05/0.83429.12/0.84526.07/0.72526.38/0.73526.82/0.755
Kodak24 [10]28.86/0.81129.78/0.82529.93/0.83626.75/0.71627.15/0.73227.72/0.754
Set14 [54]27.22/0.80028.04/0.80628.09/0.81625.40/0.71525.65/0.72626.23/0.749
Average28.08/0.81128.96/0.82229.05/0.83226.07/0.71926.39/0.73126.92/0.753
", + "bbox": [ + 516, + 472, + 903, + 537 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce965cd10e45179dd4d9939556f357eab38320bce3fb7fb813cdf9111496e616.jpg", + "table_caption": [ + "Table 2. N2V: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + ], + "table_footnote": [], + "table_body": "
MethodPoissonPoisson30Poisson&GaussianPeppersaltSpeckle
N2V30.82/0.91232.38/0.96127.91/0.82123.93/0.78226.03/0.775
N2V-EQ(ours)31.22/0.90433.59/0.96628.84/0.83524.53/0.79226.42/0.736
N2V-EQ+(ours)32.19/0.92435.96/0.97629.34/0.85024.83/0.81124.85/0.676
", + "bbox": [ + 516, + 584, + 903, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison on diverse scenarios.", + "bbox": [ + 545, + 641, + 870, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rotation Equivariant R2R: The experimental results are shown in Tab. 5. The parameter counts for methods R2R, R2R-EQ, and $\\mathrm{R2R - EQ^{+}}$ are $0.67\\mathrm{M}$ , $0.17\\mathrm{M}$ and $0.84\\mathrm{M}$ , respectively. Due to the larger number of adaptive network parameters, it is expected that the performance is inferior to our EQ version when the training dataset is small.", + "bbox": [ + 511, + 659, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 757, + 668, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selection of rotation-equivariant networks: We conducted ablation experiments with mainstream rotation-equivariant networks, including Cohen et al. [8], Weiler et al. [45] and Shen et al. [39]. Tab. 4 confirms that the chosen Fconv [49] approach is the most effective method in the field of low-level vision.", + "bbox": [ + 511, + 779, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{1}||[L_{R}\\Phi (f) - \\Phi L_{R}(f)\\|_{2}^{2} / \\| L_{R}\\Phi (f)\\|_{2}^{2},$ where $\\Phi (\\cdot)$ represents the network and $L_{R}(\\cdot)$ denotes the rotation transformation.", + "bbox": [ + 511, + 873, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12726", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ee78c577a3df50bb6e388fa230b322be39c148b208478a573bad530bb0dff85.jpg", + "image_caption": [ + "(a) GT" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 295, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73867c6504ddad7660955e61abdd893746df94c4422a1c2df1385b20913df872.jpg", + "image_caption": [ + "(b) $\\mathrm{N2N} / 28.82\\mathrm{dB}$" + ], + "image_footnote": [], + "bbox": [ + 297, + 90, + 496, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6213ffc2364035cbc42f3849df0eb07d84e560978b3f3abdbfc2c2258b1695b8.jpg", + "image_caption": [ + "(c) N2N-EQ / 29.26dB" + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 697, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b5a2bd1a9e40d618f220a64c3eff73ffd574047cb77d888301d31bd88fda90b.jpg", + "image_caption": [ + "(d) $\\mathrm{N2N - EQ^{+} / 29.41dB}$" + ], + "image_footnote": [], + "bbox": [ + 699, + 90, + 901, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3e1d546fcccbe4bab8f090f386a2f802fd3c6f912ea19b4b80ce061aa3fb473.jpg", + "image_caption": [ + "(a) GT", + "Figure 6. N2V: image denoising results of one image from BSD500 with $\\sigma = 25$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 255, + 295, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e93c5a7eb50567340e261f1fb21988e0e5563f7838029bddd8144c2fbbff87d.jpg", + "image_caption": [ + "Figure 5. N2N: image denoising results of one image from kodak with $\\sigma = 50$", + "(b) N2V / 29.97dB" + ], + "image_footnote": [], + "bbox": [ + 297, + 255, + 496, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/363f7db599bc4ddb91f4e9fba65acb8101c5ba036f1eabdf9d9820cb33097030.jpg", + "image_caption": [ + "(c) N2V-EQ / 32.65dB" + ], + "image_footnote": [], + "bbox": [ + 500, + 255, + 699, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbaeac135546f953abed5acfec3749effa8e5d8e288191e0282beda1a5908ee7.jpg", + "image_caption": [ + "(d) $\\mathrm{N2V - EQ^{+} / 33.47dB}$" + ], + "image_footnote": [], + "bbox": [ + 702, + 255, + 903, + 359 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/da5e391b8fea6fb1d2e94bfc8a8cb7771c9412fbba64aca3a9cdbc2ffce2c57a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Gaussian25Gaussian50
MethodKodak24 [10]BSDS300 [31]Set14 [54]AverageKodak24 [10]BSDS300 [31]Set14 [54]Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
CNN31.470.87430.180.86930.020.85130.560.86528.290.77827.020.76227.160.76827.490.769
G-CNN31.390.87230.230.86930.020.85230.550.86428.040.77426.850.75726.880.76227.260.764
E2-CNN31.230.86930.020.86429.700.84530.320.85928.040.76826.860.75426.830.75827.240.760
PDO-eConv31.420.87430.150.86929.830.84930.470.86428.340.78327.060.76527.050.76827.480.772
Fconv31.600.87830.280.87230.060.85430.650.86828.580.79027.240.77127.320.77527.710.779
", + "bbox": [ + 91, + 414, + 906, + 523 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/512a8b7119259ff3ba66d5813946411b9ea714543d0f6e50ff65d6f8398d465d.jpg", + "table_caption": [ + "Table 4. Ablation on rotation-equivariant networks: PSNR and SSIM results for Gaussian25 and Gaussian50." + ], + "table_footnote": [], + "table_body": "
DatasetσR2R [33]R2R-EQR2R-EQ+
BSD682529.03/0.82229.14/0.82529.10/0.826
BSD685026.01/0.70026.14/0.71126.00/0.700
", + "bbox": [ + 93, + 573, + 480, + 632 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1069335b18d3957c9f9f16ce4540eadfc76bc4cf8ea76efc57006255c063c445.jpg", + "table_caption": [ + "Table 5. R2R: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + ], + "table_footnote": [], + "table_body": "
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]23.09/0.75828.19/0.81129.05/0.84722.20/0.66326.06/0.71826.47/0.745
Kodak24 [10]22.77/0.75828.98/0.80929.82/0.84022.72/0.65726.96/0.72127.32/0.743
Set14 [54]22.03/0.73627.07/0.78728.04/0.82019.65/0.60925.27/0.70725.67/0.731
Average22.63/0.75128.08/0.80228.97/0.83621.52/0.64326.10/0.71526.49/0.740
", + "bbox": [ + 93, + 675, + 480, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. N2V w/o rotation augmentation: three networks were tested under conditions of Gaussian noise at levels 25 and 50.", + "bbox": [ + 89, + 751, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data rotation augmentation: Rotational data augmentation on training data is a commonly used method to enhance model performance and robustness. We conducted ablation experiments with rotation augmentation in the N2V method. Tab. 2 shows the experiments on the U-Net network with rotation augmentation, while Tab. 6 presents our ablation experiments without augmentation.", + "bbox": [ + 89, + 789, + 482, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 574, + 633, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we first explore and introduce rotation equivariant image prior into the self-supervised image denoising task at the network architecture level. Through rigorous theoretical analysis, we prove that simply replacing the convolution layers with ECNNs leads to a rotational equivariant version of the network. Building on this, we propose AdaReNet, an adaptive rotation equivariant network, which further enhances performance. Extensive experiments demonstrate the effectiveness of our approach, confirming that incorporating rotation equivariant prior significantly improves denoising results. Overall, our work underscores the importance of leveraging rotation invariance for self-supervised learning and sets a foundation for future research in this field.", + "bbox": [ + 509, + 595, + 906, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research was supported by NSFC project under contract U21A6005; the Major Key Project of PCL under Grant PCL2024A06; Tianyuan Fund for Mathematics of the National Natural Science Foundation of China (Grant No.12426105) and Key Research and Development Program (Grant No. 2024YFA1012000).", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12727", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josue Anaya and Adrian Barbu. Renoir-a dataset for real low-light image noise reduction. Journal of Visual Communication and Image Representation, 51:144-154, 2018. 1", + "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 1, 3", + "[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533. PMLR, 2019. 1, 3", + "[4] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 60-65. IEEE, 2005. 2", + "[5] Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. Non-local means denoising. Image Processing On Line, 1: 208-212, 2011. 2", + "[6] Sungmin Cha, Taeon Park, and Taesup Moon. Gan2gan: Generative noise learning for blind image denoising with single noisy images. arXiv preprint arXiv:1905.10488, 3, 2019. 1, 2, 3", + "[7] Meng Chang, Qi Li, Huajun Feng, and Zhihai Xu. Spatial-adaptive network for single image denoising. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 171-187. Springer, 2020. 1, 3", + "[8] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3, 7", + "[9] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2", + "[10] Rich Franzen. Kodak lossless true color image suite. source: http://rOk.us/graphics/kodak, 4(2):9, 1999. 7, 8", + "[11] Jiahong Fu, Qi Xie, Deyu Meng, and Zongben Xu. Rotation equivariant proximal operator for deep unfolding methods in image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 4", + "[12] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2", + "[13] Shuhang Gu, Qi Xie, Deyu Meng, Wangmeng Zuo, Xiangchu Feng, and Lei Zhang. Weighted nuclear norm minimization and its applications to low level vision. International journal of computer vision, 121:183-208, 2017. 2", + "[14] Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 1, 3", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[16] R Mark Henkelman. Measurement of signal intensities in the presence of noise in mr images. Medical physics, 12(2): 232-233, 1985. 1", + "[17] Emiel Hoogeboom, Jorn WT Peters, Taco S Cohen, and Max Welling. Hexaconv. In International Conference on Learning Representations, 2018. 3", + "[18] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 1, 2, 3", + "[19] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6054-6063, 2019. 3", + "[20] Yoonsik Kim, Jae Woong Soh, Gu Yong Park, and Nam Ik Cho. Transfer learning from synthetic to real-noise denoising with adaptive instance normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3482-3492, 2020. 1", + "[21] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 1, 3, 7, 8", + "[22] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. Frontiers in Computer Science, 2:5, 2020. 1", + "[23] Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4):541-551, 1989. 4", + "[24] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4", + "[25] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 3", + "[26] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 3, 5, 7", + "[27] Jizhihui Liu, Qixun Teng, and Junjun Jiang. Fm2s: Self-supervised fluorescence microscopy denoising with single noisy image. arXiv preprint arXiv:2412.10031, 2024. 7", + "[28] Zhiliang Liu and Liyuan Ren. Shaking noise exploration and elimination for detecting local flaws of steel wire ropes based on magnetic flux leakages. IEEE Transactions on Industrial Electronics, 70(4):4206-4216, 2022. 1" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12728", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ymir Mäkinen, Lucio Azzari, and Alessandro Foi. Exact transform-domain noise variance for collaborative filtering of stationary correlated noise. In 2019 IEEE international conference on image processing (ICIP), pages 185-189. IEEE, 2019. 2", + "[30] Diego Marcos, Michele Volpi, Nikos Komodakis, and Devis Tuia. Rotation equivariant vector field networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5048-5057, 2017. 3", + "[31] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7, 8", + "[32] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 1, 2, 3", + "[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2043-2052, 2021. 1, 2, 3, 7, 8", + "[34] Bumjun Park, Songhyun Yu, and Jechang Jeong. Densely connected hierarchical network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1", + "[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 1", + "[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1890-1898, 2020. 3", + "[37] Yuhui Quan, Yixin Chen, Yizhen Shao, Huan Teng, Yong Xu, and Hui Ji. Image denoising using complex-valued deep cnn. Pattern Recognition, 111:107639, 2021. 3", + "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 7", + "[39] Zhengyang Shen, Lingshen He, Zhouchen Lin, and Jinwen Ma. Pdo-econvs: Partial differential operator based equivariant convolutions. In International Conference on Machine Learning, pages 8697-8706. PMLR, 2020. 2, 3, 7", + "[40] Zhengyang Shen, Tiancheng Shen, Zhouchen Lin, and Jinwen Ma. Pdo-es2cnns: Partial differential operator based equivariant spherical cnns. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9585–9593, 2021. 3", + "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 9446-9454, 2018.3", + "[42] Raviteja Vemulapalli, Oncel Tuzel, and Ming-Yu Liu. Deep gaussian conditional random field network: A model-based deep network for discriminative denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4801-4809, 2016. 1", + "[43] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 7", + "[44] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2027-2036, 2022, 1, 3", + "[45] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. Advances in neural information processing systems, 32, 2019. 3, 7", + "[46] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 849-858, 2018. 2, 3", + "[47] Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5028-5037, 2017. 3", + "[48] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European conference on computer vision, pages 352-368. Springer, 2020. 1", + "[49] Qi Xie, Qian Zhao, Zongben Xu, and Deyu Meng. Fourier series expansion based filter parametrization for equivariant convolutions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4537-4551, 2022. 2, 3, 4, 7", + "[50] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 1", + "[51] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 1, 2, 3", + "[52] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 41-58. Springer, 2020. 3", + "[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 492-511. Springer, 2020. 3", + "[54] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12729", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 7, 8", + "[55] Dan Zhang, Fangfang Zhou, Felix Albu, Yuanzhou Wei, Xiao Yang, Yuan Gu, and Qiang Li. Unleashing the power of self-supervised image denoising: A comprehensive review. arXiv preprint arXiv:2308.00247, 2023. 1, 3", + "[56] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 3", + "[57] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1, 3", + "[58] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 519-528, 2017. 3" + ], + "bbox": [ + 91, + 90, + 483, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12730", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_model.json b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4871edfc971dbf1d3808ad2c2f04c1fc959e8635 --- /dev/null +++ b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_model.json @@ -0,0 +1,2851 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Rotation-Equivariant Self-Supervised Method in Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.179, + 0.764, + 0.199 + ], + "angle": 0, + "content": "Hanze Liu1 Jiahong Fu1 Qi Xie1,† Deyu Meng1,2,3" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.199, + 0.663, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\) Xi'an Jiaotong University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.217, + 0.666, + 0.234 + ], + "angle": 0, + "content": "2 Pengcheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.234, + 0.743, + 0.253 + ], + "angle": 0, + "content": "3 Macau University of Science and Technology, Taipa, Macao" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.771 + ], + "angle": 0, + "content": "Self-supervised image denoising methods have garnered significant research attention in recent years, for this kind of method reduces the requirement of large training datasets. Compared to supervised methods, self-supervised methods rely more on the prior embedded in deep networks themselves. As a result, most of the self-supervised methods are designed with Convolution Neural Networks (CNNs) architectures, which well capture one of the most important image prior, translation equivariant prior. Inspired by the great success achieved by the introduction of translational equivariance, in this paper, we explore the way to further incorporate another important image prior. Specifically, we first apply high-accuracy rotation equivariant convolution to self-supervised image denoising. Through rigorous theoretical analysis, we have proved that simply replacing all the convolution layers with rotation equivariant convolution layers would modify the network into its rotation equivariant version. To the best of our knowledge, this is the first time that rotation equivariant image prior is introduced to self-supervised image denoising at the network architecture level with a comprehensive theoretical analysis of equivariance errors, which offers a new perspective to the field of self-supervised image denoising. Moreover, to further improve the performance, we design a new mask mechanism to fusion the output of rotation equivariant network and vanilla CNN-based network, and construct an adaptive rotation equivariant framework. Through extensive experiments on three typical methods, we have demonstrated the effectiveness of the proposed method. The code is available at: https://github.com/liuhanze623/AdaReNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.222, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.876 + ], + "angle": 0, + "content": "During the process of capturing and transmitting images, unexpected noises are frequently introduced [16, 28]. Such noise can severely degrade image quality and disrupt subsequent image processing tasks. Image denoising is a tech-" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.289, + 0.647, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.388, + 0.613, + 0.397 + ], + "angle": 0, + "content": "(a) Input image" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.29, + 0.773, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.388, + 0.744, + 0.397 + ], + "angle": 0, + "content": "(b) CNN output" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.29, + 0.901, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.388, + 0.882, + 0.397 + ], + "angle": 0, + "content": "(c) Equ. conv. output" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 1. Illustration of the output feature map of a typical image obtained by standard CNN and our used rotation equivariant convolution neural network. Both networks are initialized randomly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.541 + ], + "angle": 0, + "content": "nique designed to address the standard inverse problem in image processing and is widely utilized across various fields. With the rapid advancement of deep learning (DL), denoisers based on different deep networks have achieved outstanding results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.683 + ], + "angle": 0, + "content": "Early, it was believed that the success of deep learning is primarily due to the exhaustive utilization of training data. Thus most early DL-based image denoising are in supervised manner [2, 7, 14, 20, 34, 42, 56, 57], where models are trained on extensive datasets that include pairs of clean and noisy images, thereby learning the transformation from noisy to clean images. However, in reality, compiling such comprehensive datasets is both costly and time-intensive, posing substantial challenges [1, 35, 50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Self-supervised approaches thus achieved significant research attention in recent years. These approaches rely less on extensive supervised datasets, but pay more attention to the full utilization of the prior information inherent in deep networks [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Notably, Lehtinen et al. introduced the innovative self-supervised learning approach called Noise2Noise [26], which enables training on pairs of noisy images depicting the identical scene. [26] was the first self-supervised algorithm that achieved performance comparable to supervised image denoising. Other methods involve training denoising models on single noisy images by designing blind-spot networks [3, 21, 22, 48], while some approaches devise strategies to generate training image pairs from noisy im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.246, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12720" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.108 + ], + "angle": 0, + "content": "ages [6, 18, 32, 33, 51] and have achieved excellent results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.109, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Nowadays, it has become common sense that the prior information inherent in deep networks plays an important role in improving the performance of DL-based methods in image processing tasks [11, 39, 46, 49]. The most classic example should be the CNN, which well captures the translation-equivariant prior of natural images. In CNNs, shifting an input image of CNN is equivalent to shifting all of its intermediate feature maps and the output image. Compared with fully-connected neural networks, this translational equivariance property brings in rational weight sharing, makes the network parameters being used more efficiently, and thus leads to substantially better performance. Inspired by the significant success brought about by the introduction of translational equivariance, another essential image prior (As illustrated in Fig. 1), rotation-equivariant prior has to be taken into consideration very recently. Rotation-equivariant CNNs have been applied and achieved notable performance improvement in multiple supervised image processing tasks [11, 49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Compared with supervised learning, self-supervised learning approaches indeed depend more on the prior knowledge embedded within the network, for the information that can be learned from the dataset is less. As a result, most current self-supervised image denoising methods are all based on the CNN architectures, which is largely due to the reliance on the prior of translational equivariance. Therefore, it holds even greater significance to incorporate rotational equivariant design into the deep networks for self-supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.808 + ], + "angle": 0, + "content": "However, there are two critical issues that need to be addressed when introducing equivariant priors into self-supervised image denoising. On the one hand, although, Xie et al. have recently constructed a rotation equivariant CNN suitable for image processing [49] and Fu et al. further analyzed the global rotation equivariance error of the entire network [11], current rotation equivariant designs for image processing are usually in ResNet[15] structure without upsampling or downsampling. For example, for the superresolution network in [49] where upsampling module is inevitable, only the portion of the network before upsampling is rotation equivariant and the equivariance for the entire network can not be guaranteed. Since self-supervised networks often utilize U-Net structures, which include multiple upsampling and downsampling modules, it is imperative to first explore the impact of these modules on rotation equivariance of the network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "On the other hand, rotation equivariant design for deep networks inevitably introduces parameter sharing and convolution kernel parameterization, which usually degrades the representation accuracy of the network. However, image denoising tasks often have high requirements for the network's representation accuracy, since important high" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "frequency components need to be reconstructed. Besides, not all areas of natural images strictly comply with rigid rotation equivariance, indiscriminately adopting a rotation equivariant network for the entire image can often be detrimental to the reconstruction performance of images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.228 + ], + "angle": 0, + "content": "This study primarily focuses on integrating rotation equivariant prior to existing self-supervised denoising techniques while solving the aforementioned issues. The key contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.907, + 0.472 + ], + "angle": 0, + "content": "- We explore the way for introducing rotation equivariant prior into self-supervised image denoising frameworks at the network architecture level. Particularly, for the first time, we rigorously analyze the impact of upsampling and downsampling on equivariant networks theoretically. The equivariant errors of upsampling and downsampling layer indeed approach zero when the resolution of the input image increases, though the approach rate \\((O(h), h\\) denoted the mesh size of image) is slower than equivariant convolutional layer (whose equivariant error approaches zero at a rate of \\(O(h^{2})\\)). Then, by taking U-Net as an example, we further analyze the rotation equivariant error of the entire network, showing that by simply replacing all the convolution layers with rotation equivariant convolutions [49], we can indeed achieve a reliable rotation equivariant network for self-supervised image denoising." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.473, + 0.906, + 0.608 + ], + "angle": 0, + "content": "- We have further developed an adaptive rotation equivariant network to enhance the representation accuracy. Specifically, we design a fusion module for integrating the advantages of both rotation-equivariant and nonequivariant networks. The module can automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. This design provides greater flexibility and yields improved denoising results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.608, + 0.906, + 0.699 + ], + "angle": 0, + "content": "- We conducted comprehensive experiments across various self-supervised denoising methods, demonstrating the effectiveness of integrating rotation-equivariant image priors into neural networks for self-supervised techniques. Our approach provides a novel perspective in the field of self-supervised image denoising." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.473, + 0.906, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.839, + 0.733 + ], + "angle": 0, + "content": "2. Related Work and Prior Knowledge" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.681, + 0.758 + ], + "angle": 0, + "content": "2.1. Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Non-learning Denoising Methods: Conventional denoising methods predominantly depend on the statistical characteristics of images and mathematical modeling, often employing image priors instead of learned denoisers. Notably, NLM [4, 5] and BM3D [9, 29] have been proposed that can effectively remove noise based on the exploitation of image self-similarity. WNNM [12, 13] treated image denoising as a low-rank matrix approximation problem and achieved great performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12721" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "Supervised Denoising Methods: The majority of DL-based denoising algorithms are supervised [2, 7, 14, 52, 56, 57]. DnCNN [56] employed an end-to-end training strategy to learn the mapping between noisy images and the residual components, markedly improving denoising efficacy. FFDNet [57] proposed a versatile denoising architecture by incorporating the noise level as a network parameter. Nevertheless, to train a proficient model, an extensive collection of paired noisy and clean images is necessary to comprehensively capture the range of image contents and noise variations [2, 14, 19, 25, 37, 53]. The acquisition of such training datasets can be prohibitively costly and challenging, and in the case of certain medical images, it may be virtually unattainable." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.485, + 0.668 + ], + "angle": 0, + "content": "Self-supervised Denoising Methods: Self-supervised deep learning methods have garnered significant interest due to their independence from clean reference images [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Noise2Noise [26], proposed by Lehtinen et al., was the first self-supervised method that achieved performance comparable to supervised methods only using paired noisy images. Noise2Void [21] and Noise2Self [3] used a single noisy image and employed a blind-spot network to predict the clean pixel values based on neighboring pixels, thereby avoiding collapse to an identity mapping. [6, 18, 32, 33, 51] devise strategies to generate training image pairs from noisy images to train a network. R2R [33] has demonstrated that the cost function defined on the noisy/noisy image pairs generated by this method is statistically equivalent to the supervised method. Single-image denoising methods [36, 41] capitalize on the statistical properties of the image itself, showing powerful denoising capabilities without extensive training datasets. These methods lack clean images for supervision and rely more on the prior knowledge embedded within the network, which is the reason for the success of these CNN-based self-supervised image denoising methods. Inspired by this, we explore the way to further incorporate rotation equivariant image prior to these methods with rigorous theoretical analysis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.303, + 0.696 + ], + "angle": 0, + "content": "2.2. Rotation Equivariance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Equivariant Convolution Neural Networks (ECNNs) have drawn substantial interest within the field of computer vision these years [8, 17, 30, 39, 40, 45-47, 58]. Their key strength stems from their ability to effectively handle image rotations through architectural design, which significantly enhances the model's generalization and robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Cohen et al. introduced Group Equivariant Convolutional Networks (G-CNNs) [8] and first integrated \\(\\frac{\\pi}{2}\\) degree rotation equivariance into the neural network. Recently, the filter parametrization technique has been widely employed and Weiler et al. used harmonics as steerable filters to achieve exact equivariance [45, 46]. However, a notable limitation of these filter parameterization techniques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "is the inaccuracy in their representation, which significantly affects low-level vision tasks. Fortunately, Xie et al. addressed this issue by proposing Fourier series expansion-based filter parametrization, which has relatively high expression accuracy [49]. The proposed Fconv exhibits precise equivariance in the continuous domain, degrading to approximation only after discretization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Equivariant convolutions, distinct from data augmentation techniques, inherently integrate rotational symmetry image priors into the network architecture, thereby guaranteeing the network's inherent equivariance and offering superior interpretability and generalizability. However, in the U-Net structure, there is no theoretical guarantee for the design of rotation equivariance. To the best of our knowledge, this study represents the first application of rotation equivariant convolution in the field of self-supervised image denoising and offers theoretical assurance of equivariance for networks based on the U-Net architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.837, + 0.387 + ], + "angle": 0, + "content": "2.3. Prior Knowledge about Equivariance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.906, + 0.499 + ], + "angle": 0, + "content": "Equivalence to a transformation indicates that the application of a transformation to the input results in a corresponding, predictable transformation of the output [39, 45, 49]. Concretely, consider a mapping \\(\\Psi\\) that transforms the input feature space to the output feature space, and let \\(G\\) denote a set of transformations. For any \\(g\\in G\\), the following relationship holds:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.5, + 0.905, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left[ \\pi_ {g} [ f ] \\right] = \\pi_ {g} ^ {\\prime} [ \\Psi [ f ] ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.906, + 0.571 + ], + "angle": 0, + "content": "where \\( f \\) represents any feature map within the input feature space, and \\( \\pi_g \\) and \\( \\pi_g' \\) describe the action of the transformation \\( g \\) on the input and output features, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.688, + 0.599 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.889, + 0.623 + ], + "angle": 0, + "content": "3.1. ECNNs for Self-supervised Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Self-supervised image denoising methods lack clean images for supervision and rely more on the prior knowledge embedded within the network. Therefore, introducing rotation equivariant image prior to this field is reasonable, which can be achieved by ECNNs [49]. The U-Net can effectively restore high-frequency details in images with an encoder-decoder architecture, making it widely used in this task. Consequently, to construct an equivariant self-supervised denoising network, we need to discuss the impact of the essential upsampling and downsampling operators in the U-Net network on equivariance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.906, + 0.9 + ], + "angle": 0, + "content": "In the subsequent section, we first present some notations and concepts, and then theoretically analyze the impact of upsampling and downsampling operators on equivariant networks. Furthermore, we deduce the equivariant error for the complete U-Net network and provide theoretical guarantees for the implementation of equivariance in U-Net architectures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12722" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Figure 2. The network architecture of the equivariant N2N method. The network can be divided into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.313, + 0.304, + 0.328 + ], + "angle": 0, + "content": "3.1.1. Notations and Concepts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.483, + 0.362 + ], + "angle": 0, + "content": "We first introduce some necessary notations and preliminaries as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.514 + ], + "angle": 0, + "content": "We consider the equivariance on the orthogonal group \\( O(2) \\). Formally, \\( O(2) = \\{A \\in \\mathbb{R}^{2 \\times 2} | A^T A = I_{2 \\times 2}\\} \\), which contains all rotation and reflection matrices. Without ambiguity, we use \\( A \\) to parameterize \\( O(2) \\). The Euclidean group \\( E(2) = \\mathbb{R}^2 \\rtimes O(2) \\) (\\(\\rtimes\\) is a semidirect-product), whose element is represented as \\((x, A)\\). Restricting the domain of \\( A \\) and \\( x \\), we can also use this representation to parameterize any subgroup of \\( E(2) \\). In practice, the subgroup is usually assumed to contain \\( t \\) rotations with \\( \\frac{2\\pi}{t} \\) degree for an integer \\( t \\in \\mathbb{N}_+ \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.576 + ], + "angle": 0, + "content": "An image \\(I \\in R^{n \\times n}\\) is viewed as a two-dimensional discretization of a smooth function \\(r: \\mathbb{R}^2 \\to \\mathbb{R}\\), at the cell-center of a regular grid with \\(n \\times n\\) cells, i.e., for \\(i, j = 1, 2, \\dots, n\\)," + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.575, + 0.482, + 0.591 + ], + "angle": 0, + "content": "\\[\nI _ {i j} = r \\left(x _ {i j}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.619 + ], + "angle": 0, + "content": "where \\( x_{ij} = \\left(\\left(i - \\frac{n + 1}{2}\\right)h, \\left(j - \\frac{n + 1}{2}\\right)h\\right)^T \\), \\( h \\) is the mesh size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.725 + ], + "angle": 0, + "content": "An intermediate feature map \\( F \\in \\mathbb{R}^{n \\times n \\times t} \\) in equivariant networks is a multi-channel tensor, which can be viewed as the discretization of a continuous function defined on \\( \\tilde{E} = \\mathbb{R}^2 \\rtimes S \\), where \\( S \\) is a subgroup of \\( O(2) \\) and \\( t \\) is the number of elements in \\( S \\). Formally, \\( F \\) can be represented as a three-dimensional grid tensor sampled from a smooth function \\( e: \\mathbb{R}^2 \\times S \\to \\mathbb{R} \\), i.e., for \\( i, j = 1, 2, \\dots, n \\)," + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.732, + 0.483, + 0.753 + ], + "angle": 0, + "content": "\\[\nF _ {i j} ^ {\\backslash A} = e (x _ {i j}, A), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.763, + 0.351, + 0.777 + ], + "angle": 0, + "content": "where \\( x_{ij} \\) is defined in (2) and \\( A \\in S \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.777, + 0.483, + 0.852 + ], + "angle": 0, + "content": "With above notations, the transformations on the input and feature maps can be mathematically formulated. Specifically, in the continuous domain, for an input \\( r \\in C^{\\infty}(\\mathbb{R}^{2}) \\) and feature map \\( e \\in C^{\\infty}(E(2)) \\), the transformation \\( \\tilde{A} \\in O(2) \\) acts on \\( r \\) and \\( e \\) respectively by:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.864, + 0.482, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\pi_ {\\tilde {A}} ^ {R} [ r ] (x) = r (\\tilde {A} ^ {- 1} x), \\forall x \\in \\mathbb {R} ^ {2}, \\\\ \\pi_ {\\tilde {A}} ^ {E} [ e ] (x, A) = e \\left(\\tilde {A} ^ {- 1} x, \\tilde {A} ^ {- 1} A\\right), \\forall (x, A) \\in E (2). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.154 + ], + "angle": 0, + "content": "In particular, if \\( A_{\\theta} \\in O(2) \\) is the rotation matrix \\( \\left[ \\begin{array}{c}\\cos \\theta ,\\sin \\theta \\\\ -\\sin \\theta ,\\cos \\theta \\end{array} \\right] \\), then the corresponding rotation operators can be expressed by \\( \\pi_{\\theta}^{R} \\) and \\( \\pi_{\\theta}^{E} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.155, + 0.906, + 0.199 + ], + "angle": 0, + "content": "Besides, in the discrete domain, we can also define the transformation \\(\\tilde{A} \\in S\\) on the input image and feature map as followings:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.212, + 0.905, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\tilde {\\pi} _ {\\bar {A}} ^ {R} (I)\\right) _ {i j} = \\pi_ {\\bar {A}} ^ {R} [ r ] (x _ {i j}), \\\\ \\left(\\tilde {\\pi} _ {\\tilde {A}} ^ {E} (F)\\right) _ {i j} ^ {A} = \\pi_ {\\tilde {A}} ^ {E} [ e ] \\left(x _ {i j}, A\\right), \\tag {5} \\\\ \\forall i, j = 1, 2, \\dots , n, A \\in S. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.902, + 0.311 + ], + "angle": 0, + "content": "Similarly, rotation operators can be denoted as \\(\\tilde{\\pi}_{\\theta}^{R}\\) and \\(\\tilde{\\pi}_{\\theta}^{E}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.898, + 0.336 + ], + "angle": 0, + "content": "3.1.2. Equivalence of Downsampling and Upsampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.906, + 0.46 + ], + "angle": 0, + "content": "As shown in Fig. 2, the U-Net architecture commonly used in self-supervised image denoising typically integrates upsampling and downsampling operators. In specific applications, these layers will inevitably affect the equivariance of the network. Recent studies [11, 49] have analyzed the equivariance properties of the group convolution layer. However, the effects of upsampling and downsampling on network equivariance remain unexplored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.491 + ], + "angle": 0, + "content": "First of all, We provide the definitions of commonly used downsampling methods in the continuous domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.596 + ], + "angle": 0, + "content": "Maxpooling Downsampling. Maxpooling is a commonly used downsampling method in CNNs, which reduces the spatial dimensions of feature maps by sliding a fixed-size window over the feature map and selecting the maximum value within each region as the output [23]. In the continuous domain, we can define maxpooling operator \\( M(\\cdot) \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.599, + 0.905, + 0.618 + ], + "angle": 0, + "content": "\\[\n[ M (F) ] (x, A) = \\max _ {\\Omega_ {x}} F _ {i j} ^ {A}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.627, + 0.905, + 0.673 + ], + "angle": 0, + "content": "where \\(x = [x_{1},x_{2}]^{T}\\in \\mathbb{R}^{2}\\) denotes the spatial coordinates, and \\(x_{1}\\in [i,i + 1]\\), \\(x_{2}\\in [j,j + 1]\\), \\(\\Omega_{x} = \\{(i,j),(i + 1,j),(i,j + 1),(i + 1,j + 1)\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Stride Downsampling. Stride Downsampling is also a widely used downsampling operator which reduces the size of the feature map by adjusting the stride of the convolution operation [24]. In the continuous domain, we can define stride downsampling operator \\( S(\\cdot) \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.762, + 0.905, + 0.781 + ], + "angle": 0, + "content": "\\[\n[ S (F) ] (x, A) = F _ {i, j + 1} ^ {A}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.794, + 0.905, + 0.84 + ], + "angle": 0, + "content": "where \\(x = [x_{1},x_{2}]^{T}\\in \\mathbb{R}^{2}\\) denotes the spatial coordinates, and \\(x_{1}\\in [i,i + 1],x_{2}\\in [j,j + 1]\\) \\(\\Omega_{x} = \\{(i,j),(i+\\) \\(1,j),(i,j + 1),(i + 1,j + 1)\\} .\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the above two common downsampling operators, it is of great significance to analyze their impact on the equivariant network. Therefore, we constructed their equivariant errors under the rotational equivariant structure." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12723" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.485, + 0.166 + ], + "angle": 0, + "content": "Theorem 1 Assume that a feature map \\( F \\in \\mathbb{R}^{n \\times n \\times t} \\) is discretized from the smooth function \\( e: \\mathbb{R}^2 \\times S \\to \\mathbb{R} \\), \\( |S| = t \\), the mesh size is \\( h \\), \\( D(\\cdot) \\) is the downsampling operator. If for any \\( A, B \\in S, x \\in \\mathbb{R}^2 \\), the following conditions are satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.167, + 0.483, + 0.184 + ], + "angle": 0, + "content": "\\[\n\\| \\nabla e (x, A) \\| \\leq G, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.188, + 0.351, + 0.204 + ], + "angle": 0, + "content": "then the following results are satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.21, + 0.483, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\left| D \\left[ \\tilde {\\pi} _ {B} ^ {E} \\right] (F) (x, A) - \\pi_ {B} ^ {E} [ D (F) ] (x, A) \\right| \\leq 2 \\sqrt {2} G h. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.484, + 0.312 + ], + "angle": 0, + "content": "The downsampling operator \\( D(\\cdot) \\) can be either \\( M(\\cdot) \\) or \\( S(\\cdot) \\). Theorem 1 reveals that the equivariant error of the downsample operator is primarily influenced by the mesh size \\( h \\), and it indeed approach zero when the resolution of the input image increases with the approach rate \\( O(h) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.342 + ], + "angle": 0, + "content": "Then, we provide the definitions of commonly used upsampling methods in the continuous domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Nearest Neighbor Upsampling. Nearest neighbor interpolation is an image scaling method that fills the pixels of the interpolated image by selecting the original pixel value closest to the target pixel position. In the continuous domain, we can define the nearest neighbor operator \\( N(\\cdot) \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.433, + 0.482, + 0.451 + ], + "angle": 0, + "content": "\\[\n[ N (F) ] (x, A) = F _ {i ^ {*} j ^ {*}} ^ {A}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.359, + 0.47 + ], + "angle": 0, + "content": "where \\((i^{*},j^{*}) = arg\\min_{ij}||x_{ij} - x||_2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Bilinear Upsampling. Bilinear interpolation calculates the new pixel value by taking the weighted average of the four surrounding known pixel values. In the continuous domain, we can define the bilinear interpolation operator \\( B(\\cdot) \\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.544, + 0.483, + 0.585 + ], + "angle": 0, + "content": "\\[\n[ B (F) ] (x, A) = \\sum_ {i = 1} ^ {2} \\sum_ {j = 1} ^ {2} \\lambda_ {i j} f \\left(Q _ {i j}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.65 + ], + "angle": 0, + "content": "where \\(\\lambda_{ij}\\) are the coefficients of bilinear interpolation and \\(f(Q_{ij})\\) represent the grid points, \\(x = [x_1,x_2]^T\\in \\mathbb{R}^2\\) denotes the 2D spatial coordinates, \\(x_{1}\\in [i,i + 1]\\), \\(x_{2}\\in [j,j + 1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.739 + ], + "angle": 0, + "content": "Both of the aforementioned upsampling operators are widely utilized across various network architectures, making it essential to analyze their mathematical properties within rotational equivariant networks. Accordingly, we evaluated their equivariant errors under a rotational equivariant framework." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.749, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Theorem 2 Assume that a feature map \\( F \\in \\mathbb{R}^{n \\times n \\times t} \\) is discretized from the smooth function \\( e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}, |S| = t \\), the mesh size is \\( h \\), \\( U(\\cdot) \\) is the upsampling operator. If for any \\( A, B \\in S, x \\in \\mathbb{R}^2 \\), the following conditions are satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.825, + 0.483, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\| \\nabla e (x, A) \\| \\leq G, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.846, + 0.35, + 0.861 + ], + "angle": 0, + "content": "then the following results are satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.869, + 0.484, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\left| U \\left[ \\tilde {\\pi} _ {B} ^ {E} \\right] (F) (x, A) - \\pi_ {B} ^ {E} [ U (F) ] (x, A) \\right| \\leq 2 (\\sqrt {2} + 1) G h. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "The upsampling operator \\( U(\\cdot) \\) can be either \\( N(\\cdot) \\) or \\( B(\\cdot) \\). It is worth noting that the above conclusions indicate that the error introduced by the upsampling operator in the rotational equivariant network is related to \\( h \\), aligning with established understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.173, + 0.819, + 0.187 + ], + "angle": 0, + "content": "3.1.3. Analysis of Complete U-Net Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.191, + 0.907, + 0.356 + ], + "angle": 0, + "content": "We take the network design of the N2N method [26] as an example to give the derivation of the rotation equivariant error for the entire U-Net architecture. As shown in Fig. 2, we decompose the network into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer (Equivariant Convolution) and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers, we provide the equivariant error for each block and subsequently derive the equivariant error bounds for the complete network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.906, + 0.464 + ], + "angle": 0, + "content": "Theorem 3 demonstrates the equivariance error of the complete U-Net network under discrete angles, and the corollary further provides the equivariance error of the complete U-Net network under any rotation angle. Note that the mesh size of upsampling and downsampling are different, we define the mesh size of the original picture to be \\( h \\), the mesh size after a \\( \\times 2 \\) downsampling is \\( 2h \\), and so on." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Theorem 3 For an image \\(X\\) with size \\(H\\times W\\times n_0\\), and a \\(N\\)-layer rotation equivariant \\(U\\)-Net network \\(\\mathrm{UNet}_{eq}(\\cdot)\\) whose channel number of the \\(l^{th}\\) layer is \\(n_l\\), rotation equivariant subgroup is \\(S\\leqslant O(2),|S| = t\\), and activation function is set as ReLU. If the latent continuous function of the \\(c^{th}\\) channel of \\(X\\) denoted as \\(r_c:\\mathbb{R}^2\\to \\mathbb{R}\\), and the latent continuous function of any convolution filters in the \\(l^{th}\\) layer denoted as \\(\\phi^l:\\mathbb{R}^2\\to \\mathbb{R}\\), \\(DB_{i}\\) and \\(UB_{i}\\) represent the downsampling block and the upsampling block, respectively. \\(\\hat{\\Psi}\\), \\(\\hat{\\Phi}\\), and \\(\\hat{\\Upsilon}\\) represent the convolutional layers in the input, middle, and output stages, respectively. We define:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.648, + 0.906, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\operatorname {U N e t} _ {e q} (\\cdot) = \\hat {\\Upsilon} \\left[ \\hat {U} B _ {m} \\dots \\hat {U} B _ {1} \\left[ \\hat {\\Phi} \\left[ \\hat {D} B _ {m} \\dots \\hat {D} B _ {1} \\left[ \\hat {\\Psi} \\right] \\right] \\dots \\right] (\\cdot), \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.685, + 0.766, + 0.7 + ], + "angle": 0, + "content": "the following conditions are satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.705, + 0.906, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| r _ {c} (x) \\right| \\leq F _ {0}, \\| \\nabla_ {x} r _ {c} (x) \\| \\leq G _ {0}, \\| \\nabla_ {x} ^ {2} r _ {c} (x) \\| \\leq H _ {0}, \\\\ \\left| \\phi^ {l} (x) \\right| \\leq F _ {l}, \\| \\nabla_ {x} \\phi^ {l} (x) \\| \\leq G _ {l}, \\| \\nabla_ {x} ^ {2} \\phi^ {l} (x) \\| \\leq H _ {l}, \\tag {15} \\\\ \\forall \\| x \\| \\geq (p + 1) h / 2, \\phi_ {l} (x) = 0, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.771, + 0.906, + 0.818 + ], + "angle": 0, + "content": "where \\( p \\) is the filter size, \\( h \\) is the mesh size, \\( \\theta_{k} = \\frac{2k\\pi}{t} \\), \\( k = 1,2,\\dots ,t \\). \\( \\nabla_{x} \\) and \\( \\nabla_{x}^{2} \\) denote the operators of gradient and Hessian matrix, respectively. We have" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.823, + 0.905, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathrm {U N e t} _ {e q} \\left[ \\tilde {\\pi} _ {\\theta_ {k}} ^ {R} \\right] (X) - \\tilde {\\pi} _ {\\theta_ {k}} ^ {R} \\left[ \\mathrm {U N e t} _ {e q} \\right] (X) \\right| \\leq R _ {1} h + R _ {2} h ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(R_{1}, R_{2}\\) are two constants with respect to \\(N, n_{l}\\) and the upper bound in (15), their specific values can be found in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12724" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.91, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 3. Illustrations of our proposed adaptive network AdaReNet. Specifically, \\( I \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times \\mathrm{C}} \\) represents a noisy image, where \\( H \\) and \\( W \\) represent the spatial dimensions, and \\( C \\) denotes the channel dimension. The Vanilla Module and EQ Module each produce their respective preliminary denoising results, denoted as \\( f_c \\) and \\( f_e \\). The Fusion Module \\( Mask(\\cdot) \\) automatically decides which areas of the image to use more EQ Module would gain more benefit. After adaptive fusion by \\( Mask(\\cdot) \\) and correction by the Self-correcting Module \\( S_c(\\cdot) \\), the final denoised image \\( \\bar{I} \\) is output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Corollary 1 Under the same condition as Theorem 3, for an arbitrary \\(\\theta \\in [0,2\\pi ]\\), let \\(\\pi_{\\theta}\\) denote the rotation transformation, then \\(\\forall \\theta\\) we have" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.486, + 0.484, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathrm {U N e t} _ {e q} \\left[ \\tilde {\\pi} _ {\\theta} ^ {R} \\right] (X) - \\tilde {\\pi} _ {\\theta} ^ {R} \\left[ \\mathrm {U N e t} _ {e q} \\right] (X) \\right| \\leq R _ {1} h + R _ {2} h ^ {2} + R _ {3} t ^ {- 1} h, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.548 + ], + "angle": 0, + "content": "where \\( R_{1}, R_{2}, R_{3} \\) are constants that can be found in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.577, + 0.192, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.652, + 0.151, + 0.661 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.577, + 0.288, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.652, + 0.249, + 0.661 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.577, + 0.384, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.652, + 0.345, + 0.661 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.577, + 0.48, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.652, + 0.441, + 0.661 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Figure 4. (a) An image from the Kodak dataset, (b) the heatmap of the low-frequency component, (c) the heatmap of the high-frequency component, (d) the output of our proposed MaskNetwork (the brighter area indicates the use of more Vanilla Module)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.757, + 0.43, + 0.774 + ], + "angle": 0, + "content": "3.2. Proposed Adaptive Network AdaReNet" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Rotation equivariant design usually degrades the representation accuracy of the network because of parameter sharing and filter parameterization and not all areas of natural images strictly comply with rigid rotation equivariance. In order to solve the above problems, the proposed network predominantly comprises four primary modules, as depicted in Fig. 3, which are the Vanilla Module, EQ Module, Fusion Module, and Self-correcting Module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Vanilla Module. This module utilizes a conventional CNN architecture, maintaining the original design where only translation equivariance is incorporated into the network. We define the result of the input image \\(I\\) after passing through the Vanilla Module as \\(f_{c}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.517, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\nf _ {c} = \\operatorname {V M} (I). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.636 + ], + "angle": 0, + "content": "EQ Module. This module employs the same architecture as the previous one, with the conventional convolution replaced by the rotation equivariant network to incorporate rotation equivariance prior into the architecture. We define the result of input image \\(I\\) after passing through the EQ Module as \\(f_{e}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.665, + 0.638, + 0.905, + 0.653 + ], + "angle": 0, + "content": "\\[\nf _ {e} = \\operatorname {E Q} (I). \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.662, + 0.907, + 0.799 + ], + "angle": 0, + "content": "Fusion Module. This module introduces a MaskNetwork \\(Mask(\\cdot)\\), a specialized network designed to merge the outputs from the Vanilla Module and the EQ Module. It employs several layers of standard convolutions to automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. The output of \\(Mask(\\cdot)\\) is shown in Fig. 4(d). We define the result of input image \\(I\\) after passing through MaskNetwork as \\(M_{f}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.811, + 0.905, + 0.827 + ], + "angle": 0, + "content": "\\[\nM _ {f} = \\operatorname {M a s k} (I). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Self-correcting Module. Following the Fusion Module, a self-correcting module is applied to refine the fused result. This is achieved through the use of simple ResNet Blocks. We define the Self-correcting Module as \\( S_{c}(\\cdot) \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12725" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "Based on the network presented in Fig. 3, we can derive the following mathematical representation to encapsulate the training and inference processes of the entire network:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.166, + 0.483, + 0.187 + ], + "angle": 0, + "content": "\\[\n\\hat {I} = M _ {f} \\odot f _ {c} + (1 - M _ {f}) \\odot f _ {e}, \\bar {I} = S _ {c} (\\hat {I}), \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.484, + 0.261 + ], + "angle": 0, + "content": "where \\(\\hat{I}\\) denotes the results of adaptive fusion, and \\(\\bar{I}\\) represents the final restoration result after going through the Self-correcting Module, \\(\\odot\\) denote element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.484, + 0.309 + ], + "angle": 0, + "content": "Loss Function. We incorporate the loss of the two subnetworks as regularization terms into the main loss. The loss function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} = \\| \\bar {I} - \\text {t a r g e t} \\| _ {2} + \\alpha_ {1} \\| f _ {c} - \\text {t a r g e t} \\| _ {2} + \\alpha_ {2} \\| f _ {e} - \\text {t a r g e t} \\| _ {2}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.385 + ], + "angle": 0, + "content": "where \\(\\alpha_{1},\\alpha_{2}\\) are hyperparameters and we empirically set \\(\\alpha_{1} = \\alpha_{2} = 0.1\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Remark. Embedding rotational equivariance into the network is highly effective for self-supervised image denoising task. Since natural images do not adhere to strict rotational equivariance, our proposed adaptive rotational equivariant network, named AdaReNet, can automatically decide which regions of the image to apply the rotation-equivariant network, thereby further enhancing performance. By observing the output of the Fusion Module in Fig. 4(d), it is noted that the mask values are larger at the edge details of the pattern, indicating that the adaptive equivariant network tends to use the outputs from the Vanilla Module more frequently in these areas. Conversely, for the restoration of the majority of low-frequency components in the image, the network predominantly uses the outputs from the EQ Module. This aligns with the common sense that convolutions are more adept at fitting high-frequency information[43]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.648, + 0.224, + 0.665 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we conducted experiments based on the existing setup and validated our method in three classic approaches. For the Noise2Noise [26] and Noise2Void [21] methods, we showed experiments on U-Net [38], which can significantly speed up the training process while maintaining acceptable performance, and demonstrated the effectiveness of our method. Besides, we also conducted the R2R experiments based on DnCNN [33], further validating the superiority of the proposed method. Due to space limitations, we only present partial results. Implementation details and further experiments concerning different models, various datasets, model parameter counts, as well as experiments in the field of self-supervised fluorescence microscopy denoising[27] can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.831, + 0.108 + ], + "angle": 0, + "content": "4.1. Experiments on Compared Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.234 + ], + "angle": 0, + "content": "Rotation Equivariant N2N: We selected the most classical Gaussian noise for our experiments, and randomized the standard deviation \\(\\sigma \\in [0,50]\\) of noise for each training example individually. N2N-EQ denotes the rotation equivariant network, while N2N-EQ\\(^+\\) serves as the adaptive rotation equivariant network. The notation in other experiments is similar. The results are shown in Tab. 1. The superiority of our method can also be observed from Fig. 5." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.245, + 0.905, + 0.3 + ], + "angle": 0, + "content": "
DatasetGaussian25Gaussian50
N2N [26]N2N-EQN2N-EQ+N2N [26]N2N-EQN2N-EQ+
Kodak [10]31.47/0.87431.60/0.87831.72/0.88028.29/0.77828.58/0.79028.69/0.791
BSD300 [31]30.18/0.86930.28/0.87230.36/0.87327.02/0.76227.24/0.77127.31/0.772
Set14 [54]30.02/0.85130.06/0.85430.19/0.85527.16/0.76827.32/0.77527.44/0.777
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.338 + ], + "angle": 0, + "content": "Table 1. N2N: three networks with U-Net architecture were tested under conditions of Gaussian noise at levels 25 and 50." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Rotation Equivariant N2V: We conducted experiments with U-Net architectures. The results are presented in Tab. 2 and Fig. 6. The equivariant error1 of networks N2V, N2VEQ, and \\(\\mathrm{N2V - EQ^{+}}\\) are 0.233, 0.068, and 0.076, respectively. This verifies that the improvements are achieved by reducing equivariant errors. Besides, Tab. 3 further performed diverse scenarios trained on the BSD500 dataset. Our method consistently achieved superior results." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.539 + ], + "angle": 0, + "content": "
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]28.17/0.82029.05/0.83429.12/0.84526.07/0.72526.38/0.73526.82/0.755
Kodak24 [10]28.86/0.81129.78/0.82529.93/0.83626.75/0.71627.15/0.73227.72/0.754
Set14 [54]27.22/0.80028.04/0.80628.09/0.81625.40/0.71525.65/0.72626.23/0.749
Average28.08/0.81128.96/0.82229.05/0.83226.07/0.71926.39/0.73126.92/0.753
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.55, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Table 2. N2V: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.633 + ], + "angle": 0, + "content": "
MethodPoissonPoisson30Poisson&GaussianPeppersaltSpeckle
N2V30.82/0.91232.38/0.96127.91/0.82123.93/0.78226.03/0.775
N2V-EQ(ours)31.22/0.90433.59/0.96628.84/0.83524.53/0.79226.42/0.736
N2V-EQ+(ours)32.19/0.92435.96/0.97629.34/0.85024.83/0.81124.85/0.676
" + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.642, + 0.872, + 0.656 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Rotation Equivariant R2R: The experimental results are shown in Tab. 5. The parameter counts for methods R2R, R2R-EQ, and \\(\\mathrm{R2R - EQ^{+}}\\) are \\(0.67\\mathrm{M}\\), \\(0.17\\mathrm{M}\\) and \\(0.84\\mathrm{M}\\), respectively. Due to the larger number of adaptive network parameters, it is expected that the performance is inferior to our EQ version when the training dataset is small." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.669, + 0.774 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Selection of rotation-equivariant networks: We conducted ablation experiments with mainstream rotation-equivariant networks, including Cohen et al. [8], Weiler et al. [45] and Shen et al. [39]. Tab. 4 confirms that the chosen Fconv [49] approach is the most effective method in the field of low-level vision." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}||[L_{R}\\Phi (f) - \\Phi L_{R}(f)\\|_{2}^{2} / \\| L_{R}\\Phi (f)\\|_{2}^{2},\\) where \\(\\Phi (\\cdot)\\) represents the network and \\(L_{R}(\\cdot)\\) denotes the rotation transformation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.297, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.199, + 0.216, + 0.21 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.092, + 0.498, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.199, + 0.455, + 0.21 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{N2N} / 28.82\\mathrm{dB}\\)" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.092, + 0.699, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.199, + 0.663, + 0.21 + ], + "angle": 0, + "content": "(c) N2N-EQ / 29.26dB" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.092, + 0.902, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.198, + 0.866, + 0.21 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{N2N - EQ^{+} / 29.41dB}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.225, + 0.733, + 0.24 + ], + "angle": 0, + "content": "Figure 5. N2N: image denoising results of one image from kodak with \\(\\sigma = 50\\)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.256, + 0.296, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.361, + 0.221, + 0.373 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.256, + 0.498, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.361, + 0.46, + 0.372 + ], + "angle": 0, + "content": "(b) N2V / 29.97dB" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.256, + 0.7, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.361, + 0.675, + 0.373 + ], + "angle": 0, + "content": "(c) N2V-EQ / 32.65dB" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.256, + 0.904, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.361, + 0.876, + 0.373 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{N2V - EQ^{+} / 33.47dB}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.388, + 0.741, + 0.403 + ], + "angle": 0, + "content": "Figure 6. N2V: image denoising results of one image from BSD500 with \\(\\sigma = 25\\)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.415, + 0.907, + 0.525 + ], + "angle": 0, + "content": "
Gaussian25Gaussian50
MethodKodak24 [10]BSDS300 [31]Set14 [54]AverageKodak24 [10]BSDS300 [31]Set14 [54]Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
CNN31.470.87430.180.86930.020.85130.560.86528.290.77827.020.76227.160.76827.490.769
G-CNN31.390.87230.230.86930.020.85230.550.86428.040.77426.850.75726.880.76227.260.764
E2-CNN31.230.86930.020.86429.700.84530.320.85928.040.76826.860.75426.830.75827.240.760
PDO-eConv31.420.87430.150.86929.830.84930.470.86428.340.78327.060.76527.050.76827.480.772
Fconv31.600.87830.280.87230.060.85430.650.86828.580.79027.240.77127.320.77527.710.779
" + }, + { + "type": "table_caption", + "bbox": [ + 0.174, + 0.536, + 0.821, + 0.55 + ], + "angle": 0, + "content": "Table 4. Ablation on rotation-equivariant networks: PSNR and SSIM results for Gaussian25 and Gaussian50." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.574, + 0.481, + 0.633 + ], + "angle": 0, + "content": "
DatasetσR2R [33]R2R-EQR2R-EQ+
BSD682529.03/0.82229.14/0.82529.10/0.826
BSD685026.01/0.70026.14/0.71126.00/0.700
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.643, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Table 5. R2R: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.676, + 0.481, + 0.742 + ], + "angle": 0, + "content": "
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]23.09/0.75828.19/0.81129.05/0.84722.20/0.66326.06/0.71826.47/0.745
Kodak24 [10]22.77/0.75828.98/0.80929.82/0.84022.72/0.65726.96/0.72127.32/0.743
Set14 [54]22.03/0.73627.07/0.78728.04/0.82019.65/0.60925.27/0.70725.67/0.731
Average22.63/0.75128.08/0.80228.97/0.83621.52/0.64326.10/0.71526.49/0.740
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.752, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Table 6. N2V w/o rotation augmentation: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Data rotation augmentation: Rotational data augmentation on training data is a commonly used method to enhance model performance and robustness. We conducted ablation experiments with rotation augmentation in the N2V method. Tab. 2 shows the experiments on the U-Net network with rotation augmentation, while Tab. 6 presents our ablation experiments without augmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.634, + 0.59 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.907, + 0.806 + ], + "angle": 0, + "content": "In this work, we first explore and introduce rotation equivariant image prior into the self-supervised image denoising task at the network architecture level. Through rigorous theoretical analysis, we prove that simply replacing the convolution layers with ECNNs leads to a rotational equivariant version of the network. Building on this, we propose AdaReNet, an adaptive rotation equivariant network, which further enhances performance. Extensive experiments demonstrate the effectiveness of our approach, confirming that incorporating rotation equivariant prior significantly improves denoising results. Overall, our work underscores the importance of leveraging rotation invariance for self-supervised learning and sets a foundation for future research in this field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. This research was supported by NSFC project under contract U21A6005; the Major Key Project of PCL under Grant PCL2024A06; Tianyuan Fund for Mathematics of the National Natural Science Foundation of China (Grant No.12426105) and Key Research and Development Program (Grant No. 2024YFA1012000)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12727" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Josue Anaya and Adrian Barbu. Renoir-a dataset for real low-light image noise reduction. Journal of Visual Communication and Image Representation, 51:144-154, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533. PMLR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[4] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 60-65. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[5] Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. Non-local means denoising. Image Processing On Line, 1: 208-212, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[6] Sungmin Cha, Taeon Park, and Taesup Moon. Gan2gan: Generative noise learning for blind image denoising with single noisy images. arXiv preprint arXiv:1905.10488, 3, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[7] Meng Chang, Qi Li, Huajun Feng, and Zhihai Xu. Spatial-adaptive network for single image denoising. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 171-187. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.529 + ], + "angle": 0, + "content": "[8] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[9] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[10] Rich Franzen. Kodak lossless true color image suite. source: http://rOk.us/graphics/kodak, 4(2):9, 1999. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[11] Jiahong Fu, Qi Xie, Deyu Meng, and Zongben Xu. Rotation equivariant proximal operator for deep unfolding methods in image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[12] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[13] Shuhang Gu, Qi Xie, Deyu Meng, Wangmeng Zuo, Xiangchu Feng, and Lei Zhang. Weighted nuclear norm minimization and its applications to low level vision. International journal of computer vision, 121:183-208, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[14] Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[16] R Mark Henkelman. Measurement of signal intensities in the presence of noise in mr images. Medical physics, 12(2): 232-233, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.905, + 0.209 + ], + "angle": 0, + "content": "[17] Emiel Hoogeboom, Jorn WT Peters, Taco S Cohen, and Max Welling. Hexaconv. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.212, + 0.907, + 0.28 + ], + "angle": 0, + "content": "[18] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.283, + 0.906, + 0.351 + ], + "angle": 0, + "content": "[19] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6054-6063, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.355, + 0.907, + 0.424 + ], + "angle": 0, + "content": "[20] Yoonsik Kim, Jae Woong Soh, Gu Yong Park, and Nam Ik Cho. Transfer learning from synthetic to real-noise denoising with adaptive instance normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3482-3492, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.906, + 0.495 + ], + "angle": 0, + "content": "[21] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[22] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. Frontiers in Computer Science, 2:5, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.556, + 0.905, + 0.624 + ], + "angle": 0, + "content": "[23] Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4):541-551, 1989. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[24] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[25] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[26] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[27] Jizhihui Liu, Qixun Teng, and Junjun Jiang. Fm2s: Self-supervised fluorescence microscopy denoising with single noisy image. arXiv preprint arXiv:2412.10031, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Zhiliang Liu and Liyuan Ren. Shaking noise exploration and elimination for detecting local flaws of steel wire ropes based on magnetic flux leakages. IEEE Transactions on Industrial Electronics, 70(4):4206-4216, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12728" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[29] Ymir Mäkinen, Lucio Azzari, and Alessandro Foi. Exact transform-domain noise variance for collaborative filtering of stationary correlated noise. In 2019 IEEE international conference on image processing (ICIP), pages 185-189. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[30] Diego Marcos, Michele Volpi, Nikos Komodakis, and Devis Tuia. Rotation equivariant vector field networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5048-5057, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[31] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[32] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2043-2052, 2021. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[34] Bumjun Park, Songhyun Yu, and Jechang Jeong. Densely connected hierarchical network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1890-1898, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[37] Yuhui Quan, Yixin Chen, Yizhen Shao, Huan Teng, Yong Xu, and Hui Ji. Image denoising using complex-valued deep cnn. Pattern Recognition, 111:107639, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[39] Zhengyang Shen, Lingshen He, Zhouchen Lin, and Jinwen Ma. Pdo-econvs: Partial differential operator based equivariant convolutions. In International Conference on Machine Learning, pages 8697-8706. PMLR, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[40] Zhengyang Shen, Tiancheng Shen, Zhouchen Lin, and Jinwen Ma. Pdo-es2cnns: Partial differential operator based equivariant spherical cnns. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9585–9593, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 9446-9454, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[42] Raviteja Vemulapalli, Oncel Tuzel, and Ming-Yu Liu. Deep gaussian conditional random field network: A model-based deep network for discriminative denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4801-4809, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[43] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.327 + ], + "angle": 0, + "content": "[44] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2027-2036, 2022, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[45] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. Advances in neural information processing systems, 32, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[46] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 849-858, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[47] Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5028-5037, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[48] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European conference on computer vision, pages 352-368. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[49] Qi Xie, Qian Zhao, Zongben Xu, and Deyu Meng. Fourier series expansion based filter parametrization for equivariant convolutions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4537-4551, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[50] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[51] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[52] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 41-58. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 492-511. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[54] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12729" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[55] Dan Zhang, Fangfang Zhou, Felix Albu, Yuanzhou Wei, Xiao Yang, Yuan Gu, and Qiang Li. Unleashing the power of self-supervised image denoising: A comprehensive review. arXiv preprint arXiv:2308.00247, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[56] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[57] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[58] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 519-528, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12730" + } + ] +] \ No newline at end of file diff --git a/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_origin.pdf b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a2feef5b119adc7ed99b2b8ef413780b5ccb71c --- /dev/null +++ b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/144f4971-bf0f-4465-8bbc-8b32834cb27e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa5fb62aba7b8c6f4cb1b89c7ccc6c4e9c84993b03a7040a7ba8c910f11f0876 +size 1283742 diff --git a/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/full.md b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..90694cc26986e7512d240cd5d140e856ad110a5a --- /dev/null +++ b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/full.md @@ -0,0 +1,437 @@ +# Rotation-Equivariant Self-Supervised Method in Image Denoising + +Hanze Liu1 Jiahong Fu1 Qi Xie1,† Deyu Meng1,2,3 + +$^{1}$ Xi'an Jiaotong University, Xi'an, China + +2 Pengcheng Laboratory, Shenzhen, China + +3 Macau University of Science and Technology, Taipa, Macao + +# Abstract + +Self-supervised image denoising methods have garnered significant research attention in recent years, for this kind of method reduces the requirement of large training datasets. Compared to supervised methods, self-supervised methods rely more on the prior embedded in deep networks themselves. As a result, most of the self-supervised methods are designed with Convolution Neural Networks (CNNs) architectures, which well capture one of the most important image prior, translation equivariant prior. Inspired by the great success achieved by the introduction of translational equivariance, in this paper, we explore the way to further incorporate another important image prior. Specifically, we first apply high-accuracy rotation equivariant convolution to self-supervised image denoising. Through rigorous theoretical analysis, we have proved that simply replacing all the convolution layers with rotation equivariant convolution layers would modify the network into its rotation equivariant version. To the best of our knowledge, this is the first time that rotation equivariant image prior is introduced to self-supervised image denoising at the network architecture level with a comprehensive theoretical analysis of equivariance errors, which offers a new perspective to the field of self-supervised image denoising. Moreover, to further improve the performance, we design a new mask mechanism to fusion the output of rotation equivariant network and vanilla CNN-based network, and construct an adaptive rotation equivariant framework. Through extensive experiments on three typical methods, we have demonstrated the effectiveness of the proposed method. The code is available at: https://github.com/liuhanze623/AdaReNet. + +# 1. Introduction + +During the process of capturing and transmitting images, unexpected noises are frequently introduced [16, 28]. Such noise can severely degrade image quality and disrupt subsequent image processing tasks. Image denoising is a tech- + +![](images/031d15c4d8f19aae03236bff2d1623d9e7b08992e58ce746422a9b091da14bed.jpg) +(a) Input image + +![](images/82dee4da83fe713d603bbe8467bb6e8460ead0480fdf9a923f01e053748543e7.jpg) +(b) CNN output + +![](images/a6e72eba01528aa8564a56f09f52d5397e01e56f6fc3a12ca51c691363bb8afd.jpg) +(c) Equ. conv. output +Figure 1. Illustration of the output feature map of a typical image obtained by standard CNN and our used rotation equivariant convolution neural network. Both networks are initialized randomly. + +nique designed to address the standard inverse problem in image processing and is widely utilized across various fields. With the rapid advancement of deep learning (DL), denoisers based on different deep networks have achieved outstanding results. + +Early, it was believed that the success of deep learning is primarily due to the exhaustive utilization of training data. Thus most early DL-based image denoising are in supervised manner [2, 7, 14, 20, 34, 42, 56, 57], where models are trained on extensive datasets that include pairs of clean and noisy images, thereby learning the transformation from noisy to clean images. However, in reality, compiling such comprehensive datasets is both costly and time-intensive, posing substantial challenges [1, 35, 50]. + +Self-supervised approaches thus achieved significant research attention in recent years. These approaches rely less on extensive supervised datasets, but pay more attention to the full utilization of the prior information inherent in deep networks [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Notably, Lehtinen et al. introduced the innovative self-supervised learning approach called Noise2Noise [26], which enables training on pairs of noisy images depicting the identical scene. [26] was the first self-supervised algorithm that achieved performance comparable to supervised image denoising. Other methods involve training denoising models on single noisy images by designing blind-spot networks [3, 21, 22, 48], while some approaches devise strategies to generate training image pairs from noisy im + +ages [6, 18, 32, 33, 51] and have achieved excellent results. + +Nowadays, it has become common sense that the prior information inherent in deep networks plays an important role in improving the performance of DL-based methods in image processing tasks [11, 39, 46, 49]. The most classic example should be the CNN, which well captures the translation-equivariant prior of natural images. In CNNs, shifting an input image of CNN is equivalent to shifting all of its intermediate feature maps and the output image. Compared with fully-connected neural networks, this translational equivariance property brings in rational weight sharing, makes the network parameters being used more efficiently, and thus leads to substantially better performance. Inspired by the significant success brought about by the introduction of translational equivariance, another essential image prior (As illustrated in Fig. 1), rotation-equivariant prior has to be taken into consideration very recently. Rotation-equivariant CNNs have been applied and achieved notable performance improvement in multiple supervised image processing tasks [11, 49]. + +Compared with supervised learning, self-supervised learning approaches indeed depend more on the prior knowledge embedded within the network, for the information that can be learned from the dataset is less. As a result, most current self-supervised image denoising methods are all based on the CNN architectures, which is largely due to the reliance on the prior of translational equivariance. Therefore, it holds even greater significance to incorporate rotational equivariant design into the deep networks for self-supervised methods. + +However, there are two critical issues that need to be addressed when introducing equivariant priors into self-supervised image denoising. On the one hand, although, Xie et al. have recently constructed a rotation equivariant CNN suitable for image processing [49] and Fu et al. further analyzed the global rotation equivariance error of the entire network [11], current rotation equivariant designs for image processing are usually in ResNet[15] structure without upsampling or downsampling. For example, for the superresolution network in [49] where upsampling module is inevitable, only the portion of the network before upsampling is rotation equivariant and the equivariance for the entire network can not be guaranteed. Since self-supervised networks often utilize U-Net structures, which include multiple upsampling and downsampling modules, it is imperative to first explore the impact of these modules on rotation equivariance of the network. + +On the other hand, rotation equivariant design for deep networks inevitably introduces parameter sharing and convolution kernel parameterization, which usually degrades the representation accuracy of the network. However, image denoising tasks often have high requirements for the network's representation accuracy, since important high + +frequency components need to be reconstructed. Besides, not all areas of natural images strictly comply with rigid rotation equivariance, indiscriminately adopting a rotation equivariant network for the entire image can often be detrimental to the reconstruction performance of images. + +This study primarily focuses on integrating rotation equivariant prior to existing self-supervised denoising techniques while solving the aforementioned issues. The key contributions can be summarized as follows: + +- We explore the way for introducing rotation equivariant prior into self-supervised image denoising frameworks at the network architecture level. Particularly, for the first time, we rigorously analyze the impact of upsampling and downsampling on equivariant networks theoretically. The equivariant errors of upsampling and downsampling layer indeed approach zero when the resolution of the input image increases, though the approach rate $(O(h), h$ denoted the mesh size of image) is slower than equivariant convolutional layer (whose equivariant error approaches zero at a rate of $O(h^{2})$ ). Then, by taking U-Net as an example, we further analyze the rotation equivariant error of the entire network, showing that by simply replacing all the convolution layers with rotation equivariant convolutions [49], we can indeed achieve a reliable rotation equivariant network for self-supervised image denoising. + +- We have further developed an adaptive rotation equivariant network to enhance the representation accuracy. Specifically, we design a fusion module for integrating the advantages of both rotation-equivariant and nonequivariant networks. The module can automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. This design provides greater flexibility and yields improved denoising results. +- We conducted comprehensive experiments across various self-supervised denoising methods, demonstrating the effectiveness of integrating rotation-equivariant image priors into neural networks for self-supervised techniques. Our approach provides a novel perspective in the field of self-supervised image denoising. + +# 2. Related Work and Prior Knowledge + +# 2.1. Image Denoising + +Non-learning Denoising Methods: Conventional denoising methods predominantly depend on the statistical characteristics of images and mathematical modeling, often employing image priors instead of learned denoisers. Notably, NLM [4, 5] and BM3D [9, 29] have been proposed that can effectively remove noise based on the exploitation of image self-similarity. WNNM [12, 13] treated image denoising as a low-rank matrix approximation problem and achieved great performance. + +Supervised Denoising Methods: The majority of DL-based denoising algorithms are supervised [2, 7, 14, 52, 56, 57]. DnCNN [56] employed an end-to-end training strategy to learn the mapping between noisy images and the residual components, markedly improving denoising efficacy. FFDNet [57] proposed a versatile denoising architecture by incorporating the noise level as a network parameter. Nevertheless, to train a proficient model, an extensive collection of paired noisy and clean images is necessary to comprehensively capture the range of image contents and noise variations [2, 14, 19, 25, 37, 53]. The acquisition of such training datasets can be prohibitively costly and challenging, and in the case of certain medical images, it may be virtually unattainable. + +Self-supervised Denoising Methods: Self-supervised deep learning methods have garnered significant interest due to their independence from clean reference images [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Noise2Noise [26], proposed by Lehtinen et al., was the first self-supervised method that achieved performance comparable to supervised methods only using paired noisy images. Noise2Void [21] and Noise2Self [3] used a single noisy image and employed a blind-spot network to predict the clean pixel values based on neighboring pixels, thereby avoiding collapse to an identity mapping. [6, 18, 32, 33, 51] devise strategies to generate training image pairs from noisy images to train a network. R2R [33] has demonstrated that the cost function defined on the noisy/noisy image pairs generated by this method is statistically equivalent to the supervised method. Single-image denoising methods [36, 41] capitalize on the statistical properties of the image itself, showing powerful denoising capabilities without extensive training datasets. These methods lack clean images for supervision and rely more on the prior knowledge embedded within the network, which is the reason for the success of these CNN-based self-supervised image denoising methods. Inspired by this, we explore the way to further incorporate rotation equivariant image prior to these methods with rigorous theoretical analysis. + +# 2.2. Rotation Equivariance + +Equivariant Convolution Neural Networks (ECNNs) have drawn substantial interest within the field of computer vision these years [8, 17, 30, 39, 40, 45-47, 58]. Their key strength stems from their ability to effectively handle image rotations through architectural design, which significantly enhances the model's generalization and robustness. + +Cohen et al. introduced Group Equivariant Convolutional Networks (G-CNNs) [8] and first integrated $\frac{\pi}{2}$ degree rotation equivariance into the neural network. Recently, the filter parametrization technique has been widely employed and Weiler et al. used harmonics as steerable filters to achieve exact equivariance [45, 46]. However, a notable limitation of these filter parameterization techniques + +is the inaccuracy in their representation, which significantly affects low-level vision tasks. Fortunately, Xie et al. addressed this issue by proposing Fourier series expansion-based filter parametrization, which has relatively high expression accuracy [49]. The proposed Fconv exhibits precise equivariance in the continuous domain, degrading to approximation only after discretization. + +Equivariant convolutions, distinct from data augmentation techniques, inherently integrate rotational symmetry image priors into the network architecture, thereby guaranteeing the network's inherent equivariance and offering superior interpretability and generalizability. However, in the U-Net structure, there is no theoretical guarantee for the design of rotation equivariance. To the best of our knowledge, this study represents the first application of rotation equivariant convolution in the field of self-supervised image denoising and offers theoretical assurance of equivariance for networks based on the U-Net architecture. + +# 2.3. Prior Knowledge about Equivariance + +Equivalence to a transformation indicates that the application of a transformation to the input results in a corresponding, predictable transformation of the output [39, 45, 49]. Concretely, consider a mapping $\Psi$ that transforms the input feature space to the output feature space, and let $G$ denote a set of transformations. For any $g\in G$ , the following relationship holds: + +$$ +\Psi \left[ \pi_ {g} [ f ] \right] = \pi_ {g} ^ {\prime} [ \Psi [ f ] ], \tag {1} +$$ + +where $f$ represents any feature map within the input feature space, and $\pi_g$ and $\pi_g'$ describe the action of the transformation $g$ on the input and output features, respectively. + +# 3. Proposed Method + +# 3.1. ECNNs for Self-supervised Image Denoising + +Self-supervised image denoising methods lack clean images for supervision and rely more on the prior knowledge embedded within the network. Therefore, introducing rotation equivariant image prior to this field is reasonable, which can be achieved by ECNNs [49]. The U-Net can effectively restore high-frequency details in images with an encoder-decoder architecture, making it widely used in this task. Consequently, to construct an equivariant self-supervised denoising network, we need to discuss the impact of the essential upsampling and downsampling operators in the U-Net network on equivariance. + +In the subsequent section, we first present some notations and concepts, and then theoretically analyze the impact of upsampling and downsampling operators on equivariant networks. Furthermore, we deduce the equivariant error for the complete U-Net network and provide theoretical guarantees for the implementation of equivariance in U-Net architectures. + +![](images/0bb10b52cb4f3715d20db0c3873d303d685ca87942634912483bed24003a532f.jpg) +Figure 2. The network architecture of the equivariant N2N method. The network can be divided into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers. + +# 3.1.1. Notations and Concepts + +We first introduce some necessary notations and preliminaries as follows. + +We consider the equivariance on the orthogonal group $O(2)$ . Formally, $O(2) = \{A \in \mathbb{R}^{2 \times 2} | A^T A = I_{2 \times 2}\}$ , which contains all rotation and reflection matrices. Without ambiguity, we use $A$ to parameterize $O(2)$ . The Euclidean group $E(2) = \mathbb{R}^2 \rtimes O(2)$ ( $\rtimes$ is a semidirect-product), whose element is represented as $(x, A)$ . Restricting the domain of $A$ and $x$ , we can also use this representation to parameterize any subgroup of $E(2)$ . In practice, the subgroup is usually assumed to contain $t$ rotations with $\frac{2\pi}{t}$ degree for an integer $t \in \mathbb{N}_+$ . + +An image $I \in R^{n \times n}$ is viewed as a two-dimensional discretization of a smooth function $r: \mathbb{R}^2 \to \mathbb{R}$ , at the cell-center of a regular grid with $n \times n$ cells, i.e., for $i, j = 1, 2, \dots, n$ , + +$$ +I _ {i j} = r \left(x _ {i j}\right), \tag {2} +$$ + +where $x_{ij} = \left(\left(i - \frac{n + 1}{2}\right)h, \left(j - \frac{n + 1}{2}\right)h\right)^T$ , $h$ is the mesh size. + +An intermediate feature map $F \in \mathbb{R}^{n \times n \times t}$ in equivariant networks is a multi-channel tensor, which can be viewed as the discretization of a continuous function defined on $\tilde{E} = \mathbb{R}^2 \rtimes S$ , where $S$ is a subgroup of $O(2)$ and $t$ is the number of elements in $S$ . Formally, $F$ can be represented as a three-dimensional grid tensor sampled from a smooth function $e: \mathbb{R}^2 \times S \to \mathbb{R}$ , i.e., for $i, j = 1, 2, \dots, n$ , + +$$ +F _ {i j} ^ {\backslash A} = e (x _ {i j}, A), \tag {3} +$$ + +where $x_{ij}$ is defined in (2) and $A \in S$ . + +With above notations, the transformations on the input and feature maps can be mathematically formulated. Specifically, in the continuous domain, for an input $r \in C^{\infty}(\mathbb{R}^{2})$ and feature map $e \in C^{\infty}(E(2))$ , the transformation $\tilde{A} \in O(2)$ acts on $r$ and $e$ respectively by: + +$$ +\begin{array}{l} \pi_ {\tilde {A}} ^ {R} [ r ] (x) = r (\tilde {A} ^ {- 1} x), \forall x \in \mathbb {R} ^ {2}, \\ \pi_ {\tilde {A}} ^ {E} [ e ] (x, A) = e \left(\tilde {A} ^ {- 1} x, \tilde {A} ^ {- 1} A\right), \forall (x, A) \in E (2). \\ \end{array} +$$ + +In particular, if $A_{\theta} \in O(2)$ is the rotation matrix $\left[ \begin{array}{c}\cos \theta ,\sin \theta \\ -\sin \theta ,\cos \theta \end{array} \right]$ , then the corresponding rotation operators can be expressed by $\pi_{\theta}^{R}$ and $\pi_{\theta}^{E}$ . + +Besides, in the discrete domain, we can also define the transformation $\tilde{A} \in S$ on the input image and feature map as followings: + +$$ +\begin{array}{l} \left(\tilde {\pi} _ {\bar {A}} ^ {R} (I)\right) _ {i j} = \pi_ {\bar {A}} ^ {R} [ r ] (x _ {i j}), \\ \left(\tilde {\pi} _ {\tilde {A}} ^ {E} (F)\right) _ {i j} ^ {A} = \pi_ {\tilde {A}} ^ {E} [ e ] \left(x _ {i j}, A\right), \tag {5} \\ \forall i, j = 1, 2, \dots , n, A \in S. \\ \end{array} +$$ + +Similarly, rotation operators can be denoted as $\tilde{\pi}_{\theta}^{R}$ and $\tilde{\pi}_{\theta}^{E}$ . + +# 3.1.2. Equivalence of Downsampling and Upsampling + +As shown in Fig. 2, the U-Net architecture commonly used in self-supervised image denoising typically integrates upsampling and downsampling operators. In specific applications, these layers will inevitably affect the equivariance of the network. Recent studies [11, 49] have analyzed the equivariance properties of the group convolution layer. However, the effects of upsampling and downsampling on network equivariance remain unexplored. + +First of all, We provide the definitions of commonly used downsampling methods in the continuous domain. + +Maxpooling Downsampling. Maxpooling is a commonly used downsampling method in CNNs, which reduces the spatial dimensions of feature maps by sliding a fixed-size window over the feature map and selecting the maximum value within each region as the output [23]. In the continuous domain, we can define maxpooling operator $M(\cdot)$ as follows, + +$$ +[ M (F) ] (x, A) = \max _ {\Omega_ {x}} F _ {i j} ^ {A}, \tag {6} +$$ + +where $x = [x_{1},x_{2}]^{T}\in \mathbb{R}^{2}$ denotes the spatial coordinates, and $x_{1}\in [i,i + 1]$ , $x_{2}\in [j,j + 1]$ , $\Omega_{x} = \{(i,j),(i + 1,j),(i,j + 1),(i + 1,j + 1)\}$ . + +Stride Downsampling. Stride Downsampling is also a widely used downsampling operator which reduces the size of the feature map by adjusting the stride of the convolution operation [24]. In the continuous domain, we can define stride downsampling operator $S(\cdot)$ as follows, + +$$ +[ S (F) ] (x, A) = F _ {i, j + 1} ^ {A}, \tag {7} +$$ + +where $x = [x_{1},x_{2}]^{T}\in \mathbb{R}^{2}$ denotes the spatial coordinates, and $x_{1}\in [i,i + 1],x_{2}\in [j,j + 1]$ $\Omega_{x} = \{(i,j),(i+$ $1,j),(i,j + 1),(i + 1,j + 1)\} .$ + +For the above two common downsampling operators, it is of great significance to analyze their impact on the equivariant network. Therefore, we constructed their equivariant errors under the rotational equivariant structure. + +Theorem 1 Assume that a feature map $F \in \mathbb{R}^{n \times n \times t}$ is discretized from the smooth function $e: \mathbb{R}^2 \times S \to \mathbb{R}$ , $|S| = t$ , the mesh size is $h$ , $D(\cdot)$ is the downsampling operator. If for any $A, B \in S, x \in \mathbb{R}^2$ , the following conditions are satisfied: + +$$ +\| \nabla e (x, A) \| \leq G, \tag {8} +$$ + +then the following results are satisfied: + +$$ +\left| D \left[ \tilde {\pi} _ {B} ^ {E} \right] (F) (x, A) - \pi_ {B} ^ {E} [ D (F) ] (x, A) \right| \leq 2 \sqrt {2} G h. \tag {9} +$$ + +The downsampling operator $D(\cdot)$ can be either $M(\cdot)$ or $S(\cdot)$ . Theorem 1 reveals that the equivariant error of the downsample operator is primarily influenced by the mesh size $h$ , and it indeed approach zero when the resolution of the input image increases with the approach rate $O(h)$ . + +Then, we provide the definitions of commonly used upsampling methods in the continuous domain. + +Nearest Neighbor Upsampling. Nearest neighbor interpolation is an image scaling method that fills the pixels of the interpolated image by selecting the original pixel value closest to the target pixel position. In the continuous domain, we can define the nearest neighbor operator $N(\cdot)$ as follows, + +$$ +[ N (F) ] (x, A) = F _ {i ^ {*} j ^ {*}} ^ {A}, \tag {10} +$$ + +where $(i^{*},j^{*}) = arg\min_{ij}||x_{ij} - x||_2^2$ + +Bilinear Upsampling. Bilinear interpolation calculates the new pixel value by taking the weighted average of the four surrounding known pixel values. In the continuous domain, we can define the bilinear interpolation operator $B(\cdot)$ as follows, + +$$ +[ B (F) ] (x, A) = \sum_ {i = 1} ^ {2} \sum_ {j = 1} ^ {2} \lambda_ {i j} f \left(Q _ {i j}\right), \tag {11} +$$ + +where $\lambda_{ij}$ are the coefficients of bilinear interpolation and $f(Q_{ij})$ represent the grid points, $x = [x_1,x_2]^T\in \mathbb{R}^2$ denotes the 2D spatial coordinates, $x_{1}\in [i,i + 1]$ , $x_{2}\in [j,j + 1]$ . + +Both of the aforementioned upsampling operators are widely utilized across various network architectures, making it essential to analyze their mathematical properties within rotational equivariant networks. Accordingly, we evaluated their equivariant errors under a rotational equivariant framework. + +Theorem 2 Assume that a feature map $F \in \mathbb{R}^{n \times n \times t}$ is discretized from the smooth function $e: \mathbb{R}^2 \times S \to \mathbb{R}, |S| = t$ , the mesh size is $h$ , $U(\cdot)$ is the upsampling operator. If for any $A, B \in S, x \in \mathbb{R}^2$ , the following conditions are satisfied: + +$$ +\| \nabla e (x, A) \| \leq G, \tag {12} +$$ + +then the following results are satisfied: + +$$ +\left| U \left[ \tilde {\pi} _ {B} ^ {E} \right] (F) (x, A) - \pi_ {B} ^ {E} [ U (F) ] (x, A) \right| \leq 2 (\sqrt {2} + 1) G h. \tag {13} +$$ + +The upsampling operator $U(\cdot)$ can be either $N(\cdot)$ or $B(\cdot)$ . It is worth noting that the above conclusions indicate that the error introduced by the upsampling operator in the rotational equivariant network is related to $h$ , aligning with established understanding. + +# 3.1.3. Analysis of Complete U-Net Network + +We take the network design of the N2N method [26] as an example to give the derivation of the rotation equivariant error for the entire U-Net architecture. As shown in Fig. 2, we decompose the network into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer (Equivariant Convolution) and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers, we provide the equivariant error for each block and subsequently derive the equivariant error bounds for the complete network. + +Theorem 3 demonstrates the equivariance error of the complete U-Net network under discrete angles, and the corollary further provides the equivariance error of the complete U-Net network under any rotation angle. Note that the mesh size of upsampling and downsampling are different, we define the mesh size of the original picture to be $h$ , the mesh size after a $\times 2$ downsampling is $2h$ , and so on. + +Theorem 3 For an image $X$ with size $H\times W\times n_0$ , and a $N$ -layer rotation equivariant $U$ -Net network $\mathrm{UNet}_{eq}(\cdot)$ whose channel number of the $l^{th}$ layer is $n_l$ , rotation equivariant subgroup is $S\leqslant O(2),|S| = t$ , and activation function is set as ReLU. If the latent continuous function of the $c^{th}$ channel of $X$ denoted as $r_c:\mathbb{R}^2\to \mathbb{R}$ , and the latent continuous function of any convolution filters in the $l^{th}$ layer denoted as $\phi^l:\mathbb{R}^2\to \mathbb{R}$ , $DB_{i}$ and $UB_{i}$ represent the downsampling block and the upsampling block, respectively. $\hat{\Psi}$ , $\hat{\Phi}$ , and $\hat{\Upsilon}$ represent the convolutional layers in the input, middle, and output stages, respectively. We define: + +$$ +\operatorname {U N e t} _ {e q} (\cdot) = \hat {\Upsilon} \left[ \hat {U} B _ {m} \dots \hat {U} B _ {1} \left[ \hat {\Phi} \left[ \hat {D} B _ {m} \dots \hat {D} B _ {1} \left[ \hat {\Psi} \right] \right] \dots \right] (\cdot), \right. \tag {14} +$$ + +the following conditions are satisfied: + +$$ +\begin{array}{l} \left| r _ {c} (x) \right| \leq F _ {0}, \| \nabla_ {x} r _ {c} (x) \| \leq G _ {0}, \| \nabla_ {x} ^ {2} r _ {c} (x) \| \leq H _ {0}, \\ \left| \phi^ {l} (x) \right| \leq F _ {l}, \| \nabla_ {x} \phi^ {l} (x) \| \leq G _ {l}, \| \nabla_ {x} ^ {2} \phi^ {l} (x) \| \leq H _ {l}, \tag {15} \\ \forall \| x \| \geq (p + 1) h / 2, \phi_ {l} (x) = 0, \\ \end{array} +$$ + +where $p$ is the filter size, $h$ is the mesh size, $\theta_{k} = \frac{2k\pi}{t}$ , $k = 1,2,\dots ,t$ . $\nabla_{x}$ and $\nabla_{x}^{2}$ denote the operators of gradient and Hessian matrix, respectively. We have + +$$ +\left| \mathrm {U N e t} _ {e q} \left[ \tilde {\pi} _ {\theta_ {k}} ^ {R} \right] (X) - \tilde {\pi} _ {\theta_ {k}} ^ {R} \left[ \mathrm {U N e t} _ {e q} \right] (X) \right| \leq R _ {1} h + R _ {2} h ^ {2}. \tag {16} +$$ + +where $R_{1}, R_{2}$ are two constants with respect to $N, n_{l}$ and the upper bound in (15), their specific values can be found in the supplementary materials. + +![](images/ca00c297445f151b65f928c399921d23eb8f3c23b76c42722455854486b3af48.jpg) +Figure 3. Illustrations of our proposed adaptive network AdaReNet. Specifically, $I \in \mathbb{R}^{\mathrm{H} \times \mathrm{W} \times \mathrm{C}}$ represents a noisy image, where $H$ and $W$ represent the spatial dimensions, and $C$ denotes the channel dimension. The Vanilla Module and EQ Module each produce their respective preliminary denoising results, denoted as $f_c$ and $f_e$ . The Fusion Module $Mask(\cdot)$ automatically decides which areas of the image to use more EQ Module would gain more benefit. After adaptive fusion by $Mask(\cdot)$ and correction by the Self-correcting Module $S_c(\cdot)$ , the final denoised image $\bar{I}$ is output. + +Corollary 1 Under the same condition as Theorem 3, for an arbitrary $\theta \in [0,2\pi ]$ , let $\pi_{\theta}$ denote the rotation transformation, then $\forall \theta$ we have + +$$ +\left| \mathrm {U N e t} _ {e q} \left[ \tilde {\pi} _ {\theta} ^ {R} \right] (X) - \tilde {\pi} _ {\theta} ^ {R} \left[ \mathrm {U N e t} _ {e q} \right] (X) \right| \leq R _ {1} h + R _ {2} h ^ {2} + R _ {3} t ^ {- 1} h, \tag {17} +$$ + +where $R_{1}, R_{2}, R_{3}$ are constants that can be found in the supplementary materials. + +![](images/8f2c85e6f5a7481384f6cadf7273fa12fe9f52168c394457b3782ee1e9b10cde.jpg) +(a) + +![](images/311276b789292fa4a963cd43fc3403acd7f9ba656973b27f4455f5b711df2367.jpg) +(b) + +![](images/d197b7e75b11fd15a10152a02f592a8b63f043f80544922bf958c0da424558e8.jpg) +(c) +Figure 4. (a) An image from the Kodak dataset, (b) the heatmap of the low-frequency component, (c) the heatmap of the high-frequency component, (d) the output of our proposed MaskNetwork (the brighter area indicates the use of more Vanilla Module). + +![](images/08f1bfb0ff3129c1b59f28fb1aafee7177e32e0251253f1ea3cdc541044854a2.jpg) +(d) + +# 3.2. Proposed Adaptive Network AdaReNet + +Rotation equivariant design usually degrades the representation accuracy of the network because of parameter sharing and filter parameterization and not all areas of natural images strictly comply with rigid rotation equivariance. In order to solve the above problems, the proposed network predominantly comprises four primary modules, as depicted in Fig. 3, which are the Vanilla Module, EQ Module, Fusion Module, and Self-correcting Module. + +Vanilla Module. This module utilizes a conventional CNN architecture, maintaining the original design where only translation equivariance is incorporated into the network. We define the result of the input image $I$ after passing through the Vanilla Module as $f_{c}$ . + +$$ +f _ {c} = \operatorname {V M} (I). \tag {18} +$$ + +EQ Module. This module employs the same architecture as the previous one, with the conventional convolution replaced by the rotation equivariant network to incorporate rotation equivariance prior into the architecture. We define the result of input image $I$ after passing through the EQ Module as $f_{e}$ , + +$$ +f _ {e} = \operatorname {E Q} (I). \tag {19} +$$ + +Fusion Module. This module introduces a MaskNetwork $Mask(\cdot)$ , a specialized network designed to merge the outputs from the Vanilla Module and the EQ Module. It employs several layers of standard convolutions to automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. The output of $Mask(\cdot)$ is shown in Fig. 4(d). We define the result of input image $I$ after passing through MaskNetwork as $M_{f}$ , + +$$ +M _ {f} = \operatorname {M a s k} (I). \tag {20} +$$ + +Self-correcting Module. Following the Fusion Module, a self-correcting module is applied to refine the fused result. This is achieved through the use of simple ResNet Blocks. We define the Self-correcting Module as $S_{c}(\cdot)$ . + +Based on the network presented in Fig. 3, we can derive the following mathematical representation to encapsulate the training and inference processes of the entire network: + +$$ +\hat {I} = M _ {f} \odot f _ {c} + (1 - M _ {f}) \odot f _ {e}, \bar {I} = S _ {c} (\hat {I}), \tag {21} +$$ + +where $\hat{I}$ denotes the results of adaptive fusion, and $\bar{I}$ represents the final restoration result after going through the Self-correcting Module, $\odot$ denote element-wise multiplication. + +Loss Function. We incorporate the loss of the two subnetworks as regularization terms into the main loss. The loss function is defined as follows: + +$$ +\mathrm {L} = \| \bar {I} - \text {t a r g e t} \| _ {2} + \alpha_ {1} \| f _ {c} - \text {t a r g e t} \| _ {2} + \alpha_ {2} \| f _ {e} - \text {t a r g e t} \| _ {2}, \tag {22} +$$ + +where $\alpha_{1},\alpha_{2}$ are hyperparameters and we empirically set $\alpha_{1} = \alpha_{2} = 0.1$ + +Remark. Embedding rotational equivariance into the network is highly effective for self-supervised image denoising task. Since natural images do not adhere to strict rotational equivariance, our proposed adaptive rotational equivariant network, named AdaReNet, can automatically decide which regions of the image to apply the rotation-equivariant network, thereby further enhancing performance. By observing the output of the Fusion Module in Fig. 4(d), it is noted that the mask values are larger at the edge details of the pattern, indicating that the adaptive equivariant network tends to use the outputs from the Vanilla Module more frequently in these areas. Conversely, for the restoration of the majority of low-frequency components in the image, the network predominantly uses the outputs from the EQ Module. This aligns with the common sense that convolutions are more adept at fitting high-frequency information[43]. + +# 4. Experiments + +In this section, we conducted experiments based on the existing setup and validated our method in three classic approaches. For the Noise2Noise [26] and Noise2Void [21] methods, we showed experiments on U-Net [38], which can significantly speed up the training process while maintaining acceptable performance, and demonstrated the effectiveness of our method. Besides, we also conducted the R2R experiments based on DnCNN [33], further validating the superiority of the proposed method. Due to space limitations, we only present partial results. Implementation details and further experiments concerning different models, various datasets, model parameter counts, as well as experiments in the field of self-supervised fluorescence microscopy denoising[27] can be found in supplementary materials. + +# 4.1. Experiments on Compared Methods + +Rotation Equivariant N2N: We selected the most classical Gaussian noise for our experiments, and randomized the standard deviation $\sigma \in [0,50]$ of noise for each training example individually. N2N-EQ denotes the rotation equivariant network, while N2N-EQ $^+$ serves as the adaptive rotation equivariant network. The notation in other experiments is similar. The results are shown in Tab. 1. The superiority of our method can also be observed from Fig. 5. + +
DatasetGaussian25Gaussian50
N2N [26]N2N-EQN2N-EQ+N2N [26]N2N-EQN2N-EQ+
Kodak [10]31.47/0.87431.60/0.87831.72/0.88028.29/0.77828.58/0.79028.69/0.791
BSD300 [31]30.18/0.86930.28/0.87230.36/0.87327.02/0.76227.24/0.77127.31/0.772
Set14 [54]30.02/0.85130.06/0.85430.19/0.85527.16/0.76827.32/0.77527.44/0.777
+ +Rotation Equivariant N2V: We conducted experiments with U-Net architectures. The results are presented in Tab. 2 and Fig. 6. The equivariant error1 of networks N2V, N2VEQ, and $\mathrm{N2V - EQ^{+}}$ are 0.233, 0.068, and 0.076, respectively. This verifies that the improvements are achieved by reducing equivariant errors. Besides, Tab. 3 further performed diverse scenarios trained on the BSD500 dataset. Our method consistently achieved superior results. + +Table 1. N2N: three networks with U-Net architecture were tested under conditions of Gaussian noise at levels 25 and 50. + +
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]28.17/0.82029.05/0.83429.12/0.84526.07/0.72526.38/0.73526.82/0.755
Kodak24 [10]28.86/0.81129.78/0.82529.93/0.83626.75/0.71627.15/0.73227.72/0.754
Set14 [54]27.22/0.80028.04/0.80628.09/0.81625.40/0.71525.65/0.72626.23/0.749
Average28.08/0.81128.96/0.82229.05/0.83226.07/0.71926.39/0.73126.92/0.753
+ +Table 2. N2V: three networks were tested under conditions of Gaussian noise at levels 25 and 50. + +
MethodPoissonPoisson30Poisson&GaussianPeppersaltSpeckle
N2V30.82/0.91232.38/0.96127.91/0.82123.93/0.78226.03/0.775
N2V-EQ(ours)31.22/0.90433.59/0.96628.84/0.83524.53/0.79226.42/0.736
N2V-EQ+(ours)32.19/0.92435.96/0.97629.34/0.85024.83/0.81124.85/0.676
+ +Table 3. Quantitative comparison on diverse scenarios. + +Rotation Equivariant R2R: The experimental results are shown in Tab. 5. The parameter counts for methods R2R, R2R-EQ, and $\mathrm{R2R - EQ^{+}}$ are $0.67\mathrm{M}$ , $0.17\mathrm{M}$ and $0.84\mathrm{M}$ , respectively. Due to the larger number of adaptive network parameters, it is expected that the performance is inferior to our EQ version when the training dataset is small. + +# 4.2. Ablation Study + +Selection of rotation-equivariant networks: We conducted ablation experiments with mainstream rotation-equivariant networks, including Cohen et al. [8], Weiler et al. [45] and Shen et al. [39]. Tab. 4 confirms that the chosen Fconv [49] approach is the most effective method in the field of low-level vision. + +![](images/5ee78c577a3df50bb6e388fa230b322be39c148b208478a573bad530bb0dff85.jpg) +(a) GT + +![](images/73867c6504ddad7660955e61abdd893746df94c4422a1c2df1385b20913df872.jpg) +(b) $\mathrm{N2N} / 28.82\mathrm{dB}$ + +![](images/6213ffc2364035cbc42f3849df0eb07d84e560978b3f3abdbfc2c2258b1695b8.jpg) +(c) N2N-EQ / 29.26dB + +![](images/3b5a2bd1a9e40d618f220a64c3eff73ffd574047cb77d888301d31bd88fda90b.jpg) +(d) $\mathrm{N2N - EQ^{+} / 29.41dB}$ + +![](images/b3e1d546fcccbe4bab8f090f386a2f802fd3c6f912ea19b4b80ce061aa3fb473.jpg) +(a) GT +Figure 6. N2V: image denoising results of one image from BSD500 with $\sigma = 25$ . + +![](images/6e93c5a7eb50567340e261f1fb21988e0e5563f7838029bddd8144c2fbbff87d.jpg) +Figure 5. N2N: image denoising results of one image from kodak with $\sigma = 50$ +(b) N2V / 29.97dB + +![](images/363f7db599bc4ddb91f4e9fba65acb8101c5ba036f1eabdf9d9820cb33097030.jpg) +(c) N2V-EQ / 32.65dB + +![](images/bbaeac135546f953abed5acfec3749effa8e5d8e288191e0282beda1a5908ee7.jpg) +(d) $\mathrm{N2V - EQ^{+} / 33.47dB}$ + +
Gaussian25Gaussian50
MethodKodak24 [10]BSDS300 [31]Set14 [54]AverageKodak24 [10]BSDS300 [31]Set14 [54]Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
CNN31.470.87430.180.86930.020.85130.560.86528.290.77827.020.76227.160.76827.490.769
G-CNN31.390.87230.230.86930.020.85230.550.86428.040.77426.850.75726.880.76227.260.764
E2-CNN31.230.86930.020.86429.700.84530.320.85928.040.76826.860.75426.830.75827.240.760
PDO-eConv31.420.87430.150.86929.830.84930.470.86428.340.78327.060.76527.050.76827.480.772
Fconv31.600.87830.280.87230.060.85430.650.86828.580.79027.240.77127.320.77527.710.779
+ +Table 4. Ablation on rotation-equivariant networks: PSNR and SSIM results for Gaussian25 and Gaussian50. + +
DatasetσR2R [33]R2R-EQR2R-EQ+
BSD682529.03/0.82229.14/0.82529.10/0.826
BSD685026.01/0.70026.14/0.71126.00/0.700
+ +Table 5. R2R: three networks were tested under conditions of Gaussian noise at levels 25 and 50. + +
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]23.09/0.75828.19/0.81129.05/0.84722.20/0.66326.06/0.71826.47/0.745
Kodak24 [10]22.77/0.75828.98/0.80929.82/0.84022.72/0.65726.96/0.72127.32/0.743
Set14 [54]22.03/0.73627.07/0.78728.04/0.82019.65/0.60925.27/0.70725.67/0.731
Average22.63/0.75128.08/0.80228.97/0.83621.52/0.64326.10/0.71526.49/0.740
+ +Table 6. N2V w/o rotation augmentation: three networks were tested under conditions of Gaussian noise at levels 25 and 50. + +Data rotation augmentation: Rotational data augmentation on training data is a commonly used method to enhance model performance and robustness. We conducted ablation experiments with rotation augmentation in the N2V method. Tab. 2 shows the experiments on the U-Net network with rotation augmentation, while Tab. 6 presents our ablation experiments without augmentation. + +# 5. Conclusion + +In this work, we first explore and introduce rotation equivariant image prior into the self-supervised image denoising task at the network architecture level. Through rigorous theoretical analysis, we prove that simply replacing the convolution layers with ECNNs leads to a rotational equivariant version of the network. Building on this, we propose AdaReNet, an adaptive rotation equivariant network, which further enhances performance. Extensive experiments demonstrate the effectiveness of our approach, confirming that incorporating rotation equivariant prior significantly improves denoising results. Overall, our work underscores the importance of leveraging rotation invariance for self-supervised learning and sets a foundation for future research in this field. + +Acknowledgement. This research was supported by NSFC project under contract U21A6005; the Major Key Project of PCL under Grant PCL2024A06; Tianyuan Fund for Mathematics of the National Natural Science Foundation of China (Grant No.12426105) and Key Research and Development Program (Grant No. 2024YFA1012000). + +# References + +[1] Josue Anaya and Adrian Barbu. Renoir-a dataset for real low-light image noise reduction. Journal of Visual Communication and Image Representation, 51:144-154, 2018. 1 +[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 1, 3 +[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533. PMLR, 2019. 1, 3 +[4] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 60-65. IEEE, 2005. 2 +[5] Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. Non-local means denoising. Image Processing On Line, 1: 208-212, 2011. 2 +[6] Sungmin Cha, Taeon Park, and Taesup Moon. Gan2gan: Generative noise learning for blind image denoising with single noisy images. arXiv preprint arXiv:1905.10488, 3, 2019. 1, 2, 3 +[7] Meng Chang, Qi Li, Huajun Feng, and Zhihai Xu. Spatial-adaptive network for single image denoising. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 171-187. Springer, 2020. 1, 3 +[8] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3, 7 +[9] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2 +[10] Rich Franzen. Kodak lossless true color image suite. source: http://rOk.us/graphics/kodak, 4(2):9, 1999. 7, 8 +[11] Jiahong Fu, Qi Xie, Deyu Meng, and Zongben Xu. Rotation equivariant proximal operator for deep unfolding methods in image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 4 +[12] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2 +[13] Shuhang Gu, Qi Xie, Deyu Meng, Wangmeng Zuo, Xiangchu Feng, and Lei Zhang. Weighted nuclear norm minimization and its applications to low level vision. International journal of computer vision, 121:183-208, 2017. 2 +[14] Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 1, 3 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[16] R Mark Henkelman. Measurement of signal intensities in the presence of noise in mr images. Medical physics, 12(2): 232-233, 1985. 1 +[17] Emiel Hoogeboom, Jorn WT Peters, Taco S Cohen, and Max Welling. Hexaconv. In International Conference on Learning Representations, 2018. 3 +[18] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 1, 2, 3 +[19] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6054-6063, 2019. 3 +[20] Yoonsik Kim, Jae Woong Soh, Gu Yong Park, and Nam Ik Cho. Transfer learning from synthetic to real-noise denoising with adaptive instance normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3482-3492, 2020. 1 +[21] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 1, 3, 7, 8 +[22] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. Frontiers in Computer Science, 2:5, 2020. 1 +[23] Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4):541-551, 1989. 4 +[24] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4 +[25] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 3 +[26] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 3, 5, 7 +[27] Jizhihui Liu, Qixun Teng, and Junjun Jiang. Fm2s: Self-supervised fluorescence microscopy denoising with single noisy image. arXiv preprint arXiv:2412.10031, 2024. 7 +[28] Zhiliang Liu and Liyuan Ren. Shaking noise exploration and elimination for detecting local flaws of steel wire ropes based on magnetic flux leakages. IEEE Transactions on Industrial Electronics, 70(4):4206-4216, 2022. 1 + +[29] Ymir Mäkinen, Lucio Azzari, and Alessandro Foi. Exact transform-domain noise variance for collaborative filtering of stationary correlated noise. In 2019 IEEE international conference on image processing (ICIP), pages 185-189. IEEE, 2019. 2 +[30] Diego Marcos, Michele Volpi, Nikos Komodakis, and Devis Tuia. Rotation equivariant vector field networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5048-5057, 2017. 3 +[31] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7, 8 +[32] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 1, 2, 3 +[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2043-2052, 2021. 1, 2, 3, 7, 8 +[34] Bumjun Park, Songhyun Yu, and Jechang Jeong. Densely connected hierarchical network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1 +[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 1 +[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1890-1898, 2020. 3 +[37] Yuhui Quan, Yixin Chen, Yizhen Shao, Huan Teng, Yong Xu, and Hui Ji. Image denoising using complex-valued deep cnn. Pattern Recognition, 111:107639, 2021. 3 +[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 7 +[39] Zhengyang Shen, Lingshen He, Zhouchen Lin, and Jinwen Ma. Pdo-econvs: Partial differential operator based equivariant convolutions. In International Conference on Machine Learning, pages 8697-8706. PMLR, 2020. 2, 3, 7 +[40] Zhengyang Shen, Tiancheng Shen, Zhouchen Lin, and Jinwen Ma. Pdo-es2cnns: Partial differential operator based equivariant spherical cnns. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9585–9593, 2021. 3 +[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on + +computer vision and pattern recognition, pages 9446-9454, 2018.3 +[42] Raviteja Vemulapalli, Oncel Tuzel, and Ming-Yu Liu. Deep gaussian conditional random field network: A model-based deep network for discriminative denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4801-4809, 2016. 1 +[43] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 7 +[44] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2027-2036, 2022, 1, 3 +[45] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. Advances in neural information processing systems, 32, 2019. 3, 7 +[46] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 849-858, 2018. 2, 3 +[47] Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5028-5037, 2017. 3 +[48] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European conference on computer vision, pages 352-368. Springer, 2020. 1 +[49] Qi Xie, Qian Zhao, Zongben Xu, and Deyu Meng. Fourier series expansion based filter parametrization for equivariant convolutions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4537-4551, 2022. 2, 3, 4, 7 +[50] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 1 +[51] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 1, 2, 3 +[52] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 41-58. Springer, 2020. 3 +[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 492-511. Springer, 2020. 3 +[54] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and + +Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 7, 8 +[55] Dan Zhang, Fangfang Zhou, Felix Albu, Yuanzhou Wei, Xiao Yang, Yuan Gu, and Qiang Li. Unleashing the power of self-supervised image denoising: A comprehensive review. arXiv preprint arXiv:2308.00247, 2023. 1, 3 +[56] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 3 +[57] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1, 3 +[58] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 519-528, 2017. 3 \ No newline at end of file diff --git a/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/images.zip b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..78505c145fabda898d19c4bd448739c68d117a75 --- /dev/null +++ b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0abc0b868f5d76cce05f32fc018a6ff3fd266bae85245793258c9496ca314bc8 +size 564904 diff --git a/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/layout.json b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee4d9233154d540778fa49fc81d6a4a42831cd90 --- /dev/null +++ b/2025/Rotation-Equivariant Self-Supervised Method in Image Denoising/layout.json @@ -0,0 +1,11948 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 510, + 121 + ], + "type": "text", + "content": "Rotation-Equivariant Self-Supervised Method in Image Denoising" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 141, + 467, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 141, + 467, + 157 + ], + "spans": [ + { + "bbox": [ + 143, + 141, + 467, + 157 + ], + "type": "text", + "content": "Hanze Liu1 Jiahong Fu1 Qi Xie1,† Deyu Meng1,2,3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 157, + 405, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 157, + 405, + 171 + ], + "spans": [ + { + "bbox": [ + 204, + 157, + 405, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 157, + 405, + 171 + ], + "type": "text", + "content": " Xi'an Jiaotong University, Xi'an, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 203, + 171, + 407, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 171, + 407, + 185 + ], + "spans": [ + { + "bbox": [ + 203, + 171, + 407, + 185 + ], + "type": "text", + "content": "2 Pengcheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 185, + 454, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 185, + 454, + 200 + ], + "spans": [ + { + "bbox": [ + 156, + 185, + 454, + 200 + ], + "type": "text", + "content": "3 Macau University of Science and Technology, Taipa, Macao" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 251, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 610 + ], + "type": "text", + "content": "Self-supervised image denoising methods have garnered significant research attention in recent years, for this kind of method reduces the requirement of large training datasets. Compared to supervised methods, self-supervised methods rely more on the prior embedded in deep networks themselves. As a result, most of the self-supervised methods are designed with Convolution Neural Networks (CNNs) architectures, which well capture one of the most important image prior, translation equivariant prior. Inspired by the great success achieved by the introduction of translational equivariance, in this paper, we explore the way to further incorporate another important image prior. Specifically, we first apply high-accuracy rotation equivariant convolution to self-supervised image denoising. Through rigorous theoretical analysis, we have proved that simply replacing all the convolution layers with rotation equivariant convolution layers would modify the network into its rotation equivariant version. To the best of our knowledge, this is the first time that rotation equivariant image prior is introduced to self-supervised image denoising at the network architecture level with a comprehensive theoretical analysis of equivariance errors, which offers a new perspective to the field of self-supervised image denoising. Moreover, to further improve the performance, we design a new mask mechanism to fusion the output of rotation equivariant network and vanilla CNN-based network, and construct an adaptive rotation equivariant framework. Through extensive experiments on three typical methods, we have demonstrated the effectiveness of the proposed method. The code is available at: https://github.com/liuhanze623/AdaReNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "content": "During the process of capturing and transmitting images, unexpected noises are frequently introduced [16, 28]. Such noise can severely degrade image quality and disrupt subsequent image processing tasks. Image denoising is a tech-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 395, + 306 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 395, + 306 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 395, + 306 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 395, + 306 + ], + "type": "image", + "image_path": "031d15c4d8f19aae03236bff2d1623d9e7b08992e58ce746422a9b091da14bed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 307, + 375, + 314 + ], + "lines": [ + { + "bbox": [ + 334, + 307, + 375, + 314 + ], + "spans": [ + { + "bbox": [ + 334, + 307, + 375, + 314 + ], + "type": "text", + "content": "(a) Input image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 396, + 229, + 473, + 306 + ], + "blocks": [ + { + "bbox": [ + 396, + 229, + 473, + 306 + ], + "lines": [ + { + "bbox": [ + 396, + 229, + 473, + 306 + ], + "spans": [ + { + "bbox": [ + 396, + 229, + 473, + 306 + ], + "type": "image", + "image_path": "82dee4da83fe713d603bbe8467bb6e8460ead0480fdf9a923f01e053748543e7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 307, + 455, + 314 + ], + "lines": [ + { + "bbox": [ + 413, + 307, + 455, + 314 + ], + "spans": [ + { + "bbox": [ + 413, + 307, + 455, + 314 + ], + "type": "text", + "content": "(b) CNN output" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 474, + 229, + 551, + 306 + ], + "blocks": [ + { + "bbox": [ + 474, + 229, + 551, + 306 + ], + "lines": [ + { + "bbox": [ + 474, + 229, + 551, + 306 + ], + "spans": [ + { + "bbox": [ + 474, + 229, + 551, + 306 + ], + "type": "image", + "image_path": "a6e72eba01528aa8564a56f09f52d5397e01e56f6fc3a12ca51c691363bb8afd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 307, + 539, + 314 + ], + "lines": [ + { + "bbox": [ + 482, + 307, + 539, + 314 + ], + "spans": [ + { + "bbox": [ + 482, + 307, + 539, + 314 + ], + "type": "text", + "content": "(c) Equ. conv. output" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 326, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 360 + ], + "type": "text", + "content": "Figure 1. Illustration of the output feature map of a typical image obtained by standard CNN and our used rotation equivariant convolution neural network. Both networks are initialized randomly." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 367, + 555, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 428 + ], + "type": "text", + "content": "nique designed to address the standard inverse problem in image processing and is widely utilized across various fields. With the rapid advancement of deep learning (DL), denoisers based on different deep networks have achieved outstanding results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 433, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 540 + ], + "type": "text", + "content": "Early, it was believed that the success of deep learning is primarily due to the exhaustive utilization of training data. Thus most early DL-based image denoising are in supervised manner [2, 7, 14, 20, 34, 42, 56, 57], where models are trained on extensive datasets that include pairs of clean and noisy images, thereby learning the transformation from noisy to clean images. However, in reality, compiling such comprehensive datasets is both costly and time-intensive, posing substantial challenges [1, 35, 50]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": "Self-supervised approaches thus achieved significant research attention in recent years. These approaches rely less on extensive supervised datasets, but pay more attention to the full utilization of the prior information inherent in deep networks [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Notably, Lehtinen et al. introduced the innovative self-supervised learning approach called Noise2Noise [26], which enables training on pairs of noisy images depicting the identical scene. [26] was the first self-supervised algorithm that achieved performance comparable to supervised image denoising. Other methods involve training denoising models on single noisy images by designing blind-spot networks [3, 21, 22, 48], while some approaches devise strategies to generate training image pairs from noisy im" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 150, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12720" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "ages [6, 18, 32, 33, 51] and have achieved excellent results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 86, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 86, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 86, + 295, + 314 + ], + "type": "text", + "content": "Nowadays, it has become common sense that the prior information inherent in deep networks plays an important role in improving the performance of DL-based methods in image processing tasks [11, 39, 46, 49]. The most classic example should be the CNN, which well captures the translation-equivariant prior of natural images. In CNNs, shifting an input image of CNN is equivalent to shifting all of its intermediate feature maps and the output image. Compared with fully-connected neural networks, this translational equivariance property brings in rational weight sharing, makes the network parameters being used more efficiently, and thus leads to substantially better performance. Inspired by the significant success brought about by the introduction of translational equivariance, another essential image prior (As illustrated in Fig. 1), rotation-equivariant prior has to be taken into consideration very recently. Rotation-equivariant CNNs have been applied and achieved notable performance improvement in multiple supervised image processing tasks [11, 49]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 434 + ], + "type": "text", + "content": "Compared with supervised learning, self-supervised learning approaches indeed depend more on the prior knowledge embedded within the network, for the information that can be learned from the dataset is less. As a result, most current self-supervised image denoising methods are all based on the CNN architectures, which is largely due to the reliance on the prior of translational equivariance. Therefore, it holds even greater significance to incorporate rotational equivariant design into the deep networks for self-supervised methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 639 + ], + "type": "text", + "content": "However, there are two critical issues that need to be addressed when introducing equivariant priors into self-supervised image denoising. On the one hand, although, Xie et al. have recently constructed a rotation equivariant CNN suitable for image processing [49] and Fu et al. further analyzed the global rotation equivariance error of the entire network [11], current rotation equivariant designs for image processing are usually in ResNet[15] structure without upsampling or downsampling. For example, for the superresolution network in [49] where upsampling module is inevitable, only the portion of the network before upsampling is rotation equivariant and the equivariance for the entire network can not be guaranteed. Since self-supervised networks often utilize U-Net structures, which include multiple upsampling and downsampling modules, it is imperative to first explore the impact of these modules on rotation equivariance of the network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "On the other hand, rotation equivariant design for deep networks inevitably introduces parameter sharing and convolution kernel parameterization, which usually degrades the representation accuracy of the network. However, image denoising tasks often have high requirements for the network's representation accuracy, since important high" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "frequency components need to be reconstructed. Besides, not all areas of natural images strictly comply with rigid rotation equivariance, indiscriminately adopting a rotation equivariant network for the entire image can often be detrimental to the reconstruction performance of images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 133, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 180 + ], + "type": "text", + "content": "This study primarily focuses on integrating rotation equivariant prior to existing self-supervised denoising techniques while solving the aforementioned issues. The key contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": "- We explore the way for introducing rotation equivariant prior into self-supervised image denoising frameworks at the network architecture level. Particularly, for the first time, we rigorously analyze the impact of upsampling and downsampling on equivariant networks theoretically. The equivariant errors of upsampling and downsampling layer indeed approach zero when the resolution of the input image increases, though the approach rate " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "(O(h), h" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": " denoted the mesh size of image) is slower than equivariant convolutional layer (whose equivariant error approaches zero at a rate of " + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "inline_equation", + "content": "O(h^{2})" + }, + { + "bbox": [ + 313, + 182, + 555, + 373 + ], + "type": "text", + "content": "). Then, by taking U-Net as an example, we further analyze the rotation equivariant error of the entire network, showing that by simply replacing all the convolution layers with rotation equivariant convolutions [49], we can indeed achieve a reliable rotation equivariant network for self-supervised image denoising." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 374, + 554, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 374, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 374, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 374, + 554, + 481 + ], + "type": "text", + "content": "- We have further developed an adaptive rotation equivariant network to enhance the representation accuracy. Specifically, we design a fusion module for integrating the advantages of both rotation-equivariant and nonequivariant networks. The module can automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. This design provides greater flexibility and yields improved denoising results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": "- We conducted comprehensive experiments across various self-supervised denoising methods, demonstrating the effectiveness of integrating rotation-equivariant image priors into neural networks for self-supervised techniques. Our approach provides a novel perspective in the field of self-supervised image denoising." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 567, + 513, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 513, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 513, + 580 + ], + "type": "text", + "content": "2. Related Work and Prior Knowledge" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 587, + 416, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 416, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 416, + 600 + ], + "type": "text", + "content": "2.1. Image Denoising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": "Non-learning Denoising Methods: Conventional denoising methods predominantly depend on the statistical characteristics of images and mathematical modeling, often employing image priors instead of learned denoisers. Notably, NLM [4, 5] and BM3D [9, 29] have been proposed that can effectively remove noise based on the exploitation of image self-similarity. WNNM [12, 13] treated image denoising as a low-rank matrix approximation problem and achieved great performance." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12721" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "Supervised Denoising Methods: The majority of DL-based denoising algorithms are supervised [2, 7, 14, 52, 56, 57]. DnCNN [56] employed an end-to-end training strategy to learn the mapping between noisy images and the residual components, markedly improving denoising efficacy. FFDNet [57] proposed a versatile denoising architecture by incorporating the noise level as a network parameter. Nevertheless, to train a proficient model, an extensive collection of paired noisy and clean images is necessary to comprehensively capture the range of image contents and noise variations [2, 14, 19, 25, 37, 53]. The acquisition of such training datasets can be prohibitively costly and challenging, and in the case of certain medical images, it may be virtually unattainable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 529 + ], + "type": "text", + "content": "Self-supervised Denoising Methods: Self-supervised deep learning methods have garnered significant interest due to their independence from clean reference images [3, 6, 18, 21, 26, 32, 33, 44, 51, 55]. Noise2Noise [26], proposed by Lehtinen et al., was the first self-supervised method that achieved performance comparable to supervised methods only using paired noisy images. Noise2Void [21] and Noise2Self [3] used a single noisy image and employed a blind-spot network to predict the clean pixel values based on neighboring pixels, thereby avoiding collapse to an identity mapping. [6, 18, 32, 33, 51] devise strategies to generate training image pairs from noisy images to train a network. R2R [33] has demonstrated that the cost function defined on the noisy/noisy image pairs generated by this method is statistically equivalent to the supervised method. Single-image denoising methods [36, 41] capitalize on the statistical properties of the image itself, showing powerful denoising capabilities without extensive training datasets. These methods lack clean images for supervision and rely more on the prior knowledge embedded within the network, which is the reason for the success of these CNN-based self-supervised image denoising methods. Inspired by this, we explore the way to further incorporate rotation equivariant image prior to these methods with rigorous theoretical analysis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 538, + 185, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 185, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 185, + 551 + ], + "type": "text", + "content": "2.2. Rotation Equivariance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 556, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 628 + ], + "type": "text", + "content": "Equivariant Convolution Neural Networks (ECNNs) have drawn substantial interest within the field of computer vision these years [8, 17, 30, 39, 40, 45-47, 58]. Their key strength stems from their ability to effectively handle image rotations through architectural design, which significantly enhances the model's generalization and robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Cohen et al. introduced Group Equivariant Convolutional Networks (G-CNNs) [8] and first integrated " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\pi}{2}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " degree rotation equivariance into the neural network. Recently, the filter parametrization technique has been widely employed and Weiler et al. used harmonics as steerable filters to achieve exact equivariance [45, 46]. However, a notable limitation of these filter parameterization techniques" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "is the inaccuracy in their representation, which significantly affects low-level vision tasks. Fortunately, Xie et al. addressed this issue by proposing Fourier series expansion-based filter parametrization, which has relatively high expression accuracy [49]. The proposed Fconv exhibits precise equivariance in the continuous domain, degrading to approximation only after discretization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 157, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 287 + ], + "type": "text", + "content": "Equivariant convolutions, distinct from data augmentation techniques, inherently integrate rotational symmetry image priors into the network architecture, thereby guaranteeing the network's inherent equivariance and offering superior interpretability and generalizability. However, in the U-Net structure, there is no theoretical guarantee for the design of rotation equivariance. To the best of our knowledge, this study represents the first application of rotation equivariant convolution in the field of self-supervised image denoising and offers theoretical assurance of equivariance for networks based on the U-Net architecture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 293, + 512, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 512, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 512, + 306 + ], + "type": "text", + "content": "2.3. Prior Knowledge about Equivariance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "text", + "content": "Equivalence to a transformation indicates that the application of a transformation to the input results in a corresponding, predictable transformation of the output [39, 45, 49]. Concretely, consider a mapping " + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "text", + "content": " that transforms the input feature space to the output feature space, and let " + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "text", + "content": " denote a set of transformations. For any " + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "inline_equation", + "content": "g\\in G" + }, + { + "bbox": [ + 313, + 311, + 554, + 395 + ], + "type": "text", + "content": ", the following relationship holds:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 387, + 396, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 396, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 387, + 396, + 553, + 410 + ], + "type": "interline_equation", + "content": "\\Psi \\left[ \\pi_ {g} [ f ] \\right] = \\pi_ {g} ^ {\\prime} [ \\Psi [ f ] ], \\tag {1}", + "image_path": "22158fb504d98030bbd58c569714773351591886ff08c9bb433afb28ffb3796d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "text", + "content": " represents any feature map within the input feature space, and " + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "inline_equation", + "content": "\\pi_g" + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "inline_equation", + "content": "\\pi_g'" + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "text", + "content": " describe the action of the transformation " + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 415, + 554, + 452 + ], + "type": "text", + "content": " on the input and output features, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 421, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 421, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 421, + 474 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 480, + 544, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 544, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 544, + 493 + ], + "type": "text", + "content": "3.1. ECNNs for Self-supervised Image Denoising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": "Self-supervised image denoising methods lack clean images for supervision and rely more on the prior knowledge embedded within the network. Therefore, introducing rotation equivariant image prior to this field is reasonable, which can be achieved by ECNNs [49]. The U-Net can effectively restore high-frequency details in images with an encoder-decoder architecture, making it widely used in this task. Consequently, to construct an equivariant self-supervised denoising network, we need to discuss the impact of the essential upsampling and downsampling operators in the U-Net network on equivariance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 631, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 554, + 712 + ], + "type": "text", + "content": "In the subsequent section, we first present some notations and concepts, and then theoretically analyze the impact of upsampling and downsampling operators on equivariant networks. Furthermore, we deduce the equivariant error for the complete U-Net network and provide theoretical guarantees for the implementation of equivariance in U-Net architectures." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12722" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 165 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 165 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 165 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 165 + ], + "type": "image", + "image_path": "0bb10b52cb4f3715d20db0c3873d303d685ca87942634912483bed24003a532f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 173, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 240 + ], + "type": "text", + "content": "Figure 2. The network architecture of the equivariant N2N method. The network can be divided into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 186, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 186, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 186, + 259 + ], + "type": "text", + "content": "3.1.1. Notations and Concepts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 295, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 286 + ], + "type": "text", + "content": "We first introduce some necessary notations and preliminaries as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": "We consider the equivariance on the orthogonal group " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "O(2)" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": ". Formally, " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "O(2) = \\{A \\in \\mathbb{R}^{2 \\times 2} | A^T A = I_{2 \\times 2}\\}" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": ", which contains all rotation and reflection matrices. Without ambiguity, we use " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": " to parameterize " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "O(2)" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": ". The Euclidean group " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "E(2) = \\mathbb{R}^2 \\rtimes O(2)" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\rtimes" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": " is a semidirect-product), whose element is represented as " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "(x, A)" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": ". Restricting the domain of " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": ", we can also use this representation to parameterize any subgroup of " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "E(2)" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": ". In practice, the subgroup is usually assumed to contain " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": " rotations with " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\frac{2\\pi}{t}" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": " degree for an integer " + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{N}_+" + }, + { + "bbox": [ + 55, + 288, + 296, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": "An image " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "I \\in R^{n \\times n}" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " is viewed as a two-dimensional discretization of a smooth function " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "r: \\mathbb{R}^2 \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": ", at the cell-center of a regular grid with " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " cells, i.e., for " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "i, j = 1, 2, \\dots, n" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 148, + 455, + 294, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 455, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 148, + 455, + 294, + 468 + ], + "type": "interline_equation", + "content": "I _ {i j} = r \\left(x _ {i j}\\right), \\tag {2}", + "image_path": "c80fa46647d290310da88218c02234df3f39fefa662b97813d347cdeb91f2d2d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "type": "inline_equation", + "content": "x_{ij} = \\left(\\left(i - \\frac{n + 1}{2}\\right)h, \\left(j - \\frac{n + 1}{2}\\right)h\\right)^T" + }, + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 474, + 295, + 490 + ], + "type": "text", + "content": " is the mesh size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": "An intermediate feature map " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{n \\times n \\times t}" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": " in equivariant networks is a multi-channel tensor, which can be viewed as the discretization of a continuous function defined on " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\tilde{E} = \\mathbb{R}^2 \\rtimes S" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": " is a subgroup of " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "O(2)" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": " is the number of elements in " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": ". Formally, " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": " can be represented as a three-dimensional grid tensor sampled from a smooth function " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": ", i.e., for " + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "inline_equation", + "content": "i, j = 1, 2, \\dots, n" + }, + { + "bbox": [ + 55, + 490, + 296, + 574 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 579, + 295, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 579, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 140, + 579, + 295, + 596 + ], + "type": "interline_equation", + "content": "F _ {i j} ^ {\\backslash A} = e (x _ {i j}, A), \\tag {3}", + "image_path": "8717d415d8ed6b5dd6f26560abf38e7cc77a68e4e7dd2a601c2e27e20e51f30a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "spans": [ + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "type": "text", + "content": " is defined in (2) and " + }, + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "type": "inline_equation", + "content": "A \\in S" + }, + { + "bbox": [ + 58, + 604, + 214, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "content": "With above notations, the transformations on the input and feature maps can be mathematically formulated. Specifically, in the continuous domain, for an input " + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "inline_equation", + "content": "r \\in C^{\\infty}(\\mathbb{R}^{2})" + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "content": " and feature map " + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "inline_equation", + "content": "e \\in C^{\\infty}(E(2))" + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "content": ", the transformation " + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{A} \\in O(2)" + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "content": " acts on " + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 615, + 295, + 674 + ], + "type": "text", + "content": " respectively by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 684, + 294, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 684, + 294, + 716 + ], + "spans": [ + { + "bbox": [ + 69, + 684, + 294, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\pi_ {\\tilde {A}} ^ {R} [ r ] (x) = r (\\tilde {A} ^ {- 1} x), \\forall x \\in \\mathbb {R} ^ {2}, \\\\ \\pi_ {\\tilde {A}} ^ {E} [ e ] (x, A) = e \\left(\\tilde {A} ^ {- 1} x, \\tilde {A} ^ {- 1} A\\right), \\forall (x, A) \\in E (2). \\\\ \\end{array}", + "image_path": "d5cb5e696a0e318f98d92146717652a4e81c194c08ebec8d1fadd6fa66073fbd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "text", + "content": "In particular, if " + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "inline_equation", + "content": "A_{\\theta} \\in O(2)" + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "text", + "content": " is the rotation matrix " + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\left[ \\begin{array}{c}\\cos \\theta ,\\sin \\theta \\\\ -\\sin \\theta ,\\cos \\theta \\end{array} \\right]" + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "text", + "content": ", then the corresponding rotation operators can be expressed by " + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}^{R}" + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}^{E}" + }, + { + "bbox": [ + 314, + 72, + 553, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 122, + 554, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 554, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 554, + 157 + ], + "type": "text", + "content": "Besides, in the discrete domain, we can also define the transformation " + }, + { + "bbox": [ + 313, + 122, + 554, + 157 + ], + "type": "inline_equation", + "content": "\\tilde{A} \\in S" + }, + { + "bbox": [ + 313, + 122, + 554, + 157 + ], + "type": "text", + "content": " on the input image and feature map as followings:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 167, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 167, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 376, + 167, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\tilde {\\pi} _ {\\bar {A}} ^ {R} (I)\\right) _ {i j} = \\pi_ {\\bar {A}} ^ {R} [ r ] (x _ {i j}), \\\\ \\left(\\tilde {\\pi} _ {\\tilde {A}} ^ {E} (F)\\right) _ {i j} ^ {A} = \\pi_ {\\tilde {A}} ^ {E} [ e ] \\left(x _ {i j}, A\\right), \\tag {5} \\\\ \\forall i, j = 1, 2, \\dots , n, A \\in S. \\\\ \\end{array}", + "image_path": "3c4957bff65bb1e7b096ededb09a1b3ae192871483e072090ead2ab1b67d8a58.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "type": "text", + "content": "Similarly, rotation operators can be denoted as " + }, + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "type": "inline_equation", + "content": "\\tilde{\\pi}_{\\theta}^{R}" + }, + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "type": "inline_equation", + "content": "\\tilde{\\pi}_{\\theta}^{E}" + }, + { + "bbox": [ + 313, + 232, + 552, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 253, + 549, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 549, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 549, + 266 + ], + "type": "text", + "content": "3.1.2. Equivalence of Downsampling and Upsampling" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 268, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 554, + 364 + ], + "type": "text", + "content": "As shown in Fig. 2, the U-Net architecture commonly used in self-supervised image denoising typically integrates upsampling and downsampling operators. In specific applications, these layers will inevitably affect the equivariance of the network. Recent studies [11, 49] have analyzed the equivariance properties of the group convolution layer. However, the effects of upsampling and downsampling on network equivariance remain unexplored." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 365, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 388 + ], + "type": "text", + "content": "First of all, We provide the definitions of commonly used downsampling methods in the continuous domain." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 389, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 472 + ], + "type": "text", + "content": "Maxpooling Downsampling. Maxpooling is a commonly used downsampling method in CNNs, which reduces the spatial dimensions of feature maps by sliding a fixed-size window over the feature map and selecting the maximum value within each region as the output [23]. In the continuous domain, we can define maxpooling operator " + }, + { + "bbox": [ + 313, + 389, + 554, + 472 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 313, + 389, + 554, + 472 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 374, + 474, + 553, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 474, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 374, + 474, + 553, + 489 + ], + "type": "interline_equation", + "content": "[ M (F) ] (x, A) = \\max _ {\\Omega_ {x}} F _ {i j} ^ {A}, \\tag {6}", + "image_path": "13d2ec9ea4c0d9c5c7a4557e4a17610e39b2af6a7305c43c33fe2b390519f9fb.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "inline_equation", + "content": "x = [x_{1},x_{2}]^{T}\\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": " denotes the spatial coordinates, and " + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "inline_equation", + "content": "x_{1}\\in [i,i + 1]" + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "inline_equation", + "content": "x_{2}\\in [j,j + 1]" + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\Omega_{x} = \\{(i,j),(i + 1,j),(i,j + 1),(i + 1,j + 1)\\}" + }, + { + "bbox": [ + 313, + 496, + 553, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "type": "text", + "content": "Stride Downsampling. Stride Downsampling is also a widely used downsampling operator which reduces the size of the feature map by adjusting the stride of the convolution operation [24]. In the continuous domain, we can define stride downsampling operator " + }, + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 313, + 533, + 554, + 594 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 385, + 603, + 553, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 603, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 385, + 603, + 553, + 618 + ], + "type": "interline_equation", + "content": "[ S (F) ] (x, A) = F _ {i, j + 1} ^ {A}, \\tag {7}", + "image_path": "954c41c1835a0ed7278deff23fb38fcb9082084a89f16a7cf6ef4b77ee97da25.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "inline_equation", + "content": "x = [x_{1},x_{2}]^{T}\\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "text", + "content": " denotes the spatial coordinates, and " + }, + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "inline_equation", + "content": "x_{1}\\in [i,i + 1],x_{2}\\in [j,j + 1]" + }, + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "inline_equation", + "content": "\\Omega_{x} = \\{(i,j),(i+" + }, + { + "bbox": [ + 313, + 628, + 553, + 665 + ], + "type": "inline_equation", + "content": "1,j),(i,j + 1),(i + 1,j + 1)\\} ." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "For the above two common downsampling operators, it is of great significance to analyze their impact on the equivariant network. Therefore, we constructed their equivariant errors under the rotational equivariant structure." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12723" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": "Theorem 1 Assume that a feature map " + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{n \\times n \\times t}" + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": " is discretized from the smooth function " + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}" + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "|S| = t" + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": ", the mesh size is " + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": " is the downsampling operator. If for any " + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "inline_equation", + "content": "A, B \\in S, x \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 54, + 72, + 296, + 131 + ], + "type": "text", + "content": ", the following conditions are satisfied:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 136, + 132, + 295, + 145 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 132, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 136, + 132, + 295, + 145 + ], + "type": "interline_equation", + "content": "\\| \\nabla e (x, A) \\| \\leq G, \\tag {8}", + "image_path": "e87e885b172e52003052a958e59cfee900a4b3bd8f0afd710686c74de8879341.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 214, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 214, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 214, + 161 + ], + "type": "text", + "content": "then the following results are satisfied:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 166, + 295, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 295, + 182 + ], + "type": "interline_equation", + "content": "\\left| D \\left[ \\tilde {\\pi} _ {B} ^ {E} \\right] (F) (x, A) - \\pi_ {B} ^ {E} [ D (F) ] (x, A) \\right| \\leq 2 \\sqrt {2} G h. \\tag {9}", + "image_path": "6dae9cc17308ec81588a814ce2a611d2fa807d4030878e90a04fa5b216007a7c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "content": "The downsampling operator " + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "content": " can be either " + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "inline_equation", + "content": "M(\\cdot)" + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "content": ". Theorem 1 reveals that the equivariant error of the downsample operator is primarily influenced by the mesh size " + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "content": ", and it indeed approach zero when the resolution of the input image increases with the approach rate " + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "inline_equation", + "content": "O(h)" + }, + { + "bbox": [ + 55, + 187, + 296, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 247, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 270 + ], + "type": "text", + "content": "Then, we provide the definitions of commonly used upsampling methods in the continuous domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "content": "Nearest Neighbor Upsampling. Nearest neighbor interpolation is an image scaling method that fills the pixels of the interpolated image by selecting the original pixel value closest to the target pixel position. In the continuous domain, we can define the nearest neighbor operator " + }, + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "inline_equation", + "content": "N(\\cdot)" + }, + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 342, + 294, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 342, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 127, + 342, + 294, + 357 + ], + "type": "interline_equation", + "content": "[ N (F) ] (x, A) = F _ {i ^ {*} j ^ {*}} ^ {A}, \\tag {10}", + "image_path": "7d57328d98c7c2b8de26d6082850d55d1f056a213470de58fa71012ed6f190cd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 359, + 219, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 219, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 219, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 359, + 219, + 372 + ], + "type": "inline_equation", + "content": "(i^{*},j^{*}) = arg\\min_{ij}||x_{ij} - x||_2^2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 372, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 430 + ], + "type": "text", + "content": "Bilinear Upsampling. Bilinear interpolation calculates the new pixel value by taking the weighted average of the four surrounding known pixel values. In the continuous domain, we can define the bilinear interpolation operator " + }, + { + "bbox": [ + 55, + 372, + 296, + 430 + ], + "type": "inline_equation", + "content": "B(\\cdot)" + }, + { + "bbox": [ + 55, + 372, + 296, + 430 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 430, + 295, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 430, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 102, + 430, + 295, + 463 + ], + "type": "interline_equation", + "content": "[ B (F) ] (x, A) = \\sum_ {i = 1} ^ {2} \\sum_ {j = 1} ^ {2} \\lambda_ {i j} f \\left(Q _ {i j}\\right), \\tag {11}", + "image_path": "9fa8cf3cce2ccfa6ccd50e6b38cb499e87dc950715edb157fab789305098f544.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\lambda_{ij}" + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "content": " are the coefficients of bilinear interpolation and " + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "inline_equation", + "content": "f(Q_{ij})" + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "content": " represent the grid points, " + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "inline_equation", + "content": "x = [x_1,x_2]^T\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "content": " denotes the 2D spatial coordinates, " + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "inline_equation", + "content": "x_{1}\\in [i,i + 1]" + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "inline_equation", + "content": "x_{2}\\in [j,j + 1]" + }, + { + "bbox": [ + 55, + 466, + 296, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 514, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 585 + ], + "type": "text", + "content": "Both of the aforementioned upsampling operators are widely utilized across various network architectures, making it essential to analyze their mathematical properties within rotational equivariant networks. Accordingly, we evaluated their equivariant errors under a rotational equivariant framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "content": "Theorem 2 Assume that a feature map " + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{n \\times n \\times t}" + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "content": " is discretized from the smooth function " + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "e: \\mathbb{R}^2 \\times S \\to \\mathbb{R}, |S| = t" + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "content": ", the mesh size is " + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "content": " is the upsampling operator. If for any " + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "inline_equation", + "content": "A, B \\in S, x \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 54, + 593, + 296, + 653 + ], + "type": "text", + "content": ", the following conditions are satisfied:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 136, + 653, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 653, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 136, + 653, + 295, + 666 + ], + "type": "interline_equation", + "content": "\\| \\nabla e (x, A) \\| \\leq G, \\tag {12}", + "image_path": "91892bc08ab3b71e062366dfa11178e8e2e1f7329b71df6a2ee8f430c0249141.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 670, + 214, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 214, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 214, + 681 + ], + "type": "text", + "content": "then the following results are satisfied:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 688, + 296, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 688, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 296, + 713 + ], + "type": "interline_equation", + "content": "\\left| U \\left[ \\tilde {\\pi} _ {B} ^ {E} \\right] (F) (x, A) - \\pi_ {B} ^ {E} [ U (F) ] (x, A) \\right| \\leq 2 (\\sqrt {2} + 1) G h. \\tag {13}", + "image_path": "765ec4e9494fd9b7421912fc325dce71a7a380cb6f6fe4ba007c5cc6154587da.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "The upsampling operator " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "U(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " can be either " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "N(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "B(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". It is worth noting that the above conclusions indicate that the error introduced by the upsampling operator in the rotational equivariant network is related to " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", aligning with established understanding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 137, + 501, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 501, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 501, + 148 + ], + "type": "text", + "content": "3.1.3. Analysis of Complete U-Net Network" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 151, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 555, + 281 + ], + "type": "text", + "content": "We take the network design of the N2N method [26] as an example to give the derivation of the rotation equivariant error for the entire U-Net architecture. As shown in Fig. 2, we decompose the network into multiple upsampling and downsampling blocks. Each downsampling block (DB) consists of one E-Conv layer (Equivariant Convolution) and a downsampling operator, while each upsampling block (UB) is composed of an upsampling operator and two E-Conv layers, we provide the equivariant error for each block and subsequently derive the equivariant error bounds for the complete network." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": "Theorem 3 demonstrates the equivariance error of the complete U-Net network under discrete angles, and the corollary further provides the equivariance error of the complete U-Net network under any rotation angle. Note that the mesh size of upsampling and downsampling are different, we define the mesh size of the original picture to be " + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": ", the mesh size after a " + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": " downsampling is " + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "inline_equation", + "content": "2h" + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": ", and so on." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": "Theorem 3 For an image " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "H\\times W\\times n_0" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", and a " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": "-layer rotation equivariant " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": "-Net network " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{UNet}_{eq}(\\cdot)" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " whose channel number of the " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " layer is " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "n_l" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", rotation equivariant subgroup is " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "S\\leqslant O(2),|S| = t" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", and activation function is set as ReLU. If the latent continuous function of the " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "c^{th}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " channel of " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " denoted as " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "r_c:\\mathbb{R}^2\\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", and the latent continuous function of any convolution filters in the " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " layer denoted as " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\phi^l:\\mathbb{R}^2\\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "DB_{i}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "UB_{i}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " represent the downsampling block and the upsampling block, respectively. " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\hat{\\Upsilon}" + }, + { + "bbox": [ + 313, + 377, + 555, + 510 + ], + "type": "text", + "content": " represent the convolutional layers in the input, middle, and output stages, respectively. We define:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 513, + 554, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 513, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 315, + 513, + 554, + 541 + ], + "type": "interline_equation", + "content": "\\operatorname {U N e t} _ {e q} (\\cdot) = \\hat {\\Upsilon} \\left[ \\hat {U} B _ {m} \\dots \\hat {U} B _ {1} \\left[ \\hat {\\Phi} \\left[ \\hat {D} B _ {m} \\dots \\hat {D} B _ {1} \\left[ \\hat {\\Psi} \\right] \\right] \\dots \\right] (\\cdot), \\right. \\tag {14}", + "image_path": "81782741ae89664adfb4e9e5253c44a6bfd707ff0fc0ce993edb15bd1f78d158.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 542, + 468, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 468, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 468, + 554 + ], + "type": "text", + "content": "the following conditions are satisfied:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 558, + 554, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 558, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 321, + 558, + 554, + 604 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| r _ {c} (x) \\right| \\leq F _ {0}, \\| \\nabla_ {x} r _ {c} (x) \\| \\leq G _ {0}, \\| \\nabla_ {x} ^ {2} r _ {c} (x) \\| \\leq H _ {0}, \\\\ \\left| \\phi^ {l} (x) \\right| \\leq F _ {l}, \\| \\nabla_ {x} \\phi^ {l} (x) \\| \\leq G _ {l}, \\| \\nabla_ {x} ^ {2} \\phi^ {l} (x) \\| \\leq H _ {l}, \\tag {15} \\\\ \\forall \\| x \\| \\geq (p + 1) h / 2, \\phi_ {l} (x) = 0, \\\\ \\end{array}", + "image_path": "5cdad615008579e37fdb2a268b61502db03411aac26b5cf038478505e0d9d142.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": " is the filter size, " + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": " is the mesh size, " + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\theta_{k} = \\frac{2k\\pi}{t}" + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots ,t" + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\nabla_{x}" + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\nabla_{x}^{2}" + }, + { + "bbox": [ + 313, + 610, + 554, + 647 + ], + "type": "text", + "content": " denote the operators of gradient and Hessian matrix, respectively. We have" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 651, + 553, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 651, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 651, + 553, + 677 + ], + "type": "interline_equation", + "content": "\\left| \\mathrm {U N e t} _ {e q} \\left[ \\tilde {\\pi} _ {\\theta_ {k}} ^ {R} \\right] (X) - \\tilde {\\pi} _ {\\theta_ {k}} ^ {R} \\left[ \\mathrm {U N e t} _ {e q} \\right] (X) \\right| \\leq R _ {1} h + R _ {2} h ^ {2}. \\tag {16}", + "image_path": "444311bb846cd4757021bcdcf4f0c1954d27b72e3b9c5111e555f6ef528d9673.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "R_{1}, R_{2}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " are two constants with respect to " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "N, n_{l}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " and the upper bound in (15), their specific values can be found in the supplementary materials." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12724" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 556, + 255 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 556, + 255 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 556, + 255 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 556, + 255 + ], + "type": "image", + "image_path": "ca00c297445f151b65f928c399921d23eb8f3c23b76c42722455854486b3af48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. Illustrations of our proposed adaptive network AdaReNet. Specifically, " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{\\mathrm{H} \\times \\mathrm{W} \\times \\mathrm{C}}" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " represents a noisy image, where " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " represent the spatial dimensions, and " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " denotes the channel dimension. The Vanilla Module and EQ Module each produce their respective preliminary denoising results, denoted as " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "f_c" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "f_e" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": ". The Fusion Module " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "Mask(\\cdot)" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " automatically decides which areas of the image to use more EQ Module would gain more benefit. After adaptive fusion by " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "Mask(\\cdot)" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " and correction by the Self-correcting Module " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "S_c(\\cdot)" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": ", the final denoised image " + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 54, + 261, + 555, + 319 + ], + "type": "text", + "content": " is output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": "Corollary 1 Under the same condition as Theorem 3, for an arbitrary " + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\theta \\in [0,2\\pi ]" + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": " denote the rotation transformation, then " + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\forall \\theta" + }, + { + "bbox": [ + 55, + 338, + 296, + 374 + ], + "type": "text", + "content": " we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 409 + ], + "type": "interline_equation", + "content": "\\left| \\mathrm {U N e t} _ {e q} \\left[ \\tilde {\\pi} _ {\\theta} ^ {R} \\right] (X) - \\tilde {\\pi} _ {\\theta} ^ {R} \\left[ \\mathrm {U N e t} _ {e q} \\right] (X) \\right| \\leq R _ {1} h + R _ {2} h ^ {2} + R _ {3} t ^ {- 1} h, \\tag {17}", + "image_path": "af125de14282593637e53032c8aa5997eddf18ad40858b16c465cf9e4a7978a5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "type": "inline_equation", + "content": "R_{1}, R_{2}, R_{3}" + }, + { + "bbox": [ + 55, + 410, + 296, + 434 + ], + "type": "text", + "content": " are constants that can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 456, + 117, + 515 + ], + "blocks": [ + { + "bbox": [ + 58, + 456, + 117, + 515 + ], + "lines": [ + { + "bbox": [ + 58, + 456, + 117, + 515 + ], + "spans": [ + { + "bbox": [ + 58, + 456, + 117, + 515 + ], + "type": "image", + "image_path": "8f2c85e6f5a7481384f6cadf7273fa12fe9f52168c394457b3782ee1e9b10cde.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 516, + 92, + 523 + ], + "lines": [ + { + "bbox": [ + 82, + 516, + 92, + 523 + ], + "spans": [ + { + "bbox": [ + 82, + 516, + 92, + 523 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 117, + 456, + 176, + 515 + ], + "blocks": [ + { + "bbox": [ + 117, + 456, + 176, + 515 + ], + "lines": [ + { + "bbox": [ + 117, + 456, + 176, + 515 + ], + "spans": [ + { + "bbox": [ + 117, + 456, + 176, + 515 + ], + "type": "image", + "image_path": "311276b789292fa4a963cd43fc3403acd7f9ba656973b27f4455f5b711df2367.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 516, + 152, + 523 + ], + "lines": [ + { + "bbox": [ + 142, + 516, + 152, + 523 + ], + "spans": [ + { + "bbox": [ + 142, + 516, + 152, + 523 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 176, + 456, + 235, + 515 + ], + "blocks": [ + { + "bbox": [ + 176, + 456, + 235, + 515 + ], + "lines": [ + { + "bbox": [ + 176, + 456, + 235, + 515 + ], + "spans": [ + { + "bbox": [ + 176, + 456, + 235, + 515 + ], + "type": "image", + "image_path": "d197b7e75b11fd15a10152a02f592a8b63f043f80544922bf958c0da424558e8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 516, + 211, + 523 + ], + "lines": [ + { + "bbox": [ + 201, + 516, + 211, + 523 + ], + "spans": [ + { + "bbox": [ + 201, + 516, + 211, + 523 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 578 + ], + "type": "text", + "content": "Figure 4. (a) An image from the Kodak dataset, (b) the heatmap of the low-frequency component, (c) the heatmap of the high-frequency component, (d) the output of our proposed MaskNetwork (the brighter area indicates the use of more Vanilla Module)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 235, + 456, + 293, + 515 + ], + "blocks": [ + { + "bbox": [ + 235, + 456, + 293, + 515 + ], + "lines": [ + { + "bbox": [ + 235, + 456, + 293, + 515 + ], + "spans": [ + { + "bbox": [ + 235, + 456, + 293, + 515 + ], + "type": "image", + "image_path": "08f1bfb0ff3129c1b59f28fb1aafee7177e32e0251253f1ea3cdc541044854a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 516, + 269, + 523 + ], + "lines": [ + { + "bbox": [ + 259, + 516, + 269, + 523 + ], + "spans": [ + { + "bbox": [ + 259, + 516, + 269, + 523 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 599, + 263, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 263, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 263, + 613 + ], + "type": "text", + "content": "3.2. Proposed Adaptive Network AdaReNet" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Rotation equivariant design usually degrades the representation accuracy of the network because of parameter sharing and filter parameterization and not all areas of natural images strictly comply with rigid rotation equivariance. In order to solve the above problems, the proposed network predominantly comprises four primary modules, as depicted in Fig. 3, which are the Vanilla Module, EQ Module, Fusion Module, and Self-correcting Module." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "type": "text", + "content": "Vanilla Module. This module utilizes a conventional CNN architecture, maintaining the original design where only translation equivariance is incorporated into the network. We define the result of the input image " + }, + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "type": "text", + "content": " after passing through the Vanilla Module as " + }, + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 313, + 338, + 555, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 405, + 409, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 409, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 405, + 409, + 553, + 421 + ], + "type": "interline_equation", + "content": "f _ {c} = \\operatorname {V M} (I). \\tag {18}", + "image_path": "81662787aac0ec756aec32e0b1cbbfbf7129b434dd28b42a4c8f9916b361975b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": "EQ Module. This module employs the same architecture as the previous one, with the conventional convolution replaced by the rotation equivariant network to incorporate rotation equivariance prior into the architecture. We define the result of input image " + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": " after passing through the EQ Module as " + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 406, + 505, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 505, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 406, + 505, + 553, + 517 + ], + "type": "interline_equation", + "content": "f _ {e} = \\operatorname {E Q} (I). \\tag {19}", + "image_path": "30886321f1a1cdfc47bb46d85a0e3f906638256bc60c4ab8a61abf5bf94f125b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "text", + "content": "Fusion Module. This module introduces a MaskNetwork " + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "inline_equation", + "content": "Mask(\\cdot)" + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "text", + "content": ", a specialized network designed to merge the outputs from the Vanilla Module and the EQ Module. It employs several layers of standard convolutions to automatically determine which regions of the image would benefit more from a rotation-equivariant network compared to a normal CNN-based network. The output of " + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "inline_equation", + "content": "Mask(\\cdot)" + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "text", + "content": " is shown in Fig. 4(d). We define the result of input image " + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "text", + "content": " after passing through MaskNetwork as " + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 313, + 524, + 555, + 632 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 399, + 642, + 553, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 642, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 399, + 642, + 553, + 654 + ], + "type": "interline_equation", + "content": "M _ {f} = \\operatorname {M a s k} (I). \\tag {20}", + "image_path": "d33d48cfd489aaac4b9e04623fff07f0d2b8e697d56542bee197e4e4b133388e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Self-correcting Module. Following the Fusion Module, a self-correcting module is applied to refine the fused result. This is achieved through the use of simple ResNet Blocks. We define the Self-correcting Module as " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "S_{c}(\\cdot)" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12725" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "Based on the network presented in Fig. 3, we can derive the following mathematical representation to encapsulate the training and inference processes of the entire network:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 131, + 295, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 131, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 76, + 131, + 295, + 148 + ], + "type": "interline_equation", + "content": "\\hat {I} = M _ {f} \\odot f _ {c} + (1 - M _ {f}) \\odot f _ {e}, \\bar {I} = S _ {c} (\\hat {I}), \\tag {21}", + "image_path": "e5d674a7325850ad9928973fb87f1245710b84049483a91abf66994d111619df.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "text", + "content": " denotes the results of adaptive fusion, and " + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "text", + "content": " represents the final restoration result after going through the Self-correcting Module, " + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 158, + 296, + 206 + ], + "type": "text", + "content": " denote element-wise multiplication." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 208, + 296, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 296, + 244 + ], + "type": "text", + "content": "Loss Function. We incorporate the loss of the two subnetworks as regularization terms into the main loss. The loss function is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 280 + ], + "type": "interline_equation", + "content": "\\mathrm {L} = \\| \\bar {I} - \\text {t a r g e t} \\| _ {2} + \\alpha_ {1} \\| f _ {c} - \\text {t a r g e t} \\| _ {2} + \\alpha_ {2} \\| f _ {e} - \\text {t a r g e t} \\| _ {2}, \\tag {22}", + "image_path": "13df695e44df06839472b0a4d26ef7a58fc46714b1c643c0d4a01233f6e39b39.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 281, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 281, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + }, + { + "bbox": [ + 55, + 281, + 296, + 304 + ], + "type": "text", + "content": " are hyperparameters and we empirically set " + }, + { + "bbox": [ + 55, + 281, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = \\alpha_{2} = 0.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 306, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 498 + ], + "type": "text", + "content": "Remark. Embedding rotational equivariance into the network is highly effective for self-supervised image denoising task. Since natural images do not adhere to strict rotational equivariance, our proposed adaptive rotational equivariant network, named AdaReNet, can automatically decide which regions of the image to apply the rotation-equivariant network, thereby further enhancing performance. By observing the output of the Fusion Module in Fig. 4(d), it is noted that the mask values are larger at the edge details of the pattern, indicating that the adaptive equivariant network tends to use the outputs from the Vanilla Module more frequently in these areas. Conversely, for the restoration of the majority of low-frequency components in the image, the network predominantly uses the outputs from the EQ Module. This aligns with the common sense that convolutions are more adept at fitting high-frequency information[43]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 513, + 137, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 137, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 137, + 526 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "In this section, we conducted experiments based on the existing setup and validated our method in three classic approaches. For the Noise2Noise [26] and Noise2Void [21] methods, we showed experiments on U-Net [38], which can significantly speed up the training process while maintaining acceptable performance, and demonstrated the effectiveness of our method. Besides, we also conducted the R2R experiments based on DnCNN [33], further validating the superiority of the proposed method. Due to space limitations, we only present partial results. Implementation details and further experiments concerning different models, various datasets, model parameter counts, as well as experiments in the field of self-supervised fluorescence microscopy denoising[27] can be found in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 508, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 508, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 508, + 85 + ], + "type": "text", + "content": "4.1. Experiments on Compared Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "text", + "content": "Rotation Equivariant N2N: We selected the most classical Gaussian noise for our experiments, and randomized the standard deviation " + }, + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\sigma \\in [0,50]" + }, + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "text", + "content": " of noise for each training example individually. N2N-EQ denotes the rotation equivariant network, while N2N-EQ" + }, + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "text", + "content": " serves as the adaptive rotation equivariant network. The notation in other experiments is similar. The results are shown in Tab. 1. The superiority of our method can also be observed from Fig. 5." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 194, + 553, + 237 + ], + "blocks": [ + { + "bbox": [ + 315, + 194, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 315, + 194, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 315, + 194, + 553, + 237 + ], + "type": "table", + "html": "
DatasetGaussian25Gaussian50
N2N [26]N2N-EQN2N-EQ+N2N [26]N2N-EQN2N-EQ+
Kodak [10]31.47/0.87431.60/0.87831.72/0.88028.29/0.77828.58/0.79028.69/0.791
BSD300 [31]30.18/0.86930.28/0.87230.36/0.87327.02/0.76227.24/0.77127.31/0.772
Set14 [54]30.02/0.85130.06/0.85430.19/0.85527.16/0.76827.32/0.77527.44/0.777
", + "image_path": "ffef4fbc996fce9776ca9a9c3814b4d88e958f1c5c3ef29972d54537cbeb587e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "type": "text", + "content": "Rotation Equivariant N2V: We conducted experiments with U-Net architectures. The results are presented in Tab. 2 and Fig. 6. The equivariant error1 of networks N2V, N2VEQ, and " + }, + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "type": "inline_equation", + "content": "\\mathrm{N2V - EQ^{+}}" + }, + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "type": "text", + "content": " are 0.233, 0.068, and 0.076, respectively. This verifies that the improvements are achieved by reducing equivariant errors. Besides, Tab. 3 further performed diverse scenarios trained on the BSD500 dataset. Our method consistently achieved superior results." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 374, + 553, + 426 + ], + "blocks": [ + { + "bbox": [ + 313, + 245, + 555, + 267 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 267 + ], + "type": "text", + "content": "Table 1. N2N: three networks with U-Net architecture were tested under conditions of Gaussian noise at levels 25 and 50." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 374, + 553, + 426 + ], + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 426 + ], + "type": "table", + "html": "
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]28.17/0.82029.05/0.83429.12/0.84526.07/0.72526.38/0.73526.82/0.755
Kodak24 [10]28.86/0.81129.78/0.82529.93/0.83626.75/0.71627.15/0.73227.72/0.754
Set14 [54]27.22/0.80028.04/0.80628.09/0.81625.40/0.71525.65/0.72626.23/0.749
Average28.08/0.81128.96/0.82229.05/0.83226.07/0.71926.39/0.73126.92/0.753
", + "image_path": "d71bb2efadd643e5809733f4b6c5e951579ff0cb20db2ae92f720d128f8cef5f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 316, + 463, + 553, + 501 + ], + "blocks": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "type": "text", + "content": "Table 2. N2V: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 463, + 553, + 501 + ], + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 501 + ], + "type": "table", + "html": "
MethodPoissonPoisson30Poisson&GaussianPeppersaltSpeckle
N2V30.82/0.91232.38/0.96127.91/0.82123.93/0.78226.03/0.775
N2V-EQ(ours)31.22/0.90433.59/0.96628.84/0.83524.53/0.79226.42/0.736
N2V-EQ+(ours)32.19/0.92435.96/0.97629.34/0.85024.83/0.81124.85/0.676
", + "image_path": "ce965cd10e45179dd4d9939556f357eab38320bce3fb7fb813cdf9111496e616.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 508, + 533, + 519 + ], + "lines": [ + { + "bbox": [ + 334, + 508, + 533, + 519 + ], + "spans": [ + { + "bbox": [ + 334, + 508, + 533, + 519 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on diverse scenarios." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": "Rotation Equivariant R2R: The experimental results are shown in Tab. 5. The parameter counts for methods R2R, R2R-EQ, and " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{R2R - EQ^{+}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "0.67\\mathrm{M}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "0.17\\mathrm{M}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "0.84\\mathrm{M}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": ", respectively. Due to the larger number of adaptive network parameters, it is expected that the performance is inferior to our EQ version when the training dataset is small." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 600, + 409, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 409, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 409, + 613 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": "Selection of rotation-equivariant networks: We conducted ablation experiments with mainstream rotation-equivariant networks, including Cohen et al. [8], Weiler et al. [45] and Shen et al. [39]. Tab. 4 confirms that the chosen Fconv [49] approach is the most effective method in the field of low-level vision." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "inline_equation", + "content": "^{1}||[L_{R}\\Phi (f) - \\Phi L_{R}(f)\\|_{2}^{2} / \\| L_{R}\\Phi (f)\\|_{2}^{2}," + }, + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Phi (\\cdot)" + }, + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "text", + "content": " represents the network and " + }, + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{R}(\\cdot)" + }, + { + "bbox": [ + 313, + 692, + 555, + 713 + ], + "type": "text", + "content": " denotes the rotation transformation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12726" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 181, + 156 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 181, + 156 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 181, + 156 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 181, + 156 + ], + "type": "image", + "image_path": "5ee78c577a3df50bb6e388fa230b322be39c148b208478a573bad530bb0dff85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 157, + 132, + 166 + ], + "lines": [ + { + "bbox": [ + 107, + 157, + 132, + 166 + ], + "spans": [ + { + "bbox": [ + 107, + 157, + 132, + 166 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 304, + 156 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 304, + 156 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 304, + 156 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 304, + 156 + ], + "type": "image", + "image_path": "73867c6504ddad7660955e61abdd893746df94c4422a1c2df1385b20913df872.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 157, + 278, + 166 + ], + "lines": [ + { + "bbox": [ + 216, + 157, + 278, + 166 + ], + "spans": [ + { + "bbox": [ + 216, + 157, + 278, + 166 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 216, + 157, + 278, + 166 + ], + "type": "inline_equation", + "content": "\\mathrm{N2N} / 28.82\\mathrm{dB}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 427, + 156 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 427, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 427, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 427, + 156 + ], + "type": "image", + "image_path": "6213ffc2364035cbc42f3849df0eb07d84e560978b3f3abdbfc2c2258b1695b8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 157, + 405, + 166 + ], + "lines": [ + { + "bbox": [ + 331, + 157, + 405, + 166 + ], + "spans": [ + { + "bbox": [ + 331, + 157, + 405, + 166 + ], + "type": "text", + "content": "(c) N2N-EQ / 29.26dB" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 428, + 72, + 552, + 156 + ], + "blocks": [ + { + "bbox": [ + 428, + 72, + 552, + 156 + ], + "lines": [ + { + "bbox": [ + 428, + 72, + 552, + 156 + ], + "spans": [ + { + "bbox": [ + 428, + 72, + 552, + 156 + ], + "type": "image", + "image_path": "3b5a2bd1a9e40d618f220a64c3eff73ffd574047cb77d888301d31bd88fda90b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 156, + 529, + 166 + ], + "lines": [ + { + "bbox": [ + 453, + 156, + 529, + 166 + ], + "spans": [ + { + "bbox": [ + 453, + 156, + 529, + 166 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 453, + 156, + 529, + 166 + ], + "type": "inline_equation", + "content": "\\mathrm{N2N - EQ^{+} / 29.41dB}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 202, + 181, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 202, + 181, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 202, + 181, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 181, + 285 + ], + "type": "image", + "image_path": "b3e1d546fcccbe4bab8f090f386a2f802fd3c6f912ea19b4b80ce061aa3fb473.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 285, + 135, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 135, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 135, + 295 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 307, + 453, + 319 + ], + "lines": [ + { + "bbox": [ + 156, + 307, + 453, + 319 + ], + "spans": [ + { + "bbox": [ + 156, + 307, + 453, + 319 + ], + "type": "text", + "content": "Figure 6. N2V: image denoising results of one image from BSD500 with " + }, + { + "bbox": [ + 156, + 307, + 453, + 319 + ], + "type": "inline_equation", + "content": "\\sigma = 25" + }, + { + "bbox": [ + 156, + 307, + 453, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 182, + 202, + 304, + 285 + ], + "blocks": [ + { + "bbox": [ + 160, + 178, + 448, + 190 + ], + "lines": [ + { + "bbox": [ + 160, + 178, + 448, + 190 + ], + "spans": [ + { + "bbox": [ + 160, + 178, + 448, + 190 + ], + "type": "text", + "content": "Figure 5. N2N: image denoising results of one image from kodak with " + }, + { + "bbox": [ + 160, + 178, + 448, + 190 + ], + "type": "inline_equation", + "content": "\\sigma = 50" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 202, + 304, + 285 + ], + "lines": [ + { + "bbox": [ + 182, + 202, + 304, + 285 + ], + "spans": [ + { + "bbox": [ + 182, + 202, + 304, + 285 + ], + "type": "image", + "image_path": "6e93c5a7eb50567340e261f1fb21988e0e5563f7838029bddd8144c2fbbff87d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 285, + 281, + 294 + ], + "lines": [ + { + "bbox": [ + 208, + 285, + 281, + 294 + ], + "spans": [ + { + "bbox": [ + 208, + 285, + 281, + 294 + ], + "type": "text", + "content": "(b) N2V / 29.97dB" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 202, + 428, + 285 + ], + "blocks": [ + { + "bbox": [ + 306, + 202, + 428, + 285 + ], + "lines": [ + { + "bbox": [ + 306, + 202, + 428, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 428, + 285 + ], + "type": "image", + "image_path": "363f7db599bc4ddb91f4e9fba65acb8101c5ba036f1eabdf9d9820cb33097030.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 285, + 413, + 295 + ], + "lines": [ + { + "bbox": [ + 326, + 285, + 413, + 295 + ], + "spans": [ + { + "bbox": [ + 326, + 285, + 413, + 295 + ], + "type": "text", + "content": "(c) N2V-EQ / 32.65dB" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 202, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 430, + 202, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 430, + 202, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 430, + 202, + 553, + 285 + ], + "type": "image", + "image_path": "bbaeac135546f953abed5acfec3749effa8e5d8e288191e0282beda1a5908ee7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 285, + 536, + 295 + ], + "lines": [ + { + "bbox": [ + 446, + 285, + 536, + 295 + ], + "spans": [ + { + "bbox": [ + 446, + 285, + 536, + 295 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 446, + 285, + 536, + 295 + ], + "type": "inline_equation", + "content": "\\mathrm{N2V - EQ^{+} / 33.47dB}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 56, + 328, + 555, + 415 + ], + "blocks": [ + { + "bbox": [ + 56, + 328, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 328, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 555, + 415 + ], + "type": "table", + "html": "
Gaussian25Gaussian50
MethodKodak24 [10]BSDS300 [31]Set14 [54]AverageKodak24 [10]BSDS300 [31]Set14 [54]Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
CNN31.470.87430.180.86930.020.85130.560.86528.290.77827.020.76227.160.76827.490.769
G-CNN31.390.87230.230.86930.020.85230.550.86428.040.77426.850.75726.880.76227.260.764
E2-CNN31.230.86930.020.86429.700.84530.320.85928.040.76826.860.75426.830.75827.240.760
PDO-eConv31.420.87430.150.86929.830.84930.470.86428.340.78327.060.76527.050.76827.480.772
Fconv31.600.87830.280.87230.060.85430.650.86828.580.79027.240.77127.320.77527.710.779
", + "image_path": "da5e391b8fea6fb1d2e94bfc8a8cb7771c9412fbba64aca3a9cdbc2ffce2c57a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 57, + 454, + 294, + 501 + ], + "blocks": [ + { + "bbox": [ + 106, + 424, + 502, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 502, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 502, + 435 + ], + "type": "text", + "content": "Table 4. Ablation on rotation-equivariant networks: PSNR and SSIM results for Gaussian25 and Gaussian50." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 454, + 294, + 501 + ], + "lines": [ + { + "bbox": [ + 57, + 454, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 294, + 501 + ], + "type": "table", + "html": "
DatasetσR2R [33]R2R-EQR2R-EQ+
BSD682529.03/0.82229.14/0.82529.10/0.826
BSD685026.01/0.70026.14/0.71126.00/0.700
", + "image_path": "512a8b7119259ff3ba66d5813946411b9ea714543d0f6e50ff65d6f8398d465d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 57, + 535, + 294, + 587 + ], + "blocks": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 531 + ], + "type": "text", + "content": "Table 5. R2R: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 535, + 294, + 587 + ], + "lines": [ + { + "bbox": [ + 57, + 535, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 294, + 587 + ], + "type": "table", + "html": "
DatasetGaussian25Gaussian50
N2V [21]N2V-EQN2V-EQ+N2V [21]N2V-EQN2V-EQ+
BSD500 [31]23.09/0.75828.19/0.81129.05/0.84722.20/0.66326.06/0.71826.47/0.745
Kodak24 [10]22.77/0.75828.98/0.80929.82/0.84022.72/0.65726.96/0.72127.32/0.743
Set14 [54]22.03/0.73627.07/0.78728.04/0.82019.65/0.60925.27/0.70725.67/0.731
Average22.63/0.75128.08/0.80228.97/0.83621.52/0.64326.10/0.71526.49/0.740
", + "image_path": "1069335b18d3957c9f9f16ce4540eadfc76bc4cf8ea76efc57006255c063c445.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 595, + 295, + 617 + ], + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 617 + ], + "type": "text", + "content": "Table 6. N2V w/o rotation augmentation: three networks were tested under conditions of Gaussian noise at levels 25 and 50." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 625, + 295, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 295, + 710 + ], + "type": "text", + "content": "Data rotation augmentation: Rotational data augmentation on training data is a commonly used method to enhance model performance and robustness. We conducted ablation experiments with rotation augmentation in the N2V method. Tab. 2 shows the experiments on the U-Net network with rotation augmentation, while Tab. 6 presents our ablation experiments without augmentation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 455, + 388, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 388, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 388, + 467 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 472, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 638 + ], + "type": "text", + "content": "In this work, we first explore and introduce rotation equivariant image prior into the self-supervised image denoising task at the network architecture level. Through rigorous theoretical analysis, we prove that simply replacing the convolution layers with ECNNs leads to a rotational equivariant version of the network. Building on this, we propose AdaReNet, an adaptive rotation equivariant network, which further enhances performance. Extensive experiments demonstrate the effectiveness of our approach, confirming that incorporating rotation equivariant prior significantly improves denoising results. Overall, our work underscores the importance of leveraging rotation invariance for self-supervised learning and sets a foundation for future research in this field." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Acknowledgement. This research was supported by NSFC project under contract U21A6005; the Major Key Project of PCL under Grant PCL2024A06; Tianyuan Fund for Mathematics of the National Natural Science Foundation of China (Grant No.12426105) and Key Research and Development Program (Grant No. 2024YFA1012000)." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12727" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Josue Anaya and Adrian Barbu. Renoir-a dataset for real low-light image noise reduction. Journal of Visual Communication and Image Representation, 51:144-154, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "text", + "content": "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3155-3164, 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 203 + ], + "type": "text", + "content": "[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533. PMLR, 2019. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "type": "text", + "content": "[4] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 60-65. IEEE, 2005. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 282 + ], + "type": "text", + "content": "[5] Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. Non-local means denoising. Image Processing On Line, 1: 208-212, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 327 + ], + "type": "text", + "content": "[6] Sungmin Cha, Taeon Park, and Taesup Moon. Gan2gan: Generative noise learning for blind image denoising with single noisy images. arXiv preprint arXiv:1905.10488, 3, 2019. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 384 + ], + "type": "text", + "content": "[7] Meng Chang, Qi Li, Huajun Feng, and Zhihai Xu. Spatial-adaptive network for single image denoising. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 171-187. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 418 + ], + "type": "text", + "content": "[8] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990-2999. PMLR, 2016. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "text", + "content": "[9] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on image processing, 16(8):2080-2095, 2007. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 294, + 487 + ], + "type": "text", + "content": "[10] Rich Franzen. Kodak lossless true color image suite. source: http://rOk.us/graphics/kodak, 4(2):9, 1999. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "type": "text", + "content": "[11] Jiahong Fu, Qi Xie, Deyu Meng, and Zongben Xu. Rotation equivariant proximal operator for deep unfolding methods in image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 534, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 295, + 588 + ], + "type": "text", + "content": "[12] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862-2869, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "type": "text", + "content": "[13] Shuhang Gu, Qi Xie, Deyu Meng, Wangmeng Zuo, Xiangchu Feng, and Lei Zhang. Weighted nuclear norm minimization and its applications to low level vision. International journal of computer vision, 121:183-208, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[14] Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1712-1722, 2019. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 98, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 129 + ], + "type": "text", + "content": "[16] R Mark Henkelman. Measurement of signal intensities in the presence of noise in mr images. Medical physics, 12(2): 232-233, 1985. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 132, + 553, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 553, + 165 + ], + "type": "text", + "content": "[17] Emiel Hoogeboom, Jorn WT Peters, Taco S Cohen, and Max Welling. Hexaconv. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 167, + 555, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 167, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 555, + 221 + ], + "type": "text", + "content": "[18] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 224, + 554, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 224, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 224, + 554, + 277 + ], + "type": "text", + "content": "[19] Xixi Jia, Sanyang Liu, Xiangchu Feng, and Lei Zhang. Focnet: A fractional optimal control network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6054-6063, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 281, + 555, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 555, + 335 + ], + "type": "text", + "content": "[20] Yoonsik Kim, Jae Woong Soh, Gu Yong Park, and Nam Ik Cho. Transfer learning from synthetic to real-noise denoising with adaptive instance normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3482-3492, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "type": "text", + "content": "[21] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 1, 3, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[22] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. Frontiers in Computer Science, 2:5, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 440, + 553, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 440, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 317, + 440, + 553, + 494 + ], + "type": "text", + "content": "[23] Yann LeCun, Bernhard Boser, John S Denker, Donnie Henderson, Richard E Howard, Wayne Hubbard, and Lawrence D Jackel. Backpropagation applied to handwritten zip code recognition. Neural computation, 1(4):541-551, 1989. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "text", + "content": "[24] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "text", + "content": "[25] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3204-3213, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 589, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 553, + 632 + ], + "type": "text", + "content": "[26] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 3, 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[27] Jizhihui Liu, Qixun Teng, and Junjun Jiang. Fm2s: Self-supervised fluorescence microscopy denoising with single noisy image. arXiv preprint arXiv:2412.10031, 2024. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Zhiliang Liu and Liyuan Ren. Shaking noise exploration and elimination for detecting local flaws of steel wire ropes based on magnetic flux leakages. IEEE Transactions on Industrial Electronics, 70(4):4206-4216, 2022. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12728" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[29] Ymir Mäkinen, Lucio Azzari, and Alessandro Foi. Exact transform-domain noise variance for collaborative filtering of stationary correlated noise. In 2019 IEEE international conference on image processing (ICIP), pages 185-189. IEEE, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[30] Diego Marcos, Michele Volpi, Nikos Komodakis, and Devis Tuia. Rotation equivariant vector field networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 5048-5057, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "text", + "content": "[31] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 296 + ], + "type": "text", + "content": "[32] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12064-12072, 2020. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-recorrupted: unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2043-2052, 2021. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "text", + "content": "[34] Bumjun Park, Songhyun Yu, and Jechang Jeong. Densely connected hierarchical network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1586-1595, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "type": "text", + "content": "[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2self with dropout: Learning self-supervised denoising from single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1890-1898, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "text", + "content": "[37] Yuhui Quan, Yixin Chen, Yizhen Shao, Huan Teng, Yong Xu, and Hui Ji. Image denoising using complex-valued deep cnn. Pattern Recognition, 111:107639, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 600 + ], + "type": "text", + "content": "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[39] Zhengyang Shen, Lingshen He, Zhouchen Lin, and Jinwen Ma. Pdo-econvs: Partial differential operator based equivariant convolutions. In International Conference on Machine Learning, pages 8697-8706. PMLR, 2020. 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[40] Zhengyang Shen, Tiancheng Shen, Zhouchen Lin, and Jinwen Ma. Pdo-es2cnns: Partial differential operator based equivariant spherical cnns. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9585–9593, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 9446-9454, 2018.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 149 + ], + "type": "text", + "content": "[42] Raviteja Vemulapalli, Oncel Tuzel, and Ming-Yu Liu. Deep gaussian conditional random field network: A model-based deep network for discriminative denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4801-4809, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 555, + 205 + ], + "type": "text", + "content": "[43] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 554, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 258 + ], + "type": "text", + "content": "[44] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2027-2036, 2022, 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "type": "text", + "content": "[45] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. Advances in neural information processing systems, 32, 2019. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "type": "text", + "content": "[46] Maurice Weiler, Fred A Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 849-858, 2018. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "type": "text", + "content": "[47] Daniel E Worrall, Stephan J Garbin, Daniyar Turmukhambetov, and Gabriel J Brostow. Harmonic networks: Deep translation and rotation equivariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5028-5037, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 437 + ], + "type": "text", + "content": "[48] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European conference on computer vision, pages 352-368. Springer, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 480 + ], + "type": "text", + "content": "[49] Qi Xie, Qian Zhao, Zongben Xu, and Deyu Meng. Fourier series expansion based filter parametrization for equivariant convolutions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4537-4551, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 514 + ], + "type": "text", + "content": "[50] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "type": "text", + "content": "[51] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-clean: Learning self-supervised denoising from corrupted image. IEEE Transactions on Image Processing, 29:9316-9329, 2020. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "text", + "content": "[52] Zongsheng Yue, Qian Zhao, Lei Zhang, and Deyu Meng. Dual adversarial network: Toward real-world noise removal and noise generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 41-58. Springer, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 615, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 554, + 689 + ], + "type": "text", + "content": "[53] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 492-511. Springer, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[54] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12729" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 285 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[55] Dan Zhang, Fangfang Zhou, Felix Albu, Yuanzhou Wei, Xiao Yang, Yuan Gu, and Qiang Li. Unleashing the power of self-supervised image denoising: A comprehensive review. arXiv preprint arXiv:2308.00247, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[56] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 239 + ], + "type": "text", + "content": "[57] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 285 + ], + "type": "text", + "content": "[58] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 519-528, 2017. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12730" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_content_list.json b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3eb031f61edbde9b48c7d5a8d29e6372bf2bf7cc --- /dev/null +++ b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_content_list.json @@ -0,0 +1,1258 @@ +[ + { + "type": "text", + "text": "S2D-LFE: Sparse-to-Dense Light Field Event Generation", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yutong Liu Wenming Weng Yueyi Zhang Zhiwei Xiong* University of Science and Technology of China {ustclyt, wmweng}@mail.ustc.edu.cn {zhyuey, zwxiong}@ustc.edu.cn", + "bbox": [ + 207, + 180, + 782, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present S2D-LFE, an innovative approach for sparse-to-dense light field event generation. For the first time to our knowledge, S2D-LFE enables controllable novel view synthesis only from sparse-view light field event (LFE) data, and addresses three critical challenges for the LFE generation task: simplicity, controllability, and consistency. The simplicity aspect eliminates the dependency on frame-based modality, which often suffers from motion blur and low frame-rate limitations. The controllability aspect enables precise view synthesis under sparse LFE conditions with view-related constraints. The consistency aspect ensures both cross-view and temporal coherence in the generated results. To realize S2D-LFE, we develop a novel diffusion-based generation network with two key components. First, we design an LFE-customized variational auto-encoder that effectively compresses and reconstructs LFE by integrating cross-view information. Second, we design an LFE-aware injection adaptor to extract comprehensive geometric and texture priors. Furthermore, we construct a large-scale synthetic LFE dataset containing 162 one-minute sequences using simulator, and capture a real-world testset using our custom-built sparse LFE acquisition system, covering diverse indoor and outdoor scenes. Extensive experiments demonstrate that S2D-LFE successfully generates up to $9 \\times 9$ dense LFE from $2 \\times 2$ sparse inputs and outperforms existing methods on both synthetic and real-world data. The datasets and code are available at https://github.com/Yutong2022/S2D-LFE.", + "bbox": [ + 91, + 301, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 729, + 223, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras capture per-pixel brightness changes asynchronously with microsecond temporal resolution [1], while Light Field (LF) cameras record both the accumulated intensity and direction of light rays, thereby revealing scene geometric structure [48]. The integration of these technologies leads to Light Field Event (LFE) cameras, which capture multi-view event streams arranged in a fixed baseline array. This combination enables simultaneous recording", + "bbox": [ + 89, + 753, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of high-quality spatial depth information and dynamic motion data with exceptional temporal resolution. Such capabilities demonstrate significant potential for various applications, including 6-degree-of-freedom virtual reality [23] and real-time dynamic depth estimation [40]. The rich spatio-temporal information provided by LFE data enables effective capture of structural information in high-speed dynamic scenes.", + "bbox": [ + 511, + 268, + 906, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to address a largely unexplored task of LFE generation. To date, the only relevant work is Ev-LFV [29], which combines an RGB LF camera with an event camera aligned to the LF center-view for LFE synthesis. However, this approach faces several critical challenges: (1) Simplicity. Ev-LFV's reliance on multi-view LF frame data as auxiliary information introduces complications. The RGB data is vulnerable to motion blur and exposure issues, potentially compromising LFE generation quality. Moreover, achieving precise spatial-temporal alignment between the LF and event cameras presents substantial technical challenges. (2) Controllability. The fixed view synthesis constraint of Ev-LFV prevents the generation of free viewpoints under sparse LFE conditions with view-related constraints. (3) Consistency. Due to the limitations of low frame-rate RGB data and insufficient structural modeling, Ev-LFV struggles to maintain both angular and temporal continuity in the generated LFE sequences.", + "bbox": [ + 511, + 393, + 908, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose S2D-LFE for Sparse-to-Dense Light Field Event generation, which is a novel approach for LFE generation. S2D-LFE demonstrates the capability to generate temporally consistent LFE from free viewpoints using only sparse LFE input, making it particularly suitable for real-world applications without the need for frame cameras. Our approach comprises two key components: an LFE-customized variational auto-encoder for efficient LFE compression and reconstruction, and a LFE-aware injection adapter integrated into a diffusion-based generation network that learns to generate the latent representation of LFE.", + "bbox": [ + 511, + 670, + 910, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first key component, LFE-VAE, is designed to effectively compress high-resolution LFE while preserving spatial and structural information. Specifically, we introduce", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11207", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two significant improvements over conventional pretrained VAE architectures used in diffusion models [25, 26, 35, 36, 53]. First, we incorporate angular interaction blocks into the VAE decoder to fully exploit cross-view information within LFE. Second, we implement a view-enhancement mechanism that utilizes input sparse LFE views to improve the quality of reconstructed novel views. These enhancements effectively leverage the inherent cross-view information in LFE and mitigate compression artifacts, resulting in high-fidelity reconstruction. The second key component, LFE-adapter, is designed to utilize geometric and texture priors from the input sparse LFE to guide the diffusion process. Specifically, LFE-adapter generates pyramid condition signals that provide targeted conditioning at multiple resolution scales for the U-Net during diffusion. This multi-scale conditioning enables more precise injection of geometric and texture priors, thereby achieving superior cross-view consistency.", + "bbox": [ + 89, + 90, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we develop the first sparse LFE capture system to facilitate the acquisition of continuous light field event data. The system consists of four Davis346 cameras arranged in a $2 \\times 2$ array configuration. Using this setup, we collect a diverse dataset comprising 25 sequences, including 15 indoor scenes and 10 autonomous driving scenarios. To support comprehensive model training, we additionally construct a large-scale synthetic dataset using Carla simulator [14], generating 162 one-minute dense LFE sequences.", + "bbox": [ + 89, + 363, + 480, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper can be summarized as follows:", + "bbox": [ + 89, + 500, + 480, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose S2D-LFE, a novel approach for generating free-viewpoint and temporally consistent light field events from sparse-view LFE input.", + "- Our S2D-LFE incorporates two key components: an LFE-VAE for efficient LFE compression, and an LFE-adapter integrated into a diffusion-based generation network that reconstructs the latent representation of LFE.", + "- We develop the first sparse LFE capture system and collect a dataset of 25 real-world sparse LFE sequences. Additionally, we synthesize 162 dense LFE sequences using the Carla simulator to facilitate comprehensive training.", + "- Our approach outperforms existing methods on both synthetic and real-world datasets, while presenting enhanced robustness in generalizing to free-viewpoint synthesis." + ], + "bbox": [ + 89, + 529, + 482, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 753, + 232, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view Event Processing. While event cameras offer exceptional temporal resolution and low latency [1], monocular event data alone is insufficient for capturing comprehensive structural information of dynamic scenes. Various approaches have been proposed to address this limitation: 1) NeRF-based scene representation learning [4], 2) hybrid systems combining events with conventional cameras [9, 29], and 3) stereo event systems [10, 19, 40, 55].", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF-based methods, while promising, require per-scene training and multiple viewpoints for effective scene representation. For example, EvDNeRF [4] introduces an event-based dynamic NeRF pipeline for reconstructing event streams in rapidly deforming scenes, but its performance degrades significantly with limited viewpoints $(< 8)$ . Hybrid systems face challenges in reconciling the temporal resolution disparity between different modalities. For instance, Lu et al. [29] encounter practical limitations in real-world applications due to the substantial difference in temporal resolution between LF and event cameras. While stereo event systems can capture geometric information, the resulting geometry tends to be sparse. In contrast, the Light Field event modality offers distinct advantages by capturing dense structural information of dynamic scenes without requiring per-scene training, while enabling multiple angular resolution generation.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light Field View Synthesis. The reconstruction of dense Light Fields from sparse views, commonly known as LF angular super-resolution, remains an active research area. Existing approaches can be broadly categorized into two groups: depth-dependent [5, 24, 43, 52] and depth-independent methods [12, 21, 23, 44]. Depth-dependent methods rely on precise disparity estimation but are vulnerable to occlusions, while depth-independent methods often struggle with generalization across varying disparity ranges. Recent advances in Neural Light Field representations [2, 3, 15, 28, 38, 42] have brought significant improvements to LF view synthesis by addressing limitations of conventional angular super-resolution methods. Compared with the widely adopted learnable 3D representation methods [7, 11, 34, 41], these approaches can more effectively exploit the spatial priors inherent in LF. However, these NeRF-based approaches are constrained by their requirement for numerous input views and scene-specific training. In contrast, our proposed S2D-LFE conquers these limitations by enabling free-viewpoint synthesis from sparse inputs without the need for per-scene training, while maintaining high recovery quality and temporal consistency.", + "bbox": [ + 511, + 359, + 903, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based Image Generation. Diffusion models have shown remarkable progress in novel view synthesis [6, 8, 20, 39, 46, 50, 51]. For example, NeRDi [20] leverages language priors for multi-view synthesis, effectively connecting image semantics to appearance reconstruction. These advances demonstrate the robust capability of diffusion models in modeling multi-view distributions. Furthermore, diffusion models have achieved impressive results across diverse data modalities, including depth [26], point clouds [30], optical flow [26, 37], and human pose [25]. Motivated by this versatility and strong fitting capability across different modalities, we adopt diffusion models for dense LFE generation.", + "bbox": [ + 511, + 703, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11208", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Real-world and Synthetic LFE Datasets", + "text_level": 1, + "bbox": [ + 89, + 89, + 446, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. LFE Capture System and Real-world Dataset", + "text_level": 1, + "bbox": [ + 89, + 116, + 475, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "System Setup. To facilitate the collection of real-world LFE data and validate our approach, we develop a custom LFE capture system. Given the lack of commercially available LFE cameras, our system design prioritizes precise spatial alignment and temporal synchronization while maintaining cost-effectiveness. The system, shown in Fig. 1, consists of a $2 \\times 2$ array of Davis346 cameras mounted on a square aluminum alloy frame, with uniform horizontal and vertical baseline lengths of $6\\mathrm{cm}$ . To ensure consistent imaging characteristics across views, all four cameras are equipped with identical lenses and configured with matching parameters, including focal length and exposure time. For precise temporal synchronization, we implement a master-slave configuration using external cables (depicted as black lines in Fig.1). The upper-left DAVIS346 serves as the master camera, transmitting trigger signal pulses to the three slave cameras through these cables. This system captures event data at a spatial resolution of $346 \\times 260$ pixels, which can be readily upgraded to a higher resolution by updating event cameras. Detailed parameter configurations are provided in the supplementary materials.", + "bbox": [ + 89, + 140, + 482, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera Calibration. Precise camera calibration is crucial for accurate Light Field Event capture, as it ensures geometric consistency across multiple views and enables reliable reconstruction of scene structure. This is particularly important for our system where accurate spatial relationships between cameras directly impact the quality of captured LFE data. We utilize the Kalibr toolbox [16] to estimate camera intrinsics and extrinsics by capturing fixed-rate frames of a checkerboard pattern. Since the DAVIS346 cameras output both frames and event streams through the same optical system, the calibration parameters obtained from frames are directly applicable to event streams. Following calibration, we perform systematic rectification of the captured sparse LFE. Using the upper left camera as reference, we first apply perspective transformation to project the upper right camera onto a virtual plane parallel to the reference camera's horizontal plane, establishing the baseline length between them. We then align the two lower cameras through perspective transformations to their corresponding positions in the sparse LF, ensuring uniform baseline lengths across the array. This comprehensive calibration and rectification process ensures accurate structural information capture of dynamic scenes.", + "bbox": [ + 89, + 460, + 482, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-world LFE Dataset. Following system calibration and rectification, we collect a real LFE dataset comprising 25 dynamic scenes using our LFE capture system, which is termed \"LFE-real\". The dataset consists of two main categories: 15 indoor sequences captured by moving our camera array through various indoor environments (featuring sub", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9ab701736589bc8e08a689fcfe9bb5fe4453fbf480f2cb1727bc7001c85226cf.jpg", + "image_caption": [ + "Figure 1. The proposed LFE capturing system." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 890, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "jects such as people, plants, and interior layouts), and 10 autonomous driving sequences recorded from displayed autonomous driving footage. Each sequence spans between 3 to 5 minutes, ensuring sufficient temporal coverage of the dynamic events. The details about scenes are in the supplementary material.", + "bbox": [ + 511, + 296, + 906, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Synthetic Dataset", + "text_level": 1, + "bbox": [ + 511, + 397, + 684, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that physical capture devices can only provide sparse view inputs without dense multi-view supervision, we leverage the Carla simulator [14] to generate a large synthetic LFE dataset for training and evaluation, which is termed \"LFE-syn\". We utilize 18 different maps provided by the Carla environment for data simulation. Each map is divided into 9 regions, with a designated starting point in each region. At these starting points, we deploy virtual vehicles equipped with $5 \\times 5$ event camera arrays. The baseline lengths between adjacent cameras are randomly set between $3\\mathrm{cm}$ and $9\\mathrm{cm}$ . The event trigger thresholds are configured asymmetrically: 0.1 to 0.15 for positive events and -0.15 to -0.1 for negative events. Using autonomous navigation, each vehicle generates one-minute LFE sequences, resulting in a total of 162 synthetic sequences. Additionally, to facilitate the evaluation, our synthetic dataset share the same resolution with the real-world dataset. Detailed parameter configurations are provided in the supplementary materials.", + "bbox": [ + 511, + 419, + 906, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. S2D-LFE Generation Network", + "text_level": 1, + "bbox": [ + 511, + 705, + 795, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our custom-built system captures sparse LFE with coarse geometric information of dynamic scenes. To enhance the structural representation capabilities, we propose S2D-LFE, a novel approach for synthesizing accurate dense LFEs from sparse inputs. We detail our approach in the following sections.", + "bbox": [ + 511, + 731, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 832, + 624, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our setting, each event stream view in the LFE is accumulated as an Event Count Map (ECM) [17] with shapes of $\\mathbb{R}^{B\\times TB\\times 2\\times H\\times W}$ , where $B$ represents the batch size,", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11209", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/768a75bddd89b0bb7371aee53791228bcb5956d7aa419df5233b3818c1697558.jpg", + "image_caption": [ + "Figure 2. The second training stage of the LFE-VAE architecture. The input to the LFE-VAE supports multiple angular resolutions of LFE. Here, we take the case of an input with 25 views as an example. The LFE-VAE is composed of the encoder $V_{enc}$ , the decoder $\\hat{V}_{dec}$ inserted by angular interaction blocks, and the following angular resblocks. The training at this stage is aimed at updating the decoder $\\hat{V}_{dec}$ and the angular resblocks." + ], + "image_footnote": [], + "bbox": [ + 150, + 85, + 841, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$TB$ represents the temporal bins, 2 denotes the polarities, and $H\\times W$ denotes the spatial resolution. To facilitate the subsequent processing of the VAE, we transform the two-polarity channels of the event stream into three channels based on [49]. A LFE can be represented as $L\\in \\mathbb{R}^{B\\times 3\\times M\\times N\\times H\\times W}$ , where $M\\times N$ denotes the angular resolution. The LFE synthesis task aims to generate a dense LFE $L_{D}\\in \\mathbb{R}^{B\\times 3\\times M\\times N\\times H\\times W}$ from a sparse LFE $L_{S}\\in \\mathbb{R}^{B\\times 3\\times m\\times n\\times H\\times W}$ , where $m\\times n$ represents a lower angular resolution. To effectively model the complex distribution of LFE data, we adopt a diffusion-based pipeline. Given the high-dimensional nature of LFE data, we first introduce LFE-VAE to encode LFE into a compact latent representation. The encoded sparse LFE, together with the encoded angular coordinates, serve as conditioning signals for the diffusion model to synthesize views at specified angular positions. To effectively incorporate these conditioning signals into the diffusion backbone, we design an LFE-adapter for optimal signal processing. Through iterative generation of all unknown views, we achieve the transformation from sparse to dense LFE. Notably, our diffusion network demonstrates the capability to synthesize novel views at arbitrary angles beyond the training set by accepting different angular coordinates as input.", + "bbox": [ + 91, + 325, + 483, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. LFE-VAE", + "text_level": 1, + "bbox": [ + 89, + 696, + 205, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle the high spatial resolution of LFE data efficiently, we employ Variational Autoencoder (VAE) architectures commonly used in diffusion models [25, 26, 35, 36, 53] for data compression. However, directly applying conventional VAEs to LFE data presents two significant challenges: 1) Traditional VAEs used in stable diffusion are designed for single-view processing, failing to leverage the inherent cross-view structural information in LFE data, and 2) The lossy nature of VAE compression inevitably results in information loss in the event streams.", + "bbox": [ + 89, + 719, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these limitations, we propose LFE-VAE, building upon the pretrained VAE model from SDXL [35].", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our architecture introduces two key innovations: (1) angular interaction blocks in the decoder to incorporate cross-view complementary information, and (2) a view-enhancement mechanism utilizing four known views to minimize compression artifacts.", + "bbox": [ + 511, + 325, + 906, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In specific, for each dense LFE input $L_{in}$ , the encoder first generates a latent code $Z_{s}$ . This latent representation is processed by our enhanced decoder, which embeds cross-view geometric information through angular interaction blocks. Each block consists of two cascaded layers of spatial and angular convolutions. The corresponding positions of the decoded target views are replaced with four known sparse LFE views, and then the LFE is refined through several cascaded resblocks defined on the LFE angular slices.", + "bbox": [ + 511, + 402, + 908, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our training process follows a two-stage approach. First, we fine-tune the stable diffusion VAE using single-view events to obtain updated weights for the encoder $V_{enc}$ and decoder $V_{dec}$ . Subsequently, while keeping $V_{enc}$ frozen, we fine-tune the proposed decoder $\\hat{V}_{dec}$ and angular resblocks. The second stage of the LFE-VAE architecture is illustrated in detail in Fig. 2. Specifically, we first select the sparse LFE views $L_{S}$ , which represent the condition signal for the subsequent diffusion pipeline. The input LFE $L_{in}$ is then encoded by $V_{enc}$ and decoded by the updated $\\hat{V}_{dec}$ , resulting in an intermediate reconstructed LFE. Next, we replace the four corner views of the recovered LFE, with the four corresponding views in $L_{S}$ selected in $L_{in}$ . Finally, the compression degradation is restored through the utilization of angular resblocks. After training, the encoder $V_{enc}$ is integrated into the diffusion training pipeline to compress input sparse LFE, enabling the subsequent diffusion network to handle higher-resolution LFE generation.", + "bbox": [ + 511, + 553, + 908, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Diffusion Pipeline", + "text_level": 1, + "bbox": [ + 511, + 833, + 689, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following mainstream diffusion architectures [13, 22, 25, 26, 35, 36, 53], our diffusion backbone adopts a UNet structure. While existing diffusion models primarily han-", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11210", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42608d48f053ec2b126a00dfc3f2f39a848f22e45896ac8fa657580dbd757254.jpg", + "image_caption": [ + "Figure 3. The diffusion pipeline of S2D-LFE at the t-th time step. The pipeline primarily consists of two parts: (1) The left part, highlighted with a blue background, represents the main pipeline diagram. The diffusion pipeline of S2D-LFE adopts a UNet backbone, where the latent features $Z_{s}$ from four viewpoints are injected into the network as conditional signals via the LFE-adapter. In addition to the time embedding, our prompt incorporates positional encodings of both horizontal and vertical coordinates. (2) The right section, highlighted with an orange background, depicts the LFE-adapter block within the LFE-adapter. After cascading the condition information through the ContBlock and GeoBlock, the output features are embedded into the corresponding layers of the UNet. The LFE-adapter consists of four LFE-adapter blocks, each producing a different conditional signal $\\hat{Z}_{\\mathrm{s1}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s2}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s3}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s4}}^{\\downarrow \\downarrow}$ corresponding to the respective layers of the UNet." + ], + "image_footnote": [], + "bbox": [ + 130, + 87, + 553, + 344 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5deff887cfac4342786502fd0dc7242c93cf4c8a19fcad00ba2fcc41352313db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 88, + 864, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dle unstructured conditioning data such as images, videos, and text, the potential of structured conditioning data, which contains rich structural and texture priors, remains largely unexplored. To address this gap, we propose a specialized diffusion pipeline tailored for LFE data. At its core is the novel LFE-adapter, which serves two key functions: 1) extracting features that encode both structural and texture information from input sparse LFE and 2) injecting these extracted features into the diffusion backbone. With the structural and texture information effectively embedded, the diffusion pipeline generates specific LFE views based on input view prompts. Our approach employs continuous, linear angular coordinate encoding for view prompts, enabling the pipeline to generalize to sampling LFE views across varying densities and provide flexibility in LFE generation. The complete architecture of our diffusion pipeline at the t-th time step is illustrated in Fig. 3. Then, we will detail each of these innovations.", + "bbox": [ + 88, + 460, + 482, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LFE Angular Prompt. Accurate positional priors of target views are essential for novel LF view generation, which we achieve through angular coordinate prompting. As illustrated in Fig. 3, we process 2D coordinates through a two-stage encoding scheme. First, we independently encode row and column coordinates. These encoded coordinates are then transformed through cascaded fully connected layers to obtain the raw and column embeddings, which are then concatenated and fed into a linear layer to get the angular embedding $F_{ang}$ . It serves dual purposes in our pipeline. It is combined with time embedding as input to each UNet layer,", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "while simultaneously providing angular positional priors to the LFE-adapter for extracting view-specific conditioning information for the target view generation.", + "bbox": [ + 511, + 460, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LFE-adapter. Our pipeline leverages priors from a sparse LFE containing four known views. The design incorporates both texture and geometry priors from these input views to provide specific conditioning signals for novel view generation. Drawing inspiration from [33], we propose an LFE-adapter to effectively inject structural and textural information from the known views. The architecture of our LFE-adapter is depicted in the orange region of Fig. 3. The processing pipeline begins by using the $V_{\\mathrm{enc}}$ of LFE-VAE to obtain compressed latent codes $Z_{s}$ from the sparse LFE. We then integrate the angle embedding $F_{\\mathrm{ang}}$ with the latent features of the four input views to obtain the updated latent feature $Z_{s}^{*}$ . This feature is processed through a content extraction block (contblock) that extracts textures and details using spatial and angular convolutions. The spatial convolution embeds individual view information, while the angular convolution captures complementary cross-view information. A geometric extraction block (geoblock) further processes the features using cascaded Epipolar Plane Image (EPI) convolutions [5, 24, 43, 52] to embed geometric information, and thus we obtain the updated feature $\\hat{Z}_{s}$ . This feature undergoes spatial downsampling to produce $\\hat{Z}_{s}^{\\downarrow}$ for subsequent network updates, followed by angular down-sampling to generate $\\hat{Z}_{s}^{\\downarrow \\downarrow}$ as conditioning signals for corresponding UNet layers. The LFE-adapter ultimately generates a set of condition signals $\\hat{Z}_{s1}^{\\downarrow \\downarrow}, \\hat{Z}_{s2}^{\\downarrow \\downarrow}, \\hat{Z}_{s3}^{\\downarrow \\downarrow}, \\hat{Z}_{s4}^{\\downarrow \\downarrow}$ . Each sig", + "bbox": [ + 511, + 508, + 908, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11211", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5638f0904acb1d8906a8f1a6eae9b68345c14c597a9cf08868ff01f4eda1a1e0.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of event-based view synthesis on the in-training-scale setting. We tested various methods under synthetic and real settings and evaluated their performance across three metrics. The best results are highlighted in red, the second-best results are highlighted in blue, and the third-best results are highlighted in green. ‘↑’: the higher the better performance, ‘↓’: the opposite." + ], + "table_footnote": [], + "table_body": "
CategoryMethodPSNR ↑Synthetic SSIM ↑LPIPS ↓NIQE ↓Real-world BRISQUE ↓MUSIQ ↑
AngularSRJing et al. [23]21.120.5800.36231.5667.3228.14
DistgASR [44]23.320.6760.27828.8863.9529.90
SAV [12]22.840.6140.28229.1562.7729.25
Guo et al. [21]23.460.6870.29028.2359.1330.72
HybridET-Net [47]+SAV [12]+Vid2E [18]20.230.5740.37332.1167.1125.45
ET-Net [47]+Guo et al.+Vid2E [18]20.880.5890.37531.9566.5825.59
NeRF-basedR2L [42]19.880.5630.40128.4460.2530.45
GenerationOurs24.060.7010.23927.6553.4531.14
", + "bbox": [ + 91, + 132, + 903, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nal is added to the corresponding same-size input features of the UNet blocks.", + "bbox": [ + 89, + 299, + 482, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LFE View Sampling. Through our LFE-adapter, the diffusion pipeline is enriched with comprehensive structural and textural information from the input LFE. The angular prompts effectively encode target view positions, ensuring angular continuity in the generated LFE views. The pipeline samples individual views sequentially, reconstructing the complete continuous LFE by iteratively processing coordinates for all target views. The framework accommodates LFEs of varying densities through relative coordinate mapping, where $(0,0)$ represents the top-left view and $(1,1)$ represents the bottom-right view. For example, generating a view at position $(i,j)$ in an $N\\times N$ LFE would use normalized coordinates $(i / (N - 1),j / (N - 1))$ . This normalized coordinate system provides a unified approach for sampling views across different LFE densities while maintaining consistent spatial relationships.", + "bbox": [ + 88, + 330, + 482, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 590, + 215, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 618, + 284, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our networks are trained on a computing cluster equipped with eight NVIDIA A100 GPUs. Following [44], we employ data augmentation techniques including random horizontal flipping, vertical flipping, and $90^{\\circ}$ rotation to enhance model robustness. The optimization process utilizes the Adam optimizer with hyperparameters $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ , processing batches of size 32. We adopt a learning rate scheduling strategy where the initial rate of $2 \\times 10^{-4}$ is halved every 400 epochs. Training concludes after 1500 epochs when convergence is achieved. To ensure fairness in the experiments, all metrics are calculated excluding the input views.", + "bbox": [ + 89, + 642, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. For experimental validation, we utilize two distinct datasets: a synthetic dataset generated using Carla for training and testing, and a real-world dataset captured by our system for additional testing. We use two sets of complementary metrics. For data with ground-truth", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "references, we utilize PSNR measured in dB, SSIM [45], and LPIPS [54]. For data without ground-truth references, we employ NIQE [32], BRISQUE [31], and MUSIQ [27].", + "bbox": [ + 511, + 299, + 906, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods. We establish comparisons with three categories of related approaches1. The first category comprises LF angular super-resolution methods, including Jing et al.[23], DistgASR[44], SAV [12], and Guo et al.[21]. The second category involves a three-stage approach: we first convert sparse LFE to sparse RGB LF using event-based video reconstruction methods like ET-Net[47], then perform angular super-resolution to get dense RGB LF, and finally convert the dense RGB LF back to dense LFE using Vid2E [18]. The third category explores neural radiance field-based reconstruction using R2L [42]. All the aforementioned baseline methods are re-trained on our dataset loading their original pre-trained parameters.", + "bbox": [ + 511, + 345, + 908, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison Results", + "text_level": 1, + "bbox": [ + 511, + 553, + 705, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-training-scale Comparisons. We first conduct in-distribution evaluation on $5 \\times 5$ LFE generation, which matches our training configuration. As shown in Table 1, our S2D-LFE outperforms all baselines across all metrics on both synthetic and real-world testsets. On the synthetic testset, S2D-LFE achieves a notable PSNR improvement of 0.60 dB over the second-best method Guo et al.[21], demonstrating superior reconstruction accuracy. The strong LPIPS performance further indicates enhanced perceptual quality, leveraging the advantages of generative modeling. This is visually evident in Fig.4, where S2D-LFE reconstructs building details with minimal artifacts and better ground truth alignment. In addition, although R2L [42] achieves decent visual results in Fig.4, it fails to reconstruct in certain scenes, leading to lower metric performance on the synthetic dataset. Meanwhile, S2D-LFE maintains its superior performance across all metrics on the real-world testset, demonstrating effective generalization to real-world scenarios. This quantitative advantage is corroborated by", + "bbox": [ + 511, + 575, + 908, + 864 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1Since EV-LFV [29] and our S2D-LFE utilize different input modalities, a direct comparison between the two methods is not feasible.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11212", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5785cd29695ed92bf8fc992bc554d3e67472ab68881926467ca60cecfb7ceef7.jpg", + "image_caption": [ + "Figure 4. The figure presents a qualitative comparison on the central view of the generated LFE, using various methods for both synthetic (top row) and real-world (bottom row) testsets. The proposed S2D-LFE method (labeled as \"Ours\") is compared against other existing techniques, including SAV [12], Guo et al. [21], ET-Net [47] + Guo et al. + Vid2E [18], R2L [42], DistgASR [44], and the ground-truth. The highlighted regions (yellow and green boxes) magnify specific areas to emphasize the differences in event generation quality, with S2D-LFE exhibiting improved fidelity and preservation of fine details in both synthetic and real-world scenarios." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 893, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the qualitative results shown in Fig. 4, where S2D-LFE produces higher quality reconstructions with better preservation of details. Due to incorrect estimation of view position, SAV [12] and DistgASR [44] exhibit misaligned results compared to other methods. Furthermore, the ET-Net $[47] + \\mathrm{SAV} + \\mathrm{Vid2E}$ [18] approach exhibits severe view aliasing in this scene.", + "bbox": [ + 89, + 426, + 483, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Out-of-training-scale Comparisons. We extend our evaluation to test the model's generalization capability beyond its training distribution of $5 \\times 5$ LFE, examining its performance on denser LFEs. Among the seven baselines we selected, only R2L [42] and Guo et al. [21] are capable of handling out-of-training-scale LFE generation. Therefore, we only compare our method with these two approaches in this section. As shown in Table 2, our S2D-LFE consistently outperforms baseline methods across all tested LFE densities. While all methods show performance degradation as LFE density increases, S2D-LFE demonstrates greater resilience with a slower degradation rate. The qualitative results in Fig.5 further validate S2D-LFE's superior performance in maintaining perceptual quality and detail preservation for out-of-distribution scenarios. Additional visual comparisons are provided in the supplementary material.", + "bbox": [ + 89, + 535, + 483, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Angular Consistency. In this section, we analyze and compare the angular continuity of the LFE generated by our S2D-LFE. Given that the angular continuity of LFE data is strongly correlated with the gradient of the Epipolar Plane Image (EPI) [44, 48], we compute metrics (PSNR, SSIM) on the EPI gradient to evaluate the angular continuity of Guo et al. [21], R2L [42], and S2D-LFE when generating LFEs with different numbers of views (ranging from $5 \\times 5$", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac77e20a8a93f083f1158f6c4133177e879179a75866efbb6e47cc28cb31d062.jpg", + "image_caption": [ + "Figure 5. The figure presents the out-of-training-scale qualitative comparison using different methods including Guo et al. [21], R2L [42], and our proposed S2D-LFE (Ours). Besides, synthetic-scene (1) is the (1,1) view from the generated $8 \\times 8$ LFE, and realworld-scene (2) is the (7,7) view from the generated $8 \\times 8$ LFE. The yellow and green boxes highlight specific regions to better illustrate the differences in event generation quality." + ], + "image_footnote": [], + "bbox": [ + 517, + 422, + 906, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to $9 \\times 9$ ). It can be observed from Fig. 6 that as the number of views increases, all three methods exhibit a decreasing trend in both PSNR and SSIM metrics while the decline for our S2D-LFE is more gradual. It can be interpreted as S2D-LFE exhibits better angular consistency and robustness compared to the other methods. Moreover, we exhibit a qualitative comparison when generating $9 \\times 9$ LFE in Fig. 7. It can be observed that S2D-LFE preserves fidelity details while producing EPI slices that more closely align with the ground truth.", + "bbox": [ + 511, + 753, + 908, + 905 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11213", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5dc7c61df9a89b7591cfc11925c1ec68e0f1fb2f3519d0c3f1f05bf6c63bdb3e.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of event-based view synthesis on the out-of-training-scale setting. We tested various methods under three different settings (generating 36 views, 49 views, and 64 views) and evaluated their performance across three metrics. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite." + ], + "table_footnote": [], + "table_body": "
MethodNIQE↓BRISQUE↓MUSIQ↑
36-view49-views64-views36-views49-views64-views36-views49-views64-views
SynGuo et al. [21]16.8518.6220.1059.4564.1270.7831.2830.0228.31
R2L [42]16.9619.0521.5461.7265.9371.5630.2729.0827.29
Ours15.8417.6519.6754.1558.2163.4932.3031.8930.34
RealGuo et al. [21]28.8129.5830.2962.8366.1570.5830.2729.7829.30
R2L [42]28.9429.7230.6164.0068.1871.3729.9929.4829.02
Ours27.9928.4729.0156.1758.9161.2930.9330.6530.21
", + "bbox": [ + 94, + 132, + 906, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b5a34f7c996ad3d86ec5977f4acf01db5be8fb186671be013d2b6eae0e87ee2.jpg", + "image_caption": [ + "Figure 6. The figure presents a quantitative comparison of angular consistency across different methods, including Guo et al., R2L, and our proposed S2D-LFE. Subfigures (a) and (b) depict PSNR (dB) and SSIM metrics respectively, when generating different numbers of views ranging from $5 \\times 5$ to $9 \\times 9$ . All methods exhibit a decline in both metrics as the number of views increases, but S2D-LFE exhibits a relatively slower decline." + ], + "image_footnote": [], + "bbox": [ + 117, + 256, + 434, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd465763e4f846f4ddbce5eb4de5cb969309fa04060ec249da2121ad107df416.jpg", + "image_caption": [ + "Figure 7. The figure presents a qualitative comparison when generating $9 \\times 9$ LFE using Guo et al., R2L, and our proposed S2D-LFE. The green boxes highlight spatial slices of view (1,1), while the yellow boxes highlight EPI slices." + ], + "image_footnote": [], + "bbox": [ + 91, + 508, + 187, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4e4c381ae88f3504a8219a51a2dbaeb7b3b729709f900c8f571a74e35b41423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 508, + 261, + 601 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/776ed85c92cb1415e633d2b19013c5cb3e4ca180cace6bb3cf36cc21f50140e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 508, + 331, + 599 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d582cd8063f1da7edd2263868b583abc7d117eb57850fe23eb4ab8202f485688.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 508, + 403, + 598 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a2ae177fc9f089aa628ad1a696ffa3e9151915f30ad0fdd39a004e1557011a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 508, + 475, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 675, + 241, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LFE-VAE. We examine the effectiveness of our LFE-VAE by making comparisons with two variants: one using a standard VAE (S-VAE) and another using a fine-tuned VAE (FT-VAE) with single-view event data. As shown in Table 3, S-VAE variant shows performance degradation across all metrics, indicating reduced reconstruction quality and perceptual similarity. While FT-VAE variant shows modest improvements with PSNR of 37.12 dB and SSIM of 0.865, its LPIPS score of 0.151 still indicates suboptimal perceptual quality. In contrast, our LFE-VAE better captures the structural and perceptual characteristics of light field events.", + "bbox": [ + 89, + 702, + 482, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LFE-adapter. To validate the effectiveness of our LFE-adapter design, we make comparisons with a variant that", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/142ee98062b26eeccf434dfc16a0c1651a6993ee76530b1b9aadadc7367a30dc.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation of the ablation on the LFE-VAE. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. '↑': the higher the better performance, '↓': the opposite." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
S-VAE36.890.8590.158
FT-VAE37.120.8650.151
LFE-VAE (Ours)38.540.8810.112
", + "bbox": [ + 517, + 311, + 905, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5cac460d3c2d066d7b655a56976c1f32d381131a1697ef187838bddd7716936d.jpg", + "table_caption": [ + "Table 4. Quantitative evaluation of the ablation on the LFE-adapter. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
Concat22.540.6310.275
LFE-adapter24.060.7010.239
", + "bbox": [ + 517, + 454, + 905, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "replaces the LFE-adapter with simple feature concatenation (concat). As shown in Table 4, the concatenation variant shows substantial performance degradation across all metrics, indicating significant reductions in reconstruction accuracy, structural fidelity, and perceptual quality. The comparison clearly validates that our LFE-adapter provides more effective feature integration compared to naive concatenation.", + "bbox": [ + 511, + 515, + 905, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 646, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present S2D-LFE, a novel approach for generating consistent, full-view LFE. Our key contributions include developing the first real-world LFE capture system and proposing effective technical components: an LFE-VAE for LFE compression and an LFE-Adapter for prior injection. S2D-LFE achieves simplicity by eliminating multi-view RGB dependencies, controllability through precise view synthesis, and consistency via cross-view and temporal coherence. These advances establish S2D-LFE as an effective solution for dense LFE generation.", + "bbox": [ + 511, + 667, + 906, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 829, + 671, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001.", + "bbox": [ + 511, + 849, + 905, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11214", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A N Angelopoulos, J N Martel, A P Kohli, J Conradt, and G Wetzstein. Event based, near eye gaze tracking beyond 10,000 hz. IEEE Transactions on Visualization and Computer Graphics, page 1, 2020. 1, 2", + "[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19819-19829, 2022. 2", + "[3] Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel, and Tobias Ritschel. X-fields: Implicit neural view-, light- and time-image interpolation. ACM TOG, 39(6):1-15, 2020. 2", + "[4] A Bhattacharya, R Madaan, F Cladera, et al. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5846-5855, 2024. 2", + "[5] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. pages 497-504. 2001. 2, 5", + "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2", + "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2", + "[8] H Chen, J Gu, A Chen, et al. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2", + "[9] P Chen, W Guan, F Huang, et al. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 2023. 2", + "[10] P Chen, W Guan, and P Lu. Espvio: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8 (6):3661-3668, 2023. 2", + "[11] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2", + "[12] Zhen Cheng, Yutong Liu, and Zhiwei Xiong. Spatial-angular versatile convolution for light field reconstruction. IEEE Transactions on Computational Imaging, 8:1131-1144, 2022. 2, 6, 7", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4", + "[14] A Dosovitskiy, G Ros, F Codevilla, et al. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Brandon Yushan Feng and Amitabh Varshney. Signet: Efficient neural representation for light fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14224-14233, 2021. 2", + "[16] P Furgale, J Rehder, and R Siegwart. Unified temporal and spatial calibration for multi-sensor systems. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1280-1286. IEEE, 2013. 3", + "[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3", + "[18] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 6, 7", + "[19] M Gehrig, W Aarents, D Gehrig, and D Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2", + "[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023, 2", + "[21] Mantang Guo, Junhui Hou, Jing Jin, Hui Liu, Huanqiang Zeng, and Jiwen Lu. Content-aware warping for view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9486-9503, 2023. 2, 6, 7, 8", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4", + "[23] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 1, 2, 6", + "[24] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM TOG, 35(6):1-10, 2016. 2, 5", + "[25] J Karras, A Holynski, T C Wang, et al. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 2, 4", + "[26] B Ke, A Obukhov, S Huang, et al. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492–9502, 2024. 2, 4", + "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6", + "[28] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Zhicheng Lu, Xiaoming Chen, Vera Yuk Ying Chung, Weidong Cai, and Yiran Shen. Ev-lfv: Synthesizing light field event streams from an event camera and multiple rgb cameras. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2, 6", + "[30] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2", + "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6", + "[32] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6", + "[33] C Mou, X Wang, L Xie, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 5", + "[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2", + "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4", + "[36] R Rombach, A Blattmann, D Lorenz, et al. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4", + "[37] S Saxena, C Herrmann, J Hur, et al. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[38] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. NeurIPS, 34:19313-19325, 2021. 2", + "[39] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2", + "[40] S M N Uddin, S H Ahmed, and Y J Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 1, 2", + "[41] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In European Conference on Computer Vision, pages 612-629. Springer, 2022. 2, 6, 7, 8", + "[43] Yunlong Wang, Fei Liu, Zilei Wang, Guangqi Hou, Zhenan Sun, and Tieniu Tan. End-to-end view synthesis for light field imaging with pseudo 4dcnn. In ECCV, pages 333-348, 2018. 2, 5", + "[44] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2022. 2, 6, 7", + "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2", + "[47] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 6, 7", + "[48] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 1, 7", + "[49] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 4", + "[50] J Wynn and D Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 2", + "[51] J Xiang, J Yang, B Huang, et al. 3d-aware image generation using 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2383-2393, 2023. 2", + "[52] Henry Wing Fung Yeung, Junhui Hou, Jie Chen, Yuk Ying Chung, and Xiaoming Chen. Fast light field reconstruction with deep coarse-to-fine modeling of spatial-angular clues. In ECCV, pages 137-152, 2018. 2, 5", + "[53] L Zhang, A Rao, and M Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[55] Y Zhou, G Gallego, and S Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37(5):1433-1450, 2021. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11216", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_model.json b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ffd9d1de5ca8ba4bae3773d49e39fb0aaa8307bc --- /dev/null +++ b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.153 + ], + "angle": 0, + "content": "S2D-LFE: Sparse-to-Dense Light Field Event Generation" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.181, + 0.784, + 0.234 + ], + "angle": 0, + "content": "Yutong Liu Wenming Weng Yueyi Zhang Zhiwei Xiong* University of Science and Technology of China {ustclyt, wmweng}@mail.ustc.edu.cn {zhyuey, zwxiong}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.302, + 0.485, + 0.725 + ], + "angle": 0, + "content": "In this paper, we present S2D-LFE, an innovative approach for sparse-to-dense light field event generation. For the first time to our knowledge, S2D-LFE enables controllable novel view synthesis only from sparse-view light field event (LFE) data, and addresses three critical challenges for the LFE generation task: simplicity, controllability, and consistency. The simplicity aspect eliminates the dependency on frame-based modality, which often suffers from motion blur and low frame-rate limitations. The controllability aspect enables precise view synthesis under sparse LFE conditions with view-related constraints. The consistency aspect ensures both cross-view and temporal coherence in the generated results. To realize S2D-LFE, we develop a novel diffusion-based generation network with two key components. First, we design an LFE-customized variational auto-encoder that effectively compresses and reconstructs LFE by integrating cross-view information. Second, we design an LFE-aware injection adaptor to extract comprehensive geometric and texture priors. Furthermore, we construct a large-scale synthetic LFE dataset containing 162 one-minute sequences using simulator, and capture a real-world testset using our custom-built sparse LFE acquisition system, covering diverse indoor and outdoor scenes. Extensive experiments demonstrate that S2D-LFE successfully generates up to \\(9 \\times 9\\) dense LFE from \\(2 \\times 2\\) sparse inputs and outperforms existing methods on both synthetic and real-world data. The datasets and code are available at https://github.com/Yutong2022/S2D-LFE." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.73, + 0.224, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Event cameras capture per-pixel brightness changes asynchronously with microsecond temporal resolution [1], while Light Field (LF) cameras record both the accumulated intensity and direction of light rays, thereby revealing scene geometric structure [48]. The integration of these technologies leads to Light Field Event (LFE) cameras, which capture multi-view event streams arranged in a fixed baseline array. This combination enables simultaneous recording" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.391 + ], + "angle": 0, + "content": "of high-quality spatial depth information and dynamic motion data with exceptional temporal resolution. Such capabilities demonstrate significant potential for various applications, including 6-degree-of-freedom virtual reality [23] and real-time dynamic depth estimation [40]. The rich spatio-temporal information provided by LFE data enables effective capture of structural information in high-speed dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.909, + 0.668 + ], + "angle": 0, + "content": "In this paper, we aim to address a largely unexplored task of LFE generation. To date, the only relevant work is Ev-LFV [29], which combines an RGB LF camera with an event camera aligned to the LF center-view for LFE synthesis. However, this approach faces several critical challenges: (1) Simplicity. Ev-LFV's reliance on multi-view LF frame data as auxiliary information introduces complications. The RGB data is vulnerable to motion blur and exposure issues, potentially compromising LFE generation quality. Moreover, achieving precise spatial-temporal alignment between the LF and event cameras presents substantial technical challenges. (2) Controllability. The fixed view synthesis constraint of Ev-LFV prevents the generation of free viewpoints under sparse LFE conditions with view-related constraints. (3) Consistency. Due to the limitations of low frame-rate RGB data and insufficient structural modeling, Ev-LFV struggles to maintain both angular and temporal continuity in the generated LFE sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.911, + 0.852 + ], + "angle": 0, + "content": "To address these challenges, we propose S2D-LFE for Sparse-to-Dense Light Field Event generation, which is a novel approach for LFE generation. S2D-LFE demonstrates the capability to generate temporally consistent LFE from free viewpoints using only sparse LFE input, making it particularly suitable for real-world applications without the need for frame cameras. Our approach comprises two key components: an LFE-customized variational auto-encoder for efficient LFE compression and reconstruction, and a LFE-aware injection adapter integrated into a diffusion-based generation network that learns to generate the latent representation of LFE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The first key component, LFE-VAE, is designed to effectively compress high-resolution LFE while preserving spatial and structural information. Specifically, we introduce" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.363 + ], + "angle": 0, + "content": "two significant improvements over conventional pretrained VAE architectures used in diffusion models [25, 26, 35, 36, 53]. First, we incorporate angular interaction blocks into the VAE decoder to fully exploit cross-view information within LFE. Second, we implement a view-enhancement mechanism that utilizes input sparse LFE views to improve the quality of reconstructed novel views. These enhancements effectively leverage the inherent cross-view information in LFE and mitigate compression artifacts, resulting in high-fidelity reconstruction. The second key component, LFE-adapter, is designed to utilize geometric and texture priors from the input sparse LFE to guide the diffusion process. Specifically, LFE-adapter generates pyramid condition signals that provide targeted conditioning at multiple resolution scales for the U-Net during diffusion. This multi-scale conditioning enables more precise injection of geometric and texture priors, thereby achieving superior cross-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.482, + 0.499 + ], + "angle": 0, + "content": "Moreover, we develop the first sparse LFE capture system to facilitate the acquisition of continuous light field event data. The system consists of four Davis346 cameras arranged in a \\(2 \\times 2\\) array configuration. Using this setup, we collect a diverse dataset comprising 25 sequences, including 15 indoor scenes and 10 autonomous driving scenarios. To support comprehensive model training, we additionally construct a large-scale synthetic dataset using Carla simulator [14], generating 162 one-minute dense LFE sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.482, + 0.528 + ], + "angle": 0, + "content": "The main contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.483, + 0.575 + ], + "angle": 0, + "content": "- We propose S2D-LFE, a novel approach for generating free-viewpoint and temporally consistent light field events from sparse-view LFE input." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.576, + 0.483, + 0.635 + ], + "angle": 0, + "content": "- Our S2D-LFE incorporates two key components: an LFE-VAE for efficient LFE compression, and an LFE-adapter integrated into a diffusion-based generation network that reconstructs the latent representation of LFE." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.695 + ], + "angle": 0, + "content": "- We develop the first sparse LFE capture system and collect a dataset of 25 real-world sparse LFE sequences. Additionally, we synthesize 162 dense LFE sequences using the Carla simulator to facilitate comprehensive training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.697, + 0.483, + 0.741 + ], + "angle": 0, + "content": "- Our approach outperforms existing methods on both synthetic and real-world datasets, while presenting enhanced robustness in generalizing to free-viewpoint synthesis." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.53, + 0.483, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.233, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multi-view Event Processing. While event cameras offer exceptional temporal resolution and low latency [1], monocular event data alone is insufficient for capturing comprehensive structural information of dynamic scenes. Various approaches have been proposed to address this limitation: 1) NeRF-based scene representation learning [4], 2) hybrid systems combining events with conventional cameras [9, 29], and 3) stereo event systems [10, 19, 40, 55]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.349 + ], + "angle": 0, + "content": "NeRF-based methods, while promising, require per-scene training and multiple viewpoints for effective scene representation. For example, EvDNeRF [4] introduces an event-based dynamic NeRF pipeline for reconstructing event streams in rapidly deforming scenes, but its performance degrades significantly with limited viewpoints \\((< 8)\\). Hybrid systems face challenges in reconciling the temporal resolution disparity between different modalities. For instance, Lu et al. [29] encounter practical limitations in real-world applications due to the substantial difference in temporal resolution between LF and event cameras. While stereo event systems can capture geometric information, the resulting geometry tends to be sparse. In contrast, the Light Field event modality offers distinct advantages by capturing dense structural information of dynamic scenes without requiring per-scene training, while enabling multiple angular resolution generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.905, + 0.693 + ], + "angle": 0, + "content": "Light Field View Synthesis. The reconstruction of dense Light Fields from sparse views, commonly known as LF angular super-resolution, remains an active research area. Existing approaches can be broadly categorized into two groups: depth-dependent [5, 24, 43, 52] and depth-independent methods [12, 21, 23, 44]. Depth-dependent methods rely on precise disparity estimation but are vulnerable to occlusions, while depth-independent methods often struggle with generalization across varying disparity ranges. Recent advances in Neural Light Field representations [2, 3, 15, 28, 38, 42] have brought significant improvements to LF view synthesis by addressing limitations of conventional angular super-resolution methods. Compared with the widely adopted learnable 3D representation methods [7, 11, 34, 41], these approaches can more effectively exploit the spatial priors inherent in LF. However, these NeRF-based approaches are constrained by their requirement for numerous input views and scene-specific training. In contrast, our proposed S2D-LFE conquers these limitations by enabling free-viewpoint synthesis from sparse inputs without the need for per-scene training, while maintaining high recovery quality and temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion-based Image Generation. Diffusion models have shown remarkable progress in novel view synthesis [6, 8, 20, 39, 46, 50, 51]. For example, NeRDi [20] leverages language priors for multi-view synthesis, effectively connecting image semantics to appearance reconstruction. These advances demonstrate the robust capability of diffusion models in modeling multi-view distributions. Furthermore, diffusion models have achieved impressive results across diverse data modalities, including depth [26], point clouds [30], optical flow [26, 37], and human pose [25]. Motivated by this versatility and strong fitting capability across different modalities, we adopt diffusion models for dense LFE generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11208" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.447, + 0.108 + ], + "angle": 0, + "content": "3. Real-world and Synthetic LFE Datasets" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.117, + 0.476, + 0.133 + ], + "angle": 0, + "content": "3.1. LFE Capture System and Real-world Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.459 + ], + "angle": 0, + "content": "System Setup. To facilitate the collection of real-world LFE data and validate our approach, we develop a custom LFE capture system. Given the lack of commercially available LFE cameras, our system design prioritizes precise spatial alignment and temporal synchronization while maintaining cost-effectiveness. The system, shown in Fig. 1, consists of a \\(2 \\times 2\\) array of Davis346 cameras mounted on a square aluminum alloy frame, with uniform horizontal and vertical baseline lengths of \\(6\\mathrm{cm}\\). To ensure consistent imaging characteristics across views, all four cameras are equipped with identical lenses and configured with matching parameters, including focal length and exposure time. For precise temporal synchronization, we implement a master-slave configuration using external cables (depicted as black lines in Fig.1). The upper-left DAVIS346 serves as the master camera, transmitting trigger signal pulses to the three slave cameras through these cables. This system captures event data at a spatial resolution of \\(346 \\times 260\\) pixels, which can be readily upgraded to a higher resolution by updating event cameras. Detailed parameter configurations are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.808 + ], + "angle": 0, + "content": "Camera Calibration. Precise camera calibration is crucial for accurate Light Field Event capture, as it ensures geometric consistency across multiple views and enables reliable reconstruction of scene structure. This is particularly important for our system where accurate spatial relationships between cameras directly impact the quality of captured LFE data. We utilize the Kalibr toolbox [16] to estimate camera intrinsics and extrinsics by capturing fixed-rate frames of a checkerboard pattern. Since the DAVIS346 cameras output both frames and event streams through the same optical system, the calibration parameters obtained from frames are directly applicable to event streams. Following calibration, we perform systematic rectification of the captured sparse LFE. Using the upper left camera as reference, we first apply perspective transformation to project the upper right camera onto a virtual plane parallel to the reference camera's horizontal plane, establishing the baseline length between them. We then align the two lower cameras through perspective transformations to their corresponding positions in the sparse LF, ensuring uniform baseline lengths across the array. This comprehensive calibration and rectification process ensures accurate structural information capture of dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Real-world LFE Dataset. Following system calibration and rectification, we collect a real LFE dataset comprising 25 dynamic scenes using our LFE capture system, which is termed \"LFE-real\". The dataset consists of two main categories: 15 indoor sequences captured by moving our camera array through various indoor environments (featuring sub" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.892, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.263, + 0.848, + 0.277 + ], + "angle": 0, + "content": "Figure 1. The proposed LFE capturing system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.388 + ], + "angle": 0, + "content": "jects such as people, plants, and interior layouts), and 10 autonomous driving sequences recorded from displayed autonomous driving footage. Each sequence spans between 3 to 5 minutes, ensuring sufficient temporal coverage of the dynamic events. The details about scenes are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.686, + 0.414 + ], + "angle": 0, + "content": "3.2. Synthetic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.693 + ], + "angle": 0, + "content": "Given that physical capture devices can only provide sparse view inputs without dense multi-view supervision, we leverage the Carla simulator [14] to generate a large synthetic LFE dataset for training and evaluation, which is termed \"LFE-syn\". We utilize 18 different maps provided by the Carla environment for data simulation. Each map is divided into 9 regions, with a designated starting point in each region. At these starting points, we deploy virtual vehicles equipped with \\(5 \\times 5\\) event camera arrays. The baseline lengths between adjacent cameras are randomly set between \\(3\\mathrm{cm}\\) and \\(9\\mathrm{cm}\\). The event trigger thresholds are configured asymmetrically: 0.1 to 0.15 for positive events and -0.15 to -0.1 for negative events. Using autonomous navigation, each vehicle generates one-minute LFE sequences, resulting in a total of 162 synthetic sequences. Additionally, to facilitate the evaluation, our synthetic dataset share the same resolution with the real-world dataset. Detailed parameter configurations are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.796, + 0.723 + ], + "angle": 0, + "content": "4. S2D-LFE Generation Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Our custom-built system captures sparse LFE with coarse geometric information of dynamic scenes. To enhance the structural representation capabilities, we propose S2D-LFE, a novel approach for synthesizing accurate dense LFEs from sparse inputs. We detail our approach in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.625, + 0.849 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In our setting, each event stream view in the LFE is accumulated as an Event Count Map (ECM) [17] with shapes of \\(\\mathbb{R}^{B\\times TB\\times 2\\times H\\times W}\\), where \\(B\\) represents the batch size," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11209" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.087, + 0.842, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.262, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 2. The second training stage of the LFE-VAE architecture. The input to the LFE-VAE supports multiple angular resolutions of LFE. Here, we take the case of an input with 25 views as an example. The LFE-VAE is composed of the encoder \\( V_{enc} \\), the decoder \\( \\hat{V}_{dec} \\) inserted by angular interaction blocks, and the following angular resblocks. The training at this stage is aimed at updating the decoder \\( \\hat{V}_{dec} \\) and the angular resblocks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.327, + 0.485, + 0.691 + ], + "angle": 0, + "content": "\\(TB\\) represents the temporal bins, 2 denotes the polarities, and \\(H\\times W\\) denotes the spatial resolution. To facilitate the subsequent processing of the VAE, we transform the two-polarity channels of the event stream into three channels based on [49]. A LFE can be represented as \\(L\\in \\mathbb{R}^{B\\times 3\\times M\\times N\\times H\\times W}\\), where \\(M\\times N\\) denotes the angular resolution. The LFE synthesis task aims to generate a dense LFE \\(L_{D}\\in \\mathbb{R}^{B\\times 3\\times M\\times N\\times H\\times W}\\) from a sparse LFE \\(L_{S}\\in \\mathbb{R}^{B\\times 3\\times m\\times n\\times H\\times W}\\), where \\(m\\times n\\) represents a lower angular resolution. To effectively model the complex distribution of LFE data, we adopt a diffusion-based pipeline. Given the high-dimensional nature of LFE data, we first introduce LFE-VAE to encode LFE into a compact latent representation. The encoded sparse LFE, together with the encoded angular coordinates, serve as conditioning signals for the diffusion model to synthesize views at specified angular positions. To effectively incorporate these conditioning signals into the diffusion backbone, we design an LFE-adapter for optimal signal processing. Through iterative generation of all unknown views, we achieve the transformation from sparse to dense LFE. Notably, our diffusion network demonstrates the capability to synthesize novel views at arbitrary angles beyond the training set by accepting different angular coordinates as input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.206, + 0.713 + ], + "angle": 0, + "content": "4.2. LFE-VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To handle the high spatial resolution of LFE data efficiently, we employ Variational Autoencoder (VAE) architectures commonly used in diffusion models [25, 26, 35, 36, 53] for data compression. However, directly applying conventional VAEs to LFE data presents two significant challenges: 1) Traditional VAEs used in stable diffusion are designed for single-view processing, failing to leverage the inherent cross-view structural information in LFE data, and 2) The lossy nature of VAE compression inevitably results in information loss in the event streams." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we propose LFE-VAE, building upon the pretrained VAE model from SDXL [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.401 + ], + "angle": 0, + "content": "Our architecture introduces two key innovations: (1) angular interaction blocks in the decoder to incorporate cross-view complementary information, and (2) a view-enhancement mechanism utilizing four known views to minimize compression artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.909, + 0.552 + ], + "angle": 0, + "content": "In specific, for each dense LFE input \\(L_{in}\\), the encoder first generates a latent code \\(Z_{s}\\). This latent representation is processed by our enhanced decoder, which embeds cross-view geometric information through angular interaction blocks. Each block consists of two cascaded layers of spatial and angular convolutions. The corresponding positions of the decoded target views are replaced with four known sparse LFE views, and then the LFE is refined through several cascaded resblocks defined on the LFE angular slices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.826 + ], + "angle": 0, + "content": "Our training process follows a two-stage approach. First, we fine-tune the stable diffusion VAE using single-view events to obtain updated weights for the encoder \\(V_{enc}\\) and decoder \\(V_{dec}\\). Subsequently, while keeping \\(V_{enc}\\) frozen, we fine-tune the proposed decoder \\(\\hat{V}_{dec}\\) and angular resblocks. The second stage of the LFE-VAE architecture is illustrated in detail in Fig. 2. Specifically, we first select the sparse LFE views \\(L_{S}\\), which represent the condition signal for the subsequent diffusion pipeline. The input LFE \\(L_{in}\\) is then encoded by \\(V_{enc}\\) and decoded by the updated \\(\\hat{V}_{dec}\\), resulting in an intermediate reconstructed LFE. Next, we replace the four corner views of the recovered LFE, with the four corresponding views in \\(L_{S}\\) selected in \\(L_{in}\\). Finally, the compression degradation is restored through the utilization of angular resblocks. After training, the encoder \\(V_{enc}\\) is integrated into the diffusion training pipeline to compress input sparse LFE, enabling the subsequent diffusion network to handle higher-resolution LFE generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.691, + 0.85 + ], + "angle": 0, + "content": "4.3. Diffusion Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Following mainstream diffusion architectures [13, 22, 25, 26, 35, 36, 53], our diffusion backbone adopts a UNet structure. While existing diffusion models primarily han-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.088, + 0.554, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.089, + 0.865, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.351, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 3. The diffusion pipeline of S2D-LFE at the t-th time step. The pipeline primarily consists of two parts: (1) The left part, highlighted with a blue background, represents the main pipeline diagram. The diffusion pipeline of S2D-LFE adopts a UNet backbone, where the latent features \\( Z_{s} \\) from four viewpoints are injected into the network as conditional signals via the LFE-adapter. In addition to the time embedding, our prompt incorporates positional encodings of both horizontal and vertical coordinates. (2) The right section, highlighted with an orange background, depicts the LFE-adapter block within the LFE-adapter. After cascading the condition information through the ContBlock and GeoBlock, the output features are embedded into the corresponding layers of the UNet. The LFE-adapter consists of four LFE-adapter blocks, each producing a different conditional signal \\( \\hat{Z}_{\\mathrm{s1}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s2}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s3}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s4}}^{\\downarrow \\downarrow} \\) corresponding to the respective layers of the UNet." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.732 + ], + "angle": 0, + "content": "dle unstructured conditioning data such as images, videos, and text, the potential of structured conditioning data, which contains rich structural and texture priors, remains largely unexplored. To address this gap, we propose a specialized diffusion pipeline tailored for LFE data. At its core is the novel LFE-adapter, which serves two key functions: 1) extracting features that encode both structural and texture information from input sparse LFE and 2) injecting these extracted features into the diffusion backbone. With the structural and texture information effectively embedded, the diffusion pipeline generates specific LFE views based on input view prompts. Our approach employs continuous, linear angular coordinate encoding for view prompts, enabling the pipeline to generalize to sampling LFE views across varying densities and provide flexibility in LFE generation. The complete architecture of our diffusion pipeline at the t-th time step is illustrated in Fig. 3. Then, we will detail each of these innovations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "LFE Angular Prompt. Accurate positional priors of target views are essential for novel LF view generation, which we achieve through angular coordinate prompting. As illustrated in Fig. 3, we process 2D coordinates through a two-stage encoding scheme. First, we independently encode row and column coordinates. These encoded coordinates are then transformed through cascaded fully connected layers to obtain the raw and column embeddings, which are then concatenated and fed into a linear layer to get the angular embedding \\( F_{ang} \\). It serves dual purposes in our pipeline. It is combined with time embedding as input to each UNet layer," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.507 + ], + "angle": 0, + "content": "while simultaneously providing angular positional priors to the LFE-adapter for extracting view-specific conditioning information for the target view generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.905 + ], + "angle": 0, + "content": "LFE-adapter. Our pipeline leverages priors from a sparse LFE containing four known views. The design incorporates both texture and geometry priors from these input views to provide specific conditioning signals for novel view generation. Drawing inspiration from [33], we propose an LFE-adapter to effectively inject structural and textural information from the known views. The architecture of our LFE-adapter is depicted in the orange region of Fig. 3. The processing pipeline begins by using the \\( V_{\\mathrm{enc}} \\) of LFE-VAE to obtain compressed latent codes \\( Z_{s} \\) from the sparse LFE. We then integrate the angle embedding \\( F_{\\mathrm{ang}} \\) with the latent features of the four input views to obtain the updated latent feature \\( Z_{s}^{*} \\). This feature is processed through a content extraction block (contblock) that extracts textures and details using spatial and angular convolutions. The spatial convolution embeds individual view information, while the angular convolution captures complementary cross-view information. A geometric extraction block (geoblock) further processes the features using cascaded Epipolar Plane Image (EPI) convolutions [5, 24, 43, 52] to embed geometric information, and thus we obtain the updated feature \\( \\hat{Z}_{s} \\). This feature undergoes spatial downsampling to produce \\( \\hat{Z}_{s}^{\\downarrow} \\) for subsequent network updates, followed by angular down-sampling to generate \\( \\hat{Z}_{s}^{\\downarrow \\downarrow} \\) as conditioning signals for corresponding UNet layers. The LFE-adapter ultimately generates a set of condition signals \\( \\hat{Z}_{s1}^{\\downarrow \\downarrow}, \\hat{Z}_{s2}^{\\downarrow \\downarrow}, \\hat{Z}_{s3}^{\\downarrow \\downarrow}, \\hat{Z}_{s4}^{\\downarrow \\downarrow} \\). Each sig" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11211" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of event-based view synthesis on the in-training-scale setting. We tested various methods under synthetic and real settings and evaluated their performance across three metrics. The best results are highlighted in red, the second-best results are highlighted in blue, and the third-best results are highlighted in green. ‘↑’: the higher the better performance, ‘↓’: the opposite." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.133, + 0.905, + 0.288 + ], + "angle": 0, + "content": "
CategoryMethodPSNR ↑Synthetic SSIM ↑LPIPS ↓NIQE ↓Real-world BRISQUE ↓MUSIQ ↑
AngularSRJing et al. [23]21.120.5800.36231.5667.3228.14
DistgASR [44]23.320.6760.27828.8863.9529.90
SAV [12]22.840.6140.28229.1562.7729.25
Guo et al. [21]23.460.6870.29028.2359.1330.72
HybridET-Net [47]+SAV [12]+Vid2E [18]20.230.5740.37332.1167.1125.45
ET-Net [47]+Guo et al.+Vid2E [18]20.880.5890.37531.9566.5825.59
NeRF-basedR2L [42]19.880.5630.40128.4460.2530.45
GenerationOurs24.060.7010.23927.6553.4531.14
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.483, + 0.329 + ], + "angle": 0, + "content": "nal is added to the corresponding same-size input features of the UNet blocks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.331, + 0.483, + 0.574 + ], + "angle": 0, + "content": "LFE View Sampling. Through our LFE-adapter, the diffusion pipeline is enriched with comprehensive structural and textural information from the input LFE. The angular prompts effectively encode target view positions, ensuring angular continuity in the generated LFE views. The pipeline samples individual views sequentially, reconstructing the complete continuous LFE by iteratively processing coordinates for all target views. The framework accommodates LFEs of varying densities through relative coordinate mapping, where \\((0,0)\\) represents the top-left view and \\((1,1)\\) represents the bottom-right view. For example, generating a view at position \\((i,j)\\) in an \\(N\\times N\\) LFE would use normalized coordinates \\((i / (N - 1),j / (N - 1))\\). This normalized coordinate system provides a unified approach for sampling views across different LFE densities while maintaining consistent spatial relationships." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.216, + 0.609 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.285, + 0.635 + ], + "angle": 0, + "content": "5.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Implementation Details. Our networks are trained on a computing cluster equipped with eight NVIDIA A100 GPUs. Following [44], we employ data augmentation techniques including random horizontal flipping, vertical flipping, and \\(90^{\\circ}\\) rotation to enhance model robustness. The optimization process utilizes the Adam optimizer with hyperparameters \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\), processing batches of size 32. We adopt a learning rate scheduling strategy where the initial rate of \\(2 \\times 10^{-4}\\) is halved every 400 epochs. Training concludes after 1500 epochs when convergence is achieved. To ensure fairness in the experiments, all metrics are calculated excluding the input views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Datasets and Metrics. For experimental validation, we utilize two distinct datasets: a synthetic dataset generated using Carla for training and testing, and a real-world dataset captured by our system for additional testing. We use two sets of complementary metrics. For data with ground-truth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.345 + ], + "angle": 0, + "content": "references, we utilize PSNR measured in dB, SSIM [45], and LPIPS [54]. For data without ground-truth references, we employ NIQE [32], BRISQUE [31], and MUSIQ [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.909, + 0.543 + ], + "angle": 0, + "content": "Comparison Methods. We establish comparisons with three categories of related approaches1. The first category comprises LF angular super-resolution methods, including Jing et al.[23], DistgASR[44], SAV [12], and Guo et al.[21]. The second category involves a three-stage approach: we first convert sparse LFE to sparse RGB LF using event-based video reconstruction methods like ET-Net[47], then perform angular super-resolution to get dense RGB LF, and finally convert the dense RGB LF back to dense LFE using Vid2E [18]. The third category explores neural radiance field-based reconstruction using R2L [42]. All the aforementioned baseline methods are re-trained on our dataset loading their original pre-trained parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.554, + 0.707, + 0.57 + ], + "angle": 0, + "content": "5.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.865 + ], + "angle": 0, + "content": "In-training-scale Comparisons. We first conduct in-distribution evaluation on \\(5 \\times 5\\) LFE generation, which matches our training configuration. As shown in Table 1, our S2D-LFE outperforms all baselines across all metrics on both synthetic and real-world testsets. On the synthetic testset, S2D-LFE achieves a notable PSNR improvement of 0.60 dB over the second-best method Guo et al.[21], demonstrating superior reconstruction accuracy. The strong LPIPS performance further indicates enhanced perceptual quality, leveraging the advantages of generative modeling. This is visually evident in Fig.4, where S2D-LFE reconstructs building details with minimal artifacts and better ground truth alignment. In addition, although R2L [42] achieves decent visual results in Fig.4, it fails to reconstruct in certain scenes, leading to lower metric performance on the synthetic dataset. Meanwhile, S2D-LFE maintains its superior performance across all metrics on the real-world testset, demonstrating effective generalization to real-world scenarios. This quantitative advantage is corroborated by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1Since EV-LFV [29] and our S2D-LFE utilize different input modalities, a direct comparison between the two methods is not feasible." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11212" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.087, + 0.895, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 4. The figure presents a qualitative comparison on the central view of the generated LFE, using various methods for both synthetic (top row) and real-world (bottom row) testsets. The proposed S2D-LFE method (labeled as \"Ours\") is compared against other existing techniques, including SAV [12], Guo et al. [21], ET-Net [47] + Guo et al. + Vid2E [18], R2L [42], DistgASR [44], and the ground-truth. The highlighted regions (yellow and green boxes) magnify specific areas to emphasize the differences in event generation quality, with S2D-LFE exhibiting improved fidelity and preservation of fine details in both synthetic and real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.534 + ], + "angle": 0, + "content": "the qualitative results shown in Fig. 4, where S2D-LFE produces higher quality reconstructions with better preservation of details. Due to incorrect estimation of view position, SAV [12] and DistgASR [44] exhibit misaligned results compared to other methods. Furthermore, the ET-Net \\([47] + \\mathrm{SAV} + \\mathrm{Vid2E}\\) [18] approach exhibits severe view aliasing in this scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Out-of-training-scale Comparisons. We extend our evaluation to test the model's generalization capability beyond its training distribution of \\(5 \\times 5\\) LFE, examining its performance on denser LFEs. Among the seven baselines we selected, only R2L [42] and Guo et al. [21] are capable of handling out-of-training-scale LFE generation. Therefore, we only compare our method with these two approaches in this section. As shown in Table 2, our S2D-LFE consistently outperforms baseline methods across all tested LFE densities. While all methods show performance degradation as LFE density increases, S2D-LFE demonstrates greater resilience with a slower degradation rate. The qualitative results in Fig.5 further validate S2D-LFE's superior performance in maintaining perceptual quality and detail preservation for out-of-distribution scenarios. Additional visual comparisons are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Angular Consistency. In this section, we analyze and compare the angular continuity of the LFE generated by our S2D-LFE. Given that the angular continuity of LFE data is strongly correlated with the gradient of the Epipolar Plane Image (EPI) [44, 48], we compute metrics (PSNR, SSIM) on the EPI gradient to evaluate the angular continuity of Guo et al. [21], R2L [42], and S2D-LFE when generating LFEs with different numbers of views (ranging from \\(5 \\times 5\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.424, + 0.907, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.738 + ], + "angle": 0, + "content": "Figure 5. The figure presents the out-of-training-scale qualitative comparison using different methods including Guo et al. [21], R2L [42], and our proposed S2D-LFE (Ours). Besides, synthetic-scene (1) is the (1,1) view from the generated \\(8 \\times 8\\) LFE, and realworld-scene (2) is the (7,7) view from the generated \\(8 \\times 8\\) LFE. The yellow and green boxes highlight specific regions to better illustrate the differences in event generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.909, + 0.906 + ], + "angle": 0, + "content": "to \\(9 \\times 9\\)). It can be observed from Fig. 6 that as the number of views increases, all three methods exhibit a decreasing trend in both PSNR and SSIM metrics while the decline for our S2D-LFE is more gradual. It can be interpreted as S2D-LFE exhibits better angular consistency and robustness compared to the other methods. Moreover, we exhibit a qualitative comparison when generating \\(9 \\times 9\\) LFE in Fig. 7. It can be observed that S2D-LFE preserves fidelity details while producing EPI slices that more closely align with the ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11213" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of event-based view synthesis on the out-of-training-scale setting. We tested various methods under three different settings (generating 36 views, 49 views, and 64 views) and evaluated their performance across three metrics. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.133, + 0.907, + 0.24 + ], + "angle": 0, + "content": "
MethodNIQE↓BRISQUE↓MUSIQ↑
36-view49-views64-views36-views49-views64-views36-views49-views64-views
SynGuo et al. [21]16.8518.6220.1059.4564.1270.7831.2830.0228.31
R2L [42]16.9619.0521.5461.7265.9371.5630.2729.0827.29
Ours15.8417.6519.6754.1558.2163.4932.3031.8930.34
RealGuo et al. [21]28.8129.5830.2962.8366.1570.5830.2729.7829.30
R2L [42]28.9429.7230.6164.0068.1871.3729.9929.4829.02
Ours27.9928.4729.0156.1758.9161.2930.9330.6530.21
" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.257, + 0.436, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Figure 6. The figure presents a quantitative comparison of angular consistency across different methods, including Guo et al., R2L, and our proposed S2D-LFE. Subfigures (a) and (b) depict PSNR (dB) and SSIM metrics respectively, when generating different numbers of views ranging from \\(5 \\times 5\\) to \\(9 \\times 9\\). All methods exhibit a decline in both metrics as the number of views increases, but S2D-LFE exhibits a relatively slower decline." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.51, + 0.189, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.51, + 0.263, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.51, + 0.333, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.51, + 0.404, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.51, + 0.476, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Figure 7. The figure presents a qualitative comparison when generating \\(9 \\times 9\\) LFE using Guo et al., R2L, and our proposed S2D-LFE. The green boxes highlight spatial slices of view (1,1), while the yellow boxes highlight EPI slices." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.676, + 0.242, + 0.693 + ], + "angle": 0, + "content": "6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.868 + ], + "angle": 0, + "content": "LFE-VAE. We examine the effectiveness of our LFE-VAE by making comparisons with two variants: one using a standard VAE (S-VAE) and another using a fine-tuned VAE (FT-VAE) with single-view event data. As shown in Table 3, S-VAE variant shows performance degradation across all metrics, indicating reduced reconstruction quality and perceptual similarity. While FT-VAE variant shows modest improvements with PSNR of 37.12 dB and SSIM of 0.865, its LPIPS score of 0.151 still indicates suboptimal perceptual quality. In contrast, our LFE-VAE better captures the structural and perceptual characteristics of light field events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "LFE-adapter. To validate the effectiveness of our LFE-adapter design, we make comparisons with a variant that" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.254, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of the ablation on the LFE-VAE. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. '↑': the higher the better performance, '↓': the opposite." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.312, + 0.906, + 0.371 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
S-VAE36.890.8590.158
FT-VAE37.120.8650.151
LFE-VAE (Ours)38.540.8810.112
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.377, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation of the ablation on the LFE-adapter. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.455, + 0.906, + 0.507 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
Concat22.540.6310.275
LFE-adapter24.060.7010.239
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.636 + ], + "angle": 0, + "content": "replaces the LFE-adapter with simple feature concatenation (concat). As shown in Table 4, the concatenation variant shows substantial performance degradation across all metrics, indicating significant reductions in reconstruction accuracy, structural fidelity, and perceptual quality. The comparison clearly validates that our LFE-adapter provides more effective feature integration compared to naive concatenation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.647, + 0.634, + 0.661 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.82 + ], + "angle": 0, + "content": "In this paper, we present S2D-LFE, a novel approach for generating consistent, full-view LFE. Our key contributions include developing the first real-world LFE capture system and proposing effective technical components: an LFE-VAE for LFE compression and an LFE-Adapter for prior injection. S2D-LFE achieves simplicity by eliminating multi-view RGB dependencies, controllability through precise view synthesis, and consistency via cross-view and temporal coherence. These advances establish S2D-LFE as an effective solution for dense LFE generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.673, + 0.846 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.851, + 0.906, + 0.896 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11214" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] A N Angelopoulos, J N Martel, A P Kohli, J Conradt, and G Wetzstein. Event based, near eye gaze tracking beyond 10,000 hz. IEEE Transactions on Visualization and Computer Graphics, page 1, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19819-19829, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel, and Tobias Ritschel. X-fields: Implicit neural view-, light- and time-image interpolation. ACM TOG, 39(6):1-15, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] A Bhattacharya, R Madaan, F Cladera, et al. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5846-5855, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[5] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. pages 497-504. 2001. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[8] H Chen, J Gu, A Chen, et al. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[9] P Chen, W Guan, F Huang, et al. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.644, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] P Chen, W Guan, and P Lu. Espvio: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8 (6):3661-3668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[11] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[12] Zhen Cheng, Yutong Liu, and Zhiwei Xiong. Spatial-angular versatile convolution for light field reconstruction. IEEE Transactions on Computational Imaging, 8:1131-1144, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] A Dosovitskiy, G Ros, F Codevilla, et al. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Brandon Yushan Feng and Amitabh Varshney. Signet: Efficient neural representation for light fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14224-14233, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[16] P Furgale, J Rehder, and R Siegwart. Unified temporal and spatial calibration for multi-sensor systems. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1280-1286. IEEE, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[18] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[19] M Gehrig, W Aarents, D Gehrig, and D Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[21] Mantang Guo, Junhui Hou, Jing Jin, Hui Liu, Huanqiang Zeng, and Jiwen Lu. Content-aware warping for view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9486-9503, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[23] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[24] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM TOG, 35(6):1-10, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[25] J Karras, A Holynski, T C Wang, et al. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[26] B Ke, A Obukhov, S Huang, et al. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492–9502, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11215" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[29] Zhicheng Lu, Xiaoming Chen, Vera Yuk Ying Chung, Weidong Cai, and Yiran Shen. Ev-lfv: Synthesizing light field event streams from an event camera and multiple rgb cameras. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[30] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[32] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[33] C Mou, X Wang, L Xie, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[36] R Rombach, A Blattmann, D Lorenz, et al. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[37] S Saxena, C Herrmann, J Hur, et al. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[38] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. NeurIPS, 34:19313-19325, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[39] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[40] S M N Uddin, S H Ahmed, and Y J Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[41] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[42] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In European Conference on Computer Vision, pages 612-629. Springer, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[43] Yunlong Wang, Fei Liu, Zilei Wang, Guangqi Hou, Zhenan Sun, and Tieniu Tan. End-to-end view synthesis for light field imaging with pseudo 4dcnn. In ECCV, pages 333-348, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[44] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[47] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[48] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[49] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[50] J Wynn and D Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[51] J Xiang, J Yang, B Huang, et al. 3d-aware image generation using 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2383-2393, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[52] Henry Wing Fung Yeung, Junhui Hou, Jie Chen, Yuk Ying Chung, and Xiaoming Chen. Fast light field reconstruction with deep coarse-to-fine modeling of spatial-angular clues. In ECCV, pages 137-152, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[53] L Zhang, A Rao, and M Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[55] Y Zhou, G Gallego, and S Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37(5):1433-1450, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11216" + } + ] +] \ No newline at end of file diff --git a/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_origin.pdf b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bdfaf7ebd90192fd367b75877956171584cd9048 --- /dev/null +++ b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/f17a4408-3f4e-4639-92d7-cce01abd5b00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8fb302cbca121b4538b6e6a4acd12897a2b12700cb45893dbdf7fb48f003abb +size 3207339 diff --git a/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/full.md b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..55d37714b896387607ef802fed791b9362fda1a2 --- /dev/null +++ b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/full.md @@ -0,0 +1,242 @@ +# S2D-LFE: Sparse-to-Dense Light Field Event Generation + +Yutong Liu Wenming Weng Yueyi Zhang Zhiwei Xiong* University of Science and Technology of China {ustclyt, wmweng}@mail.ustc.edu.cn {zhyuey, zwxiong}@ustc.edu.cn + +# Abstract + +In this paper, we present S2D-LFE, an innovative approach for sparse-to-dense light field event generation. For the first time to our knowledge, S2D-LFE enables controllable novel view synthesis only from sparse-view light field event (LFE) data, and addresses three critical challenges for the LFE generation task: simplicity, controllability, and consistency. The simplicity aspect eliminates the dependency on frame-based modality, which often suffers from motion blur and low frame-rate limitations. The controllability aspect enables precise view synthesis under sparse LFE conditions with view-related constraints. The consistency aspect ensures both cross-view and temporal coherence in the generated results. To realize S2D-LFE, we develop a novel diffusion-based generation network with two key components. First, we design an LFE-customized variational auto-encoder that effectively compresses and reconstructs LFE by integrating cross-view information. Second, we design an LFE-aware injection adaptor to extract comprehensive geometric and texture priors. Furthermore, we construct a large-scale synthetic LFE dataset containing 162 one-minute sequences using simulator, and capture a real-world testset using our custom-built sparse LFE acquisition system, covering diverse indoor and outdoor scenes. Extensive experiments demonstrate that S2D-LFE successfully generates up to $9 \times 9$ dense LFE from $2 \times 2$ sparse inputs and outperforms existing methods on both synthetic and real-world data. The datasets and code are available at https://github.com/Yutong2022/S2D-LFE. + +# 1. Introduction + +Event cameras capture per-pixel brightness changes asynchronously with microsecond temporal resolution [1], while Light Field (LF) cameras record both the accumulated intensity and direction of light rays, thereby revealing scene geometric structure [48]. The integration of these technologies leads to Light Field Event (LFE) cameras, which capture multi-view event streams arranged in a fixed baseline array. This combination enables simultaneous recording + +of high-quality spatial depth information and dynamic motion data with exceptional temporal resolution. Such capabilities demonstrate significant potential for various applications, including 6-degree-of-freedom virtual reality [23] and real-time dynamic depth estimation [40]. The rich spatio-temporal information provided by LFE data enables effective capture of structural information in high-speed dynamic scenes. + +In this paper, we aim to address a largely unexplored task of LFE generation. To date, the only relevant work is Ev-LFV [29], which combines an RGB LF camera with an event camera aligned to the LF center-view for LFE synthesis. However, this approach faces several critical challenges: (1) Simplicity. Ev-LFV's reliance on multi-view LF frame data as auxiliary information introduces complications. The RGB data is vulnerable to motion blur and exposure issues, potentially compromising LFE generation quality. Moreover, achieving precise spatial-temporal alignment between the LF and event cameras presents substantial technical challenges. (2) Controllability. The fixed view synthesis constraint of Ev-LFV prevents the generation of free viewpoints under sparse LFE conditions with view-related constraints. (3) Consistency. Due to the limitations of low frame-rate RGB data and insufficient structural modeling, Ev-LFV struggles to maintain both angular and temporal continuity in the generated LFE sequences. + +To address these challenges, we propose S2D-LFE for Sparse-to-Dense Light Field Event generation, which is a novel approach for LFE generation. S2D-LFE demonstrates the capability to generate temporally consistent LFE from free viewpoints using only sparse LFE input, making it particularly suitable for real-world applications without the need for frame cameras. Our approach comprises two key components: an LFE-customized variational auto-encoder for efficient LFE compression and reconstruction, and a LFE-aware injection adapter integrated into a diffusion-based generation network that learns to generate the latent representation of LFE. + +The first key component, LFE-VAE, is designed to effectively compress high-resolution LFE while preserving spatial and structural information. Specifically, we introduce + +two significant improvements over conventional pretrained VAE architectures used in diffusion models [25, 26, 35, 36, 53]. First, we incorporate angular interaction blocks into the VAE decoder to fully exploit cross-view information within LFE. Second, we implement a view-enhancement mechanism that utilizes input sparse LFE views to improve the quality of reconstructed novel views. These enhancements effectively leverage the inherent cross-view information in LFE and mitigate compression artifacts, resulting in high-fidelity reconstruction. The second key component, LFE-adapter, is designed to utilize geometric and texture priors from the input sparse LFE to guide the diffusion process. Specifically, LFE-adapter generates pyramid condition signals that provide targeted conditioning at multiple resolution scales for the U-Net during diffusion. This multi-scale conditioning enables more precise injection of geometric and texture priors, thereby achieving superior cross-view consistency. + +Moreover, we develop the first sparse LFE capture system to facilitate the acquisition of continuous light field event data. The system consists of four Davis346 cameras arranged in a $2 \times 2$ array configuration. Using this setup, we collect a diverse dataset comprising 25 sequences, including 15 indoor scenes and 10 autonomous driving scenarios. To support comprehensive model training, we additionally construct a large-scale synthetic dataset using Carla simulator [14], generating 162 one-minute dense LFE sequences. + +The main contributions of this paper can be summarized as follows: + +- We propose S2D-LFE, a novel approach for generating free-viewpoint and temporally consistent light field events from sparse-view LFE input. +- Our S2D-LFE incorporates two key components: an LFE-VAE for efficient LFE compression, and an LFE-adapter integrated into a diffusion-based generation network that reconstructs the latent representation of LFE. +- We develop the first sparse LFE capture system and collect a dataset of 25 real-world sparse LFE sequences. Additionally, we synthesize 162 dense LFE sequences using the Carla simulator to facilitate comprehensive training. +- Our approach outperforms existing methods on both synthetic and real-world datasets, while presenting enhanced robustness in generalizing to free-viewpoint synthesis. + +# 2. Related Work + +Multi-view Event Processing. While event cameras offer exceptional temporal resolution and low latency [1], monocular event data alone is insufficient for capturing comprehensive structural information of dynamic scenes. Various approaches have been proposed to address this limitation: 1) NeRF-based scene representation learning [4], 2) hybrid systems combining events with conventional cameras [9, 29], and 3) stereo event systems [10, 19, 40, 55]. + +NeRF-based methods, while promising, require per-scene training and multiple viewpoints for effective scene representation. For example, EvDNeRF [4] introduces an event-based dynamic NeRF pipeline for reconstructing event streams in rapidly deforming scenes, but its performance degrades significantly with limited viewpoints $(< 8)$ . Hybrid systems face challenges in reconciling the temporal resolution disparity between different modalities. For instance, Lu et al. [29] encounter practical limitations in real-world applications due to the substantial difference in temporal resolution between LF and event cameras. While stereo event systems can capture geometric information, the resulting geometry tends to be sparse. In contrast, the Light Field event modality offers distinct advantages by capturing dense structural information of dynamic scenes without requiring per-scene training, while enabling multiple angular resolution generation. + +Light Field View Synthesis. The reconstruction of dense Light Fields from sparse views, commonly known as LF angular super-resolution, remains an active research area. Existing approaches can be broadly categorized into two groups: depth-dependent [5, 24, 43, 52] and depth-independent methods [12, 21, 23, 44]. Depth-dependent methods rely on precise disparity estimation but are vulnerable to occlusions, while depth-independent methods often struggle with generalization across varying disparity ranges. Recent advances in Neural Light Field representations [2, 3, 15, 28, 38, 42] have brought significant improvements to LF view synthesis by addressing limitations of conventional angular super-resolution methods. Compared with the widely adopted learnable 3D representation methods [7, 11, 34, 41], these approaches can more effectively exploit the spatial priors inherent in LF. However, these NeRF-based approaches are constrained by their requirement for numerous input views and scene-specific training. In contrast, our proposed S2D-LFE conquers these limitations by enabling free-viewpoint synthesis from sparse inputs without the need for per-scene training, while maintaining high recovery quality and temporal consistency. + +Diffusion-based Image Generation. Diffusion models have shown remarkable progress in novel view synthesis [6, 8, 20, 39, 46, 50, 51]. For example, NeRDi [20] leverages language priors for multi-view synthesis, effectively connecting image semantics to appearance reconstruction. These advances demonstrate the robust capability of diffusion models in modeling multi-view distributions. Furthermore, diffusion models have achieved impressive results across diverse data modalities, including depth [26], point clouds [30], optical flow [26, 37], and human pose [25]. Motivated by this versatility and strong fitting capability across different modalities, we adopt diffusion models for dense LFE generation. + +# 3. Real-world and Synthetic LFE Datasets + +# 3.1. LFE Capture System and Real-world Dataset + +System Setup. To facilitate the collection of real-world LFE data and validate our approach, we develop a custom LFE capture system. Given the lack of commercially available LFE cameras, our system design prioritizes precise spatial alignment and temporal synchronization while maintaining cost-effectiveness. The system, shown in Fig. 1, consists of a $2 \times 2$ array of Davis346 cameras mounted on a square aluminum alloy frame, with uniform horizontal and vertical baseline lengths of $6\mathrm{cm}$ . To ensure consistent imaging characteristics across views, all four cameras are equipped with identical lenses and configured with matching parameters, including focal length and exposure time. For precise temporal synchronization, we implement a master-slave configuration using external cables (depicted as black lines in Fig.1). The upper-left DAVIS346 serves as the master camera, transmitting trigger signal pulses to the three slave cameras through these cables. This system captures event data at a spatial resolution of $346 \times 260$ pixels, which can be readily upgraded to a higher resolution by updating event cameras. Detailed parameter configurations are provided in the supplementary materials. + +Camera Calibration. Precise camera calibration is crucial for accurate Light Field Event capture, as it ensures geometric consistency across multiple views and enables reliable reconstruction of scene structure. This is particularly important for our system where accurate spatial relationships between cameras directly impact the quality of captured LFE data. We utilize the Kalibr toolbox [16] to estimate camera intrinsics and extrinsics by capturing fixed-rate frames of a checkerboard pattern. Since the DAVIS346 cameras output both frames and event streams through the same optical system, the calibration parameters obtained from frames are directly applicable to event streams. Following calibration, we perform systematic rectification of the captured sparse LFE. Using the upper left camera as reference, we first apply perspective transformation to project the upper right camera onto a virtual plane parallel to the reference camera's horizontal plane, establishing the baseline length between them. We then align the two lower cameras through perspective transformations to their corresponding positions in the sparse LF, ensuring uniform baseline lengths across the array. This comprehensive calibration and rectification process ensures accurate structural information capture of dynamic scenes. + +Real-world LFE Dataset. Following system calibration and rectification, we collect a real LFE dataset comprising 25 dynamic scenes using our LFE capture system, which is termed "LFE-real". The dataset consists of two main categories: 15 indoor sequences captured by moving our camera array through various indoor environments (featuring sub + +![](images/9ab701736589bc8e08a689fcfe9bb5fe4453fbf480f2cb1727bc7001c85226cf.jpg) +Figure 1. The proposed LFE capturing system. + +jects such as people, plants, and interior layouts), and 10 autonomous driving sequences recorded from displayed autonomous driving footage. Each sequence spans between 3 to 5 minutes, ensuring sufficient temporal coverage of the dynamic events. The details about scenes are in the supplementary material. + +# 3.2. Synthetic Dataset + +Given that physical capture devices can only provide sparse view inputs without dense multi-view supervision, we leverage the Carla simulator [14] to generate a large synthetic LFE dataset for training and evaluation, which is termed "LFE-syn". We utilize 18 different maps provided by the Carla environment for data simulation. Each map is divided into 9 regions, with a designated starting point in each region. At these starting points, we deploy virtual vehicles equipped with $5 \times 5$ event camera arrays. The baseline lengths between adjacent cameras are randomly set between $3\mathrm{cm}$ and $9\mathrm{cm}$ . The event trigger thresholds are configured asymmetrically: 0.1 to 0.15 for positive events and -0.15 to -0.1 for negative events. Using autonomous navigation, each vehicle generates one-minute LFE sequences, resulting in a total of 162 synthetic sequences. Additionally, to facilitate the evaluation, our synthetic dataset share the same resolution with the real-world dataset. Detailed parameter configurations are provided in the supplementary materials. + +# 4. S2D-LFE Generation Network + +Our custom-built system captures sparse LFE with coarse geometric information of dynamic scenes. To enhance the structural representation capabilities, we propose S2D-LFE, a novel approach for synthesizing accurate dense LFEs from sparse inputs. We detail our approach in the following sections. + +# 4.1. Overview + +In our setting, each event stream view in the LFE is accumulated as an Event Count Map (ECM) [17] with shapes of $\mathbb{R}^{B\times TB\times 2\times H\times W}$ , where $B$ represents the batch size, + +![](images/768a75bddd89b0bb7371aee53791228bcb5956d7aa419df5233b3818c1697558.jpg) +Figure 2. The second training stage of the LFE-VAE architecture. The input to the LFE-VAE supports multiple angular resolutions of LFE. Here, we take the case of an input with 25 views as an example. The LFE-VAE is composed of the encoder $V_{enc}$ , the decoder $\hat{V}_{dec}$ inserted by angular interaction blocks, and the following angular resblocks. The training at this stage is aimed at updating the decoder $\hat{V}_{dec}$ and the angular resblocks. + +$TB$ represents the temporal bins, 2 denotes the polarities, and $H\times W$ denotes the spatial resolution. To facilitate the subsequent processing of the VAE, we transform the two-polarity channels of the event stream into three channels based on [49]. A LFE can be represented as $L\in \mathbb{R}^{B\times 3\times M\times N\times H\times W}$ , where $M\times N$ denotes the angular resolution. The LFE synthesis task aims to generate a dense LFE $L_{D}\in \mathbb{R}^{B\times 3\times M\times N\times H\times W}$ from a sparse LFE $L_{S}\in \mathbb{R}^{B\times 3\times m\times n\times H\times W}$ , where $m\times n$ represents a lower angular resolution. To effectively model the complex distribution of LFE data, we adopt a diffusion-based pipeline. Given the high-dimensional nature of LFE data, we first introduce LFE-VAE to encode LFE into a compact latent representation. The encoded sparse LFE, together with the encoded angular coordinates, serve as conditioning signals for the diffusion model to synthesize views at specified angular positions. To effectively incorporate these conditioning signals into the diffusion backbone, we design an LFE-adapter for optimal signal processing. Through iterative generation of all unknown views, we achieve the transformation from sparse to dense LFE. Notably, our diffusion network demonstrates the capability to synthesize novel views at arbitrary angles beyond the training set by accepting different angular coordinates as input. + +# 4.2. LFE-VAE + +To handle the high spatial resolution of LFE data efficiently, we employ Variational Autoencoder (VAE) architectures commonly used in diffusion models [25, 26, 35, 36, 53] for data compression. However, directly applying conventional VAEs to LFE data presents two significant challenges: 1) Traditional VAEs used in stable diffusion are designed for single-view processing, failing to leverage the inherent cross-view structural information in LFE data, and 2) The lossy nature of VAE compression inevitably results in information loss in the event streams. + +To address these limitations, we propose LFE-VAE, building upon the pretrained VAE model from SDXL [35]. + +Our architecture introduces two key innovations: (1) angular interaction blocks in the decoder to incorporate cross-view complementary information, and (2) a view-enhancement mechanism utilizing four known views to minimize compression artifacts. + +In specific, for each dense LFE input $L_{in}$ , the encoder first generates a latent code $Z_{s}$ . This latent representation is processed by our enhanced decoder, which embeds cross-view geometric information through angular interaction blocks. Each block consists of two cascaded layers of spatial and angular convolutions. The corresponding positions of the decoded target views are replaced with four known sparse LFE views, and then the LFE is refined through several cascaded resblocks defined on the LFE angular slices. + +Our training process follows a two-stage approach. First, we fine-tune the stable diffusion VAE using single-view events to obtain updated weights for the encoder $V_{enc}$ and decoder $V_{dec}$ . Subsequently, while keeping $V_{enc}$ frozen, we fine-tune the proposed decoder $\hat{V}_{dec}$ and angular resblocks. The second stage of the LFE-VAE architecture is illustrated in detail in Fig. 2. Specifically, we first select the sparse LFE views $L_{S}$ , which represent the condition signal for the subsequent diffusion pipeline. The input LFE $L_{in}$ is then encoded by $V_{enc}$ and decoded by the updated $\hat{V}_{dec}$ , resulting in an intermediate reconstructed LFE. Next, we replace the four corner views of the recovered LFE, with the four corresponding views in $L_{S}$ selected in $L_{in}$ . Finally, the compression degradation is restored through the utilization of angular resblocks. After training, the encoder $V_{enc}$ is integrated into the diffusion training pipeline to compress input sparse LFE, enabling the subsequent diffusion network to handle higher-resolution LFE generation. + +# 4.3. Diffusion Pipeline + +Following mainstream diffusion architectures [13, 22, 25, 26, 35, 36, 53], our diffusion backbone adopts a UNet structure. While existing diffusion models primarily han- + +![](images/42608d48f053ec2b126a00dfc3f2f39a848f22e45896ac8fa657580dbd757254.jpg) +Figure 3. The diffusion pipeline of S2D-LFE at the t-th time step. The pipeline primarily consists of two parts: (1) The left part, highlighted with a blue background, represents the main pipeline diagram. The diffusion pipeline of S2D-LFE adopts a UNet backbone, where the latent features $Z_{s}$ from four viewpoints are injected into the network as conditional signals via the LFE-adapter. In addition to the time embedding, our prompt incorporates positional encodings of both horizontal and vertical coordinates. (2) The right section, highlighted with an orange background, depicts the LFE-adapter block within the LFE-adapter. After cascading the condition information through the ContBlock and GeoBlock, the output features are embedded into the corresponding layers of the UNet. The LFE-adapter consists of four LFE-adapter blocks, each producing a different conditional signal $\hat{Z}_{\mathrm{s1}}^{\downarrow \downarrow}, \hat{Z}_{\mathrm{s2}}^{\downarrow \downarrow}, \hat{Z}_{\mathrm{s3}}^{\downarrow \downarrow}, \hat{Z}_{\mathrm{s4}}^{\downarrow \downarrow}$ corresponding to the respective layers of the UNet. + +![](images/5deff887cfac4342786502fd0dc7242c93cf4c8a19fcad00ba2fcc41352313db.jpg) + +dle unstructured conditioning data such as images, videos, and text, the potential of structured conditioning data, which contains rich structural and texture priors, remains largely unexplored. To address this gap, we propose a specialized diffusion pipeline tailored for LFE data. At its core is the novel LFE-adapter, which serves two key functions: 1) extracting features that encode both structural and texture information from input sparse LFE and 2) injecting these extracted features into the diffusion backbone. With the structural and texture information effectively embedded, the diffusion pipeline generates specific LFE views based on input view prompts. Our approach employs continuous, linear angular coordinate encoding for view prompts, enabling the pipeline to generalize to sampling LFE views across varying densities and provide flexibility in LFE generation. The complete architecture of our diffusion pipeline at the t-th time step is illustrated in Fig. 3. Then, we will detail each of these innovations. + +LFE Angular Prompt. Accurate positional priors of target views are essential for novel LF view generation, which we achieve through angular coordinate prompting. As illustrated in Fig. 3, we process 2D coordinates through a two-stage encoding scheme. First, we independently encode row and column coordinates. These encoded coordinates are then transformed through cascaded fully connected layers to obtain the raw and column embeddings, which are then concatenated and fed into a linear layer to get the angular embedding $F_{ang}$ . It serves dual purposes in our pipeline. It is combined with time embedding as input to each UNet layer, + +while simultaneously providing angular positional priors to the LFE-adapter for extracting view-specific conditioning information for the target view generation. + +LFE-adapter. Our pipeline leverages priors from a sparse LFE containing four known views. The design incorporates both texture and geometry priors from these input views to provide specific conditioning signals for novel view generation. Drawing inspiration from [33], we propose an LFE-adapter to effectively inject structural and textural information from the known views. The architecture of our LFE-adapter is depicted in the orange region of Fig. 3. The processing pipeline begins by using the $V_{\mathrm{enc}}$ of LFE-VAE to obtain compressed latent codes $Z_{s}$ from the sparse LFE. We then integrate the angle embedding $F_{\mathrm{ang}}$ with the latent features of the four input views to obtain the updated latent feature $Z_{s}^{*}$ . This feature is processed through a content extraction block (contblock) that extracts textures and details using spatial and angular convolutions. The spatial convolution embeds individual view information, while the angular convolution captures complementary cross-view information. A geometric extraction block (geoblock) further processes the features using cascaded Epipolar Plane Image (EPI) convolutions [5, 24, 43, 52] to embed geometric information, and thus we obtain the updated feature $\hat{Z}_{s}$ . This feature undergoes spatial downsampling to produce $\hat{Z}_{s}^{\downarrow}$ for subsequent network updates, followed by angular down-sampling to generate $\hat{Z}_{s}^{\downarrow \downarrow}$ as conditioning signals for corresponding UNet layers. The LFE-adapter ultimately generates a set of condition signals $\hat{Z}_{s1}^{\downarrow \downarrow}, \hat{Z}_{s2}^{\downarrow \downarrow}, \hat{Z}_{s3}^{\downarrow \downarrow}, \hat{Z}_{s4}^{\downarrow \downarrow}$ . Each sig + +Table 1. Quantitative evaluation of event-based view synthesis on the in-training-scale setting. We tested various methods under synthetic and real settings and evaluated their performance across three metrics. The best results are highlighted in red, the second-best results are highlighted in blue, and the third-best results are highlighted in green. ‘↑’: the higher the better performance, ‘↓’: the opposite. + +
CategoryMethodPSNR ↑Synthetic SSIM ↑LPIPS ↓NIQE ↓Real-world BRISQUE ↓MUSIQ ↑
AngularSRJing et al. [23]21.120.5800.36231.5667.3228.14
DistgASR [44]23.320.6760.27828.8863.9529.90
SAV [12]22.840.6140.28229.1562.7729.25
Guo et al. [21]23.460.6870.29028.2359.1330.72
HybridET-Net [47]+SAV [12]+Vid2E [18]20.230.5740.37332.1167.1125.45
ET-Net [47]+Guo et al.+Vid2E [18]20.880.5890.37531.9566.5825.59
NeRF-basedR2L [42]19.880.5630.40128.4460.2530.45
GenerationOurs24.060.7010.23927.6553.4531.14
+ +nal is added to the corresponding same-size input features of the UNet blocks. + +LFE View Sampling. Through our LFE-adapter, the diffusion pipeline is enriched with comprehensive structural and textural information from the input LFE. The angular prompts effectively encode target view positions, ensuring angular continuity in the generated LFE views. The pipeline samples individual views sequentially, reconstructing the complete continuous LFE by iteratively processing coordinates for all target views. The framework accommodates LFEs of varying densities through relative coordinate mapping, where $(0,0)$ represents the top-left view and $(1,1)$ represents the bottom-right view. For example, generating a view at position $(i,j)$ in an $N\times N$ LFE would use normalized coordinates $(i / (N - 1),j / (N - 1))$ . This normalized coordinate system provides a unified approach for sampling views across different LFE densities while maintaining consistent spatial relationships. + +# 5. Experiment + +# 5.1. Experiment Settings + +Implementation Details. Our networks are trained on a computing cluster equipped with eight NVIDIA A100 GPUs. Following [44], we employ data augmentation techniques including random horizontal flipping, vertical flipping, and $90^{\circ}$ rotation to enhance model robustness. The optimization process utilizes the Adam optimizer with hyperparameters $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ , processing batches of size 32. We adopt a learning rate scheduling strategy where the initial rate of $2 \times 10^{-4}$ is halved every 400 epochs. Training concludes after 1500 epochs when convergence is achieved. To ensure fairness in the experiments, all metrics are calculated excluding the input views. + +Datasets and Metrics. For experimental validation, we utilize two distinct datasets: a synthetic dataset generated using Carla for training and testing, and a real-world dataset captured by our system for additional testing. We use two sets of complementary metrics. For data with ground-truth + +references, we utilize PSNR measured in dB, SSIM [45], and LPIPS [54]. For data without ground-truth references, we employ NIQE [32], BRISQUE [31], and MUSIQ [27]. + +Comparison Methods. We establish comparisons with three categories of related approaches1. The first category comprises LF angular super-resolution methods, including Jing et al.[23], DistgASR[44], SAV [12], and Guo et al.[21]. The second category involves a three-stage approach: we first convert sparse LFE to sparse RGB LF using event-based video reconstruction methods like ET-Net[47], then perform angular super-resolution to get dense RGB LF, and finally convert the dense RGB LF back to dense LFE using Vid2E [18]. The third category explores neural radiance field-based reconstruction using R2L [42]. All the aforementioned baseline methods are re-trained on our dataset loading their original pre-trained parameters. + +# 5.2. Comparison Results + +In-training-scale Comparisons. We first conduct in-distribution evaluation on $5 \times 5$ LFE generation, which matches our training configuration. As shown in Table 1, our S2D-LFE outperforms all baselines across all metrics on both synthetic and real-world testsets. On the synthetic testset, S2D-LFE achieves a notable PSNR improvement of 0.60 dB over the second-best method Guo et al.[21], demonstrating superior reconstruction accuracy. The strong LPIPS performance further indicates enhanced perceptual quality, leveraging the advantages of generative modeling. This is visually evident in Fig.4, where S2D-LFE reconstructs building details with minimal artifacts and better ground truth alignment. In addition, although R2L [42] achieves decent visual results in Fig.4, it fails to reconstruct in certain scenes, leading to lower metric performance on the synthetic dataset. Meanwhile, S2D-LFE maintains its superior performance across all metrics on the real-world testset, demonstrating effective generalization to real-world scenarios. This quantitative advantage is corroborated by + +![](images/5785cd29695ed92bf8fc992bc554d3e67472ab68881926467ca60cecfb7ceef7.jpg) +Figure 4. The figure presents a qualitative comparison on the central view of the generated LFE, using various methods for both synthetic (top row) and real-world (bottom row) testsets. The proposed S2D-LFE method (labeled as "Ours") is compared against other existing techniques, including SAV [12], Guo et al. [21], ET-Net [47] + Guo et al. + Vid2E [18], R2L [42], DistgASR [44], and the ground-truth. The highlighted regions (yellow and green boxes) magnify specific areas to emphasize the differences in event generation quality, with S2D-LFE exhibiting improved fidelity and preservation of fine details in both synthetic and real-world scenarios. + +the qualitative results shown in Fig. 4, where S2D-LFE produces higher quality reconstructions with better preservation of details. Due to incorrect estimation of view position, SAV [12] and DistgASR [44] exhibit misaligned results compared to other methods. Furthermore, the ET-Net $[47] + \mathrm{SAV} + \mathrm{Vid2E}$ [18] approach exhibits severe view aliasing in this scene. + +Out-of-training-scale Comparisons. We extend our evaluation to test the model's generalization capability beyond its training distribution of $5 \times 5$ LFE, examining its performance on denser LFEs. Among the seven baselines we selected, only R2L [42] and Guo et al. [21] are capable of handling out-of-training-scale LFE generation. Therefore, we only compare our method with these two approaches in this section. As shown in Table 2, our S2D-LFE consistently outperforms baseline methods across all tested LFE densities. While all methods show performance degradation as LFE density increases, S2D-LFE demonstrates greater resilience with a slower degradation rate. The qualitative results in Fig.5 further validate S2D-LFE's superior performance in maintaining perceptual quality and detail preservation for out-of-distribution scenarios. Additional visual comparisons are provided in the supplementary material. + +Angular Consistency. In this section, we analyze and compare the angular continuity of the LFE generated by our S2D-LFE. Given that the angular continuity of LFE data is strongly correlated with the gradient of the Epipolar Plane Image (EPI) [44, 48], we compute metrics (PSNR, SSIM) on the EPI gradient to evaluate the angular continuity of Guo et al. [21], R2L [42], and S2D-LFE when generating LFEs with different numbers of views (ranging from $5 \times 5$ + +![](images/ac77e20a8a93f083f1158f6c4133177e879179a75866efbb6e47cc28cb31d062.jpg) +Figure 5. The figure presents the out-of-training-scale qualitative comparison using different methods including Guo et al. [21], R2L [42], and our proposed S2D-LFE (Ours). Besides, synthetic-scene (1) is the (1,1) view from the generated $8 \times 8$ LFE, and realworld-scene (2) is the (7,7) view from the generated $8 \times 8$ LFE. The yellow and green boxes highlight specific regions to better illustrate the differences in event generation quality. + +to $9 \times 9$ ). It can be observed from Fig. 6 that as the number of views increases, all three methods exhibit a decreasing trend in both PSNR and SSIM metrics while the decline for our S2D-LFE is more gradual. It can be interpreted as S2D-LFE exhibits better angular consistency and robustness compared to the other methods. Moreover, we exhibit a qualitative comparison when generating $9 \times 9$ LFE in Fig. 7. It can be observed that S2D-LFE preserves fidelity details while producing EPI slices that more closely align with the ground truth. + +Table 2. Quantitative evaluation of event-based view synthesis on the out-of-training-scale setting. We tested various methods under three different settings (generating 36 views, 49 views, and 64 views) and evaluated their performance across three metrics. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite. + +
MethodNIQE↓BRISQUE↓MUSIQ↑
36-view49-views64-views36-views49-views64-views36-views49-views64-views
SynGuo et al. [21]16.8518.6220.1059.4564.1270.7831.2830.0228.31
R2L [42]16.9619.0521.5461.7265.9371.5630.2729.0827.29
Ours15.8417.6519.6754.1558.2163.4932.3031.8930.34
RealGuo et al. [21]28.8129.5830.2962.8366.1570.5830.2729.7829.30
R2L [42]28.9429.7230.6164.0068.1871.3729.9929.4829.02
Ours27.9928.4729.0156.1758.9161.2930.9330.6530.21
+ +![](images/0b5a34f7c996ad3d86ec5977f4acf01db5be8fb186671be013d2b6eae0e87ee2.jpg) +Figure 6. The figure presents a quantitative comparison of angular consistency across different methods, including Guo et al., R2L, and our proposed S2D-LFE. Subfigures (a) and (b) depict PSNR (dB) and SSIM metrics respectively, when generating different numbers of views ranging from $5 \times 5$ to $9 \times 9$ . All methods exhibit a decline in both metrics as the number of views increases, but S2D-LFE exhibits a relatively slower decline. + +![](images/cd465763e4f846f4ddbce5eb4de5cb969309fa04060ec249da2121ad107df416.jpg) +Figure 7. The figure presents a qualitative comparison when generating $9 \times 9$ LFE using Guo et al., R2L, and our proposed S2D-LFE. The green boxes highlight spatial slices of view (1,1), while the yellow boxes highlight EPI slices. + +![](images/f4e4c381ae88f3504a8219a51a2dbaeb7b3b729709f900c8f571a74e35b41423.jpg) + +![](images/776ed85c92cb1415e633d2b19013c5cb3e4ca180cace6bb3cf36cc21f50140e0.jpg) + +![](images/d582cd8063f1da7edd2263868b583abc7d117eb57850fe23eb4ab8202f485688.jpg) + +![](images/9a2ae177fc9f089aa628ad1a696ffa3e9151915f30ad0fdd39a004e1557011a6.jpg) + +# 6. Ablation Study + +LFE-VAE. We examine the effectiveness of our LFE-VAE by making comparisons with two variants: one using a standard VAE (S-VAE) and another using a fine-tuned VAE (FT-VAE) with single-view event data. As shown in Table 3, S-VAE variant shows performance degradation across all metrics, indicating reduced reconstruction quality and perceptual similarity. While FT-VAE variant shows modest improvements with PSNR of 37.12 dB and SSIM of 0.865, its LPIPS score of 0.151 still indicates suboptimal perceptual quality. In contrast, our LFE-VAE better captures the structural and perceptual characteristics of light field events. + +LFE-adapter. To validate the effectiveness of our LFE-adapter design, we make comparisons with a variant that + +Table 3. Quantitative evaluation of the ablation on the LFE-VAE. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. '↑': the higher the better performance, '↓': the opposite. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
S-VAE36.890.8590.158
FT-VAE37.120.8650.151
LFE-VAE (Ours)38.540.8810.112
+ +Table 4. Quantitative evaluation of the ablation on the LFE-adapter. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
Concat22.540.6310.275
LFE-adapter24.060.7010.239
+ +replaces the LFE-adapter with simple feature concatenation (concat). As shown in Table 4, the concatenation variant shows substantial performance degradation across all metrics, indicating significant reductions in reconstruction accuracy, structural fidelity, and perceptual quality. The comparison clearly validates that our LFE-adapter provides more effective feature integration compared to naive concatenation. + +# 7. Conclusion + +In this paper, we present S2D-LFE, a novel approach for generating consistent, full-view LFE. Our key contributions include developing the first real-world LFE capture system and proposing effective technical components: an LFE-VAE for LFE compression and an LFE-Adapter for prior injection. S2D-LFE achieves simplicity by eliminating multi-view RGB dependencies, controllability through precise view synthesis, and consistency via cross-view and temporal coherence. These advances establish S2D-LFE as an effective solution for dense LFE generation. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001. + +# References + +[1] A N Angelopoulos, J N Martel, A P Kohli, J Conradt, and G Wetzstein. Event based, near eye gaze tracking beyond 10,000 hz. IEEE Transactions on Visualization and Computer Graphics, page 1, 2020. 1, 2 +[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19819-19829, 2022. 2 +[3] Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel, and Tobias Ritschel. X-fields: Implicit neural view-, light- and time-image interpolation. ACM TOG, 39(6):1-15, 2020. 2 +[4] A Bhattacharya, R Madaan, F Cladera, et al. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5846-5855, 2024. 2 +[5] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. pages 497-504. 2001. 2, 5 +[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2 +[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2 +[8] H Chen, J Gu, A Chen, et al. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2 +[9] P Chen, W Guan, F Huang, et al. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 2023. 2 +[10] P Chen, W Guan, and P Lu. Espvio: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8 (6):3661-3668, 2023. 2 +[11] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2 +[12] Zhen Cheng, Yutong Liu, and Zhiwei Xiong. Spatial-angular versatile convolution for light field reconstruction. IEEE Transactions on Computational Imaging, 8:1131-1144, 2022. 2, 6, 7 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4 +[14] A Dosovitskiy, G Ros, F Codevilla, et al. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 3 + +[15] Brandon Yushan Feng and Amitabh Varshney. Signet: Efficient neural representation for light fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14224-14233, 2021. 2 +[16] P Furgale, J Rehder, and R Siegwart. Unified temporal and spatial calibration for multi-sensor systems. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1280-1286. IEEE, 2013. 3 +[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3 +[18] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 6, 7 +[19] M Gehrig, W Aarents, D Gehrig, and D Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2 +[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023, 2 +[21] Mantang Guo, Junhui Hou, Jing Jin, Hui Liu, Huanqiang Zeng, and Jiwen Lu. Content-aware warping for view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9486-9503, 2023. 2, 6, 7, 8 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 +[23] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 1, 2, 6 +[24] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM TOG, 35(6):1-10, 2016. 2, 5 +[25] J Karras, A Holynski, T C Wang, et al. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 2, 4 +[26] B Ke, A Obukhov, S Huang, et al. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492–9502, 2024. 2, 4 +[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[28] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2 + +[29] Zhicheng Lu, Xiaoming Chen, Vera Yuk Ying Chung, Weidong Cai, and Yiran Shen. Ev-lfv: Synthesizing light field event streams from an event camera and multiple rgb cameras. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2, 6 +[30] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2 +[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6 +[32] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6 +[33] C Mou, X Wang, L Xie, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 5 +[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2 +[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4 +[36] R Rombach, A Blattmann, D Lorenz, et al. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4 +[37] S Saxena, C Herrmann, J Hur, et al. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[38] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. NeurIPS, 34:19313-19325, 2021. 2 +[39] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2 +[40] S M N Uddin, S H Ahmed, and Y J Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 1, 2 +[41] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2 + +[42] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In European Conference on Computer Vision, pages 612-629. Springer, 2022. 2, 6, 7, 8 +[43] Yunlong Wang, Fei Liu, Zilei Wang, Guangqi Hou, Zhenan Sun, and Tieniu Tan. End-to-end view synthesis for light field imaging with pseudo 4dcnn. In ECCV, pages 333-348, 2018. 2, 5 +[44] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2022. 2, 6, 7 +[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2 +[47] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 6, 7 +[48] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 1, 7 +[49] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 4 +[50] J Wynn and D Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 2 +[51] J Xiang, J Yang, B Huang, et al. 3d-aware image generation using 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2383-2393, 2023. 2 +[52] Henry Wing Fung Yeung, Junhui Hou, Jie Chen, Yuk Ying Chung, and Xiaoming Chen. Fast light field reconstruction with deep coarse-to-fine modeling of spatial-angular clues. In ECCV, pages 137-152, 2018. 2, 5 +[53] L Zhang, A Rao, and M Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[55] Y Zhou, G Gallego, and S Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37(5):1433-1450, 2021. 2 \ No newline at end of file diff --git a/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/images.zip b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fbca06ee7b4aa36764fb8df84655f5c739b5fc03 --- /dev/null +++ b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10327d9f1dc7f401f21e4a7e8020ec98e54dbcf3e8a3af6408560453f6b934a8 +size 554562 diff --git a/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/layout.json b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae02453e3c7e054e60108fa9b3f10d74ff1ad171 --- /dev/null +++ b/2025/S2D-LFE_ Sparse-to-Dense Light Field Event Generation/layout.json @@ -0,0 +1,7160 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "S2D-LFE: Sparse-to-Dense Light Field Event Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 143, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 143, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 127, + 143, + 479, + 185 + ], + "type": "text", + "content": "Yutong Liu Wenming Weng Yueyi Zhang Zhiwei Xiong* University of Science and Technology of China {ustclyt, wmweng}@mail.ustc.edu.cn {zhyuey, zwxiong}@ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "content": "In this paper, we present S2D-LFE, an innovative approach for sparse-to-dense light field event generation. For the first time to our knowledge, S2D-LFE enables controllable novel view synthesis only from sparse-view light field event (LFE) data, and addresses three critical challenges for the LFE generation task: simplicity, controllability, and consistency. The simplicity aspect eliminates the dependency on frame-based modality, which often suffers from motion blur and low frame-rate limitations. The controllability aspect enables precise view synthesis under sparse LFE conditions with view-related constraints. The consistency aspect ensures both cross-view and temporal coherence in the generated results. To realize S2D-LFE, we develop a novel diffusion-based generation network with two key components. First, we design an LFE-customized variational auto-encoder that effectively compresses and reconstructs LFE by integrating cross-view information. Second, we design an LFE-aware injection adaptor to extract comprehensive geometric and texture priors. Furthermore, we construct a large-scale synthetic LFE dataset containing 162 one-minute sequences using simulator, and capture a real-world testset using our custom-built sparse LFE acquisition system, covering diverse indoor and outdoor scenes. Extensive experiments demonstrate that S2D-LFE successfully generates up to " + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "content": " dense LFE from " + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 56, + 239, + 296, + 574 + ], + "type": "text", + "content": " sparse inputs and outperforms existing methods on both synthetic and real-world data. The datasets and code are available at https://github.com/Yutong2022/S2D-LFE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 578, + 137, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 137, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 137, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 597, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 694 + ], + "type": "text", + "content": "Event cameras capture per-pixel brightness changes asynchronously with microsecond temporal resolution [1], while Light Field (LF) cameras record both the accumulated intensity and direction of light rays, thereby revealing scene geometric structure [48]. The integration of these technologies leads to Light Field Event (LFE) cameras, which capture multi-view event streams arranged in a fixed baseline array. This combination enables simultaneous recording" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 555, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 309 + ], + "type": "text", + "content": "of high-quality spatial depth information and dynamic motion data with exceptional temporal resolution. Such capabilities demonstrate significant potential for various applications, including 6-degree-of-freedom virtual reality [23] and real-time dynamic depth estimation [40]. The rich spatio-temporal information provided by LFE data enables effective capture of structural information in high-speed dynamic scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 529 + ], + "type": "text", + "content": "In this paper, we aim to address a largely unexplored task of LFE generation. To date, the only relevant work is Ev-LFV [29], which combines an RGB LF camera with an event camera aligned to the LF center-view for LFE synthesis. However, this approach faces several critical challenges: (1) Simplicity. Ev-LFV's reliance on multi-view LF frame data as auxiliary information introduces complications. The RGB data is vulnerable to motion blur and exposure issues, potentially compromising LFE generation quality. Moreover, achieving precise spatial-temporal alignment between the LF and event cameras presents substantial technical challenges. (2) Controllability. The fixed view synthesis constraint of Ev-LFV prevents the generation of free viewpoints under sparse LFE conditions with view-related constraints. (3) Consistency. Due to the limitations of low frame-rate RGB data and insufficient structural modeling, Ev-LFV struggles to maintain both angular and temporal continuity in the generated LFE sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 531, + 557, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 557, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 557, + 674 + ], + "type": "text", + "content": "To address these challenges, we propose S2D-LFE for Sparse-to-Dense Light Field Event generation, which is a novel approach for LFE generation. S2D-LFE demonstrates the capability to generate temporally consistent LFE from free viewpoints using only sparse LFE input, making it particularly suitable for real-world applications without the need for frame cameras. Our approach comprises two key components: an LFE-customized variational auto-encoder for efficient LFE compression and reconstruction, and a LFE-aware injection adapter integrated into a diffusion-based generation network that learns to generate the latent representation of LFE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "The first key component, LFE-VAE, is designed to effectively compress high-resolution LFE while preserving spatial and structural information. Specifically, we introduce" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11207" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "content": "two significant improvements over conventional pretrained VAE architectures used in diffusion models [25, 26, 35, 36, 53]. First, we incorporate angular interaction blocks into the VAE decoder to fully exploit cross-view information within LFE. Second, we implement a view-enhancement mechanism that utilizes input sparse LFE views to improve the quality of reconstructed novel views. These enhancements effectively leverage the inherent cross-view information in LFE and mitigate compression artifacts, resulting in high-fidelity reconstruction. The second key component, LFE-adapter, is designed to utilize geometric and texture priors from the input sparse LFE to guide the diffusion process. Specifically, LFE-adapter generates pyramid condition signals that provide targeted conditioning at multiple resolution scales for the U-Net during diffusion. This multi-scale conditioning enables more precise injection of geometric and texture priors, thereby achieving superior cross-view consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 294, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 294, + 395 + ], + "type": "text", + "content": "Moreover, we develop the first sparse LFE capture system to facilitate the acquisition of continuous light field event data. The system consists of four Davis346 cameras arranged in a " + }, + { + "bbox": [ + 55, + 288, + 294, + 395 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 288, + 294, + 395 + ], + "type": "text", + "content": " array configuration. Using this setup, we collect a diverse dataset comprising 25 sequences, including 15 indoor scenes and 10 autonomous driving scenarios. To support comprehensive model training, we additionally construct a large-scale synthetic dataset using Carla simulator [14], generating 162 one-minute dense LFE sequences." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 294, + 418 + ], + "type": "text", + "content": "The main contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 295, + 586 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 419, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 455 + ], + "type": "text", + "content": "- We propose S2D-LFE, a novel approach for generating free-viewpoint and temporally consistent light field events from sparse-view LFE input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 502 + ], + "type": "text", + "content": "- Our S2D-LFE incorporates two key components: an LFE-VAE for efficient LFE compression, and an LFE-adapter integrated into a diffusion-based generation network that reconstructs the latent representation of LFE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "type": "text", + "content": "- We develop the first sparse LFE capture system and collect a dataset of 25 real-world sparse LFE sequences. Additionally, we synthesize 162 dense LFE sequences using the Carla simulator to facilitate comprehensive training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 586 + ], + "type": "text", + "content": "- Our approach outperforms existing methods on both synthetic and real-world datasets, while presenting enhanced robustness in generalizing to free-viewpoint synthesis." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Multi-view Event Processing. While event cameras offer exceptional temporal resolution and low latency [1], monocular event data alone is insufficient for capturing comprehensive structural information of dynamic scenes. Various approaches have been proposed to address this limitation: 1) NeRF-based scene representation learning [4], 2) hybrid systems combining events with conventional cameras [9, 29], and 3) stereo event systems [10, 19, 40, 55]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "NeRF-based methods, while promising, require per-scene training and multiple viewpoints for effective scene representation. For example, EvDNeRF [4] introduces an event-based dynamic NeRF pipeline for reconstructing event streams in rapidly deforming scenes, but its performance degrades significantly with limited viewpoints " + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "inline_equation", + "content": "(< 8)" + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": ". Hybrid systems face challenges in reconciling the temporal resolution disparity between different modalities. For instance, Lu et al. [29] encounter practical limitations in real-world applications due to the substantial difference in temporal resolution between LF and event cameras. While stereo event systems can capture geometric information, the resulting geometry tends to be sparse. In contrast, the Light Field event modality offers distinct advantages by capturing dense structural information of dynamic scenes without requiring per-scene training, while enabling multiple angular resolution generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 285, + 553, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 548 + ], + "type": "text", + "content": "Light Field View Synthesis. The reconstruction of dense Light Fields from sparse views, commonly known as LF angular super-resolution, remains an active research area. Existing approaches can be broadly categorized into two groups: depth-dependent [5, 24, 43, 52] and depth-independent methods [12, 21, 23, 44]. Depth-dependent methods rely on precise disparity estimation but are vulnerable to occlusions, while depth-independent methods often struggle with generalization across varying disparity ranges. Recent advances in Neural Light Field representations [2, 3, 15, 28, 38, 42] have brought significant improvements to LF view synthesis by addressing limitations of conventional angular super-resolution methods. Compared with the widely adopted learnable 3D representation methods [7, 11, 34, 41], these approaches can more effectively exploit the spatial priors inherent in LF. However, these NeRF-based approaches are constrained by their requirement for numerous input views and scene-specific training. In contrast, our proposed S2D-LFE conquers these limitations by enabling free-viewpoint synthesis from sparse inputs without the need for per-scene training, while maintaining high recovery quality and temporal consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "type": "text", + "content": "Diffusion-based Image Generation. Diffusion models have shown remarkable progress in novel view synthesis [6, 8, 20, 39, 46, 50, 51]. For example, NeRDi [20] leverages language priors for multi-view synthesis, effectively connecting image semantics to appearance reconstruction. These advances demonstrate the robust capability of diffusion models in modeling multi-view distributions. Furthermore, diffusion models have achieved impressive results across diverse data modalities, including depth [26], point clouds [30], optical flow [26, 37], and human pose [25]. Motivated by this versatility and strong fitting capability across different modalities, we adopt diffusion models for dense LFE generation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11208" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 273, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 273, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 273, + 85 + ], + "type": "text", + "content": "3. Real-world and Synthetic LFE Datasets" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 291, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 291, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 291, + 105 + ], + "type": "text", + "content": "3.1. LFE Capture System and Real-world Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "text", + "content": "System Setup. To facilitate the collection of real-world LFE data and validate our approach, we develop a custom LFE capture system. Given the lack of commercially available LFE cameras, our system design prioritizes precise spatial alignment and temporal synchronization while maintaining cost-effectiveness. The system, shown in Fig. 1, consists of a " + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "text", + "content": " array of Davis346 cameras mounted on a square aluminum alloy frame, with uniform horizontal and vertical baseline lengths of " + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "inline_equation", + "content": "6\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "text", + "content": ". To ensure consistent imaging characteristics across views, all four cameras are equipped with identical lenses and configured with matching parameters, including focal length and exposure time. For precise temporal synchronization, we implement a master-slave configuration using external cables (depicted as black lines in Fig.1). The upper-left DAVIS346 serves as the master camera, transmitting trigger signal pulses to the three slave cameras through these cables. This system captures event data at a spatial resolution of " + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "inline_equation", + "content": "346 \\times 260" + }, + { + "bbox": [ + 55, + 111, + 295, + 363 + ], + "type": "text", + "content": " pixels, which can be readily upgraded to a higher resolution by updating event cameras. Detailed parameter configurations are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 639 + ], + "type": "text", + "content": "Camera Calibration. Precise camera calibration is crucial for accurate Light Field Event capture, as it ensures geometric consistency across multiple views and enables reliable reconstruction of scene structure. This is particularly important for our system where accurate spatial relationships between cameras directly impact the quality of captured LFE data. We utilize the Kalibr toolbox [16] to estimate camera intrinsics and extrinsics by capturing fixed-rate frames of a checkerboard pattern. Since the DAVIS346 cameras output both frames and event streams through the same optical system, the calibration parameters obtained from frames are directly applicable to event streams. Following calibration, we perform systematic rectification of the captured sparse LFE. Using the upper left camera as reference, we first apply perspective transformation to project the upper right camera onto a virtual plane parallel to the reference camera's horizontal plane, establishing the baseline length between them. We then align the two lower cameras through perspective transformations to their corresponding positions in the sparse LF, ensuring uniform baseline lengths across the array. This comprehensive calibration and rectification process ensures accurate structural information capture of dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Real-world LFE Dataset. Following system calibration and rectification, we collect a real LFE dataset comprising 25 dynamic scenes using our LFE capture system, which is termed \"LFE-real\". The dataset consists of two main categories: 15 indoor sequences captured by moving our camera array through various indoor environments (featuring sub" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 545, + 205 + ], + "type": "image", + "image_path": "9ab701736589bc8e08a689fcfe9bb5fe4453fbf480f2cb1727bc7001c85226cf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 208, + 518, + 219 + ], + "lines": [ + { + "bbox": [ + 348, + 208, + 518, + 219 + ], + "spans": [ + { + "bbox": [ + 348, + 208, + 518, + 219 + ], + "type": "text", + "content": "Figure 1. The proposed LFE capturing system." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 235, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 307 + ], + "type": "text", + "content": "jects such as people, plants, and interior layouts), and 10 autonomous driving sequences recorded from displayed autonomous driving footage. Each sequence spans between 3 to 5 minutes, ensuring sufficient temporal coverage of the dynamic events. The details about scenes are in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 419, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 419, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 419, + 327 + ], + "type": "text", + "content": "3.2. Synthetic Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "content": "Given that physical capture devices can only provide sparse view inputs without dense multi-view supervision, we leverage the Carla simulator [14] to generate a large synthetic LFE dataset for training and evaluation, which is termed \"LFE-syn\". We utilize 18 different maps provided by the Carla environment for data simulation. Each map is divided into 9 regions, with a designated starting point in each region. At these starting points, we deploy virtual vehicles equipped with " + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "content": " event camera arrays. The baseline lengths between adjacent cameras are randomly set between " + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "inline_equation", + "content": "3\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "inline_equation", + "content": "9\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 332, + 555, + 548 + ], + "type": "text", + "content": ". The event trigger thresholds are configured asymmetrically: 0.1 to 0.15 for positive events and -0.15 to -0.1 for negative events. Using autonomous navigation, each vehicle generates one-minute LFE sequences, resulting in a total of 162 synthetic sequences. Additionally, to facilitate the evaluation, our synthetic dataset share the same resolution with the real-world dataset. Detailed parameter configurations are provided in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 559, + 487, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 487, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 487, + 572 + ], + "type": "text", + "content": "4. S2D-LFE Generation Network" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 579, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 651 + ], + "type": "text", + "content": "Our custom-built system captures sparse LFE with coarse geometric information of dynamic scenes. To enhance the structural representation capabilities, we propose S2D-LFE, a novel approach for synthesizing accurate dense LFEs from sparse inputs. We detail our approach in the following sections." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 659, + 382, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 382, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 382, + 672 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In our setting, each event stream view in the LFE is accumulated as an Event Count Map (ECM) [17] with shapes of " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{B\\times TB\\times 2\\times H\\times W}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " represents the batch size," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11209" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 68, + 515, + 207 + ], + "blocks": [ + { + "bbox": [ + 92, + 68, + 515, + 207 + ], + "lines": [ + { + "bbox": [ + 92, + 68, + 515, + 207 + ], + "spans": [ + { + "bbox": [ + 92, + 68, + 515, + 207 + ], + "type": "image", + "image_path": "768a75bddd89b0bb7371aee53791228bcb5956d7aa419df5233b3818c1697558.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. The second training stage of the LFE-VAE architecture. The input to the LFE-VAE supports multiple angular resolutions of LFE. Here, we take the case of an input with 25 views as an example. The LFE-VAE is composed of the encoder " + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "inline_equation", + "content": "V_{enc}" + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "text", + "content": ", the decoder " + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\hat{V}_{dec}" + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "text", + "content": " inserted by angular interaction blocks, and the following angular resblocks. The training at this stage is aimed at updating the decoder " + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\hat{V}_{dec}" + }, + { + "bbox": [ + 54, + 207, + 555, + 251 + ], + "type": "text", + "content": " and the angular resblocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "TB" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": " represents the temporal bins, 2 denotes the polarities, and " + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": " denotes the spatial resolution. To facilitate the subsequent processing of the VAE, we transform the two-polarity channels of the event stream into three channels based on [49]. A LFE can be represented as " + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "L\\in \\mathbb{R}^{B\\times 3\\times M\\times N\\times H\\times W}" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "M\\times N" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": " denotes the angular resolution. The LFE synthesis task aims to generate a dense LFE " + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "L_{D}\\in \\mathbb{R}^{B\\times 3\\times M\\times N\\times H\\times W}" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": " from a sparse LFE " + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "L_{S}\\in \\mathbb{R}^{B\\times 3\\times m\\times n\\times H\\times W}" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "inline_equation", + "content": "m\\times n" + }, + { + "bbox": [ + 56, + 258, + 296, + 547 + ], + "type": "text", + "content": " represents a lower angular resolution. To effectively model the complex distribution of LFE data, we adopt a diffusion-based pipeline. Given the high-dimensional nature of LFE data, we first introduce LFE-VAE to encode LFE into a compact latent representation. The encoded sparse LFE, together with the encoded angular coordinates, serve as conditioning signals for the diffusion model to synthesize views at specified angular positions. To effectively incorporate these conditioning signals into the diffusion backbone, we design an LFE-adapter for optimal signal processing. Through iterative generation of all unknown views, we achieve the transformation from sparse to dense LFE. Notably, our diffusion network demonstrates the capability to synthesize novel views at arbitrary angles beyond the training set by accepting different angular coordinates as input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 552, + 126, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 126, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 126, + 564 + ], + "type": "text", + "content": "4.2. LFE-VAE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 688 + ], + "type": "text", + "content": "To handle the high spatial resolution of LFE data efficiently, we employ Variational Autoencoder (VAE) architectures commonly used in diffusion models [25, 26, 35, 36, 53] for data compression. However, directly applying conventional VAEs to LFE data presents two significant challenges: 1) Traditional VAEs used in stable diffusion are designed for single-view processing, failing to leverage the inherent cross-view structural information in LFE data, and 2) The lossy nature of VAE compression inevitably results in information loss in the event streams." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To address these limitations, we propose LFE-VAE, building upon the pretrained VAE model from SDXL [35]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 258, + 555, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 317 + ], + "type": "text", + "content": "Our architecture introduces two key innovations: (1) angular interaction blocks in the decoder to incorporate cross-view complementary information, and (2) a view-enhancement mechanism utilizing four known views to minimize compression artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "type": "text", + "content": "In specific, for each dense LFE input " + }, + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "type": "inline_equation", + "content": "L_{in}" + }, + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "type": "text", + "content": ", the encoder first generates a latent code " + }, + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 313, + 319, + 556, + 437 + ], + "type": "text", + "content": ". This latent representation is processed by our enhanced decoder, which embeds cross-view geometric information through angular interaction blocks. Each block consists of two cascaded layers of spatial and angular convolutions. The corresponding positions of the decoded target views are replaced with four known sparse LFE views, and then the LFE is refined through several cascaded resblocks defined on the LFE angular slices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": "Our training process follows a two-stage approach. First, we fine-tune the stable diffusion VAE using single-view events to obtain updated weights for the encoder " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "V_{enc}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "V_{dec}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": ". Subsequently, while keeping " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "V_{enc}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " frozen, we fine-tune the proposed decoder " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\hat{V}_{dec}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " and angular resblocks. The second stage of the LFE-VAE architecture is illustrated in detail in Fig. 2. Specifically, we first select the sparse LFE views " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "L_{S}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": ", which represent the condition signal for the subsequent diffusion pipeline. The input LFE " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "L_{in}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " is then encoded by " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "V_{enc}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " and decoded by the updated " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\hat{V}_{dec}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": ", resulting in an intermediate reconstructed LFE. Next, we replace the four corner views of the recovered LFE, with the four corresponding views in " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "L_{S}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " selected in " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "L_{in}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": ". Finally, the compression degradation is restored through the utilization of angular resblocks. After training, the encoder " + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "inline_equation", + "content": "V_{enc}" + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": " is integrated into the diffusion training pipeline to compress input sparse LFE, enabling the subsequent diffusion network to handle higher-resolution LFE generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 422, + 673 + ], + "type": "text", + "content": "4.3. Diffusion Pipeline" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Following mainstream diffusion architectures [13, 22, 25, 26, 35, 36, 53], our diffusion backbone adopts a UNet structure. While existing diffusion models primarily han-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11210" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 69, + 339, + 273 + ], + "blocks": [ + { + "bbox": [ + 80, + 69, + 339, + 273 + ], + "lines": [ + { + "bbox": [ + 80, + 69, + 339, + 273 + ], + "spans": [ + { + "bbox": [ + 80, + 69, + 339, + 273 + ], + "type": "image", + "image_path": "42608d48f053ec2b126a00dfc3f2f39a848f22e45896ac8fa657580dbd757254.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "type": "text", + "content": "Figure 3. The diffusion pipeline of S2D-LFE at the t-th time step. The pipeline primarily consists of two parts: (1) The left part, highlighted with a blue background, represents the main pipeline diagram. The diffusion pipeline of S2D-LFE adopts a UNet backbone, where the latent features " + }, + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "type": "text", + "content": " from four viewpoints are injected into the network as conditional signals via the LFE-adapter. In addition to the time embedding, our prompt incorporates positional encodings of both horizontal and vertical coordinates. (2) The right section, highlighted with an orange background, depicts the LFE-adapter block within the LFE-adapter. After cascading the condition information through the ContBlock and GeoBlock, the output features are embedded into the corresponding layers of the UNet. The LFE-adapter consists of four LFE-adapter blocks, each producing a different conditional signal " + }, + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\hat{Z}_{\\mathrm{s1}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s2}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s3}}^{\\downarrow \\downarrow}, \\hat{Z}_{\\mathrm{s4}}^{\\downarrow \\downarrow}" + }, + { + "bbox": [ + 54, + 277, + 555, + 357 + ], + "type": "text", + "content": " corresponding to the respective layers of the UNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 339, + 70, + 529, + 273 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 529, + 273 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 529, + 273 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 529, + 273 + ], + "type": "image", + "image_path": "5deff887cfac4342786502fd0dc7242c93cf4c8a19fcad00ba2fcc41352313db.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 365, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 579 + ], + "type": "text", + "content": "dle unstructured conditioning data such as images, videos, and text, the potential of structured conditioning data, which contains rich structural and texture priors, remains largely unexplored. To address this gap, we propose a specialized diffusion pipeline tailored for LFE data. At its core is the novel LFE-adapter, which serves two key functions: 1) extracting features that encode both structural and texture information from input sparse LFE and 2) injecting these extracted features into the diffusion backbone. With the structural and texture information effectively embedded, the diffusion pipeline generates specific LFE views based on input view prompts. Our approach employs continuous, linear angular coordinate encoding for view prompts, enabling the pipeline to generalize to sampling LFE views across varying densities and provide flexibility in LFE generation. The complete architecture of our diffusion pipeline at the t-th time step is illustrated in Fig. 3. Then, we will detail each of these innovations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "LFE Angular Prompt. Accurate positional priors of target views are essential for novel LF view generation, which we achieve through angular coordinate prompting. As illustrated in Fig. 3, we process 2D coordinates through a two-stage encoding scheme. First, we independently encode row and column coordinates. These encoded coordinates are then transformed through cascaded fully connected layers to obtain the raw and column embeddings, which are then concatenated and fed into a linear layer to get the angular embedding " + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "inline_equation", + "content": "F_{ang}" + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": ". It serves dual purposes in our pipeline. It is combined with time embedding as input to each UNet layer," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 401 + ], + "type": "text", + "content": "while simultaneously providing angular positional priors to the LFE-adapter for extracting view-specific conditioning information for the target view generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": "LFE-adapter. Our pipeline leverages priors from a sparse LFE containing four known views. The design incorporates both texture and geometry priors from these input views to provide specific conditioning signals for novel view generation. Drawing inspiration from [33], we propose an LFE-adapter to effectively inject structural and textural information from the known views. The architecture of our LFE-adapter is depicted in the orange region of Fig. 3. The processing pipeline begins by using the " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": " of LFE-VAE to obtain compressed latent codes " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": " from the sparse LFE. We then integrate the angle embedding " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ang}}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": " with the latent features of the four input views to obtain the updated latent feature " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "Z_{s}^{*}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": ". This feature is processed through a content extraction block (contblock) that extracts textures and details using spatial and angular convolutions. The spatial convolution embeds individual view information, while the angular convolution captures complementary cross-view information. A geometric extraction block (geoblock) further processes the features using cascaded Epipolar Plane Image (EPI) convolutions [5, 24, 43, 52] to embed geometric information, and thus we obtain the updated feature " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\hat{Z}_{s}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": ". This feature undergoes spatial downsampling to produce " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\hat{Z}_{s}^{\\downarrow}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": " for subsequent network updates, followed by angular down-sampling to generate " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\hat{Z}_{s}^{\\downarrow \\downarrow}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": " as conditioning signals for corresponding UNet layers. The LFE-adapter ultimately generates a set of condition signals " + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\hat{Z}_{s1}^{\\downarrow \\downarrow}, \\hat{Z}_{s2}^{\\downarrow \\downarrow}, \\hat{Z}_{s3}^{\\downarrow \\downarrow}, \\hat{Z}_{s4}^{\\downarrow \\downarrow}" + }, + { + "bbox": [ + 313, + 403, + 556, + 716 + ], + "type": "text", + "content": ". Each sig" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11211" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 105, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of event-based view synthesis on the in-training-scale setting. We tested various methods under synthetic and real settings and evaluated their performance across three metrics. The best results are highlighted in red, the second-best results are highlighted in blue, and the third-best results are highlighted in green. ‘↑’: the higher the better performance, ‘↓’: the opposite." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 105, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 105, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 105, + 553, + 228 + ], + "type": "table", + "html": "
CategoryMethodPSNR ↑Synthetic SSIM ↑LPIPS ↓NIQE ↓Real-world BRISQUE ↓MUSIQ ↑
AngularSRJing et al. [23]21.120.5800.36231.5667.3228.14
DistgASR [44]23.320.6760.27828.8863.9529.90
SAV [12]22.840.6140.28229.1562.7729.25
Guo et al. [21]23.460.6870.29028.2359.1330.72
HybridET-Net [47]+SAV [12]+Vid2E [18]20.230.5740.37332.1167.1125.45
ET-Net [47]+Guo et al.+Vid2E [18]20.880.5890.37531.9566.5825.59
NeRF-basedR2L [42]19.880.5630.40128.4460.2530.45
GenerationOurs24.060.7010.23927.6553.4531.14
", + "image_path": "5638f0904acb1d8906a8f1a6eae9b68345c14c597a9cf08868ff01f4eda1a1e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 237, + 295, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 295, + 260 + ], + "type": "text", + "content": "nal is added to the corresponding same-size input features of the UNet blocks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "content": "LFE View Sampling. Through our LFE-adapter, the diffusion pipeline is enriched with comprehensive structural and textural information from the input LFE. The angular prompts effectively encode target view positions, ensuring angular continuity in the generated LFE views. The pipeline samples individual views sequentially, reconstructing the complete continuous LFE by iteratively processing coordinates for all target views. The framework accommodates LFEs of varying densities through relative coordinate mapping, where " + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "content": " represents the top-left view and " + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "inline_equation", + "content": "(1,1)" + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "content": " represents the bottom-right view. For example, generating a view at position " + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "content": " in an " + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "content": " LFE would use normalized coordinates " + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "inline_equation", + "content": "(i / (N - 1),j / (N - 1))" + }, + { + "bbox": [ + 54, + 262, + 295, + 454 + ], + "type": "text", + "content": ". This normalized coordinate system provides a unified approach for sampling views across different LFE densities while maintaining consistent spatial relationships." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 132, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 132, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 132, + 482 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 490, + 174, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 174, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 174, + 502 + ], + "type": "text", + "content": "5.1. Experiment Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": "Implementation Details. Our networks are trained on a computing cluster equipped with eight NVIDIA A100 GPUs. Following [44], we employ data augmentation techniques including random horizontal flipping, vertical flipping, and " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " rotation to enhance model robustness. The optimization process utilizes the Adam optimizer with hyperparameters " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": ", processing batches of size 32. We adopt a learning rate scheduling strategy where the initial rate of " + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": " is halved every 400 epochs. Training concludes after 1500 epochs when convergence is achieved. To ensure fairness in the experiments, all metrics are calculated excluding the input views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Datasets and Metrics. For experimental validation, we utilize two distinct datasets: a synthetic dataset generated using Carla for training and testing, and a real-world dataset captured by our system for additional testing. We use two sets of complementary metrics. For data with ground-truth" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 237, + 555, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 273 + ], + "type": "text", + "content": "references, we utilize PSNR measured in dB, SSIM [45], and LPIPS [54]. For data without ground-truth references, we employ NIQE [32], BRISQUE [31], and MUSIQ [27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 274, + 556, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 556, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 556, + 430 + ], + "type": "text", + "content": "Comparison Methods. We establish comparisons with three categories of related approaches1. The first category comprises LF angular super-resolution methods, including Jing et al.[23], DistgASR[44], SAV [12], and Guo et al.[21]. The second category involves a three-stage approach: we first convert sparse LFE to sparse RGB LF using event-based video reconstruction methods like ET-Net[47], then perform angular super-resolution to get dense RGB LF, and finally convert the dense RGB LF back to dense LFE using Vid2E [18]. The third category explores neural radiance field-based reconstruction using R2L [42]. All the aforementioned baseline methods are re-trained on our dataset loading their original pre-trained parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 432, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 432, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 432, + 451 + ], + "type": "text", + "content": "5.2. Comparison Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": "In-training-scale Comparisons. We first conduct in-distribution evaluation on " + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 456, + 556, + 685 + ], + "type": "text", + "content": " LFE generation, which matches our training configuration. As shown in Table 1, our S2D-LFE outperforms all baselines across all metrics on both synthetic and real-world testsets. On the synthetic testset, S2D-LFE achieves a notable PSNR improvement of 0.60 dB over the second-best method Guo et al.[21], demonstrating superior reconstruction accuracy. The strong LPIPS performance further indicates enhanced perceptual quality, leveraging the advantages of generative modeling. This is visually evident in Fig.4, where S2D-LFE reconstructs building details with minimal artifacts and better ground truth alignment. In addition, although R2L [42] achieves decent visual results in Fig.4, it fails to reconstruct in certain scenes, leading to lower metric performance on the synthetic dataset. Meanwhile, S2D-LFE maintains its superior performance across all metrics on the real-world testset, demonstrating effective generalization to real-world scenarios. This quantitative advantage is corroborated by" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "1Since EV-LFV [29] and our S2D-LFE utilize different input modalities, a direct comparison between the two methods is not feasible." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11212" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 547, + 261 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 547, + 261 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 547, + 261 + ], + "type": "image", + "image_path": "5785cd29695ed92bf8fc992bc554d3e67472ab68881926467ca60cecfb7ceef7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 265, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 555, + 319 + ], + "type": "text", + "content": "Figure 4. The figure presents a qualitative comparison on the central view of the generated LFE, using various methods for both synthetic (top row) and real-world (bottom row) testsets. The proposed S2D-LFE method (labeled as \"Ours\") is compared against other existing techniques, including SAV [12], Guo et al. [21], ET-Net [47] + Guo et al. + Vid2E [18], R2L [42], DistgASR [44], and the ground-truth. The highlighted regions (yellow and green boxes) magnify specific areas to emphasize the differences in event generation quality, with S2D-LFE exhibiting improved fidelity and preservation of fine details in both synthetic and real-world scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "content": "the qualitative results shown in Fig. 4, where S2D-LFE produces higher quality reconstructions with better preservation of details. Due to incorrect estimation of view position, SAV [12] and DistgASR [44] exhibit misaligned results compared to other methods. Furthermore, the ET-Net " + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "inline_equation", + "content": "[47] + \\mathrm{SAV} + \\mathrm{Vid2E}" + }, + { + "bbox": [ + 55, + 338, + 296, + 422 + ], + "type": "text", + "content": " [18] approach exhibits severe view aliasing in this scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 424, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 616 + ], + "type": "text", + "content": "Out-of-training-scale Comparisons. We extend our evaluation to test the model's generalization capability beyond its training distribution of " + }, + { + "bbox": [ + 55, + 424, + 296, + 616 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 55, + 424, + 296, + 616 + ], + "type": "text", + "content": " LFE, examining its performance on denser LFEs. Among the seven baselines we selected, only R2L [42] and Guo et al. [21] are capable of handling out-of-training-scale LFE generation. Therefore, we only compare our method with these two approaches in this section. As shown in Table 2, our S2D-LFE consistently outperforms baseline methods across all tested LFE densities. While all methods show performance degradation as LFE density increases, S2D-LFE demonstrates greater resilience with a slower degradation rate. The qualitative results in Fig.5 further validate S2D-LFE's superior performance in maintaining perceptual quality and detail preservation for out-of-distribution scenarios. Additional visual comparisons are provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Angular Consistency. In this section, we analyze and compare the angular continuity of the LFE generated by our S2D-LFE. Given that the angular continuity of LFE data is strongly correlated with the gradient of the Epipolar Plane Image (EPI) [44, 48], we compute metrics (PSNR, SSIM) on the EPI gradient to evaluate the angular continuity of Guo et al. [21], R2L [42], and S2D-LFE when generating LFEs with different numbers of views (ranging from " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "5 \\times 5" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 335, + 555, + 504 + ], + "blocks": [ + { + "bbox": [ + 317, + 335, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 317, + 335, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 555, + 504 + ], + "type": "image", + "image_path": "ac77e20a8a93f083f1158f6c4133177e879179a75866efbb6e47cc28cb31d062.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "type": "text", + "content": "Figure 5. The figure presents the out-of-training-scale qualitative comparison using different methods including Guo et al. [21], R2L [42], and our proposed S2D-LFE (Ours). Besides, synthetic-scene (1) is the (1,1) view from the generated " + }, + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "type": "text", + "content": " LFE, and realworld-scene (2) is the (7,7) view from the generated " + }, + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 507, + 555, + 584 + ], + "type": "text", + "content": " LFE. The yellow and green boxes highlight specific regions to better illustrate the differences in event generation quality." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "type": "text", + "content": "). It can be observed from Fig. 6 that as the number of views increases, all three methods exhibit a decreasing trend in both PSNR and SSIM metrics while the decline for our S2D-LFE is more gradual. It can be interpreted as S2D-LFE exhibits better angular consistency and robustness compared to the other methods. Moreover, we exhibit a qualitative comparison when generating " + }, + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 313, + 597, + 556, + 717 + ], + "type": "text", + "content": " LFE in Fig. 7. It can be observed that S2D-LFE preserves fidelity details while producing EPI slices that more closely align with the ground truth." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11213" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 105, + 555, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of event-based view synthesis on the out-of-training-scale setting. We tested various methods under three different settings (generating 36 views, 49 views, and 64 views) and evaluated their performance across three metrics. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 105, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 105, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 105, + 555, + 190 + ], + "type": "table", + "html": "
MethodNIQE↓BRISQUE↓MUSIQ↑
36-view49-views64-views36-views49-views64-views36-views49-views64-views
SynGuo et al. [21]16.8518.6220.1059.4564.1270.7831.2830.0228.31
R2L [42]16.9619.0521.5461.7265.9371.5630.2729.0827.29
Ours15.8417.6519.6754.1558.2163.4932.3031.8930.34
RealGuo et al. [21]28.8129.5830.2962.8366.1570.5830.2729.7829.30
R2L [42]28.9429.7230.6164.0068.1871.3729.9929.4829.02
Ours27.9928.4729.0156.1758.9161.2930.9330.6530.21
", + "image_path": "5dc7c61df9a89b7591cfc11925c1ec68e0f1fb2f3519d0c3f1f05bf6c63bdb3e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 203, + 266, + 319 + ], + "blocks": [ + { + "bbox": [ + 72, + 203, + 266, + 319 + ], + "lines": [ + { + "bbox": [ + 72, + 203, + 266, + 319 + ], + "spans": [ + { + "bbox": [ + 72, + 203, + 266, + 319 + ], + "type": "image", + "image_path": "0b5a34f7c996ad3d86ec5977f4acf01db5be8fb186671be013d2b6eae0e87ee2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "type": "text", + "content": "Figure 6. The figure presents a quantitative comparison of angular consistency across different methods, including Guo et al., R2L, and our proposed S2D-LFE. Subfigures (a) and (b) depict PSNR (dB) and SSIM metrics respectively, when generating different numbers of views ranging from " + }, + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 55, + 319, + 296, + 396 + ], + "type": "text", + "content": ". All methods exhibit a decline in both metrics as the number of views increases, but S2D-LFE exhibits a relatively slower decline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 403, + 115, + 478 + ], + "blocks": [ + { + "bbox": [ + 56, + 403, + 115, + 478 + ], + "lines": [ + { + "bbox": [ + 56, + 403, + 115, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 403, + 115, + 478 + ], + "type": "image", + "image_path": "cd465763e4f846f4ddbce5eb4de5cb969309fa04060ec249da2121ad107df416.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": "Figure 7. The figure presents a qualitative comparison when generating " + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": " LFE using Guo et al., R2L, and our proposed S2D-LFE. The green boxes highlight spatial slices of view (1,1), while the yellow boxes highlight EPI slices." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 116, + 403, + 160, + 476 + ], + "blocks": [ + { + "bbox": [ + 116, + 403, + 160, + 476 + ], + "lines": [ + { + "bbox": [ + 116, + 403, + 160, + 476 + ], + "spans": [ + { + "bbox": [ + 116, + 403, + 160, + 476 + ], + "type": "image", + "image_path": "f4e4c381ae88f3504a8219a51a2dbaeb7b3b729709f900c8f571a74e35b41423.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 160, + 403, + 203, + 475 + ], + "blocks": [ + { + "bbox": [ + 160, + 403, + 203, + 475 + ], + "lines": [ + { + "bbox": [ + 160, + 403, + 203, + 475 + ], + "spans": [ + { + "bbox": [ + 160, + 403, + 203, + 475 + ], + "type": "image", + "image_path": "776ed85c92cb1415e633d2b19013c5cb3e4ca180cace6bb3cf36cc21f50140e0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 203, + 403, + 247, + 474 + ], + "blocks": [ + { + "bbox": [ + 203, + 403, + 247, + 474 + ], + "lines": [ + { + "bbox": [ + 203, + 403, + 247, + 474 + ], + "spans": [ + { + "bbox": [ + 203, + 403, + 247, + 474 + ], + "type": "image", + "image_path": "d582cd8063f1da7edd2263868b583abc7d117eb57850fe23eb4ab8202f485688.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 247, + 403, + 291, + 474 + ], + "blocks": [ + { + "bbox": [ + 247, + 403, + 291, + 474 + ], + "lines": [ + { + "bbox": [ + 247, + 403, + 291, + 474 + ], + "spans": [ + { + "bbox": [ + 247, + 403, + 291, + 474 + ], + "type": "image", + "image_path": "9a2ae177fc9f089aa628ad1a696ffa3e9151915f30ad0fdd39a004e1557011a6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 535, + 148, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 148, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 148, + 548 + ], + "type": "text", + "content": "6. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 556, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 687 + ], + "type": "text", + "content": "LFE-VAE. We examine the effectiveness of our LFE-VAE by making comparisons with two variants: one using a standard VAE (S-VAE) and another using a fine-tuned VAE (FT-VAE) with single-view event data. As shown in Table 3, S-VAE variant shows performance degradation across all metrics, indicating reduced reconstruction quality and perceptual similarity. While FT-VAE variant shows modest improvements with PSNR of 37.12 dB and SSIM of 0.865, its LPIPS score of 0.151 still indicates suboptimal perceptual quality. In contrast, our LFE-VAE better captures the structural and perceptual characteristics of light field events." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "LFE-adapter. To validate the effectiveness of our LFE-adapter design, we make comparisons with a variant that" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 247, + 554, + 293 + ], + "blocks": [ + { + "bbox": [ + 313, + 201, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 245 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of the ablation on the LFE-VAE. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. '↑': the higher the better performance, '↓': the opposite." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 247, + 554, + 293 + ], + "lines": [ + { + "bbox": [ + 317, + 247, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 247, + 554, + 293 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
S-VAE36.890.8590.158
FT-VAE37.120.8650.151
LFE-VAE (Ours)38.540.8810.112
", + "image_path": "142ee98062b26eeccf434dfc16a0c1651a6993ee76530b1b9aadadc7367a30dc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 360, + 554, + 401 + ], + "blocks": [ + { + "bbox": [ + 313, + 298, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 354 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation of the ablation on the LFE-adapter. The table compares the performance under a setting of generating 25 views in the synthetic dataset. The best results are highlighted in bold. ‘↑’: the higher the better performance, ‘↓’: the opposite." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 360, + 554, + 401 + ], + "lines": [ + { + "bbox": [ + 317, + 360, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 554, + 401 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
Concat22.540.6310.275
LFE-adapter24.060.7010.239
", + "image_path": "5cac460d3c2d066d7b655a56976c1f32d381131a1697ef187838bddd7716936d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 408, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 503 + ], + "type": "text", + "content": "replaces the LFE-adapter with simple feature concatenation (concat). As shown in Table 4, the concatenation variant shows substantial performance degradation across all metrics, indicating significant reductions in reconstruction accuracy, structural fidelity, and perceptual quality. The comparison clearly validates that our LFE-adapter provides more effective feature integration compared to naive concatenation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 512, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 512, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 512, + 388, + 523 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 529, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 649 + ], + "type": "text", + "content": "In this paper, we present S2D-LFE, a novel approach for generating consistent, full-view LFE. Our key contributions include developing the first real-world LFE capture system and proposing effective technical components: an LFE-VAE for LFE compression and an LFE-Adapter for prior injection. S2D-LFE achieves simplicity by eliminating multi-view RGB dependencies, controllability through precise view synthesis, and consistency via cross-view and temporal coherence. These advances establish S2D-LFE as an effective solution for dense LFE generation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 657, + 411, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 411, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 411, + 670 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 673, + 554, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 554, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 554, + 709 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11214" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] A N Angelopoulos, J N Martel, A P Kohli, J Conradt, and G Wetzstein. Event based, near eye gaze tracking beyond 10,000 hz. IEEE Transactions on Visualization and Computer Graphics, page 1, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Benjamin Attal, Jia-Bin Huang, Michael Zollhöfer, Johannes Kopf, and Changil Kim. Learning neural light fields with ray-space embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19819-19829, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 235 + ], + "type": "text", + "content": "[3] Mojtaba Bemana, Karol Myszkowski, Hans-Peter Seidel, and Tobias Ritschel. X-fields: Implicit neural view-, light- and time-image interpolation. ACM TOG, 39(6):1-15, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 282 + ], + "type": "text", + "content": "[4] A Bhattacharya, R Madaan, F Cladera, et al. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5846-5855, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "type": "text", + "content": "[5] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. pages 497-504. 2001. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "type": "text", + "content": "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. arXiv preprint arXiv:2304.02602, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 429 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "type": "text", + "content": "[8] H Chen, J Gu, A Chen, et al. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2416-2425, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 475, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 295, + 508 + ], + "type": "text", + "content": "[9] P Chen, W Guan, F Huang, et al. Ecmd: An event-centric multisensory driving dataset for slam. IEEE Transactions on Intelligent Vehicles, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "text", + "content": "[10] P Chen, W Guan, and P Lu. Espvio: Event-based stereo visual inertial odometry. IEEE Robotics and Automation Letters, 8 (6):3661-3668, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[11] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 643 + ], + "type": "text", + "content": "[12] Zhen Cheng, Yutong Liu, and Zhiwei Xiong. Spatial-angular versatile convolution for light field reconstruction. IEEE Transactions on Computational Imaging, 8:1131-1144, 2022. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 678 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[14] A Dosovitskiy, G Ros, F Codevilla, et al. Carla: An open urban driving simulator. In Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Brandon Yushan Feng and Amitabh Varshney. Signet: Efficient neural representation for light fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14224-14233, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[16] P Furgale, J Rehder, and R Siegwart. Unified temporal and spatial calibration for multi-sensor systems. In 2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1280-1286. IEEE, 2013. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5633-5643, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "text", + "content": "[18] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "type": "text", + "content": "[19] M Gehrig, W Aarents, D Gehrig, and D Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 309, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 373 + ], + "type": "text", + "content": "[20] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808–11826. PMLR, 2023, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 419 + ], + "type": "text", + "content": "[21] Mantang Guo, Junhui Hou, Jing Jin, Hui Liu, Huanqiang Zeng, and Jiwen Lu. Content-aware warping for view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9486-9503, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 453 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "text", + "content": "[23] Jing Jin, Junhui Hou, Hui Yuan, and Sam Kwong. Learning light field angular super-resolution via a geometry-aware network. In Proceedings of the AAAI conference on artificial intelligence, pages 11141-11148, 2020. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 501, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 532 + ], + "type": "text", + "content": "[24] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM TOG, 35(6):1-10, 2016. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[25] J Karras, A Holynski, T C Wang, et al. Dreampose: Fashion image-to-video synthesis via stable diffusion. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22623-22633, 2023. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 632 + ], + "type": "text", + "content": "[26] B Ke, A Obukhov, S Huang, et al. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492–9502, 2024. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Zhong Li, Liangchen Song, Celong Liu, Junsong Yuan, and Yi Xu. Neulf: Efficient novel view synthesis with neural 4d light field. arXiv preprint arXiv:2105.07112, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11215" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[29] Zhicheng Lu, Xiaoming Chen, Vera Yuk Ying Chung, Weidong Cai, and Yiran Shen. Ev-lfv: Synthesizing light field event streams from an event camera and multiple rgb cameras. IEEE Transactions on Visualization and Computer Graphics, 2023. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[30] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 252 + ], + "type": "text", + "content": "[32] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[33] C Mou, X Wang, L Xie, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 363 + ], + "type": "text", + "content": "[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "type": "text", + "content": "[35] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "text", + "content": "[36] R Rombach, A Blattmann, D Lorenz, et al. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "type": "text", + "content": "[37] S Saxena, C Herrmann, J Hur, et al. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "text", + "content": "[38] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. NeurIPS, 34:19313-19325, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "text", + "content": "[39] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "text", + "content": "[40] S M N Uddin, S H Ahmed, and Y J Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[41] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[42] Huan Wang, Jian Ren, Zeng Huang, Kyle Olszewski, Mengei Chai, Yun Fu, and Sergey Tulyakov. R2l: Distilling neural radiance field to neural light field for efficient novel view synthesis. In European Conference on Computer Vision, pages 612-629. Springer, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "text", + "content": "[43] Yunlong Wang, Fei Liu, Zilei Wang, Guangqi Hou, Zhenan Sun, and Tieniu Tan. End-to-end view synthesis for light field imaging with pseudo 4dcnn. In ECCV, pages 333-348, 2018. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "type": "text", + "content": "[44] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2022. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[46] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[47] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "type": "text", + "content": "[48] Gaochang Wu, Belen Masia, Adrian Jarabo, Yuchen Zhang, Liangyong Wang, Qionghai Dai, Tianyou Chai, and Yebin Liu. Light field image processing: An overview. IEEE Journal of Selected Topics in Signal Processing, 11(7):926-954, 2017. 1, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "text", + "content": "[49] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[50] J Wynn and D Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[51] J Xiang, J Yang, B Huang, et al. 3d-aware image generation using 2d diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2383-2393, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[52] Henry Wing Fung Yeung, Junhui Hou, Jie Chen, Yuk Ying Chung, and Xiaoming Chen. Fast light field reconstruction with deep coarse-to-fine modeling of spatial-angular clues. In ECCV, pages 137-152, 2018. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[53] L Zhang, A Rao, and M Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[55] Y Zhou, G Gallego, and S Shen. Event-based stereo visual odometry. IEEE Transactions on Robotics, 37(5):1433-1450, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11216" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_content_list.json b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..22f54a8e13f1530f185d50c05e87709fde17f829 --- /dev/null +++ b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_content_list.json @@ -0,0 +1,1931 @@ +[ + { + "type": "text", + "text": "S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 161, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yecong Wan $^{1}$ Mingwen Shao $^{1*}$ Yuanshuo Cheng $^{1}$ Wangmeng Zuo $^{2}$", + "bbox": [ + 207, + 179, + 790, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software, Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China)", + "bbox": [ + 101, + 199, + 897, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Faculty of Computing, Harbin Institute of Technology", + "bbox": [ + 281, + 233, + 715, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "smw278@126.com, {yecongwan, cys1294414023, cswmzuo} $@$ gmail.com", + "bbox": [ + 200, + 252, + 795, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/630e80a6bae84f426c3564fed8b99b3652d7d081c6fea98f4b2eb133a9c3eca0.jpg", + "image_caption": [ + "Figure 1. We present S2Gaussian, a novel framework capable of reconstructing high-quality 3D scenes for immersive rendering with only sparse and low-resolution input views. S2Gaussian demonstrates superior performance and yields high-fidelity and high-resolution scene reconstruction with sharp geometric and detailed textures, thus enjoying better functionality and practicality in realistic applications." + ], + "image_footnote": [], + "bbox": [ + 112, + 303, + 619, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06238b75f28db2475544560de0f253bf0587c6b4f521b2be8a701effcb220f03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 311, + 901, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 513, + 326, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim ambitiously for a realistic yet challenging problem, namely, how to reconstruct high-quality 3D scenes from sparse low-resolution views that simultaneously suffer from deficient perspectives and clarity. Whereas existing methods only deal with either sparse views or low-resolution observations, they fail to handle such hybrid and complicated scenarios. To this end, we propose a novel Sparse-view Super-resolution 3D Gaussian Splatting framework, dubbed S2Gaussian, that can reconstruct structure-accurate and detail-faithful 3D scenes with only sparse and low-resolution views. The S2Gaussian operates in a two-stage fashion. In the first stage, we initially optimize a low-resolution Gaussian representation with depth regularization and densify it to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation. In the second stage, we refine the high-resolution Gaussians with the super-resolved images generated from both original sparse views and pseudoviews rendered by the low-resolution Gaussians. In which a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are elaborately designed to mitigate multi-view inconsistency and eliminate error", + "bbox": [ + 89, + 540, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "neous updates caused by imperfect supervision. Extensive experiments demonstrate superior results and in particular establishing new state-of-the-art performances with more consistent geometry and finer details. Project Page https://jeasco.github.io/S2Gaussian/.", + "bbox": [ + 511, + 516, + 908, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 603, + 643, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of virtual reality and metaverse technologies has significantly heightened the demand for realistic 3D scene reconstruction, which holds vast potential across various application domains, including medicine, education, entertainment, etc. Among the recent notable developments, Neural Radiance Fields (NeRF) [25] have demonstrated remarkable capability in synthesizing photorealistic images; however, they still face considerable rendering and training costs, despite ongoing improvements [4, 10, 11, 26]. More recently, 3D Gaussian Splatting (3DGS) [18] has emerged as a prominent approach, characterized by its high quality, rapid reconstruction speed, and support for real-time rendering. Subsequent research efforts [5, 6, 23, 32, 39, 45] have focused on extending the applicability of 3DGS across various scenarios.", + "bbox": [ + 509, + 628, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, these methods typically rely heavily on well-captured dense and high-resolution images for impressive novel view synthesis, which is cumbersome and some", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ * Corresponding authors.", + "bbox": [ + 107, + 886, + 251, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "711", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "times impractical in real-world applications. To address this challenge, a variety of approaches have emerged that focus on reconstructing 3D scenes from sparse [7, 8, 22, 27, 35, 40, 42, 50] and low-resolution views [9, 12, 21, 31, 43, 44], yielding worth-celebrating successes. However, these two obstacles, i.e., sparsity and low resolution, have long been treated as isolated problems and tackled separately, whereas input views in practical applications often suffer from both deficient perspectives and clarity. Particularly in contexts such as robotics and internet-collected imagery, where the available views are usually sparse and compounded by low resolutions due to environmental constraints and hardware transfer limitations. This interplay of sparsity and low-resolution poses significant challenges for existing reconstruction frameworks. Furthermore, merely incorporating existing methods remains subject to inherent incompatibility, as super-resolution necessitates dense supervision to recover fine details while sparse-view regularization not only fails to introduce details but tends to lead smoothing.", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the aforementioned problems, in this paper, we propose a novel 3D reconstruction framework S2Gaussian, capable of reconstructing geometrically precise and richly detailed 3D scenes with only sparse and low-resolution views. Specifically, S2Gaussian consists of two main phases: the high-resolution Gaussians (HR GS) initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the estimated depth information. The optimized low-resolution Gaussians are then utilized to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation which can provide more compact Gaussian primitives to facilitate fine-grained details reconstruction in high-resolution scenes. In the HR GS optimization stage, we use super-resolved images generated by pretrained super-resolution model to refine the high-resolution Gaussians, leveraging both the original sparse views and the pseudo-views rendered from the low-resolution Gaussians. Notably, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are introduced to mitigate the impacts of multi-view inconsistency and imperfect supervision, ultimately leading to more detailed and higher-quality scene reconstruction. Comprehensive experiments substantiate the superiority of our S2Gaussian beyond alternatives on a variety of benchmarks.", + "bbox": [ + 91, + 381, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, the main contributions are as follows:", + "bbox": [ + 109, + 777, + 455, + 791 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an innovative two-stage framework, termed S2Gaussian, designed to reconstruct 3D scenes that are both structurally accurate and rich in detail using only sparse and low-resolution input views.", + "- A dedicated Gaussian Shuffle Split operation is designed to initialize more compact Gaussian primitives for representing fine-grained details and textures in high" + ], + "bbox": [ + 91, + 795, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "resolution scenes.", + "bbox": [ + 529, + 92, + 651, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a tailored blur-free inconsistency modeling scheme alongside a 3D robust optimization strategy to address multi-view inconsistencies and rectify erroneous updates arising from imperfect supervision.", + "- Our S2Gaussian reconstructs more fine-grained and high-quality Gaussian representations significantly outperforming existing methods across multiple benchmarks." + ], + "bbox": [ + 514, + 106, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Novel View Synthesis using Radiance Fields", + "text_level": 1, + "bbox": [ + 511, + 250, + 883, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Techniques for novel view synthesis typically involve learning a 3D representation using a limited number of input views and generating images from arbitrary novel perspectives. Recent advancements in Neural Radiance Fields (NeRF) [25] has demonstrated encouraging progress in this field, which learns an implicit neural scene representation for novel view synthesis via coordinate-based neural networks and volume rendering function. Subsequent research has focused extensively on enhancing NeRF's rendering quality [1-3, 33], improving efficiency [4, 10, 11, 26], advancing scene understanding [19, 48, 49], and facilitating 3D content generation [15, 28, 30]. Unfortunately, while NeRF significantly improves the quality of novel view rendering, its expensive training time as well as slow rendering speed hinder broader practical applications. Recently, Kerbl et al. [18] introduced a groundbreaking approach with their 3D Gaussian Splatting (3DGS) method, boosting rendering efficiency by utilizing explicit Gaussian representations combined with differentiable rasterization techniques. Building upon 3D Gaussian Splatting representation, tremendous following efforts [6, 23, 32, 39, 45] concentrate extending 3DGS under various scenarios. Nevertheless, these methods typically require dense and high-quality input views for impressive novel view synthesis, and they will tend to overfit the training data and compromise representations when provided with only sparse or degraded views, leading to a substantial decline in performance for the intended scene reconstruction.", + "bbox": [ + 511, + 271, + 906, + 693 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Sparse Novel View Synthesis", + "text_level": 1, + "bbox": [ + 511, + 699, + 769, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the availability in practical usage, numerous algorithms have been developed to surmount the reliance on dense perspectives. For instance, Depth-NeRF [8] and Reg-NeRF [27] advocate utilizing depth constraints to improve the accuracy of the sparse scene representation. FreeN-erF [42], on the other hand, employs a frequency regularization strategy to manage the frequency range and penalize density fields near the camera. Additionally, SparseN-erF [35] leverages depth priors from inaccurate observations, implementing local depth ranking and spatial continuity constraints to bolster neural radiance fields. More recently, a variety of 3DGS-based approaches have emerged", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "712", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[7, 22, 40, 50], paving the way for new blueprints in the field. These methods typically focus on various strategies that utilize depth prior to constrain the optimization of Gaussian primitives, such as SfM depth alignment [7], Pearson correlation depth distribution loss [50], global-local depth normalization [22], and patch-based depth correlation [40]. Unlike these approaches, we spotlight more arduous sparse reconstruction under low-resolution conditions, where available information is more scarce.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Super-Resolution Novel View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 239, + 426, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Super-resolution novel view synthesis aims to reconstruct high-resolution 3D scenes utilizing only low-resolution multi-view inputs. Unlike anti-aliasing methods [1, 2, 45] which are incapable of rendering intricate details of the novel views and are limited by the input image's information level, this field is more focused on fine-grained and detailed high-resolution rendering of potential high-resolution scenes. As a pioneer in the field, NeRF-SR [34] optimizes high-resolution NeRF through a super-sampling strategy, ensuring that the values of low-resolution pixels match the mean value of high-resolution sub-pixels. Follow-up works are designed to utilize either high-resolution reference images [16] or pre-trained 2D models [12, 21, 43] to generate multi-view consistent details. Concurrently, SRGS [9] and GaussianSR [44] are proposed to utilize 2D super-resolution models for texture injection or diffusion prior exploration, respectively. SuperGaussian [31] and SM [20] further prove two profile frameworks to repurpose video upsampling models to achieve 3D super-resolution.", + "bbox": [ + 89, + 261, + 482, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the aforementioned methods, in this work, we ambitiously tackle the challenge of joint sparse-view and super-resolution novel view synthesis by proposing a novel two-stage framework, which cleverly solves the problem of perspectives and clarity deficient to reconstruct geometry-accurate and fine-detailed 3D scenes, making it more desirable and feasible in real-world applications.", + "bbox": [ + 89, + 549, + 482, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 670, + 223, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 696, + 230, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) as an emerged powerful technique for novel view synthesis, models a 3D scene using a collection of $G$ colored Gaussians:", + "bbox": [ + 89, + 719, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} (x) = \\exp \\left(- \\frac {1}{2} \\left(x - \\mu_ {i}\\right) ^ {\\top} \\Sigma_ {i} ^ {- 1} \\left(x - \\mu_ {i}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 763, + 480, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $1 \\leq i \\leq G$ , $\\mu_i \\in \\mathbb{R}^3$ represents the mean or center of the Gaussian, and $\\Sigma_i \\in \\mathbb{R}^{3 \\times 3}$ denotes Gaussian's covariance matrix, specifying its shape and size, which is described by scaling vector $s \\in \\mathbb{R}_+^3$ and rotation vector $q \\in \\mathbb{R}^4$ for implementation. Additionally, each Gaussian is characterized by an opacity $\\sigma_i \\in \\mathbb{R}_+$ and a view-dependent color $c_i(v) \\in \\mathbb{R}^{k_c}$ (where $k_c$ is changed) for rasterization.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To faithfully reconstruct the 3D scene, 3DGS also employs densification operation, i.e., the split, clone, and prune operation, to adaptively regulate the number and density of Gaussian primitives. For simplicity, we deploy the same adaptive density control as the original 3DGS and will not elaborate further in the subsequent presentation.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. S2Gaussian Overview", + "text_level": 1, + "bbox": [ + 511, + 186, + 718, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The schematic illustration of the proposed S2Gaussian is depicted in Fig. 2. S2Gaussian is primarily composed of two phases i.e., the HR GS initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the predicted depth information, which is further utilized to initialize the high-resolution Gaussian representation by densifying and detailing the Gaussian primitives through our tailored Gaussian Shuffle Split operation. In the HR GS optimization stage, we further optimize the high-resolution Gaussians with the super-resolved images generated by pre-trained super-resolution model from both original sparse views and the pseudo-views rendered by LR Gaussians. In particular, a dedicated blur-free inconsistency modeling scheme and a 3D robust optimization strategy are advanced to eliminate the effects of multi-view inconsistency and imperfect supervision to reconstruct more detailed and high-quality 3D scenes.", + "bbox": [ + 511, + 205, + 903, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Stage 1: HR GS Initialization", + "text_level": 1, + "bbox": [ + 511, + 501, + 774, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Directly reconstructing a high-resolution 3D scene from sparse low-resolution viewpoints is extremely challenging, as it is not only prone to overfitting specific viewpoints but also lacks high-resolution detail constraints. To alleviate such a reconstruction dilemma, we advocate initially reconstructing a low-resolution Gaussian representation as an auxiliary to facilitate the optimization of high-resolution Gaussians. On the one hand, the low-resolution Gaussian can be used to initialize high-resolution Gaussians to provide reliable base Gaussian ellipsoids and 3D scene structures. On the other hand, it can provide pseudo-views as additional supervision to avoid overfitting on specific viewpoints as well as provide more potential details and textures for high-resolution scenes.", + "bbox": [ + 511, + 523, + 903, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, we first obtain the monocular depth maps for the sparse low-resolution training views by utilizing the handy pre-trained depth estimation model [29]. Then we utilize the RGB images as well as the depth maps to co-optimize a low-resolution Gaussian representation with mature depth regularization technologies [22, 50], where Pearson correlation loss [50] is employed in this paper as a baseline. Thereafter, a tailored Gaussian Shuffle Split regime is proposed to densify and detail the sparse Gaussian primitives so as to initialize HR Gaussians for better high-resolution fine-grained details representation.", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "713", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/424ae6196f961eb89564199daa0b6f9c870f8b18c80aa756da848c16ed02fbba.jpg", + "image_caption": [ + "Figure 2. Overview of S2Gaussian. The S2Gaussian initially optimizes an LR GS and densifies it to initialize the HR GS through a tailored Gaussian Shuffle Split operation. Then, the original sparse views along with the pseudo views rendered by the LR GS are super-resolved together to refine the high-resolution texture with 3D robust optimization. In which an inconsistency modeling module (IM) and a blur proposal module are incorporated to mitigate inconsistency and blurriness, aiming to create 3D scenes with high-fidelity texture details." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 911, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian Shuffle Split. A precise and detailed initialization is crucial for reconstructing scene details [18]. Although the structure and layout of the entire 3D scene are basically guaranteed after the optimization of the low-resolution views, the initialized sparse and coarse Gaussian ellipsoids struggle to simulate high-resolution details which require denser Gaussians for finer representation [41]. Even equipped with adaptive density control [18], it falls short in our settings as there are neither dense perspectives nor high-quality details for supervision. To address this issue, we propose a training-free local Gaussian densification strategy Gaussian Shuffle Split which can provide more mobilizable Gaussian primitives by replacing the original large Gaussian with six small Gaussians to facilitate more comprehensive simulation of fine-grained details and textures in high-resolution scenes.", + "bbox": [ + 88, + 388, + 485, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, given a Gaussian primitive with attributes $\\{\\mu, s, q, \\sigma, c\\}$ , we first generate six duplicates, each center shifted along one of the six directions aligned with the original Gaussian's three principal axes, i.e., the positive and negative directions of each axis. The shift for each axis is set to $\\alpha$ (0.5 by default) times the corresponding axis's scaling $s = [s_1, s_2, s_3]$ value. Thus the new position of the six sub-Gaussians can be obtained:", + "bbox": [ + 89, + 633, + 483, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {s u b 1} = \\mu + R \\otimes [ \\alpha s _ {1}, 0, 0 ] ^ {T}, \\mu_ {s u b 2} = \\mu + R \\otimes [ - \\alpha s _ {1}, 0, 0 ] ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 753, + 483, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {s u b 3} = \\mu + R \\otimes [ 0, \\alpha s _ {2}, 0 ] ^ {T}, \\mu_ {s u b 4} = \\mu + R \\otimes [ 0, - \\alpha s _ {2}, 0 ] ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 773, + 483, + 791 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {s u b 5} = \\mu + R \\otimes [ 0, 0, \\alpha s _ {3} ] ^ {T}, \\mu_ {s u b 6} = \\mu + R \\otimes [ 0, 0, - \\alpha s _ {3} ] ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 794, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R$ represents the rotation matrix calculated from $q$ to align with the original rotation of the Gaussian. Additionally, for each shifted sub-Gaussian, we scale their $s$ value corresponding to the offset axis to $\\frac{1}{4}$ of its original value, i.e., reducing the scale along that axis to shrink the size.", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The remaining two $s$ values are scaled by an another factor of $\\lambda$ relative to their original values:", + "bbox": [ + 511, + 388, + 906, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {s u b 1} = s _ {s u b 2} = \\left[ s _ {1} / 4, s _ {2} / \\lambda , s _ {3} / \\lambda \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 419, + 831, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {s u b 3} = s _ {s u b 4} = \\left[ s _ {1} / \\lambda , s _ {2} / 4, s _ {3} / \\lambda \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 438, + 903, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {s u b 5} = s _ {s u b 6} = \\left[ s _ {1} / \\lambda , s _ {2} / \\lambda , s _ {3} / 4 \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 458, + 828, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is experimentally set to 1.9 which we found this ratio ensures that the combined representation of the sub-Gaussians better approximate the original Gaussian, preserving the integrity of the 3D scene. All other attributes, i.e., rotation $q$ , opacity $\\alpha$ , and color $c$ remain identical to the original Gaussian ellipsoid. The combined six sub-Gaussians are utilized to substitute the original large Gaussians for densified 3D representation.", + "bbox": [ + 511, + 474, + 905, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We deployed Gaussian Shuffle Split only for Gaussian primitives with opacity greater than 0.5 since these are more likely to be responsible for representing object surfaces and key structures which require a more detailed representation in high-resolution scenes. The opacity of all primitives is set close to zero after the Gaussian Shuffle Split to facilitate the automatic elimination of redundant Gaussians with low opacity.", + "bbox": [ + 511, + 595, + 905, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Stage 2: HR GS Optimization", + "text_level": 1, + "bbox": [ + 511, + 726, + 779, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the initialized high-resolution Gaussians, in this stage, we aim to continue optimizing it with the super-resolved images generated from both original sparse views and the pseudo-views rendered by the LR Gaussians. We synthesize pseudo views by interpolating virtual views between two known views as in [50]. Considering the prohibitively expensive data costs associated with 3D super-resolution models and their computational complexity, we advocate for utilizing off-the-shelf pre-trained 2D super-resolution models to enhance low-resolution views. Unfortunately, directly", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "714", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8d1d65b005d2cf70f02cb48aac04cbcb9537458b39d5a23bad55934c96c644b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 287, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37b4a927806a721ce09fe26efd8781ac1131801c1e38c04c91b657546719425f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 89, + 473, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc2bccd44a7fa650901abd438e94a4477771f78b0af386aed1255496e6b16189.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 171, + 282, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/210db062e178916ae853f4fa18e1883858beada1c679997cfb80b0dd6e6e02d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 171, + 475, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ccbe3c2ca2f95c9a518b4c35466701dfe919d19b8fb450a4b7ee85dfa71770d2.jpg", + "image_caption": [ + "Figure 3. Illustration of Gaussian Shuffle Split that utilizes a combination of six compact Gaussians to replace the original large Gaussian primitive, which can densify the 3D scene virtually without damaging the original 3D representation." + ], + "image_footnote": [], + "bbox": [ + 96, + 253, + 282, + 354 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4327336c60bc6d1aebf227d0ea27a3beb11bf0ca073a5e3fe584d07520f274e8.jpg", + "image_caption": [ + "Figure 4. Top: Visualization of the gradient variation under different quality views supervision and our proposed 3D robust optimization. Bottom: Illustration of the proposed 3D robust optimization strategy that takes the gradient of scaling $s$ as an example." + ], + "image_footnote": [], + "bbox": [ + 290, + 253, + 477, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "leveraging 2D super-resolution model cannot ensure multiview consistency, and the low-resolution pseudo-views inevitably exhibit artifacts since there are areas that have not been optimized before. The Gaussian primitives, in their attempt to accommodate these inconsistencies and erroneous representations, will lead to the optimized 3D scene converging toward inaccurate structures and blurriness.", + "bbox": [ + 89, + 429, + 483, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During optimization, two crucial aspects contribute fundamentally to these problems, i.e., inconsistent supervision and erroneous updates caused by artifacts. Accordingly, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are proposed in the following to alleviate potential inaccurate reconstruction.", + "bbox": [ + 89, + 537, + 483, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blur-Free Inconsistency Modeling. To mitigate the inconsistency effort of the single-image super-resolution model, we employ a learnable inconsistency modeling module (IM), i.e., two residual blocks [13], after the pre-trained super-resolution model to simulate the inconsistency across different views $I_{SR}^{IM} = I_{SR} + IM(I_{SR})$ , thus avoiding making Gaussians to represent such inconsistency. However, we empirically found that this module tends to lose details during optimization to win better consistency, resulting in the loosing of sharp textures. Therefore, we further propose a blur proposal module (BP) to blur the rendered image which is then constrained with modified super-resolution image $I_{SR}^{IM}$ , thereby compensating for the loss of fine-grained details and textures caused by IM. The blur proposal module, i.e., a four-layer convolutional network, predicts per-pixel blur kernels $\\mathcal{B}_k \\in \\mathbb{R}^{K \\times K \\times H \\times W}$ ( $K$ is the blur kernel size, 5 for $4 \\times$ task) with the gradient detached rendered image $R_{HR}^{\\mathrm{detach}}$ as input $\\mathcal{B}_k = BP(R_{HR}^{\\mathrm{detach}})$ . Then the blurred", + "bbox": [ + 89, + 628, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e83f08560595106cc399211177c1368b590b1eff527934c11e78508cc8cef801.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f73887a9fdfcab8e10049b4ee6b9d9eeab197fb1c26da2a1ab6d63374ac4707b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 194, + 903, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "high-resolution image is obtained via $R_{HR}^{blur} = R_{HR} * \\mathcal{B}_k$ which is utilized to compute loss with $I_{SR}^{IM}$ , avoiding the smoothing imposed by directly constraining $R_{HR}$ :", + "bbox": [ + 511, + 402, + 903, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S R} = (1 - \\beta) \\mathcal {L} _ {1} \\left(R _ {H R} ^ {\\text {b l u r}}, I _ {S R} ^ {\\text {I M}}\\right) + \\beta \\mathcal {L} _ {D - S S I M} \\left(R _ {H R} ^ {\\text {b l u r}}, I _ {S R} ^ {\\text {I M}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 450, + 903, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is set to 0.2 similar to the original 3DGS [18].", + "bbox": [ + 511, + 479, + 874, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Robust Optimization. Albeit the above scheme could surmount the detrimental effects of multi-view inconsistencies, the optimization of Gaussians still unavoidably suffers from improper supervision of pseudo-views where some areas are not sufficiently optimized. To this end, we empirically observed that optimizing Gaussian primitives utilizing normal high-quality views results in stable and overall consistent gradients (top left of Fig. 4) whereas corrupted views can lead to significant perturbation and gradient confusion (top middle of Fig. 4), ultimately resulting in blurry rendering. Intrigued by this observation, we aim to design a more robust optimization strategy to mitigate the fluctuations caused by imperfect supervision.", + "bbox": [ + 511, + 496, + 905, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More specifically, we augment the canonical Gaussian primitives", + "bbox": [ + 511, + 693, + 903, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\mu , s, q, \\sigma , c \\} \\cup \\{\\nabla \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 724, + 903, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with an additional per-Gaussian attribute flag gradient $\\nabla \\in \\mathbb{R}^{11 + k_c}$ . Where $\\nabla$ is used to record the gradient trends of other attributes of the Gaussian, i.e., $\\mu, s, q, \\sigma$ , and the changed $k_{c}$ SH coefficients. As shown in Fig. 4, for each Gaussian in one optimization step, we first compute the cosine similarity between the current gradient for each attribute $g_i^c$ ( $i$ indicate different attribute) and the corresponding flag gradient $\\nabla (g_i)$ stored in $\\nabla$ . If the cosine similarity is greater than zero, indicating that the gradient update trends are aligned, the gradient $g_i^c$ will be directly used to", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "715", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cfd46f92fa656f57060930e3d639172e162b61add564e4ac2cd43ba2527472c0.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on Blender $\\times 4$ (8 views),LLFF $\\times 4$ (3 views),and Mip-NeRF360 $\\times 4$ (24 views). The best,second best, and third best entries are marked in red, orange, and yellow, respectively." + ], + "table_footnote": [], + "table_body": "
MethodBlender ×4 (8 views)LLFF ×4 (3 views)Mip-NeRF 360 ×4 (24 views)
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
NeRF-SR [34]12.410.7440.51593.799.280.2260.617169.5810.260.2690.628151.62
RegNeRF [27]20.680.8410.12941.2815.780.4320.422116.7417.280.4170.449105.45
SparseNeRF [35]21.320.8440.13642.6216.320.4460.428127.6417.380.4320.446109.27
3DGS [18]20.650.8310.14142.5813.010.2720.489138.5316.250.3420.474108.32
SRGS [9]22.640.8620.11636.2417.580.4720.291102.4318.370.4570.39791.62
Mip-Splatting [45]22.780.8680.11933.6315.650.4330.372114.0318.080.4480.42186.17
DNGaussian [22]20.970.8440.13541.1516.270.4300.346129.2617.210.4220.452112.22
FSGS [50]21.450.8540.13340.3618.350.5050.24797.0718.640.4580.39494.21
FSGS [50]+SRGS [9]22.680.8650.11435.4219.490.5900.17562.6319.830.5130.44884.12
S2Gaussian (Ours)24.190.8790.08928.4720.450.6540.13945.8922.050.6870.29643.51
", + "bbox": [ + 91, + 128, + 911, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimize the parameter, and the flag gradient value $\\nabla (g_i)$ will be updated to the average one:", + "bbox": [ + 89, + 313, + 480, + 343 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} ^ {c} = g _ {i} ^ {c}, \\nabla (g _ {i}) = (\\nabla (g _ {i}) + g _ {i} ^ {c}) / 2. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 344, + 480, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conversely, if the cosine similarity is less than or equal to zero, indicating potential disturbances, the current gradient will be scaled down by a factor of $\\epsilon$ to slow the parameter updates, and $\\nabla (g_i)$ will be updated as follows:", + "bbox": [ + 89, + 361, + 482, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} ^ {c} = \\epsilon g _ {i} ^ {c}, \\nabla (g _ {i}) = (1 - \\epsilon) \\nabla (g _ {i}) + \\epsilon g _ {i} ^ {c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 422, + 480, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon$ is set to 0.1 to attenuate potentially erroneous updates and preserve the ability to shift update trends. In practice, such strategy can be combined with any gradient-based optimizer, e.g., Adam, by simply passing the modified gradients to the respective optimizer. Notably, $\\nabla$ will be eliminated after the optimization and will not affect the rendering speed and storage consumption.", + "bbox": [ + 89, + 439, + 483, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in the top right of Fig. 4, our proposed 3D robust optimization scheme can significantly improve the stability of Gaussian primitive optimization by eliminating erroneous updates and perturbations thus reconstructing more accurate and high-quality 3D representations.", + "bbox": [ + 89, + 550, + 483, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total Objective. Apart from $\\mathcal{L}_{SR}$ , an auxiliary loss combined with total variation (TV) loss and sub-pixel constraint is also deployed to stabilize the training:", + "bbox": [ + 89, + 631, + 483, + 675 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A U X} = \\mathcal {L} _ {T V} \\left(R _ {H R}\\right) + \\mathcal {L} _ {1} \\left(R _ {H R} \\downarrow , I _ {L R}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 676, + 480, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\downarrow$ indicates area average downsampling and $I_{LR}$ denotes the low-resolution images before super-resolution.", + "bbox": [ + 89, + 694, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the final loss $\\mathcal{L}$ for high-resolution Gaussians optimization is defined as:", + "bbox": [ + 89, + 729, + 482, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {A U X} + \\mathcal {L} _ {S R}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 761, + 480, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 797, + 223, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following prior literature [21, 44], we primarily experiment on $4 \\times$ super-resolution tasks in the maintext since most real-world 2D super-resolution models [36, 46] are mainly trained on $\\times 4$ data, more diagnostic experiments on other scaling factors are provided in the Suppl.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 311, + 730, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implemented S2Gaussian upon the original 3DGS repository and all experiments are conducted on a single RTX3090 GPU. For the optimization of LR GS, the pretrained DPT [29] model is adapted for depth estimation and the total iterations are set to 10,000. During the HR GS optimization stage, the Gaussian primitives are further optimized with additional 10,000 iterations. As for the off-the-shelf 2D super-resolution model, we opt for powerful diffusion-based ResShift [46] as our backbone. We conduct experiments on three benchmark datasets with SfM for initialization, advanced techniques like DUSt3R [37] can facilitate the initialization but that is not the main purpose of our current work.", + "bbox": [ + 511, + 338, + 906, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Blender Datasets [25] contain 8 detailed objects synthesized by Blender with a resolution of $800 \\times 800$ . Each object initially includes 100 training images and 200 testing images. We uniformly sample 8 images from the original 100 training images and use the low-resolution of $200 \\times 200$ pixels for training.", + "bbox": [ + 511, + 539, + 906, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLFF Datasets [24] consist of 8 forward-facing real-world scenes. Following the previous setup, we select every eighth image as the test set, and evenly sample 3 sparse views from the remaining images for training. We leverage low-resolution images at $252 \\times 189$ pixels for training and render at a higher resolution of $1008 \\times 756$ pixels.", + "bbox": [ + 511, + 633, + 905, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mip-NeRF 360 Datasets [2] consist of 9 real-world scenes with 5 outdoors and 4 indoors. Similar to LLFF, we use 1/8 of the full view for testing and uniformly sample 24 images from the remaining views for training. We utilize the images with downsampling rate of 4 (about 1K resolution) as the target size and further downsample the training images by a factor 4 to obtain sparse low-resolution inputs.", + "bbox": [ + 511, + 729, + 905, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The following measures are used to evaluate the quantitative performance: Peak Signal-to-Noise Ratio (PSNR) [17], Structural Similarity Index (SSIM) [38], LPIPS (VGG) [47], and Fréchet Inception Distance (FID) [14].", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "716", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ee8c502f3b802987b70ab119c737720662ddc2cdebbc68f92962419ecd13f59.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 215, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b49094f07a701c2d6a7b46b477bfa9d2960f8acc79682e4963934c06f4f887bc.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 227, + 88, + 349, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fb05117f51f7b03ccadcf6d66f97d63728bd41f1a98fa01ad53f284e48a719d.jpg", + "image_caption": [ + "Mip-Splatting" + ], + "image_footnote": [], + "bbox": [ + 364, + 88, + 485, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5f38332568d3d2881aa01c46eaf6599cc58f4a96d280c0e2cfb5a8d7885226d.jpg", + "image_caption": [ + "FSGS" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 622, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8c8350b7aead2d9694aed8280321ffe00f9a74b3565aa6faf892fcfe17e1b9a.jpg", + "image_caption": [ + "FSGS+SRGS" + ], + "image_footnote": [], + "bbox": [ + 637, + 88, + 758, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f039759540863a34aa76cfdacaaa258acf62e8e6caf0efeff67f54866e42d0f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 772, + 88, + 893, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c08e8af8a74d4467922fd5174e34ae9b9f1a48cc6867b891a19121dcf7aaa510.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons on Blender $\\times 4$ datasets with 8 input views. Our method produces more visually appealing results, successfully reconstructing intricate thin structures with fine-grained details. Best viewed at screen!", + "Ground Truth", + "Figure 6. Qualitative comparisons on LLFF $\\times 4$ datasets with 24 input views. Our method still generates the most accurate super-resolution 3D scenes, preserving both photo-realistic texture and geometric-level details. Best viewed at screen!" + ], + "image_footnote": [], + "bbox": [ + 91, + 332, + 222, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10a48e4441ef3e9c61defbdf525bb939f5c4e06dc1d2b381714dda376a8ea0c5.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 227, + 332, + 359, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfbd85935a5f295d6a5a37833305338f7ce8875dce0b0d940dc481982f4c26dd.jpg", + "image_caption": [ + "Mip-Splatting" + ], + "image_footnote": [], + "bbox": [ + 364, + 332, + 495, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91cc7d8a5d21a3636a211561149a2b2b91b3631719cdf976f17bd5d2f3b207b0.jpg", + "image_caption": [ + "FSGS" + ], + "image_footnote": [], + "bbox": [ + 501, + 332, + 630, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c4171e4c843d0cdfc5e96c427ea8764903e6bfcbc2f6ea9e53e6a053dc0377e.jpg", + "image_caption": [ + "FSGS+SRGS" + ], + "image_footnote": [], + "bbox": [ + 637, + 332, + 767, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74a43a383cd9b8066316b297ae70efa477dd759003cbae6a8ae4d239ecbde568.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 772, + 332, + 903, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative and Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 89, + 550, + 454, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To rigorously validate the effectiveness of our proposed S2Gaussian, we conduct a comparative analysis against a range of prior methods, including NeRF-SR [34], RegNeRF [27], SparseNeRF [35], vanilla 3DGS [18], SRGS [9], Mip-Splatting [45], DNGaussian [22], and FSGS [50]. Besides, we also include an enhanced version of FSGS with SRGS to provide a standard upsampling baseline. While NeRF-SR and FSGS+SRGS employ super-resolution techniques directly, the remaining methods are trained on low-resolution input views and subsequently rendered at higher resolutions.", + "bbox": [ + 88, + 577, + 482, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. Tab. 1 presents quantitative comparison results for $4 \\times$ sparse-view super-resolution novel view synthesis task on the Blender [25], LLFF [24], and Mip-NeRF 360 datasets [2]. It is observed that our S2Gaussian delivers remarkable performance gains and outperforms all competitive method significantly in terms of PSNR, SSIM, LPIPS, and FID metrics. Especially for Mip-NeRF 360 datasets, S2Gaussian exceeds the top-performing combinatorial method FSGS+SRGS by an impressive margin of $2.22\\mathrm{dB}$ PSNR, underscoring its superior capability", + "bbox": [ + 88, + 750, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in reconstructing large, real-world scenes. In addition to delivering higher reconstruction fidelity, reflected by PSNR and SSIM, S2Gaussian also achieves substantial improvements in perceptual metrics LPIPS and FID, validating the effectiveness of S2Gaussian in converging towards greater perception quality with faithful and fidelity details.", + "bbox": [ + 511, + 551, + 906, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We also demonstrate visual comparisons in Fig. 5, Fig. 6, and Fig. 7. As suggested, both 3DGS and Mip-Splatting struggle with the extremely sparse view problem, while FSGS is limited to characterizing high-resolution details. Despite the combinatorial method FSGS+SRGS can reconstruct a reasonable overall structure but suffers from inconsistency and sparsity leading to missing details and blurring. In comparison, S2Gaussian recovers more coherent fine structures with natural textures and harmonious global tone while other methods tend to introduce improper structures and unnatural details, especially for the extremely sparse scenes (Fig. 6) as well as large-scale scenarios (Fig. 7).", + "bbox": [ + 511, + 642, + 908, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 848, + 679, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we conduct ablation experiments on the components introduced in S2Gaussian. The effectiveness of each", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "717", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd461b963b07174c7903fac95d89133e1cdf2b09ca355d94a9d20624aa8a44fc.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 222, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae35a48c4d2055831c70e91db80e3286ae2afcf4fcae4638c8aa47a7f4917458.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 227, + 88, + 359, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b173ee3032253ce4ba54f4be93dea1b5636e32814f525a01b66d2ba46051123.jpg", + "image_caption": [ + "Mip-Splatting", + "Figure 7. Qualitative comparisons on Mip-NeRF360 $\\times 4$ datasets with 24 input views. Our method consistently delivers more coherent fine structures than other methods in large-scale scenes. Best viewed at screen!" + ], + "image_footnote": [], + "bbox": [ + 364, + 88, + 495, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37df40f5e31b1489a555c74e6bcee8ada0c18e164d9ff25da4afc5f41fd5883a.jpg", + "image_caption": [ + "FSGS" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 632, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb8fd95d8db285470d17ff8c64ed1efd30710a80784b817a7ab525b7a91fa13c.jpg", + "image_caption": [ + "FSGS+SRGS" + ], + "image_footnote": [], + "bbox": [ + 637, + 88, + 767, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0078b2c7752849c4c36e74686b81059db4266b55d8bfa0daeaf1fb2da9eafef8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 774, + 88, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proposed component in S2Gaussian is evaluated by gradually integrating them into the model, revealing their individual contributions to the overall performance. More detailed analyses are as below.", + "bbox": [ + 89, + 289, + 483, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cae117d966f0fcde7b6dd22c6d2e81840587286beff4d269063072400965018e.jpg", + "table_caption": [ + "Table 2. Ablation studies on Mip-NeRF ${360} \\times 4$ (24 views)." + ], + "table_footnote": [], + "table_body": "
VariantPSNR↑SSIM↑LPIPS↓FID↓
Baseline19.790.5200.42987.22
+ Two-Stage20.060.5280.41580.64
+ Gaussian Shuffle Split20.620.5590.37671.29
+ Blur-Free Inconsistency Modeling21.120.6240.34659.63
+ 3D Robust Optimization22.050.6870.29643.51
", + "bbox": [ + 91, + 367, + 480, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Gaussian Shuffle Split. As shown in Tab. 3, fewer sub-Gaussians lead to suboptimal detail modeling (lower PSNR), while 6 offers the best trade-off, while more sub-Gaussians will not only increase the optimization complexity but also introduce extra memory overhead. Additionally, 6 sub-Gaussians can be symmetrically located along the three principal axes of the original Gaussian, perfectly approximating the original Gaussian representation.", + "bbox": [ + 89, + 462, + 483, + 583 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f43729a0ba35848a2bc53aabe3cf50c63a883cba955d1b377c6558521953e602.jpg", + "table_caption": [ + "Table 3. Ablation of sub-Gaussian number on Mip-NeRF 360." + ], + "table_footnote": [], + "table_body": "
Number456789
Storage Memory (M)6789112137169202
Final PSNR21.7621.9522.0522.0322.0922.08
", + "bbox": [ + 91, + 607, + 478, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Blur-Free Inconsistency Modeling. As illustrated in Fig. 8, IM models the inconsistency of super-resolved images, and while this would introduce blurring, we model this effect explicitly through blur kernel learning, thus reconstructing results that are blur-free and with fine structural details.", + "bbox": [ + 89, + 659, + 483, + 750 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f4cd3efacf71d147c91fe1e6eed51861d7680b170044425e5c817470f88c385.jpg", + "image_caption": [ + "Super-Resolved Image", + "Figure 8. Visualization of the super-resolved image, the residual output of IM, the blurred reconstructed image with predicted blur kernels, and the original reconstructed output." + ], + "image_footnote": [], + "bbox": [ + 94, + 760, + 187, + 833 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ba83d76d48cd3c062a222554670627ad166ead0556eea84dabb7b2d7b993052.jpg", + "image_caption": [ + "Inconsistency Modeling" + ], + "image_footnote": [], + "bbox": [ + 191, + 760, + 285, + 833 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8af871797106e7d04a37d6a51e043db3a524b0c63658f0b1b470bf813e5c0da9.jpg", + "image_caption": [ + "Blur Modeling" + ], + "image_footnote": [], + "bbox": [ + 289, + 760, + 382, + 833 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a5c457213a92274fc2c19b1f534fc660d679d215ceaf394fd0d9099bdf8009e.jpg", + "image_caption": [ + "Reconstructed Output" + ], + "image_footnote": [], + "bbox": [ + 385, + 760, + 480, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of 3D Robust Optimization. Fig. 9 showcases visual comparison with or without 3D robust optimization under corrupted training views. It is observed that 3D robust optimization can effectively avoid unfavorable or erroneous supervision to optimize towards higher quality whereas its absence results in Gaussian primitives simulating corrupted representations.", + "bbox": [ + 511, + 289, + 906, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13b1add5f6b4ed73179bfd78eecd1cd9db316bf870b146973bd34a7fc6492ccd.jpg", + "image_caption": [ + "Figure 9. Visual comparison with or without 3D robust optimization (3DRO) under corrupted training views." + ], + "image_footnote": [], + "bbox": [ + 517, + 407, + 702, + 483 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4527f81b8a4cac880d10897f73e490148498549bd80caa799a3c42425409c09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 407, + 903, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Concluding Remarks", + "text_level": 1, + "bbox": [ + 511, + 536, + 718, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose S2Gaussian, an innovative framework that can efficiently reconstruct high-quality 3D scenes with only sparse and low-resolution input views. In contrast to existing frameworks, S2Gaussian can handle both view sparse and clarity deficient simultaneously, which is more functional and practical in real-world scenarios. The proposed framework begins by optimizing a low-resolution Gaussian representation, which is then densified to initialize high-resolution Gaussians via tailored Gaussian Shuffle Split operation. Subsequently, it cleverly solves the view sparsity problem by rendering pseudo-views from low-resolution Gaussians for super-resolution supervision. Wherein a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are further proposed to mitigate multi-view inconsistency and eliminate erroneous updates for better geometric structure and fine-grained details reconstruction. Extensive experiments on multiple benchmarks manifest the effectiveness, superiority, and robustness of our method. We expect this work to provide insights into more complicated 3D reconstruction and steer future research on this Gordian knot.", + "bbox": [ + 511, + 561, + 906, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "718", + "bbox": [ + 486, + 944, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors are very indebted to the anonymous referees for their critical comments and suggestions for the improvement of this paper. This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285 and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260).", + "bbox": [ + 89, + 116, + 485, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 253, + 187, + 271 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021.", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022.", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023.", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022.", + "[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024.", + "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In *Forty-first International Conference on Machine Learning*, 2024.", + "[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splattering in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024.", + "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022.", + "[9] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Zhenzhong Kuang, Yu Jun, Jianping Fan, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024.", + "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022." + ], + "bbox": [ + 93, + 280, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021.", + "[12] Yuqi Han, Tao Yu, Xiaohang Yu, Yuwang Wang, and Qionghai Dai. Super-nerf: View-consistent detail generation for nerf super-resolution. arXiv preprint arXiv:2304.13518, 2023.", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.", + "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023.", + "[16] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023.", + "[17] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008.", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023.", + "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023.", + "[20] Hyun-kyu Ko, Dongheok Park, Youngin Park, Byeonghyeon Lee, Juhee Han, and Eunbyung Park. Sequence matters: Harnessing video models in super-resolution. arXiv preprint arXiv:2412.11525, 2024.", + "[21] Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20561–20570, 2024.", + "[22] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024.", + "[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "719", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024.", + "[24] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019.", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021.", + "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022.", + "[27] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022.", + "[28] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022.", + "[29] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021.", + "[30] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023.", + "[31] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024.", + "[32] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.", + "[33] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022.", + "[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022.", + "[35] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065–9076, 2023.", + "[36] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024.", + "[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024.", + "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.", + "[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024.", + "[40] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 $\\{\\backslash \\mathrm{deg}\\}$ sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023.", + "[41] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024.", + "[42] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023.", + "[43] Youngho Yoon and Kuk-Jin Yoon. Cross-guided optimization of radiance fields with multi-view image superresolution for high-resolution novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12428-12438, 2023.", + "[44] Xiqian Yu, Hanxin Zhu, Tianyu He, and Zhibo Chen. Gaussiansr: 3d gaussian super-resolution with 2d diffusion priors. arXiv preprint arXiv:2406.10111, 2024.", + "[45] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024.", + "[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024.", + "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018.", + "[48] Xiaoyun Zheng, Liwei Liao, Jianbo Jiao, Feng Gao, and Ronggang Wang. Surface-sos: Self-supervised object segmentation via neural surface representation. IEEE Transactions on Image Processing, 2024.", + "[49] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "720", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021.", + "[50] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025." + ], + "bbox": [ + 91, + 90, + 483, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "721", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_model.json b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1a4f042281aeedb6ab2545dbe36f70cad869aeb9 --- /dev/null +++ b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.18, + 0.791, + 0.2 + ], + "angle": 0, + "content": "Yecong Wan\\(^{1}\\) Mingwen Shao\\(^{1*}\\) Yuanshuo Cheng\\(^{1}\\) Wangmeng Zuo\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.2, + 0.898, + 0.234 + ], + "angle": 0, + "content": "1Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software, Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China)" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.234, + 0.717, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Faculty of Computing, Harbin Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.253, + 0.796, + 0.271 + ], + "angle": 0, + "content": "smw278@126.com, {yecongwan, cys1294414023, cswmzuo} \\(@\\) gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.304, + 0.62, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.312, + 0.902, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.459, + 0.908, + 0.503 + ], + "angle": 0, + "content": "Figure 1. We present S2Gaussian, a novel framework capable of reconstructing high-quality 3D scenes for immersive rendering with only sparse and low-resolution input views. S2Gaussian demonstrates superior performance and yields high-fidelity and high-resolution scene reconstruction with sharp geometric and detailed textures, thus enjoying better functionality and practicality in realistic applications." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.515, + 0.327, + 0.53 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.486, + 0.875 + ], + "angle": 0, + "content": "In this paper, we aim ambitiously for a realistic yet challenging problem, namely, how to reconstruct high-quality 3D scenes from sparse low-resolution views that simultaneously suffer from deficient perspectives and clarity. Whereas existing methods only deal with either sparse views or low-resolution observations, they fail to handle such hybrid and complicated scenarios. To this end, we propose a novel Sparse-view Super-resolution 3D Gaussian Splatting framework, dubbed S2Gaussian, that can reconstruct structure-accurate and detail-faithful 3D scenes with only sparse and low-resolution views. The S2Gaussian operates in a two-stage fashion. In the first stage, we initially optimize a low-resolution Gaussian representation with depth regularization and densify it to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation. In the second stage, we refine the high-resolution Gaussians with the super-resolved images generated from both original sparse views and pseudoviews rendered by the low-resolution Gaussians. In which a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are elaborately designed to mitigate multi-view inconsistency and eliminate error" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.909, + 0.592 + ], + "angle": 0, + "content": "neous updates caused by imperfect supervision. Extensive experiments demonstrate superior results and in particular establishing new state-of-the-art performances with more consistent geometry and finer details. Project Page https://jeasco.github.io/S2Gaussian/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.604, + 0.645, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.854 + ], + "angle": 0, + "content": "The rapid advancement of virtual reality and metaverse technologies has significantly heightened the demand for realistic 3D scene reconstruction, which holds vast potential across various application domains, including medicine, education, entertainment, etc. Among the recent notable developments, Neural Radiance Fields (NeRF) [25] have demonstrated remarkable capability in synthesizing photorealistic images; however, they still face considerable rendering and training costs, despite ongoing improvements [4, 10, 11, 26]. More recently, 3D Gaussian Splatting (3DGS) [18] has emerged as a prominent approach, characterized by its high quality, rapid reconstruction speed, and support for real-time rendering. Subsequent research efforts [5, 6, 23, 32, 39, 45] have focused on extending the applicability of 3DGS across various scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Unfortunately, these methods typically rely heavily on well-captured dense and high-resolution images for impressive novel view synthesis, which is cumbersome and some" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.253, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\) * Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "711" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.379 + ], + "angle": 0, + "content": "times impractical in real-world applications. To address this challenge, a variety of approaches have emerged that focus on reconstructing 3D scenes from sparse [7, 8, 22, 27, 35, 40, 42, 50] and low-resolution views [9, 12, 21, 31, 43, 44], yielding worth-celebrating successes. However, these two obstacles, i.e., sparsity and low resolution, have long been treated as isolated problems and tackled separately, whereas input views in practical applications often suffer from both deficient perspectives and clarity. Particularly in contexts such as robotics and internet-collected imagery, where the available views are usually sparse and compounded by low resolutions due to environmental constraints and hardware transfer limitations. This interplay of sparsity and low-resolution poses significant challenges for existing reconstruction frameworks. Furthermore, merely incorporating existing methods remains subject to inherent incompatibility, as super-resolution necessitates dense supervision to recover fine details while sparse-view regularization not only fails to introduce details but tends to lead smoothing." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.382, + 0.483, + 0.775 + ], + "angle": 0, + "content": "To tackle the aforementioned problems, in this paper, we propose a novel 3D reconstruction framework S2Gaussian, capable of reconstructing geometrically precise and richly detailed 3D scenes with only sparse and low-resolution views. Specifically, S2Gaussian consists of two main phases: the high-resolution Gaussians (HR GS) initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the estimated depth information. The optimized low-resolution Gaussians are then utilized to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation which can provide more compact Gaussian primitives to facilitate fine-grained details reconstruction in high-resolution scenes. In the HR GS optimization stage, we use super-resolved images generated by pretrained super-resolution model to refine the high-resolution Gaussians, leveraging both the original sparse views and the pseudo-views rendered from the low-resolution Gaussians. Notably, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are introduced to mitigate the impacts of multi-view inconsistency and imperfect supervision, ultimately leading to more detailed and higher-quality scene reconstruction. Comprehensive experiments substantiate the superiority of our S2Gaussian beyond alternatives on a variety of benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.778, + 0.457, + 0.792 + ], + "angle": 0, + "content": "In conclusion, the main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We propose an innovative two-stage framework, termed S2Gaussian, designed to reconstruct 3D scenes that are both structurally accurate and rich in detail using only sparse and low-resolution input views." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- A dedicated Gaussian Shuffle Split operation is designed to initialize more compact Gaussian primitives for representing fine-grained details and textures in high" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.093, + 0.652, + 0.106 + ], + "angle": 0, + "content": "resolution scenes." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.107, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- We introduce a tailored blur-free inconsistency modeling scheme alongside a 3D robust optimization strategy to address multi-view inconsistencies and rectify erroneous updates arising from imperfect supervision." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.168, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- Our S2Gaussian reconstructs more fine-grained and high-quality Gaussian representations significantly outperforming existing methods across multiple benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.107, + 0.905, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.885, + 0.267 + ], + "angle": 0, + "content": "2.1. Novel View Synthesis using Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.272, + 0.908, + 0.694 + ], + "angle": 0, + "content": "Techniques for novel view synthesis typically involve learning a 3D representation using a limited number of input views and generating images from arbitrary novel perspectives. Recent advancements in Neural Radiance Fields (NeRF) [25] has demonstrated encouraging progress in this field, which learns an implicit neural scene representation for novel view synthesis via coordinate-based neural networks and volume rendering function. Subsequent research has focused extensively on enhancing NeRF's rendering quality [1-3, 33], improving efficiency [4, 10, 11, 26], advancing scene understanding [19, 48, 49], and facilitating 3D content generation [15, 28, 30]. Unfortunately, while NeRF significantly improves the quality of novel view rendering, its expensive training time as well as slow rendering speed hinder broader practical applications. Recently, Kerbl et al. [18] introduced a groundbreaking approach with their 3D Gaussian Splatting (3DGS) method, boosting rendering efficiency by utilizing explicit Gaussian representations combined with differentiable rasterization techniques. Building upon 3D Gaussian Splatting representation, tremendous following efforts [6, 23, 32, 39, 45] concentrate extending 3DGS under various scenarios. Nevertheless, these methods typically require dense and high-quality input views for impressive novel view synthesis, and they will tend to overfit the training data and compromise representations when provided with only sparse or degraded views, leading to a substantial decline in performance for the intended scene reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.7, + 0.771, + 0.717 + ], + "angle": 0, + "content": "2.2. Sparse Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To enhance the availability in practical usage, numerous algorithms have been developed to surmount the reliance on dense perspectives. For instance, Depth-NeRF [8] and Reg-NeRF [27] advocate utilizing depth constraints to improve the accuracy of the sparse scene representation. FreeN-erF [42], on the other hand, employs a frequency regularization strategy to manage the frequency range and penalize density fields near the camera. Additionally, SparseN-erF [35] leverages depth priors from inaccurate observations, implementing local depth ranking and spatial continuity constraints to bolster neural radiance fields. More recently, a variety of 3DGS-based approaches have emerged" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "712" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "[7, 22, 40, 50], paving the way for new blueprints in the field. These methods typically focus on various strategies that utilize depth prior to constrain the optimization of Gaussian primitives, such as SfM depth alignment [7], Pearson correlation depth distribution loss [50], global-local depth normalization [22], and patch-based depth correlation [40]. Unlike these approaches, we spotlight more arduous sparse reconstruction under low-resolution conditions, where available information is more scarce." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.428, + 0.256 + ], + "angle": 0, + "content": "2.3. Super-Resolution Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.548 + ], + "angle": 0, + "content": "Super-resolution novel view synthesis aims to reconstruct high-resolution 3D scenes utilizing only low-resolution multi-view inputs. Unlike anti-aliasing methods [1, 2, 45] which are incapable of rendering intricate details of the novel views and are limited by the input image's information level, this field is more focused on fine-grained and detailed high-resolution rendering of potential high-resolution scenes. As a pioneer in the field, NeRF-SR [34] optimizes high-resolution NeRF through a super-sampling strategy, ensuring that the values of low-resolution pixels match the mean value of high-resolution sub-pixels. Follow-up works are designed to utilize either high-resolution reference images [16] or pre-trained 2D models [12, 21, 43] to generate multi-view consistent details. Concurrently, SRGS [9] and GaussianSR [44] are proposed to utilize 2D super-resolution models for texture injection or diffusion prior exploration, respectively. SuperGaussian [31] and SM [20] further prove two profile frameworks to repurpose video upsampling models to achieve 3D super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Different from the aforementioned methods, in this work, we ambitiously tackle the challenge of joint sparse-view and super-resolution novel view synthesis by proposing a novel two-stage framework, which cleverly solves the problem of perspectives and clarity deficient to reconstruct geometry-accurate and fine-detailed 3D scenes, making it more desirable and feasible in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.225, + 0.689 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.231, + 0.711 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.763 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) as an emerged powerful technique for novel view synthesis, models a 3D scene using a collection of \\(G\\) colored Gaussians:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.765, + 0.482, + 0.794 + ], + "angle": 0, + "content": "\\[\ng _ {i} (x) = \\exp \\left(- \\frac {1}{2} \\left(x - \\mu_ {i}\\right) ^ {\\top} \\Sigma_ {i} ^ {- 1} \\left(x - \\mu_ {i}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(1 \\leq i \\leq G\\), \\(\\mu_i \\in \\mathbb{R}^3\\) represents the mean or center of the Gaussian, and \\(\\Sigma_i \\in \\mathbb{R}^{3 \\times 3}\\) denotes Gaussian's covariance matrix, specifying its shape and size, which is described by scaling vector \\(s \\in \\mathbb{R}_+^3\\) and rotation vector \\(q \\in \\mathbb{R}^4\\) for implementation. Additionally, each Gaussian is characterized by an opacity \\(\\sigma_i \\in \\mathbb{R}_+\\) and a view-dependent color \\(c_i(v) \\in \\mathbb{R}^{k_c}\\) (where \\(k_c\\) is changed) for rasterization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "To faithfully reconstruct the 3D scene, 3DGS also employs densification operation, i.e., the split, clone, and prune operation, to adaptively regulate the number and density of Gaussian primitives. For simplicity, we deploy the same adaptive density control as the original 3DGS and will not elaborate further in the subsequent presentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.188, + 0.719, + 0.202 + ], + "angle": 0, + "content": "3.2. S2Gaussian Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.905, + 0.491 + ], + "angle": 0, + "content": "The schematic illustration of the proposed S2Gaussian is depicted in Fig. 2. S2Gaussian is primarily composed of two phases i.e., the HR GS initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the predicted depth information, which is further utilized to initialize the high-resolution Gaussian representation by densifying and detailing the Gaussian primitives through our tailored Gaussian Shuffle Split operation. In the HR GS optimization stage, we further optimize the high-resolution Gaussians with the super-resolved images generated by pre-trained super-resolution model from both original sparse views and the pseudo-views rendered by LR Gaussians. In particular, a dedicated blur-free inconsistency modeling scheme and a 3D robust optimization strategy are advanced to eliminate the effects of multi-view inconsistency and imperfect supervision to reconstruct more detailed and high-quality 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.776, + 0.517 + ], + "angle": 0, + "content": "3.3. Stage 1: HR GS Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Directly reconstructing a high-resolution 3D scene from sparse low-resolution viewpoints is extremely challenging, as it is not only prone to overfitting specific viewpoints but also lacks high-resolution detail constraints. To alleviate such a reconstruction dilemma, we advocate initially reconstructing a low-resolution Gaussian representation as an auxiliary to facilitate the optimization of high-resolution Gaussians. On the one hand, the low-resolution Gaussian can be used to initialize high-resolution Gaussians to provide reliable base Gaussian ellipsoids and 3D scene structures. On the other hand, it can provide pseudo-views as additional supervision to avoid overfitting on specific viewpoints as well as provide more potential details and textures for high-resolution scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Concretely, we first obtain the monocular depth maps for the sparse low-resolution training views by utilizing the handy pre-trained depth estimation model [29]. Then we utilize the RGB images as well as the depth maps to co-optimize a low-resolution Gaussian representation with mature depth regularization technologies [22, 50], where Pearson correlation loss [50] is employed in this paper as a baseline. Thereafter, a tailored Gaussian Shuffle Split regime is proposed to densify and detail the sparse Gaussian primitives so as to initialize HR Gaussians for better high-resolution fine-grained details representation." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.912, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.318, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Overview of S2Gaussian. The S2Gaussian initially optimizes an LR GS and densifies it to initialize the HR GS through a tailored Gaussian Shuffle Split operation. Then, the original sparse views along with the pseudo views rendered by the LR GS are super-resolved together to refine the high-resolution texture with 3D robust optimization. In which an inconsistency modeling module (IM) and a blur proposal module are incorporated to mitigate inconsistency and blurriness, aiming to create 3D scenes with high-fidelity texture details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.389, + 0.486, + 0.63 + ], + "angle": 0, + "content": "Gaussian Shuffle Split. A precise and detailed initialization is crucial for reconstructing scene details [18]. Although the structure and layout of the entire 3D scene are basically guaranteed after the optimization of the low-resolution views, the initialized sparse and coarse Gaussian ellipsoids struggle to simulate high-resolution details which require denser Gaussians for finer representation [41]. Even equipped with adaptive density control [18], it falls short in our settings as there are neither dense perspectives nor high-quality details for supervision. To address this issue, we propose a training-free local Gaussian densification strategy Gaussian Shuffle Split which can provide more mobilizable Gaussian primitives by replacing the original large Gaussian with six small Gaussians to facilitate more comprehensive simulation of fine-grained details and textures in high-resolution scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.754 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, given a Gaussian primitive with attributes \\(\\{\\mu, s, q, \\sigma, c\\}\\), we first generate six duplicates, each center shifted along one of the six directions aligned with the original Gaussian's three principal axes, i.e., the positive and negative directions of each axis. The shift for each axis is set to \\(\\alpha\\) (0.5 by default) times the corresponding axis's scaling \\(s = [s_1, s_2, s_3]\\) value. Thus the new position of the six sub-Gaussians can be obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.754, + 0.484, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mu_ {s u b 1} = \\mu + R \\otimes [ \\alpha s _ {1}, 0, 0 ] ^ {T}, \\mu_ {s u b 2} = \\mu + R \\otimes [ - \\alpha s _ {1}, 0, 0 ] ^ {T},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.792 + ], + "angle": 0, + "content": "\\[\n\\mu_ {s u b 3} = \\mu + R \\otimes [ 0, \\alpha s _ {2}, 0 ] ^ {T}, \\mu_ {s u b 4} = \\mu + R \\otimes [ 0, - \\alpha s _ {2}, 0 ] ^ {T},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.795, + 0.484, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mu_ {s u b 5} = \\mu + R \\otimes [ 0, 0, \\alpha s _ {3} ] ^ {T}, \\mu_ {s u b 6} = \\mu + R \\otimes [ 0, 0, - \\alpha s _ {3} ] ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( R \\) represents the rotation matrix calculated from \\( q \\) to align with the original rotation of the Gaussian. Additionally, for each shifted sub-Gaussian, we scale their \\( s \\) value corresponding to the offset axis to \\( \\frac{1}{4} \\) of its original value, i.e., reducing the scale along that axis to shrink the size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.418 + ], + "angle": 0, + "content": "The remaining two \\(s\\) values are scaled by an another factor of \\(\\lambda\\) relative to their original values:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.42, + 0.832, + 0.437 + ], + "angle": 0, + "content": "\\[\ns _ {s u b 1} = s _ {s u b 2} = \\left[ s _ {1} / 4, s _ {2} / \\lambda , s _ {3} / \\lambda \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.439, + 0.905, + 0.456 + ], + "angle": 0, + "content": "\\[\ns _ {s u b 3} = s _ {s u b 4} = \\left[ s _ {1} / \\lambda , s _ {2} / 4, s _ {3} / \\lambda \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.459, + 0.83, + 0.474 + ], + "angle": 0, + "content": "\\[\ns _ {s u b 5} = s _ {s u b 6} = \\left[ s _ {1} / \\lambda , s _ {2} / \\lambda , s _ {3} / 4 \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.596 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is experimentally set to 1.9 which we found this ratio ensures that the combined representation of the sub-Gaussians better approximate the original Gaussian, preserving the integrity of the 3D scene. All other attributes, i.e., rotation \\(q\\), opacity \\(\\alpha\\), and color \\(c\\) remain identical to the original Gaussian ellipsoid. The combined six sub-Gaussians are utilized to substitute the original large Gaussians for densified 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.718 + ], + "angle": 0, + "content": "We deployed Gaussian Shuffle Split only for Gaussian primitives with opacity greater than 0.5 since these are more likely to be responsible for representing object surfaces and key structures which require a more detailed representation in high-resolution scenes. The opacity of all primitives is set close to zero after the Gaussian Shuffle Split to facilitate the automatic elimination of redundant Gaussians with low opacity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.78, + 0.744 + ], + "angle": 0, + "content": "3.4. Stage 2: HR GS Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "With the initialized high-resolution Gaussians, in this stage, we aim to continue optimizing it with the super-resolved images generated from both original sparse views and the pseudo-views rendered by the LR Gaussians. We synthesize pseudo views by interpolating virtual views between two known views as in [50]. Considering the prohibitively expensive data costs associated with 3D super-resolution models and their computational complexity, we advocate for utilizing off-the-shelf pre-trained 2D super-resolution models to enhance low-resolution views. Unfortunately, directly" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.288, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.09, + 0.474, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.172, + 0.284, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.172, + 0.476, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.254, + 0.283, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.254, + 0.478, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.422 + ], + "angle": 0, + "content": "Figure 3. Illustration of Gaussian Shuffle Split that utilizes a combination of six compact Gaussians to replace the original large Gaussian primitive, which can densify the 3D scene virtually without damaging the original 3D representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.536 + ], + "angle": 0, + "content": "leveraging 2D super-resolution model cannot ensure multiview consistency, and the low-resolution pseudo-views inevitably exhibit artifacts since there are areas that have not been optimized before. The Gaussian primitives, in their attempt to accommodate these inconsistencies and erroneous representations, will lead to the optimized 3D scene converging toward inaccurate structures and blurriness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.627 + ], + "angle": 0, + "content": "During optimization, two crucial aspects contribute fundamentally to these problems, i.e., inconsistent supervision and erroneous updates caused by artifacts. Accordingly, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are proposed in the following to alleviate potential inaccurate reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Blur-Free Inconsistency Modeling. To mitigate the inconsistency effort of the single-image super-resolution model, we employ a learnable inconsistency modeling module (IM), i.e., two residual blocks [13], after the pre-trained super-resolution model to simulate the inconsistency across different views \\( I_{SR}^{IM} = I_{SR} + IM(I_{SR}) \\), thus avoiding making Gaussians to represent such inconsistency. However, we empirically found that this module tends to lose details during optimization to win better consistency, resulting in the loosing of sharp textures. Therefore, we further propose a blur proposal module (BP) to blur the rendered image which is then constrained with modified super-resolution image \\( I_{SR}^{IM} \\), thereby compensating for the loss of fine-grained details and textures caused by IM. The blur proposal module, i.e., a four-layer convolutional network, predicts per-pixel blur kernels \\( \\mathcal{B}_k \\in \\mathbb{R}^{K \\times K \\times H \\times W} \\) (\\( K \\) is the blur kernel size, 5 for \\( 4 \\times \\) task) with the gradient detached rendered image \\( R_{HR}^{\\mathrm{detach}} \\) as input \\( \\mathcal{B}_k = BP(R_{HR}^{\\mathrm{detach}}) \\). Then the blurred" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.195, + 0.905, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.391 + ], + "angle": 0, + "content": "Figure 4. Top: Visualization of the gradient variation under different quality views supervision and our proposed 3D robust optimization. Bottom: Illustration of the proposed 3D robust optimization strategy that takes the gradient of scaling \\( s \\) as an example." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.905, + 0.451 + ], + "angle": 0, + "content": "high-resolution image is obtained via \\( R_{HR}^{blur} = R_{HR} * \\mathcal{B}_k \\) which is utilized to compute loss with \\( I_{SR}^{IM} \\), avoiding the smoothing imposed by directly constraining \\( R_{HR} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.451, + 0.905, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S R} = (1 - \\beta) \\mathcal {L} _ {1} \\left(R _ {H R} ^ {\\text {b l u r}}, I _ {S R} ^ {\\text {I M}}\\right) + \\beta \\mathcal {L} _ {D - S S I M} \\left(R _ {H R} ^ {\\text {b l u r}}, I _ {S R} ^ {\\text {I M}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.481, + 0.875, + 0.496 + ], + "angle": 0, + "content": "where \\(\\beta\\) is set to 0.2 similar to the original 3DGS [18]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.693 + ], + "angle": 0, + "content": "3D Robust Optimization. Albeit the above scheme could surmount the detrimental effects of multi-view inconsistencies, the optimization of Gaussians still unavoidably suffers from improper supervision of pseudo-views where some areas are not sufficiently optimized. To this end, we empirically observed that optimizing Gaussian primitives utilizing normal high-quality views results in stable and overall consistent gradients (top left of Fig. 4) whereas corrupted views can lead to significant perturbation and gradient confusion (top middle of Fig. 4), ultimately resulting in blurry rendering. Intrigued by this observation, we aim to design a more robust optimization strategy to mitigate the fluctuations caused by imperfect supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.905, + 0.723 + ], + "angle": 0, + "content": "More specifically, we augment the canonical Gaussian primitives" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.725, + 0.905, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\{\\mu , s, q, \\sigma , c \\} \\cup \\{\\nabla \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "with an additional per-Gaussian attribute flag gradient \\(\\nabla \\in \\mathbb{R}^{11 + k_c}\\). Where \\(\\nabla\\) is used to record the gradient trends of other attributes of the Gaussian, i.e., \\(\\mu, s, q, \\sigma\\), and the changed \\(k_{c}\\) SH coefficients. As shown in Fig. 4, for each Gaussian in one optimization step, we first compute the cosine similarity between the current gradient for each attribute \\(g_i^c\\) (\\(i\\) indicate different attribute) and the corresponding flag gradient \\(\\nabla (g_i)\\) stored in \\(\\nabla\\). If the cosine similarity is greater than zero, indicating that the gradient update trends are aligned, the gradient \\(g_i^c\\) will be directly used to" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "715" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on Blender \\( \\times 4 \\) (8 views),LLFF \\( \\times 4 \\) (3 views),and Mip-NeRF360 \\( \\times 4 \\) (24 views). The best,second best, and third best entries are marked in red, orange, and yellow, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.13, + 0.912, + 0.297 + ], + "angle": 0, + "content": "
MethodBlender ×4 (8 views)LLFF ×4 (3 views)Mip-NeRF 360 ×4 (24 views)
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
NeRF-SR [34]12.410.7440.51593.799.280.2260.617169.5810.260.2690.628151.62
RegNeRF [27]20.680.8410.12941.2815.780.4320.422116.7417.280.4170.449105.45
SparseNeRF [35]21.320.8440.13642.6216.320.4460.428127.6417.380.4320.446109.27
3DGS [18]20.650.8310.14142.5813.010.2720.489138.5316.250.3420.474108.32
SRGS [9]22.640.8620.11636.2417.580.4720.291102.4318.370.4570.39791.62
Mip-Splatting [45]22.780.8680.11933.6315.650.4330.372114.0318.080.4480.42186.17
DNGaussian [22]20.970.8440.13541.1516.270.4300.346129.2617.210.4220.452112.22
FSGS [50]21.450.8540.13340.3618.350.5050.24797.0718.640.4580.39494.21
FSGS [50]+SRGS [9]22.680.8650.11435.4219.490.5900.17562.6319.830.5130.44884.12
S2Gaussian (Ours)24.190.8790.08928.4720.450.6540.13945.8922.050.6870.29643.51
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.482, + 0.344 + ], + "angle": 0, + "content": "optimize the parameter, and the flag gradient value \\(\\nabla (g_i)\\) will be updated to the average one:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.345, + 0.482, + 0.361 + ], + "angle": 0, + "content": "\\[\ng _ {i} ^ {c} = g _ {i} ^ {c}, \\nabla (g _ {i}) = (\\nabla (g _ {i}) + g _ {i} ^ {c}) / 2. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.422 + ], + "angle": 0, + "content": "Conversely, if the cosine similarity is less than or equal to zero, indicating potential disturbances, the current gradient will be scaled down by a factor of \\(\\epsilon\\) to slow the parameter updates, and \\(\\nabla (g_i)\\) will be updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.423, + 0.482, + 0.44 + ], + "angle": 0, + "content": "\\[\ng _ {i} ^ {c} = \\epsilon g _ {i} ^ {c}, \\nabla (g _ {i}) = (1 - \\epsilon) \\nabla (g _ {i}) + \\epsilon g _ {i} ^ {c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.546 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is set to 0.1 to attenuate potentially erroneous updates and preserve the ability to shift update trends. In practice, such strategy can be combined with any gradient-based optimizer, e.g., Adam, by simply passing the modified gradients to the respective optimizer. Notably, \\(\\nabla\\) will be eliminated after the optimization and will not affect the rendering speed and storage consumption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.627 + ], + "angle": 0, + "content": "As illustrated in the top right of Fig. 4, our proposed 3D robust optimization scheme can significantly improve the stability of Gaussian primitive optimization by eliminating erroneous updates and perturbations thus reconstructing more accurate and high-quality 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Total Objective. Apart from \\(\\mathcal{L}_{SR}\\), an auxiliary loss combined with total variation (TV) loss and sub-pixel constraint is also deployed to stabilize the training:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.678, + 0.482, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A U X} = \\mathcal {L} _ {T V} \\left(R _ {H R}\\right) + \\mathcal {L} _ {1} \\left(R _ {H R} \\downarrow , I _ {L R}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.725 + ], + "angle": 0, + "content": "where \\(\\downarrow\\) indicates area average downsampling and \\(I_{LR}\\) denotes the low-resolution images before super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.76 + ], + "angle": 0, + "content": "In summary, the final loss \\(\\mathcal{L}\\) for high-resolution Gaussians optimization is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.762, + 0.482, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {A U X} + \\mathcal {L} _ {S R}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.224, + 0.816 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Following prior literature [21, 44], we primarily experiment on \\(4 \\times\\) super-resolution tasks in the maintext since most real-world 2D super-resolution models [36, 46] are mainly trained on \\(\\times 4\\) data, more diagnostic experiments on other scaling factors are provided in the Suppl." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.313, + 0.731, + 0.329 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.535 + ], + "angle": 0, + "content": "We implemented S2Gaussian upon the original 3DGS repository and all experiments are conducted on a single RTX3090 GPU. For the optimization of LR GS, the pretrained DPT [29] model is adapted for depth estimation and the total iterations are set to 10,000. During the HR GS optimization stage, the Gaussian primitives are further optimized with additional 10,000 iterations. As for the off-the-shelf 2D super-resolution model, we opt for powerful diffusion-based ResShift [46] as our backbone. We conduct experiments on three benchmark datasets with SfM for initialization, advanced techniques like DUSt3R [37] can facilitate the initialization but that is not the main purpose of our current work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.632 + ], + "angle": 0, + "content": "Blender Datasets [25] contain 8 detailed objects synthesized by Blender with a resolution of \\(800 \\times 800\\). Each object initially includes 100 training images and 200 testing images. We uniformly sample 8 images from the original 100 training images and use the low-resolution of \\(200 \\times 200\\) pixels for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.906, + 0.727 + ], + "angle": 0, + "content": "LLFF Datasets [24] consist of 8 forward-facing real-world scenes. Following the previous setup, we select every eighth image as the test set, and evenly sample 3 sparse views from the remaining images for training. We leverage low-resolution images at \\(252 \\times 189\\) pixels for training and render at a higher resolution of \\(1008 \\times 756\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Mip-NeRF 360 Datasets [2] consist of 9 real-world scenes with 5 outdoors and 4 indoors. Similar to LLFF, we use 1/8 of the full view for testing and uniformly sample 24 images from the remaining views for training. We utilize the images with downsampling rate of 4 (about 1K resolution) as the target size and further downsample the training images by a factor 4 to obtain sparse low-resolution inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The following measures are used to evaluate the quantitative performance: Peak Signal-to-Noise Ratio (PSNR) [17], Structural Similarity Index (SSIM) [38], LPIPS (VGG) [47], and Fréchet Inception Distance (FID) [14]." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "716" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.216, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.276, + 0.206, + 0.288 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.351, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.276, + 0.315, + 0.288 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.486, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.276, + 0.479, + 0.29 + ], + "angle": 0, + "content": "Mip-Splatting" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.623, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.276, + 0.587, + 0.288 + ], + "angle": 0, + "content": "FSGS" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.089, + 0.759, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.276, + 0.747, + 0.288 + ], + "angle": 0, + "content": "FSGS+SRGS" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.089, + 0.895, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.276, + 0.857, + 0.288 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons on Blender \\(\\times 4\\) datasets with 8 input views. Our method produces more visually appealing results, successfully reconstructing intricate thin structures with fine-grained details. Best viewed at screen!" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.333, + 0.223, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.49, + 0.206, + 0.502 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.333, + 0.36, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.491, + 0.315, + 0.502 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.333, + 0.496, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.49, + 0.478, + 0.504 + ], + "angle": 0, + "content": "Mip-Splatting" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.333, + 0.632, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.49, + 0.587, + 0.502 + ], + "angle": 0, + "content": "FSGS" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.333, + 0.769, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.49, + 0.746, + 0.502 + ], + "angle": 0, + "content": "FSGS+SRGS" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.333, + 0.905, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.49, + 0.857, + 0.502 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.509, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons on LLFF \\(\\times 4\\) datasets with 24 input views. Our method still generates the most accurate super-resolution 3D scenes, preserving both photo-realistic texture and geometric-level details. Best viewed at screen!" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.551, + 0.455, + 0.568 + ], + "angle": 0, + "content": "4.2. Quantitative and Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.578, + 0.483, + 0.744 + ], + "angle": 0, + "content": "To rigorously validate the effectiveness of our proposed S2Gaussian, we conduct a comparative analysis against a range of prior methods, including NeRF-SR [34], RegNeRF [27], SparseNeRF [35], vanilla 3DGS [18], SRGS [9], Mip-Splatting [45], DNGaussian [22], and FSGS [50]. Besides, we also include an enhanced version of FSGS with SRGS to provide a standard upsampling baseline. While NeRF-SR and FSGS+SRGS employ super-resolution techniques directly, the remaining methods are trained on low-resolution input views and subsequently rendered at higher resolutions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Quantitative Evaluation. Tab. 1 presents quantitative comparison results for \\(4 \\times\\) sparse-view super-resolution novel view synthesis task on the Blender [25], LLFF [24], and Mip-NeRF 360 datasets [2]. It is observed that our S2Gaussian delivers remarkable performance gains and outperforms all competitive method significantly in terms of PSNR, SSIM, LPIPS, and FID metrics. Especially for Mip-NeRF 360 datasets, S2Gaussian exceeds the top-performing combinatorial method FSGS+SRGS by an impressive margin of \\(2.22\\mathrm{dB}\\) PSNR, underscoring its superior capability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.642 + ], + "angle": 0, + "content": "in reconstructing large, real-world scenes. In addition to delivering higher reconstruction fidelity, reflected by PSNR and SSIM, S2Gaussian also achieves substantial improvements in perceptual metrics LPIPS and FID, validating the effectiveness of S2Gaussian in converging towards greater perception quality with faithful and fidelity details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Qualitative Evaluation. We also demonstrate visual comparisons in Fig. 5, Fig. 6, and Fig. 7. As suggested, both 3DGS and Mip-Splatting struggle with the extremely sparse view problem, while FSGS is limited to characterizing high-resolution details. Despite the combinatorial method FSGS+SRGS can reconstruct a reasonable overall structure but suffers from inconsistency and sparsity leading to missing details and blurring. In comparison, S2Gaussian recovers more coherent fine structures with natural textures and harmonious global tone while other methods tend to introduce improper structures and unnatural details, especially for the extremely sparse scenes (Fig. 6) as well as large-scale scenarios (Fig. 7)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.681, + 0.863 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In Tab. 2, we conduct ablation experiments on the components introduced in S2Gaussian. The effectiveness of each" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.223, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.228, + 0.207, + 0.24 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.36, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.228, + 0.316, + 0.239 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.089, + 0.496, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.228, + 0.479, + 0.242 + ], + "angle": 0, + "content": "Mip-Splatting" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.633, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.228, + 0.588, + 0.239 + ], + "angle": 0, + "content": "FSGS" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.089, + 0.769, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.228, + 0.747, + 0.239 + ], + "angle": 0, + "content": "FSGS+SRGS" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.228, + 0.858, + 0.24 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.247, + 0.908, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons on Mip-NeRF360 \\(\\times 4\\) datasets with 24 input views. Our method consistently delivers more coherent fine structures than other methods in large-scale scenes. Best viewed at screen!" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.349 + ], + "angle": 0, + "content": "proposed component in S2Gaussian is evaluated by gradually integrating them into the model, revealing their individual contributions to the overall performance. More detailed analyses are as below." + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.353, + 0.466, + 0.367 + ], + "angle": 0, + "content": "Table 2. Ablation studies on Mip-NeRF \\( {360} \\times 4 \\) (24 views)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.368, + 0.481, + 0.456 + ], + "angle": 0, + "content": "
VariantPSNR↑SSIM↑LPIPS↓FID↓
Baseline19.790.5200.42987.22
+ Two-Stage20.060.5280.41580.64
+ Gaussian Shuffle Split20.620.5590.37671.29
+ Blur-Free Inconsistency Modeling21.120.6240.34659.63
+ 3D Robust Optimization22.050.6870.29643.51
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Effect of Gaussian Shuffle Split. As shown in Tab. 3, fewer sub-Gaussians lead to suboptimal detail modeling (lower PSNR), while 6 offers the best trade-off, while more sub-Gaussians will not only increase the optimization complexity but also introduce extra memory overhead. Additionally, 6 sub-Gaussians can be symmetrically located along the three principal axes of the original Gaussian, perfectly approximating the original Gaussian representation." + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.594, + 0.472, + 0.607 + ], + "angle": 0, + "content": "Table 3. Ablation of sub-Gaussian number on Mip-NeRF 360." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.608, + 0.48, + 0.658 + ], + "angle": 0, + "content": "
Number456789
Storage Memory (M)6789112137169202
Final PSNR21.7621.9522.0522.0322.0922.08
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Effect of Blur-Free Inconsistency Modeling. As illustrated in Fig. 8, IM models the inconsistency of super-resolved images, and while this would introduce blurring, we model this effect explicitly through blur kernel learning, thus reconstructing results that are blur-free and with fine structural details." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.761, + 0.189, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.834, + 0.182, + 0.852 + ], + "angle": 0, + "content": "Super-Resolved Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.761, + 0.287, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.835, + 0.274, + 0.852 + ], + "angle": 0, + "content": "Inconsistency Modeling" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.761, + 0.383, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.835, + 0.361, + 0.852 + ], + "angle": 0, + "content": "Blur Modeling" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.761, + 0.481, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.835, + 0.469, + 0.852 + ], + "angle": 0, + "content": "Reconstructed Output" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.898 + ], + "angle": 0, + "content": "Figure 8. Visualization of the super-resolved image, the residual output of IM, the blurred reconstructed image with predicted blur kernels, and the original reconstructed output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Effect of 3D Robust Optimization. Fig. 9 showcases visual comparison with or without 3D robust optimization under corrupted training views. It is observed that 3D robust optimization can effectively avoid unfavorable or erroneous supervision to optimize towards higher quality whereas its absence results in Gaussian primitives simulating corrupted representations." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.409, + 0.703, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.409, + 0.905, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.52 + ], + "angle": 0, + "content": "Figure 9. Visual comparison with or without 3D robust optimization (3DRO) under corrupted training views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.719, + 0.554 + ], + "angle": 0, + "content": "5. Concluding Remarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.879 + ], + "angle": 0, + "content": "In this paper, we propose S2Gaussian, an innovative framework that can efficiently reconstruct high-quality 3D scenes with only sparse and low-resolution input views. In contrast to existing frameworks, S2Gaussian can handle both view sparse and clarity deficient simultaneously, which is more functional and practical in real-world scenarios. The proposed framework begins by optimizing a low-resolution Gaussian representation, which is then densified to initialize high-resolution Gaussians via tailored Gaussian Shuffle Split operation. Subsequently, it cleverly solves the view sparsity problem by rendering pseudo-views from low-resolution Gaussians for super-resolution supervision. Wherein a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are further proposed to mitigate multi-view inconsistency and eliminate erroneous updates for better geometric structure and fine-grained details reconstruction. Extensive experiments on multiple benchmarks manifest the effectiveness, superiority, and robustness of our method. We expect this work to provide insights into more complicated 3D reconstruction and steer future research on this Gordian knot." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "718" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.239 + ], + "angle": 0, + "content": "The authors are very indebted to the anonymous referees for their critical comments and suggestions for the improvement of this paper. This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285 and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.255, + 0.188, + 0.272 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.281, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.367, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.439, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In *Forty-first International Conference on Machine Learning*, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.644, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splattering in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[9] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Zhenzhong Kuang, Yu Jun, Jianping Fan, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.281, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[11] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[12] Yuqi Han, Tao Yu, Xiaohang Yu, Yuwang Wang, and Qionghai Dai. Super-nerf: View-consistent detail generation for nerf super-resolution. arXiv preprint arXiv:2304.13518, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.908, + 0.417 + ], + "angle": 0, + "content": "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[16] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[17] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[20] Hyun-kyu Ko, Dongheok Park, Youngin Park, Byeonghyeon Lee, Juhee Han, and Eunbyung Park. Sequence matters: Harnessing video models in super-resolution. arXiv preprint arXiv:2412.11525, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[21] Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20561–20570, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[22] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "719" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[27] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[28] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[29] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[30] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[31] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[32] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[33] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[35] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065–9076, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[40] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 \\(\\{\\backslash \\mathrm{deg}\\}\\) sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[41] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[42] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[43] Youngho Yoon and Kuk-Jin Yoon. Cross-guided optimization of radiance fields with multi-view image superresolution for high-resolution novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12428-12438, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[44] Xiqian Yu, Hanxin Zhu, Tianyu He, and Zhibo Chen. Gaussiansr: 3d gaussian super-resolution with 2d diffusion priors. arXiv preprint arXiv:2406.10111, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[45] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Xiaoyun Zheng, Liwei Liao, Jianbo Jiao, Feng Gao, and Ronggang Wang. Surface-sos: Self-supervised object segmentation via neural surface representation. IEEE Transactions on Image Processing, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[50] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "721" + } + ] +] \ No newline at end of file diff --git a/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_origin.pdf b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3fe28d53a92321e5feb9e40d028f11024a94e82f --- /dev/null +++ b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/f341023e-9ff1-40bd-a8ab-93db55d08132_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15006ae7dc48e75d2fa0c98e0c02be4334713dc37f75a27497d147b49e473a77 +size 8883212 diff --git a/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/full.md b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e84f703e23546ba36102c64d2efb1671747e6408 --- /dev/null +++ b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/full.md @@ -0,0 +1,390 @@ +# S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splitting + +Yecong Wan $^{1}$ Mingwen Shao $^{1*}$ Yuanshuo Cheng $^{1}$ Wangmeng Zuo $^{2}$ + +1Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software, Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China) + +$^{2}$ Faculty of Computing, Harbin Institute of Technology + +smw278@126.com, {yecongwan, cys1294414023, cswmzuo} $@$ gmail.com + +![](images/630e80a6bae84f426c3564fed8b99b3652d7d081c6fea98f4b2eb133a9c3eca0.jpg) +Figure 1. We present S2Gaussian, a novel framework capable of reconstructing high-quality 3D scenes for immersive rendering with only sparse and low-resolution input views. S2Gaussian demonstrates superior performance and yields high-fidelity and high-resolution scene reconstruction with sharp geometric and detailed textures, thus enjoying better functionality and practicality in realistic applications. + +![](images/06238b75f28db2475544560de0f253bf0587c6b4f521b2be8a701effcb220f03.jpg) + +# Abstract + +In this paper, we aim ambitiously for a realistic yet challenging problem, namely, how to reconstruct high-quality 3D scenes from sparse low-resolution views that simultaneously suffer from deficient perspectives and clarity. Whereas existing methods only deal with either sparse views or low-resolution observations, they fail to handle such hybrid and complicated scenarios. To this end, we propose a novel Sparse-view Super-resolution 3D Gaussian Splatting framework, dubbed S2Gaussian, that can reconstruct structure-accurate and detail-faithful 3D scenes with only sparse and low-resolution views. The S2Gaussian operates in a two-stage fashion. In the first stage, we initially optimize a low-resolution Gaussian representation with depth regularization and densify it to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation. In the second stage, we refine the high-resolution Gaussians with the super-resolved images generated from both original sparse views and pseudoviews rendered by the low-resolution Gaussians. In which a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are elaborately designed to mitigate multi-view inconsistency and eliminate error + +neous updates caused by imperfect supervision. Extensive experiments demonstrate superior results and in particular establishing new state-of-the-art performances with more consistent geometry and finer details. Project Page https://jeasco.github.io/S2Gaussian/. + +# 1. Introduction + +The rapid advancement of virtual reality and metaverse technologies has significantly heightened the demand for realistic 3D scene reconstruction, which holds vast potential across various application domains, including medicine, education, entertainment, etc. Among the recent notable developments, Neural Radiance Fields (NeRF) [25] have demonstrated remarkable capability in synthesizing photorealistic images; however, they still face considerable rendering and training costs, despite ongoing improvements [4, 10, 11, 26]. More recently, 3D Gaussian Splatting (3DGS) [18] has emerged as a prominent approach, characterized by its high quality, rapid reconstruction speed, and support for real-time rendering. Subsequent research efforts [5, 6, 23, 32, 39, 45] have focused on extending the applicability of 3DGS across various scenarios. + +Unfortunately, these methods typically rely heavily on well-captured dense and high-resolution images for impressive novel view synthesis, which is cumbersome and some + +times impractical in real-world applications. To address this challenge, a variety of approaches have emerged that focus on reconstructing 3D scenes from sparse [7, 8, 22, 27, 35, 40, 42, 50] and low-resolution views [9, 12, 21, 31, 43, 44], yielding worth-celebrating successes. However, these two obstacles, i.e., sparsity and low resolution, have long been treated as isolated problems and tackled separately, whereas input views in practical applications often suffer from both deficient perspectives and clarity. Particularly in contexts such as robotics and internet-collected imagery, where the available views are usually sparse and compounded by low resolutions due to environmental constraints and hardware transfer limitations. This interplay of sparsity and low-resolution poses significant challenges for existing reconstruction frameworks. Furthermore, merely incorporating existing methods remains subject to inherent incompatibility, as super-resolution necessitates dense supervision to recover fine details while sparse-view regularization not only fails to introduce details but tends to lead smoothing. + +To tackle the aforementioned problems, in this paper, we propose a novel 3D reconstruction framework S2Gaussian, capable of reconstructing geometrically precise and richly detailed 3D scenes with only sparse and low-resolution views. Specifically, S2Gaussian consists of two main phases: the high-resolution Gaussians (HR GS) initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the estimated depth information. The optimized low-resolution Gaussians are then utilized to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation which can provide more compact Gaussian primitives to facilitate fine-grained details reconstruction in high-resolution scenes. In the HR GS optimization stage, we use super-resolved images generated by pretrained super-resolution model to refine the high-resolution Gaussians, leveraging both the original sparse views and the pseudo-views rendered from the low-resolution Gaussians. Notably, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are introduced to mitigate the impacts of multi-view inconsistency and imperfect supervision, ultimately leading to more detailed and higher-quality scene reconstruction. Comprehensive experiments substantiate the superiority of our S2Gaussian beyond alternatives on a variety of benchmarks. + +In conclusion, the main contributions are as follows: + +- We propose an innovative two-stage framework, termed S2Gaussian, designed to reconstruct 3D scenes that are both structurally accurate and rich in detail using only sparse and low-resolution input views. +- A dedicated Gaussian Shuffle Split operation is designed to initialize more compact Gaussian primitives for representing fine-grained details and textures in high + +resolution scenes. + +- We introduce a tailored blur-free inconsistency modeling scheme alongside a 3D robust optimization strategy to address multi-view inconsistencies and rectify erroneous updates arising from imperfect supervision. +- Our S2Gaussian reconstructs more fine-grained and high-quality Gaussian representations significantly outperforming existing methods across multiple benchmarks. + +# 2. Related Work + +# 2.1. Novel View Synthesis using Radiance Fields + +Techniques for novel view synthesis typically involve learning a 3D representation using a limited number of input views and generating images from arbitrary novel perspectives. Recent advancements in Neural Radiance Fields (NeRF) [25] has demonstrated encouraging progress in this field, which learns an implicit neural scene representation for novel view synthesis via coordinate-based neural networks and volume rendering function. Subsequent research has focused extensively on enhancing NeRF's rendering quality [1-3, 33], improving efficiency [4, 10, 11, 26], advancing scene understanding [19, 48, 49], and facilitating 3D content generation [15, 28, 30]. Unfortunately, while NeRF significantly improves the quality of novel view rendering, its expensive training time as well as slow rendering speed hinder broader practical applications. Recently, Kerbl et al. [18] introduced a groundbreaking approach with their 3D Gaussian Splatting (3DGS) method, boosting rendering efficiency by utilizing explicit Gaussian representations combined with differentiable rasterization techniques. Building upon 3D Gaussian Splatting representation, tremendous following efforts [6, 23, 32, 39, 45] concentrate extending 3DGS under various scenarios. Nevertheless, these methods typically require dense and high-quality input views for impressive novel view synthesis, and they will tend to overfit the training data and compromise representations when provided with only sparse or degraded views, leading to a substantial decline in performance for the intended scene reconstruction. + +# 2.2. Sparse Novel View Synthesis + +To enhance the availability in practical usage, numerous algorithms have been developed to surmount the reliance on dense perspectives. For instance, Depth-NeRF [8] and Reg-NeRF [27] advocate utilizing depth constraints to improve the accuracy of the sparse scene representation. FreeN-erF [42], on the other hand, employs a frequency regularization strategy to manage the frequency range and penalize density fields near the camera. Additionally, SparseN-erF [35] leverages depth priors from inaccurate observations, implementing local depth ranking and spatial continuity constraints to bolster neural radiance fields. More recently, a variety of 3DGS-based approaches have emerged + +[7, 22, 40, 50], paving the way for new blueprints in the field. These methods typically focus on various strategies that utilize depth prior to constrain the optimization of Gaussian primitives, such as SfM depth alignment [7], Pearson correlation depth distribution loss [50], global-local depth normalization [22], and patch-based depth correlation [40]. Unlike these approaches, we spotlight more arduous sparse reconstruction under low-resolution conditions, where available information is more scarce. + +# 2.3. Super-Resolution Novel View Synthesis + +Super-resolution novel view synthesis aims to reconstruct high-resolution 3D scenes utilizing only low-resolution multi-view inputs. Unlike anti-aliasing methods [1, 2, 45] which are incapable of rendering intricate details of the novel views and are limited by the input image's information level, this field is more focused on fine-grained and detailed high-resolution rendering of potential high-resolution scenes. As a pioneer in the field, NeRF-SR [34] optimizes high-resolution NeRF through a super-sampling strategy, ensuring that the values of low-resolution pixels match the mean value of high-resolution sub-pixels. Follow-up works are designed to utilize either high-resolution reference images [16] or pre-trained 2D models [12, 21, 43] to generate multi-view consistent details. Concurrently, SRGS [9] and GaussianSR [44] are proposed to utilize 2D super-resolution models for texture injection or diffusion prior exploration, respectively. SuperGaussian [31] and SM [20] further prove two profile frameworks to repurpose video upsampling models to achieve 3D super-resolution. + +Different from the aforementioned methods, in this work, we ambitiously tackle the challenge of joint sparse-view and super-resolution novel view synthesis by proposing a novel two-stage framework, which cleverly solves the problem of perspectives and clarity deficient to reconstruct geometry-accurate and fine-detailed 3D scenes, making it more desirable and feasible in real-world applications. + +# 3. Methodology + +# 3.1. Preliminaries + +3D Gaussian Splatting (3DGS) as an emerged powerful technique for novel view synthesis, models a 3D scene using a collection of $G$ colored Gaussians: + +$$ +g _ {i} (x) = \exp \left(- \frac {1}{2} \left(x - \mu_ {i}\right) ^ {\top} \Sigma_ {i} ^ {- 1} \left(x - \mu_ {i}\right)\right), \tag {1} +$$ + +where $1 \leq i \leq G$ , $\mu_i \in \mathbb{R}^3$ represents the mean or center of the Gaussian, and $\Sigma_i \in \mathbb{R}^{3 \times 3}$ denotes Gaussian's covariance matrix, specifying its shape and size, which is described by scaling vector $s \in \mathbb{R}_+^3$ and rotation vector $q \in \mathbb{R}^4$ for implementation. Additionally, each Gaussian is characterized by an opacity $\sigma_i \in \mathbb{R}_+$ and a view-dependent color $c_i(v) \in \mathbb{R}^{k_c}$ (where $k_c$ is changed) for rasterization. + +To faithfully reconstruct the 3D scene, 3DGS also employs densification operation, i.e., the split, clone, and prune operation, to adaptively regulate the number and density of Gaussian primitives. For simplicity, we deploy the same adaptive density control as the original 3DGS and will not elaborate further in the subsequent presentation. + +# 3.2. S2Gaussian Overview + +The schematic illustration of the proposed S2Gaussian is depicted in Fig. 2. S2Gaussian is primarily composed of two phases i.e., the HR GS initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the predicted depth information, which is further utilized to initialize the high-resolution Gaussian representation by densifying and detailing the Gaussian primitives through our tailored Gaussian Shuffle Split operation. In the HR GS optimization stage, we further optimize the high-resolution Gaussians with the super-resolved images generated by pre-trained super-resolution model from both original sparse views and the pseudo-views rendered by LR Gaussians. In particular, a dedicated blur-free inconsistency modeling scheme and a 3D robust optimization strategy are advanced to eliminate the effects of multi-view inconsistency and imperfect supervision to reconstruct more detailed and high-quality 3D scenes. + +# 3.3. Stage 1: HR GS Initialization + +Directly reconstructing a high-resolution 3D scene from sparse low-resolution viewpoints is extremely challenging, as it is not only prone to overfitting specific viewpoints but also lacks high-resolution detail constraints. To alleviate such a reconstruction dilemma, we advocate initially reconstructing a low-resolution Gaussian representation as an auxiliary to facilitate the optimization of high-resolution Gaussians. On the one hand, the low-resolution Gaussian can be used to initialize high-resolution Gaussians to provide reliable base Gaussian ellipsoids and 3D scene structures. On the other hand, it can provide pseudo-views as additional supervision to avoid overfitting on specific viewpoints as well as provide more potential details and textures for high-resolution scenes. + +Concretely, we first obtain the monocular depth maps for the sparse low-resolution training views by utilizing the handy pre-trained depth estimation model [29]. Then we utilize the RGB images as well as the depth maps to co-optimize a low-resolution Gaussian representation with mature depth regularization technologies [22, 50], where Pearson correlation loss [50] is employed in this paper as a baseline. Thereafter, a tailored Gaussian Shuffle Split regime is proposed to densify and detail the sparse Gaussian primitives so as to initialize HR Gaussians for better high-resolution fine-grained details representation. + +![](images/424ae6196f961eb89564199daa0b6f9c870f8b18c80aa756da848c16ed02fbba.jpg) +Figure 2. Overview of S2Gaussian. The S2Gaussian initially optimizes an LR GS and densifies it to initialize the HR GS through a tailored Gaussian Shuffle Split operation. Then, the original sparse views along with the pseudo views rendered by the LR GS are super-resolved together to refine the high-resolution texture with 3D robust optimization. In which an inconsistency modeling module (IM) and a blur proposal module are incorporated to mitigate inconsistency and blurriness, aiming to create 3D scenes with high-fidelity texture details. + +Gaussian Shuffle Split. A precise and detailed initialization is crucial for reconstructing scene details [18]. Although the structure and layout of the entire 3D scene are basically guaranteed after the optimization of the low-resolution views, the initialized sparse and coarse Gaussian ellipsoids struggle to simulate high-resolution details which require denser Gaussians for finer representation [41]. Even equipped with adaptive density control [18], it falls short in our settings as there are neither dense perspectives nor high-quality details for supervision. To address this issue, we propose a training-free local Gaussian densification strategy Gaussian Shuffle Split which can provide more mobilizable Gaussian primitives by replacing the original large Gaussian with six small Gaussians to facilitate more comprehensive simulation of fine-grained details and textures in high-resolution scenes. + +As illustrated in Fig. 3, given a Gaussian primitive with attributes $\{\mu, s, q, \sigma, c\}$ , we first generate six duplicates, each center shifted along one of the six directions aligned with the original Gaussian's three principal axes, i.e., the positive and negative directions of each axis. The shift for each axis is set to $\alpha$ (0.5 by default) times the corresponding axis's scaling $s = [s_1, s_2, s_3]$ value. Thus the new position of the six sub-Gaussians can be obtained: + +$$ +\mu_ {s u b 1} = \mu + R \otimes [ \alpha s _ {1}, 0, 0 ] ^ {T}, \mu_ {s u b 2} = \mu + R \otimes [ - \alpha s _ {1}, 0, 0 ] ^ {T}, +$$ + +$$ +\mu_ {s u b 3} = \mu + R \otimes [ 0, \alpha s _ {2}, 0 ] ^ {T}, \mu_ {s u b 4} = \mu + R \otimes [ 0, - \alpha s _ {2}, 0 ] ^ {T}, +$$ + +$$ +\mu_ {s u b 5} = \mu + R \otimes [ 0, 0, \alpha s _ {3} ] ^ {T}, \mu_ {s u b 6} = \mu + R \otimes [ 0, 0, - \alpha s _ {3} ] ^ {T}, \tag {2} +$$ + +where $R$ represents the rotation matrix calculated from $q$ to align with the original rotation of the Gaussian. Additionally, for each shifted sub-Gaussian, we scale their $s$ value corresponding to the offset axis to $\frac{1}{4}$ of its original value, i.e., reducing the scale along that axis to shrink the size. + +The remaining two $s$ values are scaled by an another factor of $\lambda$ relative to their original values: + +$$ +s _ {s u b 1} = s _ {s u b 2} = \left[ s _ {1} / 4, s _ {2} / \lambda , s _ {3} / \lambda \right], +$$ + +$$ +s _ {s u b 3} = s _ {s u b 4} = \left[ s _ {1} / \lambda , s _ {2} / 4, s _ {3} / \lambda \right], \tag {3} +$$ + +$$ +s _ {s u b 5} = s _ {s u b 6} = \left[ s _ {1} / \lambda , s _ {2} / \lambda , s _ {3} / 4 \right], +$$ + +where $\lambda$ is experimentally set to 1.9 which we found this ratio ensures that the combined representation of the sub-Gaussians better approximate the original Gaussian, preserving the integrity of the 3D scene. All other attributes, i.e., rotation $q$ , opacity $\alpha$ , and color $c$ remain identical to the original Gaussian ellipsoid. The combined six sub-Gaussians are utilized to substitute the original large Gaussians for densified 3D representation. + +We deployed Gaussian Shuffle Split only for Gaussian primitives with opacity greater than 0.5 since these are more likely to be responsible for representing object surfaces and key structures which require a more detailed representation in high-resolution scenes. The opacity of all primitives is set close to zero after the Gaussian Shuffle Split to facilitate the automatic elimination of redundant Gaussians with low opacity. + +# 3.4. Stage 2: HR GS Optimization + +With the initialized high-resolution Gaussians, in this stage, we aim to continue optimizing it with the super-resolved images generated from both original sparse views and the pseudo-views rendered by the LR Gaussians. We synthesize pseudo views by interpolating virtual views between two known views as in [50]. Considering the prohibitively expensive data costs associated with 3D super-resolution models and their computational complexity, we advocate for utilizing off-the-shelf pre-trained 2D super-resolution models to enhance low-resolution views. Unfortunately, directly + +![](images/f8d1d65b005d2cf70f02cb48aac04cbcb9537458b39d5a23bad55934c96c644b.jpg) + +![](images/37b4a927806a721ce09fe26efd8781ac1131801c1e38c04c91b657546719425f.jpg) + +![](images/dc2bccd44a7fa650901abd438e94a4477771f78b0af386aed1255496e6b16189.jpg) + +![](images/210db062e178916ae853f4fa18e1883858beada1c679997cfb80b0dd6e6e02d3.jpg) + +![](images/ccbe3c2ca2f95c9a518b4c35466701dfe919d19b8fb450a4b7ee85dfa71770d2.jpg) +Figure 3. Illustration of Gaussian Shuffle Split that utilizes a combination of six compact Gaussians to replace the original large Gaussian primitive, which can densify the 3D scene virtually without damaging the original 3D representation. + +![](images/4327336c60bc6d1aebf227d0ea27a3beb11bf0ca073a5e3fe584d07520f274e8.jpg) +Figure 4. Top: Visualization of the gradient variation under different quality views supervision and our proposed 3D robust optimization. Bottom: Illustration of the proposed 3D robust optimization strategy that takes the gradient of scaling $s$ as an example. + +leveraging 2D super-resolution model cannot ensure multiview consistency, and the low-resolution pseudo-views inevitably exhibit artifacts since there are areas that have not been optimized before. The Gaussian primitives, in their attempt to accommodate these inconsistencies and erroneous representations, will lead to the optimized 3D scene converging toward inaccurate structures and blurriness. + +During optimization, two crucial aspects contribute fundamentally to these problems, i.e., inconsistent supervision and erroneous updates caused by artifacts. Accordingly, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are proposed in the following to alleviate potential inaccurate reconstruction. + +Blur-Free Inconsistency Modeling. To mitigate the inconsistency effort of the single-image super-resolution model, we employ a learnable inconsistency modeling module (IM), i.e., two residual blocks [13], after the pre-trained super-resolution model to simulate the inconsistency across different views $I_{SR}^{IM} = I_{SR} + IM(I_{SR})$ , thus avoiding making Gaussians to represent such inconsistency. However, we empirically found that this module tends to lose details during optimization to win better consistency, resulting in the loosing of sharp textures. Therefore, we further propose a blur proposal module (BP) to blur the rendered image which is then constrained with modified super-resolution image $I_{SR}^{IM}$ , thereby compensating for the loss of fine-grained details and textures caused by IM. The blur proposal module, i.e., a four-layer convolutional network, predicts per-pixel blur kernels $\mathcal{B}_k \in \mathbb{R}^{K \times K \times H \times W}$ ( $K$ is the blur kernel size, 5 for $4 \times$ task) with the gradient detached rendered image $R_{HR}^{\mathrm{detach}}$ as input $\mathcal{B}_k = BP(R_{HR}^{\mathrm{detach}})$ . Then the blurred + +![](images/e83f08560595106cc399211177c1368b590b1eff527934c11e78508cc8cef801.jpg) + +![](images/f73887a9fdfcab8e10049b4ee6b9d9eeab197fb1c26da2a1ab6d63374ac4707b.jpg) + +high-resolution image is obtained via $R_{HR}^{blur} = R_{HR} * \mathcal{B}_k$ which is utilized to compute loss with $I_{SR}^{IM}$ , avoiding the smoothing imposed by directly constraining $R_{HR}$ : + +$$ +\mathcal {L} _ {S R} = (1 - \beta) \mathcal {L} _ {1} \left(R _ {H R} ^ {\text {b l u r}}, I _ {S R} ^ {\text {I M}}\right) + \beta \mathcal {L} _ {D - S S I M} \left(R _ {H R} ^ {\text {b l u r}}, I _ {S R} ^ {\text {I M}}\right), \tag {4} +$$ + +where $\beta$ is set to 0.2 similar to the original 3DGS [18]. + +3D Robust Optimization. Albeit the above scheme could surmount the detrimental effects of multi-view inconsistencies, the optimization of Gaussians still unavoidably suffers from improper supervision of pseudo-views where some areas are not sufficiently optimized. To this end, we empirically observed that optimizing Gaussian primitives utilizing normal high-quality views results in stable and overall consistent gradients (top left of Fig. 4) whereas corrupted views can lead to significant perturbation and gradient confusion (top middle of Fig. 4), ultimately resulting in blurry rendering. Intrigued by this observation, we aim to design a more robust optimization strategy to mitigate the fluctuations caused by imperfect supervision. + +More specifically, we augment the canonical Gaussian primitives + +$$ +\{\mu , s, q, \sigma , c \} \cup \{\nabla \}, \tag {5} +$$ + +with an additional per-Gaussian attribute flag gradient $\nabla \in \mathbb{R}^{11 + k_c}$ . Where $\nabla$ is used to record the gradient trends of other attributes of the Gaussian, i.e., $\mu, s, q, \sigma$ , and the changed $k_{c}$ SH coefficients. As shown in Fig. 4, for each Gaussian in one optimization step, we first compute the cosine similarity between the current gradient for each attribute $g_i^c$ ( $i$ indicate different attribute) and the corresponding flag gradient $\nabla (g_i)$ stored in $\nabla$ . If the cosine similarity is greater than zero, indicating that the gradient update trends are aligned, the gradient $g_i^c$ will be directly used to + +Table 1. Quantitative comparison on Blender $\times 4$ (8 views),LLFF $\times 4$ (3 views),and Mip-NeRF360 $\times 4$ (24 views). The best,second best, and third best entries are marked in red, orange, and yellow, respectively. + +
MethodBlender ×4 (8 views)LLFF ×4 (3 views)Mip-NeRF 360 ×4 (24 views)
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
NeRF-SR [34]12.410.7440.51593.799.280.2260.617169.5810.260.2690.628151.62
RegNeRF [27]20.680.8410.12941.2815.780.4320.422116.7417.280.4170.449105.45
SparseNeRF [35]21.320.8440.13642.6216.320.4460.428127.6417.380.4320.446109.27
3DGS [18]20.650.8310.14142.5813.010.2720.489138.5316.250.3420.474108.32
SRGS [9]22.640.8620.11636.2417.580.4720.291102.4318.370.4570.39791.62
Mip-Splatting [45]22.780.8680.11933.6315.650.4330.372114.0318.080.4480.42186.17
DNGaussian [22]20.970.8440.13541.1516.270.4300.346129.2617.210.4220.452112.22
FSGS [50]21.450.8540.13340.3618.350.5050.24797.0718.640.4580.39494.21
FSGS [50]+SRGS [9]22.680.8650.11435.4219.490.5900.17562.6319.830.5130.44884.12
S2Gaussian (Ours)24.190.8790.08928.4720.450.6540.13945.8922.050.6870.29643.51
+ +optimize the parameter, and the flag gradient value $\nabla (g_i)$ will be updated to the average one: + +$$ +g _ {i} ^ {c} = g _ {i} ^ {c}, \nabla (g _ {i}) = (\nabla (g _ {i}) + g _ {i} ^ {c}) / 2. \tag {6} +$$ + +Conversely, if the cosine similarity is less than or equal to zero, indicating potential disturbances, the current gradient will be scaled down by a factor of $\epsilon$ to slow the parameter updates, and $\nabla (g_i)$ will be updated as follows: + +$$ +g _ {i} ^ {c} = \epsilon g _ {i} ^ {c}, \nabla (g _ {i}) = (1 - \epsilon) \nabla (g _ {i}) + \epsilon g _ {i} ^ {c}, \tag {7} +$$ + +where $\epsilon$ is set to 0.1 to attenuate potentially erroneous updates and preserve the ability to shift update trends. In practice, such strategy can be combined with any gradient-based optimizer, e.g., Adam, by simply passing the modified gradients to the respective optimizer. Notably, $\nabla$ will be eliminated after the optimization and will not affect the rendering speed and storage consumption. + +As illustrated in the top right of Fig. 4, our proposed 3D robust optimization scheme can significantly improve the stability of Gaussian primitive optimization by eliminating erroneous updates and perturbations thus reconstructing more accurate and high-quality 3D representations. + +Total Objective. Apart from $\mathcal{L}_{SR}$ , an auxiliary loss combined with total variation (TV) loss and sub-pixel constraint is also deployed to stabilize the training: + +$$ +\mathcal {L} _ {A U X} = \mathcal {L} _ {T V} \left(R _ {H R}\right) + \mathcal {L} _ {1} \left(R _ {H R} \downarrow , I _ {L R}\right), \tag {8} +$$ + +where $\downarrow$ indicates area average downsampling and $I_{LR}$ denotes the low-resolution images before super-resolution. + +In summary, the final loss $\mathcal{L}$ for high-resolution Gaussians optimization is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {A U X} + \mathcal {L} _ {S R}. \tag {9} +$$ + +# 4. Experiments + +Following prior literature [21, 44], we primarily experiment on $4 \times$ super-resolution tasks in the maintext since most real-world 2D super-resolution models [36, 46] are mainly trained on $\times 4$ data, more diagnostic experiments on other scaling factors are provided in the Suppl. + +# 4.1. Implementation Details + +We implemented S2Gaussian upon the original 3DGS repository and all experiments are conducted on a single RTX3090 GPU. For the optimization of LR GS, the pretrained DPT [29] model is adapted for depth estimation and the total iterations are set to 10,000. During the HR GS optimization stage, the Gaussian primitives are further optimized with additional 10,000 iterations. As for the off-the-shelf 2D super-resolution model, we opt for powerful diffusion-based ResShift [46] as our backbone. We conduct experiments on three benchmark datasets with SfM for initialization, advanced techniques like DUSt3R [37] can facilitate the initialization but that is not the main purpose of our current work. + +Blender Datasets [25] contain 8 detailed objects synthesized by Blender with a resolution of $800 \times 800$ . Each object initially includes 100 training images and 200 testing images. We uniformly sample 8 images from the original 100 training images and use the low-resolution of $200 \times 200$ pixels for training. + +LLFF Datasets [24] consist of 8 forward-facing real-world scenes. Following the previous setup, we select every eighth image as the test set, and evenly sample 3 sparse views from the remaining images for training. We leverage low-resolution images at $252 \times 189$ pixels for training and render at a higher resolution of $1008 \times 756$ pixels. + +Mip-NeRF 360 Datasets [2] consist of 9 real-world scenes with 5 outdoors and 4 indoors. Similar to LLFF, we use 1/8 of the full view for testing and uniformly sample 24 images from the remaining views for training. We utilize the images with downsampling rate of 4 (about 1K resolution) as the target size and further downsample the training images by a factor 4 to obtain sparse low-resolution inputs. + +The following measures are used to evaluate the quantitative performance: Peak Signal-to-Noise Ratio (PSNR) [17], Structural Similarity Index (SSIM) [38], LPIPS (VGG) [47], and Fréchet Inception Distance (FID) [14]. + +![](images/5ee8c502f3b802987b70ab119c737720662ddc2cdebbc68f92962419ecd13f59.jpg) +Ground Truth + +![](images/b49094f07a701c2d6a7b46b477bfa9d2960f8acc79682e4963934c06f4f887bc.jpg) +3DGS + +![](images/8fb05117f51f7b03ccadcf6d66f97d63728bd41f1a98fa01ad53f284e48a719d.jpg) +Mip-Splatting + +![](images/a5f38332568d3d2881aa01c46eaf6599cc58f4a96d280c0e2cfb5a8d7885226d.jpg) +FSGS + +![](images/b8c8350b7aead2d9694aed8280321ffe00f9a74b3565aa6faf892fcfe17e1b9a.jpg) +FSGS+SRGS + +![](images/1f039759540863a34aa76cfdacaaa258acf62e8e6caf0efeff67f54866e42d0f.jpg) +Ours + +![](images/c08e8af8a74d4467922fd5174e34ae9b9f1a48cc6867b891a19121dcf7aaa510.jpg) +Figure 5. Qualitative comparisons on Blender $\times 4$ datasets with 8 input views. Our method produces more visually appealing results, successfully reconstructing intricate thin structures with fine-grained details. Best viewed at screen! +Ground Truth +Figure 6. Qualitative comparisons on LLFF $\times 4$ datasets with 24 input views. Our method still generates the most accurate super-resolution 3D scenes, preserving both photo-realistic texture and geometric-level details. Best viewed at screen! + +![](images/10a48e4441ef3e9c61defbdf525bb939f5c4e06dc1d2b381714dda376a8ea0c5.jpg) +3DGS + +![](images/dfbd85935a5f295d6a5a37833305338f7ce8875dce0b0d940dc481982f4c26dd.jpg) +Mip-Splatting + +![](images/91cc7d8a5d21a3636a211561149a2b2b91b3631719cdf976f17bd5d2f3b207b0.jpg) +FSGS + +![](images/0c4171e4c843d0cdfc5e96c427ea8764903e6bfcbc2f6ea9e53e6a053dc0377e.jpg) +FSGS+SRGS + +![](images/74a43a383cd9b8066316b297ae70efa477dd759003cbae6a8ae4d239ecbde568.jpg) +Ours + +# 4.2. Quantitative and Qualitative Comparisons + +To rigorously validate the effectiveness of our proposed S2Gaussian, we conduct a comparative analysis against a range of prior methods, including NeRF-SR [34], RegNeRF [27], SparseNeRF [35], vanilla 3DGS [18], SRGS [9], Mip-Splatting [45], DNGaussian [22], and FSGS [50]. Besides, we also include an enhanced version of FSGS with SRGS to provide a standard upsampling baseline. While NeRF-SR and FSGS+SRGS employ super-resolution techniques directly, the remaining methods are trained on low-resolution input views and subsequently rendered at higher resolutions. + +Quantitative Evaluation. Tab. 1 presents quantitative comparison results for $4 \times$ sparse-view super-resolution novel view synthesis task on the Blender [25], LLFF [24], and Mip-NeRF 360 datasets [2]. It is observed that our S2Gaussian delivers remarkable performance gains and outperforms all competitive method significantly in terms of PSNR, SSIM, LPIPS, and FID metrics. Especially for Mip-NeRF 360 datasets, S2Gaussian exceeds the top-performing combinatorial method FSGS+SRGS by an impressive margin of $2.22\mathrm{dB}$ PSNR, underscoring its superior capability + +in reconstructing large, real-world scenes. In addition to delivering higher reconstruction fidelity, reflected by PSNR and SSIM, S2Gaussian also achieves substantial improvements in perceptual metrics LPIPS and FID, validating the effectiveness of S2Gaussian in converging towards greater perception quality with faithful and fidelity details. + +Qualitative Evaluation. We also demonstrate visual comparisons in Fig. 5, Fig. 6, and Fig. 7. As suggested, both 3DGS and Mip-Splatting struggle with the extremely sparse view problem, while FSGS is limited to characterizing high-resolution details. Despite the combinatorial method FSGS+SRGS can reconstruct a reasonable overall structure but suffers from inconsistency and sparsity leading to missing details and blurring. In comparison, S2Gaussian recovers more coherent fine structures with natural textures and harmonious global tone while other methods tend to introduce improper structures and unnatural details, especially for the extremely sparse scenes (Fig. 6) as well as large-scale scenarios (Fig. 7). + +# 4.3. Ablation Studies + +In Tab. 2, we conduct ablation experiments on the components introduced in S2Gaussian. The effectiveness of each + +![](images/dd461b963b07174c7903fac95d89133e1cdf2b09ca355d94a9d20624aa8a44fc.jpg) +Ground Truth + +![](images/ae35a48c4d2055831c70e91db80e3286ae2afcf4fcae4638c8aa47a7f4917458.jpg) +3DGS + +![](images/5b173ee3032253ce4ba54f4be93dea1b5636e32814f525a01b66d2ba46051123.jpg) +Mip-Splatting +Figure 7. Qualitative comparisons on Mip-NeRF360 $\times 4$ datasets with 24 input views. Our method consistently delivers more coherent fine structures than other methods in large-scale scenes. Best viewed at screen! + +![](images/37df40f5e31b1489a555c74e6bcee8ada0c18e164d9ff25da4afc5f41fd5883a.jpg) +FSGS + +![](images/bb8fd95d8db285470d17ff8c64ed1efd30710a80784b817a7ab525b7a91fa13c.jpg) +FSGS+SRGS + +![](images/0078b2c7752849c4c36e74686b81059db4266b55d8bfa0daeaf1fb2da9eafef8.jpg) +Ours + +proposed component in S2Gaussian is evaluated by gradually integrating them into the model, revealing their individual contributions to the overall performance. More detailed analyses are as below. + +Table 2. Ablation studies on Mip-NeRF ${360} \times 4$ (24 views). + +
VariantPSNR↑SSIM↑LPIPS↓FID↓
Baseline19.790.5200.42987.22
+ Two-Stage20.060.5280.41580.64
+ Gaussian Shuffle Split20.620.5590.37671.29
+ Blur-Free Inconsistency Modeling21.120.6240.34659.63
+ 3D Robust Optimization22.050.6870.29643.51
+ +Effect of Gaussian Shuffle Split. As shown in Tab. 3, fewer sub-Gaussians lead to suboptimal detail modeling (lower PSNR), while 6 offers the best trade-off, while more sub-Gaussians will not only increase the optimization complexity but also introduce extra memory overhead. Additionally, 6 sub-Gaussians can be symmetrically located along the three principal axes of the original Gaussian, perfectly approximating the original Gaussian representation. + +Table 3. Ablation of sub-Gaussian number on Mip-NeRF 360. + +
Number456789
Storage Memory (M)6789112137169202
Final PSNR21.7621.9522.0522.0322.0922.08
+ +Effect of Blur-Free Inconsistency Modeling. As illustrated in Fig. 8, IM models the inconsistency of super-resolved images, and while this would introduce blurring, we model this effect explicitly through blur kernel learning, thus reconstructing results that are blur-free and with fine structural details. + +![](images/5f4cd3efacf71d147c91fe1e6eed51861d7680b170044425e5c817470f88c385.jpg) +Super-Resolved Image +Figure 8. Visualization of the super-resolved image, the residual output of IM, the blurred reconstructed image with predicted blur kernels, and the original reconstructed output. + +![](images/0ba83d76d48cd3c062a222554670627ad166ead0556eea84dabb7b2d7b993052.jpg) +Inconsistency Modeling + +![](images/8af871797106e7d04a37d6a51e043db3a524b0c63658f0b1b470bf813e5c0da9.jpg) +Blur Modeling + +![](images/3a5c457213a92274fc2c19b1f534fc660d679d215ceaf394fd0d9099bdf8009e.jpg) +Reconstructed Output + +Effect of 3D Robust Optimization. Fig. 9 showcases visual comparison with or without 3D robust optimization under corrupted training views. It is observed that 3D robust optimization can effectively avoid unfavorable or erroneous supervision to optimize towards higher quality whereas its absence results in Gaussian primitives simulating corrupted representations. + +![](images/13b1add5f6b4ed73179bfd78eecd1cd9db316bf870b146973bd34a7fc6492ccd.jpg) +Figure 9. Visual comparison with or without 3D robust optimization (3DRO) under corrupted training views. + +![](images/d4527f81b8a4cac880d10897f73e490148498549bd80caa799a3c42425409c09.jpg) + +# 5. Concluding Remarks + +In this paper, we propose S2Gaussian, an innovative framework that can efficiently reconstruct high-quality 3D scenes with only sparse and low-resolution input views. In contrast to existing frameworks, S2Gaussian can handle both view sparse and clarity deficient simultaneously, which is more functional and practical in real-world scenarios. The proposed framework begins by optimizing a low-resolution Gaussian representation, which is then densified to initialize high-resolution Gaussians via tailored Gaussian Shuffle Split operation. Subsequently, it cleverly solves the view sparsity problem by rendering pseudo-views from low-resolution Gaussians for super-resolution supervision. Wherein a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are further proposed to mitigate multi-view inconsistency and eliminate erroneous updates for better geometric structure and fine-grained details reconstruction. Extensive experiments on multiple benchmarks manifest the effectiveness, superiority, and robustness of our method. We expect this work to provide insights into more complicated 3D reconstruction and steer future research on this Gordian knot. + +# Acknowledgements + +The authors are very indebted to the anonymous referees for their critical comments and suggestions for the improvement of this paper. This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285 and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260). + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. +[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. +[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In *Forty-first International Conference on Machine Learning*, 2024. +[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splattering in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. +[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. +[9] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Zhenzhong Kuang, Yu Jun, Jianping Fan, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024. +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. + +[11] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021. +[12] Yuqi Han, Tao Yu, Xiaohang Yu, Yuwang Wang, and Qionghai Dai. Super-nerf: View-consistent detail generation for nerf super-resolution. arXiv preprint arXiv:2304.13518, 2023. +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. +[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. +[16] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023. +[17] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008. +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. +[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. +[20] Hyun-kyu Ko, Dongheok Park, Youngin Park, Byeonghyeon Lee, Juhee Han, and Eunbyung Park. Sequence matters: Harnessing video models in super-resolution. arXiv preprint arXiv:2412.11525, 2024. +[21] Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20561–20570, 2024. +[22] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. +[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. +[24] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. +[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. +[27] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. +[28] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. +[29] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. +[30] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. +[31] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024. +[32] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. +[33] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. +[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022. +[35] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065–9076, 2023. +[36] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for + +real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. +[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. +[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. +[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. +[40] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 $\{\backslash \mathrm{deg}\}$ sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. +[41] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. +[42] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. +[43] Youngho Yoon and Kuk-Jin Yoon. Cross-guided optimization of radiance fields with multi-view image superresolution for high-resolution novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12428-12438, 2023. +[44] Xiqian Yu, Hanxin Zhu, Tianyu He, and Zhibo Chen. Gaussiansr: 3d gaussian super-resolution with 2d diffusion priors. arXiv preprint arXiv:2406.10111, 2024. +[45] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. +[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. +[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[48] Xiaoyun Zheng, Liwei Liao, Jianbo Jiao, Feng Gao, and Ronggang Wang. Surface-sos: Self-supervised object segmentation via neural surface representation. IEEE Transactions on Image Processing, 2024. +[49] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding + +with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. +[50] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. \ No newline at end of file diff --git a/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/images.zip b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a7ff74176f20de5e1dbc3199f4680833b51aedb7 --- /dev/null +++ b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ec39a8eae2a841f87f5e222c355dde22368fa6f529fd449818249289e0cce46 +size 833315 diff --git a/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/layout.json b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..096aff9ab487f5032d0a20517521881b862fa232 --- /dev/null +++ b/2025/S2Gaussian_ Sparse-View Super-Resolution 3D Gaussian Splatting/layout.json @@ -0,0 +1,9988 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 512, + 121 + ], + "type": "text", + "content": "S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "text", + "content": "Yecong Wan" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "text", + "content": " Mingwen Shao" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "text", + "content": " Yuanshuo Cheng" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "text", + "content": " Wangmeng Zuo" + }, + { + "bbox": [ + 127, + 142, + 484, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 158, + 549, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 549, + 185 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 549, + 185 + ], + "type": "text", + "content": "1Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software, Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 185, + 438, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 185, + 438, + 200 + ], + "spans": [ + { + "bbox": [ + 172, + 185, + 438, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 172, + 185, + 438, + 200 + ], + "type": "text", + "content": "Faculty of Computing, Harbin Institute of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 200, + 487, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 200, + 487, + 214 + ], + "spans": [ + { + "bbox": [ + 123, + 200, + 487, + 214 + ], + "type": "text", + "content": "smw278@126.com, {yecongwan, cys1294414023, cswmzuo} " + }, + { + "bbox": [ + 123, + 200, + 487, + 214 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 123, + 200, + 487, + 214 + ], + "type": "text", + "content": " gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 240, + 379, + 356 + ], + "blocks": [ + { + "bbox": [ + 69, + 240, + 379, + 356 + ], + "lines": [ + { + "bbox": [ + 69, + 240, + 379, + 356 + ], + "spans": [ + { + "bbox": [ + 69, + 240, + 379, + 356 + ], + "type": "image", + "image_path": "630e80a6bae84f426c3564fed8b99b3652d7d081c6fea98f4b2eb133a9c3eca0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 363, + 555, + 398 + ], + "lines": [ + { + "bbox": [ + 54, + 363, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 555, + 398 + ], + "type": "text", + "content": "Figure 1. We present S2Gaussian, a novel framework capable of reconstructing high-quality 3D scenes for immersive rendering with only sparse and low-resolution input views. S2Gaussian demonstrates superior performance and yields high-fidelity and high-resolution scene reconstruction with sharp geometric and detailed textures, thus enjoying better functionality and practicality in realistic applications." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 391, + 247, + 552, + 353 + ], + "blocks": [ + { + "bbox": [ + 391, + 247, + 552, + 353 + ], + "lines": [ + { + "bbox": [ + 391, + 247, + 552, + 353 + ], + "spans": [ + { + "bbox": [ + 391, + 247, + 552, + 353 + ], + "type": "image", + "image_path": "06238b75f28db2475544560de0f253bf0587c6b4f521b2be8a701effcb220f03.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 407, + 200, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 407, + 200, + 419 + ], + "spans": [ + { + "bbox": [ + 151, + 407, + 200, + 419 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 428, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 297, + 693 + ], + "type": "text", + "content": "In this paper, we aim ambitiously for a realistic yet challenging problem, namely, how to reconstruct high-quality 3D scenes from sparse low-resolution views that simultaneously suffer from deficient perspectives and clarity. Whereas existing methods only deal with either sparse views or low-resolution observations, they fail to handle such hybrid and complicated scenarios. To this end, we propose a novel Sparse-view Super-resolution 3D Gaussian Splatting framework, dubbed S2Gaussian, that can reconstruct structure-accurate and detail-faithful 3D scenes with only sparse and low-resolution views. The S2Gaussian operates in a two-stage fashion. In the first stage, we initially optimize a low-resolution Gaussian representation with depth regularization and densify it to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation. In the second stage, we refine the high-resolution Gaussians with the super-resolved images generated from both original sparse views and pseudoviews rendered by the low-resolution Gaussians. In which a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are elaborately designed to mitigate multi-view inconsistency and eliminate error" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 409, + 556, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 556, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 556, + 468 + ], + "type": "text", + "content": "neous updates caused by imperfect supervision. Extensive experiments demonstrate superior results and in particular establishing new state-of-the-art performances with more consistent geometry and finer details. Project Page https://jeasco.github.io/S2Gaussian/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 478, + 394, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 478, + 394, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 478, + 394, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 498, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 676 + ], + "type": "text", + "content": "The rapid advancement of virtual reality and metaverse technologies has significantly heightened the demand for realistic 3D scene reconstruction, which holds vast potential across various application domains, including medicine, education, entertainment, etc. Among the recent notable developments, Neural Radiance Fields (NeRF) [25] have demonstrated remarkable capability in synthesizing photorealistic images; however, they still face considerable rendering and training costs, despite ongoing improvements [4, 10, 11, 26]. More recently, 3D Gaussian Splatting (3DGS) [18] has emerged as a prominent approach, characterized by its high quality, rapid reconstruction speed, and support for real-time rendering. Subsequent research efforts [5, 6, 23, 32, 39, 45] have focused on extending the applicability of 3DGS across various scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Unfortunately, these methods typically rely heavily on well-captured dense and high-resolution images for impressive novel view synthesis, which is cumbersome and some" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 154, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 154, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 154, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 154, + 713 + ], + "type": "text", + "content": " * Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "711" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "times impractical in real-world applications. To address this challenge, a variety of approaches have emerged that focus on reconstructing 3D scenes from sparse [7, 8, 22, 27, 35, 40, 42, 50] and low-resolution views [9, 12, 21, 31, 43, 44], yielding worth-celebrating successes. However, these two obstacles, i.e., sparsity and low resolution, have long been treated as isolated problems and tackled separately, whereas input views in practical applications often suffer from both deficient perspectives and clarity. Particularly in contexts such as robotics and internet-collected imagery, where the available views are usually sparse and compounded by low resolutions due to environmental constraints and hardware transfer limitations. This interplay of sparsity and low-resolution poses significant challenges for existing reconstruction frameworks. Furthermore, merely incorporating existing methods remains subject to inherent incompatibility, as super-resolution necessitates dense supervision to recover fine details while sparse-view regularization not only fails to introduce details but tends to lead smoothing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 302, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 295, + 613 + ], + "type": "text", + "content": "To tackle the aforementioned problems, in this paper, we propose a novel 3D reconstruction framework S2Gaussian, capable of reconstructing geometrically precise and richly detailed 3D scenes with only sparse and low-resolution views. Specifically, S2Gaussian consists of two main phases: the high-resolution Gaussians (HR GS) initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the estimated depth information. The optimized low-resolution Gaussians are then utilized to initialize the high-resolution Gaussians through a tailored Gaussian Shuffle Split operation which can provide more compact Gaussian primitives to facilitate fine-grained details reconstruction in high-resolution scenes. In the HR GS optimization stage, we use super-resolved images generated by pretrained super-resolution model to refine the high-resolution Gaussians, leveraging both the original sparse views and the pseudo-views rendered from the low-resolution Gaussians. Notably, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are introduced to mitigate the impacts of multi-view inconsistency and imperfect supervision, ultimately leading to more detailed and higher-quality scene reconstruction. Comprehensive experiments substantiate the superiority of our S2Gaussian beyond alternatives on a variety of benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 616, + 279, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 279, + 627 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 279, + 627 + ], + "type": "text", + "content": "In conclusion, the main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 630, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "type": "text", + "content": "- We propose an innovative two-stage framework, termed S2Gaussian, designed to reconstruct 3D scenes that are both structurally accurate and rich in detail using only sparse and low-resolution input views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "content": "- A dedicated Gaussian Shuffle Split operation is designed to initialize more compact Gaussian primitives for representing fine-grained details and textures in high" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 73, + 399, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 73, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 399, + 83 + ], + "type": "text", + "content": "resolution scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 84, + 553, + 168 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 84, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 84, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 553, + 132 + ], + "type": "text", + "content": "- We introduce a tailored blur-free inconsistency modeling scheme alongside a 3D robust optimization strategy to address multi-view inconsistencies and rectify erroneous updates arising from imperfect supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 133, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 133, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 315, + 133, + 553, + 168 + ], + "type": "text", + "content": "- Our S2Gaussian reconstructs more fine-grained and high-quality Gaussian representations significantly outperforming existing methods across multiple benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 198, + 541, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 541, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 541, + 211 + ], + "type": "text", + "content": "2.1. Novel View Synthesis using Radiance Fields" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 215, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 549 + ], + "type": "text", + "content": "Techniques for novel view synthesis typically involve learning a 3D representation using a limited number of input views and generating images from arbitrary novel perspectives. Recent advancements in Neural Radiance Fields (NeRF) [25] has demonstrated encouraging progress in this field, which learns an implicit neural scene representation for novel view synthesis via coordinate-based neural networks and volume rendering function. Subsequent research has focused extensively on enhancing NeRF's rendering quality [1-3, 33], improving efficiency [4, 10, 11, 26], advancing scene understanding [19, 48, 49], and facilitating 3D content generation [15, 28, 30]. Unfortunately, while NeRF significantly improves the quality of novel view rendering, its expensive training time as well as slow rendering speed hinder broader practical applications. Recently, Kerbl et al. [18] introduced a groundbreaking approach with their 3D Gaussian Splatting (3DGS) method, boosting rendering efficiency by utilizing explicit Gaussian representations combined with differentiable rasterization techniques. Building upon 3D Gaussian Splatting representation, tremendous following efforts [6, 23, 32, 39, 45] concentrate extending 3DGS under various scenarios. Nevertheless, these methods typically require dense and high-quality input views for impressive novel view synthesis, and they will tend to overfit the training data and compromise representations when provided with only sparse or degraded views, leading to a substantial decline in performance for the intended scene reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 554, + 471, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 471, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 471, + 567 + ], + "type": "text", + "content": "2.2. Sparse Novel View Synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "To enhance the availability in practical usage, numerous algorithms have been developed to surmount the reliance on dense perspectives. For instance, Depth-NeRF [8] and Reg-NeRF [27] advocate utilizing depth constraints to improve the accuracy of the sparse scene representation. FreeN-erF [42], on the other hand, employs a frequency regularization strategy to manage the frequency range and penalize density fields near the camera. Additionally, SparseN-erF [35] leverages depth priors from inaccurate observations, implementing local depth ranking and spatial continuity constraints to bolster neural radiance fields. More recently, a variety of 3DGS-based approaches have emerged" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "712" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "[7, 22, 40, 50], paving the way for new blueprints in the field. These methods typically focus on various strategies that utilize depth prior to constrain the optimization of Gaussian primitives, such as SfM depth alignment [7], Pearson correlation depth distribution loss [50], global-local depth normalization [22], and patch-based depth correlation [40]. Unlike these approaches, we spotlight more arduous sparse reconstruction under low-resolution conditions, where available information is more scarce." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 261, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 261, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 261, + 202 + ], + "type": "text", + "content": "2.3. Super-Resolution Novel View Synthesis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 434 + ], + "type": "text", + "content": "Super-resolution novel view synthesis aims to reconstruct high-resolution 3D scenes utilizing only low-resolution multi-view inputs. Unlike anti-aliasing methods [1, 2, 45] which are incapable of rendering intricate details of the novel views and are limited by the input image's information level, this field is more focused on fine-grained and detailed high-resolution rendering of potential high-resolution scenes. As a pioneer in the field, NeRF-SR [34] optimizes high-resolution NeRF through a super-sampling strategy, ensuring that the values of low-resolution pixels match the mean value of high-resolution sub-pixels. Follow-up works are designed to utilize either high-resolution reference images [16] or pre-trained 2D models [12, 21, 43] to generate multi-view consistent details. Concurrently, SRGS [9] and GaussianSR [44] are proposed to utilize 2D super-resolution models for texture injection or diffusion prior exploration, respectively. SuperGaussian [31] and SM [20] further prove two profile frameworks to repurpose video upsampling models to achieve 3D super-resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "type": "text", + "content": "Different from the aforementioned methods, in this work, we ambitiously tackle the challenge of joint sparse-view and super-resolution novel view synthesis by proposing a novel two-stage framework, which cleverly solves the problem of perspectives and clarity deficient to reconstruct geometry-accurate and fine-detailed 3D scenes, making it more desirable and feasible in real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 531, + 137, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 137, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 137, + 545 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 141, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 141, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 141, + 563 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) as an emerged powerful technique for novel view synthesis, models a 3D scene using a collection of " + }, + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "type": "text", + "content": " colored Gaussians:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 605, + 294, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 605, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 90, + 605, + 294, + 628 + ], + "type": "interline_equation", + "content": "g _ {i} (x) = \\exp \\left(- \\frac {1}{2} \\left(x - \\mu_ {i}\\right) ^ {\\top} \\Sigma_ {i} ^ {- 1} \\left(x - \\mu_ {i}\\right)\\right), \\tag {1}", + "image_path": "eb4dcbe3d60c2028e9016a2535e67bb8537b520856e42ea7a62a7ad3731fe105.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq G" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mu_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " represents the mean or center of the Gaussian, and " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_i \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " denotes Gaussian's covariance matrix, specifying its shape and size, which is described by scaling vector " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " and rotation vector " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " for implementation. Additionally, each Gaussian is characterized by an opacity " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_i \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " and a view-dependent color " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "c_i(v) \\in \\mathbb{R}^{k_c}" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "k_c" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " is changed) for rasterization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "To faithfully reconstruct the 3D scene, 3DGS also employs densification operation, i.e., the split, clone, and prune operation, to adaptively regulate the number and density of Gaussian primitives. For simplicity, we deploy the same adaptive density control as the original 3DGS and will not elaborate further in the subsequent presentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 148, + 440, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 440, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 440, + 159 + ], + "type": "text", + "content": "3.2. S2Gaussian Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 163, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 388 + ], + "type": "text", + "content": "The schematic illustration of the proposed S2Gaussian is depicted in Fig. 2. S2Gaussian is primarily composed of two phases i.e., the HR GS initialization stage and the HR GS optimization stage. In the HR GS initialization stage, we first optimize a low-resolution Gaussian representation using the sparse views as well as the predicted depth information, which is further utilized to initialize the high-resolution Gaussian representation by densifying and detailing the Gaussian primitives through our tailored Gaussian Shuffle Split operation. In the HR GS optimization stage, we further optimize the high-resolution Gaussians with the super-resolved images generated by pre-trained super-resolution model from both original sparse views and the pseudo-views rendered by LR Gaussians. In particular, a dedicated blur-free inconsistency modeling scheme and a 3D robust optimization strategy are advanced to eliminate the effects of multi-view inconsistency and imperfect supervision to reconstruct more detailed and high-quality 3D scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 397, + 474, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 474, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 474, + 409 + ], + "type": "text", + "content": "3.3. Stage 1: HR GS Initialization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 581 + ], + "type": "text", + "content": "Directly reconstructing a high-resolution 3D scene from sparse low-resolution viewpoints is extremely challenging, as it is not only prone to overfitting specific viewpoints but also lacks high-resolution detail constraints. To alleviate such a reconstruction dilemma, we advocate initially reconstructing a low-resolution Gaussian representation as an auxiliary to facilitate the optimization of high-resolution Gaussians. On the one hand, the low-resolution Gaussian can be used to initialize high-resolution Gaussians to provide reliable base Gaussian ellipsoids and 3D scene structures. On the other hand, it can provide pseudo-views as additional supervision to avoid overfitting on specific viewpoints as well as provide more potential details and textures for high-resolution scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Concretely, we first obtain the monocular depth maps for the sparse low-resolution training views by utilizing the handy pre-trained depth estimation model [29]. Then we utilize the RGB images as well as the depth maps to co-optimize a low-resolution Gaussian representation with mature depth regularization technologies [22, 50], where Pearson correlation loss [50] is employed in this paper as a baseline. Thereafter, a tailored Gaussian Shuffle Split regime is proposed to densify and detail the sparse Gaussian primitives so as to initialize HR Gaussians for better high-resolution fine-grained details representation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "713" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 558, + 247 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 558, + 247 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 558, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 558, + 247 + ], + "type": "image", + "image_path": "424ae6196f961eb89564199daa0b6f9c870f8b18c80aa756da848c16ed02fbba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 251, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Overview of S2Gaussian. The S2Gaussian initially optimizes an LR GS and densifies it to initialize the HR GS through a tailored Gaussian Shuffle Split operation. Then, the original sparse views along with the pseudo views rendered by the LR GS are super-resolved together to refine the high-resolution texture with 3D robust optimization. In which an inconsistency modeling module (IM) and a blur proposal module are incorporated to mitigate inconsistency and blurriness, aiming to create 3D scenes with high-fidelity texture details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 297, + 498 + ], + "type": "text", + "content": "Gaussian Shuffle Split. A precise and detailed initialization is crucial for reconstructing scene details [18]. Although the structure and layout of the entire 3D scene are basically guaranteed after the optimization of the low-resolution views, the initialized sparse and coarse Gaussian ellipsoids struggle to simulate high-resolution details which require denser Gaussians for finer representation [41]. Even equipped with adaptive density control [18], it falls short in our settings as there are neither dense perspectives nor high-quality details for supervision. To address this issue, we propose a training-free local Gaussian densification strategy Gaussian Shuffle Split which can provide more mobilizable Gaussian primitives by replacing the original large Gaussian with six small Gaussians to facilitate more comprehensive simulation of fine-grained details and textures in high-resolution scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "text", + "content": "As illustrated in Fig. 3, given a Gaussian primitive with attributes " + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\{\\mu, s, q, \\sigma, c\\}" + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "text", + "content": ", we first generate six duplicates, each center shifted along one of the six directions aligned with the original Gaussian's three principal axes, i.e., the positive and negative directions of each axis. The shift for each axis is set to " + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "text", + "content": " (0.5 by default) times the corresponding axis's scaling " + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "inline_equation", + "content": "s = [s_1, s_2, s_3]" + }, + { + "bbox": [ + 55, + 502, + 296, + 597 + ], + "type": "text", + "content": " value. Thus the new position of the six sub-Gaussians can be obtained:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 597, + 296, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 611 + ], + "type": "interline_equation", + "content": "\\mu_ {s u b 1} = \\mu + R \\otimes [ \\alpha s _ {1}, 0, 0 ] ^ {T}, \\mu_ {s u b 2} = \\mu + R \\otimes [ - \\alpha s _ {1}, 0, 0 ] ^ {T},", + "image_path": "b1b62b959f319da41790c0a0fe508434a4d6b18f0b0f6baed4a712c945a201bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 613, + 296, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 627 + ], + "type": "interline_equation", + "content": "\\mu_ {s u b 3} = \\mu + R \\otimes [ 0, \\alpha s _ {2}, 0 ] ^ {T}, \\mu_ {s u b 4} = \\mu + R \\otimes [ 0, - \\alpha s _ {2}, 0 ] ^ {T},", + "image_path": "8b87f31d09d6fb816a624e3d88abf1a8ff6f34ed99448344af42a6ca4b791f61.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 629, + 296, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 629, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 629, + 296, + 653 + ], + "type": "interline_equation", + "content": "\\mu_ {s u b 5} = \\mu + R \\otimes [ 0, 0, \\alpha s _ {3} ] ^ {T}, \\mu_ {s u b 6} = \\mu + R \\otimes [ 0, 0, - \\alpha s _ {3} ] ^ {T}, \\tag {2}", + "image_path": "bd2e43b612f5cc00b2a2a92af450538ba3945b8e3ed82555d5abcbadae54fc6a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " represents the rotation matrix calculated from " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " to align with the original rotation of the Gaussian. Additionally, for each shifted sub-Gaussian, we scale their " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " value corresponding to the offset axis to " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " of its original value, i.e., reducing the scale along that axis to shrink the size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "text", + "content": "The remaining two " + }, + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "text", + "content": " values are scaled by an another factor of " + }, + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 308, + 555, + 331 + ], + "type": "text", + "content": " relative to their original values:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 332, + 509, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 332, + 509, + 346 + ], + "spans": [ + { + "bbox": [ + 359, + 332, + 509, + 346 + ], + "type": "interline_equation", + "content": "s _ {s u b 1} = s _ {s u b 2} = \\left[ s _ {1} / 4, s _ {2} / \\lambda , s _ {3} / \\lambda \\right],", + "image_path": "e3e539ab87c25aa7c531dee252db77163edfb1c9ad0132618740d82bed463b55.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 360, + 347, + 553, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 347, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 360, + 347, + 553, + 361 + ], + "type": "interline_equation", + "content": "s _ {s u b 3} = s _ {s u b 4} = \\left[ s _ {1} / \\lambda , s _ {2} / 4, s _ {3} / \\lambda \\right], \\tag {3}", + "image_path": "ca2c8471209510dd424c7bca78e6253416981d15a31b2a31ff193f098d8a409a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 360, + 363, + 507, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 363, + 507, + 375 + ], + "spans": [ + { + "bbox": [ + 360, + 363, + 507, + 375 + ], + "type": "interline_equation", + "content": "s _ {s u b 5} = s _ {s u b 6} = \\left[ s _ {1} / \\lambda , s _ {2} / \\lambda , s _ {3} / 4 \\right],", + "image_path": "22c4b05f86127e70442fe1589349ce49c3cff39e8f9cd069ab2dbf515d255d22.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " is experimentally set to 1.9 which we found this ratio ensures that the combined representation of the sub-Gaussians better approximate the original Gaussian, preserving the integrity of the 3D scene. All other attributes, i.e., rotation " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " remain identical to the original Gaussian ellipsoid. The combined six sub-Gaussians are utilized to substitute the original large Gaussians for densified 3D representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 568 + ], + "type": "text", + "content": "We deployed Gaussian Shuffle Split only for Gaussian primitives with opacity greater than 0.5 since these are more likely to be responsible for representing object surfaces and key structures which require a more detailed representation in high-resolution scenes. The opacity of all primitives is set close to zero after the Gaussian Shuffle Split to facilitate the automatic elimination of redundant Gaussians with low opacity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 575, + 477, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 477, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 477, + 589 + ], + "type": "text", + "content": "3.4. Stage 2: HR GS Optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "With the initialized high-resolution Gaussians, in this stage, we aim to continue optimizing it with the super-resolved images generated from both original sparse views and the pseudo-views rendered by the LR Gaussians. We synthesize pseudo views by interpolating virtual views between two known views as in [50]. Considering the prohibitively expensive data costs associated with 3D super-resolution models and their computational complexity, we advocate for utilizing off-the-shelf pre-trained 2D super-resolution models to enhance low-resolution views. Unfortunately, directly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "714" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 176, + 135 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 176, + 135 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 176, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 176, + 135 + ], + "type": "image", + "image_path": "f8d1d65b005d2cf70f02cb48aac04cbcb9537458b39d5a23bad55934c96c644b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 71, + 290, + 135 + ], + "blocks": [ + { + "bbox": [ + 178, + 71, + 290, + 135 + ], + "lines": [ + { + "bbox": [ + 178, + 71, + 290, + 135 + ], + "spans": [ + { + "bbox": [ + 178, + 71, + 290, + 135 + ], + "type": "image", + "image_path": "37b4a927806a721ce09fe26efd8781ac1131801c1e38c04c91b657546719425f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 136, + 173, + 200 + ], + "blocks": [ + { + "bbox": [ + 60, + 136, + 173, + 200 + ], + "lines": [ + { + "bbox": [ + 60, + 136, + 173, + 200 + ], + "spans": [ + { + "bbox": [ + 60, + 136, + 173, + 200 + ], + "type": "image", + "image_path": "dc2bccd44a7fa650901abd438e94a4477771f78b0af386aed1255496e6b16189.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 136, + 291, + 200 + ], + "blocks": [ + { + "bbox": [ + 176, + 136, + 291, + 200 + ], + "lines": [ + { + "bbox": [ + 176, + 136, + 291, + 200 + ], + "spans": [ + { + "bbox": [ + 176, + 136, + 291, + 200 + ], + "type": "image", + "image_path": "210db062e178916ae853f4fa18e1883858beada1c679997cfb80b0dd6e6e02d3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 201, + 173, + 281 + ], + "blocks": [ + { + "bbox": [ + 59, + 201, + 173, + 281 + ], + "lines": [ + { + "bbox": [ + 59, + 201, + 173, + 281 + ], + "spans": [ + { + "bbox": [ + 59, + 201, + 173, + 281 + ], + "type": "image", + "image_path": "ccbe3c2ca2f95c9a518b4c35466701dfe919d19b8fb450a4b7ee85dfa71770d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 295, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 334 + ], + "type": "text", + "content": "Figure 3. Illustration of Gaussian Shuffle Split that utilizes a combination of six compact Gaussians to replace the original large Gaussian primitive, which can densify the 3D scene virtually without damaging the original 3D representation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 201, + 292, + 281 + ], + "blocks": [ + { + "bbox": [ + 178, + 201, + 292, + 281 + ], + "lines": [ + { + "bbox": [ + 178, + 201, + 292, + 281 + ], + "spans": [ + { + "bbox": [ + 178, + 201, + 292, + 281 + ], + "type": "image", + "image_path": "4327336c60bc6d1aebf227d0ea27a3beb11bf0ca073a5e3fe584d07520f274e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 265, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 309 + ], + "type": "text", + "content": "Figure 4. Top: Visualization of the gradient variation under different quality views supervision and our proposed 3D robust optimization. Bottom: Illustration of the proposed 3D robust optimization strategy that takes the gradient of scaling " + }, + { + "bbox": [ + 313, + 265, + 555, + 309 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 265, + 555, + 309 + ], + "type": "text", + "content": " as an example." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 340, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 424 + ], + "type": "text", + "content": "leveraging 2D super-resolution model cannot ensure multiview consistency, and the low-resolution pseudo-views inevitably exhibit artifacts since there are areas that have not been optimized before. The Gaussian primitives, in their attempt to accommodate these inconsistencies and erroneous representations, will lead to the optimized 3D scene converging toward inaccurate structures and blurriness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 496 + ], + "type": "text", + "content": "During optimization, two crucial aspects contribute fundamentally to these problems, i.e., inconsistent supervision and erroneous updates caused by artifacts. Accordingly, a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are proposed in the following to alleviate potential inaccurate reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": "Blur-Free Inconsistency Modeling. To mitigate the inconsistency effort of the single-image super-resolution model, we employ a learnable inconsistency modeling module (IM), i.e., two residual blocks [13], after the pre-trained super-resolution model to simulate the inconsistency across different views " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{SR}^{IM} = I_{SR} + IM(I_{SR})" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ", thus avoiding making Gaussians to represent such inconsistency. However, we empirically found that this module tends to lose details during optimization to win better consistency, resulting in the loosing of sharp textures. Therefore, we further propose a blur proposal module (BP) to blur the rendered image which is then constrained with modified super-resolution image " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_{SR}^{IM}" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ", thereby compensating for the loss of fine-grained details and textures caused by IM. The blur proposal module, i.e., a four-layer convolutional network, predicts per-pixel blur kernels " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_k \\in \\mathbb{R}^{K \\times K \\times H \\times W}" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " is the blur kernel size, 5 for " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " task) with the gradient detached rendered image " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "R_{HR}^{\\mathrm{detach}}" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " as input " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_k = BP(R_{HR}^{\\mathrm{detach}})" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ". Then the blurred" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 152 + ], + "type": "image", + "image_path": "e83f08560595106cc399211177c1368b590b1eff527934c11e78508cc8cef801.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 154, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 317, + 154, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 317, + 154, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 553, + 257 + ], + "type": "image", + "image_path": "f73887a9fdfcab8e10049b4ee6b9d9eeab197fb1c26da2a1ab6d63374ac4707b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "text", + "content": "high-resolution image is obtained via " + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "inline_equation", + "content": "R_{HR}^{blur} = R_{HR} * \\mathcal{B}_k" + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "text", + "content": " which is utilized to compute loss with " + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "inline_equation", + "content": "I_{SR}^{IM}" + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "text", + "content": ", avoiding the smoothing imposed by directly constraining " + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "inline_equation", + "content": "R_{HR}" + }, + { + "bbox": [ + 313, + 319, + 553, + 357 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 357, + 553, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 357, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 553, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S R} = (1 - \\beta) \\mathcal {L} _ {1} \\left(R _ {H R} ^ {\\text {b l u r}}, I _ {S R} ^ {\\text {I M}}\\right) + \\beta \\mathcal {L} _ {D - S S I M} \\left(R _ {H R} ^ {\\text {b l u r}}, I _ {S R} ^ {\\text {I M}}\\right), \\tag {4}", + "image_path": "ae3ad5cf56edf3d11ea36c38465ad40b66058a6cceb9a19931eb79c4a1781610.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 380, + 535, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 535, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 535, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 535, + 392 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 380, + 535, + 392 + ], + "type": "text", + "content": " is set to 0.2 similar to the original 3DGS [18]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 393, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 548 + ], + "type": "text", + "content": "3D Robust Optimization. Albeit the above scheme could surmount the detrimental effects of multi-view inconsistencies, the optimization of Gaussians still unavoidably suffers from improper supervision of pseudo-views where some areas are not sufficiently optimized. To this end, we empirically observed that optimizing Gaussian primitives utilizing normal high-quality views results in stable and overall consistent gradients (top left of Fig. 4) whereas corrupted views can lead to significant perturbation and gradient confusion (top middle of Fig. 4), ultimately resulting in blurry rendering. Intrigued by this observation, we aim to design a more robust optimization strategy to mitigate the fluctuations caused by imperfect supervision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 549, + 553, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 553, + 572 + ], + "type": "text", + "content": "More specifically, we augment the canonical Gaussian primitives" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 390, + 574, + 553, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 574, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 390, + 574, + 553, + 587 + ], + "type": "interline_equation", + "content": "\\{\\mu , s, q, \\sigma , c \\} \\cup \\{\\nabla \\}, \\tag {5}", + "image_path": "216b92c4fc64b0abf00d5893bf88042574ff851f5091f6b2737e18edf486d1fb.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "with an additional per-Gaussian attribute flag gradient " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\nabla \\in \\mathbb{R}^{11 + k_c}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " is used to record the gradient trends of other attributes of the Gaussian, i.e., " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mu, s, q, \\sigma" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", and the changed " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "k_{c}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " SH coefficients. As shown in Fig. 4, for each Gaussian in one optimization step, we first compute the cosine similarity between the current gradient for each attribute " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "g_i^c" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " indicate different attribute) and the corresponding flag gradient " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\nabla (g_i)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " stored in " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". If the cosine similarity is greater than zero, indicating that the gradient update trends are aligned, the gradient " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "g_i^c" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " will be directly used to" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "715" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 558, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on Blender " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " (8 views),LLFF " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " (3 views),and Mip-NeRF360 " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " (24 views). The best,second best, and third best entries are marked in red, orange, and yellow, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 558, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 558, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 558, + 235 + ], + "type": "table", + "html": "
MethodBlender ×4 (8 views)LLFF ×4 (3 views)Mip-NeRF 360 ×4 (24 views)
PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓PSNR↑SSIM↑LPIPS↓FID↓
NeRF-SR [34]12.410.7440.51593.799.280.2260.617169.5810.260.2690.628151.62
RegNeRF [27]20.680.8410.12941.2815.780.4320.422116.7417.280.4170.449105.45
SparseNeRF [35]21.320.8440.13642.6216.320.4460.428127.6417.380.4320.446109.27
3DGS [18]20.650.8310.14142.5813.010.2720.489138.5316.250.3420.474108.32
SRGS [9]22.640.8620.11636.2417.580.4720.291102.4318.370.4570.39791.62
Mip-Splatting [45]22.780.8680.11933.6315.650.4330.372114.0318.080.4480.42186.17
DNGaussian [22]20.970.8440.13541.1516.270.4300.346129.2617.210.4220.452112.22
FSGS [50]21.450.8540.13340.3618.350.5050.24797.0718.640.4580.39494.21
FSGS [50]+SRGS [9]22.680.8650.11435.4219.490.5900.17562.6319.830.5130.44884.12
S2Gaussian (Ours)24.190.8790.08928.4720.450.6540.13945.8922.050.6870.29643.51
", + "image_path": "cfd46f92fa656f57060930e3d639172e162b61add564e4ac2cd43ba2527472c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "content": "optimize the parameter, and the flag gradient value " + }, + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "inline_equation", + "content": "\\nabla (g_i)" + }, + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "content": " will be updated to the average one:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 273, + 294, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 273, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 102, + 273, + 294, + 285 + ], + "type": "interline_equation", + "content": "g _ {i} ^ {c} = g _ {i} ^ {c}, \\nabla (g _ {i}) = (\\nabla (g _ {i}) + g _ {i} ^ {c}) / 2. \\tag {6}", + "image_path": "2a1960a73a19a70a1a0de6ed45770d195fe11121f8741cac01bb530f15a25133.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "type": "text", + "content": "Conversely, if the cosine similarity is less than or equal to zero, indicating potential disturbances, the current gradient will be scaled down by a factor of " + }, + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "type": "text", + "content": " to slow the parameter updates, and " + }, + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "type": "inline_equation", + "content": "\\nabla (g_i)" + }, + { + "bbox": [ + 55, + 286, + 295, + 334 + ], + "type": "text", + "content": " will be updated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 335, + 294, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 335, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 93, + 335, + 294, + 348 + ], + "type": "interline_equation", + "content": "g _ {i} ^ {c} = \\epsilon g _ {i} ^ {c}, \\nabla (g _ {i}) = (1 - \\epsilon) \\nabla (g _ {i}) + \\epsilon g _ {i} ^ {c}, \\tag {7}", + "image_path": "f6cf59a4896249995575baff30e4dabb0a991a91014adbf66d3ee26a912b12e6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "type": "text", + "content": " is set to 0.1 to attenuate potentially erroneous updates and preserve the ability to shift update trends. In practice, such strategy can be combined with any gradient-based optimizer, e.g., Adam, by simply passing the modified gradients to the respective optimizer. Notably, " + }, + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 55, + 348, + 296, + 432 + ], + "type": "text", + "content": " will be eliminated after the optimization and will not affect the rendering speed and storage consumption." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "content": "As illustrated in the top right of Fig. 4, our proposed 3D robust optimization scheme can significantly improve the stability of Gaussian primitive optimization by eliminating erroneous updates and perturbations thus reconstructing more accurate and high-quality 3D representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "content": "Total Objective. Apart from " + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SR}" + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "content": ", an auxiliary loss combined with total variation (TV) loss and sub-pixel constraint is also deployed to stabilize the training:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 536, + 294, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 536, + 294, + 549 + ], + "spans": [ + { + "bbox": [ + 88, + 536, + 294, + 549 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A U X} = \\mathcal {L} _ {T V} \\left(R _ {H R}\\right) + \\mathcal {L} _ {1} \\left(R _ {H R} \\downarrow , I _ {L R}\\right), \\tag {8}", + "image_path": "674f72c7ba6845852180b8c0e14ad00081a584f0d6c340bca946fc4e23310869.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "content": " indicates area average downsampling and " + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 55, + 550, + 295, + 574 + ], + "type": "text", + "content": " denotes the low-resolution images before super-resolution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "text", + "content": "In summary, the final loss " + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "text", + "content": " for high-resolution Gaussians optimization is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 133, + 603, + 294, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 603, + 294, + 615 + ], + "spans": [ + { + "bbox": [ + 133, + 603, + 294, + 615 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {A U X} + \\mathcal {L} _ {S R}. \\tag {9}", + "image_path": "e19c790f8542cfdb2882cea77ec8f75000aac597e97b7c12a4cff51039fe2656.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 632, + 137, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 137, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 137, + 646 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Following prior literature [21, 44], we primarily experiment on " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " super-resolution tasks in the maintext since most real-world 2D super-resolution models [36, 46] are mainly trained on " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " data, more diagnostic experiments on other scaling factors are provided in the Suppl." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 247, + 447, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 447, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 447, + 260 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 268, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 423 + ], + "type": "text", + "content": "We implemented S2Gaussian upon the original 3DGS repository and all experiments are conducted on a single RTX3090 GPU. For the optimization of LR GS, the pretrained DPT [29] model is adapted for depth estimation and the total iterations are set to 10,000. During the HR GS optimization stage, the Gaussian primitives are further optimized with additional 10,000 iterations. As for the off-the-shelf 2D super-resolution model, we opt for powerful diffusion-based ResShift [46] as our backbone. We conduct experiments on three benchmark datasets with SfM for initialization, advanced techniques like DUSt3R [37] can facilitate the initialization but that is not the main purpose of our current work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "text", + "content": "Blender Datasets [25] contain 8 detailed objects synthesized by Blender with a resolution of " + }, + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "text", + "content": ". Each object initially includes 100 training images and 200 testing images. We uniformly sample 8 images from the original 100 training images and use the low-resolution of " + }, + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "text", + "content": " pixels for training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "type": "text", + "content": "LLFF Datasets [24] consist of 8 forward-facing real-world scenes. Following the previous setup, we select every eighth image as the test set, and evenly sample 3 sparse views from the remaining images for training. We leverage low-resolution images at " + }, + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "type": "inline_equation", + "content": "252 \\times 189" + }, + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "type": "text", + "content": " pixels for training and render at a higher resolution of " + }, + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "type": "inline_equation", + "content": "1008 \\times 756" + }, + { + "bbox": [ + 313, + 502, + 554, + 575 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 578, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 662 + ], + "type": "text", + "content": "Mip-NeRF 360 Datasets [2] consist of 9 real-world scenes with 5 outdoors and 4 indoors. Similar to LLFF, we use 1/8 of the full view for testing and uniformly sample 24 images from the remaining views for training. We utilize the images with downsampling rate of 4 (about 1K resolution) as the target size and further downsample the training images by a factor 4 to obtain sparse low-resolution inputs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "The following measures are used to evaluate the quantitative performance: Peak Signal-to-Noise Ratio (PSNR) [17], Structural Similarity Index (SSIM) [38], LPIPS (VGG) [47], and Fréchet Inception Distance (FID) [14]." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "716" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 132, + 217 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 132, + 217 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 132, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 132, + 217 + ], + "type": "image", + "image_path": "5ee8c502f3b802987b70ab119c737720662ddc2cdebbc68f92962419ecd13f59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 218, + 126, + 228 + ], + "lines": [ + { + "bbox": [ + 67, + 218, + 126, + 228 + ], + "spans": [ + { + "bbox": [ + 67, + 218, + 126, + 228 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 214, + 217 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 214, + 217 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 214, + 217 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 214, + 217 + ], + "type": "image", + "image_path": "b49094f07a701c2d6a7b46b477bfa9d2960f8acc79682e4963934c06f4f887bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 218, + 192, + 228 + ], + "lines": [ + { + "bbox": [ + 167, + 218, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 167, + 218, + 192, + 228 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 297, + 217 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 297, + 217 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 297, + 217 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 297, + 217 + ], + "type": "image", + "image_path": "8fb05117f51f7b03ccadcf6d66f97d63728bd41f1a98fa01ad53f284e48a719d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 218, + 293, + 229 + ], + "lines": [ + { + "bbox": [ + 235, + 218, + 293, + 229 + ], + "spans": [ + { + "bbox": [ + 235, + 218, + 293, + 229 + ], + "type": "text", + "content": "Mip-Splatting" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 381, + 217 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 381, + 217 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 381, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 381, + 217 + ], + "type": "image", + "image_path": "a5f38332568d3d2881aa01c46eaf6599cc58f4a96d280c0e2cfb5a8d7885226d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 218, + 359, + 228 + ], + "lines": [ + { + "bbox": [ + 334, + 218, + 359, + 228 + ], + "spans": [ + { + "bbox": [ + 334, + 218, + 359, + 228 + ], + "type": "text", + "content": "FSGS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 390, + 70, + 464, + 217 + ], + "blocks": [ + { + "bbox": [ + 390, + 70, + 464, + 217 + ], + "lines": [ + { + "bbox": [ + 390, + 70, + 464, + 217 + ], + "spans": [ + { + "bbox": [ + 390, + 70, + 464, + 217 + ], + "type": "image", + "image_path": "b8c8350b7aead2d9694aed8280321ffe00f9a74b3565aa6faf892fcfe17e1b9a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 218, + 457, + 228 + ], + "lines": [ + { + "bbox": [ + 403, + 218, + 457, + 228 + ], + "spans": [ + { + "bbox": [ + 403, + 218, + 457, + 228 + ], + "type": "text", + "content": "FSGS+SRGS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 473, + 70, + 547, + 217 + ], + "blocks": [ + { + "bbox": [ + 473, + 70, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 473, + 70, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 473, + 70, + 547, + 217 + ], + "type": "image", + "image_path": "1f039759540863a34aa76cfdacaaa258acf62e8e6caf0efeff67f54866e42d0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 218, + 524, + 228 + ], + "lines": [ + { + "bbox": [ + 503, + 218, + 524, + 228 + ], + "spans": [ + { + "bbox": [ + 503, + 218, + 524, + 228 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 263, + 136, + 388 + ], + "blocks": [ + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons on Blender " + }, + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "type": "text", + "content": " datasets with 8 input views. Our method produces more visually appealing results, successfully reconstructing intricate thin structures with fine-grained details. Best viewed at screen!" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 263, + 136, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 263, + 136, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 136, + 388 + ], + "type": "image", + "image_path": "c08e8af8a74d4467922fd5174e34ae9b9f1a48cc6867b891a19121dcf7aaa510.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 388, + 126, + 397 + ], + "lines": [ + { + "bbox": [ + 67, + 388, + 126, + 397 + ], + "spans": [ + { + "bbox": [ + 67, + 388, + 126, + 397 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons on LLFF " + }, + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "type": "text", + "content": " datasets with 24 input views. Our method still generates the most accurate super-resolution 3D scenes, preserving both photo-realistic texture and geometric-level details. Best viewed at screen!" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 139, + 263, + 220, + 388 + ], + "blocks": [ + { + "bbox": [ + 139, + 263, + 220, + 388 + ], + "lines": [ + { + "bbox": [ + 139, + 263, + 220, + 388 + ], + "spans": [ + { + "bbox": [ + 139, + 263, + 220, + 388 + ], + "type": "image", + "image_path": "10a48e4441ef3e9c61defbdf525bb939f5c4e06dc1d2b381714dda376a8ea0c5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 388, + 192, + 397 + ], + "lines": [ + { + "bbox": [ + 167, + 388, + 192, + 397 + ], + "spans": [ + { + "bbox": [ + 167, + 388, + 192, + 397 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 223, + 263, + 303, + 388 + ], + "blocks": [ + { + "bbox": [ + 223, + 263, + 303, + 388 + ], + "lines": [ + { + "bbox": [ + 223, + 263, + 303, + 388 + ], + "spans": [ + { + "bbox": [ + 223, + 263, + 303, + 388 + ], + "type": "image", + "image_path": "dfbd85935a5f295d6a5a37833305338f7ce8875dce0b0d940dc481982f4c26dd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 388, + 292, + 399 + ], + "lines": [ + { + "bbox": [ + 235, + 388, + 292, + 399 + ], + "spans": [ + { + "bbox": [ + 235, + 388, + 292, + 399 + ], + "type": "text", + "content": "Mip-Splatting" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 263, + 386, + 387 + ], + "blocks": [ + { + "bbox": [ + 307, + 263, + 386, + 387 + ], + "lines": [ + { + "bbox": [ + 307, + 263, + 386, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 386, + 387 + ], + "type": "image", + "image_path": "91cc7d8a5d21a3636a211561149a2b2b91b3631719cdf976f17bd5d2f3b207b0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 388, + 359, + 397 + ], + "lines": [ + { + "bbox": [ + 334, + 388, + 359, + 397 + ], + "spans": [ + { + "bbox": [ + 334, + 388, + 359, + 397 + ], + "type": "text", + "content": "FSGS" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 390, + 263, + 470, + 387 + ], + "blocks": [ + { + "bbox": [ + 390, + 263, + 470, + 387 + ], + "lines": [ + { + "bbox": [ + 390, + 263, + 470, + 387 + ], + "spans": [ + { + "bbox": [ + 390, + 263, + 470, + 387 + ], + "type": "image", + "image_path": "0c4171e4c843d0cdfc5e96c427ea8764903e6bfcbc2f6ea9e53e6a053dc0377e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 388, + 456, + 397 + ], + "lines": [ + { + "bbox": [ + 403, + 388, + 456, + 397 + ], + "spans": [ + { + "bbox": [ + 403, + 388, + 456, + 397 + ], + "type": "text", + "content": "FSGS+SRGS" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 473, + 263, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 473, + 263, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 473, + 263, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 473, + 263, + 553, + 388 + ], + "type": "image", + "image_path": "74a43a383cd9b8066316b297ae70efa477dd759003cbae6a8ae4d239ecbde568.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 388, + 524, + 397 + ], + "lines": [ + { + "bbox": [ + 502, + 388, + 524, + 397 + ], + "spans": [ + { + "bbox": [ + 502, + 388, + 524, + 397 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 436, + 278, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 278, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 278, + 449 + ], + "type": "text", + "content": "4.2. Quantitative and Qualitative Comparisons" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 54, + 457, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 457, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 457, + 295, + 589 + ], + "type": "text", + "content": "To rigorously validate the effectiveness of our proposed S2Gaussian, we conduct a comparative analysis against a range of prior methods, including NeRF-SR [34], RegNeRF [27], SparseNeRF [35], vanilla 3DGS [18], SRGS [9], Mip-Splatting [45], DNGaussian [22], and FSGS [50]. Besides, we also include an enhanced version of FSGS with SRGS to provide a standard upsampling baseline. While NeRF-SR and FSGS+SRGS employ super-resolution techniques directly, the remaining methods are trained on low-resolution input views and subsequently rendered at higher resolutions." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": "Quantitative Evaluation. Tab. 1 presents quantitative comparison results for " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " sparse-view super-resolution novel view synthesis task on the Blender [25], LLFF [24], and Mip-NeRF 360 datasets [2]. It is observed that our S2Gaussian delivers remarkable performance gains and outperforms all competitive method significantly in terms of PSNR, SSIM, LPIPS, and FID metrics. Especially for Mip-NeRF 360 datasets, S2Gaussian exceeds the top-performing combinatorial method FSGS+SRGS by an impressive margin of " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "2.22\\mathrm{dB}" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " PSNR, underscoring its superior capability" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 508 + ], + "type": "text", + "content": "in reconstructing large, real-world scenes. In addition to delivering higher reconstruction fidelity, reflected by PSNR and SSIM, S2Gaussian also achieves substantial improvements in perceptual metrics LPIPS and FID, validating the effectiveness of S2Gaussian in converging towards greater perception quality with faithful and fidelity details." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 509, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 665 + ], + "type": "text", + "content": "Qualitative Evaluation. We also demonstrate visual comparisons in Fig. 5, Fig. 6, and Fig. 7. As suggested, both 3DGS and Mip-Splatting struggle with the extremely sparse view problem, while FSGS is limited to characterizing high-resolution details. Despite the combinatorial method FSGS+SRGS can reconstruct a reasonable overall structure but suffers from inconsistency and sparsity leading to missing details and blurring. In comparison, S2Gaussian recovers more coherent fine structures with natural textures and harmonious global tone while other methods tend to introduce improper structures and unnatural details, especially for the extremely sparse scenes (Fig. 6) as well as large-scale scenarios (Fig. 7)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 672, + 416, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 416, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 416, + 683 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In Tab. 2, we conduct ablation experiments on the components introduced in S2Gaussian. The effectiveness of each" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "717" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 136, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 136, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 136, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 136, + 178 + ], + "type": "image", + "image_path": "dd461b963b07174c7903fac95d89133e1cdf2b09ca355d94a9d20624aa8a44fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 180, + 126, + 190 + ], + "lines": [ + { + "bbox": [ + 67, + 180, + 126, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 126, + 190 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 220, + 178 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 220, + 178 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 220, + 178 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 220, + 178 + ], + "type": "image", + "image_path": "ae35a48c4d2055831c70e91db80e3286ae2afcf4fcae4638c8aa47a7f4917458.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 180, + 193, + 189 + ], + "lines": [ + { + "bbox": [ + 167, + 180, + 193, + 189 + ], + "spans": [ + { + "bbox": [ + 167, + 180, + 193, + 189 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 303, + 178 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 303, + 178 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 303, + 178 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 303, + 178 + ], + "type": "image", + "image_path": "5b173ee3032253ce4ba54f4be93dea1b5636e32814f525a01b66d2ba46051123.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 180, + 293, + 191 + ], + "lines": [ + { + "bbox": [ + 235, + 180, + 293, + 191 + ], + "spans": [ + { + "bbox": [ + 235, + 180, + 293, + 191 + ], + "type": "text", + "content": "Mip-Splatting" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 195, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 555, + 217 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons on Mip-NeRF360 " + }, + { + "bbox": [ + 55, + 195, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 195, + 555, + 217 + ], + "type": "text", + "content": " datasets with 24 input views. Our method consistently delivers more coherent fine structures than other methods in large-scale scenes. Best viewed at screen!" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 387, + 178 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 387, + 178 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 387, + 178 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 387, + 178 + ], + "type": "image", + "image_path": "37df40f5e31b1489a555c74e6bcee8ada0c18e164d9ff25da4afc5f41fd5883a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 180, + 359, + 189 + ], + "lines": [ + { + "bbox": [ + 334, + 180, + 359, + 189 + ], + "spans": [ + { + "bbox": [ + 334, + 180, + 359, + 189 + ], + "type": "text", + "content": "FSGS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 390, + 70, + 470, + 178 + ], + "blocks": [ + { + "bbox": [ + 390, + 70, + 470, + 178 + ], + "lines": [ + { + "bbox": [ + 390, + 70, + 470, + 178 + ], + "spans": [ + { + "bbox": [ + 390, + 70, + 470, + 178 + ], + "type": "image", + "image_path": "bb8fd95d8db285470d17ff8c64ed1efd30710a80784b817a7ab525b7a91fa13c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 180, + 457, + 189 + ], + "lines": [ + { + "bbox": [ + 403, + 180, + 457, + 189 + ], + "spans": [ + { + "bbox": [ + 403, + 180, + 457, + 189 + ], + "type": "text", + "content": "FSGS+SRGS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "0078b2c7752849c4c36e74686b81059db4266b55d8bfa0daeaf1fb2da9eafef8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 180, + 525, + 190 + ], + "lines": [ + { + "bbox": [ + 503, + 180, + 525, + 190 + ], + "spans": [ + { + "bbox": [ + 503, + 180, + 525, + 190 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 229, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 276 + ], + "type": "text", + "content": "proposed component in S2Gaussian is evaluated by gradually integrating them into the model, revealing their individual contributions to the overall performance. More detailed analyses are as below." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 56, + 291, + 294, + 361 + ], + "blocks": [ + { + "bbox": [ + 65, + 279, + 285, + 290 + ], + "lines": [ + { + "bbox": [ + 65, + 279, + 285, + 290 + ], + "spans": [ + { + "bbox": [ + 65, + 279, + 285, + 290 + ], + "type": "text", + "content": "Table 2. Ablation studies on Mip-NeRF " + }, + { + "bbox": [ + 65, + 279, + 285, + 290 + ], + "type": "inline_equation", + "content": "{360} \\times 4" + }, + { + "bbox": [ + 65, + 279, + 285, + 290 + ], + "type": "text", + "content": " (24 views)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 291, + 294, + 361 + ], + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 361 + ], + "type": "table", + "html": "
VariantPSNR↑SSIM↑LPIPS↓FID↓
Baseline19.790.5200.42987.22
+ Two-Stage20.060.5280.41580.64
+ Gaussian Shuffle Split20.620.5590.37671.29
+ Blur-Free Inconsistency Modeling21.120.6240.34659.63
+ 3D Robust Optimization22.050.6870.29643.51
", + "image_path": "cae117d966f0fcde7b6dd22c6d2e81840587286beff4d269063072400965018e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 366, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 462 + ], + "type": "text", + "content": "Effect of Gaussian Shuffle Split. As shown in Tab. 3, fewer sub-Gaussians lead to suboptimal detail modeling (lower PSNR), while 6 offers the best trade-off, while more sub-Gaussians will not only increase the optimization complexity but also introduce extra memory overhead. Additionally, 6 sub-Gaussians can be symmetrically located along the three principal axes of the original Gaussian, perfectly approximating the original Gaussian representation." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 56, + 481, + 293, + 521 + ], + "blocks": [ + { + "bbox": [ + 61, + 470, + 288, + 480 + ], + "lines": [ + { + "bbox": [ + 61, + 470, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 61, + 470, + 288, + 480 + ], + "type": "text", + "content": "Table 3. Ablation of sub-Gaussian number on Mip-NeRF 360." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 481, + 293, + 521 + ], + "lines": [ + { + "bbox": [ + 56, + 481, + 293, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 293, + 521 + ], + "type": "table", + "html": "
Number456789
Storage Memory (M)6789112137169202
Final PSNR21.7621.9522.0522.0322.0922.08
", + "image_path": "f43729a0ba35848a2bc53aabe3cf50c63a883cba955d1b377c6558521953e602.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": "Effect of Blur-Free Inconsistency Modeling. As illustrated in Fig. 8, IM models the inconsistency of super-resolved images, and while this would introduce blurring, we model this effect explicitly through blur kernel learning, thus reconstructing results that are blur-free and with fine structural details." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 58, + 602, + 115, + 660 + ], + "blocks": [ + { + "bbox": [ + 58, + 602, + 115, + 660 + ], + "lines": [ + { + "bbox": [ + 58, + 602, + 115, + 660 + ], + "spans": [ + { + "bbox": [ + 58, + 602, + 115, + 660 + ], + "type": "image", + "image_path": "5f4cd3efacf71d147c91fe1e6eed51861d7680b170044425e5c817470f88c385.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 660, + 111, + 674 + ], + "lines": [ + { + "bbox": [ + 62, + 660, + 111, + 674 + ], + "spans": [ + { + "bbox": [ + 62, + 660, + 111, + 674 + ], + "type": "text", + "content": "Super-Resolved Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "type": "text", + "content": "Figure 8. Visualization of the super-resolved image, the residual output of IM, the blurred reconstructed image with predicted blur kernels, and the original reconstructed output." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 117, + 602, + 175, + 660 + ], + "blocks": [ + { + "bbox": [ + 117, + 602, + 175, + 660 + ], + "lines": [ + { + "bbox": [ + 117, + 602, + 175, + 660 + ], + "spans": [ + { + "bbox": [ + 117, + 602, + 175, + 660 + ], + "type": "image", + "image_path": "0ba83d76d48cd3c062a222554670627ad166ead0556eea84dabb7b2d7b993052.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 661, + 167, + 674 + ], + "lines": [ + { + "bbox": [ + 125, + 661, + 167, + 674 + ], + "spans": [ + { + "bbox": [ + 125, + 661, + 167, + 674 + ], + "type": "text", + "content": "Inconsistency Modeling" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 177, + 602, + 234, + 660 + ], + "blocks": [ + { + "bbox": [ + 177, + 602, + 234, + 660 + ], + "lines": [ + { + "bbox": [ + 177, + 602, + 234, + 660 + ], + "spans": [ + { + "bbox": [ + 177, + 602, + 234, + 660 + ], + "type": "image", + "image_path": "8af871797106e7d04a37d6a51e043db3a524b0c63658f0b1b470bf813e5c0da9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 661, + 220, + 674 + ], + "lines": [ + { + "bbox": [ + 190, + 661, + 220, + 674 + ], + "spans": [ + { + "bbox": [ + 190, + 661, + 220, + 674 + ], + "type": "text", + "content": "Blur Modeling" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 236, + 602, + 294, + 660 + ], + "blocks": [ + { + "bbox": [ + 236, + 602, + 294, + 660 + ], + "lines": [ + { + "bbox": [ + 236, + 602, + 294, + 660 + ], + "spans": [ + { + "bbox": [ + 236, + 602, + 294, + 660 + ], + "type": "image", + "image_path": "3a5c457213a92274fc2c19b1f534fc660d679d215ceaf394fd0d9099bdf8009e.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 661, + 287, + 674 + ], + "lines": [ + { + "bbox": [ + 242, + 661, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 242, + 661, + 287, + 674 + ], + "type": "text", + "content": "Reconstructed Output" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 229, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 312 + ], + "type": "text", + "content": "Effect of 3D Robust Optimization. Fig. 9 showcases visual comparison with or without 3D robust optimization under corrupted training views. It is observed that 3D robust optimization can effectively avoid unfavorable or erroneous supervision to optimize towards higher quality whereas its absence results in Gaussian primitives simulating corrupted representations." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 317, + 323, + 430, + 383 + ], + "blocks": [ + { + "bbox": [ + 317, + 323, + 430, + 383 + ], + "lines": [ + { + "bbox": [ + 317, + 323, + 430, + 383 + ], + "spans": [ + { + "bbox": [ + 317, + 323, + 430, + 383 + ], + "type": "image", + "image_path": "13b1add5f6b4ed73179bfd78eecd1cd9db316bf870b146973bd34a7fc6492ccd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 388, + 554, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 411 + ], + "type": "text", + "content": "Figure 9. Visual comparison with or without 3D robust optimization (3DRO) under corrupted training views." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 432, + 323, + 553, + 383 + ], + "blocks": [ + { + "bbox": [ + 432, + 323, + 553, + 383 + ], + "lines": [ + { + "bbox": [ + 432, + 323, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 432, + 323, + 553, + 383 + ], + "type": "image", + "image_path": "d4527f81b8a4cac880d10897f73e490148498549bd80caa799a3c42425409c09.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 425, + 440, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 440, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 440, + 438 + ], + "type": "text", + "content": "5. Concluding Remarks" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 445, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 696 + ], + "type": "text", + "content": "In this paper, we propose S2Gaussian, an innovative framework that can efficiently reconstruct high-quality 3D scenes with only sparse and low-resolution input views. In contrast to existing frameworks, S2Gaussian can handle both view sparse and clarity deficient simultaneously, which is more functional and practical in real-world scenarios. The proposed framework begins by optimizing a low-resolution Gaussian representation, which is then densified to initialize high-resolution Gaussians via tailored Gaussian Shuffle Split operation. Subsequently, it cleverly solves the view sparsity problem by rendering pseudo-views from low-resolution Gaussians for super-resolution supervision. Wherein a customized blur-free inconsistency modeling scheme and a 3D robust optimization strategy are further proposed to mitigate multi-view inconsistency and eliminate erroneous updates for better geometric structure and fine-grained details reconstruction. Extensive experiments on multiple benchmarks manifest the effectiveness, superiority, and robustness of our method. We expect this work to provide insights into more complicated 3D reconstruction and steer future research on this Gordian knot." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "text", + "content": "718" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 189 + ], + "type": "text", + "content": "The authors are very indebted to the anonymous referees for their critical comments and suggestions for the improvement of this paper. This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285 and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 201, + 115, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 115, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 115, + 215 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 222, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 222, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 296, + 287 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 290, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 290, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 61, + 290, + 296, + 345 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 347, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 347, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 61, + 347, + 296, + 402 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 404, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 437 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 440, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 294, + 461 + ], + "type": "text", + "content": "[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "type": "text", + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In *Forty-first International Conference on Machine Learning*, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 294, + 563 + ], + "type": "text", + "content": "[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splattering in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 620 + ], + "type": "text", + "content": "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 657 + ], + "type": "text", + "content": "[9] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Zhenzhong Kuang, Yu Jun, Jianping Fan, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[11] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "text", + "content": "[12] Yuqi Han, Tao Yu, Xiaohang Yu, Yuwang Wang, and Qionghai Dai. Super-nerf: View-consistent detail generation for nerf super-resolution. arXiv preprint arXiv:2304.13518, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 218 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[15] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 385 + ], + "type": "text", + "content": "[16] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. Refsr-nerf: Towards high fidelity and super resolution view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8244–8253, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 420 + ], + "type": "text", + "content": "[17] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800–801, 2008." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 467, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 554, + 520 + ], + "type": "text", + "content": "[19] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "text", + "content": "[20] Hyun-kyu Ko, Dongheok Park, Youngin Park, Byeonghyeon Lee, Juhee Han, and Eunbyung Park. Sequence matters: Harnessing video models in super-resolution. arXiv preprint arXiv:2412.11525, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "text", + "content": "[21] Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20561–20570, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 677 + ], + "type": "text", + "content": "[22] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 714 + ], + "type": "text", + "content": "[23] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "719" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 238 + ], + "type": "text", + "content": "[26] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 303 + ], + "type": "text", + "content": "[27] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "type": "text", + "content": "[28] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 381 + ], + "type": "text", + "content": "[29] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "text", + "content": "[30] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[31] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 524 + ], + "type": "text", + "content": "[32] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splattering for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "text", + "content": "[33] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 635 + ], + "type": "text", + "content": "[34] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High quality neural radiance fields using supersampling. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6445-6454, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065–9076, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "text", + "content": "[36] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 150 + ], + "type": "text", + "content": "[37] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 252 + ], + "type": "text", + "content": "[39] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[40] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 " + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "inline_equation", + "content": "\\{\\backslash \\mathrm{deg}\\}" + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": " sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[41] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "text", + "content": "[42] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[43] Youngho Yoon and Kuk-Jin Yoon. Cross-guided optimization of radiance fields with multi-view image superresolution for high-resolution novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12428-12438, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[44] Xiqian Yu, Hanxin Zhu, Tianyu He, and Zhibo Chen. Gaussiansr: 3d gaussian super-resolution with 2d diffusion priors. arXiv preprint arXiv:2406.10111, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 489, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 543 + ], + "type": "text", + "content": "[45] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[46] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[48] Xiaoyun Zheng, Liwei Liao, Jianbo Jiao, Feng Gao, and Ronggang Wang. Surface-sos: Self-supervised object segmentation via neural surface representation. IEEE Transactions on Image Processing, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "720" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "text", + "content": "[50] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "721" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_content_list.json b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a0894a73d2de49b479434d506f2897eb39ac1eaf --- /dev/null +++ b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_content_list.json @@ -0,0 +1,2026 @@ +[ + { + "type": "text", + "text": "S4-Driver: Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation", + "text_level": 1, + "bbox": [ + 91, + 128, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yichen Xie $^{1,\\dagger,*}$ Runsheng Xu $^{2,*}$ Tong He $^{2}$ Jyh-Jing Hwang $^{2}$ Katie Luo $^{3,\\dagger}$", + "bbox": [ + 169, + 200, + 823, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingwei Ji² Hubert Lin² Letian Chen⁴,† Yiren Lu² Zhaoqi Leng²", + "bbox": [ + 218, + 220, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dragomir Anguelov² Mingxing Tan²", + "bbox": [ + 352, + 239, + 663, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 UC Berkeley, 2 Waymo LLC, 3 Cornell University, 4 Georgia Institute of Technology", + "bbox": [ + 158, + 256, + 836, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 324, + 326, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The latest advancements in multi-modal large language models (MLLMs) have spurred a strong renewed interest in end-to-end motion planning approaches for autonomous driving. Many end-to-end approaches rely on human annotations to learn intermediate perception and prediction tasks, while purely self-supervised approaches—which directly learn from sensor inputs to generate planning trajectories without human annotations—often underperform the state of the art. We observe a key gap in the input representation space: end-to-end approaches built on MLLMs are often pretrained with reasoning tasks in 2D image space rather than the native 3D space in which autonomous vehicles plan. To this end, we propose S4-Driver, a scalable self-supervised motion planning algorithm with spatio-temporal visual representation, based on the popular PaLI [9] multimodal large language model. S4-Driver uses a novel sparse volume strategy to seamlessly transform the strong visual representation of MLLMs from perspective view to 3D space without the need to finetune the vision encoder. This representation aggregates multi-view and multiframe visual inputs and enables better prediction of planning trajectories in 3D space. To validate our method, we run experiments on both nuScenes and Waymo Open Motion Dataset (with in-house camera data). Results show that S4-Driver performs favorably against existing supervised multi-task approaches while requiring no human annotations. It also demonstrates great scalability when pretrained on large volumes of unannotated driving logs.", + "bbox": [ + 86, + 356, + 483, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 808, + 220, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The exploration of end-to-end autonomous driving dates back to the 1980s [43], where the motion planning model", + "bbox": [ + 89, + 834, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3a2e1d187e9baec01ed63e52bf09edb33d55e8b892eec6d7070f9f3c53b5193.jpg", + "image_caption": [ + "(a) Pretraining data scaling-up of S4-Driver." + ], + "image_footnote": [], + "bbox": [ + 547, + 311, + 870, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/829ab413ed8ba1f08f3aaac646d8963e7e37fbb6bf9506d1712950fadf5cd6a6.jpg", + "image_caption": [ + "(b) Multi-" + ], + "image_footnote": [], + "bbox": [ + 516, + 444, + 584, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9541e43b9face5cc35a2f5eaa0360ea10c06e441f3875553272b4cbef8206b3a.jpg", + "image_caption": [ + "Multi-task Results" + ], + "image_footnote": [], + "bbox": [ + 586, + 444, + 656, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57397cda1cf2eba4009f403bda065818ec1ff1132dc0e76999395fc21f870468.jpg", + "image_caption": [ + "Figure 1. S4-Driver scaling up with abundant unlabeled driving logs (1a). Different from past efforts that improve MLLM planning via multi-task learning and CoT reasoning (1b), S4-Driver focuses on self-supervised motion planning (1c) without human annotations. While a vanilla PaLI approach with self-supervised training achieves good performance, the full S4-Driver with spatio-temporal visual representation achieves the best performance and scalability (1a).", + "(c) Self-supervised planning framework." + ], + "image_footnote": [], + "bbox": [ + 712, + 444, + 782, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b85419079a725bc40ad065be321cf7c75bf867e9b539565abad03f24d3cf937.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 444, + 825, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90ae78a5d1e2f64a66812289e1cf162e095afc983a8e6c6195acad45ff40a2a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 444, + 897, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "directly predicts control signals or future waypoints based on raw sensor inputs. Due to limited robustness, these early attempts struggle in complex urban situations. Recent neural network capability advances have led to renewed interest in this field [25, 28, 57], especially given the strong generalization ability of Multimodal Large Language Models (MLLMs) [10, 42, 51].", + "bbox": [ + 511, + 625, + 906, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, vanilla application of MLLMs to end-to-end motion planning can hardly exploit their strong visual understanding and reasoning capabilities given the significant difference between motion planning and MLLM pretraining tasks, resulting in inferior planning performance [53, 61]. To narrow this gap, as shown in Fig. 1b, previous methods resort to either multi-task learning, which incorporates diverse perception and prediction tasks into training and inference, or supervised perception pretraining, which utilizes the pretrained autonomous driving perception models as visual tokenizers. However, fine-grained human annotations", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done when they were interns at Waymo", + "bbox": [ + 112, + 875, + 367, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 116, + 888, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1622", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "serve as the bottleneck for both strategies. In contrast, pure self-supervised approaches, while capable of learning directly from sensor inputs and utilizing the abundant unlabeled driving logs, often under-perform the state of the art.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To challenge this status quo, we aim to boost the performance of self-supervised motion planning orthogonal to prior work (Fig. 1c). It only requires raw sensor data and ego-vehicle trajectories for self-supervised model training, without the need of expensive human labels. First of all, we identify the following two main obstacles in this direction:", + "bbox": [ + 89, + 152, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Suboptimal representation format. MLLMs [9, 10] are typically designed for tasks in the 2D image plane. This image space representation limits their 3D reasoning ability from the combination of multi-view images.", + "- Limited data scale. Although it is one of the most widely adopted benchmarks for end-to-end planning, nuScenes [5] only contains fewer than 1k sequences and lacks driving behavior diversities. This limited scales leads to severe over-fitting issues in the finetuning of MLLMs with billion-parameter size." + ], + "bbox": [ + 91, + 244, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we handle the above challenges to better explore the potential of MLLMs. Firstly, we propose S4-Driver, a simple yet effective Scalable Self-Supervised motion planning method with Spatio-temporal visual representation. Built upon a general multimodal large language model [9, 10], we directly predict ego-vehicle waypoints from camera images, eliminating the need for intermediate perception and prediction tasks, and thus facilitating the scale-up of model pretraining leveraging massive unannotated driving logs (Fig. 1a). To address the obstacle of suboptimal visual format, we propose a novel sparse volume representation that enables aggregating visual information from multi-view and multi-frame images. Moreover, this spatio-temporal representation is designed to work with frozen features from the perspective-view vision encoder. It boosts 3D spatio-temporal reasoning capabilities of the model on motion planning, and seamlessly preserves world knowledge in the pretrained visual embeddings of MLLMs. Secondly, to rigorously evaluate our method with sufficient training data, we also utilize large-scale WOMD-PlanningADE benchmark [18] with in-house camera sensor data besides the popular nuScenes dataset [5]. WOMD-PlanningADE is approximately $100 \\times$ larger than nuScenes, so it serves as a more comprehensive benchmark to evaluate model performance in a scalable manner.", + "bbox": [ + 89, + 398, + 482, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include:", + "bbox": [ + 107, + 777, + 369, + 791 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We advocate for a simple self-supervised end-to-end motion planning with MLLMs, solely relying on ego-vehicle trajectory supervision. This approach simplifies the system architecture and facilitates data scaling for training.", + "- We propose a novel sparse volume representation to effectively aggregate multi-view and multi-frame visual information, enhancing the 3D spatio-temporal reasoning" + ], + "bbox": [ + 91, + 795, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capabilities of the MLLM for motion planning.", + "bbox": [ + 527, + 90, + 841, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our algorithm achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks. It also demonstrate great scalability with pretraining.", + "bbox": [ + 514, + 107, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 165, + 663, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs). MLLMs with both language and images modalities have garnered significant attention in recent years [1, 2, 30, 45, 51]. Prior research has focused on integrating powerful Large Language Models (LLMs) [12, 64, 64] with advanced image encoders [19, 45, 63] such as LLaVA [34], PaLI [9, 10], PaliGemma [4], and InstructBLIP [14]. Through instruction tuning or multi-modal finetuning, these models demonstrate impressive performance in multimodal understanding and reasoning. Current developments are leveraging increasingly larger multimodal datasets to further enhance their capabilities in complex perception and generalization tasks [3, 42, 59]. However, despite their strengths, these models have shown limitations in 3D spatial reasoning [11, 22, 40], posing challenges for their application in autonomous driving.", + "bbox": [ + 511, + 190, + 903, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end Autonomous Driving. While the key final step of autonomous driving is planning, traditional systems are often assembled with multiple modules for perception [33, 35, 60] and prediction [41, 54, 65] tasks. To mitigate the information loss and error accumulation across modules, end-to-end driving systems [6, 24, 25, 28, 32, 66] utilize a unified model to predict the ego-vehicle future waypoints or control signals directly from raw sensor inputs. While prioritizing planning, these systems typically still incorporate perception and prediction, requiring explicit supervision for each. Although some prior works [13, 46, 58] have explored pure motion planning without any intermediate tasks, they struggle in complex urban scenarios due to modeling capability limitations.", + "bbox": [ + 511, + 448, + 905, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLMs for Driving. The superior reasoning and generalization abilities of large-scale models are desirable properties for autonomous driving. Some efforts transform the driving scenario into large language model textual prompts [8, 38, 47] or directly process camera images with vision-language models [15, 48, 61]. However, their potential is limited to the small scale of existing benchmarks [5], which only allow partial finetuning [23]. Meanwhile, closed-loop simulators [16] struggle to provide realistic sensor data for end-to-end tasks. Consequently, multi-task joint finetuning [49, 53, 61] or chain-of-though (CoT) [56] inference [37-39] are widely adopted to simplify the reasoning despite the extra annotation requirements brought by perception and prediction modules. Alternatively, some work [15, 52] integrates pretrained perception models [31] to extract BEV features and sends to", + "bbox": [ + 511, + 659, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1623", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "language models as vision tokens. Recently, EMMA [26] leverages powerful Gemini [51] for self-supervised motion planning. Additionally, they develop a mixture of training tasks, including motion planning, 3D object detection, and road graph estimation as well as additional reasoning process for trajectory generation. In contrast, our work focuses on the enhancement of self-supervised motion planning without extra human labels.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Enhancing MLLM for Motion Planning", + "text_level": 1, + "bbox": [ + 89, + 224, + 449, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we adapt a pretrained MLLM for the end-to-end motion planning task. Without loss of generality, we leverage PaLI [10]. However, these techniques are also compatible with other more powerful MLLMs. We introduce a set of design improvements to the MLLM setup and demonstrate correspondingly improved performance on WOMD-Planning-ADE (Fig. 3).", + "bbox": [ + 89, + 250, + 480, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vanilla PaLI as Planner", + "text_level": 1, + "bbox": [ + 89, + 364, + 310, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "End-to-end motion planning models determine the future trajectory $\\mathcal{O}_{T_f}$ of ego-vehicle based on multi-view camera images $\\mathcal{C}$ and high-level behavior command $b$ . The future trajectory consists of the ego-vehicle's location at each future time step in the BEV coordinate, i.e. $\\mathcal{O}_{T_f} = [(x_t,y_t)]_{t=1}^{T_f}$ . The high-level behavior command is necessary to guide the driving direction, similar to a navigation system. Additionally, as verified in prior works [32, 62], the historical states of ego-vehicle $\\mathcal{H}_{T_h}$ are also critical for a smooth and feasible planning result, where we consider the historical locations, velocities, and accelerations as $\\mathcal{H}_{T_h} = [\\mathbf{l}_t,\\mathbf{v}_t,\\mathbf{a}_t]_{t=-1}^{-T_h}$ , i.e.", + "bbox": [ + 89, + 388, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} _ {T _ {f}} = P (\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 584, + 482, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P(\\cdot)$ is the planning model. We provide both historical ego states and high-level commands to the model as text prompts. Locations, velocities, and accelerations are represented directly as floating numbers with two decimals. The predicted future trajectories are then extracted from the decoded text outputs of the model. Example prompts and targets are provided in the supplementary materials.", + "bbox": [ + 89, + 612, + 482, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without preceding perception and prediction tasks, the vanilla PaLI, fine-tuned in a self-supervised manner, achieves reasonable but not ideal performance in motion planning (Fig. 3).", + "bbox": [ + 89, + 718, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hierarchical Planning with Meta-Decision", + "text_level": 1, + "bbox": [ + 89, + 787, + 449, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Directly outputting a future trajectory without any reasoning is challenging for MLLMs. To address this, we apply a coarse-to-fine approach, inspired by Chain-of-Thought (CoT) [56]. It adopts a hierarchical planning method [27], going from semantic decision to numeric planning. We prompt our model to provide a high-level estimation of", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the future acceleration state $D$ first, effectively decomposing the motion planning task into two steps:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = P _ {s 1} \\left(\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b\\right), \\quad \\mathcal {O} _ {T _ {f}} = P _ {s 2} \\left(\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b; \\mathcal {D}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 127, + 903, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define $D$ to encompass four meta-decisions: keep stationary, keep speed, accelerate and decelerate. Notably, unlike previous CoT applications in VLM-based planning that require human annotation for training, such as the scene analysis in DriveVLM [53], we incorporate these meta-decisions as a \"free lunch\" to simplify the motion planning process without requiring any additional annotations. The ground-truth decisions are generated by heuristic rules based on the future ego velocity and acceleration (details in the supplementary materials). Fig. 3 shows this simple design brings non-trivial improvements in planning performance.", + "bbox": [ + 511, + 148, + 906, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Scene Representation in 3D Space", + "text_level": 1, + "bbox": [ + 511, + 337, + 808, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-quality motion planning requires a robust understanding of the surrounding 3D scene, including both static and dynamic elements. While traditionally achieved through separate perception and prediction modules, our self-supervised end-to-end framework relies on the MLLM to implicitly learn this understanding without explicit supervision. However, despite strong 2D reasoning capabilities, MLLMs often struggle with 3D spatial reasoning due to limitations in their perspective view representation and the lack of depth-related tasks in pretraining [7, 36].", + "bbox": [ + 511, + 359, + 906, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. 3D Visual Representation with Dense Volumes", + "text_level": 1, + "bbox": [ + 511, + 516, + 880, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the above limitations, we draw inspiration from previous successful perception tasks [31, 35] and adopt a 3D volume representation. The multi-view feature maps $\\mathcal{F}_{2D} = \\{\\mathbf{f}_{2D}^v\\}_{v=1}^V (\\mathbf{f}_{2D}^v \\in \\mathbb{R}^{H \\times W \\times C})$ are extracted by visual encoder of MLLM, where $V$ is the number of views. We construct a 3D feature volume $\\mathbf{f}_{3D}^{vol} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}$ centered at the ego-vehicle based on the multi-view image features. To avoid introducing complex modules that might disrupt the pretrained MLLM and misalign visual features with the subsequent multimodal encoder-decoder, we employ a lightweight projection method, similar to SimpleBEV [20]. Specifically, for each voxel in the 3D volume, we project its $(x, y, z)$ coordinates to each perspective view $v$ , obtaining corresponding 2D coordinates $(u_v, v_v)$ . We then bilinearly sample the local features from each view at these projected locations. Finally, the voxel's feature representation is computed as the average of the local semantic features from all views where the voxel projects within the image bounds. This process effectively incorporates 3D spatial information while maintaining compatibility with the pretrained MLLM.", + "bbox": [ + 511, + 535, + 906, + 851 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {3 D, (x, y, z)} ^ {v o l} = \\mathbf {f} _ {(x, y, z)} ^ {s e m} + \\operatorname {P o s E m b} (x, y, z) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 856, + 903, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {(x, y, z)} ^ {s e m} = \\operatorname {A v g} \\left(\\left\\{\\text {B i l i n e a r} \\left(\\mathbf {f} _ {2 D} ^ {v}; \\left(u _ {v}, v _ {v}\\right)\\right) \\right\\} _ {v}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 878, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1624", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/739a83147b635f3fa2284d23b820d3da079b7878aa3588d0fc23335bcb8956a3.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed S4-Driver algorithm. We enhance the PaLI model for motion planning by incorporating meta-decision, spatio-temporal visual representation, and multi-decoding aggregation." + ], + "image_footnote": [], + "bbox": [ + 114, + 93, + 879, + 361 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba84bc19dd0f40b56251e9153e600bae71d43fff973a1a56a1bae8e5095dc5d3.jpg", + "image_caption": [ + "Figure 3. A roadmap for enhancing MLLM for planning. The techniques are adopted step-by-step from top to bottom, while shadow items are not adopted in the subsequent steps. We show the performance on WOMD-Planning-ADE (Sec. 4) after including each module." + ], + "image_footnote": [], + "bbox": [ + 117, + 398, + 460, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This simple and efficient projection strategy ensures that the 3D volume features $\\mathbf{f}_{3D}^{vol}$ share a similar distribution with the original multi-view features $\\mathcal{F}_{2D}$ . This similarity facilitates seamless integration with the subsequent multimodal encoder-decoder of the MLLM.", + "bbox": [ + 89, + 746, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, this 3D volume representation leads to some improvement in motion planning performance. Alternatively, we find that reducing the $Z$ -axis using a fully connected layer to obtain a BEV representation yields slightly worse performance since this reduction operation", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "may introduce ambiguity to the scene representation.", + "bbox": [ + 511, + 406, + 862, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Sparse Volume Representation", + "text_level": 1, + "bbox": [ + 511, + 428, + 772, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the 3D volume representation effectively captures spatial information, inherently most of the surrounding 3D space is empty. Moreover, for motion planning, detailed information about objects distant from the road, such as buildings and trees, is less critical. Based on this observation, we propose a sparse volume representation to reduce the number of voxels, enabling higher resolution given memory constraints and improving efficiency.", + "bbox": [ + 511, + 446, + 906, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To determine useful volumes in each scene based on their positions and semantics, we define a gate value $g_{(x,y,z)}^{vol} \\in (0,1)$ for each volume at coordinates $(x,y,z)$ . To obtain this gate, starting from multi-view image features $\\{\\mathbf{f}_{2D}^v\\}_{v=1}^V$ , we reduce its dimension through a fully connect (FC) layer as", + "bbox": [ + 511, + 569, + 908, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {2 D} ^ {\\text {g a t e}, v} = \\operatorname {F C} \\left(\\mathbf {f} _ {2 D} ^ {v}\\right), \\quad \\mathbf {f} _ {2 D} ^ {\\text {g a t e}, v} \\in \\mathbb {R} ^ {H \\times W \\times C ^ {\\prime}}, C ^ {\\prime} \\ll C. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 671, + 906, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we construct a reduced-dimension volume feature $\\mathbf{f}_{3D}^{gate} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C'}$ from $\\{\\mathbf{f}_{2D}^{gate,v}\\}_{v=1}^V$ same as Eq. 3. A smaller channel number allows for larger volume resolution and is sufficient to indicate whether a volume is related to motion planning. Afterward, the gate is derived from $\\mathbf{f}_{3D}^{gate}$ through a small MLP module as follows.", + "bbox": [ + 511, + 702, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {(x, y, z)} ^ {v o l} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\mathbf {f} _ {3 D, (x, y, z)} ^ {g a t e}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 804, + 903, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we can easily select $M$ volumes $(M\\ll X\\cdot Y\\cdot Z)$ with largest gate values with coordinates $\\{(x_i,y_i,z_i)\\}_{i = 1}^M$", + "bbox": [ + 511, + 839, + 903, + 871 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we do not have access to ground-truth occupancy state, we propose learning the gate values implicitly. We", + "bbox": [ + 511, + 871, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1625", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "assume that regions with small gate values should be empty or irrelevant to planning. For these vacant space, we assign a learnable feature $\\mathbf{f}_{vac} \\in \\mathbb{R}^C$ . We expect the model can learn the gate value via the trade-off between the semantic features and this vacant feature at each 3D position, so we fetch features for selected sparse volumes as", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} = \\hat {\\mathbf {f}} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} + \\operatorname {P o s E m b} (x, y, z) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 190, + 482, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} _ {3 D, (x, y, z)} ^ {s p a r s e} = g _ {(x, y, z)} ^ {v o l} \\cdot \\mathbf {f} _ {(x, y, z)} ^ {s e m} + \\left(1 - g _ {(x, y, z)} ^ {v o l}\\right) \\cdot \\mathbf {f} _ {v a c} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 214, + 482, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{(x,y,z)}^{sem}$ is same as Eq. 4 and $(x,y,z)\\in \\{(x_i,y_i,z_i)\\}_{i = 1}^M$ are selected volumes with large gate values. When the sparse volume features $\\mathbf{f}_{3D,(x,y,z)}^{sparse}$ are fed into the subsequent multimodal encoder, they notably boost the planning performance (Fig. 3) by simultaneously providing explicit 3D spatial cues of surrounding scenes and concentrate on the important regions adaptively.", + "bbox": [ + 89, + 244, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Local Feature Aggregation in 3D Space", + "text_level": 1, + "bbox": [ + 89, + 361, + 410, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the lack of depth information, the lifting process in Eq. 3 or 7 results in duplicate volume features along each camera ray. This spatial ambiguity can be mitigated by 3D local operations like convolutions [20] or deformable attention [31] as demonstrated in previous work. However, in our MLLM framework, it is less economical to insert extra local operations. Instead, we inject some relative position bias into the existing multimodal encoder by tailoring the self-attention as follows.", + "bbox": [ + 89, + 378, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} - b (D)\\right) V \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 522, + 482, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $M$ sparse volumes with coordinates $\\{(x_i, y_i, z_i)\\}_{i=1}^M$ , the distance matrix $D \\in \\mathbb{R}^{M \\times M \\times 3}$ is the distance along $xyz$ -axes between each pair of sparse volumes, i.e. $D_{ij} = (x_j - x_i, y_j - y_i, z_j - z_i)$ . The bias is computed with the function $b(\\Delta x, \\Delta y, \\Delta z) = b_x(\\Delta x) + b_y(\\Delta y) + b_z(\\Delta z)$ , where $\\Delta x, \\Delta y, \\Delta z$ are divided into several bins and mapped to a learnable bias value for each bin through $b_x(\\cdot), b_y(\\cdot), b_z(\\cdot)$ . We also apply a separate bias between the text tokens with 1D position $\\{p_i\\}$ .", + "bbox": [ + 89, + 566, + 483, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This relative position bias elegantly inserts local inductive bias to pretrained global self-attention modules with minimal additional costs. It leads to a notable performance gain in Fig. 3 by facilitating the aggregation of local information in 3D space, enhancing scene understanding and spatial reasoning.", + "bbox": [ + 89, + 704, + 483, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-frame Temporal Fusion", + "text_level": 1, + "bbox": [ + 89, + 803, + 357, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-frame input combination helps to compensate for the lack of depth cues in camera images. We extend our sparse volume representation to aggregate mult-frame temporal information by incorporating $T$ historical frames with an interval of 0.5s. Given images from the total $T + 1$ frames,", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 5 is applied to each frame separately to obtain the multi-view image features $\\mathbf{f}_{2D,t}^{gate,v}$ , $t = 0, -1, \\dots, -T$ , $v = 1, \\dots, V$ . After ego-motion compensation, we construct a gate feature volume $\\mathbf{f}_{3D,t}^{gate}$ based on the current ego-vehicle coordinates and per-frame image features. Multi-frame volume features are concatenated along the channel dimension to generate gate value.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. g _ {(x, y, z)} ^ {v o l} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\operatorname {C o n c a t} \\left([ \\mathbf {f} _ {3 D, t, (x, y, z)} ^ {g a t e} ] _ {t = 0} ^ {- T}\\right)\\right)\\right) \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 203, + 903, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar with Sec. 3.3.2, we select $M$ volumes based on the gate value. Following Eq. 7, the volume features are fetched from each frame image feature map separately as $\\mathbf{f}_{3D,t,(x,y,z)}^{sparse}$ , $t = 0,\\dots , - T$ , which are fused with an FC-layer as the sparse volume feature with temporal awareness.", + "bbox": [ + 511, + 242, + 905, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} = \\operatorname {F C} \\left(\\operatorname {C o n c a t} \\left(\\left[ \\mathbf {f} _ {3 D, t, (x, y, z)} ^ {\\text {s p a r s e}} \\right] _ {t = 0} ^ {- T}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 325, + 903, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 3, temporal fusion helps to improve motion planning performance by facilitating the environment understanding.", + "bbox": [ + 511, + 366, + 905, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Voting for Planning via Multi-Decoding", + "text_level": 1, + "bbox": [ + 511, + 404, + 852, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MLLMs are prone to assigning high confidence to the simple future behaviors in motion planning, e.g., keeping stationary. To relieve this bias, we aggregate multiple outputs and use voting to obtain the final planning output. We resort to nucleus sampling [21] to produce multiple future trajectories for the ego-vehicle, denoted as $\\{\\mathcal{O}_{T_f}^k\\}_{k=1}^K$ . They are simply averaged to produce a unique planning result as follows: $\\mathcal{O}_{T_t} = [(x_t, y_t)]_{t=1}^{T_f}$ .", + "bbox": [ + 511, + 426, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(x _ {t}, y _ {t}\\right) = \\left(\\frac {1}{K} \\sum_ {k = 1} ^ {K} x _ {t} ^ {k}, \\frac {1}{K} \\sum_ {k = 1} ^ {K} y _ {t} ^ {k}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 556, + 903, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This unweighted average mitigates the MLLM's bias towards simple behaviors. Fig. 3 demonstrates the notable performance gain with this simple multi-decoding aggregation approach.", + "bbox": [ + 511, + 606, + 905, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Scaling to Large-scale Raw Driving Logs", + "text_level": 1, + "bbox": [ + 511, + 674, + 862, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The self-supervised training allows our proposed S4-Driver to scale to large-scale driving logs amounts without requiring human annotations. To exploit the potential of MLLM-based planners, we pretrain the model on our internal dataset. Results in Fig. 3 demonstrate that S4-Driver achieves notable performance gains in challenging tail behaviors due to large-scale pretraining.", + "bbox": [ + 511, + 696, + 905, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Waymo Open Motion Dataset for Planning", + "text_level": 1, + "bbox": [ + 511, + 814, + 895, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the large-scale training and evaluation of planning algorithms with large models, we design a WOMD-Planning-ADE benchmark based on WOMD dataset [18], and compare with the existing datasets in Tab. 1.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1626", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d1d00dedf02e423018ade78b06c1e52ac24687376d81ee985d8793eb8284c51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkSizeData FormatMetrics
SequenceLengthFrequencyPlanning horizonSample-wiseBehavior-wise
BDD-X [29]7k77 h10 Hzcontrol signal✓✗
nuScenes [5]1k5.5 h2 Hz3s✓✗
WOMD-Planning-ADE103k574 h10 Hz5s✓✓
", + "bbox": [ + 93, + 88, + 480, + 141 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparisons between WOMD-Planning-ADE and existing counterparts. WOMD-Planning-ADE has a significantly larger size and more comprehensive behavior-wise metrics.", + "bbox": [ + 89, + 142, + 482, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f6553fb36d69809aef5a53d32a4bb7f8d2a4628d9d0ae019004db293332b883.jpg", + "image_caption": [ + "Figure 4. High-level behavior commands for motion planning." + ], + "image_footnote": [], + "bbox": [ + 114, + 202, + 454, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7fcdefac43c64057dd57f7e4490847e95b28d7981f80f3001cb1004791109d8f.jpg", + "image_caption": [ + "(a) Trajectory" + ], + "image_footnote": [], + "bbox": [ + 96, + 318, + 161, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/137fa8d8e21b0ffbd4d73a760e190f6ce2e9ca7d7478bd2ad8a03286a5c07c36.jpg", + "image_caption": [ + "(b) Velocity" + ], + "image_footnote": [], + "bbox": [ + 166, + 328, + 267, + 390 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10e1af860f183a1bf01a55bc3beda5fa15b02ead192d6b3ee01b34eef48c9a38.jpg", + "image_caption": [ + "(c) Acceleration", + "Figure 5. Statistical distributions of WOMD-Planning-ADE. We show the distributions of ego-vehicle trajectories, velocities, accelerations and each behavior scenario sizes in the validation set." + ], + "image_footnote": [], + "bbox": [ + 272, + 328, + 380, + 390 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/889b62746d6df0156b0bc997472c14a36afe55e270441e34d622b4e2796f0259.jpg", + "image_caption": [ + "(d) Behavior" + ], + "image_footnote": [], + "bbox": [ + 383, + 316, + 478, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview. It contains $103k$ real-world driving scenarios, covering diverse urban and suburban scenes. They are further divided into 9s samples with 1s history and 8s future. For end-to-end planning evaluation, besides the ego-vehicle trajectories in each sample as groundtruth, we also have the following key items in the dataset:", + "bbox": [ + 89, + 474, + 482, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(In-house) Camera Data. Most end-to-end planning methods [28, 49, 53, 61] relies on the camera images as model inputs. In our dataset, each frame contains images captured by eight multi-view cameras.", + "bbox": [ + 89, + 566, + 483, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-level Behavior Commands. Like navigation systems for humans (e.g. Google Map), the end-to-end planning system also requires navigation signals to indicate the direction to go. We consider six high-level behavior commands (Fig. 4), namely go straight forward, go straight left, go straight right, do left turn, do right turn, do left U-turn. They can cover diverse driving situations in real world, e.g. \"go straight right\" describes the case to leave the high-way. Instead of simply considering the last step position [6, 25], we decide the behavior commands based on long-term future trajectories, which can handle the low-speed or stopping situations. Details of this heuristic rule are included in the supplementary materials.", + "bbox": [ + 89, + 628, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The imbalance of data distribution is inevitable in real driving scenarios. For example, in the WOMD-Planning-ADE benchmark, moving straight and stopping account for more than $70\\%$ of all the samples, as reflected by Fig. 5. In this case, we argue that current widely", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "used sample-wise average displacement error and collision rates cannot comprehensively reflect the performance of motion planning algorithms since challenging but less frequent behaviors, like turning, are overwhelmed by simple straightforward moving scenarios. To this end, we propose novel behavior-wise metrics similar to the mAP metric in prediction [18]. For example, we denote the behavior-wise average displacement error as $bADE$ defined as follows.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nb A D E = \\frac {1}{| \\mathcal {B} |} \\sum_ {b \\in \\mathcal {B}} A D E _ {b} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 220, + 906, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $ADE_{b}$ is the ADE metric for a specific behavior. Concretely, $|\\mathcal{B}| = 7$ behaviors are considered - including the six high-level commands (Fig. 4) and an additional stop behavior1. These behavior-wise metrics make up for the shortcomings of previous sample-wise metrics by paying more attention to those scarce but safety-critical behaviors.", + "bbox": [ + 511, + 265, + 906, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 369, + 645, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 393, + 730, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model and finetuning. Unless otherwise specified, we construct our model from a pretrained PaLI3-5B model [10], which includes a ViT-G (2B) vision encoder [17] and multimodal encoder-decoder (3B) [50]. Our default setting freezes the ViT encoder and finetune only the inserted modules and the multimodal encoder-decoder. Finetuning the model on the WOMD-Planning-ADE dataset requires approximately 2.5 days on 128 Google Cloud TPU v4 using a batch size of 256 and a learning rate of 3e-3.", + "bbox": [ + 511, + 416, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate S4-Driver on both WOMD-Planning-ADE (Sec. 4) and nuScenes [5] benchmarks. Given the limited size of nuScenes, which is insufficient for training a multimodal large language model, we also attempt to finetune the model on nuScenes using weights pretrained on WOMD-Planning-ADE.", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results and Comparison", + "text_level": 1, + "bbox": [ + 511, + 651, + 784, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NuScenes dataset. To widely compare with existing end-to-end planning methods in the community, we evaluate S4-Driver on nuScenes dataset [5]. Tab. 2 show that S4-Driver notably outperforms all the previous algorithms despite its self-supervised training format. Unlike existing works, S4-Driver does not require any perception pretraining or human annotations. This self-supervision nature enables S4-Driver to utilize all the accessible raw trajectory data, whereas these easily collectable raw driving logs offer limited benefit to other algorithms without a labeling process.", + "bbox": [ + 511, + 672, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WOMD-Planning-ADE benchmark. In Tab. 3, we primarily compare S4-Driver with the vanilla PaLI3-5B base", + "bbox": [ + 511, + 824, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We do not include \"stop\" in the high-level behavior commands since navigation system cannot provide it and this may leak future information like traffic lights. Details are explained in supplementary materials.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1627", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8967db0a8878d56840888be7c15210ed7adfbf558acf0a973740365f29719164.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsHuman AnnotationsExtra DataL2(m)
perception predictiondata labels1.0s2.0s3.0sAvg1,2,3s
Multi-task end-to-end approaches
UniAD [25]✓✓XX0.420.640.910.66
VAD [28]✓✓XX0.170.340.600.37
PARA-Drive [57]✓✓XX0.250.460.740.48
Multimodal large language model approaches
GPT-Driver [38]✓✓XX0.200.400.700.44
DriveVLM [53]✓✓✓✓0.180.340.680.40
OmniDrive [55]✓✓✓✓0.140.290.550.33
Self-supervised approaches
DriveVLM (w/o CoT) [53]XX✓X0.190.410.890.49
S4-Driver (ours)XXXX0.160.340.630.38
S4-Driver* (ours)XX✓X0.130.280.510.31
", + "bbox": [ + 142, + 88, + 852, + 364 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/86cb56c90a344896416dd477797aec9c48f8c38a7211e6cf60259da25c3cd353.jpg", + "table_caption": [ + "Table 2. NuScenes benchmark results. Most previous works require human annotations as additional supervision. S4-Driver* is pretrained on WOMD-Planning-ADE, while many prior MLLM-based methods also utilize extra data and labels due to the limited scale of nuScenes." + ], + "table_footnote": [], + "table_body": "
MethodsSample-wise MetricsBehavior-wise Metrics
ADE@1sADE@3sADE@5sbADE@1sbADE@3sbADE@5s
Vanilla PaLI0.0340.2770.7980.0510.3921.069
S4-Driver (ours)0.0310.2470.6930.0490.3500.928
S4-Driver* (ours)0.0290.2330.6550.0440.3130.830
Models that utilize high-quality objects, tracks, and roadgraph as inputs rather than raw camera images.
MotionLM [46]0.0450.2660.6970.0660.3880.978
", + "bbox": [ + 140, + 402, + 854, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. WOMD-Planning-ADE benchmark results. “*” denotes methods pretrained on the internal dataset.", + "bbox": [ + 178, + 537, + 815, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line (Sec. 3.1) and modular algorithm MotionLM [46]. Compared with vanilla PaLI3-5B, a notable gap is witnessed for both sample-wise and behavior-wise metrics. For reference, we also adapt a latest motion prediction algorithms MotionLM [46] (enhanced internally reproduced version) for planning task by only predicting the future trajectories of ego-vehicle and injecting the high-level command to the model. Because it utilizes auto-labelled high-quality objects, tracklets, and roadgraph signals as model inputs [44], a direct comparison with our end-to-end approaches is not equitable. However, as shown in Tab. 3, S4-Driver still achieves favorable performance compared to MotionLM, particularly in behavior-wise metrics even if S4-Driver uses only raw camera images as inputs.", + "bbox": [ + 88, + 563, + 482, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 89, + 787, + 192, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 6 visualizes planning results in diverse scenarios. Our proposed S4-Driver can determine the future ego-behavior based on the traffic lights and road lane. It can also complete challenging behaviors and deal with different lighting conditions. We will include more visualizations in supplementary materials.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meta-decision reliability. Fig. 7 demonstrates the accuracy of meta-decision prediction (Sec. 3.2) on WOMD-Planning-ADE validation set. Across all behaviors, the model provides reliable meta-decision estimations. Without any human annotations, this preliminary prediction can simplify the derivation of numeric motion planning.", + "bbox": [ + 511, + 563, + 906, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Distribution of sparse volumes. We visualize the distribution of self-supervised learned sparse volumes along x-axis and y-axis over WOMD-Planning-ADE validation set in Fig. 8. From back to front, the sparse volumes concentrate on the close front region. From left to right, the sparse volumes cover all the regions, since there are turning scenarios, but most volumes focus on the middle region. These distributions are consistent with human driving experiences.", + "bbox": [ + 511, + 655, + 908, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 787, + 679, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section presents experiments conducted on the WOMD planning benchmark to clarify the roles of specific modules within our S4-Driver algorithm. More ablation studies are available in the supplementary material.", + "bbox": [ + 511, + 809, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MLLM inputs. In Tab. 5, we analyze the roles of camera images and historical ego-states separately using the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1628", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b57a38af5edcfefae42efab50e00d420fbd94f9e9ba31ad33318457a33442860.jpg", + "image_caption": [ + "(a) The ego-vehicle decelerates to wait for the red lights." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 218, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a242d6bd677ef62f071aa003b89b063187116f455eb4ce94d24c104620834821.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 90, + 346, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dca1783964db38094cc306b06a4261880974014f040987404bef33518bfbac5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 90, + 473, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b394c33a6a30dddd779d79a8784f8076677617c2f3ae8eda6cfbf391c74dedd6.jpg", + "image_caption": [ + "(b) The ego-vehicle conducts a U-turn at the intersection." + ], + "image_footnote": [], + "bbox": [ + 96, + 154, + 218, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a4b374df4ad16769abdb80c57e1b29e54b5901114937703b0a8d855d8b3c3a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 154, + 346, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82111de3a2686af243440848c0e561e66a92d6f806e1df59e29d41c085dbb96b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 154, + 473, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48a82813b9bf5ff020bd71902c308c3fd7f7e1d2298d0f0598d156d79cb8183e.jpg", + "image_caption": [ + "(c) The ego-vehicle keeps the lane under bad lighting condition." + ], + "image_footnote": [], + "bbox": [ + 96, + 218, + 218, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b38df7a6646238cd0619c3f1243d02fcfba905e8207dd20506fb2a94e170d6a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 219, + 346, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b62687f28a1e32765934a27c0f71aa93d9728c5f39a7c1ee383978c374e334d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 219, + 473, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/761714070865e371bac2e28b98f29cb4dff11e4cab9163eda65c17ad9a56b946.jpg", + "image_caption": [ + "Figure 6. Qualitative results of motion planning. We show the front left, front, front right cameras for each case.", + "Figure 7. Meta-decision prediction accuracy for each behavior.", + "Figure 8. Distribution of sparse volumes along x- and y-axes." + ], + "image_footnote": [], + "bbox": [ + 96, + 315, + 279, + 448 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2c00654c44c01c12d3ddc039617e87f147d09d33d691adbef46ea6d5e4173f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 315, + 475, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vanilla PaLI model. Consistent with [32, 62], the importance of ego-states is acknowledged. Meanwhile, different from their observations on nuScenes benchmark, camera inputs also make great differences on WOMD-Planning-ADE. We assume that WOMD-Planning-ADE covers more diverse driving scenarios including many severe changes of speed and directions, which attaches importance to the sensor data. This also showcases the advantages of WOMD-Planning-ADE for comprehensive evaluation. Tab. 5 also shows that random initialized model cannot converge without MLLM pretraining. Despite different domains, S4-Driver can profit from the strong reasoning ability from large-scale MLLM pretraining on general tasks.", + "bbox": [ + 88, + 489, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MLLM capability. In addition to the PaLI3-5B [10] used in other parts, we also tailor PaLI2-3B [9] for motion planning with the same techniques. PaLI2-3B is weaker in parameter numbers (3B vs. 5B), architecture, pretraining manner, and input resolution (224 vs. 448). As in Tab. 4, S4-Driver based on PaLI2-3B demonstrates much inferior performance compared to that based on PaLI3-5B. We conduct experiments on two different training data scales on WOMD-Planning-ADE, i.e. 20k (nuScenes scale) vs. 400k (full WOMD-Planning-ADE). The gap is particularly large given enough training data. It also justifies the necessity to conduct experiments on large-scale dataset, which can fully unleash the potential of powerful MLLMs.", + "bbox": [ + 89, + 686, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse volume resolution. Tab. 6 shows the results with", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a77ea038f66a33afb74740f795c69a487e6bbb33f67d9115e8fd1f4430a183af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MLLMWOMD-Planning-ADE Subset (20k)Full WOMD-Planning-ADE (400k)
ADE@5sbADE@5sADE@5sbADE@5s
PaLI2-3B [9]1.2411.8461.0351.522
PaLI3-5B [10]1.0751.4950.6930.928
", + "bbox": [ + 516, + 89, + 903, + 135 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b0161bd8d2aec398decb879da728f54d9c464a7a9c1f1668f8e3f1922dc8fd78.jpg", + "table_caption": [ + "Table 4. Ablation studies on planning performance scalability of our methods w.r.t. model and data sizes. We experiment with two MLLMs on two different training data sizes." + ], + "table_footnote": [], + "table_body": "
MethodADE@5sbADE@5s
vanilla PaLI0.7981.069
w/o historical states2.0092.250
w/o camera images1.0811.448
w/o MLLM pretrainingfail to converge
", + "bbox": [ + 532, + 190, + 883, + 294 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cc79f90d70258c4873acae40eab2651bd842535e023f71f7d68bfc793bb80573.jpg", + "table_caption": [ + "Table 5. Ablation studies on MLLM inputs." + ], + "table_footnote": [], + "table_body": "
ResolutionADE@5sbADE@5s
2m × 2m × 2m0.7981.201
1m × 1m × 2m0.7501.005
1m × 1m × 1m0.7651.051
", + "bbox": [ + 553, + 321, + 861, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies on sparse volume resolutions.", + "bbox": [ + 544, + 398, + 872, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different resolution for sparse volumes with same number of sparse volumes ( $M = 6000$ ). In line with Fig. 3, low resolution leads to relatively worse performance since it limits the precision of 3D spatial reasoning. Interestingly, higher resolution along $z$ -axis does not definitely improve the model performance since the motion planning mainly works on the $xy$ -plane and too low sparsity ratio tends to make the optimization unstable.", + "bbox": [ + 511, + 424, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 558, + 782, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces S4-Driver, a scalable self-supervised motion planning framework for autonomous driving that utilizes Multimodal Large Language Models (MLLMs). To enhance 3D reasoning in MLLMs, we propose a novel sparse volume representation, which enables improved spatio-temporal reasoning through the aggregation of multi-view and multi-frame image inputs. Additionally, we design behavior-wise metrics for comprehensive evaluation on the large-scale WOMD-Planning-ADE benchmark. S4-Driver achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks, without requiring any human annotations. This demonstrates the potential of self-supervised learning for end-to-end autonomous driving. Future work will involve applying our methods to other powerful MLLM architectures. Our contributions are orthogonal to previous works, such as those that explore multi-task learning or CoT reasoning [49, 55]. Combining our large-scale self-supervised learning approach with supervised finetuning on targeted small-scale labeled data may further enhance the final performance and interpretability of the system.", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1629", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2", + "[4] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 2", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6", + "[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2, 6", + "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 3", + "[8] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2", + "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations. 1, 2, 8", + "[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2, 3, 6, 8", + "[11] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 2", + "[12] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2", + "[13] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2", + "[15] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's-eye-view injected multi-modal large models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2024. 2", + "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[18] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5, 6", + "[19] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2", + "[20] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-bev: What really matters for multi-sensor bev perception? In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2759-2765. IEEE, 2023. 3, 5", + "[21] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5, 1", + "[22] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances" + ], + "bbox": [ + 514, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1630", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in Neural Information Processing Systems, 36:20482-20494, 2023. 2", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[24] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2", + "[25] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 6, 7", + "[26] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to-end multimodal model for autonomous driving, 2024. 3", + "[27] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 3", + "[28] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 6, 7", + "[29] Jinkyu Kim, Anna Rohrbach, Trevor Darrell, John Canny, and Zeynep Akata. Textual explanations for self-driving vehicles. In Proceedings of the European conference on computer vision (ECCV), pages 563-578, 2018. 6", + "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2", + "[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 5", + "[32] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 2, 3, 8", + "[33] Bencheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr:" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Structured modeling and learning for online vectorized HD map construction. In The Eleventh International Conference on Learning Representations. 2", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[35] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3", + "[36] Xianzheng Ma, Yash Bhalgat, Brandon Smart, Shuai Chen, Xinghui Li, Jian Ding, Jindong Gu, Dave Zhenyu Chen, Songyou Peng, Jia-Wang Bian, et al. When llms step into the 3d world: A survey and meta-analysis of 3d tasks via multi-modal large language models. arXiv preprint arXiv:2405.10255, 2024. 3", + "[37] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2", + "[38] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 7", + "[39] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 2", + "[40] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2", + "[41] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 2", + "[42] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2", + "[43] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 1", + "[44] C. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa T. Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6130-6140, 2021. 7", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[46] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 7, 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1631", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2", + "[48] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2", + "[49] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In First Vision and Language for Autonomous Driving and Robotics Workshop. 2, 6, 8, 4", + "[50] Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, et al. UI2: Unifying language learning paradigms. arXiv preprint arXiv:2205.05131, 2022. 6", + "[51] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 3", + "[52] Thomas Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. In Workshop on Language and Robot Learning: Language as an Interface. 2", + "[53] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7", + "[54] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 2", + "[55] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 7, 8", + "[56] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3", + "[57] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 7", + "[58] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2", + "[59] xAI. Grok-1.5 vision preview, 2024. 2", + "[60] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 2", + "[61] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 1, 2, 6", + "[62] Jiang-Tian Zhai, Ze Feng, Jihao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 3, 8", + "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2", + "[64] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2", + "[65] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 2", + "[66] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 656 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1632", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_model.json b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed5fdaecd726cc9409d2209051b0d14606e1a60f --- /dev/null +++ b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_model.json @@ -0,0 +1,2884 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.905, + 0.177 + ], + "angle": 0, + "content": "S4-Driver: Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.202, + 0.825, + 0.222 + ], + "angle": 0, + "content": "Yichen Xie\\(^{1,\\dagger,*}\\) Runsheng Xu\\(^{2,*}\\) Tong He\\(^{2}\\) Jyh-Jing Hwang\\(^{2}\\) Katie Luo\\(^{3,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.222, + 0.797, + 0.24 + ], + "angle": 0, + "content": "Jingwei Ji² Hubert Lin² Letian Chen⁴,† Yiren Lu² Zhaoqi Leng²" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.24, + 0.665, + 0.258 + ], + "angle": 0, + "content": "Dragomir Anguelov² Mingxing Tan²" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.257, + 0.838, + 0.276 + ], + "angle": 0, + "content": "1 UC Berkeley, 2 Waymo LLC, 3 Cornell University, 4 Georgia Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.325, + 0.327, + 0.34 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.357, + 0.485, + 0.782 + ], + "angle": 0, + "content": "The latest advancements in multi-modal large language models (MLLMs) have spurred a strong renewed interest in end-to-end motion planning approaches for autonomous driving. Many end-to-end approaches rely on human annotations to learn intermediate perception and prediction tasks, while purely self-supervised approaches—which directly learn from sensor inputs to generate planning trajectories without human annotations—often underperform the state of the art. We observe a key gap in the input representation space: end-to-end approaches built on MLLMs are often pretrained with reasoning tasks in 2D image space rather than the native 3D space in which autonomous vehicles plan. To this end, we propose S4-Driver, a scalable self-supervised motion planning algorithm with spatio-temporal visual representation, based on the popular PaLI [9] multimodal large language model. S4-Driver uses a novel sparse volume strategy to seamlessly transform the strong visual representation of MLLMs from perspective view to 3D space without the need to finetune the vision encoder. This representation aggregates multi-view and multiframe visual inputs and enables better prediction of planning trajectories in 3D space. To validate our method, we run experiments on both nuScenes and Waymo Open Motion Dataset (with in-house camera data). Results show that S4-Driver performs favorably against existing supervised multi-task approaches while requiring no human annotations. It also demonstrates great scalability when pretrained on large volumes of unannotated driving logs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.222, + 0.824 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.483, + 0.866 + ], + "angle": 0, + "content": "The exploration of end-to-end autonomous driving dates back to the 1980s [43], where the motion planning model" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.313, + 0.871, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.432, + 0.813, + 0.443 + ], + "angle": 0, + "content": "(a) Pretraining data scaling-up of S4-Driver." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.445, + 0.586, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.487, + 0.576, + 0.497 + ], + "angle": 0, + "content": "(b) Multi-" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.445, + 0.658, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.481, + 0.697, + 0.487 + ], + "angle": 0, + "content": "Multi-task Results" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.445, + 0.783, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.445, + 0.826, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.445, + 0.898, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.499, + 0.906, + 0.61 + ], + "angle": 0, + "content": "Figure 1. S4-Driver scaling up with abundant unlabeled driving logs (1a). Different from past efforts that improve MLLM planning via multi-task learning and CoT reasoning (1b), S4-Driver focuses on self-supervised motion planning (1c) without human annotations. While a vanilla PaLI approach with self-supervised training achieves good performance, the full S4-Driver with spatio-temporal visual representation achieves the best performance and scalability (1a)." + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.488, + 0.9, + 0.499 + ], + "angle": 0, + "content": "(c) Self-supervised planning framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.732 + ], + "angle": 0, + "content": "directly predicts control signals or future waypoints based on raw sensor inputs. Due to limited robustness, these early attempts struggle in complex urban situations. Recent neural network capability advances have led to renewed interest in this field [25, 28, 57], especially given the strong generalization ability of Multimodal Large Language Models (MLLMs) [10, 42, 51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, vanilla application of MLLMs to end-to-end motion planning can hardly exploit their strong visual understanding and reasoning capabilities given the significant difference between motion planning and MLLM pretraining tasks, resulting in inferior planning performance [53, 61]. To narrow this gap, as shown in Fig. 1b, previous methods resort to either multi-task learning, which incorporates diverse perception and prediction tasks into training and inference, or supervised perception pretraining, which utilizes the pretrained autonomous driving perception models as visual tokenizers. However, fine-grained human annotations" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.368, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done when they were interns at Waymo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.889, + 0.227, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.368, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1622" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "serve as the bottleneck for both strategies. In contrast, pure self-supervised approaches, while capable of learning directly from sensor inputs and utilizing the abundant unlabeled driving logs, often under-perform the state of the art." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.243 + ], + "angle": 0, + "content": "To challenge this status quo, we aim to boost the performance of self-supervised motion planning orthogonal to prior work (Fig. 1c). It only requires raw sensor data and ego-vehicle trajectories for self-supervised model training, without the need of expensive human labels. First of all, we identify the following two main obstacles in this direction:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.246, + 0.483, + 0.306 + ], + "angle": 0, + "content": "- Suboptimal representation format. MLLMs [9, 10] are typically designed for tasks in the 2D image plane. This image space representation limits their 3D reasoning ability from the combination of multi-view images." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.307, + 0.483, + 0.397 + ], + "angle": 0, + "content": "- Limited data scale. Although it is one of the most widely adopted benchmarks for end-to-end planning, nuScenes [5] only contains fewer than 1k sequences and lacks driving behavior diversities. This limited scales leads to severe over-fitting issues in the finetuning of MLLMs with billion-parameter size." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.246, + 0.483, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In this paper, we handle the above challenges to better explore the potential of MLLMs. Firstly, we propose S4-Driver, a simple yet effective Scalable Self-Supervised motion planning method with Spatio-temporal visual representation. Built upon a general multimodal large language model [9, 10], we directly predict ego-vehicle waypoints from camera images, eliminating the need for intermediate perception and prediction tasks, and thus facilitating the scale-up of model pretraining leveraging massive unannotated driving logs (Fig. 1a). To address the obstacle of suboptimal visual format, we propose a novel sparse volume representation that enables aggregating visual information from multi-view and multi-frame images. Moreover, this spatio-temporal representation is designed to work with frozen features from the perspective-view vision encoder. It boosts 3D spatio-temporal reasoning capabilities of the model on motion planning, and seamlessly preserves world knowledge in the pretrained visual embeddings of MLLMs. Secondly, to rigorously evaluate our method with sufficient training data, we also utilize large-scale WOMD-PlanningADE benchmark [18] with in-house camera sensor data besides the popular nuScenes dataset [5]. WOMD-PlanningADE is approximately \\(100 \\times\\) larger than nuScenes, so it serves as a more comprehensive benchmark to evaluate model performance in a scalable manner." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.778, + 0.37, + 0.792 + ], + "angle": 0, + "content": "In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We advocate for a simple self-supervised end-to-end motion planning with MLLMs, solely relying on ego-vehicle trajectory supervision. This approach simplifies the system architecture and facilitates data scaling for training." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.857, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel sparse volume representation to effectively aggregate multi-view and multi-frame visual information, enhancing the 3D spatio-temporal reasoning" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.092, + 0.842, + 0.108 + ], + "angle": 0, + "content": "capabilities of the MLLM for motion planning." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.108, + 0.905, + 0.152 + ], + "angle": 0, + "content": "- Our algorithm achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks. It also demonstrate great scalability with pretraining." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.166, + 0.664, + 0.182 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.191, + 0.905, + 0.447 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs). MLLMs with both language and images modalities have garnered significant attention in recent years [1, 2, 30, 45, 51]. Prior research has focused on integrating powerful Large Language Models (LLMs) [12, 64, 64] with advanced image encoders [19, 45, 63] such as LLaVA [34], PaLI [9, 10], PaliGemma [4], and InstructBLIP [14]. Through instruction tuning or multi-modal finetuning, these models demonstrate impressive performance in multimodal understanding and reasoning. Current developments are leveraging increasingly larger multimodal datasets to further enhance their capabilities in complex perception and generalization tasks [3, 42, 59]. However, despite their strengths, these models have shown limitations in 3D spatial reasoning [11, 22, 40], posing challenges for their application in autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.906, + 0.659 + ], + "angle": 0, + "content": "End-to-end Autonomous Driving. While the key final step of autonomous driving is planning, traditional systems are often assembled with multiple modules for perception [33, 35, 60] and prediction [41, 54, 65] tasks. To mitigate the information loss and error accumulation across modules, end-to-end driving systems [6, 24, 25, 28, 32, 66] utilize a unified model to predict the ego-vehicle future waypoints or control signals directly from raw sensor inputs. While prioritizing planning, these systems typically still incorporate perception and prediction, requiring explicit supervision for each. Although some prior works [13, 46, 58] have explored pure motion planning without any intermediate tasks, they struggle in complex urban scenarios due to modeling capability limitations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.906, + 0.901 + ], + "angle": 0, + "content": "MLLMs for Driving. The superior reasoning and generalization abilities of large-scale models are desirable properties for autonomous driving. Some efforts transform the driving scenario into large language model textual prompts [8, 38, 47] or directly process camera images with vision-language models [15, 48, 61]. However, their potential is limited to the small scale of existing benchmarks [5], which only allow partial finetuning [23]. Meanwhile, closed-loop simulators [16] struggle to provide realistic sensor data for end-to-end tasks. Consequently, multi-task joint finetuning [49, 53, 61] or chain-of-though (CoT) [56] inference [37-39] are widely adopted to simplify the reasoning despite the extra annotation requirements brought by perception and prediction modules. Alternatively, some work [15, 52] integrates pretrained perception models [31] to extract BEV features and sends to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1623" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "language models as vision tokens. Recently, EMMA [26] leverages powerful Gemini [51] for self-supervised motion planning. Additionally, they develop a mixture of training tasks, including motion planning, 3D object detection, and road graph estimation as well as additional reasoning process for trajectory generation. In contrast, our work focuses on the enhancement of self-supervised motion planning without extra human labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.226, + 0.45, + 0.243 + ], + "angle": 0, + "content": "3. Enhancing MLLM for Motion Planning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.482, + 0.357 + ], + "angle": 0, + "content": "In this section, we adapt a pretrained MLLM for the end-to-end motion planning task. Without loss of generality, we leverage PaLI [10]. However, these techniques are also compatible with other more powerful MLLMs. We introduce a set of design improvements to the MLLM setup and demonstrate correspondingly improved performance on WOMD-Planning-ADE (Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.311, + 0.381 + ], + "angle": 0, + "content": "3.1. Vanilla PaLI as Planner" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.483, + 0.575 + ], + "angle": 0, + "content": "End-to-end motion planning models determine the future trajectory \\(\\mathcal{O}_{T_f}\\) of ego-vehicle based on multi-view camera images \\(\\mathcal{C}\\) and high-level behavior command \\(b\\). The future trajectory consists of the ego-vehicle's location at each future time step in the BEV coordinate, i.e. \\(\\mathcal{O}_{T_f} = [(x_t,y_t)]_{t=1}^{T_f}\\). The high-level behavior command is necessary to guide the driving direction, similar to a navigation system. Additionally, as verified in prior works [32, 62], the historical states of ego-vehicle \\(\\mathcal{H}_{T_h}\\) are also critical for a smooth and feasible planning result, where we consider the historical locations, velocities, and accelerations as \\(\\mathcal{H}_{T_h} = [\\mathbf{l}_t,\\mathbf{v}_t,\\mathbf{a}_t]_{t=-1}^{-T_h}\\), i.e." + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.585, + 0.483, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} _ {T _ {f}} = P (\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.718 + ], + "angle": 0, + "content": "where \\( P(\\cdot) \\) is the planning model. We provide both historical ego states and high-level commands to the model as text prompts. Locations, velocities, and accelerations are represented directly as floating numbers with two decimals. The predicted future trajectories are then extracted from the decoded text outputs of the model. Example prompts and targets are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Without preceding perception and prediction tasks, the vanilla PaLI, fine-tuned in a self-supervised manner, achieves reasonable but not ideal performance in motion planning (Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.45, + 0.804 + ], + "angle": 0, + "content": "3.2. Hierarchical Planning with Meta-Decision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Directly outputting a future trajectory without any reasoning is challenging for MLLMs. To address this, we apply a coarse-to-fine approach, inspired by Chain-of-Thought (CoT) [56]. It adopts a hierarchical planning method [27], going from semantic decision to numeric planning. We prompt our model to provide a high-level estimation of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "the future acceleration state \\( D \\) first, effectively decomposing the motion planning task into two steps:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.128, + 0.905, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = P _ {s 1} \\left(\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b\\right), \\quad \\mathcal {O} _ {T _ {f}} = P _ {s 2} \\left(\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b; \\mathcal {D}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.15, + 0.907, + 0.33 + ], + "angle": 0, + "content": "We define \\(D\\) to encompass four meta-decisions: keep stationary, keep speed, accelerate and decelerate. Notably, unlike previous CoT applications in VLM-based planning that require human annotation for training, such as the scene analysis in DriveVLM [53], we incorporate these meta-decisions as a \"free lunch\" to simplify the motion planning process without requiring any additional annotations. The ground-truth decisions are generated by heuristic rules based on the future ego velocity and acceleration (details in the supplementary materials). Fig. 3 shows this simple design brings non-trivial improvements in planning performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.338, + 0.81, + 0.355 + ], + "angle": 0, + "content": "3.3. Scene Representation in 3D Space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.512 + ], + "angle": 0, + "content": "High-quality motion planning requires a robust understanding of the surrounding 3D scene, including both static and dynamic elements. While traditionally achieved through separate perception and prediction modules, our self-supervised end-to-end framework relies on the MLLM to implicitly learn this understanding without explicit supervision. However, despite strong 2D reasoning capabilities, MLLMs often struggle with 3D spatial reasoning due to limitations in their perspective view representation and the lack of depth-related tasks in pretraining [7, 36]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.882, + 0.532 + ], + "angle": 0, + "content": "3.3.1. 3D Visual Representation with Dense Volumes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.852 + ], + "angle": 0, + "content": "To overcome the above limitations, we draw inspiration from previous successful perception tasks [31, 35] and adopt a 3D volume representation. The multi-view feature maps \\(\\mathcal{F}_{2D} = \\{\\mathbf{f}_{2D}^v\\}_{v=1}^V (\\mathbf{f}_{2D}^v \\in \\mathbb{R}^{H \\times W \\times C})\\) are extracted by visual encoder of MLLM, where \\(V\\) is the number of views. We construct a 3D feature volume \\(\\mathbf{f}_{3D}^{vol} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}\\) centered at the ego-vehicle based on the multi-view image features. To avoid introducing complex modules that might disrupt the pretrained MLLM and misalign visual features with the subsequent multimodal encoder-decoder, we employ a lightweight projection method, similar to SimpleBEV [20]. Specifically, for each voxel in the 3D volume, we project its \\((x, y, z)\\) coordinates to each perspective view \\(v\\), obtaining corresponding 2D coordinates \\((u_v, v_v)\\). We then bilinearly sample the local features from each view at these projected locations. Finally, the voxel's feature representation is computed as the average of the local semantic features from all views where the voxel projects within the image bounds. This process effectively incorporates 3D spatial information while maintaining compatibility with the pretrained MLLM." + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.857, + 0.905, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {3 D, (x, y, z)} ^ {v o l} = \\mathbf {f} _ {(x, y, z)} ^ {s e m} + \\operatorname {P o s E m b} (x, y, z) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.879, + 0.905, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {(x, y, z)} ^ {s e m} = \\operatorname {A v g} \\left(\\left\\{\\text {B i l i n e a r} \\left(\\mathbf {f} _ {2 D} ^ {v}; \\left(u _ {v}, v _ {v}\\right)\\right) \\right\\} _ {v}\\right) \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.094, + 0.88, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed S4-Driver algorithm. We enhance the PaLI model for motion planning by incorporating meta-decision, spatio-temporal visual representation, and multi-decoding aggregation." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.4, + 0.462, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.738 + ], + "angle": 0, + "content": "Figure 3. A roadmap for enhancing MLLM for planning. The techniques are adopted step-by-step from top to bottom, while shadow items are not adopted in the subsequent steps. We show the performance on WOMD-Planning-ADE (Sec. 4) after including each module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.822 + ], + "angle": 0, + "content": "This simple and efficient projection strategy ensures that the 3D volume features \\(\\mathbf{f}_{3D}^{vol}\\) share a similar distribution with the original multi-view features \\(\\mathcal{F}_{2D}\\). This similarity facilitates seamless integration with the subsequent multimodal encoder-decoder of the MLLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3, this 3D volume representation leads to some improvement in motion planning performance. Alternatively, we find that reducing the \\(Z\\)-axis using a fully connected layer to obtain a BEV representation yields slightly worse performance since this reduction operation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.407, + 0.864, + 0.422 + ], + "angle": 0, + "content": "may introduce ambiguity to the scene representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.429, + 0.774, + 0.445 + ], + "angle": 0, + "content": "3.3.2. Sparse Volume Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.569 + ], + "angle": 0, + "content": "While the 3D volume representation effectively captures spatial information, inherently most of the surrounding 3D space is empty. Moreover, for motion planning, detailed information about objects distant from the road, such as buildings and trees, is less critical. Based on this observation, we propose a sparse volume representation to reduce the number of voxels, enabling higher resolution given memory constraints and improving efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.909, + 0.662 + ], + "angle": 0, + "content": "To determine useful volumes in each scene based on their positions and semantics, we define a gate value \\(g_{(x,y,z)}^{vol} \\in (0,1)\\) for each volume at coordinates \\((x,y,z)\\). To obtain this gate, starting from multi-view image features \\(\\{\\mathbf{f}_{2D}^v\\}_{v=1}^V\\), we reduce its dimension through a fully connect (FC) layer as" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.672, + 0.907, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {2 D} ^ {\\text {g a t e}, v} = \\operatorname {F C} \\left(\\mathbf {f} _ {2 D} ^ {v}\\right), \\quad \\mathbf {f} _ {2 D} ^ {\\text {g a t e}, v} \\in \\mathbb {R} ^ {H \\times W \\times C ^ {\\prime}}, C ^ {\\prime} \\ll C. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Then we construct a reduced-dimension volume feature \\(\\mathbf{f}_{3D}^{gate} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C'}\\) from \\(\\{\\mathbf{f}_{2D}^{gate,v}\\}_{v=1}^V\\) same as Eq. 3. A smaller channel number allows for larger volume resolution and is sufficient to indicate whether a volume is related to motion planning. Afterward, the gate is derived from \\(\\mathbf{f}_{3D}^{gate}\\) through a small MLP module as follows." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.805, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\ng _ {(x, y, z)} ^ {v o l} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\mathbf {f} _ {3 D, (x, y, z)} ^ {g a t e}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.872 + ], + "angle": 0, + "content": "Thus, we can easily select \\(M\\) volumes \\((M\\ll X\\cdot Y\\cdot Z)\\) with largest gate values with coordinates \\(\\{(x_i,y_i,z_i)\\}_{i = 1}^M\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.872, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As we do not have access to ground-truth occupancy state, we propose learning the gate values implicitly. We" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "assume that regions with small gate values should be empty or irrelevant to planning. For these vacant space, we assign a learnable feature \\(\\mathbf{f}_{vac} \\in \\mathbb{R}^C\\). We expect the model can learn the gate value via the trade-off between the semantic features and this vacant feature at each 3D position, so we fetch features for selected sparse volumes as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.191, + 0.483, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} = \\hat {\\mathbf {f}} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} + \\operatorname {P o s E m b} (x, y, z) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.215, + 0.483, + 0.237 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} _ {3 D, (x, y, z)} ^ {s p a r s e} = g _ {(x, y, z)} ^ {v o l} \\cdot \\mathbf {f} _ {(x, y, z)} ^ {s e m} + \\left(1 - g _ {(x, y, z)} ^ {v o l}\\right) \\cdot \\mathbf {f} _ {v a c} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.485, + 0.355 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{(x,y,z)}^{sem}\\) is same as Eq. 4 and \\((x,y,z)\\in \\{(x_i,y_i,z_i)\\}_{i = 1}^M\\) are selected volumes with large gate values. When the sparse volume features \\(\\mathbf{f}_{3D,(x,y,z)}^{sparse}\\) are fed into the subsequent multimodal encoder, they notably boost the planning performance (Fig. 3) by simultaneously providing explicit 3D spatial cues of surrounding scenes and concentrate on the important regions adaptively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.411, + 0.377 + ], + "angle": 0, + "content": "3.3.3. Local Feature Aggregation in 3D Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Due to the lack of depth information, the lifting process in Eq. 3 or 7 results in duplicate volume features along each camera ray. This spatial ambiguity can be mitigated by 3D local operations like convolutions [20] or deformable attention [31] as demonstrated in previous work. However, in our MLLM framework, it is less economical to insert extra local operations. Instead, we inject some relative position bias into the existing multimodal encoder by tailoring the self-attention as follows." + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.523, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} - b (D)\\right) V \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Given \\(M\\) sparse volumes with coordinates \\(\\{(x_i, y_i, z_i)\\}_{i=1}^M\\), the distance matrix \\(D \\in \\mathbb{R}^{M \\times M \\times 3}\\) is the distance along \\(xyz\\)-axes between each pair of sparse volumes, i.e. \\(D_{ij} = (x_j - x_i, y_j - y_i, z_j - z_i)\\). The bias is computed with the function \\(b(\\Delta x, \\Delta y, \\Delta z) = b_x(\\Delta x) + b_y(\\Delta y) + b_z(\\Delta z)\\), where \\(\\Delta x, \\Delta y, \\Delta z\\) are divided into several bins and mapped to a learnable bias value for each bin through \\(b_x(\\cdot), b_y(\\cdot), b_z(\\cdot)\\). We also apply a separate bias between the text tokens with 1D position \\(\\{p_i\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.795 + ], + "angle": 0, + "content": "This relative position bias elegantly inserts local inductive bias to pretrained global self-attention modules with minimal additional costs. It leads to a notable performance gain in Fig. 3 by facilitating the aggregation of local information in 3D space, enhancing scene understanding and spatial reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.358, + 0.82 + ], + "angle": 0, + "content": "3.4. Multi-frame Temporal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Multi-frame input combination helps to compensate for the lack of depth cues in camera images. We extend our sparse volume representation to aggregate mult-frame temporal information by incorporating \\( T \\) historical frames with an interval of 0.5s. Given images from the total \\( T + 1 \\) frames," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "Eq. 5 is applied to each frame separately to obtain the multi-view image features \\(\\mathbf{f}_{2D,t}^{gate,v}\\), \\(t = 0, -1, \\dots, -T\\), \\(v = 1, \\dots, V\\). After ego-motion compensation, we construct a gate feature volume \\(\\mathbf{f}_{3D,t}^{gate}\\) based on the current ego-vehicle coordinates and per-frame image features. Multi-frame volume features are concatenated along the channel dimension to generate gate value." + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.204, + 0.905, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\left. g _ {(x, y, z)} ^ {v o l} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\operatorname {C o n c a t} \\left([ \\mathbf {f} _ {3 D, t, (x, y, z)} ^ {g a t e} ] _ {t = 0} ^ {- T}\\right)\\right)\\right) \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.906, + 0.32 + ], + "angle": 0, + "content": "Similar with Sec. 3.3.2, we select \\(M\\) volumes based on the gate value. Following Eq. 7, the volume features are fetched from each frame image feature map separately as \\(\\mathbf{f}_{3D,t,(x,y,z)}^{sparse}\\), \\(t = 0,\\dots , - T\\), which are fused with an FC-layer as the sparse volume feature with temporal awareness." + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.326, + 0.905, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} = \\operatorname {F C} \\left(\\operatorname {C o n c a t} \\left(\\left[ \\mathbf {f} _ {3 D, t, (x, y, z)} ^ {\\text {s p a r s e}} \\right] _ {t = 0} ^ {- T}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.397 + ], + "angle": 0, + "content": "In Fig. 3, temporal fusion helps to improve motion planning performance by facilitating the environment understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.854, + 0.421 + ], + "angle": 0, + "content": "3.5. Voting for Planning via Multi-Decoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.552 + ], + "angle": 0, + "content": "MLLMs are prone to assigning high confidence to the simple future behaviors in motion planning, e.g., keeping stationary. To relieve this bias, we aggregate multiple outputs and use voting to obtain the final planning output. We resort to nucleus sampling [21] to produce multiple future trajectories for the ego-vehicle, denoted as \\(\\{\\mathcal{O}_{T_f}^k\\}_{k=1}^K\\). They are simply averaged to produce a unique planning result as follows: \\(\\mathcal{O}_{T_t} = [(x_t, y_t)]_{t=1}^{T_f}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.558, + 0.905, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\left(x _ {t}, y _ {t}\\right) = \\left(\\frac {1}{K} \\sum_ {k = 1} ^ {K} x _ {t} ^ {k}, \\frac {1}{K} \\sum_ {k = 1} ^ {K} y _ {t} ^ {k}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.667 + ], + "angle": 0, + "content": "This unweighted average mitigates the MLLM's bias towards simple behaviors. Fig. 3 demonstrates the notable performance gain with this simple multi-decoding aggregation approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.675, + 0.864, + 0.692 + ], + "angle": 0, + "content": "3.6. Scaling to Large-scale Raw Driving Logs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.803 + ], + "angle": 0, + "content": "The self-supervised training allows our proposed S4-Driver to scale to large-scale driving logs amounts without requiring human annotations. To exploit the potential of MLLM-based planners, we pretrain the model on our internal dataset. Results in Fig. 3 demonstrate that S4-Driver achieves notable performance gains in challenging tail behaviors due to large-scale pretraining." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.897, + 0.833 + ], + "angle": 0, + "content": "4. Waymo Open Motion Dataset for Planning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the large-scale training and evaluation of planning algorithms with large models, we design a WOMD-Planning-ADE benchmark based on WOMD dataset [18], and compare with the existing datasets in Tab. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1626" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.142 + ], + "angle": 0, + "content": "
BenchmarkSizeData FormatMetrics
SequenceLengthFrequencyPlanning horizonSample-wiseBehavior-wise
BDD-X [29]7k77 h10 Hzcontrol signal✓✗
nuScenes [5]1k5.5 h2 Hz3s✓✗
WOMD-Planning-ADE103k574 h10 Hz5s✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.143, + 0.483, + 0.184 + ], + "angle": 0, + "content": "Table 1. Comparisons between WOMD-Planning-ADE and existing counterparts. WOMD-Planning-ADE has a significantly larger size and more comprehensive behavior-wise metrics." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.203, + 0.455, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.287, + 0.473, + 0.302 + ], + "angle": 0, + "content": "Figure 4. High-level behavior commands for motion planning." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.319, + 0.162, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.394, + 0.162, + 0.405 + ], + "angle": 0, + "content": "(a) Trajectory" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.329, + 0.269, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.394, + 0.246, + 0.405 + ], + "angle": 0, + "content": "(b) Velocity" + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.329, + 0.382, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.394, + 0.365, + 0.404 + ], + "angle": 0, + "content": "(c) Acceleration" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.317, + 0.48, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.394, + 0.462, + 0.405 + ], + "angle": 0, + "content": "(d) Behavior" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.46 + ], + "angle": 0, + "content": "Figure 5. Statistical distributions of WOMD-Planning-ADE. We show the distributions of ego-vehicle trajectories, velocities, accelerations and each behavior scenario sizes in the validation set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Overview. It contains \\(103k\\) real-world driving scenarios, covering diverse urban and suburban scenes. They are further divided into 9s samples with 1s history and 8s future. For end-to-end planning evaluation, besides the ego-vehicle trajectories in each sample as groundtruth, we also have the following key items in the dataset:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.627 + ], + "angle": 0, + "content": "(In-house) Camera Data. Most end-to-end planning methods [28, 49, 53, 61] relies on the camera images as model inputs. In our dataset, each frame contains images captured by eight multi-view cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.824 + ], + "angle": 0, + "content": "High-level Behavior Commands. Like navigation systems for humans (e.g. Google Map), the end-to-end planning system also requires navigation signals to indicate the direction to go. We consider six high-level behavior commands (Fig. 4), namely go straight forward, go straight left, go straight right, do left turn, do right turn, do left U-turn. They can cover diverse driving situations in real world, e.g. \"go straight right\" describes the case to leave the high-way. Instead of simply considering the last step position [6, 25], we decide the behavior commands based on long-term future trajectories, which can handle the low-speed or stopping situations. Details of this heuristic rule are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. The imbalance of data distribution is inevitable in real driving scenarios. For example, in the WOMD-Planning-ADE benchmark, moving straight and stopping account for more than \\(70\\%\\) of all the samples, as reflected by Fig. 5. In this case, we argue that current widely" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "used sample-wise average displacement error and collision rates cannot comprehensively reflect the performance of motion planning algorithms since challenging but less frequent behaviors, like turning, are overwhelmed by simple straightforward moving scenarios. To this end, we propose novel behavior-wise metrics similar to the mAP metric in prediction [18]. For example, we denote the behavior-wise average displacement error as \\( bADE \\) defined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.221, + 0.907, + 0.258 + ], + "angle": 0, + "content": "\\[\nb A D E = \\frac {1}{| \\mathcal {B} |} \\sum_ {b \\in \\mathcal {B}} A D E _ {b} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.266, + 0.907, + 0.357 + ], + "angle": 0, + "content": "where \\( ADE_{b} \\) is the ADE metric for a specific behavior. Concretely, \\( |\\mathcal{B}| = 7 \\) behaviors are considered - including the six high-level commands (Fig. 4) and an additional stop behavior1. These behavior-wise metrics make up for the shortcomings of previous sample-wise metrics by paying more attention to those scarce but safety-critical behaviors." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.37, + 0.646, + 0.387 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.395, + 0.731, + 0.411 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Model and finetuning. Unless otherwise specified, we construct our model from a pretrained PaLI3-5B model [10], which includes a ViT-G (2B) vision encoder [17] and multimodal encoder-decoder (3B) [50]. Our default setting freezes the ViT encoder and finetune only the inserted modules and the multimodal encoder-decoder. Finetuning the model on the WOMD-Planning-ADE dataset requires approximately 2.5 days on 128 Google Cloud TPU v4 using a batch size of 256 and a learning rate of 3e-3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Datasets. We evaluate S4-Driver on both WOMD-Planning-ADE (Sec. 4) and nuScenes [5] benchmarks. Given the limited size of nuScenes, which is insufficient for training a multimodal large language model, we also attempt to finetune the model on nuScenes using weights pretrained on WOMD-Planning-ADE." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.785, + 0.669 + ], + "angle": 0, + "content": "5.2. Main Results and Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.825 + ], + "angle": 0, + "content": "NuScenes dataset. To widely compare with existing end-to-end planning methods in the community, we evaluate S4-Driver on nuScenes dataset [5]. Tab. 2 show that S4-Driver notably outperforms all the previous algorithms despite its self-supervised training format. Unlike existing works, S4-Driver does not require any perception pretraining or human annotations. This self-supervision nature enables S4-Driver to utilize all the accessible raw trajectory data, whereas these easily collectable raw driving logs offer limited benefit to other algorithms without a labeling process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.906, + 0.856 + ], + "angle": 0, + "content": "WOMD-Planning-ADE benchmark. In Tab. 3, we primarily compare S4-Driver with the vanilla PaLI3-5B base" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1We do not include \"stop\" in the high-level behavior commands since navigation system cannot provide it and this may leak future information like traffic lights. Details are explained in supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1627" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.854, + 0.366 + ], + "angle": 0, + "content": "
MethodsHuman AnnotationsExtra DataL2(m)
perception predictiondata labels1.0s2.0s3.0sAvg1,2,3s
Multi-task end-to-end approaches
UniAD [25]✓✓XX0.420.640.910.66
VAD [28]✓✓XX0.170.340.600.37
PARA-Drive [57]✓✓XX0.250.460.740.48
Multimodal large language model approaches
GPT-Driver [38]✓✓XX0.200.400.700.44
DriveVLM [53]✓✓✓✓0.180.340.680.40
OmniDrive [55]✓✓✓✓0.140.290.550.33
Self-supervised approaches
DriveVLM (w/o CoT) [53]XX✓X0.190.410.890.49
S4-Driver (ours)XXXX0.160.340.630.38
S4-Driver* (ours)XX✓X0.130.280.510.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.368, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Table 2. NuScenes benchmark results. Most previous works require human annotations as additional supervision. S4-Driver* is pretrained on WOMD-Planning-ADE, while many prior MLLM-based methods also utilize extra data and labels due to the limited scale of nuScenes." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.404, + 0.855, + 0.539 + ], + "angle": 0, + "content": "
MethodsSample-wise MetricsBehavior-wise Metrics
ADE@1sADE@3sADE@5sbADE@1sbADE@3sbADE@5s
Vanilla PaLI0.0340.2770.7980.0510.3921.069
S4-Driver (ours)0.0310.2470.6930.0490.3500.928
S4-Driver* (ours)0.0290.2330.6550.0440.3130.830
Models that utilize high-quality objects, tracks, and roadgraph as inputs rather than raw camera images.
MotionLM [46]0.0450.2660.6970.0660.3880.978
" + }, + { + "type": "table_caption", + "bbox": [ + 0.179, + 0.539, + 0.816, + 0.553 + ], + "angle": 0, + "content": "Table 3. WOMD-Planning-ADE benchmark results. “*” denotes methods pretrained on the internal dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.564, + 0.483, + 0.776 + ], + "angle": 0, + "content": "line (Sec. 3.1) and modular algorithm MotionLM [46]. Compared with vanilla PaLI3-5B, a notable gap is witnessed for both sample-wise and behavior-wise metrics. For reference, we also adapt a latest motion prediction algorithms MotionLM [46] (enhanced internally reproduced version) for planning task by only predicting the future trajectories of ego-vehicle and injecting the high-level command to the model. Because it utilizes auto-labelled high-quality objects, tracklets, and roadgraph signals as model inputs [44], a direct comparison with our end-to-end approaches is not equitable. However, as shown in Tab. 3, S4-Driver still achieves favorable performance compared to MotionLM, particularly in behavior-wise metrics even if S4-Driver uses only raw camera images as inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.194, + 0.803 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. Fig. 6 visualizes planning results in diverse scenarios. Our proposed S4-Driver can determine the future ego-behavior based on the traffic lights and road lane. It can also complete challenging behaviors and deal with different lighting conditions. We will include more visualizations in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Meta-decision reliability. Fig. 7 demonstrates the accuracy of meta-decision prediction (Sec. 3.2) on WOMD-Planning-ADE validation set. Across all behaviors, the model provides reliable meta-decision estimations. Without any human annotations, this preliminary prediction can simplify the derivation of numeric motion planning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Distribution of sparse volumes. We visualize the distribution of self-supervised learned sparse volumes along x-axis and y-axis over WOMD-Planning-ADE validation set in Fig. 8. From back to front, the sparse volumes concentrate on the close front region. From left to right, the sparse volumes cover all the regions, since there are turning scenarios, but most volumes focus on the middle region. These distributions are consistent with human driving experiences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.68, + 0.802 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.87 + ], + "angle": 0, + "content": "This section presents experiments conducted on the WOMD planning benchmark to clarify the roles of specific modules within our S4-Driver algorithm. More ablation studies are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "MLLM inputs. In Tab. 5, we analyze the roles of camera images and historical ego-states separately using the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1628" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.22, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.091, + 0.347, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.091, + 0.474, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.144, + 0.418, + 0.153 + ], + "angle": 0, + "content": "(a) The ego-vehicle decelerates to wait for the red lights." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.155, + 0.22, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.155, + 0.347, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.155, + 0.474, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.209, + 0.418, + 0.217 + ], + "angle": 0, + "content": "(b) The ego-vehicle conducts a U-turn at the intersection." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.219, + 0.22, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.22, + 0.347, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.22, + 0.474, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.272, + 0.435, + 0.282 + ], + "angle": 0, + "content": "(c) The ego-vehicle keeps the lane under bad lighting condition." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.283, + 0.482, + 0.31 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of motion planning. We show the front left, front, front right cameras for each case." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.316, + 0.28, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.316, + 0.477, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.452, + 0.282, + 0.48 + ], + "angle": 0, + "content": "Figure 7. Meta-decision prediction accuracy for each behavior." + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.453, + 0.479, + 0.48 + ], + "angle": 0, + "content": "Figure 8. Distribution of sparse volumes along x- and y-axes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.483, + 0.687 + ], + "angle": 0, + "content": "vanilla PaLI model. Consistent with [32, 62], the importance of ego-states is acknowledged. Meanwhile, different from their observations on nuScenes benchmark, camera inputs also make great differences on WOMD-Planning-ADE. We assume that WOMD-Planning-ADE covers more diverse driving scenarios including many severe changes of speed and directions, which attaches importance to the sensor data. This also showcases the advantages of WOMD-Planning-ADE for comprehensive evaluation. Tab. 5 also shows that random initialized model cannot converge without MLLM pretraining. Despite different domains, S4-Driver can profit from the strong reasoning ability from large-scale MLLM pretraining on general tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.884 + ], + "angle": 0, + "content": "MLLM capability. In addition to the PaLI3-5B [10] used in other parts, we also tailor PaLI2-3B [9] for motion planning with the same techniques. PaLI2-3B is weaker in parameter numbers (3B vs. 5B), architecture, pretraining manner, and input resolution (224 vs. 448). As in Tab. 4, S4-Driver based on PaLI2-3B demonstrates much inferior performance compared to that based on PaLI3-5B. We conduct experiments on two different training data scales on WOMD-Planning-ADE, i.e. 20k (nuScenes scale) vs. 400k (full WOMD-Planning-ADE). The gap is particularly large given enough training data. It also justifies the necessity to conduct experiments on large-scale dataset, which can fully unleash the potential of powerful MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Sparse volume resolution. Tab. 6 shows the results with" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.136 + ], + "angle": 0, + "content": "
MLLMWOMD-Planning-ADE Subset (20k)Full WOMD-Planning-ADE (400k)
ADE@5sbADE@5sADE@5sbADE@5s
PaLI2-3B [9]1.2411.8461.0351.522
PaLI3-5B [10]1.0751.4950.6930.928
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.178 + ], + "angle": 0, + "content": "Table 4. Ablation studies on planning performance scalability of our methods w.r.t. model and data sizes. We experiment with two MLLMs on two different training data sizes." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.191, + 0.885, + 0.295 + ], + "angle": 0, + "content": "
MethodADE@5sbADE@5s
vanilla PaLI0.7981.069
w/o historical states2.0092.250
w/o camera images1.0811.448
w/o MLLM pretrainingfail to converge
" + }, + { + "type": "table_caption", + "bbox": [ + 0.579, + 0.295, + 0.84, + 0.308 + ], + "angle": 0, + "content": "Table 5. Ablation studies on MLLM inputs." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.322, + 0.862, + 0.398 + ], + "angle": 0, + "content": "
ResolutionADE@5sbADE@5s
2m × 2m × 2m0.7981.201
1m × 1m × 2m0.7501.005
1m × 1m × 1m0.7651.051
" + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.399, + 0.873, + 0.411 + ], + "angle": 0, + "content": "Table 6. Ablation studies on sparse volume resolutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.545 + ], + "angle": 0, + "content": "different resolution for sparse volumes with same number of sparse volumes (\\(M = 6000\\)). In line with Fig. 3, low resolution leads to relatively worse performance since it limits the precision of 3D spatial reasoning. Interestingly, higher resolution along \\(z\\)-axis does not definitely improve the model performance since the motion planning mainly works on the \\(xy\\)-plane and too low sparsity ratio tends to make the optimization unstable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.783, + 0.574 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper introduces S4-Driver, a scalable self-supervised motion planning framework for autonomous driving that utilizes Multimodal Large Language Models (MLLMs). To enhance 3D reasoning in MLLMs, we propose a novel sparse volume representation, which enables improved spatio-temporal reasoning through the aggregation of multi-view and multi-frame image inputs. Additionally, we design behavior-wise metrics for comprehensive evaluation on the large-scale WOMD-Planning-ADE benchmark. S4-Driver achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks, without requiring any human annotations. This demonstrates the potential of self-supervised learning for end-to-end autonomous driving. Future work will involve applying our methods to other powerful MLLM architectures. Our contributions are orthogonal to previous works, such as those that explore multi-task learning or CoT reasoning [49, 55]. Combining our large-scale self-supervised learning approach with supervised finetuning on targeted small-scale labeled data may further enhance the final performance and interpretability of the system." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1629" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[4] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[8] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[12] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[13] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[15] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's-eye-view injected multi-modal large models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[18] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[19] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[20] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-bev: What really matters for multi-sensor bev perception? In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2759-2765. IEEE, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.816, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[21] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[22] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1630" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "in Neural Information Processing Systems, 36:20482-20494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[24] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[25] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[26] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to-end multimodal model for autonomous driving, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[27] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[28] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[29] Jinkyu Kim, Anna Rohrbach, Trevor Darrell, John Canny, and Zeynep Akata. Textual explanations for self-driving vehicles. In Proceedings of the European conference on computer vision (ECCV), pages 563-578, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[32] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Bencheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Structured modeling and learning for online vectorized HD map construction. In The Eleventh International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[35] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[36] Xianzheng Ma, Yash Bhalgat, Brandon Smart, Shuai Chen, Xinghui Li, Jian Ding, Jindong Gu, Dave Zhenyu Chen, Songyou Peng, Jia-Wang Bian, et al. When llms step into the 3d world: A survey and meta-analysis of 3d tasks via multi-modal large language models. arXiv preprint arXiv:2405.10255, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[37] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[38] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[39] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[40] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[41] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.828, + 0.619 + ], + "angle": 0, + "content": "[42] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.621, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[43] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.663, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[44] C. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa T. Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6130-6140, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[46] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 7, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1631" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[47] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[48] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[49] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In First Vision and Language for Autonomous Driving and Robotics Workshop. 2, 6, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[50] Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, et al. UI2: Unifying language learning paradigms. arXiv preprint arXiv:2205.05131, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[51] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[52] Thomas Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. In Workshop on Language and Robot Learning: Language as an Interface. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[53] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[54] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[55] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[56] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[57] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[58] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.779, + 0.206 + ], + "angle": 0, + "content": "[59] xAI. Grok-1.5 vision preview, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[60] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[61] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[62] Jiang-Tian Zhai, Ze Feng, Jihao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[64] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[65] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[66] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1632" + } + ] +] \ No newline at end of file diff --git a/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_origin.pdf b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..28a1c71be98853c785db54d44b42387ab3797763 --- /dev/null +++ b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/20e2679c-f6b9-4a84-9623-55af78c15f58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:879480606b4ab71e16c4f69c960795741b4cdef673fc87aa0f02d84468e3b7d7 +size 6903577 diff --git a/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/full.md b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..53db9e33d6d7cfb555859159765ec97e0a768d9b --- /dev/null +++ b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/full.md @@ -0,0 +1,406 @@ +# S4-Driver: Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation + +Yichen Xie $^{1,\dagger,*}$ Runsheng Xu $^{2,*}$ Tong He $^{2}$ Jyh-Jing Hwang $^{2}$ Katie Luo $^{3,\dagger}$ + +Jingwei Ji² Hubert Lin² Letian Chen⁴,† Yiren Lu² Zhaoqi Leng² + +Dragomir Anguelov² Mingxing Tan² + +1 UC Berkeley, 2 Waymo LLC, 3 Cornell University, 4 Georgia Institute of Technology + +# Abstract + +The latest advancements in multi-modal large language models (MLLMs) have spurred a strong renewed interest in end-to-end motion planning approaches for autonomous driving. Many end-to-end approaches rely on human annotations to learn intermediate perception and prediction tasks, while purely self-supervised approaches—which directly learn from sensor inputs to generate planning trajectories without human annotations—often underperform the state of the art. We observe a key gap in the input representation space: end-to-end approaches built on MLLMs are often pretrained with reasoning tasks in 2D image space rather than the native 3D space in which autonomous vehicles plan. To this end, we propose S4-Driver, a scalable self-supervised motion planning algorithm with spatio-temporal visual representation, based on the popular PaLI [9] multimodal large language model. S4-Driver uses a novel sparse volume strategy to seamlessly transform the strong visual representation of MLLMs from perspective view to 3D space without the need to finetune the vision encoder. This representation aggregates multi-view and multiframe visual inputs and enables better prediction of planning trajectories in 3D space. To validate our method, we run experiments on both nuScenes and Waymo Open Motion Dataset (with in-house camera data). Results show that S4-Driver performs favorably against existing supervised multi-task approaches while requiring no human annotations. It also demonstrates great scalability when pretrained on large volumes of unannotated driving logs. + +# 1. Introduction + +The exploration of end-to-end autonomous driving dates back to the 1980s [43], where the motion planning model + +![](images/d3a2e1d187e9baec01ed63e52bf09edb33d55e8b892eec6d7070f9f3c53b5193.jpg) +(a) Pretraining data scaling-up of S4-Driver. + +![](images/829ab413ed8ba1f08f3aaac646d8963e7e37fbb6bf9506d1712950fadf5cd6a6.jpg) +(b) Multi- + +![](images/9541e43b9face5cc35a2f5eaa0360ea10c06e441f3875553272b4cbef8206b3a.jpg) +Multi-task Results + +![](images/57397cda1cf2eba4009f403bda065818ec1ff1132dc0e76999395fc21f870468.jpg) +Figure 1. S4-Driver scaling up with abundant unlabeled driving logs (1a). Different from past efforts that improve MLLM planning via multi-task learning and CoT reasoning (1b), S4-Driver focuses on self-supervised motion planning (1c) without human annotations. While a vanilla PaLI approach with self-supervised training achieves good performance, the full S4-Driver with spatio-temporal visual representation achieves the best performance and scalability (1a). +(c) Self-supervised planning framework. + +![](images/2b85419079a725bc40ad065be321cf7c75bf867e9b539565abad03f24d3cf937.jpg) + +![](images/90ae78a5d1e2f64a66812289e1cf162e095afc983a8e6c6195acad45ff40a2a8.jpg) + +directly predicts control signals or future waypoints based on raw sensor inputs. Due to limited robustness, these early attempts struggle in complex urban situations. Recent neural network capability advances have led to renewed interest in this field [25, 28, 57], especially given the strong generalization ability of Multimodal Large Language Models (MLLMs) [10, 42, 51]. + +However, vanilla application of MLLMs to end-to-end motion planning can hardly exploit their strong visual understanding and reasoning capabilities given the significant difference between motion planning and MLLM pretraining tasks, resulting in inferior planning performance [53, 61]. To narrow this gap, as shown in Fig. 1b, previous methods resort to either multi-task learning, which incorporates diverse perception and prediction tasks into training and inference, or supervised perception pretraining, which utilizes the pretrained autonomous driving perception models as visual tokenizers. However, fine-grained human annotations + +serve as the bottleneck for both strategies. In contrast, pure self-supervised approaches, while capable of learning directly from sensor inputs and utilizing the abundant unlabeled driving logs, often under-perform the state of the art. + +To challenge this status quo, we aim to boost the performance of self-supervised motion planning orthogonal to prior work (Fig. 1c). It only requires raw sensor data and ego-vehicle trajectories for self-supervised model training, without the need of expensive human labels. First of all, we identify the following two main obstacles in this direction: + +- Suboptimal representation format. MLLMs [9, 10] are typically designed for tasks in the 2D image plane. This image space representation limits their 3D reasoning ability from the combination of multi-view images. +- Limited data scale. Although it is one of the most widely adopted benchmarks for end-to-end planning, nuScenes [5] only contains fewer than 1k sequences and lacks driving behavior diversities. This limited scales leads to severe over-fitting issues in the finetuning of MLLMs with billion-parameter size. + +In this paper, we handle the above challenges to better explore the potential of MLLMs. Firstly, we propose S4-Driver, a simple yet effective Scalable Self-Supervised motion planning method with Spatio-temporal visual representation. Built upon a general multimodal large language model [9, 10], we directly predict ego-vehicle waypoints from camera images, eliminating the need for intermediate perception and prediction tasks, and thus facilitating the scale-up of model pretraining leveraging massive unannotated driving logs (Fig. 1a). To address the obstacle of suboptimal visual format, we propose a novel sparse volume representation that enables aggregating visual information from multi-view and multi-frame images. Moreover, this spatio-temporal representation is designed to work with frozen features from the perspective-view vision encoder. It boosts 3D spatio-temporal reasoning capabilities of the model on motion planning, and seamlessly preserves world knowledge in the pretrained visual embeddings of MLLMs. Secondly, to rigorously evaluate our method with sufficient training data, we also utilize large-scale WOMD-PlanningADE benchmark [18] with in-house camera sensor data besides the popular nuScenes dataset [5]. WOMD-PlanningADE is approximately $100 \times$ larger than nuScenes, so it serves as a more comprehensive benchmark to evaluate model performance in a scalable manner. + +In summary, our contributions include: + +- We advocate for a simple self-supervised end-to-end motion planning with MLLMs, solely relying on ego-vehicle trajectory supervision. This approach simplifies the system architecture and facilitates data scaling for training. +- We propose a novel sparse volume representation to effectively aggregate multi-view and multi-frame visual information, enhancing the 3D spatio-temporal reasoning + +capabilities of the MLLM for motion planning. + +- Our algorithm achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks. It also demonstrate great scalability with pretraining. + +# 2. Related Works + +Multimodal Large Language Models (MLLMs). MLLMs with both language and images modalities have garnered significant attention in recent years [1, 2, 30, 45, 51]. Prior research has focused on integrating powerful Large Language Models (LLMs) [12, 64, 64] with advanced image encoders [19, 45, 63] such as LLaVA [34], PaLI [9, 10], PaliGemma [4], and InstructBLIP [14]. Through instruction tuning or multi-modal finetuning, these models demonstrate impressive performance in multimodal understanding and reasoning. Current developments are leveraging increasingly larger multimodal datasets to further enhance their capabilities in complex perception and generalization tasks [3, 42, 59]. However, despite their strengths, these models have shown limitations in 3D spatial reasoning [11, 22, 40], posing challenges for their application in autonomous driving. + +End-to-end Autonomous Driving. While the key final step of autonomous driving is planning, traditional systems are often assembled with multiple modules for perception [33, 35, 60] and prediction [41, 54, 65] tasks. To mitigate the information loss and error accumulation across modules, end-to-end driving systems [6, 24, 25, 28, 32, 66] utilize a unified model to predict the ego-vehicle future waypoints or control signals directly from raw sensor inputs. While prioritizing planning, these systems typically still incorporate perception and prediction, requiring explicit supervision for each. Although some prior works [13, 46, 58] have explored pure motion planning without any intermediate tasks, they struggle in complex urban scenarios due to modeling capability limitations. + +MLLMs for Driving. The superior reasoning and generalization abilities of large-scale models are desirable properties for autonomous driving. Some efforts transform the driving scenario into large language model textual prompts [8, 38, 47] or directly process camera images with vision-language models [15, 48, 61]. However, their potential is limited to the small scale of existing benchmarks [5], which only allow partial finetuning [23]. Meanwhile, closed-loop simulators [16] struggle to provide realistic sensor data for end-to-end tasks. Consequently, multi-task joint finetuning [49, 53, 61] or chain-of-though (CoT) [56] inference [37-39] are widely adopted to simplify the reasoning despite the extra annotation requirements brought by perception and prediction modules. Alternatively, some work [15, 52] integrates pretrained perception models [31] to extract BEV features and sends to + +language models as vision tokens. Recently, EMMA [26] leverages powerful Gemini [51] for self-supervised motion planning. Additionally, they develop a mixture of training tasks, including motion planning, 3D object detection, and road graph estimation as well as additional reasoning process for trajectory generation. In contrast, our work focuses on the enhancement of self-supervised motion planning without extra human labels. + +# 3. Enhancing MLLM for Motion Planning + +In this section, we adapt a pretrained MLLM for the end-to-end motion planning task. Without loss of generality, we leverage PaLI [10]. However, these techniques are also compatible with other more powerful MLLMs. We introduce a set of design improvements to the MLLM setup and demonstrate correspondingly improved performance on WOMD-Planning-ADE (Fig. 3). + +# 3.1. Vanilla PaLI as Planner + +End-to-end motion planning models determine the future trajectory $\mathcal{O}_{T_f}$ of ego-vehicle based on multi-view camera images $\mathcal{C}$ and high-level behavior command $b$ . The future trajectory consists of the ego-vehicle's location at each future time step in the BEV coordinate, i.e. $\mathcal{O}_{T_f} = [(x_t,y_t)]_{t=1}^{T_f}$ . The high-level behavior command is necessary to guide the driving direction, similar to a navigation system. Additionally, as verified in prior works [32, 62], the historical states of ego-vehicle $\mathcal{H}_{T_h}$ are also critical for a smooth and feasible planning result, where we consider the historical locations, velocities, and accelerations as $\mathcal{H}_{T_h} = [\mathbf{l}_t,\mathbf{v}_t,\mathbf{a}_t]_{t=-1}^{-T_h}$ , i.e. + +$$ +\mathcal {O} _ {T _ {f}} = P (\mathcal {C}, \mathcal {H} _ {T _ {h}}, b) \tag {1} +$$ + +where $P(\cdot)$ is the planning model. We provide both historical ego states and high-level commands to the model as text prompts. Locations, velocities, and accelerations are represented directly as floating numbers with two decimals. The predicted future trajectories are then extracted from the decoded text outputs of the model. Example prompts and targets are provided in the supplementary materials. + +Without preceding perception and prediction tasks, the vanilla PaLI, fine-tuned in a self-supervised manner, achieves reasonable but not ideal performance in motion planning (Fig. 3). + +# 3.2. Hierarchical Planning with Meta-Decision + +Directly outputting a future trajectory without any reasoning is challenging for MLLMs. To address this, we apply a coarse-to-fine approach, inspired by Chain-of-Thought (CoT) [56]. It adopts a hierarchical planning method [27], going from semantic decision to numeric planning. We prompt our model to provide a high-level estimation of + +the future acceleration state $D$ first, effectively decomposing the motion planning task into two steps: + +$$ +\mathcal {D} = P _ {s 1} \left(\mathcal {C}, \mathcal {H} _ {T _ {h}}, b\right), \quad \mathcal {O} _ {T _ {f}} = P _ {s 2} \left(\mathcal {C}, \mathcal {H} _ {T _ {h}}, b; \mathcal {D}\right) \tag {2} +$$ + +We define $D$ to encompass four meta-decisions: keep stationary, keep speed, accelerate and decelerate. Notably, unlike previous CoT applications in VLM-based planning that require human annotation for training, such as the scene analysis in DriveVLM [53], we incorporate these meta-decisions as a "free lunch" to simplify the motion planning process without requiring any additional annotations. The ground-truth decisions are generated by heuristic rules based on the future ego velocity and acceleration (details in the supplementary materials). Fig. 3 shows this simple design brings non-trivial improvements in planning performance. + +# 3.3. Scene Representation in 3D Space + +High-quality motion planning requires a robust understanding of the surrounding 3D scene, including both static and dynamic elements. While traditionally achieved through separate perception and prediction modules, our self-supervised end-to-end framework relies on the MLLM to implicitly learn this understanding without explicit supervision. However, despite strong 2D reasoning capabilities, MLLMs often struggle with 3D spatial reasoning due to limitations in their perspective view representation and the lack of depth-related tasks in pretraining [7, 36]. + +# 3.3.1. 3D Visual Representation with Dense Volumes + +To overcome the above limitations, we draw inspiration from previous successful perception tasks [31, 35] and adopt a 3D volume representation. The multi-view feature maps $\mathcal{F}_{2D} = \{\mathbf{f}_{2D}^v\}_{v=1}^V (\mathbf{f}_{2D}^v \in \mathbb{R}^{H \times W \times C})$ are extracted by visual encoder of MLLM, where $V$ is the number of views. We construct a 3D feature volume $\mathbf{f}_{3D}^{vol} \in \mathbb{R}^{X \times Y \times Z \times C}$ centered at the ego-vehicle based on the multi-view image features. To avoid introducing complex modules that might disrupt the pretrained MLLM and misalign visual features with the subsequent multimodal encoder-decoder, we employ a lightweight projection method, similar to SimpleBEV [20]. Specifically, for each voxel in the 3D volume, we project its $(x, y, z)$ coordinates to each perspective view $v$ , obtaining corresponding 2D coordinates $(u_v, v_v)$ . We then bilinearly sample the local features from each view at these projected locations. Finally, the voxel's feature representation is computed as the average of the local semantic features from all views where the voxel projects within the image bounds. This process effectively incorporates 3D spatial information while maintaining compatibility with the pretrained MLLM. + +$$ +\mathbf {f} _ {3 D, (x, y, z)} ^ {v o l} = \mathbf {f} _ {(x, y, z)} ^ {s e m} + \operatorname {P o s E m b} (x, y, z) \tag {3} +$$ + +$$ +\mathbf {f} _ {(x, y, z)} ^ {s e m} = \operatorname {A v g} \left(\left\{\text {B i l i n e a r} \left(\mathbf {f} _ {2 D} ^ {v}; \left(u _ {v}, v _ {v}\right)\right) \right\} _ {v}\right) \tag {4} +$$ + +![](images/739a83147b635f3fa2284d23b820d3da079b7878aa3588d0fc23335bcb8956a3.jpg) +Figure 2. Overview of our proposed S4-Driver algorithm. We enhance the PaLI model for motion planning by incorporating meta-decision, spatio-temporal visual representation, and multi-decoding aggregation. + +![](images/ba84bc19dd0f40b56251e9153e600bae71d43fff973a1a56a1bae8e5095dc5d3.jpg) +Figure 3. A roadmap for enhancing MLLM for planning. The techniques are adopted step-by-step from top to bottom, while shadow items are not adopted in the subsequent steps. We show the performance on WOMD-Planning-ADE (Sec. 4) after including each module. + +This simple and efficient projection strategy ensures that the 3D volume features $\mathbf{f}_{3D}^{vol}$ share a similar distribution with the original multi-view features $\mathcal{F}_{2D}$ . This similarity facilitates seamless integration with the subsequent multimodal encoder-decoder of the MLLM. + +As shown in Fig. 3, this 3D volume representation leads to some improvement in motion planning performance. Alternatively, we find that reducing the $Z$ -axis using a fully connected layer to obtain a BEV representation yields slightly worse performance since this reduction operation + +may introduce ambiguity to the scene representation. + +# 3.3.2. Sparse Volume Representation + +While the 3D volume representation effectively captures spatial information, inherently most of the surrounding 3D space is empty. Moreover, for motion planning, detailed information about objects distant from the road, such as buildings and trees, is less critical. Based on this observation, we propose a sparse volume representation to reduce the number of voxels, enabling higher resolution given memory constraints and improving efficiency. + +To determine useful volumes in each scene based on their positions and semantics, we define a gate value $g_{(x,y,z)}^{vol} \in (0,1)$ for each volume at coordinates $(x,y,z)$ . To obtain this gate, starting from multi-view image features $\{\mathbf{f}_{2D}^v\}_{v=1}^V$ , we reduce its dimension through a fully connect (FC) layer as + +$$ +\mathbf {f} _ {2 D} ^ {\text {g a t e}, v} = \operatorname {F C} \left(\mathbf {f} _ {2 D} ^ {v}\right), \quad \mathbf {f} _ {2 D} ^ {\text {g a t e}, v} \in \mathbb {R} ^ {H \times W \times C ^ {\prime}}, C ^ {\prime} \ll C. \tag {5} +$$ + +Then we construct a reduced-dimension volume feature $\mathbf{f}_{3D}^{gate} \in \mathbb{R}^{X \times Y \times Z \times C'}$ from $\{\mathbf{f}_{2D}^{gate,v}\}_{v=1}^V$ same as Eq. 3. A smaller channel number allows for larger volume resolution and is sufficient to indicate whether a volume is related to motion planning. Afterward, the gate is derived from $\mathbf{f}_{3D}^{gate}$ through a small MLP module as follows. + +$$ +g _ {(x, y, z)} ^ {v o l} = \operatorname {s i g m o i d} \left(\mathrm {M L P} \left(\mathbf {f} _ {3 D, (x, y, z)} ^ {g a t e}\right)\right) \tag {6} +$$ + +Thus, we can easily select $M$ volumes $(M\ll X\cdot Y\cdot Z)$ with largest gate values with coordinates $\{(x_i,y_i,z_i)\}_{i = 1}^M$ + +As we do not have access to ground-truth occupancy state, we propose learning the gate values implicitly. We + +assume that regions with small gate values should be empty or irrelevant to planning. For these vacant space, we assign a learnable feature $\mathbf{f}_{vac} \in \mathbb{R}^C$ . We expect the model can learn the gate value via the trade-off between the semantic features and this vacant feature at each 3D position, so we fetch features for selected sparse volumes as + +$$ +\mathbf {f} _ {3 D, (x, y, z)} ^ {\text {s p a r s e}} = \hat {\mathbf {f}} _ {3 D, (x, y, z)} ^ {\text {s p a r s e}} + \operatorname {P o s E m b} (x, y, z) \tag {7} +$$ + +$$ +\hat {\mathbf {f}} _ {3 D, (x, y, z)} ^ {s p a r s e} = g _ {(x, y, z)} ^ {v o l} \cdot \mathbf {f} _ {(x, y, z)} ^ {s e m} + \left(1 - g _ {(x, y, z)} ^ {v o l}\right) \cdot \mathbf {f} _ {v a c} \tag {8} +$$ + +where $\mathbf{f}_{(x,y,z)}^{sem}$ is same as Eq. 4 and $(x,y,z)\in \{(x_i,y_i,z_i)\}_{i = 1}^M$ are selected volumes with large gate values. When the sparse volume features $\mathbf{f}_{3D,(x,y,z)}^{sparse}$ are fed into the subsequent multimodal encoder, they notably boost the planning performance (Fig. 3) by simultaneously providing explicit 3D spatial cues of surrounding scenes and concentrate on the important regions adaptively. + +# 3.3.3. Local Feature Aggregation in 3D Space + +Due to the lack of depth information, the lifting process in Eq. 3 or 7 results in duplicate volume features along each camera ray. This spatial ambiguity can be mitigated by 3D local operations like convolutions [20] or deformable attention [31] as demonstrated in previous work. However, in our MLLM framework, it is less economical to insert extra local operations. Instead, we inject some relative position bias into the existing multimodal encoder by tailoring the self-attention as follows. + +$$ +\operatorname {A t t n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}} - b (D)\right) V \tag {9} +$$ + +Given $M$ sparse volumes with coordinates $\{(x_i, y_i, z_i)\}_{i=1}^M$ , the distance matrix $D \in \mathbb{R}^{M \times M \times 3}$ is the distance along $xyz$ -axes between each pair of sparse volumes, i.e. $D_{ij} = (x_j - x_i, y_j - y_i, z_j - z_i)$ . The bias is computed with the function $b(\Delta x, \Delta y, \Delta z) = b_x(\Delta x) + b_y(\Delta y) + b_z(\Delta z)$ , where $\Delta x, \Delta y, \Delta z$ are divided into several bins and mapped to a learnable bias value for each bin through $b_x(\cdot), b_y(\cdot), b_z(\cdot)$ . We also apply a separate bias between the text tokens with 1D position $\{p_i\}$ . + +This relative position bias elegantly inserts local inductive bias to pretrained global self-attention modules with minimal additional costs. It leads to a notable performance gain in Fig. 3 by facilitating the aggregation of local information in 3D space, enhancing scene understanding and spatial reasoning. + +# 3.4. Multi-frame Temporal Fusion + +Multi-frame input combination helps to compensate for the lack of depth cues in camera images. We extend our sparse volume representation to aggregate mult-frame temporal information by incorporating $T$ historical frames with an interval of 0.5s. Given images from the total $T + 1$ frames, + +Eq. 5 is applied to each frame separately to obtain the multi-view image features $\mathbf{f}_{2D,t}^{gate,v}$ , $t = 0, -1, \dots, -T$ , $v = 1, \dots, V$ . After ego-motion compensation, we construct a gate feature volume $\mathbf{f}_{3D,t}^{gate}$ based on the current ego-vehicle coordinates and per-frame image features. Multi-frame volume features are concatenated along the channel dimension to generate gate value. + +$$ +\left. g _ {(x, y, z)} ^ {v o l} = \operatorname {s i g m o i d} \left(\mathrm {M L P} \left(\operatorname {C o n c a t} \left([ \mathbf {f} _ {3 D, t, (x, y, z)} ^ {g a t e} ] _ {t = 0} ^ {- T}\right)\right)\right) \right. \tag {10} +$$ + +Similar with Sec. 3.3.2, we select $M$ volumes based on the gate value. Following Eq. 7, the volume features are fetched from each frame image feature map separately as $\mathbf{f}_{3D,t,(x,y,z)}^{sparse}$ , $t = 0,\dots , - T$ , which are fused with an FC-layer as the sparse volume feature with temporal awareness. + +$$ +\mathbf {f} _ {3 D, (x, y, z)} ^ {\text {s p a r s e}} = \operatorname {F C} \left(\operatorname {C o n c a t} \left(\left[ \mathbf {f} _ {3 D, t, (x, y, z)} ^ {\text {s p a r s e}} \right] _ {t = 0} ^ {- T}\right)\right) \tag {11} +$$ + +In Fig. 3, temporal fusion helps to improve motion planning performance by facilitating the environment understanding. + +# 3.5. Voting for Planning via Multi-Decoding + +MLLMs are prone to assigning high confidence to the simple future behaviors in motion planning, e.g., keeping stationary. To relieve this bias, we aggregate multiple outputs and use voting to obtain the final planning output. We resort to nucleus sampling [21] to produce multiple future trajectories for the ego-vehicle, denoted as $\{\mathcal{O}_{T_f}^k\}_{k=1}^K$ . They are simply averaged to produce a unique planning result as follows: $\mathcal{O}_{T_t} = [(x_t, y_t)]_{t=1}^{T_f}$ . + +$$ +\left(x _ {t}, y _ {t}\right) = \left(\frac {1}{K} \sum_ {k = 1} ^ {K} x _ {t} ^ {k}, \frac {1}{K} \sum_ {k = 1} ^ {K} y _ {t} ^ {k}\right) \tag {12} +$$ + +This unweighted average mitigates the MLLM's bias towards simple behaviors. Fig. 3 demonstrates the notable performance gain with this simple multi-decoding aggregation approach. + +# 3.6. Scaling to Large-scale Raw Driving Logs + +The self-supervised training allows our proposed S4-Driver to scale to large-scale driving logs amounts without requiring human annotations. To exploit the potential of MLLM-based planners, we pretrain the model on our internal dataset. Results in Fig. 3 demonstrate that S4-Driver achieves notable performance gains in challenging tail behaviors due to large-scale pretraining. + +# 4. Waymo Open Motion Dataset for Planning + +For the large-scale training and evaluation of planning algorithms with large models, we design a WOMD-Planning-ADE benchmark based on WOMD dataset [18], and compare with the existing datasets in Tab. 1. + +
BenchmarkSizeData FormatMetrics
SequenceLengthFrequencyPlanning horizonSample-wiseBehavior-wise
BDD-X [29]7k77 h10 Hzcontrol signal✓✗
nuScenes [5]1k5.5 h2 Hz3s✓✗
WOMD-Planning-ADE103k574 h10 Hz5s✓✓
+ +Table 1. Comparisons between WOMD-Planning-ADE and existing counterparts. WOMD-Planning-ADE has a significantly larger size and more comprehensive behavior-wise metrics. + +![](images/2f6553fb36d69809aef5a53d32a4bb7f8d2a4628d9d0ae019004db293332b883.jpg) +Figure 4. High-level behavior commands for motion planning. + +![](images/7fcdefac43c64057dd57f7e4490847e95b28d7981f80f3001cb1004791109d8f.jpg) +(a) Trajectory + +![](images/137fa8d8e21b0ffbd4d73a760e190f6ce2e9ca7d7478bd2ad8a03286a5c07c36.jpg) +(b) Velocity + +![](images/10e1af860f183a1bf01a55bc3beda5fa15b02ead192d6b3ee01b34eef48c9a38.jpg) +(c) Acceleration +Figure 5. Statistical distributions of WOMD-Planning-ADE. We show the distributions of ego-vehicle trajectories, velocities, accelerations and each behavior scenario sizes in the validation set. + +![](images/889b62746d6df0156b0bc997472c14a36afe55e270441e34d622b4e2796f0259.jpg) +(d) Behavior + +Overview. It contains $103k$ real-world driving scenarios, covering diverse urban and suburban scenes. They are further divided into 9s samples with 1s history and 8s future. For end-to-end planning evaluation, besides the ego-vehicle trajectories in each sample as groundtruth, we also have the following key items in the dataset: + +(In-house) Camera Data. Most end-to-end planning methods [28, 49, 53, 61] relies on the camera images as model inputs. In our dataset, each frame contains images captured by eight multi-view cameras. + +High-level Behavior Commands. Like navigation systems for humans (e.g. Google Map), the end-to-end planning system also requires navigation signals to indicate the direction to go. We consider six high-level behavior commands (Fig. 4), namely go straight forward, go straight left, go straight right, do left turn, do right turn, do left U-turn. They can cover diverse driving situations in real world, e.g. "go straight right" describes the case to leave the high-way. Instead of simply considering the last step position [6, 25], we decide the behavior commands based on long-term future trajectories, which can handle the low-speed or stopping situations. Details of this heuristic rule are included in the supplementary materials. + +Evaluation Metrics. The imbalance of data distribution is inevitable in real driving scenarios. For example, in the WOMD-Planning-ADE benchmark, moving straight and stopping account for more than $70\%$ of all the samples, as reflected by Fig. 5. In this case, we argue that current widely + +used sample-wise average displacement error and collision rates cannot comprehensively reflect the performance of motion planning algorithms since challenging but less frequent behaviors, like turning, are overwhelmed by simple straightforward moving scenarios. To this end, we propose novel behavior-wise metrics similar to the mAP metric in prediction [18]. For example, we denote the behavior-wise average displacement error as $bADE$ defined as follows. + +$$ +b A D E = \frac {1}{| \mathcal {B} |} \sum_ {b \in \mathcal {B}} A D E _ {b} \tag {13} +$$ + +where $ADE_{b}$ is the ADE metric for a specific behavior. Concretely, $|\mathcal{B}| = 7$ behaviors are considered - including the six high-level commands (Fig. 4) and an additional stop behavior1. These behavior-wise metrics make up for the shortcomings of previous sample-wise metrics by paying more attention to those scarce but safety-critical behaviors. + +# 5. Experiments + +# 5.1. Implementation Details + +Model and finetuning. Unless otherwise specified, we construct our model from a pretrained PaLI3-5B model [10], which includes a ViT-G (2B) vision encoder [17] and multimodal encoder-decoder (3B) [50]. Our default setting freezes the ViT encoder and finetune only the inserted modules and the multimodal encoder-decoder. Finetuning the model on the WOMD-Planning-ADE dataset requires approximately 2.5 days on 128 Google Cloud TPU v4 using a batch size of 256 and a learning rate of 3e-3. + +Datasets. We evaluate S4-Driver on both WOMD-Planning-ADE (Sec. 4) and nuScenes [5] benchmarks. Given the limited size of nuScenes, which is insufficient for training a multimodal large language model, we also attempt to finetune the model on nuScenes using weights pretrained on WOMD-Planning-ADE. + +# 5.2. Main Results and Comparison + +NuScenes dataset. To widely compare with existing end-to-end planning methods in the community, we evaluate S4-Driver on nuScenes dataset [5]. Tab. 2 show that S4-Driver notably outperforms all the previous algorithms despite its self-supervised training format. Unlike existing works, S4-Driver does not require any perception pretraining or human annotations. This self-supervision nature enables S4-Driver to utilize all the accessible raw trajectory data, whereas these easily collectable raw driving logs offer limited benefit to other algorithms without a labeling process. + +WOMD-Planning-ADE benchmark. In Tab. 3, we primarily compare S4-Driver with the vanilla PaLI3-5B base + +
MethodsHuman AnnotationsExtra DataL2(m)
perception predictiondata labels1.0s2.0s3.0sAvg1,2,3s
Multi-task end-to-end approaches
UniAD [25]✓✓XX0.420.640.910.66
VAD [28]✓✓XX0.170.340.600.37
PARA-Drive [57]✓✓XX0.250.460.740.48
Multimodal large language model approaches
GPT-Driver [38]✓✓XX0.200.400.700.44
DriveVLM [53]✓✓✓✓0.180.340.680.40
OmniDrive [55]✓✓✓✓0.140.290.550.33
Self-supervised approaches
DriveVLM (w/o CoT) [53]XX✓X0.190.410.890.49
S4-Driver (ours)XXXX0.160.340.630.38
S4-Driver* (ours)XX✓X0.130.280.510.31
+ +Table 2. NuScenes benchmark results. Most previous works require human annotations as additional supervision. S4-Driver* is pretrained on WOMD-Planning-ADE, while many prior MLLM-based methods also utilize extra data and labels due to the limited scale of nuScenes. + +
MethodsSample-wise MetricsBehavior-wise Metrics
ADE@1sADE@3sADE@5sbADE@1sbADE@3sbADE@5s
Vanilla PaLI0.0340.2770.7980.0510.3921.069
S4-Driver (ours)0.0310.2470.6930.0490.3500.928
S4-Driver* (ours)0.0290.2330.6550.0440.3130.830
Models that utilize high-quality objects, tracks, and roadgraph as inputs rather than raw camera images.
MotionLM [46]0.0450.2660.6970.0660.3880.978
+ +Table 3. WOMD-Planning-ADE benchmark results. “*” denotes methods pretrained on the internal dataset. + +line (Sec. 3.1) and modular algorithm MotionLM [46]. Compared with vanilla PaLI3-5B, a notable gap is witnessed for both sample-wise and behavior-wise metrics. For reference, we also adapt a latest motion prediction algorithms MotionLM [46] (enhanced internally reproduced version) for planning task by only predicting the future trajectories of ego-vehicle and injecting the high-level command to the model. Because it utilizes auto-labelled high-quality objects, tracklets, and roadgraph signals as model inputs [44], a direct comparison with our end-to-end approaches is not equitable. However, as shown in Tab. 3, S4-Driver still achieves favorable performance compared to MotionLM, particularly in behavior-wise metrics even if S4-Driver uses only raw camera images as inputs. + +# 5.3. Analysis + +Qualitative results. Fig. 6 visualizes planning results in diverse scenarios. Our proposed S4-Driver can determine the future ego-behavior based on the traffic lights and road lane. It can also complete challenging behaviors and deal with different lighting conditions. We will include more visualizations in supplementary materials. + +Meta-decision reliability. Fig. 7 demonstrates the accuracy of meta-decision prediction (Sec. 3.2) on WOMD-Planning-ADE validation set. Across all behaviors, the model provides reliable meta-decision estimations. Without any human annotations, this preliminary prediction can simplify the derivation of numeric motion planning. + +Distribution of sparse volumes. We visualize the distribution of self-supervised learned sparse volumes along x-axis and y-axis over WOMD-Planning-ADE validation set in Fig. 8. From back to front, the sparse volumes concentrate on the close front region. From left to right, the sparse volumes cover all the regions, since there are turning scenarios, but most volumes focus on the middle region. These distributions are consistent with human driving experiences. + +# 5.4. Ablation Studies + +This section presents experiments conducted on the WOMD planning benchmark to clarify the roles of specific modules within our S4-Driver algorithm. More ablation studies are available in the supplementary material. + +MLLM inputs. In Tab. 5, we analyze the roles of camera images and historical ego-states separately using the + +![](images/b57a38af5edcfefae42efab50e00d420fbd94f9e9ba31ad33318457a33442860.jpg) +(a) The ego-vehicle decelerates to wait for the red lights. + +![](images/a242d6bd677ef62f071aa003b89b063187116f455eb4ce94d24c104620834821.jpg) + +![](images/dca1783964db38094cc306b06a4261880974014f040987404bef33518bfbac5a.jpg) + +![](images/b394c33a6a30dddd779d79a8784f8076677617c2f3ae8eda6cfbf391c74dedd6.jpg) +(b) The ego-vehicle conducts a U-turn at the intersection. + +![](images/9a4b374df4ad16769abdb80c57e1b29e54b5901114937703b0a8d855d8b3c3a0.jpg) + +![](images/82111de3a2686af243440848c0e561e66a92d6f806e1df59e29d41c085dbb96b.jpg) + +![](images/48a82813b9bf5ff020bd71902c308c3fd7f7e1d2298d0f0598d156d79cb8183e.jpg) +(c) The ego-vehicle keeps the lane under bad lighting condition. + +![](images/b38df7a6646238cd0619c3f1243d02fcfba905e8207dd20506fb2a94e170d6a8.jpg) + +![](images/b62687f28a1e32765934a27c0f71aa93d9728c5f39a7c1ee383978c374e334d2.jpg) + +![](images/761714070865e371bac2e28b98f29cb4dff11e4cab9163eda65c17ad9a56b946.jpg) +Figure 6. Qualitative results of motion planning. We show the front left, front, front right cameras for each case. +Figure 7. Meta-decision prediction accuracy for each behavior. +Figure 8. Distribution of sparse volumes along x- and y-axes. + +![](images/c2c00654c44c01c12d3ddc039617e87f147d09d33d691adbef46ea6d5e4173f6.jpg) + +vanilla PaLI model. Consistent with [32, 62], the importance of ego-states is acknowledged. Meanwhile, different from their observations on nuScenes benchmark, camera inputs also make great differences on WOMD-Planning-ADE. We assume that WOMD-Planning-ADE covers more diverse driving scenarios including many severe changes of speed and directions, which attaches importance to the sensor data. This also showcases the advantages of WOMD-Planning-ADE for comprehensive evaluation. Tab. 5 also shows that random initialized model cannot converge without MLLM pretraining. Despite different domains, S4-Driver can profit from the strong reasoning ability from large-scale MLLM pretraining on general tasks. + +MLLM capability. In addition to the PaLI3-5B [10] used in other parts, we also tailor PaLI2-3B [9] for motion planning with the same techniques. PaLI2-3B is weaker in parameter numbers (3B vs. 5B), architecture, pretraining manner, and input resolution (224 vs. 448). As in Tab. 4, S4-Driver based on PaLI2-3B demonstrates much inferior performance compared to that based on PaLI3-5B. We conduct experiments on two different training data scales on WOMD-Planning-ADE, i.e. 20k (nuScenes scale) vs. 400k (full WOMD-Planning-ADE). The gap is particularly large given enough training data. It also justifies the necessity to conduct experiments on large-scale dataset, which can fully unleash the potential of powerful MLLMs. + +Sparse volume resolution. Tab. 6 shows the results with + +
MLLMWOMD-Planning-ADE Subset (20k)Full WOMD-Planning-ADE (400k)
ADE@5sbADE@5sADE@5sbADE@5s
PaLI2-3B [9]1.2411.8461.0351.522
PaLI3-5B [10]1.0751.4950.6930.928
+ +Table 4. Ablation studies on planning performance scalability of our methods w.r.t. model and data sizes. We experiment with two MLLMs on two different training data sizes. + +
MethodADE@5sbADE@5s
vanilla PaLI0.7981.069
w/o historical states2.0092.250
w/o camera images1.0811.448
w/o MLLM pretrainingfail to converge
+ +Table 5. Ablation studies on MLLM inputs. + +
ResolutionADE@5sbADE@5s
2m × 2m × 2m0.7981.201
1m × 1m × 2m0.7501.005
1m × 1m × 1m0.7651.051
+ +Table 6. Ablation studies on sparse volume resolutions. + +different resolution for sparse volumes with same number of sparse volumes ( $M = 6000$ ). In line with Fig. 3, low resolution leads to relatively worse performance since it limits the precision of 3D spatial reasoning. Interestingly, higher resolution along $z$ -axis does not definitely improve the model performance since the motion planning mainly works on the $xy$ -plane and too low sparsity ratio tends to make the optimization unstable. + +# 6. Conclusion and Future Work + +This paper introduces S4-Driver, a scalable self-supervised motion planning framework for autonomous driving that utilizes Multimodal Large Language Models (MLLMs). To enhance 3D reasoning in MLLMs, we propose a novel sparse volume representation, which enables improved spatio-temporal reasoning through the aggregation of multi-view and multi-frame image inputs. Additionally, we design behavior-wise metrics for comprehensive evaluation on the large-scale WOMD-Planning-ADE benchmark. S4-Driver achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks, without requiring any human annotations. This demonstrates the potential of self-supervised learning for end-to-end autonomous driving. Future work will involve applying our methods to other powerful MLLM architectures. Our contributions are orthogonal to previous works, such as those that explore multi-task learning or CoT reasoning [49, 55]. Combining our large-scale self-supervised learning approach with supervised finetuning on targeted small-scale labeled data may further enhance the final performance and interpretability of the system. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2 +[4] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 2 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6 +[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2, 6 +[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 3 +[8] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2 +[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations. 1, 2, 8 +[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2, 3, 6, 8 +[11] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial- + +rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 2 +[12] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2 +[13] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2 +[15] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's-eye-view injected multi-modal large models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2024. 2 +[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[18] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5, 6 +[19] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2 +[20] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-bev: What really matters for multi-sensor bev perception? In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2759-2765. IEEE, 2023. 3, 5 +[21] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5, 1 +[22] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances + +in Neural Information Processing Systems, 36:20482-20494, 2023. 2 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[24] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2 +[25] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 6, 7 +[26] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to-end multimodal model for autonomous driving, 2024. 3 +[27] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 3 +[28] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 6, 7 +[29] Jinkyu Kim, Anna Rohrbach, Trevor Darrell, John Canny, and Zeynep Akata. Textual explanations for self-driving vehicles. In Proceedings of the European conference on computer vision (ECCV), pages 563-578, 2018. 6 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2 +[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 5 +[32] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 2, 3, 8 +[33] Bencheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: + +Structured modeling and learning for online vectorized HD map construction. In The Eleventh International Conference on Learning Representations. 2 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[35] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3 +[36] Xianzheng Ma, Yash Bhalgat, Brandon Smart, Shuai Chen, Xinghui Li, Jian Ding, Jindong Gu, Dave Zhenyu Chen, Songyou Peng, Jia-Wang Bian, et al. When llms step into the 3d world: A survey and meta-analysis of 3d tasks via multi-modal large language models. arXiv preprint arXiv:2405.10255, 2024. 3 +[37] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2 +[38] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 7 +[39] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 2 +[40] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2 +[41] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 2 +[42] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2 +[43] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 1 +[44] C. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa T. Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6130-6140, 2021. 7 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[46] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 7, 4 + +[47] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2 +[48] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2 +[49] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In First Vision and Language for Autonomous Driving and Robotics Workshop. 2, 6, 8, 4 +[50] Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, et al. UI2: Unifying language learning paradigms. arXiv preprint arXiv:2205.05131, 2022. 6 +[51] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 3 +[52] Thomas Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. In Workshop on Language and Robot Learning: Language as an Interface. 2 +[53] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7 +[54] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 2 +[55] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 7, 8 +[56] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3 +[57] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 7 +[58] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2 +[59] xAI. Grok-1.5 vision preview, 2024. 2 +[60] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 2 +[61] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 1, 2, 6 +[62] Jiang-Tian Zhai, Ze Feng, Jihao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 3, 8 +[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2 +[64] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2 +[65] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 2 +[66] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2 \ No newline at end of file diff --git a/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/images.zip b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e1e53007e22ccdcea68282ec9892e1d8b9775db --- /dev/null +++ b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47e62b372606a0df31bf19d101914a3b09ec3e103a1b3e0b074a96cb1543f0a4 +size 562052 diff --git a/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/layout.json b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f5b7829ce9ed028b5d9bfb0b3938d163df91a897 --- /dev/null +++ b/2025/S4-Driver_ Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation/layout.json @@ -0,0 +1,10191 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 553, + 140 + ], + "type": "text", + "content": "S4-Driver: Scalable Self-Supervised Driving Multimodal Large Language Model with Spatio-Temporal Visual Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "text", + "content": "Yichen Xie" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\dagger,*}" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "text", + "content": " Runsheng Xu" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "text", + "content": " Tong He" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "text", + "content": " Jyh-Jing Hwang" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "text", + "content": " Katie Luo" + }, + { + "bbox": [ + 104, + 159, + 504, + 175 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 175, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 175, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 134, + 175, + 487, + 190 + ], + "type": "text", + "content": "Jingwei Ji² Hubert Lin² Letian Chen⁴,† Yiren Lu² Zhaoqi Leng²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 216, + 190, + 406, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 190, + 406, + 204 + ], + "spans": [ + { + "bbox": [ + 216, + 190, + 406, + 204 + ], + "type": "text", + "content": "Dragomir Anguelov² Mingxing Tan²" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 203, + 512, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 203, + 512, + 218 + ], + "spans": [ + { + "bbox": [ + 97, + 203, + 512, + 218 + ], + "type": "text", + "content": "1 UC Berkeley, 2 Waymo LLC, 3 Cornell University, 4 Georgia Institute of Technology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 257, + 200, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 257, + 200, + 269 + ], + "spans": [ + { + "bbox": [ + 152, + 257, + 200, + 269 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 282, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 296, + 619 + ], + "type": "text", + "content": "The latest advancements in multi-modal large language models (MLLMs) have spurred a strong renewed interest in end-to-end motion planning approaches for autonomous driving. Many end-to-end approaches rely on human annotations to learn intermediate perception and prediction tasks, while purely self-supervised approaches—which directly learn from sensor inputs to generate planning trajectories without human annotations—often underperform the state of the art. We observe a key gap in the input representation space: end-to-end approaches built on MLLMs are often pretrained with reasoning tasks in 2D image space rather than the native 3D space in which autonomous vehicles plan. To this end, we propose S4-Driver, a scalable self-supervised motion planning algorithm with spatio-temporal visual representation, based on the popular PaLI [9] multimodal large language model. S4-Driver uses a novel sparse volume strategy to seamlessly transform the strong visual representation of MLLMs from perspective view to 3D space without the need to finetune the vision encoder. This representation aggregates multi-view and multiframe visual inputs and enables better prediction of planning trajectories in 3D space. To validate our method, we run experiments on both nuScenes and Waymo Open Motion Dataset (with in-house camera data). Results show that S4-Driver performs favorably against existing supervised multi-task approaches while requiring no human annotations. It also demonstrates great scalability when pretrained on large volumes of unannotated driving logs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 640, + 135, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 135, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 135, + 652 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 295, + 685 + ], + "type": "text", + "content": "The exploration of end-to-end autonomous driving dates back to the 1980s [43], where the motion planning model" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 335, + 247, + 533, + 342 + ], + "blocks": [ + { + "bbox": [ + 335, + 247, + 533, + 342 + ], + "lines": [ + { + "bbox": [ + 335, + 247, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 335, + 247, + 533, + 342 + ], + "type": "image", + "image_path": "d3a2e1d187e9baec01ed63e52bf09edb33d55e8b892eec6d7070f9f3c53b5193.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 342, + 497, + 350 + ], + "lines": [ + { + "bbox": [ + 372, + 342, + 497, + 350 + ], + "spans": [ + { + "bbox": [ + 372, + 342, + 497, + 350 + ], + "type": "text", + "content": "(a) Pretraining data scaling-up of S4-Driver." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 352, + 358, + 380 + ], + "blocks": [ + { + "bbox": [ + 316, + 352, + 358, + 380 + ], + "lines": [ + { + "bbox": [ + 316, + 352, + 358, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 358, + 380 + ], + "type": "image", + "image_path": "829ab413ed8ba1f08f3aaac646d8963e7e37fbb6bf9506d1712950fadf5cd6a6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 385, + 352, + 393 + ], + "lines": [ + { + "bbox": [ + 324, + 385, + 352, + 393 + ], + "spans": [ + { + "bbox": [ + 324, + 385, + 352, + 393 + ], + "type": "text", + "content": "(b) Multi-" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 359, + 352, + 402, + 382 + ], + "blocks": [ + { + "bbox": [ + 359, + 352, + 402, + 382 + ], + "lines": [ + { + "bbox": [ + 359, + 352, + 402, + 382 + ], + "spans": [ + { + "bbox": [ + 359, + 352, + 402, + 382 + ], + "type": "image", + "image_path": "9541e43b9face5cc35a2f5eaa0360ea10c06e441f3875553272b4cbef8206b3a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 380, + 426, + 385 + ], + "lines": [ + { + "bbox": [ + 391, + 380, + 426, + 385 + ], + "spans": [ + { + "bbox": [ + 391, + 380, + 426, + 385 + ], + "type": "text", + "content": "Multi-task Results" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 436, + 352, + 479, + 384 + ], + "blocks": [ + { + "bbox": [ + 436, + 352, + 479, + 384 + ], + "lines": [ + { + "bbox": [ + 436, + 352, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 436, + 352, + 479, + 384 + ], + "type": "image", + "image_path": "57397cda1cf2eba4009f403bda065818ec1ff1132dc0e76999395fc21f870468.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 395, + 554, + 483 + ], + "lines": [ + { + "bbox": [ + 313, + 395, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 554, + 483 + ], + "type": "text", + "content": "Figure 1. S4-Driver scaling up with abundant unlabeled driving logs (1a). Different from past efforts that improve MLLM planning via multi-task learning and CoT reasoning (1b), S4-Driver focuses on self-supervised motion planning (1c) without human annotations. While a vanilla PaLI approach with self-supervised training achieves good performance, the full S4-Driver with spatio-temporal visual representation achieves the best performance and scalability (1a)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 386, + 550, + 395 + ], + "lines": [ + { + "bbox": [ + 436, + 386, + 550, + 395 + ], + "spans": [ + { + "bbox": [ + 436, + 386, + 550, + 395 + ], + "type": "text", + "content": "(c) Self-supervised planning framework." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 479, + 352, + 505, + 384 + ], + "blocks": [ + { + "bbox": [ + 479, + 352, + 505, + 384 + ], + "lines": [ + { + "bbox": [ + 479, + 352, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 479, + 352, + 505, + 384 + ], + "type": "image", + "image_path": "2b85419079a725bc40ad065be321cf7c75bf867e9b539565abad03f24d3cf937.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 507, + 352, + 549, + 384 + ], + "blocks": [ + { + "bbox": [ + 507, + 352, + 549, + 384 + ], + "lines": [ + { + "bbox": [ + 507, + 352, + 549, + 384 + ], + "spans": [ + { + "bbox": [ + 507, + 352, + 549, + 384 + ], + "type": "image", + "image_path": "90ae78a5d1e2f64a66812289e1cf162e095afc983a8e6c6195acad45ff40a2a8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 495, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 579 + ], + "type": "text", + "content": "directly predicts control signals or future waypoints based on raw sensor inputs. Due to limited robustness, these early attempts struggle in complex urban situations. Recent neural network capability advances have led to renewed interest in this field [25, 28, 57], especially given the strong generalization ability of Multimodal Large Language Models (MLLMs) [10, 42, 51]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "However, vanilla application of MLLMs to end-to-end motion planning can hardly exploit their strong visual understanding and reasoning capabilities given the significant difference between motion planning and MLLM pretraining tasks, resulting in inferior planning performance [53, 61]. To narrow this gap, as shown in Fig. 1b, previous methods resort to either multi-task learning, which incorporates diverse perception and prediction tasks into training and inference, or supervised perception pretraining, which utilizes the pretrained autonomous driving perception models as visual tokenizers. However, fine-grained human annotations" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 693, + 225, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 225, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 225, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 693, + 225, + 703 + ], + "type": "text", + "content": " Work done when they were interns at Waymo" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 71, + 704, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 704, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 704, + 138, + 712 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1622" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "serve as the bottleneck for both strategies. In contrast, pure self-supervised approaches, while capable of learning directly from sensor inputs and utilizing the abundant unlabeled driving logs, often under-perform the state of the art." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 192 + ], + "type": "text", + "content": "To challenge this status quo, we aim to boost the performance of self-supervised motion planning orthogonal to prior work (Fig. 1c). It only requires raw sensor data and ego-vehicle trajectories for self-supervised model training, without the need of expensive human labels. First of all, we identify the following two main obstacles in this direction:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 295, + 314 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 194, + 295, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 242 + ], + "type": "text", + "content": "- Suboptimal representation format. MLLMs [9, 10] are typically designed for tasks in the 2D image plane. This image space representation limits their 3D reasoning ability from the combination of multi-view images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 243, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 314 + ], + "type": "text", + "content": "- Limited data scale. Although it is one of the most widely adopted benchmarks for end-to-end planning, nuScenes [5] only contains fewer than 1k sequences and lacks driving behavior diversities. This limited scales leads to severe over-fitting issues in the finetuning of MLLMs with billion-parameter size." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 316, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 615 + ], + "type": "text", + "content": "In this paper, we handle the above challenges to better explore the potential of MLLMs. Firstly, we propose S4-Driver, a simple yet effective Scalable Self-Supervised motion planning method with Spatio-temporal visual representation. Built upon a general multimodal large language model [9, 10], we directly predict ego-vehicle waypoints from camera images, eliminating the need for intermediate perception and prediction tasks, and thus facilitating the scale-up of model pretraining leveraging massive unannotated driving logs (Fig. 1a). To address the obstacle of suboptimal visual format, we propose a novel sparse volume representation that enables aggregating visual information from multi-view and multi-frame images. Moreover, this spatio-temporal representation is designed to work with frozen features from the perspective-view vision encoder. It boosts 3D spatio-temporal reasoning capabilities of the model on motion planning, and seamlessly preserves world knowledge in the pretrained visual embeddings of MLLMs. Secondly, to rigorously evaluate our method with sufficient training data, we also utilize large-scale WOMD-PlanningADE benchmark [18] with in-house camera sensor data besides the popular nuScenes dataset [5]. WOMD-PlanningADE is approximately " + }, + { + "bbox": [ + 55, + 316, + 295, + 615 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 55, + 316, + 295, + 615 + ], + "type": "text", + "content": " larger than nuScenes, so it serves as a more comprehensive benchmark to evaluate model performance in a scalable manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 616, + 226, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 616, + 226, + 627 + ], + "spans": [ + { + "bbox": [ + 66, + 616, + 226, + 627 + ], + "type": "text", + "content": "In summary, our contributions include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 630, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "type": "text", + "content": "- We advocate for a simple self-supervised end-to-end motion planning with MLLMs, solely relying on ego-vehicle trajectory supervision. This approach simplifies the system architecture and facilitates data scaling for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 678, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 678, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 678, + 295, + 714 + ], + "type": "text", + "content": "- We propose a novel sparse volume representation to effectively aggregate multi-view and multi-frame visual information, enhancing the 3D spatio-temporal reasoning" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 515, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 515, + 85 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 515, + 85 + ], + "type": "text", + "content": "capabilities of the MLLM for motion planning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 85, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 85, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 315, + 85, + 553, + 120 + ], + "type": "text", + "content": "- Our algorithm achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks. It also demonstrate great scalability with pretraining." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 131, + 406, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 131, + 406, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 131, + 406, + 144 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 151, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 553, + 354 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs). MLLMs with both language and images modalities have garnered significant attention in recent years [1, 2, 30, 45, 51]. Prior research has focused on integrating powerful Large Language Models (LLMs) [12, 64, 64] with advanced image encoders [19, 45, 63] such as LLaVA [34], PaLI [9, 10], PaliGemma [4], and InstructBLIP [14]. Through instruction tuning or multi-modal finetuning, these models demonstrate impressive performance in multimodal understanding and reasoning. Current developments are leveraging increasingly larger multimodal datasets to further enhance their capabilities in complex perception and generalization tasks [3, 42, 59]. However, despite their strengths, these models have shown limitations in 3D spatial reasoning [11, 22, 40], posing challenges for their application in autonomous driving." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 355, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 521 + ], + "type": "text", + "content": "End-to-end Autonomous Driving. While the key final step of autonomous driving is planning, traditional systems are often assembled with multiple modules for perception [33, 35, 60] and prediction [41, 54, 65] tasks. To mitigate the information loss and error accumulation across modules, end-to-end driving systems [6, 24, 25, 28, 32, 66] utilize a unified model to predict the ego-vehicle future waypoints or control signals directly from raw sensor inputs. While prioritizing planning, these systems typically still incorporate perception and prediction, requiring explicit supervision for each. Although some prior works [13, 46, 58] have explored pure motion planning without any intermediate tasks, they struggle in complex urban scenarios due to modeling capability limitations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "type": "text", + "content": "MLLMs for Driving. The superior reasoning and generalization abilities of large-scale models are desirable properties for autonomous driving. Some efforts transform the driving scenario into large language model textual prompts [8, 38, 47] or directly process camera images with vision-language models [15, 48, 61]. However, their potential is limited to the small scale of existing benchmarks [5], which only allow partial finetuning [23]. Meanwhile, closed-loop simulators [16] struggle to provide realistic sensor data for end-to-end tasks. Consequently, multi-task joint finetuning [49, 53, 61] or chain-of-though (CoT) [56] inference [37-39] are widely adopted to simplify the reasoning despite the extra annotation requirements brought by perception and prediction modules. Alternatively, some work [15, 52] integrates pretrained perception models [31] to extract BEV features and sends to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1623" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "language models as vision tokens. Recently, EMMA [26] leverages powerful Gemini [51] for self-supervised motion planning. Additionally, they develop a mixture of training tasks, including motion planning, 3D object detection, and road graph estimation as well as additional reasoning process for trajectory generation. In contrast, our work focuses on the enhancement of self-supervised motion planning without extra human labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 275, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 275, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 275, + 192 + ], + "type": "text", + "content": "3. Enhancing MLLM for Motion Planning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 294, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 294, + 282 + ], + "type": "text", + "content": "In this section, we adapt a pretrained MLLM for the end-to-end motion planning task. Without loss of generality, we leverage PaLI [10]. However, these techniques are also compatible with other more powerful MLLMs. We introduce a set of design improvements to the MLLM setup and demonstrate correspondingly improved performance on WOMD-Planning-ADE (Fig. 3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 190, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 190, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 190, + 301 + ], + "type": "text", + "content": "3.1. Vanilla PaLI as Planner" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": "End-to-end motion planning models determine the future trajectory " + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{T_f}" + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": " of ego-vehicle based on multi-view camera images " + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": " and high-level behavior command " + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": ". The future trajectory consists of the ego-vehicle's location at each future time step in the BEV coordinate, i.e. " + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{T_f} = [(x_t,y_t)]_{t=1}^{T_f}" + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": ". The high-level behavior command is necessary to guide the driving direction, similar to a navigation system. Additionally, as verified in prior works [32, 62], the historical states of ego-vehicle " + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{T_h}" + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": " are also critical for a smooth and feasible planning result, where we consider the historical locations, velocities, and accelerations as " + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{T_h} = [\\mathbf{l}_t,\\mathbf{v}_t,\\mathbf{a}_t]_{t=-1}^{-T_h}" + }, + { + "bbox": [ + 55, + 308, + 295, + 455 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 463, + 295, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 463, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 132, + 463, + 295, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {O} _ {T _ {f}} = P (\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b) \\tag {1}", + "image_path": "0820460492f93c1cb0afea2967d4760481309bcc3862f3fca29cc9daa3066070.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 295, + 568 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 55, + 485, + 295, + 568 + ], + "type": "text", + "content": " is the planning model. We provide both historical ego states and high-level commands to the model as text prompts. Locations, velocities, and accelerations are represented directly as floating numbers with two decimals. The predicted future trajectories are then extracted from the decoded text outputs of the model. Example prompts and targets are provided in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 617 + ], + "type": "text", + "content": "Without preceding perception and prediction tasks, the vanilla PaLI, fine-tuned in a self-supervised manner, achieves reasonable but not ideal performance in motion planning (Fig. 3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 275, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 275, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 275, + 636 + ], + "type": "text", + "content": "3.2. Hierarchical Planning with Meta-Decision" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Directly outputting a future trajectory without any reasoning is challenging for MLLMs. To address this, we apply a coarse-to-fine approach, inspired by Chain-of-Thought (CoT) [56]. It adopts a hierarchical planning method [27], going from semantic decision to numeric planning. We prompt our model to provide a high-level estimation of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "the future acceleration state " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " first, effectively decomposing the motion planning task into two steps:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 101, + 553, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 101, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 327, + 101, + 553, + 114 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = P _ {s 1} \\left(\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b\\right), \\quad \\mathcal {O} _ {T _ {f}} = P _ {s 2} \\left(\\mathcal {C}, \\mathcal {H} _ {T _ {h}}, b; \\mathcal {D}\\right) \\tag {2}", + "image_path": "ed882bd6452cd662f61af5466376f7fd4a8729c565b1cd05b6aa38610a2097e4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 118, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 261 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 313, + 118, + 555, + 261 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 118, + 555, + 261 + ], + "type": "text", + "content": " to encompass four meta-decisions: keep stationary, keep speed, accelerate and decelerate. Notably, unlike previous CoT applications in VLM-based planning that require human annotation for training, such as the scene analysis in DriveVLM [53], we incorporate these meta-decisions as a \"free lunch\" to simplify the motion planning process without requiring any additional annotations. The ground-truth decisions are generated by heuristic rules based on the future ego velocity and acceleration (details in the supplementary materials). Fig. 3 shows this simple design brings non-trivial improvements in planning performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 267, + 495, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 495, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 495, + 281 + ], + "type": "text", + "content": "3.3. Scene Representation in 3D Space" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 285, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 405 + ], + "type": "text", + "content": "High-quality motion planning requires a robust understanding of the surrounding 3D scene, including both static and dynamic elements. While traditionally achieved through separate perception and prediction modules, our self-supervised end-to-end framework relies on the MLLM to implicitly learn this understanding without explicit supervision. However, despite strong 2D reasoning capabilities, MLLMs often struggle with 3D spatial reasoning due to limitations in their perspective view representation and the lack of depth-related tasks in pretraining [7, 36]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 409, + 539, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 539, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 539, + 421 + ], + "type": "text", + "content": "3.3.1. 3D Visual Representation with Dense Volumes" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": "To overcome the above limitations, we draw inspiration from previous successful perception tasks [31, 35] and adopt a 3D volume representation. The multi-view feature maps " + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{2D} = \\{\\mathbf{f}_{2D}^v\\}_{v=1}^V (\\mathbf{f}_{2D}^v \\in \\mathbb{R}^{H \\times W \\times C})" + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": " are extracted by visual encoder of MLLM, where " + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": " is the number of views. We construct a 3D feature volume " + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D}^{vol} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}" + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": " centered at the ego-vehicle based on the multi-view image features. To avoid introducing complex modules that might disrupt the pretrained MLLM and misalign visual features with the subsequent multimodal encoder-decoder, we employ a lightweight projection method, similar to SimpleBEV [20]. Specifically, for each voxel in the 3D volume, we project its " + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": " coordinates to each perspective view " + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": ", obtaining corresponding 2D coordinates " + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "inline_equation", + "content": "(u_v, v_v)" + }, + { + "bbox": [ + 313, + 424, + 555, + 674 + ], + "type": "text", + "content": ". We then bilinearly sample the local features from each view at these projected locations. Finally, the voxel's feature representation is computed as the average of the local semantic features from all views where the voxel projects within the image bounds. This process effectively incorporates 3D spatial information while maintaining compatibility with the pretrained MLLM." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 351, + 678, + 553, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 678, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 351, + 678, + 553, + 694 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {3 D, (x, y, z)} ^ {v o l} = \\mathbf {f} _ {(x, y, z)} ^ {s e m} + \\operatorname {P o s E m b} (x, y, z) \\tag {3}", + "image_path": "16bd78f13f7679a9ad2afc5909677ff4e568dcecfaf095d86a4052d7c9ae7003.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 696, + 553, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 696, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 345, + 696, + 553, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {(x, y, z)} ^ {s e m} = \\operatorname {A v g} \\left(\\left\\{\\text {B i l i n e a r} \\left(\\mathbf {f} _ {2 D} ^ {v}; \\left(u _ {v}, v _ {v}\\right)\\right) \\right\\} _ {v}\\right) \\tag {4}", + "image_path": "673634a05b4ce32b9e56034d1645c7e7bad603112bdaa379d53f8a30784fe9d8.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1624" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 74, + 538, + 286 + ], + "blocks": [ + { + "bbox": [ + 70, + 74, + 538, + 286 + ], + "lines": [ + { + "bbox": [ + 70, + 74, + 538, + 286 + ], + "spans": [ + { + "bbox": [ + 70, + 74, + 538, + 286 + ], + "type": "image", + "image_path": "739a83147b635f3fa2284d23b820d3da079b7878aa3588d0fc23335bcb8956a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 288, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 553, + 312 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed S4-Driver algorithm. We enhance the PaLI model for motion planning by incorporating meta-decision, spatio-temporal visual representation, and multi-decoding aggregation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 316, + 282, + 529 + ], + "blocks": [ + { + "bbox": [ + 72, + 316, + 282, + 529 + ], + "lines": [ + { + "bbox": [ + 72, + 316, + 282, + 529 + ], + "spans": [ + { + "bbox": [ + 72, + 316, + 282, + 529 + ], + "type": "image", + "image_path": "ba84bc19dd0f40b56251e9153e600bae71d43fff973a1a56a1bae8e5095dc5d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 529, + 295, + 584 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 584 + ], + "type": "text", + "content": "Figure 3. A roadmap for enhancing MLLM for planning. The techniques are adopted step-by-step from top to bottom, while shadow items are not adopted in the subsequent steps. We show the performance on WOMD-Planning-ADE (Sec. 4) after including each module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "type": "text", + "content": "This simple and efficient projection strategy ensures that the 3D volume features " + }, + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D}^{vol}" + }, + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "type": "text", + "content": " share a similar distribution with the original multi-view features " + }, + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{2D}" + }, + { + "bbox": [ + 55, + 591, + 295, + 651 + ], + "type": "text", + "content": ". This similarity facilitates seamless integration with the subsequent multimodal encoder-decoder of the MLLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3, this 3D volume representation leads to some improvement in motion planning performance. Alternatively, we find that reducing the " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "-axis using a fully connected layer to obtain a BEV representation yields slightly worse performance since this reduction operation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 322, + 528, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 528, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 528, + 334 + ], + "type": "text", + "content": "may introduce ambiguity to the scene representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 339, + 473, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 473, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 473, + 352 + ], + "type": "text", + "content": "3.3.2. Sparse Volume Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "content": "While the 3D volume representation effectively captures spatial information, inherently most of the surrounding 3D space is empty. Moreover, for motion planning, detailed information about objects distant from the road, such as buildings and trees, is less critical. Based on this observation, we propose a sparse volume representation to reduce the number of voxels, enabling higher resolution given memory constraints and improving efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "text", + "content": "To determine useful volumes in each scene based on their positions and semantics, we define a gate value " + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "inline_equation", + "content": "g_{(x,y,z)}^{vol} \\in (0,1)" + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "text", + "content": " for each volume at coordinates " + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "text", + "content": ". To obtain this gate, starting from multi-view image features " + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_{2D}^v\\}_{v=1}^V" + }, + { + "bbox": [ + 313, + 451, + 556, + 524 + ], + "type": "text", + "content": ", we reduce its dimension through a fully connect (FC) layer as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 532, + 555, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 532, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 319, + 532, + 555, + 548 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {2 D} ^ {\\text {g a t e}, v} = \\operatorname {F C} \\left(\\mathbf {f} _ {2 D} ^ {v}\\right), \\quad \\mathbf {f} _ {2 D} ^ {\\text {g a t e}, v} \\in \\mathbb {R} ^ {H \\times W \\times C ^ {\\prime}}, C ^ {\\prime} \\ll C. \\tag {5}", + "image_path": "eba6ec9d581d34a6c095c235d60c3764db825afd7365e3b602838fb20834bc0c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": "Then we construct a reduced-dimension volume feature " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D}^{gate} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C'}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_{2D}^{gate,v}\\}_{v=1}^V" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": " same as Eq. 3. A smaller channel number allows for larger volume resolution and is sufficient to indicate whether a volume is related to motion planning. Afterward, the gate is derived from " + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D}^{gate}" + }, + { + "bbox": [ + 313, + 556, + 555, + 628 + ], + "type": "text", + "content": " through a small MLP module as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 637, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 637, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 353, + 637, + 553, + 658 + ], + "type": "interline_equation", + "content": "g _ {(x, y, z)} ^ {v o l} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\mathbf {f} _ {3 D, (x, y, z)} ^ {g a t e}\\right)\\right) \\tag {6}", + "image_path": "0b78aa1b0a973c20101e863ee7cb7acd59b85c874fd7719367a89f43ac416f03.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "text", + "content": "Thus, we can easily select " + }, + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "text", + "content": " volumes " + }, + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "inline_equation", + "content": "(M\\ll X\\cdot Y\\cdot Z)" + }, + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "text", + "content": " with largest gate values with coordinates " + }, + { + "bbox": [ + 313, + 665, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\{(x_i,y_i,z_i)\\}_{i = 1}^M" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "content": "As we do not have access to ground-truth occupancy state, we propose learning the gate values implicitly. We" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1625" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "assume that regions with small gate values should be empty or irrelevant to planning. For these vacant space, we assign a learnable feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{vac} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ". We expect the model can learn the gate value via the trade-off between the semantic features and this vacant feature at each 3D position, so we fetch features for selected sparse volumes as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 151, + 295, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 151, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 86, + 151, + 295, + 168 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} = \\hat {\\mathbf {f}} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} + \\operatorname {P o s E m b} (x, y, z) \\tag {7}", + "image_path": "194b612c8866b01e5275ffbec83a76141c0670432f205705c4739288939a709d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 170, + 295, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 170, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 64, + 170, + 295, + 187 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} _ {3 D, (x, y, z)} ^ {s p a r s e} = g _ {(x, y, z)} ^ {v o l} \\cdot \\mathbf {f} _ {(x, y, z)} ^ {s e m} + \\left(1 - g _ {(x, y, z)} ^ {v o l}\\right) \\cdot \\mathbf {f} _ {v a c} \\tag {8}", + "image_path": "eead4e766b572446d2dd1f337c0e59bbdfdf00e8fd740037e10d22943c91f0f6.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{(x,y,z)}^{sem}" + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "text", + "content": " is same as Eq. 4 and " + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "inline_equation", + "content": "(x,y,z)\\in \\{(x_i,y_i,z_i)\\}_{i = 1}^M" + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "text", + "content": " are selected volumes with large gate values. When the sparse volume features " + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D,(x,y,z)}^{sparse}" + }, + { + "bbox": [ + 55, + 194, + 296, + 281 + ], + "type": "text", + "content": " are fed into the subsequent multimodal encoder, they notably boost the planning performance (Fig. 3) by simultaneously providing explicit 3D spatial cues of surrounding scenes and concentrate on the important regions adaptively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 251, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 251, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 251, + 298 + ], + "type": "text", + "content": "3.3.3. Local Feature Aggregation in 3D Space" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": "Due to the lack of depth information, the lifting process in Eq. 3 or 7 results in duplicate volume features along each camera ray. This spatial ambiguity can be mitigated by 3D local operations like convolutions [20] or deformable attention [31] as demonstrated in previous work. However, in our MLLM framework, it is less economical to insert extra local operations. Instead, we inject some relative position bias into the existing multimodal encoder by tailoring the self-attention as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 414, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 414, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 72, + 414, + 295, + 442 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}} - b (D)\\right) V \\tag {9}", + "image_path": "49ebb1efd7f2628a45e98935575506018d5555840ab18886a1b359f1c7a56c87.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": " sparse volumes with coordinates " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\{(x_i, y_i, z_i)\\}_{i=1}^M" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": ", the distance matrix " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{M \\times M \\times 3}" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": " is the distance along " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "xyz" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": "-axes between each pair of sparse volumes, i.e. " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "D_{ij} = (x_j - x_i, y_j - y_i, z_j - z_i)" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": ". The bias is computed with the function " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "b(\\Delta x, \\Delta y, \\Delta z) = b_x(\\Delta x) + b_y(\\Delta y) + b_z(\\Delta z)" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\Delta x, \\Delta y, \\Delta z" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": " are divided into several bins and mapped to a learnable bias value for each bin through " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "b_x(\\cdot), b_y(\\cdot), b_z(\\cdot)" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": ". We also apply a separate bias between the text tokens with 1D position " + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "inline_equation", + "content": "\\{p_i\\}" + }, + { + "bbox": [ + 55, + 449, + 296, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 629 + ], + "type": "text", + "content": "This relative position bias elegantly inserts local inductive bias to pretrained global self-attention modules with minimal additional costs. It leads to a notable performance gain in Fig. 3 by facilitating the aggregation of local information in 3D space, enhancing scene understanding and spatial reasoning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 219, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 219, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 219, + 649 + ], + "type": "text", + "content": "3.4. Multi-frame Temporal Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Multi-frame input combination helps to compensate for the lack of depth cues in camera images. We extend our sparse volume representation to aggregate mult-frame temporal information by incorporating " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " historical frames with an interval of 0.5s. Given images from the total " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "T + 1" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " frames," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "Eq. 5 is applied to each frame separately to obtain the multi-view image features " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{2D,t}^{gate,v}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "t = 0, -1, \\dots, -T" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "v = 1, \\dots, V" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ". After ego-motion compensation, we construct a gate feature volume " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D,t}^{gate}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " based on the current ego-vehicle coordinates and per-frame image features. Multi-frame volume features are concatenated along the channel dimension to generate gate value." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 161, + 553, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 161, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 161, + 553, + 191 + ], + "type": "interline_equation", + "content": "\\left. g _ {(x, y, z)} ^ {v o l} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\operatorname {C o n c a t} \\left([ \\mathbf {f} _ {3 D, t, (x, y, z)} ^ {g a t e} ] _ {t = 0} ^ {- T}\\right)\\right)\\right) \\right. \\tag {10}", + "image_path": "42d8930a74a9fd5ca71b0f9472cf2a3e8dc5ccf0d3efc8b3b6e2e0e1f75fb702.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "text", + "content": "Similar with Sec. 3.3.2, we select " + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "text", + "content": " volumes based on the gate value. Following Eq. 7, the volume features are fetched from each frame image feature map separately as " + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{3D,t,(x,y,z)}^{sparse}" + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "inline_equation", + "content": "t = 0,\\dots , - T" + }, + { + "bbox": [ + 313, + 192, + 554, + 253 + ], + "type": "text", + "content": ", which are fused with an FC-layer as the sparse volume feature with temporal awareness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 258, + 553, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 258, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 325, + 258, + 553, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {3 D, (x, y, z)} ^ {\\text {s p a r s e}} = \\operatorname {F C} \\left(\\operatorname {C o n c a t} \\left(\\left[ \\mathbf {f} _ {3 D, t, (x, y, z)} ^ {\\text {s p a r s e}} \\right] _ {t = 0} ^ {- T}\\right)\\right) \\tag {11}", + "image_path": "ca4b5b85362a7d288b7e4da283f0b5cbce7933da3e97163576eff482f4db2ea2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 290, + 554, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 314 + ], + "type": "text", + "content": "In Fig. 3, temporal fusion helps to improve motion planning performance by facilitating the environment understanding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 320, + 522, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 522, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 522, + 333 + ], + "type": "text", + "content": "3.5. Voting for Planning via Multi-Decoding" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "type": "text", + "content": "MLLMs are prone to assigning high confidence to the simple future behaviors in motion planning, e.g., keeping stationary. To relieve this bias, we aggregate multiple outputs and use voting to obtain the final planning output. We resort to nucleus sampling [21] to produce multiple future trajectories for the ego-vehicle, denoted as " + }, + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}_{T_f}^k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "type": "text", + "content": ". They are simply averaged to produce a unique planning result as follows: " + }, + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{T_t} = [(x_t, y_t)]_{t=1}^{T_f}" + }, + { + "bbox": [ + 313, + 338, + 554, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 366, + 441, + 553, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 441, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 366, + 441, + 553, + 475 + ], + "type": "interline_equation", + "content": "\\left(x _ {t}, y _ {t}\\right) = \\left(\\frac {1}{K} \\sum_ {k = 1} ^ {K} x _ {t} ^ {k}, \\frac {1}{K} \\sum_ {k = 1} ^ {K} y _ {t} ^ {k}\\right) \\tag {12}", + "image_path": "338700e82a22d50bbe7a777b328135d6f076f3804e98959f0ec0d32e6012e1f7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "type": "text", + "content": "This unweighted average mitigates the MLLM's bias towards simple behaviors. Fig. 3 demonstrates the notable performance gain with this simple multi-decoding aggregation approach." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 534, + 528, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 528, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 528, + 548 + ], + "type": "text", + "content": "3.6. Scaling to Large-scale Raw Driving Logs" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 635 + ], + "type": "text", + "content": "The self-supervised training allows our proposed S4-Driver to scale to large-scale driving logs amounts without requiring human annotations. To exploit the potential of MLLM-based planners, we pretrain the model on our internal dataset. Results in Fig. 3 demonstrate that S4-Driver achieves notable performance gains in challenging tail behaviors due to large-scale pretraining." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 645, + 548, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 548, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 548, + 659 + ], + "type": "text", + "content": "4. Waymo Open Motion Dataset for Planning" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "For the large-scale training and evaluation of planning algorithms with large models, we design a WOMD-Planning-ADE benchmark based on WOMD dataset [18], and compare with the existing datasets in Tab. 1." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1626" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 112 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 112 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 112 + ], + "type": "table", + "html": "
BenchmarkSizeData FormatMetrics
SequenceLengthFrequencyPlanning horizonSample-wiseBehavior-wise
BDD-X [29]7k77 h10 Hzcontrol signal✓✗
nuScenes [5]1k5.5 h2 Hz3s✓✗
WOMD-Planning-ADE103k574 h10 Hz5s✓✓
", + "image_path": "6d1d00dedf02e423018ade78b06c1e52ac24687376d81ee985d8793eb8284c51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 113, + 295, + 145 + ], + "lines": [ + { + "bbox": [ + 55, + 113, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 113, + 295, + 145 + ], + "type": "text", + "content": "Table 1. Comparisons between WOMD-Planning-ADE and existing counterparts. WOMD-Planning-ADE has a significantly larger size and more comprehensive behavior-wise metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 160, + 278, + 223 + ], + "blocks": [ + { + "bbox": [ + 70, + 160, + 278, + 223 + ], + "lines": [ + { + "bbox": [ + 70, + 160, + 278, + 223 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 278, + 223 + ], + "type": "image", + "image_path": "2f6553fb36d69809aef5a53d32a4bb7f8d2a4628d9d0ae019004db293332b883.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 227, + 289, + 239 + ], + "lines": [ + { + "bbox": [ + 61, + 227, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 227, + 289, + 239 + ], + "type": "text", + "content": "Figure 4. High-level behavior commands for motion planning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 252, + 99, + 311 + ], + "blocks": [ + { + "bbox": [ + 59, + 252, + 99, + 311 + ], + "lines": [ + { + "bbox": [ + 59, + 252, + 99, + 311 + ], + "spans": [ + { + "bbox": [ + 59, + 252, + 99, + 311 + ], + "type": "image", + "image_path": "7fcdefac43c64057dd57f7e4490847e95b28d7981f80f3001cb1004791109d8f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 312, + 99, + 320 + ], + "lines": [ + { + "bbox": [ + 59, + 312, + 99, + 320 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 99, + 320 + ], + "type": "text", + "content": "(a) Trajectory" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 102, + 260, + 164, + 309 + ], + "blocks": [ + { + "bbox": [ + 102, + 260, + 164, + 309 + ], + "lines": [ + { + "bbox": [ + 102, + 260, + 164, + 309 + ], + "spans": [ + { + "bbox": [ + 102, + 260, + 164, + 309 + ], + "type": "image", + "image_path": "137fa8d8e21b0ffbd4d73a760e190f6ce2e9ca7d7478bd2ad8a03286a5c07c36.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 312, + 150, + 320 + ], + "lines": [ + { + "bbox": [ + 116, + 312, + 150, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 312, + 150, + 320 + ], + "type": "text", + "content": "(b) Velocity" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 167, + 260, + 233, + 309 + ], + "blocks": [ + { + "bbox": [ + 167, + 260, + 233, + 309 + ], + "lines": [ + { + "bbox": [ + 167, + 260, + 233, + 309 + ], + "spans": [ + { + "bbox": [ + 167, + 260, + 233, + 309 + ], + "type": "image", + "image_path": "10e1af860f183a1bf01a55bc3beda5fa15b02ead192d6b3ee01b34eef48c9a38.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 312, + 223, + 319 + ], + "lines": [ + { + "bbox": [ + 176, + 312, + 223, + 319 + ], + "spans": [ + { + "bbox": [ + 176, + 312, + 223, + 319 + ], + "type": "text", + "content": "(c) Acceleration" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 331, + 296, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 364 + ], + "type": "text", + "content": "Figure 5. Statistical distributions of WOMD-Planning-ADE. We show the distributions of ego-vehicle trajectories, velocities, accelerations and each behavior scenario sizes in the validation set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 251, + 293, + 308 + ], + "blocks": [ + { + "bbox": [ + 235, + 251, + 293, + 308 + ], + "lines": [ + { + "bbox": [ + 235, + 251, + 293, + 308 + ], + "spans": [ + { + "bbox": [ + 235, + 251, + 293, + 308 + ], + "type": "image", + "image_path": "889b62746d6df0156b0bc997472c14a36afe55e270441e34d622b4e2796f0259.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 312, + 282, + 320 + ], + "lines": [ + { + "bbox": [ + 246, + 312, + 282, + 320 + ], + "spans": [ + { + "bbox": [ + 246, + 312, + 282, + 320 + ], + "type": "text", + "content": "(d) Behavior" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "type": "text", + "content": "Overview. It contains " + }, + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "type": "inline_equation", + "content": "103k" + }, + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "type": "text", + "content": " real-world driving scenarios, covering diverse urban and suburban scenes. They are further divided into 9s samples with 1s history and 8s future. For end-to-end planning evaluation, besides the ego-vehicle trajectories in each sample as groundtruth, we also have the following key items in the dataset:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 449, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 496 + ], + "type": "text", + "content": "(In-house) Camera Data. Most end-to-end planning methods [28, 49, 53, 61] relies on the camera images as model inputs. In our dataset, each frame contains images captured by eight multi-view cameras." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "content": "High-level Behavior Commands. Like navigation systems for humans (e.g. Google Map), the end-to-end planning system also requires navigation signals to indicate the direction to go. We consider six high-level behavior commands (Fig. 4), namely go straight forward, go straight left, go straight right, do left turn, do right turn, do left U-turn. They can cover diverse driving situations in real world, e.g. \"go straight right\" describes the case to leave the high-way. Instead of simply considering the last step position [6, 25], we decide the behavior commands based on long-term future trajectories, which can handle the low-speed or stopping situations. Details of this heuristic rule are included in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. The imbalance of data distribution is inevitable in real driving scenarios. For example, in the WOMD-Planning-ADE benchmark, moving straight and stopping account for more than " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " of all the samples, as reflected by Fig. 5. In this case, we argue that current widely" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "used sample-wise average displacement error and collision rates cannot comprehensively reflect the performance of motion planning algorithms since challenging but less frequent behaviors, like turning, are overwhelmed by simple straightforward moving scenarios. To this end, we propose novel behavior-wise metrics similar to the mAP metric in prediction [18]. For example, we denote the behavior-wise average displacement error as " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "bADE" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " defined as follows." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 381, + 175, + 555, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 175, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 381, + 175, + 555, + 204 + ], + "type": "interline_equation", + "content": "b A D E = \\frac {1}{| \\mathcal {B} |} \\sum_ {b \\in \\mathcal {B}} A D E _ {b} \\tag {13}", + "image_path": "d8c00fcb52b0382f75f2be5b568c1a3adec9d075c6ad50734224e3690215450c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "inline_equation", + "content": "ADE_{b}" + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "content": " is the ADE metric for a specific behavior. Concretely, " + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}| = 7" + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "content": " behaviors are considered - including the six high-level commands (Fig. 4) and an additional stop behavior1. These behavior-wise metrics make up for the shortcomings of previous sample-wise metrics by paying more attention to those scarce but safety-critical behaviors." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 293, + 395, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 293, + 395, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 395, + 306 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 312, + 447, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 447, + 325 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 447, + 325 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 330, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 437 + ], + "type": "text", + "content": "Model and finetuning. Unless otherwise specified, we construct our model from a pretrained PaLI3-5B model [10], which includes a ViT-G (2B) vision encoder [17] and multimodal encoder-decoder (3B) [50]. Our default setting freezes the ViT encoder and finetune only the inserted modules and the multimodal encoder-decoder. Finetuning the model on the WOMD-Planning-ADE dataset requires approximately 2.5 days on 128 Google Cloud TPU v4 using a batch size of 256 and a learning rate of 3e-3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "Datasets. We evaluate S4-Driver on both WOMD-Planning-ADE (Sec. 4) and nuScenes [5] benchmarks. Given the limited size of nuScenes, which is insufficient for training a multimodal large language model, we also attempt to finetune the model on nuScenes using weights pretrained on WOMD-Planning-ADE." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 516, + 480, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 480, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 480, + 529 + ], + "type": "text", + "content": "5.2. Main Results and Comparison" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "text", + "content": "NuScenes dataset. To widely compare with existing end-to-end planning methods in the community, we evaluate S4-Driver on nuScenes dataset [5]. Tab. 2 show that S4-Driver notably outperforms all the previous algorithms despite its self-supervised training format. Unlike existing works, S4-Driver does not require any perception pretraining or human annotations. This self-supervision nature enables S4-Driver to utilize all the accessible raw trajectory data, whereas these easily collectable raw driving logs offer limited benefit to other algorithms without a labeling process." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 653, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 677 + ], + "type": "text", + "content": "WOMD-Planning-ADE benchmark. In Tab. 3, we primarily compare S4-Driver with the vanilla PaLI3-5B base" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "1We do not include \"stop\" in the high-level behavior commands since navigation system cannot provide it and this may leak future information like traffic lights. Details are explained in supplementary materials." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1627" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 522, + 289 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 522, + 289 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 522, + 289 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 522, + 289 + ], + "type": "table", + "html": "
MethodsHuman AnnotationsExtra DataL2(m)
perception predictiondata labels1.0s2.0s3.0sAvg1,2,3s
Multi-task end-to-end approaches
UniAD [25]✓✓XX0.420.640.910.66
VAD [28]✓✓XX0.170.340.600.37
PARA-Drive [57]✓✓XX0.250.460.740.48
Multimodal large language model approaches
GPT-Driver [38]✓✓XX0.200.400.700.44
DriveVLM [53]✓✓✓✓0.180.340.680.40
OmniDrive [55]✓✓✓✓0.140.290.550.33
Self-supervised approaches
DriveVLM (w/o CoT) [53]XX✓X0.190.410.890.49
S4-Driver (ours)XXXX0.160.340.630.38
S4-Driver* (ours)XX✓X0.130.280.510.31
", + "image_path": "8967db0a8878d56840888be7c15210ed7adfbf558acf0a973740365f29719164.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 319, + 523, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 291, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 555, + 313 + ], + "type": "text", + "content": "Table 2. NuScenes benchmark results. Most previous works require human annotations as additional supervision. S4-Driver* is pretrained on WOMD-Planning-ADE, while many prior MLLM-based methods also utilize extra data and labels due to the limited scale of nuScenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 319, + 523, + 426 + ], + "lines": [ + { + "bbox": [ + 86, + 319, + 523, + 426 + ], + "spans": [ + { + "bbox": [ + 86, + 319, + 523, + 426 + ], + "type": "table", + "html": "
MethodsSample-wise MetricsBehavior-wise Metrics
ADE@1sADE@3sADE@5sbADE@1sbADE@3sbADE@5s
Vanilla PaLI0.0340.2770.7980.0510.3921.069
S4-Driver (ours)0.0310.2470.6930.0490.3500.928
S4-Driver* (ours)0.0290.2330.6550.0440.3130.830
Models that utilize high-quality objects, tracks, and roadgraph as inputs rather than raw camera images.
MotionLM [46]0.0450.2660.6970.0660.3880.978
", + "image_path": "86cb56c90a344896416dd477797aec9c48f8c38a7211e6cf60259da25c3cd353.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 426, + 499, + 437 + ], + "lines": [ + { + "bbox": [ + 109, + 426, + 499, + 437 + ], + "spans": [ + { + "bbox": [ + 109, + 426, + 499, + 437 + ], + "type": "text", + "content": "Table 3. WOMD-Planning-ADE benchmark results. “*” denotes methods pretrained on the internal dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 446, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 446, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 446, + 295, + 614 + ], + "type": "text", + "content": "line (Sec. 3.1) and modular algorithm MotionLM [46]. Compared with vanilla PaLI3-5B, a notable gap is witnessed for both sample-wise and behavior-wise metrics. For reference, we also adapt a latest motion prediction algorithms MotionLM [46] (enhanced internally reproduced version) for planning task by only predicting the future trajectories of ego-vehicle and injecting the high-level command to the model. Because it utilizes auto-labelled high-quality objects, tracklets, and roadgraph signals as model inputs [44], a direct comparison with our end-to-end approaches is not equitable. However, as shown in Tab. 3, S4-Driver still achieves favorable performance compared to MotionLM, particularly in behavior-wise metrics even if S4-Driver uses only raw camera images as inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 118, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 118, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 118, + 635 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Qualitative results. Fig. 6 visualizes planning results in diverse scenarios. Our proposed S4-Driver can determine the future ego-behavior based on the traffic lights and road lane. It can also complete challenging behaviors and deal with different lighting conditions. We will include more visualizations in supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 446, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 518 + ], + "type": "text", + "content": "Meta-decision reliability. Fig. 7 demonstrates the accuracy of meta-decision prediction (Sec. 3.2) on WOMD-Planning-ADE validation set. Across all behaviors, the model provides reliable meta-decision estimations. Without any human annotations, this preliminary prediction can simplify the derivation of numeric motion planning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 519, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 556, + 615 + ], + "type": "text", + "content": "Distribution of sparse volumes. We visualize the distribution of self-supervised learned sparse volumes along x-axis and y-axis over WOMD-Planning-ADE validation set in Fig. 8. From back to front, the sparse volumes concentrate on the close front region. From left to right, the sparse volumes cover all the regions, since there are turning scenarios, but most volumes focus on the middle region. These distributions are consistent with human driving experiences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 624, + 416, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 416, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 416, + 635 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 689 + ], + "type": "text", + "content": "This section presents experiments conducted on the WOMD planning benchmark to clarify the roles of specific modules within our S4-Driver algorithm. More ablation studies are available in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "MLLM inputs. In Tab. 5, we analyze the roles of camera images and historical ego-states separately using the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1628" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 134, + 114 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 134, + 114 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 134, + 114 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 134, + 114 + ], + "type": "image", + "image_path": "b57a38af5edcfefae42efab50e00d420fbd94f9e9ba31ad33318457a33442860.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 114, + 255, + 121 + ], + "lines": [ + { + "bbox": [ + 96, + 114, + 255, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 114, + 255, + 121 + ], + "type": "text", + "content": "(a) The ego-vehicle decelerates to wait for the red lights." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 72, + 212, + 114 + ], + "blocks": [ + { + "bbox": [ + 137, + 72, + 212, + 114 + ], + "lines": [ + { + "bbox": [ + 137, + 72, + 212, + 114 + ], + "spans": [ + { + "bbox": [ + 137, + 72, + 212, + 114 + ], + "type": "image", + "image_path": "a242d6bd677ef62f071aa003b89b063187116f455eb4ce94d24c104620834821.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 290, + 114 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 290, + 114 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 290, + 114 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 290, + 114 + ], + "type": "image", + "image_path": "dca1783964db38094cc306b06a4261880974014f040987404bef33518bfbac5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 122, + 134, + 165 + ], + "blocks": [ + { + "bbox": [ + 59, + 122, + 134, + 165 + ], + "lines": [ + { + "bbox": [ + 59, + 122, + 134, + 165 + ], + "spans": [ + { + "bbox": [ + 59, + 122, + 134, + 165 + ], + "type": "image", + "image_path": "b394c33a6a30dddd779d79a8784f8076677617c2f3ae8eda6cfbf391c74dedd6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 165, + 255, + 171 + ], + "lines": [ + { + "bbox": [ + 96, + 165, + 255, + 171 + ], + "spans": [ + { + "bbox": [ + 96, + 165, + 255, + 171 + ], + "type": "text", + "content": "(b) The ego-vehicle conducts a U-turn at the intersection." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 122, + 212, + 164 + ], + "blocks": [ + { + "bbox": [ + 137, + 122, + 212, + 164 + ], + "lines": [ + { + "bbox": [ + 137, + 122, + 212, + 164 + ], + "spans": [ + { + "bbox": [ + 137, + 122, + 212, + 164 + ], + "type": "image", + "image_path": "9a4b374df4ad16769abdb80c57e1b29e54b5901114937703b0a8d855d8b3c3a0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 216, + 122, + 290, + 165 + ], + "blocks": [ + { + "bbox": [ + 216, + 122, + 290, + 165 + ], + "lines": [ + { + "bbox": [ + 216, + 122, + 290, + 165 + ], + "spans": [ + { + "bbox": [ + 216, + 122, + 290, + 165 + ], + "type": "image", + "image_path": "82111de3a2686af243440848c0e561e66a92d6f806e1df59e29d41c085dbb96b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 173, + 134, + 216 + ], + "blocks": [ + { + "bbox": [ + 59, + 173, + 134, + 216 + ], + "lines": [ + { + "bbox": [ + 59, + 173, + 134, + 216 + ], + "spans": [ + { + "bbox": [ + 59, + 173, + 134, + 216 + ], + "type": "image", + "image_path": "48a82813b9bf5ff020bd71902c308c3fd7f7e1d2298d0f0598d156d79cb8183e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 215, + 266, + 223 + ], + "lines": [ + { + "bbox": [ + 85, + 215, + 266, + 223 + ], + "spans": [ + { + "bbox": [ + 85, + 215, + 266, + 223 + ], + "type": "text", + "content": "(c) The ego-vehicle keeps the lane under bad lighting condition." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 174, + 212, + 216 + ], + "blocks": [ + { + "bbox": [ + 137, + 174, + 212, + 216 + ], + "lines": [ + { + "bbox": [ + 137, + 174, + 212, + 216 + ], + "spans": [ + { + "bbox": [ + 137, + 174, + 212, + 216 + ], + "type": "image", + "image_path": "b38df7a6646238cd0619c3f1243d02fcfba905e8207dd20506fb2a94e170d6a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 216, + 174, + 290, + 217 + ], + "blocks": [ + { + "bbox": [ + 216, + 174, + 290, + 217 + ], + "lines": [ + { + "bbox": [ + 216, + 174, + 290, + 217 + ], + "spans": [ + { + "bbox": [ + 216, + 174, + 290, + 217 + ], + "type": "image", + "image_path": "b62687f28a1e32765934a27c0f71aa93d9728c5f39a7c1ee383978c374e334d2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 250, + 171, + 355 + ], + "blocks": [ + { + "bbox": [ + 54, + 224, + 294, + 245 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 294, + 245 + ], + "type": "text", + "content": "Figure 6. Qualitative results of motion planning. We show the front left, front, front right cameras for each case." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 250, + 171, + 355 + ], + "lines": [ + { + "bbox": [ + 59, + 250, + 171, + 355 + ], + "spans": [ + { + "bbox": [ + 59, + 250, + 171, + 355 + ], + "type": "image", + "image_path": "761714070865e371bac2e28b98f29cb4dff11e4cab9163eda65c17ad9a56b946.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 172, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 172, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 172, + 380 + ], + "type": "text", + "content": "Figure 7. Meta-decision prediction accuracy for each behavior." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 358, + 293, + 380 + ], + "lines": [ + { + "bbox": [ + 176, + 358, + 293, + 380 + ], + "spans": [ + { + "bbox": [ + 176, + 358, + 293, + 380 + ], + "type": "text", + "content": "Figure 8. Distribution of sparse volumes along x- and y-axes." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 178, + 250, + 291, + 350 + ], + "blocks": [ + { + "bbox": [ + 178, + 250, + 291, + 350 + ], + "lines": [ + { + "bbox": [ + 178, + 250, + 291, + 350 + ], + "spans": [ + { + "bbox": [ + 178, + 250, + 291, + 350 + ], + "type": "image", + "image_path": "c2c00654c44c01c12d3ddc039617e87f147d09d33d691adbef46ea6d5e4173f6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 388, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 544 + ], + "type": "text", + "content": "vanilla PaLI model. Consistent with [32, 62], the importance of ego-states is acknowledged. Meanwhile, different from their observations on nuScenes benchmark, camera inputs also make great differences on WOMD-Planning-ADE. We assume that WOMD-Planning-ADE covers more diverse driving scenarios including many severe changes of speed and directions, which attaches importance to the sensor data. This also showcases the advantages of WOMD-Planning-ADE for comprehensive evaluation. Tab. 5 also shows that random initialized model cannot converge without MLLM pretraining. Despite different domains, S4-Driver can profit from the strong reasoning ability from large-scale MLLM pretraining on general tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 700 + ], + "type": "text", + "content": "MLLM capability. In addition to the PaLI3-5B [10] used in other parts, we also tailor PaLI2-3B [9] for motion planning with the same techniques. PaLI2-3B is weaker in parameter numbers (3B vs. 5B), architecture, pretraining manner, and input resolution (224 vs. 448). As in Tab. 4, S4-Driver based on PaLI2-3B demonstrates much inferior performance compared to that based on PaLI3-5B. We conduct experiments on two different training data scales on WOMD-Planning-ADE, i.e. 20k (nuScenes scale) vs. 400k (full WOMD-Planning-ADE). The gap is particularly large given enough training data. It also justifies the necessity to conduct experiments on large-scale dataset, which can fully unleash the potential of powerful MLLMs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Sparse volume resolution. Tab. 6 shows the results with" + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 107 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 107 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 107 + ], + "type": "table", + "html": "
MLLMWOMD-Planning-ADE Subset (20k)Full WOMD-Planning-ADE (400k)
ADE@5sbADE@5sADE@5sbADE@5s
PaLI2-3B [9]1.2411.8461.0351.522
PaLI3-5B [10]1.0751.4950.6930.928
", + "image_path": "a77ea038f66a33afb74740f795c69a487e6bbb33f67d9115e8fd1f4430a183af.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 326, + 151, + 541, + 233 + ], + "blocks": [ + { + "bbox": [ + 313, + 108, + 555, + 140 + ], + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 140 + ], + "type": "text", + "content": "Table 4. Ablation studies on planning performance scalability of our methods w.r.t. model and data sizes. We experiment with two MLLMs on two different training data sizes." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 151, + 541, + 233 + ], + "lines": [ + { + "bbox": [ + 326, + 151, + 541, + 233 + ], + "spans": [ + { + "bbox": [ + 326, + 151, + 541, + 233 + ], + "type": "table", + "html": "
MethodADE@5sbADE@5s
vanilla PaLI0.7981.069
w/o historical states2.0092.250
w/o camera images1.0811.448
w/o MLLM pretrainingfail to converge
", + "image_path": "b0161bd8d2aec398decb879da728f54d9c464a7a9c1f1668f8e3f1922dc8fd78.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 339, + 255, + 527, + 315 + ], + "blocks": [ + { + "bbox": [ + 354, + 233, + 514, + 243 + ], + "lines": [ + { + "bbox": [ + 354, + 233, + 514, + 243 + ], + "spans": [ + { + "bbox": [ + 354, + 233, + 514, + 243 + ], + "type": "text", + "content": "Table 5. Ablation studies on MLLM inputs." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 255, + 527, + 315 + ], + "lines": [ + { + "bbox": [ + 339, + 255, + 527, + 315 + ], + "spans": [ + { + "bbox": [ + 339, + 255, + 527, + 315 + ], + "type": "table", + "html": "
ResolutionADE@5sbADE@5s
2m × 2m × 2m0.7981.201
1m × 1m × 2m0.7501.005
1m × 1m × 1m0.7651.051
", + "image_path": "cc79f90d70258c4873acae40eab2651bd842535e023f71f7d68bfc793bb80573.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 333, + 316, + 534, + 325 + ], + "lines": [ + { + "bbox": [ + 333, + 316, + 534, + 325 + ], + "spans": [ + { + "bbox": [ + 333, + 316, + 534, + 325 + ], + "type": "text", + "content": "Table 6. Ablation studies on sparse volume resolutions." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "text", + "content": "different resolution for sparse volumes with same number of sparse volumes (" + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "inline_equation", + "content": "M = 6000" + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "text", + "content": "). In line with Fig. 3, low resolution leads to relatively worse performance since it limits the precision of 3D spatial reasoning. Interestingly, higher resolution along " + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "text", + "content": "-axis does not definitely improve the model performance since the motion planning mainly works on the " + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 313, + 336, + 555, + 431 + ], + "type": "text", + "content": "-plane and too low sparsity ratio tends to make the optimization unstable." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 442, + 479, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 479, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 479, + 454 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "This paper introduces S4-Driver, a scalable self-supervised motion planning framework for autonomous driving that utilizes Multimodal Large Language Models (MLLMs). To enhance 3D reasoning in MLLMs, we propose a novel sparse volume representation, which enables improved spatio-temporal reasoning through the aggregation of multi-view and multi-frame image inputs. Additionally, we design behavior-wise metrics for comprehensive evaluation on the large-scale WOMD-Planning-ADE benchmark. S4-Driver achieves state-of-the-art performance on both nuScenes and WOMD-Planning-ADE benchmarks, without requiring any human annotations. This demonstrates the potential of self-supervised learning for end-to-end autonomous driving. Future work will involve applying our methods to other powerful MLLM architectures. Our contributions are orthogonal to previous works, such as those that explore multi-task learning or CoT reasoning [49, 55]. Combining our large-scale self-supervised learning approach with supervised finetuning on targeted small-scale labeled data may further enhance the final performance and interpretability of the system." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1629" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 268 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 323 + ], + "type": "text", + "content": "[4] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 296, + 390 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 391, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 435 + ], + "type": "text", + "content": "[6] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "type": "text", + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 568 + ], + "type": "text", + "content": "[8] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14093-14100. IEEE, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 569, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 295, + 635 + ], + "type": "text", + "content": "[9] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. In The Eleventh International Conference on Learning Representations. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 635, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 690 + ], + "type": "text", + "content": "[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[11] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[12] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[13] Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. In 2018 IEEE international conference on robotics and automation (ICRA), pages 4693-4700. IEEE, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "text", + "content": "[15] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's-eye-view injected multi-modal large models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668-13677, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 444, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 520 + ], + "type": "text", + "content": "[18] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9710-9719, 2021. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 522, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 588 + ], + "type": "text", + "content": "[19] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19358-19369, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "type": "text", + "content": "[20] Adam W Harley, Zhaoyuan Fang, Jie Li, Rares Ambrus, and Katerina Fragkiadaki. Simple-bev: What really matters for multi-sensor bev perception? In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2759-2765. IEEE, 2023. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 646, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 646, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 315, + 646, + 553, + 679 + ], + "type": "text", + "content": "[21] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 5, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[22] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1630" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "in Neural Information Processing Systems, 36:20482-20494, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 196 + ], + "type": "text", + "content": "[24] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "text", + "content": "[25] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[26] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to-end multimodal model for autonomous driving, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 396 + ], + "type": "text", + "content": "[27] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 465 + ], + "type": "text", + "content": "[28] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[29] Jinkyu Kim, Anna Rohrbach, Trevor Darrell, John Canny, and Zeynep Akata. Textual explanations for self-driving vehicles. In Proceedings of the European conference on computer vision (ECCV), pages 563-578, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "type": "text", + "content": "[31] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[32] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864-14873, 2024. 2, 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[33] Bencheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr:" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Structured modeling and learning for online vectorized HD map construction. In The Eleventh International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "text", + "content": "[35] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "text", + "content": "[36] Xianzheng Ma, Yash Bhalgat, Brandon Smart, Shuai Chen, Xinghui Li, Jian Ding, Jindong Gu, Dave Zhenyu Chen, Songyou Peng, Jia-Wang Bian, et al. When llms step into the 3d world: A survey and meta-analysis of 3d tasks via multi-modal large language models. arXiv preprint arXiv:2405.10255, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[37] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 331 + ], + "type": "text", + "content": "[38] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "type": "text", + "content": "[39] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "text", + "content": "[40] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "type": "text", + "content": "[41] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 479, + 506, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 506, + 490 + ], + "type": "text", + "content": "[42] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 491, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 523 + ], + "type": "text", + "content": "[43] Dean A Pomerleau. Alvinn: An autonomous land vehicle in a neural network. Advances in neural information processing systems, 1, 1988. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 525, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 579 + ], + "type": "text", + "content": "[44] C. Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa T. Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6130-6140, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 581, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 645 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[46] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 7, 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1631" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[47] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. *Languagempc: Large language models as decision makers for autonomous driving.* arXiv preprint arXiv:2310.03026, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 193 + ], + "type": "text", + "content": "[48] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "type": "text", + "content": "[49] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In First Vision and Language for Autonomous Driving and Robotics Workshop. 2, 6, 8, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "text", + "content": "[50] Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, et al. UI2: Unifying language learning paradigms. arXiv preprint arXiv:2205.05131, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[51] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "type": "text", + "content": "[52] Thomas Tian, Boyi Li, Xinshuo Weng, Yuxiao Chen, Edward Schmerling, Yue Wang, Boris Ivanovic, and Marco Pavone. Tokenize the world into object-level knowledge to address long-tail events in autonomous driving. In Workshop on Language and Robot Learning: Language as an Interface. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[53] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 492, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 569 + ], + "type": "text", + "content": "[54] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[55] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[56] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[57] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 520 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 1, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[58] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 152, + 476, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 476, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 476, + 163 + ], + "type": "text", + "content": "[59] xAI. Grok-1.5 vision preview, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 229 + ], + "type": "text", + "content": "[60] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 285 + ], + "type": "text", + "content": "[61] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[62] Jiang-Tian Zhai, Ze Feng, Jihao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "text", + "content": "[64] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 442, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 487 + ], + "type": "text", + "content": "[65] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "type": "text", + "content": "[66] Wenzhao Zheng, Ruiqi Song, Xianda Guo, and Long Chen. Genad: Generative end-to-end autonomous driving. arXiv preprint arXiv:2402.11502, 2024. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1632" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_content_list.json b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ab58836a4d4d109a97523016e15c7af24c0c86e --- /dev/null +++ b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_content_list.json @@ -0,0 +1,1506 @@ +[ + { + "type": "text", + "text": "SACB-Net: Spatial-awareness Convolutions for Medical Image Registration", + "text_level": 1, + "bbox": [ + 116, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinxing Cheng $^{1}$ , Tianyang Zhang $^{1}$ , Wenqi Lu $^{2}$ , Qingjie Meng $^{1,5}$ , Alejandro F. Frangi $^{3,4}$ , Jinming Duan $^{1,3,4*}$", + "bbox": [ + 240, + 179, + 756, + 215 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ School of Computer Science, University of Birmingham, UK", + "$^{2}$ Department of Computing and Mathematics, Manchester Metropolitan University, UK", + "3 Division of Informatics, Imaging and Data Sciences, University of Manchester, UK" + ], + "bbox": [ + 151, + 215, + 844, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Centre for Computational Imaging and Modelling in Medicine, University of Manchester, UK \n5 Department of Computing, Imperial College London, UK", + "bbox": [ + 122, + 268, + 874, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xxc142, txz009}@student.bham.ac.uk, W.Lu@mmu.ac.uk, m.qingjie@bham.ac.uk", + "bbox": [ + 174, + 306, + 821, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alejandro.frangi@manchester.ac.uk, jinming.duan@manchester.ac.uk", + "bbox": [ + 210, + 325, + 779, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 375, + 326, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based image registration methods have shown state-of-the-art performance and rapid inference speeds. Despite these advances, many existing approaches fall short in capturing spatially varying information in non-local regions of feature maps due to the reliance on spatially-shared convolution kernels. This limitation leads to suboptimal estimation of deformation fields. In this paper, we propose a 3D Spatial-Awareness Convolution Block (SACB) to enhance the spatial information within feature representations. Our SACB estimates the spatial clusters within feature maps by leveraging feature similarity and subsequently parameterizes the adaptive convolution kernels across diverse regions. This adaptive mechanism generates the convolution kernels (weights and biases) tailored to spatial variations, thereby enabling the network to effectively capture spatially varying information. Building on SACB, we introduce a pyramid flow estimator (named SACB-Net) that integrates SACBs to facilitate multi-scale flow composition, particularly addressing large deformations. Experimental results on the brain IXI and LPBA datasets as well as Abdomen CT datasets demonstrate the effectiveness of SACB and the superiority of SACB-Net over the state-of-the-art learning-based registration methods. The code is available at https://github.com/x-xc/SACB_Net.", + "bbox": [ + 86, + 407, + 483, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 816, + 220, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deformable image registration aims to align the corresponding anatomical structures between a moving and fixed", + "bbox": [ + 89, + 843, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14b37902be4244111f335d78f924e127a49c1699ba8c0882e2d5dcea27f7853e.jpg", + "image_caption": [ + "Vanilla Convolution" + ], + "image_footnote": [], + "bbox": [ + 531, + 393, + 720, + 539 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26fe5989a3ba2b8110d8af6a587b535fbd415434a1709e850f053283625aea62.jpg", + "image_caption": [ + "* Deformation is usually related to tissue shapes *", + "Spatial-awareness Convolution", + "Figure 1. Spatial-awareness Convolution. Since deformation is usually related to tissue shape in medical image registration, voxels/features from different regions should be given varying levels of attention. However, vanilla convolution methods apply shared kernel weights across all regions, leading to a suboptimal estimation of deformation fields. SAC mechanisms, on the other hand, apply distinct kernel attention weights for different spatial clusters." + ], + "image_footnote": [], + "bbox": [ + 722, + 391, + 885, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image pair with the estimated dense non-linear deformation field, which plays a crucial role in various medical imaging tasks [38, 48]. In the past, conventional methods like LDDMM [6], DARTEL [1], SyN [2], Demons [45], and ADMM [42] are time-consuming and computationally intensive, primarily due to the instance-level (pair-wise) iterative optimization required. Moreover, such methods may involve sophisticated hyper-parameter tuning, limiting their applications in large-scale volumetric registration. To overcome these challenges, learning-based image registration approaches have demonstrated significant success in achieving competitive accuracy along with fast inference speed. Learning-based approaches can be divided into supervised and unsupervised categories. Supervised methods", + "bbox": [ + 511, + 689, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5227", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[7, 34, 49] often rely on the high quality of the synthetic ground truth deformation field or the annotated map [35] to predict a deformation field. Therefore, unsupervised image registration is more popular for estimating an optimal deformation field of pairwise images by minimizing the distance between the warped moving image and the fixed image with a regularization term.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the classical unsupervised learning-based framework VoxelMorph [5], some U-Net-based methods focus on the performance improvements by adding stronger constraints over the deformation field such as inverse/cycle consistency [24, 50], gradient inverse consistency [43] and diffeomorphisms [13, 14, 29]. Some existing methods have introduced more advanced neural architectures, such as large-kernel network [20], vision transformers [8, 10, 17, 30, 37, 51]. To enhance the efficiency of the U-Net architecture and accelerate inference speed, several model-driven methods, including B-Spline [32], Fourier-Net [21] and WiNet [12], have been proposed to learn a low-dimensional representation of displacement. The lightweight approach, NCA-Morph [33], utilizes Neural Cellular Automata to improve the efficiency of image registration. Another track of recent works tends to progressively estimate the large and complex deformations using either multiple cascades [15, 31, 52, 53] or pyramid methods [18, 23, 25, 26, 28, 46] compositions of multi-scale flows derived from multi-scale features. The cascaded methods and pyramid coarse-to-fine methods are capable of effectively improving registration accuracy for estimating large deformation.", + "bbox": [ + 91, + 200, + 483, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing learning-based methods typically rely on networks that learn local feature correlations by applying shared convolution across all spatial dimensions within each layer. As illustrated in Fig. 1, deformations are often linked to the morphological characteristics of tissue in medical image registration, where voxels/features from various anatomical regions need to be paid different attention. However, plain convolutional methods utilize shared kernel weights uniformly across all regions, resulting in suboptimal estimation of deformation fields. Specifically, 3D medical images encapsulate complex intensity information relevant to anatomy, which makes it hard to capture spatially varying information in non-local regions of feature maps via spatially-shared convolution kernels. This key challenge is effectively incorporating spatially adaptive information, which can be addressed using region-based adaptive convolution techniques. Inspired by [16], which explores content-adaptive convolution in 2D image pansharpening, we extend this approach to 3D for medical image registration. This adaptation improves spatial awareness by modifying convolutional kernels based on regional content, thereby enhancing registration accuracy.", + "bbox": [ + 89, + 535, + 483, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel 3D pyramid medical image registration network, named SACB-Net, which in", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tegrates spatial-awareness convolution block (SACB) with a similarity matching module to form a multi-scale flow estimator. The proposed SACB employs the unsupervised KMeans to cluster the learned feature maps. Each feature clustering is believed to capture a specific intensity or structural pattern of the feature maps and then subsequently used to learn the spatial-aware adaptive convolution kernels. The adaptive kernel generation in SACB allows the network to effectively learn refined feature representations for more accurate flow estimation. We summarize our main contributions as follows:", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We consider the spatial variance between two input images in the context of image registration, which is overlooked by vanilla convolutional-based approaches. The adaptive convolutions in our SACB-Net consider the changing of spatial regions to enhance registration accuracy. To the best of our knowledge, our method is the first work that employs 3D spatial-awareness convolution in 3D medical registration.", + "- The proposed SACB aims to boost the moving and fixed feature maps using adaptive convolution kernels for different spatial regions, where the spatial regions are divided based on the clustering of features, and the convolution kernels are learned from the centroids of each feature clustering.", + "- We propose an innovative pyramid flow estimator that integrates SACB and similarity matching module to perform multi-scale flow estimation, which can not only estimate local small deformation but also handle large deformation.", + "- Experimental results on atlas-based and inter-subject registration using two brain datasets (IXI and LPBA) demonstrate the effectiveness of SACB-Net. Further evaluation on inter-subject abdominal CT registration shows improved performance in handling large deformations." + ], + "bbox": [ + 513, + 258, + 903, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 638, + 661, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. U-Net-based Registration Method", + "text_level": 1, + "bbox": [ + 513, + 665, + 808, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VoxelMorph [4, 5], a general framework for learning-based image registration, demonstrates the effectiveness of unsupervised learning in pairwise image registration using a simplified U-Net architecture. Building on this foundation, recent methods have enhanced registration performance by incorporating stronger constraints on the deformation field, such as inverse/cycle consistency [24, 50], gradient inverse consistency [43], and diffeomorphism constraints [13, 14, 29]. Another track of methods tried to capture long-range dependencies of feature correlations using either large kernel convolution [20] or vision transformer [8, 10, 17, 30, 37, 51]. TransMorph [10] incorporates Swin transformer blocks in the encoder, improving the network's capacity to capture long-range dependencies, though this", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5228", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "comes with increased computational cost. LKU [20], on the other hand, employs a large kernel convolutional encoder to model long-range relationships with fewer parameters and reduced computational load. NCA-Morph [33] employs Neural Cellular Automata mechanisms, drawing inspiration from biological processes, to improve registration efficiency. While U-Net-based methods have shown comparable performance for small deformations, they struggle to accurately estimate large deformations or register images that are not well pre-aligned.", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Model-driven Registration Method", + "text_level": 1, + "bbox": [ + 89, + 253, + 393, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce redundant convolution operations in the U-Net decoder and enhance training and inference efficiency, model-driven approaches such as B-Spline [32] and Fourier-Net [21] focus on learning a low-dimensional representation of the displacement field. However, displacement fields are not inherently smooth, and as a result, methods that rely solely on low-frequency components [21] or interpolate displacement from regular control points [32] may fail to fully and accurately capture the entire displacement field. To address this limitation, WiNet [12] introduces a wavelet-based decoder to preserve high-frequency information in the deformation field while maintaining network efficiency. Nonetheless, similar to U-Net-based methods, model-driven approaches also face challenges in effectively estimating large and complex deformations.", + "bbox": [ + 89, + 275, + 482, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Cascaded and Pyramid-based Method", + "text_level": 1, + "bbox": [ + 89, + 512, + 419, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works that estimate large deformation can be either network cascade [15, 31, 52, 53] or coarse-to-fine pyramid [18, 23, 25, 26, 28, 46]. LapIRN [31] utilizes a Laplacian pyramid network to capture large deformations by cascading registered flows at three different scales. PRNet++ [23] utilizes a dual-stream pyramid architecture for coarse-to-fine flow estimation, performing sequential warping on multi-scale feature maps and incorporating local 3D correlation layers. Im2grid [25] embedded the coordinate information on extracted multi-scale feature maps and estimated the scale flow by calculating the matching score on the neighborhood feature. ModeT [46] leverages neighbourhood attention mechanisms to estimate multi-head, multiscale sub-deformation flows from pyramid feature maps. It then applies weighting modules to fuse these multi-head flows at each scale, generating the final deformation by sequential composition. To solve the insufficient decomposition problems, RDN [18] compose the multi-scale flows by performing level-wise recursion at each resolution level and stage-wise recursion at feature pyramid level.", + "bbox": [ + 89, + 534, + 482, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Adaptive Convolution", + "text_level": 1, + "bbox": [ + 89, + 848, + 294, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike standard convolution operators that use shared convolution kernels, adaptive convolution operators apply dif", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ferent kernels based on variations in the input, providing enhanced feature extraction capabilities and greater flexibility. Following [19], recent methods have applied adaptive convolution, demonstrating its effectiveness in various 2D computer vision tasks, such as image upsampling [39], classification [11, 54], segmentation [9] and remote sensing pansharpening [16, 22]. However, spatially adaptive convolution has not yet been explored in 3D medical image registration tasks.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 239, + 604, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Network Overview", + "text_level": 1, + "bbox": [ + 511, + 263, + 692, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pair of moving image $I_{m}$ and fixed image $I_{f}$ over the spatial domain $\\Omega \\subset \\mathcal{R}^3$ , the SACB-Net $(I_m, I_f; \\pmb{\\theta})$ network aims to learn the optimal parameters $\\pmb{\\theta}_{*}$ that estimate a deformation field $\\phi_{m \\to f}: \\mathcal{R}^3 \\to \\mathcal{R}^3$ , mapping the corresponding coordinates between the paired images. In unsupervised setting, the loss function contains $\\mathcal{L}_{sim}(\\cdot)$ measuring the similarity between the warped moving image $I_{m} \\circ \\phi$ and $I_{f}$ , and the regularization term. In our experiments, the smoothness term $\\mathcal{L}_{reg}(\\phi) = \\| \\nabla \\phi(\\pmb{\\theta}) \\|_2^2$ is computed using the $L_{2}$ -norm of the gradient of the deformation field. The objective function of our network is", + "bbox": [ + 511, + 286, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\boldsymbol {\\theta}) = \\mathcal {L} _ {s i m} \\left(I _ {m} \\circ \\left(\\phi (\\boldsymbol {\\theta}) + \\operatorname {I d}\\right), I _ {f}\\right) + \\lambda \\mathcal {L} _ {r e g} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 465, + 906, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the learnable network parameters, $\\circ$ represents the warping operator, $\\mathrm{Id}$ is the identity grid, $\\nabla$ refers to the first-order gradient, and $\\lambda$ is the weight of the regularization term. We employ Normalized Cross-Correlation (NCC) to measure the similarity.", + "bbox": [ + 511, + 492, + 905, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, SACB-Net consists of a shared encoder and a pyramid flow estimator. The shared encoder is designed to extract multi-scale features $\\{F_m^i\\}$ and $\\{F_f^i\\}$ for the moving image $I_{m}$ and the fixed image $I_{f}$ , respectively, in a pyramid structure. Here, $i \\in \\{1,2,3,4,5\\}$ represents the scale of the features, with resolutions of $\\left[\\frac{D}{2^{i - 1}},\\frac{H}{2^{i - 1}},\\frac{W}{2^{i - 1}}\\right]$ for the depth, height and width. The shared encoder consists of five 3D convolutional layers and four down-sampling layers. The convolutional layer is composed of 3D convolution operation, instance normalization [44] and LeakyReLU activations [27] with the 0.1 slope. Each convolutional layer is followed by an average pooling operation with a kernel size of 2. The extracted features $\\{F_m^i\\}$ and $\\{F_f^i\\}$ are then input into our pyramid flow estimators, where they are used to perform flow estimation and composition as outlined in Sections 3.2 and 3.3.", + "bbox": [ + 511, + 568, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Spatial-awareness Convolution Block", + "text_level": 1, + "bbox": [ + 511, + 818, + 861, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively extract region-based context information for flow estimation, the proposed 3D SAC incorporates prior knowledge of spatial context through a spatial context estimation module and adaptive kernel generator. First, the", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5229", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd194211dc1b162d63785852eac2971782c5ea01fe74df893c0e57d5eb87a122.jpg", + "image_caption": [ + "Figure 2. Illustration of a 5-level pyramid SACB-Net. SACB-Net includes a shared encoder that extracts multi-scale feature maps $\\{F_m^i\\}$ and $\\{F_f^i\\}$ for the moving image $I_{m}$ and the fixed image $I_{f}$ , as well as pyramid flow estimators at each scale. At the lowest level, the flow estimator learns deformation $(\\varphi_5)$ from the extracted moving and fixed image features $(F_m^5$ and $F_{f}^{5})$ . The following flow estimators take the level-wise features and the output deformation from its preceding level to compose the deformation. Each flow estimator includes a Spatial-Awareness Convolution Block (SACB) to enhance spatially adaptive feature representation, along with a similarity matching module for flow estimation." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 906, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spatial context estimation module identifies spatial regions of interest, producing spatial contextual information. This contextual data is subsequently utilized by the adaptive kernel generator to dynamically produce convolutional kernels tailored to capture non-local, region-specific information.", + "bbox": [ + 88, + 468, + 482, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Context Estimation. After obtaining the multiscale features $\\{F_m^i\\}$ and $\\{F_f^i\\}$ , prior spatial region information is incorporated to refine these features. To achieve this, we employ unsupervised clustering method K-Means to estimate spatial region for feature maps. The reason not to use label information is due to potential inconsistencies in the labeling of $I_{m}$ and $I_{f}$ within the feature space, as well as to address the challenge of limited label availability. Let $\\mathbf{F} \\in \\mathbb{R}^{C_{in} \\times D \\times H \\times W}$ denote the input feature map, where $C_{in}$ represents the number of input channels, and $D$ , $H$ , and $W$ correspond to the depth, height, and width of the spatial dimensions, respectively. As shown in Fig. 3, first, the feature map $\\mathbf{F}$ is unfolded into local patches to capture neighborhood information using sliding windows of size $k$ . A spatial mean is then applied to each patch to reduce its spatial dimensions. For each voxel $V_{(d,h,w)}$ located at coordinates $(d,h,w)$ within $\\mathbf{F}$ , suppose $p(\\cdot)$ is the unfolding operation, the corresponding unfolded patch is defined as:", + "bbox": [ + 89, + 551, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\left(V _ {\\left(d, h, w\\right)}\\right)} = \\left\\{V _ {\\left(d + d ^ {\\prime}, h + h ^ {\\prime}, w + w ^ {\\prime}\\right)} \\mid d ^ {\\prime}, h ^ {\\prime}, w ^ {\\prime} \\in \\left[ - \\frac {k}{2}, \\frac {k}{2} \\right] \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 851, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $p_{(V_{(d,h,w)})}$ represents a patch of voxels centered at $V_{(d,h,w)}$ , with offsets $d', h', w'$ that range from $-\\frac{k}{2}$ to $\\frac{k}{2}$ (assuming $k$ is odd). The unfolded feature denotes as $p(F) \\in \\mathbb{R}^{C_{in} \\times (D \\times H \\times W) \\times k^3}$ . The spatial mean of each local patch $\\overline{p}_{(V_{(d,h,w)})}$ based on the neighborhood voxel is calculated as:", + "bbox": [ + 511, + 468, + 906, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {p} _ {\\left(V _ {(d, h, w)}\\right)} = \\frac {1}{k ^ {3}} \\sum_ {d ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} \\sum_ {h ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} \\sum_ {w ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} p _ {\\left(V _ {(d, h, w)}\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 565, + 906, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, the preprocessed feature $\\mathbf{F}$ is reshaped to dimensions $[D\\times H\\times W,C_{in}]$ to estimate a cluster index matrix $S\\in \\mathbb{Z}^{D\\times H\\times W}$ , which assigns each voxel to a cluster $S_{n} = \\{(d,w,h)|S_{(d,w,h)} = n\\};n\\in \\{1,\\dots ,N\\}$ . Here, $N$ represents the maximum number of clusters used for GPU-based KMeans clustering. Additionally, the centroid voxel of each cluster $S_{n}^{c}\\in \\mathbb{R}^{C_{in}}$ is computed as", + "bbox": [ + 511, + 618, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n} ^ {c} = \\frac {1}{\\left| S _ {n} \\right|} \\sum_ {(d, w, h) \\in S _ {n}} V _ {(d, w, h)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 728, + 906, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|S_{n}|$ denotes the total number of voxels belonging to the $n^{\\mathrm{th}}$ spatial region. The cluster index matrix $S$ and the centroid voxels $S^{c}$ for each cluster are then utilized to generate adaptive kernel weights for convolution.", + "bbox": [ + 511, + 771, + 906, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Kernel Generator. To adaptively generate a spatial convolution kernel that incorporates region-based information, we employ a multi-layer perception to model the", + "bbox": [ + 511, + 835, + 906, + 881 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "The choices of $\\overline{p}$ are investigated in Sec. 6.", + "bbox": [ + 531, + 886, + 764, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5230", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1f954c9bea66e0b0eb0090554b79cbb10551f766c38721260705c96384710ce.jpg", + "image_caption": [ + "Figure 3. Architecture of the 3D spatial-awareness convolution block (example in three clusters). This block aims to refine the input feature $\\mathbf{F}$ to $\\hat{\\mathbf{F}}$ by adaptive convolution learned from spatial feature clustering. SACB consists of three parts: 1) the spatial context estimation module employs KMeans to cluster similar spatial features on the patched unfolding features, in which the features belonging to the same cluster centroid $S_{n}^{c}$ will be indicated with a cluster index map $S_{n}$ . 2) the adaptive kernel generator leverages each cluster centroid to generate cluster-specific spatial weights and bias via two MLPs, the resulting weights and bias will be used to form the final spatial adaptive convolution kernel, which will be imposed on the cluster-indexed unfolding features, again highlighting the spatial-awareness. 3) the original features and spatial-aware features are residually connected to form the final refined $\\hat{\\mathbf{F}}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mapping $\\mathcal{F}_w$ , which produces an adaptive convolution kernel $\\mathbf{W}_n \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times k^3}$ for the $n^{\\text{th}}$ spatial region based on $S_n^c$ . To simplify the process, we first initialize a global learnable convolution kernel $\\mathbf{W} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times k^3}$ and then use the mapping function $\\mathcal{F}_w$ to generate a spatial weight that can be applied to the global kernel $\\mathbf{W}$ . Thus, $\\mathbf{W}_n$ is obtained as", + "bbox": [ + 88, + 472, + 480, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {n} = \\mathcal {F} _ {w} \\left(S _ {n} ^ {c}\\right) \\odot \\mathbf {W}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 590, + 480, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise dot product. Then, a mapping function $\\mathcal{F}_b$ using a multi-layer perception is applied to compute the convolution bias via obtained centroid $S_{n}^{c}$ . Consequently, the region-based spatially adaptive convolution is represented as:", + "bbox": [ + 89, + 616, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS A C \\left(\\mathbf {F} _ {(d, w, h)}\\right) = p _ {\\left(V _ {(d, h, w)}\\right)} \\otimes \\mathbf {W} _ {n} + \\mathcal {F} _ {b} \\left(S _ {n} ^ {c}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 700, + 480, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(\\cdot)$ denotes the unfolding operation outlined in the spatial context estimation section. Subsequently, the process of applying the SAC block to produce refined features is as follows", + "bbox": [ + 89, + 726, + 483, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} = S A C B (\\mathbf {F}) = \\mathbf {F} + \\sigma (S A C (\\mathbf {F})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 794, + 482, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ denotes activation function.", + "bbox": [ + 89, + 821, + 349, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Pyramid Flow Estimator", + "text_level": 1, + "bbox": [ + 89, + 844, + 316, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\hat{F_m}^i$ and $\\hat{F_f}^i$ represent the multi-scale features refined by the proposed SACB at the $i^{th}$ level, the sub-deformation", + "bbox": [ + 89, + 864, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "flows at each scale are then estimated based on the matching score using the refined features. Specifically, we calculate the similarity matching score between each voxel in the fixed image features and corresponding neighbouring voxels in the moving image features. Similar to [25, 40], this operation can be efficiently implemented as a matrix operation using the dot product. For instance, at the coarsest level $i = 5$ , the matching score can be computed as", + "bbox": [ + 511, + 472, + 906, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s i m} \\left(\\hat {F _ {f}} ^ {5}, \\hat {F _ {m}} ^ {5}\\right) = \\operatorname {S o f t m a x} \\left(\\hat {F _ {f}} ^ {5} p \\left(\\hat {F _ {m}} ^ {5}\\right) ^ {T}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 601, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(\\cdot)$ represents the unfolding operation as described in Section 3.2. For each voxel $V$ at $\\hat{F_m}^5$ , the similarity matching score can be calculated as", + "bbox": [ + 511, + 632, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s i m} \\left(\\hat {F} _ {f} ^ {5} (V), \\hat {F} _ {m} ^ {5} (V)\\right) = \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 689, + 903, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S o f t m a x} \\left(\\left\\{\\left\\langle \\hat {F _ {f}} ^ {5} (V), \\hat {F _ {m}} ^ {5} (V _ {c}) \\right\\rangle \\right\\} _ {V _ {c} \\in \\mathcal {N} (V)}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 714, + 862, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\langle \\cdot, \\cdot \\rangle$ denotes the inner product, and $\\mathcal{N}(V)$ represents the set of voxels within an $k \\times k \\times k$ neighborhood centered around voxel $V$ , $k$ represents the window size of the search region of $\\hat{F}_m^5$ . We set $k = 3$ in our experiments. Then, the sub-deformation flow at the $5^{\\text{th}}$ scale can be obtained as", + "bbox": [ + 511, + 755, + 906, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi^ {5} = M _ {s i m} \\left(\\hat {F _ {f}} ^ {5}, \\hat {F _ {m}} ^ {5}\\right) G \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 837, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$G \\in \\mathbb{R}^{(k \\times k \\times k) \\times 3}$ is a 3D grid that encodes the relative positions with respect to the centroid voxel location.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5231", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the $i = 4$ level, we first up scaling the flow $\\varphi^5$ to form $\\hat{\\varphi}^5$ , then $\\hat{F_m}^4$ is warped with $\\hat{\\varphi}^5$ . Sequentially, the $M_{sim}$ need to be calculated between $\\hat{F_f}^4$ and $\\hat{F_m}^4 \\circ \\hat{\\varphi}^5$ for obtaining $\\Delta \\varphi^4$ . Afterward, the overall process can be formulated as:", + "bbox": [ + 89, + 90, + 482, + 159 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\hat {\\varphi} ^ {i} = \\mathrm {u p} _ {2 \\times} \\left(\\varphi^ {i}\\right), \\\\ \\hat {F} _ {f} ^ {i - 1} = S A C B \\left(F _ {f} ^ {i - 1}\\right), \\\\ \\hat {F} _ {m} ^ {i - 1} = S A C B \\left(F _ {m} ^ {i - 1} \\circ \\hat {\\varphi} ^ {i}\\right), \\quad i \\in \\{5, 4, 3, 2 \\}. \\\\ \\Delta \\varphi^ {i - 1} = M _ {s i m} \\left(\\hat {F} _ {f} ^ {i - 1}, \\hat {F} _ {m} ^ {i - 1}\\right) G, \\\\ \\varphi^ {i - 1} = \\hat {\\varphi} ^ {i} \\circ \\Delta \\varphi^ {i - 1}, \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 169, + 480, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The $\\mathrm{up}_{2\\times}(\\cdot)$ is flow rescaling operation with $2\\times$ resolution.", + "bbox": [ + 89, + 280, + 482, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that the matching score calculation was not applied for flow estimation at the $1^{st}$ scale, as the search region with $k = 3$ can only determine a maximum displacement of 1 voxel using full-resolution features. However, the calculations at the $1^{st}$ scale significantly increase memory and computational costs, yielding limited benefit. Therefore, we used a simple approach with two convolutional layers for flow estimation at the final scale.", + "bbox": [ + 88, + 296, + 483, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 429, + 222, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our experiments use two publicly available 3D brain MRI datasets $\\mathrm{IXI}^2$ and LPBA [36], along with one public abdominal CT dataset [47]. The preprocessed IXI dataset [10] includes 576 MRI scans $(160\\times 192\\times 224)$ from healthy subjects. We followed the same protocol as [10], partitioning the data into 403, 58, and 115 scans for training, validation, and testing sets, respectively. Atlas-based brain registration was conducted using a generated atlas from [24]. The LPBA dataset consists of 40 T1-weighted MRI volumes. Each MRI scan has been rigidly pre-aligned to the MNI305 template and manually annotated with 54 regions of interest (ROIs). We center-crop all volumes to dimensions of $160\\times 192\\times 160$ with voxel size $1\\times 1\\times 1mm^3$ . LPBA dataset was used to perform inter-subject registration with divided three subsets: 25 volumes $(25\\times 24$ pairs) for training, 5 $(5\\times 4$ pairs) for validation, and 10 $(10\\times 9$ pairs) for testing. The Abdomen CT dataset consists of 30 abdominal scans with 13 manually labelled anatomical structures. We clamp the Hounsfield Units (HU) to $[-1000, 1000]$ and normalize each scan to $[0, 1]$ . Each scan is center-cropped from $192\\times 160\\times 256$ to $192\\times 160\\times 224$ with voxel size $2\\times 2\\times 2mm^3$ . The dataset is divided into 19 training volumes $(19\\times 18$ pairs), 3 validation volumes $(3\\times 2$ pairs) and 8 testing volumes $(8\\times 7$ pairs) for inter-subject registration.", + "bbox": [ + 91, + 454, + 483, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our method is implemented using PyTorch. Both the training and testing phases are deployed on an A100 GPU with 40GB VRAM. All models", + "bbox": [ + 89, + 820, + 483, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are optimized using Adam followed with the same learning rate $10^{-4}$ and batch size 1. A fast GPU K-Means implementation is adopted.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To qualify the registration performance, we use the Dice similarity coefficient (Dice) to evaluate the region of interest (ROI) segmentation overlap, the average symmetric surface distance (ASSD) [41] and the 95 percentile of the Hausdorff distance (HD95) to measure the registration accuracy of ROI structures. For evaluating the diffeomorphism of deformation fields, we use the percentage of negative values of the Jacobian determinant $(|J|_{<0\\%})$ to calculate the percentage of folding voxels.", + "bbox": [ + 511, + 138, + 906, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparative Methods. We compare our method against state-of-the-art approaches, including the conventional SyN method [2], U-Net-based models such as VoxelMorph (VM) [5], LKU [20], and TransMorph [10], model-driven methods like B-Spline-Diff [32] and Fourier-Net [21], the lightweight NCA-Morph [33], the cascaded method LapIRN [31], and four pyramid-based methods: PR++ [23], Im2Grid [25], RDN [18] and ModeT [46]. All models were trained using their officially released code, with optimal hyperparameters tuned on the validation sets.", + "bbox": [ + 511, + 279, + 908, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 444, + 601, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our SACB-Net against comparative methods on three tasks: atlas-based brain registration using the IXI dataset, inter-subject brain registration using the LPBA dataset, and large deformation estimation for inter-subject abdominal CT registration.", + "bbox": [ + 511, + 469, + 905, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Atlas-based Brain Registration", + "text_level": 1, + "bbox": [ + 511, + 554, + 787, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the quantitative results of atlas-based brain registration for 30 anatomical structures evaluated on the IXI dataset. It can be seen that SACB-Net outperforms all comparable methods in terms of Dice score while achieving plausible deformations and third-best results on HD95 and ASSD. Compared to the best-performing U-Net-based method (LKU), our SACB-Net achieves a $0.4\\%$ improvement in Dice score while using 0.98M fewer parameters. In terms of model-driven methods, SACB-Net surpasses Fourier-Net by a margin of $0.6\\%$ in the Dice score and uses 3.09M fewer parameters. For cascade and pyramid-based methods, those methods have a similar number of parameters apart from RDN. Our SACB-Net outperforms the second-best cascade method LapIRN with $0.6\\%$ Dice, 0.038mm HD95, and 0.019mm ASSD, respectively.", + "bbox": [ + 511, + 575, + 906, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Inter-subject Brain Registration", + "text_level": 1, + "bbox": [ + 511, + 811, + 797, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The inter-subject registration of the LPBA dataset tends to have a slightly larger deformation than the atlas-based registration on the IXI dataset. As shown in Table 1, most UNet", + "bbox": [ + 511, + 833, + 905, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://brain-development.org/ixi-dataset/", + "bbox": [ + 106, + 875, + 447, + 886 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://learn2reg.grand-challenge.org/", + "bbox": [ + 106, + 886, + 416, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4https://github.com/jokofa/torch_KMeans", + "bbox": [ + 529, + 886, + 838, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5232", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d35b96ac74ffe40a9c256db201fd6230b703030ca125937eb5a0db93bb1f620.jpg", + "table_caption": [ + "Table 1. Registration performance comparison on the IXI and LPBA datasets." + ], + "table_footnote": [], + "table_body": "
MethodIXI (30 ROIs)LPBA (54 ROIs)#Param
Dice↑HD95↓ASSD↓|J|<0%↓Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.386±0.1956.479±0.6662.445±0.280-0.525±0.0478.039±0.8612.586±0.350--
SyN [3]0.645±0.1526.394±1.0481.551±0.286<0.00010.707±0.0166.254±0.4441.479±0.131<0.0001-
VM-1 [5]0.729±0.1293.798±0.7570.937±0.1621.590±0.3390.664±0.0256.873±0.6541.717±0.2000.649±0.2610.27M
VM-2 [5]0.732±0.1233.723±0.6800.926±0.1581.522±0.3360.669±0.0256.847±0.6591.698±0.2000.591±0.2420.30M
NCA-Morph [33]0.753±0.1363.109±0.5250.796±0.1210.506±0.1900.679±0.0236.666±0.6341.631±0.1880.264±0.1210.37M
TransMorph [10]0.754±0.1243.543±0.7210.862±0.1681.579±0.3280.695±0.0226.564±0.6191.559±0.1820.474±0.17646.77M
LKU [20]0.765±0.1292.967±0.4940.757±0.1140.109±0.0540.706±0.0326.452±0.9511.603±0.2500.594±0.2032.09M
B-Spline-Diff [32]0.742±0.1283.256±0.5380.832±0.117<0.00010.665±0.0236.792±0.6221.713±0.1860.0±0.00.27M
Fourier-Net [21]0.763±0.1292.857±0.4560.748±0.1140.024±0.0190.672±0.0226.716±0.6011.666±0.1800.216±0.1044.20M
LapIRN [31]0.763±0.1333.166±0.6080.779±0.1260.312±0.1060.716±0.0166.116±0.4541.426±0.1330.024±0.0091.20M
PRNet++ [23]0.755±0.1303.593±0.7480.857±0.1621.052±0.3020.701±0.0216.492±0.5971.520±0.1770.072±0.0271.24M
ModeT [46]0.758±0.1253.496±0.7320.828±0.1510.114±0.0570.721±0.0135.969±0.4161.375±0.1100.010±0.0041.03M
Im2Grid [25]0.761±0.1273.316±0.6680.799±0.128<0.00020.713±0.0146.062±0.4281.419±0.1180.007±0.0030.89M
RDN [18]0.759±0.1233.476±0.8020.823±0.161<0.00010.713±0.0176.208±0.4971.436±0.142<0.000228.65M
Ours0.769±0.1273.128±0.6310.760±0.1250.083±0.0450.731±0.0125.862±0.4361.326±0.1140.018±0.0061.11M
", + "bbox": [ + 91, + 114, + 903, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based and model-driven methods do not perform as well as cascade and pyramid-based methods. Only LKU surpasses the pyramid-based method PR++, achieving improvements of $0.5\\%$ in Dice, $0.04 \\mathrm{~mm}$ in HD95 and $0.083 \\mathrm{~mm}$ in ASSD, respectively. Our SACB outperforms all comparable methods in Dice, HD95 and ASSD with plausible deformations. Compared to the second-best method, ModeT, SACB achieves improvements of $1\\%$ in Dice, $0.107 \\mathrm{~mm}$ in HD95, and $0.049 \\mathrm{~mm}$ in ASSD, respectively. The visualization in Fig. 4 shows that our SACB-Net captures more detailed correspondences in the registration process, as highlighted in the red box.", + "bbox": [ + 88, + 373, + 480, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Inter-subject Abdomen Registration", + "text_level": 1, + "bbox": [ + 89, + 571, + 405, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The primary challenge in abdominal CT registration tasks is caused by the substantial differences in the spatial distribution of organs across scans, making it difficult to accurately estimate large deformations during the registration process. Table 2 demonstrates the numerical results of the inter-subject abdomen registration task. Our results show that our method outperforms all compared approaches in terms of Dice, HD95, and ASSD. Notably, U-Net-based and model-driven methods perform inferior to pyramid and cascaded approaches for estimating large deformations. Even the best-performing method in such approaches, TransMorph, still exhibits a significant gap compared to our SACB-Net with $14.4\\%$ difference in Dice score, $5.934\\mathrm{mm}$ in HD95 and $3.153\\mathrm{mm}$ in ASSD, respectively. Compared to the second-best pyramid method, ModeT, SACB-Net achieves improvements of $3.8\\%$ in Dice, $2.098\\mathrm{mm}$ in HD95, and $0.781\\mathrm{mm}$ in ASSD, respectively.", + "bbox": [ + 91, + 595, + 482, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The visualization of the wrapped moving images and labels, as shown in Fig. 4, illustrates that most parts of the kidneys are registered with our method. However, Trans", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/20969baea225e298cf3a33b1ae4ebb8ceeab10844c5cf0310d942a444fd51829.jpg", + "table_caption": [ + "Table 2. Registration performance comparison on Abdomen CT." + ], + "table_footnote": [], + "table_body": "
MethodAbdomen CT (13 ROIs)
Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.305±0.04727.790±4.86810.869±2.075-
VM-2 [5]0.406±0.05825.226±4.9428.890±2.0663.882±1.512
NCA-Morph [33]0.406±0.06125.959±4.9829.083±2.2235.197±1.166
TransMorph [10]0.444±0.06524.187±4.9558.156±2.0656.045±1.750
LKU [20]0.423±0.05824.252±4.6048.395±1.9464.316±1.778
B-Spline-Diff [32]0.375±0.05325.338±4.8669.217±1.9600.0±0.0
Fourier-Net [21]0.417±0.04823.084±4.2018.000±1.5322.167±2.391
LapIRN [31]0.522±0.05720.728±4.3186.274±1.3793.203±0.848
PRNet++ [23]0.478±0.06023.791±4.6917.524±1.7712.252±1.192
ModeT [46]0.550±0.05420.351±4.2315.784±1.2362.239±0.906
Im2Grid [25]0.530±0.04720.763±3.7866.028±1.1462.421±0.653
RDN [18]0.527±0.05521.782±4.7226.461±1.6131.529±0.989
Ours0.588±0.04918.253±3.6105.003±1.0113.930±1.416
", + "bbox": [ + 513, + 396, + 903, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Morph, $\\mathrm{PR} + +$ , and RDN failed to register the kidneys, leaving a large portion missing on the warped mask.", + "bbox": [ + 511, + 619, + 903, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 681, + 663, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we note that $\\overline{p}(\\cdot)$ in Eq. (3) can also be defined on the channel dimension or both the spatial+channel (Mix) dimension. As shown in Table 4, using the spatial mean of unfolded patches achieves slightly better performance compared to the channel mean and the Mix setting.", + "bbox": [ + 511, + 713, + 903, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then determine the optimal number of scales at which SACB should be applied and the ideal number of clusters to achieve accurate registration performance. As shown in Table 3, the optimal configuration is applying SACB from scale $5$ to scale $2$ . Additionally, increasing the $N$ from 5 to 7 boosts the performance while a too-large value slightly decreases performance (e.g. 11).", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5233", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6657bc9b803d28710cf20fc911dc5db93f9c4665709293a01646f168d4228e48.jpg", + "image_caption": [ + "Figure 4. Visual comparisons on the Brain LPBA dataset (top two rows) and the Abdomen CT dataset (bottom two rows). Columns 3-10 display warped moving images (row 1 and 3), displacement fields as RGB images (row 2) and warped segmentation masks (row 4)." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 906, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/41915a550506467991623e5375ad9ad642d9857c1142ca00a2fad68be2ce84c2.jpg", + "table_caption": [ + "Table 3. Ablation studies on the use of SACB at different scale levels in SACB-Net and varying numbers of spatial clusters $(N)$ , along with the corresponding average training time (in seconds), including data loading time, and GPU training memory usage (in MiB). The SACB configurations are based on the best results shown in Table 4." + ], + "table_footnote": [], + "table_body": "
NWith SACBIXILPBAAbdomen CT
Scale5Scale4Scale3Scale2Dice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTraining
-XXXX0.76430.059158380.7620.71410.020115740.5450.53742.907158380.764
5✓XXX0.76680.072158800.7990.72170.087116640.5790.54443.243158800.797
5✓✓XX0.76790.076161040.8790.72410.104117700.6310.54982.619161040.864
5✓✓✓X0.76710.060169600.9450.72660.076124180.7140.56852.781169600.950
5✓✓✓✓0.76830.086216381.2550.72940.017157480.9180.58493.548216381.234
7✓✓✓✓0.76910.083216381.3760.73090.018157781.0840.58813.930216621.360
9✓✓✓✓0.76850.069216801.5500.73040.016157921.1100.58754.405216801.562
11✓✓✓✓0.76840.084217321.6790.73000.015158061.1930.58754.419217181.728
", + "bbox": [ + 93, + 481, + 903, + 612 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a4c239f6de6b11c9bbea850c87e003bf9d817b6320feb21639d813753c380db1.jpg", + "table_caption": [ + "Table 4. Ablations of the SACB configurations. After padding and unfolding on the feature map $\\mathbf{F}$ , the cluster maps can be calculated based on spatial mean on the local unfolded patch, channel mean and Mix (spatial+channel). In this study, we set $N = 5$ and apply SACBs from scale $_5$ to scale $_2$ ." + ], + "table_footnote": [], + "table_body": "
SACBLPBA
SpatialChannelMixDice ↑|J|<0% ↓
✓✗✗0.72940.017
✗✓✗0.72810.020
✗✗✓0.72850.021
", + "bbox": [ + 140, + 715, + 431, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 828, + 209, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced SACB-Net, which incorporates a 3D Spatial Awareness Convolution Block (SACB) that adaptively generates spatial convolution kernels based on the feature", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clusters, thereby improving the network's ability to capture spatially varying information. Additionally, our proposed flow estimators combine SACB with a similarity-matching mechanism to estimate displacements in a pyramid coarse-to-fine manner. Results on two brain datasets and one abdomen dataset demonstrate that SACB-Net effectively handles both small local and large deformations. Furthermore, the proposed flow estimator is a plug-in module that can be applied in other networks for better feature representation learning.", + "bbox": [ + 511, + 637, + 906, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The research was conducted using the Baskerville Tier 2 HPC service, which was funded by the EPSRC and UKRI through the World Class Labs scheme (EP/T022221/1) and the Digital Research Infrastructure programme (EP/W032244/1). Baskerville is operated by Advanced Research Computing at the University of Birmingham.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5234", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] John Ashburner. A fast diffeomorphic image registration algorithm. Neuroimage, 38(1):95-113, 2007. 1", + "[2] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal, 12(1):26-41, 2008. 1, 6", + "[3] Brian B Avants, Nicholas J Tustison, Gang Song, Philip A Cook, Arno Klein, and James C Gee. A reproducible evaluation of ants similarity metric performance in brain image registration. Neuroimage, 54(3):2033-2044, 2011. 7", + "[4] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9252-9260, 2018. 2", + "[5] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 2, 6, 7", + "[6] M Faisal Beg, Michael I Miller, Alain Trouve, and Laurent Younes. Computing large deformation metric mappings via geodesic flows of diffeomorphisms. International journal of computer vision, 61:139-157, 2005. 1", + "[7] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Qian Wang, Pew-Thian Yap, and Dinggang Shen. Deformable image registration using a cue-aware deep regression network. IEEE Transactions on Biomedical Engineering, 65(9):1900-1911, 2018. 2", + "[8] Junyu Chen, Yufan He, Eric C Frey, Ye Li, and Yong Du. Vit-v-net: Vision transformer for unsupervised volumetric medical image registration. arXiv preprint arXiv:2104.06468, 2021. 2", + "[9] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8064-8073, 2021. 3", + "[10] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 2, 6, 7", + "[11] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11030-11039, 2020. 3", + "[12] Xinxing Cheng, Xi Jia, Wenqi Lu, Qiufu Li, Linlin Shen, Alexander Krull, and Jinming Duan. Winet: Wavelet-based incremental learning for efficient medical image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 761-771. Springer, 2024. 2, 3", + "[13] Adrian Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces. Medical image analysis, 57:226-236, 2019. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Adrian V Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning for fast probabilistic diffeomorphic registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 729-738. Springer, 2018. 2", + "[15] Bob D De Vos, Floris F Berendsen, Max A Viergever, Hessam Sokooti, Marius Staring, and Ivana Isgum. A deep learning framework for unsupervised affine and deformable image registration. Medical image analysis, 52:128-143, 2019. 2, 3", + "[16] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 2, 3", + "[17] Morteza Ghahremani, Mohammad Khateri, Bailiang Jian, Benedikt Wiestler, Ehsan Adeli, and Christian Wachinger. H-vit: A hierarchical vision transformer for deformable image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11513-11523, 2024. 2", + "[18] Bo Hu, Shenglong Zhou, Zhiwei Xiong, and Feng Wu. Recursive decomposition network for deformable image registration. IEEE Journal of Biomedical and Health Informatics, 26(10):5130-5141, 2022. 2, 3, 6, 7", + "[19] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 3", + "[20] Xi Jia, Joseph Bartlett, Tianyang Zhang, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. U-net vs transformer: Is u-net outdated in medical image registration? In International Workshop on Machine Learning in Medical Imaging, pages 151-160. Springer, 2022. 2, 3, 6, 7", + "[21] Xi Jia, Joseph Bartlett, Wei Chen, Siyang Song, Tianyang Zhang, Xinxing Cheng, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. Fourier-net: Fast image registration with band-limited deformation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1015-1023, 2023. 2, 3, 6, 7", + "[22] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI conference on artificial intelligence, pages 1113-1121, 2022. 3", + "[23] Miao Kang, Xiaojun Hu, Weilin Huang, Matthew R Scott, and Mauricio Reyes. Dual-stream pyramid registration network. Medical image analysis, 78:102379, 2022. 2, 3, 6, 7", + "[24] Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. Cycle morph: cycle consistent unsupervised deformable image registration. Medical image analysis, 71:102036, 2021. 2, 6", + "[25] Yihao Liu, Lianrui Zuo, Shuo Han, Yuan Xue, Jerry L Prince, and Aaron Carass. Coordinate translator for learning deformable medical image registration. In International workshop on multiscale multimodal medical imaging, pages 98-109. Springer, 2022. 2, 3, 5, 6, 7", + "[26] Tai Ma, Suwei Zhang, Jiafeng Li, and Ying Wen. Iirp-net: Iterative inference residual pyramid network for enhanced" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5235", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11555, 2024. 2, 3", + "[27] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 3", + "[28] Mingyuan Meng, Dagan Feng, Lei Bi, and Jinman Kim. Correlation-aware coarse-to-fine mlp's for deformable medical image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9645-9654, 2024. 2, 3", + "[29] Tony CW Mok and Albert Chung. Fast symmetric diffeomorphic image registration with convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4644-4653, 2020. 2", + "[30] Tony CW Mok and Albert Chung. Affine medical image registration with coarse-to-fine vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20835-20844, 2022. 2", + "[31] Tony CW Mok and Albert CS Chung. Large deformation diffeomorphic image registration with laplacian pyramid networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 211-221. Springer, 2020. 2, 3, 6, 7", + "[32] Huaqi Qiu, Chen Qin, Andreas Schuh, Kerstin Hammernik, and Daniel Rueckert. Learning diffeomorphic and modality-invariant registration using b-splines. In Medical Imaging with Deep Learning, 2021. 2, 3, 6, 7", + "[33] Amin Ranem, John Kalkhof, and Anirban Mukhopadhyay. Nca-morph: Medical image registration with neural cellular automata. arXiv preprint arXiv:2410.22265, 2024. 2, 3, 6, 7", + "[34] Marc-Michel Rohe, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: learning deformable image registration using shape matching. In Medical Image Computing and Computer Assisted Intervention- MIC-CAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part I 20, pages 266-274. Springer, 2017. 2", + "[35] Yudi Sang, Xianglei Xing, Yingnian Wu, and Dan Ruan. Imposing implicit feasibility constraints on deformable image registration using a statistical generative model. Journal of Medical Imaging, 7(6):064005-064005, 2020. 2", + "[36] David W Shattuck, Mubeena Mirza, Vitria Adisetiyo, Cornelius Hojatkashani, Georges Salamon, Katherine L Narr, Russell A Poldrack, Robert M Bilder, and Arthur W Toga. Construction of a 3d probabilistic atlas of human cortical structures. Neuroimage, 39(3):1064-1080, 2008. 6", + "[37] Jiacheng Shi, Yuting He, Youyong Kong, Jean-Louis Coatrieux, Huazhong Shu, Guanyu Yang, and Shuo Li. Xmorph: Full transformer for deformable medical image registration via cross attention. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 217-226. Springer, 2022. 2", + "[38] Aristeidis Sotiras, Christos Davatzikos, and Nikos Paragios. Deformable medical image registration: A survey. IEEE transactions on medical imaging, 32(7):1153-1190, 2013. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 3", + "[40] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 5", + "[41] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15:1-28, 2015. 6", + "[42] Alexander Thorley, Xi Jia, Hyung Jin Chang, Boyang Liu, Karina Bunting, Victoria Stoll, Antonio de Marvao, Declan P O'Regan, Georgios Gkoutos, Dipak Kotecha, et al. Nesterov accelerated admm for fast diffeomorphic image registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 150-160. Springer, 2021. 1", + "[43] Lin Tian, Hastings Greer, François-Xavier Vialard, Roland Kwitt, Raul San José Estépar, Richard Jarrett Rushmore, Nikolaos Makris, Sylvain Bouix, and Marc Niethammer. Gradicon: Approximate diffeomorphisms via gradient inverse consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18084-18094, 2023. 2", + "[44] D Ulyanov. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016.3", + "[45] Tom Vercauteren, Xavier Pennec, Aymeric Perchant, and Nicholas Ayache. Diffeomorphic demons: Efficient nonparametric image registration. NeuroImage, 45(1):S61-S72, 2009. 1", + "[46] Haiqiao Wang and Yi Ni, Dongand Wang. Model: Learning deformable image registration via motion decomposition transformer. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2023, pages 740-749, 2023. 2, 3, 6, 7", + "[47] Zhoubing Xu, Christopher P Lee, Mattias P Heinrich, Marc Modat, Daniel Rueckert, Sebastien Ourselin, Richard G Abramson, and Bennett A Landman. Evaluation of six registration methods for the human abdomen on clinically acquired ct. IEEE Transactions on Biomedical Engineering, 63(8):1563-1572, 2016. 6", + "[48] Xiaofeng Yang, Ning Wu, Guanghui Cheng, Zhengyang Zhou, S Yu David, Jonathan J Beitler, Walter J Curran, and Tian Liu. Automated segmentation of the parotid gland based on atlas registration and machine learning: a longitudinal mri study in head-and-neck radiation therapy. International Journal of Radiation Oncology*Biology*Physics, 90 (5):1225-1233, 2014. 1", + "[49] Xiao Yang, Roland Kwitt, Martin Styner, and Marc Niethammer. Quicksilver: Fast predictive image registration-a deep learning approach. NeuroImage, 158:378-396, 2017. 2", + "[50] Jun Zhang. Inverse-consistent deep networks for unsu" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5236", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pervised deformable image registration. arXiv preprint arXiv:1809.03443, 2018. 2", + "[51] Yungeng Zhang, Yuru Pei, and Hongbin Zha. Learning dual transformer network for diffeomorphic registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 129-138. Springer, 2021. 2", + "[52] Shengyu Zhao, Yue Dong, Eric I Chang, Yan Xu, et al. Recursive cascaded networks for unsupervised medical image registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10600-10610, 2019. 2, 3", + "[53] Shengyu Zhao, Tingfung Lau, Ji Luo, I Eric, Chao Chang, and Yan Xu. Unsupervised 3d end-to-end medical image registration with volume tweening network. IEEE journal of biomedical and health informatics, 24(5):1394-1404, 2019. 2, 3", + "[54] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6647-6656, 2021. 3" + ], + "bbox": [ + 91, + 90, + 483, + 412 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5237", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_model.json b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4733f4c59d7aa74b05f730c0784258b81c70fe6b --- /dev/null +++ b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.131, + 0.88, + 0.153 + ], + "angle": 0, + "content": "SACB-Net: Spatial-awareness Convolutions for Medical Image Registration" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.18, + 0.757, + 0.217 + ], + "angle": 0, + "content": "Xinxing Cheng\\(^{1}\\), Tianyang Zhang\\(^{1}\\), Wenqi Lu\\(^{2}\\), Qingjie Meng\\(^{1,5}\\), Alejandro F. Frangi\\(^{3,4}\\), Jinming Duan\\(^{1,3,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.217, + 0.747, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science, University of Birmingham, UK" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.234, + 0.845, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computing and Mathematics, Manchester Metropolitan University, UK" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.252, + 0.833, + 0.269 + ], + "angle": 0, + "content": "3 Division of Informatics, Imaging and Data Sciences, University of Manchester, UK" + }, + { + "type": "list", + "bbox": [ + 0.153, + 0.217, + 0.845, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.27, + 0.875, + 0.306 + ], + "angle": 0, + "content": "4 Centre for Computational Imaging and Modelling in Medicine, University of Manchester, UK \n5 Department of Computing, Imperial College London, UK" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.308, + 0.822, + 0.322 + ], + "angle": 0, + "content": "{xxc142, txz009}@student.bham.ac.uk, W.Lu@mmu.ac.uk, m.qingjie@bham.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.326, + 0.78, + 0.34 + ], + "angle": 0, + "content": "alejandro.frangi@manchester.ac.uk, jinming.duan@manchester.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.375, + 0.327, + 0.391 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.408, + 0.485, + 0.786 + ], + "angle": 0, + "content": "Deep learning-based image registration methods have shown state-of-the-art performance and rapid inference speeds. Despite these advances, many existing approaches fall short in capturing spatially varying information in non-local regions of feature maps due to the reliance on spatially-shared convolution kernels. This limitation leads to suboptimal estimation of deformation fields. In this paper, we propose a 3D Spatial-Awareness Convolution Block (SACB) to enhance the spatial information within feature representations. Our SACB estimates the spatial clusters within feature maps by leveraging feature similarity and subsequently parameterizes the adaptive convolution kernels across diverse regions. This adaptive mechanism generates the convolution kernels (weights and biases) tailored to spatial variations, thereby enabling the network to effectively capture spatially varying information. Building on SACB, we introduce a pyramid flow estimator (named SACB-Net) that integrates SACBs to facilitate multi-scale flow composition, particularly addressing large deformations. Experimental results on the brain IXI and LPBA datasets as well as Abdomen CT datasets demonstrate the effectiveness of SACB and the superiority of SACB-Net over the state-of-the-art learning-based registration methods. The code is available at https://github.com/x-xc/SACB_Net." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.817, + 0.222, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Deformable image registration aims to align the corresponding anatomical structures between a moving and fixed" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.373, + 0.824, + 0.386 + ], + "angle": 0, + "content": "* Deformation is usually related to tissue shapes *" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.395, + 0.721, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.543, + 0.666, + 0.553 + ], + "angle": 0, + "content": "Vanilla Convolution" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.392, + 0.887, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.544, + 0.887, + 0.554 + ], + "angle": 0, + "content": "Spatial-awareness Convolution" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Figure 1. Spatial-awareness Convolution. Since deformation is usually related to tissue shape in medical image registration, voxels/features from different regions should be given varying levels of attention. However, vanilla convolution methods apply shared kernel weights across all regions, leading to a suboptimal estimation of deformation fields. SAC mechanisms, on the other hand, apply distinct kernel attention weights for different spatial clusters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.903 + ], + "angle": 0, + "content": "image pair with the estimated dense non-linear deformation field, which plays a crucial role in various medical imaging tasks [38, 48]. In the past, conventional methods like LDDMM [6], DARTEL [1], SyN [2], Demons [45], and ADMM [42] are time-consuming and computationally intensive, primarily due to the instance-level (pair-wise) iterative optimization required. Moreover, such methods may involve sophisticated hyper-parameter tuning, limiting their applications in large-scale volumetric registration. To overcome these challenges, learning-based image registration approaches have demonstrated significant success in achieving competitive accuracy along with fast inference speed. Learning-based approaches can be divided into supervised and unsupervised categories. Supervised methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5227" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "[7, 34, 49] often rely on the high quality of the synthetic ground truth deformation field or the annotated map [35] to predict a deformation field. Therefore, unsupervised image registration is more popular for estimating an optimal deformation field of pairwise images by minimizing the distance between the warped moving image and the fixed image with a regularization term." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.485, + 0.533 + ], + "angle": 0, + "content": "Following the classical unsupervised learning-based framework VoxelMorph [5], some U-Net-based methods focus on the performance improvements by adding stronger constraints over the deformation field such as inverse/cycle consistency [24, 50], gradient inverse consistency [43] and diffeomorphisms [13, 14, 29]. Some existing methods have introduced more advanced neural architectures, such as large-kernel network [20], vision transformers [8, 10, 17, 30, 37, 51]. To enhance the efficiency of the U-Net architecture and accelerate inference speed, several model-driven methods, including B-Spline [32], Fourier-Net [21] and WiNet [12], have been proposed to learn a low-dimensional representation of displacement. The lightweight approach, NCA-Morph [33], utilizes Neural Cellular Automata to improve the efficiency of image registration. Another track of recent works tends to progressively estimate the large and complex deformations using either multiple cascades [15, 31, 52, 53] or pyramid methods [18, 23, 25, 26, 28, 46] compositions of multi-scale flows derived from multi-scale features. The cascaded methods and pyramid coarse-to-fine methods are capable of effectively improving registration accuracy for estimating large deformation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.868 + ], + "angle": 0, + "content": "However, existing learning-based methods typically rely on networks that learn local feature correlations by applying shared convolution across all spatial dimensions within each layer. As illustrated in Fig. 1, deformations are often linked to the morphological characteristics of tissue in medical image registration, where voxels/features from various anatomical regions need to be paid different attention. However, plain convolutional methods utilize shared kernel weights uniformly across all regions, resulting in suboptimal estimation of deformation fields. Specifically, 3D medical images encapsulate complex intensity information relevant to anatomy, which makes it hard to capture spatially varying information in non-local regions of feature maps via spatially-shared convolution kernels. This key challenge is effectively incorporating spatially adaptive information, which can be addressed using region-based adaptive convolution techniques. Inspired by [16], which explores content-adaptive convolution in 2D image pansharpening, we extend this approach to 3D for medical image registration. This adaptation improves spatial awareness by modifying convolutional kernels based on regional content, thereby enhancing registration accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel 3D pyramid medical image registration network, named SACB-Net, which in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "tegrates spatial-awareness convolution block (SACB) with a similarity matching module to form a multi-scale flow estimator. The proposed SACB employs the unsupervised KMeans to cluster the learned feature maps. Each feature clustering is believed to capture a specific intensity or structural pattern of the feature maps and then subsequently used to learn the spatial-aware adaptive convolution kernels. The adaptive kernel generation in SACB allows the network to effectively learn refined feature representations for more accurate flow estimation. We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.259, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- We consider the spatial variance between two input images in the context of image registration, which is overlooked by vanilla convolutional-based approaches. The adaptive convolutions in our SACB-Net consider the changing of spatial regions to enhance registration accuracy. To the best of our knowledge, our method is the first work that employs 3D spatial-awareness convolution in 3D medical registration." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.38, + 0.905, + 0.47 + ], + "angle": 0, + "content": "- The proposed SACB aims to boost the moving and fixed feature maps using adaptive convolution kernels for different spatial regions, where the spatial regions are divided based on the clustering of features, and the convolution kernels are learned from the centroids of each feature clustering." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.471, + 0.904, + 0.545 + ], + "angle": 0, + "content": "- We propose an innovative pyramid flow estimator that integrates SACB and similarity matching module to perform multi-scale flow estimation, which can not only estimate local small deformation but also handle large deformation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.546, + 0.905, + 0.622 + ], + "angle": 0, + "content": "- Experimental results on atlas-based and inter-subject registration using two brain datasets (IXI and LPBA) demonstrate the effectiveness of SACB-Net. Further evaluation on inter-subject abdominal CT registration shows improved performance in handling large deformations." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.259, + 0.905, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.64, + 0.663, + 0.655 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.666, + 0.809, + 0.682 + ], + "angle": 0, + "content": "2.1. U-Net-based Registration Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "VoxelMorph [4, 5], a general framework for learning-based image registration, demonstrates the effectiveness of unsupervised learning in pairwise image registration using a simplified U-Net architecture. Building on this foundation, recent methods have enhanced registration performance by incorporating stronger constraints on the deformation field, such as inverse/cycle consistency [24, 50], gradient inverse consistency [43], and diffeomorphism constraints [13, 14, 29]. Another track of methods tried to capture long-range dependencies of feature correlations using either large kernel convolution [20] or vision transformer [8, 10, 17, 30, 37, 51]. TransMorph [10] incorporates Swin transformer blocks in the encoder, improving the network's capacity to capture long-range dependencies, though this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5228" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "comes with increased computational cost. LKU [20], on the other hand, employs a large kernel convolutional encoder to model long-range relationships with fewer parameters and reduced computational load. NCA-Morph [33] employs Neural Cellular Automata mechanisms, drawing inspiration from biological processes, to improve registration efficiency. While U-Net-based methods have shown comparable performance for small deformations, they struggle to accurately estimate large deformations or register images that are not well pre-aligned." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.395, + 0.27 + ], + "angle": 0, + "content": "2.2. Model-driven Registration Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.502 + ], + "angle": 0, + "content": "To reduce redundant convolution operations in the U-Net decoder and enhance training and inference efficiency, model-driven approaches such as B-Spline [32] and Fourier-Net [21] focus on learning a low-dimensional representation of the displacement field. However, displacement fields are not inherently smooth, and as a result, methods that rely solely on low-frequency components [21] or interpolate displacement from regular control points [32] may fail to fully and accurately capture the entire displacement field. To address this limitation, WiNet [12] introduces a wavelet-based decoder to preserve high-frequency information in the deformation field while maintaining network efficiency. Nonetheless, similar to U-Net-based methods, model-driven approaches also face challenges in effectively estimating large and complex deformations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.421, + 0.529 + ], + "angle": 0, + "content": "2.3. Cascaded and Pyramid-based Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Recent works that estimate large deformation can be either network cascade [15, 31, 52, 53] or coarse-to-fine pyramid [18, 23, 25, 26, 28, 46]. LapIRN [31] utilizes a Laplacian pyramid network to capture large deformations by cascading registered flows at three different scales. PRNet++ [23] utilizes a dual-stream pyramid architecture for coarse-to-fine flow estimation, performing sequential warping on multi-scale feature maps and incorporating local 3D correlation layers. Im2grid [25] embedded the coordinate information on extracted multi-scale feature maps and estimated the scale flow by calculating the matching score on the neighborhood feature. ModeT [46] leverages neighbourhood attention mechanisms to estimate multi-head, multiscale sub-deformation flows from pyramid feature maps. It then applies weighting modules to fuse these multi-head flows at each scale, generating the final deformation by sequential composition. To solve the insufficient decomposition problems, RDN [18] compose the multi-scale flows by performing level-wise recursion at each resolution level and stage-wise recursion at feature pyramid level." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.295, + 0.865 + ], + "angle": 0, + "content": "2.4. Adaptive Convolution" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Unlike standard convolution operators that use shared convolution kernels, adaptive convolution operators apply dif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "ferent kernels based on variations in the input, providing enhanced feature extraction capabilities and greater flexibility. Following [19], recent methods have applied adaptive convolution, demonstrating its effectiveness in various 2D computer vision tasks, such as image upsampling [39], classification [11, 54], segmentation [9] and remote sensing pansharpening [16, 22]. However, spatially adaptive convolution has not yet been explored in 3D medical image registration tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.241, + 0.605, + 0.256 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.265, + 0.694, + 0.28 + ], + "angle": 0, + "content": "3.1. Network Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.287, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Given a pair of moving image \\( I_{m} \\) and fixed image \\( I_{f} \\) over the spatial domain \\( \\Omega \\subset \\mathcal{R}^3 \\), the SACB-Net \\( (I_m, I_f; \\pmb{\\theta}) \\) network aims to learn the optimal parameters \\( \\pmb{\\theta}_{*} \\) that estimate a deformation field \\( \\phi_{m \\to f}: \\mathcal{R}^3 \\to \\mathcal{R}^3 \\), mapping the corresponding coordinates between the paired images. In unsupervised setting, the loss function contains \\( \\mathcal{L}_{sim}(\\cdot) \\) measuring the similarity between the warped moving image \\( I_{m} \\circ \\phi \\) and \\( I_{f} \\), and the regularization term. In our experiments, the smoothness term \\( \\mathcal{L}_{reg}(\\phi) = \\| \\nabla \\phi(\\pmb{\\theta}) \\|_2^2 \\) is computed using the \\( L_{2} \\)-norm of the gradient of the deformation field. The objective function of our network is" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.466, + 0.907, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\boldsymbol {\\theta}) = \\mathcal {L} _ {s i m} \\left(I _ {m} \\circ \\left(\\phi (\\boldsymbol {\\theta}) + \\operatorname {I d}\\right), I _ {f}\\right) + \\lambda \\mathcal {L} _ {r e g} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.568 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the learnable network parameters, \\(\\circ\\) represents the warping operator, \\(\\mathrm{Id}\\) is the identity grid, \\(\\nabla\\) refers to the first-order gradient, and \\(\\lambda\\) is the weight of the regularization term. We employ Normalized Cross-Correlation (NCC) to measure the similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.81 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, SACB-Net consists of a shared encoder and a pyramid flow estimator. The shared encoder is designed to extract multi-scale features \\(\\{F_m^i\\}\\) and \\(\\{F_f^i\\}\\) for the moving image \\(I_{m}\\) and the fixed image \\(I_{f}\\), respectively, in a pyramid structure. Here, \\(i \\in \\{1,2,3,4,5\\}\\) represents the scale of the features, with resolutions of \\(\\left[\\frac{D}{2^{i - 1}},\\frac{H}{2^{i - 1}},\\frac{W}{2^{i - 1}}\\right]\\) for the depth, height and width. The shared encoder consists of five 3D convolutional layers and four down-sampling layers. The convolutional layer is composed of 3D convolution operation, instance normalization [44] and LeakyReLU activations [27] with the 0.1 slope. Each convolutional layer is followed by an average pooling operation with a kernel size of 2. The extracted features \\(\\{F_m^i\\}\\) and \\(\\{F_f^i\\}\\) are then input into our pyramid flow estimators, where they are used to perform flow estimation and composition as outlined in Sections 3.2 and 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.862, + 0.835 + ], + "angle": 0, + "content": "3.2. 3D Spatial-awareness Convolution Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To effectively extract region-based context information for flow estimation, the proposed 3D SAC incorporates prior knowledge of spatial context through a spatial context estimation module and adaptive kernel generator. First, the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5229" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.088, + 0.907, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.358, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Illustration of a 5-level pyramid SACB-Net. SACB-Net includes a shared encoder that extracts multi-scale feature maps \\(\\{F_m^i\\}\\) and \\(\\{F_f^i\\}\\) for the moving image \\(I_{m}\\) and the fixed image \\(I_{f}\\), as well as pyramid flow estimators at each scale. At the lowest level, the flow estimator learns deformation \\((\\varphi_5)\\) from the extracted moving and fixed image features \\((F_m^5\\) and \\(F_{f}^{5})\\). The following flow estimators take the level-wise features and the output deformation from its preceding level to compose the deformation. Each flow estimator includes a Spatial-Awareness Convolution Block (SACB) to enhance spatially adaptive feature representation, along with a similarity matching module for flow estimation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.483, + 0.545 + ], + "angle": 0, + "content": "spatial context estimation module identifies spatial regions of interest, producing spatial contextual information. This contextual data is subsequently utilized by the adaptive kernel generator to dynamically produce convolutional kernels tailored to capture non-local, region-specific information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.826 + ], + "angle": 0, + "content": "Spatial Context Estimation. After obtaining the multiscale features \\(\\{F_m^i\\}\\) and \\(\\{F_f^i\\}\\), prior spatial region information is incorporated to refine these features. To achieve this, we employ unsupervised clustering method K-Means to estimate spatial region for feature maps. The reason not to use label information is due to potential inconsistencies in the labeling of \\(I_{m}\\) and \\(I_{f}\\) within the feature space, as well as to address the challenge of limited label availability. Let \\(\\mathbf{F} \\in \\mathbb{R}^{C_{in} \\times D \\times H \\times W}\\) denote the input feature map, where \\(C_{in}\\) represents the number of input channels, and \\(D\\), \\(H\\), and \\(W\\) correspond to the depth, height, and width of the spatial dimensions, respectively. As shown in Fig. 3, first, the feature map \\(\\mathbf{F}\\) is unfolded into local patches to capture neighborhood information using sliding windows of size \\(k\\). A spatial mean is then applied to each patch to reduce its spatial dimensions. For each voxel \\(V_{(d,h,w)}\\) located at coordinates \\((d,h,w)\\) within \\(\\mathbf{F}\\), suppose \\(p(\\cdot)\\) is the unfolding operation, the corresponding unfolded patch is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.852, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\np _ {\\left(V _ {\\left(d, h, w\\right)}\\right)} = \\left\\{V _ {\\left(d + d ^ {\\prime}, h + h ^ {\\prime}, w + w ^ {\\prime}\\right)} \\mid d ^ {\\prime}, h ^ {\\prime}, w ^ {\\prime} \\in \\left[ - \\frac {k}{2}, \\frac {k}{2} \\right] \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Here, \\(p_{(V_{(d,h,w)})}\\) represents a patch of voxels centered at \\(V_{(d,h,w)}\\), with offsets \\(d', h', w'\\) that range from \\(-\\frac{k}{2}\\) to \\(\\frac{k}{2}\\) (assuming \\(k\\) is odd). The unfolded feature denotes as \\(p(F) \\in \\mathbb{R}^{C_{in} \\times (D \\times H \\times W) \\times k^3}\\). The spatial mean of each local patch \\(\\overline{p}_{(V_{(d,h,w)})}\\) based on the neighborhood voxel is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.566, + 0.907, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\bar {p} _ {\\left(V _ {(d, h, w)}\\right)} = \\frac {1}{k ^ {3}} \\sum_ {d ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} \\sum_ {h ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} \\sum_ {w ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} p _ {\\left(V _ {(d, h, w)}\\right)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Afterward, the preprocessed feature \\(\\mathbf{F}\\) is reshaped to dimensions \\([D\\times H\\times W,C_{in}]\\) to estimate a cluster index matrix \\(S\\in \\mathbb{Z}^{D\\times H\\times W}\\), which assigns each voxel to a cluster \\(S_{n} = \\{(d,w,h)|S_{(d,w,h)} = n\\};n\\in \\{1,\\dots ,N\\}\\). Here, \\(N\\) represents the maximum number of clusters used for GPU-based KMeans clustering. Additionally, the centroid voxel of each cluster \\(S_{n}^{c}\\in \\mathbb{R}^{C_{in}}\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.729, + 0.907, + 0.768 + ], + "angle": 0, + "content": "\\[\nS _ {n} ^ {c} = \\frac {1}{\\left| S _ {n} \\right|} \\sum_ {(d, w, h) \\in S _ {n}} V _ {(d, w, h)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.907, + 0.834 + ], + "angle": 0, + "content": "where \\( |S_{n}| \\) denotes the total number of voxels belonging to the \\( n^{\\mathrm{th}} \\) spatial region. The cluster index matrix \\( S \\) and the centroid voxels \\( S^{c} \\) for each cluster are then utilized to generate adaptive kernel weights for convolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.836, + 0.907, + 0.882 + ], + "angle": 0, + "content": "Adaptive Kernel Generator. To adaptively generate a spatial convolution kernel that incorporates region-based information, we employ a multi-layer perception to model the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.766, + 0.901 + ], + "angle": 0, + "content": "The choices of \\(\\overline{p}\\) are investigated in Sec. 6." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5230" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.349, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 3. Architecture of the 3D spatial-awareness convolution block (example in three clusters). This block aims to refine the input feature \\(\\mathbf{F}\\) to \\(\\hat{\\mathbf{F}}\\) by adaptive convolution learned from spatial feature clustering. SACB consists of three parts: 1) the spatial context estimation module employs KMeans to cluster similar spatial features on the patched unfolding features, in which the features belonging to the same cluster centroid \\(S_{n}^{c}\\) will be indicated with a cluster index map \\(S_{n}\\). 2) the adaptive kernel generator leverages each cluster centroid to generate cluster-specific spatial weights and bias via two MLPs, the resulting weights and bias will be used to form the final spatial adaptive convolution kernel, which will be imposed on the cluster-indexed unfolding features, again highlighting the spatial-awareness. 3) the original features and spatial-aware features are residually connected to form the final refined \\(\\hat{\\mathbf{F}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.482, + 0.581 + ], + "angle": 0, + "content": "mapping \\(\\mathcal{F}_w\\), which produces an adaptive convolution kernel \\(\\mathbf{W}_n \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times k^3}\\) for the \\(n^{\\text{th}}\\) spatial region based on \\(S_n^c\\). To simplify the process, we first initialize a global learnable convolution kernel \\(\\mathbf{W} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times k^3}\\) and then use the mapping function \\(\\mathcal{F}_w\\) to generate a spatial weight that can be applied to the global kernel \\(\\mathbf{W}\\). Thus, \\(\\mathbf{W}_n\\) is obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.591, + 0.482, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {n} = \\mathcal {F} _ {w} \\left(S _ {n} ^ {c}\\right) \\odot \\mathbf {W}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.692 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise dot product. Then, a mapping function \\(\\mathcal{F}_b\\) using a multi-layer perception is applied to compute the convolution bias via obtained centroid \\(S_{n}^{c}\\). Consequently, the region-based spatially adaptive convolution is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.701, + 0.482, + 0.72 + ], + "angle": 0, + "content": "\\[\nS A C \\left(\\mathbf {F} _ {(d, w, h)}\\right) = p _ {\\left(V _ {(d, h, w)}\\right)} \\otimes \\mathbf {W} _ {n} + \\mathcal {F} _ {b} \\left(S _ {n} ^ {c}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.484, + 0.786 + ], + "angle": 0, + "content": "where \\( p(\\cdot) \\) denotes the unfolding operation outlined in the spatial context estimation section. Subsequently, the process of applying the SAC block to produce refined features is as follows" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.795, + 0.483, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} = S A C B (\\mathbf {F}) = \\mathbf {F} + \\sigma (S A C (\\mathbf {F})), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.822, + 0.351, + 0.837 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) denotes activation function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.318, + 0.861 + ], + "angle": 0, + "content": "3.3. Pyramid Flow Estimator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.866, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Let \\(\\hat{F_m}^i\\) and \\(\\hat{F_f}^i\\) represent the multi-scale features refined by the proposed SACB at the \\(i^{th}\\) level, the sub-deformation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.595 + ], + "angle": 0, + "content": "flows at each scale are then estimated based on the matching score using the refined features. Specifically, we calculate the similarity matching score between each voxel in the fixed image features and corresponding neighbouring voxels in the moving image features. Similar to [25, 40], this operation can be efficiently implemented as a matrix operation using the dot product. For instance, at the coarsest level \\( i = 5 \\), the matching score can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.602, + 0.905, + 0.625 + ], + "angle": 0, + "content": "\\[\nM _ {s i m} \\left(\\hat {F _ {f}} ^ {5}, \\hat {F _ {m}} ^ {5}\\right) = \\operatorname {S o f t m a x} \\left(\\hat {F _ {f}} ^ {5} p \\left(\\hat {F _ {m}} ^ {5}\\right) ^ {T}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.906, + 0.682 + ], + "angle": 0, + "content": "where \\(p(\\cdot)\\) represents the unfolding operation as described in Section 3.2. For each voxel \\(V\\) at \\(\\hat{F_m}^5\\), the similarity matching score can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.69, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\nM _ {s i m} \\left(\\hat {F} _ {f} ^ {5} (V), \\hat {F} _ {m} ^ {5} (V)\\right) = \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.715, + 0.864, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S o f t m a x} \\left(\\left\\{\\left\\langle \\hat {F _ {f}} ^ {5} (V), \\hat {F _ {m}} ^ {5} (V _ {c}) \\right\\rangle \\right\\} _ {V _ {c} \\in \\mathcal {N} (V)}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.907, + 0.832 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot, \\cdot \\rangle\\) denotes the inner product, and \\(\\mathcal{N}(V)\\) represents the set of voxels within an \\(k \\times k \\times k\\) neighborhood centered around voxel \\(V\\), \\(k\\) represents the window size of the search region of \\(\\hat{F}_m^5\\). We set \\(k = 3\\) in our experiments. Then, the sub-deformation flow at the \\(5^{\\text{th}}\\) scale can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.838, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\varphi^ {5} = M _ {s i m} \\left(\\hat {F _ {f}} ^ {5}, \\hat {F _ {m}} ^ {5}\\right) G \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\(G \\in \\mathbb{R}^{(k \\times k \\times k) \\times 3}\\) is a 3D grid that encodes the relative positions with respect to the centroid voxel location." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5231" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.16 + ], + "angle": 0, + "content": "At the \\(i = 4\\) level, we first up scaling the flow \\(\\varphi^5\\) to form \\(\\hat{\\varphi}^5\\), then \\(\\hat{F_m}^4\\) is warped with \\(\\hat{\\varphi}^5\\). Sequentially, the \\(M_{sim}\\) need to be calculated between \\(\\hat{F_f}^4\\) and \\(\\hat{F_m}^4 \\circ \\hat{\\varphi}^5\\) for obtaining \\(\\Delta \\varphi^4\\). Afterward, the overall process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.17, + 0.482, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\hat {\\varphi} ^ {i} = \\mathrm {u p} _ {2 \\times} \\left(\\varphi^ {i}\\right), \\\\ \\hat {F} _ {f} ^ {i - 1} = S A C B \\left(F _ {f} ^ {i - 1}\\right), \\\\ \\hat {F} _ {m} ^ {i - 1} = S A C B \\left(F _ {m} ^ {i - 1} \\circ \\hat {\\varphi} ^ {i}\\right), \\quad i \\in \\{5, 4, 3, 2 \\}. \\\\ \\Delta \\varphi^ {i - 1} = M _ {s i m} \\left(\\hat {F} _ {f} ^ {i - 1}, \\hat {F} _ {m} ^ {i - 1}\\right) G, \\\\ \\varphi^ {i - 1} = \\hat {\\varphi} ^ {i} \\circ \\Delta \\varphi^ {i - 1}, \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.297 + ], + "angle": 0, + "content": "The \\(\\mathrm{up}_{2\\times}(\\cdot)\\) is flow rescaling operation with \\(2\\times\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.297, + 0.484, + 0.417 + ], + "angle": 0, + "content": "We note that the matching score calculation was not applied for flow estimation at the \\(1^{st}\\) scale, as the search region with \\(k = 3\\) can only determine a maximum displacement of 1 voxel using full-resolution features. However, the calculations at the \\(1^{st}\\) scale significantly increase memory and computational costs, yielding limited benefit. Therefore, we used a simple approach with two convolutional layers for flow estimation at the final scale." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.43, + 0.223, + 0.447 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.455, + 0.485, + 0.818 + ], + "angle": 0, + "content": "Datasets. Our experiments use two publicly available 3D brain MRI datasets \\(\\mathrm{IXI}^2\\) and LPBA [36], along with one public abdominal CT dataset [47]. The preprocessed IXI dataset [10] includes 576 MRI scans \\((160\\times 192\\times 224)\\) from healthy subjects. We followed the same protocol as [10], partitioning the data into 403, 58, and 115 scans for training, validation, and testing sets, respectively. Atlas-based brain registration was conducted using a generated atlas from [24]. The LPBA dataset consists of 40 T1-weighted MRI volumes. Each MRI scan has been rigidly pre-aligned to the MNI305 template and manually annotated with 54 regions of interest (ROIs). We center-crop all volumes to dimensions of \\(160\\times 192\\times 160\\) with voxel size \\(1\\times 1\\times 1mm^3\\). LPBA dataset was used to perform inter-subject registration with divided three subsets: 25 volumes \\((25\\times 24\\) pairs) for training, 5 \\((5\\times 4\\) pairs) for validation, and 10 \\((10\\times 9\\) pairs) for testing. The Abdomen CT dataset consists of 30 abdominal scans with 13 manually labelled anatomical structures. We clamp the Hounsfield Units (HU) to \\([-1000, 1000]\\) and normalize each scan to \\([0, 1]\\). Each scan is center-cropped from \\(192\\times 160\\times 256\\) to \\(192\\times 160\\times 224\\) with voxel size \\(2\\times 2\\times 2mm^3\\). The dataset is divided into 19 training volumes \\((19\\times 18\\) pairs), 3 validation volumes \\((3\\times 2\\) pairs) and 8 testing volumes \\((8\\times 7\\) pairs) for inter-subject registration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Implementation Details. Our method is implemented using PyTorch. Both the training and testing phases are deployed on an A100 GPU with 40GB VRAM. All models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "are optimized using Adam followed with the same learning rate \\(10^{-4}\\) and batch size 1. A fast GPU K-Means implementation is adopted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.14, + 0.907, + 0.277 + ], + "angle": 0, + "content": "Evaluation Metrics. To qualify the registration performance, we use the Dice similarity coefficient (Dice) to evaluate the region of interest (ROI) segmentation overlap, the average symmetric surface distance (ASSD) [41] and the 95 percentile of the Hausdorff distance (HD95) to measure the registration accuracy of ROI structures. For evaluating the diffeomorphism of deformation fields, we use the percentage of negative values of the Jacobian determinant \\((|J|_{<0\\%})\\) to calculate the percentage of folding voxels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.909, + 0.432 + ], + "angle": 0, + "content": "Comparative Methods. We compare our method against state-of-the-art approaches, including the conventional SyN method [2], U-Net-based models such as VoxelMorph (VM) [5], LKU [20], and TransMorph [10], model-driven methods like B-Spline-Diff [32] and Fourier-Net [21], the lightweight NCA-Morph [33], the cascaded method LapIRN [31], and four pyramid-based methods: PR++ [23], Im2Grid [25], RDN [18] and ModeT [46]. All models were trained using their officially released code, with optimal hyperparameters tuned on the validation sets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.445, + 0.602, + 0.46 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.545 + ], + "angle": 0, + "content": "In this section, we evaluate our SACB-Net against comparative methods on three tasks: atlas-based brain registration using the IXI dataset, inter-subject brain registration using the LPBA dataset, and large deformation estimation for inter-subject abdominal CT registration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.555, + 0.788, + 0.57 + ], + "angle": 0, + "content": "5.1. Atlas-based Brain Registration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.907, + 0.803 + ], + "angle": 0, + "content": "Table 1 presents the quantitative results of atlas-based brain registration for 30 anatomical structures evaluated on the IXI dataset. It can be seen that SACB-Net outperforms all comparable methods in terms of Dice score while achieving plausible deformations and third-best results on HD95 and ASSD. Compared to the best-performing U-Net-based method (LKU), our SACB-Net achieves a \\(0.4\\%\\) improvement in Dice score while using 0.98M fewer parameters. In terms of model-driven methods, SACB-Net surpasses Fourier-Net by a margin of \\(0.6\\%\\) in the Dice score and uses 3.09M fewer parameters. For cascade and pyramid-based methods, those methods have a similar number of parameters apart from RDN. Our SACB-Net outperforms the second-best cascade method LapIRN with \\(0.6\\%\\) Dice, 0.038mm HD95, and 0.019mm ASSD, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.812, + 0.798, + 0.829 + ], + "angle": 0, + "content": "5.2. Inter-subject Brain Registration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.834, + 0.906, + 0.879 + ], + "angle": 0, + "content": "The inter-subject registration of the LPBA dataset tends to have a slightly larger deformation than the atlas-based registration on the IXI dataset. As shown in Table 1, most UNet" + }, + { + "type": "page_footnote", + "bbox": [ + 0.107, + 0.875, + 0.449, + 0.887 + ], + "angle": 0, + "content": "2https://brain-development.org/ixi-dataset/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.107, + 0.887, + 0.418, + 0.901 + ], + "angle": 0, + "content": "3https://learn2reg.grand-challenge.org/" + }, + { + "type": "list", + "bbox": [ + 0.107, + 0.875, + 0.449, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.839, + 0.901 + ], + "angle": 0, + "content": "4https://github.com/jokofa/torch_KMeans" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5232" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.268, + 0.09, + 0.728, + 0.104 + ], + "angle": 0, + "content": "Table 1. Registration performance comparison on the IXI and LPBA datasets." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.115, + 0.905, + 0.348 + ], + "angle": 0, + "content": "
MethodIXI (30 ROIs)LPBA (54 ROIs)#Param
Dice↑HD95↓ASSD↓|J|<0%↓Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.386±0.1956.479±0.6662.445±0.280-0.525±0.0478.039±0.8612.586±0.350--
SyN [3]0.645±0.1526.394±1.0481.551±0.286<0.00010.707±0.0166.254±0.4441.479±0.131<0.0001-
VM-1 [5]0.729±0.1293.798±0.7570.937±0.1621.590±0.3390.664±0.0256.873±0.6541.717±0.2000.649±0.2610.27M
VM-2 [5]0.732±0.1233.723±0.6800.926±0.1581.522±0.3360.669±0.0256.847±0.6591.698±0.2000.591±0.2420.30M
NCA-Morph [33]0.753±0.1363.109±0.5250.796±0.1210.506±0.1900.679±0.0236.666±0.6341.631±0.1880.264±0.1210.37M
TransMorph [10]0.754±0.1243.543±0.7210.862±0.1681.579±0.3280.695±0.0226.564±0.6191.559±0.1820.474±0.17646.77M
LKU [20]0.765±0.1292.967±0.4940.757±0.1140.109±0.0540.706±0.0326.452±0.9511.603±0.2500.594±0.2032.09M
B-Spline-Diff [32]0.742±0.1283.256±0.5380.832±0.117<0.00010.665±0.0236.792±0.6221.713±0.1860.0±0.00.27M
Fourier-Net [21]0.763±0.1292.857±0.4560.748±0.1140.024±0.0190.672±0.0226.716±0.6011.666±0.1800.216±0.1044.20M
LapIRN [31]0.763±0.1333.166±0.6080.779±0.1260.312±0.1060.716±0.0166.116±0.4541.426±0.1330.024±0.0091.20M
PRNet++ [23]0.755±0.1303.593±0.7480.857±0.1621.052±0.3020.701±0.0216.492±0.5971.520±0.1770.072±0.0271.24M
ModeT [46]0.758±0.1253.496±0.7320.828±0.1510.114±0.0570.721±0.0135.969±0.4161.375±0.1100.010±0.0041.03M
Im2Grid [25]0.761±0.1273.316±0.6680.799±0.128<0.00020.713±0.0146.062±0.4281.419±0.1180.007±0.0030.89M
RDN [18]0.759±0.1233.476±0.8020.823±0.161<0.00010.713±0.0176.208±0.4971.436±0.142<0.000228.65M
Ours0.769±0.1273.128±0.6310.760±0.1250.083±0.0450.731±0.0125.862±0.4361.326±0.1140.018±0.0061.11M
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.482, + 0.555 + ], + "angle": 0, + "content": "based and model-driven methods do not perform as well as cascade and pyramid-based methods. Only LKU surpasses the pyramid-based method PR++, achieving improvements of \\(0.5\\%\\) in Dice, \\(0.04 \\mathrm{~mm}\\) in HD95 and \\(0.083 \\mathrm{~mm}\\) in ASSD, respectively. Our SACB outperforms all comparable methods in Dice, HD95 and ASSD with plausible deformations. Compared to the second-best method, ModeT, SACB achieves improvements of \\(1\\%\\) in Dice, \\(0.107 \\mathrm{~mm}\\) in HD95, and \\(0.049 \\mathrm{~mm}\\) in ASSD, respectively. The visualization in Fig. 4 shows that our SACB-Net captures more detailed correspondences in the registration process, as highlighted in the red box." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.572, + 0.406, + 0.588 + ], + "angle": 0, + "content": "5.3. Inter-subject Abdomen Registration" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.597, + 0.483, + 0.853 + ], + "angle": 0, + "content": "The primary challenge in abdominal CT registration tasks is caused by the substantial differences in the spatial distribution of organs across scans, making it difficult to accurately estimate large deformations during the registration process. Table 2 demonstrates the numerical results of the inter-subject abdomen registration task. Our results show that our method outperforms all compared approaches in terms of Dice, HD95, and ASSD. Notably, U-Net-based and model-driven methods perform inferior to pyramid and cascaded approaches for estimating large deformations. Even the best-performing method in such approaches, TransMorph, still exhibits a significant gap compared to our SACB-Net with \\(14.4\\%\\) difference in Dice score, \\(5.934\\mathrm{mm}\\) in HD95 and \\(3.153\\mathrm{mm}\\) in ASSD, respectively. Compared to the second-best pyramid method, ModeT, SACB-Net achieves improvements of \\(3.8\\%\\) in Dice, \\(2.098\\mathrm{mm}\\) in HD95, and \\(0.781\\mathrm{mm}\\) in ASSD, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The visualization of the wrapped moving images and labels, as shown in Fig. 4, illustrates that most parts of the kidneys are registered with our method. However, Trans" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.373, + 0.902, + 0.387 + ], + "angle": 0, + "content": "Table 2. Registration performance comparison on Abdomen CT." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.397, + 0.905, + 0.582 + ], + "angle": 0, + "content": "
MethodAbdomen CT (13 ROIs)
Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.305±0.04727.790±4.86810.869±2.075-
VM-2 [5]0.406±0.05825.226±4.9428.890±2.0663.882±1.512
NCA-Morph [33]0.406±0.06125.959±4.9829.083±2.2235.197±1.166
TransMorph [10]0.444±0.06524.187±4.9558.156±2.0656.045±1.750
LKU [20]0.423±0.05824.252±4.6048.395±1.9464.316±1.778
B-Spline-Diff [32]0.375±0.05325.338±4.8669.217±1.9600.0±0.0
Fourier-Net [21]0.417±0.04823.084±4.2018.000±1.5322.167±2.391
LapIRN [31]0.522±0.05720.728±4.3186.274±1.3793.203±0.848
PRNet++ [23]0.478±0.06023.791±4.6917.524±1.7712.252±1.192
ModeT [46]0.550±0.05420.351±4.2315.784±1.2362.239±0.906
Im2Grid [25]0.530±0.04720.763±3.7866.028±1.1462.421±0.653
RDN [18]0.527±0.05521.782±4.7226.461±1.6131.529±0.989
Ours0.588±0.04918.253±3.6105.003±1.0113.930±1.416
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.621, + 0.905, + 0.652 + ], + "angle": 0, + "content": "Morph, \\(\\mathrm{PR} + +\\), and RDN failed to register the kidneys, leaving a large portion missing on the warped mask." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.665, + 0.7 + ], + "angle": 0, + "content": "6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.905, + 0.79 + ], + "angle": 0, + "content": "First, we note that \\(\\overline{p}(\\cdot)\\) in Eq. (3) can also be defined on the channel dimension or both the spatial+channel (Mix) dimension. As shown in Table 4, using the spatial mean of unfolded patches achieves slightly better performance compared to the channel mean and the Mix setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We then determine the optimal number of scales at which SACB should be applied and the ideal number of clusters to achieve accurate registration performance. As shown in Table 3, the optimal configuration is applying SACB from scale \\(5\\) to scale \\(2\\). Additionally, increasing the \\(N\\) from 5 to 7 boosts the performance while a too-large value slightly decreases performance (e.g. 11)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5233" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.907, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.387, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons on the Brain LPBA dataset (top two rows) and the Abdomen CT dataset (bottom two rows). Columns 3-10 display warped moving images (row 1 and 3), displacement fields as RGB images (row 2) and warped segmentation masks (row 4)." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.43, + 0.908, + 0.472 + ], + "angle": 0, + "content": "Table 3. Ablation studies on the use of SACB at different scale levels in SACB-Net and varying numbers of spatial clusters \\((N)\\), along with the corresponding average training time (in seconds), including data loading time, and GPU training memory usage (in MiB). The SACB configurations are based on the best results shown in Table 4." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.482, + 0.905, + 0.613 + ], + "angle": 0, + "content": "
NWith SACBIXILPBAAbdomen CT
Scale5Scale4Scale3Scale2Dice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTraining
-XXXX0.76430.059158380.7620.71410.020115740.5450.53742.907158380.764
5✓XXX0.76680.072158800.7990.72170.087116640.5790.54443.243158800.797
5✓✓XX0.76790.076161040.8790.72410.104117700.6310.54982.619161040.864
5✓✓✓X0.76710.060169600.9450.72660.076124180.7140.56852.781169600.950
5✓✓✓✓0.76830.086216381.2550.72940.017157480.9180.58493.548216381.234
7✓✓✓✓0.76910.083216381.3760.73090.018157781.0840.58813.930216621.360
9✓✓✓✓0.76850.069216801.5500.73040.016157921.1100.58754.405216801.562
11✓✓✓✓0.76840.084217321.6790.73000.015158061.1930.58754.419217181.728
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.635, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Table 4. Ablations of the SACB configurations. After padding and unfolding on the feature map \\(\\mathbf{F}\\), the cluster maps can be calculated based on spatial mean on the local unfolded patch, channel mean and Mix (spatial+channel). In this study, we set \\(N = 5\\) and apply SACBs from scale\\(_5\\) to scale\\(_2\\)." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.716, + 0.432, + 0.803 + ], + "angle": 0, + "content": "
SACBLPBA
SpatialChannelMixDice ↑|J|<0% ↓
✓✗✗0.72940.017
✗✓✗0.72810.020
✗✗✓0.72850.021
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.21, + 0.844 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We introduced SACB-Net, which incorporates a 3D Spatial Awareness Convolution Block (SACB) that adaptively generates spatial convolution kernels based on the feature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.789 + ], + "angle": 0, + "content": "clusters, thereby improving the network's ability to capture spatially varying information. Additionally, our proposed flow estimators combine SACB with a similarity-matching mechanism to estimate displacements in a pyramid coarse-to-fine manner. Results on two brain datasets and one abdomen dataset demonstrate that SACB-Net effectively handles both small local and large deformations. Furthermore, the proposed flow estimator is a plug-in module that can be applied in other networks for better feature representation learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. The research was conducted using the Baskerville Tier 2 HPC service, which was funded by the EPSRC and UKRI through the World Class Labs scheme (EP/T022221/1) and the Digital Research Infrastructure programme (EP/W032244/1). Baskerville is operated by Advanced Research Computing at the University of Birmingham." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5234" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.142 + ], + "angle": 0, + "content": "[1] John Ashburner. A fast diffeomorphic image registration algorithm. Neuroimage, 38(1):95-113, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal, 12(1):26-41, 2008. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Brian B Avants, Nicholas J Tustison, Gang Song, Philip A Cook, Arno Klein, and James C Gee. A reproducible evaluation of ants similarity metric performance in brain image registration. Neuroimage, 54(3):2033-2044, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[4] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9252-9260, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[6] M Faisal Beg, Michael I Miller, Alain Trouve, and Laurent Younes. Computing large deformation metric mappings via geodesic flows of diffeomorphisms. International journal of computer vision, 61:139-157, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[7] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Qian Wang, Pew-Thian Yap, and Dinggang Shen. Deformable image registration using a cue-aware deep regression network. IEEE Transactions on Biomedical Engineering, 65(9):1900-1911, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[8] Junyu Chen, Yufan He, Eric C Frey, Ye Li, and Yong Du. Vit-v-net: Vision transformer for unsupervised volumetric medical image registration. arXiv preprint arXiv:2104.06468, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[9] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8064-8073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[10] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[11] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11030-11039, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Xinxing Cheng, Xi Jia, Wenqi Lu, Qiufu Li, Linlin Shen, Alexander Krull, and Jinming Duan. Winet: Wavelet-based incremental learning for efficient medical image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 761-771. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Adrian Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces. Medical image analysis, 57:226-236, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Adrian V Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning for fast probabilistic diffeomorphic registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 729-738. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[15] Bob D De Vos, Floris F Berendsen, Max A Viergever, Hessam Sokooti, Marius Staring, and Ivana Isgum. A deep learning framework for unsupervised affine and deformable image registration. Medical image analysis, 52:128-143, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[16] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.302, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[17] Morteza Ghahremani, Mohammad Khateri, Bailiang Jian, Benedikt Wiestler, Ehsan Adeli, and Christian Wachinger. H-vit: A hierarchical vision transformer for deformable image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11513-11523, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.904, + 0.439 + ], + "angle": 0, + "content": "[18] Bo Hu, Shenglong Zhou, Zhiwei Xiong, and Feng Wu. Recursive decomposition network for deformable image registration. IEEE Journal of Biomedical and Health Informatics, 26(10):5130-5141, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[19] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Xi Jia, Joseph Bartlett, Tianyang Zhang, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. U-net vs transformer: Is u-net outdated in medical image registration? In International Workshop on Machine Learning in Medical Imaging, pages 151-160. Springer, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[21] Xi Jia, Joseph Bartlett, Wei Chen, Siyang Song, Tianyang Zhang, Xinxing Cheng, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. Fourier-net: Fast image registration with band-limited deformation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1015-1023, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[22] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI conference on artificial intelligence, pages 1113-1121, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[23] Miao Kang, Xiaojun Hu, Weilin Huang, Matthew R Scott, and Mauricio Reyes. Dual-stream pyramid registration network. Medical image analysis, 78:102379, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[24] Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. Cycle morph: cycle consistent unsupervised deformable image registration. Medical image analysis, 71:102036, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Yihao Liu, Lianrui Zuo, Shuo Han, Yuan Xue, Jerry L Prince, and Aaron Carass. Coordinate translator for learning deformable medical image registration. In International workshop on multiscale multimodal medical imaging, pages 98-109. Springer, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Tai Ma, Suwei Zhang, Jiafeng Li, and Ying Wen. Iirp-net: Iterative inference residual pyramid network for enhanced" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5235" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11555, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[27] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[28] Mingyuan Meng, Dagan Feng, Lei Bi, and Jinman Kim. Correlation-aware coarse-to-fine mlp's for deformable medical image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9645-9654, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[29] Tony CW Mok and Albert Chung. Fast symmetric diffeomorphic image registration with convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4644-4653, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[30] Tony CW Mok and Albert Chung. Affine medical image registration with coarse-to-fine vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20835-20844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[31] Tony CW Mok and Albert CS Chung. Large deformation diffeomorphic image registration with laplacian pyramid networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 211-221. Springer, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.512 + ], + "angle": 0, + "content": "[32] Huaqi Qiu, Chen Qin, Andreas Schuh, Kerstin Hammernik, and Daniel Rueckert. Learning diffeomorphic and modality-invariant registration using b-splines. In Medical Imaging with Deep Learning, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[33] Amin Ranem, John Kalkhof, and Anirban Mukhopadhyay. Nca-morph: Medical image registration with neural cellular automata. arXiv preprint arXiv:2410.22265, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[34] Marc-Michel Rohe, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: learning deformable image registration using shape matching. In Medical Image Computing and Computer Assisted Intervention- MIC-CAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part I 20, pages 266-274. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[35] Yudi Sang, Xianglei Xing, Yingnian Wu, and Dan Ruan. Imposing implicit feasibility constraints on deformable image registration using a statistical generative model. Journal of Medical Imaging, 7(6):064005-064005, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[36] David W Shattuck, Mubeena Mirza, Vitria Adisetiyo, Cornelius Hojatkashani, Georges Salamon, Katherine L Narr, Russell A Poldrack, Robert M Bilder, and Arthur W Toga. Construction of a 3d probabilistic atlas of human cortical structures. Neuroimage, 39(3):1064-1080, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[37] Jiacheng Shi, Yuting He, Youyong Kong, Jean-Louis Coatrieux, Huazhong Shu, Guanyu Yang, and Shuo Li. Xmorph: Full transformer for deformable medical image registration via cross attention. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 217-226. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Aristeidis Sotiras, Christos Davatzikos, and Nikos Paragios. Deformable medical image registration: A survey. IEEE transactions on medical imaging, 32(7):1153-1190, 2013. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[39] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[40] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[41] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15:1-28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[42] Alexander Thorley, Xi Jia, Hyung Jin Chang, Boyang Liu, Karina Bunting, Victoria Stoll, Antonio de Marvao, Declan P O'Regan, Georgios Gkoutos, Dipak Kotecha, et al. Nesterov accelerated admm for fast diffeomorphic image registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 150-160. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[43] Lin Tian, Hastings Greer, François-Xavier Vialard, Roland Kwitt, Raul San José Estépar, Richard Jarrett Rushmore, Nikolaos Makris, Sylvain Bouix, and Marc Niethammer. Gradicon: Approximate diffeomorphisms via gradient inverse consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18084-18094, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[44] D Ulyanov. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[45] Tom Vercauteren, Xavier Pennec, Aymeric Perchant, and Nicholas Ayache. Diffeomorphic demons: Efficient nonparametric image registration. NeuroImage, 45(1):S61-S72, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[46] Haiqiao Wang and Yi Ni, Dongand Wang. Model: Learning deformable image registration via motion decomposition transformer. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2023, pages 740-749, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[47] Zhoubing Xu, Christopher P Lee, Mattias P Heinrich, Marc Modat, Daniel Rueckert, Sebastien Ourselin, Richard G Abramson, and Bennett A Landman. Evaluation of six registration methods for the human abdomen on clinically acquired ct. IEEE Transactions on Biomedical Engineering, 63(8):1563-1572, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[48] Xiaofeng Yang, Ning Wu, Guanghui Cheng, Zhengyang Zhou, S Yu David, Jonathan J Beitler, Walter J Curran, and Tian Liu. Automated segmentation of the parotid gland based on atlas registration and machine learning: a longitudinal mri study in head-and-neck radiation therapy. International Journal of Radiation Oncology*Biology*Physics, 90 (5):1225-1233, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.843, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[49] Xiao Yang, Roland Kwitt, Martin Styner, and Marc Niethammer. Quicksilver: Fast predictive image registration-a deep learning approach. NeuroImage, 158:378-396, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Jun Zhang. Inverse-consistent deep networks for unsu" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5236" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "pervised deformable image registration. arXiv preprint arXiv:1809.03443, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[51] Yungeng Zhang, Yuru Pei, and Hongbin Zha. Learning dual transformer network for diffeomorphic registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 129-138. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[52] Shengyu Zhao, Yue Dong, Eric I Chang, Yan Xu, et al. Recursive cascaded networks for unsupervised medical image registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10600-10610, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[53] Shengyu Zhao, Tingfung Lau, Ji Luo, I Eric, Chao Chang, and Yan Xu. Unsupervised 3d end-to-end medical image registration with volume tweening network. IEEE journal of biomedical and health informatics, 24(5):1394-1404, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.484, + 0.414 + ], + "angle": 0, + "content": "[54] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6647-6656, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5237" + } + ] +] \ No newline at end of file diff --git a/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_origin.pdf b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5315d587cfc942d7337517537ffe0a652fece074 --- /dev/null +++ b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/91fc8aba-6a9b-4595-ad8c-fb4d819f6f44_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:962a3249151614d1a8a60c24b2167b5abbd2bb8ec385b6dde03318e1672cb4d6 +size 1167472 diff --git a/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/full.md b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e56e1d91ffbb40baf0e4271a79593879bc40e92d --- /dev/null +++ b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/full.md @@ -0,0 +1,305 @@ +# SACB-Net: Spatial-awareness Convolutions for Medical Image Registration + +Xinxing Cheng $^{1}$ , Tianyang Zhang $^{1}$ , Wenqi Lu $^{2}$ , Qingjie Meng $^{1,5}$ , Alejandro F. Frangi $^{3,4}$ , Jinming Duan $^{1,3,4*}$ + +$^{1}$ School of Computer Science, University of Birmingham, UK +$^{2}$ Department of Computing and Mathematics, Manchester Metropolitan University, UK +3 Division of Informatics, Imaging and Data Sciences, University of Manchester, UK + +4 Centre for Computational Imaging and Modelling in Medicine, University of Manchester, UK +5 Department of Computing, Imperial College London, UK + +{xxc142, txz009}@student.bham.ac.uk, W.Lu@mmu.ac.uk, m.qingjie@bham.ac.uk + +alejandro.frangi@manchester.ac.uk, jinming.duan@manchester.ac.uk + +# Abstract + +Deep learning-based image registration methods have shown state-of-the-art performance and rapid inference speeds. Despite these advances, many existing approaches fall short in capturing spatially varying information in non-local regions of feature maps due to the reliance on spatially-shared convolution kernels. This limitation leads to suboptimal estimation of deformation fields. In this paper, we propose a 3D Spatial-Awareness Convolution Block (SACB) to enhance the spatial information within feature representations. Our SACB estimates the spatial clusters within feature maps by leveraging feature similarity and subsequently parameterizes the adaptive convolution kernels across diverse regions. This adaptive mechanism generates the convolution kernels (weights and biases) tailored to spatial variations, thereby enabling the network to effectively capture spatially varying information. Building on SACB, we introduce a pyramid flow estimator (named SACB-Net) that integrates SACBs to facilitate multi-scale flow composition, particularly addressing large deformations. Experimental results on the brain IXI and LPBA datasets as well as Abdomen CT datasets demonstrate the effectiveness of SACB and the superiority of SACB-Net over the state-of-the-art learning-based registration methods. The code is available at https://github.com/x-xc/SACB_Net. + +# 1. Introduction + +Deformable image registration aims to align the corresponding anatomical structures between a moving and fixed + +![](images/14b37902be4244111f335d78f924e127a49c1699ba8c0882e2d5dcea27f7853e.jpg) +Vanilla Convolution + +![](images/26fe5989a3ba2b8110d8af6a587b535fbd415434a1709e850f053283625aea62.jpg) +* Deformation is usually related to tissue shapes * +Spatial-awareness Convolution +Figure 1. Spatial-awareness Convolution. Since deformation is usually related to tissue shape in medical image registration, voxels/features from different regions should be given varying levels of attention. However, vanilla convolution methods apply shared kernel weights across all regions, leading to a suboptimal estimation of deformation fields. SAC mechanisms, on the other hand, apply distinct kernel attention weights for different spatial clusters. + +image pair with the estimated dense non-linear deformation field, which plays a crucial role in various medical imaging tasks [38, 48]. In the past, conventional methods like LDDMM [6], DARTEL [1], SyN [2], Demons [45], and ADMM [42] are time-consuming and computationally intensive, primarily due to the instance-level (pair-wise) iterative optimization required. Moreover, such methods may involve sophisticated hyper-parameter tuning, limiting their applications in large-scale volumetric registration. To overcome these challenges, learning-based image registration approaches have demonstrated significant success in achieving competitive accuracy along with fast inference speed. Learning-based approaches can be divided into supervised and unsupervised categories. Supervised methods + +[7, 34, 49] often rely on the high quality of the synthetic ground truth deformation field or the annotated map [35] to predict a deformation field. Therefore, unsupervised image registration is more popular for estimating an optimal deformation field of pairwise images by minimizing the distance between the warped moving image and the fixed image with a regularization term. + +Following the classical unsupervised learning-based framework VoxelMorph [5], some U-Net-based methods focus on the performance improvements by adding stronger constraints over the deformation field such as inverse/cycle consistency [24, 50], gradient inverse consistency [43] and diffeomorphisms [13, 14, 29]. Some existing methods have introduced more advanced neural architectures, such as large-kernel network [20], vision transformers [8, 10, 17, 30, 37, 51]. To enhance the efficiency of the U-Net architecture and accelerate inference speed, several model-driven methods, including B-Spline [32], Fourier-Net [21] and WiNet [12], have been proposed to learn a low-dimensional representation of displacement. The lightweight approach, NCA-Morph [33], utilizes Neural Cellular Automata to improve the efficiency of image registration. Another track of recent works tends to progressively estimate the large and complex deformations using either multiple cascades [15, 31, 52, 53] or pyramid methods [18, 23, 25, 26, 28, 46] compositions of multi-scale flows derived from multi-scale features. The cascaded methods and pyramid coarse-to-fine methods are capable of effectively improving registration accuracy for estimating large deformation. + +However, existing learning-based methods typically rely on networks that learn local feature correlations by applying shared convolution across all spatial dimensions within each layer. As illustrated in Fig. 1, deformations are often linked to the morphological characteristics of tissue in medical image registration, where voxels/features from various anatomical regions need to be paid different attention. However, plain convolutional methods utilize shared kernel weights uniformly across all regions, resulting in suboptimal estimation of deformation fields. Specifically, 3D medical images encapsulate complex intensity information relevant to anatomy, which makes it hard to capture spatially varying information in non-local regions of feature maps via spatially-shared convolution kernels. This key challenge is effectively incorporating spatially adaptive information, which can be addressed using region-based adaptive convolution techniques. Inspired by [16], which explores content-adaptive convolution in 2D image pansharpening, we extend this approach to 3D for medical image registration. This adaptation improves spatial awareness by modifying convolutional kernels based on regional content, thereby enhancing registration accuracy. + +In this paper, we propose a novel 3D pyramid medical image registration network, named SACB-Net, which in + +tegrates spatial-awareness convolution block (SACB) with a similarity matching module to form a multi-scale flow estimator. The proposed SACB employs the unsupervised KMeans to cluster the learned feature maps. Each feature clustering is believed to capture a specific intensity or structural pattern of the feature maps and then subsequently used to learn the spatial-aware adaptive convolution kernels. The adaptive kernel generation in SACB allows the network to effectively learn refined feature representations for more accurate flow estimation. We summarize our main contributions as follows: + +- We consider the spatial variance between two input images in the context of image registration, which is overlooked by vanilla convolutional-based approaches. The adaptive convolutions in our SACB-Net consider the changing of spatial regions to enhance registration accuracy. To the best of our knowledge, our method is the first work that employs 3D spatial-awareness convolution in 3D medical registration. +- The proposed SACB aims to boost the moving and fixed feature maps using adaptive convolution kernels for different spatial regions, where the spatial regions are divided based on the clustering of features, and the convolution kernels are learned from the centroids of each feature clustering. +- We propose an innovative pyramid flow estimator that integrates SACB and similarity matching module to perform multi-scale flow estimation, which can not only estimate local small deformation but also handle large deformation. +- Experimental results on atlas-based and inter-subject registration using two brain datasets (IXI and LPBA) demonstrate the effectiveness of SACB-Net. Further evaluation on inter-subject abdominal CT registration shows improved performance in handling large deformations. + +# 2. Related Works + +# 2.1. U-Net-based Registration Method + +VoxelMorph [4, 5], a general framework for learning-based image registration, demonstrates the effectiveness of unsupervised learning in pairwise image registration using a simplified U-Net architecture. Building on this foundation, recent methods have enhanced registration performance by incorporating stronger constraints on the deformation field, such as inverse/cycle consistency [24, 50], gradient inverse consistency [43], and diffeomorphism constraints [13, 14, 29]. Another track of methods tried to capture long-range dependencies of feature correlations using either large kernel convolution [20] or vision transformer [8, 10, 17, 30, 37, 51]. TransMorph [10] incorporates Swin transformer blocks in the encoder, improving the network's capacity to capture long-range dependencies, though this + +comes with increased computational cost. LKU [20], on the other hand, employs a large kernel convolutional encoder to model long-range relationships with fewer parameters and reduced computational load. NCA-Morph [33] employs Neural Cellular Automata mechanisms, drawing inspiration from biological processes, to improve registration efficiency. While U-Net-based methods have shown comparable performance for small deformations, they struggle to accurately estimate large deformations or register images that are not well pre-aligned. + +# 2.2. Model-driven Registration Method + +To reduce redundant convolution operations in the U-Net decoder and enhance training and inference efficiency, model-driven approaches such as B-Spline [32] and Fourier-Net [21] focus on learning a low-dimensional representation of the displacement field. However, displacement fields are not inherently smooth, and as a result, methods that rely solely on low-frequency components [21] or interpolate displacement from regular control points [32] may fail to fully and accurately capture the entire displacement field. To address this limitation, WiNet [12] introduces a wavelet-based decoder to preserve high-frequency information in the deformation field while maintaining network efficiency. Nonetheless, similar to U-Net-based methods, model-driven approaches also face challenges in effectively estimating large and complex deformations. + +# 2.3. Cascaded and Pyramid-based Method + +Recent works that estimate large deformation can be either network cascade [15, 31, 52, 53] or coarse-to-fine pyramid [18, 23, 25, 26, 28, 46]. LapIRN [31] utilizes a Laplacian pyramid network to capture large deformations by cascading registered flows at three different scales. PRNet++ [23] utilizes a dual-stream pyramid architecture for coarse-to-fine flow estimation, performing sequential warping on multi-scale feature maps and incorporating local 3D correlation layers. Im2grid [25] embedded the coordinate information on extracted multi-scale feature maps and estimated the scale flow by calculating the matching score on the neighborhood feature. ModeT [46] leverages neighbourhood attention mechanisms to estimate multi-head, multiscale sub-deformation flows from pyramid feature maps. It then applies weighting modules to fuse these multi-head flows at each scale, generating the final deformation by sequential composition. To solve the insufficient decomposition problems, RDN [18] compose the multi-scale flows by performing level-wise recursion at each resolution level and stage-wise recursion at feature pyramid level. + +# 2.4. Adaptive Convolution + +Unlike standard convolution operators that use shared convolution kernels, adaptive convolution operators apply dif + +ferent kernels based on variations in the input, providing enhanced feature extraction capabilities and greater flexibility. Following [19], recent methods have applied adaptive convolution, demonstrating its effectiveness in various 2D computer vision tasks, such as image upsampling [39], classification [11, 54], segmentation [9] and remote sensing pansharpening [16, 22]. However, spatially adaptive convolution has not yet been explored in 3D medical image registration tasks. + +# 3. Method + +# 3.1. Network Overview + +Given a pair of moving image $I_{m}$ and fixed image $I_{f}$ over the spatial domain $\Omega \subset \mathcal{R}^3$ , the SACB-Net $(I_m, I_f; \pmb{\theta})$ network aims to learn the optimal parameters $\pmb{\theta}_{*}$ that estimate a deformation field $\phi_{m \to f}: \mathcal{R}^3 \to \mathcal{R}^3$ , mapping the corresponding coordinates between the paired images. In unsupervised setting, the loss function contains $\mathcal{L}_{sim}(\cdot)$ measuring the similarity between the warped moving image $I_{m} \circ \phi$ and $I_{f}$ , and the regularization term. In our experiments, the smoothness term $\mathcal{L}_{reg}(\phi) = \| \nabla \phi(\pmb{\theta}) \|_2^2$ is computed using the $L_{2}$ -norm of the gradient of the deformation field. The objective function of our network is + +$$ +\mathcal {L} (\boldsymbol {\theta}) = \mathcal {L} _ {s i m} \left(I _ {m} \circ \left(\phi (\boldsymbol {\theta}) + \operatorname {I d}\right), I _ {f}\right) + \lambda \mathcal {L} _ {r e g} \tag {1} +$$ + +where $\theta$ denotes the learnable network parameters, $\circ$ represents the warping operator, $\mathrm{Id}$ is the identity grid, $\nabla$ refers to the first-order gradient, and $\lambda$ is the weight of the regularization term. We employ Normalized Cross-Correlation (NCC) to measure the similarity. + +As illustrated in Fig. 2, SACB-Net consists of a shared encoder and a pyramid flow estimator. The shared encoder is designed to extract multi-scale features $\{F_m^i\}$ and $\{F_f^i\}$ for the moving image $I_{m}$ and the fixed image $I_{f}$ , respectively, in a pyramid structure. Here, $i \in \{1,2,3,4,5\}$ represents the scale of the features, with resolutions of $\left[\frac{D}{2^{i - 1}},\frac{H}{2^{i - 1}},\frac{W}{2^{i - 1}}\right]$ for the depth, height and width. The shared encoder consists of five 3D convolutional layers and four down-sampling layers. The convolutional layer is composed of 3D convolution operation, instance normalization [44] and LeakyReLU activations [27] with the 0.1 slope. Each convolutional layer is followed by an average pooling operation with a kernel size of 2. The extracted features $\{F_m^i\}$ and $\{F_f^i\}$ are then input into our pyramid flow estimators, where they are used to perform flow estimation and composition as outlined in Sections 3.2 and 3.3. + +# 3.2. 3D Spatial-awareness Convolution Block + +To effectively extract region-based context information for flow estimation, the proposed 3D SAC incorporates prior knowledge of spatial context through a spatial context estimation module and adaptive kernel generator. First, the + +![](images/fd194211dc1b162d63785852eac2971782c5ea01fe74df893c0e57d5eb87a122.jpg) +Figure 2. Illustration of a 5-level pyramid SACB-Net. SACB-Net includes a shared encoder that extracts multi-scale feature maps $\{F_m^i\}$ and $\{F_f^i\}$ for the moving image $I_{m}$ and the fixed image $I_{f}$ , as well as pyramid flow estimators at each scale. At the lowest level, the flow estimator learns deformation $(\varphi_5)$ from the extracted moving and fixed image features $(F_m^5$ and $F_{f}^{5})$ . The following flow estimators take the level-wise features and the output deformation from its preceding level to compose the deformation. Each flow estimator includes a Spatial-Awareness Convolution Block (SACB) to enhance spatially adaptive feature representation, along with a similarity matching module for flow estimation. + +spatial context estimation module identifies spatial regions of interest, producing spatial contextual information. This contextual data is subsequently utilized by the adaptive kernel generator to dynamically produce convolutional kernels tailored to capture non-local, region-specific information. + +Spatial Context Estimation. After obtaining the multiscale features $\{F_m^i\}$ and $\{F_f^i\}$ , prior spatial region information is incorporated to refine these features. To achieve this, we employ unsupervised clustering method K-Means to estimate spatial region for feature maps. The reason not to use label information is due to potential inconsistencies in the labeling of $I_{m}$ and $I_{f}$ within the feature space, as well as to address the challenge of limited label availability. Let $\mathbf{F} \in \mathbb{R}^{C_{in} \times D \times H \times W}$ denote the input feature map, where $C_{in}$ represents the number of input channels, and $D$ , $H$ , and $W$ correspond to the depth, height, and width of the spatial dimensions, respectively. As shown in Fig. 3, first, the feature map $\mathbf{F}$ is unfolded into local patches to capture neighborhood information using sliding windows of size $k$ . A spatial mean is then applied to each patch to reduce its spatial dimensions. For each voxel $V_{(d,h,w)}$ located at coordinates $(d,h,w)$ within $\mathbf{F}$ , suppose $p(\cdot)$ is the unfolding operation, the corresponding unfolded patch is defined as: + +$$ +p _ {\left(V _ {\left(d, h, w\right)}\right)} = \left\{V _ {\left(d + d ^ {\prime}, h + h ^ {\prime}, w + w ^ {\prime}\right)} \mid d ^ {\prime}, h ^ {\prime}, w ^ {\prime} \in \left[ - \frac {k}{2}, \frac {k}{2} \right] \right\}. \tag {2} +$$ + +Here, $p_{(V_{(d,h,w)})}$ represents a patch of voxels centered at $V_{(d,h,w)}$ , with offsets $d', h', w'$ that range from $-\frac{k}{2}$ to $\frac{k}{2}$ (assuming $k$ is odd). The unfolded feature denotes as $p(F) \in \mathbb{R}^{C_{in} \times (D \times H \times W) \times k^3}$ . The spatial mean of each local patch $\overline{p}_{(V_{(d,h,w)})}$ based on the neighborhood voxel is calculated as: + +$$ +\bar {p} _ {\left(V _ {(d, h, w)}\right)} = \frac {1}{k ^ {3}} \sum_ {d ^ {\prime} = - \lfloor \frac {k}{2} \rfloor} ^ {\lfloor \frac {k}{2} \rfloor} \sum_ {h ^ {\prime} = - \lfloor \frac {k}{2} \rfloor} ^ {\lfloor \frac {k}{2} \rfloor} \sum_ {w ^ {\prime} = - \lfloor \frac {k}{2} \rfloor} ^ {\lfloor \frac {k}{2} \rfloor} p _ {\left(V _ {(d, h, w)}\right)}. \tag {3} +$$ + +Afterward, the preprocessed feature $\mathbf{F}$ is reshaped to dimensions $[D\times H\times W,C_{in}]$ to estimate a cluster index matrix $S\in \mathbb{Z}^{D\times H\times W}$ , which assigns each voxel to a cluster $S_{n} = \{(d,w,h)|S_{(d,w,h)} = n\};n\in \{1,\dots ,N\}$ . Here, $N$ represents the maximum number of clusters used for GPU-based KMeans clustering. Additionally, the centroid voxel of each cluster $S_{n}^{c}\in \mathbb{R}^{C_{in}}$ is computed as + +$$ +S _ {n} ^ {c} = \frac {1}{\left| S _ {n} \right|} \sum_ {(d, w, h) \in S _ {n}} V _ {(d, w, h)}, \tag {4} +$$ + +where $|S_{n}|$ denotes the total number of voxels belonging to the $n^{\mathrm{th}}$ spatial region. The cluster index matrix $S$ and the centroid voxels $S^{c}$ for each cluster are then utilized to generate adaptive kernel weights for convolution. + +Adaptive Kernel Generator. To adaptively generate a spatial convolution kernel that incorporates region-based information, we employ a multi-layer perception to model the + +![](images/c1f954c9bea66e0b0eb0090554b79cbb10551f766c38721260705c96384710ce.jpg) +Figure 3. Architecture of the 3D spatial-awareness convolution block (example in three clusters). This block aims to refine the input feature $\mathbf{F}$ to $\hat{\mathbf{F}}$ by adaptive convolution learned from spatial feature clustering. SACB consists of three parts: 1) the spatial context estimation module employs KMeans to cluster similar spatial features on the patched unfolding features, in which the features belonging to the same cluster centroid $S_{n}^{c}$ will be indicated with a cluster index map $S_{n}$ . 2) the adaptive kernel generator leverages each cluster centroid to generate cluster-specific spatial weights and bias via two MLPs, the resulting weights and bias will be used to form the final spatial adaptive convolution kernel, which will be imposed on the cluster-indexed unfolding features, again highlighting the spatial-awareness. 3) the original features and spatial-aware features are residually connected to form the final refined $\hat{\mathbf{F}}$ . + +mapping $\mathcal{F}_w$ , which produces an adaptive convolution kernel $\mathbf{W}_n \in \mathbb{R}^{C_{out} \times C_{in} \times k^3}$ for the $n^{\text{th}}$ spatial region based on $S_n^c$ . To simplify the process, we first initialize a global learnable convolution kernel $\mathbf{W} \in \mathbb{R}^{C_{out} \times C_{in} \times k^3}$ and then use the mapping function $\mathcal{F}_w$ to generate a spatial weight that can be applied to the global kernel $\mathbf{W}$ . Thus, $\mathbf{W}_n$ is obtained as + +$$ +\mathbf {W} _ {n} = \mathcal {F} _ {w} \left(S _ {n} ^ {c}\right) \odot \mathbf {W}, \tag {5} +$$ + +where $\odot$ denotes element-wise dot product. Then, a mapping function $\mathcal{F}_b$ using a multi-layer perception is applied to compute the convolution bias via obtained centroid $S_{n}^{c}$ . Consequently, the region-based spatially adaptive convolution is represented as: + +$$ +S A C \left(\mathbf {F} _ {(d, w, h)}\right) = p _ {\left(V _ {(d, h, w)}\right)} \otimes \mathbf {W} _ {n} + \mathcal {F} _ {b} \left(S _ {n} ^ {c}\right), \tag {6} +$$ + +where $p(\cdot)$ denotes the unfolding operation outlined in the spatial context estimation section. Subsequently, the process of applying the SAC block to produce refined features is as follows + +$$ +\hat {\mathbf {F}} = S A C B (\mathbf {F}) = \mathbf {F} + \sigma (S A C (\mathbf {F})), \tag {7} +$$ + +where $\sigma (\cdot)$ denotes activation function. + +# 3.3. Pyramid Flow Estimator + +Let $\hat{F_m}^i$ and $\hat{F_f}^i$ represent the multi-scale features refined by the proposed SACB at the $i^{th}$ level, the sub-deformation + +flows at each scale are then estimated based on the matching score using the refined features. Specifically, we calculate the similarity matching score between each voxel in the fixed image features and corresponding neighbouring voxels in the moving image features. Similar to [25, 40], this operation can be efficiently implemented as a matrix operation using the dot product. For instance, at the coarsest level $i = 5$ , the matching score can be computed as + +$$ +M _ {s i m} \left(\hat {F _ {f}} ^ {5}, \hat {F _ {m}} ^ {5}\right) = \operatorname {S o f t m a x} \left(\hat {F _ {f}} ^ {5} p \left(\hat {F _ {m}} ^ {5}\right) ^ {T}\right), \tag {8} +$$ + +where $p(\cdot)$ represents the unfolding operation as described in Section 3.2. For each voxel $V$ at $\hat{F_m}^5$ , the similarity matching score can be calculated as + +$$ +M _ {s i m} \left(\hat {F} _ {f} ^ {5} (V), \hat {F} _ {m} ^ {5} (V)\right) = \tag {9} +$$ + +$$ +\operatorname {S o f t m a x} \left(\left\{\left\langle \hat {F _ {f}} ^ {5} (V), \hat {F _ {m}} ^ {5} (V _ {c}) \right\rangle \right\} _ {V _ {c} \in \mathcal {N} (V)}\right), +$$ + +where $\langle \cdot, \cdot \rangle$ denotes the inner product, and $\mathcal{N}(V)$ represents the set of voxels within an $k \times k \times k$ neighborhood centered around voxel $V$ , $k$ represents the window size of the search region of $\hat{F}_m^5$ . We set $k = 3$ in our experiments. Then, the sub-deformation flow at the $5^{\text{th}}$ scale can be obtained as + +$$ +\varphi^ {5} = M _ {s i m} \left(\hat {F _ {f}} ^ {5}, \hat {F _ {m}} ^ {5}\right) G \tag {10} +$$ + +$G \in \mathbb{R}^{(k \times k \times k) \times 3}$ is a 3D grid that encodes the relative positions with respect to the centroid voxel location. + +At the $i = 4$ level, we first up scaling the flow $\varphi^5$ to form $\hat{\varphi}^5$ , then $\hat{F_m}^4$ is warped with $\hat{\varphi}^5$ . Sequentially, the $M_{sim}$ need to be calculated between $\hat{F_f}^4$ and $\hat{F_m}^4 \circ \hat{\varphi}^5$ for obtaining $\Delta \varphi^4$ . Afterward, the overall process can be formulated as: + +$$ +\left\{ \begin{array}{l} \hat {\varphi} ^ {i} = \mathrm {u p} _ {2 \times} \left(\varphi^ {i}\right), \\ \hat {F} _ {f} ^ {i - 1} = S A C B \left(F _ {f} ^ {i - 1}\right), \\ \hat {F} _ {m} ^ {i - 1} = S A C B \left(F _ {m} ^ {i - 1} \circ \hat {\varphi} ^ {i}\right), \quad i \in \{5, 4, 3, 2 \}. \\ \Delta \varphi^ {i - 1} = M _ {s i m} \left(\hat {F} _ {f} ^ {i - 1}, \hat {F} _ {m} ^ {i - 1}\right) G, \\ \varphi^ {i - 1} = \hat {\varphi} ^ {i} \circ \Delta \varphi^ {i - 1}, \end{array} \right. \tag {11} +$$ + +The $\mathrm{up}_{2\times}(\cdot)$ is flow rescaling operation with $2\times$ resolution. + +We note that the matching score calculation was not applied for flow estimation at the $1^{st}$ scale, as the search region with $k = 3$ can only determine a maximum displacement of 1 voxel using full-resolution features. However, the calculations at the $1^{st}$ scale significantly increase memory and computational costs, yielding limited benefit. Therefore, we used a simple approach with two convolutional layers for flow estimation at the final scale. + +# 4. Experiments + +Datasets. Our experiments use two publicly available 3D brain MRI datasets $\mathrm{IXI}^2$ and LPBA [36], along with one public abdominal CT dataset [47]. The preprocessed IXI dataset [10] includes 576 MRI scans $(160\times 192\times 224)$ from healthy subjects. We followed the same protocol as [10], partitioning the data into 403, 58, and 115 scans for training, validation, and testing sets, respectively. Atlas-based brain registration was conducted using a generated atlas from [24]. The LPBA dataset consists of 40 T1-weighted MRI volumes. Each MRI scan has been rigidly pre-aligned to the MNI305 template and manually annotated with 54 regions of interest (ROIs). We center-crop all volumes to dimensions of $160\times 192\times 160$ with voxel size $1\times 1\times 1mm^3$ . LPBA dataset was used to perform inter-subject registration with divided three subsets: 25 volumes $(25\times 24$ pairs) for training, 5 $(5\times 4$ pairs) for validation, and 10 $(10\times 9$ pairs) for testing. The Abdomen CT dataset consists of 30 abdominal scans with 13 manually labelled anatomical structures. We clamp the Hounsfield Units (HU) to $[-1000, 1000]$ and normalize each scan to $[0, 1]$ . Each scan is center-cropped from $192\times 160\times 256$ to $192\times 160\times 224$ with voxel size $2\times 2\times 2mm^3$ . The dataset is divided into 19 training volumes $(19\times 18$ pairs), 3 validation volumes $(3\times 2$ pairs) and 8 testing volumes $(8\times 7$ pairs) for inter-subject registration. + +Implementation Details. Our method is implemented using PyTorch. Both the training and testing phases are deployed on an A100 GPU with 40GB VRAM. All models + +are optimized using Adam followed with the same learning rate $10^{-4}$ and batch size 1. A fast GPU K-Means implementation is adopted. + +Evaluation Metrics. To qualify the registration performance, we use the Dice similarity coefficient (Dice) to evaluate the region of interest (ROI) segmentation overlap, the average symmetric surface distance (ASSD) [41] and the 95 percentile of the Hausdorff distance (HD95) to measure the registration accuracy of ROI structures. For evaluating the diffeomorphism of deformation fields, we use the percentage of negative values of the Jacobian determinant $(|J|_{<0\%})$ to calculate the percentage of folding voxels. + +Comparative Methods. We compare our method against state-of-the-art approaches, including the conventional SyN method [2], U-Net-based models such as VoxelMorph (VM) [5], LKU [20], and TransMorph [10], model-driven methods like B-Spline-Diff [32] and Fourier-Net [21], the lightweight NCA-Morph [33], the cascaded method LapIRN [31], and four pyramid-based methods: PR++ [23], Im2Grid [25], RDN [18] and ModeT [46]. All models were trained using their officially released code, with optimal hyperparameters tuned on the validation sets. + +# 5. Results + +In this section, we evaluate our SACB-Net against comparative methods on three tasks: atlas-based brain registration using the IXI dataset, inter-subject brain registration using the LPBA dataset, and large deformation estimation for inter-subject abdominal CT registration. + +# 5.1. Atlas-based Brain Registration + +Table 1 presents the quantitative results of atlas-based brain registration for 30 anatomical structures evaluated on the IXI dataset. It can be seen that SACB-Net outperforms all comparable methods in terms of Dice score while achieving plausible deformations and third-best results on HD95 and ASSD. Compared to the best-performing U-Net-based method (LKU), our SACB-Net achieves a $0.4\%$ improvement in Dice score while using 0.98M fewer parameters. In terms of model-driven methods, SACB-Net surpasses Fourier-Net by a margin of $0.6\%$ in the Dice score and uses 3.09M fewer parameters. For cascade and pyramid-based methods, those methods have a similar number of parameters apart from RDN. Our SACB-Net outperforms the second-best cascade method LapIRN with $0.6\%$ Dice, 0.038mm HD95, and 0.019mm ASSD, respectively. + +# 5.2. Inter-subject Brain Registration + +The inter-subject registration of the LPBA dataset tends to have a slightly larger deformation than the atlas-based registration on the IXI dataset. As shown in Table 1, most UNet + +Table 1. Registration performance comparison on the IXI and LPBA datasets. + +
MethodIXI (30 ROIs)LPBA (54 ROIs)#Param
Dice↑HD95↓ASSD↓|J|<0%↓Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.386±0.1956.479±0.6662.445±0.280-0.525±0.0478.039±0.8612.586±0.350--
SyN [3]0.645±0.1526.394±1.0481.551±0.286<0.00010.707±0.0166.254±0.4441.479±0.131<0.0001-
VM-1 [5]0.729±0.1293.798±0.7570.937±0.1621.590±0.3390.664±0.0256.873±0.6541.717±0.2000.649±0.2610.27M
VM-2 [5]0.732±0.1233.723±0.6800.926±0.1581.522±0.3360.669±0.0256.847±0.6591.698±0.2000.591±0.2420.30M
NCA-Morph [33]0.753±0.1363.109±0.5250.796±0.1210.506±0.1900.679±0.0236.666±0.6341.631±0.1880.264±0.1210.37M
TransMorph [10]0.754±0.1243.543±0.7210.862±0.1681.579±0.3280.695±0.0226.564±0.6191.559±0.1820.474±0.17646.77M
LKU [20]0.765±0.1292.967±0.4940.757±0.1140.109±0.0540.706±0.0326.452±0.9511.603±0.2500.594±0.2032.09M
B-Spline-Diff [32]0.742±0.1283.256±0.5380.832±0.117<0.00010.665±0.0236.792±0.6221.713±0.1860.0±0.00.27M
Fourier-Net [21]0.763±0.1292.857±0.4560.748±0.1140.024±0.0190.672±0.0226.716±0.6011.666±0.1800.216±0.1044.20M
LapIRN [31]0.763±0.1333.166±0.6080.779±0.1260.312±0.1060.716±0.0166.116±0.4541.426±0.1330.024±0.0091.20M
PRNet++ [23]0.755±0.1303.593±0.7480.857±0.1621.052±0.3020.701±0.0216.492±0.5971.520±0.1770.072±0.0271.24M
ModeT [46]0.758±0.1253.496±0.7320.828±0.1510.114±0.0570.721±0.0135.969±0.4161.375±0.1100.010±0.0041.03M
Im2Grid [25]0.761±0.1273.316±0.6680.799±0.128<0.00020.713±0.0146.062±0.4281.419±0.1180.007±0.0030.89M
RDN [18]0.759±0.1233.476±0.8020.823±0.161<0.00010.713±0.0176.208±0.4971.436±0.142<0.000228.65M
Ours0.769±0.1273.128±0.6310.760±0.1250.083±0.0450.731±0.0125.862±0.4361.326±0.1140.018±0.0061.11M
+ +based and model-driven methods do not perform as well as cascade and pyramid-based methods. Only LKU surpasses the pyramid-based method PR++, achieving improvements of $0.5\%$ in Dice, $0.04 \mathrm{~mm}$ in HD95 and $0.083 \mathrm{~mm}$ in ASSD, respectively. Our SACB outperforms all comparable methods in Dice, HD95 and ASSD with plausible deformations. Compared to the second-best method, ModeT, SACB achieves improvements of $1\%$ in Dice, $0.107 \mathrm{~mm}$ in HD95, and $0.049 \mathrm{~mm}$ in ASSD, respectively. The visualization in Fig. 4 shows that our SACB-Net captures more detailed correspondences in the registration process, as highlighted in the red box. + +# 5.3. Inter-subject Abdomen Registration + +The primary challenge in abdominal CT registration tasks is caused by the substantial differences in the spatial distribution of organs across scans, making it difficult to accurately estimate large deformations during the registration process. Table 2 demonstrates the numerical results of the inter-subject abdomen registration task. Our results show that our method outperforms all compared approaches in terms of Dice, HD95, and ASSD. Notably, U-Net-based and model-driven methods perform inferior to pyramid and cascaded approaches for estimating large deformations. Even the best-performing method in such approaches, TransMorph, still exhibits a significant gap compared to our SACB-Net with $14.4\%$ difference in Dice score, $5.934\mathrm{mm}$ in HD95 and $3.153\mathrm{mm}$ in ASSD, respectively. Compared to the second-best pyramid method, ModeT, SACB-Net achieves improvements of $3.8\%$ in Dice, $2.098\mathrm{mm}$ in HD95, and $0.781\mathrm{mm}$ in ASSD, respectively. + +The visualization of the wrapped moving images and labels, as shown in Fig. 4, illustrates that most parts of the kidneys are registered with our method. However, Trans + +Table 2. Registration performance comparison on Abdomen CT. + +
MethodAbdomen CT (13 ROIs)
Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.305±0.04727.790±4.86810.869±2.075-
VM-2 [5]0.406±0.05825.226±4.9428.890±2.0663.882±1.512
NCA-Morph [33]0.406±0.06125.959±4.9829.083±2.2235.197±1.166
TransMorph [10]0.444±0.06524.187±4.9558.156±2.0656.045±1.750
LKU [20]0.423±0.05824.252±4.6048.395±1.9464.316±1.778
B-Spline-Diff [32]0.375±0.05325.338±4.8669.217±1.9600.0±0.0
Fourier-Net [21]0.417±0.04823.084±4.2018.000±1.5322.167±2.391
LapIRN [31]0.522±0.05720.728±4.3186.274±1.3793.203±0.848
PRNet++ [23]0.478±0.06023.791±4.6917.524±1.7712.252±1.192
ModeT [46]0.550±0.05420.351±4.2315.784±1.2362.239±0.906
Im2Grid [25]0.530±0.04720.763±3.7866.028±1.1462.421±0.653
RDN [18]0.527±0.05521.782±4.7226.461±1.6131.529±0.989
Ours0.588±0.04918.253±3.6105.003±1.0113.930±1.416
+ +Morph, $\mathrm{PR} + +$ , and RDN failed to register the kidneys, leaving a large portion missing on the warped mask. + +# 6. Ablation Study + +First, we note that $\overline{p}(\cdot)$ in Eq. (3) can also be defined on the channel dimension or both the spatial+channel (Mix) dimension. As shown in Table 4, using the spatial mean of unfolded patches achieves slightly better performance compared to the channel mean and the Mix setting. + +We then determine the optimal number of scales at which SACB should be applied and the ideal number of clusters to achieve accurate registration performance. As shown in Table 3, the optimal configuration is applying SACB from scale $5$ to scale $2$ . Additionally, increasing the $N$ from 5 to 7 boosts the performance while a too-large value slightly decreases performance (e.g. 11). + +![](images/6657bc9b803d28710cf20fc911dc5db93f9c4665709293a01646f168d4228e48.jpg) +Figure 4. Visual comparisons on the Brain LPBA dataset (top two rows) and the Abdomen CT dataset (bottom two rows). Columns 3-10 display warped moving images (row 1 and 3), displacement fields as RGB images (row 2) and warped segmentation masks (row 4). + +Table 3. Ablation studies on the use of SACB at different scale levels in SACB-Net and varying numbers of spatial clusters $(N)$ , along with the corresponding average training time (in seconds), including data loading time, and GPU training memory usage (in MiB). The SACB configurations are based on the best results shown in Table 4. + +
NWith SACBIXILPBAAbdomen CT
Scale5Scale4Scale3Scale2Dice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTraining
-XXXX0.76430.059158380.7620.71410.020115740.5450.53742.907158380.764
5✓XXX0.76680.072158800.7990.72170.087116640.5790.54443.243158800.797
5✓✓XX0.76790.076161040.8790.72410.104117700.6310.54982.619161040.864
5✓✓✓X0.76710.060169600.9450.72660.076124180.7140.56852.781169600.950
5✓✓✓✓0.76830.086216381.2550.72940.017157480.9180.58493.548216381.234
7✓✓✓✓0.76910.083216381.3760.73090.018157781.0840.58813.930216621.360
9✓✓✓✓0.76850.069216801.5500.73040.016157921.1100.58754.405216801.562
11✓✓✓✓0.76840.084217321.6790.73000.015158061.1930.58754.419217181.728
+ +Table 4. Ablations of the SACB configurations. After padding and unfolding on the feature map $\mathbf{F}$ , the cluster maps can be calculated based on spatial mean on the local unfolded patch, channel mean and Mix (spatial+channel). In this study, we set $N = 5$ and apply SACBs from scale $_5$ to scale $_2$ . + +
SACBLPBA
SpatialChannelMixDice ↑|J|<0% ↓
✓✗✗0.72940.017
✗✓✗0.72810.020
✗✗✓0.72850.021
+ +# 7. Conclusion + +We introduced SACB-Net, which incorporates a 3D Spatial Awareness Convolution Block (SACB) that adaptively generates spatial convolution kernels based on the feature + +clusters, thereby improving the network's ability to capture spatially varying information. Additionally, our proposed flow estimators combine SACB with a similarity-matching mechanism to estimate displacements in a pyramid coarse-to-fine manner. Results on two brain datasets and one abdomen dataset demonstrate that SACB-Net effectively handles both small local and large deformations. Furthermore, the proposed flow estimator is a plug-in module that can be applied in other networks for better feature representation learning. + +Acknowledgements. The research was conducted using the Baskerville Tier 2 HPC service, which was funded by the EPSRC and UKRI through the World Class Labs scheme (EP/T022221/1) and the Digital Research Infrastructure programme (EP/W032244/1). Baskerville is operated by Advanced Research Computing at the University of Birmingham. + +# References + +[1] John Ashburner. A fast diffeomorphic image registration algorithm. Neuroimage, 38(1):95-113, 2007. 1 +[2] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal, 12(1):26-41, 2008. 1, 6 +[3] Brian B Avants, Nicholas J Tustison, Gang Song, Philip A Cook, Arno Klein, and James C Gee. A reproducible evaluation of ants similarity metric performance in brain image registration. Neuroimage, 54(3):2033-2044, 2011. 7 +[4] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9252-9260, 2018. 2 +[5] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 2, 6, 7 +[6] M Faisal Beg, Michael I Miller, Alain Trouve, and Laurent Younes. Computing large deformation metric mappings via geodesic flows of diffeomorphisms. International journal of computer vision, 61:139-157, 2005. 1 +[7] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Qian Wang, Pew-Thian Yap, and Dinggang Shen. Deformable image registration using a cue-aware deep regression network. IEEE Transactions on Biomedical Engineering, 65(9):1900-1911, 2018. 2 +[8] Junyu Chen, Yufan He, Eric C Frey, Ye Li, and Yong Du. Vit-v-net: Vision transformer for unsupervised volumetric medical image registration. arXiv preprint arXiv:2104.06468, 2021. 2 +[9] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8064-8073, 2021. 3 +[10] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 2, 6, 7 +[11] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11030-11039, 2020. 3 +[12] Xinxing Cheng, Xi Jia, Wenqi Lu, Qiufu Li, Linlin Shen, Alexander Krull, and Jinming Duan. Winet: Wavelet-based incremental learning for efficient medical image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 761-771. Springer, 2024. 2, 3 +[13] Adrian Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces. Medical image analysis, 57:226-236, 2019. 2 + +[14] Adrian V Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning for fast probabilistic diffeomorphic registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 729-738. Springer, 2018. 2 +[15] Bob D De Vos, Floris F Berendsen, Max A Viergever, Hessam Sokooti, Marius Staring, and Ivana Isgum. A deep learning framework for unsupervised affine and deformable image registration. Medical image analysis, 52:128-143, 2019. 2, 3 +[16] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 2, 3 +[17] Morteza Ghahremani, Mohammad Khateri, Bailiang Jian, Benedikt Wiestler, Ehsan Adeli, and Christian Wachinger. H-vit: A hierarchical vision transformer for deformable image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11513-11523, 2024. 2 +[18] Bo Hu, Shenglong Zhou, Zhiwei Xiong, and Feng Wu. Recursive decomposition network for deformable image registration. IEEE Journal of Biomedical and Health Informatics, 26(10):5130-5141, 2022. 2, 3, 6, 7 +[19] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 3 +[20] Xi Jia, Joseph Bartlett, Tianyang Zhang, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. U-net vs transformer: Is u-net outdated in medical image registration? In International Workshop on Machine Learning in Medical Imaging, pages 151-160. Springer, 2022. 2, 3, 6, 7 +[21] Xi Jia, Joseph Bartlett, Wei Chen, Siyang Song, Tianyang Zhang, Xinxing Cheng, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. Fourier-net: Fast image registration with band-limited deformation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1015-1023, 2023. 2, 3, 6, 7 +[22] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI conference on artificial intelligence, pages 1113-1121, 2022. 3 +[23] Miao Kang, Xiaojun Hu, Weilin Huang, Matthew R Scott, and Mauricio Reyes. Dual-stream pyramid registration network. Medical image analysis, 78:102379, 2022. 2, 3, 6, 7 +[24] Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. Cycle morph: cycle consistent unsupervised deformable image registration. Medical image analysis, 71:102036, 2021. 2, 6 +[25] Yihao Liu, Lianrui Zuo, Shuo Han, Yuan Xue, Jerry L Prince, and Aaron Carass. Coordinate translator for learning deformable medical image registration. In International workshop on multiscale multimodal medical imaging, pages 98-109. Springer, 2022. 2, 3, 5, 6, 7 +[26] Tai Ma, Suwei Zhang, Jiafeng Li, and Ying Wen. Iirp-net: Iterative inference residual pyramid network for enhanced + +image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11555, 2024. 2, 3 +[27] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 3 +[28] Mingyuan Meng, Dagan Feng, Lei Bi, and Jinman Kim. Correlation-aware coarse-to-fine mlp's for deformable medical image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9645-9654, 2024. 2, 3 +[29] Tony CW Mok and Albert Chung. Fast symmetric diffeomorphic image registration with convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4644-4653, 2020. 2 +[30] Tony CW Mok and Albert Chung. Affine medical image registration with coarse-to-fine vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20835-20844, 2022. 2 +[31] Tony CW Mok and Albert CS Chung. Large deformation diffeomorphic image registration with laplacian pyramid networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 211-221. Springer, 2020. 2, 3, 6, 7 +[32] Huaqi Qiu, Chen Qin, Andreas Schuh, Kerstin Hammernik, and Daniel Rueckert. Learning diffeomorphic and modality-invariant registration using b-splines. In Medical Imaging with Deep Learning, 2021. 2, 3, 6, 7 +[33] Amin Ranem, John Kalkhof, and Anirban Mukhopadhyay. Nca-morph: Medical image registration with neural cellular automata. arXiv preprint arXiv:2410.22265, 2024. 2, 3, 6, 7 +[34] Marc-Michel Rohe, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: learning deformable image registration using shape matching. In Medical Image Computing and Computer Assisted Intervention- MIC-CAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part I 20, pages 266-274. Springer, 2017. 2 +[35] Yudi Sang, Xianglei Xing, Yingnian Wu, and Dan Ruan. Imposing implicit feasibility constraints on deformable image registration using a statistical generative model. Journal of Medical Imaging, 7(6):064005-064005, 2020. 2 +[36] David W Shattuck, Mubeena Mirza, Vitria Adisetiyo, Cornelius Hojatkashani, Georges Salamon, Katherine L Narr, Russell A Poldrack, Robert M Bilder, and Arthur W Toga. Construction of a 3d probabilistic atlas of human cortical structures. Neuroimage, 39(3):1064-1080, 2008. 6 +[37] Jiacheng Shi, Yuting He, Youyong Kong, Jean-Louis Coatrieux, Huazhong Shu, Guanyu Yang, and Shuo Li. Xmorph: Full transformer for deformable medical image registration via cross attention. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 217-226. Springer, 2022. 2 +[38] Aristeidis Sotiras, Christos Davatzikos, and Nikos Paragios. Deformable medical image registration: A survey. IEEE transactions on medical imaging, 32(7):1153-1190, 2013. 1 + +[39] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 3 +[40] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 5 +[41] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15:1-28, 2015. 6 +[42] Alexander Thorley, Xi Jia, Hyung Jin Chang, Boyang Liu, Karina Bunting, Victoria Stoll, Antonio de Marvao, Declan P O'Regan, Georgios Gkoutos, Dipak Kotecha, et al. Nesterov accelerated admm for fast diffeomorphic image registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 150-160. Springer, 2021. 1 +[43] Lin Tian, Hastings Greer, François-Xavier Vialard, Roland Kwitt, Raul San José Estépar, Richard Jarrett Rushmore, Nikolaos Makris, Sylvain Bouix, and Marc Niethammer. Gradicon: Approximate diffeomorphisms via gradient inverse consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18084-18094, 2023. 2 +[44] D Ulyanov. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016.3 +[45] Tom Vercauteren, Xavier Pennec, Aymeric Perchant, and Nicholas Ayache. Diffeomorphic demons: Efficient nonparametric image registration. NeuroImage, 45(1):S61-S72, 2009. 1 +[46] Haiqiao Wang and Yi Ni, Dongand Wang. Model: Learning deformable image registration via motion decomposition transformer. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2023, pages 740-749, 2023. 2, 3, 6, 7 +[47] Zhoubing Xu, Christopher P Lee, Mattias P Heinrich, Marc Modat, Daniel Rueckert, Sebastien Ourselin, Richard G Abramson, and Bennett A Landman. Evaluation of six registration methods for the human abdomen on clinically acquired ct. IEEE Transactions on Biomedical Engineering, 63(8):1563-1572, 2016. 6 +[48] Xiaofeng Yang, Ning Wu, Guanghui Cheng, Zhengyang Zhou, S Yu David, Jonathan J Beitler, Walter J Curran, and Tian Liu. Automated segmentation of the parotid gland based on atlas registration and machine learning: a longitudinal mri study in head-and-neck radiation therapy. International Journal of Radiation Oncology*Biology*Physics, 90 (5):1225-1233, 2014. 1 +[49] Xiao Yang, Roland Kwitt, Martin Styner, and Marc Niethammer. Quicksilver: Fast predictive image registration-a deep learning approach. NeuroImage, 158:378-396, 2017. 2 +[50] Jun Zhang. Inverse-consistent deep networks for unsu + +pervised deformable image registration. arXiv preprint arXiv:1809.03443, 2018. 2 +[51] Yungeng Zhang, Yuru Pei, and Hongbin Zha. Learning dual transformer network for diffeomorphic registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 129-138. Springer, 2021. 2 +[52] Shengyu Zhao, Yue Dong, Eric I Chang, Yan Xu, et al. Recursive cascaded networks for unsupervised medical image registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10600-10610, 2019. 2, 3 +[53] Shengyu Zhao, Tingfung Lau, Ji Luo, I Eric, Chao Chang, and Yan Xu. Unsupervised 3d end-to-end medical image registration with volume tweening network. IEEE journal of biomedical and health informatics, 24(5):1394-1404, 2019. 2, 3 +[54] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6647-6656, 2021. 3 \ No newline at end of file diff --git a/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/images.zip b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4e966e869b582c0f4676033321bc5469b113f1e --- /dev/null +++ b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3da605df700f45a3069356c0b2c013faa085fec3f1868abc396977bbb92ae213 +size 682999 diff --git a/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/layout.json b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1625b98071a128fbf168f8bc9dca72543c3d060b --- /dev/null +++ b/2025/SACB-Net_ Spatial-awareness Convolutions for Medical Image Registration/layout.json @@ -0,0 +1,9839 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 71, + 103, + 538, + 121 + ], + "type": "text", + "content": "SACB-Net: Spatial-awareness Convolutions for Medical Image Registration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "spans": [ + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "content": "Xinxing Cheng" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Tianyang Zhang" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Wenqi Lu" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Qingjie Meng" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Alejandro F. Frangi" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Jinming Duan" + }, + { + "bbox": [ + 147, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1,3,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 171, + 517, + 213 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 153, + 171, + 457, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 171, + 457, + 185 + ], + "spans": [ + { + "bbox": [ + 153, + 171, + 457, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 171, + 457, + 185 + ], + "type": "text", + "content": " School of Computer Science, University of Birmingham, UK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 185, + 517, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 185, + 517, + 199 + ], + "spans": [ + { + "bbox": [ + 93, + 185, + 517, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 185, + 517, + 199 + ], + "type": "text", + "content": " Department of Computing and Mathematics, Manchester Metropolitan University, UK" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 199, + 509, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 199, + 509, + 213 + ], + "spans": [ + { + "bbox": [ + 93, + 199, + 509, + 213 + ], + "type": "text", + "content": "3 Division of Informatics, Imaging and Data Sciences, University of Manchester, UK" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 75, + 213, + 535, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 213, + 535, + 242 + ], + "spans": [ + { + "bbox": [ + 75, + 213, + 535, + 242 + ], + "type": "text", + "content": "4 Centre for Computational Imaging and Modelling in Medicine, University of Manchester, UK \n5 Department of Computing, Imperial College London, UK" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 243, + 503, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 243, + 503, + 255 + ], + "spans": [ + { + "bbox": [ + 107, + 243, + 503, + 255 + ], + "type": "text", + "content": "{xxc142, txz009}@student.bham.ac.uk, W.Lu@mmu.ac.uk, m.qingjie@bham.ac.uk" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 258, + 477, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 258, + 477, + 269 + ], + "spans": [ + { + "bbox": [ + 129, + 258, + 477, + 269 + ], + "type": "text", + "content": "alejandro.frangi@manchester.ac.uk, jinming.duan@manchester.ac.uk" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 297, + 200, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 297, + 200, + 309 + ], + "spans": [ + { + "bbox": [ + 151, + 297, + 200, + 309 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 323, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 296, + 622 + ], + "type": "text", + "content": "Deep learning-based image registration methods have shown state-of-the-art performance and rapid inference speeds. Despite these advances, many existing approaches fall short in capturing spatially varying information in non-local regions of feature maps due to the reliance on spatially-shared convolution kernels. This limitation leads to suboptimal estimation of deformation fields. In this paper, we propose a 3D Spatial-Awareness Convolution Block (SACB) to enhance the spatial information within feature representations. Our SACB estimates the spatial clusters within feature maps by leveraging feature similarity and subsequently parameterizes the adaptive convolution kernels across diverse regions. This adaptive mechanism generates the convolution kernels (weights and biases) tailored to spatial variations, thereby enabling the network to effectively capture spatially varying information. Building on SACB, we introduce a pyramid flow estimator (named SACB-Net) that integrates SACBs to facilitate multi-scale flow composition, particularly addressing large deformations. Experimental results on the brain IXI and LPBA datasets as well as Abdomen CT datasets demonstrate the effectiveness of SACB and the superiority of SACB-Net over the state-of-the-art learning-based registration methods. The code is available at https://github.com/x-xc/SACB_Net." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 135, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 295, + 693 + ], + "type": "text", + "content": "Deformable image registration aims to align the corresponding anatomical structures between a moving and fixed" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 325, + 312, + 441, + 427 + ], + "blocks": [ + { + "bbox": [ + 325, + 312, + 441, + 427 + ], + "lines": [ + { + "bbox": [ + 325, + 312, + 441, + 427 + ], + "spans": [ + { + "bbox": [ + 325, + 312, + 441, + 427 + ], + "type": "image", + "image_path": "14b37902be4244111f335d78f924e127a49c1699ba8c0882e2d5dcea27f7853e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 430, + 407, + 437 + ], + "lines": [ + { + "bbox": [ + 348, + 430, + 407, + 437 + ], + "spans": [ + { + "bbox": [ + 348, + 430, + 407, + 437 + ], + "type": "text", + "content": "Vanilla Convolution" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 442, + 310, + 542, + 428 + ], + "blocks": [ + { + "bbox": [ + 374, + 295, + 504, + 305 + ], + "lines": [ + { + "bbox": [ + 374, + 295, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 374, + 295, + 504, + 305 + ], + "type": "text", + "content": "* Deformation is usually related to tissue shapes *" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 310, + 542, + 428 + ], + "lines": [ + { + "bbox": [ + 442, + 310, + 542, + 428 + ], + "spans": [ + { + "bbox": [ + 442, + 310, + 542, + 428 + ], + "type": "image", + "image_path": "26fe5989a3ba2b8110d8af6a587b535fbd415434a1709e850f053283625aea62.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 430, + 542, + 438 + ], + "lines": [ + { + "bbox": [ + 451, + 430, + 542, + 438 + ], + "spans": [ + { + "bbox": [ + 451, + 430, + 542, + 438 + ], + "type": "text", + "content": "Spatial-awareness Convolution" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 447, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 525 + ], + "type": "text", + "content": "Figure 1. Spatial-awareness Convolution. Since deformation is usually related to tissue shape in medical image registration, voxels/features from different regions should be given varying levels of attention. However, vanilla convolution methods apply shared kernel weights across all regions, leading to a suboptimal estimation of deformation fields. SAC mechanisms, on the other hand, apply distinct kernel attention weights for different spatial clusters." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 546, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 715 + ], + "type": "text", + "content": "image pair with the estimated dense non-linear deformation field, which plays a crucial role in various medical imaging tasks [38, 48]. In the past, conventional methods like LDDMM [6], DARTEL [1], SyN [2], Demons [45], and ADMM [42] are time-consuming and computationally intensive, primarily due to the instance-level (pair-wise) iterative optimization required. Moreover, such methods may involve sophisticated hyper-parameter tuning, limiting their applications in large-scale volumetric registration. To overcome these challenges, learning-based image registration approaches have demonstrated significant success in achieving competitive accuracy along with fast inference speed. Learning-based approaches can be divided into supervised and unsupervised categories. Supervised methods" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5227" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "[7, 34, 49] often rely on the high quality of the synthetic ground truth deformation field or the annotated map [35] to predict a deformation field. Therefore, unsupervised image registration is more popular for estimating an optimal deformation field of pairwise images by minimizing the distance between the warped moving image and the fixed image with a regularization term." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 296, + 422 + ], + "type": "text", + "content": "Following the classical unsupervised learning-based framework VoxelMorph [5], some U-Net-based methods focus on the performance improvements by adding stronger constraints over the deformation field such as inverse/cycle consistency [24, 50], gradient inverse consistency [43] and diffeomorphisms [13, 14, 29]. Some existing methods have introduced more advanced neural architectures, such as large-kernel network [20], vision transformers [8, 10, 17, 30, 37, 51]. To enhance the efficiency of the U-Net architecture and accelerate inference speed, several model-driven methods, including B-Spline [32], Fourier-Net [21] and WiNet [12], have been proposed to learn a low-dimensional representation of displacement. The lightweight approach, NCA-Morph [33], utilizes Neural Cellular Automata to improve the efficiency of image registration. Another track of recent works tends to progressively estimate the large and complex deformations using either multiple cascades [15, 31, 52, 53] or pyramid methods [18, 23, 25, 26, 28, 46] compositions of multi-scale flows derived from multi-scale features. The cascaded methods and pyramid coarse-to-fine methods are capable of effectively improving registration accuracy for estimating large deformation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 424, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 687 + ], + "type": "text", + "content": "However, existing learning-based methods typically rely on networks that learn local feature correlations by applying shared convolution across all spatial dimensions within each layer. As illustrated in Fig. 1, deformations are often linked to the morphological characteristics of tissue in medical image registration, where voxels/features from various anatomical regions need to be paid different attention. However, plain convolutional methods utilize shared kernel weights uniformly across all regions, resulting in suboptimal estimation of deformation fields. Specifically, 3D medical images encapsulate complex intensity information relevant to anatomy, which makes it hard to capture spatially varying information in non-local regions of feature maps via spatially-shared convolution kernels. This key challenge is effectively incorporating spatially adaptive information, which can be addressed using region-based adaptive convolution techniques. Inspired by [16], which explores content-adaptive convolution in 2D image pansharpening, we extend this approach to 3D for medical image registration. This adaptation improves spatial awareness by modifying convolutional kernels based on regional content, thereby enhancing registration accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel 3D pyramid medical image registration network, named SACB-Net, which in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "tegrates spatial-awareness convolution block (SACB) with a similarity matching module to form a multi-scale flow estimator. The proposed SACB employs the unsupervised KMeans to cluster the learned feature maps. Each feature clustering is believed to capture a specific intensity or structural pattern of the feature maps and then subsequently used to learn the spatial-aware adaptive convolution kernels. The adaptive kernel generation in SACB allows the network to effectively learn refined feature representations for more accurate flow estimation. We summarize our main contributions as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 205, + 553, + 492 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 205, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 553, + 300 + ], + "type": "text", + "content": "- We consider the spatial variance between two input images in the context of image registration, which is overlooked by vanilla convolutional-based approaches. The adaptive convolutions in our SACB-Net consider the changing of spatial regions to enhance registration accuracy. To the best of our knowledge, our method is the first work that employs 3D spatial-awareness convolution in 3D medical registration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 300, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 372 + ], + "type": "text", + "content": "- The proposed SACB aims to boost the moving and fixed feature maps using adaptive convolution kernels for different spatial regions, where the spatial regions are divided based on the clustering of features, and the convolution kernels are learned from the centroids of each feature clustering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 373, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 553, + 431 + ], + "type": "text", + "content": "- We propose an innovative pyramid flow estimator that integrates SACB and similarity matching module to perform multi-scale flow estimation, which can not only estimate local small deformation but also handle large deformation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 432, + 553, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 432, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 314, + 432, + 553, + 492 + ], + "type": "text", + "content": "- Experimental results on atlas-based and inter-subject registration using two brain datasets (IXI and LPBA) demonstrate the effectiveness of SACB-Net. Further evaluation on inter-subject abdominal CT registration shows improved performance in handling large deformations." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 506, + 405, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 405, + 518 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 405, + 518 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 527, + 495, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 527, + 495, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 527, + 495, + 540 + ], + "type": "text", + "content": "2.1. U-Net-based Registration Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "VoxelMorph [4, 5], a general framework for learning-based image registration, demonstrates the effectiveness of unsupervised learning in pairwise image registration using a simplified U-Net architecture. Building on this foundation, recent methods have enhanced registration performance by incorporating stronger constraints on the deformation field, such as inverse/cycle consistency [24, 50], gradient inverse consistency [43], and diffeomorphism constraints [13, 14, 29]. Another track of methods tried to capture long-range dependencies of feature correlations using either large kernel convolution [20] or vision transformer [8, 10, 17, 30, 37, 51]. TransMorph [10] incorporates Swin transformer blocks in the encoder, improving the network's capacity to capture long-range dependencies, though this" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5228" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": "comes with increased computational cost. LKU [20], on the other hand, employs a large kernel convolutional encoder to model long-range relationships with fewer parameters and reduced computational load. NCA-Morph [33] employs Neural Cellular Automata mechanisms, drawing inspiration from biological processes, to improve registration efficiency. While U-Net-based methods have shown comparable performance for small deformations, they struggle to accurately estimate large deformations or register images that are not well pre-aligned." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 241, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 241, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 241, + 213 + ], + "type": "text", + "content": "2.2. Model-driven Registration Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 397 + ], + "type": "text", + "content": "To reduce redundant convolution operations in the U-Net decoder and enhance training and inference efficiency, model-driven approaches such as B-Spline [32] and Fourier-Net [21] focus on learning a low-dimensional representation of the displacement field. However, displacement fields are not inherently smooth, and as a result, methods that rely solely on low-frequency components [21] or interpolate displacement from regular control points [32] may fail to fully and accurately capture the entire displacement field. To address this limitation, WiNet [12] introduces a wavelet-based decoder to preserve high-frequency information in the deformation field while maintaining network efficiency. Nonetheless, similar to U-Net-based methods, model-driven approaches also face challenges in effectively estimating large and complex deformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 257, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 257, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 257, + 418 + ], + "type": "text", + "content": "2.3. Cascaded and Pyramid-based Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 663 + ], + "type": "text", + "content": "Recent works that estimate large deformation can be either network cascade [15, 31, 52, 53] or coarse-to-fine pyramid [18, 23, 25, 26, 28, 46]. LapIRN [31] utilizes a Laplacian pyramid network to capture large deformations by cascading registered flows at three different scales. PRNet++ [23] utilizes a dual-stream pyramid architecture for coarse-to-fine flow estimation, performing sequential warping on multi-scale feature maps and incorporating local 3D correlation layers. Im2grid [25] embedded the coordinate information on extracted multi-scale feature maps and estimated the scale flow by calculating the matching score on the neighborhood feature. ModeT [46] leverages neighbourhood attention mechanisms to estimate multi-head, multiscale sub-deformation flows from pyramid feature maps. It then applies weighting modules to fuse these multi-head flows at each scale, generating the final deformation by sequential composition. To solve the insufficient decomposition problems, RDN [18] compose the multi-scale flows by performing level-wise recursion at each resolution level and stage-wise recursion at feature pyramid level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 180, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 180, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 180, + 685 + ], + "type": "text", + "content": "2.4. Adaptive Convolution" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Unlike standard convolution operators that use shared convolution kernels, adaptive convolution operators apply dif" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "ferent kernels based on variations in the input, providing enhanced feature extraction capabilities and greater flexibility. Following [19], recent methods have applied adaptive convolution, demonstrating its effectiveness in various 2D computer vision tasks, such as image upsampling [39], classification [11, 54], segmentation [9] and remote sensing pansharpening [16, 22]. However, spatially adaptive convolution has not yet been explored in 3D medical image registration tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 190, + 370, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 370, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 370, + 202 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 209, + 424, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 424, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 424, + 221 + ], + "type": "text", + "content": "3.1. Network Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": "Given a pair of moving image " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " and fixed image " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " over the spatial domain " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathcal{R}^3" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": ", the SACB-Net " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "(I_m, I_f; \\pmb{\\theta})" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " network aims to learn the optimal parameters " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{*}" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " that estimate a deformation field " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\phi_{m \\to f}: \\mathcal{R}^3 \\to \\mathcal{R}^3" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": ", mapping the corresponding coordinates between the paired images. In unsupervised setting, the loss function contains " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}(\\cdot)" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " measuring the similarity between the warped moving image " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "I_{m} \\circ \\phi" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": ", and the regularization term. In our experiments, the smoothness term " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}(\\phi) = \\| \\nabla \\phi(\\pmb{\\theta}) \\|_2^2" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": " is computed using the " + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 227, + 555, + 361 + ], + "type": "text", + "content": "-norm of the gradient of the deformation field. The objective function of our network is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 369, + 555, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 369, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 340, + 369, + 555, + 383 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\boldsymbol {\\theta}) = \\mathcal {L} _ {s i m} \\left(I _ {m} \\circ \\left(\\phi (\\boldsymbol {\\theta}) + \\operatorname {I d}\\right), I _ {f}\\right) + \\lambda \\mathcal {L} _ {r e g} \\tag {1}", + "image_path": "e4a665abf91059c204743dc7db99d8c3e6e07bbee01caf8fdaff77fd151b30c8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "content": " denotes the learnable network parameters, " + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "content": " represents the warping operator, " + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{Id}" + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "content": " is the identity grid, " + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "content": " refers to the first-order gradient, and " + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 390, + 554, + 449 + ], + "type": "text", + "content": " is the weight of the regularization term. We employ Normalized Cross-Correlation (NCC) to measure the similarity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": "As illustrated in Fig. 2, SACB-Net consists of a shared encoder and a pyramid flow estimator. The shared encoder is designed to extract multi-scale features " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\{F_m^i\\}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\{F_f^i\\}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " for the moving image " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " and the fixed image " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": ", respectively, in a pyramid structure. Here, " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,3,4,5\\}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " represents the scale of the features, with resolutions of " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\left[\\frac{D}{2^{i - 1}},\\frac{H}{2^{i - 1}},\\frac{W}{2^{i - 1}}\\right]" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " for the depth, height and width. The shared encoder consists of five 3D convolutional layers and four down-sampling layers. The convolutional layer is composed of 3D convolution operation, instance normalization [44] and LeakyReLU activations [27] with the 0.1 slope. Each convolutional layer is followed by an average pooling operation with a kernel size of 2. The extracted features " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\{F_m^i\\}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\{F_f^i\\}" + }, + { + "bbox": [ + 313, + 450, + 555, + 641 + ], + "type": "text", + "content": " are then input into our pyramid flow estimators, where they are used to perform flow estimation and composition as outlined in Sections 3.2 and 3.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 648, + 527, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 527, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 527, + 661 + ], + "type": "text", + "content": "3.2. 3D Spatial-awareness Convolution Block" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "To effectively extract region-based context information for flow estimation, the proposed 3D SAC incorporates prior knowledge of spatial context through a spatial context estimation module and adaptive kernel generator. First, the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5229" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 555, + 277 + ], + "type": "image", + "image_path": "fd194211dc1b162d63785852eac2971782c5ea01fe74df893c0e57d5eb87a122.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Illustration of a 5-level pyramid SACB-Net. SACB-Net includes a shared encoder that extracts multi-scale feature maps " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\{F_m^i\\}" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\{F_f^i\\}" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " for the moving image " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " and the fixed image " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": ", as well as pyramid flow estimators at each scale. At the lowest level, the flow estimator learns deformation " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "(\\varphi_5)" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " from the extracted moving and fixed image features " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "(F_m^5" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "inline_equation", + "content": "F_{f}^{5})" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": ". The following flow estimators take the level-wise features and the output deformation from its preceding level to compose the deformation. Each flow estimator includes a Spatial-Awareness Convolution Block (SACB) to enhance spatially adaptive feature representation, along with a similarity matching module for flow estimation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 371, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 295, + 431 + ], + "type": "text", + "content": "spatial context estimation module identifies spatial regions of interest, producing spatial contextual information. This contextual data is subsequently utilized by the adaptive kernel generator to dynamically produce convolutional kernels tailored to capture non-local, region-specific information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": "Spatial Context Estimation. After obtaining the multiscale features " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\{F_m^i\\}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\{F_f^i\\}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": ", prior spatial region information is incorporated to refine these features. To achieve this, we employ unsupervised clustering method K-Means to estimate spatial region for feature maps. The reason not to use label information is due to potential inconsistencies in the labeling of " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " within the feature space, as well as to address the challenge of limited label availability. Let " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{C_{in} \\times D \\times H \\times W}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " denote the input feature map, where " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " represents the number of input channels, and " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " correspond to the depth, height, and width of the spatial dimensions, respectively. As shown in Fig. 3, first, the feature map " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " is unfolded into local patches to capture neighborhood information using sliding windows of size " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": ". A spatial mean is then applied to each patch to reduce its spatial dimensions. For each voxel " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "V_{(d,h,w)}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " located at coordinates " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "(d,h,w)" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": ", suppose " + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 55, + 437, + 296, + 654 + ], + "type": "text", + "content": " is the unfolding operation, the corresponding unfolded patch is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 674, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 295, + 714 + ], + "type": "interline_equation", + "content": "p _ {\\left(V _ {\\left(d, h, w\\right)}\\right)} = \\left\\{V _ {\\left(d + d ^ {\\prime}, h + h ^ {\\prime}, w + w ^ {\\prime}\\right)} \\mid d ^ {\\prime}, h ^ {\\prime}, w ^ {\\prime} \\in \\left[ - \\frac {k}{2}, \\frac {k}{2} \\right] \\right\\}. \\tag {2}", + "image_path": "6e1ff7768dce67358f06cb5eedc7e9ed7ef2e752ec2b1b6d7284ea444b9f1ad5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "p_{(V_{(d,h,w)})}" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": " represents a patch of voxels centered at " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "V_{(d,h,w)}" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": ", with offsets " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "d', h', w'" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": " that range from " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "-\\frac{k}{2}" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\frac{k}{2}" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": " (assuming " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": " is odd). The unfolded feature denotes as " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "p(F) \\in \\mathbb{R}^{C_{in} \\times (D \\times H \\times W) \\times k^3}" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": ". The spatial mean of each local patch " + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\overline{p}_{(V_{(d,h,w)})}" + }, + { + "bbox": [ + 313, + 371, + 555, + 445 + ], + "type": "text", + "content": " based on the neighborhood voxel is calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 448, + 555, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 448, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 448, + 555, + 487 + ], + "type": "interline_equation", + "content": "\\bar {p} _ {\\left(V _ {(d, h, w)}\\right)} = \\frac {1}{k ^ {3}} \\sum_ {d ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} \\sum_ {h ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} \\sum_ {w ^ {\\prime} = - \\lfloor \\frac {k}{2} \\rfloor} ^ {\\lfloor \\frac {k}{2} \\rfloor} p _ {\\left(V _ {(d, h, w)}\\right)}. \\tag {3}", + "image_path": "9865fa4c867585c3f53e2db81fc7061f5663bd954af4e6e15d28c15580d28f70.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": "Afterward, the preprocessed feature " + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": " is reshaped to dimensions " + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "inline_equation", + "content": "[D\\times H\\times W,C_{in}]" + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": " to estimate a cluster index matrix " + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{Z}^{D\\times H\\times W}" + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": ", which assigns each voxel to a cluster " + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "inline_equation", + "content": "S_{n} = \\{(d,w,h)|S_{(d,w,h)} = n\\};n\\in \\{1,\\dots ,N\\}" + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": " represents the maximum number of clusters used for GPU-based KMeans clustering. Additionally, the centroid voxel of each cluster " + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "inline_equation", + "content": "S_{n}^{c}\\in \\mathbb{R}^{C_{in}}" + }, + { + "bbox": [ + 313, + 490, + 555, + 574 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 370, + 577, + 555, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 577, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 370, + 577, + 555, + 608 + ], + "type": "interline_equation", + "content": "S _ {n} ^ {c} = \\frac {1}{\\left| S _ {n} \\right|} \\sum_ {(d, w, h) \\in S _ {n}} V _ {(d, w, h)}, \\tag {4}", + "image_path": "02d644377b66825a56848eaad4116ef8bf004d55d46944b7f79f6bf103cd1f51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "inline_equation", + "content": "|S_{n}|" + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "text", + "content": " denotes the total number of voxels belonging to the " + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "inline_equation", + "content": "n^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "text", + "content": " spatial region. The cluster index matrix " + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "text", + "content": " and the centroid voxels " + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "inline_equation", + "content": "S^{c}" + }, + { + "bbox": [ + 313, + 611, + 555, + 660 + ], + "type": "text", + "content": " for each cluster are then utilized to generate adaptive kernel weights for convolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 662, + 555, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 555, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 555, + 698 + ], + "type": "text", + "content": "Adaptive Kernel Generator. To adaptively generate a spatial convolution kernel that incorporates region-based information, we employ a multi-layer perception to model the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 468, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 468, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 468, + 713 + ], + "type": "text", + "content": "The choices of " + }, + { + "bbox": [ + 325, + 702, + 468, + 713 + ], + "type": "inline_equation", + "content": "\\overline{p}" + }, + { + "bbox": [ + 325, + 702, + 468, + 713 + ], + "type": "text", + "content": " are investigated in Sec. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5230" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 266 + ], + "type": "image", + "image_path": "c1f954c9bea66e0b0eb0090554b79cbb10551f766c38721260705c96384710ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "text", + "content": "Figure 3. Architecture of the 3D spatial-awareness convolution block (example in three clusters). This block aims to refine the input feature " + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "text", + "content": " by adaptive convolution learned from spatial feature clustering. SACB consists of three parts: 1) the spatial context estimation module employs KMeans to cluster similar spatial features on the patched unfolding features, in which the features belonging to the same cluster centroid " + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "inline_equation", + "content": "S_{n}^{c}" + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "text", + "content": " will be indicated with a cluster index map " + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "text", + "content": ". 2) the adaptive kernel generator leverages each cluster centroid to generate cluster-specific spatial weights and bias via two MLPs, the resulting weights and bias will be used to form the final spatial adaptive convolution kernel, which will be imposed on the cluster-indexed unfolding features, again highlighting the spatial-awareness. 3) the original features and spatial-aware features are residually connected to form the final refined " + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": "mapping " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": ", which produces an adaptive convolution kernel " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_n \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times k^3}" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "n^{\\text{th}}" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": " spatial region based on " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "S_n^c" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": ". To simplify the process, we first initialize a global learnable convolution kernel " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{C_{out} \\times C_{in} \\times k^3}" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": " and then use the mapping function " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": " to generate a spatial weight that can be applied to the global kernel " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_n" + }, + { + "bbox": [ + 54, + 374, + 294, + 460 + ], + "type": "text", + "content": " is obtained as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 468, + 294, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 468, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 128, + 468, + 294, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {n} = \\mathcal {F} _ {w} \\left(S _ {n} ^ {c}\\right) \\odot \\mathbf {W}, \\tag {5}", + "image_path": "aa6f3557da12eb40691750683a7f060e826b661436f95005103ff6a30050d36b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " denotes element-wise dot product. Then, a mapping function " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_b" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " using a multi-layer perception is applied to compute the convolution bias via obtained centroid " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "S_{n}^{c}" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": ". Consequently, the region-based spatially adaptive convolution is represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 555, + 294, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 555, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 72, + 555, + 294, + 570 + ], + "type": "interline_equation", + "content": "S A C \\left(\\mathbf {F} _ {(d, w, h)}\\right) = p _ {\\left(V _ {(d, h, w)}\\right)} \\otimes \\mathbf {W} _ {n} + \\mathcal {F} _ {b} \\left(S _ {n} ^ {c}\\right), \\tag {6}", + "image_path": "e96e4b0722c69364542564c2406cb239fb7be78a6591f46ddaae5011759cdb78.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 575, + 296, + 622 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 55, + 575, + 296, + 622 + ], + "type": "text", + "content": " denotes the unfolding operation outlined in the spatial context estimation section. Subsequently, the process of applying the SAC block to produce refined features is as follows" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 629, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 629, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 97, + 629, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} = S A C B (\\mathbf {F}) = \\mathbf {F} + \\sigma (S A C (\\mathbf {F})), \\tag {7}", + "image_path": "d230b75f9c4c4554a8286d35709162e95a5581b9c0183d0c664250a1cdc190fa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 651, + 214, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 214, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 214, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 651, + 214, + 662 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 651, + 214, + 662 + ], + "type": "text", + "content": " denotes activation function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 669, + 194, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 194, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 194, + 681 + ], + "type": "text", + "content": "3.3. Pyramid Flow Estimator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F_m}^i" + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F_f}^i" + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "text", + "content": " represent the multi-scale features refined by the proposed SACB at the " + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 685, + 296, + 713 + ], + "type": "text", + "content": " level, the sub-deformation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": "flows at each scale are then estimated based on the matching score using the refined features. Specifically, we calculate the similarity matching score between each voxel in the fixed image features and corresponding neighbouring voxels in the moving image features. Similar to [25, 40], this operation can be efficiently implemented as a matrix operation using the dot product. For instance, at the coarsest level " + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "inline_equation", + "content": "i = 5" + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": ", the matching score can be computed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 476, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 476, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 341, + 476, + 553, + 495 + ], + "type": "interline_equation", + "content": "M _ {s i m} \\left(\\hat {F _ {f}} ^ {5}, \\hat {F _ {m}} ^ {5}\\right) = \\operatorname {S o f t m a x} \\left(\\hat {F _ {f}} ^ {5} p \\left(\\hat {F _ {m}} ^ {5}\\right) ^ {T}\\right), \\tag {8}", + "image_path": "25112c42054d34f845809142a9bf3310938cc4af4d078cccb3491458d9e61d90.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "text", + "content": " represents the unfolding operation as described in Section 3.2. For each voxel " + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\hat{F_m}^5" + }, + { + "bbox": [ + 313, + 501, + 554, + 540 + ], + "type": "text", + "content": ", the similarity matching score can be calculated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 546, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 546, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 339, + 546, + 553, + 564 + ], + "type": "interline_equation", + "content": "M _ {s i m} \\left(\\hat {F} _ {f} ^ {5} (V), \\hat {F} _ {m} ^ {5} (V)\\right) = \\tag {9}", + "image_path": "d5a3f3f8cbd1395185a6d1f48e0661dec469f3194ebd00d7a438411f21d9f5bd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 340, + 566, + 528, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 566, + 528, + 593 + ], + "spans": [ + { + "bbox": [ + 340, + 566, + 528, + 593 + ], + "type": "interline_equation", + "content": "\\operatorname {S o f t m a x} \\left(\\left\\{\\left\\langle \\hat {F _ {f}} ^ {5} (V), \\hat {F _ {m}} ^ {5} (V _ {c}) \\right\\rangle \\right\\} _ {V _ {c} \\in \\mathcal {N} (V)}\\right),", + "image_path": "8f0cdb156495123f9afa6c9a1d767ec13e41d281176564dccba0578749f0f502.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": " denotes the inner product, and " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(V)" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": " represents the set of voxels within an " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "k \\times k \\times k" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": " neighborhood centered around voxel " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": " represents the window size of the search region of " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\hat{F}_m^5" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": " in our experiments. Then, the sub-deformation flow at the " + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "inline_equation", + "content": "5^{\\text{th}}" + }, + { + "bbox": [ + 313, + 598, + 555, + 658 + ], + "type": "text", + "content": " scale can be obtained as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 381, + 663, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 663, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 381, + 663, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\varphi^ {5} = M _ {s i m} \\left(\\hat {F _ {f}} ^ {5}, \\hat {F _ {m}} ^ {5}\\right) G \\tag {10}", + "image_path": "d1e5c7bb278a50106dccdf5db8d182f17e09b69586a3ad5d367bb1a0da397829.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{(k \\times k \\times k) \\times 3}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " is a 3D grid that encodes the relative positions with respect to the centroid voxel location." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5231" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": "At the " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "i = 4" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": " level, we first up scaling the flow " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\varphi^5" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\hat{\\varphi}^5" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\hat{F_m}^4" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": " is warped with " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\hat{\\varphi}^5" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": ". Sequentially, the " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "M_{sim}" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": " need to be calculated between " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\hat{F_f}^4" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\hat{F_m}^4 \\circ \\hat{\\varphi}^5" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": " for obtaining " + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "inline_equation", + "content": "\\Delta \\varphi^4" + }, + { + "bbox": [ + 55, + 72, + 295, + 126 + ], + "type": "text", + "content": ". Afterward, the overall process can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 134, + 294, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 134, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 64, + 134, + 294, + 221 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\hat {\\varphi} ^ {i} = \\mathrm {u p} _ {2 \\times} \\left(\\varphi^ {i}\\right), \\\\ \\hat {F} _ {f} ^ {i - 1} = S A C B \\left(F _ {f} ^ {i - 1}\\right), \\\\ \\hat {F} _ {m} ^ {i - 1} = S A C B \\left(F _ {m} ^ {i - 1} \\circ \\hat {\\varphi} ^ {i}\\right), \\quad i \\in \\{5, 4, 3, 2 \\}. \\\\ \\Delta \\varphi^ {i - 1} = M _ {s i m} \\left(\\hat {F} _ {f} ^ {i - 1}, \\hat {F} _ {m} ^ {i - 1}\\right) G, \\\\ \\varphi^ {i - 1} = \\hat {\\varphi} ^ {i} \\circ \\Delta \\varphi^ {i - 1}, \\end{array} \\right. \\tag {11}", + "image_path": "ad8e4ebd6b53612ba4037bcbb2a72ab5e001db637f35fd99cf7f5d5c650df976.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "type": "inline_equation", + "content": "\\mathrm{up}_{2\\times}(\\cdot)" + }, + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "type": "text", + "content": " is flow rescaling operation with " + }, + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 55, + 222, + 295, + 235 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "text", + "content": "We note that the matching score calculation was not applied for flow estimation at the " + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "text", + "content": " scale, as the search region with " + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "text", + "content": " can only determine a maximum displacement of 1 voxel using full-resolution features. However, the calculations at the " + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 54, + 235, + 296, + 330 + ], + "type": "text", + "content": " scale significantly increase memory and computational costs, yielding limited benefit. Therefore, we used a simple approach with two convolutional layers for flow estimation at the final scale." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 136, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 136, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 136, + 354 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": "Datasets. Our experiments use two publicly available 3D brain MRI datasets " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathrm{IXI}^2" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " and LPBA [36], along with one public abdominal CT dataset [47]. The preprocessed IXI dataset [10] includes 576 MRI scans " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(160\\times 192\\times 224)" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " from healthy subjects. We followed the same protocol as [10], partitioning the data into 403, 58, and 115 scans for training, validation, and testing sets, respectively. Atlas-based brain registration was conducted using a generated atlas from [24]. The LPBA dataset consists of 40 T1-weighted MRI volumes. Each MRI scan has been rigidly pre-aligned to the MNI305 template and manually annotated with 54 regions of interest (ROIs). We center-crop all volumes to dimensions of " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "160\\times 192\\times 160" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " with voxel size " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "1\\times 1\\times 1mm^3" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": ". LPBA dataset was used to perform inter-subject registration with divided three subsets: 25 volumes " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(25\\times 24" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " pairs) for training, 5 " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(5\\times 4" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " pairs) for validation, and 10 " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(10\\times 9" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " pairs) for testing. The Abdomen CT dataset consists of 30 abdominal scans with 13 manually labelled anatomical structures. We clamp the Hounsfield Units (HU) to " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "[-1000, 1000]" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " and normalize each scan to " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": ". Each scan is center-cropped from " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "192\\times 160\\times 256" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "192\\times 160\\times 224" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " with voxel size " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "2\\times 2\\times 2mm^3" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": ". The dataset is divided into 19 training volumes " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(19\\times 18" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " pairs), 3 validation volumes " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(3\\times 2" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " pairs) and 8 testing volumes " + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "inline_equation", + "content": "(8\\times 7" + }, + { + "bbox": [ + 56, + 360, + 296, + 647 + ], + "type": "text", + "content": " pairs) for inter-subject registration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "content": "Implementation Details. Our method is implemented using PyTorch. Both the training and testing phases are deployed on an A100 GPU with 40GB VRAM. All models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "are optimized using Adam followed with the same learning rate " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " and batch size 1. A fast GPU K-Means implementation is adopted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "type": "text", + "content": "Evaluation Metrics. To qualify the registration performance, we use the Dice similarity coefficient (Dice) to evaluate the region of interest (ROI) segmentation overlap, the average symmetric surface distance (ASSD) [41] and the 95 percentile of the Hausdorff distance (HD95) to measure the registration accuracy of ROI structures. For evaluating the diffeomorphism of deformation fields, we use the percentage of negative values of the Jacobian determinant " + }, + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "type": "inline_equation", + "content": "(|J|_{<0\\%})" + }, + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "type": "text", + "content": " to calculate the percentage of folding voxels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 221, + 556, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 556, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 556, + 342 + ], + "type": "text", + "content": "Comparative Methods. We compare our method against state-of-the-art approaches, including the conventional SyN method [2], U-Net-based models such as VoxelMorph (VM) [5], LKU [20], and TransMorph [10], model-driven methods like B-Spline-Diff [32] and Fourier-Net [21], the lightweight NCA-Morph [33], the cascaded method LapIRN [31], and four pyramid-based methods: PR++ [23], Im2Grid [25], RDN [18] and ModeT [46]. All models were trained using their officially released code, with optimal hyperparameters tuned on the validation sets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 352, + 368, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 352, + 368, + 364 + ], + "spans": [ + { + "bbox": [ + 314, + 352, + 368, + 364 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 372, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 431 + ], + "type": "text", + "content": "In this section, we evaluate our SACB-Net against comparative methods on three tasks: atlas-based brain registration using the IXI dataset, inter-subject brain registration using the LPBA dataset, and large deformation estimation for inter-subject abdominal CT registration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 439, + 482, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 482, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 482, + 451 + ], + "type": "text", + "content": "5.1. Atlas-based Brain Registration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "text", + "content": "Table 1 presents the quantitative results of atlas-based brain registration for 30 anatomical structures evaluated on the IXI dataset. It can be seen that SACB-Net outperforms all comparable methods in terms of Dice score while achieving plausible deformations and third-best results on HD95 and ASSD. Compared to the best-performing U-Net-based method (LKU), our SACB-Net achieves a " + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "text", + "content": " improvement in Dice score while using 0.98M fewer parameters. In terms of model-driven methods, SACB-Net surpasses Fourier-Net by a margin of " + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "text", + "content": " in the Dice score and uses 3.09M fewer parameters. For cascade and pyramid-based methods, those methods have a similar number of parameters apart from RDN. Our SACB-Net outperforms the second-best cascade method LapIRN with " + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 456, + 555, + 635 + ], + "type": "text", + "content": " Dice, 0.038mm HD95, and 0.019mm ASSD, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 643, + 488, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 488, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 488, + 656 + ], + "type": "text", + "content": "5.2. Inter-subject Brain Registration" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 660, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 554, + 696 + ], + "type": "text", + "content": "The inter-subject registration of the LPBA dataset tends to have a slightly larger deformation than the atlas-based registration on the IXI dataset. As shown in Table 1, most UNet" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 693, + 274, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 693, + 274, + 702 + ], + "spans": [ + { + "bbox": [ + 65, + 693, + 274, + 702 + ], + "type": "text", + "content": "2https://brain-development.org/ixi-dataset/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 65, + 702, + 255, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 702, + 255, + 713 + ], + "spans": [ + { + "bbox": [ + 65, + 702, + 255, + 713 + ], + "type": "text", + "content": "3https://learn2reg.grand-challenge.org/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 702, + 513, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 513, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 513, + 713 + ], + "type": "text", + "content": "4https://github.com/jokofa/torch_KMeans" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5232" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 164, + 71, + 445, + 82 + ], + "lines": [ + { + "bbox": [ + 164, + 71, + 445, + 82 + ], + "spans": [ + { + "bbox": [ + 164, + 71, + 445, + 82 + ], + "type": "text", + "content": "Table 1. Registration performance comparison on the IXI and LPBA datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 553, + 275 + ], + "type": "table", + "html": "
MethodIXI (30 ROIs)LPBA (54 ROIs)#Param
Dice↑HD95↓ASSD↓|J|<0%↓Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.386±0.1956.479±0.6662.445±0.280-0.525±0.0478.039±0.8612.586±0.350--
SyN [3]0.645±0.1526.394±1.0481.551±0.286<0.00010.707±0.0166.254±0.4441.479±0.131<0.0001-
VM-1 [5]0.729±0.1293.798±0.7570.937±0.1621.590±0.3390.664±0.0256.873±0.6541.717±0.2000.649±0.2610.27M
VM-2 [5]0.732±0.1233.723±0.6800.926±0.1581.522±0.3360.669±0.0256.847±0.6591.698±0.2000.591±0.2420.30M
NCA-Morph [33]0.753±0.1363.109±0.5250.796±0.1210.506±0.1900.679±0.0236.666±0.6341.631±0.1880.264±0.1210.37M
TransMorph [10]0.754±0.1243.543±0.7210.862±0.1681.579±0.3280.695±0.0226.564±0.6191.559±0.1820.474±0.17646.77M
LKU [20]0.765±0.1292.967±0.4940.757±0.1140.109±0.0540.706±0.0326.452±0.9511.603±0.2500.594±0.2032.09M
B-Spline-Diff [32]0.742±0.1283.256±0.5380.832±0.117<0.00010.665±0.0236.792±0.6221.713±0.1860.0±0.00.27M
Fourier-Net [21]0.763±0.1292.857±0.4560.748±0.1140.024±0.0190.672±0.0226.716±0.6011.666±0.1800.216±0.1044.20M
LapIRN [31]0.763±0.1333.166±0.6080.779±0.1260.312±0.1060.716±0.0166.116±0.4541.426±0.1330.024±0.0091.20M
PRNet++ [23]0.755±0.1303.593±0.7480.857±0.1621.052±0.3020.701±0.0216.492±0.5971.520±0.1770.072±0.0271.24M
ModeT [46]0.758±0.1253.496±0.7320.828±0.1510.114±0.0570.721±0.0135.969±0.4161.375±0.1100.010±0.0041.03M
Im2Grid [25]0.761±0.1273.316±0.6680.799±0.128<0.00020.713±0.0146.062±0.4281.419±0.1180.007±0.0030.89M
RDN [18]0.759±0.1233.476±0.8020.823±0.161<0.00010.713±0.0176.208±0.4971.436±0.142<0.000228.65M
Ours0.769±0.1273.128±0.6310.760±0.1250.083±0.0450.731±0.0125.862±0.4361.326±0.1140.018±0.0061.11M
", + "image_path": "2d35b96ac74ffe40a9c256db201fd6230b703030ca125937eb5a0db93bb1f620.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": "based and model-driven methods do not perform as well as cascade and pyramid-based methods. Only LKU surpasses the pyramid-based method PR++, achieving improvements of " + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": " in Dice, " + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "inline_equation", + "content": "0.04 \\mathrm{~mm}" + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": " in HD95 and " + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "inline_equation", + "content": "0.083 \\mathrm{~mm}" + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": " in ASSD, respectively. Our SACB outperforms all comparable methods in Dice, HD95 and ASSD with plausible deformations. Compared to the second-best method, ModeT, SACB achieves improvements of " + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": " in Dice, " + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "inline_equation", + "content": "0.107 \\mathrm{~mm}" + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": " in HD95, and " + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "inline_equation", + "content": "0.049 \\mathrm{~mm}" + }, + { + "bbox": [ + 54, + 296, + 294, + 439 + ], + "type": "text", + "content": " in ASSD, respectively. The visualization in Fig. 4 shows that our SACB-Net captures more detailed correspondences in the registration process, as highlighted in the red box." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 248, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 248, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 248, + 465 + ], + "type": "text", + "content": "5.3. Inter-subject Abdomen Registration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": "The primary challenge in abdominal CT registration tasks is caused by the substantial differences in the spatial distribution of organs across scans, making it difficult to accurately estimate large deformations during the registration process. Table 2 demonstrates the numerical results of the inter-subject abdomen registration task. Our results show that our method outperforms all compared approaches in terms of Dice, HD95, and ASSD. Notably, U-Net-based and model-driven methods perform inferior to pyramid and cascaded approaches for estimating large deformations. Even the best-performing method in such approaches, TransMorph, still exhibits a significant gap compared to our SACB-Net with " + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": " difference in Dice score, " + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "inline_equation", + "content": "5.934\\mathrm{mm}" + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": " in HD95 and " + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "inline_equation", + "content": "3.153\\mathrm{mm}" + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": " in ASSD, respectively. Compared to the second-best pyramid method, ModeT, SACB-Net achieves improvements of " + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": " in Dice, " + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "inline_equation", + "content": "2.098\\mathrm{mm}" + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": " in HD95, and " + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "inline_equation", + "content": "0.781\\mathrm{mm}" + }, + { + "bbox": [ + 56, + 472, + 295, + 675 + ], + "type": "text", + "content": " in ASSD, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The visualization of the wrapped moving images and labels, as shown in Fig. 4, illustrates that most parts of the kidneys are registered with our method. However, Trans" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 314, + 553, + 460 + ], + "blocks": [ + { + "bbox": [ + 317, + 295, + 552, + 306 + ], + "lines": [ + { + "bbox": [ + 317, + 295, + 552, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 552, + 306 + ], + "type": "text", + "content": "Table 2. Registration performance comparison on Abdomen CT." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 314, + 553, + 460 + ], + "lines": [ + { + "bbox": [ + 314, + 314, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 314, + 553, + 460 + ], + "type": "table", + "html": "
MethodAbdomen CT (13 ROIs)
Dice↑HD95↓ASSD↓|J|<0%↓
Affine0.305±0.04727.790±4.86810.869±2.075-
VM-2 [5]0.406±0.05825.226±4.9428.890±2.0663.882±1.512
NCA-Morph [33]0.406±0.06125.959±4.9829.083±2.2235.197±1.166
TransMorph [10]0.444±0.06524.187±4.9558.156±2.0656.045±1.750
LKU [20]0.423±0.05824.252±4.6048.395±1.9464.316±1.778
B-Spline-Diff [32]0.375±0.05325.338±4.8669.217±1.9600.0±0.0
Fourier-Net [21]0.417±0.04823.084±4.2018.000±1.5322.167±2.391
LapIRN [31]0.522±0.05720.728±4.3186.274±1.3793.203±0.848
PRNet++ [23]0.478±0.06023.791±4.6917.524±1.7712.252±1.192
ModeT [46]0.550±0.05420.351±4.2315.784±1.2362.239±0.906
Im2Grid [25]0.530±0.04720.763±3.7866.028±1.1462.421±0.653
RDN [18]0.527±0.05521.782±4.7226.461±1.6131.529±0.989
Ours0.588±0.04918.253±3.6105.003±1.0113.930±1.416
", + "image_path": "20969baea225e298cf3a33b1ae4ebb8ceeab10844c5cf0310d942a444fd51829.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 491, + 553, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 516 + ], + "type": "text", + "content": "Morph, " + }, + { + "bbox": [ + 313, + 491, + 553, + 516 + ], + "type": "inline_equation", + "content": "\\mathrm{PR} + +" + }, + { + "bbox": [ + 313, + 491, + 553, + 516 + ], + "type": "text", + "content": ", and RDN failed to register the kidneys, leaving a large portion missing on the warped mask." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 540, + 406, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 406, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 406, + 554 + ], + "type": "text", + "content": "6. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 565, + 553, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 553, + 625 + ], + "type": "text", + "content": "First, we note that " + }, + { + "bbox": [ + 313, + 565, + 553, + 625 + ], + "type": "inline_equation", + "content": "\\overline{p}(\\cdot)" + }, + { + "bbox": [ + 313, + 565, + 553, + 625 + ], + "type": "text", + "content": " in Eq. (3) can also be defined on the channel dimension or both the spatial+channel (Mix) dimension. As shown in Table 4, using the spatial mean of unfolded patches achieves slightly better performance compared to the channel mean and the Mix setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "We then determine the optimal number of scales at which SACB should be applied and the ideal number of clusters to achieve accurate registration performance. As shown in Table 3, the optimal configuration is applying SACB from scale " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "5" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " to scale " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "2" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ". Additionally, increasing the " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " from 5 to 7 boosts the performance while a too-large value slightly decreases performance (e.g. 11)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5233" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 555, + 297 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 555, + 297 + ], + "type": "image", + "image_path": "6657bc9b803d28710cf20fc911dc5db93f9c4665709293a01646f168d4228e48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "type": "text", + "content": "Figure 4. Visual comparisons on the Brain LPBA dataset (top two rows) and the Abdomen CT dataset (bottom two rows). Columns 3-10 display warped moving images (row 1 and 3), displacement fields as RGB images (row 2) and warped segmentation masks (row 4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 381, + 553, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "type": "text", + "content": "Table 3. Ablation studies on the use of SACB at different scale levels in SACB-Net and varying numbers of spatial clusters " + }, + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 55, + 340, + 555, + 373 + ], + "type": "text", + "content": ", along with the corresponding average training time (in seconds), including data loading time, and GPU training memory usage (in MiB). The SACB configurations are based on the best results shown in Table 4." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 381, + 553, + 485 + ], + "lines": [ + { + "bbox": [ + 57, + 381, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 381, + 553, + 485 + ], + "type": "table", + "html": "
NWith SACBIXILPBAAbdomen CT
Scale5Scale4Scale3Scale2Dice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTrainingDice ↑|J|<0% ↓MemoryTraining
-XXXX0.76430.059158380.7620.71410.020115740.5450.53742.907158380.764
5✓XXX0.76680.072158800.7990.72170.087116640.5790.54443.243158800.797
5✓✓XX0.76790.076161040.8790.72410.104117700.6310.54982.619161040.864
5✓✓✓X0.76710.060169600.9450.72660.076124180.7140.56852.781169600.950
5✓✓✓✓0.76830.086216381.2550.72940.017157480.9180.58493.548216381.234
7✓✓✓✓0.76910.083216381.3760.73090.018157781.0840.58813.930216621.360
9✓✓✓✓0.76850.069216801.5500.73040.016157921.1100.58754.405216801.562
11✓✓✓✓0.76840.084217321.6790.73000.015158061.1930.58754.419217181.728
", + "image_path": "41915a550506467991623e5375ad9ad642d9857c1142ca00a2fad68be2ce84c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 86, + 567, + 264, + 635 + ], + "blocks": [ + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "text", + "content": "Table 4. Ablations of the SACB configurations. After padding and unfolding on the feature map " + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "text", + "content": ", the cluster maps can be calculated based on spatial mean on the local unfolded patch, channel mean and Mix (spatial+channel). In this study, we set " + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "text", + "content": " and apply SACBs from scale" + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "inline_equation", + "content": "_5" + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "text", + "content": " to scale" + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "inline_equation", + "content": "_2" + }, + { + "bbox": [ + 55, + 502, + 295, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 567, + 264, + 635 + ], + "lines": [ + { + "bbox": [ + 86, + 567, + 264, + 635 + ], + "spans": [ + { + "bbox": [ + 86, + 567, + 264, + 635 + ], + "type": "table", + "html": "
SACBLPBA
SpatialChannelMixDice ↑|J|<0% ↓
✓✗✗0.72940.017
✗✓✗0.72810.020
✗✗✓0.72850.021
", + "image_path": "a4c239f6de6b11c9bbea850c87e003bf9d817b6320feb21639d813753c380db1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 656, + 128, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 128, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 128, + 668 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We introduced SACB-Net, which incorporates a 3D Spatial Awareness Convolution Block (SACB) that adaptively generates spatial convolution kernels based on the feature" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 505, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 624 + ], + "type": "text", + "content": "clusters, thereby improving the network's ability to capture spatially varying information. Additionally, our proposed flow estimators combine SACB with a similarity-matching mechanism to estimate displacements in a pyramid coarse-to-fine manner. Results on two brain datasets and one abdomen dataset demonstrate that SACB-Net effectively handles both small local and large deformations. Furthermore, the proposed flow estimator is a plug-in module that can be applied in other networks for better feature representation learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgements. The research was conducted using the Baskerville Tier 2 HPC service, which was funded by the EPSRC and UKRI through the World Class Labs scheme (EP/T022221/1) and the Digital Research Infrastructure programme (EP/W032244/1). Baskerville is operated by Advanced Research Computing at the University of Birmingham." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5234" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "text", + "content": "[1] John Ashburner. A fast diffeomorphic image registration algorithm. Neuroimage, 38(1):95-113, 2007. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 167 + ], + "type": "text", + "content": "[2] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain. Med Image Anal, 12(1):26-41, 2008. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 213 + ], + "type": "text", + "content": "[3] Brian B Avants, Nicholas J Tustison, Gang Song, Philip A Cook, Arno Klein, and James C Gee. A reproducible evaluation of ants similarity metric performance in brain image registration. Neuroimage, 54(3):2033-2044, 2011. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 296, + 267 + ], + "type": "text", + "content": "[4] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. An unsupervised learning model for deformable medical image registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9252-9260, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "text", + "content": "[5] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "text", + "content": "[6] M Faisal Beg, Michael I Miller, Alain Trouve, and Laurent Younes. Computing large deformation metric mappings via geodesic flows of diffeomorphisms. International journal of computer vision, 61:139-157, 2005. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 411 + ], + "type": "text", + "content": "[7] Xiaohuan Cao, Jianhua Yang, Jun Zhang, Qian Wang, Pew-Thian Yap, and Dinggang Shen. Deformable image registration using a cue-aware deep regression network. IEEE Transactions on Biomedical Engineering, 65(9):1900-1911, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "type": "text", + "content": "[8] Junyu Chen, Yufan He, Eric C Frey, Ye Li, and Yong Du. Vit-v-net: Vision transformer for unsupervised volumetric medical image registration. arXiv preprint arXiv:2104.06468, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 502 + ], + "type": "text", + "content": "[9] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8064-8073, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 503, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 546 + ], + "type": "text", + "content": "[10] Junyu Chen, Eric C Frey, Yufan He, William P Segars, Ye Li, and Yong Du. Transmorph: Transformer for unsupervised medical image registration. Medical image analysis, 82:102615, 2022. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 602 + ], + "type": "text", + "content": "[11] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11030-11039, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 668 + ], + "type": "text", + "content": "[12] Xinxing Cheng, Xi Jia, Wenqi Lu, Qiufu Li, Linlin Shen, Alexander Krull, and Jinming Duan. Winet: Wavelet-based incremental learning for efficient medical image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 761-771. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Adrian Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces. Medical image analysis, 57:226-236, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Adrian V Dalca, Guha Balakrishnan, John Guttag, and Mert R Sabuncu. Unsupervised learning for fast probabilistic diffeomorphic registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 729-738. Springer, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "type": "text", + "content": "[15] Bob D De Vos, Floris F Berendsen, Max A Viergever, Hessam Sokooti, Marius Staring, and Ivana Isgum. A deep learning framework for unsupervised affine and deformable image registration. Medical image analysis, 52:128-143, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[16] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 239, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 303 + ], + "type": "text", + "content": "[17] Morteza Ghahremani, Mohammad Khateri, Bailiang Jian, Benedikt Wiestler, Ehsan Adeli, and Christian Wachinger. H-vit: A hierarchical vision transformer for deformable image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11513-11523, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 347 + ], + "type": "text", + "content": "[18] Bo Hu, Shenglong Zhou, Zhiwei Xiong, and Feng Wu. Recursive decomposition network for deformable image registration. IEEE Journal of Biomedical and Health Informatics, 26(10):5130-5141, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[19] Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. Dynamic filter networks. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "text", + "content": "[20] Xi Jia, Joseph Bartlett, Tianyang Zhang, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. U-net vs transformer: Is u-net outdated in medical image registration? In International Workshop on Machine Learning in Medical Imaging, pages 151-160. Springer, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "text", + "content": "[21] Xi Jia, Joseph Bartlett, Wei Chen, Siyang Song, Tianyang Zhang, Xinxing Cheng, Wenqi Lu, Zhaowen Qiu, and Jinming Duan. Fourier-net: Fast image registration with band-limited deformation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1015-1023, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "text", + "content": "[22] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI conference on artificial intelligence, pages 1113-1121, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "text", + "content": "[23] Miao Kang, Xiaojun Hu, Weilin Huang, Matthew R Scott, and Mauricio Reyes. Dual-stream pyramid registration network. Medical image analysis, 78:102379, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 635 + ], + "type": "text", + "content": "[24] Boah Kim, Dong Hwan Kim, Seong Ho Park, Jieun Kim, June-Goo Lee, and Jong Chul Ye. Cycle morph: cycle consistent unsupervised deformable image registration. Medical image analysis, 71:102036, 2021. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[25] Yihao Liu, Lianrui Zuo, Shuo Han, Yuan Xue, Jerry L Prince, and Aaron Carass. Coordinate translator for learning deformable medical image registration. In International workshop on multiscale multimodal medical imaging, pages 98-109. Springer, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Tai Ma, Suwei Zhang, Jiafeng Li, and Ying Wen. Iirp-net: Iterative inference residual pyramid network for enhanced" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5235" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11546-11555, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 139 + ], + "type": "text", + "content": "[27] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[28] Mingyuan Meng, Dagan Feng, Lei Bi, and Jinman Kim. Correlation-aware coarse-to-fine mlp's for deformable medical image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9645-9654, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "text", + "content": "[29] Tony CW Mok and Albert Chung. Fast symmetric diffeomorphic image registration with convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4644-4653, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[30] Tony CW Mok and Albert Chung. Affine medical image registration with coarse-to-fine vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20835-20844, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 360 + ], + "type": "text", + "content": "[31] Tony CW Mok and Albert CS Chung. Large deformation diffeomorphic image registration with laplacian pyramid networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part III 23, pages 211-221. Springer, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 405 + ], + "type": "text", + "content": "[32] Huaqi Qiu, Chen Qin, Andreas Schuh, Kerstin Hammernik, and Daniel Rueckert. Learning diffeomorphic and modality-invariant registration using b-splines. In Medical Imaging with Deep Learning, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "type": "text", + "content": "[33] Amin Ranem, John Kalkhof, and Anirban Mukhopadhyay. Nca-morph: Medical image registration with neural cellular automata. arXiv preprint arXiv:2410.22265, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 514 + ], + "type": "text", + "content": "[34] Marc-Michel Rohe, Manasi Datar, Tobias Heimann, Maxime Sermesant, and Xavier Pennec. Svf-net: learning deformable image registration using shape matching. In Medical Image Computing and Computer Assisted Intervention- MIC-CAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part I 20, pages 266-274. Springer, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 559 + ], + "type": "text", + "content": "[35] Yudi Sang, Xianglei Xing, Yingnian Wu, and Dan Ruan. Imposing implicit feasibility constraints on deformable image registration using a statistical generative model. Journal of Medical Imaging, 7(6):064005-064005, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[36] David W Shattuck, Mubeena Mirza, Vitria Adisetiyo, Cornelius Hojatkashani, Georges Salamon, Katherine L Narr, Russell A Poldrack, Robert M Bilder, and Arthur W Toga. Construction of a 3d probabilistic atlas of human cortical structures. Neuroimage, 39(3):1064-1080, 2008. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 680 + ], + "type": "text", + "content": "[37] Jiacheng Shi, Yuting He, Youyong Kong, Jean-Louis Coatrieux, Huazhong Shu, Guanyu Yang, and Shuo Li. Xmorph: Full transformer for deformable medical image registration via cross attention. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 217-226. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[38] Aristeidis Sotiras, Christos Davatzikos, and Nikos Paragios. Deformable medical image registration: A survey. IEEE transactions on medical imaging, 32(7):1153-1190, 2013. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[39] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[40] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[41] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15:1-28, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 308 + ], + "type": "text", + "content": "[42] Alexander Thorley, Xi Jia, Hyung Jin Chang, Boyang Liu, Karina Bunting, Victoria Stoll, Antonio de Marvao, Declan P O'Regan, Georgios Gkoutos, Dipak Kotecha, et al. Nesterov accelerated admm for fast diffeomorphic image registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 150-160. Springer, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 385 + ], + "type": "text", + "content": "[43] Lin Tian, Hastings Greer, François-Xavier Vialard, Roland Kwitt, Raul San José Estépar, Richard Jarrett Rushmore, Nikolaos Makris, Sylvain Bouix, and Marc Niethammer. Gradicon: Approximate diffeomorphisms via gradient inverse consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18084-18094, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 419 + ], + "type": "text", + "content": "[44] D Ulyanov. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[45] Tom Vercauteren, Xavier Pennec, Aymeric Perchant, and Nicholas Ayache. Diffeomorphic demons: Efficient nonparametric image registration. NeuroImage, 45(1):S61-S72, 2009. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "text", + "content": "[46] Haiqiao Wang and Yi Ni, Dongand Wang. Model: Learning deformable image registration via motion decomposition transformer. In Medical Image Computing and Computer Assisted Intervention - MICCAI 2023, pages 740-749, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 588 + ], + "type": "text", + "content": "[47] Zhoubing Xu, Christopher P Lee, Mattias P Heinrich, Marc Modat, Daniel Rueckert, Sebastien Ourselin, Richard G Abramson, and Bennett A Landman. Evaluation of six registration methods for the human abdomen on clinically acquired ct. IEEE Transactions on Biomedical Engineering, 63(8):1563-1572, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 666 + ], + "type": "text", + "content": "[48] Xiaofeng Yang, Ning Wu, Guanghui Cheng, Zhengyang Zhou, S Yu David, Jonathan J Beitler, Walter J Curran, and Tian Liu. Automated segmentation of the parotid gland based on atlas registration and machine learning: a longitudinal mri study in head-and-neck radiation therapy. International Journal of Radiation Oncology*Biology*Physics, 90 (5):1225-1233, 2014. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 667, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 667, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 667, + 553, + 700 + ], + "type": "text", + "content": "[49] Xiao Yang, Roland Kwitt, Martin Styner, and Marc Niethammer. Quicksilver: Fast predictive image registration-a deep learning approach. NeuroImage, 158:378-396, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "text", + "content": "[50] Jun Zhang. Inverse-consistent deep networks for unsu" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5236" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 327 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "pervised deformable image registration. arXiv preprint arXiv:1809.03443, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "type": "text", + "content": "[51] Yungeng Zhang, Yuru Pei, and Hongbin Zha. Learning dual transformer network for diffeomorphic registration. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part IV 24, pages 129-138. Springer, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[52] Shengyu Zhao, Yue Dong, Eric I Chang, Yan Xu, et al. Recursive cascaded networks for unsupervised medical image registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10600-10610, 2019. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 272 + ], + "type": "text", + "content": "[53] Shengyu Zhao, Tingfung Lau, Ji Luo, I Eric, Chao Chang, and Yan Xu. Unsupervised 3d end-to-end medical image registration with volume tweening network. IEEE journal of biomedical and health informatics, 24(5):1394-1404, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 296, + 327 + ], + "type": "text", + "content": "[54] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6647-6656, 2021. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5237" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_content_list.json b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..95105dc648dd6d79ed96a4af0d94a98a1e254a07 --- /dev/null +++ b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_content_list.json @@ -0,0 +1,1962 @@ +[ + { + "type": "text", + "text": "SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining", + "text_level": 1, + "bbox": [ + 124, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingjin Zhang $^{1}$ , Xiaolong Li $^{1*}$ , Fei Gao $^{1}$ , Jie Guo $^{1*}$ , Xinbo Gao $^{1,2}$ , Jing Zhang $^{3}$ $^{1}$ Xidian University, China $^{2}$ Chongqing University of Posts and Telecommunications, China \n $^{3}$ School of Computer Science & School of Artificial Intelligence, Wuhan University, China \n{mjinzhang, fgao}@xidian.edu.cn, 23011210739@stu.xidian.edu.cn \njguo, xbgao@mail.xidian.edu.cn, jingzhang.csv@gmail.com", + "bbox": [ + 130, + 202, + 869, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared Small Target Detection (IRSTD) aims to identify low signal-to-noise ratio small targets in infrared images with complex backgrounds, which is crucial for various applications. However, existing IRSTD methods typically rely solely on image modalities for processing, which fail to fully capture contextual information, leading to limited detection accuracy and adaptability in complex environments. Inspired by vision-language models, this paper proposes a novel framework, SAIST, which integrates textual information with image modalities to enhance IRSTD performance. The framework consists of two main components: Scene Recognition Contrastive Language-Image Pretraining (SR-CLIP) and CLIP-guided Segment Anything Model (CG-SAM). SR-CLIP generates a set of visual descriptions through object-object similarity and object-scene relevance, embedding them into learnable prompts to refine the textual description set. This reduces the domain gap between vision and language, generating precise textual and visual prompts. CG-SAM utilizes the prompts generated by SR-CLIP to accurately guide the Mask Decoder in learning prior knowledge of background features, while incorporating infrared imaging equations to improve small target recognition in complex backgrounds and significantly reduce the false alarm rate. Additionally, this paper introduces the first multimodal IRSTD dataset, MIRSTD, which contains abundant image-text pairs. Experimental results demonstrate that the proposed SAIST method outperforms existing state-of-the-art approaches.", + "bbox": [ + 89, + 358, + 485, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 815, + 223, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared small target detection (IRSTD) is crucial for applications like traffic management and maritime rescue", + "bbox": [ + 89, + 840, + 483, + 873 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fba36fd60f51c873be7c05f05db2569a5dad0cfd6f2db93bed58be0c2ea4c6c.jpg", + "image_caption": [ + "Figure 1. Illustration of the advantages of text-guided detection over image-only detection in identifying small infrared targets in complex backgrounds. The textual description offers valuable context, helping to reduce interference and distinguish the target from the background effectively." + ], + "image_footnote": [], + "bbox": [ + 535, + 325, + 885, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[5, 18, 29]. Unlike visible light, infrared sensors detect thermal radiation, enabling target acquisition in adverse conditions such as fog, rain, snow, and at night [44]. However, due to the long-range imaging mechanism, infrared images often lack distinctive features like color, shape, and texture, while the background typically shows strong structural features with higher intensity and sharper contours [9, 43]. These challenges—complex backgrounds, clutter, and noise—hinder accurate detection, complicating IRSTD.", + "bbox": [ + 511, + 583, + 908, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IRSTD methods can be divided into traditional algorithms and deep learning-based ones. Early methods focused on filtering and image enhancement [1, 6, 12], which performed well in simpler scenarios but struggled with complex environments due to reliance on prior knowledge and handcrafted features. With the rise of deep learning and large IRSTD datasets, newer approaches using Generative Adversarial Networks (GANs) [30], U-Net [7, 21, 38], and transformers [31, 35, 37] have been introduced.", + "bbox": [ + 511, + 719, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite advances, existing methods are limited by their reliance on single-modality inputs. Recently, the CLIP-SAM model, which combines Contrastive Language-", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 109, + 886, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9549", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Pretraining (CLIP) and the Segment Anything Model (SAM), has shown promise in enhancing cross-modal feature fusion for various tasks [20]. CLIP [24] captures semantic relationships between images and text, offering rich supervision, while SAM [16] provides precise segmentation. However, applying CLIP-SAM to IRSTD remains challenging due to the differences between infrared and natural images. Specifically: 1) Infrared targets lack typical visual features, limiting the generalization of the CLIP model; 2) Small, weak infrared targets and noisy backgrounds complicate target detection, making SAM's application difficult. To address these issues, we propose using CLIP to extract semantically rich prior knowledge from text, guiding SAM to better capture target details amidst complex backgrounds.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper introduces a novel IRSTD approach named SAIST by combining text and image modalities within a multimodal learning framework. First, we design the Scene Recognition CLIP (SR-CLIP) to capture both global scene and local object features. SR-CLIP creates similarity matrices for objects and scenes, facilitating key information extraction. It also integrates a cross-modal interaction mechanism, incorporating scene-specific visual cues into learnable prompts that refine text features and mitigate the vision-language domain gap. Building on this, we diverse CG-SAM to utilize the text prompts from SR-CLIP to guide SAM in learning background features, enhancing accurate identification and extraction across diverse scenes. By combining global semantic and background features, and leveraging the radiative transfer properties of targets and backgrounds, CG-SAM achieves effective target-background separation. Additionally, we present MIRSTD, the first multimodal IRSTD dataset, which combines imagedataset pairs from NUAA-SIRST, NUDT-SIRST, and IRSTD1K datasets. Text descriptions are generated by GPT-4V and carefully reviewed manually. Experiments validate the superiority of SAIST over state-of-the-art (SOTA) methods.", + "bbox": [ + 91, + 303, + 482, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 109, + 637, + 447, + 651 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SAIST, the first approach to integrate multimodal vision and language for IRSTD. On challenging benchmarks, it outperforms SOTA methods, excelling in both objective metrics and subjective evaluations.", + "- We design two key components, SR-CLIP and CG-SAM, to effectively fuse multimodal information and separate targets and backgrounds in complex scenes.", + "- We create a novel multimodal IRSTD dataset combining infrared images and textual data, offering rich resources for training and evaluating multimodal learning methods." + ], + "bbox": [ + 91, + 652, + 482, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 91, + 821, + 227, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Infrared small target detection", + "text_level": 1, + "bbox": [ + 91, + 847, + 362, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional IRSTD methods use image processing or handcrafted features, such as low-rank (e.g., RIFT [6],", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PSTNN [36]), human visual system-based (e.g., TLLCM [4], WSLCM [12]), and filter-based approaches (e.g., tophat [1]). Low-rank methods perform well in low SNR but struggle with false alarms in low-contrast areas. Human visual methods are effective with large objects but not in complex scenes, and filter-based techniques suppress uniform noise but fail in complex backgrounds. Deep learning IRSTD methods include CNN-based models (e.g., MDvsFA [30], ISNet [39], Dim2Clear [40], CSRNet [21], DMFNet [11], GCI-Net [41]) that focus on local features, with some improving SNR, detection, or small target extraction, and hybrid models (e.g., IAANet [31] and RKformer [37]) that combine CNNs and Vision Transformers (ViTs) to capture both local and global context. While IRSTD has made significant strides, current methods rely mainly on image modalities, which fail to capture full contextual information, limiting detection accuracy in complex backgrounds. Integrating textual information can offer substantial benefits, providing richer scene understanding and aiding target extraction in challenging environments.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. CLIP and SAM Collaboration", + "text_level": 1, + "bbox": [ + 513, + 405, + 782, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, foundational models like CLIP and SAM have garnered significant attention for their exceptional performance across various tasks. CLIP excels at aligning multimodal features, while SAM specializes in segmentation through diverse prompts. Recent research has explored synergies between these models, such as CRIS [32] integrates both models to boost performance. ClipSAM [20] uses CLIP's semantic understanding for anomaly localization and coarse segmentation, later refined by SAM. OpenVocabulary SAM enhances SAM's recognition by transferring knowledge from SAM to CLIP via distillation and adapters. However, applying pre-trained CLIP-SAM models, originally trained on natural images, to IRSTD faces challenges due to domain differences. In this work, we leverage CLIP's semantic understanding from scene descriptions to guide SAM in extracting target information from complex infrared backgrounds.", + "bbox": [ + 511, + 428, + 906, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Datasets for IRSTD", + "text_level": 1, + "bbox": [ + 513, + 696, + 700, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior studies have developed large-scale datasets [7, 19, 39], which have advanced IRSTD research. However, these datasets only include image data with target masks, limiting the detection of small targets in complex backgrounds. Recent multimodal approaches, such as text-guided detection [17, 22, 25], have shown performance improvements. Inspired by this, we create the MIRSTD dataset, which combines both image and text modalities. The text modality in MIRSTD provides descriptions of scenes in infrared images from datasets like NUAA-SIRST, IRSTD-1k, and NUTDSIRST, helping to mitigate the challenge of weak target signals against strong background noise.", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9550", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/188d26818404e83d33891b2183ffbf1284b1fc20818ed0467c2e13364219a0c1.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed encoder-decoder architecture. The CLIP text encoder converts textual descriptions into feature representations, while the frozen image encoder processes infrared images without target. SR-CLIP leverages cross-modal interactions to refine textual features, bridging the visual-textual domain gap. The SAM image encoder extracts image embeddings, fine-tuned with LoRA to enhance task adaptability. The generated textual prompts and image embeddings are fed into the SAM mask decoder, along with an infrared imaging model, to effectively suppress background noise and detect small targets." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 483, + 305 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87092d3acab5275f7887276516369979861d590fd40b72b0f416b13cea58645c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 905, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 410, + 225, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 436, + 285, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of SAIST is illustrated in Fig. 2. In this framework, the text encoder of SR-CLIP converts textual descriptions into feature representations, while the image encoder is frozen. Next, SR-CLIP employs a cross-modal interaction mechanism to integrate the visual and textual description sets, refining the textual features and reducing the domain gap between vision and language, thus generating precise textual and visual prompts. Subsequently, the pre-trained and frozen SAM image encoder extracts image embeddings, which are fine-tuned using Low-Rank Adaptation (LoRA) [14] to enhance the model's adaptability to the specific task. Finally, the generated textual prompts and image embeddings are fed into the SAM mask decoder, in conjunction with a specially designed infrared imaging model, to effectively suppress background noise and accurately detect small targets.", + "bbox": [ + 89, + 459, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. SR-CLIP", + "text_level": 1, + "bbox": [ + 89, + 710, + 199, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In infrared imaging, due to the inherent characteristics of remote sensing, the target's radiative intensity is typically weak, while the background's intensity is higher and more prominent. In this case, the information related to the target is difficult to describe clearly and accurately. In contrast, the background's radiative intensity is more stable and easier to identify, allowing the overall structure and features of the scene to be effectively reflected through intuitive textual descriptions. There, we design SR-CLIP to fully leverage the complementary relationship between background textual descriptions and visual features, generating precise", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompts that guide SAM to effectively suppress background interference in infrared images.", + "bbox": [ + 511, + 412, + 906, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Visual / Text description set", + "text_level": 1, + "bbox": [ + 511, + 449, + 751, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the scene recognition task in infrared images, scenes are composed of multiple semantic concepts, including distinct objects and backgrounds. These features necessitate the consideration of various semantic information when learning discriminative patterns. To capture comprehensive scene information, we introduce a novel form of textual prompting, referred to as scene prompts. This prompt creates a textual description set:", + "bbox": [ + 511, + 468, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "{aphotof[scene]with[object],[object],and[object]},", + "bbox": [ + 517, + 599, + 901, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which provides both scene-level and object-level textual descriptions. Simultaneously, to bridge the image-text domain gap, we extract local object features from Clip's image encoder, acquire global visual features with the help of a learnable projection layer, and generate a set of visual feature descriptions using the image captioning [13, 34].", + "bbox": [ + 511, + 626, + 905, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\left[ \\text {s c e n e} \\right], \\left[ \\text {o b j e c t} \\right], \\left[ \\text {o b j e c t} \\right] \\left[ \\text {o b j e c t} \\right] \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 727, + 836, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which provides both scene-level and object-level visual descriptions.", + "bbox": [ + 511, + 753, + 905, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2. Image Captioning", + "text_level": 1, + "bbox": [ + 511, + 791, + 687, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In scene recognition, objects often share considerable similarity, and there is a strong correlation between objects and the scene, which reflects the unique characteristics of that particular scene. By integrating both interactions, the model can capture more granular semantic information, thereby improving its discriminative power in specific contexts. To", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9551", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b939f16afc102edb76b83a6db7b66ef9e50933e872a869cf306af8abc1410ee2.jpg", + "image_caption": [ + "Figure 3. Overview of the Cross-modal Interaction Mechanism. Image captioning generates visual feature descriptions that are processed by a text encoder. The text decoder then maps visual information into the text, enhancing text feature expressiveness. Finally, the visual decoder recovers visual information from the encoded text to improve image-text alignment." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 486, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this end, we construct both an object-object similarity matrix and a scene-object correlation matrix:", + "bbox": [ + 89, + 378, + 483, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object-Object Similarity Matrix:", + "bbox": [ + 109, + 410, + 349, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c c c} o _ {1 1} & \\dots & o _ {1 n} \\\\ \\vdots & \\ddots & \\vdots \\\\ o _ {n 1} & \\dots & o _ {n n} \\end{array} \\right] = \\frac {\\exp \\left(\\frac {o _ {i} ^ {\\top} o _ {j}}{\\| o _ {i} \\| \\| o _ {j} \\|}\\right)}{\\sum_ {k = 1} ^ {n} \\exp \\left(\\frac {o _ {i} ^ {\\top} o _ {k}}{\\| o _ {i} \\| \\| o _ {k} \\|}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 434, + 482, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $o_{i,j}$ represents the similarity between the feature vector of object $i$ and object $j$ .", + "bbox": [ + 89, + 501, + 482, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene-Object Correlation Matrix:", + "bbox": [ + 109, + 531, + 352, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l l l} s _ {1} & \\dots & s _ {n} \\end{array} \\right] = \\frac {\\exp \\left(\\frac {s ^ {\\top} o _ {i}}{\\| s \\| \\| o _ {i} \\|}\\right)}{\\sum_ {k = 1} ^ {n} \\exp \\left(\\frac {s ^ {\\top} o _ {k}}{\\| s \\| \\| o _ {k} \\|}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 555, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_i$ represents the correlation between the feature vector of the scene and the feature vector of object $j$ .", + "bbox": [ + 89, + 618, + 482, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, based on the similarity and correlation matrices, we design the Object-to-Object and Scene-to-Object modules to effectively capture the interrelationships between objects and facilitate the integration of scene and object details. Meanwhile, by incorporating image captioning techniques, we generate precise visual description.", + "bbox": [ + 89, + 648, + 483, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Cross-Modal Interaction Mechanism", + "text_level": 1, + "bbox": [ + 89, + 746, + 393, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CoOp [42] can be viewed as a textual domain prompt, where the learnable context is simply a trainable vector without incorporating any visual information. Inspired by CoOp, we propose a cross-modal interaction mechanism that introduces the visual description set into the text description set to refine the textual features. This approach mitigates the domain gap between the visual and textual modalities, thereby enhancing the performance of the CLIP model in downstream tasks. Initially, as shown in Fig. 2", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we integrate the visual and textual description sets to form a comprehensive input description set, which is then processed by the CLIP text encoder. Meanwhile, to further enhance alignment, we expand the output dimension of the final learnable linear projection layer in CLIP, enabling it to generate more expressive and informative textual features.", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {D e s c r i p t i o n} = \\operatorname {C o n c a t} \\left(T _ {\\text {s c e n e}}, T _ {\\text {o b j e c t s}}, V _ {\\text {s c e n e}}, V _ {\\text {o b j e c t s}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 196, + 906, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {T e x t} = \\text {C L I P T e x t E n c o d e r (D e s c r i p t i o n)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 244, + 906, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively transfer fine-grained semantic features from the visual description set to the textual description set, we designed a text decoder. As illustrated in Fig. 3, the text decoder first employs the Text-to-Text module to refine and extract textual information. Then, based on the correlation between image features and textual descriptions, it leverages the Visual-to-Text module to dynamically map visual information onto textual descriptions. This process facilitates the recovery of semantic details relevant to the image content, thereby enhancing the expressiveness and informativeness of the textual descriptions.", + "bbox": [ + 511, + 272, + 906, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\text {P r o m p t}} = \\text {T e x t D e c o d e r (V i s u a l , T e x t)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 457, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, in order to recover the visual information from the set of text descriptions encoded by CLIP text encoder, we designed a visual decoder. As illustrated in Fig. 3, the decoder effectively extracts the visual features of the background through the Visual-to-Visual module. It then leverages the Text-to-Visual module to perform cross-modal conversion, capturing the semantic relationships between the image and textual descriptions. This process enhances the mutual understanding and alignment between visual and textual modalities, improving the coherence and expressiveness of the descriptions.", + "bbox": [ + 511, + 483, + 905, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} _ {\\text {P r o m p t}} = \\text {V i s u a l D e c o d e r} (\\text {V i s u a l}, \\text {T e x t}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 669, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Function: In cross-modal tasks, textual and visual description sets come from different modalities, and their feature spaces may differ significantly. To enable effective use of both modalities, it's crucial to align their distributions in a shared semantic space. The Maximum Mean Discrepancy (MMD)[2, 27] measures the squared distance between the means of two distributions $X_{S}$ and $X_{T}$ in the feature space:", + "bbox": [ + 511, + 695, + 905, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mathcal {H}} \\left(X _ {S}, X _ {T}\\right) = \\sup _ {\\| \\phi \\| _ {\\mathcal {H}} \\leq 1} \\| \\mathbb {E} _ {S} [ \\phi (X _ {S}) ] - \\mathbb {E} _ {T} [ \\phi (X _ {T}) ] \\| _ {\\mathcal {H}} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 829, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi$ is the feature map, and $\\mathbb{E}_S[\\phi (X_S)]$ and $\\mathbb{E}_T[\\phi (X_T)]$ are the expected values of the mapped samples. To align", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9552", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "textual and visual prompts, we propose an MMD loss:", + "bbox": [ + 89, + 90, + 449, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} (D _ {\\mathcal {H}} (X _ {S}, X _ {T})) = \\frac {1}{| X _ {n} ^ {s} | ^ {2}} \\sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {s} |} k \\Big (x _ {i} ^ {(s)}, x _ {j} ^ {(s)} \\Big) \\\\ - \\frac {2}{| X _ {n} ^ {s} | | X _ {n} ^ {t} |} \\sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \\left(x _ {i} ^ {(s)}, x _ {j} ^ {(t)}\\right) \\\\ + \\frac {1}{| X _ {n} ^ {t} | ^ {2}} \\sum_ {i = 1} ^ {| X _ {n} ^ {t} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \\left(x _ {i} ^ {(t)}, x _ {j} ^ {(t)}\\right). \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 112, + 482, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual prompt and Visual prompt feature. The samples $x_{i}^{(s)}$ and $x_{j}^{(s)}$ represent the feature descriptions in the source domain $X_{S}$ , and these features are used to represent the Text prompt. Similarly, $x_{i}^{(t)}$ and $x_{j}^{(t)}$ represent the feature descriptions in the target domain $X_{T}$ for the Visual prompt. These features are utilized in the MMD calculation to minimize the distributional gap between textual and visual representations, optimizing the alignment of both modalities for improved performance in cross-modal matching and generation tasks.", + "bbox": [ + 89, + 268, + 482, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Kernel function for similarity calculation. $k(x_{i},x_{j})$ calculates the similarity between samples $x_{i}$ and $x_{j}$ . The kernel function is used to compute similarity, which helps in measuring the relationships between samples in the source and target domains.", + "bbox": [ + 89, + 422, + 483, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Number of samples in the source and target domain. $|X_{n}^{s}|$ , $|X_{n}^{t}|$ denotes the number of samples in the source domain and target domain. It is used to normalize the number of samples, ensuring that the influence of each sample is balanced in the calculation.", + "bbox": [ + 89, + 498, + 483, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimizing this loss reduces the distributional gap between the textual and visual prompt, aligning the information for better performance in image-text matching.", + "bbox": [ + 89, + 589, + 483, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. CG-SAM", + "text_level": 1, + "bbox": [ + 89, + 642, + 200, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Infrared small targets often exhibit limited visual features, while the background is typically filled with significant clutter and noise. In this context, the SAM encoder faces an inherent trade-off between providing detailed target information and suppressing background noise. Specifically, the early layers of the ViT architecture retain more local details, which is crucial for recognizing small targets with limited visual features. However, these early layers also preserve a substantial amount of background noise, increasing the complexity of target detection. Therefore, directly using the Mask Decoder to recover clear target masks from complex backgrounds presents a significant challenge.", + "bbox": [ + 89, + 664, + 483, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Infrared Imaging Equation", + "text_level": 1, + "bbox": [ + 89, + 851, + 328, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve SAM's performance in IRSTD, we propose a novel approach based on the fundamental principles of in-", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frared imaging. The Beer-Lambert law [23, 28], a fundamental principle in optics and spectroscopy, is widely applied in the study of various light transmission problems. Ideally, the transmission of infrared radiation from a target to a sensor can be described by the Beer-Lambert law. Thus, by drawing on the Beer-Lambert law and integrating it with the core principles of infrared imaging, we propose a novel approach to model the propagation of infrared radiation and its effects on imaging, especially in scenarios where background interference leads to degradation of the observed target. In the basic model of infrared imaging, the fundamental equation is defined as:", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI (z) = J (z) \\cdot t (z) + A (z) \\cdot (1 - t (z)), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 279, + 906, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where:", + "bbox": [ + 513, + 303, + 563, + 315 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $I(z)$ : The pixel intensity of the infrared observation image at position $z$ , which includes background radiation and interference.", + "- $J(z)$ : The true infrared radiation intensity of the target at position $z$ , representing the target's brightness in the infrared spectrum.", + "- $t(z)$ : The transmission factor, representing the energy attenuation of infrared radiation from the target to the observer, with values in the range [0, 1].", + "- $A(z)$ : The background radiation refers to the infrared intensity from the background at position $z$ ." + ], + "bbox": [ + 513, + 316, + 903, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The definition of the transmission factor is given by:", + "bbox": [ + 513, + 484, + 859, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt (z) = e ^ {- \\beta \\cdot d (z)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 503, + 903, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where:", + "bbox": [ + 513, + 529, + 563, + 542 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\beta$ : The infrared radiation coefficient, determining the absorption level of infrared radiation by different components. This value typically varies depending on the infrared wavelength (e.g., near-infrared, shortwave infrared, or mid-to-far infrared).", + "- $d(z)$ : The scene depth, representing the distance between the target and the imaging device. In infrared imaging, a greater distance results in lower transmission, as infrared radiation undergoes greater loss in the atmosphere and other media." + ], + "bbox": [ + 513, + 544, + 903, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a 2D image, we need a function to calculate its corresponding depth. This process can be expressed as:", + "bbox": [ + 511, + 695, + 903, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd (z) = F (I (z)). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 732, + 903, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the scene's varying distances, classification techniques can be employed. Cao et al.[3] treated the depth estimation problem as a pixel-level classification task.", + "bbox": [ + 511, + 755, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF (I (z)) = \\operatorname {R e s N e t} (I (x)). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 815, + 903, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the above equations, we can calculate the infrared target's radiation intensity:", + "bbox": [ + 511, + 835, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ (z) = \\frac {I (z)}{t (z)} - \\frac {A (z)}{t (z)} + A (z). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 869, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9553", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Substituting the transmission factor yields:", + "bbox": [ + 109, + 90, + 393, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nJ (z) = I (z) \\cdot e ^ {\\beta \\cdot F (I (z))} - A (z) \\cdot \\left(e ^ {\\beta \\cdot F (I (z))} - 1\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 111, + 482, + 130 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through the formulations outlined above, we can more effectively separate the target from the background, thereby improving the performance of infrared small target detection. This approach constructs an interpretable mathematical model based on the fundamental principles of infrared imaging, providing a robust and explainable solution for infrared small target detection.", + "bbox": [ + 89, + 136, + 482, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2. CG-SAM", + "text_level": 1, + "bbox": [ + 89, + 247, + 202, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the infrared imaging equation (Eq. 14), we achieve precise infrared small target detection by inversely separating infrared degraded images from background images. As shown in Fig. 2, the text_prompt and visual_prompt are input into the SAM mask decoder, where they are fused with the early and late global semantic feature maps extracted by the ViT encoder. Building upon this, text_prompt and visual_prompt effectively extract and identify background features $A(z)$ from different scenes through the scene prompting mechanism. Similarly, we feed the learnable output token from SAM and the image embeddings from the SAM encoder into the mask decoder to further refine the target mask representation. This process gradually recovers target features, but due to background interference, it ultimately generates the infrared degraded image $I(z)$ with background noise.", + "bbox": [ + 89, + 267, + 482, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, based on the degraded image $I(z)$ and the background features $A(z)$ , we apply the infrared imaging equation to achieve precise detection of infrared small targets in complex backgrounds while effectively suppressing background interference. Through this integrated process, the CG-SAM can accurately identify small targets and significantly improve the accuracy of target detection in infrared images with strong background noise.", + "bbox": [ + 89, + 508, + 482, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Datasets", + "text_level": 1, + "bbox": [ + 89, + 637, + 192, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The goal of data generation is to create descriptive text for infrared images to support small target detection in complex backgrounds. To achieve this, we enhanced the IRSTD-1k, NUAA-SIRST, and NUDT-SIRST datasets by offering textual descriptions, resulting in the multimodal image-text dataset collections MIRSTD. Each infrared image, with the target mask removed, is encoded as a Base64 string and used as a text prompt for a visual-language model. ChatGPT generates concise descriptions that summarize key scene information, which are checked manually. The expert's prompt follows the format: a photo of a scene with [object], [object], and [object] offering a brief description of the scene and its objects. The final output is a 25-word description that captures the image's essential background information.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 114, + 720, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: We evaluate our model on the MIRSTD dataset, which consists of 427 and 1,000 real infrared images with one or more small targets, and 1,327 synthetic infrared images with small targets. All images are resized to $1024 \\times 1024$ . For each dataset, we split the images into three disjoint subsets: $50\\%$ for training, $30\\%$ for validation, and $20\\%$ for testing.", + "bbox": [ + 511, + 137, + 903, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: Our SAIST model is trained using the Adam optimizer with an initial learning rate of 0.0001, and a Cosine Decay Learning Rate Scheduler over 400 epochs. We use a batch size of 4 and train on a single Nvidia GeForce A800 GPU with 80 GB of memory. For comparison, we evaluate against state-of-the-art CNN-based methods for IRSTD, including ISNet [39], GCINet [41], UIUNet [33], DNANet [19], Dim2Clear [40], ALCNet [8], and ACMNet [7], as well as hybrid methods such as SCTransNet [35], IAANet [31], SAM [16] and SAM-HQ [15]. Traditional methods Top-Hat [1] and IPI [10] are also included. We adopt Intersection over Union (IoU) for pixel-level evaluation, and Probability of Detection (Pd) and False Alarm Rate (Fa) for object-level evaluation.", + "bbox": [ + 511, + 243, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparative Experiments", + "text_level": 1, + "bbox": [ + 511, + 465, + 751, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results: As shown in Tab. 1, traditional methods struggle with complex scenarios, especially CNNs, which fail to distinguish targets from backgrounds, resulting in higher false alarm and missed detection rates. The proposed SAIST outperforms existing SOTA methods across all metrics, particularly in reducing false alarms. SR-CLIP improves textual prompt generation through visual-textual alignment, while CG-SAM leverages both SR-CLIP's guidance and infrared imaging equations to enhance small target recognition in complex backgrounds.", + "bbox": [ + 511, + 488, + 905, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Results: Fig. 4 shows detection results for SR-CLIP and CG-SAM, outperforming other IRSTD methods. SR-CLIP benefits from precise visual and textual guidance, while CG-SAM effectively integrates infrared imaging equations. Both methods excel in identifying small targets in complex backgrounds, reducing $F_{a}$ and background noise, thus enhancing detection accuracy and robustness.", + "bbox": [ + 511, + 640, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ROC: As shown in Fig. 5, our SAIST significantly outperforms the other methods, where the Area Under the ROC Curve (AUC) of our SAIST is larger than that of the traditional and deep learning based methods.", + "bbox": [ + 511, + 746, + 903, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 816, + 666, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the components of SAIST, we conducted ablation studies on the NUAA-SIRST dataset, using CLIP + SAM as the baseline model. We systematically introduced the CG-SAM and SR-CLIP modules. As shown in the", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9554", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4dfb73c28be454c58d3113c4a255a58091801bb1df9f8f3296c9771721057c80.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNUDT-SIRSTIRSTD-1kNUAA-SIRSTType
IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓
IPI [10]30.9381.9817.9927.9281.3716.181.0987.0530467Traditional
Top-Hat [1]22.4089.90174.110.0675.1114321.50879.7416456Traditional
MSLSTIPT [26]8.3447.4088111.4379.0315241.0800.0528183Traditional
ACM [7]68.9097.0511.2962.4191.4435.5870.7793.089.33CNN
ALCNet [8]81.4096.519.2662.0590.5821.7874.3173.1220.21CNN
FC3-Net [38]78.5693.8623.92265.0791.5415.5572.4498.1410.85CNN
ISNet [39]81.7796.344.4768.7795.5615.3980.0299.024.61CNN
DNA-Net [19]88.1998.629.0069.3893.311.6679.2698.485.35CNN
UIU-Net [33]92.1997.7715.4469.9693.9822.0778.2597.454.29CNN
Dim2Clear[40]81.3796.239.1766.3493.7520.9377.2999.106.72CNN
GCI-Net [41]92.4398.258.9667.7593.8912.8478.8199.344.11CNN
IAANet [31]90.2297.268.3266.2593.1514.2074.2293.5322.70Hybrid
SCTransNet [35]94.0996.954.2968.0393.2710.7477.5096.9513.92Hybrid
SAM [16]84.2097.3813.2868.1293.648.3875.2197.079.78SAM
SAM-HQ [15]88.0298.3114.4868.8593.549.5675.2797.226.87SAM
SAIST (Ours)95.2399.281.3172.1496.184.7680.8299.560.87CLIP + SAM
", + "bbox": [ + 171, + 88, + 826, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison with existing IRSTD approaches on the NUDT-SIRST, IRSTD-1k and NUAA-SIRST datasets. The evaluation metrics are IoU $(10^{-2})$ , $P_d$ $(10^{-2})$ and $F_a$ $(10^{-6})$ , the best results are highlighted.", + "bbox": [ + 89, + 349, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d96c6c8b8b965a304eef7db628288b70e95e8bddac893540e0cfc8e45487107.jpg", + "image_caption": [ + "Figure 4. Visual results of different IRSTD methods. The boxes in red, yellow, and blue represent correct, missed, and false detections, respectively. Close-up views are shown in the corners." + ], + "image_footnote": [], + "bbox": [ + 99, + 382, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1, removing CG-SAM significantly reduces the IoU and Pd scores while increasing the Fa score. Additionally, as illustrated in the Fig. 7, compared to other methods, CG-SAM can precisely identify small targets in complex backgrounds. These results validate the effectiveness of CG-SAM in enhancing small target representation and suppressing background interference. Similarly, removing SR-CLIP increases the Fa score. Meanwhile, Fig. 6 demonstrates that SR-CLIP effectively suppresses background interference and noise. These results suggest that SR-CLIP gen", + "bbox": [ + 89, + 742, + 480, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "erates accurate visual-textual prompts for CG-SAM. Ultimately, as shown in Table. 3, although SAIST does not have an advantage in inference time and the number of trainable parameters, it achieves the best performance across all evaluation metrics, demonstrating the excellence of its design.", + "bbox": [ + 511, + 743, + 906, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1. Impact of the SR-CLIP", + "text_level": 1, + "bbox": [ + 511, + 833, + 723, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this ablation study, we provide identical text prompts to CLIP, CoOp, and SR-CLIP, utilizing their respective vision and text encoders for image understanding and caption gen", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9555", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfe7a2b668ffabef3593e520a8dd0f6ac35b767ddae03f6a5838912d9f621362.jpg", + "image_caption": [ + "Figure 5. ROC curves on NUAA-SIRST database." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 460, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e702909692955b2b2b67b9aea117129b26a46fa79c68cfc1799b601c5df843a7.jpg", + "image_caption": [ + "A photo of tower crane A photo of a wall" + ], + "image_footnote": [], + "bbox": [ + 93, + 357, + 166, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19ced1e25a62092afca593c20a4afc9dd36b4a340ee179e2211b13a2b29a2e6d.jpg", + "image_caption": [ + "CLIP" + ], + "image_footnote": [], + "bbox": [ + 169, + 357, + 238, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/675c6d3dab60bd9efd15c93a298b96d9cab9c38f7a3ff67c8d0c0d7d74a0db52.jpg", + "image_caption": [ + "CoOP" + ], + "image_footnote": [], + "bbox": [ + 243, + 357, + 321, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d44ec6f18c5ddd409b739eb8b4077a410887ee4e56270ed29fa4d20ae15d961e.jpg", + "image_caption": [ + "SR-CLIP" + ], + "image_footnote": [], + "bbox": [ + 323, + 357, + 395, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6394df9ca117440a0bc7933bded9da623484a4a71010364615f35ba004cf4731.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 357, + 475, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d01c28366d9cc90b9b54b0ddd7f82c554a57fac39e2750848b8c98ce30f140d1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 415, + 166, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b4f1b0b6103cb42bbc6e60f0c3c1b10eb4a816b246b4961949cf4ac21516995.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 415, + 241, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad1f2031f3a9d88e9628bd2926ef9353f5fd0f1777988b6a9a750ed8fc60c990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 415, + 321, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6bb55e26c3ec5a69d8ef0eebcef5a8abfc4d8689238ba02d2bdd51a376cb319.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 415, + 395, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da05fe9a18f62eb3ab9e3ddef3ee523cb8e110cc9a3eec9aaf9326a186528bda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 414, + 475, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89f2d620ad127f7e123b498f4892c36c068f100f01e347b321e720cff8e450d0.jpg", + "image_caption": [ + "Figure 6. Visual comparison of CLIP, CoOp and SR-CLIP." + ], + "image_footnote": [], + "bbox": [ + 94, + 491, + 169, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3578ce996adb9a9f918e0fecfc2a0c9d62d798839c51ecbf9bd9843f4aa72ddf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 173, + 491, + 248, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e12787b2d4a3710072db5241d5f49fb0267389a0ac999a16c1feed588f9995d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 491, + 325, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c9345d8838947389a8bdd9cd4d9407712c2661c1122c1023788ab796cc4bf90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 491, + 397, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6291f93359b9f51d5f85b57127e6dd492c31470f2fed12f190decc93e0df023.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 491, + 475, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b94aec4c3f1ed7f92b0a5b20232eae067c1dadb7d0f3a27bc5596116ed2bc42.jpg", + "image_caption": [ + "Figure 7. Visual comparison of SAM, HQ-SAM and CG-SAM." + ], + "image_footnote": [], + "bbox": [ + 94, + 542, + 169, + 583 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8dc122fa5ce9c4e2b9ef0be4463f77dd8cdc700e7c387c5774c6c272d36aab9.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 173, + 542, + 248, + 583 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25b87a03c11f31397f13730562f1edda62012238a344a47e6d568dd1e4121eeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 542, + 325, + 583 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d5e002284ffe5308c687460b78a40594678cbc24666e72bc1e8b0d917ec2252.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 542, + 397, + 583 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f94aa8fe9f0d285296037f813fc0f85d2233b149acce2a92916231c5c758633a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 542, + 475, + 583 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96bb796f5a291b6b725c2a946ace03e33b22dcc4f7efdf29013d4b7e13799de6.jpg", + "table_caption": [ + "Table 2. Ablation study of SR-CLIP and CG-SAM in $IoU(\\%)$ , $P_d(\\%)$ , $F_a(10^{-6})$ on on NUAA-SIRST." + ], + "table_footnote": [], + "table_body": "
MethodIoUPdFa
SAIST80.8299.560.87
w/o SR-CLIP80.0298.776.61
w/o CG-SAM78.8797.428.53
w/o CG-SAM & SR-CLIP76.8296.5012.37
", + "bbox": [ + 130, + 661, + 439, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eration, to systematically evaluate the performance of SR-CLIP. As shown in Fig. 6 and the Tab. 4, we demonstrate that SR-CLIP effectively suppresses background interference and noise, enabling the generation of precise visual-textual prompts for CG-SAM. These results highlight the efficacy of SR-CLIP in improving the alignment between the visual and textual modalities.", + "bbox": [ + 89, + 738, + 482, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Impact of the CG-SAM", + "text_level": 1, + "bbox": [ + 89, + 851, + 302, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this ablation study, we utilize the same SAM encoder while employing different Mask Decoders: SAM, HQ-", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0d3bec773a67fd9e20a61c72dcc105355e7f66cd14e631b2360ecd51636ca63f.jpg", + "table_caption": [ + "Table 3. Comparison of different models in $IoU(\\%)$ , $P_d(\\%)$ , $F_a(10^{-6})$ on MIRSTD." + ], + "table_footnote": [], + "table_body": "
ModelTrainable Params (M)Inference (s)Avg IoUAvg PdAvg Fa
SAM + CLIP383.280.07978.4996.678.45
SAIST389.570.08182.7398.342.31
", + "bbox": [ + 514, + 125, + 903, + 157 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99b4b75587a5be8876901e585cb88111bcb2280ec028deb80dd7ec9b6291e82a.jpg", + "table_caption": [ + "Table 4. Ablation study of CLIP, CoOp, and SR-CLIP in $IoU(\\%)$ , $P_d(\\%)$ , $F_a(10^{-6})$ on NUAA-SIRST." + ], + "table_footnote": [], + "table_body": "
MethodIoUPdFa
CLIP + CG-SAM80.0298.776.61
CoOp + CG-SAM80.3298.564.87
SR-CLIP + CG-SAM80.8299.560.87
", + "bbox": [ + 566, + 205, + 849, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3b181f07a215539ecb8bb8fdc9bdb96f35b0ed3a4167ebf4e7d005384a80b123.jpg", + "table_caption": [ + "Table 5. Ablation study of SAM, HQ-SAM, and CG-SAM in $IoU(\\%)$ , $P_d(\\%)$ , $F_a(10^{-6})$ on NUAA-SIRST." + ], + "table_footnote": [], + "table_body": "
MethodIoUPdFa
SAM + SR-CLIP78.8797.425.23
HQ-SAM + SR-CLIP79.4598.894.56
CG-SAM + SR-CLIP80.8299.560.87
", + "bbox": [ + 566, + 306, + 849, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SAM, and CG-SAM. Identical fine-grained visual and textual prompts, generated by SR-CLIP, are provided as inputs. The performance of CG-SAM is systematically evaluated. As demonstrated in Fig. 7 and the Tab. 5, our results confirm that CG-SAM is capable of accurately identifying small targets in complex backgrounds. These findings validate that, with the aid of infrared imaging equations, CG-SAM enhances the representation of small targets while effectively suppressing background interference.", + "bbox": [ + 511, + 375, + 906, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 527, + 633, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents SAIST, a novel framework that integrates visual and textual modalities for enhanced IRSTD. Leveraging SR-CLIP to capture cross-modal relationships, SAIST generates precise prompts to overcome the limitations of image-only methods. Guided by these prompts, CG-SAM employs interpretable infrared imaging equations to extract targets while suppressing noise. We also introduce MIRSTD, the first multimodal IRSTD dataset with rich image-text pairs. Experiments demonstrate the effectiveness of multimodal integration in IRSTD.", + "bbox": [ + 511, + 554, + 906, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 723, + 692, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62272363, Grant 92470108, Grant 62371362, and Grant U21A20514; in part by the Joint Laboratory for Innovation in Satellite-Borne Computers and Electronics Technology Open Fund 2023 under Grant 2024KFKT001-1; in part by the Proof of Concept Foundation of Xidian University Hangzhou of Technology under Grant No. GNYZ2023YL0301; and part by the Shaanxi Provincial Key R&D Program Project under Grant 2024SF-YBXM-330.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9556", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xiangzhi Bai and Fugen Zhou. Analysis of new top-hat transformation and the application for infrared dim small target detection. Pattern Recognition, 43(6):2145-2156, 2010. 1, 2, 6, 7", + "[2] Karsten M Borgwardt, Arthur Gretton, Malte J Rasch, Hans-Peter Kriegel, Bernhard Scholkopf, and Alex J Smola. Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49-e57, 2006. 4", + "[3] Yuanzhouhan Cao, Zifeng Wu, and Chunhua Shen. Estimating depth from monocular images as classification using deep fully convolutional residual networks. IEEE Transactions on Circuits and Systems for Video Technology, 28(11): 3174-3182, 2018. 5", + "[4] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE transactions on geoscience and remote sensing, 52(1):574-581, 2013. 2", + "[5] G Cucurullo, L Giordano, D Albanese, Luciano Cinquanta, and M Di Matteo. Infrared thermography assisted control for apples microwave drying. Journal of food engineering, 112 (4):319-325, 2012. 1", + "[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE journal of selected topics in applied earth observations and remote sensing, 10(8): 3752-3767, 2017. 1, 2", + "[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 950–959, 2021. 1, 2, 6, 7", + "[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 59(11):9813-9824, 2021. 6, 7", + "[9] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G. Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing, 22(12):4996-5009, 2013. 1", + "[10] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 6, 7", + "[11] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multistage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024. 2", + "[12] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters, 18(9):1670-1674, 2020. 1, 2", + "[13] MD. Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learn" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing for image captioning. ACM Comput. Surv., 51(6), 2019. 3", + "[14] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3", + "[15] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 6, 7", + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 6, 7", + "[17] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 2", + "[18] Wing-Cheung Law, Zhourui Xu, Ken-Tye Yong, Xin Liu, Mark T Swihart, Mukund Seshadri, and Paras N Prasad. Manganese-doped near-infrared emitting nanocrystals for in vivo biomedical imaging. Optics express, 24(16):17553-17561, 2016. 1", + "[19] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 2022. 2, 6, 7", + "[20] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2", + "[21] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1, 2", + "[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2", + "[23] Alexandre Mallet, Roumiana Tsenkova, Jelena Muncan, Cyrille Charnier, Éric Latrille, Ryad Bendoula, Jean-Philippe Steyer, and Jean-Michel Roger. Relating near-infrared light path-length modifications to the water content of scattering media in near-infrared spectroscopy: toward a new bouguer-beer-lambert law. Analytical Chemistry, 93 (17):6817-6823, 2021. 5", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[25] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9557", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2", + "[26] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing, 59(5):3737-3752, 2021. 7", + "[27] Danica J Sutherland, Hsiao-Yu Tung, Heiko Strathmann, Soumyajit De, Aaditya Ramdas, Alex Smola, and Arthur Gretton. Generative models and model criticism via optimized maximum mean discrepancy. arXiv preprint arXiv:1611.04488, 2016. 4", + "[28] Donald F Swinehart. The beer-lambert law. Journal of chemical education, 39(7):333, 1962. 5", + "[29] Michael Teutsch and Wolfgang Krüger. Classification of small boats in infrared images for maritime surveillance. In 2010 International WaterSide Security Conference, pages 1-7. IEEE, 2010. 1", + "[30] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8509-8518, 2019. 1, 2", + "[31] Kewei Wang, Shuaiyuan Du, Chengxin Liu, and Zhiguo Cao. Interior attention-aware network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2022. 1, 2, 6, 7", + "[32] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2", + "[33] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 6, 7", + "[34] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3", + "[35] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 6, 7", + "[36] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 2", + "[37] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1730–1738, 2022. 1, 2", + "[38] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1857-1865, 2022. 1, 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 2, 6, 7", + "[40] Mingjin Zhang, Rui Zhang, Jing Zhang, Jie Guo, Yunsong Li, and Xinbo Gao. Dim2clear network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-14, 2023. 2, 6, 7", + "[41] Mingjin Zhang, Ke Yue, Boyang Li, Jie Guo, Yunsong Li, and Xinbo Gao. Single-frame infrared small target detection via gaussian curvature inspired network. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2, 6, 7", + "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4", + "[43] Hu Zhu, Haopeng Ni, Shiming Liu, Guoxia Xu, and Lizhen Deng. TnIrs: Target-aware non-local low-rank modeling with saliency filtering regularization for infrared small target detection. IEEE Transactions on Image Processing, 29: 9546-9558, 2020. 1", + "[44] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 444 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9558", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_model.json b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8ca6e4560509369dc206270a984b2f3f7f6a1a --- /dev/null +++ b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_model.json @@ -0,0 +1,2563 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.877, + 0.178 + ], + "angle": 0, + "content": "SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.203, + 0.87, + 0.294 + ], + "angle": 0, + "content": "Mingjin Zhang\\(^{1}\\), Xiaolong Li\\(^{1*}\\), Fei Gao\\(^{1}\\), Jie Guo\\(^{1*}\\), Xinbo Gao\\(^{1,2}\\), Jing Zhang\\(^{3}\\) \n\\(^{1}\\)Xidian University, China \\(^{2}\\)Chongqing University of Posts and Telecommunications, China \n\\(^{3}\\)School of Computer Science & School of Artificial Intelligence, Wuhan University, China \n{mjinzhang, fgao}@xidian.edu.cn, 23011210739@stu.xidian.edu.cn \njguo, xbgao@mail.xidian.edu.cn, jingzhang.csv@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.328, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Infrared Small Target Detection (IRSTD) aims to identify low signal-to-noise ratio small targets in infrared images with complex backgrounds, which is crucial for various applications. However, existing IRSTD methods typically rely solely on image modalities for processing, which fail to fully capture contextual information, leading to limited detection accuracy and adaptability in complex environments. Inspired by vision-language models, this paper proposes a novel framework, SAIST, which integrates textual information with image modalities to enhance IRSTD performance. The framework consists of two main components: Scene Recognition Contrastive Language-Image Pretraining (SR-CLIP) and CLIP-guided Segment Anything Model (CG-SAM). SR-CLIP generates a set of visual descriptions through object-object similarity and object-scene relevance, embedding them into learnable prompts to refine the textual description set. This reduces the domain gap between vision and language, generating precise textual and visual prompts. CG-SAM utilizes the prompts generated by SR-CLIP to accurately guide the Mask Decoder in learning prior knowledge of background features, while incorporating infrared imaging equations to improve small target recognition in complex backgrounds and significantly reduce the false alarm rate. Additionally, this paper introduces the first multimodal IRSTD dataset, MIRSTD, which contains abundant image-text pairs. Experimental results demonstrate that the proposed SAIST method outperforms existing state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.224, + 0.832 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.842, + 0.485, + 0.874 + ], + "angle": 0, + "content": "Infrared small target detection (IRSTD) is crucial for applications like traffic management and maritime rescue" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.326, + 0.887, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.579 + ], + "angle": 0, + "content": "Figure 1. Illustration of the advantages of text-guided detection over image-only detection in identifying small infrared targets in complex backgrounds. The textual description offers valuable context, helping to reduce interference and distinguish the target from the background effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.719 + ], + "angle": 0, + "content": "[5, 18, 29]. Unlike visible light, infrared sensors detect thermal radiation, enabling target acquisition in adverse conditions such as fog, rain, snow, and at night [44]. However, due to the long-range imaging mechanism, infrared images often lack distinctive features like color, shape, and texture, while the background typically shows strong structural features with higher intensity and sharper contours [9, 43]. These challenges—complex backgrounds, clutter, and noise—hinder accurate detection, complicating IRSTD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.854 + ], + "angle": 0, + "content": "IRSTD methods can be divided into traditional algorithms and deep learning-based ones. Early methods focused on filtering and image enhancement [1, 6, 12], which performed well in simpler scenarios but struggled with complex environments due to reliance on prior knowledge and handcrafted features. With the rise of deep learning and large IRSTD datasets, newer approaches using Generative Adversarial Networks (GANs) [30], U-Net [7, 21, 38], and transformers [31, 35, 37] have been introduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Despite advances, existing methods are limited by their reliance on single-modality inputs. Recently, the CLIP-SAM model, which combines Contrastive Language-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.24, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "Image Pretraining (CLIP) and the Segment Anything Model (SAM), has shown promise in enhancing cross-modal feature fusion for various tasks [20]. CLIP [24] captures semantic relationships between images and text, offering rich supervision, while SAM [16] provides precise segmentation. However, applying CLIP-SAM to IRSTD remains challenging due to the differences between infrared and natural images. Specifically: 1) Infrared targets lack typical visual features, limiting the generalization of the CLIP model; 2) Small, weak infrared targets and noisy backgrounds complicate target detection, making SAM's application difficult. To address these issues, we propose using CLIP to extract semantically rich prior knowledge from text, guiding SAM to better capture target details amidst complex backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.64 + ], + "angle": 0, + "content": "This paper introduces a novel IRSTD approach named SAIST by combining text and image modalities within a multimodal learning framework. First, we design the Scene Recognition CLIP (SR-CLIP) to capture both global scene and local object features. SR-CLIP creates similarity matrices for objects and scenes, facilitating key information extraction. It also integrates a cross-modal interaction mechanism, incorporating scene-specific visual cues into learnable prompts that refine text features and mitigate the vision-language domain gap. Building on this, we diverse CG-SAM to utilize the text prompts from SR-CLIP to guide SAM in learning background features, enhancing accurate identification and extraction across diverse scenes. By combining global semantic and background features, and leveraging the radiative transfer properties of targets and backgrounds, CG-SAM achieves effective target-background separation. Additionally, we present MIRSTD, the first multimodal IRSTD dataset, which combines imagedataset pairs from NUAA-SIRST, NUDT-SIRST, and IRSTD1K datasets. Text descriptions are generated by GPT-4V and carefully reviewed manually. Experiments validate the superiority of SAIST over state-of-the-art (SOTA) methods." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.638, + 0.449, + 0.652 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.654, + 0.483, + 0.715 + ], + "angle": 0, + "content": "- We introduce SAIST, the first approach to integrate multimodal vision and language for IRSTD. On challenging benchmarks, it outperforms SOTA methods, excelling in both objective metrics and subjective evaluations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.715, + 0.483, + 0.759 + ], + "angle": 0, + "content": "- We design two key components, SR-CLIP and CG-SAM, to effectively fuse multimodal information and separate targets and backgrounds in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.76, + 0.483, + 0.805 + ], + "angle": 0, + "content": "- We create a novel multimodal IRSTD dataset combining infrared images and textual data, offering rich resources for training and evaluating multimodal learning methods." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.654, + 0.483, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.822, + 0.228, + 0.837 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.848, + 0.363, + 0.864 + ], + "angle": 0, + "content": "2.1. Infrared small target detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Traditional IRSTD methods use image processing or handcrafted features, such as low-rank (e.g., RIFT [6]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.395 + ], + "angle": 0, + "content": "PSTNN [36]), human visual system-based (e.g., TLLCM [4], WSLCM [12]), and filter-based approaches (e.g., tophat [1]). Low-rank methods perform well in low SNR but struggle with false alarms in low-contrast areas. Human visual methods are effective with large objects but not in complex scenes, and filter-based techniques suppress uniform noise but fail in complex backgrounds. Deep learning IRSTD methods include CNN-based models (e.g., MDvsFA [30], ISNet [39], Dim2Clear [40], CSRNet [21], DMFNet [11], GCI-Net [41]) that focus on local features, with some improving SNR, detection, or small target extraction, and hybrid models (e.g., IAANet [31] and RKformer [37]) that combine CNNs and Vision Transformers (ViTs) to capture both local and global context. While IRSTD has made significant strides, current methods rely mainly on image modalities, which fail to capture full contextual information, limiting detection accuracy in complex backgrounds. Integrating textual information can offer substantial benefits, providing richer scene understanding and aiding target extraction in challenging environments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.783, + 0.421 + ], + "angle": 0, + "content": "2.2. CLIP and SAM Collaboration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.685 + ], + "angle": 0, + "content": "In recent years, foundational models like CLIP and SAM have garnered significant attention for their exceptional performance across various tasks. CLIP excels at aligning multimodal features, while SAM specializes in segmentation through diverse prompts. Recent research has explored synergies between these models, such as CRIS [32] integrates both models to boost performance. ClipSAM [20] uses CLIP's semantic understanding for anomaly localization and coarse segmentation, later refined by SAM. OpenVocabulary SAM enhances SAM's recognition by transferring knowledge from SAM to CLIP via distillation and adapters. However, applying pre-trained CLIP-SAM models, originally trained on natural images, to IRSTD faces challenges due to domain differences. In this work, we leverage CLIP's semantic understanding from scene descriptions to guide SAM in extracting target information from complex infrared backgrounds." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.697, + 0.702, + 0.712 + ], + "angle": 0, + "content": "2.3. Datasets for IRSTD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Prior studies have developed large-scale datasets [7, 19, 39], which have advanced IRSTD research. However, these datasets only include image data with target masks, limiting the detection of small targets in complex backgrounds. Recent multimodal approaches, such as text-guided detection [17, 22, 25], have shown performance improvements. Inspired by this, we create the MIRSTD dataset, which combines both image and text modalities. The text modality in MIRSTD provides descriptions of scenes in infrared images from datasets like NUAA-SIRST, IRSTD-1k, and NUTDSIRST, helping to mitigate the challenge of weak target signals against strong background noise." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9550" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.485, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.906, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed encoder-decoder architecture. The CLIP text encoder converts textual descriptions into feature representations, while the frozen image encoder processes infrared images without target. SR-CLIP leverages cross-modal interactions to refine textual features, bridging the visual-textual domain gap. The SAM image encoder extracts image embeddings, fine-tuned with LoRA to enhance task adaptability. The generated textual prompts and image embeddings are fed into the SAM mask decoder, along with an infrared imaging model, to effectively suppress background noise and detect small targets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.226, + 0.429 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.437, + 0.287, + 0.451 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.702 + ], + "angle": 0, + "content": "The overall architecture of SAIST is illustrated in Fig. 2. In this framework, the text encoder of SR-CLIP converts textual descriptions into feature representations, while the image encoder is frozen. Next, SR-CLIP employs a cross-modal interaction mechanism to integrate the visual and textual description sets, refining the textual features and reducing the domain gap between vision and language, thus generating precise textual and visual prompts. Subsequently, the pre-trained and frozen SAM image encoder extracts image embeddings, which are fine-tuned using Low-Rank Adaptation (LoRA) [14] to enhance the model's adaptability to the specific task. Finally, the generated textual prompts and image embeddings are fed into the SAM mask decoder, in conjunction with a specially designed infrared imaging model, to effectively suppress background noise and accurately detect small targets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.2, + 0.726 + ], + "angle": 0, + "content": "3.2. SR-CLIP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In infrared imaging, due to the inherent characteristics of remote sensing, the target's radiative intensity is typically weak, while the background's intensity is higher and more prominent. In this case, the information related to the target is difficult to describe clearly and accurately. In contrast, the background's radiative intensity is more stable and easier to identify, allowing the overall structure and features of the scene to be effectively reflected through intuitive textual descriptions. There, we design SR-CLIP to fully leverage the complementary relationship between background textual descriptions and visual features, generating precise" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.413, + 0.907, + 0.444 + ], + "angle": 0, + "content": "prompts that guide SAM to effectively suppress background interference in infrared images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.753, + 0.466 + ], + "angle": 0, + "content": "3.2.1. Visual / Text description set" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.591 + ], + "angle": 0, + "content": "For the scene recognition task in infrared images, scenes are composed of multiple semantic concepts, including distinct objects and backgrounds. These features necessitate the consideration of various semantic information when learning discriminative patterns. To capture comprehensive scene information, we introduce a novel form of textual prompting, referred to as scene prompts. This prompt creates a textual description set:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.6, + 0.902, + 0.617 + ], + "angle": 0, + "content": "{aphotof[scene]with[object],[object],and[object]}," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.718 + ], + "angle": 0, + "content": "which provides both scene-level and object-level textual descriptions. Simultaneously, to bridge the image-text domain gap, we extract local object features from Clip's image encoder, acquire global visual features with the help of a learnable projection layer, and generate a set of visual feature descriptions using the image captioning [13, 34]." + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.728, + 0.838, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\left[ \\text {s c e n e} \\right], \\left[ \\text {o b j e c t} \\right], \\left[ \\text {o b j e c t} \\right] \\left[ \\text {o b j e c t} \\right] \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.755, + 0.906, + 0.785 + ], + "angle": 0, + "content": "which provides both scene-level and object-level visual descriptions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.688, + 0.807 + ], + "angle": 0, + "content": "3.2.2. Image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In scene recognition, objects often share considerable similarity, and there is a strong correlation between objects and the scene, which reflects the unique characteristics of that particular scene. By integrating both interactions, the model can capture more granular semantic information, thereby improving its discriminative power in specific contexts. To" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9551" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.487, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.485, + 0.364 + ], + "angle": 0, + "content": "Figure 3. Overview of the Cross-modal Interaction Mechanism. Image captioning generates visual feature descriptions that are processed by a text encoder. The text decoder then maps visual information into the text, enhancing text feature expressiveness. Finally, the visual decoder recovers visual information from the encoded text to improve image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.41 + ], + "angle": 0, + "content": "this end, we construct both an object-object similarity matrix and a scene-object correlation matrix:" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.411, + 0.351, + 0.427 + ], + "angle": 0, + "content": "Object-Object Similarity Matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.435, + 0.483, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c c c} o _ {1 1} & \\dots & o _ {1 n} \\\\ \\vdots & \\ddots & \\vdots \\\\ o _ {n 1} & \\dots & o _ {n n} \\end{array} \\right] = \\frac {\\exp \\left(\\frac {o _ {i} ^ {\\top} o _ {j}}{\\| o _ {i} \\| \\| o _ {j} \\|}\\right)}{\\sum_ {k = 1} ^ {n} \\exp \\left(\\frac {o _ {i} ^ {\\top} o _ {k}}{\\| o _ {i} \\| \\| o _ {k} \\|}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.532 + ], + "angle": 0, + "content": "where \\(o_{i,j}\\) represents the similarity between the feature vector of object \\(i\\) and object \\(j\\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.532, + 0.353, + 0.547 + ], + "angle": 0, + "content": "Scene-Object Correlation Matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.556, + 0.483, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l l l} s _ {1} & \\dots & s _ {n} \\end{array} \\right] = \\frac {\\exp \\left(\\frac {s ^ {\\top} o _ {i}}{\\| s \\| \\| o _ {i} \\|}\\right)}{\\sum_ {k = 1} ^ {n} \\exp \\left(\\frac {s ^ {\\top} o _ {k}}{\\| s \\| \\| o _ {k} \\|}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.648 + ], + "angle": 0, + "content": "where \\( s_i \\) represents the correlation between the feature vector of the scene and the feature vector of object \\( j \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.74 + ], + "angle": 0, + "content": "Therefore, based on the similarity and correlation matrices, we design the Object-to-Object and Scene-to-Object modules to effectively capture the interrelationships between objects and facilitate the integration of scene and object details. Meanwhile, by incorporating image captioning techniques, we generate precise visual description." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.395, + 0.761 + ], + "angle": 0, + "content": "3.2.3. Cross-Modal Interaction Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "CoOp [42] can be viewed as a textual domain prompt, where the learnable context is simply a trainable vector without incorporating any visual information. Inspired by CoOp, we propose a cross-modal interaction mechanism that introduces the visual description set into the text description set to refine the textual features. This approach mitigates the domain gap between the visual and textual modalities, thereby enhancing the performance of the CLIP model in downstream tasks. Initially, as shown in Fig. 2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "we integrate the visual and textual description sets to form a comprehensive input description set, which is then processed by the CLIP text encoder. Meanwhile, to further enhance alignment, we expand the output dimension of the final learnable linear projection layer in CLIP, enabling it to generate more expressive and informative textual features." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.197, + 0.907, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\text {D e s c r i p t i o n} = \\operatorname {C o n c a t} \\left(T _ {\\text {s c e n e}}, T _ {\\text {o b j e c t s}}, V _ {\\text {s c e n e}}, V _ {\\text {o b j e c t s}}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.246, + 0.907, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\text {T e x t} = \\text {C L I P T e x t E n c o d e r (D e s c r i p t i o n)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.44 + ], + "angle": 0, + "content": "To effectively transfer fine-grained semantic features from the visual description set to the textual description set, we designed a text decoder. As illustrated in Fig. 3, the text decoder first employs the Text-to-Text module to refine and extract textual information. Then, based on the correlation between image features and textual descriptions, it leverages the Visual-to-Text module to dynamically map visual information onto textual descriptions. This process facilitates the recovery of semantic details relevant to the image content, thereby enhancing the expressiveness and informativeness of the textual descriptions." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.458, + 0.907, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\text {P r o m p t}} = \\text {T e x t D e c o d e r (V i s u a l , T e x t)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.906, + 0.651 + ], + "angle": 0, + "content": "Meanwhile, in order to recover the visual information from the set of text descriptions encoded by CLIP text encoder, we designed a visual decoder. As illustrated in Fig. 3, the decoder effectively extracts the visual features of the background through the Visual-to-Visual module. It then leverages the Text-to-Visual module to perform cross-modal conversion, capturing the semantic relationships between the image and textual descriptions. This process enhances the mutual understanding and alignment between visual and textual modalities, improving the coherence and expressiveness of the descriptions." + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.67, + 0.907, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} _ {\\text {P r o m p t}} = \\text {V i s u a l D e c o d e r} (\\text {V i s u a l}, \\text {T e x t}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.803 + ], + "angle": 0, + "content": "Loss Function: In cross-modal tasks, textual and visual description sets come from different modalities, and their feature spaces may differ significantly. To enable effective use of both modalities, it's crucial to align their distributions in a shared semantic space. The Maximum Mean Discrepancy (MMD)[2, 27] measures the squared distance between the means of two distributions \\( X_{S} \\) and \\( X_{T} \\) in the feature space:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.83, + 0.905, + 0.871 + ], + "angle": 0, + "content": "\\[\nD _ {\\mathcal {H}} \\left(X _ {S}, X _ {T}\\right) = \\sup _ {\\| \\phi \\| _ {\\mathcal {H}} \\leq 1} \\| \\mathbb {E} _ {S} [ \\phi (X _ {S}) ] - \\mathbb {E} _ {T} [ \\phi (X _ {T}) ] \\| _ {\\mathcal {H}} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\phi\\) is the feature map, and \\(\\mathbb{E}_S[\\phi (X_S)]\\) and \\(\\mathbb{E}_T[\\phi (X_T)]\\) are the expected values of the mapped samples. To align" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9552" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.45, + 0.107 + ], + "angle": 0, + "content": "textual and visual prompts, we propose an MMD loss:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.113, + 0.483, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} (D _ {\\mathcal {H}} (X _ {S}, X _ {T})) = \\frac {1}{| X _ {n} ^ {s} | ^ {2}} \\sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {s} |} k \\Big (x _ {i} ^ {(s)}, x _ {j} ^ {(s)} \\Big) \\\\ - \\frac {2}{| X _ {n} ^ {s} | | X _ {n} ^ {t} |} \\sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \\left(x _ {i} ^ {(s)}, x _ {j} ^ {(t)}\\right) \\\\ + \\frac {1}{| X _ {n} ^ {t} | ^ {2}} \\sum_ {i = 1} ^ {| X _ {n} ^ {t} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \\left(x _ {i} ^ {(t)}, x _ {j} ^ {(t)}\\right). \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Textual prompt and Visual prompt feature. The samples \\( x_{i}^{(s)} \\) and \\( x_{j}^{(s)} \\) represent the feature descriptions in the source domain \\( X_{S} \\), and these features are used to represent the Text prompt. Similarly, \\( x_{i}^{(t)} \\) and \\( x_{j}^{(t)} \\) represent the feature descriptions in the target domain \\( X_{T} \\) for the Visual prompt. These features are utilized in the MMD calculation to minimize the distributional gap between textual and visual representations, optimizing the alignment of both modalities for improved performance in cross-modal matching and generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Kernel function for similarity calculation. \\( k(x_{i},x_{j}) \\) calculates the similarity between samples \\( x_{i} \\) and \\( x_{j} \\). The kernel function is used to compute similarity, which helps in measuring the relationships between samples in the source and target domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Number of samples in the source and target domain. \\( |X_{n}^{s}| \\), \\( |X_{n}^{t}| \\) denotes the number of samples in the source domain and target domain. It is used to normalize the number of samples, ensuring that the influence of each sample is balanced in the calculation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Minimizing this loss reduces the distributional gap between the textual and visual prompt, aligning the information for better performance in image-text matching." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.202, + 0.658 + ], + "angle": 0, + "content": "3.3. CG-SAM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.847 + ], + "angle": 0, + "content": "Infrared small targets often exhibit limited visual features, while the background is typically filled with significant clutter and noise. In this context, the SAM encoder faces an inherent trade-off between providing detailed target information and suppressing background noise. Specifically, the early layers of the ViT architecture retain more local details, which is crucial for recognizing small targets with limited visual features. However, these early layers also preserve a substantial amount of background noise, increasing the complexity of target detection. Therefore, directly using the Mask Decoder to recover clear target masks from complex backgrounds presents a significant challenge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.329, + 0.868 + ], + "angle": 0, + "content": "3.3.1. Infrared Imaging Equation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To improve SAM's performance in IRSTD, we propose a novel approach based on the fundamental principles of in-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "frared imaging. The Beer-Lambert law [23, 28], a fundamental principle in optics and spectroscopy, is widely applied in the study of various light transmission problems. Ideally, the transmission of infrared radiation from a target to a sensor can be described by the Beer-Lambert law. Thus, by drawing on the Beer-Lambert law and integrating it with the core principles of infrared imaging, we propose a novel approach to model the propagation of infrared radiation and its effects on imaging, especially in scenarios where background interference leads to degradation of the observed target. In the basic model of infrared imaging, the fundamental equation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.28, + 0.907, + 0.298 + ], + "angle": 0, + "content": "\\[\nI (z) = J (z) \\cdot t (z) + A (z) \\cdot (1 - t (z)), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.304, + 0.565, + 0.316 + ], + "angle": 0, + "content": "where:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.318, + 0.905, + 0.361 + ], + "angle": 0, + "content": "- \\( I(z) \\): The pixel intensity of the infrared observation image at position \\( z \\), which includes background radiation and interference." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.363, + 0.905, + 0.408 + ], + "angle": 0, + "content": "- \\( J(z) \\): The true infrared radiation intensity of the target at position \\( z \\), representing the target's brightness in the infrared spectrum." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.409, + 0.905, + 0.453 + ], + "angle": 0, + "content": "- \\( t(z) \\): The transmission factor, representing the energy attenuation of infrared radiation from the target to the observer, with values in the range [0, 1]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.454, + 0.905, + 0.484 + ], + "angle": 0, + "content": "- \\( A(z) \\): The background radiation refers to the infrared intensity from the background at position \\( z \\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.318, + 0.905, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.485, + 0.86, + 0.5 + ], + "angle": 0, + "content": "The definition of the transmission factor is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.504, + 0.905, + 0.523 + ], + "angle": 0, + "content": "\\[\nt (z) = e ^ {- \\beta \\cdot d (z)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.53, + 0.565, + 0.543 + ], + "angle": 0, + "content": "where:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.545, + 0.905, + 0.62 + ], + "angle": 0, + "content": "- \\(\\beta\\): The infrared radiation coefficient, determining the absorption level of infrared radiation by different components. This value typically varies depending on the infrared wavelength (e.g., near-infrared, shortwave infrared, or mid-to-far infrared)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.62, + 0.905, + 0.695 + ], + "angle": 0, + "content": "- \\(d(z)\\): The scene depth, representing the distance between the target and the imaging device. In infrared imaging, a greater distance results in lower transmission, as infrared radiation undergoes greater loss in the atmosphere and other media." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.545, + 0.905, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.696, + 0.905, + 0.726 + ], + "angle": 0, + "content": "Given a 2D image, we need a function to calculate its corresponding depth. This process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.733, + 0.905, + 0.75 + ], + "angle": 0, + "content": "\\[\nd (z) = F (I (z)). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.907, + 0.802 + ], + "angle": 0, + "content": "Considering the scene's varying distances, classification techniques can be employed. Cao et al.[3] treated the depth estimation problem as a pixel-level classification task." + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.816, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\nF (I (z)) = \\operatorname {R e s N e t} (I (x)). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.836, + 0.905, + 0.866 + ], + "angle": 0, + "content": "Using the above equations, we can calculate the infrared target's radiation intensity:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.871, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nJ (z) = \\frac {I (z)}{t (z)} - \\frac {A (z)}{t (z)} + A (z). \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9553" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.395, + 0.106 + ], + "angle": 0, + "content": "Substituting the transmission factor yields:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.112, + 0.483, + 0.131 + ], + "angle": 0, + "content": "\\[\nJ (z) = I (z) \\cdot e ^ {\\beta \\cdot F (I (z))} - A (z) \\cdot \\left(e ^ {\\beta \\cdot F (I (z))} - 1\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.244 + ], + "angle": 0, + "content": "Through the formulations outlined above, we can more effectively separate the target from the background, thereby improving the performance of infrared small target detection. This approach constructs an interpretable mathematical model based on the fundamental principles of infrared imaging, providing a robust and explainable solution for infrared small target detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.248, + 0.204, + 0.262 + ], + "angle": 0, + "content": "3.3.2. CG-SAM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.509 + ], + "angle": 0, + "content": "Based on the infrared imaging equation (Eq. 14), we achieve precise infrared small target detection by inversely separating infrared degraded images from background images. As shown in Fig. 2, the text_prompt and visual_prompt are input into the SAM mask decoder, where they are fused with the early and late global semantic feature maps extracted by the ViT encoder. Building upon this, text_prompt and visual_prompt effectively extract and identify background features \\( A(z) \\) from different scenes through the scene prompting mechanism. Similarly, we feed the learnable output token from SAM and the image embeddings from the SAM encoder into the mask decoder to further refine the target mask representation. This process gradually recovers target features, but due to background interference, it ultimately generates the infrared degraded image \\( I(z) \\) with background noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Finally, based on the degraded image \\( I(z) \\) and the background features \\( A(z) \\), we apply the infrared imaging equation to achieve precise detection of infrared small targets in complex backgrounds while effectively suppressing background interference. Through this integrated process, the CG-SAM can accurately identify small targets and significantly improve the accuracy of target detection in infrared images with strong background noise." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.194, + 0.652 + ], + "angle": 0, + "content": "3.4. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The goal of data generation is to create descriptive text for infrared images to support small target detection in complex backgrounds. To achieve this, we enhanced the IRSTD-1k, NUAA-SIRST, and NUDT-SIRST datasets by offering textual descriptions, resulting in the multimodal image-text dataset collections MIRSTD. Each infrared image, with the target mask removed, is encoded as a Base64 string and used as a text prompt for a visual-language model. ChatGPT generates concise descriptions that summarize key scene information, which are checked manually. The expert's prompt follows the format: a photo of a scene with [object], [object], and [object] offering a brief description of the scene and its objects. The final output is a 25-word description that captures the image's essential background information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.107 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.115, + 0.722, + 0.132 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.243 + ], + "angle": 0, + "content": "Dataset: We evaluate our model on the MIRSTD dataset, which consists of 427 and 1,000 real infrared images with one or more small targets, and 1,327 synthetic infrared images with small targets. All images are resized to \\(1024 \\times 1024\\). For each dataset, we split the images into three disjoint subsets: \\(50\\%\\) for training, \\(30\\%\\) for validation, and \\(20\\%\\) for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Implementation Details: Our SAIST model is trained using the Adam optimizer with an initial learning rate of 0.0001, and a Cosine Decay Learning Rate Scheduler over 400 epochs. We use a batch size of 4 and train on a single Nvidia GeForce A800 GPU with 80 GB of memory. For comparison, we evaluate against state-of-the-art CNN-based methods for IRSTD, including ISNet [39], GCINet [41], UIUNet [33], DNANet [19], Dim2Clear [40], ALCNet [8], and ACMNet [7], as well as hybrid methods such as SCTransNet [35], IAANet [31], SAM [16] and SAM-HQ [15]. Traditional methods Top-Hat [1] and IPI [10] are also included. We adopt Intersection over Union (IoU) for pixel-level evaluation, and Probability of Detection (Pd) and False Alarm Rate (Fa) for object-level evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.753, + 0.483 + ], + "angle": 0, + "content": "4.2. Comparative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Quantitative results: As shown in Tab. 1, traditional methods struggle with complex scenarios, especially CNNs, which fail to distinguish targets from backgrounds, resulting in higher false alarm and missed detection rates. The proposed SAIST outperforms existing SOTA methods across all metrics, particularly in reducing false alarms. SR-CLIP improves textual prompt generation through visual-textual alignment, while CG-SAM leverages both SR-CLIP's guidance and infrared imaging equations to enhance small target recognition in complex backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.746 + ], + "angle": 0, + "content": "Visual Results: Fig. 4 shows detection results for SR-CLIP and CG-SAM, outperforming other IRSTD methods. SR-CLIP benefits from precise visual and textual guidance, while CG-SAM effectively integrates infrared imaging equations. Both methods excel in identifying small targets in complex backgrounds, reducing \\( F_{a} \\) and background noise, thus enhancing detection accuracy and robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.905, + 0.807 + ], + "angle": 0, + "content": "ROC: As shown in Fig. 5, our SAIST significantly outperforms the other methods, where the Area Under the ROC Curve (AUC) of our SAIST is larger than that of the traditional and deep learning based methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.667, + 0.834 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To evaluate the components of SAIST, we conducted ablation studies on the NUAA-SIRST dataset, using CLIP + SAM as the baseline model. We systematically introduced the CG-SAM and SR-CLIP modules. As shown in the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9554" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.172, + 0.089, + 0.828, + 0.339 + ], + "angle": 0, + "content": "
MethodNUDT-SIRSTIRSTD-1kNUAA-SIRSTType
IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓
IPI [10]30.9381.9817.9927.9281.3716.181.0987.0530467Traditional
Top-Hat [1]22.4089.90174.110.0675.1114321.50879.7416456Traditional
MSLSTIPT [26]8.3447.4088111.4379.0315241.0800.0528183Traditional
ACM [7]68.9097.0511.2962.4191.4435.5870.7793.089.33CNN
ALCNet [8]81.4096.519.2662.0590.5821.7874.3173.1220.21CNN
FC3-Net [38]78.5693.8623.92265.0791.5415.5572.4498.1410.85CNN
ISNet [39]81.7796.344.4768.7795.5615.3980.0299.024.61CNN
DNA-Net [19]88.1998.629.0069.3893.311.6679.2698.485.35CNN
UIU-Net [33]92.1997.7715.4469.9693.9822.0778.2597.454.29CNN
Dim2Clear[40]81.3796.239.1766.3493.7520.9377.2999.106.72CNN
GCI-Net [41]92.4398.258.9667.7593.8912.8478.8199.344.11CNN
IAANet [31]90.2297.268.3266.2593.1514.2074.2293.5322.70Hybrid
SCTransNet [35]94.0996.954.2968.0393.2710.7477.5096.9513.92Hybrid
SAM [16]84.2097.3813.2868.1293.648.3875.2197.079.78SAM
SAM-HQ [15]88.0298.3114.4868.8593.549.5675.2797.226.87SAM
SAIST (Ours)95.2399.281.3172.1496.184.7680.8299.560.87CLIP + SAM
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.35, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Table 1. Comparison with existing IRSTD approaches on the NUDT-SIRST, IRSTD-1k and NUAA-SIRST datasets. The evaluation metrics are IoU \\((10^{-2})\\), \\(P_d\\) \\((10^{-2})\\) and \\(F_a\\) \\((10^{-6})\\), the best results are highlighted." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.383, + 0.907, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.688, + 0.908, + 0.718 + ], + "angle": 0, + "content": "Figure 4. Visual results of different IRSTD methods. The boxes in red, yellow, and blue represent correct, missed, and false detections, respectively. Close-up views are shown in the corners." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.482, + 0.895 + ], + "angle": 0, + "content": "Tab. 1, removing CG-SAM significantly reduces the IoU and Pd scores while increasing the Fa score. Additionally, as illustrated in the Fig. 7, compared to other methods, CG-SAM can precisely identify small targets in complex backgrounds. These results validate the effectiveness of CG-SAM in enhancing small target representation and suppressing background interference. Similarly, removing SR-CLIP increases the Fa score. Meanwhile, Fig. 6 demonstrates that SR-CLIP effectively suppresses background interference and noise. These results suggest that SR-CLIP gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.82 + ], + "angle": 0, + "content": "erates accurate visual-textual prompts for CG-SAM. Ultimately, as shown in Table. 3, although SAIST does not have an advantage in inference time and the number of trainable parameters, it achieves the best performance across all evaluation metrics, demonstrating the excellence of its design." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.724, + 0.849 + ], + "angle": 0, + "content": "4.3.1. Impact of the SR-CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this ablation study, we provide identical text prompts to CLIP, CoOp, and SR-CLIP, utilizing their respective vision and text encoders for image understanding and caption gen" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.092, + 0.462, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.33, + 0.438, + 0.343 + ], + "angle": 0, + "content": "Figure 5. ROC curves on NUAA-SIRST database." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.358, + 0.168, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.399, + 0.169, + 0.415 + ], + "angle": 0, + "content": "A photo of tower crane A photo of a wall" + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.358, + 0.24, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.399, + 0.24, + 0.414 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.358, + 0.322, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.399, + 0.314, + 0.414 + ], + "angle": 0, + "content": "CLIP" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.358, + 0.397, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.399, + 0.397, + 0.414 + ], + "angle": 0, + "content": "CoOP" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.358, + 0.477, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.4, + 0.461, + 0.414 + ], + "angle": 0, + "content": "SR-CLIP" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.416, + 0.168, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.416, + 0.243, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.416, + 0.322, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.416, + 0.397, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.415, + 0.477, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.477, + 0.461, + 0.491 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of CLIP, CoOp and SR-CLIP." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.492, + 0.171, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.492, + 0.249, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.492, + 0.326, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.492, + 0.398, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.492, + 0.477, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.533, + 0.234, + 0.543 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.544, + 0.171, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.544, + 0.249, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.544, + 0.326, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.544, + 0.398, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.544, + 0.477, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.597, + 0.476, + 0.612 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of SAM, HQ-SAM and CG-SAM." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.626, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Table 2. Ablation study of SR-CLIP and CG-SAM in \\( IoU(\\%) \\), \\( P_d(\\%) \\), \\( F_a(10^{-6}) \\) on on NUAA-SIRST." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.662, + 0.441, + 0.728 + ], + "angle": 0, + "content": "
MethodIoUPdFa
SAIST80.8299.560.87
w/o SR-CLIP80.0298.776.61
w/o CG-SAM78.8797.428.53
w/o CG-SAM & SR-CLIP76.8296.5012.37
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.844 + ], + "angle": 0, + "content": "eration, to systematically evaluate the performance of SR-CLIP. As shown in Fig. 6 and the Tab. 4, we demonstrate that SR-CLIP effectively suppresses background interference and noise, enabling the generation of precise visual-textual prompts for CG-SAM. These results highlight the efficacy of SR-CLIP in improving the alignment between the visual and textual modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.303, + 0.867 + ], + "angle": 0, + "content": "4.3.2. Impact of the CG-SAM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this ablation study, we utilize the same SAM encoder while employing different Mask Decoders: SAM, HQ-" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 3. Comparison of different models in \\( IoU(\\%) \\), \\( P_d(\\%) \\), \\( F_a(10^{-6}) \\) on MIRSTD." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.126, + 0.905, + 0.158 + ], + "angle": 0, + "content": "
ModelTrainable Params (M)Inference (s)Avg IoUAvg PdAvg Fa
SAM + CLIP383.280.07978.4996.678.45
SAIST389.570.08182.7398.342.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.171, + 0.907, + 0.199 + ], + "angle": 0, + "content": "Table 4. Ablation study of CLIP, CoOp, and SR-CLIP in \\( IoU(\\%) \\), \\( P_d(\\%) \\), \\( F_a(10^{-6}) \\) on NUAA-SIRST." + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.207, + 0.851, + 0.26 + ], + "angle": 0, + "content": "
MethodIoUPdFa
CLIP + CG-SAM80.0298.776.61
CoOp + CG-SAM80.3298.564.87
SR-CLIP + CG-SAM80.8299.560.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.272, + 0.907, + 0.301 + ], + "angle": 0, + "content": "Table 5. Ablation study of SAM, HQ-SAM, and CG-SAM in \\( IoU(\\%) \\), \\( P_d(\\%) \\), \\( F_a(10^{-6}) \\) on NUAA-SIRST." + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.308, + 0.851, + 0.361 + ], + "angle": 0, + "content": "
MethodIoUPdFa
SAM + SR-CLIP78.8797.425.23
HQ-SAM + SR-CLIP79.4598.894.56
CG-SAM + SR-CLIP80.8299.560.87
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.512 + ], + "angle": 0, + "content": "SAM, and CG-SAM. Identical fine-grained visual and textual prompts, generated by SR-CLIP, are provided as inputs. The performance of CG-SAM is systematically evaluated. As demonstrated in Fig. 7 and the Tab. 5, our results confirm that CG-SAM is capable of accurately identifying small targets in complex backgrounds. These findings validate that, with the aid of infrared imaging equations, CG-SAM enhances the representation of small targets while effectively suppressing background interference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.634, + 0.544 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.707 + ], + "angle": 0, + "content": "This paper presents SAIST, a novel framework that integrates visual and textual modalities for enhanced IRSTD. Leveraging SR-CLIP to capture cross-modal relationships, SAIST generates precise prompts to overcome the limitations of image-only methods. Guided by these prompts, CG-SAM employs interpretable infrared imaging equations to extract targets while suppressing noise. We also introduce MIRSTD, the first multimodal IRSTD dataset with rich image-text pairs. Experiments demonstrate the effectiveness of multimodal integration in IRSTD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.693, + 0.741 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62272363, Grant 92470108, Grant 62371362, and Grant U21A20514; in part by the Joint Laboratory for Innovation in Satellite-Borne Computers and Electronics Technology Open Fund 2023 under Grant 2024KFKT001-1; in part by the Proof of Concept Foundation of Xidian University Hangzhou of Technology under Grant No. GNYZ2023YL0301; and part by the Shaanxi Provincial Key R&D Program Project under Grant 2024SF-YBXM-330." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9556" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Xiangzhi Bai and Fugen Zhou. Analysis of new top-hat transformation and the application for infrared dim small target detection. Pattern Recognition, 43(6):2145-2156, 2010. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Karsten M Borgwardt, Arthur Gretton, Malte J Rasch, Hans-Peter Kriegel, Bernhard Scholkopf, and Alex J Smola. Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49-e57, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Yuanzhouhan Cao, Zifeng Wu, and Chunhua Shen. Estimating depth from monocular images as classification using deep fully convolutional residual networks. IEEE Transactions on Circuits and Systems for Video Technology, 28(11): 3174-3182, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE transactions on geoscience and remote sensing, 52(1):574-581, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[5] G Cucurullo, L Giordano, D Albanese, Luciano Cinquanta, and M Di Matteo. Infrared thermography assisted control for apples microwave drying. Journal of food engineering, 112 (4):319-325, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE journal of selected topics in applied earth observations and remote sensing, 10(8): 3752-3767, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 950–959, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 59(11):9813-9824, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[9] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G. Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing, 22(12):4996-5009, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multistage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters, 18(9):1670-1674, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] MD. Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ing for image captioning. ACM Comput. Surv., 51(6), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[17] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[18] Wing-Cheung Law, Zhourui Xu, Ken-Tye Yong, Xin Liu, Mark T Swihart, Mukund Seshadri, and Paras N Prasad. Manganese-doped near-infrared emitting nanocrystals for in vivo biomedical imaging. Optics express, 24(16):17553-17561, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[19] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[20] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[21] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[23] Alexandre Mallet, Roumiana Tsenkova, Jelena Muncan, Cyrille Charnier, Éric Latrille, Ryad Bendoula, Jean-Philippe Steyer, and Jean-Michel Roger. Relating near-infrared light path-length modifications to the water content of scattering media in near-infrared spectroscopy: toward a new bouguer-beer-lambert law. Analytical Chemistry, 93 (17):6817-6823, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9557" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing, 59(5):3737-3752, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Danica J Sutherland, Hsiao-Yu Tung, Heiko Strathmann, Soumyajit De, Aaditya Ramdas, Alex Smola, and Arthur Gretton. Generative models and model criticism via optimized maximum mean discrepancy. arXiv preprint arXiv:1611.04488, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[28] Donald F Swinehart. The beer-lambert law. Journal of chemical education, 39(7):333, 1962. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[29] Michael Teutsch and Wolfgang Krüger. Classification of small boats in infrared images for maritime surveillance. In 2010 International WaterSide Security Conference, pages 1-7. IEEE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[30] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8509-8518, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[31] Kewei Wang, Shuaiyuan Du, Chengxin Liu, and Zhiguo Cao. Interior attention-aware network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.478, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[32] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.548, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[33] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.591, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[34] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.649, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[35] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[36] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[37] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1730–1738, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1857-1865, 2022. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[39] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Mingjin Zhang, Rui Zhang, Jing Zhang, Jie Guo, Yunsong Li, and Xinbo Gao. Dim2clear network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-14, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[41] Mingjin Zhang, Ke Yue, Boyang Li, Jie Guo, Yunsong Li, and Xinbo Gao. Single-frame infrared small target detection via gaussian curvature inspired network. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[43] Hu Zhu, Haopeng Ni, Shiming Liu, Guoxia Xu, and Lizhen Deng. TnIrs: Target-aware non-local low-rank modeling with saliency filtering regularization for infrared small target detection. IEEE Transactions on Image Processing, 29: 9546-9558, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[44] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9558" + } + ] +] \ No newline at end of file diff --git a/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_origin.pdf b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1393f7197c30252413c35e186f616894e8a758fb --- /dev/null +++ b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/230aeda6-22e8-4ab6-b4a7-19b2bc75fac3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d1181aeafc9bbd77cc2235f01d3cdbaa2fdcd0e091f214855bc95f2f4cc30a8 +size 2103572 diff --git a/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/full.md b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37e78219d742c0bbb4f169c585392a0bed51f7bf --- /dev/null +++ b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/full.md @@ -0,0 +1,393 @@ +# SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining + +Mingjin Zhang $^{1}$ , Xiaolong Li $^{1*}$ , Fei Gao $^{1}$ , Jie Guo $^{1*}$ , Xinbo Gao $^{1,2}$ , Jing Zhang $^{3}$ $^{1}$ Xidian University, China $^{2}$ Chongqing University of Posts and Telecommunications, China + $^{3}$ School of Computer Science & School of Artificial Intelligence, Wuhan University, China +{mjinzhang, fgao}@xidian.edu.cn, 23011210739@stu.xidian.edu.cn +jguo, xbgao@mail.xidian.edu.cn, jingzhang.csv@gmail.com + +# Abstract + +Infrared Small Target Detection (IRSTD) aims to identify low signal-to-noise ratio small targets in infrared images with complex backgrounds, which is crucial for various applications. However, existing IRSTD methods typically rely solely on image modalities for processing, which fail to fully capture contextual information, leading to limited detection accuracy and adaptability in complex environments. Inspired by vision-language models, this paper proposes a novel framework, SAIST, which integrates textual information with image modalities to enhance IRSTD performance. The framework consists of two main components: Scene Recognition Contrastive Language-Image Pretraining (SR-CLIP) and CLIP-guided Segment Anything Model (CG-SAM). SR-CLIP generates a set of visual descriptions through object-object similarity and object-scene relevance, embedding them into learnable prompts to refine the textual description set. This reduces the domain gap between vision and language, generating precise textual and visual prompts. CG-SAM utilizes the prompts generated by SR-CLIP to accurately guide the Mask Decoder in learning prior knowledge of background features, while incorporating infrared imaging equations to improve small target recognition in complex backgrounds and significantly reduce the false alarm rate. Additionally, this paper introduces the first multimodal IRSTD dataset, MIRSTD, which contains abundant image-text pairs. Experimental results demonstrate that the proposed SAIST method outperforms existing state-of-the-art approaches. + +# 1. Introduction + +Infrared small target detection (IRSTD) is crucial for applications like traffic management and maritime rescue + +![](images/4fba36fd60f51c873be7c05f05db2569a5dad0cfd6f2db93bed58be0c2ea4c6c.jpg) +Figure 1. Illustration of the advantages of text-guided detection over image-only detection in identifying small infrared targets in complex backgrounds. The textual description offers valuable context, helping to reduce interference and distinguish the target from the background effectively. + +[5, 18, 29]. Unlike visible light, infrared sensors detect thermal radiation, enabling target acquisition in adverse conditions such as fog, rain, snow, and at night [44]. However, due to the long-range imaging mechanism, infrared images often lack distinctive features like color, shape, and texture, while the background typically shows strong structural features with higher intensity and sharper contours [9, 43]. These challenges—complex backgrounds, clutter, and noise—hinder accurate detection, complicating IRSTD. + +IRSTD methods can be divided into traditional algorithms and deep learning-based ones. Early methods focused on filtering and image enhancement [1, 6, 12], which performed well in simpler scenarios but struggled with complex environments due to reliance on prior knowledge and handcrafted features. With the rise of deep learning and large IRSTD datasets, newer approaches using Generative Adversarial Networks (GANs) [30], U-Net [7, 21, 38], and transformers [31, 35, 37] have been introduced. + +Despite advances, existing methods are limited by their reliance on single-modality inputs. Recently, the CLIP-SAM model, which combines Contrastive Language- + +Image Pretraining (CLIP) and the Segment Anything Model (SAM), has shown promise in enhancing cross-modal feature fusion for various tasks [20]. CLIP [24] captures semantic relationships between images and text, offering rich supervision, while SAM [16] provides precise segmentation. However, applying CLIP-SAM to IRSTD remains challenging due to the differences between infrared and natural images. Specifically: 1) Infrared targets lack typical visual features, limiting the generalization of the CLIP model; 2) Small, weak infrared targets and noisy backgrounds complicate target detection, making SAM's application difficult. To address these issues, we propose using CLIP to extract semantically rich prior knowledge from text, guiding SAM to better capture target details amidst complex backgrounds. + +This paper introduces a novel IRSTD approach named SAIST by combining text and image modalities within a multimodal learning framework. First, we design the Scene Recognition CLIP (SR-CLIP) to capture both global scene and local object features. SR-CLIP creates similarity matrices for objects and scenes, facilitating key information extraction. It also integrates a cross-modal interaction mechanism, incorporating scene-specific visual cues into learnable prompts that refine text features and mitigate the vision-language domain gap. Building on this, we diverse CG-SAM to utilize the text prompts from SR-CLIP to guide SAM in learning background features, enhancing accurate identification and extraction across diverse scenes. By combining global semantic and background features, and leveraging the radiative transfer properties of targets and backgrounds, CG-SAM achieves effective target-background separation. Additionally, we present MIRSTD, the first multimodal IRSTD dataset, which combines imagedataset pairs from NUAA-SIRST, NUDT-SIRST, and IRSTD1K datasets. Text descriptions are generated by GPT-4V and carefully reviewed manually. Experiments validate the superiority of SAIST over state-of-the-art (SOTA) methods. + +In summary, our main contributions are as follows: + +- We introduce SAIST, the first approach to integrate multimodal vision and language for IRSTD. On challenging benchmarks, it outperforms SOTA methods, excelling in both objective metrics and subjective evaluations. +- We design two key components, SR-CLIP and CG-SAM, to effectively fuse multimodal information and separate targets and backgrounds in complex scenes. +- We create a novel multimodal IRSTD dataset combining infrared images and textual data, offering rich resources for training and evaluating multimodal learning methods. + +# 2. Related work + +# 2.1. Infrared small target detection + +Traditional IRSTD methods use image processing or handcrafted features, such as low-rank (e.g., RIFT [6], + +PSTNN [36]), human visual system-based (e.g., TLLCM [4], WSLCM [12]), and filter-based approaches (e.g., tophat [1]). Low-rank methods perform well in low SNR but struggle with false alarms in low-contrast areas. Human visual methods are effective with large objects but not in complex scenes, and filter-based techniques suppress uniform noise but fail in complex backgrounds. Deep learning IRSTD methods include CNN-based models (e.g., MDvsFA [30], ISNet [39], Dim2Clear [40], CSRNet [21], DMFNet [11], GCI-Net [41]) that focus on local features, with some improving SNR, detection, or small target extraction, and hybrid models (e.g., IAANet [31] and RKformer [37]) that combine CNNs and Vision Transformers (ViTs) to capture both local and global context. While IRSTD has made significant strides, current methods rely mainly on image modalities, which fail to capture full contextual information, limiting detection accuracy in complex backgrounds. Integrating textual information can offer substantial benefits, providing richer scene understanding and aiding target extraction in challenging environments. + +# 2.2. CLIP and SAM Collaboration + +In recent years, foundational models like CLIP and SAM have garnered significant attention for their exceptional performance across various tasks. CLIP excels at aligning multimodal features, while SAM specializes in segmentation through diverse prompts. Recent research has explored synergies between these models, such as CRIS [32] integrates both models to boost performance. ClipSAM [20] uses CLIP's semantic understanding for anomaly localization and coarse segmentation, later refined by SAM. OpenVocabulary SAM enhances SAM's recognition by transferring knowledge from SAM to CLIP via distillation and adapters. However, applying pre-trained CLIP-SAM models, originally trained on natural images, to IRSTD faces challenges due to domain differences. In this work, we leverage CLIP's semantic understanding from scene descriptions to guide SAM in extracting target information from complex infrared backgrounds. + +# 2.3. Datasets for IRSTD + +Prior studies have developed large-scale datasets [7, 19, 39], which have advanced IRSTD research. However, these datasets only include image data with target masks, limiting the detection of small targets in complex backgrounds. Recent multimodal approaches, such as text-guided detection [17, 22, 25], have shown performance improvements. Inspired by this, we create the MIRSTD dataset, which combines both image and text modalities. The text modality in MIRSTD provides descriptions of scenes in infrared images from datasets like NUAA-SIRST, IRSTD-1k, and NUTDSIRST, helping to mitigate the challenge of weak target signals against strong background noise. + +![](images/188d26818404e83d33891b2183ffbf1284b1fc20818ed0467c2e13364219a0c1.jpg) +Figure 2. Overview of the proposed encoder-decoder architecture. The CLIP text encoder converts textual descriptions into feature representations, while the frozen image encoder processes infrared images without target. SR-CLIP leverages cross-modal interactions to refine textual features, bridging the visual-textual domain gap. The SAM image encoder extracts image embeddings, fine-tuned with LoRA to enhance task adaptability. The generated textual prompts and image embeddings are fed into the SAM mask decoder, along with an infrared imaging model, to effectively suppress background noise and detect small targets. + +![](images/87092d3acab5275f7887276516369979861d590fd40b72b0f416b13cea58645c.jpg) + +# 3. Methodology + +# 3.1. Overall Architecture + +The overall architecture of SAIST is illustrated in Fig. 2. In this framework, the text encoder of SR-CLIP converts textual descriptions into feature representations, while the image encoder is frozen. Next, SR-CLIP employs a cross-modal interaction mechanism to integrate the visual and textual description sets, refining the textual features and reducing the domain gap between vision and language, thus generating precise textual and visual prompts. Subsequently, the pre-trained and frozen SAM image encoder extracts image embeddings, which are fine-tuned using Low-Rank Adaptation (LoRA) [14] to enhance the model's adaptability to the specific task. Finally, the generated textual prompts and image embeddings are fed into the SAM mask decoder, in conjunction with a specially designed infrared imaging model, to effectively suppress background noise and accurately detect small targets. + +# 3.2. SR-CLIP + +In infrared imaging, due to the inherent characteristics of remote sensing, the target's radiative intensity is typically weak, while the background's intensity is higher and more prominent. In this case, the information related to the target is difficult to describe clearly and accurately. In contrast, the background's radiative intensity is more stable and easier to identify, allowing the overall structure and features of the scene to be effectively reflected through intuitive textual descriptions. There, we design SR-CLIP to fully leverage the complementary relationship between background textual descriptions and visual features, generating precise + +prompts that guide SAM to effectively suppress background interference in infrared images. + +# 3.2.1. Visual / Text description set + +For the scene recognition task in infrared images, scenes are composed of multiple semantic concepts, including distinct objects and backgrounds. These features necessitate the consideration of various semantic information when learning discriminative patterns. To capture comprehensive scene information, we introduce a novel form of textual prompting, referred to as scene prompts. This prompt creates a textual description set: + +{aphotof[scene]with[object],[object],and[object]}, + +which provides both scene-level and object-level textual descriptions. Simultaneously, to bridge the image-text domain gap, we extract local object features from Clip's image encoder, acquire global visual features with the help of a learnable projection layer, and generate a set of visual feature descriptions using the image captioning [13, 34]. + +$$ +\left\{\left[ \text {s c e n e} \right], \left[ \text {o b j e c t} \right], \left[ \text {o b j e c t} \right] \left[ \text {o b j e c t} \right] \right\}, +$$ + +which provides both scene-level and object-level visual descriptions. + +# 3.2.2. Image Captioning + +In scene recognition, objects often share considerable similarity, and there is a strong correlation between objects and the scene, which reflects the unique characteristics of that particular scene. By integrating both interactions, the model can capture more granular semantic information, thereby improving its discriminative power in specific contexts. To + +![](images/b939f16afc102edb76b83a6db7b66ef9e50933e872a869cf306af8abc1410ee2.jpg) +Figure 3. Overview of the Cross-modal Interaction Mechanism. Image captioning generates visual feature descriptions that are processed by a text encoder. The text decoder then maps visual information into the text, enhancing text feature expressiveness. Finally, the visual decoder recovers visual information from the encoded text to improve image-text alignment. + +this end, we construct both an object-object similarity matrix and a scene-object correlation matrix: + +Object-Object Similarity Matrix: + +$$ +\left[ \begin{array}{c c c} o _ {1 1} & \dots & o _ {1 n} \\ \vdots & \ddots & \vdots \\ o _ {n 1} & \dots & o _ {n n} \end{array} \right] = \frac {\exp \left(\frac {o _ {i} ^ {\top} o _ {j}}{\| o _ {i} \| \| o _ {j} \|}\right)}{\sum_ {k = 1} ^ {n} \exp \left(\frac {o _ {i} ^ {\top} o _ {k}}{\| o _ {i} \| \| o _ {k} \|}\right)}, \tag {1} +$$ + +where $o_{i,j}$ represents the similarity between the feature vector of object $i$ and object $j$ . + +Scene-Object Correlation Matrix: + +$$ +\left[ \begin{array}{l l l} s _ {1} & \dots & s _ {n} \end{array} \right] = \frac {\exp \left(\frac {s ^ {\top} o _ {i}}{\| s \| \| o _ {i} \|}\right)}{\sum_ {k = 1} ^ {n} \exp \left(\frac {s ^ {\top} o _ {k}}{\| s \| \| o _ {k} \|}\right)}, \tag {2} +$$ + +where $s_i$ represents the correlation between the feature vector of the scene and the feature vector of object $j$ . + +Therefore, based on the similarity and correlation matrices, we design the Object-to-Object and Scene-to-Object modules to effectively capture the interrelationships between objects and facilitate the integration of scene and object details. Meanwhile, by incorporating image captioning techniques, we generate precise visual description. + +# 3.2.3. Cross-Modal Interaction Mechanism + +CoOp [42] can be viewed as a textual domain prompt, where the learnable context is simply a trainable vector without incorporating any visual information. Inspired by CoOp, we propose a cross-modal interaction mechanism that introduces the visual description set into the text description set to refine the textual features. This approach mitigates the domain gap between the visual and textual modalities, thereby enhancing the performance of the CLIP model in downstream tasks. Initially, as shown in Fig. 2 + +we integrate the visual and textual description sets to form a comprehensive input description set, which is then processed by the CLIP text encoder. Meanwhile, to further enhance alignment, we expand the output dimension of the final learnable linear projection layer in CLIP, enabling it to generate more expressive and informative textual features. + +$$ +\text {D e s c r i p t i o n} = \operatorname {C o n c a t} \left(T _ {\text {s c e n e}}, T _ {\text {o b j e c t s}}, V _ {\text {s c e n e}}, V _ {\text {o b j e c t s}}\right), \tag {3} +$$ + +$$ +\text {T e x t} = \text {C L I P T e x t E n c o d e r (D e s c r i p t i o n)}. \tag {4} +$$ + +To effectively transfer fine-grained semantic features from the visual description set to the textual description set, we designed a text decoder. As illustrated in Fig. 3, the text decoder first employs the Text-to-Text module to refine and extract textual information. Then, based on the correlation between image features and textual descriptions, it leverages the Visual-to-Text module to dynamically map visual information onto textual descriptions. This process facilitates the recovery of semantic details relevant to the image content, thereby enhancing the expressiveness and informativeness of the textual descriptions. + +$$ +\mathcal {T} _ {\text {P r o m p t}} = \text {T e x t D e c o d e r (V i s u a l , T e x t)}, \tag {5} +$$ + +Meanwhile, in order to recover the visual information from the set of text descriptions encoded by CLIP text encoder, we designed a visual decoder. As illustrated in Fig. 3, the decoder effectively extracts the visual features of the background through the Visual-to-Visual module. It then leverages the Text-to-Visual module to perform cross-modal conversion, capturing the semantic relationships between the image and textual descriptions. This process enhances the mutual understanding and alignment between visual and textual modalities, improving the coherence and expressiveness of the descriptions. + +$$ +\mathcal {V} _ {\text {P r o m p t}} = \text {V i s u a l D e c o d e r} (\text {V i s u a l}, \text {T e x t}). \tag {6} +$$ + +Loss Function: In cross-modal tasks, textual and visual description sets come from different modalities, and their feature spaces may differ significantly. To enable effective use of both modalities, it's crucial to align their distributions in a shared semantic space. The Maximum Mean Discrepancy (MMD)[2, 27] measures the squared distance between the means of two distributions $X_{S}$ and $X_{T}$ in the feature space: + +$$ +D _ {\mathcal {H}} \left(X _ {S}, X _ {T}\right) = \sup _ {\| \phi \| _ {\mathcal {H}} \leq 1} \| \mathbb {E} _ {S} [ \phi (X _ {S}) ] - \mathbb {E} _ {T} [ \phi (X _ {T}) ] \| _ {\mathcal {H}} ^ {2}, \tag {7} +$$ + +where $\phi$ is the feature map, and $\mathbb{E}_S[\phi (X_S)]$ and $\mathbb{E}_T[\phi (X_T)]$ are the expected values of the mapped samples. To align + +textual and visual prompts, we propose an MMD loss: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} (D _ {\mathcal {H}} (X _ {S}, X _ {T})) = \frac {1}{| X _ {n} ^ {s} | ^ {2}} \sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \sum_ {j = 1} ^ {| X _ {n} ^ {s} |} k \Big (x _ {i} ^ {(s)}, x _ {j} ^ {(s)} \Big) \\ - \frac {2}{| X _ {n} ^ {s} | | X _ {n} ^ {t} |} \sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \left(x _ {i} ^ {(s)}, x _ {j} ^ {(t)}\right) \\ + \frac {1}{| X _ {n} ^ {t} | ^ {2}} \sum_ {i = 1} ^ {| X _ {n} ^ {t} |} \sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \left(x _ {i} ^ {(t)}, x _ {j} ^ {(t)}\right). \tag {8} \\ \end{array} +$$ + +Textual prompt and Visual prompt feature. The samples $x_{i}^{(s)}$ and $x_{j}^{(s)}$ represent the feature descriptions in the source domain $X_{S}$ , and these features are used to represent the Text prompt. Similarly, $x_{i}^{(t)}$ and $x_{j}^{(t)}$ represent the feature descriptions in the target domain $X_{T}$ for the Visual prompt. These features are utilized in the MMD calculation to minimize the distributional gap between textual and visual representations, optimizing the alignment of both modalities for improved performance in cross-modal matching and generation tasks. + +Kernel function for similarity calculation. $k(x_{i},x_{j})$ calculates the similarity between samples $x_{i}$ and $x_{j}$ . The kernel function is used to compute similarity, which helps in measuring the relationships between samples in the source and target domains. + +Number of samples in the source and target domain. $|X_{n}^{s}|$ , $|X_{n}^{t}|$ denotes the number of samples in the source domain and target domain. It is used to normalize the number of samples, ensuring that the influence of each sample is balanced in the calculation. + +Minimizing this loss reduces the distributional gap between the textual and visual prompt, aligning the information for better performance in image-text matching. + +# 3.3. CG-SAM + +Infrared small targets often exhibit limited visual features, while the background is typically filled with significant clutter and noise. In this context, the SAM encoder faces an inherent trade-off between providing detailed target information and suppressing background noise. Specifically, the early layers of the ViT architecture retain more local details, which is crucial for recognizing small targets with limited visual features. However, these early layers also preserve a substantial amount of background noise, increasing the complexity of target detection. Therefore, directly using the Mask Decoder to recover clear target masks from complex backgrounds presents a significant challenge. + +# 3.3.1. Infrared Imaging Equation + +To improve SAM's performance in IRSTD, we propose a novel approach based on the fundamental principles of in- + +frared imaging. The Beer-Lambert law [23, 28], a fundamental principle in optics and spectroscopy, is widely applied in the study of various light transmission problems. Ideally, the transmission of infrared radiation from a target to a sensor can be described by the Beer-Lambert law. Thus, by drawing on the Beer-Lambert law and integrating it with the core principles of infrared imaging, we propose a novel approach to model the propagation of infrared radiation and its effects on imaging, especially in scenarios where background interference leads to degradation of the observed target. In the basic model of infrared imaging, the fundamental equation is defined as: + +$$ +I (z) = J (z) \cdot t (z) + A (z) \cdot (1 - t (z)), \tag {9} +$$ + +where: + +- $I(z)$ : The pixel intensity of the infrared observation image at position $z$ , which includes background radiation and interference. +- $J(z)$ : The true infrared radiation intensity of the target at position $z$ , representing the target's brightness in the infrared spectrum. +- $t(z)$ : The transmission factor, representing the energy attenuation of infrared radiation from the target to the observer, with values in the range [0, 1]. +- $A(z)$ : The background radiation refers to the infrared intensity from the background at position $z$ . + +The definition of the transmission factor is given by: + +$$ +t (z) = e ^ {- \beta \cdot d (z)}, \tag {10} +$$ + +where: + +- $\beta$ : The infrared radiation coefficient, determining the absorption level of infrared radiation by different components. This value typically varies depending on the infrared wavelength (e.g., near-infrared, shortwave infrared, or mid-to-far infrared). +- $d(z)$ : The scene depth, representing the distance between the target and the imaging device. In infrared imaging, a greater distance results in lower transmission, as infrared radiation undergoes greater loss in the atmosphere and other media. + +Given a 2D image, we need a function to calculate its corresponding depth. This process can be expressed as: + +$$ +d (z) = F (I (z)). \tag {11} +$$ + +Considering the scene's varying distances, classification techniques can be employed. Cao et al.[3] treated the depth estimation problem as a pixel-level classification task. + +$$ +F (I (z)) = \operatorname {R e s N e t} (I (x)). \tag {12} +$$ + +Using the above equations, we can calculate the infrared target's radiation intensity: + +$$ +J (z) = \frac {I (z)}{t (z)} - \frac {A (z)}{t (z)} + A (z). \tag {13} +$$ + +Substituting the transmission factor yields: + +$$ +J (z) = I (z) \cdot e ^ {\beta \cdot F (I (z))} - A (z) \cdot \left(e ^ {\beta \cdot F (I (z))} - 1\right). \tag {14} +$$ + +Through the formulations outlined above, we can more effectively separate the target from the background, thereby improving the performance of infrared small target detection. This approach constructs an interpretable mathematical model based on the fundamental principles of infrared imaging, providing a robust and explainable solution for infrared small target detection. + +# 3.3.2. CG-SAM + +Based on the infrared imaging equation (Eq. 14), we achieve precise infrared small target detection by inversely separating infrared degraded images from background images. As shown in Fig. 2, the text_prompt and visual_prompt are input into the SAM mask decoder, where they are fused with the early and late global semantic feature maps extracted by the ViT encoder. Building upon this, text_prompt and visual_prompt effectively extract and identify background features $A(z)$ from different scenes through the scene prompting mechanism. Similarly, we feed the learnable output token from SAM and the image embeddings from the SAM encoder into the mask decoder to further refine the target mask representation. This process gradually recovers target features, but due to background interference, it ultimately generates the infrared degraded image $I(z)$ with background noise. + +Finally, based on the degraded image $I(z)$ and the background features $A(z)$ , we apply the infrared imaging equation to achieve precise detection of infrared small targets in complex backgrounds while effectively suppressing background interference. Through this integrated process, the CG-SAM can accurately identify small targets and significantly improve the accuracy of target detection in infrared images with strong background noise. + +# 3.4. Datasets + +The goal of data generation is to create descriptive text for infrared images to support small target detection in complex backgrounds. To achieve this, we enhanced the IRSTD-1k, NUAA-SIRST, and NUDT-SIRST datasets by offering textual descriptions, resulting in the multimodal image-text dataset collections MIRSTD. Each infrared image, with the target mask removed, is encoded as a Base64 string and used as a text prompt for a visual-language model. ChatGPT generates concise descriptions that summarize key scene information, which are checked manually. The expert's prompt follows the format: a photo of a scene with [object], [object], and [object] offering a brief description of the scene and its objects. The final output is a 25-word description that captures the image's essential background information. + +# 4. Experiments + +# 4.1. Experimental Settings + +Dataset: We evaluate our model on the MIRSTD dataset, which consists of 427 and 1,000 real infrared images with one or more small targets, and 1,327 synthetic infrared images with small targets. All images are resized to $1024 \times 1024$ . For each dataset, we split the images into three disjoint subsets: $50\%$ for training, $30\%$ for validation, and $20\%$ for testing. + +Implementation Details: Our SAIST model is trained using the Adam optimizer with an initial learning rate of 0.0001, and a Cosine Decay Learning Rate Scheduler over 400 epochs. We use a batch size of 4 and train on a single Nvidia GeForce A800 GPU with 80 GB of memory. For comparison, we evaluate against state-of-the-art CNN-based methods for IRSTD, including ISNet [39], GCINet [41], UIUNet [33], DNANet [19], Dim2Clear [40], ALCNet [8], and ACMNet [7], as well as hybrid methods such as SCTransNet [35], IAANet [31], SAM [16] and SAM-HQ [15]. Traditional methods Top-Hat [1] and IPI [10] are also included. We adopt Intersection over Union (IoU) for pixel-level evaluation, and Probability of Detection (Pd) and False Alarm Rate (Fa) for object-level evaluation. + +# 4.2. Comparative Experiments + +Quantitative results: As shown in Tab. 1, traditional methods struggle with complex scenarios, especially CNNs, which fail to distinguish targets from backgrounds, resulting in higher false alarm and missed detection rates. The proposed SAIST outperforms existing SOTA methods across all metrics, particularly in reducing false alarms. SR-CLIP improves textual prompt generation through visual-textual alignment, while CG-SAM leverages both SR-CLIP's guidance and infrared imaging equations to enhance small target recognition in complex backgrounds. + +Visual Results: Fig. 4 shows detection results for SR-CLIP and CG-SAM, outperforming other IRSTD methods. SR-CLIP benefits from precise visual and textual guidance, while CG-SAM effectively integrates infrared imaging equations. Both methods excel in identifying small targets in complex backgrounds, reducing $F_{a}$ and background noise, thus enhancing detection accuracy and robustness. + +ROC: As shown in Fig. 5, our SAIST significantly outperforms the other methods, where the Area Under the ROC Curve (AUC) of our SAIST is larger than that of the traditional and deep learning based methods. + +# 4.3. Ablation Study + +To evaluate the components of SAIST, we conducted ablation studies on the NUAA-SIRST dataset, using CLIP + SAM as the baseline model. We systematically introduced the CG-SAM and SR-CLIP modules. As shown in the + +
MethodNUDT-SIRSTIRSTD-1kNUAA-SIRSTType
IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓
IPI [10]30.9381.9817.9927.9281.3716.181.0987.0530467Traditional
Top-Hat [1]22.4089.90174.110.0675.1114321.50879.7416456Traditional
MSLSTIPT [26]8.3447.4088111.4379.0315241.0800.0528183Traditional
ACM [7]68.9097.0511.2962.4191.4435.5870.7793.089.33CNN
ALCNet [8]81.4096.519.2662.0590.5821.7874.3173.1220.21CNN
FC3-Net [38]78.5693.8623.92265.0791.5415.5572.4498.1410.85CNN
ISNet [39]81.7796.344.4768.7795.5615.3980.0299.024.61CNN
DNA-Net [19]88.1998.629.0069.3893.311.6679.2698.485.35CNN
UIU-Net [33]92.1997.7715.4469.9693.9822.0778.2597.454.29CNN
Dim2Clear[40]81.3796.239.1766.3493.7520.9377.2999.106.72CNN
GCI-Net [41]92.4398.258.9667.7593.8912.8478.8199.344.11CNN
IAANet [31]90.2297.268.3266.2593.1514.2074.2293.5322.70Hybrid
SCTransNet [35]94.0996.954.2968.0393.2710.7477.5096.9513.92Hybrid
SAM [16]84.2097.3813.2868.1293.648.3875.2197.079.78SAM
SAM-HQ [15]88.0298.3114.4868.8593.549.5675.2797.226.87SAM
SAIST (Ours)95.2399.281.3172.1496.184.7680.8299.560.87CLIP + SAM
+ +Table 1. Comparison with existing IRSTD approaches on the NUDT-SIRST, IRSTD-1k and NUAA-SIRST datasets. The evaluation metrics are IoU $(10^{-2})$ , $P_d$ $(10^{-2})$ and $F_a$ $(10^{-6})$ , the best results are highlighted. + +![](images/1d96c6c8b8b965a304eef7db628288b70e95e8bddac893540e0cfc8e45487107.jpg) +Figure 4. Visual results of different IRSTD methods. The boxes in red, yellow, and blue represent correct, missed, and false detections, respectively. Close-up views are shown in the corners. + +Tab. 1, removing CG-SAM significantly reduces the IoU and Pd scores while increasing the Fa score. Additionally, as illustrated in the Fig. 7, compared to other methods, CG-SAM can precisely identify small targets in complex backgrounds. These results validate the effectiveness of CG-SAM in enhancing small target representation and suppressing background interference. Similarly, removing SR-CLIP increases the Fa score. Meanwhile, Fig. 6 demonstrates that SR-CLIP effectively suppresses background interference and noise. These results suggest that SR-CLIP gen + +erates accurate visual-textual prompts for CG-SAM. Ultimately, as shown in Table. 3, although SAIST does not have an advantage in inference time and the number of trainable parameters, it achieves the best performance across all evaluation metrics, demonstrating the excellence of its design. + +# 4.3.1. Impact of the SR-CLIP + +In this ablation study, we provide identical text prompts to CLIP, CoOp, and SR-CLIP, utilizing their respective vision and text encoders for image understanding and caption gen + +![](images/bfe7a2b668ffabef3593e520a8dd0f6ac35b767ddae03f6a5838912d9f621362.jpg) +Figure 5. ROC curves on NUAA-SIRST database. + +![](images/e702909692955b2b2b67b9aea117129b26a46fa79c68cfc1799b601c5df843a7.jpg) +A photo of tower crane A photo of a wall + +![](images/19ced1e25a62092afca593c20a4afc9dd36b4a340ee179e2211b13a2b29a2e6d.jpg) +CLIP + +![](images/675c6d3dab60bd9efd15c93a298b96d9cab9c38f7a3ff67c8d0c0d7d74a0db52.jpg) +CoOP + +![](images/d44ec6f18c5ddd409b739eb8b4077a410887ee4e56270ed29fa4d20ae15d961e.jpg) +SR-CLIP + +![](images/6394df9ca117440a0bc7933bded9da623484a4a71010364615f35ba004cf4731.jpg) + +![](images/d01c28366d9cc90b9b54b0ddd7f82c554a57fac39e2750848b8c98ce30f140d1.jpg) +Input + +![](images/9b4f1b0b6103cb42bbc6e60f0c3c1b10eb4a816b246b4961949cf4ac21516995.jpg) + +![](images/ad1f2031f3a9d88e9628bd2926ef9353f5fd0f1777988b6a9a750ed8fc60c990.jpg) + +![](images/b6bb55e26c3ec5a69d8ef0eebcef5a8abfc4d8689238ba02d2bdd51a376cb319.jpg) + +![](images/da05fe9a18f62eb3ab9e3ddef3ee523cb8e110cc9a3eec9aaf9326a186528bda.jpg) + +![](images/89f2d620ad127f7e123b498f4892c36c068f100f01e347b321e720cff8e450d0.jpg) +Figure 6. Visual comparison of CLIP, CoOp and SR-CLIP. + +![](images/3578ce996adb9a9f918e0fecfc2a0c9d62d798839c51ecbf9bd9843f4aa72ddf.jpg) + +![](images/e12787b2d4a3710072db5241d5f49fb0267389a0ac999a16c1feed588f9995d3.jpg) + +![](images/6c9345d8838947389a8bdd9cd4d9407712c2661c1122c1023788ab796cc4bf90.jpg) + +![](images/f6291f93359b9f51d5f85b57127e6dd492c31470f2fed12f190decc93e0df023.jpg) + +![](images/8b94aec4c3f1ed7f92b0a5b20232eae067c1dadb7d0f3a27bc5596116ed2bc42.jpg) +Figure 7. Visual comparison of SAM, HQ-SAM and CG-SAM. + +![](images/c8dc122fa5ce9c4e2b9ef0be4463f77dd8cdc700e7c387c5774c6c272d36aab9.jpg) +Input + +![](images/25b87a03c11f31397f13730562f1edda62012238a344a47e6d568dd1e4121eeb.jpg) + +![](images/5d5e002284ffe5308c687460b78a40594678cbc24666e72bc1e8b0d917ec2252.jpg) + +![](images/f94aa8fe9f0d285296037f813fc0f85d2233b149acce2a92916231c5c758633a.jpg) + +Table 2. Ablation study of SR-CLIP and CG-SAM in $IoU(\%)$ , $P_d(\%)$ , $F_a(10^{-6})$ on on NUAA-SIRST. + +
MethodIoUPdFa
SAIST80.8299.560.87
w/o SR-CLIP80.0298.776.61
w/o CG-SAM78.8797.428.53
w/o CG-SAM & SR-CLIP76.8296.5012.37
+ +eration, to systematically evaluate the performance of SR-CLIP. As shown in Fig. 6 and the Tab. 4, we demonstrate that SR-CLIP effectively suppresses background interference and noise, enabling the generation of precise visual-textual prompts for CG-SAM. These results highlight the efficacy of SR-CLIP in improving the alignment between the visual and textual modalities. + +# 4.3.2. Impact of the CG-SAM + +In this ablation study, we utilize the same SAM encoder while employing different Mask Decoders: SAM, HQ- + +Table 3. Comparison of different models in $IoU(\%)$ , $P_d(\%)$ , $F_a(10^{-6})$ on MIRSTD. + +
ModelTrainable Params (M)Inference (s)Avg IoUAvg PdAvg Fa
SAM + CLIP383.280.07978.4996.678.45
SAIST389.570.08182.7398.342.31
+ +Table 4. Ablation study of CLIP, CoOp, and SR-CLIP in $IoU(\%)$ , $P_d(\%)$ , $F_a(10^{-6})$ on NUAA-SIRST. + +
MethodIoUPdFa
CLIP + CG-SAM80.0298.776.61
CoOp + CG-SAM80.3298.564.87
SR-CLIP + CG-SAM80.8299.560.87
+ +Table 5. Ablation study of SAM, HQ-SAM, and CG-SAM in $IoU(\%)$ , $P_d(\%)$ , $F_a(10^{-6})$ on NUAA-SIRST. + +
MethodIoUPdFa
SAM + SR-CLIP78.8797.425.23
HQ-SAM + SR-CLIP79.4598.894.56
CG-SAM + SR-CLIP80.8299.560.87
+ +SAM, and CG-SAM. Identical fine-grained visual and textual prompts, generated by SR-CLIP, are provided as inputs. The performance of CG-SAM is systematically evaluated. As demonstrated in Fig. 7 and the Tab. 5, our results confirm that CG-SAM is capable of accurately identifying small targets in complex backgrounds. These findings validate that, with the aid of infrared imaging equations, CG-SAM enhances the representation of small targets while effectively suppressing background interference. + +# 5. Conclusion + +This paper presents SAIST, a novel framework that integrates visual and textual modalities for enhanced IRSTD. Leveraging SR-CLIP to capture cross-modal relationships, SAIST generates precise prompts to overcome the limitations of image-only methods. Guided by these prompts, CG-SAM employs interpretable infrared imaging equations to extract targets while suppressing noise. We also introduce MIRSTD, the first multimodal IRSTD dataset with rich image-text pairs. Experiments demonstrate the effectiveness of multimodal integration in IRSTD. + +# 6. Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grant 62272363, Grant 92470108, Grant 62371362, and Grant U21A20514; in part by the Joint Laboratory for Innovation in Satellite-Borne Computers and Electronics Technology Open Fund 2023 under Grant 2024KFKT001-1; in part by the Proof of Concept Foundation of Xidian University Hangzhou of Technology under Grant No. GNYZ2023YL0301; and part by the Shaanxi Provincial Key R&D Program Project under Grant 2024SF-YBXM-330. + +# References + +[1] Xiangzhi Bai and Fugen Zhou. Analysis of new top-hat transformation and the application for infrared dim small target detection. Pattern Recognition, 43(6):2145-2156, 2010. 1, 2, 6, 7 +[2] Karsten M Borgwardt, Arthur Gretton, Malte J Rasch, Hans-Peter Kriegel, Bernhard Scholkopf, and Alex J Smola. Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49-e57, 2006. 4 +[3] Yuanzhouhan Cao, Zifeng Wu, and Chunhua Shen. Estimating depth from monocular images as classification using deep fully convolutional residual networks. IEEE Transactions on Circuits and Systems for Video Technology, 28(11): 3174-3182, 2018. 5 +[4] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE transactions on geoscience and remote sensing, 52(1):574-581, 2013. 2 +[5] G Cucurullo, L Giordano, D Albanese, Luciano Cinquanta, and M Di Matteo. Infrared thermography assisted control for apples microwave drying. Journal of food engineering, 112 (4):319-325, 2012. 1 +[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE journal of selected topics in applied earth observations and remote sensing, 10(8): 3752-3767, 2017. 1, 2 +[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 950–959, 2021. 1, 2, 6, 7 +[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 59(11):9813-9824, 2021. 6, 7 +[9] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G. Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing, 22(12):4996-5009, 2013. 1 +[10] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 6, 7 +[11] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multistage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024. 2 +[12] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters, 18(9):1670-1674, 2020. 1, 2 +[13] MD. Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learn + +ing for image captioning. ACM Comput. Surv., 51(6), 2019. 3 +[14] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3 +[15] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 6, 7 +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 6, 7 +[17] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 2 +[18] Wing-Cheung Law, Zhourui Xu, Ken-Tye Yong, Xin Liu, Mark T Swihart, Mukund Seshadri, and Paras N Prasad. Manganese-doped near-infrared emitting nanocrystals for in vivo biomedical imaging. Optics express, 24(16):17553-17561, 2016. 1 +[19] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 2022. 2, 6, 7 +[20] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2 +[21] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1, 2 +[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2 +[23] Alexandre Mallet, Roumiana Tsenkova, Jelena Muncan, Cyrille Charnier, Éric Latrille, Ryad Bendoula, Jean-Philippe Steyer, and Jean-Michel Roger. Relating near-infrared light path-length modifications to the water content of scattering media in near-infrared spectroscopy: toward a new bouguer-beer-lambert law. Analytical Chemistry, 93 (17):6817-6823, 2021. 5 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[25] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng + +Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2 +[26] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing, 59(5):3737-3752, 2021. 7 +[27] Danica J Sutherland, Hsiao-Yu Tung, Heiko Strathmann, Soumyajit De, Aaditya Ramdas, Alex Smola, and Arthur Gretton. Generative models and model criticism via optimized maximum mean discrepancy. arXiv preprint arXiv:1611.04488, 2016. 4 +[28] Donald F Swinehart. The beer-lambert law. Journal of chemical education, 39(7):333, 1962. 5 +[29] Michael Teutsch and Wolfgang Krüger. Classification of small boats in infrared images for maritime surveillance. In 2010 International WaterSide Security Conference, pages 1-7. IEEE, 2010. 1 +[30] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8509-8518, 2019. 1, 2 +[31] Kewei Wang, Shuaiyuan Du, Chengxin Liu, and Zhiguo Cao. Interior attention-aware network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2022. 1, 2, 6, 7 +[32] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2 +[33] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 6, 7 +[34] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[35] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 6, 7 +[36] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 2 +[37] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1730–1738, 2022. 1, 2 +[38] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1857-1865, 2022. 1, 7 + +[39] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 2, 6, 7 +[40] Mingjin Zhang, Rui Zhang, Jing Zhang, Jie Guo, Yunsong Li, and Xinbo Gao. Dim2clear network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-14, 2023. 2, 6, 7 +[41] Mingjin Zhang, Ke Yue, Boyang Li, Jie Guo, Yunsong Li, and Xinbo Gao. Single-frame infrared small target detection via gaussian curvature inspired network. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2, 6, 7 +[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4 +[43] Hu Zhu, Haopeng Ni, Shiming Liu, Guoxia Xu, and Lizhen Deng. TnIrs: Target-aware non-local low-rank modeling with saliency filtering regularization for infrared small target detection. IEEE Transactions on Image Processing, 29: 9546-9558, 2020. 1 +[44] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 1 \ No newline at end of file diff --git a/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/images.zip b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42592d1856b7ffe6067c6adc693feed9cbb8382d --- /dev/null +++ b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e9bc21c63c9e5acd3f21929e68b78d431bad85a6e48e98f940fbf9633805065 +size 587884 diff --git a/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/layout.json b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ed606953a094bfebd228dca4b953b3dbb7d880ac --- /dev/null +++ b/2025/SAIST_ Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining/layout.json @@ -0,0 +1,9268 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "type": "text", + "content": "SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": "Mingjin Zhang" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": ", Xiaolong Li" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": ", Fei Gao" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": ", Jie Guo" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": ", Xinbo Gao" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": ", Jing Zhang" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": "Xidian University, China " + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": "Chongqing University of Posts and Telecommunications, China \n" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 80, + 160, + 532, + 232 + ], + "type": "text", + "content": "School of Computer Science & School of Artificial Intelligence, Wuhan University, China \n{mjinzhang, fgao}@xidian.edu.cn, 23011210739@stu.xidian.edu.cn \njguo, xbgao@mail.xidian.edu.cn, jingzhang.csv@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 622 + ], + "type": "text", + "content": "Infrared Small Target Detection (IRSTD) aims to identify low signal-to-noise ratio small targets in infrared images with complex backgrounds, which is crucial for various applications. However, existing IRSTD methods typically rely solely on image modalities for processing, which fail to fully capture contextual information, leading to limited detection accuracy and adaptability in complex environments. Inspired by vision-language models, this paper proposes a novel framework, SAIST, which integrates textual information with image modalities to enhance IRSTD performance. The framework consists of two main components: Scene Recognition Contrastive Language-Image Pretraining (SR-CLIP) and CLIP-guided Segment Anything Model (CG-SAM). SR-CLIP generates a set of visual descriptions through object-object similarity and object-scene relevance, embedding them into learnable prompts to refine the textual description set. This reduces the domain gap between vision and language, generating precise textual and visual prompts. CG-SAM utilizes the prompts generated by SR-CLIP to accurately guide the Mask Decoder in learning prior knowledge of background features, while incorporating infrared imaging equations to improve small target recognition in complex backgrounds and significantly reduce the false alarm rate. Additionally, this paper introduces the first multimodal IRSTD dataset, MIRSTD, which contains abundant image-text pairs. Experimental results demonstrate that the proposed SAIST method outperforms existing state-of-the-art approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 646, + 137, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 137, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 137, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 692 + ], + "type": "text", + "content": "Infrared small target detection (IRSTD) is crucial for applications like traffic management and maritime rescue" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 258, + 542, + 395 + ], + "blocks": [ + { + "bbox": [ + 328, + 258, + 542, + 395 + ], + "lines": [ + { + "bbox": [ + 328, + 258, + 542, + 395 + ], + "spans": [ + { + "bbox": [ + 328, + 258, + 542, + 395 + ], + "type": "image", + "image_path": "4fba36fd60f51c873be7c05f05db2569a5dad0cfd6f2db93bed58be0c2ea4c6c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 458 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 458 + ], + "type": "text", + "content": "Figure 1. Illustration of the advantages of text-guided detection over image-only detection in identifying small infrared targets in complex backgrounds. The textual description offers valuable context, helping to reduce interference and distinguish the target from the background effectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 569 + ], + "type": "text", + "content": "[5, 18, 29]. Unlike visible light, infrared sensors detect thermal radiation, enabling target acquisition in adverse conditions such as fog, rain, snow, and at night [44]. However, due to the long-range imaging mechanism, infrared images often lack distinctive features like color, shape, and texture, while the background typically shows strong structural features with higher intensity and sharper contours [9, 43]. These challenges—complex backgrounds, clutter, and noise—hinder accurate detection, complicating IRSTD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 676 + ], + "type": "text", + "content": "IRSTD methods can be divided into traditional algorithms and deep learning-based ones. Early methods focused on filtering and image enhancement [1, 6, 12], which performed well in simpler scenarios but struggled with complex environments due to reliance on prior knowledge and handcrafted features. With the rise of deep learning and large IRSTD datasets, newer approaches using Generative Adversarial Networks (GANs) [30], U-Net [7, 21, 38], and transformers [31, 35, 37] have been introduced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Despite advances, existing methods are limited by their reliance on single-modality inputs. Recently, the CLIP-SAM model, which combines Contrastive Language-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9549" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "Image Pretraining (CLIP) and the Segment Anything Model (SAM), has shown promise in enhancing cross-modal feature fusion for various tasks [20]. CLIP [24] captures semantic relationships between images and text, offering rich supervision, while SAM [16] provides precise segmentation. However, applying CLIP-SAM to IRSTD remains challenging due to the differences between infrared and natural images. Specifically: 1) Infrared targets lack typical visual features, limiting the generalization of the CLIP model; 2) Small, weak infrared targets and noisy backgrounds complicate target detection, making SAM's application difficult. To address these issues, we propose using CLIP to extract semantically rich prior knowledge from text, guiding SAM to better capture target details amidst complex backgrounds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 506 + ], + "type": "text", + "content": "This paper introduces a novel IRSTD approach named SAIST by combining text and image modalities within a multimodal learning framework. First, we design the Scene Recognition CLIP (SR-CLIP) to capture both global scene and local object features. SR-CLIP creates similarity matrices for objects and scenes, facilitating key information extraction. It also integrates a cross-modal interaction mechanism, incorporating scene-specific visual cues into learnable prompts that refine text features and mitigate the vision-language domain gap. Building on this, we diverse CG-SAM to utilize the text prompts from SR-CLIP to guide SAM in learning background features, enhancing accurate identification and extraction across diverse scenes. By combining global semantic and background features, and leveraging the radiative transfer properties of targets and backgrounds, CG-SAM achieves effective target-background separation. Additionally, we present MIRSTD, the first multimodal IRSTD dataset, which combines imagedataset pairs from NUAA-SIRST, NUDT-SIRST, and IRSTD1K datasets. Text descriptions are generated by GPT-4V and carefully reviewed manually. Experiments validate the superiority of SAIST over state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 505, + 274, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 505, + 274, + 516 + ], + "spans": [ + { + "bbox": [ + 67, + 505, + 274, + 516 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 517, + 295, + 637 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 517, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 517, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 295, + 566 + ], + "type": "text", + "content": "- We introduce SAIST, the first approach to integrate multimodal vision and language for IRSTD. On challenging benchmarks, it outperforms SOTA methods, excelling in both objective metrics and subjective evaluations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 566, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 295, + 601 + ], + "type": "text", + "content": "- We design two key components, SR-CLIP and CG-SAM, to effectively fuse multimodal information and separate targets and backgrounds in complex scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 601, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 637 + ], + "type": "text", + "content": "- We create a novel multimodal IRSTD dataset combining infrared images and textual data, offering rich resources for training and evaluating multimodal learning methods." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 651, + 139, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 139, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 139, + 662 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 671, + 222, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 671, + 222, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 671, + 222, + 684 + ], + "type": "text", + "content": "2.1. Infrared small target detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Traditional IRSTD methods use image processing or handcrafted features, such as low-rank (e.g., RIFT [6]," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "PSTNN [36]), human visual system-based (e.g., TLLCM [4], WSLCM [12]), and filter-based approaches (e.g., tophat [1]). Low-rank methods perform well in low SNR but struggle with false alarms in low-contrast areas. Human visual methods are effective with large objects but not in complex scenes, and filter-based techniques suppress uniform noise but fail in complex backgrounds. Deep learning IRSTD methods include CNN-based models (e.g., MDvsFA [30], ISNet [39], Dim2Clear [40], CSRNet [21], DMFNet [11], GCI-Net [41]) that focus on local features, with some improving SNR, detection, or small target extraction, and hybrid models (e.g., IAANet [31] and RKformer [37]) that combine CNNs and Vision Transformers (ViTs) to capture both local and global context. While IRSTD has made significant strides, current methods rely mainly on image modalities, which fail to capture full contextual information, limiting detection accuracy in complex backgrounds. Integrating textual information can offer substantial benefits, providing richer scene understanding and aiding target extraction in challenging environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 321, + 479, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 479, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 479, + 333 + ], + "type": "text", + "content": "2.2. CLIP and SAM Collaboration" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 542 + ], + "type": "text", + "content": "In recent years, foundational models like CLIP and SAM have garnered significant attention for their exceptional performance across various tasks. CLIP excels at aligning multimodal features, while SAM specializes in segmentation through diverse prompts. Recent research has explored synergies between these models, such as CRIS [32] integrates both models to boost performance. ClipSAM [20] uses CLIP's semantic understanding for anomaly localization and coarse segmentation, later refined by SAM. OpenVocabulary SAM enhances SAM's recognition by transferring knowledge from SAM to CLIP via distillation and adapters. However, applying pre-trained CLIP-SAM models, originally trained on natural images, to IRSTD faces challenges due to domain differences. In this work, we leverage CLIP's semantic understanding from scene descriptions to guide SAM in extracting target information from complex infrared backgrounds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 552, + 429, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 429, + 563 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 429, + 563 + ], + "type": "text", + "content": "2.3. Datasets for IRSTD" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Prior studies have developed large-scale datasets [7, 19, 39], which have advanced IRSTD research. However, these datasets only include image data with target masks, limiting the detection of small targets in complex backgrounds. Recent multimodal approaches, such as text-guided detection [17, 22, 25], have shown performance improvements. Inspired by this, we create the MIRSTD dataset, which combines both image and text modalities. The text modality in MIRSTD provides descriptions of scenes in infrared images from datasets like NUAA-SIRST, IRSTD-1k, and NUTDSIRST, helping to mitigate the challenge of weak target signals against strong background noise." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9550" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 296, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 296, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 296, + 242 + ], + "type": "image", + "image_path": "188d26818404e83d33891b2183ffbf1284b1fc20818ed0467c2e13364219a0c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed encoder-decoder architecture. The CLIP text encoder converts textual descriptions into feature representations, while the frozen image encoder processes infrared images without target. SR-CLIP leverages cross-modal interactions to refine textual features, bridging the visual-textual domain gap. The SAM image encoder extracts image embeddings, fine-tuned with LoRA to enhance task adaptability. The generated textual prompts and image embeddings are fed into the SAM mask decoder, along with an infrared imaging model, to effectively suppress background noise and detect small targets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 554, + 242 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 554, + 242 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 554, + 242 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 554, + 242 + ], + "type": "image", + "image_path": "87092d3acab5275f7887276516369979861d590fd40b72b0f416b13cea58645c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 325, + 138, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 138, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 138, + 339 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 175, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 175, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 175, + 357 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 364, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 555 + ], + "type": "text", + "content": "The overall architecture of SAIST is illustrated in Fig. 2. In this framework, the text encoder of SR-CLIP converts textual descriptions into feature representations, while the image encoder is frozen. Next, SR-CLIP employs a cross-modal interaction mechanism to integrate the visual and textual description sets, refining the textual features and reducing the domain gap between vision and language, thus generating precise textual and visual prompts. Subsequently, the pre-trained and frozen SAM image encoder extracts image embeddings, which are fine-tuned using Low-Rank Adaptation (LoRA) [14] to enhance the model's adaptability to the specific task. Finally, the generated textual prompts and image embeddings are fed into the SAM mask decoder, in conjunction with a specially designed infrared imaging model, to effectively suppress background noise and accurately detect small targets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 122, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 122, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 122, + 574 + ], + "type": "text", + "content": "3.2. SR-CLIP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "In infrared imaging, due to the inherent characteristics of remote sensing, the target's radiative intensity is typically weak, while the background's intensity is higher and more prominent. In this case, the information related to the target is difficult to describe clearly and accurately. In contrast, the background's radiative intensity is more stable and easier to identify, allowing the overall structure and features of the scene to be effectively reflected through intuitive textual descriptions. There, we design SR-CLIP to fully leverage the complementary relationship between background textual descriptions and visual features, generating precise" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 351 + ], + "type": "text", + "content": "prompts that guide SAM to effectively suppress background interference in infrared images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 356, + 460, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 460, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 460, + 369 + ], + "type": "text", + "content": "3.2.1. Visual / Text description set" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 371, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 468 + ], + "type": "text", + "content": "For the scene recognition task in infrared images, scenes are composed of multiple semantic concepts, including distinct objects and backgrounds. These features necessitate the consideration of various semantic information when learning discriminative patterns. To capture comprehensive scene information, we introduce a novel form of textual prompting, referred to as scene prompts. This prompt creates a textual description set:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 475, + 552, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 552, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 552, + 488 + ], + "type": "text", + "content": "{aphotof[scene]with[object],[object],and[object]}," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 496, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 568 + ], + "type": "text", + "content": "which provides both scene-level and object-level textual descriptions. Simultaneously, to bridge the image-text domain gap, we extract local object features from Clip's image encoder, acquire global visual features with the help of a learnable projection layer, and generate a set of visual feature descriptions using the image captioning [13, 34]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 576, + 512, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 576, + 512, + 589 + ], + "spans": [ + { + "bbox": [ + 357, + 576, + 512, + 589 + ], + "type": "interline_equation", + "content": "\\left\\{\\left[ \\text {s c e n e} \\right], \\left[ \\text {o b j e c t} \\right], \\left[ \\text {o b j e c t} \\right] \\left[ \\text {o b j e c t} \\right] \\right\\},", + "image_path": "b4eca6545627898bee66c57c5d51f45c5233933a47803b8b7c7538ca9116f3b4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "type": "text", + "content": "which provides both scene-level and object-level visual descriptions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 627, + 421, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 421, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 421, + 639 + ], + "type": "text", + "content": "3.2.2. Image Captioning" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "In scene recognition, objects often share considerable similarity, and there is a strong correlation between objects and the scene, which reflects the unique characteristics of that particular scene. By integrating both interactions, the model can capture more granular semantic information, thereby improving its discriminative power in specific contexts. To" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9551" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 298, + 213 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 298, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 298, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 298, + 213 + ], + "type": "image", + "image_path": "b939f16afc102edb76b83a6db7b66ef9e50933e872a869cf306af8abc1410ee2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 288 + ], + "type": "text", + "content": "Figure 3. Overview of the Cross-modal Interaction Mechanism. Image captioning generates visual feature descriptions that are processed by a text encoder. The text decoder then maps visual information into the text, enhancing text feature expressiveness. Finally, the visual decoder recovers visual information from the encoded text to improve image-text alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 324 + ], + "type": "text", + "content": "this end, we construct both an object-object similarity matrix and a scene-object correlation matrix:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 325, + 214, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 325, + 214, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 325, + 214, + 338 + ], + "type": "text", + "content": "Object-Object Similarity Matrix:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 344, + 295, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 344, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 82, + 344, + 295, + 389 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c c c} o _ {1 1} & \\dots & o _ {1 n} \\\\ \\vdots & \\ddots & \\vdots \\\\ o _ {n 1} & \\dots & o _ {n n} \\end{array} \\right] = \\frac {\\exp \\left(\\frac {o _ {i} ^ {\\top} o _ {j}}{\\| o _ {i} \\| \\| o _ {j} \\|}\\right)}{\\sum_ {k = 1} ^ {n} \\exp \\left(\\frac {o _ {i} ^ {\\top} o _ {k}}{\\| o _ {i} \\| \\| o _ {k} \\|}\\right)}, \\tag {1}", + "image_path": "e0593d382c70fcb897503bbbc02eb2a21364cc122e5863f0ee60dd55b83bd295.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "inline_equation", + "content": "o_{i,j}" + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "text", + "content": " represents the similarity between the feature vector of object " + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 397, + 295, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 421, + 216, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 216, + 433 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 216, + 433 + ], + "type": "text", + "content": "Scene-Object Correlation Matrix:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 440, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 440, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 91, + 440, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l l l} s _ {1} & \\dots & s _ {n} \\end{array} \\right] = \\frac {\\exp \\left(\\frac {s ^ {\\top} o _ {i}}{\\| s \\| \\| o _ {i} \\|}\\right)}{\\sum_ {k = 1} ^ {n} \\exp \\left(\\frac {s ^ {\\top} o _ {k}}{\\| s \\| \\| o _ {k} \\|}\\right)}, \\tag {2}", + "image_path": "639a511c18b4d37ae00416e6e961005424fed2bcfce0b22c75b3fb03bd7cdcff.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "type": "text", + "content": " represents the correlation between the feature vector of the scene and the feature vector of object " + }, + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 490, + 295, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 586 + ], + "type": "text", + "content": "Therefore, based on the similarity and correlation matrices, we design the Object-to-Object and Scene-to-Object modules to effectively capture the interrelationships between objects and facilitate the integration of scene and object details. Meanwhile, by incorporating image captioning techniques, we generate precise visual description." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 591, + 241, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 241, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 241, + 602 + ], + "type": "text", + "content": "3.2.3. Cross-Modal Interaction Mechanism" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "CoOp [42] can be viewed as a textual domain prompt, where the learnable context is simply a trainable vector without incorporating any visual information. Inspired by CoOp, we propose a cross-modal interaction mechanism that introduces the visual description set into the text description set to refine the textual features. This approach mitigates the domain gap between the visual and textual modalities, thereby enhancing the performance of the CLIP model in downstream tasks. Initially, as shown in Fig. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "we integrate the visual and textual description sets to form a comprehensive input description set, which is then processed by the CLIP text encoder. Meanwhile, to further enhance alignment, we expand the output dimension of the final learnable linear projection layer in CLIP, enabling it to generate more expressive and informative textual features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 156, + 555, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 156, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 321, + 156, + 555, + 169 + ], + "type": "interline_equation", + "content": "\\text {D e s c r i p t i o n} = \\operatorname {C o n c a t} \\left(T _ {\\text {s c e n e}}, T _ {\\text {o b j e c t s}}, V _ {\\text {s c e n e}}, V _ {\\text {o b j e c t s}}\\right), \\tag {3}", + "image_path": "8481407c819105c85a41afacd815a73969da51a8acb16f80f02644f276440858.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 194, + 555, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 194, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 350, + 194, + 555, + 207 + ], + "type": "interline_equation", + "content": "\\text {T e x t} = \\text {C L I P T e x t E n c o d e r (D e s c r i p t i o n)}. \\tag {4}", + "image_path": "bed47088f6310731ee91e18a2c58c44667c2af6123f77f1664a87598ef214a98.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": "To effectively transfer fine-grained semantic features from the visual description set to the textual description set, we designed a text decoder. As illustrated in Fig. 3, the text decoder first employs the Text-to-Text module to refine and extract textual information. Then, based on the correlation between image features and textual descriptions, it leverages the Visual-to-Text module to dynamically map visual information onto textual descriptions. This process facilitates the recovery of semantic details relevant to the image content, thereby enhancing the expressiveness and informativeness of the textual descriptions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 362, + 555, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 362, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 353, + 362, + 555, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\text {P r o m p t}} = \\text {T e x t D e c o d e r (V i s u a l , T e x t)}, \\tag {5}", + "image_path": "07368928c62a69cdb5a0558c62f3c155337b3bba3a22ffe8adb92b59a658d251.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 383, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 554, + 515 + ], + "type": "text", + "content": "Meanwhile, in order to recover the visual information from the set of text descriptions encoded by CLIP text encoder, we designed a visual decoder. As illustrated in Fig. 3, the decoder effectively extracts the visual features of the background through the Visual-to-Visual module. It then leverages the Text-to-Visual module to perform cross-modal conversion, capturing the semantic relationships between the image and textual descriptions. This process enhances the mutual understanding and alignment between visual and textual modalities, improving the coherence and expressiveness of the descriptions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 348, + 530, + 555, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 530, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 348, + 530, + 555, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {V} _ {\\text {P r o m p t}} = \\text {V i s u a l D e c o d e r} (\\text {V i s u a l}, \\text {T e x t}). \\tag {6}", + "image_path": "8d2d647b769246859e4dc84b1c4487e0cf853e4426c1f7eef15f49cbd510ed47.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "type": "text", + "content": "Loss Function: In cross-modal tasks, textual and visual description sets come from different modalities, and their feature spaces may differ significantly. To enable effective use of both modalities, it's crucial to align their distributions in a shared semantic space. The Maximum Mean Discrepancy (MMD)[2, 27] measures the squared distance between the means of two distributions " + }, + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "type": "inline_equation", + "content": "X_{S}" + }, + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 313, + 551, + 554, + 635 + ], + "type": "text", + "content": " in the feature space:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 657, + 553, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 319, + 657, + 553, + 689 + ], + "type": "interline_equation", + "content": "D _ {\\mathcal {H}} \\left(X _ {S}, X _ {T}\\right) = \\sup _ {\\| \\phi \\| _ {\\mathcal {H}} \\leq 1} \\| \\mathbb {E} _ {S} [ \\phi (X _ {S}) ] - \\mathbb {E} _ {T} [ \\phi (X _ {T}) ] \\| _ {\\mathcal {H}} ^ {2}, \\tag {7}", + "image_path": "994d537441feb75564903345d753c6f291e1ec5d10084a38b9696a2b9f093277.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " is the feature map, and " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_S[\\phi (X_S)]" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_T[\\phi (X_T)]" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " are the expected values of the mapped samples. To align" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9552" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 275, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 275, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 275, + 84 + ], + "type": "text", + "content": "textual and visual prompts, we propose an MMD loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 89, + 295, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 89, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 89, + 295, + 213 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} (D _ {\\mathcal {H}} (X _ {S}, X _ {T})) = \\frac {1}{| X _ {n} ^ {s} | ^ {2}} \\sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {s} |} k \\Big (x _ {i} ^ {(s)}, x _ {j} ^ {(s)} \\Big) \\\\ - \\frac {2}{| X _ {n} ^ {s} | | X _ {n} ^ {t} |} \\sum_ {i = 1} ^ {| X _ {n} ^ {s} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \\left(x _ {i} ^ {(s)}, x _ {j} ^ {(t)}\\right) \\\\ + \\frac {1}{| X _ {n} ^ {t} | ^ {2}} \\sum_ {i = 1} ^ {| X _ {n} ^ {t} |} \\sum_ {j = 1} ^ {| X _ {n} ^ {t} |} k \\left(x _ {i} ^ {(t)}, x _ {j} ^ {(t)}\\right). \\tag {8} \\\\ \\end{array}", + "image_path": "df19ac788c37beb50b4145d6a24c6d1f045c85aab205750fa700d5800bcd0891.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": "Textual prompt and Visual prompt feature. The samples " + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "inline_equation", + "content": "x_{i}^{(s)}" + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "inline_equation", + "content": "x_{j}^{(s)}" + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": " represent the feature descriptions in the source domain " + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "inline_equation", + "content": "X_{S}" + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": ", and these features are used to represent the Text prompt. Similarly, " + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "inline_equation", + "content": "x_{i}^{(t)}" + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "inline_equation", + "content": "x_{j}^{(t)}" + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": " represent the feature descriptions in the target domain " + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "inline_equation", + "content": "X_{T}" + }, + { + "bbox": [ + 55, + 213, + 295, + 335 + ], + "type": "text", + "content": " for the Visual prompt. These features are utilized in the MMD calculation to minimize the distributional gap between textual and visual representations, optimizing the alignment of both modalities for improved performance in cross-modal matching and generation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "text", + "content": "Kernel function for similarity calculation. " + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "inline_equation", + "content": "k(x_{i},x_{j})" + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "text", + "content": " calculates the similarity between samples " + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 55, + 335, + 296, + 395 + ], + "type": "text", + "content": ". The kernel function is used to compute similarity, which helps in measuring the relationships between samples in the source and target domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": "Number of samples in the source and target domain. " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "|X_{n}^{s}|" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "|X_{n}^{t}|" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": " denotes the number of samples in the source domain and target domain. It is used to normalize the number of samples, ensuring that the influence of each sample is balanced in the calculation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 467, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 504 + ], + "type": "text", + "content": "Minimizing this loss reduces the distributional gap between the textual and visual prompt, aligning the information for better performance in image-text matching." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 123, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 123, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 123, + 521 + ], + "type": "text", + "content": "3.3. CG-SAM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 526, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 670 + ], + "type": "text", + "content": "Infrared small targets often exhibit limited visual features, while the background is typically filled with significant clutter and noise. In this context, the SAM encoder faces an inherent trade-off between providing detailed target information and suppressing background noise. Specifically, the early layers of the ViT architecture retain more local details, which is crucial for recognizing small targets with limited visual features. However, these early layers also preserve a substantial amount of background noise, increasing the complexity of target detection. Therefore, directly using the Mask Decoder to recover clear target masks from complex backgrounds presents a significant challenge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 674, + 201, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 201, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 201, + 687 + ], + "type": "text", + "content": "3.3.1. Infrared Imaging Equation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To improve SAM's performance in IRSTD, we propose a novel approach based on the fundamental principles of in-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "frared imaging. The Beer-Lambert law [23, 28], a fundamental principle in optics and spectroscopy, is widely applied in the study of various light transmission problems. Ideally, the transmission of infrared radiation from a target to a sensor can be described by the Beer-Lambert law. Thus, by drawing on the Beer-Lambert law and integrating it with the core principles of infrared imaging, we propose a novel approach to model the propagation of infrared radiation and its effects on imaging, especially in scenarios where background interference leads to degradation of the observed target. In the basic model of infrared imaging, the fundamental equation is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 221, + 555, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 221, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 353, + 221, + 555, + 236 + ], + "type": "interline_equation", + "content": "I (z) = J (z) \\cdot t (z) + A (z) \\cdot (1 - t (z)), \\tag {9}", + "image_path": "4d5b2fe6c74b35fc49b18c0edf0d242b4cb303062a4480e952176034f46a7227.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 240, + 345, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 345, + 250 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 345, + 250 + ], + "type": "text", + "content": "where:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 251, + 553, + 383 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "type": "inline_equation", + "content": "I(z)" + }, + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "type": "text", + "content": ": The pixel intensity of the infrared observation image at position " + }, + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 314, + 251, + 553, + 285 + ], + "type": "text", + "content": ", which includes background radiation and interference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "type": "inline_equation", + "content": "J(z)" + }, + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "type": "text", + "content": ": The true infrared radiation intensity of the target at position " + }, + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 314, + 287, + 553, + 323 + ], + "type": "text", + "content": ", representing the target's brightness in the infrared spectrum." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "type": "inline_equation", + "content": "t(z)" + }, + { + "bbox": [ + 314, + 323, + 553, + 358 + ], + "type": "text", + "content": ": The transmission factor, representing the energy attenuation of infrared radiation from the target to the observer, with values in the range [0, 1]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "type": "inline_equation", + "content": "A(z)" + }, + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "type": "text", + "content": ": The background radiation refers to the infrared intensity from the background at position " + }, + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 314, + 359, + 553, + 383 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 384, + 526, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 526, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 526, + 396 + ], + "type": "text", + "content": "The definition of the transmission factor is given by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 399, + 399, + 553, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 399, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 399, + 399, + 553, + 414 + ], + "type": "interline_equation", + "content": "t (z) = e ^ {- \\beta \\cdot d (z)}, \\tag {10}", + "image_path": "974bb1a166e09825c1e07f6417b65bc3b486dbdc559b60128da4a1a3f37da286.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 419, + 345, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 345, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 345, + 430 + ], + "type": "text", + "content": "where:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 431, + 553, + 550 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "type": "text", + "content": ": The infrared radiation coefficient, determining the absorption level of infrared radiation by different components. This value typically varies depending on the infrared wavelength (e.g., near-infrared, shortwave infrared, or mid-to-far infrared)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 491, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 553, + 550 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 491, + 553, + 550 + ], + "type": "inline_equation", + "content": "d(z)" + }, + { + "bbox": [ + 314, + 491, + 553, + 550 + ], + "type": "text", + "content": ": The scene depth, representing the distance between the target and the imaging device. In infrared imaging, a greater distance results in lower transmission, as infrared radiation undergoes greater loss in the atmosphere and other media." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 551, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 553, + 574 + ], + "type": "text", + "content": "Given a 2D image, we need a function to calculate its corresponding depth. This process can be expressed as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 399, + 580, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 580, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 399, + 580, + 553, + 594 + ], + "type": "interline_equation", + "content": "d (z) = F (I (z)). \\tag {11}", + "image_path": "2006fced6a969c471e6a2ca4c0e7744e4f1a22626dba1589ca1fe4d6ad786044.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 598, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 635 + ], + "type": "text", + "content": "Considering the scene's varying distances, classification techniques can be employed. Cao et al.[3] treated the depth estimation problem as a pixel-level classification task." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 380, + 646, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 646, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 380, + 646, + 553, + 659 + ], + "type": "interline_equation", + "content": "F (I (z)) = \\operatorname {R e s N e t} (I (x)). \\tag {12}", + "image_path": "415415d6bb32e988a6044db2d96dba1cb351a042db190bebc4e1f760bb382f3d.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 662, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 553, + 685 + ], + "type": "text", + "content": "Using the above equations, we can calculate the infrared target's radiation intensity:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 370, + 689, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 689, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 370, + 689, + 553, + 717 + ], + "type": "interline_equation", + "content": "J (z) = \\frac {I (z)}{t (z)} - \\frac {A (z)}{t (z)} + A (z). \\tag {13}", + "image_path": "36190b5927bd601a20c673ed1ca089ad14e3dacd5f024a23182710c0c130b7bd.jpg" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9553" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 241, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 241, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 241, + 83 + ], + "type": "text", + "content": "Substituting the transmission factor yields:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 88, + 295, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 88, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 64, + 88, + 295, + 103 + ], + "type": "interline_equation", + "content": "J (z) = I (z) \\cdot e ^ {\\beta \\cdot F (I (z))} - A (z) \\cdot \\left(e ^ {\\beta \\cdot F (I (z))} - 1\\right). \\tag {14}", + "image_path": "093e228bbbab1870e1537ef18813b4e25e98bc1bddee962f6e0c652f942918d4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 295, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 193 + ], + "type": "text", + "content": "Through the formulations outlined above, we can more effectively separate the target from the background, thereby improving the performance of infrared small target detection. This approach constructs an interpretable mathematical model based on the fundamental principles of infrared imaging, providing a robust and explainable solution for infrared small target detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 196, + 124, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 124, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 124, + 207 + ], + "type": "text", + "content": "3.3.2. CG-SAM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "type": "text", + "content": "Based on the infrared imaging equation (Eq. 14), we achieve precise infrared small target detection by inversely separating infrared degraded images from background images. As shown in Fig. 2, the text_prompt and visual_prompt are input into the SAM mask decoder, where they are fused with the early and late global semantic feature maps extracted by the ViT encoder. Building upon this, text_prompt and visual_prompt effectively extract and identify background features " + }, + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "type": "inline_equation", + "content": "A(z)" + }, + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "type": "text", + "content": " from different scenes through the scene prompting mechanism. Similarly, we feed the learnable output token from SAM and the image embeddings from the SAM encoder into the mask decoder to further refine the target mask representation. This process gradually recovers target features, but due to background interference, it ultimately generates the infrared degraded image " + }, + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "type": "inline_equation", + "content": "I(z)" + }, + { + "bbox": [ + 55, + 212, + 295, + 403 + ], + "type": "text", + "content": " with background noise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "type": "text", + "content": "Finally, based on the degraded image " + }, + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "I(z)" + }, + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "type": "text", + "content": " and the background features " + }, + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "A(z)" + }, + { + "bbox": [ + 55, + 403, + 295, + 499 + ], + "type": "text", + "content": ", we apply the infrared imaging equation to achieve precise detection of infrared small targets in complex backgrounds while effectively suppressing background interference. Through this integrated process, the CG-SAM can accurately identify small targets and significantly improve the accuracy of target detection in infrared images with strong background noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 118, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 118, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 118, + 516 + ], + "type": "text", + "content": "3.4. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "The goal of data generation is to create descriptive text for infrared images to support small target detection in complex backgrounds. To achieve this, we enhanced the IRSTD-1k, NUAA-SIRST, and NUDT-SIRST datasets by offering textual descriptions, resulting in the multimodal image-text dataset collections MIRSTD. Each infrared image, with the target mask removed, is encoded as a Base64 string and used as a text prompt for a visual-language model. ChatGPT generates concise descriptions that summarize key scene information, which are checked manually. The expert's prompt follows the format: a photo of a scene with [object], [object], and [object] offering a brief description of the scene and its objects. The final output is a 25-word description that captures the image's essential background information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 441, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 441, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 441, + 104 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "text", + "content": "Dataset: We evaluate our model on the MIRSTD dataset, which consists of 427 and 1,000 real infrared images with one or more small targets, and 1,327 synthetic infrared images with small targets. All images are resized to " + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "text", + "content": ". For each dataset, we split the images into three disjoint subsets: " + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "text", + "content": " for training, " + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "text", + "content": " for validation, and " + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 109, + 553, + 192 + ], + "type": "text", + "content": " for testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 193, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 361 + ], + "type": "text", + "content": "Implementation Details: Our SAIST model is trained using the Adam optimizer with an initial learning rate of 0.0001, and a Cosine Decay Learning Rate Scheduler over 400 epochs. We use a batch size of 4 and train on a single Nvidia GeForce A800 GPU with 80 GB of memory. For comparison, we evaluate against state-of-the-art CNN-based methods for IRSTD, including ISNet [39], GCINet [41], UIUNet [33], DNANet [19], Dim2Clear [40], ALCNet [8], and ACMNet [7], as well as hybrid methods such as SCTransNet [35], IAANet [31], SAM [16] and SAM-HQ [15]. Traditional methods Top-Hat [1] and IPI [10] are also included. We adopt Intersection over Union (IoU) for pixel-level evaluation, and Probability of Detection (Pd) and False Alarm Rate (Fa) for object-level evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 369, + 460, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 460, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 460, + 382 + ], + "type": "text", + "content": "4.2. Comparative Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 387, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 506 + ], + "type": "text", + "content": "Quantitative results: As shown in Tab. 1, traditional methods struggle with complex scenarios, especially CNNs, which fail to distinguish targets from backgrounds, resulting in higher false alarm and missed detection rates. The proposed SAIST outperforms existing SOTA methods across all metrics, particularly in reducing false alarms. SR-CLIP improves textual prompt generation through visual-textual alignment, while CG-SAM leverages both SR-CLIP's guidance and infrared imaging equations to enhance small target recognition in complex backgrounds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 507, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 590 + ], + "type": "text", + "content": "Visual Results: Fig. 4 shows detection results for SR-CLIP and CG-SAM, outperforming other IRSTD methods. SR-CLIP benefits from precise visual and textual guidance, while CG-SAM effectively integrates infrared imaging equations. Both methods excel in identifying small targets in complex backgrounds, reducing " + }, + { + "bbox": [ + 313, + 507, + 554, + 590 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 313, + 507, + 554, + 590 + ], + "type": "text", + "content": " and background noise, thus enhancing detection accuracy and robustness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 591, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 639 + ], + "type": "text", + "content": "ROC: As shown in Fig. 5, our SAIST significantly outperforms the other methods, where the Area Under the ROC Curve (AUC) of our SAIST is larger than that of the traditional and deep learning based methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 408, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 408, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 408, + 660 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "To evaluate the components of SAIST, we conducted ablation studies on the NUAA-SIRST dataset, using CLIP + SAM as the baseline model. We systematically introduced the CG-SAM and SR-CLIP modules. As shown in the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9554" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 70, + 506, + 268 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 506, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 506, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 506, + 268 + ], + "type": "table", + "html": "
MethodNUDT-SIRSTIRSTD-1kNUAA-SIRSTType
IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓IoU ↑Pd ↑Fa ↓
IPI [10]30.9381.9817.9927.9281.3716.181.0987.0530467Traditional
Top-Hat [1]22.4089.90174.110.0675.1114321.50879.7416456Traditional
MSLSTIPT [26]8.3447.4088111.4379.0315241.0800.0528183Traditional
ACM [7]68.9097.0511.2962.4191.4435.5870.7793.089.33CNN
ALCNet [8]81.4096.519.2662.0590.5821.7874.3173.1220.21CNN
FC3-Net [38]78.5693.8623.92265.0791.5415.5572.4498.1410.85CNN
ISNet [39]81.7796.344.4768.7795.5615.3980.0299.024.61CNN
DNA-Net [19]88.1998.629.0069.3893.311.6679.2698.485.35CNN
UIU-Net [33]92.1997.7715.4469.9693.9822.0778.2597.454.29CNN
Dim2Clear[40]81.3796.239.1766.3493.7520.9377.2999.106.72CNN
GCI-Net [41]92.4398.258.9667.7593.8912.8478.8199.344.11CNN
IAANet [31]90.2297.268.3266.2593.1514.2074.2293.5322.70Hybrid
SCTransNet [35]94.0996.954.2968.0393.2710.7477.5096.9513.92Hybrid
SAM [16]84.2097.3813.2868.1293.648.3875.2197.079.78SAM
SAM-HQ [15]88.0298.3114.4868.8593.549.5675.2797.226.87SAM
SAIST (Ours)95.2399.281.3172.1496.184.7680.8299.560.87CLIP + SAM
", + "image_path": "4dfb73c28be454c58d3113c4a255a58091801bb1df9f8f3296c9771721057c80.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": "Table 1. Comparison with existing IRSTD approaches on the NUDT-SIRST, IRSTD-1k and NUAA-SIRST datasets. The evaluation metrics are IoU " + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "(10^{-2})" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "P_d" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "(10^{-2})" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "F_a" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "(10^{-6})" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": ", the best results are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 303, + 555, + 533 + ], + "blocks": [ + { + "bbox": [ + 61, + 303, + 555, + 533 + ], + "lines": [ + { + "bbox": [ + 61, + 303, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 555, + 533 + ], + "type": "image", + "image_path": "1d96c6c8b8b965a304eef7db628288b70e95e8bddac893540e0cfc8e45487107.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 555, + 568 + ], + "type": "text", + "content": "Figure 4. Visual results of different IRSTD methods. The boxes in red, yellow, and blue represent correct, missed, and false detections, respectively. Close-up views are shown in the corners." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 588, + 294, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 294, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 294, + 708 + ], + "type": "text", + "content": "Tab. 1, removing CG-SAM significantly reduces the IoU and Pd scores while increasing the Fa score. Additionally, as illustrated in the Fig. 7, compared to other methods, CG-SAM can precisely identify small targets in complex backgrounds. These results validate the effectiveness of CG-SAM in enhancing small target representation and suppressing background interference. Similarly, removing SR-CLIP increases the Fa score. Meanwhile, Fig. 6 demonstrates that SR-CLIP effectively suppresses background interference and noise. These results suggest that SR-CLIP gen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 589, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 649 + ], + "type": "text", + "content": "erates accurate visual-textual prompts for CG-SAM. Ultimately, as shown in Table. 3, although SAIST does not have an advantage in inference time and the number of trainable parameters, it achieves the best performance across all evaluation metrics, demonstrating the excellence of its design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 660, + 443, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 443, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 443, + 672 + ], + "type": "text", + "content": "4.3.1. Impact of the SR-CLIP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "In this ablation study, we provide identical text prompts to CLIP, CoOp, and SR-CLIP, utilizing their respective vision and text encoders for image understanding and caption gen" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9555" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 282, + 260 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 282, + 260 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 282, + 260 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 282, + 260 + ], + "type": "image", + "image_path": "bfe7a2b668ffabef3593e520a8dd0f6ac35b767ddae03f6a5838912d9f621362.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 261, + 268, + 271 + ], + "lines": [ + { + "bbox": [ + 83, + 261, + 268, + 271 + ], + "spans": [ + { + "bbox": [ + 83, + 261, + 268, + 271 + ], + "type": "text", + "content": "Figure 5. ROC curves on NUAA-SIRST database." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 283, + 102, + 315 + ], + "blocks": [ + { + "bbox": [ + 57, + 283, + 102, + 315 + ], + "lines": [ + { + "bbox": [ + 57, + 283, + 102, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 283, + 102, + 315 + ], + "type": "image", + "image_path": "e702909692955b2b2b67b9aea117129b26a46fa79c68cfc1799b601c5df843a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 316, + 103, + 328 + ], + "lines": [ + { + "bbox": [ + 58, + 316, + 103, + 328 + ], + "spans": [ + { + "bbox": [ + 58, + 316, + 103, + 328 + ], + "type": "text", + "content": "A photo of tower crane A photo of a wall" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 104, + 283, + 146, + 315 + ], + "blocks": [ + { + "bbox": [ + 104, + 283, + 146, + 315 + ], + "lines": [ + { + "bbox": [ + 104, + 283, + 146, + 315 + ], + "spans": [ + { + "bbox": [ + 104, + 283, + 146, + 315 + ], + "type": "image", + "image_path": "19ced1e25a62092afca593c20a4afc9dd36b4a340ee179e2211b13a2b29a2e6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 316, + 192, + 327 + ], + "lines": [ + { + "bbox": [ + 149, + 316, + 192, + 327 + ], + "spans": [ + { + "bbox": [ + 149, + 316, + 192, + 327 + ], + "type": "text", + "content": "CLIP" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 149, + 283, + 197, + 314 + ], + "blocks": [ + { + "bbox": [ + 149, + 283, + 197, + 314 + ], + "lines": [ + { + "bbox": [ + 149, + 283, + 197, + 314 + ], + "spans": [ + { + "bbox": [ + 149, + 283, + 197, + 314 + ], + "type": "image", + "image_path": "675c6d3dab60bd9efd15c93a298b96d9cab9c38f7a3ff67c8d0c0d7d74a0db52.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 316, + 242, + 327 + ], + "lines": [ + { + "bbox": [ + 198, + 316, + 242, + 327 + ], + "spans": [ + { + "bbox": [ + 198, + 316, + 242, + 327 + ], + "type": "text", + "content": "CoOP" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 198, + 283, + 242, + 314 + ], + "blocks": [ + { + "bbox": [ + 198, + 283, + 242, + 314 + ], + "lines": [ + { + "bbox": [ + 198, + 283, + 242, + 314 + ], + "spans": [ + { + "bbox": [ + 198, + 283, + 242, + 314 + ], + "type": "image", + "image_path": "d44ec6f18c5ddd409b739eb8b4077a410887ee4e56270ed29fa4d20ae15d961e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 316, + 282, + 327 + ], + "lines": [ + { + "bbox": [ + 244, + 316, + 282, + 327 + ], + "spans": [ + { + "bbox": [ + 244, + 316, + 282, + 327 + ], + "type": "text", + "content": "SR-CLIP" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 244, + 283, + 291, + 314 + ], + "blocks": [ + { + "bbox": [ + 244, + 283, + 291, + 314 + ], + "lines": [ + { + "bbox": [ + 244, + 283, + 291, + 314 + ], + "spans": [ + { + "bbox": [ + 244, + 283, + 291, + 314 + ], + "type": "image", + "image_path": "6394df9ca117440a0bc7933bded9da623484a4a71010364615f35ba004cf4731.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 57, + 329, + 102, + 361 + ], + "blocks": [ + { + "bbox": [ + 104, + 316, + 146, + 327 + ], + "lines": [ + { + "bbox": [ + 104, + 316, + 146, + 327 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 146, + 327 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 329, + 102, + 361 + ], + "lines": [ + { + "bbox": [ + 57, + 329, + 102, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 329, + 102, + 361 + ], + "type": "image", + "image_path": "d01c28366d9cc90b9b54b0ddd7f82c554a57fac39e2750848b8c98ce30f140d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 104, + 329, + 148, + 361 + ], + "blocks": [ + { + "bbox": [ + 104, + 329, + 148, + 361 + ], + "lines": [ + { + "bbox": [ + 104, + 329, + 148, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 329, + 148, + 361 + ], + "type": "image", + "image_path": "9b4f1b0b6103cb42bbc6e60f0c3c1b10eb4a816b246b4961949cf4ac21516995.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 150, + 329, + 197, + 361 + ], + "blocks": [ + { + "bbox": [ + 150, + 329, + 197, + 361 + ], + "lines": [ + { + "bbox": [ + 150, + 329, + 197, + 361 + ], + "spans": [ + { + "bbox": [ + 150, + 329, + 197, + 361 + ], + "type": "image", + "image_path": "ad1f2031f3a9d88e9628bd2926ef9353f5fd0f1777988b6a9a750ed8fc60c990.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 198, + 329, + 242, + 361 + ], + "blocks": [ + { + "bbox": [ + 198, + 329, + 242, + 361 + ], + "lines": [ + { + "bbox": [ + 198, + 329, + 242, + 361 + ], + "spans": [ + { + "bbox": [ + 198, + 329, + 242, + 361 + ], + "type": "image", + "image_path": "b6bb55e26c3ec5a69d8ef0eebcef5a8abfc4d8689238ba02d2bdd51a376cb319.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 244, + 328, + 291, + 361 + ], + "blocks": [ + { + "bbox": [ + 244, + 328, + 291, + 361 + ], + "lines": [ + { + "bbox": [ + 244, + 328, + 291, + 361 + ], + "spans": [ + { + "bbox": [ + 244, + 328, + 291, + 361 + ], + "type": "image", + "image_path": "da05fe9a18f62eb3ab9e3ddef3ee523cb8e110cc9a3eec9aaf9326a186528bda.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 389, + 104, + 421 + ], + "blocks": [ + { + "bbox": [ + 68, + 377, + 282, + 388 + ], + "lines": [ + { + "bbox": [ + 68, + 377, + 282, + 388 + ], + "spans": [ + { + "bbox": [ + 68, + 377, + 282, + 388 + ], + "type": "text", + "content": "Figure 6. Visual comparison of CLIP, CoOp and SR-CLIP." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 389, + 104, + 421 + ], + "lines": [ + { + "bbox": [ + 58, + 389, + 104, + 421 + ], + "spans": [ + { + "bbox": [ + 58, + 389, + 104, + 421 + ], + "type": "image", + "image_path": "89f2d620ad127f7e123b498f4892c36c068f100f01e347b321e720cff8e450d0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 106, + 389, + 152, + 421 + ], + "blocks": [ + { + "bbox": [ + 106, + 389, + 152, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 152, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 152, + 421 + ], + "type": "image", + "image_path": "3578ce996adb9a9f918e0fecfc2a0c9d62d798839c51ecbf9bd9843f4aa72ddf.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 154, + 389, + 199, + 421 + ], + "blocks": [ + { + "bbox": [ + 154, + 389, + 199, + 421 + ], + "lines": [ + { + "bbox": [ + 154, + 389, + 199, + 421 + ], + "spans": [ + { + "bbox": [ + 154, + 389, + 199, + 421 + ], + "type": "image", + "image_path": "e12787b2d4a3710072db5241d5f49fb0267389a0ac999a16c1feed588f9995d3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 200, + 389, + 243, + 421 + ], + "blocks": [ + { + "bbox": [ + 200, + 389, + 243, + 421 + ], + "lines": [ + { + "bbox": [ + 200, + 389, + 243, + 421 + ], + "spans": [ + { + "bbox": [ + 200, + 389, + 243, + 421 + ], + "type": "image", + "image_path": "6c9345d8838947389a8bdd9cd4d9407712c2661c1122c1023788ab796cc4bf90.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 244, + 389, + 291, + 421 + ], + "blocks": [ + { + "bbox": [ + 244, + 389, + 291, + 421 + ], + "lines": [ + { + "bbox": [ + 244, + 389, + 291, + 421 + ], + "spans": [ + { + "bbox": [ + 244, + 389, + 291, + 421 + ], + "type": "image", + "image_path": "f6291f93359b9f51d5f85b57127e6dd492c31470f2fed12f190decc93e0df023.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 430, + 104, + 462 + ], + "blocks": [ + { + "bbox": [ + 58, + 430, + 104, + 462 + ], + "lines": [ + { + "bbox": [ + 58, + 430, + 104, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 430, + 104, + 462 + ], + "type": "image", + "image_path": "8b94aec4c3f1ed7f92b0a5b20232eae067c1dadb7d0f3a27bc5596116ed2bc42.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 472, + 291, + 484 + ], + "lines": [ + { + "bbox": [ + 59, + 472, + 291, + 484 + ], + "spans": [ + { + "bbox": [ + 59, + 472, + 291, + 484 + ], + "type": "text", + "content": "Figure 7. Visual comparison of SAM, HQ-SAM and CG-SAM." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 106, + 430, + 152, + 462 + ], + "blocks": [ + { + "bbox": [ + 115, + 422, + 143, + 430 + ], + "lines": [ + { + "bbox": [ + 115, + 422, + 143, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 422, + 143, + 430 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 430, + 152, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 152, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 152, + 462 + ], + "type": "image", + "image_path": "c8dc122fa5ce9c4e2b9ef0be4463f77dd8cdc700e7c387c5774c6c272d36aab9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 154, + 430, + 199, + 462 + ], + "blocks": [ + { + "bbox": [ + 154, + 430, + 199, + 462 + ], + "lines": [ + { + "bbox": [ + 154, + 430, + 199, + 462 + ], + "spans": [ + { + "bbox": [ + 154, + 430, + 199, + 462 + ], + "type": "image", + "image_path": "25b87a03c11f31397f13730562f1edda62012238a344a47e6d568dd1e4121eeb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 200, + 430, + 243, + 462 + ], + "blocks": [ + { + "bbox": [ + 200, + 430, + 243, + 462 + ], + "lines": [ + { + "bbox": [ + 200, + 430, + 243, + 462 + ], + "spans": [ + { + "bbox": [ + 200, + 430, + 243, + 462 + ], + "type": "image", + "image_path": "5d5e002284ffe5308c687460b78a40594678cbc24666e72bc1e8b0d917ec2252.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 244, + 430, + 291, + 462 + ], + "blocks": [ + { + "bbox": [ + 244, + 430, + 291, + 462 + ], + "lines": [ + { + "bbox": [ + 244, + 430, + 291, + 462 + ], + "spans": [ + { + "bbox": [ + 244, + 430, + 291, + 462 + ], + "type": "image", + "image_path": "f94aa8fe9f0d285296037f813fc0f85d2233b149acce2a92916231c5c758633a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 80, + 524, + 269, + 576 + ], + "blocks": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": "Table 2. Ablation study of SR-CLIP and CG-SAM in " + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "inline_equation", + "content": "IoU(\\%)" + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "inline_equation", + "content": "P_d(\\%)" + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "inline_equation", + "content": "F_a(10^{-6})" + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": " on on NUAA-SIRST." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 524, + 269, + 576 + ], + "lines": [ + { + "bbox": [ + 80, + 524, + 269, + 576 + ], + "spans": [ + { + "bbox": [ + 80, + 524, + 269, + 576 + ], + "type": "table", + "html": "
MethodIoUPdFa
SAIST80.8299.560.87
w/o SR-CLIP80.0298.776.61
w/o CG-SAM78.8797.428.53
w/o CG-SAM & SR-CLIP76.8296.5012.37
", + "image_path": "96bb796f5a291b6b725c2a946ace03e33b22dcc4f7efdf29013d4b7e13799de6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 585, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 668 + ], + "type": "text", + "content": "eration, to systematically evaluate the performance of SR-CLIP. As shown in Fig. 6 and the Tab. 4, we demonstrate that SR-CLIP effectively suppresses background interference and noise, enabling the generation of precise visual-textual prompts for CG-SAM. These results highlight the efficacy of SR-CLIP in improving the alignment between the visual and textual modalities." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 674, + 185, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 185, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 185, + 686 + ], + "type": "text", + "content": "4.3.2. Impact of the CG-SAM" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In this ablation study, we utilize the same SAM encoder while employing different Mask Decoders: SAM, HQ-" + } + ] + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 315, + 99, + 553, + 125 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Comparison of different models in " + }, + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "inline_equation", + "content": "IoU(\\%)" + }, + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "inline_equation", + "content": "P_d(\\%)" + }, + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "inline_equation", + "content": "F_a(10^{-6})" + }, + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "text", + "content": " on MIRSTD." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 99, + 553, + 125 + ], + "lines": [ + { + "bbox": [ + 315, + 99, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 315, + 99, + 553, + 125 + ], + "type": "table", + "html": "
ModelTrainable Params (M)Inference (s)Avg IoUAvg PdAvg Fa
SAM + CLIP383.280.07978.4996.678.45
SAIST389.570.08182.7398.342.31
", + "image_path": "0d3bec773a67fd9e20a61c72dcc105355e7f66cd14e631b2360ecd51636ca63f.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_body" + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 347, + 163, + 520, + 205 + ], + "blocks": [ + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "lines": [ + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "text", + "content": "Table 4. Ablation study of CLIP, CoOp, and SR-CLIP in " + }, + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "inline_equation", + "content": "IoU(\\%)" + }, + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "inline_equation", + "content": "P_d(\\%)" + }, + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "inline_equation", + "content": "F_a(10^{-6})" + }, + { + "bbox": [ + 314, + 135, + 555, + 157 + ], + "type": "text", + "content": " on NUAA-SIRST." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 163, + 520, + 205 + ], + "lines": [ + { + "bbox": [ + 347, + 163, + 520, + 205 + ], + "spans": [ + { + "bbox": [ + 347, + 163, + 520, + 205 + ], + "type": "table", + "html": "
MethodIoUPdFa
CLIP + CG-SAM80.0298.776.61
CoOp + CG-SAM80.3298.564.87
SR-CLIP + CG-SAM80.8299.560.87
", + "image_path": "99b4b75587a5be8876901e585cb88111bcb2280ec028deb80dd7ec9b6291e82a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_body" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 347, + 243, + 520, + 285 + ], + "blocks": [ + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "lines": [ + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "text", + "content": "Table 5. Ablation study of SAM, HQ-SAM, and CG-SAM in " + }, + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "inline_equation", + "content": "IoU(\\%)" + }, + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "inline_equation", + "content": "P_d(\\%)" + }, + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "inline_equation", + "content": "F_a(10^{-6})" + }, + { + "bbox": [ + 314, + 215, + 555, + 238 + ], + "type": "text", + "content": " on NUAA-SIRST." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 243, + 520, + 285 + ], + "lines": [ + { + "bbox": [ + 347, + 243, + 520, + 285 + ], + "spans": [ + { + "bbox": [ + 347, + 243, + 520, + 285 + ], + "type": "table", + "html": "
MethodIoUPdFa
SAM + SR-CLIP78.8797.425.23
HQ-SAM + SR-CLIP79.4598.894.56
CG-SAM + SR-CLIP80.8299.560.87
", + "image_path": "3b181f07a215539ecb8bb8fdc9bdb96f35b0ed3a4167ebf4e7d005384a80b123.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 297, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 405 + ], + "type": "text", + "content": "SAM, and CG-SAM. Identical fine-grained visual and textual prompts, generated by SR-CLIP, are provided as inputs. The performance of CG-SAM is systematically evaluated. As demonstrated in Fig. 7 and the Tab. 5, our results confirm that CG-SAM is capable of accurately identifying small targets in complex backgrounds. These findings validate that, with the aid of infrared imaging equations, CG-SAM enhances the representation of small targets while effectively suppressing background interference." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 314, + 418, + 388, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 388, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 388, + 430 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 439, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 559 + ], + "type": "text", + "content": "This paper presents SAIST, a novel framework that integrates visual and textual modalities for enhanced IRSTD. Leveraging SR-CLIP to capture cross-modal relationships, SAIST generates precise prompts to overcome the limitations of image-only methods. Guided by these prompts, CG-SAM employs interpretable infrared imaging equations to extract targets while suppressing noise. We also introduce MIRSTD, the first multimodal IRSTD dataset with rich image-text pairs. Experiments demonstrate the effectiveness of multimodal integration in IRSTD." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 573, + 424, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 424, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 424, + 586 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62272363, Grant 92470108, Grant 62371362, and Grant U21A20514; in part by the Joint Laboratory for Innovation in Satellite-Borne Computers and Electronics Technology Open Fund 2023 under Grant 2024KFKT001-1; in part by the Proof of Concept Foundation of Xidian University Hangzhou of Technology under Grant No. GNYZ2023YL0301; and part by the Shaanxi Provincial Key R&D Program Project under Grant 2024SF-YBXM-330." + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9556" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Xiangzhi Bai and Fugen Zhou. Analysis of new top-hat transformation and the application for infrared dim small target detection. Pattern Recognition, 43(6):2145-2156, 2010. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "type": "text", + "content": "[2] Karsten M Borgwardt, Arthur Gretton, Malte J Rasch, Hans-Peter Kriegel, Bernhard Scholkopf, and Alex J Smola. Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49-e57, 2006. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Yuanzhouhan Cao, Zifeng Wu, and Chunhua Shen. Estimating depth from monocular images as classification using deep fully convolutional residual networks. IEEE Transactions on Circuits and Systems for Video Technology, 28(11): 3174-3182, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "text", + "content": "[4] CL Philip Chen, Hong Li, Yantao Wei, Tian Xia, and Yuan Yan Tang. A local contrast method for small infrared target detection. IEEE transactions on geoscience and remote sensing, 52(1):574-581, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 323 + ], + "type": "text", + "content": "[5] G Cucurullo, L Giordano, D Albanese, Luciano Cinquanta, and M Di Matteo. Infrared thermography assisted control for apples microwave drying. Journal of food engineering, 112 (4):319-325, 2012. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "text", + "content": "[6] Yimian Dai and Yiquan Wu. Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection. IEEE journal of selected topics in applied earth observations and remote sensing, 10(8): 3752-3767, 2017. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "text", + "content": "[7] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 950–959, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[8] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 59(11):9813-9824, 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 533 + ], + "type": "text", + "content": "[9] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G. Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE Transactions on Image Processing, 22(12):4996-5009, 2013. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[10] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multistage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1–14, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[12] Jinhui Han, Saed Moradi, Iman Faramarzi, Honghui Zhang, Qian Zhao, Xiaojian Zhang, and Nan Li. Infrared small target detection based on the weighted strengthened local contrast measure. IEEE Geoscience and Remote Sensing Letters, 18(9):1670-1674, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] MD. Zakir Hossain, Ferdous Sohel, Mohd Fairuz Shiratuddin, and Hamid Laga. A comprehensive survey of deep learn" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "ing for image captioning. ACM Comput. Surv., 51(6), 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[14] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[15] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[17] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 353 + ], + "type": "text", + "content": "[18] Wing-Cheung Law, Zhourui Xu, Ken-Tye Yong, Xin Liu, Mark T Swihart, Mukund Seshadri, and Paras N Prasad. Manganese-doped near-infrared emitting nanocrystals for in vivo biomedical imaging. Optics express, 24(16):17553-17561, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "text", + "content": "[19] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 2022. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[20] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 489 + ], + "type": "text", + "content": "[21] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "text", + "content": "[22] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 546, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 622 + ], + "type": "text", + "content": "[23] Alexandre Mallet, Roumiana Tsenkova, Jelena Muncan, Cyrille Charnier, Éric Latrille, Ryad Bendoula, Jean-Philippe Steyer, and Jean-Michel Roger. Relating near-infrared light path-length modifications to the water content of scattering media in near-infrared spectroscopy: toward a new bouguer-beer-lambert law. Analytical Chemistry, 93 (17):6817-6823, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9557" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[26] Yang Sun, Jungang Yang, and Wei An. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor model. IEEE Transactions on Geoscience and Remote Sensing, 59(5):3737-3752, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[27] Danica J Sutherland, Hsiao-Yu Tung, Heiko Strathmann, Soumyajit De, Aaditya Ramdas, Alex Smola, and Arthur Gretton. Generative models and model criticism via optimized maximum mean discrepancy. arXiv preprint arXiv:1611.04488, 2016. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 294, + 229 + ], + "type": "text", + "content": "[28] Donald F Swinehart. The beer-lambert law. Journal of chemical education, 39(7):333, 1962. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 295, + 274 + ], + "type": "text", + "content": "[29] Michael Teutsch and Wolfgang Krüger. Classification of small boats in infrared images for maritime surveillance. In 2010 International WaterSide Security Conference, pages 1-7. IEEE, 2010. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "type": "text", + "content": "[30] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8509-8518, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[31] Kewei Wang, Shuaiyuan Du, Chengxin Liu, and Zhiguo Cao. Interior attention-aware network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 378, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 378, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 378, + 295, + 432 + ], + "type": "text", + "content": "[32] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 434, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 294, + 466 + ], + "type": "text", + "content": "[33] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 468, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 468, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 57, + 468, + 294, + 511 + ], + "type": "text", + "content": "[34] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 555 + ], + "type": "text", + "content": "[35] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 558, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 590 + ], + "type": "text", + "content": "[36] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "text", + "content": "[37] Mingjin Zhang, Haichen Bai, Jing Zhang, Rui Zhang, Chaoyue Wang, Jie Guo, and Xinbo Gao. Rkformer: Runge-kutta transformer with random-connection attention for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1730–1738, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Mingjin Zhang, Ke Yue, Jing Zhang, Yunsong Li, and Xinbo Gao. Exploring feature compensation and cross-level correlation for infrared small target detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1857-1865, 2022. 1, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 352 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[39] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[40] Mingjin Zhang, Rui Zhang, Jing Zhang, Jie Guo, Yunsong Li, and Xinbo Gao. Dim2clear network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-14, 2023. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[41] Mingjin Zhang, Ke Yue, Boyang Li, Jie Guo, Yunsong Li, and Xinbo Gao. Single-frame infrared small target detection via gaussian curvature inspired network. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 261 + ], + "type": "text", + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[43] Hu Zhu, Haopeng Ni, Shiming Liu, Guoxia Xu, and Lizhen Deng. TnIrs: Target-aware non-local low-rank modeling with saliency filtering regularization for infrared small target detection. IEEE Transactions on Image Processing, 29: 9546-9558, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[44] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 2023. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9558" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_content_list.json b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7eaabfd9fb1f5e08fc2992078552d85c449b7fea --- /dev/null +++ b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_content_list.json @@ -0,0 +1,1538 @@ +[ + { + "type": "text", + "text": "SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing", + "text_level": 1, + "bbox": [ + 259, + 130, + 738, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seokhyeon Hong Chaelin Kim Serin Yoon Junghyun Nam Sihun Cha Junyong Noh Visual Media Lab, KAIST", + "bbox": [ + 120, + 203, + 875, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ghd3079, chaelin.kim, serinyoon, ys4990, chacorp, junyongnoh}@kaist.ac.kr", + "bbox": [ + 222, + 241, + 769, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-driven motion generation has advanced significantly with the rise of denoising diffusion models. However, previous methods often oversimplify representations for the skeletal joints, temporal frames, and textual words, limiting their ability to fully capture the information within each modality and their interactions. Moreover, when using pretrained models for downstream tasks, such as editing, they typically require additional efforts, including manual interventions, optimization, or fine-tuning. In this paper, we introduce a skeleton-aware latent diffusion (SALAD), a model that explicitly captures the intricate inter-relationships between joints, frames, and words. Furthermore, by leveraging cross-attention maps produced during the generation process, we enable attention-based zero-shot text-driven motion editing using a pre-trained SALAD model, requiring no additional user input beyond text prompts. Our approach significantly outperforms previous methods in terms of text-motion alignment without compromising generation quality, and demonstrates practical versatility by providing diverse editing capabilities beyond generation. Code is available at project page.", + "bbox": [ + 89, + 324, + 485, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 660, + 220, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Character animation is a crucial component in various computer graphics and vision applications, including games, films, and interactive media. Traditional methods, such as keyframing and motion capture, typically require extensive manual effort, which are time-consuming and expensive. Recently, deep generative models have been introduced to mitigate these challenges. In particular, diffusion models have shown promising results in text-to-motion generation, enabling intuitive and efficient animation workflows.", + "bbox": [ + 89, + 686, + 482, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, fully capturing the intricate relationships among frames, body parts, and textual descriptions remains a complex challenge in text-driven motion generation. Previous approaches typically represent a pose as a single vector and primarily focus on temporal re", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "relationships between poses across frames, while neglecting spatial interactions between skeletal joints. Furthermore, these models often compress a sentence into a single vector for conditioning, which can overlook the important nuances of word-level variations. This over-simplification of interactions among skeletal, temporal, and textual components may lead to missing details in the generated results, underscoring the need for models that faithfully capture these complex dependencies.", + "bbox": [ + 511, + 292, + 906, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, learning meaningful representations is a key factor to enable zero-shot downstream tasks with pre-trained models. For example, zero-shot text-driven image and video editing methods [3, 7, 13, 15, 25, 35, 43] leverage attention modulation from pre-trained models for intuitive manipulation. Unfortunately, motion generation models typically lack such interpretable and manipulatable intermediate representations. This limitation stems from the over-simplification of motion and text features mentioned above, which restricts rich interactions between them. Although some methods exploit pre-trained models with manual masks, optimization, or fine-tuning [21-23, 42, 45, 47], these methods require additional effort, time, and cost to achieve desired results. Therefore, developing interpretable representations during motion synthesis could improve versatility and flexibility for downstream tasks beyond generation, such as zero-shot motion editing.", + "bbox": [ + 511, + 430, + 908, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a skeleton-aware latent diffusion, which we call SALAD, for text-driven motion generation within a skeletotemporally structured latent space. We first train a variational autoencoder (VAE) [24] to construct a motion latent space that decouples spatial and temporal dimensions. To this end, we employ skeletotemporal convolution layers to facilitate information exchange between adjacent joints and frames. We also employ skeletotemporal pooling layers to produce a compact representation by reducing the dimensionality, which can reduce the computational complexity during the sampling process of the diffusion model. Subsequently, we train a diffusion model within this latent space to generate motion features conditioned on text. To train the denoiser, we leverage both", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7158", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "skeletal and temporal attention blocks, which can effectively enable skeletotemporal coherence. Additionally, we employ cross-attention to capture fine-grained interactions between individual words and the skeletotemporal units of the motion latents.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond motion generation, we introduce an attention-based zero-shot, text-driven motion editing method that leverages the pre-trained SALAD model, requiring no additional optimization or fine-tuning. We demonstrate that the intermediate cross-attention maps of SALAD capture the relationship between text and motion features. By modulating these cross-attention maps, we enable text-driven motion editing that allows users to edit generated motion through text input alone in a training-free manner. We also propose a novel attention modulation method for motion-specific tasks, demonstrating the practical potential of our approach.", + "bbox": [ + 89, + 167, + 480, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 349, + 410, + 364 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SALAD, a novel skeleton-aware latent diffusion model for text-driven motion generation within a skeletotemporally structured latent space.", + "- We interpret the intermediate representations in the generation process, allowing for a clear understanding of the relationship between text inputs and generated motions.", + "- We present an attention-based zero-shot text-driven motion editing method using cross-attention modulation in generative models." + ], + "bbox": [ + 89, + 366, + 483, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 518, + 232, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-driven Motion Generation", + "text_level": 1, + "bbox": [ + 89, + 544, + 366, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of deep neural networks, generative models for human motion synthesis based on natural language descriptions have been extensively explored. Early studies focused on generating motions from action labels, such as throw and kick [9, 32]. More advanced methods projected both text and motion features into a shared latent space, enabling the generation of fixed motion sequences from identical input text [2]. To increase the diversity of generated results, T2M [10] employed a temporal VAE that can generate varied motion sequences that align with the given text input. Subsequent studies leveraged large pretrained text encoders, such as CLIP [36], to improve textual understanding for motion generation [4, 33, 41]. Autoregressive models that behave like language models have also been proposed, which first encode motion features into discrete tokens through vector quantization [44] and then auto-regressively predict the next motion token given the text prompt [11, 12, 20, 49, 52-54]. Recently, denoising diffusion models [17, 39] have become the dominant framework for generative models, prompting text-driven motion generation methods to adopt diffusion models as their backbone [8, 23, 27, 42, 46, 50, 51]. In this study, we build", + "bbox": [ + 89, + 569, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our framework on the diffusion model formulation and enhance the performance of text-driven motion generation through explicit modeling of the relationships among joints, frames, and words. Moreover, we introduce an attention-based zero-shot text-driven editing method, leveraging our pre-trained SALAD model as a generator, demonstrating a novel exploration beyond mere generation.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Skeleton-aware Processing of Motion Data", + "text_level": 1, + "bbox": [ + 511, + 208, + 874, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Because motion data consists of two separate skeletal and temporal dimensions, some methods have been developed to respect each dimension when processing the motion data. Yan et al. [48] presented spatial-temporal graph convolution networks that can learn both the spatial and temporal patterns of dynamic skeletons. Skeleton-aware networks [1] employed skeleto-temporal convolution networks and topology-preserving skeletal pooling layers to construct a shared motion latent space, enabling deep motion retargeting across homeomorphic skeletons. Motion style transfer has also been enabled in a part-wise manner by employing skeleto-temporal convolution modules [19, 30]. In the realm of text-to-motion generation, ParCo [54] and AttT2M [53] utilized the skeleto-temporal network architecture during the motion quantization process. In this work, we adopt a similar approach by encoding motion features using skeleto-temporal convolution networks and pooling layers to derive an effective skeleton-aware motion latent space with compact dimensionality. Additionally, we facilitate the interaction between this skeleton-aware motion latent space and textual features during the text-conditional generation process.", + "bbox": [ + 511, + 231, + 906, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Zero-shot Editing with Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 574, + 859, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By leveraging the generative capabilities of diffusion models, zero-shot editing methods have been proposed. SDEdit [28] first adds noise to the input data for a few diffusion timesteps, then denoises the noisy data with an editing condition to achieve high-fidelity zero-shot editing by leveraging a pre-trained diffusion model. Prompt-to-Prompt [15] verified that cross-attention maps establish connections between word tokens and the spatial layout of the generated image. In addition, it introduced attention modulation methods that can be utilized during the generation process, enabling the generation of high-fidelity editing results in a zero-shot manner. Null-text inversion [29] extended this method to real-images by improving the diffusion inversion method. Similarly, image and video editing methods via attention modulation have been extensively studied [3, 7, 13, 25, 35, 43]. Motion diffusion models often employ joint masks [23, 42] or optimization [21, 22, 47] to enable motion editing using pre-trained diffusion models, although they require user intervention or additional computation for the desired goal. On the other hand, CoMo [18]", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7159", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45f5cbd20cfd69199d653edb9d49ca94a573eba579db4e95c3cffc35baec46d6.jpg", + "image_caption": [ + "Figure 1. Architecture of the skeletotemporal VAE network. The encoder maps motion features into a skeletotemporal latent space, and the decoder restores the skeletotemporal latent variables into motion features." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 906, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "achieved zero-shot text-driven motion editing using large language models to manipulate semantic pose tokens. We demonstrate that the intermediate cross-attention maps of our SALAD model can capture the relationship between motion and text, similarly to image diffusion models. Furthermore, by modulating these cross-attention maps, we introduce an attention-based zero-shot text-driven motion editing method that eliminates the need for manual masks, fine-tuning, or optimization.", + "bbox": [ + 88, + 256, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 405, + 181, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to generate a motion sequence $\\mathbf{m}_{1:N}$ conditioned on a text prompt $c$ , where $N$ denotes the number of frames. We first construct a skeleton-aware motion latent representation that captures both skeletal and temporal dynamics of a motion sequence (Sec. 3.1). Using this motion representation, we train a diffusion model by modeling the complex interactions between skeletal joints, temporal frames, and textual descriptions (Sec. 3.2). Finally, we extend our model for zero-shot text-driven motion editing by modulating cross-attention maps between text and motion (Sec. 3.3). Additional details on the network architectures will be featured in the supplementary material.", + "bbox": [ + 88, + 431, + 483, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Skeleton-aware VAE", + "text_level": 1, + "bbox": [ + 89, + 619, + 285, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To derive skeleton-aware motion features, we train a VAE composed of skeletotemporal convolution networks and pooling layers [1, 48], as illustrated in Fig. 1. The core module is the skeletotemporal convolution (STConv), which decouples joint and frame dimensions while facilitating information exchange between adjacent components within each dimension. While the skeletotemporal separation of motion sequences allows for sophisticated modeling of motion data, it introduces an additional dimension for skeletal joints, unlike previous text-to-motion generation methods that treat a pose as a single vector. This expanded data space increases the risk of curse of dimensionality and leads to longer computation times for diffusion models, which rely on solving stochastic differential equations within this higher-dimensional data space. To address this issue, we apply skeletotemporal pooling (STPool) layers in the encoder to reduce the latent space dimensionality, effectively", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "compressing the representation. In the decoder, skeletotemporal unpooling (STUnpool) layers are used to reconstruct the compressed latent features back to raw motion.", + "bbox": [ + 511, + 256, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder. We separate each pose vector of $\\mathbf{m}_{1:N}$ into jointwise features and propagate them through joint-wise multilayer perceptron (MLP) layers, resulting in $\\mathbf{h} \\in \\mathbb{R}^{N \\times J \\times D}$ , where $J$ denotes the number of joints and $D$ denotes the latent dimension. Subsequently, we apply STConv to $\\mathbf{h}$ to facilitate information exchange between adjacent frames and joints:", + "bbox": [ + 511, + 301, + 905, + 407 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S T C o n v} (\\mathbf {h}) := \\operatorname {S k e l C o n v} (\\mathbf {h}) + \\operatorname {T e m p C o n v} (\\mathbf {h}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 412, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where SkelConv is a graph convolution network over the joint dimension, while TempConv is a 1D convolution network over the frame dimension.", + "bbox": [ + 511, + 435, + 905, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We apply the STPool operator to the output of the STConv layers to reduce its dimensionality by pooling over joints and frames:", + "bbox": [ + 511, + 479, + 905, + 525 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S T P o o l} (\\mathbf {h}) := \\operatorname {T e m p P o o l} (\\operatorname {S k e l P o o l} (\\mathbf {h})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 531, + 903, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where SkelPool represents the pooling layer over the joint dimension while TempPool represents the pooling layer over the temporal dimension. Specifically, SkelPool aggregates adjacent joints while preserving its skeletal topology, resulting in fewer joints with a homeomorphic structure. In contrast, TempPool performs 1D pooling over frames. These pooling operations are commutative because they operate within independent dimensions.", + "bbox": [ + 511, + 551, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consequently, we obtain the skeleto-temporally compressed latent vector $\\mathbf{z} \\in \\mathbb{R}^{N' \\times J' \\times D}$ , where $N' < N$ and $J' < J$ . This representation effectively encodes the skeleton-aware dynamics of the motion sequence within a lower dimension. Notably, we maintain 7 atomic joints for this motion latent: root, spine, head, both arms, and both legs $(J' = 7)$ .", + "bbox": [ + 511, + 674, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder. To reconstruct $\\mathbf{z}$ into raw motion features, we employ STConv and STUnpool layers in the decoder. The decoder mirrors the encoder, facilitating information exchange between adjacent joints and frames using STConv layers and increasing the skeleto-temporal resolution using STUnpool layers. Finally, the joint-wise features are processed by their respective joint-wise MLP layers, resulting in the reconstructed motion feature $\\hat{\\mathbf{m}}_{1:N}$ .", + "bbox": [ + 511, + 780, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7160", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/723af0545820b071a09c43637c46d70284980d660693d8582d64bc1969f90f87.jpg", + "image_caption": [ + "Figure 2. (Left) Overall network architecture of the denoiser. (Right) The architecture of each transformer block consisting of TempAttn, SkelAttn, CrossAttn, and FFN, along with the FiLM following each module." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 444, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9591fe8ed372fd2599f635b889b8e84cdbf6b5fa236dfaf626402a8714601f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 85, + 897, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. The VAE is trained with the following objectives:", + "bbox": [ + 89, + 267, + 482, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V A E}} = \\mathcal {L} _ {\\mathbf {m}} + \\lambda_ {\\text {p o s}} \\mathcal {L} _ {\\text {p o s}} + \\lambda_ {\\text {v e l}} \\mathcal {L} _ {\\text {v e l}} + \\lambda_ {\\mathrm {k l}} \\mathcal {L} _ {\\mathrm {k l}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 294, + 482, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathbf{m}},\\mathcal{L}_{\\mathrm{pos}},\\mathcal{L}_{\\mathrm{vel}}$ are L1 reconstruction loss terms of the motion features, joint positions, and joint velocities, while $\\mathcal{L}_{\\mathrm{kl}}$ is the Kullback-Leibler (KL) divergence regularization that encourages a structured latent space.", + "bbox": [ + 89, + 320, + 483, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Skeleton-aware Denoiser", + "text_level": 1, + "bbox": [ + 89, + 390, + 316, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network Architecture. Given the skeleton-aware motion latent vector $\\mathbf{z} \\in \\mathbb{R}^{N' \\times J' \\times D}$ , we train a diffusion model that denoises the noised latent vector conditioned on text prompt $c$ . As shown in Fig. 2, our denoiser consists of a positional embedding, MLP-based encoder and decoder, and $L$ transformer-based layers. Each transformer layer is composed of a temporal attention (TempAttn), skeletal attention (SkelAttn), cross-attention (CrossAttn), and feedforward network (FFN). Each component involves a residual connection [14], layer normalization (LN) [6], and a feature-wise linear modulation (FiLM) [31] that modulates intermediate features based on a diffusion timestep. We also incorporate skip connection [37] within the stack of transformer layers.", + "bbox": [ + 89, + 412, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input motion latent at diffusion timestep $t$ given to the $l$ -th transformer layer, denoted as $\\mathbf{z}_t^l$ , is processed by three consecutive attention blocks: temporal, skeletal, and cross-attention layers. The purpose of this architecture design is to model frame-wise interactions through TempAttn, joint-wise interactions through SkelAttn, and motion-text interactions through CrossAttn. As a result, $\\mathbf{z}_t^l$ is consecutively updated at each layer as follows:", + "bbox": [ + 89, + 625, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} ^ {l} \\gets \\mathbf {z} _ {t} ^ {l} + \\mathrm {F i L M} (\\mathrm {T e m p A t t n} (\\mathrm {L N} (\\mathbf {z} _ {t} ^ {l}))), \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 753, + 482, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} ^ {l} \\leftarrow \\mathbf {z} _ {t} ^ {l} + \\operatorname {F i L M} (\\operatorname {S k e l A t t n} (\\operatorname {L N} (\\mathbf {z} _ {t} ^ {l}))), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 775, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} ^ {l} \\leftarrow \\mathbf {z} _ {t} ^ {l} + \\mathrm {F i L M} (\\operatorname {C r o s s A t t n} (\\operatorname {L N} (\\mathbf {z} _ {t} ^ {l}), \\operatorname {C L I P} (c))). \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 796, + 482, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the CLIP text encoder [36] used in cross-attention, we use a pre-trained model and freeze its weights during training.", + "bbox": [ + 89, + 824, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Parametrization. For diffusion parametrization, our model predicts the diffusion velocity [38], denoted as", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{v}$ , as follows:", + "bbox": [ + 513, + 268, + 607, + 281 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {t} = \\alpha_ {t} \\boldsymbol {\\epsilon} - \\sigma_ {t} \\mathbf {x}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 284, + 903, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ and $\\mathbf{x}$ represent the noise and sample, respectively, while $\\alpha_{t}$ and $\\sigma_{t}$ are noise scheduling parameters at diffusion timestep $t$ . While $\\epsilon$ - and $\\mathbf{x}$ -prediction are widely used in diffusion models, $\\mathbf{v}$ -prediction combines both parametrizations, implicitly balancing the information contributed by each component. This improves the stability of generation, especially at high noise levels in which the relationship between $\\epsilon$ and $\\mathbf{x}$ becomes disrupted.", + "bbox": [ + 511, + 305, + 903, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training and Inference. The denoiser is trained to predict the diffusion velocity $\\mathbf{v}_t$ at each diffusion timestep $t$ . The training objective for the denoiser is defined as follows:", + "bbox": [ + 511, + 426, + 903, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e n o i s e r}} = \\left\\| \\hat {\\mathbf {v}} _ {t} - \\mathbf {v} _ {t} \\right\\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 481, + 903, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{v}}_t$ represents the predicted diffusion velocity by the denoiser. During training, we randomly drop the text condition with a probability of $p_{\\mathrm{uncond}}$ , enabling unconditional generation when $c$ is a null text, which is represented as $\\varnothing$ .", + "bbox": [ + 511, + 508, + 903, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To the trained denoiser, we apply classifier-free guidance (CFG) [16] for text-conditioned generation:", + "bbox": [ + 511, + 570, + 903, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {v}} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, c\\right) := \\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\varnothing\\right) + w \\left(\\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, c\\right) - \\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\varnothing\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 611, + 906, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{v}_{\\theta}$ denotes the pre-trained denoiser, $w$ is the CFG weight, and $\\hat{\\mathbf{v}}_{\\theta}$ is the modified velocity used during the denoising process. Additionally, we employ the DDIM sampling formulation [40] during inference, which effectively reduces the number of sampling steps without compromising generation quality.", + "bbox": [ + 511, + 642, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Zero-shot Text-driven Motion Editing", + "text_level": 1, + "bbox": [ + 511, + 741, + 839, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe the attention modulation methods that enable zero-shot text-driven motion editing with a pre-trained SALAD model. We introduce four distinct modulation strategies applied to the text-motion cross attention maps as shown in Fig. 3: word swap, prompt refinement, attention re-weighting, and attention mirroring. The first three are adapted from the Prompt-to-Prompt framework [15], while the attention mirroring serves as a validation method to assess whether the cross-attention map", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7161", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0084cfd6109f28f04c9bc385f2fd4b8d3f30a6917842177aa51fc34a5ae1f5cb.jpg", + "image_caption": [ + "Figure 3. Attention modulation methods applied to the cross-attention maps to enable text-driven motion editing." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "contributes to the activation of spatio-temporal dynamics in the motion generation process. Word swap exchanges attention maps between a source and target prompt, enabling the model to leverage the attention context of the source text when generating output for the target text. Prompt refinement enhances the original attention map by integrating additional attention maps for appended word tokens, providing richer semantic information compared to the base prompt. Attention re-weighting adjusts the attention values assigned to specific user-selected words, either amplifying or diminishing their influence on the generated motion. Lastly, attention mirroring swaps attention values between symmetrical body parts, such as the left and right arms, to produce mirrored motions. These methods collectively empower the generative capacity of our model to produce dynamic and contextually relevant motion edits based solely on text prompts. For more details on these modulation functions, please refer to the supplementary material.", + "bbox": [ + 88, + 279, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 566, + 222, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted evaluations of our method on two widely used motion-language benchmarks: HumanML3D [10] and KIT-ML [34]. HumanML3D dataset consists of 14,616 motion sequences with a variety of human actions, such as locomotions, sports, and acrobatics, along with 44,970 text descriptions in total. KIT-ML dataset consists of 3,911 motion sequences along with 6,278 text descriptions. Both datasets are augmented by mirroring, and divided into training, testing, and validation sets with the ratio of 0.8, 0.15, and 0.05. Evaluation results are reported in the following sections. Please refer to the supplementary material for additional quantitative experiment results with details, and the supplementary video for the animated results.", + "bbox": [ + 89, + 593, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 801, + 307, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method was executed on a single NVIDIA V100 GPU. We used the AdamW optimizer [26], and the VAE and denoiser were trained for 50 and 500 epochs, respectively. The VAE and denoiser were trained with a batch size of 64 for HumanML3D, which required 35 and 17 hours, re", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "spectively. For the KIT-ML dataset, we used a batch size of 16, with training times of 20 hours for the VAE and 7 hours for the denoiser. We trained the denoiser with 1000 diffusion steps, employing 50 steps for DDIM sampling during inference. For the CFG weight, we set $w = 7.5$ unless mentioned otherwise.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 191, + 733, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For quantitative evaluations, we adopted metrics from T2M [10]. First, $FID$ is used to evaluate the quality of generated motions by measuring the difference between the feature distributions of generated motions and those of real motions. Additionally, $R$ -precision and $MM-Dist$ are used to evaluate the semantic alignment between the input text and the generated motions. $Diversity$ , which indicates the variance in generated motions, and $MultiModality$ , which assesses the diversity of motions generated from the same text description, are also adopted as secondary metrics. They are considered less important than the generation quality and alignment with the input text.", + "bbox": [ + 511, + 215, + 905, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the practice of T2M [10], each experiment was repeated 20 times on the test sets of HumanML3D and KIT-ML datasets, and the mean with a $95\\%$ confidence interval is reported in Tab. 1. Notably, SALAD consistently achieved superior text-motion alignment across both datasets, regardless of the diffusion parametrization applied during denoiser training. Moreover, SALAD maintained high generation quality, as indicated by the best or second-best FID scores among diffusion-based methods. These results confirm that explicitly modeling interactions between skeletal joints, temporal frames, and textual words is effective in improving text-to-motion generation quality and fidelity to the input text.", + "bbox": [ + 511, + 398, + 905, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 604, + 723, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compared the qualitative results of SALAD with those generated by MDM [42] and MoMask [12], and the comparison results are shown in Fig. 4. Both MDM and MoMask reflected parts of the input text prompts, but they failed to handle all the descriptions in the text. For example, as shown in the top of Fig. 4-(a), MDM generated a character walking forward, turning, and walking back, but the locomotion style was not marching, despite the input prompt specifying \"marches\". Similarly, in Fig. 4-(b), MoMask accurately reflected the arm movements of raising above the head and throwing them down, but it missed the instruction to turn to the right. In case of Fig. 4-(c), both methods failed to perfectly reflect the input text despite its short length. In contrast, SALAD consistently incorporated all the textual descriptions into the generated motions, regardless of the length and complexity of the text prompts. This demonstrates its effectiveness in generating motions that are highly aligned with the input text.", + "bbox": [ + 511, + 628, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7162", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e8f97a58fad51e2e4edabed5feb9a8e248b1343b72a047477a7e5118f1f5d7eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
T2M [10]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
AttT2M [53]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo [54]0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
MoMask [12]0.521±.0020.713±.0020.807±.0020.045±.0022.958±.008-1.241±.040
MDM [42]0.320±.0050.498±.0040.611±.0070.544±.0045.566±.0279.559±.0862.799±.072
MLD [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse [51]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse [50]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
SALAD (Ours)0.581±.0030.769±.0030.857±.0020.076±.0022.649±.0099.696±.0961.751±.062
", + "bbox": [ + 174, + 88, + 823, + 253 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8b821bfd3b3f60d21ba8a056b61ff788cd17fc7858f8adca181349c8f2ca1b23.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
T2M [10]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [53]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
ParCo [54]0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
MoMask [12]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
MDM [42]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.847±.1091.907±.2T4
MLD [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse [51]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse [50]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
SALAD (Ours)0.477±.0060.711±.0050.828±.0050.296±.0122.585±.01611.097±.0951.004±.040
", + "bbox": [ + 173, + 256, + 821, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation results on the test sets of HumanML3D (top) and KIT-ML (bottom). $\\uparrow$ and $\\downarrow$ denote that higher and lower values are better, respectively, while $\\rightarrow$ denotes that the values closer to the real motion are better. Methods above the dotted line are auto-regressive models based on VAE or VQ-VAE, while those below are diffusion-based generative models. Red and blue colors indicate the best and the second best results, respectively.", + "bbox": [ + 89, + 425, + 906, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5cada3a7308061c93108297ac11eef69b63565e13fd4d8ff16ac892e30e1af71.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of results generated by various methods, including MDM [42], MoMask [12], and our approach." + ], + "image_footnote": [], + "bbox": [ + 109, + 484, + 883, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 707, + 243, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Skeleton-aware Latent Space. To prove the effectiveness of using skeleton-aware latent space, we compared the performance of different VAE structures employed in MoMask [12] and ParCo [54] with ours. Specifically, MoMask employs a VQ-VAE that compresses each pose into a single vector, while ParCo employs a part-aware VQ-VAE that processes different body parts independently. Because the role of VAE is to compress the motion features into latent variables, we compared FID and mean per joint position error (MPJPE) that assess the quality and accuracy of the reconstructed motion features, respectively.", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, our skeleton-aware VAE outperformed baselines with a significant margin for both metrics. Notably, our method required much fewer number of trainable parameters than the others. This indicates that skeleton-aware latent space that decouples both dimensions and facilitates information exchange between adjacent components is effective in constructing meaningful motion latent space, which is a crucial prior to train the generator.", + "bbox": [ + 511, + 708, + 906, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we conducted an ablation study that validates the effectiveness of the skeleto-temporal latent (ST-Latent) and cross-attention, which are two core components of SALAD, on the text-to-motion generation perfor", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0c81f77e5ae7716502a15a437e21ad12f3af624d248d41c2dd95d88070f515ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsNum of ParamsFID ↓MPJPE ↓
MoMask [12]19.44M0.020±.0000.030±.000
ParCo [54]6.35M0.021±.0000.108±.000
SALAD (Ours)0.16M0.003±.0000.016±.000
", + "bbox": [ + 94, + 88, + 478, + 157 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc13b2277dc3ac609c64a0161b58416dcde0bed70d532194600c095e3fcb67b5.jpg", + "table_caption": [ + "Table 2. Quantitative results on the quality and accuracy of reconstructed motion features of VAE models from different methods, along with the number of trainable parameters, measured on the test set of HumanML3D." + ], + "table_footnote": [], + "table_body": "
MethodR-Precision (Top-3) ↑FID ↓
Ours (Full model)0.857±.0020.076±.002
w/o ST-Latent0.816±.0020.433±.006
w/o CrossAttn0.778±.0020.274±.007
w/o both0.752±.0020.345±.007
", + "bbox": [ + 94, + 224, + 478, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on the VAE and denoiser.", + "bbox": [ + 133, + 316, + 439, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance. For this experiment, we evaluated three variations: (i) replacing ST-Latent by training a conventional VAE that treats each pose as a single vector, (ii) replacing cross-attention with self-attention by concatenating the sentence-level CLIP feature to the motion features, and (iii) combining (i) and (ii). As shown in Tab. 3, alteration of ST-Latent and cross-attention substantially degraded generation quality and text-motion alignment, as reflected in the higher FID and lower R-precision, respectively. This demonstrates the importance of rich information exchange between text and motion by cross-attention within the skeletotemporally structured latent space. This indicates that the skeletotemporally disentangled motion latent facilitates accurate text-motion interaction.", + "bbox": [ + 88, + 338, + 482, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CFG Weights. To analyze the influence of CFG weights on generation results, we plotted FID and R-precision metric values in Fig. 5. The figure shows that low CFG weights led to suboptimal generation quality and text-motion alignment. The performance of SALAD improved across both metrics as the weight values increased, but excessively high weight values resulted in a decreased performance for both metrics again. Based on these findings, we determined that a value of $w = 7.5$ produces the best balance between generation quality and alignment with the text prompts.", + "bbox": [ + 88, + 549, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Zero-shot Text-driven Motion Editing", + "text_level": 1, + "bbox": [ + 89, + 710, + 416, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the capability of SALAD in capturing relationships between skeleto-temporal latents and text prompts, we visualized the cross-attention maps for each word in Fig. 6. The figure reveals that high attention values were assigned to proper frames and body parts that semantically correspond to each word. For example, \"jumps forward\" was activated for the entire body at early frames, \"walks forward\" focused on the root and legs at later frames, and \"waving arms\" was activated for the arms throughout the entire frames. These results indicate that cross-attention maps produced during motion generation ef", + "bbox": [ + 88, + 734, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ab25601c73dcc4dc025aa37c182e9e71d310c4af33b8628f4f58a7f3aab3708.jpg", + "image_caption": [ + "Figure 5. Ablation study results on different CFG weights. Lower FID scores indicate better motion quality, and higher R-precision (Top-3) scores indicate better semantic alignment." + ], + "image_footnote": [], + "bbox": [ + 555, + 75, + 861, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a1e1bcdad2dcb9062350b0e8d40f722193758e7e4cb6a444d1cca8e32ad492c.jpg", + "image_caption": [ + "Figure 6. Visualizations of cross-attention maps of SALAD. Each row corresponds to a specific body part, and each column represents temporal frames. Attention maps were computed by averaging the output attention maps across all transformer layers and all diffusion timesteps." + ], + "image_footnote": [ + "a man jumps forward and then walks forward, while waving his arms" + ], + "bbox": [ + 516, + 303, + 901, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fectively capture the rich semantic relationships between text and motion.", + "bbox": [ + 511, + 611, + 903, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on these observations, we demonstrate zero-shot text-driven motion editing results using a pre-trained SALAD model by modulating cross-attention maps, and the results are shown in Fig. 7. The edited motions align well with the given instructions while preserving the original movements, such as walking phases. Notably, the changes for editing were localized to specific frames and body parts relevant to the editing prompt, while unaffected regions remain intact, without using explicit masking. Furthermore, our SALAD model understands not only the high-level correspondence between text and motion but also its intensity and direction, which is demonstrated by different hand heights when modulating the word \"high\" in the reweighting case. These results suggest that the explicit modeling of skeletal-temporal-textual relationships in SALAD enhances its understanding of text-driven motion generation, leading to effective learning representations and fine", + "bbox": [ + 511, + 643, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c1f4109ad5038b2d13b111cf439c198e1bb50d536b1eac80dc9b346736e39e6.jpg", + "image_caption": [ + "Figure 7. Text-driven motion editing results via attention modulation using a pre-trained SALAD model. Texts colored in green represent the editing instructions." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 500, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98a10733bf4aaf03205648c52782d2dd771e92d8cb512ea2136c984bdfc48095.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 88, + 903, + 352 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6706a7d42f53903114f64560bb4f9a89d54360eb7421fafe461a221ae6133bf2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPreservationSemantic AlignmentOverall Quality
MDM [42]3.729±.1532.758±.1923.196±.147
MotionFix [5]3.358±.1733.388±.1893.421±.166
SALAD (Ours)4.596±.0844.654±.0874.596±.083
", + "bbox": [ + 94, + 387, + 480, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. User study results. The red color indicates the best result.", + "bbox": [ + 89, + 453, + 480, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "grained control in both generation and editing.", + "bbox": [ + 89, + 474, + 398, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the perceptual quality of text-driven motion editing, we conducted a user study with 16 participants. We compared SALAD against MDM [42], which edits motion by masking and regenerating specified body parts, and MotionFix [5], a method specifically designed for motion editing. Participants were shown with 15 videos per method and asked to rate them based on three criteria: (i) preservation of the original motion, (ii) semantic alignment with the target text, and (iii) overall quality. Ratings were given on a 5-point Likert scale, with 5 indicating the best.", + "bbox": [ + 88, + 489, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 4, SALAD outperformed both baselines across all three aspects by a large margin. Notably, despite not using explicit masking, SALAD achieved higher preservation scores. Furthermore, while SALAD is not specifically designed for editing, it achieved higher preferences for semantic alignment and overall quality, highlighting the effectiveness of attention-based editing. These results demonstrate that leveraging cross-attention maps enables high-quality and perceptually appealing text-driven motion editing without requiring additional training.", + "bbox": [ + 88, + 641, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 799, + 204, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While SALAD produced outperforming results in text-driven motion generation compared to the existing generative models, along with the capability of zero-shot text-driven motion editing, it has several limitations. First, the numerical results for Diversity and MultiModality are lim", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ited, despite the strong capability of diffusion models to generate diverse outputs. This can be due to the trade-off between high-quality with faithfulness to the text and diversity. While we put the diversity as a second priority, finding a way to handle both could be an interesting approach. Additionally, SALAD is restricted to generating a single person action with limited text and motion length. Therefore, enabling multi-character interactions, crowd animations, generation from longer texts, and creating extended motion sequences could be an interesting direction for future work. Furthermore, although our method does not directly address editing real motions, integrating diffusion inversion methods could enable text-driven editing of real motion sequences.", + "bbox": [ + 511, + 390, + 906, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented SALAD, a novel diffusion model framework for text-driven motion generation that models the complex interactions between skeletal joints, temporal frames, and textual words. We first train a skeletotemporal VAE that effectively summarizes motion features by decoupling skeletal and temporal dimensions, and we subsequently train a skeletotemporal denoiser that enables efficient and expressive generation of motions conditioned on text input within this compact skeletotemporal latent space. Furthermore, we demonstrated the first text-driven zero-shot motion editing using a pre-trained SALAD model by modulating cross-attention maps during generation. By leveraging the outperforming generative capacity compared to previous methods, along with the interpretable learning representations during generation, we anticipate that our work can benefit future research and applications in the context of text-driven motion generation and editing.", + "bbox": [ + 509, + 643, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We appreciate the anonymous reviewers for their invaluable discussion. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00333478).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2, 3", + "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2", + "[3] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1, 2", + "[4] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2", + "[5] Nikos Athanasiou, Alpár Cseke, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 8", + "[6] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2", + "[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2, 6", + "[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2", + "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 5, 6", + "[11] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + ], + "bbox": [ + 93, + 215, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 5, 6, 7", + "[13] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 1, 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 4", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[18] Yiming Huang, Weilin Wan, Yue Yang, Chris Callison-Burch, Mark Yatskar, and Lingjie Liu. Como: Controllable motion generation through language guided pose code editing. In European Conference on Computer Vision, pages 180-196. Springer, 2024. 2", + "[19] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2", + "[20] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2", + "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 1, 2", + "[22] Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwajanakorn, and Siyu Tang. Optimizing diffusion noise can serve as universal motion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1345, 2024. 2", + "[23] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 1, 2", + "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1", + "[25] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7166", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5", + "[27] Yunhong Lou, Linchao Zhu, Yaxiong Wang, Xiaohan Wang, and Yi Yang. Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372, 2023. 2", + "[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[29] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2", + "[30] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on computer graphics and interactive techniques, 4(3):1-17, 2021. 2", + "[31] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 4", + "[32] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2", + "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2", + "[34] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 5", + "[35] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1, 2", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4", + "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4", + "[41] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2", + "[42] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 8", + "[43] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2", + "[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2", + "[45] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 1", + "[46] Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22035-22044, 2023. 2", + "[47] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 1, 2", + "[48] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3", + "[49] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2", + "[50] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2, 6", + "[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7167", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7368-7376, 2024. 2", + "[53] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2, 6", + "[54] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2, 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7168", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_model.json b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd68192495d8f42fcb6c2ddee14fee1f36555ace --- /dev/null +++ b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.131, + 0.74, + 0.177 + ], + "angle": 0, + "content": "SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.204, + 0.877, + 0.239 + ], + "angle": 0, + "content": "Seokhyeon Hong Chaelin Kim Serin Yoon Junghyun Nam Sihun Cha Junyong Noh Visual Media Lab, KAIST" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.242, + 0.771, + 0.258 + ], + "angle": 0, + "content": "{ghd3079, chaelin.kim, serinyoon, ys4990, chacorp, junyongnoh}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.486, + 0.643 + ], + "angle": 0, + "content": "Text-driven motion generation has advanced significantly with the rise of denoising diffusion models. However, previous methods often oversimplify representations for the skeletal joints, temporal frames, and textual words, limiting their ability to fully capture the information within each modality and their interactions. Moreover, when using pretrained models for downstream tasks, such as editing, they typically require additional efforts, including manual interventions, optimization, or fine-tuning. In this paper, we introduce a skeleton-aware latent diffusion (SALAD), a model that explicitly captures the intricate inter-relationships between joints, frames, and words. Furthermore, by leveraging cross-attention maps produced during the generation process, we enable attention-based zero-shot text-driven motion editing using a pre-trained SALAD model, requiring no additional user input beyond text prompts. Our approach significantly outperforms previous methods in terms of text-motion alignment without compromising generation quality, and demonstrates practical versatility by providing diverse editing capabilities beyond generation. Code is available at project page." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.661, + 0.222, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Character animation is a crucial component in various computer graphics and vision applications, including games, films, and interactive media. Traditional methods, such as keyframing and motion capture, typically require extensive manual effort, which are time-consuming and expensive. Recently, deep generative models have been introduced to mitigate these challenges. In particular, diffusion models have shown promising results in text-to-motion generation, enabling intuitive and efficient animation workflows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Despite these advancements, fully capturing the intricate relationships among frames, body parts, and textual descriptions remains a complex challenge in text-driven motion generation. Previous approaches typically represent a pose as a single vector and primarily focus on temporal re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.429 + ], + "angle": 0, + "content": "relationships between poses across frames, while neglecting spatial interactions between skeletal joints. Furthermore, these models often compress a sentence into a single vector for conditioning, which can overlook the important nuances of word-level variations. This over-simplification of interactions among skeletal, temporal, and textual components may lead to missing details in the generated results, underscoring the need for models that faithfully capture these complex dependencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.688 + ], + "angle": 0, + "content": "On the other hand, learning meaningful representations is a key factor to enable zero-shot downstream tasks with pre-trained models. For example, zero-shot text-driven image and video editing methods [3, 7, 13, 15, 25, 35, 43] leverage attention modulation from pre-trained models for intuitive manipulation. Unfortunately, motion generation models typically lack such interpretable and manipulatable intermediate representations. This limitation stems from the over-simplification of motion and text features mentioned above, which restricts rich interactions between them. Although some methods exploit pre-trained models with manual masks, optimization, or fine-tuning [21-23, 42, 45, 47], these methods require additional effort, time, and cost to achieve desired results. Therefore, developing interpretable representations during motion synthesis could improve versatility and flexibility for downstream tasks beyond generation, such as zero-shot motion editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a skeleton-aware latent diffusion, which we call SALAD, for text-driven motion generation within a skeletotemporally structured latent space. We first train a variational autoencoder (VAE) [24] to construct a motion latent space that decouples spatial and temporal dimensions. To this end, we employ skeletotemporal convolution layers to facilitate information exchange between adjacent joints and frames. We also employ skeletotemporal pooling layers to produce a compact representation by reducing the dimensionality, which can reduce the computational complexity during the sampling process of the diffusion model. Subsequently, we train a diffusion model within this latent space to generate motion features conditioned on text. To train the denoiser, we leverage both" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7158" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "skeletal and temporal attention blocks, which can effectively enable skeletotemporal coherence. Additionally, we employ cross-attention to capture fine-grained interactions between individual words and the skeletotemporal units of the motion latents." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.482, + 0.349 + ], + "angle": 0, + "content": "Beyond motion generation, we introduce an attention-based zero-shot, text-driven motion editing method that leverages the pre-trained SALAD model, requiring no additional optimization or fine-tuning. We demonstrate that the intermediate cross-attention maps of SALAD capture the relationship between text and motion features. By modulating these cross-attention maps, we enable text-driven motion editing that allows users to edit generated motion through text input alone in a training-free manner. We also propose a novel attention modulation method for motion-specific tasks, demonstrating the practical potential of our approach." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.351, + 0.411, + 0.365 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.367, + 0.483, + 0.412 + ], + "angle": 0, + "content": "- We propose SALAD, a novel skeleton-aware latent diffusion model for text-driven motion generation within a skeletotemporally structured latent space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.412, + 0.483, + 0.457 + ], + "angle": 0, + "content": "- We interpret the intermediate representations in the generation process, allowing for a clear understanding of the relationship between text inputs and generated motions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.457, + 0.484, + 0.502 + ], + "angle": 0, + "content": "- We present an attention-based zero-shot text-driven motion editing method using cross-attention modulation in generative models." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.367, + 0.484, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.519, + 0.233, + 0.535 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.367, + 0.56 + ], + "angle": 0, + "content": "2.1. Text-driven Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With the rise of deep neural networks, generative models for human motion synthesis based on natural language descriptions have been extensively explored. Early studies focused on generating motions from action labels, such as throw and kick [9, 32]. More advanced methods projected both text and motion features into a shared latent space, enabling the generation of fixed motion sequences from identical input text [2]. To increase the diversity of generated results, T2M [10] employed a temporal VAE that can generate varied motion sequences that align with the given text input. Subsequent studies leveraged large pretrained text encoders, such as CLIP [36], to improve textual understanding for motion generation [4, 33, 41]. Autoregressive models that behave like language models have also been proposed, which first encode motion features into discrete tokens through vector quantization [44] and then auto-regressively predict the next motion token given the text prompt [11, 12, 20, 49, 52-54]. Recently, denoising diffusion models [17, 39] have become the dominant framework for generative models, prompting text-driven motion generation methods to adopt diffusion models as their backbone [8, 23, 27, 42, 46, 50, 51]. In this study, we build" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "our framework on the diffusion model formulation and enhance the performance of text-driven motion generation through explicit modeling of the relationships among joints, frames, and words. Moreover, we introduce an attention-based zero-shot text-driven editing method, leveraging our pre-trained SALAD model as a generator, demonstrating a novel exploration beyond mere generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.209, + 0.875, + 0.224 + ], + "angle": 0, + "content": "2.2. Skeleton-aware Processing of Motion Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Because motion data consists of two separate skeletal and temporal dimensions, some methods have been developed to respect each dimension when processing the motion data. Yan et al. [48] presented spatial-temporal graph convolution networks that can learn both the spatial and temporal patterns of dynamic skeletons. Skeleton-aware networks [1] employed skeleto-temporal convolution networks and topology-preserving skeletal pooling layers to construct a shared motion latent space, enabling deep motion retargeting across homeomorphic skeletons. Motion style transfer has also been enabled in a part-wise manner by employing skeleto-temporal convolution modules [19, 30]. In the realm of text-to-motion generation, ParCo [54] and AttT2M [53] utilized the skeleto-temporal network architecture during the motion quantization process. In this work, we adopt a similar approach by encoding motion features using skeleto-temporal convolution networks and pooling layers to derive an effective skeleton-aware motion latent space with compact dimensionality. Additionally, we facilitate the interaction between this skeleton-aware motion latent space and textual features during the text-conditional generation process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.861, + 0.591 + ], + "angle": 0, + "content": "2.3. Zero-shot Editing with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "By leveraging the generative capabilities of diffusion models, zero-shot editing methods have been proposed. SDEdit [28] first adds noise to the input data for a few diffusion timesteps, then denoises the noisy data with an editing condition to achieve high-fidelity zero-shot editing by leveraging a pre-trained diffusion model. Prompt-to-Prompt [15] verified that cross-attention maps establish connections between word tokens and the spatial layout of the generated image. In addition, it introduced attention modulation methods that can be utilized during the generation process, enabling the generation of high-fidelity editing results in a zero-shot manner. Null-text inversion [29] extended this method to real-images by improving the diffusion inversion method. Similarly, image and video editing methods via attention modulation have been extensively studied [3, 7, 13, 25, 35, 43]. Motion diffusion models often employ joint masks [23, 42] or optimization [21, 22, 47] to enable motion editing using pre-trained diffusion models, although they require user intervention or additional computation for the desired goal. On the other hand, CoMo [18]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7159" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.907, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.907, + 0.25 + ], + "angle": 0, + "content": "Figure 1. Architecture of the skeletotemporal VAE network. The encoder maps motion features into a skeletotemporal latent space, and the decoder restores the skeletotemporal latent variables into motion features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.257, + 0.485, + 0.395 + ], + "angle": 0, + "content": "achieved zero-shot text-driven motion editing using large language models to manipulate semantic pose tokens. We demonstrate that the intermediate cross-attention maps of our SALAD model can capture the relationship between motion and text, similarly to image diffusion models. Furthermore, by modulating these cross-attention maps, we introduce an attention-based zero-shot text-driven motion editing method that eliminates the need for manual masks, fine-tuning, or optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.406, + 0.182, + 0.421 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.432, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Our goal is to generate a motion sequence \\(\\mathbf{m}_{1:N}\\) conditioned on a text prompt \\(c\\), where \\(N\\) denotes the number of frames. We first construct a skeleton-aware motion latent representation that captures both skeletal and temporal dynamics of a motion sequence (Sec. 3.1). Using this motion representation, we train a diffusion model by modeling the complex interactions between skeletal joints, temporal frames, and textual descriptions (Sec. 3.2). Finally, we extend our model for zero-shot text-driven motion editing by modulating cross-attention maps between text and motion (Sec. 3.3). Additional details on the network architectures will be featured in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.286, + 0.636 + ], + "angle": 0, + "content": "3.1. Skeleton-aware VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To derive skeleton-aware motion features, we train a VAE composed of skeletotemporal convolution networks and pooling layers [1, 48], as illustrated in Fig. 1. The core module is the skeletotemporal convolution (STConv), which decouples joint and frame dimensions while facilitating information exchange between adjacent components within each dimension. While the skeletotemporal separation of motion sequences allows for sophisticated modeling of motion data, it introduces an additional dimension for skeletal joints, unlike previous text-to-motion generation methods that treat a pose as a single vector. This expanded data space increases the risk of curse of dimensionality and leads to longer computation times for diffusion models, which rely on solving stochastic differential equations within this higher-dimensional data space. To address this issue, we apply skeletotemporal pooling (STPool) layers in the encoder to reduce the latent space dimensionality, effectively" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.907, + 0.302 + ], + "angle": 0, + "content": "compressing the representation. In the decoder, skeletotemporal unpooling (STUnpool) layers are used to reconstruct the compressed latent features back to raw motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.906, + 0.408 + ], + "angle": 0, + "content": "Encoder. We separate each pose vector of \\(\\mathbf{m}_{1:N}\\) into jointwise features and propagate them through joint-wise multilayer perceptron (MLP) layers, resulting in \\(\\mathbf{h} \\in \\mathbb{R}^{N \\times J \\times D}\\), where \\(J\\) denotes the number of joints and \\(D\\) denotes the latent dimension. Subsequently, we apply STConv to \\(\\mathbf{h}\\) to facilitate information exchange between adjacent frames and joints:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.414, + 0.907, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S T C o n v} (\\mathbf {h}) := \\operatorname {S k e l C o n v} (\\mathbf {h}) + \\operatorname {T e m p C o n v} (\\mathbf {h}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.48 + ], + "angle": 0, + "content": "where SkelConv is a graph convolution network over the joint dimension, while TempConv is a 1D convolution network over the frame dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.526 + ], + "angle": 0, + "content": "We apply the STPool operator to the output of the STConv layers to reduce its dimensionality by pooling over joints and frames:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.532, + 0.905, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S T P o o l} (\\mathbf {h}) := \\operatorname {T e m p P o o l} (\\operatorname {S k e l P o o l} (\\mathbf {h})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.674 + ], + "angle": 0, + "content": "where SkelPool represents the pooling layer over the joint dimension while TempPool represents the pooling layer over the temporal dimension. Specifically, SkelPool aggregates adjacent joints while preserving its skeletal topology, resulting in fewer joints with a homeomorphic structure. In contrast, TempPool performs 1D pooling over frames. These pooling operations are commutative because they operate within independent dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Consequently, we obtain the skeleto-temporally compressed latent vector \\(\\mathbf{z} \\in \\mathbb{R}^{N' \\times J' \\times D}\\), where \\(N' < N\\) and \\(J' < J\\). This representation effectively encodes the skeleton-aware dynamics of the motion sequence within a lower dimension. Notably, we maintain 7 atomic joints for this motion latent: root, spine, head, both arms, and both legs \\((J' = 7)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Decoder. To reconstruct \\(\\mathbf{z}\\) into raw motion features, we employ STConv and STUnpool layers in the decoder. The decoder mirrors the encoder, facilitating information exchange between adjacent joints and frames using STConv layers and increasing the skeleto-temporal resolution using STUnpool layers. Finally, the joint-wise features are processed by their respective joint-wise MLP layers, resulting in the reconstructed motion feature \\(\\hat{\\mathbf{m}}_{1:N}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7160" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.086, + 0.445, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.087, + 0.898, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.232, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Figure 2. (Left) Overall network architecture of the denoiser. (Right) The architecture of each transformer block consisting of TempAttn, SkelAttn, CrossAttn, and FFN, along with the FiLM following each module." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.268, + 0.483, + 0.284 + ], + "angle": 0, + "content": "Training. The VAE is trained with the following objectives:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.295, + 0.483, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V A E}} = \\mathcal {L} _ {\\mathbf {m}} + \\lambda_ {\\text {p o s}} \\mathcal {L} _ {\\text {p o s}} + \\lambda_ {\\text {v e l}} \\mathcal {L} _ {\\text {v e l}} + \\lambda_ {\\mathrm {k l}} \\mathcal {L} _ {\\mathrm {k l}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.381 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathbf{m}},\\mathcal{L}_{\\mathrm{pos}},\\mathcal{L}_{\\mathrm{vel}}\\) are L1 reconstruction loss terms of the motion features, joint positions, and joint velocities, while \\(\\mathcal{L}_{\\mathrm{kl}}\\) is the Kullback-Leibler (KL) divergence regularization that encourages a structured latent space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.318, + 0.405 + ], + "angle": 0, + "content": "3.2. Skeleton-aware Denoiser" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Network Architecture. Given the skeleton-aware motion latent vector \\(\\mathbf{z} \\in \\mathbb{R}^{N' \\times J' \\times D}\\), we train a diffusion model that denoises the noised latent vector conditioned on text prompt \\(c\\). As shown in Fig. 2, our denoiser consists of a positional embedding, MLP-based encoder and decoder, and \\(L\\) transformer-based layers. Each transformer layer is composed of a temporal attention (TempAttn), skeletal attention (SkelAttn), cross-attention (CrossAttn), and feedforward network (FFN). Each component involves a residual connection [14], layer normalization (LN) [6], and a feature-wise linear modulation (FiLM) [31] that modulates intermediate features based on a diffusion timestep. We also incorporate skip connection [37] within the stack of transformer layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.746 + ], + "angle": 0, + "content": "The input motion latent at diffusion timestep \\(t\\) given to the \\(l\\)-th transformer layer, denoted as \\(\\mathbf{z}_t^l\\), is processed by three consecutive attention blocks: temporal, skeletal, and cross-attention layers. The purpose of this architecture design is to model frame-wise interactions through TempAttn, joint-wise interactions through SkelAttn, and motion-text interactions through CrossAttn. As a result, \\(\\mathbf{z}_t^l\\) is consecutively updated at each layer as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.755, + 0.483, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} ^ {l} \\gets \\mathbf {z} _ {t} ^ {l} + \\mathrm {F i L M} (\\mathrm {T e m p A t t n} (\\mathrm {L N} (\\mathbf {z} _ {t} ^ {l}))), \\qquad (4)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.776, + 0.483, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} ^ {l} \\leftarrow \\mathbf {z} _ {t} ^ {l} + \\operatorname {F i L M} (\\operatorname {S k e l A t t n} (\\operatorname {L N} (\\mathbf {z} _ {t} ^ {l}))), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.797, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} ^ {l} \\leftarrow \\mathbf {z} _ {t} ^ {l} + \\mathrm {F i L M} (\\operatorname {C r o s s A t t n} (\\operatorname {L N} (\\mathbf {z} _ {t} ^ {l}), \\operatorname {C L I P} (c))). \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.87 + ], + "angle": 0, + "content": "For the CLIP text encoder [36] used in cross-attention, we use a pre-trained model and freeze its weights during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Diffusion Parametrization. For diffusion parametrization, our model predicts the diffusion velocity [38], denoted as" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.269, + 0.608, + 0.282 + ], + "angle": 0, + "content": "\\(\\mathbf{v}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.285, + 0.905, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {t} = \\alpha_ {t} \\boldsymbol {\\epsilon} - \\sigma_ {t} \\mathbf {x}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.426 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) and \\(\\mathbf{x}\\) represent the noise and sample, respectively, while \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are noise scheduling parameters at diffusion timestep \\(t\\). While \\(\\epsilon\\)- and \\(\\mathbf{x}\\)-prediction are widely used in diffusion models, \\(\\mathbf{v}\\)-prediction combines both parametrizations, implicitly balancing the information contributed by each component. This improves the stability of generation, especially at high noise levels in which the relationship between \\(\\epsilon\\) and \\(\\mathbf{x}\\) becomes disrupted." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.905, + 0.471 + ], + "angle": 0, + "content": "Training and Inference. The denoiser is trained to predict the diffusion velocity \\(\\mathbf{v}_t\\) at each diffusion timestep \\(t\\). The training objective for the denoiser is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.482, + 0.905, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e n o i s e r}} = \\left\\| \\hat {\\mathbf {v}} _ {t} - \\mathbf {v} _ {t} \\right\\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.905, + 0.57 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{v}}_t\\) represents the predicted diffusion velocity by the denoiser. During training, we randomly drop the text condition with a probability of \\(p_{\\mathrm{uncond}}\\), enabling unconditional generation when \\(c\\) is a null text, which is represented as \\(\\varnothing\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.905, + 0.601 + ], + "angle": 0, + "content": "To the trained denoiser, we apply classifier-free guidance (CFG) [16] for text-conditioned generation:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.612, + 0.908, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {v}} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, c\\right) := \\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\varnothing\\right) + w \\left(\\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, c\\right) - \\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\varnothing\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{v}_{\\theta}\\) denotes the pre-trained denoiser, \\(w\\) is the CFG weight, and \\(\\hat{\\mathbf{v}}_{\\theta}\\) is the modified velocity used during the denoising process. Additionally, we employ the DDIM sampling formulation [40] during inference, which effectively reduces the number of sampling steps without compromising generation quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.84, + 0.759 + ], + "angle": 0, + "content": "3.3. Zero-shot Text-driven Motion Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this section, we describe the attention modulation methods that enable zero-shot text-driven motion editing with a pre-trained SALAD model. We introduce four distinct modulation strategies applied to the text-motion cross attention maps as shown in Fig. 3: word swap, prompt refinement, attention re-weighting, and attention mirroring. The first three are adapted from the Prompt-to-Prompt framework [15], while the attention mirroring serves as a validation method to assess whether the cross-attention map" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7161" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Attention modulation methods applied to the cross-attention maps to enable text-driven motion editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.28, + 0.483, + 0.552 + ], + "angle": 0, + "content": "contributes to the activation of spatio-temporal dynamics in the motion generation process. Word swap exchanges attention maps between a source and target prompt, enabling the model to leverage the attention context of the source text when generating output for the target text. Prompt refinement enhances the original attention map by integrating additional attention maps for appended word tokens, providing richer semantic information compared to the base prompt. Attention re-weighting adjusts the attention values assigned to specific user-selected words, either amplifying or diminishing their influence on the generated motion. Lastly, attention mirroring swaps attention values between symmetrical body parts, such as the left and right arms, to produce mirrored motions. These methods collectively empower the generative capacity of our model to produce dynamic and contextually relevant motion edits based solely on text prompts. For more details on these modulation functions, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.223, + 0.584 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.79 + ], + "angle": 0, + "content": "We conducted evaluations of our method on two widely used motion-language benchmarks: HumanML3D [10] and KIT-ML [34]. HumanML3D dataset consists of 14,616 motion sequences with a variety of human actions, such as locomotions, sports, and acrobatics, along with 44,970 text descriptions in total. KIT-ML dataset consists of 3,911 motion sequences along with 6,278 text descriptions. Both datasets are augmented by mirroring, and divided into training, testing, and validation sets with the ratio of 0.8, 0.15, and 0.05. Evaluation results are reported in the following sections. Please refer to the supplementary material for additional quantitative experiment results with details, and the supplementary video for the animated results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.308, + 0.817 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our method was executed on a single NVIDIA V100 GPU. We used the AdamW optimizer [26], and the VAE and denoiser were trained for 50 and 500 epochs, respectively. The VAE and denoiser were trained with a batch size of 64 for HumanML3D, which required 35 and 17 hours, re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.182 + ], + "angle": 0, + "content": "spectively. For the KIT-ML dataset, we used a batch size of 16, with training times of 20 hours for the VAE and 7 hours for the denoiser. We trained the denoiser with 1000 diffusion steps, employing 50 steps for DDIM sampling during inference. For the CFG weight, we set \\( w = 7.5 \\) unless mentioned otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.193, + 0.735, + 0.209 + ], + "angle": 0, + "content": "4.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.906, + 0.398 + ], + "angle": 0, + "content": "For quantitative evaluations, we adopted metrics from T2M [10]. First, \\( FID \\) is used to evaluate the quality of generated motions by measuring the difference between the feature distributions of generated motions and those of real motions. Additionally, \\( R \\)-precision and \\( MM-Dist \\) are used to evaluate the semantic alignment between the input text and the generated motions. \\( Diversity \\), which indicates the variance in generated motions, and \\( MultiModality \\), which assesses the diversity of motions generated from the same text description, are also adopted as secondary metrics. They are considered less important than the generation quality and alignment with the input text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Following the practice of T2M [10], each experiment was repeated 20 times on the test sets of HumanML3D and KIT-ML datasets, and the mean with a \\(95\\%\\) confidence interval is reported in Tab. 1. Notably, SALAD consistently achieved superior text-motion alignment across both datasets, regardless of the diffusion parametrization applied during denoiser training. Moreover, SALAD maintained high generation quality, as indicated by the best or second-best FID scores among diffusion-based methods. These results confirm that explicitly modeling interactions between skeletal joints, temporal frames, and textual words is effective in improving text-to-motion generation quality and fidelity to the input text." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.724, + 0.621 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We compared the qualitative results of SALAD with those generated by MDM [42] and MoMask [12], and the comparison results are shown in Fig. 4. Both MDM and MoMask reflected parts of the input text prompts, but they failed to handle all the descriptions in the text. For example, as shown in the top of Fig. 4-(a), MDM generated a character walking forward, turning, and walking back, but the locomotion style was not marching, despite the input prompt specifying \"marches\". Similarly, in Fig. 4-(b), MoMask accurately reflected the arm movements of raising above the head and throwing them down, but it missed the instruction to turn to the right. In case of Fig. 4-(c), both methods failed to perfectly reflect the input text despite its short length. In contrast, SALAD consistently incorporated all the textual descriptions into the generated motions, regardless of the length and complexity of the text prompts. This demonstrates its effectiveness in generating motions that are highly aligned with the input text." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7162" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.176, + 0.089, + 0.825, + 0.254 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
T2M [10]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
AttT2M [53]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo [54]0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
MoMask [12]0.521±.0020.713±.0020.807±.0020.045±.0022.958±.008-1.241±.040
MDM [42]0.320±.0050.498±.0040.611±.0070.544±.0045.566±.0279.559±.0862.799±.072
MLD [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse [51]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse [50]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
SALAD (Ours)0.581±.0030.769±.0030.857±.0020.076±.0022.649±.0099.696±.0961.751±.062
" + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.257, + 0.822, + 0.422 + ], + "angle": 0, + "content": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
T2M [10]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [53]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
ParCo [54]0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
MoMask [12]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
MDM [42]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.847±.1091.907±.2T4
MLD [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse [51]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse [50]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
SALAD (Ours)0.477±.0060.711±.0050.828±.0050.296±.0122.585±.01611.097±.0951.004±.040
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.426, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation results on the test sets of HumanML3D (top) and KIT-ML (bottom). \\(\\uparrow\\) and \\(\\downarrow\\) denote that higher and lower values are better, respectively, while \\(\\rightarrow\\) denotes that the values closer to the real motion are better. Methods above the dotted line are auto-regressive models based on VAE or VQ-VAE, while those below are diffusion-based generative models. Red and blue colors indicate the best and the second best results, respectively." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.486, + 0.885, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.686, + 0.882, + 0.701 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of results generated by various methods, including MDM [42], MoMask [12], and our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.245, + 0.725 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Skeleton-aware Latent Space. To prove the effectiveness of using skeleton-aware latent space, we compared the performance of different VAE structures employed in MoMask [12] and ParCo [54] with ours. Specifically, MoMask employs a VQ-VAE that compresses each pose into a single vector, while ParCo employs a part-aware VQ-VAE that processes different body parts independently. Because the role of VAE is to compress the motion features into latent variables, we compared FID and mean per joint position error (MPJPE) that assess the quality and accuracy of the reconstructed motion features, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.907, + 0.831 + ], + "angle": 0, + "content": "As shown in Tab. 2, our skeleton-aware VAE outperformed baselines with a significant margin for both metrics. Notably, our method required much fewer number of trainable parameters than the others. This indicates that skeleton-aware latent space that decouples both dimensions and facilitates information exchange between adjacent components is effective in constructing meaningful motion latent space, which is a crucial prior to train the generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, we conducted an ablation study that validates the effectiveness of the skeleto-temporal latent (ST-Latent) and cross-attention, which are two core components of SALAD, on the text-to-motion generation perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7163" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.159 + ], + "angle": 0, + "content": "
MethodsNum of ParamsFID ↓MPJPE ↓
MoMask [12]19.44M0.020±.0000.030±.000
ParCo [54]6.35M0.021±.0000.108±.000
SALAD (Ours)0.16M0.003±.0000.016±.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.163, + 0.483, + 0.218 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the quality and accuracy of reconstructed motion features of VAE models from different methods, along with the number of trainable parameters, measured on the test set of HumanML3D." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.226, + 0.48, + 0.317 + ], + "angle": 0, + "content": "
MethodR-Precision (Top-3) ↑FID ↓
Ours (Full model)0.857±.0020.076±.002
w/o ST-Latent0.816±.0020.433±.006
w/o CrossAttn0.778±.0020.274±.007
w/o both0.752±.0020.345±.007
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.318, + 0.44, + 0.332 + ], + "angle": 0, + "content": "Table 3. Ablation studies on the VAE and denoiser." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.548 + ], + "angle": 0, + "content": "mance. For this experiment, we evaluated three variations: (i) replacing ST-Latent by training a conventional VAE that treats each pose as a single vector, (ii) replacing cross-attention with self-attention by concatenating the sentence-level CLIP feature to the motion features, and (iii) combining (i) and (ii). As shown in Tab. 3, alteration of ST-Latent and cross-attention substantially degraded generation quality and text-motion alignment, as reflected in the higher FID and lower R-precision, respectively. This demonstrates the importance of rich information exchange between text and motion by cross-attention within the skeletotemporally structured latent space. This indicates that the skeletotemporally disentangled motion latent facilitates accurate text-motion interaction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.483, + 0.703 + ], + "angle": 0, + "content": "CFG Weights. To analyze the influence of CFG weights on generation results, we plotted FID and R-precision metric values in Fig. 5. The figure shows that low CFG weights led to suboptimal generation quality and text-motion alignment. The performance of SALAD improved across both metrics as the weight values increased, but excessively high weight values resulted in a decreased performance for both metrics again. Based on these findings, we determined that a value of \\( w = 7.5 \\) produces the best balance between generation quality and alignment with the text prompts." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.711, + 0.418, + 0.728 + ], + "angle": 0, + "content": "4.5. Zero-shot Text-driven Motion Editing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To demonstrate the capability of SALAD in capturing relationships between skeleto-temporal latents and text prompts, we visualized the cross-attention maps for each word in Fig. 6. The figure reveals that high attention values were assigned to proper frames and body parts that semantically correspond to each word. For example, \"jumps forward\" was activated for the entire body at early frames, \"walks forward\" focused on the root and legs at later frames, and \"waving arms\" was activated for the arms throughout the entire frames. These results indicate that cross-attention maps produced during motion generation ef" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.077, + 0.862, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.237, + 0.905, + 0.28 + ], + "angle": 0, + "content": "Figure 5. Ablation study results on different CFG weights. Lower FID scores indicate better motion quality, and higher R-precision (Top-3) scores indicate better semantic alignment." + }, + { + "type": "image_footnote", + "bbox": [ + 0.553, + 0.287, + 0.868, + 0.298 + ], + "angle": 0, + "content": "a man jumps forward and then walks forward, while waving his arms" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.304, + 0.903, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.529, + 0.906, + 0.6 + ], + "angle": 0, + "content": "Figure 6. Visualizations of cross-attention maps of SALAD. Each row corresponds to a specific body part, and each column represents temporal frames. Attention maps were computed by averaging the output attention maps across all transformer layers and all diffusion timesteps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.905, + 0.642 + ], + "angle": 0, + "content": "fectively capture the rich semantic relationships between text and motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Based on these observations, we demonstrate zero-shot text-driven motion editing results using a pre-trained SALAD model by modulating cross-attention maps, and the results are shown in Fig. 7. The edited motions align well with the given instructions while preserving the original movements, such as walking phases. Notably, the changes for editing were localized to specific frames and body parts relevant to the editing prompt, while unaffected regions remain intact, without using explicit masking. Furthermore, our SALAD model understands not only the high-level correspondence between text and motion but also its intensity and direction, which is demonstrated by different hand heights when modulating the word \"high\" in the reweighting case. These results suggest that the explicit modeling of skeletal-temporal-textual relationships in SALAD enhances its understanding of text-driven motion generation, leading to effective learning representations and fine" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7164" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.5, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.905, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.384 + ], + "angle": 0, + "content": "Figure 7. Text-driven motion editing results via attention modulation using a pre-trained SALAD model. Texts colored in green represent the editing instructions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.388, + 0.481, + 0.449 + ], + "angle": 0, + "content": "
MethodsPreservationSemantic AlignmentOverall Quality
MDM [42]3.729±.1532.758±.1923.196±.147
MotionFix [5]3.358±.1733.388±.1893.421±.166
SALAD (Ours)4.596±.0844.654±.0874.596±.083
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.454, + 0.482, + 0.468 + ], + "angle": 0, + "content": "Table 4. User study results. The red color indicates the best result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.399, + 0.491 + ], + "angle": 0, + "content": "grained control in both generation and editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.483, + 0.642 + ], + "angle": 0, + "content": "To evaluate the perceptual quality of text-driven motion editing, we conducted a user study with 16 participants. We compared SALAD against MDM [42], which edits motion by masking and regenerating specified body parts, and MotionFix [5], a method specifically designed for motion editing. Participants were shown with 15 videos per method and asked to rate them based on three criteria: (i) preservation of the original motion, (ii) semantic alignment with the target text, and (iii) overall quality. Ratings were given on a 5-point Likert scale, with 5 indicating the best." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.483, + 0.793 + ], + "angle": 0, + "content": "As shown in Tab. 4, SALAD outperformed both baselines across all three aspects by a large margin. Notably, despite not using explicit masking, SALAD achieved higher preservation scores. Furthermore, while SALAD is not specifically designed for editing, it achieved higher preferences for semantic alignment and overall quality, highlighting the effectiveness of attention-based editing. These results demonstrate that leveraging cross-attention maps enables high-quality and perceptually appealing text-driven motion editing without requiring additional training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.205, + 0.816 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "While SALAD produced outperforming results in text-driven motion generation compared to the existing generative models, along with the capability of zero-shot text-driven motion editing, it has several limitations. First, the numerical results for Diversity and MultiModality are lim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.908, + 0.603 + ], + "angle": 0, + "content": "ited, despite the strong capability of diffusion models to generate diverse outputs. This can be due to the trade-off between high-quality with faithfulness to the text and diversity. While we put the diversity as a second priority, finding a way to handle both could be an interesting approach. Additionally, SALAD is restricted to generating a single person action with limited text and motion length. Therefore, enabling multi-character interactions, crowd animations, generation from longer texts, and creating extended motion sequences could be an interesting direction for future work. Furthermore, although our method does not directly address editing real motions, integrating diffusion inversion methods could enable text-driven editing of real motion sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.634, + 0.634 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this paper, we presented SALAD, a novel diffusion model framework for text-driven motion generation that models the complex interactions between skeletal joints, temporal frames, and textual words. We first train a skeletotemporal VAE that effectively summarizes motion features by decoupling skeletal and temporal dimensions, and we subsequently train a skeletotemporal denoiser that enables efficient and expressive generation of motions conditioned on text input within this compact skeletotemporal latent space. Furthermore, we demonstrated the first text-driven zero-shot motion editing using a pre-trained SALAD model by modulating cross-attention maps during generation. By leveraging the outperforming generative capacity compared to previous methods, along with the interpretable learning representations during generation, we anticipate that our work can benefit future research and applications in the context of text-driven motion generation and editing." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7165" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.177 + ], + "angle": 0, + "content": "We appreciate the anonymous reviewers for their invaluable discussion. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00333478)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.188, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[3] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[4] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[5] Nikos Athanasiou, Alpár Cseke, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[6] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[12] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.352, + 0.904, + 0.38 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.426, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[18] Yiming Huang, Weilin Wan, Yue Yang, Chris Callison-Burch, Mark Yatskar, and Lingjie Liu. Como: Controllable motion generation through language guided pose code editing. In European Conference on Computer Vision, pages 180-196. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[19] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[20] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.601, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[22] Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwajanakorn, and Siyu Tang. Optimizing diffusion noise can serve as universal motion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1345, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[25] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7166" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[27] Yunhong Lou, Linchao Zhu, Yaxiong Wang, Xiaohan Wang, and Yi Yang. Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[29] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[30] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on computer graphics and interactive techniques, 4(3):1-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[31] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[32] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[34] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[35] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[41] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[42] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[43] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.434 + ], + "angle": 0, + "content": "[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.436, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[45] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[46] Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22035-22044, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[47] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[48] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[49] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[50] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7167" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[52] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7368-7376, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[53] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[54] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7168" + } + ] +] \ No newline at end of file diff --git a/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_origin.pdf b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a37d64f96f5c767428beb58af9faa7113370b4c7 --- /dev/null +++ b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/56cf0cf2-143c-481b-820b-cc5bf0c85862_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34368d8cbe017da24fde9656001dceca5babf76303ba8b67838886185d69a253 +size 3740482 diff --git a/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/full.md b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1d90d9b351366fe3cec9eef4a74f68abc23c17cd --- /dev/null +++ b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/full.md @@ -0,0 +1,304 @@ +# SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing + +Seokhyeon Hong Chaelin Kim Serin Yoon Junghyun Nam Sihun Cha Junyong Noh Visual Media Lab, KAIST + +{ghd3079, chaelin.kim, serinyoon, ys4990, chacorp, junyongnoh}@kaist.ac.kr + +# Abstract + +Text-driven motion generation has advanced significantly with the rise of denoising diffusion models. However, previous methods often oversimplify representations for the skeletal joints, temporal frames, and textual words, limiting their ability to fully capture the information within each modality and their interactions. Moreover, when using pretrained models for downstream tasks, such as editing, they typically require additional efforts, including manual interventions, optimization, or fine-tuning. In this paper, we introduce a skeleton-aware latent diffusion (SALAD), a model that explicitly captures the intricate inter-relationships between joints, frames, and words. Furthermore, by leveraging cross-attention maps produced during the generation process, we enable attention-based zero-shot text-driven motion editing using a pre-trained SALAD model, requiring no additional user input beyond text prompts. Our approach significantly outperforms previous methods in terms of text-motion alignment without compromising generation quality, and demonstrates practical versatility by providing diverse editing capabilities beyond generation. Code is available at project page. + +# 1. Introduction + +Character animation is a crucial component in various computer graphics and vision applications, including games, films, and interactive media. Traditional methods, such as keyframing and motion capture, typically require extensive manual effort, which are time-consuming and expensive. Recently, deep generative models have been introduced to mitigate these challenges. In particular, diffusion models have shown promising results in text-to-motion generation, enabling intuitive and efficient animation workflows. + +Despite these advancements, fully capturing the intricate relationships among frames, body parts, and textual descriptions remains a complex challenge in text-driven motion generation. Previous approaches typically represent a pose as a single vector and primarily focus on temporal re + +relationships between poses across frames, while neglecting spatial interactions between skeletal joints. Furthermore, these models often compress a sentence into a single vector for conditioning, which can overlook the important nuances of word-level variations. This over-simplification of interactions among skeletal, temporal, and textual components may lead to missing details in the generated results, underscoring the need for models that faithfully capture these complex dependencies. + +On the other hand, learning meaningful representations is a key factor to enable zero-shot downstream tasks with pre-trained models. For example, zero-shot text-driven image and video editing methods [3, 7, 13, 15, 25, 35, 43] leverage attention modulation from pre-trained models for intuitive manipulation. Unfortunately, motion generation models typically lack such interpretable and manipulatable intermediate representations. This limitation stems from the over-simplification of motion and text features mentioned above, which restricts rich interactions between them. Although some methods exploit pre-trained models with manual masks, optimization, or fine-tuning [21-23, 42, 45, 47], these methods require additional effort, time, and cost to achieve desired results. Therefore, developing interpretable representations during motion synthesis could improve versatility and flexibility for downstream tasks beyond generation, such as zero-shot motion editing. + +In this paper, we propose a skeleton-aware latent diffusion, which we call SALAD, for text-driven motion generation within a skeletotemporally structured latent space. We first train a variational autoencoder (VAE) [24] to construct a motion latent space that decouples spatial and temporal dimensions. To this end, we employ skeletotemporal convolution layers to facilitate information exchange between adjacent joints and frames. We also employ skeletotemporal pooling layers to produce a compact representation by reducing the dimensionality, which can reduce the computational complexity during the sampling process of the diffusion model. Subsequently, we train a diffusion model within this latent space to generate motion features conditioned on text. To train the denoiser, we leverage both + +skeletal and temporal attention blocks, which can effectively enable skeletotemporal coherence. Additionally, we employ cross-attention to capture fine-grained interactions between individual words and the skeletotemporal units of the motion latents. + +Beyond motion generation, we introduce an attention-based zero-shot, text-driven motion editing method that leverages the pre-trained SALAD model, requiring no additional optimization or fine-tuning. We demonstrate that the intermediate cross-attention maps of SALAD capture the relationship between text and motion features. By modulating these cross-attention maps, we enable text-driven motion editing that allows users to edit generated motion through text input alone in a training-free manner. We also propose a novel attention modulation method for motion-specific tasks, demonstrating the practical potential of our approach. + +In summary, our contributions are as follows: + +- We propose SALAD, a novel skeleton-aware latent diffusion model for text-driven motion generation within a skeletotemporally structured latent space. +- We interpret the intermediate representations in the generation process, allowing for a clear understanding of the relationship between text inputs and generated motions. +- We present an attention-based zero-shot text-driven motion editing method using cross-attention modulation in generative models. + +# 2. Related Work + +# 2.1. Text-driven Motion Generation + +With the rise of deep neural networks, generative models for human motion synthesis based on natural language descriptions have been extensively explored. Early studies focused on generating motions from action labels, such as throw and kick [9, 32]. More advanced methods projected both text and motion features into a shared latent space, enabling the generation of fixed motion sequences from identical input text [2]. To increase the diversity of generated results, T2M [10] employed a temporal VAE that can generate varied motion sequences that align with the given text input. Subsequent studies leveraged large pretrained text encoders, such as CLIP [36], to improve textual understanding for motion generation [4, 33, 41]. Autoregressive models that behave like language models have also been proposed, which first encode motion features into discrete tokens through vector quantization [44] and then auto-regressively predict the next motion token given the text prompt [11, 12, 20, 49, 52-54]. Recently, denoising diffusion models [17, 39] have become the dominant framework for generative models, prompting text-driven motion generation methods to adopt diffusion models as their backbone [8, 23, 27, 42, 46, 50, 51]. In this study, we build + +our framework on the diffusion model formulation and enhance the performance of text-driven motion generation through explicit modeling of the relationships among joints, frames, and words. Moreover, we introduce an attention-based zero-shot text-driven editing method, leveraging our pre-trained SALAD model as a generator, demonstrating a novel exploration beyond mere generation. + +# 2.2. Skeleton-aware Processing of Motion Data + +Because motion data consists of two separate skeletal and temporal dimensions, some methods have been developed to respect each dimension when processing the motion data. Yan et al. [48] presented spatial-temporal graph convolution networks that can learn both the spatial and temporal patterns of dynamic skeletons. Skeleton-aware networks [1] employed skeleto-temporal convolution networks and topology-preserving skeletal pooling layers to construct a shared motion latent space, enabling deep motion retargeting across homeomorphic skeletons. Motion style transfer has also been enabled in a part-wise manner by employing skeleto-temporal convolution modules [19, 30]. In the realm of text-to-motion generation, ParCo [54] and AttT2M [53] utilized the skeleto-temporal network architecture during the motion quantization process. In this work, we adopt a similar approach by encoding motion features using skeleto-temporal convolution networks and pooling layers to derive an effective skeleton-aware motion latent space with compact dimensionality. Additionally, we facilitate the interaction between this skeleton-aware motion latent space and textual features during the text-conditional generation process. + +# 2.3. Zero-shot Editing with Diffusion Models + +By leveraging the generative capabilities of diffusion models, zero-shot editing methods have been proposed. SDEdit [28] first adds noise to the input data for a few diffusion timesteps, then denoises the noisy data with an editing condition to achieve high-fidelity zero-shot editing by leveraging a pre-trained diffusion model. Prompt-to-Prompt [15] verified that cross-attention maps establish connections between word tokens and the spatial layout of the generated image. In addition, it introduced attention modulation methods that can be utilized during the generation process, enabling the generation of high-fidelity editing results in a zero-shot manner. Null-text inversion [29] extended this method to real-images by improving the diffusion inversion method. Similarly, image and video editing methods via attention modulation have been extensively studied [3, 7, 13, 25, 35, 43]. Motion diffusion models often employ joint masks [23, 42] or optimization [21, 22, 47] to enable motion editing using pre-trained diffusion models, although they require user intervention or additional computation for the desired goal. On the other hand, CoMo [18] + +![](images/45f5cbd20cfd69199d653edb9d49ca94a573eba579db4e95c3cffc35baec46d6.jpg) +Figure 1. Architecture of the skeletotemporal VAE network. The encoder maps motion features into a skeletotemporal latent space, and the decoder restores the skeletotemporal latent variables into motion features. + +achieved zero-shot text-driven motion editing using large language models to manipulate semantic pose tokens. We demonstrate that the intermediate cross-attention maps of our SALAD model can capture the relationship between motion and text, similarly to image diffusion models. Furthermore, by modulating these cross-attention maps, we introduce an attention-based zero-shot text-driven motion editing method that eliminates the need for manual masks, fine-tuning, or optimization. + +# 3. Method + +Our goal is to generate a motion sequence $\mathbf{m}_{1:N}$ conditioned on a text prompt $c$ , where $N$ denotes the number of frames. We first construct a skeleton-aware motion latent representation that captures both skeletal and temporal dynamics of a motion sequence (Sec. 3.1). Using this motion representation, we train a diffusion model by modeling the complex interactions between skeletal joints, temporal frames, and textual descriptions (Sec. 3.2). Finally, we extend our model for zero-shot text-driven motion editing by modulating cross-attention maps between text and motion (Sec. 3.3). Additional details on the network architectures will be featured in the supplementary material. + +# 3.1. Skeleton-aware VAE + +To derive skeleton-aware motion features, we train a VAE composed of skeletotemporal convolution networks and pooling layers [1, 48], as illustrated in Fig. 1. The core module is the skeletotemporal convolution (STConv), which decouples joint and frame dimensions while facilitating information exchange between adjacent components within each dimension. While the skeletotemporal separation of motion sequences allows for sophisticated modeling of motion data, it introduces an additional dimension for skeletal joints, unlike previous text-to-motion generation methods that treat a pose as a single vector. This expanded data space increases the risk of curse of dimensionality and leads to longer computation times for diffusion models, which rely on solving stochastic differential equations within this higher-dimensional data space. To address this issue, we apply skeletotemporal pooling (STPool) layers in the encoder to reduce the latent space dimensionality, effectively + +compressing the representation. In the decoder, skeletotemporal unpooling (STUnpool) layers are used to reconstruct the compressed latent features back to raw motion. + +Encoder. We separate each pose vector of $\mathbf{m}_{1:N}$ into jointwise features and propagate them through joint-wise multilayer perceptron (MLP) layers, resulting in $\mathbf{h} \in \mathbb{R}^{N \times J \times D}$ , where $J$ denotes the number of joints and $D$ denotes the latent dimension. Subsequently, we apply STConv to $\mathbf{h}$ to facilitate information exchange between adjacent frames and joints: + +$$ +\operatorname {S T C o n v} (\mathbf {h}) := \operatorname {S k e l C o n v} (\mathbf {h}) + \operatorname {T e m p C o n v} (\mathbf {h}), \tag {1} +$$ + +where SkelConv is a graph convolution network over the joint dimension, while TempConv is a 1D convolution network over the frame dimension. + +We apply the STPool operator to the output of the STConv layers to reduce its dimensionality by pooling over joints and frames: + +$$ +\operatorname {S T P o o l} (\mathbf {h}) := \operatorname {T e m p P o o l} (\operatorname {S k e l P o o l} (\mathbf {h})), \tag {2} +$$ + +where SkelPool represents the pooling layer over the joint dimension while TempPool represents the pooling layer over the temporal dimension. Specifically, SkelPool aggregates adjacent joints while preserving its skeletal topology, resulting in fewer joints with a homeomorphic structure. In contrast, TempPool performs 1D pooling over frames. These pooling operations are commutative because they operate within independent dimensions. + +Consequently, we obtain the skeleto-temporally compressed latent vector $\mathbf{z} \in \mathbb{R}^{N' \times J' \times D}$ , where $N' < N$ and $J' < J$ . This representation effectively encodes the skeleton-aware dynamics of the motion sequence within a lower dimension. Notably, we maintain 7 atomic joints for this motion latent: root, spine, head, both arms, and both legs $(J' = 7)$ . + +Decoder. To reconstruct $\mathbf{z}$ into raw motion features, we employ STConv and STUnpool layers in the decoder. The decoder mirrors the encoder, facilitating information exchange between adjacent joints and frames using STConv layers and increasing the skeleto-temporal resolution using STUnpool layers. Finally, the joint-wise features are processed by their respective joint-wise MLP layers, resulting in the reconstructed motion feature $\hat{\mathbf{m}}_{1:N}$ . + +![](images/723af0545820b071a09c43637c46d70284980d660693d8582d64bc1969f90f87.jpg) +Figure 2. (Left) Overall network architecture of the denoiser. (Right) The architecture of each transformer block consisting of TempAttn, SkelAttn, CrossAttn, and FFN, along with the FiLM following each module. + +![](images/9591fe8ed372fd2599f635b889b8e84cdbf6b5fa236dfaf626402a8714601f38.jpg) + +Training. The VAE is trained with the following objectives: + +$$ +\mathcal {L} _ {\mathrm {V A E}} = \mathcal {L} _ {\mathbf {m}} + \lambda_ {\text {p o s}} \mathcal {L} _ {\text {p o s}} + \lambda_ {\text {v e l}} \mathcal {L} _ {\text {v e l}} + \lambda_ {\mathrm {k l}} \mathcal {L} _ {\mathrm {k l}}, \tag {3} +$$ + +where $\mathcal{L}_{\mathbf{m}},\mathcal{L}_{\mathrm{pos}},\mathcal{L}_{\mathrm{vel}}$ are L1 reconstruction loss terms of the motion features, joint positions, and joint velocities, while $\mathcal{L}_{\mathrm{kl}}$ is the Kullback-Leibler (KL) divergence regularization that encourages a structured latent space. + +# 3.2. Skeleton-aware Denoiser + +Network Architecture. Given the skeleton-aware motion latent vector $\mathbf{z} \in \mathbb{R}^{N' \times J' \times D}$ , we train a diffusion model that denoises the noised latent vector conditioned on text prompt $c$ . As shown in Fig. 2, our denoiser consists of a positional embedding, MLP-based encoder and decoder, and $L$ transformer-based layers. Each transformer layer is composed of a temporal attention (TempAttn), skeletal attention (SkelAttn), cross-attention (CrossAttn), and feedforward network (FFN). Each component involves a residual connection [14], layer normalization (LN) [6], and a feature-wise linear modulation (FiLM) [31] that modulates intermediate features based on a diffusion timestep. We also incorporate skip connection [37] within the stack of transformer layers. + +The input motion latent at diffusion timestep $t$ given to the $l$ -th transformer layer, denoted as $\mathbf{z}_t^l$ , is processed by three consecutive attention blocks: temporal, skeletal, and cross-attention layers. The purpose of this architecture design is to model frame-wise interactions through TempAttn, joint-wise interactions through SkelAttn, and motion-text interactions through CrossAttn. As a result, $\mathbf{z}_t^l$ is consecutively updated at each layer as follows: + +$$ +\mathbf {z} _ {t} ^ {l} \gets \mathbf {z} _ {t} ^ {l} + \mathrm {F i L M} (\mathrm {T e m p A t t n} (\mathrm {L N} (\mathbf {z} _ {t} ^ {l}))), \qquad (4) +$$ + +$$ +\mathbf {z} _ {t} ^ {l} \leftarrow \mathbf {z} _ {t} ^ {l} + \operatorname {F i L M} (\operatorname {S k e l A t t n} (\operatorname {L N} (\mathbf {z} _ {t} ^ {l}))), \tag {5} +$$ + +$$ +\mathbf {z} _ {t} ^ {l} \leftarrow \mathbf {z} _ {t} ^ {l} + \mathrm {F i L M} (\operatorname {C r o s s A t t n} (\operatorname {L N} (\mathbf {z} _ {t} ^ {l}), \operatorname {C L I P} (c))). \quad (6) +$$ + +For the CLIP text encoder [36] used in cross-attention, we use a pre-trained model and freeze its weights during training. + +Diffusion Parametrization. For diffusion parametrization, our model predicts the diffusion velocity [38], denoted as + +$\mathbf{v}$ , as follows: + +$$ +\mathbf {v} _ {t} = \alpha_ {t} \boldsymbol {\epsilon} - \sigma_ {t} \mathbf {x}, \tag {7} +$$ + +where $\epsilon$ and $\mathbf{x}$ represent the noise and sample, respectively, while $\alpha_{t}$ and $\sigma_{t}$ are noise scheduling parameters at diffusion timestep $t$ . While $\epsilon$ - and $\mathbf{x}$ -prediction are widely used in diffusion models, $\mathbf{v}$ -prediction combines both parametrizations, implicitly balancing the information contributed by each component. This improves the stability of generation, especially at high noise levels in which the relationship between $\epsilon$ and $\mathbf{x}$ becomes disrupted. + +Training and Inference. The denoiser is trained to predict the diffusion velocity $\mathbf{v}_t$ at each diffusion timestep $t$ . The training objective for the denoiser is defined as follows: + +$$ +\mathcal {L} _ {\text {d e n o i s e r}} = \left\| \hat {\mathbf {v}} _ {t} - \mathbf {v} _ {t} \right\| _ {2} ^ {2}, \tag {8} +$$ + +where $\hat{\mathbf{v}}_t$ represents the predicted diffusion velocity by the denoiser. During training, we randomly drop the text condition with a probability of $p_{\mathrm{uncond}}$ , enabling unconditional generation when $c$ is a null text, which is represented as $\varnothing$ . + +To the trained denoiser, we apply classifier-free guidance (CFG) [16] for text-conditioned generation: + +$$ +\hat {\mathbf {v}} _ {\theta} \left(\mathbf {z} _ {t}, t, c\right) := \mathbf {v} _ {\theta} \left(\mathbf {z} _ {t}, t, \varnothing\right) + w \left(\mathbf {v} _ {\theta} \left(\mathbf {z} _ {t}, t, c\right) - \mathbf {v} _ {\theta} \left(\mathbf {z} _ {t}, t, \varnothing\right)\right). \tag {9} +$$ + +Here, $\mathbf{v}_{\theta}$ denotes the pre-trained denoiser, $w$ is the CFG weight, and $\hat{\mathbf{v}}_{\theta}$ is the modified velocity used during the denoising process. Additionally, we employ the DDIM sampling formulation [40] during inference, which effectively reduces the number of sampling steps without compromising generation quality. + +# 3.3. Zero-shot Text-driven Motion Editing + +In this section, we describe the attention modulation methods that enable zero-shot text-driven motion editing with a pre-trained SALAD model. We introduce four distinct modulation strategies applied to the text-motion cross attention maps as shown in Fig. 3: word swap, prompt refinement, attention re-weighting, and attention mirroring. The first three are adapted from the Prompt-to-Prompt framework [15], while the attention mirroring serves as a validation method to assess whether the cross-attention map + +![](images/0084cfd6109f28f04c9bc385f2fd4b8d3f30a6917842177aa51fc34a5ae1f5cb.jpg) +Figure 3. Attention modulation methods applied to the cross-attention maps to enable text-driven motion editing. + +contributes to the activation of spatio-temporal dynamics in the motion generation process. Word swap exchanges attention maps between a source and target prompt, enabling the model to leverage the attention context of the source text when generating output for the target text. Prompt refinement enhances the original attention map by integrating additional attention maps for appended word tokens, providing richer semantic information compared to the base prompt. Attention re-weighting adjusts the attention values assigned to specific user-selected words, either amplifying or diminishing their influence on the generated motion. Lastly, attention mirroring swaps attention values between symmetrical body parts, such as the left and right arms, to produce mirrored motions. These methods collectively empower the generative capacity of our model to produce dynamic and contextually relevant motion edits based solely on text prompts. For more details on these modulation functions, please refer to the supplementary material. + +# 4. Experiments + +We conducted evaluations of our method on two widely used motion-language benchmarks: HumanML3D [10] and KIT-ML [34]. HumanML3D dataset consists of 14,616 motion sequences with a variety of human actions, such as locomotions, sports, and acrobatics, along with 44,970 text descriptions in total. KIT-ML dataset consists of 3,911 motion sequences along with 6,278 text descriptions. Both datasets are augmented by mirroring, and divided into training, testing, and validation sets with the ratio of 0.8, 0.15, and 0.05. Evaluation results are reported in the following sections. Please refer to the supplementary material for additional quantitative experiment results with details, and the supplementary video for the animated results. + +# 4.1. Implementation Details + +Our method was executed on a single NVIDIA V100 GPU. We used the AdamW optimizer [26], and the VAE and denoiser were trained for 50 and 500 epochs, respectively. The VAE and denoiser were trained with a batch size of 64 for HumanML3D, which required 35 and 17 hours, re + +spectively. For the KIT-ML dataset, we used a batch size of 16, with training times of 20 hours for the VAE and 7 hours for the denoiser. We trained the denoiser with 1000 diffusion steps, employing 50 steps for DDIM sampling during inference. For the CFG weight, we set $w = 7.5$ unless mentioned otherwise. + +# 4.2.Quantitative Evaluation + +For quantitative evaluations, we adopted metrics from T2M [10]. First, $FID$ is used to evaluate the quality of generated motions by measuring the difference between the feature distributions of generated motions and those of real motions. Additionally, $R$ -precision and $MM-Dist$ are used to evaluate the semantic alignment between the input text and the generated motions. $Diversity$ , which indicates the variance in generated motions, and $MultiModality$ , which assesses the diversity of motions generated from the same text description, are also adopted as secondary metrics. They are considered less important than the generation quality and alignment with the input text. + +Following the practice of T2M [10], each experiment was repeated 20 times on the test sets of HumanML3D and KIT-ML datasets, and the mean with a $95\%$ confidence interval is reported in Tab. 1. Notably, SALAD consistently achieved superior text-motion alignment across both datasets, regardless of the diffusion parametrization applied during denoiser training. Moreover, SALAD maintained high generation quality, as indicated by the best or second-best FID scores among diffusion-based methods. These results confirm that explicitly modeling interactions between skeletal joints, temporal frames, and textual words is effective in improving text-to-motion generation quality and fidelity to the input text. + +# 4.3. Qualitative Evaluation + +We compared the qualitative results of SALAD with those generated by MDM [42] and MoMask [12], and the comparison results are shown in Fig. 4. Both MDM and MoMask reflected parts of the input text prompts, but they failed to handle all the descriptions in the text. For example, as shown in the top of Fig. 4-(a), MDM generated a character walking forward, turning, and walking back, but the locomotion style was not marching, despite the input prompt specifying "marches". Similarly, in Fig. 4-(b), MoMask accurately reflected the arm movements of raising above the head and throwing them down, but it missed the instruction to turn to the right. In case of Fig. 4-(c), both methods failed to perfectly reflect the input text despite its short length. In contrast, SALAD consistently incorporated all the textual descriptions into the generated motions, regardless of the length and complexity of the text prompts. This demonstrates its effectiveness in generating motions that are highly aligned with the input text. + +
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
T2M [10]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
AttT2M [53]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo [54]0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
MoMask [12]0.521±.0020.713±.0020.807±.0020.045±.0022.958±.008-1.241±.040
MDM [42]0.320±.0050.498±.0040.611±.0070.544±.0045.566±.0279.559±.0862.799±.072
MLD [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse [51]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse [50]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
SALAD (Ours)0.581±.0030.769±.0030.857±.0020.076±.0022.649±.0099.696±.0961.751±.062
+ +
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
T2M [10]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [53]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
ParCo [54]0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
MoMask [12]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
MDM [42]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.847±.1091.907±.2T4
MLD [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse [51]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse [50]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
SALAD (Ours)0.477±.0060.711±.0050.828±.0050.296±.0122.585±.01611.097±.0951.004±.040
+ +Table 1. Quantitative evaluation results on the test sets of HumanML3D (top) and KIT-ML (bottom). $\uparrow$ and $\downarrow$ denote that higher and lower values are better, respectively, while $\rightarrow$ denotes that the values closer to the real motion are better. Methods above the dotted line are auto-regressive models based on VAE or VQ-VAE, while those below are diffusion-based generative models. Red and blue colors indicate the best and the second best results, respectively. + +![](images/5cada3a7308061c93108297ac11eef69b63565e13fd4d8ff16ac892e30e1af71.jpg) +Figure 4. Qualitative comparison of results generated by various methods, including MDM [42], MoMask [12], and our approach. + +# 4.4. Ablation Study + +Effectiveness of Skeleton-aware Latent Space. To prove the effectiveness of using skeleton-aware latent space, we compared the performance of different VAE structures employed in MoMask [12] and ParCo [54] with ours. Specifically, MoMask employs a VQ-VAE that compresses each pose into a single vector, while ParCo employs a part-aware VQ-VAE that processes different body parts independently. Because the role of VAE is to compress the motion features into latent variables, we compared FID and mean per joint position error (MPJPE) that assess the quality and accuracy of the reconstructed motion features, respectively. + +As shown in Tab. 2, our skeleton-aware VAE outperformed baselines with a significant margin for both metrics. Notably, our method required much fewer number of trainable parameters than the others. This indicates that skeleton-aware latent space that decouples both dimensions and facilitates information exchange between adjacent components is effective in constructing meaningful motion latent space, which is a crucial prior to train the generator. + +Furthermore, we conducted an ablation study that validates the effectiveness of the skeleto-temporal latent (ST-Latent) and cross-attention, which are two core components of SALAD, on the text-to-motion generation perfor + +
MethodsNum of ParamsFID ↓MPJPE ↓
MoMask [12]19.44M0.020±.0000.030±.000
ParCo [54]6.35M0.021±.0000.108±.000
SALAD (Ours)0.16M0.003±.0000.016±.000
+ +Table 2. Quantitative results on the quality and accuracy of reconstructed motion features of VAE models from different methods, along with the number of trainable parameters, measured on the test set of HumanML3D. + +
MethodR-Precision (Top-3) ↑FID ↓
Ours (Full model)0.857±.0020.076±.002
w/o ST-Latent0.816±.0020.433±.006
w/o CrossAttn0.778±.0020.274±.007
w/o both0.752±.0020.345±.007
+ +Table 3. Ablation studies on the VAE and denoiser. + +mance. For this experiment, we evaluated three variations: (i) replacing ST-Latent by training a conventional VAE that treats each pose as a single vector, (ii) replacing cross-attention with self-attention by concatenating the sentence-level CLIP feature to the motion features, and (iii) combining (i) and (ii). As shown in Tab. 3, alteration of ST-Latent and cross-attention substantially degraded generation quality and text-motion alignment, as reflected in the higher FID and lower R-precision, respectively. This demonstrates the importance of rich information exchange between text and motion by cross-attention within the skeletotemporally structured latent space. This indicates that the skeletotemporally disentangled motion latent facilitates accurate text-motion interaction. + +CFG Weights. To analyze the influence of CFG weights on generation results, we plotted FID and R-precision metric values in Fig. 5. The figure shows that low CFG weights led to suboptimal generation quality and text-motion alignment. The performance of SALAD improved across both metrics as the weight values increased, but excessively high weight values resulted in a decreased performance for both metrics again. Based on these findings, we determined that a value of $w = 7.5$ produces the best balance between generation quality and alignment with the text prompts. + +# 4.5. Zero-shot Text-driven Motion Editing + +To demonstrate the capability of SALAD in capturing relationships between skeleto-temporal latents and text prompts, we visualized the cross-attention maps for each word in Fig. 6. The figure reveals that high attention values were assigned to proper frames and body parts that semantically correspond to each word. For example, "jumps forward" was activated for the entire body at early frames, "walks forward" focused on the root and legs at later frames, and "waving arms" was activated for the arms throughout the entire frames. These results indicate that cross-attention maps produced during motion generation ef + +![](images/9ab25601c73dcc4dc025aa37c182e9e71d310c4af33b8628f4f58a7f3aab3708.jpg) +Figure 5. Ablation study results on different CFG weights. Lower FID scores indicate better motion quality, and higher R-precision (Top-3) scores indicate better semantic alignment. + +Figure 6. Visualizations of cross-attention maps of SALAD. Each row corresponds to a specific body part, and each column represents temporal frames. Attention maps were computed by averaging the output attention maps across all transformer layers and all diffusion timesteps. +![](images/8a1e1bcdad2dcb9062350b0e8d40f722193758e7e4cb6a444d1cca8e32ad492c.jpg) +a man jumps forward and then walks forward, while waving his arms + +fectively capture the rich semantic relationships between text and motion. + +Based on these observations, we demonstrate zero-shot text-driven motion editing results using a pre-trained SALAD model by modulating cross-attention maps, and the results are shown in Fig. 7. The edited motions align well with the given instructions while preserving the original movements, such as walking phases. Notably, the changes for editing were localized to specific frames and body parts relevant to the editing prompt, while unaffected regions remain intact, without using explicit masking. Furthermore, our SALAD model understands not only the high-level correspondence between text and motion but also its intensity and direction, which is demonstrated by different hand heights when modulating the word "high" in the reweighting case. These results suggest that the explicit modeling of skeletal-temporal-textual relationships in SALAD enhances its understanding of text-driven motion generation, leading to effective learning representations and fine + +![](images/5c1f4109ad5038b2d13b111cf439c198e1bb50d536b1eac80dc9b346736e39e6.jpg) +Figure 7. Text-driven motion editing results via attention modulation using a pre-trained SALAD model. Texts colored in green represent the editing instructions. + +![](images/98a10733bf4aaf03205648c52782d2dd771e92d8cb512ea2136c984bdfc48095.jpg) + +
MethodsPreservationSemantic AlignmentOverall Quality
MDM [42]3.729±.1532.758±.1923.196±.147
MotionFix [5]3.358±.1733.388±.1893.421±.166
SALAD (Ours)4.596±.0844.654±.0874.596±.083
+ +Table 4. User study results. The red color indicates the best result. + +grained control in both generation and editing. + +To evaluate the perceptual quality of text-driven motion editing, we conducted a user study with 16 participants. We compared SALAD against MDM [42], which edits motion by masking and regenerating specified body parts, and MotionFix [5], a method specifically designed for motion editing. Participants were shown with 15 videos per method and asked to rate them based on three criteria: (i) preservation of the original motion, (ii) semantic alignment with the target text, and (iii) overall quality. Ratings were given on a 5-point Likert scale, with 5 indicating the best. + +As shown in Tab. 4, SALAD outperformed both baselines across all three aspects by a large margin. Notably, despite not using explicit masking, SALAD achieved higher preservation scores. Furthermore, while SALAD is not specifically designed for editing, it achieved higher preferences for semantic alignment and overall quality, highlighting the effectiveness of attention-based editing. These results demonstrate that leveraging cross-attention maps enables high-quality and perceptually appealing text-driven motion editing without requiring additional training. + +# 5. Discussion + +While SALAD produced outperforming results in text-driven motion generation compared to the existing generative models, along with the capability of zero-shot text-driven motion editing, it has several limitations. First, the numerical results for Diversity and MultiModality are lim + +ited, despite the strong capability of diffusion models to generate diverse outputs. This can be due to the trade-off between high-quality with faithfulness to the text and diversity. While we put the diversity as a second priority, finding a way to handle both could be an interesting approach. Additionally, SALAD is restricted to generating a single person action with limited text and motion length. Therefore, enabling multi-character interactions, crowd animations, generation from longer texts, and creating extended motion sequences could be an interesting direction for future work. Furthermore, although our method does not directly address editing real motions, integrating diffusion inversion methods could enable text-driven editing of real motion sequences. + +# 6. Conclusion + +In this paper, we presented SALAD, a novel diffusion model framework for text-driven motion generation that models the complex interactions between skeletal joints, temporal frames, and textual words. We first train a skeletotemporal VAE that effectively summarizes motion features by decoupling skeletal and temporal dimensions, and we subsequently train a skeletotemporal denoiser that enables efficient and expressive generation of motions conditioned on text input within this compact skeletotemporal latent space. Furthermore, we demonstrated the first text-driven zero-shot motion editing using a pre-trained SALAD model by modulating cross-attention maps during generation. By leveraging the outperforming generative capacity compared to previous methods, along with the interpretable learning representations during generation, we anticipate that our work can benefit future research and applications in the context of text-driven motion generation and editing. + +# Acknowledgements + +We appreciate the anonymous reviewers for their invaluable discussion. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00333478). + +# References + +[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2, 3 +[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2 +[3] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1, 2 +[4] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2 +[5] Nikos Athanasiou, Alpár Cseke, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 8 +[6] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2 +[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2, 6 +[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2 +[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 5, 6 +[11] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2 + +[12] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 5, 6, 7 +[13] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 1, 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 4 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[18] Yiming Huang, Weilin Wan, Yue Yang, Chris Callison-Burch, Mark Yatskar, and Lingjie Liu. Como: Controllable motion generation through language guided pose code editing. In European Conference on Computer Vision, pages 180-196. Springer, 2024. 2 +[19] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2 +[20] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2 +[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 1, 2 +[22] Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwajanakorn, and Siyu Tang. Optimizing diffusion noise can serve as universal motion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1345, 2024. 2 +[23] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 1, 2 +[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1 +[25] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1, 2 + +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5 +[27] Yunhong Lou, Linchao Zhu, Yaxiong Wang, Xiaohan Wang, and Yi Yang. Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372, 2023. 2 +[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[29] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2 +[30] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on computer graphics and interactive techniques, 4(3):1-17, 2021. 2 +[31] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 4 +[32] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2 +[33] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2 +[34] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 5 +[35] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1, 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4 +[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4 + +[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4 +[41] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2 +[42] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 8 +[43] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2 +[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[45] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 1 +[46] Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22035-22044, 2023. 2 +[47] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 1, 2 +[48] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3 +[49] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2 +[50] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2, 6 +[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 6 + +[52] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7368-7376, 2024. 2 +[53] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2, 6 +[54] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2, 6, 7 \ No newline at end of file diff --git a/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/images.zip b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..073591734daf26830342e5b3d62a91e33166e65f --- /dev/null +++ b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f866a57d72399bc3de198aca07a4fef96a91f30811897c537d0a744ea895935b +size 676459 diff --git a/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/layout.json b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9340b581506c7bcb18f57a0f4a30844bbcb15115 --- /dev/null +++ b/2025/SALAD_ Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing/layout.json @@ -0,0 +1,7601 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 452, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 452, + 140 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 452, + 140 + ], + "type": "text", + "content": "SALAD: Skeleton-aware Latent Diffusion for Text-driven Motion Generation and Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 161, + 536, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 161, + 536, + 189 + ], + "spans": [ + { + "bbox": [ + 74, + 161, + 536, + 189 + ], + "type": "text", + "content": "Seokhyeon Hong Chaelin Kim Serin Yoon Junghyun Nam Sihun Cha Junyong Noh Visual Media Lab, KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 191, + 471, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 191, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 191, + 471, + 204 + ], + "type": "text", + "content": "{ghd3079, chaelin.kim, serinyoon, ys4990, chacorp, junyongnoh}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 257, + 297, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 297, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 297, + 509 + ], + "type": "text", + "content": "Text-driven motion generation has advanced significantly with the rise of denoising diffusion models. However, previous methods often oversimplify representations for the skeletal joints, temporal frames, and textual words, limiting their ability to fully capture the information within each modality and their interactions. Moreover, when using pretrained models for downstream tasks, such as editing, they typically require additional efforts, including manual interventions, optimization, or fine-tuning. In this paper, we introduce a skeleton-aware latent diffusion (SALAD), a model that explicitly captures the intricate inter-relationships between joints, frames, and words. Furthermore, by leveraging cross-attention maps produced during the generation process, we enable attention-based zero-shot text-driven motion editing using a pre-trained SALAD model, requiring no additional user input beyond text prompts. Our approach significantly outperforms previous methods in terms of text-motion alignment without compromising generation quality, and demonstrates practical versatility by providing diverse editing capabilities beyond generation. Code is available at project page." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 523, + 135, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 135, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 135, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "content": "Character animation is a crucial component in various computer graphics and vision applications, including games, films, and interactive media. Traditional methods, such as keyframing and motion capture, typically require extensive manual effort, which are time-consuming and expensive. Recently, deep generative models have been introduced to mitigate these challenges. In particular, diffusion models have shown promising results in text-to-motion generation, enabling intuitive and efficient animation workflows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Despite these advancements, fully capturing the intricate relationships among frames, body parts, and textual descriptions remains a complex challenge in text-driven motion generation. Previous approaches typically represent a pose as a single vector and primarily focus on temporal re" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "type": "text", + "content": "relationships between poses across frames, while neglecting spatial interactions between skeletal joints. Furthermore, these models often compress a sentence into a single vector for conditioning, which can overlook the important nuances of word-level variations. This over-simplification of interactions among skeletal, temporal, and textual components may lead to missing details in the generated results, underscoring the need for models that faithfully capture these complex dependencies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 544 + ], + "type": "text", + "content": "On the other hand, learning meaningful representations is a key factor to enable zero-shot downstream tasks with pre-trained models. For example, zero-shot text-driven image and video editing methods [3, 7, 13, 15, 25, 35, 43] leverage attention modulation from pre-trained models for intuitive manipulation. Unfortunately, motion generation models typically lack such interpretable and manipulatable intermediate representations. This limitation stems from the over-simplification of motion and text features mentioned above, which restricts rich interactions between them. Although some methods exploit pre-trained models with manual masks, optimization, or fine-tuning [21-23, 42, 45, 47], these methods require additional effort, time, and cost to achieve desired results. Therefore, developing interpretable representations during motion synthesis could improve versatility and flexibility for downstream tasks beyond generation, such as zero-shot motion editing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we propose a skeleton-aware latent diffusion, which we call SALAD, for text-driven motion generation within a skeletotemporally structured latent space. We first train a variational autoencoder (VAE) [24] to construct a motion latent space that decouples spatial and temporal dimensions. To this end, we employ skeletotemporal convolution layers to facilitate information exchange between adjacent joints and frames. We also employ skeletotemporal pooling layers to produce a compact representation by reducing the dimensionality, which can reduce the computational complexity during the sampling process of the diffusion model. Subsequently, we train a diffusion model within this latent space to generate motion features conditioned on text. To train the denoiser, we leverage both" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7158" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "skeletal and temporal attention blocks, which can effectively enable skeletotemporal coherence. Additionally, we employ cross-attention to capture fine-grained interactions between individual words and the skeletotemporal units of the motion latents." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 276 + ], + "type": "text", + "content": "Beyond motion generation, we introduce an attention-based zero-shot, text-driven motion editing method that leverages the pre-trained SALAD model, requiring no additional optimization or fine-tuning. We demonstrate that the intermediate cross-attention maps of SALAD capture the relationship between text and motion features. By modulating these cross-attention maps, we enable text-driven motion editing that allows users to edit generated motion through text input alone in a training-free manner. We also propose a novel attention modulation method for motion-specific tasks, demonstrating the practical potential of our approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 277, + 251, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 277, + 251, + 289 + ], + "spans": [ + { + "bbox": [ + 67, + 277, + 251, + 289 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 296, + 397 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "text", + "content": "- We propose SALAD, a novel skeleton-aware latent diffusion model for text-driven motion generation within a skeletotemporally structured latent space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 326, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 361 + ], + "type": "text", + "content": "- We interpret the intermediate representations in the generation process, allowing for a clear understanding of the relationship between text inputs and generated motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "text", + "content": "- We present an attention-based zero-shot text-driven motion editing method using cross-attention modulation in generative models." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 411, + 142, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 142, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 142, + 423 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 431, + 224, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 224, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 224, + 443 + ], + "type": "text", + "content": "2.1. Text-driven Motion Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 451, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 713 + ], + "type": "text", + "content": "With the rise of deep neural networks, generative models for human motion synthesis based on natural language descriptions have been extensively explored. Early studies focused on generating motions from action labels, such as throw and kick [9, 32]. More advanced methods projected both text and motion features into a shared latent space, enabling the generation of fixed motion sequences from identical input text [2]. To increase the diversity of generated results, T2M [10] employed a temporal VAE that can generate varied motion sequences that align with the given text input. Subsequent studies leveraged large pretrained text encoders, such as CLIP [36], to improve textual understanding for motion generation [4, 33, 41]. Autoregressive models that behave like language models have also been proposed, which first encode motion features into discrete tokens through vector quantization [44] and then auto-regressively predict the next motion token given the text prompt [11, 12, 20, 49, 52-54]. Recently, denoising diffusion models [17, 39] have become the dominant framework for generative models, prompting text-driven motion generation methods to adopt diffusion models as their backbone [8, 23, 27, 42, 46, 50, 51]. In this study, we build" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "our framework on the diffusion model formulation and enhance the performance of text-driven motion generation through explicit modeling of the relationships among joints, frames, and words. Moreover, we introduce an attention-based zero-shot text-driven editing method, leveraging our pre-trained SALAD model as a generator, demonstrating a novel exploration beyond mere generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 165, + 535, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 535, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 535, + 177 + ], + "type": "text", + "content": "2.2. Skeleton-aware Processing of Motion Data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 183, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 447 + ], + "type": "text", + "content": "Because motion data consists of two separate skeletal and temporal dimensions, some methods have been developed to respect each dimension when processing the motion data. Yan et al. [48] presented spatial-temporal graph convolution networks that can learn both the spatial and temporal patterns of dynamic skeletons. Skeleton-aware networks [1] employed skeleto-temporal convolution networks and topology-preserving skeletal pooling layers to construct a shared motion latent space, enabling deep motion retargeting across homeomorphic skeletons. Motion style transfer has also been enabled in a part-wise manner by employing skeleto-temporal convolution modules [19, 30]. In the realm of text-to-motion generation, ParCo [54] and AttT2M [53] utilized the skeleto-temporal network architecture during the motion quantization process. In this work, we adopt a similar approach by encoding motion features using skeleto-temporal convolution networks and pooling layers to derive an effective skeleton-aware motion latent space with compact dimensionality. Additionally, we facilitate the interaction between this skeleton-aware motion latent space and textual features during the text-conditional generation process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 526, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 526, + 468 + ], + "type": "text", + "content": "2.3. Zero-shot Editing with Diffusion Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "By leveraging the generative capabilities of diffusion models, zero-shot editing methods have been proposed. SDEdit [28] first adds noise to the input data for a few diffusion timesteps, then denoises the noisy data with an editing condition to achieve high-fidelity zero-shot editing by leveraging a pre-trained diffusion model. Prompt-to-Prompt [15] verified that cross-attention maps establish connections between word tokens and the spatial layout of the generated image. In addition, it introduced attention modulation methods that can be utilized during the generation process, enabling the generation of high-fidelity editing results in a zero-shot manner. Null-text inversion [29] extended this method to real-images by improving the diffusion inversion method. Similarly, image and video editing methods via attention modulation have been extensively studied [3, 7, 13, 25, 35, 43]. Motion diffusion models often employ joint masks [23, 42] or optimization [21, 22, 47] to enable motion editing using pre-trained diffusion models, although they require user intervention or additional computation for the desired goal. On the other hand, CoMo [18]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7159" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 555, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 555, + 171 + ], + "type": "image", + "image_path": "45f5cbd20cfd69199d653edb9d49ca94a573eba579db4e95c3cffc35baec46d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 555, + 198 + ], + "type": "text", + "content": "Figure 1. Architecture of the skeletotemporal VAE network. The encoder maps motion features into a skeletotemporal latent space, and the decoder restores the skeletotemporal latent variables into motion features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 203, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 203, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 296, + 312 + ], + "type": "text", + "content": "achieved zero-shot text-driven motion editing using large language models to manipulate semantic pose tokens. We demonstrate that the intermediate cross-attention maps of our SALAD model can capture the relationship between motion and text, similarly to image diffusion models. Furthermore, by modulating these cross-attention maps, we introduce an attention-based zero-shot text-driven motion editing method that eliminates the need for manual masks, fine-tuning, or optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 111, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 111, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 111, + 333 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "text", + "content": "Our goal is to generate a motion sequence " + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{1:N}" + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "text", + "content": " conditioned on a text prompt " + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 342, + 296, + 485 + ], + "type": "text", + "content": " denotes the number of frames. We first construct a skeleton-aware motion latent representation that captures both skeletal and temporal dynamics of a motion sequence (Sec. 3.1). Using this motion representation, we train a diffusion model by modeling the complex interactions between skeletal joints, temporal frames, and textual descriptions (Sec. 3.2). Finally, we extend our model for zero-shot text-driven motion editing by modulating cross-attention maps between text and motion (Sec. 3.3). Additional details on the network architectures will be featured in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 175, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 175, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 175, + 503 + ], + "type": "text", + "content": "3.1. Skeleton-aware VAE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "To derive skeleton-aware motion features, we train a VAE composed of skeletotemporal convolution networks and pooling layers [1, 48], as illustrated in Fig. 1. The core module is the skeletotemporal convolution (STConv), which decouples joint and frame dimensions while facilitating information exchange between adjacent components within each dimension. While the skeletotemporal separation of motion sequences allows for sophisticated modeling of motion data, it introduces an additional dimension for skeletal joints, unlike previous text-to-motion generation methods that treat a pose as a single vector. This expanded data space increases the risk of curse of dimensionality and leads to longer computation times for diffusion models, which rely on solving stochastic differential equations within this higher-dimensional data space. To address this issue, we apply skeletotemporal pooling (STPool) layers in the encoder to reduce the latent space dimensionality, effectively" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "type": "text", + "content": "compressing the representation. In the decoder, skeletotemporal unpooling (STUnpool) layers are used to reconstruct the compressed latent features back to raw motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": "Encoder. We separate each pose vector of " + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{1:N}" + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": " into jointwise features and propagate them through joint-wise multilayer perceptron (MLP) layers, resulting in " + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{h} \\in \\mathbb{R}^{N \\times J \\times D}" + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": " denotes the number of joints and " + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": " denotes the latent dimension. Subsequently, we apply STConv to " + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": " to facilitate information exchange between adjacent frames and joints:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 327, + 555, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 327, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 328, + 327, + 555, + 340 + ], + "type": "interline_equation", + "content": "\\operatorname {S T C o n v} (\\mathbf {h}) := \\operatorname {S k e l C o n v} (\\mathbf {h}) + \\operatorname {T e m p C o n v} (\\mathbf {h}), \\tag {1}", + "image_path": "79694de86c2b4a21ece72603acaf9b3da5cfc4ffc100f4fab01e625e1741b78b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 345, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 380 + ], + "type": "text", + "content": "where SkelConv is a graph convolution network over the joint dimension, while TempConv is a 1D convolution network over the frame dimension." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 416 + ], + "type": "text", + "content": "We apply the STPool operator to the output of the STConv layers to reduce its dimensionality by pooling over joints and frames:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 421, + 553, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 421, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 348, + 421, + 553, + 434 + ], + "type": "interline_equation", + "content": "\\operatorname {S T P o o l} (\\mathbf {h}) := \\operatorname {T e m p P o o l} (\\operatorname {S k e l P o o l} (\\mathbf {h})), \\tag {2}", + "image_path": "7bea8ecdae0acc8bc95d2ca92aafd961201994ef9b8c1b899a119531feeba17f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": "where SkelPool represents the pooling layer over the joint dimension while TempPool represents the pooling layer over the temporal dimension. Specifically, SkelPool aggregates adjacent joints while preserving its skeletal topology, resulting in fewer joints with a homeomorphic structure. In contrast, TempPool performs 1D pooling over frames. These pooling operations are commutative because they operate within independent dimensions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "text", + "content": "Consequently, we obtain the skeleto-temporally compressed latent vector " + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{N' \\times J' \\times D}" + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "inline_equation", + "content": "N' < N" + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "inline_equation", + "content": "J' < J" + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "text", + "content": ". This representation effectively encodes the skeleton-aware dynamics of the motion sequence within a lower dimension. Notably, we maintain 7 atomic joints for this motion latent: root, spine, head, both arms, and both legs " + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "inline_equation", + "content": "(J' = 7)" + }, + { + "bbox": [ + 313, + 534, + 554, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "content": "Decoder. To reconstruct " + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "content": " into raw motion features, we employ STConv and STUnpool layers in the decoder. The decoder mirrors the encoder, facilitating information exchange between adjacent joints and frames using STConv layers and increasing the skeleto-temporal resolution using STUnpool layers. Finally, the joint-wise features are processed by their respective joint-wise MLP layers, resulting in the reconstructed motion feature " + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_{1:N}" + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7160" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 272, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 272, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 272, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 272, + 178 + ], + "type": "image", + "image_path": "723af0545820b071a09c43637c46d70284980d660693d8582d64bc1969f90f87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "type": "text", + "content": "Figure 2. (Left) Overall network architecture of the denoiser. (Right) The architecture of each transformer block consisting of TempAttn, SkelAttn, CrossAttn, and FFN, along with the FiLM following each module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 280, + 68, + 549, + 178 + ], + "blocks": [ + { + "bbox": [ + 280, + 68, + 549, + 178 + ], + "lines": [ + { + "bbox": [ + 280, + 68, + 549, + 178 + ], + "spans": [ + { + "bbox": [ + 280, + 68, + 549, + 178 + ], + "type": "image", + "image_path": "9591fe8ed372fd2599f635b889b8e84cdbf6b5fa236dfaf626402a8714601f38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 295, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 224 + ], + "type": "text", + "content": "Training. The VAE is trained with the following objectives:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 233, + 295, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 233, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 80, + 233, + 295, + 246 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V A E}} = \\mathcal {L} _ {\\mathbf {m}} + \\lambda_ {\\text {p o s}} \\mathcal {L} _ {\\text {p o s}} + \\lambda_ {\\text {v e l}} \\mathcal {L} _ {\\text {v e l}} + \\lambda_ {\\mathrm {k l}} \\mathcal {L} _ {\\mathrm {k l}}, \\tag {3}", + "image_path": "d62335e107c600c347a6f865309fc185226e257361a7b50129ccb463d30ef352.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathbf{m}},\\mathcal{L}_{\\mathrm{pos}},\\mathcal{L}_{\\mathrm{vel}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "text", + "content": " are L1 reconstruction loss terms of the motion features, joint positions, and joint velocities, while " + }, + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kl}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "text", + "content": " is the Kullback-Leibler (KL) divergence regularization that encourages a structured latent space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 309, + 194, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 194, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 194, + 320 + ], + "type": "text", + "content": "3.2. Skeleton-aware Denoiser" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "content": "Network Architecture. Given the skeleton-aware motion latent vector " + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{N' \\times J' \\times D}" + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "content": ", we train a diffusion model that denoises the noised latent vector conditioned on text prompt " + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "content": ". As shown in Fig. 2, our denoiser consists of a positional embedding, MLP-based encoder and decoder, and " + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "content": " transformer-based layers. Each transformer layer is composed of a temporal attention (TempAttn), skeletal attention (SkelAttn), cross-attention (CrossAttn), and feedforward network (FFN). Each component involves a residual connection [14], layer normalization (LN) [6], and a feature-wise linear modulation (FiLM) [31] that modulates intermediate features based on a diffusion timestep. We also incorporate skip connection [37] within the stack of transformer layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "text", + "content": "The input motion latent at diffusion timestep " + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "text", + "content": " given to the " + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "text", + "content": "-th transformer layer, denoted as " + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^l" + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "text", + "content": ", is processed by three consecutive attention blocks: temporal, skeletal, and cross-attention layers. The purpose of this architecture design is to model frame-wise interactions through TempAttn, joint-wise interactions through SkelAttn, and motion-text interactions through CrossAttn. As a result, " + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^l" + }, + { + "bbox": [ + 55, + 495, + 296, + 590 + ], + "type": "text", + "content": " is consecutively updated at each layer as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 597, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 597, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 91, + 597, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} ^ {l} \\gets \\mathbf {z} _ {t} ^ {l} + \\mathrm {F i L M} (\\mathrm {T e m p A t t n} (\\mathrm {L N} (\\mathbf {z} _ {t} ^ {l}))), \\qquad (4)", + "image_path": "b34eea0d45262205bd562572e60fbd948570844358b381685e7dd2f6cb207b16.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 95, + 614, + 295, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 614, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 95, + 614, + 295, + 629 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} ^ {l} \\leftarrow \\mathbf {z} _ {t} ^ {l} + \\operatorname {F i L M} (\\operatorname {S k e l A t t n} (\\operatorname {L N} (\\mathbf {z} _ {t} ^ {l}))), \\tag {5}", + "image_path": "a0f16f6e0b29e2a5d0eec691d7ab78f61978321bea7386ead331782faa5d5933.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 631, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 631, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 66, + 631, + 295, + 646 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} ^ {l} \\leftarrow \\mathbf {z} _ {t} ^ {l} + \\mathrm {F i L M} (\\operatorname {C r o s s A t t n} (\\operatorname {L N} (\\mathbf {z} _ {t} ^ {l}), \\operatorname {C L I P} (c))). \\quad (6)", + "image_path": "a107704b3141241d7e5f7fa87024bf6d3f71b1cd02872f6b0deaae3199e8ae01.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": "For the CLIP text encoder [36] used in cross-attention, we use a pre-trained model and freeze its weights during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Diffusion Parametrization. For diffusion parametrization, our model predicts the diffusion velocity [38], denoted as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 213, + 372, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 213, + 372, + 223 + ], + "spans": [ + { + "bbox": [ + 314, + 213, + 372, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 314, + 213, + 372, + 223 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 399, + 225, + 553, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 225, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 399, + 225, + 553, + 236 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {t} = \\alpha_ {t} \\boldsymbol {\\epsilon} - \\sigma_ {t} \\mathbf {x}, \\tag {7}", + "image_path": "d2365283d22363ae127398a34334c7fd35ee0db28bfa67c48640e2a01e06a8ad.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": " represent the noise and sample, respectively, while " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": " are noise scheduling parameters at diffusion timestep " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": ". While " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": "-prediction are widely used in diffusion models, " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": "-prediction combines both parametrizations, implicitly balancing the information contributed by each component. This improves the stability of generation, especially at high noise levels in which the relationship between " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 242, + 553, + 337 + ], + "type": "text", + "content": " becomes disrupted." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "text", + "content": "Training and Inference. The denoiser is trained to predict the diffusion velocity " + }, + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "text", + "content": " at each diffusion timestep " + }, + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "text", + "content": ". The training objective for the denoiser is defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 381, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 381, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 383, + 381, + 553, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e n o i s e r}} = \\left\\| \\hat {\\mathbf {v}} _ {t} - \\mathbf {v} _ {t} \\right\\| _ {2} ^ {2}, \\tag {8}", + "image_path": "01e6ada822f1a0ab183daf072277157afd896222b518e203e1ccb802915b376d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_t" + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "text", + "content": " represents the predicted diffusion velocity by the denoiser. During training, we randomly drop the text condition with a probability of " + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{uncond}}" + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "text", + "content": ", enabling unconditional generation when " + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "text", + "content": " is a null text, which is represented as " + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 313, + 403, + 553, + 451 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 452, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 553, + 475 + ], + "type": "text", + "content": "To the trained denoiser, we apply classifier-free guidance (CFG) [16] for text-conditioned generation:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 484, + 555, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 484, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 314, + 484, + 555, + 508 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {v}} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, c\\right) := \\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\varnothing\\right) + w \\left(\\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, c\\right) - \\mathbf {v} _ {\\theta} \\left(\\mathbf {z} _ {t}, t, \\varnothing\\right)\\right). \\tag {9}", + "image_path": "978ac00bd433325b5c671e3e6e498fe18cdb8598d901f3c7576683e8d98a8a0e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "content": " denotes the pre-trained denoiser, " + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "content": " is the CFG weight, and " + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 554, + 581 + ], + "type": "text", + "content": " is the modified velocity used during the denoising process. Additionally, we employ the DDIM sampling formulation [40] during inference, which effectively reduces the number of sampling steps without compromising generation quality." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 587, + 514, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 514, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 514, + 601 + ], + "type": "text", + "content": "3.3. Zero-shot Text-driven Motion Editing" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "In this section, we describe the attention modulation methods that enable zero-shot text-driven motion editing with a pre-trained SALAD model. We introduce four distinct modulation strategies applied to the text-motion cross attention maps as shown in Fig. 3: word swap, prompt refinement, attention re-weighting, and attention mirroring. The first three are adapted from the Prompt-to-Prompt framework [15], while the attention mirroring serves as a validation method to assess whether the cross-attention map" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7161" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "type": "image", + "image_path": "0084cfd6109f28f04c9bc385f2fd4b8d3f30a6917842177aa51fc34a5ae1f5cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 214 + ], + "type": "text", + "content": "Figure 3. Attention modulation methods applied to the cross-attention maps to enable text-driven motion editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 221, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 221, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 295, + 437 + ], + "type": "text", + "content": "contributes to the activation of spatio-temporal dynamics in the motion generation process. Word swap exchanges attention maps between a source and target prompt, enabling the model to leverage the attention context of the source text when generating output for the target text. Prompt refinement enhances the original attention map by integrating additional attention maps for appended word tokens, providing richer semantic information compared to the base prompt. Attention re-weighting adjusts the attention values assigned to specific user-selected words, either amplifying or diminishing their influence on the generated motion. Lastly, attention mirroring swaps attention values between symmetrical body parts, such as the left and right arms, to produce mirrored motions. These methods collectively empower the generative capacity of our model to produce dynamic and contextually relevant motion edits based solely on text prompts. For more details on these modulation functions, please refer to the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 136, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 136, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 136, + 462 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 625 + ], + "type": "text", + "content": "We conducted evaluations of our method on two widely used motion-language benchmarks: HumanML3D [10] and KIT-ML [34]. HumanML3D dataset consists of 14,616 motion sequences with a variety of human actions, such as locomotions, sports, and acrobatics, along with 44,970 text descriptions in total. KIT-ML dataset consists of 3,911 motion sequences along with 6,278 text descriptions. Both datasets are augmented by mirroring, and divided into training, testing, and validation sets with the ratio of 0.8, 0.15, and 0.05. Evaluation results are reported in the following sections. Please refer to the supplementary material for additional quantitative experiment results with details, and the supplementary video for the animated results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 188, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 188, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 188, + 647 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Our method was executed on a single NVIDIA V100 GPU. We used the AdamW optimizer [26], and the VAE and denoiser were trained for 50 and 500 epochs, respectively. The VAE and denoiser were trained with a batch size of 64 for HumanML3D, which required 35 and 17 hours, re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "spectively. For the KIT-ML dataset, we used a batch size of 16, with training times of 20 hours for the VAE and 7 hours for the denoiser. We trained the denoiser with 1000 diffusion steps, employing 50 steps for DDIM sampling during inference. For the CFG weight, we set " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "w = 7.5" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " unless mentioned otherwise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 152, + 449, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 449, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 449, + 165 + ], + "type": "text", + "content": "4.2.Quantitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "content": "For quantitative evaluations, we adopted metrics from T2M [10]. First, " + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "inline_equation", + "content": "FID" + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "content": " is used to evaluate the quality of generated motions by measuring the difference between the feature distributions of generated motions and those of real motions. Additionally, " + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "content": "-precision and " + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "inline_equation", + "content": "MM-Dist" + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "content": " are used to evaluate the semantic alignment between the input text and the generated motions. " + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "inline_equation", + "content": "Diversity" + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "content": ", which indicates the variance in generated motions, and " + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "inline_equation", + "content": "MultiModality" + }, + { + "bbox": [ + 313, + 171, + 554, + 315 + ], + "type": "text", + "content": ", which assesses the diversity of motions generated from the same text description, are also adopted as secondary metrics. They are considered less important than the generation quality and alignment with the input text." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 471 + ], + "type": "text", + "content": "Following the practice of T2M [10], each experiment was repeated 20 times on the test sets of HumanML3D and KIT-ML datasets, and the mean with a " + }, + { + "bbox": [ + 313, + 316, + 554, + 471 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 316, + 554, + 471 + ], + "type": "text", + "content": " confidence interval is reported in Tab. 1. Notably, SALAD consistently achieved superior text-motion alignment across both datasets, regardless of the diffusion parametrization applied during denoiser training. Moreover, SALAD maintained high generation quality, as indicated by the best or second-best FID scores among diffusion-based methods. These results confirm that explicitly modeling interactions between skeletal joints, temporal frames, and textual words is effective in improving text-to-motion generation quality and fidelity to the input text." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 443, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 443, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 443, + 491 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "content": "We compared the qualitative results of SALAD with those generated by MDM [42] and MoMask [12], and the comparison results are shown in Fig. 4. Both MDM and MoMask reflected parts of the input text prompts, but they failed to handle all the descriptions in the text. For example, as shown in the top of Fig. 4-(a), MDM generated a character walking forward, turning, and walking back, but the locomotion style was not marching, despite the input prompt specifying \"marches\". Similarly, in Fig. 4-(b), MoMask accurately reflected the arm movements of raising above the head and throwing them down, but it missed the instruction to turn to the right. In case of Fig. 4-(c), both methods failed to perfectly reflect the input text despite its short length. In contrast, SALAD consistently incorporated all the textual descriptions into the generated motions, regardless of the length and complexity of the text prompts. This demonstrates its effectiveness in generating motions that are highly aligned with the input text." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7162" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 201 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 504, + 201 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 201 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 201 + ], + "type": "table", + "html": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.511±.0030.703±.0030.797±.0020.002±.0002.974±.0089.503±.065-
T2M [10]0.457±.0020.639±.0030.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
AttT2M [53]0.499±.0030.690±.0020.786±.0020.112±.0063.038±.0079.700±.0902.452±.051
ParCo [54]0.515±.0030.706±.0030.801±.0020.109±.0052.927±.0089.576±.0881.382±.060
MoMask [12]0.521±.0020.713±.0020.807±.0020.045±.0022.958±.008-1.241±.040
MDM [42]0.320±.0050.498±.0040.611±.0070.544±.0045.566±.0279.559±.0862.799±.072
MLD [8]0.481±.0030.673±.0030.772±.0020.473±.0133.196±.0109.724±.0822.413±.079
MotionDiffuse [51]0.491±.0010.681±.0010.782±.0010.630±.0013.113±.0019.410±.0491.553±.042
ReMoDiffuse [50]0.510±.0050.698±.0060.795±.0040.103±.0042.974±.0169.018±.0751.795±.043
SALAD (Ours)0.581±.0030.769±.0030.857±.0020.076±.0022.649±.0099.696±.0961.751±.062
", + "image_path": "e8f97a58fad51e2e4edabed5feb9a8e248b1343b72a047477a7e5118f1f5d7eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 106, + 203, + 503, + 334 + ], + "blocks": [ + { + "bbox": [ + 106, + 203, + 503, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 503, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 503, + 334 + ], + "type": "table", + "html": "
MethodsR-Precision ↑FID ↓MM-Dist ↓Diversity →MultiModality ↑
Top-1Top-2Top-3
Real motion0.424±.0050.649±.0060.779±.0060.031±.0042.788±.01211.08±.097-
T2M [10]0.370±.0050.569±.0070.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
AttT2M [53]0.413±.0060.632±.0060.751±.0060.870±.0393.039±.02110.96±.1232.281±.047
ParCo [54]0.430±.0040.649±.0070.772±.0060.453±.0272.820±.02810.95±.0941.245±.022
MoMask [12]0.433±.0070.656±.0050.781±.0050.204±.0112.779±.022-1.131±.043
MDM [42]0.164±.0040.291±.0040.396±.0040.497±.0219.191±.02210.847±.1091.907±.2T4
MLD [8]0.390±.0080.609±.0080.734±.0070.404±.0273.204±.02710.80±.1172.192±.071
MotionDiffuse [51]0.417±.0040.621±.0040.739±.0041.954±.0622.958±.00511.10±.1430.730±.013
ReMoDiffuse [50]0.427±.0140.641±.0040.765±.0550.155±.0062.814±.01210.80±.1051.239±.028
SALAD (Ours)0.477±.0060.711±.0050.828±.0050.296±.0122.585±.01611.097±.0951.004±.040
", + "image_path": "8b821bfd3b3f60d21ba8a056b61ff788cd17fc7858f8adca181349c8f2ca1b23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation results on the test sets of HumanML3D (top) and KIT-ML (bottom). " + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "text", + "content": " denote that higher and lower values are better, respectively, while " + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "text", + "content": " denotes that the values closer to the real motion are better. Methods above the dotted line are auto-regressive models based on VAE or VQ-VAE, while those below are diffusion-based generative models. Red and blue colors indicate the best and the second best results, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 384, + 541, + 540 + ], + "blocks": [ + { + "bbox": [ + 67, + 384, + 541, + 540 + ], + "lines": [ + { + "bbox": [ + 67, + 384, + 541, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 384, + 541, + 540 + ], + "type": "image", + "image_path": "5cada3a7308061c93108297ac11eef69b63565e13fd4d8ff16ac892e30e1af71.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 543, + 539, + 555 + ], + "lines": [ + { + "bbox": [ + 69, + 543, + 539, + 555 + ], + "spans": [ + { + "bbox": [ + 69, + 543, + 539, + 555 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of results generated by various methods, including MDM [42], MoMask [12], and our approach." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 560, + 149, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 149, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 149, + 574 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Effectiveness of Skeleton-aware Latent Space. To prove the effectiveness of using skeleton-aware latent space, we compared the performance of different VAE structures employed in MoMask [12] and ParCo [54] with ours. Specifically, MoMask employs a VQ-VAE that compresses each pose into a single vector, while ParCo employs a part-aware VQ-VAE that processes different body parts independently. Because the role of VAE is to compress the motion features into latent variables, we compared FID and mean per joint position error (MPJPE) that assess the quality and accuracy of the reconstructed motion features, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 561, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 658 + ], + "type": "text", + "content": "As shown in Tab. 2, our skeleton-aware VAE outperformed baselines with a significant margin for both metrics. Notably, our method required much fewer number of trainable parameters than the others. This indicates that skeleton-aware latent space that decouples both dimensions and facilitates information exchange between adjacent components is effective in constructing meaningful motion latent space, which is a crucial prior to train the generator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, we conducted an ablation study that validates the effectiveness of the skeleto-temporal latent (ST-Latent) and cross-attention, which are two core components of SALAD, on the text-to-motion generation perfor" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7163" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 125 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 125 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 125 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 125 + ], + "type": "table", + "html": "
MethodsNum of ParamsFID ↓MPJPE ↓
MoMask [12]19.44M0.020±.0000.030±.000
ParCo [54]6.35M0.021±.0000.108±.000
SALAD (Ours)0.16M0.003±.0000.016±.000
", + "image_path": "0c81f77e5ae7716502a15a437e21ad12f3af624d248d41c2dd95d88070f515ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 178, + 293, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 129, + 295, + 172 + ], + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 172 + ], + "type": "text", + "content": "Table 2. Quantitative results on the quality and accuracy of reconstructed motion features of VAE models from different methods, along with the number of trainable parameters, measured on the test set of HumanML3D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 178, + 293, + 251 + ], + "lines": [ + { + "bbox": [ + 58, + 178, + 293, + 251 + ], + "spans": [ + { + "bbox": [ + 58, + 178, + 293, + 251 + ], + "type": "table", + "html": "
MethodR-Precision (Top-3) ↑FID ↓
Ours (Full model)0.857±.0020.076±.002
w/o ST-Latent0.816±.0020.433±.006
w/o CrossAttn0.778±.0020.274±.007
w/o both0.752±.0020.345±.007
", + "image_path": "cc13b2277dc3ac609c64a0161b58416dcde0bed70d532194600c095e3fcb67b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 251, + 269, + 262 + ], + "lines": [ + { + "bbox": [ + 82, + 251, + 269, + 262 + ], + "spans": [ + { + "bbox": [ + 82, + 251, + 269, + 262 + ], + "type": "text", + "content": "Table 3. Ablation studies on the VAE and denoiser." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 268, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 434 + ], + "type": "text", + "content": "mance. For this experiment, we evaluated three variations: (i) replacing ST-Latent by training a conventional VAE that treats each pose as a single vector, (ii) replacing cross-attention with self-attention by concatenating the sentence-level CLIP feature to the motion features, and (iii) combining (i) and (ii). As shown in Tab. 3, alteration of ST-Latent and cross-attention substantially degraded generation quality and text-motion alignment, as reflected in the higher FID and lower R-precision, respectively. This demonstrates the importance of rich information exchange between text and motion by cross-attention within the skeletotemporally structured latent space. This indicates that the skeletotemporally disentangled motion latent facilitates accurate text-motion interaction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "content": "CFG Weights. To analyze the influence of CFG weights on generation results, we plotted FID and R-precision metric values in Fig. 5. The figure shows that low CFG weights led to suboptimal generation quality and text-motion alignment. The performance of SALAD improved across both metrics as the weight values increased, but excessively high weight values resulted in a decreased performance for both metrics again. Based on these findings, we determined that a value of " + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "inline_equation", + "content": "w = 7.5" + }, + { + "bbox": [ + 54, + 435, + 295, + 556 + ], + "type": "text", + "content": " produces the best balance between generation quality and alignment with the text prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 255, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 255, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 255, + 576 + ], + "type": "text", + "content": "4.5. Zero-shot Text-driven Motion Editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": "To demonstrate the capability of SALAD in capturing relationships between skeleto-temporal latents and text prompts, we visualized the cross-attention maps for each word in Fig. 6. The figure reveals that high attention values were assigned to proper frames and body parts that semantically correspond to each word. For example, \"jumps forward\" was activated for the entire body at early frames, \"walks forward\" focused on the root and legs at later frames, and \"waving arms\" was activated for the arms throughout the entire frames. These results indicate that cross-attention maps produced during motion generation ef" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 60, + 527, + 186 + ], + "blocks": [ + { + "bbox": [ + 340, + 60, + 527, + 186 + ], + "lines": [ + { + "bbox": [ + 340, + 60, + 527, + 186 + ], + "spans": [ + { + "bbox": [ + 340, + 60, + 527, + 186 + ], + "type": "image", + "image_path": "9ab25601c73dcc4dc025aa37c182e9e71d310c4af33b8628f4f58a7f3aab3708.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 187, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 221 + ], + "type": "text", + "content": "Figure 5. Ablation study results on different CFG weights. Lower FID scores indicate better motion quality, and higher R-precision (Top-3) scores indicate better semantic alignment." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 240, + 552, + 413 + ], + "blocks": [ + { + "bbox": [ + 338, + 227, + 531, + 236 + ], + "lines": [ + { + "bbox": [ + 338, + 227, + 531, + 236 + ], + "spans": [ + { + "bbox": [ + 338, + 227, + 531, + 236 + ], + "type": "text", + "content": "a man jumps forward and then walks forward, while waving his arms" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 316, + 240, + 552, + 413 + ], + "lines": [ + { + "bbox": [ + 316, + 240, + 552, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 552, + 413 + ], + "type": "image", + "image_path": "8a1e1bcdad2dcb9062350b0e8d40f722193758e7e4cb6a444d1cca8e32ad492c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 418, + 554, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 475 + ], + "type": "text", + "content": "Figure 6. Visualizations of cross-attention maps of SALAD. Each row corresponds to a specific body part, and each column represents temporal frames. Attention maps were computed by averaging the output attention maps across all transformer layers and all diffusion timesteps." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 553, + 508 + ], + "type": "text", + "content": "fectively capture the rich semantic relationships between text and motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": "Based on these observations, we demonstrate zero-shot text-driven motion editing results using a pre-trained SALAD model by modulating cross-attention maps, and the results are shown in Fig. 7. The edited motions align well with the given instructions while preserving the original movements, such as walking phases. Notably, the changes for editing were localized to specific frames and body parts relevant to the editing prompt, while unaffected regions remain intact, without using explicit masking. Furthermore, our SALAD model understands not only the high-level correspondence between text and motion but also its intensity and direction, which is demonstrated by different hand heights when modulating the word \"high\" in the reweighting case. These results suggest that the explicit modeling of skeletal-temporal-textual relationships in SALAD enhances its understanding of text-driven motion generation, leading to effective learning representations and fine" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7164" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 306, + 278 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 306, + 278 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 306, + 278 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 306, + 278 + ], + "type": "image", + "image_path": "5c1f4109ad5038b2d13b111cf439c198e1bb50d536b1eac80dc9b346736e39e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": "Figure 7. Text-driven motion editing results via attention modulation using a pre-trained SALAD model. Texts colored in green represent the editing instructions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 553, + 279 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 553, + 279 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 553, + 279 + ], + "type": "image", + "image_path": "98a10733bf4aaf03205648c52782d2dd771e92d8cb512ea2136c984bdfc48095.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 307, + 294, + 355 + ], + "blocks": [ + { + "bbox": [ + 58, + 307, + 294, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 307, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 307, + 294, + 355 + ], + "type": "table", + "html": "
MethodsPreservationSemantic AlignmentOverall Quality
MDM [42]3.729±.1532.758±.1923.196±.147
MotionFix [5]3.358±.1733.388±.1893.421±.166
SALAD (Ours)4.596±.0844.654±.0874.596±.083
", + "image_path": "6706a7d42f53903114f64560bb4f9a89d54360eb7421fafe461a221ae6133bf2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 359, + 294, + 370 + ], + "lines": [ + { + "bbox": [ + 55, + 359, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 294, + 370 + ], + "type": "text", + "content": "Table 4. User study results. The red color indicates the best result." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 376, + 244, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 244, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 244, + 388 + ], + "type": "text", + "content": "grained control in both generation and editing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 388, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 508 + ], + "type": "text", + "content": "To evaluate the perceptual quality of text-driven motion editing, we conducted a user study with 16 participants. We compared SALAD against MDM [42], which edits motion by masking and regenerating specified body parts, and MotionFix [5], a method specifically designed for motion editing. Participants were shown with 15 videos per method and asked to rate them based on three criteria: (i) preservation of the original motion, (ii) semantic alignment with the target text, and (iii) overall quality. Ratings were given on a 5-point Likert scale, with 5 indicating the best." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 508, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 295, + 628 + ], + "type": "text", + "content": "As shown in Tab. 4, SALAD outperformed both baselines across all three aspects by a large margin. Notably, despite not using explicit masking, SALAD achieved higher preservation scores. Furthermore, while SALAD is not specifically designed for editing, it achieved higher preferences for semantic alignment and overall quality, highlighting the effectiveness of attention-based editing. These results demonstrate that leveraging cross-attention maps enables high-quality and perceptually appealing text-driven motion editing without requiring additional training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 633, + 125, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 125, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 125, + 646 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "While SALAD produced outperforming results in text-driven motion generation compared to the existing generative models, along with the capability of zero-shot text-driven motion editing, it has several limitations. First, the numerical results for Diversity and MultiModality are lim" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 309, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 477 + ], + "type": "text", + "content": "ited, despite the strong capability of diffusion models to generate diverse outputs. This can be due to the trade-off between high-quality with faithfulness to the text and diversity. While we put the diversity as a second priority, finding a way to handle both could be an interesting approach. Additionally, SALAD is restricted to generating a single person action with limited text and motion length. Therefore, enabling multi-character interactions, crowd animations, generation from longer texts, and creating extended motion sequences could be an interesting direction for future work. Furthermore, although our method does not directly address editing real motions, integrating diffusion inversion methods could enable text-driven editing of real motion sequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 714 + ], + "type": "text", + "content": "In this paper, we presented SALAD, a novel diffusion model framework for text-driven motion generation that models the complex interactions between skeletal joints, temporal frames, and textual words. We first train a skeletotemporal VAE that effectively summarizes motion features by decoupling skeletal and temporal dimensions, and we subsequently train a skeletotemporal denoiser that enables efficient and expressive generation of motions conditioned on text input within this compact skeletotemporal latent space. Furthermore, we demonstrated the first text-driven zero-shot motion editing using a pre-trained SALAD model by modulating cross-attention maps during generation. By leveraging the outperforming generative capacity compared to previous methods, along with the interpretable learning representations during generation, we anticipate that our work can benefit future research and applications in the context of text-driven motion generation and editing." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7165" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "We appreciate the anonymous reviewers for their invaluable discussion. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS-2024-00333478)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 216 + ], + "type": "text", + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 218, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 296, + 261 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 306 + ], + "type": "text", + "content": "[3] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 352 + ], + "type": "text", + "content": "[4] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 354, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 295, + 397 + ], + "type": "text", + "content": "[5] Nikos Athanasiou, Alpár Cseke, Markos Diomataris, Michael J Black, and Gül Varol. Motionfix: Text-driven 3d human motion editing. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 430 + ], + "type": "text", + "content": "[6] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 433, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 295, + 498 + ], + "type": "text", + "content": "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 501, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 295, + 555 + ], + "type": "text", + "content": "[8] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 612 + ], + "type": "text", + "content": "[9] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "text", + "content": "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "text", + "content": "[11] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "type": "text", + "content": "[12] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[13] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 231 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 232, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 276 + ], + "type": "text", + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 278, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 278, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 278, + 553, + 300 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 303, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 553, + 335 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 337, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 337, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 337, + 553, + 392 + ], + "type": "text", + "content": "[18] Yiming Huang, Weilin Wan, Yue Yang, Chris Callison-Burch, Mark Yatskar, and Lingjie Liu. Como: Controllable motion generation through language guided pose code editing. In European Conference on Computer Vision, pages 180-196. Springer, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 427 + ], + "type": "text", + "content": "[19] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "type": "text", + "content": "[20] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 529 + ], + "type": "text", + "content": "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2151-2162, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "type": "text", + "content": "[22] Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwajanakorn, and Siyu Tang. Optimizing diffusion noise can serve as universal motion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1334-1345, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 589, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 553, + 633 + ], + "type": "text", + "content": "[23] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 657 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[25] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7166" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "type": "text", + "content": "[27] Yunhong Lou, Linchao Zhu, Yaxiong Wang, Xiaohan Wang, and Yi Yang. Diversemotion: Towards diverse human motion generation via discrete diffusion. arXiv preprint arXiv:2309.01372, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[28] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[29] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[30] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on computer graphics and interactive techniques, 4(3):1-17, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "text", + "content": "[31] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "text", + "content": "[32] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[34] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236-252, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 545 + ], + "type": "text", + "content": "[35] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 612 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 679 + ], + "type": "text", + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[39] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[41] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[42] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[43] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "text", + "content": "[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 345, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 345, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 345, + 553, + 388 + ], + "type": "text", + "content": "[45] Weilin Wan, Zhiyang Dou, Taku Komura, Wenping Wang, Dinesh Jayaraman, and Lingjie Liu. Tlcontrol: Trajectory and language control for human motion synthesis. arXiv preprint arXiv:2311.17135, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 389, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 444 + ], + "type": "text", + "content": "[46] Yin Wang, Zhiying Leng, Frederick WB Li, Shun-Cheng Wu, and Xiaohui Liang. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22035-22044, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[47] Yiming Xie, Varun Jampani, Lei Zhong, Deqing Sun, and Huaizu Jiang. Omnicontrol: Control any joint at any time for human motion generation. arXiv preprint arXiv:2310.08580, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 490, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 534 + ], + "type": "text", + "content": "[48] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 601 + ], + "type": "text", + "content": "[49] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14730-14740, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "text", + "content": "[50] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7167" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[52] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7368-7376, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[53] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 509-519, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[54] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2, 6, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7168" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_content_list.json b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a6c66a50da06e2c431aef3284da4e6415ab9367 --- /dev/null +++ b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_content_list.json @@ -0,0 +1,1417 @@ +[ + { + "type": "text", + "text": "SALOVA: Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junho Kim* Hyunjun Kim* Hosu Lee Yong Man Ro† Integrated Vision and Language Lab, KAIST, South Korea {arkimjh, kimhj709, leehosu01, ymro}@kaist.ac.kr", + "bbox": [ + 232, + 203, + 772, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite advances in Large Multi-modal Models, applying them to long and untrimmed video content remains challenging due to limitations in context length and substantial memory overhead. These constraints often lead to significant information loss and reduced relevance in the model responses. With the exponential growth of video data across web platforms, understanding long-form video is crucial for advancing generalized intelligence. In this paper, we introduce SALOVA: Segment-Augmented LOng Video Assistant, a novel video-LLM framework designed to enhance the comprehension of lengthy video content through targeted retrieval process. We address two main challenges to achieve it: (i) We present the SceneWalk dataset, a high-quality collection of 87.8K long videos, each densely captioned at the segment level to enable models to capture scene continuity and maintain rich descriptive context. (ii) We develop robust architectural designs integrating dynamic routing mechanism and spatio-temporal projector to efficiently retrieve and process relevant video segments based on user queries. Our framework mitigates the limitations of current video-LMMs by allowing for precise identification and retrieval of relevant video segments in response to queries, thereby improving the contextual relevance of the generated responses. Through extensive experiments, SALOVA demonstrates enhanced capability in processing complex long-form videos, showing significant capability to maintain contextual integrity across extended sequences.", + "bbox": [ + 89, + 323, + 485, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Large Language Models (LLMs) [22, 43, 44] have brought us one step closer to achieving Artificial General Intelligence (AGI). Next following step, the current trend is shifting toward modular systems that integrate various multi-modality, leveraging", + "bbox": [ + 89, + 787, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the exceptional generalization and reasoning capabilities of LLMs to evolve into Large Multi-modal Models (LMMs). Accordingly, users can unrestrictedly interact with the models across various modalities beyond text, expanding the scope of machine understanding and enhancing user engagement. Especially, considering the widespread adoption of long-form videos across various web platforms, the importance of understanding long, untrimmed video has become increasingly prominent in the multi-modal domain.", + "bbox": [ + 511, + 292, + 906, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "After the pioneer works [14, 37, 38] utilizing visual instruction tuning to augment vision perception into LLMs, remarkable strides [11, 17, 65] have been made in aligning cross-modal consistency—especially between vision and language domains. Albeit more recent models [29, 69] integrate various vision modalities all at once, current approaches still face significant challenges in understanding untrimmed and long-form video content. The main challenge is attributed to the limited context length of LMMs, which is an inherent structural limitation that restricts the models to process only a finite number of tokens as the input sequences. We can exemplify that LLaVA series [29, 30], when processing a video data, require 144 visual tokens per each frame, where numerical approximation is only maximum of $\\sim 56$ frames using 8K max context length LMMs, which is still limited to handle long sequence data.", + "bbox": [ + 511, + 430, + 908, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accordingly, current video-LMMs [12, 35, 39] rely on (i) sparse frame sampling to represent entire videos [26, 72], (ii) dense compression of visual tokens into a smaller size to manage the excessive number of frames [34, 41], and (iii) adaptive pooling strategies [59, 60] based on the SlowFast approach [19], all aimed at fitting the long video sequences within the limited context window of LMMs. Several studies focusing on the long video understanding task have presented memory-augmented generation [23, 53] utilizing an additional buffer to embed long-term information, or have extended the context using RoPE-based frequency extension during the training [70]. Despite of such endeavors, when handling massive video frames, previous works still confront restricted context size and significant memory overhead, which leads to substantial visual information", + "bbox": [ + 511, + 674, + 910, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author. Code is available at https://ivy-lvlm.github.io/SALOVA", + "bbox": [ + 89, + 875, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3352", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "loss. As critical events may be overlooked by the models, this hinders their ability to fully capture context changes in lengthy videos, resulting in inaccurate and irrelevant responses for the user queries.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Starting from the intuitive insight outlined below, in this paper, we propose a retrieval-driven approach for long video understanding with LLMs. Analogous to the recent Retriever-Augmented Generation (RAG) systems [28] (widely adopted in LLMs), which retrieve relevant information from external factual knowledge, humans naturally employ similar strategies when seeking specific information, efficiently locating and referring necessary materials to answer targeted questions—e.g., imagine that we are taking open-book exams or searching for a certain recipe in a cookbook. Given a long and untrimmed video, mirroring the targeted retrieval processes, we introduce a novel framework, Segment-Augmented LOng Video Assistant (SALOVA) to effectively handle the long sequence visual inputs by retrieving the relevant video segments.", + "bbox": [ + 88, + 152, + 482, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To construct our video segments retrieval framework, central challenge hinges on establishing two main components: (i) Densely captioned video data, which consists of video-caption pairs with progressively-captioned descriptions that change throughout each video used to train the model to accurately identify relevant video segments. (ii) Dynamic routing mechanism, which selects pertinent video segments for the queries, followed by being connected to LLMs. To address that, our approach is outlined as follows.", + "bbox": [ + 89, + 378, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data. (§3) Recently several video-text paired datasets [4, 6, 8, 10, 56, 62] have been released, but they are inadequate for handling long and untrimmed video data, where only partial video moments are described with limited word length as compared in Fig. 1(a). To handle such insufficiency of detailed descriptions within the videos and the short durations of both videos and texts, we introduce the SceneWalk dataset, a new high-quality video dataset with thorough captioning for each video. It includes dense and detailed descriptions for every video segment across the entire scene context. The SceneWalk dataset, sourced from long and untrimmed 87.8K YouTube videos (avg. 486 seconds each), features frequent scene transitions across a total of 11.8K hrs video duration and 1.3M massively segmented video clips. Each video segment in the dataset is provided with a detailed description (avg. 137.5 word length), generated by combining pre-trained models [54, 73] and manual curation from human.", + "bbox": [ + 91, + 518, + 482, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Architecture. (§4) Utilizing the constructed video dataset, SALOVA learns to identify relevant video segments for the given queries within each video source and then auto-regressively predicts the next token. To do so, we present two architectural designs to seamlessly incorporate the retrieved segments in an end-to-end training: Spatio-Temporal Connector and Segment Retrieval Router. By fo", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cusing on the relevant segments, our framework can perform deeper reasoning without being constrained by context length limitations. Additionally, we present FocusFast approach, which intensively analyzes the selected segments for detailed comprehension (focus pathway), while quickly accessing overall contextual information with routing tokens obtained from the entire video segments (fast pathway). The strategy ensures SALOVA to maintain comprehensive video understanding while prioritizing details where it is most needed, effectively enhancing long and untrimmed video interpretation.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments and analyses, we corroborate that competitive performance of SALOVA to the existing video-LMM models in understanding complex long-form videos. Also, our results show significant reductions in the loss of crucial visual information and a lower risk of omitting important events, demonstrating the effectiveness of our proposed method across various video benchmarks.", + "bbox": [ + 511, + 258, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution can be summarized into three-fold:", + "bbox": [ + 532, + 364, + 879, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the SceneWalk dataset, a high-quality and densely-captioned video dataset with detailed segment-level descriptions from 87.8K long and untrimmed video sources. The proposed dataset provides rich context and scene continuity, enabling effective training for long-form video understanding.", + "- We propose Segment-Augmented LOng Video Assistant (SALOVA), a novel video-LMM framework designed to enhance long video comprehension by targeting relevant video segments in lengthy videos, optimizing the model's focus on essential segment targets for the given queries.", + "- Through extensive evaluation, we validate that SALOVA improves overall long video understanding capabilities by effectively integrating relevant video segments, thus optimizing to handle long and untrimmed video content." + ], + "bbox": [ + 513, + 381, + 903, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 625, + 653, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Multi-modal Models", + "text_level": 1, + "bbox": [ + 513, + 650, + 756, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After the emergence of LLMs [5, 55], which can actively interact with users through back-and-forth conversations, as a next leap, various research efforts [2, 25, 31] integrate different modalities into the LLMs, utilizing their core reasoning and zero-shot capabilities. Building on the open-sourced models [13, 55], seminal works [14, 38, 65] have bridged image and text modality under the visual instruction tuning and presented multi-modal assistant models that possess visual perception and QA capabilities. Since then, numerous research studies have been introduced to (i) enhance vision understanding with advanced architectures [17] or higher resolutions [30, 37], (ii) implement more sophisticated alignment layers [7, 42] between modalities, and (iii) train the models with more high-quality data and larger model parameters.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3353", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/115411666638185eab77d20c4648d4db87a8318b2b9f8c5453a123aed7b53ac5.jpg", + "image_caption": [ + "(a) Video-Text Dataset Comparison" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 346, + 248 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2aca787b5795f792bf26420366df55c4fd7b063a91006caa4fd7148418b58bdf.jpg", + "image_caption": [ + "(b) SceneWalk Statistics" + ], + "image_footnote": [], + "bbox": [ + 357, + 87, + 552, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a4c4cbb8373c76c5c6851d3f3659752931644de12660eb98c28b8f161aba9adc.jpg", + "image_caption": [ + "(c) Overall Pipeline for Data Collection", + "Figure 1: The overview of the SceneWalk dataset includes (a) dataset comparison, (b) detailed statistics, and (c) the annotation pipeline for description and score collection. Note that the scale of circles in Fig. 1(a) indicates the data size, and the color distribution in Fig. 1(b) denotes the video duration in each video category—brighter colors correspond to shorter video durations. Further details about the dataset are provided in Appendix A." + ], + "image_footnote": [], + "bbox": [ + 563, + 98, + 905, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent focus has shifted towards more unified modality processing following the release of omnivorous models [46]. Some recent omni-versions of LMMs [29, 64] can handle combinatorial subsets from various modality sources, such as images, videos, audio, speech, and depth. However, the current LMMs for video [32, 35, 41] still lack of capturing the necessary details to effectively process video information due to their sparse frame sampling strategy. While such approach is seemingly adequate for relatively shorter videos, it may fail to capture comprehensive spatio-temporal information, potentially compromising the accuracy of model responses to user queries. In this paper, SALOVA first retrieves relevant video segments, then concentrates on more granular video cues. Such targeted focus allows the model to effectively understand complex analysis within the videos, significantly improving its ability to provide contextual-aware and accurate responses.", + "bbox": [ + 91, + 351, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Long Video Understanding", + "text_level": 1, + "bbox": [ + 89, + 619, + 334, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In parallel, we dually introduce video-specialized models [12, 23, 53] integrated with LLMs, which have also been widely explored these days to enhance video understanding and reasoning. Here, the most challenging part of current video-LMMs lies in handling long video sequences, mainly due to the limited context length of the LLMs. This limitation compels the models to sparsely sample the video frames in only limited sizes (e.g., typically 8 or 16 frames), potentially missing important spatial and temporal information. To address this, several studies have focused on compressing visual tokens into a more manageable size, proposing aggregation [34, 41] or pooling methods [35, 59] with advanced vision encoder structures [67, 73]. In addition, memory-augmented methods [23, 53] first stored long-term information in a memory bank, then responded to specific queries by loading memory features from the stored buffer.", + "bbox": [ + 89, + 642, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, among more recent approaches, Li", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "et al. [70] have directly extended the LLMs' context length by exploiting RoPE-based frequency interpolation, and Xue et al. [61] have introduced sequence parallelism that can be implemented on multiple GPUs by modifying backend systems. However, we argue that such approaches inherently cannot be free from the fixed context length and provoke intensive memory demands when processing more longer videos. Instead, by focusing on the relevant segments within the entire video, SALOVA can efficiently handle the limited context length, enabling targeted processing of key moments without the need for excessive memory, thereby enhancing performance on longer video sequences.", + "bbox": [ + 511, + 351, + 906, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SceneWalk Dataset", + "text_level": 1, + "bbox": [ + 511, + 547, + 697, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on how we collected the SceneWalk dataset. The overall pipeline for building the dataset and summarized statistics are illustrated in Fig. 1. While several video SFT datasets [32, 41] are widely used during the instruction tuning stage, they often fail to capture comprehensive details within the scenes. This stems from the nature of instruction-type questions, which provide only partial information, and the brief lengths of both videos and texts in QAs. In contrast, the SceneWalk dataset offers densely captioned video-text pairs that cover long sequence videos in full details, as shown in Fig. 1(a). For further detailed data statistics, please see Appendix A.", + "bbox": [ + 511, + 574, + 906, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Gathering and Processing", + "text_level": 1, + "bbox": [ + 511, + 767, + 790, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Source & Filtering. For the long and untrimmed video sources, we primarily focus on three key aspects to build densely captioned video dataset: (i) extensive video length with diverse video source categories, (ii) high-quality video contents, (iii) frequent scene transitions within each video. Accordingly, our data collection is mainly sourced from YouTube, ensuring rich dynamic content that bet-", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3354", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ter reflects real-world complexities experienced by global users—here, because our main goal for video gathering is on complex scene understanding, we exclude low-quality and user-uploaded aesthetic videos (e.g., WebVid, Pixabay, Pexels, and etc.,) that are rather beneficial for video generation tasks, despite their merits for easy collection. We have collected YouTubeurls from [27] and downloaded the whole video in untrimmed states. Among the total 32 coarse and diverse video categories YouTube API provided, we selectively curated 10 categories, excluding categories such as News & Politics, Classics, and Documentary, due to their static nature, which provides only sparse temporal information in the videos. We further supplemented the dataset with additional Movie & Drama videos sourced from [21, 53], totaling 87,867 video sources with 11.87 Khrs video duration (avg. 486.5-seconds).", + "bbox": [ + 89, + 90, + 480, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segmenting Video into Clips. Next, for the collected long and untrimmed video sources, we cut the lengthy videos into small segments to densely caption the entire video in next phase. Instead of adopting the bottom-up approach used in the ShareGPT4Video dataset [8], which segments videos into fixed time intervals (2-seconds) in advance and then merges adjacent frames based on their CLIP similarity [48], we directly employed PysceneDetect 1 to segment the videos, dynamically adjusting the threshold based on the raw-level video information to reliably detect scene changes. At the end, the total number of 1.29M of video segments with 33.11-seconds average video length is extracted from the original video sources.", + "bbox": [ + 89, + 337, + 482, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Captioning and Scoring", + "text_level": 1, + "bbox": [ + 89, + 544, + 310, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dense Segment Captioning. After obtaining the massive video segments, our next goal is to caption each segment with visual details and narrative context to capture the scene-specific explanations, which can enrich scene-level interpretation. To achieve this, we plan to utilize pre-trained LMMs to generate detailed descriptions for the partial video segments. As the captioner, we empirically found that VILA-1.5 (13B) [36] shows competent descriptive quality than other open-sourced models, and used to generate dense captions for each video segment with randomly sampled instructions for detailed descriptions. As a result, we acquire 1.29M pairs of detailed descriptions corresponding to the video segments, each description with average 137.5 word length. Please see instruction details and qualitative examples of generated captions in Appendix A.", + "bbox": [ + 89, + 571, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scoring Video-Text Correspondence. Lastly, we score the correspondence between the video segments and the paired dense descriptions, which will later be used as explicit supervision to robustly train our retrieval framework. What", + "bbox": [ + 89, + 803, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we must not overlook here is that the paired video-text relationship is not solely a one-to-one correspondence but is more akin to generalized bipartite matching. That is, within the long and untrimmed video source, each video segment can be connected to other descriptions with additional edges. Therefore, for the $N_v$ number of video segments and their paired segments, we can construct a $\\{N_v\\}^2$ correspondence matrix between video-text (V2T). To measure each correspondence, we employ LanguageBind [73] due to its competitive alignment capabilities across various modalities. In addition, we build another $\\{N_v\\}^2$ matrix to provide a doubly robust measure for the correspondence scores among adjacent descriptions (T2T) by comparing similarity within the textual context using the SBERT model [54].", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Segment-Augmented LOng Video Assistant", + "text_level": 1, + "bbox": [ + 511, + 316, + 898, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network Overview. For a given set of $N_v$ video segments sampled at 1 FPS $v = \\{v_i\\}^{N_v}$ , where each segment $v_i \\in \\mathbb{R}^{T_i \\times H \\times W \\times C}$ has varying video length (summing up to the total time $T$ of a long and untrimmed video), SA-LOVA consists of four main architecture components as illustrated in Fig. 2:", + "bbox": [ + 511, + 348, + 903, + 438 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vision Encoder: We use CLIP [48] or SigLIP [67] to extract visual features, followed by 2x2 average pooling, resulting in 144 or 196 visual tokens per frame.", + "- Spatio-Temporal Connector: To handle spatio-temporal features of varying lengths from the vision encoder, we employ the Perceiver Resampler [2], which consists of 2-layer Transformer architecture followed by a 2-layer MLP with GELU activation as projector. This resampler embeds each video segment feature into fixed size latent features that are connected to LLMs.", + "- Segment Retrieval Router: For the given textual queries, a retrieval structure (2-layer Transformer) gathers representative information (i.e., routing tokens) from each video segment and then routes the query-relevant video features into the LLMs. Note that the router architecture is trained in an end-to-end manner.", + "- Large Language Model: For LM backbones, We utilize three open-sourced LLMs with varying parameter sizes, LLaMA-3.2 (3B) [18], Phi-3.5 (3.8B) [1] and Qwen-2.5 (7B) [63], all of which are instruction-tuned models that possess QA assistant capabilities." + ], + "bbox": [ + 513, + 439, + 905, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Long Video Processing and Pipeline", + "text_level": 1, + "bbox": [ + 511, + 767, + 823, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Spatio-Temporal Connector", + "text_level": 1, + "bbox": [ + 511, + 790, + 754, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first component of our model, Spatio-Temporal Connector, efficiently handles long and variable-length input video segments by extracting each segment's visual semantics in a fixed-size latent vector. As illustrated in Fig. 2(b), we first sample video frames at 1 FPS from each video segment, then visual features are acquired with $2 \\times 2$ pooling", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We use AdaptiveDetector with default setup in https://github. com/Breakthrough/PySceneDetect", + "bbox": [ + 89, + 875, + 480, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3355", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c393cd9f4355eeb974eb9dbfc2dad22383bb863765689a7956ae0455230c5844.jpg", + "image_caption": [ + "Figure 2: The network overview of SALOVA. Our framework consists of four structural components: vision encoder, ST-connector, SR-router, and LLMs. Using the FocusFast strategy, our model can concentrate on more detailed local information while maintaining context awareness." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(thus, 196 tokens from each frame). After that, the visual features are flattened and fed into the ST-Connector with additional positional and temporal encoding. Here, when the long video is processed, the number of unfolded patch tokens becomes extremely large, leading to exhaustive computations. To address this, we employ a dynamic token drop technique to reduce computational load.", + "bbox": [ + 88, + 420, + 480, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Token Drop. To effectively manage long video sequences, the token drop has been utilized in video generation tasks [16, 40]. Expanding such approach, in our framework, the dropout rate is dynamically adjusted based on the length of the input sequence $T_{i}$ in the input visual feature $f_{v} \\sim T_{i} \\times H_{p}W_{p} \\times d$ , which allows for more efficient processing of longer sequences by reducing computational demands, while still preserving dense visual semantics in shorter videos. Additionally, to retain spatio-temporal information from the dropped patches, we add positional embeddings separately along the spatial and temporal axes. This enables more refined extraction of spatio-temporal visual semantics even after reducing the number of tokens.", + "bbox": [ + 88, + 532, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Segment Retrieval Router", + "text_level": 1, + "bbox": [ + 89, + 743, + 316, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, the key to conveying the pertinent video information to LLMs is retrieving relevant video segments by querying sentence. To densely cue the similarities between the video and sentence information, we introduce a routing framework, Segment Retrieval Router, which consists of 2-layer Transformer as illustrated in Fig. 2(c). After obtaining the routing tokens $R = \\{R_{i}\\}^{N_{v}} \\in \\mathbb{R}^{N_{v} \\times D}$ from entire video segments, we aggregate them and feed into the SRRouter as queries. For the given sentence, we employ the", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "same text encoder used for the vision encoder and project it into the shared embedding space to obtain sentence features $S \\in \\mathbb{R}^{N_t \\times D}$ , where $N_{t}$ indicates textual length.", + "bbox": [ + 511, + 420, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the cross attention mechanism (q: $R$ ; k/v: $S$ ), we can estimate similarity scores between the video segments and given sentence queries (i.e., V-T similarity). The scores enable the SR-Router to prioritize and select the most relevant video segments that align with the sentence query.", + "bbox": [ + 511, + 465, + 906, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval Objective. To seamlessly train the SR-Router with the mainstream flows of SALOVA in an end-to-end manner, we have designed a similarity loss function $\\mathcal{L}_{\\mathrm{sim}}$ that minimizes the distance between the high-dimensional embeddings of the video segments and sentence queries. Here, we use the correspondence scores (aforementioned in Sec. 3.2) as a retrieval supervision signal $y_{i}$ , after applying one-hot encoding. We incorporate a simple margin-based loss, commonly used in contrastive learning settings [33], which enables the model to learn off-diagonal relaxation in the correspondence matrices between video segments and sentences. As mentioned earlier, the relationship between paired videos and sentences is closer to generalized bipartite matching than to one-to-one matching, so relaxation learning helps to accommodate the inherent complexity in aligning correspondence. In conclusion, with the binary cross-entropy loss and the score margin loss, we can formulate the similarity loss as follows:", + "bbox": [ + 511, + 547, + 908, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m}} = \\underbrace {\\mathcal {L} _ {\\text {b c e}} \\left(y _ {i} , s _ {i}\\right) _ {i = 1} ^ {N _ {v}}} _ {\\text {p o i n t - w i s e C E}} + \\underbrace {\\frac {1}{N _ {s}} \\sum_ {j} \\max \\left(0 , \\delta - \\left(s _ {j} ^ {p} - s _ {j} ^ {n}\\right)\\right)} _ {\\text {s c o r e m a r g i n l o s s}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 830, + 903, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_j^p$ and $s_j^n$ indicate randomly sampled scores from", + "bbox": [ + 511, + 885, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3356", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "positive and negative pairs, respectively, and $\\delta$ denotes the margin parameter (set as 0.2). Note that the similarity loss is trained in conjunction with the auto-regressive loss $\\mathcal{L}_{\\mathrm{ar}}$ from subsequent LLMs in an end-to-end manner.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3. FocusFast Pathways: Integration to LLMs", + "text_level": 1, + "bbox": [ + 89, + 161, + 431, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the routing tokens, we can calculate the similarities of each video segment for the given query. Leveraging the similarities, SALOVA efficiently retrieves the specific video features that exhibit the highest relevance score to the textual query, where the indexed video features are then directly integrated into the LLM architecture. Here, extending the SlowFast pathways concept [19], we present the FocusFast mechanism to effectively manage the processing pathway for the retrieved video segments: (i) Focus pathway concatenates the top-K most pertinent features to construct a comprehensive video representation, capturing local details across retrieved segments and enabling detailed interactions with textual queries to enhance handling complex video information. (ii) Fast pathway focuses on the more broader-level context by employing segment-wide routing tokens as the condensed global representation. It effectively contains dynamic spatio-temporal changes throughout the video stream, thereby allowing SALOVA to understand the overall video content and scene-level continuity awareness.", + "bbox": [ + 89, + 180, + 482, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the most pertinent features are retrieved, they are delivered to the LM backbone for the final processing as in Fig. 2(a), integrating video-specific details into the models' responses. By effectively handling long and untrimmed videos with the proposed retrieval and routing mechanism, SALOVA can maintain the flow of salient information without the processing overhead for less related data, thus generating more context-aware responses.", + "bbox": [ + 89, + 469, + 482, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Training Strategies", + "text_level": 1, + "bbox": [ + 89, + 599, + 274, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The current training strategies for LMMs predominantly consist of two-step training: (i) cross-modal alignment and (ii) visual instruction tuning. Recently, Li et al. [29] have emphasized the importance of high-quality knowledge learning between the two training stages (thus stage 1.5), pointing out that the models cannot enoughly learn necessary knowledge during the alignment with the low-qualitative web-scale image-text data. As the similar approach of using rephrased descriptions for additional knowledge learning [29], we employ the newly collected SceneWalk dataset as the parametric knowledge injection step, which enables the SALOVA to learn detailed spatial and temporal representation from the long sequence video data before the instruction tuning. Accordingly, our training recipes and data configuration can be divided into three steps as follow (Please see training details in Appendix B):", + "bbox": [ + 89, + 623, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 1: Cross-modality Alignment. For the initial step in modality alignment, we utilize 790K image/video-text", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "paired dataset: (i) 558K image-text pairs from the CC3M dataset [52], filtered by LLaVA [38] and (ii) video-text pairs sampled from the WebVid 2.5M subset [4]. We freeze vision encoder and LLMs during the training, and mainly focus on optimizing the connector and router to map the visual information into the textual space.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 1.5: Long Video Knowledge Injection. As an intermediate training step, we use the SceneWalk dataset to train the SALOVA, unfreezing all trainable parameters except for the vision encoder. During training, we input the long and untrimmed video instances and follow the processing pipeline shown in Fig. 2. By training the model with densely captioned video-description pairs, it acquires high-quality parametric knowledge of both spatial and temporal information. In addition, through the aforementioned retrieval process, the model learns to target video segments that are mostly relevant to the video description.", + "bbox": [ + 511, + 186, + 906, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 2: Video Instruction Tuning. To possess QA capabilities in SALOVA, we use extensive video instruction-tuning data as the final training step. The instruction data are mainly sourced from four different datasets: LLaVA-Video-178K [71], NeXT-QA [58], ActivityNetQA [66], and PerceptionTest [47]—comprising a total of 1.4M video-instruction QA data, including caption entries, open-ended QA, and multiple-choice QA. Note that we train all the network parameters during this stage and auto-regressively update the instruction-following assistant's response for the next word prediction.", + "bbox": [ + 511, + 358, + 908, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 540, + 645, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Details", + "text_level": 1, + "bbox": [ + 511, + 566, + 712, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. For the vision and text encoders of the SR-Router, we utilize CLIP [48] for the small-size models and SigLIP [67] for the frontier model, with resolution sizes of 336 and 384, respectively. We employ a 2-layer transformer with a head size of 2 for the ST-Connector, which has a latent dimension of 256. The token drop mechanism is dynamically applied based on video length, with varying maximum drop rates at each training stage—Stage 1 has no token drop, Stage 1.5 allows up to 0.7, and Stage 2 allows up to 0.4. For the configuration of SR-Router, we set 2-layer of transformers with a single head, and top-K number is set to 5 during the stage 2. Following the recent works [30, 37], the projector layer consists of 2-layer MLP with GELU. We employ three LLMs: (i) 3B: Llama-3.2-3B [18], (ii) 3.8B: Phi-3.5 [1], and (iii) 7B: Qwen2.5-7B [63].", + "bbox": [ + 511, + 593, + 906, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. For the each training stage, we train SA-LOVA for 1 epoch with 1 node of 8 A100 GPUs. The total training hours for 3B (3.8B) and 7B models roughly take 5 and 7 days, respectively. We employ FlashAttention-2 [15], gradient checkpointing [9], and ZeRO-2 [49] to minimize", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3357", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aa9e7fe8608057e062a9c38b990f48994589c3c4515a6c69a1917beee70f07a2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Detailed results for the Video-MME benchmark (w/o subtitles) and LongVideoBench. The best results are highlighted in **bold** and the runner-up results are **underlined**." + ], + "table_body": "
Model#paramVideo-MMELVBench Acc. (val)
ShortMediumLongOverall
Proprietary LMMs
GPT-4V [45]n/a70.555.853.559.9-
GPT-4o [46]n/a80.070.365.371.966.7
Gemini 1.5 Pro [50]n/a81.774.367.475.064.0
Open-sourced LMMs
ST-LLM [39]7B45.736.831.337.9-
VideoChat2 [32]7B48.337.033.239.539.3
ShareGPT4Video [8]8B48.336.335.039.939.7
Video-LLaVA [35]7B45.338.036.239.939.1
Chat-UniVi-V1.5 [26]7B45.740.335.840.6-
Qwen-VL-Chat [3]7B46.938.737.841.1-
ShareGemini [51]7B49.141.339.143.2-
SliME [72]8B53.342.739.845.3-
PLLAVA [59]7B----40.2
VideoLLaMA2 [12]8B56.045.442.147.9-
LongVA [70]7B61.150.446.252.6-
Ours
SALOVA-3B3B48.346.341.145.341.4
SALOVA-3.8B3.8B47.148.844.146.741.6
SALOVA-7B7B59.450.449.453.144.6
", + "bbox": [ + 93, + 88, + 478, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the memory footprint associated with model components (i.e., gradient, activation, and optimizer states). Additionally, we fine-tune the trainable parameters at each step without employing LoRA [24]. For the extended training config, we have attached the details in Appendix C.", + "bbox": [ + 89, + 375, + 482, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Benchmarks. We evaluate our model using two types of video analysis benchmarks—long video understanding and general video understanding, categorized based on the video length. For the long video benchmark, we primarily utilize Video-MME [20] and LongVideoBench [57], both of benchmarks include videos up to two hours long duration. As the general video analysis evaluations, we employ various benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32]. Note that the same pipeline is used to obtain video segments for each benchmark, and all benchmarks are sampled at 1 FPS without token drop during inference. As comparison baselines, considering academic budget constraints, we evaluate against models that have similar parameter size.", + "bbox": [ + 88, + 455, + 482, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 676, + 290, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Long Video Understanding. Video-MME [20] evaluates LMMs with a focus on video analysis across a variety of video types and durations. We primarily compare the benchmark results in settings without subtitles, relying solely on video frames. Therefore it can assess the LMMs' visual comprehension capabilities rigorously, based purely on visual content. Also, LongVideoBench [57] is designed to assess LMMs' understanding of long-duration videos up to two hours. It includes a diverse collection of videos, challenging the models' ability to process and interpret extensive visual and contextual information across a variety of themes. As shown in Tab. 1, our model shows competent video understanding performance across all video", + "bbox": [ + 89, + 703, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d9f276eb632b189d6883e2cfbadec4a71ca3b70f2eac85b1e25561e6b22481f6.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2: Comparison results for generic video understanding benchmarks. The best results are highlighted in bold and the runner-up results are underlined." + ], + "table_body": "
Model#paramActivityNetQAVideoChatGPTMVBench
test (acc/score)test (acc)test (acc)
Proprietary LMMs
GPT-4V [45]n/a57.0-4.0643.5
GPT-4o [46]n/a61.9---
Gemini 1.5 Pro [50]n/a57.5---
Open-sourced LMMs
VideoLLaMA [68]7B12.41.12.1634.1
VideoChatGPT [41]7B35.22.72.4232.7
MovieChat [53]7B45.7-2.67-
Chat-UniVi [26]7B46.13.22.99-
LLaMA-VID [34]7B47.43.32.8941.3
VideoChat2 [32]7B49.13.32.9851.1
VideoLLaMA2 [12]8B50.23.33.1354.6
Ours
SALOVA-3B3B52.63.43.0851.7
SALOVA-3.8B3.8B51.13.52.8346.4
SALOVA-7B7B53.63.53.1353.5
", + "bbox": [ + 517, + 88, + 903, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "length distributions in Video-MME and lengthy video instances in LongVideoBench. Notably, we highlight that SALOVA achieved significant performance in the medium (average 562.7 seconds) and long (average 2385.8 seconds) length categories in Video-MME benchmark, even with more smaller size of backbone LM parameters compared with the baseline models.", + "bbox": [ + 511, + 347, + 906, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Such performance gains in long video instances are attributed to our model's dynamic capability to retrieve and process only the relevant video segments, enabling it to handle lengthy video content efficiently without being constrained by the limited context length. Especially, the routing mechanism in SALOVA strategically prioritizes video segments that are likely to contain crucial visual and contextual cues relevant to the query. This selective routing mechanism reduces the computational load and minimizes the information loss that commonly occurs in current videolMMs trying to process extensive video data in entirety.", + "bbox": [ + 511, + 455, + 908, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on General Video Understanding. For general video understanding benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32], SALOVA was evaluated across various video types to assess its general video understanding capabilities. As shown in Tab. 2, SALOVA demonstrated competent performance, comparable to existing video-LMMs, especially in dynamic and shorter video sequences. On ActivityNetQA, the model effectively utilized its segment retrieval strategy to provide focused and contextually appropriate responses, which helped maintain accuracy. This approach was similarly effective in the multi-modal settings of VideoChatGPT and MVBench, where SALOVA showed consistent performance in handling dialogues and visual cues. These outcomes highlight SALOVA's capability to process general video content efficiently through its dynamic routing mechanism, offering a reliable solution that balances computational resources with output quality.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3358", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/584fd63f52401e38264afc94bb4baff635d1f2b542065e4520c4a0d1f6f2f65d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Ablation studies on SALOVA configuration. We utilize SALOVA-3B for the resource efficiency." + ], + "table_body": "
AblationVideo-MME
Short: ≤2mMid: 4-15mLong: 30-60mOverall
frm sample : Frame sampling rate (w/o SR-Router)
8 frm48.342.037.242.5
16 frm50.042.838.043.6
1 FPS48.346.341.145.3
1 / 1.5 / 2 : Train stage - Long video knowledge injection
✓ × ✓45.643.740.243.6
✓ ✓ ✓48.346.341.145.3
FocusFast : Local-global video representation
×36.438.635.636.9
✓48.346.341.145.3
", + "bbox": [ + 94, + 88, + 478, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Additional Analyses on SALOVA", + "text_level": 1, + "bbox": [ + 89, + 294, + 382, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablation studies on three components as follows: (i) different video frame sampling strategies, (ii) intermediate training stage for long video knowledge injection, and (iii) the FocusFast mechanism to understand branched local-global representation in videos.", + "bbox": [ + 89, + 321, + 482, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As in Table 3, We first observe that using more frames with SR-Router significantly enhances performance, particularly in long-form videos. This aligns with our key contributions on managing long videos through an effective routing mechanism, suggesting that retrieving more frames can provide richer spatio-temporal information and improve the model's responses without losing contextual coherence. Additionally, we compare with a baseline trained with stage 1-2 (skipping stage 1.5). Here, we highlight the effectiveness of the SceneWalk dataset as an intermediate training step to enhance parametric knowledge for the long video analysis by allowing the model to learn from high-quality and densely captioned scene-level information, which is crucial for adapting to various lengths and contexts. Lastly, we conduct an analysis on the FocusFast method and demonstrate its efficacy in analyzing not only local details from relevant video segments but also in understanding the global video context through the simultaneous use of routing tokens, thereby facilitating a more comprehensive understanding of video content.", + "bbox": [ + 89, + 397, + 482, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Retrieving Segments. By retrieving relevant video segments for the given queries, SALOVA can effectively target salient information in the long video and retain long context information. To further demonstrate the model's targeting capabilities beyond numerical performance in long video analysis, we explore our model's application in the Visual Needle-In-A-Haystack (V-NIAH) task [70], which extends the Needle-in-a-Haystack (NIAH) evaluation for LLMs to a vision-level benchmark. This task is particularly challenging as it requires models to not only detect but also precisely retrieve the sparse yet crucial visual cues scattered across lengthy videos.", + "bbox": [ + 89, + 704, + 482, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As in Fig. 3, we compare SALOVA against a base", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca1a620b76da1ac252ac4fb290fa5ce8566e7356fabd8bf012e83a0c19238014.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61df37e920f86f5a27194cc2fe924acd75ca4571035844b0909ad159f7fe714a.jpg", + "image_caption": [ + "Figure 3: Comparison results of V-NIAH. The x/y-axis indicates the total video frames and the location of needle image within the video, respectively." + ], + "image_footnote": [], + "bbox": [ + 514, + 218, + 903, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "line trained on sparsely sampled frames (16 frm w/o SR-Router). Our framework effectively identifies and extracts relevant video segments from densely packed content, even when handling long context lengths. These results highlight SALOVA's robustness in managing complex, long-form videos, maintaining contextual continuity by strategically focusing on relevant segments to user queries.", + "bbox": [ + 511, + 402, + 905, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 520, + 761, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. Despite SALOVA's competence in handling extended video sequences, it is important to recognize scenarios where its complex architecture may not be necessary. Specifically, for shorter videos where sparse sampling suffices to capture essential spatio-temporal information, simpler models could potentially outperform the efficiency of SALOVA without necessitating its extensive processing capabilities. This suggests a future avenue for integrating a hybrid approach based on our framework by dynamically adjusting the complexity of the retrieval and processing mechanisms based on the video length and content density.", + "bbox": [ + 511, + 547, + 906, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we introduce SALOVA, a novel framework designed to enhance the comprehension of long and untrimmed video by leveraging a retrieval-driven approach with new densely captioned dataset, the SceneWalk dataset. SALOVA strategically targets and processes only the relevant video segments, effectively addressing the structural limitations of current Video-LMMs with its Spatio-Temporal Connector and Segment Retrieval Router. Through extensive evaluation on various benchmarks, SALOVA exhibits its robust performance in interpreting complex video content, enhancing efficiency, and improving the understanding of extended videos.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3359", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by two funds: IITP grant funded by the Korea government (MSIT) (RS-2022-II220984) and IITP grant funded by the Korea government (MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-Term Visual Memory Network)", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 4, 6", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4, 3", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 7", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2", + "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2", + "[7] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. arXiv preprint arXiv:2312.06742, 2023. 2", + "[8] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 4, 7, 3", + "[9] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 6", + "[10] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 232, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1", + "[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 3, 7", + "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 2", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2", + "[15] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 6", + "[16] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 1, 2", + "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4, 6", + "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 6", + "[20] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7, 6", + "[21] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (sfd): A benchmark for story-level video understanding. arXiv preprint arXiv:2406.10221, 2024. 4, 1", + "[22] Google. Gemini, 2023. 1", + "[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3360", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 1, 3", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7", + "[25] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. Language is not all you need: Aligning perception with language models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[26] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7", + "[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 4", + "[28] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 2", + "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3, 6", + "[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 2, 6", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning. PMLR, 2023. 2", + "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 7", + "[33] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. Advances in neural information processing systems, 36, 2024. 5", + "[34] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 323-340. Springer, 2025. 1, 3, 7", + "[35] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 3, 7", + "[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 4, 1", + "[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 2, 6, 3", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2, 6", + "[39] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2025. 1, 7", + "[40] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 5", + "[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 3, 7", + "[42] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2", + "[43] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023.1", + "[44] OpenAI. Gpt-4 technical report, 2023. 1, 3", + "[45] OpenAI. GPT-4V(ision) System Card, 2023. 7", + "[46] OpenAI. Hello gpt-4o, 2024. 3, 7", + "[47] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6, 3", + "[49] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 6" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3361", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 3", + "[51] Share. Sharegemini: Scaling up video caption data for multimodal large language models, 2024. 7", + "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6", + "[53] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 3, 4, 7", + "[54] Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. Augmented SBERT: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 296–310, Online, 2021. Association for Computational Linguistics. 2, 4", + "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[56] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2", + "[57] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. arXiv preprint arXiv:2407.15754, 2024. 7, 3", + "[58] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6", + "[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3, 7", + "[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 1", + "[61] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 3", + "[62] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2", + "[63] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4, 6", + "[64] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 3", + "[65] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2", + "[66] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 6, 7", + "[67] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 4, 6", + "[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7", + "[69] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1", + "[70] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 3, 7, 8", + "[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 6", + "[72] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 7", + "[73] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations. 2, 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3362", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_model.json b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c48570d5beb2e09381922b9d841ddecb026eee8a --- /dev/null +++ b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "SALOVA: Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.204, + 0.773, + 0.258 + ], + "angle": 0, + "content": "Junho Kim* Hyunjun Kim* Hosu Lee Yong Man Ro† Integrated Vision and Language Lab, KAIST, South Korea {arkimjh, kimhj709, leehosu01, ymro}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Despite advances in Large Multi-modal Models, applying them to long and untrimmed video content remains challenging due to limitations in context length and substantial memory overhead. These constraints often lead to significant information loss and reduced relevance in the model responses. With the exponential growth of video data across web platforms, understanding long-form video is crucial for advancing generalized intelligence. In this paper, we introduce SALOVA: Segment-Augmented LOng Video Assistant, a novel video-LLM framework designed to enhance the comprehension of lengthy video content through targeted retrieval process. We address two main challenges to achieve it: (i) We present the SceneWalk dataset, a high-quality collection of 87.8K long videos, each densely captioned at the segment level to enable models to capture scene continuity and maintain rich descriptive context. (ii) We develop robust architectural designs integrating dynamic routing mechanism and spatio-temporal projector to efficiently retrieve and process relevant video segments based on user queries. Our framework mitigates the limitations of current video-LMMs by allowing for precise identification and retrieval of relevant video segments in response to queries, thereby improving the contextual relevance of the generated responses. Through extensive experiments, SALOVA demonstrates enhanced capability in processing complex long-form videos, showing significant capability to maintain contextual integrity across extended sequences." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Recent advancements in Large Language Models (LLMs) [22, 43, 44] have brought us one step closer to achieving Artificial General Intelligence (AGI). Next following step, the current trend is shifting toward modular systems that integrate various multi-modality, leveraging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.43 + ], + "angle": 0, + "content": "the exceptional generalization and reasoning capabilities of LLMs to evolve into Large Multi-modal Models (LMMs). Accordingly, users can unrestrictedly interact with the models across various modalities beyond text, expanding the scope of machine understanding and enhancing user engagement. Especially, considering the widespread adoption of long-form videos across various web platforms, the importance of understanding long, untrimmed video has become increasingly prominent in the multi-modal domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.674 + ], + "angle": 0, + "content": "After the pioneer works [14, 37, 38] utilizing visual instruction tuning to augment vision perception into LLMs, remarkable strides [11, 17, 65] have been made in aligning cross-modal consistency—especially between vision and language domains. Albeit more recent models [29, 69] integrate various vision modalities all at once, current approaches still face significant challenges in understanding untrimmed and long-form video content. The main challenge is attributed to the limited context length of LMMs, which is an inherent structural limitation that restricts the models to process only a finite number of tokens as the input sequences. We can exemplify that LLaVA series [29, 30], when processing a video data, require 144 visual tokens per each frame, where numerical approximation is only maximum of \\(\\sim 56\\) frames using 8K max context length LMMs, which is still limited to handle long sequence data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Accordingly, current video-LMMs [12, 35, 39] rely on (i) sparse frame sampling to represent entire videos [26, 72], (ii) dense compression of visual tokens into a smaller size to manage the excessive number of frames [34, 41], and (iii) adaptive pooling strategies [59, 60] based on the SlowFast approach [19], all aimed at fitting the long video sequences within the limited context window of LMMs. Several studies focusing on the long video understanding task have presented memory-augmented generation [23, 53] utilizing an additional buffer to embed long-term information, or have extended the context using RoPE-based frequency extension during the training [70]. Despite of such endeavors, when handling massive video frames, previous works still confront restricted context size and significant memory overhead, which leads to substantial visual information" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.876, + 0.388, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author. Code is available at https://ivy-lvlm.github.io/SALOVA" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3352" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "loss. As critical events may be overlooked by the models, this hinders their ability to fully capture context changes in lengthy videos, resulting in inaccurate and irrelevant responses for the user queries." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.483, + 0.379 + ], + "angle": 0, + "content": "Starting from the intuitive insight outlined below, in this paper, we propose a retrieval-driven approach for long video understanding with LLMs. Analogous to the recent Retriever-Augmented Generation (RAG) systems [28] (widely adopted in LLMs), which retrieve relevant information from external factual knowledge, humans naturally employ similar strategies when seeking specific information, efficiently locating and referring necessary materials to answer targeted questions—e.g., imagine that we are taking open-book exams or searching for a certain recipe in a cookbook. Given a long and untrimmed video, mirroring the targeted retrieval processes, we introduce a novel framework, Segment-Augmented LOng Video Assistant (SALOVA) to effectively handle the long sequence visual inputs by retrieving the relevant video segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.515 + ], + "angle": 0, + "content": "To construct our video segments retrieval framework, central challenge hinges on establishing two main components: (i) Densely captioned video data, which consists of video-caption pairs with progressively-captioned descriptions that change throughout each video used to train the model to accurately identify relevant video segments. (ii) Dynamic routing mechanism, which selects pertinent video segments for the queries, followed by being connected to LLMs. To address that, our approach is outlined as follows." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Data. (§3) Recently several video-text paired datasets [4, 6, 8, 10, 56, 62] have been released, but they are inadequate for handling long and untrimmed video data, where only partial video moments are described with limited word length as compared in Fig. 1(a). To handle such insufficiency of detailed descriptions within the videos and the short durations of both videos and texts, we introduce the SceneWalk dataset, a new high-quality video dataset with thorough captioning for each video. It includes dense and detailed descriptions for every video segment across the entire scene context. The SceneWalk dataset, sourced from long and untrimmed 87.8K YouTube videos (avg. 486 seconds each), features frequent scene transitions across a total of 11.8K hrs video duration and 1.3M massively segmented video clips. Each video segment in the dataset is provided with a detailed description (avg. 137.5 word length), generated by combining pre-trained models [54, 73] and manual curation from human." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Architecture. (§4) Utilizing the constructed video dataset, SALOVA learns to identify relevant video segments for the given queries within each video source and then auto-regressively predicts the next token. To do so, we present two architectural designs to seamlessly incorporate the retrieved segments in an end-to-end training: Spatio-Temporal Connector and Segment Retrieval Router. By fo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "cusing on the relevant segments, our framework can perform deeper reasoning without being constrained by context length limitations. Additionally, we present FocusFast approach, which intensively analyzes the selected segments for detailed comprehension (focus pathway), while quickly accessing overall contextual information with routing tokens obtained from the entire video segments (fast pathway). The strategy ensures SALOVA to maintain comprehensive video understanding while prioritizing details where it is most needed, effectively enhancing long and untrimmed video interpretation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Through extensive experiments and analyses, we corroborate that competitive performance of SALOVA to the existing video-LMM models in understanding complex long-form videos. Also, our results show significant reductions in the loss of crucial visual information and a lower risk of omitting important events, demonstrating the effectiveness of our proposed method across various video benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.366, + 0.88, + 0.38 + ], + "angle": 0, + "content": "Our contribution can be summarized into three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.382, + 0.905, + 0.472 + ], + "angle": 0, + "content": "- We introduce the SceneWalk dataset, a high-quality and densely-captioned video dataset with detailed segment-level descriptions from 87.8K long and untrimmed video sources. The proposed dataset provides rich context and scene continuity, enabling effective training for long-form video understanding." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.473, + 0.905, + 0.547 + ], + "angle": 0, + "content": "- We propose Segment-Augmented LOng Video Assistant (SALOVA), a novel video-LMM framework designed to enhance long video comprehension by targeting relevant video segments in lengthy videos, optimizing the model's focus on essential segment targets for the given queries." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.548, + 0.905, + 0.608 + ], + "angle": 0, + "content": "- Through extensive evaluation, we validate that SALOVA improves overall long video understanding capabilities by effectively integrating relevant video segments, thus optimizing to handle long and untrimmed video content." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.382, + 0.905, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.625, + 0.655, + 0.641 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.651, + 0.757, + 0.667 + ], + "angle": 0, + "content": "2.1. Large Multi-modal Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "After the emergence of LLMs [5, 55], which can actively interact with users through back-and-forth conversations, as a next leap, various research efforts [2, 25, 31] integrate different modalities into the LLMs, utilizing their core reasoning and zero-shot capabilities. Building on the open-sourced models [13, 55], seminal works [14, 38, 65] have bridged image and text modality under the visual instruction tuning and presented multi-modal assistant models that possess visual perception and QA capabilities. Since then, numerous research studies have been introduced to (i) enhance vision understanding with advanced architectures [17] or higher resolutions [30, 37], (ii) implement more sophisticated alignment layers [7, 42] between modalities, and (iii) train the models with more high-quality data and larger model parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3353" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.348, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.257, + 0.346, + 0.273 + ], + "angle": 0, + "content": "(a) Video-Text Dataset Comparison" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.088, + 0.553, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.257, + 0.536, + 0.271 + ], + "angle": 0, + "content": "(b) SceneWalk Statistics" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.099, + 0.906, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.258, + 0.866, + 0.272 + ], + "angle": 0, + "content": "(c) Overall Pipeline for Data Collection" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.277, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 1: The overview of the SceneWalk dataset includes (a) dataset comparison, (b) detailed statistics, and (c) the annotation pipeline for description and score collection. Note that the scale of circles in Fig. 1(a) indicates the data size, and the color distribution in Fig. 1(b) denotes the video duration in each video category—brighter colors correspond to shorter video durations. Further details about the dataset are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.352, + 0.485, + 0.61 + ], + "angle": 0, + "content": "Recent focus has shifted towards more unified modality processing following the release of omnivorous models [46]. Some recent omni-versions of LMMs [29, 64] can handle combinatorial subsets from various modality sources, such as images, videos, audio, speech, and depth. However, the current LMMs for video [32, 35, 41] still lack of capturing the necessary details to effectively process video information due to their sparse frame sampling strategy. While such approach is seemingly adequate for relatively shorter videos, it may fail to capture comprehensive spatio-temporal information, potentially compromising the accuracy of model responses to user queries. In this paper, SALOVA first retrieves relevant video segments, then concentrates on more granular video cues. Such targeted focus allows the model to effectively understand complex analysis within the videos, significantly improving its ability to provide contextual-aware and accurate responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.336, + 0.637 + ], + "angle": 0, + "content": "2.2. Long Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.885 + ], + "angle": 0, + "content": "In parallel, we dually introduce video-specialized models [12, 23, 53] integrated with LLMs, which have also been widely explored these days to enhance video understanding and reasoning. Here, the most challenging part of current video-LMMs lies in handling long video sequences, mainly due to the limited context length of the LLMs. This limitation compels the models to sparsely sample the video frames in only limited sizes (e.g., typically 8 or 16 frames), potentially missing important spatial and temporal information. To address this, several studies have focused on compressing visual tokens into a more manageable size, proposing aggregation [34, 41] or pooling methods [35, 59] with advanced vision encoder structures [67, 73]. In addition, memory-augmented methods [23, 53] first stored long-term information in a memory bank, then responded to specific queries by loading memory features from the stored buffer." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "On the other hand, among more recent approaches, Li" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.534 + ], + "angle": 0, + "content": "et al. [70] have directly extended the LLMs' context length by exploiting RoPE-based frequency interpolation, and Xue et al. [61] have introduced sequence parallelism that can be implemented on multiple GPUs by modifying backend systems. However, we argue that such approaches inherently cannot be free from the fixed context length and provoke intensive memory demands when processing more longer videos. Instead, by focusing on the relevant segments within the entire video, SALOVA can efficiently handle the limited context length, enabling targeted processing of key moments without the need for excessive memory, thereby enhancing performance on longer video sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.549, + 0.699, + 0.565 + ], + "angle": 0, + "content": "3. SceneWalk Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.756 + ], + "angle": 0, + "content": "In this section, we elaborate on how we collected the SceneWalk dataset. The overall pipeline for building the dataset and summarized statistics are illustrated in Fig. 1. While several video SFT datasets [32, 41] are widely used during the instruction tuning stage, they often fail to capture comprehensive details within the scenes. This stems from the nature of instruction-type questions, which provide only partial information, and the brief lengths of both videos and texts in QAs. In contrast, the SceneWalk dataset offers densely captioned video-text pairs that cover long sequence videos in full details, as shown in Fig. 1(a). For further detailed data statistics, please see Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.791, + 0.785 + ], + "angle": 0, + "content": "3.1. Data Gathering and Processing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Video Source & Filtering. For the long and untrimmed video sources, we primarily focus on three key aspects to build densely captioned video dataset: (i) extensive video length with diverse video source categories, (ii) high-quality video contents, (iii) frequent scene transitions within each video. Accordingly, our data collection is mainly sourced from YouTube, ensuring rich dynamic content that bet-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3354" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": "ter reflects real-world complexities experienced by global users—here, because our main goal for video gathering is on complex scene understanding, we exclude low-quality and user-uploaded aesthetic videos (e.g., WebVid, Pixabay, Pexels, and etc.,) that are rather beneficial for video generation tasks, despite their merits for easy collection. We have collected YouTubeurls from [27] and downloaded the whole video in untrimmed states. Among the total 32 coarse and diverse video categories YouTube API provided, we selectively curated 10 categories, excluding categories such as News & Politics, Classics, and Documentary, due to their static nature, which provides only sparse temporal information in the videos. We further supplemented the dataset with additional Movie & Drama videos sourced from [21, 53], totaling 87,867 video sources with 11.87 Khrs video duration (avg. 486.5-seconds)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Segmenting Video into Clips. Next, for the collected long and untrimmed video sources, we cut the lengthy videos into small segments to densely caption the entire video in next phase. Instead of adopting the bottom-up approach used in the ShareGPT4Video dataset [8], which segments videos into fixed time intervals (2-seconds) in advance and then merges adjacent frames based on their CLIP similarity [48], we directly employed PysceneDetect 1 to segment the videos, dynamically adjusting the threshold based on the raw-level video information to reliably detect scene changes. At the end, the total number of 1.29M of video segments with 33.11-seconds average video length is extracted from the original video sources." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.312, + 0.563 + ], + "angle": 0, + "content": "3.2. Captioning and Scoring" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.799 + ], + "angle": 0, + "content": "Dense Segment Captioning. After obtaining the massive video segments, our next goal is to caption each segment with visual details and narrative context to capture the scene-specific explanations, which can enrich scene-level interpretation. To achieve this, we plan to utilize pre-trained LMMs to generate detailed descriptions for the partial video segments. As the captioner, we empirically found that VILA-1.5 (13B) [36] shows competent descriptive quality than other open-sourced models, and used to generate dense captions for each video segment with randomly sampled instructions for detailed descriptions. As a result, we acquire 1.29M pairs of detailed descriptions corresponding to the video segments, each description with average 137.5 word length. Please see instruction details and qualitative examples of generated captions in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Scoring Video-Text Correspondence. Lastly, we score the correspondence between the video segments and the paired dense descriptions, which will later be used as explicit supervision to robustly train our retrieval framework. What" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "we must not overlook here is that the paired video-text relationship is not solely a one-to-one correspondence but is more akin to generalized bipartite matching. That is, within the long and untrimmed video source, each video segment can be connected to other descriptions with additional edges. Therefore, for the \\( N_v \\) number of video segments and their paired segments, we can construct a \\( \\{N_v\\}^2 \\) correspondence matrix between video-text (V2T). To measure each correspondence, we employ LanguageBind [73] due to its competitive alignment capabilities across various modalities. In addition, we build another \\( \\{N_v\\}^2 \\) matrix to provide a doubly robust measure for the correspondence scores among adjacent descriptions (T2T) by comparing similarity within the textual context using the SBERT model [54]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.9, + 0.337 + ], + "angle": 0, + "content": "4. Segment-Augmented LOng Video Assistant" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.439 + ], + "angle": 0, + "content": "Network Overview. For a given set of \\(N_v\\) video segments sampled at 1 FPS \\(v = \\{v_i\\}^{N_v}\\), where each segment \\(v_i \\in \\mathbb{R}^{T_i \\times H \\times W \\times C}\\) has varying video length (summing up to the total time \\(T\\) of a long and untrimmed video), SA-LOVA consists of four main architecture components as illustrated in Fig. 2:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.485 + ], + "angle": 0, + "content": "- Vision Encoder: We use CLIP [48] or SigLIP [67] to extract visual features, followed by 2x2 average pooling, resulting in 144 or 196 visual tokens per frame." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.485, + 0.906, + 0.59 + ], + "angle": 0, + "content": "- Spatio-Temporal Connector: To handle spatio-temporal features of varying lengths from the vision encoder, we employ the Perceiver Resampler [2], which consists of 2-layer Transformer architecture followed by a 2-layer MLP with GELU activation as projector. This resampler embeds each video segment feature into fixed size latent features that are connected to LLMs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.591, + 0.906, + 0.68 + ], + "angle": 0, + "content": "- Segment Retrieval Router: For the given textual queries, a retrieval structure (2-layer Transformer) gathers representative information (i.e., routing tokens) from each video segment and then routes the query-relevant video features into the LLMs. Note that the router architecture is trained in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.682, + 0.905, + 0.757 + ], + "angle": 0, + "content": "- Large Language Model: For LM backbones, We utilize three open-sourced LLMs with varying parameter sizes, LLaMA-3.2 (3B) [18], Phi-3.5 (3.8B) [1] and Qwen-2.5 (7B) [63], all of which are instruction-tuned models that possess QA assistant capabilities." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.44, + 0.906, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.825, + 0.785 + ], + "angle": 0, + "content": "4.1. Long Video Processing and Pipeline" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.755, + 0.806 + ], + "angle": 0, + "content": "4.1.1. Spatio-Temporal Connector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The first component of our model, Spatio-Temporal Connector, efficiently handles long and variable-length input video segments by extracting each segment's visual semantics in a fixed-size latent vector. As illustrated in Fig. 2(b), we first sample video frames at 1 FPS from each video segment, then visual features are acquired with \\(2 \\times 2\\) pooling" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.482, + 0.901 + ], + "angle": 0, + "content": "1We use AdaptiveDetector with default setup in https://github. com/Breakthrough/PySceneDetect" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3355" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.361, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 2: The network overview of SALOVA. Our framework consists of four structural components: vision encoder, ST-connector, SR-router, and LLMs. Using the FocusFast strategy, our model can concentrate on more detailed local information while maintaining context awareness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.482, + 0.526 + ], + "angle": 0, + "content": "(thus, 196 tokens from each frame). After that, the visual features are flattened and fed into the ST-Connector with additional positional and temporal encoding. Here, when the long video is processed, the number of unfolded patch tokens becomes extremely large, leading to exhaustive computations. To address this, we employ a dynamic token drop technique to reduce computational load." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.533, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Dynamic Token Drop. To effectively manage long video sequences, the token drop has been utilized in video generation tasks [16, 40]. Expanding such approach, in our framework, the dropout rate is dynamically adjusted based on the length of the input sequence \\( T_{i} \\) in the input visual feature \\( f_{v} \\sim T_{i} \\times H_{p}W_{p} \\times d \\), which allows for more efficient processing of longer sequences by reducing computational demands, while still preserving dense visual semantics in shorter videos. Additionally, to retain spatio-temporal information from the dropped patches, we add positional embeddings separately along the spatial and temporal axes. This enables more refined extraction of spatio-temporal visual semantics even after reducing the number of tokens." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.744, + 0.318, + 0.759 + ], + "angle": 0, + "content": "4.1.2. Segment Retrieval Router" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Next, the key to conveying the pertinent video information to LLMs is retrieving relevant video segments by querying sentence. To densely cue the similarities between the video and sentence information, we introduce a routing framework, Segment Retrieval Router, which consists of 2-layer Transformer as illustrated in Fig. 2(c). After obtaining the routing tokens \\( R = \\{R_{i}\\}^{N_{v}} \\in \\mathbb{R}^{N_{v} \\times D} \\) from entire video segments, we aggregate them and feed into the SRRouter as queries. For the given sentence, we employ the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.907, + 0.466 + ], + "angle": 0, + "content": "same text encoder used for the vision encoder and project it into the shared embedding space to obtain sentence features \\( S \\in \\mathbb{R}^{N_t \\times D} \\), where \\( N_{t} \\) indicates textual length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Using the cross attention mechanism (q: \\( R \\); k/v: \\( S \\)), we can estimate similarity scores between the video segments and given sentence queries (i.e., V-T similarity). The scores enable the SR-Router to prioritize and select the most relevant video segments that align with the sentence query." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Retrieval Objective. To seamlessly train the SR-Router with the mainstream flows of SALOVA in an end-to-end manner, we have designed a similarity loss function \\(\\mathcal{L}_{\\mathrm{sim}}\\) that minimizes the distance between the high-dimensional embeddings of the video segments and sentence queries. Here, we use the correspondence scores (aforementioned in Sec. 3.2) as a retrieval supervision signal \\(y_{i}\\), after applying one-hot encoding. We incorporate a simple margin-based loss, commonly used in contrastive learning settings [33], which enables the model to learn off-diagonal relaxation in the correspondence matrices between video segments and sentences. As mentioned earlier, the relationship between paired videos and sentences is closer to generalized bipartite matching than to one-to-one matching, so relaxation learning helps to accommodate the inherent complexity in aligning correspondence. In conclusion, with the binary cross-entropy loss and the score margin loss, we can formulate the similarity loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.832, + 0.905, + 0.886 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m}} = \\underbrace {\\mathcal {L} _ {\\text {b c e}} \\left(y _ {i} , s _ {i}\\right) _ {i = 1} ^ {N _ {v}}} _ {\\text {p o i n t - w i s e C E}} + \\underbrace {\\frac {1}{N _ {s}} \\sum_ {j} \\max \\left(0 , \\delta - \\left(s _ {j} ^ {p} - s _ {j} ^ {n}\\right)\\right)} _ {\\text {s c o r e m a r g i n l o s s}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.904 + ], + "angle": 0, + "content": "where \\( s_j^p \\) and \\( s_j^n \\) indicate randomly sampled scores from" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "positive and negative pairs, respectively, and \\(\\delta\\) denotes the margin parameter (set as 0.2). Note that the similarity loss is trained in conjunction with the auto-regressive loss \\(\\mathcal{L}_{\\mathrm{ar}}\\) from subsequent LLMs in an end-to-end manner." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.162, + 0.433, + 0.177 + ], + "angle": 0, + "content": "4.1.3. FocusFast Pathways: Integration to LLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.468 + ], + "angle": 0, + "content": "Using the routing tokens, we can calculate the similarities of each video segment for the given query. Leveraging the similarities, SALOVA efficiently retrieves the specific video features that exhibit the highest relevance score to the textual query, where the indexed video features are then directly integrated into the LLM architecture. Here, extending the SlowFast pathways concept [19], we present the FocusFast mechanism to effectively manage the processing pathway for the retrieved video segments: (i) Focus pathway concatenates the top-K most pertinent features to construct a comprehensive video representation, capturing local details across retrieved segments and enabling detailed interactions with textual queries to enhance handling complex video information. (ii) Fast pathway focuses on the more broader-level context by employing segment-wide routing tokens as the condensed global representation. It effectively contains dynamic spatio-temporal changes throughout the video stream, thereby allowing SALOVA to understand the overall video content and scene-level continuity awareness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Once the most pertinent features are retrieved, they are delivered to the LM backbone for the final processing as in Fig. 2(a), integrating video-specific details into the models' responses. By effectively handling long and untrimmed videos with the proposed retrieval and routing mechanism, SALOVA can maintain the flow of salient information without the processing overhead for less related data, thus generating more context-aware responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.275, + 0.617 + ], + "angle": 0, + "content": "4.2. Training Strategies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.866 + ], + "angle": 0, + "content": "The current training strategies for LMMs predominantly consist of two-step training: (i) cross-modal alignment and (ii) visual instruction tuning. Recently, Li et al. [29] have emphasized the importance of high-quality knowledge learning between the two training stages (thus stage 1.5), pointing out that the models cannot enoughly learn necessary knowledge during the alignment with the low-qualitative web-scale image-text data. As the similar approach of using rephrased descriptions for additional knowledge learning [29], we employ the newly collected SceneWalk dataset as the parametric knowledge injection step, which enables the SALOVA to learn detailed spatial and temporal representation from the long sequence video data before the instruction tuning. Accordingly, our training recipes and data configuration can be divided into three steps as follow (Please see training details in Appendix B):" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Stage 1: Cross-modality Alignment. For the initial step in modality alignment, we utilize 790K image/video-text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "paired dataset: (i) 558K image-text pairs from the CC3M dataset [52], filtered by LLaVA [38] and (ii) video-text pairs sampled from the WebVid 2.5M subset [4]. We freeze vision encoder and LLMs during the training, and mainly focus on optimizing the connector and router to map the visual information into the textual space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.907, + 0.354 + ], + "angle": 0, + "content": "Stage 1.5: Long Video Knowledge Injection. As an intermediate training step, we use the SceneWalk dataset to train the SALOVA, unfreezing all trainable parameters except for the vision encoder. During training, we input the long and untrimmed video instances and follow the processing pipeline shown in Fig. 2. By training the model with densely captioned video-description pairs, it acquires high-quality parametric knowledge of both spatial and temporal information. In addition, through the aforementioned retrieval process, the model learns to target video segments that are mostly relevant to the video description." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.909, + 0.526 + ], + "angle": 0, + "content": "Stage 2: Video Instruction Tuning. To possess QA capabilities in SALOVA, we use extensive video instruction-tuning data as the final training step. The instruction data are mainly sourced from four different datasets: LLaVA-Video-178K [71], NeXT-QA [58], ActivityNetQA [66], and PerceptionTest [47]—comprising a total of 1.4M video-instruction QA data, including caption entries, open-ended QA, and multiple-choice QA. Note that we train all the network parameters during this stage and auto-regressively update the instruction-following assistant's response for the next word prediction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.646, + 0.558 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.714, + 0.582 + ], + "angle": 0, + "content": "5.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.821 + ], + "angle": 0, + "content": "Implementation. For the vision and text encoders of the SR-Router, we utilize CLIP [48] for the small-size models and SigLIP [67] for the frontier model, with resolution sizes of 336 and 384, respectively. We employ a 2-layer transformer with a head size of 2 for the ST-Connector, which has a latent dimension of 256. The token drop mechanism is dynamically applied based on video length, with varying maximum drop rates at each training stage—Stage 1 has no token drop, Stage 1.5 allows up to 0.7, and Stage 2 allows up to 0.4. For the configuration of SR-Router, we set 2-layer of transformers with a single head, and top-K number is set to 5 during the stage 2. Following the recent works [30, 37], the projector layer consists of 2-layer MLP with GELU. We employ three LLMs: (i) 3B: Llama-3.2-3B [18], (ii) 3.8B: Phi-3.5 [1], and (iii) 7B: Qwen2.5-7B [63]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Training Details. For the each training stage, we train SA-LOVA for 1 epoch with 1 node of 8 A100 GPUs. The total training hours for 3B (3.8B) and 7B models roughly take 5 and 7 days, respectively. We employ FlashAttention-2 [15], gradient checkpointing [9], and ZeRO-2 [49] to minimize" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3357" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.48, + 0.319 + ], + "angle": 0, + "content": "
Model#paramVideo-MMELVBench Acc. (val)
ShortMediumLongOverall
Proprietary LMMs
GPT-4V [45]n/a70.555.853.559.9-
GPT-4o [46]n/a80.070.365.371.966.7
Gemini 1.5 Pro [50]n/a81.774.367.475.064.0
Open-sourced LMMs
ST-LLM [39]7B45.736.831.337.9-
VideoChat2 [32]7B48.337.033.239.539.3
ShareGPT4Video [8]8B48.336.335.039.939.7
Video-LLaVA [35]7B45.338.036.239.939.1
Chat-UniVi-V1.5 [26]7B45.740.335.840.6-
Qwen-VL-Chat [3]7B46.938.737.841.1-
ShareGemini [51]7B49.141.339.143.2-
SliME [72]8B53.342.739.845.3-
PLLAVA [59]7B----40.2
VideoLLaMA2 [12]8B56.045.442.147.9-
LongVA [70]7B61.150.446.252.6-
Ours
SALOVA-3B3B48.346.341.145.341.4
SALOVA-3.8B3.8B47.148.844.146.741.6
SALOVA-7B7B59.450.449.453.144.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Table 1: Detailed results for the Video-MME benchmark (w/o subtitles) and LongVideoBench. The best results are highlighted in **bold** and the runner-up results are **underlined**." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.483, + 0.452 + ], + "angle": 0, + "content": "the memory footprint associated with model components (i.e., gradient, activation, and optimizer states). Additionally, we fine-tune the trainable parameters at each step without employing LoRA [24]. For the extended training config, we have attached the details in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.456, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Evaluation Benchmarks. We evaluate our model using two types of video analysis benchmarks—long video understanding and general video understanding, categorized based on the video length. For the long video benchmark, we primarily utilize Video-MME [20] and LongVideoBench [57], both of benchmarks include videos up to two hours long duration. As the general video analysis evaluations, we employ various benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32]. Note that the same pipeline is used to obtain video segments for each benchmark, and all benchmarks are sampled at 1 FPS without token drop during inference. As comparison baselines, considering academic budget constraints, we evaluate against models that have similar parameter size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.292, + 0.695 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on Long Video Understanding. Video-MME [20] evaluates LMMs with a focus on video analysis across a variety of video types and durations. We primarily compare the benchmark results in settings without subtitles, relying solely on video frames. Therefore it can assess the LMMs' visual comprehension capabilities rigorously, based purely on visual content. Also, LongVideoBench [57] is designed to assess LMMs' understanding of long-duration videos up to two hours. It includes a diverse collection of videos, challenging the models' ability to process and interpret extensive visual and contextual information across a variety of themes. As shown in Tab. 1, our model shows competent video understanding performance across all video" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.287 + ], + "angle": 0, + "content": "
Model#paramActivityNetQAVideoChatGPTMVBench
test (acc/score)test (acc)test (acc)
Proprietary LMMs
GPT-4V [45]n/a57.0-4.0643.5
GPT-4o [46]n/a61.9---
Gemini 1.5 Pro [50]n/a57.5---
Open-sourced LMMs
VideoLLaMA [68]7B12.41.12.1634.1
VideoChatGPT [41]7B35.22.72.4232.7
MovieChat [53]7B45.7-2.67-
Chat-UniVi [26]7B46.13.22.99-
LLaMA-VID [34]7B47.43.32.8941.3
VideoChat2 [32]7B49.13.32.9851.1
VideoLLaMA2 [12]8B50.23.33.1354.6
Ours
SALOVA-3B3B52.63.43.0851.7
SALOVA-3.8B3.8B51.13.52.8346.4
SALOVA-7B7B53.63.53.1353.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.288, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Table 2: Comparison results for generic video understanding benchmarks. The best results are highlighted in bold and the runner-up results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.453 + ], + "angle": 0, + "content": "length distributions in Video-MME and lengthy video instances in LongVideoBench. Notably, we highlight that SALOVA achieved significant performance in the medium (average 562.7 seconds) and long (average 2385.8 seconds) length categories in Video-MME benchmark, even with more smaller size of backbone LM parameters compared with the baseline models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.624 + ], + "angle": 0, + "content": "Such performance gains in long video instances are attributed to our model's dynamic capability to retrieve and process only the relevant video segments, enabling it to handle lengthy video content efficiently without being constrained by the limited context length. Especially, the routing mechanism in SALOVA strategically prioritizes video segments that are likely to contain crucial visual and contextual cues relevant to the query. This selective routing mechanism reduces the computational load and minimizes the information loss that commonly occurs in current videolMMs trying to process extensive video data in entirety." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results on General Video Understanding. For general video understanding benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32], SALOVA was evaluated across various video types to assess its general video understanding capabilities. As shown in Tab. 2, SALOVA demonstrated competent performance, comparable to existing video-LMMs, especially in dynamic and shorter video sequences. On ActivityNetQA, the model effectively utilized its segment retrieval strategy to provide focused and contextually appropriate responses, which helped maintain accuracy. This approach was similarly effective in the multi-modal settings of VideoChatGPT and MVBench, where SALOVA showed consistent performance in handling dialogues and visual cues. These outcomes highlight SALOVA's capability to process general video content efficiently through its dynamic routing mechanism, offering a reliable solution that balances computational resources with output quality." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3358" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.25 + ], + "angle": 0, + "content": "
AblationVideo-MME
Short: ≤2mMid: 4-15mLong: 30-60mOverall
frm sample : Frame sampling rate (w/o SR-Router)
8 frm48.342.037.242.5
16 frm50.042.838.043.6
1 FPS48.346.341.145.3
1 / 1.5 / 2 : Train stage - Long video knowledge injection
✓ × ✓45.643.740.243.6
✓ ✓ ✓48.346.341.145.3
FocusFast : Local-global video representation
×36.438.635.636.9
✓48.346.341.145.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.251, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Table 3: Ablation studies on SALOVA configuration. We utilize SALOVA-3B for the resource efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.295, + 0.383, + 0.312 + ], + "angle": 0, + "content": "5.3. Additional Analyses on SALOVA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.398 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablation studies on three components as follows: (i) different video frame sampling strategies, (ii) intermediate training stage for long video knowledge injection, and (iii) the FocusFast mechanism to understand branched local-global representation in videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.7 + ], + "angle": 0, + "content": "As in Table 3, We first observe that using more frames with SR-Router significantly enhances performance, particularly in long-form videos. This aligns with our key contributions on managing long videos through an effective routing mechanism, suggesting that retrieving more frames can provide richer spatio-temporal information and improve the model's responses without losing contextual coherence. Additionally, we compare with a baseline trained with stage 1-2 (skipping stage 1.5). Here, we highlight the effectiveness of the SceneWalk dataset as an intermediate training step to enhance parametric knowledge for the long video analysis by allowing the model to learn from high-quality and densely captioned scene-level information, which is crucial for adapting to various lengths and contexts. Lastly, we conduct an analysis on the FocusFast method and demonstrate its efficacy in analyzing not only local details from relevant video segments but also in understanding the global video context through the simultaneous use of routing tokens, thereby facilitating a more comprehensive understanding of video content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Analysis of Retrieving Segments. By retrieving relevant video segments for the given queries, SALOVA can effectively target salient information in the long video and retain long context information. To further demonstrate the model's targeting capabilities beyond numerical performance in long video analysis, we explore our model's application in the Visual Needle-In-A-Haystack (V-NIAH) task [70], which extends the Needle-in-a-Haystack (NIAH) evaluation for LLMs to a vision-level benchmark. This task is particularly challenging as it requires models to not only detect but also precisely retrieve the sparse yet crucial visual cues scattered across lengthy videos." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As in Fig. 3, we compare SALOVA against a base" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.219, + 0.904, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.348, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Figure 3: Comparison results of V-NIAH. The x/y-axis indicates the total video frames and the location of needle image within the video, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.906, + 0.51 + ], + "angle": 0, + "content": "line trained on sparsely sampled frames (16 frm w/o SR-Router). Our framework effectively identifies and extracts relevant video segments from densely packed content, even when handling long context lengths. These results highlight SALOVA's robustness in managing complex, long-form videos, maintaining contextual continuity by strategically focusing on relevant segments to user queries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.521, + 0.763, + 0.537 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Discussion. Despite SALOVA's competence in handling extended video sequences, it is important to recognize scenarios where its complex architecture may not be necessary. Specifically, for shorter videos where sparse sampling suffices to capture essential spatio-temporal information, simpler models could potentially outperform the efficiency of SALOVA without necessitating its extensive processing capabilities. This suggests a future avenue for integrating a hybrid approach based on our framework by dynamically adjusting the complexity of the retrieval and processing mechanisms based on the video length and content density." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Conclusion. In this paper, we introduce SALOVA, a novel framework designed to enhance the comprehension of long and untrimmed video by leveraging a retrieval-driven approach with new densely captioned dataset, the SceneWalk dataset. SALOVA strategically targets and processes only the relevant video segments, effectively addressing the structural limitations of current Video-LMMs with its Spatio-Temporal Connector and Segment Retrieval Router. Through extensive evaluation on various benchmarks, SALOVA exhibits its robust performance in interpreting complex video content, enhancing efficiency, and improving the understanding of extended videos." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3359" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.193 + ], + "angle": 0, + "content": "This work was partially supported by two funds: IITP grant funded by the Korea government (MSIT) (RS-2022-II220984) and IITP grant funded by the Korea government (MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-Term Visual Memory Network)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[7] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. arXiv preprint arXiv:2312.06742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[8] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 4, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[9] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[15] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[16] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[20] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[21] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (sfd): A benchmark for story-level video understanding. arXiv preprint arXiv:2406.10221, 2024. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.699, + 0.872 + ], + "angle": 0, + "content": "[22] Google. Gemini, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3360" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[25] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. Language is not all you need: Aligning perception with language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[26] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[28] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.801, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[33] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. Advances in neural information processing systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[34] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 323-340. Springer, 2025. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[35] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[39] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2025. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[40] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[42] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[43] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.805, + 0.632 + ], + "angle": 0, + "content": "[44] OpenAI. Gpt-4 technical report, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.826, + 0.647 + ], + "angle": 0, + "content": "[45] OpenAI. GPT-4V(ision) System Card, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.75, + 0.662 + ], + "angle": 0, + "content": "[46] OpenAI. Hello gpt-4o, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.663, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[47] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[49] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3361" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[50] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.176, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[51] Share. Sharegemini: Scaling up video caption data for multimodal large language models, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[53] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[54] Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. Augmented SBERT: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 296–310, Online, 2021. Association for Computational Linguistics. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[56] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[57] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. arXiv preprint arXiv:2407.15754, 2024. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[58] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[61] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[62] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[63] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[64] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[65] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[66] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[67] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[69] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[70] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[72] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[73] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3362" + } + ] +] \ No newline at end of file diff --git a/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_origin.pdf b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7d991a6c75ab6d0d39c65e32fb5a37ef16da5d8 --- /dev/null +++ b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/df919c36-d04e-4ca1-a80f-80ab910d4f79_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f03a56e1d97ee0ecee08141925d379d4c79d800c640ef030428f53f3995e44b8 +size 2138741 diff --git a/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/full.md b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fbf1f80c32a9c1e453db8d21fc6a3bd7df7e8283 --- /dev/null +++ b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/full.md @@ -0,0 +1,287 @@ +# SALOVA: Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis + +Junho Kim* Hyunjun Kim* Hosu Lee Yong Man Ro† Integrated Vision and Language Lab, KAIST, South Korea {arkimjh, kimhj709, leehosu01, ymro}@kaist.ac.kr + +# Abstract + +Despite advances in Large Multi-modal Models, applying them to long and untrimmed video content remains challenging due to limitations in context length and substantial memory overhead. These constraints often lead to significant information loss and reduced relevance in the model responses. With the exponential growth of video data across web platforms, understanding long-form video is crucial for advancing generalized intelligence. In this paper, we introduce SALOVA: Segment-Augmented LOng Video Assistant, a novel video-LLM framework designed to enhance the comprehension of lengthy video content through targeted retrieval process. We address two main challenges to achieve it: (i) We present the SceneWalk dataset, a high-quality collection of 87.8K long videos, each densely captioned at the segment level to enable models to capture scene continuity and maintain rich descriptive context. (ii) We develop robust architectural designs integrating dynamic routing mechanism and spatio-temporal projector to efficiently retrieve and process relevant video segments based on user queries. Our framework mitigates the limitations of current video-LMMs by allowing for precise identification and retrieval of relevant video segments in response to queries, thereby improving the contextual relevance of the generated responses. Through extensive experiments, SALOVA demonstrates enhanced capability in processing complex long-form videos, showing significant capability to maintain contextual integrity across extended sequences. + +# 1. Introduction + +Recent advancements in Large Language Models (LLMs) [22, 43, 44] have brought us one step closer to achieving Artificial General Intelligence (AGI). Next following step, the current trend is shifting toward modular systems that integrate various multi-modality, leveraging + +the exceptional generalization and reasoning capabilities of LLMs to evolve into Large Multi-modal Models (LMMs). Accordingly, users can unrestrictedly interact with the models across various modalities beyond text, expanding the scope of machine understanding and enhancing user engagement. Especially, considering the widespread adoption of long-form videos across various web platforms, the importance of understanding long, untrimmed video has become increasingly prominent in the multi-modal domain. + +After the pioneer works [14, 37, 38] utilizing visual instruction tuning to augment vision perception into LLMs, remarkable strides [11, 17, 65] have been made in aligning cross-modal consistency—especially between vision and language domains. Albeit more recent models [29, 69] integrate various vision modalities all at once, current approaches still face significant challenges in understanding untrimmed and long-form video content. The main challenge is attributed to the limited context length of LMMs, which is an inherent structural limitation that restricts the models to process only a finite number of tokens as the input sequences. We can exemplify that LLaVA series [29, 30], when processing a video data, require 144 visual tokens per each frame, where numerical approximation is only maximum of $\sim 56$ frames using 8K max context length LMMs, which is still limited to handle long sequence data. + +Accordingly, current video-LMMs [12, 35, 39] rely on (i) sparse frame sampling to represent entire videos [26, 72], (ii) dense compression of visual tokens into a smaller size to manage the excessive number of frames [34, 41], and (iii) adaptive pooling strategies [59, 60] based on the SlowFast approach [19], all aimed at fitting the long video sequences within the limited context window of LMMs. Several studies focusing on the long video understanding task have presented memory-augmented generation [23, 53] utilizing an additional buffer to embed long-term information, or have extended the context using RoPE-based frequency extension during the training [70]. Despite of such endeavors, when handling massive video frames, previous works still confront restricted context size and significant memory overhead, which leads to substantial visual information + +loss. As critical events may be overlooked by the models, this hinders their ability to fully capture context changes in lengthy videos, resulting in inaccurate and irrelevant responses for the user queries. + +Starting from the intuitive insight outlined below, in this paper, we propose a retrieval-driven approach for long video understanding with LLMs. Analogous to the recent Retriever-Augmented Generation (RAG) systems [28] (widely adopted in LLMs), which retrieve relevant information from external factual knowledge, humans naturally employ similar strategies when seeking specific information, efficiently locating and referring necessary materials to answer targeted questions—e.g., imagine that we are taking open-book exams or searching for a certain recipe in a cookbook. Given a long and untrimmed video, mirroring the targeted retrieval processes, we introduce a novel framework, Segment-Augmented LOng Video Assistant (SALOVA) to effectively handle the long sequence visual inputs by retrieving the relevant video segments. + +To construct our video segments retrieval framework, central challenge hinges on establishing two main components: (i) Densely captioned video data, which consists of video-caption pairs with progressively-captioned descriptions that change throughout each video used to train the model to accurately identify relevant video segments. (ii) Dynamic routing mechanism, which selects pertinent video segments for the queries, followed by being connected to LLMs. To address that, our approach is outlined as follows. + +Data. (§3) Recently several video-text paired datasets [4, 6, 8, 10, 56, 62] have been released, but they are inadequate for handling long and untrimmed video data, where only partial video moments are described with limited word length as compared in Fig. 1(a). To handle such insufficiency of detailed descriptions within the videos and the short durations of both videos and texts, we introduce the SceneWalk dataset, a new high-quality video dataset with thorough captioning for each video. It includes dense and detailed descriptions for every video segment across the entire scene context. The SceneWalk dataset, sourced from long and untrimmed 87.8K YouTube videos (avg. 486 seconds each), features frequent scene transitions across a total of 11.8K hrs video duration and 1.3M massively segmented video clips. Each video segment in the dataset is provided with a detailed description (avg. 137.5 word length), generated by combining pre-trained models [54, 73] and manual curation from human. + +Architecture. (§4) Utilizing the constructed video dataset, SALOVA learns to identify relevant video segments for the given queries within each video source and then auto-regressively predicts the next token. To do so, we present two architectural designs to seamlessly incorporate the retrieved segments in an end-to-end training: Spatio-Temporal Connector and Segment Retrieval Router. By fo + +cusing on the relevant segments, our framework can perform deeper reasoning without being constrained by context length limitations. Additionally, we present FocusFast approach, which intensively analyzes the selected segments for detailed comprehension (focus pathway), while quickly accessing overall contextual information with routing tokens obtained from the entire video segments (fast pathway). The strategy ensures SALOVA to maintain comprehensive video understanding while prioritizing details where it is most needed, effectively enhancing long and untrimmed video interpretation. + +Through extensive experiments and analyses, we corroborate that competitive performance of SALOVA to the existing video-LMM models in understanding complex long-form videos. Also, our results show significant reductions in the loss of crucial visual information and a lower risk of omitting important events, demonstrating the effectiveness of our proposed method across various video benchmarks. + +Our contribution can be summarized into three-fold: + +- We introduce the SceneWalk dataset, a high-quality and densely-captioned video dataset with detailed segment-level descriptions from 87.8K long and untrimmed video sources. The proposed dataset provides rich context and scene continuity, enabling effective training for long-form video understanding. +- We propose Segment-Augmented LOng Video Assistant (SALOVA), a novel video-LMM framework designed to enhance long video comprehension by targeting relevant video segments in lengthy videos, optimizing the model's focus on essential segment targets for the given queries. +- Through extensive evaluation, we validate that SALOVA improves overall long video understanding capabilities by effectively integrating relevant video segments, thus optimizing to handle long and untrimmed video content. + +# 2. Related Work + +# 2.1. Large Multi-modal Models + +After the emergence of LLMs [5, 55], which can actively interact with users through back-and-forth conversations, as a next leap, various research efforts [2, 25, 31] integrate different modalities into the LLMs, utilizing their core reasoning and zero-shot capabilities. Building on the open-sourced models [13, 55], seminal works [14, 38, 65] have bridged image and text modality under the visual instruction tuning and presented multi-modal assistant models that possess visual perception and QA capabilities. Since then, numerous research studies have been introduced to (i) enhance vision understanding with advanced architectures [17] or higher resolutions [30, 37], (ii) implement more sophisticated alignment layers [7, 42] between modalities, and (iii) train the models with more high-quality data and larger model parameters. + +![](images/115411666638185eab77d20c4648d4db87a8318b2b9f8c5453a123aed7b53ac5.jpg) +(a) Video-Text Dataset Comparison + +![](images/2aca787b5795f792bf26420366df55c4fd7b063a91006caa4fd7148418b58bdf.jpg) +(b) SceneWalk Statistics + +![](images/a4c4cbb8373c76c5c6851d3f3659752931644de12660eb98c28b8f161aba9adc.jpg) +(c) Overall Pipeline for Data Collection +Figure 1: The overview of the SceneWalk dataset includes (a) dataset comparison, (b) detailed statistics, and (c) the annotation pipeline for description and score collection. Note that the scale of circles in Fig. 1(a) indicates the data size, and the color distribution in Fig. 1(b) denotes the video duration in each video category—brighter colors correspond to shorter video durations. Further details about the dataset are provided in Appendix A. + +Recent focus has shifted towards more unified modality processing following the release of omnivorous models [46]. Some recent omni-versions of LMMs [29, 64] can handle combinatorial subsets from various modality sources, such as images, videos, audio, speech, and depth. However, the current LMMs for video [32, 35, 41] still lack of capturing the necessary details to effectively process video information due to their sparse frame sampling strategy. While such approach is seemingly adequate for relatively shorter videos, it may fail to capture comprehensive spatio-temporal information, potentially compromising the accuracy of model responses to user queries. In this paper, SALOVA first retrieves relevant video segments, then concentrates on more granular video cues. Such targeted focus allows the model to effectively understand complex analysis within the videos, significantly improving its ability to provide contextual-aware and accurate responses. + +# 2.2. Long Video Understanding + +In parallel, we dually introduce video-specialized models [12, 23, 53] integrated with LLMs, which have also been widely explored these days to enhance video understanding and reasoning. Here, the most challenging part of current video-LMMs lies in handling long video sequences, mainly due to the limited context length of the LLMs. This limitation compels the models to sparsely sample the video frames in only limited sizes (e.g., typically 8 or 16 frames), potentially missing important spatial and temporal information. To address this, several studies have focused on compressing visual tokens into a more manageable size, proposing aggregation [34, 41] or pooling methods [35, 59] with advanced vision encoder structures [67, 73]. In addition, memory-augmented methods [23, 53] first stored long-term information in a memory bank, then responded to specific queries by loading memory features from the stored buffer. + +On the other hand, among more recent approaches, Li + +et al. [70] have directly extended the LLMs' context length by exploiting RoPE-based frequency interpolation, and Xue et al. [61] have introduced sequence parallelism that can be implemented on multiple GPUs by modifying backend systems. However, we argue that such approaches inherently cannot be free from the fixed context length and provoke intensive memory demands when processing more longer videos. Instead, by focusing on the relevant segments within the entire video, SALOVA can efficiently handle the limited context length, enabling targeted processing of key moments without the need for excessive memory, thereby enhancing performance on longer video sequences. + +# 3. SceneWalk Dataset + +In this section, we elaborate on how we collected the SceneWalk dataset. The overall pipeline for building the dataset and summarized statistics are illustrated in Fig. 1. While several video SFT datasets [32, 41] are widely used during the instruction tuning stage, they often fail to capture comprehensive details within the scenes. This stems from the nature of instruction-type questions, which provide only partial information, and the brief lengths of both videos and texts in QAs. In contrast, the SceneWalk dataset offers densely captioned video-text pairs that cover long sequence videos in full details, as shown in Fig. 1(a). For further detailed data statistics, please see Appendix A. + +# 3.1. Data Gathering and Processing + +Video Source & Filtering. For the long and untrimmed video sources, we primarily focus on three key aspects to build densely captioned video dataset: (i) extensive video length with diverse video source categories, (ii) high-quality video contents, (iii) frequent scene transitions within each video. Accordingly, our data collection is mainly sourced from YouTube, ensuring rich dynamic content that bet- + +ter reflects real-world complexities experienced by global users—here, because our main goal for video gathering is on complex scene understanding, we exclude low-quality and user-uploaded aesthetic videos (e.g., WebVid, Pixabay, Pexels, and etc.,) that are rather beneficial for video generation tasks, despite their merits for easy collection. We have collected YouTubeurls from [27] and downloaded the whole video in untrimmed states. Among the total 32 coarse and diverse video categories YouTube API provided, we selectively curated 10 categories, excluding categories such as News & Politics, Classics, and Documentary, due to their static nature, which provides only sparse temporal information in the videos. We further supplemented the dataset with additional Movie & Drama videos sourced from [21, 53], totaling 87,867 video sources with 11.87 Khrs video duration (avg. 486.5-seconds). + +Segmenting Video into Clips. Next, for the collected long and untrimmed video sources, we cut the lengthy videos into small segments to densely caption the entire video in next phase. Instead of adopting the bottom-up approach used in the ShareGPT4Video dataset [8], which segments videos into fixed time intervals (2-seconds) in advance and then merges adjacent frames based on their CLIP similarity [48], we directly employed PysceneDetect 1 to segment the videos, dynamically adjusting the threshold based on the raw-level video information to reliably detect scene changes. At the end, the total number of 1.29M of video segments with 33.11-seconds average video length is extracted from the original video sources. + +# 3.2. Captioning and Scoring + +Dense Segment Captioning. After obtaining the massive video segments, our next goal is to caption each segment with visual details and narrative context to capture the scene-specific explanations, which can enrich scene-level interpretation. To achieve this, we plan to utilize pre-trained LMMs to generate detailed descriptions for the partial video segments. As the captioner, we empirically found that VILA-1.5 (13B) [36] shows competent descriptive quality than other open-sourced models, and used to generate dense captions for each video segment with randomly sampled instructions for detailed descriptions. As a result, we acquire 1.29M pairs of detailed descriptions corresponding to the video segments, each description with average 137.5 word length. Please see instruction details and qualitative examples of generated captions in Appendix A. + +Scoring Video-Text Correspondence. Lastly, we score the correspondence between the video segments and the paired dense descriptions, which will later be used as explicit supervision to robustly train our retrieval framework. What + +we must not overlook here is that the paired video-text relationship is not solely a one-to-one correspondence but is more akin to generalized bipartite matching. That is, within the long and untrimmed video source, each video segment can be connected to other descriptions with additional edges. Therefore, for the $N_v$ number of video segments and their paired segments, we can construct a $\{N_v\}^2$ correspondence matrix between video-text (V2T). To measure each correspondence, we employ LanguageBind [73] due to its competitive alignment capabilities across various modalities. In addition, we build another $\{N_v\}^2$ matrix to provide a doubly robust measure for the correspondence scores among adjacent descriptions (T2T) by comparing similarity within the textual context using the SBERT model [54]. + +# 4. Segment-Augmented LOng Video Assistant + +Network Overview. For a given set of $N_v$ video segments sampled at 1 FPS $v = \{v_i\}^{N_v}$ , where each segment $v_i \in \mathbb{R}^{T_i \times H \times W \times C}$ has varying video length (summing up to the total time $T$ of a long and untrimmed video), SA-LOVA consists of four main architecture components as illustrated in Fig. 2: + +- Vision Encoder: We use CLIP [48] or SigLIP [67] to extract visual features, followed by 2x2 average pooling, resulting in 144 or 196 visual tokens per frame. +- Spatio-Temporal Connector: To handle spatio-temporal features of varying lengths from the vision encoder, we employ the Perceiver Resampler [2], which consists of 2-layer Transformer architecture followed by a 2-layer MLP with GELU activation as projector. This resampler embeds each video segment feature into fixed size latent features that are connected to LLMs. +- Segment Retrieval Router: For the given textual queries, a retrieval structure (2-layer Transformer) gathers representative information (i.e., routing tokens) from each video segment and then routes the query-relevant video features into the LLMs. Note that the router architecture is trained in an end-to-end manner. +- Large Language Model: For LM backbones, We utilize three open-sourced LLMs with varying parameter sizes, LLaMA-3.2 (3B) [18], Phi-3.5 (3.8B) [1] and Qwen-2.5 (7B) [63], all of which are instruction-tuned models that possess QA assistant capabilities. + +# 4.1. Long Video Processing and Pipeline + +# 4.1.1. Spatio-Temporal Connector + +The first component of our model, Spatio-Temporal Connector, efficiently handles long and variable-length input video segments by extracting each segment's visual semantics in a fixed-size latent vector. As illustrated in Fig. 2(b), we first sample video frames at 1 FPS from each video segment, then visual features are acquired with $2 \times 2$ pooling + +![](images/c393cd9f4355eeb974eb9dbfc2dad22383bb863765689a7956ae0455230c5844.jpg) +Figure 2: The network overview of SALOVA. Our framework consists of four structural components: vision encoder, ST-connector, SR-router, and LLMs. Using the FocusFast strategy, our model can concentrate on more detailed local information while maintaining context awareness. + +(thus, 196 tokens from each frame). After that, the visual features are flattened and fed into the ST-Connector with additional positional and temporal encoding. Here, when the long video is processed, the number of unfolded patch tokens becomes extremely large, leading to exhaustive computations. To address this, we employ a dynamic token drop technique to reduce computational load. + +Dynamic Token Drop. To effectively manage long video sequences, the token drop has been utilized in video generation tasks [16, 40]. Expanding such approach, in our framework, the dropout rate is dynamically adjusted based on the length of the input sequence $T_{i}$ in the input visual feature $f_{v} \sim T_{i} \times H_{p}W_{p} \times d$ , which allows for more efficient processing of longer sequences by reducing computational demands, while still preserving dense visual semantics in shorter videos. Additionally, to retain spatio-temporal information from the dropped patches, we add positional embeddings separately along the spatial and temporal axes. This enables more refined extraction of spatio-temporal visual semantics even after reducing the number of tokens. + +# 4.1.2. Segment Retrieval Router + +Next, the key to conveying the pertinent video information to LLMs is retrieving relevant video segments by querying sentence. To densely cue the similarities between the video and sentence information, we introduce a routing framework, Segment Retrieval Router, which consists of 2-layer Transformer as illustrated in Fig. 2(c). After obtaining the routing tokens $R = \{R_{i}\}^{N_{v}} \in \mathbb{R}^{N_{v} \times D}$ from entire video segments, we aggregate them and feed into the SRRouter as queries. For the given sentence, we employ the + +same text encoder used for the vision encoder and project it into the shared embedding space to obtain sentence features $S \in \mathbb{R}^{N_t \times D}$ , where $N_{t}$ indicates textual length. + +Using the cross attention mechanism (q: $R$ ; k/v: $S$ ), we can estimate similarity scores between the video segments and given sentence queries (i.e., V-T similarity). The scores enable the SR-Router to prioritize and select the most relevant video segments that align with the sentence query. + +Retrieval Objective. To seamlessly train the SR-Router with the mainstream flows of SALOVA in an end-to-end manner, we have designed a similarity loss function $\mathcal{L}_{\mathrm{sim}}$ that minimizes the distance between the high-dimensional embeddings of the video segments and sentence queries. Here, we use the correspondence scores (aforementioned in Sec. 3.2) as a retrieval supervision signal $y_{i}$ , after applying one-hot encoding. We incorporate a simple margin-based loss, commonly used in contrastive learning settings [33], which enables the model to learn off-diagonal relaxation in the correspondence matrices between video segments and sentences. As mentioned earlier, the relationship between paired videos and sentences is closer to generalized bipartite matching than to one-to-one matching, so relaxation learning helps to accommodate the inherent complexity in aligning correspondence. In conclusion, with the binary cross-entropy loss and the score margin loss, we can formulate the similarity loss as follows: + +$$ +\mathcal {L} _ {\text {s i m}} = \underbrace {\mathcal {L} _ {\text {b c e}} \left(y _ {i} , s _ {i}\right) _ {i = 1} ^ {N _ {v}}} _ {\text {p o i n t - w i s e C E}} + \underbrace {\frac {1}{N _ {s}} \sum_ {j} \max \left(0 , \delta - \left(s _ {j} ^ {p} - s _ {j} ^ {n}\right)\right)} _ {\text {s c o r e m a r g i n l o s s}}, \tag {1} +$$ + +where $s_j^p$ and $s_j^n$ indicate randomly sampled scores from + +positive and negative pairs, respectively, and $\delta$ denotes the margin parameter (set as 0.2). Note that the similarity loss is trained in conjunction with the auto-regressive loss $\mathcal{L}_{\mathrm{ar}}$ from subsequent LLMs in an end-to-end manner. + +# 4.1.3. FocusFast Pathways: Integration to LLMs + +Using the routing tokens, we can calculate the similarities of each video segment for the given query. Leveraging the similarities, SALOVA efficiently retrieves the specific video features that exhibit the highest relevance score to the textual query, where the indexed video features are then directly integrated into the LLM architecture. Here, extending the SlowFast pathways concept [19], we present the FocusFast mechanism to effectively manage the processing pathway for the retrieved video segments: (i) Focus pathway concatenates the top-K most pertinent features to construct a comprehensive video representation, capturing local details across retrieved segments and enabling detailed interactions with textual queries to enhance handling complex video information. (ii) Fast pathway focuses on the more broader-level context by employing segment-wide routing tokens as the condensed global representation. It effectively contains dynamic spatio-temporal changes throughout the video stream, thereby allowing SALOVA to understand the overall video content and scene-level continuity awareness. + +Once the most pertinent features are retrieved, they are delivered to the LM backbone for the final processing as in Fig. 2(a), integrating video-specific details into the models' responses. By effectively handling long and untrimmed videos with the proposed retrieval and routing mechanism, SALOVA can maintain the flow of salient information without the processing overhead for less related data, thus generating more context-aware responses. + +# 4.2. Training Strategies + +The current training strategies for LMMs predominantly consist of two-step training: (i) cross-modal alignment and (ii) visual instruction tuning. Recently, Li et al. [29] have emphasized the importance of high-quality knowledge learning between the two training stages (thus stage 1.5), pointing out that the models cannot enoughly learn necessary knowledge during the alignment with the low-qualitative web-scale image-text data. As the similar approach of using rephrased descriptions for additional knowledge learning [29], we employ the newly collected SceneWalk dataset as the parametric knowledge injection step, which enables the SALOVA to learn detailed spatial and temporal representation from the long sequence video data before the instruction tuning. Accordingly, our training recipes and data configuration can be divided into three steps as follow (Please see training details in Appendix B): + +Stage 1: Cross-modality Alignment. For the initial step in modality alignment, we utilize 790K image/video-text + +paired dataset: (i) 558K image-text pairs from the CC3M dataset [52], filtered by LLaVA [38] and (ii) video-text pairs sampled from the WebVid 2.5M subset [4]. We freeze vision encoder and LLMs during the training, and mainly focus on optimizing the connector and router to map the visual information into the textual space. + +Stage 1.5: Long Video Knowledge Injection. As an intermediate training step, we use the SceneWalk dataset to train the SALOVA, unfreezing all trainable parameters except for the vision encoder. During training, we input the long and untrimmed video instances and follow the processing pipeline shown in Fig. 2. By training the model with densely captioned video-description pairs, it acquires high-quality parametric knowledge of both spatial and temporal information. In addition, through the aforementioned retrieval process, the model learns to target video segments that are mostly relevant to the video description. + +Stage 2: Video Instruction Tuning. To possess QA capabilities in SALOVA, we use extensive video instruction-tuning data as the final training step. The instruction data are mainly sourced from four different datasets: LLaVA-Video-178K [71], NeXT-QA [58], ActivityNetQA [66], and PerceptionTest [47]—comprising a total of 1.4M video-instruction QA data, including caption entries, open-ended QA, and multiple-choice QA. Note that we train all the network parameters during this stage and auto-regressively update the instruction-following assistant's response for the next word prediction. + +# 5. Experiments + +# 5.1. Experimental Details + +Implementation. For the vision and text encoders of the SR-Router, we utilize CLIP [48] for the small-size models and SigLIP [67] for the frontier model, with resolution sizes of 336 and 384, respectively. We employ a 2-layer transformer with a head size of 2 for the ST-Connector, which has a latent dimension of 256. The token drop mechanism is dynamically applied based on video length, with varying maximum drop rates at each training stage—Stage 1 has no token drop, Stage 1.5 allows up to 0.7, and Stage 2 allows up to 0.4. For the configuration of SR-Router, we set 2-layer of transformers with a single head, and top-K number is set to 5 during the stage 2. Following the recent works [30, 37], the projector layer consists of 2-layer MLP with GELU. We employ three LLMs: (i) 3B: Llama-3.2-3B [18], (ii) 3.8B: Phi-3.5 [1], and (iii) 7B: Qwen2.5-7B [63]. + +Training Details. For the each training stage, we train SA-LOVA for 1 epoch with 1 node of 8 A100 GPUs. The total training hours for 3B (3.8B) and 7B models roughly take 5 and 7 days, respectively. We employ FlashAttention-2 [15], gradient checkpointing [9], and ZeRO-2 [49] to minimize + +
Model#paramVideo-MMELVBench Acc. (val)
ShortMediumLongOverall
Proprietary LMMs
GPT-4V [45]n/a70.555.853.559.9-
GPT-4o [46]n/a80.070.365.371.966.7
Gemini 1.5 Pro [50]n/a81.774.367.475.064.0
Open-sourced LMMs
ST-LLM [39]7B45.736.831.337.9-
VideoChat2 [32]7B48.337.033.239.539.3
ShareGPT4Video [8]8B48.336.335.039.939.7
Video-LLaVA [35]7B45.338.036.239.939.1
Chat-UniVi-V1.5 [26]7B45.740.335.840.6-
Qwen-VL-Chat [3]7B46.938.737.841.1-
ShareGemini [51]7B49.141.339.143.2-
SliME [72]8B53.342.739.845.3-
PLLAVA [59]7B----40.2
VideoLLaMA2 [12]8B56.045.442.147.9-
LongVA [70]7B61.150.446.252.6-
Ours
SALOVA-3B3B48.346.341.145.341.4
SALOVA-3.8B3.8B47.148.844.146.741.6
SALOVA-7B7B59.450.449.453.144.6
+ +Table 1: Detailed results for the Video-MME benchmark (w/o subtitles) and LongVideoBench. The best results are highlighted in **bold** and the runner-up results are **underlined**. + +the memory footprint associated with model components (i.e., gradient, activation, and optimizer states). Additionally, we fine-tune the trainable parameters at each step without employing LoRA [24]. For the extended training config, we have attached the details in Appendix C. + +Evaluation Benchmarks. We evaluate our model using two types of video analysis benchmarks—long video understanding and general video understanding, categorized based on the video length. For the long video benchmark, we primarily utilize Video-MME [20] and LongVideoBench [57], both of benchmarks include videos up to two hours long duration. As the general video analysis evaluations, we employ various benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32]. Note that the same pipeline is used to obtain video segments for each benchmark, and all benchmarks are sampled at 1 FPS without token drop during inference. As comparison baselines, considering academic budget constraints, we evaluate against models that have similar parameter size. + +# 5.2. Experimental Results + +Results on Long Video Understanding. Video-MME [20] evaluates LMMs with a focus on video analysis across a variety of video types and durations. We primarily compare the benchmark results in settings without subtitles, relying solely on video frames. Therefore it can assess the LMMs' visual comprehension capabilities rigorously, based purely on visual content. Also, LongVideoBench [57] is designed to assess LMMs' understanding of long-duration videos up to two hours. It includes a diverse collection of videos, challenging the models' ability to process and interpret extensive visual and contextual information across a variety of themes. As shown in Tab. 1, our model shows competent video understanding performance across all video + +
Model#paramActivityNetQAVideoChatGPTMVBench
test (acc/score)test (acc)test (acc)
Proprietary LMMs
GPT-4V [45]n/a57.0-4.0643.5
GPT-4o [46]n/a61.9---
Gemini 1.5 Pro [50]n/a57.5---
Open-sourced LMMs
VideoLLaMA [68]7B12.41.12.1634.1
VideoChatGPT [41]7B35.22.72.4232.7
MovieChat [53]7B45.7-2.67-
Chat-UniVi [26]7B46.13.22.99-
LLaMA-VID [34]7B47.43.32.8941.3
VideoChat2 [32]7B49.13.32.9851.1
VideoLLaMA2 [12]8B50.23.33.1354.6
Ours
SALOVA-3B3B52.63.43.0851.7
SALOVA-3.8B3.8B51.13.52.8346.4
SALOVA-7B7B53.63.53.1353.5
+ +Table 2: Comparison results for generic video understanding benchmarks. The best results are highlighted in bold and the runner-up results are underlined. + +length distributions in Video-MME and lengthy video instances in LongVideoBench. Notably, we highlight that SALOVA achieved significant performance in the medium (average 562.7 seconds) and long (average 2385.8 seconds) length categories in Video-MME benchmark, even with more smaller size of backbone LM parameters compared with the baseline models. + +Such performance gains in long video instances are attributed to our model's dynamic capability to retrieve and process only the relevant video segments, enabling it to handle lengthy video content efficiently without being constrained by the limited context length. Especially, the routing mechanism in SALOVA strategically prioritizes video segments that are likely to contain crucial visual and contextual cues relevant to the query. This selective routing mechanism reduces the computational load and minimizes the information loss that commonly occurs in current videolMMs trying to process extensive video data in entirety. + +Results on General Video Understanding. For general video understanding benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32], SALOVA was evaluated across various video types to assess its general video understanding capabilities. As shown in Tab. 2, SALOVA demonstrated competent performance, comparable to existing video-LMMs, especially in dynamic and shorter video sequences. On ActivityNetQA, the model effectively utilized its segment retrieval strategy to provide focused and contextually appropriate responses, which helped maintain accuracy. This approach was similarly effective in the multi-modal settings of VideoChatGPT and MVBench, where SALOVA showed consistent performance in handling dialogues and visual cues. These outcomes highlight SALOVA's capability to process general video content efficiently through its dynamic routing mechanism, offering a reliable solution that balances computational resources with output quality. + +
AblationVideo-MME
Short: ≤2mMid: 4-15mLong: 30-60mOverall
frm sample : Frame sampling rate (w/o SR-Router)
8 frm48.342.037.242.5
16 frm50.042.838.043.6
1 FPS48.346.341.145.3
1 / 1.5 / 2 : Train stage - Long video knowledge injection
✓ × ✓45.643.740.243.6
✓ ✓ ✓48.346.341.145.3
FocusFast : Local-global video representation
×36.438.635.636.9
✓48.346.341.145.3
+ +Table 3: Ablation studies on SALOVA configuration. We utilize SALOVA-3B for the resource efficiency. + +# 5.3. Additional Analyses on SALOVA + +Ablation Study. We conduct ablation studies on three components as follows: (i) different video frame sampling strategies, (ii) intermediate training stage for long video knowledge injection, and (iii) the FocusFast mechanism to understand branched local-global representation in videos. + +As in Table 3, We first observe that using more frames with SR-Router significantly enhances performance, particularly in long-form videos. This aligns with our key contributions on managing long videos through an effective routing mechanism, suggesting that retrieving more frames can provide richer spatio-temporal information and improve the model's responses without losing contextual coherence. Additionally, we compare with a baseline trained with stage 1-2 (skipping stage 1.5). Here, we highlight the effectiveness of the SceneWalk dataset as an intermediate training step to enhance parametric knowledge for the long video analysis by allowing the model to learn from high-quality and densely captioned scene-level information, which is crucial for adapting to various lengths and contexts. Lastly, we conduct an analysis on the FocusFast method and demonstrate its efficacy in analyzing not only local details from relevant video segments but also in understanding the global video context through the simultaneous use of routing tokens, thereby facilitating a more comprehensive understanding of video content. + +Analysis of Retrieving Segments. By retrieving relevant video segments for the given queries, SALOVA can effectively target salient information in the long video and retain long context information. To further demonstrate the model's targeting capabilities beyond numerical performance in long video analysis, we explore our model's application in the Visual Needle-In-A-Haystack (V-NIAH) task [70], which extends the Needle-in-a-Haystack (NIAH) evaluation for LLMs to a vision-level benchmark. This task is particularly challenging as it requires models to not only detect but also precisely retrieve the sparse yet crucial visual cues scattered across lengthy videos. + +As in Fig. 3, we compare SALOVA against a base + +![](images/ca1a620b76da1ac252ac4fb290fa5ce8566e7356fabd8bf012e83a0c19238014.jpg) + +![](images/61df37e920f86f5a27194cc2fe924acd75ca4571035844b0909ad159f7fe714a.jpg) +Figure 3: Comparison results of V-NIAH. The x/y-axis indicates the total video frames and the location of needle image within the video, respectively. + +line trained on sparsely sampled frames (16 frm w/o SR-Router). Our framework effectively identifies and extracts relevant video segments from densely packed content, even when handling long context lengths. These results highlight SALOVA's robustness in managing complex, long-form videos, maintaining contextual continuity by strategically focusing on relevant segments to user queries. + +# 6. Discussion and Conclusion + +Discussion. Despite SALOVA's competence in handling extended video sequences, it is important to recognize scenarios where its complex architecture may not be necessary. Specifically, for shorter videos where sparse sampling suffices to capture essential spatio-temporal information, simpler models could potentially outperform the efficiency of SALOVA without necessitating its extensive processing capabilities. This suggests a future avenue for integrating a hybrid approach based on our framework by dynamically adjusting the complexity of the retrieval and processing mechanisms based on the video length and content density. + +Conclusion. In this paper, we introduce SALOVA, a novel framework designed to enhance the comprehension of long and untrimmed video by leveraging a retrieval-driven approach with new densely captioned dataset, the SceneWalk dataset. SALOVA strategically targets and processes only the relevant video segments, effectively addressing the structural limitations of current Video-LMMs with its Spatio-Temporal Connector and Segment Retrieval Router. Through extensive evaluation on various benchmarks, SALOVA exhibits its robust performance in interpreting complex video content, enhancing efficiency, and improving the understanding of extended videos. + +# Acknowledgments + +This work was partially supported by two funds: IITP grant funded by the Korea government (MSIT) (RS-2022-II220984) and IITP grant funded by the Korea government (MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-Term Visual Memory Network) + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 4, 6 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4, 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 7 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2 +[7] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. arXiv preprint arXiv:2312.06742, 2023. 2 +[8] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 4, 7, 3 +[9] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 6 +[10] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1 +[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 3, 7 +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 2 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2 +[15] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 6 +[16] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 5 +[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 1, 2 +[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4, 6 +[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 6 +[20] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7, 6 +[21] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (sfd): A benchmark for story-level video understanding. arXiv preprint arXiv:2406.10221, 2024. 4, 1 +[22] Google. Gemini, 2023. 1 +[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam + +Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 1, 3 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7 +[25] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. Language is not all you need: Aligning perception with language models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[26] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7 +[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 4 +[28] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 2 +[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3, 6 +[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 2, 6 +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning. PMLR, 2023. 2 +[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 7 +[33] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. Advances in neural information processing systems, 36, 2024. 5 +[34] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In + +European Conference on Computer Vision, pages 323-340. Springer, 2025. 1, 3, 7 +[35] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 3, 7 +[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 4, 1 +[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 2, 6, 3 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2, 6 +[39] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2025. 1, 7 +[40] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 5 +[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 3, 7 +[42] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2 +[43] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023.1 +[44] OpenAI. Gpt-4 technical report, 2023. 1, 3 +[45] OpenAI. GPT-4V(ision) System Card, 2023. 7 +[46] OpenAI. Hello gpt-4o, 2024. 3, 7 +[47] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 6 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6, 3 +[49] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 6 + +[50] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 3 +[51] Share. Sharegemini: Scaling up video caption data for multimodal large language models, 2024. 7 +[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6 +[53] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 3, 4, 7 +[54] Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. Augmented SBERT: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 296–310, Online, 2021. Association for Computational Linguistics. 2, 4 +[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[56] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2 +[57] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. arXiv preprint arXiv:2407.15754, 2024. 7, 3 +[58] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6 +[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3, 7 +[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 1 +[61] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang + +Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 3 +[62] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2 +[63] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4, 6 +[64] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 3 +[65] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2 +[66] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 6, 7 +[67] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 4, 6 +[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7 +[69] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1 +[70] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 3, 7, 8 +[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 6 +[72] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 7 +[73] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations. 2, 3, 4 \ No newline at end of file diff --git a/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/images.zip b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b90f34ab2fb679165ab87b38bf9e8faaa122772 --- /dev/null +++ b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34f2e00665cd3081a2da6876bb53f05e26c57d7e175783935d3f2535b2d603d7 +size 403379 diff --git a/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/layout.json b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fffbba5fe5c231b27cfb9bf3f7799148674dd4a1 --- /dev/null +++ b/2025/SALOVA_ Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis/layout.json @@ -0,0 +1,7378 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "text", + "content": "SALOVA: Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 161, + 473, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 161, + 473, + 204 + ], + "spans": [ + { + "bbox": [ + 142, + 161, + 473, + 204 + ], + "type": "text", + "content": "Junho Kim* Hyunjun Kim* Hosu Lee Yong Man Ro† Integrated Vision and Language Lab, KAIST, South Korea {arkimjh, kimhj709, leehosu01, ymro}@kaist.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "content": "Despite advances in Large Multi-modal Models, applying them to long and untrimmed video content remains challenging due to limitations in context length and substantial memory overhead. These constraints often lead to significant information loss and reduced relevance in the model responses. With the exponential growth of video data across web platforms, understanding long-form video is crucial for advancing generalized intelligence. In this paper, we introduce SALOVA: Segment-Augmented LOng Video Assistant, a novel video-LLM framework designed to enhance the comprehension of lengthy video content through targeted retrieval process. We address two main challenges to achieve it: (i) We present the SceneWalk dataset, a high-quality collection of 87.8K long videos, each densely captioned at the segment level to enable models to capture scene continuity and maintain rich descriptive context. (ii) We develop robust architectural designs integrating dynamic routing mechanism and spatio-temporal projector to efficiently retrieve and process relevant video segments based on user queries. Our framework mitigates the limitations of current video-LMMs by allowing for precise identification and retrieval of relevant video segments in response to queries, thereby improving the contextual relevance of the generated responses. Through extensive experiments, SALOVA demonstrates enhanced capability in processing complex long-form videos, showing significant capability to maintain contextual integrity across extended sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "content": "Recent advancements in Large Language Models (LLMs) [22, 43, 44] have brought us one step closer to achieving Artificial General Intelligence (AGI). Next following step, the current trend is shifting toward modular systems that integrate various multi-modality, leveraging" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "type": "text", + "content": "the exceptional generalization and reasoning capabilities of LLMs to evolve into Large Multi-modal Models (LMMs). Accordingly, users can unrestrictedly interact with the models across various modalities beyond text, expanding the scope of machine understanding and enhancing user engagement. Especially, considering the widespread adoption of long-form videos across various web platforms, the importance of understanding long, untrimmed video has become increasingly prominent in the multi-modal domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 341, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 533 + ], + "type": "text", + "content": "After the pioneer works [14, 37, 38] utilizing visual instruction tuning to augment vision perception into LLMs, remarkable strides [11, 17, 65] have been made in aligning cross-modal consistency—especially between vision and language domains. Albeit more recent models [29, 69] integrate various vision modalities all at once, current approaches still face significant challenges in understanding untrimmed and long-form video content. The main challenge is attributed to the limited context length of LMMs, which is an inherent structural limitation that restricts the models to process only a finite number of tokens as the input sequences. We can exemplify that LLaVA series [29, 30], when processing a video data, require 144 visual tokens per each frame, where numerical approximation is only maximum of " + }, + { + "bbox": [ + 313, + 341, + 556, + 533 + ], + "type": "inline_equation", + "content": "\\sim 56" + }, + { + "bbox": [ + 313, + 341, + 556, + 533 + ], + "type": "text", + "content": " frames using 8K max context length LMMs, which is still limited to handle long sequence data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": "Accordingly, current video-LMMs [12, 35, 39] rely on (i) sparse frame sampling to represent entire videos [26, 72], (ii) dense compression of visual tokens into a smaller size to manage the excessive number of frames [34, 41], and (iii) adaptive pooling strategies [59, 60] based on the SlowFast approach [19], all aimed at fitting the long video sequences within the limited context window of LMMs. Several studies focusing on the long video understanding task have presented memory-augmented generation [23, 53] utilizing an additional buffer to embed long-term information, or have extended the context using RoPE-based frequency extension during the training [70]. Despite of such endeavors, when handling massive video frames, previous works still confront restricted context size and significant memory overhead, which leads to substantial visual information" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 237, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author. Code is available at https://ivy-lvlm.github.io/SALOVA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3352" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "loss. As critical events may be overlooked by the models, this hinders their ability to fully capture context changes in lengthy videos, resulting in inaccurate and irrelevant responses for the user queries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 300 + ], + "type": "text", + "content": "Starting from the intuitive insight outlined below, in this paper, we propose a retrieval-driven approach for long video understanding with LLMs. Analogous to the recent Retriever-Augmented Generation (RAG) systems [28] (widely adopted in LLMs), which retrieve relevant information from external factual knowledge, humans naturally employ similar strategies when seeking specific information, efficiently locating and referring necessary materials to answer targeted questions—e.g., imagine that we are taking open-book exams or searching for a certain recipe in a cookbook. Given a long and untrimmed video, mirroring the targeted retrieval processes, we introduce a novel framework, Segment-Augmented LOng Video Assistant (SALOVA) to effectively handle the long sequence visual inputs by retrieving the relevant video segments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 407 + ], + "type": "text", + "content": "To construct our video segments retrieval framework, central challenge hinges on establishing two main components: (i) Densely captioned video data, which consists of video-caption pairs with progressively-captioned descriptions that change throughout each video used to train the model to accurately identify relevant video segments. (ii) Dynamic routing mechanism, which selects pertinent video segments for the queries, followed by being connected to LLMs. To address that, our approach is outlined as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 411, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 625 + ], + "type": "text", + "content": "Data. (§3) Recently several video-text paired datasets [4, 6, 8, 10, 56, 62] have been released, but they are inadequate for handling long and untrimmed video data, where only partial video moments are described with limited word length as compared in Fig. 1(a). To handle such insufficiency of detailed descriptions within the videos and the short durations of both videos and texts, we introduce the SceneWalk dataset, a new high-quality video dataset with thorough captioning for each video. It includes dense and detailed descriptions for every video segment across the entire scene context. The SceneWalk dataset, sourced from long and untrimmed 87.8K YouTube videos (avg. 486 seconds each), features frequent scene transitions across a total of 11.8K hrs video duration and 1.3M massively segmented video clips. Each video segment in the dataset is provided with a detailed description (avg. 137.5 word length), generated by combining pre-trained models [54, 73] and manual curation from human." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Architecture. (§4) Utilizing the constructed video dataset, SALOVA learns to identify relevant video segments for the given queries within each video source and then auto-regressively predicts the next token. To do so, we present two architectural designs to seamlessly incorporate the retrieved segments in an end-to-end training: Spatio-Temporal Connector and Segment Retrieval Router. By fo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "cusing on the relevant segments, our framework can perform deeper reasoning without being constrained by context length limitations. Additionally, we present FocusFast approach, which intensively analyzes the selected segments for detailed comprehension (focus pathway), while quickly accessing overall contextual information with routing tokens obtained from the entire video segments (fast pathway). The strategy ensures SALOVA to maintain comprehensive video understanding while prioritizing details where it is most needed, effectively enhancing long and untrimmed video interpretation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "content": "Through extensive experiments and analyses, we corroborate that competitive performance of SALOVA to the existing video-LMM models in understanding complex long-form videos. Also, our results show significant reductions in the loss of crucial visual information and a lower risk of omitting important events, demonstrating the effectiveness of our proposed method across various video benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 289, + 538, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 289, + 538, + 300 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 538, + 300 + ], + "type": "text", + "content": "Our contribution can be summarized into three-fold:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 302, + 553, + 481 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 302, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 553, + 373 + ], + "type": "text", + "content": "- We introduce the SceneWalk dataset, a high-quality and densely-captioned video dataset with detailed segment-level descriptions from 87.8K long and untrimmed video sources. The proposed dataset provides rich context and scene continuity, enabling effective training for long-form video understanding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 374, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 374, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 374, + 553, + 433 + ], + "type": "text", + "content": "- We propose Segment-Augmented LOng Video Assistant (SALOVA), a novel video-LMM framework designed to enhance long video comprehension by targeting relevant video segments in lengthy videos, optimizing the model's focus on essential segment targets for the given queries." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 434, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 553, + 481 + ], + "type": "text", + "content": "- Through extensive evaluation, we validate that SALOVA improves overall long video understanding capabilities by effectively integrating relevant video segments, thus optimizing to handle long and untrimmed video content." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 515, + 463, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 515, + 463, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 515, + 463, + 528 + ], + "type": "text", + "content": "2.1. Large Multi-modal Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "After the emergence of LLMs [5, 55], which can actively interact with users through back-and-forth conversations, as a next leap, various research efforts [2, 25, 31] integrate different modalities into the LLMs, utilizing their core reasoning and zero-shot capabilities. Building on the open-sourced models [13, 55], seminal works [14, 38, 65] have bridged image and text modality under the visual instruction tuning and presented multi-modal assistant models that possess visual perception and QA capabilities. Since then, numerous research studies have been introduced to (i) enhance vision understanding with advanced architectures [17] or higher resolutions [30, 37], (ii) implement more sophisticated alignment layers [7, 42] between modalities, and (iii) train the models with more high-quality data and larger model parameters." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3353" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 212, + 197 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 212, + 197 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 212, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 212, + 197 + ], + "type": "image", + "image_path": "115411666638185eab77d20c4648d4db87a8318b2b9f8c5453a123aed7b53ac5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 203, + 211, + 216 + ], + "lines": [ + { + "bbox": [ + 67, + 203, + 211, + 216 + ], + "spans": [ + { + "bbox": [ + 67, + 203, + 211, + 216 + ], + "type": "text", + "content": "(a) Video-Text Dataset Comparison" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 69, + 338, + 201 + ], + "blocks": [ + { + "bbox": [ + 219, + 69, + 338, + 201 + ], + "lines": [ + { + "bbox": [ + 219, + 69, + 338, + 201 + ], + "spans": [ + { + "bbox": [ + 219, + 69, + 338, + 201 + ], + "type": "image", + "image_path": "2aca787b5795f792bf26420366df55c4fd7b063a91006caa4fd7148418b58bdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 203, + 328, + 214 + ], + "lines": [ + { + "bbox": [ + 228, + 203, + 328, + 214 + ], + "spans": [ + { + "bbox": [ + 228, + 203, + 328, + 214 + ], + "type": "text", + "content": "(b) SceneWalk Statistics" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 78, + 554, + 202 + ], + "blocks": [ + { + "bbox": [ + 345, + 78, + 554, + 202 + ], + "lines": [ + { + "bbox": [ + 345, + 78, + 554, + 202 + ], + "spans": [ + { + "bbox": [ + 345, + 78, + 554, + 202 + ], + "type": "image", + "image_path": "a4c4cbb8373c76c5c6851d3f3659752931644de12660eb98c28b8f161aba9adc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 204, + 529, + 215 + ], + "lines": [ + { + "bbox": [ + 370, + 204, + 529, + 215 + ], + "spans": [ + { + "bbox": [ + 370, + 204, + 529, + 215 + ], + "type": "text", + "content": "(c) Overall Pipeline for Data Collection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 219, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 54, + 219, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 555, + 268 + ], + "type": "text", + "content": "Figure 1: The overview of the SceneWalk dataset includes (a) dataset comparison, (b) detailed statistics, and (c) the annotation pipeline for description and score collection. Note that the scale of circles in Fig. 1(a) indicates the data size, and the color distribution in Fig. 1(b) denotes the video duration in each video category—brighter colors correspond to shorter video durations. Further details about the dataset are provided in Appendix A." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 278, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 483 + ], + "type": "text", + "content": "Recent focus has shifted towards more unified modality processing following the release of omnivorous models [46]. Some recent omni-versions of LMMs [29, 64] can handle combinatorial subsets from various modality sources, such as images, videos, audio, speech, and depth. However, the current LMMs for video [32, 35, 41] still lack of capturing the necessary details to effectively process video information due to their sparse frame sampling strategy. While such approach is seemingly adequate for relatively shorter videos, it may fail to capture comprehensive spatio-temporal information, potentially compromising the accuracy of model responses to user queries. In this paper, SALOVA first retrieves relevant video segments, then concentrates on more granular video cues. Such targeted focus allows the model to effectively understand complex analysis within the videos, significantly improving its ability to provide contextual-aware and accurate responses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 205, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 205, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 205, + 504 + ], + "type": "text", + "content": "2.2. Long Video Understanding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "content": "In parallel, we dually introduce video-specialized models [12, 23, 53] integrated with LLMs, which have also been widely explored these days to enhance video understanding and reasoning. Here, the most challenging part of current video-LMMs lies in handling long video sequences, mainly due to the limited context length of the LLMs. This limitation compels the models to sparsely sample the video frames in only limited sizes (e.g., typically 8 or 16 frames), potentially missing important spatial and temporal information. To address this, several studies have focused on compressing visual tokens into a more manageable size, proposing aggregation [34, 41] or pooling methods [35, 59] with advanced vision encoder structures [67, 73]. In addition, memory-augmented methods [23, 53] first stored long-term information in a memory bank, then responded to specific queries by loading memory features from the stored buffer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "On the other hand, among more recent approaches, Li" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "content": "et al. [70] have directly extended the LLMs' context length by exploiting RoPE-based frequency interpolation, and Xue et al. [61] have introduced sequence parallelism that can be implemented on multiple GPUs by modifying backend systems. However, we argue that such approaches inherently cannot be free from the fixed context length and provoke intensive memory demands when processing more longer videos. Instead, by focusing on the relevant segments within the entire video, SALOVA can efficiently handle the limited context length, enabling targeted processing of key moments without the need for excessive memory, thereby enhancing performance on longer video sequences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 427, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 427, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 427, + 447 + ], + "type": "text", + "content": "3. SceneWalk Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 598 + ], + "type": "text", + "content": "In this section, we elaborate on how we collected the SceneWalk dataset. The overall pipeline for building the dataset and summarized statistics are illustrated in Fig. 1. While several video SFT datasets [32, 41] are widely used during the instruction tuning stage, they often fail to capture comprehensive details within the scenes. This stems from the nature of instruction-type questions, which provide only partial information, and the brief lengths of both videos and texts in QAs. In contrast, the SceneWalk dataset offers densely captioned video-text pairs that cover long sequence videos in full details, as shown in Fig. 1(a). For further detailed data statistics, please see Appendix A." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 608, + 484, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 484, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 484, + 621 + ], + "type": "text", + "content": "3.1. Data Gathering and Processing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Video Source & Filtering. For the long and untrimmed video sources, we primarily focus on three key aspects to build densely captioned video dataset: (i) extensive video length with diverse video source categories, (ii) high-quality video contents, (iii) frequent scene transitions within each video. Accordingly, our data collection is mainly sourced from YouTube, ensuring rich dynamic content that bet-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3354" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": "ter reflects real-world complexities experienced by global users—here, because our main goal for video gathering is on complex scene understanding, we exclude low-quality and user-uploaded aesthetic videos (e.g., WebVid, Pixabay, Pexels, and etc.,) that are rather beneficial for video generation tasks, despite their merits for easy collection. We have collected YouTubeurls from [27] and downloaded the whole video in untrimmed states. Among the total 32 coarse and diverse video categories YouTube API provided, we selectively curated 10 categories, excluding categories such as News & Politics, Classics, and Documentary, due to their static nature, which provides only sparse temporal information in the videos. We further supplemented the dataset with additional Movie & Drama videos sourced from [21, 53], totaling 87,867 video sources with 11.87 Khrs video duration (avg. 486.5-seconds)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 267, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 422 + ], + "type": "text", + "content": "Segmenting Video into Clips. Next, for the collected long and untrimmed video sources, we cut the lengthy videos into small segments to densely caption the entire video in next phase. Instead of adopting the bottom-up approach used in the ShareGPT4Video dataset [8], which segments videos into fixed time intervals (2-seconds) in advance and then merges adjacent frames based on their CLIP similarity [48], we directly employed PysceneDetect 1 to segment the videos, dynamically adjusting the threshold based on the raw-level video information to reliably detect scene changes. At the end, the total number of 1.29M of video segments with 33.11-seconds average video length is extracted from the original video sources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 431, + 190, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 190, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 190, + 445 + ], + "type": "text", + "content": "3.2. Captioning and Scoring" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 632 + ], + "type": "text", + "content": "Dense Segment Captioning. After obtaining the massive video segments, our next goal is to caption each segment with visual details and narrative context to capture the scene-specific explanations, which can enrich scene-level interpretation. To achieve this, we plan to utilize pre-trained LMMs to generate detailed descriptions for the partial video segments. As the captioner, we empirically found that VILA-1.5 (13B) [36] shows competent descriptive quality than other open-sourced models, and used to generate dense captions for each video segment with randomly sampled instructions for detailed descriptions. As a result, we acquire 1.29M pairs of detailed descriptions corresponding to the video segments, each description with average 137.5 word length. Please see instruction details and qualitative examples of generated captions in Appendix A." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 636, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 295, + 685 + ], + "type": "text", + "content": "Scoring Video-Text Correspondence. Lastly, we score the correspondence between the video segments and the paired dense descriptions, which will later be used as explicit supervision to robustly train our retrieval framework. What" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "we must not overlook here is that the paired video-text relationship is not solely a one-to-one correspondence but is more akin to generalized bipartite matching. That is, within the long and untrimmed video source, each video segment can be connected to other descriptions with additional edges. Therefore, for the " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " number of video segments and their paired segments, we can construct a " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{N_v\\}^2" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " correspondence matrix between video-text (V2T). To measure each correspondence, we employ LanguageBind [73] due to its competitive alignment capabilities across various modalities. In addition, we build another " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{N_v\\}^2" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " matrix to provide a doubly robust measure for the correspondence scores among adjacent descriptions (T2T) by comparing similarity within the textual context using the SBERT model [54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 550, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 550, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 550, + 266 + ], + "type": "text", + "content": "4. Segment-Augmented LOng Video Assistant" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": "Network Overview. For a given set of " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": " video segments sampled at 1 FPS " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "v = \\{v_i\\}^{N_v}" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": ", where each segment " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "v_i \\in \\mathbb{R}^{T_i \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": " has varying video length (summing up to the total time " + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 276, + 553, + 347 + ], + "type": "text", + "content": " of a long and untrimmed video), SA-LOVA consists of four main architecture components as illustrated in Fig. 2:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 348, + 554, + 599 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 348, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 384 + ], + "type": "text", + "content": "- Vision Encoder: We use CLIP [48] or SigLIP [67] to extract visual features, followed by 2x2 average pooling, resulting in 144 or 196 visual tokens per frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 384, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 554, + 467 + ], + "type": "text", + "content": "- Spatio-Temporal Connector: To handle spatio-temporal features of varying lengths from the vision encoder, we employ the Perceiver Resampler [2], which consists of 2-layer Transformer architecture followed by a 2-layer MLP with GELU activation as projector. This resampler embeds each video segment feature into fixed size latent features that are connected to LLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 468, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 554, + 538 + ], + "type": "text", + "content": "- Segment Retrieval Router: For the given textual queries, a retrieval structure (2-layer Transformer) gathers representative information (i.e., routing tokens) from each video segment and then routes the query-relevant video features into the LLMs. Note that the router architecture is trained in an end-to-end manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 540, + 553, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 553, + 599 + ], + "type": "text", + "content": "- Large Language Model: For LM backbones, We utilize three open-sourced LLMs with varying parameter sizes, LLaMA-3.2 (3B) [18], Phi-3.5 (3.8B) [1] and Qwen-2.5 (7B) [63], all of which are instruction-tuned models that possess QA assistant capabilities." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 608, + 504, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 504, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 504, + 621 + ], + "type": "text", + "content": "4.1. Long Video Processing and Pipeline" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 626, + 462, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 462, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 462, + 638 + ], + "type": "text", + "content": "4.1.1. Spatio-Temporal Connector" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "The first component of our model, Spatio-Temporal Connector, efficiently handles long and variable-length input video segments by extracting each segment's visual semantics in a fixed-size latent vector. As illustrated in Fig. 2(b), we first sample video frames at 1 FPS from each video segment, then visual features are acquired with " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " pooling" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 294, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 294, + 713 + ], + "type": "text", + "content": "1We use AdaptiveDetector with default setup in https://github. com/Breakthrough/PySceneDetect" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3355" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 281 + ], + "type": "image", + "image_path": "c393cd9f4355eeb974eb9dbfc2dad22383bb863765689a7956ae0455230c5844.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 322 + ], + "type": "text", + "content": "Figure 2: The network overview of SALOVA. Our framework consists of four structural components: vision encoder, ST-connector, SR-router, and LLMs. Using the FocusFast strategy, our model can concentrate on more detailed local information while maintaining context awareness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 333, + 294, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 294, + 416 + ], + "type": "text", + "content": "(thus, 196 tokens from each frame). After that, the visual features are flattened and fed into the ST-Connector with additional positional and temporal encoding. Here, when the long video is processed, the number of unfolded patch tokens becomes extremely large, leading to exhaustive computations. To address this, we employ a dynamic token drop technique to reduce computational load." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "type": "text", + "content": "Dynamic Token Drop. To effectively manage long video sequences, the token drop has been utilized in video generation tasks [16, 40]. Expanding such approach, in our framework, the dropout rate is dynamically adjusted based on the length of the input sequence " + }, + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "type": "text", + "content": " in the input visual feature " + }, + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "type": "inline_equation", + "content": "f_{v} \\sim T_{i} \\times H_{p}W_{p} \\times d" + }, + { + "bbox": [ + 54, + 422, + 295, + 578 + ], + "type": "text", + "content": ", which allows for more efficient processing of longer sequences by reducing computational demands, while still preserving dense visual semantics in shorter videos. Additionally, to retain spatio-temporal information from the dropped patches, we add positional embeddings separately along the spatial and temporal axes. This enables more refined extraction of spatio-temporal visual semantics even after reducing the number of tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 589, + 194, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 194, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 194, + 601 + ], + "type": "text", + "content": "4.1.2. Segment Retrieval Router" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Next, the key to conveying the pertinent video information to LLMs is retrieving relevant video segments by querying sentence. To densely cue the similarities between the video and sentence information, we introduce a routing framework, Segment Retrieval Router, which consists of 2-layer Transformer as illustrated in Fig. 2(c). After obtaining the routing tokens " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "R = \\{R_{i}\\}^{N_{v}} \\in \\mathbb{R}^{N_{v} \\times D}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " from entire video segments, we aggregate them and feed into the SRRouter as queries. For the given sentence, we employ the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": "same text encoder used for the vision encoder and project it into the shared embedding space to obtain sentence features " + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N_t \\times D}" + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": " indicates textual length." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "Using the cross attention mechanism (q: " + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "; k/v: " + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "), we can estimate similarity scores between the video segments and given sentence queries (i.e., V-T similarity). The scores enable the SR-Router to prioritize and select the most relevant video segments that align with the sentence query." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "type": "text", + "content": "Retrieval Objective. To seamlessly train the SR-Router with the mainstream flows of SALOVA in an end-to-end manner, we have designed a similarity loss function " + }, + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sim}}" + }, + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "type": "text", + "content": " that minimizes the distance between the high-dimensional embeddings of the video segments and sentence queries. Here, we use the correspondence scores (aforementioned in Sec. 3.2) as a retrieval supervision signal " + }, + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 434, + 556, + 650 + ], + "type": "text", + "content": ", after applying one-hot encoding. We incorporate a simple margin-based loss, commonly used in contrastive learning settings [33], which enables the model to learn off-diagonal relaxation in the correspondence matrices between video segments and sentences. As mentioned earlier, the relationship between paired videos and sentences is closer to generalized bipartite matching than to one-to-one matching, so relaxation learning helps to accommodate the inherent complexity in aligning correspondence. In conclusion, with the binary cross-entropy loss and the score margin loss, we can formulate the similarity loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 658, + 553, + 701 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 658, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 319, + 658, + 553, + 701 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m}} = \\underbrace {\\mathcal {L} _ {\\text {b c e}} \\left(y _ {i} , s _ {i}\\right) _ {i = 1} ^ {N _ {v}}} _ {\\text {p o i n t - w i s e C E}} + \\underbrace {\\frac {1}{N _ {s}} \\sum_ {j} \\max \\left(0 , \\delta - \\left(s _ {j} ^ {p} - s _ {j} ^ {n}\\right)\\right)} _ {\\text {s c o r e m a r g i n l o s s}}, \\tag {1}", + "image_path": "11c3fd8164b688880b88d7c37f8fb155caca58a0fdb30ad7358d93808a15db19.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "inline_equation", + "content": "s_j^p" + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "inline_equation", + "content": "s_j^n" + }, + { + "bbox": [ + 313, + 701, + 555, + 715 + ], + "type": "text", + "content": " indicate randomly sampled scores from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3356" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "positive and negative pairs, respectively, and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " denotes the margin parameter (set as 0.2). Note that the similarity loss is trained in conjunction with the auto-regressive loss " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ar}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " from subsequent LLMs in an end-to-end manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 128, + 264, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 264, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 264, + 140 + ], + "type": "text", + "content": "4.1.3. FocusFast Pathways: Integration to LLMs" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 370 + ], + "type": "text", + "content": "Using the routing tokens, we can calculate the similarities of each video segment for the given query. Leveraging the similarities, SALOVA efficiently retrieves the specific video features that exhibit the highest relevance score to the textual query, where the indexed video features are then directly integrated into the LLM architecture. Here, extending the SlowFast pathways concept [19], we present the FocusFast mechanism to effectively manage the processing pathway for the retrieved video segments: (i) Focus pathway concatenates the top-K most pertinent features to construct a comprehensive video representation, capturing local details across retrieved segments and enabling detailed interactions with textual queries to enhance handling complex video information. (ii) Fast pathway focuses on the more broader-level context by employing segment-wide routing tokens as the condensed global representation. It effectively contains dynamic spatio-temporal changes throughout the video stream, thereby allowing SALOVA to understand the overall video content and scene-level continuity awareness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "type": "text", + "content": "Once the most pertinent features are retrieved, they are delivered to the LM backbone for the final processing as in Fig. 2(a), integrating video-specific details into the models' responses. By effectively handling long and untrimmed videos with the proposed retrieval and routing mechanism, SALOVA can maintain the flow of salient information without the processing overhead for less related data, thus generating more context-aware responses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 168, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 168, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 168, + 488 + ], + "type": "text", + "content": "4.2. Training Strategies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 685 + ], + "type": "text", + "content": "The current training strategies for LMMs predominantly consist of two-step training: (i) cross-modal alignment and (ii) visual instruction tuning. Recently, Li et al. [29] have emphasized the importance of high-quality knowledge learning between the two training stages (thus stage 1.5), pointing out that the models cannot enoughly learn necessary knowledge during the alignment with the low-qualitative web-scale image-text data. As the similar approach of using rephrased descriptions for additional knowledge learning [29], we employ the newly collected SceneWalk dataset as the parametric knowledge injection step, which enables the SALOVA to learn detailed spatial and temporal representation from the long sequence video data before the instruction tuning. Accordingly, our training recipes and data configuration can be divided into three steps as follow (Please see training details in Appendix B):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Stage 1: Cross-modality Alignment. For the initial step in modality alignment, we utilize 790K image/video-text" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "paired dataset: (i) 558K image-text pairs from the CC3M dataset [52], filtered by LLaVA [38] and (ii) video-text pairs sampled from the WebVid 2.5M subset [4]. We freeze vision encoder and LLMs during the training, and mainly focus on optimizing the connector and router to map the visual information into the textual space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 148, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 280 + ], + "type": "text", + "content": "Stage 1.5: Long Video Knowledge Injection. As an intermediate training step, we use the SceneWalk dataset to train the SALOVA, unfreezing all trainable parameters except for the vision encoder. During training, we input the long and untrimmed video instances and follow the processing pipeline shown in Fig. 2. By training the model with densely captioned video-description pairs, it acquires high-quality parametric knowledge of both spatial and temporal information. In addition, through the aforementioned retrieval process, the model learns to target video segments that are mostly relevant to the video description." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 284, + 556, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 556, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 556, + 416 + ], + "type": "text", + "content": "Stage 2: Video Instruction Tuning. To possess QA capabilities in SALOVA, we use extensive video instruction-tuning data as the final training step. The instruction data are mainly sourced from four different datasets: LLaVA-Video-178K [71], NeXT-QA [58], ActivityNetQA [66], and PerceptionTest [47]—comprising a total of 1.4M video-instruction QA data, including caption entries, open-ended QA, and multiple-choice QA. Note that we train all the network parameters during this stage and auto-regressively update the instruction-following assistant's response for the next word prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 428, + 395, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 395, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 395, + 441 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 436, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 436, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 436, + 460 + ], + "type": "text", + "content": "5.1. Experimental Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 650 + ], + "type": "text", + "content": "Implementation. For the vision and text encoders of the SR-Router, we utilize CLIP [48] for the small-size models and SigLIP [67] for the frontier model, with resolution sizes of 336 and 384, respectively. We employ a 2-layer transformer with a head size of 2 for the ST-Connector, which has a latent dimension of 256. The token drop mechanism is dynamically applied based on video length, with varying maximum drop rates at each training stage—Stage 1 has no token drop, Stage 1.5 allows up to 0.7, and Stage 2 allows up to 0.4. For the configuration of SR-Router, we set 2-layer of transformers with a single head, and top-K number is set to 5 during the stage 2. Following the recent works [30, 37], the projector layer consists of 2-layer MLP with GELU. We employ three LLMs: (i) 3B: Llama-3.2-3B [18], (ii) 3.8B: Phi-3.5 [1], and (iii) 7B: Qwen2.5-7B [63]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Training Details. For the each training stage, we train SA-LOVA for 1 epoch with 1 node of 8 A100 GPUs. The total training hours for 3B (3.8B) and 7B models roughly take 5 and 7 days, respectively. We employ FlashAttention-2 [15], gradient checkpointing [9], and ZeRO-2 [49] to minimize" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3357" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 293, + 252 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 293, + 252 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 293, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 293, + 252 + ], + "type": "table", + "html": "
Model#paramVideo-MMELVBench Acc. (val)
ShortMediumLongOverall
Proprietary LMMs
GPT-4V [45]n/a70.555.853.559.9-
GPT-4o [46]n/a80.070.365.371.966.7
Gemini 1.5 Pro [50]n/a81.774.367.475.064.0
Open-sourced LMMs
ST-LLM [39]7B45.736.831.337.9-
VideoChat2 [32]7B48.337.033.239.539.3
ShareGPT4Video [8]8B48.336.335.039.939.7
Video-LLaVA [35]7B45.338.036.239.939.1
Chat-UniVi-V1.5 [26]7B45.740.335.840.6-
Qwen-VL-Chat [3]7B46.938.737.841.1-
ShareGemini [51]7B49.141.339.143.2-
SliME [72]8B53.342.739.845.3-
PLLAVA [59]7B----40.2
VideoLLaMA2 [12]8B56.045.442.147.9-
LongVA [70]7B61.150.446.252.6-
Ours
SALOVA-3B3B48.346.341.145.341.4
SALOVA-3.8B3.8B47.148.844.146.741.6
SALOVA-7B7B59.450.449.453.144.6
", + "image_path": "aa9e7fe8608057e062a9c38b990f48994589c3c4515a6c69a1917beee70f07a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 253, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 289 + ], + "type": "text", + "content": "Table 1: Detailed results for the Video-MME benchmark (w/o subtitles) and LongVideoBench. The best results are highlighted in **bold** and the runner-up results are **underlined**." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 297, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 357 + ], + "type": "text", + "content": "the memory footprint associated with model components (i.e., gradient, activation, and optimizer states). Additionally, we fine-tune the trainable parameters at each step without employing LoRA [24]. For the extended training config, we have attached the details in Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 361, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 529 + ], + "type": "text", + "content": "Evaluation Benchmarks. We evaluate our model using two types of video analysis benchmarks—long video understanding and general video understanding, categorized based on the video length. For the long video benchmark, we primarily utilize Video-MME [20] and LongVideoBench [57], both of benchmarks include videos up to two hours long duration. As the general video analysis evaluations, we employ various benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32]. Note that the same pipeline is used to obtain video segments for each benchmark, and all benchmarks are sampled at 1 FPS without token drop during inference. As comparison baselines, considering academic budget constraints, we evaluate against models that have similar parameter size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 178, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 178, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 178, + 550 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": "Results on Long Video Understanding. Video-MME [20] evaluates LMMs with a focus on video analysis across a variety of video types and durations. We primarily compare the benchmark results in settings without subtitles, relying solely on video frames. Therefore it can assess the LMMs' visual comprehension capabilities rigorously, based purely on visual content. Also, LongVideoBench [57] is designed to assess LMMs' understanding of long-duration videos up to two hours. It includes a diverse collection of videos, challenging the models' ability to process and interpret extensive visual and contextual information across a variety of themes. As shown in Tab. 1, our model shows competent video understanding performance across all video" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 227 + ], + "type": "table", + "html": "
Model#paramActivityNetQAVideoChatGPTMVBench
test (acc/score)test (acc)test (acc)
Proprietary LMMs
GPT-4V [45]n/a57.0-4.0643.5
GPT-4o [46]n/a61.9---
Gemini 1.5 Pro [50]n/a57.5---
Open-sourced LMMs
VideoLLaMA [68]7B12.41.12.1634.1
VideoChatGPT [41]7B35.22.72.4232.7
MovieChat [53]7B45.7-2.67-
Chat-UniVi [26]7B46.13.22.99-
LLaMA-VID [34]7B47.43.32.8941.3
VideoChat2 [32]7B49.13.32.9851.1
VideoLLaMA2 [12]8B50.23.33.1354.6
Ours
SALOVA-3B3B52.63.43.0851.7
SALOVA-3.8B3.8B51.13.52.8346.4
SALOVA-7B7B53.63.53.1353.5
", + "image_path": "d9f276eb632b189d6883e2cfbadec4a71ca3b70f2eac85b1e25561e6b22481f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": "Table 2: Comparison results for generic video understanding benchmarks. The best results are highlighted in bold and the runner-up results are underlined." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 275, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 358 + ], + "type": "text", + "content": "length distributions in Video-MME and lengthy video instances in LongVideoBench. Notably, we highlight that SALOVA achieved significant performance in the medium (average 562.7 seconds) and long (average 2385.8 seconds) length categories in Video-MME benchmark, even with more smaller size of backbone LM parameters compared with the baseline models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 494 + ], + "type": "text", + "content": "Such performance gains in long video instances are attributed to our model's dynamic capability to retrieve and process only the relevant video segments, enabling it to handle lengthy video content efficiently without being constrained by the limited context length. Especially, the routing mechanism in SALOVA strategically prioritizes video segments that are likely to contain crucial visual and contextual cues relevant to the query. This selective routing mechanism reduces the computational load and minimizes the information loss that commonly occurs in current videolMMs trying to process extensive video data in entirety." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "Results on General Video Understanding. For general video understanding benchmarks such as ActivityNetQA [66], VideoChatGPT [41], and MVBench [32], SALOVA was evaluated across various video types to assess its general video understanding capabilities. As shown in Tab. 2, SALOVA demonstrated competent performance, comparable to existing video-LMMs, especially in dynamic and shorter video sequences. On ActivityNetQA, the model effectively utilized its segment retrieval strategy to provide focused and contextually appropriate responses, which helped maintain accuracy. This approach was similarly effective in the multi-modal settings of VideoChatGPT and MVBench, where SALOVA showed consistent performance in handling dialogues and visual cues. These outcomes highlight SALOVA's capability to process general video content efficiently through its dynamic routing mechanism, offering a reliable solution that balances computational resources with output quality." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3358" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 198 + ], + "type": "table", + "html": "
AblationVideo-MME
Short: ≤2mMid: 4-15mLong: 30-60mOverall
frm sample : Frame sampling rate (w/o SR-Router)
8 frm48.342.037.242.5
16 frm50.042.838.043.6
1 FPS48.346.341.145.3
1 / 1.5 / 2 : Train stage - Long video knowledge injection
✓ × ✓45.643.740.243.6
✓ ✓ ✓48.346.341.145.3
FocusFast : Local-global video representation
×36.438.635.636.9
✓48.346.341.145.3
", + "image_path": "584fd63f52401e38264afc94bb4baff635d1f2b542065e4520c4a0d1f6f2f65d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 198, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 224 + ], + "type": "text", + "content": "Table 3: Ablation studies on SALOVA configuration. We utilize SALOVA-3B for the resource efficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 234, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 234, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 234, + 247 + ], + "type": "text", + "content": "5.3. Additional Analyses on SALOVA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 295, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 315 + ], + "type": "text", + "content": "Ablation Study. We conduct ablation studies on three components as follows: (i) different video frame sampling strategies, (ii) intermediate training stage for long video knowledge injection, and (iii) the FocusFast mechanism to understand branched local-global representation in videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 554 + ], + "type": "text", + "content": "As in Table 3, We first observe that using more frames with SR-Router significantly enhances performance, particularly in long-form videos. This aligns with our key contributions on managing long videos through an effective routing mechanism, suggesting that retrieving more frames can provide richer spatio-temporal information and improve the model's responses without losing contextual coherence. Additionally, we compare with a baseline trained with stage 1-2 (skipping stage 1.5). Here, we highlight the effectiveness of the SceneWalk dataset as an intermediate training step to enhance parametric knowledge for the long video analysis by allowing the model to learn from high-quality and densely captioned scene-level information, which is crucial for adapting to various lengths and contexts. Lastly, we conduct an analysis on the FocusFast method and demonstrate its efficacy in analyzing not only local details from relevant video segments but also in understanding the global video context through the simultaneous use of routing tokens, thereby facilitating a more comprehensive understanding of video content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 701 + ], + "type": "text", + "content": "Analysis of Retrieving Segments. By retrieving relevant video segments for the given queries, SALOVA can effectively target salient information in the long video and retain long context information. To further demonstrate the model's targeting capabilities beyond numerical performance in long video analysis, we explore our model's application in the Visual Needle-In-A-Haystack (V-NIAH) task [70], which extends the Needle-in-a-Haystack (NIAH) evaluation for LLMs to a vision-level benchmark. This task is particularly challenging as it requires models to not only detect but also precisely retrieve the sparse yet crucial visual cues scattered across lengthy videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "As in Fig. 3, we compare SALOVA against a base" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 173 + ], + "type": "image", + "image_path": "ca1a620b76da1ac252ac4fb290fa5ce8566e7356fabd8bf012e83a0c19238014.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 173, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 315, + 173, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 315, + 173, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 553, + 275 + ], + "type": "image", + "image_path": "61df37e920f86f5a27194cc2fe924acd75ca4571035844b0909ad159f7fe714a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "text", + "content": "Figure 3: Comparison results of V-NIAH. The x/y-axis indicates the total video frames and the location of needle image within the video, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 319, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 403 + ], + "type": "text", + "content": "line trained on sparsely sampled frames (16 frm w/o SR-Router). Our framework effectively identifies and extracts relevant video segments from densely packed content, even when handling long context lengths. These results highlight SALOVA's robustness in managing complex, long-form videos, maintaining contextual continuity by strategically focusing on relevant segments to user queries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 412, + 466, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 466, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 466, + 425 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 567 + ], + "type": "text", + "content": "Discussion. Despite SALOVA's competence in handling extended video sequences, it is important to recognize scenarios where its complex architecture may not be necessary. Specifically, for shorter videos where sparse sampling suffices to capture essential spatio-temporal information, simpler models could potentially outperform the efficiency of SALOVA without necessitating its extensive processing capabilities. This suggests a future avenue for integrating a hybrid approach based on our framework by dynamically adjusting the complexity of the retrieval and processing mechanisms based on the video length and content density." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Conclusion. In this paper, we introduce SALOVA, a novel framework designed to enhance the comprehension of long and untrimmed video by leveraging a retrieval-driven approach with new densely captioned dataset, the SceneWalk dataset. SALOVA strategically targets and processes only the relevant video segments, effectively addressing the structural limitations of current Video-LMMs with its Spatio-Temporal Connector and Segment Retrieval Router. Through extensive evaluation on various benchmarks, SALOVA exhibits its robust performance in interpreting complex video content, enhancing efficiency, and improving the understanding of extended videos." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3359" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was partially supported by two funds: IITP grant funded by the Korea government (MSIT) (RS-2022-II220984) and IITP grant funded by the Korea government (MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-Term Visual Memory Network)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "text", + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 240, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 305 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2, 4, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 361 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 475 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 294, + 532 + ], + "type": "text", + "content": "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "type": "text", + "content": "[7] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. arXiv preprint arXiv:2312.06742, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 622 + ], + "type": "text", + "content": "[8] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 4, 7, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 657 + ], + "type": "text", + "content": "[9] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[10] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "text", + "content": "[12] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 348 + ], + "type": "text", + "content": "[15] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "text", + "content": "[16] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "type": "text", + "content": "[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "type": "text", + "content": "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[20] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 678 + ], + "type": "text", + "content": "[21] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (sfd): A benchmark for story-level video understanding. arXiv preprint arXiv:2406.10221, 2024. 4, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 427, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 427, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 427, + 690 + ], + "type": "text", + "content": "[22] Google. Gemini, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[23] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3360" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13504-13514, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 163 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 229 + ], + "type": "text", + "content": "[25] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, et al. Language is not all you need: Aligning perception with language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 297 + ], + "type": "text", + "content": "[26] Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13700-13710, 2024. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 353 + ], + "type": "text", + "content": "[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions. arXiv preprint arXiv:2407.06358, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 420 + ], + "type": "text", + "content": "[28] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 421, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 295, + 465 + ], + "type": "text", + "content": "[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "text", + "content": "[30] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 566 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning. PMLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 633 + ], + "type": "text", + "content": "[32] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 634, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 295, + 689 + ], + "type": "text", + "content": "[33] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. Advances in neural information processing systems, 36, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[34] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 323-340. Springer, 2025. 1, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[35] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 4, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 2, 6, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 263 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[39] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2025. 1, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 365 + ], + "type": "text", + "content": "[40] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 410 + ], + "type": "text", + "content": "[41] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 1, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[42] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 488 + ], + "type": "text", + "content": "[43] OpenAI. ChatGPT. https://openai.com/blog/chatgpt/, 2023.1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 492, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 492, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 492, + 500 + ], + "type": "text", + "content": "[44] OpenAI. Gpt-4 technical report, 2023. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 505, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 505, + 512 + ], + "type": "text", + "content": "[45] OpenAI. GPT-4V(ision) System Card, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 513, + 459, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 459, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 459, + 524 + ], + "type": "text", + "content": "[46] OpenAI. Hello gpt-4o, 2024. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 525, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 590 + ], + "type": "text", + "content": "[47] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 657 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[49] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3361" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[50] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 7, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 139, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 294, + 161 + ], + "type": "text", + "content": "[51] Share. Sharegemini: Scaling up video caption data for multimodal large language models, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 226 + ], + "type": "text", + "content": "[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "type": "text", + "content": "[53] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1, 3, 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 380 + ], + "type": "text", + "content": "[54] Nandan Thakur, Nils Reimers, Johannes Daxenberger, and Iryna Gurevych. Augmented SBERT: Data augmentation method for improving bi-encoders for pairwise sentence scoring tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 296–310, Online, 2021. Association for Computational Linguistics. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "text", + "content": "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[56] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 492, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 535 + ], + "type": "text", + "content": "[57] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. arXiv preprint arXiv:2407.15754, 2024. 7, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "text", + "content": "[58] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[59] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[60] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[61] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "text", + "content": "[62] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "text", + "content": "[63] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[64] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "text", + "content": "[65] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[66] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9127–9134, 2019. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[67] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "text", + "content": "[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[69] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[70] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "text", + "content": "[71] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "text", + "content": "[72] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 1, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[73] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations. 2, 3, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3362" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_content_list.json b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b550ae4868908b8c6628498458c386f50cc97e9 --- /dev/null +++ b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_content_list.json @@ -0,0 +1,1482 @@ +[ + { + "type": "text", + "text": "SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than $0.2\\%$ Training Cost", + "text_level": 1, + "bbox": [ + 135, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiyang Mei, Pengyu Zhang, Mike Zheng Shou* Show Lab, National University of Singapore", + "bbox": [ + 295, + 203, + 699, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models like the Segment Anything Model (SAM) have significantly advanced promptable image segmentation in computer vision. However, extending these capabilities to videos presents substantial challenges, particularly in ensuring precise and temporally consistent mask propagation in dynamic scenes. SAM 2 attempts to address this by training a model on massive image and video data from scratch to learn complex spatiotemporal associations, resulting in huge training costs that hinder research and practical deployment. In this paper, we introduce SAM-I2V, an effective image-to-video upgrading method for cultivating a promptable video segmentation (PVS) model. Our approach strategically upgrades the pretrained SAM to support PVS, significantly reducing training complexity and resource requirements. To achieve this, we introduce three key innovations: (i) an image-to-video feature extraction upgrader built upon SAM's static image encoder to enable spatiotemporal video perception, (ii) a memory filtering strategy that selects the most relevant past frames for more effective utilization of historical information, and (iii) a memory-as-prompt mechanism leveraging object memory to ensure temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that our method achieves over $90\\%$ of SAM 2's performance while using only $0.2\\%$ of its training cost. Our work presents a resource-efficient pathway to PVS, lowering barriers for further research in PVS model design and enabling broader applications and advancements in the field. Code and model will be available at: https://github.com/showlab/SAM-I2V.", + "bbox": [ + 86, + 305, + 485, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 789, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models have revolutionized artificial intelligence by leveraging massive datasets and large-scale models to achieve remarkable performance across various tasks. In computer vision, the Segment Anything Model", + "bbox": [ + 89, + 814, + 483, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3a48662775cefd0497c75ed6faf677746e6f994772e0fbb663a58c459526cbe.jpg", + "image_caption": [ + "Figure 1. Comparison on promptable video segmentation (PVS). We depict the online-PVS performance (in $J \\& F$ ) averaged on four benchmark datasets [1, 18, 36, 41] (vertical axis) with respect to model training cost (horizontal axis). The size of circles indicates model parameters (in M). SAM 2.1 [36] is a strong PVS model but needs huge training cost (i.e., 256 A100 GPUs for 108 hours). We propose to cultivate a PVS model via image-to-video model upgrading. Building on the pre-trained SAM [21, 40], incorporating existing video object segmentation models (e.g., XMem++ [1]) or temporal association modules (e.g., temporal adapter [36]) significantly reduces training costs compared to SAM 2.1 [36]. Furthermore, our designed SAM-I2V helps achieve over $90\\%$ of SAM 2.1's performance while using only $0.2\\%$ of its training cost." + ], + "image_footnote": [], + "bbox": [ + 522, + 277, + 903, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(SAM) [21] has emerged as a prominent foundation model for general-purpose promptable image segmentation, enabling users to segment objects/regions in images based on prompts such as points, boxes, or masks. While SAM demonstrates impressive capabilities on static images, extending these abilities to video data introduces significant challenges due to the temporal dynamics and consistency required in video segmentation tasks.", + "bbox": [ + 511, + 686, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, SAM 2 [36] has been developed to address promptable video segmentation (PVS), aiming to extend SAM's image segmentation capabilities to videos. SAM 2 is trained from scratch on a massive collection of images and videos to learn complex spatiotemporal associations. However, it demands exorbitant computational resources,", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 114, + 887, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3417", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "requiring 256 A100 GPUs over 108 hours for training. Such high computational costs create substantial barriers to exploring alternative architectures and advancing research in PVS, limiting accessibility for researchers and practitioners.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This raises a critical question: Can we achieve promptable video segmentation through a resource-efficient approach? If so, it would open opportunities for developing diverse PVS foundation models, benefiting a wide range of downstream applications. Addressing this challenge, we propose SAM-I2V, an efficient method to upgrade the pretrained image-based SAM [21] model to support promptable video segmentation with affordable training cost (Figure 1). Our approach strategically leverages SAM's strong promptable image segmentation capabilities and develops image-to-video upgrading techniques to extend functionality from images to videos.", + "bbox": [ + 88, + 152, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge the gap between promptable image and video segmentation, our SAM-I2V includes three key components: a temporal feature integrator TFI, a memory selective associator MSA, and a memory prompt generator MPG. (i) The TFI enhances SAM's image encoder by integrating temporal information, enabling the extraction of spatiotemporal features essential for understanding dynamic scenes; (ii) The MSA retrieves relevant segmentation information from preceding video frames through a new memory filtering mechanism, selectively associating historical information to guide the current frame's segmentation; (iii) The MPG incorporates memory as an additional prompt, utilizing object memory to ensure precise and temporally consistent mask propagation across frames. Comprehensive experiments validate that our method significantly reduces training complexity and resource requirements, achieving over $90\\%$ of SAM 2's performance with only $0.2\\%$ of its training cost. In summary, our contributions are:", + "bbox": [ + 91, + 332, + 482, + 603 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We present SAM-I2V, the first attempt to upgrade a pre-trained image-based SAM model to support promptable video segmentation, enabling PVS capabilities with substantially reduced training costs.", + "2. We design an upgrade method that extends SAM's static image encoder into a video perception module, enabling spatiotemporal feature extraction for PVS.", + "3. We propose a memory filtering strategy that selects the most relevant past frames to facilitate more effective utilization of historical information for memory-guided video segmentation.", + "4. We introduce a memory-as-prompt mechanism that leverages object memory to ensure temporally consistent mask propagation in dynamic scenes." + ], + "bbox": [ + 102, + 604, + 478, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 829, + 238, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Promptable Visual Segmentation. The Segment Anything Model (SAM) [21] has set a significant milestone in promptable segmentation. SAM is designed to gen", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erate a valid segmentation mask for an object of interest in an image when provided with input prompts such as bounding boxes or reference points. Trained on the extensive SA-1B dataset, SAM facilitates zero-shot segmentation with its versatile prompting mechanism, making it highly adaptable for diverse downstream tasks. This foundational work has spurred further innovations aimed at enhancing SAM's quality and efficiency. For instance, models like HQ-SAM [20] have focused on improving output precision through high-quality token integration and fine-tuning with detailed masks. Efficiency-oriented models such as EfficientSAM [47], MobileSAM [53, 54], FastSAM [56], TinySAM [40], and SlimSAM [6] have broadened the scope of SAM's applications, enabling its deployment in more resource-constrained and real-world scenarios, including medical imaging [25, 27, 43] and remote sensing [3, 37, 45]. Building upon the success of SAM [21], SAM 2 [36] represents an evolution toward promptable video segmentation, addressing the complexities of propagating segmentation across multiple frames. This allows for interaction at any frame of a video, accepting prompts such as positive/negative clicks, bounding boxes, or masks to define or refine object segmentation. Unlike its predecessor, SAM 2 not only responds to initial prompts on single frames but also propagates these segmentation cues throughout the entire video, maintaining temporal consistency and adaptability across frames. Evaluated through interactive online and offline settings, SAM 2's integration of promptable video interaction ensures comprehensive and dynamic segmentation, pushing the boundaries of practical real-world applicability [5, 15, 23, 39, 46, 58].", + "bbox": [ + 511, + 90, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interactive Video Object Segmentation. Interactive video object segmentation (IVOS) is an essential task for enabling real-time object tracking and segmentation in videos through user input, such as clicks or bounding boxes. Recent studies [8, 14, 17] have aimed to develop modular approaches that allow segmenting an object in a single frame and then propagating this segmentation across subsequent frames to maintain consistency. Building upon this, many IVOS methods leverage interactive mechanisms to guide segmentation over multiple frames. While SAM has been used for video object segmentation combined with video trackers based on masks [10, 12, 49] or points [35], these approaches face certain limitations: trackers may not be effective for all objects, SAM may perform suboptimally on certain video frames, and there is no interactive mechanism for correcting model errors aside from re-annotating erroneous frames and restarting the tracking process from that point. This highlights expanded functionality of SAM 2 that provides a flexible solution to enable interactive refinement.", + "bbox": [ + 511, + 563, + 903, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-Supervised Video Object Segmentation. Semi-supervised video object segmentation (Semi-VOS) typically begins with a provided mask for the first frame,", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3418", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which acts as a reference for tracking the object through the video [32]. The term \"semi-supervised\" refers to using the initial mask as supervision, with no further manual input for subsequent frames. This task has become vital for various applications requiring automatic tracking, such as video editing, robotics, and automatic background removal. Early Semi-VOS techniques involved fine-tuning models on the first frame [2, 26, 38] or incorporating offline-trained models that predicted object masks for the entire video [28, 50, 52]. More advanced strategies integrated RNNs [48] and cross-attention [1, 7, 9, 11] to extend segmentation from the first frame throughout the sequence, improving object tracking accuracy. Recent methods [44, 55] have enhanced a single vision transformers to jointly handle the current frame alongside all prior frames and their corresponding predictions, simplifying the architecture but incuring a significant inference cost. Semi-VOS process is analogous to a specialized case of promptable video segmentation, where providing an high-quality mask prompt in the first frame is crucial but practically challenging and time-consuming, limiting the practical application of Semi-VOS in real-world scenarios.", + "bbox": [ + 89, + 90, + 480, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-to-Video Transfer Learning. In recent years, the importance of image-to-video transfer learning has grown due to its significant potential for reducing training costs and enhancing adaptability in video-based tasks. To facilitate efficient image-to-video transfer, various methods have been introduced, including spatio-temporal adapters [29, 51], dual-path adaptation [30], disentangled spatial and temporal learning [33], and learnable prompts [19, 22]. These methods incorporate lightweight, trainable components that leverage pre-trained image foundation models such as ViT [16] and CLIP [34] while integrating spatiotemporal dynamics, striking a balance between computational efficiency and performance to enhance video applications. These methods, although centered around imaged-to-video transfer for video recognition and understanding tasks, offer insights into our efficient cultivation of a promtable video segmentation model (i.e., SAM 2 [36]) via upgrading the pre-trained image-based SAM model [21] to a video one. These approaches provide a solid foundation for spatio-temporal information learning in our image-to-video upgradation but lack direct applicability for addressing segmentation-specific challenge, i.e., ensuring precise and temporally consistent mask propagation in scene dynamics. This makes our focus on SAM to SAM 2 upgradation a distinct and challenging endeavor.", + "bbox": [ + 91, + 422, + 482, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 814, + 225, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While SAM 2 demonstrates strong capabilities, its exorbitant training demands—requiring 256 A100 GPUs over 108 hours—present a significant barrier to exploring alternative architectures and advancing research in PVS. To overcome", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "this challenge, we aim to develop an efficient method for cultivating a PVS model with affordable training costs. Our strategy leverages the pre-trained SAM model, which already performs well on single frames, and upgrades it to SAM 2. We design three components to achieve this transformation: a Temporal Feature Integrator (Section 3.2), a Memory Associator (3.3), and a Memory Prompt Generator (3.4), which collectively upgrade image-based SAM into a promptable video segmentation model.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 247, + 624, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our SAM-I2V enables image-based SAM model can process video input for promtable video segmentation. As illustrated in Figure 2, the video frames are passed through a backbone feature extractor, which processes the current frame using a series of image encoder layers. These encoder outputs are enriched with temporal context through a temporal feature integrator, enabling the capture of dynamic information across consecutive frames. The enhanced features are subsequently input into the memory selective associator and memory prompt generator modules. These components manage memory states and generate target object prompts informed by past frames data, ensuring continuity and contextual awareness throughout the video. Optional user-provided prompts (e.g., masks, points, or bounding boxes) are processed by the prompt encoder to guide the segmentation with user-specific inputs. Finally, the integrated information is directed into the mask decoder, which produces the final segmentation mask for the current frame.", + "bbox": [ + 511, + 272, + 906, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Temporal Feature Integrator", + "text_level": 1, + "bbox": [ + 511, + 565, + 772, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance SAM's image encoder, which is limited to extracting only spatial features, into a video perception module, we introduce a Temporal Feature Integrator (TFI). The TFI comprises a temporal branch and an integration branch that operates synergistically to iteratively extract and integrate temporal features, enabling a comprehensive spatiotemporal feature extraction.", + "bbox": [ + 511, + 590, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, given the input video $X \\in \\mathbb{R}^{c \\times n \\times h \\times w}$ , where $c$ , $n$ , $h$ and $w$ are the frame channel, number, height and width, respectively, SAM's image encoder processes the current frame $X_{t}$ via a 2D stem and $L$ spatial blocks, generating multi-stage spatial features $S_{l} \\in \\mathbb{R}^{c_{l} \\times h_{l} \\times w_{l}}$ , where $c_{l}$ , $h_{l}$ and $w_{l}$ denote the features channel, height and width at stage $l \\in \\{1, \\dots, L\\}$ , respectively. The extracted features are fed into our TFI to obtain spatiotemporal features $I_{l}$ .", + "bbox": [ + 511, + 700, + 903, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the temporal branch of TFI processes the video input $X$ through a 3D stem followed by $L$ temporal blocks to capture temporal contexts. In $l$ -th temporal block, we employ 3D convolutions $\\mathcal{C}_{3d}$ to progressively", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3419", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3e7b726652c6bb910033a2e025a5d4bd92ef5d6d0c5761217bc340afe0fca57e.jpg", + "image_caption": [ + "Figure 2. Overview of SAM-I2V. The input video is fed into a backbone feature extractor, which passes the current frame through a series of image encoder layers. The output of these layers is enhanced with temporal information by a temporal feature integrator to capture dynamic context across frames. The extracted features are then fed into the memory associator and the memory prompt generator, which manage memory states and generate target object prompts based on past frame information. The prompt encoder processes optional user input (e.g., masks, points, or bounding boxes) to guide the segmentation. Finally, the processed information flows into the mask decoder, which outputs the segmentation mask for current frame, enabling user-intention-followed or memory conditioned promptable video segmentation." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model temporal correspondences between frames:", + "bbox": [ + 89, + 315, + 421, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {l} ^ {\\prime} = T ^ {(3)},\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 339, + 271, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {(i)} = T ^ {(i - 1)} + \\mathcal {C} _ {3 d} ^ {(i)} \\big (T ^ {(i - 1)} \\big),\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 362, + 388, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T^{(0)} = T_{l - 1}$ is the input to the $l$ -th temporal block. To further enrich temporal features $T_{l}^{\\prime}$ with spatial details, we fuse it with intermediate features $I_{l}^{\\prime}$ from TFI's integration branch:", + "bbox": [ + 89, + 393, + 482, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {l} = \\mathcal {S T} (I _ {l} ^ {\\prime}, T _ {l} ^ {\\prime}),\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 454, + 346, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{ST}$ denotes the spatial-to-temporal fusion operation that slices out current-frame temporal features from $T_{l}^{\\prime}$ (i.e., the last slice in temporal dimension), concatenates it with current-frame spatial features $I_{l}^{\\prime}$ along the channel dimension, applies a $1\\times 1$ convolution to integrate concatenated features, and uses the resulting output to overwrite original current-frame temporal features in $T_{l}^{\\prime}$ .", + "bbox": [ + 89, + 477, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, TFI's integration branch combines spatial features $S_{l}$ and temporal features $T_{l}^{\\prime}$ to get spatiotemporal features:", + "bbox": [ + 89, + 583, + 483, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l} ^ {\\prime} = I _ {l - 1} + S _ {l},\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 625, + 318, + 641 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l} = \\mathcal {C} _ {2 d} \\big (\\mathcal {T S} (I _ {l} ^ {\\prime}, T _ {l} ^ {\\prime}) \\big),\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 645, + 362, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{l-1}$ is the input to the $l$ -th integration block. The temporal-to-spatial fusion operation $\\mathcal{T}S$ merges current-frame temporal features in $T_l'$ with $I_l'$ , producing temporal-enriched spatial features. These features then pass through a 2D convolution $C_{2d}$ , yielding the enhanced spatiotemporal output $I_l$ . Finally, the temporal and integration branches of TFI iteratively refine the features over multiple stages, enabling the model to progressively capture rich spatiotemporal context for the subsequent segmentation process.", + "bbox": [ + 89, + 672, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Memory Selective Associator", + "text_level": 1, + "bbox": [ + 89, + 818, + 349, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While image-based SAM [21] segments the target object based on user prompts, in the PVS task, segmentations from preceding video frames can serve as target indicators to guide the segmentation of the target in the current frame.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f1783dfe9a12cadfa1d1fb48e1a00d0a5e1d7e8f588574fdecfbd486caf115b7.jpg", + "image_caption": [ + "Figure 3. Structural details of temporal feature integrator (TFI). It consists of a temporal branch and an integration branch, enhancing SAM's image encoder to function as a video perceptron." + ], + "image_footnote": [], + "bbox": [ + 526, + 316, + 897, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This motivates our Memory Selective Associator (MSA), designed to retrieve relevant and beneficial segmentation information from preceding video frames, leveraging this memory to enhance the current frame's features.", + "bbox": [ + 511, + 536, + 905, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MSA first utilizes convolution to fuse the features from the previous frames $\\{F_{t - p}\\}_{p = 1}^{P}$ and their corresponding segmentation masks $\\{M_{t - p}\\}_{p = 1}^{P}$ . These fused features are then passed to the memory selector, which chooses the most relevant past frame features. The selected features are used in the memory attention module, where cross-attention between the current frame features $I_{t}$ and the selected previous frame features is applied, producing the memory-enhanced features $I_{t}^{\\prime}$ .", + "bbox": [ + 511, + 597, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The memory encoder and attention mechanism in our MSA follow a similar structure as in SAM 2 [36]. Unlike SAM 2 [36], which selects the previous $z$ consecutive frames leading up to the current frame, our memory selector chooses $z$ frames based on their relevance, as determined by similarity to the current frame. Specifically, it first chooses the prompt frame based on the user input, and then divides the previous frames into $X$ short-term (local) frames and $Y$ long-term (global) frames. The similarity between the current frame features $I_{t} \\in \\mathbb{R}^{N}$ and $X/Y$ previous frame features $F \\in \\mathbb{R}^{X/Y \\times N}$ in each set is computed to identify", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3420", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65d921fc0f56c614ab9d3e4067f5d19a7eab7737565ad4fdbaa17bbfe7733f2a.jpg", + "image_caption": [ + "Figure 4. Illustration of our memory selective associator (MSA). It consists of a memory encoder, a memory selector, and memory attention." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 906, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the most relevant local/global frames:", + "bbox": [ + 89, + 276, + 341, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {l o c} = F ^ {l o c} \\cdot I _ {t} \\in \\mathbb {R} ^ {X}, S ^ {g l o} = F ^ {g l o} \\cdot I _ {t} \\in \\mathbb {R} ^ {Y},\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 303, + 452, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N = c_{L} \\times h_{L} \\times w_{L}$ . Next, we normalize the similarity scores to obtain the probability distribution $D^{loc}$ and $D^{glo}$ , respectively:", + "bbox": [ + 89, + 335, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {l o c} = \\frac {e ^ {S ^ {l o c}}}{\\sum_ {j} e ^ {S _ {j} ^ {l o c}}} \\in \\mathbb {R} ^ {X}, D ^ {g l o} = \\frac {e ^ {S ^ {g l o}}}{\\sum_ {j} e ^ {S _ {j} ^ {g l o}}} \\in \\mathbb {R} ^ {Y}.\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 393, + 465, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then sample $x$ local frames and $y = z - x$ global frames based on $D^{loc}$ and $D^{glo}$ :", + "bbox": [ + 89, + 450, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ni _ {k} ^ {l o c} \\sim P (i _ {k} ^ {l o c} = i) = D ^ {l o c} (i), \\quad k \\in \\{1, 2, \\ldots , X \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 493, + 460, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ni _ {k} ^ {g l o} \\sim P (i _ {k} ^ {g l o} = i) = D ^ {g l o} (i), \\quad k \\in \\{1, 2, \\ldots , Y \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 527, + 460, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i_k^{loc}$ and $i_k^{glo}$ denote the indices of the selected local and global frames, respectively, and $D^{loc}(i)$ and $D^{glo}(i)$ are the corresponding probabilities for each frame. The selected frames are subsequently used to enhance the segmentation of the current frame through memory attention, resulting in memory-associated features $I_t'$ .", + "bbox": [ + 89, + 559, + 483, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Memory Prompt Generator", + "text_level": 1, + "bbox": [ + 89, + 664, + 341, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SAM's mask decoder segments images using image features and user prompts to achieve intention-followed results. Motivated by its prompt integration capability, we propose incorporating memory as an additional prompt. This enables the model to leverage segmentation information from previous frames, enhancing temporal consistency across the video sequence.", + "bbox": [ + 89, + 686, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 illustrates the proposed Memory Prompt Generator (MPG). We introduce a set of learnable memory prompt tokens $G \\in \\mathbb{R}^{g \\times d}$ , where $g$ is the number of memory prompts and $d$ is the embedding dimension. These tokens interact with the flattened memory features $F_{\\mathrm{mem}} \\in \\mathbb{R}^{(mh_Lw_L) \\times c_L}$ , where $m$ is the number of memory frames, guided by the corresponding binarized segmentation masks", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef25dd9d326ff9b7bbca6d6489a16bd817eddf0871266fd60879856f56c63604.jpg", + "image_caption": [ + "Figure 5. Illustration of the memory prompt generator (MPG)." + ], + "image_footnote": [], + "bbox": [ + 524, + 279, + 897, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$M_{\\mathrm{mem}} \\in \\mathbb{R}^{(mh_Lw_L) \\times 1}$ via masked cross-attention:", + "bbox": [ + 513, + 411, + 846, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG ^ {\\prime} = \\operatorname {S o f t m a x} (\\mathcal {M} + Q K ^ {\\top}) V + G, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 438, + 906, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\psi (G), \\quad K = \\psi \\left(F _ {\\mathrm {m e m}}\\right), \\quad V = \\psi \\left(F _ {\\mathrm {m e m}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 458, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (x, y) = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} M _ {\\mathrm {m e m}} (x, y) = 1, \\\\ - \\infty , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 479, + 905, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi$ is the learnable linear transformation. The updated memory prompt tokens $G^{\\prime}$ , having captured foreground context from memories, are then refined through self-attention to incorporate intra-token dependencies:", + "bbox": [ + 511, + 529, + 906, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG ^ {\\prime \\prime} = \\operatorname {S e l f A t t n} \\left(G ^ {\\prime}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 602, + 777, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the refined memory prompt embeddings $G''$ are passed through a multi-layer perceptron (MLP) to produce the final updated memory prompt tokens:", + "bbox": [ + 511, + 631, + 906, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG ^ {\\prime \\prime \\prime} = \\operatorname {M L P} \\left(G ^ {\\prime \\prime}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 688, + 769, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Such an iterative attention-based token generation enables the MPG to selectively capture target object information from memories, providing guidance for the mask decoder.", + "bbox": [ + 511, + 718, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 777, + 645, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 803, + 702, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. We train our SAM-I2V jointly on 10,000 SA-1B [21] images and 50,583 SAV [36] videos. The SA-1B dataset [21] is a comprehensive collection consisting of over 1.1 billion high-quality segmentation masks from 11 million licensed and privacy-preserving images.", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3421", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This dataset was assembled using a multi-phase data engine where initial stages involved human annotation aided by interactive tools, and subsequent phases transitioned to fully automated mask generation through the SAM model. The rich variety of mask annotations allows for robust training that supports zero-shot generalization and diverse segmentation tasks, proving effective in many downstream applications. The Segment Anything Video (SA-V) dataset [36] complements the image-focused SA-1B by providing extensive coverage for video segmentation. Its training set includes 50,583 videos and 642,036 masklets. SA-V is tailored for video scenarios involving complex object motions, occlusions, and reappearance, making it invaluable for advancing video-based models like SAM 2 [36].", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Configurations. Following [36], to simulate an interactive training setting, we sample 8-frame sequences, designating the first frame as the prompt frame and randomly selecting up to 1 additional frame from the remaining seven frames for corrective clicks. The initial prompt for the first frame can be the ground-truth mask (50% probability), a positive click from the ground-truth mask (25%), or a bounding box input (25%). The corrective clicks are sampled from the error region between the model prediction and the ground-truth mask (90%) or directly from the ground-truth mask (10%). Our SAM-I2V model is implemented using PyTorch [31] and trained for 10 epochs on 8 RTX A5000 GPUs (24GB memory). We employ the AdamW optimizer [24] and apply layer-wise decay [13] specifically to the image encoder. The initial learning rates are set to $3 \\times 10^{-4}$ for the image encoder and $6 \\times 10^{-5}$ for the other components, respectively. These learning rates are subsequently decayed using a cosine scheduler. We supervise the model's predictions using a linear combination of mask prediction loss for the output mask, meanabsolute-error (MAE) loss for the IoU prediction, and cross-entropy loss for object prediction with a ratio of 20:1:1, respectively. The mask prediction loss $\\mathcal{L}_m$ is defined as the combination of the weighted binary cross-entropy (BCE) loss $\\ell_{wbce}$ [42] and the weighted IoU loss $\\ell_{wiou}$ [42], i.e., $\\mathcal{L}_m = \\ell_{wbce} + \\ell_{wiou}$ , to ensure the model pays more attention to potential distraction regions such as object boundaries, elongated areas, or holes. Notably, all SAM-I2V modules except the TinySAM [40] backbone are randomly initialized and trained from scratch without leveraging any weights from SAM 2 [36], ensuring complete independence from SAM 2 [36] in the training process.", + "bbox": [ + 91, + 306, + 483, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Settings. To comprehensively assess the capabilities of our SAM-I2V, we perform comparisons on four benchmark datasets: ESD [18] is a high-quality 3D spatial and temporal dataset for object segmentation in an indoor cluttered environment; PUMaVOS [1] contains videos with segments around object parts such as a person's cheek; LV-VIS [41] contains videos from a diverse set of open-", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "vocabulary object categories; and SAV-Test [36] is the test set of SA-V dataset, consist of 150 videos and 278 masklets. For these evaluations, we report the standard $J \\& F$ metric [32] which combines region similarity (Jaccard index) and boundary accuracy (F-measure). We conducted evaluation on the promptable video segmentation. In this task, the model works in an interactive manner, where the point prompts are used to identify the target of interest or to refine the current segmentation result. Following [36], two settings are included in the comparison, namely offline evaluation (PVS-offline) and online evaluation (PVS-online). In both settings, the target is firstly initialized via $N_{\\mathrm{click}}$ point prompts. As for PVS-offline, the model segments the target throughout the entire video in $N_{\\mathrm{pass}}$ passes, where the interactive frame is selected based on the largest model error (i.e., with lowest IoU between the prediction and ground truth masks) and $N_{\\mathrm{click}}$ points are added on such frame to refine the mask in each pass. All the prompts are used to guide the segmentation process in the next pass. PVS-online follows one pass evaluation protocol, where the segmentation process is paused and refined when the result is not reliable (IoU $< 0.75$ with ground-truth). Also, $N_{\\mathrm{click}}$ points are added on the refined frame to generate corrected mask, which is used for subsequent frames. The refine process is repeated $N_{\\mathrm{frame}}$ times.", + "bbox": [ + 511, + 90, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 482, + 803, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed SAM-I2V, we compared it against seven state-of-the-art methods on both online and offline PVS tasks under different click settings, as reported in Table 1 and 2. Besides SAM 2.1 [36], we also include upgradation methods that combine SAM (both base [21] and tiny [40] variants) with existing state-of-the-art video object segmentation models, specifically Xmem++ [1] and Cutie [11], as well as approaches that upgrade SAM by incorporating temporal adaptation (TA) [36] and spatial adaptation (SA) [4] modules.", + "bbox": [ + 511, + 505, + 906, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Comparison with SAM 2.1. While SAM 2.1 [36] serves as a strong baseline with an overall accuracy of 71.9 (online) and 73.3 (offline), it requires substantial computational resources (2.2 million $G \\cdot$ Hour). In contrast, our SAM-I2V achieves its $91\\%$ (i.e., 65.3, online) and $90\\%$ (i.e., 66.0, offline) overall scores, with its $0.2\\%$ (i.e., 4.6k $G \\cdot$ Hour) training cost and its $49\\%$ (i.e., 18.9m vs. 38.9m) model parameters.", + "bbox": [ + 511, + 657, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness over Existing SAM Upgradation Methods. Compared to SAM [21] + XMem++ [1] (157.0M parameters) and TinySAM [40] + XMem++ (72.3M parameters), our SAM-I2V offers a substantial reduction in model size (18.9M) while delivering comparable or superior performance. For example, under the online setting, SAM-I2V helps achieve an overall score of 65.3, outperforming SAM + XMem++ (60.4) and SAM + Cutie [11] (61.7).", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3422", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f084450992f4ef6f84dd30bcd61238e1eadbc57e153a7a57f6d11067760f19fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.645.859.366.386.866.879.575.288.870.182.877.279.760.974.972.971.9
SAM [21] + XMem++ [1]157.0-78.349.368.042.478.846.273.041.979.345.478.443.378.847.073.142.560.4
SAM [21] + Cutie [11]129.8-78.438.369.451.279.044.373.951.879.844.676.952.979.142.473.452.061.7
TinySAM [40] + XMem++ [1]72.3-79.355.268.547.382.959.874.948.784.260.476.951.682.158.573.449.265.8
TinySAM [40] + Cutie [11]45.1-80.653.071.158.483.352.676.259.984.456.278.863.182.853.975.460.568.1
TinySAM [40] + TA [36]17.84.6k68.728.856.242.980.245.172.349.182.149.876.550.977.041.268.447.658.6
TinySAM [40] + SA [4] + TA [36]18.74.6k54.834.652.844.179.352.772.950.681.959.476.851.972.048.967.548.959.3
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.039.153.850.584.760.873.458.286.465.978.359.581.455.368.556.165.3
", + "bbox": [ + 91, + 90, + 906, + 284 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61a879d80f89d8df6d03d86781ebf118d7efd728020b3e8c72b723e3f69bc888.jpg", + "table_caption": [ + "Table 1. Comparisons of the online promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance." + ], + "table_footnote": [], + "table_body": "
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.946.261.366.887.767.383.976.290.170.787.578.180.661.477.673.773.3
SAM [21] + XMem++ [1]157.0-49.536.850.141.164.049.353.744.165.354.755.244.959.646.953.043.450.7
TinySAM [40] + XMem++ [1]72.3-47.746.047.549.173.655.452.955.575.962.553.758.265.754.651.454.356.5
TinySAM [40] + TA [36]17.84.6k68.929.558.345.381.149.376.352.482.955.681.154.977.644.871.950.961.3
TinySAM [40] + SA [4] + TA [36]18.74.6k57.037.054.944.880.956.676.652.583.462.181.253.673.851.970.950.361.7
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.437.855.748.984.661.277.858.586.465.183.059.881.554.772.255.766.0
", + "bbox": [ + 91, + 352, + 906, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparisons of the offline promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance. Cutie [11] is excluded from the comparison as it only supports single-frame mask inputs, rendering it unsuitable for offline PVS settings.", + "bbox": [ + 89, + 521, + 906, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robust Performance Across Interaction Levels. The online and offline results presented in Table 1 and 2 reflect that the performance of all methods generally improves as the number of interactions increases. Our SAM-I2V consistently delivers superiority across 1-click, 3-click, and 5-click settings, underscoring its robustness to varying levels of user interaction.", + "bbox": [ + 89, + 603, + 482, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robust Online and Offline Performance. SAM-I2V's overall online performance in Table 1 is slightly lower than the offline performance in Table 2. This discrepancy arises because, unlike in the offline evaluation, in the online setting the new prompts affect only the frames following the current paused frame and not the preceding frames. Compared to other methods, our approach maintains an advantage in both online and offline settings.", + "bbox": [ + 89, + 714, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above results confirm that our SAM-I2V not only matches but often surpasses existing upgrading methods in performance with reduced computational costs, making SAM-I2V a practical choice for PVS tasks.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 603, + 666, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an extensive ablation study to validate the extensibility of SAM-I2V and the effectiveness of its each key component. Table 3, 4, 5, 6 summarize our findings.", + "bbox": [ + 511, + 626, + 903, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Applicability of SAM-I2V to Upgrade Different SAM Models. In the first ablation study, we investigate the performance of our SAM-I2V to upgrade different SAM models. From the results presented in Table 3, it is evident that our SAM-I2V effectively upgrades different SAM variants to a PVS model, demonstrate that SAM-I2V is a robust and versatile image-to-video promptable segmentation upgradation approach.", + "bbox": [ + 511, + 672, + 905, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the Temporal Feature Integrator. As shown in Table 4, our proposed TFI module outperforms the baseline (TinySAM [40] + Temporal Associator [36]) and other adaptation methods. While previous approaches like ST-Adapter [29] and Parallel Adapter [4] perform adaptation within each attention block, our TFI is specifically designed with temporal branch and integration branch to iter", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3423", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f32aac6d113b2db29f8dac568123a743ddde1f760e7617f31e7edf2b1191573e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsES[18]PU[1]LV[41]SV[36]OA
TinySAM [40] + TA [36]80.245.172.349.161.7
TinySAM [40] + SAM-I2V84.760.873.458.269.3
EdgeSAM [57] + TA [36]78.341.469.644.658.5
EdgeSAM [57] + SAM-I2V83.251.971.654.465.1
", + "bbox": [ + 96, + 89, + 477, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/72b2954c1c2b0c9f6880c4abd46ae18bb8a8418a7fc189c03ad68f4cc554c417.jpg", + "table_caption": [ + "Table 3. Ablation on the applicability of SAM-I2V to upgrade different SAM models for PVS task under the online, 3-click setting." + ], + "table_footnote": [], + "table_body": "
MethodsES[18]PU[1]LV[41]SV[36]OA
Base80.245.172.349.161.7
Base + ST-Adapter [29]79.354.173.350.164.2
Base + Parallel Adapter [4]79.352.772.950.663.9
Base + TFI81.358.172.452.566.1
", + "bbox": [ + 96, + 217, + 477, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "atively extract and integrate spatial-temporal features. This allows our model to capture richer temporal dependencies, leading to superior performance.", + "bbox": [ + 88, + 354, + 482, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Memory Selective Associator. Relying on the previous $z = 6$ consecutive frames for segmentation is equivalent to setting $x$ and $y$ in MSA to 0 and 6, respectively. By varying the number of selected global $(x)$ and local $(y)$ frames in our MSA, we observe that the segmentation performance improves when the MSA selectively retrieves the most relevant frames based on their similarity to the current frame. Specifically, the configuration with $x = 3$ global frames and $y = 3$ local frames performs better, highlighting the importance of incorporating both short-term and long-term memory. In addition, the ability of MSA to select frames based on relevance rather than uniform sampling allows the model to focus on the most relevant past frames, leading to improved performance.", + "bbox": [ + 89, + 401, + 482, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Memory Prompt Generator. From Table 6, we can observe that, (i) introducing MPG $(w/$ MPG-1/3/5) consistently improves performance over the one without MPG, demonstrating the effectiveness of the memory-as-prompt strategy in enhancing segmentation accuracy; (ii) varying the number of memory prompt tokens among 1, 3, and 5 yields comparable results, indicating that the MPG is robust to the choice of token quantity within this range; and (iii) removing the Masked Cross-Attention (MCA) mechanism (i.e., $w/$ MPG-3 ( $w/o$ MCA)) leads to a performance drop, highlighting the importance of MCA in focusing on relevant foreground information, eliminating background interference, and thus enhancing the quality of memory prompt generation.", + "bbox": [ + 89, + 612, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 89, + 832, + 348, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One limitation of our approach is that its accuracy does not well match that of SAM 2.1 [36]. Note that the released SAM 2.1 model was trained not only on the SAM-1B", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bd2b304d6882f461978c210308c6e852a37ac68018d502f6170ee95af47a21f1.jpg", + "table_caption": [ + "Table 4. Ablation on the effectiveness of the temporal feature integrator TFI for online PVS task under the 3-click setting." + ], + "table_footnote": [], + "table_body": "
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MSA (i.e., x=0, y=6)84.054.473.256.967.1
w/ MSA (x=1, y=5)84.754.873.457.967.7
w/ MSA (x=2, y=4)84.556.373.456.967.8
w/ MSA (x=3, y=3)84.760.873.458.269.3
w/ MSA (x=3, y=3) w/ UFS84.156.073.056.567.4
", + "bbox": [ + 519, + 90, + 901, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/575cb2d12fb303d1edd990969327f5332ac2fb487d4d299dc67a74212a0ff6ed.jpg", + "table_caption": [ + "Table 5. Ablation on the effectiveness of the memory selective associator MSA for online PVS task under the 3-click setting. \"x\" and \"y\" denote the number of selected global and local frames, respectively. \"UFS\" indicates the uniform frame sampling strategy." + ], + "table_footnote": [], + "table_body": "
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MPG82.157.772.153.266.3
w/ MPG-184.058.473.055.667.8
w/ MPG-384.760.873.458.269.3
w/ MPG-584.558.673.657.168.5
w/ MPG-3 (w/o MCA)84.958.172.855.167.7
", + "bbox": [ + 521, + 258, + 898, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation on the effectiveness of the memory prompt generator MPG for online PVS task under the 3-click setting. \"MPG-g\" denotes the MPG with $g$ learnable memory prompt tokens. \"MCA\" indicates the masked cross attention.", + "bbox": [ + 511, + 357, + 906, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and SAV datasets but also on 62.9k proprietary, non-public videos, which our method was unable to utilize. Despite this, our method demonstrates superior performance compared to baseline approaches and creates targeted model design opportunities for downstream applications in resource-constrained academic scenarios. Future work will focus on enhancing our model by incorporating additional publicly available datasets and exploring advanced training techniques to bridge the accuracy gap with SAM 2.1 [36].", + "bbox": [ + 511, + 439, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced SAM-I2V, a training-efficient framework that upgrades the pre-trained Segment Anything Model (SAM) to support promptable video segmentation (PVS). SAM-I2V consists of three key innovations: an image-to-video feature extraction upgrader that extends SAM's static image encoder for spatiotemporal video perception; a memory filtering strategy that selects the most relevant past frames to enhance the utilization of historical information; and a memory-as-prompt mechanism that leverages object memory for temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that SAM-I2V achieved over $90\\%$ of SAM 2's performance at only $0.2\\%$ of its training cost. This remarkable training efficiency significantly lowers the barriers for developing diverse PVS models, facilitating rapid innovation and broader practical applications.", + "bbox": [ + 511, + 613, + 906, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012).", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maksym Bekuzarov, Ariana Bermudez, Joon-Young Lee, and Hao Li. Xmem++: Production-level video segmentation from few annotated frames. In ICCV, 2023. 1, 3, 6, 7, 8", + "[2] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, 2017. 3", + "[3] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2", + "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 6, 7, 8", + "[5] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv:2408.04579, 2024. 2", + "[6] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. $0.1\\%$ data makes segment anything slim. NeurIPS, 2024. 2", + "[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3", + "[8] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 2", + "[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. NeurIPS, 2021. 3", + "[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 2", + "[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In CVPR, 2024. 3, 6, 7", + "[12] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv:2305.06558, 2023. 2", + "[13] K Clark. Electra: Pre-training text encoders as discriminators rather than generators. arXiv:2003.10555, 2020. 6", + "[14] Thanos Delatolas, Vicky Kalogeiton, and Dim P Papadopoulos. Learning the what and how of annotation in video object segmentation. In WACV, 2024. 2", + "[15] Haoyu Dong, Hanxue Gu, Yaqian Chen, Jichen Yang, and Maciej A Mazurowski. Segment anything model 2: an application to 2d and 3d medical images. arXiv:2408.00756, 2024. 2", + "[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Yuk Heo, Yeong Jun Koh, and Chang-Su Kim. Interactive video object segmentation using global and local transfer modules. In ECCV, 2020. 2", + "[18] Xiaoqian Huang, Kachole Sanket, Abdulla Ayyad, Fariborz Baghaei Naeini, Dimitrios Makris, and Yahya Zweiri. A neuromorphic dataset for object segmentation in indoor cluttered environment. arXiv:2302.06301, 2023. 1, 6, 7, 8", + "[19] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3", + "[20] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. NeurIPS, 2023. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7", + "[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 3", + "[23] Haofeng Liu, Erli Zhang, Junde Wu, Mingxuan Hong, and Yueming Jin. Surgical sam 2: Real-time segment anything in surgical video by efficient frame pruning. arXiv:2408.07931, 2024. 2", + "[24] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6", + "[25] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 2024. 2", + "[26] K-K Maninis, Sergi Caelles, Yuhua Chen, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. Video object segmentation without temporal information. IEEE TPAMI, 2018. 3", + "[27] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 2023. 2", + "[28] Seoung Wug Oh, Joon-Young Lee, Kalyan Sunkavalli, and Seon Joo Kim. Fast video object segmentation by reference-guided mask propagation. In CVPR, 2018. 3", + "[29] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. NeurIPS, 2022. 3, 7, 8", + "[30] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dualpath adaptation from image to video transformers. In CVPR, 2023. 3", + "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6", + "[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In ICCV, 2023. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[35] Frano Rajic, Lei Ke, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu. Segment anything meets point tracking. arXiv:2307.01197, 2023. 2", + "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[37] Simiao Ren, Francesco Luzi, Saad Lahrichi, Kaleb Kassaw, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Segment anything, from space? In WACV, 2024. 2", + "[38] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In CVPR, 2020. 3", + "[39] Chuyun Shen, Wenhao Li, Yuhang Shi, and Xiangfeng Wang. Interactive 3d medical image segmentation with sam 2. arXiv:2408.02635, 2024. 2", + "[40] Han Shu, Wenshuo Li, Yehui Tang, Yiman Zhang, Yihao Chen, Houqiang Li, Yunhe Wang, and Xinghao Chen. Tinysam: Pushing the envelope for efficient segment anything model. arXiv:2312.13789, 2023. 1, 2, 6, 7, 8", + "[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 6, 7, 8", + "[42] Jun Wei, Shuhui Wang, and Qingming Huang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 6", + "[43] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv:2304.12620, 2023. 2", + "[44] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, 2023. 3", + "[45] Aoran Xiao, Weihao Xuan, Heli Qi, Yun Xing, Ruijie Ren, Xiaojin Zhang, Ling Shao, and Shijian Lu. Cat-sam: Conditional tuning for few-shot adaptation of segment anything model. In ECCV, 2024. 2", + "[46] Xinyu Xiong, Zihuang Wu, Shangyi Tan, Wenxue Li, Feilong Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation. arXiv:2408.08870, 2024. 2", + "[47] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "masked image pretraining for efficient segment anything. In CVPR, 2024. 2", + "[48] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 3", + "[49] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv:2304.11968, 2023. 2", + "[50] Linjie Yang, Yanran Wang, Xuehan Xiong, Jianchao Yang, and Aggelos Katsaggellos. Efficient video object segmentation via network modulation. In CVPR, 2018. 3", + "[51] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In ICLR, 2023. 3", + "[52] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3", + "[53] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv:2306.14289, 2023. 2", + "[54] Chaoning Zhang, Dongshen Han, Sheng Zheng, Jinwoo Choi, Tae-Ho Kim, and Choong Seon Hong. Mobilesamv2: Faster segment anything to everything. arXiv:2312.09579, 2023. 2", + "[55] Jiaming Zhang, Yutao Cui, Gangshan Wu, and Limin Wang. Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv:2308.13505, 2023. 3", + "[56] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv:2306.12156, 2023. 2", + "[57] Chong Zhou, Xiangtai Li, Chen Change Loy, and Bo Dai. Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam. arXiv:2312.06660, 2023. 8", + "[58] Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv:2408.00874, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 643 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3426", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_model.json b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28be80cf6e54e965ee82c8f08074782b45e11eca --- /dev/null +++ b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than \\(0.2\\%\\) Training Cost" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.204, + 0.7, + 0.24 + ], + "angle": 0, + "content": "Haiyang Mei, Pengyu Zhang, Mike Zheng Shou* Show Lab, National University of Singapore" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.306, + 0.486, + 0.761 + ], + "angle": 0, + "content": "Foundation models like the Segment Anything Model (SAM) have significantly advanced promptable image segmentation in computer vision. However, extending these capabilities to videos presents substantial challenges, particularly in ensuring precise and temporally consistent mask propagation in dynamic scenes. SAM 2 attempts to address this by training a model on massive image and video data from scratch to learn complex spatiotemporal associations, resulting in huge training costs that hinder research and practical deployment. In this paper, we introduce SAM-I2V, an effective image-to-video upgrading method for cultivating a promptable video segmentation (PVS) model. Our approach strategically upgrades the pretrained SAM to support PVS, significantly reducing training complexity and resource requirements. To achieve this, we introduce three key innovations: (i) an image-to-video feature extraction upgrader built upon SAM's static image encoder to enable spatiotemporal video perception, (ii) a memory filtering strategy that selects the most relevant past frames for more effective utilization of historical information, and (iii) a memory-as-prompt mechanism leveraging object memory to ensure temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that our method achieves over \\(90\\%\\) of SAM 2's performance while using only \\(0.2\\%\\) of its training cost. Our work presents a resource-efficient pathway to PVS, lowering barriers for further research in PVS model design and enabling broader applications and advancements in the field. Code and model will be available at: https://github.com/showlab/SAM-I2V." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.79, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.485, + 0.878 + ], + "angle": 0, + "content": "Foundation models have revolutionized artificial intelligence by leveraging massive datasets and large-scale models to achieve remarkable performance across various tasks. In computer vision, the Segment Anything Model" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.279, + 0.905, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.671 + ], + "angle": 0, + "content": "Figure 1. Comparison on promptable video segmentation (PVS). We depict the online-PVS performance (in \\( J \\& F \\)) averaged on four benchmark datasets [1, 18, 36, 41] (vertical axis) with respect to model training cost (horizontal axis). The size of circles indicates model parameters (in M). SAM 2.1 [36] is a strong PVS model but needs huge training cost (i.e., 256 A100 GPUs for 108 hours). We propose to cultivate a PVS model via image-to-video model upgrading. Building on the pre-trained SAM [21, 40], incorporating existing video object segmentation models (e.g., XMem++ [1]) or temporal association modules (e.g., temporal adapter [36]) significantly reduces training costs compared to SAM 2.1 [36]. Furthermore, our designed SAM-I2V helps achieve over \\( 90\\% \\) of SAM 2.1's performance while using only \\( 0.2\\% \\) of its training cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.809 + ], + "angle": 0, + "content": "(SAM) [21] has emerged as a prominent foundation model for general-purpose promptable image segmentation, enabling users to segment objects/regions in images based on prompts such as points, boxes, or masks. While SAM demonstrates impressive capabilities on static images, extending these abilities to video data introduces significant challenges due to the temporal dynamics and consistency required in video segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, SAM 2 [36] has been developed to address promptable video segmentation (PVS), aiming to extend SAM's image segmentation capabilities to videos. SAM 2 is trained from scratch on a massive collection of images and videos to learn complex spatiotemporal associations. However, it demands exorbitant computational resources," + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.245, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "requiring 256 A100 GPUs over 108 hours for training. Such high computational costs create substantial barriers to exploring alternative architectures and advancing research in PVS, limiting accessibility for researchers and practitioners." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.482, + 0.332 + ], + "angle": 0, + "content": "This raises a critical question: Can we achieve promptable video segmentation through a resource-efficient approach? If so, it would open opportunities for developing diverse PVS foundation models, benefiting a wide range of downstream applications. Addressing this challenge, we propose SAM-I2V, an efficient method to upgrade the pretrained image-based SAM [21] model to support promptable video segmentation with affordable training cost (Figure 1). Our approach strategically leverages SAM's strong promptable image segmentation capabilities and develops image-to-video upgrading techniques to extend functionality from images to videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.333, + 0.483, + 0.604 + ], + "angle": 0, + "content": "To bridge the gap between promptable image and video segmentation, our SAM-I2V includes three key components: a temporal feature integrator TFI, a memory selective associator MSA, and a memory prompt generator MPG. (i) The TFI enhances SAM's image encoder by integrating temporal information, enabling the extraction of spatiotemporal features essential for understanding dynamic scenes; (ii) The MSA retrieves relevant segmentation information from preceding video frames through a new memory filtering mechanism, selectively associating historical information to guide the current frame's segmentation; (iii) The MPG incorporates memory as an additional prompt, utilizing object memory to ensure precise and temporally consistent mask propagation across frames. Comprehensive experiments validate that our method significantly reduces training complexity and resource requirements, achieving over \\(90\\%\\) of SAM 2's performance with only \\(0.2\\%\\) of its training cost. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.605, + 0.48, + 0.664 + ], + "angle": 0, + "content": "1. We present SAM-I2V, the first attempt to upgrade a pre-trained image-based SAM model to support promptable video segmentation, enabling PVS capabilities with substantially reduced training costs." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.666, + 0.48, + 0.709 + ], + "angle": 0, + "content": "2. We design an upgrade method that extends SAM's static image encoder into a video perception module, enabling spatiotemporal feature extraction for PVS." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.712, + 0.48, + 0.769 + ], + "angle": 0, + "content": "3. We propose a memory filtering strategy that selects the most relevant past frames to facilitate more effective utilization of historical information for memory-guided video segmentation." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.772, + 0.48, + 0.815 + ], + "angle": 0, + "content": "4. We introduce a memory-as-prompt mechanism that leverages object memory to ensure temporally consistent mask propagation in dynamic scenes." + }, + { + "type": "list", + "bbox": [ + 0.103, + 0.605, + 0.48, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.24, + 0.845 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Promptable Visual Segmentation. The Segment Anything Model (SAM) [21] has set a significant milestone in promptable segmentation. SAM is designed to gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.56 + ], + "angle": 0, + "content": "erate a valid segmentation mask for an object of interest in an image when provided with input prompts such as bounding boxes or reference points. Trained on the extensive SA-1B dataset, SAM facilitates zero-shot segmentation with its versatile prompting mechanism, making it highly adaptable for diverse downstream tasks. This foundational work has spurred further innovations aimed at enhancing SAM's quality and efficiency. For instance, models like HQ-SAM [20] have focused on improving output precision through high-quality token integration and fine-tuning with detailed masks. Efficiency-oriented models such as EfficientSAM [47], MobileSAM [53, 54], FastSAM [56], TinySAM [40], and SlimSAM [6] have broadened the scope of SAM's applications, enabling its deployment in more resource-constrained and real-world scenarios, including medical imaging [25, 27, 43] and remote sensing [3, 37, 45]. Building upon the success of SAM [21], SAM 2 [36] represents an evolution toward promptable video segmentation, addressing the complexities of propagating segmentation across multiple frames. This allows for interaction at any frame of a video, accepting prompts such as positive/negative clicks, bounding boxes, or masks to define or refine object segmentation. Unlike its predecessor, SAM 2 not only responds to initial prompts on single frames but also propagates these segmentation cues throughout the entire video, maintaining temporal consistency and adaptability across frames. Evaluated through interactive online and offline settings, SAM 2's integration of promptable video interaction ensures comprehensive and dynamic segmentation, pushing the boundaries of practical real-world applicability [5, 15, 23, 39, 46, 58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.905, + 0.852 + ], + "angle": 0, + "content": "Interactive Video Object Segmentation. Interactive video object segmentation (IVOS) is an essential task for enabling real-time object tracking and segmentation in videos through user input, such as clicks or bounding boxes. Recent studies [8, 14, 17] have aimed to develop modular approaches that allow segmenting an object in a single frame and then propagating this segmentation across subsequent frames to maintain consistency. Building upon this, many IVOS methods leverage interactive mechanisms to guide segmentation over multiple frames. While SAM has been used for video object segmentation combined with video trackers based on masks [10, 12, 49] or points [35], these approaches face certain limitations: trackers may not be effective for all objects, SAM may perform suboptimally on certain video frames, and there is no interactive mechanism for correcting model errors aside from re-annotating erroneous frames and restarting the tracking process from that point. This highlights expanded functionality of SAM 2 that provides a flexible solution to enable interactive refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Semi-Supervised Video Object Segmentation. Semi-supervised video object segmentation (Semi-VOS) typically begins with a provided mask for the first frame," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.422 + ], + "angle": 0, + "content": "which acts as a reference for tracking the object through the video [32]. The term \"semi-supervised\" refers to using the initial mask as supervision, with no further manual input for subsequent frames. This task has become vital for various applications requiring automatic tracking, such as video editing, robotics, and automatic background removal. Early Semi-VOS techniques involved fine-tuning models on the first frame [2, 26, 38] or incorporating offline-trained models that predicted object masks for the entire video [28, 50, 52]. More advanced strategies integrated RNNs [48] and cross-attention [1, 7, 9, 11] to extend segmentation from the first frame throughout the sequence, improving object tracking accuracy. Recent methods [44, 55] have enhanced a single vision transformers to jointly handle the current frame alongside all prior frames and their corresponding predictions, simplifying the architecture but incuring a significant inference cost. Semi-VOS process is analogous to a specialized case of promptable video segmentation, where providing an high-quality mask prompt in the first frame is crucial but practically challenging and time-consuming, limiting the practical application of Semi-VOS in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.424, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Image-to-Video Transfer Learning. In recent years, the importance of image-to-video transfer learning has grown due to its significant potential for reducing training costs and enhancing adaptability in video-based tasks. To facilitate efficient image-to-video transfer, various methods have been introduced, including spatio-temporal adapters [29, 51], dual-path adaptation [30], disentangled spatial and temporal learning [33], and learnable prompts [19, 22]. These methods incorporate lightweight, trainable components that leverage pre-trained image foundation models such as ViT [16] and CLIP [34] while integrating spatiotemporal dynamics, striking a balance between computational efficiency and performance to enhance video applications. These methods, although centered around imaged-to-video transfer for video recognition and understanding tasks, offer insights into our efficient cultivation of a promtable video segmentation model (i.e., SAM 2 [36]) via upgrading the pre-trained image-based SAM model [21] to a video one. These approaches provide a solid foundation for spatio-temporal information learning in our image-to-video upgradation but lack direct applicability for addressing segmentation-specific challenge, i.e., ensuring precise and temporally consistent mask propagation in scene dynamics. This makes our focus on SAM to SAM 2 upgradation a distinct and challenging endeavor." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.226, + 0.833 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While SAM 2 demonstrates strong capabilities, its exorbitant training demands—requiring 256 A100 GPUs over 108 hours—present a significant barrier to exploring alternative architectures and advancing research in PVS. To overcome" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "this challenge, we aim to develop an efficient method for cultivating a PVS model with affordable training costs. Our strategy leverages the pre-trained SAM model, which already performs well on single frames, and upgrades it to SAM 2. We design three components to achieve this transformation: a Temporal Feature Integrator (Section 3.2), a Memory Associator (3.3), and a Memory Prompt Generator (3.4), which collectively upgrade image-based SAM into a promptable video segmentation model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.248, + 0.625, + 0.263 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Our SAM-I2V enables image-based SAM model can process video input for promtable video segmentation. As illustrated in Figure 2, the video frames are passed through a backbone feature extractor, which processes the current frame using a series of image encoder layers. These encoder outputs are enriched with temporal context through a temporal feature integrator, enabling the capture of dynamic information across consecutive frames. The enhanced features are subsequently input into the memory selective associator and memory prompt generator modules. These components manage memory states and generate target object prompts informed by past frames data, ensuring continuity and contextual awareness throughout the video. Optional user-provided prompts (e.g., masks, points, or bounding boxes) are processed by the prompt encoder to guide the segmentation with user-specific inputs. Finally, the integrated information is directed into the mask decoder, which produces the final segmentation mask for the current frame." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.773, + 0.582 + ], + "angle": 0, + "content": "3.2. Temporal Feature Integrator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.905, + 0.697 + ], + "angle": 0, + "content": "To enhance SAM's image encoder, which is limited to extracting only spatial features, into a video perception module, we introduce a Temporal Feature Integrator (TFI). The TFI comprises a temporal branch and an integration branch that operates synergistically to iteratively extract and integrate temporal features, enabling a comprehensive spatiotemporal feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.701, + 0.905, + 0.838 + ], + "angle": 0, + "content": "As illustrated in Figure 3, given the input video \\(X \\in \\mathbb{R}^{c \\times n \\times h \\times w}\\), where \\(c\\), \\(n\\), \\(h\\) and \\(w\\) are the frame channel, number, height and width, respectively, SAM's image encoder processes the current frame \\(X_{t}\\) via a 2D stem and \\(L\\) spatial blocks, generating multi-stage spatial features \\(S_{l} \\in \\mathbb{R}^{c_{l} \\times h_{l} \\times w_{l}}\\), where \\(c_{l}\\), \\(h_{l}\\) and \\(w_{l}\\) denote the features channel, height and width at stage \\(l \\in \\{1, \\dots, L\\}\\), respectively. The extracted features are fed into our TFI to obtain spatiotemporal features \\(I_{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Specifically, the temporal branch of TFI processes the video input \\(X\\) through a 3D stem followed by \\(L\\) temporal blocks to capture temporal contexts. In \\(l\\)-th temporal block, we employ 3D convolutions \\(\\mathcal{C}_{3d}\\) to progressively" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3419" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.216, + 0.908, + 0.302 + ], + "angle": 0, + "content": "Figure 2. Overview of SAM-I2V. The input video is fed into a backbone feature extractor, which passes the current frame through a series of image encoder layers. The output of these layers is enhanced with temporal information by a temporal feature integrator to capture dynamic context across frames. The extracted features are then fed into the memory associator and the memory prompt generator, which manage memory states and generate target object prompts based on past frame information. The prompt encoder processes optional user input (e.g., masks, points, or bounding boxes) to guide the segmentation. Finally, the processed information flows into the mask decoder, which outputs the segmentation mask for current frame, enabling user-intention-followed or memory conditioned promptable video segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.423, + 0.332 + ], + "angle": 0, + "content": "model temporal correspondences between frames:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.34, + 0.272, + 0.36 + ], + "angle": 0, + "content": "\\[\nT _ {l} ^ {\\prime} = T ^ {(3)},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.363, + 0.39, + 0.384 + ], + "angle": 0, + "content": "\\[\nT ^ {(i)} = T ^ {(i - 1)} + \\mathcal {C} _ {3 d} ^ {(i)} \\big (T ^ {(i - 1)} \\big),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.455 + ], + "angle": 0, + "content": "where \\( T^{(0)} = T_{l - 1} \\) is the input to the \\( l \\)-th temporal block. To further enrich temporal features \\( T_{l}^{\\prime} \\) with spatial details, we fuse it with intermediate features \\( I_{l}^{\\prime} \\) from TFI's integration branch:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.455, + 0.347, + 0.473 + ], + "angle": 0, + "content": "\\[\nT _ {l} = \\mathcal {S T} (I _ {l} ^ {\\prime}, T _ {l} ^ {\\prime}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathcal{ST}\\) denotes the spatial-to-temporal fusion operation that slices out current-frame temporal features from \\(T_{l}^{\\prime}\\) (i.e., the last slice in temporal dimension), concatenates it with current-frame spatial features \\(I_{l}^{\\prime}\\) along the channel dimension, applies a \\(1\\times 1\\) convolution to integrate concatenated features, and uses the resulting output to overwrite original current-frame temporal features in \\(T_{l}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.584, + 0.484, + 0.615 + ], + "angle": 0, + "content": "Then, TFI's integration branch combines spatial features \\( S_{l} \\) and temporal features \\( T_{l}^{\\prime} \\) to get spatiotemporal features:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.626, + 0.319, + 0.642 + ], + "angle": 0, + "content": "\\[\nI _ {l} ^ {\\prime} = I _ {l - 1} + S _ {l},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.646, + 0.364, + 0.664 + ], + "angle": 0, + "content": "\\[\nI _ {l} = \\mathcal {C} _ {2 d} \\big (\\mathcal {T S} (I _ {l} ^ {\\prime}, T _ {l} ^ {\\prime}) \\big),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.81 + ], + "angle": 0, + "content": "where \\( I_{l-1} \\) is the input to the \\( l \\)-th integration block. The temporal-to-spatial fusion operation \\( \\mathcal{T}S \\) merges current-frame temporal features in \\( T_l' \\) with \\( I_l' \\), producing temporal-enriched spatial features. These features then pass through a 2D convolution \\( C_{2d} \\), yielding the enhanced spatiotemporal output \\( I_l \\). Finally, the temporal and integration branches of TFI iteratively refine the features over multiple stages, enabling the model to progressively capture rich spatiotemporal context for the subsequent segmentation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.351, + 0.835 + ], + "angle": 0, + "content": "3.3. Memory Selective Associator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While image-based SAM [21] segments the target object based on user prompts, in the PVS task, segmentations from preceding video frames can serve as target indicators to guide the segmentation of the target in the current frame." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.318, + 0.898, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.478, + 0.908, + 0.521 + ], + "angle": 0, + "content": "Figure 3. Structural details of temporal feature integrator (TFI). It consists of a temporal branch and an integration branch, enhancing SAM's image encoder to function as a video perceptron." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.597 + ], + "angle": 0, + "content": "This motivates our Memory Selective Associator (MSA), designed to retrieve relevant and beneficial segmentation information from preceding video frames, leveraging this memory to enhance the current frame's features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.734 + ], + "angle": 0, + "content": "MSA first utilizes convolution to fuse the features from the previous frames \\(\\{F_{t - p}\\}_{p = 1}^{P}\\) and their corresponding segmentation masks \\(\\{M_{t - p}\\}_{p = 1}^{P}\\). These fused features are then passed to the memory selector, which chooses the most relevant past frame features. The selected features are used in the memory attention module, where cross-attention between the current frame features \\(I_{t}\\) and the selected previous frame features is applied, producing the memory-enhanced features \\(I_{t}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The memory encoder and attention mechanism in our MSA follow a similar structure as in SAM 2 [36]. Unlike SAM 2 [36], which selects the previous \\(z\\) consecutive frames leading up to the current frame, our memory selector chooses \\(z\\) frames based on their relevance, as determined by similarity to the current frame. Specifically, it first chooses the prompt frame based on the user input, and then divides the previous frames into \\(X\\) short-term (local) frames and \\(Y\\) long-term (global) frames. The similarity between the current frame features \\(I_{t} \\in \\mathbb{R}^{N}\\) and \\(X/Y\\) previous frame features \\(F \\in \\mathbb{R}^{X/Y \\times N}\\) in each set is computed to identify" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.907, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Figure 4. Illustration of our memory selective associator (MSA). It consists of a memory encoder, a memory selector, and memory attention." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.277, + 0.343, + 0.292 + ], + "angle": 0, + "content": "the most relevant local/global frames:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.304, + 0.454, + 0.323 + ], + "angle": 0, + "content": "\\[\nS ^ {l o c} = F ^ {l o c} \\cdot I _ {t} \\in \\mathbb {R} ^ {X}, S ^ {g l o} = F ^ {g l o} \\cdot I _ {t} \\in \\mathbb {R} ^ {Y},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.385 + ], + "angle": 0, + "content": "where \\(N = c_{L} \\times h_{L} \\times w_{L}\\). Next, we normalize the similarity scores to obtain the probability distribution \\(D^{loc}\\) and \\(D^{glo}\\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.395, + 0.466, + 0.44 + ], + "angle": 0, + "content": "\\[\nD ^ {l o c} = \\frac {e ^ {S ^ {l o c}}}{\\sum_ {j} e ^ {S _ {j} ^ {l o c}}} \\in \\mathbb {R} ^ {X}, D ^ {g l o} = \\frac {e ^ {S ^ {g l o}}}{\\sum_ {j} e ^ {S _ {j} ^ {g l o}}} \\in \\mathbb {R} ^ {Y}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.48 + ], + "angle": 0, + "content": "We then sample \\(x\\) local frames and \\(y = z - x\\) global frames based on \\(D^{loc}\\) and \\(D^{glo}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.494, + 0.462, + 0.514 + ], + "angle": 0, + "content": "\\[\ni _ {k} ^ {l o c} \\sim P (i _ {k} ^ {l o c} = i) = D ^ {l o c} (i), \\quad k \\in \\{1, 2, \\ldots , X \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.529, + 0.462, + 0.55 + ], + "angle": 0, + "content": "\\[\ni _ {k} ^ {g l o} \\sim P (i _ {k} ^ {g l o} = i) = D ^ {g l o} (i), \\quad k \\in \\{1, 2, \\ldots , Y \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.652 + ], + "angle": 0, + "content": "where \\( i_k^{loc} \\) and \\( i_k^{glo} \\) denote the indices of the selected local and global frames, respectively, and \\( D^{loc}(i) \\) and \\( D^{glo}(i) \\) are the corresponding probabilities for each frame. The selected frames are subsequently used to enhance the segmentation of the current frame through memory attention, resulting in memory-associated features \\( I_t' \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.342, + 0.681 + ], + "angle": 0, + "content": "3.4. Memory Prompt Generator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.794 + ], + "angle": 0, + "content": "SAM's mask decoder segments images using image features and user prompts to achieve intention-followed results. Motivated by its prompt integration capability, we propose incorporating memory as an additional prompt. This enables the model to leverage segmentation information from previous frames, enhancing temporal consistency across the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Figure 5 illustrates the proposed Memory Prompt Generator (MPG). We introduce a set of learnable memory prompt tokens \\( G \\in \\mathbb{R}^{g \\times d} \\), where \\( g \\) is the number of memory prompts and \\( d \\) is the embedding dimension. These tokens interact with the flattened memory features \\( F_{\\mathrm{mem}} \\in \\mathbb{R}^{(mh_Lw_L) \\times c_L} \\), where \\( m \\) is the number of memory frames, guided by the corresponding binarized segmentation masks" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.28, + 0.898, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.379, + 0.896, + 0.394 + ], + "angle": 0, + "content": "Figure 5. Illustration of the memory prompt generator (MPG)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.412, + 0.847, + 0.428 + ], + "angle": 0, + "content": "\\(M_{\\mathrm{mem}} \\in \\mathbb{R}^{(mh_Lw_L) \\times 1}\\) via masked cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.439, + 0.907, + 0.457 + ], + "angle": 0, + "content": "\\[\nG ^ {\\prime} = \\operatorname {S o f t m a x} (\\mathcal {M} + Q K ^ {\\top}) V + G, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.459, + 0.907, + 0.476 + ], + "angle": 0, + "content": "\\[\nQ = \\psi (G), \\quad K = \\psi \\left(F _ {\\mathrm {m e m}}\\right), \\quad V = \\psi \\left(F _ {\\mathrm {m e m}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.48, + 0.906, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (x, y) = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} M _ {\\mathrm {m e m}} (x, y) = 1, \\\\ - \\infty , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.591 + ], + "angle": 0, + "content": "where \\(\\psi\\) is the learnable linear transformation. The updated memory prompt tokens \\(G^{\\prime}\\), having captured foreground context from memories, are then refined through self-attention to incorporate intra-token dependencies:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.603, + 0.779, + 0.62 + ], + "angle": 0, + "content": "\\[\nG ^ {\\prime \\prime} = \\operatorname {S e l f A t t n} \\left(G ^ {\\prime}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.632, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Finally, the refined memory prompt embeddings \\( G'' \\) are passed through a multi-layer perceptron (MLP) to produce the final updated memory prompt tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.689, + 0.771, + 0.706 + ], + "angle": 0, + "content": "\\[\nG ^ {\\prime \\prime \\prime} = \\operatorname {M L P} \\left(G ^ {\\prime \\prime}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Such an iterative attention-based token generation enables the MPG to selectively capture target object information from memories, providing guidance for the mask decoder." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.778, + 0.646, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.704, + 0.82 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Training Datasets. We train our SAM-I2V jointly on 10,000 SA-1B [21] images and 50,583 SAV [36] videos. The SA-1B dataset [21] is a comprehensive collection consisting of over 1.1 billion high-quality segmentation masks from 11 million licensed and privacy-preserving images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3421" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "This dataset was assembled using a multi-phase data engine where initial stages involved human annotation aided by interactive tools, and subsequent phases transitioned to fully automated mask generation through the SAM model. The rich variety of mask annotations allows for robust training that supports zero-shot generalization and diverse segmentation tasks, proving effective in many downstream applications. The Segment Anything Video (SA-V) dataset [36] complements the image-focused SA-1B by providing extensive coverage for video segmentation. Its training set includes 50,583 videos and 642,036 masklets. SA-V is tailored for video scenarios involving complex object motions, occlusions, and reappearance, making it invaluable for advancing video-based models like SAM 2 [36]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.485, + 0.791 + ], + "angle": 0, + "content": "Training Configurations. Following [36], to simulate an interactive training setting, we sample 8-frame sequences, designating the first frame as the prompt frame and randomly selecting up to 1 additional frame from the remaining seven frames for corrective clicks. The initial prompt for the first frame can be the ground-truth mask (50% probability), a positive click from the ground-truth mask (25%), or a bounding box input (25%). The corrective clicks are sampled from the error region between the model prediction and the ground-truth mask (90%) or directly from the ground-truth mask (10%). Our SAM-I2V model is implemented using PyTorch [31] and trained for 10 epochs on 8 RTX A5000 GPUs (24GB memory). We employ the AdamW optimizer [24] and apply layer-wise decay [13] specifically to the image encoder. The initial learning rates are set to \\(3 \\times 10^{-4}\\) for the image encoder and \\(6 \\times 10^{-5}\\) for the other components, respectively. These learning rates are subsequently decayed using a cosine scheduler. We supervise the model's predictions using a linear combination of mask prediction loss for the output mask, meanabsolute-error (MAE) loss for the IoU prediction, and cross-entropy loss for object prediction with a ratio of 20:1:1, respectively. The mask prediction loss \\(\\mathcal{L}_m\\) is defined as the combination of the weighted binary cross-entropy (BCE) loss \\(\\ell_{wbce}\\) [42] and the weighted IoU loss \\(\\ell_{wiou}\\) [42], i.e., \\(\\mathcal{L}_m = \\ell_{wbce} + \\ell_{wiou}\\), to ensure the model pays more attention to potential distraction regions such as object boundaries, elongated areas, or holes. Notably, all SAM-I2V modules except the TinySAM [40] backbone are randomly initialized and trained from scratch without leveraging any weights from SAM 2 [36], ensuring complete independence from SAM 2 [36] in the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Evaluation Settings. To comprehensively assess the capabilities of our SAM-I2V, we perform comparisons on four benchmark datasets: ESD [18] is a high-quality 3D spatial and temporal dataset for object segmentation in an indoor cluttered environment; PUMaVOS [1] contains videos with segments around object parts such as a person's cheek; LV-VIS [41] contains videos from a diverse set of open-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.471 + ], + "angle": 0, + "content": "vocabulary object categories; and SAV-Test [36] is the test set of SA-V dataset, consist of 150 videos and 278 masklets. For these evaluations, we report the standard \\( J \\& F \\) metric [32] which combines region similarity (Jaccard index) and boundary accuracy (F-measure). We conducted evaluation on the promptable video segmentation. In this task, the model works in an interactive manner, where the point prompts are used to identify the target of interest or to refine the current segmentation result. Following [36], two settings are included in the comparison, namely offline evaluation (PVS-offline) and online evaluation (PVS-online). In both settings, the target is firstly initialized via \\( N_{\\mathrm{click}} \\) point prompts. As for PVS-offline, the model segments the target throughout the entire video in \\( N_{\\mathrm{pass}} \\) passes, where the interactive frame is selected based on the largest model error (i.e., with lowest IoU between the prediction and ground truth masks) and \\( N_{\\mathrm{click}} \\) points are added on such frame to refine the mask in each pass. All the prompts are used to guide the segmentation process in the next pass. PVS-online follows one pass evaluation protocol, where the segmentation process is paused and refined when the result is not reliable (IoU \\( < 0.75 \\) with ground-truth). Also, \\( N_{\\mathrm{click}} \\) points are added on the refined frame to generate corrected mask, which is used for subsequent frames. The refine process is repeated \\( N_{\\mathrm{frame}} \\) times." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.804, + 0.499 + ], + "angle": 0, + "content": "4.2. Comparison to Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.656 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed SAM-I2V, we compared it against seven state-of-the-art methods on both online and offline PVS tasks under different click settings, as reported in Table 1 and 2. Besides SAM 2.1 [36], we also include upgradation methods that combine SAM (both base [21] and tiny [40] variants) with existing state-of-the-art video object segmentation models, specifically Xmem++ [1] and Cutie [11], as well as approaches that upgrade SAM by incorporating temporal adaptation (TA) [36] and spatial adaptation (SA) [4] modules." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Performance Comparison with SAM 2.1. While SAM 2.1 [36] serves as a strong baseline with an overall accuracy of 71.9 (online) and 73.3 (offline), it requires substantial computational resources (2.2 million \\(G \\cdot\\) Hour). In contrast, our SAM-I2V achieves its \\(91\\%\\) (i.e., 65.3, online) and \\(90\\%\\) (i.e., 66.0, offline) overall scores, with its \\(0.2\\%\\) (i.e., 4.6k \\(G \\cdot\\) Hour) training cost and its \\(49\\%\\) (i.e., 18.9m vs. 38.9m) model parameters." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Effectiveness over Existing SAM Upgradation Methods. Compared to SAM [21] + XMem++ [1] (157.0M parameters) and TinySAM [40] + XMem++ (72.3M parameters), our SAM-I2V offers a substantial reduction in model size (18.9M) while delivering comparable or superior performance. For example, under the online setting, SAM-I2V helps achieve an overall score of 65.3, outperforming SAM + XMem++ (60.4) and SAM + Cutie [11] (61.7)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3422" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.091, + 0.907, + 0.285 + ], + "angle": 0, + "content": "
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.645.859.366.386.866.879.575.288.870.182.877.279.760.974.972.971.9
SAM [21] + XMem++ [1]157.0-78.349.368.042.478.846.273.041.979.345.478.443.378.847.073.142.560.4
SAM [21] + Cutie [11]129.8-78.438.369.451.279.044.373.951.879.844.676.952.979.142.473.452.061.7
TinySAM [40] + XMem++ [1]72.3-79.355.268.547.382.959.874.948.784.260.476.951.682.158.573.449.265.8
TinySAM [40] + Cutie [11]45.1-80.653.071.158.483.352.676.259.984.456.278.863.182.853.975.460.568.1
TinySAM [40] + TA [36]17.84.6k68.728.856.242.980.245.172.349.182.149.876.550.977.041.268.447.658.6
TinySAM [40] + SA [4] + TA [36]18.74.6k54.834.652.844.179.352.772.950.681.959.476.851.972.048.967.548.959.3
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.039.153.850.584.760.873.458.286.465.978.359.581.455.368.556.165.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.907, + 0.339 + ], + "angle": 0, + "content": "Table 1. Comparisons of the online promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.353, + 0.907, + 0.51 + ], + "angle": 0, + "content": "
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.946.261.366.887.767.383.976.290.170.787.578.180.661.477.673.773.3
SAM [21] + XMem++ [1]157.0-49.536.850.141.164.049.353.744.165.354.755.244.959.646.953.043.450.7
TinySAM [40] + XMem++ [1]72.3-47.746.047.549.173.655.452.955.575.962.553.758.265.754.651.454.356.5
TinySAM [40] + TA [36]17.84.6k68.929.558.345.381.149.376.352.482.955.681.154.977.644.871.950.961.3
TinySAM [40] + SA [4] + TA [36]18.74.6k57.037.054.944.880.956.676.652.583.462.181.253.673.851.970.950.361.7
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.437.855.748.984.661.277.858.586.465.183.059.881.554.772.255.766.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.522, + 0.907, + 0.579 + ], + "angle": 0, + "content": "Table 2. Comparisons of the offline promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance. Cutie [11] is excluded from the comparison as it only supports single-frame mask inputs, rendering it unsuitable for offline PVS settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Robust Performance Across Interaction Levels. The online and offline results presented in Table 1 and 2 reflect that the performance of all methods generally improves as the number of interactions increases. Our SAM-I2V consistently delivers superiority across 1-click, 3-click, and 5-click settings, underscoring its robustness to varying levels of user interaction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Robust Online and Offline Performance. SAM-I2V's overall online performance in Table 1 is slightly lower than the offline performance in Table 2. This discrepancy arises because, unlike in the offline evaluation, in the online setting the new prompts affect only the frames following the current paused frame and not the preceding frames. Compared to other methods, our approach maintains an advantage in both online and offline settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The above results confirm that our SAM-I2V not only matches but often surpasses existing upgrading methods in performance with reduced computational costs, making SAM-I2V a practical choice for PVS tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.667, + 0.62 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.672 + ], + "angle": 0, + "content": "We conduct an extensive ablation study to validate the extensibility of SAM-I2V and the effectiveness of its each key component. Table 3, 4, 5, 6 summarize our findings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Applicability of SAM-I2V to Upgrade Different SAM Models. In the first ablation study, we investigate the performance of our SAM-I2V to upgrade different SAM models. From the results presented in Table 3, it is evident that our SAM-I2V effectively upgrades different SAM variants to a PVS model, demonstrate that SAM-I2V is a robust and versatile image-to-video promptable segmentation upgradation approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the Temporal Feature Integrator. As shown in Table 4, our proposed TFI module outperforms the baseline (TinySAM [40] + Temporal Associator [36]) and other adaptation methods. While previous approaches like ST-Adapter [29] and Parallel Adapter [4] perform adaptation within each attention block, our TFI is specifically designed with temporal branch and integration branch to iter" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3423" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.09, + 0.478, + 0.164 + ], + "angle": 0, + "content": "
MethodsES[18]PU[1]LV[41]SV[36]OA
TinySAM [40] + TA [36]80.245.172.349.161.7
TinySAM [40] + SAM-I2V84.760.873.458.269.3
EdgeSAM [57] + TA [36]78.341.469.644.658.5
EdgeSAM [57] + SAM-I2V83.251.971.654.465.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.175, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Table 3. Ablation on the applicability of SAM-I2V to upgrade different SAM models for PVS task under the online, 3-click setting." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.218, + 0.478, + 0.294 + ], + "angle": 0, + "content": "
MethodsES[18]PU[1]LV[41]SV[36]OA
Base80.245.172.349.161.7
Base + ST-Adapter [29]79.354.173.350.164.2
Base + Parallel Adapter [4]79.352.772.950.663.9
Base + TFI81.358.172.452.566.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.305, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Table 4. Ablation on the effectiveness of the temporal feature integrator TFI for online PVS task under the 3-click setting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.356, + 0.483, + 0.401 + ], + "angle": 0, + "content": "atively extract and integrate spatial-temporal features. This allows our model to capture richer temporal dependencies, leading to superior performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Effectiveness of the Memory Selective Associator. Relying on the previous \\( z = 6 \\) consecutive frames for segmentation is equivalent to setting \\( x \\) and \\( y \\) in MSA to 0 and 6, respectively. By varying the number of selected global \\( (x) \\) and local \\( (y) \\) frames in our MSA, we observe that the segmentation performance improves when the MSA selectively retrieves the most relevant frames based on their similarity to the current frame. Specifically, the configuration with \\( x = 3 \\) global frames and \\( y = 3 \\) local frames performs better, highlighting the importance of incorporating both short-term and long-term memory. In addition, the ability of MSA to select frames based on relevance rather than uniform sampling allows the model to focus on the most relevant past frames, leading to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Effectiveness of the Memory Prompt Generator. From Table 6, we can observe that, (i) introducing MPG \\((w/\\) MPG-1/3/5) consistently improves performance over the one without MPG, demonstrating the effectiveness of the memory-as-prompt strategy in enhancing segmentation accuracy; (ii) varying the number of memory prompt tokens among 1, 3, and 5 yields comparable results, indicating that the MPG is robust to the choice of token quantity within this range; and (iii) removing the Masked Cross-Attention (MCA) mechanism (i.e., \\( w/\\) MPG-3 (\\( w/o \\) MCA)) leads to a performance drop, highlighting the importance of MCA in focusing on relevant foreground information, eliminating background interference, and thus enhancing the quality of memory prompt generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.349, + 0.849 + ], + "angle": 0, + "content": "4.4. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "One limitation of our approach is that its accuracy does not well match that of SAM 2.1 [36]. Note that the released SAM 2.1 model was trained not only on the SAM-1B" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.091, + 0.902, + 0.178 + ], + "angle": 0, + "content": "
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MSA (i.e., x=0, y=6)84.054.473.256.967.1
w/ MSA (x=1, y=5)84.754.873.457.967.7
w/ MSA (x=2, y=4)84.556.373.456.967.8
w/ MSA (x=3, y=3)84.760.873.458.269.3
w/ MSA (x=3, y=3) w/ UFS84.156.073.056.567.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.188, + 0.907, + 0.247 + ], + "angle": 0, + "content": "Table 5. Ablation on the effectiveness of the memory selective associator MSA for online PVS task under the 3-click setting. \"x\" and \"y\" denote the number of selected global and local frames, respectively. \"UFS\" indicates the uniform frame sampling strategy." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.259, + 0.9, + 0.348 + ], + "angle": 0, + "content": "
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MPG82.157.772.153.266.3
w/ MPG-184.058.473.055.667.8
w/ MPG-384.760.873.458.269.3
w/ MPG-584.558.673.657.168.5
w/ MPG-3 (w/o MCA)84.958.172.855.167.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.358, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Table 6. Ablation on the effectiveness of the memory prompt generator MPG for online PVS task under the 3-click setting. \"MPG-g\" denotes the MPG with \\( g \\) learnable memory prompt tokens. \"MCA\" indicates the masked cross attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.577 + ], + "angle": 0, + "content": "and SAV datasets but also on 62.9k proprietary, non-public videos, which our method was unable to utilize. Despite this, our method demonstrates superior performance compared to baseline approaches and creates targeted model design opportunities for downstream applications in resource-constrained academic scenarios. Future work will focus on enhancing our model by incorporating additional publicly available datasets and exploring advanced training techniques to bridge the accuracy gap with SAM 2.1 [36]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.856 + ], + "angle": 0, + "content": "In this work, we introduced SAM-I2V, a training-efficient framework that upgrades the pre-trained Segment Anything Model (SAM) to support promptable video segmentation (PVS). SAM-I2V consists of three key innovations: an image-to-video feature extraction upgrader that extends SAM's static image encoder for spatiotemporal video perception; a memory filtering strategy that selects the most relevant past frames to enhance the utilization of historical information; and a memory-as-prompt mechanism that leverages object memory for temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that SAM-I2V achieved over \\(90\\%\\) of SAM 2's performance at only \\(0.2\\%\\) of its training cost. This remarkable training efficiency significantly lowers the barriers for developing diverse PVS models, facilitating rapid innovation and broader practical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3424" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Maksym Bekuzarov, Ariana Bermudez, Joon-Young Lee, and Hao Li. Xmem++: Production-level video segmentation from few annotated frames. In ICCV, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.174, + 0.484, + 0.215 + ], + "angle": 0, + "content": "[2] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[3] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.29, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[5] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv:2408.04579, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[6] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. \\(0.1\\%\\) data makes segment anything slim. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[8] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.596, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.639, + 0.482, + 0.68 + ], + "angle": 0, + "content": "[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In CVPR, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.684, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[12] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv:2305.06558, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.728, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[13] K Clark. Electra: Pre-training text encoders as discriminators rather than generators. arXiv:2003.10555, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[14] Thanos Delatolas, Vicky Kalogeiton, and Dim P Papadopoulos. Learning the what and how of annotation in video object segmentation. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[15] Haoyu Dong, Hanxue Gu, Yaqian Chen, Jichen Yang, and Maciej A Mazurowski. Segment anything model 2: an application to 2d and 3d medical images. arXiv:2408.00756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[17] Yuk Heo, Yeong Jun Koh, and Chang-Su Kim. Interactive video object segmentation using global and local transfer modules. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[18] Xiaoqian Huang, Kachole Sanket, Abdulla Ayyad, Fariborz Baghaei Naeini, Dimitrios Makris, and Yahya Zweiri. A neuromorphic dataset for object segmentation in indoor cluttered environment. arXiv:2302.06301, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[19] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[20] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[23] Haofeng Liu, Erli Zhang, Junde Wu, Mingxuan Hong, and Yueming Jin. Surgical sam 2: Real-time segment anything in surgical video by efficient frame pruning. arXiv:2408.07931, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.904, + 0.481 + ], + "angle": 0, + "content": "[24] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[25] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.528, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[26] K-K Maninis, Sergi Caelles, Yuhua Chen, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. Video object segmentation without temporal information. IEEE TPAMI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.586, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[27] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[28] Seoung Wug Oh, Joon-Young Lee, Kalyan Sunkavalli, and Seon Joo Kim. Fast video object segmentation by reference-guided mask propagation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[29] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. NeurIPS, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[30] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dualpath adaptation from image to video transformers. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3425" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[33] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[35] Frano Rajic, Lei Ke, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu. Segment anything meets point tracking. arXiv:2307.01197, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[37] Simiao Ren, Francesco Luzi, Saad Lahrichi, Kaleb Kassaw, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Segment anything, from space? In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[38] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[39] Chuyun Shen, Wenhao Li, Yuhang Shi, and Xiangfeng Wang. Interactive 3d medical image segmentation with sam 2. arXiv:2408.02635, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[40] Han Shu, Wenshuo Li, Yehui Tang, Yiman Zhang, Yihao Chen, Houqiang Li, Yunhe Wang, and Xinghao Chen. Tinysam: Pushing the envelope for efficient segment anything model. arXiv:2312.13789, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[42] Jun Wei, Shuhui Wang, and Qingming Huang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[43] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv:2304.12620, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[44] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[45] Aoran Xiao, Weihao Xuan, Heli Qi, Yun Xing, Ruijie Ren, Xiaojin Zhang, Ling Shao, and Shijian Lu. Cat-sam: Conditional tuning for few-shot adaptation of segment anything model. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[46] Xinyu Xiong, Zihuang Wu, Shangyi Tan, Wenxue Li, Feilong Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation. arXiv:2408.08870, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[47] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "masked image pretraining for efficient segment anything. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[48] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[49] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv:2304.11968, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[50] Linjie Yang, Yanran Wang, Xuehan Xiong, Jianchao Yang, and Aggelos Katsaggellos. Efficient video object segmentation via network modulation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[51] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[52] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[53] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv:2306.14289, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[54] Chaoning Zhang, Dongshen Han, Sheng Zheng, Jinwoo Choi, Tae-Ho Kim, and Choong Seon Hong. Mobilesamv2: Faster segment anything to everything. arXiv:2312.09579, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[55] Jiaming Zhang, Yutao Cui, Gangshan Wu, and Limin Wang. Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv:2308.13505, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[56] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv:2306.12156, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[57] Chong Zhou, Xiangtai Li, Chen Change Loy, and Bo Dai. Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam. arXiv:2312.06660, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[58] Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv:2408.00874, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3426" + } + ] +] \ No newline at end of file diff --git a/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_origin.pdf b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..29f95739acdf0fca0b90dc2a2026c3c197525053 --- /dev/null +++ b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/ca4e7e79-4820-4add-9b9e-7558abdc9127_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:746fd589a0448faf526337a9d0a38a9b85e02e46dfd7e6148ccf882a77545ad1 +size 1301169 diff --git a/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/full.md b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/full.md new file mode 100644 index 0000000000000000000000000000000000000000..874bb5316259ca2b4e7160eab9a0b8281f0611db --- /dev/null +++ b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/full.md @@ -0,0 +1,314 @@ +# SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than $0.2\%$ Training Cost + +Haiyang Mei, Pengyu Zhang, Mike Zheng Shou* Show Lab, National University of Singapore + +# Abstract + +Foundation models like the Segment Anything Model (SAM) have significantly advanced promptable image segmentation in computer vision. However, extending these capabilities to videos presents substantial challenges, particularly in ensuring precise and temporally consistent mask propagation in dynamic scenes. SAM 2 attempts to address this by training a model on massive image and video data from scratch to learn complex spatiotemporal associations, resulting in huge training costs that hinder research and practical deployment. In this paper, we introduce SAM-I2V, an effective image-to-video upgrading method for cultivating a promptable video segmentation (PVS) model. Our approach strategically upgrades the pretrained SAM to support PVS, significantly reducing training complexity and resource requirements. To achieve this, we introduce three key innovations: (i) an image-to-video feature extraction upgrader built upon SAM's static image encoder to enable spatiotemporal video perception, (ii) a memory filtering strategy that selects the most relevant past frames for more effective utilization of historical information, and (iii) a memory-as-prompt mechanism leveraging object memory to ensure temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that our method achieves over $90\%$ of SAM 2's performance while using only $0.2\%$ of its training cost. Our work presents a resource-efficient pathway to PVS, lowering barriers for further research in PVS model design and enabling broader applications and advancements in the field. Code and model will be available at: https://github.com/showlab/SAM-I2V. + +# 1. Introduction + +Foundation models have revolutionized artificial intelligence by leveraging massive datasets and large-scale models to achieve remarkable performance across various tasks. In computer vision, the Segment Anything Model + +![](images/a3a48662775cefd0497c75ed6faf677746e6f994772e0fbb663a58c459526cbe.jpg) +Figure 1. Comparison on promptable video segmentation (PVS). We depict the online-PVS performance (in $J \& F$ ) averaged on four benchmark datasets [1, 18, 36, 41] (vertical axis) with respect to model training cost (horizontal axis). The size of circles indicates model parameters (in M). SAM 2.1 [36] is a strong PVS model but needs huge training cost (i.e., 256 A100 GPUs for 108 hours). We propose to cultivate a PVS model via image-to-video model upgrading. Building on the pre-trained SAM [21, 40], incorporating existing video object segmentation models (e.g., XMem++ [1]) or temporal association modules (e.g., temporal adapter [36]) significantly reduces training costs compared to SAM 2.1 [36]. Furthermore, our designed SAM-I2V helps achieve over $90\%$ of SAM 2.1's performance while using only $0.2\%$ of its training cost. + +(SAM) [21] has emerged as a prominent foundation model for general-purpose promptable image segmentation, enabling users to segment objects/regions in images based on prompts such as points, boxes, or masks. While SAM demonstrates impressive capabilities on static images, extending these abilities to video data introduces significant challenges due to the temporal dynamics and consistency required in video segmentation tasks. + +Recently, SAM 2 [36] has been developed to address promptable video segmentation (PVS), aiming to extend SAM's image segmentation capabilities to videos. SAM 2 is trained from scratch on a massive collection of images and videos to learn complex spatiotemporal associations. However, it demands exorbitant computational resources, + +requiring 256 A100 GPUs over 108 hours for training. Such high computational costs create substantial barriers to exploring alternative architectures and advancing research in PVS, limiting accessibility for researchers and practitioners. + +This raises a critical question: Can we achieve promptable video segmentation through a resource-efficient approach? If so, it would open opportunities for developing diverse PVS foundation models, benefiting a wide range of downstream applications. Addressing this challenge, we propose SAM-I2V, an efficient method to upgrade the pretrained image-based SAM [21] model to support promptable video segmentation with affordable training cost (Figure 1). Our approach strategically leverages SAM's strong promptable image segmentation capabilities and develops image-to-video upgrading techniques to extend functionality from images to videos. + +To bridge the gap between promptable image and video segmentation, our SAM-I2V includes three key components: a temporal feature integrator TFI, a memory selective associator MSA, and a memory prompt generator MPG. (i) The TFI enhances SAM's image encoder by integrating temporal information, enabling the extraction of spatiotemporal features essential for understanding dynamic scenes; (ii) The MSA retrieves relevant segmentation information from preceding video frames through a new memory filtering mechanism, selectively associating historical information to guide the current frame's segmentation; (iii) The MPG incorporates memory as an additional prompt, utilizing object memory to ensure precise and temporally consistent mask propagation across frames. Comprehensive experiments validate that our method significantly reduces training complexity and resource requirements, achieving over $90\%$ of SAM 2's performance with only $0.2\%$ of its training cost. In summary, our contributions are: + +1. We present SAM-I2V, the first attempt to upgrade a pre-trained image-based SAM model to support promptable video segmentation, enabling PVS capabilities with substantially reduced training costs. +2. We design an upgrade method that extends SAM's static image encoder into a video perception module, enabling spatiotemporal feature extraction for PVS. +3. We propose a memory filtering strategy that selects the most relevant past frames to facilitate more effective utilization of historical information for memory-guided video segmentation. +4. We introduce a memory-as-prompt mechanism that leverages object memory to ensure temporally consistent mask propagation in dynamic scenes. + +# 2. Related Works + +Promptable Visual Segmentation. The Segment Anything Model (SAM) [21] has set a significant milestone in promptable segmentation. SAM is designed to gen + +erate a valid segmentation mask for an object of interest in an image when provided with input prompts such as bounding boxes or reference points. Trained on the extensive SA-1B dataset, SAM facilitates zero-shot segmentation with its versatile prompting mechanism, making it highly adaptable for diverse downstream tasks. This foundational work has spurred further innovations aimed at enhancing SAM's quality and efficiency. For instance, models like HQ-SAM [20] have focused on improving output precision through high-quality token integration and fine-tuning with detailed masks. Efficiency-oriented models such as EfficientSAM [47], MobileSAM [53, 54], FastSAM [56], TinySAM [40], and SlimSAM [6] have broadened the scope of SAM's applications, enabling its deployment in more resource-constrained and real-world scenarios, including medical imaging [25, 27, 43] and remote sensing [3, 37, 45]. Building upon the success of SAM [21], SAM 2 [36] represents an evolution toward promptable video segmentation, addressing the complexities of propagating segmentation across multiple frames. This allows for interaction at any frame of a video, accepting prompts such as positive/negative clicks, bounding boxes, or masks to define or refine object segmentation. Unlike its predecessor, SAM 2 not only responds to initial prompts on single frames but also propagates these segmentation cues throughout the entire video, maintaining temporal consistency and adaptability across frames. Evaluated through interactive online and offline settings, SAM 2's integration of promptable video interaction ensures comprehensive and dynamic segmentation, pushing the boundaries of practical real-world applicability [5, 15, 23, 39, 46, 58]. + +Interactive Video Object Segmentation. Interactive video object segmentation (IVOS) is an essential task for enabling real-time object tracking and segmentation in videos through user input, such as clicks or bounding boxes. Recent studies [8, 14, 17] have aimed to develop modular approaches that allow segmenting an object in a single frame and then propagating this segmentation across subsequent frames to maintain consistency. Building upon this, many IVOS methods leverage interactive mechanisms to guide segmentation over multiple frames. While SAM has been used for video object segmentation combined with video trackers based on masks [10, 12, 49] or points [35], these approaches face certain limitations: trackers may not be effective for all objects, SAM may perform suboptimally on certain video frames, and there is no interactive mechanism for correcting model errors aside from re-annotating erroneous frames and restarting the tracking process from that point. This highlights expanded functionality of SAM 2 that provides a flexible solution to enable interactive refinement. + +Semi-Supervised Video Object Segmentation. Semi-supervised video object segmentation (Semi-VOS) typically begins with a provided mask for the first frame, + +which acts as a reference for tracking the object through the video [32]. The term "semi-supervised" refers to using the initial mask as supervision, with no further manual input for subsequent frames. This task has become vital for various applications requiring automatic tracking, such as video editing, robotics, and automatic background removal. Early Semi-VOS techniques involved fine-tuning models on the first frame [2, 26, 38] or incorporating offline-trained models that predicted object masks for the entire video [28, 50, 52]. More advanced strategies integrated RNNs [48] and cross-attention [1, 7, 9, 11] to extend segmentation from the first frame throughout the sequence, improving object tracking accuracy. Recent methods [44, 55] have enhanced a single vision transformers to jointly handle the current frame alongside all prior frames and their corresponding predictions, simplifying the architecture but incuring a significant inference cost. Semi-VOS process is analogous to a specialized case of promptable video segmentation, where providing an high-quality mask prompt in the first frame is crucial but practically challenging and time-consuming, limiting the practical application of Semi-VOS in real-world scenarios. + +Image-to-Video Transfer Learning. In recent years, the importance of image-to-video transfer learning has grown due to its significant potential for reducing training costs and enhancing adaptability in video-based tasks. To facilitate efficient image-to-video transfer, various methods have been introduced, including spatio-temporal adapters [29, 51], dual-path adaptation [30], disentangled spatial and temporal learning [33], and learnable prompts [19, 22]. These methods incorporate lightweight, trainable components that leverage pre-trained image foundation models such as ViT [16] and CLIP [34] while integrating spatiotemporal dynamics, striking a balance between computational efficiency and performance to enhance video applications. These methods, although centered around imaged-to-video transfer for video recognition and understanding tasks, offer insights into our efficient cultivation of a promtable video segmentation model (i.e., SAM 2 [36]) via upgrading the pre-trained image-based SAM model [21] to a video one. These approaches provide a solid foundation for spatio-temporal information learning in our image-to-video upgradation but lack direct applicability for addressing segmentation-specific challenge, i.e., ensuring precise and temporally consistent mask propagation in scene dynamics. This makes our focus on SAM to SAM 2 upgradation a distinct and challenging endeavor. + +# 3. Methodology + +While SAM 2 demonstrates strong capabilities, its exorbitant training demands—requiring 256 A100 GPUs over 108 hours—present a significant barrier to exploring alternative architectures and advancing research in PVS. To overcome + +this challenge, we aim to develop an efficient method for cultivating a PVS model with affordable training costs. Our strategy leverages the pre-trained SAM model, which already performs well on single frames, and upgrades it to SAM 2. We design three components to achieve this transformation: a Temporal Feature Integrator (Section 3.2), a Memory Associator (3.3), and a Memory Prompt Generator (3.4), which collectively upgrade image-based SAM into a promptable video segmentation model. + +# 3.1. Overview + +Our SAM-I2V enables image-based SAM model can process video input for promtable video segmentation. As illustrated in Figure 2, the video frames are passed through a backbone feature extractor, which processes the current frame using a series of image encoder layers. These encoder outputs are enriched with temporal context through a temporal feature integrator, enabling the capture of dynamic information across consecutive frames. The enhanced features are subsequently input into the memory selective associator and memory prompt generator modules. These components manage memory states and generate target object prompts informed by past frames data, ensuring continuity and contextual awareness throughout the video. Optional user-provided prompts (e.g., masks, points, or bounding boxes) are processed by the prompt encoder to guide the segmentation with user-specific inputs. Finally, the integrated information is directed into the mask decoder, which produces the final segmentation mask for the current frame. + +# 3.2. Temporal Feature Integrator + +To enhance SAM's image encoder, which is limited to extracting only spatial features, into a video perception module, we introduce a Temporal Feature Integrator (TFI). The TFI comprises a temporal branch and an integration branch that operates synergistically to iteratively extract and integrate temporal features, enabling a comprehensive spatiotemporal feature extraction. + +As illustrated in Figure 3, given the input video $X \in \mathbb{R}^{c \times n \times h \times w}$ , where $c$ , $n$ , $h$ and $w$ are the frame channel, number, height and width, respectively, SAM's image encoder processes the current frame $X_{t}$ via a 2D stem and $L$ spatial blocks, generating multi-stage spatial features $S_{l} \in \mathbb{R}^{c_{l} \times h_{l} \times w_{l}}$ , where $c_{l}$ , $h_{l}$ and $w_{l}$ denote the features channel, height and width at stage $l \in \{1, \dots, L\}$ , respectively. The extracted features are fed into our TFI to obtain spatiotemporal features $I_{l}$ . + +Specifically, the temporal branch of TFI processes the video input $X$ through a 3D stem followed by $L$ temporal blocks to capture temporal contexts. In $l$ -th temporal block, we employ 3D convolutions $\mathcal{C}_{3d}$ to progressively + +![](images/3e7b726652c6bb910033a2e025a5d4bd92ef5d6d0c5761217bc340afe0fca57e.jpg) +Figure 2. Overview of SAM-I2V. The input video is fed into a backbone feature extractor, which passes the current frame through a series of image encoder layers. The output of these layers is enhanced with temporal information by a temporal feature integrator to capture dynamic context across frames. The extracted features are then fed into the memory associator and the memory prompt generator, which manage memory states and generate target object prompts based on past frame information. The prompt encoder processes optional user input (e.g., masks, points, or bounding boxes) to guide the segmentation. Finally, the processed information flows into the mask decoder, which outputs the segmentation mask for current frame, enabling user-intention-followed or memory conditioned promptable video segmentation. + +model temporal correspondences between frames: + +$$ +T _ {l} ^ {\prime} = T ^ {(3)}, +$$ + +$$ +T ^ {(i)} = T ^ {(i - 1)} + \mathcal {C} _ {3 d} ^ {(i)} \big (T ^ {(i - 1)} \big), +$$ + +where $T^{(0)} = T_{l - 1}$ is the input to the $l$ -th temporal block. To further enrich temporal features $T_{l}^{\prime}$ with spatial details, we fuse it with intermediate features $I_{l}^{\prime}$ from TFI's integration branch: + +$$ +T _ {l} = \mathcal {S T} (I _ {l} ^ {\prime}, T _ {l} ^ {\prime}), +$$ + +where $\mathcal{ST}$ denotes the spatial-to-temporal fusion operation that slices out current-frame temporal features from $T_{l}^{\prime}$ (i.e., the last slice in temporal dimension), concatenates it with current-frame spatial features $I_{l}^{\prime}$ along the channel dimension, applies a $1\times 1$ convolution to integrate concatenated features, and uses the resulting output to overwrite original current-frame temporal features in $T_{l}^{\prime}$ . + +Then, TFI's integration branch combines spatial features $S_{l}$ and temporal features $T_{l}^{\prime}$ to get spatiotemporal features: + +$$ +I _ {l} ^ {\prime} = I _ {l - 1} + S _ {l}, +$$ + +$$ +I _ {l} = \mathcal {C} _ {2 d} \big (\mathcal {T S} (I _ {l} ^ {\prime}, T _ {l} ^ {\prime}) \big), +$$ + +where $I_{l-1}$ is the input to the $l$ -th integration block. The temporal-to-spatial fusion operation $\mathcal{T}S$ merges current-frame temporal features in $T_l'$ with $I_l'$ , producing temporal-enriched spatial features. These features then pass through a 2D convolution $C_{2d}$ , yielding the enhanced spatiotemporal output $I_l$ . Finally, the temporal and integration branches of TFI iteratively refine the features over multiple stages, enabling the model to progressively capture rich spatiotemporal context for the subsequent segmentation process. + +# 3.3. Memory Selective Associator + +While image-based SAM [21] segments the target object based on user prompts, in the PVS task, segmentations from preceding video frames can serve as target indicators to guide the segmentation of the target in the current frame. + +![](images/f1783dfe9a12cadfa1d1fb48e1a00d0a5e1d7e8f588574fdecfbd486caf115b7.jpg) +Figure 3. Structural details of temporal feature integrator (TFI). It consists of a temporal branch and an integration branch, enhancing SAM's image encoder to function as a video perceptron. + +This motivates our Memory Selective Associator (MSA), designed to retrieve relevant and beneficial segmentation information from preceding video frames, leveraging this memory to enhance the current frame's features. + +MSA first utilizes convolution to fuse the features from the previous frames $\{F_{t - p}\}_{p = 1}^{P}$ and their corresponding segmentation masks $\{M_{t - p}\}_{p = 1}^{P}$ . These fused features are then passed to the memory selector, which chooses the most relevant past frame features. The selected features are used in the memory attention module, where cross-attention between the current frame features $I_{t}$ and the selected previous frame features is applied, producing the memory-enhanced features $I_{t}^{\prime}$ . + +The memory encoder and attention mechanism in our MSA follow a similar structure as in SAM 2 [36]. Unlike SAM 2 [36], which selects the previous $z$ consecutive frames leading up to the current frame, our memory selector chooses $z$ frames based on their relevance, as determined by similarity to the current frame. Specifically, it first chooses the prompt frame based on the user input, and then divides the previous frames into $X$ short-term (local) frames and $Y$ long-term (global) frames. The similarity between the current frame features $I_{t} \in \mathbb{R}^{N}$ and $X/Y$ previous frame features $F \in \mathbb{R}^{X/Y \times N}$ in each set is computed to identify + +![](images/65d921fc0f56c614ab9d3e4067f5d19a7eab7737565ad4fdbaa17bbfe7733f2a.jpg) +Figure 4. Illustration of our memory selective associator (MSA). It consists of a memory encoder, a memory selector, and memory attention. + +the most relevant local/global frames: + +$$ +S ^ {l o c} = F ^ {l o c} \cdot I _ {t} \in \mathbb {R} ^ {X}, S ^ {g l o} = F ^ {g l o} \cdot I _ {t} \in \mathbb {R} ^ {Y}, +$$ + +where $N = c_{L} \times h_{L} \times w_{L}$ . Next, we normalize the similarity scores to obtain the probability distribution $D^{loc}$ and $D^{glo}$ , respectively: + +$$ +D ^ {l o c} = \frac {e ^ {S ^ {l o c}}}{\sum_ {j} e ^ {S _ {j} ^ {l o c}}} \in \mathbb {R} ^ {X}, D ^ {g l o} = \frac {e ^ {S ^ {g l o}}}{\sum_ {j} e ^ {S _ {j} ^ {g l o}}} \in \mathbb {R} ^ {Y}. +$$ + +We then sample $x$ local frames and $y = z - x$ global frames based on $D^{loc}$ and $D^{glo}$ : + +$$ +i _ {k} ^ {l o c} \sim P (i _ {k} ^ {l o c} = i) = D ^ {l o c} (i), \quad k \in \{1, 2, \ldots , X \}, +$$ + +$$ +i _ {k} ^ {g l o} \sim P (i _ {k} ^ {g l o} = i) = D ^ {g l o} (i), \quad k \in \{1, 2, \ldots , Y \}, +$$ + +where $i_k^{loc}$ and $i_k^{glo}$ denote the indices of the selected local and global frames, respectively, and $D^{loc}(i)$ and $D^{glo}(i)$ are the corresponding probabilities for each frame. The selected frames are subsequently used to enhance the segmentation of the current frame through memory attention, resulting in memory-associated features $I_t'$ . + +# 3.4. Memory Prompt Generator + +SAM's mask decoder segments images using image features and user prompts to achieve intention-followed results. Motivated by its prompt integration capability, we propose incorporating memory as an additional prompt. This enables the model to leverage segmentation information from previous frames, enhancing temporal consistency across the video sequence. + +Figure 5 illustrates the proposed Memory Prompt Generator (MPG). We introduce a set of learnable memory prompt tokens $G \in \mathbb{R}^{g \times d}$ , where $g$ is the number of memory prompts and $d$ is the embedding dimension. These tokens interact with the flattened memory features $F_{\mathrm{mem}} \in \mathbb{R}^{(mh_Lw_L) \times c_L}$ , where $m$ is the number of memory frames, guided by the corresponding binarized segmentation masks + +![](images/ef25dd9d326ff9b7bbca6d6489a16bd817eddf0871266fd60879856f56c63604.jpg) +Figure 5. Illustration of the memory prompt generator (MPG). + +$M_{\mathrm{mem}} \in \mathbb{R}^{(mh_Lw_L) \times 1}$ via masked cross-attention: + +$$ +G ^ {\prime} = \operatorname {S o f t m a x} (\mathcal {M} + Q K ^ {\top}) V + G, \tag {1} +$$ + +$$ +Q = \psi (G), \quad K = \psi \left(F _ {\mathrm {m e m}}\right), \quad V = \psi \left(F _ {\mathrm {m e m}}\right), \tag {2} +$$ + +$$ +\mathcal {M} (x, y) = \left\{ \begin{array}{l l} 0, & \text {i f} M _ {\mathrm {m e m}} (x, y) = 1, \\ - \infty , & \text {o t h e r w i s e}, \end{array} \right. \tag {3} +$$ + +where $\psi$ is the learnable linear transformation. The updated memory prompt tokens $G^{\prime}$ , having captured foreground context from memories, are then refined through self-attention to incorporate intra-token dependencies: + +$$ +G ^ {\prime \prime} = \operatorname {S e l f A t t n} \left(G ^ {\prime}\right). +$$ + +Finally, the refined memory prompt embeddings $G''$ are passed through a multi-layer perceptron (MLP) to produce the final updated memory prompt tokens: + +$$ +G ^ {\prime \prime \prime} = \operatorname {M L P} \left(G ^ {\prime \prime}\right). +$$ + +Such an iterative attention-based token generation enables the MPG to selectively capture target object information from memories, providing guidance for the mask decoder. + +# 4. Experiments + +# 4.1. Experimental Setup + +Training Datasets. We train our SAM-I2V jointly on 10,000 SA-1B [21] images and 50,583 SAV [36] videos. The SA-1B dataset [21] is a comprehensive collection consisting of over 1.1 billion high-quality segmentation masks from 11 million licensed and privacy-preserving images. + +This dataset was assembled using a multi-phase data engine where initial stages involved human annotation aided by interactive tools, and subsequent phases transitioned to fully automated mask generation through the SAM model. The rich variety of mask annotations allows for robust training that supports zero-shot generalization and diverse segmentation tasks, proving effective in many downstream applications. The Segment Anything Video (SA-V) dataset [36] complements the image-focused SA-1B by providing extensive coverage for video segmentation. Its training set includes 50,583 videos and 642,036 masklets. SA-V is tailored for video scenarios involving complex object motions, occlusions, and reappearance, making it invaluable for advancing video-based models like SAM 2 [36]. + +Training Configurations. Following [36], to simulate an interactive training setting, we sample 8-frame sequences, designating the first frame as the prompt frame and randomly selecting up to 1 additional frame from the remaining seven frames for corrective clicks. The initial prompt for the first frame can be the ground-truth mask (50% probability), a positive click from the ground-truth mask (25%), or a bounding box input (25%). The corrective clicks are sampled from the error region between the model prediction and the ground-truth mask (90%) or directly from the ground-truth mask (10%). Our SAM-I2V model is implemented using PyTorch [31] and trained for 10 epochs on 8 RTX A5000 GPUs (24GB memory). We employ the AdamW optimizer [24] and apply layer-wise decay [13] specifically to the image encoder. The initial learning rates are set to $3 \times 10^{-4}$ for the image encoder and $6 \times 10^{-5}$ for the other components, respectively. These learning rates are subsequently decayed using a cosine scheduler. We supervise the model's predictions using a linear combination of mask prediction loss for the output mask, meanabsolute-error (MAE) loss for the IoU prediction, and cross-entropy loss for object prediction with a ratio of 20:1:1, respectively. The mask prediction loss $\mathcal{L}_m$ is defined as the combination of the weighted binary cross-entropy (BCE) loss $\ell_{wbce}$ [42] and the weighted IoU loss $\ell_{wiou}$ [42], i.e., $\mathcal{L}_m = \ell_{wbce} + \ell_{wiou}$ , to ensure the model pays more attention to potential distraction regions such as object boundaries, elongated areas, or holes. Notably, all SAM-I2V modules except the TinySAM [40] backbone are randomly initialized and trained from scratch without leveraging any weights from SAM 2 [36], ensuring complete independence from SAM 2 [36] in the training process. + +Evaluation Settings. To comprehensively assess the capabilities of our SAM-I2V, we perform comparisons on four benchmark datasets: ESD [18] is a high-quality 3D spatial and temporal dataset for object segmentation in an indoor cluttered environment; PUMaVOS [1] contains videos with segments around object parts such as a person's cheek; LV-VIS [41] contains videos from a diverse set of open- + +vocabulary object categories; and SAV-Test [36] is the test set of SA-V dataset, consist of 150 videos and 278 masklets. For these evaluations, we report the standard $J \& F$ metric [32] which combines region similarity (Jaccard index) and boundary accuracy (F-measure). We conducted evaluation on the promptable video segmentation. In this task, the model works in an interactive manner, where the point prompts are used to identify the target of interest or to refine the current segmentation result. Following [36], two settings are included in the comparison, namely offline evaluation (PVS-offline) and online evaluation (PVS-online). In both settings, the target is firstly initialized via $N_{\mathrm{click}}$ point prompts. As for PVS-offline, the model segments the target throughout the entire video in $N_{\mathrm{pass}}$ passes, where the interactive frame is selected based on the largest model error (i.e., with lowest IoU between the prediction and ground truth masks) and $N_{\mathrm{click}}$ points are added on such frame to refine the mask in each pass. All the prompts are used to guide the segmentation process in the next pass. PVS-online follows one pass evaluation protocol, where the segmentation process is paused and refined when the result is not reliable (IoU $< 0.75$ with ground-truth). Also, $N_{\mathrm{click}}$ points are added on the refined frame to generate corrected mask, which is used for subsequent frames. The refine process is repeated $N_{\mathrm{frame}}$ times. + +# 4.2. Comparison to Existing Methods + +To evaluate the effectiveness of our proposed SAM-I2V, we compared it against seven state-of-the-art methods on both online and offline PVS tasks under different click settings, as reported in Table 1 and 2. Besides SAM 2.1 [36], we also include upgradation methods that combine SAM (both base [21] and tiny [40] variants) with existing state-of-the-art video object segmentation models, specifically Xmem++ [1] and Cutie [11], as well as approaches that upgrade SAM by incorporating temporal adaptation (TA) [36] and spatial adaptation (SA) [4] modules. + +Performance Comparison with SAM 2.1. While SAM 2.1 [36] serves as a strong baseline with an overall accuracy of 71.9 (online) and 73.3 (offline), it requires substantial computational resources (2.2 million $G \cdot$ Hour). In contrast, our SAM-I2V achieves its $91\%$ (i.e., 65.3, online) and $90\%$ (i.e., 66.0, offline) overall scores, with its $0.2\%$ (i.e., 4.6k $G \cdot$ Hour) training cost and its $49\%$ (i.e., 18.9m vs. 38.9m) model parameters. + +Effectiveness over Existing SAM Upgradation Methods. Compared to SAM [21] + XMem++ [1] (157.0M parameters) and TinySAM [40] + XMem++ (72.3M parameters), our SAM-I2V offers a substantial reduction in model size (18.9M) while delivering comparable or superior performance. For example, under the online setting, SAM-I2V helps achieve an overall score of 65.3, outperforming SAM + XMem++ (60.4) and SAM + Cutie [11] (61.7). + +
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.645.859.366.386.866.879.575.288.870.182.877.279.760.974.972.971.9
SAM [21] + XMem++ [1]157.0-78.349.368.042.478.846.273.041.979.345.478.443.378.847.073.142.560.4
SAM [21] + Cutie [11]129.8-78.438.369.451.279.044.373.951.879.844.676.952.979.142.473.452.061.7
TinySAM [40] + XMem++ [1]72.3-79.355.268.547.382.959.874.948.784.260.476.951.682.158.573.449.265.8
TinySAM [40] + Cutie [11]45.1-80.653.071.158.483.352.676.259.984.456.278.863.182.853.975.460.568.1
TinySAM [40] + TA [36]17.84.6k68.728.856.242.980.245.172.349.182.149.876.550.977.041.268.447.658.6
TinySAM [40] + SA [4] + TA [36]18.74.6k54.834.652.844.179.352.772.950.681.959.476.851.972.048.967.548.959.3
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.039.153.850.584.760.873.458.286.465.978.359.581.455.368.556.165.3
+ +Table 1. Comparisons of the online promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance. + +
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.946.261.366.887.767.383.976.290.170.787.578.180.661.477.673.773.3
SAM [21] + XMem++ [1]157.0-49.536.850.141.164.049.353.744.165.354.755.244.959.646.953.043.450.7
TinySAM [40] + XMem++ [1]72.3-47.746.047.549.173.655.452.955.575.962.553.758.265.754.651.454.356.5
TinySAM [40] + TA [36]17.84.6k68.929.558.345.381.149.376.352.482.955.681.154.977.644.871.950.961.3
TinySAM [40] + SA [4] + TA [36]18.74.6k57.037.054.944.880.956.676.652.583.462.181.253.673.851.970.950.361.7
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.437.855.748.984.661.277.858.586.465.183.059.881.554.772.255.766.0
+ +Table 2. Comparisons of the offline promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance. Cutie [11] is excluded from the comparison as it only supports single-frame mask inputs, rendering it unsuitable for offline PVS settings. + +Robust Performance Across Interaction Levels. The online and offline results presented in Table 1 and 2 reflect that the performance of all methods generally improves as the number of interactions increases. Our SAM-I2V consistently delivers superiority across 1-click, 3-click, and 5-click settings, underscoring its robustness to varying levels of user interaction. + +Robust Online and Offline Performance. SAM-I2V's overall online performance in Table 1 is slightly lower than the offline performance in Table 2. This discrepancy arises because, unlike in the offline evaluation, in the online setting the new prompts affect only the frames following the current paused frame and not the preceding frames. Compared to other methods, our approach maintains an advantage in both online and offline settings. + +The above results confirm that our SAM-I2V not only matches but often surpasses existing upgrading methods in performance with reduced computational costs, making SAM-I2V a practical choice for PVS tasks. + +# 4.3. Ablation Study + +We conduct an extensive ablation study to validate the extensibility of SAM-I2V and the effectiveness of its each key component. Table 3, 4, 5, 6 summarize our findings. + +Applicability of SAM-I2V to Upgrade Different SAM Models. In the first ablation study, we investigate the performance of our SAM-I2V to upgrade different SAM models. From the results presented in Table 3, it is evident that our SAM-I2V effectively upgrades different SAM variants to a PVS model, demonstrate that SAM-I2V is a robust and versatile image-to-video promptable segmentation upgradation approach. + +Effectiveness of the Temporal Feature Integrator. As shown in Table 4, our proposed TFI module outperforms the baseline (TinySAM [40] + Temporal Associator [36]) and other adaptation methods. While previous approaches like ST-Adapter [29] and Parallel Adapter [4] perform adaptation within each attention block, our TFI is specifically designed with temporal branch and integration branch to iter + +
MethodsES[18]PU[1]LV[41]SV[36]OA
TinySAM [40] + TA [36]80.245.172.349.161.7
TinySAM [40] + SAM-I2V84.760.873.458.269.3
EdgeSAM [57] + TA [36]78.341.469.644.658.5
EdgeSAM [57] + SAM-I2V83.251.971.654.465.1
+ +Table 3. Ablation on the applicability of SAM-I2V to upgrade different SAM models for PVS task under the online, 3-click setting. + +
MethodsES[18]PU[1]LV[41]SV[36]OA
Base80.245.172.349.161.7
Base + ST-Adapter [29]79.354.173.350.164.2
Base + Parallel Adapter [4]79.352.772.950.663.9
Base + TFI81.358.172.452.566.1
+ +atively extract and integrate spatial-temporal features. This allows our model to capture richer temporal dependencies, leading to superior performance. + +Effectiveness of the Memory Selective Associator. Relying on the previous $z = 6$ consecutive frames for segmentation is equivalent to setting $x$ and $y$ in MSA to 0 and 6, respectively. By varying the number of selected global $(x)$ and local $(y)$ frames in our MSA, we observe that the segmentation performance improves when the MSA selectively retrieves the most relevant frames based on their similarity to the current frame. Specifically, the configuration with $x = 3$ global frames and $y = 3$ local frames performs better, highlighting the importance of incorporating both short-term and long-term memory. In addition, the ability of MSA to select frames based on relevance rather than uniform sampling allows the model to focus on the most relevant past frames, leading to improved performance. + +Effectiveness of the Memory Prompt Generator. From Table 6, we can observe that, (i) introducing MPG $(w/$ MPG-1/3/5) consistently improves performance over the one without MPG, demonstrating the effectiveness of the memory-as-prompt strategy in enhancing segmentation accuracy; (ii) varying the number of memory prompt tokens among 1, 3, and 5 yields comparable results, indicating that the MPG is robust to the choice of token quantity within this range; and (iii) removing the Masked Cross-Attention (MCA) mechanism (i.e., $w/$ MPG-3 ( $w/o$ MCA)) leads to a performance drop, highlighting the importance of MCA in focusing on relevant foreground information, eliminating background interference, and thus enhancing the quality of memory prompt generation. + +# 4.4. Limitation and Future Work + +One limitation of our approach is that its accuracy does not well match that of SAM 2.1 [36]. Note that the released SAM 2.1 model was trained not only on the SAM-1B + +Table 4. Ablation on the effectiveness of the temporal feature integrator TFI for online PVS task under the 3-click setting. + +
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MSA (i.e., x=0, y=6)84.054.473.256.967.1
w/ MSA (x=1, y=5)84.754.873.457.967.7
w/ MSA (x=2, y=4)84.556.373.456.967.8
w/ MSA (x=3, y=3)84.760.873.458.269.3
w/ MSA (x=3, y=3) w/ UFS84.156.073.056.567.4
+ +Table 5. Ablation on the effectiveness of the memory selective associator MSA for online PVS task under the 3-click setting. "x" and "y" denote the number of selected global and local frames, respectively. "UFS" indicates the uniform frame sampling strategy. + +
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MPG82.157.772.153.266.3
w/ MPG-184.058.473.055.667.8
w/ MPG-384.760.873.458.269.3
w/ MPG-584.558.673.657.168.5
w/ MPG-3 (w/o MCA)84.958.172.855.167.7
+ +Table 6. Ablation on the effectiveness of the memory prompt generator MPG for online PVS task under the 3-click setting. "MPG-g" denotes the MPG with $g$ learnable memory prompt tokens. "MCA" indicates the masked cross attention. + +and SAV datasets but also on 62.9k proprietary, non-public videos, which our method was unable to utilize. Despite this, our method demonstrates superior performance compared to baseline approaches and creates targeted model design opportunities for downstream applications in resource-constrained academic scenarios. Future work will focus on enhancing our model by incorporating additional publicly available datasets and exploring advanced training techniques to bridge the accuracy gap with SAM 2.1 [36]. + +# 5. Conclusion + +In this work, we introduced SAM-I2V, a training-efficient framework that upgrades the pre-trained Segment Anything Model (SAM) to support promptable video segmentation (PVS). SAM-I2V consists of three key innovations: an image-to-video feature extraction upgrader that extends SAM's static image encoder for spatiotemporal video perception; a memory filtering strategy that selects the most relevant past frames to enhance the utilization of historical information; and a memory-as-prompt mechanism that leverages object memory for temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that SAM-I2V achieved over $90\%$ of SAM 2's performance at only $0.2\%$ of its training cost. This remarkable training efficiency significantly lowers the barriers for developing diverse PVS models, facilitating rapid innovation and broader practical applications. + +Acknowledgments. This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012). + +# References + +[1] Maksym Bekuzarov, Ariana Bermudez, Joon-Young Lee, and Hao Li. Xmem++: Production-level video segmentation from few annotated frames. In ICCV, 2023. 1, 3, 6, 7, 8 +[2] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, 2017. 3 +[3] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2 +[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 6, 7, 8 +[5] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv:2408.04579, 2024. 2 +[6] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. $0.1\%$ data makes segment anything slim. NeurIPS, 2024. 2 +[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3 +[8] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 2 +[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. NeurIPS, 2021. 3 +[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 2 +[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In CVPR, 2024. 3, 6, 7 +[12] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv:2305.06558, 2023. 2 +[13] K Clark. Electra: Pre-training text encoders as discriminators rather than generators. arXiv:2003.10555, 2020. 6 +[14] Thanos Delatolas, Vicky Kalogeiton, and Dim P Papadopoulos. Learning the what and how of annotation in video object segmentation. In WACV, 2024. 2 +[15] Haoyu Dong, Hanxue Gu, Yaqian Chen, Jichen Yang, and Maciej A Mazurowski. Segment anything model 2: an application to 2d and 3d medical images. arXiv:2408.00756, 2024. 2 +[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3 + +[17] Yuk Heo, Yeong Jun Koh, and Chang-Su Kim. Interactive video object segmentation using global and local transfer modules. In ECCV, 2020. 2 +[18] Xiaoqian Huang, Kachole Sanket, Abdulla Ayyad, Fariborz Baghaei Naeini, Dimitrios Makris, and Yahya Zweiri. A neuromorphic dataset for object segmentation in indoor cluttered environment. arXiv:2302.06301, 2023. 1, 6, 7, 8 +[19] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3 +[20] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. NeurIPS, 2023. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7 +[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 3 +[23] Haofeng Liu, Erli Zhang, Junde Wu, Mingxuan Hong, and Yueming Jin. Surgical sam 2: Real-time segment anything in surgical video by efficient frame pruning. arXiv:2408.07931, 2024. 2 +[24] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6 +[25] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 2024. 2 +[26] K-K Maninis, Sergi Caelles, Yuhua Chen, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. Video object segmentation without temporal information. IEEE TPAMI, 2018. 3 +[27] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 2023. 2 +[28] Seoung Wug Oh, Joon-Young Lee, Kalyan Sunkavalli, and Seon Joo Kim. Fast video object segmentation by reference-guided mask propagation. In CVPR, 2018. 3 +[29] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. NeurIPS, 2022. 3, 7, 8 +[30] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dualpath adaptation from image to video transformers. In CVPR, 2023. 3 +[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6 +[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 3, 6 + +[33] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In ICCV, 2023. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[35] Frano Rajic, Lei Ke, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu. Segment anything meets point tracking. arXiv:2307.01197, 2023. 2 +[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[37] Simiao Ren, Francesco Luzi, Saad Lahrichi, Kaleb Kassaw, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Segment anything, from space? In WACV, 2024. 2 +[38] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In CVPR, 2020. 3 +[39] Chuyun Shen, Wenhao Li, Yuhang Shi, and Xiangfeng Wang. Interactive 3d medical image segmentation with sam 2. arXiv:2408.02635, 2024. 2 +[40] Han Shu, Wenshuo Li, Yehui Tang, Yiman Zhang, Yihao Chen, Houqiang Li, Yunhe Wang, and Xinghao Chen. Tinysam: Pushing the envelope for efficient segment anything model. arXiv:2312.13789, 2023. 1, 2, 6, 7, 8 +[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 6, 7, 8 +[42] Jun Wei, Shuhui Wang, and Qingming Huang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 6 +[43] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv:2304.12620, 2023. 2 +[44] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, 2023. 3 +[45] Aoran Xiao, Weihao Xuan, Heli Qi, Yun Xing, Ruijie Ren, Xiaojin Zhang, Ling Shao, and Shijian Lu. Cat-sam: Conditional tuning for few-shot adaptation of segment anything model. In ECCV, 2024. 2 +[46] Xinyu Xiong, Zihuang Wu, Shangyi Tan, Wenxue Li, Feilong Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation. arXiv:2408.08870, 2024. 2 +[47] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged + +masked image pretraining for efficient segment anything. In CVPR, 2024. 2 +[48] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 3 +[49] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv:2304.11968, 2023. 2 +[50] Linjie Yang, Yanran Wang, Xuehan Xiong, Jianchao Yang, and Aggelos Katsaggellos. Efficient video object segmentation via network modulation. In CVPR, 2018. 3 +[51] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In ICLR, 2023. 3 +[52] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3 +[53] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv:2306.14289, 2023. 2 +[54] Chaoning Zhang, Dongshen Han, Sheng Zheng, Jinwoo Choi, Tae-Ho Kim, and Choong Seon Hong. Mobilesamv2: Faster segment anything to everything. arXiv:2312.09579, 2023. 2 +[55] Jiaming Zhang, Yutao Cui, Gangshan Wu, and Limin Wang. Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv:2308.13505, 2023. 3 +[56] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv:2306.12156, 2023. 2 +[57] Chong Zhou, Xiangtai Li, Chen Change Loy, and Bo Dai. Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam. arXiv:2312.06660, 2023. 8 +[58] Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv:2408.00874, 2024. 2 \ No newline at end of file diff --git a/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/images.zip b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2993eb3a54460cb9ac65c1be21990c68c0faaa30 --- /dev/null +++ b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd0324f592873012478626f9a79087785e8c334a11f01d03a9fec6afaf8d49c1 +size 508335 diff --git a/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/layout.json b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72317e95d3bfc9224e436f3f4c4ec9caf6cd9fa2 --- /dev/null +++ b/2025/SAM-I2V_ Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost/layout.json @@ -0,0 +1,8692 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "type": "text", + "content": "SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than " + }, + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "type": "text", + "content": " Training Cost" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 161, + 428, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 161, + 428, + 190 + ], + "spans": [ + { + "bbox": [ + 181, + 161, + 428, + 190 + ], + "type": "text", + "content": "Haiyang Mei, Pengyu Zhang, Mike Zheng Shou* Show Lab, National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "type": "text", + "content": "Foundation models like the Segment Anything Model (SAM) have significantly advanced promptable image segmentation in computer vision. However, extending these capabilities to videos presents substantial challenges, particularly in ensuring precise and temporally consistent mask propagation in dynamic scenes. SAM 2 attempts to address this by training a model on massive image and video data from scratch to learn complex spatiotemporal associations, resulting in huge training costs that hinder research and practical deployment. In this paper, we introduce SAM-I2V, an effective image-to-video upgrading method for cultivating a promptable video segmentation (PVS) model. Our approach strategically upgrades the pretrained SAM to support PVS, significantly reducing training complexity and resource requirements. To achieve this, we introduce three key innovations: (i) an image-to-video feature extraction upgrader built upon SAM's static image encoder to enable spatiotemporal video perception, (ii) a memory filtering strategy that selects the most relevant past frames for more effective utilization of historical information, and (iii) a memory-as-prompt mechanism leveraging object memory to ensure temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that our method achieves over " + }, + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "type": "text", + "content": " of SAM 2's performance while using only " + }, + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 53, + 242, + 297, + 602 + ], + "type": "text", + "content": " of its training cost. Our work presents a resource-efficient pathway to PVS, lowering barriers for further research in PVS model design and enabling broader applications and advancements in the field. Code and model will be available at: https://github.com/showlab/SAM-I2V." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 625, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 645, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 695 + ], + "type": "text", + "content": "Foundation models have revolutionized artificial intelligence by leveraging massive datasets and large-scale models to achieve remarkable performance across various tasks. In computer vision, the Segment Anything Model" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 220, + 553, + 381 + ], + "blocks": [ + { + "bbox": [ + 320, + 220, + 553, + 381 + ], + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 381 + ], + "type": "image", + "image_path": "a3a48662775cefd0497c75ed6faf677746e6f994772e0fbb663a58c459526cbe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "text", + "content": "Figure 1. Comparison on promptable video segmentation (PVS). We depict the online-PVS performance (in " + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "inline_equation", + "content": "J \\& F" + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "text", + "content": ") averaged on four benchmark datasets [1, 18, 36, 41] (vertical axis) with respect to model training cost (horizontal axis). The size of circles indicates model parameters (in M). SAM 2.1 [36] is a strong PVS model but needs huge training cost (i.e., 256 A100 GPUs for 108 hours). We propose to cultivate a PVS model via image-to-video model upgrading. Building on the pre-trained SAM [21, 40], incorporating existing video object segmentation models (e.g., XMem++ [1]) or temporal association modules (e.g., temporal adapter [36]) significantly reduces training costs compared to SAM 2.1 [36]. Furthermore, our designed SAM-I2V helps achieve over " + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "text", + "content": " of SAM 2.1's performance while using only " + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 387, + 556, + 531 + ], + "type": "text", + "content": " of its training cost." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 544, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 640 + ], + "type": "text", + "content": "(SAM) [21] has emerged as a prominent foundation model for general-purpose promptable image segmentation, enabling users to segment objects/regions in images based on prompts such as points, boxes, or masks. While SAM demonstrates impressive capabilities on static images, extending these abilities to video data introduces significant challenges due to the temporal dynamics and consistency required in video segmentation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Recently, SAM 2 [36] has been developed to address promptable video segmentation (PVS), aiming to extend SAM's image segmentation capabilities to videos. SAM 2 is trained from scratch on a massive collection of images and videos to learn complex spatiotemporal associations. However, it demands exorbitant computational resources," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3417" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "requiring 256 A100 GPUs over 108 hours for training. Such high computational costs create substantial barriers to exploring alternative architectures and advancing research in PVS, limiting accessibility for researchers and practitioners." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 294, + 262 + ], + "type": "text", + "content": "This raises a critical question: Can we achieve promptable video segmentation through a resource-efficient approach? If so, it would open opportunities for developing diverse PVS foundation models, benefiting a wide range of downstream applications. Addressing this challenge, we propose SAM-I2V, an efficient method to upgrade the pretrained image-based SAM [21] model to support promptable video segmentation with affordable training cost (Figure 1). Our approach strategically leverages SAM's strong promptable image segmentation capabilities and develops image-to-video upgrading techniques to extend functionality from images to videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "type": "text", + "content": "To bridge the gap between promptable image and video segmentation, our SAM-I2V includes three key components: a temporal feature integrator TFI, a memory selective associator MSA, and a memory prompt generator MPG. (i) The TFI enhances SAM's image encoder by integrating temporal information, enabling the extraction of spatiotemporal features essential for understanding dynamic scenes; (ii) The MSA retrieves relevant segmentation information from preceding video frames through a new memory filtering mechanism, selectively associating historical information to guide the current frame's segmentation; (iii) The MPG incorporates memory as an additional prompt, utilizing object memory to ensure precise and temporally consistent mask propagation across frames. Comprehensive experiments validate that our method significantly reduces training complexity and resource requirements, achieving over " + }, + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "type": "text", + "content": " of SAM 2's performance with only " + }, + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 56, + 263, + 295, + 478 + ], + "type": "text", + "content": " of its training cost. In summary, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 479, + 293, + 645 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 63, + 479, + 293, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 479, + 293, + 525 + ], + "spans": [ + { + "bbox": [ + 63, + 479, + 293, + 525 + ], + "type": "text", + "content": "1. We present SAM-I2V, the first attempt to upgrade a pre-trained image-based SAM model to support promptable video segmentation, enabling PVS capabilities with substantially reduced training costs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 527, + 293, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 527, + 293, + 561 + ], + "spans": [ + { + "bbox": [ + 63, + 527, + 293, + 561 + ], + "type": "text", + "content": "2. We design an upgrade method that extends SAM's static image encoder into a video perception module, enabling spatiotemporal feature extraction for PVS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 563, + 293, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 563, + 293, + 609 + ], + "spans": [ + { + "bbox": [ + 63, + 563, + 293, + 609 + ], + "type": "text", + "content": "3. We propose a memory filtering strategy that selects the most relevant past frames to facilitate more effective utilization of historical information for memory-guided video segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 611, + 293, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 611, + 293, + 645 + ], + "spans": [ + { + "bbox": [ + 63, + 611, + 293, + 645 + ], + "type": "text", + "content": "4. We introduce a memory-as-prompt mechanism that leverages object memory to ensure temporally consistent mask propagation in dynamic scenes." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 657, + 146, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 146, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 146, + 669 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "Promptable Visual Segmentation. The Segment Anything Model (SAM) [21] has set a significant milestone in promptable segmentation. SAM is designed to gen" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 443 + ], + "type": "text", + "content": "erate a valid segmentation mask for an object of interest in an image when provided with input prompts such as bounding boxes or reference points. Trained on the extensive SA-1B dataset, SAM facilitates zero-shot segmentation with its versatile prompting mechanism, making it highly adaptable for diverse downstream tasks. This foundational work has spurred further innovations aimed at enhancing SAM's quality and efficiency. For instance, models like HQ-SAM [20] have focused on improving output precision through high-quality token integration and fine-tuning with detailed masks. Efficiency-oriented models such as EfficientSAM [47], MobileSAM [53, 54], FastSAM [56], TinySAM [40], and SlimSAM [6] have broadened the scope of SAM's applications, enabling its deployment in more resource-constrained and real-world scenarios, including medical imaging [25, 27, 43] and remote sensing [3, 37, 45]. Building upon the success of SAM [21], SAM 2 [36] represents an evolution toward promptable video segmentation, addressing the complexities of propagating segmentation across multiple frames. This allows for interaction at any frame of a video, accepting prompts such as positive/negative clicks, bounding boxes, or masks to define or refine object segmentation. Unlike its predecessor, SAM 2 not only responds to initial prompts on single frames but also propagates these segmentation cues throughout the entire video, maintaining temporal consistency and adaptability across frames. Evaluated through interactive online and offline settings, SAM 2's integration of promptable video interaction ensures comprehensive and dynamic segmentation, pushing the boundaries of practical real-world applicability [5, 15, 23, 39, 46, 58]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 446, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 553, + 674 + ], + "type": "text", + "content": "Interactive Video Object Segmentation. Interactive video object segmentation (IVOS) is an essential task for enabling real-time object tracking and segmentation in videos through user input, such as clicks or bounding boxes. Recent studies [8, 14, 17] have aimed to develop modular approaches that allow segmenting an object in a single frame and then propagating this segmentation across subsequent frames to maintain consistency. Building upon this, many IVOS methods leverage interactive mechanisms to guide segmentation over multiple frames. While SAM has been used for video object segmentation combined with video trackers based on masks [10, 12, 49] or points [35], these approaches face certain limitations: trackers may not be effective for all objects, SAM may perform suboptimally on certain video frames, and there is no interactive mechanism for correcting model errors aside from re-annotating erroneous frames and restarting the tracking process from that point. This highlights expanded functionality of SAM 2 that provides a flexible solution to enable interactive refinement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Semi-Supervised Video Object Segmentation. Semi-supervised video object segmentation (Semi-VOS) typically begins with a provided mask for the first frame," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3418" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 334 + ], + "type": "text", + "content": "which acts as a reference for tracking the object through the video [32]. The term \"semi-supervised\" refers to using the initial mask as supervision, with no further manual input for subsequent frames. This task has become vital for various applications requiring automatic tracking, such as video editing, robotics, and automatic background removal. Early Semi-VOS techniques involved fine-tuning models on the first frame [2, 26, 38] or incorporating offline-trained models that predicted object masks for the entire video [28, 50, 52]. More advanced strategies integrated RNNs [48] and cross-attention [1, 7, 9, 11] to extend segmentation from the first frame throughout the sequence, improving object tracking accuracy. Recent methods [44, 55] have enhanced a single vision transformers to jointly handle the current frame alongside all prior frames and their corresponding predictions, simplifying the architecture but incuring a significant inference cost. Semi-VOS process is analogous to a specialized case of promptable video segmentation, where providing an high-quality mask prompt in the first frame is crucial but practically challenging and time-consuming, limiting the practical application of Semi-VOS in real-world scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 335, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 295, + 635 + ], + "type": "text", + "content": "Image-to-Video Transfer Learning. In recent years, the importance of image-to-video transfer learning has grown due to its significant potential for reducing training costs and enhancing adaptability in video-based tasks. To facilitate efficient image-to-video transfer, various methods have been introduced, including spatio-temporal adapters [29, 51], dual-path adaptation [30], disentangled spatial and temporal learning [33], and learnable prompts [19, 22]. These methods incorporate lightweight, trainable components that leverage pre-trained image foundation models such as ViT [16] and CLIP [34] while integrating spatiotemporal dynamics, striking a balance between computational efficiency and performance to enhance video applications. These methods, although centered around imaged-to-video transfer for video recognition and understanding tasks, offer insights into our efficient cultivation of a promtable video segmentation model (i.e., SAM 2 [36]) via upgrading the pre-trained image-based SAM model [21] to a video one. These approaches provide a solid foundation for spatio-temporal information learning in our image-to-video upgradation but lack direct applicability for addressing segmentation-specific challenge, i.e., ensuring precise and temporally consistent mask propagation in scene dynamics. This makes our focus on SAM to SAM 2 upgradation a distinct and challenging endeavor." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 645, + 138, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 138, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 138, + 659 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "While SAM 2 demonstrates strong capabilities, its exorbitant training demands—requiring 256 A100 GPUs over 108 hours—present a significant barrier to exploring alternative architectures and advancing research in PVS. To overcome" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "this challenge, we aim to develop an efficient method for cultivating a PVS model with affordable training costs. Our strategy leverages the pre-trained SAM model, which already performs well on single frames, and upgrades it to SAM 2. We design three components to achieve this transformation: a Temporal Feature Integrator (Section 3.2), a Memory Associator (3.3), and a Memory Prompt Generator (3.4), which collectively upgrade image-based SAM into a promptable video segmentation model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 196, + 382, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 382, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 382, + 208 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": "Our SAM-I2V enables image-based SAM model can process video input for promtable video segmentation. As illustrated in Figure 2, the video frames are passed through a backbone feature extractor, which processes the current frame using a series of image encoder layers. These encoder outputs are enriched with temporal context through a temporal feature integrator, enabling the capture of dynamic information across consecutive frames. The enhanced features are subsequently input into the memory selective associator and memory prompt generator modules. These components manage memory states and generate target object prompts informed by past frames data, ensuring continuity and contextual awareness throughout the video. Optional user-provided prompts (e.g., masks, points, or bounding boxes) are processed by the prompt encoder to guide the segmentation with user-specific inputs. Finally, the integrated information is directed into the mask decoder, which produces the final segmentation mask for the current frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 448, + 473, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 473, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 473, + 460 + ], + "type": "text", + "content": "3.2. Temporal Feature Integrator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 468, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 552 + ], + "type": "text", + "content": "To enhance SAM's image encoder, which is limited to extracting only spatial features, into a video perception module, we introduce a Temporal Feature Integrator (TFI). The TFI comprises a temporal branch and an integration branch that operates synergistically to iteratively extract and integrate temporal features, enabling a comprehensive spatiotemporal feature extraction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": "As illustrated in Figure 3, given the input video " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{c \\times n \\times h \\times w}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": " are the frame channel, number, height and width, respectively, SAM's image encoder processes the current frame " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": " via a 2D stem and " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": " spatial blocks, generating multi-stage spatial features " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "S_{l} \\in \\mathbb{R}^{c_{l} \\times h_{l} \\times w_{l}}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "c_{l}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "h_{l}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "w_{l}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": " denote the features channel, height and width at stage " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "l \\in \\{1, \\dots, L\\}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": ", respectively. The extracted features are fed into our TFI to obtain spatiotemporal features " + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 313, + 555, + 553, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "Specifically, the temporal branch of TFI processes the video input " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " through a 3D stem followed by " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " temporal blocks to capture temporal contexts. In " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "-th temporal block, we employ 3D convolutions " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{3d}" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " to progressively" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3419" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 166 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 166 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 166 + ], + "type": "image", + "image_path": "3e7b726652c6bb910033a2e025a5d4bd92ef5d6d0c5761217bc340afe0fca57e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 171, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 54, + 171, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 555, + 239 + ], + "type": "text", + "content": "Figure 2. Overview of SAM-I2V. The input video is fed into a backbone feature extractor, which passes the current frame through a series of image encoder layers. The output of these layers is enhanced with temporal information by a temporal feature integrator to capture dynamic context across frames. The extracted features are then fed into the memory associator and the memory prompt generator, which manage memory states and generate target object prompts based on past frame information. The prompt encoder processes optional user input (e.g., masks, points, or bounding boxes) to guide the segmentation. Finally, the processed information flows into the mask decoder, which outputs the segmentation mask for current frame, enabling user-intention-followed or memory conditioned promptable video segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 258, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 258, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 258, + 262 + ], + "type": "text", + "content": "model temporal correspondences between frames:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 269, + 166, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 269, + 166, + 285 + ], + "spans": [ + { + "bbox": [ + 118, + 269, + 166, + 285 + ], + "type": "interline_equation", + "content": "T _ {l} ^ {\\prime} = T ^ {(3)},", + "image_path": "3d08c5e82844869276eefbcfeffd99f11d59cd84661d9ac33ff56e97eb167494.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 287, + 238, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 287, + 238, + 304 + ], + "spans": [ + { + "bbox": [ + 112, + 287, + 238, + 304 + ], + "type": "interline_equation", + "content": "T ^ {(i)} = T ^ {(i - 1)} + \\mathcal {C} _ {3 d} ^ {(i)} \\big (T ^ {(i - 1)} \\big),", + "image_path": "c38245207c19b289c8db1f945456b3cf6c2b73ddbb5e7cb3e3ba76171b08268e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "inline_equation", + "content": "T^{(0)} = T_{l - 1}" + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": " is the input to the " + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": "-th temporal block. To further enrich temporal features " + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "inline_equation", + "content": "T_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": " with spatial details, we fuse it with intermediate features " + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "inline_equation", + "content": "I_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": " from TFI's integration branch:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 360, + 212, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 360, + 212, + 374 + ], + "spans": [ + { + "bbox": [ + 138, + 360, + 212, + 374 + ], + "type": "interline_equation", + "content": "T _ {l} = \\mathcal {S T} (I _ {l} ^ {\\prime}, T _ {l} ^ {\\prime}),", + "image_path": "70d730250a30eff675a5ced9a6a36a673e1a11849c6367fcfd74b1892c0d24e5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{ST}" + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": " denotes the spatial-to-temporal fusion operation that slices out current-frame temporal features from " + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "inline_equation", + "content": "T_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": " (i.e., the last slice in temporal dimension), concatenates it with current-frame spatial features " + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "inline_equation", + "content": "I_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": " along the channel dimension, applies a " + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": " convolution to integrate concatenated features, and uses the resulting output to overwrite original current-frame temporal features in " + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "inline_equation", + "content": "T_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 378, + 296, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "type": "text", + "content": "Then, TFI's integration branch combines spatial features " + }, + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "type": "text", + "content": " and temporal features " + }, + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "type": "inline_equation", + "content": "T_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 462, + 296, + 487 + ], + "type": "text", + "content": " to get spatiotemporal features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 495, + 195, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 495, + 195, + 508 + ], + "spans": [ + { + "bbox": [ + 127, + 495, + 195, + 508 + ], + "type": "interline_equation", + "content": "I _ {l} ^ {\\prime} = I _ {l - 1} + S _ {l},", + "image_path": "eda7c12723775bbcd0418870cc6f9b57584891a456ec92b62c9c3f7ffc4ad455.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 511, + 222, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 511, + 222, + 525 + ], + "spans": [ + { + "bbox": [ + 129, + 511, + 222, + 525 + ], + "type": "interline_equation", + "content": "I _ {l} = \\mathcal {C} _ {2 d} \\big (\\mathcal {T S} (I _ {l} ^ {\\prime}, T _ {l} ^ {\\prime}) \\big),", + "image_path": "b02be6bf7ae73b5260641e8b7f00d32cc4157a36f43b6c6cda3537961e5d7028.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "I_{l-1}" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": " is the input to the " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": "-th integration block. The temporal-to-spatial fusion operation " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{T}S" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": " merges current-frame temporal features in " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "T_l'" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "I_l'" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": ", producing temporal-enriched spatial features. These features then pass through a 2D convolution " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "C_{2d}" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": ", yielding the enhanced spatiotemporal output " + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 55, + 533, + 296, + 641 + ], + "type": "text", + "content": ". Finally, the temporal and integration branches of TFI iteratively refine the features over multiple stages, enabling the model to progressively capture rich spatiotemporal context for the subsequent segmentation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 648, + 214, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 214, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 214, + 661 + ], + "type": "text", + "content": "3.3. Memory Selective Associator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "While image-based SAM [21] segments the target object based on user prompts, in the PVS task, segmentations from preceding video frames can serve as target indicators to guide the segmentation of the target in the current frame." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 322, + 251, + 549, + 370 + ], + "blocks": [ + { + "bbox": [ + 322, + 251, + 549, + 370 + ], + "lines": [ + { + "bbox": [ + 322, + 251, + 549, + 370 + ], + "spans": [ + { + "bbox": [ + 322, + 251, + 549, + 370 + ], + "type": "image", + "image_path": "f1783dfe9a12cadfa1d1fb48e1a00d0a5e1d7e8f588574fdecfbd486caf115b7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 378, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 412 + ], + "type": "text", + "content": "Figure 3. Structural details of temporal feature integrator (TFI). It consists of a temporal branch and an integration branch, enhancing SAM's image encoder to function as a video perceptron." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 425, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 472 + ], + "type": "text", + "content": "This motivates our Memory Selective Associator (MSA), designed to retrieve relevant and beneficial segmentation information from preceding video frames, leveraging this memory to enhance the current frame's features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": "MSA first utilizes convolution to fuse the features from the previous frames " + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\{F_{t - p}\\}_{p = 1}^{P}" + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": " and their corresponding segmentation masks " + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\{M_{t - p}\\}_{p = 1}^{P}" + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": ". These fused features are then passed to the memory selector, which chooses the most relevant past frame features. The selected features are used in the memory attention module, where cross-attention between the current frame features " + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": " and the selected previous frame features is applied, producing the memory-enhanced features " + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "inline_equation", + "content": "I_{t}^{\\prime}" + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "The memory encoder and attention mechanism in our MSA follow a similar structure as in SAM 2 [36]. Unlike SAM 2 [36], which selects the previous " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " consecutive frames leading up to the current frame, our memory selector chooses " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " frames based on their relevance, as determined by similarity to the current frame. Specifically, it first chooses the prompt frame based on the user input, and then divides the previous frames into " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " short-term (local) frames and " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " long-term (global) frames. The similarity between the current frame features " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "I_{t} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "X/Y" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " previous frame features " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{X/Y \\times N}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " in each set is computed to identify" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3420" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 555, + 186 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 555, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 555, + 186 + ], + "type": "image", + "image_path": "65d921fc0f56c614ab9d3e4067f5d19a7eab7737565ad4fdbaa17bbfe7733f2a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 206 + ], + "type": "text", + "content": "Figure 4. Illustration of our memory selective associator (MSA). It consists of a memory encoder, a memory selector, and memory attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 209, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 209, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 209, + 231 + ], + "type": "text", + "content": "the most relevant local/global frames:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 240, + 277, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 240, + 277, + 255 + ], + "spans": [ + { + "bbox": [ + 74, + 240, + 277, + 255 + ], + "type": "interline_equation", + "content": "S ^ {l o c} = F ^ {l o c} \\cdot I _ {t} \\in \\mathbb {R} ^ {X}, S ^ {g l o} = F ^ {g l o} \\cdot I _ {t} \\in \\mathbb {R} ^ {Y},", + "image_path": "53927a4643c5e129fb81c5d6d4e30ae80a58d39233ae5b09c264532536891af4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "N = c_{L} \\times h_{L} \\times w_{L}" + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "text", + "content": ". Next, we normalize the similarity scores to obtain the probability distribution " + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "D^{loc}" + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "inline_equation", + "content": "D^{glo}" + }, + { + "bbox": [ + 55, + 266, + 296, + 304 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 312, + 285, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 312, + 285, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 285, + 348 + ], + "type": "interline_equation", + "content": "D ^ {l o c} = \\frac {e ^ {S ^ {l o c}}}{\\sum_ {j} e ^ {S _ {j} ^ {l o c}}} \\in \\mathbb {R} ^ {X}, D ^ {g l o} = \\frac {e ^ {S ^ {g l o}}}{\\sum_ {j} e ^ {S _ {j} ^ {g l o}}} \\in \\mathbb {R} ^ {Y}.", + "image_path": "870ce9c9582f2127b4bc892a260cbf85bca5f20089b92a16a06b9af3d1c96679.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "text", + "content": "We then sample " + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "text", + "content": " local frames and " + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "inline_equation", + "content": "y = z - x" + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "text", + "content": " global frames based on " + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "inline_equation", + "content": "D^{loc}" + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "inline_equation", + "content": "D^{glo}" + }, + { + "bbox": [ + 55, + 357, + 296, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 391, + 282, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 391, + 282, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 391, + 282, + 407 + ], + "type": "interline_equation", + "content": "i _ {k} ^ {l o c} \\sim P (i _ {k} ^ {l o c} = i) = D ^ {l o c} (i), \\quad k \\in \\{1, 2, \\ldots , X \\},", + "image_path": "cf33eb3c4bba4cc77c2dd9b1c4000e700885e4448dae11b525935f41318d48ce.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 418, + 282, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 418, + 282, + 435 + ], + "spans": [ + { + "bbox": [ + 69, + 418, + 282, + 435 + ], + "type": "interline_equation", + "content": "i _ {k} ^ {g l o} \\sim P (i _ {k} ^ {g l o} = i) = D ^ {g l o} (i), \\quad k \\in \\{1, 2, \\ldots , Y \\},", + "image_path": "5bc26c74200d83508c104630a7dd77e37d79419b123a73112092fe687a698a15.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "inline_equation", + "content": "i_k^{loc}" + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "inline_equation", + "content": "i_k^{glo}" + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "content": " denote the indices of the selected local and global frames, respectively, and " + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "inline_equation", + "content": "D^{loc}(i)" + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "inline_equation", + "content": "D^{glo}(i)" + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "content": " are the corresponding probabilities for each frame. The selected frames are subsequently used to enhance the segmentation of the current frame through memory attention, resulting in memory-associated features " + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "inline_equation", + "content": "I_t'" + }, + { + "bbox": [ + 55, + 443, + 296, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 526, + 209, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 209, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 209, + 539 + ], + "type": "text", + "content": "3.4. Memory Prompt Generator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": "SAM's mask decoder segments images using image features and user prompts to achieve intention-followed results. Motivated by its prompt integration capability, we propose incorporating memory as an additional prompt. This enables the model to leverage segmentation information from previous frames, enhancing temporal consistency across the video sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Figure 5 illustrates the proposed Memory Prompt Generator (MPG). We introduce a set of learnable memory prompt tokens " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{g \\times d}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " is the number of memory prompts and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " is the embedding dimension. These tokens interact with the flattened memory features " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{mem}} \\in \\mathbb{R}^{(mh_Lw_L) \\times c_L}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " is the number of memory frames, guided by the corresponding binarized segmentation masks" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 221, + 549, + 293 + ], + "blocks": [ + { + "bbox": [ + 321, + 221, + 549, + 293 + ], + "lines": [ + { + "bbox": [ + 321, + 221, + 549, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 221, + 549, + 293 + ], + "type": "image", + "image_path": "ef25dd9d326ff9b7bbca6d6489a16bd817eddf0871266fd60879856f56c63604.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 300, + 548, + 312 + ], + "lines": [ + { + "bbox": [ + 320, + 300, + 548, + 312 + ], + "spans": [ + { + "bbox": [ + 320, + 300, + 548, + 312 + ], + "type": "text", + "content": "Figure 5. Illustration of the memory prompt generator (MPG)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 326, + 518, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 518, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 518, + 338 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{mem}} \\in \\mathbb{R}^{(mh_Lw_L) \\times 1}" + }, + { + "bbox": [ + 314, + 326, + 518, + 338 + ], + "type": "text", + "content": " via masked cross-attention:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 352, + 347, + 555, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 347, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 352, + 347, + 555, + 361 + ], + "type": "interline_equation", + "content": "G ^ {\\prime} = \\operatorname {S o f t m a x} (\\mathcal {M} + Q K ^ {\\top}) V + G, \\tag {1}", + "image_path": "dcd8650e3d09f80b14f3489b2bc9520cb942a9895dbeb8acbe304697e2674361.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 335, + 363, + 555, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 363, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 335, + 363, + 555, + 376 + ], + "type": "interline_equation", + "content": "Q = \\psi (G), \\quad K = \\psi \\left(F _ {\\mathrm {m e m}}\\right), \\quad V = \\psi \\left(F _ {\\mathrm {m e m}}\\right), \\tag {2}", + "image_path": "0de7e9a357f72f733948d9d9aaa7675674c45d38c74db3ccf9ddb86582132e68.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 380, + 554, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 380, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 328, + 380, + 554, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (x, y) = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} M _ {\\mathrm {m e m}} (x, y) = 1, \\\\ - \\infty , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {3}", + "image_path": "1384443142e6b8a54b92cb4ec02e6e2d116f6788289c4311853ffa1bb34e89c4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "type": "text", + "content": " is the learnable linear transformation. The updated memory prompt tokens " + }, + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "type": "inline_equation", + "content": "G^{\\prime}" + }, + { + "bbox": [ + 313, + 419, + 555, + 468 + ], + "type": "text", + "content": ", having captured foreground context from memories, are then refined through self-attention to incorporate intra-token dependencies:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 392, + 477, + 476, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 477, + 476, + 491 + ], + "spans": [ + { + "bbox": [ + 392, + 477, + 476, + 491 + ], + "type": "interline_equation", + "content": "G ^ {\\prime \\prime} = \\operatorname {S e l f A t t n} \\left(G ^ {\\prime}\\right).", + "image_path": "0102d8cd932c26e9f1e5683f88611b61cf3cbd703dd132010f922346b202883e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 500, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 555, + 537 + ], + "type": "text", + "content": "Finally, the refined memory prompt embeddings " + }, + { + "bbox": [ + 313, + 500, + 555, + 537 + ], + "type": "inline_equation", + "content": "G''" + }, + { + "bbox": [ + 313, + 500, + 555, + 537 + ], + "type": "text", + "content": " are passed through a multi-layer perceptron (MLP) to produce the final updated memory prompt tokens:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 396, + 545, + 471, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 545, + 471, + 559 + ], + "spans": [ + { + "bbox": [ + 396, + 545, + 471, + 559 + ], + "type": "interline_equation", + "content": "G ^ {\\prime \\prime \\prime} = \\operatorname {M L P} \\left(G ^ {\\prime \\prime}\\right).", + "image_path": "5a95a592776b9e6e192bdc0ca6c350e83c63349e56fdba4b2921c46cddfcd7b2.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 569, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 605 + ], + "type": "text", + "content": "Such an iterative attention-based token generation enables the MPG to selectively capture target object information from memories, providing guidance for the mask decoder." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 636, + 430, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 430, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 430, + 649 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "Training Datasets. We train our SAM-I2V jointly on 10,000 SA-1B [21] images and 50,583 SAV [36] videos. The SA-1B dataset [21] is a comprehensive collection consisting of over 1.1 billion high-quality segmentation masks from 11 million licensed and privacy-preserving images." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3421" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "This dataset was assembled using a multi-phase data engine where initial stages involved human annotation aided by interactive tools, and subsequent phases transitioned to fully automated mask generation through the SAM model. The rich variety of mask annotations allows for robust training that supports zero-shot generalization and diverse segmentation tasks, proving effective in many downstream applications. The Segment Anything Video (SA-V) dataset [36] complements the image-focused SA-1B by providing extensive coverage for video segmentation. Its training set includes 50,583 videos and 642,036 masklets. SA-V is tailored for video scenarios involving complex object motions, occlusions, and reappearance, making it invaluable for advancing video-based models like SAM 2 [36]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": "Training Configurations. Following [36], to simulate an interactive training setting, we sample 8-frame sequences, designating the first frame as the prompt frame and randomly selecting up to 1 additional frame from the remaining seven frames for corrective clicks. The initial prompt for the first frame can be the ground-truth mask (50% probability), a positive click from the ground-truth mask (25%), or a bounding box input (25%). The corrective clicks are sampled from the error region between the model prediction and the ground-truth mask (90%) or directly from the ground-truth mask (10%). Our SAM-I2V model is implemented using PyTorch [31] and trained for 10 epochs on 8 RTX A5000 GPUs (24GB memory). We employ the AdamW optimizer [24] and apply layer-wise decay [13] specifically to the image encoder. The initial learning rates are set to " + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": " for the image encoder and " + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-5}" + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": " for the other components, respectively. These learning rates are subsequently decayed using a cosine scheduler. We supervise the model's predictions using a linear combination of mask prediction loss for the output mask, meanabsolute-error (MAE) loss for the IoU prediction, and cross-entropy loss for object prediction with a ratio of 20:1:1, respectively. The mask prediction loss " + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": " is defined as the combination of the weighted binary cross-entropy (BCE) loss " + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\ell_{wbce}" + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": " [42] and the weighted IoU loss " + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\ell_{wiou}" + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": " [42], i.e., " + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m = \\ell_{wbce} + \\ell_{wiou}" + }, + { + "bbox": [ + 56, + 243, + 296, + 626 + ], + "type": "text", + "content": ", to ensure the model pays more attention to potential distraction regions such as object boundaries, elongated areas, or holes. Notably, all SAM-I2V modules except the TinySAM [40] backbone are randomly initialized and trained from scratch without leveraging any weights from SAM 2 [36], ensuring complete independence from SAM 2 [36] in the training process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Evaluation Settings. To comprehensively assess the capabilities of our SAM-I2V, we perform comparisons on four benchmark datasets: ESD [18] is a high-quality 3D spatial and temporal dataset for object segmentation in an indoor cluttered environment; PUMaVOS [1] contains videos with segments around object parts such as a person's cheek; LV-VIS [41] contains videos from a diverse set of open-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": "vocabulary object categories; and SAV-Test [36] is the test set of SA-V dataset, consist of 150 videos and 278 masklets. For these evaluations, we report the standard " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "J \\& F" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " metric [32] which combines region similarity (Jaccard index) and boundary accuracy (F-measure). We conducted evaluation on the promptable video segmentation. In this task, the model works in an interactive manner, where the point prompts are used to identify the target of interest or to refine the current segmentation result. Following [36], two settings are included in the comparison, namely offline evaluation (PVS-offline) and online evaluation (PVS-online). In both settings, the target is firstly initialized via " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{click}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " point prompts. As for PVS-offline, the model segments the target throughout the entire video in " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{pass}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " passes, where the interactive frame is selected based on the largest model error (i.e., with lowest IoU between the prediction and ground truth masks) and " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{click}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " points are added on such frame to refine the mask in each pass. All the prompts are used to guide the segmentation process in the next pass. PVS-online follows one pass evaluation protocol, where the segmentation process is paused and refined when the result is not reliable (IoU " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "< 0.75" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " with ground-truth). Also, " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{click}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " points are added on the refined frame to generate corrected mask, which is used for subsequent frames. The refine process is repeated " + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{frame}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 382, + 492, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 492, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 492, + 395 + ], + "type": "text", + "content": "4.2. Comparison to Existing Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 400, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 519 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed SAM-I2V, we compared it against seven state-of-the-art methods on both online and offline PVS tasks under different click settings, as reported in Table 1 and 2. Besides SAM 2.1 [36], we also include upgradation methods that combine SAM (both base [21] and tiny [40] variants) with existing state-of-the-art video object segmentation models, specifically Xmem++ [1] and Cutie [11], as well as approaches that upgrade SAM by incorporating temporal adaptation (TA) [36] and spatial adaptation (SA) [4] modules." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": "Performance Comparison with SAM 2.1. While SAM 2.1 [36] serves as a strong baseline with an overall accuracy of 71.9 (online) and 73.3 (offline), it requires substantial computational resources (2.2 million " + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "inline_equation", + "content": "G \\cdot" + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": " Hour). In contrast, our SAM-I2V achieves its " + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": " (i.e., 65.3, online) and " + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": " (i.e., 66.0, offline) overall scores, with its " + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": " (i.e., 4.6k " + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "inline_equation", + "content": "G \\cdot" + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": " Hour) training cost and its " + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": " (i.e., 18.9m vs. 38.9m) model parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Effectiveness over Existing SAM Upgradation Methods. Compared to SAM [21] + XMem++ [1] (157.0M parameters) and TinySAM [40] + XMem++ (72.3M parameters), our SAM-I2V offers a substantial reduction in model size (18.9M) while delivering comparable or superior performance. For example, under the online setting, SAM-I2V helps achieve an overall score of 65.3, outperforming SAM + XMem++ (60.4) and SAM + Cutie [11] (61.7)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3422" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 72, + 555, + 225 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 555, + 225 + ], + "type": "table", + "html": "
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.645.859.366.386.866.879.575.288.870.182.877.279.760.974.972.971.9
SAM [21] + XMem++ [1]157.0-78.349.368.042.478.846.273.041.979.345.478.443.378.847.073.142.560.4
SAM [21] + Cutie [11]129.8-78.438.369.451.279.044.373.951.879.844.676.952.979.142.473.452.061.7
TinySAM [40] + XMem++ [1]72.3-79.355.268.547.382.959.874.948.784.260.476.951.682.158.573.449.265.8
TinySAM [40] + Cutie [11]45.1-80.653.071.158.483.352.676.259.984.456.278.863.182.853.975.460.568.1
TinySAM [40] + TA [36]17.84.6k68.728.856.242.980.245.172.349.182.149.876.550.977.041.268.447.658.6
TinySAM [40] + SA [4] + TA [36]18.74.6k54.834.652.844.179.352.772.950.681.959.476.851.972.048.967.548.959.3
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.039.153.850.584.760.873.458.286.465.978.359.581.455.368.556.165.3
", + "image_path": "f084450992f4ef6f84dd30bcd61238e1eadbc57e153a7a57f6d11067760f19fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 279, + 555, + 403 + ], + "blocks": [ + { + "bbox": [ + 55, + 233, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 268 + ], + "type": "text", + "content": "Table 1. Comparisons of the online promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 279, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 56, + 279, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 555, + 403 + ], + "type": "table", + "html": "
MethodsPara. (M)Cost1-click3-click5-clickAverageOA
ESPULVSVESPULVSVESPULVSVESPULVSV
SAM 2.1 [36]38.92.2m63.946.261.366.887.767.383.976.290.170.787.578.180.661.477.673.773.3
SAM [21] + XMem++ [1]157.0-49.536.850.141.164.049.353.744.165.354.755.244.959.646.953.043.450.7
TinySAM [40] + XMem++ [1]72.3-47.746.047.549.173.655.452.955.575.962.553.758.265.754.651.454.356.5
TinySAM [40] + TA [36]17.84.6k68.929.558.345.381.149.376.352.482.955.681.154.977.644.871.950.961.3
TinySAM [40] + SA [4] + TA [36]18.74.6k57.037.054.944.880.956.676.652.583.462.181.253.673.851.970.950.361.7
TinySAM [40] + SAM-I2V (Ours)18.94.6k73.437.855.748.984.661.277.858.586.465.183.059.881.554.772.255.766.0
", + "image_path": "61a879d80f89d8df6d03d86781ebf118d7efd728020b3e8c72b723e3f69bc888.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 555, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 555, + 458 + ], + "type": "text", + "content": "Table 2. Comparisons of the offline promptable video segmentation with 1-click, 3-click, and 5-click settings on four benchmark datasets (i.e., ES [18], PU [1], LV [41], and SV [36]). “-” indicates directly combining existing pre-trained models for inference. “Cost” is calculated by GPU number × GPU memory × training hours. “OA” denotes the overall segmentation performance. Cutie [11] is excluded from the comparison as it only supports single-frame mask inputs, rendering it unsuitable for offline PVS settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 478, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 295, + 562 + ], + "type": "text", + "content": "Robust Performance Across Interaction Levels. The online and offline results presented in Table 1 and 2 reflect that the performance of all methods generally improves as the number of interactions increases. Our SAM-I2V consistently delivers superiority across 1-click, 3-click, and 5-click settings, underscoring its robustness to varying levels of user interaction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 662 + ], + "type": "text", + "content": "Robust Online and Offline Performance. SAM-I2V's overall online performance in Table 1 is slightly lower than the offline performance in Table 2. This discrepancy arises because, unlike in the offline evaluation, in the online setting the new prompts affect only the frames following the current paused frame and not the preceding frames. Compared to other methods, our approach maintains an advantage in both online and offline settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "The above results confirm that our SAM-I2V not only matches but often surpasses existing upgrading methods in performance with reduced computational costs, making SAM-I2V a practical choice for PVS tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 478, + 408, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 408, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 408, + 491 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 496, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 532 + ], + "type": "text", + "content": "We conduct an extensive ablation study to validate the extensibility of SAM-I2V and the effectiveness of its each key component. Table 3, 4, 5, 6 summarize our findings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 628 + ], + "type": "text", + "content": "Applicability of SAM-I2V to Upgrade Different SAM Models. In the first ablation study, we investigate the performance of our SAM-I2V to upgrade different SAM models. From the results presented in Table 3, it is evident that our SAM-I2V effectively upgrades different SAM variants to a PVS model, demonstrate that SAM-I2V is a robust and versatile image-to-video promptable segmentation upgradation approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Effectiveness of the Temporal Feature Integrator. As shown in Table 4, our proposed TFI module outperforms the baseline (TinySAM [40] + Temporal Associator [36]) and other adaptation methods. While previous approaches like ST-Adapter [29] and Parallel Adapter [4] perform adaptation within each attention block, our TFI is specifically designed with temporal branch and integration branch to iter" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3423" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 292, + 129 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 292, + 129 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 292, + 129 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 292, + 129 + ], + "type": "table", + "html": "
MethodsES[18]PU[1]LV[41]SV[36]OA
TinySAM [40] + TA [36]80.245.172.349.161.7
TinySAM [40] + SAM-I2V84.760.873.458.269.3
EdgeSAM [57] + TA [36]78.341.469.644.658.5
EdgeSAM [57] + SAM-I2V83.251.971.654.465.1
", + "image_path": "f32aac6d113b2db29f8dac568123a743ddde1f760e7617f31e7edf2b1191573e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 172, + 292, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "content": "Table 3. Ablation on the applicability of SAM-I2V to upgrade different SAM models for PVS task under the online, 3-click setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 172, + 292, + 232 + ], + "lines": [ + { + "bbox": [ + 59, + 172, + 292, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 172, + 292, + 232 + ], + "type": "table", + "html": "
MethodsES[18]PU[1]LV[41]SV[36]OA
Base80.245.172.349.161.7
Base + ST-Adapter [29]79.354.173.350.164.2
Base + Parallel Adapter [4]79.352.772.950.663.9
Base + TFI81.358.172.452.566.1
", + "image_path": "72b2954c1c2b0c9f6880c4abd46ae18bb8a8418a7fc189c03ad68f4cc554c417.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 281, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 317 + ], + "type": "text", + "content": "atively extract and integrate spatial-temporal features. This allows our model to capture richer temporal dependencies, leading to superior performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": "Effectiveness of the Memory Selective Associator. Relying on the previous " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "z = 6" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " consecutive frames for segmentation is equivalent to setting " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " in MSA to 0 and 6, respectively. By varying the number of selected global " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "(x)" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " and local " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "(y)" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " frames in our MSA, we observe that the segmentation performance improves when the MSA selectively retrieves the most relevant frames based on their similarity to the current frame. Specifically, the configuration with " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "x = 3" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " global frames and " + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "inline_equation", + "content": "y = 3" + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": " local frames performs better, highlighting the importance of incorporating both short-term and long-term memory. In addition, the ability of MSA to select frames based on relevance rather than uniform sampling allows the model to focus on the most relevant past frames, leading to improved performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "text", + "content": "Effectiveness of the Memory Prompt Generator. From Table 6, we can observe that, (i) introducing MPG " + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "inline_equation", + "content": "(w/" + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "text", + "content": " MPG-1/3/5) consistently improves performance over the one without MPG, demonstrating the effectiveness of the memory-as-prompt strategy in enhancing segmentation accuracy; (ii) varying the number of memory prompt tokens among 1, 3, and 5 yields comparable results, indicating that the MPG is robust to the choice of token quantity within this range; and (iii) removing the Masked Cross-Attention (MCA) mechanism (i.e., " + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "text", + "content": " MPG-3 (" + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 55, + 485, + 295, + 653 + ], + "type": "text", + "content": " MCA)) leads to a performance drop, highlighting the importance of MCA in focusing on relevant foreground information, eliminating background interference, and thus enhancing the quality of memory prompt generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 213, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 213, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 213, + 672 + ], + "type": "text", + "content": "4.4. Limitation and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "One limitation of our approach is that its accuracy does not well match that of SAM 2.1 [36]. Note that the released SAM 2.1 model was trained not only on the SAM-1B" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 72, + 552, + 140 + ], + "blocks": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "text", + "content": "Table 4. Ablation on the effectiveness of the temporal feature integrator TFI for online PVS task under the 3-click setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 72, + 552, + 140 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 552, + 140 + ], + "type": "table", + "html": "
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MSA (i.e., x=0, y=6)84.054.473.256.967.1
w/ MSA (x=1, y=5)84.754.873.457.967.7
w/ MSA (x=2, y=4)84.556.373.456.967.8
w/ MSA (x=3, y=3)84.760.873.458.269.3
w/ MSA (x=3, y=3) w/ UFS84.156.073.056.567.4
", + "image_path": "bd2b304d6882f461978c210308c6e852a37ac68018d502f6170ee95af47a21f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 205, + 550, + 275 + ], + "blocks": [ + { + "bbox": [ + 313, + 148, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 195 + ], + "type": "text", + "content": "Table 5. Ablation on the effectiveness of the memory selective associator MSA for online PVS task under the 3-click setting. \"x\" and \"y\" denote the number of selected global and local frames, respectively. \"UFS\" indicates the uniform frame sampling strategy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 205, + 550, + 275 + ], + "lines": [ + { + "bbox": [ + 319, + 205, + 550, + 275 + ], + "spans": [ + { + "bbox": [ + 319, + 205, + 550, + 275 + ], + "type": "table", + "html": "
TinySAM [40] + SAM-I2VES[18]PU[1]LV[41]SV[36]OA
w/o MPG82.157.772.153.266.3
w/ MPG-184.058.473.055.667.8
w/ MPG-384.760.873.458.269.3
w/ MPG-584.558.673.657.168.5
w/ MPG-3 (w/o MCA)84.958.172.855.167.7
", + "image_path": "575cb2d12fb303d1edd990969327f5332ac2fb487d4d299dc67a74212a0ff6ed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "type": "text", + "content": "Table 6. Ablation on the effectiveness of the memory prompt generator MPG for online PVS task under the 3-click setting. \"MPG-g\" denotes the MPG with " + }, + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "type": "text", + "content": " learnable memory prompt tokens. \"MCA\" indicates the masked cross attention." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 348, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 456 + ], + "type": "text", + "content": "and SAV datasets but also on 62.9k proprietary, non-public videos, which our method was unable to utilize. Despite this, our method demonstrates superior performance compared to baseline approaches and creates targeted model design opportunities for downstream applications in resource-constrained academic scenarios. Future work will focus on enhancing our model by incorporating additional publicly available datasets and exploring advanced training techniques to bridge the accuracy gap with SAM 2.1 [36]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "content": "In this work, we introduced SAM-I2V, a training-efficient framework that upgrades the pre-trained Segment Anything Model (SAM) to support promptable video segmentation (PVS). SAM-I2V consists of three key innovations: an image-to-video feature extraction upgrader that extends SAM's static image encoder for spatiotemporal video perception; a memory filtering strategy that selects the most relevant past frames to enhance the utilization of historical information; and a memory-as-prompt mechanism that leverages object memory for temporally consistent mask propagation in dynamic scenes. Comprehensive experiments demonstrate that SAM-I2V achieved over " + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "content": " of SAM 2's performance at only " + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 486, + 555, + 677 + ], + "type": "text", + "content": " of its training cost. This remarkable training efficiency significantly lowers the barriers for developing diverse PVS models, facilitating rapid innovation and broader practical applications." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. This research is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award No: MOE-T2EP20124-0012)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3424" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Maksym Bekuzarov, Ariana Bermudez, Joon-Young Lee, and Hao Li. Xmem++: Production-level video segmentation from few annotated frames. In ICCV, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 170 + ], + "type": "text", + "content": "[2] Sergi Caelles, Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. One-shot video object segmentation. In CVPR, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 172, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 295, + 227 + ], + "type": "text", + "content": "[3] Keyan Chen, Chenyang Liu, Hao Chen, Haotian Zhang, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsprompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 229, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 295, + 272 + ], + "type": "text", + "content": "[4] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, 2022. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 329 + ], + "type": "text", + "content": "[5] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv:2408.04579, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 294, + 363 + ], + "type": "text", + "content": "[6] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. " + }, + { + "bbox": [ + 62, + 332, + 294, + 363 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 62, + 332, + 294, + 363 + ], + "type": "text", + "content": " data makes segment anything slim. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 367, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 294, + 399 + ], + "type": "text", + "content": "[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 294, + 434 + ], + "type": "text", + "content": "[8] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Modular interactive video object segmentation: Interaction-to-mask, propagation and difference-aware fusion. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 294, + 468 + ], + "type": "text", + "content": "[9] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 472, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 472, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 294, + 503 + ], + "type": "text", + "content": "[10] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 506, + 294, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 506, + 294, + 538 + ], + "spans": [ + { + "bbox": [ + 57, + 506, + 294, + 538 + ], + "type": "text", + "content": "[11] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Joon-Young Lee, and Alexander Schwing. Putting the object back into video object segmentation. In CVPR, 2024. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 541, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 541, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 57, + 541, + 294, + 573 + ], + "type": "text", + "content": "[12] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv:2305.06558, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 576, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 576, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 576, + 294, + 597 + ], + "type": "text", + "content": "[13] K Clark. Electra: Pre-training text encoders as discriminators rather than generators. arXiv:2003.10555, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "type": "text", + "content": "[14] Thanos Delatolas, Vicky Kalogeiton, and Dim P Papadopoulos. Learning the what and how of annotation in video object segmentation. In WACV, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 677 + ], + "type": "text", + "content": "[15] Haoyu Dong, Hanxue Gu, Yaqian Chen, Jichen Yang, and Maciej A Mazurowski. Segment anything model 2: an application to 2d and 3d medical images. arXiv:2408.00756, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Yuk Heo, Yeong Jun Koh, and Chang-Su Kim. Interactive video object segmentation using global and local transfer modules. In ECCV, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[18] Xiaoqian Huang, Kachole Sanket, Abdulla Ayyad, Fariborz Baghaei Naeini, Dimitrios Makris, and Yahya Zweiri. A neuromorphic dataset for object segmentation in indoor cluttered environment. arXiv:2302.06301, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 154, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 186 + ], + "type": "text", + "content": "[19] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 189, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 220 + ], + "type": "text", + "content": "[20] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. NeurIPS, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 269, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 311 + ], + "type": "text", + "content": "[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 314, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 356 + ], + "type": "text", + "content": "[23] Haofeng Liu, Erli Zhang, Junde Wu, Mingxuan Hong, and Yueming Jin. Surgical sam 2: Real-time segment anything in surgical video by efficient frame pruning. arXiv:2408.07931, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 360, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 380 + ], + "type": "text", + "content": "[24] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "text", + "content": "[25] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 553, + 460 + ], + "type": "text", + "content": "[26] K-K Maninis, Sergi Caelles, Yuhua Chen, Jordi Pont-Tuset, Laura Leal-Taixe, Daniel Cremers, and Luc Van Gool. Video object segmentation without temporal information. IEEE TPAMI, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 464, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 506 + ], + "type": "text", + "content": "[27] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "type": "text", + "content": "[28] Seoung Wug Oh, Joon-Young Lee, Kalyan Sunkavalli, and Seon Joo Kim. Fast video object segmentation by reference-guided mask propagation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "text", + "content": "[29] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. NeurIPS, 2022. 3, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "text", + "content": "[30] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dualpath adaptation from image to video transformers. In CVPR, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "text", + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 3, 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3425" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[33] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In ICCV, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "text", + "content": "[35] Frano Rajic, Lei Ke, Yu-Wing Tai, Chi-Keung Tang, Martin Danelljan, and Fisher Yu. Segment anything meets point tracking. arXiv:2307.01197, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 297 + ], + "type": "text", + "content": "[37] Simiao Ren, Francesco Luzi, Saad Lahrichi, Kaleb Kassaw, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Segment anything, from space? In WACV, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 341 + ], + "type": "text", + "content": "[38] Andreas Robinson, Felix Jaremo Lawin, Martin Danelljan, Fahad Shahbaz Khan, and Michael Felsberg. Learning fast and robust target models for video object segmentation. In CVPR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "text", + "content": "[39] Chuyun Shen, Wenhao Li, Yuhang Shi, and Xiangfeng Wang. Interactive 3d medical image segmentation with sam 2. arXiv:2408.02635, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[40] Han Shu, Wenshuo Li, Yehui Tang, Yiman Zhang, Yihao Chen, Houqiang Li, Yunhe Wang, and Xinghao Chen. Tinysam: Pushing the envelope for efficient segment anything model. arXiv:2312.13789, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[41] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 498 + ], + "type": "text", + "content": "[42] Jun Wei, Shuhui Wang, and Qingming Huang. F3net: Fusion, feedback and focus for salient object detection. In AAAI, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[43] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv:2304.12620, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[44] Qiangqiang Wu, Tianyu Yang, Wei Wu, and Antoni B Chan. Scalable video object segmentation with simplified framework. In ICCV, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[45] Aoran Xiao, Weihao Xuan, Heli Qi, Yun Xing, Ruijie Ren, Xiaojin Zhang, Ling Shao, and Shijian Lu. Cat-sam: Conditional tuning for few-shot adaptation of segment anything model. In ECCV, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "text", + "content": "[46] Xinyu Xiong, Zihuang Wu, Shangyi Tan, Wenxue Li, Feilong Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation. arXiv:2408.08870, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[47] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 510 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "masked image pretraining for efficient segment anything. In CVPR, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[48] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In ECCV, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[49] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv:2304.11968, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[50] Linjie Yang, Yanran Wang, Xuehan Xiong, Jianchao Yang, and Aggelos Katsaggellos. Efficient video object segmentation via network modulation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "text", + "content": "[51] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. AIM: Adapting image models for efficient video action recognition. In ICLR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[52] Zongxin Yang, Yunchao Wei, and Yi Yang. Collaborative video object segmentation by foreground-background integration. In ECCV, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[53] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv:2306.14289, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[54] Chaoning Zhang, Dongshen Han, Sheng Zheng, Jinwoo Choi, Tae-Ho Kim, and Choong Seon Hong. Mobilesamv2: Faster segment anything to everything. arXiv:2312.09579, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[55] Jiaming Zhang, Yutao Cui, Gangshan Wu, and Limin Wang. Joint modeling of feature, correspondence, and a compressed memory for video object segmentation. arXiv:2308.13505, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[56] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv:2306.12156, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 477 + ], + "type": "text", + "content": "[57] Chong Zhou, Xiangtai Li, Chen Change Loy, and Bo Dai. Edgesam: Prompt-in-the-loop distillation for on-device deployment of sam. arXiv:2312.06660, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 479, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 510 + ], + "type": "text", + "content": "[58] Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2. arXiv:2408.00874, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3426" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_content_list.json b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..160d06b026ba4ebbe668b46d589b16e71af3a6e8 --- /dev/null +++ b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_content_list.json @@ -0,0 +1,1807 @@ +[ + { + "type": "text", + "text": "SAM-REF: Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model", + "text_level": 1, + "bbox": [ + 107, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chongkai $\\mathrm{Yu}^{1,\\dagger}$ , Ting Liu $^{1,\\dagger,\\boxtimes}$ , Anqi Li $^{2}$ , Xiaochao Qu $^{1}$ , Chengjing Wu $^{1}$ , Luoqi Liu $^{1}$ , Xiaolin Hu $^{3,\\boxtimes}$", + "bbox": [ + 106, + 200, + 892, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MT Lab, Meitu Inc., Beijing 100083, China", + "bbox": [ + 364, + 223, + 633, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Beijing Institute of Technology", + "bbox": [ + 398, + 239, + 596, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Department of Computer Science and Technology", + "bbox": [ + 344, + 258, + 651, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "BNRist, IDG/McGovern Institute for Brain Research", + "bbox": [ + 339, + 277, + 656, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University, Beijing 100084, China", + "bbox": [ + 367, + 295, + 630, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yck, lt, qxc, ethan, llq5}@meitu.com", + "bbox": [ + 336, + 311, + 661, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xIhu@tsinghua.edu.cn", + "bbox": [ + 408, + 330, + 586, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†Joint first authors. ♕Joint corresponding author.", + "bbox": [ + 339, + 345, + 651, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 373, + 326, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interactive segmentation is to segment the mask of the target object according to the user's interactive prompts. There are two mainstream strategies: early fusion and late fusion. Current specialist models utilize the early fusion strategy that encodes the combination of images and prompts to target the prompted objects, yet repetitive complex computations on the images result in high latency. Late fusion models extract image embeddings once and merge them with the prompts in later interactions. This strategy avoids redundant image feature extraction and improves efficiency significantly. A recent milestone is the Segment Anything Model (SAM). However, this strategy limits the models' ability to extract detailed information from the prompted target zone. To address this issue, we propose SAM-REF, a two-stage refinement framework that fully integrates images and prompts by using a lightweight refiner into the interaction of late fusion, which combines the accuracy of early fusion and maintains the efficiency of late fusion. Through extensive experiments, we show that our SAM-REF model outperforms the current state-of-the-art method in most metrics on segmentation quality without compromising efficiency.", + "bbox": [ + 89, + 406, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interactive segmentation [4, 12, 13, 15, 22] is an important task in computer vision that has garnered significant attention from both academia and industry. It leverages user input through simple interactions, such as scribbles, boxes, and clicks, to conveniently obtain accurate masks. Based on the timing of prompt-image fusion, existing methods can be categorized into early fusion strategies and late fusion strategies, as shown in Fig. 1.", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ec9a8b6a4b94d9f3237422b8e0b5b11e234b6cd2da12ab2a17d7fe18d7ab65f.jpg", + "image_caption": [ + "(a) Early fusion" + ], + "image_footnote": [], + "bbox": [ + 540, + 375, + 895, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b6e05b855797dd35fa3958322f0ad7a15d5b1274a61ddbd32a8713289bcfb9e.jpg", + "image_caption": [ + "(b) Late fusion" + ], + "image_footnote": [], + "bbox": [ + 540, + 481, + 897, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a727f9eb6c01befdbe92b6c005875ac7e748ab68457061952a5834829641c2bb.jpg", + "image_caption": [ + "(c) Our method", + "Figure 1. Three pipelines for interactive image segmentation. (a) Early fusion pipeline. (b) Late fusion pipeline. (c) Our proposed pipeline, which takes advantages of both early fusion and late fusion pipelines. The orange arrows indicate the interaction between the user and the systems." + ], + "image_footnote": [], + "bbox": [ + 540, + 578, + 895, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early fusion methods such as FocalClick [4], and SimpleClick [21] jointly encode the image and prompts during interaction (Fig. 1a). Since the encoder can generate features specifically for target objects, these early fusion pipelines can achieve accurate segmentation results, espe", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19356", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cially for complex and challenging cases like thin objects. However, encoding the image in each interaction leads to low efficiency. In contrast, late fusion methods like the Segment Anything Model (SAM) [15] and InterFormer [12] extract features of the image before interaction, and fuse the prompt with the features during each interaction (Fig. 1b). These methods often utilize a powerful image encoder, Vision Transformer (ViT) [5], to generate embeddings with rich information. SAM [15] also used a billion-scale large dataset for training to enhance the encoder's extraction capability. Late fusion strategy avoids redundant image feature extraction, improving training and inference efficiency significantly. However, despite the ability of a powerful image encoder to focus on all objects, it struggles to capture the fine details of the target objects, because it discards the interaction between the original image and the prompt. This always leads to unsatisfactory results in many complex scenarios. Subsequent finetune approaches [13, 14, 29] based on SAM are also limited by the image embeddings.", + "bbox": [ + 89, + 90, + 480, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a method called SAM-REF which takes advantage of both early fusion and late fusion pipelines. The key idea is to utilize the late fusion pipeline for its high efficiency but integrate the prompt and the input image during the user's interaction with the system, which provides precise instructions of the user and fine details of the image. To achieve this goal, we design a lightweight refiner and place it after SAM (Fig. 1c). The overall process is as follows. First, SAM generates the image embeddings with strong semantic representation capability and the prompt embeddings, and then combines them into the decoder to generate the coarse mask. Then the lightweight refiner uses a coarse-to-fine structure that can utilize global and local information. It obtains the semantic information of the image from SAM's embeddings, and processes the details by reintroducing image-prompt synergy. The light refiner generates a refined mask as the final output.", + "bbox": [ + 89, + 378, + 482, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimentally, SAM-REF fully exploits the potential of SAM for interactive segmentation. On widely used benchmark datasets (GrabCut [25], Berkeley [23], DAVIS [24], and SBD [8]) and high-quality benchmark dataset HQSeg-44K [14], our method achieves state-of-the-art performance in most metrics while maintaining SAM's low latency advantage.", + "bbox": [ + 89, + 635, + 482, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose SAM-REF, a new pipeline that leverages the power of SAM but integrates the prompt and the image during the interaction process to capture object details, leading to both high accuracy and high efficiency.", + "bbox": [ + 89, + 739, + 482, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 814, + 232, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on interactive segmentation. There are mainly two methods, early fusion and late fusion interactive segmentation method, divided based on the timing of prompt-image fusion.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Early Fusion Methods", + "text_level": 1, + "bbox": [ + 513, + 90, + 720, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early fusion interactive segmentation [1] means combining different information in the early stages of networks. It combines the image with interactive information before processing them through deeper network layers, e.g. ViT [5], which is capable of producing more refined and contextually aware segmentation results. The first deep-learning-based interactive segmentation model, DiOS [30], takes the early fusion strategy by embedding positive and negative clicks into distance maps and concatenating them with the original image as input. RITM [27] introduces concatenating the previous mask with point maps, and later works mostly follow this approach. Many methods have been used to improve accuracy. For example, FocalClick [4] and FocusCut [19] use local refinement to optimize the details. CDNet [3] proposes a pixel diffusion module to fully fuse the prompt and the image. SimpleClick [21] up-samples the image feature, and PseudoClick [20] simulates an additional prompt. Though high precision, encoding the image and prompts every interaction leads to the high latency of early fusion, which has always been a problem. Some work has tried to reduce latency. FocalClick [4] compresses the size of the feature to speed up, but this method loses some accuracy.", + "bbox": [ + 511, + 113, + 906, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Late Fusion Methods", + "text_level": 1, + "bbox": [ + 513, + 472, + 714, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Late fusion strategy extracts image embeddings once and combines them with interactive prompts in later interactions. Interformer [12] and SAM [15] decouple the light prompt encoder and the decoder from the heavy encoder, accelerating the speed. However, late fusion cannot catch details according to the focusless embeddings, leading to low precision. To promote accuracy, SegNext [22] adds dense representation and fusion of visual prompts to extract more visual information, FocSAM [13] introduces window-based attention to focus on the target zone of embeddings, while HQ-SAM [14] uses global-local fusion and proposed HQ-SAM dataset with high quality. These methods aim to explore more information from the embeddings. However, the embeddings may not have enough detailed information about the target zone, so these methods cannot achieve significant improvements in accuracy. In contrast, our method combines early fusion and late fusion, extracting detailed information from the image directly.", + "bbox": [ + 511, + 494, + 906, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 782, + 604, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since our proposed method is built upon SAM, we first briefly review the architecture of SAM. Then we describe our method SAM-REF, which has three major components, Global Fusion Refiner (GFR), Local Fusion Refiner (LFR), and a Dynamic Selector (DS). Finally, we introduce the training loss.", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19357", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52679d1be08940b73dce2643e6dcc5ba886fd28a022a6574148e32fc29051b24.jpg", + "image_caption": [ + "(a) Pipeline" + ], + "image_footnote": [], + "bbox": [ + 147, + 92, + 883, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41c7a046a775ee7f56f6877742a0f824168f9ae61d36b2862ef61ab33cbd54cc.jpg", + "image_caption": [ + "(b) GFR" + ], + "image_footnote": [], + "bbox": [ + 127, + 303, + 848, + 439 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c422981d3509b9a9c0bd00e536a6a15e2a03de8edf23474d163804bbaed8702c.jpg", + "image_caption": [ + "(c) LFR", + "Figure 2. The framework of the proposed SAM-REF. We propose a decoupled lightweight refiner while maintaining the inherent structure of SAM [15] (a). The refiner mainly has three modules, the Global Fusion Refiner (GFR), the Local Fusion Refiner (LFR), and the Dynamic Selector (DS). In GFR (b) and LFR (c), Conv $\\times n$ represents an n-layer convolutional network." + ], + "image_footnote": [ + ": Channel-wise concatenation", + "$\\oplus$ : Element-wise summation", + "$\\odot$ : Element-wise multiplication", + "@: Sigmoid" + ], + "bbox": [ + 178, + 468, + 844, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: SAM", + "text_level": 1, + "bbox": [ + 89, + 705, + 282, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM consists of three modules: an image encode, a prompt encode, and a mask decode. Image encoder: It is a heavy ViT-based backbone, structured into four stages with equal depth. Each stage employs window-based attention to enhance computational efficiency [16]. Prompt encoder: It is a lightweight structure consisting of MLP and CNN. In the interaction phase, it transforms the positional information of the input boxes, clicks, and masks into embeddings. However, the sparse representations of visual prompts, similar to those used in linguistic prompts, may inhibit the ability to capture image details [22]. Mask decoder: It is a", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lightweight structure composed of a Transformer [28] decoder and a convolutional mask head. It combines image embeddings and prompt embeddings through a two-layer transformer-based decoder to predict the final masks.", + "bbox": [ + 511, + 705, + 906, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the interaction process, the prompt encoder and mask decoder are lightweight, and the heavy image encoder only needs to be performed once before the interaction. Therefore, the entire inference pipeline is very efficient.", + "bbox": [ + 511, + 768, + 906, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Our Pipeline of the SAM-REF", + "text_level": 1, + "bbox": [ + 511, + 845, + 782, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To preserve the inherent structure of SAM while fully exploiting its potential, we adopt a decoupled design approach", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19358", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "wherever possible. To this end, we introduce two modules GFR and LFR to refine the mask progressively.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall architecture of SAM-REF is shown in Fig. 2. First, we use a late fusion approach to combine the encoded image with the prompt, producing a coarse mask, just like SAM [15]. Then we use the lightweight refiner, a decoupled module from SAM, to refine the coarse mask in a coarse-to-fine manner. In addition to the coarse mask and the prompt, the image and the feature from the encoder are also input to the lightweight refiner. Inside the lightweight refiner, we first use the global fusion refiner (GFR) to directly combine the prompt with the image and extract detailed information about the prompted target with a lightweight convolutional network. Meanwhile, it progressively incorporates SAM's image embeddings at each stage to preserve SAM's robust deep semantic capabilities, thus refining the overall objects.", + "bbox": [ + 88, + 121, + 482, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further refine the local details of the target region, we propose the local fusion refiner (LFR). It crops the target region of the mask from GFR according to the prompt, refines the cropped mask, and pastes it back to the mask. We propose a dynamic selector to adaptively determine which mask to paste on. The cropped mask can be pasted on the predicted result of GFR or the previous mask as the output. The former can re-predict the full mask, while the latter can only refine the target local region without modifying other regions.", + "bbox": [ + 88, + 333, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Global Fusion Refiner (GFR)", + "text_level": 1, + "bbox": [ + 89, + 494, + 352, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Fusion of Image, Prompt, Coarse Mask and Encoder Features", + "text_level": 1, + "bbox": [ + 89, + 517, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As stated in Sec. 1, we need to efficiently fuse the prompt and the image. Unlike SAM [15] that represents the prompts with sparse one-dimension vectors, we use a dense presentation of prompts (disks with a small radius) to match the spatial dimensions of the image [2, 4, 22, 27], which better preserving the detailed spatial attributes of the visual prompts. We encode the prompts and the coarse mask $M_{t}^{S}$ ( $S$ refers to SAM, and $t$ refers to the $t$ -th interaction) to generate a dense map $D$ . The dense map is the concatenation of the positive prompt map, the negative prompt map, and the coarse mask $M_{t}^{S}$ . We adopt the method from RITM [27], utilizing two convolutions to align the scale and channels of the image and the dense map from $I_{ori} \\in \\mathbb{R}^{3 \\times 1024 \\times 1024}$ , $D_{ori} \\in \\mathbb{R}^{3 \\times 256 \\times 256}$ to $I \\in \\mathbb{R}^{64 \\times 256 \\times 256}$ , $D \\in \\mathbb{R}^{64 \\times 256 \\times 256}$ .", + "bbox": [ + 89, + 551, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assume that SAM, relying on the powerful ViT architecture, has already learned deep semantic information [11]. The primary deficiency lies in its understanding of low-level details. To maintain SAM's semantic understanding and design an efficient detail extraction module, we employ a scale-invariant convolutional structure and integrate it with SAM image embeddings $\\pmb{F}_{Enc} \\in \\mathbb{R}^{256 \\times 64 \\times 64}$ .", + "bbox": [ + 89, + 779, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as illustrated in Fig. 2b, we progressively", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "upscale the SAM image embeddings $\\mathbf{F}_{Enc}$ to $\\hat{\\mathbf{F}}_{Enc} \\in \\mathbb{R}^{64 \\times 256 \\times 256}$ . Then we fuse the image, prompts, coarse mask, and encoder features by adding them together:", + "bbox": [ + 511, + 89, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} (0) = \\boldsymbol {I} + \\boldsymbol {D} + \\hat {\\boldsymbol {F}} _ {E n c}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 148, + 906, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $D$ contains the information of prompts and the coarse mask. Subsequently, we utilize a scale-invariant ResBasicBlock [9] to capture fine details. This process is repeated $N$ times, with each iteration merging with the SAM image embeddings to prevent the loss of semantic information from SAM:", + "bbox": [ + 511, + 181, + 906, + 271 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} (i + 1) = \\operatorname {R e s B l o c k} (\\boldsymbol {F} (i)) + \\hat {\\boldsymbol {F}} _ {E n c}, i \\in [ 0, N - 1 ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 285, + 906, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\pmb{F}(N)$ contains the information of image, prompt, coarse mask, and encoder features.", + "bbox": [ + 511, + 316, + 905, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Fusion of Coarse Mask and Refined Mask", + "text_level": 1, + "bbox": [ + 511, + 359, + 852, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the coarse mask $M_{t}^{S}$ is the prediction of SAM, it is generally accurate. We want to identify the inaccurate areas of $M_{t}^{S}$ , and only modify these areas. GFR predicts a mask result $\\bar{M}_{r}^{G}$ and an error map $E^{G}$ , which represents error-prone regions. We replace these error-prone regions in $M_{t}^{S}$ with the corresponding regions in $M_{r}^{G}$ . The fusion process is formulated as follows.", + "bbox": [ + 511, + 378, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {r} ^ {G} = \\sigma (\\boldsymbol {E} ^ {G}) \\odot \\bar {\\boldsymbol {M}} _ {r} ^ {G} + (1 - \\sigma (\\boldsymbol {E} ^ {G})) \\odot \\boldsymbol {M} _ {t} ^ {S}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 497, + 906, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma(\\cdot)$ represents Sigmoid function. $E^G$ is supervised by the error regions between the ground truth and the binary mask. After GFR, we obtain the global refinement of the objects.", + "bbox": [ + 511, + 529, + 905, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Local Fusion Refiner (LFR)", + "text_level": 1, + "bbox": [ + 511, + 604, + 764, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further refine the local details of the target region, we propose LFR to refine the local region in $M_r^G$ . In this module, we perform prompt and image fusion again for the user-interacted local regions to refine the local details of the objects. Inspired by FocalClick [4], we calculate the max connected component of the difference between the GFR result and the previous mask. Subsequently, we generate the external box for the connected component that includes the last prompt and proportionally expand it. The specific design of the LFR is illustrated in Fig. 2c.", + "bbox": [ + 511, + 627, + 905, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to FocalClick, we use RoiAlign [10] to crop the logits and the fused feature $\\pmb{F}(N)$ from GFR. Similar to GFR, our two prediction heads generate the local mask result $\\bar{M}_r^L$ and the error map $\\pmb{E}^L$ based on the prediction of GFR $M_r^G$ . The local refined mask $M_r^L$ fuses $\\bar{M}_r^L$ and $M_r^G$ as follows:", + "bbox": [ + 511, + 779, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {r} ^ {L} = \\sigma (\\boldsymbol {E} ^ {L}) \\odot \\bar {\\boldsymbol {M}} _ {r} ^ {L} + (1 - \\sigma (\\boldsymbol {E} ^ {L})) \\odot \\boldsymbol {M} _ {r} ^ {G}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19359", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Dynamic Selector", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As stated in FocalClick [4], when the previous mask $M_{t-1}$ is already well-refined, especially when it has been refined for many rounds, re-predicting the full mask may destroy areas that have already been well refined. FocalClick [4] always pastes the refined local mask onto the previous mask after 10 interactions, as by this point the previous mask typically converges to good segmentation quality. In this paper, we use the dynamic selector to select which mask should the local result $M_r^L$ be pasted onto, the global refined mask $M_r^G$ or the previous mask $M_{t-1}$ . Taking the global result $M_r^G$ pasted with $M_r^L$ as the output can refine the local area based on the refined full mask, which is suitable when re-predicting the full mask will lead to less error. And when re-predicting the full mask leads to more error, which often happens when $M_{t-1}$ is already precise enough, we paste $M_r^L$ onto $M_{t-1}$ to only refine the target local area and avoid overwriting other areas.", + "bbox": [ + 89, + 112, + 483, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adaptively decide which mask to paste onto by comparing the changes in error maps $\\pmb{E}^{G}$ between each iteration of GFR. We use the error ratio $R_{e}$ , formulated as:", + "bbox": [ + 89, + 369, + 482, + 414 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {e} = \\left\\| H \\left(\\boldsymbol {E} _ {t} ^ {G}\\right) \\right\\| _ {0} / \\left\\| H \\left(\\boldsymbol {E} _ {t - 1} ^ {G}\\right) \\right\\| _ {0}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 421, + 482, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$H(\\cdot)$ returns 1 when the value is greater than 0, else returns 0. If the error ratio $R_{e}$ is larger than a threshold $\\theta$ , it indicates that the modification on the full mask will lead to more error, and we choose to paste $M_r^L$ onto $M_{t - 1}$ as the output $M_t$ . Otherwise, paste it onto $M_r^G$ as $M_t$ . The output $M_t$ will become a previous mask $M_{t - 1}$ in next interaction. In our experiment, the threshold $\\theta$ is set to be 1.1.", + "bbox": [ + 89, + 446, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Training supervision", + "text_level": 1, + "bbox": [ + 89, + 560, + 285, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we adopted Normalized Focal Loss $L_{nfl}$ proposed in RITM [27] to finetune the SAM decoder. Then, we used Dice Loss $L_{dice}$ and Binary Cross Entropy Loss $L_{bce}$ to supervise the error heads of GFR and LFR. $L_{wnfl}$ denotes weighted NFL loss for the refined predictions. We calculate the error regions between the ground truth and the binary mask (for GFR it is $M_t^S$ , and for LFR it is $M_r^G$ ). We calculate the NFL loss and add weight (1.5 in our experiment) for these error regions to get $L_{wnfl}$ .", + "bbox": [ + 89, + 582, + 483, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {L} = \\boldsymbol {L} _ {n f l} + \\boldsymbol {L} _ {d i c e} ^ {G} + \\boldsymbol {L} _ {b c e} ^ {G} + \\boldsymbol {L} _ {w n f l} ^ {G} \\tag {6} \\\\ + \\boldsymbol {L} _ {d i c e} ^ {L} + \\boldsymbol {L} _ {b c e} ^ {L} + \\boldsymbol {L} _ {w n f l} ^ {L} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 724, + 482, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the superscripts $G$ and $L$ indicate the supervision for GFR and LFR respectively.", + "bbox": [ + 89, + 772, + 482, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 815, + 220, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a series of experiments to test the performance of our model. We contrast our method with previous works. The experimental settings and the results of the experiments are as follows.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 513, + 90, + 714, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Dataset.", + "text_level": 1, + "bbox": [ + 511, + 112, + 616, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model on 2 different dataset configurations. For one configuration we train our model on COCO [18] + LVIS [6] datasets, following previous works [4, 12, 13, 21, 27]. For another configuration, we test the performance of our method on high-precision datasets. We train our model on HQSeg-44K [14, 22] datasets. HQSeg-44K dataset is a high-quality and well-annotated dataset with $44K$ extremely accurate image mask annotations. We evaluate the models' zero-shot interactive segmentation capabilities on GrabCut [25], Berkeley [23], DAVIS [24], SBD [8], and HQSeg-44K validation set, aligning with previous studies.", + "bbox": [ + 511, + 131, + 906, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Implementation details.", + "text_level": 1, + "bbox": [ + 511, + 303, + 723, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the pre-trained ViT-H and ViT-B from SAM [15] as the backbone with the prompt encoder and decoder. The GlobalFusion Refiner and the LocalFusion Refiner are light-weighted convolutional neural networks, as shown in Fig. 2. We use $N = 3$ ResBasicBlocks to achieve the best results. Training and evaluations are performed on a server with 4 NVIDIA Tesla V100-PCIE-32GB GPUs and Intel(R) Xeon(R) Gold 6278C CPU.", + "bbox": [ + 511, + 321, + 905, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Training strategy.", + "text_level": 1, + "bbox": [ + 511, + 449, + 683, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the training process, we adopt the two-stage training strategy of FocSAM [13] that firstly fine-tune the SAM decoder for 320k iterations at a batch size of 4 with frozen image encoder and prompt encoder, and then train SAM-REF with the frozen decoder and frozen encoders in the same settings for 80k iterations. We adopt InterFormer's click simulation strategy [12] for interactive simulation before loss computation. Moreover, we use the image encoder to preextract and store the image embeddings of the datasets to reduce computational costs. This strategy greatly improves the training speed.", + "bbox": [ + 511, + 468, + 906, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.4. Evaluation.", + "text_level": 1, + "bbox": [ + 511, + 640, + 638, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report the results on metrics of mean Intersection over Union (mIoU), number of clicks (NoC), number of failure cases (NoF), and the Seconds Per Click (SPC) metrics, following previous work [4, 13, 15, 21, 22]. And we test the latency of Segmentation Anything Task (SAT Latency) following SegNext [22]. mIoU measures the average intersection over union (IoU) given a fixed number of consecutive interactions. We report the mIoU on 5 clicks. NoC measures the number of clicks required to achieve a target IoU. We report NoC90 and NoC95 to measure the segmentation ability. NoC90 is a commonly used evaluation metric. With the increasing demand for high-quality segmentation, we provide the stricter NoC95 to demonstrate the capability to handle the details of our method. NoF measures the number of failure cases that cannot achieve the target IoU in 20 clicks. We report the failure cases on NoF95. SPC is a", + "bbox": [ + 511, + 657, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19360", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5714866e8131f353c1f39c503719408bca69fd7a5171a4c7f2a645cc54f96f9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain dataBackboneSPC/s↓Grabcut [25]Berkeley [23]DAVIS [24]SBD [8]
NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓
Early Fusion
F-BRS [26]CVPR20COCO+LVISResNet-2.72-4.57-7.41-7.73-
CDNet [3]ICCV21COCO+LVISResNet-1.52-2.06-7.04-5.56-
EdgeFlow [7]ICCVW21COCO+LVISHRNet18-1.72-2.40---5.77-
RITM [27]ICIP22COCO+LVISHRNet32-1.562.482.105.415.3411.525.7112.00
PseudoClick [20]ECCV22COCO+LVISHRNet32-1.50-2.08-5.11-5.54-
FocalClick [4]CVPR22COCO+LVISSegFB3-S23840.2321.681.921.714.214.9010.405.5911.91
AdaptiveClick [17]TNNLS2024COCO+LVISViT-H1.691.38-1.64-4.60-4.68-
SimpleClick [21]ICCV23COCO+LVISViT-H8.241.501.661.754.344.788.884.7010.76
Late Fusion
InterFormer [12]ICCV23COCO+LVISViT-B0.1531.50-3.14-6.19-6.34-
InterFormer [12]ICCV23COCO+LVISViT-L0.2711.36-2.53-5.21-5.51-
SAM [15]ICCV23SA-1BViT-H0.4131.882.282.095.145.1910.007.6215.03
HQ-SAM [14]NeurIPS24HQSeg-44KViT-H*0.4431.862.262.144.945.069.657.9115.08
FocSAM [13]CVPR24COCO+LVISViT-H*0.4671.442.601.503.444.7610.775.0711.67
SAM-REFHQSeg-44KViT-H*0.5111.762.221.984.574.789.386.8112.18
SAM-REFCOCO+LVISViT-B*0.2781.402.441.583.734.7510.455.5611.76
SAM-REFCOCO+LVISViT-H*0.5111.362.201.433.184.549.104.4410.41
", + "bbox": [ + 91, + 88, + 950, + 358 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2cb587d6592d7fd5bdc87736b882c841b9580eb231e67692e0de8a205b109bd1.jpg", + "table_caption": [ + "Table 1. Comparison of SPC, NoC90 and NoC95 with previous methods. We report the results on the four mainstream datasets. The superscript * indicates frozen backbone." + ], + "table_footnote": [], + "table_body": "
MethodTrain dataBackboneLatency(s)↓5-mIoU↑NoC90↓NoC95↓NoF95↓
FocalClick [4]CVPR22COCO+LVISSegFB3-S225630.8584.638.1212.63835
FocalClick [4]CVPR22COCO+LVISSegFB3-S238440.1485.457.0310.74649
SimpleClick [21]ICCV23COCO+LVISViT-B88.585.117.4712.39797
InterFormer [12]ICCV23COCO+LVISViT-B30.582.627.1710.77658
SegNext [22]CVPR24COCO+LVISViT-B22.185.717.1811.52700
SAM [15]ICCV23SA-1BViT-B4.086.167.4612.42811
SAM [15]ICCV23SA-1BViT-H4.2188.06.5010.53653
HQ-SAM [14]NeurIPS24HQSeg-44KViT-B*4.789.456.4910.79671
FocSAM [13]CVPR24COCO+LVISViT-H*18.688.65.749.44580
SAM-REFHQSeg-44KViT-B*5.190.755.9010.08607
SAM-REFCOCO+LVISViT-B*5.188.56.199.84603
SAM-REFCOCO+LVISViT-H*5.2289.65.449.16566
", + "bbox": [ + 132, + 410, + 867, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on high-quality datasets. All of the models are evaluated on the HQSeg-44K validation set [14]. SAM-REF achieves better results than SAM-based and other methods. The superscript * indicates frozen backbone.", + "bbox": [ + 89, + 612, + 906, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "commonly used metric to measure the segmentation speed, which represents the time consumed of each click. SPC is tested on CPU. Following SegNext [22], SAT Latency measures the latency for the Segment Anything Task (SAT). We fix the input image size at $1024 \\times 1024$ and prompt the model with a grid of $16 \\times 16$ points. We compute the total time required to process all points by inputting one point at a time with our previous mask. SAT Latency is tested on a single GPU. Compared to SPC, SAT Latency calculates the overall latency but not one-prompt latency, which favors late fusion methods where the image is only encoded once.", + "bbox": [ + 88, + 667, + 485, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 666, + 653, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on widely used benchmarks. In Tab. 1, we present the main results of SAM-REF, benchmarked against mainstream methods. Our method is based on SAM, where we freeze the SAM encoder and train only a few remaining parameters to mitigate bias inherent. We compare our model with previous early fusion and late fusion methods. As reported, SAM-REF achieves state-of-the-art performance on the NoC90 in all mainstream evaluation datasets. In addition, when applying the stricter criteria evaluation standard NoC95, our method is slightly weaker than SimpleClick only on the DAVIS and Grabcut datasets. In the remaining datasets, SAM-REF achieves state-of-the-art performance in both NOC90 and NOC95 metrics, us", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19361", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acab688e6e97f5af9fd54a592a23b1d836fd4ac2789f5c17912d0ce15ea40ca9.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 114, + 102, + 287, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f25d27ec3707083c2d66620a4e1c3a71388aafaf0e63bd70f4e279b425a6461c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 103, + 424, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b509f082a2c2c90c6260448ec9a905047eedf220a1233017e6b066743498cd3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 176, + 424, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea794b6c9c977c3fff44665c611f35f12825ae804a17681059c1225b662e98d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 103, + 560, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6cb6b63cdebd7c68cac9e965fd2a6f73a7a5988aabedad6e0c93decc585a684.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 176, + 558, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b82bf923370b7e448d05962b093e8527b85a588d264da59a71f2598726127007.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 103, + 694, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e5c9d9402330593fcaa9694cd98b254e4578532d5182fe13fb14e2f475272f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 176, + 694, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0877529792eead23cc0daf2ddbf1c68c9bc47b68cf6427478b1c4696478eeae8.jpg", + "image_caption": [ + "SAM" + ], + "image_footnote": [], + "bbox": [ + 710, + 103, + 828, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a487c2c4681487141f4b51ee07634b51e347047cf1e00c0cb2920aeaf4c2e426.jpg", + "image_caption": [ + "FocSAM" + ], + "image_footnote": [], + "bbox": [ + 710, + 176, + 828, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3fb9e5fb07c0f22f23feb1a35428e21270d72947d97b32312fa69b9747765fe.jpg", + "image_caption": [ + "GroundTruth", + "Figure 3. Qualitative analysis on a challenge Example. The far-left image presents a challenging example featuring the image and ground truth (GT) with blue masks. The top, middle, and bottom rows on the right respectively display the segmentation results of SAM, FocSAM, and our method, at the 5th, 10th, 15th, and 20th clicks." + ], + "image_footnote": [], + "bbox": [ + 114, + 222, + 285, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47704268c36e4f69999936d72cc627545350e458c2fb4f34f7b95b0468055761.jpg", + "image_caption": [ + "5st click" + ], + "image_footnote": [], + "bbox": [ + 303, + 250, + 424, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a50e4f75c861028fa567763c2a29f7c2b99e20a559fa258fbceb27dffe264010.jpg", + "image_caption": [ + "$10^{\\mathrm{st}}$ click" + ], + "image_footnote": [], + "bbox": [ + 439, + 250, + 560, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f82255e2ddeac8fba554051311aa15effc76c94fab6c6ea464e69e9e062a736.jpg", + "image_caption": [ + "$15^{\\mathrm{st}}$ click" + ], + "image_footnote": [], + "bbox": [ + 575, + 250, + 694, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c3c6fec5c4e0bb1267f2988f54ee3cfaeed765c0ef32500e513a9d26de50709.jpg", + "image_caption": [ + "$20^{\\mathrm{st}}$ click", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 710, + 250, + 828, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing ViT-H as the backbone. However, as the early fusion method, SimpleClick requires a heavy image encoding at each interaction, whereas our approach is much more time-efficient, with only $6.2\\%$ of SimpleClick's SPC. Additionally, we provide ViT-B results, showing it matches the speed (0.278 SPC) and accuracy (4.75 NoC90 on DAVIS) of the fastest early fusion method FocalClick [4] (0.232 SPC and 4.90 NoC90 on DAVIS).", + "bbox": [ + 88, + 433, + 483, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By reintroducing prompt-image synergy on late fusion, our method achieves state-of-the-art accuracy compared to all late fusion methods on these datasets. Specifically, under the same settings, we also outperform other SAM-based methods, such as FocSAM [13] and HQ-SAM [14]. We only add 0.1s SPC compared to SAM, which still maintain the high efficiency of late fusion.", + "bbox": [ + 88, + 561, + 483, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on high-quality benchmark. In Tab. 2, we compare our method with state-of-the-art approaches on high-quality dataset HQSeg-44K [14](1537 samples). We compare two types of baselines: SAM-based methods and other methods. As reported, SAM-REF achieves state-ofthe-art performance to the two baselines, which demonstrate that SAM-REF can effectively enhances SAM's performance to top-tier levels in high-quality interactive segmentation while maintaining the low latency. Other latefusion methods (InterFormer [12] and SegNext [22]), which rely solely on the dense presentation of visual prompts, are more costly than SAM's use of the sparse representation [22]. Our method combines the benefits of both prompt types. It only focuses on detail extraction using dense maps of visual prompts, without needing complex fusion of visual", + "bbox": [ + 89, + 672, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prompts and image embeddings. This demonstrates that our method can effectively enhance the performance of late fusion in complex scenarios, as evidenced by Qualitative Results.", + "bbox": [ + 511, + 433, + 906, + 492 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8fd3d304029826c67977c4bbee8a81be0cd71db3049d0ac287100baf1958a093.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParams/MBFPSMem/G
SAM [15]635.61.703.7
SAM-REF636.31.673.7
", + "bbox": [ + 566, + 505, + 854, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Computation analysis for SAM and SAM-REF", + "bbox": [ + 534, + 566, + 883, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computation analysis. In Tab 3, we report the comparison of model parameters (Params), inference time per image (FPS), and GPU memory (Mem) usage between ViT-H based on SAM and SAM-REF. While SAM-REF produces substantially better segmentation quality, our method adds just 0.7MB to the model parameters compared to SAM, with negligible increases in GPU memory usage and inference time per image.", + "bbox": [ + 511, + 602, + 906, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Experiments", + "text_level": 1, + "bbox": [ + 511, + 738, + 723, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct plenty of ablation studies, reporting experimental results for ViT-H models trained on COCO+LVIS and tested on DAVIS.", + "bbox": [ + 511, + 763, + 905, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Module Analysis. To demonstrate the effect of our model's modules, i.e., GFR, LFR, and the Dynamic Selector, we perform an ablation study on whether and which mask to paste $M_r^L$ . We have set up 4 experiments. One directly outputs $M_r^G$ , one always pastes $M_r^L$ onto $M_r^G$ , one always pastes $M_r^L$ onto $M_{t-1}$ , and one", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19362", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f70f52c0b1d852954b1f6f4f4d0bb8baa965249102bd5081272002e63bf8317.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNoC90NoC95
1no \\( M_r^L \\) only \\( M_r^G \\)4.669.53
2paste \\( M_r^L \\) to \\( M_{t-1} \\)4.809.75
3paste \\( M_r^L \\) to \\( M_r^G \\)4.609.40
4Dynamic Selector4.549.10
", + "bbox": [ + 133, + 88, + 442, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pastes $M_r^L$ according to the Dynamic Selector. For each experiment, we test NoC90 and NoC95 to evaluate the accuracy. We have shown the results in Tab. 4. The first experiment has no local refinement and always modifies the full mask, which leads to mutual influence between interactions and cannot achieve high accuracy. The second experiment only refines a local region but not the full mask in each interaction, which makes it modify slowly and needs more interactions to achieve the same accuracy. The third experiment with local refinement has a higher accuracy based on 1-th, but still cannot solve the mutual influence. The fourth experiment pasting $M_r^L$ with Dynamic Selector avoids the interference and refines local regions, achieving the best accuracy.", + "bbox": [ + 88, + 239, + 483, + 454 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bd7295180ba865b73c3ba4fd22c8c9371d3346123285914a4e90bd09887a6312.jpg", + "table_caption": [ + "Table 4. Ablation study for the refiners and the Dynamic Se-. lector. The methods indicate which mask to paste $M_r^L$ onto." + ], + "table_footnote": [], + "table_body": "
MethodSPCNoC90NoC95
ResBlock×10.4834.579.94
ResBlock×30.5114.549.10
ResBlock×50.5454.669.88
ResBlock×80.7104.6110.61
", + "bbox": [ + 150, + 465, + 424, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study for the number of ResBlocks.", + "bbox": [ + 117, + 556, + 454, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of ResBlocks. Tab. 5 shows the impact of the number of ResBlocks on accuracy and speed. With three ResBlocks, our method achieves the best results. Fewer blocks fail to capture details effectively, while more blocks reduce speed and harm the semantic capabilities of SAM image embeddings, thus decreasing accuracy. This experiment demonstrates the rationale for setting the number of ResBlocks to 3.", + "bbox": [ + 89, + 588, + 483, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Qualitative Result", + "text_level": 1, + "bbox": [ + 89, + 724, + 267, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 3, we show the qualitative results compared to SAM [15] and FocSAM [13] on a challenge Example. This visualization demonstrates SAM-REF's enhanced precision. Our method achieves the highest mIoU in all clicks. Both FocSAM and our method are trained on $\\mathrm{COCO + LVIS}$ and freeze the SAM's image encoder parameters. In the later clicks, our method can segment the detailed shape of the wheel, while SAM and FocSAM cannot. Based on challenging scenarios, the original SAM results are unsatisfactory. While FocSAM enhances stability and prevents pro", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac332cd1beb006877e680b08b9847a0de0b3108e64c1e0f913b47c9a71d68e7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 92, + 614, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e96dac20c4b5f8ba666d37d3bcedd089466fb93feb9b22227972b36ef7acdd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 92, + 710, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce41af0e62a013835694fba5b9e85d22eb461cbdc7a3ee8ee40949c00ca14f0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 92, + 794, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82ef2b81520563c9a83260a09becc5fd15a1139d9b98fa035a1b8e6305839ab4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 92, + 880, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9b9e0ceeff6c8cc946b023e9c63d11d76b0b5152100378a0c3f616c36bb245e.jpg", + "image_caption": [ + "Figure 4. Qualitative results for the effectiveness of GlobalFusion and LocalFusion Refiner. The first row shows the global refinement results of GFR. The red box indicates the region for local refinement. The second row shows the local refinement results of LFR." + ], + "image_footnote": [], + "bbox": [ + 532, + 172, + 614, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6acf7bd555576c2a1b5fca038e6c775e26e46380d08dd887593dbf8389b6ffb.jpg", + "image_caption": [ + "Coarse mask Error map" + ], + "image_footnote": [], + "bbox": [ + 629, + 172, + 710, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f597bf79370772830588684d7015f7f106f982a327e8d0b0be521f0e3c59e9bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 172, + 795, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6acd82e30ebdacc43f7a3a231aca03434d0d857417329cef311efe427c1c0323.jpg", + "image_caption": [ + "Refined mask" + ], + "image_footnote": [], + "bbox": [ + 802, + 172, + 883, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gressive deterioration, it remains limited by image embeddings, potentially missing object details. Our method retrieves lost object details from the original image through image-prompt synergy.", + "bbox": [ + 511, + 368, + 906, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the results at each stage of our method in Fig. 4. The coarse mask is generated by the SAM decoder. The first row demonstrates how GFR refines the global segmentation process. The error map presents the error regions within SAM's predicted areas. In this scenario, the error regions are concentrated around the object edges and thin structures. Then, GFR specifically refines these error regions. The second row demonstrates the refinement process of LFR. It focuses on refining the areas with the largest errors between the interactions. In this image, the errors are mainly concentrated in the hollow regions.", + "bbox": [ + 511, + 431, + 908, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 616, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In view that early fusion strategies in the interactive segmentation can achieve high accuracy and late fusion strategies can achieve high efficiency, we propose a new method called SAM-REF, which combines the advantages of the two strategies. Our method redesigns the SAM-based interactive segmentation pipeline, fusing images and prompts from both global and local perspectives. This adaptation fully taps into SAM's potential for interactive segmentation. SAM-REF outperforms current state-of-the-art methods while maintaining SAM's low latency, even in challenging scenarios.", + "bbox": [ + 511, + 645, + 908, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 830, + 692, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Xiaolin Hu was supported by the National Natural Science Foundation of China (No. U2341228).", + "bbox": [ + 511, + 857, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19363", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] George Barnum, Sabera Talukder, and Yisong Yue. On the benefits of early fusion in multimodal representation learning. arXiv preprint arXiv:2011.07191, 2020. 2", + "[2] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 4", + "[3] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7345-7354, 2021. 2, 6", + "[4] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: Towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022. 1, 2, 4, 5, 6, 7", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[6] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5", + "[7] Yuying Hao, Yi Liu, Zewu Wu, Lin Han, Yizhou Chen, Guowei Chen, Lutao Chu, Shiyu Tang, Zhiliang Yu, Zeyu Chen, et al. Edgeflow: Achieving practical interactive segmentation with edge-guided flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1551-1560, 2021. 6", + "[8] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 2, 5, 6", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4", + "[11] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4", + "[12] You Huang, Hao Yang, Ke Sun, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, and Rongrong Ji. Interformer: Real-time interactive image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22301-22311, 2023. 1, 2, 5, 6, 7", + "[13] You Huang, Zongyu Lan, Liujuan Cao, Xianming Lin, Shengchuan Zhang, Guannan Jiang, and Rongrong Ji. Foc" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sam: Delving deeply into focused objects in segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3120-3130, 2024. 1, 2, 5, 6, 7, 8", + "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7", + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[16] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3", + "[17] Jiacheng Lin, Jiajun Chen, Kailun Yang, Alina Roitberg, Siyu Li, Zhiyong Li, and Shutao Li. Adaptiveclick: Click-aware transformer with adaptive focal loss for interactive image segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6", + "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[19] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. Focuscut: Diving into a focus view in interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2637-2646, 2022. 2", + "[20] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. Pseudoclick: Interactive image segmentation with click imitation. In European Conference on Computer Vision, pages 728-745. Springer, 2022. 2, 6", + "[21] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 1, 2, 5, 6", + "[22] Qin Liu, Jaemin Cho, Mohit Bansal, and Marc Niethammer. Rethinking interactive image segmentation with low latency high quality and diverse prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3782, 2024. 1, 2, 3, 4, 5, 6, 7", + "[23] Kevin McGuinness and Noel E O'connor. A comparative evaluation of interactive segmentation algorithms. Pattern Recognition, 43(2):434-444, 2010. 2, 5, 6", + "[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2, 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19364", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 2, 5, 6", + "[26] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8623-8632, 2020. 6", + "[27] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 2, 4, 5, 6", + "[28] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[29] Zhaozhi Xie, Bochen Guan, Weihao Jiang, Muyang Yi, Yue Ding, Hongtao Lu, and Lei Zhang. Pa-sam: Prompt adapter sam for high-quality image segmentation. arXiv preprint arXiv:2401.13051, 2024. 2", + "[30] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 2" + ], + "bbox": [ + 91, + 92, + 480, + 430 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19365", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_model.json b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4c4b3d22842c2cb02fc3d64caaf4aae12b487bf5 --- /dev/null +++ b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "SAM-REF: Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.202, + 0.893, + 0.223 + ], + "angle": 0, + "content": "Chongkai \\(\\mathrm{Yu}^{1,\\dagger}\\), Ting Liu\\(^{1,\\dagger,\\boxtimes}\\), Anqi Li\\(^{2}\\), Xiaochao Qu\\(^{1}\\), Chengjing Wu\\(^{1}\\), Luoqi Liu\\(^{1}\\), Xiaolin Hu\\(^{3,\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.224, + 0.635, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\) MT Lab, Meitu Inc., Beijing 100083, China" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.241, + 0.598, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.259, + 0.652, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Department of Computer Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.278, + 0.657, + 0.291 + ], + "angle": 0, + "content": "BNRist, IDG/McGovern Institute for Brain Research" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.296, + 0.631, + 0.31 + ], + "angle": 0, + "content": "Tsinghua University, Beijing 100084, China" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.313, + 0.663, + 0.327 + ], + "angle": 0, + "content": "{yck, lt, qxc, ethan, llq5}@meitu.com" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.332, + 0.588, + 0.344 + ], + "angle": 0, + "content": "xIhu@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.347, + 0.652, + 0.363 + ], + "angle": 0, + "content": "†Joint first authors. ♕Joint corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.374, + 0.327, + 0.39 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.485, + 0.725 + ], + "angle": 0, + "content": "Interactive segmentation is to segment the mask of the target object according to the user's interactive prompts. There are two mainstream strategies: early fusion and late fusion. Current specialist models utilize the early fusion strategy that encodes the combination of images and prompts to target the prompted objects, yet repetitive complex computations on the images result in high latency. Late fusion models extract image embeddings once and merge them with the prompts in later interactions. This strategy avoids redundant image feature extraction and improves efficiency significantly. A recent milestone is the Segment Anything Model (SAM). However, this strategy limits the models' ability to extract detailed information from the prompted target zone. To address this issue, we propose SAM-REF, a two-stage refinement framework that fully integrates images and prompts by using a lightweight refiner into the interaction of late fusion, which combines the accuracy of early fusion and maintains the efficiency of late fusion. Through extensive experiments, we show that our SAM-REF model outperforms the current state-of-the-art method in most metrics on segmentation quality without compromising efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.222, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Interactive segmentation [4, 12, 13, 15, 22] is an important task in computer vision that has garnered significant attention from both academia and industry. It leverages user input through simple interactions, such as scribbles, boxes, and clicks, to conveniently obtain accurate masks. Based on the timing of prompt-image fusion, existing methods can be categorized into early fusion strategies and late fusion strategies, as shown in Fig. 1." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.376, + 0.897, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.459, + 0.756, + 0.472 + ], + "angle": 0, + "content": "(a) Early fusion" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.482, + 0.898, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.559, + 0.753, + 0.571 + ], + "angle": 0, + "content": "(b) Late fusion" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.579, + 0.897, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.687, + 0.757, + 0.699 + ], + "angle": 0, + "content": "(c) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.715, + 0.907, + 0.786 + ], + "angle": 0, + "content": "Figure 1. Three pipelines for interactive image segmentation. (a) Early fusion pipeline. (b) Late fusion pipeline. (c) Our proposed pipeline, which takes advantages of both early fusion and late fusion pipelines. The orange arrows indicate the interaction between the user and the systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Early fusion methods such as FocalClick [4], and SimpleClick [21] jointly encode the image and prompts during interaction (Fig. 1a). Since the encoder can generate features specifically for target objects, these early fusion pipelines can achieve accurate segmentation results, espe" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.378 + ], + "angle": 0, + "content": "cially for complex and challenging cases like thin objects. However, encoding the image in each interaction leads to low efficiency. In contrast, late fusion methods like the Segment Anything Model (SAM) [15] and InterFormer [12] extract features of the image before interaction, and fuse the prompt with the features during each interaction (Fig. 1b). These methods often utilize a powerful image encoder, Vision Transformer (ViT) [5], to generate embeddings with rich information. SAM [15] also used a billion-scale large dataset for training to enhance the encoder's extraction capability. Late fusion strategy avoids redundant image feature extraction, improving training and inference efficiency significantly. However, despite the ability of a powerful image encoder to focus on all objects, it struggles to capture the fine details of the target objects, because it discards the interaction between the original image and the prompt. This always leads to unsatisfactory results in many complex scenarios. Subsequent finetune approaches [13, 14, 29] based on SAM are also limited by the image embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.635 + ], + "angle": 0, + "content": "In this paper, we present a method called SAM-REF which takes advantage of both early fusion and late fusion pipelines. The key idea is to utilize the late fusion pipeline for its high efficiency but integrate the prompt and the input image during the user's interaction with the system, which provides precise instructions of the user and fine details of the image. To achieve this goal, we design a lightweight refiner and place it after SAM (Fig. 1c). The overall process is as follows. First, SAM generates the image embeddings with strong semantic representation capability and the prompt embeddings, and then combines them into the decoder to generate the coarse mask. Then the lightweight refiner uses a coarse-to-fine structure that can utilize global and local information. It obtains the semantic information of the image from SAM's embeddings, and processes the details by reintroducing image-prompt synergy. The light refiner generates a refined mask as the final output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Experimentally, SAM-REF fully exploits the potential of SAM for interactive segmentation. On widely used benchmark datasets (GrabCut [25], Berkeley [23], DAVIS [24], and SBD [8]) and high-quality benchmark dataset HQSeg-44K [14], our method achieves state-of-the-art performance in most metrics while maintaining SAM's low latency advantage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.802 + ], + "angle": 0, + "content": "In summary, we propose SAM-REF, a new pipeline that leverages the power of SAM but integrates the prompt and the image during the interaction process to capture object details, leading to both high accuracy and high efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.233, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we focus on interactive segmentation. There are mainly two methods, early fusion and late fusion interactive segmentation method, divided based on the timing of prompt-image fusion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.722, + 0.107 + ], + "angle": 0, + "content": "2.1. Early Fusion Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Early fusion interactive segmentation [1] means combining different information in the early stages of networks. It combines the image with interactive information before processing them through deeper network layers, e.g. ViT [5], which is capable of producing more refined and contextually aware segmentation results. The first deep-learning-based interactive segmentation model, DiOS [30], takes the early fusion strategy by embedding positive and negative clicks into distance maps and concatenating them with the original image as input. RITM [27] introduces concatenating the previous mask with point maps, and later works mostly follow this approach. Many methods have been used to improve accuracy. For example, FocalClick [4] and FocusCut [19] use local refinement to optimize the details. CDNet [3] proposes a pixel diffusion module to fully fuse the prompt and the image. SimpleClick [21] up-samples the image feature, and PseudoClick [20] simulates an additional prompt. Though high precision, encoding the image and prompts every interaction leads to the high latency of early fusion, which has always been a problem. Some work has tried to reduce latency. FocalClick [4] compresses the size of the feature to speed up, but this method loses some accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.473, + 0.715, + 0.488 + ], + "angle": 0, + "content": "2.2. Late Fusion Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.496, + 0.907, + 0.768 + ], + "angle": 0, + "content": "Late fusion strategy extracts image embeddings once and combines them with interactive prompts in later interactions. Interformer [12] and SAM [15] decouple the light prompt encoder and the decoder from the heavy encoder, accelerating the speed. However, late fusion cannot catch details according to the focusless embeddings, leading to low precision. To promote accuracy, SegNext [22] adds dense representation and fusion of visual prompts to extract more visual information, FocSAM [13] introduces window-based attention to focus on the target zone of embeddings, while HQ-SAM [14] uses global-local fusion and proposed HQ-SAM dataset with high quality. These methods aim to explore more information from the embeddings. However, the embeddings may not have enough detailed information about the target zone, so these methods cannot achieve significant improvements in accuracy. In contrast, our method combines early fusion and late fusion, extracting detailed information from the image directly." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.784, + 0.605, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Since our proposed method is built upon SAM, we first briefly review the architecture of SAM. Then we describe our method SAM-REF, which has three major components, Global Fusion Refiner (GFR), Local Fusion Refiner (LFR), and a Dynamic Selector (DS). Finally, we introduce the training loss." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.093, + 0.885, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.467, + 0.277, + 0.536, + 0.291 + ], + "angle": 0, + "content": "(a) Pipeline" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.304, + 0.849, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.442, + 0.522, + 0.454 + ], + "angle": 0, + "content": "(b) GFR" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.469, + 0.846, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.59, + 0.523, + 0.602 + ], + "angle": 0, + "content": "(c) LFR" + }, + { + "type": "image_footnote", + "bbox": [ + 0.153, + 0.604, + 0.332, + 0.616 + ], + "angle": 0, + "content": ": Channel-wise concatenation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.354, + 0.604, + 0.515, + 0.617 + ], + "angle": 0, + "content": "\\(\\oplus\\) : Element-wise summation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.528, + 0.604, + 0.704, + 0.617 + ], + "angle": 0, + "content": "\\(\\odot\\) : Element-wise multiplication" + }, + { + "type": "image_footnote", + "bbox": [ + 0.72, + 0.604, + 0.791, + 0.617 + ], + "angle": 0, + "content": "@: Sigmoid" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.638, + 0.908, + 0.681 + ], + "angle": 0, + "content": "Figure 2. The framework of the proposed SAM-REF. We propose a decoupled lightweight refiner while maintaining the inherent structure of SAM [15] (a). The refiner mainly has three modules, the Global Fusion Refiner (GFR), the Local Fusion Refiner (LFR), and the Dynamic Selector (DS). In GFR (b) and LFR (c), Conv \\(\\times n\\) represents an n-layer convolutional network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.284, + 0.721 + ], + "angle": 0, + "content": "3.1. Preliminaries: SAM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "SAM consists of three modules: an image encode, a prompt encode, and a mask decode. Image encoder: It is a heavy ViT-based backbone, structured into four stages with equal depth. Each stage employs window-based attention to enhance computational efficiency [16]. Prompt encoder: It is a lightweight structure consisting of MLP and CNN. In the interaction phase, it transforms the positional information of the input boxes, clicks, and masks into embeddings. However, the sparse representations of visual prompts, similar to those used in linguistic prompts, may inhibit the ability to capture image details [22]. Mask decoder: It is a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.907, + 0.767 + ], + "angle": 0, + "content": "lightweight structure composed of a Transformer [28] decoder and a convolutional mask head. It combines image embeddings and prompt embeddings through a two-layer transformer-based decoder to predict the final masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.907, + 0.83 + ], + "angle": 0, + "content": "During the interaction process, the prompt encoder and mask decoder are lightweight, and the heavy image encoder only needs to be performed once before the interaction. Therefore, the entire inference pipeline is very efficient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.783, + 0.862 + ], + "angle": 0, + "content": "3.2. Our Pipeline of the SAM-REF" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To preserve the inherent structure of SAM while fully exploiting its potential, we adopt a decoupled design approach" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19358" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "wherever possible. To this end, we introduce two modules GFR and LFR to refine the mask progressively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.483, + 0.333 + ], + "angle": 0, + "content": "The overall architecture of SAM-REF is shown in Fig. 2. First, we use a late fusion approach to combine the encoded image with the prompt, producing a coarse mask, just like SAM [15]. Then we use the lightweight refiner, a decoupled module from SAM, to refine the coarse mask in a coarse-to-fine manner. In addition to the coarse mask and the prompt, the image and the feature from the encoder are also input to the lightweight refiner. Inside the lightweight refiner, we first use the global fusion refiner (GFR) to directly combine the prompt with the image and extract detailed information about the prompted target with a lightweight convolutional network. Meanwhile, it progressively incorporates SAM's image embeddings at each stage to preserve SAM's robust deep semantic capabilities, thus refining the overall objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.484, + 0.485 + ], + "angle": 0, + "content": "To further refine the local details of the target region, we propose the local fusion refiner (LFR). It crops the target region of the mask from GFR according to the prompt, refines the cropped mask, and pastes it back to the mask. We propose a dynamic selector to adaptively determine which mask to paste on. The cropped mask can be pasted on the predicted result of GFR or the previous mask as the output. The former can re-predict the full mask, while the latter can only refine the target local region without modifying other regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.353, + 0.512 + ], + "angle": 0, + "content": "3.3. Global Fusion Refiner (GFR)" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.547 + ], + "angle": 0, + "content": "3.3.1. Fusion of Image, Prompt, Coarse Mask and Encoder Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.779 + ], + "angle": 0, + "content": "As stated in Sec. 1, we need to efficiently fuse the prompt and the image. Unlike SAM [15] that represents the prompts with sparse one-dimension vectors, we use a dense presentation of prompts (disks with a small radius) to match the spatial dimensions of the image [2, 4, 22, 27], which better preserving the detailed spatial attributes of the visual prompts. We encode the prompts and the coarse mask \\( M_{t}^{S} \\) (\\( S \\) refers to SAM, and \\( t \\) refers to the \\( t \\)-th interaction) to generate a dense map \\( D \\). The dense map is the concatenation of the positive prompt map, the negative prompt map, and the coarse mask \\( M_{t}^{S} \\). We adopt the method from RITM [27], utilizing two convolutions to align the scale and channels of the image and the dense map from \\( I_{ori} \\in \\mathbb{R}^{3 \\times 1024 \\times 1024} \\), \\( D_{ori} \\in \\mathbb{R}^{3 \\times 256 \\times 256} \\) to \\( I \\in \\mathbb{R}^{64 \\times 256 \\times 256} \\), \\( D \\in \\mathbb{R}^{64 \\times 256 \\times 256} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.885 + ], + "angle": 0, + "content": "We assume that SAM, relying on the powerful ViT architecture, has already learned deep semantic information [11]. The primary deficiency lies in its understanding of low-level details. To maintain SAM's semantic understanding and design an efficient detail extraction module, we employ a scale-invariant convolutional structure and integrate it with SAM image embeddings \\( \\pmb{F}_{Enc} \\in \\mathbb{R}^{256 \\times 64 \\times 64} \\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Specifically, as illustrated in Fig. 2b, we progressively" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.138 + ], + "angle": 0, + "content": "upscale the SAM image embeddings \\( \\mathbf{F}_{Enc} \\) to \\( \\hat{\\mathbf{F}}_{Enc} \\in \\mathbb{R}^{64 \\times 256 \\times 256} \\). Then we fuse the image, prompts, coarse mask, and encoder features by adding them together:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.15, + 0.907, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} (0) = \\boldsymbol {I} + \\boldsymbol {D} + \\hat {\\boldsymbol {F}} _ {E n c}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.907, + 0.272 + ], + "angle": 0, + "content": "Here \\( D \\) contains the information of prompts and the coarse mask. Subsequently, we utilize a scale-invariant ResBasicBlock [9] to capture fine details. This process is repeated \\( N \\) times, with each iteration merging with the SAM image embeddings to prevent the loss of semantic information from SAM:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.286, + 0.907, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} (i + 1) = \\operatorname {R e s B l o c k} (\\boldsymbol {F} (i)) + \\hat {\\boldsymbol {F}} _ {E n c}, i \\in [ 0, N - 1 ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.906, + 0.348 + ], + "angle": 0, + "content": "\\(\\pmb{F}(N)\\) contains the information of image, prompt, coarse mask, and encoder features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.853, + 0.374 + ], + "angle": 0, + "content": "3.3.2. Fusion of Coarse Mask and Refined Mask" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Since the coarse mask \\( M_{t}^{S} \\) is the prediction of SAM, it is generally accurate. We want to identify the inaccurate areas of \\( M_{t}^{S} \\), and only modify these areas. GFR predicts a mask result \\( \\bar{M}_{r}^{G} \\) and an error map \\( E^{G} \\), which represents error-prone regions. We replace these error-prone regions in \\( M_{t}^{S} \\) with the corresponding regions in \\( M_{r}^{G} \\). The fusion process is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.498, + 0.907, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {r} ^ {G} = \\sigma (\\boldsymbol {E} ^ {G}) \\odot \\bar {\\boldsymbol {M}} _ {r} ^ {G} + (1 - \\sigma (\\boldsymbol {E} ^ {G})) \\odot \\boldsymbol {M} _ {t} ^ {S}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.592 + ], + "angle": 0, + "content": "where \\(\\sigma(\\cdot)\\) represents Sigmoid function. \\(E^G\\) is supervised by the error regions between the ground truth and the binary mask. After GFR, we obtain the global refinement of the objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.765, + 0.62 + ], + "angle": 0, + "content": "3.4. Local Fusion Refiner (LFR)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.779 + ], + "angle": 0, + "content": "To further refine the local details of the target region, we propose LFR to refine the local region in \\( M_r^G \\). In this module, we perform prompt and image fusion again for the user-interacted local regions to refine the local details of the objects. Inspired by FocalClick [4], we calculate the max connected component of the difference between the GFR result and the previous mask. Subsequently, we generate the external box for the connected component that includes the last prompt and proportionally expand it. The specific design of the LFR is illustrated in Fig. 2c." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.871 + ], + "angle": 0, + "content": "In contrast to FocalClick, we use RoiAlign [10] to crop the logits and the fused feature \\( \\pmb{F}(N) \\) from GFR. Similar to GFR, our two prediction heads generate the local mask result \\( \\bar{M}_r^L \\) and the error map \\( \\pmb{E}^L \\) based on the prediction of GFR \\( M_r^G \\). The local refined mask \\( M_r^L \\) fuses \\( \\bar{M}_r^L \\) and \\( M_r^G \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.884, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {r} ^ {L} = \\sigma (\\boldsymbol {E} ^ {L}) \\odot \\bar {\\boldsymbol {M}} _ {r} ^ {L} + (1 - \\sigma (\\boldsymbol {E} ^ {L})) \\odot \\boldsymbol {M} _ {r} ^ {G}. \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19359" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.106 + ], + "angle": 0, + "content": "3.5. Dynamic Selector" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.113, + 0.484, + 0.369 + ], + "angle": 0, + "content": "As stated in FocalClick [4], when the previous mask \\( M_{t-1} \\) is already well-refined, especially when it has been refined for many rounds, re-predicting the full mask may destroy areas that have already been well refined. FocalClick [4] always pastes the refined local mask onto the previous mask after 10 interactions, as by this point the previous mask typically converges to good segmentation quality. In this paper, we use the dynamic selector to select which mask should the local result \\( M_r^L \\) be pasted onto, the global refined mask \\( M_r^G \\) or the previous mask \\( M_{t-1} \\). Taking the global result \\( M_r^G \\) pasted with \\( M_r^L \\) as the output can refine the local area based on the refined full mask, which is suitable when re-predicting the full mask will lead to less error. And when re-predicting the full mask leads to more error, which often happens when \\( M_{t-1} \\) is already precise enough, we paste \\( M_r^L \\) onto \\( M_{t-1} \\) to only refine the target local area and avoid overwriting other areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.415 + ], + "angle": 0, + "content": "We adaptively decide which mask to paste onto by comparing the changes in error maps \\( \\pmb{E}^{G} \\) between each iteration of GFR. We use the error ratio \\( R_{e} \\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.422, + 0.483, + 0.44 + ], + "angle": 0, + "content": "\\[\nR _ {e} = \\left\\| H \\left(\\boldsymbol {E} _ {t} ^ {G}\\right) \\right\\| _ {0} / \\left\\| H \\left(\\boldsymbol {E} _ {t - 1} ^ {G}\\right) \\right\\| _ {0}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.553 + ], + "angle": 0, + "content": "\\(H(\\cdot)\\) returns 1 when the value is greater than 0, else returns 0. If the error ratio \\(R_{e}\\) is larger than a threshold \\(\\theta\\), it indicates that the modification on the full mask will lead to more error, and we choose to paste \\(M_r^L\\) onto \\(M_{t - 1}\\) as the output \\(M_t\\). Otherwise, paste it onto \\(M_r^G\\) as \\(M_t\\). The output \\(M_t\\) will become a previous mask \\(M_{t - 1}\\) in next interaction. In our experiment, the threshold \\(\\theta\\) is set to be 1.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.287, + 0.577 + ], + "angle": 0, + "content": "3.6. Training supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.72 + ], + "angle": 0, + "content": "First, we adopted Normalized Focal Loss \\( L_{nfl} \\) proposed in RITM [27] to finetune the SAM decoder. Then, we used Dice Loss \\( L_{dice} \\) and Binary Cross Entropy Loss \\( L_{bce} \\) to supervise the error heads of GFR and LFR. \\( L_{wnfl} \\) denotes weighted NFL loss for the refined predictions. We calculate the error regions between the ground truth and the binary mask (for GFR it is \\( M_t^S \\), and for LFR it is \\( M_r^G \\)). We calculate the NFL loss and add weight (1.5 in our experiment) for these error regions to get \\( L_{wnfl} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.726, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {L} = \\boldsymbol {L} _ {n f l} + \\boldsymbol {L} _ {d i c e} ^ {G} + \\boldsymbol {L} _ {b c e} ^ {G} + \\boldsymbol {L} _ {w n f l} ^ {G} \\tag {6} \\\\ + \\boldsymbol {L} _ {d i c e} ^ {L} + \\boldsymbol {L} _ {b c e} ^ {L} + \\boldsymbol {L} _ {w n f l} ^ {L} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.804 + ], + "angle": 0, + "content": "where the superscripts \\(G\\) and \\(L\\) indicate the supervision for GFR and LFR respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.222, + 0.833 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "We conduct a series of experiments to test the performance of our model. We contrast our method with previous works. The experimental settings and the results of the experiments are as follows." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.715, + 0.107 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.113, + 0.617, + 0.127 + ], + "angle": 0, + "content": "4.1.1. Dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.132, + 0.907, + 0.299 + ], + "angle": 0, + "content": "We train our model on 2 different dataset configurations. For one configuration we train our model on COCO [18] + LVIS [6] datasets, following previous works [4, 12, 13, 21, 27]. For another configuration, we test the performance of our method on high-precision datasets. We train our model on HQSeg-44K [14, 22] datasets. HQSeg-44K dataset is a high-quality and well-annotated dataset with \\(44K\\) extremely accurate image mask annotations. We evaluate the models' zero-shot interactive segmentation capabilities on GrabCut [25], Berkeley [23], DAVIS [24], SBD [8], and HQSeg-44K validation set, aligning with previous studies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.304, + 0.724, + 0.319 + ], + "angle": 0, + "content": "4.1.2. Implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.906, + 0.444 + ], + "angle": 0, + "content": "We utilize the pre-trained ViT-H and ViT-B from SAM [15] as the backbone with the prompt encoder and decoder. The GlobalFusion Refiner and the LocalFusion Refiner are light-weighted convolutional neural networks, as shown in Fig. 2. We use \\( N = 3 \\) ResBasicBlocks to achieve the best results. Training and evaluations are performed on a server with 4 NVIDIA Tesla V100-PCIE-32GB GPUs and Intel(R) Xeon(R) Gold 6278C CPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.684, + 0.465 + ], + "angle": 0, + "content": "4.1.3. Training strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.636 + ], + "angle": 0, + "content": "In the training process, we adopt the two-stage training strategy of FocSAM [13] that firstly fine-tune the SAM decoder for 320k iterations at a batch size of 4 with frozen image encoder and prompt encoder, and then train SAM-REF with the frozen decoder and frozen encoders in the same settings for 80k iterations. We adopt InterFormer's click simulation strategy [12] for interactive simulation before loss computation. Moreover, we use the image encoder to preextract and store the image embeddings of the datasets to reduce computational costs. This strategy greatly improves the training speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.64, + 0.655 + ], + "angle": 0, + "content": "4.1.4. Evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We report the results on metrics of mean Intersection over Union (mIoU), number of clicks (NoC), number of failure cases (NoF), and the Seconds Per Click (SPC) metrics, following previous work [4, 13, 15, 21, 22]. And we test the latency of Segmentation Anything Task (SAT Latency) following SegNext [22]. mIoU measures the average intersection over union (IoU) given a fixed number of consecutive interactions. We report the mIoU on 5 clicks. NoC measures the number of clicks required to achieve a target IoU. We report NoC90 and NoC95 to measure the segmentation ability. NoC90 is a commonly used evaluation metric. With the increasing demand for high-quality segmentation, we provide the stricter NoC95 to demonstrate the capability to handle the details of our method. NoF measures the number of failure cases that cannot achieve the target IoU in 20 clicks. We report the failure cases on NoF95. SPC is a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19360" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.952, + 0.359 + ], + "angle": 0, + "content": "
MethodTrain dataBackboneSPC/s↓Grabcut [25]Berkeley [23]DAVIS [24]SBD [8]
NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓
Early Fusion
F-BRS [26]CVPR20COCO+LVISResNet-2.72-4.57-7.41-7.73-
CDNet [3]ICCV21COCO+LVISResNet-1.52-2.06-7.04-5.56-
EdgeFlow [7]ICCVW21COCO+LVISHRNet18-1.72-2.40---5.77-
RITM [27]ICIP22COCO+LVISHRNet32-1.562.482.105.415.3411.525.7112.00
PseudoClick [20]ECCV22COCO+LVISHRNet32-1.50-2.08-5.11-5.54-
FocalClick [4]CVPR22COCO+LVISSegFB3-S23840.2321.681.921.714.214.9010.405.5911.91
AdaptiveClick [17]TNNLS2024COCO+LVISViT-H1.691.38-1.64-4.60-4.68-
SimpleClick [21]ICCV23COCO+LVISViT-H8.241.501.661.754.344.788.884.7010.76
Late Fusion
InterFormer [12]ICCV23COCO+LVISViT-B0.1531.50-3.14-6.19-6.34-
InterFormer [12]ICCV23COCO+LVISViT-L0.2711.36-2.53-5.21-5.51-
SAM [15]ICCV23SA-1BViT-H0.4131.882.282.095.145.1910.007.6215.03
HQ-SAM [14]NeurIPS24HQSeg-44KViT-H*0.4431.862.262.144.945.069.657.9115.08
FocSAM [13]CVPR24COCO+LVISViT-H*0.4671.442.601.503.444.7610.775.0711.67
SAM-REFHQSeg-44KViT-H*0.5111.762.221.984.574.789.386.8112.18
SAM-REFCOCO+LVISViT-B*0.2781.402.441.583.734.7510.455.5611.76
SAM-REFCOCO+LVISViT-H*0.5111.362.201.433.184.549.104.4410.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.369, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Table 1. Comparison of SPC, NoC90 and NoC95 with previous methods. We report the results on the four mainstream datasets. The superscript * indicates frozen backbone." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.411, + 0.868, + 0.603 + ], + "angle": 0, + "content": "
MethodTrain dataBackboneLatency(s)↓5-mIoU↑NoC90↓NoC95↓NoF95↓
FocalClick [4]CVPR22COCO+LVISSegFB3-S225630.8584.638.1212.63835
FocalClick [4]CVPR22COCO+LVISSegFB3-S238440.1485.457.0310.74649
SimpleClick [21]ICCV23COCO+LVISViT-B88.585.117.4712.39797
InterFormer [12]ICCV23COCO+LVISViT-B30.582.627.1710.77658
SegNext [22]CVPR24COCO+LVISViT-B22.185.717.1811.52700
SAM [15]ICCV23SA-1BViT-B4.086.167.4612.42811
SAM [15]ICCV23SA-1BViT-H4.2188.06.5010.53653
HQ-SAM [14]NeurIPS24HQSeg-44KViT-B*4.789.456.4910.79671
FocSAM [13]CVPR24COCO+LVISViT-H*18.688.65.749.44580
SAM-REFHQSeg-44KViT-B*5.190.755.9010.08607
SAM-REFCOCO+LVISViT-B*5.188.56.199.84603
SAM-REFCOCO+LVISViT-H*5.2289.65.449.16566
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.613, + 0.908, + 0.642 + ], + "angle": 0, + "content": "Table 2. Results on high-quality datasets. All of the models are evaluated on the HQSeg-44K validation set [14]. SAM-REF achieves better results than SAM-based and other methods. The superscript * indicates frozen backbone." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.668, + 0.486, + 0.851 + ], + "angle": 0, + "content": "commonly used metric to measure the segmentation speed, which represents the time consumed of each click. SPC is tested on CPU. Following SegNext [22], SAT Latency measures the latency for the Segment Anything Task (SAT). We fix the input image size at \\(1024 \\times 1024\\) and prompt the model with a grid of \\(16 \\times 16\\) points. We compute the total time required to process all points by inputting one point at a time with our previous mask. SAT Latency is tested on a single GPU. Compared to SPC, SAT Latency calculates the overall latency but not one-prompt latency, which favors late fusion methods where the image is only encoded once." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.667, + 0.655, + 0.683 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Performance on widely used benchmarks. In Tab. 1, we present the main results of SAM-REF, benchmarked against mainstream methods. Our method is based on SAM, where we freeze the SAM encoder and train only a few remaining parameters to mitigate bias inherent. We compare our model with previous early fusion and late fusion methods. As reported, SAM-REF achieves state-of-the-art performance on the NoC90 in all mainstream evaluation datasets. In addition, when applying the stricter criteria evaluation standard NoC95, our method is slightly weaker than SimpleClick only on the DAVIS and Grabcut datasets. In the remaining datasets, SAM-REF achieves state-of-the-art performance in both NOC90 and NOC95 metrics, us" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.103, + 0.288, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.204, + 0.22, + 0.218 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.104, + 0.425, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.178, + 0.425, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.104, + 0.561, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.178, + 0.56, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.104, + 0.696, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.178, + 0.696, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.104, + 0.83, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.178, + 0.83, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.833, + 0.132, + 0.871, + 0.144 + ], + "angle": 0, + "content": "SAM" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.223, + 0.286, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.323, + 0.242, + 0.336 + ], + "angle": 0, + "content": "GroundTruth" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.251, + 0.425, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.323, + 0.396, + 0.336 + ], + "angle": 0, + "content": "5st click" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.251, + 0.561, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.323, + 0.537, + 0.336 + ], + "angle": 0, + "content": "\\(10^{\\mathrm{st}}\\) click" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.251, + 0.696, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.323, + 0.67, + 0.336 + ], + "angle": 0, + "content": "\\(15^{\\mathrm{st}}\\) click" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.251, + 0.83, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.323, + 0.801, + 0.336 + ], + "angle": 0, + "content": "\\(20^{\\mathrm{st}}\\) click" + }, + { + "type": "image_caption", + "bbox": [ + 0.833, + 0.206, + 0.894, + 0.219 + ], + "angle": 0, + "content": "FocSAM" + }, + { + "type": "image_caption", + "bbox": [ + 0.834, + 0.28, + 0.868, + 0.292 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.365, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Qualitative analysis on a challenge Example. The far-left image presents a challenging example featuring the image and ground truth (GT) with blue masks. The top, middle, and bottom rows on the right respectively display the segmentation results of SAM, FocSAM, and our method, at the 5th, 10th, 15th, and 20th clicks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.434, + 0.484, + 0.554 + ], + "angle": 0, + "content": "ing ViT-H as the backbone. However, as the early fusion method, SimpleClick requires a heavy image encoding at each interaction, whereas our approach is much more time-efficient, with only \\(6.2\\%\\) of SimpleClick's SPC. Additionally, we provide ViT-B results, showing it matches the speed (0.278 SPC) and accuracy (4.75 NoC90 on DAVIS) of the fastest early fusion method FocalClick [4] (0.232 SPC and 4.90 NoC90 on DAVIS)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.562, + 0.484, + 0.669 + ], + "angle": 0, + "content": "By reintroducing prompt-image synergy on late fusion, our method achieves state-of-the-art accuracy compared to all late fusion methods on these datasets. Specifically, under the same settings, we also outperform other SAM-based methods, such as FocSAM [13] and HQ-SAM [14]. We only add 0.1s SPC compared to SAM, which still maintain the high efficiency of late fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Performance on high-quality benchmark. In Tab. 2, we compare our method with state-of-the-art approaches on high-quality dataset HQSeg-44K [14](1537 samples). We compare two types of baselines: SAM-based methods and other methods. As reported, SAM-REF achieves state-ofthe-art performance to the two baselines, which demonstrate that SAM-REF can effectively enhances SAM's performance to top-tier levels in high-quality interactive segmentation while maintaining the low latency. Other latefusion methods (InterFormer [12] and SegNext [22]), which rely solely on the dense presentation of visual prompts, are more costly than SAM's use of the sparse representation [22]. Our method combines the benefits of both prompt types. It only focuses on detail extraction using dense maps of visual prompts, without needing complex fusion of visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.493 + ], + "angle": 0, + "content": "prompts and image embeddings. This demonstrates that our method can effectively enhance the performance of late fusion in complex scenarios, as evidenced by Qualitative Results." + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.506, + 0.855, + 0.558 + ], + "angle": 0, + "content": "
MethodParams/MBFPSMem/G
SAM [15]635.61.703.7
SAM-REF636.31.673.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.535, + 0.568, + 0.885, + 0.582 + ], + "angle": 0, + "content": "Table 3. Computation analysis for SAM and SAM-REF" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.907, + 0.724 + ], + "angle": 0, + "content": "Computation analysis. In Tab 3, we report the comparison of model parameters (Params), inference time per image (FPS), and GPU memory (Mem) usage between ViT-H based on SAM and SAM-REF. While SAM-REF produces substantially better segmentation quality, our method adds just 0.7MB to the model parameters compared to SAM, with negligible increases in GPU memory usage and inference time per image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.724, + 0.756 + ], + "angle": 0, + "content": "5. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.809 + ], + "angle": 0, + "content": "We conduct plenty of ablation studies, reporting experimental results for ViT-H models trained on COCO+LVIS and tested on DAVIS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Component Module Analysis. To demonstrate the effect of our model's modules, i.e., GFR, LFR, and the Dynamic Selector, we perform an ablation study on whether and which mask to paste \\( M_r^L \\). We have set up 4 experiments. One directly outputs \\( M_r^G \\), one always pastes \\( M_r^L \\) onto \\( M_r^G \\), one always pastes \\( M_r^L \\) onto \\( M_{t-1} \\), and one" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19362" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.444, + 0.176 + ], + "angle": 0, + "content": "
MethodNoC90NoC95
1no \\( M_r^L \\) only \\( M_r^G \\)4.669.53
2paste \\( M_r^L \\) to \\( M_{t-1} \\)4.809.75
3paste \\( M_r^L \\) to \\( M_r^G \\)4.609.40
4Dynamic Selector4.549.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.215 + ], + "angle": 0, + "content": "Table 4. Ablation study for the refiners and the Dynamic Se-. lector. The methods indicate which mask to paste \\(M_r^L\\) onto." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.241, + 0.485, + 0.455 + ], + "angle": 0, + "content": "pastes \\( M_r^L \\) according to the Dynamic Selector. For each experiment, we test NoC90 and NoC95 to evaluate the accuracy. We have shown the results in Tab. 4. The first experiment has no local refinement and always modifies the full mask, which leads to mutual influence between interactions and cannot achieve high accuracy. The second experiment only refines a local region but not the full mask in each interaction, which makes it modify slowly and needs more interactions to achieve the same accuracy. The third experiment with local refinement has a higher accuracy based on 1-th, but still cannot solve the mutual influence. The fourth experiment pasting \\( M_r^L \\) with Dynamic Selector avoids the interference and refines local regions, achieving the best accuracy." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.466, + 0.426, + 0.547 + ], + "angle": 0, + "content": "
MethodSPCNoC90NoC95
ResBlock×10.4834.579.94
ResBlock×30.5114.549.10
ResBlock×50.5454.669.88
ResBlock×80.7104.6110.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.557, + 0.455, + 0.572 + ], + "angle": 0, + "content": "Table 5. Ablation study for the number of ResBlocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.709 + ], + "angle": 0, + "content": "Number of ResBlocks. Tab. 5 shows the impact of the number of ResBlocks on accuracy and speed. With three ResBlocks, our method achieves the best results. Fewer blocks fail to capture details effectively, while more blocks reduce speed and harm the semantic capabilities of SAM image embeddings, thus decreasing accuracy. This experiment demonstrates the rationale for setting the number of ResBlocks to 3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.725, + 0.268, + 0.74 + ], + "angle": 0, + "content": "6. Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Fig. 3, we show the qualitative results compared to SAM [15] and FocSAM [13] on a challenge Example. This visualization demonstrates SAM-REF's enhanced precision. Our method achieves the highest mIoU in all clicks. Both FocSAM and our method are trained on \\(\\mathrm{COCO + LVIS}\\) and freeze the SAM's image encoder parameters. In the later clicks, our method can segment the detailed shape of the wheel, while SAM and FocSAM cannot. Based on challenging scenarios, the original SAM results are unsatisfactory. While FocSAM enhances stability and prevents pro" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.093, + 0.616, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.093, + 0.712, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.093, + 0.795, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.093, + 0.882, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.174, + 0.616, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.174, + 0.712, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.174, + 0.796, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.174, + 0.884, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.242, + 0.792, + 0.256 + ], + "angle": 0, + "content": "Coarse mask Error map" + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.242, + 0.892, + 0.256 + ], + "angle": 0, + "content": "Refined mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for the effectiveness of GlobalFusion and LocalFusion Refiner. The first row shows the global refinement results of GFR. The red box indicates the region for local refinement. The second row shows the local refinement results of LFR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.43 + ], + "angle": 0, + "content": "gressive deterioration, it remains limited by image embeddings, potentially missing object details. Our method retrieves lost object details from the original image through image-prompt synergy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.598 + ], + "angle": 0, + "content": "We present the results at each stage of our method in Fig. 4. The coarse mask is generated by the SAM decoder. The first row demonstrates how GFR refines the global segmentation process. The error map presents the error regions within SAM's predicted areas. In this scenario, the error regions are concentrated around the object edges and thin structures. Then, GFR specifically refines these error regions. The second row demonstrates the refinement process of LFR. It focuses on refining the areas with the largest errors between the interactions. In this image, the errors are mainly concentrated in the hollow regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.634, + 0.634 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.812 + ], + "angle": 0, + "content": "In view that early fusion strategies in the interactive segmentation can achieve high accuracy and late fusion strategies can achieve high efficiency, we propose a new method called SAM-REF, which combines the advantages of the two strategies. Our method redesigns the SAM-based interactive segmentation pipeline, fusing images and prompts from both global and local perspectives. This adaptation fully taps into SAM's potential for interactive segmentation. SAM-REF outperforms current state-of-the-art methods while maintaining SAM's low latency, even in challenging scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.694, + 0.849 + ], + "angle": 0, + "content": "8. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Xiaolin Hu was supported by the National Natural Science Foundation of China (No. U2341228)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19363" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] George Barnum, Sabera Talukder, and Yisong Yue. On the benefits of early fusion in multimodal representation learning. arXiv preprint arXiv:2011.07191, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7345-7354, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[4] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: Towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[6] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[7] Yuying Hao, Yi Liu, Zewu Wu, Lin Han, Yizhou Chen, Guowei Chen, Lutao Chu, Shiyu Tang, Zhiliang Yu, Zeyu Chen, et al. Edgeflow: Achieving practical interactive segmentation with edge-guided flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1551-1560, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[8] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[11] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[12] You Huang, Hao Yang, Ke Sun, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, and Rongrong Ji. Interformer: Real-time interactive image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22301-22311, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] You Huang, Zongyu Lan, Liujuan Cao, Xianming Lin, Shengchuan Zhang, Guannan Jiang, and Rongrong Ji. Foc" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "sam: Delving deeply into focused objects in segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3120-3130, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[16] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[17] Jiacheng Lin, Jiajun Chen, Kailun Yang, Alina Roitberg, Siyu Li, Zhiyong Li, and Shutao Li. Adaptiveclick: Click-aware transformer with adaptive focal loss for interactive image segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[19] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. Focuscut: Diving into a focus view in interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2637-2646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[20] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. Pseudoclick: Interactive image segmentation with click imitation. In European Conference on Computer Vision, pages 728-745. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[21] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.776 + ], + "angle": 0, + "content": "[22] Qin Liu, Jaemin Cho, Mohit Bansal, and Marc Niethammer. Rethinking interactive image segmentation with low latency high quality and diverse prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3782, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[23] Kevin McGuinness and Noel E O'connor. A comparative evaluation of interactive segmentation algorithms. Pattern Recognition, 43(2):434-444, 2010. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19364" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8623-8632, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[27] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[29] Zhaozhi Xie, Bochen Guan, Weihao Jiang, Muyang Yi, Yue Ding, Hongtao Lu, and Lei Zhang. Pa-sam: Prompt adapter sam for high-quality image segmentation. arXiv preprint arXiv:2401.13051, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[30] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19365" + } + ] +] \ No newline at end of file diff --git a/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_origin.pdf b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c45ae731f31a47c8f1a455972b77d938dd2f177c --- /dev/null +++ b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/a959cab0-5bb3-4e06-9751-4a034dc7813b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ceaac2691c9ec605f191563d003b686073e1fb14addbd32cb523f7aabcec2387 +size 2489007 diff --git a/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/full.md b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ed4168764edfefc45a08af444bdd487e4b6f4fc --- /dev/null +++ b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/full.md @@ -0,0 +1,338 @@ +# SAM-REF: Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model + +Chongkai $\mathrm{Yu}^{1,\dagger}$ , Ting Liu $^{1,\dagger,\boxtimes}$ , Anqi Li $^{2}$ , Xiaochao Qu $^{1}$ , Chengjing Wu $^{1}$ , Luoqi Liu $^{1}$ , Xiaolin Hu $^{3,\boxtimes}$ + +$^{1}$ MT Lab, Meitu Inc., Beijing 100083, China + +$^{2}$ Beijing Institute of Technology + +$^{3}$ Department of Computer Science and Technology + +BNRist, IDG/McGovern Institute for Brain Research + +Tsinghua University, Beijing 100084, China + +{yck, lt, qxc, ethan, llq5}@meitu.com + +xIhu@tsinghua.edu.cn + +†Joint first authors. ♕Joint corresponding author. + +# Abstract + +Interactive segmentation is to segment the mask of the target object according to the user's interactive prompts. There are two mainstream strategies: early fusion and late fusion. Current specialist models utilize the early fusion strategy that encodes the combination of images and prompts to target the prompted objects, yet repetitive complex computations on the images result in high latency. Late fusion models extract image embeddings once and merge them with the prompts in later interactions. This strategy avoids redundant image feature extraction and improves efficiency significantly. A recent milestone is the Segment Anything Model (SAM). However, this strategy limits the models' ability to extract detailed information from the prompted target zone. To address this issue, we propose SAM-REF, a two-stage refinement framework that fully integrates images and prompts by using a lightweight refiner into the interaction of late fusion, which combines the accuracy of early fusion and maintains the efficiency of late fusion. Through extensive experiments, we show that our SAM-REF model outperforms the current state-of-the-art method in most metrics on segmentation quality without compromising efficiency. + +# 1. Introduction + +Interactive segmentation [4, 12, 13, 15, 22] is an important task in computer vision that has garnered significant attention from both academia and industry. It leverages user input through simple interactions, such as scribbles, boxes, and clicks, to conveniently obtain accurate masks. Based on the timing of prompt-image fusion, existing methods can be categorized into early fusion strategies and late fusion strategies, as shown in Fig. 1. + +![](images/4ec9a8b6a4b94d9f3237422b8e0b5b11e234b6cd2da12ab2a17d7fe18d7ab65f.jpg) +(a) Early fusion + +![](images/7b6e05b855797dd35fa3958322f0ad7a15d5b1274a61ddbd32a8713289bcfb9e.jpg) +(b) Late fusion + +![](images/a727f9eb6c01befdbe92b6c005875ac7e748ab68457061952a5834829641c2bb.jpg) +(c) Our method +Figure 1. Three pipelines for interactive image segmentation. (a) Early fusion pipeline. (b) Late fusion pipeline. (c) Our proposed pipeline, which takes advantages of both early fusion and late fusion pipelines. The orange arrows indicate the interaction between the user and the systems. + +Early fusion methods such as FocalClick [4], and SimpleClick [21] jointly encode the image and prompts during interaction (Fig. 1a). Since the encoder can generate features specifically for target objects, these early fusion pipelines can achieve accurate segmentation results, espe + +cially for complex and challenging cases like thin objects. However, encoding the image in each interaction leads to low efficiency. In contrast, late fusion methods like the Segment Anything Model (SAM) [15] and InterFormer [12] extract features of the image before interaction, and fuse the prompt with the features during each interaction (Fig. 1b). These methods often utilize a powerful image encoder, Vision Transformer (ViT) [5], to generate embeddings with rich information. SAM [15] also used a billion-scale large dataset for training to enhance the encoder's extraction capability. Late fusion strategy avoids redundant image feature extraction, improving training and inference efficiency significantly. However, despite the ability of a powerful image encoder to focus on all objects, it struggles to capture the fine details of the target objects, because it discards the interaction between the original image and the prompt. This always leads to unsatisfactory results in many complex scenarios. Subsequent finetune approaches [13, 14, 29] based on SAM are also limited by the image embeddings. + +In this paper, we present a method called SAM-REF which takes advantage of both early fusion and late fusion pipelines. The key idea is to utilize the late fusion pipeline for its high efficiency but integrate the prompt and the input image during the user's interaction with the system, which provides precise instructions of the user and fine details of the image. To achieve this goal, we design a lightweight refiner and place it after SAM (Fig. 1c). The overall process is as follows. First, SAM generates the image embeddings with strong semantic representation capability and the prompt embeddings, and then combines them into the decoder to generate the coarse mask. Then the lightweight refiner uses a coarse-to-fine structure that can utilize global and local information. It obtains the semantic information of the image from SAM's embeddings, and processes the details by reintroducing image-prompt synergy. The light refiner generates a refined mask as the final output. + +Experimentally, SAM-REF fully exploits the potential of SAM for interactive segmentation. On widely used benchmark datasets (GrabCut [25], Berkeley [23], DAVIS [24], and SBD [8]) and high-quality benchmark dataset HQSeg-44K [14], our method achieves state-of-the-art performance in most metrics while maintaining SAM's low latency advantage. + +In summary, we propose SAM-REF, a new pipeline that leverages the power of SAM but integrates the prompt and the image during the interaction process to capture object details, leading to both high accuracy and high efficiency. + +# 2. Related Work + +In this paper, we focus on interactive segmentation. There are mainly two methods, early fusion and late fusion interactive segmentation method, divided based on the timing of prompt-image fusion. + +# 2.1. Early Fusion Methods + +Early fusion interactive segmentation [1] means combining different information in the early stages of networks. It combines the image with interactive information before processing them through deeper network layers, e.g. ViT [5], which is capable of producing more refined and contextually aware segmentation results. The first deep-learning-based interactive segmentation model, DiOS [30], takes the early fusion strategy by embedding positive and negative clicks into distance maps and concatenating them with the original image as input. RITM [27] introduces concatenating the previous mask with point maps, and later works mostly follow this approach. Many methods have been used to improve accuracy. For example, FocalClick [4] and FocusCut [19] use local refinement to optimize the details. CDNet [3] proposes a pixel diffusion module to fully fuse the prompt and the image. SimpleClick [21] up-samples the image feature, and PseudoClick [20] simulates an additional prompt. Though high precision, encoding the image and prompts every interaction leads to the high latency of early fusion, which has always been a problem. Some work has tried to reduce latency. FocalClick [4] compresses the size of the feature to speed up, but this method loses some accuracy. + +# 2.2. Late Fusion Methods + +Late fusion strategy extracts image embeddings once and combines them with interactive prompts in later interactions. Interformer [12] and SAM [15] decouple the light prompt encoder and the decoder from the heavy encoder, accelerating the speed. However, late fusion cannot catch details according to the focusless embeddings, leading to low precision. To promote accuracy, SegNext [22] adds dense representation and fusion of visual prompts to extract more visual information, FocSAM [13] introduces window-based attention to focus on the target zone of embeddings, while HQ-SAM [14] uses global-local fusion and proposed HQ-SAM dataset with high quality. These methods aim to explore more information from the embeddings. However, the embeddings may not have enough detailed information about the target zone, so these methods cannot achieve significant improvements in accuracy. In contrast, our method combines early fusion and late fusion, extracting detailed information from the image directly. + +# 3. Method + +Since our proposed method is built upon SAM, we first briefly review the architecture of SAM. Then we describe our method SAM-REF, which has three major components, Global Fusion Refiner (GFR), Local Fusion Refiner (LFR), and a Dynamic Selector (DS). Finally, we introduce the training loss. + +![](images/52679d1be08940b73dce2643e6dcc5ba886fd28a022a6574148e32fc29051b24.jpg) +(a) Pipeline + +![](images/41c7a046a775ee7f56f6877742a0f824168f9ae61d36b2862ef61ab33cbd54cc.jpg) +(b) GFR + +(c) LFR +Figure 2. The framework of the proposed SAM-REF. We propose a decoupled lightweight refiner while maintaining the inherent structure of SAM [15] (a). The refiner mainly has three modules, the Global Fusion Refiner (GFR), the Local Fusion Refiner (LFR), and the Dynamic Selector (DS). In GFR (b) and LFR (c), Conv $\times n$ represents an n-layer convolutional network. +![](images/c422981d3509b9a9c0bd00e536a6a15e2a03de8edf23474d163804bbaed8702c.jpg) +: Channel-wise concatenation +$\oplus$ : Element-wise summation +$\odot$ : Element-wise multiplication +@: Sigmoid + +# 3.1. Preliminaries: SAM + +SAM consists of three modules: an image encode, a prompt encode, and a mask decode. Image encoder: It is a heavy ViT-based backbone, structured into four stages with equal depth. Each stage employs window-based attention to enhance computational efficiency [16]. Prompt encoder: It is a lightweight structure consisting of MLP and CNN. In the interaction phase, it transforms the positional information of the input boxes, clicks, and masks into embeddings. However, the sparse representations of visual prompts, similar to those used in linguistic prompts, may inhibit the ability to capture image details [22]. Mask decoder: It is a + +lightweight structure composed of a Transformer [28] decoder and a convolutional mask head. It combines image embeddings and prompt embeddings through a two-layer transformer-based decoder to predict the final masks. + +During the interaction process, the prompt encoder and mask decoder are lightweight, and the heavy image encoder only needs to be performed once before the interaction. Therefore, the entire inference pipeline is very efficient. + +# 3.2. Our Pipeline of the SAM-REF + +To preserve the inherent structure of SAM while fully exploiting its potential, we adopt a decoupled design approach + +wherever possible. To this end, we introduce two modules GFR and LFR to refine the mask progressively. + +The overall architecture of SAM-REF is shown in Fig. 2. First, we use a late fusion approach to combine the encoded image with the prompt, producing a coarse mask, just like SAM [15]. Then we use the lightweight refiner, a decoupled module from SAM, to refine the coarse mask in a coarse-to-fine manner. In addition to the coarse mask and the prompt, the image and the feature from the encoder are also input to the lightweight refiner. Inside the lightweight refiner, we first use the global fusion refiner (GFR) to directly combine the prompt with the image and extract detailed information about the prompted target with a lightweight convolutional network. Meanwhile, it progressively incorporates SAM's image embeddings at each stage to preserve SAM's robust deep semantic capabilities, thus refining the overall objects. + +To further refine the local details of the target region, we propose the local fusion refiner (LFR). It crops the target region of the mask from GFR according to the prompt, refines the cropped mask, and pastes it back to the mask. We propose a dynamic selector to adaptively determine which mask to paste on. The cropped mask can be pasted on the predicted result of GFR or the previous mask as the output. The former can re-predict the full mask, while the latter can only refine the target local region without modifying other regions. + +# 3.3. Global Fusion Refiner (GFR) + +# 3.3.1. Fusion of Image, Prompt, Coarse Mask and Encoder Features + +As stated in Sec. 1, we need to efficiently fuse the prompt and the image. Unlike SAM [15] that represents the prompts with sparse one-dimension vectors, we use a dense presentation of prompts (disks with a small radius) to match the spatial dimensions of the image [2, 4, 22, 27], which better preserving the detailed spatial attributes of the visual prompts. We encode the prompts and the coarse mask $M_{t}^{S}$ ( $S$ refers to SAM, and $t$ refers to the $t$ -th interaction) to generate a dense map $D$ . The dense map is the concatenation of the positive prompt map, the negative prompt map, and the coarse mask $M_{t}^{S}$ . We adopt the method from RITM [27], utilizing two convolutions to align the scale and channels of the image and the dense map from $I_{ori} \in \mathbb{R}^{3 \times 1024 \times 1024}$ , $D_{ori} \in \mathbb{R}^{3 \times 256 \times 256}$ to $I \in \mathbb{R}^{64 \times 256 \times 256}$ , $D \in \mathbb{R}^{64 \times 256 \times 256}$ . + +We assume that SAM, relying on the powerful ViT architecture, has already learned deep semantic information [11]. The primary deficiency lies in its understanding of low-level details. To maintain SAM's semantic understanding and design an efficient detail extraction module, we employ a scale-invariant convolutional structure and integrate it with SAM image embeddings $\pmb{F}_{Enc} \in \mathbb{R}^{256 \times 64 \times 64}$ . + +Specifically, as illustrated in Fig. 2b, we progressively + +upscale the SAM image embeddings $\mathbf{F}_{Enc}$ to $\hat{\mathbf{F}}_{Enc} \in \mathbb{R}^{64 \times 256 \times 256}$ . Then we fuse the image, prompts, coarse mask, and encoder features by adding them together: + +$$ +\boldsymbol {F} (0) = \boldsymbol {I} + \boldsymbol {D} + \hat {\boldsymbol {F}} _ {E n c}. \tag {1} +$$ + +Here $D$ contains the information of prompts and the coarse mask. Subsequently, we utilize a scale-invariant ResBasicBlock [9] to capture fine details. This process is repeated $N$ times, with each iteration merging with the SAM image embeddings to prevent the loss of semantic information from SAM: + +$$ +\boldsymbol {F} (i + 1) = \operatorname {R e s B l o c k} (\boldsymbol {F} (i)) + \hat {\boldsymbol {F}} _ {E n c}, i \in [ 0, N - 1 ]. \tag {2} +$$ + +$\pmb{F}(N)$ contains the information of image, prompt, coarse mask, and encoder features. + +# 3.3.2. Fusion of Coarse Mask and Refined Mask + +Since the coarse mask $M_{t}^{S}$ is the prediction of SAM, it is generally accurate. We want to identify the inaccurate areas of $M_{t}^{S}$ , and only modify these areas. GFR predicts a mask result $\bar{M}_{r}^{G}$ and an error map $E^{G}$ , which represents error-prone regions. We replace these error-prone regions in $M_{t}^{S}$ with the corresponding regions in $M_{r}^{G}$ . The fusion process is formulated as follows. + +$$ +\boldsymbol {M} _ {r} ^ {G} = \sigma (\boldsymbol {E} ^ {G}) \odot \bar {\boldsymbol {M}} _ {r} ^ {G} + (1 - \sigma (\boldsymbol {E} ^ {G})) \odot \boldsymbol {M} _ {t} ^ {S}, \tag {3} +$$ + +where $\sigma(\cdot)$ represents Sigmoid function. $E^G$ is supervised by the error regions between the ground truth and the binary mask. After GFR, we obtain the global refinement of the objects. + +# 3.4. Local Fusion Refiner (LFR) + +To further refine the local details of the target region, we propose LFR to refine the local region in $M_r^G$ . In this module, we perform prompt and image fusion again for the user-interacted local regions to refine the local details of the objects. Inspired by FocalClick [4], we calculate the max connected component of the difference between the GFR result and the previous mask. Subsequently, we generate the external box for the connected component that includes the last prompt and proportionally expand it. The specific design of the LFR is illustrated in Fig. 2c. + +In contrast to FocalClick, we use RoiAlign [10] to crop the logits and the fused feature $\pmb{F}(N)$ from GFR. Similar to GFR, our two prediction heads generate the local mask result $\bar{M}_r^L$ and the error map $\pmb{E}^L$ based on the prediction of GFR $M_r^G$ . The local refined mask $M_r^L$ fuses $\bar{M}_r^L$ and $M_r^G$ as follows: + +$$ +\boldsymbol {M} _ {r} ^ {L} = \sigma (\boldsymbol {E} ^ {L}) \odot \bar {\boldsymbol {M}} _ {r} ^ {L} + (1 - \sigma (\boldsymbol {E} ^ {L})) \odot \boldsymbol {M} _ {r} ^ {G}. \tag {4} +$$ + +# 3.5. Dynamic Selector + +As stated in FocalClick [4], when the previous mask $M_{t-1}$ is already well-refined, especially when it has been refined for many rounds, re-predicting the full mask may destroy areas that have already been well refined. FocalClick [4] always pastes the refined local mask onto the previous mask after 10 interactions, as by this point the previous mask typically converges to good segmentation quality. In this paper, we use the dynamic selector to select which mask should the local result $M_r^L$ be pasted onto, the global refined mask $M_r^G$ or the previous mask $M_{t-1}$ . Taking the global result $M_r^G$ pasted with $M_r^L$ as the output can refine the local area based on the refined full mask, which is suitable when re-predicting the full mask will lead to less error. And when re-predicting the full mask leads to more error, which often happens when $M_{t-1}$ is already precise enough, we paste $M_r^L$ onto $M_{t-1}$ to only refine the target local area and avoid overwriting other areas. + +We adaptively decide which mask to paste onto by comparing the changes in error maps $\pmb{E}^{G}$ between each iteration of GFR. We use the error ratio $R_{e}$ , formulated as: + +$$ +R _ {e} = \left\| H \left(\boldsymbol {E} _ {t} ^ {G}\right) \right\| _ {0} / \left\| H \left(\boldsymbol {E} _ {t - 1} ^ {G}\right) \right\| _ {0}. \tag {5} +$$ + +$H(\cdot)$ returns 1 when the value is greater than 0, else returns 0. If the error ratio $R_{e}$ is larger than a threshold $\theta$ , it indicates that the modification on the full mask will lead to more error, and we choose to paste $M_r^L$ onto $M_{t - 1}$ as the output $M_t$ . Otherwise, paste it onto $M_r^G$ as $M_t$ . The output $M_t$ will become a previous mask $M_{t - 1}$ in next interaction. In our experiment, the threshold $\theta$ is set to be 1.1. + +# 3.6. Training supervision + +First, we adopted Normalized Focal Loss $L_{nfl}$ proposed in RITM [27] to finetune the SAM decoder. Then, we used Dice Loss $L_{dice}$ and Binary Cross Entropy Loss $L_{bce}$ to supervise the error heads of GFR and LFR. $L_{wnfl}$ denotes weighted NFL loss for the refined predictions. We calculate the error regions between the ground truth and the binary mask (for GFR it is $M_t^S$ , and for LFR it is $M_r^G$ ). We calculate the NFL loss and add weight (1.5 in our experiment) for these error regions to get $L_{wnfl}$ . + +$$ +\begin{array}{l} \boldsymbol {L} = \boldsymbol {L} _ {n f l} + \boldsymbol {L} _ {d i c e} ^ {G} + \boldsymbol {L} _ {b c e} ^ {G} + \boldsymbol {L} _ {w n f l} ^ {G} \tag {6} \\ + \boldsymbol {L} _ {d i c e} ^ {L} + \boldsymbol {L} _ {b c e} ^ {L} + \boldsymbol {L} _ {w n f l} ^ {L} \\ \end{array} +$$ + +where the superscripts $G$ and $L$ indicate the supervision for GFR and LFR respectively. + +# 4. Experiments + +We conduct a series of experiments to test the performance of our model. We contrast our method with previous works. The experimental settings and the results of the experiments are as follows. + +# 4.1. Experimental Setting + +# 4.1.1. Dataset. + +We train our model on 2 different dataset configurations. For one configuration we train our model on COCO [18] + LVIS [6] datasets, following previous works [4, 12, 13, 21, 27]. For another configuration, we test the performance of our method on high-precision datasets. We train our model on HQSeg-44K [14, 22] datasets. HQSeg-44K dataset is a high-quality and well-annotated dataset with $44K$ extremely accurate image mask annotations. We evaluate the models' zero-shot interactive segmentation capabilities on GrabCut [25], Berkeley [23], DAVIS [24], SBD [8], and HQSeg-44K validation set, aligning with previous studies. + +# 4.1.2. Implementation details. + +We utilize the pre-trained ViT-H and ViT-B from SAM [15] as the backbone with the prompt encoder and decoder. The GlobalFusion Refiner and the LocalFusion Refiner are light-weighted convolutional neural networks, as shown in Fig. 2. We use $N = 3$ ResBasicBlocks to achieve the best results. Training and evaluations are performed on a server with 4 NVIDIA Tesla V100-PCIE-32GB GPUs and Intel(R) Xeon(R) Gold 6278C CPU. + +# 4.1.3. Training strategy. + +In the training process, we adopt the two-stage training strategy of FocSAM [13] that firstly fine-tune the SAM decoder for 320k iterations at a batch size of 4 with frozen image encoder and prompt encoder, and then train SAM-REF with the frozen decoder and frozen encoders in the same settings for 80k iterations. We adopt InterFormer's click simulation strategy [12] for interactive simulation before loss computation. Moreover, we use the image encoder to preextract and store the image embeddings of the datasets to reduce computational costs. This strategy greatly improves the training speed. + +# 4.1.4. Evaluation. + +We report the results on metrics of mean Intersection over Union (mIoU), number of clicks (NoC), number of failure cases (NoF), and the Seconds Per Click (SPC) metrics, following previous work [4, 13, 15, 21, 22]. And we test the latency of Segmentation Anything Task (SAT Latency) following SegNext [22]. mIoU measures the average intersection over union (IoU) given a fixed number of consecutive interactions. We report the mIoU on 5 clicks. NoC measures the number of clicks required to achieve a target IoU. We report NoC90 and NoC95 to measure the segmentation ability. NoC90 is a commonly used evaluation metric. With the increasing demand for high-quality segmentation, we provide the stricter NoC95 to demonstrate the capability to handle the details of our method. NoF measures the number of failure cases that cannot achieve the target IoU in 20 clicks. We report the failure cases on NoF95. SPC is a + +
MethodTrain dataBackboneSPC/s↓Grabcut [25]Berkeley [23]DAVIS [24]SBD [8]
NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓
Early Fusion
F-BRS [26]CVPR20COCO+LVISResNet-2.72-4.57-7.41-7.73-
CDNet [3]ICCV21COCO+LVISResNet-1.52-2.06-7.04-5.56-
EdgeFlow [7]ICCVW21COCO+LVISHRNet18-1.72-2.40---5.77-
RITM [27]ICIP22COCO+LVISHRNet32-1.562.482.105.415.3411.525.7112.00
PseudoClick [20]ECCV22COCO+LVISHRNet32-1.50-2.08-5.11-5.54-
FocalClick [4]CVPR22COCO+LVISSegFB3-S23840.2321.681.921.714.214.9010.405.5911.91
AdaptiveClick [17]TNNLS2024COCO+LVISViT-H1.691.38-1.64-4.60-4.68-
SimpleClick [21]ICCV23COCO+LVISViT-H8.241.501.661.754.344.788.884.7010.76
Late Fusion
InterFormer [12]ICCV23COCO+LVISViT-B0.1531.50-3.14-6.19-6.34-
InterFormer [12]ICCV23COCO+LVISViT-L0.2711.36-2.53-5.21-5.51-
SAM [15]ICCV23SA-1BViT-H0.4131.882.282.095.145.1910.007.6215.03
HQ-SAM [14]NeurIPS24HQSeg-44KViT-H*0.4431.862.262.144.945.069.657.9115.08
FocSAM [13]CVPR24COCO+LVISViT-H*0.4671.442.601.503.444.7610.775.0711.67
SAM-REFHQSeg-44KViT-H*0.5111.762.221.984.574.789.386.8112.18
SAM-REFCOCO+LVISViT-B*0.2781.402.441.583.734.7510.455.5611.76
SAM-REFCOCO+LVISViT-H*0.5111.362.201.433.184.549.104.4410.41
+ +Table 1. Comparison of SPC, NoC90 and NoC95 with previous methods. We report the results on the four mainstream datasets. The superscript * indicates frozen backbone. + +
MethodTrain dataBackboneLatency(s)↓5-mIoU↑NoC90↓NoC95↓NoF95↓
FocalClick [4]CVPR22COCO+LVISSegFB3-S225630.8584.638.1212.63835
FocalClick [4]CVPR22COCO+LVISSegFB3-S238440.1485.457.0310.74649
SimpleClick [21]ICCV23COCO+LVISViT-B88.585.117.4712.39797
InterFormer [12]ICCV23COCO+LVISViT-B30.582.627.1710.77658
SegNext [22]CVPR24COCO+LVISViT-B22.185.717.1811.52700
SAM [15]ICCV23SA-1BViT-B4.086.167.4612.42811
SAM [15]ICCV23SA-1BViT-H4.2188.06.5010.53653
HQ-SAM [14]NeurIPS24HQSeg-44KViT-B*4.789.456.4910.79671
FocSAM [13]CVPR24COCO+LVISViT-H*18.688.65.749.44580
SAM-REFHQSeg-44KViT-B*5.190.755.9010.08607
SAM-REFCOCO+LVISViT-B*5.188.56.199.84603
SAM-REFCOCO+LVISViT-H*5.2289.65.449.16566
+ +Table 2. Results on high-quality datasets. All of the models are evaluated on the HQSeg-44K validation set [14]. SAM-REF achieves better results than SAM-based and other methods. The superscript * indicates frozen backbone. + +commonly used metric to measure the segmentation speed, which represents the time consumed of each click. SPC is tested on CPU. Following SegNext [22], SAT Latency measures the latency for the Segment Anything Task (SAT). We fix the input image size at $1024 \times 1024$ and prompt the model with a grid of $16 \times 16$ points. We compute the total time required to process all points by inputting one point at a time with our previous mask. SAT Latency is tested on a single GPU. Compared to SPC, SAT Latency calculates the overall latency but not one-prompt latency, which favors late fusion methods where the image is only encoded once. + +# 4.2. Main Results + +Performance on widely used benchmarks. In Tab. 1, we present the main results of SAM-REF, benchmarked against mainstream methods. Our method is based on SAM, where we freeze the SAM encoder and train only a few remaining parameters to mitigate bias inherent. We compare our model with previous early fusion and late fusion methods. As reported, SAM-REF achieves state-of-the-art performance on the NoC90 in all mainstream evaluation datasets. In addition, when applying the stricter criteria evaluation standard NoC95, our method is slightly weaker than SimpleClick only on the DAVIS and Grabcut datasets. In the remaining datasets, SAM-REF achieves state-of-the-art performance in both NOC90 and NOC95 metrics, us + +![](images/acab688e6e97f5af9fd54a592a23b1d836fd4ac2789f5c17912d0ce15ea40ca9.jpg) +Image + +![](images/f25d27ec3707083c2d66620a4e1c3a71388aafaf0e63bd70f4e279b425a6461c.jpg) + +![](images/b509f082a2c2c90c6260448ec9a905047eedf220a1233017e6b066743498cd3c.jpg) + +![](images/ea794b6c9c977c3fff44665c611f35f12825ae804a17681059c1225b662e98d8.jpg) + +![](images/e6cb6b63cdebd7c68cac9e965fd2a6f73a7a5988aabedad6e0c93decc585a684.jpg) + +![](images/b82bf923370b7e448d05962b093e8527b85a588d264da59a71f2598726127007.jpg) + +![](images/0e5c9d9402330593fcaa9694cd98b254e4578532d5182fe13fb14e2f475272f1.jpg) + +![](images/0877529792eead23cc0daf2ddbf1c68c9bc47b68cf6427478b1c4696478eeae8.jpg) +SAM + +![](images/a487c2c4681487141f4b51ee07634b51e347047cf1e00c0cb2920aeaf4c2e426.jpg) +FocSAM + +![](images/e3fb9e5fb07c0f22f23feb1a35428e21270d72947d97b32312fa69b9747765fe.jpg) +GroundTruth +Figure 3. Qualitative analysis on a challenge Example. The far-left image presents a challenging example featuring the image and ground truth (GT) with blue masks. The top, middle, and bottom rows on the right respectively display the segmentation results of SAM, FocSAM, and our method, at the 5th, 10th, 15th, and 20th clicks. + +![](images/47704268c36e4f69999936d72cc627545350e458c2fb4f34f7b95b0468055761.jpg) +5st click + +![](images/a50e4f75c861028fa567763c2a29f7c2b99e20a559fa258fbceb27dffe264010.jpg) +$10^{\mathrm{st}}$ click + +![](images/9f82255e2ddeac8fba554051311aa15effc76c94fab6c6ea464e69e9e062a736.jpg) +$15^{\mathrm{st}}$ click + +![](images/5c3c6fec5c4e0bb1267f2988f54ee3cfaeed765c0ef32500e513a9d26de50709.jpg) +$20^{\mathrm{st}}$ click +Ours + +ing ViT-H as the backbone. However, as the early fusion method, SimpleClick requires a heavy image encoding at each interaction, whereas our approach is much more time-efficient, with only $6.2\%$ of SimpleClick's SPC. Additionally, we provide ViT-B results, showing it matches the speed (0.278 SPC) and accuracy (4.75 NoC90 on DAVIS) of the fastest early fusion method FocalClick [4] (0.232 SPC and 4.90 NoC90 on DAVIS). + +By reintroducing prompt-image synergy on late fusion, our method achieves state-of-the-art accuracy compared to all late fusion methods on these datasets. Specifically, under the same settings, we also outperform other SAM-based methods, such as FocSAM [13] and HQ-SAM [14]. We only add 0.1s SPC compared to SAM, which still maintain the high efficiency of late fusion. + +Performance on high-quality benchmark. In Tab. 2, we compare our method with state-of-the-art approaches on high-quality dataset HQSeg-44K [14](1537 samples). We compare two types of baselines: SAM-based methods and other methods. As reported, SAM-REF achieves state-ofthe-art performance to the two baselines, which demonstrate that SAM-REF can effectively enhances SAM's performance to top-tier levels in high-quality interactive segmentation while maintaining the low latency. Other latefusion methods (InterFormer [12] and SegNext [22]), which rely solely on the dense presentation of visual prompts, are more costly than SAM's use of the sparse representation [22]. Our method combines the benefits of both prompt types. It only focuses on detail extraction using dense maps of visual prompts, without needing complex fusion of visual + +prompts and image embeddings. This demonstrates that our method can effectively enhance the performance of late fusion in complex scenarios, as evidenced by Qualitative Results. + +
MethodParams/MBFPSMem/G
SAM [15]635.61.703.7
SAM-REF636.31.673.7
+ +Table 3. Computation analysis for SAM and SAM-REF + +Computation analysis. In Tab 3, we report the comparison of model parameters (Params), inference time per image (FPS), and GPU memory (Mem) usage between ViT-H based on SAM and SAM-REF. While SAM-REF produces substantially better segmentation quality, our method adds just 0.7MB to the model parameters compared to SAM, with negligible increases in GPU memory usage and inference time per image. + +# 5. Ablation Experiments + +We conduct plenty of ablation studies, reporting experimental results for ViT-H models trained on COCO+LVIS and tested on DAVIS. + +Component Module Analysis. To demonstrate the effect of our model's modules, i.e., GFR, LFR, and the Dynamic Selector, we perform an ablation study on whether and which mask to paste $M_r^L$ . We have set up 4 experiments. One directly outputs $M_r^G$ , one always pastes $M_r^L$ onto $M_r^G$ , one always pastes $M_r^L$ onto $M_{t-1}$ , and one + +
MethodNoC90NoC95
1no \( M_r^L \) only \( M_r^G \)4.669.53
2paste \( M_r^L \) to \( M_{t-1} \)4.809.75
3paste \( M_r^L \) to \( M_r^G \)4.609.40
4Dynamic Selector4.549.10
+ +pastes $M_r^L$ according to the Dynamic Selector. For each experiment, we test NoC90 and NoC95 to evaluate the accuracy. We have shown the results in Tab. 4. The first experiment has no local refinement and always modifies the full mask, which leads to mutual influence between interactions and cannot achieve high accuracy. The second experiment only refines a local region but not the full mask in each interaction, which makes it modify slowly and needs more interactions to achieve the same accuracy. The third experiment with local refinement has a higher accuracy based on 1-th, but still cannot solve the mutual influence. The fourth experiment pasting $M_r^L$ with Dynamic Selector avoids the interference and refines local regions, achieving the best accuracy. + +Table 4. Ablation study for the refiners and the Dynamic Se-. lector. The methods indicate which mask to paste $M_r^L$ onto. + +
MethodSPCNoC90NoC95
ResBlock×10.4834.579.94
ResBlock×30.5114.549.10
ResBlock×50.5454.669.88
ResBlock×80.7104.6110.61
+ +Table 5. Ablation study for the number of ResBlocks. + +Number of ResBlocks. Tab. 5 shows the impact of the number of ResBlocks on accuracy and speed. With three ResBlocks, our method achieves the best results. Fewer blocks fail to capture details effectively, while more blocks reduce speed and harm the semantic capabilities of SAM image embeddings, thus decreasing accuracy. This experiment demonstrates the rationale for setting the number of ResBlocks to 3. + +# 6. Qualitative Result + +In Fig. 3, we show the qualitative results compared to SAM [15] and FocSAM [13] on a challenge Example. This visualization demonstrates SAM-REF's enhanced precision. Our method achieves the highest mIoU in all clicks. Both FocSAM and our method are trained on $\mathrm{COCO + LVIS}$ and freeze the SAM's image encoder parameters. In the later clicks, our method can segment the detailed shape of the wheel, while SAM and FocSAM cannot. Based on challenging scenarios, the original SAM results are unsatisfactory. While FocSAM enhances stability and prevents pro + +![](images/ac332cd1beb006877e680b08b9847a0de0b3108e64c1e0f913b47c9a71d68e7e.jpg) + +![](images/1e96dac20c4b5f8ba666d37d3bcedd089466fb93feb9b22227972b36ef7acdd1.jpg) + +![](images/ce41af0e62a013835694fba5b9e85d22eb461cbdc7a3ee8ee40949c00ca14f0f.jpg) + +![](images/82ef2b81520563c9a83260a09becc5fd15a1139d9b98fa035a1b8e6305839ab4.jpg) + +![](images/c9b9e0ceeff6c8cc946b023e9c63d11d76b0b5152100378a0c3f616c36bb245e.jpg) +Figure 4. Qualitative results for the effectiveness of GlobalFusion and LocalFusion Refiner. The first row shows the global refinement results of GFR. The red box indicates the region for local refinement. The second row shows the local refinement results of LFR. + +![](images/b6acf7bd555576c2a1b5fca038e6c775e26e46380d08dd887593dbf8389b6ffb.jpg) +Coarse mask Error map + +![](images/f597bf79370772830588684d7015f7f106f982a327e8d0b0be521f0e3c59e9bd.jpg) + +![](images/6acd82e30ebdacc43f7a3a231aca03434d0d857417329cef311efe427c1c0323.jpg) +Refined mask + +gressive deterioration, it remains limited by image embeddings, potentially missing object details. Our method retrieves lost object details from the original image through image-prompt synergy. + +We present the results at each stage of our method in Fig. 4. The coarse mask is generated by the SAM decoder. The first row demonstrates how GFR refines the global segmentation process. The error map presents the error regions within SAM's predicted areas. In this scenario, the error regions are concentrated around the object edges and thin structures. Then, GFR specifically refines these error regions. The second row demonstrates the refinement process of LFR. It focuses on refining the areas with the largest errors between the interactions. In this image, the errors are mainly concentrated in the hollow regions. + +# 7. Conclusion + +In view that early fusion strategies in the interactive segmentation can achieve high accuracy and late fusion strategies can achieve high efficiency, we propose a new method called SAM-REF, which combines the advantages of the two strategies. Our method redesigns the SAM-based interactive segmentation pipeline, fusing images and prompts from both global and local perspectives. This adaptation fully taps into SAM's potential for interactive segmentation. SAM-REF outperforms current state-of-the-art methods while maintaining SAM's low latency, even in challenging scenarios. + +# 8. Acknowledgement + +Xiaolin Hu was supported by the National Natural Science Foundation of China (No. U2341228). + +# References + +[1] George Barnum, Sabera Talukder, and Yisong Yue. On the benefits of early fusion in multimodal representation learning. arXiv preprint arXiv:2011.07191, 2020. 2 +[2] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 4 +[3] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7345-7354, 2021. 2, 6 +[4] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: Towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022. 1, 2, 4, 5, 6, 7 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[6] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5 +[7] Yuying Hao, Yi Liu, Zewu Wu, Lin Han, Yizhou Chen, Guowei Chen, Lutao Chu, Shiyu Tang, Zhiliang Yu, Zeyu Chen, et al. Edgeflow: Achieving practical interactive segmentation with edge-guided flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1551-1560, 2021. 6 +[8] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 2, 5, 6 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4 +[11] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4 +[12] You Huang, Hao Yang, Ke Sun, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, and Rongrong Ji. Interformer: Real-time interactive image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22301-22311, 2023. 1, 2, 5, 6, 7 +[13] You Huang, Zongyu Lan, Liujuan Cao, Xianming Lin, Shengchuan Zhang, Guannan Jiang, and Rongrong Ji. Foc + +sam: Delving deeply into focused objects in segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3120-3130, 2024. 1, 2, 5, 6, 7, 8 +[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7 +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[16] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3 +[17] Jiacheng Lin, Jiajun Chen, Kailun Yang, Alina Roitberg, Siyu Li, Zhiyong Li, and Shutao Li. Adaptiveclick: Click-aware transformer with adaptive focal loss for interactive image segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6 +[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[19] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. Focuscut: Diving into a focus view in interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2637-2646, 2022. 2 +[20] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. Pseudoclick: Interactive image segmentation with click imitation. In European Conference on Computer Vision, pages 728-745. Springer, 2022. 2, 6 +[21] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 1, 2, 5, 6 +[22] Qin Liu, Jaemin Cho, Mohit Bansal, and Marc Niethammer. Rethinking interactive image segmentation with low latency high quality and diverse prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3782, 2024. 1, 2, 3, 4, 5, 6, 7 +[23] Kevin McGuinness and Noel E O'connor. A comparative evaluation of interactive segmentation algorithms. Pattern Recognition, 43(2):434-444, 2010. 2, 5, 6 +[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2, 5, 6 + +[25] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. "grabcut" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 2, 5, 6 +[26] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8623-8632, 2020. 6 +[27] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 2, 4, 5, 6 +[28] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[29] Zhaozhi Xie, Bochen Guan, Weihao Jiang, Muyang Yi, Yue Ding, Hongtao Lu, and Lei Zhang. Pa-sam: Prompt adapter sam for high-quality image segmentation. arXiv preprint arXiv:2401.13051, 2024. 2 +[30] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 2 \ No newline at end of file diff --git a/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/images.zip b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc1c927d1fb2ff403dd84ab9179e38a3a156cfc0 --- /dev/null +++ b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00c2ac1cbc232c4d8417e7fff16580dafe7a602ffa089bf7cd7e34314422205a +size 564712 diff --git a/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/layout.json b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d85c02d6671dcd64c53a78b3c8bea09f3168723 --- /dev/null +++ b/2025/SAM-REF_ Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model/layout.json @@ -0,0 +1,9001 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "type": "text", + "content": "SAM-REF: Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "spans": [ + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": "Chongkai " + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{1,\\dagger}" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": ", Ting Liu" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "^{1,\\dagger,\\boxtimes}" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": ", Anqi Li" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": ", Xiaochao Qu" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": ", Chengjing Wu" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": ", Luoqi Liu" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "text", + "content": ", Xiaolin Hu" + }, + { + "bbox": [ + 65, + 159, + 546, + 176 + ], + "type": "inline_equation", + "content": "^{3,\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 177, + 388, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 177, + 388, + 189 + ], + "spans": [ + { + "bbox": [ + 223, + 177, + 388, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 223, + 177, + 388, + 189 + ], + "type": "text", + "content": " MT Lab, Meitu Inc., Beijing 100083, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "spans": [ + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "type": "text", + "content": " Beijing Institute of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 205, + 399, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 205, + 399, + 217 + ], + "spans": [ + { + "bbox": [ + 211, + 205, + 399, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 211, + 205, + 399, + 217 + ], + "type": "text", + "content": "Department of Computer Science and Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 220, + 402, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 220, + 402, + 230 + ], + "spans": [ + { + "bbox": [ + 208, + 220, + 402, + 230 + ], + "type": "text", + "content": "BNRist, IDG/McGovern Institute for Brain Research" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 225, + 234, + 386, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 234, + 386, + 245 + ], + "spans": [ + { + "bbox": [ + 225, + 234, + 386, + 245 + ], + "type": "text", + "content": "Tsinghua University, Beijing 100084, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 206, + 247, + 405, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 247, + 405, + 258 + ], + "spans": [ + { + "bbox": [ + 206, + 247, + 405, + 258 + ], + "type": "text", + "content": "{yck, lt, qxc, ethan, llq5}@meitu.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 250, + 262, + 359, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 262, + 359, + 272 + ], + "spans": [ + { + "bbox": [ + 250, + 262, + 359, + 272 + ], + "type": "text", + "content": "xIhu@tsinghua.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 208, + 274, + 399, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 274, + 399, + 287 + ], + "spans": [ + { + "bbox": [ + 208, + 274, + 399, + 287 + ], + "type": "text", + "content": "†Joint first authors. ♕Joint corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 296, + 200, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 296, + 200, + 308 + ], + "spans": [ + { + "bbox": [ + 151, + 296, + 200, + 308 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 322, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 574 + ], + "type": "text", + "content": "Interactive segmentation is to segment the mask of the target object according to the user's interactive prompts. There are two mainstream strategies: early fusion and late fusion. Current specialist models utilize the early fusion strategy that encodes the combination of images and prompts to target the prompted objects, yet repetitive complex computations on the images result in high latency. Late fusion models extract image embeddings once and merge them with the prompts in later interactions. This strategy avoids redundant image feature extraction and improves efficiency significantly. A recent milestone is the Segment Anything Model (SAM). However, this strategy limits the models' ability to extract detailed information from the prompted target zone. To address this issue, we propose SAM-REF, a two-stage refinement framework that fully integrates images and prompts by using a lightweight refiner into the interaction of late fusion, which combines the accuracy of early fusion and maintains the efficiency of late fusion. Through extensive experiments, we show that our SAM-REF model outperforms the current state-of-the-art method in most metrics on segmentation quality without compromising efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Interactive segmentation [4, 12, 13, 15, 22] is an important task in computer vision that has garnered significant attention from both academia and industry. It leverages user input through simple interactions, such as scribbles, boxes, and clicks, to conveniently obtain accurate masks. Based on the timing of prompt-image fusion, existing methods can be categorized into early fusion strategies and late fusion strategies, as shown in Fig. 1." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 331, + 297, + 548, + 363 + ], + "blocks": [ + { + "bbox": [ + 331, + 297, + 548, + 363 + ], + "lines": [ + { + "bbox": [ + 331, + 297, + 548, + 363 + ], + "spans": [ + { + "bbox": [ + 331, + 297, + 548, + 363 + ], + "type": "image", + "image_path": "4ec9a8b6a4b94d9f3237422b8e0b5b11e234b6cd2da12ab2a17d7fe18d7ab65f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 363, + 462, + 373 + ], + "lines": [ + { + "bbox": [ + 406, + 363, + 462, + 373 + ], + "spans": [ + { + "bbox": [ + 406, + 363, + 462, + 373 + ], + "type": "text", + "content": "(a) Early fusion" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 331, + 381, + 549, + 441 + ], + "blocks": [ + { + "bbox": [ + 331, + 381, + 549, + 441 + ], + "lines": [ + { + "bbox": [ + 331, + 381, + 549, + 441 + ], + "spans": [ + { + "bbox": [ + 331, + 381, + 549, + 441 + ], + "type": "image", + "image_path": "7b6e05b855797dd35fa3958322f0ad7a15d5b1274a61ddbd32a8713289bcfb9e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 442, + 460, + 452 + ], + "lines": [ + { + "bbox": [ + 408, + 442, + 460, + 452 + ], + "spans": [ + { + "bbox": [ + 408, + 442, + 460, + 452 + ], + "type": "text", + "content": "(b) Late fusion" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 331, + 458, + 548, + 541 + ], + "blocks": [ + { + "bbox": [ + 331, + 458, + 548, + 541 + ], + "lines": [ + { + "bbox": [ + 331, + 458, + 548, + 541 + ], + "spans": [ + { + "bbox": [ + 331, + 458, + 548, + 541 + ], + "type": "image", + "image_path": "a727f9eb6c01befdbe92b6c005875ac7e748ab68457061952a5834829641c2bb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 544, + 463, + 553 + ], + "lines": [ + { + "bbox": [ + 408, + 544, + 463, + 553 + ], + "spans": [ + { + "bbox": [ + 408, + 544, + 463, + 553 + ], + "type": "text", + "content": "(c) Our method" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 566, + 555, + 622 + ], + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 622 + ], + "type": "text", + "content": "Figure 1. Three pipelines for interactive image segmentation. (a) Early fusion pipeline. (b) Late fusion pipeline. (c) Our proposed pipeline, which takes advantages of both early fusion and late fusion pipelines. The orange arrows indicate the interaction between the user and the systems." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Early fusion methods such as FocalClick [4], and SimpleClick [21] jointly encode the image and prompts during interaction (Fig. 1a). Since the encoder can generate features specifically for target objects, these early fusion pipelines can achieve accurate segmentation results, espe" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19356" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "content": "cially for complex and challenging cases like thin objects. However, encoding the image in each interaction leads to low efficiency. In contrast, late fusion methods like the Segment Anything Model (SAM) [15] and InterFormer [12] extract features of the image before interaction, and fuse the prompt with the features during each interaction (Fig. 1b). These methods often utilize a powerful image encoder, Vision Transformer (ViT) [5], to generate embeddings with rich information. SAM [15] also used a billion-scale large dataset for training to enhance the encoder's extraction capability. Late fusion strategy avoids redundant image feature extraction, improving training and inference efficiency significantly. However, despite the ability of a powerful image encoder to focus on all objects, it struggles to capture the fine details of the target objects, because it discards the interaction between the original image and the prompt. This always leads to unsatisfactory results in many complex scenarios. Subsequent finetune approaches [13, 14, 29] based on SAM are also limited by the image embeddings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 502 + ], + "type": "text", + "content": "In this paper, we present a method called SAM-REF which takes advantage of both early fusion and late fusion pipelines. The key idea is to utilize the late fusion pipeline for its high efficiency but integrate the prompt and the input image during the user's interaction with the system, which provides precise instructions of the user and fine details of the image. To achieve this goal, we design a lightweight refiner and place it after SAM (Fig. 1c). The overall process is as follows. First, SAM generates the image embeddings with strong semantic representation capability and the prompt embeddings, and then combines them into the decoder to generate the coarse mask. Then the lightweight refiner uses a coarse-to-fine structure that can utilize global and local information. It obtains the semantic information of the image from SAM's embeddings, and processes the details by reintroducing image-prompt synergy. The light refiner generates a refined mask as the final output." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 586 + ], + "type": "text", + "content": "Experimentally, SAM-REF fully exploits the potential of SAM for interactive segmentation. On widely used benchmark datasets (GrabCut [25], Berkeley [23], DAVIS [24], and SBD [8]) and high-quality benchmark dataset HQSeg-44K [14], our method achieves state-of-the-art performance in most metrics while maintaining SAM's low latency advantage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 586, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 635 + ], + "type": "text", + "content": "In summary, we propose SAM-REF, a new pipeline that leverages the power of SAM but integrates the prompt and the image during the interaction process to capture object details, leading to both high accuracy and high efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 142, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we focus on interactive segmentation. There are mainly two methods, early fusion and late fusion interactive segmentation method, divided based on the timing of prompt-image fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "type": "text", + "content": "2.1. Early Fusion Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 90, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 365 + ], + "type": "text", + "content": "Early fusion interactive segmentation [1] means combining different information in the early stages of networks. It combines the image with interactive information before processing them through deeper network layers, e.g. ViT [5], which is capable of producing more refined and contextually aware segmentation results. The first deep-learning-based interactive segmentation model, DiOS [30], takes the early fusion strategy by embedding positive and negative clicks into distance maps and concatenating them with the original image as input. RITM [27] introduces concatenating the previous mask with point maps, and later works mostly follow this approach. Many methods have been used to improve accuracy. For example, FocalClick [4] and FocusCut [19] use local refinement to optimize the details. CDNet [3] proposes a pixel diffusion module to fully fuse the prompt and the image. SimpleClick [21] up-samples the image feature, and PseudoClick [20] simulates an additional prompt. Though high precision, encoding the image and prompts every interaction leads to the high latency of early fusion, which has always been a problem. Some work has tried to reduce latency. FocalClick [4] compresses the size of the feature to speed up, but this method loses some accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 374, + 437, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 374, + 437, + 386 + ], + "spans": [ + { + "bbox": [ + 314, + 374, + 437, + 386 + ], + "type": "text", + "content": "2.2. Late Fusion Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 392, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 608 + ], + "type": "text", + "content": "Late fusion strategy extracts image embeddings once and combines them with interactive prompts in later interactions. Interformer [12] and SAM [15] decouple the light prompt encoder and the decoder from the heavy encoder, accelerating the speed. However, late fusion cannot catch details according to the focusless embeddings, leading to low precision. To promote accuracy, SegNext [22] adds dense representation and fusion of visual prompts to extract more visual information, FocSAM [13] introduces window-based attention to focus on the target zone of embeddings, while HQ-SAM [14] uses global-local fusion and proposed HQ-SAM dataset with high quality. These methods aim to explore more information from the embeddings. However, the embeddings may not have enough detailed information about the target zone, so these methods cannot achieve significant improvements in accuracy. In contrast, our method combines early fusion and late fusion, extracting detailed information from the image directly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 620, + 370, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 620, + 370, + 633 + ], + "spans": [ + { + "bbox": [ + 314, + 620, + 370, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "Since our proposed method is built upon SAM, we first briefly review the architecture of SAM. Then we describe our method SAM-REF, which has three major components, Global Fusion Refiner (GFR), Local Fusion Refiner (LFR), and a Dynamic Selector (DS). Finally, we introduce the training loss." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19357" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 73, + 541, + 217 + ], + "blocks": [ + { + "bbox": [ + 90, + 73, + 541, + 217 + ], + "lines": [ + { + "bbox": [ + 90, + 73, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 90, + 73, + 541, + 217 + ], + "type": "image", + "image_path": "52679d1be08940b73dce2643e6dcc5ba886fd28a022a6574148e32fc29051b24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 219, + 328, + 230 + ], + "lines": [ + { + "bbox": [ + 285, + 219, + 328, + 230 + ], + "spans": [ + { + "bbox": [ + 285, + 219, + 328, + 230 + ], + "type": "text", + "content": "(a) Pipeline" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 240, + 519, + 348 + ], + "blocks": [ + { + "bbox": [ + 78, + 240, + 519, + 348 + ], + "lines": [ + { + "bbox": [ + 78, + 240, + 519, + 348 + ], + "spans": [ + { + "bbox": [ + 78, + 240, + 519, + 348 + ], + "type": "image", + "image_path": "41c7a046a775ee7f56f6877742a0f824168f9ae61d36b2862ef61ab33cbd54cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 350, + 319, + 359 + ], + "lines": [ + { + "bbox": [ + 288, + 350, + 319, + 359 + ], + "spans": [ + { + "bbox": [ + 288, + 350, + 319, + 359 + ], + "type": "text", + "content": "(b) GFR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 109, + 371, + 517, + 463 + ], + "blocks": [ + { + "bbox": [ + 109, + 371, + 517, + 463 + ], + "lines": [ + { + "bbox": [ + 109, + 371, + 517, + 463 + ], + "spans": [ + { + "bbox": [ + 109, + 371, + 517, + 463 + ], + "type": "image", + "image_path": "c422981d3509b9a9c0bd00e536a6a15e2a03de8edf23474d163804bbaed8702c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 467, + 320, + 476 + ], + "lines": [ + { + "bbox": [ + 290, + 467, + 320, + 476 + ], + "spans": [ + { + "bbox": [ + 290, + 467, + 320, + 476 + ], + "type": "text", + "content": "(c) LFR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 478, + 203, + 487 + ], + "lines": [ + { + "bbox": [ + 93, + 478, + 203, + 487 + ], + "spans": [ + { + "bbox": [ + 93, + 478, + 203, + 487 + ], + "type": "text", + "content": ": Channel-wise concatenation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 216, + 478, + 315, + 488 + ], + "lines": [ + { + "bbox": [ + 216, + 478, + 315, + 488 + ], + "spans": [ + { + "bbox": [ + 216, + 478, + 315, + 488 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 216, + 478, + 315, + 488 + ], + "type": "text", + "content": " : Element-wise summation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 323, + 478, + 430, + 488 + ], + "lines": [ + { + "bbox": [ + 323, + 478, + 430, + 488 + ], + "spans": [ + { + "bbox": [ + 323, + 478, + 430, + 488 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 323, + 478, + 430, + 488 + ], + "type": "text", + "content": " : Element-wise multiplication" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 440, + 478, + 484, + 488 + ], + "lines": [ + { + "bbox": [ + 440, + 478, + 484, + 488 + ], + "spans": [ + { + "bbox": [ + 440, + 478, + 484, + 488 + ], + "type": "text", + "content": "@: Sigmoid" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 54, + 505, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 54, + 505, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 555, + 539 + ], + "type": "text", + "content": "Figure 2. The framework of the proposed SAM-REF. We propose a decoupled lightweight refiner while maintaining the inherent structure of SAM [15] (a). The refiner mainly has three modules, the Global Fusion Refiner (GFR), the Local Fusion Refiner (LFR), and the Dynamic Selector (DS). In GFR (b) and LFR (c), Conv " + }, + { + "bbox": [ + 54, + 505, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\times n" + }, + { + "bbox": [ + 54, + 505, + 555, + 539 + ], + "type": "text", + "content": " represents an n-layer convolutional network." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 173, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 173, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 173, + 571 + ], + "type": "text", + "content": "3.1. Preliminaries: SAM" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "SAM consists of three modules: an image encode, a prompt encode, and a mask decode. Image encoder: It is a heavy ViT-based backbone, structured into four stages with equal depth. Each stage employs window-based attention to enhance computational efficiency [16]. Prompt encoder: It is a lightweight structure consisting of MLP and CNN. In the interaction phase, it transforms the positional information of the input boxes, clicks, and masks into embeddings. However, the sparse representations of visual prompts, similar to those used in linguistic prompts, may inhibit the ability to capture image details [22]. Mask decoder: It is a" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 559, + 555, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 607 + ], + "type": "text", + "content": "lightweight structure composed of a Transformer [28] decoder and a convolutional mask head. It combines image embeddings and prompt embeddings through a two-layer transformer-based decoder to predict the final masks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 609, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 555, + 657 + ], + "type": "text", + "content": "During the interaction process, the prompt encoder and mask decoder are lightweight, and the heavy image encoder only needs to be performed once before the interaction. Therefore, the entire inference pipeline is very efficient." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 670, + 479, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 479, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 479, + 682 + ], + "type": "text", + "content": "3.2. Our Pipeline of the SAM-REF" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "To preserve the inherent structure of SAM while fully exploiting its potential, we adopt a decoupled design approach" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19358" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "wherever possible. To this end, we introduce two modules GFR and LFR to refine the mask progressively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 295, + 263 + ], + "type": "text", + "content": "The overall architecture of SAM-REF is shown in Fig. 2. First, we use a late fusion approach to combine the encoded image with the prompt, producing a coarse mask, just like SAM [15]. Then we use the lightweight refiner, a decoupled module from SAM, to refine the coarse mask in a coarse-to-fine manner. In addition to the coarse mask and the prompt, the image and the feature from the encoder are also input to the lightweight refiner. Inside the lightweight refiner, we first use the global fusion refiner (GFR) to directly combine the prompt with the image and extract detailed information about the prompted target with a lightweight convolutional network. Meanwhile, it progressively incorporates SAM's image embeddings at each stage to preserve SAM's robust deep semantic capabilities, thus refining the overall objects." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 264, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 296, + 384 + ], + "type": "text", + "content": "To further refine the local details of the target region, we propose the local fusion refiner (LFR). It crops the target region of the mask from GFR according to the prompt, refines the cropped mask, and pastes it back to the mask. We propose a dynamic selector to adaptively determine which mask to paste on. The cropped mask can be pasted on the predicted result of GFR or the previous mask as the output. The former can re-predict the full mask, while the latter can only refine the target local region without modifying other regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 216, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 216, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 216, + 405 + ], + "type": "text", + "content": "3.3. Global Fusion Refiner (GFR)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 433 + ], + "type": "text", + "content": "3.3.1. Fusion of Image, Prompt, Coarse Mask and Encoder Features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": "As stated in Sec. 1, we need to efficiently fuse the prompt and the image. Unlike SAM [15] that represents the prompts with sparse one-dimension vectors, we use a dense presentation of prompts (disks with a small radius) to match the spatial dimensions of the image [2, 4, 22, 27], which better preserving the detailed spatial attributes of the visual prompts. We encode the prompts and the coarse mask " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "M_{t}^{S}" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": " refers to SAM, and " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": "-th interaction) to generate a dense map " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": ". The dense map is the concatenation of the positive prompt map, the negative prompt map, and the coarse mask " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "M_{t}^{S}" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": ". We adopt the method from RITM [27], utilizing two convolutions to align the scale and channels of the image and the dense map from " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "I_{ori} \\in \\mathbb{R}^{3 \\times 1024 \\times 1024}" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "D_{ori} \\in \\mathbb{R}^{3 \\times 256 \\times 256}" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{64 \\times 256 \\times 256}" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{64 \\times 256 \\times 256}" + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": "We assume that SAM, relying on the powerful ViT architecture, has already learned deep semantic information [11]. The primary deficiency lies in its understanding of low-level details. To maintain SAM's semantic understanding and design an efficient detail extraction module, we employ a scale-invariant convolutional structure and integrate it with SAM image embeddings " + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{Enc} \\in \\mathbb{R}^{256 \\times 64 \\times 64}" + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Specifically, as illustrated in Fig. 2b, we progressively" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "type": "text", + "content": "upscale the SAM image embeddings " + }, + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{Enc}" + }, + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{Enc} \\in \\mathbb{R}^{64 \\times 256 \\times 256}" + }, + { + "bbox": [ + 313, + 71, + 555, + 109 + ], + "type": "text", + "content": ". Then we fuse the image, prompts, coarse mask, and encoder features by adding them together:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 382, + 118, + 555, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 118, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 382, + 118, + 555, + 133 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} (0) = \\boldsymbol {I} + \\boldsymbol {D} + \\hat {\\boldsymbol {F}} _ {E n c}. \\tag {1}", + "image_path": "348c7c6cf907ea039ec07dc7202fd08df20674ca31b1f84f75a6c2a64f762364.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "type": "text", + "content": " contains the information of prompts and the coarse mask. Subsequently, we utilize a scale-invariant ResBasicBlock [9] to capture fine details. This process is repeated " + }, + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 144, + 555, + 215 + ], + "type": "text", + "content": " times, with each iteration merging with the SAM image embeddings to prevent the loss of semantic information from SAM:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 226, + 555, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 226, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 319, + 226, + 555, + 242 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} (i + 1) = \\operatorname {R e s B l o c k} (\\boldsymbol {F} (i)) + \\hat {\\boldsymbol {F}} _ {E n c}, i \\in [ 0, N - 1 ]. \\tag {2}", + "image_path": "58c58b14121a13d4de2026df5729d5d097fde3e755619f6aa76b09a198688239.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 554, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 275 + ], + "type": "inline_equation", + "content": "\\pmb{F}(N)" + }, + { + "bbox": [ + 313, + 251, + 554, + 275 + ], + "type": "text", + "content": " contains the information of image, prompt, coarse mask, and encoder features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 522, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 522, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 522, + 296 + ], + "type": "text", + "content": "3.3.2. Fusion of Coarse Mask and Refined Mask" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": "Since the coarse mask " + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "inline_equation", + "content": "M_{t}^{S}" + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": " is the prediction of SAM, it is generally accurate. We want to identify the inaccurate areas of " + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "inline_equation", + "content": "M_{t}^{S}" + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": ", and only modify these areas. GFR predicts a mask result " + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\bar{M}_{r}^{G}" + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": " and an error map " + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "inline_equation", + "content": "E^{G}" + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": ", which represents error-prone regions. We replace these error-prone regions in " + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "inline_equation", + "content": "M_{t}^{S}" + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": " with the corresponding regions in " + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "inline_equation", + "content": "M_{r}^{G}" + }, + { + "bbox": [ + 313, + 300, + 555, + 384 + ], + "type": "text", + "content": ". The fusion process is formulated as follows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 394, + 555, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 394, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 337, + 394, + 555, + 410 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {r} ^ {G} = \\sigma (\\boldsymbol {E} ^ {G}) \\odot \\bar {\\boldsymbol {M}} _ {r} ^ {G} + (1 - \\sigma (\\boldsymbol {E} ^ {G})) \\odot \\boldsymbol {M} _ {t} ^ {S}, \\tag {3}", + "image_path": "18441e5c3c097623ede912976fbd1e2e39e7029557b8412656dc70a6e4ef4171.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "type": "text", + "content": " represents Sigmoid function. " + }, + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "type": "inline_equation", + "content": "E^G" + }, + { + "bbox": [ + 313, + 419, + 554, + 468 + ], + "type": "text", + "content": " is supervised by the error regions between the ground truth and the binary mask. After GFR, we obtain the global refinement of the objects." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 479, + 468, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 468, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 468, + 491 + ], + "type": "text", + "content": "3.4. Local Fusion Refiner (LFR)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 616 + ], + "type": "text", + "content": "To further refine the local details of the target region, we propose LFR to refine the local region in " + }, + { + "bbox": [ + 313, + 497, + 554, + 616 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 313, + 497, + 554, + 616 + ], + "type": "text", + "content": ". In this module, we perform prompt and image fusion again for the user-interacted local regions to refine the local details of the objects. Inspired by FocalClick [4], we calculate the max connected component of the difference between the GFR result and the previous mask. Subsequently, we generate the external box for the connected component that includes the last prompt and proportionally expand it. The specific design of the LFR is illustrated in Fig. 2c." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": "In contrast to FocalClick, we use RoiAlign [10] to crop the logits and the fused feature " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{F}(N)" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": " from GFR. Similar to GFR, our two prediction heads generate the local mask result " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\bar{M}_r^L" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": " and the error map " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{E}^L" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": " based on the prediction of GFR " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": ". The local refined mask " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": " fuses " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\bar{M}_r^L" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 313, + 617, + 555, + 689 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 338, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 338, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {r} ^ {L} = \\sigma (\\boldsymbol {E} ^ {L}) \\odot \\bar {\\boldsymbol {M}} _ {r} ^ {L} + (1 - \\sigma (\\boldsymbol {E} ^ {L})) \\odot \\boldsymbol {M} _ {r} ^ {G}. \\tag {4}", + "image_path": "2c3ca1324a969af465a7332a63b412f69120e777a48b0eedbf2d921bfce90d21.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19359" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 83 + ], + "type": "text", + "content": "3.5. Dynamic Selector" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": "As stated in FocalClick [4], when the previous mask " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " is already well-refined, especially when it has been refined for many rounds, re-predicting the full mask may destroy areas that have already been well refined. FocalClick [4] always pastes the refined local mask onto the previous mask after 10 interactions, as by this point the previous mask typically converges to good segmentation quality. In this paper, we use the dynamic selector to select which mask should the local result " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " be pasted onto, the global refined mask " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " or the previous mask " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": ". Taking the global result " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " pasted with " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " as the output can refine the local area based on the refined full mask, which is suitable when re-predicting the full mask will lead to less error. And when re-predicting the full mask leads to more error, which often happens when " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " is already precise enough, we paste " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 55, + 89, + 296, + 292 + ], + "type": "text", + "content": " to only refine the target local area and avoid overwriting other areas." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "type": "text", + "content": "We adaptively decide which mask to paste onto by comparing the changes in error maps " + }, + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{G}" + }, + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "type": "text", + "content": " between each iteration of GFR. We use the error ratio " + }, + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "type": "inline_equation", + "content": "R_{e}" + }, + { + "bbox": [ + 55, + 293, + 295, + 328 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 334, + 295, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 334, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 295, + 348 + ], + "type": "interline_equation", + "content": "R _ {e} = \\left\\| H \\left(\\boldsymbol {E} _ {t} ^ {G}\\right) \\right\\| _ {0} / \\left\\| H \\left(\\boldsymbol {E} _ {t - 1} ^ {G}\\right) \\right\\| _ {0}. \\tag {5}", + "image_path": "5cf292faba4717af78cdbc82e33ee12950c44b0b18e6a5a8b7487e516efd2784.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " returns 1 when the value is greater than 0, else returns 0. If the error ratio " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "R_{e}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " is larger than a threshold " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ", it indicates that the modification on the full mask will lead to more error, and we choose to paste " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_{t - 1}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " as the output " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ". Otherwise, paste it onto " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_t" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " will become a previous mask " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "M_{t - 1}" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " in next interaction. In our experiment, the threshold " + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": " is set to be 1.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 175, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 175, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 175, + 456 + ], + "type": "text", + "content": "3.6. Training supervision" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": "First, we adopted Normalized Focal Loss " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "L_{nfl}" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": " proposed in RITM [27] to finetune the SAM decoder. Then, we used Dice Loss " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "L_{dice}" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": " and Binary Cross Entropy Loss " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "L_{bce}" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": " to supervise the error heads of GFR and LFR. " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "L_{wnfl}" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": " denotes weighted NFL loss for the refined predictions. We calculate the error regions between the ground truth and the binary mask (for GFR it is " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "M_t^S" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": ", and for LFR it is " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": "). We calculate the NFL loss and add weight (1.5 in our experiment) for these error regions to get " + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "inline_equation", + "content": "L_{wnfl}" + }, + { + "bbox": [ + 55, + 461, + 296, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 574, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 574, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 295, + 607 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {L} = \\boldsymbol {L} _ {n f l} + \\boldsymbol {L} _ {d i c e} ^ {G} + \\boldsymbol {L} _ {b c e} ^ {G} + \\boldsymbol {L} _ {w n f l} ^ {G} \\tag {6} \\\\ + \\boldsymbol {L} _ {d i c e} ^ {L} + \\boldsymbol {L} _ {b c e} ^ {L} + \\boldsymbol {L} _ {w n f l} ^ {L} \\\\ \\end{array}", + "image_path": "890213009cf7402bffd3e26a580bdbf4b9f10c60f064cf9361db0ff13942bc2d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "content": "where the superscripts " + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "content": " indicate the supervision for GFR and LFR respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 646, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 135, + 659 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "We conduct a series of experiments to test the performance of our model. We contrast our method with previous works. The experimental settings and the results of the experiments are as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 437, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 437, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 437, + 84 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 89, + 377, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 377, + 100 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 377, + 100 + ], + "type": "text", + "content": "4.1.1. Dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 104, + 555, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 555, + 236 + ], + "type": "text", + "content": "We train our model on 2 different dataset configurations. For one configuration we train our model on COCO [18] + LVIS [6] datasets, following previous works [4, 12, 13, 21, 27]. For another configuration, we test the performance of our method on high-precision datasets. We train our model on HQSeg-44K [14, 22] datasets. HQSeg-44K dataset is a high-quality and well-annotated dataset with " + }, + { + "bbox": [ + 313, + 104, + 555, + 236 + ], + "type": "inline_equation", + "content": "44K" + }, + { + "bbox": [ + 313, + 104, + 555, + 236 + ], + "type": "text", + "content": " extremely accurate image mask annotations. We evaluate the models' zero-shot interactive segmentation capabilities on GrabCut [25], Berkeley [23], DAVIS [24], SBD [8], and HQSeg-44K validation set, aligning with previous studies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 240, + 443, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 443, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 443, + 252 + ], + "type": "text", + "content": "4.1.2. Implementation details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 255, + 554, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 351 + ], + "type": "text", + "content": "We utilize the pre-trained ViT-H and ViT-B from SAM [15] as the backbone with the prompt encoder and decoder. The GlobalFusion Refiner and the LocalFusion Refiner are light-weighted convolutional neural networks, as shown in Fig. 2. We use " + }, + { + "bbox": [ + 313, + 255, + 554, + 351 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 313, + 255, + 554, + 351 + ], + "type": "text", + "content": " ResBasicBlocks to achieve the best results. Training and evaluations are performed on a server with 4 NVIDIA Tesla V100-PCIE-32GB GPUs and Intel(R) Xeon(R) Gold 6278C CPU." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 356, + 418, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 418, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 418, + 368 + ], + "type": "text", + "content": "4.1.3. Training strategy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 371, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 503 + ], + "type": "text", + "content": "In the training process, we adopt the two-stage training strategy of FocSAM [13] that firstly fine-tune the SAM decoder for 320k iterations at a batch size of 4 with frozen image encoder and prompt encoder, and then train SAM-REF with the frozen decoder and frozen encoders in the same settings for 80k iterations. We adopt InterFormer's click simulation strategy [12] for interactive simulation before loss computation. Moreover, we use the image encoder to preextract and store the image embeddings of the datasets to reduce computational costs. This strategy greatly improves the training speed." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 507, + 391, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 391, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 391, + 518 + ], + "type": "text", + "content": "4.1.4. Evaluation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": "We report the results on metrics of mean Intersection over Union (mIoU), number of clicks (NoC), number of failure cases (NoF), and the Seconds Per Click (SPC) metrics, following previous work [4, 13, 15, 21, 22]. And we test the latency of Segmentation Anything Task (SAT Latency) following SegNext [22]. mIoU measures the average intersection over union (IoU) given a fixed number of consecutive interactions. We report the mIoU on 5 clicks. NoC measures the number of clicks required to achieve a target IoU. We report NoC90 and NoC95 to measure the segmentation ability. NoC90 is a commonly used evaluation metric. With the increasing demand for high-quality segmentation, we provide the stricter NoC95 to demonstrate the capability to handle the details of our method. NoF measures the number of failure cases that cannot achieve the target IoU in 20 clicks. We report the failure cases on NoF95. SPC is a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19360" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 582, + 284 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 582, + 284 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 582, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 582, + 284 + ], + "type": "table", + "html": "
MethodTrain dataBackboneSPC/s↓Grabcut [25]Berkeley [23]DAVIS [24]SBD [8]
NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓NoC90↓NoC95↓
Early Fusion
F-BRS [26]CVPR20COCO+LVISResNet-2.72-4.57-7.41-7.73-
CDNet [3]ICCV21COCO+LVISResNet-1.52-2.06-7.04-5.56-
EdgeFlow [7]ICCVW21COCO+LVISHRNet18-1.72-2.40---5.77-
RITM [27]ICIP22COCO+LVISHRNet32-1.562.482.105.415.3411.525.7112.00
PseudoClick [20]ECCV22COCO+LVISHRNet32-1.50-2.08-5.11-5.54-
FocalClick [4]CVPR22COCO+LVISSegFB3-S23840.2321.681.921.714.214.9010.405.5911.91
AdaptiveClick [17]TNNLS2024COCO+LVISViT-H1.691.38-1.64-4.60-4.68-
SimpleClick [21]ICCV23COCO+LVISViT-H8.241.501.661.754.344.788.884.7010.76
Late Fusion
InterFormer [12]ICCV23COCO+LVISViT-B0.1531.50-3.14-6.19-6.34-
InterFormer [12]ICCV23COCO+LVISViT-L0.2711.36-2.53-5.21-5.51-
SAM [15]ICCV23SA-1BViT-H0.4131.882.282.095.145.1910.007.6215.03
HQ-SAM [14]NeurIPS24HQSeg-44KViT-H*0.4431.862.262.144.945.069.657.9115.08
FocSAM [13]CVPR24COCO+LVISViT-H*0.4671.442.601.503.444.7610.775.0711.67
SAM-REFHQSeg-44KViT-H*0.5111.762.221.984.574.789.386.8112.18
SAM-REFCOCO+LVISViT-B*0.2781.402.441.583.734.7510.455.5611.76
SAM-REFCOCO+LVISViT-H*0.5111.362.201.433.184.549.104.4410.41
", + "image_path": "5714866e8131f353c1f39c503719408bca69fd7a5171a4c7f2a645cc54f96f9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 81, + 325, + 531, + 477 + ], + "blocks": [ + { + "bbox": [ + 55, + 292, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 555, + 314 + ], + "type": "text", + "content": "Table 1. Comparison of SPC, NoC90 and NoC95 with previous methods. We report the results on the four mainstream datasets. The superscript * indicates frozen backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 325, + 531, + 477 + ], + "lines": [ + { + "bbox": [ + 81, + 325, + 531, + 477 + ], + "spans": [ + { + "bbox": [ + 81, + 325, + 531, + 477 + ], + "type": "table", + "html": "
MethodTrain dataBackboneLatency(s)↓5-mIoU↑NoC90↓NoC95↓NoF95↓
FocalClick [4]CVPR22COCO+LVISSegFB3-S225630.8584.638.1212.63835
FocalClick [4]CVPR22COCO+LVISSegFB3-S238440.1485.457.0310.74649
SimpleClick [21]ICCV23COCO+LVISViT-B88.585.117.4712.39797
InterFormer [12]ICCV23COCO+LVISViT-B30.582.627.1710.77658
SegNext [22]CVPR24COCO+LVISViT-B22.185.717.1811.52700
SAM [15]ICCV23SA-1BViT-B4.086.167.4612.42811
SAM [15]ICCV23SA-1BViT-H4.2188.06.5010.53653
HQ-SAM [14]NeurIPS24HQSeg-44KViT-B*4.789.456.4910.79671
FocSAM [13]CVPR24COCO+LVISViT-H*18.688.65.749.44580
SAM-REFHQSeg-44KViT-B*5.190.755.9010.08607
SAM-REFCOCO+LVISViT-B*5.188.56.199.84603
SAM-REFCOCO+LVISViT-H*5.2289.65.449.16566
", + "image_path": "2cb587d6592d7fd5bdc87736b882c841b9580eb231e67692e0de8a205b109bd1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 485, + 555, + 508 + ], + "lines": [ + { + "bbox": [ + 55, + 485, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 555, + 508 + ], + "type": "text", + "content": "Table 2. Results on high-quality datasets. All of the models are evaluated on the HQSeg-44K validation set [14]. SAM-REF achieves better results than SAM-based and other methods. The superscript * indicates frozen backbone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "type": "text", + "content": "commonly used metric to measure the segmentation speed, which represents the time consumed of each click. SPC is tested on CPU. Following SegNext [22], SAT Latency measures the latency for the Segment Anything Task (SAT). We fix the input image size at " + }, + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "type": "text", + "content": " and prompt the model with a grid of " + }, + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 54, + 529, + 297, + 673 + ], + "type": "text", + "content": " points. We compute the total time required to process all points by inputting one point at a time with our previous mask. SAT Latency is tested on a single GPU. Compared to SPC, SAT Latency calculates the overall latency but not one-prompt latency, which favors late fusion methods where the image is only encoded once." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 528, + 400, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 528, + 400, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 528, + 400, + 540 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "Performance on widely used benchmarks. In Tab. 1, we present the main results of SAM-REF, benchmarked against mainstream methods. Our method is based on SAM, where we freeze the SAM encoder and train only a few remaining parameters to mitigate bias inherent. We compare our model with previous early fusion and late fusion methods. As reported, SAM-REF achieves state-of-the-art performance on the NoC90 in all mainstream evaluation datasets. In addition, when applying the stricter criteria evaluation standard NoC95, our method is slightly weaker than SimpleClick only on the DAVIS and Grabcut datasets. In the remaining datasets, SAM-REF achieves state-of-the-art performance in both NOC90 and NOC95 metrics, us" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19361" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 81, + 176, + 160 + ], + "blocks": [ + { + "bbox": [ + 70, + 81, + 176, + 160 + ], + "lines": [ + { + "bbox": [ + 70, + 81, + 176, + 160 + ], + "spans": [ + { + "bbox": [ + 70, + 81, + 176, + 160 + ], + "type": "image", + "image_path": "acab688e6e97f5af9fd54a592a23b1d836fd4ac2789f5c17912d0ce15ea40ca9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 161, + 134, + 172 + ], + "lines": [ + { + "bbox": [ + 107, + 161, + 134, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 161, + 134, + 172 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 82, + 260, + 137 + ], + "blocks": [ + { + "bbox": [ + 186, + 82, + 260, + 137 + ], + "lines": [ + { + "bbox": [ + 186, + 82, + 260, + 137 + ], + "spans": [ + { + "bbox": [ + 186, + 82, + 260, + 137 + ], + "type": "image", + "image_path": "f25d27ec3707083c2d66620a4e1c3a71388aafaf0e63bd70f4e279b425a6461c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 140, + 260, + 194 + ], + "blocks": [ + { + "bbox": [ + 186, + 140, + 260, + 194 + ], + "lines": [ + { + "bbox": [ + 186, + 140, + 260, + 194 + ], + "spans": [ + { + "bbox": [ + 186, + 140, + 260, + 194 + ], + "type": "image", + "image_path": "b509f082a2c2c90c6260448ec9a905047eedf220a1233017e6b066743498cd3c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 269, + 82, + 343, + 137 + ], + "blocks": [ + { + "bbox": [ + 269, + 82, + 343, + 137 + ], + "lines": [ + { + "bbox": [ + 269, + 82, + 343, + 137 + ], + "spans": [ + { + "bbox": [ + 269, + 82, + 343, + 137 + ], + "type": "image", + "image_path": "ea794b6c9c977c3fff44665c611f35f12825ae804a17681059c1225b662e98d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 269, + 140, + 342, + 194 + ], + "blocks": [ + { + "bbox": [ + 269, + 140, + 342, + 194 + ], + "lines": [ + { + "bbox": [ + 269, + 140, + 342, + 194 + ], + "spans": [ + { + "bbox": [ + 269, + 140, + 342, + 194 + ], + "type": "image", + "image_path": "e6cb6b63cdebd7c68cac9e965fd2a6f73a7a5988aabedad6e0c93decc585a684.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 352, + 82, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 352, + 82, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 352, + 82, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 352, + 82, + 425, + 137 + ], + "type": "image", + "image_path": "b82bf923370b7e448d05962b093e8527b85a588d264da59a71f2598726127007.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 352, + 140, + 425, + 194 + ], + "blocks": [ + { + "bbox": [ + 352, + 140, + 425, + 194 + ], + "lines": [ + { + "bbox": [ + 352, + 140, + 425, + 194 + ], + "spans": [ + { + "bbox": [ + 352, + 140, + 425, + 194 + ], + "type": "image", + "image_path": "0e5c9d9402330593fcaa9694cd98b254e4578532d5182fe13fb14e2f475272f1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 435, + 82, + 507, + 137 + ], + "blocks": [ + { + "bbox": [ + 435, + 82, + 507, + 137 + ], + "lines": [ + { + "bbox": [ + 435, + 82, + 507, + 137 + ], + "spans": [ + { + "bbox": [ + 435, + 82, + 507, + 137 + ], + "type": "image", + "image_path": "0877529792eead23cc0daf2ddbf1c68c9bc47b68cf6427478b1c4696478eeae8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 104, + 533, + 114 + ], + "lines": [ + { + "bbox": [ + 509, + 104, + 533, + 114 + ], + "spans": [ + { + "bbox": [ + 509, + 104, + 533, + 114 + ], + "type": "text", + "content": "SAM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 140, + 507, + 194 + ], + "blocks": [ + { + "bbox": [ + 435, + 140, + 507, + 194 + ], + "lines": [ + { + "bbox": [ + 435, + 140, + 507, + 194 + ], + "spans": [ + { + "bbox": [ + 435, + 140, + 507, + 194 + ], + "type": "image", + "image_path": "a487c2c4681487141f4b51ee07634b51e347047cf1e00c0cb2920aeaf4c2e426.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 163, + 547, + 173 + ], + "lines": [ + { + "bbox": [ + 509, + 163, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 509, + 163, + 547, + 173 + ], + "type": "text", + "content": "FocSAM" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 70, + 176, + 175, + 255 + ], + "blocks": [ + { + "bbox": [ + 70, + 176, + 175, + 255 + ], + "lines": [ + { + "bbox": [ + 70, + 176, + 175, + 255 + ], + "spans": [ + { + "bbox": [ + 70, + 176, + 175, + 255 + ], + "type": "image", + "image_path": "e3fb9e5fb07c0f22f23feb1a35428e21270d72947d97b32312fa69b9747765fe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 255, + 148, + 266 + ], + "lines": [ + { + "bbox": [ + 94, + 255, + 148, + 266 + ], + "spans": [ + { + "bbox": [ + 94, + 255, + 148, + 266 + ], + "type": "text", + "content": "GroundTruth" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 289, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 54, + 289, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 555, + 322 + ], + "type": "text", + "content": "Figure 3. Qualitative analysis on a challenge Example. The far-left image presents a challenging example featuring the image and ground truth (GT) with blue masks. The top, middle, and bottom rows on the right respectively display the segmentation results of SAM, FocSAM, and our method, at the 5th, 10th, 15th, and 20th clicks." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 186, + 198, + 260, + 254 + ], + "blocks": [ + { + "bbox": [ + 186, + 198, + 260, + 254 + ], + "lines": [ + { + "bbox": [ + 186, + 198, + 260, + 254 + ], + "spans": [ + { + "bbox": [ + 186, + 198, + 260, + 254 + ], + "type": "image", + "image_path": "47704268c36e4f69999936d72cc627545350e458c2fb4f34f7b95b0468055761.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 255, + 242, + 266 + ], + "lines": [ + { + "bbox": [ + 209, + 255, + 242, + 266 + ], + "spans": [ + { + "bbox": [ + 209, + 255, + 242, + 266 + ], + "type": "text", + "content": "5st click" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 269, + 198, + 343, + 254 + ], + "blocks": [ + { + "bbox": [ + 269, + 198, + 343, + 254 + ], + "lines": [ + { + "bbox": [ + 269, + 198, + 343, + 254 + ], + "spans": [ + { + "bbox": [ + 269, + 198, + 343, + 254 + ], + "type": "image", + "image_path": "a50e4f75c861028fa567763c2a29f7c2b99e20a559fa258fbceb27dffe264010.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 255, + 328, + 266 + ], + "lines": [ + { + "bbox": [ + 291, + 255, + 328, + 266 + ], + "spans": [ + { + "bbox": [ + 291, + 255, + 328, + 266 + ], + "type": "inline_equation", + "content": "10^{\\mathrm{st}}" + }, + { + "bbox": [ + 291, + 255, + 328, + 266 + ], + "type": "text", + "content": " click" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 352, + 198, + 425, + 254 + ], + "blocks": [ + { + "bbox": [ + 352, + 198, + 425, + 254 + ], + "lines": [ + { + "bbox": [ + 352, + 198, + 425, + 254 + ], + "spans": [ + { + "bbox": [ + 352, + 198, + 425, + 254 + ], + "type": "image", + "image_path": "9f82255e2ddeac8fba554051311aa15effc76c94fab6c6ea464e69e9e062a736.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 255, + 410, + 266 + ], + "lines": [ + { + "bbox": [ + 373, + 255, + 410, + 266 + ], + "spans": [ + { + "bbox": [ + 373, + 255, + 410, + 266 + ], + "type": "inline_equation", + "content": "15^{\\mathrm{st}}" + }, + { + "bbox": [ + 373, + 255, + 410, + 266 + ], + "type": "text", + "content": " click" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 435, + 198, + 507, + 254 + ], + "blocks": [ + { + "bbox": [ + 435, + 198, + 507, + 254 + ], + "lines": [ + { + "bbox": [ + 435, + 198, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 435, + 198, + 507, + 254 + ], + "type": "image", + "image_path": "5c3c6fec5c4e0bb1267f2988f54ee3cfaeed765c0ef32500e513a9d26de50709.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 255, + 490, + 266 + ], + "lines": [ + { + "bbox": [ + 452, + 255, + 490, + 266 + ], + "spans": [ + { + "bbox": [ + 452, + 255, + 490, + 266 + ], + "type": "inline_equation", + "content": "20^{\\mathrm{st}}" + }, + { + "bbox": [ + 452, + 255, + 490, + 266 + ], + "type": "text", + "content": " click" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 510, + 221, + 531, + 231 + ], + "lines": [ + { + "bbox": [ + 510, + 221, + 531, + 231 + ], + "spans": [ + { + "bbox": [ + 510, + 221, + 531, + 231 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 343, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 343, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 343, + 296, + 438 + ], + "type": "text", + "content": "ing ViT-H as the backbone. However, as the early fusion method, SimpleClick requires a heavy image encoding at each interaction, whereas our approach is much more time-efficient, with only " + }, + { + "bbox": [ + 54, + 343, + 296, + 438 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 54, + 343, + 296, + 438 + ], + "type": "text", + "content": " of SimpleClick's SPC. Additionally, we provide ViT-B results, showing it matches the speed (0.278 SPC) and accuracy (4.75 NoC90 on DAVIS) of the fastest early fusion method FocalClick [4] (0.232 SPC and 4.90 NoC90 on DAVIS)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 445, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 296, + 529 + ], + "type": "text", + "content": "By reintroducing prompt-image synergy on late fusion, our method achieves state-of-the-art accuracy compared to all late fusion methods on these datasets. Specifically, under the same settings, we also outperform other SAM-based methods, such as FocSAM [13] and HQ-SAM [14]. We only add 0.1s SPC compared to SAM, which still maintain the high efficiency of late fusion." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": "Performance on high-quality benchmark. In Tab. 2, we compare our method with state-of-the-art approaches on high-quality dataset HQSeg-44K [14](1537 samples). We compare two types of baselines: SAM-based methods and other methods. As reported, SAM-REF achieves state-ofthe-art performance to the two baselines, which demonstrate that SAM-REF can effectively enhances SAM's performance to top-tier levels in high-quality interactive segmentation while maintaining the low latency. Other latefusion methods (InterFormer [12] and SegNext [22]), which rely solely on the dense presentation of visual prompts, are more costly than SAM's use of the sparse representation [22]. Our method combines the benefits of both prompt types. It only focuses on detail extraction using dense maps of visual prompts, without needing complex fusion of visual" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 343, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 390 + ], + "type": "text", + "content": "prompts and image embeddings. This demonstrates that our method can effectively enhance the performance of late fusion in complex scenarios, as evidenced by Qualitative Results." + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 347, + 400, + 523, + 441 + ], + "blocks": [ + { + "bbox": [ + 347, + 400, + 523, + 441 + ], + "lines": [ + { + "bbox": [ + 347, + 400, + 523, + 441 + ], + "spans": [ + { + "bbox": [ + 347, + 400, + 523, + 441 + ], + "type": "table", + "html": "
MethodParams/MBFPSMem/G
SAM [15]635.61.703.7
SAM-REF636.31.673.7
", + "image_path": "8fd3d304029826c67977c4bbee8a81be0cd71db3049d0ac287100baf1958a093.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 327, + 449, + 541, + 460 + ], + "lines": [ + { + "bbox": [ + 327, + 449, + 541, + 460 + ], + "spans": [ + { + "bbox": [ + 327, + 449, + 541, + 460 + ], + "type": "text", + "content": "Table 3. Computation analysis for SAM and SAM-REF" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "text", + "content": "Computation analysis. In Tab 3, we report the comparison of model parameters (Params), inference time per image (FPS), and GPU memory (Mem) usage between ViT-H based on SAM and SAM-REF. While SAM-REF produces substantially better segmentation quality, our method adds just 0.7MB to the model parameters compared to SAM, with negligible increases in GPU memory usage and inference time per image." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 585, + 443, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 443, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 443, + 598 + ], + "type": "text", + "content": "5. Ablation Experiments" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "content": "We conduct plenty of ablation studies, reporting experimental results for ViT-H models trained on COCO+LVIS and tested on DAVIS." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "Component Module Analysis. To demonstrate the effect of our model's modules, i.e., GFR, LFR, and the Dynamic Selector, we perform an ablation study on whether and which mask to paste " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ". We have set up 4 experiments. One directly outputs " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", one always pastes " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_r^G" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", one always pastes " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_{t-1}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", and one" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19362" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 271, + 139 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 271, + 139 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 271, + 139 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 271, + 139 + ], + "type": "table", + "html": "
MethodNoC90NoC95
1no \\( M_r^L \\) only \\( M_r^G \\)4.669.53
2paste \\( M_r^L \\) to \\( M_{t-1} \\)4.809.75
3paste \\( M_r^L \\) to \\( M_r^G \\)4.609.40
4Dynamic Selector4.549.10
", + "image_path": "5f70f52c0b1d852954b1f6f4f4d0bb8baa965249102bd5081272002e63bf8317.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "type": "text", + "content": "pastes " + }, + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "type": "text", + "content": " according to the Dynamic Selector. For each experiment, we test NoC90 and NoC95 to evaluate the accuracy. We have shown the results in Tab. 4. The first experiment has no local refinement and always modifies the full mask, which leads to mutual influence between interactions and cannot achieve high accuracy. The second experiment only refines a local region but not the full mask in each interaction, which makes it modify slowly and needs more interactions to achieve the same accuracy. The third experiment with local refinement has a higher accuracy based on 1-th, but still cannot solve the mutual influence. The fourth experiment pasting " + }, + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 54, + 190, + 296, + 360 + ], + "type": "text", + "content": " with Dynamic Selector avoids the interference and refines local regions, achieving the best accuracy." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 92, + 369, + 260, + 433 + ], + "blocks": [ + { + "bbox": [ + 55, + 147, + 295, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 170 + ], + "type": "text", + "content": "Table 4. Ablation study for the refiners and the Dynamic Se-. lector. The methods indicate which mask to paste " + }, + { + "bbox": [ + 55, + 147, + 295, + 170 + ], + "type": "inline_equation", + "content": "M_r^L" + }, + { + "bbox": [ + 55, + 147, + 295, + 170 + ], + "type": "text", + "content": " onto." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 369, + 260, + 433 + ], + "lines": [ + { + "bbox": [ + 92, + 369, + 260, + 433 + ], + "spans": [ + { + "bbox": [ + 92, + 369, + 260, + 433 + ], + "type": "table", + "html": "
MethodSPCNoC90NoC95
ResBlock×10.4834.579.94
ResBlock×30.5114.549.10
ResBlock×50.5454.669.88
ResBlock×80.7104.6110.61
", + "image_path": "bd7295180ba865b73c3ba4fd22c8c9371d3346123285914a4e90bd09887a6312.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 441, + 278, + 453 + ], + "lines": [ + { + "bbox": [ + 72, + 441, + 278, + 453 + ], + "spans": [ + { + "bbox": [ + 72, + 441, + 278, + 453 + ], + "type": "text", + "content": "Table 5. Ablation study for the number of ResBlocks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 466, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 561 + ], + "type": "text", + "content": "Number of ResBlocks. Tab. 5 shows the impact of the number of ResBlocks on accuracy and speed. With three ResBlocks, our method achieves the best results. Fewer blocks fail to capture details effectively, while more blocks reduce speed and harm the semantic capabilities of SAM image embeddings, thus decreasing accuracy. This experiment demonstrates the rationale for setting the number of ResBlocks to 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 164, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 164, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 164, + 586 + ], + "type": "text", + "content": "6. Qualitative Result" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "In Fig. 3, we show the qualitative results compared to SAM [15] and FocSAM [13] on a challenge Example. This visualization demonstrates SAM-REF's enhanced precision. Our method achieves the highest mIoU in all clicks. Both FocSAM and our method are trained on " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO + LVIS}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " and freeze the SAM's image encoder parameters. In the later clicks, our method can segment the detailed shape of the wheel, while SAM and FocSAM cannot. Based on challenging scenarios, the original SAM results are unsatisfactory. While FocSAM enhances stability and prevents pro" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 73, + 376, + 127 + ], + "blocks": [ + { + "bbox": [ + 326, + 73, + 376, + 127 + ], + "lines": [ + { + "bbox": [ + 326, + 73, + 376, + 127 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 376, + 127 + ], + "type": "image", + "image_path": "ac332cd1beb006877e680b08b9847a0de0b3108e64c1e0f913b47c9a71d68e7e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 385, + 73, + 435, + 128 + ], + "blocks": [ + { + "bbox": [ + 385, + 73, + 435, + 128 + ], + "lines": [ + { + "bbox": [ + 385, + 73, + 435, + 128 + ], + "spans": [ + { + "bbox": [ + 385, + 73, + 435, + 128 + ], + "type": "image", + "image_path": "1e96dac20c4b5f8ba666d37d3bcedd089466fb93feb9b22227972b36ef7acdd1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 438, + 73, + 486, + 128 + ], + "blocks": [ + { + "bbox": [ + 438, + 73, + 486, + 128 + ], + "lines": [ + { + "bbox": [ + 438, + 73, + 486, + 128 + ], + "spans": [ + { + "bbox": [ + 438, + 73, + 486, + 128 + ], + "type": "image", + "image_path": "ce41af0e62a013835694fba5b9e85d22eb461cbdc7a3ee8ee40949c00ca14f0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 492, + 73, + 539, + 128 + ], + "blocks": [ + { + "bbox": [ + 492, + 73, + 539, + 128 + ], + "lines": [ + { + "bbox": [ + 492, + 73, + 539, + 128 + ], + "spans": [ + { + "bbox": [ + 492, + 73, + 539, + 128 + ], + "type": "image", + "image_path": "82ef2b81520563c9a83260a09becc5fd15a1139d9b98fa035a1b8e6305839ab4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 137, + 376, + 191 + ], + "blocks": [ + { + "bbox": [ + 326, + 137, + 376, + 191 + ], + "lines": [ + { + "bbox": [ + 326, + 137, + 376, + 191 + ], + "spans": [ + { + "bbox": [ + 326, + 137, + 376, + 191 + ], + "type": "image", + "image_path": "c9b9e0ceeff6c8cc946b023e9c63d11d76b0b5152100378a0c3f616c36bb245e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 266 + ], + "type": "text", + "content": "Figure 4. Qualitative results for the effectiveness of GlobalFusion and LocalFusion Refiner. The first row shows the global refinement results of GFR. The red box indicates the region for local refinement. The second row shows the local refinement results of LFR." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 385, + 137, + 435, + 191 + ], + "blocks": [ + { + "bbox": [ + 385, + 137, + 435, + 191 + ], + "lines": [ + { + "bbox": [ + 385, + 137, + 435, + 191 + ], + "spans": [ + { + "bbox": [ + 385, + 137, + 435, + 191 + ], + "type": "image", + "image_path": "b6acf7bd555576c2a1b5fca038e6c775e26e46380d08dd887593dbf8389b6ffb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 191, + 484, + 202 + ], + "lines": [ + { + "bbox": [ + 384, + 191, + 484, + 202 + ], + "spans": [ + { + "bbox": [ + 384, + 191, + 484, + 202 + ], + "type": "text", + "content": "Coarse mask Error map" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 438, + 137, + 487, + 191 + ], + "blocks": [ + { + "bbox": [ + 438, + 137, + 487, + 191 + ], + "lines": [ + { + "bbox": [ + 438, + 137, + 487, + 191 + ], + "spans": [ + { + "bbox": [ + 438, + 137, + 487, + 191 + ], + "type": "image", + "image_path": "f597bf79370772830588684d7015f7f106f982a327e8d0b0be521f0e3c59e9bd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 491, + 137, + 541, + 191 + ], + "blocks": [ + { + "bbox": [ + 491, + 137, + 541, + 191 + ], + "lines": [ + { + "bbox": [ + 491, + 137, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 491, + 137, + 541, + 191 + ], + "type": "image", + "image_path": "6acd82e30ebdacc43f7a3a231aca03434d0d857417329cef311efe427c1c0323.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 191, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 488, + 191, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 488, + 191, + 545, + 202 + ], + "type": "text", + "content": "Refined mask" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 292, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 340 + ], + "type": "text", + "content": "gressive deterioration, it remains limited by image embeddings, potentially missing object details. Our method retrieves lost object details from the original image through image-prompt synergy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 342, + 556, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 473 + ], + "type": "text", + "content": "We present the results at each stage of our method in Fig. 4. The coarse mask is generated by the SAM decoder. The first row demonstrates how GFR refines the global segmentation process. The error map presents the error regions within SAM's predicted areas. In this scenario, the error regions are concentrated around the object edges and thin structures. Then, GFR specifically refines these error regions. The second row demonstrates the refinement process of LFR. It focuses on refining the areas with the largest errors between the interactions. In this image, the errors are mainly concentrated in the hollow regions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 488, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 388, + 502 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 511, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 643 + ], + "type": "text", + "content": "In view that early fusion strategies in the interactive segmentation can achieve high accuracy and late fusion strategies can achieve high efficiency, we propose a new method called SAM-REF, which combines the advantages of the two strategies. Our method redesigns the SAM-based interactive segmentation pipeline, fusing images and prompts from both global and local perspectives. This adaptation fully taps into SAM's potential for interactive segmentation. SAM-REF outperforms current state-of-the-art methods while maintaining SAM's low latency, even in challenging scenarios." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 658, + 424, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 424, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 424, + 672 + ], + "type": "text", + "content": "8. Acknowledgement" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 679, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 703 + ], + "type": "text", + "content": "Xiaolin Hu was supported by the National Natural Science Foundation of China (No. U2341228)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19363" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] George Barnum, Sabera Talukder, and Yisong Yue. On the benefits of early fusion in multimodal representation learning. arXiv preprint arXiv:2011.07191, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "text", + "content": "[2] Rodrigo Benenson, Stefan Popov, and Vittorio Ferrari. Large-scale interactive object segmentation with human annotators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11700-11709, 2019. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 224 + ], + "type": "text", + "content": "[3] Xi Chen, Zhiyan Zhao, Feiwu Yu, Yilei Zhang, and Manni Duan. Conditional diffusion for interactive segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7345-7354, 2021. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 296, + 280 + ], + "type": "text", + "content": "[4] Xi Chen, Zhiyan Zhao, Yilei Zhang, Manni Duan, Donglian Qi, and Hengshuang Zhao. Focalclick: Towards practical interactive image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1300-1309, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 346 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 391 + ], + "type": "text", + "content": "[6] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 457 + ], + "type": "text", + "content": "[7] Yuying Hao, Yi Liu, Zewu Wu, Lin Han, Yizhou Chen, Guowei Chen, Lutao Chu, Shiyu Tang, Zhiliang Yu, Zeyu Chen, et al. Edgeflow: Achieving practical interactive segmentation with edge-guided flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1551-1560, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "type": "text", + "content": "[8] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In 2011 international conference on computer vision, pages 991-998. IEEE, 2011. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 580 + ], + "type": "text", + "content": "[10] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "text", + "content": "[11] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "type": "text", + "content": "[12] You Huang, Hao Yang, Ke Sun, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, and Rongrong Ji. Interformer: Real-time interactive image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22301-22311, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[13] You Huang, Zongyu Lan, Liujuan Cao, Xianming Lin, Shengchuan Zhang, Guannan Jiang, and Rongrong Ji. Foc" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "sam: Delving deeply into focused objects in segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3120-3130, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 272 + ], + "type": "text", + "content": "[16] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 327 + ], + "type": "text", + "content": "[17] Jiacheng Lin, Jiajun Chen, Kailun Yang, Alina Roitberg, Siyu Li, Zhiyong Li, and Shutao Li. Adaptiveclick: Click-aware transformer with adaptive focal loss for interactive image segmentation. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 327, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 393 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 393, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 553, + 449 + ], + "type": "text", + "content": "[19] Zheng Lin, Zheng-Peng Duan, Zhao Zhang, Chun-Le Guo, and Ming-Ming Cheng. Focuscut: Diving into a focus view in interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2637-2646, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 504 + ], + "type": "text", + "content": "[20] Qin Liu, Meng Zheng, Benjamin Planche, Srikrishna Karanam, Terrence Chen, Marc Niethammer, and Ziyan Wu. Pseudoclick: Interactive image segmentation with click imitation. In European Conference on Computer Vision, pages 728-745. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 558 + ], + "type": "text", + "content": "[21] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. Simpleclick: Interactive image segmentation with simple vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22290-22300, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 614 + ], + "type": "text", + "content": "[22] Qin Liu, Jaemin Cho, Mohit Bansal, and Marc Niethammer. Rethinking interactive image segmentation with low latency high quality and diverse prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3773-3782, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 647 + ], + "type": "text", + "content": "[23] Kevin McGuinness and Noel E O'connor. A comparative evaluation of interactive segmentation algorithms. Pattern Recognition, 43(2):434-444, 2010. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 2, 5, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19364" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[25] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[26] Konstantin Sofiuk, Ilia Petrov, Olga Barinova, and Anton Konushin. f-brs: Rethinking backpropagating refinement for interactive segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8623-8632, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[27] Konstantin Sofiuk, Ilya A Petrov, and Anton Konushin. Reviving iterative training with mask guidance for interactive segmentation. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3141-3145. IEEE, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 251 + ], + "type": "text", + "content": "[28] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[29] Zhaozhi Xie, Bochen Guan, Weihao Jiang, Muyang Yi, Yue Ding, Hongtao Lu, and Lei Zhang. Pa-sam: Prompt adapter sam for high-quality image segmentation. arXiv preprint arXiv:2401.13051, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[30] Ning Xu, Brian Price, Scott Cohen, Jimei Yang, and Thomas S Huang. Deep interactive object selection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 373-381, 2016. 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19365" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_content_list.json b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4346ae80673609030f3adb9236c3fe255896f08f --- /dev/null +++ b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_content_list.json @@ -0,0 +1,1376 @@ +[ + { + "type": "text", + "text": "SAM2-LOVE: Segment Anything Model 2 in Language-aided Audio-Visual Scenes", + "text_level": 1, + "bbox": [ + 187, + 128, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuji Wang $^{1*}$ , Haoran Xu $^{2*}$ , Yong Liu $^{1}$ , Jiaze Li $^{2}$ , Yansong Tang $^{1\\dagger}$", + "bbox": [ + 243, + 202, + 756, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University, $^{2}$ ZJU", + "bbox": [ + 186, + 220, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuji-wan24@mails.tsinghua.edu.cn, †tang.yansong@sz.tsinghua.edu.cn", + "bbox": [ + 202, + 241, + 787, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reference Audio-Visual Segmentation (Ref-AVS) aims to provide a pixel-wise scene understanding in Language-aided Audio-Visual Scenes (LAVS). This task requires the model to continuously segment objects referred to by text and audio from a video. Previous dual-modality methods always fail due to the lack of a third modality and the existing triple-modality method struggles with spatio-temporal consistency, leading to the target shift of different frames. In this work, we introduce a novel framework, termed SAM2-LOVE, which integrates textual, audio, and visual representations into a learnable token to prompt and align SAM2 for achieving Ref-AVS in the LAVS. Technically, our approach includes a multimodal fusion module aimed at improving multimodal understanding of SAM2, as well as token propagation and accumulation strategies designed to enhance spatio-temporal consistency without forgetting historical information. We conducted extensive experiments to demonstrate that SAM2-LOVE outperforms the SOTA by $8.5\\%$ in $\\mathcal{J}\\&\\mathcal{F}$ on the Ref-AVS benchmark and showcase the simplicity and effectiveness of the components. Our code will be available here.", + "bbox": [ + 88, + 323, + 485, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 671, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the process of human perception of the world, visual signals play a primary role, often accompanied by other multimodal signals such as audio and text. These additional multimodal cues are not only essential to locating and separating objects and areas of interest but also to understanding the changing behaviors of the objects. The dynamic scene that integrates visual elements with diverse multimodal cues including audio and text is regarded as the Language-aided Audio-Visual Scene (LAVS). To mimic the effect from humans, prior works such as Music-AVQA [10] and AVQA [39] have primarily focused on question-answering", + "bbox": [ + 89, + 696, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d43998c6c8420773f1678dab1fa83c90aa7d0dd726802510d6d658a6cdbe7cc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 296, + 720, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06d5c19d02f822cd0b15cd3d50fa6411708bdf91eb3db7d58eb3f748a96c7946.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 296, + 916, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bcfeafbdb1a56205f9621532a4faaed69a074b05c106f118d0bea3553142c43.jpg", + "image_caption": [ + "Figure 1. The comparison with dual-modality SAM pipelines. (a) The text-visual SAM struggles to distinguish the sounding object in the silent scenes. (b) The audio-visual SAM fails to understand the dynamic control from the text input. (c) Our model fuses three modalities' information and expands to the video due to SAM2." + ], + "image_footnote": [], + "bbox": [ + 519, + 493, + 911, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks. Nevertheless, these approaches lack a fine-grained understanding of LVAS at the pixel level. As shown in Figure 2, Reference Audio-Visual Segmentation (Ref-AVS) task [31] makes a good example to improve the pixel-wise", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28932", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16be766b43381123fdbeb8a4fffae0994bc3b2499bb896eabe1d7763492a799e.jpg", + "image_caption": [ + "Figure 2. Illustration of the Ref-AVS task under varying modality cues. The expression can be treated as a kind of control signal to condition the referred object on audio. Different expressions can test the models' pixel-wise scene understanding capabilities in the LAVS with different dimensions. The absence angle of a triangle means that modality is not the main test dimension." + ], + "image_footnote": [], + "bbox": [ + 116, + 92, + 879, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene understanding of the model in the LAVS. The previous duality-referring segmentation works struggle with the Ref-AVS task where the interested object can only be located under triple modalities, shown in Figure 1. In neither Figure 1 (a) text-visual EVF-SAM [43] nor Figure 1 (b) audio-visual GAVS [30], the truly referred target is accurately segmented due to the absence of additional help from the other third modality. Additionally, they are constrained in the image level, which can not be directly adapted to the dynamic nature of video content. Therefore, methods for effectively fusing information from the three modalities for segmentation have garnered widespread attention.", + "bbox": [ + 88, + 398, + 482, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior work, EEMC [31], simultaneously models three modalities' representations to achieve excellent results on the Ref-AVS benchmark. However, it showcases the relatively insufficient spatio-temporal consistency. Specifically, despite the existence of memory caches, the model fails to continuously track the target's position and shape, leading to a shift in the segmented area over time. Recently, SAM2 [22] has been proposed to generalize the segmenting anything capabilities of the image to the video, which demonstrates a powerful video tracking capability in interframe consistency. SAM2 follows the previous promptable visual segmentation of SAM [8] that requires a strong manual annotation to the interested target of the starting frame (not only limited to the first frame) and then propagates to the whole video based on the crucial frame. Nevertheless, the lack of dependency on other modalities (text or audio) constraints its further application in the LAVS.", + "bbox": [ + 89, + 580, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, in this paper, we propose a novel framework, namely SAM2-LOVE, which means Segment Anything Model 2 in Language-aided audio-O-Visual scEnes. To enhance SAM2's understanding of multimodal information,", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we design the multimodal fusion module to compress all the modalities' information into a learnable token that aligns the SAM2 for segmenting the referring target. Additionally, we incorporate token propagation and accumulation strategies to strengthen spatio-temporal understanding of the video while avoiding forgetting historical frames. Our main contributions can be summarized as follows:", + "bbox": [ + 511, + 398, + 906, + 503 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel framework, SAM2-LOVE that firstly leverages SAM2 to achieve pixel-wise understanding in the LAVS by designing a multimodal fusion module.", + "- We develop creative token propagation and accumulation strategies to improve spatio-temporal comprehension of the promtable token.", + "- Extensive experiments on Ref-AVS dataset demonstrate the superiority of our method, with ablation studies highlighting the simplicity and effectiveness of its modules." + ], + "bbox": [ + 513, + 506, + 903, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 661, + 661, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Segmentation via Language or Audio. Unlike semantic segmentation [1, 17, 29, 32], visual segmentation via language starts with referring expression segmentation task [7, 18, 21, 33, 34, 36, 40, 41] which performs target object segmentation based on given natural language expressions. X-Decoder [47] bridges vision and language, unifying multiple tasks within a single model. Moreover, GRES [13] defines a new task called generalized reference expression segmentation. With the development of multimodal large language models (MLLMs) such as LLaVA [14, 15], LISA [9] proposes the concept of reasoning segmentation with combination of pixel-level understanding and MLLMs. GSVA [37] and PixelLM [23] propose multi-target and null-target reasoning segmentation. On the other", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28933", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hand, audio-visual segmentation (AVS) is also constantly improving. Audio-Visual Segmentation [45, 46] proposes the AVS problem and AVSBench dataset. Numerous subsequent methods [2, 4, 27] have been developed and refined based on the proposed benchmark.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segment Anything Model. SAM [8] is an interactive segmentation model based on various types of prompts (points, boxes, coarse masks). It shows significant generalization ability in open-world segmentation. Several works strive to reduce the massive computation cost of SAM, such as EfficientSAM [38], FastSAM [44], and MobileSAM [42]. The advent of SAM2 [22] extends SAM's functionality to include video segmentation. While SAM and SAM2 perform effectively in visual segmentation tasks using box, point, or mask prompts, they currently do not possess language understanding capabilities, rendering the direct use of text prompts for referring segmentation infeasible. EVF-SAM [43] extends SAM's capabilities with text-prompted segmentation, while GAVS [30] equips the SAM with image-audio fusion to achieve AVS.", + "bbox": [ + 89, + 167, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-aided Audio-Visual Scene. There is presently a scarcity of publicly available tasks that provide datasets for audio-visual scene comprehension with language-based support. The complexity of language-aided audio-visual scene understanding arises from its reliance on three modalities (vision, audio, and text) necessitating massive matched data and their corresponding labels, posing significant challenges for data collection and model design. MusicAVQA [10] and AVQA [39] are two of the classic audio-visual question-answering (QA) datasets. However, the above work focuses on multimodal QA and spatio-temporal reasoning of audio-visual scenes involving various relationships between objects or activities. Ref-AVS [30, 31] begins pixel-level understanding of dynamic audio-visual scenes.", + "bbox": [ + 89, + 395, + 482, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, no methods exist for integrating the foundational visual segmentation model SAM2 into language-based audio-visual scenes. Therefore, to achieve pixel-level understanding in the LAVS, we introduce SAM2-LOVE.", + "bbox": [ + 89, + 607, + 482, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 684, + 181, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of SAM2", + "text_level": 1, + "bbox": [ + 89, + 709, + 292, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM2 can be treated as a generalization of SAM to segment anything in images and videos. It also supports visual prompts (point, box, and mask) to indicate the objects to be segmented in an individual frame (image as a single frame). Then, it generalizes across the whole video based on the prompted frame to achieve the video segmentation.", + "bbox": [ + 89, + 733, + 482, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 3 (Inference), this model begins with a hierarchical image encoder that encodes each frame into multilevel embeddings. Unlike SAM which directly feeds these features to the mask decoder for predictions, SAM2 incorporates a series of transformer blocks (i.e. memory", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attention) to condition the current image feature on memories of previous and prompted frames stored in a memory bank that is maintained by a FIFO (First In First Out) queue. Each memory entry is composed of two kinds of memories: (1) spatial feature maps that are generated through an element-wise summation of the downsampled and encoded mask features from the memory encoder and the unconditioned frame embedding from the image encoder, and (2) the object pointers that are obtained from the output tokens within the mask decoder contain high-level semantic information of the interested target. Therefore, the memory attention operation in the transformer makes the current frame's feature cross-attend both spatial fine-grained memory and abstract temporal object pointers. Second, the mask decoder of SAM2 largely follows the design of SAM, which is a \"two-way\" transformer that takes the prompt embeddings from the prompt encoder and frame embeddings as inputs. The difference lies in the additional occlusion score that aims to judge the object's presence in the current frame.", + "bbox": [ + 511, + 90, + 903, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thanks to the strong and generalized capability of SAM2 across images and videos, we can first compress all the information contained in three modalities into a single token and then use it to prompt SAM2 on a single frame. Finally, starting from this crucial frame, we can propagate through video to obtain the desired masks of all frames.", + "bbox": [ + 511, + 377, + 903, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. SAM2-LOVE", + "text_level": 1, + "bbox": [ + 511, + 477, + 653, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of our model is shown in Figure 3, which integrates multimodal information in the LAVS into a representative token and uses it to provide a visual clue for SAM2 to segment the reference targets.", + "bbox": [ + 511, + 500, + 903, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Representations. Following the previous work [31], for each modality, we use three popular encoders to separately extract their representations, which were not aligned before in their pretraining stages. VGGish [5, 6] encodes an audio clip into audio embeddings $F_{A} \\in R^{N \\times d_{A}}$ , where $N$ denotes the duration of the audio in seconds and corresponds to the total number of frames in the video. The sampled $N$ frames are encoded into $F_{V} \\in R^{N \\times d_{V} \\times h \\times w}$ by a ViT [3] pre-trained on ImageNet. For the text expressions, we utilize DistilRoBERTa [16, 24] to obtain the textual representations $F_{T} \\in R^{P \\times d_{T}}$ , where $P$ is the number of textual tokens. To unify and align all embeddings into the same semantic space, we use three MLPs to project them into a shared dimension respectively, shown in Equation (1).", + "bbox": [ + 511, + 560, + 905, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {F} _ {Z} = \\operatorname {L i n e a r} (\\operatorname {R e L U} (\\operatorname {L i n e a r} (F _ {Z}))), Z \\in \\{A, V, T \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 780, + 903, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Fusion. As shown in Figure 3, a multimodal fusion module is proposed to compress all three modalities' information into a single token and then leverage its embedding to represent the fused features. We define this token as a learnable segmentation token, [seg], designed to prompt SAM2 as a clue to retrieve the targeted", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28934", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4768acffa4521db7f0bbbd66a3a58d4a7e5f2a3e62165d8db31ba80981dfa202.jpg", + "image_caption": [ + "Figure 3. The pipeline of SAM2-LOVE. We harness a fusion transformer to compress the tripe modality representations into a learnable token [seg]. For the input, [aud] and [vis] are added to indicate their corresponding modality sequences, and [his] is to provide additional information about former frames. In the training stage, we use the first frame to align the SAM2 with the compressed and fused representation of [seg]. During inference, we harness the strong zero-shot capabilities of SAM2 to track the referred target at the pixel level." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 890, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "object. The fusion module is a naive transformer encoder architecture with $L$ stacked bi-directional transformer block that contains a multi-head self-attention (MSA), Layernorm (LN), and MLP, formulated below:", + "bbox": [ + 89, + 585, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {l} ^ {\\prime} = M S A \\left(L N \\left(z _ {l - 1}\\right)\\right) + z _ {l - 1}, \\quad l = 1 \\dots L \\tag {2} \\\\ z _ {l} = M L P (L N (z _ {l} ^ {\\prime})) + z _ {l} ^ {\\prime}, \\quad l = 1 \\dots L \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 650, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{l-1}$ is the input sequence of $l$ -th layer.", + "bbox": [ + 89, + 688, + 393, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Propagation. The behavior of the [seg] token is a self-propagation process within the whole video, which adaptively captures the inner-frame spatial features and the strong continuity among consecutive frames to achieve spatio-temporal consistency. To be more specific, after obtaining the three projected multimodal representations $\\widehat{F}_A$ , $\\widehat{F}_T$ , and $\\widehat{F}_V$ , we first prepend the [seg] token to the head of the sequence of modalities' embeddings to obtain the multimodal sequence $F_M^i$ for the $i$ -th frame, shown as follows:", + "bbox": [ + 89, + 703, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {M} ^ {i} = \\operatorname {C o n c a t} ([ [ s e g ]; \\widehat {F} _ {A}; [ a u d ]; \\widehat {F} _ {T}; [ v i s ]; \\widehat {F} _ {V} ^ {i} ]), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 845, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{Concat}$ denotes the concatenation operation in the feature dimension axis, two fixed tokens, [aud] and [vis]", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "indicate the end and beginning of their corresponding sequences, respectively. This concatenated sequence will be fed into the fusion transformer for updating the status of the [seg] token, allowing the token to capture the visual information of the current frame along with the integral audio and text representations. Since the transformer is the bidirectional encoder that enables all the tokens to interact with each other, we can harness the single token to serve as the multimodal representation of the whole sequence. We take the first element from the output of the fusion transformer $Q_{i} = \\text{Transformer}(F_{M}^{i})$ , as the updated [seg] token $[seg] = Q_{i}^{0}$ , shown in Figure 4. Therefore, the token not only contains the current frame information but also the compressed information of the previous frames. Our proposed token propagation strategy allows it to model the video from both the spatial and temporal aspects and aggregate the information within the text and audio.", + "bbox": [ + 511, + 584, + 906, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Accumulation. Although the token propagation strategy can effectively obtain the multimodal characteristics and be used to prompt the SAM2, it still faces the issue", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28935", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a23f14fa9ce23cfdbee60a3f04933d9d121a88dac2bab81b561f7f256804cba7.jpg", + "image_caption": [ + "Figure 4. The token propagation and token accumulation strategy in the fusion transformer. For the token propagation, the output [seg] of the current step serves as the input for the next step, which interacts with different frame features but the same audio and text features. To offer additional historical information, the token accumulation strategy appends the global representations of all the previous frames to update the maintained history token [his]." + ], + "image_footnote": [], + "bbox": [ + 99, + 95, + 898, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that the information of the first several frames can be diluted and forgotten with the increasing propagation times. To address the issue, we propose the [cls] token accumulation strategy, in which we keep a history of token sequences, [his] to store the global and abstract representation of historical frames, shown in Figure 4. It is achieved that when propagating to the $i$ -th frame, we provide additional global information of all the previous frames by appending and concatenating the [his] tokens into $F_{M}^{i}$ . The updation of the history token [his] can be formulated by Equation (4).", + "bbox": [ + 88, + 284, + 480, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ h i s ] ^ {i} = C o n c a t \\left(\\left[ [ c l s ] ^ {0}; [ c l s ] ^ {1}; \\dots ; [ c l s ] ^ {i} \\right]\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 446, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[cls]^i$ is the global representation of $i$ -th frame from ViT. These historical tokens can be referred to by the $[seg]$ of the current stage, promising the information of all the previous frames is not lost.", + "bbox": [ + 89, + 474, + 483, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unified Perspective of Token Behaviors. We interpret the two token operations described above from a unified knowledge transfer perspective. Token propagation of the [seg] token is a direct flow, which expects the knowledge from different frames to be gradually connected through the propagation process. This is the forward knowledge transfer. In contrast, token accumulation is the replay of [his] tokens, a reversed flow to prevent the [seg] token from forgetting the knowledge of historical frames as it propagates forward. This is the backward knowledge transfer. We propose that these two token operations are functionally opposed yet conceptually unified.", + "bbox": [ + 89, + 536, + 483, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 727, + 194, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trainable Modules. The encoders used for multimodal representations for audio, video, and text are fully frozen, perceiving the pre-trained knowledge. However, the multimodal fusion transformer and [seg] token are fully trainable during our training process, allowing them to learn how to generate the fused and aligned embeddings tailored for prompting SAM2. For SAM2, we keep the image encoder and memory attention modules frozen while training the prompt encoder and mask decoder. Additionally, all the MLPs for projecting are trainable.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy. Our training strategy initially activates SAM2's multimodal prompting capability at the image level and then leverages its extensive pre-trained knowledge for zero-shot video segmentation. In this setting, the fusion transformer and [seg] token are trained with full access to the entire video sequence, while SAM2 is exposed only to the first frame during training. Specifically, we consolidate multimodal information (video, audio, and text) through the fusion module, which compresses these features into tokens that serve as prompts for SAM2 on the first frame, shown in Figure 3 (Training). Then, a supervised loss is applied to this frame to train SAM2's referring segmentation abilities. During testing, both the fusion module and SAM2 receive the entire video, allowing for comprehensive segmentation across all frames. This approach enables SAM2 to effectively utilize multimodal prompts on the initial frame and then generalize to the complete video sequence through its zero-shot Video Object Segmentation (VOS) capabilities, shown in Figure 3 (Inference).", + "bbox": [ + 511, + 284, + 906, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. The model is trained end-to-end under the supervised mask loss $\\mathcal{L}_{mask}$ , to encourage the SAM2 to predict the correct localized masks. To compute this loss, we combine the per-pixel binary cross-entropy (BCE) loss and DICE loss with their corresponding weights $\\lambda_{bce}$ and $\\lambda_{dice}$ , which can be formulated by:", + "bbox": [ + 511, + 571, + 906, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathbf {B C E} (\\hat {M}, M) + \\lambda_ {d i c e} \\mathbf {D I C E} (\\hat {M}, M) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 667, + 906, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{M}$ is a single mask prediction for the first frame and $M$ is the label of the first frame for supervision.", + "bbox": [ + 511, + 690, + 905, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 733, + 645, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 757, + 720, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We evaluate our method on the Ref-AVS benchmark [31], a substantial collection of 4,000 videos with manual pixel-level annotations and 20,000 expressions reflecting the scenarios in audio, vision, and time dimensions. Additionally, the collected videos contain abundant and diverse visual elements. Each video encompasses a high average number of objects about 1.72, a wide variety of audible objects in 48 categories, and 3 kinds of static objects.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28936", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7a66c8bbd9c4c2535713681bb3b04da20bb1cf2488d4538f5afc6b74b40c9583.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSeen (%)Unseen (%)Mix (S+U)NULLVENUE
JFJ&FJFJ&FJFJ&Fs
AVSBench [45]21.245.733.527.750.939.324.548.336.4-ECCV'2022
+text23.251.137.232.454.743.527.852.940.30.208
AVSegFormer [4]29.242.335.732.745.138.930.943.737.3-AAAI'2024
+text33.547.040.236.150.143.134.848.641.70.171
GAVS [28]24.845.535.227.846.036.926.345.736.0-AAAI'2024
+text28.949.839.429.849.739.829.449.839.60.190
ReferFormer [35]27.447.737.524.950.437.626.249.137.6-CVPR'2022
+audio31.350.140.730.448.839.630.949.540.20.176
R2VOS [11]21.330.926.123.943.133.522.637.029.8-ICCV'2023
+audio25.041.033.027.949.838.926.545.435.90.183
EEMC [31]34.251.342.849.564.857.241.958.150.00.007ECCV'2024
SAM2-LOVE (ours)43.551.947.766.572.369.455.062.158.50.23CVPR2025
vs. prev. SOTA+9.3+0.6+4.9+17.0+7.5+12.2+13.1+4.0+8.5--
", + "bbox": [ + 96, + 88, + 911, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison across different methods in Seen, Unseen, Mix (S+U), and NULL settings of Ref-AVS benchmark. The mix indicates the average value of seen and unseen splits.", + "bbox": [ + 89, + 348, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the tested scenarios are more challenging with multiple sound sources and multiple semantics. The dataset is divided into three splits, a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three distinct subsets: 1) seen splits that are composed of the same 39 categories in the training set. 2) unseen splits with 13 additional categories that do not appear in the training stage. 3) null splits to test the model when an expression refers to an object that either does not exist or is not visible in the given context.", + "bbox": [ + 88, + 402, + 480, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with six challenging baselines: AVSBench [45], AVSegFormer [4], GAVS [28], ReferFormer [35], R2VOS [11] and EEMC [31]. AVSBench, AVSegFormer, and GAVS are designed for Audio-Visual Segmentation (AVS) tasks while ReferFormer and R2VOS claim for Referring Video Object Segmentation(Ref-VOS) tasks. We also report the results achieved by offering the third their absent modalities, which is consistent with EEMC. EEMC is specially designed to achieve fine-grained semantic understanding in the LAVS and also proposes the Ref-AVS task.", + "bbox": [ + 88, + 561, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For the training recipe, we adopt the DeepSpeed training engine with ZeRO-2 to train our model on 2 NVIDIA A100 GPUs. The AdamW [19] optimizer is harnessed with the learning rate and weight decay set to 1e-4 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the BCE loss $\\lambda_{bce}$ and Dice loss $\\lambda_{dice}$ are set to 1.0 and 1.0, respectively. The batch size is set to 8 and the gradient accumulation step is set to 2. For the network architecture, the fusion transformer is the stacked 6 transformer blocks with bi-directional self", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "attention. The hidden dimension of the transformer is 1024 and the number of [seg] tokens is set to 1. SAM2 is initialized with the weight of SAM2-Hiera-large.", + "bbox": [ + 511, + 402, + 906, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following [20, 31], we assess the performance of the Ref-AVS benchmark using the Jaccard index $(\\mathcal{I})$ , the F-score $(\\mathcal{F})$ , and their average value $(\\mathcal{I} \\& \\mathcal{F})$ as key evaluation metrics.", + "bbox": [ + 511, + 449, + 906, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 516, + 653, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the comparative results of our method on Ref-AVS benchmarks. Our SAM2-LOVE achieves state-of-the-art (SOTA) performance and significantly outperforms the second-best method EEMC on both seen and unseen test splits. The results indicate that our model has highly generalized capabilities that can be directly used for out-of-domain scenarios. We attribute two reasons leading to these huge performance improvements. Firstly, the fusion transformer can effectively step-by-step compress the multimodal cues into the learnable tokens through the [seg] token propagation strategy and [cls] token accumulation. The generated token contains high semantics as well as positional information, which can accurately prompt SAM2 to segment the inserted targets. Secondly, SAM2 showcases strong zero-shot VOS capabilities, accurately tracking the referred object despite the circumstance where it suddenly disappears and occurs again or similar objects have interacted together.", + "bbox": [ + 511, + 537, + 908, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quality Assessment", + "text_level": 1, + "bbox": [ + 511, + 818, + 702, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 shows the visualization comparative results with the previous SOTA EEMC. In the first case, we can observe that the SAM2 shows strong consistency within different frames, which can track the referred object with high", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28937", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37ca21691a1762d603ff43b7d5ce5cad8e504b1d657f29f9b3d44624c5268885.jpg", + "image_caption": [ + "Figure 5. The visualization results of the referred objects in the Ref-AVS. Notably, we use the legends of the trumpet to represent the different sound volumes of the audio objects, from silent to increasing loudness. More cases can be found in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 99, + 98, + 890, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bd04c94dc8ef108f8a881274ff8583763c4a3664bad4ab55a20c3786af8a57b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIP-T\nCLIP-VRoBERTa\nViT[cls] AccSeen (%)Unseen (%)
JFJ&FJFJ&F
✓✓42.551.246.865.171.468.2
✓✓43.551.947.766.572.369.4
✓42.650.946.766.371.669.0
", + "bbox": [ + 94, + 496, + 500, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mask quality. Additionally, it also demonstrates that the [seg] token can accurately locate the small referring target, while EEMC fails in the first several frames. The second case shows that our SAM2-LOVE is more robust regarding changes in audible objects. EEMC may not distinguish the difference or malfunction in understanding the hidden meaning in the expressions (i.e. \"shorter duration\" instead of the obvious \"the man\").", + "bbox": [ + 89, + 637, + 483, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 771, + 243, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Backbone. Table 2 shows the ablation study of backbones. Compared to CLIP vision and text encoder, RoBERTa and ViT achieve better performance, which can be the difference from the pre-trained paradigm. To align with visual elements, the CLIP textual encoder focuses more on the text with specific semantics such as cat and person. However, in the LAVS scenes, the massive expres", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7af92675ffdf280234d87ad36f1a3d90a6e767032a7883fbc1513a453a30cb69.jpg", + "table_caption": [ + "Table 2. Ablations on backbone design and our proposed [cls] token accumulation strategy." + ], + "table_footnote": [], + "table_body": "
Number of transformer blocksSeen (%)Unseen (%)
JFJ&FJFJ&F
1 layer41.549.445.465.270.968.1
6 layers42.650.946.766.371.669.0
12 layers42.050.746.467.773.370.5
", + "bbox": [ + 540, + 496, + 879, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablations on the number of fusion transformer blocks.", + "bbox": [ + 521, + 580, + 895, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sions tend to provide a kind of control signal without referred category names, e.g., “the one making the sound.” and “the one making the most intense sound.” Therefore, the RoBERTa may demonstrate greater strength in understanding instructions due to expertise in natural language understanding during training.", + "bbox": [ + 511, + 625, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of [cls] Token Accumulation. Since we harness the propagated token to prompt the first frame for SAM2, the attention for the first several frames may weaken during propagation, causing the token to depend only on the last frame. Therefore, in the fusion transformer, we attach the visual global [cls] tokens of previous frames to the current frame to update the history token, avoiding the [seg] token forgetting the information of former frames. Compared to the last row and second row in Table 2, we can observe that this strategy can effectively improve performance for free without additional design or resource consumption.", + "bbox": [ + 511, + 717, + 908, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Fusion Transformer Design. The design of", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28938", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e3cd72259512e56a5024b23afa4e2babf818b3625e20a4eeb7527589598c100.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of [seg] tokensSeen (%)Unseen (%)
JFJ&FJFJ&F
142.650.946.766.371.669.0
441.350.846.065.271.068.1
840.749.845.267.072.269.6
", + "bbox": [ + 102, + 88, + 470, + 175 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf3294315dd1774eb1c641d4962c0712d13976e4c12320c69b109c185b498a3c.jpg", + "table_caption": [ + "Table 4. Ablations on the numbers of [seg] tokens." + ], + "table_footnote": [], + "table_body": "
Fusion StrategySeen (%)Unseen (%)
JFJ&FJFJ&F
learnable token42.650.946.766.371.669.0
mean41.650.145.865.971.668.7
", + "bbox": [ + 104, + 215, + 468, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our fusion transformer is a naive transformer architecture without additional operations such as ROPE [26] or attention enhancement [25]. Table 3 shows the different performances under different numbers of transformer blocks. Increasing the model's volume can lead to gains in the unseen test set and comparable performance in the seen test set. The single layer can also achieve satisfying results, demonstrating that the simple multimodal fusion still makes a huge impact. However, considering the additional resource consumption and balance of efficiency, we use 6 layers of transformer instead of 12.", + "bbox": [ + 88, + 354, + 482, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Number of [seg] Tokens. We only use a single token to promote SAM2 for localization of the correct target, but the number of tokens entering the propagation process can be multiple. As shown in Table 4, increasing the number of tokens can improve significant performances in the unseen split, while this sacrifices the performances of the seen category. We think that it fits the insight from LIBERO [12]. The additional tokens can be regarded as kinds of registers that store additional global information, and this prior knowledge would facilitate learning on new tasks (Unseen tasks). On the other hand, utilizing a single token implicitly encodes global information into that token, which can enhance performance on prior tasks (Seen tasks) but could negatively impact performance on new tasks (Unseen tasks). Despite the representation learned by a single token exhibiting a stronger inductive bias, it is still useful and we set it by default.", + "bbox": [ + 91, + 521, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Fusion Strategy. The token used to prompt the SAM2 can be generated by different strategies. Besides using a learnable token, we can directly take the mean of whole the sequence that is generated by concatenating all the frames' embeddings with those of audio and text. As shown in Table 5, our learnable token can behave better since it can aggregate information from different frames gradually. In contrast, the direct mean operation may lose a", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60eabab4e00472357ad724dd63969fc802da3a575aa2d2ea1ba4f1b856a93504.jpg", + "table_caption": [ + "Table 5. Ablations on the different fusion strategies for generating the prompt token for SAM2." + ], + "table_footnote": [], + "table_body": "
Mask decoderPrompt encoderSeen (%)
JFJ&F
**34.844.839.8
✓*42.250.746.5
✓✓42.550.946.7
", + "bbox": [ + 558, + 88, + 862, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablations on the different training modules in SAM2. \\* denotes the frozen module while $\\sqrt{}$ makes the module trainable.", + "bbox": [ + 511, + 196, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "large amount of inter-frame information.", + "bbox": [ + 511, + 256, + 782, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Training Modules in SAM2. To preserve the strong VOS capabilities of SAM2, we freeze the memory attention modules but train the prompt encoder and mask decoder to further adapt the SAM2 to the multimodal clue reference, which can achieve the best results shown in Table 6. Without training the mask decoder, it will cause a huge performance drop due to the low capability to fit the multimodal representation of the [seg] token.", + "bbox": [ + 511, + 272, + 906, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 407, + 638, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we discuss the limitations of our method. As shown in Table 1 NULL split, our method has a relatively low capability to reject the users' requirements for a null reference. It is the inherent limitation of the prompting paradigm for localization that SAM2 attends to segment something or some irrelevant regions in the image without any meaning. SAM2-LOVE lags behind the EEMC in the null test with 0.2 in $S$ score, achieving 0.23 similar results with the AVSBench plus text method. The $S$ denotes the square root of the ratio of the number of foreground pixels over the background pixels. This can be our future work on how to solve the null reference problem.", + "bbox": [ + 511, + 431, + 906, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 631, + 633, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces SAM2-LOVE, a novel framework for segmenting reference objects in Language-aided Audio-Visual Scenes, overcoming the limitations of previous dual-modality SAM approaches in handling scenarios that integrate three modalities. We propose a multimodal fusion module that compresses information from all modalities into a single token, which then prompts SAM2 for video segmentation. This compressed [seg] token propagates forward through the video to capture both the spatial and temporal features, while also referring to the accumulated historical tokens backward preserving information from previous frames. Extensive experiments demonstrate the superiority of our model on the Ref-AVS benchmark and highlight the efficiency of our design. Future work will focus on addressing scenarios where the three modality references are misaligned, such as the null reference problem.", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28939", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62206153), and in part by Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2", + "[2] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26497–26507, 2024. 3", + "[3] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[4] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12155–12163, 2024. 3, 6", + "[5] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 776-780. IEEE, 2017. 3", + "[6] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 131-135. IEEE, 2017. 3", + "[7] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing, pages 787-798, 2014. 2", + "[8] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3", + "[9] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2", + "[10] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19108-19118, 2022. 1, 3" + ], + "bbox": [ + 93, + 215, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22236-22245, 2023. 6", + "[12] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[13] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2", + "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[16] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364, 2019. 3", + "[17] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3500, 2024. 2", + "[18] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 2", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[20] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In IEEE European Conference on Computer Vision, pages 792-807. Springer, 2016. 2", + "[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dolar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3", + "[23] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2", + "[24] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28940", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.3", + "[25] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 8", + "[26] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 8", + "[27] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Autoacid: A large-scale dataset for audio-language representation learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5025-5034, 2024. 3", + "[28] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929, 2023. 6", + "[29] Yuji Wang, Ruojun Zhao, and Zijun Sun. Efficient remote sensing transformer for coastline detection with sentinel-2 satellite imagery. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium, pages 5439-5442. IEEE, 2023. 2", + "[30] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5669–5677, 2024. 2, 3", + "[31] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision, 2024. 1, 2, 3, 5, 6", + "[32] Yuji Wang, Ruojun Zhao, Shicai Wei, Jingchen Ni, Meng Wu, Yang Luo, and Chunbo Luo. Convolution meets transformer: Efficient hybrid transformer for semantic segmentation with very high resolution imagery. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, pages 9688-9691. IEEE, 2024. 2", + "[33] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. arXiv preprint arXiv:2503.00936, 2025. 2", + "[34] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2", + "[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6", + "[36] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024. 2", + "[38] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16111-16121, 2024. 3", + "[39] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audiovisual question answering on videos. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3480-3491, 2022. 1, 3", + "[40] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[41] Zhao Yang, Jiaqi Wang, Xubing Ye, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Language-aware vision transformer for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[42] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3", + "[43] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Lei Liu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Evf-sam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2, 3", + "[44] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 3", + "[45] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 3, 6", + "[46] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audiovisual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 3", + "[47] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 849 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28941", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_model.json b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..526eebb19a7887689b0b107ff9fbf36e302b3a0f --- /dev/null +++ b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.13, + 0.812, + 0.175 + ], + "angle": 0, + "content": "SAM2-LOVE: Segment Anything Model 2 in Language-aided Audio-Visual Scenes" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.203, + 0.758, + 0.222 + ], + "angle": 0, + "content": "Yuji Wang\\(^{1*}\\), Haoran Xu\\(^{2*}\\), Yong Liu\\(^{1}\\), Jiaze Li\\(^{2}\\), Yansong Tang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.222, + 0.812, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua Shenzhen International Graduate School, Tsinghua University, \\(^{2}\\)ZJU" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.242, + 0.788, + 0.256 + ], + "angle": 0, + "content": "yuji-wan24@mails.tsinghua.edu.cn, †tang.yansong@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.641 + ], + "angle": 0, + "content": "Reference Audio-Visual Segmentation (Ref-AVS) aims to provide a pixel-wise scene understanding in Language-aided Audio-Visual Scenes (LAVS). This task requires the model to continuously segment objects referred to by text and audio from a video. Previous dual-modality methods always fail due to the lack of a third modality and the existing triple-modality method struggles with spatio-temporal consistency, leading to the target shift of different frames. In this work, we introduce a novel framework, termed SAM2-LOVE, which integrates textual, audio, and visual representations into a learnable token to prompt and align SAM2 for achieving Ref-AVS in the LAVS. Technically, our approach includes a multimodal fusion module aimed at improving multimodal understanding of SAM2, as well as token propagation and accumulation strategies designed to enhance spatio-temporal consistency without forgetting historical information. We conducted extensive experiments to demonstrate that SAM2-LOVE outperforms the SOTA by \\(8.5\\%\\) in \\(\\mathcal{J}\\&\\mathcal{F}\\) on the Ref-AVS benchmark and showcase the simplicity and effectiveness of the components. Our code will be available here." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.672, + 0.222, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.864 + ], + "angle": 0, + "content": "In the process of human perception of the world, visual signals play a primary role, often accompanied by other multimodal signals such as audio and text. These additional multimodal cues are not only essential to locating and separating objects and areas of interest but also to understanding the changing behaviors of the objects. The dynamic scene that integrates visual elements with diverse multimodal cues including audio and text is regarded as the Language-aided Audio-Visual Scene (LAVS). To mimic the effect from humans, prior works such as Music-AVQA [10] and AVQA [39] have primarily focused on question-answering" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.297, + 0.721, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.297, + 0.917, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.494, + 0.912, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.729, + 0.908, + 0.8 + ], + "angle": 0, + "content": "Figure 1. The comparison with dual-modality SAM pipelines. (a) The text-visual SAM struggles to distinguish the sounding object in the silent scenes. (b) The audio-visual SAM fails to understand the dynamic control from the text input. (c) Our model fuses three modalities' information and expands to the video due to SAM2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "tasks. Nevertheless, these approaches lack a fine-grained understanding of LVAS at the pixel level. As shown in Figure 2, Reference Audio-Visual Segmentation (Ref-AVS) task [31] makes a good example to improve the pixel-wise" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.217, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.234, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28932" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.093, + 0.88, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.33, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Ref-AVS task under varying modality cues. The expression can be treated as a kind of control signal to condition the referred object on audio. Different expressions can test the models' pixel-wise scene understanding capabilities in the LAVS with different dimensions. The absence angle of a triangle means that modality is not the main test dimension." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.483, + 0.581 + ], + "angle": 0, + "content": "scene understanding of the model in the LAVS. The previous duality-referring segmentation works struggle with the Ref-AVS task where the interested object can only be located under triple modalities, shown in Figure 1. In neither Figure 1 (a) text-visual EVF-SAM [43] nor Figure 1 (b) audio-visual GAVS [30], the truly referred target is accurately segmented due to the absence of additional help from the other third modality. Additionally, they are constrained in the image level, which can not be directly adapted to the dynamic nature of video content. Therefore, methods for effectively fusing information from the three modalities for segmentation have garnered widespread attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Prior work, EEMC [31], simultaneously models three modalities' representations to achieve excellent results on the Ref-AVS benchmark. However, it showcases the relatively insufficient spatio-temporal consistency. Specifically, despite the existence of memory caches, the model fails to continuously track the target's position and shape, leading to a shift in the segmented area over time. Recently, SAM2 [22] has been proposed to generalize the segmenting anything capabilities of the image to the video, which demonstrates a powerful video tracking capability in interframe consistency. SAM2 follows the previous promptable visual segmentation of SAM [8] that requires a strong manual annotation to the interested target of the starting frame (not only limited to the first frame) and then propagates to the whole video based on the crucial frame. Nevertheless, the lack of dependency on other modalities (text or audio) constraints its further application in the LAVS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Thus, in this paper, we propose a novel framework, namely SAM2-LOVE, which means Segment Anything Model 2 in Language-aided audio-O-Visual scEnes. To enhance SAM2's understanding of multimodal information," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.504 + ], + "angle": 0, + "content": "we design the multimodal fusion module to compress all the modalities' information into a learnable token that aligns the SAM2 for segmenting the referring target. Additionally, we incorporate token propagation and accumulation strategies to strengthen spatio-temporal understanding of the video while avoiding forgetting historical frames. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.507, + 0.905, + 0.551 + ], + "angle": 0, + "content": "- We propose a novel framework, SAM2-LOVE that firstly leverages SAM2 to achieve pixel-wise understanding in the LAVS by designing a multimodal fusion module." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.553, + 0.905, + 0.597 + ], + "angle": 0, + "content": "- We develop creative token propagation and accumulation strategies to improve spatio-temporal comprehension of the promtable token." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.598, + 0.905, + 0.643 + ], + "angle": 0, + "content": "- Extensive experiments on Ref-AVS dataset demonstrate the superiority of our method, with ablation studies highlighting the simplicity and effectiveness of its modules." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.507, + 0.905, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.662, + 0.663, + 0.679 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Segmentation via Language or Audio. Unlike semantic segmentation [1, 17, 29, 32], visual segmentation via language starts with referring expression segmentation task [7, 18, 21, 33, 34, 36, 40, 41] which performs target object segmentation based on given natural language expressions. X-Decoder [47] bridges vision and language, unifying multiple tasks within a single model. Moreover, GRES [13] defines a new task called generalized reference expression segmentation. With the development of multimodal large language models (MLLMs) such as LLaVA [14, 15], LISA [9] proposes the concept of reasoning segmentation with combination of pixel-level understanding and MLLMs. GSVA [37] and PixelLM [23] propose multi-target and null-target reasoning segmentation. On the other" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28933" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "hand, audio-visual segmentation (AVS) is also constantly improving. Audio-Visual Segmentation [45, 46] proposes the AVS problem and AVSBench dataset. Numerous subsequent methods [2, 4, 27] have been developed and refined based on the proposed benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.394 + ], + "angle": 0, + "content": "Segment Anything Model. SAM [8] is an interactive segmentation model based on various types of prompts (points, boxes, coarse masks). It shows significant generalization ability in open-world segmentation. Several works strive to reduce the massive computation cost of SAM, such as EfficientSAM [38], FastSAM [44], and MobileSAM [42]. The advent of SAM2 [22] extends SAM's functionality to include video segmentation. While SAM and SAM2 perform effectively in visual segmentation tasks using box, point, or mask prompts, they currently do not possess language understanding capabilities, rendering the direct use of text prompts for referring segmentation infeasible. EVF-SAM [43] extends SAM's capabilities with text-prompted segmentation, while GAVS [30] equips the SAM with image-audio fusion to achieve AVS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Language-aided Audio-Visual Scene. There is presently a scarcity of publicly available tasks that provide datasets for audio-visual scene comprehension with language-based support. The complexity of language-aided audio-visual scene understanding arises from its reliance on three modalities (vision, audio, and text) necessitating massive matched data and their corresponding labels, posing significant challenges for data collection and model design. MusicAVQA [10] and AVQA [39] are two of the classic audio-visual question-answering (QA) datasets. However, the above work focuses on multimodal QA and spatio-temporal reasoning of audio-visual scenes involving various relationships between objects or activities. Ref-AVS [30, 31] begins pixel-level understanding of dynamic audio-visual scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.669 + ], + "angle": 0, + "content": "In summary, no methods exist for integrating the foundational visual segmentation model SAM2 into language-based audio-visual scenes. Therefore, to achieve pixel-level understanding in the LAVS, we introduce SAM2-LOVE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.182, + 0.7 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.294, + 0.726 + ], + "angle": 0, + "content": "3.1. Preliminary of SAM2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.824 + ], + "angle": 0, + "content": "SAM2 can be treated as a generalization of SAM to segment anything in images and videos. It also supports visual prompts (point, box, and mask) to indicate the objects to be segmented in an individual frame (image as a single frame). Then, it generalizes across the whole video based on the prompted frame to achieve the video segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 3 (Inference), this model begins with a hierarchical image encoder that encodes each frame into multilevel embeddings. Unlike SAM which directly feeds these features to the mask decoder for predictions, SAM2 incorporates a series of transformer blocks (i.e. memory" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.378 + ], + "angle": 0, + "content": "attention) to condition the current image feature on memories of previous and prompted frames stored in a memory bank that is maintained by a FIFO (First In First Out) queue. Each memory entry is composed of two kinds of memories: (1) spatial feature maps that are generated through an element-wise summation of the downsampled and encoded mask features from the memory encoder and the unconditioned frame embedding from the image encoder, and (2) the object pointers that are obtained from the output tokens within the mask decoder contain high-level semantic information of the interested target. Therefore, the memory attention operation in the transformer makes the current frame's feature cross-attend both spatial fine-grained memory and abstract temporal object pointers. Second, the mask decoder of SAM2 largely follows the design of SAM, which is a \"two-way\" transformer that takes the prompt embeddings from the prompt encoder and frame embeddings as inputs. The difference lies in the additional occlusion score that aims to judge the object's presence in the current frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.905, + 0.468 + ], + "angle": 0, + "content": "Thanks to the strong and generalized capability of SAM2 across images and videos, we can first compress all the information contained in three modalities into a single token and then use it to prompt SAM2 on a single frame. Finally, starting from this crucial frame, we can propagate through video to obtain the desired masks of all frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.478, + 0.655, + 0.492 + ], + "angle": 0, + "content": "3.2. SAM2-LOVE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.905, + 0.56 + ], + "angle": 0, + "content": "The overall pipeline of our model is shown in Figure 3, which integrates multimodal information in the LAVS into a representative token and uses it to provide a visual clue for SAM2 to segment the reference targets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.906, + 0.772 + ], + "angle": 0, + "content": "Multimodal Representations. Following the previous work [31], for each modality, we use three popular encoders to separately extract their representations, which were not aligned before in their pretraining stages. VGGish [5, 6] encodes an audio clip into audio embeddings \\( F_{A} \\in R^{N \\times d_{A}} \\), where \\( N \\) denotes the duration of the audio in seconds and corresponds to the total number of frames in the video. The sampled \\( N \\) frames are encoded into \\( F_{V} \\in R^{N \\times d_{V} \\times h \\times w} \\) by a ViT [3] pre-trained on ImageNet. For the text expressions, we utilize DistilRoBERTa [16, 24] to obtain the textual representations \\( F_{T} \\in R^{P \\times d_{T}} \\), where \\( P \\) is the number of textual tokens. To unify and align all embeddings into the same semantic space, we use three MLPs to project them into a shared dimension respectively, shown in Equation (1)." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.781, + 0.905, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\widehat {F} _ {Z} = \\operatorname {L i n e a r} (\\operatorname {R e L U} (\\operatorname {L i n e a r} (F _ {Z}))), Z \\in \\{A, V, T \\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Multimodal Fusion. As shown in Figure 3, a multimodal fusion module is proposed to compress all three modalities' information into a single token and then leverage its embedding to represent the fused features. We define this token as a learnable segmentation token, [seg], designed to prompt SAM2 as a clue to retrieve the targeted" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28934" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.892, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.502, + 0.908, + 0.561 + ], + "angle": 0, + "content": "Figure 3. The pipeline of SAM2-LOVE. We harness a fusion transformer to compress the tripe modality representations into a learnable token [seg]. For the input, [aud] and [vis] are added to indicate their corresponding modality sequences, and [his] is to provide additional information about former frames. In the training stage, we use the first frame to align the SAM2 with the compressed and fused representation of [seg]. During inference, we harness the strong zero-shot capabilities of SAM2 to track the referred target at the pixel level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.645 + ], + "angle": 0, + "content": "object. The fusion module is a naive transformer encoder architecture with \\( L \\) stacked bi-directional transformer block that contains a multi-head self-attention (MSA), Layernorm (LN), and MLP, formulated below:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.651, + 0.483, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {l} ^ {\\prime} = M S A \\left(L N \\left(z _ {l - 1}\\right)\\right) + z _ {l - 1}, \\quad l = 1 \\dots L \\tag {2} \\\\ z _ {l} = M L P (L N (z _ {l} ^ {\\prime})) + z _ {l} ^ {\\prime}, \\quad l = 1 \\dots L \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.689, + 0.395, + 0.703 + ], + "angle": 0, + "content": "where \\( z_{l-1} \\) is the input sequence of \\( l \\)-th layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Token Propagation. The behavior of the [seg] token is a self-propagation process within the whole video, which adaptively captures the inner-frame spatial features and the strong continuity among consecutive frames to achieve spatio-temporal consistency. To be more specific, after obtaining the three projected multimodal representations \\(\\widehat{F}_A\\), \\(\\widehat{F}_T\\), and \\(\\widehat{F}_V\\), we first prepend the [seg] token to the head of the sequence of modalities' embeddings to obtain the multimodal sequence \\(F_M^i\\) for the \\(i\\)-th frame, shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.846, + 0.483, + 0.866 + ], + "angle": 0, + "content": "\\[\nF _ {M} ^ {i} = \\operatorname {C o n c a t} ([ [ s e g ]; \\widehat {F} _ {A}; [ a u d ]; \\widehat {F} _ {T}; [ v i s ]; \\widehat {F} _ {V} ^ {i} ]), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\( \\text{Concat} \\) denotes the concatenation operation in the feature dimension axis, two fixed tokens, [aud] and [vis]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.844 + ], + "angle": 0, + "content": "indicate the end and beginning of their corresponding sequences, respectively. This concatenated sequence will be fed into the fusion transformer for updating the status of the [seg] token, allowing the token to capture the visual information of the current frame along with the integral audio and text representations. Since the transformer is the bidirectional encoder that enables all the tokens to interact with each other, we can harness the single token to serve as the multimodal representation of the whole sequence. We take the first element from the output of the fusion transformer \\( Q_{i} = \\text{Transformer}(F_{M}^{i}) \\), as the updated [seg] token \\( [seg] = Q_{i}^{0} \\), shown in Figure 4. Therefore, the token not only contains the current frame information but also the compressed information of the previous frames. Our proposed token propagation strategy allows it to model the video from both the spatial and temporal aspects and aggregate the information within the text and audio." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Token Accumulation. Although the token propagation strategy can effectively obtain the multimodal characteristics and be used to prompt the SAM2, it still faces the issue" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28935" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.096, + 0.9, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.202, + 0.908, + 0.26 + ], + "angle": 0, + "content": "Figure 4. The token propagation and token accumulation strategy in the fusion transformer. For the token propagation, the output [seg] of the current step serves as the input for the next step, which interacts with different frame features but the same audio and text features. To offer additional historical information, the token accumulation strategy appends the global representations of all the previous frames to update the maintained history token [his]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.285, + 0.482, + 0.436 + ], + "angle": 0, + "content": "that the information of the first several frames can be diluted and forgotten with the increasing propagation times. To address the issue, we propose the [cls] token accumulation strategy, in which we keep a history of token sequences, [his] to store the global and abstract representation of historical frames, shown in Figure 4. It is achieved that when propagating to the \\(i\\)-th frame, we provide additional global information of all the previous frames by appending and concatenating the [his] tokens into \\(F_{M}^{i}\\). The updation of the history token [his] can be formulated by Equation (4)." + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.447, + 0.483, + 0.467 + ], + "angle": 0, + "content": "\\[\n[ h i s ] ^ {i} = C o n c a t \\left(\\left[ [ c l s ] ^ {0}; [ c l s ] ^ {1}; \\dots ; [ c l s ] ^ {i} \\right]\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.536 + ], + "angle": 0, + "content": "where \\([cls]^i\\) is the global representation of \\(i\\)-th frame from ViT. These historical tokens can be referred to by the \\([seg]\\) of the current stage, promising the information of all the previous frames is not lost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Unified Perspective of Token Behaviors. We interpret the two token operations described above from a unified knowledge transfer perspective. Token propagation of the [seg] token is a direct flow, which expects the knowledge from different frames to be gradually connected through the propagation process. This is the forward knowledge transfer. In contrast, token accumulation is the replay of [his] tokens, a reversed flow to prevent the [seg] token from forgetting the knowledge of historical frames as it propagates forward. This is the backward knowledge transfer. We propose that these two token operations are functionally opposed yet conceptually unified." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.196, + 0.744 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Trainable Modules. The encoders used for multimodal representations for audio, video, and text are fully frozen, perceiving the pre-trained knowledge. However, the multimodal fusion transformer and [seg] token are fully trainable during our training process, allowing them to learn how to generate the fused and aligned embeddings tailored for prompting SAM2. For SAM2, we keep the image encoder and memory attention modules frozen while training the prompt encoder and mask decoder. Additionally, all the MLPs for projecting are trainable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Training Strategy. Our training strategy initially activates SAM2's multimodal prompting capability at the image level and then leverages its extensive pre-trained knowledge for zero-shot video segmentation. In this setting, the fusion transformer and [seg] token are trained with full access to the entire video sequence, while SAM2 is exposed only to the first frame during training. Specifically, we consolidate multimodal information (video, audio, and text) through the fusion module, which compresses these features into tokens that serve as prompts for SAM2 on the first frame, shown in Figure 3 (Training). Then, a supervised loss is applied to this frame to train SAM2's referring segmentation abilities. During testing, both the fusion module and SAM2 receive the entire video, allowing for comprehensive segmentation across all frames. This approach enables SAM2 to effectively utilize multimodal prompts on the initial frame and then generalize to the complete video sequence through its zero-shot Video Object Segmentation (VOS) capabilities, shown in Figure 3 (Inference)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Loss Function. The model is trained end-to-end under the supervised mask loss \\(\\mathcal{L}_{mask}\\), to encourage the SAM2 to predict the correct localized masks. To compute this loss, we combine the per-pixel binary cross-entropy (BCE) loss and DICE loss with their corresponding weights \\(\\lambda_{bce}\\) and \\(\\lambda_{dice}\\), which can be formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.668, + 0.907, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathbf {B C E} (\\hat {M}, M) + \\lambda_ {d i c e} \\mathbf {D I C E} (\\hat {M}, M) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.691, + 0.906, + 0.723 + ], + "angle": 0, + "content": "where \\(\\hat{M}\\) is a single mask prediction for the first frame and \\(M\\) is the label of the first frame for supervision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.734, + 0.646, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.721, + 0.774 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Dataset. We evaluate our method on the Ref-AVS benchmark [31], a substantial collection of 4,000 videos with manual pixel-level annotations and 20,000 expressions reflecting the scenarios in audio, vision, and time dimensions. Additionally, the collected videos contain abundant and diverse visual elements. Each video encompasses a high average number of objects about 1.72, a wide variety of audible objects in 48 categories, and 3 kinds of static objects." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28936" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.912, + 0.338 + ], + "angle": 0, + "content": "
MethodSeen (%)Unseen (%)Mix (S+U)NULLVENUE
JFJ&FJFJ&FJFJ&Fs
AVSBench [45]21.245.733.527.750.939.324.548.336.4-ECCV'2022
+text23.251.137.232.454.743.527.852.940.30.208
AVSegFormer [4]29.242.335.732.745.138.930.943.737.3-AAAI'2024
+text33.547.040.236.150.143.134.848.641.70.171
GAVS [28]24.845.535.227.846.036.926.345.736.0-AAAI'2024
+text28.949.839.429.849.739.829.449.839.60.190
ReferFormer [35]27.447.737.524.950.437.626.249.137.6-CVPR'2022
+audio31.350.140.730.448.839.630.949.540.20.176
R2VOS [11]21.330.926.123.943.133.522.637.029.8-ICCV'2023
+audio25.041.033.027.949.838.926.545.435.90.183
EEMC [31]34.251.342.849.564.857.241.958.150.00.007ECCV'2024
SAM2-LOVE (ours)43.551.947.766.572.369.455.062.158.50.23CVPR2025
vs. prev. SOTA+9.3+0.6+4.9+17.0+7.5+12.2+13.1+4.0+8.5--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.349, + 0.908, + 0.379 + ], + "angle": 0, + "content": "Table 1. Performance comparison across different methods in Seen, Unseen, Mix (S+U), and NULL settings of Ref-AVS benchmark. The mix indicates the average value of seen and unseen splits." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.404, + 0.482, + 0.556 + ], + "angle": 0, + "content": "Therefore, the tested scenarios are more challenging with multiple sound sources and multiple semantics. The dataset is divided into three splits, a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three distinct subsets: 1) seen splits that are composed of the same 39 categories in the training set. 2) unseen splits with 13 additional categories that do not appear in the training stage. 3) null splits to test the model when an expression refers to an object that either does not exist or is not visible in the given context." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.562, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Baselines. We compare our method with six challenging baselines: AVSBench [45], AVSegFormer [4], GAVS [28], ReferFormer [35], R2VOS [11] and EEMC [31]. AVSBench, AVSegFormer, and GAVS are designed for Audio-Visual Segmentation (AVS) tasks while ReferFormer and R2VOS claim for Referring Video Object Segmentation(Ref-VOS) tasks. We also report the results achieved by offering the third their absent modalities, which is consistent with EEMC. EEMC is specially designed to achieve fine-grained semantic understanding in the LAVS and also proposes the Ref-AVS task." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For the training recipe, we adopt the DeepSpeed training engine with ZeRO-2 to train our model on 2 NVIDIA A100 GPUs. The AdamW [19] optimizer is harnessed with the learning rate and weight decay set to 1e-4 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the BCE loss \\(\\lambda_{bce}\\) and Dice loss \\(\\lambda_{dice}\\) are set to 1.0 and 1.0, respectively. The batch size is set to 8 and the gradient accumulation step is set to 2. For the network architecture, the fusion transformer is the stacked 6 transformer blocks with bi-directional self" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.449 + ], + "angle": 0, + "content": "attention. The hidden dimension of the transformer is 1024 and the number of [seg] tokens is set to 1. SAM2 is initialized with the weight of SAM2-Hiera-large." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Metrics. Following [20, 31], we assess the performance of the Ref-AVS benchmark using the Jaccard index \\((\\mathcal{I})\\), the F-score \\((\\mathcal{F})\\), and their average value \\((\\mathcal{I} \\& \\mathcal{F})\\) as key evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.655, + 0.532 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.812 + ], + "angle": 0, + "content": "Table 1 shows the comparative results of our method on Ref-AVS benchmarks. Our SAM2-LOVE achieves state-of-the-art (SOTA) performance and significantly outperforms the second-best method EEMC on both seen and unseen test splits. The results indicate that our model has highly generalized capabilities that can be directly used for out-of-domain scenarios. We attribute two reasons leading to these huge performance improvements. Firstly, the fusion transformer can effectively step-by-step compress the multimodal cues into the learnable tokens through the [seg] token propagation strategy and [cls] token accumulation. The generated token contains high semantics as well as positional information, which can accurately prompt SAM2 to segment the inserted targets. Secondly, SAM2 showcases strong zero-shot VOS capabilities, accurately tracking the referred object despite the circumstance where it suddenly disappears and occurs again or similar objects have interacted together." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.703, + 0.835 + ], + "angle": 0, + "content": "4.3. Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Figure 5 shows the visualization comparative results with the previous SOTA EEMC. In the first case, we can observe that the SAM2 shows strong consistency within different frames, which can track the referred object with high" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28937" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.099, + 0.892, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.446, + 0.908, + 0.476 + ], + "angle": 0, + "content": "Figure 5. The visualization results of the referred objects in the Ref-AVS. Notably, we use the legends of the trumpet to represent the different sound volumes of the audio objects, from silent to increasing loudness. More cases can be found in the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.497, + 0.501, + 0.571 + ], + "angle": 0, + "content": "
CLIP-T\nCLIP-VRoBERTa\nViT[cls] AccSeen (%)Unseen (%)
JFJ&FJFJ&F
✓✓42.551.246.865.171.468.2
✓✓43.551.947.766.572.369.4
✓42.650.946.766.371.669.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Table 2. Ablations on backbone design and our proposed [cls] token accumulation strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.484, + 0.759 + ], + "angle": 0, + "content": "mask quality. Additionally, it also demonstrates that the [seg] token can accurately locate the small referring target, while EEMC fails in the first several frames. The second case shows that our SAM2-LOVE is more robust regarding changes in audible objects. EEMC may not distinguish the difference or malfunction in understanding the hidden meaning in the expressions (i.e. \"shorter duration\" instead of the obvious \"the man\")." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.245, + 0.789 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effect of Backbone. Table 2 shows the ablation study of backbones. Compared to CLIP vision and text encoder, RoBERTa and ViT achieve better performance, which can be the difference from the pre-trained paradigm. To align with visual elements, the CLIP textual encoder focuses more on the text with specific semantics such as cat and person. However, in the LAVS scenes, the massive expres" + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.497, + 0.88, + 0.571 + ], + "angle": 0, + "content": "
Number of transformer blocksSeen (%)Unseen (%)
JFJ&FJFJ&F
1 layer41.549.445.465.270.968.1
6 layers42.650.946.766.371.669.0
12 layers42.050.746.467.773.370.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.581, + 0.897, + 0.594 + ], + "angle": 0, + "content": "Table 3. Ablations on the number of fusion transformer blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.717 + ], + "angle": 0, + "content": "sions tend to provide a kind of control signal without referred category names, e.g., “the one making the sound.” and “the one making the most intense sound.” Therefore, the RoBERTa may demonstrate greater strength in understanding instructions due to expertise in natural language understanding during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Effect of [cls] Token Accumulation. Since we harness the propagated token to prompt the first frame for SAM2, the attention for the first several frames may weaken during propagation, causing the token to depend only on the last frame. Therefore, in the fusion transformer, we attach the visual global [cls] tokens of previous frames to the current frame to update the history token, avoiding the [seg] token forgetting the information of former frames. Compared to the last row and second row in Table 2, we can observe that this strategy can effectively improve performance for free without additional design or resource consumption." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Effect of Fusion Transformer Design. The design of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28938" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.472, + 0.176 + ], + "angle": 0, + "content": "
Number of [seg] tokensSeen (%)Unseen (%)
JFJ&FJFJ&F
142.650.946.766.371.669.0
441.350.846.065.271.068.1
840.749.845.267.072.269.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.136, + 0.186, + 0.438, + 0.201 + ], + "angle": 0, + "content": "Table 4. Ablations on the numbers of [seg] tokens." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.217, + 0.47, + 0.288 + ], + "angle": 0, + "content": "
Fusion StrategySeen (%)Unseen (%)
JFJ&FJFJ&F
learnable token42.650.946.766.371.669.0
mean41.650.145.865.971.668.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Table 5. Ablations on the different fusion strategies for generating the prompt token for SAM2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.483, + 0.52 + ], + "angle": 0, + "content": "our fusion transformer is a naive transformer architecture without additional operations such as ROPE [26] or attention enhancement [25]. Table 3 shows the different performances under different numbers of transformer blocks. Increasing the model's volume can lead to gains in the unseen test set and comparable performance in the seen test set. The single layer can also achieve satisfying results, demonstrating that the simple multimodal fusion still makes a huge impact. However, considering the additional resource consumption and balance of efficiency, we use 6 layers of transformer instead of 12." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.522, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Effect of the Number of [seg] Tokens. We only use a single token to promote SAM2 for localization of the correct target, but the number of tokens entering the propagation process can be multiple. As shown in Table 4, increasing the number of tokens can improve significant performances in the unseen split, while this sacrifices the performances of the seen category. We think that it fits the insight from LIBERO [12]. The additional tokens can be regarded as kinds of registers that store additional global information, and this prior knowledge would facilitate learning on new tasks (Unseen tasks). On the other hand, utilizing a single token implicitly encodes global information into that token, which can enhance performance on prior tasks (Seen tasks) but could negatively impact performance on new tasks (Unseen tasks). Despite the representation learned by a single token exhibiting a stronger inductive bias, it is still useful and we set it by default." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Effect of Fusion Strategy. The token used to prompt the SAM2 can be generated by different strategies. Besides using a learnable token, we can directly take the mean of whole the sequence that is generated by concatenating all the frames' embeddings with those of audio and text. As shown in Table 5, our learnable token can behave better since it can aggregate information from different frames gradually. In contrast, the direct mean operation may lose a" + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.089, + 0.863, + 0.187 + ], + "angle": 0, + "content": "
Mask decoderPrompt encoderSeen (%)
JFJ&F
**34.844.839.8
✓*42.250.746.5
✓✓42.550.946.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.197, + 0.905, + 0.226 + ], + "angle": 0, + "content": "Table 6. Ablations on the different training modules in SAM2. \\* denotes the frozen module while \\(\\sqrt{}\\) makes the module trainable." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.257, + 0.783, + 0.271 + ], + "angle": 0, + "content": "large amount of inter-frame information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Effect of Training Modules in SAM2. To preserve the strong VOS capabilities of SAM2, we freeze the memory attention modules but train the prompt encoder and mask decoder to further adapt the SAM2 to the multimodal clue reference, which can achieve the best results shown in Table 6. Without training the mask decoder, it will cause a huge performance drop due to the low capability to fit the multimodal representation of the [seg] token." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.409, + 0.64, + 0.423 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.614 + ], + "angle": 0, + "content": "In this section, we discuss the limitations of our method. As shown in Table 1 NULL split, our method has a relatively low capability to reject the users' requirements for a null reference. It is the inherent limitation of the prompting paradigm for localization that SAM2 attends to segment something or some irrelevant regions in the image without any meaning. SAM2-LOVE lags behind the EEMC in the null test with 0.2 in \\( S \\) score, achieving 0.23 similar results with the AVSBench plus text method. The \\( S \\) denotes the square root of the ratio of the number of foreground pixels over the background pixels. This can be our future work on how to solve the null reference problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.634, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper introduces SAM2-LOVE, a novel framework for segmenting reference objects in Language-aided Audio-Visual Scenes, overcoming the limitations of previous dual-modality SAM approaches in handling scenarios that integrate three modalities. We propose a multimodal fusion module that compresses information from all modalities into a single token, which then prompts SAM2 for video segmentation. This compressed [seg] token propagates forward through the video to capture both the spatial and temporal features, while also referring to the accumulated historical tokens backward preserving information from previous frames. Extensive experiments demonstrate the superiority of our model on the Ref-AVS benchmark and highlight the efficiency of our design. Future work will focus on addressing scenarios where the three modality references are misaligned, such as the null reference problem." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28939" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.178 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62206153), and in part by Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.193, + 0.188, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.217, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.261, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[2] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26497–26507, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[3] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[4] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12155–12163, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.448, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[5] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 776-780. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[6] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 131-135. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing, pages 787-798, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[9] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19108-19118, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[11] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22236-22245, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[12] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[13] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.33 + ], + "angle": 0, + "content": "[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.4 + ], + "angle": 0, + "content": "[16] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[17] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3500, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[18] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[20] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.706 + ], + "angle": 0, + "content": "[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In IEEE European Conference on Computer Vision, pages 792-807. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dolar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.908, + 0.873 + ], + "angle": 0, + "content": "[23] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28940" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[25] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[26] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[27] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Autoacid: A large-scale dataset for audio-language representation learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5025-5034, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[28] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[29] Yuji Wang, Ruojun Zhao, and Zijun Sun. Efficient remote sensing transformer for coastline detection with sentinel-2 satellite imagery. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium, pages 5439-5442. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[30] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5669–5677, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[31] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[32] Yuji Wang, Ruojun Zhao, Shicai Wei, Jingchen Ni, Meng Wu, Yang Luo, and Chunbo Luo. Convolution meets transformer: Efficient hybrid transformer for semantic segmentation with very high resolution imagery. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, pages 9688-9691. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.638, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[33] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. arXiv preprint arXiv:2503.00936, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[34] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[37] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.246 + ], + "angle": 0, + "content": "[38] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16111-16121, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[39] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audiovisual question answering on videos. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3480-3491, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[40] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.906, + 0.457 + ], + "angle": 0, + "content": "[41] Zhao Yang, Jiaqi Wang, Xubing Ye, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Language-aware vision transformer for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.906, + 0.527 + ], + "angle": 0, + "content": "[42] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.906, + 0.598 + ], + "angle": 0, + "content": "[43] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Lei Liu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Evf-sam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[44] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.906, + 0.698 + ], + "angle": 0, + "content": "[45] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.906, + 0.767 + ], + "angle": 0, + "content": "[46] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audiovisual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.906, + 0.851 + ], + "angle": 0, + "content": "[47] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28941" + } + ] +] \ No newline at end of file diff --git a/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_origin.pdf b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..338c1f1776c6f5807cc1ebc13cc7db5eee7dc181 --- /dev/null +++ b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/dc4d562e-b5c0-4f65-86f2-b983c6d324de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:220086e2c8d986ce90cbaf59304efe132552b28aa43e075866c7ebad8553f9f4 +size 1285174 diff --git a/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/full.md b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7a97cf121f1f94b4c1bc4cef16b466eecaa4c56 --- /dev/null +++ b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/full.md @@ -0,0 +1,266 @@ +# SAM2-LOVE: Segment Anything Model 2 in Language-aided Audio-Visual Scenes + +Yuji Wang $^{1*}$ , Haoran Xu $^{2*}$ , Yong Liu $^{1}$ , Jiaze Li $^{2}$ , Yansong Tang $^{1\dagger}$ + +$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University, $^{2}$ ZJU + +yuji-wan24@mails.tsinghua.edu.cn, †tang.yansong@sz.tsinghua.edu.cn + +# Abstract + +Reference Audio-Visual Segmentation (Ref-AVS) aims to provide a pixel-wise scene understanding in Language-aided Audio-Visual Scenes (LAVS). This task requires the model to continuously segment objects referred to by text and audio from a video. Previous dual-modality methods always fail due to the lack of a third modality and the existing triple-modality method struggles with spatio-temporal consistency, leading to the target shift of different frames. In this work, we introduce a novel framework, termed SAM2-LOVE, which integrates textual, audio, and visual representations into a learnable token to prompt and align SAM2 for achieving Ref-AVS in the LAVS. Technically, our approach includes a multimodal fusion module aimed at improving multimodal understanding of SAM2, as well as token propagation and accumulation strategies designed to enhance spatio-temporal consistency without forgetting historical information. We conducted extensive experiments to demonstrate that SAM2-LOVE outperforms the SOTA by $8.5\%$ in $\mathcal{J}\&\mathcal{F}$ on the Ref-AVS benchmark and showcase the simplicity and effectiveness of the components. Our code will be available here. + +# 1. Introduction + +In the process of human perception of the world, visual signals play a primary role, often accompanied by other multimodal signals such as audio and text. These additional multimodal cues are not only essential to locating and separating objects and areas of interest but also to understanding the changing behaviors of the objects. The dynamic scene that integrates visual elements with diverse multimodal cues including audio and text is regarded as the Language-aided Audio-Visual Scene (LAVS). To mimic the effect from humans, prior works such as Music-AVQA [10] and AVQA [39] have primarily focused on question-answering + +![](images/d43998c6c8420773f1678dab1fa83c90aa7d0dd726802510d6d658a6cdbe7cc9.jpg) + +![](images/06d5c19d02f822cd0b15cd3d50fa6411708bdf91eb3db7d58eb3f748a96c7946.jpg) + +![](images/2bcfeafbdb1a56205f9621532a4faaed69a074b05c106f118d0bea3553142c43.jpg) +Figure 1. The comparison with dual-modality SAM pipelines. (a) The text-visual SAM struggles to distinguish the sounding object in the silent scenes. (b) The audio-visual SAM fails to understand the dynamic control from the text input. (c) Our model fuses three modalities' information and expands to the video due to SAM2. + +tasks. Nevertheless, these approaches lack a fine-grained understanding of LVAS at the pixel level. As shown in Figure 2, Reference Audio-Visual Segmentation (Ref-AVS) task [31] makes a good example to improve the pixel-wise + +![](images/16be766b43381123fdbeb8a4fffae0994bc3b2499bb896eabe1d7763492a799e.jpg) +Figure 2. Illustration of the Ref-AVS task under varying modality cues. The expression can be treated as a kind of control signal to condition the referred object on audio. Different expressions can test the models' pixel-wise scene understanding capabilities in the LAVS with different dimensions. The absence angle of a triangle means that modality is not the main test dimension. + +scene understanding of the model in the LAVS. The previous duality-referring segmentation works struggle with the Ref-AVS task where the interested object can only be located under triple modalities, shown in Figure 1. In neither Figure 1 (a) text-visual EVF-SAM [43] nor Figure 1 (b) audio-visual GAVS [30], the truly referred target is accurately segmented due to the absence of additional help from the other third modality. Additionally, they are constrained in the image level, which can not be directly adapted to the dynamic nature of video content. Therefore, methods for effectively fusing information from the three modalities for segmentation have garnered widespread attention. + +Prior work, EEMC [31], simultaneously models three modalities' representations to achieve excellent results on the Ref-AVS benchmark. However, it showcases the relatively insufficient spatio-temporal consistency. Specifically, despite the existence of memory caches, the model fails to continuously track the target's position and shape, leading to a shift in the segmented area over time. Recently, SAM2 [22] has been proposed to generalize the segmenting anything capabilities of the image to the video, which demonstrates a powerful video tracking capability in interframe consistency. SAM2 follows the previous promptable visual segmentation of SAM [8] that requires a strong manual annotation to the interested target of the starting frame (not only limited to the first frame) and then propagates to the whole video based on the crucial frame. Nevertheless, the lack of dependency on other modalities (text or audio) constraints its further application in the LAVS. + +Thus, in this paper, we propose a novel framework, namely SAM2-LOVE, which means Segment Anything Model 2 in Language-aided audio-O-Visual scEnes. To enhance SAM2's understanding of multimodal information, + +we design the multimodal fusion module to compress all the modalities' information into a learnable token that aligns the SAM2 for segmenting the referring target. Additionally, we incorporate token propagation and accumulation strategies to strengthen spatio-temporal understanding of the video while avoiding forgetting historical frames. Our main contributions can be summarized as follows: + +- We propose a novel framework, SAM2-LOVE that firstly leverages SAM2 to achieve pixel-wise understanding in the LAVS by designing a multimodal fusion module. +- We develop creative token propagation and accumulation strategies to improve spatio-temporal comprehension of the promtable token. +- Extensive experiments on Ref-AVS dataset demonstrate the superiority of our method, with ablation studies highlighting the simplicity and effectiveness of its modules. + +# 2. Related Works + +Segmentation via Language or Audio. Unlike semantic segmentation [1, 17, 29, 32], visual segmentation via language starts with referring expression segmentation task [7, 18, 21, 33, 34, 36, 40, 41] which performs target object segmentation based on given natural language expressions. X-Decoder [47] bridges vision and language, unifying multiple tasks within a single model. Moreover, GRES [13] defines a new task called generalized reference expression segmentation. With the development of multimodal large language models (MLLMs) such as LLaVA [14, 15], LISA [9] proposes the concept of reasoning segmentation with combination of pixel-level understanding and MLLMs. GSVA [37] and PixelLM [23] propose multi-target and null-target reasoning segmentation. On the other + +hand, audio-visual segmentation (AVS) is also constantly improving. Audio-Visual Segmentation [45, 46] proposes the AVS problem and AVSBench dataset. Numerous subsequent methods [2, 4, 27] have been developed and refined based on the proposed benchmark. + +Segment Anything Model. SAM [8] is an interactive segmentation model based on various types of prompts (points, boxes, coarse masks). It shows significant generalization ability in open-world segmentation. Several works strive to reduce the massive computation cost of SAM, such as EfficientSAM [38], FastSAM [44], and MobileSAM [42]. The advent of SAM2 [22] extends SAM's functionality to include video segmentation. While SAM and SAM2 perform effectively in visual segmentation tasks using box, point, or mask prompts, they currently do not possess language understanding capabilities, rendering the direct use of text prompts for referring segmentation infeasible. EVF-SAM [43] extends SAM's capabilities with text-prompted segmentation, while GAVS [30] equips the SAM with image-audio fusion to achieve AVS. + +Language-aided Audio-Visual Scene. There is presently a scarcity of publicly available tasks that provide datasets for audio-visual scene comprehension with language-based support. The complexity of language-aided audio-visual scene understanding arises from its reliance on three modalities (vision, audio, and text) necessitating massive matched data and their corresponding labels, posing significant challenges for data collection and model design. MusicAVQA [10] and AVQA [39] are two of the classic audio-visual question-answering (QA) datasets. However, the above work focuses on multimodal QA and spatio-temporal reasoning of audio-visual scenes involving various relationships between objects or activities. Ref-AVS [30, 31] begins pixel-level understanding of dynamic audio-visual scenes. + +In summary, no methods exist for integrating the foundational visual segmentation model SAM2 into language-based audio-visual scenes. Therefore, to achieve pixel-level understanding in the LAVS, we introduce SAM2-LOVE. + +# 3. Method + +# 3.1. Preliminary of SAM2 + +SAM2 can be treated as a generalization of SAM to segment anything in images and videos. It also supports visual prompts (point, box, and mask) to indicate the objects to be segmented in an individual frame (image as a single frame). Then, it generalizes across the whole video based on the prompted frame to achieve the video segmentation. + +As shown in Figure 3 (Inference), this model begins with a hierarchical image encoder that encodes each frame into multilevel embeddings. Unlike SAM which directly feeds these features to the mask decoder for predictions, SAM2 incorporates a series of transformer blocks (i.e. memory + +attention) to condition the current image feature on memories of previous and prompted frames stored in a memory bank that is maintained by a FIFO (First In First Out) queue. Each memory entry is composed of two kinds of memories: (1) spatial feature maps that are generated through an element-wise summation of the downsampled and encoded mask features from the memory encoder and the unconditioned frame embedding from the image encoder, and (2) the object pointers that are obtained from the output tokens within the mask decoder contain high-level semantic information of the interested target. Therefore, the memory attention operation in the transformer makes the current frame's feature cross-attend both spatial fine-grained memory and abstract temporal object pointers. Second, the mask decoder of SAM2 largely follows the design of SAM, which is a "two-way" transformer that takes the prompt embeddings from the prompt encoder and frame embeddings as inputs. The difference lies in the additional occlusion score that aims to judge the object's presence in the current frame. + +Thanks to the strong and generalized capability of SAM2 across images and videos, we can first compress all the information contained in three modalities into a single token and then use it to prompt SAM2 on a single frame. Finally, starting from this crucial frame, we can propagate through video to obtain the desired masks of all frames. + +# 3.2. SAM2-LOVE + +The overall pipeline of our model is shown in Figure 3, which integrates multimodal information in the LAVS into a representative token and uses it to provide a visual clue for SAM2 to segment the reference targets. + +Multimodal Representations. Following the previous work [31], for each modality, we use three popular encoders to separately extract their representations, which were not aligned before in their pretraining stages. VGGish [5, 6] encodes an audio clip into audio embeddings $F_{A} \in R^{N \times d_{A}}$ , where $N$ denotes the duration of the audio in seconds and corresponds to the total number of frames in the video. The sampled $N$ frames are encoded into $F_{V} \in R^{N \times d_{V} \times h \times w}$ by a ViT [3] pre-trained on ImageNet. For the text expressions, we utilize DistilRoBERTa [16, 24] to obtain the textual representations $F_{T} \in R^{P \times d_{T}}$ , where $P$ is the number of textual tokens. To unify and align all embeddings into the same semantic space, we use three MLPs to project them into a shared dimension respectively, shown in Equation (1). + +$$ +\widehat {F} _ {Z} = \operatorname {L i n e a r} (\operatorname {R e L U} (\operatorname {L i n e a r} (F _ {Z}))), Z \in \{A, V, T \} \tag {1} +$$ + +Multimodal Fusion. As shown in Figure 3, a multimodal fusion module is proposed to compress all three modalities' information into a single token and then leverage its embedding to represent the fused features. We define this token as a learnable segmentation token, [seg], designed to prompt SAM2 as a clue to retrieve the targeted + +![](images/4768acffa4521db7f0bbbd66a3a58d4a7e5f2a3e62165d8db31ba80981dfa202.jpg) +Figure 3. The pipeline of SAM2-LOVE. We harness a fusion transformer to compress the tripe modality representations into a learnable token [seg]. For the input, [aud] and [vis] are added to indicate their corresponding modality sequences, and [his] is to provide additional information about former frames. In the training stage, we use the first frame to align the SAM2 with the compressed and fused representation of [seg]. During inference, we harness the strong zero-shot capabilities of SAM2 to track the referred target at the pixel level. + +object. The fusion module is a naive transformer encoder architecture with $L$ stacked bi-directional transformer block that contains a multi-head self-attention (MSA), Layernorm (LN), and MLP, formulated below: + +$$ +\begin{array}{l} z _ {l} ^ {\prime} = M S A \left(L N \left(z _ {l - 1}\right)\right) + z _ {l - 1}, \quad l = 1 \dots L \tag {2} \\ z _ {l} = M L P (L N (z _ {l} ^ {\prime})) + z _ {l} ^ {\prime}, \quad l = 1 \dots L \\ \end{array} +$$ + +where $z_{l-1}$ is the input sequence of $l$ -th layer. + +Token Propagation. The behavior of the [seg] token is a self-propagation process within the whole video, which adaptively captures the inner-frame spatial features and the strong continuity among consecutive frames to achieve spatio-temporal consistency. To be more specific, after obtaining the three projected multimodal representations $\widehat{F}_A$ , $\widehat{F}_T$ , and $\widehat{F}_V$ , we first prepend the [seg] token to the head of the sequence of modalities' embeddings to obtain the multimodal sequence $F_M^i$ for the $i$ -th frame, shown as follows: + +$$ +F _ {M} ^ {i} = \operatorname {C o n c a t} ([ [ s e g ]; \widehat {F} _ {A}; [ a u d ]; \widehat {F} _ {T}; [ v i s ]; \widehat {F} _ {V} ^ {i} ]), \tag {3} +$$ + +where $\text{Concat}$ denotes the concatenation operation in the feature dimension axis, two fixed tokens, [aud] and [vis] + +indicate the end and beginning of their corresponding sequences, respectively. This concatenated sequence will be fed into the fusion transformer for updating the status of the [seg] token, allowing the token to capture the visual information of the current frame along with the integral audio and text representations. Since the transformer is the bidirectional encoder that enables all the tokens to interact with each other, we can harness the single token to serve as the multimodal representation of the whole sequence. We take the first element from the output of the fusion transformer $Q_{i} = \text{Transformer}(F_{M}^{i})$ , as the updated [seg] token $[seg] = Q_{i}^{0}$ , shown in Figure 4. Therefore, the token not only contains the current frame information but also the compressed information of the previous frames. Our proposed token propagation strategy allows it to model the video from both the spatial and temporal aspects and aggregate the information within the text and audio. + +Token Accumulation. Although the token propagation strategy can effectively obtain the multimodal characteristics and be used to prompt the SAM2, it still faces the issue + +![](images/a23f14fa9ce23cfdbee60a3f04933d9d121a88dac2bab81b561f7f256804cba7.jpg) +Figure 4. The token propagation and token accumulation strategy in the fusion transformer. For the token propagation, the output [seg] of the current step serves as the input for the next step, which interacts with different frame features but the same audio and text features. To offer additional historical information, the token accumulation strategy appends the global representations of all the previous frames to update the maintained history token [his]. + +that the information of the first several frames can be diluted and forgotten with the increasing propagation times. To address the issue, we propose the [cls] token accumulation strategy, in which we keep a history of token sequences, [his] to store the global and abstract representation of historical frames, shown in Figure 4. It is achieved that when propagating to the $i$ -th frame, we provide additional global information of all the previous frames by appending and concatenating the [his] tokens into $F_{M}^{i}$ . The updation of the history token [his] can be formulated by Equation (4). + +$$ +[ h i s ] ^ {i} = C o n c a t \left(\left[ [ c l s ] ^ {0}; [ c l s ] ^ {1}; \dots ; [ c l s ] ^ {i} \right]\right) \tag {4} +$$ + +where $[cls]^i$ is the global representation of $i$ -th frame from ViT. These historical tokens can be referred to by the $[seg]$ of the current stage, promising the information of all the previous frames is not lost. + +Unified Perspective of Token Behaviors. We interpret the two token operations described above from a unified knowledge transfer perspective. Token propagation of the [seg] token is a direct flow, which expects the knowledge from different frames to be gradually connected through the propagation process. This is the forward knowledge transfer. In contrast, token accumulation is the replay of [his] tokens, a reversed flow to prevent the [seg] token from forgetting the knowledge of historical frames as it propagates forward. This is the backward knowledge transfer. We propose that these two token operations are functionally opposed yet conceptually unified. + +# 3.3. Training + +Trainable Modules. The encoders used for multimodal representations for audio, video, and text are fully frozen, perceiving the pre-trained knowledge. However, the multimodal fusion transformer and [seg] token are fully trainable during our training process, allowing them to learn how to generate the fused and aligned embeddings tailored for prompting SAM2. For SAM2, we keep the image encoder and memory attention modules frozen while training the prompt encoder and mask decoder. Additionally, all the MLPs for projecting are trainable. + +Training Strategy. Our training strategy initially activates SAM2's multimodal prompting capability at the image level and then leverages its extensive pre-trained knowledge for zero-shot video segmentation. In this setting, the fusion transformer and [seg] token are trained with full access to the entire video sequence, while SAM2 is exposed only to the first frame during training. Specifically, we consolidate multimodal information (video, audio, and text) through the fusion module, which compresses these features into tokens that serve as prompts for SAM2 on the first frame, shown in Figure 3 (Training). Then, a supervised loss is applied to this frame to train SAM2's referring segmentation abilities. During testing, both the fusion module and SAM2 receive the entire video, allowing for comprehensive segmentation across all frames. This approach enables SAM2 to effectively utilize multimodal prompts on the initial frame and then generalize to the complete video sequence through its zero-shot Video Object Segmentation (VOS) capabilities, shown in Figure 3 (Inference). + +Loss Function. The model is trained end-to-end under the supervised mask loss $\mathcal{L}_{mask}$ , to encourage the SAM2 to predict the correct localized masks. To compute this loss, we combine the per-pixel binary cross-entropy (BCE) loss and DICE loss with their corresponding weights $\lambda_{bce}$ and $\lambda_{dice}$ , which can be formulated by: + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \mathbf {B C E} (\hat {M}, M) + \lambda_ {d i c e} \mathbf {D I C E} (\hat {M}, M) \tag {5} +$$ + +where $\hat{M}$ is a single mask prediction for the first frame and $M$ is the label of the first frame for supervision. + +# 4. Experiments + +# 4.1. Experimental Settings + +Dataset. We evaluate our method on the Ref-AVS benchmark [31], a substantial collection of 4,000 videos with manual pixel-level annotations and 20,000 expressions reflecting the scenarios in audio, vision, and time dimensions. Additionally, the collected videos contain abundant and diverse visual elements. Each video encompasses a high average number of objects about 1.72, a wide variety of audible objects in 48 categories, and 3 kinds of static objects. + +
MethodSeen (%)Unseen (%)Mix (S+U)NULLVENUE
JFJ&FJFJ&FJFJ&Fs
AVSBench [45]21.245.733.527.750.939.324.548.336.4-ECCV'2022
+text23.251.137.232.454.743.527.852.940.30.208
AVSegFormer [4]29.242.335.732.745.138.930.943.737.3-AAAI'2024
+text33.547.040.236.150.143.134.848.641.70.171
GAVS [28]24.845.535.227.846.036.926.345.736.0-AAAI'2024
+text28.949.839.429.849.739.829.449.839.60.190
ReferFormer [35]27.447.737.524.950.437.626.249.137.6-CVPR'2022
+audio31.350.140.730.448.839.630.949.540.20.176
R2VOS [11]21.330.926.123.943.133.522.637.029.8-ICCV'2023
+audio25.041.033.027.949.838.926.545.435.90.183
EEMC [31]34.251.342.849.564.857.241.958.150.00.007ECCV'2024
SAM2-LOVE (ours)43.551.947.766.572.369.455.062.158.50.23CVPR2025
vs. prev. SOTA+9.3+0.6+4.9+17.0+7.5+12.2+13.1+4.0+8.5--
+ +Table 1. Performance comparison across different methods in Seen, Unseen, Mix (S+U), and NULL settings of Ref-AVS benchmark. The mix indicates the average value of seen and unseen splits. + +Therefore, the tested scenarios are more challenging with multiple sound sources and multiple semantics. The dataset is divided into three splits, a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three distinct subsets: 1) seen splits that are composed of the same 39 categories in the training set. 2) unseen splits with 13 additional categories that do not appear in the training stage. 3) null splits to test the model when an expression refers to an object that either does not exist or is not visible in the given context. + +Baselines. We compare our method with six challenging baselines: AVSBench [45], AVSegFormer [4], GAVS [28], ReferFormer [35], R2VOS [11] and EEMC [31]. AVSBench, AVSegFormer, and GAVS are designed for Audio-Visual Segmentation (AVS) tasks while ReferFormer and R2VOS claim for Referring Video Object Segmentation(Ref-VOS) tasks. We also report the results achieved by offering the third their absent modalities, which is consistent with EEMC. EEMC is specially designed to achieve fine-grained semantic understanding in the LAVS and also proposes the Ref-AVS task. + +Implementation Details. For the training recipe, we adopt the DeepSpeed training engine with ZeRO-2 to train our model on 2 NVIDIA A100 GPUs. The AdamW [19] optimizer is harnessed with the learning rate and weight decay set to 1e-4 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the BCE loss $\lambda_{bce}$ and Dice loss $\lambda_{dice}$ are set to 1.0 and 1.0, respectively. The batch size is set to 8 and the gradient accumulation step is set to 2. For the network architecture, the fusion transformer is the stacked 6 transformer blocks with bi-directional self + +attention. The hidden dimension of the transformer is 1024 and the number of [seg] tokens is set to 1. SAM2 is initialized with the weight of SAM2-Hiera-large. + +Metrics. Following [20, 31], we assess the performance of the Ref-AVS benchmark using the Jaccard index $(\mathcal{I})$ , the F-score $(\mathcal{F})$ , and their average value $(\mathcal{I} \& \mathcal{F})$ as key evaluation metrics. + +# 4.2. Main Results + +Table 1 shows the comparative results of our method on Ref-AVS benchmarks. Our SAM2-LOVE achieves state-of-the-art (SOTA) performance and significantly outperforms the second-best method EEMC on both seen and unseen test splits. The results indicate that our model has highly generalized capabilities that can be directly used for out-of-domain scenarios. We attribute two reasons leading to these huge performance improvements. Firstly, the fusion transformer can effectively step-by-step compress the multimodal cues into the learnable tokens through the [seg] token propagation strategy and [cls] token accumulation. The generated token contains high semantics as well as positional information, which can accurately prompt SAM2 to segment the inserted targets. Secondly, SAM2 showcases strong zero-shot VOS capabilities, accurately tracking the referred object despite the circumstance where it suddenly disappears and occurs again or similar objects have interacted together. + +# 4.3. Quality Assessment + +Figure 5 shows the visualization comparative results with the previous SOTA EEMC. In the first case, we can observe that the SAM2 shows strong consistency within different frames, which can track the referred object with high + +![](images/37ca21691a1762d603ff43b7d5ce5cad8e504b1d657f29f9b3d44624c5268885.jpg) +Figure 5. The visualization results of the referred objects in the Ref-AVS. Notably, we use the legends of the trumpet to represent the different sound volumes of the audio objects, from silent to increasing loudness. More cases can be found in the supplementary. + +
CLIP-T +CLIP-VRoBERTa +ViT[cls] AccSeen (%)Unseen (%)
JFJ&FJFJ&F
✓✓42.551.246.865.171.468.2
✓✓43.551.947.766.572.369.4
✓42.650.946.766.371.669.0
+ +mask quality. Additionally, it also demonstrates that the [seg] token can accurately locate the small referring target, while EEMC fails in the first several frames. The second case shows that our SAM2-LOVE is more robust regarding changes in audible objects. EEMC may not distinguish the difference or malfunction in understanding the hidden meaning in the expressions (i.e. "shorter duration" instead of the obvious "the man"). + +# 4.4. Ablation Study + +Effect of Backbone. Table 2 shows the ablation study of backbones. Compared to CLIP vision and text encoder, RoBERTa and ViT achieve better performance, which can be the difference from the pre-trained paradigm. To align with visual elements, the CLIP textual encoder focuses more on the text with specific semantics such as cat and person. However, in the LAVS scenes, the massive expres + +Table 2. Ablations on backbone design and our proposed [cls] token accumulation strategy. + +
Number of transformer blocksSeen (%)Unseen (%)
JFJ&FJFJ&F
1 layer41.549.445.465.270.968.1
6 layers42.650.946.766.371.669.0
12 layers42.050.746.467.773.370.5
+ +Table 3. Ablations on the number of fusion transformer blocks. + +sions tend to provide a kind of control signal without referred category names, e.g., “the one making the sound.” and “the one making the most intense sound.” Therefore, the RoBERTa may demonstrate greater strength in understanding instructions due to expertise in natural language understanding during training. + +Effect of [cls] Token Accumulation. Since we harness the propagated token to prompt the first frame for SAM2, the attention for the first several frames may weaken during propagation, causing the token to depend only on the last frame. Therefore, in the fusion transformer, we attach the visual global [cls] tokens of previous frames to the current frame to update the history token, avoiding the [seg] token forgetting the information of former frames. Compared to the last row and second row in Table 2, we can observe that this strategy can effectively improve performance for free without additional design or resource consumption. + +Effect of Fusion Transformer Design. The design of + +
Number of [seg] tokensSeen (%)Unseen (%)
JFJ&FJFJ&F
142.650.946.766.371.669.0
441.350.846.065.271.068.1
840.749.845.267.072.269.6
+ +Table 4. Ablations on the numbers of [seg] tokens. + +
Fusion StrategySeen (%)Unseen (%)
JFJ&FJFJ&F
learnable token42.650.946.766.371.669.0
mean41.650.145.865.971.668.7
+ +our fusion transformer is a naive transformer architecture without additional operations such as ROPE [26] or attention enhancement [25]. Table 3 shows the different performances under different numbers of transformer blocks. Increasing the model's volume can lead to gains in the unseen test set and comparable performance in the seen test set. The single layer can also achieve satisfying results, demonstrating that the simple multimodal fusion still makes a huge impact. However, considering the additional resource consumption and balance of efficiency, we use 6 layers of transformer instead of 12. + +Effect of the Number of [seg] Tokens. We only use a single token to promote SAM2 for localization of the correct target, but the number of tokens entering the propagation process can be multiple. As shown in Table 4, increasing the number of tokens can improve significant performances in the unseen split, while this sacrifices the performances of the seen category. We think that it fits the insight from LIBERO [12]. The additional tokens can be regarded as kinds of registers that store additional global information, and this prior knowledge would facilitate learning on new tasks (Unseen tasks). On the other hand, utilizing a single token implicitly encodes global information into that token, which can enhance performance on prior tasks (Seen tasks) but could negatively impact performance on new tasks (Unseen tasks). Despite the representation learned by a single token exhibiting a stronger inductive bias, it is still useful and we set it by default. + +Effect of Fusion Strategy. The token used to prompt the SAM2 can be generated by different strategies. Besides using a learnable token, we can directly take the mean of whole the sequence that is generated by concatenating all the frames' embeddings with those of audio and text. As shown in Table 5, our learnable token can behave better since it can aggregate information from different frames gradually. In contrast, the direct mean operation may lose a + +Table 5. Ablations on the different fusion strategies for generating the prompt token for SAM2. + +
Mask decoderPrompt encoderSeen (%)
JFJ&F
**34.844.839.8
✓*42.250.746.5
✓✓42.550.946.7
+ +Table 6. Ablations on the different training modules in SAM2. \* denotes the frozen module while $\sqrt{}$ makes the module trainable. + +large amount of inter-frame information. + +Effect of Training Modules in SAM2. To preserve the strong VOS capabilities of SAM2, we freeze the memory attention modules but train the prompt encoder and mask decoder to further adapt the SAM2 to the multimodal clue reference, which can achieve the best results shown in Table 6. Without training the mask decoder, it will cause a huge performance drop due to the low capability to fit the multimodal representation of the [seg] token. + +# 4.5. Limitations + +In this section, we discuss the limitations of our method. As shown in Table 1 NULL split, our method has a relatively low capability to reject the users' requirements for a null reference. It is the inherent limitation of the prompting paradigm for localization that SAM2 attends to segment something or some irrelevant regions in the image without any meaning. SAM2-LOVE lags behind the EEMC in the null test with 0.2 in $S$ score, achieving 0.23 similar results with the AVSBench plus text method. The $S$ denotes the square root of the ratio of the number of foreground pixels over the background pixels. This can be our future work on how to solve the null reference problem. + +# 5. Conclusion + +This paper introduces SAM2-LOVE, a novel framework for segmenting reference objects in Language-aided Audio-Visual Scenes, overcoming the limitations of previous dual-modality SAM approaches in handling scenarios that integrate three modalities. We propose a multimodal fusion module that compresses information from all modalities into a single token, which then prompts SAM2 for video segmentation. This compressed [seg] token propagates forward through the video to capture both the spatial and temporal features, while also referring to the accumulated historical tokens backward preserving information from previous frames. Extensive experiments demonstrate the superiority of our model on the Ref-AVS benchmark and highlight the efficiency of our design. Future work will focus on addressing scenarios where the three modality references are misaligned, such as the null reference problem. + +# Acknowledgement + +This work was supported in part by National Natural Science Foundation of China (Grant No. 62206153), and in part by Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012). + +# References + +[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2 +[2] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26497–26507, 2024. 3 +[3] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[4] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12155–12163, 2024. 3, 6 +[5] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 776-780. IEEE, 2017. 3 +[6] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 131-135. IEEE, 2017. 3 +[7] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing, pages 787-798, 2014. 2 +[8] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3 +[9] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2 +[10] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19108-19118, 2022. 1, 3 + +[11] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22236-22245, 2023. 6 +[12] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 8 +[13] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2 +[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[16] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364, 2019. 3 +[17] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3500, 2024. 2 +[18] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 2 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[20] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 6 +[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In IEEE European Conference on Computer Vision, pages 792-807. Springer, 2016. 2 +[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dolar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3 +[23] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2 +[24] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, + +faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.3 +[25] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 8 +[26] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 8 +[27] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Autoacid: A large-scale dataset for audio-language representation learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5025-5034, 2024. 3 +[28] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929, 2023. 6 +[29] Yuji Wang, Ruojun Zhao, and Zijun Sun. Efficient remote sensing transformer for coastline detection with sentinel-2 satellite imagery. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium, pages 5439-5442. IEEE, 2023. 2 +[30] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5669–5677, 2024. 2, 3 +[31] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision, 2024. 1, 2, 3, 5, 6 +[32] Yuji Wang, Ruojun Zhao, Shicai Wei, Jingchen Ni, Meng Wu, Yang Luo, and Chunbo Luo. Convolution meets transformer: Efficient hybrid transformer for semantic segmentation with very high resolution imagery. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, pages 9688-9691. IEEE, 2024. 2 +[33] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. arXiv preprint arXiv:2503.00936, 2025. 2 +[34] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2 +[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6 +[36] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2 + +[37] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024. 2 +[38] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16111-16121, 2024. 3 +[39] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audiovisual question answering on videos. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3480-3491, 2022. 1, 3 +[40] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[41] Zhao Yang, Jiaqi Wang, Xubing Ye, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Language-aware vision transformer for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[42] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3 +[43] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Lei Liu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Evf-sam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2, 3 +[44] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 3 +[45] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 3, 6 +[46] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audiovisual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 3 +[47] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2 \ No newline at end of file diff --git a/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/images.zip b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6cd1ee22615c9d2c8c548cb3b724a32bc3daa3ee --- /dev/null +++ b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87633e4fc2c7b45618bd1b4de1a2d4b5928a5759fee0517148c30daaaef8403e +size 647033 diff --git a/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/layout.json b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6cec0856a26c016d1859193c45671fe91d7f64be --- /dev/null +++ b/2025/SAM2-LOVE_ Segment Anything Model 2 in Language-aided Audio-Visual Scenes/layout.json @@ -0,0 +1,6724 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 102, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 102, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 102, + 496, + 138 + ], + "type": "text", + "content": "SAM2-LOVE: Segment Anything Model 2 in Language-aided Audio-Visual Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "text", + "content": "Yuji Wang" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Haoran Xu" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Yong Liu" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Jiaze Li" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "text", + "content": ", Yansong Tang" + }, + { + "bbox": [ + 149, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 175, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 175, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 114, + 175, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 175, + 496, + 190 + ], + "type": "text", + "content": "Tsinghua Shenzhen International Graduate School, Tsinghua University, " + }, + { + "bbox": [ + 114, + 175, + 496, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 175, + 496, + 190 + ], + "type": "text", + "content": "ZJU" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 191, + 482, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 191, + 482, + 202 + ], + "spans": [ + { + "bbox": [ + 124, + 191, + 482, + 202 + ], + "type": "text", + "content": "yuji-wan24@mails.tsinghua.edu.cn, †tang.yansong@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "type": "text", + "content": "Reference Audio-Visual Segmentation (Ref-AVS) aims to provide a pixel-wise scene understanding in Language-aided Audio-Visual Scenes (LAVS). This task requires the model to continuously segment objects referred to by text and audio from a video. Previous dual-modality methods always fail due to the lack of a third modality and the existing triple-modality method struggles with spatio-temporal consistency, leading to the target shift of different frames. In this work, we introduce a novel framework, termed SAM2-LOVE, which integrates textual, audio, and visual representations into a learnable token to prompt and align SAM2 for achieving Ref-AVS in the LAVS. Technically, our approach includes a multimodal fusion module aimed at improving multimodal understanding of SAM2, as well as token propagation and accumulation strategies designed to enhance spatio-temporal consistency without forgetting historical information. We conducted extensive experiments to demonstrate that SAM2-LOVE outperforms the SOTA by " + }, + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "type": "inline_equation", + "content": "8.5\\%" + }, + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 54, + 256, + 297, + 507 + ], + "type": "text", + "content": " on the Ref-AVS benchmark and showcase the simplicity and effectiveness of the components. Our code will be available here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 532, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 684 + ], + "type": "text", + "content": "In the process of human perception of the world, visual signals play a primary role, often accompanied by other multimodal signals such as audio and text. These additional multimodal cues are not only essential to locating and separating objects and areas of interest but also to understanding the changing behaviors of the objects. The dynamic scene that integrates visual elements with diverse multimodal cues including audio and text is regarded as the Language-aided Audio-Visual Scene (LAVS). To mimic the effect from humans, prior works such as Music-AVQA [10] and AVQA [39] have primarily focused on question-answering" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 235, + 441, + 389 + ], + "blocks": [ + { + "bbox": [ + 318, + 235, + 441, + 389 + ], + "lines": [ + { + "bbox": [ + 318, + 235, + 441, + 389 + ], + "spans": [ + { + "bbox": [ + 318, + 235, + 441, + 389 + ], + "type": "image", + "image_path": "d43998c6c8420773f1678dab1fa83c90aa7d0dd726802510d6d658a6cdbe7cc9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 446, + 235, + 561, + 389 + ], + "blocks": [ + { + "bbox": [ + 446, + 235, + 561, + 389 + ], + "lines": [ + { + "bbox": [ + 446, + 235, + 561, + 389 + ], + "spans": [ + { + "bbox": [ + 446, + 235, + 561, + 389 + ], + "type": "image", + "image_path": "06d5c19d02f822cd0b15cd3d50fa6411708bdf91eb3db7d58eb3f748a96c7946.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 391, + 558, + 566 + ], + "blocks": [ + { + "bbox": [ + 318, + 391, + 558, + 566 + ], + "lines": [ + { + "bbox": [ + 318, + 391, + 558, + 566 + ], + "spans": [ + { + "bbox": [ + 318, + 391, + 558, + 566 + ], + "type": "image", + "image_path": "2bcfeafbdb1a56205f9621532a4faaed69a074b05c106f118d0bea3553142c43.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 577, + 555, + 633 + ], + "lines": [ + { + "bbox": [ + 313, + 577, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 555, + 633 + ], + "type": "text", + "content": "Figure 1. The comparison with dual-modality SAM pipelines. (a) The text-visual SAM struggles to distinguish the sounding object in the silent scenes. (b) The audio-visual SAM fails to understand the dynamic control from the text input. (c) Our model fuses three modalities' information and expands to the video due to SAM2." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "tasks. Nevertheless, these approaches lack a fine-grained understanding of LVAS at the pixel level. As shown in Figure 2, Reference Audio-Visual Segmentation (Ref-AVS) task [31] makes a good example to improve the pixel-wise" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28932" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 73, + 538, + 250 + ], + "blocks": [ + { + "bbox": [ + 71, + 73, + 538, + 250 + ], + "lines": [ + { + "bbox": [ + 71, + 73, + 538, + 250 + ], + "spans": [ + { + "bbox": [ + 71, + 73, + 538, + 250 + ], + "type": "image", + "image_path": "16be766b43381123fdbeb8a4fffae0994bc3b2499bb896eabe1d7763492a799e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 261, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 297 + ], + "type": "text", + "content": "Figure 2. Illustration of the Ref-AVS task under varying modality cues. The expression can be treated as a kind of control signal to condition the referred object on audio. Different expressions can test the models' pixel-wise scene understanding capabilities in the LAVS with different dimensions. The absence angle of a triangle means that modality is not the main test dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": "scene understanding of the model in the LAVS. The previous duality-referring segmentation works struggle with the Ref-AVS task where the interested object can only be located under triple modalities, shown in Figure 1. In neither Figure 1 (a) text-visual EVF-SAM [43] nor Figure 1 (b) audio-visual GAVS [30], the truly referred target is accurately segmented due to the absence of additional help from the other third modality. Additionally, they are constrained in the image level, which can not be directly adapted to the dynamic nature of video content. Therefore, methods for effectively fusing information from the three modalities for segmentation have garnered widespread attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 664 + ], + "type": "text", + "content": "Prior work, EEMC [31], simultaneously models three modalities' representations to achieve excellent results on the Ref-AVS benchmark. However, it showcases the relatively insufficient spatio-temporal consistency. Specifically, despite the existence of memory caches, the model fails to continuously track the target's position and shape, leading to a shift in the segmented area over time. Recently, SAM2 [22] has been proposed to generalize the segmenting anything capabilities of the image to the video, which demonstrates a powerful video tracking capability in interframe consistency. SAM2 follows the previous promptable visual segmentation of SAM [8] that requires a strong manual annotation to the interested target of the starting frame (not only limited to the first frame) and then propagates to the whole video based on the crucial frame. Nevertheless, the lack of dependency on other modalities (text or audio) constraints its further application in the LAVS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Thus, in this paper, we propose a novel framework, namely SAM2-LOVE, which means Segment Anything Model 2 in Language-aided audio-O-Visual scEnes. To enhance SAM2's understanding of multimodal information," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 399 + ], + "type": "text", + "content": "we design the multimodal fusion module to compress all the modalities' information into a learnable token that aligns the SAM2 for segmenting the referring target. Additionally, we incorporate token propagation and accumulation strategies to strengthen spatio-temporal understanding of the video while avoiding forgetting historical frames. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 401, + 553, + 509 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 401, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 401, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 401, + 553, + 436 + ], + "type": "text", + "content": "- We propose a novel framework, SAM2-LOVE that firstly leverages SAM2 to achieve pixel-wise understanding in the LAVS by designing a multimodal fusion module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 437, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 553, + 472 + ], + "type": "text", + "content": "- We develop creative token propagation and accumulation strategies to improve spatio-temporal comprehension of the promtable token." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 473, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 553, + 509 + ], + "type": "text", + "content": "- Extensive experiments on Ref-AVS dataset demonstrate the superiority of our method, with ablation studies highlighting the simplicity and effectiveness of its modules." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 524, + 405, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 524, + 405, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 524, + 405, + 537 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "Segmentation via Language or Audio. Unlike semantic segmentation [1, 17, 29, 32], visual segmentation via language starts with referring expression segmentation task [7, 18, 21, 33, 34, 36, 40, 41] which performs target object segmentation based on given natural language expressions. X-Decoder [47] bridges vision and language, unifying multiple tasks within a single model. Moreover, GRES [13] defines a new task called generalized reference expression segmentation. With the development of multimodal large language models (MLLMs) such as LLaVA [14, 15], LISA [9] proposes the concept of reasoning segmentation with combination of pixel-level understanding and MLLMs. GSVA [37] and PixelLM [23] propose multi-target and null-target reasoning segmentation. On the other" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28933" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "hand, audio-visual segmentation (AVS) is also constantly improving. Audio-Visual Segmentation [45, 46] proposes the AVS problem and AVSBench dataset. Numerous subsequent methods [2, 4, 27] have been developed and refined based on the proposed benchmark." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 312 + ], + "type": "text", + "content": "Segment Anything Model. SAM [8] is an interactive segmentation model based on various types of prompts (points, boxes, coarse masks). It shows significant generalization ability in open-world segmentation. Several works strive to reduce the massive computation cost of SAM, such as EfficientSAM [38], FastSAM [44], and MobileSAM [42]. The advent of SAM2 [22] extends SAM's functionality to include video segmentation. While SAM and SAM2 perform effectively in visual segmentation tasks using box, point, or mask prompts, they currently do not possess language understanding capabilities, rendering the direct use of text prompts for referring segmentation infeasible. EVF-SAM [43] extends SAM's capabilities with text-prompted segmentation, while GAVS [30] equips the SAM with image-audio fusion to achieve AVS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 313, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 479 + ], + "type": "text", + "content": "Language-aided Audio-Visual Scene. There is presently a scarcity of publicly available tasks that provide datasets for audio-visual scene comprehension with language-based support. The complexity of language-aided audio-visual scene understanding arises from its reliance on three modalities (vision, audio, and text) necessitating massive matched data and their corresponding labels, posing significant challenges for data collection and model design. MusicAVQA [10] and AVQA [39] are two of the classic audio-visual question-answering (QA) datasets. However, the above work focuses on multimodal QA and spatio-temporal reasoning of audio-visual scenes involving various relationships between objects or activities. Ref-AVS [30, 31] begins pixel-level understanding of dynamic audio-visual scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 481, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 529 + ], + "type": "text", + "content": "In summary, no methods exist for integrating the foundational visual segmentation model SAM2 into language-based audio-visual scenes. Therefore, to achieve pixel-level understanding in the LAVS, we introduce SAM2-LOVE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 542, + 111, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 111, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 111, + 554 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 562, + 179, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 179, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 179, + 574 + ], + "type": "text", + "content": "3.1. Preliminary of SAM2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 652 + ], + "type": "text", + "content": "SAM2 can be treated as a generalization of SAM to segment anything in images and videos. It also supports visual prompts (point, box, and mask) to indicate the objects to be segmented in an individual frame (image as a single frame). Then, it generalizes across the whole video based on the prompted frame to achieve the video segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "As shown in Figure 3 (Inference), this model begins with a hierarchical image encoder that encodes each frame into multilevel embeddings. Unlike SAM which directly feeds these features to the mask decoder for predictions, SAM2 incorporates a series of transformer blocks (i.e. memory" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "content": "attention) to condition the current image feature on memories of previous and prompted frames stored in a memory bank that is maintained by a FIFO (First In First Out) queue. Each memory entry is composed of two kinds of memories: (1) spatial feature maps that are generated through an element-wise summation of the downsampled and encoded mask features from the memory encoder and the unconditioned frame embedding from the image encoder, and (2) the object pointers that are obtained from the output tokens within the mask decoder contain high-level semantic information of the interested target. Therefore, the memory attention operation in the transformer makes the current frame's feature cross-attend both spatial fine-grained memory and abstract temporal object pointers. Second, the mask decoder of SAM2 largely follows the design of SAM, which is a \"two-way\" transformer that takes the prompt embeddings from the prompt encoder and frame embeddings as inputs. The difference lies in the additional occlusion score that aims to judge the object's presence in the current frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 299, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 553, + 370 + ], + "type": "text", + "content": "Thanks to the strong and generalized capability of SAM2 across images and videos, we can first compress all the information contained in three modalities into a single token and then use it to prompt SAM2 on a single frame. Finally, starting from this crucial frame, we can propagate through video to obtain the desired masks of all frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 378, + 400, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 400, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 400, + 389 + ], + "type": "text", + "content": "3.2. SAM2-LOVE" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 443 + ], + "type": "text", + "content": "The overall pipeline of our model is shown in Figure 3, which integrates multimodal information in the LAVS into a representative token and uses it to provide a visual clue for SAM2 to segment the reference targets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": "Multimodal Representations. Following the previous work [31], for each modality, we use three popular encoders to separately extract their representations, which were not aligned before in their pretraining stages. VGGish [5, 6] encodes an audio clip into audio embeddings " + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "inline_equation", + "content": "F_{A} \\in R^{N \\times d_{A}}" + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": " denotes the duration of the audio in seconds and corresponds to the total number of frames in the video. The sampled " + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": " frames are encoded into " + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "inline_equation", + "content": "F_{V} \\in R^{N \\times d_{V} \\times h \\times w}" + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": " by a ViT [3] pre-trained on ImageNet. For the text expressions, we utilize DistilRoBERTa [16, 24] to obtain the textual representations " + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "inline_equation", + "content": "F_{T} \\in R^{P \\times d_{T}}" + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 444, + 554, + 611 + ], + "type": "text", + "content": " is the number of textual tokens. To unify and align all embeddings into the same semantic space, we use three MLPs to project them into a shared dimension respectively, shown in Equation (1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 618, + 553, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 618, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 320, + 618, + 553, + 633 + ], + "type": "interline_equation", + "content": "\\widehat {F} _ {Z} = \\operatorname {L i n e a r} (\\operatorname {R e L U} (\\operatorname {L i n e a r} (F _ {Z}))), Z \\in \\{A, V, T \\} \\tag {1}", + "image_path": "ee8238acdf04268fa3b54003ebecf106d5cbc1069f1ca93b96c910ca1b4ce791.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Multimodal Fusion. As shown in Figure 3, a multimodal fusion module is proposed to compress all three modalities' information into a single token and then leverage its embedding to represent the fused features. We define this token as a learnable segmentation token, [seg], designed to prompt SAM2 as a clue to retrieve the targeted" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28934" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 545, + 383 + ], + "type": "image", + "image_path": "4768acffa4521db7f0bbbd66a3a58d4a7e5f2a3e62165d8db31ba80981dfa202.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 397, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 54, + 397, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 54, + 397, + 555, + 444 + ], + "type": "text", + "content": "Figure 3. The pipeline of SAM2-LOVE. We harness a fusion transformer to compress the tripe modality representations into a learnable token [seg]. For the input, [aud] and [vis] are added to indicate their corresponding modality sequences, and [his] is to provide additional information about former frames. In the training stage, we use the first frame to align the SAM2 with the compressed and fused representation of [seg]. During inference, we harness the strong zero-shot capabilities of SAM2 to track the referred target at the pixel level." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 464, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 510 + ], + "type": "text", + "content": "object. The fusion module is a naive transformer encoder architecture with " + }, + { + "bbox": [ + 55, + 464, + 295, + 510 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 464, + 295, + 510 + ], + "type": "text", + "content": " stacked bi-directional transformer block that contains a multi-head self-attention (MSA), Layernorm (LN), and MLP, formulated below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 515, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 515, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 87, + 515, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {l} ^ {\\prime} = M S A \\left(L N \\left(z _ {l - 1}\\right)\\right) + z _ {l - 1}, \\quad l = 1 \\dots L \\tag {2} \\\\ z _ {l} = M L P (L N (z _ {l} ^ {\\prime})) + z _ {l} ^ {\\prime}, \\quad l = 1 \\dots L \\\\ \\end{array}", + "image_path": "b57fd71335c17f65257439612f06410026f1bbeb43f32f9d4f31be87c1903515.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "type": "inline_equation", + "content": "z_{l-1}" + }, + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "type": "text", + "content": " is the input sequence of " + }, + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 545, + 241, + 556 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": "Token Propagation. The behavior of the [seg] token is a self-propagation process within the whole video, which adaptively captures the inner-frame spatial features and the strong continuity among consecutive frames to achieve spatio-temporal consistency. To be more specific, after obtaining the three projected multimodal representations " + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\widehat{F}_A" + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\widehat{F}_T" + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\widehat{F}_V" + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": ", we first prepend the [seg] token to the head of the sequence of modalities' embeddings to obtain the multimodal sequence " + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "inline_equation", + "content": "F_M^i" + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 557, + 295, + 666 + ], + "type": "text", + "content": "-th frame, shown as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 670, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 670, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 670, + 295, + 685 + ], + "type": "interline_equation", + "content": "F _ {M} ^ {i} = \\operatorname {C o n c a t} ([ [ s e g ]; \\widehat {F} _ {A}; [ a u d ]; \\widehat {F} _ {T}; [ v i s ]; \\widehat {F} _ {V} ^ {i} ]), \\tag {3}", + "image_path": "66ff4fd61ee623855d460bec8b0f86ab09cd6bbb6bf93cfd060db199259420d3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\text{Concat}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " denotes the concatenation operation in the feature dimension axis, two fixed tokens, [aud] and [vis]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": "indicate the end and beginning of their corresponding sequences, respectively. This concatenated sequence will be fed into the fusion transformer for updating the status of the [seg] token, allowing the token to capture the visual information of the current frame along with the integral audio and text representations. Since the transformer is the bidirectional encoder that enables all the tokens to interact with each other, we can harness the single token to serve as the multimodal representation of the whole sequence. We take the first element from the output of the fusion transformer " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "Q_{i} = \\text{Transformer}(F_{M}^{i})" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": ", as the updated [seg] token " + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "inline_equation", + "content": "[seg] = Q_{i}^{0}" + }, + { + "bbox": [ + 313, + 463, + 555, + 668 + ], + "type": "text", + "content": ", shown in Figure 4. Therefore, the token not only contains the current frame information but also the compressed information of the previous frames. Our proposed token propagation strategy allows it to model the video from both the spatial and temporal aspects and aggregate the information within the text and audio." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Token Accumulation. Although the token propagation strategy can effectively obtain the multimodal characteristics and be used to prompt the SAM2, it still faces the issue" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28935" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 76, + 550, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 550, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 550, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 550, + 150 + ], + "type": "image", + "image_path": "a23f14fa9ce23cfdbee60a3f04933d9d121a88dac2bab81b561f7f256804cba7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 159, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 54, + 159, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 555, + 205 + ], + "type": "text", + "content": "Figure 4. The token propagation and token accumulation strategy in the fusion transformer. For the token propagation, the output [seg] of the current step serves as the input for the next step, which interacts with different frame features but the same audio and text features. To offer additional historical information, the token accumulation strategy appends the global representations of all the previous frames to update the maintained history token [his]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "type": "text", + "content": "that the information of the first several frames can be diluted and forgotten with the increasing propagation times. To address the issue, we propose the [cls] token accumulation strategy, in which we keep a history of token sequences, [his] to store the global and abstract representation of historical frames, shown in Figure 4. It is achieved that when propagating to the " + }, + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "type": "text", + "content": "-th frame, we provide additional global information of all the previous frames by appending and concatenating the [his] tokens into " + }, + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "type": "inline_equation", + "content": "F_{M}^{i}" + }, + { + "bbox": [ + 54, + 225, + 294, + 345 + ], + "type": "text", + "content": ". The updation of the history token [his] can be formulated by Equation (4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 354, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 354, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 89, + 354, + 295, + 369 + ], + "type": "interline_equation", + "content": "[ h i s ] ^ {i} = C o n c a t \\left(\\left[ [ c l s ] ^ {0}; [ c l s ] ^ {1}; \\dots ; [ c l s ] ^ {i} \\right]\\right) \\tag {4}", + "image_path": "1a608b69115607fe2e84d6ff07edb2807aa0740188b8ee587ec36254386eee0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "inline_equation", + "content": "[cls]^i" + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "text", + "content": " is the global representation of " + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "text", + "content": "-th frame from ViT. These historical tokens can be referred to by the " + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "inline_equation", + "content": "[seg]" + }, + { + "bbox": [ + 55, + 376, + 296, + 424 + ], + "type": "text", + "content": " of the current stage, promising the information of all the previous frames is not lost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 568 + ], + "type": "text", + "content": "Unified Perspective of Token Behaviors. We interpret the two token operations described above from a unified knowledge transfer perspective. Token propagation of the [seg] token is a direct flow, which expects the knowledge from different frames to be gradually connected through the propagation process. This is the forward knowledge transfer. In contrast, token accumulation is the replay of [his] tokens, a reversed flow to prevent the [seg] token from forgetting the knowledge of historical frames as it propagates forward. This is the backward knowledge transfer. We propose that these two token operations are functionally opposed yet conceptually unified." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 119, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 119, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 119, + 589 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Trainable Modules. The encoders used for multimodal representations for audio, video, and text are fully frozen, perceiving the pre-trained knowledge. However, the multimodal fusion transformer and [seg] token are fully trainable during our training process, allowing them to learn how to generate the fused and aligned embeddings tailored for prompting SAM2. For SAM2, we keep the image encoder and memory attention modules frozen while training the prompt encoder and mask decoder. Additionally, all the MLPs for projecting are trainable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 225, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 452 + ], + "type": "text", + "content": "Training Strategy. Our training strategy initially activates SAM2's multimodal prompting capability at the image level and then leverages its extensive pre-trained knowledge for zero-shot video segmentation. In this setting, the fusion transformer and [seg] token are trained with full access to the entire video sequence, while SAM2 is exposed only to the first frame during training. Specifically, we consolidate multimodal information (video, audio, and text) through the fusion module, which compresses these features into tokens that serve as prompts for SAM2 on the first frame, shown in Figure 3 (Training). Then, a supervised loss is applied to this frame to train SAM2's referring segmentation abilities. During testing, both the fusion module and SAM2 receive the entire video, allowing for comprehensive segmentation across all frames. This approach enables SAM2 to effectively utilize multimodal prompts on the initial frame and then generalize to the complete video sequence through its zero-shot Video Object Segmentation (VOS) capabilities, shown in Figure 3 (Inference)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "text", + "content": "Loss Function. The model is trained end-to-end under the supervised mask loss " + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "text", + "content": ", to encourage the SAM2 to predict the correct localized masks. To compute this loss, we combine the per-pixel binary cross-entropy (BCE) loss and DICE loss with their corresponding weights " + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 313, + 453, + 555, + 525 + ], + "type": "text", + "content": ", which can be formulated by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 529, + 555, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 529, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 325, + 529, + 555, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathbf {B C E} (\\hat {M}, M) + \\lambda_ {d i c e} \\mathbf {D I C E} (\\hat {M}, M) \\tag {5}", + "image_path": "98eaa3965c0d97d2599b09551a9cdd422d15fa9134a68323884859dc03c7b96d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "type": "text", + "content": " is a single mask prediction for the first frame and " + }, + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 547, + 554, + 572 + ], + "type": "text", + "content": " is the label of the first frame for supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 600, + 441, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 441, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 441, + 613 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Dataset. We evaluate our method on the Ref-AVS benchmark [31], a substantial collection of 4,000 videos with manual pixel-level annotations and 20,000 expressions reflecting the scenarios in audio, vision, and time dimensions. Additionally, the collected videos contain abundant and diverse visual elements. Each video encompasses a high average number of objects about 1.72, a wide variety of audible objects in 48 categories, and 3 kinds of static objects." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28936" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 558, + 267 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 558, + 267 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 558, + 267 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 558, + 267 + ], + "type": "table", + "html": "
MethodSeen (%)Unseen (%)Mix (S+U)NULLVENUE
JFJ&FJFJ&FJFJ&Fs
AVSBench [45]21.245.733.527.750.939.324.548.336.4-ECCV'2022
+text23.251.137.232.454.743.527.852.940.30.208
AVSegFormer [4]29.242.335.732.745.138.930.943.737.3-AAAI'2024
+text33.547.040.236.150.143.134.848.641.70.171
GAVS [28]24.845.535.227.846.036.926.345.736.0-AAAI'2024
+text28.949.839.429.849.739.829.449.839.60.190
ReferFormer [35]27.447.737.524.950.437.626.249.137.6-CVPR'2022
+audio31.350.140.730.448.839.630.949.540.20.176
R2VOS [11]21.330.926.123.943.133.522.637.029.8-ICCV'2023
+audio25.041.033.027.949.838.926.545.435.90.183
EEMC [31]34.251.342.849.564.857.241.958.150.00.007ECCV'2024
SAM2-LOVE (ours)43.551.947.766.572.369.455.062.158.50.23CVPR2025
vs. prev. SOTA+9.3+0.6+4.9+17.0+7.5+12.2+13.1+4.0+8.5--
", + "image_path": "7a66c8bbd9c4c2535713681bb3b04da20bb1cf2488d4538f5afc6b74b40c9583.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 555, + 300 + ], + "type": "text", + "content": "Table 1. Performance comparison across different methods in Seen, Unseen, Mix (S+U), and NULL settings of Ref-AVS benchmark. The mix indicates the average value of seen and unseen splits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 319, + 294, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 294, + 440 + ], + "type": "text", + "content": "Therefore, the tested scenarios are more challenging with multiple sound sources and multiple semantics. The dataset is divided into three splits, a training set of 2908 videos, a validation set of 276 videos, and a test set of 818 videos. The test set is further divided into three distinct subsets: 1) seen splits that are composed of the same 39 categories in the training set. 2) unseen splits with 13 additional categories that do not appear in the training stage. 3) null splits to test the model when an expression refers to an object that either does not exist or is not visible in the given context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 445, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 577 + ], + "type": "text", + "content": "Baselines. We compare our method with six challenging baselines: AVSBench [45], AVSegFormer [4], GAVS [28], ReferFormer [35], R2VOS [11] and EEMC [31]. AVSBench, AVSegFormer, and GAVS are designed for Audio-Visual Segmentation (AVS) tasks while ReferFormer and R2VOS claim for Referring Video Object Segmentation(Ref-VOS) tasks. We also report the results achieved by offering the third their absent modalities, which is consistent with EEMC. EEMC is specially designed to achieve fine-grained semantic understanding in the LAVS and also proposes the Ref-AVS task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. For the training recipe, we adopt the DeepSpeed training engine with ZeRO-2 to train our model on 2 NVIDIA A100 GPUs. The AdamW [19] optimizer is harnessed with the learning rate and weight decay set to 1e-4 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the BCE loss " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " and Dice loss " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " are set to 1.0 and 1.0, respectively. The batch size is set to 8 and the gradient accumulation step is set to 2. For the network architecture, the fusion transformer is the stacked 6 transformer blocks with bi-directional self" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 355 + ], + "type": "text", + "content": "attention. The hidden dimension of the transformer is 1024 and the number of [seg] tokens is set to 1. SAM2 is initialized with the weight of SAM2-Hiera-large." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "text", + "content": "Metrics. Following [20, 31], we assess the performance of the Ref-AVS benchmark using the Jaccard index " + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "text", + "content": ", the F-score " + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "text", + "content": ", and their average value " + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "inline_equation", + "content": "(\\mathcal{I} \\& \\mathcal{F})" + }, + { + "bbox": [ + 313, + 356, + 555, + 403 + ], + "type": "text", + "content": " as key evaluation metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 409, + 400, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 400, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 400, + 421 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 556, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 643 + ], + "type": "text", + "content": "Table 1 shows the comparative results of our method on Ref-AVS benchmarks. Our SAM2-LOVE achieves state-of-the-art (SOTA) performance and significantly outperforms the second-best method EEMC on both seen and unseen test splits. The results indicate that our model has highly generalized capabilities that can be directly used for out-of-domain scenarios. We attribute two reasons leading to these huge performance improvements. Firstly, the fusion transformer can effectively step-by-step compress the multimodal cues into the learnable tokens through the [seg] token propagation strategy and [cls] token accumulation. The generated token contains high semantics as well as positional information, which can accurately prompt SAM2 to segment the inserted targets. Secondly, SAM2 showcases strong zero-shot VOS capabilities, accurately tracking the referred object despite the circumstance where it suddenly disappears and occurs again or similar objects have interacted together." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 648, + 430, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 430, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 430, + 661 + ], + "type": "text", + "content": "4.3. Quality Assessment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "Figure 5 shows the visualization comparative results with the previous SOTA EEMC. In the first case, we can observe that the SAM2 shows strong consistency within different frames, which can track the referred object with high" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28937" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 78, + 545, + 340 + ], + "blocks": [ + { + "bbox": [ + 61, + 78, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 61, + 78, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 61, + 78, + 545, + 340 + ], + "type": "image", + "image_path": "37ca21691a1762d603ff43b7d5ce5cad8e504b1d657f29f9b3d44624c5268885.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 353, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 353, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 555, + 376 + ], + "type": "text", + "content": "Figure 5. The visualization results of the referred objects in the Ref-AVS. Notably, we use the legends of the trumpet to represent the different sound volumes of the audio objects, from silent to increasing loudness. More cases can be found in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 393, + 306, + 452 + ], + "blocks": [ + { + "bbox": [ + 58, + 393, + 306, + 452 + ], + "lines": [ + { + "bbox": [ + 58, + 393, + 306, + 452 + ], + "spans": [ + { + "bbox": [ + 58, + 393, + 306, + 452 + ], + "type": "table", + "html": "
CLIP-T\nCLIP-VRoBERTa\nViT[cls] AccSeen (%)Unseen (%)
JFJ&FJFJ&F
✓✓42.551.246.865.171.468.2
✓✓43.551.947.766.572.369.4
✓42.650.946.766.371.669.0
", + "image_path": "3bd04c94dc8ef108f8a881274ff8583763c4a3664bad4ab55a20c3786af8a57b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 505, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 601 + ], + "type": "text", + "content": "mask quality. Additionally, it also demonstrates that the [seg] token can accurately locate the small referring target, while EEMC fails in the first several frames. The second case shows that our SAM2-LOVE is more robust regarding changes in audible objects. EEMC may not distinguish the difference or malfunction in understanding the hidden meaning in the expressions (i.e. \"shorter duration\" instead of the obvious \"the man\")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 149, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 149, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 149, + 624 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Effect of Backbone. Table 2 shows the ablation study of backbones. Compared to CLIP vision and text encoder, RoBERTa and ViT achieve better performance, which can be the difference from the pre-trained paradigm. To align with visual elements, the CLIP textual encoder focuses more on the text with specific semantics such as cat and person. However, in the LAVS scenes, the massive expres" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 331, + 393, + 538, + 452 + ], + "blocks": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": "Table 2. Ablations on backbone design and our proposed [cls] token accumulation strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 393, + 538, + 452 + ], + "lines": [ + { + "bbox": [ + 331, + 393, + 538, + 452 + ], + "spans": [ + { + "bbox": [ + 331, + 393, + 538, + 452 + ], + "type": "table", + "html": "
Number of transformer blocksSeen (%)Unseen (%)
JFJ&FJFJ&F
1 layer41.549.445.465.270.968.1
6 layers42.650.946.766.371.669.0
12 layers42.050.746.467.773.370.5
", + "image_path": "7af92675ffdf280234d87ad36f1a3d90a6e767032a7883fbc1513a453a30cb69.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 460, + 548, + 470 + ], + "lines": [ + { + "bbox": [ + 319, + 460, + 548, + 470 + ], + "spans": [ + { + "bbox": [ + 319, + 460, + 548, + 470 + ], + "type": "text", + "content": "Table 3. Ablations on the number of fusion transformer blocks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 495, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 567 + ], + "type": "text", + "content": "sions tend to provide a kind of control signal without referred category names, e.g., “the one making the sound.” and “the one making the most intense sound.” Therefore, the RoBERTa may demonstrate greater strength in understanding instructions due to expertise in natural language understanding during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 568, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 556, + 700 + ], + "type": "text", + "content": "Effect of [cls] Token Accumulation. Since we harness the propagated token to prompt the first frame for SAM2, the attention for the first several frames may weaken during propagation, causing the token to depend only on the last frame. Therefore, in the fusion transformer, we attach the visual global [cls] tokens of previous frames to the current frame to update the history token, avoiding the [seg] token forgetting the information of former frames. Compared to the last row and second row in Table 2, we can observe that this strategy can effectively improve performance for free without additional design or resource consumption." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "Effect of Fusion Transformer Design. The design of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28938" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 288, + 139 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 139 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 139 + ], + "type": "table", + "html": "
Number of [seg] tokensSeen (%)Unseen (%)
JFJ&FJFJ&F
142.650.946.766.371.669.0
441.350.846.065.271.068.1
840.749.845.267.072.269.6
", + "image_path": "6e3cd72259512e56a5024b23afa4e2babf818b3625e20a4eeb7527589598c100.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 171, + 287, + 228 + ], + "blocks": [ + { + "bbox": [ + 83, + 147, + 268, + 159 + ], + "lines": [ + { + "bbox": [ + 83, + 147, + 268, + 159 + ], + "spans": [ + { + "bbox": [ + 83, + 147, + 268, + 159 + ], + "type": "text", + "content": "Table 4. Ablations on the numbers of [seg] tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 171, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 64, + 171, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 64, + 171, + 287, + 228 + ], + "type": "table", + "html": "
Fusion StrategySeen (%)Unseen (%)
JFJ&FJFJ&F
learnable token42.650.946.766.371.669.0
mean41.650.145.865.971.668.7
", + "image_path": "cf3294315dd1774eb1c641d4962c0712d13976e4c12320c69b109c185b498a3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 281, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 411 + ], + "type": "text", + "content": "our fusion transformer is a naive transformer architecture without additional operations such as ROPE [26] or attention enhancement [25]. Table 3 shows the different performances under different numbers of transformer blocks. Increasing the model's volume can lead to gains in the unseen test set and comparable performance in the seen test set. The single layer can also achieve satisfying results, demonstrating that the simple multimodal fusion still makes a huge impact. However, considering the additional resource consumption and balance of efficiency, we use 6 layers of transformer instead of 12." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 413, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 413, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 413, + 295, + 616 + ], + "type": "text", + "content": "Effect of the Number of [seg] Tokens. We only use a single token to promote SAM2 for localization of the correct target, but the number of tokens entering the propagation process can be multiple. As shown in Table 4, increasing the number of tokens can improve significant performances in the unseen split, while this sacrifices the performances of the seen category. We think that it fits the insight from LIBERO [12]. The additional tokens can be regarded as kinds of registers that store additional global information, and this prior knowledge would facilitate learning on new tasks (Unseen tasks). On the other hand, utilizing a single token implicitly encodes global information into that token, which can enhance performance on prior tasks (Seen tasks) but could negatively impact performance on new tasks (Unseen tasks). Despite the representation learned by a single token exhibiting a stronger inductive bias, it is still useful and we set it by default." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Effect of Fusion Strategy. The token used to prompt the SAM2 can be generated by different strategies. Besides using a learnable token, we can directly take the mean of whole the sequence that is generated by concatenating all the frames' embeddings with those of audio and text. As shown in Table 5, our learnable token can behave better since it can aggregate information from different frames gradually. In contrast, the direct mean operation may lose a" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 342, + 70, + 528, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 235, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 258 + ], + "type": "text", + "content": "Table 5. Ablations on the different fusion strategies for generating the prompt token for SAM2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 70, + 528, + 148 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 528, + 148 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 528, + 148 + ], + "type": "table", + "html": "
Mask decoderPrompt encoderSeen (%)
JFJ&F
**34.844.839.8
✓*42.250.746.5
✓✓42.550.946.7
", + "image_path": "60eabab4e00472357ad724dd63969fc802da3a575aa2d2ea1ba4f1b856a93504.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "type": "text", + "content": "Table 6. Ablations on the different training modules in SAM2. \\* denotes the frozen module while " + }, + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "type": "inline_equation", + "content": "\\sqrt{}" + }, + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "type": "text", + "content": " makes the module trainable." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 203, + 479, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 479, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 479, + 214 + ], + "type": "text", + "content": "large amount of inter-frame information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "content": "Effect of Training Modules in SAM2. To preserve the strong VOS capabilities of SAM2, we freeze the memory attention modules but train the prompt encoder and mask decoder to further adapt the SAM2 to the multimodal clue reference, which can achieve the best results shown in Table 6. Without training the mask decoder, it will cause a huge performance drop due to the low capability to fit the multimodal representation of the [seg] token." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 391, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 391, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 391, + 335 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": "In this section, we discuss the limitations of our method. As shown in Table 1 NULL split, our method has a relatively low capability to reject the users' requirements for a null reference. It is the inherent limitation of the prompting paradigm for localization that SAM2 attends to segment something or some irrelevant regions in the image without any meaning. SAM2-LOVE lags behind the EEMC in the null test with 0.2 in " + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": " score, achieving 0.23 similar results with the AVSBench plus text method. The " + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 342, + 555, + 486 + ], + "type": "text", + "content": " denotes the square root of the ratio of the number of foreground pixels over the background pixels. This can be our future work on how to solve the null reference problem." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 500, + 388, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 388, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 388, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "This paper introduces SAM2-LOVE, a novel framework for segmenting reference objects in Language-aided Audio-Visual Scenes, overcoming the limitations of previous dual-modality SAM approaches in handling scenarios that integrate three modalities. We propose a multimodal fusion module that compresses information from all modalities into a single token, which then prompts SAM2 for video segmentation. This compressed [seg] token propagates forward through the video to capture both the spatial and temporal features, while also referring to the accumulated historical tokens backward preserving information from previous frames. Extensive experiments demonstrate the superiority of our model on the Ref-AVS benchmark and highlight the efficiency of our design. Future work will focus on addressing scenarios where the three modality references are misaligned, such as the null reference problem." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28939" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (Grant No. 62206153), and in part by Guangdong Natural Science Funds for Distinguished Young Scholar (No. 2025B1515020012)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 205 + ], + "type": "text", + "content": "[1] Sule Bai, Yong Liu, Yifei Han, Haoji Zhang, and Yansong Tang. Self-calibrated clip for training-free open-vocabulary segmentation. arXiv preprint arXiv:2411.15869, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 206, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 271 + ], + "type": "text", + "content": "[2] Yuanhong Chen, Yuyuan Liu, Hu Wang, Fengbei Liu, Chong Wang, Helen Frazer, and Gustavo Carneiro. Unraveling instance associations: A closer look for audio-visual segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26497–26507, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "type": "text", + "content": "[3] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 353 + ], + "type": "text", + "content": "[4] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12155–12163, 2024. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 354, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 420 + ], + "type": "text", + "content": "[5] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 776-780. IEEE, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "text", + "content": "[6] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 IEEE international conference on acoustics, speech and signal processing, pages 131-135. IEEE, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "type": "text", + "content": "[7] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing, pages 787-798, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "text", + "content": "[8] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "type": "text", + "content": "[9] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "text", + "content": "[10] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19108-19118, 2022. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[11] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22236-22245, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "text", + "content": "[12] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 216 + ], + "type": "text", + "content": "[13] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 554, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 261 + ], + "type": "text", + "content": "[14] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 554, + 294 + ], + "type": "text", + "content": "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 316 + ], + "type": "text", + "content": "[16] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 317, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 554, + 370 + ], + "type": "text", + "content": "[17] Yong Liu, Sule Bai, Guanbin Li, Yitong Wang, and Yansong Tang. Open-vocabulary segmentation with semantic-assisted calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3491-3500, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "type": "text", + "content": "[18] Yong Liu, Cairong Zhang, Yitong Wang, Jiahao Wang, Yujiu Yang, and Yansong Tang. Universal segmentation at arbitrary granularity with language instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3459-3469, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 459 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "text", + "content": "[20] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 554, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 559 + ], + "type": "text", + "content": "[21] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In IEEE European Conference on Computer Vision, pages 792-807. Springer, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 635 + ], + "type": "text", + "content": "[22] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dolar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "text", + "content": "[23] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[24] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28940" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[25] Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531-3539, 2021. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 193 + ], + "type": "text", + "content": "[26] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[27] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Autoacid: A large-scale dataset for audio-language representation learning. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5025-5034, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[28] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. arXiv preprint arXiv:2309.07929, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[29] Yuji Wang, Ruojun Zhao, and Zijun Sun. Efficient remote sensing transformer for coastline detection with sentinel-2 satellite imagery. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium, pages 5439-5442. IEEE, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[30] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting segmentation with sound is generalizable audio-visual source localizer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5669–5677, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[31] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-avs: Refer and segment objects in audio-visual scenes. IEEE European Conference on Computer Vision, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 503 + ], + "type": "text", + "content": "[32] Yuji Wang, Ruojun Zhao, Shicai Wei, Jingchen Ni, Meng Wu, Yang Luo, and Chunbo Luo. Convolution meets transformer: Efficient hybrid transformer for semantic segmentation with very high resolution imagery. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, pages 9688-9691. IEEE, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 505, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 294, + 548 + ], + "type": "text", + "content": "[33] Yuji Wang, Jingchen Ni, Yong Liu, Chun Yuan, and Yansong Tang. Iterprime: Zero-shot referring image segmentation with iterative grad-cam refinement and primary word emphasis. arXiv preprint arXiv:2503.00936, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "text", + "content": "[34] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. Onlinerefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "text", + "content": "[35] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 673 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[37] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3858-3869, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "type": "text", + "content": "[38] Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16111-16121, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 554, + 250 + ], + "type": "text", + "content": "[39] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audiovisual question answering on videos. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3480-3491, 2022. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 306 + ], + "type": "text", + "content": "[40] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 554, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 361 + ], + "type": "text", + "content": "[41] Zhao Yang, Jiaqi Wang, Xubing Ye, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Language-aware vision transformer for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 363, + 554, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 554, + 417 + ], + "type": "text", + "content": "[42] Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mobile applications. arXiv preprint arXiv:2306.14289, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 419, + 554, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 554, + 473 + ], + "type": "text", + "content": "[43] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Lei Liu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Evf-sam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 475, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 554, + 507 + ], + "type": "text", + "content": "[44] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 509, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 554, + 552 + ], + "type": "text", + "content": "[45] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 554, + 554, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 554, + 607 + ], + "type": "text", + "content": "[46] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audiovisual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 609, + 554, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 554, + 673 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 554, + 673 + ], + "type": "text", + "content": "[47] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28941" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_content_list.json b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0602044745d0081b25611b0057591b76a24ba34b --- /dev/null +++ b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_content_list.json @@ -0,0 +1,1473 @@ +[ + { + "type": "text", + "text": "SAM2Object: Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 194, + 128, + 810, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jihuai Zhao Junbao Zhuo* Jiansheng Chen Huimin Ma* University of Science and Technology Beijing", + "bbox": [ + 254, + 203, + 748, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jihuaizhao@xs.ustb.edu.cn, {junbaozhuo, jschen, mhmpub}@ustb.edu.cn", + "bbox": [ + 200, + 241, + 792, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of zero-shot 3D instance segmentation, existing 2D-to-3D lifting methods typically obtain 2D segmentation across multiple RGB frames using vision foundation models, which are then projected and merged into 3D space. However, since the inference of vision foundation models on a single frame is not integrated with adjacent frames, the masks of the same object may vary across different frames, leading to a lack of view consistency in the 2D segmentation. Furthermore, current lifting methods average the 2D segmentation from multiple views during the projection into 3D space, causing low-quality masks and high-quality masks to share the same weight. These factors can lead to fragmented 3D segmentation. In this paper, we present SAM2Object, a novel zero-shot 3D instance segmentation method that effectively utilizes the Segment Anything Model 2 to segment and track objects, consolidating view consistency across frames. Our approach combines these consistent 2D masks with 3D geometric priors, improving the robustness of 3D segmentation. Additionally, we introduce mask consolidation module to filter out low-quality masks across frames, which enables more precise 2D-to-3D matching. Comprehensive evaluations on ScanNetV2, ScanNet++ and ScanNet200 demonstrate the robustness and effectiveness of SAM2Object, showcasing its ability to outperform previous methods. Our project page is at https://jihuaizhaohd.github.io/SAM2Object.", + "bbox": [ + 89, + 323, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 744, + 223, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D instance segmentation plays a crucial role in various applications such as autonomous driving, robotics and augmented reality. Most of previous works [16, 20, 25, 41, 45] have predominantly relied on supervised learning, requiring high-quality 3D data and annotations for training. The annotation process for 3D data is often labor-intensive and costly, leading to the scarcity of high-quality 3D datasets.", + "bbox": [ + 89, + 771, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ceac3aed7ab363987ac077132dca74ea9ea4a52e1e02d6d610e023403fe1bc8c.jpg", + "image_caption": [ + "Figure 1. View Consistency. Compared with performing frame-by-frame segmentation on multi-view images, our method connects objects across multiple frames, ensuring that segmentation results remain consistent across various views." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 906, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This scarcity, in turn, greatly limits the development and advancement of 3D vision models, often leading to models with weak generalization capabilities, particularly in terms of zero-shot generalization. Zero-shot capability is essential, particularly in indoor navigation scenarios [1, 29]. For example, a robot misidentifying an unknown object, like a stove, could pose a fire hazard and endanger people.", + "bbox": [ + 511, + 508, + 906, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The relative ease of acquiring 2D annotated data has led to the creation of numerous high-quality datasets, accelerating the advancement of 2D vision foundation models [3, 4, 19, 35]. Segment Anything Model (SAM) [19] is a prime example of this progress. SAM is trained on the extensive and high-quality SA-1B dataset [19], exhibiting outstanding zero-shot performance across various types of 2D vision tasks.", + "bbox": [ + 511, + 614, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consequently, some methods [11, 14, 42, 46, 47] have attempted to transfer the robust generalization ability of SAM to 3D tasks by lifting its 2D segmentation into the 3D space. However, these methods often perform segmentation on a frame-by-frame manner, lacking inter-view consistency. Even slight changes in viewing angle can result in inconsistent 2D masks for the same object across different views. As shown in Fig. 1, the mask completely encapsulates the object in \"View 1\", but only partially covers it in \"View i\" and \"View n\". Therefore, conflicts can arise when merging 2D masks from different views, leading to", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19325", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fragmented 3D segmentation and decreased performance.", + "bbox": [ + 89, + 90, + 470, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides, existing work has not effectively addressed the errors introduced by long-distance projections in the 2D-to-3D matching process, as objects at different projection distances are assigned the same weight. Compared with the close-distance scene in Fig. 2 (a), the long-distance scene in Fig. 2 (b) erroneously projects the 3D points onto the \"wall\", causing a 3D object to incorrectly match with the 2D mask of another object. Therefore, the two originally separate objects are wrongly segmented as one. These errors increase as objects move farther from the camera, leading to incorrect 3D segmentation.", + "bbox": [ + 88, + 106, + 482, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the inter-view inconsistency, in this work, we introduce SAM2Object, a novel approach that harnesses the strong generalization capability of SAM2 [35] and its multiview consistency. SAM2Object consolidate the consistency of 2D masks across multiple views and integrate 3D geometric priors to achieve superior 3D instance segmentation.", + "bbox": [ + 89, + 272, + 483, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, to solve the problem of long-distance projection error, we propose a mask consolidation algorithm that applies weighting based on the distance of object from the camera across different views. During the 2D-to-3D matching process, the same object is prioritized to match with the view closer to the camera, thereby enhancing view consistency. The proposed SAM2Object conducts zero-shot 3D instance segmentation without the need for annotated data or additional training.", + "bbox": [ + 89, + 363, + 483, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 500, + 410, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We proposed SAM2Object, a novel zero-shot 3D instance segmentation method that leverages SAM2 to consolidate view consistency, effectively improving the robustness of 3D segmentation.", + "- We proposed a mask consistency metric to evaluate and maintain the 2D mask consistency across different views, enabling integration of 3D geometric priors and high-quality 2D masks for better 3D segmentation.", + "- We achieve more robust 3D instance segmentation on ScanNetV2 [5], ScanNet++ [48] and ScanNet200 [36]. Our method demonstrate superior performance compared with previous methods and outperform supervised models in zero-shot task." + ], + "bbox": [ + 89, + 513, + 482, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 723, + 232, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closed-vocabulary 3D segmentation. 3D segmentation methods have rapidly advanced due to the emergence of high-quality 3D datasets [5, 48]. The task of 3D semantic segmentation [9, 23, 26, 33] distinguishes various categories of objects within a scene. While 3D instance segmentation [6, 16, 21, 40] is required to further differentiate between various instances within each category. Point-Group [16] is a CNN-based method that segments objects using both the original and offset-shifted point sets, which groups nearby points with the same label and pro", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/afa3285739e91c2c417835b7cbbdd4b6d0d45fd9f5e5863bec96d3aa476a9771.jpg", + "image_caption": [ + "Figure 2. The quality of superpoint projection with different distances. In a scene, when an object is closer to the camera, the projection of 3D points onto the 2D image is more accurate in (a). However, if the object is farther from the camera in $(\\mathrm{b})\\sim (\\mathrm{d})$ , projection errors are more likely to occur (the blue dots represent the 3D points projected onto the 2D image, colored masks correspond to projection targets)." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gressively expands the group. Mask3D [38] introduced the first transformer-based 3D segmentation method, which directly predicts instance masks while avoiding manually selected voting schemes, achieving state-of-the-art performance. However, Mask3D still requires a substantial volume of 3D annotated data for training, which suffer poor performance on unseen data outside their domain-specific datasets. In this work, We directly leverage the zero-shot capability of SAM2 for 3D segmentation, eliminating the need for model training or 3D annotation.", + "bbox": [ + 511, + 356, + 906, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary 3D segmentation. Recently, vision foundation models have achieved excellent performance in 2D understanding tasks [3, 4, 19, 35], leading to the emergence of notable work in open-vocabulary 2D segmentation tasks [12, 15, 43, 44]. Inspired by these efforts, some works has focused on exploring open-vocabulary capabilities in 3D tasks. OpenMask3D [42] and OpenIns3D [14] leverage Mask3D [38] to extract 3D instances in 3D space and then directly project them onto different 2D views for extracting open-vocabulary features. These methods suffer poor performance in recognizing small or detailed objects. LERF [18] integrates language embeddings into NeRF [31], by performing voxel rendering of CLIP [34] embeddings along training rays and supervising these embeddings across training views. LERF focuses on roughly locating specific objects through text prompts rather than achieving fine-grained 3D segmentation and has shown to be inadequate in complex scenes. In this study, we focus on exploring how to achieve better fine-grained 3D segmentation solely leveraging SAM2, without relying on any 3D pre-trained models.", + "bbox": [ + 511, + 508, + 908, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Segment Anything Model. The Segment Anything Model (SAM) [19] has made significant breakthroughs in 2D segmentation tasks, demonstrating a strong generalization ability to segment any object in any image. The upgraded version of SAM, SAM2 [35], has further expanded", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19326", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into the domain of video segmentation. The image predictor of SAM2 remains consistent with SAM, while the video predictor is used to segment objects within videos. It is equipped with a memory that stores information about the object and previous interactions, enabling it to generate and correct predictions throughout the video. Recently, several 2D-to-3D lifting methods leverage SAM to perform frame-by-frame segmentation of multi-view images and project them onto 3D space. SAM3D [47] leads the way in extending SAM onto 3D segmentation by projecting segmentation data from 2D images onto 3D space. SAI3D [49] and SAM-graph [11] similarly adopts the approach of segmenting each frame independently, projecting the 2D segmentations onto superpoints and then merges superpoints using either region growing or Graph Neural Networks (GNNs). However, these methods do not link adjacent frames, potentially causing inconsistencies in 2D segmentation across consecutive frames, which can result in fragmented 3D segmentation. In this study, we effectively leverage SAM2 to establish connections across multi-view images, ensuring consistency in 2D segmentation between views, thereby achieving more robust 3D instance segmentation.", + "bbox": [ + 89, + 90, + 483, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 438, + 191, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation and method overview", + "text_level": 1, + "bbox": [ + 89, + 463, + 457, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a series of posed color images $\\{I_1, I_2, \\dots, I_T\\}$ and the point cloud P, our goal is to output the 3D instance segmentation of the scene. In this paper, we propose a novel zero-shot 3D instance segmentation framework, as illustrated in Fig. 3. We first use SAM2 [35] to segment and track objects, ensuring that the masks accurately cover objects even as the viewpoint changes (Sec. 3.2), thereby consolidating multi-view consistency. Then, we combine 2D mask proposals with over-segmented 3D superpoints to construct a graph which is described in Sec. 3.3. In Sec. 3.4, we introduce an iterative graph clustering technique that hierarchically clusters different superpoints based on their view consistency levels, ultimately achieving 3D segmentation.", + "bbox": [ + 89, + 484, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 2D mask proposal", + "text_level": 1, + "bbox": [ + 89, + 722, + 266, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Keyframes extraction", + "text_level": 1, + "bbox": [ + 89, + 744, + 287, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define keyframe as the frame where new objects appear. An illustration of the keyframe can be found in Fig. 4. In order to segment and track new objects, it is essential to extract the keyframes and perform segmentation and tracking on them. Given a set of color images, we can treat it as a video. When there is a scene change or a new object appears, the pixel differences between the previous frame $I_{pre}$ and the current frame $I_{cur}$ tend to increase significantly. We use these frame differences to identify keyframes. The", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frame differences $I_{dif}$ between two adjacent frames can be defined as", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {d i f} = \\frac {\\left| I _ {p r e} - I _ {c u r} \\right|}{I _ {h} \\cdot I _ {w}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 117, + 906, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$I_h$ and $I_w$ represent the height and width of the color images, respectively. After extracting the differences between each pair of adjacent frames across the entire sequence of images, we apply a Hanning window [10] to smooth the frame difference array in order to mitigate the interference of noise. The window weight can be calculated as", + "bbox": [ + 511, + 154, + 906, + 247 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw _ {n} = \\left\\{ \\begin{array}{c c} 0. 5 - 0. 5 \\cos \\left(\\frac {2 \\pi n}{N - 1}\\right), & 0 \\leq n \\leq N - 1, \\\\ 0, & o t h e r w i s e. \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 256, + 906, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$N$ is the window length, $n$ represents the index within the window, and $w_{n}$ denotes the value of the window function at position $n$ . We define the smoothed array as $S$ , where the $j$ -th position in array $S$ can be defined as", + "bbox": [ + 511, + 301, + 905, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {j} = \\frac {\\sum_ {n = 0} ^ {N - 1} w _ {n} \\cdot I _ {d i f} [ j - n ]}{\\sum_ {n = 0} ^ {N - 1} w _ {n}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 372, + 906, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $I_{dif}[j - n]$ denotes the value at the $j - n$ position in $I_{dif}$ . The indices corresponding to the local maxima in the smoothed array $S$ represent the keyframes. We control the average distance between all keyframes to be around 100 frames by setting the window size appropriately.", + "bbox": [ + 511, + 422, + 906, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2. Segmentation and tracking", + "text_level": 1, + "bbox": [ + 511, + 507, + 750, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ SAM2 [35] for object segmentation and tracking to consolidate view consistency. SAM2 utilizes a sophisticated memory mechanism to track objects across frames, ensuring continuity even when objects are temporarily occluded or exit and re-enter the scene. Although SAM2's auto-segmentation technique does not require any points, boxes, or mask prompts, these prompts are necessary for object tracking. As shown in Fig. 3 (b), given a series of images and keyframes, we treat the image sequences between every two keyframes as video clips, with the first and last frames of the clip being the keyframe. We use the image predictor of SAM2 to automatically segment the keyframes and generate 2D masks for objects. These masks then serve as prompts for the video predictor of SAM2 to track the objects within the video clips. We segment new objects from the keyframes and track them in the video clips between key frames. Finally, at the edge of adjacent video clips, we link the same objects across different clips according to IoU metric. In the end, we obtain the segmentation and tracking results for all objects throughout the entire video.", + "bbox": [ + 511, + 525, + 906, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.3. Merging segmentation", + "text_level": 1, + "bbox": [ + 511, + 851, + 715, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Even though we can segment keyframes to obtain new objects, keyframes are not always the exact frames where new", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19327", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/671d5dceac898b074dbfb3dde64f8439eb1d3e887cfa3848b6cb94eeaa0b2e68.jpg", + "image_caption": [ + "Figure 3. Method overview. Our method consists of six steps. We over-segment the 3D mesh / point cloud into superpoints. (a) We extract keyframes from posed images. (b) We leverage SAM2 to segment the keyframes, and its segmentation serve as mask prompts for SAM2 to track objects throughout the entire set of posed images. (c) We consolidate the consistency of the masks obtained by SAM2 to filter out low-quality mask proposals. (d) We combine the masks with 3D geometric priors to construct a graph, where the nodes are superpoints and the edges represent the affinity scores between masks of superpoints that have undergone consistency consolidation. (e) We perform iterative graph clustering on the graph to achieve 3D instance segmentation." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 338, + 428 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4987034fe8843e3b0b0d2e331f325f27b8c7952ed73d46e1448e7dd3e2e695e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 88, + 895, + 429 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8c0d699823022c649db5ac5de128ce124e312cb8e1a318ec1239bf42daceb6c.jpg", + "image_caption": [ + "Figure 4. Segmentation and tracking. (a) We extract keyframes from posed images. (b) We use the keyframes segmentation as mask prompts to guide SAM2 in tracking objects in both directions. (c) We merge the results to ensure that no objects are missed." + ], + "image_footnote": [], + "bbox": [ + 94, + 540, + 480, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "objects appear. If the new objects appear before the key frames, the new objects may be missed or overlooked. To address this issue, we adopt a bidirectional tracking strategy", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(forward and backward). As shown in Fig. 4 (b), objects missed during forward tracking are identified in backward tracking (such as the \"trash can\" in the second column of images). Similarly, objects missed in backward tracking are captured in forward tracking (such as the \"door\" in the fourth column of images). The tracking results from the different directions complement each other. So we merge the 2D masks generated from both directions. An illustration of merging segmentation can be found in Fig. 4 (c). The segmentation we obtain maintain consistency across multiple views, which is crucial for subsequently combining 3D geometric priors to achieve more robust 3D instance segmentation.", + "bbox": [ + 511, + 540, + 906, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Graph construction", + "text_level": 1, + "bbox": [ + 511, + 750, + 702, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Building the superpoint graph", + "text_level": 1, + "bbox": [ + 511, + 773, + 769, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a point cloud $P \\in \\mathbb{R}^{N \\times 3}$ , we represent the scene or point cloud as a graph $G = (V, E)$ , where $V$ denotes the set of superpoints $\\{S_i\\}_{i=1}^{N_s}$ in the point cloud and $E$ denotes the relationships between these superpoints. Following recent works [47, 49], we employ a graph-cut algorithm [8] to over-segment the whole point cloud based on 3D geometric attributes, resulting in a series of superpoints. Each", + "bbox": [ + 511, + 792, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19328", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "node in the graph represents the multiple points in the point cloud that share similar geometric features, and the number of superpoints is substantially lower than the number of points in the point cloud. By performing calculations at the superpoint-level, we achieve efficient computation while also extracting geometric priors.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (d), all color images from the different views are fed into SAM2 to obtain 2D masks, and we then project superpoints onto these 2D masks in all views to match superpoints with these masks. We use a common pinhole camera matrix to perform the 3D-to-2D projection, utilizing the corresponding camera intrinsic parameter $R_{k}$ and pose parameter $E_{k}$ to project the $i$ -th 3D superpoint $S_{i}$ onto the $k$ -th image, and get its corresponding 2D mask proposals:", + "bbox": [ + 89, + 181, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i, k} ^ {2 D} = \\phi \\left(R _ {k}, E _ {k}, S _ {i}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 316, + 482, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\phi(\\cdot)$ is the projection operator. The relationships between superpoints are represented by the affinity scores between their corresponding 2D masks, which are used to construct the edges of the graph. Specifically, we compute a normalized histogram based on the 2D mask proposals $S_{i,k}^{2D}$ and convert it into a vector, denoted as $\\mathbf{e}_i^k = \\text{vec(normalized}(S_{i,k}^{2D}))$ . This histogram vector represents the 2D class-agnostic instance features of the 3D superpoint, and we get the affinity score between different superpoints on the $k$ -th image by computing the cosine similarity of their 2D class-agnostic instance features:", + "bbox": [ + 89, + 340, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i, j} ^ {k} = \\frac {\\mathbf {e} _ {i} ^ {k} \\cdot \\mathbf {e} _ {j} ^ {k}}{\\left| \\mathbf {e} _ {i} ^ {k} \\right| \\left| \\mathbf {e} _ {j} ^ {k} \\right|}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 520, + 482, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Mask consolidation", + "text_level": 1, + "bbox": [ + 89, + 564, + 272, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traversing all the images will yield the overall affinity score between $S_{i}$ and $S_{j}$ , but the distribution of 2D mask proposals covered by the corresponding superpoint may vary due to changes in the shape of the same object from different views, as well as factors such as occlusion, motion blur, and projection errors. To address this issue, we introduce a mask consistency weight to filter out low-quality mask proposals, thereby clarifying the relationships between superpoints in the graph.", + "bbox": [ + 89, + 582, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We calculate the visibility of 3D points on 2D images to address object occlusion and visibility issues. To obtain the visible points, we project the points from the point cloud onto the image, filter out the points that fall outside the image, and determine occlusion between objects based on their depth values. We define $V_{i}^{k}\\in (0,1)$ as the ratio of the visible point count of $S_{i}$ on the $k$ -th image to its total point count in the point cloud. Thus, we denote the visibility edge weight between the $i$ -th superpoint and the $j$ -th superpoint as", + "bbox": [ + 89, + 719, + 483, + 867 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {i, j} ^ {k} = V _ {i} ^ {k} \\cdot V _ {j} ^ {k}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The projection accuracy of the same superpoint in the point cloud can differ across various views. Projections in views closer to the camera tend to be more accurate than those in views farther from the camera. These errors can lead to mismatches between the 2D mask proposals $S_{i,k}^{2D}$ and the superpoint $S_{i}$ (as shown in Fig. 2, the superpoint that should have been matched to the \"table\" ended up being matched to the \"wall\"), which in turn leads to errors in the segmentation results. In addition, objects that are farther away from the camera appear smaller in the image, which also makes it easy to generate inaccurate 2D masks. We define $D_{i}^{k} \\in (0,1)$ as the distance weight to perform weighted averaging of the 2D mask proposals $S_{i,k}^{2D}$ based on the distance from the $k$ -th camera:", + "bbox": [ + 511, + 90, + 906, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} ^ {k} = 0. 9 \\cdot \\frac {d _ {m a x} ^ {k} - d _ {i} ^ {k}}{d _ {m a x} ^ {k} - d _ {m i n} ^ {k}} + 0. 1, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 306, + 906, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_i^k$ represents the distance of visible points from the $k$ -th camera within $S_{i,k}^{2D}$ , $d_{max}^k$ and $d_{min}^k$ denotes the maximum and minimum distances of all visible points from the $k$ -th camera. So, we denote the distance edge weight between the $i$ -th superpoint and the $j$ -th superpoint as $\\psi_{i,j}^k$ , which can be calculated as", + "bbox": [ + 511, + 349, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {i, j} ^ {k} = D _ {i} ^ {k} \\cdot D _ {j} ^ {k}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 454, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the number of 2D mask proposals $S_{i,k}^{2D}$ covered by the corresponding superpoint is defined as its purity. If a superpoint covers more 2D mask proposals, it is considered a difficult sample, and we reduce its weight and merge it only during the later stages of graph clustering in Sec. 3.4. We define $L_i^k \\in (0,1)$ as the maximum ratio of all points within a superpoint projected onto different 2D mask proposals of $k$ -th image:", + "bbox": [ + 511, + 477, + 905, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i} ^ {k} = \\max \\left(\\frac {N _ {i , 1} ^ {k}}{N _ {i} ^ {k}}, \\frac {N _ {i , 2} ^ {k}}{N _ {i} ^ {k}}, \\dots , \\frac {N _ {i , n} ^ {k}}{N _ {i} ^ {k}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 604, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$N_{i}^{k}$ represents the total number of visible points of the $i$ -th superpoint in the $k$ -th image, and $N_{i,j}^{k}$ denotes the number of points belonging to the mask label with a value of $j$ , $n$ is the total number of 2D mask proposals. So, we denote the purity edge weight between the $i$ -th superpoint and the $j$ -th superpoint as $\\Phi_{i,j}^{k}$ , which can be calculated as", + "bbox": [ + 511, + 648, + 906, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {i, j} ^ {k} = L _ {i} ^ {k} \\cdot L _ {j} ^ {k}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 753, + 903, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we combine these three weights to form the mask consistency weight as", + "bbox": [ + 511, + 776, + 905, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i, j} ^ {k} = \\varphi_ {i, j} ^ {k} \\cdot \\Phi_ {i, j} ^ {k} \\cdot \\psi_ {i, j} ^ {k}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 813, + 903, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, the final affinity score is calculated as", + "bbox": [ + 511, + 840, + 803, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i, j} = \\frac {\\sum_ {k = 1} ^ {M} \\left(W _ {i , j} ^ {k} A _ {i , j} ^ {k}\\right)}{\\sum_ {k = 1} ^ {M} W _ {i , j} ^ {k}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 863, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19329", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Iterative graph clustering", + "text_level": 1, + "bbox": [ + 89, + 90, + 323, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building upon the superpoint graph, we introduce an iterative graph clustering technique to merge superpoints and update the graph. In each iteration, we employ a region growing method to merge superpoints, considering not only the affinity scores but also weighting based on the distance between superpoints. From a geometric perspective, superpoints that are far apart are less likely to be similar. However, once projected, they may appear closer on images from specific viewpoints. According to Sec. 3.3.2, their 2D mask proposals tend to be of lower quality, resulting in incorrect affinity scores. So we set the edge weight decreases for superpoints that are farther apart.", + "bbox": [ + 89, + 112, + 480, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the introduction of the Sec. 3.3.2, some superpoints covering low-quality 2D mask proposals have lower view consistency and affinity scores. These superpoints may be located at the junction of two objects. Therefore, in each iteration, we set different thresholds for merging superpoints based on the affinity scores. We initially merge small and adjacent regions with high threshold, gradually lowering the threshold over iterations to merge superpoints with lower confidence. This iterative graph clustering approach allows for hierarchical merging of superpoints with varying levels of view consistency. Superpoints with stronger view consistency are prioritized for merging. Subsequently, superpoints with lower view consistency are merged, ultimately achieving robust 3D segmentation.", + "bbox": [ + 89, + 295, + 482, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 522, + 222, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our SAM2Object on multiple datasets to demonstrate its effectiveness and robustness in zero-shot 3D instance segmentation. We also compare our SAM2Object with leading methods in the open-vocabulary segmentation field and the state-of-the-art closed-vocabulary segmentation methods which require training on annotated datasets.", + "bbox": [ + 89, + 547, + 482, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment settings", + "text_level": 1, + "bbox": [ + 89, + 651, + 279, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. ScanNetV2 [5] is a large-scale dataset tailored for indoor scene understanding and segmentation. It comprises RGB and depth images, as well as 3D point cloud data derived from indoor scenes. It includes over 2.5 million extensive views from more than 1,500 scans, along with semantic labels for 21 categories. ScanNet200 [36] is an extension of ScanNetV2, increasing the number of semantic categories to 200. Furthermore, ScanNet++ [48] is a newly released indoor dataset contains 280 indoor scenes and couples geometric and color data of high-quality, commodity-level indoor scenes. Each scene is captured by laser scanners with higher-resolution and a DSLR camera with a fisheye lens, as well as RGB-D streams obtained from iPhone 13 Pro. We evaluate our method on the well-studied ScanNetV2, ScanNet++ and ScanNet200 datasets. Following SAI3D [49],", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When generating 2D mask proposals, we use only $20\\%$ of the images from the ScanNetV2 dataset and only $5\\%$ from the ScanNet++ dataset.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1b0022115836474c132a6d27974e9f551b0d4954b0b4bf01c6c322a607e73145.jpg", + "table_caption": [ + "Table 1. Class-agnostic 3D instance segmentation on ScanNetV2 dataset. Bold and underlined values indicate the best and second best scores, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTraining SetmAP\\( AP_{50} \\)\\( AP_{25} \\)
Closed-vocabulary
Mask3D [38]S3DIS31.144.958.0
Open-vocabulary
HDBSCAN [30]-1.65.532.1
UCIS [32]-2.37.330.5
EGIS [7]-5.012.738.9
Part2Object [39]-12.626.855.1
SAM-graph [11]-15.133.359.1
UnScene3D [37]-15.932.258.5
SAMPro3D [46]-18.032.856.3
SAM3D [47]-20.234.053.3
SAI3D [49]-30.850.570.6
SAM2Object (Ours)-34.052.770.3
", + "bbox": [ + 514, + 220, + 903, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate the quantitative results with the widely adopted Average Precision (AP) score. Following [11, 38, 42, 49], we calculate AP with IoU thresholds of $50\\%$ and $25\\%$ $(\\mathrm{AP}_{50}, \\mathrm{AP}_{25})$ and AP averaged with thresholds from $50\\%$ to $95\\%$ with a step size of $5\\%$ (mAP). We implemented two distinct evaluation settings to comprehensively validate the performance of our method. Firstly, we evaluate class-agnostic instance segmentation which disregards category information and solely evaluates the accuracy of instance masks. Secondly, semantic instance segmentation builds on this by considering the semantic label of each instance, emphasizing the relationship between an instance and its category.", + "bbox": [ + 511, + 492, + 906, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with both open-vocabulary and closed-vocabulary baselines. For closed-vocabulary methods, we compare with the state-of-the-art method, Mask3D [38], which is a supervised model requiring training data. For open-vocabulary methods, we compare with the recent lifting methods based on leveraging SAM, including SAM3D [47], SAMPro3D [46], Segment3D [13], SAM-graph [11], SAI3D [49] and OVIR-3D [27]. We also compared with OpenMask3D [42], which generates masks using Mask3D and assigns semantic labels. Additionally, three traditional methods, referred to as HDBSCAN [30], EGIS [7] and UCIS [32], as well as the unsupervised methods UnScene3D [37] and Part2Object [39] are also included in our comparison.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19330", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/344987d868fc372fd83f4589512ccce7cd7cf2e0d7906e39aee1fc5da4314757.jpg", + "table_caption": [ + "Table 2. Class-agnostic 3D instance segmentation on ScanNet++ dataset. Bold and underlined values indicate the best and second best scores, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTraining SetmAP\\( AP_{50} \\)\\( AP_{25} \\)
Closed-vocabulary
Mask3DScanNetV29.917.325.8
Open-vocabulary
EGIS [7]-4.19.225.3
SAM3D [47]-7.214.229.4
Segment3D [13]-12.022.737.8
SAM-graph [11]-12.925.343.6
SAI3D [49]-17.131.149.5
SAM2Object (Ours)-20.234.148.7
", + "bbox": [ + 91, + 142, + 480, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 332, + 184, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. We conduct the experiments on ScanNetV2, ScanNet++ and ScanNet200. Table 1 displays the quantitative results of our method on class-agnostic instance segmentation within ScanNetV2. Our method outperforms previous works in all metrics and leads in mAP and $\\mathrm{AP}_{50}$ compared with the state-of-the-art methods SAI3D, trailing by only 0.3 in the $\\mathrm{AP}_{25}$ metric. Notably, SAI3D requires adaptation of different 2D segmentation models for different datasets (using Semantic-SAM [22] on ScanNetV2 and HQ-SAM [17] on ScanNet++), whereas we consistently use SAM2 for validation across different datasets. Furthermore, our method outperforms both unsupervised and closed-vocabulary methods. The latter requires training data, so to conduct a fair zero-shot evaluation, we replace its training data with S3DIS [2] and validate on ScanNetV2.", + "bbox": [ + 88, + 354, + 482, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 reports the class-agnostic instance segmentation results on the challenging ScanNet++, we still lead in the mAP and $\\mathrm{AP}_{50}$ metric. In lifting methods, the performance of the 2D segmentation model is crucial to the final 3D instance segmentation, SAI3D uses the optimal segmentation model for each dataset, while we consistently use SAM2. We simply replace the optimal segmentation model in SAI3D with SAM2, resulting in a mAP of $28.23\\%$ on ScanNetV2 and $16.02\\%$ on ScanNet++, both lower than SAI3D. This demonstrates the effectiveness of our method.", + "bbox": [ + 88, + 582, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following OpenMask3D [42] and SAI3D [49], we validate the semantic instance segmentation results on ScanNet200, as shown in Table 3. We assign semantic labels to our segmented 3D instances using the method from OpenMask3D. We exhibit outstanding performance in both mAP and $\\mathrm{AP}_{50}$ , indicating that our method demonstrates greater generalization when handling different semantic labels.", + "bbox": [ + 88, + 733, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method has lower $\\mathrm{AP}_{25}$ than SAI3D across the three datasets. SAI3D lacks view consistency, resulting in fragmented and over-segmented outputs, which partially contributes to its higher $\\mathrm{AP}_{25}$ . Additionally, SAI3D selects", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the most suitable visual foundation model for each dataset, achieving fine-grained compared with SAM2, which explains their higher $\\mathrm{AP}_{25}$ . In contrast, we prefer to design a method capable of generalizing across multiple datasets.", + "bbox": [ + 511, + 90, + 905, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a6536d3eeb886557943040dcd1f8449086af0e1462625b27c590da2988d805d.jpg", + "table_caption": [ + "Table 3. Semantic instance segmentation on ScanNet200 dataset. \"OV\" and \"Sup\" represent \"Open-vocabulary\" and \"Supervised\" respectively. Bold and underlined values indicate the best and second best scores, respectively." + ], + "table_footnote": [], + "table_body": "
MethodmAP\\( AP_{50} \\)\\( AP_{25} \\)
Supmasked + OVsemantic OpenMask3D [42]15.419.923.1
OVmasked + OVsemantic
OVIR-3D [27]9.318.725.0
SAM3D [47]9.815.220.7
SAI3D [49]12.718.824.1
SAM2Object (Ours)13.319.023.8
", + "bbox": [ + 522, + 232, + 890, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. We qualitatively compared our method with the SOTA method SAI3D on ScanNetV2 and ScanNet++. SAI3D is unable to handle the problem of projection errors, and the mismatched projections in Fig. 3 (c) and Fig. 2 result in incorrect 3D segmentation in Fig. 5 (a) and Fig. 5 (b). Moreover, SAI3D employs a frame-by-frame segmentation approach across multiple views, which lacks view consistency. This inconsistency in Fig. 1 leads to fragmented 3D segmentation in Fig. 5 (a), and similar fragmentation can be observed in Fig. 5 (c). Furthermore, the under-segmentation is also evident in Fig. 5 (d).", + "bbox": [ + 511, + 385, + 905, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, our approach consolidates view consistency and filters out low-quality 2D masks, enabling us to achieve comprehensive and robust 3D segmentation. More results can be found in the supplementary material.", + "bbox": [ + 511, + 551, + 905, + 612 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aca2723b771b1c45b7955d73222bc74b5c3e24fed10f25cc207207948d040eec.jpg", + "table_caption": [ + "Table 4. Ablations of main components in our method." + ], + "table_footnote": [], + "table_body": "
Seg. & TrackMask consolidationmAPAP50AP25
28.244.765.2
✓29.347.666.0
✓32.149.968.3
✓✓34.052.770.3
", + "bbox": [ + 516, + 651, + 898, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation and analysis", + "text_level": 1, + "bbox": [ + 511, + 757, + 718, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 4, we performed ablation experiments on ScanNetV2 to verify the effectiveness of each component of our method. Starting from the baseline, similar to existing 2D lifting methods, we use the image predictor of SAM2 to perform frame-by-frame segmentation on multi-view images, setting the weight of all visible views equally. This approach attempts to achieve view consistency by averaging results across all visible views. However, since there is", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19331", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d14d51dfb7d1993d178680d78d70fdf464523301c83c2ce610d907d2a63b29c.jpg", + "image_caption": [ + "Figure 5. Qualitative results comparisons between our method with SAI3D [49] on ScanNetV2 and ScanNet++. Due to projection errors resulting in incorrect 2D-to-3D matching relationships, SAI3D mistakenly segments part of the chair as the wall in (a), and segments the table and wall as a single object in (b). Due to the lack of view consistency, SAI3D exhibits fragmented segmentation in (c), as well as under-segmentation in (d). Our method shows the most robust and accurate class-agnostic instance segmentation results." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 890, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "no consistency between the masks of consecutive frames, it leads to fragmented segmentation. By applying mask consistency weighting through the mask consolidation module to results from multiple views, this strategy reduces low-quality mask interference during graph clustering (Sec. 3.4), resulting in a slight performance improvement. Employing the video predictor of SAM2 for object tracking ensures mask consistency across multi-view images. This results in more precise matching with 3D superpoints in Sec. 3.3, leading to improved performance. Finally, combining these two strategies fully leverages the benefits of consolidated view consistency, achieving further improved performance.", + "bbox": [ + 88, + 546, + 486, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 89, + 741, + 218, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method enhances view consistency by tracking objects across multiple views using SAM2 but requires continuous and dense multi-view images. Furthermore, although vision foundation models exhibit strong generalization, they are only capable of performing inference on 2D data and cannot fully understand scenes from a 3D scene. Although our approach connects adjacent frames, it is less effective than methods that integrate 3D priors with 2D semantics to establish view consistency. Future work should leverage 3D", + "bbox": [ + 88, + 763, + 486, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "geometric priors to enhance visual foundation models in understanding 3D scenes from multi-view images, improving their generalization for 3D tasks.", + "bbox": [ + 511, + 546, + 908, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 607, + 633, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel zero-shot 3D instance segmentation method, SAM2Object. The key idea is to achieve more accurate 2D-to-3D matching by consolidating view consistency. We employ SAM2 to generate consistent masks across multiple views and introduce a metric to measure mask consistency. After constructing the superpoint graph, we perform iterative clustering based on view consistency levels, achieving more accurate segmentation. Comprehensive evaluations across multiple datasets demonstrate the effectiveness of SAM2Object, highlighting its ability to outperform previous methods and rival fully supervised techniques in zero-shot instance segmentation tasks.", + "bbox": [ + 511, + 632, + 908, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 829, + 692, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China: 62402033, 62227801 and 62376024.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19332", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 1", + "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 7", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 2", + "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 3", + "[6] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2", + "[7] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 6, 7", + "[8] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4", + "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2", + "[10] Thomas Grandke. Interpolation algorithms for discrete fourier transforms of weighted signals. IEEE transactions on instrumentation and measurement, 32(2):350-355, 1983. 3", + "[11] Haoyu Guo, He Zhu, Sida Peng, Yuang Wang, Yujun Shen, Ruizhen Hu, and Xiaowei Zhou. Sam-guided graph cut for 3d instance segmentation. In ECCV, 2024. 1, 3, 6, 7", + "[12] Shuting He, Henghui Ding, and Wei Jiang. Semantic-promoted debiasing and background disambiguation for zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19498-19507, 2023. 2", + "[13] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation without manual labels. In European Conference on Computer Vision, pages 278-295. Springer, 2025. 6, 7", + "[14] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616, 2023. 1, 2", + "[15] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2", + "[16] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2", + "[17] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2", + "[20] Maksim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Top-down beats bottom-up in 3d instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3566-3574, 2024. 1", + "[21] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 2", + "[22] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 7", + "[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2", + "[24] Wenteng Liang, Feng Xue, Yihao Liu, Guofeng Zhong, and Anlong Ming. Unknown sniffer for object detection: Don't turn a blind eye to unknown objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3230-3239, 2023. 1", + "[25] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19333", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "international conference on computer vision, pages 2783-2792, 2021. 1", + "[26] Haizhuang Liu, Junbao Zhuo, Chen Liang, Jiansheng Chen, and Huimin Ma. Affinity3d: Propagating instance-level semantic affinity for zero-shot point cloud semantic segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9019-9028, 2024. 2", + "[27] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data. In Conference on Robot Learning, pages 1610–1620. PMLR, 2023. 6, 7", + "[28] Weijie Lyu, Xueting Li, Abhijit Kundu, Yi-Hsuan Tsai, and Ming-Hsuan Yang. Gaga: Group any gaussians via 3d-aware memory bank. arXiv preprint arXiv:2404.07977, 2024. 2", + "[29] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. In Advances in Neural Information Processing Systems, pages 32340-32352. Curran Associates, Inc., 2022. 1", + "[30] Leland McInnes and John Healy. Accelerated hierarchical density based clustering. In 2017 IEEE international conference on data mining workshops (ICDMW), pages 33-42. IEEE, 2017. 6", + "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[32] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters, 7(4):8713-8720, 2022. 6", + "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 1", + "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3", + "[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 6, 1, 5", + "[37] David Rozenberszki, Or Litany, and Angela Dai. Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19957-19967, 2024. 6" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6", + "[39] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6", + "[40] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 2", + "[41] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1", + "[42] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 6, 7", + "[43] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2", + "[44] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635–3647, 2024. 2", + "[45] Yizhou Wang, Longguang Wang, Qingyong Hu, Yan Liu, Ye Zhang, and Yulan Guo. Panoptic segmentation of 3d point clouds with gaussian mixture model in outdoor scenes. Visual Intelligence, 2(1):10, 2024. 1", + "[46] Mutian Xu, Xingyilang Yin, Lingteng Qiu, Yang Liu, Xin Tong, and Xiaoguang Han. Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation. arXiv preprint arXiv:2311.17707, 2023. 1, 6", + "[47] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 1, 3, 4, 6, 7, 2", + "[48] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 6, 4", + "[49] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3292–3302, 2024. 3, 4, 6, 7, 8, 2" + ], + "bbox": [ + 516, + 92, + 906, + 867 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19334", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_model.json b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72835ed63f4a73969e4c601b62e0eef122709af1 --- /dev/null +++ b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.13, + 0.812, + 0.177 + ], + "angle": 0, + "content": "SAM2Object: Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.204, + 0.749, + 0.24 + ], + "angle": 0, + "content": "Jihuai Zhao Junbao Zhuo* Jiansheng Chen Huimin Ma* University of Science and Technology Beijing" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.242, + 0.794, + 0.258 + ], + "angle": 0, + "content": "jihuaizhao@xs.ustb.edu.cn, {junbaozhuo, jschen, mhmpub}@ustb.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.718 + ], + "angle": 0, + "content": "In the field of zero-shot 3D instance segmentation, existing 2D-to-3D lifting methods typically obtain 2D segmentation across multiple RGB frames using vision foundation models, which are then projected and merged into 3D space. However, since the inference of vision foundation models on a single frame is not integrated with adjacent frames, the masks of the same object may vary across different frames, leading to a lack of view consistency in the 2D segmentation. Furthermore, current lifting methods average the 2D segmentation from multiple views during the projection into 3D space, causing low-quality masks and high-quality masks to share the same weight. These factors can lead to fragmented 3D segmentation. In this paper, we present SAM2Object, a novel zero-shot 3D instance segmentation method that effectively utilizes the Segment Anything Model 2 to segment and track objects, consolidating view consistency across frames. Our approach combines these consistent 2D masks with 3D geometric priors, improving the robustness of 3D segmentation. Additionally, we introduce mask consolidation module to filter out low-quality masks across frames, which enables more precise 2D-to-3D matching. Comprehensive evaluations on ScanNetV2, ScanNet++ and ScanNet200 demonstrate the robustness and effectiveness of SAM2Object, showcasing its ability to outperform previous methods. Our project page is at https://jihuaizhaohd.github.io/SAM2Object." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.746, + 0.224, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.485, + 0.879 + ], + "angle": 0, + "content": "3D instance segmentation plays a crucial role in various applications such as autonomous driving, robotics and augmented reality. Most of previous works [16, 20, 25, 41, 45] have predominantly relied on supervised learning, requiring high-quality 3D data and annotations for training. The annotation process for 3D data is often labor-intensive and costly, leading to the scarcity of high-quality 3D datasets." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.426, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 1. View Consistency. Compared with performing frame-by-frame segmentation on multi-view images, our method connects objects across multiple frames, ensuring that segmentation results remain consistent across various views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.614 + ], + "angle": 0, + "content": "This scarcity, in turn, greatly limits the development and advancement of 3D vision models, often leading to models with weak generalization capabilities, particularly in terms of zero-shot generalization. Zero-shot capability is essential, particularly in indoor navigation scenarios [1, 29]. For example, a robot misidentifying an unknown object, like a stove, could pose a fire hazard and endanger people." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.909, + 0.734 + ], + "angle": 0, + "content": "The relative ease of acquiring 2D annotated data has led to the creation of numerous high-quality datasets, accelerating the advancement of 2D vision foundation models [3, 4, 19, 35]. Segment Anything Model (SAM) [19] is a prime example of this progress. SAM is trained on the extensive and high-quality SA-1B dataset [19], exhibiting outstanding zero-shot performance across various types of 2D vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Consequently, some methods [11, 14, 42, 46, 47] have attempted to transfer the robust generalization ability of SAM to 3D tasks by lifting its 2D segmentation into the 3D space. However, these methods often perform segmentation on a frame-by-frame manner, lacking inter-view consistency. Even slight changes in viewing angle can result in inconsistent 2D masks for the same object across different views. As shown in Fig. 1, the mask completely encapsulates the object in \"View 1\", but only partially covers it in \"View i\" and \"View n\". Therefore, conflicts can arise when merging 2D masks from different views, leading to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.243, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "19325" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.471, + 0.106 + ], + "angle": 0, + "content": "fragmented 3D segmentation and decreased performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.483, + 0.272 + ], + "angle": 0, + "content": "Besides, existing work has not effectively addressed the errors introduced by long-distance projections in the 2D-to-3D matching process, as objects at different projection distances are assigned the same weight. Compared with the close-distance scene in Fig. 2 (a), the long-distance scene in Fig. 2 (b) erroneously projects the 3D points onto the \"wall\", causing a 3D object to incorrectly match with the 2D mask of another object. Therefore, the two originally separate objects are wrongly segmented as one. These errors increase as objects move farther from the camera, leading to incorrect 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.363 + ], + "angle": 0, + "content": "To address the inter-view inconsistency, in this work, we introduce SAM2Object, a novel approach that harnesses the strong generalization capability of SAM2 [35] and its multiview consistency. SAM2Object consolidate the consistency of 2D masks across multiple views and integrate 3D geometric priors to achieve superior 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Additionally, to solve the problem of long-distance projection error, we propose a mask consolidation algorithm that applies weighting based on the distance of object from the camera across different views. During the 2D-to-3D matching process, the same object is prioritized to match with the view closer to the camera, thereby enhancing view consistency. The proposed SAM2Object conducts zero-shot 3D instance segmentation without the need for annotated data or additional training." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.5, + 0.411, + 0.514 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.515, + 0.483, + 0.575 + ], + "angle": 0, + "content": "- We proposed SAM2Object, a novel zero-shot 3D instance segmentation method that leverages SAM2 to consolidate view consistency, effectively improving the robustness of 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.576, + 0.483, + 0.635 + ], + "angle": 0, + "content": "- We proposed a mask consistency metric to evaluate and maintain the 2D mask consistency across different views, enabling integration of 3D geometric priors and high-quality 2D masks for better 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.71 + ], + "angle": 0, + "content": "- We achieve more robust 3D instance segmentation on ScanNetV2 [5], ScanNet++ [48] and ScanNet200 [36]. Our method demonstrate superior performance compared with previous methods and outperform supervised models in zero-shot task." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.515, + 0.483, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.724, + 0.233, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Closed-vocabulary 3D segmentation. 3D segmentation methods have rapidly advanced due to the emergence of high-quality 3D datasets [5, 48]. The task of 3D semantic segmentation [9, 23, 26, 33] distinguishes various categories of objects within a scene. While 3D instance segmentation [6, 16, 21, 40] is required to further differentiate between various instances within each category. Point-Group [16] is a CNN-based method that segments objects using both the original and offset-shifted point sets, which groups nearby points with the same label and pro" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.232, + 0.907, + 0.33 + ], + "angle": 0, + "content": "Figure 2. The quality of superpoint projection with different distances. In a scene, when an object is closer to the camera, the projection of 3D points onto the 2D image is more accurate in (a). However, if the object is farther from the camera in \\((\\mathrm{b})\\sim (\\mathrm{d})\\), projection errors are more likely to occur (the blue dots represent the 3D points projected onto the 2D image, colored masks correspond to projection targets)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.508 + ], + "angle": 0, + "content": "gressively expands the group. Mask3D [38] introduced the first transformer-based 3D segmentation method, which directly predicts instance masks while avoiding manually selected voting schemes, achieving state-of-the-art performance. However, Mask3D still requires a substantial volume of 3D annotated data for training, which suffer poor performance on unseen data outside their domain-specific datasets. In this work, We directly leverage the zero-shot capability of SAM2 for 3D segmentation, eliminating the need for model training or 3D annotation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Open-vocabulary 3D segmentation. Recently, vision foundation models have achieved excellent performance in 2D understanding tasks [3, 4, 19, 35], leading to the emergence of notable work in open-vocabulary 2D segmentation tasks [12, 15, 43, 44]. Inspired by these efforts, some works has focused on exploring open-vocabulary capabilities in 3D tasks. OpenMask3D [42] and OpenIns3D [14] leverage Mask3D [38] to extract 3D instances in 3D space and then directly project them onto different 2D views for extracting open-vocabulary features. These methods suffer poor performance in recognizing small or detailed objects. LERF [18] integrates language embeddings into NeRF [31], by performing voxel rendering of CLIP [34] embeddings along training rays and supervising these embeddings across training views. LERF focuses on roughly locating specific objects through text prompts rather than achieving fine-grained 3D segmentation and has shown to be inadequate in complex scenes. In this study, we focus on exploring how to achieve better fine-grained 3D segmentation solely leveraging SAM2, without relying on any 3D pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Segment Anything Model. The Segment Anything Model (SAM) [19] has made significant breakthroughs in 2D segmentation tasks, demonstrating a strong generalization ability to segment any object in any image. The upgraded version of SAM, SAM2 [35], has further expanded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19326" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.424 + ], + "angle": 0, + "content": "into the domain of video segmentation. The image predictor of SAM2 remains consistent with SAM, while the video predictor is used to segment objects within videos. It is equipped with a memory that stores information about the object and previous interactions, enabling it to generate and correct predictions throughout the video. Recently, several 2D-to-3D lifting methods leverage SAM to perform frame-by-frame segmentation of multi-view images and project them onto 3D space. SAM3D [47] leads the way in extending SAM onto 3D segmentation by projecting segmentation data from 2D images onto 3D space. SAI3D [49] and SAM-graph [11] similarly adopts the approach of segmenting each frame independently, projecting the 2D segmentations onto superpoints and then merges superpoints using either region growing or Graph Neural Networks (GNNs). However, these methods do not link adjacent frames, potentially causing inconsistencies in 2D segmentation across consecutive frames, which can result in fragmented 3D segmentation. In this study, we effectively leverage SAM2 to establish connections across multi-view images, ensuring consistency in 2D segmentation between views, thereby achieving more robust 3D instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.192, + 0.454 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.458, + 0.479 + ], + "angle": 0, + "content": "3.1. Problem formulation and method overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Given a series of posed color images \\(\\{I_1, I_2, \\dots, I_T\\}\\) and the point cloud P, our goal is to output the 3D instance segmentation of the scene. In this paper, we propose a novel zero-shot 3D instance segmentation framework, as illustrated in Fig. 3. We first use SAM2 [35] to segment and track objects, ensuring that the masks accurately cover objects even as the viewpoint changes (Sec. 3.2), thereby consolidating multi-view consistency. Then, we combine 2D mask proposals with over-segmented 3D superpoints to construct a graph which is described in Sec. 3.3. In Sec. 3.4, we introduce an iterative graph clustering technique that hierarchically clusters different superpoints based on their view consistency levels, ultimately achieving 3D segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.267, + 0.739 + ], + "angle": 0, + "content": "3.2. 2D mask proposal" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.288, + 0.761 + ], + "angle": 0, + "content": "3.2.1. Keyframes extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We define keyframe as the frame where new objects appear. An illustration of the keyframe can be found in Fig. 4. In order to segment and track new objects, it is essential to extract the keyframes and perform segmentation and tracking on them. Given a set of color images, we can treat it as a video. When there is a scene change or a new object appears, the pixel differences between the previous frame \\( I_{pre} \\) and the current frame \\( I_{cur} \\) tend to increase significantly. We use these frame differences to identify keyframes. The" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "frame differences \\( I_{dif} \\) between two adjacent frames can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.118, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\[\nI _ {d i f} = \\frac {\\left| I _ {p r e} - I _ {c u r} \\right|}{I _ {h} \\cdot I _ {w}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.155, + 0.907, + 0.248 + ], + "angle": 0, + "content": "\\(I_h\\) and \\(I_w\\) represent the height and width of the color images, respectively. After extracting the differences between each pair of adjacent frames across the entire sequence of images, we apply a Hanning window [10] to smooth the frame difference array in order to mitigate the interference of noise. The window weight can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.257, + 0.907, + 0.291 + ], + "angle": 0, + "content": "\\[\nw _ {n} = \\left\\{ \\begin{array}{c c} 0. 5 - 0. 5 \\cos \\left(\\frac {2 \\pi n}{N - 1}\\right), & 0 \\leq n \\leq N - 1, \\\\ 0, & o t h e r w i s e. \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.302, + 0.906, + 0.363 + ], + "angle": 0, + "content": "\\(N\\) is the window length, \\(n\\) represents the index within the window, and \\(w_{n}\\) denotes the value of the window function at position \\(n\\). We define the smoothed array as \\(S\\), where the \\(j\\)-th position in array \\(S\\) can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.373, + 0.907, + 0.415 + ], + "angle": 0, + "content": "\\[\nS _ {j} = \\frac {\\sum_ {n = 0} ^ {N - 1} w _ {n} \\cdot I _ {d i f} [ j - n ]}{\\sum_ {n = 0} ^ {N - 1} w _ {n}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.423, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Where \\( I_{dif}[j - n] \\) denotes the value at the \\( j - n \\) position in \\( I_{dif} \\). The indices corresponding to the local maxima in the smoothed array \\( S \\) represent the keyframes. We control the average distance between all keyframes to be around 100 frames by setting the window size appropriately." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.508, + 0.75, + 0.523 + ], + "angle": 0, + "content": "3.2.2. Segmentation and tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.844 + ], + "angle": 0, + "content": "We employ SAM2 [35] for object segmentation and tracking to consolidate view consistency. SAM2 utilizes a sophisticated memory mechanism to track objects across frames, ensuring continuity even when objects are temporarily occluded or exit and re-enter the scene. Although SAM2's auto-segmentation technique does not require any points, boxes, or mask prompts, these prompts are necessary for object tracking. As shown in Fig. 3 (b), given a series of images and keyframes, we treat the image sequences between every two keyframes as video clips, with the first and last frames of the clip being the keyframe. We use the image predictor of SAM2 to automatically segment the keyframes and generate 2D masks for objects. These masks then serve as prompts for the video predictor of SAM2 to track the objects within the video clips. We segment new objects from the keyframes and track them in the video clips between key frames. Finally, at the edge of adjacent video clips, we link the same objects across different clips according to IoU metric. In the end, we obtain the segmentation and tracking results for all objects throughout the entire video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.717, + 0.867 + ], + "angle": 0, + "content": "3.2.3. Merging segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Even though we can segment keyframes to obtain new objects, keyframes are not always the exact frames where new" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19327" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.339, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.089, + 0.896, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.442, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 3. Method overview. Our method consists of six steps. We over-segment the 3D mesh / point cloud into superpoints. (a) We extract keyframes from posed images. (b) We leverage SAM2 to segment the keyframes, and its segmentation serve as mask prompts for SAM2 to track objects throughout the entire set of posed images. (c) We consolidate the consistency of the masks obtained by SAM2 to filter out low-quality mask proposals. (d) We combine the masks with 3D geometric priors to construct a graph, where the nodes are superpoints and the edges represent the affinity scores between masks of superpoints that have undergone consistency consolidation. (e) We perform iterative graph clustering on the graph to achieve 3D instance segmentation." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.541, + 0.481, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.789, + 0.483, + 0.845 + ], + "angle": 0, + "content": "Figure 4. Segmentation and tracking. (a) We extract keyframes from posed images. (b) We use the keyframes segmentation as mask prompts to guide SAM2 in tracking objects in both directions. (c) We merge the results to ensure that no objects are missed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "objects appear. If the new objects appear before the key frames, the new objects may be missed or overlooked. To address this issue, we adopt a bidirectional tracking strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.737 + ], + "angle": 0, + "content": "(forward and backward). As shown in Fig. 4 (b), objects missed during forward tracking are identified in backward tracking (such as the \"trash can\" in the second column of images). Similarly, objects missed in backward tracking are captured in forward tracking (such as the \"door\" in the fourth column of images). The tracking results from the different directions complement each other. So we merge the 2D masks generated from both directions. An illustration of merging segmentation can be found in Fig. 4 (c). The segmentation we obtain maintain consistency across multiple views, which is crucial for subsequently combining 3D geometric priors to achieve more robust 3D instance segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.75, + 0.704, + 0.767 + ], + "angle": 0, + "content": "3.3. Graph construction" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.771, + 0.79 + ], + "angle": 0, + "content": "3.3.1. Building the superpoint graph" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Given a point cloud \\( P \\in \\mathbb{R}^{N \\times 3} \\), we represent the scene or point cloud as a graph \\( G = (V, E) \\), where \\( V \\) denotes the set of superpoints \\( \\{S_i\\}_{i=1}^{N_s} \\) in the point cloud and \\( E \\) denotes the relationships between these superpoints. Following recent works [47, 49], we employ a graph-cut algorithm [8] to over-segment the whole point cloud based on 3D geometric attributes, resulting in a series of superpoints. Each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19328" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "node in the graph represents the multiple points in the point cloud that share similar geometric features, and the number of superpoints is substantially lower than the number of points in the point cloud. By performing calculations at the superpoint-level, we achieve efficient computation while also extracting geometric priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.317 + ], + "angle": 0, + "content": "As shown in Fig. 3 (d), all color images from the different views are fed into SAM2 to obtain 2D masks, and we then project superpoints onto these 2D masks in all views to match superpoints with these masks. We use a common pinhole camera matrix to perform the 3D-to-2D projection, utilizing the corresponding camera intrinsic parameter \\( R_{k} \\) and pose parameter \\( E_{k} \\) to project the \\( i \\)-th 3D superpoint \\( S_{i} \\) onto the \\( k \\)-th image, and get its corresponding 2D mask proposals:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.318, + 0.483, + 0.335 + ], + "angle": 0, + "content": "\\[\nS _ {i, k} ^ {2 D} = \\phi \\left(R _ {k}, E _ {k}, S _ {i}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.51 + ], + "angle": 0, + "content": "\\(\\phi(\\cdot)\\) is the projection operator. The relationships between superpoints are represented by the affinity scores between their corresponding 2D masks, which are used to construct the edges of the graph. Specifically, we compute a normalized histogram based on the 2D mask proposals \\(S_{i,k}^{2D}\\) and convert it into a vector, denoted as \\(\\mathbf{e}_i^k = \\text{vec(normalized}(S_{i,k}^{2D}))\\). This histogram vector represents the 2D class-agnostic instance features of the 3D superpoint, and we get the affinity score between different superpoints on the \\(k\\)-th image by computing the cosine similarity of their 2D class-agnostic instance features:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.521, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\nA _ {i, j} ^ {k} = \\frac {\\mathbf {e} _ {i} ^ {k} \\cdot \\mathbf {e} _ {j} ^ {k}}{\\left| \\mathbf {e} _ {i} ^ {k} \\right| \\left| \\mathbf {e} _ {j} ^ {k} \\right|}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.273, + 0.578 + ], + "angle": 0, + "content": "3.3.2. Mask consolidation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Traversing all the images will yield the overall affinity score between \\( S_{i} \\) and \\( S_{j} \\), but the distribution of 2D mask proposals covered by the corresponding superpoint may vary due to changes in the shape of the same object from different views, as well as factors such as occlusion, motion blur, and projection errors. To address this issue, we introduce a mask consistency weight to filter out low-quality mask proposals, thereby clarifying the relationships between superpoints in the graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.868 + ], + "angle": 0, + "content": "We calculate the visibility of 3D points on 2D images to address object occlusion and visibility issues. To obtain the visible points, we project the points from the point cloud onto the image, filter out the points that fall outside the image, and determine occlusion between objects based on their depth values. We define \\( V_{i}^{k}\\in (0,1) \\) as the ratio of the visible point count of \\( S_{i} \\) on the \\( k \\)-th image to its total point count in the point cloud. Thus, we denote the visibility edge weight between the \\( i \\)-th superpoint and the \\( j \\)-th superpoint as" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {i, j} ^ {k} = V _ {i} ^ {k} \\cdot V _ {j} ^ {k}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.302 + ], + "angle": 0, + "content": "The projection accuracy of the same superpoint in the point cloud can differ across various views. Projections in views closer to the camera tend to be more accurate than those in views farther from the camera. These errors can lead to mismatches between the 2D mask proposals \\( S_{i,k}^{2D} \\) and the superpoint \\( S_{i} \\) (as shown in Fig. 2, the superpoint that should have been matched to the \"table\" ended up being matched to the \"wall\"), which in turn leads to errors in the segmentation results. In addition, objects that are farther away from the camera appear smaller in the image, which also makes it easy to generate inaccurate 2D masks. We define \\( D_{i}^{k} \\in (0,1) \\) as the distance weight to perform weighted averaging of the 2D mask proposals \\( S_{i,k}^{2D} \\) based on the distance from the \\( k \\)-th camera:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.308, + 0.907, + 0.345 + ], + "angle": 0, + "content": "\\[\nD _ {i} ^ {k} = 0. 9 \\cdot \\frac {d _ {m a x} ^ {k} - d _ {i} ^ {k}}{d _ {m a x} ^ {k} - d _ {m i n} ^ {k}} + 0. 1, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.351, + 0.907, + 0.442 + ], + "angle": 0, + "content": "where \\(d_i^k\\) represents the distance of visible points from the \\(k\\)-th camera within \\(S_{i,k}^{2D}\\), \\(d_{max}^k\\) and \\(d_{min}^k\\) denotes the maximum and minimum distances of all visible points from the \\(k\\)-th camera. So, we denote the distance edge weight between the \\(i\\)-th superpoint and the \\(j\\)-th superpoint as \\(\\psi_{i,j}^k\\), which can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.455, + 0.907, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\psi_ {i, j} ^ {k} = D _ {i} ^ {k} \\cdot D _ {j} ^ {k}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.906, + 0.599 + ], + "angle": 0, + "content": "As shown in Fig. 1, the number of 2D mask proposals \\( S_{i,k}^{2D} \\) covered by the corresponding superpoint is defined as its purity. If a superpoint covers more 2D mask proposals, it is considered a difficult sample, and we reduce its weight and merge it only during the later stages of graph clustering in Sec. 3.4. We define \\( L_i^k \\in (0,1) \\) as the maximum ratio of all points within a superpoint projected onto different 2D mask proposals of \\( k \\)-th image:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.606, + 0.906, + 0.644 + ], + "angle": 0, + "content": "\\[\nL _ {i} ^ {k} = \\max \\left(\\frac {N _ {i , 1} ^ {k}}{N _ {i} ^ {k}}, \\frac {N _ {i , 2} ^ {k}}{N _ {i} ^ {k}}, \\dots , \\frac {N _ {i , n} ^ {k}}{N _ {i} ^ {k}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.65, + 0.907, + 0.744 + ], + "angle": 0, + "content": "\\(N_{i}^{k}\\) represents the total number of visible points of the \\(i\\)-th superpoint in the \\(k\\)-th image, and \\(N_{i,j}^{k}\\) denotes the number of points belonging to the mask label with a value of \\(j\\), \\(n\\) is the total number of 2D mask proposals. So, we denote the purity edge weight between the \\(i\\)-th superpoint and the \\(j\\)-th superpoint as \\(\\Phi_{i,j}^{k}\\), which can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.755, + 0.905, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {i, j} ^ {k} = L _ {i} ^ {k} \\cdot L _ {j} ^ {k}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.777, + 0.906, + 0.808 + ], + "angle": 0, + "content": "Finally, we combine these three weights to form the mask consistency weight as" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.814, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\nW _ {i, j} ^ {k} = \\varphi_ {i, j} ^ {k} \\cdot \\Phi_ {i, j} ^ {k} \\cdot \\psi_ {i, j} ^ {k}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.805, + 0.856 + ], + "angle": 0, + "content": "Thus, the final affinity score is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.864, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nA _ {i, j} = \\frac {\\sum_ {k = 1} ^ {M} \\left(W _ {i , j} ^ {k} A _ {i , j} ^ {k}\\right)}{\\sum_ {k = 1} ^ {M} W _ {i , j} ^ {k}}. \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19329" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.325, + 0.108 + ], + "angle": 0, + "content": "3.4. Iterative graph clustering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.482, + 0.295 + ], + "angle": 0, + "content": "Building upon the superpoint graph, we introduce an iterative graph clustering technique to merge superpoints and update the graph. In each iteration, we employ a region growing method to merge superpoints, considering not only the affinity scores but also weighting based on the distance between superpoints. From a geometric perspective, superpoints that are far apart are less likely to be similar. However, once projected, they may appear closer on images from specific viewpoints. According to Sec. 3.3.2, their 2D mask proposals tend to be of lower quality, resulting in incorrect affinity scores. So we set the edge weight decreases for superpoints that are farther apart." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.508 + ], + "angle": 0, + "content": "Due to the introduction of the Sec. 3.3.2, some superpoints covering low-quality 2D mask proposals have lower view consistency and affinity scores. These superpoints may be located at the junction of two objects. Therefore, in each iteration, we set different thresholds for merging superpoints based on the affinity scores. We initially merge small and adjacent regions with high threshold, gradually lowering the threshold over iterations to merge superpoints with lower confidence. This iterative graph clustering approach allows for hierarchical merging of superpoints with varying levels of view consistency. Superpoints with stronger view consistency are prioritized for merging. Subsequently, superpoints with lower view consistency are merged, ultimately achieving robust 3D segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.523, + 0.223, + 0.54 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.641 + ], + "angle": 0, + "content": "We evaluate our SAM2Object on multiple datasets to demonstrate its effectiveness and robustness in zero-shot 3D instance segmentation. We also compare our SAM2Object with leading methods in the open-vocabulary segmentation field and the state-of-the-art closed-vocabulary segmentation methods which require training on annotated datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.281, + 0.669 + ], + "angle": 0, + "content": "4.1. Experiment settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. ScanNetV2 [5] is a large-scale dataset tailored for indoor scene understanding and segmentation. It comprises RGB and depth images, as well as 3D point cloud data derived from indoor scenes. It includes over 2.5 million extensive views from more than 1,500 scans, along with semantic labels for 21 categories. ScanNet200 [36] is an extension of ScanNetV2, increasing the number of semantic categories to 200. Furthermore, ScanNet++ [48] is a newly released indoor dataset contains 280 indoor scenes and couples geometric and color data of high-quality, commodity-level indoor scenes. Each scene is captured by laser scanners with higher-resolution and a DSLR camera with a fisheye lens, as well as RGB-D streams obtained from iPhone 13 Pro. We evaluate our method on the well-studied ScanNetV2, ScanNet++ and ScanNet200 datasets. Following SAI3D [49]," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "When generating 2D mask proposals, we use only \\(20\\%\\) of the images from the ScanNetV2 dataset and only \\(5\\%\\) from the ScanNet++ dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.21 + ], + "angle": 0, + "content": "Table 1. Class-agnostic 3D instance segmentation on ScanNetV2 dataset. Bold and underlined values indicate the best and second best scores, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.221, + 0.905, + 0.457 + ], + "angle": 0, + "content": "
MethodTraining SetmAP\\( AP_{50} \\)\\( AP_{25} \\)
Closed-vocabulary
Mask3D [38]S3DIS31.144.958.0
Open-vocabulary
HDBSCAN [30]-1.65.532.1
UCIS [32]-2.37.330.5
EGIS [7]-5.012.738.9
Part2Object [39]-12.626.855.1
SAM-graph [11]-15.133.359.1
UnScene3D [37]-15.932.258.5
SAMPro3D [46]-18.032.856.3
SAM3D [47]-20.234.053.3
SAI3D [49]-30.850.570.6
SAM2Object (Ours)-34.052.770.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate the quantitative results with the widely adopted Average Precision (AP) score. Following [11, 38, 42, 49], we calculate AP with IoU thresholds of \\(50\\%\\) and \\(25\\%\\) \\((\\mathrm{AP}_{50}, \\mathrm{AP}_{25})\\) and AP averaged with thresholds from \\(50\\%\\) to \\(95\\%\\) with a step size of \\(5\\%\\) (mAP). We implemented two distinct evaluation settings to comprehensively validate the performance of our method. Firstly, we evaluate class-agnostic instance segmentation which disregards category information and solely evaluates the accuracy of instance masks. Secondly, semantic instance segmentation builds on this by considering the semantic label of each instance, emphasizing the relationship between an instance and its category." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our method with both open-vocabulary and closed-vocabulary baselines. For closed-vocabulary methods, we compare with the state-of-the-art method, Mask3D [38], which is a supervised model requiring training data. For open-vocabulary methods, we compare with the recent lifting methods based on leveraging SAM, including SAM3D [47], SAMPro3D [46], Segment3D [13], SAM-graph [11], SAI3D [49] and OVIR-3D [27]. We also compared with OpenMask3D [42], which generates masks using Mask3D and assigns semantic labels. Additionally, three traditional methods, referred to as HDBSCAN [30], EGIS [7] and UCIS [32], as well as the unsupervised methods UnScene3D [37] and Part2Object [39] are also included in our comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19330" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 2. Class-agnostic 3D instance segmentation on ScanNet++ dataset. Bold and underlined values indicate the best and second best scores, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.482, + 0.317 + ], + "angle": 0, + "content": "
MethodTraining SetmAP\\( AP_{50} \\)\\( AP_{25} \\)
Closed-vocabulary
Mask3DScanNetV29.917.325.8
Open-vocabulary
EGIS [7]-4.19.225.3
SAM3D [47]-7.214.229.4
Segment3D [13]-12.022.737.8
SAM-graph [11]-12.925.343.6
SAI3D [49]-17.131.149.5
SAM2Object (Ours)-20.234.148.7
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.333, + 0.185, + 0.347 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Quantitative results. We conduct the experiments on ScanNetV2, ScanNet++ and ScanNet200. Table 1 displays the quantitative results of our method on class-agnostic instance segmentation within ScanNetV2. Our method outperforms previous works in all metrics and leads in mAP and \\(\\mathrm{AP}_{50}\\) compared with the state-of-the-art methods SAI3D, trailing by only 0.3 in the \\(\\mathrm{AP}_{25}\\) metric. Notably, SAI3D requires adaptation of different 2D segmentation models for different datasets (using Semantic-SAM [22] on ScanNetV2 and HQ-SAM [17] on ScanNet++), whereas we consistently use SAM2 for validation across different datasets. Furthermore, our method outperforms both unsupervised and closed-vocabulary methods. The latter requires training data, so to conduct a fair zero-shot evaluation, we replace its training data with S3DIS [2] and validate on ScanNetV2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.583, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Table 2 reports the class-agnostic instance segmentation results on the challenging ScanNet++, we still lead in the mAP and \\(\\mathrm{AP}_{50}\\) metric. In lifting methods, the performance of the 2D segmentation model is crucial to the final 3D instance segmentation, SAI3D uses the optimal segmentation model for each dataset, while we consistently use SAM2. We simply replace the optimal segmentation model in SAI3D with SAM2, resulting in a mAP of \\(28.23\\%\\) on ScanNetV2 and \\(16.02\\%\\) on ScanNet++, both lower than SAI3D. This demonstrates the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.734, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Following OpenMask3D [42] and SAI3D [49], we validate the semantic instance segmentation results on ScanNet200, as shown in Table 3. We assign semantic labels to our segmented 3D instances using the method from OpenMask3D. We exhibit outstanding performance in both mAP and \\(\\mathrm{AP}_{50}\\), indicating that our method demonstrates greater generalization when handling different semantic labels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our method has lower \\(\\mathrm{AP}_{25}\\) than SAI3D across the three datasets. SAI3D lacks view consistency, resulting in fragmented and over-segmented outputs, which partially contributes to its higher \\(\\mathrm{AP}_{25}\\). Additionally, SAI3D selects" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.153 + ], + "angle": 0, + "content": "the most suitable visual foundation model for each dataset, achieving fine-grained compared with SAM2, which explains their higher \\(\\mathrm{AP}_{25}\\). In contrast, we prefer to design a method capable of generalizing across multiple datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.165, + 0.906, + 0.222 + ], + "angle": 0, + "content": "Table 3. Semantic instance segmentation on ScanNet200 dataset. \"OV\" and \"Sup\" represent \"Open-vocabulary\" and \"Supervised\" respectively. Bold and underlined values indicate the best and second best scores, respectively." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.233, + 0.892, + 0.376 + ], + "angle": 0, + "content": "
MethodmAP\\( AP_{50} \\)\\( AP_{25} \\)
Supmasked + OVsemantic OpenMask3D [42]15.419.923.1
OVmasked + OVsemantic
OVIR-3D [27]9.318.725.0
SAM3D [47]9.815.220.7
SAI3D [49]12.718.824.1
SAM2Object (Ours)13.319.023.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Qualitative results. We qualitatively compared our method with the SOTA method SAI3D on ScanNetV2 and ScanNet++. SAI3D is unable to handle the problem of projection errors, and the mismatched projections in Fig. 3 (c) and Fig. 2 result in incorrect 3D segmentation in Fig. 5 (a) and Fig. 5 (b). Moreover, SAI3D employs a frame-by-frame segmentation approach across multiple views, which lacks view consistency. This inconsistency in Fig. 1 leads to fragmented 3D segmentation in Fig. 5 (a), and similar fragmentation can be observed in Fig. 5 (c). Furthermore, the under-segmentation is also evident in Fig. 5 (d)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.613 + ], + "angle": 0, + "content": "In contrast, our approach consolidates view consistency and filters out low-quality 2D masks, enabling us to achieve comprehensive and robust 3D segmentation. More results can be found in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.626, + 0.882, + 0.641 + ], + "angle": 0, + "content": "Table 4. Ablations of main components in our method." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.652, + 0.9, + 0.745 + ], + "angle": 0, + "content": "
Seg. & TrackMask consolidationmAPAP50AP25
28.244.765.2
✓29.347.666.0
✓32.149.968.3
✓✓34.052.770.3
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.719, + 0.774 + ], + "angle": 0, + "content": "4.3. Ablation and analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In Table 4, we performed ablation experiments on ScanNetV2 to verify the effectiveness of each component of our method. Starting from the baseline, similar to existing 2D lifting methods, we use the image predictor of SAM2 to perform frame-by-frame segmentation on multi-view images, setting the weight of all visible views equally. This approach attempts to achieve view consistency by averaging results across all visible views. However, since there is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.481, + 0.908, + 0.539 + ], + "angle": 0, + "content": "Figure 5. Qualitative results comparisons between our method with SAI3D [49] on ScanNetV2 and ScanNet++. Due to projection errors resulting in incorrect 2D-to-3D matching relationships, SAI3D mistakenly segments part of the chair as the wall in (a), and segments the table and wall as a single object in (b). Due to the lack of view consistency, SAI3D exhibits fragmented segmentation in (c), as well as under-segmentation in (d). Our method shows the most robust and accurate class-agnostic instance segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.487, + 0.73 + ], + "angle": 0, + "content": "no consistency between the masks of consecutive frames, it leads to fragmented segmentation. By applying mask consistency weighting through the mask consolidation module to results from multiple views, this strategy reduces low-quality mask interference during graph clustering (Sec. 3.4), resulting in a slight performance improvement. Employing the video predictor of SAM2 for object tracking ensures mask consistency across multi-view images. This results in more precise matching with 3D superpoints in Sec. 3.3, leading to improved performance. Finally, combining these two strategies fully leverages the benefits of consolidated view consistency, achieving further improved performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.742, + 0.219, + 0.756 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Our method enhances view consistency by tracking objects across multiple views using SAM2 but requires continuous and dense multi-view images. Furthermore, although vision foundation models exhibit strong generalization, they are only capable of performing inference on 2D data and cannot fully understand scenes from a 3D scene. Although our approach connects adjacent frames, it is less effective than methods that integrate 3D priors with 2D semantics to establish view consistency. Future work should leverage 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.593 + ], + "angle": 0, + "content": "geometric priors to enhance visual foundation models in understanding 3D scenes from multi-view images, improving their generalization for 3D tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.608, + 0.634, + 0.623 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.909, + 0.816 + ], + "angle": 0, + "content": "In this paper, we introduce a novel zero-shot 3D instance segmentation method, SAM2Object. The key idea is to achieve more accurate 2D-to-3D matching by consolidating view consistency. We employ SAM2 to generate consistent masks across multiple views and introduce a metric to measure mask consistency. After constructing the superpoint graph, we perform iterative clustering based on view consistency levels, achieving more accurate segmentation. Comprehensive evaluations across multiple datasets demonstrate the effectiveness of SAM2Object, highlighting its ability to outperform previous methods and rival fully supervised techniques in zero-shot instance segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.693, + 0.848 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China: 62402033, 62227801 and 62376024." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19332" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[7] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[8] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Thomas Grandke. Interpolation algorithms for discrete fourier transforms of weighted signals. IEEE transactions on instrumentation and measurement, 32(2):350-355, 1983. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[11] Haoyu Guo, He Zhu, Sida Peng, Yuang Wang, Yujun Shen, Ruizhen Hu, and Xiaowei Zhou. Sam-guided graph cut for 3d instance segmentation. In ECCV, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Shuting He, Henghui Ding, and Wei Jiang. Semantic-promoted debiasing and background disambiguation for zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19498-19507, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "segmentation without manual labels. In European Conference on Computer Vision, pages 278-295. Springer, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[15] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[16] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[17] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.461, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[20] Maksim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Top-down beats bottom-up in 3d instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3566-3574, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[21] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[22] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[24] Wenteng Liang, Feng Xue, Yihao Liu, Guofeng Zhong, and Anlong Ming. Unknown sniffer for object detection: Don't turn a blind eye to unknown objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3230-3239, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19333" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "international conference on computer vision, pages 2783-2792, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Haizhuang Liu, Junbao Zhuo, Chen Liang, Jiansheng Chen, and Huimin Ma. Affinity3d: Propagating instance-level semantic affinity for zero-shot point cloud semantic segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9019-9028, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data. In Conference on Robot Learning, pages 1610–1620. PMLR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[28] Weijie Lyu, Xueting Li, Abhijit Kundu, Yi-Hsuan Tsai, and Ming-Hsuan Yang. Gaga: Group any gaussians via 3d-aware memory bank. arXiv preprint arXiv:2404.07977, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[29] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. In Advances in Neural Information Processing Systems, pages 32340-32352. Curran Associates, Inc., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[30] Leland McInnes and John Healy. Accelerated hierarchical density based clustering. In 2017 IEEE international conference on data mining workshops (ICDMW), pages 33-42. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[32] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters, 7(4):8713-8720, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 6, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[37] David Rozenberszki, Or Litany, and Angela Dai. Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19957-19967, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[39] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[40] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[41] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[42] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[43] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.907, + 0.585 + ], + "angle": 0, + "content": "[44] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635–3647, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[45] Yizhou Wang, Longguang Wang, Qingyong Hu, Yan Liu, Ye Zhang, and Yulan Guo. Panoptic segmentation of 3d point clouds with gaussian mixture model in outdoor scenes. Visual Intelligence, 2(1):10, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[46] Mutian Xu, Xingyilang Yin, Lingteng Qiu, Yang Liu, Xin Tong, and Xiaoguang Han. Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation. arXiv preprint arXiv:2311.17707, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[47] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 1, 3, 4, 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[48] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.799, + 0.905, + 0.868 + ], + "angle": 0, + "content": "[49] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3292–3302, 2024. 3, 4, 6, 7, 8, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19334" + } + ] +] \ No newline at end of file diff --git a/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_origin.pdf b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a93e614d43ab29a8f9f11c6333537d794b048715 --- /dev/null +++ b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/663b1e37-f7c9-4cbb-87f6-8d16e84f1d5e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50b031859f92f2960d9f1c4f49b3652b2aaa87ffdc2fbee06d8154e5630f733b +size 9743432 diff --git a/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/full.md b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26aa7f58e0135203c530dbd6bf1687f94ac00a11 --- /dev/null +++ b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/full.md @@ -0,0 +1,301 @@ +# SAM2Object: Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation + +Jihuai Zhao Junbao Zhuo* Jiansheng Chen Huimin Ma* University of Science and Technology Beijing + +jihuaizhao@xs.ustb.edu.cn, {junbaozhuo, jschen, mhmpub}@ustb.edu.cn + +# Abstract + +In the field of zero-shot 3D instance segmentation, existing 2D-to-3D lifting methods typically obtain 2D segmentation across multiple RGB frames using vision foundation models, which are then projected and merged into 3D space. However, since the inference of vision foundation models on a single frame is not integrated with adjacent frames, the masks of the same object may vary across different frames, leading to a lack of view consistency in the 2D segmentation. Furthermore, current lifting methods average the 2D segmentation from multiple views during the projection into 3D space, causing low-quality masks and high-quality masks to share the same weight. These factors can lead to fragmented 3D segmentation. In this paper, we present SAM2Object, a novel zero-shot 3D instance segmentation method that effectively utilizes the Segment Anything Model 2 to segment and track objects, consolidating view consistency across frames. Our approach combines these consistent 2D masks with 3D geometric priors, improving the robustness of 3D segmentation. Additionally, we introduce mask consolidation module to filter out low-quality masks across frames, which enables more precise 2D-to-3D matching. Comprehensive evaluations on ScanNetV2, ScanNet++ and ScanNet200 demonstrate the robustness and effectiveness of SAM2Object, showcasing its ability to outperform previous methods. Our project page is at https://jihuaizhaohd.github.io/SAM2Object. + +# 1. Introduction + +3D instance segmentation plays a crucial role in various applications such as autonomous driving, robotics and augmented reality. Most of previous works [16, 20, 25, 41, 45] have predominantly relied on supervised learning, requiring high-quality 3D data and annotations for training. The annotation process for 3D data is often labor-intensive and costly, leading to the scarcity of high-quality 3D datasets. + +![](images/ceac3aed7ab363987ac077132dca74ea9ea4a52e1e02d6d610e023403fe1bc8c.jpg) +Figure 1. View Consistency. Compared with performing frame-by-frame segmentation on multi-view images, our method connects objects across multiple frames, ensuring that segmentation results remain consistent across various views. + +This scarcity, in turn, greatly limits the development and advancement of 3D vision models, often leading to models with weak generalization capabilities, particularly in terms of zero-shot generalization. Zero-shot capability is essential, particularly in indoor navigation scenarios [1, 29]. For example, a robot misidentifying an unknown object, like a stove, could pose a fire hazard and endanger people. + +The relative ease of acquiring 2D annotated data has led to the creation of numerous high-quality datasets, accelerating the advancement of 2D vision foundation models [3, 4, 19, 35]. Segment Anything Model (SAM) [19] is a prime example of this progress. SAM is trained on the extensive and high-quality SA-1B dataset [19], exhibiting outstanding zero-shot performance across various types of 2D vision tasks. + +Consequently, some methods [11, 14, 42, 46, 47] have attempted to transfer the robust generalization ability of SAM to 3D tasks by lifting its 2D segmentation into the 3D space. However, these methods often perform segmentation on a frame-by-frame manner, lacking inter-view consistency. Even slight changes in viewing angle can result in inconsistent 2D masks for the same object across different views. As shown in Fig. 1, the mask completely encapsulates the object in "View 1", but only partially covers it in "View i" and "View n". Therefore, conflicts can arise when merging 2D masks from different views, leading to + +fragmented 3D segmentation and decreased performance. + +Besides, existing work has not effectively addressed the errors introduced by long-distance projections in the 2D-to-3D matching process, as objects at different projection distances are assigned the same weight. Compared with the close-distance scene in Fig. 2 (a), the long-distance scene in Fig. 2 (b) erroneously projects the 3D points onto the "wall", causing a 3D object to incorrectly match with the 2D mask of another object. Therefore, the two originally separate objects are wrongly segmented as one. These errors increase as objects move farther from the camera, leading to incorrect 3D segmentation. + +To address the inter-view inconsistency, in this work, we introduce SAM2Object, a novel approach that harnesses the strong generalization capability of SAM2 [35] and its multiview consistency. SAM2Object consolidate the consistency of 2D masks across multiple views and integrate 3D geometric priors to achieve superior 3D instance segmentation. + +Additionally, to solve the problem of long-distance projection error, we propose a mask consolidation algorithm that applies weighting based on the distance of object from the camera across different views. During the 2D-to-3D matching process, the same object is prioritized to match with the view closer to the camera, thereby enhancing view consistency. The proposed SAM2Object conducts zero-shot 3D instance segmentation without the need for annotated data or additional training. + +In summary, our contributions are as follows: + +- We proposed SAM2Object, a novel zero-shot 3D instance segmentation method that leverages SAM2 to consolidate view consistency, effectively improving the robustness of 3D segmentation. +- We proposed a mask consistency metric to evaluate and maintain the 2D mask consistency across different views, enabling integration of 3D geometric priors and high-quality 2D masks for better 3D segmentation. +- We achieve more robust 3D instance segmentation on ScanNetV2 [5], ScanNet++ [48] and ScanNet200 [36]. Our method demonstrate superior performance compared with previous methods and outperform supervised models in zero-shot task. + +# 2. Related Work + +Closed-vocabulary 3D segmentation. 3D segmentation methods have rapidly advanced due to the emergence of high-quality 3D datasets [5, 48]. The task of 3D semantic segmentation [9, 23, 26, 33] distinguishes various categories of objects within a scene. While 3D instance segmentation [6, 16, 21, 40] is required to further differentiate between various instances within each category. Point-Group [16] is a CNN-based method that segments objects using both the original and offset-shifted point sets, which groups nearby points with the same label and pro + +![](images/afa3285739e91c2c417835b7cbbdd4b6d0d45fd9f5e5863bec96d3aa476a9771.jpg) +Figure 2. The quality of superpoint projection with different distances. In a scene, when an object is closer to the camera, the projection of 3D points onto the 2D image is more accurate in (a). However, if the object is farther from the camera in $(\mathrm{b})\sim (\mathrm{d})$ , projection errors are more likely to occur (the blue dots represent the 3D points projected onto the 2D image, colored masks correspond to projection targets). + +gressively expands the group. Mask3D [38] introduced the first transformer-based 3D segmentation method, which directly predicts instance masks while avoiding manually selected voting schemes, achieving state-of-the-art performance. However, Mask3D still requires a substantial volume of 3D annotated data for training, which suffer poor performance on unseen data outside their domain-specific datasets. In this work, We directly leverage the zero-shot capability of SAM2 for 3D segmentation, eliminating the need for model training or 3D annotation. + +Open-vocabulary 3D segmentation. Recently, vision foundation models have achieved excellent performance in 2D understanding tasks [3, 4, 19, 35], leading to the emergence of notable work in open-vocabulary 2D segmentation tasks [12, 15, 43, 44]. Inspired by these efforts, some works has focused on exploring open-vocabulary capabilities in 3D tasks. OpenMask3D [42] and OpenIns3D [14] leverage Mask3D [38] to extract 3D instances in 3D space and then directly project them onto different 2D views for extracting open-vocabulary features. These methods suffer poor performance in recognizing small or detailed objects. LERF [18] integrates language embeddings into NeRF [31], by performing voxel rendering of CLIP [34] embeddings along training rays and supervising these embeddings across training views. LERF focuses on roughly locating specific objects through text prompts rather than achieving fine-grained 3D segmentation and has shown to be inadequate in complex scenes. In this study, we focus on exploring how to achieve better fine-grained 3D segmentation solely leveraging SAM2, without relying on any 3D pre-trained models. + +The Segment Anything Model. The Segment Anything Model (SAM) [19] has made significant breakthroughs in 2D segmentation tasks, demonstrating a strong generalization ability to segment any object in any image. The upgraded version of SAM, SAM2 [35], has further expanded + +into the domain of video segmentation. The image predictor of SAM2 remains consistent with SAM, while the video predictor is used to segment objects within videos. It is equipped with a memory that stores information about the object and previous interactions, enabling it to generate and correct predictions throughout the video. Recently, several 2D-to-3D lifting methods leverage SAM to perform frame-by-frame segmentation of multi-view images and project them onto 3D space. SAM3D [47] leads the way in extending SAM onto 3D segmentation by projecting segmentation data from 2D images onto 3D space. SAI3D [49] and SAM-graph [11] similarly adopts the approach of segmenting each frame independently, projecting the 2D segmentations onto superpoints and then merges superpoints using either region growing or Graph Neural Networks (GNNs). However, these methods do not link adjacent frames, potentially causing inconsistencies in 2D segmentation across consecutive frames, which can result in fragmented 3D segmentation. In this study, we effectively leverage SAM2 to establish connections across multi-view images, ensuring consistency in 2D segmentation between views, thereby achieving more robust 3D instance segmentation. + +# 3. Methods + +# 3.1. Problem formulation and method overview + +Given a series of posed color images $\{I_1, I_2, \dots, I_T\}$ and the point cloud P, our goal is to output the 3D instance segmentation of the scene. In this paper, we propose a novel zero-shot 3D instance segmentation framework, as illustrated in Fig. 3. We first use SAM2 [35] to segment and track objects, ensuring that the masks accurately cover objects even as the viewpoint changes (Sec. 3.2), thereby consolidating multi-view consistency. Then, we combine 2D mask proposals with over-segmented 3D superpoints to construct a graph which is described in Sec. 3.3. In Sec. 3.4, we introduce an iterative graph clustering technique that hierarchically clusters different superpoints based on their view consistency levels, ultimately achieving 3D segmentation. + +# 3.2. 2D mask proposal + +# 3.2.1. Keyframes extraction + +We define keyframe as the frame where new objects appear. An illustration of the keyframe can be found in Fig. 4. In order to segment and track new objects, it is essential to extract the keyframes and perform segmentation and tracking on them. Given a set of color images, we can treat it as a video. When there is a scene change or a new object appears, the pixel differences between the previous frame $I_{pre}$ and the current frame $I_{cur}$ tend to increase significantly. We use these frame differences to identify keyframes. The + +frame differences $I_{dif}$ between two adjacent frames can be defined as + +$$ +I _ {d i f} = \frac {\left| I _ {p r e} - I _ {c u r} \right|}{I _ {h} \cdot I _ {w}}. \tag {1} +$$ + +$I_h$ and $I_w$ represent the height and width of the color images, respectively. After extracting the differences between each pair of adjacent frames across the entire sequence of images, we apply a Hanning window [10] to smooth the frame difference array in order to mitigate the interference of noise. The window weight can be calculated as + +$$ +w _ {n} = \left\{ \begin{array}{c c} 0. 5 - 0. 5 \cos \left(\frac {2 \pi n}{N - 1}\right), & 0 \leq n \leq N - 1, \\ 0, & o t h e r w i s e. \end{array} \right. \tag {2} +$$ + +$N$ is the window length, $n$ represents the index within the window, and $w_{n}$ denotes the value of the window function at position $n$ . We define the smoothed array as $S$ , where the $j$ -th position in array $S$ can be defined as + +$$ +S _ {j} = \frac {\sum_ {n = 0} ^ {N - 1} w _ {n} \cdot I _ {d i f} [ j - n ]}{\sum_ {n = 0} ^ {N - 1} w _ {n}}. \tag {3} +$$ + +Where $I_{dif}[j - n]$ denotes the value at the $j - n$ position in $I_{dif}$ . The indices corresponding to the local maxima in the smoothed array $S$ represent the keyframes. We control the average distance between all keyframes to be around 100 frames by setting the window size appropriately. + +# 3.2.2. Segmentation and tracking + +We employ SAM2 [35] for object segmentation and tracking to consolidate view consistency. SAM2 utilizes a sophisticated memory mechanism to track objects across frames, ensuring continuity even when objects are temporarily occluded or exit and re-enter the scene. Although SAM2's auto-segmentation technique does not require any points, boxes, or mask prompts, these prompts are necessary for object tracking. As shown in Fig. 3 (b), given a series of images and keyframes, we treat the image sequences between every two keyframes as video clips, with the first and last frames of the clip being the keyframe. We use the image predictor of SAM2 to automatically segment the keyframes and generate 2D masks for objects. These masks then serve as prompts for the video predictor of SAM2 to track the objects within the video clips. We segment new objects from the keyframes and track them in the video clips between key frames. Finally, at the edge of adjacent video clips, we link the same objects across different clips according to IoU metric. In the end, we obtain the segmentation and tracking results for all objects throughout the entire video. + +# 3.2.3. Merging segmentation + +Even though we can segment keyframes to obtain new objects, keyframes are not always the exact frames where new + +![](images/671d5dceac898b074dbfb3dde64f8439eb1d3e887cfa3848b6cb94eeaa0b2e68.jpg) +Figure 3. Method overview. Our method consists of six steps. We over-segment the 3D mesh / point cloud into superpoints. (a) We extract keyframes from posed images. (b) We leverage SAM2 to segment the keyframes, and its segmentation serve as mask prompts for SAM2 to track objects throughout the entire set of posed images. (c) We consolidate the consistency of the masks obtained by SAM2 to filter out low-quality mask proposals. (d) We combine the masks with 3D geometric priors to construct a graph, where the nodes are superpoints and the edges represent the affinity scores between masks of superpoints that have undergone consistency consolidation. (e) We perform iterative graph clustering on the graph to achieve 3D instance segmentation. + +![](images/4987034fe8843e3b0b0d2e331f325f27b8c7952ed73d46e1448e7dd3e2e695e6.jpg) + +![](images/c8c0d699823022c649db5ac5de128ce124e312cb8e1a318ec1239bf42daceb6c.jpg) +Figure 4. Segmentation and tracking. (a) We extract keyframes from posed images. (b) We use the keyframes segmentation as mask prompts to guide SAM2 in tracking objects in both directions. (c) We merge the results to ensure that no objects are missed. + +objects appear. If the new objects appear before the key frames, the new objects may be missed or overlooked. To address this issue, we adopt a bidirectional tracking strategy + +(forward and backward). As shown in Fig. 4 (b), objects missed during forward tracking are identified in backward tracking (such as the "trash can" in the second column of images). Similarly, objects missed in backward tracking are captured in forward tracking (such as the "door" in the fourth column of images). The tracking results from the different directions complement each other. So we merge the 2D masks generated from both directions. An illustration of merging segmentation can be found in Fig. 4 (c). The segmentation we obtain maintain consistency across multiple views, which is crucial for subsequently combining 3D geometric priors to achieve more robust 3D instance segmentation. + +# 3.3. Graph construction + +# 3.3.1. Building the superpoint graph + +Given a point cloud $P \in \mathbb{R}^{N \times 3}$ , we represent the scene or point cloud as a graph $G = (V, E)$ , where $V$ denotes the set of superpoints $\{S_i\}_{i=1}^{N_s}$ in the point cloud and $E$ denotes the relationships between these superpoints. Following recent works [47, 49], we employ a graph-cut algorithm [8] to over-segment the whole point cloud based on 3D geometric attributes, resulting in a series of superpoints. Each + +node in the graph represents the multiple points in the point cloud that share similar geometric features, and the number of superpoints is substantially lower than the number of points in the point cloud. By performing calculations at the superpoint-level, we achieve efficient computation while also extracting geometric priors. + +As shown in Fig. 3 (d), all color images from the different views are fed into SAM2 to obtain 2D masks, and we then project superpoints onto these 2D masks in all views to match superpoints with these masks. We use a common pinhole camera matrix to perform the 3D-to-2D projection, utilizing the corresponding camera intrinsic parameter $R_{k}$ and pose parameter $E_{k}$ to project the $i$ -th 3D superpoint $S_{i}$ onto the $k$ -th image, and get its corresponding 2D mask proposals: + +$$ +S _ {i, k} ^ {2 D} = \phi \left(R _ {k}, E _ {k}, S _ {i}\right). \tag {4} +$$ + +$\phi(\cdot)$ is the projection operator. The relationships between superpoints are represented by the affinity scores between their corresponding 2D masks, which are used to construct the edges of the graph. Specifically, we compute a normalized histogram based on the 2D mask proposals $S_{i,k}^{2D}$ and convert it into a vector, denoted as $\mathbf{e}_i^k = \text{vec(normalized}(S_{i,k}^{2D}))$ . This histogram vector represents the 2D class-agnostic instance features of the 3D superpoint, and we get the affinity score between different superpoints on the $k$ -th image by computing the cosine similarity of their 2D class-agnostic instance features: + +$$ +A _ {i, j} ^ {k} = \frac {\mathbf {e} _ {i} ^ {k} \cdot \mathbf {e} _ {j} ^ {k}}{\left| \mathbf {e} _ {i} ^ {k} \right| \left| \mathbf {e} _ {j} ^ {k} \right|}. \tag {5} +$$ + +# 3.3.2. Mask consolidation + +Traversing all the images will yield the overall affinity score between $S_{i}$ and $S_{j}$ , but the distribution of 2D mask proposals covered by the corresponding superpoint may vary due to changes in the shape of the same object from different views, as well as factors such as occlusion, motion blur, and projection errors. To address this issue, we introduce a mask consistency weight to filter out low-quality mask proposals, thereby clarifying the relationships between superpoints in the graph. + +We calculate the visibility of 3D points on 2D images to address object occlusion and visibility issues. To obtain the visible points, we project the points from the point cloud onto the image, filter out the points that fall outside the image, and determine occlusion between objects based on their depth values. We define $V_{i}^{k}\in (0,1)$ as the ratio of the visible point count of $S_{i}$ on the $k$ -th image to its total point count in the point cloud. Thus, we denote the visibility edge weight between the $i$ -th superpoint and the $j$ -th superpoint as + +$$ +\varphi_ {i, j} ^ {k} = V _ {i} ^ {k} \cdot V _ {j} ^ {k}. \tag {6} +$$ + +The projection accuracy of the same superpoint in the point cloud can differ across various views. Projections in views closer to the camera tend to be more accurate than those in views farther from the camera. These errors can lead to mismatches between the 2D mask proposals $S_{i,k}^{2D}$ and the superpoint $S_{i}$ (as shown in Fig. 2, the superpoint that should have been matched to the "table" ended up being matched to the "wall"), which in turn leads to errors in the segmentation results. In addition, objects that are farther away from the camera appear smaller in the image, which also makes it easy to generate inaccurate 2D masks. We define $D_{i}^{k} \in (0,1)$ as the distance weight to perform weighted averaging of the 2D mask proposals $S_{i,k}^{2D}$ based on the distance from the $k$ -th camera: + +$$ +D _ {i} ^ {k} = 0. 9 \cdot \frac {d _ {m a x} ^ {k} - d _ {i} ^ {k}}{d _ {m a x} ^ {k} - d _ {m i n} ^ {k}} + 0. 1, \tag {7} +$$ + +where $d_i^k$ represents the distance of visible points from the $k$ -th camera within $S_{i,k}^{2D}$ , $d_{max}^k$ and $d_{min}^k$ denotes the maximum and minimum distances of all visible points from the $k$ -th camera. So, we denote the distance edge weight between the $i$ -th superpoint and the $j$ -th superpoint as $\psi_{i,j}^k$ , which can be calculated as + +$$ +\psi_ {i, j} ^ {k} = D _ {i} ^ {k} \cdot D _ {j} ^ {k}. \tag {8} +$$ + +As shown in Fig. 1, the number of 2D mask proposals $S_{i,k}^{2D}$ covered by the corresponding superpoint is defined as its purity. If a superpoint covers more 2D mask proposals, it is considered a difficult sample, and we reduce its weight and merge it only during the later stages of graph clustering in Sec. 3.4. We define $L_i^k \in (0,1)$ as the maximum ratio of all points within a superpoint projected onto different 2D mask proposals of $k$ -th image: + +$$ +L _ {i} ^ {k} = \max \left(\frac {N _ {i , 1} ^ {k}}{N _ {i} ^ {k}}, \frac {N _ {i , 2} ^ {k}}{N _ {i} ^ {k}}, \dots , \frac {N _ {i , n} ^ {k}}{N _ {i} ^ {k}}\right). \tag {9} +$$ + +$N_{i}^{k}$ represents the total number of visible points of the $i$ -th superpoint in the $k$ -th image, and $N_{i,j}^{k}$ denotes the number of points belonging to the mask label with a value of $j$ , $n$ is the total number of 2D mask proposals. So, we denote the purity edge weight between the $i$ -th superpoint and the $j$ -th superpoint as $\Phi_{i,j}^{k}$ , which can be calculated as + +$$ +\Phi_ {i, j} ^ {k} = L _ {i} ^ {k} \cdot L _ {j} ^ {k}. \tag {10} +$$ + +Finally, we combine these three weights to form the mask consistency weight as + +$$ +W _ {i, j} ^ {k} = \varphi_ {i, j} ^ {k} \cdot \Phi_ {i, j} ^ {k} \cdot \psi_ {i, j} ^ {k}. \tag {11} +$$ + +Thus, the final affinity score is calculated as + +$$ +A _ {i, j} = \frac {\sum_ {k = 1} ^ {M} \left(W _ {i , j} ^ {k} A _ {i , j} ^ {k}\right)}{\sum_ {k = 1} ^ {M} W _ {i , j} ^ {k}}. \tag {12} +$$ + +# 3.4. Iterative graph clustering + +Building upon the superpoint graph, we introduce an iterative graph clustering technique to merge superpoints and update the graph. In each iteration, we employ a region growing method to merge superpoints, considering not only the affinity scores but also weighting based on the distance between superpoints. From a geometric perspective, superpoints that are far apart are less likely to be similar. However, once projected, they may appear closer on images from specific viewpoints. According to Sec. 3.3.2, their 2D mask proposals tend to be of lower quality, resulting in incorrect affinity scores. So we set the edge weight decreases for superpoints that are farther apart. + +Due to the introduction of the Sec. 3.3.2, some superpoints covering low-quality 2D mask proposals have lower view consistency and affinity scores. These superpoints may be located at the junction of two objects. Therefore, in each iteration, we set different thresholds for merging superpoints based on the affinity scores. We initially merge small and adjacent regions with high threshold, gradually lowering the threshold over iterations to merge superpoints with lower confidence. This iterative graph clustering approach allows for hierarchical merging of superpoints with varying levels of view consistency. Superpoints with stronger view consistency are prioritized for merging. Subsequently, superpoints with lower view consistency are merged, ultimately achieving robust 3D segmentation. + +# 4. Experiments + +We evaluate our SAM2Object on multiple datasets to demonstrate its effectiveness and robustness in zero-shot 3D instance segmentation. We also compare our SAM2Object with leading methods in the open-vocabulary segmentation field and the state-of-the-art closed-vocabulary segmentation methods which require training on annotated datasets. + +# 4.1. Experiment settings + +Datasets. ScanNetV2 [5] is a large-scale dataset tailored for indoor scene understanding and segmentation. It comprises RGB and depth images, as well as 3D point cloud data derived from indoor scenes. It includes over 2.5 million extensive views from more than 1,500 scans, along with semantic labels for 21 categories. ScanNet200 [36] is an extension of ScanNetV2, increasing the number of semantic categories to 200. Furthermore, ScanNet++ [48] is a newly released indoor dataset contains 280 indoor scenes and couples geometric and color data of high-quality, commodity-level indoor scenes. Each scene is captured by laser scanners with higher-resolution and a DSLR camera with a fisheye lens, as well as RGB-D streams obtained from iPhone 13 Pro. We evaluate our method on the well-studied ScanNetV2, ScanNet++ and ScanNet200 datasets. Following SAI3D [49], + +When generating 2D mask proposals, we use only $20\%$ of the images from the ScanNetV2 dataset and only $5\%$ from the ScanNet++ dataset. + +Table 1. Class-agnostic 3D instance segmentation on ScanNetV2 dataset. Bold and underlined values indicate the best and second best scores, respectively. + +
MethodTraining SetmAP\( AP_{50} \)\( AP_{25} \)
Closed-vocabulary
Mask3D [38]S3DIS31.144.958.0
Open-vocabulary
HDBSCAN [30]-1.65.532.1
UCIS [32]-2.37.330.5
EGIS [7]-5.012.738.9
Part2Object [39]-12.626.855.1
SAM-graph [11]-15.133.359.1
UnScene3D [37]-15.932.258.5
SAMPro3D [46]-18.032.856.3
SAM3D [47]-20.234.053.3
SAI3D [49]-30.850.570.6
SAM2Object (Ours)-34.052.770.3
+ +Evaluation metrics. We evaluate the quantitative results with the widely adopted Average Precision (AP) score. Following [11, 38, 42, 49], we calculate AP with IoU thresholds of $50\%$ and $25\%$ $(\mathrm{AP}_{50}, \mathrm{AP}_{25})$ and AP averaged with thresholds from $50\%$ to $95\%$ with a step size of $5\%$ (mAP). We implemented two distinct evaluation settings to comprehensively validate the performance of our method. Firstly, we evaluate class-agnostic instance segmentation which disregards category information and solely evaluates the accuracy of instance masks. Secondly, semantic instance segmentation builds on this by considering the semantic label of each instance, emphasizing the relationship between an instance and its category. + +Baselines. We compare our method with both open-vocabulary and closed-vocabulary baselines. For closed-vocabulary methods, we compare with the state-of-the-art method, Mask3D [38], which is a supervised model requiring training data. For open-vocabulary methods, we compare with the recent lifting methods based on leveraging SAM, including SAM3D [47], SAMPro3D [46], Segment3D [13], SAM-graph [11], SAI3D [49] and OVIR-3D [27]. We also compared with OpenMask3D [42], which generates masks using Mask3D and assigns semantic labels. Additionally, three traditional methods, referred to as HDBSCAN [30], EGIS [7] and UCIS [32], as well as the unsupervised methods UnScene3D [37] and Part2Object [39] are also included in our comparison. + +Table 2. Class-agnostic 3D instance segmentation on ScanNet++ dataset. Bold and underlined values indicate the best and second best scores, respectively. + +
MethodTraining SetmAP\( AP_{50} \)\( AP_{25} \)
Closed-vocabulary
Mask3DScanNetV29.917.325.8
Open-vocabulary
EGIS [7]-4.19.225.3
SAM3D [47]-7.214.229.4
Segment3D [13]-12.022.737.8
SAM-graph [11]-12.925.343.6
SAI3D [49]-17.131.149.5
SAM2Object (Ours)-20.234.148.7
+ +# 4.2. Results + +Quantitative results. We conduct the experiments on ScanNetV2, ScanNet++ and ScanNet200. Table 1 displays the quantitative results of our method on class-agnostic instance segmentation within ScanNetV2. Our method outperforms previous works in all metrics and leads in mAP and $\mathrm{AP}_{50}$ compared with the state-of-the-art methods SAI3D, trailing by only 0.3 in the $\mathrm{AP}_{25}$ metric. Notably, SAI3D requires adaptation of different 2D segmentation models for different datasets (using Semantic-SAM [22] on ScanNetV2 and HQ-SAM [17] on ScanNet++), whereas we consistently use SAM2 for validation across different datasets. Furthermore, our method outperforms both unsupervised and closed-vocabulary methods. The latter requires training data, so to conduct a fair zero-shot evaluation, we replace its training data with S3DIS [2] and validate on ScanNetV2. + +Table 2 reports the class-agnostic instance segmentation results on the challenging ScanNet++, we still lead in the mAP and $\mathrm{AP}_{50}$ metric. In lifting methods, the performance of the 2D segmentation model is crucial to the final 3D instance segmentation, SAI3D uses the optimal segmentation model for each dataset, while we consistently use SAM2. We simply replace the optimal segmentation model in SAI3D with SAM2, resulting in a mAP of $28.23\%$ on ScanNetV2 and $16.02\%$ on ScanNet++, both lower than SAI3D. This demonstrates the effectiveness of our method. + +Following OpenMask3D [42] and SAI3D [49], we validate the semantic instance segmentation results on ScanNet200, as shown in Table 3. We assign semantic labels to our segmented 3D instances using the method from OpenMask3D. We exhibit outstanding performance in both mAP and $\mathrm{AP}_{50}$ , indicating that our method demonstrates greater generalization when handling different semantic labels. + +Our method has lower $\mathrm{AP}_{25}$ than SAI3D across the three datasets. SAI3D lacks view consistency, resulting in fragmented and over-segmented outputs, which partially contributes to its higher $\mathrm{AP}_{25}$ . Additionally, SAI3D selects + +the most suitable visual foundation model for each dataset, achieving fine-grained compared with SAM2, which explains their higher $\mathrm{AP}_{25}$ . In contrast, we prefer to design a method capable of generalizing across multiple datasets. + +Table 3. Semantic instance segmentation on ScanNet200 dataset. "OV" and "Sup" represent "Open-vocabulary" and "Supervised" respectively. Bold and underlined values indicate the best and second best scores, respectively. + +
MethodmAP\( AP_{50} \)\( AP_{25} \)
Supmasked + OVsemantic OpenMask3D [42]15.419.923.1
OVmasked + OVsemantic
OVIR-3D [27]9.318.725.0
SAM3D [47]9.815.220.7
SAI3D [49]12.718.824.1
SAM2Object (Ours)13.319.023.8
+ +Qualitative results. We qualitatively compared our method with the SOTA method SAI3D on ScanNetV2 and ScanNet++. SAI3D is unable to handle the problem of projection errors, and the mismatched projections in Fig. 3 (c) and Fig. 2 result in incorrect 3D segmentation in Fig. 5 (a) and Fig. 5 (b). Moreover, SAI3D employs a frame-by-frame segmentation approach across multiple views, which lacks view consistency. This inconsistency in Fig. 1 leads to fragmented 3D segmentation in Fig. 5 (a), and similar fragmentation can be observed in Fig. 5 (c). Furthermore, the under-segmentation is also evident in Fig. 5 (d). + +In contrast, our approach consolidates view consistency and filters out low-quality 2D masks, enabling us to achieve comprehensive and robust 3D segmentation. More results can be found in the supplementary material. + +Table 4. Ablations of main components in our method. + +
Seg. & TrackMask consolidationmAPAP50AP25
28.244.765.2
✓29.347.666.0
✓32.149.968.3
✓✓34.052.770.3
+ +# 4.3. Ablation and analysis + +In Table 4, we performed ablation experiments on ScanNetV2 to verify the effectiveness of each component of our method. Starting from the baseline, similar to existing 2D lifting methods, we use the image predictor of SAM2 to perform frame-by-frame segmentation on multi-view images, setting the weight of all visible views equally. This approach attempts to achieve view consistency by averaging results across all visible views. However, since there is + +![](images/5d14d51dfb7d1993d178680d78d70fdf464523301c83c2ce610d907d2a63b29c.jpg) +Figure 5. Qualitative results comparisons between our method with SAI3D [49] on ScanNetV2 and ScanNet++. Due to projection errors resulting in incorrect 2D-to-3D matching relationships, SAI3D mistakenly segments part of the chair as the wall in (a), and segments the table and wall as a single object in (b). Due to the lack of view consistency, SAI3D exhibits fragmented segmentation in (c), as well as under-segmentation in (d). Our method shows the most robust and accurate class-agnostic instance segmentation results. + +no consistency between the masks of consecutive frames, it leads to fragmented segmentation. By applying mask consistency weighting through the mask consolidation module to results from multiple views, this strategy reduces low-quality mask interference during graph clustering (Sec. 3.4), resulting in a slight performance improvement. Employing the video predictor of SAM2 for object tracking ensures mask consistency across multi-view images. This results in more precise matching with 3D superpoints in Sec. 3.3, leading to improved performance. Finally, combining these two strategies fully leverages the benefits of consolidated view consistency, achieving further improved performance. + +# 4.4. Limitations + +Our method enhances view consistency by tracking objects across multiple views using SAM2 but requires continuous and dense multi-view images. Furthermore, although vision foundation models exhibit strong generalization, they are only capable of performing inference on 2D data and cannot fully understand scenes from a 3D scene. Although our approach connects adjacent frames, it is less effective than methods that integrate 3D priors with 2D semantics to establish view consistency. Future work should leverage 3D + +geometric priors to enhance visual foundation models in understanding 3D scenes from multi-view images, improving their generalization for 3D tasks. + +# 5. Conclusion + +In this paper, we introduce a novel zero-shot 3D instance segmentation method, SAM2Object. The key idea is to achieve more accurate 2D-to-3D matching by consolidating view consistency. We employ SAM2 to generate consistent masks across multiple views and introduce a metric to measure mask consistency. After constructing the superpoint graph, we perform iterative clustering based on view consistency levels, achieving more accurate segmentation. Comprehensive evaluations across multiple datasets demonstrate the effectiveness of SAM2Object, highlighting its ability to outperform previous methods and rival fully supervised techniques in zero-shot instance segmentation tasks. + +# 6. Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China: 62402033, 62227801 and 62376024. + +# References + +[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 1 +[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 7 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 2 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 3 +[6] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2 +[7] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 6, 7 +[8] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4 +[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2 +[10] Thomas Grandke. Interpolation algorithms for discrete fourier transforms of weighted signals. IEEE transactions on instrumentation and measurement, 32(2):350-355, 1983. 3 +[11] Haoyu Guo, He Zhu, Sida Peng, Yuang Wang, Yujun Shen, Ruizhen Hu, and Xiaowei Zhou. Sam-guided graph cut for 3d instance segmentation. In ECCV, 2024. 1, 3, 6, 7 +[12] Shuting He, Henghui Ding, and Wei Jiang. Semantic-promoted debiasing and background disambiguation for zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19498-19507, 2023. 2 +[13] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d + +segmentation without manual labels. In European Conference on Computer Vision, pages 278-295. Springer, 2025. 6, 7 +[14] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616, 2023. 1, 2 +[15] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2 +[16] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2 +[17] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 7 +[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2 +[20] Maksim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Top-down beats bottom-up in 3d instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3566-3574, 2024. 1 +[21] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 2 +[22] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 7 +[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2 +[24] Wenteng Liang, Feng Xue, Yihao Liu, Guofeng Zhong, and Anlong Ming. Unknown sniffer for object detection: Don't turn a blind eye to unknown objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3230-3239, 2023. 1 +[25] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF + +international conference on computer vision, pages 2783-2792, 2021. 1 +[26] Haizhuang Liu, Junbao Zhuo, Chen Liang, Jiansheng Chen, and Huimin Ma. Affinity3d: Propagating instance-level semantic affinity for zero-shot point cloud semantic segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9019-9028, 2024. 2 +[27] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data. In Conference on Robot Learning, pages 1610–1620. PMLR, 2023. 6, 7 +[28] Weijie Lyu, Xueting Li, Abhijit Kundu, Yi-Hsuan Tsai, and Ming-Hsuan Yang. Gaga: Group any gaussians via 3d-aware memory bank. arXiv preprint arXiv:2404.07977, 2024. 2 +[29] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. In Advances in Neural Information Processing Systems, pages 32340-32352. Curran Associates, Inc., 2022. 1 +[30] Leland McInnes and John Healy. Accelerated hierarchical density based clustering. In 2017 IEEE international conference on data mining workshops (ICDMW), pages 33-42. IEEE, 2017. 6 +[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[32] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters, 7(4):8713-8720, 2022. 6 +[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 1 +[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3 +[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 6, 1, 5 +[37] David Rozenberszki, Or Litany, and Angela Dai. Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19957-19967, 2024. 6 + +[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6 +[39] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6 +[40] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 2 +[41] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1 +[42] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 6, 7 +[43] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2 +[44] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635–3647, 2024. 2 +[45] Yizhou Wang, Longguang Wang, Qingyong Hu, Yan Liu, Ye Zhang, and Yulan Guo. Panoptic segmentation of 3d point clouds with gaussian mixture model in outdoor scenes. Visual Intelligence, 2(1):10, 2024. 1 +[46] Mutian Xu, Xingyilang Yin, Lingteng Qiu, Yang Liu, Xin Tong, and Xiaoguang Han. Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation. arXiv preprint arXiv:2311.17707, 2023. 1, 6 +[47] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 1, 3, 4, 6, 7, 2 +[48] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 6, 4 +[49] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3292–3302, 2024. 3, 4, 6, 7, 8, 2 \ No newline at end of file diff --git a/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/images.zip b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa3e2850db4004b55d047fca7b1f9440ad7471ad --- /dev/null +++ b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d93ebf0f0e5d0befea99f989cc558333291ba97e457a91841f3bd50478fba7d +size 630309 diff --git a/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/layout.json b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b52aab826053489c7e5a81f340601cf8df3e653 --- /dev/null +++ b/2025/SAM2Object_ Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation/layout.json @@ -0,0 +1,7819 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 102, + 496, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 102, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 102, + 496, + 140 + ], + "type": "text", + "content": "SAM2Object: Consolidating View Consistency via SAM2 for Zero-Shot 3D Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 161, + 458, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 161, + 458, + 190 + ], + "spans": [ + { + "bbox": [ + 156, + 161, + 458, + 190 + ], + "type": "text", + "content": "Jihuai Zhao Junbao Zhuo* Jiansheng Chen Huimin Ma* University of Science and Technology Beijing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 191, + 485, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 191, + 485, + 204 + ], + "spans": [ + { + "bbox": [ + 123, + 191, + 485, + 204 + ], + "type": "text", + "content": "jihuaizhao@xs.ustb.edu.cn, {junbaozhuo, jschen, mhmpub}@ustb.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "content": "In the field of zero-shot 3D instance segmentation, existing 2D-to-3D lifting methods typically obtain 2D segmentation across multiple RGB frames using vision foundation models, which are then projected and merged into 3D space. However, since the inference of vision foundation models on a single frame is not integrated with adjacent frames, the masks of the same object may vary across different frames, leading to a lack of view consistency in the 2D segmentation. Furthermore, current lifting methods average the 2D segmentation from multiple views during the projection into 3D space, causing low-quality masks and high-quality masks to share the same weight. These factors can lead to fragmented 3D segmentation. In this paper, we present SAM2Object, a novel zero-shot 3D instance segmentation method that effectively utilizes the Segment Anything Model 2 to segment and track objects, consolidating view consistency across frames. Our approach combines these consistent 2D masks with 3D geometric priors, improving the robustness of 3D segmentation. Additionally, we introduce mask consolidation module to filter out low-quality masks across frames, which enables more precise 2D-to-3D matching. Comprehensive evaluations on ScanNetV2, ScanNet++ and ScanNet200 demonstrate the robustness and effectiveness of SAM2Object, showcasing its ability to outperform previous methods. Our project page is at https://jihuaizhaohd.github.io/SAM2Object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 590, + 137, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 137, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 137, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "content": "3D instance segmentation plays a crucial role in various applications such as autonomous driving, robotics and augmented reality. Most of previous works [16, 20, 25, 41, 45] have predominantly relied on supervised learning, requiring high-quality 3D data and annotations for training. The annotation process for 3D data is often labor-intensive and costly, leading to the scarcity of high-quality 3D datasets." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 555, + 329 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 329 + ], + "type": "image", + "image_path": "ceac3aed7ab363987ac077132dca74ea9ea4a52e1e02d6d610e023403fe1bc8c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 337, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 381 + ], + "type": "text", + "content": "Figure 1. View Consistency. Compared with performing frame-by-frame segmentation on multi-view images, our method connects objects across multiple frames, ensuring that segmentation results remain consistent across various views." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 486 + ], + "type": "text", + "content": "This scarcity, in turn, greatly limits the development and advancement of 3D vision models, often leading to models with weak generalization capabilities, particularly in terms of zero-shot generalization. Zero-shot capability is essential, particularly in indoor navigation scenarios [1, 29]. For example, a robot misidentifying an unknown object, like a stove, could pose a fire hazard and endanger people." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 487, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 581 + ], + "type": "text", + "content": "The relative ease of acquiring 2D annotated data has led to the creation of numerous high-quality datasets, accelerating the advancement of 2D vision foundation models [3, 4, 19, 35]. Segment Anything Model (SAM) [19] is a prime example of this progress. SAM is trained on the extensive and high-quality SA-1B dataset [19], exhibiting outstanding zero-shot performance across various types of 2D vision tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Consequently, some methods [11, 14, 42, 46, 47] have attempted to transfer the robust generalization ability of SAM to 3D tasks by lifting its 2D segmentation into the 3D space. However, these methods often perform segmentation on a frame-by-frame manner, lacking inter-view consistency. Even slight changes in viewing angle can result in inconsistent 2D masks for the same object across different views. As shown in Fig. 1, the mask completely encapsulates the object in \"View 1\", but only partially covers it in \"View i\" and \"View n\". Therefore, conflicts can arise when merging 2D masks from different views, leading to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "19325" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "type": "text", + "content": "fragmented 3D segmentation and decreased performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 295, + 215 + ], + "type": "text", + "content": "Besides, existing work has not effectively addressed the errors introduced by long-distance projections in the 2D-to-3D matching process, as objects at different projection distances are assigned the same weight. Compared with the close-distance scene in Fig. 2 (a), the long-distance scene in Fig. 2 (b) erroneously projects the 3D points onto the \"wall\", causing a 3D object to incorrectly match with the 2D mask of another object. Therefore, the two originally separate objects are wrongly segmented as one. These errors increase as objects move farther from the camera, leading to incorrect 3D segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "text", + "content": "To address the inter-view inconsistency, in this work, we introduce SAM2Object, a novel approach that harnesses the strong generalization capability of SAM2 [35] and its multiview consistency. SAM2Object consolidate the consistency of 2D masks across multiple views and integrate 3D geometric priors to achieve superior 3D instance segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 395 + ], + "type": "text", + "content": "Additionally, to solve the problem of long-distance projection error, we propose a mask consolidation algorithm that applies weighting based on the distance of object from the camera across different views. During the 2D-to-3D matching process, the same object is prioritized to match with the view closer to the camera, thereby enhancing view consistency. The proposed SAM2Object conducts zero-shot 3D instance segmentation without the need for annotated data or additional training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 396, + 251, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 396, + 251, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 396, + 251, + 407 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 295, + 562 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "type": "text", + "content": "- We proposed SAM2Object, a novel zero-shot 3D instance segmentation method that leverages SAM2 to consolidate view consistency, effectively improving the robustness of 3D segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 502 + ], + "type": "text", + "content": "- We proposed a mask consistency metric to evaluate and maintain the 2D mask consistency across different views, enabling integration of 3D geometric priors and high-quality 2D masks for better 3D segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 503, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 562 + ], + "type": "text", + "content": "- We achieve more robust 3D instance segmentation on ScanNetV2 [5], ScanNet++ [48] and ScanNet200 [36]. Our method demonstrate superior performance compared with previous methods and outperform supervised models in zero-shot task." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 573, + 142, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 142, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 142, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "Closed-vocabulary 3D segmentation. 3D segmentation methods have rapidly advanced due to the emergence of high-quality 3D datasets [5, 48]. The task of 3D semantic segmentation [9, 23, 26, 33] distinguishes various categories of objects within a scene. While 3D instance segmentation [6, 16, 21, 40] is required to further differentiate between various instances within each category. Point-Group [16] is a CNN-based method that segments objects using both the original and offset-shifted point sets, which groups nearby points with the same label and pro" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 175 + ], + "type": "image", + "image_path": "afa3285739e91c2c417835b7cbbdd4b6d0d45fd9f5e5863bec96d3aa476a9771.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 183, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 261 + ], + "type": "text", + "content": "Figure 2. The quality of superpoint projection with different distances. In a scene, when an object is closer to the camera, the projection of 3D points onto the 2D image is more accurate in (a). However, if the object is farther from the camera in " + }, + { + "bbox": [ + 313, + 183, + 555, + 261 + ], + "type": "inline_equation", + "content": "(\\mathrm{b})\\sim (\\mathrm{d})" + }, + { + "bbox": [ + 313, + 183, + 555, + 261 + ], + "type": "text", + "content": ", projection errors are more likely to occur (the blue dots represent the 3D points projected onto the 2D image, colored masks correspond to projection targets)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 282, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 402 + ], + "type": "text", + "content": "gressively expands the group. Mask3D [38] introduced the first transformer-based 3D segmentation method, which directly predicts instance masks while avoiding manually selected voting schemes, achieving state-of-the-art performance. However, Mask3D still requires a substantial volume of 3D annotated data for training, which suffer poor performance on unseen data outside their domain-specific datasets. In this work, We directly leverage the zero-shot capability of SAM2 for 3D segmentation, eliminating the need for model training or 3D annotation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 403, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 653 + ], + "type": "text", + "content": "Open-vocabulary 3D segmentation. Recently, vision foundation models have achieved excellent performance in 2D understanding tasks [3, 4, 19, 35], leading to the emergence of notable work in open-vocabulary 2D segmentation tasks [12, 15, 43, 44]. Inspired by these efforts, some works has focused on exploring open-vocabulary capabilities in 3D tasks. OpenMask3D [42] and OpenIns3D [14] leverage Mask3D [38] to extract 3D instances in 3D space and then directly project them onto different 2D views for extracting open-vocabulary features. These methods suffer poor performance in recognizing small or detailed objects. LERF [18] integrates language embeddings into NeRF [31], by performing voxel rendering of CLIP [34] embeddings along training rays and supervising these embeddings across training views. LERF focuses on roughly locating specific objects through text prompts rather than achieving fine-grained 3D segmentation and has shown to be inadequate in complex scenes. In this study, we focus on exploring how to achieve better fine-grained 3D segmentation solely leveraging SAM2, without relying on any 3D pre-trained models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The Segment Anything Model. The Segment Anything Model (SAM) [19] has made significant breakthroughs in 2D segmentation tasks, demonstrating a strong generalization ability to segment any object in any image. The upgraded version of SAM, SAM2 [35], has further expanded" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19326" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 335 + ], + "type": "text", + "content": "into the domain of video segmentation. The image predictor of SAM2 remains consistent with SAM, while the video predictor is used to segment objects within videos. It is equipped with a memory that stores information about the object and previous interactions, enabling it to generate and correct predictions throughout the video. Recently, several 2D-to-3D lifting methods leverage SAM to perform frame-by-frame segmentation of multi-view images and project them onto 3D space. SAM3D [47] leads the way in extending SAM onto 3D segmentation by projecting segmentation data from 2D images onto 3D space. SAI3D [49] and SAM-graph [11] similarly adopts the approach of segmenting each frame independently, projecting the 2D segmentations onto superpoints and then merges superpoints using either region growing or Graph Neural Networks (GNNs). However, these methods do not link adjacent frames, potentially causing inconsistencies in 2D segmentation across consecutive frames, which can result in fragmented 3D segmentation. In this study, we effectively leverage SAM2 to establish connections across multi-view images, ensuring consistency in 2D segmentation between views, thereby achieving more robust 3D instance segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 347, + 117, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 117, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 117, + 359 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 367, + 280, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 280, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 280, + 379 + ], + "type": "text", + "content": "3.1. Problem formulation and method overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 553 + ], + "type": "text", + "content": "Given a series of posed color images " + }, + { + "bbox": [ + 55, + 384, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\{I_1, I_2, \\dots, I_T\\}" + }, + { + "bbox": [ + 55, + 384, + 296, + 553 + ], + "type": "text", + "content": " and the point cloud P, our goal is to output the 3D instance segmentation of the scene. In this paper, we propose a novel zero-shot 3D instance segmentation framework, as illustrated in Fig. 3. We first use SAM2 [35] to segment and track objects, ensuring that the masks accurately cover objects even as the viewpoint changes (Sec. 3.2), thereby consolidating multi-view consistency. Then, we combine 2D mask proposals with over-segmented 3D superpoints to construct a graph which is described in Sec. 3.3. In Sec. 3.4, we introduce an iterative graph clustering technique that hierarchically clusters different superpoints based on their view consistency levels, ultimately achieving 3D segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 572, + 163, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 163, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 163, + 585 + ], + "type": "text", + "content": "3.2. 2D mask proposal" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 590, + 176, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 176, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 176, + 602 + ], + "type": "text", + "content": "3.2.1. Keyframes extraction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We define keyframe as the frame where new objects appear. An illustration of the keyframe can be found in Fig. 4. In order to segment and track new objects, it is essential to extract the keyframes and perform segmentation and tracking on them. Given a set of color images, we can treat it as a video. When there is a scene change or a new object appears, the pixel differences between the previous frame " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{pre}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and the current frame " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{cur}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " tend to increase significantly. We use these frame differences to identify keyframes. The" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "frame differences " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "I_{dif}" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": " between two adjacent frames can be defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 389, + 93, + 555, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 93, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 389, + 93, + 555, + 119 + ], + "type": "interline_equation", + "content": "I _ {d i f} = \\frac {\\left| I _ {p r e} - I _ {c u r} \\right|}{I _ {h} \\cdot I _ {w}}. \\tag {1}", + "image_path": "72b1ec7819e3f29f6b2f1e7e442fbb2ac00646c5eb36e3f7a580fa76a7a7947f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "inline_equation", + "content": "I_h" + }, + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "inline_equation", + "content": "I_w" + }, + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "text", + "content": " represent the height and width of the color images, respectively. After extracting the differences between each pair of adjacent frames across the entire sequence of images, we apply a Hanning window [10] to smooth the frame difference array in order to mitigate the interference of noise. The window weight can be calculated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 203, + 555, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 203, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 324, + 203, + 555, + 230 + ], + "type": "interline_equation", + "content": "w _ {n} = \\left\\{ \\begin{array}{c c} 0. 5 - 0. 5 \\cos \\left(\\frac {2 \\pi n}{N - 1}\\right), & 0 \\leq n \\leq N - 1, \\\\ 0, & o t h e r w i s e. \\end{array} \\right. \\tag {2}", + "image_path": "99d084bc39e6220d4a7266053e405e49653fbd8e6248d0fe4a6c9a79b769e053.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": " is the window length, " + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": " represents the index within the window, and " + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": " denotes the value of the window function at position " + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": ". We define the smoothed array as " + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": "-th position in array " + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 239, + 554, + 287 + ], + "type": "text", + "content": " can be defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 295, + 555, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 295, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 370, + 295, + 555, + 328 + ], + "type": "interline_equation", + "content": "S _ {j} = \\frac {\\sum_ {n = 0} ^ {N - 1} w _ {n} \\cdot I _ {d i f} [ j - n ]}{\\sum_ {n = 0} ^ {N - 1} w _ {n}}. \\tag {3}", + "image_path": "19ffc1fe72efae4e386a09b5f50fb44b390030e9b176abf9fc5408b4b994df6f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "I_{dif}[j - n]" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " denotes the value at the " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "j - n" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " position in " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "I_{dif}" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": ". The indices corresponding to the local maxima in the smoothed array " + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 335, + 555, + 396 + ], + "type": "text", + "content": " represent the keyframes. We control the average distance between all keyframes to be around 100 frames by setting the window size appropriately." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 459, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 459, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 459, + 414 + ], + "type": "text", + "content": "3.2.2. Segmentation and tracking" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 416, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 668 + ], + "type": "text", + "content": "We employ SAM2 [35] for object segmentation and tracking to consolidate view consistency. SAM2 utilizes a sophisticated memory mechanism to track objects across frames, ensuring continuity even when objects are temporarily occluded or exit and re-enter the scene. Although SAM2's auto-segmentation technique does not require any points, boxes, or mask prompts, these prompts are necessary for object tracking. As shown in Fig. 3 (b), given a series of images and keyframes, we treat the image sequences between every two keyframes as video clips, with the first and last frames of the clip being the keyframe. We use the image predictor of SAM2 to automatically segment the keyframes and generate 2D masks for objects. These masks then serve as prompts for the video predictor of SAM2 to track the objects within the video clips. We segment new objects from the keyframes and track them in the video clips between key frames. Finally, at the edge of adjacent video clips, we link the same objects across different clips according to IoU metric. In the end, we obtain the segmentation and tracking results for all objects throughout the entire video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 674, + 438, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 438, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 438, + 686 + ], + "type": "text", + "content": "3.2.3. Merging segmentation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Even though we can segment keyframes to obtain new objects, keyframes are not always the exact frames where new" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19327" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 207, + 339 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 207, + 339 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 207, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 207, + 339 + ], + "type": "image", + "image_path": "671d5dceac898b074dbfb3dde64f8439eb1d3e887cfa3848b6cb94eeaa0b2e68.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 416 + ], + "type": "text", + "content": "Figure 3. Method overview. Our method consists of six steps. We over-segment the 3D mesh / point cloud into superpoints. (a) We extract keyframes from posed images. (b) We leverage SAM2 to segment the keyframes, and its segmentation serve as mask prompts for SAM2 to track objects throughout the entire set of posed images. (c) We consolidate the consistency of the masks obtained by SAM2 to filter out low-quality mask proposals. (d) We combine the masks with 3D geometric priors to construct a graph, where the nodes are superpoints and the edges represent the affinity scores between masks of superpoints that have undergone consistency consolidation. (e) We perform iterative graph clustering on the graph to achieve 3D instance segmentation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 210, + 70, + 548, + 340 + ], + "blocks": [ + { + "bbox": [ + 210, + 70, + 548, + 340 + ], + "lines": [ + { + "bbox": [ + 210, + 70, + 548, + 340 + ], + "spans": [ + { + "bbox": [ + 210, + 70, + 548, + 340 + ], + "type": "image", + "image_path": "4987034fe8843e3b0b0d2e331f325f27b8c7952ed73d46e1448e7dd3e2e695e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 428, + 294, + 616 + ], + "blocks": [ + { + "bbox": [ + 58, + 428, + 294, + 616 + ], + "lines": [ + { + "bbox": [ + 58, + 428, + 294, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 428, + 294, + 616 + ], + "type": "image", + "image_path": "c8c0d699823022c649db5ac5de128ce124e312cb8e1a318ec1239bf42daceb6c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 624, + 295, + 669 + ], + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 669 + ], + "type": "text", + "content": "Figure 4. Segmentation and tracking. (a) We extract keyframes from posed images. (b) We use the keyframes segmentation as mask prompts to guide SAM2 in tracking objects in both directions. (c) We merge the results to ensure that no objects are missed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "objects appear. If the new objects appear before the key frames, the new objects may be missed or overlooked. To address this issue, we adopt a bidirectional tracking strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 428, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 583 + ], + "type": "text", + "content": "(forward and backward). As shown in Fig. 4 (b), objects missed during forward tracking are identified in backward tracking (such as the \"trash can\" in the second column of images). Similarly, objects missed in backward tracking are captured in forward tracking (such as the \"door\" in the fourth column of images). The tracking results from the different directions complement each other. So we merge the 2D masks generated from both directions. An illustration of merging segmentation can be found in Fig. 4 (c). The segmentation we obtain maintain consistency across multiple views, which is crucial for subsequently combining 3D geometric priors to achieve more robust 3D instance segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 430, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 430, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 430, + 607 + ], + "type": "text", + "content": "3.3. Graph construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 613, + 471, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 471, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 471, + 625 + ], + "type": "text", + "content": "3.3.1. Building the superpoint graph" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", we represent the scene or point cloud as a graph " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " denotes the set of superpoints " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{S_i\\}_{i=1}^{N_s}" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " in the point cloud and " + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 628, + 555, + 713 + ], + "type": "text", + "content": " denotes the relationships between these superpoints. Following recent works [47, 49], we employ a graph-cut algorithm [8] to over-segment the whole point cloud based on 3D geometric attributes, resulting in a series of superpoints. Each" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19328" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "node in the graph represents the multiple points in the point cloud that share similar geometric features, and the number of superpoints is substantially lower than the number of points in the point cloud. By performing calculations at the superpoint-level, we achieve efficient computation while also extracting geometric priors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": "As shown in Fig. 3 (d), all color images from the different views are fed into SAM2 to obtain 2D masks, and we then project superpoints onto these 2D masks in all views to match superpoints with these masks. We use a common pinhole camera matrix to perform the 3D-to-2D projection, utilizing the corresponding camera intrinsic parameter " + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "inline_equation", + "content": "R_{k}" + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": " and pose parameter " + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "inline_equation", + "content": "E_{k}" + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": " to project the " + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": "-th 3D superpoint " + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": " onto the " + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": "-th image, and get its corresponding 2D mask proposals:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 129, + 251, + 295, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 251, + 295, + 265 + ], + "spans": [ + { + "bbox": [ + 129, + 251, + 295, + 265 + ], + "type": "interline_equation", + "content": "S _ {i, k} ^ {2 D} = \\phi \\left(R _ {k}, E _ {k}, S _ {i}\\right). \\tag {4}", + "image_path": "a9cd80b4b1582c870fc6cf5b29389c206633836823af4e73d5c775a9f933a52a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "text", + "content": " is the projection operator. The relationships between superpoints are represented by the affinity scores between their corresponding 2D masks, which are used to construct the edges of the graph. Specifically, we compute a normalized histogram based on the 2D mask proposals " + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "inline_equation", + "content": "S_{i,k}^{2D}" + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "text", + "content": " and convert it into a vector, denoted as " + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_i^k = \\text{vec(normalized}(S_{i,k}^{2D}))" + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "text", + "content": ". This histogram vector represents the 2D class-agnostic instance features of the 3D superpoint, and we get the affinity score between different superpoints on the " + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 270, + 296, + 403 + ], + "type": "text", + "content": "-th image by computing the cosine similarity of their 2D class-agnostic instance features:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 412, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 412, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 139, + 412, + 295, + 442 + ], + "type": "interline_equation", + "content": "A _ {i, j} ^ {k} = \\frac {\\mathbf {e} _ {i} ^ {k} \\cdot \\mathbf {e} _ {j} ^ {k}}{\\left| \\mathbf {e} _ {i} ^ {k} \\right| \\left| \\mathbf {e} _ {j} ^ {k} \\right|}. \\tag {5}", + "image_path": "6b0baab06827b7119d683e28f56bc1e036f8c609f7d29d06d51b0bd861c7f8a4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 167, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 167, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 167, + 457 + ], + "type": "text", + "content": "3.3.2. Mask consolidation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "type": "text", + "content": "Traversing all the images will yield the overall affinity score between " + }, + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 55, + 461, + 295, + 569 + ], + "type": "text", + "content": ", but the distribution of 2D mask proposals covered by the corresponding superpoint may vary due to changes in the shape of the same object from different views, as well as factors such as occlusion, motion blur, and projection errors. To address this issue, we introduce a mask consistency weight to filter out low-quality mask proposals, thereby clarifying the relationships between superpoints in the graph." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "content": "We calculate the visibility of 3D points on 2D images to address object occlusion and visibility issues. To obtain the visible points, we project the points from the point cloud onto the image, filter out the points that fall outside the image, and determine occlusion between objects based on their depth values. We define " + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "inline_equation", + "content": "V_{i}^{k}\\in (0,1)" + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "content": " as the ratio of the visible point count of " + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "content": " on the " + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "content": "-th image to its total point count in the point cloud. Thus, we denote the visibility edge weight between the " + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "content": "-th superpoint and the " + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 570, + 296, + 687 + ], + "type": "text", + "content": "-th superpoint as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 141, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\varphi_ {i, j} ^ {k} = V _ {i} ^ {k} \\cdot V _ {j} ^ {k}. \\tag {6}", + "image_path": "91d48496d6a3da89f5f3016d23043f3c1f0ce3ebeb9e99d98adb66146b1410a0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "The projection accuracy of the same superpoint in the point cloud can differ across various views. Projections in views closer to the camera tend to be more accurate than those in views farther from the camera. These errors can lead to mismatches between the 2D mask proposals " + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "inline_equation", + "content": "S_{i,k}^{2D}" + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": " and the superpoint " + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": " (as shown in Fig. 2, the superpoint that should have been matched to the \"table\" ended up being matched to the \"wall\"), which in turn leads to errors in the segmentation results. In addition, objects that are farther away from the camera appear smaller in the image, which also makes it easy to generate inaccurate 2D masks. We define " + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "inline_equation", + "content": "D_{i}^{k} \\in (0,1)" + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": " as the distance weight to perform weighted averaging of the 2D mask proposals " + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "inline_equation", + "content": "S_{i,k}^{2D}" + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": " based on the distance from the " + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "-th camera:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 243, + 555, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 243, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 367, + 243, + 555, + 273 + ], + "type": "interline_equation", + "content": "D _ {i} ^ {k} = 0. 9 \\cdot \\frac {d _ {m a x} ^ {k} - d _ {i} ^ {k}}{d _ {m a x} ^ {k} - d _ {m i n} ^ {k}} + 0. 1, \\tag {7}", + "image_path": "5b94477133db6cc723d918b47477a779215c773673cfe937394f76e515a6fb4e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "d_i^k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " represents the distance of visible points from the " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "-th camera within " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "S_{i,k}^{2D}" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "d_{max}^k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "d_{min}^k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " denotes the maximum and minimum distances of all visible points from the " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "-th camera. So, we denote the distance edge weight between the " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "-th superpoint and the " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "-th superpoint as " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\psi_{i,j}^k" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ", which can be calculated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 360, + 555, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 360, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 399, + 360, + 555, + 376 + ], + "type": "interline_equation", + "content": "\\psi_ {i, j} ^ {k} = D _ {i} ^ {k} \\cdot D _ {j} ^ {k}. \\tag {8}", + "image_path": "d0d2151b5527e2cea9dd1e04cf71816dec450a68c62db9f33053731491c1f055.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "text", + "content": "As shown in Fig. 1, the number of 2D mask proposals " + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "inline_equation", + "content": "S_{i,k}^{2D}" + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "text", + "content": " covered by the corresponding superpoint is defined as its purity. If a superpoint covers more 2D mask proposals, it is considered a difficult sample, and we reduce its weight and merge it only during the later stages of graph clustering in Sec. 3.4. We define " + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "inline_equation", + "content": "L_i^k \\in (0,1)" + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "text", + "content": " as the maximum ratio of all points within a superpoint projected onto different 2D mask proposals of " + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 378, + 554, + 474 + ], + "type": "text", + "content": "-th image:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 479, + 554, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 479, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 361, + 479, + 554, + 510 + ], + "type": "interline_equation", + "content": "L _ {i} ^ {k} = \\max \\left(\\frac {N _ {i , 1} ^ {k}}{N _ {i} ^ {k}}, \\frac {N _ {i , 2} ^ {k}}{N _ {i} ^ {k}}, \\dots , \\frac {N _ {i , n} ^ {k}}{N _ {i} ^ {k}}\\right). \\tag {9}", + "image_path": "bdbb885d9d44cdefcf830793f2893ec5c4ab51c2877266e9b36668eac3217d92.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "N_{i}^{k}" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": " represents the total number of visible points of the " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": "-th superpoint in the " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": "-th image, and " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "N_{i,j}^{k}" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": " denotes the number of points belonging to the mask label with a value of " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": " is the total number of 2D mask proposals. So, we denote the purity edge weight between the " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": "-th superpoint and the " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": "-th superpoint as " + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\Phi_{i,j}^{k}" + }, + { + "bbox": [ + 313, + 514, + 555, + 589 + ], + "type": "text", + "content": ", which can be calculated as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 400, + 597, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 597, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 400, + 597, + 553, + 613 + ], + "type": "interline_equation", + "content": "\\Phi_ {i, j} ^ {k} = L _ {i} ^ {k} \\cdot L _ {j} ^ {k}. \\tag {10}", + "image_path": "065fa96fce15268d90c6618e4ff880b61da5185bed149ae5f94ca00b913f45c2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "type": "text", + "content": "Finally, we combine these three weights to form the mask consistency weight as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 644, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 644, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 383, + 644, + 553, + 661 + ], + "type": "interline_equation", + "content": "W _ {i, j} ^ {k} = \\varphi_ {i, j} ^ {k} \\cdot \\Phi_ {i, j} ^ {k} \\cdot \\psi_ {i, j} ^ {k}. \\tag {11}", + "image_path": "658460295da0187e8842c94722ebba2e0261569892ab84ed1f6c2f239ce29b99.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 666, + 492, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 492, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 492, + 677 + ], + "type": "text", + "content": "Thus, the final affinity score is calculated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 380, + 684, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 684, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 380, + 684, + 553, + 717 + ], + "type": "interline_equation", + "content": "A _ {i, j} = \\frac {\\sum_ {k = 1} ^ {M} \\left(W _ {i , j} ^ {k} A _ {i , j} ^ {k}\\right)}{\\sum_ {k = 1} ^ {M} W _ {i , j} ^ {k}}. \\tag {12}", + "image_path": "cd456197ac2ae1b00df3fa80971a175cb03dd2f3c4423cce55b6c27526678002.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19329" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "type": "text", + "content": "3.4. Iterative graph clustering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 294, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 294, + 233 + ], + "type": "text", + "content": "Building upon the superpoint graph, we introduce an iterative graph clustering technique to merge superpoints and update the graph. In each iteration, we employ a region growing method to merge superpoints, considering not only the affinity scores but also weighting based on the distance between superpoints. From a geometric perspective, superpoints that are far apart are less likely to be similar. However, once projected, they may appear closer on images from specific viewpoints. According to Sec. 3.3.2, their 2D mask proposals tend to be of lower quality, resulting in incorrect affinity scores. So we set the edge weight decreases for superpoints that are farther apart." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 402 + ], + "type": "text", + "content": "Due to the introduction of the Sec. 3.3.2, some superpoints covering low-quality 2D mask proposals have lower view consistency and affinity scores. These superpoints may be located at the junction of two objects. Therefore, in each iteration, we set different thresholds for merging superpoints based on the affinity scores. We initially merge small and adjacent regions with high threshold, gradually lowering the threshold over iterations to merge superpoints with lower confidence. This iterative graph clustering approach allows for hierarchical merging of superpoints with varying levels of view consistency. Superpoints with stronger view consistency are prioritized for merging. Subsequently, superpoints with lower view consistency are merged, ultimately achieving robust 3D segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 414, + 136, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 136, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 136, + 427 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 507 + ], + "type": "text", + "content": "We evaluate our SAM2Object on multiple datasets to demonstrate its effectiveness and robustness in zero-shot 3D instance segmentation. We also compare our SAM2Object with leading methods in the open-vocabulary segmentation field and the state-of-the-art closed-vocabulary segmentation methods which require training on annotated datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 516, + 171, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 171, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 171, + 529 + ], + "type": "text", + "content": "4.1. Experiment settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Datasets. ScanNetV2 [5] is a large-scale dataset tailored for indoor scene understanding and segmentation. It comprises RGB and depth images, as well as 3D point cloud data derived from indoor scenes. It includes over 2.5 million extensive views from more than 1,500 scans, along with semantic labels for 21 categories. ScanNet200 [36] is an extension of ScanNetV2, increasing the number of semantic categories to 200. Furthermore, ScanNet++ [48] is a newly released indoor dataset contains 280 indoor scenes and couples geometric and color data of high-quality, commodity-level indoor scenes. Each scene is captured by laser scanners with higher-resolution and a DSLR camera with a fisheye lens, as well as RGB-D streams obtained from iPhone 13 Pro. We evaluate our method on the well-studied ScanNetV2, ScanNet++ and ScanNet200 datasets. Following SAI3D [49]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "When generating 2D mask proposals, we use only " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " of the images from the ScanNetV2 dataset and only " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " from the ScanNet++ dataset." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 175, + 553, + 361 + ], + "blocks": [ + { + "bbox": [ + 313, + 133, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 166 + ], + "type": "text", + "content": "Table 1. Class-agnostic 3D instance segmentation on ScanNetV2 dataset. Bold and underlined values indicate the best and second best scores, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 175, + 553, + 361 + ], + "lines": [ + { + "bbox": [ + 315, + 175, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 553, + 361 + ], + "type": "table", + "html": "
MethodTraining SetmAP\\( AP_{50} \\)\\( AP_{25} \\)
Closed-vocabulary
Mask3D [38]S3DIS31.144.958.0
Open-vocabulary
HDBSCAN [30]-1.65.532.1
UCIS [32]-2.37.330.5
EGIS [7]-5.012.738.9
Part2Object [39]-12.626.855.1
SAM-graph [11]-15.133.359.1
UnScene3D [37]-15.932.258.5
SAMPro3D [46]-18.032.856.3
SAM3D [47]-20.234.053.3
SAI3D [49]-30.850.570.6
SAM2Object (Ours)-34.052.770.3
", + "image_path": "1b0022115836474c132a6d27974e9f551b0d4954b0b4bf01c6c322a607e73145.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate the quantitative results with the widely adopted Average Precision (AP) score. Following [11, 38, 42, 49], we calculate AP with IoU thresholds of " + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{50}, \\mathrm{AP}_{25})" + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "content": " and AP averaged with thresholds from " + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "content": " with a step size of " + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 390, + 555, + 545 + ], + "type": "text", + "content": " (mAP). We implemented two distinct evaluation settings to comprehensively validate the performance of our method. Firstly, we evaluate class-agnostic instance segmentation which disregards category information and solely evaluates the accuracy of instance masks. Secondly, semantic instance segmentation builds on this by considering the semantic label of each instance, emphasizing the relationship between an instance and its category." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Baselines. We compare our method with both open-vocabulary and closed-vocabulary baselines. For closed-vocabulary methods, we compare with the state-of-the-art method, Mask3D [38], which is a supervised model requiring training data. For open-vocabulary methods, we compare with the recent lifting methods based on leveraging SAM, including SAM3D [47], SAMPro3D [46], Segment3D [13], SAM-graph [11], SAI3D [49] and OVIR-3D [27]. We also compared with OpenMask3D [42], which generates masks using Mask3D and assigns semantic labels. Additionally, three traditional methods, referred to as HDBSCAN [30], EGIS [7] and UCIS [32], as well as the unsupervised methods UnScene3D [37] and Part2Object [39] are also included in our comparison." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19330" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 294, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 2. Class-agnostic 3D instance segmentation on ScanNet++ dataset. Bold and underlined values indicate the best and second best scores, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 294, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 294, + 251 + ], + "type": "table", + "html": "
MethodTraining SetmAP\\( AP_{50} \\)\\( AP_{25} \\)
Closed-vocabulary
Mask3DScanNetV29.917.325.8
Open-vocabulary
EGIS [7]-4.19.225.3
SAM3D [47]-7.214.229.4
Segment3D [13]-12.022.737.8
SAM-graph [11]-12.925.343.6
SAI3D [49]-17.131.149.5
SAM2Object (Ours)-20.234.148.7
", + "image_path": "344987d868fc372fd83f4589512ccce7cd7cf2e0d7906e39aee1fc5da4314757.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 113, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 113, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 113, + 274 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "type": "text", + "content": "Quantitative results. We conduct the experiments on ScanNetV2, ScanNet++ and ScanNet200. Table 1 displays the quantitative results of our method on class-agnostic instance segmentation within ScanNetV2. Our method outperforms previous works in all metrics and leads in mAP and " + }, + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "type": "text", + "content": " compared with the state-of-the-art methods SAI3D, trailing by only 0.3 in the " + }, + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{25}" + }, + { + "bbox": [ + 54, + 281, + 295, + 460 + ], + "type": "text", + "content": " metric. Notably, SAI3D requires adaptation of different 2D segmentation models for different datasets (using Semantic-SAM [22] on ScanNetV2 and HQ-SAM [17] on ScanNet++), whereas we consistently use SAM2 for validation across different datasets. Furthermore, our method outperforms both unsupervised and closed-vocabulary methods. The latter requires training data, so to conduct a fair zero-shot evaluation, we replace its training data with S3DIS [2] and validate on ScanNetV2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "text", + "content": "Table 2 reports the class-agnostic instance segmentation results on the challenging ScanNet++, we still lead in the mAP and " + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "text", + "content": " metric. In lifting methods, the performance of the 2D segmentation model is crucial to the final 3D instance segmentation, SAI3D uses the optimal segmentation model for each dataset, while we consistently use SAM2. We simply replace the optimal segmentation model in SAI3D with SAM2, resulting in a mAP of " + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "28.23\\%" + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "text", + "content": " on ScanNetV2 and " + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "inline_equation", + "content": "16.02\\%" + }, + { + "bbox": [ + 54, + 461, + 295, + 581 + ], + "type": "text", + "content": " on ScanNet++, both lower than SAI3D. This demonstrates the effectiveness of our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "content": "Following OpenMask3D [42] and SAI3D [49], we validate the semantic instance segmentation results on ScanNet200, as shown in Table 3. We assign semantic labels to our segmented 3D instances using the method from OpenMask3D. We exhibit outstanding performance in both mAP and " + }, + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "content": ", indicating that our method demonstrates greater generalization when handling different semantic labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our method has lower " + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{25}" + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": " than SAI3D across the three datasets. SAI3D lacks view consistency, resulting in fragmented and over-segmented outputs, which partially contributes to its higher " + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{25}" + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Additionally, SAI3D selects" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": "the most suitable visual foundation model for each dataset, achieving fine-grained compared with SAM2, which explains their higher " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{25}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": ". In contrast, we prefer to design a method capable of generalizing across multiple datasets." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 320, + 184, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 313, + 130, + 554, + 175 + ], + "lines": [ + { + "bbox": [ + 313, + 130, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 554, + 175 + ], + "type": "text", + "content": "Table 3. Semantic instance segmentation on ScanNet200 dataset. \"OV\" and \"Sup\" represent \"Open-vocabulary\" and \"Supervised\" respectively. Bold and underlined values indicate the best and second best scores, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 184, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 320, + 184, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 545, + 297 + ], + "type": "table", + "html": "
MethodmAP\\( AP_{50} \\)\\( AP_{25} \\)
Supmasked + OVsemantic OpenMask3D [42]15.419.923.1
OVmasked + OVsemantic
OVIR-3D [27]9.318.725.0
SAM3D [47]9.815.220.7
SAI3D [49]12.718.824.1
SAM2Object (Ours)13.319.023.8
", + "image_path": "8a6536d3eeb886557943040dcd1f8449086af0e1462625b27c590da2988d805d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": "Qualitative results. We qualitatively compared our method with the SOTA method SAI3D on ScanNetV2 and ScanNet++. SAI3D is unable to handle the problem of projection errors, and the mismatched projections in Fig. 3 (c) and Fig. 2 result in incorrect 3D segmentation in Fig. 5 (a) and Fig. 5 (b). Moreover, SAI3D employs a frame-by-frame segmentation approach across multiple views, which lacks view consistency. This inconsistency in Fig. 1 leads to fragmented 3D segmentation in Fig. 5 (a), and similar fragmentation can be observed in Fig. 5 (c). Furthermore, the under-segmentation is also evident in Fig. 5 (d)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 485 + ], + "type": "text", + "content": "In contrast, our approach consolidates view consistency and filters out low-quality 2D masks, enabling us to achieve comprehensive and robust 3D segmentation. More results can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 516, + 550, + 590 + ], + "blocks": [ + { + "bbox": [ + 329, + 495, + 539, + 507 + ], + "lines": [ + { + "bbox": [ + 329, + 495, + 539, + 507 + ], + "spans": [ + { + "bbox": [ + 329, + 495, + 539, + 507 + ], + "type": "text", + "content": "Table 4. Ablations of main components in our method." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 516, + 550, + 590 + ], + "lines": [ + { + "bbox": [ + 316, + 516, + 550, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 550, + 590 + ], + "type": "table", + "html": "
Seg. & TrackMask consolidationmAPAP50AP25
28.244.765.2
✓29.347.666.0
✓32.149.968.3
✓✓34.052.770.3
", + "image_path": "aca2723b771b1c45b7955d73222bc74b5c3e24fed10f25cc207207948d040eec.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 440, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 440, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 440, + 613 + ], + "type": "text", + "content": "4.3. Ablation and analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "In Table 4, we performed ablation experiments on ScanNetV2 to verify the effectiveness of each component of our method. Starting from the baseline, similar to existing 2D lifting methods, we use the image predictor of SAM2 to perform frame-by-frame segmentation on multi-view images, setting the weight of all visible views equally. This approach attempts to achieve view consistency by averaging results across all visible views. However, since there is" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19331" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 373 + ], + "type": "image", + "image_path": "5d14d51dfb7d1993d178680d78d70fdf464523301c83c2ce610d907d2a63b29c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 380, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 54, + 380, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 555, + 426 + ], + "type": "text", + "content": "Figure 5. Qualitative results comparisons between our method with SAI3D [49] on ScanNetV2 and ScanNet++. Due to projection errors resulting in incorrect 2D-to-3D matching relationships, SAI3D mistakenly segments part of the chair as the wall in (a), and segments the table and wall as a single object in (b). Due to the lack of view consistency, SAI3D exhibits fragmented segmentation in (c), as well as under-segmentation in (d). Our method shows the most robust and accurate class-agnostic instance segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 433, + 298, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 298, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 298, + 578 + ], + "type": "text", + "content": "no consistency between the masks of consecutive frames, it leads to fragmented segmentation. By applying mask consistency weighting through the mask consolidation module to results from multiple views, this strategy reduces low-quality mask interference during graph clustering (Sec. 3.4), resulting in a slight performance improvement. Employing the video predictor of SAM2 for object tracking ensures mask consistency across multi-view images. This results in more precise matching with 3D superpoints in Sec. 3.3, leading to improved performance. Finally, combining these two strategies fully leverages the benefits of consolidated view consistency, achieving further improved performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 587, + 134, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 134, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 134, + 598 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 605, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 298, + 715 + ], + "type": "text", + "content": "Our method enhances view consistency by tracking objects across multiple views using SAM2 but requires continuous and dense multi-view images. Furthermore, although vision foundation models exhibit strong generalization, they are only capable of performing inference on 2D data and cannot fully understand scenes from a 3D scene. Although our approach connects adjacent frames, it is less effective than methods that integrate 3D priors with 2D semantics to establish view consistency. Future work should leverage 3D" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 433, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 469 + ], + "type": "text", + "content": "geometric priors to enhance visual foundation models in understanding 3D scenes from multi-view images, improving their generalization for 3D tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 481, + 388, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 388, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 388, + 493 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 501, + 556, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 556, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 556, + 646 + ], + "type": "text", + "content": "In this paper, we introduce a novel zero-shot 3D instance segmentation method, SAM2Object. The key idea is to achieve more accurate 2D-to-3D matching by consolidating view consistency. We employ SAM2 to generate consistent masks across multiple views and introduce a metric to measure mask consistency. After constructing the superpoint graph, we perform iterative clustering based on view consistency levels, achieving more accurate segmentation. Comprehensive evaluations across multiple datasets demonstrate the effectiveness of SAM2Object, highlighting its ability to outperform previous methods and rival fully supervised techniques in zero-shot instance segmentation tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 657, + 424, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 424, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 424, + 671 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China: 62402033, 62227801 and 62376024." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19332" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 368 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "text", + "content": "[6] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 296, + 456 + ], + "type": "text", + "content": "[7] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 490 + ], + "type": "text", + "content": "[8] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 545 + ], + "type": "text", + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[10] Thomas Grandke. Interpolation algorithms for discrete fourier transforms of weighted signals. IEEE transactions on instrumentation and measurement, 32(2):350-355, 1983. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "text", + "content": "[11] Haoyu Guo, He Zhu, Sida Peng, Yuang Wang, Yujun Shen, Ruizhen Hu, and Xiaowei Zhou. Sam-guided graph cut for 3d instance segmentation. In ECCV, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[12] Shuting He, Henghui Ding, and Wei Jiang. Semantic-promoted debiasing and background disambiguation for zero-shot instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19498-19507, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "segmentation without manual labels. In European Conference on Computer Vision, pages 278-295. Springer, 2025. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[14] Zhening Huang, Xiaoyang Wu, Xi Chen, Hengshuang Zhao, Lei Zhu, and Joan Lasenby. Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. arXiv preprint arXiv:2309.00616, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "type": "text", + "content": "[15] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 262 + ], + "type": "text", + "content": "[16] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[17] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "text", + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 365, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 421 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[20] Maksim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Top-down beats bottom-up in 3d instance segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3566-3574, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "text", + "content": "[21] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-attention-free transformer for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3693–3703, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "type": "text", + "content": "[22] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 579, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 621 + ], + "type": "text", + "content": "[23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 555, + 679 + ], + "type": "text", + "content": "[24] Wenteng Liang, Feng Xue, Yihao Liu, Guofeng Zhong, and Anlong Ming. Unknown sniffer for object detection: Don't turn a blind eye to unknown objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3230-3239, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19333" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "international conference on computer vision, pages 2783-2792, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Haizhuang Liu, Junbao Zhuo, Chen Liang, Jiansheng Chen, and Huimin Ma. Affinity3d: Propagating instance-level semantic affinity for zero-shot point cloud semantic segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9019-9028, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[27] Shiyang Lu, Haonan Chang, Eric Pu Jing, Abdeslam Boularias, and Kostas Bekris. Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data. In Conference on Robot Learning, pages 1610–1620. PMLR, 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 231 + ], + "type": "text", + "content": "[28] Weijie Lyu, Xueting Li, Abhijit Kundu, Yi-Hsuan Tsai, and Ming-Hsuan Yang. Gaga: Group any gaussians via 3d-aware memory bank. arXiv preprint arXiv:2404.07977, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "text", + "content": "[29] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. In Advances in Neural Information Processing Systems, pages 32340-32352. Curran Associates, Inc., 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 331 + ], + "type": "text", + "content": "[30] Leland McInnes and John Healy. Accelerated hierarchical density based clustering. In 2017 IEEE international conference on data mining workshops (ICDMW), pages 33-42. IEEE, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 386 + ], + "type": "text", + "content": "[31] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[32] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised class-agnostic instance segmentation of 3d lidar data for autonomous vehicles. IEEE Robotics and Automation Letters, 7(4):8713-8720, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 489 + ], + "type": "text", + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 555 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[36] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 2, 6, 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[37] David Rozenberszki, Or Litany, and Angela Dai. Unscene3d: Unsupervised 3d instance segmentation for indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19957-19967, 2024. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 687 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[38] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[39] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[40] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[41] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[42] Ayca Takmaz, Elisabetta Fedele, Robert Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 384 + ], + "type": "text", + "content": "[43] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 555, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 555, + 463 + ], + "type": "text", + "content": "[44] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635–3647, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "text", + "content": "[45] Yizhou Wang, Longguang Wang, Qingyong Hu, Yan Liu, Ye Zhang, and Yulan Guo. Panoptic segmentation of 3d point clouds with gaussian mixture model in outdoor scenes. Visual Intelligence, 2(1):10, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[46] Mutian Xu, Xingyilang Yin, Lingteng Qiu, Yang Liu, Xin Tong, and Xiaoguang Han. Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation. arXiv preprint arXiv:2311.17707, 2023. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 586 + ], + "type": "text", + "content": "[47] Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv preprint arXiv:2306.03908, 2023. 1, 3, 4, 6, 7, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "type": "text", + "content": "[48] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 6, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 632, + 553, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 553, + 687 + ], + "type": "text", + "content": "[49] Yingda Yin, Yuzheng Liu, Yang Xiao, Daniel Cohen-Or, Jingwei Huang, and Baoquan Chen. Sai3d: Segment any instance in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3292–3302, 2024. 3, 4, 6, 7, 8, 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19334" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_content_list.json b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f3a8891ff11ddefe8ce40b44bbd6de5bec32c46 --- /dev/null +++ b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity", + "text_level": 1, + "bbox": [ + 174, + 128, + 821, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengzhi Wu1", + "bbox": [ + 220, + 202, + 343, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyun Zhong", + "bbox": [ + 222, + 220, + 339, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxin Wan1*", + "bbox": [ + 383, + 203, + 488, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junwei Zheng $^{1\\dagger}$", + "bbox": [ + 361, + 220, + 488, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao $\\mathbf{F u}^{1*}$", + "bbox": [ + 522, + 203, + 596, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaming Zhang", + "bbox": [ + 503, + 220, + 630, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julius Pfrommer2", + "bbox": [ + 630, + 203, + 772, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jirgen Beyerer $^{1,2}$", + "bbox": [ + 651, + 220, + 792, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Karlsruhe Institute of Technology, Germany", + "bbox": [ + 181, + 244, + 540, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Fraunhofer IOSB, Germany", + "bbox": [ + 578, + 244, + 808, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chengzhi.wu, zeyun.zhong, junwei.zheng, jiangming.zhang}@kit.edu,", + "bbox": [ + 233, + 267, + 751, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuxin.wan, hao.fu}@student.kit.edu, {julius.pfrommer, juergen.beyerer}@iosb.fraunhofer.de", + "bbox": [ + 135, + 284, + 854, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 330, + 326, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Driven by the increasing demand for accurate and efficient representation of 3D data in various domains, point cloud sampling has emerged as a pivotal research topic in 3D computer vision. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on sharp edge details. Moreover, they all overlook the natural variations in point distribution across different shapes, applying a similar sampling strategy to all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes. SAMBLE effectively achieves an improved balance between sampling edge points for local details and preserving uniformity in the global shape, resulting in superior performance across multiple common point cloud downstream tasks, even in scenarios with few-point sampling.", + "bbox": [ + 88, + 361, + 485, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 686, + 220, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud sampling is a less explored research area within the realm of this data representation. Traditional random sampling (RS) and farthest point sampling (FPS) remain the most commonly employed methods when sampling is required for point cloud learning. With the advancement of neural networks, several methods have emerged for point cloud sampling in a downstream task-oriented learning framework, including S-Net [7], SampleNet [14], MOPS-Net [34], etc. However, these methods first generate a new, smaller-sized point cloud as a proxy rather than directly sampling points from the original input, rendering the", + "bbox": [ + 89, + 710, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16a07d5cb8a61a3c346fea446e86a7c46c17ebb8d56e938aed7cda5bd58c4b45.jpg", + "image_caption": [ + "Figure 1. Our method achieves an improved trade-off between sampling local details and preserving global uniformity." + ], + "image_footnote": [], + "bbox": [ + 516, + 327, + 910, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "techniques akin to black-box neural network models with limited interpretability. Consequently, discerning geometric patterns in their qualitative results becomes challenging, as their outcomes closely resemble those obtained through random sampling. More recently, APES [46] pioneers the direction of using neural networks to learn point-wise sampling scores, with which it subsequently samples points whose scores are higher. However, with its score computation design and the Top-M sampling strategy, APES excessively focuses on local details of edge points, resulting in a deficiency in preserving good global uniformity of the input shapes. Consequently, the interpolation operation becomes impractical during the upsampling process, and the sampling quality of few-point sampling is notably subpar (see Fig. 1). In this paper, we introduce a novel point cloud sampling method that addresses the limitations of prior approaches, aiming to achieve a refined balance between capturing local details and preserving global uniformity.", + "bbox": [ + 509, + 582, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The concept originates from rethinking the mathematical characteristics of local details within point cloud shapes. Typically, these local details are represented by edge points", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 107, + 886, + 377, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1342", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a42dd2e458f64fa44e2e9873c157dc9584f431b0abf426ffc333f1f4c0b0269.jpg", + "image_caption": [ + "Figure 2. When selecting an equal number of neighbors for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies (numbers on the right)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that define the shape's outline and sharpest features. Is there a point property that can easily distinguish between different categories, such as edge points and non-edge points? The answer is affirmative. In our investigation, we have uncovered an extremely fundamental yet crucial observation: if point $\\mathbf{p}_i$ is one of the $k$ -nearest neighbors of point $\\mathbf{p}_j$ , it does not necessarily imply that $\\mathbf{p}_j$ is also among the $k$ -nearest neighbors of $\\mathbf{p}_i$ . Consequently, it leads to the conclusion that the frequency of each point being chosen as a neighbor exhibits variation across a single point cloud.", + "bbox": [ + 88, + 262, + 482, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We explore and demonstrate the importance of this point property with a simple example as illustrated in Fig. 2. Assume the input point cloud is a simple grid. When selecting 5 neighbors for each point, all three possible cases are given on the left (center point is self-contained as a neighbor). Note that in the triangular and rectangular cases, they each has a \"quantum-entangled\" twin point pair, in which two points share the possibility of being chosen as the neighbor. While an equal number of neighbors is selected for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies, as listed on the right of Fig. 2. From it, we can observe that in addition to the edge point and non-edge point categories, there is also another noteworthy point category of close-to-edge points. Moreover, within each category, the points can be further grouped into more sub-categories. Overall, this point property effectively captures the local characteristics of a shape, especially for shape outline and sharp details. Building on it, we propose Sparse Attention Map (SAM) and introduce new methods for computing point-wise sampling scores to effectively balance the trade-off between local and global sampling. See more details in Sec. 3.2.", + "bbox": [ + 91, + 415, + 482, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, after the point-wise sampling scores are computed, previous methods employ a Top-M sampling strategy for all point cloud shapes, which exacerbates the issue of oversampling edge points. We argue that the naive top-M sampling strategy may not be optimal across all point cloud shapes for downstream tasks. For example, sampling more non-edge points enhances global uniformity, while sampling more close-to-edge points \"thickens\" the edge, both of which can potentially improve the performance on downstream tasks [46]. To address this, we introduce a", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "novel bin-based method to explore better sampling strategies shape-specifically by leveraging all point categories. This approach enables the sampling of points with smaller sampling scores, further optimizing the local-global tradeoff. As a result, our method dynamically adjusts the sampling strategy for each shape, leading to a more tailored and interpretable sampling process for improved performance.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 529, + 196, + 870, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a sparse attention map that directly integrates shape local and global information at the attention map level for point cloud sampling, introducing multiple methods for computing point-wise sampling scores.", + "- We present a novel approach for learning bin boundaries to partition points within individual shapes, enabling shape-specific sampling strategies by incorporating additional bin tokens during the attention computation.", + "- Our method successfully achieves an improved trade-off between capturing local details and preserving global uniformity for the sampling process, resulting in enhanced performance both qualitatively and quantitatively." + ], + "bbox": [ + 511, + 212, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 404, + 653, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point Cloud Sampling. Point cloud sampling is a key process in 3D data handling for simplifying high-resolution dense point clouds. Over the past decades, non-learning-based methods [8, 10, 26] have predominantly been used for point cloud sampling. While Farthest Point Sampling (FPS) [8] is the most widely used one [18, 31, 33, 49, 60], Random Sampling (RS) has also been frequently adopted [10, 32, 61]. More recently, learning-based sampling methods have shown superior performance with task-oriented training. S-Net [7] represents a pioneering work of generating new point coordinates from global representations, while SampleNet [14] introduces a soft projection operation for better point approximation. Following S-Net, multiple learning-based methods have been proposed [21, 28, 40, 41]. MOPS-Net [34] learns a transformation matrix and multiplies it with the original point cloud to generate the sampled one. By employing the attention mechanism to learn point-wise sampling scores, APES [46] captures the edge points in the input point clouds with a strong focus.", + "bbox": [ + 511, + 429, + 906, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Learning on Point Clouds. In contrast to the voxelization-based methods [12, 17, 25] and multi-view-based methods [2, 3, 16, 37], point-based methods deal directly with point clouds. The pioneer studies of PointNet [30] and PointNet++ [31] tackle point clouds through pointwise Multi-Layer Perceptrons (MLPs) and max-pooling operations. Subsequently, other research shifts focus towards constructing more efficient building blocks for local feature extraction, such as convolution-based ones [1, 18, 20, 38, 49, 50, 62] and graph-based ones [4, 15, 22, 24, 36, 42, 54, 58]. More recently, while MLP-based methods like PointNeXt [33] and PointMetaBase [19] have rekin", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1343", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/462afef2c30ba826d6a0784e920c27b971073e864edbe4ea303523083b525c3b.jpg", + "image_caption": [ + "Figure 3. A brief pipeline of our proposed method SAMBLE. It learns shape-specific sampling strategies for point cloud shapes." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dled people's interest, the application of attention mechanisms to point cloud analysis has also garnered widespread attention [9, 11, 39, 43, 45, 48, 57, 60]. For example, PT [51, 52, 60] series improve the model performance by introducing subtraction-based attention blocks, and [48] performs a large ablation study over attention module designs for point cloud processing. In addition, approaches that apply Transformers for point cloud self-supervised learning have also been proposed and explored [23, 29, 47, 57, 59].", + "bbox": [ + 89, + 308, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 457, + 225, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A brief pipeline of SAMBLE is illustrated in Fig. 3. It consists of three key steps: constructing a sparse attention map, computing point-wise sampling scores, and learning shape-specific sampling strategies through bin partitioning.", + "bbox": [ + 89, + 482, + 483, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sparse Attention Map", + "text_level": 1, + "bbox": [ + 89, + 551, + 297, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local and Global Attention Maps. Both local and global attention maps are widely used in point cloud analysis. A global attention map is derived from the application of classical self-attention to point features of all points, while a local attention map concentrates on a point-centered area wherein cross-attention is specifically applied to the central point and its neighbors.", + "bbox": [ + 89, + 573, + 483, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote $\\mathcal{S}_i$ as the set of $k$ -nearest neighbors of point $\\mathbf{p}_i$ , the local attention map for $\\mathbf{p}_i$ is defined as", + "bbox": [ + 89, + 679, + 482, + 709 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {i} ^ {l} = \\operatorname {s o f t m a x} \\left(Q (\\mathbf {p} _ {i}) K \\left(\\mathbf {p} _ {i j} - \\mathbf {p} _ {i}\\right) _ {j \\in \\mathcal {S} _ {i}} ^ {\\top} / \\sqrt {d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 714, + 482, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q$ and $K$ stand for the linear layers applied on the query and key input, and the square root of the feature dimension count $\\sqrt{d}$ serves as a scaling factor [39].", + "bbox": [ + 89, + 743, + 482, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the global attention map which is equivalent to taking all points as neighbors for each point, it is defined as", + "bbox": [ + 89, + 789, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {g} = \\operatorname {s o f t m a x} \\left(Q (\\mathbf {p} _ {i}) K (\\mathbf {p} _ {j}) _ {i, j \\in \\mathcal {S}} ^ {\\top} / \\sqrt {d}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 823, + 482, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S$ denotes the set of all input points.", + "bbox": [ + 89, + 852, + 374, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse Attention Map. Instead of using local or global attention maps solely, we propose sparse attention map,", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c10a1eaf649b8be3a09d1ddfe454ae27f25e8eda27df7a4609583cf5dca31d4.jpg", + "image_caption": [ + "Figure 4. Sparse attention map. In each row, $k$ cells are selected based on the KNN neighbor indexes for each point. The values of other non-selected cells are all set to 0. Note that the number of cells selected within each column is variable." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 893, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which combines the knowledge from both local and global information, to compute point-wise sampling scores. The idea is illustrated in Fig. 4. After obtaining the global attention map with Eq. 2, $k$ -NN is employed locally to find $k$ neighbors for each point. In this case, $k$ cells are being selected in each row. However, as discussed before, please notice that each point is chosen as a neighbor with varying frequencies. This means while for each row $k$ cells are selected, for each column, the number of selected cells varies. The selected cells are then \"carved\" out to form the sparse attention map, with the values of other non-selected cells being set to 0. More vividly, consider the global attention map as a grid stone slab of size $N \\times N$ . For carve-based SAM, values of all cells are pre-computed and hidden in the slab grid cells, and only the selected cells are carved out.", + "bbox": [ + 511, + 367, + 906, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Computing Point-wise Sampling Score", + "text_level": 1, + "bbox": [ + 511, + 604, + 844, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Indexing Mode. When sampling points, the points are indexed based on the computed point-wise sampling scores. We call the method of computing point-wise sampling scores from the full/sparse attention map as Indexing Mode. With the original full attention map, following APES, there are two possible indexing modes: (i) row standard deviation; and (ii) column sum. For a global attention map $\\mathbf{M}^g$ of size $N\\times N$ , denote $m_{ij}$ as the value of $i$ th row and $j$ th column in $\\mathbf{M}^g$ . These two indexing modes are formulated as modes (i) and (ii) in Tab. 1. To avoid possible confusion, we use notation $\\mathbf{p}_o$ to denote a point only in this subsection.", + "bbox": [ + 511, + 628, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the proposed sparse attention map, there are many other possible indexing modes. As discussed in Sec. 1, to achieve an improved trade-off between sampling edge points and preserving global uniformity, the frequency of each point being chosen as a neighbor, i.e., the number of selected cells in each column is the key. The following indexing modes are designed and explored for comparison:", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1344", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e2905a1e55f9d12452fc2400c93400f50a9134928df767ff20733dc85df2f39a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Indexing ModeAttention MapFormulaRemark
(i) Row standard deviationFullap0 = fstd{m0j|j = 1,2,...,N}fstd: Computing standard deviation for a set of values
(ii) Column sumFullap0 = ∑i=1N mio
(iii) Row standard deviationSparseap0 = fstd{msj|j ∈ So}So: Set of indices of selected cells in oth row
(iv) Row sumSparseap0 = ∑j=1N msjNon-selected cells are all of 0s
(v) Column sumSparseap0 = ∑i=1N msi0
(vi) Column averageSparseap0 = ∑i=1N msio/n0n0: Number of selected cells in oth column
(vii) Column square-dividedSparseap0 = ∑i=1N msio/n02n0: Number of selected cells in oth column
", + "bbox": [ + 109, + 88, + 885, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Proposed different indexing modes for computing point-wise sampling scores.", + "bbox": [ + 238, + 229, + 756, + 243 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b8fdc4e0f0630112c3f6176e6b8ab1097fc70150a7d4112f1a71acc928103360.jpg", + "image_caption": [ + "Figure 5. Point sampling score heatmaps under different indexing modes. Scores are normalized to $\\mathcal{N}(0.5,1)$ for better visualization." + ], + "image_footnote": [], + "bbox": [ + 106, + 258, + 888, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(iii) sparse row standard deviation; (iv) sparse row sum; (v) sparse column sum; (vi) sparse column average; and (vii) sparse column square-divided. Again, for a sparse attention map $\\mathbf{M}^s$ of size $N\\times N$ , denote $m_{ij}^{s}$ as the value of $i$ th row and $j$ th column in $\\mathbf{M}^s$ . For point $\\mathbf{p}_o$ , we denote the set of indexes of the selected $k$ cells (indexes of $k$ -nearest neighbors) in $o$ th row as $S_o$ , and denote the number of selected cells in $o$ th column as $n_o$ . Details and respective formulas of the proposed indexing modes are listed in Tab. 1.", + "bbox": [ + 88, + 500, + 482, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Heatmap. To analyze the behavior of each indexing mode, we train a separate model for each mode, ensuring that all other settings remain consistent. The sampling score distributions are visualized as heatmaps in Fig. 5, enhancing the interpretability of our method. From these heatmaps, we can see that both row-standard-deviation-based modes (modes i and iii) concentrate heavily on edge points. However, because they consistently prioritize thin or detailed regions, some areas may be overlooked. In contrast, modes ii and iv show less emphasis on edge points and instead distribute focus across a broader range of points, with a tendency toward other non-edge regions in a biased manner.", + "bbox": [ + 88, + 641, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More interestingly, the comparison of modes v, vi, and vii, which utilize column-wise information from SAM, reveals distinct sampling preferences and strategies across different point categories. Mode v prioritizes non-edge points, mode vi emphasizes the global shape, and mode", + "bbox": [ + 88, + 824, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vii focuses slightly more on edge points. This is because edge points typically have a smaller number of $n_o$ . Despite these differences and unique characteristics, all three modes capture the overall shape more uniformly compared to the former four. In our case, we aim to sample edge points without over-emphasizing them. For instance, when sampling detailed areas like chair legs, we want to capture some edge points without selecting them all, while also ensuring that non-edge points are sampled to preserve better global uniformity. Given this balance, we chose mode vii as the primary indexing mode for most of the experiments in the following sections. The detailed ablation study over different indexing modes is presented in Sec. 4.4.", + "bbox": [ + 511, + 500, + 906, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sampling with Bins", + "text_level": 1, + "bbox": [ + 511, + 708, + 700, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After point-wise sampling scores are computed with SAM, points are sampled according to certain rules. The simplest approach is top-M sampling, where points with the highest scores are sampled. In our case, as we aim to enhance the local-global trade-off and leverage all point categories during the sampling process, we suggest employing a bin-based sampling strategy to allow for the possible sampling of certain close-to-edge points or even non-edge points.", + "bbox": [ + 511, + 731, + 906, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bin Partitioning. The process begins with processing the distribution of normalized point-wise sampling scores $a_{\\mathbf{p}_i}$ across the shapes within the current batch. Let $n_b$ repre", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1345", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sent the number of bins used for partitioning, from which $n_b - 1$ bin boundary values need to be derived from the distribution. During each training step, a vector $\\nu_c = (\\nu_1, \\nu_2, \\dots, \\nu_{n_b - 1})$ is computed based on the point score distribution, ensuring an equitable division of points across all shapes in the current batch. Note that while $\\nu_c$ facilitates an even division at the batch level, the points within each individual shape are not necessarily evenly partitioned according to the batch-based bin boundary values.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training, for the first iteration, we directly use the boundary values derived from the first batch of data as the dynamic boundary values. Subsequently, since the second iteration, boundaries are updated adaptively in a momentum-based manner:", + "bbox": [ + 89, + 229, + 482, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\nu} _ {t} = \\gamma \\boldsymbol {\\nu} _ {t - 1} + (1 - \\gamma) \\boldsymbol {\\nu} _ {c}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 311, + 480, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nu_{t - 1}$ stands for the bin partitioning boundaries used in the last iteration, and $\\nu_{t}$ is the updated dynamic boundaries used for the current iteration. $\\gamma \\in (0,1)$ is the momentum update factor. With updated boundary values $\\nu_{t}$ , points in each shape are divided into $n_b$ subsets of $\\{\\mathcal{B}_1,\\mathcal{B}_2,\\dots ,\\mathcal{B}_{n_b}\\}$ based on their sampling scores.", + "bbox": [ + 89, + 332, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The core idea presented here is that, instead of using pre-engineered bin boundary values, we employ adaptive learning for bin boundaries, and they are gradually learned from the entire training dataset. These values are intended to evenly partition the distribution of point-wise sampling scores across all shapes and points in the training dataset. Consequently, for each individual shape, the acquired boundary values can effectively partition its points into bins with a shape-specific strategy, capturing the unique characteristics of the shape while maintaining a degree of proximity to other shapes within the dataset.", + "bbox": [ + 88, + 425, + 483, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tokens for Learning Bin Weights. With points already being partitioned into bins for each shape, the next step is to learn a shape-specific sampling strategy, i.e., to learn shape-specific sampling weights for each bin. Inspired by ViT [6], ViT [13], and Mask3D [35] —which leverage additional tokens during the computation of attention maps to extract and convey information across the entire feature map or specific groups of points or pixels — we introduce additional tokens specifically for learning bin sampling weights. In our case, attention maps are computed shape-specific during the downsampling process, facilitating the learning of bin sampling weights also in a shape-specific manner.", + "bbox": [ + 89, + 595, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the former proposed bin partitioning method, points in each shape are partitioned into $n_b$ subsets of $\\{\\mathcal{B}_1, \\mathcal{B}_2, \\dots, \\mathcal{B}_{n_b}\\}$ . The sampling weight $\\omega_j$ for bin $\\mathcal{B}_j (j = 1, 2, \\dots, n_b)$ is established based on the distinctive features of each shape. Fig. 6 gives the network structure of our proposed downsampling layer and illustrates the idea of using additional tokens. $n_b$ bin tokens are introduced during the attention computation, where each token corresponds to a", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bf79882e586b50f754d2fbdcdb06df97f3a7b6a6656f82b05bb01cb76aa2178.jpg", + "image_caption": [ + "Figure 6. Network structure of our proposed downsampling layer. Block $\\mathcal{A}$ : Points in each shape are partitioned into $n_b$ bins. Block $\\mathcal{B}$ : Masking the split-out point-to-token sub-attention map. Block $\\mathcal{C}$ : Learned bin sampling weights." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 733, + 376 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1418140f004eac1091bc0941f5abc2e6d88060e319a3ecdb4005f751aabad8f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 89, + 901, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "specific bin. As shown in Fig. 6, the bin tokens are initially concatenated with the input point-wise features for Key and Value. Subsequently, the combined features are subjected to a cross-attention mechanism with the original point-wise features as Query. The attention map is split into two parts of a point-to-point sub-attention map and a point-to-token sub-attention map. For the point-to-point attention map, the methods proposed in Sec. 3.1 and Sec. 3.2 are applied to it to obtain point-wise sampling scores. Note that in this case, the row-wise sum is not exactly equal to 1 but still very close to 1 since $n_b$ is of a very small quantity compared to $N$ . With computed point scores, dynamic boundary values $\\mathbf{v}_t$ are obtained for bin partitioning. Using the information regarding the allocation of points to respective bins, a mask operation is performed on the point-to-token sub-attention map as illustrated in Block B of Fig. 6. The sampling weights $\\omega_j$ are then subsequently acquired with", + "bbox": [ + 511, + 462, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {j} = \\operatorname {R e L U} \\left(\\frac {1}{\\beta_ {j}} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {B} _ {j}} m _ {\\mathbf {p} _ {i}, \\mathcal {B} _ {j}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 724, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta_{j}$ stands for the number of points in bin $\\mathcal{B}_j$ , and $m_{\\mathbf{p}_i,\\mathcal{B}_j}$ represents the element in the energy matrix corresponding to point $\\mathbf{p}_i$ in row and $\\mathcal{B}_j$ in column.", + "bbox": [ + 511, + 777, + 906, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In-Bin Point Sampling. For each shape, by considering the number of points contained within bins $\\beta = (\\beta_{1},\\beta_{2},\\ldots ,\\beta_{n_{b}})$ alongside the determined bin sampling weights $\\omega = (\\omega_{1},\\omega_{2},\\dots ,\\omega_{n_{b}})$ , the specific numbers of points to be sampled from each bin $\\kappa = (\\kappa_{1},\\kappa_{2},\\dots ,\\kappa_{n_{b}})$", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1346", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04e6a6edfeedc4ef67d6f258093d03c472da16ebc5780c957ab8eb6faf6d04c7.jpg", + "image_caption": [ + "Figure 7. Qualitative results of our proposed SAMBLE. Apart from the sampled results, sampling score heatmaps and bin histograms along with bin sampling ratios are also given. All shapes are from the test set. Zoom in for optimal visual clarity." + ], + "image_footnote": [], + "bbox": [ + 102, + 87, + 897, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "need to be determined. Direct multiplication of $\\beta$ and $\\omega$ does not yield a sum that aligns with the total number of down-sampled points $M$ required by the network structure. To address this discrepancy, a scaling method is applied to first scale bin sampling weights $\\omega_{j}$ . Furthermore, to prevent $\\kappa_{j}$ from surpassing the available point number $\\beta_{j}$ in any bin, any excess points are proportionately redistributed to other bins that have not been fully sampled. The detailed algorithm is presented in the supplementary materials.", + "bbox": [ + 88, + 320, + 480, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, within bin $\\mathcal{B}_j$ , $\\kappa_{j}$ points are selected through random sampling with priors. The sampling probability $\\rho_{\\mathbf{p}_i}$ is obtained via softmax operation over the normalized point sampling score $a_{\\mathbf{p}_i}$ with a temperature parameter $\\tau$ :", + "bbox": [ + 88, + 455, + 482, + 517 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\rho_ {\\mathbf {p} _ {i}} = \\frac {e ^ {a _ {\\mathbf {p} _ {i}} / \\tau}}{\\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {B} _ {j}} e ^ {a _ {\\mathbf {p} _ {i}} / \\tau}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 521, + 482, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 571, + 223, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Like most related works, such as S-NET [7], SampleNet [14], and APES [46], SAMBLE is specifically designed for point cloud shapes. To ensure a fair comparison, we conduct experiments using the same base network architecture as APES [46] on standard point cloud shape datasets, including ModelNet40 and ShapeNet-Part. It is important to note that point cloud sampling is not a standalone task; its effectiveness must be validated through downstream tasks.", + "bbox": [ + 88, + 597, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Classification", + "text_level": 1, + "bbox": [ + 89, + 727, + 230, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment Setting. ModelNet40 classification benchmark [53] includes 12,311 CAD models across 40 categories. For a fair comparison, we use the official train-test split, with 9,843 for training and 2,468 for testing. Points are uniformly sampled from the mesh surface and normalized to the unit sphere. Only 3D coordinates are used as input, with random scaling, rotation, and shifting applied for data augmentation. We use $n_b = 6$ bins for point partitioning. The momentum update factor $\\gamma = 0.99$ for updating bin boundary values. The temperature parameter $\\tau = 0.1$ .", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0ac59e965714554d93dbfc4650af35feb213ab46315fbd77a215393bc2ea6f9d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCls. OA (%)Seg.
Cat. mIoU (%)Ins. mIoU (%)
PointNet [30]89.280.483.7
PointNet++ [31]91.981.985.1
SpiderCNN [55]92.482.485.3
DGCNN [42]92.982.385.2
PointConv [49]92.582.885.7
PT1 [9]92.8-85.9
PT2 [60]93.783.786.6
PCT [11]93.2-86.4
PRA-Net [5]93.783.786.3
CurveNet [27]93.8-86.6
DeltaConv [44]93.8-86.6
PointNeXt [33]93.284.486.7
PointMetaBase [19]-84.386.7
APES (local) [46]93.583.185.6
APES (global) [46]93.883.785.8
SAMBLE94.284.586.7
", + "bbox": [ + 532, + 316, + 890, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Classification and segmentation results on the ModelNet40 and ShapeNet-Part benchmarks. In comparison with other SOTA methods that also only use raw point cloud data as input.", + "bbox": [ + 511, + 569, + 906, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative and Quantitative Results. Qualitative results of SAMBLE are presented in Fig. 7, including sampling score heatmaps, learned bin partitioning strategy with bin sampling ratios, and final sampled results. From it, we can observe that SAMBLE effectively samples sufficient edge points to capture the shape structure. Moreover, it maintains better global uniformity by avoiding excessive focus on edge points, particularly in those thin or sharp regions like chair legs. Logged shape bin histograms confirm the learning of shape-specific sampling strategies. More visualization results are provided in the supplementary materials, highlighting an intriguing pattern where shapes of the same category exhibit similar histogram distributions and sampling strategies. Overall, SAMBLE successfully achieves an improved trade-off between sampling edge points and preserving shape global uniformity. Quantitative results are provided in Tab. 2. Our method performs better than previous approaches and achieves state-of-the-art performance.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1347", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e3dd4e4ef9d27d6a8291912b297c5aa9817c11124f0f4783b423fde1932411a.jpg", + "image_caption": [ + "Figure 8. Segmentation results of our proposed SAMBLE in comparison with APES. All shapes are from the test set." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 486, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29278ad8cb1c14ef6b139995f649ebac3f2094f4b154848e29e7e8195e2df9b2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPointNeXtSAMBLE
Point Number2048102451220481024512
Cat. mIoU (%)84.4083.7982.7784.5184.8485.04
Ins. mIoU (%)86.7086.1885.1886.6786.9387.12
", + "bbox": [ + 112, + 361, + 460, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Additional segmentation performances evaluated on the intermediate downsampled sub-point clouds.", + "bbox": [ + 89, + 436, + 483, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Segmentation", + "text_level": 1, + "bbox": [ + 89, + 492, + 233, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiment setting. The ShapeNet-Part dataset [56] is used for 3D object part segmentation. It includes 16,880 3D models across 16 categories, with 14,006 models for training and 2,874 for testing. Each category contains 2-6 parts, totaling 50 distinct parts. We use the sampled point sets produced in [30] for a fair comparison with prior work. For evaluation metrics, we report both category mIoU and instance mIoU. We use $n_b = 4$ bins for point partitioning. The momentum update factor $\\gamma = 0.99$ for updating bin boundary values. The temperature parameter $\\tau = 0.1$ .", + "bbox": [ + 89, + 516, + 483, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative and Quantitative Results. Qualitative results are presented in Fig. 8, where we observe that, compared to APES, which tends to focus excessively on edge points, SAMBLE achieves a significantly improved balance between sampling edge points and preserving shape global uniformity. For example, SAMBLE demonstrates a more balanced use of non-edge points, as seen in the chair seat and lamp base, reflecting a thoughtful sampling strategy that accounts for different point categories and provides a more comprehensive representation of the overall shape. The quantitative results in Tab. 2 further highlight that SAMBLE achieves state-of-the-art performance.", + "bbox": [ + 89, + 672, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the part segmentation benchmark, we further report the performance on the intermediate downsampled subpoint clouds in Tab. 3. Additionally, results from PointNeXt", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edf37c3da05d789880d8cc624a33c280b3651673a443ef1789dfd2d35c68e30a.jpg", + "image_caption": [ + "Figure 9. Sampled results of few-point sampling in comparison with APES. Zoom in for optimal clarity." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[33] are also presented, which is a prominent point cloud learning method that employs FPS for downsampling. It is evident that FPS-based methods exhibit poorer performance when evaluated on intermediate downsampled subpoint clouds. In contrast, a notable observation is that SAMBLE achieves superior performance on intermediate downsampled points. This indicates that the learned sampled points contribute more to the overall performance, while the upsampling layer cannot fully reconstruct the features of the discarded points. Although SAMBLE enables the interpolation-based upsampling and it outperforms the upsampling layer used in APES (see details in Sec. 4.4), there remains potential for further improvement by designing a more meticulously crafted upsampling layer. However, this is another topic and beyond the scope of this paper.", + "bbox": [ + 511, + 368, + 906, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Few-Point Sampling", + "text_level": 1, + "bbox": [ + 511, + 604, + 705, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiment setting. We further compare our sampling method to previous approaches, including RS, FPS, and more recent learning-based methods such as S-Net, SampleNet, LightTN, APES, and others. The evaluation follows the same framework as in [7, 41, 46]. First, the point cloud is downsampled to a limited number of points, and the resulting subset is then fed into a task network for evaluation. For this comparison, we use the ModelNet40 classification task with the vanilla PointNet network. All sampling methods are evaluated across multiple sampling sizes of $M$ .", + "bbox": [ + 511, + 626, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative and Quantitative Results. Quantitative results are presented in Tab. 4. Note that APES [46] uses FPS to pre-process the input into $2M$ points while we do not. For a fair comparison, additional results of APES without the preprocessing step are also tested and reported. Nonetheless, even without pre-processing, SAMBLE achieves state-of-the-art results in the few-point sampling task as the number of sampled points decreases to extremely smaller ones.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1348", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1d630604f315e47f28ce83c0e1e8c8e66e13dd857030c4c486cc6585ab4d9e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MVoxelRSFPS [8]S-NET [7]PST-NET [40]SampleNet [14]MOPS-Net [34]DA-Net [21]LightTN [41]APES [46] (w/ pre-pro.)APES [46] (w/o pre-pro.)SAMBLE
51273.8287.5288.3487.8087.9488.1686.6789.0189.9190.8189.8190.58
25673.5077.0983.6482.3883.1584.2786.6386.2488.2190.4086.7890.18
12868.1556.4470.3477.5380.1180.7586.0685.6786.2689.7784.8790.02
6458.3131.6946.4270.4576.0679.8685.2585.5586.5189.5779.2389.81
3220.0216.3526.5860.7063.9277.3184.2885.1186.1888.5675.6389.45
", + "bbox": [ + 94, + 88, + 903, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6b08dc1c9efe62c2ce706ab0d7b3a28792b3984b5ecf2e6031b30151ae0cdf0f.jpg", + "table_caption": [ + "Table 4. Comparison with other sampling methods. Evaluated on the ModelNet40 classification benchmark with multiple sampling sizes. For APES, we additionally report its performance when pre-processing is not performed for a fair comparison." + ], + "table_footnote": [], + "table_body": "
Indexing Modeiiiiiiivvvivii
Cls.OA (%)93.9293.7893.6393.6693.4094.1194.08
Seg.Cat. mIoU (%)83.9883.8583.6283.5183.4784.1284.22
Ins. mIoU (%)86.1685.9985.7485.6085.4986.3886.46
", + "bbox": [ + 93, + 231, + 483, + 290 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/23fed35f2519d6752a38d4325e3e162bf29c94da11f66231f06355b674a1edae.jpg", + "table_caption": [ + "Table 5. Classification and segmentation performance with different indexing modes." + ], + "table_footnote": [], + "table_body": "
Number of Bins124681012
Cls.OA (%)93.9593.9193.9894.1894.0293.8093.84
Seg.Cat. mIoU (%)84.2284.1484.5184.4084.1983.9884.36
Ins. mIoU (%)86.4686.2886.6786.6186.4886.2386.43
", + "bbox": [ + 93, + 340, + 483, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results are presented in Fig. 9. For few-point sampling, APES relies on FPS to pre-sample the input into $2M$ points due to its limitations. In contrast, our method preserves better global uniformity, allowing direct few-point sampling from the input while still achieving satisfactory sampled results, as demonstrated in Fig. 9. When sampling very few points, APES tends to concentrate on the sharpest regions, whereas our SAmBLE method preserves better global uniformity throughout the point cloud shape.", + "bbox": [ + 88, + 455, + 482, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 601, + 243, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, our emphasis is directed toward the novel designs introduced within this paper, excluding common topics such as network width. More ablation studies and further design justifications are provided in the supplementary materials to enhance our method's interpretability.", + "bbox": [ + 89, + 623, + 482, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different Indexing Modes. Apart from the visualized heatmaps given in Fig. 5, we also report their respective experimental results in Tab. 5. The tests are performed using top-M as the sampling strategy. From it, we can observe that indexing modes vi and vii achieve best performances.", + "bbox": [ + 89, + 700, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Bins. As a key parameter in SAMBLE, an ablation study is performed over the number of bins $n_b$ . The results are presented in Tab. 6. Remarkably, increasing the number of bins does not yield improved performance. This phenomenon is likely attributable to the subdivision of shapes into an excessive number of point categories, leading to the gradual diminishment of score disparities across the bins. In our case, $n_b = 6$ and 4 yield the best performance", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/186fc3701a3cbe427de15843bd892601a6a7f6b3aea46debe8a81259d51973ab.jpg", + "table_caption": [ + "Table 6. Classification and segmentation performance with different number of bins." + ], + "table_footnote": [], + "table_body": "
UpsampleInterpolationCross-Attention
Kup=3Kup=8Kup=16
APES (local)82.89 / 85.4082.95 / 85.4482.96 / 85.4283.11 / 85.58
APES (global)83.16 / 85.5383.19 / 85.5983.17 / 85.5583.67 / 85.81
SAMBLE84.51 / 86.6784.35 / 86.4884.31 / 86.4384.36 / 86.44
", + "bbox": [ + 513, + 231, + 905, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Segmentation results with different upsampling layers on ShapeNet-Part. The number before “/” is the category mIoU, and the number after is the instance mIoU.", + "bbox": [ + 511, + 316, + 906, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for the classification and segmentation tasks respectively, and we use it for the corresponding experiments.", + "bbox": [ + 511, + 378, + 903, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Upsampling layer. An important aspect to highlight is the upsampling layer. Most point cloud network models employ neighbor-based interpolation [31, 33, 60] for upsampling, as FPS is used during the downsampling process. However, APES introduces a cross-attention layer for upsampling to address its limitations of overemphasizing edge points, which renders traditional neighbor-based interpolation impractical. In contrast, our method achieves an improved balance between sampling edge points and preserving global uniformity, allowing the use of interpolation operations during upsampling. An ablation study for evaluating various upsampling layers and interpolation with different $K_{up}$ values is conducted, and the results are presented in Table 7. The results show a performance drop for APES when interpolation is used in place of cross-attention, while SAMBLE demonstrates superior performance with it.", + "bbox": [ + 511, + 411, + 906, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 667, + 633, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, a novel point cloud sampling method SAMBLE is proposed to learn shape-specific sampling strategies for point cloud shapes. Based on a sparse attention map that integrates both local and global information, multiple indexing modes are designed and explored. By partitioning the points in each shape into bins and learning respective sampling ratios for each bin, shape-specific sampling strategies are acquired for individual point cloud shapes. SAMBLE achieves an optimal trade-off between sampling local details and preserving global uniformity, resulting in improved performance on downstream tasks. For future directions, advancements in upsampling layers could further improve the model's performance. Additionally, adapting the proposed method for point cloud scenes is another promising area to explore.", + "bbox": [ + 511, + 691, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1349", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pyunghwan Ahn, Juyoung Yang, Eojindl Yi, Chanho Lee, and Junmo Kim. Projection-based point convolution for efficient point cloud segmentation. IEEE Access, 10:15348-15358, 2022. 2", + "[2] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Semantic segmentation of earth observation data using multimodal and multi-scale deep networks. In *Asian Conference on Computer Vision*, pages 180–196. Springer, 2016. 2", + "[3] Alexandre Boulch, Bertrand Le Saux, and Nicolas Audebert. Unstructured point cloud semantic labeling using deep segmentation networks. 3DOR@ Eurographics, 3, 2017. 2", + "[4] Can Chen, Luca Zanotti Fragonara, and Antonios Tsourdos. GAPointNet: Graph attention based point neural network for exploiting local feature of point cloud. Neurocomputing, 438:122-132, 2021. 2", + "[5] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. PRA-Net: Point relation-aware network for 3d point cloud analysis. IEEE Transactions on Image Processing, 30: 4436-4448, 2021. 6", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2760-2769, 2019. 1, 2, 6, 7, 8", + "[8] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 8", + "[9] Nico Engel, Vasileios Belagiannis, and Klaus Dietmayer. Point transformer. IEEE Access, 9:134826-134840, 2021. 3, 6", + "[10] Fabian Groh, Patrick Wieschollek, and Hendrik PA Lensch. Flex-convolution: Million-scale point-cloud learning beyond grid-worlds. In Asian Conference on Computer Vision (ACCV), pages 105–122. Springer, 2018. 2", + "[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. PCT: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 6", + "[12] Mingyang Jiang, Yiran Wu, Tianqi Zhao, Zelin Zhao, and Cewu Lu. PointSift: A sift-like network module for 3d point cloud semantic segmentation. arXiv preprint arXiv:1807.00652, 2018. 2", + "[13] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5", + "[14] Itai Lang, Asaf Manor, and Shai Avidan. SampleNet: Differentiable point cloud sampling. In Proceedings of the" + ], + "bbox": [ + 93, + 143, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7578-7588, 2020. 1, 2, 6, 8", + "[15] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. DPC: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 2", + "[16] Felix Järemo Lawin, Martin Danelljan, Patrik Tosteberg, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Deep projective 3d semantic segmentation. In International Conference on Computer Analysis of Images and Patterns, pages 95-107. Springer, 2017. 2", + "[17] Truc Le and Ye Duan. PointGrid: A deep network for 3d shape understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9204-9214, 2018. 2", + "[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution on x-transformed points. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2", + "[19] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17682-17691, 2023. 2, 6", + "[20] Yiqun Lin, Zizheng Yan, Haibin Huang, Dong Du, Ligang Liu, Shuguang Cui, and Xiaoguang Han. FPCov: Learning local flattening for point convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4293-4302, 2020. 2", + "[21] Yanan Lin, Yan Huang, Shihao Zhou, Mengxi Jiang, Tianlong Wang, and Yunqi Lei. DA-Net: Density-adaptive down-sampling network for point cloud classification via end-to-end learning. In 2021 4th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), pages 13–18. IEEE, 2021. 2, 8", + "[22] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1800-1809, 2020. 2", + "[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision (ECCV), pages 657-675. Springer, 2022. 3", + "[24] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 2", + "[25] Daniel Maturana and Sebastian Scherer. VoxNet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2", + "[26] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 2", + "[27] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. CurveNet: Curvature-based multitask learning" + ], + "bbox": [ + 516, + 93, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1350", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6", + "[28] Ehsan Nezhadarya, Ehsan Taghavi, Ryan Razani, Bingbing Liu, and Jun Luo. Adaptive hierarchical down-sampling for point cloud classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12956-12964, 2020. 2", + "[29] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision (ECCV), pages 604-621. Springer, 2022. 3", + "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 2, 6, 7", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 6, 8", + "[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6329-6338, 2020. 2", + "[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems (NeurIPS), 35:23192-23204, 2022. 2, 6, 7, 8", + "[34] Yu Qian, Junhui Hou, Yiming Zeng, Qijian Zhang, Sam Tak Wu Kwong, and Ying He. MOPS-Net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. ArXiv, abs/2005.00383, 2020. 1, 2, 8", + "[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 5", + "[36] Martin Simonovsky and Nikos Komodakis. Dynamic edge-conditioned filters in convolutional neural networks on graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3693-3702, 2017. 2", + "[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3887-3896, 2018. 2", + "[38] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. KPCov: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3", + "[40] Xu Wang, Yi Jin, Yigang Cen, Congyan Lang, and Yidong Li. PST-Net: Point cloud sampling via point-based transformer. In 11th International Conference on Image and Graphics (ICIG), pages 57–69. Springer, 2021. 2, 8", + "[41] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. LighTN: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. IEEE Transactions on Multimedia, 2023. 2, 7, 8", + "[42] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2, 6", + "[43] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learnable skeleton-aware 3d point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, 2023. 3", + "[44] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. DeltaConv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 6", + "[45] Chengzhi Wu, Xuelei Bi, Julius Pfrommer, Alexander Cebulla, Simon Mangold, and Jürgen Beyerer. Sim2real transfer learning for point cloud segmentation: An industrial application case on autonomous disassembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4531-4540, 2023. 3", + "[46] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5333-5343, 2023. 1, 2, 6, 7, 8", + "[47] Chengzhi Wu, Qianliang Huang, Kun Jin, Julius Pfrommer, and Jürgen Beyerer. A cross branch fusion-based contrastive learning framework for point cloud self-supervised learning. In 2024 International Conference on 3D Vision (3DV), pages 528-538. IEEE, 2024. 3", + "[48] Chengzhi Wu, Kaige Wang, Zeyun Zhong, Hao Fu, Junwei Zheng, Jiaming Zhang, Julius Pfrommer, and Jurgen Beyerer. Rethinking attention module design for point cloud analysis. In International Conference on Pattern Recognition (ICPR), 2024. 3", + "[49] Wenxuan Wu, Zhongang Qi, and Li Fuxin. PointConv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9621-9630, 2019. 2, 6", + "[50] Wenxuan Wu, Li Fuxin, and Qi Shan. PointConvFormer: Revenge of the point-based convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21802-21813, 2023. 2", + "[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems (NeurIPS), 35:33330-33342, 2022. 3", + "[52] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1351", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4840-4851, 2024. 3", + "[53] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 6", + "[54] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-GCN for fast and scalable point cloud learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5661-5670, 2020. 2", + "[55] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102, 2018. 6", + "[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 7", + "[57] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 3", + "[58] Kuangen Zhang, Ming Hao, Jing Wang, Xinxing Chen, Yuquan Leng, Clarence W de Silva, and Chenglong Fu. Linked dynamic graph cnn: Learning through point cloud by linking hierarchical features. In 2021 27th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 7-12. IEEE, 2021. 2", + "[59] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in Neural Information Processing Systems (NeurIPS), 35:27061-27074, 2022. 3", + "[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 2, 3, 6, 8", + "[61] Yin Zhou and Oncel Tuzel. VoxelNet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4490-4499, 2018. 2", + "[62] Wei Zhu, Yue Ying, Jin Zhang, Xiuli Wang, and Yayu Zheng. Point cloud registration network based on convolution fusion and attention mechanism. Neural Processing Letters, pages 1-21, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 823 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1352", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_model.json b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f19561eb1c764b178c05a75c107771433c4296a6 --- /dev/null +++ b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.176, + 0.13, + 0.823, + 0.177 + ], + "angle": 0, + "content": "SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.344, + 0.219 + ], + "angle": 0, + "content": "Chengzhi Wu1" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.222, + 0.34, + 0.24 + ], + "angle": 0, + "content": "Zeyun Zhong" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.204, + 0.489, + 0.221 + ], + "angle": 0, + "content": "Yuxin Wan1*" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.222, + 0.49, + 0.24 + ], + "angle": 0, + "content": "Junwei Zheng\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.204, + 0.598, + 0.221 + ], + "angle": 0, + "content": "Hao \\(\\mathbf{F u}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.222, + 0.631, + 0.24 + ], + "angle": 0, + "content": "Jiaming Zhang" + }, + { + "type": "text", + "bbox": [ + 0.632, + 0.204, + 0.773, + 0.221 + ], + "angle": 0, + "content": "Julius Pfrommer2" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.222, + 0.793, + 0.24 + ], + "angle": 0, + "content": "Jirgen Beyerer\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.246, + 0.541, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\)Karlsruhe Institute of Technology, Germany" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.246, + 0.81, + 0.265 + ], + "angle": 0, + "content": "\\(^{2}\\)Fraunhofer IOSB, Germany" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.268, + 0.752, + 0.282 + ], + "angle": 0, + "content": "{chengzhi.wu, zeyun.zhong, junwei.zheng, jiangming.zhang}@kit.edu," + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.285, + 0.855, + 0.3 + ], + "angle": 0, + "content": "{yuxin.wan, hao.fu}@student.kit.edu, {julius.pfrommer, juergen.beyerer}@iosb.fraunhofer.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.331, + 0.327, + 0.347 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.486, + 0.665 + ], + "angle": 0, + "content": "Driven by the increasing demand for accurate and efficient representation of 3D data in various domains, point cloud sampling has emerged as a pivotal research topic in 3D computer vision. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on sharp edge details. Moreover, they all overlook the natural variations in point distribution across different shapes, applying a similar sampling strategy to all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes. SAMBLE effectively achieves an improved balance between sampling edge points for local details and preserving uniformity in the global shape, resulting in superior performance across multiple common point cloud downstream tasks, even in scenarios with few-point sampling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.222, + 0.702 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Point cloud sampling is a less explored research area within the realm of this data representation. Traditional random sampling (RS) and farthest point sampling (FPS) remain the most commonly employed methods when sampling is required for point cloud learning. With the advancement of neural networks, several methods have emerged for point cloud sampling in a downstream task-oriented learning framework, including S-Net [7], SampleNet [14], MOPS-Net [34], etc. However, these methods first generate a new, smaller-sized point cloud as a proxy rather than directly sampling points from the original input, rendering the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.328, + 0.911, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.534, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Figure 1. Our method achieves an improved trade-off between sampling local details and preserving global uniformity." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.907, + 0.854 + ], + "angle": 0, + "content": "techniques akin to black-box neural network models with limited interpretability. Consequently, discerning geometric patterns in their qualitative results becomes challenging, as their outcomes closely resemble those obtained through random sampling. More recently, APES [46] pioneers the direction of using neural networks to learn point-wise sampling scores, with which it subsequently samples points whose scores are higher. However, with its score computation design and the Top-M sampling strategy, APES excessively focuses on local details of edge points, resulting in a deficiency in preserving good global uniformity of the input shapes. Consequently, the interpolation operation becomes impractical during the upsampling process, and the sampling quality of few-point sampling is notably subpar (see Fig. 1). In this paper, we introduce a novel point cloud sampling method that addresses the limitations of prior approaches, aiming to achieve a refined balance between capturing local details and preserving global uniformity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The concept originates from rethinking the mathematical characteristics of local details within point cloud shapes. Typically, these local details are represented by edge points" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.378, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1342" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.194, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Figure 2. When selecting an equal number of neighbors for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies (numbers on the right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.263, + 0.483, + 0.414 + ], + "angle": 0, + "content": "that define the shape's outline and sharpest features. Is there a point property that can easily distinguish between different categories, such as edge points and non-edge points? The answer is affirmative. In our investigation, we have uncovered an extremely fundamental yet crucial observation: if point \\(\\mathbf{p}_i\\) is one of the \\(k\\)-nearest neighbors of point \\(\\mathbf{p}_j\\), it does not necessarily imply that \\(\\mathbf{p}_j\\) is also among the \\(k\\)-nearest neighbors of \\(\\mathbf{p}_i\\). Consequently, it leads to the conclusion that the frequency of each point being chosen as a neighbor exhibits variation across a single point cloud." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.416, + 0.483, + 0.748 + ], + "angle": 0, + "content": "We explore and demonstrate the importance of this point property with a simple example as illustrated in Fig. 2. Assume the input point cloud is a simple grid. When selecting 5 neighbors for each point, all three possible cases are given on the left (center point is self-contained as a neighbor). Note that in the triangular and rectangular cases, they each has a \"quantum-entangled\" twin point pair, in which two points share the possibility of being chosen as the neighbor. While an equal number of neighbors is selected for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies, as listed on the right of Fig. 2. From it, we can observe that in addition to the edge point and non-edge point categories, there is also another noteworthy point category of close-to-edge points. Moreover, within each category, the points can be further grouped into more sub-categories. Overall, this point property effectively captures the local characteristics of a shape, especially for shape outline and sharp details. Building on it, we propose Sparse Attention Map (SAM) and introduce new methods for computing point-wise sampling scores to effectively balance the trade-off between local and global sampling. See more details in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "On the other hand, after the point-wise sampling scores are computed, previous methods employ a Top-M sampling strategy for all point cloud shapes, which exacerbates the issue of oversampling edge points. We argue that the naive top-M sampling strategy may not be optimal across all point cloud shapes for downstream tasks. For example, sampling more non-edge points enhances global uniformity, while sampling more close-to-edge points \"thickens\" the edge, both of which can potentially improve the performance on downstream tasks [46]. To address this, we introduce a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "novel bin-based method to explore better sampling strategies shape-specifically by leveraging all point categories. This approach enables the sampling of points with smaller sampling scores, further optimizing the local-global tradeoff. As a result, our method dynamically adjusts the sampling strategy for each shape, leading to a more tailored and interpretable sampling process for improved performance." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.198, + 0.872, + 0.211 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- We propose a sparse attention map that directly integrates shape local and global information at the attention map level for point cloud sampling, introducing multiple methods for computing point-wise sampling scores." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.332 + ], + "angle": 0, + "content": "- We present a novel approach for learning bin boundaries to partition points within individual shapes, enabling shape-specific sampling strategies by incorporating additional bin tokens during the attention computation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.333, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- Our method successfully achieves an improved trade-off between capturing local details and preserving global uniformity for the sampling process, resulting in enhanced performance both qualitatively and quantitatively." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.405, + 0.655, + 0.421 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Point Cloud Sampling. Point cloud sampling is a key process in 3D data handling for simplifying high-resolution dense point clouds. Over the past decades, non-learning-based methods [8, 10, 26] have predominantly been used for point cloud sampling. While Farthest Point Sampling (FPS) [8] is the most widely used one [18, 31, 33, 49, 60], Random Sampling (RS) has also been frequently adopted [10, 32, 61]. More recently, learning-based sampling methods have shown superior performance with task-oriented training. S-Net [7] represents a pioneering work of generating new point coordinates from global representations, while SampleNet [14] introduces a soft projection operation for better point approximation. Following S-Net, multiple learning-based methods have been proposed [21, 28, 40, 41]. MOPS-Net [34] learns a transformation matrix and multiplies it with the original point cloud to generate the sampled one. By employing the attention mechanism to learn point-wise sampling scores, APES [46] captures the edge points in the input point clouds with a strong focus." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Deep Learning on Point Clouds. In contrast to the voxelization-based methods [12, 17, 25] and multi-view-based methods [2, 3, 16, 37], point-based methods deal directly with point clouds. The pioneer studies of PointNet [30] and PointNet++ [31] tackle point clouds through pointwise Multi-Layer Perceptrons (MLPs) and max-pooling operations. Subsequently, other research shifts focus towards constructing more efficient building blocks for local feature extraction, such as convolution-based ones [1, 18, 20, 38, 49, 50, 62] and graph-based ones [4, 15, 22, 24, 36, 42, 54, 58]. More recently, while MLP-based methods like PointNeXt [33] and PointMetaBase [19] have rekin" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1343" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.294 + ], + "angle": 0, + "content": "Figure 3. A brief pipeline of our proposed method SAMBLE. It learns shape-specific sampling strategies for point cloud shapes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.448 + ], + "angle": 0, + "content": "dled people's interest, the application of attention mechanisms to point cloud analysis has also garnered widespread attention [9, 11, 39, 43, 45, 48, 57, 60]. For example, PT [51, 52, 60] series improve the model performance by introducing subtraction-based attention blocks, and [48] performs a large ablation study over attention module designs for point cloud processing. In addition, approaches that apply Transformers for point cloud self-supervised learning have also been proposed and explored [23, 29, 47, 57, 59]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.458, + 0.226, + 0.475 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.484, + 0.545 + ], + "angle": 0, + "content": "A brief pipeline of SAMBLE is illustrated in Fig. 3. It consists of three key steps: constructing a sparse attention map, computing point-wise sampling scores, and learning shape-specific sampling strategies through bin partitioning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.299, + 0.569 + ], + "angle": 0, + "content": "3.1. Sparse Attention Map" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.679 + ], + "angle": 0, + "content": "Local and Global Attention Maps. Both local and global attention maps are widely used in point cloud analysis. A global attention map is derived from the application of classical self-attention to point features of all points, while a local attention map concentrates on a point-centered area wherein cross-attention is specifically applied to the central point and its neighbors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Denote \\( \\mathcal{S}_i \\) as the set of \\( k \\)-nearest neighbors of point \\( \\mathbf{p}_i \\), the local attention map for \\( \\mathbf{p}_i \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.715, + 0.483, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {i} ^ {l} = \\operatorname {s o f t m a x} \\left(Q (\\mathbf {p} _ {i}) K \\left(\\mathbf {p} _ {i j} - \\mathbf {p} _ {i}\\right) _ {j \\in \\mathcal {S} _ {i}} ^ {\\top} / \\sqrt {d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.789 + ], + "angle": 0, + "content": "where \\(Q\\) and \\(K\\) stand for the linear layers applied on the query and key input, and the square root of the feature dimension count \\(\\sqrt{d}\\) serves as a scaling factor [39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.483, + 0.82 + ], + "angle": 0, + "content": "For the global attention map which is equivalent to taking all points as neighbors for each point, it is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.824, + 0.483, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {g} = \\operatorname {s o f t m a x} \\left(Q (\\mathbf {p} _ {i}) K (\\mathbf {p} _ {j}) _ {i, j \\in \\mathcal {S}} ^ {\\top} / \\sqrt {d}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.375, + 0.869 + ], + "angle": 0, + "content": "where \\(S\\) denotes the set of all input points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Sparse Attention Map. Instead of using local or global attention maps solely, we propose sparse attention map," + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.091, + 0.895, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.291, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 4. Sparse attention map. In each row, \\( k \\) cells are selected based on the KNN neighbor indexes for each point. The values of other non-selected cells are all set to 0. Note that the number of cells selected within each column is variable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.596 + ], + "angle": 0, + "content": "which combines the knowledge from both local and global information, to compute point-wise sampling scores. The idea is illustrated in Fig. 4. After obtaining the global attention map with Eq. 2, \\( k \\)-NN is employed locally to find \\( k \\) neighbors for each point. In this case, \\( k \\) cells are being selected in each row. However, as discussed before, please notice that each point is chosen as a neighbor with varying frequencies. This means while for each row \\( k \\) cells are selected, for each column, the number of selected cells varies. The selected cells are then \"carved\" out to form the sparse attention map, with the values of other non-selected cells being set to 0. More vividly, consider the global attention map as a grid stone slab of size \\( N \\times N \\). For carve-based SAM, values of all cells are pre-computed and hidden in the slab grid cells, and only the selected cells are carved out." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.845, + 0.623 + ], + "angle": 0, + "content": "3.2. Computing Point-wise Sampling Score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Indexing Mode. When sampling points, the points are indexed based on the computed point-wise sampling scores. We call the method of computing point-wise sampling scores from the full/sparse attention map as Indexing Mode. With the original full attention map, following APES, there are two possible indexing modes: (i) row standard deviation; and (ii) column sum. For a global attention map \\(\\mathbf{M}^g\\) of size \\(N\\times N\\), denote \\(m_{ij}\\) as the value of \\(i\\)th row and \\(j\\)th column in \\(\\mathbf{M}^g\\). These two indexing modes are formulated as modes (i) and (ii) in Tab. 1. To avoid possible confusion, we use notation \\(\\mathbf{p}_o\\) to denote a point only in this subsection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "With the proposed sparse attention map, there are many other possible indexing modes. As discussed in Sec. 1, to achieve an improved trade-off between sampling edge points and preserving global uniformity, the frequency of each point being chosen as a neighbor, i.e., the number of selected cells in each column is the key. The following indexing modes are designed and explored for comparison:" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1344" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.887, + 0.22 + ], + "angle": 0, + "content": "
Indexing ModeAttention MapFormulaRemark
(i) Row standard deviationFullap0 = fstd{m0j|j = 1,2,...,N}fstd: Computing standard deviation for a set of values
(ii) Column sumFullap0 = ∑i=1N mio
(iii) Row standard deviationSparseap0 = fstd{msj|j ∈ So}So: Set of indices of selected cells in oth row
(iv) Row sumSparseap0 = ∑j=1N msjNon-selected cells are all of 0s
(v) Column sumSparseap0 = ∑i=1N msi0
(vi) Column averageSparseap0 = ∑i=1N msio/n0n0: Number of selected cells in oth column
(vii) Column square-dividedSparseap0 = ∑i=1N msio/n02n0: Number of selected cells in oth column
" + }, + { + "type": "table_caption", + "bbox": [ + 0.239, + 0.23, + 0.757, + 0.244 + ], + "angle": 0, + "content": "Table 1. Proposed different indexing modes for computing point-wise sampling scores." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.259, + 0.889, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.467, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Figure 5. Point sampling score heatmaps under different indexing modes. Scores are normalized to \\(\\mathcal{N}(0.5,1)\\) for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.501, + 0.483, + 0.637 + ], + "angle": 0, + "content": "(iii) sparse row standard deviation; (iv) sparse row sum; (v) sparse column sum; (vi) sparse column average; and (vii) sparse column square-divided. Again, for a sparse attention map \\(\\mathbf{M}^s\\) of size \\(N\\times N\\), denote \\(m_{ij}^{s}\\) as the value of \\(i\\)th row and \\(j\\)th column in \\(\\mathbf{M}^s\\). For point \\(\\mathbf{p}_o\\), we denote the set of indexes of the selected \\(k\\) cells (indexes of \\(k\\)-nearest neighbors) in \\(o\\)th row as \\(S_o\\), and denote the number of selected cells in \\(o\\)th column as \\(n_o\\). Details and respective formulas of the proposed indexing modes are listed in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Heatmap. To analyze the behavior of each indexing mode, we train a separate model for each mode, ensuring that all other settings remain consistent. The sampling score distributions are visualized as heatmaps in Fig. 5, enhancing the interpretability of our method. From these heatmaps, we can see that both row-standard-deviation-based modes (modes i and iii) concentrate heavily on edge points. However, because they consistently prioritize thin or detailed regions, some areas may be overlooked. In contrast, modes ii and iv show less emphasis on edge points and instead distribute focus across a broader range of points, with a tendency toward other non-edge regions in a biased manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "More interestingly, the comparison of modes v, vi, and vii, which utilize column-wise information from SAM, reveals distinct sampling preferences and strategies across different point categories. Mode v prioritizes non-edge points, mode vi emphasizes the global shape, and mode" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.698 + ], + "angle": 0, + "content": "vii focuses slightly more on edge points. This is because edge points typically have a smaller number of \\( n_o \\). Despite these differences and unique characteristics, all three modes capture the overall shape more uniformly compared to the former four. In our case, we aim to sample edge points without over-emphasizing them. For instance, when sampling detailed areas like chair legs, we want to capture some edge points without selecting them all, while also ensuring that non-edge points are sampled to preserve better global uniformity. Given this balance, we chose mode vii as the primary indexing mode for most of the experiments in the following sections. The detailed ablation study over different indexing modes is presented in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.702, + 0.725 + ], + "angle": 0, + "content": "3.3. Sampling with Bins" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.853 + ], + "angle": 0, + "content": "After point-wise sampling scores are computed with SAM, points are sampled according to certain rules. The simplest approach is top-M sampling, where points with the highest scores are sampled. In our case, as we aim to enhance the local-global trade-off and leverage all point categories during the sampling process, we suggest employing a bin-based sampling strategy to allow for the possible sampling of certain close-to-edge points or even non-edge points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Bin Partitioning. The process begins with processing the distribution of normalized point-wise sampling scores \\( a_{\\mathbf{p}_i} \\) across the shapes within the current batch. Let \\( n_b \\) repre" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1345" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "sent the number of bins used for partitioning, from which \\( n_b - 1 \\) bin boundary values need to be derived from the distribution. During each training step, a vector \\( \\nu_c = (\\nu_1, \\nu_2, \\dots, \\nu_{n_b - 1}) \\) is computed based on the point score distribution, ensuring an equitable division of points across all shapes in the current batch. Note that while \\( \\nu_c \\) facilitates an even division at the batch level, the points within each individual shape are not necessarily evenly partitioned according to the batch-based bin boundary values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.304 + ], + "angle": 0, + "content": "During the training, for the first iteration, we directly use the boundary values derived from the first batch of data as the dynamic boundary values. Subsequently, since the second iteration, boundaries are updated adaptively in a momentum-based manner:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.312, + 0.482, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\nu} _ {t} = \\gamma \\boldsymbol {\\nu} _ {t - 1} + (1 - \\gamma) \\boldsymbol {\\nu} _ {c}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.424 + ], + "angle": 0, + "content": "where \\(\\nu_{t - 1}\\) stands for the bin partitioning boundaries used in the last iteration, and \\(\\nu_{t}\\) is the updated dynamic boundaries used for the current iteration. \\(\\gamma \\in (0,1)\\) is the momentum update factor. With updated boundary values \\(\\nu_{t}\\), points in each shape are divided into \\(n_b\\) subsets of \\(\\{\\mathcal{B}_1,\\mathcal{B}_2,\\dots ,\\mathcal{B}_{n_b}\\}\\) based on their sampling scores." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.484, + 0.593 + ], + "angle": 0, + "content": "The core idea presented here is that, instead of using pre-engineered bin boundary values, we employ adaptive learning for bin boundaries, and they are gradually learned from the entire training dataset. These values are intended to evenly partition the distribution of point-wise sampling scores across all shapes and points in the training dataset. Consequently, for each individual shape, the acquired boundary values can effectively partition its points into bins with a shape-specific strategy, capturing the unique characteristics of the shape while maintaining a degree of proximity to other shapes within the dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Tokens for Learning Bin Weights. With points already being partitioned into bins for each shape, the next step is to learn a shape-specific sampling strategy, i.e., to learn shape-specific sampling weights for each bin. Inspired by ViT [6], ViT [13], and Mask3D [35] —which leverage additional tokens during the computation of attention maps to extract and convey information across the entire feature map or specific groups of points or pixels — we introduce additional tokens specifically for learning bin sampling weights. In our case, attention maps are computed shape-specific during the downsampling process, facilitating the learning of bin sampling weights also in a shape-specific manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Using the former proposed bin partitioning method, points in each shape are partitioned into \\( n_b \\) subsets of \\( \\{\\mathcal{B}_1, \\mathcal{B}_2, \\dots, \\mathcal{B}_{n_b}\\} \\). The sampling weight \\( \\omega_j \\) for bin \\( \\mathcal{B}_j (j = 1, 2, \\dots, n_b) \\) is established based on the distinctive features of each shape. Fig. 6 gives the network structure of our proposed downsampling layer and illustrates the idea of using additional tokens. \\( n_b \\) bin tokens are introduced during the attention computation, where each token corresponds to a" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.735, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.09, + 0.903, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.387, + 0.907, + 0.444 + ], + "angle": 0, + "content": "Figure 6. Network structure of our proposed downsampling layer. Block \\(\\mathcal{A}\\): Points in each shape are partitioned into \\(n_b\\) bins. Block \\(\\mathcal{B}\\): Masking the split-out point-to-token sub-attention map. Block \\(\\mathcal{C}\\): Learned bin sampling weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.72 + ], + "angle": 0, + "content": "specific bin. As shown in Fig. 6, the bin tokens are initially concatenated with the input point-wise features for Key and Value. Subsequently, the combined features are subjected to a cross-attention mechanism with the original point-wise features as Query. The attention map is split into two parts of a point-to-point sub-attention map and a point-to-token sub-attention map. For the point-to-point attention map, the methods proposed in Sec. 3.1 and Sec. 3.2 are applied to it to obtain point-wise sampling scores. Note that in this case, the row-wise sum is not exactly equal to 1 but still very close to 1 since \\( n_b \\) is of a very small quantity compared to \\( N \\). With computed point scores, dynamic boundary values \\( \\mathbf{v}_t \\) are obtained for bin partitioning. Using the information regarding the allocation of points to respective bins, a mask operation is performed on the point-to-token sub-attention map as illustrated in Block B of Fig. 6. The sampling weights \\( \\omega_j \\) are then subsequently acquired with" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.725, + 0.907, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\omega_ {j} = \\operatorname {R e L U} \\left(\\frac {1}{\\beta_ {j}} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {B} _ {j}} m _ {\\mathbf {p} _ {i}, \\mathcal {B} _ {j}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.778, + 0.907, + 0.824 + ], + "angle": 0, + "content": "where \\(\\beta_{j}\\) stands for the number of points in bin \\(\\mathcal{B}_j\\), and \\(m_{\\mathbf{p}_i,\\mathcal{B}_j}\\) represents the element in the energy matrix corresponding to point \\(\\mathbf{p}_i\\) in row and \\(\\mathcal{B}_j\\) in column." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In-Bin Point Sampling. For each shape, by considering the number of points contained within bins \\(\\beta = (\\beta_{1},\\beta_{2},\\ldots ,\\beta_{n_{b}})\\) alongside the determined bin sampling weights \\(\\omega = (\\omega_{1},\\omega_{2},\\dots ,\\omega_{n_{b}})\\), the specific numbers of points to be sampled from each bin \\(\\kappa = (\\kappa_{1},\\kappa_{2},\\dots ,\\kappa_{n_{b}})\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1346" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.088, + 0.898, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.273, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of our proposed SAMBLE. Apart from the sampled results, sampling score heatmaps and bin histograms along with bin sampling ratios are also given. All shapes are from the test set. Zoom in for optimal visual clarity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.321, + 0.482, + 0.456 + ], + "angle": 0, + "content": "need to be determined. Direct multiplication of \\(\\beta\\) and \\(\\omega\\) does not yield a sum that aligns with the total number of down-sampled points \\(M\\) required by the network structure. To address this discrepancy, a scaling method is applied to first scale bin sampling weights \\(\\omega_{j}\\). Furthermore, to prevent \\(\\kappa_{j}\\) from surpassing the available point number \\(\\beta_{j}\\) in any bin, any excess points are proportionately redistributed to other bins that have not been fully sampled. The detailed algorithm is presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Finally, within bin \\(\\mathcal{B}_j\\), \\(\\kappa_{j}\\) points are selected through random sampling with priors. The sampling probability \\(\\rho_{\\mathbf{p}_i}\\) is obtained via softmax operation over the normalized point sampling score \\(a_{\\mathbf{p}_i}\\) with a temperature parameter \\(\\tau\\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.522, + 0.483, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\rho_ {\\mathbf {p} _ {i}} = \\frac {e ^ {a _ {\\mathbf {p} _ {i}} / \\tau}}{\\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {B} _ {j}} e ^ {a _ {\\mathbf {p} _ {i}} / \\tau}}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.572, + 0.224, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.598, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Like most related works, such as S-NET [7], SampleNet [14], and APES [46], SAMBLE is specifically designed for point cloud shapes. To ensure a fair comparison, we conduct experiments using the same base network architecture as APES [46] on standard point cloud shape datasets, including ModelNet40 and ShapeNet-Part. It is important to note that point cloud sampling is not a standalone task; its effectiveness must be validated through downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.232, + 0.743 + ], + "angle": 0, + "content": "4.1. Classification" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experiment Setting. ModelNet40 classification benchmark [53] includes 12,311 CAD models across 40 categories. For a fair comparison, we use the official train-test split, with 9,843 for training and 2,468 for testing. Points are uniformly sampled from the mesh surface and normalized to the unit sphere. Only 3D coordinates are used as input, with random scaling, rotation, and shifting applied for data augmentation. We use \\( n_b = 6 \\) bins for point partitioning. The momentum update factor \\( \\gamma = 0.99 \\) for updating bin boundary values. The temperature parameter \\( \\tau = 0.1 \\)." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.318, + 0.891, + 0.561 + ], + "angle": 0, + "content": "
MethodCls. OA (%)Seg.
Cat. mIoU (%)Ins. mIoU (%)
PointNet [30]89.280.483.7
PointNet++ [31]91.981.985.1
SpiderCNN [55]92.482.485.3
DGCNN [42]92.982.385.2
PointConv [49]92.582.885.7
PT1 [9]92.8-85.9
PT2 [60]93.783.786.6
PCT [11]93.2-86.4
PRA-Net [5]93.783.786.3
CurveNet [27]93.8-86.6
DeltaConv [44]93.8-86.6
PointNeXt [33]93.284.486.7
PointMetaBase [19]-84.386.7
APES (local) [46]93.583.185.6
APES (global) [46]93.883.785.8
SAMBLE94.284.586.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.57, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Table 2. Classification and segmentation results on the ModelNet40 and ShapeNet-Part benchmarks. In comparison with other SOTA methods that also only use raw point cloud data as input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative and Quantitative Results. Qualitative results of SAMBLE are presented in Fig. 7, including sampling score heatmaps, learned bin partitioning strategy with bin sampling ratios, and final sampled results. From it, we can observe that SAMBLE effectively samples sufficient edge points to capture the shape structure. Moreover, it maintains better global uniformity by avoiding excessive focus on edge points, particularly in those thin or sharp regions like chair legs. Logged shape bin histograms confirm the learning of shape-specific sampling strategies. More visualization results are provided in the supplementary materials, highlighting an intriguing pattern where shapes of the same category exhibit similar histogram distributions and sampling strategies. Overall, SAMBLE successfully achieves an improved trade-off between sampling edge points and preserving shape global uniformity. Quantitative results are provided in Tab. 2. Our method performs better than previous approaches and achieves state-of-the-art performance." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1347" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.488, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.349 + ], + "angle": 0, + "content": "Figure 8. Segmentation results of our proposed SAMBLE in comparison with APES. All shapes are from the test set." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.362, + 0.462, + 0.426 + ], + "angle": 0, + "content": "
MethodPointNeXtSAMBLE
Point Number2048102451220481024512
Cat. mIoU (%)84.4083.7982.7784.5184.8485.04
Ins. mIoU (%)86.7086.1885.1886.6786.9387.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.467 + ], + "angle": 0, + "content": "Table 3. Additional segmentation performances evaluated on the intermediate downsampled sub-point clouds." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.493, + 0.235, + 0.51 + ], + "angle": 0, + "content": "4.2. Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Experiment setting. The ShapeNet-Part dataset [56] is used for 3D object part segmentation. It includes 16,880 3D models across 16 categories, with 14,006 models for training and 2,874 for testing. Each category contains 2-6 parts, totaling 50 distinct parts. We use the sampled point sets produced in [30] for a fair comparison with prior work. For evaluation metrics, we report both category mIoU and instance mIoU. We use \\( n_b = 4 \\) bins for point partitioning. The momentum update factor \\( \\gamma = 0.99 \\) for updating bin boundary values. The temperature parameter \\( \\tau = 0.1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Qualitative and Quantitative Results. Qualitative results are presented in Fig. 8, where we observe that, compared to APES, which tends to focus excessively on edge points, SAMBLE achieves a significantly improved balance between sampling edge points and preserving shape global uniformity. For example, SAMBLE demonstrates a more balanced use of non-edge points, as seen in the chair seat and lamp base, reflecting a thoughtful sampling strategy that accounts for different point categories and provides a more comprehensive representation of the overall shape. The quantitative results in Tab. 2 further highlight that SAMBLE achieves state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For the part segmentation benchmark, we further report the performance on the intermediate downsampled subpoint clouds in Tab. 3. Additionally, results from PointNeXt" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 9. Sampled results of few-point sampling in comparison with APES. Zoom in for optimal clarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.598 + ], + "angle": 0, + "content": "[33] are also presented, which is a prominent point cloud learning method that employs FPS for downsampling. It is evident that FPS-based methods exhibit poorer performance when evaluated on intermediate downsampled subpoint clouds. In contrast, a notable observation is that SAMBLE achieves superior performance on intermediate downsampled points. This indicates that the learned sampled points contribute more to the overall performance, while the upsampling layer cannot fully reconstruct the features of the discarded points. Although SAMBLE enables the interpolation-based upsampling and it outperforms the upsampling layer used in APES (see details in Sec. 4.4), there remains potential for further improvement by designing a more meticulously crafted upsampling layer. However, this is another topic and beyond the scope of this paper." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.707, + 0.621 + ], + "angle": 0, + "content": "4.3. Few-Point Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Experiment setting. We further compare our sampling method to previous approaches, including RS, FPS, and more recent learning-based methods such as S-Net, SampleNet, LightTN, APES, and others. The evaluation follows the same framework as in [7, 41, 46]. First, the point cloud is downsampled to a limited number of points, and the resulting subset is then fed into a task network for evaluation. For this comparison, we use the ModelNet40 classification task with the vanilla PointNet network. All sampling methods are evaluated across multiple sampling sizes of \\( M \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative and Quantitative Results. Quantitative results are presented in Tab. 4. Note that APES [46] uses FPS to pre-process the input into \\(2M\\) points while we do not. For a fair comparison, additional results of APES without the preprocessing step are also tested and reported. Nonetheless, even without pre-processing, SAMBLE achieves state-of-the-art results in the few-point sampling task as the number of sampled points decreases to extremely smaller ones." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1348" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.173 + ], + "angle": 0, + "content": "
MVoxelRSFPS [8]S-NET [7]PST-NET [40]SampleNet [14]MOPS-Net [34]DA-Net [21]LightTN [41]APES [46] (w/ pre-pro.)APES [46] (w/o pre-pro.)SAMBLE
51273.8287.5288.3487.8087.9488.1686.6789.0189.9190.8189.8190.58
25673.5077.0983.6482.3883.1584.2786.6386.2488.2190.4086.7890.18
12868.1556.4470.3477.5380.1180.7586.0685.6786.2689.7784.8790.02
6458.3131.6946.4270.4576.0679.8685.2585.5586.5189.5779.2389.81
3220.0216.3526.5860.7063.9277.3184.2885.1186.1888.5675.6389.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.183, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Table 4. Comparison with other sampling methods. Evaluated on the ModelNet40 classification benchmark with multiple sampling sizes. For APES, we additionally report its performance when pre-processing is not performed for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.232, + 0.484, + 0.291 + ], + "angle": 0, + "content": "
Indexing Modeiiiiiiivvvivii
Cls.OA (%)93.9293.7893.6393.6693.4094.1194.08
Seg.Cat. mIoU (%)83.9883.8583.6283.5183.4784.1284.22
Ins. mIoU (%)86.1685.9985.7485.6085.4986.3886.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.33 + ], + "angle": 0, + "content": "Table 5. Classification and segmentation performance with different indexing modes." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.342, + 0.484, + 0.401 + ], + "angle": 0, + "content": "
Number of Bins124681012
Cls.OA (%)93.9593.9193.9894.1894.0293.8093.84
Seg.Cat. mIoU (%)84.2284.1484.5184.4084.1983.9884.36
Ins. mIoU (%)86.4686.2886.6786.6186.4886.2386.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.439 + ], + "angle": 0, + "content": "Table 6. Classification and segmentation performance with different number of bins." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Qualitative results are presented in Fig. 9. For few-point sampling, APES relies on FPS to pre-sample the input into \\(2M\\) points due to its limitations. In contrast, our method preserves better global uniformity, allowing direct few-point sampling from the input while still achieving satisfactory sampled results, as demonstrated in Fig. 9. When sampling very few points, APES tends to concentrate on the sharpest regions, whereas our SAmBLE method preserves better global uniformity throughout the point cloud shape." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.245, + 0.618 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.7 + ], + "angle": 0, + "content": "In this subsection, our emphasis is directed toward the novel designs introduced within this paper, excluding common topics such as network width. More ablation studies and further design justifications are provided in the supplementary materials to enhance our method's interpretability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.778 + ], + "angle": 0, + "content": "Different Indexing Modes. Apart from the visualized heatmaps given in Fig. 5, we also report their respective experimental results in Tab. 5. The tests are performed using top-M as the sampling strategy. From it, we can observe that indexing modes vi and vii achieve best performances." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Number of Bins. As a key parameter in SAMBLE, an ablation study is performed over the number of bins \\( n_b \\). The results are presented in Tab. 6. Remarkably, increasing the number of bins does not yield improved performance. This phenomenon is likely attributable to the subdivision of shapes into an excessive number of point categories, leading to the gradual diminishment of score disparities across the bins. In our case, \\( n_b = 6 \\) and 4 yield the best performance" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.232, + 0.906, + 0.308 + ], + "angle": 0, + "content": "
UpsampleInterpolationCross-Attention
Kup=3Kup=8Kup=16
APES (local)82.89 / 85.4082.95 / 85.4482.96 / 85.4283.11 / 85.58
APES (global)83.16 / 85.5383.19 / 85.5983.17 / 85.5583.67 / 85.81
SAMBLE84.51 / 86.6784.35 / 86.4884.31 / 86.4384.36 / 86.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 7. Segmentation results with different upsampling layers on ShapeNet-Part. The number before “/” is the category mIoU, and the number after is the instance mIoU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.411 + ], + "angle": 0, + "content": "for the classification and segmentation tasks respectively, and we use it for the corresponding experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Upsampling layer. An important aspect to highlight is the upsampling layer. Most point cloud network models employ neighbor-based interpolation [31, 33, 60] for upsampling, as FPS is used during the downsampling process. However, APES introduces a cross-attention layer for upsampling to address its limitations of overemphasizing edge points, which renders traditional neighbor-based interpolation impractical. In contrast, our method achieves an improved balance between sampling edge points and preserving global uniformity, allowing the use of interpolation operations during upsampling. An ablation study for evaluating various upsampling layers and interpolation with different \\( K_{up} \\) values is conducted, and the results are presented in Table 7. The results show a performance drop for APES when interpolation is used in place of cross-attention, while SAMBLE demonstrates superior performance with it." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.634, + 0.684 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, a novel point cloud sampling method SAMBLE is proposed to learn shape-specific sampling strategies for point cloud shapes. Based on a sparse attention map that integrates both local and global information, multiple indexing modes are designed and explored. By partitioning the points in each shape into bins and learning respective sampling ratios for each bin, shape-specific sampling strategies are acquired for individual point cloud shapes. SAMBLE achieves an optimal trade-off between sampling local details and preserving global uniformity, resulting in improved performance on downstream tasks. For future directions, advancements in upsampling layers could further improve the model's performance. Additionally, adapting the proposed method for point cloud scenes is another promising area to explore." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1349" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.144, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[1] Pyunghwan Ahn, Juyoung Yang, Eojindl Yi, Chanho Lee, and Junmo Kim. Projection-based point convolution for efficient point cloud segmentation. IEEE Access, 10:15348-15358, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.483, + 0.256 + ], + "angle": 0, + "content": "[2] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Semantic segmentation of earth observation data using multimodal and multi-scale deep networks. In *Asian Conference on Computer Vision*, pages 180–196. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.482, + 0.298 + ], + "angle": 0, + "content": "[3] Alexandre Boulch, Bertrand Le Saux, and Nicolas Audebert. Unstructured point cloud semantic labeling using deep segmentation networks. 3DOR@ Eurographics, 3, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[4] Can Chen, Luca Zanotti Fragonara, and Antonios Tsourdos. GAPointNet: Graph attention based point neural network for exploiting local feature of point cloud. Neurocomputing, 438:122-132, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[5] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. PRA-Net: Point relation-aware network for 3d point cloud analysis. IEEE Transactions on Image Processing, 30: 4436-4448, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2760-2769, 2019. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[8] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[9] Nico Engel, Vasileios Belagiannis, and Klaus Dietmayer. Point transformer. IEEE Access, 9:134826-134840, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Fabian Groh, Patrick Wieschollek, and Hendrik PA Lensch. Flex-convolution: Million-scale point-cloud learning beyond grid-worlds. In Asian Conference on Computer Vision (ACCV), pages 105–122. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. PCT: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[12] Mingyang Jiang, Yiran Wu, Tianqi Zhao, Zelin Zhao, and Cewu Lu. PointSift: A sift-like network module for 3d point cloud semantic segmentation. arXiv preprint arXiv:1807.00652, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[13] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[14] Itai Lang, Asaf Manor, and Shai Avidan. SampleNet: Differentiable point cloud sampling. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.144, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7578-7588, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[15] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. DPC: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[16] Felix Järemo Lawin, Martin Danelljan, Patrik Tosteberg, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Deep projective 3d semantic segmentation. In International Conference on Computer Analysis of Images and Patterns, pages 95-107. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[17] Truc Le and Ye Duan. PointGrid: A deep network for 3d shape understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9204-9214, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.303, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution on x-transformed points. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.359, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[19] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17682-17691, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[20] Yiqun Lin, Zizheng Yan, Haibin Huang, Dong Du, Ligang Liu, Shuguang Cui, and Xiaoguang Han. FPCov: Learning local flattening for point convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4293-4302, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[21] Yanan Lin, Yan Huang, Shihao Zhou, Mengxi Jiang, Tianlong Wang, and Yunqi Lei. DA-Net: Density-adaptive down-sampling network for point cloud classification via end-to-end learning. In 2021 4th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), pages 13–18. IEEE, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[22] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1800-1809, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.65, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision (ECCV), pages 657-675. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[24] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[25] Daniel Maturana and Sebastian Scherer. VoxNet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. CurveNet: Curvature-based multitask learning" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1350" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[28] Ehsan Nezhadarya, Ehsan Taghavi, Ryan Razani, Bingbing Liu, and Jun Luo. Adaptive hierarchical down-sampling for point cloud classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12956-12964, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.483, + 0.246 + ], + "angle": 0, + "content": "[29] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision (ECCV), pages 604-621. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6329-6338, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems (NeurIPS), 35:23192-23204, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[34] Yu Qian, Junhui Hou, Yiming Zeng, Qijian Zhang, Sam Tak Wu Kwong, and Ying He. MOPS-Net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. ArXiv, abs/2005.00383, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[36] Martin Simonovsky and Nikos Komodakis. Dynamic edge-conditioned filters in convolutional neural networks on graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3693-3702, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3887-3896, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[38] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. KPCov: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[40] Xu Wang, Yi Jin, Yigang Cen, Congyan Lang, and Yidong Li. PST-Net: Point cloud sampling via point-based transformer. In 11th International Conference on Image and Graphics (ICIG), pages 57–69. Springer, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[41] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. LighTN: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. IEEE Transactions on Multimedia, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[42] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[43] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learnable skeleton-aware 3d point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[44] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. DeltaConv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.483 + ], + "angle": 0, + "content": "[45] Chengzhi Wu, Xuelei Bi, Julius Pfrommer, Alexander Cebulla, Simon Mangold, and Jürgen Beyerer. Sim2real transfer learning for point cloud segmentation: An industrial application case on autonomous disassembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4531-4540, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[46] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5333-5343, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[47] Chengzhi Wu, Qianliang Huang, Kun Jin, Julius Pfrommer, and Jürgen Beyerer. A cross branch fusion-based contrastive learning framework for point cloud self-supervised learning. In 2024 International Conference on 3D Vision (3DV), pages 528-538. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[48] Chengzhi Wu, Kaige Wang, Zeyun Zhong, Hao Fu, Junwei Zheng, Jiaming Zhang, Julius Pfrommer, and Jurgen Beyerer. Rethinking attention module design for point cloud analysis. In International Conference on Pattern Recognition (ICPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[49] Wenxuan Wu, Zhongang Qi, and Li Fuxin. PointConv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9621-9630, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[50] Wenxuan Wu, Li Fuxin, and Qi Shan. PointConvFormer: Revenge of the point-based convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21802-21813, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems (NeurIPS), 35:33330-33342, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[52] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1351" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4840-4851, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[53] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[54] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-GCN for fast and scalable point cloud learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5661-5670, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[55] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[57] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[58] Kuangen Zhang, Ming Hao, Jing Wang, Xinxing Chen, Yuquan Leng, Clarence W de Silva, and Chenglong Fu. Linked dynamic graph cnn: Learning through point cloud by linking hierarchical features. In 2021 27th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 7-12. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[59] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in Neural Information Processing Systems (NeurIPS), 35:27061-27074, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[61] Yin Zhou and Oncel Tuzel. VoxelNet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4490-4499, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.77, + 0.482, + 0.824 + ], + "angle": 0, + "content": "[62] Wei Zhu, Yue Ying, Jin Zhang, Xiuli Wang, and Yayu Zheng. Point cloud registration network based on convolution fusion and attention mechanism. Neural Processing Letters, pages 1-21, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1352" + } + ] +] \ No newline at end of file diff --git a/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_origin.pdf b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c109322decffa8a7389b8b32e5affd2e854ee39e --- /dev/null +++ b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/10d62ba6-9e83-4677-8d7a-788b8bf292ff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a64f7459cabc097d49dc87d0d41f794b766f516eee0ebf5ec905bd571ed9f861 +size 5371552 diff --git a/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/full.md b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab76bfcc8b7d7fc22dc1e0ca8962dff3320adc42 --- /dev/null +++ b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/full.md @@ -0,0 +1,315 @@ +# SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity + +Chengzhi Wu1 + +Zeyun Zhong + +Yuxin Wan1* + +Junwei Zheng $^{1\dagger}$ + +Hao $\mathbf{F u}^{1*}$ + +Jiaming Zhang + +Julius Pfrommer2 + +Jirgen Beyerer $^{1,2}$ + +$^{1}$ Karlsruhe Institute of Technology, Germany + +$^{2}$ Fraunhofer IOSB, Germany + +{chengzhi.wu, zeyun.zhong, junwei.zheng, jiangming.zhang}@kit.edu, + +{yuxin.wan, hao.fu}@student.kit.edu, {julius.pfrommer, juergen.beyerer}@iosb.fraunhofer.de + +# Abstract + +Driven by the increasing demand for accurate and efficient representation of 3D data in various domains, point cloud sampling has emerged as a pivotal research topic in 3D computer vision. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on sharp edge details. Moreover, they all overlook the natural variations in point distribution across different shapes, applying a similar sampling strategy to all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes. SAMBLE effectively achieves an improved balance between sampling edge points for local details and preserving uniformity in the global shape, resulting in superior performance across multiple common point cloud downstream tasks, even in scenarios with few-point sampling. + +# 1. Introduction + +Point cloud sampling is a less explored research area within the realm of this data representation. Traditional random sampling (RS) and farthest point sampling (FPS) remain the most commonly employed methods when sampling is required for point cloud learning. With the advancement of neural networks, several methods have emerged for point cloud sampling in a downstream task-oriented learning framework, including S-Net [7], SampleNet [14], MOPS-Net [34], etc. However, these methods first generate a new, smaller-sized point cloud as a proxy rather than directly sampling points from the original input, rendering the + +![](images/16a07d5cb8a61a3c346fea446e86a7c46c17ebb8d56e938aed7cda5bd58c4b45.jpg) +Figure 1. Our method achieves an improved trade-off between sampling local details and preserving global uniformity. + +techniques akin to black-box neural network models with limited interpretability. Consequently, discerning geometric patterns in their qualitative results becomes challenging, as their outcomes closely resemble those obtained through random sampling. More recently, APES [46] pioneers the direction of using neural networks to learn point-wise sampling scores, with which it subsequently samples points whose scores are higher. However, with its score computation design and the Top-M sampling strategy, APES excessively focuses on local details of edge points, resulting in a deficiency in preserving good global uniformity of the input shapes. Consequently, the interpolation operation becomes impractical during the upsampling process, and the sampling quality of few-point sampling is notably subpar (see Fig. 1). In this paper, we introduce a novel point cloud sampling method that addresses the limitations of prior approaches, aiming to achieve a refined balance between capturing local details and preserving global uniformity. + +The concept originates from rethinking the mathematical characteristics of local details within point cloud shapes. Typically, these local details are represented by edge points + +![](images/4a42dd2e458f64fa44e2e9873c157dc9584f431b0abf426ffc333f1f4c0b0269.jpg) +Figure 2. When selecting an equal number of neighbors for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies (numbers on the right). + +that define the shape's outline and sharpest features. Is there a point property that can easily distinguish between different categories, such as edge points and non-edge points? The answer is affirmative. In our investigation, we have uncovered an extremely fundamental yet crucial observation: if point $\mathbf{p}_i$ is one of the $k$ -nearest neighbors of point $\mathbf{p}_j$ , it does not necessarily imply that $\mathbf{p}_j$ is also among the $k$ -nearest neighbors of $\mathbf{p}_i$ . Consequently, it leads to the conclusion that the frequency of each point being chosen as a neighbor exhibits variation across a single point cloud. + +We explore and demonstrate the importance of this point property with a simple example as illustrated in Fig. 2. Assume the input point cloud is a simple grid. When selecting 5 neighbors for each point, all three possible cases are given on the left (center point is self-contained as a neighbor). Note that in the triangular and rectangular cases, they each has a "quantum-entangled" twin point pair, in which two points share the possibility of being chosen as the neighbor. While an equal number of neighbors is selected for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies, as listed on the right of Fig. 2. From it, we can observe that in addition to the edge point and non-edge point categories, there is also another noteworthy point category of close-to-edge points. Moreover, within each category, the points can be further grouped into more sub-categories. Overall, this point property effectively captures the local characteristics of a shape, especially for shape outline and sharp details. Building on it, we propose Sparse Attention Map (SAM) and introduce new methods for computing point-wise sampling scores to effectively balance the trade-off between local and global sampling. See more details in Sec. 3.2. + +On the other hand, after the point-wise sampling scores are computed, previous methods employ a Top-M sampling strategy for all point cloud shapes, which exacerbates the issue of oversampling edge points. We argue that the naive top-M sampling strategy may not be optimal across all point cloud shapes for downstream tasks. For example, sampling more non-edge points enhances global uniformity, while sampling more close-to-edge points "thickens" the edge, both of which can potentially improve the performance on downstream tasks [46]. To address this, we introduce a + +novel bin-based method to explore better sampling strategies shape-specifically by leveraging all point categories. This approach enables the sampling of points with smaller sampling scores, further optimizing the local-global tradeoff. As a result, our method dynamically adjusts the sampling strategy for each shape, leading to a more tailored and interpretable sampling process for improved performance. + +Our main contributions are summarized as follows: + +- We propose a sparse attention map that directly integrates shape local and global information at the attention map level for point cloud sampling, introducing multiple methods for computing point-wise sampling scores. +- We present a novel approach for learning bin boundaries to partition points within individual shapes, enabling shape-specific sampling strategies by incorporating additional bin tokens during the attention computation. +- Our method successfully achieves an improved trade-off between capturing local details and preserving global uniformity for the sampling process, resulting in enhanced performance both qualitatively and quantitatively. + +# 2. Related Work + +Point Cloud Sampling. Point cloud sampling is a key process in 3D data handling for simplifying high-resolution dense point clouds. Over the past decades, non-learning-based methods [8, 10, 26] have predominantly been used for point cloud sampling. While Farthest Point Sampling (FPS) [8] is the most widely used one [18, 31, 33, 49, 60], Random Sampling (RS) has also been frequently adopted [10, 32, 61]. More recently, learning-based sampling methods have shown superior performance with task-oriented training. S-Net [7] represents a pioneering work of generating new point coordinates from global representations, while SampleNet [14] introduces a soft projection operation for better point approximation. Following S-Net, multiple learning-based methods have been proposed [21, 28, 40, 41]. MOPS-Net [34] learns a transformation matrix and multiplies it with the original point cloud to generate the sampled one. By employing the attention mechanism to learn point-wise sampling scores, APES [46] captures the edge points in the input point clouds with a strong focus. + +Deep Learning on Point Clouds. In contrast to the voxelization-based methods [12, 17, 25] and multi-view-based methods [2, 3, 16, 37], point-based methods deal directly with point clouds. The pioneer studies of PointNet [30] and PointNet++ [31] tackle point clouds through pointwise Multi-Layer Perceptrons (MLPs) and max-pooling operations. Subsequently, other research shifts focus towards constructing more efficient building blocks for local feature extraction, such as convolution-based ones [1, 18, 20, 38, 49, 50, 62] and graph-based ones [4, 15, 22, 24, 36, 42, 54, 58]. More recently, while MLP-based methods like PointNeXt [33] and PointMetaBase [19] have rekin + +![](images/462afef2c30ba826d6a0784e920c27b971073e864edbe4ea303523083b525c3b.jpg) +Figure 3. A brief pipeline of our proposed method SAMBLE. It learns shape-specific sampling strategies for point cloud shapes. + +dled people's interest, the application of attention mechanisms to point cloud analysis has also garnered widespread attention [9, 11, 39, 43, 45, 48, 57, 60]. For example, PT [51, 52, 60] series improve the model performance by introducing subtraction-based attention blocks, and [48] performs a large ablation study over attention module designs for point cloud processing. In addition, approaches that apply Transformers for point cloud self-supervised learning have also been proposed and explored [23, 29, 47, 57, 59]. + +# 3. Methodology + +A brief pipeline of SAMBLE is illustrated in Fig. 3. It consists of three key steps: constructing a sparse attention map, computing point-wise sampling scores, and learning shape-specific sampling strategies through bin partitioning. + +# 3.1. Sparse Attention Map + +Local and Global Attention Maps. Both local and global attention maps are widely used in point cloud analysis. A global attention map is derived from the application of classical self-attention to point features of all points, while a local attention map concentrates on a point-centered area wherein cross-attention is specifically applied to the central point and its neighbors. + +Denote $\mathcal{S}_i$ as the set of $k$ -nearest neighbors of point $\mathbf{p}_i$ , the local attention map for $\mathbf{p}_i$ is defined as + +$$ +\mathbf {m} _ {i} ^ {l} = \operatorname {s o f t m a x} \left(Q (\mathbf {p} _ {i}) K \left(\mathbf {p} _ {i j} - \mathbf {p} _ {i}\right) _ {j \in \mathcal {S} _ {i}} ^ {\top} / \sqrt {d}\right), \tag {1} +$$ + +where $Q$ and $K$ stand for the linear layers applied on the query and key input, and the square root of the feature dimension count $\sqrt{d}$ serves as a scaling factor [39]. + +For the global attention map which is equivalent to taking all points as neighbors for each point, it is defined as + +$$ +\mathbf {M} ^ {g} = \operatorname {s o f t m a x} \left(Q (\mathbf {p} _ {i}) K (\mathbf {p} _ {j}) _ {i, j \in \mathcal {S}} ^ {\top} / \sqrt {d}\right), \tag {2} +$$ + +where $S$ denotes the set of all input points. + +Sparse Attention Map. Instead of using local or global attention maps solely, we propose sparse attention map, + +![](images/0c10a1eaf649b8be3a09d1ddfe454ae27f25e8eda27df7a4609583cf5dca31d4.jpg) +Figure 4. Sparse attention map. In each row, $k$ cells are selected based on the KNN neighbor indexes for each point. The values of other non-selected cells are all set to 0. Note that the number of cells selected within each column is variable. + +which combines the knowledge from both local and global information, to compute point-wise sampling scores. The idea is illustrated in Fig. 4. After obtaining the global attention map with Eq. 2, $k$ -NN is employed locally to find $k$ neighbors for each point. In this case, $k$ cells are being selected in each row. However, as discussed before, please notice that each point is chosen as a neighbor with varying frequencies. This means while for each row $k$ cells are selected, for each column, the number of selected cells varies. The selected cells are then "carved" out to form the sparse attention map, with the values of other non-selected cells being set to 0. More vividly, consider the global attention map as a grid stone slab of size $N \times N$ . For carve-based SAM, values of all cells are pre-computed and hidden in the slab grid cells, and only the selected cells are carved out. + +# 3.2. Computing Point-wise Sampling Score + +Indexing Mode. When sampling points, the points are indexed based on the computed point-wise sampling scores. We call the method of computing point-wise sampling scores from the full/sparse attention map as Indexing Mode. With the original full attention map, following APES, there are two possible indexing modes: (i) row standard deviation; and (ii) column sum. For a global attention map $\mathbf{M}^g$ of size $N\times N$ , denote $m_{ij}$ as the value of $i$ th row and $j$ th column in $\mathbf{M}^g$ . These two indexing modes are formulated as modes (i) and (ii) in Tab. 1. To avoid possible confusion, we use notation $\mathbf{p}_o$ to denote a point only in this subsection. + +With the proposed sparse attention map, there are many other possible indexing modes. As discussed in Sec. 1, to achieve an improved trade-off between sampling edge points and preserving global uniformity, the frequency of each point being chosen as a neighbor, i.e., the number of selected cells in each column is the key. The following indexing modes are designed and explored for comparison: + +
Indexing ModeAttention MapFormulaRemark
(i) Row standard deviationFullap0 = fstd{m0j|j = 1,2,...,N}fstd: Computing standard deviation for a set of values
(ii) Column sumFullap0 = ∑i=1N mio
(iii) Row standard deviationSparseap0 = fstd{msj|j ∈ So}So: Set of indices of selected cells in oth row
(iv) Row sumSparseap0 = ∑j=1N msjNon-selected cells are all of 0s
(v) Column sumSparseap0 = ∑i=1N msi0
(vi) Column averageSparseap0 = ∑i=1N msio/n0n0: Number of selected cells in oth column
(vii) Column square-dividedSparseap0 = ∑i=1N msio/n02n0: Number of selected cells in oth column
+ +Table 1. Proposed different indexing modes for computing point-wise sampling scores. + +![](images/b8fdc4e0f0630112c3f6176e6b8ab1097fc70150a7d4112f1a71acc928103360.jpg) +Figure 5. Point sampling score heatmaps under different indexing modes. Scores are normalized to $\mathcal{N}(0.5,1)$ for better visualization. + +(iii) sparse row standard deviation; (iv) sparse row sum; (v) sparse column sum; (vi) sparse column average; and (vii) sparse column square-divided. Again, for a sparse attention map $\mathbf{M}^s$ of size $N\times N$ , denote $m_{ij}^{s}$ as the value of $i$ th row and $j$ th column in $\mathbf{M}^s$ . For point $\mathbf{p}_o$ , we denote the set of indexes of the selected $k$ cells (indexes of $k$ -nearest neighbors) in $o$ th row as $S_o$ , and denote the number of selected cells in $o$ th column as $n_o$ . Details and respective formulas of the proposed indexing modes are listed in Tab. 1. + +Heatmap. To analyze the behavior of each indexing mode, we train a separate model for each mode, ensuring that all other settings remain consistent. The sampling score distributions are visualized as heatmaps in Fig. 5, enhancing the interpretability of our method. From these heatmaps, we can see that both row-standard-deviation-based modes (modes i and iii) concentrate heavily on edge points. However, because they consistently prioritize thin or detailed regions, some areas may be overlooked. In contrast, modes ii and iv show less emphasis on edge points and instead distribute focus across a broader range of points, with a tendency toward other non-edge regions in a biased manner. + +More interestingly, the comparison of modes v, vi, and vii, which utilize column-wise information from SAM, reveals distinct sampling preferences and strategies across different point categories. Mode v prioritizes non-edge points, mode vi emphasizes the global shape, and mode + +vii focuses slightly more on edge points. This is because edge points typically have a smaller number of $n_o$ . Despite these differences and unique characteristics, all three modes capture the overall shape more uniformly compared to the former four. In our case, we aim to sample edge points without over-emphasizing them. For instance, when sampling detailed areas like chair legs, we want to capture some edge points without selecting them all, while also ensuring that non-edge points are sampled to preserve better global uniformity. Given this balance, we chose mode vii as the primary indexing mode for most of the experiments in the following sections. The detailed ablation study over different indexing modes is presented in Sec. 4.4. + +# 3.3. Sampling with Bins + +After point-wise sampling scores are computed with SAM, points are sampled according to certain rules. The simplest approach is top-M sampling, where points with the highest scores are sampled. In our case, as we aim to enhance the local-global trade-off and leverage all point categories during the sampling process, we suggest employing a bin-based sampling strategy to allow for the possible sampling of certain close-to-edge points or even non-edge points. + +Bin Partitioning. The process begins with processing the distribution of normalized point-wise sampling scores $a_{\mathbf{p}_i}$ across the shapes within the current batch. Let $n_b$ repre + +sent the number of bins used for partitioning, from which $n_b - 1$ bin boundary values need to be derived from the distribution. During each training step, a vector $\nu_c = (\nu_1, \nu_2, \dots, \nu_{n_b - 1})$ is computed based on the point score distribution, ensuring an equitable division of points across all shapes in the current batch. Note that while $\nu_c$ facilitates an even division at the batch level, the points within each individual shape are not necessarily evenly partitioned according to the batch-based bin boundary values. + +During the training, for the first iteration, we directly use the boundary values derived from the first batch of data as the dynamic boundary values. Subsequently, since the second iteration, boundaries are updated adaptively in a momentum-based manner: + +$$ +\boldsymbol {\nu} _ {t} = \gamma \boldsymbol {\nu} _ {t - 1} + (1 - \gamma) \boldsymbol {\nu} _ {c}, \tag {3} +$$ + +where $\nu_{t - 1}$ stands for the bin partitioning boundaries used in the last iteration, and $\nu_{t}$ is the updated dynamic boundaries used for the current iteration. $\gamma \in (0,1)$ is the momentum update factor. With updated boundary values $\nu_{t}$ , points in each shape are divided into $n_b$ subsets of $\{\mathcal{B}_1,\mathcal{B}_2,\dots ,\mathcal{B}_{n_b}\}$ based on their sampling scores. + +The core idea presented here is that, instead of using pre-engineered bin boundary values, we employ adaptive learning for bin boundaries, and they are gradually learned from the entire training dataset. These values are intended to evenly partition the distribution of point-wise sampling scores across all shapes and points in the training dataset. Consequently, for each individual shape, the acquired boundary values can effectively partition its points into bins with a shape-specific strategy, capturing the unique characteristics of the shape while maintaining a degree of proximity to other shapes within the dataset. + +Tokens for Learning Bin Weights. With points already being partitioned into bins for each shape, the next step is to learn a shape-specific sampling strategy, i.e., to learn shape-specific sampling weights for each bin. Inspired by ViT [6], ViT [13], and Mask3D [35] —which leverage additional tokens during the computation of attention maps to extract and convey information across the entire feature map or specific groups of points or pixels — we introduce additional tokens specifically for learning bin sampling weights. In our case, attention maps are computed shape-specific during the downsampling process, facilitating the learning of bin sampling weights also in a shape-specific manner. + +Using the former proposed bin partitioning method, points in each shape are partitioned into $n_b$ subsets of $\{\mathcal{B}_1, \mathcal{B}_2, \dots, \mathcal{B}_{n_b}\}$ . The sampling weight $\omega_j$ for bin $\mathcal{B}_j (j = 1, 2, \dots, n_b)$ is established based on the distinctive features of each shape. Fig. 6 gives the network structure of our proposed downsampling layer and illustrates the idea of using additional tokens. $n_b$ bin tokens are introduced during the attention computation, where each token corresponds to a + +![](images/8bf79882e586b50f754d2fbdcdb06df97f3a7b6a6656f82b05bb01cb76aa2178.jpg) +Figure 6. Network structure of our proposed downsampling layer. Block $\mathcal{A}$ : Points in each shape are partitioned into $n_b$ bins. Block $\mathcal{B}$ : Masking the split-out point-to-token sub-attention map. Block $\mathcal{C}$ : Learned bin sampling weights. + +![](images/1418140f004eac1091bc0941f5abc2e6d88060e319a3ecdb4005f751aabad8f3.jpg) + +specific bin. As shown in Fig. 6, the bin tokens are initially concatenated with the input point-wise features for Key and Value. Subsequently, the combined features are subjected to a cross-attention mechanism with the original point-wise features as Query. The attention map is split into two parts of a point-to-point sub-attention map and a point-to-token sub-attention map. For the point-to-point attention map, the methods proposed in Sec. 3.1 and Sec. 3.2 are applied to it to obtain point-wise sampling scores. Note that in this case, the row-wise sum is not exactly equal to 1 but still very close to 1 since $n_b$ is of a very small quantity compared to $N$ . With computed point scores, dynamic boundary values $\mathbf{v}_t$ are obtained for bin partitioning. Using the information regarding the allocation of points to respective bins, a mask operation is performed on the point-to-token sub-attention map as illustrated in Block B of Fig. 6. The sampling weights $\omega_j$ are then subsequently acquired with + +$$ +\omega_ {j} = \operatorname {R e L U} \left(\frac {1}{\beta_ {j}} \sum_ {\mathbf {p} _ {i} \in \mathcal {B} _ {j}} m _ {\mathbf {p} _ {i}, \mathcal {B} _ {j}}\right), \tag {4} +$$ + +where $\beta_{j}$ stands for the number of points in bin $\mathcal{B}_j$ , and $m_{\mathbf{p}_i,\mathcal{B}_j}$ represents the element in the energy matrix corresponding to point $\mathbf{p}_i$ in row and $\mathcal{B}_j$ in column. + +In-Bin Point Sampling. For each shape, by considering the number of points contained within bins $\beta = (\beta_{1},\beta_{2},\ldots ,\beta_{n_{b}})$ alongside the determined bin sampling weights $\omega = (\omega_{1},\omega_{2},\dots ,\omega_{n_{b}})$ , the specific numbers of points to be sampled from each bin $\kappa = (\kappa_{1},\kappa_{2},\dots ,\kappa_{n_{b}})$ + +![](images/04e6a6edfeedc4ef67d6f258093d03c472da16ebc5780c957ab8eb6faf6d04c7.jpg) +Figure 7. Qualitative results of our proposed SAMBLE. Apart from the sampled results, sampling score heatmaps and bin histograms along with bin sampling ratios are also given. All shapes are from the test set. Zoom in for optimal visual clarity. + +need to be determined. Direct multiplication of $\beta$ and $\omega$ does not yield a sum that aligns with the total number of down-sampled points $M$ required by the network structure. To address this discrepancy, a scaling method is applied to first scale bin sampling weights $\omega_{j}$ . Furthermore, to prevent $\kappa_{j}$ from surpassing the available point number $\beta_{j}$ in any bin, any excess points are proportionately redistributed to other bins that have not been fully sampled. The detailed algorithm is presented in the supplementary materials. + +Finally, within bin $\mathcal{B}_j$ , $\kappa_{j}$ points are selected through random sampling with priors. The sampling probability $\rho_{\mathbf{p}_i}$ is obtained via softmax operation over the normalized point sampling score $a_{\mathbf{p}_i}$ with a temperature parameter $\tau$ : + +$$ +\rho_ {\mathbf {p} _ {i}} = \frac {e ^ {a _ {\mathbf {p} _ {i}} / \tau}}{\sum_ {\mathbf {p} _ {i} \in \mathcal {B} _ {j}} e ^ {a _ {\mathbf {p} _ {i}} / \tau}}. \tag {5} +$$ + +# 4. Experiments + +Like most related works, such as S-NET [7], SampleNet [14], and APES [46], SAMBLE is specifically designed for point cloud shapes. To ensure a fair comparison, we conduct experiments using the same base network architecture as APES [46] on standard point cloud shape datasets, including ModelNet40 and ShapeNet-Part. It is important to note that point cloud sampling is not a standalone task; its effectiveness must be validated through downstream tasks. + +# 4.1. Classification + +Experiment Setting. ModelNet40 classification benchmark [53] includes 12,311 CAD models across 40 categories. For a fair comparison, we use the official train-test split, with 9,843 for training and 2,468 for testing. Points are uniformly sampled from the mesh surface and normalized to the unit sphere. Only 3D coordinates are used as input, with random scaling, rotation, and shifting applied for data augmentation. We use $n_b = 6$ bins for point partitioning. The momentum update factor $\gamma = 0.99$ for updating bin boundary values. The temperature parameter $\tau = 0.1$ . + +
MethodCls. OA (%)Seg.
Cat. mIoU (%)Ins. mIoU (%)
PointNet [30]89.280.483.7
PointNet++ [31]91.981.985.1
SpiderCNN [55]92.482.485.3
DGCNN [42]92.982.385.2
PointConv [49]92.582.885.7
PT1 [9]92.8-85.9
PT2 [60]93.783.786.6
PCT [11]93.2-86.4
PRA-Net [5]93.783.786.3
CurveNet [27]93.8-86.6
DeltaConv [44]93.8-86.6
PointNeXt [33]93.284.486.7
PointMetaBase [19]-84.386.7
APES (local) [46]93.583.185.6
APES (global) [46]93.883.785.8
SAMBLE94.284.586.7
+ +Table 2. Classification and segmentation results on the ModelNet40 and ShapeNet-Part benchmarks. In comparison with other SOTA methods that also only use raw point cloud data as input. + +Qualitative and Quantitative Results. Qualitative results of SAMBLE are presented in Fig. 7, including sampling score heatmaps, learned bin partitioning strategy with bin sampling ratios, and final sampled results. From it, we can observe that SAMBLE effectively samples sufficient edge points to capture the shape structure. Moreover, it maintains better global uniformity by avoiding excessive focus on edge points, particularly in those thin or sharp regions like chair legs. Logged shape bin histograms confirm the learning of shape-specific sampling strategies. More visualization results are provided in the supplementary materials, highlighting an intriguing pattern where shapes of the same category exhibit similar histogram distributions and sampling strategies. Overall, SAMBLE successfully achieves an improved trade-off between sampling edge points and preserving shape global uniformity. Quantitative results are provided in Tab. 2. Our method performs better than previous approaches and achieves state-of-the-art performance. + +![](images/9e3dd4e4ef9d27d6a8291912b297c5aa9817c11124f0f4783b423fde1932411a.jpg) +Figure 8. Segmentation results of our proposed SAMBLE in comparison with APES. All shapes are from the test set. + +
MethodPointNeXtSAMBLE
Point Number2048102451220481024512
Cat. mIoU (%)84.4083.7982.7784.5184.8485.04
Ins. mIoU (%)86.7086.1885.1886.6786.9387.12
+ +Table 3. Additional segmentation performances evaluated on the intermediate downsampled sub-point clouds. + +# 4.2. Segmentation + +Experiment setting. The ShapeNet-Part dataset [56] is used for 3D object part segmentation. It includes 16,880 3D models across 16 categories, with 14,006 models for training and 2,874 for testing. Each category contains 2-6 parts, totaling 50 distinct parts. We use the sampled point sets produced in [30] for a fair comparison with prior work. For evaluation metrics, we report both category mIoU and instance mIoU. We use $n_b = 4$ bins for point partitioning. The momentum update factor $\gamma = 0.99$ for updating bin boundary values. The temperature parameter $\tau = 0.1$ . + +Qualitative and Quantitative Results. Qualitative results are presented in Fig. 8, where we observe that, compared to APES, which tends to focus excessively on edge points, SAMBLE achieves a significantly improved balance between sampling edge points and preserving shape global uniformity. For example, SAMBLE demonstrates a more balanced use of non-edge points, as seen in the chair seat and lamp base, reflecting a thoughtful sampling strategy that accounts for different point categories and provides a more comprehensive representation of the overall shape. The quantitative results in Tab. 2 further highlight that SAMBLE achieves state-of-the-art performance. + +For the part segmentation benchmark, we further report the performance on the intermediate downsampled subpoint clouds in Tab. 3. Additionally, results from PointNeXt + +![](images/edf37c3da05d789880d8cc624a33c280b3651673a443ef1789dfd2d35c68e30a.jpg) +Figure 9. Sampled results of few-point sampling in comparison with APES. Zoom in for optimal clarity. + +[33] are also presented, which is a prominent point cloud learning method that employs FPS for downsampling. It is evident that FPS-based methods exhibit poorer performance when evaluated on intermediate downsampled subpoint clouds. In contrast, a notable observation is that SAMBLE achieves superior performance on intermediate downsampled points. This indicates that the learned sampled points contribute more to the overall performance, while the upsampling layer cannot fully reconstruct the features of the discarded points. Although SAMBLE enables the interpolation-based upsampling and it outperforms the upsampling layer used in APES (see details in Sec. 4.4), there remains potential for further improvement by designing a more meticulously crafted upsampling layer. However, this is another topic and beyond the scope of this paper. + +# 4.3. Few-Point Sampling + +Experiment setting. We further compare our sampling method to previous approaches, including RS, FPS, and more recent learning-based methods such as S-Net, SampleNet, LightTN, APES, and others. The evaluation follows the same framework as in [7, 41, 46]. First, the point cloud is downsampled to a limited number of points, and the resulting subset is then fed into a task network for evaluation. For this comparison, we use the ModelNet40 classification task with the vanilla PointNet network. All sampling methods are evaluated across multiple sampling sizes of $M$ . + +Qualitative and Quantitative Results. Quantitative results are presented in Tab. 4. Note that APES [46] uses FPS to pre-process the input into $2M$ points while we do not. For a fair comparison, additional results of APES without the preprocessing step are also tested and reported. Nonetheless, even without pre-processing, SAMBLE achieves state-of-the-art results in the few-point sampling task as the number of sampled points decreases to extremely smaller ones. + +
MVoxelRSFPS [8]S-NET [7]PST-NET [40]SampleNet [14]MOPS-Net [34]DA-Net [21]LightTN [41]APES [46] (w/ pre-pro.)APES [46] (w/o pre-pro.)SAMBLE
51273.8287.5288.3487.8087.9488.1686.6789.0189.9190.8189.8190.58
25673.5077.0983.6482.3883.1584.2786.6386.2488.2190.4086.7890.18
12868.1556.4470.3477.5380.1180.7586.0685.6786.2689.7784.8790.02
6458.3131.6946.4270.4576.0679.8685.2585.5586.5189.5779.2389.81
3220.0216.3526.5860.7063.9277.3184.2885.1186.1888.5675.6389.45
+ +Table 4. Comparison with other sampling methods. Evaluated on the ModelNet40 classification benchmark with multiple sampling sizes. For APES, we additionally report its performance when pre-processing is not performed for a fair comparison. + +
Indexing Modeiiiiiiivvvivii
Cls.OA (%)93.9293.7893.6393.6693.4094.1194.08
Seg.Cat. mIoU (%)83.9883.8583.6283.5183.4784.1284.22
Ins. mIoU (%)86.1685.9985.7485.6085.4986.3886.46
+ +Table 5. Classification and segmentation performance with different indexing modes. + +
Number of Bins124681012
Cls.OA (%)93.9593.9193.9894.1894.0293.8093.84
Seg.Cat. mIoU (%)84.2284.1484.5184.4084.1983.9884.36
Ins. mIoU (%)86.4686.2886.6786.6186.4886.2386.43
+ +Qualitative results are presented in Fig. 9. For few-point sampling, APES relies on FPS to pre-sample the input into $2M$ points due to its limitations. In contrast, our method preserves better global uniformity, allowing direct few-point sampling from the input while still achieving satisfactory sampled results, as demonstrated in Fig. 9. When sampling very few points, APES tends to concentrate on the sharpest regions, whereas our SAmBLE method preserves better global uniformity throughout the point cloud shape. + +# 4.4. Ablation Study + +In this subsection, our emphasis is directed toward the novel designs introduced within this paper, excluding common topics such as network width. More ablation studies and further design justifications are provided in the supplementary materials to enhance our method's interpretability. + +Different Indexing Modes. Apart from the visualized heatmaps given in Fig. 5, we also report their respective experimental results in Tab. 5. The tests are performed using top-M as the sampling strategy. From it, we can observe that indexing modes vi and vii achieve best performances. + +Number of Bins. As a key parameter in SAMBLE, an ablation study is performed over the number of bins $n_b$ . The results are presented in Tab. 6. Remarkably, increasing the number of bins does not yield improved performance. This phenomenon is likely attributable to the subdivision of shapes into an excessive number of point categories, leading to the gradual diminishment of score disparities across the bins. In our case, $n_b = 6$ and 4 yield the best performance + +Table 6. Classification and segmentation performance with different number of bins. + +
UpsampleInterpolationCross-Attention
Kup=3Kup=8Kup=16
APES (local)82.89 / 85.4082.95 / 85.4482.96 / 85.4283.11 / 85.58
APES (global)83.16 / 85.5383.19 / 85.5983.17 / 85.5583.67 / 85.81
SAMBLE84.51 / 86.6784.35 / 86.4884.31 / 86.4384.36 / 86.44
+ +Table 7. Segmentation results with different upsampling layers on ShapeNet-Part. The number before “/” is the category mIoU, and the number after is the instance mIoU. + +for the classification and segmentation tasks respectively, and we use it for the corresponding experiments. + +Upsampling layer. An important aspect to highlight is the upsampling layer. Most point cloud network models employ neighbor-based interpolation [31, 33, 60] for upsampling, as FPS is used during the downsampling process. However, APES introduces a cross-attention layer for upsampling to address its limitations of overemphasizing edge points, which renders traditional neighbor-based interpolation impractical. In contrast, our method achieves an improved balance between sampling edge points and preserving global uniformity, allowing the use of interpolation operations during upsampling. An ablation study for evaluating various upsampling layers and interpolation with different $K_{up}$ values is conducted, and the results are presented in Table 7. The results show a performance drop for APES when interpolation is used in place of cross-attention, while SAMBLE demonstrates superior performance with it. + +# 5. Conclusion + +In this paper, a novel point cloud sampling method SAMBLE is proposed to learn shape-specific sampling strategies for point cloud shapes. Based on a sparse attention map that integrates both local and global information, multiple indexing modes are designed and explored. By partitioning the points in each shape into bins and learning respective sampling ratios for each bin, shape-specific sampling strategies are acquired for individual point cloud shapes. SAMBLE achieves an optimal trade-off between sampling local details and preserving global uniformity, resulting in improved performance on downstream tasks. For future directions, advancements in upsampling layers could further improve the model's performance. Additionally, adapting the proposed method for point cloud scenes is another promising area to explore. + +# References + +[1] Pyunghwan Ahn, Juyoung Yang, Eojindl Yi, Chanho Lee, and Junmo Kim. Projection-based point convolution for efficient point cloud segmentation. IEEE Access, 10:15348-15358, 2022. 2 +[2] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Semantic segmentation of earth observation data using multimodal and multi-scale deep networks. In *Asian Conference on Computer Vision*, pages 180–196. Springer, 2016. 2 +[3] Alexandre Boulch, Bertrand Le Saux, and Nicolas Audebert. Unstructured point cloud semantic labeling using deep segmentation networks. 3DOR@ Eurographics, 3, 2017. 2 +[4] Can Chen, Luca Zanotti Fragonara, and Antonios Tsourdos. GAPointNet: Graph attention based point neural network for exploiting local feature of point cloud. Neurocomputing, 438:122-132, 2021. 2 +[5] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. PRA-Net: Point relation-aware network for 3d point cloud analysis. IEEE Transactions on Image Processing, 30: 4436-4448, 2021. 6 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2760-2769, 2019. 1, 2, 6, 7, 8 +[8] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 8 +[9] Nico Engel, Vasileios Belagiannis, and Klaus Dietmayer. Point transformer. IEEE Access, 9:134826-134840, 2021. 3, 6 +[10] Fabian Groh, Patrick Wieschollek, and Hendrik PA Lensch. Flex-convolution: Million-scale point-cloud learning beyond grid-worlds. In Asian Conference on Computer Vision (ACCV), pages 105–122. Springer, 2018. 2 +[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. PCT: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 6 +[12] Mingyang Jiang, Yiran Wu, Tianqi Zhao, Zelin Zhao, and Cewu Lu. PointSift: A sift-like network module for 3d point cloud semantic segmentation. arXiv preprint arXiv:1807.00652, 2018. 2 +[13] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5 +[14] Itai Lang, Asaf Manor, and Shai Avidan. SampleNet: Differentiable point cloud sampling. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7578-7588, 2020. 1, 2, 6, 8 +[15] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. DPC: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 2 +[16] Felix Järemo Lawin, Martin Danelljan, Patrik Tosteberg, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Deep projective 3d semantic segmentation. In International Conference on Computer Analysis of Images and Patterns, pages 95-107. Springer, 2017. 2 +[17] Truc Le and Ye Duan. PointGrid: A deep network for 3d shape understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9204-9214, 2018. 2 +[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution on x-transformed points. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2 +[19] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17682-17691, 2023. 2, 6 +[20] Yiqun Lin, Zizheng Yan, Haibin Huang, Dong Du, Ligang Liu, Shuguang Cui, and Xiaoguang Han. FPCov: Learning local flattening for point convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4293-4302, 2020. 2 +[21] Yanan Lin, Yan Huang, Shihao Zhou, Mengxi Jiang, Tianlong Wang, and Yunqi Lei. DA-Net: Density-adaptive down-sampling network for point cloud classification via end-to-end learning. In 2021 4th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), pages 13–18. IEEE, 2021. 2, 8 +[22] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1800-1809, 2020. 2 +[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision (ECCV), pages 657-675. Springer, 2022. 3 +[24] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 2 +[25] Daniel Maturana and Sebastian Scherer. VoxNet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2 +[26] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 2 +[27] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. CurveNet: Curvature-based multitask learning + +deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6 +[28] Ehsan Nezhadarya, Ehsan Taghavi, Ryan Razani, Bingbing Liu, and Jun Luo. Adaptive hierarchical down-sampling for point cloud classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12956-12964, 2020. 2 +[29] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision (ECCV), pages 604-621. Springer, 2022. 3 +[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 2, 6, 7 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 6, 8 +[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6329-6338, 2020. 2 +[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems (NeurIPS), 35:23192-23204, 2022. 2, 6, 7, 8 +[34] Yu Qian, Junhui Hou, Yiming Zeng, Qijian Zhang, Sam Tak Wu Kwong, and Ying He. MOPS-Net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. ArXiv, abs/2005.00383, 2020. 1, 2, 8 +[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 5 +[36] Martin Simonovsky and Nikos Komodakis. Dynamic edge-conditioned filters in convolutional neural networks on graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3693-3702, 2017. 2 +[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3887-3896, 2018. 2 +[38] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. KPCov: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3 +[40] Xu Wang, Yi Jin, Yigang Cen, Congyan Lang, and Yidong Li. PST-Net: Point cloud sampling via point-based transformer. In 11th International Conference on Image and Graphics (ICIG), pages 57–69. Springer, 2021. 2, 8 +[41] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. LighTN: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. IEEE Transactions on Multimedia, 2023. 2, 7, 8 +[42] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2, 6 +[43] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learnable skeleton-aware 3d point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, 2023. 3 +[44] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. DeltaConv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 6 +[45] Chengzhi Wu, Xuelei Bi, Julius Pfrommer, Alexander Cebulla, Simon Mangold, and Jürgen Beyerer. Sim2real transfer learning for point cloud segmentation: An industrial application case on autonomous disassembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4531-4540, 2023. 3 +[46] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5333-5343, 2023. 1, 2, 6, 7, 8 +[47] Chengzhi Wu, Qianliang Huang, Kun Jin, Julius Pfrommer, and Jürgen Beyerer. A cross branch fusion-based contrastive learning framework for point cloud self-supervised learning. In 2024 International Conference on 3D Vision (3DV), pages 528-538. IEEE, 2024. 3 +[48] Chengzhi Wu, Kaige Wang, Zeyun Zhong, Hao Fu, Junwei Zheng, Jiaming Zhang, Julius Pfrommer, and Jurgen Beyerer. Rethinking attention module design for point cloud analysis. In International Conference on Pattern Recognition (ICPR), 2024. 3 +[49] Wenxuan Wu, Zhongang Qi, and Li Fuxin. PointConv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9621-9630, 2019. 2, 6 +[50] Wenxuan Wu, Li Fuxin, and Qi Shan. PointConvFormer: Revenge of the point-based convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21802-21813, 2023. 2 +[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems (NeurIPS), 35:33330-33342, 2022. 3 +[52] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang + +Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4840-4851, 2024. 3 +[53] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 6 +[54] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-GCN for fast and scalable point cloud learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5661-5670, 2020. 2 +[55] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102, 2018. 6 +[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 7 +[57] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 3 +[58] Kuangen Zhang, Ming Hao, Jing Wang, Xinxing Chen, Yuquan Leng, Clarence W de Silva, and Chenglong Fu. Linked dynamic graph cnn: Learning through point cloud by linking hierarchical features. In 2021 27th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 7-12. IEEE, 2021. 2 +[59] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in Neural Information Processing Systems (NeurIPS), 35:27061-27074, 2022. 3 +[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 2, 3, 6, 8 +[61] Yin Zhou and Oncel Tuzel. VoxelNet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4490-4499, 2018. 2 +[62] Wei Zhu, Yue Ying, Jin Zhang, Xiuli Wang, and Yayu Zheng. Point cloud registration network based on convolution fusion and attention mechanism. Neural Processing Letters, pages 1-21, 2023. 2 \ No newline at end of file diff --git a/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/images.zip b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a82e534adc7059b9ef660d8d33a27454d39d8249 --- /dev/null +++ b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:458956140a644c64f44718e2e9a72ff9df4632e61f5077d22947c51db9a62671 +size 631495 diff --git a/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/layout.json b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b843118d7468fbd93523e82b34ee7c9c5e96b504 --- /dev/null +++ b/2025/SAMBLE_ Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity/layout.json @@ -0,0 +1,9106 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 102, + 503, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 102, + 503, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 503, + 140 + ], + "type": "text", + "content": "SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 210, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 210, + 173 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 210, + 173 + ], + "type": "text", + "content": "Chengzhi Wu1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 175, + 208, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 175, + 208, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 208, + 190 + ], + "type": "text", + "content": "Zeyun Zhong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 235, + 161, + 299, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 161, + 299, + 175 + ], + "spans": [ + { + "bbox": [ + 235, + 161, + 299, + 175 + ], + "type": "text", + "content": "Yuxin Wan1*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 221, + 175, + 299, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 175, + 299, + 190 + ], + "spans": [ + { + "bbox": [ + 221, + 175, + 299, + 190 + ], + "type": "text", + "content": "Junwei Zheng" + }, + { + "bbox": [ + 221, + 175, + 299, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 161, + 365, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 161, + 365, + 175 + ], + "spans": [ + { + "bbox": [ + 320, + 161, + 365, + 175 + ], + "type": "text", + "content": "Hao " + }, + { + "bbox": [ + 320, + 161, + 365, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{F u}^{1*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 175, + 386, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 386, + 190 + ], + "type": "text", + "content": "Jiaming Zhang" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 386, + 161, + 473, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 161, + 473, + 175 + ], + "spans": [ + { + "bbox": [ + 386, + 161, + 473, + 175 + ], + "type": "text", + "content": "Julius Pfrommer2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 399, + 175, + 485, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 175, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 399, + 175, + 485, + 190 + ], + "type": "text", + "content": "Jirgen Beyerer" + }, + { + "bbox": [ + 399, + 175, + 485, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 111, + 194, + 331, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 194, + 331, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 331, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 194, + 331, + 209 + ], + "type": "text", + "content": "Karlsruhe Institute of Technology, Germany" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 194, + 495, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 194, + 495, + 209 + ], + "spans": [ + { + "bbox": [ + 354, + 194, + 495, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 354, + 194, + 495, + 209 + ], + "type": "text", + "content": "Fraunhofer IOSB, Germany" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 212, + 460, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 212, + 460, + 223 + ], + "spans": [ + { + "bbox": [ + 143, + 212, + 460, + 223 + ], + "type": "text", + "content": "{chengzhi.wu, zeyun.zhong, junwei.zheng, jiangming.zhang}@kit.edu," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 83, + 225, + 523, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 225, + 523, + 237 + ], + "spans": [ + { + "bbox": [ + 83, + 225, + 523, + 237 + ], + "type": "text", + "content": "{yuxin.wan, hao.fu}@student.kit.edu, {julius.pfrommer, juergen.beyerer}@iosb.fraunhofer.de" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 151, + 262, + 200, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 262, + 200, + 274 + ], + "spans": [ + { + "bbox": [ + 151, + 262, + 200, + 274 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 286, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 297, + 526 + ], + "type": "text", + "content": "Driven by the increasing demand for accurate and efficient representation of 3D data in various domains, point cloud sampling has emerged as a pivotal research topic in 3D computer vision. Recently, learning-to-sample methods have garnered growing interest from the community, particularly for their ability to be jointly trained with downstream tasks. However, previous learning-based sampling methods either lead to unrecognizable sampling patterns by generating a new point cloud or biased sampled results by focusing excessively on sharp edge details. Moreover, they all overlook the natural variations in point distribution across different shapes, applying a similar sampling strategy to all point clouds. In this paper, we propose a Sparse Attention Map and Bin-based Learning method (termed SAMBLE) to learn shape-specific sampling strategies for point cloud shapes. SAMBLE effectively achieves an improved balance between sampling edge points for local details and preserving uniformity in the global shape, resulting in superior performance across multiple common point cloud downstream tasks, even in scenarios with few-point sampling." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 544, + 135, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 135, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 135, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": "Point cloud sampling is a less explored research area within the realm of this data representation. Traditional random sampling (RS) and farthest point sampling (FPS) remain the most commonly employed methods when sampling is required for point cloud learning. With the advancement of neural networks, several methods have emerged for point cloud sampling in a downstream task-oriented learning framework, including S-Net [7], SampleNet [14], MOPS-Net [34], etc. However, these methods first generate a new, smaller-sized point cloud as a proxy rather than directly sampling points from the original input, rendering the" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 259, + 557, + 415 + ], + "blocks": [ + { + "bbox": [ + 316, + 259, + 557, + 415 + ], + "lines": [ + { + "bbox": [ + 316, + 259, + 557, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 259, + 557, + 415 + ], + "type": "image", + "image_path": "16a07d5cb8a61a3c346fea446e86a7c46c17ebb8d56e938aed7cda5bd58c4b45.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. Our method achieves an improved trade-off between sampling local details and preserving global uniformity." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 461, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 555, + 676 + ], + "type": "text", + "content": "techniques akin to black-box neural network models with limited interpretability. Consequently, discerning geometric patterns in their qualitative results becomes challenging, as their outcomes closely resemble those obtained through random sampling. More recently, APES [46] pioneers the direction of using neural networks to learn point-wise sampling scores, with which it subsequently samples points whose scores are higher. However, with its score computation design and the Top-M sampling strategy, APES excessively focuses on local details of edge points, resulting in a deficiency in preserving good global uniformity of the input shapes. Consequently, the interpolation operation becomes impractical during the upsampling process, and the sampling quality of few-point sampling is notably subpar (see Fig. 1). In this paper, we introduce a novel point cloud sampling method that addresses the limitations of prior approaches, aiming to achieve a refined balance between capturing local details and preserving global uniformity." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "The concept originates from rethinking the mathematical characteristics of local details within point cloud shapes. Typically, these local details are represented by edge points" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 231, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 231, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 231, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1342" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 144 + ], + "type": "image", + "image_path": "4a42dd2e458f64fa44e2e9873c157dc9584f431b0abf426ffc333f1f4c0b0269.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 153, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 187 + ], + "type": "text", + "content": "Figure 2. When selecting an equal number of neighbors for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies (numbers on the right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": "that define the shape's outline and sharpest features. Is there a point property that can easily distinguish between different categories, such as edge points and non-edge points? The answer is affirmative. In our investigation, we have uncovered an extremely fundamental yet crucial observation: if point " + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": " is one of the " + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": "-nearest neighbors of point " + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": ", it does not necessarily imply that " + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": " is also among the " + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 54, + 208, + 295, + 327 + ], + "type": "text", + "content": ". Consequently, it leads to the conclusion that the frequency of each point being chosen as a neighbor exhibits variation across a single point cloud." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 329, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 329, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 329, + 295, + 592 + ], + "type": "text", + "content": "We explore and demonstrate the importance of this point property with a simple example as illustrated in Fig. 2. Assume the input point cloud is a simple grid. When selecting 5 neighbors for each point, all three possible cases are given on the left (center point is self-contained as a neighbor). Note that in the triangular and rectangular cases, they each has a \"quantum-entangled\" twin point pair, in which two points share the possibility of being chosen as the neighbor. While an equal number of neighbors is selected for each point in the input point cloud, points at different positions are chosen as neighbors with varying frequencies, as listed on the right of Fig. 2. From it, we can observe that in addition to the edge point and non-edge point categories, there is also another noteworthy point category of close-to-edge points. Moreover, within each category, the points can be further grouped into more sub-categories. Overall, this point property effectively captures the local characteristics of a shape, especially for shape outline and sharp details. Building on it, we propose Sparse Attention Map (SAM) and introduce new methods for computing point-wise sampling scores to effectively balance the trade-off between local and global sampling. See more details in Sec. 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "On the other hand, after the point-wise sampling scores are computed, previous methods employ a Top-M sampling strategy for all point cloud shapes, which exacerbates the issue of oversampling edge points. We argue that the naive top-M sampling strategy may not be optimal across all point cloud shapes for downstream tasks. For example, sampling more non-edge points enhances global uniformity, while sampling more close-to-edge points \"thickens\" the edge, both of which can potentially improve the performance on downstream tasks [46]. To address this, we introduce a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "novel bin-based method to explore better sampling strategies shape-specifically by leveraging all point categories. This approach enables the sampling of points with smaller sampling scores, further optimizing the local-global tradeoff. As a result, our method dynamically adjusts the sampling strategy for each shape, leading to a more tailored and interpretable sampling process for improved performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 156, + 533, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 156, + 533, + 167 + ], + "spans": [ + { + "bbox": [ + 324, + 156, + 533, + 167 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 168, + 553, + 312 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "content": "- We propose a sparse attention map that directly integrates shape local and global information at the attention map level for point cloud sampling, introducing multiple methods for computing point-wise sampling scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 216, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 262 + ], + "type": "text", + "content": "- We present a novel approach for learning bin boundaries to partition points within individual shapes, enabling shape-specific sampling strategies by incorporating additional bin tokens during the attention computation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 263, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 553, + 312 + ], + "type": "text", + "content": "- Our method successfully achieves an improved trade-off between capturing local details and preserving global uniformity for the sampling process, resulting in enhanced performance both qualitatively and quantitatively." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 320, + 400, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 320, + 400, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 400, + 333 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 340, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 568 + ], + "type": "text", + "content": "Point Cloud Sampling. Point cloud sampling is a key process in 3D data handling for simplifying high-resolution dense point clouds. Over the past decades, non-learning-based methods [8, 10, 26] have predominantly been used for point cloud sampling. While Farthest Point Sampling (FPS) [8] is the most widely used one [18, 31, 33, 49, 60], Random Sampling (RS) has also been frequently adopted [10, 32, 61]. More recently, learning-based sampling methods have shown superior performance with task-oriented training. S-Net [7] represents a pioneering work of generating new point coordinates from global representations, while SampleNet [14] introduces a soft projection operation for better point approximation. Following S-Net, multiple learning-based methods have been proposed [21, 28, 40, 41]. MOPS-Net [34] learns a transformation matrix and multiplies it with the original point cloud to generate the sampled one. By employing the attention mechanism to learn point-wise sampling scores, APES [46] captures the edge points in the input point clouds with a strong focus." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Deep Learning on Point Clouds. In contrast to the voxelization-based methods [12, 17, 25] and multi-view-based methods [2, 3, 16, 37], point-based methods deal directly with point clouds. The pioneer studies of PointNet [30] and PointNet++ [31] tackle point clouds through pointwise Multi-Layer Perceptrons (MLPs) and max-pooling operations. Subsequently, other research shifts focus towards constructing more efficient building blocks for local feature extraction, such as convolution-based ones [1, 18, 20, 38, 49, 50, 62] and graph-based ones [4, 15, 22, 24, 36, 42, 54, 58]. More recently, while MLP-based methods like PointNeXt [33] and PointMetaBase [19] have rekin" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1343" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 202 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 202 + ], + "type": "image", + "image_path": "462afef2c30ba826d6a0784e920c27b971073e864edbe4ea303523083b525c3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "type": "text", + "content": "Figure 3. A brief pipeline of our proposed method SAMBLE. It learns shape-specific sampling strategies for point cloud shapes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 244, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 354 + ], + "type": "text", + "content": "dled people's interest, the application of attention mechanisms to point cloud analysis has also garnered widespread attention [9, 11, 39, 43, 45, 48, 57, 60]. For example, PT [51, 52, 60] series improve the model performance by introducing subtraction-based attention blocks, and [48] performs a large ablation study over attention module designs for point cloud processing. In addition, approaches that apply Transformers for point cloud self-supervised learning have also been proposed and explored [23, 29, 47, 57, 59]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 138, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 138, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 138, + 376 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 296, + 431 + ], + "type": "text", + "content": "A brief pipeline of SAMBLE is illustrated in Fig. 3. It consists of three key steps: constructing a sparse attention map, computing point-wise sampling scores, and learning shape-specific sampling strategies through bin partitioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 182, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 182, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 182, + 450 + ], + "type": "text", + "content": "3.1. Sparse Attention Map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 454, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 537 + ], + "type": "text", + "content": "Local and Global Attention Maps. Both local and global attention maps are widely used in point cloud analysis. A global attention map is derived from the application of classical self-attention to point features of all points, while a local attention map concentrates on a point-centered area wherein cross-attention is specifically applied to the central point and its neighbors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": "Denote " + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i" + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": " as the set of " + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": "-nearest neighbors of point " + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": ", the local attention map for " + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 538, + 295, + 562 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 566, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 566, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 71, + 566, + 295, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {i} ^ {l} = \\operatorname {s o f t m a x} \\left(Q (\\mathbf {p} _ {i}) K \\left(\\mathbf {p} _ {i j} - \\mathbf {p} _ {i}\\right) _ {j \\in \\mathcal {S} _ {i}} ^ {\\top} / \\sqrt {d}\\right), \\tag {1}", + "image_path": "5c6cf9ff9861ce54720a317b92e3ffcbbac603381fd3a55dbd13ef31c501485e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "text", + "content": " stand for the linear layers applied on the query and key input, and the square root of the feature dimension count " + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\sqrt{d}" + }, + { + "bbox": [ + 55, + 589, + 295, + 624 + ], + "type": "text", + "content": " serves as a scaling factor [39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 625, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 295, + 649 + ], + "type": "text", + "content": "For the global attention map which is equivalent to taking all points as neighbors for each point, it is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 652, + 295, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 652, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 86, + 652, + 295, + 673 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {g} = \\operatorname {s o f t m a x} \\left(Q (\\mathbf {p} _ {i}) K (\\mathbf {p} _ {j}) _ {i, j \\in \\mathcal {S}} ^ {\\top} / \\sqrt {d}\\right), \\tag {2}", + "image_path": "7cf6f3f43206051b075cb29179277efe47b197690856473783da409957c3f70d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 675, + 229, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 229, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 229, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 675, + 229, + 688 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 675, + 229, + 688 + ], + "type": "text", + "content": " denotes the set of all input points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Sparse Attention Map. Instead of using local or global attention maps solely, we propose sparse attention map," + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 547, + 220 + ], + "type": "image", + "image_path": "0c10a1eaf649b8be3a09d1ddfe454ae27f25e8eda27df7a4609583cf5dca31d4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 230, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 275 + ], + "type": "text", + "content": "Figure 4. Sparse attention map. In each row, " + }, + { + "bbox": [ + 313, + 230, + 555, + 275 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 230, + 555, + 275 + ], + "type": "text", + "content": " cells are selected based on the KNN neighbor indexes for each point. The values of other non-selected cells are all set to 0. Note that the number of cells selected within each column is variable." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "content": "which combines the knowledge from both local and global information, to compute point-wise sampling scores. The idea is illustrated in Fig. 4. After obtaining the global attention map with Eq. 2, " + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "content": "-NN is employed locally to find " + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "content": " neighbors for each point. In this case, " + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "content": " cells are being selected in each row. However, as discussed before, please notice that each point is chosen as a neighbor with varying frequencies. This means while for each row " + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "content": " cells are selected, for each column, the number of selected cells varies. The selected cells are then \"carved\" out to form the sparse attention map, with the values of other non-selected cells being set to 0. More vividly, consider the global attention map as a grid stone slab of size " + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 313, + 291, + 555, + 472 + ], + "type": "text", + "content": ". For carve-based SAM, values of all cells are pre-computed and hidden in the slab grid cells, and only the selected cells are carved out." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 479, + 517, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 517, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 517, + 493 + ], + "type": "text", + "content": "3.2. Computing Point-wise Sampling Score" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": "Indexing Mode. When sampling points, the points are indexed based on the computed point-wise sampling scores. We call the method of computing point-wise sampling scores from the full/sparse attention map as Indexing Mode. With the original full attention map, following APES, there are two possible indexing modes: (i) row standard deviation; and (ii) column sum. For a global attention map " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^g" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": ", denote " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "m_{ij}" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": " as the value of " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": "th row and " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": "th column in " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^g" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": ". These two indexing modes are formulated as modes (i) and (ii) in Tab. 1. To avoid possible confusion, we use notation " + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_o" + }, + { + "bbox": [ + 313, + 498, + 555, + 629 + ], + "type": "text", + "content": " to denote a point only in this subsection." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "With the proposed sparse attention map, there are many other possible indexing modes. As discussed in Sec. 1, to achieve an improved trade-off between sampling edge points and preserving global uniformity, the frequency of each point being chosen as a neighbor, i.e., the number of selected cells in each column is the key. The following indexing modes are designed and explored for comparison:" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1344" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 542, + 174 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 542, + 174 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 542, + 174 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 542, + 174 + ], + "type": "table", + "html": "
Indexing ModeAttention MapFormulaRemark
(i) Row standard deviationFullap0 = fstd{m0j|j = 1,2,...,N}fstd: Computing standard deviation for a set of values
(ii) Column sumFullap0 = ∑i=1N mio
(iii) Row standard deviationSparseap0 = fstd{msj|j ∈ So}So: Set of indices of selected cells in oth row
(iv) Row sumSparseap0 = ∑j=1N msjNon-selected cells are all of 0s
(v) Column sumSparseap0 = ∑i=1N msi0
(vi) Column averageSparseap0 = ∑i=1N msio/n0n0: Number of selected cells in oth column
(vii) Column square-dividedSparseap0 = ∑i=1N msio/n02n0: Number of selected cells in oth column
", + "image_path": "e2905a1e55f9d12452fc2400c93400f50a9134928df767ff20733dc85df2f39a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 182, + 463, + 193 + ], + "lines": [ + { + "bbox": [ + 146, + 182, + 463, + 193 + ], + "spans": [ + { + "bbox": [ + 146, + 182, + 463, + 193 + ], + "type": "text", + "content": "Table 1. Proposed different indexing modes for computing point-wise sampling scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 65, + 205, + 544, + 358 + ], + "blocks": [ + { + "bbox": [ + 65, + 205, + 544, + 358 + ], + "lines": [ + { + "bbox": [ + 65, + 205, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 65, + 205, + 544, + 358 + ], + "type": "image", + "image_path": "b8fdc4e0f0630112c3f6176e6b8ab1097fc70150a7d4112f1a71acc928103360.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 369, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 61, + 369, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 369, + 545, + 381 + ], + "type": "text", + "content": "Figure 5. Point sampling score heatmaps under different indexing modes. Scores are normalized to " + }, + { + "bbox": [ + 61, + 369, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0.5,1)" + }, + { + "bbox": [ + 61, + 369, + 545, + 381 + ], + "type": "text", + "content": " for better visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": "(iii) sparse row standard deviation; (iv) sparse row sum; (v) sparse column sum; (vi) sparse column average; and (vii) sparse column square-divided. Again, for a sparse attention map " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^s" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": ", denote " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "m_{ij}^{s}" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": " as the value of " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": "th row and " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": "th column in " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^s" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": ". For point " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_o" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": ", we denote the set of indexes of the selected " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": " cells (indexes of " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": "-nearest neighbors) in " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": "th row as " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "S_o" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": ", and denote the number of selected cells in " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": "th column as " + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "inline_equation", + "content": "n_o" + }, + { + "bbox": [ + 54, + 396, + 295, + 504 + ], + "type": "text", + "content": ". Details and respective formulas of the proposed indexing modes are listed in Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 295, + 651 + ], + "type": "text", + "content": "Heatmap. To analyze the behavior of each indexing mode, we train a separate model for each mode, ensuring that all other settings remain consistent. The sampling score distributions are visualized as heatmaps in Fig. 5, enhancing the interpretability of our method. From these heatmaps, we can see that both row-standard-deviation-based modes (modes i and iii) concentrate heavily on edge points. However, because they consistently prioritize thin or detailed regions, some areas may be overlooked. In contrast, modes ii and iv show less emphasis on edge points and instead distribute focus across a broader range of points, with a tendency toward other non-edge regions in a biased manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": "More interestingly, the comparison of modes v, vi, and vii, which utilize column-wise information from SAM, reveals distinct sampling preferences and strategies across different point categories. Mode v prioritizes non-edge points, mode vi emphasizes the global shape, and mode" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "type": "text", + "content": "vii focuses slightly more on edge points. This is because edge points typically have a smaller number of " + }, + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "type": "inline_equation", + "content": "n_o" + }, + { + "bbox": [ + 313, + 396, + 555, + 552 + ], + "type": "text", + "content": ". Despite these differences and unique characteristics, all three modes capture the overall shape more uniformly compared to the former four. In our case, we aim to sample edge points without over-emphasizing them. For instance, when sampling detailed areas like chair legs, we want to capture some edge points without selecting them all, while also ensuring that non-edge points are sampled to preserve better global uniformity. Given this balance, we chose mode vii as the primary indexing mode for most of the experiments in the following sections. The detailed ablation study over different indexing modes is presented in Sec. 4.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 561, + 429, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 429, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 429, + 574 + ], + "type": "text", + "content": "3.3. Sampling with Bins" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 579, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 675 + ], + "type": "text", + "content": "After point-wise sampling scores are computed with SAM, points are sampled according to certain rules. The simplest approach is top-M sampling, where points with the highest scores are sampled. In our case, as we aim to enhance the local-global trade-off and leverage all point categories during the sampling process, we suggest employing a bin-based sampling strategy to allow for the possible sampling of certain close-to-edge points or even non-edge points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Bin Partitioning. The process begins with processing the distribution of normalized point-wise sampling scores " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "a_{\\mathbf{p}_i}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " across the shapes within the current batch. Let " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " repre" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1345" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "sent the number of bins used for partitioning, from which " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "n_b - 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " bin boundary values need to be derived from the distribution. During each training step, a vector " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\nu_c = (\\nu_1, \\nu_2, \\dots, \\nu_{n_b - 1})" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " is computed based on the point score distribution, ensuring an equitable division of points across all shapes in the current batch. Note that while " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\nu_c" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " facilitates an even division at the batch level, the points within each individual shape are not necessarily evenly partitioned according to the batch-based bin boundary values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 240 + ], + "type": "text", + "content": "During the training, for the first iteration, we directly use the boundary values derived from the first batch of data as the dynamic boundary values. Subsequently, since the second iteration, boundaries are updated adaptively in a momentum-based manner:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 120, + 247, + 294, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 247, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 120, + 247, + 294, + 259 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\nu} _ {t} = \\gamma \\boldsymbol {\\nu} _ {t - 1} + (1 - \\gamma) \\boldsymbol {\\nu} _ {c}, \\tag {3}", + "image_path": "df42190ace239e14ba2bc1b705370b1dfd786eac451500c5e61da3aab158d01c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\nu_{t - 1}" + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": " stands for the bin partitioning boundaries used in the last iteration, and " + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\nu_{t}" + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": " is the updated dynamic boundaries used for the current iteration. " + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\gamma \\in (0,1)" + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": " is the momentum update factor. With updated boundary values " + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\nu_{t}" + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": ", points in each shape are divided into " + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": " subsets of " + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_1,\\mathcal{B}_2,\\dots ,\\mathcal{B}_{n_b}\\}" + }, + { + "bbox": [ + 55, + 263, + 296, + 335 + ], + "type": "text", + "content": " based on their sampling scores." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 337, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 296, + 469 + ], + "type": "text", + "content": "The core idea presented here is that, instead of using pre-engineered bin boundary values, we employ adaptive learning for bin boundaries, and they are gradually learned from the entire training dataset. These values are intended to evenly partition the distribution of point-wise sampling scores across all shapes and points in the training dataset. Consequently, for each individual shape, the acquired boundary values can effectively partition its points into bins with a shape-specific strategy, capturing the unique characteristics of the shape while maintaining a degree of proximity to other shapes within the dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": "Tokens for Learning Bin Weights. With points already being partitioned into bins for each shape, the next step is to learn a shape-specific sampling strategy, i.e., to learn shape-specific sampling weights for each bin. Inspired by ViT [6], ViT [13], and Mask3D [35] —which leverage additional tokens during the computation of attention maps to extract and convey information across the entire feature map or specific groups of points or pixels — we introduce additional tokens specifically for learning bin sampling weights. In our case, attention maps are computed shape-specific during the downsampling process, facilitating the learning of bin sampling weights also in a shape-specific manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Using the former proposed bin partitioning method, points in each shape are partitioned into " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " subsets of " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_1, \\mathcal{B}_2, \\dots, \\mathcal{B}_{n_b}\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". The sampling weight " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega_j" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " for bin " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j (j = 1, 2, \\dots, n_b)" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " is established based on the distinctive features of each shape. Fig. 6 gives the network structure of our proposed downsampling layer and illustrates the idea of using additional tokens. " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " bin tokens are introduced during the attention computation, where each token corresponds to a" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 449, + 298 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 449, + 298 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 449, + 298 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 449, + 298 + ], + "type": "image", + "image_path": "8bf79882e586b50f754d2fbdcdb06df97f3a7b6a6656f82b05bb01cb76aa2178.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "text", + "content": "Figure 6. Network structure of our proposed downsampling layer. Block " + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "text", + "content": ": Points in each shape are partitioned into " + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "text", + "content": " bins. Block " + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "text", + "content": ": Masking the split-out point-to-token sub-attention map. Block " + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 306, + 555, + 351 + ], + "type": "text", + "content": ": Learned bin sampling weights." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 450, + 71, + 552, + 298 + ], + "blocks": [ + { + "bbox": [ + 450, + 71, + 552, + 298 + ], + "lines": [ + { + "bbox": [ + 450, + 71, + 552, + 298 + ], + "spans": [ + { + "bbox": [ + 450, + 71, + 552, + 298 + ], + "type": "image", + "image_path": "1418140f004eac1091bc0941f5abc2e6d88060e319a3ecdb4005f751aabad8f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "text", + "content": "specific bin. As shown in Fig. 6, the bin tokens are initially concatenated with the input point-wise features for Key and Value. Subsequently, the combined features are subjected to a cross-attention mechanism with the original point-wise features as Query. The attention map is split into two parts of a point-to-point sub-attention map and a point-to-token sub-attention map. For the point-to-point attention map, the methods proposed in Sec. 3.1 and Sec. 3.2 are applied to it to obtain point-wise sampling scores. Note that in this case, the row-wise sum is not exactly equal to 1 but still very close to 1 since " + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "text", + "content": " is of a very small quantity compared to " + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "text", + "content": ". With computed point scores, dynamic boundary values " + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "text", + "content": " are obtained for bin partitioning. Using the information regarding the allocation of points to respective bins, a mask operation is performed on the point-to-token sub-attention map as illustrated in Block B of Fig. 6. The sampling weights " + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\omega_j" + }, + { + "bbox": [ + 313, + 366, + 555, + 570 + ], + "type": "text", + "content": " are then subsequently acquired with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 574, + 555, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 574, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 361, + 574, + 555, + 612 + ], + "type": "interline_equation", + "content": "\\omega_ {j} = \\operatorname {R e L U} \\left(\\frac {1}{\\beta_ {j}} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {B} _ {j}} m _ {\\mathbf {p} _ {i}, \\mathcal {B} _ {j}}\\right), \\tag {4}", + "image_path": "0e6592939e837a844bc3016658bdbd8f0d36d75fba4deac28047b4fa93c0137c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\beta_{j}" + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "content": " stands for the number of points in bin " + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j" + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "inline_equation", + "content": "m_{\\mathbf{p}_i,\\mathcal{B}_j}" + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "content": " represents the element in the energy matrix corresponding to point " + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "content": " in row and " + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j" + }, + { + "bbox": [ + 313, + 616, + 555, + 652 + ], + "type": "text", + "content": " in column." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "In-Bin Point Sampling. For each shape, by considering the number of points contained within bins " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\beta = (\\beta_{1},\\beta_{2},\\ldots ,\\beta_{n_{b}})" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " alongside the determined bin sampling weights " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega = (\\omega_{1},\\omega_{2},\\dots ,\\omega_{n_{b}})" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", the specific numbers of points to be sampled from each bin " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\kappa = (\\kappa_{1},\\kappa_{2},\\dots ,\\kappa_{n_{b}})" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1346" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 69, + 549, + 206 + ], + "blocks": [ + { + "bbox": [ + 63, + 69, + 549, + 206 + ], + "lines": [ + { + "bbox": [ + 63, + 69, + 549, + 206 + ], + "spans": [ + { + "bbox": [ + 63, + 69, + 549, + 206 + ], + "type": "image", + "image_path": "04e6a6edfeedc4ef67d6f258093d03c472da16ebc5780c957ab8eb6faf6d04c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 216, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 216, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 555, + 240 + ], + "type": "text", + "content": "Figure 7. Qualitative results of our proposed SAMBLE. Apart from the sampled results, sampling score heatmaps and bin histograms along with bin sampling ratios are also given. All shapes are from the test set. Zoom in for optimal visual clarity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": "need to be determined. Direct multiplication of " + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": " does not yield a sum that aligns with the total number of down-sampled points " + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": " required by the network structure. To address this discrepancy, a scaling method is applied to first scale bin sampling weights " + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": ". Furthermore, to prevent " + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "inline_equation", + "content": "\\kappa_{j}" + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": " from surpassing the available point number " + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "inline_equation", + "content": "\\beta_{j}" + }, + { + "bbox": [ + 54, + 254, + 294, + 361 + ], + "type": "text", + "content": " in any bin, any excess points are proportionately redistributed to other bins that have not been fully sampled. The detailed algorithm is presented in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "content": "Finally, within bin " + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j" + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\kappa_{j}" + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "content": " points are selected through random sampling with priors. The sampling probability " + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\rho_{\\mathbf{p}_i}" + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "content": " is obtained via softmax operation over the normalized point sampling score " + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "inline_equation", + "content": "a_{\\mathbf{p}_i}" + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "content": " with a temperature parameter " + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 361, + 295, + 410 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 413, + 295, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 413, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 125, + 413, + 295, + 446 + ], + "type": "interline_equation", + "content": "\\rho_ {\\mathbf {p} _ {i}} = \\frac {e ^ {a _ {\\mathbf {p} _ {i}} / \\tau}}{\\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {B} _ {j}} e ^ {a _ {\\mathbf {p} _ {i}} / \\tau}}. \\tag {5}", + "image_path": "982d595b0c56510bcf5306e1aee68bd4180aa27113556867bf5294f670e2e6de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 473, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 295, + 568 + ], + "type": "text", + "content": "Like most related works, such as S-NET [7], SampleNet [14], and APES [46], SAMBLE is specifically designed for point cloud shapes. To ensure a fair comparison, we conduct experiments using the same base network architecture as APES [46] on standard point cloud shape datasets, including ModelNet40 and ShapeNet-Part. It is important to note that point cloud sampling is not a standalone task; its effectiveness must be validated through downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 141, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 141, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 141, + 588 + ], + "type": "text", + "content": "4.1. Classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "Experiment Setting. ModelNet40 classification benchmark [53] includes 12,311 CAD models across 40 categories. For a fair comparison, we use the official train-test split, with 9,843 for training and 2,468 for testing. Points are uniformly sampled from the mesh surface and normalized to the unit sphere. Only 3D coordinates are used as input, with random scaling, rotation, and shifting applied for data augmentation. We use " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "n_b = 6" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " bins for point partitioning. The momentum update factor " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 0.99" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " for updating bin boundary values. The temperature parameter " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 251, + 545, + 444 + ], + "blocks": [ + { + "bbox": [ + 326, + 251, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 326, + 251, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 326, + 251, + 545, + 444 + ], + "type": "table", + "html": "
MethodCls. OA (%)Seg.
Cat. mIoU (%)Ins. mIoU (%)
PointNet [30]89.280.483.7
PointNet++ [31]91.981.985.1
SpiderCNN [55]92.482.485.3
DGCNN [42]92.982.385.2
PointConv [49]92.582.885.7
PT1 [9]92.8-85.9
PT2 [60]93.783.786.6
PCT [11]93.2-86.4
PRA-Net [5]93.783.786.3
CurveNet [27]93.8-86.6
DeltaConv [44]93.8-86.6
PointNeXt [33]93.284.486.7
PointMetaBase [19]-84.386.7
APES (local) [46]93.583.185.6
APES (global) [46]93.883.785.8
SAMBLE94.284.586.7
", + "image_path": "0ac59e965714554d93dbfc4650af35feb213ab46315fbd77a215393bc2ea6f9d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 451, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 485 + ], + "type": "text", + "content": "Table 2. Classification and segmentation results on the ModelNet40 and ShapeNet-Part benchmarks. In comparison with other SOTA methods that also only use raw point cloud data as input." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Qualitative and Quantitative Results. Qualitative results of SAMBLE are presented in Fig. 7, including sampling score heatmaps, learned bin partitioning strategy with bin sampling ratios, and final sampled results. From it, we can observe that SAMBLE effectively samples sufficient edge points to capture the shape structure. Moreover, it maintains better global uniformity by avoiding excessive focus on edge points, particularly in those thin or sharp regions like chair legs. Logged shape bin histograms confirm the learning of shape-specific sampling strategies. More visualization results are provided in the supplementary materials, highlighting an intriguing pattern where shapes of the same category exhibit similar histogram distributions and sampling strategies. Overall, SAMBLE successfully achieves an improved trade-off between sampling edge points and preserving shape global uniformity. Quantitative results are provided in Tab. 2. Our method performs better than previous approaches and achieves state-of-the-art performance." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1347" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 298, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 298, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 298, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 298, + 242 + ], + "type": "image", + "image_path": "9e3dd4e4ef9d27d6a8291912b297c5aa9817c11124f0f4783b423fde1932411a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 276 + ], + "type": "text", + "content": "Figure 8. Segmentation results of our proposed SAMBLE in comparison with APES. All shapes are from the test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 286, + 282, + 337 + ], + "blocks": [ + { + "bbox": [ + 69, + 286, + 282, + 337 + ], + "lines": [ + { + "bbox": [ + 69, + 286, + 282, + 337 + ], + "spans": [ + { + "bbox": [ + 69, + 286, + 282, + 337 + ], + "type": "table", + "html": "
MethodPointNeXtSAMBLE
Point Number2048102451220481024512
Cat. mIoU (%)84.4083.7982.7784.5184.8485.04
Ins. mIoU (%)86.7086.1885.1886.6786.9387.12
", + "image_path": "29278ad8cb1c14ef6b139995f649ebac3f2094f4b154848e29e7e8195e2df9b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 296, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 369 + ], + "type": "text", + "content": "Table 3. Additional segmentation performances evaluated on the intermediate downsampled sub-point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 390, + 143, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 143, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 143, + 403 + ], + "type": "text", + "content": "4.2. Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "text", + "content": "Experiment setting. The ShapeNet-Part dataset [56] is used for 3D object part segmentation. It includes 16,880 3D models across 16 categories, with 14,006 models for training and 2,874 for testing. Each category contains 2-6 parts, totaling 50 distinct parts. We use the sampled point sets produced in [30] for a fair comparison with prior work. For evaluation metrics, we report both category mIoU and instance mIoU. We use " + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "inline_equation", + "content": "n_b = 4" + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "text", + "content": " bins for point partitioning. The momentum update factor " + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\gamma = 0.99" + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "text", + "content": " for updating bin boundary values. The temperature parameter " + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 55, + 409, + 296, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 676 + ], + "type": "text", + "content": "Qualitative and Quantitative Results. Qualitative results are presented in Fig. 8, where we observe that, compared to APES, which tends to focus excessively on edge points, SAMBLE achieves a significantly improved balance between sampling edge points and preserving shape global uniformity. For example, SAMBLE demonstrates a more balanced use of non-edge points, as seen in the chair seat and lamp base, reflecting a thoughtful sampling strategy that accounts for different point categories and provides a more comprehensive representation of the overall shape. The quantitative results in Tab. 2 further highlight that SAMBLE achieves state-of-the-art performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "For the part segmentation benchmark, we further report the performance on the intermediate downsampled subpoint clouds in Tab. 3. Additionally, results from PointNeXt" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 242 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 242 + ], + "type": "image", + "image_path": "edf37c3da05d789880d8cc624a33c280b3651673a443ef1789dfd2d35c68e30a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "type": "text", + "content": "Figure 9. Sampled results of few-point sampling in comparison with APES. Zoom in for optimal clarity." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 292, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 473 + ], + "type": "text", + "content": "[33] are also presented, which is a prominent point cloud learning method that employs FPS for downsampling. It is evident that FPS-based methods exhibit poorer performance when evaluated on intermediate downsampled subpoint clouds. In contrast, a notable observation is that SAMBLE achieves superior performance on intermediate downsampled points. This indicates that the learned sampled points contribute more to the overall performance, while the upsampling layer cannot fully reconstruct the features of the discarded points. Although SAMBLE enables the interpolation-based upsampling and it outperforms the upsampling layer used in APES (see details in Sec. 4.4), there remains potential for further improvement by designing a more meticulously crafted upsampling layer. However, this is another topic and beyond the scope of this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 432, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 432, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 432, + 491 + ], + "type": "text", + "content": "4.3. Few-Point Sampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "content": "Experiment setting. We further compare our sampling method to previous approaches, including RS, FPS, and more recent learning-based methods such as S-Net, SampleNet, LightTN, APES, and others. The evaluation follows the same framework as in [7, 41, 46]. First, the point cloud is downsampled to a limited number of points, and the resulting subset is then fed into a task network for evaluation. For this comparison, we use the ModelNet40 classification task with the vanilla PointNet network. All sampling methods are evaluated across multiple sampling sizes of " + }, + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Qualitative and Quantitative Results. Quantitative results are presented in Tab. 4. Note that APES [46] uses FPS to pre-process the input into " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " points while we do not. For a fair comparison, additional results of APES without the preprocessing step are also tested and reported. Nonetheless, even without pre-processing, SAMBLE achieves state-of-the-art results in the few-point sampling task as the number of sampled points decreases to extremely smaller ones." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1348" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 137 + ], + "type": "table", + "html": "
MVoxelRSFPS [8]S-NET [7]PST-NET [40]SampleNet [14]MOPS-Net [34]DA-Net [21]LightTN [41]APES [46] (w/ pre-pro.)APES [46] (w/o pre-pro.)SAMBLE
51273.8287.5288.3487.8087.9488.1686.6789.0189.9190.8189.8190.58
25673.5077.0983.6482.3883.1584.2786.6386.2488.2190.4086.7890.18
12868.1556.4470.3477.5380.1180.7586.0685.6786.2689.7784.8790.02
6458.3131.6946.4270.4576.0679.8685.2585.5586.5189.5779.2389.81
3220.0216.3526.5860.7063.9277.3184.2885.1186.1888.5675.6389.45
", + "image_path": "b1d630604f315e47f28ce83c0e1e8c8e66e13dd857030c4c486cc6585ab4d9e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 183, + 296, + 230 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 553, + 168 + ], + "type": "text", + "content": "Table 4. Comparison with other sampling methods. Evaluated on the ModelNet40 classification benchmark with multiple sampling sizes. For APES, we additionally report its performance when pre-processing is not performed for a fair comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 183, + 296, + 230 + ], + "lines": [ + { + "bbox": [ + 57, + 183, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 57, + 183, + 296, + 230 + ], + "type": "table", + "html": "
Indexing Modeiiiiiiivvvivii
Cls.OA (%)93.9293.7893.6393.6693.4094.1194.08
Seg.Cat. mIoU (%)83.9883.8583.6283.5183.4784.1284.22
Ins. mIoU (%)86.1685.9985.7485.6085.4986.3886.46
", + "image_path": "6b08dc1c9efe62c2ce706ab0d7b3a28792b3984b5ecf2e6031b30151ae0cdf0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 270, + 296, + 317 + ], + "blocks": [ + { + "bbox": [ + 55, + 237, + 296, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 261 + ], + "type": "text", + "content": "Table 5. Classification and segmentation performance with different indexing modes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 270, + 296, + 317 + ], + "lines": [ + { + "bbox": [ + 57, + 270, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 270, + 296, + 317 + ], + "type": "table", + "html": "
Number of Bins124681012
Cls.OA (%)93.9593.9193.9894.1894.0293.8093.84
Seg.Cat. mIoU (%)84.2284.1484.5184.4084.1983.9884.36
Ins. mIoU (%)86.4686.2886.6786.6186.4886.2386.43
", + "image_path": "23fed35f2519d6752a38d4325e3e162bf29c94da11f66231f06355b674a1edae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 361, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 469 + ], + "type": "text", + "content": "Qualitative results are presented in Fig. 9. For few-point sampling, APES relies on FPS to pre-sample the input into " + }, + { + "bbox": [ + 54, + 361, + 295, + 469 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 54, + 361, + 295, + 469 + ], + "type": "text", + "content": " points due to its limitations. In contrast, our method preserves better global uniformity, allowing direct few-point sampling from the input while still achieving satisfactory sampled results, as demonstrated in Fig. 9. When sampling very few points, APES tends to concentrate on the sharpest regions, whereas our SAmBLE method preserves better global uniformity throughout the point cloud shape." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 476, + 149, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 149, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 149, + 489 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 494, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 554 + ], + "type": "text", + "content": "In this subsection, our emphasis is directed toward the novel designs introduced within this paper, excluding common topics such as network width. More ablation studies and further design justifications are provided in the supplementary materials to enhance our method's interpretability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 616 + ], + "type": "text", + "content": "Different Indexing Modes. Apart from the visualized heatmaps given in Fig. 5, we also report their respective experimental results in Tab. 5. The tests are performed using top-M as the sampling strategy. From it, we can observe that indexing modes vi and vii achieve best performances." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Number of Bins. As a key parameter in SAMBLE, an ablation study is performed over the number of bins " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": ". The results are presented in Tab. 6. Remarkably, increasing the number of bins does not yield improved performance. This phenomenon is likely attributable to the subdivision of shapes into an excessive number of point categories, leading to the gradual diminishment of score disparities across the bins. In our case, " + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "n_b = 6" + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": " and 4 yield the best performance" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 314, + 183, + 554, + 243 + ], + "blocks": [ + { + "bbox": [ + 55, + 325, + 296, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 347 + ], + "type": "text", + "content": "Table 6. Classification and segmentation performance with different number of bins." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 183, + 554, + 243 + ], + "lines": [ + { + "bbox": [ + 314, + 183, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 183, + 554, + 243 + ], + "type": "table", + "html": "
UpsampleInterpolationCross-Attention
Kup=3Kup=8Kup=16
APES (local)82.89 / 85.4082.95 / 85.4482.96 / 85.4283.11 / 85.58
APES (global)83.16 / 85.5383.19 / 85.5983.17 / 85.5583.67 / 85.81
SAMBLE84.51 / 86.6784.35 / 86.4884.31 / 86.4384.36 / 86.44
", + "image_path": "186fc3701a3cbe427de15843bd892601a6a7f6b3aea46debe8a81259d51973ab.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 285 + ], + "type": "text", + "content": "Table 7. Segmentation results with different upsampling layers on ShapeNet-Part. The number before “/” is the category mIoU, and the number after is the instance mIoU." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 300, + 553, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 325 + ], + "type": "text", + "content": "for the classification and segmentation tasks respectively, and we use it for the corresponding experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": "Upsampling layer. An important aspect to highlight is the upsampling layer. Most point cloud network models employ neighbor-based interpolation [31, 33, 60] for upsampling, as FPS is used during the downsampling process. However, APES introduces a cross-attention layer for upsampling to address its limitations of overemphasizing edge points, which renders traditional neighbor-based interpolation impractical. In contrast, our method achieves an improved balance between sampling edge points and preserving global uniformity, allowing the use of interpolation operations during upsampling. An ablation study for evaluating various upsampling layers and interpolation with different " + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "inline_equation", + "content": "K_{up}" + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": " values is conducted, and the results are presented in Table 7. The results show a performance drop for APES when interpolation is used in place of cross-attention, while SAMBLE demonstrates superior performance with it." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 529, + 388, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 388, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 388, + 541 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 548, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 713 + ], + "type": "text", + "content": "In this paper, a novel point cloud sampling method SAMBLE is proposed to learn shape-specific sampling strategies for point cloud shapes. Based on a sparse attention map that integrates both local and global information, multiple indexing modes are designed and explored. By partitioning the points in each shape into bins and learning respective sampling ratios for each bin, shape-specific sampling strategies are acquired for individual point cloud shapes. SAMBLE achieves an optimal trade-off between sampling local details and preserving global uniformity, resulting in improved performance on downstream tasks. For future directions, advancements in upsampling layers could further improve the model's performance. Additionally, adapting the proposed method for point cloud scenes is another promising area to explore." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1349" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 114, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 156 + ], + "type": "text", + "content": "[1] Pyunghwan Ahn, Juyoung Yang, Eojindl Yi, Chanho Lee, and Junmo Kim. Projection-based point convolution for efficient point cloud segmentation. IEEE Access, 10:15348-15358, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 295, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 295, + 202 + ], + "type": "text", + "content": "[2] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Semantic segmentation of earth observation data using multimodal and multi-scale deep networks. In *Asian Conference on Computer Vision*, pages 180–196. Springer, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 294, + 236 + ], + "type": "text", + "content": "[3] Alexandre Boulch, Bertrand Le Saux, and Nicolas Audebert. Unstructured point cloud semantic labeling using deep segmentation networks. 3DOR@ Eurographics, 3, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 280 + ], + "type": "text", + "content": "[4] Can Chen, Luca Zanotti Fragonara, and Antonios Tsourdos. GAPointNet: Graph attention based point neural network for exploiting local feature of point cloud. Neurocomputing, 438:122-132, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 294, + 323 + ], + "type": "text", + "content": "[5] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. PRA-Net: Point relation-aware network for 3d point cloud analysis. IEEE Transactions on Image Processing, 30: 4436-4448, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 294, + 390 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "text", + "content": "[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2760-2769, 2019. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "text", + "content": "[8] Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y Zeevi. The farthest point strategy for progressive image sampling. IEEE Transactions on Image Processing, 6(9):1305-1315, 1997. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 512 + ], + "type": "text", + "content": "[9] Nico Engel, Vasileios Belagiannis, and Klaus Dietmayer. Point transformer. IEEE Access, 9:134826-134840, 2021. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "text", + "content": "[10] Fabian Groh, Patrick Wieschollek, and Hendrik PA Lensch. Flex-convolution: Million-scale point-cloud learning beyond grid-worlds. In Asian Conference on Computer Vision (ACCV), pages 105–122. Springer, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "text", + "content": "[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. PCT: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "text", + "content": "[12] Mingyang Jiang, Yiran Wu, Tianqi Zhao, Zelin Zhao, and Cewu Lu. PointSift: A sift-like network module for 3d point cloud semantic segmentation. arXiv preprint arXiv:1807.00652, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[13] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[14] Itai Lang, Asaf Manor, and Shai Avidan. SampleNet: Differentiable point cloud sampling. In Proceedings of the" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7578-7588, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 553, + 139 + ], + "type": "text", + "content": "[15] Itai Lang, Dvir Ginzburg, Shai Avidan, and Dan Raviv. DPC: Unsupervised deep point correspondence via cross and self construction. In 2021 International Conference on 3D Vision (3DV), pages 1442-1451. IEEE, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[16] Felix Järemo Lawin, Martin Danelljan, Patrik Tosteberg, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Deep projective 3d semantic segmentation. In International Conference on Computer Analysis of Images and Patterns, pages 95-107. Springer, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 239 + ], + "type": "text", + "content": "[17] Truc Le and Ye Duan. PointGrid: A deep network for 3d shape understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9204-9214, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 282 + ], + "type": "text", + "content": "[18] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution on x-transformed points. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 284, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 553, + 338 + ], + "type": "text", + "content": "[19] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17682-17691, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "text", + "content": "[20] Yiqun Lin, Zizheng Yan, Haibin Huang, Dong Du, Ligang Liu, Shuguang Cui, and Xiaoguang Han. FPCov: Learning local flattening for point convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4293-4302, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 393, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 553, + 458 + ], + "type": "text", + "content": "[21] Yanan Lin, Yan Huang, Shihao Zhou, Mengxi Jiang, Tianlong Wang, and Yunqi Lei. DA-Net: Density-adaptive down-sampling network for point cloud classification via end-to-end learning. In 2021 4th International Conference on Pattern Recognition and Artificial Intelligence (PRAI), pages 13–18. IEEE, 2021. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[22] Zhi-Hao Lin, Sheng-Yu Huang, and Yu-Chiang Frank Wang. Convolution in the cloud: Learning deformable kernels in 3d graph convolution networks for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1800-1809, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 514, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 558 + ], + "type": "text", + "content": "[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision (ECCV), pages 657-675. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "text", + "content": "[24] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "type": "text", + "content": "[25] Daniel Maturana and Sebastian Scherer. VoxNet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "text", + "content": "[26] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, Computer Laboratory, 2003. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[27] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. CurveNet: Curvature-based multitask learning" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1350" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 149 + ], + "type": "text", + "content": "[28] Ehsan Nezhadarya, Ehsan Taghavi, Ryan Razani, Bingbing Liu, and Jun Luo. Adaptive hierarchical down-sampling for point cloud classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12956-12964, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 295, + 194 + ], + "type": "text", + "content": "[29] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision (ECCV), pages 604-621. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "text", + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 293 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 295, + 348 + ], + "type": "text", + "content": "[32] Haozhe Qi, Chen Feng, Zhiguo Cao, Feng Zhao, and Yang Xiao. P2B: Point-to-box network for 3d object tracking in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6329-6338, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 414 + ], + "type": "text", + "content": "[33] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. PointNext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems (NeurIPS), 35:23192-23204, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 459 + ], + "type": "text", + "content": "[34] Yu Qian, Junhui Hou, Yiming Zeng, Qijian Zhang, Sam Tak Wu Kwong, and Ying He. MOPS-Net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. ArXiv, abs/2005.00383, 2020. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 514 + ], + "type": "text", + "content": "[35] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 568 + ], + "type": "text", + "content": "[36] Martin Simonovsky and Nikos Komodakis. Dynamic edge-conditioned filters in convolutional neural networks on graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3693-3702, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 623 + ], + "type": "text", + "content": "[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3887-3896, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 689 + ], + "type": "text", + "content": "[38] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. KPCov: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6411–6420, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 712 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 139 + ], + "type": "text", + "content": "[40] Xu Wang, Yi Jin, Yigang Cen, Congyan Lang, and Yidong Li. PST-Net: Point cloud sampling via point-based transformer. In 11th International Conference on Image and Graphics (ICIG), pages 57–69. Springer, 2021. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "text", + "content": "[41] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. LighTN: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. IEEE Transactions on Multimedia, 2023. 2, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 228 + ], + "type": "text", + "content": "[42] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph CNN for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "text", + "content": "[43] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learnable skeleton-aware 3d point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[44] Ruben Wiersma, Ahmad Nasikun, Elmar Eisemann, and Klaus Hildebrandt. DeltaConv: anisotropic operators for geometric deep learning on point clouds. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 555, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 382 + ], + "type": "text", + "content": "[45] Chengzhi Wu, Xuelei Bi, Julius Pfrommer, Alexander Cebulla, Simon Mangold, and Jürgen Beyerer. Sim2real transfer learning for point cloud segmentation: An industrial application case on autonomous disassembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4531-4540, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 436 + ], + "type": "text", + "content": "[46] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5333-5343, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "type": "text", + "content": "[47] Chengzhi Wu, Qianliang Huang, Kun Jin, Julius Pfrommer, and Jürgen Beyerer. A cross branch fusion-based contrastive learning framework for point cloud self-supervised learning. In 2024 International Conference on 3D Vision (3DV), pages 528-538. IEEE, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "type": "text", + "content": "[48] Chengzhi Wu, Kaige Wang, Zeyun Zhong, Hao Fu, Junwei Zheng, Jiaming Zhang, Julius Pfrommer, and Jurgen Beyerer. Rethinking attention module design for point cloud analysis. In International Conference on Pattern Recognition (ICPR), 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "type": "text", + "content": "[49] Wenxuan Wu, Zhongang Qi, and Li Fuxin. PointConv: Deep convolutional networks on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9621-9630, 2019. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "text", + "content": "[50] Wenxuan Wu, Li Fuxin, and Qi Shan. PointConvFormer: Revenge of the point-based convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21802-21813, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 689 + ], + "type": "text", + "content": "[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems (NeurIPS), 35:33330-33342, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[52] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1351" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 652 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4840-4851, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "type": "text", + "content": "[53] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[54] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-GCN for fast and scalable point cloud learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5661-5670, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 282 + ], + "type": "text", + "content": "[55] Yifan Xu, Tianqi Fan, Mingye Xu, Long Zeng, and Yu Qiao. SpiderCNN: Deep learning on point sets with parameterized convolutional filters. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "text", + "content": "[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (TOG), 35(6):1-12, 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "type": "text", + "content": "[57] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-BERT: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 462 + ], + "type": "text", + "content": "[58] Kuangen Zhang, Ming Hao, Jing Wang, Xinxing Chen, Yuquan Leng, Clarence W de Silva, and Chenglong Fu. Linked dynamic graph cnn: Learning through point cloud by linking hierarchical features. In 2021 27th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 7-12. IEEE, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "text", + "content": "[59] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. PointM2AE: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in Neural Information Processing Systems (NeurIPS), 35:27061-27074, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 565, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 294, + 608 + ], + "type": "text", + "content": "[61] Yin Zhou and Oncel Tuzel. VoxelNet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4490-4499, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 609, + 294, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 294, + 652 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 294, + 652 + ], + "type": "text", + "content": "[62] Wei Zhu, Yue Ying, Jin Zhang, Xiuli Wang, and Yayu Zheng. Point cloud registration network based on convolution fusion and attention mechanism. Neural Processing Letters, pages 1-21, 2023. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1352" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_content_list.json b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0db086dd0a1fb193698715380c9e471999ef5077 --- /dev/null +++ b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Claudia Cuttano1 Gabriele Trivigno1 Gabriele Rosi1,2 Carlo Masone1,2 Giuseppe Averta1,2 \n1 Politecnico di Torino 2 Focoos AI \n{name.surname}@polito.it {name.surname}@focoos.ai", + "bbox": [ + 102, + 179, + 892, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RVOS Streaming processing", + "text_level": 1, + "bbox": [ + 401, + 268, + 583, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63c93735b7990a4d36388399bcb1f003cc06a254889db45e3028f16e59d58cb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 287, + 905, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segment Anything 2", + "text_level": 1, + "bbox": [ + 94, + 402, + 246, + 417 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/23f7f2df1b62e57e251a3aa8998e9903248be9b7e759af68c4f014fa13e963fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SPATIAL\nPROMPTTEXT\nPROMPTCORRECTION\nMECHANISMTEMPORAL\nFEATURES
✓✗human✗
", + "bbox": [ + 96, + 420, + 357, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9f657dd9d09d5070003b0f31fc61b067dca54fdf013c62e4543661c774b0802.jpg", + "image_caption": [ + "Figure 1. SAMWISE. Our approach infuses knowledge about natural language in the Segment-Anything 2 model, adding explicit temporal cues in the feature extraction for the task of streaming-based Referring Video Segmentation (RVOS). We use a learnable mechanism to mitigate the so-called tracking bias, i.e. SAM2 tendency to overlook a correct object once it becomes identifiable, due to its ongoing tracking of a different object. Our design enables effective streaming processing for RVOS, exploiting the memory from previous frames to propagate past context. The figure shows an example where the target object is not present in the first frame, leading SAM2 to start tracking the wrong one. Afterwards, when the correct object appears, our learnable correction mechanisms guide SAM2 to switch its tracking focus. By adding in its features the notion of temporal evolution, the model is able to recognize that the new object is more aligned with the provided textual query. Finally, we exploit SAM2 tracking skills and robustness to occlusions to keep following the object." + ], + "image_footnote": [], + "bbox": [ + 375, + 388, + 638, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d848b0dc5e95e32decd1ee44d07ef7e066202465b54bf5d29b987e25ff14670.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 393, + 769, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SAMWISE (ours)", + "text_level": 1, + "bbox": [ + 769, + 401, + 901, + 416 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/5e7e7fffafc9ef4d8a777c6d6449d81fcec3b60bd01bf47da3688d003a8186aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SPATIAL\nPROMPTTEXT\nPROMPTCORRECTION\nMECHANISMTEMPORAL\nFEATURES
✓✓learnable✓
", + "bbox": [ + 653, + 417, + 903, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 609, + 326, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation (RVOS) relies on natural language expressions to segment an object in a video clip. Existing methods restrict reasoning either to independent short clips, losing global context, or process the entire video offline, impairing their application in a streaming fashion. In this work, we aim to surpass these limitations and design an RVOS method capable of effectively operating in streaming-like scenarios while retaining contextual information from past frames. We build upon the Segment-Anything 2 (SAM2) model, that provides robust segmentation and tracking capabilities and is naturally suited for streaming processing. We make SAM2 wiser, by empowering it with natural language understanding and explicit temporal modeling at the feature extraction stage, without fine-tuning its weights, and without outsourcing modality interaction to external models. To this end, we introduce a novel adapter module that injects temporal information", + "bbox": [ + 86, + 643, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and multi-modal cues in the feature extraction process. We further reveal the phenomenon of tracking bias in SAM2 and propose a learnable module to adjust its tracking focus when the current frame features suggest a new object more aligned with the caption. Our proposed method, SAMWISE, achieves state-of-the-art across various benchmarks, by adding a negligible overhead of less than $5M$ parameters. Code is available at https://github.com/ClaudiaCuttano/SAMWISE.", + "bbox": [ + 511, + 609, + 908, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring video segmentation (RVOS) [9, 17, 24, 38, 43, 48] aims at segmenting and tracking specific objects of interest within video content, guided by natural language expressions [3, 10, 28]. Existing RVOS methods are mostly based on a divide and conquer paradigm, where the video is divided into shorter clips that are processed independently [3, 39, 43]. However, as demonstrated by MeViS [7], this", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3395", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "solution fails in examples that require taking into account long-term motion and global context. As a workaround to handle this challenge, the state-of-the-art method [11] processes the entire video in an offline fashion, first modeling trajectories of all instances throughout the entire clip and then selecting the most appropriate one. Albeit effective, this approach is not applicable when the model has access only to a portion of the video, for example when the data at inference time are presented in a streaming fashion or due to limitations in the computational resources. The trade-off of these two paradigms is schematized in Fig. 1. To this end, OnlineRefer [42] introduced a context propagation scheme for online RVOS but relies solely on past context from a single frame, limiting its ability to capture long-term dependencies. In this work, we investigate how to exploit the memory from past frames to design an RVOS method capable of retaining global context while operating within a streaming paradigm, i.e., without requiring access to the whole video at once. This idea is inspired by the recent release of Segment-Anything 2 (SAM2) [35], a foundational model that has shown impressive capabilities in various Video Segmentation tasks thanks to a memory bank that allows to leverage long-range past information. Since SAM2 operates in a streaming fashion, extending this method to enable context-aware streaming processing in RVOS would appear a natural step. However, this entails some non-trivial challenges:", + "bbox": [ + 89, + 90, + 485, + 500 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "i) Text understanding. SAM2 original design accounts only for spatial prompts (e.g. points) and lacks mechanisms to interpret semantic prompts like text, which require reasoning over visual and textual modalities. While we are the first to address the challenge of adding textual prompts to SAM2, previous methods have explored this problem for SAM-1 at image-level. These solutions [19, 50] delegate visual-textual interaction to an off-the-shelf large VLM (like BEIT-3 [41], LLaVa [21]), which generates a multimodal embedding that is used to prompt SAM-1.", + "ii) Temporal modeling. To segment the referred object throughout the video, it must be first recognized and then tracked. While the latter requires matching objects visual appearance across adjacent frame, the recognition problem entails modeling temporal evolution to reason over actions that unfold over multiple frames. However, SAM2 extracts frame features independently, lacking such reasoning.", + "iii) Tracking bias. In RVOS, the target object might be unrecognizable during certain time intervals, due to occlusions, presence of multiple instances or forthcoming actions, as in the first frames of Fig. 1. In such cases, SAM2 may start tracking an incorrect object that partially matches the textual prompt, and persist in following it, leading to what we denote as tracking bias. While SAM2 original design allows for a user to manually correct the prediction by providing a new prompt, such a strategy is not applicable in" + ], + "bbox": [ + 89, + 501, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks without a human-in-the-loop like RVOS.", + "bbox": [ + 511, + 90, + 821, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim at making SAM2 wiser, by addressing these limitations without fine-tuning SAM2 weights, thereby preserving its original capabilities, and without outsourcing modality interaction to external, heavy models. To overcome challenges $i$ and $ii$ , we design a learnable Adapter [12] module, named Cross-Modal Temporal Adapter (CMT), with two key principles in mind: a) enabling mutual interaction between visual and linguistic modalities; and b) encoding temporal cues into visual features. Then, to generate a prompt, we follow [21, 51] and employ a learnable MLP to project the sentence embedding for the SAM2 Mask Decoder, which then outputs the final segmentation mask. In this way, we can exploit SAM2 tracking capability to segment an object given a textual query across the video. Finally, to mitigate the tracking bias problem iii), we introduce a lightweight Conditional Memory Encoder (CME) which detects when a candidate object, aligned with the text, appears in the frame, thus enabling SAM2 to dynamically refocus its tracking to the correct object as it becomes distinguishable.", + "bbox": [ + 511, + 107, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Summarizing, this paper contributes with the following:", + "bbox": [ + 511, + 410, + 883, + 426 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present SAMWISE, the first method that integrates natural language knowledge into SAM2 in an end-to-end solution tailored to address the challenges of RVOS. We introduce a novel adapter, namely Cross Modal Temporal (CMT) Adapter, which purposefully models temporal evolution and multi-modal interaction;", + "- We provide insight into the functioning of SAM2, highlighting the phenomenon of tracking bias, and introduce a learnable module (Conditional Memory Encoder) to adjust tracking based on new information;", + "- Our methods achieve state-of-the-art results both on traditional RVOS benchmarks (Ref-Youtu-VOS [38], RefDAVIS [17]), as well as the more challenging MeViS [7], without compromising SAM2 capabilities and adding less than $5M$ learnable parameters." + ], + "bbox": [ + 513, + 426, + 906, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 671, + 660, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Video Segmentation. In RVOS, the goal is to segment an object in a clip described with natural language [8, 17, 24]. Earlier works adapted image-based methods [2, 9, 17, 48], or used a spatio-temporal memory to attend to masks of previous frames [31, 38]. Subsequent works employ a DETR-like [4] structure to process multiple frames and text embeddings [3, 10, 29, 43]. All these methods process short clips independently, thus losing global context.", + "bbox": [ + 511, + 703, + 906, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, [7] showed how traditional RVOS benchmarks lack challenging captions that require to disambiguate between instances and their actions, as well as occlusions and dynamic queries, highlighting how they could be solved even with image-based methods. The MeViS dataset [7] tar", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3396", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/741ff6debc471c20241f49667fc2203e80c648e1e6c9e427d6dfbedc44499b16.jpg", + "image_caption": [ + "Figure 2. Overview of SAMWISE. We build on a frozen SAM2 and a frozen Text Encoder to segment images in video given a textual description. We incorporate the Cross-Modal Temporal Adapter (CMT) into the text and visual encoders at every intermediate layer $k$ to model temporal dynamics within visual features while contaminating each modality with the other. Then, we extract the [CLS] and verb embeddings, namely Contextual and Motion prompts, from the adapted textual features and project them through a learnable MLP. The final embedding is used to prompt the Mask Decoder, which outputs the segmentation mask. Finally, the Conditional Memory Encoder detects when a new candidate object, aligned with the caption, appears in the frame, enabling SAM2 to dynamically refocus its tracking." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 898, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gets these scenarios, with challenging examples that previous image or clip-based methods fail to address. To this end, a few works proposed offline methods to explicitly model multiple object trajectories [11, 28], with the latter representing the state-of-the-art on MeViS. Concurrently, OnlineRefer [42] proposed a first attempt towards an online RVOS setting, with a query propagation scheme. However, its effectiveness is limited as predictions are based on a single frame. Our method builds on this paradigm by leveraging SAM2 memory bank to encode long-range past context.", + "bbox": [ + 88, + 415, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-prompted Segment-Anything. Recent works have provided solutions to adapt SAM-1 for text-prompted segmentation. Grounded SAM [36] employs a two step pipeline where GroundingDINO [25] generates bounding boxes for SAM-1 to produce segmentation masks. Applying such pipeline in RVOS is problematic, as potential errors in the first frame are propagated throughout the whole video. To directly prompt SAM-1, RefSAM [20] exploits a projection layer to map the textual embedding into the prompt space, while [1, 19, 45] resort to large off-the-shelf VLM to generate a multi-modal embedding that is used to prompt SAM-1. Both solutions finetune the Mask Decoder, thereby compromising its capabilities on its original task. In contrast, our work is the first to propose an end-to-end model that incorporates textual knowledge within SAM2 without fine-tuning nor relying on external models.", + "bbox": [ + 89, + 574, + 483, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-Trained Knowledge Transfer. In recent years, the release of powerful pretrained models has sparked interest in the question of how to extend their skills to novel tasks, as full fine-tuning becomes increasingly impractical with growing model sizes [16, 33]. A powerful strat", + "bbox": [ + 88, + 824, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "egy to address this problem relies on using Adapters [12], small trainable modules that enable efficient adaptation of pre-trained models. Following this paradigm, recent studies have explored adapting CLIP [34] for downstream tasks. At the image level, [44] inserts Transformer Decoder blocks within CLIP encoders, which entail costly Self-Attentions on all tokens. For video tasks, [40] places independent adapter modules within each encoder, whereas [15, 16, 27, 47] rely on a weight-sharing mechanism to project both modalities in a shared sub-space. Nevertheless, as features of each modality are independently extracted, none of these adapters allows explicit feature interaction, unlike our CMT, which also incorporates temporal modeling. Lastly, all these works start from a model that already includes a text encoder (CLIP), whereas ours is the first to propose an adapter for the Segment-Anything 2 model to add textual understanding, achieving robust performances while introducing less than $5\\mathrm{M}$ parameters.", + "bbox": [ + 511, + 415, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SAMWISE", + "text_level": 1, + "bbox": [ + 513, + 698, + 633, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setting. Given an input video $\\mathcal{V} = \\{I_t\\}_{t=1}^{T_V}$ with $T_V$ frames and a referring expression, we aim to predict a set of binary masks $S = \\{s_t\\}_{t=1}^{T_V}$ , $s_t \\in \\mathbb{R}^{H \\times W}$ of the referred object. We tokenize the textual query in a set of $L$ words, $E = \\{e_l\\}_{l=1}^L$ , and add a global sentence representation token [CLS]. The tokens are then processed using a frozen text encoder to extract language features $\\mathcal{E} \\in \\mathbb{R}^{L \\times C_t}$ . We process videos in a streaming fashion, collecting clips of $T$ frames as they are available. Throughout the rest of the section, we use $T$ to indicate clip length.", + "bbox": [ + 511, + 728, + 906, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. We first provide a brief discussion of the SAM2", + "bbox": [ + 511, + 885, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3397", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03701d1352d3461a2bc78a46886ee2ef59a1fcfe00bc94c936f4597ebf4e2e32.jpg", + "image_caption": [ + "Figure 3. Architecture of our Cross Modal Temporal (CMT) Adapter, made up of Hierarchical Selective Attention (HSA) to model temporal cues, a Visual-to-Text Attention (VTA) and Text-to-Visual Attention (VTA) modules." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model (Sec. 3.1). We then outline the pipeline of our proposed SAMWISE, starting from the prompting strategy in Sec. 3.2. In Sec. 3.3, we detail our novel Cross-Modal Temporal Adapter. Lastly, in Sec. 3.5, we discuss our learnable correction strategy, named Conditional Memory Encoder, to address the issue of tracking bias.", + "bbox": [ + 89, + 318, + 483, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Background: Segment-Anything", + "text_level": 1, + "bbox": [ + 89, + 420, + 377, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Segment-Anything Model 2 (SAM2) builds upon SAM-1 [18] to tackle the task of Promptable Video Object Segmentation, i.e., tracking an object in a video given a textual prompt. Following SAM-1, it consists of an image encoder, a Prompt Encoder and a Mask Decoder, which combines the image and prompt embeddings to predict segmentation masks. To enable video processing, SAM2 comes with a few modifications: $i$ ) the original ViT backbone is replaced by Hiera [37], roughly 3 times faster, which processes frames independently to provide hierarchical visual features. Hereinafter, we refer to them as memory-less features $\\mathcal{F}$ ; $ii$ ) frame embeddings are not directly fed to the Mask Decoder, but they are first conditioned on memories of past predictions from a Memory Bank. We refer to these conditioned features as memory features $\\mathcal{F}_{mem}$ . Lastly, $iii$ ) once the mask for the current frame is predicted, the Memory Encoder updates the Memory Bank. By design, SAM2 handles video frames as they become available, progressively encoding the past in its Memory Bank. We argue that this streaming approach is especially valuable in RVOS, enabling reasoning over a wide temporal horizon.", + "bbox": [ + 89, + 441, + 483, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Prompting SAM2", + "text_level": 1, + "bbox": [ + 89, + 771, + 264, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guide the SAM2 decoder, we use a Contextual Prompt, $\\mathcal{E}_C\\in \\mathbb{R}^{1\\times C_t}$ , which encodes the high-level semantic information for the given text query, emphasizing the essential aspects of the query while downplaying less relevant elements. To this end, we employ the [CLS] embedding of text features, $\\mathcal{E}$ . Furthermore, we also introduce a second prompt, the Motion Prompt $\\mathcal{E}_M\\in \\mathbb{R}^{1\\times C_t}$ , which cap", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures action-related cues by using verb embeddings from $\\mathcal{E}$ . These prompts are concatenated and projected through a learnable three-layer MLP:", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\rho = \\mathbf {W} _ {\\text {p r o m p t}} \\left(\\operatorname {C A T} \\left[ \\mathcal {E} _ {C}, \\mathcal {E} _ {M} \\right]\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 148, + 906, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, the provided prompts encode both subject-related and motion-based information. Given that in our task the textual prompt is not referred a-priori to any particular frame, we prompt SAM2 at each frame, so that the model has to balance the influence of tracking while also considering the content of each frame. We discuss more in depth this aspect in Sec. 3.5.", + "bbox": [ + 511, + 175, + 906, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cross-Modal Temporal Adapter", + "text_level": 1, + "bbox": [ + 511, + 290, + 797, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An adapter consists of a linear down-projection $(\\mathbf{W}_{\\mathrm{down}})$ to a bottleneck dimensionality, followed by an up-projection back $(\\mathbf{W}_{\\mathrm{up}})$ in the original space, separated by a non-linear activation function $\\sigma$ . Formally, given an input feature $\\mathbf{x} \\in \\mathbb{R}^{1 \\times d}$ , the adapter function is defined as:", + "bbox": [ + 511, + 311, + 905, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA d a p t e r (\\mathbf {x}) = \\mathbf {x} + \\sigma (\\mathbf {x} \\mathbf {W} _ {\\text {d o w n}}) \\mathbf {W} _ {\\text {u p}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 398, + 906, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build on this popular Adapter framework [12] and propose a novel Cross-Modal Temporal Adapter (CMT) (see Fig. 3) which models temporal dynamics within visual features while enriching each modality with the other. Formally, given the visual features in a clip $\\mathcal{F}^k\\in \\mathbb{R}^{T\\times H_k\\times W_k\\times C_k}$ and the textual features $\\mathcal{E}^k\\in \\mathbb{R}^{L\\times C}$ extracted at layer $k$ of the image and text encoders, respectively, the CMT can be formulated as:", + "bbox": [ + 511, + 426, + 905, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA d a p t e r \\left(\\mathcal {F} ^ {k}\\right) = \\mathcal {F} ^ {k} + h \\left(\\mathcal {F} ^ {k} \\mathbf {W} _ {\\text {d o w n , v}}, \\mathcal {E} ^ {k} \\mathbf {W} _ {\\text {d o w n , t}}\\right) \\mathbf {W} _ {\\text {u p , v}}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 556, + 895, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A d p a t e r} \\left(\\mathcal {E} ^ {k}\\right) = \\mathcal {E} ^ {k} + h \\left(\\mathcal {E} ^ {k} \\mathbf {W} _ {\\text {d o w n , t}}, \\mathcal {F} ^ {k} \\mathbf {W} _ {\\text {d o w n , v}}\\right) \\mathbf {W} _ {\\text {u p , t}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 577, + 903, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathrm{down},\\mathrm{v}}$ , $\\mathbf{W}_{\\mathrm{down},\\mathrm{t}}$ , $\\mathbf{W}_{\\mathrm{up},\\mathrm{v}}$ , $\\mathbf{W}_{\\mathrm{up},\\mathrm{t}}$ are modality specific down- and up-projections weights and $h$ is our proposed adapter function. The adapter output is summed with the original features, allowing the model to retain the original encoding while incorporating temporal and cross-modal reasoning. We integrate the Cross-Modal Temporal Adapter (CMT) into the frozen text and visual encoders at every intermediate layer $k$ . In the following paragraphs we detail the temporal and cross-modal adaptation functions, which are tightly coupled in our Adapter module.", + "bbox": [ + 511, + 609, + 905, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Adaptation. Our approach aims to embed motion cues directly into the frame-level features of SAM2. Previous works based on Adapters either perform self-attention (SA) over all tokens in a clip [16], which is costly, or restrict the attention to the temporal axis for each pixel [22, 23]. We observe that, within a video, object motion across adjacent frames typically spans a localized region of the image [32]. Consequently, a given element of the feature map primarily benefits from interactions with its", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3398", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/465da27ee6136a28cf0059b52de94bf6729403a9b200dc8e7284f2576e3d77c0.jpg", + "image_caption": [ + "Figure 4. Scheme of our Hierarchical Selective Attention (HSA), modeling temporal evolution of features in our adapter." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "spatial and temporal neighbors, rather than requiring long-range connections across the entire feature map. Building on this intuition, we introduce a Hierarchical Selective Attention (HSA) mechanism, illustrated in Fig. 4. By modeling interactions among spatially and temporally proximal regions, HSA reduces unnecessary computations while capturing motion-based context.", + "bbox": [ + 88, + 294, + 483, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, at layer $k$ , given the set of feature maps for a $T$ -frames clip: $\\mathcal{F}^k \\in \\mathbb{R}^{T \\times H_k \\times W_k \\times C_k}$ , we decompose this feature volume into non-overlapping, 3-D spatio-temporal patches of size $T \\times P \\times P$ , obtaining $N = H_k W_k / P^2$ sub-volumes. These sub-volumes, considered pixelwise, can be represented as a set of tokens $F^k[n] = \\left\\{x_{i,j,t}^{k,n} \\in \\mathbb{R}^{C_k} : i \\in 1, \\dots, P, j \\in 1, \\dots, P, t \\in 1, \\dots, T\\right\\}$ . To encode spatio-temporal positioning, to each vector we add a spatial $(e[i,j])$ and a temporal $(e[t])$ sinusoidal positional embeddings, in 2-D and 1-D formats, respectively. Specifically: $x_{i,j,t}^{k,n} = x_{i,j,t}^{k,n} + e[i,j] + e[t]$ . Each sub-volume contains $M = P^2 * T$ tokens, on which we perform self-attention as follows:", + "bbox": [ + 89, + 401, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i, j, t} ^ {k, n} := S A \\left(\\left\\{\\mathbf {x} _ {i ^ {\\prime}, j ^ {\\prime}, t ^ {\\prime}} ^ {k, n} \\right\\} _ { \\begin{array}{c} i ^ {\\prime} = 1.. P \\\\ j ^ {\\prime} = 1.. P \\\\ t ^ {\\prime} = 1.. T \\end{array} }\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 618, + 482, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each layer $k$ of the feature extraction process, the patch size $P$ is progressively scaled, as depicted in Fig. 4-d. This scaling adapts the sub-volume to the hierarchy of feature resolution, encoding information at multiple scales.", + "bbox": [ + 89, + 681, + 483, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Modal Adaptation. To unify text and visual representations, we encourage modality interaction from early stages of the feature extraction process through two symmetric operations: Visual-to-Text Attention (VTA) and Text-to-Visual Attention (TVA).", + "bbox": [ + 89, + 748, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within the former, each visual feature, already enriched with temporal information through the HSA, attends to the full textual expression, allowing the model to identify candidate regions within the image based on both categorical details (e.g., the subject described in the text) and motion", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6707a5f2777873552bdefb177637112beb9491d1926fab2c9d5813c89a3701e1.jpg", + "image_caption": [ + "Figure 5. Cross Modal Temporal Adapter: we show via PCA that our CMT provides contextualized visual features based on the given textual prompt, compared to SAM2 original ones." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cues (e.g., actions), facilitating early alignment with the prompt, as visible in Fig. 5.", + "bbox": [ + 511, + 277, + 903, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, at layer $k$ , we consider the feature of each frame in the clip, i.e. $\\mathcal{F}^k[t] \\in \\mathbb{R}^{H_k \\times W_k \\times C_k}$ , $t = 1, \\ldots, T$ , and the set of textual embeddings $\\mathcal{E}^k \\in \\mathbb{R}^{L \\times C}$ to compute:", + "bbox": [ + 511, + 309, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {k} [ t ] := \\mathcal {F} ^ {k} [ t ] * C A \\left(\\mathcal {F} ^ {k} [ t ], \\mathcal {E} ^ {k}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 364, + 903, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In parallel, as the meaning of a caption can shift significantly depending on the visual content of the associated image [6], we aim at contextualizing the textual query with the semantics provided by the visual modality. To this end, the TVA progressively enriches the linguistic tokens $\\mathcal{E}^k\\in \\mathbb{R}^{L\\times C}$ with information from the visual feature maps, averaged over the video clip $\\mathcal{F}_{avg}^k$", + "bbox": [ + 511, + 393, + 905, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} ^ {k} := \\mathcal {E} ^ {k} * C A \\left(\\mathcal {E} ^ {k}, \\mathcal {F} _ {\\text {a v g}} ^ {k}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 512, + 903, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mask prediction", + "text_level": 1, + "bbox": [ + 511, + 550, + 674, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the end of the feature extraction process, we obtain the adapted visual and linguistic features, respectively $\\mathcal{E}$ and $\\mathcal{F}$ . To perform the final prediction, we extract the prompt $\\rho$ as in Eq. (1), while the Memory Attention module generates the memory features $\\mathcal{F}_{mem}$ by conditioning the visual features $\\mathcal{F}$ on past predictions from the Memory Bank. The prompt $\\rho$ is fed into the frozen Mask Decoder $\\mathcal{D}_{dec}$ , which generates the output mask $\\mathcal{P}_M \\in \\mathbb{R}^{1 \\times H \\times W}$ and the mask token $\\tau_m \\in \\mathbb{R}^{1 \\times d}$ , i.e. an embedding representing the segmented object. Formally:", + "bbox": [ + 511, + 573, + 905, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\tau_ {m}, \\mathcal {P} _ {M} = \\mathcal {D} _ {d e c} (\\mathcal {F} _ {m e m}, \\rho), \\\\ \\mathcal {Y} = \\mathcal {P} _ {M} > 0, \\end{array} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 734, + 903, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{V} \\in \\mathbb{R}^{1 \\times H \\times W}$ denotes the output binary segmentation mask. Finally, the Memory Encoder updates the memory bank with $\\mathcal{P}_M$ .", + "bbox": [ + 511, + 777, + 903, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Conditional Memory Encoder", + "text_level": 1, + "bbox": [ + 511, + 833, + 781, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We identify as tracking bias the phenomenon of SAM2 tracking the wrong object when the correct one is not yet identifiable in the video, and persist in following it. This", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3399", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/217775a55d75ecef3c6baeb7615ed5e4edead1bbcea8c5000c8cdd24470f57a4.jpg", + "image_caption": [ + "Figure 6. Effect of our Conditional Memory Encoder. The caption above requires disambiguating multiple instances of the same class (e.g., \"cat\") by identifying a specific action (e.g., \"climbing\"). Since none of the instances perform this action initially, the model begins tracking the wrong instance and fails to correct itself once the target object performs the action. The top section visualizes the effect of this tracking bias in the memory features of SAM2. Our CME detects that the cat that starts climbing is more aligned with the caption, and encodes its presence in the memory bank, allowing SAM2 to switch its focus." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bias, as exemplified in Fig. 6, is encoded in the memory features, which are propagated to subsequent frames through the Memory Encoder. On the other hand, we observe that the memory-less features: i) contain an unbiased representation of the current frames, ii) are aligned with the textual prompt via our CMT (cf. Fig. 5), and iii) can thus be used to propose candidate instances that match the prompt without being biased by past predictions. Building on these intuitions, we derive a memory-less token $\\tau_{l}$ from a cross-attention between the unbiased feature maps and the prompt. Such token represents a summary of the visual features that match the prompt. The idea is to compare it with the mask token $\\tau_{m}$ generated by the Mask Decoder, to detect when they represent different objects, i.e., to detect when SAM2 is tracking an object that is not the one currently most aligned with the caption. Formally:", + "bbox": [ + 88, + 482, + 483, + 724 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {l} = C A (\\mathcal {F}, \\rho) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 736, + 482, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that we initialize (and keep frozen) the weights of the cross-attention with those from SAM2 Mask Decoder. We introduce a small learnable module, named Conditional Memory Encoder (CME), to detect such situations. When a new object is detected, a naive solution would be to compute its mask and use it to re-prompt the model at the given frame, just like a user would do, forcing SAM2 to switch its prediction. However, since the prediction computed on the memory-less features does not have access to past video", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "context, it might generate false positives. Thus, we propose a soft assignment, obtained by encoding the masks of both objects in the memory bank. Essentially, the CME allows SAM2 to 'see' other objects beyond the currently tracked one, and balance the influence of past context with new information, to select the one that fits the prompt the most. In detail, our CME, illustrated in Fig. 6-bottom, concatenates the two tokens $\\tau_{m},\\tau_{l}$ with a learnable decision token [DEC], and performs a self-attention followed by a Linear classifier:", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ z _ {D E C}, z _ {M T}, z _ {M L} \\right] = S A \\left(\\left[ [ D E C ], \\tau_ {m} [ t ], \\tau_ {l} [ t ] \\right]\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 250, + 906, + 281 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {d e t e c t} = \\phi (z _ {D E C})\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 282, + 751, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi$ is a linear function $\\mathbb{R}^d\\to \\mathbb{R}^1$ . When detecting a candidate text-aligned object, (i.e., $p_{\\text{detect}} > 0.5$ ), instead of directly feeding the predicted output mask $\\mathcal{P}_m$ to the Memory Encoder, our module computes the unbiased output mask, namely $\\mathcal{P}_L\\in \\mathbb{R}^{1\\times H\\times W}$ , to fuse it with $\\mathcal{P}_m$ :", + "bbox": [ + 511, + 309, + 906, + 387 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {l} = \\mathcal {D} _ {d e c} \\left(\\mathcal {F}, \\rho\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 398, + 764, + 415 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (h, w) = \\mathbb {1} (h, w) [ h, w: \\mathcal {P} _ {l} > 0 ] \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 417, + 903, + 434 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} = \\lambda * \\mathcal {P} _ {l} \\circ \\mathcal {M} + \\mathcal {P} _ {m} \\circ (1 - \\mathcal {M})\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 436, + 834, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{M}(h, w)$ is a binary mask whose value is zero except for the pixels corresponding to the object, and $\\lambda$ is an hyperparameter weighing the influence of the memoryless prediction. The resulting mask $\\mathcal{P}$ is fed to the Memory Encoder. We train the CME via self-supervision with a standard Cross-Entropy loss, by providing examples where the memory-less features highlight different objects w.r.t the one currently tracked. We discuss in detail our training protocol in the Supp. Mat..", + "bbox": [ + 511, + 465, + 906, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental results", + "text_level": 1, + "bbox": [ + 511, + 616, + 714, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate our method on MeVis [7], Ref-Youtube-VOS [2] and Ref-Davis [17]. MeViS includes 2,006 videos and features a total of 28K annotations that capture various aspects of motion. Ref-Youtu-VOS enhances the original YouTube-VOS benchmark by incorporating textual descriptions. It contains a total of 3,978 videos and approximately 15K language expressions. RefDAVIS17 builds upon DAVIS17 dataset, adding more than 1.5K linguistic annotations to 90 videos.", + "bbox": [ + 511, + 647, + 906, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We utilize standard evaluation metrics, region similarity $(\\mathcal{J})$ , contour accuracy $(\\mathcal{F})$ , and their average $(\\mathcal{J} \\& \\mathcal{F})$ . For MeViS and Ref-Youtu-VOS we conduct the evaluation using the official challenge servers; for Ref-DAVIS17, we used the official evaluation code.", + "bbox": [ + 511, + 787, + 906, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We employ Hiera-B [37] as visual extractor. As text encoder, we experiment with", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3400", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/842a4e5b8d74ae0d5493958ee1aa18a061b55f9968acced1d07b13b593dc460a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVisual EncoderText EncoderTotal ParamsMeViSRef-YouTube-VOSRef-DAVIS17
J&FJFJ&FJFJ&FJF
Large VLM based
LISA [19] [CVPR'24]ViT-HLLaVa7 B37.235.139.453.953.454.364.862.267.3
VISA [45] [ECCV'24]ViT-HChat-UniVi7 B43.540.746.361.559.863.269.466.372.5
One-Token-Seg-All [1] [NIPS'24]ViT-HPhi-33.8 B42.339.445.261.760.263.367.763.871.5
MTTR [3] [CVPR'22]V-Swin TRoBERTa-30.028.831.255.354.056.6---
TCE-RVOS [14] [WACV'24]ResNet-50RoBERTa----59.658.360.859.456.562.4
ReferFormer [43] [CVPR'22]V-Swin BRoBERTa237 M31.029.832.262.961.364.661.158.164.1
SOC [28] [NIPS'23]V-Swin BRoBERTa220 M---66.064.167.964.261.067.4
OnlineRefer [42] [ICCV'23]Swin LRoBERTa232 M32.331.533.163.561.665.564.861.667.7
LMPM [7] [ICCV'23]Swin TRoBERTa195 M37.234.240.2------
RefSAM [20] [arXiv]ViT-BT53 B---58.457.459.462.159.065.3
DsHmp [11] [CVPR'24]V-Swin BRoBERTa339 M---67.165.069.164.961.768.1
DsHmp [11] [CVPR'24]Swin TRoBERTa272 M46.443.049.8------
MUTR [46] [AAAI'24]V-Swin BRoBERTa250 M---67.565.469.666.462.870.0
GroundingDINO [25]+SAM2Hiera-BBERT240 M37.734.940.557.555.659.566.462.869.9
SAMWISE (ours)Hiera-BCLIP-B150 M48.345.451.267.265.269.368.565.671.5
SAMWISE (ours)Hiera-BRoBERTa202 M49.546.652.469.267.870.670.667.474.5
", + "bbox": [ + 91, + 88, + 906, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison of SAMWISE against state-of-the-art RVOS methods on MeViS, Ref-Youtu-VOS and Ref-DAVIS datasets. We further include methods based on large VLMs for comparison. Bold and underline indicate the two top results.", + "bbox": [ + 89, + 356, + 906, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CLIP [34] and RoBERTa [26]. We note that the text encoder and SAM2 weights are entirely frozen and we train only the Adapters and the CME module (4.2M parameters when using CLIP and 4.9M with RoBERTa). Following [10, 20, 28, 42, 43], we undergo pre-training for 6 epochs on RefCOCO+/g [30, 49] with a learning rate at 1e-4 and fine-tune on Ref-Youtu-VOS [38] for 4 epochs with a learning rate of 1e-5, using the Adam optimizer. The model trained on the Ref-Youtu-VOS is directly evaluated on DAVIS-17 [17]. On MeViS [7], we train for 1 epoch. We set $T = 8$", + "bbox": [ + 88, + 397, + 480, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 89, + 561, + 230, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare against existing works in the literature, and ablate our contributions. In the Supp. Mat. we report additional qualitative results and ablations.", + "bbox": [ + 89, + 585, + 482, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. To assess the validity of our approach, we divide the experimental comparison in the following categories:", + "bbox": [ + 89, + 636, + 482, + 667 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Standard RVOS methods: we compare against recent relevant works in RVOS. The main comparison is w.r.t. the previous state-of-the-art, namely DsHmp [11];", + "- Methods with Context propagation: OnlineRefer [42] was the first to propose this setting. RefSAM [20] relies on SAM1 to provide frame-level masks, and then propagates the mask token to subsequent frames. A baseline that we propose is GroundingDINO + SAM2, where we use the popular grounded detector to provide boxes for the first frame, and let SAM2 track the object;", + "- Large VLM based: Although these methods [1, 19, 45] are not comparable to ours, or previous ones, in terms of model size, we include them in the table to provide an interesting reference of performance." + ], + "bbox": [ + 89, + 667, + 482, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with standard RVOS methods. Traditional", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RVOS methods, such as ReferFormer [43] and MTTR [3], suffer a significant performance drop on the MeViS benchmark, as they are unable to solve queries which require to model long-term context. An exception is represented by LMPM and its follow up work DsHmp, which represents the state-of-the-art: these methods process the entire video clip at once, modeling multiple trajectories for all the instances in the video to select the one that fits the prompt the most. Despite this, SAMWISE outperforms DsHmp [11] on all three datasets, improving $\\mathcal{J}\\&\\mathcal{F}$ of $+3.1\\%$ , $+2.1\\%$ , and $+5.7\\%$ , while utilizing a smaller model in terms of total parameters. Notably, we achieve this by training only $4.9\\mathrm{M}$ parameters out of $202\\mathrm{M}$ . This result is particularly impressive, as offline methods exploit information from the entire video to handle challenges such as late-appearing objects or motion-dependent disambiguation, as opposed to our streaming approach. With respect to other methods, we outperform them significantly on MeViS, whereas the gap is smaller on Ref-YouTube-VOS and Ref-DAVIS, which contain more descriptive captions, and object-centric videos. Lastly, we also experiment with the text encoder of CLIP, which achieves state-of-the-art results on MeViS, and competitive performance on other benchmarks, while providing a more compact model with just $150\\mathrm{M}$ params.", + "bbox": [ + 511, + 397, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Context Propagation methods. Our proposed baseline GroundingDINO [25] + SAM2, while obviously flawed, being forced to predict the desired instance based on the first frame only, achieves acceptable results on Ref-DAVIS, whereas on MeViS and Ref-Youtu-VOS its performance drops of $11.8\\%$ and $11.7\\%$ , respectively. Differently, SAMWISE, demonstrates excellent performance in both motion-dependent and static scenarios. Specifically,", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3401", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a98b393c94257a50abadcaa3b801981a779fd0e2e94bd11a063978af47b23ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MLP-onlyText-to-VisualVisual-to-TextHSACMEJ&F
✓45.2
✓✓47.5
✓✓48.3
✓✓✓50.3
✓✓✓✓54.2
✓✓✓✓✓55.5
", + "bbox": [ + 106, + 89, + 470, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on MeViS, we outperform OnlineRefer [42] by $+17.2\\%$ . On Ref-YouTube-VOS and Ref-DAVIS, the gap is of $+5.7\\%$ and $+5.8\\%$ , respectively.", + "bbox": [ + 89, + 262, + 482, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Large-VLM based. While comparisons with Large-VLM based approaches are not standard in RVOS evaluations, we include them in this work to provide additional context. The VLM-based solutions [1, 19, 45] are designed to leverage the extensive reasoning capabilities of VLMs to address complex textual instructions and implicit descriptions that require world knowledge [19]. This leads to improved performance in tasks like MeViS, where reasoning over motion patterns is required. However, delegating cross-modal reasoning to these VLMs incurs in significant computational overhead, whereas SAMWISE incorporates visual-text interaction directly at the feature level. Notably, SAMWISE outperforms VISA, the best VLM-based competitor by a substantial margin, respectively $+6\\%$ $+7.7\\%$ $+2.9\\%$ on the three benchmarks.", + "bbox": [ + 89, + 313, + 483, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 550, + 254, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct our ablations on MeViS, as it embodies the core challenges of online RVOS. We report results on the 'valid_u' set [7], employing CLIP-B as text encoder.", + "bbox": [ + 89, + 571, + 482, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Making SAM2 Wiser. We start by showing, in Tab. 2, how each of the core components of our CMT Adapter progressively injects wisdom (i.e. knowledge about language and temporal context) into SAM2. The first line reports the result using the 'naive' solution of aligning the textual prompt to the visual features using a single learnable MLP [51]. While effective to some extent, the results show that allowing early interaction of the two modalities grants a substantial boost $(+5.1\\%)$ with both adapters). Adding explicit temporal feature modeling provides an additional improvements of $+3.9\\%$ . These results sustain our intuition that adding frame-level alignment through a MLP is not enough to obtain robust performances, and that it is essential to couple the text and visual semantics, as well as modeling temporal context. Lastly, the table shows how adding our CME is effective in mitigating SAM2 tracking bias.", + "bbox": [ + 89, + 622, + 482, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hierarchical Selective Attention. In the top section of Tab. 3, we study the effect of the spatial patch size in our", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a9e3ac6f23a8b361755098e85787bb39d85eac5f114e9f1f94ba590e344f951b.jpg", + "table_caption": [ + "Table 2. Ablation of our Cross-Modal Temporal Adapter (CMT). We show the effect of not using CMT (i.e. MLP-only to prompt SAM2), vs. adding one at a time its core components. Lastly, the Conditional Memory Encoder (CME) is added." + ], + "table_footnote": [], + "table_body": "
HSA Patch Size
Fixed SizeHierarchical
1488/4/216/8/4
J&F49.752.353.154.253.8
CME vs Random choice
NeverAlways1 every 4CME
J&F54.250.752.455.5
", + "bbox": [ + 563, + 89, + 857, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HSA module, which models the temporal evolution of over a spatial patch of size $P$ across the temporal axis. Using $P = 1$ is equivalent to processing each pixel independently across frames. The table shows that including spatial context, up to 8 pixels, is beneficial. Using a hierarchical patch size that scales with the feature map resolution yields a gain of $+1.1\\%$ over the fixed sized alternative.", + "bbox": [ + 511, + 262, + 903, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditional Memory Encoder. The bottom section of Tab. 3 provides insight into our CME module. The CME, essentially, detects whenever an object in the unbiased feature maps of the current frame displays higher alignment with the textual prompts w.r.t. the currently tracked one, but SAM2 fails in noticing it due to the tracking bias (Fig. 6). The table compares the effect of Never applying such strategy (i.e., not using CME), doing it Always (i.e., at every frame), or once every 4 frames. The results show that increasing the frequency of artificial detection worsens performances, adding noise to the tracking, whereas the predictions of our CME are beneficial, with a boost of $+1.3\\%$ .", + "bbox": [ + 511, + 373, + 906, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptation strategy. In Tab. 4 we compare different adaptation strategies, including Full Fine-Tuning (FT), LoRa [13], AdaptFormer [5], and the proposed CMT.", + "bbox": [ + 511, + 560, + 905, + 604 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d7c9f20de28c7129c0feda7e0c143fe423fd2a94ad8874d800669735e768ba0a.jpg", + "table_caption": [ + "Table 3. Top: Ablation of the Patch size in our HSA, with the effect of a fixed size vs Hierarchical. Bottom: Effect of not predicting detections (Never), predicting them at every frame (Always), randomly (1 in 4) vs. using predictions of our CME." + ], + "table_footnote": [], + "table_body": "
Full-FT LoRa [13] AdaptFormer [5] CMT (ours)
MeViS J&F43.144.243.948.3
", + "bbox": [ + 527, + 618, + 893, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Baselines: prior adapters and full-finetune, with CLIP-B.", + "bbox": [ + 514, + 660, + 903, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced SAMWISE, a novel approach for RVOS that builds upon the SAM2 model by incorporating i) natural language understanding, ii) temporal feature modeling, and iii) a learnable strategy to adjusts tracking focus according to visual cues that emerge over time. SAMWISE achieves SOTA performance across benchmarks while adding less than 5M parameters, without modifying SAM2 weights or using external models for visual-text alignment. We obtain an effective pipeline for applications of streaming video segmentation, addressing limitations of existing RVOS approaches, which either lack long-term context or rely on single-frame context propagation.", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3402", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. Claudia Cuttano was supported by the Sustainable Mobility Center (CNMS) which received funding from the European Union Next Generation EU (Piano Nazionale di Ripresa e Resilienza (PNRR), Missione 4 Componente 2 Investimento 1.4 \"Potenziamento strutture di ricerca e creazione di \"campioni nazionali di R&S\" su alcune Key Enabling Technologies\") with grant agreement no. CN_00000023.", + "bbox": [ + 89, + 92, + 480, + 188 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gabriele Trivigno, Giuseppe Avera and Carlo Masone were supported by FAIR - Future Artificial Intelligence Research which received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them.", + "bbox": [ + 89, + 189, + 480, + 301 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources.", + "bbox": [ + 89, + 301, + 480, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 345, + 187, + 361 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 7, 8", + "[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, 82(3):4419-4438, 2023. 2, 6", + "[3] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 2, 7", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 8", + "[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7900-7916, 2022. 5", + "[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 1, 2, 6, 7, 8", + "[8] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, 8(10), 2021. 2" + ], + "bbox": [ + 99, + 371, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5958-5966, 2018. 1, 2", + "[10] Mingfei Han, Yali Wang, Zhihui Li, Lina Yao, Xiaojun Chang, and Yu Qiao. Http: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13414-13423, 2023. 1, 2, 7", + "[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024. 2, 3, 7", + "[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3, 4", + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 8", + "[14] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5574-5583, 2024. 7", + "[15] Haojun Jiang, Jianke Zhang, Rui Huang, Chunjiang Ge, Zanlin Ni, Jiwen Lu, Jie Zhou, Shiji Song, and Gao Huang. Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623, 2022. 3", + "[16] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27144-27153, 2024. 3, 4", + "[17] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 2, 6, 7", + "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4", + "[19] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3, 7, 8", + "[20] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. Refsam: Efficiently adapting segmenting anything model for referring video object segmentation, 2024. 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3403", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[22] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6555-6564, 2023. 4", + "[23] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H. Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13658-13667, 2024. 4", + "[24] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 1, 2", + "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 7", + "[26] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 7", + "[27] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 3", + "[28] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7", + "[29] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 2", + "[30] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7", + "[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9226–9235, 2019. 2", + "[32] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 4" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7", + "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2", + "[36] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3", + "[37] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 4, 6", + "[38] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In European Conference on Computer Vision, 2020. 1, 2, 7", + "[39] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1", + "[40] Mengmeng Wang, Jiazheng Xing, Boyuan Jiang, Jun Chen, Jianbiao Mei, Xingxing Zuo, Guang Dai, Jingdong Wang, and Yong Liu. A multimodal, multi-task adapting framework for video action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5517-5525, 2024. 3", + "[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2", + "[42] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. OnlineRefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2, 3, 7, 8", + "[43] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 2, 7", + "[44] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3404", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 17503-17512, 2023. 3", + "[45] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 3, 7, 8", + "[46] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6449-6457, 2024. 7", + "[47] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 3", + "[48] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 2", + "[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7", + "[50] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, Xinggang Wang, et al. Evfsam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2", + "[51] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 8" + ], + "bbox": [ + 91, + 90, + 482, + 613 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3405", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_model.json b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b69c02d124322782eab5554c5b67a427182d265 --- /dev/null +++ b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.884, + 0.154 + ], + "angle": 0, + "content": "SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.18, + 0.893, + 0.235 + ], + "angle": 0, + "content": "Claudia Cuttano1 Gabriele Trivigno1 Gabriele Rosi1,2 Carlo Masone1,2 Giuseppe Averta1,2 \n1 Politecnico di Torino 2 Focoos AI \n{name.surname}@polito.it {name.surname}@focoos.ai" + }, + { + "type": "title", + "bbox": [ + 0.402, + 0.27, + 0.584, + 0.285 + ], + "angle": 0, + "content": "RVOS Streaming processing" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.289, + 0.906, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.403, + 0.247, + 0.419 + ], + "angle": 0, + "content": "Segment Anything 2" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.421, + 0.359, + 0.461 + ], + "angle": 0, + "content": "
SPATIAL\nPROMPTTEXT\nPROMPTCORRECTION\nMECHANISMTEMPORAL\nFEATURES
✓✗human✗
" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.39, + 0.639, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.395, + 0.771, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.771, + 0.402, + 0.903, + 0.417 + ], + "angle": 0, + "content": "SAMWISE (ours)" + }, + { + "type": "table", + "bbox": [ + 0.654, + 0.419, + 0.904, + 0.468 + ], + "angle": 0, + "content": "
SPATIAL\nPROMPTTEXT\nPROMPTCORRECTION\nMECHANISMTEMPORAL\nFEATURES
✓✓learnable✓
" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.485, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Figure 1. SAMWISE. Our approach infuses knowledge about natural language in the Segment-Anything 2 model, adding explicit temporal cues in the feature extraction for the task of streaming-based Referring Video Segmentation (RVOS). We use a learnable mechanism to mitigate the so-called tracking bias, i.e. SAM2 tendency to overlook a correct object once it becomes identifiable, due to its ongoing tracking of a different object. Our design enables effective streaming processing for RVOS, exploiting the memory from previous frames to propagate past context. The figure shows an example where the target object is not present in the first frame, leading SAM2 to start tracking the wrong one. Afterwards, when the correct object appears, our learnable correction mechanisms guide SAM2 to switch its tracking focus. By adding in its features the notion of temporal evolution, the model is able to recognize that the new object is more aligned with the provided textual query. Finally, we exploit SAM2 tracking skills and robustness to occlusions to keep following the object." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.61, + 0.327, + 0.625 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.644, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Referring Video Object Segmentation (RVOS) relies on natural language expressions to segment an object in a video clip. Existing methods restrict reasoning either to independent short clips, losing global context, or process the entire video offline, impairing their application in a streaming fashion. In this work, we aim to surpass these limitations and design an RVOS method capable of effectively operating in streaming-like scenarios while retaining contextual information from past frames. We build upon the Segment-Anything 2 (SAM2) model, that provides robust segmentation and tracking capabilities and is naturally suited for streaming processing. We make SAM2 wiser, by empowering it with natural language understanding and explicit temporal modeling at the feature extraction stage, without fine-tuning its weights, and without outsourcing modality interaction to external models. To this end, we introduce a novel adapter module that injects temporal information" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.748 + ], + "angle": 0, + "content": "and multi-modal cues in the feature extraction process. We further reveal the phenomenon of tracking bias in SAM2 and propose a learnable module to adjust its tracking focus when the current frame features suggest a new object more aligned with the caption. Our proposed method, SAMWISE, achieves state-of-the-art across various benchmarks, by adding a negligible overhead of less than \\(5M\\) parameters. Code is available at https://github.com/ClaudiaCuttano/SAMWISE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Referring video segmentation (RVOS) [9, 17, 24, 38, 43, 48] aims at segmenting and tracking specific objects of interest within video content, guided by natural language expressions [3, 10, 28]. Existing RVOS methods are mostly based on a divide and conquer paradigm, where the video is divided into shorter clips that are processed independently [3, 39, 43]. However, as demonstrated by MeViS [7], this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "3395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.5 + ], + "angle": 0, + "content": "solution fails in examples that require taking into account long-term motion and global context. As a workaround to handle this challenge, the state-of-the-art method [11] processes the entire video in an offline fashion, first modeling trajectories of all instances throughout the entire clip and then selecting the most appropriate one. Albeit effective, this approach is not applicable when the model has access only to a portion of the video, for example when the data at inference time are presented in a streaming fashion or due to limitations in the computational resources. The trade-off of these two paradigms is schematized in Fig. 1. To this end, OnlineRefer [42] introduced a context propagation scheme for online RVOS but relies solely on past context from a single frame, limiting its ability to capture long-term dependencies. In this work, we investigate how to exploit the memory from past frames to design an RVOS method capable of retaining global context while operating within a streaming paradigm, i.e., without requiring access to the whole video at once. This idea is inspired by the recent release of Segment-Anything 2 (SAM2) [35], a foundational model that has shown impressive capabilities in various Video Segmentation tasks thanks to a memory bank that allows to leverage long-range past information. Since SAM2 operates in a streaming fashion, extending this method to enable context-aware streaming processing in RVOS would appear a natural step. However, this entails some non-trivial challenges:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.485, + 0.655 + ], + "angle": 0, + "content": "i) Text understanding. SAM2 original design accounts only for spatial prompts (e.g. points) and lacks mechanisms to interpret semantic prompts like text, which require reasoning over visual and textual modalities. While we are the first to address the challenge of adding textual prompts to SAM2, previous methods have explored this problem for SAM-1 at image-level. These solutions [19, 50] delegate visual-textual interaction to an off-the-shelf large VLM (like BEIT-3 [41], LLaVa [21]), which generates a multimodal embedding that is used to prompt SAM-1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.485, + 0.763 + ], + "angle": 0, + "content": "ii) Temporal modeling. To segment the referred object throughout the video, it must be first recognized and then tracked. While the latter requires matching objects visual appearance across adjacent frame, the recognition problem entails modeling temporal evolution to reason over actions that unfold over multiple frames. However, SAM2 extracts frame features independently, lacking such reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.904 + ], + "angle": 0, + "content": "iii) Tracking bias. In RVOS, the target object might be unrecognizable during certain time intervals, due to occlusions, presence of multiple instances or forthcoming actions, as in the first frames of Fig. 1. In such cases, SAM2 may start tracking an incorrect object that partially matches the textual prompt, and persist in following it, leading to what we denote as tracking bias. While SAM2 original design allows for a user to manually correct the prediction by providing a new prompt, such a strategy is not applicable in" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.502, + 0.486, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.823, + 0.107 + ], + "angle": 0, + "content": "tasks without a human-in-the-loop like RVOS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.41 + ], + "angle": 0, + "content": "In this work, we aim at making SAM2 wiser, by addressing these limitations without fine-tuning SAM2 weights, thereby preserving its original capabilities, and without outsourcing modality interaction to external, heavy models. To overcome challenges \\(i\\) and \\(ii\\), we design a learnable Adapter [12] module, named Cross-Modal Temporal Adapter (CMT), with two key principles in mind: a) enabling mutual interaction between visual and linguistic modalities; and b) encoding temporal cues into visual features. Then, to generate a prompt, we follow [21, 51] and employ a learnable MLP to project the sentence embedding for the SAM2 Mask Decoder, which then outputs the final segmentation mask. In this way, we can exploit SAM2 tracking capability to segment an object given a textual query across the video. Finally, to mitigate the tracking bias problem iii), we introduce a lightweight Conditional Memory Encoder (CME) which detects when a candidate object, aligned with the text, appears in the frame, thus enabling SAM2 to dynamically refocus its tracking to the correct object as it becomes distinguishable." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.885, + 0.427 + ], + "angle": 0, + "content": "Summarizing, this paper contributes with the following:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.428, + 0.907, + 0.517 + ], + "angle": 0, + "content": "- We present SAMWISE, the first method that integrates natural language knowledge into SAM2 in an end-to-end solution tailored to address the challenges of RVOS. We introduce a novel adapter, namely Cross Modal Temporal (CMT) Adapter, which purposefully models temporal evolution and multi-modal interaction;" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.518, + 0.907, + 0.579 + ], + "angle": 0, + "content": "- We provide insight into the functioning of SAM2, highlighting the phenomenon of tracking bias, and introduce a learnable module (Conditional Memory Encoder) to adjust tracking based on new information;" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.579, + 0.906, + 0.655 + ], + "angle": 0, + "content": "- Our methods achieve state-of-the-art results both on traditional RVOS benchmarks (Ref-Youtu-VOS [38], RefDAVIS [17]), as well as the more challenging MeViS [7], without compromising SAM2 capabilities and adding less than \\(5M\\) learnable parameters." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.428, + 0.907, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.672, + 0.661, + 0.687 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Referring Video Segmentation. In RVOS, the goal is to segment an object in a clip described with natural language [8, 17, 24]. Earlier works adapted image-based methods [2, 9, 17, 48], or used a spatio-temporal memory to attend to masks of previous frames [31, 38]. Subsequent works employ a DETR-like [4] structure to process multiple frames and text embeddings [3, 10, 29, 43]. All these methods process short clips independently, thus losing global context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recently, [7] showed how traditional RVOS benchmarks lack challenging captions that require to disambiguate between instances and their actions, as well as occlusions and dynamic queries, highlighting how they could be solved even with image-based methods. The MeViS dataset [7] tar" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.9, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.322, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 2. Overview of SAMWISE. We build on a frozen SAM2 and a frozen Text Encoder to segment images in video given a textual description. We incorporate the Cross-Modal Temporal Adapter (CMT) into the text and visual encoders at every intermediate layer \\( k \\) to model temporal dynamics within visual features while contaminating each modality with the other. Then, we extract the [CLS] and verb embeddings, namely Contextual and Motion prompts, from the adapted textual features and project them through a learnable MLP. The final embedding is used to prompt the Mask Decoder, which outputs the segmentation mask. Finally, the Conditional Memory Encoder detects when a new candidate object, aligned with the caption, appears in the frame, enabling SAM2 to dynamically refocus its tracking." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.484, + 0.568 + ], + "angle": 0, + "content": "gets these scenarios, with challenging examples that previous image or clip-based methods fail to address. To this end, a few works proposed offline methods to explicitly model multiple object trajectories [11, 28], with the latter representing the state-of-the-art on MeViS. Concurrently, OnlineRefer [42] proposed a first attempt towards an online RVOS setting, with a query propagation scheme. However, its effectiveness is limited as predictions are based on a single frame. Our method builds on this paradigm by leveraging SAM2 memory bank to encode long-range past context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.818 + ], + "angle": 0, + "content": "Text-prompted Segment-Anything. Recent works have provided solutions to adapt SAM-1 for text-prompted segmentation. Grounded SAM [36] employs a two step pipeline where GroundingDINO [25] generates bounding boxes for SAM-1 to produce segmentation masks. Applying such pipeline in RVOS is problematic, as potential errors in the first frame are propagated throughout the whole video. To directly prompt SAM-1, RefSAM [20] exploits a projection layer to map the textual embedding into the prompt space, while [1, 19, 45] resort to large off-the-shelf VLM to generate a multi-modal embedding that is used to prompt SAM-1. Both solutions finetune the Mask Decoder, thereby compromising its capabilities on its original task. In contrast, our work is the first to propose an end-to-end model that incorporates textual knowledge within SAM2 without fine-tuning nor relying on external models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Pre-Trained Knowledge Transfer. In recent years, the release of powerful pretrained models has sparked interest in the question of how to extend their skills to novel tasks, as full fine-tuning becomes increasingly impractical with growing model sizes [16, 33]. A powerful strat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.689 + ], + "angle": 0, + "content": "egy to address this problem relies on using Adapters [12], small trainable modules that enable efficient adaptation of pre-trained models. Following this paradigm, recent studies have explored adapting CLIP [34] for downstream tasks. At the image level, [44] inserts Transformer Decoder blocks within CLIP encoders, which entail costly Self-Attentions on all tokens. For video tasks, [40] places independent adapter modules within each encoder, whereas [15, 16, 27, 47] rely on a weight-sharing mechanism to project both modalities in a shared sub-space. Nevertheless, as features of each modality are independently extracted, none of these adapters allows explicit feature interaction, unlike our CMT, which also incorporates temporal modeling. Lastly, all these works start from a model that already includes a text encoder (CLIP), whereas ours is the first to propose an adapter for the Segment-Anything 2 model to add textual understanding, achieving robust performances while introducing less than \\(5\\mathrm{M}\\) parameters." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.699, + 0.634, + 0.715 + ], + "angle": 0, + "content": "3. SAMWISE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.907, + 0.882 + ], + "angle": 0, + "content": "Problem setting. Given an input video \\(\\mathcal{V} = \\{I_t\\}_{t=1}^{T_V}\\) with \\(T_V\\) frames and a referring expression, we aim to predict a set of binary masks \\(S = \\{s_t\\}_{t=1}^{T_V}\\), \\(s_t \\in \\mathbb{R}^{H \\times W}\\) of the referred object. We tokenize the textual query in a set of \\(L\\) words, \\(E = \\{e_l\\}_{l=1}^L\\), and add a global sentence representation token [CLS]. The tokens are then processed using a frozen text encoder to extract language features \\(\\mathcal{E} \\in \\mathbb{R}^{L \\times C_t}\\). We process videos in a streaming fashion, collecting clips of \\(T\\) frames as they are available. Throughout the rest of the section, we use \\(T\\) to indicate clip length." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Overview. We first provide a brief discussion of the SAM2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3397" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Architecture of our Cross Modal Temporal (CMT) Adapter, made up of Hierarchical Selective Attention (HSA) to model temporal cues, a Visual-to-Text Attention (VTA) and Text-to-Visual Attention (VTA) modules." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.41 + ], + "angle": 0, + "content": "model (Sec. 3.1). We then outline the pipeline of our proposed SAMWISE, starting from the prompting strategy in Sec. 3.2. In Sec. 3.3, we detail our novel Cross-Modal Temporal Adapter. Lastly, in Sec. 3.5, we discuss our learnable correction strategy, named Conditional Memory Encoder, to address the issue of tracking bias." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.379, + 0.437 + ], + "angle": 0, + "content": "3.1. Background: Segment-Anything" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.761 + ], + "angle": 0, + "content": "The Segment-Anything Model 2 (SAM2) builds upon SAM-1 [18] to tackle the task of Promptable Video Object Segmentation, i.e., tracking an object in a video given a textual prompt. Following SAM-1, it consists of an image encoder, a Prompt Encoder and a Mask Decoder, which combines the image and prompt embeddings to predict segmentation masks. To enable video processing, SAM2 comes with a few modifications: \\(i\\) ) the original ViT backbone is replaced by Hiera [37], roughly 3 times faster, which processes frames independently to provide hierarchical visual features. Hereinafter, we refer to them as memory-less features \\(\\mathcal{F}\\); \\(ii\\) ) frame embeddings are not directly fed to the Mask Decoder, but they are first conditioned on memories of past predictions from a Memory Bank. We refer to these conditioned features as memory features \\(\\mathcal{F}_{mem}\\). Lastly, \\(iii\\) ) once the mask for the current frame is predicted, the Memory Encoder updates the Memory Bank. By design, SAM2 handles video frames as they become available, progressively encoding the past in its Memory Bank. We argue that this streaming approach is especially valuable in RVOS, enabling reasoning over a wide temporal horizon." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.265, + 0.789 + ], + "angle": 0, + "content": "3.2. Prompting SAM2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To guide the SAM2 decoder, we use a Contextual Prompt, \\(\\mathcal{E}_C\\in \\mathbb{R}^{1\\times C_t}\\), which encodes the high-level semantic information for the given text query, emphasizing the essential aspects of the query while downplaying less relevant elements. To this end, we employ the [CLS] embedding of text features, \\(\\mathcal{E}\\). Furthermore, we also introduce a second prompt, the Motion Prompt \\(\\mathcal{E}_M\\in \\mathbb{R}^{1\\times C_t}\\), which cap" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "tures action-related cues by using verb embeddings from \\(\\mathcal{E}\\). These prompts are concatenated and projected through a learnable three-layer MLP:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.149, + 0.907, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\rho = \\mathbf {W} _ {\\text {p r o m p t}} \\left(\\operatorname {C A T} \\left[ \\mathcal {E} _ {C}, \\mathcal {E} _ {M} \\right]\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.176, + 0.907, + 0.282 + ], + "angle": 0, + "content": "In this way, the provided prompts encode both subject-related and motion-based information. Given that in our task the textual prompt is not referred a-priori to any particular frame, we prompt SAM2 at each frame, so that the model has to balance the influence of tracking while also considering the content of each frame. We discuss more in depth this aspect in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.291, + 0.798, + 0.307 + ], + "angle": 0, + "content": "3.3. Cross-Modal Temporal Adapter" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.313, + 0.906, + 0.388 + ], + "angle": 0, + "content": "An adapter consists of a linear down-projection \\((\\mathbf{W}_{\\mathrm{down}})\\) to a bottleneck dimensionality, followed by an up-projection back \\((\\mathbf{W}_{\\mathrm{up}})\\) in the original space, separated by a non-linear activation function \\(\\sigma\\). Formally, given an input feature \\(\\mathbf{x} \\in \\mathbb{R}^{1 \\times d}\\), the adapter function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.4, + 0.907, + 0.417 + ], + "angle": 0, + "content": "\\[\nA d a p t e r (\\mathbf {x}) = \\mathbf {x} + \\sigma (\\mathbf {x} \\mathbf {W} _ {\\text {d o w n}}) \\mathbf {W} _ {\\text {u p}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.906, + 0.547 + ], + "angle": 0, + "content": "We build on this popular Adapter framework [12] and propose a novel Cross-Modal Temporal Adapter (CMT) (see Fig. 3) which models temporal dynamics within visual features while enriching each modality with the other. Formally, given the visual features in a clip \\(\\mathcal{F}^k\\in \\mathbb{R}^{T\\times H_k\\times W_k\\times C_k}\\) and the textual features \\(\\mathcal{E}^k\\in \\mathbb{R}^{L\\times C}\\) extracted at layer \\(k\\) of the image and text encoders, respectively, the CMT can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.558, + 0.897, + 0.577 + ], + "angle": 0, + "content": "\\[\nA d a p t e r \\left(\\mathcal {F} ^ {k}\\right) = \\mathcal {F} ^ {k} + h \\left(\\mathcal {F} ^ {k} \\mathbf {W} _ {\\text {d o w n , v}}, \\mathcal {E} ^ {k} \\mathbf {W} _ {\\text {d o w n , t}}\\right) \\mathbf {W} _ {\\text {u p , v}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.578, + 0.905, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A d p a t e r} \\left(\\mathcal {E} ^ {k}\\right) = \\mathcal {E} ^ {k} + h \\left(\\mathcal {E} ^ {k} \\mathbf {W} _ {\\text {d o w n , t}}, \\mathcal {F} ^ {k} \\mathbf {W} _ {\\text {d o w n , v}}\\right) \\mathbf {W} _ {\\text {u p , t}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.761 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathrm{down},\\mathrm{v}}\\), \\(\\mathbf{W}_{\\mathrm{down},\\mathrm{t}}\\), \\(\\mathbf{W}_{\\mathrm{up},\\mathrm{v}}\\), \\(\\mathbf{W}_{\\mathrm{up},\\mathrm{t}}\\) are modality specific down- and up-projections weights and \\(h\\) is our proposed adapter function. The adapter output is summed with the original features, allowing the model to retain the original encoding while incorporating temporal and cross-modal reasoning. We integrate the Cross-Modal Temporal Adapter (CMT) into the frozen text and visual encoders at every intermediate layer \\(k\\). In the following paragraphs we detail the temporal and cross-modal adaptation functions, which are tightly coupled in our Adapter module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Temporal Adaptation. Our approach aims to embed motion cues directly into the frame-level features of SAM2. Previous works based on Adapters either perform self-attention (SA) over all tokens in a clip [16], which is costly, or restrict the attention to the temporal axis for each pixel [22, 23]. We observe that, within a video, object motion across adjacent frames typically spans a localized region of the image [32]. Consequently, a given element of the feature map primarily benefits from interactions with its" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3398" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.279 + ], + "angle": 0, + "content": "Figure 4. Scheme of our Hierarchical Selective Attention (HSA), modeling temporal evolution of features in our adapter." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.295, + 0.484, + 0.4 + ], + "angle": 0, + "content": "spatial and temporal neighbors, rather than requiring long-range connections across the entire feature map. Building on this intuition, we introduce a Hierarchical Selective Attention (HSA) mechanism, illustrated in Fig. 4. By modeling interactions among spatially and temporally proximal regions, HSA reduces unnecessary computations while capturing motion-based context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Formally, at layer \\(k\\), given the set of feature maps for a \\(T\\)-frames clip: \\(\\mathcal{F}^k \\in \\mathbb{R}^{T \\times H_k \\times W_k \\times C_k}\\), we decompose this feature volume into non-overlapping, 3-D spatio-temporal patches of size \\(T \\times P \\times P\\), obtaining \\(N = H_k W_k / P^2\\) sub-volumes. These sub-volumes, considered pixelwise, can be represented as a set of tokens \\(F^k[n] = \\left\\{x_{i,j,t}^{k,n} \\in \\mathbb{R}^{C_k} : i \\in 1, \\dots, P, j \\in 1, \\dots, P, t \\in 1, \\dots, T\\right\\}\\). To encode spatio-temporal positioning, to each vector we add a spatial \\((e[i,j])\\) and a temporal \\((e[t])\\) sinusoidal positional embeddings, in 2-D and 1-D formats, respectively. Specifically: \\(x_{i,j,t}^{k,n} = x_{i,j,t}^{k,n} + e[i,j] + e[t]\\). Each sub-volume contains \\(M = P^2 * T\\) tokens, on which we perform self-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.619, + 0.483, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i, j, t} ^ {k, n} := S A \\left(\\left\\{\\mathbf {x} _ {i ^ {\\prime}, j ^ {\\prime}, t ^ {\\prime}} ^ {k, n} \\right\\} _ { \\begin{array}{c} i ^ {\\prime} = 1.. P \\\\ j ^ {\\prime} = 1.. P \\\\ t ^ {\\prime} = 1.. T \\end{array} }\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.744 + ], + "angle": 0, + "content": "At each layer \\( k \\) of the feature extraction process, the patch size \\( P \\) is progressively scaled, as depicted in Fig. 4-d. This scaling adapts the sub-volume to the hierarchy of feature resolution, encoding information at multiple scales." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Cross-Modal Adaptation. To unify text and visual representations, we encourage modality interaction from early stages of the feature extraction process through two symmetric operations: Visual-to-Text Attention (VTA) and Text-to-Visual Attention (TVA)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Within the former, each visual feature, already enriched with temporal information through the HSA, attends to the full textual expression, allowing the model to identify candidate regions within the image based on both categorical details (e.g., the subject described in the text) and motion" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.227, + 0.906, + 0.269 + ], + "angle": 0, + "content": "Figure 5. Cross Modal Temporal Adapter: we show via PCA that our CMT provides contextualized visual features based on the given textual prompt, compared to SAM2 original ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.905, + 0.31 + ], + "angle": 0, + "content": "cues (e.g., actions), facilitating early alignment with the prompt, as visible in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.355 + ], + "angle": 0, + "content": "Formally, at layer \\(k\\), we consider the feature of each frame in the clip, i.e. \\(\\mathcal{F}^k[t] \\in \\mathbb{R}^{H_k \\times W_k \\times C_k}\\), \\(t = 1, \\ldots, T\\), and the set of textual embeddings \\(\\mathcal{E}^k \\in \\mathbb{R}^{L \\times C}\\) to compute:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.365, + 0.905, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {k} [ t ] := \\mathcal {F} ^ {k} [ t ] * C A \\left(\\mathcal {F} ^ {k} [ t ], \\mathcal {E} ^ {k}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.502 + ], + "angle": 0, + "content": "In parallel, as the meaning of a caption can shift significantly depending on the visual content of the associated image [6], we aim at contextualizing the textual query with the semantics provided by the visual modality. To this end, the TVA progressively enriches the linguistic tokens \\(\\mathcal{E}^k\\in \\mathbb{R}^{L\\times C}\\) with information from the visual feature maps, averaged over the video clip \\(\\mathcal{F}_{avg}^k\\)" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.513, + 0.905, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} ^ {k} := \\mathcal {E} ^ {k} * C A \\left(\\mathcal {E} ^ {k}, \\mathcal {F} _ {\\text {a v g}} ^ {k}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.676, + 0.568 + ], + "angle": 0, + "content": "3.4. Mask prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.726 + ], + "angle": 0, + "content": "At the end of the feature extraction process, we obtain the adapted visual and linguistic features, respectively \\(\\mathcal{E}\\) and \\(\\mathcal{F}\\). To perform the final prediction, we extract the prompt \\(\\rho\\) as in Eq. (1), while the Memory Attention module generates the memory features \\(\\mathcal{F}_{mem}\\) by conditioning the visual features \\(\\mathcal{F}\\) on past predictions from the Memory Bank. The prompt \\(\\rho\\) is fed into the frozen Mask Decoder \\(\\mathcal{D}_{dec}\\), which generates the output mask \\(\\mathcal{P}_M \\in \\mathbb{R}^{1 \\times H \\times W}\\) and the mask token \\(\\tau_m \\in \\mathbb{R}^{1 \\times d}\\), i.e. an embedding representing the segmented object. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.736, + 0.905, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\tau_ {m}, \\mathcal {P} _ {M} = \\mathcal {D} _ {d e c} (\\mathcal {F} _ {m e m}, \\rho), \\\\ \\mathcal {Y} = \\mathcal {P} _ {M} > 0, \\end{array} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.905, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathcal{V} \\in \\mathbb{R}^{1 \\times H \\times W}\\) denotes the output binary segmentation mask. Finally, the Memory Encoder updates the memory bank with \\(\\mathcal{P}_M\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.782, + 0.85 + ], + "angle": 0, + "content": "3.5. Conditional Memory Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We identify as tracking bias the phenomenon of SAM2 tracking the wrong object when the correct one is not yet identifiable in the video, and persist in following it. This" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.481, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.334, + 0.485, + 0.473 + ], + "angle": 0, + "content": "Figure 6. Effect of our Conditional Memory Encoder. The caption above requires disambiguating multiple instances of the same class (e.g., \"cat\") by identifying a specific action (e.g., \"climbing\"). Since none of the instances perform this action initially, the model begins tracking the wrong instance and fails to correct itself once the target object performs the action. The top section visualizes the effect of this tracking bias in the memory features of SAM2. Our CME detects that the cat that starts climbing is more aligned with the caption, and encodes its presence in the memory bank, allowing SAM2 to switch its focus." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.484, + 0.726 + ], + "angle": 0, + "content": "bias, as exemplified in Fig. 6, is encoded in the memory features, which are propagated to subsequent frames through the Memory Encoder. On the other hand, we observe that the memory-less features: i) contain an unbiased representation of the current frames, ii) are aligned with the textual prompt via our CMT (cf. Fig. 5), and iii) can thus be used to propose candidate instances that match the prompt without being biased by past predictions. Building on these intuitions, we derive a memory-less token \\(\\tau_{l}\\) from a cross-attention between the unbiased feature maps and the prompt. Such token represents a summary of the visual features that match the prompt. The idea is to compare it with the mask token \\(\\tau_{m}\\) generated by the Mask Decoder, to detect when they represent different objects, i.e., to detect when SAM2 is tracking an object that is not the one currently most aligned with the caption. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.737, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\tau_ {l} = C A (\\mathcal {F}, \\rho) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We note that we initialize (and keep frozen) the weights of the cross-attention with those from SAM2 Mask Decoder. We introduce a small learnable module, named Conditional Memory Encoder (CME), to detect such situations. When a new object is detected, a naive solution would be to compute its mask and use it to re-prompt the model at the given frame, just like a user would do, forcing SAM2 to switch its prediction. However, since the prediction computed on the memory-less features does not have access to past video" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.243 + ], + "angle": 0, + "content": "context, it might generate false positives. Thus, we propose a soft assignment, obtained by encoding the masks of both objects in the memory bank. Essentially, the CME allows SAM2 to 'see' other objects beyond the currently tracked one, and balance the influence of past context with new information, to select the one that fits the prompt the most. In detail, our CME, illustrated in Fig. 6-bottom, concatenates the two tokens \\(\\tau_{m},\\tau_{l}\\) with a learnable decision token [DEC], and performs a self-attention followed by a Linear classifier:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.251, + 0.907, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\left[ z _ {D E C}, z _ {M T}, z _ {M L} \\right] = S A \\left(\\left[ [ D E C ], \\tau_ {m} [ t ], \\tau_ {l} [ t ] \\right]\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.283, + 0.752, + 0.297 + ], + "angle": 0, + "content": "\\[\np _ {d e t e c t} = \\phi (z _ {D E C})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.388 + ], + "angle": 0, + "content": "where \\(\\phi\\) is a linear function \\(\\mathbb{R}^d\\to \\mathbb{R}^1\\). When detecting a candidate text-aligned object, (i.e., \\(p_{\\text{detect}} > 0.5\\)), instead of directly feeding the predicted output mask \\(\\mathcal{P}_m\\) to the Memory Encoder, our module computes the unbiased output mask, namely \\(\\mathcal{P}_L\\in \\mathbb{R}^{1\\times H\\times W}\\), to fuse it with \\(\\mathcal{P}_m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.399, + 0.766, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {l} = \\mathcal {D} _ {d e c} \\left(\\mathcal {F}, \\rho\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.418, + 0.905, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (h, w) = \\mathbb {1} (h, w) [ h, w: \\mathcal {P} _ {l} > 0 ] \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.437, + 0.836, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} = \\lambda * \\mathcal {P} _ {l} \\circ \\mathcal {M} + \\mathcal {P} _ {m} \\circ (1 - \\mathcal {M})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.603 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}(h, w)\\) is a binary mask whose value is zero except for the pixels corresponding to the object, and \\(\\lambda\\) is an hyperparameter weighing the influence of the memoryless prediction. The resulting mask \\(\\mathcal{P}\\) is fed to the Memory Encoder. We train the CME via self-supervision with a standard Cross-Entropy loss, by providing examples where the memory-less features highlight different objects w.r.t the one currently tracked. We discuss in detail our training protocol in the Supp. Mat.." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.715, + 0.634 + ], + "angle": 0, + "content": "4. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.784 + ], + "angle": 0, + "content": "Dataset. We evaluate our method on MeVis [7], Ref-Youtube-VOS [2] and Ref-Davis [17]. MeViS includes 2,006 videos and features a total of 28K annotations that capture various aspects of motion. Ref-Youtu-VOS enhances the original YouTube-VOS benchmark by incorporating textual descriptions. It contains a total of 3,978 videos and approximately 15K language expressions. RefDAVIS17 builds upon DAVIS17 dataset, adding more than 1.5K linguistic annotations to 90 videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Evaluation Metrics. We utilize standard evaluation metrics, region similarity \\((\\mathcal{J})\\), contour accuracy \\((\\mathcal{F})\\), and their average \\((\\mathcal{J} \\& \\mathcal{F})\\). For MeViS and Ref-Youtu-VOS we conduct the evaluation using the official challenge servers; for Ref-DAVIS17, we used the official evaluation code." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We employ Hiera-B [37] as visual extractor. As text encoder, we experiment with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3400" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.351 + ], + "angle": 0, + "content": "
MethodVisual EncoderText EncoderTotal ParamsMeViSRef-YouTube-VOSRef-DAVIS17
J&FJFJ&FJFJ&FJF
Large VLM based
LISA [19] [CVPR'24]ViT-HLLaVa7 B37.235.139.453.953.454.364.862.267.3
VISA [45] [ECCV'24]ViT-HChat-UniVi7 B43.540.746.361.559.863.269.466.372.5
One-Token-Seg-All [1] [NIPS'24]ViT-HPhi-33.8 B42.339.445.261.760.263.367.763.871.5
MTTR [3] [CVPR'22]V-Swin TRoBERTa-30.028.831.255.354.056.6---
TCE-RVOS [14] [WACV'24]ResNet-50RoBERTa----59.658.360.859.456.562.4
ReferFormer [43] [CVPR'22]V-Swin BRoBERTa237 M31.029.832.262.961.364.661.158.164.1
SOC [28] [NIPS'23]V-Swin BRoBERTa220 M---66.064.167.964.261.067.4
OnlineRefer [42] [ICCV'23]Swin LRoBERTa232 M32.331.533.163.561.665.564.861.667.7
LMPM [7] [ICCV'23]Swin TRoBERTa195 M37.234.240.2------
RefSAM [20] [arXiv]ViT-BT53 B---58.457.459.462.159.065.3
DsHmp [11] [CVPR'24]V-Swin BRoBERTa339 M---67.165.069.164.961.768.1
DsHmp [11] [CVPR'24]Swin TRoBERTa272 M46.443.049.8------
MUTR [46] [AAAI'24]V-Swin BRoBERTa250 M---67.565.469.666.462.870.0
GroundingDINO [25]+SAM2Hiera-BBERT240 M37.734.940.557.555.659.566.462.869.9
SAMWISE (ours)Hiera-BCLIP-B150 M48.345.451.267.265.269.368.565.671.5
SAMWISE (ours)Hiera-BRoBERTa202 M49.546.652.469.267.870.670.667.474.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.357, + 0.907, + 0.387 + ], + "angle": 0, + "content": "Table 1. Comparison of SAMWISE against state-of-the-art RVOS methods on MeViS, Ref-Youtu-VOS and Ref-DAVIS datasets. We further include methods based on large VLMs for comparison. Bold and underline indicate the two top results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.482, + 0.551 + ], + "angle": 0, + "content": "CLIP [34] and RoBERTa [26]. We note that the text encoder and SAM2 weights are entirely frozen and we train only the Adapters and the CME module (4.2M parameters when using CLIP and 4.9M with RoBERTa). Following [10, 20, 28, 42, 43], we undergo pre-training for 6 epochs on RefCOCO+/g [30, 49] with a learning rate at 1e-4 and fine-tune on Ref-Youtu-VOS [38] for 4 epochs with a learning rate of 1e-5, using the Adam optimizer. The model trained on the Ref-Youtu-VOS is directly evaluated on DAVIS-17 [17]. On MeViS [7], we train for 1 epoch. We set \\( T = 8 \\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.231, + 0.577 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.632 + ], + "angle": 0, + "content": "In this section, we compare against existing works in the literature, and ablate our contributions. In the Supp. Mat. we report additional qualitative results and ablations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Baselines. To assess the validity of our approach, we divide the experimental comparison in the following categories:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.669, + 0.483, + 0.713 + ], + "angle": 0, + "content": "- Standard RVOS methods: we compare against recent relevant works in RVOS. The main comparison is w.r.t. the previous state-of-the-art, namely DsHmp [11];" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.714, + 0.483, + 0.819 + ], + "angle": 0, + "content": "- Methods with Context propagation: OnlineRefer [42] was the first to propose this setting. RefSAM [20] relies on SAM1 to provide frame-level masks, and then propagates the mask token to subsequent frames. A baseline that we propose is GroundingDINO + SAM2, where we use the popular grounded detector to provide boxes for the first frame, and let SAM2 track the object;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.82, + 0.483, + 0.88 + ], + "angle": 0, + "content": "- Large VLM based: Although these methods [1, 19, 45] are not comparable to ours, or previous ones, in terms of model size, we include them in the table to provide an interesting reference of performance." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.669, + 0.483, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison with standard RVOS methods. Traditional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.908, + 0.763 + ], + "angle": 0, + "content": "RVOS methods, such as ReferFormer [43] and MTTR [3], suffer a significant performance drop on the MeViS benchmark, as they are unable to solve queries which require to model long-term context. An exception is represented by LMPM and its follow up work DsHmp, which represents the state-of-the-art: these methods process the entire video clip at once, modeling multiple trajectories for all the instances in the video to select the one that fits the prompt the most. Despite this, SAMWISE outperforms DsHmp [11] on all three datasets, improving \\(\\mathcal{J}\\&\\mathcal{F}\\) of \\(+3.1\\%\\), \\(+2.1\\%\\), and \\(+5.7\\%\\), while utilizing a smaller model in terms of total parameters. Notably, we achieve this by training only \\(4.9\\mathrm{M}\\) parameters out of \\(202\\mathrm{M}\\). This result is particularly impressive, as offline methods exploit information from the entire video to handle challenges such as late-appearing objects or motion-dependent disambiguation, as opposed to our streaming approach. With respect to other methods, we outperform them significantly on MeViS, whereas the gap is smaller on Ref-YouTube-VOS and Ref-DAVIS, which contain more descriptive captions, and object-centric videos. Lastly, we also experiment with the text encoder of CLIP, which achieves state-of-the-art results on MeViS, and competitive performance on other benchmarks, while providing a more compact model with just \\(150\\mathrm{M}\\) params." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison with Context Propagation methods. Our proposed baseline GroundingDINO [25] + SAM2, while obviously flawed, being forced to predict the desired instance based on the first frame only, achieves acceptable results on Ref-DAVIS, whereas on MeViS and Ref-Youtu-VOS its performance drops of \\(11.8\\%\\) and \\(11.7\\%\\), respectively. Differently, SAMWISE, demonstrates excellent performance in both motion-dependent and static scenarios. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3401" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.09, + 0.472, + 0.187 + ], + "angle": 0, + "content": "
MLP-onlyText-to-VisualVisual-to-TextHSACMEJ&F
✓45.2
✓✓47.5
✓✓48.3
✓✓✓50.3
✓✓✓✓54.2
✓✓✓✓✓55.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.195, + 0.483, + 0.252 + ], + "angle": 0, + "content": "Table 2. Ablation of our Cross-Modal Temporal Adapter (CMT). We show the effect of not using CMT (i.e. MLP-only to prompt SAM2), vs. adding one at a time its core components. Lastly, the Conditional Memory Encoder (CME) is added." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.309 + ], + "angle": 0, + "content": "on MeViS, we outperform OnlineRefer [42] by \\(+17.2\\%\\). On Ref-YouTube-VOS and Ref-DAVIS, the gap is of \\(+5.7\\%\\) and \\(+5.8\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.541 + ], + "angle": 0, + "content": "Comparison with Large-VLM based. While comparisons with Large-VLM based approaches are not standard in RVOS evaluations, we include them in this work to provide additional context. The VLM-based solutions [1, 19, 45] are designed to leverage the extensive reasoning capabilities of VLMs to address complex textual instructions and implicit descriptions that require world knowledge [19]. This leads to improved performance in tasks like MeViS, where reasoning over motion patterns is required. However, delegating cross-modal reasoning to these VLMs incurs in significant computational overhead, whereas SAMWISE incorporates visual-text interaction directly at the feature level. Notably, SAMWISE outperforms VISA, the best VLM-based competitor by a substantial margin, respectively \\(+6\\%\\) \\(+7.7\\%\\) \\(+2.9\\%\\) on the three benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.255, + 0.565 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.619 + ], + "angle": 0, + "content": "We conduct our ablations on MeViS, as it embodies the core challenges of online RVOS. We report results on the 'valid_u' set [7], employing CLIP-B as text encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Making SAM2 Wiser. We start by showing, in Tab. 2, how each of the core components of our CMT Adapter progressively injects wisdom (i.e. knowledge about language and temporal context) into SAM2. The first line reports the result using the 'naive' solution of aligning the textual prompt to the visual features using a single learnable MLP [51]. While effective to some extent, the results show that allowing early interaction of the two modalities grants a substantial boost \\((+5.1\\%)\\) with both adapters). Adding explicit temporal feature modeling provides an additional improvements of \\(+3.9\\%\\). These results sustain our intuition that adding frame-level alignment through a MLP is not enough to obtain robust performances, and that it is essential to couple the text and visual semantics, as well as modeling temporal context. Lastly, the table shows how adding our CME is effective in mitigating SAM2 tracking bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Hierarchical Selective Attention. In the top section of Tab. 3, we study the effect of the spatial patch size in our" + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.09, + 0.858, + 0.188 + ], + "angle": 0, + "content": "
HSA Patch Size
Fixed SizeHierarchical
1488/4/216/8/4
J&F49.752.353.154.253.8
CME vs Random choice
NeverAlways1 every 4CME
J&F54.250.752.455.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.905, + 0.252 + ], + "angle": 0, + "content": "Table 3. Top: Ablation of the Patch size in our HSA, with the effect of a fixed size vs Hierarchical. Bottom: Effect of not predicting detections (Never), predicting them at every frame (Always), randomly (1 in 4) vs. using predictions of our CME." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.263, + 0.905, + 0.368 + ], + "angle": 0, + "content": "HSA module, which models the temporal evolution of over a spatial patch of size \\( P \\) across the temporal axis. Using \\( P = 1 \\) is equivalent to processing each pixel independently across frames. The table shows that including spatial context, up to 8 pixels, is beneficial. Using a hierarchical patch size that scales with the feature map resolution yields a gain of \\( +1.1\\% \\) over the fixed sized alternative." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Conditional Memory Encoder. The bottom section of Tab. 3 provides insight into our CME module. The CME, essentially, detects whenever an object in the unbiased feature maps of the current frame displays higher alignment with the textual prompts w.r.t. the currently tracked one, but SAM2 fails in noticing it due to the tracking bias (Fig. 6). The table compares the effect of Never applying such strategy (i.e., not using CME), doing it Always (i.e., at every frame), or once every 4 frames. The results show that increasing the frequency of artificial detection worsens performances, adding noise to the tracking, whereas the predictions of our CME are beneficial, with a boost of \\(+1.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.561, + 0.906, + 0.606 + ], + "angle": 0, + "content": "Adaptation strategy. In Tab. 4 we compare different adaptation strategies, including Full Fine-Tuning (FT), LoRa [13], AdaptFormer [5], and the proposed CMT." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.619, + 0.895, + 0.655 + ], + "angle": 0, + "content": "
Full-FT LoRa [13] AdaptFormer [5] CMT (ours)
MeViS J&F43.144.243.948.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.661, + 0.904, + 0.675 + ], + "angle": 0, + "content": "Table 4. Baselines: prior adapters and full-finetune, with CLIP-B." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we introduced SAMWISE, a novel approach for RVOS that builds upon the SAM2 model by incorporating i) natural language understanding, ii) temporal feature modeling, and iii) a learnable strategy to adjusts tracking focus according to visual cues that emerge over time. SAMWISE achieves SOTA performance across benchmarks while adding less than 5M parameters, without modifying SAM2 weights or using external models for visual-text alignment. We obtain an effective pipeline for applications of streaming video segmentation, addressing limitations of existing RVOS approaches, which either lack long-term context or rely on single-frame context propagation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3402" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.093, + 0.482, + 0.189 + ], + "angle": 0, + "content": "Acknowledgements. Claudia Cuttano was supported by the Sustainable Mobility Center (CNMS) which received funding from the European Union Next Generation EU (Piano Nazionale di Ripresa e Resilienza (PNRR), Missione 4 Componente 2 Investimento 1.4 \"Potenziamento strutture di ricerca e creazione di \"campioni nazionali di R&S\" su alcune Key Enabling Technologies\") with grant agreement no. CN_00000023." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.19, + 0.482, + 0.302 + ], + "angle": 0, + "content": "Gabriele Trivigno, Giuseppe Avera and Carlo Masone were supported by FAIR - Future Artificial Intelligence Research which received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.482, + 0.331 + ], + "angle": 0, + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.347, + 0.188, + 0.362 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.372, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.513 + ], + "angle": 0, + "content": "[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, 82(3):4419-4438, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[3] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7900-7916, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[8] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, 8(10), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.372, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5958-5966, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[10] Mingfei Han, Yali Wang, Zhihui Li, Lina Yao, Xiaojun Chang, and Yu Qiao. Http: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13414-13423, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[14] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5574-5583, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[15] Haojun Jiang, Jianke Zhang, Rui Huang, Chunjiang Ge, Zanlin Ni, Jiwen Lu, Jie Zhou, Shiji Song, and Gao Huang. Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[16] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27144-27153, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[17] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[19] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[20] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. Refsam: Efficiently adapting segmenting anything model for referring video object segmentation, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3403" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[22] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6555-6564, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[23] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H. Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13658-13667, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[24] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[26] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[27] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[28] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[29] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[30] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9226–9235, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[33] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[36] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[37] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[38] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In European Conference on Computer Vision, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[39] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[40] Mengmeng Wang, Jiazheng Xing, Boyuan Jiang, Jun Chen, Jianbiao Mei, Xingxing Zuo, Guang Dai, Jingdong Wang, and Yong Liu. A multimodal, multi-task adapting framework for video action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5517-5525, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[42] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. OnlineRefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[43] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[44] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3404" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 17503-17512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[45] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[46] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6449-6457, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[47] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[48] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[50] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, Xinggang Wang, et al. Evfsam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[51] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3405" + } + ] +] \ No newline at end of file diff --git a/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_origin.pdf b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d01a74c8e3e04df68e9c0174372036eb0dd6bc38 --- /dev/null +++ b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/ded78dc8-8abb-4efd-9adb-f4bf8c96b90f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:404d2878937f854a542c2e8120ae34f078f2e5fc3460db4731e129ebb46d6cbe +size 6911454 diff --git a/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/full.md b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..49f3b758ca0b874f8282bf4e76e7bfa1cba30b7e --- /dev/null +++ b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/full.md @@ -0,0 +1,336 @@ +# SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation + +Claudia Cuttano1 Gabriele Trivigno1 Gabriele Rosi1,2 Carlo Masone1,2 Giuseppe Averta1,2 +1 Politecnico di Torino 2 Focoos AI +{name.surname}@polito.it {name.surname}@focoos.ai + +# RVOS Streaming processing + +![](images/63c93735b7990a4d36388399bcb1f003cc06a254889db45e3028f16e59d58cb3.jpg) + +# Segment Anything 2 + +
SPATIAL +PROMPTTEXT +PROMPTCORRECTION +MECHANISMTEMPORAL +FEATURES
✓✗human✗
+ +![](images/c9f657dd9d09d5070003b0f31fc61b067dca54fdf013c62e4543661c774b0802.jpg) +Figure 1. SAMWISE. Our approach infuses knowledge about natural language in the Segment-Anything 2 model, adding explicit temporal cues in the feature extraction for the task of streaming-based Referring Video Segmentation (RVOS). We use a learnable mechanism to mitigate the so-called tracking bias, i.e. SAM2 tendency to overlook a correct object once it becomes identifiable, due to its ongoing tracking of a different object. Our design enables effective streaming processing for RVOS, exploiting the memory from previous frames to propagate past context. The figure shows an example where the target object is not present in the first frame, leading SAM2 to start tracking the wrong one. Afterwards, when the correct object appears, our learnable correction mechanisms guide SAM2 to switch its tracking focus. By adding in its features the notion of temporal evolution, the model is able to recognize that the new object is more aligned with the provided textual query. Finally, we exploit SAM2 tracking skills and robustness to occlusions to keep following the object. + +![](images/6d848b0dc5e95e32decd1ee44d07ef7e066202465b54bf5d29b987e25ff14670.jpg) + +# SAMWISE (ours) + +
SPATIAL +PROMPTTEXT +PROMPTCORRECTION +MECHANISMTEMPORAL +FEATURES
✓✓learnable✓
+ +# Abstract + +Referring Video Object Segmentation (RVOS) relies on natural language expressions to segment an object in a video clip. Existing methods restrict reasoning either to independent short clips, losing global context, or process the entire video offline, impairing their application in a streaming fashion. In this work, we aim to surpass these limitations and design an RVOS method capable of effectively operating in streaming-like scenarios while retaining contextual information from past frames. We build upon the Segment-Anything 2 (SAM2) model, that provides robust segmentation and tracking capabilities and is naturally suited for streaming processing. We make SAM2 wiser, by empowering it with natural language understanding and explicit temporal modeling at the feature extraction stage, without fine-tuning its weights, and without outsourcing modality interaction to external models. To this end, we introduce a novel adapter module that injects temporal information + +and multi-modal cues in the feature extraction process. We further reveal the phenomenon of tracking bias in SAM2 and propose a learnable module to adjust its tracking focus when the current frame features suggest a new object more aligned with the caption. Our proposed method, SAMWISE, achieves state-of-the-art across various benchmarks, by adding a negligible overhead of less than $5M$ parameters. Code is available at https://github.com/ClaudiaCuttano/SAMWISE. + +# 1. Introduction + +Referring video segmentation (RVOS) [9, 17, 24, 38, 43, 48] aims at segmenting and tracking specific objects of interest within video content, guided by natural language expressions [3, 10, 28]. Existing RVOS methods are mostly based on a divide and conquer paradigm, where the video is divided into shorter clips that are processed independently [3, 39, 43]. However, as demonstrated by MeViS [7], this + +solution fails in examples that require taking into account long-term motion and global context. As a workaround to handle this challenge, the state-of-the-art method [11] processes the entire video in an offline fashion, first modeling trajectories of all instances throughout the entire clip and then selecting the most appropriate one. Albeit effective, this approach is not applicable when the model has access only to a portion of the video, for example when the data at inference time are presented in a streaming fashion or due to limitations in the computational resources. The trade-off of these two paradigms is schematized in Fig. 1. To this end, OnlineRefer [42] introduced a context propagation scheme for online RVOS but relies solely on past context from a single frame, limiting its ability to capture long-term dependencies. In this work, we investigate how to exploit the memory from past frames to design an RVOS method capable of retaining global context while operating within a streaming paradigm, i.e., without requiring access to the whole video at once. This idea is inspired by the recent release of Segment-Anything 2 (SAM2) [35], a foundational model that has shown impressive capabilities in various Video Segmentation tasks thanks to a memory bank that allows to leverage long-range past information. Since SAM2 operates in a streaming fashion, extending this method to enable context-aware streaming processing in RVOS would appear a natural step. However, this entails some non-trivial challenges: + +i) Text understanding. SAM2 original design accounts only for spatial prompts (e.g. points) and lacks mechanisms to interpret semantic prompts like text, which require reasoning over visual and textual modalities. While we are the first to address the challenge of adding textual prompts to SAM2, previous methods have explored this problem for SAM-1 at image-level. These solutions [19, 50] delegate visual-textual interaction to an off-the-shelf large VLM (like BEIT-3 [41], LLaVa [21]), which generates a multimodal embedding that is used to prompt SAM-1. +ii) Temporal modeling. To segment the referred object throughout the video, it must be first recognized and then tracked. While the latter requires matching objects visual appearance across adjacent frame, the recognition problem entails modeling temporal evolution to reason over actions that unfold over multiple frames. However, SAM2 extracts frame features independently, lacking such reasoning. +iii) Tracking bias. In RVOS, the target object might be unrecognizable during certain time intervals, due to occlusions, presence of multiple instances or forthcoming actions, as in the first frames of Fig. 1. In such cases, SAM2 may start tracking an incorrect object that partially matches the textual prompt, and persist in following it, leading to what we denote as tracking bias. While SAM2 original design allows for a user to manually correct the prediction by providing a new prompt, such a strategy is not applicable in + +tasks without a human-in-the-loop like RVOS. + +In this work, we aim at making SAM2 wiser, by addressing these limitations without fine-tuning SAM2 weights, thereby preserving its original capabilities, and without outsourcing modality interaction to external, heavy models. To overcome challenges $i$ and $ii$ , we design a learnable Adapter [12] module, named Cross-Modal Temporal Adapter (CMT), with two key principles in mind: a) enabling mutual interaction between visual and linguistic modalities; and b) encoding temporal cues into visual features. Then, to generate a prompt, we follow [21, 51] and employ a learnable MLP to project the sentence embedding for the SAM2 Mask Decoder, which then outputs the final segmentation mask. In this way, we can exploit SAM2 tracking capability to segment an object given a textual query across the video. Finally, to mitigate the tracking bias problem iii), we introduce a lightweight Conditional Memory Encoder (CME) which detects when a candidate object, aligned with the text, appears in the frame, thus enabling SAM2 to dynamically refocus its tracking to the correct object as it becomes distinguishable. + +Summarizing, this paper contributes with the following: + +- We present SAMWISE, the first method that integrates natural language knowledge into SAM2 in an end-to-end solution tailored to address the challenges of RVOS. We introduce a novel adapter, namely Cross Modal Temporal (CMT) Adapter, which purposefully models temporal evolution and multi-modal interaction; +- We provide insight into the functioning of SAM2, highlighting the phenomenon of tracking bias, and introduce a learnable module (Conditional Memory Encoder) to adjust tracking based on new information; +- Our methods achieve state-of-the-art results both on traditional RVOS benchmarks (Ref-Youtu-VOS [38], RefDAVIS [17]), as well as the more challenging MeViS [7], without compromising SAM2 capabilities and adding less than $5M$ learnable parameters. + +# 2. Related works + +Referring Video Segmentation. In RVOS, the goal is to segment an object in a clip described with natural language [8, 17, 24]. Earlier works adapted image-based methods [2, 9, 17, 48], or used a spatio-temporal memory to attend to masks of previous frames [31, 38]. Subsequent works employ a DETR-like [4] structure to process multiple frames and text embeddings [3, 10, 29, 43]. All these methods process short clips independently, thus losing global context. + +Recently, [7] showed how traditional RVOS benchmarks lack challenging captions that require to disambiguate between instances and their actions, as well as occlusions and dynamic queries, highlighting how they could be solved even with image-based methods. The MeViS dataset [7] tar + +![](images/741ff6debc471c20241f49667fc2203e80c648e1e6c9e427d6dfbedc44499b16.jpg) +Figure 2. Overview of SAMWISE. We build on a frozen SAM2 and a frozen Text Encoder to segment images in video given a textual description. We incorporate the Cross-Modal Temporal Adapter (CMT) into the text and visual encoders at every intermediate layer $k$ to model temporal dynamics within visual features while contaminating each modality with the other. Then, we extract the [CLS] and verb embeddings, namely Contextual and Motion prompts, from the adapted textual features and project them through a learnable MLP. The final embedding is used to prompt the Mask Decoder, which outputs the segmentation mask. Finally, the Conditional Memory Encoder detects when a new candidate object, aligned with the caption, appears in the frame, enabling SAM2 to dynamically refocus its tracking. + +gets these scenarios, with challenging examples that previous image or clip-based methods fail to address. To this end, a few works proposed offline methods to explicitly model multiple object trajectories [11, 28], with the latter representing the state-of-the-art on MeViS. Concurrently, OnlineRefer [42] proposed a first attempt towards an online RVOS setting, with a query propagation scheme. However, its effectiveness is limited as predictions are based on a single frame. Our method builds on this paradigm by leveraging SAM2 memory bank to encode long-range past context. + +Text-prompted Segment-Anything. Recent works have provided solutions to adapt SAM-1 for text-prompted segmentation. Grounded SAM [36] employs a two step pipeline where GroundingDINO [25] generates bounding boxes for SAM-1 to produce segmentation masks. Applying such pipeline in RVOS is problematic, as potential errors in the first frame are propagated throughout the whole video. To directly prompt SAM-1, RefSAM [20] exploits a projection layer to map the textual embedding into the prompt space, while [1, 19, 45] resort to large off-the-shelf VLM to generate a multi-modal embedding that is used to prompt SAM-1. Both solutions finetune the Mask Decoder, thereby compromising its capabilities on its original task. In contrast, our work is the first to propose an end-to-end model that incorporates textual knowledge within SAM2 without fine-tuning nor relying on external models. + +Pre-Trained Knowledge Transfer. In recent years, the release of powerful pretrained models has sparked interest in the question of how to extend their skills to novel tasks, as full fine-tuning becomes increasingly impractical with growing model sizes [16, 33]. A powerful strat + +egy to address this problem relies on using Adapters [12], small trainable modules that enable efficient adaptation of pre-trained models. Following this paradigm, recent studies have explored adapting CLIP [34] for downstream tasks. At the image level, [44] inserts Transformer Decoder blocks within CLIP encoders, which entail costly Self-Attentions on all tokens. For video tasks, [40] places independent adapter modules within each encoder, whereas [15, 16, 27, 47] rely on a weight-sharing mechanism to project both modalities in a shared sub-space. Nevertheless, as features of each modality are independently extracted, none of these adapters allows explicit feature interaction, unlike our CMT, which also incorporates temporal modeling. Lastly, all these works start from a model that already includes a text encoder (CLIP), whereas ours is the first to propose an adapter for the Segment-Anything 2 model to add textual understanding, achieving robust performances while introducing less than $5\mathrm{M}$ parameters. + +# 3. SAMWISE + +Problem setting. Given an input video $\mathcal{V} = \{I_t\}_{t=1}^{T_V}$ with $T_V$ frames and a referring expression, we aim to predict a set of binary masks $S = \{s_t\}_{t=1}^{T_V}$ , $s_t \in \mathbb{R}^{H \times W}$ of the referred object. We tokenize the textual query in a set of $L$ words, $E = \{e_l\}_{l=1}^L$ , and add a global sentence representation token [CLS]. The tokens are then processed using a frozen text encoder to extract language features $\mathcal{E} \in \mathbb{R}^{L \times C_t}$ . We process videos in a streaming fashion, collecting clips of $T$ frames as they are available. Throughout the rest of the section, we use $T$ to indicate clip length. + +Overview. We first provide a brief discussion of the SAM2 + +![](images/03701d1352d3461a2bc78a46886ee2ef59a1fcfe00bc94c936f4597ebf4e2e32.jpg) +Figure 3. Architecture of our Cross Modal Temporal (CMT) Adapter, made up of Hierarchical Selective Attention (HSA) to model temporal cues, a Visual-to-Text Attention (VTA) and Text-to-Visual Attention (VTA) modules. + +model (Sec. 3.1). We then outline the pipeline of our proposed SAMWISE, starting from the prompting strategy in Sec. 3.2. In Sec. 3.3, we detail our novel Cross-Modal Temporal Adapter. Lastly, in Sec. 3.5, we discuss our learnable correction strategy, named Conditional Memory Encoder, to address the issue of tracking bias. + +# 3.1. Background: Segment-Anything + +The Segment-Anything Model 2 (SAM2) builds upon SAM-1 [18] to tackle the task of Promptable Video Object Segmentation, i.e., tracking an object in a video given a textual prompt. Following SAM-1, it consists of an image encoder, a Prompt Encoder and a Mask Decoder, which combines the image and prompt embeddings to predict segmentation masks. To enable video processing, SAM2 comes with a few modifications: $i$ ) the original ViT backbone is replaced by Hiera [37], roughly 3 times faster, which processes frames independently to provide hierarchical visual features. Hereinafter, we refer to them as memory-less features $\mathcal{F}$ ; $ii$ ) frame embeddings are not directly fed to the Mask Decoder, but they are first conditioned on memories of past predictions from a Memory Bank. We refer to these conditioned features as memory features $\mathcal{F}_{mem}$ . Lastly, $iii$ ) once the mask for the current frame is predicted, the Memory Encoder updates the Memory Bank. By design, SAM2 handles video frames as they become available, progressively encoding the past in its Memory Bank. We argue that this streaming approach is especially valuable in RVOS, enabling reasoning over a wide temporal horizon. + +# 3.2. Prompting SAM2 + +To guide the SAM2 decoder, we use a Contextual Prompt, $\mathcal{E}_C\in \mathbb{R}^{1\times C_t}$ , which encodes the high-level semantic information for the given text query, emphasizing the essential aspects of the query while downplaying less relevant elements. To this end, we employ the [CLS] embedding of text features, $\mathcal{E}$ . Furthermore, we also introduce a second prompt, the Motion Prompt $\mathcal{E}_M\in \mathbb{R}^{1\times C_t}$ , which cap + +tures action-related cues by using verb embeddings from $\mathcal{E}$ . These prompts are concatenated and projected through a learnable three-layer MLP: + +$$ +\rho = \mathbf {W} _ {\text {p r o m p t}} \left(\operatorname {C A T} \left[ \mathcal {E} _ {C}, \mathcal {E} _ {M} \right]\right). \tag {1} +$$ + +In this way, the provided prompts encode both subject-related and motion-based information. Given that in our task the textual prompt is not referred a-priori to any particular frame, we prompt SAM2 at each frame, so that the model has to balance the influence of tracking while also considering the content of each frame. We discuss more in depth this aspect in Sec. 3.5. + +# 3.3. Cross-Modal Temporal Adapter + +An adapter consists of a linear down-projection $(\mathbf{W}_{\mathrm{down}})$ to a bottleneck dimensionality, followed by an up-projection back $(\mathbf{W}_{\mathrm{up}})$ in the original space, separated by a non-linear activation function $\sigma$ . Formally, given an input feature $\mathbf{x} \in \mathbb{R}^{1 \times d}$ , the adapter function is defined as: + +$$ +A d a p t e r (\mathbf {x}) = \mathbf {x} + \sigma (\mathbf {x} \mathbf {W} _ {\text {d o w n}}) \mathbf {W} _ {\text {u p}} \tag {2} +$$ + +We build on this popular Adapter framework [12] and propose a novel Cross-Modal Temporal Adapter (CMT) (see Fig. 3) which models temporal dynamics within visual features while enriching each modality with the other. Formally, given the visual features in a clip $\mathcal{F}^k\in \mathbb{R}^{T\times H_k\times W_k\times C_k}$ and the textual features $\mathcal{E}^k\in \mathbb{R}^{L\times C}$ extracted at layer $k$ of the image and text encoders, respectively, the CMT can be formulated as: + +$$ +A d a p t e r \left(\mathcal {F} ^ {k}\right) = \mathcal {F} ^ {k} + h \left(\mathcal {F} ^ {k} \mathbf {W} _ {\text {d o w n , v}}, \mathcal {E} ^ {k} \mathbf {W} _ {\text {d o w n , t}}\right) \mathbf {W} _ {\text {u p , v}} +$$ + +$$ +\operatorname {A d p a t e r} \left(\mathcal {E} ^ {k}\right) = \mathcal {E} ^ {k} + h \left(\mathcal {E} ^ {k} \mathbf {W} _ {\text {d o w n , t}}, \mathcal {F} ^ {k} \mathbf {W} _ {\text {d o w n , v}}\right) \mathbf {W} _ {\text {u p , t}} \tag {3} +$$ + +where $\mathbf{W}_{\mathrm{down},\mathrm{v}}$ , $\mathbf{W}_{\mathrm{down},\mathrm{t}}$ , $\mathbf{W}_{\mathrm{up},\mathrm{v}}$ , $\mathbf{W}_{\mathrm{up},\mathrm{t}}$ are modality specific down- and up-projections weights and $h$ is our proposed adapter function. The adapter output is summed with the original features, allowing the model to retain the original encoding while incorporating temporal and cross-modal reasoning. We integrate the Cross-Modal Temporal Adapter (CMT) into the frozen text and visual encoders at every intermediate layer $k$ . In the following paragraphs we detail the temporal and cross-modal adaptation functions, which are tightly coupled in our Adapter module. + +Temporal Adaptation. Our approach aims to embed motion cues directly into the frame-level features of SAM2. Previous works based on Adapters either perform self-attention (SA) over all tokens in a clip [16], which is costly, or restrict the attention to the temporal axis for each pixel [22, 23]. We observe that, within a video, object motion across adjacent frames typically spans a localized region of the image [32]. Consequently, a given element of the feature map primarily benefits from interactions with its + +![](images/465da27ee6136a28cf0059b52de94bf6729403a9b200dc8e7284f2576e3d77c0.jpg) +Figure 4. Scheme of our Hierarchical Selective Attention (HSA), modeling temporal evolution of features in our adapter. + +spatial and temporal neighbors, rather than requiring long-range connections across the entire feature map. Building on this intuition, we introduce a Hierarchical Selective Attention (HSA) mechanism, illustrated in Fig. 4. By modeling interactions among spatially and temporally proximal regions, HSA reduces unnecessary computations while capturing motion-based context. + +Formally, at layer $k$ , given the set of feature maps for a $T$ -frames clip: $\mathcal{F}^k \in \mathbb{R}^{T \times H_k \times W_k \times C_k}$ , we decompose this feature volume into non-overlapping, 3-D spatio-temporal patches of size $T \times P \times P$ , obtaining $N = H_k W_k / P^2$ sub-volumes. These sub-volumes, considered pixelwise, can be represented as a set of tokens $F^k[n] = \left\{x_{i,j,t}^{k,n} \in \mathbb{R}^{C_k} : i \in 1, \dots, P, j \in 1, \dots, P, t \in 1, \dots, T\right\}$ . To encode spatio-temporal positioning, to each vector we add a spatial $(e[i,j])$ and a temporal $(e[t])$ sinusoidal positional embeddings, in 2-D and 1-D formats, respectively. Specifically: $x_{i,j,t}^{k,n} = x_{i,j,t}^{k,n} + e[i,j] + e[t]$ . Each sub-volume contains $M = P^2 * T$ tokens, on which we perform self-attention as follows: + +$$ +\mathbf {x} _ {i, j, t} ^ {k, n} := S A \left(\left\{\mathbf {x} _ {i ^ {\prime}, j ^ {\prime}, t ^ {\prime}} ^ {k, n} \right\} _ { \begin{array}{c} i ^ {\prime} = 1.. P \\ j ^ {\prime} = 1.. P \\ t ^ {\prime} = 1.. T \end{array} }\right). \tag {4} +$$ + +At each layer $k$ of the feature extraction process, the patch size $P$ is progressively scaled, as depicted in Fig. 4-d. This scaling adapts the sub-volume to the hierarchy of feature resolution, encoding information at multiple scales. + +Cross-Modal Adaptation. To unify text and visual representations, we encourage modality interaction from early stages of the feature extraction process through two symmetric operations: Visual-to-Text Attention (VTA) and Text-to-Visual Attention (TVA). + +Within the former, each visual feature, already enriched with temporal information through the HSA, attends to the full textual expression, allowing the model to identify candidate regions within the image based on both categorical details (e.g., the subject described in the text) and motion + +![](images/6707a5f2777873552bdefb177637112beb9491d1926fab2c9d5813c89a3701e1.jpg) +Figure 5. Cross Modal Temporal Adapter: we show via PCA that our CMT provides contextualized visual features based on the given textual prompt, compared to SAM2 original ones. + +cues (e.g., actions), facilitating early alignment with the prompt, as visible in Fig. 5. + +Formally, at layer $k$ , we consider the feature of each frame in the clip, i.e. $\mathcal{F}^k[t] \in \mathbb{R}^{H_k \times W_k \times C_k}$ , $t = 1, \ldots, T$ , and the set of textual embeddings $\mathcal{E}^k \in \mathbb{R}^{L \times C}$ to compute: + +$$ +\mathcal {F} ^ {k} [ t ] := \mathcal {F} ^ {k} [ t ] * C A \left(\mathcal {F} ^ {k} [ t ], \mathcal {E} ^ {k}\right). \tag {5} +$$ + +In parallel, as the meaning of a caption can shift significantly depending on the visual content of the associated image [6], we aim at contextualizing the textual query with the semantics provided by the visual modality. To this end, the TVA progressively enriches the linguistic tokens $\mathcal{E}^k\in \mathbb{R}^{L\times C}$ with information from the visual feature maps, averaged over the video clip $\mathcal{F}_{avg}^k$ + +$$ +\mathcal {E} ^ {k} := \mathcal {E} ^ {k} * C A \left(\mathcal {E} ^ {k}, \mathcal {F} _ {\text {a v g}} ^ {k}\right). \tag {6} +$$ + +# 3.4. Mask prediction + +At the end of the feature extraction process, we obtain the adapted visual and linguistic features, respectively $\mathcal{E}$ and $\mathcal{F}$ . To perform the final prediction, we extract the prompt $\rho$ as in Eq. (1), while the Memory Attention module generates the memory features $\mathcal{F}_{mem}$ by conditioning the visual features $\mathcal{F}$ on past predictions from the Memory Bank. The prompt $\rho$ is fed into the frozen Mask Decoder $\mathcal{D}_{dec}$ , which generates the output mask $\mathcal{P}_M \in \mathbb{R}^{1 \times H \times W}$ and the mask token $\tau_m \in \mathbb{R}^{1 \times d}$ , i.e. an embedding representing the segmented object. Formally: + +$$ +\begin{array}{c} \tau_ {m}, \mathcal {P} _ {M} = \mathcal {D} _ {d e c} (\mathcal {F} _ {m e m}, \rho), \\ \mathcal {Y} = \mathcal {P} _ {M} > 0, \end{array} \tag {7} +$$ + +where $\mathcal{V} \in \mathbb{R}^{1 \times H \times W}$ denotes the output binary segmentation mask. Finally, the Memory Encoder updates the memory bank with $\mathcal{P}_M$ . + +# 3.5. Conditional Memory Encoder + +We identify as tracking bias the phenomenon of SAM2 tracking the wrong object when the correct one is not yet identifiable in the video, and persist in following it. This + +![](images/217775a55d75ecef3c6baeb7615ed5e4edead1bbcea8c5000c8cdd24470f57a4.jpg) +Figure 6. Effect of our Conditional Memory Encoder. The caption above requires disambiguating multiple instances of the same class (e.g., "cat") by identifying a specific action (e.g., "climbing"). Since none of the instances perform this action initially, the model begins tracking the wrong instance and fails to correct itself once the target object performs the action. The top section visualizes the effect of this tracking bias in the memory features of SAM2. Our CME detects that the cat that starts climbing is more aligned with the caption, and encodes its presence in the memory bank, allowing SAM2 to switch its focus. + +bias, as exemplified in Fig. 6, is encoded in the memory features, which are propagated to subsequent frames through the Memory Encoder. On the other hand, we observe that the memory-less features: i) contain an unbiased representation of the current frames, ii) are aligned with the textual prompt via our CMT (cf. Fig. 5), and iii) can thus be used to propose candidate instances that match the prompt without being biased by past predictions. Building on these intuitions, we derive a memory-less token $\tau_{l}$ from a cross-attention between the unbiased feature maps and the prompt. Such token represents a summary of the visual features that match the prompt. The idea is to compare it with the mask token $\tau_{m}$ generated by the Mask Decoder, to detect when they represent different objects, i.e., to detect when SAM2 is tracking an object that is not the one currently most aligned with the caption. Formally: + +$$ +\tau_ {l} = C A (\mathcal {F}, \rho) \tag {8} +$$ + +We note that we initialize (and keep frozen) the weights of the cross-attention with those from SAM2 Mask Decoder. We introduce a small learnable module, named Conditional Memory Encoder (CME), to detect such situations. When a new object is detected, a naive solution would be to compute its mask and use it to re-prompt the model at the given frame, just like a user would do, forcing SAM2 to switch its prediction. However, since the prediction computed on the memory-less features does not have access to past video + +context, it might generate false positives. Thus, we propose a soft assignment, obtained by encoding the masks of both objects in the memory bank. Essentially, the CME allows SAM2 to 'see' other objects beyond the currently tracked one, and balance the influence of past context with new information, to select the one that fits the prompt the most. In detail, our CME, illustrated in Fig. 6-bottom, concatenates the two tokens $\tau_{m},\tau_{l}$ with a learnable decision token [DEC], and performs a self-attention followed by a Linear classifier: + +$$ +\left[ z _ {D E C}, z _ {M T}, z _ {M L} \right] = S A \left(\left[ [ D E C ], \tau_ {m} [ t ], \tau_ {l} [ t ] \right]\right) \tag {9} +$$ + +$$ +p _ {d e t e c t} = \phi (z _ {D E C}) +$$ + +where $\phi$ is a linear function $\mathbb{R}^d\to \mathbb{R}^1$ . When detecting a candidate text-aligned object, (i.e., $p_{\text{detect}} > 0.5$ ), instead of directly feeding the predicted output mask $\mathcal{P}_m$ to the Memory Encoder, our module computes the unbiased output mask, namely $\mathcal{P}_L\in \mathbb{R}^{1\times H\times W}$ , to fuse it with $\mathcal{P}_m$ : + +$$ +\mathcal {P} _ {l} = \mathcal {D} _ {d e c} \left(\mathcal {F}, \rho\right) +$$ + +$$ +\mathcal {M} (h, w) = \mathbb {1} (h, w) [ h, w: \mathcal {P} _ {l} > 0 ] \tag {10} +$$ + +$$ +\mathcal {P} = \lambda * \mathcal {P} _ {l} \circ \mathcal {M} + \mathcal {P} _ {m} \circ (1 - \mathcal {M}) +$$ + +where $\mathcal{M}(h, w)$ is a binary mask whose value is zero except for the pixels corresponding to the object, and $\lambda$ is an hyperparameter weighing the influence of the memoryless prediction. The resulting mask $\mathcal{P}$ is fed to the Memory Encoder. We train the CME via self-supervision with a standard Cross-Entropy loss, by providing examples where the memory-less features highlight different objects w.r.t the one currently tracked. We discuss in detail our training protocol in the Supp. Mat.. + +# 4. Experimental results + +Dataset. We evaluate our method on MeVis [7], Ref-Youtube-VOS [2] and Ref-Davis [17]. MeViS includes 2,006 videos and features a total of 28K annotations that capture various aspects of motion. Ref-Youtu-VOS enhances the original YouTube-VOS benchmark by incorporating textual descriptions. It contains a total of 3,978 videos and approximately 15K language expressions. RefDAVIS17 builds upon DAVIS17 dataset, adding more than 1.5K linguistic annotations to 90 videos. + +Evaluation Metrics. We utilize standard evaluation metrics, region similarity $(\mathcal{J})$ , contour accuracy $(\mathcal{F})$ , and their average $(\mathcal{J} \& \mathcal{F})$ . For MeViS and Ref-Youtu-VOS we conduct the evaluation using the official challenge servers; for Ref-DAVIS17, we used the official evaluation code. + +Implementation Details. We employ Hiera-B [37] as visual extractor. As text encoder, we experiment with + +
MethodVisual EncoderText EncoderTotal ParamsMeViSRef-YouTube-VOSRef-DAVIS17
J&FJFJ&FJFJ&FJF
Large VLM based
LISA [19] [CVPR'24]ViT-HLLaVa7 B37.235.139.453.953.454.364.862.267.3
VISA [45] [ECCV'24]ViT-HChat-UniVi7 B43.540.746.361.559.863.269.466.372.5
One-Token-Seg-All [1] [NIPS'24]ViT-HPhi-33.8 B42.339.445.261.760.263.367.763.871.5
MTTR [3] [CVPR'22]V-Swin TRoBERTa-30.028.831.255.354.056.6---
TCE-RVOS [14] [WACV'24]ResNet-50RoBERTa----59.658.360.859.456.562.4
ReferFormer [43] [CVPR'22]V-Swin BRoBERTa237 M31.029.832.262.961.364.661.158.164.1
SOC [28] [NIPS'23]V-Swin BRoBERTa220 M---66.064.167.964.261.067.4
OnlineRefer [42] [ICCV'23]Swin LRoBERTa232 M32.331.533.163.561.665.564.861.667.7
LMPM [7] [ICCV'23]Swin TRoBERTa195 M37.234.240.2------
RefSAM [20] [arXiv]ViT-BT53 B---58.457.459.462.159.065.3
DsHmp [11] [CVPR'24]V-Swin BRoBERTa339 M---67.165.069.164.961.768.1
DsHmp [11] [CVPR'24]Swin TRoBERTa272 M46.443.049.8------
MUTR [46] [AAAI'24]V-Swin BRoBERTa250 M---67.565.469.666.462.870.0
GroundingDINO [25]+SAM2Hiera-BBERT240 M37.734.940.557.555.659.566.462.869.9
SAMWISE (ours)Hiera-BCLIP-B150 M48.345.451.267.265.269.368.565.671.5
SAMWISE (ours)Hiera-BRoBERTa202 M49.546.652.469.267.870.670.667.474.5
+ +Table 1. Comparison of SAMWISE against state-of-the-art RVOS methods on MeViS, Ref-Youtu-VOS and Ref-DAVIS datasets. We further include methods based on large VLMs for comparison. Bold and underline indicate the two top results. + +CLIP [34] and RoBERTa [26]. We note that the text encoder and SAM2 weights are entirely frozen and we train only the Adapters and the CME module (4.2M parameters when using CLIP and 4.9M with RoBERTa). Following [10, 20, 28, 42, 43], we undergo pre-training for 6 epochs on RefCOCO+/g [30, 49] with a learning rate at 1e-4 and fine-tune on Ref-Youtu-VOS [38] for 4 epochs with a learning rate of 1e-5, using the Adam optimizer. The model trained on the Ref-Youtu-VOS is directly evaluated on DAVIS-17 [17]. On MeViS [7], we train for 1 epoch. We set $T = 8$ + +# 4.1. Main Results + +In this section, we compare against existing works in the literature, and ablate our contributions. In the Supp. Mat. we report additional qualitative results and ablations. + +Baselines. To assess the validity of our approach, we divide the experimental comparison in the following categories: + +- Standard RVOS methods: we compare against recent relevant works in RVOS. The main comparison is w.r.t. the previous state-of-the-art, namely DsHmp [11]; +- Methods with Context propagation: OnlineRefer [42] was the first to propose this setting. RefSAM [20] relies on SAM1 to provide frame-level masks, and then propagates the mask token to subsequent frames. A baseline that we propose is GroundingDINO + SAM2, where we use the popular grounded detector to provide boxes for the first frame, and let SAM2 track the object; +- Large VLM based: Although these methods [1, 19, 45] are not comparable to ours, or previous ones, in terms of model size, we include them in the table to provide an interesting reference of performance. + +Comparison with standard RVOS methods. Traditional + +RVOS methods, such as ReferFormer [43] and MTTR [3], suffer a significant performance drop on the MeViS benchmark, as they are unable to solve queries which require to model long-term context. An exception is represented by LMPM and its follow up work DsHmp, which represents the state-of-the-art: these methods process the entire video clip at once, modeling multiple trajectories for all the instances in the video to select the one that fits the prompt the most. Despite this, SAMWISE outperforms DsHmp [11] on all three datasets, improving $\mathcal{J}\&\mathcal{F}$ of $+3.1\%$ , $+2.1\%$ , and $+5.7\%$ , while utilizing a smaller model in terms of total parameters. Notably, we achieve this by training only $4.9\mathrm{M}$ parameters out of $202\mathrm{M}$ . This result is particularly impressive, as offline methods exploit information from the entire video to handle challenges such as late-appearing objects or motion-dependent disambiguation, as opposed to our streaming approach. With respect to other methods, we outperform them significantly on MeViS, whereas the gap is smaller on Ref-YouTube-VOS and Ref-DAVIS, which contain more descriptive captions, and object-centric videos. Lastly, we also experiment with the text encoder of CLIP, which achieves state-of-the-art results on MeViS, and competitive performance on other benchmarks, while providing a more compact model with just $150\mathrm{M}$ params. + +Comparison with Context Propagation methods. Our proposed baseline GroundingDINO [25] + SAM2, while obviously flawed, being forced to predict the desired instance based on the first frame only, achieves acceptable results on Ref-DAVIS, whereas on MeViS and Ref-Youtu-VOS its performance drops of $11.8\%$ and $11.7\%$ , respectively. Differently, SAMWISE, demonstrates excellent performance in both motion-dependent and static scenarios. Specifically, + +
MLP-onlyText-to-VisualVisual-to-TextHSACMEJ&F
✓45.2
✓✓47.5
✓✓48.3
✓✓✓50.3
✓✓✓✓54.2
✓✓✓✓✓55.5
+ +on MeViS, we outperform OnlineRefer [42] by $+17.2\%$ . On Ref-YouTube-VOS and Ref-DAVIS, the gap is of $+5.7\%$ and $+5.8\%$ , respectively. + +Comparison with Large-VLM based. While comparisons with Large-VLM based approaches are not standard in RVOS evaluations, we include them in this work to provide additional context. The VLM-based solutions [1, 19, 45] are designed to leverage the extensive reasoning capabilities of VLMs to address complex textual instructions and implicit descriptions that require world knowledge [19]. This leads to improved performance in tasks like MeViS, where reasoning over motion patterns is required. However, delegating cross-modal reasoning to these VLMs incurs in significant computational overhead, whereas SAMWISE incorporates visual-text interaction directly at the feature level. Notably, SAMWISE outperforms VISA, the best VLM-based competitor by a substantial margin, respectively $+6\%$ $+7.7\%$ $+2.9\%$ on the three benchmarks. + +# 4.2. Ablation Studies + +We conduct our ablations on MeViS, as it embodies the core challenges of online RVOS. We report results on the 'valid_u' set [7], employing CLIP-B as text encoder. + +Making SAM2 Wiser. We start by showing, in Tab. 2, how each of the core components of our CMT Adapter progressively injects wisdom (i.e. knowledge about language and temporal context) into SAM2. The first line reports the result using the 'naive' solution of aligning the textual prompt to the visual features using a single learnable MLP [51]. While effective to some extent, the results show that allowing early interaction of the two modalities grants a substantial boost $(+5.1\%)$ with both adapters). Adding explicit temporal feature modeling provides an additional improvements of $+3.9\%$ . These results sustain our intuition that adding frame-level alignment through a MLP is not enough to obtain robust performances, and that it is essential to couple the text and visual semantics, as well as modeling temporal context. Lastly, the table shows how adding our CME is effective in mitigating SAM2 tracking bias. + +Hierarchical Selective Attention. In the top section of Tab. 3, we study the effect of the spatial patch size in our + +Table 2. Ablation of our Cross-Modal Temporal Adapter (CMT). We show the effect of not using CMT (i.e. MLP-only to prompt SAM2), vs. adding one at a time its core components. Lastly, the Conditional Memory Encoder (CME) is added. + +
HSA Patch Size
Fixed SizeHierarchical
1488/4/216/8/4
J&F49.752.353.154.253.8
CME vs Random choice
NeverAlways1 every 4CME
J&F54.250.752.455.5
+ +HSA module, which models the temporal evolution of over a spatial patch of size $P$ across the temporal axis. Using $P = 1$ is equivalent to processing each pixel independently across frames. The table shows that including spatial context, up to 8 pixels, is beneficial. Using a hierarchical patch size that scales with the feature map resolution yields a gain of $+1.1\%$ over the fixed sized alternative. + +Conditional Memory Encoder. The bottom section of Tab. 3 provides insight into our CME module. The CME, essentially, detects whenever an object in the unbiased feature maps of the current frame displays higher alignment with the textual prompts w.r.t. the currently tracked one, but SAM2 fails in noticing it due to the tracking bias (Fig. 6). The table compares the effect of Never applying such strategy (i.e., not using CME), doing it Always (i.e., at every frame), or once every 4 frames. The results show that increasing the frequency of artificial detection worsens performances, adding noise to the tracking, whereas the predictions of our CME are beneficial, with a boost of $+1.3\%$ . + +Adaptation strategy. In Tab. 4 we compare different adaptation strategies, including Full Fine-Tuning (FT), LoRa [13], AdaptFormer [5], and the proposed CMT. + +Table 3. Top: Ablation of the Patch size in our HSA, with the effect of a fixed size vs Hierarchical. Bottom: Effect of not predicting detections (Never), predicting them at every frame (Always), randomly (1 in 4) vs. using predictions of our CME. + +
Full-FT LoRa [13] AdaptFormer [5] CMT (ours)
MeViS J&F43.144.243.948.3
+ +Table 4. Baselines: prior adapters and full-finetune, with CLIP-B. + +# 5. Conclusion + +In this work, we introduced SAMWISE, a novel approach for RVOS that builds upon the SAM2 model by incorporating i) natural language understanding, ii) temporal feature modeling, and iii) a learnable strategy to adjusts tracking focus according to visual cues that emerge over time. SAMWISE achieves SOTA performance across benchmarks while adding less than 5M parameters, without modifying SAM2 weights or using external models for visual-text alignment. We obtain an effective pipeline for applications of streaming video segmentation, addressing limitations of existing RVOS approaches, which either lack long-term context or rely on single-frame context propagation. + +Acknowledgements. Claudia Cuttano was supported by the Sustainable Mobility Center (CNMS) which received funding from the European Union Next Generation EU (Piano Nazionale di Ripresa e Resilienza (PNRR), Missione 4 Componente 2 Investimento 1.4 "Potenziamento strutture di ricerca e creazione di "campioni nazionali di R&S" su alcune Key Enabling Technologies") with grant agreement no. CN_00000023. + +Gabriele Trivigno, Giuseppe Avera and Carlo Masone were supported by FAIR - Future Artificial Intelligence Research which received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them. + +We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources. + +# References + +[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 7, 8 +[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, 82(3):4419-4438, 2023. 2, 6 +[3] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 2, 7 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 8 +[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7900-7916, 2022. 5 +[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 1, 2, 6, 7, 8 +[8] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, 8(10), 2021. 2 + +[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5958-5966, 2018. 1, 2 +[10] Mingfei Han, Yali Wang, Zhihui Li, Lina Yao, Xiaojun Chang, and Yu Qiao. Http: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13414-13423, 2023. 1, 2, 7 +[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024. 2, 3, 7 +[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3, 4 +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 8 +[14] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5574-5583, 2024. 7 +[15] Haojun Jiang, Jianke Zhang, Rui Huang, Chunjiang Ge, Zanlin Ni, Jiwen Lu, Jie Zhou, Shiji Song, and Gao Huang. Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623, 2022. 3 +[16] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27144-27153, 2024. 3, 4 +[17] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 2, 6, 7 +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4 +[19] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3, 7, 8 +[20] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. Refsam: Efficiently adapting segmenting anything model for referring video object segmentation, 2024. 3, 7 + +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[22] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6555-6564, 2023. 4 +[23] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H. Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13658-13667, 2024. 4 +[24] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 1, 2 +[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 7 +[26] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 7 +[27] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 3 +[28] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7 +[29] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 2 +[30] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7 +[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9226–9235, 2019. 2 +[32] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 4 + +[33] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7 +[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2 +[36] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3 +[37] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 4, 6 +[38] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In European Conference on Computer Vision, 2020. 1, 2, 7 +[39] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1 +[40] Mengmeng Wang, Jiazheng Xing, Boyuan Jiang, Jun Chen, Jianbiao Mei, Xingxing Zuo, Guang Dai, Jingdong Wang, and Yong Liu. A multimodal, multi-task adapting framework for video action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5517-5525, 2024. 3 +[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2 +[42] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. OnlineRefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2, 3, 7, 8 +[43] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 2, 7 +[44] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 17503-17512, 2023. 3 +[45] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 3, 7, 8 +[46] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6449-6457, 2024. 7 +[47] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 3 +[48] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 2 +[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7 +[50] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, Xinggang Wang, et al. Evfsam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2 +[51] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 8 \ No newline at end of file diff --git a/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/images.zip b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3b2cda3b124975819e2dce4f4903e725c30d954 --- /dev/null +++ b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d166fc6f80e68cfb7b5d80ec019e9b4b2310c789e2297ce198f5049992913d44 +size 580489 diff --git a/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/layout.json b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3771a6d6680c2fba8e2c015ac5f330eff6c44d08 --- /dev/null +++ b/2025/SAMWISE_ Infusing Wisdom in SAM2 for Text-Driven Video Segmentation/layout.json @@ -0,0 +1,8783 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "SAMWISE: Infusing Wisdom in SAM2 for Text-Driven Video Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 142, + 546, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 142, + 546, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 142, + 546, + 186 + ], + "type": "text", + "content": "Claudia Cuttano1 Gabriele Trivigno1 Gabriele Rosi1,2 Carlo Masone1,2 Giuseppe Averta1,2 \n1 Politecnico di Torino 2 Focoos AI \n{name.surname}@polito.it {name.surname}@focoos.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 213, + 357, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 213, + 357, + 225 + ], + "spans": [ + { + "bbox": [ + 246, + 213, + 357, + 225 + ], + "type": "text", + "content": "RVOS Streaming processing" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 228, + 554, + 308 + ], + "blocks": [ + { + "bbox": [ + 58, + 228, + 554, + 308 + ], + "lines": [ + { + "bbox": [ + 58, + 228, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 58, + 228, + 554, + 308 + ], + "type": "image", + "image_path": "63c93735b7990a4d36388399bcb1f003cc06a254889db45e3028f16e59d58cb3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 319, + 151, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 319, + 151, + 331 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 151, + 331 + ], + "type": "text", + "content": "Segment Anything 2" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 59, + 333, + 219, + 365 + ], + "blocks": [ + { + "bbox": [ + 59, + 333, + 219, + 365 + ], + "lines": [ + { + "bbox": [ + 59, + 333, + 219, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 333, + 219, + 365 + ], + "type": "table", + "html": "
SPATIAL\nPROMPTTEXT\nPROMPTCORRECTION\nMECHANISMTEMPORAL\nFEATURES
✓✗human✗
", + "image_path": "23f7f2df1b62e57e251a3aa8998e9903248be9b7e759af68c4f014fa13e963fc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 230, + 308, + 391, + 371 + ], + "blocks": [ + { + "bbox": [ + 230, + 308, + 391, + 371 + ], + "lines": [ + { + "bbox": [ + 230, + 308, + 391, + 371 + ], + "spans": [ + { + "bbox": [ + 230, + 308, + 391, + 371 + ], + "type": "image", + "image_path": "c9f657dd9d09d5070003b0f31fc61b067dca54fdf013c62e4543661c774b0802.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 384, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 54, + 384, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. SAMWISE. Our approach infuses knowledge about natural language in the Segment-Anything 2 model, adding explicit temporal cues in the feature extraction for the task of streaming-based Referring Video Segmentation (RVOS). We use a learnable mechanism to mitigate the so-called tracking bias, i.e. SAM2 tendency to overlook a correct object once it becomes identifiable, due to its ongoing tracking of a different object. Our design enables effective streaming processing for RVOS, exploiting the memory from previous frames to propagate past context. The figure shows an example where the target object is not present in the first frame, leading SAM2 to start tracking the wrong one. Afterwards, when the correct object appears, our learnable correction mechanisms guide SAM2 to switch its tracking focus. By adding in its features the notion of temporal evolution, the model is able to recognize that the new object is more aligned with the provided textual query. Finally, we exploit SAM2 tracking skills and robustness to occlusions to keep following the object." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 455, + 312, + 471, + 329 + ], + "blocks": [ + { + "bbox": [ + 455, + 312, + 471, + 329 + ], + "lines": [ + { + "bbox": [ + 455, + 312, + 471, + 329 + ], + "spans": [ + { + "bbox": [ + 455, + 312, + 471, + 329 + ], + "type": "image", + "image_path": "6d848b0dc5e95e32decd1ee44d07ef7e066202465b54bf5d29b987e25ff14670.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 471, + 318, + 552, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 318, + 552, + 330 + ], + "spans": [ + { + "bbox": [ + 471, + 318, + 552, + 330 + ], + "type": "text", + "content": "SAMWISE (ours)" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 400, + 331, + 553, + 370 + ], + "blocks": [ + { + "bbox": [ + 400, + 331, + 553, + 370 + ], + "lines": [ + { + "bbox": [ + 400, + 331, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 400, + 331, + 553, + 370 + ], + "type": "table", + "html": "
SPATIAL\nPROMPTTEXT\nPROMPTCORRECTION\nMECHANISMTEMPORAL\nFEATURES
✓✓learnable✓
", + "image_path": "5e7e7fffafc9ef4d8a777c6d6449d81fcec3b60bd01bf47da3688d003a8186aa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 483, + 200, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 483, + 200, + 495 + ], + "spans": [ + { + "bbox": [ + 151, + 483, + 200, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 510, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 297, + 714 + ], + "type": "text", + "content": "Referring Video Object Segmentation (RVOS) relies on natural language expressions to segment an object in a video clip. Existing methods restrict reasoning either to independent short clips, losing global context, or process the entire video offline, impairing their application in a streaming fashion. In this work, we aim to surpass these limitations and design an RVOS method capable of effectively operating in streaming-like scenarios while retaining contextual information from past frames. We build upon the Segment-Anything 2 (SAM2) model, that provides robust segmentation and tracking capabilities and is naturally suited for streaming processing. We make SAM2 wiser, by empowering it with natural language understanding and explicit temporal modeling at the feature extraction stage, without fine-tuning its weights, and without outsourcing modality interaction to external models. To this end, we introduce a novel adapter module that injects temporal information" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 483, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 592 + ], + "type": "text", + "content": "and multi-modal cues in the feature extraction process. We further reveal the phenomenon of tracking bias in SAM2 and propose a learnable module to adjust its tracking focus when the current frame features suggest a new object more aligned with the caption. Our proposed method, SAMWISE, achieves state-of-the-art across various benchmarks, by adding a negligible overhead of less than " + }, + { + "bbox": [ + 313, + 483, + 556, + 592 + ], + "type": "inline_equation", + "content": "5M" + }, + { + "bbox": [ + 313, + 483, + 556, + 592 + ], + "type": "text", + "content": " parameters. Code is available at https://github.com/ClaudiaCuttano/SAMWISE." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Referring video segmentation (RVOS) [9, 17, 24, 38, 43, 48] aims at segmenting and tracking specific objects of interest within video content, guided by natural language expressions [3, 10, 28]. Existing RVOS methods are mostly based on a divide and conquer paradigm, where the video is divided into shorter clips that are processed independently [3, 39, 43]. However, as demonstrated by MeViS [7], this" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3395" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 396 + ], + "type": "text", + "content": "solution fails in examples that require taking into account long-term motion and global context. As a workaround to handle this challenge, the state-of-the-art method [11] processes the entire video in an offline fashion, first modeling trajectories of all instances throughout the entire clip and then selecting the most appropriate one. Albeit effective, this approach is not applicable when the model has access only to a portion of the video, for example when the data at inference time are presented in a streaming fashion or due to limitations in the computational resources. The trade-off of these two paradigms is schematized in Fig. 1. To this end, OnlineRefer [42] introduced a context propagation scheme for online RVOS but relies solely on past context from a single frame, limiting its ability to capture long-term dependencies. In this work, we investigate how to exploit the memory from past frames to design an RVOS method capable of retaining global context while operating within a streaming paradigm, i.e., without requiring access to the whole video at once. This idea is inspired by the recent release of Segment-Anything 2 (SAM2) [35], a foundational model that has shown impressive capabilities in various Video Segmentation tasks thanks to a memory bank that allows to leverage long-range past information. Since SAM2 operates in a streaming fashion, extending this method to enable context-aware streaming processing in RVOS would appear a natural step. However, this entails some non-trivial challenges:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 397, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 397, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 518 + ], + "type": "text", + "content": "i) Text understanding. SAM2 original design accounts only for spatial prompts (e.g. points) and lacks mechanisms to interpret semantic prompts like text, which require reasoning over visual and textual modalities. While we are the first to address the challenge of adding textual prompts to SAM2, previous methods have explored this problem for SAM-1 at image-level. These solutions [19, 50] delegate visual-textual interaction to an off-the-shelf large VLM (like BEIT-3 [41], LLaVa [21]), which generates a multimodal embedding that is used to prompt SAM-1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 520, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 604 + ], + "type": "text", + "content": "ii) Temporal modeling. To segment the referred object throughout the video, it must be first recognized and then tracked. While the latter requires matching objects visual appearance across adjacent frame, the recognition problem entails modeling temporal evolution to reason over actions that unfold over multiple frames. However, SAM2 extracts frame features independently, lacking such reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "iii) Tracking bias. In RVOS, the target object might be unrecognizable during certain time intervals, due to occlusions, presence of multiple instances or forthcoming actions, as in the first frames of Fig. 1. In such cases, SAM2 may start tracking an incorrect object that partially matches the textual prompt, and persist in following it, leading to what we denote as tracking bias. While SAM2 original design allows for a user to manually correct the prediction by providing a new prompt, such a strategy is not applicable in" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 503, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 503, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 503, + 84 + ], + "type": "text", + "content": "tasks without a human-in-the-loop like RVOS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "type": "text", + "content": "In this work, we aim at making SAM2 wiser, by addressing these limitations without fine-tuning SAM2 weights, thereby preserving its original capabilities, and without outsourcing modality interaction to external, heavy models. To overcome challenges " + }, + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 313, + 85, + 555, + 324 + ], + "type": "text", + "content": ", we design a learnable Adapter [12] module, named Cross-Modal Temporal Adapter (CMT), with two key principles in mind: a) enabling mutual interaction between visual and linguistic modalities; and b) encoding temporal cues into visual features. Then, to generate a prompt, we follow [21, 51] and employ a learnable MLP to project the sentence embedding for the SAM2 Mask Decoder, which then outputs the final segmentation mask. In this way, we can exploit SAM2 tracking capability to segment an object given a textual query across the video. Finally, to mitigate the tracking bias problem iii), we introduce a lightweight Conditional Memory Encoder (CME) which detects when a candidate object, aligned with the text, appears in the frame, thus enabling SAM2 to dynamically refocus its tracking to the correct object as it becomes distinguishable." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 325, + 541, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 541, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 541, + 338 + ], + "type": "text", + "content": "Summarizing, this paper contributes with the following:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 338, + 555, + 518 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 338, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 555, + 409 + ], + "type": "text", + "content": "- We present SAMWISE, the first method that integrates natural language knowledge into SAM2 in an end-to-end solution tailored to address the challenges of RVOS. We introduce a novel adapter, namely Cross Modal Temporal (CMT) Adapter, which purposefully models temporal evolution and multi-modal interaction;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 410, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 555, + 458 + ], + "type": "text", + "content": "- We provide insight into the functioning of SAM2, highlighting the phenomenon of tracking bias, and introduce a learnable module (Conditional Memory Encoder) to adjust tracking based on new information;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 458, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 458, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 458, + 554, + 518 + ], + "type": "text", + "content": "- Our methods achieve state-of-the-art results both on traditional RVOS benchmarks (Ref-Youtu-VOS [38], RefDAVIS [17]), as well as the more challenging MeViS [7], without compromising SAM2 capabilities and adding less than " + }, + { + "bbox": [ + 315, + 458, + 554, + 518 + ], + "type": "inline_equation", + "content": "5M" + }, + { + "bbox": [ + 315, + 458, + 554, + 518 + ], + "type": "text", + "content": " learnable parameters." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 532, + 404, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 532, + 404, + 544 + ], + "spans": [ + { + "bbox": [ + 314, + 532, + 404, + 544 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 653 + ], + "type": "text", + "content": "Referring Video Segmentation. In RVOS, the goal is to segment an object in a clip described with natural language [8, 17, 24]. Earlier works adapted image-based methods [2, 9, 17, 48], or used a spatio-temporal memory to attend to masks of previous frames [31, 38]. Subsequent works employ a DETR-like [4] structure to process multiple frames and text embeddings [3, 10, 29, 43]. All these methods process short clips independently, thus losing global context." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Recently, [7] showed how traditional RVOS benchmarks lack challenging captions that require to disambiguate between instances and their actions, as well as occlusions and dynamic queries, highlighting how they could be solved even with image-based methods. The MeViS dataset [7] tar" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3396" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 550, + 248 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 248 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 248 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 248 + ], + "type": "image", + "image_path": "741ff6debc471c20241f49667fc2203e80c648e1e6c9e427d6dfbedc44499b16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 322 + ], + "type": "text", + "content": "Figure 2. Overview of SAMWISE. We build on a frozen SAM2 and a frozen Text Encoder to segment images in video given a textual description. We incorporate the Cross-Modal Temporal Adapter (CMT) into the text and visual encoders at every intermediate layer " + }, + { + "bbox": [ + 54, + 255, + 555, + 322 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 255, + 555, + 322 + ], + "type": "text", + "content": " to model temporal dynamics within visual features while contaminating each modality with the other. Then, we extract the [CLS] and verb embeddings, namely Contextual and Motion prompts, from the adapted textual features and project them through a learnable MLP. The final embedding is used to prompt the Mask Decoder, which outputs the segmentation mask. Finally, the Conditional Memory Encoder detects when a new candidate object, aligned with the caption, appears in the frame, enabling SAM2 to dynamically refocus its tracking." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 329, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 296, + 449 + ], + "type": "text", + "content": "gets these scenarios, with challenging examples that previous image or clip-based methods fail to address. To this end, a few works proposed offline methods to explicitly model multiple object trajectories [11, 28], with the latter representing the state-of-the-art on MeViS. Concurrently, OnlineRefer [42] proposed a first attempt towards an online RVOS setting, with a query propagation scheme. However, its effectiveness is limited as predictions are based on a single frame. Our method builds on this paradigm by leveraging SAM2 memory bank to encode long-range past context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 647 + ], + "type": "text", + "content": "Text-prompted Segment-Anything. Recent works have provided solutions to adapt SAM-1 for text-prompted segmentation. Grounded SAM [36] employs a two step pipeline where GroundingDINO [25] generates bounding boxes for SAM-1 to produce segmentation masks. Applying such pipeline in RVOS is problematic, as potential errors in the first frame are propagated throughout the whole video. To directly prompt SAM-1, RefSAM [20] exploits a projection layer to map the textual embedding into the prompt space, while [1, 19, 45] resort to large off-the-shelf VLM to generate a multi-modal embedding that is used to prompt SAM-1. Both solutions finetune the Mask Decoder, thereby compromising its capabilities on its original task. In contrast, our work is the first to propose an end-to-end model that incorporates textual knowledge within SAM2 without fine-tuning nor relying on external models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": "Pre-Trained Knowledge Transfer. In recent years, the release of powerful pretrained models has sparked interest in the question of how to extend their skills to novel tasks, as full fine-tuning becomes increasingly impractical with growing model sizes [16, 33]. A powerful strat" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 329, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 545 + ], + "type": "text", + "content": "egy to address this problem relies on using Adapters [12], small trainable modules that enable efficient adaptation of pre-trained models. Following this paradigm, recent studies have explored adapting CLIP [34] for downstream tasks. At the image level, [44] inserts Transformer Decoder blocks within CLIP encoders, which entail costly Self-Attentions on all tokens. For video tasks, [40] places independent adapter modules within each encoder, whereas [15, 16, 27, 47] rely on a weight-sharing mechanism to project both modalities in a shared sub-space. Nevertheless, as features of each modality are independently extracted, none of these adapters allows explicit feature interaction, unlike our CMT, which also incorporates temporal modeling. Lastly, all these works start from a model that already includes a text encoder (CLIP), whereas ours is the first to propose an adapter for the Segment-Anything 2 model to add textual understanding, achieving robust performances while introducing less than " + }, + { + "bbox": [ + 313, + 329, + 555, + 545 + ], + "type": "inline_equation", + "content": "5\\mathrm{M}" + }, + { + "bbox": [ + 313, + 329, + 555, + 545 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 553, + 388, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 553, + 388, + 566 + ], + "spans": [ + { + "bbox": [ + 314, + 553, + 388, + 566 + ], + "type": "text", + "content": "3. SAMWISE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": "Problem setting. Given an input video " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{I_t\\}_{t=1}^{T_V}" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "T_V" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": " frames and a referring expression, we aim to predict a set of binary masks " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "S = \\{s_t\\}_{t=1}^{T_V}" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "s_t \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": " of the referred object. We tokenize the textual query in a set of " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": " words, " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "E = \\{e_l\\}_{l=1}^L" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": ", and add a global sentence representation token [CLS]. The tokens are then processed using a frozen text encoder to extract language features " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "\\mathcal{E} \\in \\mathbb{R}^{L \\times C_t}" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": ". We process videos in a streaming fashion, collecting clips of " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": " frames as they are available. Throughout the rest of the section, we use " + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 577, + 555, + 698 + ], + "type": "text", + "content": " to indicate clip length." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": "Overview. We first provide a brief discussion of the SAM2" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3397" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 187 + ], + "type": "image", + "image_path": "03701d1352d3461a2bc78a46886ee2ef59a1fcfe00bc94c936f4597ebf4e2e32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "type": "text", + "content": "Figure 3. Architecture of our Cross Modal Temporal (CMT) Adapter, made up of Hierarchical Selective Attention (HSA) to model temporal cues, a Visual-to-Text Attention (VTA) and Text-to-Visual Attention (VTA) modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "content": "model (Sec. 3.1). We then outline the pipeline of our proposed SAMWISE, starting from the prompting strategy in Sec. 3.2. In Sec. 3.3, we detail our novel Cross-Modal Temporal Adapter. Lastly, in Sec. 3.5, we discuss our learnable correction strategy, named Conditional Memory Encoder, to address the issue of tracking bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 333, + 231, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 231, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 231, + 346 + ], + "type": "text", + "content": "3.1. Background: Segment-Anything" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "content": "The Segment-Anything Model 2 (SAM2) builds upon SAM-1 [18] to tackle the task of Promptable Video Object Segmentation, i.e., tracking an object in a video given a textual prompt. Following SAM-1, it consists of an image encoder, a Prompt Encoder and a Mask Decoder, which combines the image and prompt embeddings to predict segmentation masks. To enable video processing, SAM2 comes with a few modifications: " + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "content": " ) the original ViT backbone is replaced by Hiera [37], roughly 3 times faster, which processes frames independently to provide hierarchical visual features. Hereinafter, we refer to them as memory-less features " + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "content": " ) frame embeddings are not directly fed to the Mask Decoder, but they are first conditioned on memories of past predictions from a Memory Bank. We refer to these conditioned features as memory features " + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{mem}" + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "content": ". Lastly, " + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "inline_equation", + "content": "iii" + }, + { + "bbox": [ + 55, + 350, + 296, + 602 + ], + "type": "text", + "content": " ) once the mask for the current frame is predicted, the Memory Encoder updates the Memory Bank. By design, SAM2 handles video frames as they become available, progressively encoding the past in its Memory Bank. We argue that this streaming approach is especially valuable in RVOS, enabling reasoning over a wide temporal horizon." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 162, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 162, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 162, + 624 + ], + "type": "text", + "content": "3.2. Prompting SAM2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "To guide the SAM2 decoder, we use a Contextual Prompt, " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_C\\in \\mathbb{R}^{1\\times C_t}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", which encodes the high-level semantic information for the given text query, emphasizing the essential aspects of the query while downplaying less relevant elements. To this end, we employ the [CLS] embedding of text features, " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ". Furthermore, we also introduce a second prompt, the Motion Prompt " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_M\\in \\mathbb{R}^{1\\times C_t}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", which cap" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "tures action-related cues by using verb embeddings from " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": ". These prompts are concatenated and projected through a learnable three-layer MLP:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 374, + 118, + 555, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 118, + 555, + 131 + ], + "spans": [ + { + "bbox": [ + 374, + 118, + 555, + 131 + ], + "type": "interline_equation", + "content": "\\rho = \\mathbf {W} _ {\\text {p r o m p t}} \\left(\\operatorname {C A T} \\left[ \\mathcal {E} _ {C}, \\mathcal {E} _ {M} \\right]\\right). \\tag {1}", + "image_path": "669ef59decd9978d189fa6dc78d1bdd7ee97b224e916da8623d5509fbedb22b7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 139, + 555, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 555, + 223 + ], + "type": "text", + "content": "In this way, the provided prompts encode both subject-related and motion-based information. Given that in our task the textual prompt is not referred a-priori to any particular frame, we prompt SAM2 at each frame, so that the model has to balance the influence of tracking while also considering the content of each frame. We discuss more in depth this aspect in Sec. 3.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 230, + 488, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 488, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 488, + 243 + ], + "type": "text", + "content": "3.3. Cross-Modal Temporal Adapter" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "text", + "content": "An adapter consists of a linear down-projection " + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_{\\mathrm{down}})" + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "text", + "content": " to a bottleneck dimensionality, followed by an up-projection back " + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_{\\mathrm{up}})" + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "text", + "content": " in the original space, separated by a non-linear activation function " + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "text", + "content": ". Formally, given an input feature " + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 313, + 247, + 554, + 307 + ], + "type": "text", + "content": ", the adapter function is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 316, + 555, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 316, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 358, + 316, + 555, + 330 + ], + "type": "interline_equation", + "content": "A d a p t e r (\\mathbf {x}) = \\mathbf {x} + \\sigma (\\mathbf {x} \\mathbf {W} _ {\\text {d o w n}}) \\mathbf {W} _ {\\text {u p}} \\tag {2}", + "image_path": "2886ab60ac3565de786b2b5fecf4f9560e80c4c6f3e2694c6c247119fdfbf077.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "text", + "content": "We build on this popular Adapter framework [12] and propose a novel Cross-Modal Temporal Adapter (CMT) (see Fig. 3) which models temporal dynamics within visual features while enriching each modality with the other. Formally, given the visual features in a clip " + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^k\\in \\mathbb{R}^{T\\times H_k\\times W_k\\times C_k}" + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "text", + "content": " and the textual features " + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^k\\in \\mathbb{R}^{L\\times C}" + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "text", + "content": " extracted at layer " + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 338, + 554, + 433 + ], + "type": "text", + "content": " of the image and text encoders, respectively, the CMT can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 441, + 548, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 441, + 548, + 456 + ], + "spans": [ + { + "bbox": [ + 319, + 441, + 548, + 456 + ], + "type": "interline_equation", + "content": "A d a p t e r \\left(\\mathcal {F} ^ {k}\\right) = \\mathcal {F} ^ {k} + h \\left(\\mathcal {F} ^ {k} \\mathbf {W} _ {\\text {d o w n , v}}, \\mathcal {E} ^ {k} \\mathbf {W} _ {\\text {d o w n , t}}\\right) \\mathbf {W} _ {\\text {u p , v}}", + "image_path": "7fff0f0e2f23ba4782a803498214f751a8bf0c362937646e1e0048a700de7153.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 457, + 553, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 457, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 322, + 457, + 553, + 482 + ], + "type": "interline_equation", + "content": "\\operatorname {A d p a t e r} \\left(\\mathcal {E} ^ {k}\\right) = \\mathcal {E} ^ {k} + h \\left(\\mathcal {E} ^ {k} \\mathbf {W} _ {\\text {d o w n , t}}, \\mathcal {F} ^ {k} \\mathbf {W} _ {\\text {d o w n , v}}\\right) \\mathbf {W} _ {\\text {u p , t}} \\tag {3}", + "image_path": "51fda84507b185d1f3f16f034005a14e1b853b63387bf4a2c3e1b20b587cbdb8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down},\\mathrm{v}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down},\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up},\\mathrm{v}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up},\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": " are modality specific down- and up-projections weights and " + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": " is our proposed adapter function. The adapter output is summed with the original features, allowing the model to retain the original encoding while incorporating temporal and cross-modal reasoning. We integrate the Cross-Modal Temporal Adapter (CMT) into the frozen text and visual encoders at every intermediate layer " + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 483, + 554, + 602 + ], + "type": "text", + "content": ". In the following paragraphs we detail the temporal and cross-modal adaptation functions, which are tightly coupled in our Adapter module." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Temporal Adaptation. Our approach aims to embed motion cues directly into the frame-level features of SAM2. Previous works based on Adapters either perform self-attention (SA) over all tokens in a clip [16], which is costly, or restrict the attention to the temporal axis for each pixel [22, 23]. We observe that, within a video, object motion across adjacent frames typically spans a localized region of the image [32]. Consequently, a given element of the feature map primarily benefits from interactions with its" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3398" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 187 + ], + "type": "image", + "image_path": "465da27ee6136a28cf0059b52de94bf6729403a9b200dc8e7284f2576e3d77c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "type": "text", + "content": "Figure 4. Scheme of our Hierarchical Selective Attention (HSA), modeling temporal evolution of features in our adapter." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 233, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 233, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 296, + 316 + ], + "type": "text", + "content": "spatial and temporal neighbors, rather than requiring long-range connections across the entire feature map. Building on this intuition, we introduce a Hierarchical Selective Attention (HSA) mechanism, illustrated in Fig. 4. By modeling interactions among spatially and temporally proximal regions, HSA reduces unnecessary computations while capturing motion-based context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": "Formally, at layer " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": ", given the set of feature maps for a " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": "-frames clip: " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^k \\in \\mathbb{R}^{T \\times H_k \\times W_k \\times C_k}" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": ", we decompose this feature volume into non-overlapping, 3-D spatio-temporal patches of size " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "T \\times P \\times P" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": ", obtaining " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "N = H_k W_k / P^2" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": " sub-volumes. These sub-volumes, considered pixelwise, can be represented as a set of tokens " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "F^k[n] = \\left\\{x_{i,j,t}^{k,n} \\in \\mathbb{R}^{C_k} : i \\in 1, \\dots, P, j \\in 1, \\dots, P, t \\in 1, \\dots, T\\right\\}" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": ". To encode spatio-temporal positioning, to each vector we add a spatial " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "(e[i,j])" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": " and a temporal " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "(e[t])" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": " sinusoidal positional embeddings, in 2-D and 1-D formats, respectively. Specifically: " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "x_{i,j,t}^{k,n} = x_{i,j,t}^{k,n} + e[i,j] + e[t]" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": ". Each sub-volume contains " + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "inline_equation", + "content": "M = P^2 * T" + }, + { + "bbox": [ + 55, + 318, + 296, + 479 + ], + "type": "text", + "content": " tokens, on which we perform self-attention as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 490, + 295, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 490, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 100, + 490, + 295, + 531 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i, j, t} ^ {k, n} := S A \\left(\\left\\{\\mathbf {x} _ {i ^ {\\prime}, j ^ {\\prime}, t ^ {\\prime}} ^ {k, n} \\right\\} _ { \\begin{array}{c} i ^ {\\prime} = 1.. P \\\\ j ^ {\\prime} = 1.. P \\\\ t ^ {\\prime} = 1.. T \\end{array} }\\right). \\tag {4}", + "image_path": "de9c28ecd72f31a8928825410c79fd1980e61635d68eef53c5beaa0d266944ba.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "type": "text", + "content": "At each layer " + }, + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "type": "text", + "content": " of the feature extraction process, the patch size " + }, + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 540, + 296, + 589 + ], + "type": "text", + "content": " is progressively scaled, as depicted in Fig. 4-d. This scaling adapts the sub-volume to the hierarchy of feature resolution, encoding information at multiple scales." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 593, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 652 + ], + "type": "text", + "content": "Cross-Modal Adaptation. To unify text and visual representations, we encourage modality interaction from early stages of the feature extraction process through two symmetric operations: Visual-to-Text Attention (VTA) and Text-to-Visual Attention (TVA)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Within the former, each visual feature, already enriched with temporal information through the HSA, attends to the full textual expression, allowing the model to identify candidate regions within the image based on both categorical details (e.g., the subject described in the text) and motion" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 167 + ], + "type": "image", + "image_path": "6707a5f2777873552bdefb177637112beb9491d1926fab2c9d5813c89a3701e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 179, + 554, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 179, + 554, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 554, + 213 + ], + "type": "text", + "content": "Figure 5. Cross Modal Temporal Adapter: we show via PCA that our CMT provides contextualized visual features based on the given textual prompt, compared to SAM2 original ones." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 553, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 245 + ], + "type": "text", + "content": "cues (e.g., actions), facilitating early alignment with the prompt, as visible in Fig. 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "text", + "content": "Formally, at layer " + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "text", + "content": ", we consider the feature of each frame in the clip, i.e. " + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^k[t] \\in \\mathbb{R}^{H_k \\times W_k \\times C_k}" + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "inline_equation", + "content": "t = 1, \\ldots, T" + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "text", + "content": ", and the set of textual embeddings " + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^k \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 313, + 245, + 553, + 281 + ], + "type": "text", + "content": " to compute:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 289, + 553, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 289, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 364, + 289, + 553, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {k} [ t ] := \\mathcal {F} ^ {k} [ t ] * C A \\left(\\mathcal {F} ^ {k} [ t ], \\mathcal {E} ^ {k}\\right). \\tag {5}", + "image_path": "35d53ec27c17140d9dcd790a23dbde0f50602396640ee49f018fa71c8df8814e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 397 + ], + "type": "text", + "content": "In parallel, as the meaning of a caption can shift significantly depending on the visual content of the associated image [6], we aim at contextualizing the textual query with the semantics provided by the visual modality. To this end, the TVA progressively enriches the linguistic tokens " + }, + { + "bbox": [ + 313, + 312, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^k\\in \\mathbb{R}^{L\\times C}" + }, + { + "bbox": [ + 313, + 312, + 554, + 397 + ], + "type": "text", + "content": " with information from the visual feature maps, averaged over the video clip " + }, + { + "bbox": [ + 313, + 312, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{avg}^k" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 406, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 406, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 376, + 406, + 553, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {E} ^ {k} := \\mathcal {E} ^ {k} * C A \\left(\\mathcal {E} ^ {k}, \\mathcal {F} _ {\\text {a v g}} ^ {k}\\right). \\tag {6}", + "image_path": "480f39c8a6ae4a764c16242721529200ba4514ec86c4048367d0f03585fc9c02.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 436, + 413, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 413, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 413, + 449 + ], + "type": "text", + "content": "3.4. Mask prediction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": "At the end of the feature extraction process, we obtain the adapted visual and linguistic features, respectively " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": ". To perform the final prediction, we extract the prompt " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": " as in Eq. (1), while the Memory Attention module generates the memory features " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{mem}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": " by conditioning the visual features " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": " on past predictions from the Memory Bank. The prompt " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": " is fed into the frozen Mask Decoder " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{dec}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": ", which generates the output mask " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_M \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": " and the mask token " + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\tau_m \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 313, + 454, + 554, + 574 + ], + "type": "text", + "content": ", i.e. an embedding representing the segmented object. Formally:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 582, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 582, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 376, + 582, + 553, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\tau_ {m}, \\mathcal {P} _ {M} = \\mathcal {D} _ {d e c} (\\mathcal {F} _ {m e m}, \\rho), \\\\ \\mathcal {Y} = \\mathcal {P} _ {M} > 0, \\end{array} \\tag {7}", + "image_path": "2c4919988c813283c76da3b21d20674949d3c2f699b55dc934ecc8a9082e4c55.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "text", + "content": " denotes the output binary segmentation mask. Finally, the Memory Encoder updates the memory bank with " + }, + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_M" + }, + { + "bbox": [ + 313, + 616, + 553, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 660, + 478, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 478, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 478, + 673 + ], + "type": "text", + "content": "3.5. Conditional Memory Encoder" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We identify as tracking bias the phenomenon of SAM2 tracking the wrong object when the correct one is not yet identifiable in the video, and persist in following it. This" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3399" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 255 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 255 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 255 + ], + "type": "image", + "image_path": "217775a55d75ecef3c6baeb7615ed5e4edead1bbcea8c5000c8cdd24470f57a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 264, + 296, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 374 + ], + "type": "text", + "content": "Figure 6. Effect of our Conditional Memory Encoder. The caption above requires disambiguating multiple instances of the same class (e.g., \"cat\") by identifying a specific action (e.g., \"climbing\"). Since none of the instances perform this action initially, the model begins tracking the wrong instance and fails to correct itself once the target object performs the action. The top section visualizes the effect of this tracking bias in the memory features of SAM2. Our CME detects that the cat that starts climbing is more aligned with the caption, and encodes its presence in the memory bank, allowing SAM2 to switch its focus." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "type": "text", + "content": "bias, as exemplified in Fig. 6, is encoded in the memory features, which are propagated to subsequent frames through the Memory Encoder. On the other hand, we observe that the memory-less features: i) contain an unbiased representation of the current frames, ii) are aligned with the textual prompt via our CMT (cf. Fig. 5), and iii) can thus be used to propose candidate instances that match the prompt without being biased by past predictions. Building on these intuitions, we derive a memory-less token " + }, + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\tau_{l}" + }, + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "type": "text", + "content": " from a cross-attention between the unbiased feature maps and the prompt. Such token represents a summary of the visual features that match the prompt. The idea is to compare it with the mask token " + }, + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\tau_{m}" + }, + { + "bbox": [ + 54, + 382, + 296, + 574 + ], + "type": "text", + "content": " generated by the Mask Decoder, to detect when they represent different objects, i.e., to detect when SAM2 is tracking an object that is not the one currently most aligned with the caption. Formally:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 583, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 583, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 141, + 583, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\tau_ {l} = C A (\\mathcal {F}, \\rho) \\tag {8}", + "image_path": "44684cd97a0870ff12563ea27aba080c03196da4014720f142e84341dfe88391.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We note that we initialize (and keep frozen) the weights of the cross-attention with those from SAM2 Mask Decoder. We introduce a small learnable module, named Conditional Memory Encoder (CME), to detect such situations. When a new object is detected, a naive solution would be to compute its mask and use it to re-prompt the model at the given frame, just like a user would do, forcing SAM2 to switch its prediction. However, since the prediction computed on the memory-less features does not have access to past video" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "context, it might generate false positives. Thus, we propose a soft assignment, obtained by encoding the masks of both objects in the memory bank. Essentially, the CME allows SAM2 to 'see' other objects beyond the currently tracked one, and balance the influence of past context with new information, to select the one that fits the prompt the most. In detail, our CME, illustrated in Fig. 6-bottom, concatenates the two tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\tau_{m},\\tau_{l}" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " with a learnable decision token [DEC], and performs a self-attention followed by a Linear classifier:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 198, + 555, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 198, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 328, + 198, + 555, + 223 + ], + "type": "interline_equation", + "content": "\\left[ z _ {D E C}, z _ {M T}, z _ {M L} \\right] = S A \\left(\\left[ [ D E C ], \\tau_ {m} [ t ], \\tau_ {l} [ t ] \\right]\\right) \\tag {9}", + "image_path": "39e587a72cbffb8d7f6aaa4b2413a07d6f93e878957a2f44c75b78179d6e912e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 378, + 224, + 460, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 224, + 460, + 235 + ], + "spans": [ + { + "bbox": [ + 378, + 224, + 460, + 235 + ], + "type": "interline_equation", + "content": "p _ {d e t e c t} = \\phi (z _ {D E C})", + "image_path": "968a2d4d4eef8ed9b6a84782efaf5e49f3b05866533b5308115f8b3ee65ecfe8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": " is a linear function " + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d\\to \\mathbb{R}^1" + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": ". When detecting a candidate text-aligned object, (i.e., " + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "inline_equation", + "content": "p_{\\text{detect}} > 0.5" + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": "), instead of directly feeding the predicted output mask " + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_m" + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": " to the Memory Encoder, our module computes the unbiased output mask, namely " + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_L\\in \\mathbb{R}^{1\\times H\\times W}" + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": ", to fuse it with " + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_m" + }, + { + "bbox": [ + 313, + 245, + 555, + 307 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 394, + 316, + 468, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 316, + 468, + 329 + ], + "spans": [ + { + "bbox": [ + 394, + 316, + 468, + 329 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {l} = \\mathcal {D} _ {d e c} \\left(\\mathcal {F}, \\rho\\right)", + "image_path": "0414d88eeab2dc11f384fbb2611e92b28551e3a32b19cc72466abb05e09c7dc8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 357, + 331, + 553, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 331, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 357, + 331, + 553, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (h, w) = \\mathbb {1} (h, w) [ h, w: \\mathcal {P} _ {l} > 0 ] \\tag {10}", + "image_path": "95758da9d9cc54e2d0d984da196e0b519eabfad1749fff7f2a6956441baea4ca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 346, + 511, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 346, + 511, + 359 + ], + "spans": [ + { + "bbox": [ + 367, + 346, + 511, + 359 + ], + "type": "interline_equation", + "content": "\\mathcal {P} = \\lambda * \\mathcal {P} _ {l} \\circ \\mathcal {M} + \\mathcal {P} _ {m} \\circ (1 - \\mathcal {M})", + "image_path": "87c28fb73c09f14af192c1e8c7ff3487f991f710069cd9cba6d85ebd714f53fb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(h, w)" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " is a binary mask whose value is zero except for the pixels corresponding to the object, and " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " is an hyperparameter weighing the influence of the memoryless prediction. The resulting mask " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " is fed to the Memory Encoder. We train the CME via self-supervision with a standard Cross-Entropy loss, by providing examples where the memory-less features highlight different objects w.r.t the one currently tracked. We discuss in detail our training protocol in the Supp. Mat.." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 488, + 437, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 437, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 437, + 502 + ], + "type": "text", + "content": "4. Experimental results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 513, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 620 + ], + "type": "text", + "content": "Dataset. We evaluate our method on MeVis [7], Ref-Youtube-VOS [2] and Ref-Davis [17]. MeViS includes 2,006 videos and features a total of 28K annotations that capture various aspects of motion. Ref-Youtu-VOS enhances the original YouTube-VOS benchmark by incorporating textual descriptions. It contains a total of 3,978 videos and approximately 15K language expressions. RefDAVIS17 builds upon DAVIS17 dataset, adding more than 1.5K linguistic annotations to 90 videos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": "Evaluation Metrics. We utilize standard evaluation metrics, region similarity " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "(\\mathcal{J})" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": ", contour accuracy " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": ", and their average " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "(\\mathcal{J} \\& \\mathcal{F})" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": ". For MeViS and Ref-Youtu-VOS we conduct the evaluation using the official challenge servers; for Ref-DAVIS17, we used the official evaluation code." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. We employ Hiera-B [37] as visual extractor. As text encoder, we experiment with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3400" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 277 + ], + "type": "table", + "html": "
MethodVisual EncoderText EncoderTotal ParamsMeViSRef-YouTube-VOSRef-DAVIS17
J&FJFJ&FJFJ&FJF
Large VLM based
LISA [19] [CVPR'24]ViT-HLLaVa7 B37.235.139.453.953.454.364.862.267.3
VISA [45] [ECCV'24]ViT-HChat-UniVi7 B43.540.746.361.559.863.269.466.372.5
One-Token-Seg-All [1] [NIPS'24]ViT-HPhi-33.8 B42.339.445.261.760.263.367.763.871.5
MTTR [3] [CVPR'22]V-Swin TRoBERTa-30.028.831.255.354.056.6---
TCE-RVOS [14] [WACV'24]ResNet-50RoBERTa----59.658.360.859.456.562.4
ReferFormer [43] [CVPR'22]V-Swin BRoBERTa237 M31.029.832.262.961.364.661.158.164.1
SOC [28] [NIPS'23]V-Swin BRoBERTa220 M---66.064.167.964.261.067.4
OnlineRefer [42] [ICCV'23]Swin LRoBERTa232 M32.331.533.163.561.665.564.861.667.7
LMPM [7] [ICCV'23]Swin TRoBERTa195 M37.234.240.2------
RefSAM [20] [arXiv]ViT-BT53 B---58.457.459.462.159.065.3
DsHmp [11] [CVPR'24]V-Swin BRoBERTa339 M---67.165.069.164.961.768.1
DsHmp [11] [CVPR'24]Swin TRoBERTa272 M46.443.049.8------
MUTR [46] [AAAI'24]V-Swin BRoBERTa250 M---67.565.469.666.462.870.0
GroundingDINO [25]+SAM2Hiera-BBERT240 M37.734.940.557.555.659.566.462.869.9
SAMWISE (ours)Hiera-BCLIP-B150 M48.345.451.267.265.269.368.565.671.5
SAMWISE (ours)Hiera-BRoBERTa202 M49.546.652.469.267.870.670.667.474.5
", + "image_path": "842a4e5b8d74ae0d5493958ee1aa18a061b55f9968acced1d07b13b593dc460a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 555, + 306 + ], + "type": "text", + "content": "Table 1. Comparison of SAMWISE against state-of-the-art RVOS methods on MeViS, Ref-Youtu-VOS and Ref-DAVIS datasets. We further include methods based on large VLMs for comparison. Bold and underline indicate the two top results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 315, + 294, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 294, + 436 + ], + "type": "text", + "content": "CLIP [34] and RoBERTa [26]. We note that the text encoder and SAM2 weights are entirely frozen and we train only the Adapters and the CME module (4.2M parameters when using CLIP and 4.9M with RoBERTa). Following [10, 20, 28, 42, 43], we undergo pre-training for 6 epochs on RefCOCO+/g [30, 49] with a learning rate at 1e-4 and fine-tune on Ref-Youtu-VOS [38] for 4 epochs with a learning rate of 1e-5, using the Adam optimizer. The model trained on the Ref-Youtu-VOS is directly evaluated on DAVIS-17 [17]. On MeViS [7], we train for 1 epoch. We set " + }, + { + "bbox": [ + 54, + 315, + 294, + 436 + ], + "type": "inline_equation", + "content": "T = 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 141, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 141, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 141, + 456 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": "In this section, we compare against existing works in the literature, and ablate our contributions. In the Supp. Mat. we report additional qualitative results and ablations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 529 + ], + "type": "text", + "content": "Baselines. To assess the validity of our approach, we divide the experimental comparison in the following categories:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 696 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "type": "text", + "content": "- Standard RVOS methods: we compare against recent relevant works in RVOS. The main comparison is w.r.t. the previous state-of-the-art, namely DsHmp [11];" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 648 + ], + "type": "text", + "content": "- Methods with Context propagation: OnlineRefer [42] was the first to propose this setting. RefSAM [20] relies on SAM1 to provide frame-level masks, and then propagates the mask token to subsequent frames. A baseline that we propose is GroundingDINO + SAM2, where we use the popular grounded detector to provide boxes for the first frame, and let SAM2 track the object;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 649, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 295, + 696 + ], + "type": "text", + "content": "- Large VLM based: Although these methods [1, 19, 45] are not comparable to ours, or previous ones, in terms of model size, we include them in the table to provide an interesting reference of performance." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Comparison with standard RVOS methods. Traditional" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": "RVOS methods, such as ReferFormer [43] and MTTR [3], suffer a significant performance drop on the MeViS benchmark, as they are unable to solve queries which require to model long-term context. An exception is represented by LMPM and its follow up work DsHmp, which represents the state-of-the-art: these methods process the entire video clip at once, modeling multiple trajectories for all the instances in the video to select the one that fits the prompt the most. Despite this, SAMWISE outperforms DsHmp [11] on all three datasets, improving " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\&\\mathcal{F}" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "+3.1\\%" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "+2.1\\%" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "+5.7\\%" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": ", while utilizing a smaller model in terms of total parameters. Notably, we achieve this by training only " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "4.9\\mathrm{M}" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": " parameters out of " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "202\\mathrm{M}" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": ". This result is particularly impressive, as offline methods exploit information from the entire video to handle challenges such as late-appearing objects or motion-dependent disambiguation, as opposed to our streaming approach. With respect to other methods, we outperform them significantly on MeViS, whereas the gap is smaller on Ref-YouTube-VOS and Ref-DAVIS, which contain more descriptive captions, and object-centric videos. Lastly, we also experiment with the text encoder of CLIP, which achieves state-of-the-art results on MeViS, and competitive performance on other benchmarks, while providing a more compact model with just " + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "inline_equation", + "content": "150\\mathrm{M}" + }, + { + "bbox": [ + 313, + 315, + 555, + 604 + ], + "type": "text", + "content": " params." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Comparison with Context Propagation methods. Our proposed baseline GroundingDINO [25] + SAM2, while obviously flawed, being forced to predict the desired instance based on the first frame only, achieves acceptable results on Ref-DAVIS, whereas on MeViS and Ref-Youtu-VOS its performance drops of " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "11.8\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", respectively. Differently, SAMWISE, demonstrates excellent performance in both motion-dependent and static scenarios. Specifically," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3401" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 71, + 288, + 148 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 288, + 148 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 288, + 148 + ], + "type": "table", + "html": "
MLP-onlyText-to-VisualVisual-to-TextHSACMEJ&F
✓45.2
✓✓47.5
✓✓48.3
✓✓✓50.3
✓✓✓✓54.2
✓✓✓✓✓55.5
", + "image_path": "4a98b393c94257a50abadcaa3b801981a779fd0e2e94bd11a063978af47b23ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "text", + "content": "on MeViS, we outperform OnlineRefer [42] by " + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "inline_equation", + "content": "+17.2\\%" + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "text", + "content": ". On Ref-YouTube-VOS and Ref-DAVIS, the gap is of " + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "inline_equation", + "content": "+5.7\\%" + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "inline_equation", + "content": "+5.8\\%" + }, + { + "bbox": [ + 55, + 208, + 295, + 244 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "type": "text", + "content": "Comparison with Large-VLM based. While comparisons with Large-VLM based approaches are not standard in RVOS evaluations, we include them in this work to provide additional context. The VLM-based solutions [1, 19, 45] are designed to leverage the extensive reasoning capabilities of VLMs to address complex textual instructions and implicit descriptions that require world knowledge [19]. This leads to improved performance in tasks like MeViS, where reasoning over motion patterns is required. However, delegating cross-modal reasoning to these VLMs incurs in significant computational overhead, whereas SAMWISE incorporates visual-text interaction directly at the feature level. Notably, SAMWISE outperforms VISA, the best VLM-based competitor by a substantial margin, respectively " + }, + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "type": "inline_equation", + "content": "+6\\%" + }, + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "type": "inline_equation", + "content": "+7.7\\%" + }, + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 55, + 248, + 296, + 428 + ], + "type": "text", + "content": " on the three benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 156, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 156, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 156, + 447 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 490 + ], + "type": "text", + "content": "We conduct our ablations on MeViS, as it embodies the core challenges of online RVOS. We report results on the 'valid_u' set [7], employing CLIP-B as text encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "type": "text", + "content": "Making SAM2 Wiser. We start by showing, in Tab. 2, how each of the core components of our CMT Adapter progressively injects wisdom (i.e. knowledge about language and temporal context) into SAM2. The first line reports the result using the 'naive' solution of aligning the textual prompt to the visual features using a single learnable MLP [51]. While effective to some extent, the results show that allowing early interaction of the two modalities grants a substantial boost " + }, + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "type": "inline_equation", + "content": "(+5.1\\%)" + }, + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "type": "text", + "content": " with both adapters). Adding explicit temporal feature modeling provides an additional improvements of " + }, + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "type": "inline_equation", + "content": "+3.9\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 685 + ], + "type": "text", + "content": ". These results sustain our intuition that adding frame-level alignment through a MLP is not enough to obtain robust performances, and that it is essential to couple the text and visual semantics, as well as modeling temporal context. Lastly, the table shows how adding our CME is effective in mitigating SAM2 tracking bias." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Hierarchical Selective Attention. In the top section of Tab. 3, we study the effect of the spatial patch size in our" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 345, + 71, + 525, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 154, + 295, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 199 + ], + "type": "text", + "content": "Table 2. Ablation of our Cross-Modal Temporal Adapter (CMT). We show the effect of not using CMT (i.e. MLP-only to prompt SAM2), vs. adding one at a time its core components. Lastly, the Conditional Memory Encoder (CME) is added." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 71, + 525, + 148 + ], + "lines": [ + { + "bbox": [ + 345, + 71, + 525, + 148 + ], + "spans": [ + { + "bbox": [ + 345, + 71, + 525, + 148 + ], + "type": "table", + "html": "
HSA Patch Size
Fixed SizeHierarchical
1488/4/216/8/4
J&F49.752.353.154.253.8
CME vs Random choice
NeverAlways1 every 4CME
J&F54.250.752.455.5
", + "image_path": "a9e3ac6f23a8b361755098e85787bb39d85eac5f114e9f1f94ba590e344f951b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "content": "HSA module, which models the temporal evolution of over a spatial patch of size " + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "content": " across the temporal axis. Using " + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "inline_equation", + "content": "P = 1" + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "content": " is equivalent to processing each pixel independently across frames. The table shows that including spatial context, up to 8 pixels, is beneficial. Using a hierarchical patch size that scales with the feature map resolution yields a gain of " + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 313, + 208, + 553, + 291 + ], + "type": "text", + "content": " over the fixed sized alternative." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 296, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 439 + ], + "type": "text", + "content": "Conditional Memory Encoder. The bottom section of Tab. 3 provides insight into our CME module. The CME, essentially, detects whenever an object in the unbiased feature maps of the current frame displays higher alignment with the textual prompts w.r.t. the currently tracked one, but SAM2 fails in noticing it due to the tracking bias (Fig. 6). The table compares the effect of Never applying such strategy (i.e., not using CME), doing it Always (i.e., at every frame), or once every 4 frames. The results show that increasing the frequency of artificial detection worsens performances, adding noise to the tracking, whereas the predictions of our CME are beneficial, with a boost of " + }, + { + "bbox": [ + 313, + 296, + 555, + 439 + ], + "type": "inline_equation", + "content": "+1.3\\%" + }, + { + "bbox": [ + 313, + 296, + 555, + 439 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 444, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 479 + ], + "type": "text", + "content": "Adaptation strategy. In Tab. 4 we compare different adaptation strategies, including Full Fine-Tuning (FT), LoRa [13], AdaptFormer [5], and the proposed CMT." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 323, + 490, + 547, + 518 + ], + "blocks": [ + { + "bbox": [ + 313, + 154, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 553, + 199 + ], + "type": "text", + "content": "Table 3. Top: Ablation of the Patch size in our HSA, with the effect of a fixed size vs Hierarchical. Bottom: Effect of not predicting detections (Never), predicting them at every frame (Always), randomly (1 in 4) vs. using predictions of our CME." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 490, + 547, + 518 + ], + "lines": [ + { + "bbox": [ + 323, + 490, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 323, + 490, + 547, + 518 + ], + "type": "table", + "html": "
Full-FT LoRa [13] AdaptFormer [5] CMT (ours)
MeViS J&F43.144.243.948.3
", + "image_path": "d7c9f20de28c7129c0feda7e0c143fe423fd2a94ad8874d800669735e768ba0a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 523, + 553, + 534 + ], + "lines": [ + { + "bbox": [ + 315, + 523, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 315, + 523, + 553, + 534 + ], + "type": "text", + "content": "Table 4. Baselines: prior adapters and full-finetune, with CLIP-B." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "In this work, we introduced SAMWISE, a novel approach for RVOS that builds upon the SAM2 model by incorporating i) natural language understanding, ii) temporal feature modeling, and iii) a learnable strategy to adjusts tracking focus according to visual cues that emerge over time. SAMWISE achieves SOTA performance across benchmarks while adding less than 5M parameters, without modifying SAM2 weights or using external models for visual-text alignment. We obtain an effective pipeline for applications of streaming video segmentation, addressing limitations of existing RVOS approaches, which either lack long-term context or rely on single-frame context propagation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3402" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 294, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 294, + 149 + ], + "type": "text", + "content": "Acknowledgements. Claudia Cuttano was supported by the Sustainable Mobility Center (CNMS) which received funding from the European Union Next Generation EU (Piano Nazionale di Ripresa e Resilienza (PNRR), Missione 4 Componente 2 Investimento 1.4 \"Potenziamento strutture di ricerca e creazione di \"campioni nazionali di R&S\" su alcune Key Enabling Technologies\") with grant agreement no. CN_00000023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 150, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 294, + 239 + ], + "type": "text", + "content": "Gabriele Trivigno, Giuseppe Avera and Carlo Masone were supported by FAIR - Future Artificial Intelligence Research which received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) - MISSIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 - D.D. 1555 11/10/2022, PE00000013). This manuscript reflects only the authors' views and opinions, neither the European Union nor the European Commission can be considered responsible for them." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 262 + ], + "type": "text", + "content": "We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 274, + 115, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 115, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 115, + 286 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 294, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 294, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 294, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 295, + 350 + ], + "type": "text", + "content": "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 351, + 295, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 406 + ], + "type": "text", + "content": "[2] Miriam Bellver, Carles Ventura, Carina Silberer, Ioannis Kazakos, Jordi Torres, and Xavier Giro-i Nieto. A closer look at referring expressions for video object segmentation. Multimedia Tools and Applications, 82(3):4419-4438, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 408, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 295, + 462 + ], + "type": "text", + "content": "[3] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 464, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 509 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 510, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 564 + ], + "type": "text", + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 295, + 610 + ], + "type": "text", + "content": "[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7900-7916, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 295, + 656 + ], + "type": "text", + "content": "[7] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A large-scale benchmark for video segmentation with motion expressions. In ICCV, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[8] Zihan Ding, Tianrui Hui, Shaofei Huang, Si Liu, Xuan Luo, Junshi Huang, and Xiaoming Wei. Progressive multimodal interaction network for referring video object segmentation. The 3rd Large-scale Video Object Segmentation Challenge, 8(10), 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5958-5966, 2018. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[10] Mingfei Han, Yali Wang, Zhihui Li, Lina Yao, Xiaojun Chang, and Yu Qiao. Http: Hybrid temporal-scale multimodal learning framework for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13414-13423, 2023. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[11] Shuting He and Henghui Ding. Decoupling static and hierarchical motion perception for referring video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13332-13341, 2024. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "text", + "content": "[14] Xiao Hu, Basavaraj Hampiholi, Heiko Neumann, and Jochen Lang. Temporal context enhanced referring video object segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5574-5583, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "text", + "content": "[15] Haojun Jiang, Jianke Zhang, Rui Huang, Chunjiang Ge, Zanlin Ni, Jiwen Lu, Jie Zhou, Shiji Song, and Gao Huang. Cross-modal adapter for text-video retrieval. arXiv preprint arXiv:2211.09623, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 437, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 502 + ], + "type": "text", + "content": "[16] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, Xueqing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27144-27153, 2024. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 504, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 568 + ], + "type": "text", + "content": "[17] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 624 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "text", + "content": "[19] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[20] Yonglin Li, Jing Zhang, Xiao Teng, Long Lan, and Xinwang Liu. Refsam: Efficiently adapting segmenting anything model for referring video object segmentation, 2024. 3, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3403" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "type": "text", + "content": "[22] Ruyang Liu, Jingjia Huang, Ge Li, Jiashi Feng, Xinglong Wu, and Thomas H Li. Revisiting temporal modeling for clip-based image-to-video knowledge transferring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6555-6564, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 219 + ], + "type": "text", + "content": "[23] Ruyang Liu, Chen Li, Yixiao Ge, Thomas H. Li, Ying Shan, and Ge Li. Bt-adapter: Video conversation is feasible without video instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13658-13667, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[24] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[26] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[27] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "text", + "content": "[28] Zhuoyan Luo, Yicheng Xiao, Yong Liu, Shuyan Li, Yitong Wang, Yansong Tang, Xiu Li, and Yujiu Yang. Soc: Semantic-assisted object cluster for referring video object segmentation. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[29] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920-930, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 610 + ], + "type": "text", + "content": "[30] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 658 + ], + "type": "text", + "content": "[31] Seoung Wug Oh, Joon-Young Lee, Ning Xu, and Seon Joo Kim. Video object segmentation using space-time memory networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9226–9235, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[32] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[33] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 171 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 227 + ], + "type": "text", + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[36] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "text", + "content": "[37] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[38] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In European Conference on Computer Vision, 2020. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[39] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "text", + "content": "[40] Mengmeng Wang, Jiazheng Xing, Boyuan Jiang, Jun Chen, Jianbiao Mei, Xingxing Zuo, Guang Dai, Jingdong Wang, and Yong Liu. A multimodal, multi-task adapting framework for video action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5517-5525, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 504, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 559 + ], + "type": "text", + "content": "[41] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[42] Dongming Wu, Tiancai Wang, Yuang Zhang, Xiangyu Zhang, and Jianbing Shen. OnlineRefer: A simple online baseline for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2761-2770, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "type": "text", + "content": "[43] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[44] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3404" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 486 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 17503-17512, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[45] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. Visa: Reasoning video object segmentation via large language models. arXiv preprint arXiv:2407.11325, 2024. 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "text", + "content": "[46] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by multi-modality: A unified temporal transformer for video object segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6449-6457, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[47] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[48] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 384 + ], + "type": "text", + "content": "[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 439 + ], + "type": "text", + "content": "[50] Yuxuan Zhang, Tianheng Cheng, Rui Hu, Heng Liu, Longjin Ran, Xiaoxin Chen, Wenyu Liu, Xinggang Wang, et al. Evfsam: Early vision-language fusion for text-prompted segment anything model. arXiv preprint arXiv:2406.20076, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 441, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 486 + ], + "type": "text", + "content": "[51] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3405" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_content_list.json b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b67719f49c2bcbe19fa44b247d98a69c04d45b4d --- /dev/null +++ b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "SAR3D: Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE", + "text_level": 1, + "bbox": [ + 145, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongwei Chen1 Yushi Lan1 Shangchen Zhou1 Tengfei Wang2 Xingang Pan1", + "bbox": [ + 156, + 202, + 833, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University", + "bbox": [ + 326, + 227, + 669, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 325, + 246, + 671, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://cyw-3d.github.io/projects/SAR3D/", + "bbox": [ + 330, + 263, + 661, + 282 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e496bf99f4c4ac90bc6ba30bbfac6f766fd35add1b2c4aa3d9495d0f1f5dc36f.jpg", + "image_caption": [ + "Figure 1. Our method, SAR3D, proposes a comprehensive framework for 3D generation and understanding via autoregressive modeling. For (a) 3D generation, given a single image or text prompt, SAR3D generates multi-scale 3D objects in an autoregressive manner. For (b) 3D understanding, SAR3D-LLM can interpret a 3D model and provide a detailed description." + ], + "image_footnote": [], + "bbox": [ + 93, + 295, + 637, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/735a1c30784266fb056fd2ffa4c090f00b169c18d99ef3b3569bc09c4991c048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 296, + 756, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Give a concise interpretation of the 3D data presented here.", + "bbox": [ + 784, + 301, + 880, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A small, cute, and round yellow Pikachu stuffed toy with a distinctive yellow color, pointy ears, and large black eyes, resembling the iconic Pokémon character.", + "bbox": [ + 661, + 343, + 895, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b) 3D Understanding", + "bbox": [ + 725, + 375, + 823, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 452, + 326, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autoregressive models have demonstrated remarkable success across various fields, from large language models (LLMs) to large multimodal models (LMMs) and 2D content generation, moving closer to artificial general intelligence (AGI). Despite these advances, applying autoregressive approaches to 3D object generation and understanding remains largely unexplored. This paper introduces Scale AutoRegressive 3D (SAR3D), a novel framework that leverages a multi-scale 3D vector-quantized variational autoencoder (VQVAE) to tokenize 3D objects for efficient autoregressive generation and detailed understanding. By predicting the next scale in a multi-scale latent representation instead of the next single token, SAR3D reduces generation time significantly, achieving fast 3D object generation in just 0.82 seconds on an A6000 GPU. Additionally, given the tokens enriched with hierarchical 3D-aware information, we finetune a pretrained LLM on them, enabling multimodal comprehension of 3D content. Our experiments show that SAR3D surpasses current 3D generation methods in both speed and quality and allows LLMs to interpret and caption 3D models comprehensively.", + "bbox": [ + 88, + 483, + 485, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 829, + 220, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autoregressive models have achieved remarkable success in various domains, including large language models (LLMs) [1, 5, 13, 48, 69, 70], 2D generation [63, 68, 86], and large", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multimodal models (LMMs) [2, 16, 64], marking significant strides toward artificial general intelligence (AGI). By predicting the next token [1] or scale [68], autoregressive models are trained using a simple cross-entropy loss and share similar architectures. This commonality allows them to easily benefit from the optimizations the community has developed over the years for LLMs. Nevertheless, there has been limited exploration of how this next-token/scale prediction approach can be applied to 3D object generation and understanding.", + "bbox": [ + 511, + 453, + 906, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previously, the scarcity of 3D data pushed researchers to rely on pretrained 2D diffusion models [55] as a prior to generate 3D objects via multi-view score distillation sampling (SDS) loss [50]. Following this, alternative approaches [25, 65] have focused on training feed-forward 3D reconstruction models for fast 3D reconstruction, enabled by large-scale 3D object datasets like Objaverse [14, 15]. These methods are capable of generating 3D assets in mere seconds. More recently, native 3D generative models [27, 32, 46, 90] have emerged, attempting to sample 3D assets from noise under various conditions (e.g., text or image). However, as most of these models rely on diffusion-based methods, they suffer from slow inference times. In parallel, mesh-based generative models [11, 60] attempt to generate 3D topology using autoregressive predictions, but they are limited in detail and require slow, face-by-face predictions. For 3D understanding, some studies [18, 24, 83] have attempted to finetune LLMs on 3D data to interpret the 3D world. However, these methods primarily use 3D point", + "bbox": [ + 511, + 613, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28371", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cloud representations, which are limited in capturing fine object details.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the immense potential of autoregressive next-token prediction paradigm and their underexplored status in 3D generation and understanding, we pose an important question: Can autoregressive models be effectively applied to achieve both fast 3D object generation and detailed understanding? Addressing this challenge requires a 3D tokenizer capable of encapsulating detailed information about 3D objects into compact tokens, as well as an efficient schedule for autoregressive prediction.", + "bbox": [ + 88, + 122, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Scale AutoRegressive 3D (SAR3D), a framework that leverages autoregressive models for both fast object generation and comprehensive understanding. Central to SAR3D is a multi-scale 3D vector-quantized variational autoencoder (VQVAE) capable of tokenizing 3D objects into hierarchical levels of tokens. These multi-scale tokens facilitate next-scale prediction training, significantly reducing the steps required for 3D generation compared to diffusion models and traditional next-token prediction methods. Furthermore, the tokens, enriched with 3D-aware information, are naturally compatible with LLM fine-tuning for detailed 3D understanding.", + "bbox": [ + 89, + 258, + 482, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our SAR3D introduces a multi-scale 3D VQVAE to encode multiview RGB images, along with their corresponding depth and camera parameters, into a multiscale latent triplane representation. For 3D generation, we train an autoregressive model to predict the next scale of this latent triplane based on previous scales, conditioned on a single image or text prompt. By predicting the next scale instead of the next single token, our approach significantly reduces generation time, achieving 3D object generation in only 0.82 seconds on an A6000 GPU. For 3D understanding, we use truncated scale tokens from our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to process multimodal inputs that combine text and 3D tokens. Notably, our finetuned LLM can interpret 3D tokens encoded by our VQVAE as well as those generated by our autoregressive model, supporting both 3D captioning and simultaneous generation and understanding.", + "bbox": [ + 91, + 441, + 482, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments show that SAR3D surpasses existing 3D generation methods in both speed and quality, and our VQVAE enables LLMs to generate detailed captions for 3D objects. Our key technical contributions are as follows:", + "bbox": [ + 89, + 699, + 480, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SAR3D, a framework designed for both fast 3D object generation and detailed 3D understanding.", + "- For 3D generation, our method utilizes a next-scale prediction approach for both text-to-3D and single-image-to-3D, achieving faster generation with higher quality compared to existing methods.", + "- For 3D understanding, we leverage truncated scale tokens generated by our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to interpret and describe 3D" + ], + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models, and showcasing the potential of our approach in multimodal applications.", + "bbox": [ + 526, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 135, + 663, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Generative Models. With the success of 2D diffusion models [22, 62], their adaptation for 3D generation has been widely explored. Score distillation sampling [9, 10, 50, 66, 75] leverages these 2D models to distill 3D content, yet it encounters challenges such as costly optimization, mode collapse, and the Janus problem. More recent approaches adopt a two-stage pipeline, generating multiview images first [40, 58, 59, 77] and then reconstructing 3D structures through feed-forward processes [25, 67, 82]. Although promising, these methods are constrained by the quality of multi-view image generation, which often lacks view consistency [39] and fails to scale to higher resolutions [58]. Additionally, this two-stage setup limits 3D editing capabilities due to the absence of a 3D-aware latent space.", + "bbox": [ + 511, + 160, + 903, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, native 3D diffusion models [32, 33, 35, 71, 73, 89, 90] have been introduced, offering high-quality, efficient, and scalable 3D generation. Native 3D diffusion pipelines use a two-stage training process: first encoding 3D objects into a VAE latent space [30, 31], followed by applying a latent diffusion model on the resulting codes. However, diffusion-based 3D generation is slow during inference, and its latent space cannot be easily renovated for 3D understanding. In parallel, mesh generative models [11, 60] generates 3D topology through autoregressive prediction. However, they lack details and require slow per-face prediction. In this study, we show that our autoregressive SAR3D achieves efficient sampling, better quality, and can naturally be used for 3D understanding by cascading a Large Language Model.", + "bbox": [ + 511, + 387, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive Visual Generation. Pioneered by PixelCNN [56], researchers have proposed to generate images as pixel sequences. Early research VQVAE [72] and VQGAN [17] further quantize image patches into discrete tokens and employ a transformer to learn the autoregressive priors, similar to language modeling [1]. The following research further improves its sampling speed [7] and tokenization efficiency [87]. To further improve the reconstruction quality, RQVAE [34] proposed multi-scale quantization and VAR [68] transforms it into next-scale prediction and significantly enhances the sampling speed. Parallel efforts are spent on scaling up autoregressive models on text-conditioned visual generation task [36, 54, 63, 74]. In 3D area, though some preliminary works [45, 88] studied 3D autoregressive modeling on toy dataset [6] without textures, the research on autoregressive 3D generation over large scale 3D dataset [14, 15] is missing.", + "bbox": [ + 511, + 614, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Multimodal Models. Inspired by the great success of large language models (LLMs) [5, 69, 70], large", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28372", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multimodal models (LMMs) are proposed to comprehend and generate a wide range of information beyond text-based data. Two prominent paradigms exist to train the models in an end-to-end strategy: training the model from scratch [43] or aligning pre-trained LLMs and unimodal encoders [2, 37]. The second strategy typically involves a two-stage process: alignment of the unimodal encoder with LLM's feature space, and instruction-based fine-tuning. Following up works also extend the LMMs to 3D understanding, specifically on point cloud [24, 51, 83, 84]. However, point clouds significantly ignore the details of the given 3D inputs. Here, we demonstrate that our 3D VQVAE can connect with an LLM for a detailed 3D understanding.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 303, + 230, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-scale Visual Autoregressive Generation", + "text_level": 1, + "bbox": [ + 89, + 328, + 473, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VAR [68] presents a multi-scale visual modeling method for image generation, shifting from \"next-token prediction\" to \"next-scale prediction\", which significantly improves the inference speed of autoregressive models. Given an encoded feature map $f \\in \\mathbb{R}^{h \\times w \\times C}$ of an input image $I$ , VAR quantizes $f$ into $K$ multi-scale token maps $R = (r_1, r_2, \\dots, r_K)$ at an increasingly higher resolution $h_k \\times w_k$ with $r_K$ matches the resolution of the input feature map $f$ . The autoregressive likelihood reads as:", + "bbox": [ + 89, + 352, + 483, + 488 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 501, + 483, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where each autoregressive unit $r_k \\in [V]^{h_k \\times w_k}$ is the token map at scale $k$ , and the sequence $(r_1, r_2, \\ldots, r_{k-1})$ serves as the prefix for $r_k$ . Tockenize the input image $I$ to multi-scale discrete token maps $R$ for the learning of next-scale prediction, VAR proposes a multi-scale VQVAE with a multi-scale quantizer $\\mathcal{Q}(\\cdot)$ :", + "bbox": [ + 89, + 555, + 483, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf = \\mathcal {E} (I), \\quad R = \\mathcal {Q} (f), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 660, + 482, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I$ denotes the raw image and $\\mathcal{E}$ is the image encoder. This quantization process will map $f$ to a sequence of multiscale token maps by looking up the nearest code [72] in codebook $Z\\in \\mathbb{R}^{V\\times C}$ :", + "bbox": [ + 89, + 689, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {k} ^ {(i, j)} = \\left(\\arg \\min _ {v \\in [ V ]} \\left\\| \\operatorname {l o o k u p} (Z, v) - r _ {k} ^ {(i, j)} \\right\\| _ {2}\\right) \\in [ V ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 760, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{lookup}(Z,v)$ means taking the $v$ -th vector in codebook $Z$ . To train the quantized autoencoder, $Z$ is looked up by every $z_{k}(i,j)$ to get $\\hat{f}$ , the approximation of original $f$ . Then a new image $\\hat{I}$ is reconstructed using the decoder $D(\\cdot)$ given $\\hat{f}$ :", + "bbox": [ + 89, + 806, + 483, + 882 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} = \\operatorname {l o o k u p} (Z, z), \\quad \\hat {I} = \\mathcal {D} (\\hat {f}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once fully trained, the autoencoder $\\{\\mathcal{E},\\mathcal{Q},\\mathcal{D}\\}$ will tokenize the incoming images for training the unidirectional autoregressive model.", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. PointLLM for Point Cloud Understanding", + "text_level": 1, + "bbox": [ + 511, + 143, + 874, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given multi-modal sentences containing both point clouds $P \\in \\mathbb{R}^{n \\times d}$ and text, where $n$ is the number of points and $d$ is the dimension of each point, PointLLM [83] aims to perform 3D point cloud understanding by finetuning a pretrained large language model [12, 70]. It consists of three main components: a pretrained point cloud encoder $\\Gamma_{\\mathrm{pe}}$ (e.g., Point-BERT [88]), a projector $\\Gamma_{\\mathrm{proj}}$ and a pretrained large language model backbone $\\Gamma_{\\mathrm{llm}}$ . The $\\Gamma_{\\mathrm{pe}}$ and $\\Gamma_{\\mathrm{proj}}$ project $P$ to a point cloud token sequence $Z_{p} \\in \\mathbb{R}^{m \\times c'}$ , where $m$ is the total number of tokens and $c'$ is the projected dimension of the point tokens. The final mixed sequence of tokens $Z_{m} = (z_{1}, z_{2}, \\dots, z_{l}) \\in \\mathbb{R}^{l \\times c}$ consists of both point tokens $Z_{p}$ and text tokens $Z_{t}$ :", + "bbox": [ + 511, + 167, + 906, + 366 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {p} = \\Gamma_ {\\mathrm {p r o j}} \\left(\\Gamma_ {\\mathrm {p e}} (P)\\right), \\quad Z _ {m} = \\operatorname {C o n c a t} \\left(Z _ {p}, Z _ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 372, + 906, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Z_{t}$ is obtained by tokenizer of $\\Gamma_{\\mathrm{llm}}$ and $\\mathrm{Concat}(\\cdot)$ means concatenation of two vectors. The LLM backbone $\\Gamma_{\\mathrm{llm}}$ is a GPT-style Transformers [5], which accepts a sequence of previous multi-modal tokens $Z_{< i} = (z_1, \\ldots, z_{i-1})$ and predict the next token:", + "bbox": [ + 511, + 395, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i} = \\Gamma_ {\\mathrm {l l m}} \\left(Z _ {< i}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 478, + 903, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The finetuning process has two stages. In the first stage, it freezes $\\{\\Gamma_{\\mathrm{pe}},\\Gamma_{\\mathrm{llm}}\\}$ and finetunes $\\Gamma_{\\mathrm{proj}}$ to align point features with the text token space. In the second stage, it freezes $\\Gamma_{\\mathrm{pe}}$ and finetunes $\\{\\Gamma_{\\mathrm{llm}},\\Gamma_{\\mathrm{proj}}\\}$ together.", + "bbox": [ + 511, + 501, + 905, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 573, + 604, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present SAR3D for high-quality 3D object generation and detailed understanding. First, we introduce a multi-scale 3D vector-quantized variational autoencoder (VQVAE) in Sec. 4.1, which tokenizes the input 3D model into multi-scale tokens. Fig. 2 illustrates the design of our 3D VQVAE. Next, in Sec. 4.2, with the whole sequence of different scales of the feature tokens, we train an autoregressive model to perform the next scale prediction given a single image or text prompt, which is only supervised by simple cross-entropy loss. Finally, in Sec. 4.3, we explore using truncated scales of the whole sequence to fine-tune a pretrained LLM [12, 69, 70] to handle multimodal input sequence containing both 3D and text tokens, thereby enabling the understanding of the input 3D model. Fig. 3 illustrates our 3D generation and understanding pipeline. Different from other methods [78] that train different encoders for generation and understanding, we train a single VQVAE and use the whole and truncated sequence for generation and understanding, respectively. Details of our SAR3D are shown below.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28373", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bc020b430e7537d276a1e7d75e15f90258d8e7dde158b0407979bd1b40af577.jpg", + "image_caption": [ + "Figure 2. Overview of Multi-scale VQVAE. Given a 3D model, we leverage multi-view RGB-D(depth) renderings and Plücker embeddings as the input to our multi-view encoder $\\mathcal{E}$ . The encoder predicts a continuous feature map that is then quantized by the multi-scale quantizer $\\mathcal{Q}$ , giving $R = (r_1, r_2, \\ldots, r_K)$ of latent tri-plane features. Each code of different scales share the same codebook. The triplane decoder then converts the quantized latent triplane features into the triplane representation through a plane-wise manner. The predicted triplane is multi-view supervised with the ground truth image, depth, and normal." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 911, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Multi-scale 3D VQVAE", + "text_level": 1, + "bbox": [ + 89, + 335, + 307, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As demonstrated by previous studies [4, 7, 32, 55], the key to high-quality visual generation lies in a compact latent space achieved by a specially designed variational autoencoder [30, 72]. To achieve both fast 3D generation and detailed understanding, we propose a multi-scale 3D VQVAE that maps a given 3D object into a discrete multi-scale latent space. To encode a 3D model, we leverage its multi-view posed RGB-D renderings as input. This approach offers a comprehensive representation of the 3D structure and enables compatibility with existing architectures [79].", + "bbox": [ + 88, + 359, + 482, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the input of VQVAE is a set of multi-view renderings of the 3D object from 6 views. Each rendering $M = (I, \\Delta, \\pi)$ captures essential 3D attributes, representing the object from a specific viewpoint: the RGB image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , the depth map $\\Delta \\in \\mathbb{R}^{H \\times W}$ , and the corresponding camera pose $\\pi$ . To standardize these 3D attributes, we transform the camera pose $\\pi$ into Plücker coordinates [61], expressed as $\\mathbf{p}_i = (\\mathbf{o} \\times \\mathbf{d}_{u,v}, \\mathbf{d}_{u,v}) \\in \\mathbb{R}^6$ , where $\\mathbf{o}_i \\in \\mathbb{R}^3$ denotes the camera origin, $\\mathbf{d}_{u,v} \\in \\mathbb{R}^3$ is the normalized ray direction, and $\\times$ represents the cross product. Consequently, the Plücker embedding of the camera $\\pi$ is represented as $\\mathbf{P} \\in \\mathbb{R}^{H \\times W \\times 6}$ . The final representation is formed by channel-wise concatenating these elements, resulting in $M = [I \\oplus \\Delta \\oplus \\mathbf{P}] \\in \\mathbb{R}^{H \\times W \\times 10}$ , where $\\oplus$ denotes concatenation.", + "bbox": [ + 89, + 510, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To maintain both geometry and texture details of $M$ , similar to LN3Diff [32], we encode the inputs through a multi-view convolutional encoder [59, 67]. For better 3D awareness, the latent space is designed to be a latent triplane [32, 80] $f \\in \\mathbb{R}^{3 \\times h \\times w \\times C}$ . Besides, this representation also has spatial inductive bias and is compatible with the scale and interpolation design in VAR [68]. After encoding, $f$ is interpolated to different scales and quantized using latent triplane quantization layer $\\mathcal{Q}$ :", + "bbox": [ + 88, + 736, + 482, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf = \\mathcal {E} (\\tilde {M}), \\quad R = \\mathcal {Q} (f), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is the encoder of our VQVAE and $R = (r_1, r_2, \\dots, r_K)$ is the scale sequence and $r_k \\in \\mathbb{R}^{3 \\times h_k \\times w_k \\times C}$ , where each sub latent plane $r_k^i \\in \\mathbb{R}^{h_k \\times w_k \\times C}$ is independently quantized and interpolated over the shared codebook $Z$ . Please refer to the Supp Mat for more quantization and interpolation details.", + "bbox": [ + 511, + 335, + 903, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, the decoder $\\mathcal{D}$ decodes discrete scales $R$ to triplane and render multiple views to calculate the reconstruction loss. To balance training stability and mesh extraction quality, we first train our model on volume rendering [44] with the loss reads as:", + "bbox": [ + 511, + 428, + 903, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {r e n d e r}} \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {\\mathrm {V Q}} \\mathcal {L} _ {\\mathrm {V Q}} + \\lambda_ {\\mathrm {G A N}} \\mathcal {L} _ {\\mathrm {G A N}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 513, + 903, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{render}}$ combines mean absolute error (MAE) and perceptual loss [91] between the rendered RGB-D images with masks and ground truth, $\\mathcal{L}_{\\mathrm{VQ}}$ includes both encoding error and commitment loss [72], and $\\mathcal{L}_{\\mathrm{GAN}}$ serves as the adversarial loss to encourage a perceptually rich latent space. $\\lambda_{*}$ are the corresponding loss weights.", + "bbox": [ + 511, + 542, + 905, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To facilitate 3D mesh extraction, we further finetune the model to the hybrid representation Flexicubes [57, 82] with the extra $\\mathcal{L}_{\\mathrm{flex}}$ loss:", + "bbox": [ + 511, + 632, + 905, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f l e x}} = \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 689, + 903, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{normal}}$ is MAE loss between rendered normal and ground truth, $\\mathcal{L}_{\\mathrm{reg}}$ is regularization term for Flexicubes parameters [57]. $\\lambda_{*}$ are the corresponding loss weights. Similar to LATTE3D [81], we only fine-tune the decoder of the VQVAE in this stage for stabilized training.", + "bbox": [ + 511, + 717, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. 3D Generation via Multi-scale Autoregressive Modeling", + "text_level": 1, + "bbox": [ + 511, + 801, + 903, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAR3D Transformer. We illustrate our generation framework in Fig. 3. Similar to VAR [68], we use standard GPT-style transformer [5] with AdaLN layer [49], with specific layer design following the simple rule of scaling law [28].", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28374", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0befbd1d39f5c8894dbd2364518c63db036cdb11e74545011dca48d7d85ae29d.jpg", + "image_caption": [ + "(a) 3D Generation", + "(b) 3D Understanding", + "Figure 3. Overview of 3D Generation and 3D Understanding. Given a 3D model, our 3D VQVAE encodes it into multi-scale discrete tokens for both 3D generation and understanding. In (a) 3D Generation, text or a single image is encoded by $\\mathrm{CLIP_T}$ or DINOv2, and the encoded condition features are integrated into the decoder-only transformer via cross attention. The transformer then causally predicts each scale of the latent triplane. In (b) 3D Understanding, truncated 3D tokens are first processed with an MLP projector. The large language model receives a multimodal sequence of text and 3D tokens and generates a detailed caption describing the input 3D model." + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 903, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt tri-plane latent for autoregressive prediction with Eq. 1, where different latent plane $r_k^i$ is differentiated with the corresponding learnable positional embeddings.", + "bbox": [ + 88, + 375, + 480, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional 3D Generation. Unlike feed-forward 3D reconstruction models [25, 67] that map input image into 3D, we achieve flexible multimodal 3D generation by introducing diverse conditions, as shown in Fig. 3. For text conditions, we use $\\mathrm{CLIP_T}$ [53] ViT-L text encoder and inject the text embeddings into the autoregressive models through cross attention. For the image-conditioned model, we use DINOv2 [47] ViT-L to extract local patch features and send them into the autoregressive model through pre-cross-attention block [25], which empirically yields better performance. Besides local patch features, we also leverage the pooled out feature of $\\mathrm{CLIP_T} / \\mathrm{DINOv2}$ as the start token of the sequence. Please refer to the Supp Mat for more details of our transformer blocks.", + "bbox": [ + 88, + 421, + 482, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classifier-free Guidance. As first proposed in diffusion models [22, 55], classifier-free guidance [21] (CFG) has shown effectiveness for improving generation quality and condition alignment. Therefore, we also enable CFG in our model by randomly dropping out $10\\%$ of the input condition by replacing it with a null unconditional embedding [49]. During inference, the logit $r_g$ of each token is calculated by $r_g = r_u + s(r_c - r_u)$ , given the conditional logit $r_c$ and the unconditional logit $r_u$ . $s$ stands for the scale of the classifier-free guidance.", + "bbox": [ + 88, + 633, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. SAR3D-LLM for 3D Object Understanding", + "text_level": 1, + "bbox": [ + 89, + 800, + 460, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our 3D VQVAE model provides a comprehensive encoding of the given 3D object, it can be naturally extended to 3D object understanding. Following PointLLM [83], we fine-tune LLaMA [12, 69, 70] in two stages: first, we fine-tune the projector $\\Gamma_{\\mathrm{proj}}$ while freezing the LLM $\\Gamma_{\\mathrm{llm}}$ to align", + "bbox": [ + 88, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D features with the text token space; second, we fine-tune both. As briefed in Sec. 3.2, the encoded 3D tokens $R$ are projected to the language latent space, and concatenated with the text instruction tokens $Z_{t}$ . Here, we directly use the output tokens from pre-trained SAR3D VQVAE to the projector $\\Gamma_{\\mathrm{proj}}$ . Since we only study the 3D captioning [41] task here, the instruction tokens $Z_{t}$ is fixed to $\\tilde{Z}_{t}$ which are tokenized from \"Give a concise interpretation of the 3D data presented here\". The final framework, SAR3D-LLM, supports both detailed 3D captioning given 3D object and simultaneous 3D generation and captioning given text or image.", + "bbox": [ + 511, + 375, + 906, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, a surprising observation here is that not all scales in $R$ are required for 3D understanding training. Empirically, we use the truncated scale latent codes $\\tilde{R} = (r_1, r_2, \\dots, r_{K-2})$ as the input to the LLM, which contains only $37.5\\%$ of the overall tokens required for training 3D generation. The final multimodal tokens $Z_m$ that serve as the input to the LLM reads as", + "bbox": [ + 511, + 556, + 906, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {p r o j} = \\Gamma_ {\\mathrm {p r o j}} (\\tilde {R}), \\quad Z _ {m} = \\operatorname {C o n c a t} \\left(Z _ {p r o j}, \\tilde {Z} _ {t}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 670, + 903, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Z_{proj}$ is the projected 3D tokens, and $\\mathrm{Concat}(\\cdot, \\cdot)$ means concatenation. A similar observation is also mentioned in Janus [78], where different features are required for multimodal understanding and generation. Furthermore, unlike other 3D captioning approaches such as Cap3D [41, 42], which separately extracts captions from 8 multi-view renderings and require post-processing to merge them into a unified caption, our method efficiently generates a detailed caption with a single encoding step.", + "bbox": [ + 511, + 696, + 906, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 844, + 645, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To train our model, we use renderings from G-Objaverse [15, 52] and select a high-quality subset of", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/67ab4c764fe35ad7ec08cb93346a7cdf429833952d4f064bb54f27cac5dd75b4.jpg", + "table_caption": [ + "Table 1. Quantitative Evaluation of Image-conditioned 3D Generation. We evaluate the quality of both 2D rendering and 3D shapes. As shown below, the proposed method demonstrates strong performance across all metrics. Although LGM, a multi-view images-to-3D approach, achieves slightly better performance on FID, it falls short on more advanced image quality assessment metrics such as MUSIQ and has significantly worse 3D shape quality. For multi-view to 3D methods, we also include the number of input views (V=\\#). The latency time is all profiled on Tesla V100 architecture." + ], + "table_footnote": [], + "table_body": "
MethodFID↓KID(%)↓MUSIQ↑COV(%)↑MMD(‰)↓Latency-V100 (s) ↓
Splatter-Image48.803.6530.3337.6630.690.83
OpenLRM38.411.8745.4639.3329.087.21
One-2-3-45 (V=12)88.396.3459.0233.3335.0959.23
Lara (V=4)43.741.9539.3739.3328.8411.93
CRM (V=6)45.531.9364.1038.8328.9122.10
LGM (V=4)19.930.5554.7850.8322.063.87
Shap-E138.5311.9531.5161.3319.179.54
LN3Diff29.080.8950.3955.1719.947.51
SAR3D-NeRF22.550.4265.7774.1713.631.64
SAR3D-Flexicubes27.300.6367.2471.5015.252.92
", + "bbox": [ + 133, + 162, + 861, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/942591ab501ea1692aa3e4b1923d1410b3835a9456aee65e413099d6399f5052.jpg", + "image_caption": [ + "Figure 4. Qualitative Comparison of Image-conditioned 3D Generation. Here, we compare with the state-of-the-art 3D generative models under different categories. As visualized here, our method achieves superior 3D consistency across views and generates intact objects without distortion. For more comparisons with other methods, please refer to the Supp Mat." + ], + "image_footnote": [], + "bbox": [ + 133, + 358, + 864, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "around $176K$ 3D instances, where each consists of 40 random views with RGB, normal, depth maps and camera poses. For text-conditioned generation and 3D understanding training, we use captions provided by 3DTopia [23]. For image-conditioned training, we select a random view of the corresponding 3D instance as the condition.", + "bbox": [ + 89, + 655, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In our multi-scale VQVAE, we use images with a resolution of $H = W = 256$ as input. The feature map is quantized across 10 scales, with sizes of $3 \\times (1^2, 2^2, 3^2, 4^2, 5^2, 6^2, 8^2, 10^2, 13^2, 16^2)$ . To enhance codebook utilization and stabilize 3D generation training, we follow [63, 85] by applying $\\ell_2$ -normalization to codebook vectors, setting a low codebook vector dimension $C = 8$ , and using a large codebook size $V = 16384$ .", + "bbox": [ + 89, + 747, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For 3D generation, we base our architecture on VAR [68], adding plane positional encoding for each plane.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For text-conditioned generation, the model has 16 transformer blocks with 16 heads, while for image-conditioned generation, it has 24 transformer blocks with 16 heads. We use the AdamW optimizer with a learning rate of $10^{-4}$ . For 3D understanding, we utilize the Vicuna-7B [12] checkpoint of LLaMA [70], following PointLLM [83]. The training was conducted on 7 NVIDIA A100 GPUs for the multiscale VQVAE with batch size 28, image-conditioned transformer with batch 63, text-conditioned transformer with batch size 52. For SAR3D-LLM, the stage-1 alignment is trained with batch size 140, and stage-2 with 112.", + "bbox": [ + 511, + 655, + 906, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Single Image to 3D", + "text_level": 1, + "bbox": [ + 511, + 830, + 694, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our SAR3D with three categories of methods: single-image to 3D methods (Splatter-Image [65], OpenLRM [19, 25]), multi-view image to 3D methods", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28376", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/394b8052b3e4104138e3fb694e22768b571f8df71e8bb00b9a9e92617ca9f4ce.jpg", + "image_caption": [ + "Figure 5. More results of image and text conditioned 3D generation of SAR3D." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 496, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7732b71c044fa6e9cc4250eddb7f89ae10707348a743c81a330ed7bfce6a1c24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 85, + 905, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3e40b44f446ca99b5b249ffe104bb2ab19425be150473176128d8661b68bade.jpg", + "image_caption": [ + "Figure 6. Comparison of Text-conditioned 3D Generation. We present text-conditioned 3D objects generated by SAR3D, displaying two views of each sample. Compared to baseline methods, our approach consistently yields better quality regarding geometry, texture, and text-3D alignment." + ], + "image_footnote": [], + "bbox": [ + 99, + 282, + 898, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(One-2-3-45 [38], Lara [8], CRM [76], LGM [67]), and native 3D diffusion models (Shap-E [27], LN3Diff-image [32]). Quantitatively, we benchmark rendering metrics with FID [20], KID [3], and MUSIQ [29, 92]. For 3D quality evaluation, we report Coverage Score (COV), and Minimum Matching Distance (MMD) score, as shown in Table 1. Our SAR3D demonstrates strong performance across all metrics.", + "bbox": [ + 88, + 513, + 482, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we also profile the generation speed. This timing covers the complete process, from input image processing to mesh extraction. Thanks to efficient next-scale prediction, SAR3D achieves exceptionally fast generation speeds, achieving 0.82 seconds and 1.46 seconds respectively on a single A6000 GPU. Since other baseline methods are tested on Tesla V100 GPUs, we scale our results by a factor of 2 for fair comparison in Table 1.", + "bbox": [ + 88, + 638, + 482, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative comparisons between SAR3D and existing methods is also included in Fig. 4. Compared to single-image to 3D methods like OpenLRM[19], and multi-view image to 3D methods like LGM [67], our approach achieves better 3D consistency across views and reduces distortion in generated 3D objects. Compared to native 3D diffusion models like LN3Diff [32], SAR3D produces more complete 3D models. Additional qualitative results are shown in Fig. 5. For more comparisons with other methods and", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "details for training data, please refer to the Supp Mat.", + "bbox": [ + 511, + 513, + 864, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Text to 3D", + "text_level": 1, + "bbox": [ + 511, + 539, + 630, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to image-to-3D generation, SAR3D also supports the creation of high-quality 3D assets from text prompts. As shown in Fig. 5, SAR3D generates diverse and detailed 3D objects based on the same text input. For instance, in the first and second samples, SAR3D produces different shapes for the cannon barrel and chair base, while in the third sample, it varies the texture of a wooden chest. In Fig. 6, we compare our method with other text-to-3D generation approaches, including Point-E [46], ShapE [27], 3DTopia [23], and LN3Diff [32]. Compared to these baselines, SAR3D achieves sharper visual results and better alignment with the input prompts. For example, in the second sample, SAR3D generates red patterns on the handle, closely matching the input text.. In contrast, Point-E [46] reverses the colors of the handle and blade, 3DTopia [23] produces a completely red sword, and Shap-E [27] yields a less detailed result.", + "bbox": [ + 509, + 563, + 906, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.3D Captioning", + "text_level": 1, + "bbox": [ + 511, + 830, + 663, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Object Captioning. In this section, we present the results of our 3D understanding model applied to various 3D models. As shown in Fig. 8, given the prompt \"Give", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28377", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/342f3dbdfffb53ef548eae675fca22b1af7b652223ed81edc39f08bccbd5d8e9.jpg", + "image_caption": [ + "6duDpuuDusrAupn uDuaeau 4ndU" + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 199, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bbdead4eb6127ecb043e45a869d7f142b5b2029a3c28c1d5f8617d3c92eebb3a.jpg", + "image_caption": [ + "A blue, curved sword with a long and slender blade, featuring a hilt or handle at the base. It has a metallic sheen and is associated with the protagonist, Link, in The Legend of Zelda video game franchise." + ], + "image_footnote": [], + "bbox": [ + 245, + 88, + 333, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e7b6f1d40038eada258e0a473210c866faa2feb38eadca6f5141a3bf72d26f10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 88, + 480, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fec023d28cea05bae7e86a90874173517b6785f05b5f4f682ed26ba27403f23e.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 596, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b6d5e6d2083e94f451cdf09f447dd4359bac8096620c4d1593760092cd2a50f.jpg", + "image_caption": [ + "with a red cap, depicted as the spotted" + ], + "image_footnote": [], + "bbox": [ + 625, + 90, + 745, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bde5fd108111525d4fa9e6d1feb0117ac7ef01e94eb2e5d37d05ff79722a799a.jpg", + "image_caption": [ + "\"Bookshelf\"", + "a white shelving unit in the shape of a building." + ], + "image_footnote": [], + "bbox": [ + 764, + 114, + 887, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 625, + 118, + 679, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/689366401793067a7315fe96d57f2aa588f59abb6f149e773a500d916bbd99bd.jpg", + "image_caption": [ + "Figure 7. Simultaneous 3D Generation and Captioning. Given a single image or text, SAR3D-LLM can generate both a 3D model and a descriptive caption for the model.", + "Figure 8. 3D Object Captioning. Given a 3D model, SAR3D-LLM can generate captions that include both category and details." + ], + "image_footnote": [], + "bbox": [ + 91, + 303, + 486, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a concise interpretation of the 3D data presented here,\" SAR3D-LLM can generate both the correct category and fine details of the input 3D models. For example, in the chair case, SAR3D accurately describes the shape (curved backrest), colors (blue and white), and components (black base, cushioned seat), whereas the ground truth text lacks these details. Furthermore, our 3D tokens enable the LLM to capture the spatial relationships between different parts of the model. For instance, in the third column of Fig. 8, SAR3D uses phrases like \"leading up to\" and \"in front of\" to describe the spatial relationship between the staircase, entrance, and patio area, while the ground truth label merely lists these parts without capturing their spatial connections. Simultaneous 3D Generation and Captioning. In addition to interpreting tokens encoded by our 3D VQVAE, SAR3D can also process 3D tokens generated by our autoregressive model to enable simultaneous 3D generation and captioning, as illustrated in Fig. 7. Given the condition input image or text, SAR3D not only generates the entire object but also detailed captions based on truncated scales of the generated 3D tokens. Notably, in text-conditioned generation and understanding, SAR3D generates additional details beyond those specified in the input text, resulting in accurate and comprehensive descriptions of the generated", + "bbox": [ + 89, + 537, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "content.", + "bbox": [ + 513, + 309, + 571, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 513, + 342, + 635, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The first limitation is that, while SAR3D can generate high-quality 3D objects and detailed interpretations, it currently relies on two separate autoregressive models. Future work could focus on developing a truly multimodal model [78] capable of processing tokens that integrate both text and 3D information, producing both 3D and text outputs. Besides, the quality of the geometry and texture is limited by volume rendering. Using more efficient 3D representations [26] or cascaded generation [90] will further boost the overall quality. Finally, our method's scalability remains unverified due to resource limits. However, 2D scaling laws in VAR [68] suggest its feasibility. We believe that with more resources, our method has the potential to demonstrate favorable scaling laws in 3D generation and understanding.", + "bbox": [ + 511, + 369, + 906, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 601, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented SAR3D, a novel framework that advances both fast 3D object generation and comprehensive 3D understanding through a multi-scale VQVAE and autoregressive modeling. By introducing a latent tri-plane next-scale prediction approach, we addressed the speed limitations of existing diffusion-based 3D generation methods, achieving sub-second generation times with high-quality results. Furthermore, our multi-scale VQVAE enables a pretrained LLM to process and interpret multimodal inputs by leveraging truncated scale 3D tokens, demonstrating the capability of LLMs for detailed 3D object captioning as well as simultaneous 3D generation and captioning. Experimental results underscore SAR3D's efficiency and effectiveness in 3D generation and understanding tasks, positioning it as a versatile tool for multimodal AI applications. Future research may further explore scalability and extend SAR3D's application to broader 3D content and multimodal understanding challenges.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28378", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award . This research is also supported by NTU SUG-NAP and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 236, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35: 23716-23736, 2022. 1, 3", + "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations (ICLR), 2018. 7", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1:8, 2024. 4", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1, 2, 3, 4", + "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2", + "[7] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11315-11325, 2022. 2, 4", + "[8] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), pages 338-355. Springer, 2024. 7", + "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2", + "[10] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets" + ], + "bbox": [ + 93, + 261, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "creation using spatially-aware diffusion guidance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 128-146. Springer, 2024. 2", + "[11] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In International Conference on Learning Representations (ICLR), 2025. 1, 2", + "[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3, 5, 6", + "[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1", + "[14] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems (NeurIPS), 36:35799-35813, 2023. 1, 2", + "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 5", + "[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. In International Conference on Machine Learning (ICML), pages 8469-8488. PMLR, 2023. 1", + "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2", + "[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 1", + "[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 6, 7", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7" + ], + "bbox": [ + 514, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28379", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 5", + "[23] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3DTopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024.6,7", + "[24] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems (NeurIPS), 36: 20482-20494, 2023. 1, 3", + "[25] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1, 2, 5, 6", + "[26] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2024. 8", + "[27] Heewoo Jun and Alex Nichol. Shap-E: Generating conditional 3D implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 7", + "[28] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. 4", + "[29] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, 2021. 7", + "[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. arXiv, 2013. 2, 4", + "[31] Adam R Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Sona Mokrá, and Danilo Jimenez Rezende. Nerf-vae: A geometry aware 3d scene generative model. In International Conference on Machine Learning (ICML), pages 5742-5752. PMLR, 2021. 2", + "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 112-130. Springer, 2024. 1, 2, 4, 7", + "[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussian anything: Interactive point cloud latent diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2025. 2", + "[34] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. CraftsMan: High-fidelity mesh generation with 3D native generation and interactive geometry refiner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2", + "[36] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 2", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems (NeurIPS), 36:34892-34916, 2023. 3", + "[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems (NeurIPS), 36:22226-22246, 2023. 7", + "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2", + "[40] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2", + "[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems (NeurIPS), 36:75307-75337, 2023. 5", + "[42] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 180-197. Springer, 2024. 5", + "[43] Tengchao Lv, Yupan Huang, Jingye Chen, Yuzhong Zhao, Yilin Jia, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 3", + "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 4", + "[45] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, 2022. 2", + "[46] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 7", + "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28380", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, pages 1-31, 2024. 5", + "[48] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (NeurIPS), 35:27730-27744, 2022. 1", + "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 4, 5", + "[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 1, 2", + "[51] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaiheng Ma. Shapeim: Universal 3d object understanding for embodied interaction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 214-238. Springer, 2024. 3", + "[52] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 5", + "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PmLR, 2021. 5", + "[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831. Pmlr, 2021. 2", + "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5", + "[56] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations (ICLR), 2017. 2", + "[57] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 4", + "[58] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Su. Zero123++: a single image to consistent multi-view diffusion base model. In arXiv, 2023. 2", + "[59] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2, 4", + "[60] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2", + "[61] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems (NeurIPS), 34:19313-19325, 2021. 4", + "[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2", + "[63] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6", + "[64] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations (ICLR), 2024. 1", + "[65] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1, 6", + "[66] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2", + "[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2024. 2, 4, 5, 7", + "[68] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in Neural Information Processing Systems (NeurIPS), 37:84839-84865, 2024. 1, 2, 3, 4, 6, 8", + "[69] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 3, 5", + "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28381", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. \nLlama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 3, 5, 6", + "[71] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems (NeurIPS), 35:10021-10039, 2022. 2", + "[72] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4", + "[73] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4563-4573, 2023. 2", + "[74] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2", + "[75] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems (NeurIPS), 36:8406-8441, 2023. 2", + "[76] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 57-74. Springer, 2024. 7", + "[77] Zhenwei Wang, Tengfei Wang, Zexin He, Gerhard Petrus Hancke, Ziwei Liu, and Rynson WH Lau. Phidias: A generative model for creating 3d content from text, image, and 3d conditions with reference-augmented diffusion. In International Conference on Learning Representations (ICLR), 2025. 2", + "[78] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 3, 5, 8", + "[79] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9065-9075, 2023. 4", + "[80] Shuang Wu, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4", + "[81] Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, James Lucas, Antonio Torralba, Sanja Fidler, and Xiaohui Zeng. Latte3d: Large-scale amortized text-to-enhanced3d synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 305-322. Springer, 2024. 4" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4", + "[83] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 131-147. Springer, 2024. 1, 3, 5, 6", + "[84] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1179-1189, 2023. 3", + "[85] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. 6", + "[86] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. In International Conference on Learning Representations (ICLR), 2024. 1", + "[87] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems (NeurIPS), 37: 128940-128966, 2024. 2", + "[88] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 2, 3", + "[89] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2", + "[90] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 8", + "[91] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4", + "[92] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems (NeurIPS), 35: 30599-30611, 2022. 7" + ], + "bbox": [ + 516, + 90, + 903, + 892 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "28382", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_model.json b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_model.json new file mode 100644 index 0000000000000000000000000000000000000000..15a366ed5de91d7ab599a024a214192873ba02a8 --- /dev/null +++ b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_model.json @@ -0,0 +1,2765 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.13, + 0.851, + 0.175 + ], + "angle": 0, + "content": "SAR3D: Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.203, + 0.834, + 0.224 + ], + "angle": 0, + "content": "Yongwei Chen1 Yushi Lan1 Shangchen Zhou1 Tengfei Wang2 Xingang Pan1" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.228, + 0.671, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.247, + 0.673, + 0.265 + ], + "angle": 0, + "content": "\\(^{2}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.265, + 0.663, + 0.283 + ], + "angle": 0, + "content": "https://cyw-3d.github.io/projects/SAR3D/" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.296, + 0.638, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.297, + 0.758, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.785, + 0.303, + 0.882, + 0.332 + ], + "angle": 0, + "content": "Give a concise interpretation of the 3D data presented here." + }, + { + "type": "text", + "bbox": [ + 0.663, + 0.344, + 0.897, + 0.373 + ], + "angle": 0, + "content": "A small, cute, and round yellow Pikachu stuffed toy with a distinctive yellow color, pointy ears, and large black eyes, resembling the iconic Pokémon character." + }, + { + "type": "text", + "bbox": [ + 0.726, + 0.376, + 0.824, + 0.388 + ], + "angle": 0, + "content": "(b) 3D Understanding" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.393, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 1. Our method, SAR3D, proposes a comprehensive framework for 3D generation and understanding via autoregressive modeling. For (a) 3D generation, given a single image or text prompt, SAR3D generates multi-scale 3D objects in an autoregressive manner. For (b) 3D understanding, SAR3D-LLM can interpret a 3D model and provide a detailed description." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.453, + 0.327, + 0.469 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.484, + 0.486, + 0.803 + ], + "angle": 0, + "content": "Autoregressive models have demonstrated remarkable success across various fields, from large language models (LLMs) to large multimodal models (LMMs) and 2D content generation, moving closer to artificial general intelligence (AGI). Despite these advances, applying autoregressive approaches to 3D object generation and understanding remains largely unexplored. This paper introduces Scale AutoRegressive 3D (SAR3D), a novel framework that leverages a multi-scale 3D vector-quantized variational autoencoder (VQVAE) to tokenize 3D objects for efficient autoregressive generation and detailed understanding. By predicting the next scale in a multi-scale latent representation instead of the next single token, SAR3D reduces generation time significantly, achieving fast 3D object generation in just 0.82 seconds on an A6000 GPU. Additionally, given the tokens enriched with hierarchical 3D-aware information, we finetune a pretrained LLM on them, enabling multimodal comprehension of 3D content. Our experiments show that SAR3D surpasses current 3D generation methods in both speed and quality and allows LLMs to interpret and caption 3D models comprehensively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.222, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Autoregressive models have achieved remarkable success in various domains, including large language models (LLMs) [1, 5, 13, 48, 69, 70], 2D generation [63, 68, 86], and large" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.606 + ], + "angle": 0, + "content": "multimodal models (LMMs) [2, 16, 64], marking significant strides toward artificial general intelligence (AGI). By predicting the next token [1] or scale [68], autoregressive models are trained using a simple cross-entropy loss and share similar architectures. This commonality allows them to easily benefit from the optimizations the community has developed over the years for LLMs. Nevertheless, there has been limited exploration of how this next-token/scale prediction approach can be applied to 3D object generation and understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Previously, the scarcity of 3D data pushed researchers to rely on pretrained 2D diffusion models [55] as a prior to generate 3D objects via multi-view score distillation sampling (SDS) loss [50]. Following this, alternative approaches [25, 65] have focused on training feed-forward 3D reconstruction models for fast 3D reconstruction, enabled by large-scale 3D object datasets like Objaverse [14, 15]. These methods are capable of generating 3D assets in mere seconds. More recently, native 3D generative models [27, 32, 46, 90] have emerged, attempting to sample 3D assets from noise under various conditions (e.g., text or image). However, as most of these models rely on diffusion-based methods, they suffer from slow inference times. In parallel, mesh-based generative models [11, 60] attempt to generate 3D topology using autoregressive predictions, but they are limited in detail and require slow, face-by-face predictions. For 3D understanding, some studies [18, 24, 83] have attempted to finetune LLMs on 3D data to interpret the 3D world. However, these methods primarily use 3D point" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "cloud representations, which are limited in capturing fine object details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.482, + 0.258 + ], + "angle": 0, + "content": "In light of the immense potential of autoregressive next-token prediction paradigm and their underexplored status in 3D generation and understanding, we pose an important question: Can autoregressive models be effectively applied to achieve both fast 3D object generation and detailed understanding? Addressing this challenge requires a 3D tokenizer capable of encapsulating detailed information about 3D objects into compact tokens, as well as an efficient schedule for autoregressive prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.44 + ], + "angle": 0, + "content": "In this work, we propose Scale AutoRegressive 3D (SAR3D), a framework that leverages autoregressive models for both fast object generation and comprehensive understanding. Central to SAR3D is a multi-scale 3D vector-quantized variational autoencoder (VQVAE) capable of tokenizing 3D objects into hierarchical levels of tokens. These multi-scale tokens facilitate next-scale prediction training, significantly reducing the steps required for 3D generation compared to diffusion models and traditional next-token prediction methods. Furthermore, the tokens, enriched with 3D-aware information, are naturally compatible with LLM fine-tuning for detailed 3D understanding." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.699 + ], + "angle": 0, + "content": "Specifically, our SAR3D introduces a multi-scale 3D VQVAE to encode multiview RGB images, along with their corresponding depth and camera parameters, into a multiscale latent triplane representation. For 3D generation, we train an autoregressive model to predict the next scale of this latent triplane based on previous scales, conditioned on a single image or text prompt. By predicting the next scale instead of the next single token, our approach significantly reduces generation time, achieving 3D object generation in only 0.82 seconds on an A6000 GPU. For 3D understanding, we use truncated scale tokens from our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to process multimodal inputs that combine text and 3D tokens. Notably, our finetuned LLM can interpret 3D tokens encoded by our VQVAE as well as those generated by our autoregressive model, supporting both 3D captioning and simultaneous generation and understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.482, + 0.76 + ], + "angle": 0, + "content": "Experiments show that SAR3D surpasses existing 3D generation methods in both speed and quality, and our VQVAE enables LLMs to generate detailed captions for 3D objects. Our key technical contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- We introduce SAR3D, a framework designed for both fast 3D object generation and detailed 3D understanding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.797, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- For 3D generation, our method utilizes a next-scale prediction approach for both text-to-3D and single-image-to-3D, achieving faster generation with higher quality compared to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.857, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- For 3D understanding, we leverage truncated scale tokens generated by our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to interpret and describe 3D" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "models, and showcasing the potential of our approach in multimodal applications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.136, + 0.664, + 0.151 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.161, + 0.905, + 0.386 + ], + "angle": 0, + "content": "3D Generative Models. With the success of 2D diffusion models [22, 62], their adaptation for 3D generation has been widely explored. Score distillation sampling [9, 10, 50, 66, 75] leverages these 2D models to distill 3D content, yet it encounters challenges such as costly optimization, mode collapse, and the Janus problem. More recent approaches adopt a two-stage pipeline, generating multiview images first [40, 58, 59, 77] and then reconstructing 3D structures through feed-forward processes [25, 67, 82]. Although promising, these methods are constrained by the quality of multi-view image generation, which often lacks view consistency [39] and fails to scale to higher resolutions [58]. Additionally, this two-stage setup limits 3D editing capabilities due to the absence of a 3D-aware latent space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.905, + 0.614 + ], + "angle": 0, + "content": "To overcome these limitations, native 3D diffusion models [32, 33, 35, 71, 73, 89, 90] have been introduced, offering high-quality, efficient, and scalable 3D generation. Native 3D diffusion pipelines use a two-stage training process: first encoding 3D objects into a VAE latent space [30, 31], followed by applying a latent diffusion model on the resulting codes. However, diffusion-based 3D generation is slow during inference, and its latent space cannot be easily renovated for 3D understanding. In parallel, mesh generative models [11, 60] generates 3D topology through autoregressive prediction. However, they lack details and require slow per-face prediction. In this study, we show that our autoregressive SAR3D achieves efficient sampling, better quality, and can naturally be used for 3D understanding by cascading a Large Language Model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Autoregressive Visual Generation. Pioneered by PixelCNN [56], researchers have proposed to generate images as pixel sequences. Early research VQVAE [72] and VQGAN [17] further quantize image patches into discrete tokens and employ a transformer to learn the autoregressive priors, similar to language modeling [1]. The following research further improves its sampling speed [7] and tokenization efficiency [87]. To further improve the reconstruction quality, RQVAE [34] proposed multi-scale quantization and VAR [68] transforms it into next-scale prediction and significantly enhances the sampling speed. Parallel efforts are spent on scaling up autoregressive models on text-conditioned visual generation task [36, 54, 63, 74]. In 3D area, though some preliminary works [45, 88] studied 3D autoregressive modeling on toy dataset [6] without textures, the research on autoregressive 3D generation over large scale 3D dataset [14, 15] is missing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Large Multimodal Models. Inspired by the great success of large language models (LLMs) [5, 69, 70], large" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28372" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "multimodal models (LMMs) are proposed to comprehend and generate a wide range of information beyond text-based data. Two prominent paradigms exist to train the models in an end-to-end strategy: training the model from scratch [43] or aligning pre-trained LLMs and unimodal encoders [2, 37]. The second strategy typically involves a two-stage process: alignment of the unimodal encoder with LLM's feature space, and instruction-based fine-tuning. Following up works also extend the LMMs to 3D understanding, specifically on point cloud [24, 51, 83, 84]. However, point clouds significantly ignore the details of the given 3D inputs. Here, we demonstrate that our 3D VQVAE can connect with an LLM for a detailed 3D understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.231, + 0.319 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.329, + 0.475, + 0.345 + ], + "angle": 0, + "content": "3.1. Multi-scale Visual Autoregressive Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.485, + 0.489 + ], + "angle": 0, + "content": "VAR [68] presents a multi-scale visual modeling method for image generation, shifting from \"next-token prediction\" to \"next-scale prediction\", which significantly improves the inference speed of autoregressive models. Given an encoded feature map \\( f \\in \\mathbb{R}^{h \\times w \\times C} \\) of an input image \\( I \\), VAR quantizes \\( f \\) into \\( K \\) multi-scale token maps \\( R = (r_1, r_2, \\dots, r_K) \\) at an increasingly higher resolution \\( h_k \\times w_k \\) with \\( r_K \\) matches the resolution of the input feature map \\( f \\). The autoregressive likelihood reads as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.502, + 0.484, + 0.544 + ], + "angle": 0, + "content": "\\[\np \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.648 + ], + "angle": 0, + "content": "where each autoregressive unit \\(r_k \\in [V]^{h_k \\times w_k}\\) is the token map at scale \\(k\\), and the sequence \\((r_1, r_2, \\ldots, r_{k-1})\\) serves as the prefix for \\(r_k\\). Tockenize the input image \\(I\\) to multi-scale discrete token maps \\(R\\) for the learning of next-scale prediction, VAR proposes a multi-scale VQVAE with a multi-scale quantizer \\(\\mathcal{Q}(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.661, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\nf = \\mathcal {E} (I), \\quad R = \\mathcal {Q} (f), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.751 + ], + "angle": 0, + "content": "where \\(I\\) denotes the raw image and \\(\\mathcal{E}\\) is the image encoder. This quantization process will map \\(f\\) to a sequence of multiscale token maps by looking up the nearest code [72] in codebook \\(Z\\in \\mathbb{R}^{V\\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.761, + 0.483, + 0.806 + ], + "angle": 0, + "content": "\\[\nz _ {k} ^ {(i, j)} = \\left(\\arg \\min _ {v \\in [ V ]} \\left\\| \\operatorname {l o o k u p} (Z, v) - r _ {k} ^ {(i, j)} \\right\\| _ {2}\\right) \\in [ V ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.484, + 0.883 + ], + "angle": 0, + "content": "where \\(\\mathrm{lookup}(Z,v)\\) means taking the \\(v\\)-th vector in codebook \\(Z\\). To train the quantized autoencoder, \\(Z\\) is looked up by every \\(z_{k}(i,j)\\) to get \\(\\hat{f}\\), the approximation of original \\(f\\). Then a new image \\(\\hat{I}\\) is reconstructed using the decoder \\(D(\\cdot)\\) given \\(\\hat{f}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {f} = \\operatorname {l o o k u p} (Z, z), \\quad \\hat {I} = \\mathcal {D} (\\hat {f}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "Once fully trained, the autoencoder \\(\\{\\mathcal{E},\\mathcal{Q},\\mathcal{D}\\}\\) will tokenize the incoming images for training the unidirectional autoregressive model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.144, + 0.875, + 0.16 + ], + "angle": 0, + "content": "3.2. PointLLM for Point Cloud Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.367 + ], + "angle": 0, + "content": "Given multi-modal sentences containing both point clouds \\( P \\in \\mathbb{R}^{n \\times d} \\) and text, where \\( n \\) is the number of points and \\( d \\) is the dimension of each point, PointLLM [83] aims to perform 3D point cloud understanding by finetuning a pretrained large language model [12, 70]. It consists of three main components: a pretrained point cloud encoder \\( \\Gamma_{\\mathrm{pe}} \\) (e.g., Point-BERT [88]), a projector \\( \\Gamma_{\\mathrm{proj}} \\) and a pretrained large language model backbone \\( \\Gamma_{\\mathrm{llm}} \\). The \\( \\Gamma_{\\mathrm{pe}} \\) and \\( \\Gamma_{\\mathrm{proj}} \\) project \\( P \\) to a point cloud token sequence \\( Z_{p} \\in \\mathbb{R}^{m \\times c'} \\), where \\( m \\) is the total number of tokens and \\( c' \\) is the projected dimension of the point tokens. The final mixed sequence of tokens \\( Z_{m} = (z_{1}, z_{2}, \\dots, z_{l}) \\in \\mathbb{R}^{l \\times c} \\) consists of both point tokens \\( Z_{p} \\) and text tokens \\( Z_{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.373, + 0.907, + 0.39 + ], + "angle": 0, + "content": "\\[\nZ _ {p} = \\Gamma_ {\\mathrm {p r o j}} \\left(\\Gamma_ {\\mathrm {p e}} (P)\\right), \\quad Z _ {m} = \\operatorname {C o n c a t} \\left(Z _ {p}, Z _ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.906, + 0.472 + ], + "angle": 0, + "content": "where \\(Z_{t}\\) is obtained by tokenizer of \\(\\Gamma_{\\mathrm{llm}}\\) and \\(\\mathrm{Concat}(\\cdot)\\) means concatenation of two vectors. The LLM backbone \\(\\Gamma_{\\mathrm{llm}}\\) is a GPT-style Transformers [5], which accepts a sequence of previous multi-modal tokens \\(Z_{< i} = (z_1, \\ldots, z_{i-1})\\) and predict the next token:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.479, + 0.905, + 0.495 + ], + "angle": 0, + "content": "\\[\nz _ {i} = \\Gamma_ {\\mathrm {l l m}} \\left(Z _ {< i}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.564 + ], + "angle": 0, + "content": "The finetuning process has two stages. In the first stage, it freezes \\(\\{\\Gamma_{\\mathrm{pe}},\\Gamma_{\\mathrm{llm}}\\}\\) and finetunes \\(\\Gamma_{\\mathrm{proj}}\\) to align point features with the text token space. In the second stage, it freezes \\(\\Gamma_{\\mathrm{pe}}\\) and finetunes \\(\\{\\Gamma_{\\mathrm{llm}},\\Gamma_{\\mathrm{proj}}\\}\\) together." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.605, + 0.589 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we present SAR3D for high-quality 3D object generation and detailed understanding. First, we introduce a multi-scale 3D vector-quantized variational autoencoder (VQVAE) in Sec. 4.1, which tokenizes the input 3D model into multi-scale tokens. Fig. 2 illustrates the design of our 3D VQVAE. Next, in Sec. 4.2, with the whole sequence of different scales of the feature tokens, we train an autoregressive model to perform the next scale prediction given a single image or text prompt, which is only supervised by simple cross-entropy loss. Finally, in Sec. 4.3, we explore using truncated scales of the whole sequence to fine-tune a pretrained LLM [12, 69, 70] to handle multimodal input sequence containing both 3D and text tokens, thereby enabling the understanding of the input 3D model. Fig. 3 illustrates our 3D generation and understanding pipeline. Different from other methods [78] that train different encoders for generation and understanding, we train a single VQVAE and use the whole and truncated sequence for generation and understanding, respectively. Details of our SAR3D are shown below." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28373" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.255, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Overview of Multi-scale VQVAE. Given a 3D model, we leverage multi-view RGB-D(depth) renderings and Plücker embeddings as the input to our multi-view encoder \\(\\mathcal{E}\\). The encoder predicts a continuous feature map that is then quantized by the multi-scale quantizer \\(\\mathcal{Q}\\), giving \\(R = (r_1, r_2, \\ldots, r_K)\\) of latent tri-plane features. Each code of different scales share the same codebook. The triplane decoder then converts the quantized latent triplane features into the triplane representation through a plane-wise manner. The predicted triplane is multi-view supervised with the ground truth image, depth, and normal." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.336, + 0.308, + 0.353 + ], + "angle": 0, + "content": "4.1. Multi-scale 3D VQVAE" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.483, + 0.51 + ], + "angle": 0, + "content": "As demonstrated by previous studies [4, 7, 32, 55], the key to high-quality visual generation lies in a compact latent space achieved by a specially designed variational autoencoder [30, 72]. To achieve both fast 3D generation and detailed understanding, we propose a multi-scale 3D VQVAE that maps a given 3D object into a discrete multi-scale latent space. To encode a 3D model, we leverage its multi-view posed RGB-D renderings as input. This approach offers a comprehensive representation of the 3D structure and enables compatibility with existing architectures [79]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.483, + 0.736 + ], + "angle": 0, + "content": "Specifically, the input of VQVAE is a set of multi-view renderings of the 3D object from 6 views. Each rendering \\( M = (I, \\Delta, \\pi) \\) captures essential 3D attributes, representing the object from a specific viewpoint: the RGB image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), the depth map \\( \\Delta \\in \\mathbb{R}^{H \\times W} \\), and the corresponding camera pose \\( \\pi \\). To standardize these 3D attributes, we transform the camera pose \\( \\pi \\) into Plücker coordinates [61], expressed as \\( \\mathbf{p}_i = (\\mathbf{o} \\times \\mathbf{d}_{u,v}, \\mathbf{d}_{u,v}) \\in \\mathbb{R}^6 \\), where \\( \\mathbf{o}_i \\in \\mathbb{R}^3 \\) denotes the camera origin, \\( \\mathbf{d}_{u,v} \\in \\mathbb{R}^3 \\) is the normalized ray direction, and \\( \\times \\) represents the cross product. Consequently, the Plücker embedding of the camera \\( \\pi \\) is represented as \\( \\mathbf{P} \\in \\mathbb{R}^{H \\times W \\times 6} \\). The final representation is formed by channel-wise concatenating these elements, resulting in \\( M = [I \\oplus \\Delta \\oplus \\mathbf{P}] \\in \\mathbb{R}^{H \\times W \\times 10} \\), where \\( \\oplus \\) denotes concatenation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.737, + 0.483, + 0.875 + ], + "angle": 0, + "content": "To maintain both geometry and texture details of \\(M\\), similar to LN3Diff [32], we encode the inputs through a multi-view convolutional encoder [59, 67]. For better 3D awareness, the latent space is designed to be a latent triplane [32, 80] \\(f \\in \\mathbb{R}^{3 \\times h \\times w \\times C}\\). Besides, this representation also has spatial inductive bias and is compatible with the scale and interpolation design in VAR [68]. After encoding, \\(f\\) is interpolated to different scales and quantized using latent triplane quantization layer \\(\\mathcal{Q}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nf = \\mathcal {E} (\\tilde {M}), \\quad R = \\mathcal {Q} (f), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.905, + 0.428 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is the encoder of our VQVAE and \\(R = (r_1, r_2, \\dots, r_K)\\) is the scale sequence and \\(r_k \\in \\mathbb{R}^{3 \\times h_k \\times w_k \\times C}\\), where each sub latent plane \\(r_k^i \\in \\mathbb{R}^{h_k \\times w_k \\times C}\\) is independently quantized and interpolated over the shared codebook \\(Z\\). Please refer to the Supp Mat for more quantization and interpolation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.905, + 0.504 + ], + "angle": 0, + "content": "Afterward, the decoder \\(\\mathcal{D}\\) decodes discrete scales \\(R\\) to triplane and render multiple views to calculate the reconstruction loss. To balance training stability and mesh extraction quality, we first train our model on volume rendering [44] with the loss reads as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.515, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {r e n d e r}} \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {\\mathrm {V Q}} \\mathcal {L} _ {\\mathrm {V Q}} + \\lambda_ {\\mathrm {G A N}} \\mathcal {L} _ {\\mathrm {G A N}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.906, + 0.633 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{render}}\\) combines mean absolute error (MAE) and perceptual loss [91] between the rendered RGB-D images with masks and ground truth, \\(\\mathcal{L}_{\\mathrm{VQ}}\\) includes both encoding error and commitment loss [72], and \\(\\mathcal{L}_{\\mathrm{GAN}}\\) serves as the adversarial loss to encourage a perceptually rich latent space. \\(\\lambda_{*}\\) are the corresponding loss weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.906, + 0.679 + ], + "angle": 0, + "content": "To facilitate 3D mesh extraction, we further finetune the model to the hybrid representation Flexicubes [57, 82] with the extra \\(\\mathcal{L}_{\\mathrm{flex}}\\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.69, + 0.905, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f l e x}} = \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.794 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{normal}}\\) is MAE loss between rendered normal and ground truth, \\(\\mathcal{L}_{\\mathrm{reg}}\\) is regularization term for Flexicubes parameters [57]. \\(\\lambda_{*}\\) are the corresponding loss weights. Similar to LATTE3D [81], we only fine-tune the decoder of the VQVAE in this stage for stabilized training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.905, + 0.833 + ], + "angle": 0, + "content": "4.2. 3D Generation via Multi-scale Autoregressive Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "SAR3D Transformer. We illustrate our generation framework in Fig. 3. Similar to VAR [68], we use standard GPT-style transformer [5] with AdaLN layer [49], with specific layer design following the simple rule of scaling law [28]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28374" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.089, + 0.384, + 0.103 + ], + "angle": 0, + "content": "(a) 3D Generation" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.106, + 0.905, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.089, + 0.788, + 0.103 + ], + "angle": 0, + "content": "(b) 3D Understanding" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Overview of 3D Generation and 3D Understanding. Given a 3D model, our 3D VQVAE encodes it into multi-scale discrete tokens for both 3D generation and understanding. In (a) 3D Generation, text or a single image is encoded by \\(\\mathrm{CLIP_T}\\) or DINOv2, and the encoded condition features are integrated into the decoder-only transformer via cross attention. The transformer then causally predicts each scale of the latent triplane. In (b) 3D Understanding, truncated 3D tokens are first processed with an MLP projector. The large language model receives a multimodal sequence of text and 3D tokens and generates a detailed caption describing the input 3D model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.482, + 0.42 + ], + "angle": 0, + "content": "We adopt tri-plane latent for autoregressive prediction with Eq. 1, where different latent plane \\( r_k^i \\) is differentiated with the corresponding learnable positional embeddings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Conditional 3D Generation. Unlike feed-forward 3D reconstruction models [25, 67] that map input image into 3D, we achieve flexible multimodal 3D generation by introducing diverse conditions, as shown in Fig. 3. For text conditions, we use \\(\\mathrm{CLIP_T}\\) [53] ViT-L text encoder and inject the text embeddings into the autoregressive models through cross attention. For the image-conditioned model, we use DINOv2 [47] ViT-L to extract local patch features and send them into the autoregressive model through pre-cross-attention block [25], which empirically yields better performance. Besides local patch features, we also leverage the pooled out feature of \\(\\mathrm{CLIP_T} / \\mathrm{DINOv2}\\) as the start token of the sequence. Please refer to the Supp Mat for more details of our transformer blocks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.635, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Classifier-free Guidance. As first proposed in diffusion models [22, 55], classifier-free guidance [21] (CFG) has shown effectiveness for improving generation quality and condition alignment. Therefore, we also enable CFG in our model by randomly dropping out \\(10\\%\\) of the input condition by replacing it with a null unconditional embedding [49]. During inference, the logit \\(r_g\\) of each token is calculated by \\(r_g = r_u + s(r_c - r_u)\\), given the conditional logit \\(r_c\\) and the unconditional logit \\(r_u\\). \\(s\\) stands for the scale of the classifier-free guidance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.801, + 0.462, + 0.817 + ], + "angle": 0, + "content": "4.3. SAR3D-LLM for 3D Object Understanding" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Since our 3D VQVAE model provides a comprehensive encoding of the given 3D object, it can be naturally extended to 3D object understanding. Following PointLLM [83], we fine-tune LLaMA [12, 69, 70] in two stages: first, we fine-tune the projector \\(\\Gamma_{\\mathrm{proj}}\\) while freezing the LLM \\(\\Gamma_{\\mathrm{llm}}\\) to align" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.908, + 0.557 + ], + "angle": 0, + "content": "3D features with the text token space; second, we fine-tune both. As briefed in Sec. 3.2, the encoded 3D tokens \\( R \\) are projected to the language latent space, and concatenated with the text instruction tokens \\( Z_{t} \\). Here, we directly use the output tokens from pre-trained SAR3D VQVAE to the projector \\( \\Gamma_{\\mathrm{proj}} \\). Since we only study the 3D captioning [41] task here, the instruction tokens \\( Z_{t} \\) is fixed to \\( \\tilde{Z}_{t} \\) which are tokenized from \"Give a concise interpretation of the 3D data presented here\". The final framework, SAR3D-LLM, supports both detailed 3D captioning given 3D object and simultaneous 3D generation and captioning given text or image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Moreover, a surprising observation here is that not all scales in \\( R \\) are required for 3D understanding training. Empirically, we use the truncated scale latent codes \\( \\tilde{R} = (r_1, r_2, \\dots, r_{K-2}) \\) as the input to the LLM, which contains only \\( 37.5\\% \\) of the overall tokens required for training 3D generation. The final multimodal tokens \\( Z_m \\) that serve as the input to the LLM reads as" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.671, + 0.905, + 0.69 + ], + "angle": 0, + "content": "\\[\nZ _ {p r o j} = \\Gamma_ {\\mathrm {p r o j}} (\\tilde {R}), \\quad Z _ {m} = \\operatorname {C o n c a t} \\left(Z _ {p r o j}, \\tilde {Z} _ {t}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.907, + 0.834 + ], + "angle": 0, + "content": "where \\( Z_{proj} \\) is the projected 3D tokens, and \\( \\mathrm{Concat}(\\cdot, \\cdot) \\) means concatenation. A similar observation is also mentioned in Janus [78], where different features are required for multimodal understanding and generation. Furthermore, unlike other 3D captioning approaches such as Cap3D [41, 42], which separately extracts captions from 8 multi-view renderings and require post-processing to merge them into a unified caption, our method efficiently generates a detailed caption with a single encoding step." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.646, + 0.862 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Datasets. To train our model, we use renderings from G-Objaverse [15, 52] and select a high-quality subset of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28375" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.082, + 0.907, + 0.152 + ], + "angle": 0, + "content": "Table 1. Quantitative Evaluation of Image-conditioned 3D Generation. We evaluate the quality of both 2D rendering and 3D shapes. As shown below, the proposed method demonstrates strong performance across all metrics. Although LGM, a multi-view images-to-3D approach, achieves slightly better performance on FID, it falls short on more advanced image quality assessment metrics such as MUSIQ and has significantly worse 3D shape quality. For multi-view to 3D methods, we also include the number of input views (V=\\#). The latency time is all profiled on Tesla V100 architecture." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.163, + 0.862, + 0.35 + ], + "angle": 0, + "content": "
MethodFID↓KID(%)↓MUSIQ↑COV(%)↑MMD(‰)↓Latency-V100 (s) ↓
Splatter-Image48.803.6530.3337.6630.690.83
OpenLRM38.411.8745.4639.3329.087.21
One-2-3-45 (V=12)88.396.3459.0233.3335.0959.23
Lara (V=4)43.741.9539.3739.3328.8411.93
CRM (V=6)45.531.9364.1038.8328.9122.10
LGM (V=4)19.930.5554.7850.8322.063.87
Shap-E138.5311.9531.5161.3319.179.54
LN3Diff29.080.8950.3955.1719.947.51
SAR3D-NeRF22.550.4265.7774.1713.631.64
SAR3D-Flexicubes27.300.6367.2471.5015.252.92
" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.359, + 0.866, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.595, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison of Image-conditioned 3D Generation. Here, we compare with the state-of-the-art 3D generative models under different categories. As visualized here, our method achieves superior 3D consistency across views and generates intact objects without distortion. For more comparisons with other methods, please refer to the Supp Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.746 + ], + "angle": 0, + "content": "around \\(176K\\) 3D instances, where each consists of 40 random views with RGB, normal, depth maps and camera poses. For text-conditioned generation and 3D understanding training, we use captions provided by 3DTopia [23]. For image-conditioned training, we select a random view of the corresponding 3D instance as the condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Implementation Details. In our multi-scale VQVAE, we use images with a resolution of \\( H = W = 256 \\) as input. The feature map is quantized across 10 scales, with sizes of \\( 3 \\times (1^2, 2^2, 3^2, 4^2, 5^2, 6^2, 8^2, 10^2, 13^2, 16^2) \\). To enhance codebook utilization and stabilize 3D generation training, we follow [63, 85] by applying \\( \\ell_2 \\)-normalization to codebook vectors, setting a low codebook vector dimension \\( C = 8 \\), and using a large codebook size \\( V = 16384 \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For 3D generation, we base our architecture on VAR [68], adding plane positional encoding for each plane." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.822 + ], + "angle": 0, + "content": "For text-conditioned generation, the model has 16 transformer blocks with 16 heads, while for image-conditioned generation, it has 24 transformer blocks with 16 heads. We use the AdamW optimizer with a learning rate of \\(10^{-4}\\). For 3D understanding, we utilize the Vicuna-7B [12] checkpoint of LLaMA [70], following PointLLM [83]. The training was conducted on 7 NVIDIA A100 GPUs for the multiscale VQVAE with batch size 28, image-conditioned transformer with batch 63, text-conditioned transformer with batch size 52. For SAR3D-LLM, the stage-1 alignment is trained with batch size 140, and stage-2 with 112." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.696, + 0.848 + ], + "angle": 0, + "content": "5.1. Single Image to 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compare our SAR3D with three categories of methods: single-image to 3D methods (Splatter-Image [65], OpenLRM [19, 25]), multi-view image to 3D methods" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28376" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.498, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.087, + 0.906, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.256, + 0.746, + 0.272 + ], + "angle": 0, + "content": "Figure 5. More results of image and text conditioned 3D generation of SAR3D." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.284, + 0.9, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.446, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Figure 6. Comparison of Text-conditioned 3D Generation. We present text-conditioned 3D objects generated by SAR3D, displaying two views of each sample. Compared to baseline methods, our approach consistently yields better quality regarding geometry, texture, and text-3D alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.514, + 0.483, + 0.634 + ], + "angle": 0, + "content": "(One-2-3-45 [38], Lara [8], CRM [76], LGM [67]), and native 3D diffusion models (Shap-E [27], LN3Diff-image [32]). Quantitatively, we benchmark rendering metrics with FID [20], KID [3], and MUSIQ [29, 92]. For 3D quality evaluation, we report Coverage Score (COV), and Minimum Matching Distance (MMD) score, as shown in Table 1. Our SAR3D demonstrates strong performance across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.64, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Furthermore, we also profile the generation speed. This timing covers the complete process, from input image processing to mesh extraction. Thanks to efficient next-scale prediction, SAR3D achieves exceptionally fast generation speeds, achieving 0.82 seconds and 1.46 seconds respectively on a single A6000 GPU. Since other baseline methods are tested on Tesla V100 GPUs, we scale our results by a factor of 2 for fair comparison in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The qualitative comparisons between SAR3D and existing methods is also included in Fig. 4. Compared to single-image to 3D methods like OpenLRM[19], and multi-view image to 3D methods like LGM [67], our approach achieves better 3D consistency across views and reduces distortion in generated 3D objects. Compared to native 3D diffusion models like LN3Diff [32], SAR3D produces more complete 3D models. Additional qualitative results are shown in Fig. 5. For more comparisons with other methods and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.866, + 0.53 + ], + "angle": 0, + "content": "details for training data, please refer to the Supp Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.54, + 0.631, + 0.555 + ], + "angle": 0, + "content": "5.2. Text to 3D" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.564, + 0.907, + 0.82 + ], + "angle": 0, + "content": "In addition to image-to-3D generation, SAR3D also supports the creation of high-quality 3D assets from text prompts. As shown in Fig. 5, SAR3D generates diverse and detailed 3D objects based on the same text input. For instance, in the first and second samples, SAR3D produces different shapes for the cannon barrel and chair base, while in the third sample, it varies the texture of a wooden chest. In Fig. 6, we compare our method with other text-to-3D generation approaches, including Point-E [46], ShapE [27], 3DTopia [23], and LN3Diff [32]. Compared to these baselines, SAR3D achieves sharper visual results and better alignment with the input prompts. For example, in the second sample, SAR3D generates red patterns on the handle, closely matching the input text.. In contrast, Point-E [46] reverses the colors of the handle and blade, 3DTopia [23] produces a completely red sword, and Shap-E [27] yields a less detailed result." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.665, + 0.849 + ], + "angle": 0, + "content": "5.3.3D Captioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D Object Captioning. In this section, we present the results of our 3D understanding model applied to various 3D models. As shown in Fig. 8, given the prompt \"Give" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28377" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.098, + 0.121, + 0.243 + ], + "angle": 270, + "content": "6duDpuuDusrAupn uDuaeau 4ndU" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.2, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.089, + 0.334, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.171, + 0.361, + 0.258 + ], + "angle": 0, + "content": "A blue, curved sword with a long and slender blade, featuring a hilt or handle at the base. It has a metallic sheen and is associated with the protagonist, Link, in The Legend of Zelda video game franchise." + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.481, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.091, + 0.597, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.091, + 0.623, + 0.109 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.091, + 0.746, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.097, + 0.854, + 0.107 + ], + "angle": 0, + "content": "\"Bookshelf\"" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.116, + 0.888, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.119, + 0.68, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.17, + 0.762, + 0.213 + ], + "angle": 0, + "content": "with a red cap, depicted as the spotted" + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.169, + 0.882, + 0.213 + ], + "angle": 0, + "content": "a white shelving unit in the shape of a building." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 7. Simultaneous 3D Generation and Captioning. Given a single image or text, SAR3D-LLM can generate both a 3D model and a descriptive caption for the model." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.304, + 0.487, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Figure 8. 3D Object Captioning. Given a 3D model, SAR3D-LLM can generate captions that include both category and details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.485, + 0.903 + ], + "angle": 0, + "content": "a concise interpretation of the 3D data presented here,\" SAR3D-LLM can generate both the correct category and fine details of the input 3D models. For example, in the chair case, SAR3D accurately describes the shape (curved backrest), colors (blue and white), and components (black base, cushioned seat), whereas the ground truth text lacks these details. Furthermore, our 3D tokens enable the LLM to capture the spatial relationships between different parts of the model. For instance, in the third column of Fig. 8, SAR3D uses phrases like \"leading up to\" and \"in front of\" to describe the spatial relationship between the staircase, entrance, and patio area, while the ground truth label merely lists these parts without capturing their spatial connections. Simultaneous 3D Generation and Captioning. In addition to interpreting tokens encoded by our 3D VQVAE, SAR3D can also process 3D tokens generated by our autoregressive model to enable simultaneous 3D generation and captioning, as illustrated in Fig. 7. Given the condition input image or text, SAR3D not only generates the entire object but also detailed captions based on truncated scales of the generated 3D tokens. Notably, in text-conditioned generation and understanding, SAR3D generates additional details beyond those specified in the input text, resulting in accurate and comprehensive descriptions of the generated" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.31, + 0.573, + 0.322 + ], + "angle": 0, + "content": "content." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.343, + 0.637, + 0.359 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.908, + 0.583 + ], + "angle": 0, + "content": "The first limitation is that, while SAR3D can generate high-quality 3D objects and detailed interpretations, it currently relies on two separate autoregressive models. Future work could focus on developing a truly multimodal model [78] capable of processing tokens that integrate both text and 3D information, producing both 3D and text outputs. Besides, the quality of the geometry and texture is limited by volume rendering. Using more efficient 3D representations [26] or cascaded generation [90] will further boost the overall quality. Finally, our method's scalability remains unverified due to resource limits. However, 2D scaling laws in VAR [68] suggest its feasibility. We believe that with more resources, our method has the potential to demonstrate favorable scaling laws in 3D generation and understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.602, + 0.634, + 0.617 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we presented SAR3D, a novel framework that advances both fast 3D object generation and comprehensive 3D understanding through a multi-scale VQVAE and autoregressive modeling. By introducing a latent tri-plane next-scale prediction approach, we addressed the speed limitations of existing diffusion-based 3D generation methods, achieving sub-second generation times with high-quality results. Furthermore, our multi-scale VQVAE enables a pretrained LLM to process and interpret multimodal inputs by leveraging truncated scale 3D tokens, demonstrating the capability of LLMs for detailed 3D object captioning as well as simultaneous 3D generation and captioning. Experimental results underscore SAR3D's efficiency and effectiveness in 3D generation and understanding tasks, positioning it as a versatile tool for multimodal AI applications. Future research may further explore scalability and extend SAR3D's application to broader 3D content and multimodal understanding challenges." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28378" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award . This research is also supported by NTU SUG-NAP and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.237, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35: 23716-23736, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations (ICLR), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1:8, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[7] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11315-11325, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[8] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), pages 338-355. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "creation using spatially-aware diffusion guidance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 128-146. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[11] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In International Conference on Learning Representations (ICLR), 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[14] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems (NeurIPS), 36:35799-35813, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. In International Conference on Machine Learning (ICML), pages 8469-8488. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28379" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[23] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3DTopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[24] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems (NeurIPS), 36: 20482-20494, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[25] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[26] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[27] Heewoo Jun and Alex Nichol. Shap-E: Generating conditional 3D implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[28] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[29] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. arXiv, 2013. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[31] Adam R Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Sona Mokrá, and Danilo Jimenez Rezende. Nerf-vae: A geometry aware 3d scene generative model. In International Conference on Machine Learning (ICML), pages 5742-5752. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 112-130. Springer, 2024. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussian anything: Interactive point cloud latent diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[35] Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. CraftsMan: High-fidelity mesh generation with 3D native generation and interactive geometry refiner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[36] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems (NeurIPS), 36:34892-34916, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems (NeurIPS), 36:22226-22246, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[40] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems (NeurIPS), 36:75307-75337, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[42] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 180-197. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[43] Tengchao Lv, Yupan Huang, Jingye Chen, Yuzhong Zhao, Yilin Jia, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[45] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[46] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28380" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, pages 1-31, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.231 + ], + "angle": 0, + "content": "[48] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (NeurIPS), 35:27730-27744, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[51] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaiheng Ma. Shapeim: Universal 3d object understanding for embodied interaction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 214-238. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[52] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PmLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831. Pmlr, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[56] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations (ICLR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[57] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[58] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Su. Zero123++: a single image to consistent multi-view diffusion base model. In arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[59] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[60] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[61] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems (NeurIPS), 34:19313-19325, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[63] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[64] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations (ICLR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[65] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[66] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2024. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[68] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in Neural Information Processing Systems (NeurIPS), 37:84839-84865, 2024. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[69] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28381" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. \nLlama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[71] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems (NeurIPS), 35:10021-10039, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[72] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[73] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4563-4573, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[74] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[75] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems (NeurIPS), 36:8406-8441, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[76] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 57-74. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[77] Zhenwei Wang, Tengfei Wang, Zexin He, Gerhard Petrus Hancke, Ziwei Liu, and Rynson WH Lau. Phidias: A generative model for creating 3d content from text, image, and 3d conditions with reference-augmented diffusion. In International Conference on Learning Representations (ICLR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[78] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[79] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9065-9075, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[80] Shuang Wu, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[81] Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, James Lucas, Antonio Torralba, Sanja Fidler, and Xiaohui Zeng. Latte3d: Large-scale amortized text-to-enhanced3d synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 305-322. Springer, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[82] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[83] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 131-147. Springer, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[84] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1179-1189, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[85] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.402, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[86] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. In International Conference on Learning Representations (ICLR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.486, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[87] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems (NeurIPS), 37: 128940-128966, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.556, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[88] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[89] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.905, + 0.753 + ], + "angle": 0, + "content": "[90] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.754, + 0.905, + 0.824 + ], + "angle": 0, + "content": "[91] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.824, + 0.905, + 0.893 + ], + "angle": 0, + "content": "[92] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems (NeurIPS), 35: 30599-30611, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28382" + } + ] +] \ No newline at end of file diff --git a/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_origin.pdf b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7df665afa660370abe2cdaee0a82dc358a0f04c4 --- /dev/null +++ b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/742f6d19-8530-4fec-8c8e-94f82f4b2720_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ac23b9415092c3ac544c25c13bdaa3bf3ee9b036059fd7fc3dfa7d496b898a2 +size 2610896 diff --git a/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/full.md b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fd4bb303fa701afac0b309e7326756212d52e9b3 --- /dev/null +++ b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/full.md @@ -0,0 +1,363 @@ +# SAR3D: Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE + +Yongwei Chen1 Yushi Lan1 Shangchen Zhou1 Tengfei Wang2 Xingang Pan1 + +$^{1}$ S-Lab, Nanyang Technological University + +$^{2}$ Shanghai Artificial Intelligence Laboratory + +https://cyw-3d.github.io/projects/SAR3D/ + +![](images/e496bf99f4c4ac90bc6ba30bbfac6f766fd35add1b2c4aa3d9495d0f1f5dc36f.jpg) +Figure 1. Our method, SAR3D, proposes a comprehensive framework for 3D generation and understanding via autoregressive modeling. For (a) 3D generation, given a single image or text prompt, SAR3D generates multi-scale 3D objects in an autoregressive manner. For (b) 3D understanding, SAR3D-LLM can interpret a 3D model and provide a detailed description. + +![](images/735a1c30784266fb056fd2ffa4c090f00b169c18d99ef3b3569bc09c4991c048.jpg) + +Give a concise interpretation of the 3D data presented here. + +A small, cute, and round yellow Pikachu stuffed toy with a distinctive yellow color, pointy ears, and large black eyes, resembling the iconic Pokémon character. + +(b) 3D Understanding + +# Abstract + +Autoregressive models have demonstrated remarkable success across various fields, from large language models (LLMs) to large multimodal models (LMMs) and 2D content generation, moving closer to artificial general intelligence (AGI). Despite these advances, applying autoregressive approaches to 3D object generation and understanding remains largely unexplored. This paper introduces Scale AutoRegressive 3D (SAR3D), a novel framework that leverages a multi-scale 3D vector-quantized variational autoencoder (VQVAE) to tokenize 3D objects for efficient autoregressive generation and detailed understanding. By predicting the next scale in a multi-scale latent representation instead of the next single token, SAR3D reduces generation time significantly, achieving fast 3D object generation in just 0.82 seconds on an A6000 GPU. Additionally, given the tokens enriched with hierarchical 3D-aware information, we finetune a pretrained LLM on them, enabling multimodal comprehension of 3D content. Our experiments show that SAR3D surpasses current 3D generation methods in both speed and quality and allows LLMs to interpret and caption 3D models comprehensively. + +# 1. Introduction + +Autoregressive models have achieved remarkable success in various domains, including large language models (LLMs) [1, 5, 13, 48, 69, 70], 2D generation [63, 68, 86], and large + +multimodal models (LMMs) [2, 16, 64], marking significant strides toward artificial general intelligence (AGI). By predicting the next token [1] or scale [68], autoregressive models are trained using a simple cross-entropy loss and share similar architectures. This commonality allows them to easily benefit from the optimizations the community has developed over the years for LLMs. Nevertheless, there has been limited exploration of how this next-token/scale prediction approach can be applied to 3D object generation and understanding. + +Previously, the scarcity of 3D data pushed researchers to rely on pretrained 2D diffusion models [55] as a prior to generate 3D objects via multi-view score distillation sampling (SDS) loss [50]. Following this, alternative approaches [25, 65] have focused on training feed-forward 3D reconstruction models for fast 3D reconstruction, enabled by large-scale 3D object datasets like Objaverse [14, 15]. These methods are capable of generating 3D assets in mere seconds. More recently, native 3D generative models [27, 32, 46, 90] have emerged, attempting to sample 3D assets from noise under various conditions (e.g., text or image). However, as most of these models rely on diffusion-based methods, they suffer from slow inference times. In parallel, mesh-based generative models [11, 60] attempt to generate 3D topology using autoregressive predictions, but they are limited in detail and require slow, face-by-face predictions. For 3D understanding, some studies [18, 24, 83] have attempted to finetune LLMs on 3D data to interpret the 3D world. However, these methods primarily use 3D point + +cloud representations, which are limited in capturing fine object details. + +In light of the immense potential of autoregressive next-token prediction paradigm and their underexplored status in 3D generation and understanding, we pose an important question: Can autoregressive models be effectively applied to achieve both fast 3D object generation and detailed understanding? Addressing this challenge requires a 3D tokenizer capable of encapsulating detailed information about 3D objects into compact tokens, as well as an efficient schedule for autoregressive prediction. + +In this work, we propose Scale AutoRegressive 3D (SAR3D), a framework that leverages autoregressive models for both fast object generation and comprehensive understanding. Central to SAR3D is a multi-scale 3D vector-quantized variational autoencoder (VQVAE) capable of tokenizing 3D objects into hierarchical levels of tokens. These multi-scale tokens facilitate next-scale prediction training, significantly reducing the steps required for 3D generation compared to diffusion models and traditional next-token prediction methods. Furthermore, the tokens, enriched with 3D-aware information, are naturally compatible with LLM fine-tuning for detailed 3D understanding. + +Specifically, our SAR3D introduces a multi-scale 3D VQVAE to encode multiview RGB images, along with their corresponding depth and camera parameters, into a multiscale latent triplane representation. For 3D generation, we train an autoregressive model to predict the next scale of this latent triplane based on previous scales, conditioned on a single image or text prompt. By predicting the next scale instead of the next single token, our approach significantly reduces generation time, achieving 3D object generation in only 0.82 seconds on an A6000 GPU. For 3D understanding, we use truncated scale tokens from our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to process multimodal inputs that combine text and 3D tokens. Notably, our finetuned LLM can interpret 3D tokens encoded by our VQVAE as well as those generated by our autoregressive model, supporting both 3D captioning and simultaneous generation and understanding. + +Experiments show that SAR3D surpasses existing 3D generation methods in both speed and quality, and our VQVAE enables LLMs to generate detailed captions for 3D objects. Our key technical contributions are as follows: + +- We introduce SAR3D, a framework designed for both fast 3D object generation and detailed 3D understanding. +- For 3D generation, our method utilizes a next-scale prediction approach for both text-to-3D and single-image-to-3D, achieving faster generation with higher quality compared to existing methods. +- For 3D understanding, we leverage truncated scale tokens generated by our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to interpret and describe 3D + +models, and showcasing the potential of our approach in multimodal applications. + +# 2. Related Works + +3D Generative Models. With the success of 2D diffusion models [22, 62], their adaptation for 3D generation has been widely explored. Score distillation sampling [9, 10, 50, 66, 75] leverages these 2D models to distill 3D content, yet it encounters challenges such as costly optimization, mode collapse, and the Janus problem. More recent approaches adopt a two-stage pipeline, generating multiview images first [40, 58, 59, 77] and then reconstructing 3D structures through feed-forward processes [25, 67, 82]. Although promising, these methods are constrained by the quality of multi-view image generation, which often lacks view consistency [39] and fails to scale to higher resolutions [58]. Additionally, this two-stage setup limits 3D editing capabilities due to the absence of a 3D-aware latent space. + +To overcome these limitations, native 3D diffusion models [32, 33, 35, 71, 73, 89, 90] have been introduced, offering high-quality, efficient, and scalable 3D generation. Native 3D diffusion pipelines use a two-stage training process: first encoding 3D objects into a VAE latent space [30, 31], followed by applying a latent diffusion model on the resulting codes. However, diffusion-based 3D generation is slow during inference, and its latent space cannot be easily renovated for 3D understanding. In parallel, mesh generative models [11, 60] generates 3D topology through autoregressive prediction. However, they lack details and require slow per-face prediction. In this study, we show that our autoregressive SAR3D achieves efficient sampling, better quality, and can naturally be used for 3D understanding by cascading a Large Language Model. + +Autoregressive Visual Generation. Pioneered by PixelCNN [56], researchers have proposed to generate images as pixel sequences. Early research VQVAE [72] and VQGAN [17] further quantize image patches into discrete tokens and employ a transformer to learn the autoregressive priors, similar to language modeling [1]. The following research further improves its sampling speed [7] and tokenization efficiency [87]. To further improve the reconstruction quality, RQVAE [34] proposed multi-scale quantization and VAR [68] transforms it into next-scale prediction and significantly enhances the sampling speed. Parallel efforts are spent on scaling up autoregressive models on text-conditioned visual generation task [36, 54, 63, 74]. In 3D area, though some preliminary works [45, 88] studied 3D autoregressive modeling on toy dataset [6] without textures, the research on autoregressive 3D generation over large scale 3D dataset [14, 15] is missing. + +Large Multimodal Models. Inspired by the great success of large language models (LLMs) [5, 69, 70], large + +multimodal models (LMMs) are proposed to comprehend and generate a wide range of information beyond text-based data. Two prominent paradigms exist to train the models in an end-to-end strategy: training the model from scratch [43] or aligning pre-trained LLMs and unimodal encoders [2, 37]. The second strategy typically involves a two-stage process: alignment of the unimodal encoder with LLM's feature space, and instruction-based fine-tuning. Following up works also extend the LMMs to 3D understanding, specifically on point cloud [24, 51, 83, 84]. However, point clouds significantly ignore the details of the given 3D inputs. Here, we demonstrate that our 3D VQVAE can connect with an LLM for a detailed 3D understanding. + +# 3. Preliminaries + +# 3.1. Multi-scale Visual Autoregressive Generation + +VAR [68] presents a multi-scale visual modeling method for image generation, shifting from "next-token prediction" to "next-scale prediction", which significantly improves the inference speed of autoregressive models. Given an encoded feature map $f \in \mathbb{R}^{h \times w \times C}$ of an input image $I$ , VAR quantizes $f$ into $K$ multi-scale token maps $R = (r_1, r_2, \dots, r_K)$ at an increasingly higher resolution $h_k \times w_k$ with $r_K$ matches the resolution of the input feature map $f$ . The autoregressive likelihood reads as: + +$$ +p \left(r _ {1}, r _ {2}, \dots , r _ {K}\right) = \prod_ {k = 1} ^ {K} p \left(r _ {k} \mid r _ {1}, r _ {2}, \dots , r _ {k - 1}\right), \tag {1} +$$ + +where each autoregressive unit $r_k \in [V]^{h_k \times w_k}$ is the token map at scale $k$ , and the sequence $(r_1, r_2, \ldots, r_{k-1})$ serves as the prefix for $r_k$ . Tockenize the input image $I$ to multi-scale discrete token maps $R$ for the learning of next-scale prediction, VAR proposes a multi-scale VQVAE with a multi-scale quantizer $\mathcal{Q}(\cdot)$ : + +$$ +f = \mathcal {E} (I), \quad R = \mathcal {Q} (f), \tag {2} +$$ + +where $I$ denotes the raw image and $\mathcal{E}$ is the image encoder. This quantization process will map $f$ to a sequence of multiscale token maps by looking up the nearest code [72] in codebook $Z\in \mathbb{R}^{V\times C}$ : + +$$ +z _ {k} ^ {(i, j)} = \left(\arg \min _ {v \in [ V ]} \left\| \operatorname {l o o k u p} (Z, v) - r _ {k} ^ {(i, j)} \right\| _ {2}\right) \in [ V ], \tag {3} +$$ + +where $\mathrm{lookup}(Z,v)$ means taking the $v$ -th vector in codebook $Z$ . To train the quantized autoencoder, $Z$ is looked up by every $z_{k}(i,j)$ to get $\hat{f}$ , the approximation of original $f$ . Then a new image $\hat{I}$ is reconstructed using the decoder $D(\cdot)$ given $\hat{f}$ : + +$$ +\hat {f} = \operatorname {l o o k u p} (Z, z), \quad \hat {I} = \mathcal {D} (\hat {f}). \tag {4} +$$ + +Once fully trained, the autoencoder $\{\mathcal{E},\mathcal{Q},\mathcal{D}\}$ will tokenize the incoming images for training the unidirectional autoregressive model. + +# 3.2. PointLLM for Point Cloud Understanding + +Given multi-modal sentences containing both point clouds $P \in \mathbb{R}^{n \times d}$ and text, where $n$ is the number of points and $d$ is the dimension of each point, PointLLM [83] aims to perform 3D point cloud understanding by finetuning a pretrained large language model [12, 70]. It consists of three main components: a pretrained point cloud encoder $\Gamma_{\mathrm{pe}}$ (e.g., Point-BERT [88]), a projector $\Gamma_{\mathrm{proj}}$ and a pretrained large language model backbone $\Gamma_{\mathrm{llm}}$ . The $\Gamma_{\mathrm{pe}}$ and $\Gamma_{\mathrm{proj}}$ project $P$ to a point cloud token sequence $Z_{p} \in \mathbb{R}^{m \times c'}$ , where $m$ is the total number of tokens and $c'$ is the projected dimension of the point tokens. The final mixed sequence of tokens $Z_{m} = (z_{1}, z_{2}, \dots, z_{l}) \in \mathbb{R}^{l \times c}$ consists of both point tokens $Z_{p}$ and text tokens $Z_{t}$ : + +$$ +Z _ {p} = \Gamma_ {\mathrm {p r o j}} \left(\Gamma_ {\mathrm {p e}} (P)\right), \quad Z _ {m} = \operatorname {C o n c a t} \left(Z _ {p}, Z _ {t}\right), \tag {5} +$$ + +where $Z_{t}$ is obtained by tokenizer of $\Gamma_{\mathrm{llm}}$ and $\mathrm{Concat}(\cdot)$ means concatenation of two vectors. The LLM backbone $\Gamma_{\mathrm{llm}}$ is a GPT-style Transformers [5], which accepts a sequence of previous multi-modal tokens $Z_{< i} = (z_1, \ldots, z_{i-1})$ and predict the next token: + +$$ +z _ {i} = \Gamma_ {\mathrm {l l m}} \left(Z _ {< i}\right). \tag {6} +$$ + +The finetuning process has two stages. In the first stage, it freezes $\{\Gamma_{\mathrm{pe}},\Gamma_{\mathrm{llm}}\}$ and finetunes $\Gamma_{\mathrm{proj}}$ to align point features with the text token space. In the second stage, it freezes $\Gamma_{\mathrm{pe}}$ and finetunes $\{\Gamma_{\mathrm{llm}},\Gamma_{\mathrm{proj}}\}$ together. + +# 4. Method + +In this section, we present SAR3D for high-quality 3D object generation and detailed understanding. First, we introduce a multi-scale 3D vector-quantized variational autoencoder (VQVAE) in Sec. 4.1, which tokenizes the input 3D model into multi-scale tokens. Fig. 2 illustrates the design of our 3D VQVAE. Next, in Sec. 4.2, with the whole sequence of different scales of the feature tokens, we train an autoregressive model to perform the next scale prediction given a single image or text prompt, which is only supervised by simple cross-entropy loss. Finally, in Sec. 4.3, we explore using truncated scales of the whole sequence to fine-tune a pretrained LLM [12, 69, 70] to handle multimodal input sequence containing both 3D and text tokens, thereby enabling the understanding of the input 3D model. Fig. 3 illustrates our 3D generation and understanding pipeline. Different from other methods [78] that train different encoders for generation and understanding, we train a single VQVAE and use the whole and truncated sequence for generation and understanding, respectively. Details of our SAR3D are shown below. + +![](images/4bc020b430e7537d276a1e7d75e15f90258d8e7dde158b0407979bd1b40af577.jpg) +Figure 2. Overview of Multi-scale VQVAE. Given a 3D model, we leverage multi-view RGB-D(depth) renderings and Plücker embeddings as the input to our multi-view encoder $\mathcal{E}$ . The encoder predicts a continuous feature map that is then quantized by the multi-scale quantizer $\mathcal{Q}$ , giving $R = (r_1, r_2, \ldots, r_K)$ of latent tri-plane features. Each code of different scales share the same codebook. The triplane decoder then converts the quantized latent triplane features into the triplane representation through a plane-wise manner. The predicted triplane is multi-view supervised with the ground truth image, depth, and normal. + +# 4.1. Multi-scale 3D VQVAE + +As demonstrated by previous studies [4, 7, 32, 55], the key to high-quality visual generation lies in a compact latent space achieved by a specially designed variational autoencoder [30, 72]. To achieve both fast 3D generation and detailed understanding, we propose a multi-scale 3D VQVAE that maps a given 3D object into a discrete multi-scale latent space. To encode a 3D model, we leverage its multi-view posed RGB-D renderings as input. This approach offers a comprehensive representation of the 3D structure and enables compatibility with existing architectures [79]. + +Specifically, the input of VQVAE is a set of multi-view renderings of the 3D object from 6 views. Each rendering $M = (I, \Delta, \pi)$ captures essential 3D attributes, representing the object from a specific viewpoint: the RGB image $I \in \mathbb{R}^{H \times W \times 3}$ , the depth map $\Delta \in \mathbb{R}^{H \times W}$ , and the corresponding camera pose $\pi$ . To standardize these 3D attributes, we transform the camera pose $\pi$ into Plücker coordinates [61], expressed as $\mathbf{p}_i = (\mathbf{o} \times \mathbf{d}_{u,v}, \mathbf{d}_{u,v}) \in \mathbb{R}^6$ , where $\mathbf{o}_i \in \mathbb{R}^3$ denotes the camera origin, $\mathbf{d}_{u,v} \in \mathbb{R}^3$ is the normalized ray direction, and $\times$ represents the cross product. Consequently, the Plücker embedding of the camera $\pi$ is represented as $\mathbf{P} \in \mathbb{R}^{H \times W \times 6}$ . The final representation is formed by channel-wise concatenating these elements, resulting in $M = [I \oplus \Delta \oplus \mathbf{P}] \in \mathbb{R}^{H \times W \times 10}$ , where $\oplus$ denotes concatenation. + +To maintain both geometry and texture details of $M$ , similar to LN3Diff [32], we encode the inputs through a multi-view convolutional encoder [59, 67]. For better 3D awareness, the latent space is designed to be a latent triplane [32, 80] $f \in \mathbb{R}^{3 \times h \times w \times C}$ . Besides, this representation also has spatial inductive bias and is compatible with the scale and interpolation design in VAR [68]. After encoding, $f$ is interpolated to different scales and quantized using latent triplane quantization layer $\mathcal{Q}$ : + +$$ +f = \mathcal {E} (\tilde {M}), \quad R = \mathcal {Q} (f), \tag {7} +$$ + +where $\mathcal{E}$ is the encoder of our VQVAE and $R = (r_1, r_2, \dots, r_K)$ is the scale sequence and $r_k \in \mathbb{R}^{3 \times h_k \times w_k \times C}$ , where each sub latent plane $r_k^i \in \mathbb{R}^{h_k \times w_k \times C}$ is independently quantized and interpolated over the shared codebook $Z$ . Please refer to the Supp Mat for more quantization and interpolation details. + +Afterward, the decoder $\mathcal{D}$ decodes discrete scales $R$ to triplane and render multiple views to calculate the reconstruction loss. To balance training stability and mesh extraction quality, we first train our model on volume rendering [44] with the loss reads as: + +$$ +\mathcal {L} = \lambda_ {\text {r e n d e r}} \mathcal {L} _ {\text {r e n d e r}} + \lambda_ {\mathrm {V Q}} \mathcal {L} _ {\mathrm {V Q}} + \lambda_ {\mathrm {G A N}} \mathcal {L} _ {\mathrm {G A N}}, \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{render}}$ combines mean absolute error (MAE) and perceptual loss [91] between the rendered RGB-D images with masks and ground truth, $\mathcal{L}_{\mathrm{VQ}}$ includes both encoding error and commitment loss [72], and $\mathcal{L}_{\mathrm{GAN}}$ serves as the adversarial loss to encourage a perceptually rich latent space. $\lambda_{*}$ are the corresponding loss weights. + +To facilitate 3D mesh extraction, we further finetune the model to the hybrid representation Flexicubes [57, 82] with the extra $\mathcal{L}_{\mathrm{flex}}$ loss: + +$$ +\mathcal {L} _ {\text {f l e x}} = \lambda_ {\text {n o r m a l}} \mathcal {L} _ {\text {n o r m a l}} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}}, \tag {9} +$$ + +where $\mathcal{L}_{\mathrm{normal}}$ is MAE loss between rendered normal and ground truth, $\mathcal{L}_{\mathrm{reg}}$ is regularization term for Flexicubes parameters [57]. $\lambda_{*}$ are the corresponding loss weights. Similar to LATTE3D [81], we only fine-tune the decoder of the VQVAE in this stage for stabilized training. + +# 4.2. 3D Generation via Multi-scale Autoregressive Modeling + +SAR3D Transformer. We illustrate our generation framework in Fig. 3. Similar to VAR [68], we use standard GPT-style transformer [5] with AdaLN layer [49], with specific layer design following the simple rule of scaling law [28]. + +![](images/0befbd1d39f5c8894dbd2364518c63db036cdb11e74545011dca48d7d85ae29d.jpg) +(a) 3D Generation +(b) 3D Understanding +Figure 3. Overview of 3D Generation and 3D Understanding. Given a 3D model, our 3D VQVAE encodes it into multi-scale discrete tokens for both 3D generation and understanding. In (a) 3D Generation, text or a single image is encoded by $\mathrm{CLIP_T}$ or DINOv2, and the encoded condition features are integrated into the decoder-only transformer via cross attention. The transformer then causally predicts each scale of the latent triplane. In (b) 3D Understanding, truncated 3D tokens are first processed with an MLP projector. The large language model receives a multimodal sequence of text and 3D tokens and generates a detailed caption describing the input 3D model. + +We adopt tri-plane latent for autoregressive prediction with Eq. 1, where different latent plane $r_k^i$ is differentiated with the corresponding learnable positional embeddings. + +Conditional 3D Generation. Unlike feed-forward 3D reconstruction models [25, 67] that map input image into 3D, we achieve flexible multimodal 3D generation by introducing diverse conditions, as shown in Fig. 3. For text conditions, we use $\mathrm{CLIP_T}$ [53] ViT-L text encoder and inject the text embeddings into the autoregressive models through cross attention. For the image-conditioned model, we use DINOv2 [47] ViT-L to extract local patch features and send them into the autoregressive model through pre-cross-attention block [25], which empirically yields better performance. Besides local patch features, we also leverage the pooled out feature of $\mathrm{CLIP_T} / \mathrm{DINOv2}$ as the start token of the sequence. Please refer to the Supp Mat for more details of our transformer blocks. + +Classifier-free Guidance. As first proposed in diffusion models [22, 55], classifier-free guidance [21] (CFG) has shown effectiveness for improving generation quality and condition alignment. Therefore, we also enable CFG in our model by randomly dropping out $10\%$ of the input condition by replacing it with a null unconditional embedding [49]. During inference, the logit $r_g$ of each token is calculated by $r_g = r_u + s(r_c - r_u)$ , given the conditional logit $r_c$ and the unconditional logit $r_u$ . $s$ stands for the scale of the classifier-free guidance. + +# 4.3. SAR3D-LLM for 3D Object Understanding + +Since our 3D VQVAE model provides a comprehensive encoding of the given 3D object, it can be naturally extended to 3D object understanding. Following PointLLM [83], we fine-tune LLaMA [12, 69, 70] in two stages: first, we fine-tune the projector $\Gamma_{\mathrm{proj}}$ while freezing the LLM $\Gamma_{\mathrm{llm}}$ to align + +3D features with the text token space; second, we fine-tune both. As briefed in Sec. 3.2, the encoded 3D tokens $R$ are projected to the language latent space, and concatenated with the text instruction tokens $Z_{t}$ . Here, we directly use the output tokens from pre-trained SAR3D VQVAE to the projector $\Gamma_{\mathrm{proj}}$ . Since we only study the 3D captioning [41] task here, the instruction tokens $Z_{t}$ is fixed to $\tilde{Z}_{t}$ which are tokenized from "Give a concise interpretation of the 3D data presented here". The final framework, SAR3D-LLM, supports both detailed 3D captioning given 3D object and simultaneous 3D generation and captioning given text or image. + +Moreover, a surprising observation here is that not all scales in $R$ are required for 3D understanding training. Empirically, we use the truncated scale latent codes $\tilde{R} = (r_1, r_2, \dots, r_{K-2})$ as the input to the LLM, which contains only $37.5\%$ of the overall tokens required for training 3D generation. The final multimodal tokens $Z_m$ that serve as the input to the LLM reads as + +$$ +Z _ {p r o j} = \Gamma_ {\mathrm {p r o j}} (\tilde {R}), \quad Z _ {m} = \operatorname {C o n c a t} \left(Z _ {p r o j}, \tilde {Z} _ {t}\right), \tag {10} +$$ + +where $Z_{proj}$ is the projected 3D tokens, and $\mathrm{Concat}(\cdot, \cdot)$ means concatenation. A similar observation is also mentioned in Janus [78], where different features are required for multimodal understanding and generation. Furthermore, unlike other 3D captioning approaches such as Cap3D [41, 42], which separately extracts captions from 8 multi-view renderings and require post-processing to merge them into a unified caption, our method efficiently generates a detailed caption with a single encoding step. + +# 5. Experiments + +Datasets. To train our model, we use renderings from G-Objaverse [15, 52] and select a high-quality subset of + +Table 1. Quantitative Evaluation of Image-conditioned 3D Generation. We evaluate the quality of both 2D rendering and 3D shapes. As shown below, the proposed method demonstrates strong performance across all metrics. Although LGM, a multi-view images-to-3D approach, achieves slightly better performance on FID, it falls short on more advanced image quality assessment metrics such as MUSIQ and has significantly worse 3D shape quality. For multi-view to 3D methods, we also include the number of input views (V=\#). The latency time is all profiled on Tesla V100 architecture. + +
MethodFID↓KID(%)↓MUSIQ↑COV(%)↑MMD(‰)↓Latency-V100 (s) ↓
Splatter-Image48.803.6530.3337.6630.690.83
OpenLRM38.411.8745.4639.3329.087.21
One-2-3-45 (V=12)88.396.3459.0233.3335.0959.23
Lara (V=4)43.741.9539.3739.3328.8411.93
CRM (V=6)45.531.9364.1038.8328.9122.10
LGM (V=4)19.930.5554.7850.8322.063.87
Shap-E138.5311.9531.5161.3319.179.54
LN3Diff29.080.8950.3955.1719.947.51
SAR3D-NeRF22.550.4265.7774.1713.631.64
SAR3D-Flexicubes27.300.6367.2471.5015.252.92
+ +![](images/942591ab501ea1692aa3e4b1923d1410b3835a9456aee65e413099d6399f5052.jpg) +Figure 4. Qualitative Comparison of Image-conditioned 3D Generation. Here, we compare with the state-of-the-art 3D generative models under different categories. As visualized here, our method achieves superior 3D consistency across views and generates intact objects without distortion. For more comparisons with other methods, please refer to the Supp Mat. + +around $176K$ 3D instances, where each consists of 40 random views with RGB, normal, depth maps and camera poses. For text-conditioned generation and 3D understanding training, we use captions provided by 3DTopia [23]. For image-conditioned training, we select a random view of the corresponding 3D instance as the condition. + +Implementation Details. In our multi-scale VQVAE, we use images with a resolution of $H = W = 256$ as input. The feature map is quantized across 10 scales, with sizes of $3 \times (1^2, 2^2, 3^2, 4^2, 5^2, 6^2, 8^2, 10^2, 13^2, 16^2)$ . To enhance codebook utilization and stabilize 3D generation training, we follow [63, 85] by applying $\ell_2$ -normalization to codebook vectors, setting a low codebook vector dimension $C = 8$ , and using a large codebook size $V = 16384$ . + +For 3D generation, we base our architecture on VAR [68], adding plane positional encoding for each plane. + +For text-conditioned generation, the model has 16 transformer blocks with 16 heads, while for image-conditioned generation, it has 24 transformer blocks with 16 heads. We use the AdamW optimizer with a learning rate of $10^{-4}$ . For 3D understanding, we utilize the Vicuna-7B [12] checkpoint of LLaMA [70], following PointLLM [83]. The training was conducted on 7 NVIDIA A100 GPUs for the multiscale VQVAE with batch size 28, image-conditioned transformer with batch 63, text-conditioned transformer with batch size 52. For SAR3D-LLM, the stage-1 alignment is trained with batch size 140, and stage-2 with 112. + +# 5.1. Single Image to 3D + +We compare our SAR3D with three categories of methods: single-image to 3D methods (Splatter-Image [65], OpenLRM [19, 25]), multi-view image to 3D methods + +![](images/394b8052b3e4104138e3fb694e22768b571f8df71e8bb00b9a9e92617ca9f4ce.jpg) +Figure 5. More results of image and text conditioned 3D generation of SAR3D. + +![](images/7732b71c044fa6e9cc4250eddb7f89ae10707348a743c81a330ed7bfce6a1c24.jpg) + +![](images/e3e40b44f446ca99b5b249ffe104bb2ab19425be150473176128d8661b68bade.jpg) +Figure 6. Comparison of Text-conditioned 3D Generation. We present text-conditioned 3D objects generated by SAR3D, displaying two views of each sample. Compared to baseline methods, our approach consistently yields better quality regarding geometry, texture, and text-3D alignment. + +(One-2-3-45 [38], Lara [8], CRM [76], LGM [67]), and native 3D diffusion models (Shap-E [27], LN3Diff-image [32]). Quantitatively, we benchmark rendering metrics with FID [20], KID [3], and MUSIQ [29, 92]. For 3D quality evaluation, we report Coverage Score (COV), and Minimum Matching Distance (MMD) score, as shown in Table 1. Our SAR3D demonstrates strong performance across all metrics. + +Furthermore, we also profile the generation speed. This timing covers the complete process, from input image processing to mesh extraction. Thanks to efficient next-scale prediction, SAR3D achieves exceptionally fast generation speeds, achieving 0.82 seconds and 1.46 seconds respectively on a single A6000 GPU. Since other baseline methods are tested on Tesla V100 GPUs, we scale our results by a factor of 2 for fair comparison in Table 1. + +The qualitative comparisons between SAR3D and existing methods is also included in Fig. 4. Compared to single-image to 3D methods like OpenLRM[19], and multi-view image to 3D methods like LGM [67], our approach achieves better 3D consistency across views and reduces distortion in generated 3D objects. Compared to native 3D diffusion models like LN3Diff [32], SAR3D produces more complete 3D models. Additional qualitative results are shown in Fig. 5. For more comparisons with other methods and + +details for training data, please refer to the Supp Mat. + +# 5.2. Text to 3D + +In addition to image-to-3D generation, SAR3D also supports the creation of high-quality 3D assets from text prompts. As shown in Fig. 5, SAR3D generates diverse and detailed 3D objects based on the same text input. For instance, in the first and second samples, SAR3D produces different shapes for the cannon barrel and chair base, while in the third sample, it varies the texture of a wooden chest. In Fig. 6, we compare our method with other text-to-3D generation approaches, including Point-E [46], ShapE [27], 3DTopia [23], and LN3Diff [32]. Compared to these baselines, SAR3D achieves sharper visual results and better alignment with the input prompts. For example, in the second sample, SAR3D generates red patterns on the handle, closely matching the input text.. In contrast, Point-E [46] reverses the colors of the handle and blade, 3DTopia [23] produces a completely red sword, and Shap-E [27] yields a less detailed result. + +# 5.3.3D Captioning + +3D Object Captioning. In this section, we present the results of our 3D understanding model applied to various 3D models. As shown in Fig. 8, given the prompt "Give + +![](images/342f3dbdfffb53ef548eae675fca22b1af7b652223ed81edc39f08bccbd5d8e9.jpg) +6duDpuuDusrAupn uDuaeau 4ndU + +![](images/bbdead4eb6127ecb043e45a869d7f142b5b2029a3c28c1d5f8617d3c92eebb3a.jpg) +A blue, curved sword with a long and slender blade, featuring a hilt or handle at the base. It has a metallic sheen and is associated with the protagonist, Link, in The Legend of Zelda video game franchise. + +![](images/e7b6f1d40038eada258e0a473210c866faa2feb38eadca6f5141a3bf72d26f10.jpg) + +![](images/fec023d28cea05bae7e86a90874173517b6785f05b5f4f682ed26ba27403f23e.jpg) + +![](images/4b6d5e6d2083e94f451cdf09f447dd4359bac8096620c4d1593760092cd2a50f.jpg) +with a red cap, depicted as the spotted + +![](images/bde5fd108111525d4fa9e6d1feb0117ac7ef01e94eb2e5d37d05ff79722a799a.jpg) +"Bookshelf" +a white shelving unit in the shape of a building. + +![](images/689366401793067a7315fe96d57f2aa588f59abb6f149e773a500d916bbd99bd.jpg) +Figure 7. Simultaneous 3D Generation and Captioning. Given a single image or text, SAR3D-LLM can generate both a 3D model and a descriptive caption for the model. +Figure 8. 3D Object Captioning. Given a 3D model, SAR3D-LLM can generate captions that include both category and details. + +a concise interpretation of the 3D data presented here," SAR3D-LLM can generate both the correct category and fine details of the input 3D models. For example, in the chair case, SAR3D accurately describes the shape (curved backrest), colors (blue and white), and components (black base, cushioned seat), whereas the ground truth text lacks these details. Furthermore, our 3D tokens enable the LLM to capture the spatial relationships between different parts of the model. For instance, in the third column of Fig. 8, SAR3D uses phrases like "leading up to" and "in front of" to describe the spatial relationship between the staircase, entrance, and patio area, while the ground truth label merely lists these parts without capturing their spatial connections. Simultaneous 3D Generation and Captioning. In addition to interpreting tokens encoded by our 3D VQVAE, SAR3D can also process 3D tokens generated by our autoregressive model to enable simultaneous 3D generation and captioning, as illustrated in Fig. 7. Given the condition input image or text, SAR3D not only generates the entire object but also detailed captions based on truncated scales of the generated 3D tokens. Notably, in text-conditioned generation and understanding, SAR3D generates additional details beyond those specified in the input text, resulting in accurate and comprehensive descriptions of the generated + +content. + +# 6. Limitations + +The first limitation is that, while SAR3D can generate high-quality 3D objects and detailed interpretations, it currently relies on two separate autoregressive models. Future work could focus on developing a truly multimodal model [78] capable of processing tokens that integrate both text and 3D information, producing both 3D and text outputs. Besides, the quality of the geometry and texture is limited by volume rendering. Using more efficient 3D representations [26] or cascaded generation [90] will further boost the overall quality. Finally, our method's scalability remains unverified due to resource limits. However, 2D scaling laws in VAR [68] suggest its feasibility. We believe that with more resources, our method has the potential to demonstrate favorable scaling laws in 3D generation and understanding. + +# 7. Conclusion + +In this work, we presented SAR3D, a novel framework that advances both fast 3D object generation and comprehensive 3D understanding through a multi-scale VQVAE and autoregressive modeling. By introducing a latent tri-plane next-scale prediction approach, we addressed the speed limitations of existing diffusion-based 3D generation methods, achieving sub-second generation times with high-quality results. Furthermore, our multi-scale VQVAE enables a pretrained LLM to process and interpret multimodal inputs by leveraging truncated scale 3D tokens, demonstrating the capability of LLMs for detailed 3D object captioning as well as simultaneous 3D generation and captioning. Experimental results underscore SAR3D's efficiency and effectiveness in 3D generation and understanding tasks, positioning it as a versatile tool for multimodal AI applications. Future research may further explore scalability and extend SAR3D's application to broader 3D content and multimodal understanding challenges. + +# Acknowledgements + +This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award . This research is also supported by NTU SUG-NAP and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35: 23716-23736, 2022. 1, 3 +[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations (ICLR), 2018. 7 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1:8, 2024. 4 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1, 2, 3, 4 +[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2 +[7] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11315-11325, 2022. 2, 4 +[8] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), pages 338-355. Springer, 2024. 7 +[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2 +[10] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets + +creation using spatially-aware diffusion guidance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 128-146. Springer, 2024. 2 +[11] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In International Conference on Learning Representations (ICLR), 2025. 1, 2 +[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3, 5, 6 +[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1 +[14] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems (NeurIPS), 36:35799-35813, 2023. 1, 2 +[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 5 +[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. In International Conference on Machine Learning (ICML), pages 8469-8488. PMLR, 2023. 1 +[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2 +[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 1 +[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 6, 7 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7 + +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 5 +[23] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3DTopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024.6,7 +[24] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems (NeurIPS), 36: 20482-20494, 2023. 1, 3 +[25] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1, 2, 5, 6 +[26] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2024. 8 +[27] Heewoo Jun and Alex Nichol. Shap-E: Generating conditional 3D implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 7 +[28] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. 4 +[29] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, 2021. 7 +[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. arXiv, 2013. 2, 4 +[31] Adam R Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Sona Mokrá, and Danilo Jimenez Rezende. Nerf-vae: A geometry aware 3d scene generative model. In International Conference on Machine Learning (ICML), pages 5742-5752. PMLR, 2021. 2 +[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 112-130. Springer, 2024. 1, 2, 4, 7 +[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussian anything: Interactive point cloud latent diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2025. 2 +[34] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2 + +[35] Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. CraftsMan: High-fidelity mesh generation with 3D native generation and interactive geometry refiner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2 +[36] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 2 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems (NeurIPS), 36:34892-34916, 2023. 3 +[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems (NeurIPS), 36:22226-22246, 2023. 7 +[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2 +[40] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2 +[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems (NeurIPS), 36:75307-75337, 2023. 5 +[42] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 180-197. Springer, 2024. 5 +[43] Tengchao Lv, Yupan Huang, Jingye Chen, Yuzhong Zhao, Yilin Jia, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 3 +[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 4 +[45] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, 2022. 2 +[46] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 7 +[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, + +Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, pages 1-31, 2024. 5 +[48] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (NeurIPS), 35:27730-27744, 2022. 1 +[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 4, 5 +[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 1, 2 +[51] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaiheng Ma. Shapeim: Universal 3d object understanding for embodied interaction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 214-238. Springer, 2024. 3 +[52] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 5 +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PmLR, 2021. 5 +[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831. Pmlr, 2021. 2 +[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5 +[56] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations (ICLR), 2017. 2 +[57] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 4 +[58] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao + +Su. Zero123++: a single image to consistent multi-view diffusion base model. In arXiv, 2023. 2 +[59] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2, 4 +[60] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2 +[61] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems (NeurIPS), 34:19313-19325, 2021. 4 +[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2 +[63] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6 +[64] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations (ICLR), 2024. 1 +[65] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1, 6 +[66] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2 +[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2024. 2, 4, 5, 7 +[68] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in Neural Information Processing Systems (NeurIPS), 37:84839-84865, 2024. 1, 2, 3, 4, 6, 8 +[69] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 3, 5 +[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, + +Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. +Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 3, 5, 6 +[71] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems (NeurIPS), 35:10021-10039, 2022. 2 +[72] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4 +[73] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4563-4573, 2023. 2 +[74] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2 +[75] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems (NeurIPS), 36:8406-8441, 2023. 2 +[76] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 57-74. Springer, 2024. 7 +[77] Zhenwei Wang, Tengfei Wang, Zexin He, Gerhard Petrus Hancke, Ziwei Liu, and Rynson WH Lau. Phidias: A generative model for creating 3d content from text, image, and 3d conditions with reference-augmented diffusion. In International Conference on Learning Representations (ICLR), 2025. 2 +[78] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 3, 5, 8 +[79] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9065-9075, 2023. 4 +[80] Shuang Wu, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4 +[81] Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, James Lucas, Antonio Torralba, Sanja Fidler, and Xiaohui Zeng. Latte3d: Large-scale amortized text-to-enhanced3d synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 305-322. Springer, 2024. 4 + +[82] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4 +[83] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 131-147. Springer, 2024. 1, 3, 5, 6 +[84] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1179-1189, 2023. 3 +[85] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. 6 +[86] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. In International Conference on Learning Representations (ICLR), 2024. 1 +[87] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems (NeurIPS), 37: 128940-128966, 2024. 2 +[88] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 2, 3 +[89] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2 +[90] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 8 +[91] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4 +[92] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems (NeurIPS), 35: 30599-30611, 2022. 7 \ No newline at end of file diff --git a/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/images.zip b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eea7617c503813c28eac3b13ec712d56bc12d0ee --- /dev/null +++ b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eab3ed7b8d5f4721c6a3c429972e0a2e2e12d5367c65ba322151867172f56038 +size 507003 diff --git a/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/layout.json b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ed32197c2066e293eb7bd07a223b620eddcdb8a8 --- /dev/null +++ b/2025/SAR3D_ Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE/layout.json @@ -0,0 +1,10827 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 89, + 102, + 520, + 138 + ], + "type": "text", + "content": "SAR3D: Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 510, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 510, + 177 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 510, + 177 + ], + "type": "text", + "content": "Yongwei Chen1 Yushi Lan1 Shangchen Zhou1 Tengfei Wang2 Xingang Pan1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 180, + 410, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 180, + 410, + 195 + ], + "spans": [ + { + "bbox": [ + 200, + 180, + 410, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 200, + 180, + 410, + 195 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 199, + 195, + 411, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 195, + 411, + 209 + ], + "spans": [ + { + "bbox": [ + 199, + 195, + 411, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 199, + 195, + 411, + 209 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 202, + 209, + 405, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 209, + 405, + 224 + ], + "spans": [ + { + "bbox": [ + 202, + 209, + 405, + 224 + ], + "type": "text", + "content": "https://cyw-3d.github.io/projects/SAR3D/" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 234, + 390, + 306 + ], + "blocks": [ + { + "bbox": [ + 57, + 234, + 390, + 306 + ], + "lines": [ + { + "bbox": [ + 57, + 234, + 390, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 234, + 390, + 306 + ], + "type": "image", + "image_path": "e496bf99f4c4ac90bc6ba30bbfac6f766fd35add1b2c4aa3d9495d0f1f5dc36f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 311, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 344 + ], + "type": "text", + "content": "Figure 1. Our method, SAR3D, proposes a comprehensive framework for 3D generation and understanding via autoregressive modeling. For (a) 3D generation, given a single image or text prompt, SAR3D generates multi-scale 3D objects in an autoregressive manner. For (b) 3D understanding, SAR3D-LLM can interpret a 3D model and provide a detailed description." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 432, + 235, + 463, + 269 + ], + "blocks": [ + { + "bbox": [ + 432, + 235, + 463, + 269 + ], + "lines": [ + { + "bbox": [ + 432, + 235, + 463, + 269 + ], + "spans": [ + { + "bbox": [ + 432, + 235, + 463, + 269 + ], + "type": "image", + "image_path": "735a1c30784266fb056fd2ffa4c090f00b169c18d99ef3b3569bc09c4991c048.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 480, + 239, + 539, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 480, + 239, + 539, + 262 + ], + "spans": [ + { + "bbox": [ + 480, + 239, + 539, + 262 + ], + "type": "text", + "content": "Give a concise interpretation of the 3D data presented here." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 405, + 272, + 548, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 272, + 548, + 295 + ], + "spans": [ + { + "bbox": [ + 405, + 272, + 548, + 295 + ], + "type": "text", + "content": "A small, cute, and round yellow Pikachu stuffed toy with a distinctive yellow color, pointy ears, and large black eyes, resembling the iconic Pokémon character." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 444, + 297, + 504, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 297, + 504, + 307 + ], + "spans": [ + { + "bbox": [ + 444, + 297, + 504, + 307 + ], + "type": "text", + "content": "(b) 3D Understanding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 358, + 200, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 358, + 200, + 371 + ], + "spans": [ + { + "bbox": [ + 151, + 358, + 200, + 371 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 383, + 297, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 383, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 54, + 383, + 297, + 635 + ], + "type": "text", + "content": "Autoregressive models have demonstrated remarkable success across various fields, from large language models (LLMs) to large multimodal models (LMMs) and 2D content generation, moving closer to artificial general intelligence (AGI). Despite these advances, applying autoregressive approaches to 3D object generation and understanding remains largely unexplored. This paper introduces Scale AutoRegressive 3D (SAR3D), a novel framework that leverages a multi-scale 3D vector-quantized variational autoencoder (VQVAE) to tokenize 3D objects for efficient autoregressive generation and detailed understanding. By predicting the next scale in a multi-scale latent representation instead of the next single token, SAR3D reduces generation time significantly, achieving fast 3D object generation in just 0.82 seconds on an A6000 GPU. Additionally, given the tokens enriched with hierarchical 3D-aware information, we finetune a pretrained LLM on them, enabling multimodal comprehension of 3D content. Our experiments show that SAR3D surpasses current 3D generation methods in both speed and quality and allows LLMs to interpret and caption 3D models comprehensively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 657, + 135, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 135, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 135, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Autoregressive models have achieved remarkable success in various domains, including large language models (LLMs) [1, 5, 13, 48, 69, 70], 2D generation [63, 68, 86], and large" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 359, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 479 + ], + "type": "text", + "content": "multimodal models (LMMs) [2, 16, 64], marking significant strides toward artificial general intelligence (AGI). By predicting the next token [1] or scale [68], autoregressive models are trained using a simple cross-entropy loss and share similar architectures. This commonality allows them to easily benefit from the optimizations the community has developed over the years for LLMs. Nevertheless, there has been limited exploration of how this next-token/scale prediction approach can be applied to 3D object generation and understanding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": "Previously, the scarcity of 3D data pushed researchers to rely on pretrained 2D diffusion models [55] as a prior to generate 3D objects via multi-view score distillation sampling (SDS) loss [50]. Following this, alternative approaches [25, 65] have focused on training feed-forward 3D reconstruction models for fast 3D reconstruction, enabled by large-scale 3D object datasets like Objaverse [14, 15]. These methods are capable of generating 3D assets in mere seconds. More recently, native 3D generative models [27, 32, 46, 90] have emerged, attempting to sample 3D assets from noise under various conditions (e.g., text or image). However, as most of these models rely on diffusion-based methods, they suffer from slow inference times. In parallel, mesh-based generative models [11, 60] attempt to generate 3D topology using autoregressive predictions, but they are limited in detail and require slow, face-by-face predictions. For 3D understanding, some studies [18, 24, 83] have attempted to finetune LLMs on 3D data to interpret the 3D world. However, these methods primarily use 3D point" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28371" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "cloud representations, which are limited in capturing fine object details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 294, + 204 + ], + "type": "text", + "content": "In light of the immense potential of autoregressive next-token prediction paradigm and their underexplored status in 3D generation and understanding, we pose an important question: Can autoregressive models be effectively applied to achieve both fast 3D object generation and detailed understanding? Addressing this challenge requires a 3D tokenizer capable of encapsulating detailed information about 3D objects into compact tokens, as well as an efficient schedule for autoregressive prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 348 + ], + "type": "text", + "content": "In this work, we propose Scale AutoRegressive 3D (SAR3D), a framework that leverages autoregressive models for both fast object generation and comprehensive understanding. Central to SAR3D is a multi-scale 3D vector-quantized variational autoencoder (VQVAE) capable of tokenizing 3D objects into hierarchical levels of tokens. These multi-scale tokens facilitate next-scale prediction training, significantly reducing the steps required for 3D generation compared to diffusion models and traditional next-token prediction methods. Furthermore, the tokens, enriched with 3D-aware information, are naturally compatible with LLM fine-tuning for detailed 3D understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 350, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 553 + ], + "type": "text", + "content": "Specifically, our SAR3D introduces a multi-scale 3D VQVAE to encode multiview RGB images, along with their corresponding depth and camera parameters, into a multiscale latent triplane representation. For 3D generation, we train an autoregressive model to predict the next scale of this latent triplane based on previous scales, conditioned on a single image or text prompt. By predicting the next scale instead of the next single token, our approach significantly reduces generation time, achieving 3D object generation in only 0.82 seconds on an A6000 GPU. For 3D understanding, we use truncated scale tokens from our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to process multimodal inputs that combine text and 3D tokens. Notably, our finetuned LLM can interpret 3D tokens encoded by our VQVAE as well as those generated by our autoregressive model, supporting both 3D captioning and simultaneous generation and understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 554, + 294, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 294, + 601 + ], + "type": "text", + "content": "Experiments show that SAR3D surpasses existing 3D generation methods in both speed and quality, and our VQVAE enables LLMs to generate detailed captions for 3D objects. Our key technical contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": "- We introduce SAR3D, a framework designed for both fast 3D object generation and detailed 3D understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 631, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 295, + 677 + ], + "type": "text", + "content": "- For 3D generation, our method utilizes a next-scale prediction approach for both text-to-3D and single-image-to-3D, achieving faster generation with higher quality compared to existing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "content": "- For 3D understanding, we leverage truncated scale tokens generated by our 3D multi-scale VQVAE to finetune a pretrained LLM, enabling it to interpret and describe 3D" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": "models, and showcasing the potential of our approach in multimodal applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 107, + 406, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 107, + 406, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 107, + 406, + 119 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 127, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 305 + ], + "type": "text", + "content": "3D Generative Models. With the success of 2D diffusion models [22, 62], their adaptation for 3D generation has been widely explored. Score distillation sampling [9, 10, 50, 66, 75] leverages these 2D models to distill 3D content, yet it encounters challenges such as costly optimization, mode collapse, and the Janus problem. More recent approaches adopt a two-stage pipeline, generating multiview images first [40, 58, 59, 77] and then reconstructing 3D structures through feed-forward processes [25, 67, 82]. Although promising, these methods are constrained by the quality of multi-view image generation, which often lacks view consistency [39] and fails to scale to higher resolutions [58]. Additionally, this two-stage setup limits 3D editing capabilities due to the absence of a 3D-aware latent space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 307, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 486 + ], + "type": "text", + "content": "To overcome these limitations, native 3D diffusion models [32, 33, 35, 71, 73, 89, 90] have been introduced, offering high-quality, efficient, and scalable 3D generation. Native 3D diffusion pipelines use a two-stage training process: first encoding 3D objects into a VAE latent space [30, 31], followed by applying a latent diffusion model on the resulting codes. However, diffusion-based 3D generation is slow during inference, and its latent space cannot be easily renovated for 3D understanding. In parallel, mesh generative models [11, 60] generates 3D topology through autoregressive prediction. However, they lack details and require slow per-face prediction. In this study, we show that our autoregressive SAR3D achieves efficient sampling, better quality, and can naturally be used for 3D understanding by cascading a Large Language Model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 487, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 689 + ], + "type": "text", + "content": "Autoregressive Visual Generation. Pioneered by PixelCNN [56], researchers have proposed to generate images as pixel sequences. Early research VQVAE [72] and VQGAN [17] further quantize image patches into discrete tokens and employ a transformer to learn the autoregressive priors, similar to language modeling [1]. The following research further improves its sampling speed [7] and tokenization efficiency [87]. To further improve the reconstruction quality, RQVAE [34] proposed multi-scale quantization and VAR [68] transforms it into next-scale prediction and significantly enhances the sampling speed. Parallel efforts are spent on scaling up autoregressive models on text-conditioned visual generation task [36, 54, 63, 74]. In 3D area, though some preliminary works [45, 88] studied 3D autoregressive modeling on toy dataset [6] without textures, the research on autoregressive 3D generation over large scale 3D dataset [14, 15] is missing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Large Multimodal Models. Inspired by the great success of large language models (LLMs) [5, 69, 70], large" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28372" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "multimodal models (LMMs) are proposed to comprehend and generate a wide range of information beyond text-based data. Two prominent paradigms exist to train the models in an end-to-end strategy: training the model from scratch [43] or aligning pre-trained LLMs and unimodal encoders [2, 37]. The second strategy typically involves a two-stage process: alignment of the unimodal encoder with LLM's feature space, and instruction-based fine-tuning. Following up works also extend the LMMs to 3D understanding, specifically on point cloud [24, 51, 83, 84]. However, point clouds significantly ignore the details of the given 3D inputs. Here, we demonstrate that our 3D VQVAE can connect with an LLM for a detailed 3D understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 240, + 141, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 141, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 141, + 252 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 260, + 290, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 290, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 290, + 273 + ], + "type": "text", + "content": "3.1. Multi-scale Visual Autoregressive Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": "VAR [68] presents a multi-scale visual modeling method for image generation, shifting from \"next-token prediction\" to \"next-scale prediction\", which significantly improves the inference speed of autoregressive models. Given an encoded feature map " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": " of an input image " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": ", VAR quantizes " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": " multi-scale token maps " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "R = (r_1, r_2, \\dots, r_K)" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": " at an increasingly higher resolution " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "h_k \\times w_k" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "r_K" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": " matches the resolution of the input feature map " + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 279, + 296, + 387 + ], + "type": "text", + "content": ". The autoregressive likelihood reads as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 397, + 296, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 397, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 82, + 397, + 296, + 430 + ], + "type": "interline_equation", + "content": "p \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}", + "image_path": "66a6ad76a0f62222e5716767889e2d77026f1b2235026bf48cc84edf555eaa5d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": "where each autoregressive unit " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "r_k \\in [V]^{h_k \\times w_k}" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " is the token map at scale " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": ", and the sequence " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "(r_1, r_2, \\ldots, r_{k-1})" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " serves as the prefix for " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": ". Tockenize the input image " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " to multi-scale discrete token maps " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " for the learning of next-scale prediction, VAR proposes a multi-scale VQVAE with a multi-scale quantizer " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 523, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 523, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 124, + 523, + 295, + 536 + ], + "type": "interline_equation", + "content": "f = \\mathcal {E} (I), \\quad R = \\mathcal {Q} (f), \\tag {2}", + "image_path": "ca2353eb7581367ba195e5319f65b2a5250400cd5176a4c4ea2f911b01cd2ebd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": " denotes the raw image and " + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": " is the image encoder. This quantization process will map " + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": " to a sequence of multiscale token maps by looking up the nearest code [72] in codebook " + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "inline_equation", + "content": "Z\\in \\mathbb{R}^{V\\times C}" + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 602, + 295, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 602, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 66, + 602, + 295, + 638 + ], + "type": "interline_equation", + "content": "z _ {k} ^ {(i, j)} = \\left(\\arg \\min _ {v \\in [ V ]} \\left\\| \\operatorname {l o o k u p} (Z, v) - r _ {k} ^ {(i, j)} \\right\\| _ {2}\\right) \\in [ V ], \\tag {3}", + "image_path": "95fed2e0644b09429fcea6b5c17a70f6228ac2c47ca41d08609da14940af53cc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\mathrm{lookup}(Z,v)" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": " means taking the " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": "-th vector in codebook " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": ". To train the quantized autoencoder, " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": " is looked up by every " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "z_{k}(i,j)" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": ", the approximation of original " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": ". Then a new image " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": " is reconstructed using the decoder " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 55, + 639, + 296, + 699 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 110, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\hat {f} = \\operatorname {l o o k u p} (Z, z), \\quad \\hat {I} = \\mathcal {D} (\\hat {f}). \\tag {4}", + "image_path": "4e1b95c9e51ce26c59c7a45edc61601434a6f6d62a599f6e4e28b4b018f953ec.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "Once fully trained, the autoencoder " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E},\\mathcal{Q},\\mathcal{D}\\}" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " will tokenize the incoming images for training the unidirectional autoregressive model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 114, + 535, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 535, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 535, + 126 + ], + "type": "text", + "content": "3.2. PointLLM for Point Cloud Understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": "Given multi-modal sentences containing both point clouds " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " and text, where " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " is the number of points and " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " is the dimension of each point, PointLLM [83] aims to perform 3D point cloud understanding by finetuning a pretrained large language model [12, 70]. It consists of three main components: a pretrained point cloud encoder " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{pe}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " (e.g., Point-BERT [88]), a projector " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{proj}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " and a pretrained large language model backbone " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{llm}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{pe}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{proj}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " project " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " to a point cloud token sequence " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "Z_{p} \\in \\mathbb{R}^{m \\times c'}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " is the total number of tokens and " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " is the projected dimension of the point tokens. The final mixed sequence of tokens " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "Z_{m} = (z_{1}, z_{2}, \\dots, z_{l}) \\in \\mathbb{R}^{l \\times c}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " consists of both point tokens " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "Z_{p}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": " and text tokens " + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 133, + 555, + 290 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 295, + 555, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 295, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 339, + 295, + 555, + 308 + ], + "type": "interline_equation", + "content": "Z _ {p} = \\Gamma_ {\\mathrm {p r o j}} \\left(\\Gamma_ {\\mathrm {p e}} (P)\\right), \\quad Z _ {m} = \\operatorname {C o n c a t} \\left(Z _ {p}, Z _ {t}\\right), \\tag {5}", + "image_path": "d567c2d5b9873caa8f3151e553055714b6a95b5fb96fe5203cb29aa4ac133c32.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "content": " is obtained by tokenizer of " + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{llm}}" + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "content": " means concatenation of two vectors. The LLM backbone " + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{llm}}" + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "content": " is a GPT-style Transformers [5], which accepts a sequence of previous multi-modal tokens " + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "inline_equation", + "content": "Z_{< i} = (z_1, \\ldots, z_{i-1})" + }, + { + "bbox": [ + 313, + 313, + 554, + 373 + ], + "type": "text", + "content": " and predict the next token:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 400, + 379, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 379, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 400, + 379, + 553, + 392 + ], + "type": "interline_equation", + "content": "z _ {i} = \\Gamma_ {\\mathrm {l l m}} \\left(Z _ {< i}\\right). \\tag {6}", + "image_path": "b6fa0b161d4e53f368b5aafc86f110dcc4a247fa64e4a6a575d4f5b9f77ac345.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "text", + "content": "The finetuning process has two stages. In the first stage, it freezes " + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\mathrm{pe}},\\Gamma_{\\mathrm{llm}}\\}" + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "text", + "content": " and finetunes " + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{proj}}" + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "text", + "content": " to align point features with the text token space. In the second stage, it freezes " + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{pe}}" + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "text", + "content": " and finetunes " + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\{\\Gamma_{\\mathrm{llm}},\\Gamma_{\\mathrm{proj}}\\}" + }, + { + "bbox": [ + 313, + 397, + 554, + 446 + ], + "type": "text", + "content": " together." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 454, + 370, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 370, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 370, + 466 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "In this section, we present SAR3D for high-quality 3D object generation and detailed understanding. First, we introduce a multi-scale 3D vector-quantized variational autoencoder (VQVAE) in Sec. 4.1, which tokenizes the input 3D model into multi-scale tokens. Fig. 2 illustrates the design of our 3D VQVAE. Next, in Sec. 4.2, with the whole sequence of different scales of the feature tokens, we train an autoregressive model to perform the next scale prediction given a single image or text prompt, which is only supervised by simple cross-entropy loss. Finally, in Sec. 4.3, we explore using truncated scales of the whole sequence to fine-tune a pretrained LLM [12, 69, 70] to handle multimodal input sequence containing both 3D and text tokens, thereby enabling the understanding of the input 3D model. Fig. 3 illustrates our 3D generation and understanding pipeline. Different from other methods [78] that train different encoders for generation and understanding, we train a single VQVAE and use the whole and truncated sequence for generation and understanding, respectively. Details of our SAR3D are shown below." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28373" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 558, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 194 + ], + "type": "image", + "image_path": "4bc020b430e7537d276a1e7d75e15f90258d8e7dde158b0407979bd1b40af577.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Overview of Multi-scale VQVAE. Given a 3D model, we leverage multi-view RGB-D(depth) renderings and Plücker embeddings as the input to our multi-view encoder " + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "text", + "content": ". The encoder predicts a continuous feature map that is then quantized by the multi-scale quantizer " + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "text", + "content": ", giving " + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "inline_equation", + "content": "R = (r_1, r_2, \\ldots, r_K)" + }, + { + "bbox": [ + 54, + 201, + 555, + 258 + ], + "type": "text", + "content": " of latent tri-plane features. Each code of different scales share the same codebook. The triplane decoder then converts the quantized latent triplane features into the triplane representation through a plane-wise manner. The predicted triplane is multi-view supervised with the ground truth image, depth, and normal." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 188, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 188, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 188, + 279 + ], + "type": "text", + "content": "4.1. Multi-scale 3D VQVAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 285, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 295, + 403 + ], + "type": "text", + "content": "As demonstrated by previous studies [4, 7, 32, 55], the key to high-quality visual generation lies in a compact latent space achieved by a specially designed variational autoencoder [30, 72]. To achieve both fast 3D generation and detailed understanding, we propose a multi-scale 3D VQVAE that maps a given 3D object into a discrete multi-scale latent space. To encode a 3D model, we leverage its multi-view posed RGB-D renderings as input. This approach offers a comprehensive representation of the 3D structure and enables compatibility with existing architectures [79]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": "Specifically, the input of VQVAE is a set of multi-view renderings of the 3D object from 6 views. Each rendering " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "M = (I, \\Delta, \\pi)" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " captures essential 3D attributes, representing the object from a specific viewpoint: the RGB image " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": ", the depth map " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": ", and the corresponding camera pose " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": ". To standardize these 3D attributes, we transform the camera pose " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " into Plücker coordinates [61], expressed as " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i = (\\mathbf{o} \\times \\mathbf{d}_{u,v}, \\mathbf{d}_{u,v}) \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " denotes the camera origin, " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{u,v} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " is the normalized ray direction, and " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " represents the cross product. Consequently, the Plücker embedding of the camera " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{H \\times W \\times 6}" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": ". The final representation is formed by channel-wise concatenating these elements, resulting in " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "M = [I \\oplus \\Delta \\oplus \\mathbf{P}] \\in \\mathbb{R}^{H \\times W \\times 10}" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 404, + 295, + 582 + ], + "type": "text", + "content": " denotes concatenation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "text", + "content": "To maintain both geometry and texture details of " + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "text", + "content": ", similar to LN3Diff [32], we encode the inputs through a multi-view convolutional encoder [59, 67]. For better 3D awareness, the latent space is designed to be a latent triplane [32, 80] " + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{3 \\times h \\times w \\times C}" + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "text", + "content": ". Besides, this representation also has spatial inductive bias and is compatible with the scale and interpolation design in VAR [68]. After encoding, " + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "text", + "content": " is interpolated to different scales and quantized using latent triplane quantization layer " + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 54, + 583, + 295, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 121, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "f = \\mathcal {E} (\\tilde {M}), \\quad R = \\mathcal {Q} (f), \\tag {7}", + "image_path": "a89abe60aecbf8748e348e24aa4b1ee1021e4004eb45aef887c70582f8b769c4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": " is the encoder of our VQVAE and " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "R = (r_1, r_2, \\dots, r_K)" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": " is the scale sequence and " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "r_k \\in \\mathbb{R}^{3 \\times h_k \\times w_k \\times C}" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": ", where each sub latent plane " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "r_k^i \\in \\mathbb{R}^{h_k \\times w_k \\times C}" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": " is independently quantized and interpolated over the shared codebook " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": ". Please refer to the Supp Mat for more quantization and interpolation details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "content": "Afterward, the decoder " + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "content": " decodes discrete scales " + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "content": " to triplane and render multiple views to calculate the reconstruction loss. To balance training stability and mesh extraction quality, we first train our model on volume rendering [44] with the loss reads as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 407, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 407, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 340, + 407, + 553, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {r e n d e r}} \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {\\mathrm {V Q}} \\mathcal {L} _ {\\mathrm {V Q}} + \\lambda_ {\\mathrm {G A N}} \\mathcal {L} _ {\\mathrm {G A N}}, \\tag {8}", + "image_path": "79e4c923add7a35fb4239dfb057b21c76c2e49cb51c7bee58f9dcdb945a85613.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{render}}" + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "text", + "content": " combines mean absolute error (MAE) and perceptual loss [91] between the rendered RGB-D images with masks and ground truth, " + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VQ}}" + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "text", + "content": " includes both encoding error and commitment loss [72], and " + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GAN}}" + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "text", + "content": " serves as the adversarial loss to encourage a perceptually rich latent space. " + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 313, + 430, + 554, + 501 + ], + "type": "text", + "content": " are the corresponding loss weights." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": "To facilitate 3D mesh extraction, we further finetune the model to the hybrid representation Flexicubes [57, 82] with the extra " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{flex}}" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 362, + 546, + 553, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 546, + 553, + 560 + ], + "spans": [ + { + "bbox": [ + 362, + 546, + 553, + 560 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f l e x}} = \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}}, \\tag {9}", + "image_path": "2d6f9c399bc559f4106469c818000d941573a884b05d999ed7d0bb88f60e0c8c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": " is MAE loss between rendered normal and ground truth, " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": " is regularization term for Flexicubes parameters [57]. " + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 313, + 568, + 554, + 628 + ], + "type": "text", + "content": " are the corresponding loss weights. Similar to LATTE3D [81], we only fine-tune the decoder of the VQVAE in this stage for stabilized training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 553, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 659 + ], + "type": "text", + "content": "4.2. 3D Generation via Multi-scale Autoregressive Modeling" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "SAR3D Transformer. We illustrate our generation framework in Fig. 3. Similar to VAR [68], we use standard GPT-style transformer [5] with AdaLN layer [49], with specific layer design following the simple rule of scaling law [28]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28374" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 235, + 81 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 235, + 81 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 235, + 81 + ], + "type": "text", + "content": "(a) 3D Generation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 553, + 223 + ], + "type": "image", + "image_path": "0befbd1d39f5c8894dbd2364518c63db036cdb11e74545011dca48d7d85ae29d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 70, + 482, + 81 + ], + "lines": [ + { + "bbox": [ + 398, + 70, + 482, + 81 + ], + "spans": [ + { + "bbox": [ + 398, + 70, + 482, + 81 + ], + "type": "text", + "content": "(b) 3D Understanding" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Overview of 3D Generation and 3D Understanding. Given a 3D model, our 3D VQVAE encodes it into multi-scale discrete tokens for both 3D generation and understanding. In (a) 3D Generation, text or a single image is encoded by " + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP_T}" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": " or DINOv2, and the encoded condition features are integrated into the decoder-only transformer via cross attention. The transformer then causally predicts each scale of the latent triplane. In (b) 3D Understanding, truncated 3D tokens are first processed with an MLP projector. The large language model receives a multimodal sequence of text and 3D tokens and generates a detailed caption describing the input 3D model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 297, + 294, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 294, + 332 + ], + "type": "text", + "content": "We adopt tri-plane latent for autoregressive prediction with Eq. 1, where different latent plane " + }, + { + "bbox": [ + 54, + 297, + 294, + 332 + ], + "type": "inline_equation", + "content": "r_k^i" + }, + { + "bbox": [ + 54, + 297, + 294, + 332 + ], + "type": "text", + "content": " is differentiated with the corresponding learnable positional embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "type": "text", + "content": "Conditional 3D Generation. Unlike feed-forward 3D reconstruction models [25, 67] that map input image into 3D, we achieve flexible multimodal 3D generation by introducing diverse conditions, as shown in Fig. 3. For text conditions, we use " + }, + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP_T}" + }, + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "type": "text", + "content": " [53] ViT-L text encoder and inject the text embeddings into the autoregressive models through cross attention. For the image-conditioned model, we use DINOv2 [47] ViT-L to extract local patch features and send them into the autoregressive model through pre-cross-attention block [25], which empirically yields better performance. Besides local patch features, we also leverage the pooled out feature of " + }, + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP_T} / \\mathrm{DINOv2}" + }, + { + "bbox": [ + 54, + 334, + 295, + 501 + ], + "type": "text", + "content": " as the start token of the sequence. Please refer to the Supp Mat for more details of our transformer blocks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": "Classifier-free Guidance. As first proposed in diffusion models [22, 55], classifier-free guidance [21] (CFG) has shown effectiveness for improving generation quality and condition alignment. Therefore, we also enable CFG in our model by randomly dropping out " + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": " of the input condition by replacing it with a null unconditional embedding [49]. During inference, the logit " + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "inline_equation", + "content": "r_g" + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": " of each token is calculated by " + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "inline_equation", + "content": "r_g = r_u + s(r_c - r_u)" + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": ", given the conditional logit " + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "inline_equation", + "content": "r_c" + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": " and the unconditional logit " + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "inline_equation", + "content": "r_u" + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 502, + 295, + 622 + ], + "type": "text", + "content": " stands for the scale of the classifier-free guidance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 282, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 282, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 282, + 647 + ], + "type": "text", + "content": "4.3. SAR3D-LLM for 3D Object Understanding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": "Since our 3D VQVAE model provides a comprehensive encoding of the given 3D object, it can be naturally extended to 3D object understanding. Following PointLLM [83], we fine-tune LLaMA [12, 69, 70] in two stages: first, we fine-tune the projector " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{proj}}" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " while freezing the LLM " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{llm}}" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " to align" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": "3D features with the text token space; second, we fine-tune both. As briefed in Sec. 3.2, the encoded 3D tokens " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " are projected to the language latent space, and concatenated with the text instruction tokens " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ". Here, we directly use the output tokens from pre-trained SAR3D VQVAE to the projector " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\Gamma_{\\mathrm{proj}}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ". Since we only study the 3D captioning [41] task here, the instruction tokens " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " is fixed to " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\tilde{Z}_{t}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " which are tokenized from \"Give a concise interpretation of the 3D data presented here\". The final framework, SAR3D-LLM, supports both detailed 3D captioning given 3D object and simultaneous 3D generation and captioning given text or image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "content": "Moreover, a surprising observation here is that not all scales in " + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "content": " are required for 3D understanding training. Empirically, we use the truncated scale latent codes " + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{R} = (r_1, r_2, \\dots, r_{K-2})" + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "content": " as the input to the LLM, which contains only " + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "content": " of the overall tokens required for training 3D generation. The final multimodal tokens " + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "inline_equation", + "content": "Z_m" + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "content": " that serve as the input to the LLM reads as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 531, + 553, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 531, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 329, + 531, + 553, + 546 + ], + "type": "interline_equation", + "content": "Z _ {p r o j} = \\Gamma_ {\\mathrm {p r o j}} (\\tilde {R}), \\quad Z _ {m} = \\operatorname {C o n c a t} \\left(Z _ {p r o j}, \\tilde {Z} _ {t}\\right), \\tag {10}", + "image_path": "18ac1cf01d583a704e9405cf85ae1296e74db10b241ff6abb9a3741dfa6e8188.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "type": "inline_equation", + "content": "Z_{proj}" + }, + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "type": "text", + "content": " is the projected 3D tokens, and " + }, + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 552, + 555, + 660 + ], + "type": "text", + "content": " means concatenation. A similar observation is also mentioned in Janus [78], where different features are required for multimodal understanding and generation. Furthermore, unlike other 3D captioning approaches such as Cap3D [41, 42], which separately extracts captions from 8 multi-view renderings and require post-processing to merge them into a unified caption, our method efficiently generates a detailed caption with a single encoding step." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 669, + 395, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 395, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 395, + 682 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Datasets. To train our model, we use renderings from G-Objaverse [15, 52] and select a high-quality subset of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28375" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 129, + 527, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 64, + 555, + 120 + ], + "lines": [ + { + "bbox": [ + 55, + 64, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 64, + 555, + 120 + ], + "type": "text", + "content": "Table 1. Quantitative Evaluation of Image-conditioned 3D Generation. We evaluate the quality of both 2D rendering and 3D shapes. As shown below, the proposed method demonstrates strong performance across all metrics. Although LGM, a multi-view images-to-3D approach, achieves slightly better performance on FID, it falls short on more advanced image quality assessment metrics such as MUSIQ and has significantly worse 3D shape quality. For multi-view to 3D methods, we also include the number of input views (V=\\#). The latency time is all profiled on Tesla V100 architecture." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 129, + 527, + 277 + ], + "lines": [ + { + "bbox": [ + 82, + 129, + 527, + 277 + ], + "spans": [ + { + "bbox": [ + 82, + 129, + 527, + 277 + ], + "type": "table", + "html": "
MethodFID↓KID(%)↓MUSIQ↑COV(%)↑MMD(‰)↓Latency-V100 (s) ↓
Splatter-Image48.803.6530.3337.6630.690.83
OpenLRM38.411.8745.4639.3329.087.21
One-2-3-45 (V=12)88.396.3459.0233.3335.0959.23
Lara (V=4)43.741.9539.3739.3328.8411.93
CRM (V=6)45.531.9364.1038.8328.9122.10
LGM (V=4)19.930.5554.7850.8322.063.87
Shap-E138.5311.9531.5161.3319.179.54
LN3Diff29.080.8950.3955.1719.947.51
SAR3D-NeRF22.550.4265.7774.1713.631.64
SAR3D-Flexicubes27.300.6367.2471.5015.252.92
", + "image_path": "67ab4c764fe35ad7ec08cb93346a7cdf429833952d4f064bb54f27cac5dd75b4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 284, + 529, + 462 + ], + "blocks": [ + { + "bbox": [ + 82, + 284, + 529, + 462 + ], + "lines": [ + { + "bbox": [ + 82, + 284, + 529, + 462 + ], + "spans": [ + { + "bbox": [ + 82, + 284, + 529, + 462 + ], + "type": "image", + "image_path": "942591ab501ea1692aa3e4b1923d1410b3835a9456aee65e413099d6399f5052.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 471, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 471, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 555, + 504 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison of Image-conditioned 3D Generation. Here, we compare with the state-of-the-art 3D generative models under different categories. As visualized here, our method achieves superior 3D consistency across views and generates intact objects without distortion. For more comparisons with other methods, please refer to the Supp Mat." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": "around " + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "inline_equation", + "content": "176K" + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": " 3D instances, where each consists of 40 random views with RGB, normal, depth maps and camera poses. For text-conditioned generation and 3D understanding training, we use captions provided by 3DTopia [23]. For image-conditioned training, we select a random view of the corresponding 3D instance as the condition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "content": "Implementation Details. In our multi-scale VQVAE, we use images with a resolution of " + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "inline_equation", + "content": "H = W = 256" + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "content": " as input. The feature map is quantized across 10 scales, with sizes of " + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "inline_equation", + "content": "3 \\times (1^2, 2^2, 3^2, 4^2, 5^2, 6^2, 8^2, 10^2, 13^2, 16^2)" + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "content": ". To enhance codebook utilization and stabilize 3D generation training, we follow [63, 85] by applying " + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "content": "-normalization to codebook vectors, setting a low codebook vector dimension " + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "inline_equation", + "content": "C = 8" + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "content": ", and using a large codebook size " + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "inline_equation", + "content": "V = 16384" + }, + { + "bbox": [ + 55, + 592, + 295, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "For 3D generation, we base our architecture on VAR [68], adding plane positional encoding for each plane." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "type": "text", + "content": "For text-conditioned generation, the model has 16 transformer blocks with 16 heads, while for image-conditioned generation, it has 24 transformer blocks with 16 heads. We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "type": "text", + "content": ". For 3D understanding, we utilize the Vicuna-7B [12] checkpoint of LLaMA [70], following PointLLM [83]. The training was conducted on 7 NVIDIA A100 GPUs for the multiscale VQVAE with batch size 28, image-conditioned transformer with batch 63, text-conditioned transformer with batch size 52. For SAR3D-LLM, the stage-1 alignment is trained with batch size 140, and stage-2 with 112." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 658, + 425, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 425, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 425, + 671 + ], + "type": "text", + "content": "5.1. Single Image to 3D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We compare our SAR3D with three categories of methods: single-image to 3D methods (Splatter-Image [65], OpenLRM [19, 25]), multi-view image to 3D methods" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28376" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 304, + 195 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 304, + 195 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 304, + 195 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 304, + 195 + ], + "type": "image", + "image_path": "394b8052b3e4104138e3fb694e22768b571f8df71e8bb00b9a9e92617ca9f4ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 202, + 456, + 215 + ], + "lines": [ + { + "bbox": [ + 153, + 202, + 456, + 215 + ], + "spans": [ + { + "bbox": [ + 153, + 202, + 456, + 215 + ], + "type": "text", + "content": "Figure 5. More results of image and text conditioned 3D generation of SAR3D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 68, + 554, + 195 + ], + "blocks": [ + { + "bbox": [ + 306, + 68, + 554, + 195 + ], + "lines": [ + { + "bbox": [ + 306, + 68, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 306, + 68, + 554, + 195 + ], + "type": "image", + "image_path": "7732b71c044fa6e9cc4250eddb7f89ae10707348a743c81a330ed7bfce6a1c24.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 224, + 550, + 345 + ], + "blocks": [ + { + "bbox": [ + 61, + 224, + 550, + 345 + ], + "lines": [ + { + "bbox": [ + 61, + 224, + 550, + 345 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 550, + 345 + ], + "type": "image", + "image_path": "e3e40b44f446ca99b5b249ffe104bb2ab19425be150473176128d8661b68bade.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 353, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 54, + 353, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 555, + 386 + ], + "type": "text", + "content": "Figure 6. Comparison of Text-conditioned 3D Generation. We present text-conditioned 3D objects generated by SAR3D, displaying two views of each sample. Compared to baseline methods, our approach consistently yields better quality regarding geometry, texture, and text-3D alignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 407, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 295, + 502 + ], + "type": "text", + "content": "(One-2-3-45 [38], Lara [8], CRM [76], LGM [67]), and native 3D diffusion models (Shap-E [27], LN3Diff-image [32]). Quantitatively, we benchmark rendering metrics with FID [20], KID [3], and MUSIQ [29, 92]. For 3D quality evaluation, we report Coverage Score (COV), and Minimum Matching Distance (MMD) score, as shown in Table 1. Our SAR3D demonstrates strong performance across all metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 506, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 295, + 602 + ], + "type": "text", + "content": "Furthermore, we also profile the generation speed. This timing covers the complete process, from input image processing to mesh extraction. Thanks to efficient next-scale prediction, SAR3D achieves exceptionally fast generation speeds, achieving 0.82 seconds and 1.46 seconds respectively on a single A6000 GPU. Since other baseline methods are tested on Tesla V100 GPUs, we scale our results by a factor of 2 for fair comparison in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "The qualitative comparisons between SAR3D and existing methods is also included in Fig. 4. Compared to single-image to 3D methods like OpenLRM[19], and multi-view image to 3D methods like LGM [67], our approach achieves better 3D consistency across views and reduces distortion in generated 3D objects. Compared to native 3D diffusion models like LN3Diff [32], SAR3D produces more complete 3D models. Additional qualitative results are shown in Fig. 5. For more comparisons with other methods and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 529, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 529, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 529, + 419 + ], + "type": "text", + "content": "details for training data, please refer to the Supp Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 427, + 386, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 386, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 386, + 439 + ], + "type": "text", + "content": "5.2. Text to 3D" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 446, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 446, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 312, + 446, + 555, + 649 + ], + "type": "text", + "content": "In addition to image-to-3D generation, SAR3D also supports the creation of high-quality 3D assets from text prompts. As shown in Fig. 5, SAR3D generates diverse and detailed 3D objects based on the same text input. For instance, in the first and second samples, SAR3D produces different shapes for the cannon barrel and chair base, while in the third sample, it varies the texture of a wooden chest. In Fig. 6, we compare our method with other text-to-3D generation approaches, including Point-E [46], ShapE [27], 3DTopia [23], and LN3Diff [32]. Compared to these baselines, SAR3D achieves sharper visual results and better alignment with the input prompts. For example, in the second sample, SAR3D generates red patterns on the handle, closely matching the input text.. In contrast, Point-E [46] reverses the colors of the handle and blade, 3DTopia [23] produces a completely red sword, and Shap-E [27] yields a less detailed result." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 406, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 406, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 406, + 672 + ], + "type": "text", + "content": "5.3.3D Captioning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "3D Object Captioning. In this section, we present the results of our 3D understanding model applied to various 3D models. As shown in Fig. 8, given the prompt \"Give" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28377" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 122, + 178 + ], + "blocks": [ + { + "bbox": [ + 63, + 77, + 74, + 192 + ], + "lines": [ + { + "bbox": [ + 63, + 77, + 74, + 192 + ], + "spans": [ + { + "bbox": [ + 63, + 77, + 74, + 192 + ], + "type": "text", + "content": "6duDpuuDusrAupn uDuaeau 4ndU" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 70, + 122, + 178 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 122, + 178 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 122, + 178 + ], + "type": "image", + "image_path": "342f3dbdfffb53ef548eae675fca22b1af7b652223ed81edc39f08bccbd5d8e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 150, + 70, + 204, + 134 + ], + "blocks": [ + { + "bbox": [ + 150, + 70, + 204, + 134 + ], + "lines": [ + { + "bbox": [ + 150, + 70, + 204, + 134 + ], + "spans": [ + { + "bbox": [ + 150, + 70, + 204, + 134 + ], + "type": "image", + "image_path": "bbdead4eb6127ecb043e45a869d7f142b5b2029a3c28c1d5f8617d3c92eebb3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 135, + 220, + 204 + ], + "lines": [ + { + "bbox": [ + 136, + 135, + 220, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 135, + 220, + 204 + ], + "type": "text", + "content": "A blue, curved sword with a long and slender blade, featuring a hilt or handle at the base. It has a metallic sheen and is associated with the protagonist, Link, in The Legend of Zelda video game franchise." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 294, + 178 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 294, + 178 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 294, + 178 + ], + "type": "image", + "image_path": "e7b6f1d40038eada258e0a473210c866faa2feb38eadca6f5141a3bf72d26f10.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 365, + 194 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 365, + 194 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 365, + 194 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 365, + 194 + ], + "type": "image", + "image_path": "fec023d28cea05bae7e86a90874173517b6785f05b5f4f682ed26ba27403f23e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 72, + 381, + 86 + ], + "lines": [ + { + "bbox": [ + 367, + 72, + 381, + 86 + ], + "spans": [ + { + "bbox": [ + 367, + 72, + 381, + 86 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 383, + 72, + 456, + 169 + ], + "blocks": [ + { + "bbox": [ + 383, + 72, + 456, + 169 + ], + "lines": [ + { + "bbox": [ + 383, + 72, + 456, + 169 + ], + "spans": [ + { + "bbox": [ + 383, + 72, + 456, + 169 + ], + "type": "image", + "image_path": "4b6d5e6d2083e94f451cdf09f447dd4359bac8096620c4d1593760092cd2a50f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 134, + 466, + 168 + ], + "lines": [ + { + "bbox": [ + 425, + 134, + 466, + 168 + ], + "spans": [ + { + "bbox": [ + 425, + 134, + 466, + 168 + ], + "type": "text", + "content": "with a red cap, depicted as the spotted" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 468, + 91, + 543, + 132 + ], + "blocks": [ + { + "bbox": [ + 485, + 76, + 522, + 84 + ], + "lines": [ + { + "bbox": [ + 485, + 76, + 522, + 84 + ], + "spans": [ + { + "bbox": [ + 485, + 76, + 522, + 84 + ], + "type": "text", + "content": "\"Bookshelf\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 91, + 543, + 132 + ], + "lines": [ + { + "bbox": [ + 468, + 91, + 543, + 132 + ], + "spans": [ + { + "bbox": [ + 468, + 91, + 543, + 132 + ], + "type": "image", + "image_path": "bde5fd108111525d4fa9e6d1feb0117ac7ef01e94eb2e5d37d05ff79722a799a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 133, + 539, + 168 + ], + "lines": [ + { + "bbox": [ + 468, + 133, + 539, + 168 + ], + "spans": [ + { + "bbox": [ + 468, + 133, + 539, + 168 + ], + "type": "text", + "content": "a white shelving unit in the shape of a building." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 383, + 94, + 416, + 132 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 94, + 416, + 132 + ], + "spans": [ + { + "bbox": [ + 383, + 94, + 416, + 132 + ], + "type": "image", + "image_path": "6df6634c943d4045fe52a5889667a119e68b3daab929553d91c83e848c53b2e1.jpg" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 240, + 298, + 378 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 232 + ], + "type": "text", + "content": "Figure 7. Simultaneous 3D Generation and Captioning. Given a single image or text, SAR3D-LLM can generate both a 3D model and a descriptive caption for the model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 240, + 298, + 378 + ], + "lines": [ + { + "bbox": [ + 56, + 240, + 298, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 298, + 378 + ], + "type": "image", + "image_path": "689366401793067a7315fe96d57f2aa588f59abb6f149e773a500d916bbd99bd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 390, + 295, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 413 + ], + "type": "text", + "content": "Figure 8. 3D Object Captioning. Given a 3D model, SAR3D-LLM can generate captions that include both category and details." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 426, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 715 + ], + "type": "text", + "content": "a concise interpretation of the 3D data presented here,\" SAR3D-LLM can generate both the correct category and fine details of the input 3D models. For example, in the chair case, SAR3D accurately describes the shape (curved backrest), colors (blue and white), and components (black base, cushioned seat), whereas the ground truth text lacks these details. Furthermore, our 3D tokens enable the LLM to capture the spatial relationships between different parts of the model. For instance, in the third column of Fig. 8, SAR3D uses phrases like \"leading up to\" and \"in front of\" to describe the spatial relationship between the staircase, entrance, and patio area, while the ground truth label merely lists these parts without capturing their spatial connections. Simultaneous 3D Generation and Captioning. In addition to interpreting tokens encoded by our 3D VQVAE, SAR3D can also process 3D tokens generated by our autoregressive model to enable simultaneous 3D generation and captioning, as illustrated in Fig. 7. Given the condition input image or text, SAR3D not only generates the entire object but also detailed captions based on truncated scales of the generated 3D tokens. Notably, in text-conditioned generation and understanding, SAR3D generates additional details beyond those specified in the input text, resulting in accurate and comprehensive descriptions of the generated" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 245, + 350, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 245, + 350, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 245, + 350, + 255 + ], + "type": "text", + "content": "content." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 271, + 389, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 271, + 389, + 284 + ], + "spans": [ + { + "bbox": [ + 314, + 271, + 389, + 284 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 293, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 461 + ], + "type": "text", + "content": "The first limitation is that, while SAR3D can generate high-quality 3D objects and detailed interpretations, it currently relies on two separate autoregressive models. Future work could focus on developing a truly multimodal model [78] capable of processing tokens that integrate both text and 3D information, producing both 3D and text outputs. Besides, the quality of the geometry and texture is limited by volume rendering. Using more efficient 3D representations [26] or cascaded generation [90] will further boost the overall quality. Finally, our method's scalability remains unverified due to resource limits. However, 2D scaling laws in VAR [68] suggest its feasibility. We believe that with more resources, our method has the potential to demonstrate favorable scaling laws in 3D generation and understanding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 476, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 476, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 314, + 476, + 388, + 488 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "In this work, we presented SAR3D, a novel framework that advances both fast 3D object generation and comprehensive 3D understanding through a multi-scale VQVAE and autoregressive modeling. By introducing a latent tri-plane next-scale prediction approach, we addressed the speed limitations of existing diffusion-based 3D generation methods, achieving sub-second generation times with high-quality results. Furthermore, our multi-scale VQVAE enables a pretrained LLM to process and interpret multimodal inputs by leveraging truncated scale 3D tokens, demonstrating the capability of LLMs for detailed 3D object captioning as well as simultaneous 3D generation and captioning. Experimental results underscore SAR3D's efficiency and effectiveness in 3D generation and understanding tasks, positioning it as a versatile tool for multimodal AI applications. Future research may further explore scalability and extend SAR3D's application to broader 3D content and multimodal understanding challenges." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28378" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award . This research is also supported by NTU SUG-NAP and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 261 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 296, + 329 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35: 23716-23736, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 331, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 364 + ], + "type": "text", + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In International Conference on Learning Representations (ICLR), 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 409 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI Blog, 1:8, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 533 + ], + "type": "text", + "content": "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 588 + ], + "type": "text", + "content": "[7] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11315-11325, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 635 + ], + "type": "text", + "content": "[8] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), pages 338-355. Springer, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "text", + "content": "[9] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[10] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "creation using spatially-aware diffusion guidance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 128-146. Springer, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "text", + "content": "[11] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In International Conference on Learning Representations (ICLR), 2025. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "text", + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 304 + ], + "type": "text", + "content": "[13] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240): 1-113, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "text", + "content": "[14] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems (NeurIPS), 36:35799-35813, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 371, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 437 + ], + "type": "text", + "content": "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 438, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 503 + ], + "type": "text", + "content": "[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. In International Conference on Machine Learning (ICML), pages 8469-8488. PMLR, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 503, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 557 + ], + "type": "text", + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "type": "text", + "content": "[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28379" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[23] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3DTopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024.6,7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[24] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems (NeurIPS), 36: 20482-20494, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[25] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[26] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Proceedings, pages 1-11, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[27] Heewoo Jun and Alex Nichol. Shap-E: Generating conditional 3D implicit functions. arXiv preprint arXiv:2305.02463, 2023. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[28] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[29] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 501 + ], + "type": "text", + "content": "[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. arXiv, 2013. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 556 + ], + "type": "text", + "content": "[31] Adam R Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Sona Mokrá, and Danilo Jimenez Rezende. Nerf-vae: A geometry aware 3d scene generative model. In International Conference on Machine Learning (ICML), pages 5742-5752. PMLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 558, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 623 + ], + "type": "text", + "content": "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 112-130. Springer, 2024. 1, 2, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussian anything: Interactive point cloud latent diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[34] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[35] Weiyu Li, Jiarui Liu, Hongyu Yan, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. CraftsMan: High-fidelity mesh generation with 3D native generation and interactive geometry refiner. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 193 + ], + "type": "text", + "content": "[36] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems (NeurIPS), 36:34892-34916, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[38] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems (NeurIPS), 36:22226-22246, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 337 + ], + "type": "text", + "content": "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9298–9309, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "type": "text", + "content": "[40] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems (NeurIPS), 36:75307-75337, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[42] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In Proceedings of the European Conference on Computer Vision (ECCV), pages 180-197. Springer, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "text", + "content": "[43] Tengchao Lv, Yupan Huang, Jingye Chen, Yuzhong Zhao, Yilin Jia, Lei Cui, Shuming Ma, Yaoyao Chang, Shaohan Huang, Wenhui Wang, et al. Kosmos-2.5: A multimodal literate model. arXiv preprint arXiv:2309.11419, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "type": "text", + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 646 + ], + "type": "text", + "content": "[45] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 306-315, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[46] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28380" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, pages 1-31, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 182 + ], + "type": "text", + "content": "[48] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems (NeurIPS), 35:27730-27744, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 295, + 227 + ], + "type": "text", + "content": "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195-4205, 2023. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 270 + ], + "type": "text", + "content": "[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "type": "text", + "content": "[51] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaiheng Ma. Shapeim: Universal 3d object understanding for embodied interaction. In Proceedings of the European Conference on Computer Vision (ECCV), pages 214-238. Springer, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 403 + ], + "type": "text", + "content": "[52] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "text", + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PmLR, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), pages 8821-8831. Pmlr, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "text", + "content": "[56] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations (ICLR), 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 691 + ], + "type": "text", + "content": "[57] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[58] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Su. Zero123++: a single image to consistent multi-view diffusion base model. In arXiv, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[59] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "type": "text", + "content": "[60] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[61] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems (NeurIPS), 34:19313-19325, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[63] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[64] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations (ICLR), 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[65] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[66] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22819–22829, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 581 + ], + "type": "text", + "content": "[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 1-18. Springer, 2024. 2, 4, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[68] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in Neural Information Processing Systems (NeurIPS), 37:84839-84865, 2024. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "type": "text", + "content": "[69] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[70] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28381" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. \nLlama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[71] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems (NeurIPS), 35:10021-10039, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "type": "text", + "content": "[72] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 261 + ], + "type": "text", + "content": "[73] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4563-4573, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 305 + ], + "type": "text", + "content": "[74] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "text", + "content": "[75] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems (NeurIPS), 36:8406-8441, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 426 + ], + "type": "text", + "content": "[76] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In Proceedings of the European Conference on Computer Vision (ECCV), pages 57-74. Springer, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[77] Zhenwei Wang, Tengfei Wang, Zexin He, Gerhard Petrus Hancke, Ziwei Liu, and Rynson WH Lau. Phidias: A generative model for creating 3d content from text, image, and 3d conditions with reference-augmented diffusion. In International Conference on Learning Representations (ICLR), 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 548 + ], + "type": "text", + "content": "[78] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 3, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "text", + "content": "[79] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9065-9075, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[80] Shuang Wu, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 714 + ], + "type": "text", + "content": "[81] Kevin Xie, Jonathan Lorraine, Tianshi Cao, Jun Gao, James Lucas, Antonio Torralba, Sanja Fidler, and Xiaohui Zeng. Latte3d: Large-scale amortized text-to-enhanced3d synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 305-322. Springer, 2024. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 707 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[82] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[83] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 131-147. Springer, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 260 + ], + "type": "text", + "content": "[84] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1179-1189, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 261, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 317 + ], + "type": "text", + "content": "[85] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 318, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 318, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 553, + 384 + ], + "type": "text", + "content": "[86] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. In International Conference on Learning Representations (ICLR), 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 384, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 439 + ], + "type": "text", + "content": "[87] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems (NeurIPS), 37: 128940-128966, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 440, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 553, + 496 + ], + "type": "text", + "content": "[88] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "text", + "content": "[89] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 541, + 553, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 553, + 596 + ], + "type": "text", + "content": "[90] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1, 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 597, + 553, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 597, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 597, + 553, + 652 + ], + "type": "text", + "content": "[91] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 652, + 553, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 652, + 553, + 707 + ], + "spans": [ + { + "bbox": [ + 316, + 652, + 553, + 707 + ], + "type": "text", + "content": "[92] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. Advances in Neural Information Processing Systems (NeurIPS), 35: 30599-30611, 2022. 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28382" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_content_list.json b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..643bb1d5f906269b5fa49b17c38f0da2c1bf645f --- /dev/null +++ b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_content_list.json @@ -0,0 +1,1425 @@ +[ + { + "type": "text", + "text": "SASep: Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds", + "text_level": 1, + "bbox": [ + 133, + 130, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinfeng Xu $^{1}$ Xianzhi Li $^{1,2}$ Yuan Tang $^{1}$ Xu Han $^{1}$ Qiao Yu $^{1}$ Yixue Hao $^{1}$ Long Hu $^{1*}$ Min Chen $^{3,4}$", + "bbox": [ + 86, + 202, + 898, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Huazhong University of Science and Technology \n3South China University of Technology", + "bbox": [ + 122, + 220, + 514, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Guangdong Intelligent Robotics Institute \n $^{4}$ Pazhou Laboratory", + "bbox": [ + 532, + 220, + 864, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jinfengx, xzli, yuan_tang, xhanxu, qiaoyu_epic, yixuehao, hulong, minchen}@hust.edu.cn", + "bbox": [ + 109, + 258, + 870, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning have greatly enhanced 3D object recognition, but most models are limited to closed-set scenarios, unable to handle unknown samples in real-world applications. Open-set recognition (OSR) addresses this limitation by enabling models to both classify known classes and identify novel classes. However, current OSR methods rely on global features to differentiate known and unknown classes, treating the entire object uniformly and overlooking the varying semantic importance of its different parts. To address this gap, we propose Salience-Aware Structured Separation (SASep), which includes (i) a tunable semantic decomposition (TSD) module to semantically decompose objects into important and unimportant parts, (ii) a geometric synthesis strategy (GSS) to generate pseudo-unknown objects by combining these unimportant parts, and (iii) a synth-aided margin separation (SMS) module to enhance feature-level separation by expanding the feature distributions between classes. Together, these components improve both geometric and feature representations, enhancing the model's ability to effectively distinguish known and unknown classes. Experimental results show that SASep achieves superior performance in 3D OSR, outperforming existing state-of-the-art methods. The codes are available at https://github.com/JinfengX/SASep.", + "bbox": [ + 88, + 340, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 731, + 220, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, deep learning has significantly advanced 3D object recognition, enabling its application in diverse areas such as autonomous driving, robotics, augmented/virtual reality, etc. Most existing 3D recognition models [31, 44, 46] are designed for closed-set scenarios, where all test classes are assumed to have been seen during training. However, real-world applications usually present dynamic environments where unknown or unseen samples outside the", + "bbox": [ + 89, + 756, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a42e8c55a35204c0012fb44801e060d7e311b1752c466c643d9252a8cd2845c.jpg", + "image_caption": [ + "Figure 1. Given input objects, the TSD module semantically decomposes them into important and unimportant parts based on each point's semantic relevance to classification. The important parts enhance feature learning for known classes, while the unimportant parts are mixed in GSS to generate pseudo-unknown samples. The SMS module then constrains the features of objects and their decomposed parts to improve inter-class separation and intraclass compactness." + ], + "image_footnote": [], + "bbox": [ + 517, + 306, + 906, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "set of known classes may be encountered. These out-of-distribution samples pose a significant challenge, as the closed-set models are unable to handle these unknown samples appropriately, often misclassifying them as one of the predefined known classes. To address this challenge, open-set recognition (OSR) task has gained increasing attention recently, as it aims to extend the capabilities of closed-set models by enabling them to classify known classes and identify unknown ones simultaneously.", + "bbox": [ + 509, + 733, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early 2D OSR approaches [13, 14, 24] aim to distinguish between known and unknown classes by using discrimina", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27295", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tive functions to compute confidence scores from network predictions for each input. Samples are identified as unknown when their confidence scores are low. Thus, designing a robust open-set solver hinges on ensuring the confidence score is discriminative, meaning scores for known classes should clearly exceed those for unknown classes.", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although recent 2D OSR works [4, 36, 42] have achieved significant progress in the 2D area, research on OSR for 3D data remains relatively underdeveloped. Unlike 2D OSR, the unique spatial characteristics of 3D data limit the performance of directly adapting 2D OSR methods to 3D, as they fall short of leveraging 3D-specific structural and topological information. Recent works [2, 27, 43] propose to address the 3D OSR task from the perspective of geometric reconstruction, feature embeddings, and sample synthesis. However, these approaches, similar to many 2D OSR methods, primarily focus on assigning confidence scores based on whole-object features, often overlooking the fact that different parts of an object contribute differently to classification. For example, specific regions, such as the wheels of a car or the wings of an airplane, carry essential features that are more informative for classification than other parts.", + "bbox": [ + 91, + 185, + 483, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this observation, we propose a novel method for 3D open-set recognition on point clouds, named Salience-Aware Structured Separation (SASep), which tackles the open-set point cloud learning task by highlighting the semantically important geometric structures of objects. As illustrated in Fig. 1, SASep first decomposes the input objects into important and unimportant parts with a tunable semantic decomposition (TSD) module based on the relevance of each region to the class semantics. Then, we augment the feature learning of known classes by incorporating important parts into the training dataset. For the unimportant parts, we design a geometric synthesis strategy (GSS) to generate synthetic samples, serving as pseudo-unknown objects to simulate geometric structures of unknown classes. Furthermore, we refine the feature space with synthie-aided margin separation module (SMS) by applying constraints that reduce the feature distance between the object and its important parts, while increasing the feature distance between the object and those from other classes. This process enhances the feature representations of known classes, thus reducing the likelihood of misclassifying unknown class samples as known classes. The experimental results demonstrate that our proposed method outperforms existing OSR approaches, achieving state-of-the-art (SOTA) performance in 3D open-set recognition. Overall, our contributions can be summarized as follows:", + "bbox": [ + 91, + 445, + 483, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a novel approach for 3D open-set recognition on point clouds, called Salience-Aware Structured Separation (SASep), which enhances the distinction between known and unknown classes by utilizing the se", + "bbox": [ + 89, + 839, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mantically decomposed parts of objects.", + "bbox": [ + 527, + 90, + 792, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel semantic decomposition method TSD that decomposes an object into semantically important and unimportant regions, enabling the model to utilize structural information from different parts of the objects.", + "- We design GSS and SMS to utilize the decomposed parts of TSD to explore feature distribution of the unknown class and enhance feature distinction of the known classes, respectively." + ], + "bbox": [ + 513, + 106, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 242, + 650, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Closed-set point cloud learning", + "text_level": 1, + "bbox": [ + 511, + 267, + 787, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closed-set point cloud learning has made significant progress, focusing on accurate classification and segmentation of 3D objects from predefined categories. Existing approaches can be broadly categorized into view-based, voxel-based, and point-based methods. View-based methods [5, 21, 35, 37] classify 3D shapes by using 2D CNN to extract features from the projected 2D images of input. Voxel-based methods [6, 11, 23, 28, 39] process point clouds by dividing 3D space into voxel grids, allowing 3D CNNs to learn the spatial features. Pioneering point-based methods [31] use Multi-Layer Perceptrons (MLPs) with symmetric functions to directly process raw point cloud data. The following point-based research [12, 20, 26, 32, 40] further improved the performance by enhancing the models' capacity to aggregate the local and global features. While these methods have achieved significant success in recognizing known classes, they face notable limitations when dealing with unseen classes, as the models usually tend to misclassify the unknown objects as one of the predefined known classes. This limitation restricts the models' ability to generalize to real-world applications, where novel or unknown classes may be encountered during inference.", + "bbox": [ + 511, + 292, + 906, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Open-set 2D recognition", + "text_level": 1, + "bbox": [ + 513, + 635, + 735, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitation of closed-set learning, open-set recognition (OSR) enables models to classify known objects and identify unknown or novel classes during inference. The straightforward approaches [16, 24, 36, 41] utilize discriminative functions to evaluate confidence scores for identifying unknown classes. MSP [13] established a robust baseline by using Maximum Softmax Probability to identify samples as unknown when they exhibit low maximum softmax scores. Similarly, Maximum Logit Score (MLS) [14] directly uses the highest logit score as a confidence measure. Beyond these discriminative-based approaches, reconstruction-based methods [8, 17, 19] identify unknown objects by comparing the original inputs with their reconstructions from the generative models. Recent research in 2D OSR aims to learn feature embeddings that form compact clusters for known classes. For instance,", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/020c2a250dadfcae2541720591b7e24079188a34324344fab22d0ba69f57615b.jpg", + "image_caption": [ + "Figure 2. Overall architecture of our SASep. The architecture includes a forward process where the input point cloud $\\mathcal{P}_{in}$ is encoded with backbone, followed by a backward process utilizing the tunable semantic decomposition module to decompose $\\mathcal{P}_{in}$ into high- and low-saliency parts. Then, the high-saliency parts are used in the synth-aided margin separation module for separating feature distribution; the low-saliency parts are combined for synthetic objects with a geometric synthesis strategy." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ARPL [4] introduces reciprocal points in the embedding space to minimize the overlap of known and unknown distributions. Fontanel et al. [10] proposed a method that uses learnable class prototypes to detect anomalies by measuring the cosine distance of inputs from learned class prototypes. Furthermore, several studies [9, 25, 30] have investigated using synthetic samples to expose potential unknown classes to the model, thereby enhancing the model's generalization on samples out of the training dataset.", + "bbox": [ + 88, + 410, + 480, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Though 2D OSR has seen significant advancements, applying these methods to 3D domains presents limitations, particularly in leveraging the unique spatial and geometric complexities within 3D data. Thus, directly shifting a 2D OSR method to 3D data often fails to effectively distinguish known and unknown samples.", + "bbox": [ + 89, + 547, + 482, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Open-set 3D recognition", + "text_level": 1, + "bbox": [ + 89, + 648, + 313, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to the widely explored 2D OSR, research on 3D OSR remains relatively underdeveloped. Masuda et al. [27] developed a reconstruction-based method to obtain anomaly scores by measuring reconstruction errors between model input and output. To enhance the learning of distribution for unknown classes, Bhardwaj et al. [2] proposed generating pseudo-unknown samples by mixing objects from known classes. To further refine the synthesis strategy, Lee et al. [22] and Hong et al. [15] designed more sophisticated algorithms to synthesize high-quality pseudo unknown samples, considering the local geometric structure of objects. More recently, PartCom [43] introduced a novel part codebook to learn 3D object representations by capturing distinct part compositions across classes. Similarly, Rabino et al. [33] introduced a training-free method lever", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "aging part-based reasoning. Building on the work of 2D and 3D OSR, 3DOS [1] provided a comprehensive benchmark for 3D open-set learning, revealing the strengths and limitations of existing approaches in the 3D open-set context.", + "bbox": [ + 511, + 410, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing open-set recognition methods focus on distinguishing known and unknown objects or features as a whole, while overlooking the essential semantic information carried by different parts in an object. For example, the wheels of vehicles often exhibit more significant class-relevant information than the other parts. Humans can recognize objects even with a partial view of such recognizable geometric or topological structures. Building on this observation, we develop a method to address the challenges of open-set learning by focusing on the semantically important parts within 3D objects. Our proposed method strengthens the association between class semantics and object partial structures, ultimately improving the accuracy and robustness of 3D open-set recognition.", + "bbox": [ + 511, + 472, + 908, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 699, + 604, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture overview", + "text_level": 1, + "bbox": [ + 511, + 724, + 720, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As introduced in Sec. 1, we propose the Saliency-Aware Separation (SASep) method to address the limitations of current methods, which overlook the different contributions of object parts to classification. Figure 2 illustrates the architecture overview of our SASep. Building upon the basic pipeline of using an encoder to lift the input point cloud $P_{in}$ into the global feature $\\mathcal{F}_{in}$ , the closed-set model output $\\mathcal{Y}_{in} \\in \\mathbb{R}^{C}$ can be formulated as the softmax of $g(\\mathcal{F}_{in})$ where $g(\\cdot)$ maps the global feature to a set of logits of $C$ known classes. Rather than using the commonly adopted", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27297", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "linear layer to implement $g(\\cdot)$ , we formulate $\\mathcal{Y}_{in}$ by employing a prototype-based method [10] (see Eq. (8)), aiming to help capture important geometric characteristics of each class. Then, we calculate the confidence score $\\mathcal{Q}$ by extracting the maximum logit score (MLS) from $\\mathcal{Y}_{in}$ . Objects with a $\\mathcal{Q}$ value below a predefined threshold are classified as belonging to the unknown class.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance this basic pipeline, we first design a tunable semantic (TSD) module (marked by a green dash box) to evaluate the semantic saliency of each point, thereby dividing objects into important (high-saliency) and unimportant (low-saliency) parts based on the level of semantic saliency. For the high-saliency parts, we directly treat them as new instances of the same known class as the original objects and re-input them into the feature encoder to enhance the network's learning ability for this category. Note that, for the low-saliency parts that carry less known-class-specific information, we do not discard them. Instead, we develop a geometric synthesis strategy (GSS) to create unknown objects by combining low-saliency parts from various known objects. The rationale behind GSS lies in the fact that unknown- and known-class objects are likely to share similar local geometric structures. By re-inputting these synthetic objects into the network, the network's discriminative ability for similar regions can be enhanced.", + "bbox": [ + 89, + 198, + 482, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please note that the above designs are processed at the geometric level. Given the difficulty in accurately recognizing unknown classes, we take a further step by designing a synth-aided margin separation (SMS) module (marked by a blue dash box) that operates $\\mathcal{F}_{in}$ at the feature level, with the key idea of enlarging the feature distribution margins between classes. In the following sections, we will provide detailed information about each of our designs.", + "bbox": [ + 89, + 469, + 482, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tunable semantic decomposition module", + "text_level": 1, + "bbox": [ + 89, + 597, + 437, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The tunable semantic decomposition (TSD) module aims to decompose the objects into high-saliency parts and low-saliency parts, based on each point's contribution to classification predictions. The pipeline of our proposed TSD module is illustrated in Fig. 3. For the input point cloud $\\mathcal{P}_{in}$ , the model extracts its per-point features $\\mathcal{A}$ to predict the output $\\mathcal{Y}_{in}$ . Since $\\mathcal{Y}_{in}$ reflects the class probabilities for $\\mathcal{P}_{in}$ , we aim to estimate the contribution of each point to the corresponding object class probability in $\\mathcal{Y}_{in}$ . To achieve this goal, we adopt the Grad-CAM approach [34] to perform semantic evaluation, which aggregates each channel of the point-wise feature $\\mathcal{A}$ with associated gradient information to generate a saliency map $S$ :", + "bbox": [ + 89, + 619, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = R e L U \\left(\\sum_ {k} \\left(\\frac {1}{N} \\sum_ {i} \\frac {\\partial \\mathcal {Y} _ {i n} ^ {c}}{\\partial \\mathcal {A} _ {i} ^ {k}}\\right) \\mathcal {A} ^ {k}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 821, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_i^k$ denotes the $k$ -th channel of $\\mathcal{A}$ corresponding to the $i$ -th point and $N$ is the number of points. Generally,", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c6118ba15fd7a5d01856a6ff0d54aec440e86990ebe2a8631ded8a55912d1c57.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed tunable semantic decomposition (TSD) module. The TSD module decomposes the input objects into high- and low-saliency parts using the semantically evaluated saliency map $S$ . It also incorporates random partial views to enhance geometric information in a tunable manner." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a point with a higher saliency score in $S$ indicates that the point has a greater influence on the classification result, i.e., it is more important. Then, we sort $S$ in ascending order, designating the first $\\lfloor N / M \\rfloor$ points as the low-saliency part, while the remaining points are marked as the high-saliency part. Here, $M$ is a parameter that determines the number of low-saliency parts used in the geometric synthesis strategy.", + "bbox": [ + 511, + 361, + 903, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because the saliency map $\\mathcal{S}$ focuses on evaluating the semantic relevance, the saliency-based decomposition tends to generate relatively fixed decomposed parts. However, diverse scenarios, such as occlusion, noise, or corrupted data, are common in real-world point cloud data, which makes it difficult for the saliency-based decomposition method to effectively capture the key parts. To improve the robustness and generalization of the model in such challenging conditions, we introduce a geometric augmentation strategy to the TSD module. Specifically, by estimating visibility from view poses with random radii around the input points [29], we obtain the cropped parts of objects from these partial views. For each partial view, the saliency scores of each point are retrieved from $\\mathcal{S}$ Unlike simply augmenting the dataset by appending additional partial view data, we replace low- and high-saliency parts with partial views that have overall scores above $\\tau_h$ and below $\\tau_l$ ( $\\tau_h > \\tau_l$ ), respectively. By fine-tuning $\\tau_h$ and $\\tau_l$ , the TSD module can flexibly shift the decomposition strategy from semantically focused (when $\\tau_h$ is high and $\\tau_l$ is low) to geometrically focused (when $\\tau_h$ is low and $\\tau_l$ is high). This adaptability allows the TSD module to generate different samples that capture both semantic importance and geometric diversity.", + "bbox": [ + 511, + 469, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Geometric synthesis strategy", + "text_level": 1, + "bbox": [ + 511, + 830, + 769, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the scarcity of unknown class data during training, the geometric synthesis strategy (GSS) synthesizes pseudo-unknown samples by mixing multiple low-saliency", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27298", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "parts generated from the TSD module, thereby expanding the dataset to explore potential combinations of local structures that unknown classes may encompass. Unlike methods that synthesize data using random parts of known objects, which risk introducing class-specific geometric structures, GSS limits impact on classification by leveraging less influential low-saliency parts.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the low-saliency parts undergo standard transformations, followed by a mixing operation where $M$ randomly selected parts from different objects are combined. The mixed objects are then encoded to generate outputs $\\mathcal{V}_s$ , which are supervised using pseudo-labels $\\hat{\\mathcal{V}}_s$ set to $C + 1$ . To reduce overconfidence, label smoothing with base smooth weight $\\epsilon$ is applied to $\\hat{\\mathcal{V}}_s$ . Additional smooth weight $\\epsilon_h$ is also assigned to the involved object classes, ensuring the mixed objects retain characteristics of their originals. Assuming that the mixed parts are from classes $\\mathcal{K}_s = \\{c_1,\\dots ,c_n\\}$ , we define $\\hat{\\mathcal{V}}_s$ as:", + "bbox": [ + 89, + 196, + 482, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {Y}} _ {s} ^ {i} = \\left\\{ \\begin{array}{l l} 1 - \\epsilon - \\epsilon_ {h}, & i f i = C + 1 \\\\ \\epsilon / C, & i f i \\in \\mathcal {K} a n d i \\notin \\mathcal {K} _ {s} \\\\ \\epsilon / C + \\epsilon_ {h} M _ {c _ {i}} / M, & i f i \\in \\mathcal {K} _ {s}, \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 373, + 482, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{c_i}$ is the number of mixed objects belonging to $c_i$ . Finally, the loss of GSS is formulated as:", + "bbox": [ + 89, + 438, + 482, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = - \\sum_ {i = 1} ^ {C + 1} \\hat {\\mathcal {Y}} _ {s} ^ {i} \\log \\frac {e ^ {\\mathcal {Y} _ {s} ^ {i}}}{\\sum_ {j = 1} ^ {C + 1} e ^ {\\mathcal {Y} _ {s} ^ {j}}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 478, + 482, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{V}_s^i$ denotes the $i$ -th class probability of $\\mathcal{V}_s$ .", + "bbox": [ + 89, + 529, + 416, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Synth-aided margin separation module", + "text_level": 1, + "bbox": [ + 89, + 554, + 426, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The synth-aided margin separation (SMS) module is designed to improve the compactness within each class and the separability between different classes. It achieves this by applying feature constraints to individual object features. The key idea of SMS is illustrated in Fig. 4. We denote the global feature of the input point cloud $\\mathcal{P}_{c1}$ and its high-saliency part $\\mathcal{P}_h$ as $\\mathcal{F}_{c1}$ and $\\mathcal{F}_h$ , respectively. To highlight the semantically-relevant local geometric structures in $\\mathcal{P}_{c1}$ , we design a dual alignment strategy comprising attraction and repulsion forces. First, $\\mathcal{P}_h$ is designated as a positive exemplar, and its feature $\\mathcal{F}_h$ is aligned towards $\\mathcal{F}_{c1}$ to consolidate intra-class similarity. Concurrently, to enforce inter-class discrepancy, the negative exemplar $\\mathcal{P}_{c2}$ ( $c1 \\neq c2$ ) is pushed away via its feature $\\mathcal{F}_{c2}$ in the embedding space. By using both positive and negative exemplars, the feature constraint $\\mathcal{L}_m$ is defined through a triplet loss framework to ensure that $d(\\mathcal{F}_{c1}, \\mathcal{F}_h)$ is smaller than $d(\\mathcal{F}_{c1}, \\mathcal{F}_{c2})$ .", + "bbox": [ + 89, + 577, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {h}\\right) - d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {c 2}\\right) + \\tau_ {m}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 859, + 482, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ is the distance metric function.", + "bbox": [ + 89, + 885, + 370, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/94c1f375387e88583ffd2dda6e97e932dc802c0577717f8ed141c7ef22926bf1.jpg", + "image_caption": [ + "Figure 4. Illustration of our proposed synth-aided margin separation (SMS) module. The SMS module regularizes the feature space with a triplet loss, which employs the pseudo-features to help form the positive examples and negative examples." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 901, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although $\\mathcal{F}_h$ and $\\mathcal{F}_{c2}$ are appropriate for constructing feature constraints, their performance may be limited by data scarcity due to the sparse distribution of feature samples. To increase the availability of feature pairs, we generate synthetic features by adding Gaussian noise to $\\mathcal{F}_{c1}$ .", + "bbox": [ + 511, + 340, + 905, + 416 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {F} _ {s} = \\mathcal {F} _ {c 1} + \\mathcal {N} (0, \\mathcal {W}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 426, + 903, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W} = \\{w_1, \\dots, w_n\\}$ is a list of weights for Gaussian noise. Then, we filter $\\mathbb{F}_s$ to retain only the samples whose predicted results match the ground-truth labels. This step enhances the reliability of synthetic features. From these filtered samples, one is randomly selected to serve as the pseudo feature. This pseudo feature is then utilized for either random positive or negative replacements by substituting $\\mathcal{F}_h$ or $\\mathcal{F}_{c2}$ , resulting in the creation of new positive exemplars $\\mathcal{F}_p$ or negative exemplars $\\mathcal{F}_n$ . In the case of positive replacement, the distance $d(\\mathcal{F}_{c1}, \\mathcal{F}_h)$ is adjusted to accommodate a broader variability in $d(\\mathcal{F}_{c1}, \\mathcal{F}_p)$ , allowing for greater flexibility compared to the original distance. Conversely, negative replacement is designed to enhance the robustness of object features by introducing disturbances directly into the feature space, applying force on $\\mathcal{F}_{c1}$ with $d(\\mathcal{F}_{c1}, \\mathcal{F}_n)$ as highlighted in Eq. (4). It's important to note that only one of the positive or negative replacements will be executed for each input feature. Finally, the $\\mathcal{L}_m$ is refined with a weighted triplet loss:", + "bbox": [ + 511, + 453, + 906, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\rho d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {p}\\right) - \\eta d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {n}\\right) + \\tau_ {m}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 751, + 903, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho$ and $\\eta$ are the balance weights and $\\tau_{m}$ is the margin threshold. Given that the positive exemplar pairs are already constrained by the classification loss, we set $\\rho < \\eta$ to prioritize the enhancement of $d(\\mathcal{F}_{c1},\\mathcal{F}_n)$ .", + "bbox": [ + 511, + 777, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss function", + "text_level": 1, + "bbox": [ + 511, + 847, + 653, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss of SAsep integrates four components: (1) the primary classification loss $\\mathcal{L}_{cls}$ for raw inputs $\\mathcal{P}_{in}$ , (2) the", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27299", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "high-saliency loss $\\mathcal{L}_h$ for supervising TSD-generated high-saliency parts $\\mathcal{P}_h$ , (3) the synthesis loss $\\mathcal{L}_s$ (Eq. (3)) for synthetic samples from the GSS module, and (4) the feature margin loss $\\mathcal{L}_m$ (Eq. (6)). For $\\mathcal{P}_{in}$ of class $c$ , $\\mathcal{L}_{cls}$ applies cross-entropy to the model output $\\mathcal{Y}_{in}$ :", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\log \\frac {e ^ {\\mathcal {Y} _ {i n} ^ {c}}}{\\sum_ {i = 1} ^ {C + 1} e ^ {\\mathcal {Y} _ {i n} ^ {i}}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 176, + 482, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{Y}_{in}^{i}$ is the cosine similarity between the encoded feature $\\mathcal{F}_{in} \\in \\mathbb{R}^d$ of $\\mathcal{P}_{in}$ and learnable prototypes $\\mathcal{M} = \\{\\mathbf{m}_1, \\dots, \\mathbf{m}_C, \\mathbf{m}_{C+1}\\} \\in \\mathbb{R}^{(C+1) \\times d}$ , with $\\mathbf{m}_{C+1}$ for unknown classes. Hence, $\\mathcal{Y}_{in}^{i}$ is defined as follows:", + "bbox": [ + 89, + 227, + 483, + 289 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} _ {i n} ^ {i} = g \\left(\\mathcal {F} _ {i n}; \\mathbf {m} _ {i}\\right) = \\frac {\\mathcal {F} _ {i n} \\cdot \\mathbf {m} _ {i}}{\\| \\mathcal {F} _ {i n} \\| \\| \\mathbf {m} _ {i} \\|}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 299, + 482, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ denotes the $L2$ -norm. Note that the prototypes $\\mathcal{M}$ are initialized randomly and optimized during training. As aforementioned in Sec. 3.1, we consider TSD-generated high-saliency parts $\\mathcal{P}_h$ as new instances belonging to the same class as $\\mathcal{P}_{in}$ and directly feed $\\mathcal{P}_h$ into the model. Consequently, we compute $\\mathcal{L}_h$ for $\\mathcal{P}_h$ using the same formulation outlined in Eq. (7). The overall open-set classification loss $\\mathcal{L}$ is thus defined as:", + "bbox": [ + 89, + 342, + 483, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {h} + \\beta \\mathcal {L} _ {s} + \\gamma \\mathcal {L} _ {m}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 477, + 482, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta$ and $\\gamma$ are balance parameters.", + "bbox": [ + 89, + 505, + 367, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 535, + 223, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 560, + 192, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments using the benchmarks proposed by 3DOS [1], which are based on three widely used 3D object datasets: ShapeNetCore [3], ModelNet40 [45], and ScanObjectNN [38]. The ShapeNetCore dataset is a large synthetic 3D dataset that contains 51,127 3D mesh instances across 55 object categories. Approximately $70\\%$ of the data is allocated for training, $10\\%$ for validation, and $20\\%$ for testing. The ModelNet40 dataset consists of 12,311 3D CAD-generated meshes from 40 categories. This dataset is divided into an official training set of 9,843 samples and a test set of 2,468 samples. The corresponding point cloud data points are uniformly sampled from the surfaces of the CAD meshes and scaled to fit with a unit sphere. ScanObjectNN is a real-world dataset comprising 2,902 3D scans across 15 categories. It presents significant challenges due to artifacts, such as background clutter, occlusions, missing parts, and non-uniform density.", + "bbox": [ + 89, + 583, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the original experimental settings for open-set 3D recognition defined by 3DOS benchmarks, which include three tracks designed to evaluate the performance under different conditions of semantic and domain shifts.", + "bbox": [ + 89, + 840, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the synthetic benchmark, the ShapeNetCore dataset is divided into three distinct category subgroups, designated as SN1, SN2 and SN3, each containing 18 classes with no overlap. Each group rotates as the set of known classes, while the other two are treated as unknown classes, creating three unique experimental configurations. Similarly, the real-to-real benchmark follows the ScanObjectNN dataset, which is also divided into three subgroups (RR1, RR2, and RR3) to create three corresponding experimental configurations under real-scanned conditions. The synthetic-to-real benchmark establishes cross-domain scenarios (SR1, SR2), wherein synthetic samples from ModelNet40 serve as the known class set, while real scans from ScanObjectNN are designated as the unknown class set. Given the significant domain shifts between the known and unknown categories, this synthetic-to-real benchmark presents greater challenges compared to the other two benchmarks.", + "bbox": [ + 511, + 90, + 906, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation metrics", + "text_level": 1, + "bbox": [ + 511, + 364, + 694, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open-set 3D recognition includes two main tasks: closed-set classification and identification of unknown samples. For the closed-set classification task, we utilize accuracy (ACC) and mean accuracy (mACC) to evaluate the model's performance in correctly classifying known samples. To assess performance in identifying unknown samples, we utilize AUROC and FPR95 metrics. The AUROC (Area Under the Receiver Operating Characteristic Curve) measures a model's ability to differentiate between known and unknown samples. The ROC curve plots the true positive rate against the false positive rate across various thresholds to reject unknown samples. A higher AUROC value indicates better performance in identifying unknown samples. FPR95(False Positive Rate at $95\\%$ True Positive Rate) measures the rate of false positives while keeping the true positive rate fixed at $95\\%$ . This metric offers insights into the model's capacity in reducing misclassifications, particularly in preventing the incorrect identification of unknown samples as known.", + "bbox": [ + 511, + 388, + 906, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 694, + 725, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed SASep method is implemented using the PyTorch framework and evaluated with DGCNN [40] and PointNet++ [32]. To simplify the training process, saliency maps for each object were generated using Grad-CAM [34] with the pre-trained closed-set backbones. All models are trained on 4 NVIDIA RTX 3090 GPUs for 200 epochs, with a total batch size of 120. For both DGCNN and PointNet++, we adopt the Adam optimizer and CosLR scheduler with an initial learning rate set to 0.001. The balance parameters $\\alpha$ , $\\beta$ , and $\\gamma$ in Eq. (9) are assigned to 0.1, 0.01, and 0.3, respectively. In Eq. (6), the weights $\\rho$ and $\\eta$ are set to 0.01 and 1.0, respectively, with $\\tau_{m}$ fixed at 10.", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27300", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/732e334ad88cb5554969d6f063e55d5d5d6be96db0edf94b519461e4eb747054.jpg", + "table_caption": [ + "Table 1. The results of unknown sample identification task on the synthetic, synthetic-to-real, and real-to-real benchmarks. Each column shows the average results of each benchmark's subgroups. The best results are in bold in each metric." + ], + "table_footnote": [], + "table_body": "
MethodsDGCNNPointNet++
SyntheticSynthetic-to-RealReal-to-RealSyntheticSynthetic-to-RealReal-to-Real
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
MSP [13]85.263.266.790.670.882.881.371.075.683.280.877.2
MLS [14]86.251.865.790.572.677.381.961.874.881.783.067.4
ODIN [24]86.351.365.790.672.677.380.564.376.080.883.365.5
Energy [41]86.251.765.690.872.678.182.061.874.882.482.667.3
GradNorm [16]72.976.663.491.567.284.562.683.373.083.280.467.6
ReAct [36]88.443.965.690.575.175.386.152.974.681.482.767.4
VAE [27]77.064.263.384.6--------
NF88.546.771.382.370.282.181.266.776.484.381.465.8
OE+mixup [22]85.256.565.390.869.885.080.169.965.791.669.086.6
PointCaM [15]--71.186.6--------
ARPL+CS [4]84.473.667.189.8--80.367.875.482.1--
Cosine proto [10]86.548.957.991.078.068.286.955.778.276.183.664.9
CE (L2)89.143.666.189.274.974.488.644.277.782.379.175.8
SupCon [18]84.364.5----83.262.9----
SubArcFace [7]89.348.071.686.776.772.983.758.176.983.879.973.2
Ours92.528.777.975.482.062.392.829.179.767.584.855.0
", + "bbox": [ + 94, + 128, + 901, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison on unknown sample identification", + "text_level": 1, + "bbox": [ + 89, + 440, + 480, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our proposed SASep method with various existing open-set recognition methods, including both 2D and 3D techniques, across all tracks on synthetic, synth-to-real, and real-to-real benchmarks. The results of each benchmark's subgroups are aggregated to provide a comprehensive evaluation. As shown in Tab. 1, our proposed method consistently demonstrates the highest AUROC scores and the lowest FPR95 scores across all benchmarks and backbones. Specifically, SASep surpasses the best-performing methods in each benchmark with AUROC and FPR95 improvements of at least $2\\%$ and $6\\%$ , respectively. Notably, SASep achieves the highest improvements in the AUROC and FPR95 metrics on synthetic-to-real and synthetic benchmarks, with gains of $6.3\\%$ and $15.1\\%$ , respectively. Furthermore, SASep demonstrates higher improvements in the synthetic benchmark compared to the other benchmarks. We believe this is due to the characteristics of the synthetic data, which has more complete geometric structures and less real-world noise. This allows SASep to effectively capture the class-relevant structures, leading to improved performance. Apart from the synthetic benchmark, the improvements observed in the synthetic-to-real and real-to-real benchmarks further highlight the robustness of our method in addressing the challenges presented by the cross-domain gap between known and unknown categories.", + "bbox": [ + 88, + 470, + 482, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To provide a more intuitive comparison of our method with other representative methods, we visualize the AUROC and FPR95 metrics for both DGCNN and PointNet++", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "across all subgroups of each benchmark using radar charts, as shown in Fig. 5. Clearly, our proposed method (highlighted in orange) outperforms other methods in nearly all experiments, demonstrating its superior performance. Please refer to the Supplemental Material for more detailed experimental results.", + "bbox": [ + 511, + 441, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison on closed-set classification", + "text_level": 1, + "bbox": [ + 511, + 544, + 848, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the closed-set classification performance of our method against baseline closed-set models with DGCNN and PointNet++ backbones. The closed-set models are trained and tested solely on known classes. Although our method further performs the unknown sample identification task, we only compare the classification accuracy. Tab. 2 shows the accuracy (ACC) of each method across three subgroups: SN1, SR1, and RR1. For the DGCNN backbone, our method achieves improved performance in SR1 and RR1, reaching $81.3\\%$ and $94\\%$ , respectively, compared to the closed-set model's $72.1\\%$ and $91.5\\%$ , with a slight decrease in SN1. The improvements observed in SR1 and RR1 demonstrate the enhanced robustness of our method when applied to real-world scan data. This can be attributed to our design, which focuses on capturing semantically important geometric structures, thereby improving the model's resilience to real-world scenarios. For the PointNet++ backbone, our method achieves results comparable to the closed-set model, with only a slight decrease in performance. This may be due to PointNet++'s stronger capacity for feature extraction using hierarchical network structures. Overall, these results suggest that our method not only competes", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da296c652ce4e3a39ea2cea4e0085a5da61d0bf7c94e03d9e0997a2d82bc2444.jpg", + "image_caption": [ + "Figure 5. Comparisons on all the subgroups of synthetic, synthetic-to-real, real-to-real benchmarks. The AUROC and FPR95 metrics on the DGCNN and PointNet++ are drawn in the radar chart (a)-(d). The results of our proposed SASEp are highlighted in orange." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0abdbf831b52d4f774894afe98da23b0f9704c1a336f7b863523a44c0af3ce8c.jpg", + "table_caption": [ + "Table 2. Results of closed-set classification task on the subgroup (SN1, SR1, RR1) of synthetic, synthetic-to-real, and real-to-real benchmarks. The ACC metric is used as the evaluation metric." + ], + "table_footnote": [], + "table_body": "
MethodsDGCNNPointNet++
SN1SR1RR1SN1SR1RR1
closed-set84.772.191.585.980.095.6
Our83.381.394.085.480.094.8
", + "bbox": [ + 122, + 382, + 450, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with closed-set models in terms of accuracy but also provides increased robustness against real-world challenges.", + "bbox": [ + 89, + 474, + 482, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 513, + 241, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the contribution of each component in our proposed method, we conduct ablation studies on SN2 of the synthetic benchmark with DGCNN backbone. As shown in Tab. 3, we report the AUROC, FPR95, ACC, and mACC metrics across various configurations.", + "bbox": [ + 89, + 537, + 482, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When all modules (TSD, GSS, SMS) are enabled, the model achieves its best performance, with an AUROC of 94.8, the lowest FPR95 at 24.0, and the highest ACC and mACC at 90.5 and 79.8, respectively. This indicates that the full combination maximizes performance for both known class classification and unknown class identification tasks.", + "bbox": [ + 89, + 613, + 483, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Removing each module clarifies specific strengths. Excluding TSD (with only GSS and SMS active) leads to a decreased AUROC and mACC, demonstrating TSD's significant role in both tasks of open-set learning. When GSS is removed, the AUROC slightly increases, but the FPR95 drops to 29.1, indicating GSS's key role in controlling false positives. Without SMS, all the metrics are slightly decreased, suggesting SMS's impact on overall accuracy.", + "bbox": [ + 89, + 704, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Individual module analysis highlights TSD's strength, achieving an AUROC of 94.2 and the lowest FPR95 of 25.9. GSS alone enhances mACC but raises FPR95 to 30.6, while SMS alone has the highest FPR95, indicating its lower effectiveness in isolation.", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7c0c5c40786d50b33157f5b8dbba047e58d07b6cce431b0263d879dfd3e45efa.jpg", + "table_caption": [ + "Table 3. Comparing the contribution of major components in our proposed SASep on SN2 of synthetic benchmark using DGCNN." + ], + "table_footnote": [], + "table_body": "
TSDGSSSMSAUROC↑FPR95↓ACC↑mACC↑
✓✓93.728.388.374.0
✓✓94.129.189.978.4
✓✓93.826.989.278.5
✓94.225.990.077.6
✓93.630.688.679.0
✓93.036.189.675.9
✓✓✓94.824.090.579.8
", + "bbox": [ + 517, + 369, + 903, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 523, + 633, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel open-set point cloud learning method, SASep, which identifies unknown classes by emphasizing the importance of utilizing the geometric structures of objects' different parts, rather than treating them as a whole. Specifically, we design a tunable semantic decomposition module to decompose objects into the semantically important and unimportant parts for classification. Then, we introduce a geometric synthesis strategy and a synth-aided margin separation module to leverage these decomposed parts, improving feature learning for known classes and enhancing class distinction. Experimental results show that SASep significantly outperforms the state-of-the-art method in various datasets. In scenarios with severe corruption, our method may be limited due to challenges in accurately capturing semantics, motivating us to explore a more robust method for semantic decomposition. We hope SASep offers some new insights into this under-explored area.", + "bbox": [ + 509, + 549, + 906, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 833, + 681, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Key Research and Development Program for Young Scientists of China (2024YFB3310100), the China National Nat", + "bbox": [ + 511, + 857, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27302", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ural Science Foundation No. 62202182, the Guangdong Basic and Applied Basic Research Foundation 2024A1515010224, the National Key Research and Development Program of China under Grant 2023YFB4503400, and the China National Natural Science Foundation Nos. 62176101, 62276109, 62450064, and 62322205.", + "bbox": [ + 89, + 90, + 485, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning - benchmarking and understanding semantic novelty detection on point clouds. In NeurIPS, pages 21228-21240. Curran Associates, Inc., 2022. 3, 6", + "[2] Ayush Bhardwaj, Sakshee Pimpale, Saurabh Kumar, and Biplab Banerjee. Empowering knowledge distillation via open set recognition for robust 3d point cloud classification. Pattern Recognition Letters, 151:172-179, 2021. 2, 3", + "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[4] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial Reciprocal Points Learning for Open Set Recognition. IEEE TPAMI, 44(11):8065-8081, 2022. 2, 3, 7", + "[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2", + "[7] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In ECCV, pages 741-757, Cham, 2020. Springer International Publishing. 7", + "[8] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In ICLR, 2022. 2", + "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 3", + "[10] Dario Fontanel, Fabio Cermelli, Massimiliano Mancini, and Barbara Caputo. Detecting anomalies in semantic segmentation with prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 113-121, 2021. 3, 4, 7", + "[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2", + "[12] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, page 4995-5004, New York, NY, USA, 2024. Association for Computing Machinery. 2" + ], + "bbox": [ + 93, + 229, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 7", + "[14] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In Proceedings of the 39th International Conference on Machine Learning, pages 8759-8773. PMLR, 2022. 1, 2, 7", + "[15] Jie Hong, Shi Qiu, Weihao Li, Saeed Anwar, Mehrtash Harandi, Nick Barnes, and Lars Petersson. Pointcam: Cut-and-mix for open-set point cloud learning. arXiv preprint arXiv:2212.02011, 2022. 3, 7", + "[16] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, pages 677-689. Curran Associates, Inc., 2021. 2, 7", + "[17] Inhyuk Jo, Jungtaek Kim, Hyohyeong Kang, Yong-Deok Kim, and Seungjin Choi. Open set recognition by regularising classifier with fake data generated by generative adversarial networks. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2686–2690, 2018. 2", + "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, pages 18661-18673. Curran Associates, Inc., 2020. 7", + "[19] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, pages 813-822, 2021. 2", + "[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 2", + "[21] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2", + "[22] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3, 7", + "[23] Huan Lei, Naveed Akhtar, and Ajmal Mian. Octree guided cnn with spherical kernels for 3d point clouds. In CVPR, 2019. 2", + "[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 7", + "[25] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In CVPR, pages 20402-20411, 2023. 3", + "[26] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 2", + "[27] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In ICIP, pages 3118-3122, 2021. 2, 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928, 2015. 2", + "[29] Ravish Mehra, Pushkar Tripathi, Alla Sheffer, and Niloy J Mitra. Visibility of noisy point cloud data. Computers & Graphics, 34(3):219-230, 2010. 4", + "[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 3", + "[31] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2", + "[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6", + "[33] Paolo Rabino, Antonio Alliegro, and Tatiana Tommasi. 3d semantic novelty detection via large-scale pre-trained models. IEEE Access, 12:135352-135361, 2024. 3", + "[34] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4, 6", + "[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2", + "[36] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157. Curran Associates, Inc., 2021. 2, 7", + "[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2", + "[38] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6", + "[39] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4), 2017. 2", + "[40] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 2, 6", + "[41] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In ICCV, pages 9302-9311, 2021. 2, 7", + "[42] Yu Wang, Junxian Mu, Pengfei Zhu, and Qinghua Hu. Exploring diverse representations for open set recognition. AAAI, 38(6):5731-5739, 2024. 2", + "[43] Tingyu Weng, Jun Xiao, Hao Pan, and Haiyong Jiang. Partcom: Part composition learning for 3d open-set recognition. IJCV, 132(4):1393-1416, 2024. 2, 3", + "[44] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6", + "[46] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1" + ], + "bbox": [ + 514, + 92, + 903, + 189 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27304", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_model.json b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7401a36b87e2de9ab1cc3da12527b350ba14044f --- /dev/null +++ b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_model.json @@ -0,0 +1,1936 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.865, + 0.177 + ], + "angle": 0, + "content": "SASep: Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.203, + 0.899, + 0.222 + ], + "angle": 0, + "content": "Jinfeng Xu\\(^{1}\\) Xianzhi Li\\(^{1,2}\\) Yuan Tang\\(^{1}\\) Xu Han\\(^{1}\\) Qiao Yu\\(^{1}\\) Yixue Hao\\(^{1}\\) Long Hu\\(^{1*}\\) Min Chen\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.222, + 0.516, + 0.258 + ], + "angle": 0, + "content": "1Huazhong University of Science and Technology \n3South China University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.222, + 0.866, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Guangdong Intelligent Robotics Institute \n\\(^{4}\\)Pazhou Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.26, + 0.871, + 0.275 + ], + "angle": 0, + "content": "{jinfengx, xzli, yuan_tang, xhanxu, qiaoyu_epic, yixuehao, hulong, minchen}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Recent advancements in deep learning have greatly enhanced 3D object recognition, but most models are limited to closed-set scenarios, unable to handle unknown samples in real-world applications. Open-set recognition (OSR) addresses this limitation by enabling models to both classify known classes and identify novel classes. However, current OSR methods rely on global features to differentiate known and unknown classes, treating the entire object uniformly and overlooking the varying semantic importance of its different parts. To address this gap, we propose Salience-Aware Structured Separation (SASep), which includes (i) a tunable semantic decomposition (TSD) module to semantically decompose objects into important and unimportant parts, (ii) a geometric synthesis strategy (GSS) to generate pseudo-unknown objects by combining these unimportant parts, and (iii) a synth-aided margin separation (SMS) module to enhance feature-level separation by expanding the feature distributions between classes. Together, these components improve both geometric and feature representations, enhancing the model's ability to effectively distinguish known and unknown classes. Experimental results show that SASep achieves superior performance in 3D OSR, outperforming existing state-of-the-art methods. The codes are available at https://github.com/JinfengX/SASep." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.732, + 0.222, + 0.748 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In recent years, deep learning has significantly advanced 3D object recognition, enabling its application in diverse areas such as autonomous driving, robotics, augmented/virtual reality, etc. Most existing 3D recognition models [31, 44, 46] are designed for closed-set scenarios, where all test classes are assumed to have been seen during training. However, real-world applications usually present dynamic environments where unknown or unseen samples outside the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.307, + 0.907, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Figure 1. Given input objects, the TSD module semantically decomposes them into important and unimportant parts based on each point's semantic relevance to classification. The important parts enhance feature learning for known classes, while the unimportant parts are mixed in GSS to generate pseudo-unknown samples. The SMS module then constrains the features of objects and their decomposed parts to improve inter-class separation and intraclass compactness." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.734, + 0.907, + 0.87 + ], + "angle": 0, + "content": "set of known classes may be encountered. These out-of-distribution samples pose a significant challenge, as the closed-set models are unable to handle these unknown samples appropriately, often misclassifying them as one of the predefined known classes. To address this challenge, open-set recognition (OSR) task has gained increasing attention recently, as it aims to extend the capabilities of closed-set models by enabling them to classify known classes and identify unknown ones simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Early 2D OSR approaches [13, 14, 24] aim to distinguish between known and unknown classes by using discrimina" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27295" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "tive functions to compute confidence scores from network predictions for each input. Samples are identified as unknown when their confidence scores are low. Thus, designing a robust open-set solver hinges on ensuring the confidence score is discriminative, meaning scores for known classes should clearly exceed those for unknown classes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.485, + 0.442 + ], + "angle": 0, + "content": "Although recent 2D OSR works [4, 36, 42] have achieved significant progress in the 2D area, research on OSR for 3D data remains relatively underdeveloped. Unlike 2D OSR, the unique spatial characteristics of 3D data limit the performance of directly adapting 2D OSR methods to 3D, as they fall short of leveraging 3D-specific structural and topological information. Recent works [2, 27, 43] propose to address the 3D OSR task from the perspective of geometric reconstruction, feature embeddings, and sample synthesis. However, these approaches, similar to many 2D OSR methods, primarily focus on assigning confidence scores based on whole-object features, often overlooking the fact that different parts of an object contribute differently to classification. For example, specific regions, such as the wheels of a car or the wings of an airplane, carry essential features that are more informative for classification than other parts." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.446, + 0.485, + 0.837 + ], + "angle": 0, + "content": "Based on this observation, we propose a novel method for 3D open-set recognition on point clouds, named Salience-Aware Structured Separation (SASep), which tackles the open-set point cloud learning task by highlighting the semantically important geometric structures of objects. As illustrated in Fig. 1, SASep first decomposes the input objects into important and unimportant parts with a tunable semantic decomposition (TSD) module based on the relevance of each region to the class semantics. Then, we augment the feature learning of known classes by incorporating important parts into the training dataset. For the unimportant parts, we design a geometric synthesis strategy (GSS) to generate synthetic samples, serving as pseudo-unknown objects to simulate geometric structures of unknown classes. Furthermore, we refine the feature space with synthie-aided margin separation module (SMS) by applying constraints that reduce the feature distance between the object and its important parts, while increasing the feature distance between the object and those from other classes. This process enhances the feature representations of known classes, thus reducing the likelihood of misclassifying unknown class samples as known classes. The experimental results demonstrate that our proposed method outperforms existing OSR approaches, achieving state-of-the-art (SOTA) performance in 3D open-set recognition. Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We introduce a novel approach for 3D open-set recognition on point clouds, called Salience-Aware Structured Separation (SASep), which enhances the distinction between known and unknown classes by utilizing the se" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.794, + 0.107 + ], + "angle": 0, + "content": "mantically decomposed parts of objects." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- We propose a novel semantic decomposition method TSD that decomposes an object into semantically important and unimportant regions, enabling the model to utilize structural information from different parts of the objects." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- We design GSS and SMS to utilize the decomposed parts of TSD to explore feature distribution of the unknown class and enhance feature distinction of the known classes, respectively." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.243, + 0.651, + 0.258 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.268, + 0.788, + 0.284 + ], + "angle": 0, + "content": "2.1. Closed-set point cloud learning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.908, + 0.625 + ], + "angle": 0, + "content": "Closed-set point cloud learning has made significant progress, focusing on accurate classification and segmentation of 3D objects from predefined categories. Existing approaches can be broadly categorized into view-based, voxel-based, and point-based methods. View-based methods [5, 21, 35, 37] classify 3D shapes by using 2D CNN to extract features from the projected 2D images of input. Voxel-based methods [6, 11, 23, 28, 39] process point clouds by dividing 3D space into voxel grids, allowing 3D CNNs to learn the spatial features. Pioneering point-based methods [31] use Multi-Layer Perceptrons (MLPs) with symmetric functions to directly process raw point cloud data. The following point-based research [12, 20, 26, 32, 40] further improved the performance by enhancing the models' capacity to aggregate the local and global features. While these methods have achieved significant success in recognizing known classes, they face notable limitations when dealing with unseen classes, as the models usually tend to misclassify the unknown objects as one of the predefined known classes. This limitation restricts the models' ability to generalize to real-world applications, where novel or unknown classes may be encountered during inference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.636, + 0.736, + 0.652 + ], + "angle": 0, + "content": "2.2. Open-set 2D recognition" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address the limitation of closed-set learning, open-set recognition (OSR) enables models to classify known objects and identify unknown or novel classes during inference. The straightforward approaches [16, 24, 36, 41] utilize discriminative functions to evaluate confidence scores for identifying unknown classes. MSP [13] established a robust baseline by using Maximum Softmax Probability to identify samples as unknown when they exhibit low maximum softmax scores. Similarly, Maximum Logit Score (MLS) [14] directly uses the highest logit score as a confidence measure. Beyond these discriminative-based approaches, reconstruction-based methods [8, 17, 19] identify unknown objects by comparing the original inputs with their reconstructions from the generative models. Recent research in 2D OSR aims to learn feature embeddings that form compact clusters for known classes. For instance," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.91, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.335, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Overall architecture of our SASep. The architecture includes a forward process where the input point cloud \\(\\mathcal{P}_{in}\\) is encoded with backbone, followed by a backward process utilizing the tunable semantic decomposition module to decompose \\(\\mathcal{P}_{in}\\) into high- and low-saliency parts. Then, the high-saliency parts are used in the synth-aided margin separation module for separating feature distribution; the low-saliency parts are combined for synthetic objects with a geometric synthesis strategy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.411, + 0.482, + 0.546 + ], + "angle": 0, + "content": "ARPL [4] introduces reciprocal points in the embedding space to minimize the overlap of known and unknown distributions. Fontanel et al. [10] proposed a method that uses learnable class prototypes to detect anomalies by measuring the cosine distance of inputs from learned class prototypes. Furthermore, several studies [9, 25, 30] have investigated using synthetic samples to expose potential unknown classes to the model, thereby enhancing the model's generalization on samples out of the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.639 + ], + "angle": 0, + "content": "Though 2D OSR has seen significant advancements, applying these methods to 3D domains presents limitations, particularly in leveraging the unique spatial and geometric complexities within 3D data. Thus, directly shifting a 2D OSR method to 3D data often fails to effectively distinguish known and unknown samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.315, + 0.667 + ], + "angle": 0, + "content": "2.3. Open-set 3D recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In contrast to the widely explored 2D OSR, research on 3D OSR remains relatively underdeveloped. Masuda et al. [27] developed a reconstruction-based method to obtain anomaly scores by measuring reconstruction errors between model input and output. To enhance the learning of distribution for unknown classes, Bhardwaj et al. [2] proposed generating pseudo-unknown samples by mixing objects from known classes. To further refine the synthesis strategy, Lee et al. [22] and Hong et al. [15] designed more sophisticated algorithms to synthesize high-quality pseudo unknown samples, considering the local geometric structure of objects. More recently, PartCom [43] introduced a novel part codebook to learn 3D object representations by capturing distinct part compositions across classes. Similarly, Rabino et al. [33] introduced a training-free method lever" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.472 + ], + "angle": 0, + "content": "aging part-based reasoning. Building on the work of 2D and 3D OSR, 3DOS [1] provided a comprehensive benchmark for 3D open-set learning, revealing the strengths and limitations of existing approaches in the 3D open-set context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Existing open-set recognition methods focus on distinguishing known and unknown objects or features as a whole, while overlooking the essential semantic information carried by different parts in an object. For example, the wheels of vehicles often exhibit more significant class-relevant information than the other parts. Humans can recognize objects even with a partial view of such recognizable geometric or topological structures. Building on this observation, we develop a method to address the challenges of open-set learning by focusing on the semantically important parts within 3D objects. Our proposed method strengthens the association between class semantics and object partial structures, ultimately improving the accuracy and robustness of 3D open-set recognition." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.7, + 0.605, + 0.715 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.722, + 0.74 + ], + "angle": 0, + "content": "3.1. Architecture overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As introduced in Sec. 1, we propose the Saliency-Aware Separation (SASep) method to address the limitations of current methods, which overlook the different contributions of object parts to classification. Figure 2 illustrates the architecture overview of our SASep. Building upon the basic pipeline of using an encoder to lift the input point cloud \\( P_{in} \\) into the global feature \\( \\mathcal{F}_{in} \\), the closed-set model output \\( \\mathcal{Y}_{in} \\in \\mathbb{R}^{C} \\) can be formulated as the softmax of \\( g(\\mathcal{F}_{in}) \\) where \\( g(\\cdot) \\) maps the global feature to a set of logits of \\( C \\) known classes. Rather than using the commonly adopted" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27297" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "linear layer to implement \\( g(\\cdot) \\), we formulate \\( \\mathcal{Y}_{in} \\) by employing a prototype-based method [10] (see Eq. (8)), aiming to help capture important geometric characteristics of each class. Then, we calculate the confidence score \\( \\mathcal{Q} \\) by extracting the maximum logit score (MLS) from \\( \\mathcal{Y}_{in} \\). Objects with a \\( \\mathcal{Q} \\) value below a predefined threshold are classified as belonging to the unknown class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.468 + ], + "angle": 0, + "content": "To enhance this basic pipeline, we first design a tunable semantic (TSD) module (marked by a green dash box) to evaluate the semantic saliency of each point, thereby dividing objects into important (high-saliency) and unimportant (low-saliency) parts based on the level of semantic saliency. For the high-saliency parts, we directly treat them as new instances of the same known class as the original objects and re-input them into the feature encoder to enhance the network's learning ability for this category. Note that, for the low-saliency parts that carry less known-class-specific information, we do not discard them. Instead, we develop a geometric synthesis strategy (GSS) to create unknown objects by combining low-saliency parts from various known objects. The rationale behind GSS lies in the fact that unknown- and known-class objects are likely to share similar local geometric structures. By re-inputting these synthetic objects into the network, the network's discriminative ability for similar regions can be enhanced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Please note that the above designs are processed at the geometric level. Given the difficulty in accurately recognizing unknown classes, we take a further step by designing a synth-aided margin separation (SMS) module (marked by a blue dash box) that operates \\(\\mathcal{F}_{in}\\) at the feature level, with the key idea of enlarging the feature distribution margins between classes. In the following sections, we will provide detailed information about each of our designs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.598, + 0.438, + 0.613 + ], + "angle": 0, + "content": "3.2. Tunable semantic decomposition module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.817 + ], + "angle": 0, + "content": "The tunable semantic decomposition (TSD) module aims to decompose the objects into high-saliency parts and low-saliency parts, based on each point's contribution to classification predictions. The pipeline of our proposed TSD module is illustrated in Fig. 3. For the input point cloud \\(\\mathcal{P}_{in}\\), the model extracts its per-point features \\(\\mathcal{A}\\) to predict the output \\(\\mathcal{Y}_{in}\\). Since \\(\\mathcal{Y}_{in}\\) reflects the class probabilities for \\(\\mathcal{P}_{in}\\), we aim to estimate the contribution of each point to the corresponding object class probability in \\(\\mathcal{Y}_{in}\\). To achieve this goal, we adopt the Grad-CAM approach [34] to perform semantic evaluation, which aggregates each channel of the point-wise feature \\(\\mathcal{A}\\) with associated gradient information to generate a saliency map \\(S\\):" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.823, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = R e L U \\left(\\sum_ {k} \\left(\\frac {1}{N} \\sum_ {i} \\frac {\\partial \\mathcal {Y} _ {i n} ^ {c}}{\\partial \\mathcal {A} _ {i} ^ {k}}\\right) \\mathcal {A} ^ {k}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_i^k\\) denotes the \\(k\\)-th channel of \\(\\mathcal{A}\\) corresponding to the \\(i\\)-th point and \\(N\\) is the number of points. Generally," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed tunable semantic decomposition (TSD) module. The TSD module decomposes the input objects into high- and low-saliency parts using the semantically evaluated saliency map \\( S \\). It also incorporates random partial views to enhance geometric information in a tunable manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.905, + 0.468 + ], + "angle": 0, + "content": "a point with a higher saliency score in \\( S \\) indicates that the point has a greater influence on the classification result, i.e., it is more important. Then, we sort \\( S \\) in ascending order, designating the first \\( \\lfloor N / M \\rfloor \\) points as the low-saliency part, while the remaining points are marked as the high-saliency part. Here, \\( M \\) is a parameter that determines the number of low-saliency parts used in the geometric synthesis strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.905, + 0.817 + ], + "angle": 0, + "content": "Because the saliency map \\(\\mathcal{S}\\) focuses on evaluating the semantic relevance, the saliency-based decomposition tends to generate relatively fixed decomposed parts. However, diverse scenarios, such as occlusion, noise, or corrupted data, are common in real-world point cloud data, which makes it difficult for the saliency-based decomposition method to effectively capture the key parts. To improve the robustness and generalization of the model in such challenging conditions, we introduce a geometric augmentation strategy to the TSD module. Specifically, by estimating visibility from view poses with random radii around the input points [29], we obtain the cropped parts of objects from these partial views. For each partial view, the saliency scores of each point are retrieved from \\(\\mathcal{S}\\) Unlike simply augmenting the dataset by appending additional partial view data, we replace low- and high-saliency parts with partial views that have overall scores above \\(\\tau_h\\) and below \\(\\tau_l\\) (\\(\\tau_h > \\tau_l\\)), respectively. By fine-tuning \\(\\tau_h\\) and \\(\\tau_l\\), the TSD module can flexibly shift the decomposition strategy from semantically focused (when \\(\\tau_h\\) is high and \\(\\tau_l\\) is low) to geometrically focused (when \\(\\tau_h\\) is low and \\(\\tau_l\\) is high). This adaptability allows the TSD module to generate different samples that capture both semantic importance and geometric diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.771, + 0.848 + ], + "angle": 0, + "content": "3.3. Geometric synthesis strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "To address the scarcity of unknown class data during training, the geometric synthesis strategy (GSS) synthesizes pseudo-unknown samples by mixing multiple low-saliency" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27298" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "parts generated from the TSD module, thereby expanding the dataset to explore potential combinations of local structures that unknown classes may encompass. Unlike methods that synthesize data using random parts of known objects, which risk introducing class-specific geometric structures, GSS limits impact on classification by leveraging less influential low-saliency parts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Specifically, the low-saliency parts undergo standard transformations, followed by a mixing operation where \\(M\\) randomly selected parts from different objects are combined. The mixed objects are then encoded to generate outputs \\(\\mathcal{V}_s\\), which are supervised using pseudo-labels \\(\\hat{\\mathcal{V}}_s\\) set to \\(C + 1\\). To reduce overconfidence, label smoothing with base smooth weight \\(\\epsilon\\) is applied to \\(\\hat{\\mathcal{V}}_s\\). Additional smooth weight \\(\\epsilon_h\\) is also assigned to the involved object classes, ensuring the mixed objects retain characteristics of their originals. Assuming that the mixed parts are from classes \\(\\mathcal{K}_s = \\{c_1,\\dots ,c_n\\}\\), we define \\(\\hat{\\mathcal{V}}_s\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.374, + 0.483, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {Y}} _ {s} ^ {i} = \\left\\{ \\begin{array}{l l} 1 - \\epsilon - \\epsilon_ {h}, & i f i = C + 1 \\\\ \\epsilon / C, & i f i \\in \\mathcal {K} a n d i \\notin \\mathcal {K} _ {s} \\\\ \\epsilon / C + \\epsilon_ {h} M _ {c _ {i}} / M, & i f i \\in \\mathcal {K} _ {s}, \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.469 + ], + "angle": 0, + "content": "where \\( M_{c_i} \\) is the number of mixed objects belonging to \\( c_i \\). Finally, the loss of GSS is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.479, + 0.483, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = - \\sum_ {i = 1} ^ {C + 1} \\hat {\\mathcal {Y}} _ {s} ^ {i} \\log \\frac {e ^ {\\mathcal {Y} _ {s} ^ {i}}}{\\sum_ {j = 1} ^ {C + 1} e ^ {\\mathcal {Y} _ {s} ^ {j}}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.418, + 0.547 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}_s^i\\) denotes the \\(i\\)-th class probability of \\(\\mathcal{V}_s\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.555, + 0.427, + 0.571 + ], + "angle": 0, + "content": "3.4. Synth-aided margin separation module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.851 + ], + "angle": 0, + "content": "The synth-aided margin separation (SMS) module is designed to improve the compactness within each class and the separability between different classes. It achieves this by applying feature constraints to individual object features. The key idea of SMS is illustrated in Fig. 4. We denote the global feature of the input point cloud \\(\\mathcal{P}_{c1}\\) and its high-saliency part \\(\\mathcal{P}_h\\) as \\(\\mathcal{F}_{c1}\\) and \\(\\mathcal{F}_h\\), respectively. To highlight the semantically-relevant local geometric structures in \\(\\mathcal{P}_{c1}\\), we design a dual alignment strategy comprising attraction and repulsion forces. First, \\(\\mathcal{P}_h\\) is designated as a positive exemplar, and its feature \\(\\mathcal{F}_h\\) is aligned towards \\(\\mathcal{F}_{c1}\\) to consolidate intra-class similarity. Concurrently, to enforce inter-class discrepancy, the negative exemplar \\(\\mathcal{P}_{c2}\\) (\\(c1 \\neq c2\\)) is pushed away via its feature \\(\\mathcal{F}_{c2}\\) in the embedding space. By using both positive and negative exemplars, the feature constraint \\(\\mathcal{L}_m\\) is defined through a triplet loss framework to ensure that \\(d(\\mathcal{F}_{c1}, \\mathcal{F}_h)\\) is smaller than \\(d(\\mathcal{F}_{c1}, \\mathcal{F}_{c2})\\)." + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.86, + 0.483, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {h}\\right) - d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {c 2}\\right) + \\tau_ {m}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.371, + 0.902 + ], + "angle": 0, + "content": "where \\(d(\\cdot)\\) is the distance metric function." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.903, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Figure 4. Illustration of our proposed synth-aided margin separation (SMS) module. The SMS module regularizes the feature space with a triplet loss, which employs the pseudo-features to help form the positive examples and negative examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.906, + 0.417 + ], + "angle": 0, + "content": "Although \\(\\mathcal{F}_h\\) and \\(\\mathcal{F}_{c2}\\) are appropriate for constructing feature constraints, their performance may be limited by data scarcity due to the sparse distribution of feature samples. To increase the availability of feature pairs, we generate synthetic features by adding Gaussian noise to \\(\\mathcal{F}_{c1}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.428, + 0.905, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathbb {F} _ {s} = \\mathcal {F} _ {c 1} + \\mathcal {N} (0, \\mathcal {W}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.742 + ], + "angle": 0, + "content": "where \\(\\mathcal{W} = \\{w_1, \\dots, w_n\\}\\) is a list of weights for Gaussian noise. Then, we filter \\(\\mathbb{F}_s\\) to retain only the samples whose predicted results match the ground-truth labels. This step enhances the reliability of synthetic features. From these filtered samples, one is randomly selected to serve as the pseudo feature. This pseudo feature is then utilized for either random positive or negative replacements by substituting \\(\\mathcal{F}_h\\) or \\(\\mathcal{F}_{c2}\\), resulting in the creation of new positive exemplars \\(\\mathcal{F}_p\\) or negative exemplars \\(\\mathcal{F}_n\\). In the case of positive replacement, the distance \\(d(\\mathcal{F}_{c1}, \\mathcal{F}_h)\\) is adjusted to accommodate a broader variability in \\(d(\\mathcal{F}_{c1}, \\mathcal{F}_p)\\), allowing for greater flexibility compared to the original distance. Conversely, negative replacement is designed to enhance the robustness of object features by introducing disturbances directly into the feature space, applying force on \\(\\mathcal{F}_{c1}\\) with \\(d(\\mathcal{F}_{c1}, \\mathcal{F}_n)\\) as highlighted in Eq. (4). It's important to note that only one of the positive or negative replacements will be executed for each input feature. Finally, the \\(\\mathcal{L}_m\\) is refined with a weighted triplet loss:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.752, + 0.905, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\rho d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {p}\\right) - \\eta d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {n}\\right) + \\tau_ {m}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.84 + ], + "angle": 0, + "content": "where \\(\\rho\\) and \\(\\eta\\) are the balance weights and \\(\\tau_{m}\\) is the margin threshold. Given that the positive exemplar pairs are already constrained by the classification loss, we set \\(\\rho < \\eta\\) to prioritize the enhancement of \\(d(\\mathcal{F}_{c1},\\mathcal{F}_n)\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.655, + 0.862 + ], + "angle": 0, + "content": "3.5. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The total loss of SAsep integrates four components: (1) the primary classification loss \\(\\mathcal{L}_{cls}\\) for raw inputs \\(\\mathcal{P}_{in}\\), (2) the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "27299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "high-saliency loss \\(\\mathcal{L}_h\\) for supervising TSD-generated high-saliency parts \\(\\mathcal{P}_h\\), (3) the synthesis loss \\(\\mathcal{L}_s\\) (Eq. (3)) for synthetic samples from the GSS module, and (4) the feature margin loss \\(\\mathcal{L}_m\\) (Eq. (6)). For \\(\\mathcal{P}_{in}\\) of class \\(c\\), \\(\\mathcal{L}_{cls}\\) applies cross-entropy to the model output \\(\\mathcal{Y}_{in}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.178, + 0.483, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\log \\frac {e ^ {\\mathcal {Y} _ {i n} ^ {c}}}{\\sum_ {i = 1} ^ {C + 1} e ^ {\\mathcal {Y} _ {i n} ^ {i}}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.228, + 0.485, + 0.29 + ], + "angle": 0, + "content": "where \\(\\mathcal{Y}_{in}^{i}\\) is the cosine similarity between the encoded feature \\(\\mathcal{F}_{in} \\in \\mathbb{R}^d\\) of \\(\\mathcal{P}_{in}\\) and learnable prototypes \\(\\mathcal{M} = \\{\\mathbf{m}_1, \\dots, \\mathbf{m}_C, \\mathbf{m}_{C+1}\\} \\in \\mathbb{R}^{(C+1) \\times d}\\), with \\(\\mathbf{m}_{C+1}\\) for unknown classes. Hence, \\(\\mathcal{Y}_{in}^{i}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.3, + 0.483, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} _ {i n} ^ {i} = g \\left(\\mathcal {F} _ {i n}; \\mathbf {m} _ {i}\\right) = \\frac {\\mathcal {F} _ {i n} \\cdot \\mathbf {m} _ {i}}{\\| \\mathcal {F} _ {i n} \\| \\| \\mathbf {m} _ {i} \\|}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.464 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) denotes the \\( L2 \\)-norm. Note that the prototypes \\( \\mathcal{M} \\) are initialized randomly and optimized during training. As aforementioned in Sec. 3.1, we consider TSD-generated high-saliency parts \\( \\mathcal{P}_h \\) as new instances belonging to the same class as \\( \\mathcal{P}_{in} \\) and directly feed \\( \\mathcal{P}_h \\) into the model. Consequently, we compute \\( \\mathcal{L}_h \\) for \\( \\mathcal{P}_h \\) using the same formulation outlined in Eq. (7). The overall open-set classification loss \\( \\mathcal{L} \\) is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.478, + 0.483, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {h} + \\beta \\mathcal {L} _ {s} + \\gamma \\mathcal {L} _ {m}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.506, + 0.368, + 0.521 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta\\) and \\(\\gamma\\) are balance parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.224, + 0.552 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.194, + 0.574 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.84 + ], + "angle": 0, + "content": "We conduct experiments using the benchmarks proposed by 3DOS [1], which are based on three widely used 3D object datasets: ShapeNetCore [3], ModelNet40 [45], and ScanObjectNN [38]. The ShapeNetCore dataset is a large synthetic 3D dataset that contains 51,127 3D mesh instances across 55 object categories. Approximately \\(70\\%\\) of the data is allocated for training, \\(10\\%\\) for validation, and \\(20\\%\\) for testing. The ModelNet40 dataset consists of 12,311 3D CAD-generated meshes from 40 categories. This dataset is divided into an official training set of 9,843 samples and a test set of 2,468 samples. The corresponding point cloud data points are uniformly sampled from the surfaces of the CAD meshes and scaled to fit with a unit sphere. ScanObjectNN is a real-world dataset comprising 2,902 3D scans across 15 categories. It presents significant challenges due to artifacts, such as background clutter, occlusions, missing parts, and non-uniform density." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.9 + ], + "angle": 0, + "content": "We follow the original experimental settings for open-set 3D recognition defined by 3DOS benchmarks, which include three tracks designed to evaluate the performance under different conditions of semantic and domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.349 + ], + "angle": 0, + "content": "In the synthetic benchmark, the ShapeNetCore dataset is divided into three distinct category subgroups, designated as SN1, SN2 and SN3, each containing 18 classes with no overlap. Each group rotates as the set of known classes, while the other two are treated as unknown classes, creating three unique experimental configurations. Similarly, the real-to-real benchmark follows the ScanObjectNN dataset, which is also divided into three subgroups (RR1, RR2, and RR3) to create three corresponding experimental configurations under real-scanned conditions. The synthetic-to-real benchmark establishes cross-domain scenarios (SR1, SR2), wherein synthetic samples from ModelNet40 serve as the known class set, while real scans from ScanObjectNN are designated as the unknown class set. Given the significant domain shifts between the known and unknown categories, this synthetic-to-real benchmark presents greater challenges compared to the other two benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.365, + 0.696, + 0.38 + ], + "angle": 0, + "content": "4.2. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.678 + ], + "angle": 0, + "content": "Open-set 3D recognition includes two main tasks: closed-set classification and identification of unknown samples. For the closed-set classification task, we utilize accuracy (ACC) and mean accuracy (mACC) to evaluate the model's performance in correctly classifying known samples. To assess performance in identifying unknown samples, we utilize AUROC and FPR95 metrics. The AUROC (Area Under the Receiver Operating Characteristic Curve) measures a model's ability to differentiate between known and unknown samples. The ROC curve plots the true positive rate against the false positive rate across various thresholds to reject unknown samples. A higher AUROC value indicates better performance in identifying unknown samples. FPR95(False Positive Rate at \\(95\\%\\) True Positive Rate) measures the rate of false positives while keeping the true positive rate fixed at \\(95\\%\\). This metric offers insights into the model's capacity in reducing misclassifications, particularly in preventing the incorrect identification of unknown samples as known." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.727, + 0.71 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our proposed SASep method is implemented using the PyTorch framework and evaluated with DGCNN [40] and PointNet++ [32]. To simplify the training process, saliency maps for each object were generated using Grad-CAM [34] with the pre-trained closed-set backbones. All models are trained on 4 NVIDIA RTX 3090 GPUs for 200 epochs, with a total batch size of 120. For both DGCNN and PointNet++, we adopt the Adam optimizer and CosLR scheduler with an initial learning rate set to 0.001. The balance parameters \\(\\alpha\\), \\(\\beta\\), and \\(\\gamma\\) in Eq. (9) are assigned to 0.1, 0.01, and 0.3, respectively. In Eq. (6), the weights \\(\\rho\\) and \\(\\eta\\) are set to 0.01 and 1.0, respectively, with \\(\\tau_{m}\\) fixed at 10." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27300" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. The results of unknown sample identification task on the synthetic, synthetic-to-real, and real-to-real benchmarks. Each column shows the average results of each benchmark's subgroups. The best results are in bold in each metric." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.13, + 0.903, + 0.415 + ], + "angle": 0, + "content": "
MethodsDGCNNPointNet++
SyntheticSynthetic-to-RealReal-to-RealSyntheticSynthetic-to-RealReal-to-Real
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
MSP [13]85.263.266.790.670.882.881.371.075.683.280.877.2
MLS [14]86.251.865.790.572.677.381.961.874.881.783.067.4
ODIN [24]86.351.365.790.672.677.380.564.376.080.883.365.5
Energy [41]86.251.765.690.872.678.182.061.874.882.482.667.3
GradNorm [16]72.976.663.491.567.284.562.683.373.083.280.467.6
ReAct [36]88.443.965.690.575.175.386.152.974.681.482.767.4
VAE [27]77.064.263.384.6--------
NF88.546.771.382.370.282.181.266.776.484.381.465.8
OE+mixup [22]85.256.565.390.869.885.080.169.965.791.669.086.6
PointCaM [15]--71.186.6--------
ARPL+CS [4]84.473.667.189.8--80.367.875.482.1--
Cosine proto [10]86.548.957.991.078.068.286.955.778.276.183.664.9
CE (L2)89.143.666.189.274.974.488.644.277.782.379.175.8
SupCon [18]84.364.5----83.262.9----
SubArcFace [7]89.348.071.686.776.772.983.758.176.983.879.973.2
Ours92.528.777.975.482.062.392.829.179.767.584.855.0
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.441, + 0.482, + 0.459 + ], + "angle": 0, + "content": "4.4. Comparison on unknown sample identification" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.849 + ], + "angle": 0, + "content": "We compare our proposed SASep method with various existing open-set recognition methods, including both 2D and 3D techniques, across all tracks on synthetic, synth-to-real, and real-to-real benchmarks. The results of each benchmark's subgroups are aggregated to provide a comprehensive evaluation. As shown in Tab. 1, our proposed method consistently demonstrates the highest AUROC scores and the lowest FPR95 scores across all benchmarks and backbones. Specifically, SASep surpasses the best-performing methods in each benchmark with AUROC and FPR95 improvements of at least \\(2\\%\\) and \\(6\\%\\), respectively. Notably, SASep achieves the highest improvements in the AUROC and FPR95 metrics on synthetic-to-real and synthetic benchmarks, with gains of \\(6.3\\%\\) and \\(15.1\\%\\), respectively. Furthermore, SASep demonstrates higher improvements in the synthetic benchmark compared to the other benchmarks. We believe this is due to the characteristics of the synthetic data, which has more complete geometric structures and less real-world noise. This allows SASep to effectively capture the class-relevant structures, leading to improved performance. Apart from the synthetic benchmark, the improvements observed in the synthetic-to-real and real-to-real benchmarks further highlight the robustness of our method in addressing the challenges presented by the cross-domain gap between known and unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To provide a more intuitive comparison of our method with other representative methods, we visualize the AUROC and FPR95 metrics for both DGCNN and PointNet++" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.534 + ], + "angle": 0, + "content": "across all subgroups of each benchmark using radar charts, as shown in Fig. 5. Clearly, our proposed method (highlighted in orange) outperforms other methods in nearly all experiments, demonstrating its superior performance. Please refer to the Supplemental Material for more detailed experimental results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.849, + 0.56 + ], + "angle": 0, + "content": "4.5. Comparison on closed-set classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We evaluate the closed-set classification performance of our method against baseline closed-set models with DGCNN and PointNet++ backbones. The closed-set models are trained and tested solely on known classes. Although our method further performs the unknown sample identification task, we only compare the classification accuracy. Tab. 2 shows the accuracy (ACC) of each method across three subgroups: SN1, SR1, and RR1. For the DGCNN backbone, our method achieves improved performance in SR1 and RR1, reaching \\(81.3\\%\\) and \\(94\\%\\), respectively, compared to the closed-set model's \\(72.1\\%\\) and \\(91.5\\%\\), with a slight decrease in SN1. The improvements observed in SR1 and RR1 demonstrate the enhanced robustness of our method when applied to real-world scan data. This can be attributed to our design, which focuses on capturing semantically important geometric structures, thereby improving the model's resilience to real-world scenarios. For the PointNet++ backbone, our method achieves results comparable to the closed-set model, with only a slight decrease in performance. This may be due to PointNet++'s stronger capacity for feature extraction using hierarchical network structures. Overall, these results suggest that our method not only competes" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27301" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.277, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 5. Comparisons on all the subgroups of synthetic, synthetic-to-real, real-to-real benchmarks. The AUROC and FPR95 metrics on the DGCNN and PointNet++ are drawn in the radar chart (a)-(d). The results of our proposed SASEp are highlighted in orange." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.33, + 0.486, + 0.372 + ], + "angle": 0, + "content": "Table 2. Results of closed-set classification task on the subgroup (SN1, SR1, RR1) of synthetic, synthetic-to-real, and real-to-real benchmarks. The ACC metric is used as the evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.383, + 0.452, + 0.452 + ], + "angle": 0, + "content": "
MethodsDGCNNPointNet++
SN1SR1RR1SN1SR1RR1
closed-set84.772.191.585.980.095.6
Our83.381.394.085.480.094.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.507 + ], + "angle": 0, + "content": "with closed-set models in terms of accuracy but also provides increased robustness against real-world challenges." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.515, + 0.242, + 0.531 + ], + "angle": 0, + "content": "4.6. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.613 + ], + "angle": 0, + "content": "To evaluate the contribution of each component in our proposed method, we conduct ablation studies on SN2 of the synthetic benchmark with DGCNN backbone. As shown in Tab. 3, we report the AUROC, FPR95, ACC, and mACC metrics across various configurations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.704 + ], + "angle": 0, + "content": "When all modules (TSD, GSS, SMS) are enabled, the model achieves its best performance, with an AUROC of 94.8, the lowest FPR95 at 24.0, and the highest ACC and mACC at 90.5 and 79.8, respectively. This indicates that the full combination maximizes performance for both known class classification and unknown class identification tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Removing each module clarifies specific strengths. Excluding TSD (with only GSS and SMS active) leads to a decreased AUROC and mACC, demonstrating TSD's significant role in both tasks of open-set learning. When GSS is removed, the AUROC slightly increases, but the FPR95 drops to 29.1, indicating GSS's key role in controlling false positives. Without SMS, all the metrics are slightly decreased, suggesting SMS's impact on overall accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Individual module analysis highlights TSD's strength, achieving an AUROC of 94.2 and the lowest FPR95 of 25.9. GSS alone enhances mACC but raises FPR95 to 30.6, while SMS alone has the highest FPR95, indicating its lower effectiveness in isolation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.33, + 0.908, + 0.36 + ], + "angle": 0, + "content": "Table 3. Comparing the contribution of major components in our proposed SASep on SN2 of synthetic benchmark using DGCNN." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.37, + 0.905, + 0.503 + ], + "angle": 0, + "content": "
TSDGSSSMSAUROC↑FPR95↓ACC↑mACC↑
✓✓93.728.388.374.0
✓✓94.129.189.978.4
✓✓93.826.989.278.5
✓94.225.990.077.6
✓93.630.688.679.0
✓93.036.189.675.9
✓✓✓94.824.090.579.8
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.524, + 0.634, + 0.54 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.55, + 0.907, + 0.822 + ], + "angle": 0, + "content": "In this work, we propose a novel open-set point cloud learning method, SASep, which identifies unknown classes by emphasizing the importance of utilizing the geometric structures of objects' different parts, rather than treating them as a whole. Specifically, we design a tunable semantic decomposition module to decompose objects into the semantically important and unimportant parts for classification. Then, we introduce a geometric synthesis strategy and a synth-aided margin separation module to leverage these decomposed parts, improving feature learning for known classes and enhancing class distinction. Experimental results show that SASep significantly outperforms the state-of-the-art method in various datasets. In scenarios with severe corruption, our method may be limited due to challenges in accurately capturing semantics, motivating us to explore a more robust method for semantic decomposition. We hope SASep offers some new insights into this under-explored area." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.834, + 0.682, + 0.851 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.908, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Key Research and Development Program for Young Scientists of China (2024YFB3310100), the China National Nat" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "ural Science Foundation No. 62202182, the Guangdong Basic and Applied Basic Research Foundation 2024A1515010224, the National Key Research and Development Program of China under Grant 2023YFB4503400, and the China National Natural Science Foundation Nos. 62176101, 62276109, 62450064, and 62322205." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.188, + 0.221 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.23, + 0.482, + 0.299 + ], + "angle": 0, + "content": "[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning - benchmarking and understanding semantic novelty detection on point clouds. In NeurIPS, pages 21228-21240. Curran Associates, Inc., 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.301, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[2] Ayush Bhardwaj, Sakshee Pimpale, Saurabh Kumar, and Biplab Banerjee. Empowering knowledge distillation via open set recognition for robust 3d point cloud classification. Pattern Recognition Letters, 151:172-179, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.359, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.483, + 0.484 + ], + "angle": 0, + "content": "[4] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial Reciprocal Points Learning for Open Set Recognition. IEEE TPAMI, 44(11):8065-8081, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In ECCV, pages 741-757, Cham, 2020. Springer International Publishing. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[8] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Dario Fontanel, Fabio Cermelli, Massimiliano Mancini, and Barbara Caputo. Detecting anomalies in semantic segmentation with prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 113-121, 2021. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, page 4995-5004, New York, NY, USA, 2024. Association for Computing Machinery. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.23, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[14] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In Proceedings of the 39th International Conference on Machine Learning, pages 8759-8773. PMLR, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] Jie Hong, Shi Qiu, Weihao Li, Saeed Anwar, Mehrtash Harandi, Nick Barnes, and Lars Petersson. Pointcam: Cut-and-mix for open-set point cloud learning. arXiv preprint arXiv:2212.02011, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[16] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, pages 677-689. Curran Associates, Inc., 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[17] Inhyuk Jo, Jungtaek Kim, Hyohyeong Kang, Yong-Deok Kim, and Seungjin Choi. Open set recognition by regularising classifier with fake data generated by generative adversarial networks. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2686–2690, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, pages 18661-18673. Curran Associates, Inc., 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.904, + 0.515 + ], + "angle": 0, + "content": "[19] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, pages 813-822, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.904, + 0.616 + ], + "angle": 0, + "content": "[21] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[22] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[23] Huan Lei, Naveed Akhtar, and Ajmal Mian. Octree guided cnn with spherical kernels for 3d point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In CVPR, pages 20402-20411, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In ICIP, pages 3118-3122, 2021. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27303" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[29] Ravish Mehra, Pushkar Tripathi, Alla Sheffer, and Niloy J Mitra. Visibility of noisy point cloud data. Computers & Graphics, 34(3):219-230, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[31] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[33] Paolo Rabino, Antonio Alliegro, and Tatiana Tommasi. 3d semantic novelty detection via large-scale pre-trained models. IEEE Access, 12:135352-135361, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[34] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[36] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157. Curran Associates, Inc., 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[38] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[39] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[40] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[41] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In ICCV, pages 9302-9311, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[42] Yu Wang, Junxian Mu, Pengfei Zhu, and Qinghua Hu. Exploring diverse representations for open set recognition. AAAI, 38(6):5731-5739, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[43] Tingyu Weng, Jun Xiao, Hao Pan, and Haiyong Jiang. Partcom: Part composition learning for 3d open-set recognition. IJCV, 132(4):1393-1416, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[44] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[45] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27304" + } + ] +] \ No newline at end of file diff --git a/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_origin.pdf b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f867055189b8ef59dddc346173454b4bdf04f384 --- /dev/null +++ b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/eddf9d0f-a790-4fc2-8177-c2291f25615a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58a34bf0a8561d7a92f917521109dd44743f28a5ffc30d1223a459963f8ef876 +size 3348930 diff --git a/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/full.md b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4347719c05632e482ae7db5a1518d32b645f6e60 --- /dev/null +++ b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/full.md @@ -0,0 +1,283 @@ +# SASep: Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds + +Jinfeng Xu $^{1}$ Xianzhi Li $^{1,2}$ Yuan Tang $^{1}$ Xu Han $^{1}$ Qiao Yu $^{1}$ Yixue Hao $^{1}$ Long Hu $^{1*}$ Min Chen $^{3,4}$ + +1Huazhong University of Science and Technology +3South China University of Technology + +$^{2}$ Guangdong Intelligent Robotics Institute + $^{4}$ Pazhou Laboratory + +{jinfengx, xzli, yuan_tang, xhanxu, qiaoyu_epic, yixuehao, hulong, minchen}@hust.edu.cn + +# Abstract + +Recent advancements in deep learning have greatly enhanced 3D object recognition, but most models are limited to closed-set scenarios, unable to handle unknown samples in real-world applications. Open-set recognition (OSR) addresses this limitation by enabling models to both classify known classes and identify novel classes. However, current OSR methods rely on global features to differentiate known and unknown classes, treating the entire object uniformly and overlooking the varying semantic importance of its different parts. To address this gap, we propose Salience-Aware Structured Separation (SASep), which includes (i) a tunable semantic decomposition (TSD) module to semantically decompose objects into important and unimportant parts, (ii) a geometric synthesis strategy (GSS) to generate pseudo-unknown objects by combining these unimportant parts, and (iii) a synth-aided margin separation (SMS) module to enhance feature-level separation by expanding the feature distributions between classes. Together, these components improve both geometric and feature representations, enhancing the model's ability to effectively distinguish known and unknown classes. Experimental results show that SASep achieves superior performance in 3D OSR, outperforming existing state-of-the-art methods. The codes are available at https://github.com/JinfengX/SASep. + +# 1. Introduction + +In recent years, deep learning has significantly advanced 3D object recognition, enabling its application in diverse areas such as autonomous driving, robotics, augmented/virtual reality, etc. Most existing 3D recognition models [31, 44, 46] are designed for closed-set scenarios, where all test classes are assumed to have been seen during training. However, real-world applications usually present dynamic environments where unknown or unseen samples outside the + +![](images/7a42e8c55a35204c0012fb44801e060d7e311b1752c466c643d9252a8cd2845c.jpg) +Figure 1. Given input objects, the TSD module semantically decomposes them into important and unimportant parts based on each point's semantic relevance to classification. The important parts enhance feature learning for known classes, while the unimportant parts are mixed in GSS to generate pseudo-unknown samples. The SMS module then constrains the features of objects and their decomposed parts to improve inter-class separation and intraclass compactness. + +set of known classes may be encountered. These out-of-distribution samples pose a significant challenge, as the closed-set models are unable to handle these unknown samples appropriately, often misclassifying them as one of the predefined known classes. To address this challenge, open-set recognition (OSR) task has gained increasing attention recently, as it aims to extend the capabilities of closed-set models by enabling them to classify known classes and identify unknown ones simultaneously. + +Early 2D OSR approaches [13, 14, 24] aim to distinguish between known and unknown classes by using discrimina + +tive functions to compute confidence scores from network predictions for each input. Samples are identified as unknown when their confidence scores are low. Thus, designing a robust open-set solver hinges on ensuring the confidence score is discriminative, meaning scores for known classes should clearly exceed those for unknown classes. + +Although recent 2D OSR works [4, 36, 42] have achieved significant progress in the 2D area, research on OSR for 3D data remains relatively underdeveloped. Unlike 2D OSR, the unique spatial characteristics of 3D data limit the performance of directly adapting 2D OSR methods to 3D, as they fall short of leveraging 3D-specific structural and topological information. Recent works [2, 27, 43] propose to address the 3D OSR task from the perspective of geometric reconstruction, feature embeddings, and sample synthesis. However, these approaches, similar to many 2D OSR methods, primarily focus on assigning confidence scores based on whole-object features, often overlooking the fact that different parts of an object contribute differently to classification. For example, specific regions, such as the wheels of a car or the wings of an airplane, carry essential features that are more informative for classification than other parts. + +Based on this observation, we propose a novel method for 3D open-set recognition on point clouds, named Salience-Aware Structured Separation (SASep), which tackles the open-set point cloud learning task by highlighting the semantically important geometric structures of objects. As illustrated in Fig. 1, SASep first decomposes the input objects into important and unimportant parts with a tunable semantic decomposition (TSD) module based on the relevance of each region to the class semantics. Then, we augment the feature learning of known classes by incorporating important parts into the training dataset. For the unimportant parts, we design a geometric synthesis strategy (GSS) to generate synthetic samples, serving as pseudo-unknown objects to simulate geometric structures of unknown classes. Furthermore, we refine the feature space with synthie-aided margin separation module (SMS) by applying constraints that reduce the feature distance between the object and its important parts, while increasing the feature distance between the object and those from other classes. This process enhances the feature representations of known classes, thus reducing the likelihood of misclassifying unknown class samples as known classes. The experimental results demonstrate that our proposed method outperforms existing OSR approaches, achieving state-of-the-art (SOTA) performance in 3D open-set recognition. Overall, our contributions can be summarized as follows: + +- We introduce a novel approach for 3D open-set recognition on point clouds, called Salience-Aware Structured Separation (SASep), which enhances the distinction between known and unknown classes by utilizing the se + +mantically decomposed parts of objects. + +- We propose a novel semantic decomposition method TSD that decomposes an object into semantically important and unimportant regions, enabling the model to utilize structural information from different parts of the objects. +- We design GSS and SMS to utilize the decomposed parts of TSD to explore feature distribution of the unknown class and enhance feature distinction of the known classes, respectively. + +# 2. Related work + +# 2.1. Closed-set point cloud learning + +Closed-set point cloud learning has made significant progress, focusing on accurate classification and segmentation of 3D objects from predefined categories. Existing approaches can be broadly categorized into view-based, voxel-based, and point-based methods. View-based methods [5, 21, 35, 37] classify 3D shapes by using 2D CNN to extract features from the projected 2D images of input. Voxel-based methods [6, 11, 23, 28, 39] process point clouds by dividing 3D space into voxel grids, allowing 3D CNNs to learn the spatial features. Pioneering point-based methods [31] use Multi-Layer Perceptrons (MLPs) with symmetric functions to directly process raw point cloud data. The following point-based research [12, 20, 26, 32, 40] further improved the performance by enhancing the models' capacity to aggregate the local and global features. While these methods have achieved significant success in recognizing known classes, they face notable limitations when dealing with unseen classes, as the models usually tend to misclassify the unknown objects as one of the predefined known classes. This limitation restricts the models' ability to generalize to real-world applications, where novel or unknown classes may be encountered during inference. + +# 2.2. Open-set 2D recognition + +To address the limitation of closed-set learning, open-set recognition (OSR) enables models to classify known objects and identify unknown or novel classes during inference. The straightforward approaches [16, 24, 36, 41] utilize discriminative functions to evaluate confidence scores for identifying unknown classes. MSP [13] established a robust baseline by using Maximum Softmax Probability to identify samples as unknown when they exhibit low maximum softmax scores. Similarly, Maximum Logit Score (MLS) [14] directly uses the highest logit score as a confidence measure. Beyond these discriminative-based approaches, reconstruction-based methods [8, 17, 19] identify unknown objects by comparing the original inputs with their reconstructions from the generative models. Recent research in 2D OSR aims to learn feature embeddings that form compact clusters for known classes. For instance, + +![](images/020c2a250dadfcae2541720591b7e24079188a34324344fab22d0ba69f57615b.jpg) +Figure 2. Overall architecture of our SASep. The architecture includes a forward process where the input point cloud $\mathcal{P}_{in}$ is encoded with backbone, followed by a backward process utilizing the tunable semantic decomposition module to decompose $\mathcal{P}_{in}$ into high- and low-saliency parts. Then, the high-saliency parts are used in the synth-aided margin separation module for separating feature distribution; the low-saliency parts are combined for synthetic objects with a geometric synthesis strategy. + +ARPL [4] introduces reciprocal points in the embedding space to minimize the overlap of known and unknown distributions. Fontanel et al. [10] proposed a method that uses learnable class prototypes to detect anomalies by measuring the cosine distance of inputs from learned class prototypes. Furthermore, several studies [9, 25, 30] have investigated using synthetic samples to expose potential unknown classes to the model, thereby enhancing the model's generalization on samples out of the training dataset. + +Though 2D OSR has seen significant advancements, applying these methods to 3D domains presents limitations, particularly in leveraging the unique spatial and geometric complexities within 3D data. Thus, directly shifting a 2D OSR method to 3D data often fails to effectively distinguish known and unknown samples. + +# 2.3. Open-set 3D recognition + +In contrast to the widely explored 2D OSR, research on 3D OSR remains relatively underdeveloped. Masuda et al. [27] developed a reconstruction-based method to obtain anomaly scores by measuring reconstruction errors between model input and output. To enhance the learning of distribution for unknown classes, Bhardwaj et al. [2] proposed generating pseudo-unknown samples by mixing objects from known classes. To further refine the synthesis strategy, Lee et al. [22] and Hong et al. [15] designed more sophisticated algorithms to synthesize high-quality pseudo unknown samples, considering the local geometric structure of objects. More recently, PartCom [43] introduced a novel part codebook to learn 3D object representations by capturing distinct part compositions across classes. Similarly, Rabino et al. [33] introduced a training-free method lever + +aging part-based reasoning. Building on the work of 2D and 3D OSR, 3DOS [1] provided a comprehensive benchmark for 3D open-set learning, revealing the strengths and limitations of existing approaches in the 3D open-set context. + +Existing open-set recognition methods focus on distinguishing known and unknown objects or features as a whole, while overlooking the essential semantic information carried by different parts in an object. For example, the wheels of vehicles often exhibit more significant class-relevant information than the other parts. Humans can recognize objects even with a partial view of such recognizable geometric or topological structures. Building on this observation, we develop a method to address the challenges of open-set learning by focusing on the semantically important parts within 3D objects. Our proposed method strengthens the association between class semantics and object partial structures, ultimately improving the accuracy and robustness of 3D open-set recognition. + +# 3. Method + +# 3.1. Architecture overview + +As introduced in Sec. 1, we propose the Saliency-Aware Separation (SASep) method to address the limitations of current methods, which overlook the different contributions of object parts to classification. Figure 2 illustrates the architecture overview of our SASep. Building upon the basic pipeline of using an encoder to lift the input point cloud $P_{in}$ into the global feature $\mathcal{F}_{in}$ , the closed-set model output $\mathcal{Y}_{in} \in \mathbb{R}^{C}$ can be formulated as the softmax of $g(\mathcal{F}_{in})$ where $g(\cdot)$ maps the global feature to a set of logits of $C$ known classes. Rather than using the commonly adopted + +linear layer to implement $g(\cdot)$ , we formulate $\mathcal{Y}_{in}$ by employing a prototype-based method [10] (see Eq. (8)), aiming to help capture important geometric characteristics of each class. Then, we calculate the confidence score $\mathcal{Q}$ by extracting the maximum logit score (MLS) from $\mathcal{Y}_{in}$ . Objects with a $\mathcal{Q}$ value below a predefined threshold are classified as belonging to the unknown class. + +To enhance this basic pipeline, we first design a tunable semantic (TSD) module (marked by a green dash box) to evaluate the semantic saliency of each point, thereby dividing objects into important (high-saliency) and unimportant (low-saliency) parts based on the level of semantic saliency. For the high-saliency parts, we directly treat them as new instances of the same known class as the original objects and re-input them into the feature encoder to enhance the network's learning ability for this category. Note that, for the low-saliency parts that carry less known-class-specific information, we do not discard them. Instead, we develop a geometric synthesis strategy (GSS) to create unknown objects by combining low-saliency parts from various known objects. The rationale behind GSS lies in the fact that unknown- and known-class objects are likely to share similar local geometric structures. By re-inputting these synthetic objects into the network, the network's discriminative ability for similar regions can be enhanced. + +Please note that the above designs are processed at the geometric level. Given the difficulty in accurately recognizing unknown classes, we take a further step by designing a synth-aided margin separation (SMS) module (marked by a blue dash box) that operates $\mathcal{F}_{in}$ at the feature level, with the key idea of enlarging the feature distribution margins between classes. In the following sections, we will provide detailed information about each of our designs. + +# 3.2. Tunable semantic decomposition module + +The tunable semantic decomposition (TSD) module aims to decompose the objects into high-saliency parts and low-saliency parts, based on each point's contribution to classification predictions. The pipeline of our proposed TSD module is illustrated in Fig. 3. For the input point cloud $\mathcal{P}_{in}$ , the model extracts its per-point features $\mathcal{A}$ to predict the output $\mathcal{Y}_{in}$ . Since $\mathcal{Y}_{in}$ reflects the class probabilities for $\mathcal{P}_{in}$ , we aim to estimate the contribution of each point to the corresponding object class probability in $\mathcal{Y}_{in}$ . To achieve this goal, we adopt the Grad-CAM approach [34] to perform semantic evaluation, which aggregates each channel of the point-wise feature $\mathcal{A}$ with associated gradient information to generate a saliency map $S$ : + +$$ +\mathcal {S} = R e L U \left(\sum_ {k} \left(\frac {1}{N} \sum_ {i} \frac {\partial \mathcal {Y} _ {i n} ^ {c}}{\partial \mathcal {A} _ {i} ^ {k}}\right) \mathcal {A} ^ {k}\right), \tag {1} +$$ + +where $\mathcal{A}_i^k$ denotes the $k$ -th channel of $\mathcal{A}$ corresponding to the $i$ -th point and $N$ is the number of points. Generally, + +![](images/c6118ba15fd7a5d01856a6ff0d54aec440e86990ebe2a8631ded8a55912d1c57.jpg) +Figure 3. Illustration of our proposed tunable semantic decomposition (TSD) module. The TSD module decomposes the input objects into high- and low-saliency parts using the semantically evaluated saliency map $S$ . It also incorporates random partial views to enhance geometric information in a tunable manner. + +a point with a higher saliency score in $S$ indicates that the point has a greater influence on the classification result, i.e., it is more important. Then, we sort $S$ in ascending order, designating the first $\lfloor N / M \rfloor$ points as the low-saliency part, while the remaining points are marked as the high-saliency part. Here, $M$ is a parameter that determines the number of low-saliency parts used in the geometric synthesis strategy. + +Because the saliency map $\mathcal{S}$ focuses on evaluating the semantic relevance, the saliency-based decomposition tends to generate relatively fixed decomposed parts. However, diverse scenarios, such as occlusion, noise, or corrupted data, are common in real-world point cloud data, which makes it difficult for the saliency-based decomposition method to effectively capture the key parts. To improve the robustness and generalization of the model in such challenging conditions, we introduce a geometric augmentation strategy to the TSD module. Specifically, by estimating visibility from view poses with random radii around the input points [29], we obtain the cropped parts of objects from these partial views. For each partial view, the saliency scores of each point are retrieved from $\mathcal{S}$ Unlike simply augmenting the dataset by appending additional partial view data, we replace low- and high-saliency parts with partial views that have overall scores above $\tau_h$ and below $\tau_l$ ( $\tau_h > \tau_l$ ), respectively. By fine-tuning $\tau_h$ and $\tau_l$ , the TSD module can flexibly shift the decomposition strategy from semantically focused (when $\tau_h$ is high and $\tau_l$ is low) to geometrically focused (when $\tau_h$ is low and $\tau_l$ is high). This adaptability allows the TSD module to generate different samples that capture both semantic importance and geometric diversity. + +# 3.3. Geometric synthesis strategy + +To address the scarcity of unknown class data during training, the geometric synthesis strategy (GSS) synthesizes pseudo-unknown samples by mixing multiple low-saliency + +parts generated from the TSD module, thereby expanding the dataset to explore potential combinations of local structures that unknown classes may encompass. Unlike methods that synthesize data using random parts of known objects, which risk introducing class-specific geometric structures, GSS limits impact on classification by leveraging less influential low-saliency parts. + +Specifically, the low-saliency parts undergo standard transformations, followed by a mixing operation where $M$ randomly selected parts from different objects are combined. The mixed objects are then encoded to generate outputs $\mathcal{V}_s$ , which are supervised using pseudo-labels $\hat{\mathcal{V}}_s$ set to $C + 1$ . To reduce overconfidence, label smoothing with base smooth weight $\epsilon$ is applied to $\hat{\mathcal{V}}_s$ . Additional smooth weight $\epsilon_h$ is also assigned to the involved object classes, ensuring the mixed objects retain characteristics of their originals. Assuming that the mixed parts are from classes $\mathcal{K}_s = \{c_1,\dots ,c_n\}$ , we define $\hat{\mathcal{V}}_s$ as: + +$$ +\hat {\mathcal {Y}} _ {s} ^ {i} = \left\{ \begin{array}{l l} 1 - \epsilon - \epsilon_ {h}, & i f i = C + 1 \\ \epsilon / C, & i f i \in \mathcal {K} a n d i \notin \mathcal {K} _ {s} \\ \epsilon / C + \epsilon_ {h} M _ {c _ {i}} / M, & i f i \in \mathcal {K} _ {s}, \end{array} \right. \tag {2} +$$ + +where $M_{c_i}$ is the number of mixed objects belonging to $c_i$ . Finally, the loss of GSS is formulated as: + +$$ +\mathcal {L} _ {s} = - \sum_ {i = 1} ^ {C + 1} \hat {\mathcal {Y}} _ {s} ^ {i} \log \frac {e ^ {\mathcal {Y} _ {s} ^ {i}}}{\sum_ {j = 1} ^ {C + 1} e ^ {\mathcal {Y} _ {s} ^ {j}}}, \tag {3} +$$ + +where $\mathcal{V}_s^i$ denotes the $i$ -th class probability of $\mathcal{V}_s$ . + +# 3.4. Synth-aided margin separation module + +The synth-aided margin separation (SMS) module is designed to improve the compactness within each class and the separability between different classes. It achieves this by applying feature constraints to individual object features. The key idea of SMS is illustrated in Fig. 4. We denote the global feature of the input point cloud $\mathcal{P}_{c1}$ and its high-saliency part $\mathcal{P}_h$ as $\mathcal{F}_{c1}$ and $\mathcal{F}_h$ , respectively. To highlight the semantically-relevant local geometric structures in $\mathcal{P}_{c1}$ , we design a dual alignment strategy comprising attraction and repulsion forces. First, $\mathcal{P}_h$ is designated as a positive exemplar, and its feature $\mathcal{F}_h$ is aligned towards $\mathcal{F}_{c1}$ to consolidate intra-class similarity. Concurrently, to enforce inter-class discrepancy, the negative exemplar $\mathcal{P}_{c2}$ ( $c1 \neq c2$ ) is pushed away via its feature $\mathcal{F}_{c2}$ in the embedding space. By using both positive and negative exemplars, the feature constraint $\mathcal{L}_m$ is defined through a triplet loss framework to ensure that $d(\mathcal{F}_{c1}, \mathcal{F}_h)$ is smaller than $d(\mathcal{F}_{c1}, \mathcal{F}_{c2})$ . + +$$ +\mathcal {L} _ {m} = d \left(\mathcal {F} _ {c 1}, \mathcal {F} _ {h}\right) - d \left(\mathcal {F} _ {c 1}, \mathcal {F} _ {c 2}\right) + \tau_ {m}, \tag {4} +$$ + +where $d(\cdot)$ is the distance metric function. + +![](images/94c1f375387e88583ffd2dda6e97e932dc802c0577717f8ed141c7ef22926bf1.jpg) +Figure 4. Illustration of our proposed synth-aided margin separation (SMS) module. The SMS module regularizes the feature space with a triplet loss, which employs the pseudo-features to help form the positive examples and negative examples. + +Although $\mathcal{F}_h$ and $\mathcal{F}_{c2}$ are appropriate for constructing feature constraints, their performance may be limited by data scarcity due to the sparse distribution of feature samples. To increase the availability of feature pairs, we generate synthetic features by adding Gaussian noise to $\mathcal{F}_{c1}$ . + +$$ +\mathbb {F} _ {s} = \mathcal {F} _ {c 1} + \mathcal {N} (0, \mathcal {W}), \tag {5} +$$ + +where $\mathcal{W} = \{w_1, \dots, w_n\}$ is a list of weights for Gaussian noise. Then, we filter $\mathbb{F}_s$ to retain only the samples whose predicted results match the ground-truth labels. This step enhances the reliability of synthetic features. From these filtered samples, one is randomly selected to serve as the pseudo feature. This pseudo feature is then utilized for either random positive or negative replacements by substituting $\mathcal{F}_h$ or $\mathcal{F}_{c2}$ , resulting in the creation of new positive exemplars $\mathcal{F}_p$ or negative exemplars $\mathcal{F}_n$ . In the case of positive replacement, the distance $d(\mathcal{F}_{c1}, \mathcal{F}_h)$ is adjusted to accommodate a broader variability in $d(\mathcal{F}_{c1}, \mathcal{F}_p)$ , allowing for greater flexibility compared to the original distance. Conversely, negative replacement is designed to enhance the robustness of object features by introducing disturbances directly into the feature space, applying force on $\mathcal{F}_{c1}$ with $d(\mathcal{F}_{c1}, \mathcal{F}_n)$ as highlighted in Eq. (4). It's important to note that only one of the positive or negative replacements will be executed for each input feature. Finally, the $\mathcal{L}_m$ is refined with a weighted triplet loss: + +$$ +\mathcal {L} _ {m} = \rho d \left(\mathcal {F} _ {c 1}, \mathcal {F} _ {p}\right) - \eta d \left(\mathcal {F} _ {c 1}, \mathcal {F} _ {n}\right) + \tau_ {m}, \tag {6} +$$ + +where $\rho$ and $\eta$ are the balance weights and $\tau_{m}$ is the margin threshold. Given that the positive exemplar pairs are already constrained by the classification loss, we set $\rho < \eta$ to prioritize the enhancement of $d(\mathcal{F}_{c1},\mathcal{F}_n)$ . + +# 3.5. Loss function + +The total loss of SAsep integrates four components: (1) the primary classification loss $\mathcal{L}_{cls}$ for raw inputs $\mathcal{P}_{in}$ , (2) the + +high-saliency loss $\mathcal{L}_h$ for supervising TSD-generated high-saliency parts $\mathcal{P}_h$ , (3) the synthesis loss $\mathcal{L}_s$ (Eq. (3)) for synthetic samples from the GSS module, and (4) the feature margin loss $\mathcal{L}_m$ (Eq. (6)). For $\mathcal{P}_{in}$ of class $c$ , $\mathcal{L}_{cls}$ applies cross-entropy to the model output $\mathcal{Y}_{in}$ : + +$$ +\mathcal {L} _ {c l s} = - \log \frac {e ^ {\mathcal {Y} _ {i n} ^ {c}}}{\sum_ {i = 1} ^ {C + 1} e ^ {\mathcal {Y} _ {i n} ^ {i}}}, \tag {7} +$$ + +where $\mathcal{Y}_{in}^{i}$ is the cosine similarity between the encoded feature $\mathcal{F}_{in} \in \mathbb{R}^d$ of $\mathcal{P}_{in}$ and learnable prototypes $\mathcal{M} = \{\mathbf{m}_1, \dots, \mathbf{m}_C, \mathbf{m}_{C+1}\} \in \mathbb{R}^{(C+1) \times d}$ , with $\mathbf{m}_{C+1}$ for unknown classes. Hence, $\mathcal{Y}_{in}^{i}$ is defined as follows: + +$$ +\mathcal {Y} _ {i n} ^ {i} = g \left(\mathcal {F} _ {i n}; \mathbf {m} _ {i}\right) = \frac {\mathcal {F} _ {i n} \cdot \mathbf {m} _ {i}}{\| \mathcal {F} _ {i n} \| \| \mathbf {m} _ {i} \|}, \tag {8} +$$ + +where $\|\cdot\|$ denotes the $L2$ -norm. Note that the prototypes $\mathcal{M}$ are initialized randomly and optimized during training. As aforementioned in Sec. 3.1, we consider TSD-generated high-saliency parts $\mathcal{P}_h$ as new instances belonging to the same class as $\mathcal{P}_{in}$ and directly feed $\mathcal{P}_h$ into the model. Consequently, we compute $\mathcal{L}_h$ for $\mathcal{P}_h$ using the same formulation outlined in Eq. (7). The overall open-set classification loss $\mathcal{L}$ is thus defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \alpha \mathcal {L} _ {h} + \beta \mathcal {L} _ {s} + \gamma \mathcal {L} _ {m}, \tag {9} +$$ + +where $\alpha, \beta$ and $\gamma$ are balance parameters. + +# 4. Experiments + +# 4.1. Datasets + +We conduct experiments using the benchmarks proposed by 3DOS [1], which are based on three widely used 3D object datasets: ShapeNetCore [3], ModelNet40 [45], and ScanObjectNN [38]. The ShapeNetCore dataset is a large synthetic 3D dataset that contains 51,127 3D mesh instances across 55 object categories. Approximately $70\%$ of the data is allocated for training, $10\%$ for validation, and $20\%$ for testing. The ModelNet40 dataset consists of 12,311 3D CAD-generated meshes from 40 categories. This dataset is divided into an official training set of 9,843 samples and a test set of 2,468 samples. The corresponding point cloud data points are uniformly sampled from the surfaces of the CAD meshes and scaled to fit with a unit sphere. ScanObjectNN is a real-world dataset comprising 2,902 3D scans across 15 categories. It presents significant challenges due to artifacts, such as background clutter, occlusions, missing parts, and non-uniform density. + +We follow the original experimental settings for open-set 3D recognition defined by 3DOS benchmarks, which include three tracks designed to evaluate the performance under different conditions of semantic and domain shifts. + +In the synthetic benchmark, the ShapeNetCore dataset is divided into three distinct category subgroups, designated as SN1, SN2 and SN3, each containing 18 classes with no overlap. Each group rotates as the set of known classes, while the other two are treated as unknown classes, creating three unique experimental configurations. Similarly, the real-to-real benchmark follows the ScanObjectNN dataset, which is also divided into three subgroups (RR1, RR2, and RR3) to create three corresponding experimental configurations under real-scanned conditions. The synthetic-to-real benchmark establishes cross-domain scenarios (SR1, SR2), wherein synthetic samples from ModelNet40 serve as the known class set, while real scans from ScanObjectNN are designated as the unknown class set. Given the significant domain shifts between the known and unknown categories, this synthetic-to-real benchmark presents greater challenges compared to the other two benchmarks. + +# 4.2. Evaluation metrics + +Open-set 3D recognition includes two main tasks: closed-set classification and identification of unknown samples. For the closed-set classification task, we utilize accuracy (ACC) and mean accuracy (mACC) to evaluate the model's performance in correctly classifying known samples. To assess performance in identifying unknown samples, we utilize AUROC and FPR95 metrics. The AUROC (Area Under the Receiver Operating Characteristic Curve) measures a model's ability to differentiate between known and unknown samples. The ROC curve plots the true positive rate against the false positive rate across various thresholds to reject unknown samples. A higher AUROC value indicates better performance in identifying unknown samples. FPR95(False Positive Rate at $95\%$ True Positive Rate) measures the rate of false positives while keeping the true positive rate fixed at $95\%$ . This metric offers insights into the model's capacity in reducing misclassifications, particularly in preventing the incorrect identification of unknown samples as known. + +# 4.3. Implementation details + +Our proposed SASep method is implemented using the PyTorch framework and evaluated with DGCNN [40] and PointNet++ [32]. To simplify the training process, saliency maps for each object were generated using Grad-CAM [34] with the pre-trained closed-set backbones. All models are trained on 4 NVIDIA RTX 3090 GPUs for 200 epochs, with a total batch size of 120. For both DGCNN and PointNet++, we adopt the Adam optimizer and CosLR scheduler with an initial learning rate set to 0.001. The balance parameters $\alpha$ , $\beta$ , and $\gamma$ in Eq. (9) are assigned to 0.1, 0.01, and 0.3, respectively. In Eq. (6), the weights $\rho$ and $\eta$ are set to 0.01 and 1.0, respectively, with $\tau_{m}$ fixed at 10. + +Table 1. The results of unknown sample identification task on the synthetic, synthetic-to-real, and real-to-real benchmarks. Each column shows the average results of each benchmark's subgroups. The best results are in bold in each metric. + +
MethodsDGCNNPointNet++
SyntheticSynthetic-to-RealReal-to-RealSyntheticSynthetic-to-RealReal-to-Real
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
MSP [13]85.263.266.790.670.882.881.371.075.683.280.877.2
MLS [14]86.251.865.790.572.677.381.961.874.881.783.067.4
ODIN [24]86.351.365.790.672.677.380.564.376.080.883.365.5
Energy [41]86.251.765.690.872.678.182.061.874.882.482.667.3
GradNorm [16]72.976.663.491.567.284.562.683.373.083.280.467.6
ReAct [36]88.443.965.690.575.175.386.152.974.681.482.767.4
VAE [27]77.064.263.384.6--------
NF88.546.771.382.370.282.181.266.776.484.381.465.8
OE+mixup [22]85.256.565.390.869.885.080.169.965.791.669.086.6
PointCaM [15]--71.186.6--------
ARPL+CS [4]84.473.667.189.8--80.367.875.482.1--
Cosine proto [10]86.548.957.991.078.068.286.955.778.276.183.664.9
CE (L2)89.143.666.189.274.974.488.644.277.782.379.175.8
SupCon [18]84.364.5----83.262.9----
SubArcFace [7]89.348.071.686.776.772.983.758.176.983.879.973.2
Ours92.528.777.975.482.062.392.829.179.767.584.855.0
+ +# 4.4. Comparison on unknown sample identification + +We compare our proposed SASep method with various existing open-set recognition methods, including both 2D and 3D techniques, across all tracks on synthetic, synth-to-real, and real-to-real benchmarks. The results of each benchmark's subgroups are aggregated to provide a comprehensive evaluation. As shown in Tab. 1, our proposed method consistently demonstrates the highest AUROC scores and the lowest FPR95 scores across all benchmarks and backbones. Specifically, SASep surpasses the best-performing methods in each benchmark with AUROC and FPR95 improvements of at least $2\%$ and $6\%$ , respectively. Notably, SASep achieves the highest improvements in the AUROC and FPR95 metrics on synthetic-to-real and synthetic benchmarks, with gains of $6.3\%$ and $15.1\%$ , respectively. Furthermore, SASep demonstrates higher improvements in the synthetic benchmark compared to the other benchmarks. We believe this is due to the characteristics of the synthetic data, which has more complete geometric structures and less real-world noise. This allows SASep to effectively capture the class-relevant structures, leading to improved performance. Apart from the synthetic benchmark, the improvements observed in the synthetic-to-real and real-to-real benchmarks further highlight the robustness of our method in addressing the challenges presented by the cross-domain gap between known and unknown categories. + +To provide a more intuitive comparison of our method with other representative methods, we visualize the AUROC and FPR95 metrics for both DGCNN and PointNet++ + +across all subgroups of each benchmark using radar charts, as shown in Fig. 5. Clearly, our proposed method (highlighted in orange) outperforms other methods in nearly all experiments, demonstrating its superior performance. Please refer to the Supplemental Material for more detailed experimental results. + +# 4.5. Comparison on closed-set classification + +We evaluate the closed-set classification performance of our method against baseline closed-set models with DGCNN and PointNet++ backbones. The closed-set models are trained and tested solely on known classes. Although our method further performs the unknown sample identification task, we only compare the classification accuracy. Tab. 2 shows the accuracy (ACC) of each method across three subgroups: SN1, SR1, and RR1. For the DGCNN backbone, our method achieves improved performance in SR1 and RR1, reaching $81.3\%$ and $94\%$ , respectively, compared to the closed-set model's $72.1\%$ and $91.5\%$ , with a slight decrease in SN1. The improvements observed in SR1 and RR1 demonstrate the enhanced robustness of our method when applied to real-world scan data. This can be attributed to our design, which focuses on capturing semantically important geometric structures, thereby improving the model's resilience to real-world scenarios. For the PointNet++ backbone, our method achieves results comparable to the closed-set model, with only a slight decrease in performance. This may be due to PointNet++'s stronger capacity for feature extraction using hierarchical network structures. Overall, these results suggest that our method not only competes + +![](images/da296c652ce4e3a39ea2cea4e0085a5da61d0bf7c94e03d9e0997a2d82bc2444.jpg) +Figure 5. Comparisons on all the subgroups of synthetic, synthetic-to-real, real-to-real benchmarks. The AUROC and FPR95 metrics on the DGCNN and PointNet++ are drawn in the radar chart (a)-(d). The results of our proposed SASEp are highlighted in orange. + +Table 2. Results of closed-set classification task on the subgroup (SN1, SR1, RR1) of synthetic, synthetic-to-real, and real-to-real benchmarks. The ACC metric is used as the evaluation metric. + +
MethodsDGCNNPointNet++
SN1SR1RR1SN1SR1RR1
closed-set84.772.191.585.980.095.6
Our83.381.394.085.480.094.8
+ +with closed-set models in terms of accuracy but also provides increased robustness against real-world challenges. + +# 4.6. Ablation study + +To evaluate the contribution of each component in our proposed method, we conduct ablation studies on SN2 of the synthetic benchmark with DGCNN backbone. As shown in Tab. 3, we report the AUROC, FPR95, ACC, and mACC metrics across various configurations. + +When all modules (TSD, GSS, SMS) are enabled, the model achieves its best performance, with an AUROC of 94.8, the lowest FPR95 at 24.0, and the highest ACC and mACC at 90.5 and 79.8, respectively. This indicates that the full combination maximizes performance for both known class classification and unknown class identification tasks. + +Removing each module clarifies specific strengths. Excluding TSD (with only GSS and SMS active) leads to a decreased AUROC and mACC, demonstrating TSD's significant role in both tasks of open-set learning. When GSS is removed, the AUROC slightly increases, but the FPR95 drops to 29.1, indicating GSS's key role in controlling false positives. Without SMS, all the metrics are slightly decreased, suggesting SMS's impact on overall accuracy. + +Individual module analysis highlights TSD's strength, achieving an AUROC of 94.2 and the lowest FPR95 of 25.9. GSS alone enhances mACC but raises FPR95 to 30.6, while SMS alone has the highest FPR95, indicating its lower effectiveness in isolation. + +Table 3. Comparing the contribution of major components in our proposed SASep on SN2 of synthetic benchmark using DGCNN. + +
TSDGSSSMSAUROC↑FPR95↓ACC↑mACC↑
✓✓93.728.388.374.0
✓✓94.129.189.978.4
✓✓93.826.989.278.5
✓94.225.990.077.6
✓93.630.688.679.0
✓93.036.189.675.9
✓✓✓94.824.090.579.8
+ +# 5. Conclusion + +In this work, we propose a novel open-set point cloud learning method, SASep, which identifies unknown classes by emphasizing the importance of utilizing the geometric structures of objects' different parts, rather than treating them as a whole. Specifically, we design a tunable semantic decomposition module to decompose objects into the semantically important and unimportant parts for classification. Then, we introduce a geometric synthesis strategy and a synth-aided margin separation module to leverage these decomposed parts, improving feature learning for known classes and enhancing class distinction. Experimental results show that SASep significantly outperforms the state-of-the-art method in various datasets. In scenarios with severe corruption, our method may be limited due to challenges in accurately capturing semantics, motivating us to explore a more robust method for semantic decomposition. We hope SASep offers some new insights into this under-explored area. + +# Acknowledgements + +This work was supported by the National Key Research and Development Program for Young Scientists of China (2024YFB3310100), the China National Nat + +ural Science Foundation No. 62202182, the Guangdong Basic and Applied Basic Research Foundation 2024A1515010224, the National Key Research and Development Program of China under Grant 2023YFB4503400, and the China National Natural Science Foundation Nos. 62176101, 62276109, 62450064, and 62322205. + +# References + +[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning - benchmarking and understanding semantic novelty detection on point clouds. In NeurIPS, pages 21228-21240. Curran Associates, Inc., 2022. 3, 6 +[2] Ayush Bhardwaj, Sakshee Pimpale, Saurabh Kumar, and Biplab Banerjee. Empowering knowledge distillation via open set recognition for robust 3d point cloud classification. Pattern Recognition Letters, 151:172-179, 2021. 2, 3 +[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[4] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial Reciprocal Points Learning for Open Set Recognition. IEEE TPAMI, 44(11):8065-8081, 2022. 2, 3, 7 +[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2 +[7] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In ECCV, pages 741-757, Cham, 2020. Springer International Publishing. 7 +[8] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In ICLR, 2022. 2 +[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 3 +[10] Dario Fontanel, Fabio Cermelli, Massimiliano Mancini, and Barbara Caputo. Detecting anomalies in semantic segmentation with prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 113-121, 2021. 3, 4, 7 +[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2 +[12] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, page 4995-5004, New York, NY, USA, 2024. Association for Computing Machinery. 2 + +[13] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 7 +[14] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In Proceedings of the 39th International Conference on Machine Learning, pages 8759-8773. PMLR, 2022. 1, 2, 7 +[15] Jie Hong, Shi Qiu, Weihao Li, Saeed Anwar, Mehrtash Harandi, Nick Barnes, and Lars Petersson. Pointcam: Cut-and-mix for open-set point cloud learning. arXiv preprint arXiv:2212.02011, 2022. 3, 7 +[16] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, pages 677-689. Curran Associates, Inc., 2021. 2, 7 +[17] Inhyuk Jo, Jungtaek Kim, Hyohyeong Kang, Yong-Deok Kim, and Seungjin Choi. Open set recognition by regularising classifier with fake data generated by generative adversarial networks. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2686–2690, 2018. 2 +[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, pages 18661-18673. Curran Associates, Inc., 2020. 7 +[19] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, pages 813-822, 2021. 2 +[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 2 +[21] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2 +[22] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3, 7 +[23] Huan Lei, Naveed Akhtar, and Ajmal Mian. Octree guided cnn with spherical kernels for 3d point clouds. In CVPR, 2019. 2 +[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 7 +[25] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In CVPR, pages 20402-20411, 2023. 3 +[26] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 2 +[27] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In ICIP, pages 3118-3122, 2021. 2, 3, 7 + +[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928, 2015. 2 +[29] Ravish Mehra, Pushkar Tripathi, Alla Sheffer, and Niloy J Mitra. Visibility of noisy point cloud data. Computers & Graphics, 34(3):219-230, 2010. 4 +[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 3 +[31] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2 +[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6 +[33] Paolo Rabino, Antonio Alliegro, and Tatiana Tommasi. 3d semantic novelty detection via large-scale pre-trained models. IEEE Access, 12:135352-135361, 2024. 3 +[34] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4, 6 +[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2 +[36] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157. Curran Associates, Inc., 2021. 2, 7 +[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2 +[38] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6 +[39] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4), 2017. 2 +[40] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 2, 6 +[41] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In ICCV, pages 9302-9311, 2021. 2, 7 +[42] Yu Wang, Junxian Mu, Pengfei Zhu, and Qinghua Hu. Exploring diverse representations for open set recognition. AAAI, 38(6):5731-5739, 2024. 2 +[43] Tingyu Weng, Jun Xiao, Hao Pan, and Haiyong Jiang. Partcom: Part composition learning for 3d open-set recognition. IJCV, 132(4):1393-1416, 2024. 2, 3 +[44] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1 + +[45] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6 +[46] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1 \ No newline at end of file diff --git a/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/images.zip b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c6179e12b659d42691e462a59b0400a4747883b1 --- /dev/null +++ b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e3edb3daf0b34ddc1c558c40831f51b43d14b6cb70e3cd27f807102263700c1 +size 488782 diff --git a/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/layout.json b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5a4cf7bcf82ae5eda68775ec3fc1a170d37f7a5 --- /dev/null +++ b/2025/SASep_ Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds/layout.json @@ -0,0 +1,8791 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 529, + 140 + ], + "type": "text", + "content": "SASep: Saliency-Aware Structured Separation of Geometry and Feature for Open Set Learning on Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": "Jinfeng Xu" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Xianzhi Li" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Yuan Tang" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Xu Han" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Qiao Yu" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Yixue Hao" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Long Hu" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "text", + "content": " Min Chen" + }, + { + "bbox": [ + 53, + 160, + 550, + 175 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 175, + 315, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 175, + 315, + 204 + ], + "spans": [ + { + "bbox": [ + 75, + 175, + 315, + 204 + ], + "type": "text", + "content": "1Huazhong University of Science and Technology \n3South China University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 175, + 529, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 175, + 529, + 204 + ], + "spans": [ + { + "bbox": [ + 326, + 175, + 529, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 326, + 175, + 529, + 204 + ], + "type": "text", + "content": "Guangdong Intelligent Robotics Institute \n" + }, + { + "bbox": [ + 326, + 175, + 529, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 326, + 175, + 529, + 204 + ], + "type": "text", + "content": "Pazhou Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 205, + 533, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 205, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 205, + 533, + 217 + ], + "type": "text", + "content": "{jinfengx, xzli, yuan_tang, xhanxu, qiaoyu_epic, yixuehao, hulong, minchen}@hust.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 558 + ], + "type": "text", + "content": "Recent advancements in deep learning have greatly enhanced 3D object recognition, but most models are limited to closed-set scenarios, unable to handle unknown samples in real-world applications. Open-set recognition (OSR) addresses this limitation by enabling models to both classify known classes and identify novel classes. However, current OSR methods rely on global features to differentiate known and unknown classes, treating the entire object uniformly and overlooking the varying semantic importance of its different parts. To address this gap, we propose Salience-Aware Structured Separation (SASep), which includes (i) a tunable semantic decomposition (TSD) module to semantically decompose objects into important and unimportant parts, (ii) a geometric synthesis strategy (GSS) to generate pseudo-unknown objects by combining these unimportant parts, and (iii) a synth-aided margin separation (SMS) module to enhance feature-level separation by expanding the feature distributions between classes. Together, these components improve both geometric and feature representations, enhancing the model's ability to effectively distinguish known and unknown classes. Experimental results show that SASep achieves superior performance in 3D OSR, outperforming existing state-of-the-art methods. The codes are available at https://github.com/JinfengX/SASep." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "content": "In recent years, deep learning has significantly advanced 3D object recognition, enabling its application in diverse areas such as autonomous driving, robotics, augmented/virtual reality, etc. Most existing 3D recognition models [31, 44, 46] are designed for closed-set scenarios, where all test classes are assumed to have been seen during training. However, real-world applications usually present dynamic environments where unknown or unseen samples outside the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 243, + 555, + 470 + ], + "blocks": [ + { + "bbox": [ + 317, + 243, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 317, + 243, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 555, + 470 + ], + "type": "image", + "image_path": "7a42e8c55a35204c0012fb44801e060d7e311b1752c466c643d9252a8cd2845c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 476, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 564 + ], + "type": "text", + "content": "Figure 1. Given input objects, the TSD module semantically decomposes them into important and unimportant parts based on each point's semantic relevance to classification. The important parts enhance feature learning for known classes, while the unimportant parts are mixed in GSS to generate pseudo-unknown samples. The SMS module then constrains the features of objects and their decomposed parts to improve inter-class separation and intraclass compactness." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 581, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 555, + 689 + ], + "type": "text", + "content": "set of known classes may be encountered. These out-of-distribution samples pose a significant challenge, as the closed-set models are unable to handle these unknown samples appropriately, often misclassifying them as one of the predefined known classes. To address this challenge, open-set recognition (OSR) task has gained increasing attention recently, as it aims to extend the capabilities of closed-set models by enabling them to classify known classes and identify unknown ones simultaneously." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Early 2D OSR approaches [13, 14, 24] aim to distinguish between known and unknown classes by using discrimina" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27295" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "tive functions to compute confidence scores from network predictions for each input. Samples are identified as unknown when their confidence scores are low. Thus, designing a robust open-set solver hinges on ensuring the confidence score is discriminative, meaning scores for known classes should clearly exceed those for unknown classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 296, + 350 + ], + "type": "text", + "content": "Although recent 2D OSR works [4, 36, 42] have achieved significant progress in the 2D area, research on OSR for 3D data remains relatively underdeveloped. Unlike 2D OSR, the unique spatial characteristics of 3D data limit the performance of directly adapting 2D OSR methods to 3D, as they fall short of leveraging 3D-specific structural and topological information. Recent works [2, 27, 43] propose to address the 3D OSR task from the perspective of geometric reconstruction, feature embeddings, and sample synthesis. However, these approaches, similar to many 2D OSR methods, primarily focus on assigning confidence scores based on whole-object features, often overlooking the fact that different parts of an object contribute differently to classification. For example, specific regions, such as the wheels of a car or the wings of an airplane, carry essential features that are more informative for classification than other parts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 353, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 296, + 662 + ], + "type": "text", + "content": "Based on this observation, we propose a novel method for 3D open-set recognition on point clouds, named Salience-Aware Structured Separation (SASep), which tackles the open-set point cloud learning task by highlighting the semantically important geometric structures of objects. As illustrated in Fig. 1, SASep first decomposes the input objects into important and unimportant parts with a tunable semantic decomposition (TSD) module based on the relevance of each region to the class semantics. Then, we augment the feature learning of known classes by incorporating important parts into the training dataset. For the unimportant parts, we design a geometric synthesis strategy (GSS) to generate synthetic samples, serving as pseudo-unknown objects to simulate geometric structures of unknown classes. Furthermore, we refine the feature space with synthie-aided margin separation module (SMS) by applying constraints that reduce the feature distance between the object and its important parts, while increasing the feature distance between the object and those from other classes. This process enhances the feature representations of known classes, thus reducing the likelihood of misclassifying unknown class samples as known classes. The experimental results demonstrate that our proposed method outperforms existing OSR approaches, achieving state-of-the-art (SOTA) performance in 3D open-set recognition. Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "content": "- We introduce a novel approach for 3D open-set recognition on point clouds, called Salience-Aware Structured Separation (SASep), which enhances the distinction between known and unknown classes by utilizing the se" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 323, + 72, + 485, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 485, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 485, + 84 + ], + "type": "text", + "content": "mantically decomposed parts of objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 84, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 132 + ], + "type": "text", + "content": "- We propose a novel semantic decomposition method TSD that decomposes an object into semantically important and unimportant regions, enabling the model to utilize structural information from different parts of the objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "content": "- We design GSS and SMS to utilize the decomposed parts of TSD to explore feature distribution of the unknown class and enhance feature distinction of the known classes, respectively." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 192, + 398, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 398, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 398, + 204 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 212, + 482, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 482, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 482, + 224 + ], + "type": "text", + "content": "2.1. Closed-set point cloud learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 495 + ], + "type": "text", + "content": "Closed-set point cloud learning has made significant progress, focusing on accurate classification and segmentation of 3D objects from predefined categories. Existing approaches can be broadly categorized into view-based, voxel-based, and point-based methods. View-based methods [5, 21, 35, 37] classify 3D shapes by using 2D CNN to extract features from the projected 2D images of input. Voxel-based methods [6, 11, 23, 28, 39] process point clouds by dividing 3D space into voxel grids, allowing 3D CNNs to learn the spatial features. Pioneering point-based methods [31] use Multi-Layer Perceptrons (MLPs) with symmetric functions to directly process raw point cloud data. The following point-based research [12, 20, 26, 32, 40] further improved the performance by enhancing the models' capacity to aggregate the local and global features. While these methods have achieved significant success in recognizing known classes, they face notable limitations when dealing with unseen classes, as the models usually tend to misclassify the unknown objects as one of the predefined known classes. This limitation restricts the models' ability to generalize to real-world applications, where novel or unknown classes may be encountered during inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 503, + 450, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 450, + 516 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 450, + 516 + ], + "type": "text", + "content": "2.2. Open-set 2D recognition" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "To address the limitation of closed-set learning, open-set recognition (OSR) enables models to classify known objects and identify unknown or novel classes during inference. The straightforward approaches [16, 24, 36, 41] utilize discriminative functions to evaluate confidence scores for identifying unknown classes. MSP [13] established a robust baseline by using Maximum Softmax Probability to identify samples as unknown when they exhibit low maximum softmax scores. Similarly, Maximum Logit Score (MLS) [14] directly uses the highest logit score as a confidence measure. Beyond these discriminative-based approaches, reconstruction-based methods [8, 17, 19] identify unknown objects by comparing the original inputs with their reconstructions from the generative models. Recent research in 2D OSR aims to learn feature embeddings that form compact clusters for known classes. For instance," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27296" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 259 + ], + "type": "image", + "image_path": "020c2a250dadfcae2541720591b7e24079188a34324344fab22d0ba69f57615b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Overall architecture of our SASep. The architecture includes a forward process where the input point cloud " + }, + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "type": "text", + "content": " is encoded with backbone, followed by a backward process utilizing the tunable semantic decomposition module to decompose " + }, + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 54, + 265, + 555, + 312 + ], + "type": "text", + "content": " into high- and low-saliency parts. Then, the high-saliency parts are used in the synth-aided margin separation module for separating feature distribution; the low-saliency parts are combined for synthetic objects with a geometric synthesis strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 325, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 325, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 294, + 432 + ], + "type": "text", + "content": "ARPL [4] introduces reciprocal points in the embedding space to minimize the overlap of known and unknown distributions. Fontanel et al. [10] proposed a method that uses learnable class prototypes to detect anomalies by measuring the cosine distance of inputs from learned class prototypes. Furthermore, several studies [9, 25, 30] have investigated using synthetic samples to expose potential unknown classes to the model, thereby enhancing the model's generalization on samples out of the training dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": "Though 2D OSR has seen significant advancements, applying these methods to 3D domains presents limitations, particularly in leveraging the unique spatial and geometric complexities within 3D data. Thus, directly shifting a 2D OSR method to 3D data often fails to effectively distinguish known and unknown samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 192, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 192, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 192, + 528 + ], + "type": "text", + "content": "2.3. Open-set 3D recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "In contrast to the widely explored 2D OSR, research on 3D OSR remains relatively underdeveloped. Masuda et al. [27] developed a reconstruction-based method to obtain anomaly scores by measuring reconstruction errors between model input and output. To enhance the learning of distribution for unknown classes, Bhardwaj et al. [2] proposed generating pseudo-unknown samples by mixing objects from known classes. To further refine the synthesis strategy, Lee et al. [22] and Hong et al. [15] designed more sophisticated algorithms to synthesize high-quality pseudo unknown samples, considering the local geometric structure of objects. More recently, PartCom [43] introduced a novel part codebook to learn 3D object representations by capturing distinct part compositions across classes. Similarly, Rabino et al. [33] introduced a training-free method lever" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 325, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 373 + ], + "type": "text", + "content": "aging part-based reasoning. Building on the work of 2D and 3D OSR, 3DOS [1] provided a comprehensive benchmark for 3D open-set learning, revealing the strengths and limitations of existing approaches in the 3D open-set context." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 374, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 542 + ], + "type": "text", + "content": "Existing open-set recognition methods focus on distinguishing known and unknown objects or features as a whole, while overlooking the essential semantic information carried by different parts in an object. For example, the wheels of vehicles often exhibit more significant class-relevant information than the other parts. Humans can recognize objects even with a partial view of such recognizable geometric or topological structures. Building on this observation, we develop a method to address the challenges of open-set learning by focusing on the semantically important parts within 3D objects. Our proposed method strengthens the association between class semantics and object partial structures, ultimately improving the accuracy and robustness of 3D open-set recognition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 554, + 370, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 554, + 370, + 566 + ], + "spans": [ + { + "bbox": [ + 314, + 554, + 370, + 566 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 574, + 441, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 441, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 441, + 586 + ], + "type": "text", + "content": "3.1. Architecture overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "As introduced in Sec. 1, we propose the Saliency-Aware Separation (SASep) method to address the limitations of current methods, which overlook the different contributions of object parts to classification. Figure 2 illustrates the architecture overview of our SASep. Building upon the basic pipeline of using an encoder to lift the input point cloud " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "P_{in}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " into the global feature " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{in}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", the closed-set model output " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " can be formulated as the softmax of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "g(\\mathcal{F}_{in})" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " maps the global feature to a set of logits of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " known classes. Rather than using the commonly adopted" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27297" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "linear layer to implement " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", we formulate " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " by employing a prototype-based method [10] (see Eq. (8)), aiming to help capture important geometric characteristics of each class. Then, we calculate the confidence score " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " by extracting the maximum logit score (MLS) from " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". Objects with a " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " value below a predefined threshold are classified as belonging to the unknown class." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 370 + ], + "type": "text", + "content": "To enhance this basic pipeline, we first design a tunable semantic (TSD) module (marked by a green dash box) to evaluate the semantic saliency of each point, thereby dividing objects into important (high-saliency) and unimportant (low-saliency) parts based on the level of semantic saliency. For the high-saliency parts, we directly treat them as new instances of the same known class as the original objects and re-input them into the feature encoder to enhance the network's learning ability for this category. Note that, for the low-saliency parts that carry less known-class-specific information, we do not discard them. Instead, we develop a geometric synthesis strategy (GSS) to create unknown objects by combining low-saliency parts from various known objects. The rationale behind GSS lies in the fact that unknown- and known-class objects are likely to share similar local geometric structures. By re-inputting these synthetic objects into the network, the network's discriminative ability for similar regions can be enhanced." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "type": "text", + "content": "Please note that the above designs are processed at the geometric level. Given the difficulty in accurately recognizing unknown classes, we take a further step by designing a synth-aided margin separation (SMS) module (marked by a blue dash box) that operates " + }, + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{in}" + }, + { + "bbox": [ + 55, + 372, + 295, + 467 + ], + "type": "text", + "content": " at the feature level, with the key idea of enlarging the feature distribution margins between classes. In the following sections, we will provide detailed information about each of our designs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 268, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 268, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 268, + 485 + ], + "type": "text", + "content": "3.2. Tunable semantic decomposition module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": "The tunable semantic decomposition (TSD) module aims to decompose the objects into high-saliency parts and low-saliency parts, based on each point's contribution to classification predictions. The pipeline of our proposed TSD module is illustrated in Fig. 3. For the input point cloud " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": ", the model extracts its per-point features " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " to predict the output " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " reflects the class probabilities for " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": ", we aim to estimate the contribution of each point to the corresponding object class probability in " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": ". To achieve this goal, we adopt the Grad-CAM approach [34] to perform semantic evaluation, which aggregates each channel of the point-wise feature " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": " with associated gradient information to generate a saliency map " + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 491, + 295, + 647 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 651, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 651, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 90, + 651, + 295, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = R e L U \\left(\\sum_ {k} \\left(\\frac {1}{N} \\sum_ {i} \\frac {\\partial \\mathcal {Y} _ {i n} ^ {c}}{\\partial \\mathcal {A} _ {i} ^ {k}}\\right) \\mathcal {A} ^ {k}\\right), \\tag {1}", + "image_path": "f1c7c3869ccb73e54739f4c8612584a299b525f49cd26c51a2c57fe90095d6dc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i^k" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "-th channel of " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "-th point and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is the number of points. Generally," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 213 + ], + "type": "image", + "image_path": "c6118ba15fd7a5d01856a6ff0d54aec440e86990ebe2a8631ded8a55912d1c57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 216, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 270 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed tunable semantic decomposition (TSD) module. The TSD module decomposes the input objects into high- and low-saliency parts using the semantically evaluated saliency map " + }, + { + "bbox": [ + 313, + 216, + 553, + 270 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 216, + 553, + 270 + ], + "type": "text", + "content": ". It also incorporates random partial views to enhance geometric information in a tunable manner." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "text", + "content": "a point with a higher saliency score in " + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "text", + "content": " indicates that the point has a greater influence on the classification result, i.e., it is more important. Then, we sort " + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "text", + "content": " in ascending order, designating the first " + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "inline_equation", + "content": "\\lfloor N / M \\rfloor" + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "text", + "content": " points as the low-saliency part, while the remaining points are marked as the high-saliency part. Here, " + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 286, + 553, + 370 + ], + "type": "text", + "content": " is a parameter that determines the number of low-saliency parts used in the geometric synthesis strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": "Because the saliency map " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " focuses on evaluating the semantic relevance, the saliency-based decomposition tends to generate relatively fixed decomposed parts. However, diverse scenarios, such as occlusion, noise, or corrupted data, are common in real-world point cloud data, which makes it difficult for the saliency-based decomposition method to effectively capture the key parts. To improve the robustness and generalization of the model in such challenging conditions, we introduce a geometric augmentation strategy to the TSD module. Specifically, by estimating visibility from view poses with random radii around the input points [29], we obtain the cropped parts of objects from these partial views. For each partial view, the saliency scores of each point are retrieved from " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " Unlike simply augmenting the dataset by appending additional partial view data, we replace low- and high-saliency parts with partial views that have overall scores above " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_h" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " and below " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_l" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_h > \\tau_l" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": "), respectively. By fine-tuning " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_h" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_l" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": ", the TSD module can flexibly shift the decomposition strategy from semantically focused (when " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_h" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " is high and " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_l" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " is low) to geometrically focused (when " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_h" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " is low and " + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "inline_equation", + "content": "\\tau_l" + }, + { + "bbox": [ + 313, + 372, + 553, + 647 + ], + "type": "text", + "content": " is high). This adaptability allows the TSD module to generate different samples that capture both semantic importance and geometric diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 471, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 471, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 471, + 671 + ], + "type": "text", + "content": "3.3. Geometric synthesis strategy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "To address the scarcity of unknown class data during training, the geometric synthesis strategy (GSS) synthesizes pseudo-unknown samples by mixing multiple low-saliency" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27298" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "parts generated from the TSD module, thereby expanding the dataset to explore potential combinations of local structures that unknown classes may encompass. Unlike methods that synthesize data using random parts of known objects, which risk introducing class-specific geometric structures, GSS limits impact on classification by leveraging less influential low-saliency parts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": "Specifically, the low-saliency parts undergo standard transformations, followed by a mixing operation where " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": " randomly selected parts from different objects are combined. The mixed objects are then encoded to generate outputs " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_s" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": ", which are supervised using pseudo-labels " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_s" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": " set to " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": ". To reduce overconfidence, label smoothing with base smooth weight " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_s" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": ". Additional smooth weight " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\epsilon_h" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": " is also assigned to the involved object classes, ensuring the mixed objects retain characteristics of their originals. Assuming that the mixed parts are from classes " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_s = \\{c_1,\\dots ,c_n\\}" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_s" + }, + { + "bbox": [ + 55, + 156, + 295, + 288 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 296, + 295, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 296, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 67, + 296, + 295, + 339 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {Y}} _ {s} ^ {i} = \\left\\{ \\begin{array}{l l} 1 - \\epsilon - \\epsilon_ {h}, & i f i = C + 1 \\\\ \\epsilon / C, & i f i \\in \\mathcal {K} a n d i \\notin \\mathcal {K} _ {s} \\\\ \\epsilon / C + \\epsilon_ {h} M _ {c _ {i}} / M, & i f i \\in \\mathcal {K} _ {s}, \\end{array} \\right. \\tag {2}", + "image_path": "d58cf30e342ab03752e9dade5f96991c853fdbf67794b94ea12f9d2091bb3169.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "inline_equation", + "content": "M_{c_i}" + }, + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "text", + "content": " is the number of mixed objects belonging to " + }, + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "text", + "content": ". Finally, the loss of GSS is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 379, + 295, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 379, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 110, + 379, + 295, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = - \\sum_ {i = 1} ^ {C + 1} \\hat {\\mathcal {Y}} _ {s} ^ {i} \\log \\frac {e ^ {\\mathcal {Y} _ {s} ^ {i}}}{\\sum_ {j = 1} ^ {C + 1} e ^ {\\mathcal {Y} _ {s} ^ {j}}}, \\tag {3}", + "image_path": "7c1ae723bbafe76cd65222ba2237fc1138b8b095517a9880d65dbb6e6b13f17b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_s^i" + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "text", + "content": "-th class probability of " + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_s" + }, + { + "bbox": [ + 55, + 419, + 255, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 439, + 261, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 261, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 261, + 452 + ], + "type": "text", + "content": "3.4. Synth-aided margin separation module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": "The synth-aided margin separation (SMS) module is designed to improve the compactness within each class and the separability between different classes. It achieves this by applying feature constraints to individual object features. The key idea of SMS is illustrated in Fig. 4. We denote the global feature of the input point cloud " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{c1}" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " and its high-saliency part " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c1}" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": ", respectively. To highlight the semantically-relevant local geometric structures in " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{c1}" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": ", we design a dual alignment strategy comprising attraction and repulsion forces. First, " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " is designated as a positive exemplar, and its feature " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " is aligned towards " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c1}" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " to consolidate intra-class similarity. Concurrently, to enforce inter-class discrepancy, the negative exemplar " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{c2}" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "c1 \\neq c2" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": ") is pushed away via its feature " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c2}" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " in the embedding space. By using both positive and negative exemplars, the feature constraint " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " is defined through a triplet loss framework to ensure that " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "d(\\mathcal{F}_{c1}, \\mathcal{F}_h)" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": " is smaller than " + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "inline_equation", + "content": "d(\\mathcal{F}_{c1}, \\mathcal{F}_{c2})" + }, + { + "bbox": [ + 55, + 457, + 295, + 673 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 681, + 295, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 681, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 94, + 681, + 295, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {h}\\right) - d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {c 2}\\right) + \\tau_ {m}, \\tag {4}", + "image_path": "b9a0d5e28e1019cde9b97a9523eb62fa241a601ec76c98b5746ee119e8a42cdd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 227, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 227, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 227, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 227, + 714 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 55, + 701, + 227, + 714 + ], + "type": "text", + "content": " is the distance metric function." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 552, + 210 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 210 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 210 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 210 + ], + "type": "image", + "image_path": "94c1f375387e88583ffd2dda6e97e932dc802c0577717f8ed141c7ef22926bf1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": "Figure 4. Illustration of our proposed synth-aided margin separation (SMS) module. The SMS module regularizes the feature space with a triplet loss, which employs the pseudo-features to help form the positive examples and negative examples." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": "Although " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c2}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " are appropriate for constructing feature constraints, their performance may be limited by data scarcity due to the sparse distribution of feature samples. To increase the availability of feature pairs, we generate synthetic features by adding Gaussian noise to " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c1}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 338, + 553, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 338, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 386, + 338, + 553, + 352 + ], + "type": "interline_equation", + "content": "\\mathbb {F} _ {s} = \\mathcal {F} _ {c 1} + \\mathcal {N} (0, \\mathcal {W}), \\tag {5}", + "image_path": "6b7e7a194152f211f207601936009a2c0bc02a66adbff407a0ac046acecdde3f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\{w_1, \\dots, w_n\\}" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " is a list of weights for Gaussian noise. Then, we filter " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_s" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " to retain only the samples whose predicted results match the ground-truth labels. This step enhances the reliability of synthetic features. From these filtered samples, one is randomly selected to serve as the pseudo feature. This pseudo feature is then utilized for either random positive or negative replacements by substituting " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c2}" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": ", resulting in the creation of new positive exemplars " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_p" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " or negative exemplars " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_n" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": ". In the case of positive replacement, the distance " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "d(\\mathcal{F}_{c1}, \\mathcal{F}_h)" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " is adjusted to accommodate a broader variability in " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "d(\\mathcal{F}_{c1}, \\mathcal{F}_p)" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": ", allowing for greater flexibility compared to the original distance. Conversely, negative replacement is designed to enhance the robustness of object features by introducing disturbances directly into the feature space, applying force on " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{c1}" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "d(\\mathcal{F}_{c1}, \\mathcal{F}_n)" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " as highlighted in Eq. (4). It's important to note that only one of the positive or negative replacements will be executed for each input feature. Finally, the " + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 313, + 359, + 555, + 587 + ], + "type": "text", + "content": " is refined with a weighted triplet loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 595, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 595, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 349, + 595, + 553, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\rho d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {p}\\right) - \\eta d \\left(\\mathcal {F} _ {c 1}, \\mathcal {F} _ {n}\\right) + \\tau_ {m}, \\tag {6}", + "image_path": "35e8965037a854b594dfa7852adc0495e037b6a4936569406ca3d144f4798734.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": " are the balance weights and " + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\tau_{m}" + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": " is the margin threshold. Given that the positive exemplar pairs are already constrained by the classification loss, we set " + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\rho < \\eta" + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": " to prioritize the enhancement of " + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "inline_equation", + "content": "d(\\mathcal{F}_{c1},\\mathcal{F}_n)" + }, + { + "bbox": [ + 313, + 616, + 555, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 671, + 400, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 400, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 400, + 682 + ], + "type": "text", + "content": "3.5. Loss function" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "The total loss of SAsep integrates four components: (1) the primary classification loss " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " for raw inputs " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", (2) the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "27299" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "high-saliency loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " for supervising TSD-generated high-saliency parts " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", (3) the synthesis loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " (Eq. (3)) for synthetic samples from the GSS module, and (4) the feature margin loss " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " (Eq. (6)). For " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " applies cross-entropy to the model output " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 140, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 140, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 120, + 140, + 295, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\log \\frac {e ^ {\\mathcal {Y} _ {i n} ^ {c}}}{\\sum_ {i = 1} ^ {C + 1} e ^ {\\mathcal {Y} _ {i n} ^ {i}}}, \\tag {7}", + "image_path": "a173585d3c709921f08cb958d949644a0b0965f6de30c068677b0c0afee6c0ad.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}^{i}" + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": " is the cosine similarity between the encoded feature " + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{in} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": " and learnable prototypes " + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathbf{m}_1, \\dots, \\mathbf{m}_C, \\mathbf{m}_{C+1}\\} \\in \\mathbb{R}^{(C+1) \\times d}" + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{C+1}" + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": " for unknown classes. Hence, " + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in}^{i}" + }, + { + "bbox": [ + 55, + 180, + 296, + 229 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 237, + 295, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 237, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 102, + 237, + 295, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} _ {i n} ^ {i} = g \\left(\\mathcal {F} _ {i n}; \\mathbf {m} _ {i}\\right) = \\frac {\\mathcal {F} _ {i n} \\cdot \\mathbf {m} _ {i}}{\\| \\mathcal {F} _ {i n} \\| \\| \\mathbf {m} _ {i} \\|}, \\tag {8}", + "image_path": "debfce472db9dd9c474ceca7ca464b8f886d6baa9c38928ea4db33e2a54c0a89.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": "-norm. Note that the prototypes " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " are initialized randomly and optimized during training. As aforementioned in Sec. 3.1, we consider TSD-generated high-saliency parts " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " as new instances belonging to the same class as " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{in}" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " and directly feed " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " into the model. Consequently, we compute " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_h" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " using the same formulation outlined in Eq. (7). The overall open-set classification loss " + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 271, + 296, + 367 + ], + "type": "text", + "content": " is thus defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 378, + 295, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 378, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 107, + 378, + 295, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\alpha \\mathcal {L} _ {h} + \\beta \\mathcal {L} _ {s} + \\gamma \\mathcal {L} _ {m}, \\tag {9}", + "image_path": "a8969d2fa14017bf67d2a5194446de08c30f5131e557a43d5b0b3fb6a593ab73.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 400, + 225, + 412 + ], + "type": "text", + "content": " are balance parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 424, + 137, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 137, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 137, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 444, + 118, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 118, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 118, + 454 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "text", + "content": "We conduct experiments using the benchmarks proposed by 3DOS [1], which are based on three widely used 3D object datasets: ShapeNetCore [3], ModelNet40 [45], and ScanObjectNN [38]. The ShapeNetCore dataset is a large synthetic 3D dataset that contains 51,127 3D mesh instances across 55 object categories. Approximately " + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "text", + "content": " of the data is allocated for training, " + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "text", + "content": " for validation, and " + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 665 + ], + "type": "text", + "content": " for testing. The ModelNet40 dataset consists of 12,311 3D CAD-generated meshes from 40 categories. This dataset is divided into an official training set of 9,843 samples and a test set of 2,468 samples. The corresponding point cloud data points are uniformly sampled from the surfaces of the CAD meshes and scaled to fit with a unit sphere. ScanObjectNN is a real-world dataset comprising 2,902 3D scans across 15 categories. It presents significant challenges due to artifacts, such as background clutter, occlusions, missing parts, and non-uniform density." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 712 + ], + "type": "text", + "content": "We follow the original experimental settings for open-set 3D recognition defined by 3DOS benchmarks, which include three tracks designed to evaluate the performance under different conditions of semantic and domain shifts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": "In the synthetic benchmark, the ShapeNetCore dataset is divided into three distinct category subgroups, designated as SN1, SN2 and SN3, each containing 18 classes with no overlap. Each group rotates as the set of known classes, while the other two are treated as unknown classes, creating three unique experimental configurations. Similarly, the real-to-real benchmark follows the ScanObjectNN dataset, which is also divided into three subgroups (RR1, RR2, and RR3) to create three corresponding experimental configurations under real-scanned conditions. The synthetic-to-real benchmark establishes cross-domain scenarios (SR1, SR2), wherein synthetic samples from ModelNet40 serve as the known class set, while real scans from ScanObjectNN are designated as the unknown class set. Given the significant domain shifts between the known and unknown categories, this synthetic-to-real benchmark presents greater challenges compared to the other two benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 425, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 425, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 425, + 300 + ], + "type": "text", + "content": "4.2. Evaluation metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "type": "text", + "content": "Open-set 3D recognition includes two main tasks: closed-set classification and identification of unknown samples. For the closed-set classification task, we utilize accuracy (ACC) and mean accuracy (mACC) to evaluate the model's performance in correctly classifying known samples. To assess performance in identifying unknown samples, we utilize AUROC and FPR95 metrics. The AUROC (Area Under the Receiver Operating Characteristic Curve) measures a model's ability to differentiate between known and unknown samples. The ROC curve plots the true positive rate against the false positive rate across various thresholds to reject unknown samples. A higher AUROC value indicates better performance in identifying unknown samples. FPR95(False Positive Rate at " + }, + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "type": "text", + "content": " True Positive Rate) measures the rate of false positives while keeping the true positive rate fixed at " + }, + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 536 + ], + "type": "text", + "content": ". This metric offers insights into the model's capacity in reducing misclassifications, particularly in preventing the incorrect identification of unknown samples as known." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 444, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 444, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 444, + 562 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "Our proposed SASep method is implemented using the PyTorch framework and evaluated with DGCNN [40] and PointNet++ [32]. To simplify the training process, saliency maps for each object were generated using Grad-CAM [34] with the pre-trained closed-set backbones. All models are trained on 4 NVIDIA RTX 3090 GPUs for 200 epochs, with a total batch size of 120. For both DGCNN and PointNet++, we adopt the Adam optimizer and CosLR scheduler with an initial learning rate set to 0.001. The balance parameters " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " in Eq. (9) are assigned to 0.1, 0.01, and 0.3, respectively. In Eq. (6), the weights " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " are set to 0.01 and 1.0, respectively, with " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{m}" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " fixed at 10." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27300" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. The results of unknown sample identification task on the synthetic, synthetic-to-real, and real-to-real benchmarks. Each column shows the average results of each benchmark's subgroups. The best results are in bold in each metric." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 328 + ], + "type": "table", + "html": "
MethodsDGCNNPointNet++
SyntheticSynthetic-to-RealReal-to-RealSyntheticSynthetic-to-RealReal-to-Real
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
MSP [13]85.263.266.790.670.882.881.371.075.683.280.877.2
MLS [14]86.251.865.790.572.677.381.961.874.881.783.067.4
ODIN [24]86.351.365.790.672.677.380.564.376.080.883.365.5
Energy [41]86.251.765.690.872.678.182.061.874.882.482.667.3
GradNorm [16]72.976.663.491.567.284.562.683.373.083.280.467.6
ReAct [36]88.443.965.690.575.175.386.152.974.681.482.767.4
VAE [27]77.064.263.384.6--------
NF88.546.771.382.370.282.181.266.776.484.381.465.8
OE+mixup [22]85.256.565.390.869.885.080.169.965.791.669.086.6
PointCaM [15]--71.186.6--------
ARPL+CS [4]84.473.667.189.8--80.367.875.482.1--
Cosine proto [10]86.548.957.991.078.068.286.955.778.276.183.664.9
CE (L2)89.143.666.189.274.974.488.644.277.782.379.175.8
SupCon [18]84.364.5----83.262.9----
SubArcFace [7]89.348.071.686.776.772.983.758.176.983.879.973.2
Ours92.528.777.975.482.062.392.829.179.767.584.855.0
", + "image_path": "732e334ad88cb5554969d6f063e55d5d5d6be96db0edf94b519461e4eb747054.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 294, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 294, + 363 + ], + "type": "text", + "content": "4.4. Comparison on unknown sample identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "text", + "content": "We compare our proposed SASep method with various existing open-set recognition methods, including both 2D and 3D techniques, across all tracks on synthetic, synth-to-real, and real-to-real benchmarks. The results of each benchmark's subgroups are aggregated to provide a comprehensive evaluation. As shown in Tab. 1, our proposed method consistently demonstrates the highest AUROC scores and the lowest FPR95 scores across all benchmarks and backbones. Specifically, SASep surpasses the best-performing methods in each benchmark with AUROC and FPR95 improvements of at least " + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "text", + "content": ", respectively. Notably, SASep achieves the highest improvements in the AUROC and FPR95 metrics on synthetic-to-real and synthetic benchmarks, with gains of " + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 54, + 373, + 295, + 672 + ], + "type": "text", + "content": ", respectively. Furthermore, SASep demonstrates higher improvements in the synthetic benchmark compared to the other benchmarks. We believe this is due to the characteristics of the synthetic data, which has more complete geometric structures and less real-world noise. This allows SASep to effectively capture the class-relevant structures, leading to improved performance. Apart from the synthetic benchmark, the improvements observed in the synthetic-to-real and real-to-real benchmarks further highlight the robustness of our method in addressing the challenges presented by the cross-domain gap between known and unknown categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To provide a more intuitive comparison of our method with other representative methods, we visualize the AUROC and FPR95 metrics for both DGCNN and PointNet++" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 350, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 422 + ], + "type": "text", + "content": "across all subgroups of each benchmark using radar charts, as shown in Fig. 5. Clearly, our proposed method (highlighted in orange) outperforms other methods in nearly all experiments, demonstrating its superior performance. Please refer to the Supplemental Material for more detailed experimental results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 431, + 519, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 519, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 519, + 443 + ], + "type": "text", + "content": "4.5. Comparison on closed-set classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "We evaluate the closed-set classification performance of our method against baseline closed-set models with DGCNN and PointNet++ backbones. The closed-set models are trained and tested solely on known classes. Although our method further performs the unknown sample identification task, we only compare the classification accuracy. Tab. 2 shows the accuracy (ACC) of each method across three subgroups: SN1, SR1, and RR1. For the DGCNN backbone, our method achieves improved performance in SR1 and RR1, reaching " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "81.3\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": ", respectively, compared to the closed-set model's " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "72.1\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "91.5\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": ", with a slight decrease in SN1. The improvements observed in SR1 and RR1 demonstrate the enhanced robustness of our method when applied to real-world scan data. This can be attributed to our design, which focuses on capturing semantically important geometric structures, thereby improving the model's resilience to real-world scenarios. For the PointNet++ backbone, our method achieves results comparable to the closed-set model, with only a slight decrease in performance. This may be due to PointNet++'s stronger capacity for feature extraction using hierarchical network structures. Overall, these results suggest that our method not only competes" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27301" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 213 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 213 + ], + "type": "image", + "image_path": "da296c652ce4e3a39ea2cea4e0085a5da61d0bf7c94e03d9e0997a2d82bc2444.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 219, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 219, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 555, + 243 + ], + "type": "text", + "content": "Figure 5. Comparisons on all the subgroups of synthetic, synthetic-to-real, real-to-real benchmarks. The AUROC and FPR95 metrics on the DGCNN and PointNet++ are drawn in the radar chart (a)-(d). The results of our proposed SASEp are highlighted in orange." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 303, + 276, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 261, + 297, + 294 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 297, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 297, + 294 + ], + "type": "text", + "content": "Table 2. Results of closed-set classification task on the subgroup (SN1, SR1, RR1) of synthetic, synthetic-to-real, and real-to-real benchmarks. The ACC metric is used as the evaluation metric." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 303, + 276, + 357 + ], + "lines": [ + { + "bbox": [ + 75, + 303, + 276, + 357 + ], + "spans": [ + { + "bbox": [ + 75, + 303, + 276, + 357 + ], + "type": "table", + "html": "
MethodsDGCNNPointNet++
SN1SR1RR1SN1SR1RR1
closed-set84.772.191.585.980.095.6
Our83.381.394.085.480.094.8
", + "image_path": "0abdbf831b52d4f774894afe98da23b0f9704c1a336f7b863523a44c0af3ce8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "text", + "content": "with closed-set models in terms of accuracy but also provides increased robustness against real-world challenges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 148, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 148, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 148, + 420 + ], + "type": "text", + "content": "4.6. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 485 + ], + "type": "text", + "content": "To evaluate the contribution of each component in our proposed method, we conduct ablation studies on SN2 of the synthetic benchmark with DGCNN backbone. As shown in Tab. 3, we report the AUROC, FPR95, ACC, and mACC metrics across various configurations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 557 + ], + "type": "text", + "content": "When all modules (TSD, GSS, SMS) are enabled, the model achieves its best performance, with an AUROC of 94.8, the lowest FPR95 at 24.0, and the highest ACC and mACC at 90.5 and 79.8, respectively. This indicates that the full combination maximizes performance for both known class classification and unknown class identification tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "content": "Removing each module clarifies specific strengths. Excluding TSD (with only GSS and SMS active) leads to a decreased AUROC and mACC, demonstrating TSD's significant role in both tasks of open-set learning. When GSS is removed, the AUROC slightly increases, but the FPR95 drops to 29.1, indicating GSS's key role in controlling false positives. Without SMS, all the metrics are slightly decreased, suggesting SMS's impact on overall accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "Individual module analysis highlights TSD's strength, achieving an AUROC of 94.2 and the lowest FPR95 of 25.9. GSS alone enhances mACC but raises FPR95 to 30.6, while SMS alone has the highest FPR95, indicating its lower effectiveness in isolation." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 293, + 553, + 398 + ], + "blocks": [ + { + "bbox": [ + 313, + 261, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 285 + ], + "type": "text", + "content": "Table 3. Comparing the contribution of major components in our proposed SASep on SN2 of synthetic benchmark using DGCNN." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 293, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 317, + 293, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 553, + 398 + ], + "type": "table", + "html": "
TSDGSSSMSAUROC↑FPR95↓ACC↑mACC↑
✓✓93.728.388.374.0
✓✓94.129.189.978.4
✓✓93.826.989.278.5
✓94.225.990.077.6
✓93.630.688.679.0
✓93.036.189.675.9
✓✓✓94.824.090.579.8
", + "image_path": "7c0c5c40786d50b33157f5b8dbba047e58d07b6cce431b0263d879dfd3e45efa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 415, + 388, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 388, + 427 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 388, + 427 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 435, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 435, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 435, + 555, + 651 + ], + "type": "text", + "content": "In this work, we propose a novel open-set point cloud learning method, SASep, which identifies unknown classes by emphasizing the importance of utilizing the geometric structures of objects' different parts, rather than treating them as a whole. Specifically, we design a tunable semantic decomposition module to decompose objects into the semantically important and unimportant parts for classification. Then, we introduce a geometric synthesis strategy and a synth-aided margin separation module to leverage these decomposed parts, improving feature learning for known classes and enhancing class distinction. Experimental results show that SASep significantly outperforms the state-of-the-art method in various datasets. In scenarios with severe corruption, our method may be limited due to challenges in accurately capturing semantics, motivating us to explore a more robust method for semantic decomposition. We hope SASep offers some new insights into this under-explored area." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 660, + 417, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 660, + 417, + 673 + ], + "spans": [ + { + "bbox": [ + 314, + 660, + 417, + 673 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 713 + ], + "type": "text", + "content": "This work was supported by the National Key Research and Development Program for Young Scientists of China (2024YFB3310100), the China National Nat" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27302" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 139 + ], + "type": "text", + "content": "ural Science Foundation No. 62202182, the Guangdong Basic and Applied Basic Research Foundation 2024A1515010224, the National Key Research and Development Program of China under Grant 2023YFB4503400, and the China National Natural Science Foundation Nos. 62176101, 62276109, 62450064, and 62322205." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 182, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 182, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 294, + 236 + ], + "type": "text", + "content": "[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning - benchmarking and understanding semantic novelty detection on point clouds. In NeurIPS, pages 21228-21240. Curran Associates, Inc., 2022. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 238, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 295, + 282 + ], + "type": "text", + "content": "[2] Ayush Bhardwaj, Sakshee Pimpale, Saurabh Kumar, and Biplab Banerjee. Empowering knowledge distillation via open set recognition for robust 3d point cloud classification. Pattern Recognition Letters, 151:172-179, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 284, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 284, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 295, + 338 + ], + "type": "text", + "content": "[3] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 340, + 295, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 295, + 383 + ], + "type": "text", + "content": "[4] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial Reciprocal Points Learning for Open Set Recognition. IEEE TPAMI, 44(11):8065-8081, 2022. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 385, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 418 + ], + "type": "text", + "content": "[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 498 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In ECCV, pages 741-757, Cham, 2020. Springer International Publishing. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "type": "text", + "content": "[8] Xuefeng Du, Zhaoning Wang, Mu Cai, and Sharon Li. Towards unknown-aware learning with virtual outlier synthesis. In ICLR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "type": "text", + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 568, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 295, + 622 + ], + "type": "text", + "content": "[10] Dario Fontanel, Fabio Cermelli, Massimiliano Mancini, and Barbara Caputo. Detecting anomalies in semantic segmentation with prototypes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 113-121, 2021. 3, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[12] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, page 4995-5004, New York, NY, USA, 2024. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "type": "text", + "content": "[14] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In Proceedings of the 39th International Conference on Machine Learning, pages 8759-8773. PMLR, 2022. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[15] Jie Hong, Shi Qiu, Weihao Li, Saeed Anwar, Mehrtash Harandi, Nick Barnes, and Lars Petersson. Pointcam: Cut-and-mix for open-set point cloud learning. arXiv preprint arXiv:2212.02011, 2022. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[16] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, pages 677-689. Curran Associates, Inc., 2021. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 264, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 329 + ], + "type": "text", + "content": "[17] Inhyuk Jo, Jungtaek Kim, Hyohyeong Kang, Yong-Deok Kim, and Seungjin Choi. Open set recognition by regularising classifier with fake data generated by generative adversarial networks. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2686–2690, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[18] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In NeurIPS, pages 18661-18673. Curran Associates, Inc., 2020. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 407 + ], + "type": "text", + "content": "[19] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, pages 813-822, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "text", + "content": "[20] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, pages 8500-8509, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[21] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "text", + "content": "[22] Dogyoon Lee, Jaeha Lee, Junhyeop Lee, Hyeongmin Lee, Minhyeok Lee, Sungmin Woo, and Sangyoun Lee. Regularization strategy for point cloud via rigidly mixed sample. In CVPR, pages 15900-15909, 2021. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 565 + ], + "type": "text", + "content": "[23] Huan Lei, Naveed Akhtar, and Ajmal Mian. Octree guided cnn with spherical kernels for 3d point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 599 + ], + "type": "text", + "content": "[24] Shiyu Liang, Yixuan Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[25] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In CVPR, pages 20402-20411, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[26] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In ICIP, pages 3118-3122, 2021. 2, 3, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27303" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[29] Ravish Mehra, Pushkar Tripathi, Alla Sheffer, and Niloy J Mitra. Visibility of noisy point cloud data. Computers & Graphics, 34(3):219-230, 2010. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[30] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng-Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In ECCV, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[31] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[32] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[33] Paolo Rabino, Antonio Alliegro, and Tatiana Tommasi. 3d semantic novelty detection via large-scale pre-trained models. IEEE Access, 12:135352-135361, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "text", + "content": "[34] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "text", + "content": "[36] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, pages 144-157. Curran Associates, Inc., 2021. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 430 + ], + "type": "text", + "content": "[37] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[38] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 479, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 510 + ], + "type": "text", + "content": "[39] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4), 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "text", + "content": "[40] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 599 + ], + "type": "text", + "content": "[41] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In ICCV, pages 9302-9311, 2021. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 633 + ], + "type": "text", + "content": "[42] Yu Wang, Junxian Mu, Pengfei Zhu, and Qinghua Hu. Exploring diverse representations for open set recognition. AAAI, 38(6):5731-5739, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "text", + "content": "[43] Tingyu Weng, Jun Xiao, Hao Pan, and Haiyong Jiang. Partcom: Part composition learning for 3d open-set recognition. IJCV, 132(4):1393-1416, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[44] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 150 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[45] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[46] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27304" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_content_list.json b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..089f77dcba595583f86a219c007ffbebe205d0fc --- /dev/null +++ b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_content_list.json @@ -0,0 +1,1408 @@ +[ + { + "type": "text", + "text": "SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens", + "text_level": 1, + "bbox": [ + 191, + 130, + 805, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chi Su1 Xiaoxuan Ma1,☑ Jiajun Su5 Yizhou Wang1,2,3,4,☑", + "bbox": [ + 256, + 210, + 759, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Center on Frontiers of Computing Studies, School of Computer Science, Peking University \n $^{2}$ Inst. for Artificial Intelligence, Peking University $^{3}$ Nat'l Eng. Research Center of Visual Technology \n $^{4}$ State Key Laboratory of General Artificial Intelligence, Peking University \n $^{5}$ International Digital Economy Academy (IDEA)", + "bbox": [ + 101, + 231, + 915, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "suchi@stu.pku.edu.cn, {maxiaoxuan, yizhou.wang}@pku.edu.cn, sujiajun@idea.edu.cn Corresponding authors", + "bbox": [ + 153, + 315, + 861, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50f12f786f864623e3d98e0bc9316ad10c2c51febf4a4279ac54bb7aa782954e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 109, + 375, + 500, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef0882c895bd45834df76479317ffebbcc4c36391a5d247de3a5b86540791457.jpg", + "image_caption": [ + "(b)", + "Figure 1. (a) We propose scale-adaptive tokens in our one-stage framework for real-time multi-person 3D mesh estimation. Our method introduces scale-adaptive tokens, dynamically adjusted based on the relative size of individuals in the image, to more efficiently encode features, enabling real-time and accurate multi-person mesh estimation. We present a conceptual visualization of the scale-adaptive tokens. The right column visualizes the predicted meshes projected onto an image from 3DPW [49] dataset and from an elevated view. (b) Comparison of estimation error and inference time across different methods, with input resolutions in parentheses. Our method, using a mixed resolution with a base resolution of 644, achieves comparable performance to state-of-the-art methods on AGORA [33] test set while maintaining real-time inference efficiency. Code and models are available at https://ChiSu001.github.io/SAT-HMR/." + ], + "image_footnote": [], + "bbox": [ + 527, + 378, + 893, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 710, + 325, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose SAT-HMR, a one-stage framework for real-time multi-person 3D human mesh estimation from a single RGB image. While current one-stage methods, which follow a DETR-style pipeline, achieve state-of-the-art (SOTA) performance with high-resolution inputs, we observe that this particularly benefits the estimation of individuals in smaller scales of the image (e.g., those of young age or far from the camera), but at the cost of significantly increased computation overhead. To address this, we introduce scale-adaptive tokens that are dynamically adjusted based on the relative", + "bbox": [ + 88, + 750, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scale of each individual in the image within the DETR framework. Specifically, individuals in smaller scales are processed at higher resolutions, larger ones at lower resolutions, and background regions are further distilled. These scale-adaptive tokens more efficiently encode the image features, facilitating subsequent decoding to regress the human mesh, while allowing the model to allocate computational resources more effectively and focus on more challenging cases. Experiments show that our method preserves the accuracy benefits of high-resolution processing while substantially reducing computational cost, achieving real-time inference with performance comparable to SOTA methods.", + "bbox": [ + 511, + 712, + 908, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16796", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14f3515040cc45f35aa5cab6495e04b8f5095170809d1f19ee2e268cf7936c12.jpg", + "image_caption": [ + "Figure 2. Estimation errors and FPS of baselines with different resolutions and our method across individuals at various scales. The scale of an individual refers to the person's size relative to the overall image and please refer to Sec. 3.3 for mathematical definition. The colored lines show the Mean Vertex Error (MVE) errors (left y-axis) of the baselines with different resolutions (Res.) on the AGORA [33] validation set. The colored markers on the right y-axis indicate the FPS of the corresponding models. Our method adopts a mixed resolution with a base resolution of 644." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 472, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 463, + 222, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-person 3D human mesh estimation from a single RGB image aims to localize all individuals in the scene and estimate their 3D meshes, typically represented by a parametric human model such as SMPL [24]. This is a fundamental task with broad applications, spanning social interaction [42, 60], game production [61], etc. Existing methods primarily focus on addressing challenges such as monocular depth ambiguity and inter-person interactions, and can be broadly categorized into two types: multi-stage and one-stage approaches. Multi-stage methods [7, 14, 35, 55, 56] typically detect and process each individual by cropping and resizing them to a uniform, relatively high resolution, which generally improves performance. However, this overlooks the global context, such as the original scales of individuals and their relative positions within the entire image, potentially leading to suboptimal results.", + "bbox": [ + 88, + 489, + 485, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, one-stage methods gradually emerge as a promising alternative [2, 44-46]. These approaches operate on the entire image without explicit cropping, preserving global context and enabling end-to-end training. Building on the success of DETR [5] in object detection, recent works [2, 44] adopt a DETR-style pipeline. They mainly focus on designing queries and decoders to estimate the human mesh, e.g. Multi-HMR [2] uses queries to represent each person and decodes the corresponding human mesh. However, these methods rely on high-resolution inputs to achieve state-of-the-art (SOTA) results, as shown", + "bbox": [ + 89, + 734, + 486, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Fig. 1 (b), which significantly increases computational cost and limits real-time processing capabilities.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the findings of [2, 44], we observe that high-resolution input plays a crucial role in improving accuracy. To explore this, we design a naive baseline following similar architecture and analyze how its estimation error changes for individuals of different scales as we increase the input resolution. We define scale as the size of the person relative to the entire image1. As shown in Fig. 2, increasing the input resolution consistently leads to a decrease in the estimation error, particularly for smaller-scale individuals (e.g., kids, those far from the camera or those in closed-poses). Notably, for individuals in the $0 - 10\\%$ scale, we observe an error decrease of nearly $35\\mathrm{mm}$ when comparing the highest-resolution model (blueviolet) to the lowest-resolution one (yellow). However, processing the entire image at high resolution incurs significant computational costs, especially when individuals are close to the camera and occupy a large portion of the image. In such cases, these individuals are already represented by many tokens, and using more tokens with high resolution would only greatly increase the cost while offering marginal improvement (i.e., a slight error reduction in cases where individuals cover $30 + \\%$ of the image in Fig. 2). Therefore, both our best baseline model (blueviolet) and recent SOTA methods [2, 44] in Fig. 1 (b) fail to achieve real-time inference and run at about 5 FPS, highlighting a common limitation.", + "bbox": [ + 511, + 122, + 908, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we extend our baseline built on DETR [5] and shift our focus toward more efficient feature encoding. Specifically, we introduce scale-adaptive tokens, which are dynamically adjusted based on the scale of individuals in the image. We first predict a patch-level scale map, assigning a scale value to each token. Tokens corresponding to smaller-scale individuals are replaced with higher-resolution tokens. Additionally, background tokens identified by the scale map are further distilled to reduce computational cost. These scale-adaptive tokens capture individuals at varying scales with the appropriate level of detail, as shown in Fig. 1 (a). The tokens are then processed by the decoder and prediction heads to regress the human mesh. By efficiently allocating computational resources to the more challenging cases, our method strikes a better balance between accuracy and speed. As shown in Fig. 2, our method (red) achieves performance comparable to high-resolution input while maintaining real-time inference speed at 24 FPS. Experiments on benchmark datasets further demonstrate the effectiveness of our method, achieving competitive performance with high efficiency.", + "bbox": [ + 511, + 500, + 908, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conclude, our contributions are as follows:", + "bbox": [ + 532, + 801, + 836, + 815 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose scale-adaptive tokens that are dynamically adjusted to handle individuals of varying scales, enabling more efficient feature encoding.", + "- By employing scale-adaptive tokens, our method achieves" + ], + "bbox": [ + 513, + 816, + 903, + 876 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "See Eq. (1) for detailed definition of scale.", + "bbox": [ + 531, + 886, + 763, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16797", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance comparable to using full-image high resolution, but with significantly lower computational cost.", + "bbox": [ + 102, + 90, + 483, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our approach achieves SOTA performance with up to 5x the speed of the top-performing methods, making it the best real-time model for multi-person 3D mesh estimation, running at 24 FPS.", + "bbox": [ + 91, + 121, + 483, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 196, + 228, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Multi-Person Human Mesh Estimation", + "text_level": 1, + "bbox": [ + 89, + 223, + 424, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-person human mesh estimation from a single RGB image aims to regress all human meshes, typically represented by parametric models such as SMPL [24] and SMPL-X [34]. Existing methods can be categorized into multi-stage and one-stage approaches. Multi-stage methods [7, 11, 14, 35, 53, 55, 56] use off-the-shelf human detectors [38] to crop each person's region, followed by single-person mesh estimation [26, 27, 50, 57, 59]. This approach preserves relatively high-resolution images for each cropped region, generally leading to higher accuracy [44]. However, its reliance on detected bounding boxes and the lack of global context make it challenging to handle occlusion [45].", + "bbox": [ + 89, + 246, + 483, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, one-stage methods [2, 36, 44-46] estimate human meshes for all individuals by encoding the entire image and predicting simultaneously. Pioneer work like ROMP [45] and BEV [46] use CNN backbones [12] to extract global features for mesh regression. However, the low input resolution limits the expressiveness of the distilled features [44], leading to suboptimal performance. Recent methods like AiOS [44] and Multi-HMR [2] adopt DETR-style architectures to achieve SOTA performance at higher resolutions, but with significant computational overhead, making real-time use impractical (e.g., Multi-HMR runs at only 4 FPS as shown in Fig. 1). Instead, we argue that tokenizing the entire image at high resolution is unnecessary. By adjusting patch resolution based on the scale and focusing computational resources on more challenging areas, we achieve real-time inference with competitive performance.", + "bbox": [ + 89, + 428, + 483, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Transformers in Human Mesh Estimation", + "text_level": 1, + "bbox": [ + 89, + 681, + 447, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers [48], originally designed for sequential inputs, have demonstrated strong performance in various vision tasks, such as image classification with Vision Transformers (ViTs) [8]. And it has been widely adopted in human mesh estimation [2, 6, 9, 18, 19, 44, 52]. For single-person settings, methods such as [6, 9, 18, 19, 52] focus on improving the accuracy through various architectural designs. Inspired by the DETR series [5, 23, 58] in object detection, recent methods have adopted a DETR-style framework for multi-person pose and mesh estimation [2, 22, 43, 44, 51], enabling one-stage estimation for all individuals. These works primarily focus on designing decoder queries to represent humans, keypoints, and mesh parameters. For instance,", + "bbox": [ + 89, + 704, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AiOS [44] refines predictions by progressively filtering and expanding queries in the decoder, while Multi-HMR [2] uses each query to represent a detected person. Our method also builds upon the DETR framework but focuses on improving token representations in the encoder to capture scale-aware human features, enabling accurate 3D mesh regression.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Efficient Vision Transformers", + "text_level": 1, + "bbox": [ + 511, + 191, + 777, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although ViTs [8] achieves significant success across various vision tasks, their quadratic computational complexity remains a major challenge, particularly for tasks that require dense predictions such as object detection [47]. As a result, efforts to accelerate ViTs primarily focus on reducing the overall number of image tokens. For instance, [30, 37, 54] drop less informative tokens, [4, 39] merge similar tokens, and [41] proposes a mixed-resolution image tokenization scheme. Recently, TORE [9] accelerates single-person 3D mesh estimation by pruning background tokens. However, since background still provides important context, entirely discarding it leads to performance degradation (see Sec. 4.4). In this work, we propose a new perspective with scale-adaptive tokens that achieve a more effective trade-off between performance and computational cost.", + "bbox": [ + 511, + 213, + 906, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 453, + 604, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 479, + 653, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use SMPL [24] as our body model $\\mathcal{M}$ , which uses pose and shape parameters to represent the human body. Specifically, the pose parameters $\\theta \\in \\mathbb{R}^{24\\times 3}$ are the relative rotations of 24 body joints, and the shape parameters $\\beta \\in \\mathbb{R}^{10}$ describe the body shape. The model outputs a human mesh $\\mathbf{V} = \\mathcal{M}(\\pmb {\\theta},\\pmb {\\beta})\\in \\mathbb{R}^{6890\\times 3}$ and the 3D joints can be obtained via $\\mathbf{J} = \\mathbf{MV}\\in \\mathbb{R}^{J\\times 3}$ , where $\\mathbf{M}\\in \\mathbb{R}^{J\\times 6890}$ is the joint regressor, and $J$ represents the number of joints.", + "bbox": [ + 511, + 500, + 906, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 511, + 631, + 624, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on recent methods [2, 44], we utilize a DETR-style pipeline [5], consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters.", + "bbox": [ + 511, + 652, + 906, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Naive baseline. While high-resolution inputs are shown to significantly enhance estimation accuracy [2, 44], they also introduce considerable computational overhead. To further explore the trade-off between resolution and efficiency, we first introduce a simple baseline using a DETR-style architecture, as depicted in Fig. 3 (top).", + "bbox": [ + 511, + 704, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an RGB image $\\mathbf{I}_{\\mathrm{hr}} \\in \\mathbb{R}^{H_{\\mathrm{hr}} \\times W_{\\mathrm{hr}} \\times 3}$ , it is divided into regular patches of size $P \\times P$ . Patch embedding and positional embedding are then applied to these image patches, resulting in a sequence of image feature tokens $\\mathcal{T} = \\{t_1, t_2, \\dots, t_k\\}$ , where $k = H_{\\mathrm{hr}} / P \\times W_{\\mathrm{hr}} / P$ . These tokens are subsequently processed by a Transformer encoder. In the decoder, a set of human queries is initialized", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16798", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5cd82588e1232c0c4c1cedbbfe296a314ec9874064ec616bf3df315ff3f4b0f4.jpg", + "image_caption": [ + "Figure 3. Overview of (top) the baseline method and (bottom) our method with scale-adaptive tokens. Top: Our baseline method adopts a DETR-style [5] pipeline consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters. Bottom: Our method focuses on efficient feature encoding using scale-adaptive tokens. Specifically, low-resolution and high-resolution patches are extracted from the input images $\\mathbf{I}$ and $\\mathbf{I}_{\\mathrm{hr}}$ , respectively. A scale head network predicts a patch-level scale map $\\mathbf{S}$ from the low-resolution tokens, classifying them into three categories: background, small-scale, and large-scale. This scale map guides the pruning and pooling of low-resolution tokens $\\mathcal{T}_{\\mathrm{LR}}$ and indicates which patches should be replaced by high-resolution ones. By concatenating the pooled background tokens $\\mathcal{T}_{\\mathrm{B}}^{\\prime}$ , the remaining large-scale low-resolution tokens $\\mathcal{T}_{\\mathrm{LARGE}}$ , and the high-resolution tokens $\\mathcal{T}_{\\mathrm{HR}}$ , we obtain scale-adaptive tokens $\\mathcal{T}_{\\mathrm{SA}}$ . These tokens are then processed by the encoder, decoder, and multiple prediction heads to regress the human mesh." + ], + "image_footnote": [], + "bbox": [ + 114, + 90, + 885, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as $\\mathcal{Q} = \\{q_1, q_2, \\ldots, q_n\\}$ , where $n$ is a predefined hyperparameter large enough to detect all the individuals. After being processed by the decoder, these human queries are passed through multiple prediction heads: pose head $\\mathcal{H}_{\\mathrm{p}}$ , shape head $\\mathcal{H}_{\\mathrm{s}}$ , translation head $\\mathcal{H}_{\\mathrm{t}}$ , and box head $\\mathcal{H}_{\\mathrm{b}}$ to predict SMPL parameters (pose $\\theta$ , shape $\\beta$ , and 3D translation $\\pmb{t}$ ) and bounding boxes, respectively. Each of the four prediction heads is implemented as an Multi-Layer Perceptron (MLP). Additionally, a linear projection is applied to the human queries to compute confidence scores, which are used to filter valid predictions with a threshold $\\alpha_{\\mathrm{d}}$ .", + "bbox": [ + 88, + 537, + 485, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. We evaluate the baseline with different image resolutions as input (see Fig. 2, Sec. 4.4 for details) and observe that high-resolution inputs significantly improve estimation accuracy, particularly for small-scale individuals whose size is relatively small to the image. However, for large-scale individuals, e.g., those closer to the camera, the benefits of higher resolution are much less pronounced. Despite these improvements, high-resolution inputs bring a substantial increase in computational cost, which is consistent with findings from prior work [2, 44] (see Fig. 1).", + "bbox": [ + 88, + 713, + 485, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This motivates us to explore more efficient strategies that balance performance and speed and we aim to encode the im", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age feature more efficiently from an earlier stage. For large-scale individuals, additional tokens may be unnecessary; instead, computational resources should be allocated more effectively. To address this, we introduce scale-adaptive tokens, which replace the uniformly partitioned tokens used in the baseline. In the following sections, we will explain how the scale is defined and computed, and how the tokens are adjusted accordingly (Sec. 3.3), followed by a discussion of the training process (Sec. 3.4).", + "bbox": [ + 511, + 537, + 908, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Scale-Adaptive Tokens", + "text_level": 1, + "bbox": [ + 511, + 681, + 723, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (bottom), we begin with a low-resolution image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $(H, W) = (H_{\\mathrm{hr}} / 2, W_{\\mathrm{hr}} / 2)$ , and first uniformly partition it into low-resolution patches. Our goal is to adjust the resolution of each patch based on the scale of the individual it represents. To achieve this, we introduce a patch-level scale map that guides the resolution adjustments for each patch. Specifically, this map determines which low-resolution tokens should be further subdivided and replaced by high-resolution ones, and which tokens can be compressed. The high-resolution patches are extracted from a corresponding high-resolution image $\\mathbf{I}_{\\mathrm{hr}}$ . After being processed by the Transformer layers, the remaining low-resolution tokens and the newly generated high-resolution", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16799", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens are combined into scale-adaptive tokens, which provide a more efficient representation with varying levels of detail. These scale-adaptive tokens are then passed through additional Transformer layers and, finally, to the decoder to regress the SMPL parameters.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch-level scale map. We define a patch-level scale map $\\mathbf{S} \\in [0,1]^{\\frac{H}{P} \\times \\frac{W}{P} \\times 2}$ for patches extracted from the low-resolution image $\\mathbf{I}$ . For each patch, the scale map $\\mathbf{S}(i,j) = (c,s)$ contains two values: $c$ represents the confidence indicating whether the patch overlaps with an individual, where $0$ denotes the background; $s$ represents the scale of the individual within the patch, defined as", + "bbox": [ + 89, + 171, + 483, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns = \\min \\left(d _ {\\mathrm {b b}} / S _ {\\mathrm {h r}}, 1\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 280, + 483, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{\\mathrm{bb}}$ is the diagonal length of the bounding box in $\\mathbf{I}_{\\mathrm{hr}}$ , and $S_{\\mathrm{hr}} = \\max (H_{\\mathrm{hr}},W_{\\mathrm{hr}})$ . If the patch overlaps with multiple individuals, we use the scale of the individual closest to the camera, i.e., the one with the smaller depth. The scale is only defined for patches overlapping with individuals, and for background patches, the scale value is ignored.", + "bbox": [ + 89, + 303, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To predict the patch-level scale map, we pass the low-resolution patches through a shallow transformer encoder with $N_{\\mathrm{lr}}$ layers, followed by a scale head network, implemented as an MLP. As illustrated in Fig. 3, patches without color are classified as background, while colored regions represent patches overlapping with individuals. A color gradient from blue to rose indicates the scale of the individuals, with blue representing large-scale individuals and rose representing small-scale individuals.", + "bbox": [ + 89, + 393, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scale-adaptive encoder. Building on the predicted scale map S, the low-resolution tokens $\\mathcal{T}_{\\mathrm{LR}} = \\{t_{1}^{\\mathrm{lr}}, t_{2}^{\\mathrm{lr}}, \\ldots, t_{k_{\\mathrm{lr}}}^{\\mathrm{lr}}\\}$ generated from the low-resolution patches are then classified into three groups: background $\\mathcal{T}_{\\mathrm{B}}$ , small-scale $\\mathcal{T}_{\\mathrm{SMALL}}$ , and large-scale $\\mathcal{T}_{\\mathrm{LARGE}}$ tokens. Tokens identified as background, based on a confidence score $c$ that meets a threshold $\\alpha_{c}$ , are marked as background tokens $\\mathcal{T}_{\\mathrm{B}} = \\{t_{1}^{\\mathrm{b}}, t_{2}^{\\mathrm{b}}, \\ldots, t_{k_{\\mathrm{b}}}^{\\mathrm{b}}\\}$ . The remaining tokens are further categorized by their scale values $s$ using a scale threshold $\\alpha_{s}$ and are assigned to either the small-scale category, denoted by $\\mathcal{T}_{\\mathrm{SMALL}} = \\{t_{1}^{\\mathrm{small}}, t_{2}^{\\mathrm{small}}, \\ldots, t_{k_{\\mathrm{small}}}^{\\mathrm{small}}\\}$ , or the large-scale category, denoted by $\\mathcal{T}_{\\mathrm{LARGE}} = \\{t_{1}^{\\mathrm{large}}, t_{2}^{\\mathrm{large}}, \\ldots, t_{k_{\\mathrm{large}}}^{\\mathrm{large}}\\}$ .", + "bbox": [ + 89, + 535, + 483, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Small-scale tokens are pruned and replaced by their high-resolution counterparts. High-resolution tokens are generated by applying a shallow transformer encoder with $N_{\\mathrm{hr}}$ layers to the high-resolution image $\\mathbf{I}_{\\mathrm{hr}}$ , where $N_{\\mathrm{hr}} = N_{\\mathrm{lr}}$ to ensure feature alignment. The small-scale tokens are then replaced by their corresponding high-resolution tokens (Fig. 3), resulting in an expanded set of high-resolution tokens, denoted as $\\mathcal{T}_{\\mathrm{HR}} = \\{t_1^{\\mathrm{hr}}, t_2^{\\mathrm{hr}}, \\dots, t_{k_{\\mathrm{hr}}}^{\\mathrm{hr}}\\}$ , where $k_{\\mathrm{hr}} = 4k_{\\mathrm{small}}$ .", + "bbox": [ + 89, + 720, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although background can provide valuable contextual information, their feature encoding can be further optimized. To this end, we distill the background tokens $\\mathcal{T}_{\\mathrm{B}}$ by spatially pooling every four neighboring tokens, resulting in", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pooled background tokens $\\mathcal{T}_{\\mathrm{B}}^{\\prime} = \\{t_{1}^{\\mathrm{b}^{\\prime}}, t_{2}^{\\mathrm{b}^{\\prime}}, \\ldots, t_{k_{\\mathrm{b}}^{\\prime}}^{\\mathrm{b}^{\\prime}}\\}$ . To avoid irregular pooling, some low-resolution tokens that cannot be grouped remain unchanged. As a result, $k_{\\mathrm{b}}^{\\prime}$ does not necessarily equal $k_{\\mathrm{b}} / 4$ . By retaining these tokens, we better preserve their positional information, which facilitates subsequent positional encoding operations [23].", + "bbox": [ + 511, + 89, + 903, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The remaining tokens in $\\mathcal{T}_{\\mathrm{LR}}$ representing large-scale individuals, $\\mathcal{T}_{\\mathrm{LARGE}}$ , remain unchanged. Next, we integrate the different tokens, including pooled background tokens and high-resolution tokens from selected regions. This results in scale-adaptive tokens, denoted as $\\mathcal{T}_{\\mathrm{SA}} = \\{\\mathcal{T}_{\\mathrm{B}}^{\\prime}, \\mathcal{T}_{\\mathrm{LARGE}}, \\mathcal{T}_{\\mathrm{HR}}\\}$ . Compared to the uniform low-resolution tokens, this approach allocates feature details more efficiently, preserving different levels of detail for different individuals and regions.", + "bbox": [ + 511, + 184, + 906, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These tokens, $\\mathcal{T}_{\\mathrm{SA}}$ , are then processed by another Transformer encoder with $N_{\\mathrm{sa}}$ layers and further decoded in subsequent stages. Fig. 3 (bottom) provides an overview of the scale-adaptive encoder.", + "bbox": [ + 511, + 304, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoder and prediction heads. The decoder and prediction heads follow the architecture of our baseline approach as mentioned in Sec. 3.2. For detailed architectures, please refer to the supplementary material (Sec. A.1).", + "bbox": [ + 511, + 369, + 903, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Losses", + "text_level": 1, + "bbox": [ + 513, + 440, + 669, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to DETR [5], we first match our human predictions to ground-truth (GT) using Hungarian Matching, and we leverage bounding boxes, confidence scores, and projected joints during the matching. This matching step is necessary before computing the overall loss $\\mathcal{L}$ . More details on the matching process are provided in Sec. A.1. The loss function, $\\mathcal{L}$ , is a weighted sum of various terms, with each $\\lambda$ representing a hyperparameter:", + "bbox": [ + 511, + 462, + 905, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {m a p}} \\mathcal {L} _ {\\text {m a p}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {p o s e}} \\mathcal {L} _ {\\text {p o s e}} + \\lambda_ {\\text {s h a p e}} \\mathcal {L} _ {\\text {s h a p e}} \\\\ + \\lambda_ {\\mathrm {j 3 d}} \\mathcal {L} _ {\\mathrm {j 3 d}} + \\lambda_ {\\mathrm {j 2 d}} \\mathcal {L} _ {\\mathrm {j 2 d}} + \\lambda_ {\\mathrm {b o x}} \\mathcal {L} _ {\\mathrm {b o x}} + \\lambda_ {\\det} \\mathcal {L} _ {\\det}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 592, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scale map loss $\\mathcal{L}_{\\mathrm{map}}$ . To supervise the patch-level scale map, we apply focal loss [21] to the predicted confidence $c$ and L1 loss to the predicted scale $s$ . The total loss, $\\mathcal{L}_{\\mathrm{map}}$ , is the sum of these two terms. The GT for confidence and scale is precomputed, as described in Sec. 3.3.", + "bbox": [ + 511, + 638, + 905, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depth loss $\\mathcal{L}_{\\mathrm{depth}}$ . We supervise the root depth (i.e., the z-dimension) of the 3D joints $\\mathbf{J}$ , regressed from the predicted mesh $\\mathbf{V}$ (Sec. 3.1). Specifically, we normalize the depth based on the focal length, following [10]. The depth loss is calculated as $\\mathcal{L}_{\\mathrm{depth}} = \\left|\\frac{1}{\\tilde{d}} - \\frac{f}{\\tilde{f}d}\\right|$ , where $f$ is the predefined focal length of our camera, and $\\tilde{d}$ and $\\tilde{f}$ are the GT depth and focal length, respectively. See Sec. A.1 for more details.", + "bbox": [ + 511, + 719, + 905, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Other losses. For the pose parameters loss $\\mathcal{L}_{\\mathrm{pose}}$ , shape parameters loss $\\mathcal{L}_{\\mathrm{shape}}$ , 3D joints loss $\\mathcal{L}_{\\mathrm{j3d}}$ , and projected 2D joints loss $\\mathcal{L}_{\\mathrm{j2d}}$ , we use the L1 distance. For the bounding box loss $\\mathcal{L}_{\\mathrm{box}}$ , we combine L1 loss with GIoU loss [40].", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16800", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d6c6953e841e5bfbd18bb8b2c2e0666af6412a5e24f4aa5006de5500f86f62a1.jpg", + "table_caption": [ + "Table 1. Comparison with SOTA methods on AGORA test set [33]. \"Res.\" represents the input resolution. \"644*\" means we use a base resolution of 644 and adaptively scale certain regions up to a maximum resolution of 1288." + ], + "table_footnote": [], + "table_body": "
MethodRes.Time (ms)MACs (G)F1-Score ↑Precision ↑Recall ↑MPJPE ↓MVE ↓NMJE ↓NMVE ↓
ROMP [45]51238.743.60.910.950.88108.1103.4118.8113.6
BEV [46]51250.648.90.930.960.90105.3100.7113.2108.3
PSVT [36]512--0.93--97.794.1105.1101.2
AiOS [44]1333405.2314.50.940.980.9063.957.568.061.2
Multi-HMR [2]89697.42075.10.93--82.877.689.083.4
Multi-HMR [2]1288231.76104.60.950.970.9365.361.168.764.3
Ours644*42.0133.10.950.980.9167.963.371.566.6
", + "bbox": [ + 137, + 128, + 864, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The detection loss $\\mathcal{L}_{\\mathrm{det}}$ uses focal loss [21] to supervise the confidence scores of human queries.", + "bbox": [ + 89, + 268, + 482, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 311, + 225, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and metrics", + "text_level": 1, + "bbox": [ + 89, + 335, + 287, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model on multi-person datasets including AGORA [33], BEDLAM [3], COCO [20], Crowdpose [17], and MPII [1], as well as the single-person dataset H3.6M [13]. AGORA and BEDLAM are synthetic datasets with high-quality 3D mesh GTs. For other multi-person datasets, we use pseudo GTs from NeuralAnnot [31] and only supervise projected 2D joints.", + "bbox": [ + 89, + 357, + 483, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For comparison with SOTA methods, we evaluate our method on AGORA [33], MuPoTS [29], 3DPW [49], and CMU Panoptic [15] datasets. Following previous works [2, 44-46], we report Mean Vertex Error (MVE) and Mean Per-Joint Position Error (MPJPE) before and after Procrustes-Alignment (PA). On AGORA, we also report F1-Score, precision, and recall for evaluating detection accuracy; Normalized Mean Vertex Error (NMVE) and Normalized Mean Joint Error (NMJE) that considered regression accuracy with detection accuracy. All metrics above, except detection accuracy, are reported in millimeters (mm). On MuPoTS-3D [29], we report the Percentage of Correctly estimated Keypoints (PCK) using a threshold of $15\\mathrm{cm}$ . To evaluate computational costs, we report Multiply-Add Cumulation (MACs) and the average inference time (ms) following [2] on one RTX 3090 GPU.", + "bbox": [ + 89, + 463, + 483, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 712, + 307, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The input image is resized to maintain its original aspect ratio, with a base resolution of $\\max(H, W) = 644$ and $\\max(H_{\\mathrm{hr}}, W_{\\mathrm{hr}}) = 1288$ for high-resolution when needed. We denote this mixed-resolution setting as \"644*\". The Transformer layers in our encoder are initialized using pretrained DINOv2 [32] with a patch size of $P = 14$ . We use ViT-B architecture as ViT-L significantly increases inference time (42.0 ms vs. 74.2 ms). The Transformer used in the encoder has $N_{\\mathrm{lr}} = N_{\\mathrm{hr}} = 3$ , and $N_{\\mathrm{sa}} = 9$ layers, respectively. Our decoder consists of 6 Transformer layers and 50 human queries. Training is conducted on 8 RTX 3090 GPUs with a", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/faa4a6ea6f467b2c5c6c9a10fa540b838ad772b6fc04de9859b470832268d6b3.jpg", + "table_caption": [ + "Table 2. Comparison with SOTA methods on 3DPW [49] and MuPoTS [29] test sets. “Res.” represents the input resolution. “Crop” refers to cropping individuals for single-person mesh estimation, i.e., a multi-stage approach. “644*” means a base resolution of 644, with adaptive scaling up to 1288." + ], + "table_footnote": [], + "table_body": "
MethodRes.3DPWMuPoTS (PCK)
PA-MPJPE ↓MVE ↓All ↑Matched ↑
CRMH [14]832--69.172.2
3DCrowdNet [7]Crop51.598.372.773.3
ROMP [45]51247.393.469.972.2
BEV [46]51246.992.370.275.2
PSVT [36]51245.784.9--
Multi-HMR [2]89641.775.985.089.3
Ours644*41.673.789.090.1
", + "bbox": [ + 521, + 347, + 903, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "total batch size of 40 for 60 epochs with an initial learning rate of $4e - 5$ . Please refer to Sec. A.2 for more details.", + "bbox": [ + 511, + 512, + 906, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison to the State-of-the-arts", + "text_level": 1, + "bbox": [ + 511, + 558, + 823, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance and Efficiency. We conduct evaluations on the AGORA [33], 3DPW [49], MuPoTS [28], and CMU Panoptic [15] datasets. Tab. 1 shows the results of our method compared to SOTA methods on the AGORA [33] test set leaderboard. Our method, using a mixed resolution input with a base of 644, achieves performance on par with the top SOTA methods [2, 44], despite their use of higher resolutions and larger models (e.g., Multi-HMR use ViT-L [32] model). These methods, however, result in significantly higher computational complexity, including much higher MACs and slower runtime, compromising real-time performance. Besides, these methods [2, 44] estimates SMPL-X [34], from which SMPL [24] mesh is derived. In contrast, our method maintains real-time computational efficiency while improving upon real-time methods ROMP [45], by nearly $40\\%$ . This establishes our approach as the leading real-time model, setting a new standard by achieving an unparalleled balance of performance and efficiency.", + "bbox": [ + 511, + 580, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 presents the results of our method on the 3DPW dataset. Following Multi-HMR [2], we finetune on the 3DPW training set. We further evaluate the generaliza", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db5ac48be8e1bdec5fe2071f768509f4940c5e1edd4a932afaf6af9e6f3858ec.jpg", + "image_caption": [ + "Figure 4. Comparison with SOTA methods [2, 44-46] on in-the-wild images from the Internet. Red dashed circles highlight areas with incorrect estimations. The third case is left blank due to the small scale of individuals. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 151, + 89, + 846, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/765b805e0f2a4fe287b4f5fb6e4862107dd5ead812f32786ec90ac94f1914891.jpg", + "table_caption": [ + "Table 3. Comparison with SOTA methods on CMU Panoptic [15] test set. We report MPJPE for four activities and the average." + ], + "table_footnote": [], + "table_body": "
MethodHaggling ↓Mafia ↓Ultimatum ↓Pizza ↓Avg. ↓
CRMH [14]129.6133.5153.0156.7143.2
3DCrowdNet [7]109.6135.9129.8135.6127.6
ROMP [45]110.8122.8141.6137.6128.2
BEV [46]90.7103.7113.1125.2109.5
PSVT [36]88.797.9115.2121.2105.7
Ours67.978.595.894.684.2
", + "bbox": [ + 98, + 366, + 483, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6daf1a86b3c79a0f77d78ca5374e236ae0576ce473de792a0539599ecc73c5ae.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of different resolutions for our baseline. Resolution for the baseline: (a) 518, (b) 1288. Red dashed circles highlight differences; zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 114, + 486, + 460, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion performance of our method on the MuPoTS and CMU Panoptic datasets, as shown in Tab. 2 and Tab. 3. For Panoptic dataset, we follow the common evaluation protocol [36, 45, 46] and exclude Multi-HMR [2] due to potential inconsistencies in joint format. Our method consistently achieves the best results across all datasets, demonstrating exceptional generalization capability.", + "bbox": [ + 89, + 650, + 483, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 4 presents a visual comparison between our method and existing SOTA approaches. The images are sourced from the Internet. Our method provides the most accurate overall estimations, especially in the third case, where individuals are far from the camera, resulting in very small scales. Despite this challenge, our method remains highly robust, delivering precise estimations. In contrast, ROMP [45] and BEV [46] fail to detect all individuals, while AiOS [44] and Multi-HMR [2], despite using", + "bbox": [ + 88, + 763, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/82ef2f4f24d2f7c8de3c274c4317616c5599a21cceda73a161e811d17654f3d1.jpg", + "table_caption": [ + "Table 4. Ablation studies on BEDLAM [3] validation set. We report MVE for different scale ranges and the average (Avg.)." + ], + "table_footnote": [], + "table_body": "
Ablation0-20%20-40%40-60%60-80%80%+Avg.
(a) Drop all59.955.161.566.670.857.2
(b) No pooling60.354.557.059.864.156.1
(c) Pooling×260.754.557.562.166.556.3
(d) Ours60.054.457.260.362.756.0
", + "bbox": [ + 531, + 366, + 890, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "higher-resolution models, show significant estimation errors.", + "bbox": [ + 511, + 460, + 906, + 476 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 488, + 666, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies using only the AGORA [33] and sampled BEDLAM [3] datasets to validate model designs, and report results on their validation sets.", + "bbox": [ + 511, + 511, + 906, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of resolution. Fig. 2 shows baseline results with different input resolutions on the AGORA [33] validation set. As resolution increases, the error across individuals in different scale ranges consistently decreases, with the most significant drop (35mm) in the smallest scale range. In contrast, the improvement in the large-scale range is less pronounced. However, the performance gain comes at the cost of a significant drop in computational efficiency, reducing the frame rate from real-time (32 FPS) to 5 FPS. Fig. 5 shows an example, clearly illustrating that higher resolution greatly improves estimation for small-scale individuals.", + "bbox": [ + 511, + 561, + 908, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background tokens $\\mathcal{T}_{\\mathrm{B}}$ . We investigate the impact of background tokens on model performance and report the MVE across different scale ranges on the BEDLAM validation set in Tab. 4. In (a), we discard all background tokens and observe stable performance at the smallest scale, but a decline as the scale increases, indicating that background context becomes more important for larger scales. In (b), skipping pooling for background tokens yields nearly the same performance but introduces higher computational cost. In (c), applying pooling twice with the same grouping strategy results in a performance drop for large-scale instances, highlighting", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16802", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bea354c11784882fb0b6c14e84c5533b86b14f97854ccd6e4bfcacaab7e9511f.jpg", + "image_caption": [ + "Figure 6. Visualization of our predicted patch-level scale maps S and scale-adaptive tokens $\\mathcal{T}_{\\mathrm{SA}}$ . We show the predicted scale maps, scale-adaptive tokens, and the estimated meshes overlaid on the image. In the scale map, colors represent scale values, with uncolored areas indicating background patches. A gradient from blue (large-scale individuals) to rose (small-scale individuals) illustrates the scale distribution. We visualize scale-adaptive tokens as patches of different sizes and colors on the image, with low-resolution patches blurred." + ], + "image_footnote": [], + "bbox": [ + 143, + 88, + 854, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87bbf41706ae5c72aa2656c33dc729d41efd5efc2787fc1094be52253b87cdde.jpg", + "image_caption": [ + "Figure 7. Qualitative results of our method. We visualize four cases from AGORA [33], BEDLAM [3], CMU Panoptic [15] and 3DPW [49] datasets, displaying the input, the estimated mesh overlay, and an elevated view (top to bottom)." + ], + "image_footnote": [], + "bbox": [ + 142, + 383, + 854, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the significance of background context for estimation. We hypothesize that large-scale people face more truncation, relying more on global context which requires further exploration. We provide additional efficiency comparisons in Sec. B, showing that pooling once achieves the best balance between performance and efficiency.", + "bbox": [ + 89, + 685, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 787, + 272, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 visualize the predicted patch-level scale map and scale-adaptive tokens by our model, enabling dynamic adjustment of tokens for accurate estimations. Fig. 7 presents additional visualizations of predicted meshes, demonstrating the effectiveness of our method in various scenarios. Please refer to Sec. B for more qualitative results and discussion.", + "bbox": [ + 89, + 809, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 684, + 633, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel one-stage framework for real-time multiperson 3D mesh estimation from an RGB image. By introducing scale-adaptive tokens that dynamically adjust based on the relative scale of individuals, our method effectively balances efficiency and accuracy, retaining the benefits of high-resolution processing while achieving real-time performance. Experiments demonstrate that our method is the best real-time model, offering competitive performance and superior generalization compared to SOTAs. Our limitations include the lack of height-aware scale estimation, which may cause depth errors, and the current focus on body-only mesh estimation, extendable to whole-body in the future.", + "bbox": [ + 511, + 708, + 908, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16803", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by 2022ZD0114904 and NSFC-6247070125. We thank Yiyang Wang for the valuable discussions during the early stage of this work.", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6, 1", + "[2] Fabien Baradel, Matthieu Armando, Salma Galaaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas. Multi-hmr: Multi-person whole-body human mesh recovery in a single shot. In ECCV, 2024. 2, 3, 4, 6, 7, 1, 5", + "[3] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 6, 7, 8, 1, 2, 3, 4, 5", + "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In ICLR, 2023. 3", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 3, 4, 5, 1", + "[6] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In ECCV, 2022. 3", + "[7] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3d human mesh from in-the-wild crowded scenes. In CVPR, pages 1475-1484, 2022. 2, 3, 6, 7", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[9] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. Tore: Token reduction for efficient human mesh recovery with transformer. In ICCV, pages 15143-15155, 2023. 3", + "[10] Jose M Facil, Benjamin Ummenhofer, Huizhong Zhou, Luis Montesano, Thomas Brox, and Javier Civera. Cam-convs: Camera-aware multi-scale convolutions for single-view depth. In CVPR, pages 11826-11835, 2019. 5", + "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 3, 1", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3", + "[13] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predic" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6, 1", + "[14] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In CVPR, pages 5579-5588, 2020. 2, 3, 6, 7", + "[15] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, pages 3334-3342, 2015. 6, 7, 8, 2, 3", + "[16] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1", + "[17] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In CVPR, pages 10863-10872, 2019. 6, 1", + "[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 3", + "[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, pages 12939-12948, 2021. 3", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6, 1", + "[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5, 6", + "[22] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, et al. Group pose: A simple baseline for end-to-end multi-person pose estimation. In ICCV, pages 15029-15038, 2023. 3", + "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 3, 5, 1, 2", + "[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3, 6, 1", + "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 1", + "[26] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In CVPR, pages 534-543, 2023. 3", + "[27] Xiaoxuan Ma, Jiajun Su, Yuan Xu, Wentao Zhu, Chunyu Wang, and Yizhou Wang. Vmarker-pro: Probabilistic 3d human mesh estimation from virtual markers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3", + "[28] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, pages 506-516, 2017. 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16804", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 6, 1, 3", + "[30] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 3", + "[31] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralnot: Neural annotator for 3d human mesh training sets. In CVPR, pages 2299-2307, 2022. 6, 1", + "[32] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 6", + "[33] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In CVPR, pages 13468-13478, 2021. 1, 2, 6, 7, 8, 3, 4, 5", + "[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 6, 1", + "[35] Zhongwei Qiu, Qiansheng Yang, Jian Wang, and Dongmei Fu. Dynamic graph reasoning for multi-person 3d pose estimation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3521-3529, 2022. 2, 3", + "[36] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In CVPR, pages 21254-21263, 2023. 3, 6, 7", + "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3", + "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3", + "[39] Cedric Renggli, André Susano Pinto, Neil Houlsby, Basil Mustafa, Joan Puigcerver, and Carlos Riquelme. Learning to merge tokens in vision transformers. arXiv preprint arXiv:2202.12015, 2022. 3", + "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5", + "[41] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In CVPRW, pages 4613-4622, 2023. 3", + "[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In CVPR, pages 11069-11078, 2022. 3", + "[44] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi-Sing Leung, Ziwei Liu, Lei Yang, et al. Aios: All-in-one-stage expressive human pose and shape estimation. In CVPR, pages 1834-1843, 2024, 2, 3, 4, 6, 7, 1, 5", + "[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 3, 6, 7, 1, 2", + "[46] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 2, 3, 6, 7, 1, 4", + "[47] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 3", + "[48] A Vaswani. Attention is all you need. NeurIPS, 2017. 3", + "[49] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 1, 6, 8, 3, 4", + "[50] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In CVPR, pages 979–989, 2024. 3", + "[51] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 3", + "[52] Sen Yang, Wen Heng, Gang Liu, Guozhong Luo, Wankou Yang, and YU Gang. Capturing the motion of every joint: 3d human pose and shape estimation with independent tokens. In ICLR, 2023. 3", + "[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In ECCV, 2022. 3", + "[54] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, pages 10809-10818, 2022. 3", + "[55] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 2, 3", + "[56] Andrei Zanfir, Elisabeta Marinoiu, Mihai Zanfir, Alin-Ionut Popa, and Cristian Sminchisescu. Deep network for the integrated 3d sensing of multiple people in natural images. In NeurIPS, 2018. 2, 3", + "[57] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 3", + "[58] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2023. 3" + ], + "bbox": [ + 514, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In CVPR, pages 15085-15099, 2023. 3", + "[60] Wentao Zhu, Jason Qin, Yuke Lou, Hang Ye, Xiaoxuan Ma, Hai Ci, and Yizhou Wang. Social motion prediction with cognitive hierarchies. In NeurIPS, 2023. 2", + "[61] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlun Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang: Human motion generation: A survey. IEEE TPAMI, 46(4): 2430-2449, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16806", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_model.json b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e656158aa68eb9475b374986a7d3d4c414018e29 --- /dev/null +++ b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.806, + 0.176 + ], + "angle": 0, + "content": "SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.211, + 0.761, + 0.231 + ], + "angle": 0, + "content": "Chi Su1 Xiaoxuan Ma1,☑ Jiajun Su5 Yizhou Wang1,2,3,4,☑" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.232, + 0.916, + 0.311 + ], + "angle": 0, + "content": "\\(^{1}\\) Center on Frontiers of Computing Studies, School of Computer Science, Peking University \n\\(^{2}\\) Inst. for Artificial Intelligence, Peking University \\(^{3}\\) Nat'l Eng. Research Center of Visual Technology \n\\(^{4}\\) State Key Laboratory of General Artificial Intelligence, Peking University \n\\(^{5}\\) International Digital Economy Academy (IDEA)" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.316, + 0.862, + 0.351 + ], + "angle": 0, + "content": "suchi@stu.pku.edu.cn, {maxiaoxuan, yizhou.wang}@pku.edu.cn, sujiajun@idea.edu.cn Corresponding authors" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.375, + 0.501, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.583, + 0.316, + 0.595 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.379, + 0.895, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.583, + 0.725, + 0.595 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.598, + 0.907, + 0.695 + ], + "angle": 0, + "content": "Figure 1. (a) We propose scale-adaptive tokens in our one-stage framework for real-time multi-person 3D mesh estimation. Our method introduces scale-adaptive tokens, dynamically adjusted based on the relative size of individuals in the image, to more efficiently encode features, enabling real-time and accurate multi-person mesh estimation. We present a conceptual visualization of the scale-adaptive tokens. The right column visualizes the predicted meshes projected onto an image from 3DPW [49] dataset and from an elevated view. (b) Comparison of estimation error and inference time across different methods, with input resolutions in parentheses. Our method, using a mixed resolution with a base resolution of 644, achieves comparable performance to state-of-the-art methods on AGORA [33] test set while maintaining real-time inference efficiency. Code and models are available at https://ChiSu001.github.io/SAT-HMR/." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.711, + 0.326, + 0.726 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We propose SAT-HMR, a one-stage framework for real-time multi-person 3D human mesh estimation from a single RGB image. While current one-stage methods, which follow a DETR-style pipeline, achieve state-of-the-art (SOTA) performance with high-resolution inputs, we observe that this particularly benefits the estimation of individuals in smaller scales of the image (e.g., those of young age or far from the camera), but at the cost of significantly increased computation overhead. To address this, we introduce scale-adaptive tokens that are dynamically adjusted based on the relative" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.909, + 0.895 + ], + "angle": 0, + "content": "scale of each individual in the image within the DETR framework. Specifically, individuals in smaller scales are processed at higher resolutions, larger ones at lower resolutions, and background regions are further distilled. These scale-adaptive tokens more efficiently encode the image features, facilitating subsequent decoding to regress the human mesh, while allowing the model to allocate computational resources more effectively and focus on more challenging cases. Experiments show that our method preserves the accuracy benefits of high-resolution processing while substantially reducing computational cost, achieving real-time inference with performance comparable to SOTA methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.473, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.487, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Estimation errors and FPS of baselines with different resolutions and our method across individuals at various scales. The scale of an individual refers to the person's size relative to the overall image and please refer to Sec. 3.3 for mathematical definition. The colored lines show the Mean Vertex Error (MVE) errors (left y-axis) of the baselines with different resolutions (Res.) on the AGORA [33] validation set. The colored markers on the right y-axis indicate the FPS of the corresponding models. Our method adopts a mixed resolution with a base resolution of 644." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.223, + 0.48 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Multi-person 3D human mesh estimation from a single RGB image aims to localize all individuals in the scene and estimate their 3D meshes, typically represented by a parametric human model such as SMPL [24]. This is a fundamental task with broad applications, spanning social interaction [42, 60], game production [61], etc. Existing methods primarily focus on addressing challenges such as monocular depth ambiguity and inter-person interactions, and can be broadly categorized into two types: multi-stage and one-stage approaches. Multi-stage methods [7, 14, 35, 55, 56] typically detect and process each individual by cropping and resizing them to a uniform, relatively high resolution, which generally improves performance. However, this overlooks the global context, such as the original scales of individuals and their relative positions within the entire image, potentially leading to suboptimal results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.487, + 0.901 + ], + "angle": 0, + "content": "To address these limitations, one-stage methods gradually emerge as a promising alternative [2, 44-46]. These approaches operate on the entire image without explicit cropping, preserving global context and enabling end-to-end training. Building on the success of DETR [5] in object detection, recent works [2, 44] adopt a DETR-style pipeline. They mainly focus on designing queries and decoders to estimate the human mesh, e.g. Multi-HMR [2] uses queries to represent each person and decodes the corresponding human mesh. However, these methods rely on high-resolution inputs to achieve state-of-the-art (SOTA) results, as shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "in Fig. 1 (b), which significantly increases computational cost and limits real-time processing capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.909, + 0.497 + ], + "angle": 0, + "content": "Building on the findings of [2, 44], we observe that high-resolution input plays a crucial role in improving accuracy. To explore this, we design a naive baseline following similar architecture and analyze how its estimation error changes for individuals of different scales as we increase the input resolution. We define scale as the size of the person relative to the entire image1. As shown in Fig. 2, increasing the input resolution consistently leads to a decrease in the estimation error, particularly for smaller-scale individuals (e.g., kids, those far from the camera or those in closed-poses). Notably, for individuals in the \\(0 - 10\\%\\) scale, we observe an error decrease of nearly \\(35\\mathrm{mm}\\) when comparing the highest-resolution model (blueviolet) to the lowest-resolution one (yellow). However, processing the entire image at high resolution incurs significant computational costs, especially when individuals are close to the camera and occupy a large portion of the image. In such cases, these individuals are already represented by many tokens, and using more tokens with high resolution would only greatly increase the cost while offering marginal improvement (i.e., a slight error reduction in cases where individuals cover \\(30 + \\%\\) of the image in Fig. 2). Therefore, both our best baseline model (blueviolet) and recent SOTA methods [2, 44] in Fig. 1 (b) fail to achieve real-time inference and run at about 5 FPS, highlighting a common limitation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.909, + 0.801 + ], + "angle": 0, + "content": "To address this, we extend our baseline built on DETR [5] and shift our focus toward more efficient feature encoding. Specifically, we introduce scale-adaptive tokens, which are dynamically adjusted based on the scale of individuals in the image. We first predict a patch-level scale map, assigning a scale value to each token. Tokens corresponding to smaller-scale individuals are replaced with higher-resolution tokens. Additionally, background tokens identified by the scale map are further distilled to reduce computational cost. These scale-adaptive tokens capture individuals at varying scales with the appropriate level of detail, as shown in Fig. 1 (a). The tokens are then processed by the decoder and prediction heads to regress the human mesh. By efficiently allocating computational resources to the more challenging cases, our method strikes a better balance between accuracy and speed. As shown in Fig. 2, our method (red) achieves performance comparable to high-resolution input while maintaining real-time inference speed at 24 FPS. Experiments on benchmark datasets further demonstrate the effectiveness of our method, achieving competitive performance with high efficiency." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.802, + 0.838, + 0.816 + ], + "angle": 0, + "content": "To conclude, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.817, + 0.905, + 0.862 + ], + "angle": 0, + "content": "- We propose scale-adaptive tokens that are dynamically adjusted to handle individuals of varying scales, enabling more efficient feature encoding." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.863, + 0.905, + 0.877 + ], + "angle": 0, + "content": "- By employing scale-adaptive tokens, our method achieves" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.817, + 0.905, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.764, + 0.901 + ], + "angle": 0, + "content": "See Eq. (1) for detailed definition of scale." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16797" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.484, + 0.121 + ], + "angle": 0, + "content": "performance comparable to using full-image high resolution, but with significantly lower computational cost." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.122, + 0.484, + 0.182 + ], + "angle": 0, + "content": "- Our approach achieves SOTA performance with up to 5x the speed of the top-performing methods, making it the best real-time model for multi-person 3D mesh estimation, running at 24 FPS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.229, + 0.213 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.426, + 0.239 + ], + "angle": 0, + "content": "2.1. Multi-Person Human Mesh Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.485, + 0.428 + ], + "angle": 0, + "content": "Multi-person human mesh estimation from a single RGB image aims to regress all human meshes, typically represented by parametric models such as SMPL [24] and SMPL-X [34]. Existing methods can be categorized into multi-stage and one-stage approaches. Multi-stage methods [7, 11, 14, 35, 53, 55, 56] use off-the-shelf human detectors [38] to crop each person's region, followed by single-person mesh estimation [26, 27, 50, 57, 59]. This approach preserves relatively high-resolution images for each cropped region, generally leading to higher accuracy [44]. However, its reliance on detected bounding boxes and the lack of global context make it challenging to handle occlusion [45]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.671 + ], + "angle": 0, + "content": "In contrast, one-stage methods [2, 36, 44-46] estimate human meshes for all individuals by encoding the entire image and predicting simultaneously. Pioneer work like ROMP [45] and BEV [46] use CNN backbones [12] to extract global features for mesh regression. However, the low input resolution limits the expressiveness of the distilled features [44], leading to suboptimal performance. Recent methods like AiOS [44] and Multi-HMR [2] adopt DETR-style architectures to achieve SOTA performance at higher resolutions, but with significant computational overhead, making real-time use impractical (e.g., Multi-HMR runs at only 4 FPS as shown in Fig. 1). Instead, we argue that tokenizing the entire image at high resolution is unnecessary. By adjusting patch resolution based on the scale and focusing computational resources on more challenging areas, we achieve real-time inference with competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.449, + 0.697 + ], + "angle": 0, + "content": "2.2. Transformers in Human Mesh Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Transformers [48], originally designed for sequential inputs, have demonstrated strong performance in various vision tasks, such as image classification with Vision Transformers (ViTs) [8]. And it has been widely adopted in human mesh estimation [2, 6, 9, 18, 19, 44, 52]. For single-person settings, methods such as [6, 9, 18, 19, 52] focus on improving the accuracy through various architectural designs. Inspired by the DETR series [5, 23, 58] in object detection, recent methods have adopted a DETR-style framework for multi-person pose and mesh estimation [2, 22, 43, 44, 51], enabling one-stage estimation for all individuals. These works primarily focus on designing decoder queries to represent humans, keypoints, and mesh parameters. For instance," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "AiOS [44] refines predictions by progressively filtering and expanding queries in the decoder, while Multi-HMR [2] uses each query to represent a detected person. Our method also builds upon the DETR framework but focuses on improving token representations in the encoder to capture scale-aware human features, enabling accurate 3D mesh regression." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.779, + 0.207 + ], + "angle": 0, + "content": "2.3. Efficient Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Although ViTs [8] achieves significant success across various vision tasks, their quadratic computational complexity remains a major challenge, particularly for tasks that require dense predictions such as object detection [47]. As a result, efforts to accelerate ViTs primarily focus on reducing the overall number of image tokens. For instance, [30, 37, 54] drop less informative tokens, [4, 39] merge similar tokens, and [41] proposes a mixed-resolution image tokenization scheme. Recently, TORE [9] accelerates single-person 3D mesh estimation by pruning background tokens. However, since background still provides important context, entirely discarding it leads to performance degradation (see Sec. 4.4). In this work, we propose a new perspective with scale-adaptive tokens that achieve a more effective trade-off between performance and computational cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.605, + 0.47 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.655, + 0.494 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.908, + 0.623 + ], + "angle": 0, + "content": "We use SMPL [24] as our body model \\(\\mathcal{M}\\), which uses pose and shape parameters to represent the human body. Specifically, the pose parameters \\(\\theta \\in \\mathbb{R}^{24\\times 3}\\) are the relative rotations of 24 body joints, and the shape parameters \\(\\beta \\in \\mathbb{R}^{10}\\) describe the body shape. The model outputs a human mesh \\(\\mathbf{V} = \\mathcal{M}(\\pmb {\\theta},\\pmb {\\beta})\\in \\mathbb{R}^{6890\\times 3}\\) and the 3D joints can be obtained via \\(\\mathbf{J} = \\mathbf{MV}\\in \\mathbb{R}^{J\\times 3}\\), where \\(\\mathbf{M}\\in \\mathbb{R}^{J\\times 6890}\\) is the joint regressor, and \\(J\\) represents the number of joints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.625, + 0.646 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Building on recent methods [2, 44], we utilize a DETR-style pipeline [5], consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Naive baseline. While high-resolution inputs are shown to significantly enhance estimation accuracy [2, 44], they also introduce considerable computational overhead. To further explore the trade-off between resolution and efficiency, we first introduce a simple baseline using a DETR-style architecture, as depicted in Fig. 3 (top)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Given an RGB image \\(\\mathbf{I}_{\\mathrm{hr}} \\in \\mathbb{R}^{H_{\\mathrm{hr}} \\times W_{\\mathrm{hr}} \\times 3}\\), it is divided into regular patches of size \\(P \\times P\\). Patch embedding and positional embedding are then applied to these image patches, resulting in a sequence of image feature tokens \\(\\mathcal{T} = \\{t_1, t_2, \\dots, t_k\\}\\), where \\(k = H_{\\mathrm{hr}} / P \\times W_{\\mathrm{hr}} / P\\). These tokens are subsequently processed by a Transformer encoder. In the decoder, a set of human queries is initialized" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.092, + 0.887, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.41, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Overview of (top) the baseline method and (bottom) our method with scale-adaptive tokens. Top: Our baseline method adopts a DETR-style [5] pipeline consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters. Bottom: Our method focuses on efficient feature encoding using scale-adaptive tokens. Specifically, low-resolution and high-resolution patches are extracted from the input images \\(\\mathbf{I}\\) and \\(\\mathbf{I}_{\\mathrm{hr}}\\), respectively. A scale head network predicts a patch-level scale map \\(\\mathbf{S}\\) from the low-resolution tokens, classifying them into three categories: background, small-scale, and large-scale. This scale map guides the pruning and pooling of low-resolution tokens \\(\\mathcal{T}_{\\mathrm{LR}}\\) and indicates which patches should be replaced by high-resolution ones. By concatenating the pooled background tokens \\(\\mathcal{T}_{\\mathrm{B}}^{\\prime}\\), the remaining large-scale low-resolution tokens \\(\\mathcal{T}_{\\mathrm{LARGE}}\\), and the high-resolution tokens \\(\\mathcal{T}_{\\mathrm{HR}}\\), we obtain scale-adaptive tokens \\(\\mathcal{T}_{\\mathrm{SA}}\\). These tokens are then processed by the encoder, decoder, and multiple prediction heads to regress the human mesh." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.486, + 0.705 + ], + "angle": 0, + "content": "as \\(\\mathcal{Q} = \\{q_1, q_2, \\ldots, q_n\\}\\), where \\(n\\) is a predefined hyperparameter large enough to detect all the individuals. After being processed by the decoder, these human queries are passed through multiple prediction heads: pose head \\(\\mathcal{H}_{\\mathrm{p}}\\), shape head \\(\\mathcal{H}_{\\mathrm{s}}\\), translation head \\(\\mathcal{H}_{\\mathrm{t}}\\), and box head \\(\\mathcal{H}_{\\mathrm{b}}\\) to predict SMPL parameters (pose \\(\\theta\\), shape \\(\\beta\\), and 3D translation \\(\\pmb{t}\\)) and bounding boxes, respectively. Each of the four prediction heads is implemented as an Multi-Layer Perceptron (MLP). Additionally, a linear projection is applied to the human queries to compute confidence scores, which are used to filter valid predictions with a threshold \\(\\alpha_{\\mathrm{d}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.714, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Motivation. We evaluate the baseline with different image resolutions as input (see Fig. 2, Sec. 4.4 for details) and observe that high-resolution inputs significantly improve estimation accuracy, particularly for small-scale individuals whose size is relatively small to the image. However, for large-scale individuals, e.g., those closer to the camera, the benefits of higher resolution are much less pronounced. Despite these improvements, high-resolution inputs bring a substantial increase in computational cost, which is consistent with findings from prior work [2, 44] (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "This motivates us to explore more efficient strategies that balance performance and speed and we aim to encode the im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.909, + 0.676 + ], + "angle": 0, + "content": "age feature more efficiently from an earlier stage. For large-scale individuals, additional tokens may be unnecessary; instead, computational resources should be allocated more effectively. To address this, we introduce scale-adaptive tokens, which replace the uniformly partitioned tokens used in the baseline. In the following sections, we will explain how the scale is defined and computed, and how the tokens are adjusted accordingly (Sec. 3.3), followed by a discussion of the training process (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.725, + 0.699 + ], + "angle": 0, + "content": "3.3. Scale-Adaptive Tokens" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3 (bottom), we begin with a low-resolution image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where \\((H, W) = (H_{\\mathrm{hr}} / 2, W_{\\mathrm{hr}} / 2)\\), and first uniformly partition it into low-resolution patches. Our goal is to adjust the resolution of each patch based on the scale of the individual it represents. To achieve this, we introduce a patch-level scale map that guides the resolution adjustments for each patch. Specifically, this map determines which low-resolution tokens should be further subdivided and replaced by high-resolution ones, and which tokens can be compressed. The high-resolution patches are extracted from a corresponding high-resolution image \\(\\mathbf{I}_{\\mathrm{hr}}\\). After being processed by the Transformer layers, the remaining low-resolution tokens and the newly generated high-resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "tokens are combined into scale-adaptive tokens, which provide a more efficient representation with varying levels of detail. These scale-adaptive tokens are then passed through additional Transformer layers and, finally, to the decoder to regress the SMPL parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.484, + 0.278 + ], + "angle": 0, + "content": "Patch-level scale map. We define a patch-level scale map \\(\\mathbf{S} \\in [0,1]^{\\frac{H}{P} \\times \\frac{W}{P} \\times 2}\\) for patches extracted from the low-resolution image \\(\\mathbf{I}\\). For each patch, the scale map \\(\\mathbf{S}(i,j) = (c,s)\\) contains two values: \\(c\\) represents the confidence indicating whether the patch overlaps with an individual, where \\(0\\) denotes the background; \\(s\\) represents the scale of the individual within the patch, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.281, + 0.484, + 0.298 + ], + "angle": 0, + "content": "\\[\ns = \\min \\left(d _ {\\mathrm {b b}} / S _ {\\mathrm {h r}}, 1\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.394 + ], + "angle": 0, + "content": "where \\(d_{\\mathrm{bb}}\\) is the diagonal length of the bounding box in \\(\\mathbf{I}_{\\mathrm{hr}}\\), and \\(S_{\\mathrm{hr}} = \\max (H_{\\mathrm{hr}},W_{\\mathrm{hr}})\\). If the patch overlaps with multiple individuals, we use the scale of the individual closest to the camera, i.e., the one with the smaller depth. The scale is only defined for patches overlapping with individuals, and for background patches, the scale value is ignored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.485, + 0.53 + ], + "angle": 0, + "content": "To predict the patch-level scale map, we pass the low-resolution patches through a shallow transformer encoder with \\( N_{\\mathrm{lr}} \\) layers, followed by a scale head network, implemented as an MLP. As illustrated in Fig. 3, patches without color are classified as background, while colored regions represent patches overlapping with individuals. A color gradient from blue to rose indicates the scale of the individuals, with blue representing large-scale individuals and rose representing small-scale individuals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.72 + ], + "angle": 0, + "content": "Scale-adaptive encoder. Building on the predicted scale map S, the low-resolution tokens \\(\\mathcal{T}_{\\mathrm{LR}} = \\{t_{1}^{\\mathrm{lr}}, t_{2}^{\\mathrm{lr}}, \\ldots, t_{k_{\\mathrm{lr}}}^{\\mathrm{lr}}\\}\\) generated from the low-resolution patches are then classified into three groups: background \\(\\mathcal{T}_{\\mathrm{B}}\\), small-scale \\(\\mathcal{T}_{\\mathrm{SMALL}}\\), and large-scale \\(\\mathcal{T}_{\\mathrm{LARGE}}\\) tokens. Tokens identified as background, based on a confidence score \\(c\\) that meets a threshold \\(\\alpha_{c}\\), are marked as background tokens \\(\\mathcal{T}_{\\mathrm{B}} = \\{t_{1}^{\\mathrm{b}}, t_{2}^{\\mathrm{b}}, \\ldots, t_{k_{\\mathrm{b}}}^{\\mathrm{b}}\\}\\). The remaining tokens are further categorized by their scale values \\(s\\) using a scale threshold \\(\\alpha_{s}\\) and are assigned to either the small-scale category, denoted by \\(\\mathcal{T}_{\\mathrm{SMALL}} = \\{t_{1}^{\\mathrm{small}}, t_{2}^{\\mathrm{small}}, \\ldots, t_{k_{\\mathrm{small}}}^{\\mathrm{small}}\\}\\), or the large-scale category, denoted by \\(\\mathcal{T}_{\\mathrm{LARGE}} = \\{t_{1}^{\\mathrm{large}}, t_{2}^{\\mathrm{large}}, \\ldots, t_{k_{\\mathrm{large}}}^{\\mathrm{large}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Small-scale tokens are pruned and replaced by their high-resolution counterparts. High-resolution tokens are generated by applying a shallow transformer encoder with \\(N_{\\mathrm{hr}}\\) layers to the high-resolution image \\(\\mathbf{I}_{\\mathrm{hr}}\\), where \\(N_{\\mathrm{hr}} = N_{\\mathrm{lr}}\\) to ensure feature alignment. The small-scale tokens are then replaced by their corresponding high-resolution tokens (Fig. 3), resulting in an expanded set of high-resolution tokens, denoted as \\(\\mathcal{T}_{\\mathrm{HR}} = \\{t_1^{\\mathrm{hr}}, t_2^{\\mathrm{hr}}, \\dots, t_{k_{\\mathrm{hr}}}^{\\mathrm{hr}}\\}\\), where \\(k_{\\mathrm{hr}} = 4k_{\\mathrm{small}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Although background can provide valuable contextual information, their feature encoding can be further optimized. To this end, we distill the background tokens \\(\\mathcal{T}_{\\mathrm{B}}\\) by spatially pooling every four neighboring tokens, resulting in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.184 + ], + "angle": 0, + "content": "pooled background tokens \\(\\mathcal{T}_{\\mathrm{B}}^{\\prime} = \\{t_{1}^{\\mathrm{b}^{\\prime}}, t_{2}^{\\mathrm{b}^{\\prime}}, \\ldots, t_{k_{\\mathrm{b}}^{\\prime}}^{\\mathrm{b}^{\\prime}}\\}\\). To avoid irregular pooling, some low-resolution tokens that cannot be grouped remain unchanged. As a result, \\(k_{\\mathrm{b}}^{\\prime}\\) does not necessarily equal \\(k_{\\mathrm{b}} / 4\\). By retaining these tokens, we better preserve their positional information, which facilitates subsequent positional encoding operations [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.305 + ], + "angle": 0, + "content": "The remaining tokens in \\(\\mathcal{T}_{\\mathrm{LR}}\\) representing large-scale individuals, \\(\\mathcal{T}_{\\mathrm{LARGE}}\\), remain unchanged. Next, we integrate the different tokens, including pooled background tokens and high-resolution tokens from selected regions. This results in scale-adaptive tokens, denoted as \\(\\mathcal{T}_{\\mathrm{SA}} = \\{\\mathcal{T}_{\\mathrm{B}}^{\\prime}, \\mathcal{T}_{\\mathrm{LARGE}}, \\mathcal{T}_{\\mathrm{HR}}\\}\\). Compared to the uniform low-resolution tokens, this approach allocates feature details more efficiently, preserving different levels of detail for different individuals and regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.366 + ], + "angle": 0, + "content": "These tokens, \\(\\mathcal{T}_{\\mathrm{SA}}\\), are then processed by another Transformer encoder with \\(N_{\\mathrm{sa}}\\) layers and further decoded in subsequent stages. Fig. 3 (bottom) provides an overview of the scale-adaptive encoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.371, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Decoder and prediction heads. The decoder and prediction heads follow the architecture of our baseline approach as mentioned in Sec. 3.2. For detailed architectures, please refer to the supplementary material (Sec. A.1)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.441, + 0.671, + 0.457 + ], + "angle": 0, + "content": "3.4. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Similar to DETR [5], we first match our human predictions to ground-truth (GT) using Hungarian Matching, and we leverage bounding boxes, confidence scores, and projected joints during the matching. This matching step is necessary before computing the overall loss \\(\\mathcal{L}\\). More details on the matching process are provided in Sec. A.1. The loss function, \\(\\mathcal{L}\\), is a weighted sum of various terms, with each \\(\\lambda\\) representing a hyperparameter:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.593, + 0.892, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {m a p}} \\mathcal {L} _ {\\text {m a p}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {p o s e}} \\mathcal {L} _ {\\text {p o s e}} + \\lambda_ {\\text {s h a p e}} \\mathcal {L} _ {\\text {s h a p e}} \\\\ + \\lambda_ {\\mathrm {j 3 d}} \\mathcal {L} _ {\\mathrm {j 3 d}} + \\lambda_ {\\mathrm {j 2 d}} \\mathcal {L} _ {\\mathrm {j 2 d}} + \\lambda_ {\\mathrm {b o x}} \\mathcal {L} _ {\\mathrm {b o x}} + \\lambda_ {\\det} \\mathcal {L} _ {\\det}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.64, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Scale map loss \\(\\mathcal{L}_{\\mathrm{map}}\\). To supervise the patch-level scale map, we apply focal loss [21] to the predicted confidence \\(c\\) and L1 loss to the predicted scale \\(s\\). The total loss, \\(\\mathcal{L}_{\\mathrm{map}}\\), is the sum of these two terms. The GT for confidence and scale is precomputed, as described in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.836 + ], + "angle": 0, + "content": "Depth loss \\(\\mathcal{L}_{\\mathrm{depth}}\\). We supervise the root depth (i.e., the z-dimension) of the 3D joints \\(\\mathbf{J}\\), regressed from the predicted mesh \\(\\mathbf{V}\\) (Sec. 3.1). Specifically, we normalize the depth based on the focal length, following [10]. The depth loss is calculated as \\(\\mathcal{L}_{\\mathrm{depth}} = \\left|\\frac{1}{\\tilde{d}} - \\frac{f}{\\tilde{f}d}\\right|\\), where \\(f\\) is the predefined focal length of our camera, and \\(\\tilde{d}\\) and \\(\\tilde{f}\\) are the GT depth and focal length, respectively. See Sec. A.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Other losses. For the pose parameters loss \\(\\mathcal{L}_{\\mathrm{pose}}\\), shape parameters loss \\(\\mathcal{L}_{\\mathrm{shape}}\\), 3D joints loss \\(\\mathcal{L}_{\\mathrm{j3d}}\\), and projected 2D joints loss \\(\\mathcal{L}_{\\mathrm{j2d}}\\), we use the L1 distance. For the bounding box loss \\(\\mathcal{L}_{\\mathrm{box}}\\), we combine L1 loss with GIoU loss [40]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16800" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison with SOTA methods on AGORA test set [33]. \"Res.\" represents the input resolution. \"644*\" means we use a base resolution of 644 and adaptively scale certain regions up to a maximum resolution of 1288." + }, + { + "type": "table", + "bbox": [ + 0.138, + 0.129, + 0.865, + 0.254 + ], + "angle": 0, + "content": "
MethodRes.Time (ms)MACs (G)F1-Score ↑Precision ↑Recall ↑MPJPE ↓MVE ↓NMJE ↓NMVE ↓
ROMP [45]51238.743.60.910.950.88108.1103.4118.8113.6
BEV [46]51250.648.90.930.960.90105.3100.7113.2108.3
PSVT [36]512--0.93--97.794.1105.1101.2
AiOS [44]1333405.2314.50.940.980.9063.957.568.061.2
Multi-HMR [2]89697.42075.10.93--82.877.689.083.4
Multi-HMR [2]1288231.76104.60.950.970.9365.361.168.764.3
Ours644*42.0133.10.950.980.9167.963.371.566.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.301 + ], + "angle": 0, + "content": "The detection loss \\(\\mathcal{L}_{\\mathrm{det}}\\) uses focal loss [21] to supervise the confidence scores of human queries." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.226, + 0.329 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.288, + 0.351 + ], + "angle": 0, + "content": "4.1. Datasets and metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.464 + ], + "angle": 0, + "content": "We train our model on multi-person datasets including AGORA [33], BEDLAM [3], COCO [20], Crowdpose [17], and MPII [1], as well as the single-person dataset H3.6M [13]. AGORA and BEDLAM are synthetic datasets with high-quality 3D mesh GTs. For other multi-person datasets, we use pseudo GTs from NeuralAnnot [31] and only supervise projected 2D joints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.706 + ], + "angle": 0, + "content": "For comparison with SOTA methods, we evaluate our method on AGORA [33], MuPoTS [29], 3DPW [49], and CMU Panoptic [15] datasets. Following previous works [2, 44-46], we report Mean Vertex Error (MVE) and Mean Per-Joint Position Error (MPJPE) before and after Procrustes-Alignment (PA). On AGORA, we also report F1-Score, precision, and recall for evaluating detection accuracy; Normalized Mean Vertex Error (NMVE) and Normalized Mean Joint Error (NMJE) that considered regression accuracy with detection accuracy. All metrics above, except detection accuracy, are reported in millimeters (mm). On MuPoTS-3D [29], we report the Percentage of Correctly estimated Keypoints (PCK) using a threshold of \\(15\\mathrm{cm}\\). To evaluate computational costs, we report Multiply-Add Cumulation (MACs) and the average inference time (ms) following [2] on one RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.308, + 0.729 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "The input image is resized to maintain its original aspect ratio, with a base resolution of \\(\\max(H, W) = 644\\) and \\(\\max(H_{\\mathrm{hr}}, W_{\\mathrm{hr}}) = 1288\\) for high-resolution when needed. We denote this mixed-resolution setting as \"644*\". The Transformer layers in our encoder are initialized using pretrained DINOv2 [32] with a patch size of \\(P = 14\\). We use ViT-B architecture as ViT-L significantly increases inference time (42.0 ms vs. 74.2 ms). The Transformer used in the encoder has \\(N_{\\mathrm{lr}} = N_{\\mathrm{hr}} = 3\\), and \\(N_{\\mathrm{sa}} = 9\\) layers, respectively. Our decoder consists of 6 Transformer layers and 50 human queries. Training is conducted on 8 RTX 3090 GPUs with a" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.267, + 0.908, + 0.338 + ], + "angle": 0, + "content": "Table 2. Comparison with SOTA methods on 3DPW [49] and MuPoTS [29] test sets. “Res.” represents the input resolution. “Crop” refers to cropping individuals for single-person mesh estimation, i.e., a multi-stage approach. “644*” means a base resolution of 644, with adaptive scaling up to 1288." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.348, + 0.905, + 0.495 + ], + "angle": 0, + "content": "
MethodRes.3DPWMuPoTS (PCK)
PA-MPJPE ↓MVE ↓All ↑Matched ↑
CRMH [14]832--69.172.2
3DCrowdNet [7]Crop51.598.372.773.3
ROMP [45]51247.393.469.972.2
BEV [46]51246.992.370.275.2
PSVT [36]51245.784.9--
Multi-HMR [2]89641.775.985.089.3
Ours644*41.673.789.090.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.543 + ], + "angle": 0, + "content": "total batch size of 40 for 60 epochs with an initial learning rate of \\(4e - 5\\). Please refer to Sec. A.2 for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.824, + 0.574 + ], + "angle": 0, + "content": "4.3. Comparison to the State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Performance and Efficiency. We conduct evaluations on the AGORA [33], 3DPW [49], MuPoTS [28], and CMU Panoptic [15] datasets. Tab. 1 shows the results of our method compared to SOTA methods on the AGORA [33] test set leaderboard. Our method, using a mixed resolution input with a base of 644, achieves performance on par with the top SOTA methods [2, 44], despite their use of higher resolutions and larger models (e.g., Multi-HMR use ViT-L [32] model). These methods, however, result in significantly higher computational complexity, including much higher MACs and slower runtime, compromising real-time performance. Besides, these methods [2, 44] estimates SMPL-X [34], from which SMPL [24] mesh is derived. In contrast, our method maintains real-time computational efficiency while improving upon real-time methods ROMP [45], by nearly \\(40\\%\\). This establishes our approach as the leading real-time model, setting a new standard by achieving an unparalleled balance of performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Tab. 2 presents the results of our method on the 3DPW dataset. Following Multi-HMR [2], we finetune on the 3DPW training set. We further evaluate the generaliza" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.09, + 0.848, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.285, + 0.908, + 0.313 + ], + "angle": 0, + "content": "Figure 4. Comparison with SOTA methods [2, 44-46] on in-the-wild images from the Internet. Red dashed circles highlight areas with incorrect estimations. The third case is left blank due to the small scale of individuals. Please zoom in for details." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.357 + ], + "angle": 0, + "content": "Table 3. Comparison with SOTA methods on CMU Panoptic [15] test set. We report MPJPE for four activities and the average." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.367, + 0.484, + 0.478 + ], + "angle": 0, + "content": "
MethodHaggling ↓Mafia ↓Ultimatum ↓Pizza ↓Avg. ↓
CRMH [14]129.6133.5153.0156.7143.2
3DCrowdNet [7]109.6135.9129.8135.6127.6
ROMP [45]110.8122.8141.6137.6128.2
BEV [46]90.7103.7113.1125.2109.5
PSVT [36]88.797.9115.2121.2105.7
Ours67.978.595.894.684.2
" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.487, + 0.462, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.632 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of different resolutions for our baseline. Resolution for the baseline: (a) 518, (b) 1288. Red dashed circles highlight differences; zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.484, + 0.759 + ], + "angle": 0, + "content": "tion performance of our method on the MuPoTS and CMU Panoptic datasets, as shown in Tab. 2 and Tab. 3. For Panoptic dataset, we follow the common evaluation protocol [36, 45, 46] and exclude Multi-HMR [2] due to potential inconsistencies in joint format. Our method consistently achieves the best results across all datasets, demonstrating exceptional generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 4 presents a visual comparison between our method and existing SOTA approaches. The images are sourced from the Internet. Our method provides the most accurate overall estimations, especially in the third case, where individuals are far from the camera, resulting in very small scales. Despite this challenge, our method remains highly robust, delivering precise estimations. In contrast, ROMP [45] and BEV [46] fail to detect all individuals, while AiOS [44] and Multi-HMR [2], despite using" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.328, + 0.906, + 0.357 + ], + "angle": 0, + "content": "Table 4. Ablation studies on BEDLAM [3] validation set. We report MVE for different scale ranges and the average (Avg.)." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.367, + 0.892, + 0.448 + ], + "angle": 0, + "content": "
Ablation0-20%20-40%40-60%60-80%80%+Avg.
(a) Drop all59.955.161.566.670.857.2
(b) No pooling60.354.557.059.864.156.1
(c) Pooling×260.754.557.562.166.556.3
(d) Ours60.054.457.260.362.756.0
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.908, + 0.477 + ], + "angle": 0, + "content": "higher-resolution models, show significant estimation errors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.667, + 0.505 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.557 + ], + "angle": 0, + "content": "We conduct ablation studies using only the AGORA [33] and sampled BEDLAM [3] datasets to validate model designs, and report results on their validation sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.73 + ], + "angle": 0, + "content": "Impact of resolution. Fig. 2 shows baseline results with different input resolutions on the AGORA [33] validation set. As resolution increases, the error across individuals in different scale ranges consistently decreases, with the most significant drop (35mm) in the smallest scale range. In contrast, the improvement in the large-scale range is less pronounced. However, the performance gain comes at the cost of a significant drop in computational efficiency, reducing the frame rate from real-time (32 FPS) to 5 FPS. Fig. 5 shows an example, clearly illustrating that higher resolution greatly improves estimation for small-scale individuals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Background tokens \\(\\mathcal{T}_{\\mathrm{B}}\\). We investigate the impact of background tokens on model performance and report the MVE across different scale ranges on the BEDLAM validation set in Tab. 4. In (a), we discard all background tokens and observe stable performance at the smallest scale, but a decline as the scale increases, indicating that background context becomes more important for larger scales. In (b), skipping pooling for background tokens yields nearly the same performance but introduces higher computational cost. In (c), applying pooling twice with the same grouping strategy results in a performance drop for large-scale instances, highlighting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16802" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.089, + 0.855, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 6. Visualization of our predicted patch-level scale maps S and scale-adaptive tokens \\(\\mathcal{T}_{\\mathrm{SA}}\\). We show the predicted scale maps, scale-adaptive tokens, and the estimated meshes overlaid on the image. In the scale map, colors represent scale values, with uncolored areas indicating background patches. A gradient from blue (large-scale individuals) to rose (small-scale individuals) illustrates the scale distribution. We visualize scale-adaptive tokens as patches of different sizes and colors on the image, with low-resolution patches blurred." + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.384, + 0.856, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.641, + 0.908, + 0.671 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of our method. We visualize four cases from AGORA [33], BEDLAM [3], CMU Panoptic [15] and 3DPW [49] datasets, displaying the input, the estimated mesh overlay, and an elevated view (top to bottom)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.779 + ], + "angle": 0, + "content": "the significance of background context for estimation. We hypothesize that large-scale people face more truncation, relying more on global context which requires further exploration. We provide additional efficiency comparisons in Sec. B, showing that pooling once achieves the best balance between performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.274, + 0.803 + ], + "angle": 0, + "content": "4.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Fig. 6 visualize the predicted patch-level scale map and scale-adaptive tokens by our model, enabling dynamic adjustment of tokens for accurate estimations. Fig. 7 presents additional visualizations of predicted meshes, demonstrating the effectiveness of our method in various scenarios. Please refer to Sec. B for more qualitative results and discussion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.685, + 0.634, + 0.7 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.909, + 0.892 + ], + "angle": 0, + "content": "We present a novel one-stage framework for real-time multiperson 3D mesh estimation from an RGB image. By introducing scale-adaptive tokens that dynamically adjust based on the relative scale of individuals, our method effectively balances efficiency and accuracy, retaining the benefits of high-resolution processing while achieving real-time performance. Experiments demonstrate that our method is the best real-time model, offering competitive performance and superior generalization compared to SOTAs. Our limitations include the lack of height-aware scale estimation, which may cause depth errors, and the current focus on body-only mesh estimation, extendable to whole-body in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16803" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work was supported by 2022ZD0114904 and NSFC-6247070125. We thank Yiyang Wang for the valuable discussions during the early stage of this work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.256, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Fabien Baradel, Matthieu Armando, Salma Galaaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas. Multi-hmr: Multi-person whole-body human mesh recovery in a single shot. In ECCV, 2024. 2, 3, 4, 6, 7, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.366 + ], + "angle": 0, + "content": "[3] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 6, 7, 8, 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.408 + ], + "angle": 0, + "content": "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 3, 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[6] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[7] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3d human mesh from in-the-wild crowded scenes. In CVPR, pages 1475-1484, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[9] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. Tore: Token reduction for efficient human mesh recovery with transformer. In ICCV, pages 15143-15155, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Jose M Facil, Benjamin Ummenhofer, Huizhong Zhou, Luis Montesano, Thomas Brox, and Javier Civera. Cam-convs: Camera-aware multi-scale convolutions for single-view depth. In CVPR, pages 11826-11835, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predic" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "tive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[14] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In CVPR, pages 5579-5588, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[15] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, pages 3334-3342, 2015. 6, 7, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[16] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.908, + 0.357 + ], + "angle": 0, + "content": "[17] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In CVPR, pages 10863-10872, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.907, + 0.427 + ], + "angle": 0, + "content": "[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, pages 12939-12948, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[22] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, et al. Group pose: A simple baseline for end-to-end multi-person pose estimation. In ICCV, pages 15029-15038, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.907, + 0.649 + ], + "angle": 0, + "content": "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 3, 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[26] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In CVPR, pages 534-543, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.778, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[27] Xiaoxuan Ma, Jiajun Su, Yuan Xu, Wentao Zhu, Chunyu Wang, and Yizhou Wang. Vmarker-pro: Probabilistic 3d human mesh estimation from virtual markers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, pages 506-516, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[29] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[30] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.484, + 0.251 + ], + "angle": 0, + "content": "[31] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralnot: Neural annotator for 3d human mesh training sets. In CVPR, pages 2299-2307, 2022. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.253, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[32] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.324, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[33] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In CVPR, pages 13468-13478, 2021. 1, 2, 6, 7, 8, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.382, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.453, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[35] Zhongwei Qiu, Qiansheng Yang, Jian Wang, and Dongmei Fu. Dynamic graph reasoning for multi-person 3d pose estimation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3521-3529, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.484, + 0.581 + ], + "angle": 0, + "content": "[36] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In CVPR, pages 21254-21263, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.484, + 0.638 + ], + "angle": 0, + "content": "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.641, + 0.484, + 0.683 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[39] Cedric Renggli, André Susano Pinto, Neil Houlsby, Basil Mustafa, Joan Puigcerver, and Carlos Riquelme. Learning to merge tokens in vision transformers. arXiv preprint arXiv:2202.12015, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[41] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In CVPRW, pages 4613-4622, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.135 + ], + "angle": 0, + "content": "[43] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In CVPR, pages 11069-11078, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[44] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi-Sing Leung, Ziwei Liu, Lei Yang, et al. Aios: All-in-one-stage expressive human pose and shape estimation. In CVPR, pages 1834-1843, 2024, 2, 3, 4, 6, 7, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.908, + 0.25 + ], + "angle": 0, + "content": "[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 3, 6, 7, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.907, + 0.306 + ], + "angle": 0, + "content": "[46] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 2, 3, 6, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[47] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.881, + 0.353 + ], + "angle": 0, + "content": "[48] A Vaswani. Attention is all you need. NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.355, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[49] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 1, 6, 8, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.907, + 0.467 + ], + "angle": 0, + "content": "[50] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In CVPR, pages 979–989, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[51] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.907, + 0.569 + ], + "angle": 0, + "content": "[52] Sen Yang, Wen Heng, Gang Liu, Guozhong Luo, Wankou Yang, and YU Gang. Capturing the motion of every joint: 3d human pose and shape estimation with independent tokens. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.907, + 0.669 + ], + "angle": 0, + "content": "[54] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, pages 10809-10818, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.907, + 0.728 + ], + "angle": 0, + "content": "[55] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.907, + 0.785 + ], + "angle": 0, + "content": "[56] Andrei Zanfir, Elisabeta Marinoiu, Mihai Zanfir, Alin-Ionut Popa, and Cristian Sminchisescu. Deep network for the integrated 3d sensing of multiple people in natural images. In NeurIPS, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[57] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[58] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In CVPR, pages 15085-15099, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[60] Wentao Zhu, Jason Qin, Yuke Lou, Hang Ye, Xiaoxuan Ma, Hai Ci, and Yizhou Wang. Social motion prediction with cognitive hierarchies. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[61] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlun Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang: Human motion generation: A survey. IEEE TPAMI, 46(4): 2430-2449, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16806" + } + ] +] \ No newline at end of file diff --git a/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_origin.pdf b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..654f3e3298d1c606274e7cfe5fc08d234dcb546c --- /dev/null +++ b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/eab4bc33-3ff9-4b95-b35e-cb1cb0f44fba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab822be49a3636ae396b4e06d37391193df54f3d096535f5654c90d6b7cf871a +size 10008234 diff --git a/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/full.md b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/full.md new file mode 100644 index 0000000000000000000000000000000000000000..07e5d06f7755fe3f4abe5ac16301faf39bdf4763 --- /dev/null +++ b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/full.md @@ -0,0 +1,281 @@ +# SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens + +Chi Su1 Xiaoxuan Ma1,☑ Jiajun Su5 Yizhou Wang1,2,3,4,☑ + +$^{1}$ Center on Frontiers of Computing Studies, School of Computer Science, Peking University + $^{2}$ Inst. for Artificial Intelligence, Peking University $^{3}$ Nat'l Eng. Research Center of Visual Technology + $^{4}$ State Key Laboratory of General Artificial Intelligence, Peking University + $^{5}$ International Digital Economy Academy (IDEA) + +suchi@stu.pku.edu.cn, {maxiaoxuan, yizhou.wang}@pku.edu.cn, sujiajun@idea.edu.cn Corresponding authors + +![](images/50f12f786f864623e3d98e0bc9316ad10c2c51febf4a4279ac54bb7aa782954e.jpg) +(a) + +![](images/ef0882c895bd45834df76479317ffebbcc4c36391a5d247de3a5b86540791457.jpg) +(b) +Figure 1. (a) We propose scale-adaptive tokens in our one-stage framework for real-time multi-person 3D mesh estimation. Our method introduces scale-adaptive tokens, dynamically adjusted based on the relative size of individuals in the image, to more efficiently encode features, enabling real-time and accurate multi-person mesh estimation. We present a conceptual visualization of the scale-adaptive tokens. The right column visualizes the predicted meshes projected onto an image from 3DPW [49] dataset and from an elevated view. (b) Comparison of estimation error and inference time across different methods, with input resolutions in parentheses. Our method, using a mixed resolution with a base resolution of 644, achieves comparable performance to state-of-the-art methods on AGORA [33] test set while maintaining real-time inference efficiency. Code and models are available at https://ChiSu001.github.io/SAT-HMR/. + +# Abstract + +We propose SAT-HMR, a one-stage framework for real-time multi-person 3D human mesh estimation from a single RGB image. While current one-stage methods, which follow a DETR-style pipeline, achieve state-of-the-art (SOTA) performance with high-resolution inputs, we observe that this particularly benefits the estimation of individuals in smaller scales of the image (e.g., those of young age or far from the camera), but at the cost of significantly increased computation overhead. To address this, we introduce scale-adaptive tokens that are dynamically adjusted based on the relative + +scale of each individual in the image within the DETR framework. Specifically, individuals in smaller scales are processed at higher resolutions, larger ones at lower resolutions, and background regions are further distilled. These scale-adaptive tokens more efficiently encode the image features, facilitating subsequent decoding to regress the human mesh, while allowing the model to allocate computational resources more effectively and focus on more challenging cases. Experiments show that our method preserves the accuracy benefits of high-resolution processing while substantially reducing computational cost, achieving real-time inference with performance comparable to SOTA methods. + +![](images/14f3515040cc45f35aa5cab6495e04b8f5095170809d1f19ee2e268cf7936c12.jpg) +Figure 2. Estimation errors and FPS of baselines with different resolutions and our method across individuals at various scales. The scale of an individual refers to the person's size relative to the overall image and please refer to Sec. 3.3 for mathematical definition. The colored lines show the Mean Vertex Error (MVE) errors (left y-axis) of the baselines with different resolutions (Res.) on the AGORA [33] validation set. The colored markers on the right y-axis indicate the FPS of the corresponding models. Our method adopts a mixed resolution with a base resolution of 644. + +# 1. Introduction + +Multi-person 3D human mesh estimation from a single RGB image aims to localize all individuals in the scene and estimate their 3D meshes, typically represented by a parametric human model such as SMPL [24]. This is a fundamental task with broad applications, spanning social interaction [42, 60], game production [61], etc. Existing methods primarily focus on addressing challenges such as monocular depth ambiguity and inter-person interactions, and can be broadly categorized into two types: multi-stage and one-stage approaches. Multi-stage methods [7, 14, 35, 55, 56] typically detect and process each individual by cropping and resizing them to a uniform, relatively high resolution, which generally improves performance. However, this overlooks the global context, such as the original scales of individuals and their relative positions within the entire image, potentially leading to suboptimal results. + +To address these limitations, one-stage methods gradually emerge as a promising alternative [2, 44-46]. These approaches operate on the entire image without explicit cropping, preserving global context and enabling end-to-end training. Building on the success of DETR [5] in object detection, recent works [2, 44] adopt a DETR-style pipeline. They mainly focus on designing queries and decoders to estimate the human mesh, e.g. Multi-HMR [2] uses queries to represent each person and decodes the corresponding human mesh. However, these methods rely on high-resolution inputs to achieve state-of-the-art (SOTA) results, as shown + +in Fig. 1 (b), which significantly increases computational cost and limits real-time processing capabilities. + +Building on the findings of [2, 44], we observe that high-resolution input plays a crucial role in improving accuracy. To explore this, we design a naive baseline following similar architecture and analyze how its estimation error changes for individuals of different scales as we increase the input resolution. We define scale as the size of the person relative to the entire image1. As shown in Fig. 2, increasing the input resolution consistently leads to a decrease in the estimation error, particularly for smaller-scale individuals (e.g., kids, those far from the camera or those in closed-poses). Notably, for individuals in the $0 - 10\%$ scale, we observe an error decrease of nearly $35\mathrm{mm}$ when comparing the highest-resolution model (blueviolet) to the lowest-resolution one (yellow). However, processing the entire image at high resolution incurs significant computational costs, especially when individuals are close to the camera and occupy a large portion of the image. In such cases, these individuals are already represented by many tokens, and using more tokens with high resolution would only greatly increase the cost while offering marginal improvement (i.e., a slight error reduction in cases where individuals cover $30 + \%$ of the image in Fig. 2). Therefore, both our best baseline model (blueviolet) and recent SOTA methods [2, 44] in Fig. 1 (b) fail to achieve real-time inference and run at about 5 FPS, highlighting a common limitation. + +To address this, we extend our baseline built on DETR [5] and shift our focus toward more efficient feature encoding. Specifically, we introduce scale-adaptive tokens, which are dynamically adjusted based on the scale of individuals in the image. We first predict a patch-level scale map, assigning a scale value to each token. Tokens corresponding to smaller-scale individuals are replaced with higher-resolution tokens. Additionally, background tokens identified by the scale map are further distilled to reduce computational cost. These scale-adaptive tokens capture individuals at varying scales with the appropriate level of detail, as shown in Fig. 1 (a). The tokens are then processed by the decoder and prediction heads to regress the human mesh. By efficiently allocating computational resources to the more challenging cases, our method strikes a better balance between accuracy and speed. As shown in Fig. 2, our method (red) achieves performance comparable to high-resolution input while maintaining real-time inference speed at 24 FPS. Experiments on benchmark datasets further demonstrate the effectiveness of our method, achieving competitive performance with high efficiency. + +To conclude, our contributions are as follows: + +- We propose scale-adaptive tokens that are dynamically adjusted to handle individuals of varying scales, enabling more efficient feature encoding. +- By employing scale-adaptive tokens, our method achieves + +performance comparable to using full-image high resolution, but with significantly lower computational cost. + +- Our approach achieves SOTA performance with up to 5x the speed of the top-performing methods, making it the best real-time model for multi-person 3D mesh estimation, running at 24 FPS. + +# 2. Related work + +# 2.1. Multi-Person Human Mesh Estimation + +Multi-person human mesh estimation from a single RGB image aims to regress all human meshes, typically represented by parametric models such as SMPL [24] and SMPL-X [34]. Existing methods can be categorized into multi-stage and one-stage approaches. Multi-stage methods [7, 11, 14, 35, 53, 55, 56] use off-the-shelf human detectors [38] to crop each person's region, followed by single-person mesh estimation [26, 27, 50, 57, 59]. This approach preserves relatively high-resolution images for each cropped region, generally leading to higher accuracy [44]. However, its reliance on detected bounding boxes and the lack of global context make it challenging to handle occlusion [45]. + +In contrast, one-stage methods [2, 36, 44-46] estimate human meshes for all individuals by encoding the entire image and predicting simultaneously. Pioneer work like ROMP [45] and BEV [46] use CNN backbones [12] to extract global features for mesh regression. However, the low input resolution limits the expressiveness of the distilled features [44], leading to suboptimal performance. Recent methods like AiOS [44] and Multi-HMR [2] adopt DETR-style architectures to achieve SOTA performance at higher resolutions, but with significant computational overhead, making real-time use impractical (e.g., Multi-HMR runs at only 4 FPS as shown in Fig. 1). Instead, we argue that tokenizing the entire image at high resolution is unnecessary. By adjusting patch resolution based on the scale and focusing computational resources on more challenging areas, we achieve real-time inference with competitive performance. + +# 2.2. Transformers in Human Mesh Estimation + +Transformers [48], originally designed for sequential inputs, have demonstrated strong performance in various vision tasks, such as image classification with Vision Transformers (ViTs) [8]. And it has been widely adopted in human mesh estimation [2, 6, 9, 18, 19, 44, 52]. For single-person settings, methods such as [6, 9, 18, 19, 52] focus on improving the accuracy through various architectural designs. Inspired by the DETR series [5, 23, 58] in object detection, recent methods have adopted a DETR-style framework for multi-person pose and mesh estimation [2, 22, 43, 44, 51], enabling one-stage estimation for all individuals. These works primarily focus on designing decoder queries to represent humans, keypoints, and mesh parameters. For instance, + +AiOS [44] refines predictions by progressively filtering and expanding queries in the decoder, while Multi-HMR [2] uses each query to represent a detected person. Our method also builds upon the DETR framework but focuses on improving token representations in the encoder to capture scale-aware human features, enabling accurate 3D mesh regression. + +# 2.3. Efficient Vision Transformers + +Although ViTs [8] achieves significant success across various vision tasks, their quadratic computational complexity remains a major challenge, particularly for tasks that require dense predictions such as object detection [47]. As a result, efforts to accelerate ViTs primarily focus on reducing the overall number of image tokens. For instance, [30, 37, 54] drop less informative tokens, [4, 39] merge similar tokens, and [41] proposes a mixed-resolution image tokenization scheme. Recently, TORE [9] accelerates single-person 3D mesh estimation by pruning background tokens. However, since background still provides important context, entirely discarding it leads to performance degradation (see Sec. 4.4). In this work, we propose a new perspective with scale-adaptive tokens that achieve a more effective trade-off between performance and computational cost. + +# 3. Method + +# 3.1. Preliminaries + +We use SMPL [24] as our body model $\mathcal{M}$ , which uses pose and shape parameters to represent the human body. Specifically, the pose parameters $\theta \in \mathbb{R}^{24\times 3}$ are the relative rotations of 24 body joints, and the shape parameters $\beta \in \mathbb{R}^{10}$ describe the body shape. The model outputs a human mesh $\mathbf{V} = \mathcal{M}(\pmb {\theta},\pmb {\beta})\in \mathbb{R}^{6890\times 3}$ and the 3D joints can be obtained via $\mathbf{J} = \mathbf{MV}\in \mathbb{R}^{J\times 3}$ , where $\mathbf{M}\in \mathbb{R}^{J\times 6890}$ is the joint regressor, and $J$ represents the number of joints. + +# 3.2. Overview + +Building on recent methods [2, 44], we utilize a DETR-style pipeline [5], consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters. + +Naive baseline. While high-resolution inputs are shown to significantly enhance estimation accuracy [2, 44], they also introduce considerable computational overhead. To further explore the trade-off between resolution and efficiency, we first introduce a simple baseline using a DETR-style architecture, as depicted in Fig. 3 (top). + +Given an RGB image $\mathbf{I}_{\mathrm{hr}} \in \mathbb{R}^{H_{\mathrm{hr}} \times W_{\mathrm{hr}} \times 3}$ , it is divided into regular patches of size $P \times P$ . Patch embedding and positional embedding are then applied to these image patches, resulting in a sequence of image feature tokens $\mathcal{T} = \{t_1, t_2, \dots, t_k\}$ , where $k = H_{\mathrm{hr}} / P \times W_{\mathrm{hr}} / P$ . These tokens are subsequently processed by a Transformer encoder. In the decoder, a set of human queries is initialized + +![](images/5cd82588e1232c0c4c1cedbbfe296a314ec9874064ec616bf3df315ff3f4b0f4.jpg) +Figure 3. Overview of (top) the baseline method and (bottom) our method with scale-adaptive tokens. Top: Our baseline method adopts a DETR-style [5] pipeline consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters. Bottom: Our method focuses on efficient feature encoding using scale-adaptive tokens. Specifically, low-resolution and high-resolution patches are extracted from the input images $\mathbf{I}$ and $\mathbf{I}_{\mathrm{hr}}$ , respectively. A scale head network predicts a patch-level scale map $\mathbf{S}$ from the low-resolution tokens, classifying them into three categories: background, small-scale, and large-scale. This scale map guides the pruning and pooling of low-resolution tokens $\mathcal{T}_{\mathrm{LR}}$ and indicates which patches should be replaced by high-resolution ones. By concatenating the pooled background tokens $\mathcal{T}_{\mathrm{B}}^{\prime}$ , the remaining large-scale low-resolution tokens $\mathcal{T}_{\mathrm{LARGE}}$ , and the high-resolution tokens $\mathcal{T}_{\mathrm{HR}}$ , we obtain scale-adaptive tokens $\mathcal{T}_{\mathrm{SA}}$ . These tokens are then processed by the encoder, decoder, and multiple prediction heads to regress the human mesh. + +as $\mathcal{Q} = \{q_1, q_2, \ldots, q_n\}$ , where $n$ is a predefined hyperparameter large enough to detect all the individuals. After being processed by the decoder, these human queries are passed through multiple prediction heads: pose head $\mathcal{H}_{\mathrm{p}}$ , shape head $\mathcal{H}_{\mathrm{s}}$ , translation head $\mathcal{H}_{\mathrm{t}}$ , and box head $\mathcal{H}_{\mathrm{b}}$ to predict SMPL parameters (pose $\theta$ , shape $\beta$ , and 3D translation $\pmb{t}$ ) and bounding boxes, respectively. Each of the four prediction heads is implemented as an Multi-Layer Perceptron (MLP). Additionally, a linear projection is applied to the human queries to compute confidence scores, which are used to filter valid predictions with a threshold $\alpha_{\mathrm{d}}$ . + +Motivation. We evaluate the baseline with different image resolutions as input (see Fig. 2, Sec. 4.4 for details) and observe that high-resolution inputs significantly improve estimation accuracy, particularly for small-scale individuals whose size is relatively small to the image. However, for large-scale individuals, e.g., those closer to the camera, the benefits of higher resolution are much less pronounced. Despite these improvements, high-resolution inputs bring a substantial increase in computational cost, which is consistent with findings from prior work [2, 44] (see Fig. 1). + +This motivates us to explore more efficient strategies that balance performance and speed and we aim to encode the im + +age feature more efficiently from an earlier stage. For large-scale individuals, additional tokens may be unnecessary; instead, computational resources should be allocated more effectively. To address this, we introduce scale-adaptive tokens, which replace the uniformly partitioned tokens used in the baseline. In the following sections, we will explain how the scale is defined and computed, and how the tokens are adjusted accordingly (Sec. 3.3), followed by a discussion of the training process (Sec. 3.4). + +# 3.3. Scale-Adaptive Tokens + +As shown in Fig. 3 (bottom), we begin with a low-resolution image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , where $(H, W) = (H_{\mathrm{hr}} / 2, W_{\mathrm{hr}} / 2)$ , and first uniformly partition it into low-resolution patches. Our goal is to adjust the resolution of each patch based on the scale of the individual it represents. To achieve this, we introduce a patch-level scale map that guides the resolution adjustments for each patch. Specifically, this map determines which low-resolution tokens should be further subdivided and replaced by high-resolution ones, and which tokens can be compressed. The high-resolution patches are extracted from a corresponding high-resolution image $\mathbf{I}_{\mathrm{hr}}$ . After being processed by the Transformer layers, the remaining low-resolution tokens and the newly generated high-resolution + +tokens are combined into scale-adaptive tokens, which provide a more efficient representation with varying levels of detail. These scale-adaptive tokens are then passed through additional Transformer layers and, finally, to the decoder to regress the SMPL parameters. + +Patch-level scale map. We define a patch-level scale map $\mathbf{S} \in [0,1]^{\frac{H}{P} \times \frac{W}{P} \times 2}$ for patches extracted from the low-resolution image $\mathbf{I}$ . For each patch, the scale map $\mathbf{S}(i,j) = (c,s)$ contains two values: $c$ represents the confidence indicating whether the patch overlaps with an individual, where $0$ denotes the background; $s$ represents the scale of the individual within the patch, defined as + +$$ +s = \min \left(d _ {\mathrm {b b}} / S _ {\mathrm {h r}}, 1\right), \tag {1} +$$ + +where $d_{\mathrm{bb}}$ is the diagonal length of the bounding box in $\mathbf{I}_{\mathrm{hr}}$ , and $S_{\mathrm{hr}} = \max (H_{\mathrm{hr}},W_{\mathrm{hr}})$ . If the patch overlaps with multiple individuals, we use the scale of the individual closest to the camera, i.e., the one with the smaller depth. The scale is only defined for patches overlapping with individuals, and for background patches, the scale value is ignored. + +To predict the patch-level scale map, we pass the low-resolution patches through a shallow transformer encoder with $N_{\mathrm{lr}}$ layers, followed by a scale head network, implemented as an MLP. As illustrated in Fig. 3, patches without color are classified as background, while colored regions represent patches overlapping with individuals. A color gradient from blue to rose indicates the scale of the individuals, with blue representing large-scale individuals and rose representing small-scale individuals. + +Scale-adaptive encoder. Building on the predicted scale map S, the low-resolution tokens $\mathcal{T}_{\mathrm{LR}} = \{t_{1}^{\mathrm{lr}}, t_{2}^{\mathrm{lr}}, \ldots, t_{k_{\mathrm{lr}}}^{\mathrm{lr}}\}$ generated from the low-resolution patches are then classified into three groups: background $\mathcal{T}_{\mathrm{B}}$ , small-scale $\mathcal{T}_{\mathrm{SMALL}}$ , and large-scale $\mathcal{T}_{\mathrm{LARGE}}$ tokens. Tokens identified as background, based on a confidence score $c$ that meets a threshold $\alpha_{c}$ , are marked as background tokens $\mathcal{T}_{\mathrm{B}} = \{t_{1}^{\mathrm{b}}, t_{2}^{\mathrm{b}}, \ldots, t_{k_{\mathrm{b}}}^{\mathrm{b}}\}$ . The remaining tokens are further categorized by their scale values $s$ using a scale threshold $\alpha_{s}$ and are assigned to either the small-scale category, denoted by $\mathcal{T}_{\mathrm{SMALL}} = \{t_{1}^{\mathrm{small}}, t_{2}^{\mathrm{small}}, \ldots, t_{k_{\mathrm{small}}}^{\mathrm{small}}\}$ , or the large-scale category, denoted by $\mathcal{T}_{\mathrm{LARGE}} = \{t_{1}^{\mathrm{large}}, t_{2}^{\mathrm{large}}, \ldots, t_{k_{\mathrm{large}}}^{\mathrm{large}}\}$ . + +Small-scale tokens are pruned and replaced by their high-resolution counterparts. High-resolution tokens are generated by applying a shallow transformer encoder with $N_{\mathrm{hr}}$ layers to the high-resolution image $\mathbf{I}_{\mathrm{hr}}$ , where $N_{\mathrm{hr}} = N_{\mathrm{lr}}$ to ensure feature alignment. The small-scale tokens are then replaced by their corresponding high-resolution tokens (Fig. 3), resulting in an expanded set of high-resolution tokens, denoted as $\mathcal{T}_{\mathrm{HR}} = \{t_1^{\mathrm{hr}}, t_2^{\mathrm{hr}}, \dots, t_{k_{\mathrm{hr}}}^{\mathrm{hr}}\}$ , where $k_{\mathrm{hr}} = 4k_{\mathrm{small}}$ . + +Although background can provide valuable contextual information, their feature encoding can be further optimized. To this end, we distill the background tokens $\mathcal{T}_{\mathrm{B}}$ by spatially pooling every four neighboring tokens, resulting in + +pooled background tokens $\mathcal{T}_{\mathrm{B}}^{\prime} = \{t_{1}^{\mathrm{b}^{\prime}}, t_{2}^{\mathrm{b}^{\prime}}, \ldots, t_{k_{\mathrm{b}}^{\prime}}^{\mathrm{b}^{\prime}}\}$ . To avoid irregular pooling, some low-resolution tokens that cannot be grouped remain unchanged. As a result, $k_{\mathrm{b}}^{\prime}$ does not necessarily equal $k_{\mathrm{b}} / 4$ . By retaining these tokens, we better preserve their positional information, which facilitates subsequent positional encoding operations [23]. + +The remaining tokens in $\mathcal{T}_{\mathrm{LR}}$ representing large-scale individuals, $\mathcal{T}_{\mathrm{LARGE}}$ , remain unchanged. Next, we integrate the different tokens, including pooled background tokens and high-resolution tokens from selected regions. This results in scale-adaptive tokens, denoted as $\mathcal{T}_{\mathrm{SA}} = \{\mathcal{T}_{\mathrm{B}}^{\prime}, \mathcal{T}_{\mathrm{LARGE}}, \mathcal{T}_{\mathrm{HR}}\}$ . Compared to the uniform low-resolution tokens, this approach allocates feature details more efficiently, preserving different levels of detail for different individuals and regions. + +These tokens, $\mathcal{T}_{\mathrm{SA}}$ , are then processed by another Transformer encoder with $N_{\mathrm{sa}}$ layers and further decoded in subsequent stages. Fig. 3 (bottom) provides an overview of the scale-adaptive encoder. + +Decoder and prediction heads. The decoder and prediction heads follow the architecture of our baseline approach as mentioned in Sec. 3.2. For detailed architectures, please refer to the supplementary material (Sec. A.1). + +# 3.4. Training Losses + +Similar to DETR [5], we first match our human predictions to ground-truth (GT) using Hungarian Matching, and we leverage bounding boxes, confidence scores, and projected joints during the matching. This matching step is necessary before computing the overall loss $\mathcal{L}$ . More details on the matching process are provided in Sec. A.1. The loss function, $\mathcal{L}$ , is a weighted sum of various terms, with each $\lambda$ representing a hyperparameter: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {\text {m a p}} \mathcal {L} _ {\text {m a p}} + \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {\text {p o s e}} \mathcal {L} _ {\text {p o s e}} + \lambda_ {\text {s h a p e}} \mathcal {L} _ {\text {s h a p e}} \\ + \lambda_ {\mathrm {j 3 d}} \mathcal {L} _ {\mathrm {j 3 d}} + \lambda_ {\mathrm {j 2 d}} \mathcal {L} _ {\mathrm {j 2 d}} + \lambda_ {\mathrm {b o x}} \mathcal {L} _ {\mathrm {b o x}} + \lambda_ {\det} \mathcal {L} _ {\det}. \\ \end{array} +$$ + +Scale map loss $\mathcal{L}_{\mathrm{map}}$ . To supervise the patch-level scale map, we apply focal loss [21] to the predicted confidence $c$ and L1 loss to the predicted scale $s$ . The total loss, $\mathcal{L}_{\mathrm{map}}$ , is the sum of these two terms. The GT for confidence and scale is precomputed, as described in Sec. 3.3. + +Depth loss $\mathcal{L}_{\mathrm{depth}}$ . We supervise the root depth (i.e., the z-dimension) of the 3D joints $\mathbf{J}$ , regressed from the predicted mesh $\mathbf{V}$ (Sec. 3.1). Specifically, we normalize the depth based on the focal length, following [10]. The depth loss is calculated as $\mathcal{L}_{\mathrm{depth}} = \left|\frac{1}{\tilde{d}} - \frac{f}{\tilde{f}d}\right|$ , where $f$ is the predefined focal length of our camera, and $\tilde{d}$ and $\tilde{f}$ are the GT depth and focal length, respectively. See Sec. A.1 for more details. + +Other losses. For the pose parameters loss $\mathcal{L}_{\mathrm{pose}}$ , shape parameters loss $\mathcal{L}_{\mathrm{shape}}$ , 3D joints loss $\mathcal{L}_{\mathrm{j3d}}$ , and projected 2D joints loss $\mathcal{L}_{\mathrm{j2d}}$ , we use the L1 distance. For the bounding box loss $\mathcal{L}_{\mathrm{box}}$ , we combine L1 loss with GIoU loss [40]. + +Table 1. Comparison with SOTA methods on AGORA test set [33]. "Res." represents the input resolution. "644*" means we use a base resolution of 644 and adaptively scale certain regions up to a maximum resolution of 1288. + +
MethodRes.Time (ms)MACs (G)F1-Score ↑Precision ↑Recall ↑MPJPE ↓MVE ↓NMJE ↓NMVE ↓
ROMP [45]51238.743.60.910.950.88108.1103.4118.8113.6
BEV [46]51250.648.90.930.960.90105.3100.7113.2108.3
PSVT [36]512--0.93--97.794.1105.1101.2
AiOS [44]1333405.2314.50.940.980.9063.957.568.061.2
Multi-HMR [2]89697.42075.10.93--82.877.689.083.4
Multi-HMR [2]1288231.76104.60.950.970.9365.361.168.764.3
Ours644*42.0133.10.950.980.9167.963.371.566.6
+ +The detection loss $\mathcal{L}_{\mathrm{det}}$ uses focal loss [21] to supervise the confidence scores of human queries. + +# 4. Experiments + +# 4.1. Datasets and metrics + +We train our model on multi-person datasets including AGORA [33], BEDLAM [3], COCO [20], Crowdpose [17], and MPII [1], as well as the single-person dataset H3.6M [13]. AGORA and BEDLAM are synthetic datasets with high-quality 3D mesh GTs. For other multi-person datasets, we use pseudo GTs from NeuralAnnot [31] and only supervise projected 2D joints. + +For comparison with SOTA methods, we evaluate our method on AGORA [33], MuPoTS [29], 3DPW [49], and CMU Panoptic [15] datasets. Following previous works [2, 44-46], we report Mean Vertex Error (MVE) and Mean Per-Joint Position Error (MPJPE) before and after Procrustes-Alignment (PA). On AGORA, we also report F1-Score, precision, and recall for evaluating detection accuracy; Normalized Mean Vertex Error (NMVE) and Normalized Mean Joint Error (NMJE) that considered regression accuracy with detection accuracy. All metrics above, except detection accuracy, are reported in millimeters (mm). On MuPoTS-3D [29], we report the Percentage of Correctly estimated Keypoints (PCK) using a threshold of $15\mathrm{cm}$ . To evaluate computational costs, we report Multiply-Add Cumulation (MACs) and the average inference time (ms) following [2] on one RTX 3090 GPU. + +# 4.2. Implementation Details + +The input image is resized to maintain its original aspect ratio, with a base resolution of $\max(H, W) = 644$ and $\max(H_{\mathrm{hr}}, W_{\mathrm{hr}}) = 1288$ for high-resolution when needed. We denote this mixed-resolution setting as "644*". The Transformer layers in our encoder are initialized using pretrained DINOv2 [32] with a patch size of $P = 14$ . We use ViT-B architecture as ViT-L significantly increases inference time (42.0 ms vs. 74.2 ms). The Transformer used in the encoder has $N_{\mathrm{lr}} = N_{\mathrm{hr}} = 3$ , and $N_{\mathrm{sa}} = 9$ layers, respectively. Our decoder consists of 6 Transformer layers and 50 human queries. Training is conducted on 8 RTX 3090 GPUs with a + +Table 2. Comparison with SOTA methods on 3DPW [49] and MuPoTS [29] test sets. “Res.” represents the input resolution. “Crop” refers to cropping individuals for single-person mesh estimation, i.e., a multi-stage approach. “644*” means a base resolution of 644, with adaptive scaling up to 1288. + +
MethodRes.3DPWMuPoTS (PCK)
PA-MPJPE ↓MVE ↓All ↑Matched ↑
CRMH [14]832--69.172.2
3DCrowdNet [7]Crop51.598.372.773.3
ROMP [45]51247.393.469.972.2
BEV [46]51246.992.370.275.2
PSVT [36]51245.784.9--
Multi-HMR [2]89641.775.985.089.3
Ours644*41.673.789.090.1
+ +total batch size of 40 for 60 epochs with an initial learning rate of $4e - 5$ . Please refer to Sec. A.2 for more details. + +# 4.3. Comparison to the State-of-the-arts + +Performance and Efficiency. We conduct evaluations on the AGORA [33], 3DPW [49], MuPoTS [28], and CMU Panoptic [15] datasets. Tab. 1 shows the results of our method compared to SOTA methods on the AGORA [33] test set leaderboard. Our method, using a mixed resolution input with a base of 644, achieves performance on par with the top SOTA methods [2, 44], despite their use of higher resolutions and larger models (e.g., Multi-HMR use ViT-L [32] model). These methods, however, result in significantly higher computational complexity, including much higher MACs and slower runtime, compromising real-time performance. Besides, these methods [2, 44] estimates SMPL-X [34], from which SMPL [24] mesh is derived. In contrast, our method maintains real-time computational efficiency while improving upon real-time methods ROMP [45], by nearly $40\%$ . This establishes our approach as the leading real-time model, setting a new standard by achieving an unparalleled balance of performance and efficiency. + +Tab. 2 presents the results of our method on the 3DPW dataset. Following Multi-HMR [2], we finetune on the 3DPW training set. We further evaluate the generaliza + +![](images/db5ac48be8e1bdec5fe2071f768509f4940c5e1edd4a932afaf6af9e6f3858ec.jpg) +Figure 4. Comparison with SOTA methods [2, 44-46] on in-the-wild images from the Internet. Red dashed circles highlight areas with incorrect estimations. The third case is left blank due to the small scale of individuals. Please zoom in for details. + +Table 3. Comparison with SOTA methods on CMU Panoptic [15] test set. We report MPJPE for four activities and the average. + +
MethodHaggling ↓Mafia ↓Ultimatum ↓Pizza ↓Avg. ↓
CRMH [14]129.6133.5153.0156.7143.2
3DCrowdNet [7]109.6135.9129.8135.6127.6
ROMP [45]110.8122.8141.6137.6128.2
BEV [46]90.7103.7113.1125.2109.5
PSVT [36]88.797.9115.2121.2105.7
Ours67.978.595.894.684.2
+ +![](images/6daf1a86b3c79a0f77d78ca5374e236ae0576ce473de792a0539599ecc73c5ae.jpg) +Figure 5. Qualitative comparison of different resolutions for our baseline. Resolution for the baseline: (a) 518, (b) 1288. Red dashed circles highlight differences; zoom in for details. + +tion performance of our method on the MuPoTS and CMU Panoptic datasets, as shown in Tab. 2 and Tab. 3. For Panoptic dataset, we follow the common evaluation protocol [36, 45, 46] and exclude Multi-HMR [2] due to potential inconsistencies in joint format. Our method consistently achieves the best results across all datasets, demonstrating exceptional generalization capability. + +Qualitative Results. Fig. 4 presents a visual comparison between our method and existing SOTA approaches. The images are sourced from the Internet. Our method provides the most accurate overall estimations, especially in the third case, where individuals are far from the camera, resulting in very small scales. Despite this challenge, our method remains highly robust, delivering precise estimations. In contrast, ROMP [45] and BEV [46] fail to detect all individuals, while AiOS [44] and Multi-HMR [2], despite using + +Table 4. Ablation studies on BEDLAM [3] validation set. We report MVE for different scale ranges and the average (Avg.). + +
Ablation0-20%20-40%40-60%60-80%80%+Avg.
(a) Drop all59.955.161.566.670.857.2
(b) No pooling60.354.557.059.864.156.1
(c) Pooling×260.754.557.562.166.556.3
(d) Ours60.054.457.260.362.756.0
+ +higher-resolution models, show significant estimation errors. + +# 4.4. Ablation Study + +We conduct ablation studies using only the AGORA [33] and sampled BEDLAM [3] datasets to validate model designs, and report results on their validation sets. + +Impact of resolution. Fig. 2 shows baseline results with different input resolutions on the AGORA [33] validation set. As resolution increases, the error across individuals in different scale ranges consistently decreases, with the most significant drop (35mm) in the smallest scale range. In contrast, the improvement in the large-scale range is less pronounced. However, the performance gain comes at the cost of a significant drop in computational efficiency, reducing the frame rate from real-time (32 FPS) to 5 FPS. Fig. 5 shows an example, clearly illustrating that higher resolution greatly improves estimation for small-scale individuals. + +Background tokens $\mathcal{T}_{\mathrm{B}}$ . We investigate the impact of background tokens on model performance and report the MVE across different scale ranges on the BEDLAM validation set in Tab. 4. In (a), we discard all background tokens and observe stable performance at the smallest scale, but a decline as the scale increases, indicating that background context becomes more important for larger scales. In (b), skipping pooling for background tokens yields nearly the same performance but introduces higher computational cost. In (c), applying pooling twice with the same grouping strategy results in a performance drop for large-scale instances, highlighting + +![](images/bea354c11784882fb0b6c14e84c5533b86b14f97854ccd6e4bfcacaab7e9511f.jpg) +Figure 6. Visualization of our predicted patch-level scale maps S and scale-adaptive tokens $\mathcal{T}_{\mathrm{SA}}$ . We show the predicted scale maps, scale-adaptive tokens, and the estimated meshes overlaid on the image. In the scale map, colors represent scale values, with uncolored areas indicating background patches. A gradient from blue (large-scale individuals) to rose (small-scale individuals) illustrates the scale distribution. We visualize scale-adaptive tokens as patches of different sizes and colors on the image, with low-resolution patches blurred. + +![](images/87bbf41706ae5c72aa2656c33dc729d41efd5efc2787fc1094be52253b87cdde.jpg) +Figure 7. Qualitative results of our method. We visualize four cases from AGORA [33], BEDLAM [3], CMU Panoptic [15] and 3DPW [49] datasets, displaying the input, the estimated mesh overlay, and an elevated view (top to bottom). + +the significance of background context for estimation. We hypothesize that large-scale people face more truncation, relying more on global context which requires further exploration. We provide additional efficiency comparisons in Sec. B, showing that pooling once achieves the best balance between performance and efficiency. + +# 4.5. Qualitative Results + +Fig. 6 visualize the predicted patch-level scale map and scale-adaptive tokens by our model, enabling dynamic adjustment of tokens for accurate estimations. Fig. 7 presents additional visualizations of predicted meshes, demonstrating the effectiveness of our method in various scenarios. Please refer to Sec. B for more qualitative results and discussion. + +# 5. Conclusion + +We present a novel one-stage framework for real-time multiperson 3D mesh estimation from an RGB image. By introducing scale-adaptive tokens that dynamically adjust based on the relative scale of individuals, our method effectively balances efficiency and accuracy, retaining the benefits of high-resolution processing while achieving real-time performance. Experiments demonstrate that our method is the best real-time model, offering competitive performance and superior generalization compared to SOTAs. Our limitations include the lack of height-aware scale estimation, which may cause depth errors, and the current focus on body-only mesh estimation, extendable to whole-body in the future. + +# Acknowledgments + +This work was supported by 2022ZD0114904 and NSFC-6247070125. We thank Yiyang Wang for the valuable discussions during the early stage of this work. + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6, 1 +[2] Fabien Baradel, Matthieu Armando, Salma Galaaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas. Multi-hmr: Multi-person whole-body human mesh recovery in a single shot. In ECCV, 2024. 2, 3, 4, 6, 7, 1, 5 +[3] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 6, 7, 8, 1, 2, 3, 4, 5 +[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In ICLR, 2023. 3 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 3, 4, 5, 1 +[6] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In ECCV, 2022. 3 +[7] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3d human mesh from in-the-wild crowded scenes. In CVPR, pages 1475-1484, 2022. 2, 3, 6, 7 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[9] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. Tore: Token reduction for efficient human mesh recovery with transformer. In ICCV, pages 15143-15155, 2023. 3 +[10] Jose M Facil, Benjamin Ummenhofer, Huizhong Zhou, Luis Montesano, Thomas Brox, and Javier Civera. Cam-convs: Camera-aware multi-scale convolutions for single-view depth. In CVPR, pages 11826-11835, 2019. 5 +[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 3, 1 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3 +[13] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predic + +tive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6, 1 +[14] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In CVPR, pages 5579-5588, 2020. 2, 3, 6, 7 +[15] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, pages 3334-3342, 2015. 6, 7, 8, 2, 3 +[16] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1 +[17] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In CVPR, pages 10863-10872, 2019. 6, 1 +[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 3 +[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, pages 12939-12948, 2021. 3 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6, 1 +[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5, 6 +[22] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, et al. Group pose: A simple baseline for end-to-end multi-person pose estimation. In ICCV, pages 15029-15038, 2023. 3 +[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 3, 5, 1, 2 +[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3, 6, 1 +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 1 +[26] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In CVPR, pages 534-543, 2023. 3 +[27] Xiaoxuan Ma, Jiajun Su, Yuan Xu, Wentao Zhu, Chunyu Wang, and Yizhou Wang. Vmarker-pro: Probabilistic 3d human mesh estimation from virtual markers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3 +[28] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, pages 506-516, 2017. 6 + +[29] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 6, 1, 3 +[30] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 3 +[31] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralnot: Neural annotator for 3d human mesh training sets. In CVPR, pages 2299-2307, 2022. 6, 1 +[32] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 6 +[33] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In CVPR, pages 13468-13478, 2021. 1, 2, 6, 7, 8, 3, 4, 5 +[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 6, 1 +[35] Zhongwei Qiu, Qiansheng Yang, Jian Wang, and Dongmei Fu. Dynamic graph reasoning for multi-person 3d pose estimation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3521-3529, 2022. 2, 3 +[36] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In CVPR, pages 21254-21263, 2023. 3, 6, 7 +[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3 +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3 +[39] Cedric Renggli, André Susano Pinto, Neil Houlsby, Basil Mustafa, Joan Puigcerver, and Carlos Riquelme. Learning to merge tokens in vision transformers. arXiv preprint arXiv:2202.12015, 2022. 3 +[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5 +[41] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In CVPRW, pages 4613-4622, 2023. 3 +[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 2 + +[43] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In CVPR, pages 11069-11078, 2022. 3 +[44] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi-Sing Leung, Ziwei Liu, Lei Yang, et al. Aios: All-in-one-stage expressive human pose and shape estimation. In CVPR, pages 1834-1843, 2024, 2, 3, 4, 6, 7, 1, 5 +[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 3, 6, 7, 1, 2 +[46] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 2, 3, 6, 7, 1, 4 +[47] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 3 +[48] A Vaswani. Attention is all you need. NeurIPS, 2017. 3 +[49] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 1, 6, 8, 3, 4 +[50] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In CVPR, pages 979–989, 2024. 3 +[51] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 3 +[52] Sen Yang, Wen Heng, Gang Liu, Guozhong Luo, Wankou Yang, and YU Gang. Capturing the motion of every joint: 3d human pose and shape estimation with independent tokens. In ICLR, 2023. 3 +[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In ECCV, 2022. 3 +[54] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, pages 10809-10818, 2022. 3 +[55] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 2, 3 +[56] Andrei Zanfir, Elisabeta Marinoiu, Mihai Zanfir, Alin-Ionut Popa, and Cristian Sminchisescu. Deep network for the integrated 3d sensing of multiple people in natural images. In NeurIPS, 2018. 2, 3 +[57] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 3 +[58] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2023. 3 + +[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In CVPR, pages 15085-15099, 2023. 3 +[60] Wentao Zhu, Jason Qin, Yuke Lou, Hang Ye, Xiaoxuan Ma, Hai Ci, and Yizhou Wang. Social motion prediction with cognitive hierarchies. In NeurIPS, 2023. 2 +[61] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlun Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang: Human motion generation: A survey. IEEE TPAMI, 46(4): 2430-2449, 2024. 2 \ No newline at end of file diff --git a/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/images.zip b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4f25cc41d72b4599d40fd5d55e4aa1be13b4c8f5 --- /dev/null +++ b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59bb3f0d848798f7dedc20241d54a6510e9385abb04b23febcda9d3c5503eedf +size 805812 diff --git a/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/layout.json b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5f49c0952bd99d506ba6aa84ee4970ad3bfe368e --- /dev/null +++ b/2025/SAT-HMR_ Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens/layout.json @@ -0,0 +1,8556 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 493, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 493, + 139 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 493, + 139 + ], + "type": "text", + "content": "SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 167, + 465, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 167, + 465, + 182 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 465, + 182 + ], + "type": "text", + "content": "Chi Su1 Xiaoxuan Ma1,☑ Jiajun Su5 Yizhou Wang1,2,3,4,☑" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "text", + "content": " Center on Frontiers of Computing Studies, School of Computer Science, Peking University \n" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "text", + "content": " Inst. for Artificial Intelligence, Peking University " + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "text", + "content": " Nat'l Eng. Research Center of Visual Technology \n" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "text", + "content": " State Key Laboratory of General Artificial Intelligence, Peking University \n" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 183, + 560, + 246 + ], + "type": "text", + "content": " International Digital Economy Academy (IDEA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 250, + 527, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 250, + 527, + 277 + ], + "spans": [ + { + "bbox": [ + 94, + 250, + 527, + 277 + ], + "type": "text", + "content": "suchi@stu.pku.edu.cn, {maxiaoxuan, yizhou.wang}@pku.edu.cn, sujiajun@idea.edu.cn Corresponding authors" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 297, + 306, + 456 + ], + "blocks": [ + { + "bbox": [ + 67, + 297, + 306, + 456 + ], + "lines": [ + { + "bbox": [ + 67, + 297, + 306, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 306, + 456 + ], + "type": "image", + "image_path": "50f12f786f864623e3d98e0bc9316ad10c2c51febf4a4279ac54bb7aa782954e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 461, + 193, + 471 + ], + "lines": [ + { + "bbox": [ + 180, + 461, + 193, + 471 + ], + "spans": [ + { + "bbox": [ + 180, + 461, + 193, + 471 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 323, + 300, + 547, + 456 + ], + "blocks": [ + { + "bbox": [ + 323, + 300, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 323, + 300, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 323, + 300, + 547, + 456 + ], + "type": "image", + "image_path": "ef0882c895bd45834df76479317ffebbcc4c36391a5d247de3a5b86540791457.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 461, + 443, + 471 + ], + "lines": [ + { + "bbox": [ + 430, + 461, + 443, + 471 + ], + "spans": [ + { + "bbox": [ + 430, + 461, + 443, + 471 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 473, + 555, + 550 + ], + "lines": [ + { + "bbox": [ + 55, + 473, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 555, + 550 + ], + "type": "text", + "content": "Figure 1. (a) We propose scale-adaptive tokens in our one-stage framework for real-time multi-person 3D mesh estimation. Our method introduces scale-adaptive tokens, dynamically adjusted based on the relative size of individuals in the image, to more efficiently encode features, enabling real-time and accurate multi-person mesh estimation. We present a conceptual visualization of the scale-adaptive tokens. The right column visualizes the predicted meshes projected onto an image from 3DPW [49] dataset and from an elevated view. (b) Comparison of estimation error and inference time across different methods, with input resolutions in parentheses. Our method, using a mixed resolution with a base resolution of 644, achieves comparable performance to state-of-the-art methods on AGORA [33] test set while maintaining real-time inference efficiency. Code and models are available at https://ChiSu001.github.io/SAT-HMR/." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 563, + 199, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 563, + 199, + 574 + ], + "spans": [ + { + "bbox": [ + 152, + 563, + 199, + 574 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 594, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 713 + ], + "type": "text", + "content": "We propose SAT-HMR, a one-stage framework for real-time multi-person 3D human mesh estimation from a single RGB image. While current one-stage methods, which follow a DETR-style pipeline, achieve state-of-the-art (SOTA) performance with high-resolution inputs, we observe that this particularly benefits the estimation of individuals in smaller scales of the image (e.g., those of young age or far from the camera), but at the cost of significantly increased computation overhead. To address this, we introduce scale-adaptive tokens that are dynamically adjusted based on the relative" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 564, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 556, + 708 + ], + "type": "text", + "content": "scale of each individual in the image within the DETR framework. Specifically, individuals in smaller scales are processed at higher resolutions, larger ones at lower resolutions, and background regions are further distilled. These scale-adaptive tokens more efficiently encode the image features, facilitating subsequent decoding to regress the human mesh, while allowing the model to allocate computational resources more effectively and focus on more challenging cases. Experiments show that our method preserves the accuracy benefits of high-resolution processing while substantially reducing computational cost, achieving real-time inference with performance comparable to SOTA methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16796" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 289, + 249 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 289, + 249 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 289, + 249 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 289, + 249 + ], + "type": "image", + "image_path": "14f3515040cc45f35aa5cab6495e04b8f5095170809d1f19ee2e268cf7936c12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 298, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 298, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 298, + 352 + ], + "type": "text", + "content": "Figure 2. Estimation errors and FPS of baselines with different resolutions and our method across individuals at various scales. The scale of an individual refers to the person's size relative to the overall image and please refer to Sec. 3.3 for mathematical definition. The colored lines show the Mean Vertex Error (MVE) errors (left y-axis) of the baselines with different resolutions (Res.) on the AGORA [33] validation set. The colored markers on the right y-axis indicate the FPS of the corresponding models. Our method adopts a mixed resolution with a base resolution of 644." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 367, + 136, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 136, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 136, + 380 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 388, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 297, + 579 + ], + "type": "text", + "content": "Multi-person 3D human mesh estimation from a single RGB image aims to localize all individuals in the scene and estimate their 3D meshes, typically represented by a parametric human model such as SMPL [24]. This is a fundamental task with broad applications, spanning social interaction [42, 60], game production [61], etc. Existing methods primarily focus on addressing challenges such as monocular depth ambiguity and inter-person interactions, and can be broadly categorized into two types: multi-stage and one-stage approaches. Multi-stage methods [7, 14, 35, 55, 56] typically detect and process each individual by cropping and resizing them to a uniform, relatively high resolution, which generally improves performance. However, this overlooks the global context, such as the original scales of individuals and their relative positions within the entire image, potentially leading to suboptimal results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 298, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 298, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 298, + 713 + ], + "type": "text", + "content": "To address these limitations, one-stage methods gradually emerge as a promising alternative [2, 44-46]. These approaches operate on the entire image without explicit cropping, preserving global context and enabling end-to-end training. Building on the success of DETR [5] in object detection, recent works [2, 44] adopt a DETR-style pipeline. They mainly focus on designing queries and decoders to estimate the human mesh, e.g. Multi-HMR [2] uses queries to represent each person and decodes the corresponding human mesh. However, these methods rely on high-resolution inputs to achieve state-of-the-art (SOTA) results, as shown" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "in Fig. 1 (b), which significantly increases computational cost and limits real-time processing capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "text", + "content": "Building on the findings of [2, 44], we observe that high-resolution input plays a crucial role in improving accuracy. To explore this, we design a naive baseline following similar architecture and analyze how its estimation error changes for individuals of different scales as we increase the input resolution. We define scale as the size of the person relative to the entire image1. As shown in Fig. 2, increasing the input resolution consistently leads to a decrease in the estimation error, particularly for smaller-scale individuals (e.g., kids, those far from the camera or those in closed-poses). Notably, for individuals in the " + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "inline_equation", + "content": "0 - 10\\%" + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "text", + "content": " scale, we observe an error decrease of nearly " + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "inline_equation", + "content": "35\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "text", + "content": " when comparing the highest-resolution model (blueviolet) to the lowest-resolution one (yellow). However, processing the entire image at high resolution incurs significant computational costs, especially when individuals are close to the camera and occupy a large portion of the image. In such cases, these individuals are already represented by many tokens, and using more tokens with high resolution would only greatly increase the cost while offering marginal improvement (i.e., a slight error reduction in cases where individuals cover " + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "inline_equation", + "content": "30 + \\%" + }, + { + "bbox": [ + 313, + 97, + 556, + 393 + ], + "type": "text", + "content": " of the image in Fig. 2). Therefore, both our best baseline model (blueviolet) and recent SOTA methods [2, 44] in Fig. 1 (b) fail to achieve real-time inference and run at about 5 FPS, highlighting a common limitation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 634 + ], + "type": "text", + "content": "To address this, we extend our baseline built on DETR [5] and shift our focus toward more efficient feature encoding. Specifically, we introduce scale-adaptive tokens, which are dynamically adjusted based on the scale of individuals in the image. We first predict a patch-level scale map, assigning a scale value to each token. Tokens corresponding to smaller-scale individuals are replaced with higher-resolution tokens. Additionally, background tokens identified by the scale map are further distilled to reduce computational cost. These scale-adaptive tokens capture individuals at varying scales with the appropriate level of detail, as shown in Fig. 1 (a). The tokens are then processed by the decoder and prediction heads to regress the human mesh. By efficiently allocating computational resources to the more challenging cases, our method strikes a better balance between accuracy and speed. As shown in Fig. 2, our method (red) achieves performance comparable to high-resolution input while maintaining real-time inference speed at 24 FPS. Experiments on benchmark datasets further demonstrate the effectiveness of our method, achieving competitive performance with high efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 635, + 512, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 635, + 512, + 646 + ], + "spans": [ + { + "bbox": [ + 326, + 635, + 512, + 646 + ], + "type": "text", + "content": "To conclude, our contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 647, + 553, + 694 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 647, + 553, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 647, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 647, + 553, + 682 + ], + "type": "text", + "content": "- We propose scale-adaptive tokens that are dynamically adjusted to handle individuals of varying scales, enabling more efficient feature encoding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 683, + 553, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 683, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 314, + 683, + 553, + 694 + ], + "type": "text", + "content": "- By employing scale-adaptive tokens, our method achieves" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 467, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 467, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 467, + 713 + ], + "type": "text", + "content": "See Eq. (1) for detailed definition of scale." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16797" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 296, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 296, + 95 + ], + "type": "text", + "content": "performance comparable to using full-image high resolution, but with significantly lower computational cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 144 + ], + "type": "text", + "content": "- Our approach achieves SOTA performance with up to 5x the speed of the top-performing methods, making it the best real-time model for multi-person 3D mesh estimation, running at 24 FPS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 140, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 140, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 140, + 168 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 177, + 260, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 260, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 260, + 189 + ], + "type": "text", + "content": "2.1. Multi-Person Human Mesh Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 195, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 338 + ], + "type": "text", + "content": "Multi-person human mesh estimation from a single RGB image aims to regress all human meshes, typically represented by parametric models such as SMPL [24] and SMPL-X [34]. Existing methods can be categorized into multi-stage and one-stage approaches. Multi-stage methods [7, 11, 14, 35, 53, 55, 56] use off-the-shelf human detectors [38] to crop each person's region, followed by single-person mesh estimation [26, 27, 50, 57, 59]. This approach preserves relatively high-resolution images for each cropped region, generally leading to higher accuracy [44]. However, its reliance on detected bounding boxes and the lack of global context make it challenging to handle occlusion [45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 339, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 531 + ], + "type": "text", + "content": "In contrast, one-stage methods [2, 36, 44-46] estimate human meshes for all individuals by encoding the entire image and predicting simultaneously. Pioneer work like ROMP [45] and BEV [46] use CNN backbones [12] to extract global features for mesh regression. However, the low input resolution limits the expressiveness of the distilled features [44], leading to suboptimal performance. Recent methods like AiOS [44] and Multi-HMR [2] adopt DETR-style architectures to achieve SOTA performance at higher resolutions, but with significant computational overhead, making real-time use impractical (e.g., Multi-HMR runs at only 4 FPS as shown in Fig. 1). Instead, we argue that tokenizing the entire image at high resolution is unnecessary. By adjusting patch resolution based on the scale and focusing computational resources on more challenging areas, we achieve real-time inference with competitive performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 274, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 274, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 274, + 552 + ], + "type": "text", + "content": "2.2. Transformers in Human Mesh Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": "Transformers [48], originally designed for sequential inputs, have demonstrated strong performance in various vision tasks, such as image classification with Vision Transformers (ViTs) [8]. And it has been widely adopted in human mesh estimation [2, 6, 9, 18, 19, 44, 52]. For single-person settings, methods such as [6, 9, 18, 19, 52] focus on improving the accuracy through various architectural designs. Inspired by the DETR series [5, 23, 58] in object detection, recent methods have adopted a DETR-style framework for multi-person pose and mesh estimation [2, 22, 43, 44, 51], enabling one-stage estimation for all individuals. These works primarily focus on designing decoder queries to represent humans, keypoints, and mesh parameters. For instance," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "AiOS [44] refines predictions by progressively filtering and expanding queries in the decoder, while Multi-HMR [2] uses each query to represent a detected person. Our method also builds upon the DETR framework but focuses on improving token representations in the encoder to capture scale-aware human features, enabling accurate 3D mesh regression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 152, + 476, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 476, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 476, + 163 + ], + "type": "text", + "content": "2.3. Efficient Vision Transformers" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 169, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 350 + ], + "type": "text", + "content": "Although ViTs [8] achieves significant success across various vision tasks, their quadratic computational complexity remains a major challenge, particularly for tasks that require dense predictions such as object detection [47]. As a result, efforts to accelerate ViTs primarily focus on reducing the overall number of image tokens. For instance, [30, 37, 54] drop less informative tokens, [4, 39] merge similar tokens, and [41] proposes a mixed-resolution image tokenization scheme. Recently, TORE [9] accelerates single-person 3D mesh estimation by pruning background tokens. However, since background still provides important context, entirely discarding it leads to performance degradation (see Sec. 4.4). In this work, we propose a new perspective with scale-adaptive tokens that achieve a more effective trade-off between performance and computational cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 359, + 370, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 370, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 370, + 372 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 380, + 400, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 400, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 400, + 391 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": "We use SMPL [24] as our body model " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": ", which uses pose and shape parameters to represent the human body. Specifically, the pose parameters " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{24\\times 3}" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": " are the relative rotations of 24 body joints, and the shape parameters " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": " describe the body shape. The model outputs a human mesh " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathcal{M}(\\pmb {\\theta},\\pmb {\\beta})\\in \\mathbb{R}^{6890\\times 3}" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": " and the 3D joints can be obtained via " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{J} = \\mathbf{MV}\\in \\mathbb{R}^{J\\times 3}" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\in \\mathbb{R}^{J\\times 6890}" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": " is the joint regressor, and " + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 396, + 555, + 493 + ], + "type": "text", + "content": " represents the number of joints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 500, + 382, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 382, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 382, + 511 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 555 + ], + "type": "text", + "content": "Building on recent methods [2, 44], we utilize a DETR-style pipeline [5], consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 629 + ], + "type": "text", + "content": "Naive baseline. While high-resolution inputs are shown to significantly enhance estimation accuracy [2, 44], they also introduce considerable computational overhead. To further explore the trade-off between resolution and efficiency, we first introduce a simple baseline using a DETR-style architecture, as depicted in Fig. 3 (top)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Given an RGB image " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{hr}} \\in \\mathbb{R}^{H_{\\mathrm{hr}} \\times W_{\\mathrm{hr}} \\times 3}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", it is divided into regular patches of size " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". Patch embedding and positional embedding are then applied to these image patches, resulting in a sequence of image feature tokens " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_1, t_2, \\dots, t_k\\}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "k = H_{\\mathrm{hr}} / P \\times W_{\\mathrm{hr}} / P" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". These tokens are subsequently processed by a Transformer encoder. In the decoder, a set of human queries is initialized" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16798" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 542, + 321 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 542, + 321 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 542, + 321 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 542, + 321 + ], + "type": "image", + "image_path": "5cd82588e1232c0c4c1cedbbfe296a314ec9874064ec616bf3df315ff3f4b0f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": "Figure 3. Overview of (top) the baseline method and (bottom) our method with scale-adaptive tokens. Top: Our baseline method adopts a DETR-style [5] pipeline consisting of a Transformer encoder, decoder, and prediction heads for regressing SMPL parameters. Bottom: Our method focuses on efficient feature encoding using scale-adaptive tokens. Specifically, low-resolution and high-resolution patches are extracted from the input images " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{hr}}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": ", respectively. A scale head network predicts a patch-level scale map " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": " from the low-resolution tokens, classifying them into three categories: background, small-scale, and large-scale. This scale map guides the pruning and pooling of low-resolution tokens " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LR}}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": " and indicates which patches should be replaced by high-resolution ones. By concatenating the pooled background tokens " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{B}}^{\\prime}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": ", the remaining large-scale low-resolution tokens " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LARGE}}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": ", and the high-resolution tokens " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{HR}}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": ", we obtain scale-adaptive tokens " + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{SA}}" + }, + { + "bbox": [ + 54, + 324, + 555, + 413 + ], + "type": "text", + "content": ". These tokens are then processed by the encoder, decoder, and multiple prediction heads to regress the human mesh." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{q_1, q_2, \\ldots, q_n\\}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": " is a predefined hyperparameter large enough to detect all the individuals. After being processed by the decoder, these human queries are passed through multiple prediction heads: pose head " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{p}}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ", shape head " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{s}}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ", translation head " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{t}}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ", and box head " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{b}}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": " to predict SMPL parameters (pose " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ", shape " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ", and 3D translation " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": ") and bounding boxes, respectively. Each of the four prediction heads is implemented as an Multi-Layer Perceptron (MLP). Additionally, a linear projection is applied to the human queries to compute confidence scores, which are used to filter valid predictions with a threshold " + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{d}}" + }, + { + "bbox": [ + 54, + 426, + 297, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 565, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 565, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 565, + 297, + 685 + ], + "type": "text", + "content": "Motivation. We evaluate the baseline with different image resolutions as input (see Fig. 2, Sec. 4.4 for details) and observe that high-resolution inputs significantly improve estimation accuracy, particularly for small-scale individuals whose size is relatively small to the image. However, for large-scale individuals, e.g., those closer to the camera, the benefits of higher resolution are much less pronounced. Despite these improvements, high-resolution inputs bring a substantial increase in computational cost, which is consistent with findings from prior work [2, 44] (see Fig. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "This motivates us to explore more efficient strategies that balance performance and speed and we aim to encode the im" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 426, + 556, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 535 + ], + "type": "text", + "content": "age feature more efficiently from an earlier stage. For large-scale individuals, additional tokens may be unnecessary; instead, computational resources should be allocated more effectively. To address this, we introduce scale-adaptive tokens, which replace the uniformly partitioned tokens used in the baseline. In the following sections, we will explain how the scale is defined and computed, and how the tokens are adjusted accordingly (Sec. 3.3), followed by a discussion of the training process (Sec. 3.4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 540, + 443, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 443, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 443, + 553 + ], + "type": "text", + "content": "3.3. Scale-Adaptive Tokens" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3 (bottom), we begin with a low-resolution image " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "(H, W) = (H_{\\mathrm{hr}} / 2, W_{\\mathrm{hr}} / 2)" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", and first uniformly partition it into low-resolution patches. Our goal is to adjust the resolution of each patch based on the scale of the individual it represents. To achieve this, we introduce a patch-level scale map that guides the resolution adjustments for each patch. Specifically, this map determines which low-resolution tokens should be further subdivided and replaced by high-resolution ones, and which tokens can be compressed. The high-resolution patches are extracted from a corresponding high-resolution image " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{hr}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ". After being processed by the Transformer layers, the remaining low-resolution tokens and the newly generated high-resolution" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16799" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "tokens are combined into scale-adaptive tokens, which provide a more efficient representation with varying levels of detail. These scale-adaptive tokens are then passed through additional Transformer layers and, finally, to the decoder to regress the SMPL parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": "Patch-level scale map. We define a patch-level scale map " + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in [0,1]^{\\frac{H}{P} \\times \\frac{W}{P} \\times 2}" + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": " for patches extracted from the low-resolution image " + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": ". For each patch, the scale map " + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(i,j) = (c,s)" + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": " contains two values: " + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": " represents the confidence indicating whether the patch overlaps with an individual, where " + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": " denotes the background; " + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 136, + 296, + 220 + ], + "type": "text", + "content": " represents the scale of the individual within the patch, defined as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 222, + 296, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 222, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 132, + 222, + 296, + 236 + ], + "type": "interline_equation", + "content": "s = \\min \\left(d _ {\\mathrm {b b}} / S _ {\\mathrm {h r}}, 1\\right), \\tag {1}", + "image_path": "70b709e6ed4323ce432590d486e3b74155d4182c2a4699cf67e531872c4e4ca2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{bb}}" + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "text", + "content": " is the diagonal length of the bounding box in " + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{hr}}" + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{hr}} = \\max (H_{\\mathrm{hr}},W_{\\mathrm{hr}})" + }, + { + "bbox": [ + 55, + 240, + 296, + 312 + ], + "type": "text", + "content": ". If the patch overlaps with multiple individuals, we use the scale of the individual closest to the camera, i.e., the one with the smaller depth. The scale is only defined for patches overlapping with individuals, and for background patches, the scale value is ignored." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 419 + ], + "type": "text", + "content": "To predict the patch-level scale map, we pass the low-resolution patches through a shallow transformer encoder with " + }, + { + "bbox": [ + 55, + 312, + 296, + 419 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{lr}}" + }, + { + "bbox": [ + 55, + 312, + 296, + 419 + ], + "type": "text", + "content": " layers, followed by a scale head network, implemented as an MLP. As illustrated in Fig. 3, patches without color are classified as background, while colored regions represent patches overlapping with individuals. A color gradient from blue to rose indicates the scale of the individuals, with blue representing large-scale individuals and rose representing small-scale individuals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": "Scale-adaptive encoder. Building on the predicted scale map S, the low-resolution tokens " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LR}} = \\{t_{1}^{\\mathrm{lr}}, t_{2}^{\\mathrm{lr}}, \\ldots, t_{k_{\\mathrm{lr}}}^{\\mathrm{lr}}\\}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": " generated from the low-resolution patches are then classified into three groups: background " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": ", small-scale " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{SMALL}}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": ", and large-scale " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LARGE}}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": " tokens. Tokens identified as background, based on a confidence score " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": " that meets a threshold " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\alpha_{c}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": ", are marked as background tokens " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{B}} = \\{t_{1}^{\\mathrm{b}}, t_{2}^{\\mathrm{b}}, \\ldots, t_{k_{\\mathrm{b}}}^{\\mathrm{b}}\\}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": ". The remaining tokens are further categorized by their scale values " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": " using a scale threshold " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": " and are assigned to either the small-scale category, denoted by " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{SMALL}} = \\{t_{1}^{\\mathrm{small}}, t_{2}^{\\mathrm{small}}, \\ldots, t_{k_{\\mathrm{small}}}^{\\mathrm{small}}\\}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": ", or the large-scale category, denoted by " + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LARGE}} = \\{t_{1}^{\\mathrm{large}}, t_{2}^{\\mathrm{large}}, \\ldots, t_{k_{\\mathrm{large}}}^{\\mathrm{large}}\\}" + }, + { + "bbox": [ + 55, + 424, + 296, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": "Small-scale tokens are pruned and replaced by their high-resolution counterparts. High-resolution tokens are generated by applying a shallow transformer encoder with " + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{hr}}" + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": " layers to the high-resolution image " + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{hr}}" + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{hr}} = N_{\\mathrm{lr}}" + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": " to ensure feature alignment. The small-scale tokens are then replaced by their corresponding high-resolution tokens (Fig. 3), resulting in an expanded set of high-resolution tokens, denoted as " + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{HR}} = \\{t_1^{\\mathrm{hr}}, t_2^{\\mathrm{hr}}, \\dots, t_{k_{\\mathrm{hr}}}^{\\mathrm{hr}}\\}" + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{hr}} = 4k_{\\mathrm{small}}" + }, + { + "bbox": [ + 55, + 571, + 296, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Although background can provide valuable contextual information, their feature encoding can be further optimized. To this end, we distill the background tokens " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " by spatially pooling every four neighboring tokens, resulting in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "text", + "content": "pooled background tokens " + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{B}}^{\\prime} = \\{t_{1}^{\\mathrm{b}^{\\prime}}, t_{2}^{\\mathrm{b}^{\\prime}}, \\ldots, t_{k_{\\mathrm{b}}^{\\prime}}^{\\mathrm{b}^{\\prime}}\\}" + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "text", + "content": ". To avoid irregular pooling, some low-resolution tokens that cannot be grouped remain unchanged. As a result, " + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{b}}^{\\prime}" + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "text", + "content": " does not necessarily equal " + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{b}} / 4" + }, + { + "bbox": [ + 313, + 71, + 553, + 145 + ], + "type": "text", + "content": ". By retaining these tokens, we better preserve their positional information, which facilitates subsequent positional encoding operations [23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "text", + "content": "The remaining tokens in " + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LR}}" + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "text", + "content": " representing large-scale individuals, " + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{LARGE}}" + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "text", + "content": ", remain unchanged. Next, we integrate the different tokens, including pooled background tokens and high-resolution tokens from selected regions. This results in scale-adaptive tokens, denoted as " + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{SA}} = \\{\\mathcal{T}_{\\mathrm{B}}^{\\prime}, \\mathcal{T}_{\\mathrm{LARGE}}, \\mathcal{T}_{\\mathrm{HR}}\\}" + }, + { + "bbox": [ + 313, + 146, + 555, + 241 + ], + "type": "text", + "content": ". Compared to the uniform low-resolution tokens, this approach allocates feature details more efficiently, preserving different levels of detail for different individuals and regions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "type": "text", + "content": "These tokens, " + }, + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{SA}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "type": "text", + "content": ", are then processed by another Transformer encoder with " + }, + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sa}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 289 + ], + "type": "text", + "content": " layers and further decoded in subsequent stages. Fig. 3 (bottom) provides an overview of the scale-adaptive encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 342 + ], + "type": "text", + "content": "Decoder and prediction heads. The decoder and prediction heads follow the architecture of our baseline approach as mentioned in Sec. 3.2. For detailed architectures, please refer to the supplementary material (Sec. A.1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 349, + 410, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 410, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 410, + 361 + ], + "type": "text", + "content": "3.4. Training Losses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "text", + "content": "Similar to DETR [5], we first match our human predictions to ground-truth (GT) using Hungarian Matching, and we leverage bounding boxes, confidence scores, and projected joints during the matching. This matching step is necessary before computing the overall loss " + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "text", + "content": ". More details on the matching process are provided in Sec. A.1. The loss function, " + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "text", + "content": ", is a weighted sum of various terms, with each " + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 366, + 554, + 462 + ], + "type": "text", + "content": " representing a hyperparameter:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 469, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 469, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 321, + 469, + 545, + 498 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {m a p}} \\mathcal {L} _ {\\text {m a p}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {p o s e}} \\mathcal {L} _ {\\text {p o s e}} + \\lambda_ {\\text {s h a p e}} \\mathcal {L} _ {\\text {s h a p e}} \\\\ + \\lambda_ {\\mathrm {j 3 d}} \\mathcal {L} _ {\\mathrm {j 3 d}} + \\lambda_ {\\mathrm {j 2 d}} \\mathcal {L} _ {\\mathrm {j 2 d}} + \\lambda_ {\\mathrm {b o x}} \\mathcal {L} _ {\\mathrm {b o x}} + \\lambda_ {\\det} \\mathcal {L} _ {\\det}. \\\\ \\end{array}", + "image_path": "a46332af12475ee6a4f8f20d6b68e71c822aaa115220fe286e54a4bf3fcc3286.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "text", + "content": "Scale map loss " + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "text", + "content": ". To supervise the patch-level scale map, we apply focal loss [21] to the predicted confidence " + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "text", + "content": " and L1 loss to the predicted scale " + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "text", + "content": ". The total loss, " + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 506, + 554, + 566 + ], + "type": "text", + "content": ", is the sum of these two terms. The GT for confidence and scale is precomputed, as described in Sec. 3.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": "Depth loss " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": ". We supervise the root depth (i.e., the z-dimension) of the 3D joints " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": ", regressed from the predicted mesh " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": " (Sec. 3.1). Specifically, we normalize the depth based on the focal length, following [10]. The depth loss is calculated as " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}} = \\left|\\frac{1}{\\tilde{d}} - \\frac{f}{\\tilde{f}d}\\right|" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": " is the predefined focal length of our camera, and " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{d}" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{f}" + }, + { + "bbox": [ + 313, + 570, + 554, + 662 + ], + "type": "text", + "content": " are the GT depth and focal length, respectively. See Sec. A.1 for more details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Other losses. For the pose parameters loss " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pose}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", shape parameters loss " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{shape}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", 3D joints loss " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{j3d}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", and projected 2D joints loss " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{j2d}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", we use the L1 distance. For the bounding box loss " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", we combine L1 loss with GIoU loss [40]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16800" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 84, + 102, + 529, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison with SOTA methods on AGORA test set [33]. \"Res.\" represents the input resolution. \"644*\" means we use a base resolution of 644 and adaptively scale certain regions up to a maximum resolution of 1288." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 102, + 529, + 201 + ], + "lines": [ + { + "bbox": [ + 84, + 102, + 529, + 201 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 529, + 201 + ], + "type": "table", + "html": "
MethodRes.Time (ms)MACs (G)F1-Score ↑Precision ↑Recall ↑MPJPE ↓MVE ↓NMJE ↓NMVE ↓
ROMP [45]51238.743.60.910.950.88108.1103.4118.8113.6
BEV [46]51250.648.90.930.960.90105.3100.7113.2108.3
PSVT [36]512--0.93--97.794.1105.1101.2
AiOS [44]1333405.2314.50.940.980.9063.957.568.061.2
Multi-HMR [2]89697.42075.10.93--82.877.689.083.4
Multi-HMR [2]1288231.76104.60.950.970.9365.361.168.764.3
Ours644*42.0133.10.950.980.9167.963.371.566.6
", + "image_path": "d6c6953e841e5bfbd18bb8b2c2e0666af6412a5e24f4aa5006de5500f86f62a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 295, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 238 + ], + "type": "text", + "content": "The detection loss " + }, + { + "bbox": [ + 55, + 213, + 295, + 238 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}" + }, + { + "bbox": [ + 55, + 213, + 295, + 238 + ], + "type": "text", + "content": " uses focal loss [21] to supervise the confidence scores of human queries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 138, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 138, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 138, + 260 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 176, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 176, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 176, + 277 + ], + "type": "text", + "content": "4.1. Datasets and metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 367 + ], + "type": "text", + "content": "We train our model on multi-person datasets including AGORA [33], BEDLAM [3], COCO [20], Crowdpose [17], and MPII [1], as well as the single-person dataset H3.6M [13]. AGORA and BEDLAM are synthetic datasets with high-quality 3D mesh GTs. For other multi-person datasets, we use pseudo GTs from NeuralAnnot [31] and only supervise projected 2D joints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 559 + ], + "type": "text", + "content": "For comparison with SOTA methods, we evaluate our method on AGORA [33], MuPoTS [29], 3DPW [49], and CMU Panoptic [15] datasets. Following previous works [2, 44-46], we report Mean Vertex Error (MVE) and Mean Per-Joint Position Error (MPJPE) before and after Procrustes-Alignment (PA). On AGORA, we also report F1-Score, precision, and recall for evaluating detection accuracy; Normalized Mean Vertex Error (NMVE) and Normalized Mean Joint Error (NMJE) that considered regression accuracy with detection accuracy. All metrics above, except detection accuracy, are reported in millimeters (mm). On MuPoTS-3D [29], we report the Percentage of Correctly estimated Keypoints (PCK) using a threshold of " + }, + { + "bbox": [ + 55, + 367, + 296, + 559 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 367, + 296, + 559 + ], + "type": "text", + "content": ". To evaluate computational costs, we report Multiply-Add Cumulation (MACs) and the average inference time (ms) following [2] on one RTX 3090 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 188, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 188, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 188, + 577 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "The input image is resized to maintain its original aspect ratio, with a base resolution of " + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\max(H, W) = 644" + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\max(H_{\\mathrm{hr}}, W_{\\mathrm{hr}}) = 1288" + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": " for high-resolution when needed. We denote this mixed-resolution setting as \"644*\". The Transformer layers in our encoder are initialized using pretrained DINOv2 [32] with a patch size of " + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "inline_equation", + "content": "P = 14" + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": ". We use ViT-B architecture as ViT-L significantly increases inference time (42.0 ms vs. 74.2 ms). The Transformer used in the encoder has " + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{lr}} = N_{\\mathrm{hr}} = 3" + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sa}} = 9" + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": " layers, respectively. Our decoder consists of 6 Transformer layers and 50 human queries. Training is conducted on 8 RTX 3090 GPUs with a" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 275, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 313, + 211, + 555, + 267 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 267 + ], + "type": "text", + "content": "Table 2. Comparison with SOTA methods on 3DPW [49] and MuPoTS [29] test sets. “Res.” represents the input resolution. “Crop” refers to cropping individuals for single-person mesh estimation, i.e., a multi-stage approach. “644*” means a base resolution of 644, with adaptive scaling up to 1288." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 275, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 319, + 275, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 319, + 275, + 553, + 392 + ], + "type": "table", + "html": "
MethodRes.3DPWMuPoTS (PCK)
PA-MPJPE ↓MVE ↓All ↑Matched ↑
CRMH [14]832--69.172.2
3DCrowdNet [7]Crop51.598.372.773.3
ROMP [45]51247.393.469.972.2
BEV [46]51246.992.370.275.2
PSVT [36]51245.784.9--
Multi-HMR [2]89641.775.985.089.3
Ours644*41.673.789.090.1
", + "image_path": "faa4a6ea6f467b2c5c6c9a10fa540b838ad772b6fc04de9859b470832268d6b3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 406, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 430 + ], + "type": "text", + "content": "total batch size of 40 for 60 epochs with an initial learning rate of " + }, + { + "bbox": [ + 313, + 406, + 555, + 430 + ], + "type": "inline_equation", + "content": "4e - 5" + }, + { + "bbox": [ + 313, + 406, + 555, + 430 + ], + "type": "text", + "content": ". Please refer to Sec. A.2 for more details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 442, + 504, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 504, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 504, + 454 + ], + "type": "text", + "content": "4.3. Comparison to the State-of-the-arts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 676 + ], + "type": "text", + "content": "Performance and Efficiency. We conduct evaluations on the AGORA [33], 3DPW [49], MuPoTS [28], and CMU Panoptic [15] datasets. Tab. 1 shows the results of our method compared to SOTA methods on the AGORA [33] test set leaderboard. Our method, using a mixed resolution input with a base of 644, achieves performance on par with the top SOTA methods [2, 44], despite their use of higher resolutions and larger models (e.g., Multi-HMR use ViT-L [32] model). These methods, however, result in significantly higher computational complexity, including much higher MACs and slower runtime, compromising real-time performance. Besides, these methods [2, 44] estimates SMPL-X [34], from which SMPL [24] mesh is derived. In contrast, our method maintains real-time computational efficiency while improving upon real-time methods ROMP [45], by nearly " + }, + { + "bbox": [ + 313, + 460, + 556, + 676 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 676 + ], + "type": "text", + "content": ". This establishes our approach as the leading real-time model, setting a new standard by achieving an unparalleled balance of performance and efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Tab. 2 presents the results of our method on the 3DPW dataset. Following Multi-HMR [2], we finetune on the 3DPW training set. We further evaluate the generaliza" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16801" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 71, + 518, + 223 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 518, + 223 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 518, + 223 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 518, + 223 + ], + "type": "image", + "image_path": "db5ac48be8e1bdec5fe2071f768509f4940c5e1edd4a932afaf6af9e6f3858ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 225, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 225, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 555, + 247 + ], + "type": "text", + "content": "Figure 4. Comparison with SOTA methods [2, 44-46] on in-the-wild images from the Internet. Red dashed circles highlight areas with incorrect estimations. The third case is left blank due to the small scale of individuals. Please zoom in for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 290, + 296, + 378 + ], + "blocks": [ + { + "bbox": [ + 55, + 259, + 296, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 282 + ], + "type": "text", + "content": "Table 3. Comparison with SOTA methods on CMU Panoptic [15] test set. We report MPJPE for four activities and the average." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 290, + 296, + 378 + ], + "lines": [ + { + "bbox": [ + 60, + 290, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 60, + 290, + 296, + 378 + ], + "type": "table", + "html": "
MethodHaggling ↓Mafia ↓Ultimatum ↓Pizza ↓Avg. ↓
CRMH [14]129.6133.5153.0156.7143.2
3DCrowdNet [7]109.6135.9129.8135.6127.6
ROMP [45]110.8122.8141.6137.6128.2
BEV [46]90.7103.7113.1125.2109.5
PSVT [36]88.797.9115.2121.2105.7
Ours67.978.595.894.684.2
", + "image_path": "765b805e0f2a4fe287b4f5fb6e4862107dd5ead812f32786ec90ac94f1914891.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 385, + 282, + 464 + ], + "blocks": [ + { + "bbox": [ + 70, + 385, + 282, + 464 + ], + "lines": [ + { + "bbox": [ + 70, + 385, + 282, + 464 + ], + "spans": [ + { + "bbox": [ + 70, + 385, + 282, + 464 + ], + "type": "image", + "image_path": "6daf1a86b3c79a0f77d78ca5374e236ae0576ce473de792a0539599ecc73c5ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 467, + 296, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 500 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of different resolutions for our baseline. Resolution for the baseline: (a) 518, (b) 1288. Red dashed circles highlight differences; zoom in for details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 296, + 601 + ], + "type": "text", + "content": "tion performance of our method on the MuPoTS and CMU Panoptic datasets, as shown in Tab. 2 and Tab. 3. For Panoptic dataset, we follow the common evaluation protocol [36, 45, 46] and exclude Multi-HMR [2] due to potential inconsistencies in joint format. Our method consistently achieves the best results across all datasets, demonstrating exceptional generalization capability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 296, + 715 + ], + "type": "text", + "content": "Qualitative Results. Fig. 4 presents a visual comparison between our method and existing SOTA approaches. The images are sourced from the Internet. Our method provides the most accurate overall estimations, especially in the third case, where individuals are far from the camera, resulting in very small scales. Despite this challenge, our method remains highly robust, delivering precise estimations. In contrast, ROMP [45] and BEV [46] fail to detect all individuals, while AiOS [44] and Multi-HMR [2], despite using" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 325, + 290, + 545, + 354 + ], + "blocks": [ + { + "bbox": [ + 313, + 259, + 554, + 282 + ], + "lines": [ + { + "bbox": [ + 313, + 259, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 554, + 282 + ], + "type": "text", + "content": "Table 4. Ablation studies on BEDLAM [3] validation set. We report MVE for different scale ranges and the average (Avg.)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 290, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 325, + 290, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 325, + 290, + 545, + 354 + ], + "type": "table", + "html": "
Ablation0-20%20-40%40-60%60-80%80%+Avg.
(a) Drop all59.955.161.566.670.857.2
(b) No pooling60.354.557.059.864.156.1
(c) Pooling×260.754.557.562.166.556.3
(d) Ours60.054.457.260.362.756.0
", + "image_path": "82ef2f4f24d2f7c8de3c274c4317616c5599a21cceda73a161e811d17654f3d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 555, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 377 + ], + "type": "text", + "content": "higher-resolution models, show significant estimation errors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 408, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 408, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 408, + 399 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 405, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 441 + ], + "type": "text", + "content": "We conduct ablation studies using only the AGORA [33] and sampled BEDLAM [3] datasets to validate model designs, and report results on their validation sets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": "Impact of resolution. Fig. 2 shows baseline results with different input resolutions on the AGORA [33] validation set. As resolution increases, the error across individuals in different scale ranges consistently decreases, with the most significant drop (35mm) in the smallest scale range. In contrast, the improvement in the large-scale range is less pronounced. However, the performance gain comes at the cost of a significant drop in computational efficiency, reducing the frame rate from real-time (32 FPS) to 5 FPS. Fig. 5 shows an example, clearly illustrating that higher resolution greatly improves estimation for small-scale individuals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Background tokens " + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": ". We investigate the impact of background tokens on model performance and report the MVE across different scale ranges on the BEDLAM validation set in Tab. 4. In (a), we discard all background tokens and observe stable performance at the smallest scale, but a decline as the scale increases, indicating that background context becomes more important for larger scales. In (b), skipping pooling for background tokens yields nearly the same performance but introduces higher computational cost. In (c), applying pooling twice with the same grouping strategy results in a performance drop for large-scale instances, highlighting" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16802" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 523, + 252 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 523, + 252 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 523, + 252 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 523, + 252 + ], + "type": "image", + "image_path": "bea354c11784882fb0b6c14e84c5533b86b14f97854ccd6e4bfcacaab7e9511f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 301 + ], + "type": "text", + "content": "Figure 6. Visualization of our predicted patch-level scale maps S and scale-adaptive tokens " + }, + { + "bbox": [ + 55, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{SA}}" + }, + { + "bbox": [ + 55, + 255, + 555, + 301 + ], + "type": "text", + "content": ". We show the predicted scale maps, scale-adaptive tokens, and the estimated meshes overlaid on the image. In the scale map, colors represent scale values, with uncolored areas indicating background patches. A gradient from blue (large-scale individuals) to rose (small-scale individuals) illustrates the scale distribution. We visualize scale-adaptive tokens as patches of different sizes and colors on the image, with low-resolution patches blurred." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 87, + 304, + 523, + 505 + ], + "blocks": [ + { + "bbox": [ + 87, + 304, + 523, + 505 + ], + "lines": [ + { + "bbox": [ + 87, + 304, + 523, + 505 + ], + "spans": [ + { + "bbox": [ + 87, + 304, + 523, + 505 + ], + "type": "image", + "image_path": "87bbf41706ae5c72aa2656c33dc729d41efd5efc2787fc1094be52253b87cdde.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 507, + 555, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 507, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 555, + 531 + ], + "type": "text", + "content": "Figure 7. Qualitative results of our method. We visualize four cases from AGORA [33], BEDLAM [3], CMU Panoptic [15] and 3DPW [49] datasets, displaying the input, the estimated mesh overlay, and an elevated view (top to bottom)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 543, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 616 + ], + "type": "text", + "content": "the significance of background context for estimation. We hypothesize that large-scale people face more truncation, relying more on global context which requires further exploration. We provide additional efficiency comparisons in Sec. B, showing that pooling once achieves the best balance between performance and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 167, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 167, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 167, + 635 + ], + "type": "text", + "content": "4.5. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "Fig. 6 visualize the predicted patch-level scale map and scale-adaptive tokens by our model, enabling dynamic adjustment of tokens for accurate estimations. Fig. 7 presents additional visualizations of predicted meshes, demonstrating the effectiveness of our method in various scenarios. Please refer to Sec. B for more qualitative results and discussion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 542, + 388, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 542, + 388, + 554 + ], + "spans": [ + { + "bbox": [ + 314, + 542, + 388, + 554 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 561, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 556, + 706 + ], + "type": "text", + "content": "We present a novel one-stage framework for real-time multiperson 3D mesh estimation from an RGB image. By introducing scale-adaptive tokens that dynamically adjust based on the relative scale of individuals, our method effectively balances efficiency and accuracy, retaining the benefits of high-resolution processing while achieving real-time performance. Experiments demonstrate that our method is the best real-time model, offering competitive performance and superior generalization compared to SOTAs. Our limitations include the lack of height-aware scale estimation, which may cause depth errors, and the current focus on body-only mesh estimation, extendable to whole-body in the future." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16803" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work was supported by 2022ZD0114904 and NSFC-6247070125. We thank Yiyang Wang for the valuable discussions during the early stage of this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 200 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In CVPR, pages 3686-3693, 2014. 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 202, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 296, + 245 + ], + "type": "text", + "content": "[2] Fabien Baradel, Matthieu Armando, Salma Galaaoui, Romain Brégier, Philippe Weinzaepfel, Grégory Rogez, and Thomas Lucas. Multi-hmr: Multi-person whole-body human mesh recovery in a single shot. In ECCV, 2024. 2, 3, 4, 6, 7, 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 289 + ], + "type": "text", + "content": "[3] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737, 2023. 6, 7, 8, 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 323 + ], + "type": "text", + "content": "[4] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In ICLR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 324, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 296, + 367 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 3, 4, 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 402 + ], + "type": "text", + "content": "[6] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In ECCV, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 445 + ], + "type": "text", + "content": "[7] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3d human mesh from in-the-wild crowded scenes. In CVPR, pages 1475-1484, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 448, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 296, + 512 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "text", + "content": "[9] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. Tore: Token reduction for efficient human mesh recovery with transformer. In ICCV, pages 15143-15155, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "text", + "content": "[10] Jose M Facil, Benjamin Ummenhofer, Huizhong Zhou, Luis Montesano, Thomas Brox, and Javier Civera. Cam-convs: Camera-aware multi-scale convolutions for single-view depth. In CVPR, pages 11826-11835, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 656 + ], + "type": "text", + "content": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 3, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[13] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predic" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "tive methods for 3d human sensing in natural environments. IEEE TPAMI, 36(7):1325-1339, 2013. 6, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[14] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In CVPR, pages 5579-5588, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "type": "text", + "content": "[15] Hanbyul Joo, Hao Liu, Lei Tan, Lin Gui, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social motion capture. In ICCV, pages 3334-3342, 2015. 6, 7, 8, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[16] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "type": "text", + "content": "[17] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In CVPR, pages 10863-10872, 2019. 6, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 284, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 316 + ], + "type": "text", + "content": "[18] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, pages 1954-1963, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 317, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 555, + 338 + ], + "type": "text", + "content": "[19] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, pages 12939-12948, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 339, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 555, + 381 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 383, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 414 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "type": "text", + "content": "[22] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, et al. Group pose: A simple baseline for end-to-end multi-person pose estimation. In ICCV, pages 15029-15038, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 472, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 555, + 514 + ], + "type": "text", + "content": "[23] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 3, 5, 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 515, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 555, + 557 + ], + "type": "text", + "content": "[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM TOG, 34(6):1-16, 2015. 2, 3, 6, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 559, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 580 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 582, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 613 + ], + "type": "text", + "content": "[26] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual markers. In CVPR, pages 534-543, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 616, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 616, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 616, + 555, + 667 + ], + "type": "text", + "content": "[27] Xiaoxuan Ma, Jiajun Su, Yuan Xu, Wentao Zhu, Chunyu Wang, and Yizhou Wang. Vmarker-pro: Probabilistic 3d human mesh estimation from virtual markers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[28] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 3DV, pages 506-516, 2017. 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16804" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[29] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 3DV, pages 120-130, 2018. 6, 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 296, + 163 + ], + "type": "text", + "content": "[30] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 198 + ], + "type": "text", + "content": "[31] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Neuralnot: Neural annotator for 3d human mesh training sets. In CVPR, pages 2299-2307, 2022. 6, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 200, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 296, + 255 + ], + "type": "text", + "content": "[32] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 256, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 300 + ], + "type": "text", + "content": "[33] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In CVPR, pages 13468-13478, 2021. 1, 2, 6, 7, 8, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 302, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 296, + 356 + ], + "type": "text", + "content": "[34] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 3, 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 358, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 358, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 296, + 403 + ], + "type": "text", + "content": "[35] Zhongwei Qiu, Qiansheng Yang, Jian Wang, and Dongmei Fu. Dynamic graph reasoning for multi-person 3d pose estimation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3521-3529, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 296, + 460 + ], + "type": "text", + "content": "[36] Zhongwei Qiu, Qiansheng Yang, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Chang Xu, Dongmei Fu, and Jingdong Wang. Psvt: End-to-end multi-person 3d pose and shape estimation with progressive video transformers. In CVPR, pages 21254-21263, 2023. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 462, + 296, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 296, + 505 + ], + "type": "text", + "content": "[37] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 507, + 296, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 507, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 507, + 296, + 540 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "type": "text", + "content": "[39] Cedric Renggli, André Susano Pinto, Neil Houlsby, Basil Mustafa, Joan Puigcerver, and Carlos Riquelme. Learning to merge tokens in vision transformers. arXiv preprint arXiv:2202.12015, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 588, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 296, + 632 + ], + "type": "text", + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 667 + ], + "type": "text", + "content": "[41] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In CVPRW, pages 4613-4622, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 106 + ], + "type": "text", + "content": "[43] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estimation with transformers. In CVPR, pages 11069-11078, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 162 + ], + "type": "text", + "content": "[44] Qingping Sun, Yanjun Wang, Ailing Zeng, Wanqi Yin, Chen Wei, Wenjia Wang, Haiyi Mei, Chi-Sing Leung, Ziwei Liu, Lei Yang, et al. Aios: All-in-one-stage expressive human pose and shape estimation. In CVPR, pages 1834-1843, 2024, 2, 3, 4, 6, 7, 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 555, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 198 + ], + "type": "text", + "content": "[45] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In ICCV, pages 11179-11188, 2021. 3, 6, 7, 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 199, + 555, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 555, + 242 + ], + "type": "text", + "content": "[46] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In CVPR, pages 13243-13252, 2022. 2, 3, 6, 7, 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 244, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 553, + 266 + ], + "type": "text", + "content": "[47] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 268, + 539, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 539, + 279 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 539, + 279 + ], + "type": "text", + "content": "[48] A Vaswani. Attention is all you need. NeurIPS, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 281, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 553, + 325 + ], + "type": "text", + "content": "[49] Timo von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In ECCV, pages 601-617, 2018. 1, 6, 8, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "text", + "content": "[50] Yuan Xu, Xiaoxuan Ma, Jiajun Su, Wentao Zhu, Yu Qiao, and Yizhou Wang. Scorehypo: Probabilistic human mesh estimation with hypothesis scoring. In CVPR, pages 979–989, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 405 + ], + "type": "text", + "content": "[51] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. In ICLR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 407, + 555, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 555, + 450 + ], + "type": "text", + "content": "[52] Sen Yang, Wen Heng, Gang Liu, Guozhong Luo, Wankou Yang, and YU Gang. Capturing the motion of every joint: 3d human pose and shape estimation with independent tokens. In ICLR, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "type": "text", + "content": "[53] Hang Ye, Wentao Zhu, Chunyu Wang, Rujie Wu, and Yizhou Wang. Faster voxelpose: Real-time 3d human pose estimation by orthographic projection. In ECCV, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 487, + 555, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 555, + 529 + ], + "type": "text", + "content": "[54] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In CVPR, pages 10809-10818, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 555, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 576 + ], + "type": "text", + "content": "[55] Andrei Zanfir, Elisabeta Marinoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In CVPR, pages 2148-2157, 2018. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 621 + ], + "type": "text", + "content": "[56] Andrei Zanfir, Elisabeta Marinoiu, Mihai Zanfir, Alin-Ionut Popa, and Cristian Sminchisescu. Deep network for the integrated 3d sensing of multiple people in natural images. In NeurIPS, 2018. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[57] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In ICCV, pages 11446-11456, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[58] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2023. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16805" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 195 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[59] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In CVPR, pages 15085-15099, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[60] Wentao Zhu, Jason Qin, Yuke Lou, Hang Ye, Xiaoxuan Ma, Hai Ci, and Yizhou Wang. Social motion prediction with cognitive hierarchies. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[61] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlun Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang: Human motion generation: A survey. IEEE TPAMI, 46(4): 2430-2449, 2024. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16806" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_content_list.json b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e5be6d0ed36f942408e7e77fcb55e0f596b5e487 --- /dev/null +++ b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_content_list.json @@ -0,0 +1,1438 @@ +[ + { + "type": "text", + "text": "SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nick Nikzad, Yi Liao, Yongsheng Gao, Jun Zhou", + "bbox": [ + 302, + 203, + 692, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Griffith University, Queensland, Australia", + "bbox": [ + 333, + 222, + 663, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{n.nikzaddehaji, y.liao, yongsheng.gao, jun.zhou}@griffith.edu.au", + "bbox": [ + 209, + 241, + 784, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few years, vision transformers (ViTs) have consistently demonstrated remarkable performance across various visual recognition tasks. However, attempts to enhance their robustness have yielded limited success, mainly focusing on different training strategies, input patch augmentation, or network structural enhancements. These approaches often involve extensive training and fine-tuning, which are time-consuming and resource-intensive. To tackle these obstacles, we introduce a novel approach named Spatial Autocorrelation Token Analysis (SATA). By harnessing spatial relationships between token features, SATA enhances both the representational capacity and robustness of ViT models. This is achieved through the analysis and grouping of tokens according to their spatial autocorrelation scores prior to their input into the Feed-Forward Network (FFN) block of the self-attention mechanism. Importantly, SATA seamlessly integrates into existing pre-trained ViT baselines without requiring retraining or additional fine-tuning, while concurrently improving efficiency by reducing the computational load of the FFN units. Experimental results show that the baseline ViTs enhanced with SATA not only achieve a new state-of-the-art top-1 accuracy on ImageNet-1K image classification $(94.9\\%)$ but also establish new state-of-the-art performance across multiple robustness benchmarks, including ImageNet-A (top-1=63.6%), ImageNet-R (top-1=79.2%), and ImageNet-C (mCE=13.6%), all without requiring additional training or fine-tuning of baseline models.", + "bbox": [ + 91, + 325, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Availability: https://github.com/nick-nikzad/SATA", + "bbox": [ + 89, + 732, + 418, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, vision transformers (ViTs) have demonstrated exceptional performance across diverse computer vision applications [9, 18]. Drawing inspiration from the significant achievements of transformer architectures in natural language processing (NLP), ViTs divide an input image into a sequence of patches (tokens) and leverage self-attention", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "layers [48] to capture relationships between these tokens, ultimately generating rich representations for visual recognition tasks. While recent studies indicate that ViTs can exhibit greater robustness than Convolutional Networks (ConvNets), attributed to their self-attention mechanism [1, 2, 35, 37]. However, this hypothesis has been challenged. Liu et al. [30] demonstrated that a carefully constructed ConvNet can surpass ViTs in both generalization and robustness. Furthermore, while techniques such as patch augmentation [32, 38], contrastive learning strategies [16, 38], and network adjustments [32, 57] have shown promise in enhancing ViT performance and robustness, two primary limitations remain: 1) these methods require extensive retraining or fine-tuning on large datasets (e.g., ImageNet-1K, ImageNet-21K), a labour-intensive and resource-demanding process, especially for large-scale ViT architectures; and 2) the attention maps generated by the self-attention mechanism are highly sensitive to noise, which degrades robustness when inputs are corrupted.", + "bbox": [ + 511, + 292, + 906, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Nikzad et al. [36] showed the existence of spatial correlation among feature maps in convolutional neural networks (CNNs). Moreover, they observed a decrease in spatial autocorrelation among feature maps through deeper network layers, suggesting that final features exhibit reduced spatial dependency. Motivated by these findings, in this work, we first investigate spatial autocorrelation within Vision Transformer (ViT) architectures and its implications for their performance and robustness. Then, we present a novel approach named \"Spatial Autocorrelation Token Analysis\" (SATA) to tackle the identified shortcomings in the current efforts to enhance ViT robustness.", + "bbox": [ + 511, + 566, + 908, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our analysis shows that spatial autocorrelation among visual patches (tokens) diminishes through ViT networks, similar to CNNs [36], aligning with findings that spatially uncorrelated features enhance recognition performance [25]. SATA integrates seamlessly into pre-trained ViT models without retraining, boosting robustness and efficiency by reducing the computational load on FFN units. These enhancements in feature aggregation and resilience to corruption account for SATA's exceptional performance from two perspectives:", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9730", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d5fe590560b1ece0e7de5a5de5ca3f7497361aed0fddfc7abf79c14262a6cd2.jpg", + "image_caption": [ + "Figure 1. Visual comparison of class token attention maps and spatial autocorrelation score maps across three layers of Deit-Base/16 pre-trained on ImageNet-1K. The clean image is sourced from the 'goldfinch' class of the ImageNet-1K dataset, and its corresponding corrupted version, with maximum severity (5), is sourced from the ImageNet-C [20] dataset. $\\delta$ represents the cosine similarity between either the attention maps or the spatial autocorrelation score maps of a corrupted image and its corresponding clean version at each block." + ], + "image_footnote": [], + "bbox": [ + 233, + 89, + 754, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Recognition performance: Our study shows that in the later layers of ViT networks, patches with extremely high or low spatial autocorrelation scores in non-informative regions can impede recognition performance and compromise the network's robustness against corrupted inputs. Our proposed SATA method adopts a unique splitting and grouping algorithm to address this issue based on tokens' spatial relation scores in the later layers. This approach prevents the input of unnecessary tokens into the FFN block of the self-attention mechanism, resulting in richer feature aggregation. Unlike token pruning and merging techniques, SATA concatenates bypassed tokens (Residual tokens) with the output of the FFN module, restoring the original number of tokens and preventing information loss for later blocks of the ViT.", + "bbox": [ + 89, + 354, + 485, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Robustness performance: As illustrated in Figure 1, our study shows that in the layers of ViT networks, the spatial autocorrelation scores of patches are more robust against different corruptions than their attention maps. Specifically, while the cosine similarity between clean and corrupted attention maps drops significantly in the blocks of the Vision Transformer, the similarity for spatial autocorrelation scores remains consistently high. This highlights that our spatial autocorrelation analysis provides a more stable and reliable schema to achieve corruption-resistant feature representation throughout the network, offering strong robustness against various types of corruption.", + "bbox": [ + 89, + 582, + 483, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments conducted on ImageNet-1K image classification and various robustness evaluation benchmarks demonstrate the effectiveness of the proposed spatial autocorrelation paradigm in significantly improving the robustness and accuracy performance of Vision Transformers (ViTs). These findings establish a new state-of-the-art performance level, achieving a top-1 accuracy of $94.9\\%$ on ImageNet-1K image classification, as well as impressive results across multiple robustness benchmarks, including ImageNet-A [8] (top", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$1 = 63.6\\%$ ), ImageNet-R [22] (top-1=79.2%), and ImageNet-C [20] (mCE=13.6%), without requiring additional expensive fine-tuning or training. Furthermore, in-depth investigations are conducted to thoroughly explore the characteristics of the proposed Spatial Autocorrelation Token Analysis (SATA).", + "bbox": [ + 511, + 354, + 908, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 460, + 663, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision Transformer", + "text_level": 1, + "bbox": [ + 511, + 484, + 700, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the introduction of Vision Transformers (ViTs), they have achieved remarkable success in various computer vision tasks [9, 18]. Most improvements to date have focused on enhancing either the accuracy or the efficiency of ViTs. Numerous ViT variants have been proposed to boost their performance [18]. Through dedicated data augmentation [46] and advanced self-attention structures [13, 53], ViTs have demonstrated competitive or superior performance compared to convolutional neural networks (CNNs). Hybrid models like CvT [51] introduce intrinsic inductive bias into the ViT architecture by adding additional convolutional layers before the multi-head self-attention (MHSA) modules. CeiT [54] extracts low-level features through the Image-to-Token (I2T) module and enhances locality by replacing the standard feedforward network with the locally enhanced feed-forward (LeFF) layer. To enable ViTs to learn multi-scale features, CrossViT [5] employs a dual-branch transformer that combines different sizes of image patches to produce stronger image features. ViTAE [52] incorporates multi-scale context by designing reduction cells (RC) and normal cells (NC).", + "bbox": [ + 511, + 507, + 908, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To create efficient Vision Transformers, several recent works have focused on pruning [11, 33, 44] or combining [24, 28] tokens. ResT [56] introduces an efficient self-attention module using overlapping depth-wise convolutions, while T2T-ViT [55] employs a Tokens-to-Token (T2T) module for token aggregation. PiT [23] reduces spatial size with", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9731", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pooling layers, and Dynamic-ViT [40] dynamically prunes tokens during inference. CaiT [47] optimizes the ViT architecture with layer scaling and class-attention mechanisms. More recently, Bolya et al.[3] proposed a simple token merging technique that potentially does not require retraining.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Robustness of ViTs", + "text_level": 1, + "bbox": [ + 89, + 176, + 274, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While several Recent research has yielded mixed results on the robustness of Vision Transformers (ViTs) compared to Convolutional Neural Networks (CNNs). While some studies [2, 14, 35, 37] suggest ViTs are more robust against various perturbations and distribution shifts, Liu et al. [30] challenge this notion by demonstrating that a well-designed CNN can outperform ViTs in generalization and robustness.", + "bbox": [ + 89, + 199, + 483, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance ViT robustness, various methods have been proposed, including network structural adjustments, patch augmentation, and diverse training strategies [4, 14-16, 26, 32, 38, 57]. For instance, Robust Vision Transformer (RVT) [32] introduces a convolutional stem and token pooling to improve robustness, while Full Attention Net (FAN) [57] leverages an attentional channel processing design. RobustViT [4] downplays the influence of image backgrounds, and a method proposed in [14] uses temperature scaling to smooth attention weights.", + "bbox": [ + 89, + 306, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, Qin et al.[38] improve the robustness of ViTs by using images transformed with patch-based operations as negative augmentation. Li et al. [26] propose TORA-ViT, which consists of an accuracy adapter, a robustness adapter, and a gated fusion module. The accuracy adapter extracts predictive features, while the robustness adapter extracts robust features. These features are then combined by the gated fusion module. Reducing Sensitivity to Patch Corruptions (RSPC) [16] enhances the robustness of ViTs through a specialized training strategy. In [15], the Attention Diversification Loss (ADL) is introduced to encourage output tokens to aggregate information from a diverse set of input tokens. Recently, Shi et al. [43] introduced learnable tokens and employed pixel-focused attention to capture structured information between tokens. However, most of these approaches require extensive training or fine-tuning and often sacrifice performance for efficiency [3]. In contrast, our method can be applied to baseline vision transformers[9, 46] without requiring additional training and without any performance drop.", + "bbox": [ + 89, + 458, + 483, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 776, + 230, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vision Transformers: Multi-head Self Attention (MHSA)", + "text_level": 1, + "bbox": [ + 89, + 801, + 482, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A standard ViT [9] partitions an input image into $N$ patches (tokens). These patches are then transformed to generate a token embedding tensor $\\mathbf{X} \\in \\mathbb{R}^{N \\times d}$ . These tokens are then processed through a stack of transformer blocks, as", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "illustrated in Figure 2(a). ViTs leverage self-attention [48] to aggregate global information. Given the input token embedding tensor $\\mathbf{X} \\in \\mathbb{R}^{N \\times d}$ , self-attention applies linear transformations with parameters $W_{K}, W_{Q}$ , and $W_{V}$ to embed them into the key $K = W_{K}\\mathbf{X}$ , query $Q = W_{Q}\\mathbf{X}$ , and value $V = W_{V}\\mathbf{X}$ , respectively. Self-attention utilises $K$ and $Q$ to generate a pairwise attention map $\\mathbf{M}_{att} \\in \\mathbb{R}^{N \\times N}$ and then aggregates the token features using the attention map $\\mathbf{M}_{att}$ as follows:", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {a t t} = \\operatorname {S o f t m a x} \\left(Q K ^ {t} / \\sqrt {d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 231, + 906, + 250 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S e l f - A t t e n t i o n} (Q, K, V) = \\mathbf {M} _ {\\text {a t t}} V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 252, + 906, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the symbol “ $t$ ” indicates the transpose of the matrix. To achieve rich feature hierarchies, the Transformer block employs multiple self-attention heads. Specifically, $h$ heads are stacked in parallel, resulting in an output of $N \\times h \\times d$ . These concatenated features are then processed by a feedforward network (FFN) for further transformation. Finally, the FFN output of $N \\times d$ serves as the final output of the Multi-Head Self-Attention (MHSA) block within the Transformer architecture.", + "bbox": [ + 511, + 273, + 908, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Geographical Spatial Auto-correlation", + "text_level": 1, + "bbox": [ + 511, + 416, + 844, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In geographical modelling, spatial autocorrelation plays a crucial role in assessing the spatial interdependence of entities based on their locations and values. Positive spatial autocorrelation indicates that neighbouring observations share similar values, while negative spatial autocorrelation suggests that nearby observations tend to have contrasting values. Typically, two types of measures are used: global measures, which provide an overall assessment of spatial autocorrelation across all data points, and local measures, which offer insights into the spatial autocorrelation of individual locations relative to their neighbourhoods. Moran's metric [6, 34] is commonly employed in geographical analysis to compute such measurements. In this study, we employ Moran's measurement for the first time, to the best of our knowledge, to investigate spatial dependency among vision transformers' tokens (patches).", + "bbox": [ + 511, + 439, + 908, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{X}$ be a set of $N$ observations (here, tokens) presented by embedded vectors $\\mathbf{x}_i\\in \\mathbb{R}^d$ , $\\mathbf{X} = [\\mathbf{x}_1,\\mathbf{x}_2,\\dots,\\mathbf{x}_N]$ , and an associated attribute, $\\mathbf{a} = [a_1,a_2,\\dots,a_N]$ , the local Moran's I metric can be defined as:", + "bbox": [ + 511, + 680, + 906, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {l} = \\left[ \\operatorname {d i a g} \\left(\\mathbf {z z} ^ {t} \\mathbf {W}\\right) \\right] _ {N \\times 1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 746, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{diag}(.)$ returns the diagonal elements of a matrix. Symbol \"t\" indicates the transpose of the matrix. $\\mathbf{W} = [w_{ij}]_{N\\times N}$ represents spatial weight matrix, in which $w_{ij}$ denotes the degree of closeness or the contiguous relationships between $\\mathbf{x}_i$ and $\\mathbf{x}_j$ and can be computed using a dot product similarity $(\\mathbf{x}_i.\\mathbf{x}_j^t i,j = 1,2,\\dots,N)$ . $\\mathbf{z}$ refers to normalised token-wise attribute values as:", + "bbox": [ + 511, + 768, + 908, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\frac {\\mathbf {a} - \\mu}{\\sigma}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 666, + 877, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9732", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b8c83390f656fadadb70704ce1ae9d2c4be6e5819ee21ebb34b18a77254b422.jpg", + "image_caption": [ + "(a)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1bf9a59b62008c79d08a6eab62fbce5ab9b6cb713ac66f86583394ac664b22c7.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 294, + 284, + 702, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ denote mean and standard deviation of a, respectively. The final local spatial autocorrelation descriptor, s, can be defined as normalised $I_{l}$ [36]:", + "bbox": [ + 88, + 483, + 485, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} = \\frac {\\boldsymbol {I} _ {l} - \\mu_ {\\boldsymbol {I} _ {l}}}{\\sigma_ {\\boldsymbol {I} _ {l}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 539, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{I_l}$ and $\\sigma_{I_l}$ indicate mean and standard deviation of $I_{l}$ , respectively. Following [36], given a token embedding tensor $\\mathbf{X} = [\\mathbf{x}_1,\\mathbf{x}_2,\\dots,\\mathbf{x}_N]\\in \\mathbb{R}^{N\\times d}$ , its token-wise global context attribute $\\mathbf{a} = [a_1,a_2,\\dots,a_N]\\in \\mathbb{R}^{N\\times 1}$ can be defined as:", + "bbox": [ + 89, + 583, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} = \\left[ a _ {i} = \\frac {1}{d} \\sum_ {t = 1} ^ {d} \\mathbf {x} _ {i} (t) \\right] _ {N \\times 1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 667, + 483, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ denotes the spatial dimension of the tokens. $\\mathbf{x}_i(t)$ represents the $i$ -th token value at position $t$ . It's worth noting that more advanced strategies or application-specific criteria can be employed to derive the global contextual information descriptor (Eq.(6)). In this context, we adopt the same approach as [36] for the sake of simplicity.", + "bbox": [ + 89, + 723, + 483, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Spatial Autocorrelation Token Analysis", + "text_level": 1, + "bbox": [ + 89, + 829, + 444, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2(b) illustrates the overall architecture of the proposed spatial autocorrelation token analysis module, situated between attention and FFN units within a standard ViT block,", + "bbox": [ + 89, + 854, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/026e7ef143c9f562ecf37face211d3e544281d1f7ebd4a064be62ea1d499ac3b.jpg", + "image_caption": [ + "Figure 2. (a) Comparison between conventional ViT block and the augmented ViT with SATA (b) Overall architecture of the proposed SATA module. (c) Tokens Merging.", + "Figure 3. Plotting the variations of $\\mu_{\\mathrm{s}}$ , $|\\hat{\\mathbf{s}}|$ , and the lower and upper bounds across different blocks of the ViT." + ], + "image_footnote": [], + "bbox": [ + 516, + 482, + 903, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as depicted in Figure 2(a). To initiate our spatial autocorrelation token analysis, we begin with the observation of the spatial autocorrelation scores, s (Eq. (5)), for ViT's token embedding tensors over different blocks. As transformers inherently capture pairwise closeness relationships between tokens by computing the attention map $\\mathbf{M}_{att}$ in Eq.(2), we can directly set $\\mathbf{W} = \\mathbf{M}_{att}$ in Eq.(3) to enhance both efficiency and effectiveness1.", + "bbox": [ + 511, + 602, + 906, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 illustrates the alterations in the mean $(\\mu_{\\mathrm{s}})$ and the absolute value of the median $(|\\hat{\\mathbf{s}}|)$ statistics of the spatial autocorrelation (s) across different blocks of Deit-Base/16 on the ImageNet-1K validation set. It can be seen that for the later layers, specifically starting from block six, tokens tend to exhibit lower $|\\hat{\\mathbf{s}}|$ values. Drawing from the aforementioned observation, we encapsulate the proposed analysis into two sequential steps to handle these tokens to improve the ViT's robustness and performance:", + "bbox": [ + 511, + 724, + 908, + 861 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The implementation details of the proposed SATA are provided in Supplementary material", + "bbox": [ + 511, + 875, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9733", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Token Splitting", + "text_level": 1, + "bbox": [ + 89, + 90, + 246, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above findings, the overall of the proposed SATA method is illustrated in Figure 2 (b). As shown in Figure 3, we limit token processing to the latter stages of the transformer. Specifically, tokens in layers from $\\gamma \\times B$ onward ( $\\gamma > 0$ , where $B$ represents the depth, or number of blocks, of the transformer) are partitioned into two sets, $\\mathbb{A}$ and $\\mathbb{B}$ , using the spatial autocorrelation scores s as follows:", + "bbox": [ + 89, + 112, + 485, + 219 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {A} = \\left\\{\\mathbf {x} _ {i}; s _ {i} < \\alpha \\left(\\mu_ {\\mathbf {s}} - | \\hat {\\mathbf {s}} |\\right) \\text {a n d} s _ {i} > \\alpha \\left(\\mu_ {\\mathbf {s}} + | \\hat {\\mathbf {s}} |\\right) \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 241, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {B} = \\left\\{\\mathbf {x} _ {j}; \\alpha \\left(\\mu_ {\\mathbf {s}} - | \\hat {\\mathbf {s}} |\\right) < = s _ {j} < = \\alpha \\left(\\mu_ {\\mathbf {s}} + | \\hat {\\mathbf {s}} |\\right) \\right\\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 260, + 483, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha (\\mu_{\\mathbf{s}} - |\\hat{\\mathbf{s}} |)$ and $\\alpha (\\mu_{\\mathbf{s}} + |\\hat{\\mathbf{s}} |)$ represent lower and upper bounds, respectively. $\\alpha$ denotes the controlling factor (parameters $\\alpha$ and $\\gamma$ choices are discussed in Section 5.6).", + "bbox": [ + 89, + 282, + 483, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Token Grouping", + "text_level": 1, + "bbox": [ + 89, + 335, + 254, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To manage tokens falling beyond the lower and upper bounds (i.e. $\\alpha (\\mu_{\\mathrm{s}}\\pm |\\hat{\\mathbf{s}} |))$ , inspired by the Bipartite Matching algorithm [3] we employ a unique Tokens Merging algorithm to efficiently match and merge similar tokens in set A. In particular, the proposed Tokens Merging algorithm can be summarized as follows (illustrated in Figure 2(c)):", + "bbox": [ + 89, + 358, + 483, + 449 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Partition set $\\mathbb{A}$ into two sets $\\mathbb{A}_1$ and $\\mathbb{A}_2$ of roughly equal size.", + "2. Draw one edge from each token in $\\mathbb{A}_1$ to its most similar token in $\\mathbb{A}_2$ .", + "3. Merged Tokens: Merge tokens by averaging their features if they are connected.", + "4. Residual Tokens: Tokens from set $\\mathbb{A}_1$ representing maximum similarity with tokens in set $\\mathbb{A}_2$ ." + ], + "bbox": [ + 89, + 450, + 483, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed Merging algorithm includes unconnected tokens from set $\\mathbb{A}_2$ in the merged tokens output, whereas the Bipartite Matching algorithm [3] disregards these tokens. This feature reduces information loss within the overall SATA schema. As depicted in Figure 2(b), the output tokens of the proposed SATA module comprise the concatenation of tokens with spatial scores within the range of lower and upper bounds (i.e., set $\\mathbb{B}$ ) and the Merged Tokens. This output is then fed into the FFN module, reducing unnecessary token processing in the FFN block and lowering the computational load and GFLOPs. Furthermore, Residual Tokens are concatenated with the FFN output to restore the original number of tokens, $N$ , forming the final output of the new ViT block, as shown on the right side of Figure 2(a).", + "bbox": [ + 89, + 570, + 485, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment Results & Analysis", + "text_level": 1, + "bbox": [ + 89, + 792, + 377, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 818, + 277, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details All experiments were conducted on an NVIDIA V100 GPU with a $224 \\times 224$ image resolution. The batch size was set to 256 for all experiments unless otherwise specified. We integrated the proposed SATA module", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "into pre-trained generic vision transformers [9, 46] (DeiTiny/16, DeiT-Small/16, DeiT-Base/16, and Vit-Base/16), resulting in three model sizes named SATA-T, SATA-S, SATA-B, and SATA-B*, respectively.", + "bbox": [ + 511, + 90, + 908, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Benchmarks We employ the ImageNet2012-1K [7] dataset for standard performance evaluation. For robustness assessment, we evaluate the proposed SATA in three dimensions: 1) Adversarial Robustness: Testing is conducted on adversarial examples generated by white-box attack algorithms FGSM [12] and PGD [31] using the ImageNet-1K validation set. ImageNet-A [8] (IN-A) includes the ImageNet objects in unusual contexts or orientations and is utilized to assess model performance against natural adversarial examples. 2) Common Corruption Robustness: We use ImageNet-C [20] (IN-C), which includes 15 types of algorithmically generated corruptions, each with five levels of severity. 3) Out-of-Distribution Robustness: Evaluation is performed on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK). Both datasets feature images with naturally occurring distribution shifts. ImageNet-R [22] (IN-R) contains abstract or rendered versions of the objects. ImageNet-Sketch [49] consists solely of sketch images, serving to test classification capability when texture or colour information is absent.", + "bbox": [ + 511, + 166, + 908, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Standard Performance Evaluation", + "text_level": 1, + "bbox": [ + 511, + 479, + 815, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For standard performance evaluation, we compare our method with several state-of-the-art classification models, including Transformer-based models and representative CNN-based models, as shown in Table 1. Our proposed SATA significantly outperforms all other architectures, including both CNN-based and ViT-based models. Specifically, ViT models enhanced with SATA achieve new state-of-the-art top-1 accuracy of $86.5\\%$ , $89.3\\%$ , and $93.9\\%$ for the tiny, small, and base versions, respectively, all without requiring additional training, input augmentation, or fine-tuning. Notably, integrating the proposed SATA into pre-trained ViT-Base/16 [9] (SATA- $\\mathbf{B}^*$ ) results in an additional $1.0\\%$ improvement. Furthermore, comparing the computation cost (GFLOPs) of the baseline DeiTs and SATA models demonstrates that the proposed spatial autocorrelation token analysis method also improves efficiency.", + "bbox": [ + 511, + 503, + 908, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Adversarial Robustness Evaluation", + "text_level": 1, + "bbox": [ + 511, + 757, + 820, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluating white-box attack adversarial robustness, we follow [32] and adopt the single-step attack algorithm FGSM [12] and the multi-step attack algorithm PGD [31] (with 5 steps and a step size of 0.5). Both attackers perturb the input image with a maximum magnitude of $\\epsilon = 1$ . As shown in Table 1, adversarial robustness appears unrelated to standard performance. For instance, models like Swin [29], PVT [50], and TNT-S [17] achieve higher standard accuracy", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9734", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0192da7a3922d3bb99ea14f90fc61c5aed098dbe9f6693ca39b65c9fe397c4f6.jpg", + "table_caption": [ + "Table 1. Performance of SATA and several state-of-the-art (SOTA) CNNs and ViTs models on ImageNet and six robustness benchmarks: We report the mean corruption error (mCE) for ImageNet-C [20], where lower mCE values indicate higher model robustness. Our SATA models consistently outperform other counterparts in standard performance and enhance robustness across various model sizes compared to the baseline, all without requiring additional training or fine-tuning. SATA-B* refers to the integration of the proposed SATA module into the pre-trained vanilla ViT-Base/16 model [9]." + ], + "table_footnote": [], + "table_body": "
GroupModelFLOPs (G)Params (M)ImageNet-1K Top-1Top-5FGSMPGDRobustness Benchmarks
IN-C (mCE↓)IN-AIN-RIN-SK
CNNResNet50 [19]4.125.676.186.012.20.976.70.036.124.1
RegNetY-4GF[39]4.020.679.294.715.42.468.78.938.825.9
EfficientNet-B4[45]4.419.383.096.344.618.571.126.347.134.1
DeepAugment[21]4.125.675.892.727.19.553.63.946.732.6
ANT[41]4.125.676.193.017.83.163.01.139.026.3
Debiased CNN[27]4.125.676.993.420.45.567.53.540.828.4
ConvNeXt-B[30]15.48983.8---46.836.751.338.2
ViT-TinyDeiT-Ti[46]1.35.772.291.122.36.271.17.332.620.2
ConvViT-Ti[10]1.45.773.391.824.77.568.48.935.222.4
PiT-Ti[23]0.74.972.991.320.45.169.16.234.621.6
PVT-Tiny[50]1.913.275.092.510.00.579.67.933.921.5
RVT-Ti [32]1.38.678.494.234.811.758.213.343.730.0
FAN-T-ViT [57]1.37.079.2---57.5-42.5-
RVT-Ti+RSPC [16]1.310.979.2---55.716.5--
SATA-T (ours)1.05.786.598.240.010.951.114.647.325.2
ViT-SmallDeiT-S[46]4.622.179.995.040.716.754.618.942.229.4
ConvViT-S[10]5.427.881.595.841.717.249.824.545.433.1
PiT-S[23]2.923.580.995.341.016.552.521.743.630.8
PVT-Small[50]3.824.579.995.026.63.166.918.040.127.2
Swin-T[29]4.528.381.295.533.77.362.021.641.329.1
TNT-S[17]5.223.881.595.733.24.253.124.743.831.6
T2T-ViT_t-14[55]6.121.581.795.940.911.753.223.945.032.5
RVT-S [32]4.722.181.795.751.326.050.124.146.935.0
FAN-S-ViT [57]5.328.082.9---47.729.150.4-
RVT-S+RSPC [16]4.723.382.2---48.427.9--
SATA-S (ours)3.922.189.399.157.418.033.830.559.539.2
ViT-BaseDeiT-B[46]17.686.682.095.746.421.348.527.444.932.4
ConvViT-B[10]17.786.582.095.746.421.348.527.444.932.4
PiT-B[23]12.573.882.495.749.323.748.233.943.732.3
PVT-Large[50]9.861.481.795.933.17.359.826.642.730.2
Swin-B[29]15.487.883.496.449.221.354.435.846.632.4
T2T-ViT_t-24[55]15.064.182.696.146.717.548.428.947.935.4
RVT-B [32]17.786.282.596.052.327.447.327.748.235.8
FAN-B-ViT [57]10.454.083.6---44.435.451.8-
RVT-B+RSPC [16]17.791.882.8---45.732.1--
TORA-ViT-B/16(λ=0.1) [26]26.0111.284.1-48.423.331.746.557.6-
TransNeXt [43]18.489.784.8---43.550.653.941.4
SATA-B (ours)15.986.693.999.763.920.228.763.570.049.8
SATA-B* (ours)15.986.694.999.865.628.313.663.679.257.9
", + "bbox": [ + 91, + 166, + 906, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "than DeiTs corresponding, but their adversarial robustness is significantly lower, consistent with findings from [32, 42]. Our proposed SATA model achieves superior performance against both FGSM [12] and PGD [31] attacks. Specifically, SATA-S, SATA-B, and SATA-B* show over a $20\\%$ improvement on FGSM [12] compared to previous ViT variants.", + "bbox": [ + 89, + 722, + 485, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding natural adversarial robustness, the proposed SATA-T demonstrates a comparable performance of $14.6\\%$ , which is on par with some current state-of-the-art methods like RVT-Ti [32] and RVT-Ti+RSPC [16], while being about half their size. However, for models of similar size (e.g.,", + "bbox": [ + 89, + 824, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ViT-Small and ViT-Base), the proposed SATAs outperform others by about $50\\%$ , indicating the effectiveness of SATA against natural adversarial attacks.", + "bbox": [ + 511, + 722, + 906, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Common Corruption Robustness Evaluation", + "text_level": 1, + "bbox": [ + 511, + 785, + 890, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To measure model degradation on common image corruptions, we report the mean corruption error (mCE) on ImageNet-C [20] (IN-C) in Table 1. Our SATA method significantly reduces the mCE of DeiT-Ti [46] from $71.1\\%$ to $51.1\\%$ , achieving the lowest mCE among vision transformers within the ViT-Tiny group. For the other two larger", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9735", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff8967be74804351126cc4974e2ad94160563efd56dff610c314463ddc8787c5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 472, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/165bf6f2641482103d718545c0dc815d05d7541798f66d4dfef04bdebd18ab78.jpg", + "image_caption": [ + "(b)", + "Figure 4. (a) Ablation study on $\\gamma$ . (b) Ablation study on $\\alpha$ . We set $\\gamma$ and $\\alpha$ to 0.7 and 1.0, respectively, for all experiments throughout this paper. Ablation studies are conducted on the SATA-B model using the ImageNet-1K dataset. Dashed lines for both graphs represent the baseline Deit-Base/16 top-1 accuracy." + ], + "image_footnote": [], + "bbox": [ + 107, + 258, + 449, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ViT groups, the proposed SATA models achieve an mCE of approximately $28\\%$ , improving by around 20 points over all other ViT or CNN-based methods on the leaderboard, thereby establishing a new state-of-the-art. This result also suggests that spatial autocorrelation management of visual tokens can successfully handle different types of image corruption.", + "bbox": [ + 89, + 529, + 483, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Out-of-Distribution Robustness Evaluation", + "text_level": 1, + "bbox": [ + 89, + 672, + 455, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the generalization ability of SATA on out-of-distribution data by reporting the top-1 accuracy on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK) in Table 1. The generic vision transformers [9, 46] enhanced by the proposed SATA consistently outperform other ViT models on ImageNet-R [22], achieving $47.3\\%$ , $57.2\\%$ , $70.0\\%$ , $79.9\\%$ in the ViT-Tiny, ViT-Small, and ViT-Base groups, respectively. Regarding ImageNet-Sketch [49] (IN-SK), SATA demonstrates superior performance compared to other models of similar size. These results imply that the spatial autocorrelation tokens analysis effectively captures feature distribution shifts, enhancing the model's out-of-distribution generalization capabilities.", + "bbox": [ + 89, + 704, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Ablation study", + "text_level": 1, + "bbox": [ + 513, + 90, + 663, + 106 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Token Splitting and Token Grouping We evaluate the role of token splitting based on the upper and lower bounds of spatial autocorrelation scores and token grouping modules. To this end, we examine five SATA configurations, as depicted in Table 2. As shown in Table 2, utilizing only token grouping yields a top-1 accuracy of $84.4\\%$ , improving upon the reference (DeiT-Base) accuracy by $2.4\\%$ . Including either lower or upper bounds significantly enhances accuracy by about $10\\%$ of the baseline top-1 accuracy, highlighting the effectiveness of the proposed spatial autocorrelation token splitting schema. The upper bound contributes slightly more, suggesting that tokens with extremely high spatial autocorrelation scores are more likely to be filtered by the splitting process. However, incorporating upper or lower bounds also increases computations and token counts, reducing throughput despite the substantial accuracy gains. Additionally, applying bounds does not guarantee an equal token distribution across bands, as this depends on the data distribution (shown in Figure 2 of the supplementary material). Finally, adding token grouping provides a further $1.6\\%$ improvement over the splitting process alone.", + "bbox": [ + 511, + 112, + 906, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding throughput, two main components of the proposed SATA framework (i.e. spatial autocorrelation-based token splitting and token grouping) primarily utilise parallelizable matrix multiplication and dot product operations. This design ensures computational efficiency with a negligible impact on inference time, as shown in Table 2.", + "bbox": [ + 511, + 431, + 908, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Threshold of starting block $(\\gamma)$ We also examine the effect of parameter $\\gamma$ , which controls at which transformer block the SATA module is applied. As Figure 4(a) shows, applying SATA from block $0.4 \\times B$ onwards significantly improves model efficiency while exceeding baseline ViT accuracy $(82\\%)$ . Applying SATA to earlier blocks $(\\gamma < 0.4)$ degrades accuracy, suggesting that high spatial correlation within token features and the importance of all tokens in early layers is beneficial. To achieve a good trade-off between accuracy and efficiency, we use $\\gamma = 0.7$ in all our experiments, unless stated otherwise. This results in a top-1 accuracy of $93.9\\%$ and GFLOPs of 15.9.", + "bbox": [ + 511, + 542, + 908, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lower/Upper bounds controlling factor $(\\alpha)$ We further assess the influence of $\\alpha$ , the factor controlling the lower and upper bounds in SATA. Figure 4(b) shows the performance of SATA-B on ImageNet-1K with $\\alpha$ values ranging from 0.5 to 2. $\\alpha$ determines the number of tokens passed to the FFN block and setting $\\alpha = 1$ yields optimal performance.", + "bbox": [ + 511, + 744, + 908, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.7. Visualisation and Discussion", + "text_level": 1, + "bbox": [ + 511, + 847, + 767, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although the effectiveness of the proposed SATA module has been empirically demonstrated, we conduct a deeper", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9736", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6332add24d47ee44db94d11044329fd89c2b153624cdfdf548c072f542544b0.jpg", + "table_caption": [ + "Table 2. Ablation study of the spatial autocorrelation module (token splitting) and token grouping. The symbols “√” and “X” indicate whether the corresponding element is employed with the configuration or not. † represents top-1 accuracy for the baseline Deit-Base/16 [46] model." + ], + "table_footnote": [], + "table_body": "
Spt. Auto. Correlation (Tokens Splitting)Tokens GroupingTop-1Throughput (ims/s)Latency (s)
Lower bound (μs - |s|)Upper bound (μs + |s|)
×××82.0†3030.0033
××✓84.43610.0027
×✓✓92.82970.0033
✓×✓92.52710.0035
✓✓×92.32680.0037
✓✓✓93.92760.0035
", + "bbox": [ + 104, + 141, + 885, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/765594f9b73860db62910a2eaae275672e63fde7965eea7a50fd8d7a29ed6ec3.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 102, + 297, + 455, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1149a4d3cc9f7cdad93af8672b24fe378cbcfb79d7c9e7abe1df6afbbe05ef25.jpg", + "image_caption": [ + "(b)", + "Figure 5. (a) Cosine similarity between the clean and corrupted versions of the class token attention map and spatial autocorrelation scores across different blocks of SATA-B. Results are averaged across various types of image corruptions and severity levels on ImageNet-C [20]. (b) Visualisation of token splitting for a pair of clean and noisy images. Notably, the selected tokens for each set are similar for both clean and corrupted inputs." + ], + "image_footnote": [], + "bbox": [ + 101, + 476, + 470, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "investigation to better understand its behaviour. To this end, we calculate the cosine similarity between the class token attention maps and spatial autocorrelation scores of clean and corrupted image pairs from the ImageNet-1K validation set and its corresponding ImageNet-C [20], respectively. We compute these similarities for various types of image corruptions and severity levels in ImageNet-C [20], and report the average values across different blocks of the proposed SATA-B in Figure 5(a). As shown in Figure 5(a), the cosine similarity between clean and corrupted attention maps", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "drops significantly in the later blocks of the transformer. In contrast, the similarity for spatial autocorrelation scores improves at the early stages and remains consistently high, averaging above 0.8. This highlights that the proposed method can provide a more stable and reliable feature representation throughout the network, offering strong robustness against various types of corruption.", + "bbox": [ + 511, + 297, + 906, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, Figure 5 provides a visualization of tokens (patches) are split into set $\\mathbb{A}$ and set $\\mathbb{B}$ for a pair of clean and noisy images according to the proposed SATA algorithm. Notably, the similarity between corresponding sets for clean and noisy inputs is evident, further highlighting the robustness of the proposed method2.", + "bbox": [ + 511, + 406, + 908, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 518, + 633, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SATA, a novel method designed to enhance the performance and robustness of vision transformers against various types of corruption. SATA utilises a straightforward yet powerful spatial autocorrelation scheme to exploit spatial inter-dependencies among token features, significantly improving representational capacity and efficiency while reducing computational costs. Our experimental results demonstrate that SATA-enhanced vision transformers consistently deliver stable and reliable feature representations, achieving state-of-the-art performance on ImageNet-1K classification and setting new benchmarks for robustness across multiple evaluations, all without requiring additional training or fine-tuning.", + "bbox": [ + 511, + 547, + 908, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work highlights SATA's transformative potential and opens several promising avenues for future research. Key directions include adapting SATA for window-based and hybrid ViT architectures to boost performance in tasks such as object detection and segmentation. Additionally, exploring SATA's application in other transformer-based domains, such as large language models (LLMs), could further extend its impact.", + "bbox": [ + 511, + 744, + 908, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2Additional visual comparisons are included in Supplementary material.", + "bbox": [ + 529, + 886, + 903, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9737", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Australian Research Council (ARC), Australia under Discovery Grant DP180100958 and Industrial Transformation Research Hub Grant IH180100002.", + "bbox": [ + 89, + 114, + 485, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? In NeurIPS, pages 26831-26843, 2021. 1", + "[2] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In ICCV, pages 10231-10241, 2021. 1, 3", + "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3, 5", + "[4] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. In NeurIPS, pages 33618-33632, 2022. 3", + "[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2", + "[6] Yanguang Chen. New approaches for calculating Moran's index of spatial autocorrelation. *PloS one*, 8(7), 2013. 3", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5", + "[8] Minjing Dong, Yanxi Li, Yunhe Wang, and Chang Xu. Adversarily robust neural architectures. arXiv preprint arXiv:2009.00902, 2020. 2, 5", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5, 6, 7", + "[10] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, pages 2286-2296. PMLR, 2021. 6", + "[11] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, pages 396-414. Springer, 2022. 2", + "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5, 6", + "[13] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, pages 12259-12269, 2021. 2", + "[14] Jindong Gu, Volker Tresp, and Yao Qin. Are vision transformers robust to patch perturbations? In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3" + ], + "bbox": [ + 93, + 213, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yong Guo, David Stutz, and Bernt Schiele. Robustifying token attention for vision transformers. In ICCV, pages 17557-17568, 2023. 3", + "[16] Yong Guo, David Stutz, and Bernt Schiele. Improving robustness of vision transformers by reducing sensitivity to patch corruptions. In CVPR, pages 4108-4118, 2023. 1, 3, 6", + "[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, pages 15908-15919, 2021. 5, 6", + "[18] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE TPAMI, 45(1):87-110, 2022. 1, 2", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 2, 5, 6, 8", + "[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6", + "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 2, 5, 7", + "[23] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, pages 11936-11945, 2021. 2, 6", + "[24] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, pages 620-640. Springer, 2022. 2", + "[25] Xuelong Li, Han Zhang, Rui Zhang, and Feiping Nie. Discriminative and uncorrelated feature selection with constrained spectral analysis in unsupervised learning. IEEE Transactions on Image Processing, 29:2139-2149, 2019. 1", + "[26] Yanxi Li and Chang Xu. Trade-off between robustness and accuracy of vision transformers. In CVPR, pages 7558-7568, 2023. 3, 6", + "[27] Yingwei Li, Qihang Yu, Mingxing Tan, Jieru Mei, Peng Tang, Wei Shen, Alan Yuille, and Cihang Xie. Shape-texture de-biased neural network training. In In Proceedings of the International Conference on Learning Representations, 2021. 6", + "[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2", + "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 6" + ], + "bbox": [ + 516, + 92, + 908, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9738", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 3, 6", + "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 5, 6", + "[32] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In CVPR, pages 12042-12051, 2022. 1, 3, 5, 6", + "[33] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 2", + "[34] Patrick AP Moran. The interpretation of statistical maps. Journal of the Royal Statistical Society. Series B (Methodological), 10(2):243-251, 1948. 3", + "[35] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, pages 23296-23308, 2021. 1, 3", + "[36] Nick Nikzad, Yongsheng Gao, and Jun Zhou. CSA-Net: Channel-wise Spatially Autocorrelated Attention Networks. arXiv preprint arXiv:2405.05755, 2024. 1, 4", + "[37] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In AAAI, pages 2071-2081, 2022. 1, 3", + "[38] Yao Qin, Chiyuan Zhang, Ting Chen, Balaji Lakshminarayanan, Alex Beutel, and Xuezhi Wang. Understanding and improving robustness of vision transformers through patch-based negative augmentation. In NeurIPS, pages 16276-16289, 2022. 1, 3", + "[39] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 6", + "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3", + "[41] Evgenia Rusak, Lukas Schott, Roland S Zimmermann, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge, and Wieland Brendel. A simple way to make neural networks robust against diverse image corruptions. In ECCV, pages 53-69. Springer, 2020. 6", + "[42] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 6", + "[43] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In CVPR, pages 17773-17783, 2024. 3, 6", + "[44] Zhuoran Song, Yihong Xu, Zhezhi He, Li Jiang, Naifeng Jing, and Xiaoyao Liang. Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570, 2022. 2", + "[45] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6", + "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + ], + "bbox": [ + 91, + 92, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6, 7, 8", + "[47] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021. 3", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 3", + "[49] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5, 7", + "[50] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 5, 6", + "[51] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, pages 22–31, 2021. 2", + "[52] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, pages 28522-28535, 2021. 2", + "[53] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021. 2", + "[54] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In ICCV, pages 579-588, 2021. 2", + "[55] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 2, 6", + "[56] Qinglong Zhang and Yu-Bin Yang. Rest: An efficient transformer for visual recognition. In NeurIPS, pages 15475-15485, 2021. 2", + "[57] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In ICML, pages 27378-27394. PMLR, 2022. 1, 3, 6" + ], + "bbox": [ + 516, + 92, + 906, + 672 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9739", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_model.json b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d07859a83be06fa854e37868ac7722ed0e463a65 --- /dev/null +++ b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.204, + 0.694, + 0.222 + ], + "angle": 0, + "content": "Nick Nikzad, Yi Liao, Yongsheng Gao, Jun Zhou" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.223, + 0.664, + 0.239 + ], + "angle": 0, + "content": "Griffith University, Queensland, Australia" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.242, + 0.785, + 0.257 + ], + "angle": 0, + "content": "{n.nikzaddehaji, y.liao, yongsheng.gao, jun.zhou}@griffith.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.326, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Over the past few years, vision transformers (ViTs) have consistently demonstrated remarkable performance across various visual recognition tasks. However, attempts to enhance their robustness have yielded limited success, mainly focusing on different training strategies, input patch augmentation, or network structural enhancements. These approaches often involve extensive training and fine-tuning, which are time-consuming and resource-intensive. To tackle these obstacles, we introduce a novel approach named Spatial Autocorrelation Token Analysis (SATA). By harnessing spatial relationships between token features, SATA enhances both the representational capacity and robustness of ViT models. This is achieved through the analysis and grouping of tokens according to their spatial autocorrelation scores prior to their input into the Feed-Forward Network (FFN) block of the self-attention mechanism. Importantly, SATA seamlessly integrates into existing pre-trained ViT baselines without requiring retraining or additional fine-tuning, while concurrently improving efficiency by reducing the computational load of the FFN units. Experimental results show that the baseline ViTs enhanced with SATA not only achieve a new state-of-the-art top-1 accuracy on ImageNet-1K image classification \\((94.9\\%)\\) but also establish new state-of-the-art performance across multiple robustness benchmarks, including ImageNet-A (top-1=63.6%), ImageNet-R (top-1=79.2%), and ImageNet-C (mCE=13.6%), all without requiring additional training or fine-tuning of baseline models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.419, + 0.748 + ], + "angle": 0, + "content": "Availability: https://github.com/nick-nikzad/SATA" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.783, + 0.222, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In recent years, vision transformers (ViTs) have demonstrated exceptional performance across diverse computer vision applications [9, 18]. Drawing inspiration from the significant achievements of transformer architectures in natural language processing (NLP), ViTs divide an input image into a sequence of patches (tokens) and leverage self-attention" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.908, + 0.565 + ], + "angle": 0, + "content": "layers [48] to capture relationships between these tokens, ultimately generating rich representations for visual recognition tasks. While recent studies indicate that ViTs can exhibit greater robustness than Convolutional Networks (ConvNets), attributed to their self-attention mechanism [1, 2, 35, 37]. However, this hypothesis has been challenged. Liu et al. [30] demonstrated that a carefully constructed ConvNet can surpass ViTs in both generalization and robustness. Furthermore, while techniques such as patch augmentation [32, 38], contrastive learning strategies [16, 38], and network adjustments [32, 57] have shown promise in enhancing ViT performance and robustness, two primary limitations remain: 1) these methods require extensive retraining or fine-tuning on large datasets (e.g., ImageNet-1K, ImageNet-21K), a labour-intensive and resource-demanding process, especially for large-scale ViT architectures; and 2) the attention maps generated by the self-attention mechanism are highly sensitive to noise, which degrades robustness when inputs are corrupted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.747 + ], + "angle": 0, + "content": "Recently, Nikzad et al. [36] showed the existence of spatial correlation among feature maps in convolutional neural networks (CNNs). Moreover, they observed a decrease in spatial autocorrelation among feature maps through deeper network layers, suggesting that final features exhibit reduced spatial dependency. Motivated by these findings, in this work, we first investigate spatial autocorrelation within Vision Transformer (ViT) architectures and its implications for their performance and robustness. Then, we present a novel approach named \"Spatial Autocorrelation Token Analysis\" (SATA) to tackle the identified shortcomings in the current efforts to enhance ViT robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Our analysis shows that spatial autocorrelation among visual patches (tokens) diminishes through ViT networks, similar to CNNs [36], aligning with findings that spatially uncorrelated features enhance recognition performance [25]. SATA integrates seamlessly into pre-trained ViT models without retraining, boosting robustness and efficiency by reducing the computational load on FFN units. These enhancements in feature aggregation and resilience to corruption account for SATA's exceptional performance from two perspectives:" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9730" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.234, + 0.09, + 0.756, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.908, + 0.349 + ], + "angle": 0, + "content": "Figure 1. Visual comparison of class token attention maps and spatial autocorrelation score maps across three layers of Deit-Base/16 pre-trained on ImageNet-1K. The clean image is sourced from the 'goldfinch' class of the ImageNet-1K dataset, and its corresponding corrupted version, with maximum severity (5), is sourced from the ImageNet-C [20] dataset. \\(\\delta\\) represents the cosine similarity between either the attention maps or the spatial autocorrelation score maps of a corrupted image and its corresponding clean version at each block." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.356, + 0.486, + 0.581 + ], + "angle": 0, + "content": "- Recognition performance: Our study shows that in the later layers of ViT networks, patches with extremely high or low spatial autocorrelation scores in non-informative regions can impede recognition performance and compromise the network's robustness against corrupted inputs. Our proposed SATA method adopts a unique splitting and grouping algorithm to address this issue based on tokens' spatial relation scores in the later layers. This approach prevents the input of unnecessary tokens into the FFN block of the self-attention mechanism, resulting in richer feature aggregation. Unlike token pruning and merging techniques, SATA concatenates bypassed tokens (Residual tokens) with the output of the FFN module, restoring the original number of tokens and preventing information loss for later blocks of the ViT." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.583, + 0.485, + 0.764 + ], + "angle": 0, + "content": "- Robustness performance: As illustrated in Figure 1, our study shows that in the layers of ViT networks, the spatial autocorrelation scores of patches are more robust against different corruptions than their attention maps. Specifically, while the cosine similarity between clean and corrupted attention maps drops significantly in the blocks of the Vision Transformer, the similarity for spatial autocorrelation scores remains consistently high. This highlights that our spatial autocorrelation analysis provides a more stable and reliable schema to achieve corruption-resistant feature representation throughout the network, offering strong robustness against various types of corruption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Extensive experiments conducted on ImageNet-1K image classification and various robustness evaluation benchmarks demonstrate the effectiveness of the proposed spatial autocorrelation paradigm in significantly improving the robustness and accuracy performance of Vision Transformers (ViTs). These findings establish a new state-of-the-art performance level, achieving a top-1 accuracy of \\(94.9\\%\\) on ImageNet-1K image classification, as well as impressive results across multiple robustness benchmarks, including ImageNet-A [8] (top" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.909, + 0.448 + ], + "angle": 0, + "content": "\\(1 = 63.6\\%\\) ), ImageNet-R [22] (top-1=79.2%), and ImageNet-C [20] (mCE=13.6%), without requiring additional expensive fine-tuning or training. Furthermore, in-depth investigations are conducted to thoroughly explore the characteristics of the proposed Spatial Autocorrelation Token Analysis (SATA)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.665, + 0.477 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.702, + 0.501 + ], + "angle": 0, + "content": "2.1. Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Since the introduction of Vision Transformers (ViTs), they have achieved remarkable success in various computer vision tasks [9, 18]. Most improvements to date have focused on enhancing either the accuracy or the efficiency of ViTs. Numerous ViT variants have been proposed to boost their performance [18]. Through dedicated data augmentation [46] and advanced self-attention structures [13, 53], ViTs have demonstrated competitive or superior performance compared to convolutional neural networks (CNNs). Hybrid models like CvT [51] introduce intrinsic inductive bias into the ViT architecture by adding additional convolutional layers before the multi-head self-attention (MHSA) modules. CeiT [54] extracts low-level features through the Image-to-Token (I2T) module and enhances locality by replacing the standard feedforward network with the locally enhanced feed-forward (LeFF) layer. To enable ViTs to learn multi-scale features, CrossViT [5] employs a dual-branch transformer that combines different sizes of image patches to produce stronger image features. ViTAE [52] incorporates multi-scale context by designing reduction cells (RC) and normal cells (NC)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "To create efficient Vision Transformers, several recent works have focused on pruning [11, 33, 44] or combining [24, 28] tokens. ResT [56] introduces an efficient self-attention module using overlapping depth-wise convolutions, while T2T-ViT [55] employs a Tokens-to-Token (T2T) module for token aggregation. PiT [23] reduces spatial size with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9731" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "pooling layers, and Dynamic-ViT [40] dynamically prunes tokens during inference. CaiT [47] optimizes the ViT architecture with layer scaling and class-attention mechanisms. More recently, Bolya et al.[3] proposed a simple token merging technique that potentially does not require retraining." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.177, + 0.276, + 0.192 + ], + "angle": 0, + "content": "2.2. Robustness of ViTs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.485, + 0.305 + ], + "angle": 0, + "content": "While several Recent research has yielded mixed results on the robustness of Vision Transformers (ViTs) compared to Convolutional Neural Networks (CNNs). While some studies [2, 14, 35, 37] suggest ViTs are more robust against various perturbations and distribution shifts, Liu et al. [30] challenge this notion by demonstrating that a well-designed CNN can outperform ViTs in generalization and robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.457 + ], + "angle": 0, + "content": "To enhance ViT robustness, various methods have been proposed, including network structural adjustments, patch augmentation, and diverse training strategies [4, 14-16, 26, 32, 38, 57]. For instance, Robust Vision Transformer (RVT) [32] introduces a convolutional stem and token pooling to improve robustness, while Full Attention Net (FAN) [57] leverages an attentional channel processing design. RobustViT [4] downplays the influence of image backgrounds, and a method proposed in [14] uses temperature scaling to smooth attention weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.485, + 0.761 + ], + "angle": 0, + "content": "Additionally, Qin et al.[38] improve the robustness of ViTs by using images transformed with patch-based operations as negative augmentation. Li et al. [26] propose TORA-ViT, which consists of an accuracy adapter, a robustness adapter, and a gated fusion module. The accuracy adapter extracts predictive features, while the robustness adapter extracts robust features. These features are then combined by the gated fusion module. Reducing Sensitivity to Patch Corruptions (RSPC) [16] enhances the robustness of ViTs through a specialized training strategy. In [15], the Attention Diversification Loss (ADL) is introduced to encourage output tokens to aggregate information from a diverse set of input tokens. Recently, Shi et al. [43] introduced learnable tokens and employed pixel-focused attention to capture structured information between tokens. However, most of these approaches require extensive training or fine-tuning and often sacrifice performance for efficiency [3]. In contrast, our method can be applied to baseline vision transformers[9, 46] without requiring additional training and without any performance drop." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.231, + 0.792 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.483, + 0.833 + ], + "angle": 0, + "content": "3.1. Vision Transformers: Multi-head Self Attention (MHSA)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A standard ViT [9] partitions an input image into \\(N\\) patches (tokens). These patches are then transformed to generate a token embedding tensor \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times d}\\). These tokens are then processed through a stack of transformer blocks, as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "illustrated in Figure 2(a). ViTs leverage self-attention [48] to aggregate global information. Given the input token embedding tensor \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times d}\\), self-attention applies linear transformations with parameters \\(W_{K}, W_{Q}\\), and \\(W_{V}\\) to embed them into the key \\(K = W_{K}\\mathbf{X}\\), query \\(Q = W_{Q}\\mathbf{X}\\), and value \\(V = W_{V}\\mathbf{X}\\), respectively. Self-attention utilises \\(K\\) and \\(Q\\) to generate a pairwise attention map \\(\\mathbf{M}_{att} \\in \\mathbb{R}^{N \\times N}\\) and then aggregates the token features using the attention map \\(\\mathbf{M}_{att}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.232, + 0.907, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {a t t} = \\operatorname {S o f t m a x} \\left(Q K ^ {t} / \\sqrt {d}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.253, + 0.907, + 0.269 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S e l f - A t t e n t i o n} (Q, K, V) = \\mathbf {M} _ {\\text {a t t}} V, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.909, + 0.41 + ], + "angle": 0, + "content": "where the symbol “\\(t\\)” indicates the transpose of the matrix. To achieve rich feature hierarchies, the Transformer block employs multiple self-attention heads. Specifically, \\(h\\) heads are stacked in parallel, resulting in an output of \\(N \\times h \\times d\\). These concatenated features are then processed by a feedforward network (FFN) for further transformation. Finally, the FFN output of \\(N \\times d\\) serves as the final output of the Multi-Head Self-Attention (MHSA) block within the Transformer architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.417, + 0.845, + 0.433 + ], + "angle": 0, + "content": "3.2. Geographical Spatial Auto-correlation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.909, + 0.68 + ], + "angle": 0, + "content": "In geographical modelling, spatial autocorrelation plays a crucial role in assessing the spatial interdependence of entities based on their locations and values. Positive spatial autocorrelation indicates that neighbouring observations share similar values, while negative spatial autocorrelation suggests that nearby observations tend to have contrasting values. Typically, two types of measures are used: global measures, which provide an overall assessment of spatial autocorrelation across all data points, and local measures, which offer insights into the spatial autocorrelation of individual locations relative to their neighbourhoods. Moran's metric [6, 34] is commonly employed in geographical analysis to compute such measurements. In this study, we employ Moran's measurement for the first time, to the best of our knowledge, to investigate spatial dependency among vision transformers' tokens (patches)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.908, + 0.741 + ], + "angle": 0, + "content": "Let \\(\\mathbf{X}\\) be a set of \\(N\\) observations (here, tokens) presented by embedded vectors \\(\\mathbf{x}_i\\in \\mathbb{R}^d\\), \\(\\mathbf{X} = [\\mathbf{x}_1,\\mathbf{x}_2,\\dots,\\mathbf{x}_N]\\), and an associated attribute, \\(\\mathbf{a} = [a_1,a_2,\\dots,a_N]\\), the local Moran's I metric can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.747, + 0.907, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {l} = \\left[ \\operatorname {d i a g} \\left(\\mathbf {z z} ^ {t} \\mathbf {W}\\right) \\right] _ {N \\times 1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.769, + 0.909, + 0.874 + ], + "angle": 0, + "content": "where \\(\\mathrm{diag}(.)\\) returns the diagonal elements of a matrix. Symbol \"t\" indicates the transpose of the matrix. \\(\\mathbf{W} = [w_{ij}]_{N\\times N}\\) represents spatial weight matrix, in which \\(w_{ij}\\) denotes the degree of closeness or the contiguous relationships between \\(\\mathbf{x}_i\\) and \\(\\mathbf{x}_j\\) and can be computed using a dot product similarity \\((\\mathbf{x}_i.\\mathbf{x}_j^t i,j = 1,2,\\dots,N)\\). \\(\\mathbf{z}\\) refers to normalised token-wise attribute values as:" + }, + { + "type": "equation", + "bbox": [ + 0.668, + 0.878, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\frac {\\mathbf {a} - \\mu}{\\sigma}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9732" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.273, + 0.201, + 0.284 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.272, + 0.638, + 0.284 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.285, + 0.704, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.416, + 0.508, + 0.428 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.439, + 0.908, + 0.469 + ], + "angle": 0, + "content": "Figure 2. (a) Comparison between conventional ViT block and the augmented ViT with SATA (b) Overall architecture of the proposed SATA module. (c) Tokens Merging." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.484, + 0.486, + 0.53 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) denote mean and standard deviation of a, respectively. The final local spatial autocorrelation descriptor, s, can be defined as normalised \\(I_{l}\\) [36]:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.54, + 0.484, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} = \\frac {\\boldsymbol {I} _ {l} - \\mu_ {\\boldsymbol {I} _ {l}}}{\\sigma_ {\\boldsymbol {I} _ {l}}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mu_{I_l}\\) and \\(\\sigma_{I_l}\\) indicate mean and standard deviation of \\(I_{l}\\), respectively. Following [36], given a token embedding tensor \\(\\mathbf{X} = [\\mathbf{x}_1,\\mathbf{x}_2,\\dots,\\mathbf{x}_N]\\in \\mathbb{R}^{N\\times d}\\), its token-wise global context attribute \\(\\mathbf{a} = [a_1,a_2,\\dots,a_N]\\in \\mathbb{R}^{N\\times 1}\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.669, + 0.484, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} = \\left[ a _ {i} = \\frac {1}{d} \\sum_ {t = 1} ^ {d} \\mathbf {x} _ {i} (t) \\right] _ {N \\times 1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.485, + 0.815 + ], + "angle": 0, + "content": "where \\(d\\) denotes the spatial dimension of the tokens. \\(\\mathbf{x}_i(t)\\) represents the \\(i\\)-th token value at position \\(t\\). It's worth noting that more advanced strategies or application-specific criteria can be employed to derive the global contextual information descriptor (Eq.(6)). In this context, we adopt the same approach as [36] for the sake of simplicity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.83, + 0.445, + 0.848 + ], + "angle": 0, + "content": "4. Spatial Autocorrelation Token Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Figure 2(b) illustrates the overall architecture of the proposed spatial autocorrelation token analysis module, situated between attention and FFN units within a standard ViT block," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.564, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Figure 3. Plotting the variations of \\(\\mu_{\\mathrm{s}}\\), \\(|\\hat{\\mathbf{s}}|\\), and the lower and upper bounds across different blocks of the ViT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.907, + 0.724 + ], + "angle": 0, + "content": "as depicted in Figure 2(a). To initiate our spatial autocorrelation token analysis, we begin with the observation of the spatial autocorrelation scores, s (Eq. (5)), for ViT's token embedding tensors over different blocks. As transformers inherently capture pairwise closeness relationships between tokens by computing the attention map \\(\\mathbf{M}_{att}\\) in Eq.(2), we can directly set \\(\\mathbf{W} = \\mathbf{M}_{att}\\) in Eq.(3) to enhance both efficiency and effectiveness1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.725, + 0.909, + 0.862 + ], + "angle": 0, + "content": "Figure 3 illustrates the alterations in the mean \\((\\mu_{\\mathrm{s}})\\) and the absolute value of the median \\((|\\hat{\\mathbf{s}}|)\\) statistics of the spatial autocorrelation (s) across different blocks of Deit-Base/16 on the ImageNet-1K validation set. It can be seen that for the later layers, specifically starting from block six, tokens tend to exhibit lower \\(|\\hat{\\mathbf{s}}|\\) values. Drawing from the aforementioned observation, we encapsulate the proposed analysis into two sequential steps to handle these tokens to improve the ViT's robustness and performance:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.903 + ], + "angle": 0, + "content": "1The implementation details of the proposed SATA are provided in Supplementary material" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.248, + 0.108 + ], + "angle": 0, + "content": "4.1. Token Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.22 + ], + "angle": 0, + "content": "Based on the above findings, the overall of the proposed SATA method is illustrated in Figure 2 (b). As shown in Figure 3, we limit token processing to the latter stages of the transformer. Specifically, tokens in layers from \\(\\gamma \\times B\\) onward (\\(\\gamma > 0\\), where \\(B\\) represents the depth, or number of blocks, of the transformer) are partitioned into two sets, \\(\\mathbb{A}\\) and \\(\\mathbb{B}\\), using the spatial autocorrelation scores s as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.242, + 0.484, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathbb {A} = \\left\\{\\mathbf {x} _ {i}; s _ {i} < \\alpha \\left(\\mu_ {\\mathbf {s}} - | \\hat {\\mathbf {s}} |\\right) \\text {a n d} s _ {i} > \\alpha \\left(\\mu_ {\\mathbf {s}} + | \\hat {\\mathbf {s}} |\\right) \\right\\} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.261, + 0.484, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\mathbb {B} = \\left\\{\\mathbf {x} _ {j}; \\alpha \\left(\\mu_ {\\mathbf {s}} - | \\hat {\\mathbf {s}} |\\right) < = s _ {j} < = \\alpha \\left(\\mu_ {\\mathbf {s}} + | \\hat {\\mathbf {s}} |\\right) \\right\\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.485, + 0.332 + ], + "angle": 0, + "content": "where \\(\\alpha (\\mu_{\\mathbf{s}} - |\\hat{\\mathbf{s}} |)\\) and \\(\\alpha (\\mu_{\\mathbf{s}} + |\\hat{\\mathbf{s}} |)\\) represent lower and upper bounds, respectively. \\(\\alpha\\) denotes the controlling factor (parameters \\(\\alpha\\) and \\(\\gamma\\) choices are discussed in Section 5.6)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.337, + 0.255, + 0.354 + ], + "angle": 0, + "content": "4.2. Token Grouping" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.45 + ], + "angle": 0, + "content": "To manage tokens falling beyond the lower and upper bounds (i.e. \\(\\alpha (\\mu_{\\mathrm{s}}\\pm |\\hat{\\mathbf{s}} |))\\) , inspired by the Bipartite Matching algorithm [3] we employ a unique Tokens Merging algorithm to efficiently match and merge similar tokens in set A. In particular, the proposed Tokens Merging algorithm can be summarized as follows (illustrated in Figure 2(c)):" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.451, + 0.483, + 0.478 + ], + "angle": 0, + "content": "1. Partition set \\(\\mathbb{A}\\) into two sets \\(\\mathbb{A}_1\\) and \\(\\mathbb{A}_2\\) of roughly equal size." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.48, + 0.483, + 0.51 + ], + "angle": 0, + "content": "2. Draw one edge from each token in \\(\\mathbb{A}_1\\) to its most similar token in \\(\\mathbb{A}_2\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.511, + 0.484, + 0.54 + ], + "angle": 0, + "content": "3. Merged Tokens: Merge tokens by averaging their features if they are connected." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.541, + 0.484, + 0.571 + ], + "angle": 0, + "content": "4. Residual Tokens: Tokens from set \\(\\mathbb{A}_1\\) representing maximum similarity with tokens in set \\(\\mathbb{A}_2\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.451, + 0.484, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.486, + 0.783 + ], + "angle": 0, + "content": "The proposed Merging algorithm includes unconnected tokens from set \\(\\mathbb{A}_2\\) in the merged tokens output, whereas the Bipartite Matching algorithm [3] disregards these tokens. This feature reduces information loss within the overall SATA schema. As depicted in Figure 2(b), the output tokens of the proposed SATA module comprise the concatenation of tokens with spatial scores within the range of lower and upper bounds (i.e., set \\(\\mathbb{B}\\)) and the Merged Tokens. This output is then fed into the FFN module, reducing unnecessary token processing in the FFN block and lowering the computational load and GFLOPs. Furthermore, Residual Tokens are concatenated with the FFN output to restore the original number of tokens, \\(N\\), forming the final output of the new ViT block, as shown on the right side of Figure 2(a)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.379, + 0.811 + ], + "angle": 0, + "content": "5. Experiment Results & Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.279, + 0.835 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Implementation Details All experiments were conducted on an NVIDIA V100 GPU with a \\(224 \\times 224\\) image resolution. The batch size was set to 256 for all experiments unless otherwise specified. We integrated the proposed SATA module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.153 + ], + "angle": 0, + "content": "into pre-trained generic vision transformers [9, 46] (DeiTiny/16, DeiT-Small/16, DeiT-Base/16, and Vit-Base/16), resulting in three model sizes named SATA-T, SATA-S, SATA-B, and SATA-B*, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.909, + 0.47 + ], + "angle": 0, + "content": "Evaluation Benchmarks We employ the ImageNet2012-1K [7] dataset for standard performance evaluation. For robustness assessment, we evaluate the proposed SATA in three dimensions: 1) Adversarial Robustness: Testing is conducted on adversarial examples generated by white-box attack algorithms FGSM [12] and PGD [31] using the ImageNet-1K validation set. ImageNet-A [8] (IN-A) includes the ImageNet objects in unusual contexts or orientations and is utilized to assess model performance against natural adversarial examples. 2) Common Corruption Robustness: We use ImageNet-C [20] (IN-C), which includes 15 types of algorithmically generated corruptions, each with five levels of severity. 3) Out-of-Distribution Robustness: Evaluation is performed on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK). Both datasets feature images with naturally occurring distribution shifts. ImageNet-R [22] (IN-R) contains abstract or rendered versions of the objects. ImageNet-Sketch [49] consists solely of sketch images, serving to test classification capability when texture or colour information is absent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.816, + 0.496 + ], + "angle": 0, + "content": "5.2. Standard Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.747 + ], + "angle": 0, + "content": "For standard performance evaluation, we compare our method with several state-of-the-art classification models, including Transformer-based models and representative CNN-based models, as shown in Table 1. Our proposed SATA significantly outperforms all other architectures, including both CNN-based and ViT-based models. Specifically, ViT models enhanced with SATA achieve new state-of-the-art top-1 accuracy of \\(86.5\\%\\), \\(89.3\\%\\), and \\(93.9\\%\\) for the tiny, small, and base versions, respectively, all without requiring additional training, input augmentation, or fine-tuning. Notably, integrating the proposed SATA into pre-trained ViT-Base/16 [9] (SATA-\\(\\mathbf{B}^*\\)) results in an additional \\(1.0\\%\\) improvement. Furthermore, comparing the computation cost (GFLOPs) of the baseline DeiTs and SATA models demonstrates that the proposed spatial autocorrelation token analysis method also improves efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.821, + 0.773 + ], + "angle": 0, + "content": "5.3. Adversarial Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For evaluating white-box attack adversarial robustness, we follow [32] and adopt the single-step attack algorithm FGSM [12] and the multi-step attack algorithm PGD [31] (with 5 steps and a step size of 0.5). Both attackers perturb the input image with a maximum magnitude of \\(\\epsilon = 1\\). As shown in Table 1, adversarial robustness appears unrelated to standard performance. For instance, models like Swin [29], PVT [50], and TNT-S [17] achieve higher standard accuracy" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9734" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 1. Performance of SATA and several state-of-the-art (SOTA) CNNs and ViTs models on ImageNet and six robustness benchmarks: We report the mean corruption error (mCE) for ImageNet-C [20], where lower mCE values indicate higher model robustness. Our SATA models consistently outperform other counterparts in standard performance and enhance robustness across various model sizes compared to the baseline, all without requiring additional training or fine-tuning. SATA-B* refers to the integration of the proposed SATA module into the pre-trained vanilla ViT-Base/16 model [9]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.167, + 0.907, + 0.713 + ], + "angle": 0, + "content": "
GroupModelFLOPs (G)Params (M)ImageNet-1K Top-1Top-5FGSMPGDRobustness Benchmarks
IN-C (mCE↓)IN-AIN-RIN-SK
CNNResNet50 [19]4.125.676.186.012.20.976.70.036.124.1
RegNetY-4GF[39]4.020.679.294.715.42.468.78.938.825.9
EfficientNet-B4[45]4.419.383.096.344.618.571.126.347.134.1
DeepAugment[21]4.125.675.892.727.19.553.63.946.732.6
ANT[41]4.125.676.193.017.83.163.01.139.026.3
Debiased CNN[27]4.125.676.993.420.45.567.53.540.828.4
ConvNeXt-B[30]15.48983.8---46.836.751.338.2
ViT-TinyDeiT-Ti[46]1.35.772.291.122.36.271.17.332.620.2
ConvViT-Ti[10]1.45.773.391.824.77.568.48.935.222.4
PiT-Ti[23]0.74.972.991.320.45.169.16.234.621.6
PVT-Tiny[50]1.913.275.092.510.00.579.67.933.921.5
RVT-Ti [32]1.38.678.494.234.811.758.213.343.730.0
FAN-T-ViT [57]1.37.079.2---57.5-42.5-
RVT-Ti+RSPC [16]1.310.979.2---55.716.5--
SATA-T (ours)1.05.786.598.240.010.951.114.647.325.2
ViT-SmallDeiT-S[46]4.622.179.995.040.716.754.618.942.229.4
ConvViT-S[10]5.427.881.595.841.717.249.824.545.433.1
PiT-S[23]2.923.580.995.341.016.552.521.743.630.8
PVT-Small[50]3.824.579.995.026.63.166.918.040.127.2
Swin-T[29]4.528.381.295.533.77.362.021.641.329.1
TNT-S[17]5.223.881.595.733.24.253.124.743.831.6
T2T-ViT_t-14[55]6.121.581.795.940.911.753.223.945.032.5
RVT-S [32]4.722.181.795.751.326.050.124.146.935.0
FAN-S-ViT [57]5.328.082.9---47.729.150.4-
RVT-S+RSPC [16]4.723.382.2---48.427.9--
SATA-S (ours)3.922.189.399.157.418.033.830.559.539.2
ViT-BaseDeiT-B[46]17.686.682.095.746.421.348.527.444.932.4
ConvViT-B[10]17.786.582.095.746.421.348.527.444.932.4
PiT-B[23]12.573.882.495.749.323.748.233.943.732.3
PVT-Large[50]9.861.481.795.933.17.359.826.642.730.2
Swin-B[29]15.487.883.496.449.221.354.435.846.632.4
T2T-ViT_t-24[55]15.064.182.696.146.717.548.428.947.935.4
RVT-B [32]17.786.282.596.052.327.447.327.748.235.8
FAN-B-ViT [57]10.454.083.6---44.435.451.8-
RVT-B+RSPC [16]17.791.882.8---45.732.1--
TORA-ViT-B/16(λ=0.1) [26]26.0111.284.1-48.423.331.746.557.6-
TransNeXt [43]18.489.784.8---43.550.653.941.4
SATA-B (ours)15.986.693.999.763.920.228.763.570.049.8
SATA-B* (ours)15.986.694.999.865.628.313.663.679.257.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.486, + 0.814 + ], + "angle": 0, + "content": "than DeiTs corresponding, but their adversarial robustness is significantly lower, consistent with findings from [32, 42]. Our proposed SATA model achieves superior performance against both FGSM [12] and PGD [31] attacks. Specifically, SATA-S, SATA-B, and SATA-B* show over a \\(20\\%\\) improvement on FGSM [12] compared to previous ViT variants." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Regarding natural adversarial robustness, the proposed SATA-T demonstrates a comparable performance of \\(14.6\\%\\), which is on par with some current state-of-the-art methods like RVT-Ti [32] and RVT-Ti+RSPC [16], while being about half their size. However, for models of similar size (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.907, + 0.768 + ], + "angle": 0, + "content": "ViT-Small and ViT-Base), the proposed SATAs outperform others by about \\(50\\%\\), indicating the effectiveness of SATA against natural adversarial attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.892, + 0.803 + ], + "angle": 0, + "content": "5.4. Common Corruption Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To measure model degradation on common image corruptions, we report the mean corruption error (mCE) on ImageNet-C [20] (IN-C) in Table 1. Our SATA method significantly reduces the mCE of DeiT-Ti [46] from \\(71.1\\%\\) to \\(51.1\\%\\), achieving the lowest mCE among vision transformers within the ViT-Tiny group. For the other two larger" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.089, + 0.473, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.246, + 0.297, + 0.257 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.26, + 0.45, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.409, + 0.297, + 0.421 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Figure 4. (a) Ablation study on \\(\\gamma\\). (b) Ablation study on \\(\\alpha\\). We set \\(\\gamma\\) and \\(\\alpha\\) to 0.7 and 1.0, respectively, for all experiments throughout this paper. Ablation studies are conducted on the SATA-B model using the ImageNet-1K dataset. Dashed lines for both graphs represent the baseline Deit-Base/16 top-1 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.636 + ], + "angle": 0, + "content": "ViT groups, the proposed SATA models achieve an mCE of approximately \\(28\\%\\), improving by around 20 points over all other ViT or CNN-based methods on the leaderboard, thereby establishing a new state-of-the-art. This result also suggests that spatial autocorrelation management of visual tokens can successfully handle different types of image corruption." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.456, + 0.689 + ], + "angle": 0, + "content": "5.5. Out-of-Distribution Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We evaluate the generalization ability of SATA on out-of-distribution data by reporting the top-1 accuracy on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK) in Table 1. The generic vision transformers [9, 46] enhanced by the proposed SATA consistently outperform other ViT models on ImageNet-R [22], achieving \\(47.3\\%\\), \\(57.2\\%\\), \\(70.0\\%\\), \\(79.9\\%\\) in the ViT-Tiny, ViT-Small, and ViT-Base groups, respectively. Regarding ImageNet-Sketch [49] (IN-SK), SATA demonstrates superior performance compared to other models of similar size. These results imply that the spatial autocorrelation tokens analysis effectively captures feature distribution shifts, enhancing the model's out-of-distribution generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.664, + 0.107 + ], + "angle": 0, + "content": "5.6. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.908, + 0.431 + ], + "angle": 0, + "content": "Token Splitting and Token Grouping We evaluate the role of token splitting based on the upper and lower bounds of spatial autocorrelation scores and token grouping modules. To this end, we examine five SATA configurations, as depicted in Table 2. As shown in Table 2, utilizing only token grouping yields a top-1 accuracy of \\(84.4\\%\\), improving upon the reference (DeiT-Base) accuracy by \\(2.4\\%\\). Including either lower or upper bounds significantly enhances accuracy by about \\(10\\%\\) of the baseline top-1 accuracy, highlighting the effectiveness of the proposed spatial autocorrelation token splitting schema. The upper bound contributes slightly more, suggesting that tokens with extremely high spatial autocorrelation scores are more likely to be filtered by the splitting process. However, incorporating upper or lower bounds also increases computations and token counts, reducing throughput despite the substantial accuracy gains. Additionally, applying bounds does not guarantee an equal token distribution across bands, as this depends on the data distribution (shown in Figure 2 of the supplementary material). Finally, adding token grouping provides a further \\(1.6\\%\\) improvement over the splitting process alone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.522 + ], + "angle": 0, + "content": "Regarding throughput, two main components of the proposed SATA framework (i.e. spatial autocorrelation-based token splitting and token grouping) primarily utilise parallelizable matrix multiplication and dot product operations. This design ensures computational efficiency with a negligible impact on inference time, as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.726 + ], + "angle": 0, + "content": "Threshold of starting block \\((\\gamma)\\) We also examine the effect of parameter \\(\\gamma\\), which controls at which transformer block the SATA module is applied. As Figure 4(a) shows, applying SATA from block \\(0.4 \\times B\\) onwards significantly improves model efficiency while exceeding baseline ViT accuracy \\((82\\%)\\). Applying SATA to earlier blocks \\((\\gamma < 0.4)\\) degrades accuracy, suggesting that high spatial correlation within token features and the importance of all tokens in early layers is beneficial. To achieve a good trade-off between accuracy and efficiency, we use \\(\\gamma = 0.7\\) in all our experiments, unless stated otherwise. This results in a top-1 accuracy of \\(93.9\\%\\) and GFLOPs of 15.9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.909, + 0.839 + ], + "angle": 0, + "content": "Lower/Upper bounds controlling factor \\((\\alpha)\\) We further assess the influence of \\(\\alpha\\), the factor controlling the lower and upper bounds in SATA. Figure 4(b) shows the performance of SATA-B on ImageNet-1K with \\(\\alpha\\) values ranging from 0.5 to 2. \\(\\alpha\\) determines the number of tokens passed to the FFN block and setting \\(\\alpha = 1\\) yields optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.768, + 0.862 + ], + "angle": 0, + "content": "5.7. Visualisation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Although the effectiveness of the proposed SATA module has been empirically demonstrated, we conduct a deeper" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9736" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 2. Ablation study of the spatial autocorrelation module (token splitting) and token grouping. The symbols “√” and “X” indicate whether the corresponding element is employed with the configuration or not. † represents top-1 accuracy for the baseline Deit-Base/16 [46] model." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.142, + 0.887, + 0.282 + ], + "angle": 0, + "content": "
Spt. Auto. Correlation (Tokens Splitting)Tokens GroupingTop-1Throughput (ims/s)Latency (s)
Lower bound (μs - |s|)Upper bound (μs + |s|)
×××82.0†3030.0033
××✓84.43610.0027
×✓✓92.82970.0033
✓×✓92.52710.0035
✓✓×92.32680.0037
✓✓✓93.92760.0035
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.298, + 0.456, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.455, + 0.297, + 0.466 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.477, + 0.472, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.607, + 0.294, + 0.618 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.627, + 0.486, + 0.725 + ], + "angle": 0, + "content": "Figure 5. (a) Cosine similarity between the clean and corrupted versions of the class token attention map and spatial autocorrelation scores across different blocks of SATA-B. Results are averaged across various types of image corruptions and severity levels on ImageNet-C [20]. (b) Visualisation of token splitting for a pair of clean and noisy images. Notably, the selected tokens for each set are similar for both clean and corrupted inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "investigation to better understand its behaviour. To this end, we calculate the cosine similarity between the class token attention maps and spatial autocorrelation scores of clean and corrupted image pairs from the ImageNet-1K validation set and its corresponding ImageNet-C [20], respectively. We compute these similarities for various types of image corruptions and severity levels in ImageNet-C [20], and report the average values across different blocks of the proposed SATA-B in Figure 5(a). As shown in Figure 5(a), the cosine similarity between clean and corrupted attention maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.908, + 0.405 + ], + "angle": 0, + "content": "drops significantly in the later blocks of the transformer. In contrast, the similarity for spatial autocorrelation scores improves at the early stages and remains consistently high, averaging above 0.8. This highlights that the proposed method can provide a more stable and reliable feature representation throughout the network, offering strong robustness against various types of corruption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.499 + ], + "angle": 0, + "content": "Moreover, Figure 5 provides a visualization of tokens (patches) are split into set \\(\\mathbb{A}\\) and set \\(\\mathbb{B}\\) for a pair of clean and noisy images according to the proposed SATA algorithm. Notably, the similarity between corresponding sets for clean and noisy inputs is evident, further highlighting the robustness of the proposed method2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.519, + 0.634, + 0.535 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.744 + ], + "angle": 0, + "content": "In this paper, we introduce SATA, a novel method designed to enhance the performance and robustness of vision transformers against various types of corruption. SATA utilises a straightforward yet powerful spatial autocorrelation scheme to exploit spatial inter-dependencies among token features, significantly improving representational capacity and efficiency while reducing computational costs. Our experimental results demonstrate that SATA-enhanced vision transformers consistently deliver stable and reliable feature representations, achieving state-of-the-art performance on ImageNet-1K classification and setting new benchmarks for robustness across multiple evaluations, all without requiring additional training or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.909, + 0.868 + ], + "angle": 0, + "content": "This work highlights SATA's transformative potential and opens several promising avenues for future research. Key directions include adapting SATA for window-based and hybrid ViT architectures to boost performance in tasks such as object detection and segmentation. Additionally, exploring SATA's application in other transformer-based domains, such as large language models (LLMs), could further extend its impact." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "2Additional visual comparisons are included in Supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9737" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.175 + ], + "angle": 0, + "content": "This work was supported in part by the Australian Research Council (ARC), Australia under Discovery Grant DP180100958 and Industrial Transformation Research Hub Grant IH180100002." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? In NeurIPS, pages 26831-26843, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.256, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[2] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In ICCV, pages 10231-10241, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[4] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. In NeurIPS, pages 33618-33632, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[6] Yanguang Chen. New approaches for calculating Moran's index of spatial autocorrelation. *PloS one*, 8(7), 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[8] Minjing Dong, Yanxi Li, Yunhe Wang, and Chang Xu. Adversarily robust neural architectures. arXiv preprint arXiv:2009.00902, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[10] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, pages 2286-2296. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[11] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, pages 396-414. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[13] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, pages 12259-12269, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Jindong Gu, Volker Tresp, and Yao Qin. Are vision transformers robust to patch perturbations? In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.133 + ], + "angle": 0, + "content": "[15] Yong Guo, David Stutz, and Bernt Schiele. Robustifying token attention for vision transformers. In ICCV, pages 17557-17568, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.908, + 0.175 + ], + "angle": 0, + "content": "[16] Yong Guo, David Stutz, and Bernt Schiele. Improving robustness of vision transformers by reducing sensitivity to patch corruptions. In CVPR, pages 4108-4118, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, pages 15908-15919, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.908, + 0.273 + ], + "angle": 0, + "content": "[18] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE TPAMI, 45(1):87-110, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.908, + 0.356 + ], + "angle": 0, + "content": "[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.908, + 0.427 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.908, + 0.496 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.908, + 0.551 + ], + "angle": 0, + "content": "[23] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, pages 11936-11945, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[24] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, pages 620-640. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.908, + 0.677 + ], + "angle": 0, + "content": "[25] Xuelong Li, Han Zhang, Rui Zhang, and Feiping Nie. Discriminative and uncorrelated feature selection with constrained spectral analysis in unsupervised learning. IEEE Transactions on Image Processing, 29:2139-2149, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.908, + 0.719 + ], + "angle": 0, + "content": "[26] Yanxi Li and Chang Xu. Trade-off between robustness and accuracy of vision transformers. In CVPR, pages 7558-7568, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.909, + 0.787 + ], + "angle": 0, + "content": "[27] Yingwei Li, Qihang Yu, Mingxing Tan, Jieru Mei, Peng Tang, Wei Shen, Alan Yuille, and Cihang Xie. Shape-texture de-biased neural network training. In In Proceedings of the International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9738" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[32] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In CVPR, pages 12042-12051, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[33] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[34] Patrick AP Moran. The interpretation of statistical maps. Journal of the Royal Statistical Society. Series B (Methodological), 10(2):243-251, 1948. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[35] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, pages 23296-23308, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[36] Nick Nikzad, Yongsheng Gao, and Jun Zhou. CSA-Net: Channel-wise Spatially Autocorrelated Attention Networks. arXiv preprint arXiv:2405.05755, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[37] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In AAAI, pages 2071-2081, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[38] Yao Qin, Chiyuan Zhang, Ting Chen, Balaji Lakshminarayanan, Alex Beutel, and Xuezhi Wang. Understanding and improving robustness of vision transformers through patch-based negative augmentation. In NeurIPS, pages 16276-16289, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[39] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[41] Evgenia Rusak, Lukas Schott, Roland S Zimmermann, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge, and Wieland Brendel. A simple way to make neural networks robust against diverse image corruptions. In ECCV, pages 53-69. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[42] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.772 + ], + "angle": 0, + "content": "[43] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In CVPR, pages 17773-17783, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[44] Zhuoran Song, Yihong Xu, Zhezhi He, Li Jiang, Naifeng Jing, and Xiaoyao Liang. Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[45] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.22 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[49] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[50] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[51] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, pages 22–31, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[52] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, pages 28522-28535, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[53] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[54] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In ICCV, pages 579-588, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.907, + 0.574 + ], + "angle": 0, + "content": "[55] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.908, + 0.616 + ], + "angle": 0, + "content": "[56] Qinglong Zhang and Yu-Bin Yang. Rest: An efficient transformer for visual recognition. In NeurIPS, pages 15475-15485, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.908, + 0.674 + ], + "angle": 0, + "content": "[57] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In ICML, pages 27378-27394. PMLR, 2022. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9739" + } + ] +] \ No newline at end of file diff --git a/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_origin.pdf b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0928e041759c2460348de745a73da7382d507427 --- /dev/null +++ b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/5e97f7eb-5402-4736-8030-7dfbcb7e2397_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:145ee31f5a4704a9aefd469aeb67890bc4d59cc3167756ca12823e315b637b82 +size 540889 diff --git a/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/full.md b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fab454b27f4735fbf330b8d2e1d10c785e455e82 --- /dev/null +++ b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/full.md @@ -0,0 +1,294 @@ +# SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers + +Nick Nikzad, Yi Liao, Yongsheng Gao, Jun Zhou + +Griffith University, Queensland, Australia + +{n.nikzaddehaji, y.liao, yongsheng.gao, jun.zhou}@griffith.edu.au + +# Abstract + +Over the past few years, vision transformers (ViTs) have consistently demonstrated remarkable performance across various visual recognition tasks. However, attempts to enhance their robustness have yielded limited success, mainly focusing on different training strategies, input patch augmentation, or network structural enhancements. These approaches often involve extensive training and fine-tuning, which are time-consuming and resource-intensive. To tackle these obstacles, we introduce a novel approach named Spatial Autocorrelation Token Analysis (SATA). By harnessing spatial relationships between token features, SATA enhances both the representational capacity and robustness of ViT models. This is achieved through the analysis and grouping of tokens according to their spatial autocorrelation scores prior to their input into the Feed-Forward Network (FFN) block of the self-attention mechanism. Importantly, SATA seamlessly integrates into existing pre-trained ViT baselines without requiring retraining or additional fine-tuning, while concurrently improving efficiency by reducing the computational load of the FFN units. Experimental results show that the baseline ViTs enhanced with SATA not only achieve a new state-of-the-art top-1 accuracy on ImageNet-1K image classification $(94.9\%)$ but also establish new state-of-the-art performance across multiple robustness benchmarks, including ImageNet-A (top-1=63.6%), ImageNet-R (top-1=79.2%), and ImageNet-C (mCE=13.6%), all without requiring additional training or fine-tuning of baseline models. + +Availability: https://github.com/nick-nikzad/SATA + +# 1. Introduction + +In recent years, vision transformers (ViTs) have demonstrated exceptional performance across diverse computer vision applications [9, 18]. Drawing inspiration from the significant achievements of transformer architectures in natural language processing (NLP), ViTs divide an input image into a sequence of patches (tokens) and leverage self-attention + +layers [48] to capture relationships between these tokens, ultimately generating rich representations for visual recognition tasks. While recent studies indicate that ViTs can exhibit greater robustness than Convolutional Networks (ConvNets), attributed to their self-attention mechanism [1, 2, 35, 37]. However, this hypothesis has been challenged. Liu et al. [30] demonstrated that a carefully constructed ConvNet can surpass ViTs in both generalization and robustness. Furthermore, while techniques such as patch augmentation [32, 38], contrastive learning strategies [16, 38], and network adjustments [32, 57] have shown promise in enhancing ViT performance and robustness, two primary limitations remain: 1) these methods require extensive retraining or fine-tuning on large datasets (e.g., ImageNet-1K, ImageNet-21K), a labour-intensive and resource-demanding process, especially for large-scale ViT architectures; and 2) the attention maps generated by the self-attention mechanism are highly sensitive to noise, which degrades robustness when inputs are corrupted. + +Recently, Nikzad et al. [36] showed the existence of spatial correlation among feature maps in convolutional neural networks (CNNs). Moreover, they observed a decrease in spatial autocorrelation among feature maps through deeper network layers, suggesting that final features exhibit reduced spatial dependency. Motivated by these findings, in this work, we first investigate spatial autocorrelation within Vision Transformer (ViT) architectures and its implications for their performance and robustness. Then, we present a novel approach named "Spatial Autocorrelation Token Analysis" (SATA) to tackle the identified shortcomings in the current efforts to enhance ViT robustness. + +Our analysis shows that spatial autocorrelation among visual patches (tokens) diminishes through ViT networks, similar to CNNs [36], aligning with findings that spatially uncorrelated features enhance recognition performance [25]. SATA integrates seamlessly into pre-trained ViT models without retraining, boosting robustness and efficiency by reducing the computational load on FFN units. These enhancements in feature aggregation and resilience to corruption account for SATA's exceptional performance from two perspectives: + +![](images/2d5fe590560b1ece0e7de5a5de5ca3f7497361aed0fddfc7abf79c14262a6cd2.jpg) +Figure 1. Visual comparison of class token attention maps and spatial autocorrelation score maps across three layers of Deit-Base/16 pre-trained on ImageNet-1K. The clean image is sourced from the 'goldfinch' class of the ImageNet-1K dataset, and its corresponding corrupted version, with maximum severity (5), is sourced from the ImageNet-C [20] dataset. $\delta$ represents the cosine similarity between either the attention maps or the spatial autocorrelation score maps of a corrupted image and its corresponding clean version at each block. + +- Recognition performance: Our study shows that in the later layers of ViT networks, patches with extremely high or low spatial autocorrelation scores in non-informative regions can impede recognition performance and compromise the network's robustness against corrupted inputs. Our proposed SATA method adopts a unique splitting and grouping algorithm to address this issue based on tokens' spatial relation scores in the later layers. This approach prevents the input of unnecessary tokens into the FFN block of the self-attention mechanism, resulting in richer feature aggregation. Unlike token pruning and merging techniques, SATA concatenates bypassed tokens (Residual tokens) with the output of the FFN module, restoring the original number of tokens and preventing information loss for later blocks of the ViT. + +- Robustness performance: As illustrated in Figure 1, our study shows that in the layers of ViT networks, the spatial autocorrelation scores of patches are more robust against different corruptions than their attention maps. Specifically, while the cosine similarity between clean and corrupted attention maps drops significantly in the blocks of the Vision Transformer, the similarity for spatial autocorrelation scores remains consistently high. This highlights that our spatial autocorrelation analysis provides a more stable and reliable schema to achieve corruption-resistant feature representation throughout the network, offering strong robustness against various types of corruption. + +Extensive experiments conducted on ImageNet-1K image classification and various robustness evaluation benchmarks demonstrate the effectiveness of the proposed spatial autocorrelation paradigm in significantly improving the robustness and accuracy performance of Vision Transformers (ViTs). These findings establish a new state-of-the-art performance level, achieving a top-1 accuracy of $94.9\%$ on ImageNet-1K image classification, as well as impressive results across multiple robustness benchmarks, including ImageNet-A [8] (top + +$1 = 63.6\%$ ), ImageNet-R [22] (top-1=79.2%), and ImageNet-C [20] (mCE=13.6%), without requiring additional expensive fine-tuning or training. Furthermore, in-depth investigations are conducted to thoroughly explore the characteristics of the proposed Spatial Autocorrelation Token Analysis (SATA). + +# 2. Related Works + +# 2.1. Vision Transformer + +Since the introduction of Vision Transformers (ViTs), they have achieved remarkable success in various computer vision tasks [9, 18]. Most improvements to date have focused on enhancing either the accuracy or the efficiency of ViTs. Numerous ViT variants have been proposed to boost their performance [18]. Through dedicated data augmentation [46] and advanced self-attention structures [13, 53], ViTs have demonstrated competitive or superior performance compared to convolutional neural networks (CNNs). Hybrid models like CvT [51] introduce intrinsic inductive bias into the ViT architecture by adding additional convolutional layers before the multi-head self-attention (MHSA) modules. CeiT [54] extracts low-level features through the Image-to-Token (I2T) module and enhances locality by replacing the standard feedforward network with the locally enhanced feed-forward (LeFF) layer. To enable ViTs to learn multi-scale features, CrossViT [5] employs a dual-branch transformer that combines different sizes of image patches to produce stronger image features. ViTAE [52] incorporates multi-scale context by designing reduction cells (RC) and normal cells (NC). + +To create efficient Vision Transformers, several recent works have focused on pruning [11, 33, 44] or combining [24, 28] tokens. ResT [56] introduces an efficient self-attention module using overlapping depth-wise convolutions, while T2T-ViT [55] employs a Tokens-to-Token (T2T) module for token aggregation. PiT [23] reduces spatial size with + +pooling layers, and Dynamic-ViT [40] dynamically prunes tokens during inference. CaiT [47] optimizes the ViT architecture with layer scaling and class-attention mechanisms. More recently, Bolya et al.[3] proposed a simple token merging technique that potentially does not require retraining. + +# 2.2. Robustness of ViTs + +While several Recent research has yielded mixed results on the robustness of Vision Transformers (ViTs) compared to Convolutional Neural Networks (CNNs). While some studies [2, 14, 35, 37] suggest ViTs are more robust against various perturbations and distribution shifts, Liu et al. [30] challenge this notion by demonstrating that a well-designed CNN can outperform ViTs in generalization and robustness. + +To enhance ViT robustness, various methods have been proposed, including network structural adjustments, patch augmentation, and diverse training strategies [4, 14-16, 26, 32, 38, 57]. For instance, Robust Vision Transformer (RVT) [32] introduces a convolutional stem and token pooling to improve robustness, while Full Attention Net (FAN) [57] leverages an attentional channel processing design. RobustViT [4] downplays the influence of image backgrounds, and a method proposed in [14] uses temperature scaling to smooth attention weights. + +Additionally, Qin et al.[38] improve the robustness of ViTs by using images transformed with patch-based operations as negative augmentation. Li et al. [26] propose TORA-ViT, which consists of an accuracy adapter, a robustness adapter, and a gated fusion module. The accuracy adapter extracts predictive features, while the robustness adapter extracts robust features. These features are then combined by the gated fusion module. Reducing Sensitivity to Patch Corruptions (RSPC) [16] enhances the robustness of ViTs through a specialized training strategy. In [15], the Attention Diversification Loss (ADL) is introduced to encourage output tokens to aggregate information from a diverse set of input tokens. Recently, Shi et al. [43] introduced learnable tokens and employed pixel-focused attention to capture structured information between tokens. However, most of these approaches require extensive training or fine-tuning and often sacrifice performance for efficiency [3]. In contrast, our method can be applied to baseline vision transformers[9, 46] without requiring additional training and without any performance drop. + +# 3. Preliminaries + +# 3.1. Vision Transformers: Multi-head Self Attention (MHSA) + +A standard ViT [9] partitions an input image into $N$ patches (tokens). These patches are then transformed to generate a token embedding tensor $\mathbf{X} \in \mathbb{R}^{N \times d}$ . These tokens are then processed through a stack of transformer blocks, as + +illustrated in Figure 2(a). ViTs leverage self-attention [48] to aggregate global information. Given the input token embedding tensor $\mathbf{X} \in \mathbb{R}^{N \times d}$ , self-attention applies linear transformations with parameters $W_{K}, W_{Q}$ , and $W_{V}$ to embed them into the key $K = W_{K}\mathbf{X}$ , query $Q = W_{Q}\mathbf{X}$ , and value $V = W_{V}\mathbf{X}$ , respectively. Self-attention utilises $K$ and $Q$ to generate a pairwise attention map $\mathbf{M}_{att} \in \mathbb{R}^{N \times N}$ and then aggregates the token features using the attention map $\mathbf{M}_{att}$ as follows: + +$$ +\mathbf {M} _ {a t t} = \operatorname {S o f t m a x} \left(Q K ^ {t} / \sqrt {d}\right), \tag {1} +$$ + +$$ +\operatorname {S e l f - A t t e n t i o n} (Q, K, V) = \mathbf {M} _ {\text {a t t}} V, \tag {2} +$$ + +where the symbol “ $t$ ” indicates the transpose of the matrix. To achieve rich feature hierarchies, the Transformer block employs multiple self-attention heads. Specifically, $h$ heads are stacked in parallel, resulting in an output of $N \times h \times d$ . These concatenated features are then processed by a feedforward network (FFN) for further transformation. Finally, the FFN output of $N \times d$ serves as the final output of the Multi-Head Self-Attention (MHSA) block within the Transformer architecture. + +# 3.2. Geographical Spatial Auto-correlation + +In geographical modelling, spatial autocorrelation plays a crucial role in assessing the spatial interdependence of entities based on their locations and values. Positive spatial autocorrelation indicates that neighbouring observations share similar values, while negative spatial autocorrelation suggests that nearby observations tend to have contrasting values. Typically, two types of measures are used: global measures, which provide an overall assessment of spatial autocorrelation across all data points, and local measures, which offer insights into the spatial autocorrelation of individual locations relative to their neighbourhoods. Moran's metric [6, 34] is commonly employed in geographical analysis to compute such measurements. In this study, we employ Moran's measurement for the first time, to the best of our knowledge, to investigate spatial dependency among vision transformers' tokens (patches). + +Let $\mathbf{X}$ be a set of $N$ observations (here, tokens) presented by embedded vectors $\mathbf{x}_i\in \mathbb{R}^d$ , $\mathbf{X} = [\mathbf{x}_1,\mathbf{x}_2,\dots,\mathbf{x}_N]$ , and an associated attribute, $\mathbf{a} = [a_1,a_2,\dots,a_N]$ , the local Moran's I metric can be defined as: + +$$ +\boldsymbol {I} _ {l} = \left[ \operatorname {d i a g} \left(\mathbf {z z} ^ {t} \mathbf {W}\right) \right] _ {N \times 1}, \tag {3} +$$ + +where $\mathrm{diag}(.)$ returns the diagonal elements of a matrix. Symbol "t" indicates the transpose of the matrix. $\mathbf{W} = [w_{ij}]_{N\times N}$ represents spatial weight matrix, in which $w_{ij}$ denotes the degree of closeness or the contiguous relationships between $\mathbf{x}_i$ and $\mathbf{x}_j$ and can be computed using a dot product similarity $(\mathbf{x}_i.\mathbf{x}_j^t i,j = 1,2,\dots,N)$ . $\mathbf{z}$ refers to normalised token-wise attribute values as: + +$$ +\mathbf {z} = \frac {\mathbf {a} - \mu}{\sigma}, \tag {4} +$$ + +![](images/0b8c83390f656fadadb70704ce1ae9d2c4be6e5819ee21ebb34b18a77254b422.jpg) +(a) +(b) + +![](images/1bf9a59b62008c79d08a6eab62fbce5ab9b6cb713ac66f86583394ac664b22c7.jpg) +(c) + +where $\mu$ and $\sigma$ denote mean and standard deviation of a, respectively. The final local spatial autocorrelation descriptor, s, can be defined as normalised $I_{l}$ [36]: + +$$ +\mathbf {s} = \frac {\boldsymbol {I} _ {l} - \mu_ {\boldsymbol {I} _ {l}}}{\sigma_ {\boldsymbol {I} _ {l}}}, \tag {5} +$$ + +where $\mu_{I_l}$ and $\sigma_{I_l}$ indicate mean and standard deviation of $I_{l}$ , respectively. Following [36], given a token embedding tensor $\mathbf{X} = [\mathbf{x}_1,\mathbf{x}_2,\dots,\mathbf{x}_N]\in \mathbb{R}^{N\times d}$ , its token-wise global context attribute $\mathbf{a} = [a_1,a_2,\dots,a_N]\in \mathbb{R}^{N\times 1}$ can be defined as: + +$$ +\mathbf {a} = \left[ a _ {i} = \frac {1}{d} \sum_ {t = 1} ^ {d} \mathbf {x} _ {i} (t) \right] _ {N \times 1}, \tag {6} +$$ + +where $d$ denotes the spatial dimension of the tokens. $\mathbf{x}_i(t)$ represents the $i$ -th token value at position $t$ . It's worth noting that more advanced strategies or application-specific criteria can be employed to derive the global contextual information descriptor (Eq.(6)). In this context, we adopt the same approach as [36] for the sake of simplicity. + +# 4. Spatial Autocorrelation Token Analysis + +Figure 2(b) illustrates the overall architecture of the proposed spatial autocorrelation token analysis module, situated between attention and FFN units within a standard ViT block, + +![](images/026e7ef143c9f562ecf37face211d3e544281d1f7ebd4a064be62ea1d499ac3b.jpg) +Figure 2. (a) Comparison between conventional ViT block and the augmented ViT with SATA (b) Overall architecture of the proposed SATA module. (c) Tokens Merging. +Figure 3. Plotting the variations of $\mu_{\mathrm{s}}$ , $|\hat{\mathbf{s}}|$ , and the lower and upper bounds across different blocks of the ViT. + +as depicted in Figure 2(a). To initiate our spatial autocorrelation token analysis, we begin with the observation of the spatial autocorrelation scores, s (Eq. (5)), for ViT's token embedding tensors over different blocks. As transformers inherently capture pairwise closeness relationships between tokens by computing the attention map $\mathbf{M}_{att}$ in Eq.(2), we can directly set $\mathbf{W} = \mathbf{M}_{att}$ in Eq.(3) to enhance both efficiency and effectiveness1. + +Figure 3 illustrates the alterations in the mean $(\mu_{\mathrm{s}})$ and the absolute value of the median $(|\hat{\mathbf{s}}|)$ statistics of the spatial autocorrelation (s) across different blocks of Deit-Base/16 on the ImageNet-1K validation set. It can be seen that for the later layers, specifically starting from block six, tokens tend to exhibit lower $|\hat{\mathbf{s}}|$ values. Drawing from the aforementioned observation, we encapsulate the proposed analysis into two sequential steps to handle these tokens to improve the ViT's robustness and performance: + +# 4.1. Token Splitting + +Based on the above findings, the overall of the proposed SATA method is illustrated in Figure 2 (b). As shown in Figure 3, we limit token processing to the latter stages of the transformer. Specifically, tokens in layers from $\gamma \times B$ onward ( $\gamma > 0$ , where $B$ represents the depth, or number of blocks, of the transformer) are partitioned into two sets, $\mathbb{A}$ and $\mathbb{B}$ , using the spatial autocorrelation scores s as follows: + +$$ +\mathbb {A} = \left\{\mathbf {x} _ {i}; s _ {i} < \alpha \left(\mu_ {\mathbf {s}} - | \hat {\mathbf {s}} |\right) \text {a n d} s _ {i} > \alpha \left(\mu_ {\mathbf {s}} + | \hat {\mathbf {s}} |\right) \right\} \tag {7} +$$ + +$$ +\mathbb {B} = \left\{\mathbf {x} _ {j}; \alpha \left(\mu_ {\mathbf {s}} - | \hat {\mathbf {s}} |\right) < = s _ {j} < = \alpha \left(\mu_ {\mathbf {s}} + | \hat {\mathbf {s}} |\right) \right\} \tag {8} +$$ + +where $\alpha (\mu_{\mathbf{s}} - |\hat{\mathbf{s}} |)$ and $\alpha (\mu_{\mathbf{s}} + |\hat{\mathbf{s}} |)$ represent lower and upper bounds, respectively. $\alpha$ denotes the controlling factor (parameters $\alpha$ and $\gamma$ choices are discussed in Section 5.6). + +# 4.2. Token Grouping + +To manage tokens falling beyond the lower and upper bounds (i.e. $\alpha (\mu_{\mathrm{s}}\pm |\hat{\mathbf{s}} |))$ , inspired by the Bipartite Matching algorithm [3] we employ a unique Tokens Merging algorithm to efficiently match and merge similar tokens in set A. In particular, the proposed Tokens Merging algorithm can be summarized as follows (illustrated in Figure 2(c)): + +1. Partition set $\mathbb{A}$ into two sets $\mathbb{A}_1$ and $\mathbb{A}_2$ of roughly equal size. +2. Draw one edge from each token in $\mathbb{A}_1$ to its most similar token in $\mathbb{A}_2$ . +3. Merged Tokens: Merge tokens by averaging their features if they are connected. +4. Residual Tokens: Tokens from set $\mathbb{A}_1$ representing maximum similarity with tokens in set $\mathbb{A}_2$ . + +The proposed Merging algorithm includes unconnected tokens from set $\mathbb{A}_2$ in the merged tokens output, whereas the Bipartite Matching algorithm [3] disregards these tokens. This feature reduces information loss within the overall SATA schema. As depicted in Figure 2(b), the output tokens of the proposed SATA module comprise the concatenation of tokens with spatial scores within the range of lower and upper bounds (i.e., set $\mathbb{B}$ ) and the Merged Tokens. This output is then fed into the FFN module, reducing unnecessary token processing in the FFN block and lowering the computational load and GFLOPs. Furthermore, Residual Tokens are concatenated with the FFN output to restore the original number of tokens, $N$ , forming the final output of the new ViT block, as shown on the right side of Figure 2(a). + +# 5. Experiment Results & Analysis + +# 5.1. Experimental setup + +Implementation Details All experiments were conducted on an NVIDIA V100 GPU with a $224 \times 224$ image resolution. The batch size was set to 256 for all experiments unless otherwise specified. We integrated the proposed SATA module + +into pre-trained generic vision transformers [9, 46] (DeiTiny/16, DeiT-Small/16, DeiT-Base/16, and Vit-Base/16), resulting in three model sizes named SATA-T, SATA-S, SATA-B, and SATA-B*, respectively. + +Evaluation Benchmarks We employ the ImageNet2012-1K [7] dataset for standard performance evaluation. For robustness assessment, we evaluate the proposed SATA in three dimensions: 1) Adversarial Robustness: Testing is conducted on adversarial examples generated by white-box attack algorithms FGSM [12] and PGD [31] using the ImageNet-1K validation set. ImageNet-A [8] (IN-A) includes the ImageNet objects in unusual contexts or orientations and is utilized to assess model performance against natural adversarial examples. 2) Common Corruption Robustness: We use ImageNet-C [20] (IN-C), which includes 15 types of algorithmically generated corruptions, each with five levels of severity. 3) Out-of-Distribution Robustness: Evaluation is performed on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK). Both datasets feature images with naturally occurring distribution shifts. ImageNet-R [22] (IN-R) contains abstract or rendered versions of the objects. ImageNet-Sketch [49] consists solely of sketch images, serving to test classification capability when texture or colour information is absent. + +# 5.2. Standard Performance Evaluation + +For standard performance evaluation, we compare our method with several state-of-the-art classification models, including Transformer-based models and representative CNN-based models, as shown in Table 1. Our proposed SATA significantly outperforms all other architectures, including both CNN-based and ViT-based models. Specifically, ViT models enhanced with SATA achieve new state-of-the-art top-1 accuracy of $86.5\%$ , $89.3\%$ , and $93.9\%$ for the tiny, small, and base versions, respectively, all without requiring additional training, input augmentation, or fine-tuning. Notably, integrating the proposed SATA into pre-trained ViT-Base/16 [9] (SATA- $\mathbf{B}^*$ ) results in an additional $1.0\%$ improvement. Furthermore, comparing the computation cost (GFLOPs) of the baseline DeiTs and SATA models demonstrates that the proposed spatial autocorrelation token analysis method also improves efficiency. + +# 5.3. Adversarial Robustness Evaluation + +For evaluating white-box attack adversarial robustness, we follow [32] and adopt the single-step attack algorithm FGSM [12] and the multi-step attack algorithm PGD [31] (with 5 steps and a step size of 0.5). Both attackers perturb the input image with a maximum magnitude of $\epsilon = 1$ . As shown in Table 1, adversarial robustness appears unrelated to standard performance. For instance, models like Swin [29], PVT [50], and TNT-S [17] achieve higher standard accuracy + +Table 1. Performance of SATA and several state-of-the-art (SOTA) CNNs and ViTs models on ImageNet and six robustness benchmarks: We report the mean corruption error (mCE) for ImageNet-C [20], where lower mCE values indicate higher model robustness. Our SATA models consistently outperform other counterparts in standard performance and enhance robustness across various model sizes compared to the baseline, all without requiring additional training or fine-tuning. SATA-B* refers to the integration of the proposed SATA module into the pre-trained vanilla ViT-Base/16 model [9]. + +
GroupModelFLOPs (G)Params (M)ImageNet-1K Top-1Top-5FGSMPGDRobustness Benchmarks
IN-C (mCE↓)IN-AIN-RIN-SK
CNNResNet50 [19]4.125.676.186.012.20.976.70.036.124.1
RegNetY-4GF[39]4.020.679.294.715.42.468.78.938.825.9
EfficientNet-B4[45]4.419.383.096.344.618.571.126.347.134.1
DeepAugment[21]4.125.675.892.727.19.553.63.946.732.6
ANT[41]4.125.676.193.017.83.163.01.139.026.3
Debiased CNN[27]4.125.676.993.420.45.567.53.540.828.4
ConvNeXt-B[30]15.48983.8---46.836.751.338.2
ViT-TinyDeiT-Ti[46]1.35.772.291.122.36.271.17.332.620.2
ConvViT-Ti[10]1.45.773.391.824.77.568.48.935.222.4
PiT-Ti[23]0.74.972.991.320.45.169.16.234.621.6
PVT-Tiny[50]1.913.275.092.510.00.579.67.933.921.5
RVT-Ti [32]1.38.678.494.234.811.758.213.343.730.0
FAN-T-ViT [57]1.37.079.2---57.5-42.5-
RVT-Ti+RSPC [16]1.310.979.2---55.716.5--
SATA-T (ours)1.05.786.598.240.010.951.114.647.325.2
ViT-SmallDeiT-S[46]4.622.179.995.040.716.754.618.942.229.4
ConvViT-S[10]5.427.881.595.841.717.249.824.545.433.1
PiT-S[23]2.923.580.995.341.016.552.521.743.630.8
PVT-Small[50]3.824.579.995.026.63.166.918.040.127.2
Swin-T[29]4.528.381.295.533.77.362.021.641.329.1
TNT-S[17]5.223.881.595.733.24.253.124.743.831.6
T2T-ViT_t-14[55]6.121.581.795.940.911.753.223.945.032.5
RVT-S [32]4.722.181.795.751.326.050.124.146.935.0
FAN-S-ViT [57]5.328.082.9---47.729.150.4-
RVT-S+RSPC [16]4.723.382.2---48.427.9--
SATA-S (ours)3.922.189.399.157.418.033.830.559.539.2
ViT-BaseDeiT-B[46]17.686.682.095.746.421.348.527.444.932.4
ConvViT-B[10]17.786.582.095.746.421.348.527.444.932.4
PiT-B[23]12.573.882.495.749.323.748.233.943.732.3
PVT-Large[50]9.861.481.795.933.17.359.826.642.730.2
Swin-B[29]15.487.883.496.449.221.354.435.846.632.4
T2T-ViT_t-24[55]15.064.182.696.146.717.548.428.947.935.4
RVT-B [32]17.786.282.596.052.327.447.327.748.235.8
FAN-B-ViT [57]10.454.083.6---44.435.451.8-
RVT-B+RSPC [16]17.791.882.8---45.732.1--
TORA-ViT-B/16(λ=0.1) [26]26.0111.284.1-48.423.331.746.557.6-
TransNeXt [43]18.489.784.8---43.550.653.941.4
SATA-B (ours)15.986.693.999.763.920.228.763.570.049.8
SATA-B* (ours)15.986.694.999.865.628.313.663.679.257.9
+ +than DeiTs corresponding, but their adversarial robustness is significantly lower, consistent with findings from [32, 42]. Our proposed SATA model achieves superior performance against both FGSM [12] and PGD [31] attacks. Specifically, SATA-S, SATA-B, and SATA-B* show over a $20\%$ improvement on FGSM [12] compared to previous ViT variants. + +Regarding natural adversarial robustness, the proposed SATA-T demonstrates a comparable performance of $14.6\%$ , which is on par with some current state-of-the-art methods like RVT-Ti [32] and RVT-Ti+RSPC [16], while being about half their size. However, for models of similar size (e.g., + +ViT-Small and ViT-Base), the proposed SATAs outperform others by about $50\%$ , indicating the effectiveness of SATA against natural adversarial attacks. + +# 5.4. Common Corruption Robustness Evaluation + +To measure model degradation on common image corruptions, we report the mean corruption error (mCE) on ImageNet-C [20] (IN-C) in Table 1. Our SATA method significantly reduces the mCE of DeiT-Ti [46] from $71.1\%$ to $51.1\%$ , achieving the lowest mCE among vision transformers within the ViT-Tiny group. For the other two larger + +![](images/ff8967be74804351126cc4974e2ad94160563efd56dff610c314463ddc8787c5.jpg) +(a) + +![](images/165bf6f2641482103d718545c0dc815d05d7541798f66d4dfef04bdebd18ab78.jpg) +(b) +Figure 4. (a) Ablation study on $\gamma$ . (b) Ablation study on $\alpha$ . We set $\gamma$ and $\alpha$ to 0.7 and 1.0, respectively, for all experiments throughout this paper. Ablation studies are conducted on the SATA-B model using the ImageNet-1K dataset. Dashed lines for both graphs represent the baseline Deit-Base/16 top-1 accuracy. + +ViT groups, the proposed SATA models achieve an mCE of approximately $28\%$ , improving by around 20 points over all other ViT or CNN-based methods on the leaderboard, thereby establishing a new state-of-the-art. This result also suggests that spatial autocorrelation management of visual tokens can successfully handle different types of image corruption. + +# 5.5. Out-of-Distribution Robustness Evaluation + +We evaluate the generalization ability of SATA on out-of-distribution data by reporting the top-1 accuracy on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK) in Table 1. The generic vision transformers [9, 46] enhanced by the proposed SATA consistently outperform other ViT models on ImageNet-R [22], achieving $47.3\%$ , $57.2\%$ , $70.0\%$ , $79.9\%$ in the ViT-Tiny, ViT-Small, and ViT-Base groups, respectively. Regarding ImageNet-Sketch [49] (IN-SK), SATA demonstrates superior performance compared to other models of similar size. These results imply that the spatial autocorrelation tokens analysis effectively captures feature distribution shifts, enhancing the model's out-of-distribution generalization capabilities. + +# 5.6. Ablation study + +Token Splitting and Token Grouping We evaluate the role of token splitting based on the upper and lower bounds of spatial autocorrelation scores and token grouping modules. To this end, we examine five SATA configurations, as depicted in Table 2. As shown in Table 2, utilizing only token grouping yields a top-1 accuracy of $84.4\%$ , improving upon the reference (DeiT-Base) accuracy by $2.4\%$ . Including either lower or upper bounds significantly enhances accuracy by about $10\%$ of the baseline top-1 accuracy, highlighting the effectiveness of the proposed spatial autocorrelation token splitting schema. The upper bound contributes slightly more, suggesting that tokens with extremely high spatial autocorrelation scores are more likely to be filtered by the splitting process. However, incorporating upper or lower bounds also increases computations and token counts, reducing throughput despite the substantial accuracy gains. Additionally, applying bounds does not guarantee an equal token distribution across bands, as this depends on the data distribution (shown in Figure 2 of the supplementary material). Finally, adding token grouping provides a further $1.6\%$ improvement over the splitting process alone. + +Regarding throughput, two main components of the proposed SATA framework (i.e. spatial autocorrelation-based token splitting and token grouping) primarily utilise parallelizable matrix multiplication and dot product operations. This design ensures computational efficiency with a negligible impact on inference time, as shown in Table 2. + +Threshold of starting block $(\gamma)$ We also examine the effect of parameter $\gamma$ , which controls at which transformer block the SATA module is applied. As Figure 4(a) shows, applying SATA from block $0.4 \times B$ onwards significantly improves model efficiency while exceeding baseline ViT accuracy $(82\%)$ . Applying SATA to earlier blocks $(\gamma < 0.4)$ degrades accuracy, suggesting that high spatial correlation within token features and the importance of all tokens in early layers is beneficial. To achieve a good trade-off between accuracy and efficiency, we use $\gamma = 0.7$ in all our experiments, unless stated otherwise. This results in a top-1 accuracy of $93.9\%$ and GFLOPs of 15.9. + +Lower/Upper bounds controlling factor $(\alpha)$ We further assess the influence of $\alpha$ , the factor controlling the lower and upper bounds in SATA. Figure 4(b) shows the performance of SATA-B on ImageNet-1K with $\alpha$ values ranging from 0.5 to 2. $\alpha$ determines the number of tokens passed to the FFN block and setting $\alpha = 1$ yields optimal performance. + +# 5.7. Visualisation and Discussion + +Although the effectiveness of the proposed SATA module has been empirically demonstrated, we conduct a deeper + +Table 2. Ablation study of the spatial autocorrelation module (token splitting) and token grouping. The symbols “√” and “X” indicate whether the corresponding element is employed with the configuration or not. † represents top-1 accuracy for the baseline Deit-Base/16 [46] model. + +
Spt. Auto. Correlation (Tokens Splitting)Tokens GroupingTop-1Throughput (ims/s)Latency (s)
Lower bound (μs - |s|)Upper bound (μs + |s|)
×××82.0†3030.0033
××✓84.43610.0027
×✓✓92.82970.0033
✓×✓92.52710.0035
✓✓×92.32680.0037
✓✓✓93.92760.0035
+ +![](images/765594f9b73860db62910a2eaae275672e63fde7965eea7a50fd8d7a29ed6ec3.jpg) +(a) + +![](images/1149a4d3cc9f7cdad93af8672b24fe378cbcfb79d7c9e7abe1df6afbbe05ef25.jpg) +(b) +Figure 5. (a) Cosine similarity between the clean and corrupted versions of the class token attention map and spatial autocorrelation scores across different blocks of SATA-B. Results are averaged across various types of image corruptions and severity levels on ImageNet-C [20]. (b) Visualisation of token splitting for a pair of clean and noisy images. Notably, the selected tokens for each set are similar for both clean and corrupted inputs. + +investigation to better understand its behaviour. To this end, we calculate the cosine similarity between the class token attention maps and spatial autocorrelation scores of clean and corrupted image pairs from the ImageNet-1K validation set and its corresponding ImageNet-C [20], respectively. We compute these similarities for various types of image corruptions and severity levels in ImageNet-C [20], and report the average values across different blocks of the proposed SATA-B in Figure 5(a). As shown in Figure 5(a), the cosine similarity between clean and corrupted attention maps + +drops significantly in the later blocks of the transformer. In contrast, the similarity for spatial autocorrelation scores improves at the early stages and remains consistently high, averaging above 0.8. This highlights that the proposed method can provide a more stable and reliable feature representation throughout the network, offering strong robustness against various types of corruption. + +Moreover, Figure 5 provides a visualization of tokens (patches) are split into set $\mathbb{A}$ and set $\mathbb{B}$ for a pair of clean and noisy images according to the proposed SATA algorithm. Notably, the similarity between corresponding sets for clean and noisy inputs is evident, further highlighting the robustness of the proposed method2. + +# 6. Conclusion + +In this paper, we introduce SATA, a novel method designed to enhance the performance and robustness of vision transformers against various types of corruption. SATA utilises a straightforward yet powerful spatial autocorrelation scheme to exploit spatial inter-dependencies among token features, significantly improving representational capacity and efficiency while reducing computational costs. Our experimental results demonstrate that SATA-enhanced vision transformers consistently deliver stable and reliable feature representations, achieving state-of-the-art performance on ImageNet-1K classification and setting new benchmarks for robustness across multiple evaluations, all without requiring additional training or fine-tuning. + +This work highlights SATA's transformative potential and opens several promising avenues for future research. Key directions include adapting SATA for window-based and hybrid ViT architectures to boost performance in tasks such as object detection and segmentation. Additionally, exploring SATA's application in other transformer-based domains, such as large language models (LLMs), could further extend its impact. + +# 7. Acknowledgements + +This work was supported in part by the Australian Research Council (ARC), Australia under Discovery Grant DP180100958 and Industrial Transformation Research Hub Grant IH180100002. + +# References + +[1] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? In NeurIPS, pages 26831-26843, 2021. 1 +[2] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In ICCV, pages 10231-10241, 2021. 1, 3 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3, 5 +[4] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. In NeurIPS, pages 33618-33632, 2022. 3 +[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2 +[6] Yanguang Chen. New approaches for calculating Moran's index of spatial autocorrelation. *PloS one*, 8(7), 2013. 3 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5 +[8] Minjing Dong, Yanxi Li, Yunhe Wang, and Chang Xu. Adversarily robust neural architectures. arXiv preprint arXiv:2009.00902, 2020. 2, 5 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5, 6, 7 +[10] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, pages 2286-2296. PMLR, 2021. 6 +[11] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, pages 396-414. Springer, 2022. 2 +[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5, 6 +[13] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, pages 12259-12269, 2021. 2 +[14] Jindong Gu, Volker Tresp, and Yao Qin. Are vision transformers robust to patch perturbations? In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3 + +[15] Yong Guo, David Stutz, and Bernt Schiele. Robustifying token attention for vision transformers. In ICCV, pages 17557-17568, 2023. 3 +[16] Yong Guo, David Stutz, and Bernt Schiele. Improving robustness of vision transformers by reducing sensitivity to patch corruptions. In CVPR, pages 4108-4118, 2023. 1, 3, 6 +[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, pages 15908-15919, 2021. 5, 6 +[18] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE TPAMI, 45(1):87-110, 2022. 1, 2 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 2, 5, 6, 8 +[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6 +[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 2, 5, 7 +[23] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, pages 11936-11945, 2021. 2, 6 +[24] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, pages 620-640. Springer, 2022. 2 +[25] Xuelong Li, Han Zhang, Rui Zhang, and Feiping Nie. Discriminative and uncorrelated feature selection with constrained spectral analysis in unsupervised learning. IEEE Transactions on Image Processing, 29:2139-2149, 2019. 1 +[26] Yanxi Li and Chang Xu. Trade-off between robustness and accuracy of vision transformers. In CVPR, pages 7558-7568, 2023. 3, 6 +[27] Yingwei Li, Qihang Yu, Mingxing Tan, Jieru Mei, Peng Tang, Wei Shen, Alan Yuille, and Cihang Xie. Shape-texture de-biased neural network training. In In Proceedings of the International Conference on Learning Representations, 2021. 6 +[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2 +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 6 + +[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 3, 6 +[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 5, 6 +[32] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In CVPR, pages 12042-12051, 2022. 1, 3, 5, 6 +[33] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 2 +[34] Patrick AP Moran. The interpretation of statistical maps. Journal of the Royal Statistical Society. Series B (Methodological), 10(2):243-251, 1948. 3 +[35] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, pages 23296-23308, 2021. 1, 3 +[36] Nick Nikzad, Yongsheng Gao, and Jun Zhou. CSA-Net: Channel-wise Spatially Autocorrelated Attention Networks. arXiv preprint arXiv:2405.05755, 2024. 1, 4 +[37] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In AAAI, pages 2071-2081, 2022. 1, 3 +[38] Yao Qin, Chiyuan Zhang, Ting Chen, Balaji Lakshminarayanan, Alex Beutel, and Xuezhi Wang. Understanding and improving robustness of vision transformers through patch-based negative augmentation. In NeurIPS, pages 16276-16289, 2022. 1, 3 +[39] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 6 +[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3 +[41] Evgenia Rusak, Lukas Schott, Roland S Zimmermann, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge, and Wieland Brendel. A simple way to make neural networks robust against diverse image corruptions. In ECCV, pages 53-69. Springer, 2020. 6 +[42] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 6 +[43] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In CVPR, pages 17773-17783, 2024. 3, 6 +[44] Zhuoran Song, Yihong Xu, Zhezhi He, Li Jiang, Naifeng Jing, and Xiaoyao Liang. Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570, 2022. 2 +[45] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6 +[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training + +data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6, 7, 8 +[47] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021. 3 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 3 +[49] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5, 7 +[50] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 5, 6 +[51] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, pages 22–31, 2021. 2 +[52] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, pages 28522-28535, 2021. 2 +[53] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021. 2 +[54] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In ICCV, pages 579-588, 2021. 2 +[55] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 2, 6 +[56] Qinglong Zhang and Yu-Bin Yang. Rest: An efficient transformer for visual recognition. In NeurIPS, pages 15475-15485, 2021. 2 +[57] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In ICML, pages 27378-27394. PMLR, 2022. 1, 3, 6 \ No newline at end of file diff --git a/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/images.zip b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..172a345db2131f7ddd84efba476cb8d9717e6634 --- /dev/null +++ b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8e98159d0a990d156b04c6607ad6424a0f8be8b11a77363547680c434a679b3 +size 541526 diff --git a/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/layout.json b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb90713d1466fb9a07d5ef309c04d0e6e83a6dee --- /dev/null +++ b/2025/SATA_ Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers/layout.json @@ -0,0 +1,8730 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "type": "text", + "content": "SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 185, + 161, + 424, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 161, + 424, + 175 + ], + "spans": [ + { + "bbox": [ + 185, + 161, + 424, + 175 + ], + "type": "text", + "content": "Nick Nikzad, Yi Liao, Yongsheng Gao, Jun Zhou" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 176, + 406, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 176, + 406, + 189 + ], + "spans": [ + { + "bbox": [ + 204, + 176, + 406, + 189 + ], + "type": "text", + "content": "Griffith University, Queensland, Australia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 191, + 480, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 191, + 480, + 203 + ], + "spans": [ + { + "bbox": [ + 128, + 191, + 480, + 203 + ], + "type": "text", + "content": "{n.nikzaddehaji, y.liao, yongsheng.gao, jun.zhou}@griffith.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 258, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 579 + ], + "type": "text", + "content": "Over the past few years, vision transformers (ViTs) have consistently demonstrated remarkable performance across various visual recognition tasks. However, attempts to enhance their robustness have yielded limited success, mainly focusing on different training strategies, input patch augmentation, or network structural enhancements. These approaches often involve extensive training and fine-tuning, which are time-consuming and resource-intensive. To tackle these obstacles, we introduce a novel approach named Spatial Autocorrelation Token Analysis (SATA). By harnessing spatial relationships between token features, SATA enhances both the representational capacity and robustness of ViT models. This is achieved through the analysis and grouping of tokens according to their spatial autocorrelation scores prior to their input into the Feed-Forward Network (FFN) block of the self-attention mechanism. Importantly, SATA seamlessly integrates into existing pre-trained ViT baselines without requiring retraining or additional fine-tuning, while concurrently improving efficiency by reducing the computational load of the FFN units. Experimental results show that the baseline ViTs enhanced with SATA not only achieve a new state-of-the-art top-1 accuracy on ImageNet-1K image classification " + }, + { + "bbox": [ + 56, + 258, + 296, + 579 + ], + "type": "inline_equation", + "content": "(94.9\\%)" + }, + { + "bbox": [ + 56, + 258, + 296, + 579 + ], + "type": "text", + "content": " but also establish new state-of-the-art performance across multiple robustness benchmarks, including ImageNet-A (top-1=63.6%), ImageNet-R (top-1=79.2%), and ImageNet-C (mCE=13.6%), all without requiring additional training or fine-tuning of baseline models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 580, + 256, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 256, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 256, + 592 + ], + "type": "text", + "content": "Availability: https://github.com/nick-nikzad/SATA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In recent years, vision transformers (ViTs) have demonstrated exceptional performance across diverse computer vision applications [9, 18]. Drawing inspiration from the significant achievements of transformer architectures in natural language processing (NLP), ViTs divide an input image into a sequence of patches (tokens) and leverage self-attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 447 + ], + "type": "text", + "content": "layers [48] to capture relationships between these tokens, ultimately generating rich representations for visual recognition tasks. While recent studies indicate that ViTs can exhibit greater robustness than Convolutional Networks (ConvNets), attributed to their self-attention mechanism [1, 2, 35, 37]. However, this hypothesis has been challenged. Liu et al. [30] demonstrated that a carefully constructed ConvNet can surpass ViTs in both generalization and robustness. Furthermore, while techniques such as patch augmentation [32, 38], contrastive learning strategies [16, 38], and network adjustments [32, 57] have shown promise in enhancing ViT performance and robustness, two primary limitations remain: 1) these methods require extensive retraining or fine-tuning on large datasets (e.g., ImageNet-1K, ImageNet-21K), a labour-intensive and resource-demanding process, especially for large-scale ViT architectures; and 2) the attention maps generated by the self-attention mechanism are highly sensitive to noise, which degrades robustness when inputs are corrupted." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 591 + ], + "type": "text", + "content": "Recently, Nikzad et al. [36] showed the existence of spatial correlation among feature maps in convolutional neural networks (CNNs). Moreover, they observed a decrease in spatial autocorrelation among feature maps through deeper network layers, suggesting that final features exhibit reduced spatial dependency. Motivated by these findings, in this work, we first investigate spatial autocorrelation within Vision Transformer (ViT) architectures and its implications for their performance and robustness. Then, we present a novel approach named \"Spatial Autocorrelation Token Analysis\" (SATA) to tackle the identified shortcomings in the current efforts to enhance ViT robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "Our analysis shows that spatial autocorrelation among visual patches (tokens) diminishes through ViT networks, similar to CNNs [36], aligning with findings that spatially uncorrelated features enhance recognition performance [25]. SATA integrates seamlessly into pre-trained ViT models without retraining, boosting robustness and efficiency by reducing the computational load on FFN units. These enhancements in feature aggregation and resilience to corruption account for SATA's exceptional performance from two perspectives:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9730" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 71, + 462, + 219 + ], + "blocks": [ + { + "bbox": [ + 143, + 71, + 462, + 219 + ], + "lines": [ + { + "bbox": [ + 143, + 71, + 462, + 219 + ], + "spans": [ + { + "bbox": [ + 143, + 71, + 462, + 219 + ], + "type": "image", + "image_path": "2d5fe590560b1ece0e7de5a5de5ca3f7497361aed0fddfc7abf79c14262a6cd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 276 + ], + "type": "text", + "content": "Figure 1. Visual comparison of class token attention maps and spatial autocorrelation score maps across three layers of Deit-Base/16 pre-trained on ImageNet-1K. The clean image is sourced from the 'goldfinch' class of the ImageNet-1K dataset, and its corresponding corrupted version, with maximum severity (5), is sourced from the ImageNet-C [20] dataset. " + }, + { + "bbox": [ + 54, + 230, + 555, + 276 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 230, + 555, + 276 + ], + "type": "text", + "content": " represents the cosine similarity between either the attention maps or the spatial autocorrelation score maps of a corrupted image and its corresponding clean version at each block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 297, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 297, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 297, + 460 + ], + "type": "text", + "content": "- Recognition performance: Our study shows that in the later layers of ViT networks, patches with extremely high or low spatial autocorrelation scores in non-informative regions can impede recognition performance and compromise the network's robustness against corrupted inputs. Our proposed SATA method adopts a unique splitting and grouping algorithm to address this issue based on tokens' spatial relation scores in the later layers. This approach prevents the input of unnecessary tokens into the FFN block of the self-attention mechanism, resulting in richer feature aggregation. Unlike token pruning and merging techniques, SATA concatenates bypassed tokens (Residual tokens) with the output of the FFN module, restoring the original number of tokens and preventing information loss for later blocks of the ViT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 605 + ], + "type": "text", + "content": "- Robustness performance: As illustrated in Figure 1, our study shows that in the layers of ViT networks, the spatial autocorrelation scores of patches are more robust against different corruptions than their attention maps. Specifically, while the cosine similarity between clean and corrupted attention maps drops significantly in the blocks of the Vision Transformer, the similarity for spatial autocorrelation scores remains consistently high. This highlights that our spatial autocorrelation analysis provides a more stable and reliable schema to achieve corruption-resistant feature representation throughout the network, offering strong robustness against various types of corruption." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "Extensive experiments conducted on ImageNet-1K image classification and various robustness evaluation benchmarks demonstrate the effectiveness of the proposed spatial autocorrelation paradigm in significantly improving the robustness and accuracy performance of Vision Transformers (ViTs). These findings establish a new state-of-the-art performance level, achieving a top-1 accuracy of " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "94.9\\%" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": " on ImageNet-1K image classification, as well as impressive results across multiple robustness benchmarks, including ImageNet-A [8] (top" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 281, + 556, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 354 + ], + "type": "inline_equation", + "content": "1 = 63.6\\%" + }, + { + "bbox": [ + 313, + 281, + 556, + 354 + ], + "type": "text", + "content": " ), ImageNet-R [22] (top-1=79.2%), and ImageNet-C [20] (mCE=13.6%), without requiring additional expensive fine-tuning or training. Furthermore, in-depth investigations are conducted to thoroughly explore the characteristics of the proposed Spatial Autocorrelation Token Analysis (SATA)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 365, + 406, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 406, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 406, + 377 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 384, + 429, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 429, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 429, + 396 + ], + "type": "text", + "content": "2.1. Vision Transformer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 641 + ], + "type": "text", + "content": "Since the introduction of Vision Transformers (ViTs), they have achieved remarkable success in various computer vision tasks [9, 18]. Most improvements to date have focused on enhancing either the accuracy or the efficiency of ViTs. Numerous ViT variants have been proposed to boost their performance [18]. Through dedicated data augmentation [46] and advanced self-attention structures [13, 53], ViTs have demonstrated competitive or superior performance compared to convolutional neural networks (CNNs). Hybrid models like CvT [51] introduce intrinsic inductive bias into the ViT architecture by adding additional convolutional layers before the multi-head self-attention (MHSA) modules. CeiT [54] extracts low-level features through the Image-to-Token (I2T) module and enhances locality by replacing the standard feedforward network with the locally enhanced feed-forward (LeFF) layer. To enable ViTs to learn multi-scale features, CrossViT [5] employs a dual-branch transformer that combines different sizes of image patches to produce stronger image features. ViTAE [52] incorporates multi-scale context by designing reduction cells (RC) and normal cells (NC)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "To create efficient Vision Transformers, several recent works have focused on pruning [11, 33, 44] or combining [24, 28] tokens. ResT [56] introduces an efficient self-attention module using overlapping depth-wise convolutions, while T2T-ViT [55] employs a Tokens-to-Token (T2T) module for token aggregation. PiT [23] reduces spatial size with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9731" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "pooling layers, and Dynamic-ViT [40] dynamically prunes tokens during inference. CaiT [47] optimizes the ViT architecture with layer scaling and class-attention mechanisms. More recently, Bolya et al.[3] proposed a simple token merging technique that potentially does not require retraining." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 168, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 168, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 168, + 152 + ], + "type": "text", + "content": "2.2. Robustness of ViTs" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 296, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 241 + ], + "type": "text", + "content": "While several Recent research has yielded mixed results on the robustness of Vision Transformers (ViTs) compared to Convolutional Neural Networks (CNNs). While some studies [2, 14, 35, 37] suggest ViTs are more robust against various perturbations and distribution shifts, Liu et al. [30] challenge this notion by demonstrating that a well-designed CNN can outperform ViTs in generalization and robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 361 + ], + "type": "text", + "content": "To enhance ViT robustness, various methods have been proposed, including network structural adjustments, patch augmentation, and diverse training strategies [4, 14-16, 26, 32, 38, 57]. For instance, Robust Vision Transformer (RVT) [32] introduces a convolutional stem and token pooling to improve robustness, while Full Attention Net (FAN) [57] leverages an attentional channel processing design. RobustViT [4] downplays the influence of image backgrounds, and a method proposed in [14] uses temperature scaling to smooth attention weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 363, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 602 + ], + "type": "text", + "content": "Additionally, Qin et al.[38] improve the robustness of ViTs by using images transformed with patch-based operations as negative augmentation. Li et al. [26] propose TORA-ViT, which consists of an accuracy adapter, a robustness adapter, and a gated fusion module. The accuracy adapter extracts predictive features, while the robustness adapter extracts robust features. These features are then combined by the gated fusion module. Reducing Sensitivity to Patch Corruptions (RSPC) [16] enhances the robustness of ViTs through a specialized training strategy. In [15], the Attention Diversification Loss (ADL) is introduced to encourage output tokens to aggregate information from a diverse set of input tokens. Recently, Shi et al. [43] introduced learnable tokens and employed pixel-focused attention to capture structured information between tokens. However, most of these approaches require extensive training or fine-tuning and often sacrifice performance for efficiency [3]. In contrast, our method can be applied to baseline vision transformers[9, 46] without requiring additional training and without any performance drop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 615, + 141, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 141, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 141, + 627 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "text", + "content": "3.1. Vision Transformers: Multi-head Self Attention (MHSA)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "A standard ViT [9] partitions an input image into " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " patches (tokens). These patches are then transformed to generate a token embedding tensor " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". These tokens are then processed through a stack of transformer blocks, as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "illustrated in Figure 2(a). ViTs leverage self-attention [48] to aggregate global information. Given the input token embedding tensor " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", self-attention applies linear transformations with parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "W_{K}, W_{Q}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " to embed them into the key " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "K = W_{K}\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", query " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "Q = W_{Q}\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "V = W_{V}\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", respectively. Self-attention utilises " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " to generate a pairwise attention map " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{att} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and then aggregates the token features using the attention map " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{att}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 385, + 183, + 555, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 183, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 385, + 183, + 555, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {a t t} = \\operatorname {S o f t m a x} \\left(Q K ^ {t} / \\sqrt {d}\\right), \\tag {1}", + "image_path": "9b3d9f464320403e01cff485b29a43c6a3d4b8a1a26c1fec9439c12a4d2a3f8a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 200, + 555, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 200, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 361, + 200, + 555, + 213 + ], + "type": "interline_equation", + "content": "\\operatorname {S e l f - A t t e n t i o n} (Q, K, V) = \\mathbf {M} _ {\\text {a t t}} V, \\tag {2}", + "image_path": "65c97c2460ad4201708edbd5a88a207af636e0c9b4f054c14cdfcc8eb74487f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "text", + "content": "where the symbol “" + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "text", + "content": "” indicates the transpose of the matrix. To achieve rich feature hierarchies, the Transformer block employs multiple self-attention heads. Specifically, " + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "text", + "content": " heads are stacked in parallel, resulting in an output of " + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "inline_equation", + "content": "N \\times h \\times d" + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "text", + "content": ". These concatenated features are then processed by a feedforward network (FFN) for further transformation. Finally, the FFN output of " + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 313, + 217, + 556, + 324 + ], + "type": "text", + "content": " serves as the final output of the Multi-Head Self-Attention (MHSA) block within the Transformer architecture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 517, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 517, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 517, + 342 + ], + "type": "text", + "content": "3.2. Geographical Spatial Auto-correlation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 348, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 556, + 538 + ], + "type": "text", + "content": "In geographical modelling, spatial autocorrelation plays a crucial role in assessing the spatial interdependence of entities based on their locations and values. Positive spatial autocorrelation indicates that neighbouring observations share similar values, while negative spatial autocorrelation suggests that nearby observations tend to have contrasting values. Typically, two types of measures are used: global measures, which provide an overall assessment of spatial autocorrelation across all data points, and local measures, which offer insights into the spatial autocorrelation of individual locations relative to their neighbourhoods. Moran's metric [6, 34] is commonly employed in geographical analysis to compute such measurements. In this study, we employ Moran's measurement for the first time, to the best of our knowledge, to investigate spatial dependency among vision transformers' tokens (patches)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": " be a set of " + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": " observations (here, tokens) presented by embedded vectors " + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\mathbf{x}_1,\\mathbf{x}_2,\\dots,\\mathbf{x}_N]" + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": ", and an associated attribute, " + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = [a_1,a_2,\\dots,a_N]" + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": ", the local Moran's I metric can be defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 382, + 591, + 555, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 591, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 382, + 591, + 555, + 605 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {l} = \\left[ \\operatorname {d i a g} \\left(\\mathbf {z z} ^ {t} \\mathbf {W}\\right) \\right] _ {N \\times 1}, \\tag {3}", + "image_path": "cdc7f14f404e821b356b3eb2375bd9330015f6d043fecd2ad26c6837665fba9b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(.)" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": " returns the diagonal elements of a matrix. Symbol \"t\" indicates the transpose of the matrix. " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = [w_{ij}]_{N\\times N}" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": " represents spatial weight matrix, in which " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": " denotes the degree of closeness or the contiguous relationships between " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": " and can be computed using a dot product similarity " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i.\\mathbf{x}_j^t i,j = 1,2,\\dots,N)" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 609, + 556, + 692 + ], + "type": "text", + "content": " refers to normalised token-wise attribute values as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 408, + 695, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 695, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 408, + 695, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\frac {\\mathbf {a} - \\mu}{\\sigma}, \\tag {4}", + "image_path": "df0d79072c983b7dc94ddeb0cc023f65abf491c3c344111c090fdd8f6d855b9b.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9732" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 213 + ], + "type": "image", + "image_path": "0b8c83390f656fadadb70704ce1ae9d2c4be6e5819ee21ebb34b18a77254b422.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 216, + 123, + 224 + ], + "lines": [ + { + "bbox": [ + 112, + 216, + 123, + 224 + ], + "spans": [ + { + "bbox": [ + 112, + 216, + 123, + 224 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 215, + 390, + 224 + ], + "lines": [ + { + "bbox": [ + 378, + 215, + 390, + 224 + ], + "spans": [ + { + "bbox": [ + 378, + 215, + 390, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 180, + 225, + 430, + 327 + ], + "blocks": [ + { + "bbox": [ + 180, + 225, + 430, + 327 + ], + "lines": [ + { + "bbox": [ + 180, + 225, + 430, + 327 + ], + "spans": [ + { + "bbox": [ + 180, + 225, + 430, + 327 + ], + "type": "image", + "image_path": "1bf9a59b62008c79d08a6eab62fbce5ab9b6cb713ac66f86583394ac664b22c7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 329, + 310, + 338 + ], + "lines": [ + { + "bbox": [ + 299, + 329, + 310, + 338 + ], + "spans": [ + { + "bbox": [ + 299, + 329, + 310, + 338 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "text", + "content": " denote mean and standard deviation of a, respectively. The final local spatial autocorrelation descriptor, s, can be defined as normalised " + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 54, + 383, + 297, + 419 + ], + "type": "text", + "content": " [36]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 427, + 296, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 427, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 427, + 296, + 453 + ], + "type": "interline_equation", + "content": "\\mathbf {s} = \\frac {\\boldsymbol {I} _ {l} - \\mu_ {\\boldsymbol {I} _ {l}}}{\\sigma_ {\\boldsymbol {I} _ {l}}}, \\tag {5}", + "image_path": "43ad44d911fb5d00f7e82787fb463285fc636daf8d6e120622e4a03ddfb948f1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mu_{I_l}" + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\sigma_{I_l}" + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": " indicate mean and standard deviation of " + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": ", respectively. Following [36], given a token embedding tensor " + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\mathbf{x}_1,\\mathbf{x}_2,\\dots,\\mathbf{x}_N]\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": ", its token-wise global context attribute " + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = [a_1,a_2,\\dots,a_N]\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 55, + 462, + 296, + 521 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 529, + 296, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 529, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 113, + 529, + 296, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {a} = \\left[ a _ {i} = \\frac {1}{d} \\sum_ {t = 1} ^ {d} \\mathbf {x} _ {i} (t) \\right] _ {N \\times 1}, \\tag {6}", + "image_path": "01c534cdf874ea55adbb5f0f3c51680fb09d176ef684636885da2f463a7c3acb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "text", + "content": " denotes the spatial dimension of the tokens. " + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i(t)" + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "text", + "content": "-th token value at position " + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 573, + 296, + 645 + ], + "type": "text", + "content": ". It's worth noting that more advanced strategies or application-specific criteria can be employed to derive the global contextual information descriptor (Eq.(6)). In this context, we adopt the same approach as [36] for the sake of simplicity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 657, + 272, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 272, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 272, + 671 + ], + "type": "text", + "content": "4. Spatial Autocorrelation Token Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "content": "Figure 2(b) illustrates the overall architecture of the proposed spatial autocorrelation token analysis module, situated between attention and FFN units within a standard ViT block," + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 382, + 553, + 437 + ], + "blocks": [ + { + "bbox": [ + 54, + 347, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 54, + 347, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 555, + 371 + ], + "type": "text", + "content": "Figure 2. (a) Comparison between conventional ViT block and the augmented ViT with SATA (b) Overall architecture of the proposed SATA module. (c) Tokens Merging." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 437 + ], + "type": "image", + "image_path": "026e7ef143c9f562ecf37face211d3e544281d1f7ebd4a064be62ea1d499ac3b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "type": "text", + "content": "Figure 3. Plotting the variations of " + }, + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "type": "inline_equation", + "content": "|\\hat{\\mathbf{s}}|" + }, + { + "bbox": [ + 313, + 446, + 555, + 468 + ], + "type": "text", + "content": ", and the lower and upper bounds across different blocks of the ViT." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "text", + "content": "as depicted in Figure 2(a). To initiate our spatial autocorrelation token analysis, we begin with the observation of the spatial autocorrelation scores, s (Eq. (5)), for ViT's token embedding tensors over different blocks. As transformers inherently capture pairwise closeness relationships between tokens by computing the attention map " + }, + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{att}" + }, + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "text", + "content": " in Eq.(2), we can directly set " + }, + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{W} = \\mathbf{M}_{att}" + }, + { + "bbox": [ + 313, + 477, + 555, + 573 + ], + "type": "text", + "content": " in Eq.(3) to enhance both efficiency and effectiveness1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "text", + "content": "Figure 3 illustrates the alterations in the mean " + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "inline_equation", + "content": "(\\mu_{\\mathrm{s}})" + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "text", + "content": " and the absolute value of the median " + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "inline_equation", + "content": "(|\\hat{\\mathbf{s}}|)" + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "text", + "content": " statistics of the spatial autocorrelation (s) across different blocks of Deit-Base/16 on the ImageNet-1K validation set. It can be seen that for the later layers, specifically starting from block six, tokens tend to exhibit lower " + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "inline_equation", + "content": "|\\hat{\\mathbf{s}}|" + }, + { + "bbox": [ + 313, + 574, + 556, + 682 + ], + "type": "text", + "content": " values. Drawing from the aforementioned observation, we encapsulate the proposed analysis into two sequential steps to handle these tokens to improve the ViT's robustness and performance:" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "text", + "content": "1The implementation details of the proposed SATA are provided in Supplementary material" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9733" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 151, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 151, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 151, + 85 + ], + "type": "text", + "content": "4.1. Token Splitting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "content": "Based on the above findings, the overall of the proposed SATA method is illustrated in Figure 2 (b). As shown in Figure 3, we limit token processing to the latter stages of the transformer. Specifically, tokens in layers from " + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "inline_equation", + "content": "\\gamma \\times B" + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "content": " onward (" + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "inline_equation", + "content": "\\gamma > 0" + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "content": " represents the depth, or number of blocks, of the transformer) are partitioned into two sets, " + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 55, + 89, + 297, + 174 + ], + "type": "text", + "content": ", using the spatial autocorrelation scores s as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 191, + 296, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 191, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 191, + 296, + 205 + ], + "type": "interline_equation", + "content": "\\mathbb {A} = \\left\\{\\mathbf {x} _ {i}; s _ {i} < \\alpha \\left(\\mu_ {\\mathbf {s}} - | \\hat {\\mathbf {s}} |\\right) \\text {a n d} s _ {i} > \\alpha \\left(\\mu_ {\\mathbf {s}} + | \\hat {\\mathbf {s}} |\\right) \\right\\} \\tag {7}", + "image_path": "51f2f6dc438fc704719e2f9a21803229d86ae8487eb8e3d3392f005ad4051326.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 206, + 296, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 206, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 69, + 206, + 296, + 219 + ], + "type": "interline_equation", + "content": "\\mathbb {B} = \\left\\{\\mathbf {x} _ {j}; \\alpha \\left(\\mu_ {\\mathbf {s}} - | \\hat {\\mathbf {s}} |\\right) < = s _ {j} < = \\alpha \\left(\\mu_ {\\mathbf {s}} + | \\hat {\\mathbf {s}} |\\right) \\right\\} \\tag {8}", + "image_path": "f2810304814b18833274d4d5b568151d2b0a8d75300b9ec368b6d2cd69e91023.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha (\\mu_{\\mathbf{s}} - |\\hat{\\mathbf{s}} |)" + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha (\\mu_{\\mathbf{s}} + |\\hat{\\mathbf{s}} |)" + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "content": " represent lower and upper bounds, respectively. " + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "content": " denotes the controlling factor (parameters " + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 224, + 296, + 262 + ], + "type": "text", + "content": " choices are discussed in Section 5.6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 156, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 156, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 156, + 280 + ], + "type": "text", + "content": "4.2. Token Grouping" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 356 + ], + "type": "text", + "content": "To manage tokens falling beyond the lower and upper bounds (i.e. " + }, + { + "bbox": [ + 55, + 284, + 296, + 356 + ], + "type": "inline_equation", + "content": "\\alpha (\\mu_{\\mathrm{s}}\\pm |\\hat{\\mathbf{s}} |))" + }, + { + "bbox": [ + 55, + 284, + 296, + 356 + ], + "type": "text", + "content": " , inspired by the Bipartite Matching algorithm [3] we employ a unique Tokens Merging algorithm to efficiently match and merge similar tokens in set A. In particular, the proposed Tokens Merging algorithm can be summarized as follows (illustrated in Figure 2(c)):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 357, + 296, + 452 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "text", + "content": "1. Partition set " + }, + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "text", + "content": " into two sets " + }, + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_1" + }, + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_2" + }, + { + "bbox": [ + 56, + 357, + 295, + 378 + ], + "type": "text", + "content": " of roughly equal size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": "2. Draw one edge from each token in " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_1" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": " to its most similar token in " + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_2" + }, + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 404, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 427 + ], + "type": "text", + "content": "3. Merged Tokens: Merge tokens by averaging their features if they are connected." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "type": "text", + "content": "4. Residual Tokens: Tokens from set " + }, + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_1" + }, + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "type": "text", + "content": " representing maximum similarity with tokens in set " + }, + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_2" + }, + { + "bbox": [ + 55, + 428, + 296, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "text", + "content": "The proposed Merging algorithm includes unconnected tokens from set " + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_2" + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "text", + "content": " in the merged tokens output, whereas the Bipartite Matching algorithm [3] disregards these tokens. This feature reduces information loss within the overall SATA schema. As depicted in Figure 2(b), the output tokens of the proposed SATA module comprise the concatenation of tokens with spatial scores within the range of lower and upper bounds (i.e., set " + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "text", + "content": ") and the Merged Tokens. This output is then fed into the FFN module, reducing unnecessary token processing in the FFN block and lowering the computational load and GFLOPs. Furthermore, Residual Tokens are concatenated with the FFN output to restore the original number of tokens, " + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 452, + 297, + 620 + ], + "type": "text", + "content": ", forming the final output of the new ViT block, as shown on the right side of Figure 2(a)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 628, + 231, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 231, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 231, + 642 + ], + "type": "text", + "content": "5. Experiment Results & Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 648, + 170, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 170, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 170, + 661 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "Implementation Details All experiments were conducted on an NVIDIA V100 GPU with a " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": " image resolution. The batch size was set to 256 for all experiments unless otherwise specified. We integrated the proposed SATA module" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "type": "text", + "content": "into pre-trained generic vision transformers [9, 46] (DeiTiny/16, DeiT-Small/16, DeiT-Base/16, and Vit-Base/16), resulting in three model sizes named SATA-T, SATA-S, SATA-B, and SATA-B*, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 132, + 556, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 556, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 556, + 372 + ], + "type": "text", + "content": "Evaluation Benchmarks We employ the ImageNet2012-1K [7] dataset for standard performance evaluation. For robustness assessment, we evaluate the proposed SATA in three dimensions: 1) Adversarial Robustness: Testing is conducted on adversarial examples generated by white-box attack algorithms FGSM [12] and PGD [31] using the ImageNet-1K validation set. ImageNet-A [8] (IN-A) includes the ImageNet objects in unusual contexts or orientations and is utilized to assess model performance against natural adversarial examples. 2) Common Corruption Robustness: We use ImageNet-C [20] (IN-C), which includes 15 types of algorithmically generated corruptions, each with five levels of severity. 3) Out-of-Distribution Robustness: Evaluation is performed on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK). Both datasets feature images with naturally occurring distribution shifts. ImageNet-R [22] (IN-R) contains abstract or rendered versions of the objects. ImageNet-Sketch [49] consists solely of sketch images, serving to test classification capability when texture or colour information is absent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 380, + 499, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 499, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 499, + 392 + ], + "type": "text", + "content": "5.2. Standard Performance Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "content": "For standard performance evaluation, we compare our method with several state-of-the-art classification models, including Transformer-based models and representative CNN-based models, as shown in Table 1. Our proposed SATA significantly outperforms all other architectures, including both CNN-based and ViT-based models. Specifically, ViT models enhanced with SATA achieve new state-of-the-art top-1 accuracy of " + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "inline_equation", + "content": "86.5\\%" + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "inline_equation", + "content": "89.3\\%" + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "inline_equation", + "content": "93.9\\%" + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "content": " for the tiny, small, and base versions, respectively, all without requiring additional training, input augmentation, or fine-tuning. Notably, integrating the proposed SATA into pre-trained ViT-Base/16 [9] (SATA-" + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^*" + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "content": ") results in an additional " + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 399, + 556, + 591 + ], + "type": "text", + "content": " improvement. Furthermore, comparing the computation cost (GFLOPs) of the baseline DeiTs and SATA models demonstrates that the proposed spatial autocorrelation token analysis method also improves efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 600, + 502, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 502, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 502, + 612 + ], + "type": "text", + "content": "5.3. Adversarial Robustness Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "For evaluating white-box attack adversarial robustness, we follow [32] and adopt the single-step attack algorithm FGSM [12] and the multi-step attack algorithm PGD [31] (with 5 steps and a step size of 0.5). Both attackers perturb the input image with a maximum magnitude of " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 1" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ". As shown in Table 1, adversarial robustness appears unrelated to standard performance. For instance, models like Swin [29], PVT [50], and TNT-S [17] achieve higher standard accuracy" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9734" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 132, + 555, + 564 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Performance of SATA and several state-of-the-art (SOTA) CNNs and ViTs models on ImageNet and six robustness benchmarks: We report the mean corruption error (mCE) for ImageNet-C [20], where lower mCE values indicate higher model robustness. Our SATA models consistently outperform other counterparts in standard performance and enhance robustness across various model sizes compared to the baseline, all without requiring additional training or fine-tuning. SATA-B* refers to the integration of the proposed SATA module into the pre-trained vanilla ViT-Base/16 model [9]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 132, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 56, + 132, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 555, + 564 + ], + "type": "table", + "html": "
GroupModelFLOPs (G)Params (M)ImageNet-1K Top-1Top-5FGSMPGDRobustness Benchmarks
IN-C (mCE↓)IN-AIN-RIN-SK
CNNResNet50 [19]4.125.676.186.012.20.976.70.036.124.1
RegNetY-4GF[39]4.020.679.294.715.42.468.78.938.825.9
EfficientNet-B4[45]4.419.383.096.344.618.571.126.347.134.1
DeepAugment[21]4.125.675.892.727.19.553.63.946.732.6
ANT[41]4.125.676.193.017.83.163.01.139.026.3
Debiased CNN[27]4.125.676.993.420.45.567.53.540.828.4
ConvNeXt-B[30]15.48983.8---46.836.751.338.2
ViT-TinyDeiT-Ti[46]1.35.772.291.122.36.271.17.332.620.2
ConvViT-Ti[10]1.45.773.391.824.77.568.48.935.222.4
PiT-Ti[23]0.74.972.991.320.45.169.16.234.621.6
PVT-Tiny[50]1.913.275.092.510.00.579.67.933.921.5
RVT-Ti [32]1.38.678.494.234.811.758.213.343.730.0
FAN-T-ViT [57]1.37.079.2---57.5-42.5-
RVT-Ti+RSPC [16]1.310.979.2---55.716.5--
SATA-T (ours)1.05.786.598.240.010.951.114.647.325.2
ViT-SmallDeiT-S[46]4.622.179.995.040.716.754.618.942.229.4
ConvViT-S[10]5.427.881.595.841.717.249.824.545.433.1
PiT-S[23]2.923.580.995.341.016.552.521.743.630.8
PVT-Small[50]3.824.579.995.026.63.166.918.040.127.2
Swin-T[29]4.528.381.295.533.77.362.021.641.329.1
TNT-S[17]5.223.881.595.733.24.253.124.743.831.6
T2T-ViT_t-14[55]6.121.581.795.940.911.753.223.945.032.5
RVT-S [32]4.722.181.795.751.326.050.124.146.935.0
FAN-S-ViT [57]5.328.082.9---47.729.150.4-
RVT-S+RSPC [16]4.723.382.2---48.427.9--
SATA-S (ours)3.922.189.399.157.418.033.830.559.539.2
ViT-BaseDeiT-B[46]17.686.682.095.746.421.348.527.444.932.4
ConvViT-B[10]17.786.582.095.746.421.348.527.444.932.4
PiT-B[23]12.573.882.495.749.323.748.233.943.732.3
PVT-Large[50]9.861.481.795.933.17.359.826.642.730.2
Swin-B[29]15.487.883.496.449.221.354.435.846.632.4
T2T-ViT_t-24[55]15.064.182.696.146.717.548.428.947.935.4
RVT-B [32]17.786.282.596.052.327.447.327.748.235.8
FAN-B-ViT [57]10.454.083.6---44.435.451.8-
RVT-B+RSPC [16]17.791.882.8---45.732.1--
TORA-ViT-B/16(λ=0.1) [26]26.0111.284.1-48.423.331.746.557.6-
TransNeXt [43]18.489.784.8---43.550.653.941.4
SATA-B (ours)15.986.693.999.763.920.228.763.570.049.8
SATA-B* (ours)15.986.694.999.865.628.313.663.679.257.9
", + "image_path": "0192da7a3922d3bb99ea14f90fc61c5aed098dbe9f6693ca39b65c9fe397c4f6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 572, + 297, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 297, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 297, + 644 + ], + "type": "text", + "content": "than DeiTs corresponding, but their adversarial robustness is significantly lower, consistent with findings from [32, 42]. Our proposed SATA model achieves superior performance against both FGSM [12] and PGD [31] attacks. Specifically, SATA-S, SATA-B, and SATA-B* show over a " + }, + { + "bbox": [ + 55, + 572, + 297, + 644 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 572, + 297, + 644 + ], + "type": "text", + "content": " improvement on FGSM [12] compared to previous ViT variants." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": "Regarding natural adversarial robustness, the proposed SATA-T demonstrates a comparable performance of " + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": ", which is on par with some current state-of-the-art methods like RVT-Ti [32] and RVT-Ti+RSPC [16], while being about half their size. However, for models of similar size (e.g.," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 572, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 555, + 608 + ], + "type": "text", + "content": "ViT-Small and ViT-Base), the proposed SATAs outperform others by about " + }, + { + "bbox": [ + 313, + 572, + 555, + 608 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 572, + 555, + 608 + ], + "type": "text", + "content": ", indicating the effectiveness of SATA against natural adversarial attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 622, + 545, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 545, + 635 + ], + "type": "text", + "content": "5.4. Common Corruption Robustness Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "To measure model degradation on common image corruptions, we report the mean corruption error (mCE) on ImageNet-C [20] (IN-C) in Table 1. Our SATA method significantly reduces the mCE of DeiT-Ti [46] from " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "71.1\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "51.1\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", achieving the lowest mCE among vision transformers within the ViT-Tiny group. For the other two larger" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9735" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 289, + 190 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 289, + 190 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 289, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 289, + 190 + ], + "type": "image", + "image_path": "ff8967be74804351126cc4974e2ad94160563efd56dff610c314463ddc8787c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 194, + 181, + 203 + ], + "lines": [ + { + "bbox": [ + 170, + 194, + 181, + 203 + ], + "spans": [ + { + "bbox": [ + 170, + 194, + 181, + 203 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 205, + 275, + 318 + ], + "blocks": [ + { + "bbox": [ + 66, + 205, + 275, + 318 + ], + "lines": [ + { + "bbox": [ + 66, + 205, + 275, + 318 + ], + "spans": [ + { + "bbox": [ + 66, + 205, + 275, + 318 + ], + "type": "image", + "image_path": "165bf6f2641482103d718545c0dc815d05d7541798f66d4dfef04bdebd18ab78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 323, + 181, + 333 + ], + "lines": [ + { + "bbox": [ + 170, + 323, + 181, + 333 + ], + "spans": [ + { + "bbox": [ + 170, + 323, + 181, + 333 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": "Figure 4. (a) Ablation study on " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": ". (b) Ablation study on " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 342, + 296, + 399 + ], + "type": "text", + "content": " to 0.7 and 1.0, respectively, for all experiments throughout this paper. Ablation studies are conducted on the SATA-B model using the ImageNet-1K dataset. Dashed lines for both graphs represent the baseline Deit-Base/16 top-1 accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 503 + ], + "type": "text", + "content": "ViT groups, the proposed SATA models achieve an mCE of approximately " + }, + { + "bbox": [ + 55, + 419, + 296, + 503 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 55, + 419, + 296, + 503 + ], + "type": "text", + "content": ", improving by around 20 points over all other ViT or CNN-based methods on the leaderboard, thereby establishing a new state-of-the-art. This result also suggests that spatial autocorrelation management of visual tokens can successfully handle different types of image corruption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 279, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 279, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 279, + 545 + ], + "type": "text", + "content": "5.5. Out-of-Distribution Robustness Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": "We evaluate the generalization ability of SATA on out-of-distribution data by reporting the top-1 accuracy on ImageNet-R [22] (IN-R) and ImageNet-Sketch [49] (IN-SK) in Table 1. The generic vision transformers [9, 46] enhanced by the proposed SATA consistently outperform other ViT models on ImageNet-R [22], achieving " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "47.3\\%" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "57.2\\%" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "70.0\\%" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "79.9\\%" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": " in the ViT-Tiny, ViT-Small, and ViT-Base groups, respectively. Regarding ImageNet-Sketch [49] (IN-SK), SATA demonstrates superior performance compared to other models of similar size. These results imply that the spatial autocorrelation tokens analysis effectively captures feature distribution shifts, enhancing the model's out-of-distribution generalization capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "type": "text", + "content": "5.6. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": "Token Splitting and Token Grouping We evaluate the role of token splitting based on the upper and lower bounds of spatial autocorrelation scores and token grouping modules. To this end, we examine five SATA configurations, as depicted in Table 2. As shown in Table 2, utilizing only token grouping yields a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "84.4\\%" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ", improving upon the reference (DeiT-Base) accuracy by " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": ". Including either lower or upper bounds significantly enhances accuracy by about " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " of the baseline top-1 accuracy, highlighting the effectiveness of the proposed spatial autocorrelation token splitting schema. The upper bound contributes slightly more, suggesting that tokens with extremely high spatial autocorrelation scores are more likely to be filtered by the splitting process. However, incorporating upper or lower bounds also increases computations and token counts, reducing throughput despite the substantial accuracy gains. Additionally, applying bounds does not guarantee an equal token distribution across bands, as this depends on the data distribution (shown in Figure 2 of the supplementary material). Finally, adding token grouping provides a further " + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 89, + 555, + 341 + ], + "type": "text", + "content": " improvement over the splitting process alone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 413 + ], + "type": "text", + "content": "Regarding throughput, two main components of the proposed SATA framework (i.e. spatial autocorrelation-based token splitting and token grouping) primarily utilise parallelizable matrix multiplication and dot product operations. This design ensures computational efficiency with a negligible impact on inference time, as shown in Table 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": "Threshold of starting block " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": " We also examine the effect of parameter " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": ", which controls at which transformer block the SATA module is applied. As Figure 4(a) shows, applying SATA from block " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "0.4 \\times B" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": " onwards significantly improves model efficiency while exceeding baseline ViT accuracy " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "(82\\%)" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": ". Applying SATA to earlier blocks " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "(\\gamma < 0.4)" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": " degrades accuracy, suggesting that high spatial correlation within token features and the importance of all tokens in early layers is beneficial. To achieve a good trade-off between accuracy and efficiency, we use " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "\\gamma = 0.7" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": " in all our experiments, unless stated otherwise. This results in a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "inline_equation", + "content": "93.9\\%" + }, + { + "bbox": [ + 313, + 430, + 556, + 574 + ], + "type": "text", + "content": " and GFLOPs of 15.9." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "content": "Lower/Upper bounds controlling factor " + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "content": " We further assess the influence of " + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "content": ", the factor controlling the lower and upper bounds in SATA. Figure 4(b) shows the performance of SATA-B on ImageNet-1K with " + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "content": " values ranging from 0.5 to 2. " + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "content": " determines the number of tokens passed to the FFN block and setting " + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 590, + 556, + 664 + ], + "type": "text", + "content": " yields optimal performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 671, + 470, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 470, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 470, + 682 + ], + "type": "text", + "content": "5.7. Visualisation and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Although the effectiveness of the proposed SATA module has been empirically demonstrated, we conduct a deeper" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9736" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 112, + 542, + 223 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Ablation study of the spatial autocorrelation module (token splitting) and token grouping. The symbols “√” and “X” indicate whether the corresponding element is employed with the configuration or not. † represents top-1 accuracy for the baseline Deit-Base/16 [46] model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 112, + 542, + 223 + ], + "lines": [ + { + "bbox": [ + 64, + 112, + 542, + 223 + ], + "spans": [ + { + "bbox": [ + 64, + 112, + 542, + 223 + ], + "type": "table", + "html": "
Spt. Auto. Correlation (Tokens Splitting)Tokens GroupingTop-1Throughput (ims/s)Latency (s)
Lower bound (μs - |s|)Upper bound (μs + |s|)
×××82.0†3030.0033
××✓84.43610.0027
×✓✓92.82970.0033
✓×✓92.52710.0035
✓✓×92.32680.0037
✓✓✓93.92760.0035
", + "image_path": "f6332add24d47ee44db94d11044329fd89c2b153624cdfdf548c072f542544b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 236, + 279, + 357 + ], + "blocks": [ + { + "bbox": [ + 63, + 236, + 279, + 357 + ], + "lines": [ + { + "bbox": [ + 63, + 236, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 63, + 236, + 279, + 357 + ], + "type": "image", + "image_path": "765594f9b73860db62910a2eaae275672e63fde7965eea7a50fd8d7a29ed6ec3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 360, + 181, + 369 + ], + "lines": [ + { + "bbox": [ + 170, + 360, + 181, + 369 + ], + "spans": [ + { + "bbox": [ + 170, + 360, + 181, + 369 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 377, + 288, + 479 + ], + "blocks": [ + { + "bbox": [ + 62, + 377, + 288, + 479 + ], + "lines": [ + { + "bbox": [ + 62, + 377, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 288, + 479 + ], + "type": "image", + "image_path": "1149a4d3cc9f7cdad93af8672b24fe378cbcfb79d7c9e7abe1df6afbbe05ef25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 480, + 179, + 489 + ], + "lines": [ + { + "bbox": [ + 169, + 480, + 179, + 489 + ], + "spans": [ + { + "bbox": [ + 169, + 480, + 179, + 489 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 496, + 297, + 574 + ], + "lines": [ + { + "bbox": [ + 55, + 496, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 297, + 574 + ], + "type": "text", + "content": "Figure 5. (a) Cosine similarity between the clean and corrupted versions of the class token attention map and spatial autocorrelation scores across different blocks of SATA-B. Results are averaged across various types of image corruptions and severity levels on ImageNet-C [20]. (b) Visualisation of token splitting for a pair of clean and noisy images. Notably, the selected tokens for each set are similar for both clean and corrupted inputs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "investigation to better understand its behaviour. To this end, we calculate the cosine similarity between the class token attention maps and spatial autocorrelation scores of clean and corrupted image pairs from the ImageNet-1K validation set and its corresponding ImageNet-C [20], respectively. We compute these similarities for various types of image corruptions and severity levels in ImageNet-C [20], and report the average values across different blocks of the proposed SATA-B in Figure 5(a). As shown in Figure 5(a), the cosine similarity between clean and corrupted attention maps" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": "drops significantly in the later blocks of the transformer. In contrast, the similarity for spatial autocorrelation scores improves at the early stages and remains consistently high, averaging above 0.8. This highlights that the proposed method can provide a more stable and reliable feature representation throughout the network, offering strong robustness against various types of corruption." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "type": "text", + "content": "Moreover, Figure 5 provides a visualization of tokens (patches) are split into set " + }, + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "type": "text", + "content": " and set " + }, + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 322, + 556, + 395 + ], + "type": "text", + "content": " for a pair of clean and noisy images according to the proposed SATA algorithm. Notably, the similarity between corresponding sets for clean and noisy inputs is evident, further highlighting the robustness of the proposed method2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 411, + 388, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 411, + 388, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 388, + 423 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 589 + ], + "type": "text", + "content": "In this paper, we introduce SATA, a novel method designed to enhance the performance and robustness of vision transformers against various types of corruption. SATA utilises a straightforward yet powerful spatial autocorrelation scheme to exploit spatial inter-dependencies among token features, significantly improving representational capacity and efficiency while reducing computational costs. Our experimental results demonstrate that SATA-enhanced vision transformers consistently deliver stable and reliable feature representations, achieving state-of-the-art performance on ImageNet-1K classification and setting new benchmarks for robustness across multiple evaluations, all without requiring additional training or fine-tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 687 + ], + "type": "text", + "content": "This work highlights SATA's transformative potential and opens several promising avenues for future research. Key directions include adapting SATA for window-based and hybrid ViT architectures to boost performance in tasks such as object detection and segmentation. Additionally, exploring SATA's application in other transformer-based domains, such as large language models (LLMs), could further extend its impact." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 553, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 553, + 714 + ], + "type": "text", + "content": "2Additional visual comparisons are included in Supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9737" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "content": "This work was supported in part by the Australian Research Council (ARC), Australia under Discovery Grant DP180100958 and Industrial Transformation Research Hub Grant IH180100002." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "text", + "content": "[1] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? In NeurIPS, pages 26831-26843, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 202, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 296, + 246 + ], + "type": "text", + "content": "[2] Srinadh Bhojanapalli, Ayan Chakrabarti, Daniel Glasner, Daliang Li, Thomas Unterthiner, and Andreas Veit. Understanding robustness of transformers for image classification. In ICCV, pages 10231-10241, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 278 + ], + "type": "text", + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 296, + 312 + ], + "type": "text", + "content": "[4] Hila Chefer, Idan Schwartz, and Lior Wolf. Optimizing relevance maps of vision transformers improves robustness. In NeurIPS, pages 33618-33632, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 314, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 296, + 346 + ], + "type": "text", + "content": "[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 369 + ], + "type": "text", + "content": "[6] Yanguang Chen. New approaches for calculating Moran's index of spatial autocorrelation. *PloS one*, 8(7), 2013. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 403, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 435 + ], + "type": "text", + "content": "[8] Minjing Dong, Yanxi Li, Yunhe Wang, and Chang Xu. Adversarily robust neural architectures. arXiv preprint arXiv:2009.00902, 2020. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 437, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 501 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 502, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 296, + 546 + ], + "type": "text", + "content": "[10] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In ICML, pages 2286-2296. PMLR, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 547, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 296, + 602 + ], + "type": "text", + "content": "[11] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In ECCV, pages 396-414. Springer, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 634 + ], + "type": "text", + "content": "[12] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 679 + ], + "type": "text", + "content": "[13] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In ICCV, pages 12259-12269, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[14] Jindong Gu, Volker Tresp, and Yao Qin. Are vision transformers robust to patch perturbations? In European Conference on Computer Vision, pages 404-421. Springer, 2022. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 556, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 556, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 556, + 105 + ], + "type": "text", + "content": "[15] Yong Guo, David Stutz, and Bernt Schiele. Robustifying token attention for vision transformers. In ICCV, pages 17557-17568, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 138 + ], + "type": "text", + "content": "[16] Yong Guo, David Stutz, and Bernt Schiele. Improving robustness of vision transformers by reducing sensitivity to patch corruptions. In CVPR, pages 4108-4118, 2023. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 140, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 172 + ], + "type": "text", + "content": "[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, pages 15908-15919, 2021. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 174, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 555, + 216 + ], + "type": "text", + "content": "[18] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE TPAMI, 45(1):87-110, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 251, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 281 + ], + "type": "text", + "content": "[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 2, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 284, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 338 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 339, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 555, + 392 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 2, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "type": "text", + "content": "[23] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In ICCV, pages 11936-11945, 2021. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 438, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 555, + 491 + ], + "type": "text", + "content": "[24] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In ECCV, pages 620-640. Springer, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 494, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 555, + 536 + ], + "type": "text", + "content": "[25] Xuelong Li, Han Zhang, Rui Zhang, and Feiping Nie. Discriminative and uncorrelated feature selection with constrained spectral analysis in unsupervised learning. IEEE Transactions on Image Processing, 29:2139-2149, 2019. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 537, + 555, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 569 + ], + "type": "text", + "content": "[26] Yanxi Li and Chang Xu. Trade-off between robustness and accuracy of vision transformers. In CVPR, pages 7558-7568, 2023. 3, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 571, + 556, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 556, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 556, + 623 + ], + "type": "text", + "content": "[27] Yingwei Li, Qihang Yu, Mingxing Tan, Jieru Mei, Peng Tang, Wei Shen, Alan Yuille, and Cihang Xie. Shape-texture de-biased neural network training. In In Proceedings of the International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "type": "text", + "content": "[28] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5, 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9738" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 1, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "text", + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[32] Xiaofeng Mao, Gege Qi, Yuefeng Chen, Xiaodan Li, Ranjie Duan, Shaokai Ye, Yuan He, and Hui Xue. Towards robust vision transformer. In CVPR, pages 12042-12051, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[33] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In CVPR, pages 12309-12318, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 262 + ], + "type": "text", + "content": "[34] Patrick AP Moran. The interpretation of statistical maps. Journal of the Royal Statistical Society. Series B (Methodological), 10(2):243-251, 1948. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[35] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In NeurIPS, pages 23296-23308, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 342 + ], + "type": "text", + "content": "[36] Nick Nikzad, Yongsheng Gao, and Jun Zhou. CSA-Net: Channel-wise Spatially Autocorrelated Attention Networks. arXiv preprint arXiv:2405.05755, 2024. 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "type": "text", + "content": "[37] Sayak Paul and Pin-Yu Chen. Vision transformers are robust learners. In AAAI, pages 2071-2081, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "text", + "content": "[38] Yao Qin, Chiyuan Zhang, Ting Chen, Balaji Lakshminarayanan, Alex Beutel, and Xuezhi Wang. Understanding and improving robustness of vision transformers through patch-based negative augmentation. In NeurIPS, pages 16276-16289, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 454 + ], + "type": "text", + "content": "[39] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, pages 10428-10436, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 498 + ], + "type": "text", + "content": "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, pages 13937-13949, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[41] Evgenia Rusak, Lukas Schott, Roland S Zimmermann, Julian Bitterwolf, Oliver Bringmann, Matthias Bethge, and Wieland Brendel. A simple way to make neural networks robust against diverse image corruptions. In ECCV, pages 53-69. Springer, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 588 + ], + "type": "text", + "content": "[42] Rulin Shao, Zhouxing Shi, Jinfeng Yi, Pin-Yu Chen, and Cho-Jui Hsieh. On the adversarial robustness of vision transformers. arXiv preprint arXiv:2103.15670, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 296, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 611 + ], + "type": "text", + "content": "[43] Dai Shi. Transnext: Robust foveal visual perception for vision transformers. In CVPR, pages 17773-17783, 2024. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "type": "text", + "content": "[44] Zhuoran Song, Yihong Xu, Zhezhi He, Li Jiang, Naifeng Jing, and Xiaoyao Liang. Cp-vit: Cascade vision transformer pruning via progressive sparsity prediction. arXiv preprint arXiv:2203.04570, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "type": "text", + "content": "[45] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "type": "text", + "content": "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 533 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "type": "text", + "content": "[47] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In ICCV, pages 32-42, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "text", + "content": "[49] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "text", + "content": "[50] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, pages 568-578, 2021. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 264, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 297 + ], + "type": "text", + "content": "[51] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In ICCV, pages 22–31, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "type": "text", + "content": "[52] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vitae: Vision transformer advanced by exploring intrinsic inductive bias. In NeurIPS, pages 28522-28535, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "text", + "content": "[53] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "type": "text", + "content": "[54] Kun Yuan, Shaopeng Guo, Ziwei Liu, Aojun Zhou, Fengwei Yu, and Wei Wu. Incorporating convolution designs into visual transformers. In ICCV, pages 579-588, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 411, + 555, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 454 + ], + "type": "text", + "content": "[55] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. In ICCV, pages 558-567, 2021. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 456, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 487 + ], + "type": "text", + "content": "[56] Qinglong Zhang and Yu-Bin Yang. Rest: An efficient transformer for visual recognition. In NeurIPS, pages 15475-15485, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 489, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 533 + ], + "type": "text", + "content": "[57] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Animashree Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. In ICML, pages 27378-27394. PMLR, 2022. 1, 3, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9739" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_content_list.json b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c96c41a89697e3942ef2ab834bad07a8c22cae01 --- /dev/null +++ b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_content_list.json @@ -0,0 +1,1766 @@ +[ + { + "type": "text", + "text": "SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting", + "text_level": 1, + "bbox": [ + 135, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenyu Zhang*, Kunlun Xu*, Zichen Liu, Yuxin Peng, Jiahuan Zhou†", + "bbox": [ + 227, + 203, + 774, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wangxuan Institute of Computer Technology, Peking University, Beijing 100871, China", + "bbox": [ + 238, + 223, + 761, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{2200010814, xkl, lzc20180720}@stu.pku.edu.cn, {pengyuxin, jiahuanzhou}@pku.edu.cn", + "bbox": [ + 135, + 241, + 854, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) encounter considerable challenges when adapting to domain shifts stemming from changes in data distribution. Test-time adaptation (TTA) has emerged as a promising approach to enhance VLM performance under such conditions. In practice, test data often arrives in batches, leading to increasing interest in the transductive TTA setting. However, existing TTA methods primarily focus on individual test samples, overlooking crucial cross-sample correlations within a batch. While recent ViT-based TTA methods have introduced batch-level adaptation, they remain suboptimal for VLMs due to inadequate integration of the text modality. To address these limitations, we propose a novel transductive TTA framework, Supportive Clique-based Attribute Prompting (SCAP), which effectively combines visual and textual information to enhance adaptation by generating fine-grained attribute prompts across test batches. SCAP first forms supportive cliques of test samples in an unsupervised manner based on visual similarity and learns an attribute prompt for each clique, capturing shared attributes critical for adaptation. For each test sample, SCAP aggregates attribute prompts from its associated cliques, providing enriched contextual information. To ensure adaptability over time, we incorporate a retention module that dynamically updates attribute prompts and their associated attributes as new data arrives. Comprehensive experiments across multiple benchmarks demonstrate that SCAP outperforms existing state-of-the-art methods, significantly advancing VLM generalization under domain shifts. Our code is available at https://github.com/zhoujiahuan1991/CVPR2025-SCAP.", + "bbox": [ + 86, + 323, + 483, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) such as CLIP [32] have exhibited remarkable generalization capabilities across a", + "bbox": [ + 89, + 832, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3663c2c26e042dcd2e24ccfc41b98c255b2f5b41b5d7fde59ccfba92673f6a61.jpg", + "image_caption": [ + "(a) Current Prompt Learning based Test-time Adaptation Methods" + ], + "image_footnote": [], + "bbox": [ + 521, + 291, + 906, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2ecaa5d2a555735b2357b754ef3dd69353a1009afc91b7d5c969ad16e0ac72d.jpg", + "image_caption": [ + "(b) Ours", + "Figure 1. Comparison between our proposed SCAP to existing prompt learning-based TTA methods. Specifically, current TTA methods learn from instances, while our method utilizes cross-sample visual similarity information from batches to construct supportive cliques and extract attributes from them. We learn attributes from both modalities based on the cliques. For each image, SCAP jointly utilizes attribute prompts from its associated cliques, leading to more effective and accurate prompting." + ], + "image_footnote": [], + "bbox": [ + 521, + 483, + 903, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "range of downstream tasks in zero-shot settings. These advancements highlight the significant potential of CLIP in a wide range of domains. However, their performance often degrades [7, 20, 25] when downstream data exhibits substantial domain shifts [43, 46, 52, 55] from the pre", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30032", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training distribution [10, 36, 44, 54]. This limitation has sparked increasing interest in the area of Test-Time Adaptation (TTA) [1, 2, 39, 49], which aims to adapt pre-trained VLMs to mitigate domain shifts during test time.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works in TTA have primarily focused on adapting CLIP by independently learning from each sample during test time [20, 25, 36, 49]. However, in real-world scenarios, test-time data typically arrives in batches, creating a transductive TTA setting. Existing TTA methods can tackle such a scenario by treating each test sample individually [25]. Although this solution is available, it suffers from significant limitations due to disregarding the intrinsic cross-sample visual similarity within each batch. Exploiting these inter-sample similarities can greatly enhance both the effectiveness and robustness of the adaptation process. As illustrated in Figure 1, test batches often contain Cliques, groups of visually and semantically similar images that share common characteristics, referred to as Attributes in this paper. Leveraging these cliques enables a more precise and context-aware adaptation by guiding the model to focus on discriminative attributes.", + "bbox": [ + 89, + 154, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works [13, 17, 28, 31, 38] have explored transductive TTA under the Vision Transformer (ViT) backbone, often relying on batch statistics to adjust normalization parameters or select class-balanced samples for adaptation. However, these methods cannot be directly applied to CLIP since they completely neglect the adaptation of the text modality, which is crucial in VLMs. Rare CLIP-based TTA methods are designed for the transductive setting. In the latest method [14], the top-k predicted classes of each image in a batch are aggregated into a single new text prompt, which is used as pseudo labels to reclassify inputs in a transductive manner. However, the aforementioned transductive TTA methods mainly perform coarse-grained batch-level utilization of the visual similarity information. Thus, the fine-grained discriminative attribute information of cross-sample visual similarity is underutilized.", + "bbox": [ + 89, + 414, + 482, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Supportive Clique-based Attribute Prompting (SCAP), a novel CLIP-based transductive TTA approach designed to effectively leverage the cross-sample visual similarity information. Given a test batch, SCAP automatically mines distinct Supportive Cliques containing test images with shared attributes by leveraging visual and semantic similarity metrics. These cliques play a crucial role in learning Attribute Prompts that capture discriminative, attribute-specific information for each clique. For any given test sample, its aggregated attribute representation is constructed by combining the learned attribute prompts associated with the cliques to which it belongs. By utilizing these aggregated attribute prompts, SCAP enables CLIP to focus on fine-grained, instance-specific details, enhancing prediction accuracy and robustness. Additionally, we introduce a retention mechanism that dynamically up", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dates and preserves attribute prompts, ensuring long-term adaptability across tasks. Extensive experiments on two benchmarks demonstrate that SCAP significantly outperforms various current state-of-the-art methods in the transductive TTA setting, showing the effectiveness of SCAP in mining the intrinsic knowledge in batches.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this work are fourfold: (1) We propose Supportive Clique-based Prompting (SCAP), a novel approach that leverages the rich information of cross-sample relationships, improving the effectiveness of test-time adaptation under transductive setting. (2) To achieve robust and accurate fine-grained attribute prompt learning during test time, we propose to learn attribute prompts in a transductive manner based on cliques. (3) To retain the historically learned attribute prompts, we further design a retention mechanism to preserve the learned prompts for future reuse. (4) Extensive experiments on two benchmarks demonstrate that SCAP outperforms several current state-of-the-art methods in the transductive TTA setting, highlighting its effectiveness in extracting intrinsic knowledge from batches.", + "bbox": [ + 511, + 183, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 424, + 653, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Prompt Learning for CLIP", + "text_level": 1, + "bbox": [ + 511, + 450, + 758, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLMs like CLIP have demonstrated abundant knowledge learned from large-scale image-text data [5, 18, 19, 32, 33, 45]. Adapting CLIP to various downstream tasks is a valuable and challenging problem [3]. Recently, prompt learning [23, 24, 37, 48, 51] has emerged as a promising approach for enhancing model adaptability. Specifically, CoOp [57], CoCoOp [56], MaPLe [21], and Tip-Adapter [51] have achieved state-of-the-art performance. However, they rely on training data from downstream tasks. However, these approaches rely on training data annotations and often struggle when faced with substantial domain shifts at test time [38, 42].", + "bbox": [ + 511, + 473, + 906, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 511, + 666, + 718, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time adaptation (TTA) focuses on dynamically adjusting pre-trained models, such as CLIP or ViTs [22, 25], by leveraging sequentially obtained unlabeled test data [9, 11, 38, 49, 50]. Existing TTA approaches can be broadly categorized into training-free, augmentation-based, and prompt-based methods. Training-free methods aim to adapt CLIP efficiently without backpropagation [7, 20]. TDA introduces a positive and negative cache mechanism to assist prediction. Zero [7] aggregates predictions from multiple augmented versions without modifying model parameters. The augmentation-based methods [6, 49] focus on generating diverse augmented samples to integrate test distribution knowledge into the model. MTA [49] introduces a quality assessment variable for each", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30033", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a13064320d9da45957ed8d1376e7644ea1170283872273512d632c098f4664af.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed SCAP. SCAP firstly mines the supportive cliques for all images in parallel. Based on the cliques, it then learns the corresponding attribute prompts from both modalities. The learned visual attribute prompts and the text attribute prompts are separately retained to accumulate the knowledge of test domains. For each instance, we conduct inference by jointly utilizing all the associated attribute prompts and the retained knowledge to generate the final prediction." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "augmented view and models the distribution using a mixed Gaussian approach. Prompt-based methods employ prompt tuning to adapt the model while preserving generalization capabilities [12, 25, 26, 36, 47, 50]. For example, TPT [36] optimizes instance-specific text prompts by minimizing prediction entropy, whereas DART [25] refines prompts across both modalities while leveraging historically learned knowledge via prompt retention. While these methods assume a sequential data setting where test samples are processed individually [13, 38], real-world scenarios often involve test-time data arriving in batches. Consequently, existing approaches struggle to exploit the inter-sample relationships within a batch, leading to suboptimal adaptation performance.", + "bbox": [ + 88, + 535, + 485, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Transductive Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 89, + 757, + 401, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transductive TTA considers the practical scenario where test data are provided in batches rather than individual samples. Existing Transductive TTA methods are predominantly developed for ViTs pre-trained on ImageNet [13, 28, 31, 35, 39, 41]. These approaches typically rely on batch statistics to update normalization parameters [39] or select confident samples for model tuning [28]. Recent studies have demonstrated that CLIP outperforms ViTs pre-trained", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on ImageNet in TTA tasks. However, existing transductive TTA methods fail to fully exploit CLIP's capabilities, as they largely disregard the adaptation of the text modality, which is crucial for leveraging its knowledge learned from vision-language alignment. A few recent efforts have attempted to adapt CLIP for transductive TTA. For example, CLIPArTT [14] optimizes the model by jointly utilizing all images within a batch. However, since cross-instance relationships within a batch are often imbalanced, directly modeling all samples together does not effectively exploit the most informative relationships. In this work, we take a step further by identifying and leveraging the informative cliques, groups of strongly correlated samples within a batch, to facilitate robust adaptation by capturing discriminative attributes.", + "bbox": [ + 511, + 535, + 906, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 777, + 604, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting and Notations", + "text_level": 1, + "bbox": [ + 511, + 801, + 785, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In transductive TTA, test data arrive sequentially in batches. Given $t$ -th test batch $B_{t} = \\{x_{j}\\}_{j=1}^{b}$ , where $b$ is the batch size and $x_{j}$ is a test image. We generate a set of Supportive Cliques for each class $i$ , denoted as $\\mathcal{C}_i = \\{C_{ik}\\}_{k=1}^m$ . Then, we extract the corresponding clique attributes, which are de", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30034", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "noted as $\\mathcal{A}_i = \\{a_{ik} \\in \\mathbb{R}^d\\}_{k=1}^m$ where $d$ is the attribute dimension, equivalent to the feature dimension. Additionally, we introduce Attribute Prompts $\\mathcal{P}_i^A = \\{\\mathcal{P}_i^I, \\mathcal{P}_i^T\\}$ , which consist of visual attribute prompts $\\mathcal{P}_i^I = \\{P_{ik}^I\\}_{k=1}^m$ and text attribute prompts $\\mathcal{P}_i^T = \\{P_{ik}^T\\}_{k=1}^m$ . Our model $\\theta$ is built on CLIP, which comprises an image encoder $\\mathbf{E}_I$ and a text encoder $\\mathbf{E}_T$ .", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Attribute-Driven Supportive Clique Mining", + "text_level": 1, + "bbox": [ + 89, + 203, + 464, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attributes are the fine-grained characteristics shared by a group of images in the batch. Since images containing the same attribute often exhibit high visual similarity to each other, in this section, we propose to obtain Supportive Cliques for the potential attributes by evaluating cross-image similarity.", + "bbox": [ + 89, + 224, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we use CLIP to obtain the top- $k$ class predictions $topk(x_j)$ for each image $x_j$ , where $k$ is a hyperparameter. For each class $i$ , all images whose top- $k$ predictions include class $i$ are selected to form a subset $S_i$ :", + "bbox": [ + 89, + 316, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} = \\left\\{x _ {j} \\in B _ {t}: i \\in \\operatorname {t o p k} \\left(x _ {j}\\right) \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 383, + 482, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since images in $S_{i}$ share similar class-specific information, which ensures semantic relevance. To further capture more fine-grained relevance, we extract the features $F_{i}^{I}\\in$ $\\mathbb{R}^{|S_i|\\times d}$ of images within $S_{i}$ using the CLIP image encoder $\\mathbf{E}_I$ and compute the similarity matrix $M_{i}\\in \\mathbb{R}^{|S_{i}|\\times |S_{i}|}$ as follows:", + "bbox": [ + 89, + 405, + 483, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} = F _ {i} ^ {I} \\cdot \\left(F _ {i} ^ {I}\\right) ^ {\\top} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 492, + 480, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, given the $l$ -th row of $M_{i}$ , we collect elements exceeding a threshold $t$ , whose corresponding images form a Clique $C_{il}$ . Formally, the Clique generated from the $l$ -th row of $M_{i}$ is defined as:", + "bbox": [ + 89, + 515, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {i l} = \\left\\{x _ {j} ^ {i}: x _ {j} ^ {i} \\in S _ {i}, (M _ {i}) _ {l j} > t \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 589, + 480, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(M_{i})_{lj}$ represents the $l$ -th row, $j$ -th column of $M_{i}$ . Only non-trivial Cliques $(|C_{il}| > 1)$ are retained, and duplicate Cliques are removed to ensure uniqueness. Finally, we obtain a set of valid Cliques $\\mathcal{C}_i = \\{C_{ik}\\}_{k=1}^m$ .", + "bbox": [ + 89, + 609, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attribute Prompt Learning", + "text_level": 1, + "bbox": [ + 89, + 676, + 336, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we aim to learn prompts that capture attribute-specific knowledge to facilitate classification.", + "bbox": [ + 89, + 700, + 482, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for each Clique $C_{ik}$ , we learn a visual attribute prompt $P_{ik}^{A}$ from the image modality and a text attribute prompt $P_{ik}^{T}$ from the text modality. The feature $f_{j}^{I} \\in \\mathbb{R}^{d}$ for the $j$ -th image in $C_{ik}$ is extracted as:", + "bbox": [ + 89, + 729, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {j} ^ {I} = \\mathbf {E} _ {I} \\left(C _ {i k} ^ {j}, P _ {k} ^ {I}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 797, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we define the attribute feature $a_{ik} \\in \\mathbb{R}^d$ of $C_{ik}$ which is calculated by:", + "bbox": [ + 89, + 821, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {i k} = \\frac {1}{\\left| C _ {i k} \\right|} \\sum_ {j = 1} ^ {\\left| C _ {i k} \\right|} f _ {j} ^ {I} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 859, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, we compute the text feature $f_{n}^{T}$ of the $n$ -th class as follows:", + "bbox": [ + 511, + 90, + 905, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n} ^ {T} = \\mathbf {E} _ {T} \\left(c _ {n}, P _ {i k} ^ {T}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 133, + 903, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given total $N$ classes, we introduce an attribute-specific class probability distribution vector, which is obtained by:", + "bbox": [ + 511, + 157, + 905, + 189 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {P}} _ {i k} = \\left[ \\hat {p} _ {i k} ^ {1}, \\hat {p} _ {i k} ^ {2}, \\dots , \\hat {p} _ {i k} ^ {N} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 198, + 903, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 227, + 558, + 239 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {i k} ^ {n} = \\frac {\\exp \\left(\\cos \\left(a _ {i k} , f _ {n} ^ {T}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(a _ {i k} , f _ {j} ^ {T}\\right) / \\tau\\right)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 237, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $\\cos (\\cdot ,\\cdot)$ denotes cosine similarity and $\\tau$ is a temperature parameter to scale the similarity scores.", + "bbox": [ + 511, + 282, + 905, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we compute the Entropy Loss for the probability distribution of the attribute:", + "bbox": [ + 511, + 313, + 905, + 342 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {e n} \\left(C _ {i k}\\right) = - \\sum_ {n = 1} ^ {N} \\hat {p} _ {i k} ^ {n} \\log \\hat {p} _ {i k} ^ {n} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 349, + 903, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further guide the model to extract attribute-specific information and discard attribute-irrelevant information, we introduce a concentration loss $\\mathcal{L}_{con}(C_{ik})$ that forces the model to capture the shared information across the samples within a Clique. Specifically, $\\mathcal{L}_{con}(C_{ik})$ is calculated by", + "bbox": [ + 511, + 401, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} \\left(C _ {i k}\\right) = \\sum_ {j = 1} ^ {| C _ {i k} |} \\left| \\left| f _ {j} ^ {I} - a _ {i k} \\right| \\right| ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 487, + 903, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, both kinds of prompts $P_{ik}^{I}$ and $P_{ik}^{T}$ are guided to encode attribute-specific knowledge as the model optimization. Such knowledge is valuable when generating predictions for images showing the same attributes.", + "bbox": [ + 511, + 542, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Attribute Prompt Retention", + "text_level": 1, + "bbox": [ + 513, + 613, + 764, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, to retain attribute knowledge during the test process and ensure consistent performance improvement, we introduce mechanisms for retaining both text and visual attribute knowledge.", + "bbox": [ + 511, + 635, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, to retain the text attribute knowledge, we propose a Text Retention Prompt $\\tilde{P}^T$ . Specifically, when a new Clique $C_{ik}$ is learned, we update $\\tilde{P}^T$ by", + "bbox": [ + 511, + 696, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {P} ^ {T} = \\alpha_ {\\tau} \\tilde {P} ^ {T} + (1 - \\alpha_ {\\tau}) P _ {i k} ^ {T}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 750, + 903, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{\\tau} = \\tau /(1 + \\tau)$ is a decay rate based on iteration count $\\tau$", + "bbox": [ + 511, + 779, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, to retain the visual attribute knowledge, we design a Retention Cache $\\mathcal{R} = \\{R_i\\}_{i=1}^N$ where $R_i$ is a class-specific retention cache. Specifically, we store $(a_{ik}, P_{ik}^I)$ as key-value pairs in $R_i$ . We denote the retained attributes (i.e. keys) as $R_i^A = \\{a_1, \\ldots, a_L\\}$ and the corresponding attribute prompts (i.e. values) as $R_i^P = \\{P_1^I, \\ldots, P_L^I\\}$ . The", + "bbox": [ + 511, + 810, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30035", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "maximum size of each $R_i$ is set to $L$ to avoid memory explosion. When the retained key-value pair number is less than $L$ , we directly add a new $(a_{ik}, P_{ik}^I)$ to $R_i$ . Otherwise, we add $(a_{ik}, P_{ik}^I)$ to $R_i$ first, then we update all pairs and aggregate the two that are most similar to each other in $R_i$ by using a graph structure as follows.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we construct a weighted graph $G_{i} = (V_{i},E_{i})$ . The vertices are the keys in the retention cache $V_{i} = R_{i}^{A} = \\{a_{1},\\dots ,a_{L + 1}\\}$ . The edges $E_{i}$ between each two vertices are defined as the Gaussian Distance of the two attributes. Thus, the adjacency matrix $W_{i}\\in \\mathbb{R}^{(L + 1)\\times (L + 1)}$ of $G_{i}$ is defined with:", + "bbox": [ + 89, + 181, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(W _ {i}\\right) _ {k l} = \\exp \\left(- \\frac {\\left\\| a _ {k} - a _ {l} \\right\\| ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 281, + 482, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(W_{i})_{kl}$ is the $k$ -th row, $l$ -column of $W_{i}$ and $\\sigma$ is a hyperparameter. We utilize the $K$ -max values in each row of $W_{i}$ to construct a $K$ -nearest retention graph. Then we perform graph propagation following [53] and obtain the updated graph vertices $V_{i}^{\\prime} = \\{a_{1}^{\\prime},\\ldots ,a_{L + 1}^{\\prime}\\}$ , where the similar attributes are pushed closer automatically. Then, the most similar two attributes $a_{j_0}^\\prime ,a_{j_1}^\\prime$ and the corresponding prompts are fused into one respectively as follows:", + "bbox": [ + 89, + 324, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} a _ {n e w} = \\frac {1}{2} \\left(a _ {j _ {0}} ^ {\\prime} + a _ {j _ {1}} ^ {\\prime}\\right) \\\\ P _ {n e w} ^ {I} = \\frac {1}{2} \\left(P _ {j _ {0}} ^ {I} + P _ {j _ {1}} ^ {I}\\right) \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 457, + 482, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through the above process, we obtain the updated retention cache $R_{i}$ , dynamically retaining the visual and text attribute knowledge learned from previous batches into Retention Cache and Text Retention Prompt, respectively.", + "bbox": [ + 89, + 527, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optimization and Inference", + "text_level": 1, + "bbox": [ + 89, + 598, + 339, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each batch $B_{t}$ , we optimize the visual attribute prompts $\\mathcal{P}_i^I = \\{P_{ik}^I\\}_{k=1}^m$ and text attribute prompts $\\mathcal{P}_i^T = \\{P_{ik}^T\\}_{k=1}^m$ for every class $i$ by minimizing the following loss function:", + "bbox": [ + 89, + 621, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {k = 1} ^ {m} \\mathcal {L} _ {e n} \\left(C _ {i k}\\right) + \\lambda \\mathcal {L} _ {c o n} \\left(C _ {i k}\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 678, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter.", + "bbox": [ + 89, + 724, + 284, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we conduct inference for each $S_{i}$ of $B_{t}$ . To fully utilize both the Attribute Prompts learned from the current batch $B_{t}$ and the retained knowledge to improve prediction. Given the test image $x_{j} \\in S_{i}$ , we first collect all Cliques containing $x_{j}$ , namely $\\mathcal{K} = \\{k : x_{j} \\in C_{ik}\\}$ , whose corresponding visual attribute prompts are concatenated together:", + "bbox": [ + 89, + 739, + 483, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {I} = \\operatorname {c o n c a t} \\left(\\left[ P _ {i k} ^ {I}: k \\in \\mathcal {K} \\right]\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 844, + 482, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we introduce the retention prompt matching strategy utilizing the key-value pair structure of the retention", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cache. Specifically, the keys are the retained attribute features $R_{i}^{A} = \\{a_{1},\\dots ,a_{L}\\}$ and the values are the corresponding attribute prompts $R_{i}^{P} = \\{P_{1}^{I},\\dots ,P_{L}^{I}\\}$ . Given the image feature $f_{j}^{I}$ of $x_{j}$ which is already extracted in Sec. 3.3, the top-1 attribute $a^*$ is obtained by evaluating the similarity of $f_{j}^{I}$ to all keys in $R_{i}^{A}$ , i.e.", + "bbox": [ + 511, + 90, + 903, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na ^ {*} = \\underset {a _ {l} \\in R _ {i} ^ {A}} {\\operatorname {a r g m a x}} < f _ {j} ^ {I}, a _ {l} > \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 191, + 903, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the matched value corresponding to $a^*$ , i.e., $P^{I*}$ , we obtain the Composed Image Prompt $P_C^I$ by concatenating $P^I$ and $P^{I*}$ along the token dimension", + "bbox": [ + 511, + 229, + 903, + 275 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {C} ^ {I} = \\operatorname {c o n c a t} \\left(P ^ {I}, P ^ {I ^ {*}}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 282, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the text modality, we fuse $\\mathcal{P}^{\\mathcal{T}} = \\{P_{ik}^{T}\\}_{k=1}^{m}$ and $\\tilde{P}^T$ together to get the Composed Text Prompt $P_C^T$ :", + "bbox": [ + 511, + 306, + 903, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {C} ^ {T} = \\alpha_ {r} \\tilde {P} ^ {T} + (1 - \\alpha_ {r}) \\frac {1}{m} \\sum_ {k = 1} ^ {m} P _ {i k} ^ {T} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 348, + 903, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{r}$ is the weight of retained text prompts. Then, for the final prediction, the image-inference feature $\\hat{f}_j^I$ of $x_j^i$ and the text-inference feature $\\hat{f}_n^T$ of class $c_{n}$ are obtained by:", + "bbox": [ + 511, + 393, + 903, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\hat {f} _ {j} ^ {I} = \\mathbf {E} _ {I} \\left(x _ {j}, P _ {C} ^ {I}\\right) \\\\ \\hat {f} _ {n} ^ {T} = \\mathbf {E} _ {T} \\left(c _ {n}, P _ {C} ^ {T}\\right) \\end{array} \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 455, + 903, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prediction $y_{j}$ of $x_{j}$ is obtained by applying Eq.7.", + "bbox": [ + 532, + 500, + 883, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 527, + 645, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the two benchmarks used for model evaluation, followed by a detailed description of the implementation setup. Next, we present a comparative analysis of existing methods and our proposed approach. Finally, extensive ablation studies are conducted to validate the effectiveness of SCAP. Additional experimental results and in-depth discussions are provided in the Supplementary Materials.", + "bbox": [ + 511, + 553, + 905, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 681, + 648, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our evaluation is conducted on two benchmarks: an out-of-distribution (OOD) benchmark and a cross-domain benchmark. The OOD benchmark consists of four datasets derived from ImageNet [4], i.e., ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [40], and ImageNet-V2 [34]. The OOD benchmark aims to assess the model's robustness to substantial distribution shifts in the test data. The cross-domain benchmark comprises four datasets of the image classification task, each representing a distinct domain: Aircraft [27], Caltech101 [8], Flower102 [29], and Pets [30]. The cross-domain benchmark provides a measure of the model's adaptability across a diverse range of domain-specific tasks.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30036", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ca296c6e9c0ac0487a23e3afabef7c94b5aac5582d600e4f993ac8f0678bc964.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPublicationImageNet-AImageNet-RImageNet-SImageNet-V2Average
CLIP [32]ICML 202147.8773.9846.0961.8857.46
CoOp [57]IJCV 202249.7175.2147.9964.2059.28
CoCoOp [56]CVPR 202250.6376.1848.7564.0759.91
Tip-Adapter [51]ECCV 202251.0477.7648.8863.4160.27
TPT [36]NeurIPS 202254.7777.0647.9463.4560.81
C-TPT [47]ICLR 202452.9078.0048.5063.4060.70
DART [25]AAAI 202460.5679.5649.7664.0363.48
MTA [49]CVPR 202458.0678.3349.6164.2462.56
TDA [20]CVPR 202460.1180.2450.5464.6763.89
Zero [7]NeuIPS 202461.3577.2848.2964.1362.76
SCAPThis Paper64.5281.6851.6564.6565.63
", + "bbox": [ + 122, + 88, + 874, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results comparison on the out-of-distribution (OOD) Benchmark. The results of the compared methods follow the original paper or are implemented with the official code [25].", + "bbox": [ + 89, + 306, + 906, + 338 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/146f629d2924afda08ff1c390a5025955ec95f3bd2ed7b27c814f7f1780c015b.jpg", + "image_caption": [ + "Figure 3. Results on the Cross-Domain Benchmark. Comparison of SCAP with state-of-the-art methods. All methods are evaluated using CLIP-ViT-B/16 as the backbone." + ], + "image_footnote": [], + "bbox": [ + 98, + 349, + 285, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2405584b2bdab9b03f18c705ee00d940acaaf1ddcf25cef570866618cd0b1bf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 349, + 470, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30ad29488f3bb38f8f8a5250796452380dc534a87f573f09144d6a199d955143.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 349, + 663, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9f79901e069342977655652326c00b30417e3110dea6a0b5d534e4e4df88497.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 349, + 895, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We adopt CLIP with ViT-B/16 as the backbone and set the batch size to 64 for all experiments. Visual attribute prompts are initialized using a uniform distribution in the range $(-1, 1)$ , following previous TTA methods [25, 36]. For text attribute prompts, we employ an ensemble of multiple templates for the OOD benchmark, whereas dataset-specific templates are used for the cross-domain benchmark, following TDA [20]. Our key hyperparameters include the threshold $t$ for supportive clique extraction, the weight $\\alpha_{r}$ for computing $P_C^T$ , the concentration loss weight $\\lambda$ , and graph structure parameters $\\sigma$ , $L$ , and $K$ . The hyperparameter ablations of this paper are conducted on ImageNet-A, with the default values of $\\alpha_{r}$ , $\\sigma$ , $L$ , and $K$ set to 1, 0.3, 6, and 3, respectively. We set $k = 3$ in the initial top-k predictions Eq. (1). An Adam optimizer with a learning rate of 0.003 is used to update the prompts, and all experiments are conducted on a single NVIDIA 4090 GPU.", + "bbox": [ + 88, + 539, + 485, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 818, + 393, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare our proposed SCAP with state-of-the-art TTA methods designed for CLIP adaptation, including TPT [36], C-TPT [47], DART [25], MTA [49], TDA [20], and Zero [7]. Additionally, we report", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "results from the original CLIP model to highlight the knowledge gained during testing. Following previous works [20], we also include training-time adaptation methods such as CoOp [57], CoCoOp [56], and Tip-Adapter [51] for comparison to further verify the label efficiency and effectiveness of our approach.", + "bbox": [ + 511, + 539, + 906, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on the OOD Benchmark. As shown in Tab. 1, our proposed SCAP achieves state-of-the-art performance on the ImageNet-A, ImageNet-R, and ImageNet-S datasets, outperforming existing methods by at least $3.17\\%$ , $1.44\\%$ , and $1.11\\%$ , respectively. Additionally, on the challenging ImageNet-V2 dataset, SCAP performs comparably to TDA, with only a marginal $0.02\\%$ difference. While TDA benefits from a feature cache to enable joint learning across multiple samples, it incurs additional storage overhead. In contrast, SCAP achieves comparable or superior performance without maintaining historical features, making it more efficient in practice.", + "bbox": [ + 511, + 636, + 908, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, SCAP improves average accuracy across the four ImageNet variants by $1.74\\%$ , highlighting its effectiveness in handling natural object images across diverse distribution shifts. This performance gain is primarily attributed to the proposed supportive clique mining strategy, which ef", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30037", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/db48e0c2e6a6290ba8605a27506463020bddd943138f1805d57b3128916bbeed.jpg", + "table_caption": [], + "table_footnote": [ + "- Image and Text represent learning from the Image Modality and Text Modality in the Attribute Prompt Learning module, respectively. \n- Retention refers only to the Retention Cache for the visual modality while the text retention prompts are always used." + ], + "table_body": "
Components in SCAPImageNet-A
ImageTextRetentionAcc@1
XXX60.98
X✓X62.57
✓XX63.14
✓X✓63.82
✓✓✓64.52
", + "bbox": [ + 99, + 88, + 475, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study about the different components", + "bbox": [ + 120, + 210, + 452, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fectively captures diverse object attributes, such as \"head\" and \"body\", that enhance the object recognition capacity of the model.", + "bbox": [ + 89, + 321, + 483, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results over cross-domain benchmark. As shown in Fig. 3, we compare our SCAP method with state-of-the-art approaches across four cross-domain datasets. The results demonstrate that our method surpasses the state-of-the-art approaches, with improvements of at least $1.64\\%$ , $0.23\\%$ , $0.18\\%$ , and $0.24\\%$ on the Pets, Flower102, Caltech101, and Aircraft datasets, respectively. This superior performance stems from the proposed Attribute Prompt Learning and Attribute Prompt Retention mechanisms, which effectively capture and leverage generalizable knowledge shared across diverse samples.", + "bbox": [ + 89, + 369, + 483, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 545, + 254, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Influence of Each Component. To evaluate the effectiveness of our module designs across different modalities (i.e., image, text, and retention), we conduct modality-specific ablation studies. As shown in Tab. 2, leveraging both image and text modalities within the Attribute Prompt Learning module consistently improves performance over the baseline model. Additionally, incorporating the Attribute Prompt Retention module further enhances the utilization of learned image prompts, leading to substantial performance gains. Furthermore, when all modules and modalities are integrated, the model performance improves further due to joint optimization, which mitigates semantic drift as learnable parameters are updated.", + "bbox": [ + 89, + 568, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Influence of Batch Size in SCAP. Batch Size $B$ is an important factor in the transductive TTA setting, as larger batches provide richer inter-image relationships. As illustrated in Fig. 4 (Left), the model performance consistently improves with increasing batch size, demonstrating SCAP's capability to leverage the additional contextual information. We also observe that the image modality varies significantly as batch size changes, while the text modality remains relatively stable across batch size variations. On", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfa5c169664d6e9e4b56830bf72a7e68adcb82969199e98ba9a5404c343feb52.jpg", + "image_caption": [ + "Figure 4. Left: the influence of batch size on SCAP performance. Right: the average maximum Clique size generated per batch under different batch sizes." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 710, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08c1ef4d26770d845c2a2c5180111a406eaf1537cf1c15608278c7073a7778ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 90, + 901, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a95f19d56c8f23680964b96e34d658a869dd6c844d6c751be04a83d40ddf15bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 290, + 717, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/833359702d9529a1910ef1a8050c8f6b16d2bb95d3a1cbc718cf53449fa7b338.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 287, + 901, + 405 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18f79ebf935cbf8dda60b30d5f21ff9636092c14c9457729a2a7fa7cd61513bf.jpg", + "image_caption": [ + "Figure 5. Ablation study about the influence of different hyperparameters on ImageNet-A." + ], + "image_footnote": [], + "bbox": [ + 522, + 411, + 715, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ae297e278b0568dfb5c8e8d8430444865b5fcd9a8c3b082f0bce81d15b57424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 411, + 900, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the right side of Fig. 4, the results reveal that larger batches lead to increasing average clique size, indicating that more support information can be captured under such conditions.", + "bbox": [ + 511, + 597, + 903, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Influence of Different Hyperparameters. The key hyperparameters in our approach include the threshold $t$ for clique extraction, the concentration loss weight $\\lambda$ , the retention cache limit $L$ , and the degree of the retention graph $K$ . As illustrated in Fig. 5 (a), the model performance is relatively sensitive to changes in $t$ . A lower $t$ results in cliques containing unrelated images, which hinders effective attribute prompt learning, whereas a higher $t$ limits the number of support samples, leading to insufficient attribute modeling. Subsequently, we explore the impact of $\\lambda$ in Fig. 5 (b). The model performance improves consistently up to $\\lambda = 1.0$ , as the concentration loss enhances attribute learning. However, further increasing $\\lambda$ leads to overfitting, negatively affecting generalization. Then, we investigate the influence of $L$ in Fig. 5 (c). We observe that as $L$ increases, the performance increases and then becomes stable, indicating that while a larger cache enhances retention,", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30038", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d07aebf70c601d87de011877379514f89052e9dbd3e3114c855169b90d7ca043.jpg", + "image_caption": [ + "Figure 6. Visualization of our prompt-guided attention maps in comparison with the state-of-the-art prompt-based method." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36ddeec2c5066de92857751429bee77e8db6c50f7b7966cbb326c61e6d0ffffb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Combine PIkCombine PI, P*ImageNet-A
MeanConcatMeanConcatAcc@1
✓X✓X61.82
✓XX✓62.11
X✓✓X64.35
X✓X✓64.52
", + "bbox": [ + 99, + 419, + 475, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. The influence of different combination approaches.", + "bbox": [ + 107, + 535, + 464, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "excessive stored knowledge does not yield further gains. Finally, we study the influence of $K$ in Fig. 5 (d). Increasing $K$ up to 2 improves performance by strengthening graph connectivity and enhancing information propagation. However, beyond this point, performance declines due to the inclusion of irrelevant nodes, which introduces noise and reduces the discriminative power of the learned attributes.", + "bbox": [ + 89, + 579, + 482, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence of Different Prompt Combination Approaches. Two primary strategies for combining multiple prompts are considered: generating a mean prompt and concatenating prompts along the token dimension, as illustrated in Eq. (15) and in Eq. (17). Experimental results in Table Tab. 3 indicate that concatenation outperforms averaging in our SCAP framework. This advantage due to concatenation preserves the full prompt information, enabling each prompt to focus on distinct image regions. In contrast, the mean prompt approach struggles to capture diverse visual patterns, leading to suboptimal performance.", + "bbox": [ + 89, + 686, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Supportive Cliques. To demonstrate the effectiveness of our supportive clique-based attribute prompt learning approach, we visualize the attention", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "maps in comparison with the state-of-the-art prompt-based method DART [25]. As illustrated in Fig. 6, each supportive clique in our method guides the prompts to focus on specific object parts, such as the head, leg, or body. When the attribute prompts are exploited together, the obtained attention maps precisely and comprehensively capture the object information, highlighting the superior feature-learning capability of our approach over DART.", + "bbox": [ + 511, + 422, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 570, + 633, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel transductive test-time adaptation approach, Supportive Clique-based Attribute Prompting (SCAP), which leverages intrinsic cross-sample shared visual attributes to enhance model adaptation during testing. Specifically, our method identifies and models the attributes within batches through supportive cliques, sets of samples exhibiting similar patterns. Furthermore, a retention module is designed to dynamically store, aggregate, and utilize knowledge from previously learned attributes, thereby improving predictions for subsequent test images. Extensive experiments on two benchmarks demonstrate the effectiveness of our method in leveraging batch-wise information. The proposed approach offers a promising direction for fully exploiting cross-sample relationships in practical transductive TTA scenarios.", + "bbox": [ + 511, + 599, + 906, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the grants from the National Natural Science Foundation of China (62376011, 61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006).", + "bbox": [ + 511, + 829, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30039", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 2", + "[2] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, pages 295-305, 2022. 2", + "[3] Zhenyu Cui, Yuxin Peng, Xun Wang, Manyu Zhu, and Jiuhuan Zhou. Continual vision-language retrieval via dynamic knowledge rectification. In AAAI, pages 11704-11712, 2024. 2", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5", + "[5] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2", + "[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, pages 7704-7714, 2023. 2", + "[7] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, Elisa Ricci, et al. Frustratingly easy test-time adaptation of vision-language models. In NeurIPS, 2024. 1, 2, 6", + "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5", + "[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, pages 2704-2714, 2023. 2", + "[10] Lihua Fu, Yubin Du, Yu Ding, Dan Wang, Hanxu Jiang, and Haitao Zhang. Domain adaptive learning with multi-granularity features for unsupervised person re-identification. Chinese Journal of Electronics, 31(1):116-128, 2022. 2", + "[11] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 2", + "[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3", + "[13] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. NeurIPS, 36, 2024. 2, 3", + "[14] Gustavo Adolfo Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Clipartt: Light-weight adaptation of clip to new domains at test time. arXiv preprint arXiv:2405.00754, 2024. 2, 3", + "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust-" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5", + "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5", + "[17] Xuefeng Hu, Gokhan Uzunbas, Sirius Chen, Rui Wang, Ashish Shah, Ram Nevatia, and Ser-Nam Lim. Mixnorm: Test-time adaptation through online normalization estimation. arXiv preprint arXiv:2110.11478, 2021. 2", + "[18] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In CVPR, pages 17980-17989, 2022. 2", + "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2", + "[20] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, pages 14162-14171, 2024. 1, 2, 6", + "[21] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2", + "[22] Qiwei Li, Kunlun Xu, Yuxin Peng, and Jiahuan Zhou. Exemplar-free lifelong person re-identification via prompt-guided adaptive knowledge consolidation. IJCV, 132(11): 4850-4865, 2024. 2", + "[23] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. NeurIPS, 36:13448-13466, 2023. 2", + "[24] Zichen Liu, Yuxin Peng, and Jiahuan Zhou. Inspv: Efficient instance visual prompting from image itself. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6443-6452, 2024. 2", + "[25] Zichen Liu, Hongbo Sun, Yuxin Peng, and Jiahuan Zhou. Dart: dual-modal adaptive online prompting and knowledge retention for test-time adaptation. In AAAI, pages 14106-14114, 2024. 1, 2, 3, 6, 8", + "[26] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. NeurIPS, 36, 2024. 3", + "[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[28] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, pages 2555-2565, 2024. 2, 3", + "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30040", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505. IEEE, 2012. 5", + "[31] Ori Press, Steffen Schneider, Matthias Kümmerer, and Matthias Bethge. Rdumb: A simple approach that questions our progress in continual test-time adaptation. NeurIPS, 36: 39915-39935, 2023. 2, 3", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 2", + "[34] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400. PMLR, 2019. 5", + "[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 33:11539-11551, 2020. 3", + "[36] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. NeurIPS, 35:14274-14289, 2022. 2, 3, 6", + "[37] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In CVPR, pages 11920-11929, 2023. 2", + "[38] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR. 2, 3", + "[39] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3", + "[40] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5", + "[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3", + "[42] Kunlun Xu, Haozhuo Zhang, Yu Li, Yuxin Peng, and Jiahuan Zhou. Mitigate catastrophic remembering via continual knowledge purification for noisy lifelong person re-identification. In ACM MM, pages 5790–5799, 2024. 2", + "[43] Kunlun Xu, Xu Zou, Yuxin Peng, and Jiahuan Zhou. Distribution-aware knowledge prototyping for non-exemplar lifelong person re-identification. In CVPR, pages 16604–16613, 2024. 1", + "[44] Kunlun Xu, Xu Zou, and Jiahuan Zhou. Lstkc: Long short-term knowledge consolidation for lifelong person re-identification. In AAAI, pages 16202–16210, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, pages 15671-15680, 2022. 2", + "[46] Zhaoda Ye, Xiangteng He, and Yuxin Peng. Unsupervised cross-media hashing learning via knowledge graph. Chinese Journal of Electronics, 31(6):1081-1091, 2022. 1", + "[47] Hee Suk Yoon, Eunseop Yoon, Joshua Tian Jin Tee, Mark Hasegawa-Johnson, Yingzhen Li, and Chang D Yoo. C-tpt: Calibrated test-time prompt tuning for vision-language models via text feature dispersion. arXiv preprint arXiv:2403.14119, 2024. 3, 6", + "[48] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 2", + "[49] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, pages 23783-23793, 2024. 2, 6", + "[50] Ding-Chu Zhang, Zhi Zhou, and Yu-Feng Li. Robust test-time adaptation for zero-shot prompt tuning. In AAAI, pages 16714-16722, 2024. 2, 3", + "[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, pages 493-510. Springer, 2022. 2, 6", + "[52] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024. 1", + "[53] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. NeurIPS, 16, 2003. 5", + "[54] Jiahuan Zhou, Pei Yu, Wei Tang, and Ying Wu. Efficient online local metric adaptation via negative samples for person re-identification. In ICCV, pages 2420–2428, 2017. 2", + "[55] Jiahuan Zhou, Bing Su, and Ying Wu. Online joint multimetric adaptation from frequent sharing-subset mining for person re-identification. In CVPR, pages 2909–2918, 2020. 1", + "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 6", + "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 756 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30041", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_model.json b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7bd56d6ba07eaa185097382d680f1cb154ff68d --- /dev/null +++ b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_model.json @@ -0,0 +1,2376 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.204, + 0.775, + 0.222 + ], + "angle": 0, + "content": "Chenyu Zhang*, Kunlun Xu*, Zichen Liu, Yuxin Peng, Jiahuan Zhou†" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.224, + 0.763, + 0.24 + ], + "angle": 0, + "content": "Wangxuan Institute of Computer Technology, Peking University, Beijing 100871, China" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.242, + 0.856, + 0.259 + ], + "angle": 0, + "content": "{2200010814, xkl, lzc20180720}@stu.pku.edu.cn, {pengyuxin, jiahuanzhou}@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.485, + 0.779 + ], + "angle": 0, + "content": "Vision-language models (VLMs) encounter considerable challenges when adapting to domain shifts stemming from changes in data distribution. Test-time adaptation (TTA) has emerged as a promising approach to enhance VLM performance under such conditions. In practice, test data often arrives in batches, leading to increasing interest in the transductive TTA setting. However, existing TTA methods primarily focus on individual test samples, overlooking crucial cross-sample correlations within a batch. While recent ViT-based TTA methods have introduced batch-level adaptation, they remain suboptimal for VLMs due to inadequate integration of the text modality. To address these limitations, we propose a novel transductive TTA framework, Supportive Clique-based Attribute Prompting (SCAP), which effectively combines visual and textual information to enhance adaptation by generating fine-grained attribute prompts across test batches. SCAP first forms supportive cliques of test samples in an unsupervised manner based on visual similarity and learns an attribute prompt for each clique, capturing shared attributes critical for adaptation. For each test sample, SCAP aggregates attribute prompts from its associated cliques, providing enriched contextual information. To ensure adaptability over time, we incorporate a retention module that dynamically updates attribute prompts and their associated attributes as new data arrives. Comprehensive experiments across multiple benchmarks demonstrate that SCAP outperforms existing state-of-the-art methods, significantly advancing VLM generalization under domain shifts. Our code is available at https://github.com/zhoujiahuan1991/CVPR2025-SCAP." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.865 + ], + "angle": 0, + "content": "Vision-language models (VLMs) such as CLIP [32] have exhibited remarkable generalization capabilities across a" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.292, + 0.907, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.47, + 0.885, + 0.482 + ], + "angle": 0, + "content": "(a) Current Prompt Learning based Test-time Adaptation Methods" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.484, + 0.905, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.661, + 0.736, + 0.673 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.798 + ], + "angle": 0, + "content": "Figure 1. Comparison between our proposed SCAP to existing prompt learning-based TTA methods. Specifically, current TTA methods learn from instances, while our method utilizes cross-sample visual similarity information from batches to construct supportive cliques and extract attributes from them. We learn attributes from both modalities based on the cliques. For each image, SCAP jointly utilizes attribute prompts from its associated cliques, leading to more effective and accurate prompting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "range of downstream tasks in zero-shot settings. These advancements highlight the significant potential of CLIP in a wide range of domains. However, their performance often degrades [7, 20, 25] when downstream data exhibits substantial domain shifts [43, 46, 52, 55] from the pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.217, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.234, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "training distribution [10, 36, 44, 54]. This limitation has sparked increasing interest in the area of Test-Time Adaptation (TTA) [1, 2, 39, 49], which aims to adapt pre-trained VLMs to mitigate domain shifts during test time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Recent works in TTA have primarily focused on adapting CLIP by independently learning from each sample during test time [20, 25, 36, 49]. However, in real-world scenarios, test-time data typically arrives in batches, creating a transductive TTA setting. Existing TTA methods can tackle such a scenario by treating each test sample individually [25]. Although this solution is available, it suffers from significant limitations due to disregarding the intrinsic cross-sample visual similarity within each batch. Exploiting these inter-sample similarities can greatly enhance both the effectiveness and robustness of the adaptation process. As illustrated in Figure 1, test batches often contain Cliques, groups of visually and semantically similar images that share common characteristics, referred to as Attributes in this paper. Leveraging these cliques enables a more precise and context-aware adaptation by guiding the model to focus on discriminative attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Recent works [13, 17, 28, 31, 38] have explored transductive TTA under the Vision Transformer (ViT) backbone, often relying on batch statistics to adjust normalization parameters or select class-balanced samples for adaptation. However, these methods cannot be directly applied to CLIP since they completely neglect the adaptation of the text modality, which is crucial in VLMs. Rare CLIP-based TTA methods are designed for the transductive setting. In the latest method [14], the top-k predicted classes of each image in a batch are aggregated into a single new text prompt, which is used as pseudo labels to reclassify inputs in a transductive manner. However, the aforementioned transductive TTA methods mainly perform coarse-grained batch-level utilization of the visual similarity information. Thus, the fine-grained discriminative attribute information of cross-sample visual similarity is underutilized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose Supportive Clique-based Attribute Prompting (SCAP), a novel CLIP-based transductive TTA approach designed to effectively leverage the cross-sample visual similarity information. Given a test batch, SCAP automatically mines distinct Supportive Cliques containing test images with shared attributes by leveraging visual and semantic similarity metrics. These cliques play a crucial role in learning Attribute Prompts that capture discriminative, attribute-specific information for each clique. For any given test sample, its aggregated attribute representation is constructed by combining the learned attribute prompts associated with the cliques to which it belongs. By utilizing these aggregated attribute prompts, SCAP enables CLIP to focus on fine-grained, instance-specific details, enhancing prediction accuracy and robustness. Additionally, we introduce a retention mechanism that dynamically up" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "dates and preserves attribute prompts, ensuring long-term adaptability across tasks. Extensive experiments on two benchmarks demonstrate that SCAP significantly outperforms various current state-of-the-art methods in the transductive TTA setting, showing the effectiveness of SCAP in mining the intrinsic knowledge in batches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.41 + ], + "angle": 0, + "content": "In summary, the contributions of this work are fourfold: (1) We propose Supportive Clique-based Prompting (SCAP), a novel approach that leverages the rich information of cross-sample relationships, improving the effectiveness of test-time adaptation under transductive setting. (2) To achieve robust and accurate fine-grained attribute prompt learning during test time, we propose to learn attribute prompts in a transductive manner based on cliques. (3) To retain the historically learned attribute prompts, we further design a retention mechanism to preserve the learned prompts for future reuse. (4) Extensive experiments on two benchmarks demonstrate that SCAP outperforms several current state-of-the-art methods in the transductive TTA setting, highlighting its effectiveness in extracting intrinsic knowledge from batches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.655, + 0.44 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.759, + 0.467 + ], + "angle": 0, + "content": "2.1. Prompt Learning for CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.655 + ], + "angle": 0, + "content": "VLMs like CLIP have demonstrated abundant knowledge learned from large-scale image-text data [5, 18, 19, 32, 33, 45]. Adapting CLIP to various downstream tasks is a valuable and challenging problem [3]. Recently, prompt learning [23, 24, 37, 48, 51] has emerged as a promising approach for enhancing model adaptability. Specifically, CoOp [57], CoCoOp [56], MaPLe [21], and Tip-Adapter [51] have achieved state-of-the-art performance. However, they rely on training data from downstream tasks. However, these approaches rely on training data annotations and often struggle when faced with substantial domain shifts at test time [38, 42]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.719, + 0.683 + ], + "angle": 0, + "content": "2.2. Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Test-time adaptation (TTA) focuses on dynamically adjusting pre-trained models, such as CLIP or ViTs [22, 25], by leveraging sequentially obtained unlabeled test data [9, 11, 38, 49, 50]. Existing TTA approaches can be broadly categorized into training-free, augmentation-based, and prompt-based methods. Training-free methods aim to adapt CLIP efficiently without backpropagation [7, 20]. TDA introduces a positive and negative cache mechanism to assist prediction. Zero [7] aggregates predictions from multiple augmented versions without modifying model parameters. The augmentation-based methods [6, 49] focus on generating diverse augmented samples to integrate test distribution knowledge into the model. MTA [49] introduces a quality assessment variable for each" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "30033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed SCAP. SCAP firstly mines the supportive cliques for all images in parallel. Based on the cliques, it then learns the corresponding attribute prompts from both modalities. The learned visual attribute prompts and the text attribute prompts are separately retained to accumulate the knowledge of test domains. For each instance, we conduct inference by jointly utilizing all the associated attribute prompts and the retained knowledge to generate the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.486, + 0.749 + ], + "angle": 0, + "content": "augmented view and models the distribution using a mixed Gaussian approach. Prompt-based methods employ prompt tuning to adapt the model while preserving generalization capabilities [12, 25, 26, 36, 47, 50]. For example, TPT [36] optimizes instance-specific text prompts by minimizing prediction entropy, whereas DART [25] refines prompts across both modalities while leveraging historically learned knowledge via prompt retention. While these methods assume a sequential data setting where test samples are processed individually [13, 38], real-world scenarios often involve test-time data arriving in batches. Consequently, existing approaches struggle to exploit the inter-sample relationships within a batch, leading to suboptimal adaptation performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.402, + 0.775 + ], + "angle": 0, + "content": "2.3. Transductive Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Transductive TTA considers the practical scenario where test data are provided in batches rather than individual samples. Existing Transductive TTA methods are predominantly developed for ViTs pre-trained on ImageNet [13, 28, 31, 35, 39, 41]. These approaches typically rely on batch statistics to update normalization parameters [39] or select confident samples for model tuning [28]. Recent studies have demonstrated that CLIP outperforms ViTs pre-trained" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.763 + ], + "angle": 0, + "content": "on ImageNet in TTA tasks. However, existing transductive TTA methods fail to fully exploit CLIP's capabilities, as they largely disregard the adaptation of the text modality, which is crucial for leveraging its knowledge learned from vision-language alignment. A few recent efforts have attempted to adapt CLIP for transductive TTA. For example, CLIPArTT [14] optimizes the model by jointly utilizing all images within a batch. However, since cross-instance relationships within a batch are often imbalanced, directly modeling all samples together does not effectively exploit the most informative relationships. In this work, we take a step further by identifying and leveraging the informative cliques, groups of strongly correlated samples within a batch, to facilitate robust adaptation by capturing discriminative attributes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.778, + 0.605, + 0.793 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.787, + 0.819 + ], + "angle": 0, + "content": "3.1. Problem Setting and Notations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In transductive TTA, test data arrive sequentially in batches. Given \\( t \\)-th test batch \\( B_{t} = \\{x_{j}\\}_{j=1}^{b} \\), where \\( b \\) is the batch size and \\( x_{j} \\) is a test image. We generate a set of Supportive Cliques for each class \\( i \\), denoted as \\( \\mathcal{C}_i = \\{C_{ik}\\}_{k=1}^m \\). Then, we extract the corresponding clique attributes, which are de" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30034" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "noted as \\(\\mathcal{A}_i = \\{a_{ik} \\in \\mathbb{R}^d\\}_{k=1}^m\\) where \\(d\\) is the attribute dimension, equivalent to the feature dimension. Additionally, we introduce Attribute Prompts \\(\\mathcal{P}_i^A = \\{\\mathcal{P}_i^I, \\mathcal{P}_i^T\\}\\), which consist of visual attribute prompts \\(\\mathcal{P}_i^I = \\{P_{ik}^I\\}_{k=1}^m\\) and text attribute prompts \\(\\mathcal{P}_i^T = \\{P_{ik}^T\\}_{k=1}^m\\). Our model \\(\\theta\\) is built on CLIP, which comprises an image encoder \\(\\mathbf{E}_I\\) and a text encoder \\(\\mathbf{E}_T\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.204, + 0.465, + 0.222 + ], + "angle": 0, + "content": "3.2. Attribute-Driven Supportive Clique Mining" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.484, + 0.317 + ], + "angle": 0, + "content": "Attributes are the fine-grained characteristics shared by a group of images in the batch. Since images containing the same attribute often exhibit high visual similarity to each other, in this section, we propose to obtain Supportive Cliques for the potential attributes by evaluating cross-image similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.378 + ], + "angle": 0, + "content": "Specifically, we use CLIP to obtain the top-\\(k\\) class predictions \\(topk(x_j)\\) for each image \\(x_j\\), where \\(k\\) is a hyperparameter. For each class \\(i\\), all images whose top-\\(k\\) predictions include class \\(i\\) are selected to form a subset \\(S_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.384, + 0.483, + 0.401 + ], + "angle": 0, + "content": "\\[\nS _ {i} = \\left\\{x _ {j} \\in B _ {t}: i \\in \\operatorname {t o p k} \\left(x _ {j}\\right) \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.493 + ], + "angle": 0, + "content": "Since images in \\(S_{i}\\) share similar class-specific information, which ensures semantic relevance. To further capture more fine-grained relevance, we extract the features \\(F_{i}^{I}\\in\\) \\(\\mathbb{R}^{|S_i|\\times d}\\) of images within \\(S_{i}\\) using the CLIP image encoder \\(\\mathbf{E}_I\\) and compute the similarity matrix \\(M_{i}\\in \\mathbb{R}^{|S_{i}|\\times |S_{i}|}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.493, + 0.482, + 0.513 + ], + "angle": 0, + "content": "\\[\nM _ {i} = F _ {i} ^ {I} \\cdot \\left(F _ {i} ^ {I}\\right) ^ {\\top} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.576 + ], + "angle": 0, + "content": "Then, given the \\(l\\)-th row of \\(M_{i}\\), we collect elements exceeding a threshold \\(t\\), whose corresponding images form a Clique \\(C_{il}\\). Formally, the Clique generated from the \\(l\\)-th row of \\(M_{i}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.59, + 0.482, + 0.609 + ], + "angle": 0, + "content": "\\[\nC _ {i l} = \\left\\{x _ {j} ^ {i}: x _ {j} ^ {i} \\in S _ {i}, (M _ {i}) _ {l j} > t \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.673 + ], + "angle": 0, + "content": "where \\((M_{i})_{lj}\\) represents the \\(l\\)-th row, \\(j\\)-th column of \\(M_{i}\\). Only non-trivial Cliques \\((|C_{il}| > 1)\\) are retained, and duplicate Cliques are removed to ensure uniqueness. Finally, we obtain a set of valid Cliques \\(\\mathcal{C}_i = \\{C_{ik}\\}_{k=1}^m\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.338, + 0.695 + ], + "angle": 0, + "content": "3.3. Attribute Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.73 + ], + "angle": 0, + "content": "In this section, we aim to learn prompts that capture attribute-specific knowledge to facilitate classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.793 + ], + "angle": 0, + "content": "Specifically, for each Clique \\(C_{ik}\\), we learn a visual attribute prompt \\(P_{ik}^{A}\\) from the image modality and a text attribute prompt \\(P_{ik}^{T}\\) from the text modality. The feature \\(f_{j}^{I} \\in \\mathbb{R}^{d}\\) for the \\(j\\)-th image in \\(C_{ik}\\) is extracted as:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.798, + 0.483, + 0.818 + ], + "angle": 0, + "content": "\\[\nf _ {j} ^ {I} = \\mathbf {E} _ {I} \\left(C _ {i k} ^ {j}, P _ {k} ^ {I}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Then, we define the attribute feature \\(a_{ik} \\in \\mathbb{R}^d\\) of \\(C_{ik}\\) which is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.86, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\na _ {i k} = \\frac {1}{\\left| C _ {i k} \\right|} \\sum_ {j = 1} ^ {\\left| C _ {i k} \\right|} f _ {j} ^ {I} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.906, + 0.121 + ], + "angle": 0, + "content": "Besides, we compute the text feature \\( f_{n}^{T} \\) of the \\( n \\)-th class as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.135, + 0.905, + 0.154 + ], + "angle": 0, + "content": "\\[\nf _ {n} ^ {T} = \\mathbf {E} _ {T} \\left(c _ {n}, P _ {i k} ^ {T}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.159, + 0.906, + 0.19 + ], + "angle": 0, + "content": "Given total \\(N\\) classes, we introduce an attribute-specific class probability distribution vector, which is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.199, + 0.905, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {P}} _ {i k} = \\left[ \\hat {p} _ {i k} ^ {1}, \\hat {p} _ {i k} ^ {2}, \\dots , \\hat {p} _ {i k} ^ {N} \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.56, + 0.241 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.238, + 0.905, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {i k} ^ {n} = \\frac {\\exp \\left(\\cos \\left(a _ {i k} , f _ {n} ^ {T}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(a _ {i k} , f _ {j} ^ {T}\\right) / \\tau\\right)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.283, + 0.906, + 0.314 + ], + "angle": 0, + "content": "Note that \\(\\cos (\\cdot ,\\cdot)\\) denotes cosine similarity and \\(\\tau\\) is a temperature parameter to scale the similarity scores." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.314, + 0.906, + 0.343 + ], + "angle": 0, + "content": "Then, we compute the Entropy Loss for the probability distribution of the attribute:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.351, + 0.905, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {e n} \\left(C _ {i k}\\right) = - \\sum_ {n = 1} ^ {N} \\hat {p} _ {i k} ^ {n} \\log \\hat {p} _ {i k} ^ {n} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.402, + 0.906, + 0.48 + ], + "angle": 0, + "content": "To further guide the model to extract attribute-specific information and discard attribute-irrelevant information, we introduce a concentration loss \\(\\mathcal{L}_{con}(C_{ik})\\) that forces the model to capture the shared information across the samples within a Clique. Specifically, \\(\\mathcal{L}_{con}(C_{ik})\\) is calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.488, + 0.905, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} \\left(C _ {i k}\\right) = \\sum_ {j = 1} ^ {| C _ {i k} |} \\left| \\left| f _ {j} ^ {I} - a _ {i k} \\right| \\right| ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.906, + 0.606 + ], + "angle": 0, + "content": "Therefore, both kinds of prompts \\( P_{ik}^{I} \\) and \\( P_{ik}^{T} \\) are guided to encode attribute-specific knowledge as the model optimization. Such knowledge is valuable when generating predictions for images showing the same attributes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.614, + 0.766, + 0.631 + ], + "angle": 0, + "content": "3.4. Attribute Prompt Retention" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.636, + 0.906, + 0.697 + ], + "angle": 0, + "content": "In this section, to retain attribute knowledge during the test process and ensure consistent performance improvement, we introduce mechanisms for retaining both text and visual attribute knowledge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.697, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Firstly, to retain the text attribute knowledge, we propose a Text Retention Prompt \\(\\tilde{P}^T\\). Specifically, when a new Clique \\(C_{ik}\\) is learned, we update \\(\\tilde{P}^T\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.751, + 0.905, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\tilde {P} ^ {T} = \\alpha_ {\\tau} \\tilde {P} ^ {T} + (1 - \\alpha_ {\\tau}) P _ {i k} ^ {T}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.906, + 0.809 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\tau} = \\tau /(1 + \\tau)\\) is a decay rate based on iteration count \\(\\tau\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Besides, to retain the visual attribute knowledge, we design a Retention Cache \\(\\mathcal{R} = \\{R_i\\}_{i=1}^N\\) where \\(R_i\\) is a class-specific retention cache. Specifically, we store \\((a_{ik}, P_{ik}^I)\\) as key-value pairs in \\(R_i\\). We denote the retained attributes (i.e. keys) as \\(R_i^A = \\{a_1, \\ldots, a_L\\}\\) and the corresponding attribute prompts (i.e. values) as \\(R_i^P = \\{P_1^I, \\ldots, P_L^I\\}\\). The" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30035" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "maximum size of each \\( R_i \\) is set to \\( L \\) to avoid memory explosion. When the retained key-value pair number is less than \\( L \\), we directly add a new \\( (a_{ik}, P_{ik}^I) \\) to \\( R_i \\). Otherwise, we add \\( (a_{ik}, P_{ik}^I) \\) to \\( R_i \\) first, then we update all pairs and aggregate the two that are most similar to each other in \\( R_i \\) by using a graph structure as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.183, + 0.484, + 0.273 + ], + "angle": 0, + "content": "Firstly, we construct a weighted graph \\( G_{i} = (V_{i},E_{i}) \\). The vertices are the keys in the retention cache \\( V_{i} = R_{i}^{A} = \\{a_{1},\\dots ,a_{L + 1}\\} \\). The edges \\( E_{i} \\) between each two vertices are defined as the Gaussian Distance of the two attributes. Thus, the adjacency matrix \\( W_{i}\\in \\mathbb{R}^{(L + 1)\\times (L + 1)} \\) of \\( G_{i} \\) is defined with:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.282, + 0.483, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\left(W _ {i}\\right) _ {k l} = \\exp \\left(- \\frac {\\left\\| a _ {k} - a _ {l} \\right\\| ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.448 + ], + "angle": 0, + "content": "where \\((W_{i})_{kl}\\) is the \\(k\\)-th row, \\(l\\)-column of \\(W_{i}\\) and \\(\\sigma\\) is a hyperparameter. We utilize the \\(K\\)-max values in each row of \\(W_{i}\\) to construct a \\(K\\)-nearest retention graph. Then we perform graph propagation following [53] and obtain the updated graph vertices \\(V_{i}^{\\prime} = \\{a_{1}^{\\prime},\\ldots ,a_{L + 1}^{\\prime}\\}\\), where the similar attributes are pushed closer automatically. Then, the most similar two attributes \\(a_{j_0}^\\prime ,a_{j_1}^\\prime\\) and the corresponding prompts are fused into one respectively as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.458, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} a _ {n e w} = \\frac {1}{2} \\left(a _ {j _ {0}} ^ {\\prime} + a _ {j _ {1}} ^ {\\prime}\\right) \\\\ P _ {n e w} ^ {I} = \\frac {1}{2} \\left(P _ {j _ {0}} ^ {I} + P _ {j _ {1}} ^ {I}\\right) \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.589 + ], + "angle": 0, + "content": "Through the above process, we obtain the updated retention cache \\( R_{i} \\), dynamically retaining the visual and text attribute knowledge learned from previous batches into Retention Cache and Text Retention Prompt, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.599, + 0.34, + 0.615 + ], + "angle": 0, + "content": "3.5. Optimization and Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.668 + ], + "angle": 0, + "content": "For each batch \\( B_{t} \\), we optimize the visual attribute prompts \\( \\mathcal{P}_i^I = \\{P_{ik}^I\\}_{k=1}^m \\) and text attribute prompts \\( \\mathcal{P}_i^T = \\{P_{ik}^T\\}_{k=1}^m \\) for every class \\( i \\) by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.679, + 0.483, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {k = 1} ^ {m} \\mathcal {L} _ {e n} \\left(C _ {i k}\\right) + \\lambda \\mathcal {L} _ {c o n} \\left(C _ {i k}\\right) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.725, + 0.285, + 0.74 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.484, + 0.844 + ], + "angle": 0, + "content": "Then we conduct inference for each \\( S_{i} \\) of \\( B_{t} \\). To fully utilize both the Attribute Prompts learned from the current batch \\( B_{t} \\) and the retained knowledge to improve prediction. Given the test image \\( x_{j} \\in S_{i} \\), we first collect all Cliques containing \\( x_{j} \\), namely \\( \\mathcal{K} = \\{k : x_{j} \\in C_{ik}\\} \\), whose corresponding visual attribute prompts are concatenated together:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.845, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nP ^ {I} = \\operatorname {c o n c a t} \\left(\\left[ P _ {i k} ^ {I}: k \\in \\mathcal {K} \\right]\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then, we introduce the retention prompt matching strategy utilizing the key-value pair structure of the retention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.184 + ], + "angle": 0, + "content": "cache. Specifically, the keys are the retained attribute features \\( R_{i}^{A} = \\{a_{1},\\dots ,a_{L}\\} \\) and the values are the corresponding attribute prompts \\( R_{i}^{P} = \\{P_{1}^{I},\\dots ,P_{L}^{I}\\} \\). Given the image feature \\( f_{j}^{I} \\) of \\( x_{j} \\) which is already extracted in Sec. 3.3, the top-1 attribute \\( a^* \\) is obtained by evaluating the similarity of \\( f_{j}^{I} \\) to all keys in \\( R_{i}^{A} \\), i.e." + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.192, + 0.905, + 0.222 + ], + "angle": 0, + "content": "\\[\na ^ {*} = \\underset {a _ {l} \\in R _ {i} ^ {A}} {\\operatorname {a r g m a x}} < f _ {j} ^ {I}, a _ {l} > \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.905, + 0.276 + ], + "angle": 0, + "content": "Given the matched value corresponding to \\( a^* \\), i.e., \\( P^{I*} \\), we obtain the Composed Image Prompt \\( P_C^I \\) by concatenating \\( P^I \\) and \\( P^{I*} \\) along the token dimension" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.283, + 0.905, + 0.302 + ], + "angle": 0, + "content": "\\[\nP _ {C} ^ {I} = \\operatorname {c o n c a t} \\left(P ^ {I}, P ^ {I ^ {*}}\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.905, + 0.341 + ], + "angle": 0, + "content": "For the text modality, we fuse \\(\\mathcal{P}^{\\mathcal{T}} = \\{P_{ik}^{T}\\}_{k=1}^{m}\\) and \\(\\tilde{P}^T\\) together to get the Composed Text Prompt \\(P_C^T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "\\[\nP _ {C} ^ {T} = \\alpha_ {r} \\tilde {P} ^ {T} + (1 - \\alpha_ {r}) \\frac {1}{m} \\sum_ {k = 1} ^ {m} P _ {i k} ^ {T} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.457 + ], + "angle": 0, + "content": "where \\(\\alpha_{r}\\) is the weight of retained text prompts. Then, for the final prediction, the image-inference feature \\(\\hat{f}_j^I\\) of \\(x_j^i\\) and the text-inference feature \\(\\hat{f}_n^T\\) of class \\(c_{n}\\) are obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.457, + 0.905, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\hat {f} _ {j} ^ {I} = \\mathbf {E} _ {I} \\left(x _ {j}, P _ {C} ^ {I}\\right) \\\\ \\hat {f} _ {n} ^ {T} = \\mathbf {E} _ {T} \\left(c _ {n}, P _ {C} ^ {T}\\right) \\end{array} \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.501, + 0.885, + 0.518 + ], + "angle": 0, + "content": "The prediction \\( y_{j} \\) of \\( x_{j} \\) is obtained by applying Eq.7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.646, + 0.545 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.674 + ], + "angle": 0, + "content": "In this section, we first introduce the two benchmarks used for model evaluation, followed by a detailed description of the implementation setup. Next, we present a comparative analysis of existing methods and our proposed approach. Finally, extensive ablation studies are conducted to validate the effectiveness of SCAP. Additional experimental results and in-depth discussions are provided in the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.649, + 0.697 + ], + "angle": 0, + "content": "4.1. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our evaluation is conducted on two benchmarks: an out-of-distribution (OOD) benchmark and a cross-domain benchmark. The OOD benchmark consists of four datasets derived from ImageNet [4], i.e., ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [40], and ImageNet-V2 [34]. The OOD benchmark aims to assess the model's robustness to substantial distribution shifts in the test data. The cross-domain benchmark comprises four datasets of the image classification task, each representing a distinct domain: Aircraft [27], Caltech101 [8], Flower102 [29], and Pets [30]. The cross-domain benchmark provides a measure of the model's adaptability across a diverse range of domain-specific tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30036" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.875, + 0.299 + ], + "angle": 0, + "content": "
MethodPublicationImageNet-AImageNet-RImageNet-SImageNet-V2Average
CLIP [32]ICML 202147.8773.9846.0961.8857.46
CoOp [57]IJCV 202249.7175.2147.9964.2059.28
CoCoOp [56]CVPR 202250.6376.1848.7564.0759.91
Tip-Adapter [51]ECCV 202251.0477.7648.8863.4160.27
TPT [36]NeurIPS 202254.7777.0647.9463.4560.81
C-TPT [47]ICLR 202452.9078.0048.5063.4060.70
DART [25]AAAI 202460.5679.5649.7664.0363.48
MTA [49]CVPR 202458.0678.3349.6164.2462.56
TDA [20]CVPR 202460.1180.2450.5464.6763.89
Zero [7]NeuIPS 202461.3577.2848.2964.1362.76
SCAPThis Paper64.5281.6851.6564.6565.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.308, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Table 1. Results comparison on the out-of-distribution (OOD) Benchmark. The results of the compared methods follow the original paper or are implemented with the official code [25]." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.351, + 0.287, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.351, + 0.472, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.351, + 0.665, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.351, + 0.897, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.485, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 3. Results on the Cross-Domain Benchmark. Comparison of SCAP with state-of-the-art methods. All methods are evaluated using CLIP-ViT-B/16 as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.486, + 0.812 + ], + "angle": 0, + "content": "Implementation details. We adopt CLIP with ViT-B/16 as the backbone and set the batch size to 64 for all experiments. Visual attribute prompts are initialized using a uniform distribution in the range \\((-1, 1)\\), following previous TTA methods [25, 36]. For text attribute prompts, we employ an ensemble of multiple templates for the OOD benchmark, whereas dataset-specific templates are used for the cross-domain benchmark, following TDA [20]. Our key hyperparameters include the threshold \\(t\\) for supportive clique extraction, the weight \\(\\alpha_{r}\\) for computing \\(P_C^T\\), the concentration loss weight \\(\\lambda\\), and graph structure parameters \\(\\sigma\\), \\(L\\), and \\(K\\). The hyperparameter ablations of this paper are conducted on ImageNet-A, with the default values of \\(\\alpha_{r}\\), \\(\\sigma\\), \\(L\\), and \\(K\\) set to 1, 0.3, 6, and 3, respectively. We set \\(k = 3\\) in the initial top-k predictions Eq. (1). An Adam optimizer with a learning rate of 0.003 is used to update the prompts, and all experiments are conducted on a single NVIDIA 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.394, + 0.835 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Compared methods. We compare our proposed SCAP with state-of-the-art TTA methods designed for CLIP adaptation, including TPT [36], C-TPT [47], DART [25], MTA [49], TDA [20], and Zero [7]. Additionally, we report" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.632 + ], + "angle": 0, + "content": "results from the original CLIP model to highlight the knowledge gained during testing. Following previous works [20], we also include training-time adaptation methods such as CoOp [57], CoCoOp [56], and Tip-Adapter [51] for comparison to further verify the label efficiency and effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.909, + 0.819 + ], + "angle": 0, + "content": "Results on the OOD Benchmark. As shown in Tab. 1, our proposed SCAP achieves state-of-the-art performance on the ImageNet-A, ImageNet-R, and ImageNet-S datasets, outperforming existing methods by at least \\(3.17\\%\\), \\(1.44\\%\\), and \\(1.11\\%\\), respectively. Additionally, on the challenging ImageNet-V2 dataset, SCAP performs comparably to TDA, with only a marginal \\(0.02\\%\\) difference. While TDA benefits from a feature cache to enable joint learning across multiple samples, it incurs additional storage overhead. In contrast, SCAP achieves comparable or superior performance without maintaining historical features, making it more efficient in practice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Overall, SCAP improves average accuracy across the four ImageNet variants by \\(1.74\\%\\), highlighting its effectiveness in handling natural object images across diverse distribution shifts. This performance gain is primarily attributed to the proposed supportive clique mining strategy, which ef" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "30037" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.477, + 0.201 + ], + "angle": 0, + "content": "
Components in SCAPImageNet-A
ImageTextRetentionAcc@1
XXX60.98
X✓X62.57
✓XX63.14
✓X✓63.82
✓✓✓64.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.211, + 0.454, + 0.225 + ], + "angle": 0, + "content": "Table 2. Ablation study about the different components" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.235, + 0.464, + 0.285 + ], + "angle": 0, + "content": "- Image and Text represent learning from the Image Modality and Text Modality in the Attribute Prompt Learning module, respectively. \n- Retention refers only to the Retention Cache for the visual modality while the text retention prompts are always used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.367 + ], + "angle": 0, + "content": "fectively captures diverse object attributes, such as \"head\" and \"body\", that enhance the object recognition capacity of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Results over cross-domain benchmark. As shown in Fig. 3, we compare our SCAP method with state-of-the-art approaches across four cross-domain datasets. The results demonstrate that our method surpasses the state-of-the-art approaches, with improvements of at least \\(1.64\\%\\), \\(0.23\\%\\), \\(0.18\\%\\), and \\(0.24\\%\\) on the Pets, Flower102, Caltech101, and Aircraft datasets, respectively. This superior performance stems from the proposed Attribute Prompt Learning and Attribute Prompt Retention mechanisms, which effectively capture and leverage generalizable knowledge shared across diverse samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.255, + 0.56 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.764 + ], + "angle": 0, + "content": "The Influence of Each Component. To evaluate the effectiveness of our module designs across different modalities (i.e., image, text, and retention), we conduct modality-specific ablation studies. As shown in Tab. 2, leveraging both image and text modalities within the Attribute Prompt Learning module consistently improves performance over the baseline model. Additionally, incorporating the Attribute Prompt Retention module further enhances the utilization of learned image prompts, leading to substantial performance gains. Furthermore, when all modules and modalities are integrated, the model performance improves further due to joint optimization, which mitigates semantic drift as learnable parameters are updated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The Influence of Batch Size in SCAP. Batch Size \\( B \\) is an important factor in the transductive TTA setting, as larger batches provide richer inter-image relationships. As illustrated in Fig. 4 (Left), the model performance consistently improves with increasing batch size, demonstrating SCAP's capability to leverage the additional contextual information. We also observe that the image modality varies significantly as batch size changes, while the text modality remains relatively stable across batch size variations. On" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.092, + 0.712, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.092, + 0.902, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.27 + ], + "angle": 0, + "content": "Figure 4. Left: the influence of batch size on SCAP performance. Right: the average maximum Clique size generated per batch under different batch sizes." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.291, + 0.718, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.289, + 0.902, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.412, + 0.717, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.412, + 0.901, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.542, + 0.905, + 0.569 + ], + "angle": 0, + "content": "Figure 5. Ablation study about the influence of different hyperparameters on ImageNet-A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.905, + 0.643 + ], + "angle": 0, + "content": "the right side of Fig. 4, the results reveal that larger batches lead to increasing average clique size, indicating that more support information can be captured under such conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The Influence of Different Hyperparameters. The key hyperparameters in our approach include the threshold \\( t \\) for clique extraction, the concentration loss weight \\( \\lambda \\), the retention cache limit \\( L \\), and the degree of the retention graph \\( K \\). As illustrated in Fig. 5 (a), the model performance is relatively sensitive to changes in \\( t \\). A lower \\( t \\) results in cliques containing unrelated images, which hinders effective attribute prompt learning, whereas a higher \\( t \\) limits the number of support samples, leading to insufficient attribute modeling. Subsequently, we explore the impact of \\( \\lambda \\) in Fig. 5 (b). The model performance improves consistently up to \\( \\lambda = 1.0 \\), as the concentration loss enhances attribute learning. However, further increasing \\( \\lambda \\) leads to overfitting, negatively affecting generalization. Then, we investigate the influence of \\( L \\) in Fig. 5 (c). We observe that as \\( L \\) increases, the performance increases and then becomes stable, indicating that while a larger cache enhances retention," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.382, + 0.86, + 0.397 + ], + "angle": 0, + "content": "Figure 6. Visualization of our prompt-guided attention maps in comparison with the state-of-the-art prompt-based method." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.42, + 0.477, + 0.526 + ], + "angle": 0, + "content": "
Combine PIkCombine PI, P*ImageNet-A
MeanConcatMeanConcatAcc@1
✓X✓X61.82
✓XX✓62.11
X✓✓X64.35
X✓X✓64.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.536, + 0.465, + 0.55 + ], + "angle": 0, + "content": "Table 3. The influence of different combination approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.686 + ], + "angle": 0, + "content": "excessive stored knowledge does not yield further gains. Finally, we study the influence of \\( K \\) in Fig. 5 (d). Increasing \\( K \\) up to 2 improves performance by strengthening graph connectivity and enhancing information propagation. However, beyond this point, performance declines due to the inclusion of irrelevant nodes, which introduces noise and reduces the discriminative power of the learned attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Influence of Different Prompt Combination Approaches. Two primary strategies for combining multiple prompts are considered: generating a mean prompt and concatenating prompts along the token dimension, as illustrated in Eq. (15) and in Eq. (17). Experimental results in Table Tab. 3 indicate that concatenation outperforms averaging in our SCAP framework. This advantage due to concatenation preserves the full prompt information, enabling each prompt to focus on distinct image regions. In contrast, the mean prompt approach struggles to capture diverse visual patterns, leading to suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Visualization of Supportive Cliques. To demonstrate the effectiveness of our supportive clique-based attribute prompt learning approach, we visualize the attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.907, + 0.545 + ], + "angle": 0, + "content": "maps in comparison with the state-of-the-art prompt-based method DART [25]. As illustrated in Fig. 6, each supportive clique in our method guides the prompts to focus on specific object parts, such as the head, leg, or body. When the attribute prompts are exploited together, the obtained attention maps precisely and comprehensively capture the object information, highlighting the superior feature-learning capability of our approach over DART." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.634, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.827 + ], + "angle": 0, + "content": "In this paper, we propose a novel transductive test-time adaptation approach, Supportive Clique-based Attribute Prompting (SCAP), which leverages intrinsic cross-sample shared visual attributes to enhance model adaptation during testing. Specifically, our method identifies and models the attributes within batches through supportive cliques, sets of samples exhibiting similar patterns. Furthermore, a retention module is designed to dynamically store, aggregate, and utilize knowledge from previously learned attributes, thereby improving predictions for subsequent test images. Extensive experiments on two benchmarks demonstrate the effectiveness of our method in leveraging batch-wise information. The proposed approach offers a promising direction for fully exploiting cross-sample relationships in practical transductive TTA scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.83, + 0.909, + 0.888 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the grants from the National Natural Science Foundation of China (62376011, 61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30039" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, pages 295-305, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Zhenyu Cui, Yuxin Peng, Xun Wang, Manyu Zhu, and Jiuhuan Zhou. Continual vision-language retrieval via dynamic knowledge rectification. In AAAI, pages 11704-11712, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[5] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, pages 7704-7714, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[7] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, Elisa Ricci, et al. Frustratingly easy test-time adaptation of vision-language models. In NeurIPS, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, pages 2704-2714, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.56, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[10] Lihua Fu, Yubin Du, Yu Ding, Dan Wang, Hanxu Jiang, and Haitao Zhang. Domain adaptive learning with multi-granularity features for unsupervised person re-identification. Chinese Journal of Electronics, 31(1):116-128, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[11] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[13] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. NeurIPS, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Gustavo Adolfo Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Clipartt: Light-weight adaptation of clip to new domains at test time. arXiv preprint arXiv:2405.00754, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[17] Xuefeng Hu, Gokhan Uzunbas, Sirius Chen, Rui Wang, Ashish Shah, Ram Nevatia, and Ser-Nam Lim. Mixnorm: Test-time adaptation through online normalization estimation. arXiv preprint arXiv:2110.11478, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[18] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In CVPR, pages 17980-17989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.279, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[20] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, pages 14162-14171, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[21] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.904, + 0.516 + ], + "angle": 0, + "content": "[22] Qiwei Li, Kunlun Xu, Yuxin Peng, and Jiahuan Zhou. Exemplar-free lifelong person re-identification via prompt-guided adaptive knowledge consolidation. IJCV, 132(11): 4850-4865, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[23] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. NeurIPS, 36:13448-13466, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[24] Zichen Liu, Yuxin Peng, and Jiahuan Zhou. Inspv: Efficient instance visual prompting from image itself. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6443-6452, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[25] Zichen Liu, Hongbo Sun, Yuxin Peng, and Jiahuan Zhou. Dart: dual-modal adaptive online prompting and knowledge retention for test-time adaptation. In AAAI, pages 14106-14114, 2024. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[26] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. NeurIPS, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[28] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, pages 2555-2565, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30040" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[30] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In \\(CVPR\\), pages 3498-3505. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[31] Ori Press, Steffen Schneider, Matthias Kümmerer, and Matthias Bethge. Rdumb: A simple approach that questions our progress in continual test-time adaptation. NeurIPS, 36: 39915-39935, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[34] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 33:11539-11551, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[36] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. NeurIPS, 35:14274-14289, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[37] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In CVPR, pages 11920-11929, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[38] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[39] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[40] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[42] Kunlun Xu, Haozhuo Zhang, Yu Li, Yuxin Peng, and Jiahuan Zhou. Mitigate catastrophic remembering via continual knowledge purification for noisy lifelong person re-identification. In ACM MM, pages 5790–5799, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[43] Kunlun Xu, Xu Zou, Yuxin Peng, and Jiahuan Zhou. Distribution-aware knowledge prototyping for non-exemplar lifelong person re-identification. In CVPR, pages 16604–16613, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Kunlun Xu, Xu Zou, and Jiahuan Zhou. Lstkc: Long short-term knowledge consolidation for lifelong person re-identification. In AAAI, pages 16202–16210, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[45] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, pages 15671-15680, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Zhaoda Ye, Xiangteng He, and Yuxin Peng. Unsupervised cross-media hashing learning via knowledge graph. Chinese Journal of Electronics, 31(6):1081-1091, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[47] Hee Suk Yoon, Eunseop Yoon, Joshua Tian Jin Tee, Mark Hasegawa-Johnson, Yingzhen Li, and Chang D Yoo. C-tpt: Calibrated test-time prompt tuning for vision-language models via text feature dispersion. arXiv preprint arXiv:2403.14119, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[48] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[49] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, pages 23783-23793, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[50] Ding-Chu Zhang, Zhi Zhou, and Yu-Feng Li. Robust test-time adaptation for zero-shot prompt tuning. In AAAI, pages 16714-16722, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, pages 493-510. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[52] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[53] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. NeurIPS, 16, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[54] Jiahuan Zhou, Pei Yu, Wei Tang, and Ying Wu. Efficient online local metric adaptation via negative samples for person re-identification. In ICCV, pages 2420–2428, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[55] Jiahuan Zhou, Bing Su, and Ying Wu. Online joint multimetric adaptation from frequent sharing-subset mining for person re-identification. In CVPR, pages 2909–2918, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30041" + } + ] +] \ No newline at end of file diff --git a/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_origin.pdf b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dbfeeea083ff1e77eab06365779ef1685e615389 --- /dev/null +++ b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/f1fd7643-840d-469e-b4c4-32935040de8e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb3b9dc42af5f2f4fe4c605f98ca07c2297876e7f1941603581213a4bafa7563 +size 4711912 diff --git a/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/full.md b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb764353cc9014a50a636cc39a19ce006c3295ea --- /dev/null +++ b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/full.md @@ -0,0 +1,366 @@ +# SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting + +Chenyu Zhang*, Kunlun Xu*, Zichen Liu, Yuxin Peng, Jiahuan Zhou† + +Wangxuan Institute of Computer Technology, Peking University, Beijing 100871, China + +{2200010814, xkl, lzc20180720}@stu.pku.edu.cn, {pengyuxin, jiahuanzhou}@pku.edu.cn + +# Abstract + +Vision-language models (VLMs) encounter considerable challenges when adapting to domain shifts stemming from changes in data distribution. Test-time adaptation (TTA) has emerged as a promising approach to enhance VLM performance under such conditions. In practice, test data often arrives in batches, leading to increasing interest in the transductive TTA setting. However, existing TTA methods primarily focus on individual test samples, overlooking crucial cross-sample correlations within a batch. While recent ViT-based TTA methods have introduced batch-level adaptation, they remain suboptimal for VLMs due to inadequate integration of the text modality. To address these limitations, we propose a novel transductive TTA framework, Supportive Clique-based Attribute Prompting (SCAP), which effectively combines visual and textual information to enhance adaptation by generating fine-grained attribute prompts across test batches. SCAP first forms supportive cliques of test samples in an unsupervised manner based on visual similarity and learns an attribute prompt for each clique, capturing shared attributes critical for adaptation. For each test sample, SCAP aggregates attribute prompts from its associated cliques, providing enriched contextual information. To ensure adaptability over time, we incorporate a retention module that dynamically updates attribute prompts and their associated attributes as new data arrives. Comprehensive experiments across multiple benchmarks demonstrate that SCAP outperforms existing state-of-the-art methods, significantly advancing VLM generalization under domain shifts. Our code is available at https://github.com/zhoujiahuan1991/CVPR2025-SCAP. + +# 1. Introduction + +Vision-language models (VLMs) such as CLIP [32] have exhibited remarkable generalization capabilities across a + +![](images/3663c2c26e042dcd2e24ccfc41b98c255b2f5b41b5d7fde59ccfba92673f6a61.jpg) +(a) Current Prompt Learning based Test-time Adaptation Methods + +![](images/c2ecaa5d2a555735b2357b754ef3dd69353a1009afc91b7d5c969ad16e0ac72d.jpg) +(b) Ours +Figure 1. Comparison between our proposed SCAP to existing prompt learning-based TTA methods. Specifically, current TTA methods learn from instances, while our method utilizes cross-sample visual similarity information from batches to construct supportive cliques and extract attributes from them. We learn attributes from both modalities based on the cliques. For each image, SCAP jointly utilizes attribute prompts from its associated cliques, leading to more effective and accurate prompting. + +range of downstream tasks in zero-shot settings. These advancements highlight the significant potential of CLIP in a wide range of domains. However, their performance often degrades [7, 20, 25] when downstream data exhibits substantial domain shifts [43, 46, 52, 55] from the pre + +training distribution [10, 36, 44, 54]. This limitation has sparked increasing interest in the area of Test-Time Adaptation (TTA) [1, 2, 39, 49], which aims to adapt pre-trained VLMs to mitigate domain shifts during test time. + +Recent works in TTA have primarily focused on adapting CLIP by independently learning from each sample during test time [20, 25, 36, 49]. However, in real-world scenarios, test-time data typically arrives in batches, creating a transductive TTA setting. Existing TTA methods can tackle such a scenario by treating each test sample individually [25]. Although this solution is available, it suffers from significant limitations due to disregarding the intrinsic cross-sample visual similarity within each batch. Exploiting these inter-sample similarities can greatly enhance both the effectiveness and robustness of the adaptation process. As illustrated in Figure 1, test batches often contain Cliques, groups of visually and semantically similar images that share common characteristics, referred to as Attributes in this paper. Leveraging these cliques enables a more precise and context-aware adaptation by guiding the model to focus on discriminative attributes. + +Recent works [13, 17, 28, 31, 38] have explored transductive TTA under the Vision Transformer (ViT) backbone, often relying on batch statistics to adjust normalization parameters or select class-balanced samples for adaptation. However, these methods cannot be directly applied to CLIP since they completely neglect the adaptation of the text modality, which is crucial in VLMs. Rare CLIP-based TTA methods are designed for the transductive setting. In the latest method [14], the top-k predicted classes of each image in a batch are aggregated into a single new text prompt, which is used as pseudo labels to reclassify inputs in a transductive manner. However, the aforementioned transductive TTA methods mainly perform coarse-grained batch-level utilization of the visual similarity information. Thus, the fine-grained discriminative attribute information of cross-sample visual similarity is underutilized. + +In this paper, we propose Supportive Clique-based Attribute Prompting (SCAP), a novel CLIP-based transductive TTA approach designed to effectively leverage the cross-sample visual similarity information. Given a test batch, SCAP automatically mines distinct Supportive Cliques containing test images with shared attributes by leveraging visual and semantic similarity metrics. These cliques play a crucial role in learning Attribute Prompts that capture discriminative, attribute-specific information for each clique. For any given test sample, its aggregated attribute representation is constructed by combining the learned attribute prompts associated with the cliques to which it belongs. By utilizing these aggregated attribute prompts, SCAP enables CLIP to focus on fine-grained, instance-specific details, enhancing prediction accuracy and robustness. Additionally, we introduce a retention mechanism that dynamically up + +dates and preserves attribute prompts, ensuring long-term adaptability across tasks. Extensive experiments on two benchmarks demonstrate that SCAP significantly outperforms various current state-of-the-art methods in the transductive TTA setting, showing the effectiveness of SCAP in mining the intrinsic knowledge in batches. + +In summary, the contributions of this work are fourfold: (1) We propose Supportive Clique-based Prompting (SCAP), a novel approach that leverages the rich information of cross-sample relationships, improving the effectiveness of test-time adaptation under transductive setting. (2) To achieve robust and accurate fine-grained attribute prompt learning during test time, we propose to learn attribute prompts in a transductive manner based on cliques. (3) To retain the historically learned attribute prompts, we further design a retention mechanism to preserve the learned prompts for future reuse. (4) Extensive experiments on two benchmarks demonstrate that SCAP outperforms several current state-of-the-art methods in the transductive TTA setting, highlighting its effectiveness in extracting intrinsic knowledge from batches. + +# 2. Related Work + +# 2.1. Prompt Learning for CLIP + +VLMs like CLIP have demonstrated abundant knowledge learned from large-scale image-text data [5, 18, 19, 32, 33, 45]. Adapting CLIP to various downstream tasks is a valuable and challenging problem [3]. Recently, prompt learning [23, 24, 37, 48, 51] has emerged as a promising approach for enhancing model adaptability. Specifically, CoOp [57], CoCoOp [56], MaPLe [21], and Tip-Adapter [51] have achieved state-of-the-art performance. However, they rely on training data from downstream tasks. However, these approaches rely on training data annotations and often struggle when faced with substantial domain shifts at test time [38, 42]. + +# 2.2. Test-Time Adaptation + +Test-time adaptation (TTA) focuses on dynamically adjusting pre-trained models, such as CLIP or ViTs [22, 25], by leveraging sequentially obtained unlabeled test data [9, 11, 38, 49, 50]. Existing TTA approaches can be broadly categorized into training-free, augmentation-based, and prompt-based methods. Training-free methods aim to adapt CLIP efficiently without backpropagation [7, 20]. TDA introduces a positive and negative cache mechanism to assist prediction. Zero [7] aggregates predictions from multiple augmented versions without modifying model parameters. The augmentation-based methods [6, 49] focus on generating diverse augmented samples to integrate test distribution knowledge into the model. MTA [49] introduces a quality assessment variable for each + +![](images/a13064320d9da45957ed8d1376e7644ea1170283872273512d632c098f4664af.jpg) +Figure 2. Overview of our proposed SCAP. SCAP firstly mines the supportive cliques for all images in parallel. Based on the cliques, it then learns the corresponding attribute prompts from both modalities. The learned visual attribute prompts and the text attribute prompts are separately retained to accumulate the knowledge of test domains. For each instance, we conduct inference by jointly utilizing all the associated attribute prompts and the retained knowledge to generate the final prediction. + +augmented view and models the distribution using a mixed Gaussian approach. Prompt-based methods employ prompt tuning to adapt the model while preserving generalization capabilities [12, 25, 26, 36, 47, 50]. For example, TPT [36] optimizes instance-specific text prompts by minimizing prediction entropy, whereas DART [25] refines prompts across both modalities while leveraging historically learned knowledge via prompt retention. While these methods assume a sequential data setting where test samples are processed individually [13, 38], real-world scenarios often involve test-time data arriving in batches. Consequently, existing approaches struggle to exploit the inter-sample relationships within a batch, leading to suboptimal adaptation performance. + +# 2.3. Transductive Test-Time Adaptation + +Transductive TTA considers the practical scenario where test data are provided in batches rather than individual samples. Existing Transductive TTA methods are predominantly developed for ViTs pre-trained on ImageNet [13, 28, 31, 35, 39, 41]. These approaches typically rely on batch statistics to update normalization parameters [39] or select confident samples for model tuning [28]. Recent studies have demonstrated that CLIP outperforms ViTs pre-trained + +on ImageNet in TTA tasks. However, existing transductive TTA methods fail to fully exploit CLIP's capabilities, as they largely disregard the adaptation of the text modality, which is crucial for leveraging its knowledge learned from vision-language alignment. A few recent efforts have attempted to adapt CLIP for transductive TTA. For example, CLIPArTT [14] optimizes the model by jointly utilizing all images within a batch. However, since cross-instance relationships within a batch are often imbalanced, directly modeling all samples together does not effectively exploit the most informative relationships. In this work, we take a step further by identifying and leveraging the informative cliques, groups of strongly correlated samples within a batch, to facilitate robust adaptation by capturing discriminative attributes. + +# 3. Method + +# 3.1. Problem Setting and Notations + +In transductive TTA, test data arrive sequentially in batches. Given $t$ -th test batch $B_{t} = \{x_{j}\}_{j=1}^{b}$ , where $b$ is the batch size and $x_{j}$ is a test image. We generate a set of Supportive Cliques for each class $i$ , denoted as $\mathcal{C}_i = \{C_{ik}\}_{k=1}^m$ . Then, we extract the corresponding clique attributes, which are de + +noted as $\mathcal{A}_i = \{a_{ik} \in \mathbb{R}^d\}_{k=1}^m$ where $d$ is the attribute dimension, equivalent to the feature dimension. Additionally, we introduce Attribute Prompts $\mathcal{P}_i^A = \{\mathcal{P}_i^I, \mathcal{P}_i^T\}$ , which consist of visual attribute prompts $\mathcal{P}_i^I = \{P_{ik}^I\}_{k=1}^m$ and text attribute prompts $\mathcal{P}_i^T = \{P_{ik}^T\}_{k=1}^m$ . Our model $\theta$ is built on CLIP, which comprises an image encoder $\mathbf{E}_I$ and a text encoder $\mathbf{E}_T$ . + +# 3.2. Attribute-Driven Supportive Clique Mining + +Attributes are the fine-grained characteristics shared by a group of images in the batch. Since images containing the same attribute often exhibit high visual similarity to each other, in this section, we propose to obtain Supportive Cliques for the potential attributes by evaluating cross-image similarity. + +Specifically, we use CLIP to obtain the top- $k$ class predictions $topk(x_j)$ for each image $x_j$ , where $k$ is a hyperparameter. For each class $i$ , all images whose top- $k$ predictions include class $i$ are selected to form a subset $S_i$ : + +$$ +S _ {i} = \left\{x _ {j} \in B _ {t}: i \in \operatorname {t o p k} \left(x _ {j}\right) \right\} \tag {1} +$$ + +Since images in $S_{i}$ share similar class-specific information, which ensures semantic relevance. To further capture more fine-grained relevance, we extract the features $F_{i}^{I}\in$ $\mathbb{R}^{|S_i|\times d}$ of images within $S_{i}$ using the CLIP image encoder $\mathbf{E}_I$ and compute the similarity matrix $M_{i}\in \mathbb{R}^{|S_{i}|\times |S_{i}|}$ as follows: + +$$ +M _ {i} = F _ {i} ^ {I} \cdot \left(F _ {i} ^ {I}\right) ^ {\top} \tag {2} +$$ + +Then, given the $l$ -th row of $M_{i}$ , we collect elements exceeding a threshold $t$ , whose corresponding images form a Clique $C_{il}$ . Formally, the Clique generated from the $l$ -th row of $M_{i}$ is defined as: + +$$ +C _ {i l} = \left\{x _ {j} ^ {i}: x _ {j} ^ {i} \in S _ {i}, (M _ {i}) _ {l j} > t \right\}, \tag {3} +$$ + +where $(M_{i})_{lj}$ represents the $l$ -th row, $j$ -th column of $M_{i}$ . Only non-trivial Cliques $(|C_{il}| > 1)$ are retained, and duplicate Cliques are removed to ensure uniqueness. Finally, we obtain a set of valid Cliques $\mathcal{C}_i = \{C_{ik}\}_{k=1}^m$ . + +# 3.3. Attribute Prompt Learning + +In this section, we aim to learn prompts that capture attribute-specific knowledge to facilitate classification. + +Specifically, for each Clique $C_{ik}$ , we learn a visual attribute prompt $P_{ik}^{A}$ from the image modality and a text attribute prompt $P_{ik}^{T}$ from the text modality. The feature $f_{j}^{I} \in \mathbb{R}^{d}$ for the $j$ -th image in $C_{ik}$ is extracted as: + +$$ +f _ {j} ^ {I} = \mathbf {E} _ {I} \left(C _ {i k} ^ {j}, P _ {k} ^ {I}\right) \tag {4} +$$ + +Then, we define the attribute feature $a_{ik} \in \mathbb{R}^d$ of $C_{ik}$ which is calculated by: + +$$ +a _ {i k} = \frac {1}{\left| C _ {i k} \right|} \sum_ {j = 1} ^ {\left| C _ {i k} \right|} f _ {j} ^ {I} \tag {5} +$$ + +Besides, we compute the text feature $f_{n}^{T}$ of the $n$ -th class as follows: + +$$ +f _ {n} ^ {T} = \mathbf {E} _ {T} \left(c _ {n}, P _ {i k} ^ {T}\right) \tag {6} +$$ + +Given total $N$ classes, we introduce an attribute-specific class probability distribution vector, which is obtained by: + +$$ +\hat {\mathbf {P}} _ {i k} = \left[ \hat {p} _ {i k} ^ {1}, \hat {p} _ {i k} ^ {2}, \dots , \hat {p} _ {i k} ^ {N} \right] \tag {7} +$$ + +where + +$$ +\hat {p} _ {i k} ^ {n} = \frac {\exp \left(\cos \left(a _ {i k} , f _ {n} ^ {T}\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\cos \left(a _ {i k} , f _ {j} ^ {T}\right) / \tau\right)} \tag {8} +$$ + +Note that $\cos (\cdot ,\cdot)$ denotes cosine similarity and $\tau$ is a temperature parameter to scale the similarity scores. + +Then, we compute the Entropy Loss for the probability distribution of the attribute: + +$$ +\mathcal {L} _ {e n} \left(C _ {i k}\right) = - \sum_ {n = 1} ^ {N} \hat {p} _ {i k} ^ {n} \log \hat {p} _ {i k} ^ {n} \tag {9} +$$ + +To further guide the model to extract attribute-specific information and discard attribute-irrelevant information, we introduce a concentration loss $\mathcal{L}_{con}(C_{ik})$ that forces the model to capture the shared information across the samples within a Clique. Specifically, $\mathcal{L}_{con}(C_{ik})$ is calculated by + +$$ +\mathcal {L} _ {\text {c o n}} \left(C _ {i k}\right) = \sum_ {j = 1} ^ {| C _ {i k} |} \left| \left| f _ {j} ^ {I} - a _ {i k} \right| \right| ^ {2} \tag {10} +$$ + +Therefore, both kinds of prompts $P_{ik}^{I}$ and $P_{ik}^{T}$ are guided to encode attribute-specific knowledge as the model optimization. Such knowledge is valuable when generating predictions for images showing the same attributes. + +# 3.4. Attribute Prompt Retention + +In this section, to retain attribute knowledge during the test process and ensure consistent performance improvement, we introduce mechanisms for retaining both text and visual attribute knowledge. + +Firstly, to retain the text attribute knowledge, we propose a Text Retention Prompt $\tilde{P}^T$ . Specifically, when a new Clique $C_{ik}$ is learned, we update $\tilde{P}^T$ by + +$$ +\tilde {P} ^ {T} = \alpha_ {\tau} \tilde {P} ^ {T} + (1 - \alpha_ {\tau}) P _ {i k} ^ {T}, \tag {11} +$$ + +where $\alpha_{\tau} = \tau /(1 + \tau)$ is a decay rate based on iteration count $\tau$ + +Besides, to retain the visual attribute knowledge, we design a Retention Cache $\mathcal{R} = \{R_i\}_{i=1}^N$ where $R_i$ is a class-specific retention cache. Specifically, we store $(a_{ik}, P_{ik}^I)$ as key-value pairs in $R_i$ . We denote the retained attributes (i.e. keys) as $R_i^A = \{a_1, \ldots, a_L\}$ and the corresponding attribute prompts (i.e. values) as $R_i^P = \{P_1^I, \ldots, P_L^I\}$ . The + +maximum size of each $R_i$ is set to $L$ to avoid memory explosion. When the retained key-value pair number is less than $L$ , we directly add a new $(a_{ik}, P_{ik}^I)$ to $R_i$ . Otherwise, we add $(a_{ik}, P_{ik}^I)$ to $R_i$ first, then we update all pairs and aggregate the two that are most similar to each other in $R_i$ by using a graph structure as follows. + +Firstly, we construct a weighted graph $G_{i} = (V_{i},E_{i})$ . The vertices are the keys in the retention cache $V_{i} = R_{i}^{A} = \{a_{1},\dots ,a_{L + 1}\}$ . The edges $E_{i}$ between each two vertices are defined as the Gaussian Distance of the two attributes. Thus, the adjacency matrix $W_{i}\in \mathbb{R}^{(L + 1)\times (L + 1)}$ of $G_{i}$ is defined with: + +$$ +\left(W _ {i}\right) _ {k l} = \exp \left(- \frac {\left\| a _ {k} - a _ {l} \right\| ^ {2}}{2 \sigma^ {2}}\right), \tag {12} +$$ + +where $(W_{i})_{kl}$ is the $k$ -th row, $l$ -column of $W_{i}$ and $\sigma$ is a hyperparameter. We utilize the $K$ -max values in each row of $W_{i}$ to construct a $K$ -nearest retention graph. Then we perform graph propagation following [53] and obtain the updated graph vertices $V_{i}^{\prime} = \{a_{1}^{\prime},\ldots ,a_{L + 1}^{\prime}\}$ , where the similar attributes are pushed closer automatically. Then, the most similar two attributes $a_{j_0}^\prime ,a_{j_1}^\prime$ and the corresponding prompts are fused into one respectively as follows: + +$$ +\left\{ \begin{array}{l} a _ {n e w} = \frac {1}{2} \left(a _ {j _ {0}} ^ {\prime} + a _ {j _ {1}} ^ {\prime}\right) \\ P _ {n e w} ^ {I} = \frac {1}{2} \left(P _ {j _ {0}} ^ {I} + P _ {j _ {1}} ^ {I}\right) \end{array} \right. \tag {13} +$$ + +Through the above process, we obtain the updated retention cache $R_{i}$ , dynamically retaining the visual and text attribute knowledge learned from previous batches into Retention Cache and Text Retention Prompt, respectively. + +# 3.5. Optimization and Inference + +For each batch $B_{t}$ , we optimize the visual attribute prompts $\mathcal{P}_i^I = \{P_{ik}^I\}_{k=1}^m$ and text attribute prompts $\mathcal{P}_i^T = \{P_{ik}^T\}_{k=1}^m$ for every class $i$ by minimizing the following loss function: + +$$ +\mathcal {L} = \sum_ {k = 1} ^ {m} \mathcal {L} _ {e n} \left(C _ {i k}\right) + \lambda \mathcal {L} _ {c o n} \left(C _ {i k}\right) \tag {14} +$$ + +where $\lambda$ is a hyperparameter. + +Then we conduct inference for each $S_{i}$ of $B_{t}$ . To fully utilize both the Attribute Prompts learned from the current batch $B_{t}$ and the retained knowledge to improve prediction. Given the test image $x_{j} \in S_{i}$ , we first collect all Cliques containing $x_{j}$ , namely $\mathcal{K} = \{k : x_{j} \in C_{ik}\}$ , whose corresponding visual attribute prompts are concatenated together: + +$$ +P ^ {I} = \operatorname {c o n c a t} \left(\left[ P _ {i k} ^ {I}: k \in \mathcal {K} \right]\right) \tag {15} +$$ + +Then, we introduce the retention prompt matching strategy utilizing the key-value pair structure of the retention + +cache. Specifically, the keys are the retained attribute features $R_{i}^{A} = \{a_{1},\dots ,a_{L}\}$ and the values are the corresponding attribute prompts $R_{i}^{P} = \{P_{1}^{I},\dots ,P_{L}^{I}\}$ . Given the image feature $f_{j}^{I}$ of $x_{j}$ which is already extracted in Sec. 3.3, the top-1 attribute $a^*$ is obtained by evaluating the similarity of $f_{j}^{I}$ to all keys in $R_{i}^{A}$ , i.e. + +$$ +a ^ {*} = \underset {a _ {l} \in R _ {i} ^ {A}} {\operatorname {a r g m a x}} < f _ {j} ^ {I}, a _ {l} > \tag {16} +$$ + +Given the matched value corresponding to $a^*$ , i.e., $P^{I*}$ , we obtain the Composed Image Prompt $P_C^I$ by concatenating $P^I$ and $P^{I*}$ along the token dimension + +$$ +P _ {C} ^ {I} = \operatorname {c o n c a t} \left(P ^ {I}, P ^ {I ^ {*}}\right) \tag {17} +$$ + +For the text modality, we fuse $\mathcal{P}^{\mathcal{T}} = \{P_{ik}^{T}\}_{k=1}^{m}$ and $\tilde{P}^T$ together to get the Composed Text Prompt $P_C^T$ : + +$$ +P _ {C} ^ {T} = \alpha_ {r} \tilde {P} ^ {T} + (1 - \alpha_ {r}) \frac {1}{m} \sum_ {k = 1} ^ {m} P _ {i k} ^ {T} \tag {18} +$$ + +where $\alpha_{r}$ is the weight of retained text prompts. Then, for the final prediction, the image-inference feature $\hat{f}_j^I$ of $x_j^i$ and the text-inference feature $\hat{f}_n^T$ of class $c_{n}$ are obtained by: + +$$ +\left\{ \begin{array}{l} \hat {f} _ {j} ^ {I} = \mathbf {E} _ {I} \left(x _ {j}, P _ {C} ^ {I}\right) \\ \hat {f} _ {n} ^ {T} = \mathbf {E} _ {T} \left(c _ {n}, P _ {C} ^ {T}\right) \end{array} \right. \tag {19} +$$ + +The prediction $y_{j}$ of $x_{j}$ is obtained by applying Eq.7. + +# 4. Experiments + +In this section, we first introduce the two benchmarks used for model evaluation, followed by a detailed description of the implementation setup. Next, we present a comparative analysis of existing methods and our proposed approach. Finally, extensive ablation studies are conducted to validate the effectiveness of SCAP. Additional experimental results and in-depth discussions are provided in the Supplementary Materials. + +# 4.1. Benchmarks + +Our evaluation is conducted on two benchmarks: an out-of-distribution (OOD) benchmark and a cross-domain benchmark. The OOD benchmark consists of four datasets derived from ImageNet [4], i.e., ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [40], and ImageNet-V2 [34]. The OOD benchmark aims to assess the model's robustness to substantial distribution shifts in the test data. The cross-domain benchmark comprises four datasets of the image classification task, each representing a distinct domain: Aircraft [27], Caltech101 [8], Flower102 [29], and Pets [30]. The cross-domain benchmark provides a measure of the model's adaptability across a diverse range of domain-specific tasks. + +
MethodPublicationImageNet-AImageNet-RImageNet-SImageNet-V2Average
CLIP [32]ICML 202147.8773.9846.0961.8857.46
CoOp [57]IJCV 202249.7175.2147.9964.2059.28
CoCoOp [56]CVPR 202250.6376.1848.7564.0759.91
Tip-Adapter [51]ECCV 202251.0477.7648.8863.4160.27
TPT [36]NeurIPS 202254.7777.0647.9463.4560.81
C-TPT [47]ICLR 202452.9078.0048.5063.4060.70
DART [25]AAAI 202460.5679.5649.7664.0363.48
MTA [49]CVPR 202458.0678.3349.6164.2462.56
TDA [20]CVPR 202460.1180.2450.5464.6763.89
Zero [7]NeuIPS 202461.3577.2848.2964.1362.76
SCAPThis Paper64.5281.6851.6564.6565.63
+ +Table 1. Results comparison on the out-of-distribution (OOD) Benchmark. The results of the compared methods follow the original paper or are implemented with the official code [25]. + +![](images/146f629d2924afda08ff1c390a5025955ec95f3bd2ed7b27c814f7f1780c015b.jpg) +Figure 3. Results on the Cross-Domain Benchmark. Comparison of SCAP with state-of-the-art methods. All methods are evaluated using CLIP-ViT-B/16 as the backbone. + +![](images/2405584b2bdab9b03f18c705ee00d940acaaf1ddcf25cef570866618cd0b1bf4.jpg) + +![](images/30ad29488f3bb38f8f8a5250796452380dc534a87f573f09144d6a199d955143.jpg) + +![](images/e9f79901e069342977655652326c00b30417e3110dea6a0b5d534e4e4df88497.jpg) + +Implementation details. We adopt CLIP with ViT-B/16 as the backbone and set the batch size to 64 for all experiments. Visual attribute prompts are initialized using a uniform distribution in the range $(-1, 1)$ , following previous TTA methods [25, 36]. For text attribute prompts, we employ an ensemble of multiple templates for the OOD benchmark, whereas dataset-specific templates are used for the cross-domain benchmark, following TDA [20]. Our key hyperparameters include the threshold $t$ for supportive clique extraction, the weight $\alpha_{r}$ for computing $P_C^T$ , the concentration loss weight $\lambda$ , and graph structure parameters $\sigma$ , $L$ , and $K$ . The hyperparameter ablations of this paper are conducted on ImageNet-A, with the default values of $\alpha_{r}$ , $\sigma$ , $L$ , and $K$ set to 1, 0.3, 6, and 3, respectively. We set $k = 3$ in the initial top-k predictions Eq. (1). An Adam optimizer with a learning rate of 0.003 is used to update the prompts, and all experiments are conducted on a single NVIDIA 4090 GPU. + +# 4.2. Comparison with State-of-the-arts + +Compared methods. We compare our proposed SCAP with state-of-the-art TTA methods designed for CLIP adaptation, including TPT [36], C-TPT [47], DART [25], MTA [49], TDA [20], and Zero [7]. Additionally, we report + +results from the original CLIP model to highlight the knowledge gained during testing. Following previous works [20], we also include training-time adaptation methods such as CoOp [57], CoCoOp [56], and Tip-Adapter [51] for comparison to further verify the label efficiency and effectiveness of our approach. + +Results on the OOD Benchmark. As shown in Tab. 1, our proposed SCAP achieves state-of-the-art performance on the ImageNet-A, ImageNet-R, and ImageNet-S datasets, outperforming existing methods by at least $3.17\%$ , $1.44\%$ , and $1.11\%$ , respectively. Additionally, on the challenging ImageNet-V2 dataset, SCAP performs comparably to TDA, with only a marginal $0.02\%$ difference. While TDA benefits from a feature cache to enable joint learning across multiple samples, it incurs additional storage overhead. In contrast, SCAP achieves comparable or superior performance without maintaining historical features, making it more efficient in practice. + +Overall, SCAP improves average accuracy across the four ImageNet variants by $1.74\%$ , highlighting its effectiveness in handling natural object images across diverse distribution shifts. This performance gain is primarily attributed to the proposed supportive clique mining strategy, which ef + +
Components in SCAPImageNet-A
ImageTextRetentionAcc@1
XXX60.98
X✓X62.57
✓XX63.14
✓X✓63.82
✓✓✓64.52
+ +- Image and Text represent learning from the Image Modality and Text Modality in the Attribute Prompt Learning module, respectively. +- Retention refers only to the Retention Cache for the visual modality while the text retention prompts are always used. + +Table 2. Ablation study about the different components + +fectively captures diverse object attributes, such as "head" and "body", that enhance the object recognition capacity of the model. + +Results over cross-domain benchmark. As shown in Fig. 3, we compare our SCAP method with state-of-the-art approaches across four cross-domain datasets. The results demonstrate that our method surpasses the state-of-the-art approaches, with improvements of at least $1.64\%$ , $0.23\%$ , $0.18\%$ , and $0.24\%$ on the Pets, Flower102, Caltech101, and Aircraft datasets, respectively. This superior performance stems from the proposed Attribute Prompt Learning and Attribute Prompt Retention mechanisms, which effectively capture and leverage generalizable knowledge shared across diverse samples. + +# 4.3. Ablation Studies + +The Influence of Each Component. To evaluate the effectiveness of our module designs across different modalities (i.e., image, text, and retention), we conduct modality-specific ablation studies. As shown in Tab. 2, leveraging both image and text modalities within the Attribute Prompt Learning module consistently improves performance over the baseline model. Additionally, incorporating the Attribute Prompt Retention module further enhances the utilization of learned image prompts, leading to substantial performance gains. Furthermore, when all modules and modalities are integrated, the model performance improves further due to joint optimization, which mitigates semantic drift as learnable parameters are updated. + +The Influence of Batch Size in SCAP. Batch Size $B$ is an important factor in the transductive TTA setting, as larger batches provide richer inter-image relationships. As illustrated in Fig. 4 (Left), the model performance consistently improves with increasing batch size, demonstrating SCAP's capability to leverage the additional contextual information. We also observe that the image modality varies significantly as batch size changes, while the text modality remains relatively stable across batch size variations. On + +![](images/cfa5c169664d6e9e4b56830bf72a7e68adcb82969199e98ba9a5404c343feb52.jpg) +Figure 4. Left: the influence of batch size on SCAP performance. Right: the average maximum Clique size generated per batch under different batch sizes. + +![](images/08c1ef4d26770d845c2a2c5180111a406eaf1537cf1c15608278c7073a7778ee.jpg) + +![](images/a95f19d56c8f23680964b96e34d658a869dd6c844d6c751be04a83d40ddf15bd.jpg) + +![](images/833359702d9529a1910ef1a8050c8f6b16d2bb95d3a1cbc718cf53449fa7b338.jpg) + +![](images/18f79ebf935cbf8dda60b30d5f21ff9636092c14c9457729a2a7fa7cd61513bf.jpg) +Figure 5. Ablation study about the influence of different hyperparameters on ImageNet-A. + +![](images/2ae297e278b0568dfb5c8e8d8430444865b5fcd9a8c3b082f0bce81d15b57424.jpg) + +the right side of Fig. 4, the results reveal that larger batches lead to increasing average clique size, indicating that more support information can be captured under such conditions. + +The Influence of Different Hyperparameters. The key hyperparameters in our approach include the threshold $t$ for clique extraction, the concentration loss weight $\lambda$ , the retention cache limit $L$ , and the degree of the retention graph $K$ . As illustrated in Fig. 5 (a), the model performance is relatively sensitive to changes in $t$ . A lower $t$ results in cliques containing unrelated images, which hinders effective attribute prompt learning, whereas a higher $t$ limits the number of support samples, leading to insufficient attribute modeling. Subsequently, we explore the impact of $\lambda$ in Fig. 5 (b). The model performance improves consistently up to $\lambda = 1.0$ , as the concentration loss enhances attribute learning. However, further increasing $\lambda$ leads to overfitting, negatively affecting generalization. Then, we investigate the influence of $L$ in Fig. 5 (c). We observe that as $L$ increases, the performance increases and then becomes stable, indicating that while a larger cache enhances retention, + +![](images/d07aebf70c601d87de011877379514f89052e9dbd3e3114c855169b90d7ca043.jpg) +Figure 6. Visualization of our prompt-guided attention maps in comparison with the state-of-the-art prompt-based method. + +
Combine PIkCombine PI, P*ImageNet-A
MeanConcatMeanConcatAcc@1
✓X✓X61.82
✓XX✓62.11
X✓✓X64.35
X✓X✓64.52
+ +Table 3. The influence of different combination approaches. + +excessive stored knowledge does not yield further gains. Finally, we study the influence of $K$ in Fig. 5 (d). Increasing $K$ up to 2 improves performance by strengthening graph connectivity and enhancing information propagation. However, beyond this point, performance declines due to the inclusion of irrelevant nodes, which introduces noise and reduces the discriminative power of the learned attributes. + +Influence of Different Prompt Combination Approaches. Two primary strategies for combining multiple prompts are considered: generating a mean prompt and concatenating prompts along the token dimension, as illustrated in Eq. (15) and in Eq. (17). Experimental results in Table Tab. 3 indicate that concatenation outperforms averaging in our SCAP framework. This advantage due to concatenation preserves the full prompt information, enabling each prompt to focus on distinct image regions. In contrast, the mean prompt approach struggles to capture diverse visual patterns, leading to suboptimal performance. + +Visualization of Supportive Cliques. To demonstrate the effectiveness of our supportive clique-based attribute prompt learning approach, we visualize the attention + +maps in comparison with the state-of-the-art prompt-based method DART [25]. As illustrated in Fig. 6, each supportive clique in our method guides the prompts to focus on specific object parts, such as the head, leg, or body. When the attribute prompts are exploited together, the obtained attention maps precisely and comprehensively capture the object information, highlighting the superior feature-learning capability of our approach over DART. + +# 5. Conclusion + +In this paper, we propose a novel transductive test-time adaptation approach, Supportive Clique-based Attribute Prompting (SCAP), which leverages intrinsic cross-sample shared visual attributes to enhance model adaptation during testing. Specifically, our method identifies and models the attributes within batches through supportive cliques, sets of samples exhibiting similar patterns. Furthermore, a retention module is designed to dynamically store, aggregate, and utilize knowledge from previously learned attributes, thereby improving predictions for subsequent test images. Extensive experiments on two benchmarks demonstrate the effectiveness of our method in leveraging batch-wise information. The proposed approach offers a promising direction for fully exploiting cross-sample relationships in practical transductive TTA scenarios. + +Acknowledgments. This work was supported by the grants from the National Natural Science Foundation of China (62376011, 61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006). + +# References + +[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 2 +[2] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, pages 295-305, 2022. 2 +[3] Zhenyu Cui, Yuxin Peng, Xun Wang, Manyu Zhu, and Jiuhuan Zhou. Continual vision-language retrieval via dynamic knowledge rectification. In AAAI, pages 11704-11712, 2024. 2 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5 +[5] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2 +[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, pages 7704-7714, 2023. 2 +[7] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, Elisa Ricci, et al. Frustratingly easy test-time adaptation of vision-language models. In NeurIPS, 2024. 1, 2, 6 +[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5 +[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, pages 2704-2714, 2023. 2 +[10] Lihua Fu, Yubin Du, Yu Ding, Dan Wang, Hanxu Jiang, and Haitao Zhang. Domain adaptive learning with multi-granularity features for unsupervised person re-identification. Chinese Journal of Electronics, 31(1):116-128, 2022. 2 +[11] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 2 +[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3 +[13] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. NeurIPS, 36, 2024. 2, 3 +[14] Gustavo Adolfo Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Clipartt: Light-weight adaptation of clip to new domains at test time. arXiv preprint arXiv:2405.00754, 2024. 2, 3 +[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust- + +ness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5 +[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5 +[17] Xuefeng Hu, Gokhan Uzunbas, Sirius Chen, Rui Wang, Ashish Shah, Ram Nevatia, and Ser-Nam Lim. Mixnorm: Test-time adaptation through online normalization estimation. arXiv preprint arXiv:2110.11478, 2021. 2 +[18] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In CVPR, pages 17980-17989, 2022. 2 +[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2 +[20] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, pages 14162-14171, 2024. 1, 2, 6 +[21] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2 +[22] Qiwei Li, Kunlun Xu, Yuxin Peng, and Jiahuan Zhou. Exemplar-free lifelong person re-identification via prompt-guided adaptive knowledge consolidation. IJCV, 132(11): 4850-4865, 2024. 2 +[23] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. NeurIPS, 36:13448-13466, 2023. 2 +[24] Zichen Liu, Yuxin Peng, and Jiahuan Zhou. Inspv: Efficient instance visual prompting from image itself. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6443-6452, 2024. 2 +[25] Zichen Liu, Hongbo Sun, Yuxin Peng, and Jiahuan Zhou. Dart: dual-modal adaptive online prompting and knowledge retention for test-time adaptation. In AAAI, pages 14106-14114, 2024. 1, 2, 3, 6, 8 +[26] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. NeurIPS, 36, 2024. 3 +[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[28] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, pages 2555-2565, 2024. 2, 3 +[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5 + +[30] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505. IEEE, 2012. 5 +[31] Ori Press, Steffen Schneider, Matthias Kümmerer, and Matthias Bethge. Rdumb: A simple approach that questions our progress in continual test-time adaptation. NeurIPS, 36: 39915-39935, 2023. 2, 3 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 2 +[34] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400. PMLR, 2019. 5 +[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 33:11539-11551, 2020. 3 +[36] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. NeurIPS, 35:14274-14289, 2022. 2, 3, 6 +[37] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In CVPR, pages 11920-11929, 2023. 2 +[38] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR. 2, 3 +[39] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3 +[40] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5 +[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3 +[42] Kunlun Xu, Haozhuo Zhang, Yu Li, Yuxin Peng, and Jiahuan Zhou. Mitigate catastrophic remembering via continual knowledge purification for noisy lifelong person re-identification. In ACM MM, pages 5790–5799, 2024. 2 +[43] Kunlun Xu, Xu Zou, Yuxin Peng, and Jiahuan Zhou. Distribution-aware knowledge prototyping for non-exemplar lifelong person re-identification. In CVPR, pages 16604–16613, 2024. 1 +[44] Kunlun Xu, Xu Zou, and Jiahuan Zhou. Lstkc: Long short-term knowledge consolidation for lifelong person re-identification. In AAAI, pages 16202–16210, 2024. 2 + +[45] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, pages 15671-15680, 2022. 2 +[46] Zhaoda Ye, Xiangteng He, and Yuxin Peng. Unsupervised cross-media hashing learning via knowledge graph. Chinese Journal of Electronics, 31(6):1081-1091, 2022. 1 +[47] Hee Suk Yoon, Eunseop Yoon, Joshua Tian Jin Tee, Mark Hasegawa-Johnson, Yingzhen Li, and Chang D Yoo. C-tpt: Calibrated test-time prompt tuning for vision-language models via text feature dispersion. arXiv preprint arXiv:2403.14119, 2024. 3, 6 +[48] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 2 +[49] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, pages 23783-23793, 2024. 2, 6 +[50] Ding-Chu Zhang, Zhi Zhou, and Yu-Feng Li. Robust test-time adaptation for zero-shot prompt tuning. In AAAI, pages 16714-16722, 2024. 2, 3 +[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, pages 493-510. Springer, 2022. 2, 6 +[52] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024. 1 +[53] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. NeurIPS, 16, 2003. 5 +[54] Jiahuan Zhou, Pei Yu, Wei Tang, and Ying Wu. Efficient online local metric adaptation via negative samples for person re-identification. In ICCV, pages 2420–2428, 2017. 2 +[55] Jiahuan Zhou, Bing Su, and Ying Wu. Online joint multimetric adaptation from frequent sharing-subset mining for person re-identification. In CVPR, pages 2909–2918, 2020. 1 +[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 6 +[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 6 \ No newline at end of file diff --git a/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/images.zip b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..540025c457d1a7c284d4a55ff9e233fb35874901 --- /dev/null +++ b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d99813636da4696d6f6ba73f1000a7d095fe0f4a3345209ceea368799411f37 +size 725785 diff --git a/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/layout.json b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..50d27f94104adcae544c15d41f717de7a3ad29f3 --- /dev/null +++ b/2025/SCAP_ Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting/layout.json @@ -0,0 +1,10535 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 526, + 140 + ], + "type": "text", + "content": "SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 161, + 474, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 161, + 474, + 175 + ], + "spans": [ + { + "bbox": [ + 139, + 161, + 474, + 175 + ], + "type": "text", + "content": "Chenyu Zhang*, Kunlun Xu*, Zichen Liu, Yuxin Peng, Jiahuan Zhou†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 177, + 466, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 177, + 466, + 190 + ], + "spans": [ + { + "bbox": [ + 146, + 177, + 466, + 190 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University, Beijing 100871, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 191, + 523, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 191, + 523, + 205 + ], + "spans": [ + { + "bbox": [ + 83, + 191, + 523, + 205 + ], + "type": "text", + "content": "{2200010814, xkl, lzc20180720}@stu.pku.edu.cn, {pengyuxin, jiahuanzhou}@pku.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 256, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 296, + 616 + ], + "type": "text", + "content": "Vision-language models (VLMs) encounter considerable challenges when adapting to domain shifts stemming from changes in data distribution. Test-time adaptation (TTA) has emerged as a promising approach to enhance VLM performance under such conditions. In practice, test data often arrives in batches, leading to increasing interest in the transductive TTA setting. However, existing TTA methods primarily focus on individual test samples, overlooking crucial cross-sample correlations within a batch. While recent ViT-based TTA methods have introduced batch-level adaptation, they remain suboptimal for VLMs due to inadequate integration of the text modality. To address these limitations, we propose a novel transductive TTA framework, Supportive Clique-based Attribute Prompting (SCAP), which effectively combines visual and textual information to enhance adaptation by generating fine-grained attribute prompts across test batches. SCAP first forms supportive cliques of test samples in an unsupervised manner based on visual similarity and learns an attribute prompt for each clique, capturing shared attributes critical for adaptation. For each test sample, SCAP aggregates attribute prompts from its associated cliques, providing enriched contextual information. To ensure adaptability over time, we incorporate a retention module that dynamically updates attribute prompts and their associated attributes as new data arrives. Comprehensive experiments across multiple benchmarks demonstrate that SCAP outperforms existing state-of-the-art methods, significantly advancing VLM generalization under domain shifts. Our code is available at https://github.com/zhoujiahuan1991/CVPR2025-SCAP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 685 + ], + "type": "text", + "content": "Vision-language models (VLMs) such as CLIP [32] have exhibited remarkable generalization capabilities across a" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 231, + 555, + 369 + ], + "blocks": [ + { + "bbox": [ + 319, + 231, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 319, + 231, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 319, + 231, + 555, + 369 + ], + "type": "image", + "image_path": "3663c2c26e042dcd2e24ccfc41b98c255b2f5b41b5d7fde59ccfba92673f6a61.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 372, + 541, + 381 + ], + "lines": [ + { + "bbox": [ + 329, + 372, + 541, + 381 + ], + "spans": [ + { + "bbox": [ + 329, + 372, + 541, + 381 + ], + "type": "text", + "content": "(a) Current Prompt Learning based Test-time Adaptation Methods" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 383, + 553, + 521 + ], + "blocks": [ + { + "bbox": [ + 319, + 383, + 553, + 521 + ], + "lines": [ + { + "bbox": [ + 319, + 383, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 319, + 383, + 553, + 521 + ], + "type": "image", + "image_path": "c2ecaa5d2a555735b2357b754ef3dd69353a1009afc91b7d5c969ad16e0ac72d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 523, + 450, + 533 + ], + "lines": [ + { + "bbox": [ + 421, + 523, + 450, + 533 + ], + "spans": [ + { + "bbox": [ + 421, + 523, + 450, + 533 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 542, + 555, + 632 + ], + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 632 + ], + "type": "text", + "content": "Figure 1. Comparison between our proposed SCAP to existing prompt learning-based TTA methods. Specifically, current TTA methods learn from instances, while our method utilizes cross-sample visual similarity information from batches to construct supportive cliques and extract attributes from them. We learn attributes from both modalities based on the cliques. For each image, SCAP jointly utilizes attribute prompts from its associated cliques, leading to more effective and accurate prompting." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "range of downstream tasks in zero-shot settings. These advancements highlight the significant potential of CLIP in a wide range of domains. However, their performance often degrades [7, 20, 25] when downstream data exhibits substantial domain shifts [43, 46, 52, 55] from the pre" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30032" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "training distribution [10, 36, 44, 54]. This limitation has sparked increasing interest in the area of Test-Time Adaptation (TTA) [1, 2, 39, 49], which aims to adapt pre-trained VLMs to mitigate domain shifts during test time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "type": "text", + "content": "Recent works in TTA have primarily focused on adapting CLIP by independently learning from each sample during test time [20, 25, 36, 49]. However, in real-world scenarios, test-time data typically arrives in batches, creating a transductive TTA setting. Existing TTA methods can tackle such a scenario by treating each test sample individually [25]. Although this solution is available, it suffers from significant limitations due to disregarding the intrinsic cross-sample visual similarity within each batch. Exploiting these inter-sample similarities can greatly enhance both the effectiveness and robustness of the adaptation process. As illustrated in Figure 1, test batches often contain Cliques, groups of visually and semantically similar images that share common characteristics, referred to as Attributes in this paper. Leveraging these cliques enables a more precise and context-aware adaptation by guiding the model to focus on discriminative attributes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 519 + ], + "type": "text", + "content": "Recent works [13, 17, 28, 31, 38] have explored transductive TTA under the Vision Transformer (ViT) backbone, often relying on batch statistics to adjust normalization parameters or select class-balanced samples for adaptation. However, these methods cannot be directly applied to CLIP since they completely neglect the adaptation of the text modality, which is crucial in VLMs. Rare CLIP-based TTA methods are designed for the transductive setting. In the latest method [14], the top-k predicted classes of each image in a batch are aggregated into a single new text prompt, which is used as pseudo labels to reclassify inputs in a transductive manner. However, the aforementioned transductive TTA methods mainly perform coarse-grained batch-level utilization of the visual similarity information. Thus, the fine-grained discriminative attribute information of cross-sample visual similarity is underutilized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose Supportive Clique-based Attribute Prompting (SCAP), a novel CLIP-based transductive TTA approach designed to effectively leverage the cross-sample visual similarity information. Given a test batch, SCAP automatically mines distinct Supportive Cliques containing test images with shared attributes by leveraging visual and semantic similarity metrics. These cliques play a crucial role in learning Attribute Prompts that capture discriminative, attribute-specific information for each clique. For any given test sample, its aggregated attribute representation is constructed by combining the learned attribute prompts associated with the cliques to which it belongs. By utilizing these aggregated attribute prompts, SCAP enables CLIP to focus on fine-grained, instance-specific details, enhancing prediction accuracy and robustness. Additionally, we introduce a retention mechanism that dynamically up" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "dates and preserves attribute prompts, ensuring long-term adaptability across tasks. Extensive experiments on two benchmarks demonstrate that SCAP significantly outperforms various current state-of-the-art methods in the transductive TTA setting, showing the effectiveness of SCAP in mining the intrinsic knowledge in batches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 145, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 324 + ], + "type": "text", + "content": "In summary, the contributions of this work are fourfold: (1) We propose Supportive Clique-based Prompting (SCAP), a novel approach that leverages the rich information of cross-sample relationships, improving the effectiveness of test-time adaptation under transductive setting. (2) To achieve robust and accurate fine-grained attribute prompt learning during test time, we propose to learn attribute prompts in a transductive manner based on cliques. (3) To retain the historically learned attribute prompts, we further design a retention mechanism to preserve the learned prompts for future reuse. (4) Extensive experiments on two benchmarks demonstrate that SCAP outperforms several current state-of-the-art methods in the transductive TTA setting, highlighting its effectiveness in extracting intrinsic knowledge from batches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 336, + 400, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 400, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 400, + 348 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 357, + 464, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 464, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 464, + 369 + ], + "type": "text", + "content": "2.1. Prompt Learning for CLIP" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": "VLMs like CLIP have demonstrated abundant knowledge learned from large-scale image-text data [5, 18, 19, 32, 33, 45]. Adapting CLIP to various downstream tasks is a valuable and challenging problem [3]. Recently, prompt learning [23, 24, 37, 48, 51] has emerged as a promising approach for enhancing model adaptability. Specifically, CoOp [57], CoCoOp [56], MaPLe [21], and Tip-Adapter [51] have achieved state-of-the-art performance. However, they rely on training data from downstream tasks. However, these approaches rely on training data annotations and often struggle when faced with substantial domain shifts at test time [38, 42]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 528, + 440, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 440, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 440, + 540 + ], + "type": "text", + "content": "2.2. Test-Time Adaptation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Test-time adaptation (TTA) focuses on dynamically adjusting pre-trained models, such as CLIP or ViTs [22, 25], by leveraging sequentially obtained unlabeled test data [9, 11, 38, 49, 50]. Existing TTA approaches can be broadly categorized into training-free, augmentation-based, and prompt-based methods. Training-free methods aim to adapt CLIP efficiently without backpropagation [7, 20]. TDA introduces a positive and negative cache mechanism to assist prediction. Zero [7] aggregates predictions from multiple augmented versions without modifying model parameters. The augmentation-based methods [6, 49] focus on generating diverse augmented samples to integrate test distribution knowledge into the model. MTA [49] introduces a quality assessment variable for each" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30033" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 349 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 349 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 349 + ], + "type": "image", + "image_path": "a13064320d9da45957ed8d1376e7644ea1170283872273512d632c098f4664af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 358, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 555, + 404 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed SCAP. SCAP firstly mines the supportive cliques for all images in parallel. Based on the cliques, it then learns the corresponding attribute prompts from both modalities. The learned visual attribute prompts and the text attribute prompts are separately retained to accumulate the knowledge of test domains. For each instance, we conduct inference by jointly utilizing all the associated attribute prompts and the retained knowledge to generate the final prediction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 424, + 297, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 297, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 297, + 593 + ], + "type": "text", + "content": "augmented view and models the distribution using a mixed Gaussian approach. Prompt-based methods employ prompt tuning to adapt the model while preserving generalization capabilities [12, 25, 26, 36, 47, 50]. For example, TPT [36] optimizes instance-specific text prompts by minimizing prediction entropy, whereas DART [25] refines prompts across both modalities while leveraging historically learned knowledge via prompt retention. While these methods assume a sequential data setting where test samples are processed individually [13, 38], real-world scenarios often involve test-time data arriving in batches. Consequently, existing approaches struggle to exploit the inter-sample relationships within a batch, leading to suboptimal adaptation performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 600, + 246, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 246, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 246, + 613 + ], + "type": "text", + "content": "2.3. Transductive Test-Time Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Transductive TTA considers the practical scenario where test data are provided in batches rather than individual samples. Existing Transductive TTA methods are predominantly developed for ViTs pre-trained on ImageNet [13, 28, 31, 35, 39, 41]. These approaches typically rely on batch statistics to update normalization parameters [39] or select confident samples for model tuning [28]. Recent studies have demonstrated that CLIP outperforms ViTs pre-trained" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 424, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 604 + ], + "type": "text", + "content": "on ImageNet in TTA tasks. However, existing transductive TTA methods fail to fully exploit CLIP's capabilities, as they largely disregard the adaptation of the text modality, which is crucial for leveraging its knowledge learned from vision-language alignment. A few recent efforts have attempted to adapt CLIP for transductive TTA. For example, CLIPArTT [14] optimizes the model by jointly utilizing all images within a batch. However, since cross-instance relationships within a batch are often imbalanced, directly modeling all samples together does not effectively exploit the most informative relationships. In this work, we take a step further by identifying and leveraging the informative cliques, groups of strongly correlated samples within a batch, to facilitate robust adaptation by capturing discriminative attributes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 616, + 370, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 370, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 370, + 628 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 635, + 481, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 481, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 481, + 648 + ], + "type": "text", + "content": "3.1. Problem Setting and Notations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "In transductive TTA, test data arrive sequentially in batches. Given " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "-th test batch " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "B_{t} = \\{x_{j}\\}_{j=1}^{b}" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " is the batch size and " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " is a test image. We generate a set of Supportive Cliques for each class " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i = \\{C_{ik}\\}_{k=1}^m" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": ". Then, we extract the corresponding clique attributes, which are de" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30034" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "noted as " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i = \\{a_{ik} \\in \\mathbb{R}^d\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is the attribute dimension, equivalent to the feature dimension. Additionally, we introduce Attribute Prompts " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^A = \\{\\mathcal{P}_i^I, \\mathcal{P}_i^T\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", which consist of visual attribute prompts " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^I = \\{P_{ik}^I\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and text attribute prompts " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^T = \\{P_{ik}^T\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Our model " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is built on CLIP, which comprises an image encoder " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_I" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_T" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 161, + 284, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 284, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 284, + 175 + ], + "type": "text", + "content": "3.2. Attribute-Driven Supportive Clique Mining" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 251 + ], + "type": "text", + "content": "Attributes are the fine-grained characteristics shared by a group of images in the batch. Since images containing the same attribute often exhibit high visual similarity to each other, in this section, we propose to obtain Supportive Cliques for the potential attributes by evaluating cross-image similarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": "Specifically, we use CLIP to obtain the top-" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": " class predictions " + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "topk(x_j)" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "x_j" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": " is a hyperparameter. For each class " + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": ", all images whose top-" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": " predictions include class " + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": " are selected to form a subset " + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 55, + 251, + 296, + 299 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 304, + 295, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 304, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 110, + 304, + 295, + 317 + ], + "type": "interline_equation", + "content": "S _ {i} = \\left\\{x _ {j} \\in B _ {t}: i \\in \\operatorname {t o p k} \\left(x _ {j}\\right) \\right\\} \\tag {1}", + "image_path": "e45eff6ecaadb6d363d12772c4cebe01f6a12d2c3c369a529316bddd803d3749.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "content": "Since images in " + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "content": " share similar class-specific information, which ensures semantic relevance. To further capture more fine-grained relevance, we extract the features " + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "inline_equation", + "content": "F_{i}^{I}\\in" + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{|S_i|\\times d}" + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "content": " of images within " + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "content": " using the CLIP image encoder " + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_I" + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "content": " and compute the similarity matrix " + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "inline_equation", + "content": "M_{i}\\in \\mathbb{R}^{|S_{i}|\\times |S_{i}|}" + }, + { + "bbox": [ + 55, + 321, + 296, + 390 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 390, + 294, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 390, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 138, + 390, + 294, + 406 + ], + "type": "interline_equation", + "content": "M _ {i} = F _ {i} ^ {I} \\cdot \\left(F _ {i} ^ {I}\\right) ^ {\\top} \\tag {2}", + "image_path": "b69dddfeab618d1fe8abddbe58b79c2db70a3e475a6189e7dd36aea74dba4939.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": "Then, given the " + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": "-th row of " + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": ", we collect elements exceeding a threshold " + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": ", whose corresponding images form a Clique " + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "inline_equation", + "content": "C_{il}" + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": ". Formally, the Clique generated from the " + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": "-th row of " + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 467, + 294, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 467, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 104, + 467, + 294, + 482 + ], + "type": "interline_equation", + "content": "C _ {i l} = \\left\\{x _ {j} ^ {i}: x _ {j} ^ {i} \\in S _ {i}, (M _ {i}) _ {l j} > t \\right\\}, \\tag {3}", + "image_path": "b8e0eca3f2e0de2e8a0ed35f8bdbd168eb411360c65e3ea8d0f65cd9da2efe1c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "(M_{i})_{lj}" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "-th row, " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": ". Only non-trivial Cliques " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "(|C_{il}| > 1)" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": " are retained, and duplicate Cliques are removed to ensure uniqueness. Finally, we obtain a set of valid Cliques " + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i = \\{C_{ik}\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 483, + 296, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 536, + 206, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 206, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 206, + 550 + ], + "type": "text", + "content": "3.3. Attribute Prompt Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "In this section, we aim to learn prompts that capture attribute-specific knowledge to facilitate classification." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": "Specifically, for each Clique " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "C_{ik}" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": ", we learn a visual attribute prompt " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "P_{ik}^{A}" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " from the image modality and a text attribute prompt " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "P_{ik}^{T}" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " from the text modality. The feature " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "f_{j}^{I} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": "-th image in " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "C_{ik}" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " is extracted as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 136, + 632, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 632, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 136, + 632, + 295, + 647 + ], + "type": "interline_equation", + "content": "f _ {j} ^ {I} = \\mathbf {E} _ {I} \\left(C _ {i k} ^ {j}, P _ {k} ^ {I}\\right) \\tag {4}", + "image_path": "b21172e0efd97da075b0d3355e1d83cac8374c513ab0ee9fd5acb2daaa082dbc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": "Then, we define the attribute feature " + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "inline_equation", + "content": "a_{ik} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "inline_equation", + "content": "C_{ik}" + }, + { + "bbox": [ + 55, + 651, + 295, + 677 + ], + "type": "text", + "content": " which is calculated by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 133, + 681, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 681, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 133, + 681, + 295, + 715 + ], + "type": "interline_equation", + "content": "a _ {i k} = \\frac {1}{\\left| C _ {i k} \\right|} \\sum_ {j = 1} ^ {\\left| C _ {i k} \\right|} f _ {j} ^ {I} \\tag {5}", + "image_path": "c25d78569049c69bbdb9791055fd65c52a46a94fbc0bc62daa562c261f9763a1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "type": "text", + "content": "Besides, we compute the text feature " + }, + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "type": "inline_equation", + "content": "f_{n}^{T}" + }, + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 554, + 95 + ], + "type": "text", + "content": "-th class as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 395, + 106, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 106, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 395, + 106, + 553, + 121 + ], + "type": "interline_equation", + "content": "f _ {n} ^ {T} = \\mathbf {E} _ {T} \\left(c _ {n}, P _ {i k} ^ {T}\\right) \\tag {6}", + "image_path": "1c2fd71eb9fe1685aa947c960f39985a93391d51ea007dbe1debc660461c0184.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 125, + 554, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 554, + 150 + ], + "type": "text", + "content": "Given total " + }, + { + "bbox": [ + 313, + 125, + 554, + 150 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 125, + 554, + 150 + ], + "type": "text", + "content": " classes, we introduce an attribute-specific class probability distribution vector, which is obtained by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 383, + 157, + 553, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 157, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 383, + 157, + 553, + 172 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {P}} _ {i k} = \\left[ \\hat {p} _ {i k} ^ {1}, \\hat {p} _ {i k} ^ {2}, \\dots , \\hat {p} _ {i k} ^ {N} \\right] \\tag {7}", + "image_path": "1373576f109cac0167c977cd4d6cf3f6112085e4c2604e8aecb4409748d080d6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 180, + 342, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 342, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 342, + 190 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 361, + 188, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 188, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 361, + 188, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {i k} ^ {n} = \\frac {\\exp \\left(\\cos \\left(a _ {i k} , f _ {n} ^ {T}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\cos \\left(a _ {i k} , f _ {j} ^ {T}\\right) / \\tau\\right)} \\tag {8}", + "image_path": "479e128cfb6e99c32ae280f3de4dec68e62a9ce03cf023e4c74bd72fe8609d89.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": " denotes cosine similarity and " + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 224, + 554, + 248 + ], + "type": "text", + "content": " is a temperature parameter to scale the similarity scores." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 248, + 554, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 554, + 271 + ], + "type": "text", + "content": "Then, we compute the Entropy Loss for the probability distribution of the attribute:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 372, + 277, + 553, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 277, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 372, + 277, + 553, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {e n} \\left(C _ {i k}\\right) = - \\sum_ {n = 1} ^ {N} \\hat {p} _ {i k} ^ {n} \\log \\hat {p} _ {i k} ^ {n} \\tag {9}", + "image_path": "325368de8ddffbb2e9a1c09dc7e26ce83e540d900d510b569ec47e4f3068fcb8.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "type": "text", + "content": "To further guide the model to extract attribute-specific information and discard attribute-irrelevant information, we introduce a concentration loss " + }, + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}(C_{ik})" + }, + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "type": "text", + "content": " that forces the model to capture the shared information across the samples within a Clique. Specifically, " + }, + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}(C_{ik})" + }, + { + "bbox": [ + 313, + 318, + 554, + 380 + ], + "type": "text", + "content": " is calculated by" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 369, + 386, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 386, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 369, + 386, + 553, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} \\left(C _ {i k}\\right) = \\sum_ {j = 1} ^ {| C _ {i k} |} \\left| \\left| f _ {j} ^ {I} - a _ {i k} \\right| \\right| ^ {2} \\tag {10}", + "image_path": "2e06f59854555f0964991e7166c53206fc1df71d0f8b3a3abbd89b8ea1b8c414.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "type": "text", + "content": "Therefore, both kinds of prompts " + }, + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "type": "inline_equation", + "content": "P_{ik}^{I}" + }, + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "type": "inline_equation", + "content": "P_{ik}^{T}" + }, + { + "bbox": [ + 313, + 430, + 554, + 479 + ], + "type": "text", + "content": " are guided to encode attribute-specific knowledge as the model optimization. Such knowledge is valuable when generating predictions for images showing the same attributes." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 486, + 468, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 486, + 468, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 486, + 468, + 499 + ], + "type": "text", + "content": "3.4. Attribute Prompt Retention" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 503, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 552 + ], + "type": "text", + "content": "In this section, to retain attribute knowledge during the test process and ensure consistent performance improvement, we introduce mechanisms for retaining both text and visual attribute knowledge." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": "Firstly, to retain the text attribute knowledge, we propose a Text Retention Prompt " + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\tilde{P}^T" + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": ". Specifically, when a new Clique " + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "inline_equation", + "content": "C_{ik}" + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": " is learned, we update " + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\tilde{P}^T" + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 373, + 594, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 594, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 373, + 594, + 553, + 609 + ], + "type": "interline_equation", + "content": "\\tilde {P} ^ {T} = \\alpha_ {\\tau} \\tilde {P} ^ {T} + (1 - \\alpha_ {\\tau}) P _ {i k} ^ {T}, \\tag {11}", + "image_path": "8664114ea79c82233dc7299c5ebb42cc9a8ceffd25fa57b54dd9351eb7d02a6c.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\alpha_{\\tau} = \\tau /(1 + \\tau)" + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "content": " is a decay rate based on iteration count " + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": "Besides, to retain the visual attribute knowledge, we design a Retention Cache " + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{R_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": " is a class-specific retention cache. Specifically, we store " + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "inline_equation", + "content": "(a_{ik}, P_{ik}^I)" + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": " as key-value pairs in " + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": ". We denote the retained attributes (i.e. keys) as " + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "inline_equation", + "content": "R_i^A = \\{a_1, \\ldots, a_L\\}" + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": " and the corresponding attribute prompts (i.e. values) as " + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "inline_equation", + "content": "R_i^P = \\{P_1^I, \\ldots, P_L^I\\}" + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30035" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "maximum size of each " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " to avoid memory explosion. When the retained key-value pair number is less than " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", we directly add a new " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "(a_{ik}, P_{ik}^I)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". Otherwise, we add " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "(a_{ik}, P_{ik}^I)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " first, then we update all pairs and aggregate the two that are most similar to each other in " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " by using a graph structure as follows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": "Firstly, we construct a weighted graph " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "G_{i} = (V_{i},E_{i})" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": ". The vertices are the keys in the retention cache " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "V_{i} = R_{i}^{A} = \\{a_{1},\\dots ,a_{L + 1}\\}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": ". The edges " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " between each two vertices are defined as the Gaussian Distance of the two attributes. Thus, the adjacency matrix " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "W_{i}\\in \\mathbb{R}^{(L + 1)\\times (L + 1)}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 55, + 144, + 296, + 216 + ], + "type": "text", + "content": " is defined with:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 223, + 295, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 223, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 113, + 223, + 295, + 249 + ], + "type": "interline_equation", + "content": "\\left(W _ {i}\\right) _ {k l} = \\exp \\left(- \\frac {\\left\\| a _ {k} - a _ {l} \\right\\| ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {12}", + "image_path": "4821dbeae7f24829a3cbc934427fcbaea8396509a7790b4e37a40ddfbb583c7d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "(W_{i})_{kl}" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": "-th row, " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": "-column of " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": " is a hyperparameter. We utilize the " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": "-max values in each row of " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": " to construct a " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": "-nearest retention graph. Then we perform graph propagation following [53] and obtain the updated graph vertices " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "V_{i}^{\\prime} = \\{a_{1}^{\\prime},\\ldots ,a_{L + 1}^{\\prime}\\}" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": ", where the similar attributes are pushed closer automatically. Then, the most similar two attributes " + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "inline_equation", + "content": "a_{j_0}^\\prime ,a_{j_1}^\\prime" + }, + { + "bbox": [ + 55, + 257, + 296, + 354 + ], + "type": "text", + "content": " and the corresponding prompts are fused into one respectively as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 362, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 362, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 124, + 362, + 295, + 409 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} a _ {n e w} = \\frac {1}{2} \\left(a _ {j _ {0}} ^ {\\prime} + a _ {j _ {1}} ^ {\\prime}\\right) \\\\ P _ {n e w} ^ {I} = \\frac {1}{2} \\left(P _ {j _ {0}} ^ {I} + P _ {j _ {1}} ^ {I}\\right) \\end{array} \\right. \\tag {13}", + "image_path": "3469c0ef4c219fe0da1b668a49ab094cd0050474d2bf01bdbcb89e078ccf6855.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "type": "text", + "content": "Through the above process, we obtain the updated retention cache " + }, + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "type": "text", + "content": ", dynamically retaining the visual and text attribute knowledge learned from previous batches into Retention Cache and Text Retention Prompt, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 208, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 208, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 208, + 487 + ], + "type": "text", + "content": "3.5. Optimization and Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "text", + "content": "For each batch " + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "inline_equation", + "content": "B_{t}" + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "text", + "content": ", we optimize the visual attribute prompts " + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^I = \\{P_{ik}^I\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "text", + "content": " and text attribute prompts " + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^T = \\{P_{ik}^T\\}_{k=1}^m" + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "text", + "content": " for every class " + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 492, + 296, + 529 + ], + "type": "text", + "content": " by minimizing the following loss function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 537, + 295, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 537, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 295, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {k = 1} ^ {m} \\mathcal {L} _ {e n} \\left(C _ {i k}\\right) + \\lambda \\mathcal {L} _ {c o n} \\left(C _ {i k}\\right) \\tag {14}", + "image_path": "db530de222d21ed218e28ebab112951983624d485f2cad7ee4481fa06f4fef51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 574, + 174, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 174, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 174, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 574, + 174, + 586 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 574, + 174, + 586 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": "Then we conduct inference for each " + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "inline_equation", + "content": "B_{t}" + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": ". To fully utilize both the Attribute Prompts learned from the current batch " + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "inline_equation", + "content": "B_{t}" + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": " and the retained knowledge to improve prediction. Given the test image " + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "inline_equation", + "content": "x_{j} \\in S_{i}" + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": ", we first collect all Cliques containing " + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": ", namely " + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = \\{k : x_{j} \\in C_{ik}\\}" + }, + { + "bbox": [ + 55, + 586, + 296, + 668 + ], + "type": "text", + "content": ", whose corresponding visual attribute prompts are concatenated together:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 669, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 669, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 295, + 684 + ], + "type": "interline_equation", + "content": "P ^ {I} = \\operatorname {c o n c a t} \\left(\\left[ P _ {i k} ^ {I}: k \\in \\mathcal {K} \\right]\\right) \\tag {15}", + "image_path": "6efc7e938e5015b18fdc79c0008e109f0d17491db6c99d7adb7b6727387efbb5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Then, we introduce the retention prompt matching strategy utilizing the key-value pair structure of the retention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": "cache. Specifically, the keys are the retained attribute features " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "R_{i}^{A} = \\{a_{1},\\dots ,a_{L}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": " and the values are the corresponding attribute prompts " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "R_{i}^{P} = \\{P_{1}^{I},\\dots ,P_{L}^{I}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": ". Given the image feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "f_{j}^{I}" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": " which is already extracted in Sec. 3.3, the top-1 attribute " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "a^*" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": " is obtained by evaluating the similarity of " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "f_{j}^{I}" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": " to all keys in " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "R_{i}^{A}" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 152, + 553, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 152, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 381, + 152, + 553, + 175 + ], + "type": "interline_equation", + "content": "a ^ {*} = \\underset {a _ {l} \\in R _ {i} ^ {A}} {\\operatorname {a r g m a x}} < f _ {j} ^ {I}, a _ {l} > \\tag {16}", + "image_path": "69deb3e19b9506b1f17c20996ba0ca7988b6a7cc11fc66efd777d0c91504f070.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": "Given the matched value corresponding to " + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "inline_equation", + "content": "a^*" + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "inline_equation", + "content": "P^{I*}" + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": ", we obtain the Composed Image Prompt " + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "inline_equation", + "content": "P_C^I" + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": " by concatenating " + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "inline_equation", + "content": "P^I" + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "inline_equation", + "content": "P^{I*}" + }, + { + "bbox": [ + 313, + 182, + 553, + 218 + ], + "type": "text", + "content": " along the token dimension" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 384, + 224, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 224, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 384, + 224, + 553, + 239 + ], + "type": "interline_equation", + "content": "P _ {C} ^ {I} = \\operatorname {c o n c a t} \\left(P ^ {I}, P ^ {I ^ {*}}\\right) \\tag {17}", + "image_path": "cda93abdadcb4af91f5486896e7f32ba73be4b032d1c1b169dd6505d5bb3a54a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "text", + "content": "For the text modality, we fuse " + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathcal{T}} = \\{P_{ik}^{T}\\}_{k=1}^{m}" + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\tilde{P}^T" + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "text", + "content": " together to get the Composed Text Prompt " + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "inline_equation", + "content": "P_C^T" + }, + { + "bbox": [ + 313, + 243, + 553, + 270 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 276, + 553, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 361, + 276, + 553, + 307 + ], + "type": "interline_equation", + "content": "P _ {C} ^ {T} = \\alpha_ {r} \\tilde {P} ^ {T} + (1 - \\alpha_ {r}) \\frac {1}{m} \\sum_ {k = 1} ^ {m} P _ {i k} ^ {T} \\tag {18}", + "image_path": "53a444252be1e361006cc8638a50af91a5e3b28efa98a4c73f735dee64854ecb.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\alpha_{r}" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " is the weight of retained text prompts. Then, for the final prediction, the image-inference feature " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\hat{f}_j^I" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "x_j^i" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " and the text-inference feature " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\hat{f}_n^T" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " are obtained by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 391, + 361, + 553, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 361, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 391, + 361, + 553, + 395 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\hat {f} _ {j} ^ {I} = \\mathbf {E} _ {I} \\left(x _ {j}, P _ {C} ^ {I}\\right) \\\\ \\hat {f} _ {n} ^ {T} = \\mathbf {E} _ {T} \\left(c _ {n}, P _ {C} ^ {T}\\right) \\end{array} \\right. \\tag {19}", + "image_path": "a743049ce582e0c83462656d010480e9e8c5db95c1b458948859185e0d6bf316.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "spans": [ + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "type": "text", + "content": "The prediction " + }, + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 326, + 396, + 541, + 410 + ], + "type": "text", + "content": " is obtained by applying Eq.7." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 418, + 395, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 395, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 395, + 431 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 438, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 533 + ], + "type": "text", + "content": "In this section, we first introduce the two benchmarks used for model evaluation, followed by a detailed description of the implementation setup. Next, we present a comparative analysis of existing methods and our proposed approach. Finally, extensive ablation studies are conducted to validate the effectiveness of SCAP. Additional experimental results and in-depth discussions are provided in the Supplementary Materials." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 540, + 397, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 397, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 397, + 552 + ], + "type": "text", + "content": "4.1. Benchmarks" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Our evaluation is conducted on two benchmarks: an out-of-distribution (OOD) benchmark and a cross-domain benchmark. The OOD benchmark consists of four datasets derived from ImageNet [4], i.e., ImageNet-A [16], ImageNet-R [15], ImageNet-Sketch [40], and ImageNet-V2 [34]. The OOD benchmark aims to assess the model's robustness to substantial distribution shifts in the test data. The cross-domain benchmark comprises four datasets of the image classification task, each representing a distinct domain: Aircraft [27], Caltech101 [8], Flower102 [29], and Pets [30]. The cross-domain benchmark provides a measure of the model's adaptability across a diverse range of domain-specific tasks." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30036" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 535, + 236 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 535, + 236 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 535, + 236 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 535, + 236 + ], + "type": "table", + "html": "
MethodPublicationImageNet-AImageNet-RImageNet-SImageNet-V2Average
CLIP [32]ICML 202147.8773.9846.0961.8857.46
CoOp [57]IJCV 202249.7175.2147.9964.2059.28
CoCoOp [56]CVPR 202250.6376.1848.7564.0759.91
Tip-Adapter [51]ECCV 202251.0477.7648.8863.4160.27
TPT [36]NeurIPS 202254.7777.0647.9463.4560.81
C-TPT [47]ICLR 202452.9078.0048.5063.4060.70
DART [25]AAAI 202460.5679.5649.7664.0363.48
MTA [49]CVPR 202458.0678.3349.6164.2462.56
TDA [20]CVPR 202460.1180.2450.5464.6763.89
Zero [7]NeuIPS 202461.3577.2848.2964.1362.76
SCAPThis Paper64.5281.6851.6564.6565.63
", + "image_path": "ca296c6e9c0ac0487a23e3afabef7c94b5aac5582d600e4f993ac8f0678bc964.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 268 + ], + "type": "text", + "content": "Table 1. Results comparison on the out-of-distribution (OOD) Benchmark. The results of the compared methods follow the original paper or are implemented with the official code [25]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 277, + 175, + 373 + ], + "blocks": [ + { + "bbox": [ + 60, + 277, + 175, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 277, + 175, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 277, + 175, + 373 + ], + "type": "image", + "image_path": "146f629d2924afda08ff1c390a5025955ec95f3bd2ed7b27c814f7f1780c015b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 407 + ], + "type": "text", + "content": "Figure 3. Results on the Cross-Domain Benchmark. Comparison of SCAP with state-of-the-art methods. All methods are evaluated using CLIP-ViT-B/16 as the backbone." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 183, + 277, + 288, + 373 + ], + "blocks": [ + { + "bbox": [ + 183, + 277, + 288, + 373 + ], + "lines": [ + { + "bbox": [ + 183, + 277, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 183, + 277, + 288, + 373 + ], + "type": "image", + "image_path": "2405584b2bdab9b03f18c705ee00d940acaaf1ddcf25cef570866618cd0b1bf4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 295, + 277, + 406, + 373 + ], + "blocks": [ + { + "bbox": [ + 295, + 277, + 406, + 373 + ], + "lines": [ + { + "bbox": [ + 295, + 277, + 406, + 373 + ], + "spans": [ + { + "bbox": [ + 295, + 277, + 406, + 373 + ], + "type": "image", + "image_path": "30ad29488f3bb38f8f8a5250796452380dc534a87f573f09144d6a199d955143.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 414, + 277, + 548, + 373 + ], + "blocks": [ + { + "bbox": [ + 414, + 277, + 548, + 373 + ], + "lines": [ + { + "bbox": [ + 414, + 277, + 548, + 373 + ], + "spans": [ + { + "bbox": [ + 414, + 277, + 548, + 373 + ], + "type": "image", + "image_path": "e9f79901e069342977655652326c00b30417e3110dea6a0b5d534e4e4df88497.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": "Implementation details. We adopt CLIP with ViT-B/16 as the backbone and set the batch size to 64 for all experiments. Visual attribute prompts are initialized using a uniform distribution in the range " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "(-1, 1)" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", following previous TTA methods [25, 36]. For text attribute prompts, we employ an ensemble of multiple templates for the OOD benchmark, whereas dataset-specific templates are used for the cross-domain benchmark, following TDA [20]. Our key hyperparameters include the threshold " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": " for supportive clique extraction, the weight " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\alpha_{r}" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": " for computing " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "P_C^T" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", the concentration loss weight " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", and graph structure parameters " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ". The hyperparameter ablations of this paper are conducted on ImageNet-A, with the default values of " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\alpha_{r}" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": " set to 1, 0.3, 6, and 3, respectively. We set " + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 54, + 427, + 297, + 643 + ], + "type": "text", + "content": " in the initial top-k predictions Eq. (1). An Adam optimizer with a learning rate of 0.003 is used to update the prompts, and all experiments are conducted on a single NVIDIA 4090 GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 648, + 241, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 241, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 241, + 661 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "Compared methods. We compare our proposed SCAP with state-of-the-art TTA methods designed for CLIP adaptation, including TPT [36], C-TPT [47], DART [25], MTA [49], TDA [20], and Zero [7]. Additionally, we report" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 500 + ], + "type": "text", + "content": "results from the original CLIP model to highlight the knowledge gained during testing. Following previous works [20], we also include training-time adaptation methods such as CoOp [57], CoCoOp [56], and Tip-Adapter [51] for comparison to further verify the label efficiency and effectiveness of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "text", + "content": "Results on the OOD Benchmark. As shown in Tab. 1, our proposed SCAP achieves state-of-the-art performance on the ImageNet-A, ImageNet-R, and ImageNet-S datasets, outperforming existing methods by at least " + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "inline_equation", + "content": "3.17\\%" + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "inline_equation", + "content": "1.44\\%" + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "inline_equation", + "content": "1.11\\%" + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "text", + "content": ", respectively. Additionally, on the challenging ImageNet-V2 dataset, SCAP performs comparably to TDA, with only a marginal " + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 313, + 504, + 556, + 648 + ], + "type": "text", + "content": " difference. While TDA benefits from a feature cache to enable joint learning across multiple samples, it incurs additional storage overhead. In contrast, SCAP achieves comparable or superior performance without maintaining historical features, making it more efficient in practice." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Overall, SCAP improves average accuracy across the four ImageNet variants by " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.74\\%" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": ", highlighting its effectiveness in handling natural object images across diverse distribution shifts. This performance gain is primarily attributed to the proposed supportive clique mining strategy, which ef" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30037" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 291, + 159 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 291, + 159 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 291, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 291, + 159 + ], + "type": "table", + "html": "
Components in SCAPImageNet-A
ImageTextRetentionAcc@1
XXX60.98
X✓X62.57
✓XX63.14
✓X✓63.82
✓✓✓64.52
", + "image_path": "db48e0c2e6a6290ba8605a27506463020bddd943138f1805d57b3128916bbeed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 186, + 283, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 283, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 283, + 225 + ], + "type": "text", + "content": "- Image and Text represent learning from the Image Modality and Text Modality in the Attribute Prompt Learning module, respectively. \n- Retention refers only to the Retention Cache for the visual modality while the text retention prompts are always used." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 167, + 277, + 178 + ], + "lines": [ + { + "bbox": [ + 74, + 167, + 277, + 178 + ], + "spans": [ + { + "bbox": [ + 74, + 167, + 277, + 178 + ], + "type": "text", + "content": "Table 2. Ablation study about the different components" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 255, + 296, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 290 + ], + "type": "text", + "content": "fectively captures diverse object attributes, such as \"head\" and \"body\", that enhance the object recognition capacity of the model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "text", + "content": "Results over cross-domain benchmark. As shown in Fig. 3, we compare our SCAP method with state-of-the-art approaches across four cross-domain datasets. The results demonstrate that our method surpasses the state-of-the-art approaches, with improvements of at least " + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "inline_equation", + "content": "1.64\\%" + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "inline_equation", + "content": "0.23\\%" + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "inline_equation", + "content": "0.18\\%" + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "inline_equation", + "content": "0.24\\%" + }, + { + "bbox": [ + 55, + 293, + 296, + 424 + ], + "type": "text", + "content": " on the Pets, Flower102, Caltech101, and Aircraft datasets, respectively. This superior performance stems from the proposed Attribute Prompt Learning and Attribute Prompt Retention mechanisms, which effectively capture and leverage generalizable knowledge shared across diverse samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 156, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 156, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 156, + 443 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 605 + ], + "type": "text", + "content": "The Influence of Each Component. To evaluate the effectiveness of our module designs across different modalities (i.e., image, text, and retention), we conduct modality-specific ablation studies. As shown in Tab. 2, leveraging both image and text modalities within the Attribute Prompt Learning module consistently improves performance over the baseline model. Additionally, incorporating the Attribute Prompt Retention module further enhances the utilization of learned image prompts, leading to substantial performance gains. Furthermore, when all modules and modalities are integrated, the model performance improves further due to joint optimization, which mitigates semantic drift as learnable parameters are updated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "The Influence of Batch Size in SCAP. Batch Size " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " is an important factor in the transductive TTA setting, as larger batches provide richer inter-image relationships. As illustrated in Fig. 4 (Left), the model performance consistently improves with increasing batch size, demonstrating SCAP's capability to leverage the additional contextual information. We also observe that the image modality varies significantly as batch size changes, while the text modality remains relatively stable across batch size variations. On" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 435, + 167 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 435, + 167 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 435, + 167 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 435, + 167 + ], + "type": "image", + "image_path": "cfa5c169664d6e9e4b56830bf72a7e68adcb82969199e98ba9a5404c343feb52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 181, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 213 + ], + "type": "text", + "content": "Figure 4. Left: the influence of batch size on SCAP performance. Right: the average maximum Clique size generated per batch under different batch sizes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 72, + 552, + 167 + ], + "blocks": [ + { + "bbox": [ + 438, + 72, + 552, + 167 + ], + "lines": [ + { + "bbox": [ + 438, + 72, + 552, + 167 + ], + "spans": [ + { + "bbox": [ + 438, + 72, + 552, + 167 + ], + "type": "image", + "image_path": "08c1ef4d26770d845c2a2c5180111a406eaf1537cf1c15608278c7073a7778ee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 230, + 439, + 321 + ], + "blocks": [ + { + "bbox": [ + 319, + 230, + 439, + 321 + ], + "lines": [ + { + "bbox": [ + 319, + 230, + 439, + 321 + ], + "spans": [ + { + "bbox": [ + 319, + 230, + 439, + 321 + ], + "type": "image", + "image_path": "a95f19d56c8f23680964b96e34d658a869dd6c844d6c751be04a83d40ddf15bd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 442, + 228, + 552, + 321 + ], + "blocks": [ + { + "bbox": [ + 442, + 228, + 552, + 321 + ], + "lines": [ + { + "bbox": [ + 442, + 228, + 552, + 321 + ], + "spans": [ + { + "bbox": [ + 442, + 228, + 552, + 321 + ], + "type": "image", + "image_path": "833359702d9529a1910ef1a8050c8f6b16d2bb95d3a1cbc718cf53449fa7b338.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 326, + 438, + 415 + ], + "blocks": [ + { + "bbox": [ + 320, + 326, + 438, + 415 + ], + "lines": [ + { + "bbox": [ + 320, + 326, + 438, + 415 + ], + "spans": [ + { + "bbox": [ + 320, + 326, + 438, + 415 + ], + "type": "image", + "image_path": "18f79ebf935cbf8dda60b30d5f21ff9636092c14c9457729a2a7fa7cd61513bf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 429, + 553, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 553, + 450 + ], + "type": "text", + "content": "Figure 5. Ablation study about the influence of different hyperparameters on ImageNet-A." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 326, + 551, + 415 + ], + "blocks": [ + { + "bbox": [ + 442, + 326, + 551, + 415 + ], + "lines": [ + { + "bbox": [ + 442, + 326, + 551, + 415 + ], + "spans": [ + { + "bbox": [ + 442, + 326, + 551, + 415 + ], + "type": "image", + "image_path": "2ae297e278b0568dfb5c8e8d8430444865b5fcd9a8c3b082f0bce81d15b57424.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 473, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 509 + ], + "type": "text", + "content": "the right side of Fig. 4, the results reveal that larger batches lead to increasing average clique size, indicating that more support information can be captured under such conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "The Influence of Different Hyperparameters. The key hyperparameters in our approach include the threshold " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " for clique extraction, the concentration loss weight " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ", the retention cache limit " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ", and the degree of the retention graph " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ". As illustrated in Fig. 5 (a), the model performance is relatively sensitive to changes in " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ". A lower " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " results in cliques containing unrelated images, which hinders effective attribute prompt learning, whereas a higher " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " limits the number of support samples, leading to insufficient attribute modeling. Subsequently, we explore the impact of " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " in Fig. 5 (b). The model performance improves consistently up to " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ", as the concentration loss enhances attribute learning. However, further increasing " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " leads to overfitting, negatively affecting generalization. Then, we investigate the influence of " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " in Fig. 5 (c). We observe that as " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " increases, the performance increases and then becomes stable, indicating that while a larger cache enhances retention," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30038" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 293 + ], + "type": "image", + "image_path": "d07aebf70c601d87de011877379514f89052e9dbd3e3114c855169b90d7ca043.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 302, + 526, + 314 + ], + "lines": [ + { + "bbox": [ + 83, + 302, + 526, + 314 + ], + "spans": [ + { + "bbox": [ + 83, + 302, + 526, + 314 + ], + "type": "text", + "content": "Figure 6. Visualization of our prompt-guided attention maps in comparison with the state-of-the-art prompt-based method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 332, + 291, + 416 + ], + "blocks": [ + { + "bbox": [ + 61, + 332, + 291, + 416 + ], + "lines": [ + { + "bbox": [ + 61, + 332, + 291, + 416 + ], + "spans": [ + { + "bbox": [ + 61, + 332, + 291, + 416 + ], + "type": "table", + "html": "
Combine PIkCombine PI, P*ImageNet-A
MeanConcatMeanConcatAcc@1
✓X✓X61.82
✓XX✓62.11
X✓✓X64.35
X✓X✓64.52
", + "image_path": "36ddeec2c5066de92857751429bee77e8db6c50f7b7966cbb326c61e6d0ffffb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 424, + 284, + 435 + ], + "lines": [ + { + "bbox": [ + 66, + 424, + 284, + 435 + ], + "spans": [ + { + "bbox": [ + 66, + 424, + 284, + 435 + ], + "type": "text", + "content": "Table 3. The influence of different combination approaches." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "type": "text", + "content": "excessive stored knowledge does not yield further gains. Finally, we study the influence of " + }, + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "type": "text", + "content": " in Fig. 5 (d). Increasing " + }, + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 459, + 295, + 543 + ], + "type": "text", + "content": " up to 2 improves performance by strengthening graph connectivity and enhancing information propagation. However, beyond this point, performance declines due to the inclusion of irrelevant nodes, which introduces noise and reduces the discriminative power of the learned attributes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "type": "text", + "content": "Influence of Different Prompt Combination Approaches. Two primary strategies for combining multiple prompts are considered: generating a mean prompt and concatenating prompts along the token dimension, as illustrated in Eq. (15) and in Eq. (17). Experimental results in Table Tab. 3 indicate that concatenation outperforms averaging in our SCAP framework. This advantage due to concatenation preserves the full prompt information, enabling each prompt to focus on distinct image regions. In contrast, the mean prompt approach struggles to capture diverse visual patterns, leading to suboptimal performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Visualization of Supportive Cliques. To demonstrate the effectiveness of our supportive clique-based attribute prompt learning approach, we visualize the attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 431 + ], + "type": "text", + "content": "maps in comparison with the state-of-the-art prompt-based method DART [25]. As illustrated in Fig. 6, each supportive clique in our method guides the prompts to focus on specific object parts, such as the head, leg, or body. When the attribute prompts are exploited together, the obtained attention maps precisely and comprehensively capture the object information, highlighting the superior feature-learning capability of our approach over DART." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 654 + ], + "type": "text", + "content": "In this paper, we propose a novel transductive test-time adaptation approach, Supportive Clique-based Attribute Prompting (SCAP), which leverages intrinsic cross-sample shared visual attributes to enhance model adaptation during testing. Specifically, our method identifies and models the attributes within batches through supportive cliques, sets of samples exhibiting similar patterns. Furthermore, a retention module is designed to dynamically store, aggregate, and utilize knowledge from previously learned attributes, thereby improving predictions for subsequent test images. Extensive experiments on two benchmarks demonstrate the effectiveness of our method in leveraging batch-wise information. The proposed approach offers a promising direction for fully exploiting cross-sample relationships in practical transductive TTA scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 657, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 556, + 703 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the grants from the National Natural Science Foundation of China (62376011, 61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006)." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30039" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi. Contrastive test-time adaptation. In CVPR, pages 295-305, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "text", + "content": "[3] Zhenyu Cui, Yuxin Peng, Xun Wang, Manyu Zhu, and Jiuhuan Zhou. Continual vision-language retrieval via dynamic knowledge rectification. In AAAI, pages 11704-11712, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 270 + ], + "type": "text", + "content": "[5] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, pages 11162-11173, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 306 + ], + "type": "text", + "content": "[6] Mario Döbler, Robert A Marsden, and Bin Yang. Robust mean teacher for continual and gradual test-time adaptation. In CVPR, pages 7704-7714, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "text", + "content": "[7] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, Elisa Ricci, et al. Frustratingly easy test-time adaptation of vision-language models. In NeurIPS, 2024. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "text", + "content": "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178-178. IEEE, 2004. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 294, + 441 + ], + "type": "text", + "content": "[9] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In ICCV, pages 2704-2714, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 443, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 443, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 57, + 443, + 295, + 497 + ], + "type": "text", + "content": "[10] Lihua Fu, Yubin Du, Yu Ding, Dan Wang, Hanxu Jiang, and Haitao Zhang. Domain adaptive learning with multi-granularity features for unsupervised person re-identification. Chinese Journal of Electronics, 31(1):116-128, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 499, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 294, + 543 + ], + "type": "text", + "content": "[11] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "text", + "content": "[12] Yulu Gan, Yan Bai, Yihang Lou, Xianzheng Ma, Renrui Zhang, Nian Shi, and Lin Luo. Decorate the newcomers: Visual domain prompt for continual test time adaptation. In AAAI, pages 7595-7603, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 623 + ], + "type": "text", + "content": "[13] Taesik Gong, Yewon Kim, Taeckyung Lee, Sorn Chottananurak, and Sung-Ju Lee. Sotta: Robust test-time adaptation on noisy data streams. NeurIPS, 36, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "text", + "content": "[14] Gustavo Adolfo Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdanpanah, Ismail Ben Ayed, and Christian Desrosiers. Clipartt: Light-weight adaptation of clip to new domains at test time. arXiv preprint arXiv:2405.00754, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, pages 15262-15271, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 553, + 174 + ], + "type": "text", + "content": "[17] Xuefeng Hu, Gokhan Uzunbas, Sirius Chen, Rui Wang, Ashish Shah, Ram Nevatia, and Ser-Nam Lim. Mixnorm: Test-time adaptation through online normalization estimation. arXiv preprint arXiv:2110.11478, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "text", + "content": "[18] Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. Scaling up vision-language pre-training for image captioning. In CVPR, pages 17980-17989, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "text", + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[20] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In CVPR, pages 14162-14171, 2024. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "type": "text", + "content": "[21] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 408 + ], + "type": "text", + "content": "[22] Qiwei Li, Kunlun Xu, Yuxin Peng, and Jiahuan Zhou. Exemplar-free lifelong person re-identification via prompt-guided adaptive knowledge consolidation. IJCV, 132(11): 4850-4865, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "text", + "content": "[23] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. NeurIPS, 36:13448-13466, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[24] Zichen Liu, Yuxin Peng, and Jiahuan Zhou. Inspv: Efficient instance visual prompting from image itself. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6443-6452, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 544 + ], + "type": "text", + "content": "[25] Zichen Liu, Hongbo Sun, Yuxin Peng, and Jiahuan Zhou. Dart: dual-modal adaptive online prompting and knowledge retention for test-time adaptation. In AAAI, pages 14106-14114, 2024. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "text", + "content": "[26] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision-language models. NeurIPS, 36, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 580, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 621 + ], + "type": "text", + "content": "[27] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "text", + "content": "[28] Robert A Marsden, Mario Döbler, and Bin Yang. Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction. In WACV, pages 2555-2565, 2024. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30040" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[30] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In " + }, + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": ", pages 3498-3505. IEEE, 2012. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[31] Ori Press, Steffen Schneider, Matthias Kümmerer, and Matthias Bethge. Rdumb: A simple approach that questions our progress in continual test-time adaptation. NeurIPS, 36: 39915-39935, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[34] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imagenet? In ICML, pages 5389-5400. PMLR, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[35] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. NeurIPS, 33:11539-11551, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 396 + ], + "type": "text", + "content": "[36] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. NeurIPS, 35:14274-14289, 2022. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "text", + "content": "[37] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. Ecotta: Memory-efficient continual test-time adaptation via self-distilled regularization. In CVPR, pages 11920-11929, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "text", + "content": "[38] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "text", + "content": "[39] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[40] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "text", + "content": "[41] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 591, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 635 + ], + "type": "text", + "content": "[42] Kunlun Xu, Haozhuo Zhang, Yu Li, Yuxin Peng, and Jiahuan Zhou. Mitigate catastrophic remembering via continual knowledge purification for noisy lifelong person re-identification. In ACM MM, pages 5790–5799, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "text", + "content": "[43] Kunlun Xu, Xu Zou, Yuxin Peng, and Jiahuan Zhou. Distribution-aware knowledge prototyping for non-exemplar lifelong person re-identification. In CVPR, pages 16604–16613, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[44] Kunlun Xu, Xu Zou, and Jiahuan Zhou. Lstkc: Long short-term knowledge consolidation for lifelong person re-identification. In AAAI, pages 16202–16210, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 599 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[45] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In CVPR, pages 15671-15680, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 119, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 150 + ], + "type": "text", + "content": "[46] Zhaoda Ye, Xiangteng He, and Yuxin Peng. Unsupervised cross-media hashing learning via knowledge graph. Chinese Journal of Electronics, 31(6):1081-1091, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[47] Hee Suk Yoon, Eunseop Yoon, Joshua Tian Jin Tee, Mark Hasegawa-Johnson, Yingzhen Li, and Chang D Yoo. C-tpt: Calibrated test-time prompt tuning for vision-language models via text feature dispersion. arXiv preprint arXiv:2403.14119, 2024. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "text", + "content": "[48] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In CVPR, pages 10899-10909, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 284 + ], + "type": "text", + "content": "[49] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In CVPR, pages 23783-23793, 2024. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[50] Ding-Chu Zhang, Zhi Zhou, and Yu-Feng Li. Robust test-time adaptation for zero-shot prompt tuning. In AAAI, pages 16714-16722, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[51] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, pages 493-510. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 366, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 418 + ], + "type": "text", + "content": "[52] Ruru Zhang, E Haihong, and Meina Song. Fscil-eaca: Few-shot class-incremental learning network based on embedding augmentation and classifier adaptation for image classification. Chinese Journal of Electronics, 33(1):139-152, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[53] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. NeurIPS, 16, 2003. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[54] Jiahuan Zhou, Pei Yu, Wei Tang, and Ying Wu. Efficient online local metric adaptation via negative samples for person re-identification. In ICCV, pages 2420–2428, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 489, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 531 + ], + "type": "text", + "content": "[55] Jiahuan Zhou, Bing Su, and Ying Wu. Online joint multimetric adaptation from frequent sharing-subset mining for person re-identification. In CVPR, pages 2909–2918, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 568, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 599 + ], + "type": "text", + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30041" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_content_list.json b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9b2b79f42510f0a8d38f58e5257c40b8fd8a5683 --- /dev/null +++ b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_content_list.json @@ -0,0 +1,1973 @@ +[ + { + "type": "text", + "text": "SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow", + "text_level": 1, + "bbox": [ + 91, + 130, + 903, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingyuan Wang $^{1}$ , Rui Song $^{1}$ , Jiaojiao Li $^{1}$ , Kerui Cheng $^{2}$ , David Ferstl $^{3}$ , Yinlin Hu $^{3}$ $^{1}$ State Key Laboratory of ISN, Xidian University $^{2}$ Taiyuan University of Technology $^{3}$ MagicLeap", + "bbox": [ + 104, + 179, + 890, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. Most recent 6D object pose methods rely on refinement to get accurate results. However, most existing refinement methods either suffer from noises in establishing correspondences, or rely on retraining for novel objects. SCFlow2 is based on the SCFlow model designed for refinement with shape constraint, but formulates the additional depth as a regularization in the iteration via 3D scene flow for RGBD frames. The key design of SCFlow2 is an introduction of geometry constraints into the training of recurrent matching network, by combining the rigid-motion embeddings in 3D scene flow and 3D shape prior of the target. We train SCFlow2 on a combination of dataset Objaverse, GSO and ShapeNet, and evaluate on BOP datasets with novel objects. After using our method as a post-processing, most state-of-the-art methods produce significantly better results, without any retraining or fine-tuning. The source code is available at https://scflow2.github.io.", + "bbox": [ + 88, + 282, + 483, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 601, + 220, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6D object pose estimation, which determines an object's 3D rotation and 3D translation relative to a camera, is crucial in fields like robotics, augmented reality, and spatial computing [31, 48].", + "bbox": [ + 89, + 626, + 482, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing pose estimation methods rely on a refinement procedure to get accurate pose results [10, 12, 20, 23, 24, 48, 50], and most of the refinement methods are based on render-and-compare strategies [10, 20, 23, 50]. A group of refinement methods [24, 48] achieves highly accurate results based on retraining or fine-tuning the network for novel objects, which, however, is cumbersome and unfriendly to real applications. Some recent refinement methods [23, 42, 50] have shown promising results without fine-tuning for novel objects. However, they still suffer from several weaknesses.", + "bbox": [ + 89, + 688, + 482, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, most methods formulate the render-and-compare framework as a general matching problem, but there is a strong prior of the target's 3D shape that is not used, mak", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d9e7353265abbc3cc39e2b8711d1dae8cdd8e26005018cb014ec4688486c7a8.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 517, + 250, + 629, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39e9f1a017519c747248789e2f543723571f1a48d63fcda0fff46d8f36974daa.jpg", + "image_caption": [ + "Network" + ], + "image_footnote": [], + "bbox": [ + 630, + 253, + 790, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc9a0acc4c70262982e7ed8cbbfe9f28059af475cb02e8082fa5569f12da9f32.jpg", + "image_caption": [ + "Refinement" + ], + "image_footnote": [], + "bbox": [ + 794, + 250, + 906, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e26ec74bbd33569cd6465a71ed55909bf8fb3c661efcfd670244a606f7e67ed6.jpg", + "image_caption": [ + "(a) Object pose refinement with retraining", + "Initialization", + "(b) Plug-and-play SCFlow2", + "Figure 1. Object pose refinement is critical for accurate object pose estimation. (a) Most existing object pose refinement methods, including SCFlow [11], rely on retraining for novel objects to achieve high accuracy. (b) The proposed SCFlow2 achieves even higher accuracy, and more importantly, generalizes well to novel objects without any retraining or fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 517, + 369, + 629, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96dfa597e0a6fed791cd94d81be963687b6d2c25b6700f88b0a048052197a73a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 404, + 787, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e2c47419cf7f86407f32c13c84a98b02450e63278a95d4905b30a135ec27a0d.jpg", + "image_caption": [ + "", + "Refinement" + ], + "image_footnote": [], + "bbox": [ + 794, + 369, + 906, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing it usually has an unnecessarily large search space for matching and is less effective for 6D object pose estimation [11]. Second, to address the matching issue, most methods attempt to refine the pose starting from multiple pose hypotheses based on the initialization [23, 32, 36]. While this paradigm is effective, it significantly slows down the refinement process. Third, most methods formulate the comparison after the rendering as a 2D matching procedure only, and then use ICP or Kabsch [21] to consume additional depth input as a second stage, which is less effective in capturing 3D motions and local optimal within each stage.", + "bbox": [ + 511, + 609, + 906, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation, as shown in Fig. 1. SCFlow2 is based on SCFlow [11] which is a state-of-the-art refinement method embedding 3D shape constraints into the optimization for known objects. We build on the basic shape-constraint structure of SCFlow but extend it in several novel ways.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9952a3a3f8f062e595a961342928d43ba8b64f83b476d0db771ae47febb7d09c.jpg", + "image_caption": [ + "(a) The framework of SCFlow" + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 519, + 207 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/89961d668d373cd6376f0275c610c970361917171d098a01020651d5bf062d97.jpg", + "image_caption": [ + "(b) The framework of SCFlow2", + "Figure 2. Design overview of SCFlow and SCFlow2. Given the object 3D mesh, we render an image $I_{1}$ and depth map $D_{1}$ based on an initial pose, and then use networks to compare these rendered outputs with the real input $I_{2}$ and $D_{2}$ to refine the pose. (a) Although SCFlow [11] adds 3D shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, which is less effective in capturing 3D motions. On the other hand, it cannot work with RGBD images. A common practice is to use RANSAC Kabsch [21] to consume additional depth as a second stage, which however is only local optimal within each stage. (b) SCFlow2 tackles these problems. We introduce an intermediate representation based on 3D scene flow to capture 3D motions in network optimization. Furthermore, we embed depth into the loop by formulating depth as an additional regularization to guide the correlation look-up iteratively, producing an end-to-end trainable system with RGBD images." + ], + "image_footnote": [], + "bbox": [ + 535, + 92, + 910, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main design of SCFlow2 is introducing a scene flow based representation into the iterative optimization. Although SCFlow embeds shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, and suffers in the solution with a second stage of RANSAC Kabsch [21] for RGBD images. On the other hand, SCFlow cannot work with novel objects. To handle these problems, we introduce rigid-motion embeddings in 3D scene flow [47] as our intermediate representation, and formulate depth as an additional regularization to guide the network optimization iteratively, which helps reducing search space and results in an end-to-end system that generalizes to novel objects, as illustrated in Fig. 2.", + "bbox": [ + 88, + 371, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train SCFlow2 on a combination of dataset Objverse [6], GSO [8] and ShapeNet [2], and evaluate on BOP datasets [15] with novel objects. SCFlow2 achieves state-of-the-art accuracy, and can produce accurate results with only one pose hypothesis, outperforming most multiple-hypothesis based refinement methods [23, 32, 36] in both accuracy and efficiency, without any retraining.", + "bbox": [ + 89, + 569, + 482, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 691, + 232, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object pose estimation aims to predict the 3D rotation and 3D translation with respect to the camera, and has shown prominent progress with deep learning techniques [13, 27, 28, 33, 37, 45]. Most existing methods assume the existence of many real images of the target, and train a learning model based on those real images [11, 19, 20, 38, 44], based on either 2D-to-3D matching [16, 26, 53], 3D-to-2D matching [19, 38, 40], or end-to-end training with differentiable perspective-n-point (PnP) [3, 29, 43]. This type of method is accurate on the trained target in general, which however, cannot work with novel objects that were not involved in the training. Retraining or fine-tuning can adapt the net", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work to novel objects, which, however, is cumbersome and limits its applications in real scenarios. Some recent methods show promising results for novel object pose estimation [23, 32, 34, 36, 50], and one of their key components is object pose refinement, which is based on render-and compare [23, 32, 50] or feature matching [34, 36] strategies. Although these refinement methods work in general, they are less effective at utilizing prior shape information of the target. The recent method SCFlow [11] introduces a shape-constraint recurrent matching framework for object pose estimation, which embeds the target's 3D shape into the optimization. It produces accurate results for known objects, while has limited generalization ability to novel objects. SCFlow2 is built on top of the basic shape-constraint structure of SCFlow, but extends it by introducing 3D scene flow representation into the network optimization, producing an end-to-end system with RGBD images.", + "bbox": [ + 511, + 371, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical flow estimation is a fundamental computer vision task to estimate pixel-level correspondences between two frames [5, 17, 41, 52]. Most recent methods are based on deep learning models and achieve accurate results in general scenes [5, 41, 46]. The recent method RAFT [46] presents a recurrent framework for flow estimation, and RAFT-3D [47] extends it to estimate 3D scene flow with rigid motion embeddings. Although they show significant advantages over existing flow methods, estimating dense matching with these general flow methods for object pose estimation is less effective, without using the prior 3D shape information of the target. Furthermore, the general flow loss is a surrogate matching loss that does not directly reflect the final 6D object pose [11, 18, 20, 43] in the context of object pose estimation. We combine the shape-constraint formulation [11] and the rigid motion embeddings in 3D scene flow, producing a geometry-guided 6D object pose refiner that is end-to-end trainable and generalizes to novel objects.", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22046", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af008dda54be056434a6663231d3a00e3944b3a311335ebbb42f830665f96048.jpg", + "image_caption": [ + "Figure 3. Overview of SCFlow2. Given an RGBD image and an initial pose of the target, we first render the target to obtain a synthetic RGBD image as the reference, and use an RGB encoder and depth encoder to extract features from the image pair, which will be used to create a 4D correlation volume. Based on the correlation volume and GRU, We use an intermediate flow regressor to predict 3D scene flow that is represented as a dense 3D transformation field $\\mathbf{T}_k^{\\prime}$ . We then use a pose regressor to predict a global pose update $\\Delta \\mathbf{P}_k$ based on implicit voting from the pixel-wise 3D transformation field. Finally, the updated pose $\\mathbf{P}_k$ is used to compute the pose-induced flow $F_{k}$ based on the target mesh to index the correlation volume for the next iteration. Note how the depth and 3D shape of the target are embedded into the framework to guide the optimization iteratively." + ], + "image_footnote": [], + "bbox": [ + 91, + 92, + 911, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 416, + 181, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an RGBD image and the 3D mesh of the object, our goal is to estimate the 6D pose of the target object. We assume the camera intrinsic matrix is known. We first introduce the framework at a high level in Section 3.1, and then present the design of the main components in Sections 3.2 and 3.3. We finally discuss the implementation details in Section 3.5.", + "bbox": [ + 89, + 444, + 483, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 568, + 200, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, we first render an RGBD image based on an initial pose of the target as a reference, and then predict the pose update based on comparing the rendered image and the real input with a recurrent matching network.", + "bbox": [ + 89, + 592, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the RGBD image pair, we first use an RGB encoder and depth encoder to extract features, and then construct a 4D correlation volume based on computing the dot product of the fused RGBD features. Based on the constructed correlation volume, we propose to use an intermediate flow regressor to predict 3D scene flow that is represented as a dense SE3 motion field [47]. We then use a pose regressor to predict a global pose update based on implicit voting from the pixel-wise SE3 motion field. Finally, the updated pose is used to compute a pose-induced flow based on the target mesh to index the correlation volume for the next iteration.", + "bbox": [ + 89, + 656, + 482, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key of our method is the combination of the dense 3D scene flow representation and the embedding of the target 3D shape prior [11] in a unified recurrent matching framework. Fig. 3 shows the overview of this framework.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Correlation Volume Construction", + "text_level": 1, + "bbox": [ + 511, + 417, + 805, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We construct a 4D volume pyramid to represent the correlation between any two locations on the two input images. We use an RGB encoder and a depth encoder to extract features from the RGB and the depth image, respectively. For the rendered image $I_{1}$ based on the initial pose and the input RGB image $I_{2}$ , we use an encoder with shared parameters to obtain their dense feature maps $f_{1}^{i} \\in \\mathbb{R}^{H \\times W \\times C}$ and $f_{2}^{i} \\in \\mathbb{R}^{H \\times W \\times C}$ at $1/8$ of the original resolution. We use DINOv2 ViT-B [35] as our RGB feature extractor in the encoder, and freeze the pretrained weights during training. For the depths $D_{1}$ and $D_{2}$ , we first lift the two depth maps to point clouds according to the camera intrinsic matrix, and then extract their 3D structural features with PointNet++ [39].", + "bbox": [ + 511, + 439, + 906, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to DenseFusion [48], we fuse the RGB features and depth features to generate semantic and contextual features. We then compute the 4D correlation volume via computing the dot product of the fused features according to different location pairs.", + "bbox": [ + 511, + 652, + 906, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Regressing 3D Transformation Field", + "text_level": 1, + "bbox": [ + 511, + 739, + 828, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With a dense matching field as the index, we can extract a 3D correlation map from the 4D correlation volume.", + "bbox": [ + 511, + 762, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In SCFlow [11], the correlation map is used to predict the optical flow $f(u,v)$ , which maps the pixel position $\\mathbf{x}_1 = (u,v)$ in $I_1$ to the corresponding position in $I_2$ : $\\mathbf{x}_2 = (u,v) + f(u,v)$ . However, this 2D flow formulation only contains 2D information, which is less effective in capturing 3D motions.", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, we propose to predict 3D scene flow [47]", + "bbox": [ + 532, + 885, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22047", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7caf4371b06dca9632e0d2c34e391d8d2f735912e0e1466c7345a1fc4d954e3.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 210, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4eb94e7619a549af88556663cf2cedcae0c20320fcb868d6775351fb556cd580.jpg", + "image_caption": [ + "Initial pose" + ], + "image_footnote": [], + "bbox": [ + 220, + 97, + 320, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a85df86751bd2e885784022a1c176fdf79ad41166f2492361758b2385f5a697.jpg", + "image_caption": [ + "T" + ], + "image_footnote": [], + "bbox": [ + 336, + 97, + 436, + 272 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ee6d55a9d7d9eab885f3b2ab3b4413a91b393670f436f4873ddbe616223c415.jpg", + "image_caption": [ + "$\\phi$", + "Figure 4. Visualization of pose-induced flow with 3D scene flow representation. Given a real input image and the rendered image based on an initial pose, we predict 3D scene flow represented as a dense SE3 motion field. We represent the motion field as a twist field $(\\tau, \\theta)$ . In theory, the motion fields should be constant pixels for rigid objects. We use a pose regressor to predict a global object-level pose based on the noisy motion fields. We then use the updated global pose to generate a pose-induced flow based on the target 3D mesh. The pose-induced flow embeds the shape prior of the target and reduces the search space for matching." + ], + "image_footnote": [], + "bbox": [ + 454, + 97, + 552, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67fc95840ac40972906fc606e7a5ff0308f9633baea68d51a0f9f72aaa418c13.jpg", + "image_caption": [ + "Intermediate flow" + ], + "image_footnote": [], + "bbox": [ + 570, + 97, + 669, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7df50d2dc89c1441b0f5f0b8ba3b23a982951c004e0534a1eecc122df6844d6c.jpg", + "image_caption": [ + "Pose-induced flow" + ], + "image_footnote": [], + "bbox": [ + 687, + 97, + 784, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27b035c8982fe8e4555ea73464bc517e9cf36ab8bbf3f047db07848364c86a53.jpg", + "image_caption": [ + "Refined pose" + ], + "image_footnote": [], + "bbox": [ + 792, + 88, + 910, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as an intermediate representation for the network. Given $I_{1}$ and $I_{2}$ and their corresponding point clouds $C_1$ and $C_2$ , we formulate scene flow as a dense 3D transformation field $\\mathbf{T} \\in SE(3)^{H \\times W}$ to represent the 3D motion between the two point clouds in 3D space.", + "bbox": [ + 88, + 397, + 482, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for the point $X_{i}$ in $C_1$ with index $i$ and its projected pixel coordinates $\\mathbf{x}_i = (u_i,v_i,z_i)$ , we can perform the transformation $\\mathbf{T}_i\\cdot X_i$ to obtain the corresponding point $X_{i}^{\\prime}$ in $C_2$ and its projected pixel coordinates $\\mathbf{x}_i' = (u_i',v_i',z_i')$ . The scene flow vector is defined as the difference $\\mathbf{f} = \\mathbf{x}_i' - \\mathbf{x}_i$ , where the first two components represent the standard 2D optical flow and the last component describes the depth difference between the two points.", + "bbox": [ + 88, + 472, + 482, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use GRU [4] as the recurrent model for the intermediate scene flow predictor. In the $k^{th}$ iteration, the hidden state update can be expressed as:", + "bbox": [ + 88, + 593, + 482, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\operatorname {G R U} \\left(\\mathbf {L} _ {\\mathbf {C}} \\left(F _ {k - 1}\\right), \\mathbf {T} _ {k - 1}, h _ {k - 1}; \\Theta\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 648, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{L}_{\\mathbf{C}}(F_{k - 1})$ represents the correlation look-up operation guided by the global flow $F_{k - 1}$ from the $(k - 1)^{th}$ iteration, $\\mathbf{T}_{k - 1}$ is the dense 3D transformation field from the previous iteration, $\\Theta$ denotes the GRU network parameters, and $h_k$ is the hidden state feature which will be fed into a dense SE3 layer [47] to update the SE3 motion fields and generate a new transformation field $\\mathbf{T}_k^\\prime$", + "bbox": [ + 88, + 672, + 482, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Regressing Global Object Pose", + "text_level": 1, + "bbox": [ + 89, + 787, + 362, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to predict the pose residual of the single object in $I_{1}$ and $I_{2}$ . Considering the rigidity of objects in 6D pose estimation scenarios, all pixel-level 3D motions in $\\mathbf{T}_k^\\prime$ theoretically describe the same rigid object and should be the same. However, noise is inevitable in predicting the dense transformation field. Therefore, we designed a small", + "bbox": [ + 88, + 809, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "network to predict the global pose residual using the dense transformation field as input, as shown in Fig. 4. First, we represent the dense transformation field $\\mathbf{T}_k^{\\prime}$ as a $4\\times 4$ transformation matrix, and then we encode it using a three-layer 2D convolutional network. Finally, we use two fully connected layers to output the global pose residual $\\Delta \\mathbf{P}_k$ , which is represented as a 9-dim vector including 6-min for the rotation [54] and 3-dim for the normalized translation [25].", + "bbox": [ + 511, + 397, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the $k^{th}$ iteration, the pose residual $\\Delta \\mathbf{P}_k$ will be used for the pose update: $\\mathbf{P}_k = \\mathbf{P}_{k - 1} \\otimes \\Delta \\mathbf{P}_k$ , and the updated pose $\\mathbf{P}_k$ will be used to compute the pose-induced flow based on the object mesh, which imposes shape constraints for the lookup operation in the next iteration [11]. At the same time, we will generate a new dense transformation field $\\mathbf{T}_k$ based on the updated pose $\\mathbf{P}_k$ . Specifically, we calculate the pose residual between the initial pose $\\mathbf{P}_0 = [\\mathbf{R}_0|\\mathbf{t}_0]$ and $\\mathbf{P}_k = [\\mathbf{R}_k|\\mathbf{t}_k]$ as follows:", + "bbox": [ + 511, + 518, + 908, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {R} _ {k} = \\mathbf {R} _ {k} \\cdot \\mathbf {R} _ {0}, \\quad \\Delta \\mathbf {t} _ {k} = \\mathbf {t} _ {k} - \\Delta \\mathbf {R} _ {k} \\cdot \\mathbf {t} _ {0} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 672, + 903, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we update the 3D motion field $\\mathbf{T}_k\\in SE(3)^{H\\times W}$ , where the motion at every location will be updated as the same new pose residual. This new motion field will be the input for the next iteration.", + "bbox": [ + 511, + 704, + 906, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 782, + 730, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the input RGB and depth images $I_{1}$ and $D_{2}$ , we first crop the target out based on the initial pose, and then resize it to a fixed resolution of $256 \\times 256$ .", + "bbox": [ + 511, + 806, + 906, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the ground truth pose, we can compute the ground truth optical flow based on the object mesh. We use the exponentially weighted strategy [11, 46] to calculate the loss", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22048", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/afb3116f36e3cf0be29d9d7c52dd1d2a920952f02a650aaec17dd54fb33653c7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLM-OT-LESSTUD-LIC-BINITODDHBYCB-VAvg.
MegaPose [23]62.048.584.646.246.072.576.462.3
MegaPose + Ours62.454.988.348.449.576.479.065.6
FoundPose [36]61.057.069.447.940.772.369.059.6
FoundPose + Ours69.658.688.154.648.476.284.968.6
GenFlow [32]63.552.186.253.455.477.983.367.4
GenFlow + Ours68.753.788.055.456.479.285.569.6
Gigapose [34]67.855.681.156.357.579.182.568.6
Gigapose + Ours71.058.682.057.758.380.184.470.3
SAM6D [27]69.951.590.458.860.277.684.570.4
SAM6D + Ours72.152.390.860.060.577.785.171.2
FoundationPose [50]75.664.692.350.858.083.588.973.4
FoundationPose + Ours80.165.095.551.758.384.491.475.2
", + "bbox": [ + 176, + 88, + 826, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison with the state of the art. All the baselines have their own refinement strategy, and we report their final result published on BOP leaderboard. After using our method as a post-processing (“+ Ours”), the result consistently improves across all baselines on all datasets.", + "bbox": [ + 89, + 328, + 906, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in each iteration for both flow loss and pose loss:", + "bbox": [ + 89, + 397, + 413, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {k = 1} ^ {N} \\gamma^ {N - k} \\left(\\mathcal {L} _ {\\text {p o s e}} ^ {k} + \\alpha \\mathcal {L} _ {\\text {f l o w}} ^ {k}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 420, + 482, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the weight $\\gamma = 0.8$ , $N = 8$ and $\\alpha = 0.1$ . $\\mathcal{L}_{flow}$ represents the L1 loss between the first two dimensions of the predicted scene flow and the ground truth optical flow. $\\mathcal{L}_{pose}$ refers to pose loss, which is computed as the L1 distance between the 3D point clouds transformed by the predicted pose and the ground truth pose, respectively [11, 49].", + "bbox": [ + 89, + 469, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 573, + 223, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SCFlow2 on seven challenging BOP benchmark datasets in this section. Section 4.1 presents our experimental setup, including the datasets used for training and evaluation, training details and the evaluation metrics. In Section 4.2, we compare our approach with current state-of-the-art pose methods. Section 4.3 provides a detailed ablation study of the proposed framework.", + "bbox": [ + 89, + 598, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 712, + 297, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We train SCFlow2 on a combination of datasets including ShapeNet-Objects [2], Google-Scanned-Objects [8], and Objverse [6]. All these datasets only contain object meshes, in a total of about 90K objects. We reuse the synthetic images from [23] and [50] which are rendered based on those meshes, with about 3M images in total.", + "bbox": [ + 89, + 734, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate on seven datasets from the BOP benchmark [15], including LM-O [1], T-LESS [14], TUD-L [15], IC-BIN [7], ITODD [9], HB [22], and YCB-V [51]. These datasets include 108 objects having different sizes and different texture and symmetry properties. Most of them are", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "captured in challenging setups, including multiple object instances, significant clutter, and severed occlusions. These objects are not included in the training dataset and are novel to the models.", + "bbox": [ + 511, + 397, + 906, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details. We train SCFlow2 using the AdamW optimizer [30] with a batch size of 16, and use a cosine annealing learning rate scheduler. The model is trained for 200k iterations with an initial learning rate of 1e-4. During training, we generate pseudo initial poses by adding random noise to the ground-truth pose. We add rotation jitter sampled from a normal distribution with a standard deviation of 15 degrees in each 3D axis, and add translation jitter sampled with deviations of 15, 15, and 50 millimeters along xyz axes, respectively.", + "bbox": [ + 511, + 460, + 908, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics We report the result in the common BOP metric, which is the mean Average Recall (AR) of three different metrics, including Visible Surface Discrepancy (VSD), Maximum Symmetry-aware Surface Distance (MSSD), and Maximum Symmetry-aware Projection Distance (MSPD). Readers can refer to [15] for further details about these metrics.", + "bbox": [ + 511, + 614, + 908, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with the State of the Art", + "text_level": 1, + "bbox": [ + 511, + 739, + 834, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To explore the plug-and-play nature of our refinement method, we apply it to the final results of Megapose [23], FoundPose [36], GenFlow [32], GigaPose [34], SAM6D [27], and FoundationPose [50], as shown in Table 1. Although these baselines have a refinement strategy themselves, plugging our refinement method into the end of their framework further improves the results, consistently on all datasets. Most baseline results are downloaded from the BOP leaderboards.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22049", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3b417ab83ff58bf96a7884c19198e8da908705bc3089496f5c2ae8a47f0d871f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDepthLM-OYCB-V
FoundPose (init)-39.645.2
FoundPose (final)✓61.069.0
FoundPose (init) + Ours✓61.381.8
FoundPose (final) + Ours✓69.684.9
GenFlow (init)-25.027.7
GenFlow (final)✓63.583.3
GenFlow (init) + Ours✓64.483.6
GenFlow (final) + Ours✓68.785.5
FoundationPose (init)---
FoundationPose (final)✓75.688.9
FoundationPose (init) + Ours✓79.790.4
FoundationPose (final) + Ours✓80.191.4
", + "bbox": [ + 96, + 88, + 478, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc967123e8bedd2cce676a2f6911855b11fad453a4dc00381ef1190a44faa51a.jpg", + "table_caption": [ + "Table 2. Effectiveness as a plug-and-play object pose refiner. All the baselines are based on an initialization-refinement architecture. Our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner." + ], + "table_footnote": [], + "table_body": "
L3L5L10L20L30L40L50
Initialization92.783.864.53517.67.83.5
+ SCFlow291.691.591.590.888.683.270.7
", + "bbox": [ + 94, + 393, + 486, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 559, + 243, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness as a plug-and-play object pose refiner. We evaluate SCFlow2 in more detail in the context of three different state-of-the-art methods, FoundPose [36], GenFlow [32], and FoundationPose [50]. Table 2 summaries the results. All of them are based on an initialization-refinement architecture. We denote their initialization results as \"init\" in the table, and replace their refinement methods with ours in each method (denoted as \"init + ours\"). Furthermore, we also report the results of applying our refinement to their final results directly, denoted as \"final + ours\". Note that, FoundationPose outputs hundreds of pose hypotheses in the initialization stage, and there is no well-defined initialization result for it. In this case, we replace its original refinement method with SCFlow2 in its framework, and keep all other components untouched, including post-processing procedures such as pose scoring and pose selection, which is the reported result of \"FoundationPose (init) + Ours\" in the table. Overall, our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner. We show some qualitative results in Fig. 5 and 6.", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/569fdb8791e2294aa9b2714e6b1483ae5b874328f590d09f9ce12db9412c3ea8.jpg", + "table_caption": [ + "Table 3. Impact of pose initialization. We add randomly generated rotation and translation noise to GT pose as the initialization, and then use SCFlow2 to refine the pose. L3 to L50 denote different noise levels. SCFlow2 is robust to pose errors in the initialization and works well under noise level 30." + ], + "table_footnote": [], + "table_body": "
DatasetsInit.SeenUnseen
V1V1+V1++V2
LM-O61.062.969.368.969.6
T-LESS57.057.954.955.358.6
TUD-L69.475.988.886.688.1
IC-BIN47.947.954.153.754.6
ITODD40.734.249.148.148.4
HB72.374.576.576.176.2
YCB-V69.080.785.584.484.9
Avg.59.662.068.367.668.6
", + "bbox": [ + 544, + 88, + 877, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Advantages over SCFlow. We use the same initialization based on results from FoundPose (\"Init.\", and report results of SCFlow (\"V1\") and SCFlow2 (\"V2\"). SCFlow2 outperforms all \"V1\" versions without any retraining.", + "bbox": [ + 511, + 272, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of pose initialization. We add random rotation and translation noise to GT pose as the initialization, and report the results on the YCB-V dataset, as shown in Table 3. L3 to L50 denote different noise levels (rotation error from 3 to 50 degrees, translation error from 3 to $50~\\mathrm{mm}$ ). SCFlow2 works well under noise level 30.", + "bbox": [ + 511, + 356, + 906, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation of performance with different iterations. We conduct ablation studies on the trade off between the accuracy/efficiency and the iteration numbers of our method, as shown in Fig. 7. Our method produces meaningful pose refinement results after only 2 iterations and reaches saturation after 6 iterations. In our experiments, we use 8 iterations as the default setting in both training and testing.", + "bbox": [ + 511, + 446, + 905, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Advantages over SCFlow. We compare SCFlow2 and its previous version SCFlow systematically. We summarize the results in Table 4. We use the same initialization based on results from FoundPose, and refine it with SCFlow (\"V1\") and SCFlow2 (\"V2\"), respectively. We evaluate three different versions of SCFlow, where \"V1\" denotes the original version based on retraining on target objects with RGB images only, \"V1+\" denotes the version using RANSAC Kabsch to consume additional depth images. Both \"V1\" and \"V1+\" need retraining on target objects. We also evaluate a version \"V1++\" which does not rely on retraining for novel objects, but uses the same large-scale datasets as \"V2\" for training on RGB images and then refine the pose results using RANSAC Kabsch with additional depths. We can see that, without retraining, \"V1++\" produces worse results compared with \"V1+\". On the other hand, SCFlow2 even outperforms \"V1+\" without any retraining.", + "bbox": [ + 511, + 553, + 905, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation with different settings. We use the initialization results of FoundPose as our initialization. Table 5 shows the result, where \"Shape\" denotes the shape-constraint regularization during the recurrent optimization, and \"Scene\" is the 3D scene flow representation. We use the intermediate flow for the look-up without relying on the", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22050", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b7028c30c3d5d8ea2d9947bc123ee74a1fb6705470fbfa30a7a15722c91a3b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 88, + 233, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e49da28d43c17269488796d27168477f7ab0a513a507a903e96fce5506ea62a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 88, + 364, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a7642d12614e45cdd0a9e4d2fa1128957ccecc1b2dc0390af2659ce920f921f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 496, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5ba2bb65a4bff0bab725fcfbe0cf79b020266aed3b78a6eccf15b8125f98b18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 630, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7a3a6acd7f7c6bef5825e99725791fa99ee534aab1c348564223d7acfc94faf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 88, + 763, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e07ccc5ffb53a64aff587442bbd5a33a56e2edd160831414accce8057d2c7289.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 88, + 895, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97ccebaf4a43b588bfa22824cf0f63ec473fd93b1120780fab4fa77acb2f26f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 193, + 233, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4ce9c2d457f94737b8fc86dbd41793765b2687358f650cd8b070e60c518ea43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 193, + 364, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/607345a095c7a497fac7616da5bdc4f6a26b46723b6604b170f5a2c74d0cbfe0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 193, + 496, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d504ef74dcc57f2dddf986ae5b9fecd43ee69686ae45149616d634e0d91b7afe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 193, + 630, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/635c2bfe4f7fb40a28a26fdfce49c935bf417060a02cdea8c65ee8b52aba139f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 193, + 763, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c109ab539f6cd06f3a4e17ee4fdde87c7ba2297bb0d411d91cb4b57f45c934e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 193, + 895, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e765ca510dd1325d8e9011166b6fdcee7bc93e1072ab0df618f9ea18ba88f1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 296, + 232, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23279438713f379fa872ef124492582d9f90991059264e4222971e8371386615.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 296, + 364, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c301fc7c2a99f6f672dc0f3fe1debabf8923498594fb57f3cb11615c06034b64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 296, + 496, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24e41fbaef3e5ff135ffdc33407da9cd500de10a59e62e94d19900372556106e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 296, + 630, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b8b379dfabd67ed2ef49c1e7b787dada8389ebfc7a4e9640803e9ee98dbb3b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 296, + 763, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8a0c7719260f15ff5c0b156dfdb93aae41bf0047ca0669a31aded29f67f7f9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 296, + 895, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/659dc210b7221b1cbef18fadfd358f8768628a70a581300c0ef799210a095f84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 401, + 232, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c67747d55f53117ffbf55f478c95bdd44e25fe514adefb81c98869e28f47441a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 401, + 364, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/957ac03eef067decf82350030ecd1eeffdc0dea662521e32a3fe934b42637d31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 401, + 496, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ff90ee668a8848f6081688f54389c72330cc939446c58e56c7b15cc75d649cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 401, + 630, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6cc5653ccee602f61fa0abe7d6c8c3a1aef917868c38a4eb82f17e7f3200208.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 401, + 763, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60e7eae561a438b81a8beddd50d785ee221b184baa52efe9f3b10e6293f5a852.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 401, + 895, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f81ae0963f8fc4d78a853b9925684f6269620215b688ac49f5c6ccbcfe3e49a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 506, + 232, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e0b37b7df406a192939e7b086fabe80b58154deda543c015ee527eb76369b6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 506, + 364, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/332da04141ccf188cc9f1f0786c2a5ef2461ab49e4d2453082a1b7d9129ab4f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 506, + 496, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0da0edac4c1e76502e54d9a04598b59cb1399b61ea00d645951c63c21d4e49d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 506, + 630, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f573fef8b980e886d4f8fd8976a8de8dd63d6d0d60ee46e2b5f475c49ae3bbdc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 506, + 763, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2f89dce3be900307788c91a7867ca9a33f3646541e66d1e72e3350a6a0e8bd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 506, + 895, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97e578f7e4fb968a9eeec9700270085e2cc5257cdf0bbb2eb3fe58e82846b58d.jpg", + "image_caption": [ + "GFlow (init)" + ], + "image_footnote": [], + "bbox": [ + 99, + 611, + 232, + 713 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/404d233ef64ae10c59d59b6b4493846aa9ce662bcd81fba7cab67e6f01137aad.jpg", + "image_caption": [ + "GFlow (final)" + ], + "image_footnote": [], + "bbox": [ + 233, + 611, + 364, + 713 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc2e05ae4a2afc6a00c5acb7f8292429feb91199946a844b0cba183fb409e51c.jpg", + "image_caption": [ + "GFlow (init) + Ours" + ], + "image_footnote": [], + "bbox": [ + 364, + 611, + 496, + 713 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78c14e4cb765558c16daf893eba69843bd4b34706b4fced397c1d857f3552075.jpg", + "image_caption": [ + "FPose (init)" + ], + "image_footnote": [], + "bbox": [ + 500, + 611, + 630, + 713 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bca1447470d91eb08343caab2eb956aade89024ae341920a1bef03d3cd941fa.jpg", + "image_caption": [ + "FPose (final)", + "Figure 5. Qualitative results. Given the same pose initialization as that in GenFlow (\"GFlow\") and FoundPose (\"FPose\"), denoted as \"GFlow (init)\" and \"FPose (init)\" respectively, our refinement method (\"+ Ours\") produces considerably more accurate results compared to the refinement approaches in their original methods (note how our reprojected 3D mesh aligns better with the object contours)." + ], + "image_footnote": [], + "bbox": [ + 630, + 611, + 763, + 713 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c983e00021a65622b43e1a7c60b429d5150285c4b212e326f6fda2bcc1c8ac1.jpg", + "image_caption": [ + "FPose (init) + Ours" + ], + "image_footnote": [], + "bbox": [ + 763, + 611, + 895, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "target's 3D mesh when \"Shape\" is disabled, and we replace the scene flow representation with a plain MLP pose regressor when \"Scene\" is disabled. The performance becomes worse without the shape prior embedding, and deteriorates significantly without the 3D scene flow formulation.", + "bbox": [ + 89, + 811, + 485, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Timing analysis. The training of SCFlow2 needs about 108 hours on a workstation with 2 NVIDIA RTX-3090 GPU. SCFlow2 considers the pose initialization as the only hypothesis. During inference, SCFlow2 takes about 0.18 seconds to refine a pose on an RTX-3090 GPU with a typical", + "bbox": [ + 511, + 811, + 908, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c95330b737894cbb359c6010e4603e0a782e5fb10bccb5596732c1c155864bb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 287, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e740f9664a51697a2cde48f58569336b541ea139ef67a49828b420c294cc6902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 475, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55d3517f07d5b89afb8d24213ce703691f61a6ca04537979f73a80a45aa399f7.jpg", + "image_caption": [ + "Initialization" + ], + "image_footnote": [], + "bbox": [ + 98, + 200, + 285, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b04ffe006eaeee30faac55b6f08f36daaf59efd8be21a8a82765a38b4b5c753d.jpg", + "image_caption": [ + "+Ours", + "Figure 6. More qualitative results. We show two examples on YCB-V. Although the initialization has large pose error, our refinement method produces accurate results that align well with the target, without retraining on these novel objects." + ], + "image_footnote": [], + "bbox": [ + 287, + 200, + 475, + 311 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd0e9f9c9d3fea97eec4fc78dfd707611801eac618a402b54f8f368a3abbe1d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ShapeSceneLM-OYCB-V
✓-52.667.6
-✓60.080.4
✓✓61.381.8
", + "bbox": [ + 161, + 412, + 415, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Evaluation with different settings. \"Shape\" denotes the shape-constraint regularization during the optimization, and \"Scene\" is the 3D scene flow formulation. The performance deteriorates significantly without the 3D scene flow formulation.", + "bbox": [ + 89, + 500, + 482, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image from BOP dataset. Taken for comparison, the multi-hypothesis based method MegaPose needs about 1.85 seconds for a refinement on the same workstation, which is $10+$ times slower than SCFlow2.", + "bbox": [ + 89, + 584, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation discussion. In general, our refinement method can improve the accuracy of initial pose significantly. While occasionally, it produces worse results than the initial pose. Fig. 8 shows the histogram of samples along with the relative accuracy improvement after pose refinement on YCB-V and LM-O with the initial pose of FoundPose as our initialization. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases where the initial poses are already very wrong caused by severe occlusion or extreme light conditions. We will take this as one of our future work.", + "bbox": [ + 89, + 646, + 482, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 844, + 209, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. We extended", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/053909c232e707e09b3a4a3fa68e70f5b8d018233ae3daba93ef087ff50defcc.jpg", + "image_caption": [ + "Figure 7. Evaluation of performance with different iterations. Given the same pose initialization with noise level 30 as in Table 3, more recurrent iterations produce more accurate results while spending more time in inference. Nevertheless, our approach yields meaningful outcomes after just 2 iterations, and the performance reaches saturation after 6 iterations." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 705, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9855c0de2da3571d78a875667dd53ce61edf8966f336df35e5f8afa7713a4ee9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 90, + 901, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/051e7db46c7cbed92218312665a01a5baaab189b5ebb60cb75be1bcd2ff23c04.jpg", + "image_caption": [ + "Figure 8. Limitation discussion. We show the histogram of samples along with the relative accuracy improvement after pose refinement. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases." + ], + "image_footnote": [], + "bbox": [ + 617, + 345, + 805, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the RGB-based method SCFlow by formulating additional depth as a regularization in the iteration with an intermediate 3D scene flow representation for RGBD frames, and combined it with the 3D shape prior information of the target, producing a geometry-guided 6D object pose refiner end-to-end trainable and generalizable to novel objects. Experiments on BOP datasets with novel objects show that our method achieves state-of-the-art accuracy, and is one of the most efficient refinement methods.", + "bbox": [ + 511, + 601, + 905, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the Youth Innovation Team of Shaanxi Universities, the \"Scientist + Engineer\" team of the Qin Chuang Yuan in Shaanxi Province, the Xi'an Science and Technology Program, the \"Leading the Charge\" initiative for the industrialization of core technologies in key industrial chains in Shaanxi Province, the National Nature Science Foundation of China under Grant 62371359, the Key Research and Development Program of Shaanxi, the Key Research Program of the Chinese Academy of Sciences.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22052", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D Object Pose Estimation Using 3D Object Coordinates. In European Conference on Computer Vision, 2014. 5", + "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5", + "[3] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-End Learnable Geometric Vision by Backpropagating PnP Optimization. In Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[4] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv preprint arXiv:1409.1259, 2014. 4", + "[5] Javier Cuadrado, Ulysse Rancon, Benoit R Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical Flow Estimation from Event-based Cameras and Spiking Neural Networks. Frontiers in Neuroscience, 17, 2023. 2", + "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A Universe of Annotated 3D Objects. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 5", + "[7] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6D Object Pose and Predicting Next-Best-View in the Crowd. In Conference on Computer Vision and Pattern Recognition, 2016. 5", + "[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items. In International Conference on Robotics and Automation, 2022. 2, 5", + "[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing MVTec ITODDA Dataset for 3D Object Recognition in Industry. In International Conference on Computer Vision Workshops, 2017. 5", + "[10] Yang Hai, Rui Song, Jiaojiao Li, David Ferstl, and Yinlin Hu. Pseudo Flow Consistency for Self-Supervised 6D Object Pose Estimation. In International Conference on Computer Vision, 2023. 1", + "[11] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-Constraint Flow for 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5", + "[12] Rasmus Laurvig Haugaard and Anders Glent Buch. SurfEmb: Dense and Continuous Correspondence Distributions for Object Pose Estimation with Learnt Surface Embeddings. In Conference on Computer Vision and Pattern Recognition, 2022. 1", + "[13] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. FS6D: Few-Shot 6D Pose Estimation of Novel Ob" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "jects. In Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[14] Tomás Hodań, Pavel Haluza, Štepan Obdržánek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-Less Objects. In Winter Conference on Applications of Computer Vision, 2017. 5", + "[15] Tomáš Hodař, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In European Conference on Computer Vision, 2018. 2, 5", + "[16] Tomás Hodań, Daniel Barath, and Jérí Matas. EPOS: Estimating 6D Pose of Objects with Symmetries. In Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[17] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2", + "[18] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-Stage 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[19] Yinlin Hu, Sebastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-Depth-Range 6D Object Pose Estimation in Space. In Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[20] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective Flow Aggregation for Data-Limited 6D Object Pose Estimation. In European Conference on Computer Vision, 2022. 1, 2", + "[21] Wolfgang Kabsch. A Solution for the Best Rotation to Relate Two Sets of Vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 1, 2", + "[22] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D Dataset for 6D Pose Estimation of 3D Objects. In International Conference on Computer Vision Workshops, 2019. 5", + "[23] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In Conference on Robot Learning. 1, 2, 5", + "[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent Multi-View Multi-Object 6d Pose Estimation. In European Conference on Computer Vision, 2020. 1", + "[25] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep Iterative Matching for 6D Pose Estimation. In European Conference on Computer Vision, 2018. 4", + "[26] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In International Conference on Computer Vision, 2019. 2", + "[27] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yongliang Lin, Yongzhi Su, Praveen Nathan, Sandeep Inuganti, Yan Di, Martin Sundermeyer, Fabian Manhardt, Didier Stricker, Jason Rambach, and Yu Zhang. HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[29] Fulin Liu, Yinlin Hu, and Mathieu Salzmann. Linear-Covariance Loss for End-to-End Learning of 6D Pose Estimation. In International Conference on Computer Vision, 2023. 2", + "[30] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In International Conference on Learning Representations, 2019. 5", + "[31] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose Estimation for Augmented Reality: A Hands-On Survey. IEEE transactions on visualization and computer graphics, 22(12):2633-2651, 2015. 1", + "[32] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6", + "[33] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. NOPE: Novel Object Pose Estimation from a Single Image. In Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[34] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5", + "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOV2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 3", + "[36] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found-Pose: Unseen Object Pose Estimation with Foundation Features. European Conference on Computer Vision, 2024. 1, 2, 5, 6", + "[37] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. LatentFusion: End-to-End Differentiable Reconstruction and Rendering for Unseen Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[38] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-Wise Voting Network for 6DoF Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. arXiv preprint arXiv:1706.02413, 2017. 3", + "[40] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects without Using Depth. In International Conference on Computer Vision, 2017. 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[42] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A Multi-Stage One Shot Object Pose Estimation Framework. In Conference on Computer Vision and Pattern Recognition, 2022. 1", + "[43] Chen Song, Jiaru Song, and Qixing Huang. HybridPose: 6D Object Pose Estimation under Hybrid Representations. In Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[44] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation. arXiv preprint arXiv:2203.09418, 2022. 2", + "[45] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-Shot Object Pose Estimation without CAD Models. In Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[46] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In European Conference on Computer Vision, 2020. 2, 4", + "[47] Zachary Teed and Jia Deng. RAFT-3D: Scene Flow using Rigid-Motion Embeddings. In Conference on Computer Vision and Pattern Recognition, 2021. 2, 3, 4", + "[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In Conference on Computer Vision and Pattern Recognition, 2019. 1, 3", + "[49] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 5", + "[50] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6", + "[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes. In Robotics: Science and Systems, 2018. 5", + "[52] Zhaolin Xiao, Zhan Su, Fengyuan Zuo, and Haiyan Jin. Low-Light Optical Flow Estimation with Hidden Feature Supervision Using a Siamese Network. Journal of Image and Graphics, 29:0231-0242, 2024. 2", + "[53] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In International Conference on Computer Vision, 2019. 2", + "[54] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Conference on Computer Vision and Pattern Recognition, 2019. 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22054", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_model.json b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..802ee992cb5bbff8e2bfbea73d0666e054b31553 --- /dev/null +++ b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_model.json @@ -0,0 +1,2706 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.904, + 0.153 + ], + "angle": 0, + "content": "SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.18, + 0.892, + 0.22 + ], + "angle": 0, + "content": "Qingyuan Wang \\(^{1}\\), Rui Song \\(^{1}\\), Jiaojiao Li \\(^{1}\\), Kerui Cheng \\(^{2}\\), David Ferstl \\(^{3}\\), Yinlin Hu \\(^{3}\\) \n\\(^{1}\\) State Key Laboratory of ISN, Xidian University \\(^{2}\\) Taiyuan University of Technology \\(^{3}\\) MagicLeap" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.284, + 0.485, + 0.571 + ], + "angle": 0, + "content": "We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. Most recent 6D object pose methods rely on refinement to get accurate results. However, most existing refinement methods either suffer from noises in establishing correspondences, or rely on retraining for novel objects. SCFlow2 is based on the SCFlow model designed for refinement with shape constraint, but formulates the additional depth as a regularization in the iteration via 3D scene flow for RGBD frames. The key design of SCFlow2 is an introduction of geometry constraints into the training of recurrent matching network, by combining the rigid-motion embeddings in 3D scene flow and 3D shape prior of the target. We train SCFlow2 on a combination of dataset Objaverse, GSO and ShapeNet, and evaluate on BOP datasets with novel objects. After using our method as a post-processing, most state-of-the-art methods produce significantly better results, without any retraining or fine-tuning. The source code is available at https://scflow2.github.io." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.221, + 0.617 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.688 + ], + "angle": 0, + "content": "6D object pose estimation, which determines an object's 3D rotation and 3D translation relative to a camera, is crucial in fields like robotics, augmented reality, and spatial computing [31, 48]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Most existing pose estimation methods rely on a refinement procedure to get accurate pose results [10, 12, 20, 23, 24, 48, 50], and most of the refinement methods are based on render-and-compare strategies [10, 20, 23, 50]. A group of refinement methods [24, 48] achieves highly accurate results based on retraining or fine-tuning the network for novel objects, which, however, is cumbersome and unfriendly to real applications. Some recent refinement methods [23, 42, 50] have shown promising results without fine-tuning for novel objects. However, they still suffer from several weaknesses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "First, most methods formulate the render-and-compare framework as a general matching problem, but there is a strong prior of the target's 3D shape that is not used, mak" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.251, + 0.63, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.338, + 0.613, + 0.35 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.255, + 0.791, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.337, + 0.739, + 0.346 + ], + "angle": 0, + "content": "Network" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.25, + 0.907, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.338, + 0.887, + 0.35 + ], + "angle": 0, + "content": "Refinement" + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.357, + 0.836, + 0.371 + ], + "angle": 0, + "content": "(a) Object pose refinement with retraining" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.371, + 0.63, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.459, + 0.612, + 0.47 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.405, + 0.789, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.458, + 0.795, + 0.469 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.371, + 0.907, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.458, + 0.886, + 0.469 + ], + "angle": 0, + "content": "Refinement" + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.477, + 0.796, + 0.491 + ], + "angle": 0, + "content": "(b) Plug-and-play SCFlow2" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Figure 1. Object pose refinement is critical for accurate object pose estimation. (a) Most existing object pose refinement methods, including SCFlow [11], rely on retraining for novel objects to achieve high accuracy. (b) The proposed SCFlow2 achieves even higher accuracy, and more importantly, generalizes well to novel objects without any retraining or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.791 + ], + "angle": 0, + "content": "ing it usually has an unnecessarily large search space for matching and is less effective for 6D object pose estimation [11]. Second, to address the matching issue, most methods attempt to refine the pose starting from multiple pose hypotheses based on the initialization [23, 32, 36]. While this paradigm is effective, it significantly slows down the refinement process. Third, most methods formulate the comparison after the rendering as a 2D matching procedure only, and then use ICP or Kabsch [21] to consume additional depth input as a second stage, which is less effective in capturing 3D motions and local optimal within each stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation, as shown in Fig. 1. SCFlow2 is based on SCFlow [11] which is a state-of-the-art refinement method embedding 3D shape constraints into the optimization for known objects. We build on the basic shape-constraint structure of SCFlow but extend it in several novel ways." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22045" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.52, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.215, + 0.407, + 0.227 + ], + "angle": 0, + "content": "(a) The framework of SCFlow" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.093, + 0.911, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.215, + 0.832, + 0.227 + ], + "angle": 0, + "content": "(b) The framework of SCFlow2" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Figure 2. Design overview of SCFlow and SCFlow2. Given the object 3D mesh, we render an image \\( I_{1} \\) and depth map \\( D_{1} \\) based on an initial pose, and then use networks to compare these rendered outputs with the real input \\( I_{2} \\) and \\( D_{2} \\) to refine the pose. (a) Although SCFlow [11] adds 3D shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, which is less effective in capturing 3D motions. On the other hand, it cannot work with RGBD images. A common practice is to use RANSAC Kabsch [21] to consume additional depth as a second stage, which however is only local optimal within each stage. (b) SCFlow2 tackles these problems. We introduce an intermediate representation based on 3D scene flow to capture 3D motions in network optimization. Furthermore, we embed depth into the loop by formulating depth as an additional regularization to guide the correlation look-up iteratively, producing an end-to-end trainable system with RGBD images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Our main design of SCFlow2 is introducing a scene flow based representation into the iterative optimization. Although SCFlow embeds shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, and suffers in the solution with a second stage of RANSAC Kabsch [21] for RGBD images. On the other hand, SCFlow cannot work with novel objects. To handle these problems, we introduce rigid-motion embeddings in 3D scene flow [47] as our intermediate representation, and formulate depth as an additional regularization to guide the network optimization iteratively, which helps reducing search space and results in an end-to-end system that generalizes to novel objects, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.676 + ], + "angle": 0, + "content": "We train SCFlow2 on a combination of dataset Objverse [6], GSO [8] and ShapeNet [2], and evaluate on BOP datasets [15] with novel objects. SCFlow2 achieves state-of-the-art accuracy, and can produce accurate results with only one pose hypothesis, outperforming most multiple-hypothesis based refinement methods [23, 32, 36] in both accuracy and efficiency, without any retraining." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.233, + 0.708 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Object pose estimation aims to predict the 3D rotation and 3D translation with respect to the camera, and has shown prominent progress with deep learning techniques [13, 27, 28, 33, 37, 45]. Most existing methods assume the existence of many real images of the target, and train a learning model based on those real images [11, 19, 20, 38, 44], based on either 2D-to-3D matching [16, 26, 53], 3D-to-2D matching [19, 38, 40], or end-to-end training with differentiable perspective-n-point (PnP) [3, 29, 43]. This type of method is accurate on the trained target in general, which however, cannot work with novel objects that were not involved in the training. Retraining or fine-tuning can adapt the net" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.629 + ], + "angle": 0, + "content": "work to novel objects, which, however, is cumbersome and limits its applications in real scenarios. Some recent methods show promising results for novel object pose estimation [23, 32, 34, 36, 50], and one of their key components is object pose refinement, which is based on render-and compare [23, 32, 50] or feature matching [34, 36] strategies. Although these refinement methods work in general, they are less effective at utilizing prior shape information of the target. The recent method SCFlow [11] introduces a shape-constraint recurrent matching framework for object pose estimation, which embeds the target's 3D shape into the optimization. It produces accurate results for known objects, while has limited generalization ability to novel objects. SCFlow2 is built on top of the basic shape-constraint structure of SCFlow, but extends it by introducing 3D scene flow representation into the network optimization, producing an end-to-end system with RGBD images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Optical flow estimation is a fundamental computer vision task to estimate pixel-level correspondences between two frames [5, 17, 41, 52]. Most recent methods are based on deep learning models and achieve accurate results in general scenes [5, 41, 46]. The recent method RAFT [46] presents a recurrent framework for flow estimation, and RAFT-3D [47] extends it to estimate 3D scene flow with rigid motion embeddings. Although they show significant advantages over existing flow methods, estimating dense matching with these general flow methods for object pose estimation is less effective, without using the prior 3D shape information of the target. Furthermore, the general flow loss is a surrogate matching loss that does not directly reflect the final 6D object pose [11, 18, 20, 43] in the context of object pose estimation. We combine the shape-constraint formulation [11] and the rigid motion embeddings in 3D scene flow, producing a geometry-guided 6D object pose refiner that is end-to-end trainable and generalizes to novel objects." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22046" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.093, + 0.912, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.293, + 0.908, + 0.394 + ], + "angle": 0, + "content": "Figure 3. Overview of SCFlow2. Given an RGBD image and an initial pose of the target, we first render the target to obtain a synthetic RGBD image as the reference, and use an RGB encoder and depth encoder to extract features from the image pair, which will be used to create a 4D correlation volume. Based on the correlation volume and GRU, We use an intermediate flow regressor to predict 3D scene flow that is represented as a dense 3D transformation field \\(\\mathbf{T}_k^{\\prime}\\). We then use a pose regressor to predict a global pose update \\(\\Delta \\mathbf{P}_k\\) based on implicit voting from the pixel-wise 3D transformation field. Finally, the updated pose \\(\\mathbf{P}_k\\) is used to compute the pose-induced flow \\(F_{k}\\) based on the target mesh to index the correlation volume for the next iteration. Note how the depth and 3D shape of the target are embedded into the framework to guide the optimization iteratively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.182, + 0.431 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.55 + ], + "angle": 0, + "content": "Given an RGBD image and the 3D mesh of the object, our goal is to estimate the 6D pose of the target object. We assume the camera intrinsic matrix is known. We first introduce the framework at a high level in Section 3.1, and then present the design of the main components in Sections 3.2 and 3.3. We finally discuss the implementation details in Section 3.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.569, + 0.201, + 0.584 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.655 + ], + "angle": 0, + "content": "In general, we first render an RGBD image based on an initial pose of the target as a reference, and then predict the pose update based on comparing the rendered image and the real input with a recurrent matching network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.837 + ], + "angle": 0, + "content": "With the RGBD image pair, we first use an RGB encoder and depth encoder to extract features, and then construct a 4D correlation volume based on computing the dot product of the fused RGBD features. Based on the constructed correlation volume, we propose to use an intermediate flow regressor to predict 3D scene flow that is represented as a dense SE3 motion field [47]. We then use a pose regressor to predict a global pose update based on implicit voting from the pixel-wise SE3 motion field. Finally, the updated pose is used to compute a pose-induced flow based on the target mesh to index the correlation volume for the next iteration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The key of our method is the combination of the dense 3D scene flow representation and the embedding of the target 3D shape prior [11] in a unified recurrent matching framework. Fig. 3 shows the overview of this framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.418, + 0.807, + 0.432 + ], + "angle": 0, + "content": "3.2. Correlation Volume Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.652 + ], + "angle": 0, + "content": "We construct a 4D volume pyramid to represent the correlation between any two locations on the two input images. We use an RGB encoder and a depth encoder to extract features from the RGB and the depth image, respectively. For the rendered image \\( I_{1} \\) based on the initial pose and the input RGB image \\( I_{2} \\), we use an encoder with shared parameters to obtain their dense feature maps \\( f_{1}^{i} \\in \\mathbb{R}^{H \\times W \\times C} \\) and \\( f_{2}^{i} \\in \\mathbb{R}^{H \\times W \\times C} \\) at \\( 1/8 \\) of the original resolution. We use DINOv2 ViT-B [35] as our RGB feature extractor in the encoder, and freeze the pretrained weights during training. For the depths \\( D_{1} \\) and \\( D_{2} \\), we first lift the two depth maps to point clouds according to the camera intrinsic matrix, and then extract their 3D structural features with PointNet++ [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.729 + ], + "angle": 0, + "content": "Similar to DenseFusion [48], we fuse the RGB features and depth features to generate semantic and contextual features. We then compute the 4D correlation volume via computing the dot product of the fused features according to different location pairs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.83, + 0.757 + ], + "angle": 0, + "content": "3.3. Regressing 3D Transformation Field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.793 + ], + "angle": 0, + "content": "With a dense matching field as the index, we can extract a 3D correlation map from the 4D correlation volume." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.884 + ], + "angle": 0, + "content": "In SCFlow [11], the correlation map is used to predict the optical flow \\( f(u,v) \\), which maps the pixel position \\( \\mathbf{x}_1 = (u,v) \\) in \\( I_1 \\) to the corresponding position in \\( I_2 \\): \\( \\mathbf{x}_2 = (u,v) + f(u,v) \\). However, this 2D flow formulation only contains 2D information, which is less effective in capturing 3D motions." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To address this, we propose to predict 3D scene flow [47]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22047" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.212, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.275, + 0.171, + 0.288 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.098, + 0.321, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.275, + 0.304, + 0.289 + ], + "angle": 0, + "content": "Initial pose" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.098, + 0.437, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.276, + 0.391, + 0.287 + ], + "angle": 0, + "content": "T" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.098, + 0.553, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.276, + 0.508, + 0.288 + ], + "angle": 0, + "content": "\\(\\phi\\)" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.098, + 0.671, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.275, + 0.672, + 0.287 + ], + "angle": 0, + "content": "Intermediate flow" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.098, + 0.785, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.275, + 0.791, + 0.287 + ], + "angle": 0, + "content": "Pose-induced flow" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.089, + 0.911, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.276, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Refined pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.301, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Figure 4. Visualization of pose-induced flow with 3D scene flow representation. Given a real input image and the rendered image based on an initial pose, we predict 3D scene flow represented as a dense SE3 motion field. We represent the motion field as a twist field \\((\\tau, \\theta)\\). In theory, the motion fields should be constant pixels for rigid objects. We use a pose regressor to predict a global object-level pose based on the noisy motion fields. We then use the updated global pose to generate a pose-induced flow based on the target 3D mesh. The pose-induced flow embeds the shape prior of the target and reduces the search space for matching." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.483, + 0.473 + ], + "angle": 0, + "content": "as an intermediate representation for the network. Given \\( I_{1} \\) and \\( I_{2} \\) and their corresponding point clouds \\( C_1 \\) and \\( C_2 \\), we formulate scene flow as a dense 3D transformation field \\( \\mathbf{T} \\in SE(3)^{H \\times W} \\) to represent the 3D motion between the two point clouds in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Specifically, for the point \\( X_{i} \\) in \\( C_1 \\) with index \\( i \\) and its projected pixel coordinates \\( \\mathbf{x}_i = (u_i,v_i,z_i) \\), we can perform the transformation \\( \\mathbf{T}_i\\cdot X_i \\) to obtain the corresponding point \\( X_{i}^{\\prime} \\) in \\( C_2 \\) and its projected pixel coordinates \\( \\mathbf{x}_i' = (u_i',v_i',z_i') \\). The scene flow vector is defined as the difference \\( \\mathbf{f} = \\mathbf{x}_i' - \\mathbf{x}_i \\), where the first two components represent the standard 2D optical flow and the last component describes the depth difference between the two points." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.594, + 0.483, + 0.639 + ], + "angle": 0, + "content": "We use GRU [4] as the recurrent model for the intermediate scene flow predictor. In the \\(k^{th}\\) iteration, the hidden state update can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.649, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\operatorname {G R U} \\left(\\mathbf {L} _ {\\mathbf {C}} \\left(F _ {k - 1}\\right), \\mathbf {T} _ {k - 1}, h _ {k - 1}; \\Theta\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.483, + 0.781 + ], + "angle": 0, + "content": "where \\(\\mathbf{L}_{\\mathbf{C}}(F_{k - 1})\\) represents the correlation look-up operation guided by the global flow \\(F_{k - 1}\\) from the \\((k - 1)^{th}\\) iteration, \\(\\mathbf{T}_{k - 1}\\) is the dense 3D transformation field from the previous iteration, \\(\\Theta\\) denotes the GRU network parameters, and \\(h_k\\) is the hidden state feature which will be fed into a dense SE3 layer [47] to update the SE3 motion fields and generate a new transformation field \\(\\mathbf{T}_k^\\prime\\)" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.364, + 0.804 + ], + "angle": 0, + "content": "3.4. Regressing Global Object Pose" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our goal is to predict the pose residual of the single object in \\( I_{1} \\) and \\( I_{2} \\). Considering the rigidity of objects in 6D pose estimation scenarios, all pixel-level 3D motions in \\( \\mathbf{T}_k^\\prime \\) theoretically describe the same rigid object and should be the same. However, noise is inevitable in predicting the dense transformation field. Therefore, we designed a small" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.519 + ], + "angle": 0, + "content": "network to predict the global pose residual using the dense transformation field as input, as shown in Fig. 4. First, we represent the dense transformation field \\(\\mathbf{T}_k^{\\prime}\\) as a \\(4\\times 4\\) transformation matrix, and then we encode it using a three-layer 2D convolutional network. Finally, we use two fully connected layers to output the global pose residual \\(\\Delta \\mathbf{P}_k\\), which is represented as a 9-dim vector including 6-min for the rotation [54] and 3-dim for the normalized translation [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.659 + ], + "angle": 0, + "content": "In the \\(k^{th}\\) iteration, the pose residual \\(\\Delta \\mathbf{P}_k\\) will be used for the pose update: \\(\\mathbf{P}_k = \\mathbf{P}_{k - 1} \\otimes \\Delta \\mathbf{P}_k\\), and the updated pose \\(\\mathbf{P}_k\\) will be used to compute the pose-induced flow based on the object mesh, which imposes shape constraints for the lookup operation in the next iteration [11]. At the same time, we will generate a new dense transformation field \\(\\mathbf{T}_k\\) based on the updated pose \\(\\mathbf{P}_k\\). Specifically, we calculate the pose residual between the initial pose \\(\\mathbf{P}_0 = [\\mathbf{R}_0|\\mathbf{t}_0]\\) and \\(\\mathbf{P}_k = [\\mathbf{R}_k|\\mathbf{t}_k]\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.674, + 0.905, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {R} _ {k} = \\mathbf {R} _ {k} \\cdot \\mathbf {R} _ {0}, \\quad \\Delta \\mathbf {t} _ {k} = \\mathbf {t} _ {k} - \\Delta \\mathbf {R} _ {k} \\cdot \\mathbf {t} _ {0} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.767 + ], + "angle": 0, + "content": "Then, we update the 3D motion field \\(\\mathbf{T}_k\\in SE(3)^{H\\times W}\\), where the motion at every location will be updated as the same new pose residual. This new motion field will be the input for the next iteration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.783, + 0.731, + 0.8 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.907, + 0.853 + ], + "angle": 0, + "content": "For the input RGB and depth images \\( I_{1} \\) and \\( D_{2} \\), we first crop the target out based on the initial pose, and then resize it to a fixed resolution of \\( 256 \\times 256 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Given the ground truth pose, we can compute the ground truth optical flow based on the object mesh. We use the exponentially weighted strategy [11, 46] to calculate the loss" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22048" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.177, + 0.089, + 0.827, + 0.318 + ], + "angle": 0, + "content": "
MethodsLM-OT-LESSTUD-LIC-BINITODDHBYCB-VAvg.
MegaPose [23]62.048.584.646.246.072.576.462.3
MegaPose + Ours62.454.988.348.449.576.479.065.6
FoundPose [36]61.057.069.447.940.772.369.059.6
FoundPose + Ours69.658.688.154.648.476.284.968.6
GenFlow [32]63.552.186.253.455.477.983.367.4
GenFlow + Ours68.753.788.055.456.479.285.569.6
Gigapose [34]67.855.681.156.357.579.182.568.6
Gigapose + Ours71.058.682.057.758.380.184.470.3
SAM6D [27]69.951.590.458.860.277.684.570.4
SAM6D + Ours72.152.390.860.060.577.785.171.2
FoundationPose [50]75.664.692.350.858.083.588.973.4
FoundationPose + Ours80.165.095.551.758.384.491.475.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.329, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 1. Comparison with the state of the art. All the baselines have their own refinement strategy, and we report their final result published on BOP leaderboard. After using our method as a post-processing (“+ Ours”), the result consistently improves across all baselines on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.398, + 0.415, + 0.413 + ], + "angle": 0, + "content": "in each iteration for both flow loss and pose loss:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.421, + 0.483, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {k = 1} ^ {N} \\gamma^ {N - k} \\left(\\mathcal {L} _ {\\text {p o s e}} ^ {k} + \\alpha \\mathcal {L} _ {\\text {f l o w}} ^ {k}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.562 + ], + "angle": 0, + "content": "where the weight \\(\\gamma = 0.8\\), \\(N = 8\\) and \\(\\alpha = 0.1\\). \\(\\mathcal{L}_{flow}\\) represents the L1 loss between the first two dimensions of the predicted scene flow and the ground truth optical flow. \\(\\mathcal{L}_{pose}\\) refers to pose loss, which is computed as the L1 distance between the 3D point clouds transformed by the predicted pose and the ground truth pose, respectively [11, 49]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.224, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.705 + ], + "angle": 0, + "content": "We evaluate SCFlow2 on seven challenging BOP benchmark datasets in this section. Section 4.1 presents our experimental setup, including the datasets used for training and evaluation, training details and the evaluation metrics. In Section 4.2, we compare our approach with current state-of-the-art pose methods. Section 4.3 provides a detailed ablation study of the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.299, + 0.729 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Dataset. We train SCFlow2 on a combination of datasets including ShapeNet-Objects [2], Google-Scanned-Objects [8], and Objverse [6]. All these datasets only contain object meshes, in a total of about 90K objects. We reuse the synthetic images from [23] and [50] which are rendered based on those meshes, with about 3M images in total." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate on seven datasets from the BOP benchmark [15], including LM-O [1], T-LESS [14], TUD-L [15], IC-BIN [7], ITODD [9], HB [22], and YCB-V [51]. These datasets include 108 objects having different sizes and different texture and symmetry properties. Most of them are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.457 + ], + "angle": 0, + "content": "captured in challenging setups, including multiple object instances, significant clutter, and severed occlusions. These objects are not included in the training dataset and are novel to the models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.909, + 0.612 + ], + "angle": 0, + "content": "Training details. We train SCFlow2 using the AdamW optimizer [30] with a batch size of 16, and use a cosine annealing learning rate scheduler. The model is trained for 200k iterations with an initial learning rate of 1e-4. During training, we generate pseudo initial poses by adding random noise to the ground-truth pose. We add rotation jitter sampled from a normal distribution with a standard deviation of 15 degrees in each 3D axis, and add translation jitter sampled with deviations of 15, 15, and 50 millimeters along xyz axes, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Evaluation metrics We report the result in the common BOP metric, which is the mean Average Recall (AR) of three different metrics, including Visible Surface Discrepancy (VSD), Maximum Symmetry-aware Surface Distance (MSSD), and Maximum Symmetry-aware Projection Distance (MSPD). Readers can refer to [15] for further details about these metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.836, + 0.755 + ], + "angle": 0, + "content": "4.2. Comparison with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To explore the plug-and-play nature of our refinement method, we apply it to the final results of Megapose [23], FoundPose [36], GenFlow [32], GigaPose [34], SAM6D [27], and FoundationPose [50], as shown in Table 1. Although these baselines have a refinement strategy themselves, plugging our refinement method into the end of their framework further improves the results, consistently on all datasets. Most baseline results are downloaded from the BOP leaderboards." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22049" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.296 + ], + "angle": 0, + "content": "
MethodsDepthLM-OYCB-V
FoundPose (init)-39.645.2
FoundPose (final)✓61.069.0
FoundPose (init) + Ours✓61.381.8
FoundPose (final) + Ours✓69.684.9
GenFlow (init)-25.027.7
GenFlow (final)✓63.583.3
GenFlow (init) + Ours✓64.483.6
GenFlow (final) + Ours✓68.785.5
FoundationPose (init)---
FoundationPose (final)✓75.688.9
FoundationPose (init) + Ours✓79.790.4
FoundationPose (final) + Ours✓80.191.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.378 + ], + "angle": 0, + "content": "Table 2. Effectiveness as a plug-and-play object pose refiner. All the baselines are based on an initialization-refinement architecture. Our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.395, + 0.487, + 0.45 + ], + "angle": 0, + "content": "
L3L5L10L20L30L40L50
Initialization92.783.864.53517.67.83.5
+ SCFlow291.691.591.590.888.683.270.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.53 + ], + "angle": 0, + "content": "Table 3. Impact of pose initialization. We add randomly generated rotation and translation noise to GT pose as the initialization, and then use SCFlow2 to refine the pose. L3 to L50 denote different noise levels. SCFlow2 is robust to pose errors in the initialization and works well under noise level 30." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.245, + 0.577 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness as a plug-and-play object pose refiner. We evaluate SCFlow2 in more detail in the context of three different state-of-the-art methods, FoundPose [36], GenFlow [32], and FoundationPose [50]. Table 2 summaries the results. All of them are based on an initialization-refinement architecture. We denote their initialization results as \"init\" in the table, and replace their refinement methods with ours in each method (denoted as \"init + ours\"). Furthermore, we also report the results of applying our refinement to their final results directly, denoted as \"final + ours\". Note that, FoundationPose outputs hundreds of pose hypotheses in the initialization stage, and there is no well-defined initialization result for it. In this case, we replace its original refinement method with SCFlow2 in its framework, and keep all other components untouched, including post-processing procedures such as pose scoring and pose selection, which is the reported result of \"FoundationPose (init) + Ours\" in the table. Overall, our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner. We show some qualitative results in Fig. 5 and 6." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.089, + 0.878, + 0.264 + ], + "angle": 0, + "content": "
DatasetsInit.SeenUnseen
V1V1+V1++V2
LM-O61.062.969.368.969.6
T-LESS57.057.954.955.358.6
TUD-L69.475.988.886.688.1
IC-BIN47.947.954.153.754.6
ITODD40.734.249.148.148.4
HB72.374.576.576.176.2
YCB-V69.080.785.584.484.9
Avg.59.662.068.367.668.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.273, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Table 4. Advantages over SCFlow. We use the same initialization based on results from FoundPose (\"Init.\", and report results of SCFlow (\"V1\") and SCFlow2 (\"V2\"). SCFlow2 outperforms all \"V1\" versions without any retraining." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Impact of pose initialization. We add random rotation and translation noise to GT pose as the initialization, and report the results on the YCB-V dataset, as shown in Table 3. L3 to L50 denote different noise levels (rotation error from 3 to 50 degrees, translation error from 3 to \\(50~\\mathrm{mm}\\)). SCFlow2 works well under noise level 30." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Evaluation of performance with different iterations. We conduct ablation studies on the trade off between the accuracy/efficiency and the iteration numbers of our method, as shown in Fig. 7. Our method produces meaningful pose refinement results after only 2 iterations and reaches saturation after 6 iterations. In our experiments, we use 8 iterations as the default setting in both training and testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Advantages over SCFlow. We compare SCFlow2 and its previous version SCFlow systematically. We summarize the results in Table 4. We use the same initialization based on results from FoundPose, and refine it with SCFlow (\"V1\") and SCFlow2 (\"V2\"), respectively. We evaluate three different versions of SCFlow, where \"V1\" denotes the original version based on retraining on target objects with RGB images only, \"V1+\" denotes the version using RANSAC Kabsch to consume additional depth images. Both \"V1\" and \"V1+\" need retraining on target objects. We also evaluate a version \"V1++\" which does not rely on retraining for novel objects, but uses the same large-scale datasets as \"V2\" for training on RGB images and then refine the pose results using RANSAC Kabsch with additional depths. We can see that, without retraining, \"V1++\" produces worse results compared with \"V1+\". On the other hand, SCFlow2 even outperforms \"V1+\" without any retraining." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation with different settings. We use the initialization results of FoundPose as our initialization. Table 5 shows the result, where \"Shape\" denotes the shape-constraint regularization during the recurrent optimization, and \"Scene\" is the 3D scene flow representation. We use the intermediate flow for the look-up without relying on the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22050" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.235, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.089, + 0.366, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.089, + 0.498, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.631, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.089, + 0.764, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.089, + 0.897, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.194, + 0.234, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.194, + 0.366, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.194, + 0.498, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.194, + 0.631, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.194, + 0.764, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.194, + 0.897, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.297, + 0.233, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.297, + 0.366, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.297, + 0.498, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.297, + 0.631, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.297, + 0.764, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.297, + 0.897, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.402, + 0.233, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.402, + 0.366, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.402, + 0.498, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.402, + 0.631, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.402, + 0.764, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.402, + 0.897, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.507, + 0.233, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.507, + 0.366, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.507, + 0.498, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.507, + 0.631, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.507, + 0.764, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.507, + 0.897, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.612, + 0.233, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.612, + 0.366, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.612, + 0.498, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.612, + 0.631, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.612, + 0.764, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.612, + 0.897, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.718, + 0.207, + 0.731 + ], + "angle": 0, + "content": "GFlow (init)" + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.718, + 0.342, + 0.731 + ], + "angle": 0, + "content": "GFlow (final)" + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.718, + 0.498, + 0.731 + ], + "angle": 0, + "content": "GFlow (init) + Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.718, + 0.601, + 0.731 + ], + "angle": 0, + "content": "FPose (init)" + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.718, + 0.737, + 0.731 + ], + "angle": 0, + "content": "FPose (final)" + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.718, + 0.893, + 0.731 + ], + "angle": 0, + "content": "FPose (init) + Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.744, + 0.908, + 0.787 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. Given the same pose initialization as that in GenFlow (\"GFlow\") and FoundPose (\"FPose\"), denoted as \"GFlow (init)\" and \"FPose (init)\" respectively, our refinement method (\"+ Ours\") produces considerably more accurate results compared to the refinement approaches in their original methods (note how our reprojected 3D mesh aligns better with the object contours)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.486, + 0.889 + ], + "angle": 0, + "content": "target's 3D mesh when \"Shape\" is disabled, and we replace the scene flow representation with a plain MLP pose regressor when \"Scene\" is disabled. The performance becomes worse without the shape prior embedding, and deteriorates significantly without the 3D scene flow formulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.813, + 0.909, + 0.89 + ], + "angle": 0, + "content": "Timing analysis. The training of SCFlow2 needs about 108 hours on a workstation with 2 NVIDIA RTX-3090 GPU. SCFlow2 considers the pose initialization as the only hypothesis. During inference, SCFlow2 takes about 0.18 seconds to refine a pose on an RTX-3090 GPU with a typical" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.288, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.477, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.201, + 0.286, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.315, + 0.232, + 0.327 + ], + "angle": 0, + "content": "Initialization" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.201, + 0.477, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.316, + 0.405, + 0.327 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.341, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Figure 6. More qualitative results. We show two examples on YCB-V. Although the initialization has large pose error, our refinement method produces accurate results that align well with the target, without retraining on these novel objects." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.413, + 0.416, + 0.49 + ], + "angle": 0, + "content": "
ShapeSceneLM-OYCB-V
✓-52.667.6
-✓60.080.4
✓✓61.381.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Table 5. Evaluation with different settings. \"Shape\" denotes the shape-constraint regularization during the optimization, and \"Scene\" is the 3D scene flow formulation. The performance deteriorates significantly without the 3D scene flow formulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.645 + ], + "angle": 0, + "content": "image from BOP dataset. Taken for comparison, the multi-hypothesis based method MegaPose needs about 1.85 seconds for a refinement on the same workstation, which is \\(10+\\) times slower than SCFlow2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Limitation discussion. In general, our refinement method can improve the accuracy of initial pose significantly. While occasionally, it produces worse results than the initial pose. Fig. 8 shows the histogram of samples along with the relative accuracy improvement after pose refinement on YCB-V and LM-O with the initial pose of FoundPose as our initialization. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases where the initial poses are already very wrong caused by severe occlusion or extreme light conditions. We will take this as one of our future work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We have introduced SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. We extended" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.707, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.091, + 0.903, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.224, + 0.906, + 0.307 + ], + "angle": 0, + "content": "Figure 7. Evaluation of performance with different iterations. Given the same pose initialization with noise level 30 as in Table 3, more recurrent iterations produce more accurate results while spending more time in inference. Nevertheless, our approach yields meaningful outcomes after just 2 iterations, and the performance reaches saturation after 6 iterations." + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.346, + 0.807, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.481, + 0.906, + 0.55 + ], + "angle": 0, + "content": "Figure 8. Limitation discussion. We show the histogram of samples along with the relative accuracy improvement after pose refinement. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.906, + 0.737 + ], + "angle": 0, + "content": "the RGB-based method SCFlow by formulating additional depth as a regularization in the iteration with an intermediate 3D scene flow representation for RGBD frames, and combined it with the 3D shape prior information of the target, producing a geometry-guided 6D object pose refiner end-to-end trainable and generalizable to novel objects. Experiments on BOP datasets with novel objects show that our method achieves state-of-the-art accuracy, and is one of the most efficient refinement methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the Youth Innovation Team of Shaanxi Universities, the \"Scientist + Engineer\" team of the Qin Chuang Yuan in Shaanxi Province, the Xi'an Science and Technology Program, the \"Leading the Charge\" initiative for the industrialization of core technologies in key industrial chains in Shaanxi Province, the National Nature Science Foundation of China under Grant 62371359, the Key Research and Development Program of Shaanxi, the Key Research Program of the Chinese Academy of Sciences." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22052" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D Object Pose Estimation Using 3D Object Coordinates. In European Conference on Computer Vision, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-End Learnable Geometric Vision by Backpropagating PnP Optimization. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv preprint arXiv:1409.1259, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[5] Javier Cuadrado, Ulysse Rancon, Benoit R Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical Flow Estimation from Event-based Cameras and Spiking Neural Networks. Frontiers in Neuroscience, 17, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A Universe of Annotated 3D Objects. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6D Object Pose and Predicting Next-Best-View in the Crowd. In Conference on Computer Vision and Pattern Recognition, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items. In International Conference on Robotics and Automation, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing MVTec ITODDA Dataset for 3D Object Recognition in Industry. In International Conference on Computer Vision Workshops, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Yang Hai, Rui Song, Jiaojiao Li, David Ferstl, and Yinlin Hu. Pseudo Flow Consistency for Self-Supervised 6D Object Pose Estimation. In International Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-Constraint Flow for 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Rasmus Laurvig Haugaard and Anders Glent Buch. SurfEmb: Dense and Continuous Correspondence Distributions for Object Pose Estimation with Learnt Surface Embeddings. In Conference on Computer Vision and Pattern Recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. FS6D: Few-Shot 6D Pose Estimation of Novel Ob" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "jects. In Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[14] Tomás Hodań, Pavel Haluza, Štepan Obdržánek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-Less Objects. In Winter Conference on Applications of Computer Vision, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[15] Tomáš Hodař, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In European Conference on Computer Vision, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[16] Tomás Hodań, Daniel Barath, and Jérí Matas. EPOS: Estimating 6D Pose of Objects with Symmetries. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[17] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.381 + ], + "angle": 0, + "content": "[18] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-Stage 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[19] Yinlin Hu, Sebastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-Depth-Range 6D Object Pose Estimation in Space. In Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[20] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective Flow Aggregation for Data-Limited 6D Object Pose Estimation. In European Conference on Computer Vision, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[21] Wolfgang Kabsch. A Solution for the Best Rotation to Relate Two Sets of Vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.557, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[22] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D Dataset for 6D Pose Estimation of 3D Objects. In International Conference on Computer Vision Workshops, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[23] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In Conference on Robot Learning. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent Multi-View Multi-Object 6d Pose Estimation. In European Conference on Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep Iterative Matching for 6D Pose Estimation. In European Conference on Computer Vision, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22053" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[28] Yongliang Lin, Yongzhi Su, Praveen Nathan, Sandeep Inuganti, Yan Di, Martin Sundermeyer, Fabian Manhardt, Didier Stricker, Jason Rambach, and Yu Zhang. HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Fulin Liu, Yinlin Hu, and Mathieu Salzmann. Linear-Covariance Loss for End-to-End Learning of 6D Pose Estimation. In International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[31] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose Estimation for Augmented Reality: A Hands-On Survey. IEEE transactions on visualization and computer graphics, 22(12):2633-2651, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[32] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[33] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. NOPE: Novel Object Pose Estimation from a Single Image. In Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[34] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOV2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[36] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found-Pose: Unseen Object Pose Estimation with Foundation Features. European Conference on Computer Vision, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[37] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. LatentFusion: End-to-End Differentiable Reconstruction and Rendering for Unseen Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[38] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-Wise Voting Network for 6DoF Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. arXiv preprint arXiv:1706.02413, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects without Using Depth. In International Conference on Computer Vision, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[42] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A Multi-Stage One Shot Object Pose Estimation Framework. In Conference on Computer Vision and Pattern Recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[43] Chen Song, Jiaru Song, and Qixing Huang. HybridPose: 6D Object Pose Estimation under Hybrid Representations. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[44] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation. arXiv preprint arXiv:2203.09418, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[45] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-Shot Object Pose Estimation without CAD Models. In Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.454 + ], + "angle": 0, + "content": "[46] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In European Conference on Computer Vision, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[47] Zachary Teed and Jia Deng. RAFT-3D: Scene Flow using Rigid-Motion Embeddings. In Conference on Computer Vision and Pattern Recognition, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In Conference on Computer Vision and Pattern Recognition, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[49] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[50] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes. In Robotics: Science and Systems, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[52] Zhaolin Xiao, Zhan Su, Fengyuan Zuo, and Haiyan Jin. Low-Light Optical Flow Estimation with Hidden Feature Supervision Using a Siamese Network. Journal of Image and Graphics, 29:0231-0242, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[53] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Conference on Computer Vision and Pattern Recognition, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22054" + } + ] +] \ No newline at end of file diff --git a/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_origin.pdf b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd080b879deb9bd766e90d46b5fdd86287677140 --- /dev/null +++ b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/bd2181f3-adaf-406c-ac89-ebbf9c811d2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76a25a9beafce2be81e8d1ac8aba3a11e74b9a07dee953aa335d8b939f8ea6a3 +size 6416106 diff --git a/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/full.md b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..40d3a961798a85c80f3ab246fbea8b77ca0766f9 --- /dev/null +++ b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/full.md @@ -0,0 +1,377 @@ +# SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow + +Qingyuan Wang $^{1}$ , Rui Song $^{1}$ , Jiaojiao Li $^{1}$ , Kerui Cheng $^{2}$ , David Ferstl $^{3}$ , Yinlin Hu $^{3}$ $^{1}$ State Key Laboratory of ISN, Xidian University $^{2}$ Taiyuan University of Technology $^{3}$ MagicLeap + +# Abstract + +We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. Most recent 6D object pose methods rely on refinement to get accurate results. However, most existing refinement methods either suffer from noises in establishing correspondences, or rely on retraining for novel objects. SCFlow2 is based on the SCFlow model designed for refinement with shape constraint, but formulates the additional depth as a regularization in the iteration via 3D scene flow for RGBD frames. The key design of SCFlow2 is an introduction of geometry constraints into the training of recurrent matching network, by combining the rigid-motion embeddings in 3D scene flow and 3D shape prior of the target. We train SCFlow2 on a combination of dataset Objaverse, GSO and ShapeNet, and evaluate on BOP datasets with novel objects. After using our method as a post-processing, most state-of-the-art methods produce significantly better results, without any retraining or fine-tuning. The source code is available at https://scflow2.github.io. + +# 1. Introduction + +6D object pose estimation, which determines an object's 3D rotation and 3D translation relative to a camera, is crucial in fields like robotics, augmented reality, and spatial computing [31, 48]. + +Most existing pose estimation methods rely on a refinement procedure to get accurate pose results [10, 12, 20, 23, 24, 48, 50], and most of the refinement methods are based on render-and-compare strategies [10, 20, 23, 50]. A group of refinement methods [24, 48] achieves highly accurate results based on retraining or fine-tuning the network for novel objects, which, however, is cumbersome and unfriendly to real applications. Some recent refinement methods [23, 42, 50] have shown promising results without fine-tuning for novel objects. However, they still suffer from several weaknesses. + +First, most methods formulate the render-and-compare framework as a general matching problem, but there is a strong prior of the target's 3D shape that is not used, mak + +![](images/2d9e7353265abbc3cc39e2b8711d1dae8cdd8e26005018cb014ec4688486c7a8.jpg) +Initialization + +![](images/39e9f1a017519c747248789e2f543723571f1a48d63fcda0fff46d8f36974daa.jpg) +Network + +![](images/fc9a0acc4c70262982e7ed8cbbfe9f28059af475cb02e8082fa5569f12da9f32.jpg) +Refinement + +![](images/e26ec74bbd33569cd6465a71ed55909bf8fb3c661efcfd670244a606f7e67ed6.jpg) +(a) Object pose refinement with retraining +Initialization +(b) Plug-and-play SCFlow2 +Figure 1. Object pose refinement is critical for accurate object pose estimation. (a) Most existing object pose refinement methods, including SCFlow [11], rely on retraining for novel objects to achieve high accuracy. (b) The proposed SCFlow2 achieves even higher accuracy, and more importantly, generalizes well to novel objects without any retraining or fine-tuning. + +![](images/96dfa597e0a6fed791cd94d81be963687b6d2c25b6700f88b0a048052197a73a.jpg) + +![](images/9e2c47419cf7f86407f32c13c84a98b02450e63278a95d4905b30a135ec27a0d.jpg) + +Refinement + +ing it usually has an unnecessarily large search space for matching and is less effective for 6D object pose estimation [11]. Second, to address the matching issue, most methods attempt to refine the pose starting from multiple pose hypotheses based on the initialization [23, 32, 36]. While this paradigm is effective, it significantly slows down the refinement process. Third, most methods formulate the comparison after the rendering as a 2D matching procedure only, and then use ICP or Kabsch [21] to consume additional depth input as a second stage, which is less effective in capturing 3D motions and local optimal within each stage. + +We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation, as shown in Fig. 1. SCFlow2 is based on SCFlow [11] which is a state-of-the-art refinement method embedding 3D shape constraints into the optimization for known objects. We build on the basic shape-constraint structure of SCFlow but extend it in several novel ways. + +![](images/9952a3a3f8f062e595a961342928d43ba8b64f83b476d0db771ae47febb7d09c.jpg) +(a) The framework of SCFlow + +![](images/89961d668d373cd6376f0275c610c970361917171d098a01020651d5bf062d97.jpg) +(b) The framework of SCFlow2 +Figure 2. Design overview of SCFlow and SCFlow2. Given the object 3D mesh, we render an image $I_{1}$ and depth map $D_{1}$ based on an initial pose, and then use networks to compare these rendered outputs with the real input $I_{2}$ and $D_{2}$ to refine the pose. (a) Although SCFlow [11] adds 3D shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, which is less effective in capturing 3D motions. On the other hand, it cannot work with RGBD images. A common practice is to use RANSAC Kabsch [21] to consume additional depth as a second stage, which however is only local optimal within each stage. (b) SCFlow2 tackles these problems. We introduce an intermediate representation based on 3D scene flow to capture 3D motions in network optimization. Furthermore, we embed depth into the loop by formulating depth as an additional regularization to guide the correlation look-up iteratively, producing an end-to-end trainable system with RGBD images. + +Our main design of SCFlow2 is introducing a scene flow based representation into the iterative optimization. Although SCFlow embeds shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, and suffers in the solution with a second stage of RANSAC Kabsch [21] for RGBD images. On the other hand, SCFlow cannot work with novel objects. To handle these problems, we introduce rigid-motion embeddings in 3D scene flow [47] as our intermediate representation, and formulate depth as an additional regularization to guide the network optimization iteratively, which helps reducing search space and results in an end-to-end system that generalizes to novel objects, as illustrated in Fig. 2. + +We train SCFlow2 on a combination of dataset Objverse [6], GSO [8] and ShapeNet [2], and evaluate on BOP datasets [15] with novel objects. SCFlow2 achieves state-of-the-art accuracy, and can produce accurate results with only one pose hypothesis, outperforming most multiple-hypothesis based refinement methods [23, 32, 36] in both accuracy and efficiency, without any retraining. + +# 2. Related Work + +Object pose estimation aims to predict the 3D rotation and 3D translation with respect to the camera, and has shown prominent progress with deep learning techniques [13, 27, 28, 33, 37, 45]. Most existing methods assume the existence of many real images of the target, and train a learning model based on those real images [11, 19, 20, 38, 44], based on either 2D-to-3D matching [16, 26, 53], 3D-to-2D matching [19, 38, 40], or end-to-end training with differentiable perspective-n-point (PnP) [3, 29, 43]. This type of method is accurate on the trained target in general, which however, cannot work with novel objects that were not involved in the training. Retraining or fine-tuning can adapt the net + +work to novel objects, which, however, is cumbersome and limits its applications in real scenarios. Some recent methods show promising results for novel object pose estimation [23, 32, 34, 36, 50], and one of their key components is object pose refinement, which is based on render-and compare [23, 32, 50] or feature matching [34, 36] strategies. Although these refinement methods work in general, they are less effective at utilizing prior shape information of the target. The recent method SCFlow [11] introduces a shape-constraint recurrent matching framework for object pose estimation, which embeds the target's 3D shape into the optimization. It produces accurate results for known objects, while has limited generalization ability to novel objects. SCFlow2 is built on top of the basic shape-constraint structure of SCFlow, but extends it by introducing 3D scene flow representation into the network optimization, producing an end-to-end system with RGBD images. + +Optical flow estimation is a fundamental computer vision task to estimate pixel-level correspondences between two frames [5, 17, 41, 52]. Most recent methods are based on deep learning models and achieve accurate results in general scenes [5, 41, 46]. The recent method RAFT [46] presents a recurrent framework for flow estimation, and RAFT-3D [47] extends it to estimate 3D scene flow with rigid motion embeddings. Although they show significant advantages over existing flow methods, estimating dense matching with these general flow methods for object pose estimation is less effective, without using the prior 3D shape information of the target. Furthermore, the general flow loss is a surrogate matching loss that does not directly reflect the final 6D object pose [11, 18, 20, 43] in the context of object pose estimation. We combine the shape-constraint formulation [11] and the rigid motion embeddings in 3D scene flow, producing a geometry-guided 6D object pose refiner that is end-to-end trainable and generalizes to novel objects. + +![](images/af008dda54be056434a6663231d3a00e3944b3a311335ebbb42f830665f96048.jpg) +Figure 3. Overview of SCFlow2. Given an RGBD image and an initial pose of the target, we first render the target to obtain a synthetic RGBD image as the reference, and use an RGB encoder and depth encoder to extract features from the image pair, which will be used to create a 4D correlation volume. Based on the correlation volume and GRU, We use an intermediate flow regressor to predict 3D scene flow that is represented as a dense 3D transformation field $\mathbf{T}_k^{\prime}$ . We then use a pose regressor to predict a global pose update $\Delta \mathbf{P}_k$ based on implicit voting from the pixel-wise 3D transformation field. Finally, the updated pose $\mathbf{P}_k$ is used to compute the pose-induced flow $F_{k}$ based on the target mesh to index the correlation volume for the next iteration. Note how the depth and 3D shape of the target are embedded into the framework to guide the optimization iteratively. + +# 3. Method + +Given an RGBD image and the 3D mesh of the object, our goal is to estimate the 6D pose of the target object. We assume the camera intrinsic matrix is known. We first introduce the framework at a high level in Section 3.1, and then present the design of the main components in Sections 3.2 and 3.3. We finally discuss the implementation details in Section 3.5. + +# 3.1. Overview + +In general, we first render an RGBD image based on an initial pose of the target as a reference, and then predict the pose update based on comparing the rendered image and the real input with a recurrent matching network. + +With the RGBD image pair, we first use an RGB encoder and depth encoder to extract features, and then construct a 4D correlation volume based on computing the dot product of the fused RGBD features. Based on the constructed correlation volume, we propose to use an intermediate flow regressor to predict 3D scene flow that is represented as a dense SE3 motion field [47]. We then use a pose regressor to predict a global pose update based on implicit voting from the pixel-wise SE3 motion field. Finally, the updated pose is used to compute a pose-induced flow based on the target mesh to index the correlation volume for the next iteration. + +The key of our method is the combination of the dense 3D scene flow representation and the embedding of the target 3D shape prior [11] in a unified recurrent matching framework. Fig. 3 shows the overview of this framework. + +# 3.2. Correlation Volume Construction + +We construct a 4D volume pyramid to represent the correlation between any two locations on the two input images. We use an RGB encoder and a depth encoder to extract features from the RGB and the depth image, respectively. For the rendered image $I_{1}$ based on the initial pose and the input RGB image $I_{2}$ , we use an encoder with shared parameters to obtain their dense feature maps $f_{1}^{i} \in \mathbb{R}^{H \times W \times C}$ and $f_{2}^{i} \in \mathbb{R}^{H \times W \times C}$ at $1/8$ of the original resolution. We use DINOv2 ViT-B [35] as our RGB feature extractor in the encoder, and freeze the pretrained weights during training. For the depths $D_{1}$ and $D_{2}$ , we first lift the two depth maps to point clouds according to the camera intrinsic matrix, and then extract their 3D structural features with PointNet++ [39]. + +Similar to DenseFusion [48], we fuse the RGB features and depth features to generate semantic and contextual features. We then compute the 4D correlation volume via computing the dot product of the fused features according to different location pairs. + +# 3.3. Regressing 3D Transformation Field + +With a dense matching field as the index, we can extract a 3D correlation map from the 4D correlation volume. + +In SCFlow [11], the correlation map is used to predict the optical flow $f(u,v)$ , which maps the pixel position $\mathbf{x}_1 = (u,v)$ in $I_1$ to the corresponding position in $I_2$ : $\mathbf{x}_2 = (u,v) + f(u,v)$ . However, this 2D flow formulation only contains 2D information, which is less effective in capturing 3D motions. + +To address this, we propose to predict 3D scene flow [47] + +![](images/f7caf4371b06dca9632e0d2c34e391d8d2f735912e0e1466c7345a1fc4d954e3.jpg) +Input + +![](images/4eb94e7619a549af88556663cf2cedcae0c20320fcb868d6775351fb556cd580.jpg) +Initial pose + +![](images/5a85df86751bd2e885784022a1c176fdf79ad41166f2492361758b2385f5a697.jpg) +T + +![](images/1ee6d55a9d7d9eab885f3b2ab3b4413a91b393670f436f4873ddbe616223c415.jpg) +$\phi$ +Figure 4. Visualization of pose-induced flow with 3D scene flow representation. Given a real input image and the rendered image based on an initial pose, we predict 3D scene flow represented as a dense SE3 motion field. We represent the motion field as a twist field $(\tau, \theta)$ . In theory, the motion fields should be constant pixels for rigid objects. We use a pose regressor to predict a global object-level pose based on the noisy motion fields. We then use the updated global pose to generate a pose-induced flow based on the target 3D mesh. The pose-induced flow embeds the shape prior of the target and reduces the search space for matching. + +![](images/67fc95840ac40972906fc606e7a5ff0308f9633baea68d51a0f9f72aaa418c13.jpg) +Intermediate flow + +![](images/7df50d2dc89c1441b0f5f0b8ba3b23a982951c004e0534a1eecc122df6844d6c.jpg) +Pose-induced flow + +![](images/27b035c8982fe8e4555ea73464bc517e9cf36ab8bbf3f047db07848364c86a53.jpg) +Refined pose + +as an intermediate representation for the network. Given $I_{1}$ and $I_{2}$ and their corresponding point clouds $C_1$ and $C_2$ , we formulate scene flow as a dense 3D transformation field $\mathbf{T} \in SE(3)^{H \times W}$ to represent the 3D motion between the two point clouds in 3D space. + +Specifically, for the point $X_{i}$ in $C_1$ with index $i$ and its projected pixel coordinates $\mathbf{x}_i = (u_i,v_i,z_i)$ , we can perform the transformation $\mathbf{T}_i\cdot X_i$ to obtain the corresponding point $X_{i}^{\prime}$ in $C_2$ and its projected pixel coordinates $\mathbf{x}_i' = (u_i',v_i',z_i')$ . The scene flow vector is defined as the difference $\mathbf{f} = \mathbf{x}_i' - \mathbf{x}_i$ , where the first two components represent the standard 2D optical flow and the last component describes the depth difference between the two points. + +We use GRU [4] as the recurrent model for the intermediate scene flow predictor. In the $k^{th}$ iteration, the hidden state update can be expressed as: + +$$ +h _ {k} = \operatorname {G R U} \left(\mathbf {L} _ {\mathbf {C}} \left(F _ {k - 1}\right), \mathbf {T} _ {k - 1}, h _ {k - 1}; \Theta\right) \tag {1} +$$ + +where $\mathbf{L}_{\mathbf{C}}(F_{k - 1})$ represents the correlation look-up operation guided by the global flow $F_{k - 1}$ from the $(k - 1)^{th}$ iteration, $\mathbf{T}_{k - 1}$ is the dense 3D transformation field from the previous iteration, $\Theta$ denotes the GRU network parameters, and $h_k$ is the hidden state feature which will be fed into a dense SE3 layer [47] to update the SE3 motion fields and generate a new transformation field $\mathbf{T}_k^\prime$ + +# 3.4. Regressing Global Object Pose + +Our goal is to predict the pose residual of the single object in $I_{1}$ and $I_{2}$ . Considering the rigidity of objects in 6D pose estimation scenarios, all pixel-level 3D motions in $\mathbf{T}_k^\prime$ theoretically describe the same rigid object and should be the same. However, noise is inevitable in predicting the dense transformation field. Therefore, we designed a small + +network to predict the global pose residual using the dense transformation field as input, as shown in Fig. 4. First, we represent the dense transformation field $\mathbf{T}_k^{\prime}$ as a $4\times 4$ transformation matrix, and then we encode it using a three-layer 2D convolutional network. Finally, we use two fully connected layers to output the global pose residual $\Delta \mathbf{P}_k$ , which is represented as a 9-dim vector including 6-min for the rotation [54] and 3-dim for the normalized translation [25]. + +In the $k^{th}$ iteration, the pose residual $\Delta \mathbf{P}_k$ will be used for the pose update: $\mathbf{P}_k = \mathbf{P}_{k - 1} \otimes \Delta \mathbf{P}_k$ , and the updated pose $\mathbf{P}_k$ will be used to compute the pose-induced flow based on the object mesh, which imposes shape constraints for the lookup operation in the next iteration [11]. At the same time, we will generate a new dense transformation field $\mathbf{T}_k$ based on the updated pose $\mathbf{P}_k$ . Specifically, we calculate the pose residual between the initial pose $\mathbf{P}_0 = [\mathbf{R}_0|\mathbf{t}_0]$ and $\mathbf{P}_k = [\mathbf{R}_k|\mathbf{t}_k]$ as follows: + +$$ +\Delta \mathbf {R} _ {k} = \mathbf {R} _ {k} \cdot \mathbf {R} _ {0}, \quad \Delta \mathbf {t} _ {k} = \mathbf {t} _ {k} - \Delta \mathbf {R} _ {k} \cdot \mathbf {t} _ {0} \tag {2} +$$ + +Then, we update the 3D motion field $\mathbf{T}_k\in SE(3)^{H\times W}$ , where the motion at every location will be updated as the same new pose residual. This new motion field will be the input for the next iteration. + +# 3.5. Implementation Details + +For the input RGB and depth images $I_{1}$ and $D_{2}$ , we first crop the target out based on the initial pose, and then resize it to a fixed resolution of $256 \times 256$ . + +Given the ground truth pose, we can compute the ground truth optical flow based on the object mesh. We use the exponentially weighted strategy [11, 46] to calculate the loss + +
MethodsLM-OT-LESSTUD-LIC-BINITODDHBYCB-VAvg.
MegaPose [23]62.048.584.646.246.072.576.462.3
MegaPose + Ours62.454.988.348.449.576.479.065.6
FoundPose [36]61.057.069.447.940.772.369.059.6
FoundPose + Ours69.658.688.154.648.476.284.968.6
GenFlow [32]63.552.186.253.455.477.983.367.4
GenFlow + Ours68.753.788.055.456.479.285.569.6
Gigapose [34]67.855.681.156.357.579.182.568.6
Gigapose + Ours71.058.682.057.758.380.184.470.3
SAM6D [27]69.951.590.458.860.277.684.570.4
SAM6D + Ours72.152.390.860.060.577.785.171.2
FoundationPose [50]75.664.692.350.858.083.588.973.4
FoundationPose + Ours80.165.095.551.758.384.491.475.2
+ +Table 1. Comparison with the state of the art. All the baselines have their own refinement strategy, and we report their final result published on BOP leaderboard. After using our method as a post-processing (“+ Ours”), the result consistently improves across all baselines on all datasets. + +in each iteration for both flow loss and pose loss: + +$$ +\mathcal {L} = \sum_ {k = 1} ^ {N} \gamma^ {N - k} \left(\mathcal {L} _ {\text {p o s e}} ^ {k} + \alpha \mathcal {L} _ {\text {f l o w}} ^ {k}\right) \tag {3} +$$ + +where the weight $\gamma = 0.8$ , $N = 8$ and $\alpha = 0.1$ . $\mathcal{L}_{flow}$ represents the L1 loss between the first two dimensions of the predicted scene flow and the ground truth optical flow. $\mathcal{L}_{pose}$ refers to pose loss, which is computed as the L1 distance between the 3D point clouds transformed by the predicted pose and the ground truth pose, respectively [11, 49]. + +# 4. Experiments + +We evaluate SCFlow2 on seven challenging BOP benchmark datasets in this section. Section 4.1 presents our experimental setup, including the datasets used for training and evaluation, training details and the evaluation metrics. In Section 4.2, we compare our approach with current state-of-the-art pose methods. Section 4.3 provides a detailed ablation study of the proposed framework. + +# 4.1. Experimental Settings + +Dataset. We train SCFlow2 on a combination of datasets including ShapeNet-Objects [2], Google-Scanned-Objects [8], and Objverse [6]. All these datasets only contain object meshes, in a total of about 90K objects. We reuse the synthetic images from [23] and [50] which are rendered based on those meshes, with about 3M images in total. + +We evaluate on seven datasets from the BOP benchmark [15], including LM-O [1], T-LESS [14], TUD-L [15], IC-BIN [7], ITODD [9], HB [22], and YCB-V [51]. These datasets include 108 objects having different sizes and different texture and symmetry properties. Most of them are + +captured in challenging setups, including multiple object instances, significant clutter, and severed occlusions. These objects are not included in the training dataset and are novel to the models. + +Training details. We train SCFlow2 using the AdamW optimizer [30] with a batch size of 16, and use a cosine annealing learning rate scheduler. The model is trained for 200k iterations with an initial learning rate of 1e-4. During training, we generate pseudo initial poses by adding random noise to the ground-truth pose. We add rotation jitter sampled from a normal distribution with a standard deviation of 15 degrees in each 3D axis, and add translation jitter sampled with deviations of 15, 15, and 50 millimeters along xyz axes, respectively. + +Evaluation metrics We report the result in the common BOP metric, which is the mean Average Recall (AR) of three different metrics, including Visible Surface Discrepancy (VSD), Maximum Symmetry-aware Surface Distance (MSSD), and Maximum Symmetry-aware Projection Distance (MSPD). Readers can refer to [15] for further details about these metrics. + +# 4.2. Comparison with the State of the Art + +To explore the plug-and-play nature of our refinement method, we apply it to the final results of Megapose [23], FoundPose [36], GenFlow [32], GigaPose [34], SAM6D [27], and FoundationPose [50], as shown in Table 1. Although these baselines have a refinement strategy themselves, plugging our refinement method into the end of their framework further improves the results, consistently on all datasets. Most baseline results are downloaded from the BOP leaderboards. + +
MethodsDepthLM-OYCB-V
FoundPose (init)-39.645.2
FoundPose (final)✓61.069.0
FoundPose (init) + Ours✓61.381.8
FoundPose (final) + Ours✓69.684.9
GenFlow (init)-25.027.7
GenFlow (final)✓63.583.3
GenFlow (init) + Ours✓64.483.6
GenFlow (final) + Ours✓68.785.5
FoundationPose (init)---
FoundationPose (final)✓75.688.9
FoundationPose (init) + Ours✓79.790.4
FoundationPose (final) + Ours✓80.191.4
+ +Table 2. Effectiveness as a plug-and-play object pose refiner. All the baselines are based on an initialization-refinement architecture. Our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner. + +
L3L5L10L20L30L40L50
Initialization92.783.864.53517.67.83.5
+ SCFlow291.691.591.590.888.683.270.7
+ +# 4.3. Ablation Study + +Effectiveness as a plug-and-play object pose refiner. We evaluate SCFlow2 in more detail in the context of three different state-of-the-art methods, FoundPose [36], GenFlow [32], and FoundationPose [50]. Table 2 summaries the results. All of them are based on an initialization-refinement architecture. We denote their initialization results as "init" in the table, and replace their refinement methods with ours in each method (denoted as "init + ours"). Furthermore, we also report the results of applying our refinement to their final results directly, denoted as "final + ours". Note that, FoundationPose outputs hundreds of pose hypotheses in the initialization stage, and there is no well-defined initialization result for it. In this case, we replace its original refinement method with SCFlow2 in its framework, and keep all other components untouched, including post-processing procedures such as pose scoring and pose selection, which is the reported result of "FoundationPose (init) + Ours" in the table. Overall, our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner. We show some qualitative results in Fig. 5 and 6. + +Table 3. Impact of pose initialization. We add randomly generated rotation and translation noise to GT pose as the initialization, and then use SCFlow2 to refine the pose. L3 to L50 denote different noise levels. SCFlow2 is robust to pose errors in the initialization and works well under noise level 30. + +
DatasetsInit.SeenUnseen
V1V1+V1++V2
LM-O61.062.969.368.969.6
T-LESS57.057.954.955.358.6
TUD-L69.475.988.886.688.1
IC-BIN47.947.954.153.754.6
ITODD40.734.249.148.148.4
HB72.374.576.576.176.2
YCB-V69.080.785.584.484.9
Avg.59.662.068.367.668.6
+ +Table 4. Advantages over SCFlow. We use the same initialization based on results from FoundPose ("Init.", and report results of SCFlow ("V1") and SCFlow2 ("V2"). SCFlow2 outperforms all "V1" versions without any retraining. + +Impact of pose initialization. We add random rotation and translation noise to GT pose as the initialization, and report the results on the YCB-V dataset, as shown in Table 3. L3 to L50 denote different noise levels (rotation error from 3 to 50 degrees, translation error from 3 to $50~\mathrm{mm}$ ). SCFlow2 works well under noise level 30. + +Evaluation of performance with different iterations. We conduct ablation studies on the trade off between the accuracy/efficiency and the iteration numbers of our method, as shown in Fig. 7. Our method produces meaningful pose refinement results after only 2 iterations and reaches saturation after 6 iterations. In our experiments, we use 8 iterations as the default setting in both training and testing. + +Advantages over SCFlow. We compare SCFlow2 and its previous version SCFlow systematically. We summarize the results in Table 4. We use the same initialization based on results from FoundPose, and refine it with SCFlow ("V1") and SCFlow2 ("V2"), respectively. We evaluate three different versions of SCFlow, where "V1" denotes the original version based on retraining on target objects with RGB images only, "V1+" denotes the version using RANSAC Kabsch to consume additional depth images. Both "V1" and "V1+" need retraining on target objects. We also evaluate a version "V1++" which does not rely on retraining for novel objects, but uses the same large-scale datasets as "V2" for training on RGB images and then refine the pose results using RANSAC Kabsch with additional depths. We can see that, without retraining, "V1++" produces worse results compared with "V1+". On the other hand, SCFlow2 even outperforms "V1+" without any retraining. + +Evaluation with different settings. We use the initialization results of FoundPose as our initialization. Table 5 shows the result, where "Shape" denotes the shape-constraint regularization during the recurrent optimization, and "Scene" is the 3D scene flow representation. We use the intermediate flow for the look-up without relying on the + +![](images/8b7028c30c3d5d8ea2d9947bc123ee74a1fb6705470fbfa30a7a15722c91a3b2.jpg) + +![](images/4e49da28d43c17269488796d27168477f7ab0a513a507a903e96fce5506ea62a.jpg) + +![](images/7a7642d12614e45cdd0a9e4d2fa1128957ccecc1b2dc0390af2659ce920f921f.jpg) + +![](images/d5ba2bb65a4bff0bab725fcfbe0cf79b020266aed3b78a6eccf15b8125f98b18.jpg) + +![](images/e7a3a6acd7f7c6bef5825e99725791fa99ee534aab1c348564223d7acfc94faf.jpg) + +![](images/e07ccc5ffb53a64aff587442bbd5a33a56e2edd160831414accce8057d2c7289.jpg) + +![](images/97ccebaf4a43b588bfa22824cf0f63ec473fd93b1120780fab4fa77acb2f26f6.jpg) + +![](images/a4ce9c2d457f94737b8fc86dbd41793765b2687358f650cd8b070e60c518ea43.jpg) + +![](images/607345a095c7a497fac7616da5bdc4f6a26b46723b6604b170f5a2c74d0cbfe0.jpg) + +![](images/d504ef74dcc57f2dddf986ae5b9fecd43ee69686ae45149616d634e0d91b7afe.jpg) + +![](images/635c2bfe4f7fb40a28a26fdfce49c935bf417060a02cdea8c65ee8b52aba139f.jpg) + +![](images/c109ab539f6cd06f3a4e17ee4fdde87c7ba2297bb0d411d91cb4b57f45c934e8.jpg) + +![](images/8e765ca510dd1325d8e9011166b6fdcee7bc93e1072ab0df618f9ea18ba88f1f.jpg) + +![](images/23279438713f379fa872ef124492582d9f90991059264e4222971e8371386615.jpg) + +![](images/c301fc7c2a99f6f672dc0f3fe1debabf8923498594fb57f3cb11615c06034b64.jpg) + +![](images/24e41fbaef3e5ff135ffdc33407da9cd500de10a59e62e94d19900372556106e.jpg) + +![](images/9b8b379dfabd67ed2ef49c1e7b787dada8389ebfc7a4e9640803e9ee98dbb3b5.jpg) + +![](images/b8a0c7719260f15ff5c0b156dfdb93aae41bf0047ca0669a31aded29f67f7f9f.jpg) + +![](images/659dc210b7221b1cbef18fadfd358f8768628a70a581300c0ef799210a095f84.jpg) + +![](images/c67747d55f53117ffbf55f478c95bdd44e25fe514adefb81c98869e28f47441a.jpg) + +![](images/957ac03eef067decf82350030ecd1eeffdc0dea662521e32a3fe934b42637d31.jpg) + +![](images/5ff90ee668a8848f6081688f54389c72330cc939446c58e56c7b15cc75d649cc.jpg) + +![](images/c6cc5653ccee602f61fa0abe7d6c8c3a1aef917868c38a4eb82f17e7f3200208.jpg) + +![](images/60e7eae561a438b81a8beddd50d785ee221b184baa52efe9f3b10e6293f5a852.jpg) + +![](images/2f81ae0963f8fc4d78a853b9925684f6269620215b688ac49f5c6ccbcfe3e49a.jpg) + +![](images/8e0b37b7df406a192939e7b086fabe80b58154deda543c015ee527eb76369b6b.jpg) + +![](images/332da04141ccf188cc9f1f0786c2a5ef2461ab49e4d2453082a1b7d9129ab4f8.jpg) + +![](images/0da0edac4c1e76502e54d9a04598b59cb1399b61ea00d645951c63c21d4e49d5.jpg) + +![](images/f573fef8b980e886d4f8fd8976a8de8dd63d6d0d60ee46e2b5f475c49ae3bbdc.jpg) + +![](images/a2f89dce3be900307788c91a7867ca9a33f3646541e66d1e72e3350a6a0e8bd2.jpg) + +![](images/97e578f7e4fb968a9eeec9700270085e2cc5257cdf0bbb2eb3fe58e82846b58d.jpg) +GFlow (init) + +![](images/404d233ef64ae10c59d59b6b4493846aa9ce662bcd81fba7cab67e6f01137aad.jpg) +GFlow (final) + +![](images/dc2e05ae4a2afc6a00c5acb7f8292429feb91199946a844b0cba183fb409e51c.jpg) +GFlow (init) + Ours + +![](images/78c14e4cb765558c16daf893eba69843bd4b34706b4fced397c1d857f3552075.jpg) +FPose (init) + +![](images/8bca1447470d91eb08343caab2eb956aade89024ae341920a1bef03d3cd941fa.jpg) +FPose (final) +Figure 5. Qualitative results. Given the same pose initialization as that in GenFlow ("GFlow") and FoundPose ("FPose"), denoted as "GFlow (init)" and "FPose (init)" respectively, our refinement method ("+ Ours") produces considerably more accurate results compared to the refinement approaches in their original methods (note how our reprojected 3D mesh aligns better with the object contours). + +![](images/9c983e00021a65622b43e1a7c60b429d5150285c4b212e326f6fda2bcc1c8ac1.jpg) +FPose (init) + Ours + +target's 3D mesh when "Shape" is disabled, and we replace the scene flow representation with a plain MLP pose regressor when "Scene" is disabled. The performance becomes worse without the shape prior embedding, and deteriorates significantly without the 3D scene flow formulation. + +Timing analysis. The training of SCFlow2 needs about 108 hours on a workstation with 2 NVIDIA RTX-3090 GPU. SCFlow2 considers the pose initialization as the only hypothesis. During inference, SCFlow2 takes about 0.18 seconds to refine a pose on an RTX-3090 GPU with a typical + +![](images/c95330b737894cbb359c6010e4603e0a782e5fb10bccb5596732c1c155864bb1.jpg) + +![](images/e740f9664a51697a2cde48f58569336b541ea139ef67a49828b420c294cc6902.jpg) + +![](images/55d3517f07d5b89afb8d24213ce703691f61a6ca04537979f73a80a45aa399f7.jpg) +Initialization + +![](images/b04ffe006eaeee30faac55b6f08f36daaf59efd8be21a8a82765a38b4b5c753d.jpg) ++Ours +Figure 6. More qualitative results. We show two examples on YCB-V. Although the initialization has large pose error, our refinement method produces accurate results that align well with the target, without retraining on these novel objects. + +
ShapeSceneLM-OYCB-V
✓-52.667.6
-✓60.080.4
✓✓61.381.8
+ +Table 5. Evaluation with different settings. "Shape" denotes the shape-constraint regularization during the optimization, and "Scene" is the 3D scene flow formulation. The performance deteriorates significantly without the 3D scene flow formulation. + +image from BOP dataset. Taken for comparison, the multi-hypothesis based method MegaPose needs about 1.85 seconds for a refinement on the same workstation, which is $10+$ times slower than SCFlow2. + +Limitation discussion. In general, our refinement method can improve the accuracy of initial pose significantly. While occasionally, it produces worse results than the initial pose. Fig. 8 shows the histogram of samples along with the relative accuracy improvement after pose refinement on YCB-V and LM-O with the initial pose of FoundPose as our initialization. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases where the initial poses are already very wrong caused by severe occlusion or extreme light conditions. We will take this as one of our future work. + +# 5. Conclusion + +We have introduced SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. We extended + +![](images/053909c232e707e09b3a4a3fa68e70f5b8d018233ae3daba93ef087ff50defcc.jpg) +Figure 7. Evaluation of performance with different iterations. Given the same pose initialization with noise level 30 as in Table 3, more recurrent iterations produce more accurate results while spending more time in inference. Nevertheless, our approach yields meaningful outcomes after just 2 iterations, and the performance reaches saturation after 6 iterations. + +![](images/9855c0de2da3571d78a875667dd53ce61edf8966f336df35e5f8afa7713a4ee9.jpg) + +![](images/051e7db46c7cbed92218312665a01a5baaab189b5ebb60cb75be1bcd2ff23c04.jpg) +Figure 8. Limitation discussion. We show the histogram of samples along with the relative accuracy improvement after pose refinement. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases. + +the RGB-based method SCFlow by formulating additional depth as a regularization in the iteration with an intermediate 3D scene flow representation for RGBD frames, and combined it with the 3D shape prior information of the target, producing a geometry-guided 6D object pose refiner end-to-end trainable and generalizable to novel objects. Experiments on BOP datasets with novel objects show that our method achieves state-of-the-art accuracy, and is one of the most efficient refinement methods. + +Acknowledgments. This work was supported in part by the Youth Innovation Team of Shaanxi Universities, the "Scientist + Engineer" team of the Qin Chuang Yuan in Shaanxi Province, the Xi'an Science and Technology Program, the "Leading the Charge" initiative for the industrialization of core technologies in key industrial chains in Shaanxi Province, the National Nature Science Foundation of China under Grant 62371359, the Key Research and Development Program of Shaanxi, the Key Research Program of the Chinese Academy of Sciences. + +# References + +[1] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D Object Pose Estimation Using 3D Object Coordinates. In European Conference on Computer Vision, 2014. 5 +[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5 +[3] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-End Learnable Geometric Vision by Backpropagating PnP Optimization. In Conference on Computer Vision and Pattern Recognition, 2020. 2 +[4] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv preprint arXiv:1409.1259, 2014. 4 +[5] Javier Cuadrado, Ulysse Rancon, Benoit R Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical Flow Estimation from Event-based Cameras and Spiking Neural Networks. Frontiers in Neuroscience, 17, 2023. 2 +[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A Universe of Annotated 3D Objects. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 5 +[7] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6D Object Pose and Predicting Next-Best-View in the Crowd. In Conference on Computer Vision and Pattern Recognition, 2016. 5 +[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items. In International Conference on Robotics and Automation, 2022. 2, 5 +[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing MVTec ITODDA Dataset for 3D Object Recognition in Industry. In International Conference on Computer Vision Workshops, 2017. 5 +[10] Yang Hai, Rui Song, Jiaojiao Li, David Ferstl, and Yinlin Hu. Pseudo Flow Consistency for Self-Supervised 6D Object Pose Estimation. In International Conference on Computer Vision, 2023. 1 +[11] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-Constraint Flow for 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5 +[12] Rasmus Laurvig Haugaard and Anders Glent Buch. SurfEmb: Dense and Continuous Correspondence Distributions for Object Pose Estimation with Learnt Surface Embeddings. In Conference on Computer Vision and Pattern Recognition, 2022. 1 +[13] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. FS6D: Few-Shot 6D Pose Estimation of Novel Ob + +jects. In Conference on Computer Vision and Pattern Recognition, 2022. 2 +[14] Tomás Hodań, Pavel Haluza, Štepan Obdržánek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-Less Objects. In Winter Conference on Applications of Computer Vision, 2017. 5 +[15] Tomáš Hodař, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In European Conference on Computer Vision, 2018. 2, 5 +[16] Tomás Hodań, Daniel Barath, and Jérí Matas. EPOS: Estimating 6D Pose of Objects with Symmetries. In Conference on Computer Vision and Pattern Recognition, 2020. 2 +[17] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2 +[18] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-Stage 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2 +[19] Yinlin Hu, Sebastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-Depth-Range 6D Object Pose Estimation in Space. In Conference on Computer Vision and Pattern Recognition, 2021. 2 +[20] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective Flow Aggregation for Data-Limited 6D Object Pose Estimation. In European Conference on Computer Vision, 2022. 1, 2 +[21] Wolfgang Kabsch. A Solution for the Best Rotation to Relate Two Sets of Vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 1, 2 +[22] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D Dataset for 6D Pose Estimation of 3D Objects. In International Conference on Computer Vision Workshops, 2019. 5 +[23] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In Conference on Robot Learning. 1, 2, 5 +[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent Multi-View Multi-Object 6d Pose Estimation. In European Conference on Computer Vision, 2020. 1 +[25] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep Iterative Matching for 6D Pose Estimation. In European Conference on Computer Vision, 2018. 4 +[26] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In International Conference on Computer Vision, 2019. 2 +[27] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5 + +[28] Yongliang Lin, Yongzhi Su, Praveen Nathan, Sandeep Inuganti, Yan Di, Martin Sundermeyer, Fabian Manhardt, Didier Stricker, Jason Rambach, and Yu Zhang. HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2 +[29] Fulin Liu, Yinlin Hu, and Mathieu Salzmann. Linear-Covariance Loss for End-to-End Learning of 6D Pose Estimation. In International Conference on Computer Vision, 2023. 2 +[30] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In International Conference on Learning Representations, 2019. 5 +[31] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose Estimation for Augmented Reality: A Hands-On Survey. IEEE transactions on visualization and computer graphics, 22(12):2633-2651, 2015. 1 +[32] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6 +[33] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. NOPE: Novel Object Pose Estimation from a Single Image. In Conference on Computer Vision and Pattern Recognition, 2024. 2 +[34] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5 +[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOV2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 3 +[36] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found-Pose: Unseen Object Pose Estimation with Foundation Features. European Conference on Computer Vision, 2024. 1, 2, 5, 6 +[37] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. LatentFusion: End-to-End Differentiable Reconstruction and Rendering for Unseen Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2 +[38] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-Wise Voting Network for 6DoF Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2019. 2 +[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. arXiv preprint arXiv:1706.02413, 2017. 3 +[40] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects without Using Depth. In International Conference on Computer Vision, 2017. 2 + +[41] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[42] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A Multi-Stage One Shot Object Pose Estimation Framework. In Conference on Computer Vision and Pattern Recognition, 2022. 1 +[43] Chen Song, Jiaru Song, and Qixing Huang. HybridPose: 6D Object Pose Estimation under Hybrid Representations. In Conference on Computer Vision and Pattern Recognition, 2020. 2 +[44] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation. arXiv preprint arXiv:2203.09418, 2022. 2 +[45] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-Shot Object Pose Estimation without CAD Models. In Conference on Computer Vision and Pattern Recognition, 2022. 2 +[46] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In European Conference on Computer Vision, 2020. 2, 4 +[47] Zachary Teed and Jia Deng. RAFT-3D: Scene Flow using Rigid-Motion Embeddings. In Conference on Computer Vision and Pattern Recognition, 2021. 2, 3, 4 +[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In Conference on Computer Vision and Pattern Recognition, 2019. 1, 3 +[49] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 5 +[50] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6 +[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes. In Robotics: Science and Systems, 2018. 5 +[52] Zhaolin Xiao, Zhan Su, Fengyuan Zuo, and Haiyan Jin. Low-Light Optical Flow Estimation with Hidden Feature Supervision Using a Siamese Network. Journal of Image and Graphics, 29:0231-0242, 2024. 2 +[53] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In International Conference on Computer Vision, 2019. 2 +[54] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Conference on Computer Vision and Pattern Recognition, 2019. 4 \ No newline at end of file diff --git a/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/images.zip b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42e9395d69cf85940447e229ecaec95c4315c7ad --- /dev/null +++ b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4acc41ff79d0c19af9c1661ca4f4db729d8ea50cf4214b53a25a2ce367d30794 +size 847784 diff --git a/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/layout.json b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..944a872a1032341adb108fd6d11365240efa8139 --- /dev/null +++ b/2025/SCFlow2_ Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow/layout.json @@ -0,0 +1,10388 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "type": "text", + "content": "SCFlow2: Plug-and-Play Object Pose Refiner with Shape-Constraint Scene Flow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": "Qingyuan Wang " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": ", Rui Song " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": ", Jiaojiao Li " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": ", Kerui Cheng " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": ", David Ferstl " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": ", Yinlin Hu " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": " State Key Laboratory of ISN, Xidian University " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": " Taiyuan University of Technology " + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 142, + 545, + 174 + ], + "type": "text", + "content": " MagicLeap" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 224, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 296, + 452 + ], + "type": "text", + "content": "We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. Most recent 6D object pose methods rely on refinement to get accurate results. However, most existing refinement methods either suffer from noises in establishing correspondences, or rely on retraining for novel objects. SCFlow2 is based on the SCFlow model designed for refinement with shape constraint, but formulates the additional depth as a regularization in the iteration via 3D scene flow for RGBD frames. The key design of SCFlow2 is an introduction of geometry constraints into the training of recurrent matching network, by combining the rigid-motion embeddings in 3D scene flow and 3D shape prior of the target. We train SCFlow2 on a combination of dataset Objaverse, GSO and ShapeNet, and evaluate on BOP datasets with novel objects. After using our method as a post-processing, most state-of-the-art methods produce significantly better results, without any retraining or fine-tuning. The source code is available at https://scflow2.github.io." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 476, + 135, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 135, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 135, + 488 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 544 + ], + "type": "text", + "content": "6D object pose estimation, which determines an object's 3D rotation and 3D translation relative to a camera, is crucial in fields like robotics, augmented reality, and spatial computing [31, 48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 676 + ], + "type": "text", + "content": "Most existing pose estimation methods rely on a refinement procedure to get accurate pose results [10, 12, 20, 23, 24, 48, 50], and most of the refinement methods are based on render-and-compare strategies [10, 20, 23, 50]. A group of refinement methods [24, 48] achieves highly accurate results based on retraining or fine-tuning the network for novel objects, which, however, is cumbersome and unfriendly to real applications. Some recent refinement methods [23, 42, 50] have shown promising results without fine-tuning for novel objects. However, they still suffer from several weaknesses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "First, most methods formulate the render-and-compare framework as a general matching problem, but there is a strong prior of the target's 3D shape that is not used, mak" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 198, + 385, + 266 + ], + "blocks": [ + { + "bbox": [ + 317, + 198, + 385, + 266 + ], + "lines": [ + { + "bbox": [ + 317, + 198, + 385, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 385, + 266 + ], + "type": "image", + "image_path": "2d9e7353265abbc3cc39e2b8711d1dae8cdd8e26005018cb014ec4688486c7a8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 267, + 375, + 277 + ], + "lines": [ + { + "bbox": [ + 327, + 267, + 375, + 277 + ], + "spans": [ + { + "bbox": [ + 327, + 267, + 375, + 277 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 386, + 201, + 484, + 266 + ], + "blocks": [ + { + "bbox": [ + 386, + 201, + 484, + 266 + ], + "lines": [ + { + "bbox": [ + 386, + 201, + 484, + 266 + ], + "spans": [ + { + "bbox": [ + 386, + 201, + 484, + 266 + ], + "type": "image", + "image_path": "39e9f1a017519c747248789e2f543723571f1a48d63fcda0fff46d8f36974daa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 266, + 452, + 274 + ], + "lines": [ + { + "bbox": [ + 419, + 266, + 452, + 274 + ], + "spans": [ + { + "bbox": [ + 419, + 266, + 452, + 274 + ], + "type": "text", + "content": "Network" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 486, + 198, + 555, + 266 + ], + "blocks": [ + { + "bbox": [ + 486, + 198, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 486, + 198, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 486, + 198, + 555, + 266 + ], + "type": "image", + "image_path": "fc9a0acc4c70262982e7ed8cbbfe9f28059af475cb02e8082fa5569f12da9f32.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 267, + 542, + 277 + ], + "lines": [ + { + "bbox": [ + 498, + 267, + 542, + 277 + ], + "spans": [ + { + "bbox": [ + 498, + 267, + 542, + 277 + ], + "type": "text", + "content": "Refinement" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 293, + 385, + 362 + ], + "blocks": [ + { + "bbox": [ + 359, + 282, + 511, + 293 + ], + "lines": [ + { + "bbox": [ + 359, + 282, + 511, + 293 + ], + "spans": [ + { + "bbox": [ + 359, + 282, + 511, + 293 + ], + "type": "text", + "content": "(a) Object pose refinement with retraining" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 293, + 385, + 362 + ], + "lines": [ + { + "bbox": [ + 317, + 293, + 385, + 362 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 385, + 362 + ], + "type": "image", + "image_path": "e26ec74bbd33569cd6465a71ed55909bf8fb3c661efcfd670244a606f7e67ed6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 363, + 374, + 372 + ], + "lines": [ + { + "bbox": [ + 327, + 363, + 374, + 372 + ], + "spans": [ + { + "bbox": [ + 327, + 363, + 374, + 372 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 377, + 487, + 388 + ], + "lines": [ + { + "bbox": [ + 384, + 377, + 487, + 388 + ], + "spans": [ + { + "bbox": [ + 384, + 377, + 487, + 388 + ], + "type": "text", + "content": "(b) Plug-and-play SCFlow2" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 389, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 456 + ], + "type": "text", + "content": "Figure 1. Object pose refinement is critical for accurate object pose estimation. (a) Most existing object pose refinement methods, including SCFlow [11], rely on retraining for novel objects to achieve high accuracy. (b) The proposed SCFlow2 achieves even higher accuracy, and more importantly, generalizes well to novel objects without any retraining or fine-tuning." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 386, + 320, + 482, + 358 + ], + "blocks": [ + { + "bbox": [ + 386, + 320, + 482, + 358 + ], + "lines": [ + { + "bbox": [ + 386, + 320, + 482, + 358 + ], + "spans": [ + { + "bbox": [ + 386, + 320, + 482, + 358 + ], + "type": "image", + "image_path": "96dfa597e0a6fed791cd94d81be963687b6d2c25b6700f88b0a048052197a73a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 486, + 293, + 555, + 361 + ], + "blocks": [ + { + "bbox": [ + 419, + 362, + 486, + 371 + ], + "lines": [ + { + "bbox": [ + 419, + 362, + 486, + 371 + ], + "spans": [ + { + "bbox": [ + 419, + 362, + 486, + 371 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 486, + 293, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 486, + 293, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 486, + 293, + 555, + 361 + ], + "type": "image", + "image_path": "9e2c47419cf7f86407f32c13c84a98b02450e63278a95d4905b30a135ec27a0d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 362, + 542, + 371 + ], + "lines": [ + { + "bbox": [ + 498, + 362, + 542, + 371 + ], + "spans": [ + { + "bbox": [ + 498, + 362, + 542, + 371 + ], + "type": "text", + "content": "Refinement" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "content": "ing it usually has an unnecessarily large search space for matching and is less effective for 6D object pose estimation [11]. Second, to address the matching issue, most methods attempt to refine the pose starting from multiple pose hypotheses based on the initialization [23, 32, 36]. While this paradigm is effective, it significantly slows down the refinement process. Third, most methods formulate the comparison after the rendering as a 2D matching procedure only, and then use ICP or Kabsch [21] to consume additional depth input as a second stage, which is less effective in capturing 3D motions and local optimal within each stage." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "We introduce SCFlow2, a plug-and-play refinement framework for 6D object pose estimation, as shown in Fig. 1. SCFlow2 is based on SCFlow [11] which is a state-of-the-art refinement method embedding 3D shape constraints into the optimization for known objects. We build on the basic shape-constraint structure of SCFlow but extend it in several novel ways." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22045" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 318, + 164 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 318, + 164 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 318, + 164 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 318, + 164 + ], + "type": "image", + "image_path": "9952a3a3f8f062e595a961342928d43ba8b64f83b476d0db771ae47febb7d09c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 170, + 249, + 179 + ], + "lines": [ + { + "bbox": [ + 138, + 170, + 249, + 179 + ], + "spans": [ + { + "bbox": [ + 138, + 170, + 249, + 179 + ], + "type": "text", + "content": "(a) The framework of SCFlow" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 328, + 73, + 557, + 165 + ], + "blocks": [ + { + "bbox": [ + 328, + 73, + 557, + 165 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 557, + 165 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 557, + 165 + ], + "type": "image", + "image_path": "89961d668d373cd6376f0275c610c970361917171d098a01020651d5bf062d97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 170, + 509, + 179 + ], + "lines": [ + { + "bbox": [ + 392, + 170, + 509, + 179 + ], + "spans": [ + { + "bbox": [ + 392, + 170, + 509, + 179 + ], + "type": "text", + "content": "(b) The framework of SCFlow2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "text", + "content": "Figure 2. Design overview of SCFlow and SCFlow2. Given the object 3D mesh, we render an image " + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "text", + "content": " and depth map " + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "text", + "content": " based on an initial pose, and then use networks to compare these rendered outputs with the real input " + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "inline_equation", + "content": "D_{2}" + }, + { + "bbox": [ + 55, + 185, + 555, + 274 + ], + "type": "text", + "content": " to refine the pose. (a) Although SCFlow [11] adds 3D shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, which is less effective in capturing 3D motions. On the other hand, it cannot work with RGBD images. A common practice is to use RANSAC Kabsch [21] to consume additional depth as a second stage, which however is only local optimal within each stage. (b) SCFlow2 tackles these problems. We introduce an intermediate representation based on 3D scene flow to capture 3D motions in network optimization. Furthermore, we embed depth into the loop by formulating depth as an additional regularization to guide the correlation look-up iteratively, producing an end-to-end trainable system with RGBD images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 294, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 295, + 450 + ], + "type": "text", + "content": "Our main design of SCFlow2 is introducing a scene flow based representation into the iterative optimization. Although SCFlow embeds shape constraint into the optimization loop, it formulates the matching process as a pure 2D problem, and suffers in the solution with a second stage of RANSAC Kabsch [21] for RGBD images. On the other hand, SCFlow cannot work with novel objects. To handle these problems, we introduce rigid-motion embeddings in 3D scene flow [47] as our intermediate representation, and formulate depth as an additional regularization to guide the network optimization iteratively, which helps reducing search space and results in an end-to-end system that generalizes to novel objects, as illustrated in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 535 + ], + "type": "text", + "content": "We train SCFlow2 on a combination of dataset Objverse [6], GSO [8] and ShapeNet [2], and evaluate on BOP datasets [15] with novel objects. SCFlow2 achieves state-of-the-art accuracy, and can produce accurate results with only one pose hypothesis, outperforming most multiple-hypothesis based refinement methods [23, 32, 36] in both accuracy and efficiency, without any retraining." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 142, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 142, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 142, + 560 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Object pose estimation aims to predict the 3D rotation and 3D translation with respect to the camera, and has shown prominent progress with deep learning techniques [13, 27, 28, 33, 37, 45]. Most existing methods assume the existence of many real images of the target, and train a learning model based on those real images [11, 19, 20, 38, 44], based on either 2D-to-3D matching [16, 26, 53], 3D-to-2D matching [19, 38, 40], or end-to-end training with differentiable perspective-n-point (PnP) [3, 29, 43]. This type of method is accurate on the trained target in general, which however, cannot work with novel objects that were not involved in the training. Retraining or fine-tuning can adapt the net" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 294, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 498 + ], + "type": "text", + "content": "work to novel objects, which, however, is cumbersome and limits its applications in real scenarios. Some recent methods show promising results for novel object pose estimation [23, 32, 34, 36, 50], and one of their key components is object pose refinement, which is based on render-and compare [23, 32, 50] or feature matching [34, 36] strategies. Although these refinement methods work in general, they are less effective at utilizing prior shape information of the target. The recent method SCFlow [11] introduces a shape-constraint recurrent matching framework for object pose estimation, which embeds the target's 3D shape into the optimization. It produces accurate results for known objects, while has limited generalization ability to novel objects. SCFlow2 is built on top of the basic shape-constraint structure of SCFlow, but extends it by introducing 3D scene flow representation into the network optimization, producing an end-to-end system with RGBD images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Optical flow estimation is a fundamental computer vision task to estimate pixel-level correspondences between two frames [5, 17, 41, 52]. Most recent methods are based on deep learning models and achieve accurate results in general scenes [5, 41, 46]. The recent method RAFT [46] presents a recurrent framework for flow estimation, and RAFT-3D [47] extends it to estimate 3D scene flow with rigid motion embeddings. Although they show significant advantages over existing flow methods, estimating dense matching with these general flow methods for object pose estimation is less effective, without using the prior 3D shape information of the target. Furthermore, the general flow loss is a surrogate matching loss that does not directly reflect the final 6D object pose [11, 18, 20, 43] in the context of object pose estimation. We combine the shape-constraint formulation [11] and the rigid motion embeddings in 3D scene flow, producing a geometry-guided 6D object pose refiner that is end-to-end trainable and generalizes to novel objects." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22046" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 73, + 558, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 558, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 558, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 558, + 220 + ], + "type": "image", + "image_path": "af008dda54be056434a6663231d3a00e3944b3a311335ebbb42f830665f96048.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "text", + "content": "Figure 3. Overview of SCFlow2. Given an RGBD image and an initial pose of the target, we first render the target to obtain a synthetic RGBD image as the reference, and use an RGB encoder and depth encoder to extract features from the image pair, which will be used to create a 4D correlation volume. Based on the correlation volume and GRU, We use an intermediate flow regressor to predict 3D scene flow that is represented as a dense 3D transformation field " + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k^{\\prime}" + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "text", + "content": ". We then use a pose regressor to predict a global pose update " + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P}_k" + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "text", + "content": " based on implicit voting from the pixel-wise 3D transformation field. Finally, the updated pose " + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "text", + "content": " is used to compute the pose-induced flow " + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 54, + 232, + 555, + 312 + ], + "type": "text", + "content": " based on the target mesh to index the correlation volume for the next iteration. Note how the depth and 3D shape of the target are embedded into the framework to guide the optimization iteratively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 330, + 111, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 111, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 111, + 341 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "type": "text", + "content": "Given an RGBD image and the 3D mesh of the object, our goal is to estimate the 6D pose of the target object. We assume the camera intrinsic matrix is known. We first introduce the framework at a high level in Section 3.1, and then present the design of the main components in Sections 3.2 and 3.3. We finally discuss the implementation details in Section 3.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 123, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 123, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 123, + 462 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 469, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 518 + ], + "type": "text", + "content": "In general, we first render an RGBD image based on an initial pose of the target as a reference, and then predict the pose update based on comparing the rendered image and the real input with a recurrent matching network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "type": "text", + "content": "With the RGBD image pair, we first use an RGB encoder and depth encoder to extract features, and then construct a 4D correlation volume based on computing the dot product of the fused RGBD features. Based on the constructed correlation volume, we propose to use an intermediate flow regressor to predict 3D scene flow that is represented as a dense SE3 motion field [47]. We then use a pose regressor to predict a global pose update based on implicit voting from the pixel-wise SE3 motion field. Finally, the updated pose is used to compute a pose-induced flow based on the target mesh to index the correlation volume for the next iteration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The key of our method is the combination of the dense 3D scene flow representation and the embedding of the target 3D shape prior [11] in a unified recurrent matching framework. Fig. 3 shows the overview of this framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 493, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 493, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 493, + 342 + ], + "type": "text", + "content": "3.2. Correlation Volume Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": "We construct a 4D volume pyramid to represent the correlation between any two locations on the two input images. We use an RGB encoder and a depth encoder to extract features from the RGB and the depth image, respectively. For the rendered image " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": " based on the initial pose and the input RGB image " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": ", we use an encoder with shared parameters to obtain their dense feature maps " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "f_{1}^{i} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "f_{2}^{i} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": " of the original resolution. We use DINOv2 ViT-B [35] as our RGB feature extractor in the encoder, and freeze the pretrained weights during training. For the depths " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "inline_equation", + "content": "D_{2}" + }, + { + "bbox": [ + 313, + 348, + 555, + 516 + ], + "type": "text", + "content": ", we first lift the two depth maps to point clouds according to the camera intrinsic matrix, and then extract their 3D structural features with PointNet++ [39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "type": "text", + "content": "Similar to DenseFusion [48], we fuse the RGB features and depth features to generate semantic and contextual features. We then compute the 4D correlation volume via computing the dot product of the fused features according to different location pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 586, + 507, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 507, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 507, + 599 + ], + "type": "text", + "content": "3.3. Regressing 3D Transformation Field" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 604, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 628 + ], + "type": "text", + "content": "With a dense matching field as the index, we can extract a 3D correlation map from the 4D correlation volume." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "In SCFlow [11], the correlation map is used to predict the optical flow " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "f(u,v)" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": ", which maps the pixel position " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1 = (u,v)" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": " to the corresponding position in " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_2 = (u,v) + f(u,v)" + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": ". However, this 2D flow formulation only contains 2D information, which is less effective in capturing 3D motions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "content": "To address this, we propose to predict 3D scene flow [47]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22047" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 129, + 215 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 129, + 215 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 129, + 215 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 129, + 215 + ], + "type": "image", + "image_path": "f7caf4371b06dca9632e0d2c34e391d8d2f735912e0e1466c7345a1fc4d954e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 217, + 104, + 228 + ], + "lines": [ + { + "bbox": [ + 83, + 217, + 104, + 228 + ], + "spans": [ + { + "bbox": [ + 83, + 217, + 104, + 228 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 135, + 77, + 196, + 215 + ], + "blocks": [ + { + "bbox": [ + 135, + 77, + 196, + 215 + ], + "lines": [ + { + "bbox": [ + 135, + 77, + 196, + 215 + ], + "spans": [ + { + "bbox": [ + 135, + 77, + 196, + 215 + ], + "type": "image", + "image_path": "4eb94e7619a549af88556663cf2cedcae0c20320fcb868d6775351fb556cd580.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 217, + 186, + 228 + ], + "lines": [ + { + "bbox": [ + 143, + 217, + 186, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 186, + 228 + ], + "type": "text", + "content": "Initial pose" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 77, + 267, + 216 + ], + "blocks": [ + { + "bbox": [ + 206, + 77, + 267, + 216 + ], + "lines": [ + { + "bbox": [ + 206, + 77, + 267, + 216 + ], + "spans": [ + { + "bbox": [ + 206, + 77, + 267, + 216 + ], + "type": "image", + "image_path": "5a85df86751bd2e885784022a1c176fdf79ad41166f2492361758b2385f5a697.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 218, + 239, + 227 + ], + "lines": [ + { + "bbox": [ + 233, + 218, + 239, + 227 + ], + "spans": [ + { + "bbox": [ + 233, + 218, + 239, + 227 + ], + "type": "text", + "content": "T" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 278, + 77, + 338, + 215 + ], + "blocks": [ + { + "bbox": [ + 278, + 77, + 338, + 215 + ], + "lines": [ + { + "bbox": [ + 278, + 77, + 338, + 215 + ], + "spans": [ + { + "bbox": [ + 278, + 77, + 338, + 215 + ], + "type": "image", + "image_path": "1ee6d55a9d7d9eab885f3b2ab3b4413a91b393670f436f4873ddbe616223c415.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 218, + 310, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 218, + 310, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 310, + 228 + ], + "type": "inline_equation", + "content": "\\phi" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 238, + 555, + 294 + ], + "lines": [ + { + "bbox": [ + 54, + 238, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 555, + 294 + ], + "type": "text", + "content": "Figure 4. Visualization of pose-induced flow with 3D scene flow representation. Given a real input image and the rendered image based on an initial pose, we predict 3D scene flow represented as a dense SE3 motion field. We represent the motion field as a twist field " + }, + { + "bbox": [ + 54, + 238, + 555, + 294 + ], + "type": "inline_equation", + "content": "(\\tau, \\theta)" + }, + { + "bbox": [ + 54, + 238, + 555, + 294 + ], + "type": "text", + "content": ". In theory, the motion fields should be constant pixels for rigid objects. We use a pose regressor to predict a global object-level pose based on the noisy motion fields. We then use the updated global pose to generate a pose-induced flow based on the target 3D mesh. The pose-induced flow embeds the shape prior of the target and reduces the search space for matching." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 349, + 77, + 410, + 215 + ], + "blocks": [ + { + "bbox": [ + 349, + 77, + 410, + 215 + ], + "lines": [ + { + "bbox": [ + 349, + 77, + 410, + 215 + ], + "spans": [ + { + "bbox": [ + 349, + 77, + 410, + 215 + ], + "type": "image", + "image_path": "67fc95840ac40972906fc606e7a5ff0308f9633baea68d51a0f9f72aaa418c13.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 217, + 411, + 227 + ], + "lines": [ + { + "bbox": [ + 346, + 217, + 411, + 227 + ], + "spans": [ + { + "bbox": [ + 346, + 217, + 411, + 227 + ], + "type": "text", + "content": "Intermediate flow" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 421, + 77, + 480, + 215 + ], + "blocks": [ + { + "bbox": [ + 421, + 77, + 480, + 215 + ], + "lines": [ + { + "bbox": [ + 421, + 77, + 480, + 215 + ], + "spans": [ + { + "bbox": [ + 421, + 77, + 480, + 215 + ], + "type": "image", + "image_path": "7df50d2dc89c1441b0f5f0b8ba3b23a982951c004e0534a1eecc122df6844d6c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 217, + 484, + 227 + ], + "lines": [ + { + "bbox": [ + 415, + 217, + 484, + 227 + ], + "spans": [ + { + "bbox": [ + 415, + 217, + 484, + 227 + ], + "type": "text", + "content": "Pose-induced flow" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 557, + 215 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 557, + 215 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 557, + 215 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 557, + 215 + ], + "type": "image", + "image_path": "27b035c8982fe8e4555ea73464bc517e9cf36ab8bbf3f047db07848364c86a53.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 497, + 218, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 497, + 218, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 497, + 218, + 545, + 228 + ], + "type": "text", + "content": "Refined pose" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "content": "as an intermediate representation for the network. Given " + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "content": " and their corresponding point clouds " + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "content": ", we formulate scene flow as a dense 3D transformation field " + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in SE(3)^{H \\times W}" + }, + { + "bbox": [ + 54, + 315, + 295, + 374 + ], + "type": "text", + "content": " to represent the 3D motion between the two point clouds in 3D space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": "Specifically, for the point " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": " with index " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": " and its projected pixel coordinates " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = (u_i,v_i,z_i)" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": ", we can perform the transformation " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i\\cdot X_i" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": " to obtain the corresponding point " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "X_{i}^{\\prime}" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": " and its projected pixel coordinates " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i' = (u_i',v_i',z_i')" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": ". The scene flow vector is defined as the difference " + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = \\mathbf{x}_i' - \\mathbf{x}_i" + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": ", where the first two components represent the standard 2D optical flow and the last component describes the depth difference between the two points." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 470, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 470, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 470, + 295, + 506 + ], + "type": "text", + "content": "We use GRU [4] as the recurrent model for the intermediate scene flow predictor. In the " + }, + { + "bbox": [ + 54, + 470, + 295, + 506 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 54, + 470, + 295, + 506 + ], + "type": "text", + "content": " iteration, the hidden state update can be expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 92, + 514, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 514, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 92, + 514, + 295, + 526 + ], + "type": "interline_equation", + "content": "h _ {k} = \\operatorname {G R U} \\left(\\mathbf {L} _ {\\mathbf {C}} \\left(F _ {k - 1}\\right), \\mathbf {T} _ {k - 1}, h _ {k - 1}; \\Theta\\right) \\tag {1}", + "image_path": "aedc799d28a631621a20770a42c5720682ed49965b2a6e86fa4b91d11b459a36.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathbf{C}}(F_{k - 1})" + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": " represents the correlation look-up operation guided by the global flow " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "F_{k - 1}" + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "(k - 1)^{th}" + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": " iteration, " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{k - 1}" + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": " is the dense 3D transformation field from the previous iteration, " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": " denotes the GRU network parameters, and " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "text", + "content": " is the hidden state feature which will be fed into a dense SE3 layer [47] to update the SE3 motion fields and generate a new transformation field " + }, + { + "bbox": [ + 54, + 533, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k^\\prime" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 624, + 222, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 222, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 222, + 636 + ], + "type": "text", + "content": "3.4. Regressing Global Object Pose" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "text", + "content": "Our goal is to predict the pose residual of the single object in " + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "text", + "content": ". Considering the rigidity of objects in 6D pose estimation scenarios, all pixel-level 3D motions in " + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k^\\prime" + }, + { + "bbox": [ + 54, + 641, + 295, + 713 + ], + "type": "text", + "content": " theoretically describe the same rigid object and should be the same. However, noise is inevitable in predicting the dense transformation field. Therefore, we designed a small" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "content": "network to predict the global pose residual using the dense transformation field as input, as shown in Fig. 4. First, we represent the dense transformation field " + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k^{\\prime}" + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "content": " as a " + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "inline_equation", + "content": "4\\times 4" + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "content": " transformation matrix, and then we encode it using a three-layer 2D convolutional network. Finally, we use two fully connected layers to output the global pose residual " + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P}_k" + }, + { + "bbox": [ + 313, + 315, + 555, + 411 + ], + "type": "text", + "content": ", which is represented as a 9-dim vector including 6-min for the rotation [54] and 3-dim for the normalized translation [25]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": "In the " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " iteration, the pose residual " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{P}_k" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " will be used for the pose update: " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k = \\mathbf{P}_{k - 1} \\otimes \\Delta \\mathbf{P}_k" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": ", and the updated pose " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " will be used to compute the pose-induced flow based on the object mesh, which imposes shape constraints for the lookup operation in the next iteration [11]. At the same time, we will generate a new dense transformation field " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " based on the updated pose " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": ". Specifically, we calculate the pose residual between the initial pose " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_0 = [\\mathbf{R}_0|\\mathbf{t}_0]" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_k = [\\mathbf{R}_k|\\mathbf{t}_k]" + }, + { + "bbox": [ + 313, + 411, + 556, + 521 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 345, + 533, + 553, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 533, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 345, + 533, + 553, + 546 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {R} _ {k} = \\mathbf {R} _ {k} \\cdot \\mathbf {R} _ {0}, \\quad \\Delta \\mathbf {t} _ {k} = \\mathbf {t} _ {k} - \\Delta \\mathbf {R} _ {k} \\cdot \\mathbf {t} _ {0} \\tag {2}", + "image_path": "6fca03279f746c9d332837055ab3d5cf61f75ab001d115573ba9c8ff5c04f445.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 558, + 555, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 607 + ], + "type": "text", + "content": "Then, we update the 3D motion field " + }, + { + "bbox": [ + 313, + 558, + 555, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_k\\in SE(3)^{H\\times W}" + }, + { + "bbox": [ + 313, + 558, + 555, + 607 + ], + "type": "text", + "content": ", where the motion at every location will be updated as the same new pose residual. This new motion field will be the input for the next iteration." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 620, + 447, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 447, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 447, + 633 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "text", + "content": "For the input RGB and depth images " + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "inline_equation", + "content": "D_{2}" + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "text", + "content": ", we first crop the target out based on the initial pose, and then resize it to a fixed resolution of " + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 639, + 555, + 675 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Given the ground truth pose, we can compute the ground truth optical flow based on the object mesh. We use the exponentially weighted strategy [11, 46] to calculate the loss" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22048" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 506, + 251 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 506, + 251 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 506, + 251 + ], + "type": "table", + "html": "
MethodsLM-OT-LESSTUD-LIC-BINITODDHBYCB-VAvg.
MegaPose [23]62.048.584.646.246.072.576.462.3
MegaPose + Ours62.454.988.348.449.576.479.065.6
FoundPose [36]61.057.069.447.940.772.369.059.6
FoundPose + Ours69.658.688.154.648.476.284.968.6
GenFlow [32]63.552.186.253.455.477.983.367.4
GenFlow + Ours68.753.788.055.456.479.285.569.6
Gigapose [34]67.855.681.156.357.579.182.568.6
Gigapose + Ours71.058.682.057.758.380.184.470.3
SAM6D [27]69.951.590.458.860.277.684.570.4
SAM6D + Ours72.152.390.860.060.577.785.171.2
FoundationPose [50]75.664.692.350.858.083.588.973.4
FoundationPose + Ours80.165.095.551.758.384.491.475.2
", + "image_path": "afb3116f36e3cf0be29d9d7c52dd1d2a920952f02a650aaec17dd54fb33653c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 260, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 260, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 555, + 293 + ], + "type": "text", + "content": "Table 1. Comparison with the state of the art. All the baselines have their own refinement strategy, and we report their final result published on BOP leaderboard. After using our method as a post-processing (“+ Ours”), the result consistently improves across all baselines on all datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 315, + 253, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 253, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 253, + 327 + ], + "type": "text", + "content": "in each iteration for both flow loss and pose loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 333, + 295, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 333, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 295, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {k = 1} ^ {N} \\gamma^ {N - k} \\left(\\mathcal {L} _ {\\text {p o s e}} ^ {k} + \\alpha \\mathcal {L} _ {\\text {f l o w}} ^ {k}\\right) \\tag {3}", + "image_path": "ac410638bff376224c432671d9b7f97211f1bad71032b88a23e18b374c37ccf4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "content": "where the weight " + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\gamma = 0.8" + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flow}" + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "content": " represents the L1 loss between the first two dimensions of the predicted scene flow and the ground truth optical flow. " + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pose}" + }, + { + "bbox": [ + 55, + 372, + 296, + 445 + ], + "type": "text", + "content": " refers to pose loss, which is computed as the L1 distance between the 3D point clouds transformed by the predicted pose and the ground truth pose, respectively [11, 49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 454, + 137, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 137, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 137, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 558 + ], + "type": "text", + "content": "We evaluate SCFlow2 on seven challenging BOP benchmark datasets in this section. Section 4.1 presents our experimental setup, including the datasets used for training and evaluation, training details and the evaluation metrics. In Section 4.2, we compare our approach with current state-of-the-art pose methods. Section 4.3 provides a detailed ablation study of the proposed framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 182, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 182, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 182, + 577 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": "Dataset. We train SCFlow2 on a combination of datasets including ShapeNet-Objects [2], Google-Scanned-Objects [8], and Objverse [6]. All these datasets only contain object meshes, in a total of about 90K objects. We reuse the synthetic images from [23] and [50] which are rendered based on those meshes, with about 3M images in total." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "We evaluate on seven datasets from the BOP benchmark [15], including LM-O [1], T-LESS [14], TUD-L [15], IC-BIN [7], ITODD [9], HB [22], and YCB-V [51]. These datasets include 108 objects having different sizes and different texture and symmetry properties. Most of them are" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 315, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 361 + ], + "type": "text", + "content": "captured in challenging setups, including multiple object instances, significant clutter, and severed occlusions. These objects are not included in the training dataset and are novel to the models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 484 + ], + "type": "text", + "content": "Training details. We train SCFlow2 using the AdamW optimizer [30] with a batch size of 16, and use a cosine annealing learning rate scheduler. The model is trained for 200k iterations with an initial learning rate of 1e-4. During training, we generate pseudo initial poses by adding random noise to the ground-truth pose. We add rotation jitter sampled from a normal distribution with a standard deviation of 15 degrees in each 3D axis, and add translation jitter sampled with deviations of 15, 15, and 50 millimeters along xyz axes, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 487, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 570 + ], + "type": "text", + "content": "Evaluation metrics We report the result in the common BOP metric, which is the mean Average Recall (AR) of three different metrics, including Visible Surface Discrepancy (VSD), Maximum Symmetry-aware Surface Distance (MSSD), and Maximum Symmetry-aware Projection Distance (MSPD). Readers can refer to [15] for further details about these metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 511, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 511, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 511, + 597 + ], + "type": "text", + "content": "4.2. Comparison with the State of the Art" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "To explore the plug-and-play nature of our refinement method, we apply it to the final results of Megapose [23], FoundPose [36], GenFlow [32], GigaPose [34], SAM6D [27], and FoundationPose [50], as shown in Table 1. Although these baselines have a refinement strategy themselves, plugging our refinement method into the end of their framework further improves the results, consistently on all datasets. Most baseline results are downloaded from the BOP leaderboards." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22049" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 234 + ], + "type": "table", + "html": "
MethodsDepthLM-OYCB-V
FoundPose (init)-39.645.2
FoundPose (final)✓61.069.0
FoundPose (init) + Ours✓61.381.8
FoundPose (final) + Ours✓69.684.9
GenFlow (init)-25.027.7
GenFlow (final)✓63.583.3
GenFlow (init) + Ours✓64.483.6
GenFlow (final) + Ours✓68.785.5
FoundationPose (init)---
FoundationPose (final)✓75.688.9
FoundationPose (init) + Ours✓79.790.4
FoundationPose (final) + Ours✓80.191.4
", + "image_path": "3b417ab83ff58bf96a7884c19198e8da908705bc3089496f5c2ae8a47f0d871f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 312, + 298, + 356 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 295, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 299 + ], + "type": "text", + "content": "Table 2. Effectiveness as a plug-and-play object pose refiner. All the baselines are based on an initialization-refinement architecture. Our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 312, + 298, + 356 + ], + "lines": [ + { + "bbox": [ + 58, + 312, + 298, + 356 + ], + "spans": [ + { + "bbox": [ + 58, + 312, + 298, + 356 + ], + "type": "table", + "html": "
L3L5L10L20L30L40L50
Initialization92.783.864.53517.67.83.5
+ SCFlow291.691.591.590.888.683.270.7
", + "image_path": "fc967123e8bedd2cce676a2f6911855b11fad453a4dc00381ef1190a44faa51a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 149, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 149, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 149, + 456 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness as a plug-and-play object pose refiner. We evaluate SCFlow2 in more detail in the context of three different state-of-the-art methods, FoundPose [36], GenFlow [32], and FoundationPose [50]. Table 2 summaries the results. All of them are based on an initialization-refinement architecture. We denote their initialization results as \"init\" in the table, and replace their refinement methods with ours in each method (denoted as \"init + ours\"). Furthermore, we also report the results of applying our refinement to their final results directly, denoted as \"final + ours\". Note that, FoundationPose outputs hundreds of pose hypotheses in the initialization stage, and there is no well-defined initialization result for it. In this case, we replace its original refinement method with SCFlow2 in its framework, and keep all other components untouched, including post-processing procedures such as pose scoring and pose selection, which is the reported result of \"FoundationPose (init) + Ours\" in the table. Overall, our method consistently outperforms their original refinement strategies, and can also improve their final results significantly, in a plug-and-play manner. We show some qualitative results in Fig. 5 and 6." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 333, + 70, + 537, + 209 + ], + "blocks": [ + { + "bbox": [ + 55, + 364, + 295, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 419 + ], + "type": "text", + "content": "Table 3. Impact of pose initialization. We add randomly generated rotation and translation noise to GT pose as the initialization, and then use SCFlow2 to refine the pose. L3 to L50 denote different noise levels. SCFlow2 is robust to pose errors in the initialization and works well under noise level 30." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 70, + 537, + 209 + ], + "lines": [ + { + "bbox": [ + 333, + 70, + 537, + 209 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 537, + 209 + ], + "type": "table", + "html": "
DatasetsInit.SeenUnseen
V1V1+V1++V2
LM-O61.062.969.368.969.6
T-LESS57.057.954.955.358.6
TUD-L69.475.988.886.688.1
IC-BIN47.947.954.153.754.6
ITODD40.734.249.148.148.4
HB72.374.576.576.176.2
YCB-V69.080.785.584.484.9
Avg.59.662.068.367.668.6
", + "image_path": "569fdb8791e2294aa9b2714e6b1483ae5b874328f590d09f9ce12db9412c3ea8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 262 + ], + "type": "text", + "content": "Table 4. Advantages over SCFlow. We use the same initialization based on results from FoundPose (\"Init.\", and report results of SCFlow (\"V1\") and SCFlow2 (\"V2\"). SCFlow2 outperforms all \"V1\" versions without any retraining." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 282, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 353 + ], + "type": "text", + "content": "Impact of pose initialization. We add random rotation and translation noise to GT pose as the initialization, and report the results on the YCB-V dataset, as shown in Table 3. L3 to L50 denote different noise levels (rotation error from 3 to 50 degrees, translation error from 3 to " + }, + { + "bbox": [ + 313, + 282, + 555, + 353 + ], + "type": "inline_equation", + "content": "50~\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 282, + 555, + 353 + ], + "type": "text", + "content": "). SCFlow2 works well under noise level 30." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "type": "text", + "content": "Evaluation of performance with different iterations. We conduct ablation studies on the trade off between the accuracy/efficiency and the iteration numbers of our method, as shown in Fig. 7. Our method produces meaningful pose refinement results after only 2 iterations and reaches saturation after 6 iterations. In our experiments, we use 8 iterations as the default setting in both training and testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 641 + ], + "type": "text", + "content": "Advantages over SCFlow. We compare SCFlow2 and its previous version SCFlow systematically. We summarize the results in Table 4. We use the same initialization based on results from FoundPose, and refine it with SCFlow (\"V1\") and SCFlow2 (\"V2\"), respectively. We evaluate three different versions of SCFlow, where \"V1\" denotes the original version based on retraining on target objects with RGB images only, \"V1+\" denotes the version using RANSAC Kabsch to consume additional depth images. Both \"V1\" and \"V1+\" need retraining on target objects. We also evaluate a version \"V1++\" which does not rely on retraining for novel objects, but uses the same large-scale datasets as \"V2\" for training on RGB images and then refine the pose results using RANSAC Kabsch with additional depths. We can see that, without retraining, \"V1++\" produces worse results compared with \"V1+\". On the other hand, SCFlow2 even outperforms \"V1+\" without any retraining." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Evaluation with different settings. We use the initialization results of FoundPose as our initialization. Table 5 shows the result, where \"Shape\" denotes the shape-constraint regularization during the recurrent optimization, and \"Scene\" is the 3D scene flow representation. We use the intermediate flow for the look-up without relying on the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22050" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 143, + 152 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 143, + 152 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 143, + 152 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 143, + 152 + ], + "type": "image", + "image_path": "8b7028c30c3d5d8ea2d9947bc123ee74a1fb6705470fbfa30a7a15722c91a3b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 70, + 223, + 152 + ], + "blocks": [ + { + "bbox": [ + 143, + 70, + 223, + 152 + ], + "lines": [ + { + "bbox": [ + 143, + 70, + 223, + 152 + ], + "spans": [ + { + "bbox": [ + 143, + 70, + 223, + 152 + ], + "type": "image", + "image_path": "4e49da28d43c17269488796d27168477f7ab0a513a507a903e96fce5506ea62a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 304, + 152 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 304, + 152 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 304, + 152 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 304, + 152 + ], + "type": "image", + "image_path": "7a7642d12614e45cdd0a9e4d2fa1128957ccecc1b2dc0390af2659ce920f921f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 386, + 152 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 386, + 152 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 386, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 386, + 152 + ], + "type": "image", + "image_path": "d5ba2bb65a4bff0bab725fcfbe0cf79b020266aed3b78a6eccf15b8125f98b18.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 386, + 70, + 467, + 152 + ], + "blocks": [ + { + "bbox": [ + 386, + 70, + 467, + 152 + ], + "lines": [ + { + "bbox": [ + 386, + 70, + 467, + 152 + ], + "spans": [ + { + "bbox": [ + 386, + 70, + 467, + 152 + ], + "type": "image", + "image_path": "e7a3a6acd7f7c6bef5825e99725791fa99ee534aab1c348564223d7acfc94faf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 467, + 70, + 548, + 152 + ], + "blocks": [ + { + "bbox": [ + 467, + 70, + 548, + 152 + ], + "lines": [ + { + "bbox": [ + 467, + 70, + 548, + 152 + ], + "spans": [ + { + "bbox": [ + 467, + 70, + 548, + 152 + ], + "type": "image", + "image_path": "e07ccc5ffb53a64aff587442bbd5a33a56e2edd160831414accce8057d2c7289.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 153, + 143, + 233 + ], + "blocks": [ + { + "bbox": [ + 61, + 153, + 143, + 233 + ], + "lines": [ + { + "bbox": [ + 61, + 153, + 143, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 143, + 233 + ], + "type": "image", + "image_path": "97ccebaf4a43b588bfa22824cf0f63ec473fd93b1120780fab4fa77acb2f26f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 153, + 223, + 233 + ], + "blocks": [ + { + "bbox": [ + 143, + 153, + 223, + 233 + ], + "lines": [ + { + "bbox": [ + 143, + 153, + 223, + 233 + ], + "spans": [ + { + "bbox": [ + 143, + 153, + 223, + 233 + ], + "type": "image", + "image_path": "a4ce9c2d457f94737b8fc86dbd41793765b2687358f650cd8b070e60c518ea43.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 153, + 304, + 233 + ], + "blocks": [ + { + "bbox": [ + 223, + 153, + 304, + 233 + ], + "lines": [ + { + "bbox": [ + 223, + 153, + 304, + 233 + ], + "spans": [ + { + "bbox": [ + 223, + 153, + 304, + 233 + ], + "type": "image", + "image_path": "607345a095c7a497fac7616da5bdc4f6a26b46723b6604b170f5a2c74d0cbfe0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 153, + 386, + 233 + ], + "blocks": [ + { + "bbox": [ + 304, + 153, + 386, + 233 + ], + "lines": [ + { + "bbox": [ + 304, + 153, + 386, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 386, + 233 + ], + "type": "image", + "image_path": "d504ef74dcc57f2dddf986ae5b9fecd43ee69686ae45149616d634e0d91b7afe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 386, + 153, + 467, + 233 + ], + "blocks": [ + { + "bbox": [ + 386, + 153, + 467, + 233 + ], + "lines": [ + { + "bbox": [ + 386, + 153, + 467, + 233 + ], + "spans": [ + { + "bbox": [ + 386, + 153, + 467, + 233 + ], + "type": "image", + "image_path": "635c2bfe4f7fb40a28a26fdfce49c935bf417060a02cdea8c65ee8b52aba139f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 467, + 153, + 548, + 233 + ], + "blocks": [ + { + "bbox": [ + 467, + 153, + 548, + 233 + ], + "lines": [ + { + "bbox": [ + 467, + 153, + 548, + 233 + ], + "spans": [ + { + "bbox": [ + 467, + 153, + 548, + 233 + ], + "type": "image", + "image_path": "c109ab539f6cd06f3a4e17ee4fdde87c7ba2297bb0d411d91cb4b57f45c934e8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 61, + 235, + 142, + 316 + ], + "blocks": [ + { + "bbox": [ + 61, + 235, + 142, + 316 + ], + "lines": [ + { + "bbox": [ + 61, + 235, + 142, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 142, + 316 + ], + "type": "image", + "image_path": "8e765ca510dd1325d8e9011166b6fdcee7bc93e1072ab0df618f9ea18ba88f1f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 143, + 235, + 223, + 316 + ], + "blocks": [ + { + "bbox": [ + 143, + 235, + 223, + 316 + ], + "lines": [ + { + "bbox": [ + 143, + 235, + 223, + 316 + ], + "spans": [ + { + "bbox": [ + 143, + 235, + 223, + 316 + ], + "type": "image", + "image_path": "23279438713f379fa872ef124492582d9f90991059264e4222971e8371386615.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 223, + 235, + 304, + 316 + ], + "blocks": [ + { + "bbox": [ + 223, + 235, + 304, + 316 + ], + "lines": [ + { + "bbox": [ + 223, + 235, + 304, + 316 + ], + "spans": [ + { + "bbox": [ + 223, + 235, + 304, + 316 + ], + "type": "image", + "image_path": "c301fc7c2a99f6f672dc0f3fe1debabf8923498594fb57f3cb11615c06034b64.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 304, + 235, + 386, + 316 + ], + "blocks": [ + { + "bbox": [ + 304, + 235, + 386, + 316 + ], + "lines": [ + { + "bbox": [ + 304, + 235, + 386, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 386, + 316 + ], + "type": "image", + "image_path": "24e41fbaef3e5ff135ffdc33407da9cd500de10a59e62e94d19900372556106e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 386, + 235, + 467, + 316 + ], + "blocks": [ + { + "bbox": [ + 386, + 235, + 467, + 316 + ], + "lines": [ + { + "bbox": [ + 386, + 235, + 467, + 316 + ], + "spans": [ + { + "bbox": [ + 386, + 235, + 467, + 316 + ], + "type": "image", + "image_path": "9b8b379dfabd67ed2ef49c1e7b787dada8389ebfc7a4e9640803e9ee98dbb3b5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 467, + 235, + 548, + 316 + ], + "blocks": [ + { + "bbox": [ + 467, + 235, + 548, + 316 + ], + "lines": [ + { + "bbox": [ + 467, + 235, + 548, + 316 + ], + "spans": [ + { + "bbox": [ + 467, + 235, + 548, + 316 + ], + "type": "image", + "image_path": "b8a0c7719260f15ff5c0b156dfdb93aae41bf0047ca0669a31aded29f67f7f9f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 61, + 318, + 142, + 388 + ], + "blocks": [ + { + "bbox": [ + 61, + 318, + 142, + 388 + ], + "lines": [ + { + "bbox": [ + 61, + 318, + 142, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 142, + 388 + ], + "type": "image", + "image_path": "659dc210b7221b1cbef18fadfd358f8768628a70a581300c0ef799210a095f84.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 143, + 318, + 223, + 388 + ], + "blocks": [ + { + "bbox": [ + 143, + 318, + 223, + 388 + ], + "lines": [ + { + "bbox": [ + 143, + 318, + 223, + 388 + ], + "spans": [ + { + "bbox": [ + 143, + 318, + 223, + 388 + ], + "type": "image", + "image_path": "c67747d55f53117ffbf55f478c95bdd44e25fe514adefb81c98869e28f47441a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 223, + 318, + 304, + 388 + ], + "blocks": [ + { + "bbox": [ + 223, + 318, + 304, + 388 + ], + "lines": [ + { + "bbox": [ + 223, + 318, + 304, + 388 + ], + "spans": [ + { + "bbox": [ + 223, + 318, + 304, + 388 + ], + "type": "image", + "image_path": "957ac03eef067decf82350030ecd1eeffdc0dea662521e32a3fe934b42637d31.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 304, + 318, + 386, + 388 + ], + "blocks": [ + { + "bbox": [ + 304, + 318, + 386, + 388 + ], + "lines": [ + { + "bbox": [ + 304, + 318, + 386, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 386, + 388 + ], + "type": "image", + "image_path": "5ff90ee668a8848f6081688f54389c72330cc939446c58e56c7b15cc75d649cc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 386, + 318, + 467, + 388 + ], + "blocks": [ + { + "bbox": [ + 386, + 318, + 467, + 388 + ], + "lines": [ + { + "bbox": [ + 386, + 318, + 467, + 388 + ], + "spans": [ + { + "bbox": [ + 386, + 318, + 467, + 388 + ], + "type": "image", + "image_path": "c6cc5653ccee602f61fa0abe7d6c8c3a1aef917868c38a4eb82f17e7f3200208.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 467, + 318, + 548, + 388 + ], + "blocks": [ + { + "bbox": [ + 467, + 318, + 548, + 388 + ], + "lines": [ + { + "bbox": [ + 467, + 318, + 548, + 388 + ], + "spans": [ + { + "bbox": [ + 467, + 318, + 548, + 388 + ], + "type": "image", + "image_path": "60e7eae561a438b81a8beddd50d785ee221b184baa52efe9f3b10e6293f5a852.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 61, + 401, + 142, + 483 + ], + "blocks": [ + { + "bbox": [ + 61, + 401, + 142, + 483 + ], + "lines": [ + { + "bbox": [ + 61, + 401, + 142, + 483 + ], + "spans": [ + { + "bbox": [ + 61, + 401, + 142, + 483 + ], + "type": "image", + "image_path": "2f81ae0963f8fc4d78a853b9925684f6269620215b688ac49f5c6ccbcfe3e49a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 143, + 401, + 223, + 483 + ], + "blocks": [ + { + "bbox": [ + 143, + 401, + 223, + 483 + ], + "lines": [ + { + "bbox": [ + 143, + 401, + 223, + 483 + ], + "spans": [ + { + "bbox": [ + 143, + 401, + 223, + 483 + ], + "type": "image", + "image_path": "8e0b37b7df406a192939e7b086fabe80b58154deda543c015ee527eb76369b6b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 223, + 401, + 304, + 483 + ], + "blocks": [ + { + "bbox": [ + 223, + 401, + 304, + 483 + ], + "lines": [ + { + "bbox": [ + 223, + 401, + 304, + 483 + ], + "spans": [ + { + "bbox": [ + 223, + 401, + 304, + 483 + ], + "type": "image", + "image_path": "332da04141ccf188cc9f1f0786c2a5ef2461ab49e4d2453082a1b7d9129ab4f8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 304, + 401, + 386, + 483 + ], + "blocks": [ + { + "bbox": [ + 304, + 401, + 386, + 483 + ], + "lines": [ + { + "bbox": [ + 304, + 401, + 386, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 386, + 483 + ], + "type": "image", + "image_path": "0da0edac4c1e76502e54d9a04598b59cb1399b61ea00d645951c63c21d4e49d5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 386, + 401, + 467, + 483 + ], + "blocks": [ + { + "bbox": [ + 386, + 401, + 467, + 483 + ], + "lines": [ + { + "bbox": [ + 386, + 401, + 467, + 483 + ], + "spans": [ + { + "bbox": [ + 386, + 401, + 467, + 483 + ], + "type": "image", + "image_path": "f573fef8b980e886d4f8fd8976a8de8dd63d6d0d60ee46e2b5f475c49ae3bbdc.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 467, + 401, + 548, + 483 + ], + "blocks": [ + { + "bbox": [ + 467, + 401, + 548, + 483 + ], + "lines": [ + { + "bbox": [ + 467, + 401, + 548, + 483 + ], + "spans": [ + { + "bbox": [ + 467, + 401, + 548, + 483 + ], + "type": "image", + "image_path": "a2f89dce3be900307788c91a7867ca9a33f3646541e66d1e72e3350a6a0e8bd2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 61, + 484, + 142, + 565 + ], + "blocks": [ + { + "bbox": [ + 61, + 484, + 142, + 565 + ], + "lines": [ + { + "bbox": [ + 61, + 484, + 142, + 565 + ], + "spans": [ + { + "bbox": [ + 61, + 484, + 142, + 565 + ], + "type": "image", + "image_path": "97e578f7e4fb968a9eeec9700270085e2cc5257cdf0bbb2eb3fe58e82846b58d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 568, + 126, + 578 + ], + "lines": [ + { + "bbox": [ + 78, + 568, + 126, + 578 + ], + "spans": [ + { + "bbox": [ + 78, + 568, + 126, + 578 + ], + "type": "text", + "content": "GFlow (init)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 143, + 484, + 223, + 565 + ], + "blocks": [ + { + "bbox": [ + 143, + 484, + 223, + 565 + ], + "lines": [ + { + "bbox": [ + 143, + 484, + 223, + 565 + ], + "spans": [ + { + "bbox": [ + 143, + 484, + 223, + 565 + ], + "type": "image", + "image_path": "404d233ef64ae10c59d59b6b4493846aa9ce662bcd81fba7cab67e6f01137aad.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 568, + 209, + 578 + ], + "lines": [ + { + "bbox": [ + 158, + 568, + 209, + 578 + ], + "spans": [ + { + "bbox": [ + 158, + 568, + 209, + 578 + ], + "type": "text", + "content": "GFlow (final)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 223, + 484, + 304, + 565 + ], + "blocks": [ + { + "bbox": [ + 223, + 484, + 304, + 565 + ], + "lines": [ + { + "bbox": [ + 223, + 484, + 304, + 565 + ], + "spans": [ + { + "bbox": [ + 223, + 484, + 304, + 565 + ], + "type": "image", + "image_path": "dc2e05ae4a2afc6a00c5acb7f8292429feb91199946a844b0cba183fb409e51c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 568, + 304, + 578 + ], + "lines": [ + { + "bbox": [ + 228, + 568, + 304, + 578 + ], + "spans": [ + { + "bbox": [ + 228, + 568, + 304, + 578 + ], + "type": "text", + "content": "GFlow (init) + Ours" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 306, + 484, + 386, + 565 + ], + "blocks": [ + { + "bbox": [ + 306, + 484, + 386, + 565 + ], + "lines": [ + { + "bbox": [ + 306, + 484, + 386, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 386, + 565 + ], + "type": "image", + "image_path": "78c14e4cb765558c16daf893eba69843bd4b34706b4fced397c1d857f3552075.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 568, + 367, + 578 + ], + "lines": [ + { + "bbox": [ + 323, + 568, + 367, + 578 + ], + "spans": [ + { + "bbox": [ + 323, + 568, + 367, + 578 + ], + "type": "text", + "content": "FPose (init)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 386, + 484, + 467, + 565 + ], + "blocks": [ + { + "bbox": [ + 386, + 484, + 467, + 565 + ], + "lines": [ + { + "bbox": [ + 386, + 484, + 467, + 565 + ], + "spans": [ + { + "bbox": [ + 386, + 484, + 467, + 565 + ], + "type": "image", + "image_path": "8bca1447470d91eb08343caab2eb956aade89024ae341920a1bef03d3cd941fa.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 568, + 451, + 578 + ], + "lines": [ + { + "bbox": [ + 403, + 568, + 451, + 578 + ], + "spans": [ + { + "bbox": [ + 403, + 568, + 451, + 578 + ], + "type": "text", + "content": "FPose (final)" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 589, + 555, + 623 + ], + "lines": [ + { + "bbox": [ + 55, + 589, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 555, + 623 + ], + "type": "text", + "content": "Figure 5. Qualitative results. Given the same pose initialization as that in GenFlow (\"GFlow\") and FoundPose (\"FPose\"), denoted as \"GFlow (init)\" and \"FPose (init)\" respectively, our refinement method (\"+ Ours\") produces considerably more accurate results compared to the refinement approaches in their original methods (note how our reprojected 3D mesh aligns better with the object contours)." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 467, + 484, + 548, + 565 + ], + "blocks": [ + { + "bbox": [ + 467, + 484, + 548, + 565 + ], + "lines": [ + { + "bbox": [ + 467, + 484, + 548, + 565 + ], + "spans": [ + { + "bbox": [ + 467, + 484, + 548, + 565 + ], + "type": "image", + "image_path": "9c983e00021a65622b43e1a7c60b429d5150285c4b212e326f6fda2bcc1c8ac1.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 568, + 546, + 578 + ], + "lines": [ + { + "bbox": [ + 473, + 568, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 473, + 568, + 546, + 578 + ], + "type": "text", + "content": "FPose (init) + Ours" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 643, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 297, + 704 + ], + "type": "text", + "content": "target's 3D mesh when \"Shape\" is disabled, and we replace the scene flow representation with a plain MLP pose regressor when \"Scene\" is disabled. The performance becomes worse without the shape prior embedding, and deteriorates significantly without the 3D scene flow formulation." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 643, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 556, + 704 + ], + "type": "text", + "content": "Timing analysis. The training of SCFlow2 needs about 108 hours on a workstation with 2 NVIDIA RTX-3090 GPU. SCFlow2 considers the pose initialization as the only hypothesis. During inference, SCFlow2 takes about 0.18 seconds to refine a pose on an RTX-3090 GPU with a typical" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22051" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 176, + 157 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 176, + 157 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 176, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 176, + 157 + ], + "type": "image", + "image_path": "c95330b737894cbb359c6010e4603e0a782e5fb10bccb5596732c1c155864bb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 291, + 157 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 291, + 157 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 291, + 157 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 291, + 157 + ], + "type": "image", + "image_path": "e740f9664a51697a2cde48f58569336b541ea139ef67a49828b420c294cc6902.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 159, + 175, + 246 + ], + "blocks": [ + { + "bbox": [ + 60, + 159, + 175, + 246 + ], + "lines": [ + { + "bbox": [ + 60, + 159, + 175, + 246 + ], + "spans": [ + { + "bbox": [ + 60, + 159, + 175, + 246 + ], + "type": "image", + "image_path": "55d3517f07d5b89afb8d24213ce703691f61a6ca04537979f73a80a45aa399f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 249, + 141, + 258 + ], + "lines": [ + { + "bbox": [ + 94, + 249, + 141, + 258 + ], + "spans": [ + { + "bbox": [ + 94, + 249, + 141, + 258 + ], + "type": "text", + "content": "Initialization" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 159, + 291, + 247 + ], + "blocks": [ + { + "bbox": [ + 176, + 159, + 291, + 247 + ], + "lines": [ + { + "bbox": [ + 176, + 159, + 291, + 247 + ], + "spans": [ + { + "bbox": [ + 176, + 159, + 291, + 247 + ], + "type": "image", + "image_path": "b04ffe006eaeee30faac55b6f08f36daaf59efd8be21a8a82765a38b4b5c753d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 250, + 247, + 258 + ], + "lines": [ + { + "bbox": [ + 220, + 250, + 247, + 258 + ], + "spans": [ + { + "bbox": [ + 220, + 250, + 247, + 258 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 314 + ], + "type": "text", + "content": "Figure 6. More qualitative results. We show two examples on YCB-V. Although the initialization has large pose error, our refinement method produces accurate results that align well with the target, without retraining on these novel objects." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 99, + 327, + 254, + 388 + ], + "blocks": [ + { + "bbox": [ + 99, + 327, + 254, + 388 + ], + "lines": [ + { + "bbox": [ + 99, + 327, + 254, + 388 + ], + "spans": [ + { + "bbox": [ + 99, + 327, + 254, + 388 + ], + "type": "table", + "html": "
ShapeSceneLM-OYCB-V
✓-52.667.6
-✓60.080.4
✓✓61.381.8
", + "image_path": "dd0e9f9c9d3fea97eec4fc78dfd707611801eac618a402b54f8f368a3abbe1d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 396, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 440 + ], + "type": "text", + "content": "Table 5. Evaluation with different settings. \"Shape\" denotes the shape-constraint regularization during the optimization, and \"Scene\" is the 3D scene flow formulation. The performance deteriorates significantly without the 3D scene flow formulation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 463, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 510 + ], + "type": "text", + "content": "image from BOP dataset. Taken for comparison, the multi-hypothesis based method MegaPose needs about 1.85 seconds for a refinement on the same workstation, which is " + }, + { + "bbox": [ + 55, + 463, + 295, + 510 + ], + "type": "inline_equation", + "content": "10+" + }, + { + "bbox": [ + 55, + 463, + 295, + 510 + ], + "type": "text", + "content": " times slower than SCFlow2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 512, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 654 + ], + "type": "text", + "content": "Limitation discussion. In general, our refinement method can improve the accuracy of initial pose significantly. While occasionally, it produces worse results than the initial pose. Fig. 8 shows the histogram of samples along with the relative accuracy improvement after pose refinement on YCB-V and LM-O with the initial pose of FoundPose as our initialization. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases where the initial poses are already very wrong caused by severe occlusion or extreme light conditions. We will take this as one of our future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We have introduced SCFlow2, a plug-and-play refinement framework for 6D object pose estimation. We extended" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 432, + 168 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 432, + 168 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 432, + 168 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 432, + 168 + ], + "type": "image", + "image_path": "053909c232e707e09b3a4a3fa68e70f5b8d018233ae3daba93ef087ff50defcc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 177, + 554, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 554, + 243 + ], + "type": "text", + "content": "Figure 7. Evaluation of performance with different iterations. Given the same pose initialization with noise level 30 as in Table 3, more recurrent iterations produce more accurate results while spending more time in inference. Nevertheless, our approach yields meaningful outcomes after just 2 iterations, and the performance reaches saturation after 6 iterations." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 437, + 72, + 552, + 167 + ], + "blocks": [ + { + "bbox": [ + 437, + 72, + 552, + 167 + ], + "lines": [ + { + "bbox": [ + 437, + 72, + 552, + 167 + ], + "spans": [ + { + "bbox": [ + 437, + 72, + 552, + 167 + ], + "type": "image", + "image_path": "9855c0de2da3571d78a875667dd53ce61edf8966f336df35e5f8afa7713a4ee9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 274, + 493, + 370 + ], + "blocks": [ + { + "bbox": [ + 378, + 274, + 493, + 370 + ], + "lines": [ + { + "bbox": [ + 378, + 274, + 493, + 370 + ], + "spans": [ + { + "bbox": [ + 378, + 274, + 493, + 370 + ], + "type": "image", + "image_path": "051e7db46c7cbed92218312665a01a5baaab189b5ebb60cb75be1bcd2ff23c04.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 554, + 435 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 435 + ], + "type": "text", + "content": "Figure 8. Limitation discussion. We show the histogram of samples along with the relative accuracy improvement after pose refinement. In the majority of cases, our method improves the initial poses significantly. However, it still happens that the refinement makes the results worse in some extreme cases." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 476, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 554, + 583 + ], + "type": "text", + "content": "the RGB-based method SCFlow by formulating additional depth as a regularization in the iteration with an intermediate 3D scene flow representation for RGBD frames, and combined it with the 3D shape prior information of the target, producing a geometry-guided 6D object pose refiner end-to-end trainable and generalizable to novel objects. Experiments on BOP datasets with novel objects show that our method achieves state-of-the-art accuracy, and is one of the most efficient refinement methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the Youth Innovation Team of Shaanxi Universities, the \"Scientist + Engineer\" team of the Qin Chuang Yuan in Shaanxi Province, the Xi'an Science and Technology Program, the \"Leading the Charge\" initiative for the industrialization of core technologies in key industrial chains in Shaanxi Province, the National Nature Science Foundation of China under Grant 62371359, the Key Research and Development Program of Shaanxi, the Key Research Program of the Chinese Academy of Sciences." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22052" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D Object Pose Estimation Using 3D Object Coordinates. In European Conference on Computer Vision, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An Information-Rich 3D Model Repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-End Learnable Geometric Vision by Backpropagating PnP Optimization. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "type": "text", + "content": "[4] Kyunghyun Cho, Bart Van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv preprint arXiv:1409.1259, 2014. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[5] Javier Cuadrado, Ulysse Rancon, Benoit R Cotterreau, Francisco Barranco, and Timothée Masquelier. Optical Flow Estimation from Event-based Cameras and Spiking Neural Networks. Frontiers in Neuroscience, 17, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 379 + ], + "type": "text", + "content": "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A Universe of Annotated 3D Objects. In Conference on Computer Vision and Pattern Recognition, 2023. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[7] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6D Object Pose and Predicting Next-Best-View in the Crowd. In Conference on Computer Vision and Pattern Recognition, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 489 + ], + "type": "text", + "content": "[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items. In International Conference on Robotics and Automation, 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "type": "text", + "content": "[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing MVTec ITODDA Dataset for 3D Object Recognition in Industry. In International Conference on Computer Vision Workshops, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[10] Yang Hai, Rui Song, Jiaojiao Li, David Ferstl, and Yinlin Hu. Pseudo Flow Consistency for Self-Supervised 6D Object Pose Estimation. In International Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Yang Hai, Rui Song, Jiaojiao Li, and Yinlin Hu. Shape-Constraint Flow for 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[12] Rasmus Laurvig Haugaard and Anders Glent Buch. SurfEmb: Dense and Continuous Correspondence Distributions for Object Pose Estimation with Learnt Surface Embeddings. In Conference on Computer Vision and Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 712 + ], + "type": "text", + "content": "[13] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. FS6D: Few-Shot 6D Pose Estimation of Novel Ob" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "jects. In Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "type": "text", + "content": "[14] Tomás Hodań, Pavel Haluza, Štepan Obdržánek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D Dataset for 6D Pose Estimation of Texture-Less Objects. In Winter Conference on Applications of Computer Vision, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 208 + ], + "type": "text", + "content": "[15] Tomáš Hodař, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In European Conference on Computer Vision, 2018. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 210, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 553, + 243 + ], + "type": "text", + "content": "[16] Tomás Hodań, Daniel Barath, and Jérí Matas. EPOS: Estimating 6D Pose of Objects with Symmetries. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 245, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 267 + ], + "type": "text", + "content": "[17] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 269, + 553, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 301 + ], + "type": "text", + "content": "[18] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-Stage 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 303, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 553, + 346 + ], + "type": "text", + "content": "[19] Yinlin Hu, Sebastien Speierer, Wenzel Jakob, Pascal Fua, and Mathieu Salzmann. Wide-Depth-Range 6D Object Pose Estimation in Space. In Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[20] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective Flow Aggregation for Data-Limited 6D Object Pose Estimation. In European Conference on Computer Vision, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 395, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 438 + ], + "type": "text", + "content": "[21] Wolfgang Kabsch. A Solution for the Best Rotation to Relate Two Sets of Vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922-923, 1976. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "type": "text", + "content": "[22] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: RGB-D Dataset for 6D Pose Estimation of 3D Objects. In International Conference on Computer Vision Workshops, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 487, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 541 + ], + "type": "text", + "content": "[23] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In Conference on Robot Learning. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "text", + "content": "[24] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent Multi-View Multi-Object 6d Pose Estimation. In European Conference on Computer Vision, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 589, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 622 + ], + "type": "text", + "content": "[25] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep Iterative Matching for 6D Pose Estimation. In European Conference on Computer Vision, 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[26] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. SAM-6D: Segment Anything Model Meets Zero-Shot 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22053" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[28] Yongliang Lin, Yongzhi Su, Praveen Nathan, Sandeep Inuganti, Yan Di, Martin Sundermeyer, Fabian Manhardt, Didier Stricker, Jason Rambach, and Yu Zhang. HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[29] Fulin Liu, Yinlin Hu, and Mathieu Salzmann. Linear-Covariance Loss for End-to-End Learning of 6D Pose Estimation. In International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[31] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose Estimation for Augmented Reality: A Hands-On Survey. IEEE transactions on visualization and computer graphics, 22(12):2633-2651, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[32] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. GenFlow: Generalizable Recurrent Flow for 6D Pose Refinement of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "type": "text", + "content": "[33] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. NOPE: Novel Object Pose Estimation from a Single Image. In Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[34] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence. In Conference on Computer Vision and Pattern Recognition, 2024. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOV2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "text", + "content": "[36] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found-Pose: Unseen Object Pose Estimation with Foundation Features. European Conference on Computer Vision, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "type": "text", + "content": "[37] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. LatentFusion: End-to-End Differentiable Reconstruction and Rendering for Unseen Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[38] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. PVNet: Pixel-Wise Voting Network for 6DoF Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[39] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. arXiv preprint arXiv:1706.02413, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[40] Mahdi Rad and Vincent Lepetit. BB8: A Scalable, Accurate, Robust to Partial Occlusion Method for Predicting the 3D Poses of Challenging Objects without Using Depth. In International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[41] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The Surprising Effectiveness of Diffusion Models for Optical Flow and Monocular Depth Estimation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "text", + "content": "[42] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A Multi-Stage One Shot Object Pose Estimation Framework. In Conference on Computer Vision and Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "text", + "content": "[43] Chen Song, Jiaru Song, and Qixing Huang. HybridPose: 6D Object Pose Estimation under Hybrid Representations. In Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[44] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. ZebraPose: Coarse to Fine Surface Encoding for 6DoF Object Pose Estimation. arXiv preprint arXiv:2203.09418, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[45] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-Shot Object Pose Estimation without CAD Models. In Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 359 + ], + "type": "text", + "content": "[46] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In European Conference on Computer Vision, 2020. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "text", + "content": "[47] Zachary Teed and Jia Deng. RAFT-3D: Scene Flow using Rigid-Motion Embeddings. In Conference on Computer Vision and Pattern Recognition, 2021. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "text", + "content": "[48] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martín-Martín, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion. In Conference on Computer Vision and Pattern Recognition, 2019. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[49] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. In Conference on Computer Vision and Pattern Recognition, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "text", + "content": "[50] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. In Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[51] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes. In Robotics: Science and Systems, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[52] Zhaolin Xiao, Zhan Su, Fengyuan Zuo, and Haiyan Jin. Low-Light Optical Flow Estimation with Hidden Feature Supervision Using a Siamese Network. Journal of Image and Graphics, 29:0231-0242, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[53] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. DPOD: 6D Pose Object Detector and Refiner. In International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[54] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Conference on Computer Vision and Pattern Recognition, 2019. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22054" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_content_list.json b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02815f6110a4dbf27af30f333e520393dd7c61a7 --- /dev/null +++ b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_content_list.json @@ -0,0 +1,2214 @@ +[ + { + "type": "text", + "text": "SCSA: A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer", + "text_level": 1, + "bbox": [ + 107, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunnan Shang", + "bbox": [ + 272, + 203, + 403, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhejiang University", + "bbox": [ + 258, + 222, + 419, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chunnan.22@intl.zju.edu.cn", + "bbox": [ + 223, + 242, + 455, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongwei Wang", + "bbox": [ + 276, + 257, + 401, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhejiang University", + "bbox": [ + 258, + 275, + 419, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hongweiwang@intl.zju.edu.cn", + "bbox": [ + 217, + 295, + 460, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhizhong Wang†", + "bbox": [ + 581, + 203, + 718, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhejiang University", + "bbox": [ + 565, + 222, + 725, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "endywon@zju.edu.cn", + "bbox": [ + 560, + 242, + 723, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangming Meng†", + "bbox": [ + 575, + 256, + 725, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhejiang University", + "bbox": [ + 565, + 275, + 725, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiangmingmeng@intl.zju.edu.cn", + "bbox": [ + 514, + 295, + 774, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Attention-based arbitrary style transfer methods, including CNN-based, Transformer-based, and Diffusion-based, have flourished and produced high-quality stylized images. However, they perform poorly on the content and style images with the same semantics, i.e., the style of the corresponding semantic region of the generated stylized image is inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. To address this issue, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer—each query point considers certain key points in the corresponding semantic region. Specifically, semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region that reflect the overall style characteristics of that region; Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region that exhibits the specific stylistic texture of that region. By combining the two modules, the resulting SCSA aligns the overall style of the corresponding semantic regions while transferring the vivid textures of these regions. Qualitative and quantitative results demonstrate that SCSA enables attention-based arbitrary style transfer methods to produce high-quality semantic stylized images. The code can be found in https://github.com/ecn-00/SCSA.", + "bbox": [ + 89, + 376, + 483, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Initially, style transfer [11] was proposed to preserve the content of a specified content image while transferring the style from a specified style image by iteratively optimizing", + "bbox": [ + 89, + 827, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Content Inputs", + "bbox": [ + 549, + 345, + 625, + 356 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29539f9393f59eaedf8e4f3df4fe7a2d31336c799a7ea941bd5a2d88c28e2bdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 356, + 625, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Style Inputs", + "bbox": [ + 637, + 345, + 699, + 356 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1edea075c59643bdb6e64fc76bd85e63f217067de0a7c98b0dc932efea7ee423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 356, + 707, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SANet", + "bbox": [ + 732, + 345, + 767, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3d5ff1194c5f66d5f494a69365fc6802c0f2c1a2c4904050a792ba40b25d14e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 356, + 789, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69520fff0fd62612fbfa196b714b03310c301051644920d01f06395aa510dd0e.jpg", + "image_caption": [ + "SANet + SCSA" + ], + "image_footnote": [], + "bbox": [ + 792, + 356, + 870, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15e701c7f8fbe11b2a43211c375f3c3348b8c91b1ca4f31f4bffa759fc87c0db.jpg", + "image_caption": [ + "Content Inputs" + ], + "image_footnote": [], + "bbox": [ + 547, + 434, + 627, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b07a3ebe2a8e54b28e9f8d93d9acfe4ef59b7447b6dc2499044183fefc17b97d.jpg", + "image_caption": [ + "Style Inputs" + ], + "image_footnote": [], + "bbox": [ + 629, + 434, + 707, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/596510bbb4c3721d65ef49ee5f6bd73a985709a969f13ca19a8534e36b7e10c8.jpg", + "image_caption": [ + "StyTr2" + ], + "image_footnote": [], + "bbox": [ + 710, + 434, + 789, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2db4d58977d0db3840ba12e85802fbb942e8b5673f201919e2d24d92f7bca1cc.jpg", + "image_caption": [ + "$\\mathrm{StyTr}^2 + \\mathrm{SCSA}$" + ], + "image_footnote": [], + "bbox": [ + 792, + 434, + 870, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09a63a75528b1fbb8509ed47479af42211873702bebafe06105d70d17228a1e7.jpg", + "image_caption": [ + "Content Inputs", + "Figure 1. Comparisons of the Attn-AST approaches-CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]-without and with our SCSA. The stylized images generated by the Attn-AST approaches exhibit style discontinuity in adjacent regions within the identical semantic regions (e.g., the background in the 1st row), style inconsistency between corresponding semantic (e.g., the cloth color in the 3rd row), and fewer textures (e.g., the block textures in the 3rd row)." + ], + "image_footnote": [], + "bbox": [ + 547, + 510, + 627, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff63f632d36177c00a54774ebecab33fc7a25d93f3a241aab2362609c72e2a84.jpg", + "image_caption": [ + "Style Inputs" + ], + "image_footnote": [], + "bbox": [ + 629, + 510, + 707, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9531137a9b4cdacae2a503bbec2bdab64926ebc500a11531a6f7d7afe7a24a17.jpg", + "image_caption": [ + "StyleID" + ], + "image_footnote": [], + "bbox": [ + 710, + 510, + 789, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fede9470022d5f96840822978dd18f1779f9560b937c83c7137b440777a05c8.jpg", + "image_caption": [ + "StyleID + SCSA" + ], + "image_footnote": [], + "bbox": [ + 792, + 510, + 870, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the generated stylized image. However, its excessive slowness prevents practical application. Given the need for efficiency, effectiveness, and generalization, Arbitrary Style Transfer (AST) [1, 4, 8, 23, 27] has become a highly prominent research focus in the style transfer field. AST seeks to generate a new image that replicates the content of an arbitrary content image while adopting the style of an arbitrary style image. In spite of many recent studies aiming to advance the AST effect by incorporating novel knowledge [15, 36, 39] and innovative constraints [41, 43, 49], the core of AST continues to reside in feature transformation modules, which remains indispensable for all approaches. Hence, the existing AST approaches can be classified into", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 109, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two categories, statistics-based methods [5, 16, 24, 48] and attention-based methods [6, 8, 9, 27, 31, 32].", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention-based Arbitrary Style Transfer (Attn-AST) approaches, leveraging the ability to capture global information and focus on the related features of the attention mechanism [2], preserve the content of the content image while transferring the style that is relevant to the content of the content image from the style image. They encompass CNN-based [27, 31, 32, 39], Transformer-based [9, 14, 42, 46], and Diffusion-based methods [8, 35, 44]. Commonly, Attn-AST methods first establish a holistic relationship, an attention map, between each content query point of the content image features and all style key points of the style image features. Then, Attn-AST methods get the transferred style for each content point by weighting all the style points, where the weights are the softmax function outputs of a naturally calculated attention map of content and style points.", + "bbox": [ + 89, + 125, + 482, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although Attn-AST methods have thrived and produced high-quality stylized images due to the capacity mentioned above for global capture and focus, they perform poorly on the content and style image with the same semantics. As shown in Fig. 1, the style of the corresponding semantic region of the stylized image generated by Attn-AST approaches exhibits a style inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. (1) During the formulation of the attention map, Attn-AST methods bluntly and simplistically consider the relationship between each query point and all key points, as shown in Fig. 2 (c), without adequately accounting for the semantic regions to which these points belong. This oversimplification may cause Attn-AST methods to disproportionately emphasize structurally similar points across different semantic regions, resulting in artifacts or inaccurate style [27, 32], e.g., the $2nd$ row of Fig. 1. (2) Attn-AST approaches fundamentally depend on structural principles for generating a global style, as they utilize content and style image features. The subtle disparity in content structure details across adjacent regions can cause different stylization effects. Thus, the stylized images often lack stylistic continuity across adjacent regions within the semantic regions, e.g., the background in the $1st$ row of Fig. 1. (3) Attn-AST approaches acquire the transferred style by weighting all the style points. However, the most accurate representation of the original stylistic texture resides in the encoded single points. The weighted style struggles to preserve the original stylistic texture, leading to a lack of vividness, e.g., the block textures in the $3rd$ row of Fig. 1.", + "bbox": [ + 89, + 353, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To maintain the benefits of Attn-AST methods while addressing their existing issues, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer. The key insight of SCSA is that, based on the interpretability of the attention mecha", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nism, each query point considers certain key points in the corresponding semantic region. Specifically, SCSA consists of two parts: Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA). (1) SCA: Semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region by using content and style semantic map features as query and key, respectively. Thus, SCA can achieve the overall style characteristics of the corresponding semantic regions of the style image to generate the stylized image with stylistic continuity across adjacent regions within the semantic regions. (2) SSA: Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region by using content and style image features as query and key, respectively. Hence, SSA can obtain the specific stylistic texture of the corresponding semantic regions of the style image to produce the stylized image with vivid stylistic textures. The intuitive comparison between our SCSA and universal attention can be found in Fig. 2.", + "bbox": [ + 511, + 90, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By seamlessly integrating SCSA into any Attn-AST approach without training, we enable arbitrary semantic style transfer while preserving and even enhancing the original stylization effect. To our knowledge, we are the first work to extend attention-based arbitrary style transfer to arbitrary semantic style transfer in a plug-and-play way.", + "bbox": [ + 511, + 378, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In brief, our contributions are divided into three parts:", + "bbox": [ + 532, + 469, + 887, + 483 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal that Attn-AST methods struggle with content and style images with the same semantics and identify the root causes for their subpar performance.", + "- We propose a semantic continuous-sparse attention, dubbed SCSA, that can extend Attn-AST to arbitrary semantic style transfer in a plug-and-play manner.", + "- We conduct extensive experiments and comparisons to demonstrate the effectiveness and generalization of SCSA, enabling Attn-AST methods (CNN-based, Transformer-based, and Diffusion-based approaches) to perform arbitrary semantic style transfer while preserving and even enhancing the original stylization." + ], + "bbox": [ + 532, + 484, + 903, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 679, + 653, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing AST approaches can be roughly divided into two groups: statistics-based and attention-based methods.", + "bbox": [ + 511, + 704, + 903, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Statistics-based Methods. Statistics-based AST methods [1, 18, 23, 43, 49] regard style as the statistical traits inherent in feature distribution. Specifically, they achieve stylized features by aligning the feature statistics of the content image with those of the style image, making them a global style transfer approach. The two most representative approaches are AdaIN [16] and WCT [24], which match the first-order and second-order statistics of features, respectively. To more comprehensively capture the stylistic characteristics of features, higher-order statistical methods, such as CMD [19] and EFDM [48], have been continually ex", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13052", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a8a070f469eff37ca8ddaf7419da0c7d5864bf20d620a954475bf76ed0ff2d75.jpg", + "image_caption": [ + "Figure 2. Comparison between Semantic Continuous-Sparse Attention (SCSA) and Universal Attention (UA). SCSA includes two parts: (a) Semantic Continuous Attention (SCA): The query point of the content semantic map features can match all continuous key points of the style semantic map features in the same semantic region. Therefore, SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of regions with the same semantics; (b) Semantic Sparse Attention (SSA): The query point of the content image features can match the most similar sparse key point of the style image features in the same semantic region. Hence, SSA can intently concentrate on the specific stylistic texture of regions with the same semantics. In contrast, (c) Universal Attention (UA) [32]: The query point of the content image features pays attention to all key points of the style image features. As a result, UA fails to accurately convey the intricate overall stylistic characteristics and specific textures of regions that share identical semantics." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 354, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a91ff56b4b3596a3de4e931eaa36fb179c120abd4b24789adea2f56ea1cce33b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 88, + 630, + 226 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4c153bda1c4d810bec18fc672a835f943d801afa7446fb4cace1a6be8a3577a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 88, + 906, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plored and refined. Expanding on the above foundation, semantic style transfer [10, 20, 26, 29, 30, 50] has been studied by aligning feature statistics within corresponding semantic regions. STROTSS [21] employs the Earth Movers Distance [22] to reduce the feature distribution difference within semantic regions and self-similarity to preserve the semantics and spatial structure. MAST [17] aligns the manifold distributions of semantic regions to facilitate semantic style transfer while preserving the content structure.", + "bbox": [ + 88, + 362, + 480, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since feature statistics only provide a global representation, the stylized images they generate often lack fine local textures. To address this, we propose SCSA, which could yield a coherent overall style across adjacent regions by overlooking structure details, while producing vivid textures by emphasizing them via the attention mechanism.", + "bbox": [ + 88, + 503, + 482, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention-based Methods. Attention-based AST (Attn-AST) methods define style as the weighted convergence of all style features. Exactly, Attn-AST methods obtain the stylized feature of each content feature by weighting all the style features, where weights are derived from the correlation between the content feature and all style features obtained through the attention mechanism, making it a local structure style transfer. Attn-AST approaches mainly include CNN-based methods [27, 32], Transformer-based methods [9], and Diffusion-based methods [8]. SANet [32] is the pioneering work to integrate attention into AST, employing content features as query and style features as key and value to obtain stylized features via the attention mechanism. StyTr² [9] integrates the outputs of the content and style transformer [37] encoders into the transformer feature fusion module, obtaining stylized features. StyleID [8] injects the content query, style key, and style value, derived from DDIM inversion [34], into certain self-attention layers of U-Net [33] to merge style, achieving stylized features through multiple inversion steps. Even though they", + "bbox": [ + 91, + 598, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "continually improve the quality of stylized images, their performance remains poor on content and style images with identical semantics. In contrast, patch-based methods [3, 7, 12, 38, 45] achieve semantic style transfer by swapping the styles of patches with the most similar content structures. When the patch size in patch-based methods is 1, they can be roughly considered attention-based methods. Deep-Image-Analogy (DIA) [25] introduces nearest-neighbor field search to tackle the challenge of patch matching for images, which are visually distinct yet semantically similar. GLStyleNet [38] uses multi-level pyramid features for semantic optimization via patch matching, iteratively creating stylized images that meet semantic criteria. TR [40] utilizes the patch match to achieve precise semantic-guided and structure-preserving texture transfer.", + "bbox": [ + 511, + 362, + 903, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the above methods target local structural similarity, subtle changes in the content structure of adjacent regions can result in stylization discontinuity in the identical semantic region. To mitigate this concern, we introduce SCSA, which provides more stable and continuous stylization by fully comprehending semantics and transfers vivid stylistic textures by focusing on local structure. Most importantly, it can be easily integrated into existing frameworks, while other methods are either designed for specific networks or require multiple iterations for optimization.", + "bbox": [ + 511, + 590, + 905, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 758, + 604, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Semantic Continuous-Sparse Attention", + "text_level": 1, + "bbox": [ + 511, + 786, + 848, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Revisit Universal Attention in Attn-AST. Attn-AST approaches feed the encoded content feature and style feature into Universal Attention (UA), constructing intricate connections between various positions within the two feature sequences by directly utilizing the self-attention [2] computational paradigm, to capture the long-range local textures", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c3af095416a55cda99882831fec17124c534a0f2d7f8ae2ddea7d2cd03a5147.jpg", + "image_caption": [ + "Figure 3. Detailed Procedure of Semantic Continuous-Sparse Attention (SCSA). In (a), S-AdaIN denotes that AdaIN [16] is applied individually for each semantic region. It initializes $F_{c}$ , by matching the feature statistics of corresponding semantic regions in $F_{c}$ and $F_{s}$ , to ensure that $F_{c}$ has semantically aligned color style information to some extent. Subsequently, $F_{c}$ can provide a more accurate query $f_{q}(\\bar{F}_{c})$ and more pure content structure features with less interference from the original color style. SCSA in (a) includes Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA) two parts. \"G1\" in SCA sets the value of the specific product point generated by multiplying two points from different semantic categories to negative infinity. As shown in (b), only the product of points with the same semantics is retained. Therefore, SCA can fully account for the overall stylistic characteristics of regions with the same semantics. \"G2\" in SSA retains only the maximum value of the product of a specified query point with all key points in the same semantic region, while all other values are set to negative infinity, as shown in (c). Hence, SSA can intently concentrate on the specific stylistic texture of the most similar structure in the regions with the same semantics. After SCSA, the generated $F_{cs}$ from (a) not only reflects the overall stylistic characteristics of the semantic region corresponding to $F_{s}$ but also captures the specific textures of that semantic region. Meanwhile, $F_{cs}$ also includes the content structure of $F_{c}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 256, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d186ca292dded83a9b23844c0e0538c4c6b7733c94a334f8aba382c3959e2180.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 89, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relationship. Typically, given the encoded content features $F_{c}$ and style features $F_{s}$ , the query (Q), key (K), and value (V) of UA can be obtained as follows:", + "bbox": [ + 89, + 616, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = f _ {q} \\left(\\bar {F} _ {c}\\right), K = f _ {k} \\left(\\bar {F} _ {s}\\right), V = f _ {v} \\left(F _ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 672, + 482, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{F}_c$ and $\\bar{F}_s$ denote the normalized $F_{c}$ and $F_{s}$ , respectively. $f_{q}, f_{k}$ , and $f_{v}$ are the project networks for getting query, key, and value in UA of the given Attn-AST model.", + "bbox": [ + 89, + 699, + 482, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, the attention map $S$ , which represents a holistic relationship between each content query point of the content features and all style key points of the style features, can be obtained as follows:", + "bbox": [ + 89, + 746, + 483, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = Q ^ {\\mathsf {T}} \\otimes K, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 814, + 482, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes matrix multiplication.", + "bbox": [ + 89, + 844, + 352, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the stylized features $F_{cs}$ can be calculated as:", + "bbox": [ + 109, + 859, + 465, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = f _ {o} (\\text {s o f t m a x} (S) \\otimes V) + F _ {c}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{o}$ is the output project networks in UA of the specified Attn-AST model.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although Attn-AST methods utilizing UA can achieve high-quality stylization results, they still underperform on content and style images with identical semantics due to inherent limitations, as detailed analysis in Sec. 1.", + "bbox": [ + 511, + 121, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Continuous Attention. To resolve the problem of the style discontinuity in adjacent regions within the identical semantic regions of the stylized image, as well as style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we propose semantic continuous attention, termed SCA, as shown in Fig. 2 (a) and detailed in Fig. 3. This method disregards the image structures while considering the relationship between local regions and semantic regions.", + "bbox": [ + 511, + 181, + 906, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, SCA first takes the encoded content semantic map features $F_{csem}$ and style semantic map features $F_{ssem}$ as inputs for the query $(Q_1)$ and key $(K_1)$ , respectively, while the encoded style image features $F_s$ are used as inputs for the value $(V_1)$ :", + "bbox": [ + 511, + 318, + 905, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {1} = f _ {q} (\\bar {F} _ {c s e m}), K _ {1} = f _ {k} (\\bar {F} _ {s s e m}), V _ {1} = f _ {v} (F _ {s}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 401, + 903, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{F}_{csem}$ and $\\bar{F}_{ssem}$ represent the normalized form of $F_{csem}$ and $F_{ssem}$ , respectively. $f_{q}, f_{k}$ , and $f_{v}$ are the same as those in Eq. 1.", + "bbox": [ + 511, + 426, + 905, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the preliminary attention map $\\mathcal{A}$ can be obtained:", + "bbox": [ + 532, + 473, + 901, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = Q _ {1} ^ {\\mathrm {T}} \\otimes K _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 494, + 903, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes matrix multiplication. Since $F_{csem}$ and $F_{ssem}$ do not encompass the content structure of the content and style images, after attention computation, the correlations between content points within the same semantic region and all style points are identical. That is if the content query points $i$ and $j$ belong to the same semantic category, then $\\mathcal{A}_{im} = \\mathcal{A}_{jm}$ , where $m$ represents any style key point. Consequently, the issue of stylization inconsistency in adjacent areas of the same semantic region, stemming from structural differences, is effectively addressed.", + "bbox": [ + 511, + 522, + 905, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make SCA focus on regions with the same semantics to transfer a coherent overall style, we modulate the attention map $\\mathcal{A}$ with operation $G_{1}$ , setting attention weights of two points from different semantic categories to negative infinity and those within the same category unchanged:", + "bbox": [ + 511, + 674, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathcal {A}} = G _ {1} (\\mathcal {A}),\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 757, + 751, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {1} \\left(\\mathcal {A} _ {q k}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {A} _ {q k}, \\text {i f} q \\in c \\text {a n d} k \\in c \\\\ - \\infty , \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 776, + 903, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q$ and $k$ represent the query and key point of $Q_{1}$ and $K_{1}$ , respectively. $c$ is the semantic category. $\\bar{A}$ is the new attention map that captures the relationship exclusively between content query points and style key points of the same semantic category. This is visually shown in Fig. 3 (b).", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13054", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ultimately, we can obtain the stylized features $F_{sca}$ :", + "bbox": [ + 109, + 90, + 455, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s c a} = f _ {o} \\left(\\text {s o f t m a x} (\\bar {\\mathcal {A}}) \\otimes V _ {1}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 118, + 482, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{o}$ is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of different content points and all style points in the same semantic region are equal. Thus, each query point can match all continuous key points in the same semantic region. SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of the same semantic regions in the style image while maintaining style consistency in adjacent areas of the identical semantic region in the stylized image.", + "bbox": [ + 89, + 148, + 483, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Adaptive Instance Normalization. To provide more pure content structure features with less interference from the original color style and a more accurate query for computing the relationships of local structures in corresponding semantic regions of the content and style images, we apply AdaIN [16] to the encoded content image features $F_{c}$ and style image features $F_{s}$ for each semantic region:", + "bbox": [ + 89, + 316, + 483, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c s e m} \\xrightarrow {K M} M _ {c} ^ {i}, I _ {s s e m} \\xrightarrow {K M} M _ {s} ^ {i}, i = 1, \\dots , N,\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 433, + 439, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c} ^ {i} = F _ {c} ^ {i} \\odot M _ {c} ^ {i}, \\quad F _ {s} ^ {i} = F _ {s} ^ {i} \\odot M _ {s} ^ {i}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 457, + 480, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c} ^ {i} = \\frac {F _ {c} ^ {i} - \\mu (F _ {c} ^ {i})}{\\sigma (F _ {c} ^ {i})} \\times \\sigma (F _ {s} ^ {i}) + \\mu (F _ {s} ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 479, + 403, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $KM$ refers to K-Means clustering [28]. $M_c^i$ and $M_s^i$ are the masks containing only 0 and 1. $F_c^i$ and $F_s^i$ represent the features belong to corresponding semantic category $i$ . Then, we obtain the features $F_c = \\sum_i F_c^i$ whose overall styles of relevant semantic regions align with those of $F_s$ .", + "bbox": [ + 89, + 523, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Sparse Attention. To preserve the original vivid stylistic texture of the style image and address the issue of style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we introduce semantic sparse attention, known as SSA, as shown in Fig. 2 (b) and Fig. 3. This method focuses solely on preserving the style texture with the highest weights, rather than weighting all the style textures, while considering the relationship between local and semantic regions.", + "bbox": [ + 89, + 601, + 483, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SSA first employs the transformed content features $F_{c}$ as inputs for the query $(Q_{2})$ and the encoded style image features $F_{s}$ as inputs for both the key $(K_{2})$ and value $(V_{2})$ :", + "bbox": [ + 89, + 737, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {2} = f _ {q} \\left(\\bar {F} _ {c}\\right), K _ {2} = f _ {k} \\left(\\bar {F} _ {s}\\right), V _ {2} = f _ {v} \\left(F _ {s}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 795, + 482, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{F}_c$ and $\\bar{F}_s$ represent the normalized $F_{c}$ and $F_{s}$ , respectively. $f_{q}, f_{k}$ , and $f_{v}$ are the same as those in Eq. 1.", + "bbox": [ + 89, + 824, + 482, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the initial attention map $\\mathcal{B}$ of SSA can be obtained:", + "bbox": [ + 109, + 856, + 482, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} = Q _ {2} ^ {\\mathrm {T}} \\otimes K _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 883, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes matrix multiplication. Since $F_{c}$ and $F_{s}$ capture the structures of the content and style images, $\\mathcal{B}$ can consider the feature correlations based on local structures.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We claim that the most accurate representation of the original stylistic texture resides in the encoded discrete feature point and the weighted style feature points struggle to preserve the original vivid stylistic textures. Thus, to get accurate and vivid textures, we adjust attention map $\\mathcal{B}$ with operation $G_{2}$ to target the discrete feature point within related regions, setting the only maximum attention weight of a content query point and all style key points in the same semantic category unchanged and others to negative infinity:", + "bbox": [ + 511, + 136, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\boldsymbol {B}}} = G _ {2} (\\mathcal {B}),\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 277, + 750, + 296 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {2} \\left(\\mathcal {B} _ {q k}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {B} _ {q k}, \\text {i f} \\mathcal {B} _ {q k} = \\max \\left(\\mathcal {B} _ {q n}\\right), \\\\ n \\in \\{1, \\dots , \\operatorname {l e n} \\left(K _ {2}\\right) \\}, q, k, n \\in c \\\\ - \\infty , \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 299, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q$ represents the query point of $Q_{2}$ . $k$ and $n$ represent the key points of $K_{2}$ . $c$ denotes the semantic category. $\\bar{\\mathcal{B}}$ is the new attention map that can grasp the relationship between the content and style feature points, which share the most similar structure. This is visually shown in Fig. 3 (c).", + "bbox": [ + 511, + 363, + 905, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eventually, we can obtain the stylized features $F_{ssa}$ :", + "bbox": [ + 531, + 439, + 879, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s s a} = f _ {o} \\left(\\text {s o f t m a x} (\\bar {\\mathcal {B}}) \\otimes V _ {2}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 460, + 903, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{o}$ is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of each content point and the most similar style point in the same semantic region are 1 and weights of this content point and other style points in the same semantic region are 0. Thus, each query point can match sparse key points in the same semantic region. SSA can intently concentrate on the specific stylistic texture of regions with the same semantics to transfer them to the matching semantic region in the stylized image.", + "bbox": [ + 511, + 484, + 905, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Fusion. With the $F_{sca}$ and $F_{ssa}$ of SCA and SSA, the feature fusion can be executed:", + "bbox": [ + 511, + 651, + 905, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = \\alpha_ {1} \\times F _ {s c a} + \\alpha_ {2} \\times F _ {s s a} + F _ {c}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 689, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ and $\\alpha_{2}$ denote the stylization degree of the overall style and vivid textures separately. $F_{c}$ are the outputs of Eq. 8 for superior styling. $F_{cs}$ are the final stylized features.", + "bbox": [ + 511, + 710, + 905, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Overall Framework", + "text_level": 1, + "bbox": [ + 511, + 763, + 702, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Revisit Attn-AST Approaches. Given a content image $I_{c}$ and a style image $I_{s}$ , the Attn-AST model $M$ could produce a stylized image $I_{cs}$ , which preserves the content of $I_{c}$ and exhibit the style of $I_{s}$ .", + "bbox": [ + 511, + 786, + 905, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, they achieve the encoded content features $F_{c}$ and style features $F_{s}$ through the encoder $E$ of $M$ :", + "bbox": [ + 511, + 847, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c} = E \\left(I _ {c}\\right), F _ {s} = E \\left(I _ {s}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 883, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e375e3af126ce2b966761f4fba7e797de8b226be5718ef2bb3a72f1df267f0a2.jpg", + "image_caption": [ + "Figure 4. The overall frameworks of Attn-AST approaches-CNN-based, Transformer-based, and Diffusion-based methods with semantic continuous-sparse attention (SCSA). SCA depicts semantic continuous attention. SSA denotes semantic sparse attention. S-AdaIN is semantic adaptive instance normalization. UA expresses universal attention in the Attn-AST methods. The dashed line in (b) shows encoded content features $F_{c}$ are only used in the first use of the feature transformation module, and the output features $F_{cs}$ are used as new content features in subsequent transformation. The dashed line in (c) indicates S-AdaIN is used only when $t$ is the maximum time step $T$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 330, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d1a312e4e3a062108d6a8aa70b297b6d08fd5a0cb7faf5e88a1e23e60269064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 88, + 604, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d5b85a6a727b3a716728cb0fe396c7792c9c2103ab3d4962fee85755ccc32b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 85, + 893, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, they use the feature transformation module $T$ with universal attention (UA) to obtain the stylized features $F_{cs}$ :", + "bbox": [ + 89, + 340, + 482, + 372 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = T _ {U A} \\left(F _ {c}, F _ {s}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 380, + 482, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eventually, they could acquire the stylized image with the decoder $D$ of $M$ :", + "bbox": [ + 89, + 406, + 483, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c s} = D \\left(F _ {c s}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 445, + 482, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I_{cs}$ possesses the content of $I_{c}$ and the style of $I_{s}$ , but the style of its corresponding semantic region is inconsistent with that of the style image.", + "bbox": [ + 89, + 470, + 483, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. Despite the three different Attn-AST frameworks (CNN, Transformer, and Diffusion), their pipelines remain largely consistent with the above, differing only in model structures. The CNN-based method uses a simple encoder-decoder setup, while the Transformer-based approach employs two distinct encoders to process content and style separately. In contrast, the diffusion-based method also utilizes a single encoder-decoder structure but incorporates a time-step iteration mechanism for further refinement.", + "bbox": [ + 89, + 517, + 483, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attn-AST Approaches with Our SCSA. SCSA is a plug-and-play method that can be incorporated into any Attn-AST method. Thus, given a content image $I_{c}$ with its semantic map $I_{csem}$ , a style image $I_{s}$ with its semantic map $I_{ssem}$ , and a specified Attn-AST model $M$ , our objective is that $M_{SCSA}$ , the approach obtained after incorporating SCSA into $M$ , can generate a stylized image $I_{cs}$ using the quadruple $\\{I_{c}, I_{csem}, I_{s}, I_{ssem}\\}$ , exhibiting a consistent style with the style image within the same semantic regions.", + "bbox": [ + 89, + 652, + 482, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At first, we obtain the four encoded features of the quadruple through the encoder $E$ :", + "bbox": [ + 89, + 787, + 483, + 818 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c} = E \\left(I _ {c}\\right), F _ {s} = E \\left(I _ {s}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 825, + 480, + 849 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s e m} = E (I _ {c s e m}), F _ {s e m} = E (I _ {s e m}).\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 847, + 415, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, we replace UA in the feature transformation module $T$ of Eq. 15 with SCSA, which could perform semantic", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "feature conversion by enabling each query point to consider certain key points in the corresponding semantic region:", + "bbox": [ + 511, + 340, + 906, + 372 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {c s} = T _ {S C S A} \\left(F _ {c}, F _ {s}, F _ {c s e m}, F _ {s s e m}\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 378, + 903, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_{cs}$ are stylized features that meet semantic criteria.", + "bbox": [ + 511, + 402, + 900, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lastly, we get the stylized image with the decoder $D$ :", + "bbox": [ + 531, + 419, + 887, + 434 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {c s} = D \\left(F _ {c s}\\right), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 441, + 903, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the style of the semantic region of the stylized image $I_{cs}$ is consistent with that of the style image $I_{s}$ , while the content of $I_{cs}$ is the same as that of the content image $I_{c}$ .", + "bbox": [ + 511, + 465, + 905, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The frameworks of the CNN-based, Transformer-based, and Diffusion-based Attn-AST methods with our SCSA are shown in Fig. 4. They differ only in the layers where UA is replaced with SCSA and in the encoder and decoder structures, while the other operations remain largely consistent.", + "bbox": [ + 511, + 510, + 905, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 597, + 645, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 511, + 622, + 705, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select three proxy Attn-AST approaches (CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]), and five state-of-the-art (SOTA) semantic stylization methods (STROTSS [21], MAST [17], TR [40], DIA [25], and GLStyleNet [38]) as baselines. We obtain images from prior research, datasets, and the Internet, and construct their semantic maps to generate quadruple data. More details are in the supplementary materials.", + "bbox": [ + 511, + 643, + 906, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 773, + 735, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive comparison experiments to evaluate the effectiveness and generalization of our SCSA.", + "bbox": [ + 511, + 795, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 5, Attn-AST approaches combined with our SCSA can generate high-quality and vivid stylized images that better meet semantic needs compared to traditional Attn-AST methods, showcasing rich textures and vibrant colors in corresponding regions, e.g., the background", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13056", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7992688d78e741a42f957464972fac140e56acb80298280e7d258b12c9dd4495.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons among Attn-AST approaches, those with SCSA, and SOTA methods." + ], + "image_footnote": [], + "bbox": [ + 109, + 87, + 893, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ed1d6d2a8aa5156c8bee855b4c1ec214c358aa508fbe414f01dc636be6e343f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SANetSANet + SCSAStyTr2StyTr2 + SCSAStyleIDStyleID + SCSASTROTSSMASTTRDIAGLStyleNet
SSL ↓1.65830.87621.98261.22281.75381.24471.09811.73201.26311.93981.0305
FID ↓14.338513.078812.527312.396312.594412.449712.840016.516313.584620.794214.4880
CFSD ↓0.11030.08740.07520.07050.09160.11780.10080.07460.11490.11470.3991
Pref. ↑0.16850.83150.15760.84240.21920.78080.18670.11000.04500.01670.1183
", + "bbox": [ + 117, + 320, + 879, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons of Attn-AST approaches, those with SCSA, and several SOTA semantic style transfer methods. The best results for Attn-AST methods and Attn-AST methods with SCSA are in bold, while the best results among all methods are underlined.", + "bbox": [ + 89, + 388, + 906, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in the 1st row, along with the sky and ground in the 2nd row. Likewise, compared to SOTA semantic style transfer methods, the stylized images generated by SCSA demonstrate superior semantic accuracy and content preservation, e.g. the river in the 3rd row and the horses in the 4th row.", + "bbox": [ + 89, + 433, + 482, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, SCSA is a plug-and-play solution that can be integrated into Attn-AST methods, addressing their semantic issues and offering a level of capability in semantic style transfer that surpasses that of existing SOTA methods.", + "bbox": [ + 89, + 508, + 483, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 577, + 323, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct quantitative analyses for full validation.", + "bbox": [ + 89, + 599, + 465, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Style Loss (SSL). Traditional style loss [16] is unsuitable for semantic style transfer. There are two main reasons: First, it focuses on the overall region, making it difficult to capture the semantic region style; second, the inconsistency in the sizes of the semantic regions between the style and stylized images affects the calculation of global style loss. Hence, we propose a semantic style loss to calculate each region style loss similar to Eq. 8 using the style loss of [16]. From the 1st row of Tab. 1, it is clear that the Attn-AST methods utilizing SCSA not only achieve the lowest SSL compared to traditional Attn-AST methods, but when SANet incorporates SCSA, it also reaches the best performance among all semantic style transfer methods. This indicates that SCSA can effectively facilitate semantic style transfer with its effectiveness and generalization.", + "bbox": [ + 88, + 613, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Frechet Inception Distance (FID). Similar to [8], we use FID [13] to evaluate the overall style fidelity between the stylized and style images. As indicated by the 2nd row of Tab. 1, the Attn-AST methods integrated with SCSA", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieve the lowest FID value as opposed to typical Attn-AST methods. Particularly, $\\mathrm{StyTr}^2$ with SCSA secures the optimal FID value among all the semantic style transfer methods. The above indicates that SCSA not only conducts the semantic style transfer but also markedly improves the overall style quality of the stylized images.", + "bbox": [ + 511, + 433, + 906, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Content Feature Structural Distance (CFSD). As discussed in [8], content loss [16] and LPIPS [47] are affected by the stylization. Thus, we use the content feature structural distance [8] to assess content retention. As shown in the $3rd$ row of Tab. 1, the SCSA achieves the lowest CFSD value on CNN-based and Transformer-based methods. Although the CFSD value is slightly higher on the Diffusion-based approach with SCSA, both quantitative and qualitative results indicate that SCSA greatly improves semantic stylization, making this trade-off worthwhile. Besides, StyTr² with SCSA has the lowest CFSD value in all SOTA methods, showcasing its content preservation advantage.", + "bbox": [ + 511, + 526, + 908, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. Our user study is divided into two parts: the first compares traditional Attn-AST methods with SCSA-enhanced ones, and the second evaluates Attn-AST methods with SCSA against existing semantic style transfer methods. As shown in the 4th row of Tab. 1, the stylized images of the Attn-AST methods with SCSA are more favored by the public, and their popularity surpasses that of the SOTA methods. These demonstrate the superiority of SCSA.", + "bbox": [ + 511, + 709, + 908, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 845, + 666, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments to verify the crucial roles of SCA and SSA and to affirm the potency of S-AdaIN.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13057", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89f979d59d5b8f7f52558385078c333c94e608349fad7cadaf7fc7acb805a650.jpg", + "image_caption": [ + "Figure 6. Qualitative and quantitative ablation results on SCA, SSA, and S-AdaIN. Zoom in for better comparison." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 897, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/add8873588041f0364df3f3d211a37776aeb2194afdaf7a159209f12cc7a1f15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 257, + 473, + 357 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5edb33b049b1dc3d057d390562a26c3a8e38eefcfe032dd9c95bed0411116fe2.jpg", + "table_caption": [ + "Figure 7. The stylization of different content representation." + ], + "table_footnote": [], + "table_body": "
UASCASSASCSA
Runtime (milliseconds)1.2311.9172.0153.944
Memory Usage (bytes)4708479649726220
", + "bbox": [ + 102, + 386, + 468, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Computation costs for feature fusion modules.", + "bbox": [ + 120, + 441, + 450, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of SCA and SSA. We use SCA or SSA individually as the feature transformation module to assess their capabilities. As shown in Fig. 6, SCA alone facilitates overall color transfer in corresponding semantic regions but results in texture loss and reduced stylization, as reflected by the higher SSL and FID values. In contrast, SSA can effectively transfer specific textures but introduces color discrepancies between the semantic regions of the stylized and style images. This indicates that while SCA captures the overall style characteristics, SSA is crucial for transferring specific textures, highlighting the complementary roles of both.", + "bbox": [ + 88, + 470, + 482, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of S-AdaIN. To validate the effectiveness of S-AdaIN, we conducted ablation experiments by removing S-AdaIN. As shown in Fig. 6, without S-AdaIN, the stylized images exhibit less accurate global style transfer in semantic regions and reduced texture details, particularly evident for the StyleID method. Although S-AdaIN was not employed, CFSD values of StyTr² and StyleID with SCSA are lower, which comes at the expense of stylization quality. Therefore, after considering the pros and cons, S-AdaIN is essential and effective for semantic style transfer.", + "bbox": [ + 89, + 638, + 482, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 801, + 209, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform more discussions about our task and method.", + "bbox": [ + 89, + 824, + 470, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Arbitrary Semantic Style Transfer. Our plug-and-play SCSA is built on the Attn-AST approaches, unlike methods that iteratively optimize the specific content and style images or only apply a predefined style to any content im", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age, so that it can stylize between any content and style images. Besides, it is important to note that SCSA leverages the user-provided semantic maps to guide the transfer process. Therefore, as long as the user's semantic maps mark the regions where the style should be applied, we can perform the relevant region stylization, as shown in Fig. 7.", + "bbox": [ + 511, + 261, + 903, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computation Costs. Since the costs vary with the feature size, we use the features of the $512 \\times 512$ images encoded by ReLU4_1 layer of VGG-19 to discuss costs. The results are in Tab. 2. In the 1st row, SCA and SSA have slightly longer runtimes than UA due to the guidance operations. SCSA takes barely longer than the sum of SCA and SSA, as it merges both outputs. The 2nd row shows SCA and SSA use a bit more memory than UA, while SCSA uses less than SCA plus SSA. This is because some operations reduce memory usage by reusing existing variables. Despite the cost increases, this is acceptable and does not impact practical use. Cost refinements are left for future work.", + "bbox": [ + 511, + 352, + 906, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 547, + 633, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we highlight the challenge faced by existing arbitrary style transfer methods, with a particular emphasis on attention-based approaches, and provide an in-depth analysis of this issue, which is the failure to account for the relationship between local regions and semantic regions. To address this issue and generate high-quality stylized images that meet semantic needs, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA. It achieves precise semantic style transfer by allowing each query point to consider certain key points in the corresponding semantic region. Extensive experimental results demonstrate that our SCSA is effective and generalized.", + "bbox": [ + 511, + 573, + 906, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 768, + 671, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partly supported by the National Natural Science Foundation of China No. 62306277, the Fundamental Research Funds for the Zhejiang Provincial Universities Grant No. K20240090, the Research Fund for International Scientists of National Natural Science Foundation of China No. 72350710798, and the Natural Science Foundation of Zhejiang Province No. LDT23F02023F02. This", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "research was supported in part through funding from the following sources: the National Natural Science Foundation of China under Award No. 62306277, and the Fundamental Research Funds for Provincial Universities in Zhejiang Province through Grant No. K20240090.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 180, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 1, 2", + "[2] Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3", + "[3] Alex J Champandard. Semantic style transfer and turning two-bit doodles into fine artworks. arXiv preprint arXiv:1603.01768, 2016. 3", + "[4] Dar-Yen Chen. Artfusion: Arbitrary style transfer using dual conditional latent diffusion models. arXiv preprint arXiv:2306.09330, 2023. 1", + "[5] Dar-Yen Chen, Hamish Tennent, and Ching-Wen Hsu. Art-adapter: Text-to-image style transfer using multi-level style encoder and explicit adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8619-8628, 2024. 2", + "[6] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2", + "[7] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 3", + "[8] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 1, 2, 3, 6, 7", + "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 1, 2, 3, 6", + "[10] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3730-3738. IEEE. 3", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 1", + "[12] Mark Hamazaspyan and Shant Navasardyan. Diffusion-enhanced patchmatch: A framework for arbitrary style transfer with diffusion models. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 205, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 797-805, 2023. 3", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[14] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 2", + "[15] Siyu Huang, Jie An, Donglai Wei, Jiebo Luo, and Hanspeter Pfister. Quantart: Quantizing image style transfer towards high visual fidelity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5947-5956, 2023. 1", + "[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4, 5, 7", + "[17] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14861-14869, 2021. 3, 6", + "[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence, pages 4369-4376, 2020. 2", + "[19] Nikolai Kalischek, Jan D Wegner, and Konrad Schindler. In the light of feature distributions: moment matching for neural style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9391, 2021. 2", + "[20] Da-Hyun Kim, Dong-Hee Shin, and Tae-Eui Kam. Bridging the bci illiteracy gap: a subject-to-subject semantic style transfer for eeg-based motor imagery classification. Frontiers in Human Neuroscience, 17:1194751, 2023. 3", + "[21] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 3, 6", + "[22] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In International conference on machine learning, pages 957-966. PMLR, 2015. 3", + "[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer, 2018. 1, 2", + "[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13059", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. arXiv preprint arXiv:1705.01088, 2017. 3, 6", + "[26] Yi-Sheng Liao and Chun-Rong Huang. Semantic context-aware image style transfer. IEEE Transactions on Image Processing, 31:1911-1923, 2022. 3", + "[27] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 1, 2, 3", + "[28] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5", + "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5951-5960. IEEE, .3", + "[30] Ming Lu, Hao Zhao, Anbang Yao, Feng Xu, Yurong Chen, and Li Zhang. Decoder network over lightweight reconstructed feature for fast semantic style transfer. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2488-2496. IEEE, .3", + "[31] Yingnan Ma, Chenqiu Zhao, Xudong Li, and Anup Basu. Rast: Restorable arbitrary style transfer via multi-restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 331-340, 2023. 2", + "[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 1, 2, 3, 6", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[35] Onkar Susladkar, Gayatri Deshmukh, Sparsh Mittal, and Parth Shastri. D2styler: Advancing arbitrary style transfer with discrete diffusion methods. arXiv preprint arXiv:2408.03558, 2024. 2", + "[36] Gemma Canet Tarres, Dan Ruta, Tu Bui, and John Collosomasse. Parasol: Parametric style control for diffusion image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2432-2442, 2024. 1", + "[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[38] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Gliystylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 3, 6", + "[39] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: towards" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1095-1106, 2022. 1, 2", + "[40] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2624-2632, 2022. 3, 6", + "[41] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1", + "[42] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2", + "[43] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 1, 2", + "[44] Zhengtao Xiang, Xing Wan, Libo Xu, Xin Yu, and Yuhan Mao. A training-free latent diffusion style transfer method. Information, 15(10):588, 2024. 2", + "[45] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 3", + "[46] Chiyu Zhang, Xiaogang Xu, Lei Wang, Zaiyan Dai, and Jun Yang. S2wat: Image style transfer via hierarchical vision transformer using strips window attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7024-7032, 2024. 2", + "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[48] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 2", + "[49] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. A unified arbitrary style transfer framework via adaptive contrastive learning. ACM Trans. Graph., 42(5), 2023. 1, 2", + "[50] Hui-Huang Zhao, Paul L Rosin, Yu-Kun Lai, and Yao-Nan Wang. Automatic semantic style transfer using deep convolutional neural networks and soft masks. The Visual Computer, 36(7):1307-1324, 2020. 3" + ], + "bbox": [ + 516, + 92, + 906, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13060", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_model.json b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b462ed62a09e8d8fe2bc3b79c64306336638677d --- /dev/null +++ b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_model.json @@ -0,0 +1,2827 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "SCSA: A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.204, + 0.404, + 0.222 + ], + "angle": 0, + "content": "Chunnan Shang" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.223, + 0.42, + 0.241 + ], + "angle": 0, + "content": "Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.243, + 0.456, + 0.257 + ], + "angle": 0, + "content": "chunnan.22@intl.zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.258, + 0.403, + 0.275 + ], + "angle": 0, + "content": "Hongwei Wang" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.276, + 0.42, + 0.293 + ], + "angle": 0, + "content": "Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.296, + 0.461, + 0.31 + ], + "angle": 0, + "content": "hongweiwang@intl.zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.204, + 0.719, + 0.223 + ], + "angle": 0, + "content": "Zhizhong Wang†" + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.223, + 0.726, + 0.241 + ], + "angle": 0, + "content": "Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.243, + 0.724, + 0.257 + ], + "angle": 0, + "content": "endywon@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.576, + 0.257, + 0.726, + 0.275 + ], + "angle": 0, + "content": "Xiangming Meng†" + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.276, + 0.726, + 0.293 + ], + "angle": 0, + "content": "Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.296, + 0.775, + 0.31 + ], + "angle": 0, + "content": "xiangmingmeng@intl.zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.485, + 0.771 + ], + "angle": 0, + "content": "Attention-based arbitrary style transfer methods, including CNN-based, Transformer-based, and Diffusion-based, have flourished and produced high-quality stylized images. However, they perform poorly on the content and style images with the same semantics, i.e., the style of the corresponding semantic region of the generated stylized image is inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. To address this issue, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer—each query point considers certain key points in the corresponding semantic region. Specifically, semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region that reflect the overall style characteristics of that region; Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region that exhibits the specific stylistic texture of that region. By combining the two modules, the resulting SCSA aligns the overall style of the corresponding semantic regions while transferring the vivid textures of these regions. Qualitative and quantitative results demonstrate that SCSA enables attention-based arbitrary style transfer methods to produce high-quality semantic stylized images. The code can be found in https://github.com/ecn-00/SCSA." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.802, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Initially, style transfer [11] was proposed to preserve the content of a specified content image while transferring the style from a specified style image by iteratively optimizing" + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.346, + 0.626, + 0.357 + ], + "angle": 0, + "content": "Content Inputs" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.357, + 0.627, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.423, + 0.625, + 0.434 + ], + "angle": 0, + "content": "Content Inputs" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.346, + 0.7, + 0.357 + ], + "angle": 0, + "content": "Style Inputs" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.357, + 0.709, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.422, + 0.699, + 0.435 + ], + "angle": 0, + "content": "Style Inputs" + }, + { + "type": "text", + "bbox": [ + 0.733, + 0.346, + 0.768, + 0.356 + ], + "angle": 0, + "content": "SANet" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.357, + 0.79, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.422, + 0.767, + 0.434 + ], + "angle": 0, + "content": "StyTr2" + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.346, + 0.871, + 0.356 + ], + "angle": 0, + "content": "SANet + SCSA" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.357, + 0.871, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.422, + 0.87, + 0.434 + ], + "angle": 0, + "content": "\\(\\mathrm{StyTr}^2 + \\mathrm{SCSA}\\)" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.435, + 0.628, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.499, + 0.625, + 0.511 + ], + "angle": 0, + "content": "Content Inputs" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.435, + 0.709, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.499, + 0.698, + 0.511 + ], + "angle": 0, + "content": "Style Inputs" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.435, + 0.79, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.5, + 0.769, + 0.511 + ], + "angle": 0, + "content": "StyleID" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.435, + 0.871, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.499, + 0.871, + 0.51 + ], + "angle": 0, + "content": "StyleID + SCSA" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.511, + 0.628, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.511, + 0.709, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.511, + 0.79, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.511, + 0.871, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.69 + ], + "angle": 0, + "content": "Figure 1. Comparisons of the Attn-AST approaches-CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]-without and with our SCSA. The stylized images generated by the Attn-AST approaches exhibit style discontinuity in adjacent regions within the identical semantic regions (e.g., the background in the 1st row), style inconsistency between corresponding semantic (e.g., the cloth color in the 3rd row), and fewer textures (e.g., the block textures in the 3rd row)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "the generated stylized image. However, its excessive slowness prevents practical application. Given the need for efficiency, effectiveness, and generalization, Arbitrary Style Transfer (AST) [1, 4, 8, 23, 27] has become a highly prominent research focus in the style transfer field. AST seeks to generate a new image that replicates the content of an arbitrary content image while adopting the style of an arbitrary style image. In spite of many recent studies aiming to advance the AST effect by incorporating novel knowledge [15, 36, 39] and innovative constraints [41, 43, 49], the core of AST continues to reside in feature transformation modules, which remains indispensable for all approaches. Hence, the existing AST approaches can be classified into" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13051" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "two categories, statistics-based methods [5, 16, 24, 48] and attention-based methods [6, 8, 9, 27, 31, 32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Attention-based Arbitrary Style Transfer (Attn-AST) approaches, leveraging the ability to capture global information and focus on the related features of the attention mechanism [2], preserve the content of the content image while transferring the style that is relevant to the content of the content image from the style image. They encompass CNN-based [27, 31, 32, 39], Transformer-based [9, 14, 42, 46], and Diffusion-based methods [8, 35, 44]. Commonly, Attn-AST methods first establish a holistic relationship, an attention map, between each content query point of the content image features and all style key points of the style image features. Then, Attn-AST methods get the transferred style for each content point by weighting all the style points, where the weights are the softmax function outputs of a naturally calculated attention map of content and style points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Although Attn-AST methods have thrived and produced high-quality stylized images due to the capacity mentioned above for global capture and focus, they perform poorly on the content and style image with the same semantics. As shown in Fig. 1, the style of the corresponding semantic region of the stylized image generated by Attn-AST approaches exhibits a style inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. (1) During the formulation of the attention map, Attn-AST methods bluntly and simplistically consider the relationship between each query point and all key points, as shown in Fig. 2 (c), without adequately accounting for the semantic regions to which these points belong. This oversimplification may cause Attn-AST methods to disproportionately emphasize structurally similar points across different semantic regions, resulting in artifacts or inaccurate style [27, 32], e.g., the \\(2nd\\) row of Fig. 1. (2) Attn-AST approaches fundamentally depend on structural principles for generating a global style, as they utilize content and style image features. The subtle disparity in content structure details across adjacent regions can cause different stylization effects. Thus, the stylized images often lack stylistic continuity across adjacent regions within the semantic regions, e.g., the background in the \\(1st\\) row of Fig. 1. (3) Attn-AST approaches acquire the transferred style by weighting all the style points. However, the most accurate representation of the original stylistic texture resides in the encoded single points. The weighted style struggles to preserve the original stylistic texture, leading to a lack of vividness, e.g., the block textures in the \\(3rd\\) row of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To maintain the benefits of Attn-AST methods while addressing their existing issues, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer. The key insight of SCSA is that, based on the interpretability of the attention mecha" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.378 + ], + "angle": 0, + "content": "nism, each query point considers certain key points in the corresponding semantic region. Specifically, SCSA consists of two parts: Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA). (1) SCA: Semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region by using content and style semantic map features as query and key, respectively. Thus, SCA can achieve the overall style characteristics of the corresponding semantic regions of the style image to generate the stylized image with stylistic continuity across adjacent regions within the semantic regions. (2) SSA: Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region by using content and style image features as query and key, respectively. Hence, SSA can obtain the specific stylistic texture of the corresponding semantic regions of the style image to produce the stylized image with vivid stylistic textures. The intuitive comparison between our SCSA and universal attention can be found in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.469 + ], + "angle": 0, + "content": "By seamlessly integrating SCSA into any Attn-AST approach without training, we enable arbitrary semantic style transfer while preserving and even enhancing the original stylization effect. To our knowledge, we are the first work to extend attention-based arbitrary style transfer to arbitrary semantic style transfer in a plug-and-play way." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.47, + 0.888, + 0.484 + ], + "angle": 0, + "content": "In brief, our contributions are divided into three parts:" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.485, + 0.905, + 0.529 + ], + "angle": 0, + "content": "- We reveal that Attn-AST methods struggle with content and style images with the same semantics and identify the root causes for their subpar performance." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.53, + 0.905, + 0.575 + ], + "angle": 0, + "content": "- We propose a semantic continuous-sparse attention, dubbed SCSA, that can extend Attn-AST to arbitrary semantic style transfer in a plug-and-play manner." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.576, + 0.905, + 0.665 + ], + "angle": 0, + "content": "- We conduct extensive experiments and comparisons to demonstrate the effectiveness and generalization of SCSA, enabling Attn-AST methods (CNN-based, Transformer-based, and Diffusion-based approaches) to perform arbitrary semantic style transfer while preserving and even enhancing the original stylization." + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.485, + 0.905, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.655, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.734 + ], + "angle": 0, + "content": "The existing AST approaches can be roughly divided into two groups: statistics-based and attention-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Statistics-based Methods. Statistics-based AST methods [1, 18, 23, 43, 49] regard style as the statistical traits inherent in feature distribution. Specifically, they achieve stylized features by aligning the feature statistics of the content image with those of the style image, making them a global style transfer approach. The two most representative approaches are AdaIN [16] and WCT [24], which match the first-order and second-order statistics of features, respectively. To more comprehensively capture the stylistic characteristics of features, higher-order statistical methods, such as CMD [19] and EFDM [48], have been continually ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13052" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.355, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.631, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.907, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.234, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Figure 2. Comparison between Semantic Continuous-Sparse Attention (SCSA) and Universal Attention (UA). SCSA includes two parts: (a) Semantic Continuous Attention (SCA): The query point of the content semantic map features can match all continuous key points of the style semantic map features in the same semantic region. Therefore, SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of regions with the same semantics; (b) Semantic Sparse Attention (SSA): The query point of the content image features can match the most similar sparse key point of the style image features in the same semantic region. Hence, SSA can intently concentrate on the specific stylistic texture of regions with the same semantics. In contrast, (c) Universal Attention (UA) [32]: The query point of the content image features pays attention to all key points of the style image features. As a result, UA fails to accurately convey the intricate overall stylistic characteristics and specific textures of regions that share identical semantics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.363, + 0.482, + 0.499 + ], + "angle": 0, + "content": "plored and refined. Expanding on the above foundation, semantic style transfer [10, 20, 26, 29, 30, 50] has been studied by aligning feature statistics within corresponding semantic regions. STROTSS [21] employs the Earth Movers Distance [22] to reduce the feature distribution difference within semantic regions and self-similarity to preserve the semantics and spatial structure. MAST [17] aligns the manifold distributions of semantic regions to facilitate semantic style transfer while preserving the content structure." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Since feature statistics only provide a global representation, the stylized images they generate often lack fine local textures. To address this, we propose SCSA, which could yield a coherent overall style across adjacent regions by overlooking structure details, while producing vivid textures by emphasizing them via the attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.599, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Attention-based Methods. Attention-based AST (Attn-AST) methods define style as the weighted convergence of all style features. Exactly, Attn-AST methods obtain the stylized feature of each content feature by weighting all the style features, where weights are derived from the correlation between the content feature and all style features obtained through the attention mechanism, making it a local structure style transfer. Attn-AST approaches mainly include CNN-based methods [27, 32], Transformer-based methods [9], and Diffusion-based methods [8]. SANet [32] is the pioneering work to integrate attention into AST, employing content features as query and style features as key and value to obtain stylized features via the attention mechanism. StyTr² [9] integrates the outputs of the content and style transformer [37] encoders into the transformer feature fusion module, obtaining stylized features. StyleID [8] injects the content query, style key, and style value, derived from DDIM inversion [34], into certain self-attention layers of U-Net [33] to merge style, achieving stylized features through multiple inversion steps. Even though they" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.905, + 0.59 + ], + "angle": 0, + "content": "continually improve the quality of stylized images, their performance remains poor on content and style images with identical semantics. In contrast, patch-based methods [3, 7, 12, 38, 45] achieve semantic style transfer by swapping the styles of patches with the most similar content structures. When the patch size in patch-based methods is 1, they can be roughly considered attention-based methods. Deep-Image-Analogy (DIA) [25] introduces nearest-neighbor field search to tackle the challenge of patch matching for images, which are visually distinct yet semantically similar. GLStyleNet [38] uses multi-level pyramid features for semantic optimization via patch matching, iteratively creating stylized images that meet semantic criteria. TR [40] utilizes the patch match to achieve precise semantic-guided and structure-preserving texture transfer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.906, + 0.742 + ], + "angle": 0, + "content": "As the above methods target local structural similarity, subtle changes in the content structure of adjacent regions can result in stylization discontinuity in the identical semantic region. To mitigate this concern, we introduce SCSA, which provides more stable and continuous stylization by fully comprehending semantics and transfers vivid stylistic textures by focusing on local structure. Most importantly, it can be easily integrated into existing frameworks, while other methods are either designed for specific networks or require multiple iterations for optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.76, + 0.605, + 0.775 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.849, + 0.803 + ], + "angle": 0, + "content": "3.1. Semantic Continuous-Sparse Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Revisit Universal Attention in Attn-AST. Attn-AST approaches feed the encoded content feature and style feature into Universal Attention (UA), constructing intricate connections between various positions within the two feature sequences by directly utilizing the self-attention [2] computational paradigm, to capture the long-range local textures" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13053" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.258, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.09, + 0.484, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.485, + 0.601 + ], + "angle": 0, + "content": "Figure 3. Detailed Procedure of Semantic Continuous-Sparse Attention (SCSA). In (a), S-AdaIN denotes that AdaIN [16] is applied individually for each semantic region. It initializes \\( F_{c} \\), by matching the feature statistics of corresponding semantic regions in \\( F_{c} \\) and \\( F_{s} \\), to ensure that \\( F_{c} \\) has semantically aligned color style information to some extent. Subsequently, \\( F_{c} \\) can provide a more accurate query \\( f_{q}(\\bar{F}_{c}) \\) and more pure content structure features with less interference from the original color style. SCSA in (a) includes Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA) two parts. \"G1\" in SCA sets the value of the specific product point generated by multiplying two points from different semantic categories to negative infinity. As shown in (b), only the product of points with the same semantics is retained. Therefore, SCA can fully account for the overall stylistic characteristics of regions with the same semantics. \"G2\" in SSA retains only the maximum value of the product of a specified query point with all key points in the same semantic region, while all other values are set to negative infinity, as shown in (c). Hence, SSA can intently concentrate on the specific stylistic texture of the most similar structure in the regions with the same semantics. After SCSA, the generated \\( F_{cs} \\) from (a) not only reflects the overall stylistic characteristics of the semantic region corresponding to \\( F_{s} \\) but also captures the specific textures of that semantic region. Meanwhile, \\( F_{cs} \\) also includes the content structure of \\( F_{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.663 + ], + "angle": 0, + "content": "relationship. Typically, given the encoded content features \\( F_{c} \\) and style features \\( F_{s} \\), the query (Q), key (K), and value (V) of UA can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.673, + 0.483, + 0.691 + ], + "angle": 0, + "content": "\\[\nQ = f _ {q} \\left(\\bar {F} _ {c}\\right), K = f _ {k} \\left(\\bar {F} _ {s}\\right), V = f _ {v} \\left(F _ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.746 + ], + "angle": 0, + "content": "where \\(\\bar{F}_c\\) and \\(\\bar{F}_s\\) denote the normalized \\(F_{c}\\) and \\(F_{s}\\), respectively. \\(f_{q}, f_{k}\\), and \\(f_{v}\\) are the project networks for getting query, key, and value in UA of the given Attn-AST model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.806 + ], + "angle": 0, + "content": "After that, the attention map \\( S \\), which represents a holistic relationship between each content query point of the content features and all style key points of the style features, can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.815, + 0.483, + 0.834 + ], + "angle": 0, + "content": "\\[\nS = Q ^ {\\mathsf {T}} \\otimes K, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.845, + 0.353, + 0.859 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes matrix multiplication." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.86, + 0.466, + 0.874 + ], + "angle": 0, + "content": "Finally, the stylized features \\( F_{cs} \\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = f _ {o} (\\text {s o f t m a x} (S) \\otimes V) + F _ {c}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "where \\( f_{o} \\) is the output project networks in UA of the specified Attn-AST model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Although Attn-AST methods utilizing UA can achieve high-quality stylization results, they still underperform on content and style images with identical semantics due to inherent limitations, as detailed analysis in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Semantic Continuous Attention. To resolve the problem of the style discontinuity in adjacent regions within the identical semantic regions of the stylized image, as well as style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we propose semantic continuous attention, termed SCA, as shown in Fig. 2 (a) and detailed in Fig. 3. This method disregards the image structures while considering the relationship between local regions and semantic regions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.319, + 0.906, + 0.395 + ], + "angle": 0, + "content": "In particular, SCA first takes the encoded content semantic map features \\( F_{csem} \\) and style semantic map features \\( F_{ssem} \\) as inputs for the query \\( (Q_1) \\) and key \\( (K_1) \\), respectively, while the encoded style image features \\( F_s \\) are used as inputs for the value \\( (V_1) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.402, + 0.905, + 0.421 + ], + "angle": 0, + "content": "\\[\nQ _ {1} = f _ {q} (\\bar {F} _ {c s e m}), K _ {1} = f _ {k} (\\bar {F} _ {s s e m}), V _ {1} = f _ {v} (F _ {s}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.428, + 0.906, + 0.473 + ], + "angle": 0, + "content": "where \\(\\bar{F}_{csem}\\) and \\(\\bar{F}_{ssem}\\) represent the normalized form of \\(F_{csem}\\) and \\(F_{ssem}\\), respectively. \\(f_{q}, f_{k}\\), and \\(f_{v}\\) are the same as those in Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.474, + 0.902, + 0.489 + ], + "angle": 0, + "content": "Then, the preliminary attention map \\(\\mathcal{A}\\) can be obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.496, + 0.905, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = Q _ {1} ^ {\\mathrm {T}} \\otimes K _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.674 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes matrix multiplication. Since \\(F_{csem}\\) and \\(F_{ssem}\\) do not encompass the content structure of the content and style images, after attention computation, the correlations between content points within the same semantic region and all style points are identical. That is if the content query points \\(i\\) and \\(j\\) belong to the same semantic category, then \\(\\mathcal{A}_{im} = \\mathcal{A}_{jm}\\), where \\(m\\) represents any style key point. Consequently, the issue of stylization inconsistency in adjacent areas of the same semantic region, stemming from structural differences, is effectively addressed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.75 + ], + "angle": 0, + "content": "To make SCA focus on regions with the same semantics to transfer a coherent overall style, we modulate the attention map \\(\\mathcal{A}\\) with operation \\(G_{1}\\), setting attention weights of two points from different semantic categories to negative infinity and those within the same category unchanged:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.758, + 0.753, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathcal {A}} = G _ {1} (\\mathcal {A}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.777, + 0.905, + 0.819 + ], + "angle": 0, + "content": "\\[\nG _ {1} \\left(\\mathcal {A} _ {q k}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {A} _ {q k}, \\text {i f} q \\in c \\text {a n d} k \\in c \\\\ - \\infty , \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\( q \\) and \\( k \\) represent the query and key point of \\( Q_{1} \\) and \\( K_{1} \\), respectively. \\( c \\) is the semantic category. \\( \\bar{A} \\) is the new attention map that captures the relationship exclusively between content query points and style key points of the same semantic category. This is visually shown in Fig. 3 (b)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13054" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.457, + 0.107 + ], + "angle": 0, + "content": "Ultimately, we can obtain the stylized features \\( F_{sca} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.119, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\nF _ {s c a} = f _ {o} \\left(\\text {s o f t m a x} (\\bar {\\mathcal {A}}) \\otimes V _ {1}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.314 + ], + "angle": 0, + "content": "where \\( f_{o} \\) is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of different content points and all style points in the same semantic region are equal. Thus, each query point can match all continuous key points in the same semantic region. SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of the same semantic regions in the style image while maintaining style consistency in adjacent areas of the identical semantic region in the stylized image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Semantic Adaptive Instance Normalization. To provide more pure content structure features with less interference from the original color style and a more accurate query for computing the relationships of local structures in corresponding semantic regions of the content and style images, we apply AdaIN [16] to the encoded content image features \\( F_{c} \\) and style image features \\( F_{s} \\) for each semantic region:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.434, + 0.44, + 0.455 + ], + "angle": 0, + "content": "\\[\nI _ {c s e m} \\xrightarrow {K M} M _ {c} ^ {i}, I _ {s s e m} \\xrightarrow {K M} M _ {s} ^ {i}, i = 1, \\dots , N,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.458, + 0.482, + 0.48 + ], + "angle": 0, + "content": "\\[\nF _ {c} ^ {i} = F _ {c} ^ {i} \\odot M _ {c} ^ {i}, \\quad F _ {s} ^ {i} = F _ {s} ^ {i} \\odot M _ {s} ^ {i}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.48, + 0.405, + 0.514 + ], + "angle": 0, + "content": "\\[\nF _ {c} ^ {i} = \\frac {F _ {c} ^ {i} - \\mu (F _ {c} ^ {i})}{\\sigma (F _ {c} ^ {i})} \\times \\sigma (F _ {s} ^ {i}) + \\mu (F _ {s} ^ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.601 + ], + "angle": 0, + "content": "where \\(KM\\) refers to K-Means clustering [28]. \\(M_c^i\\) and \\(M_s^i\\) are the masks containing only 0 and 1. \\(F_c^i\\) and \\(F_s^i\\) represent the features belong to corresponding semantic category \\(i\\). Then, we obtain the features \\(F_c = \\sum_i F_c^i\\) whose overall styles of relevant semantic regions align with those of \\(F_s\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Semantic Sparse Attention. To preserve the original vivid stylistic texture of the style image and address the issue of style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we introduce semantic sparse attention, known as SSA, as shown in Fig. 2 (b) and Fig. 3. This method focuses solely on preserving the style texture with the highest weights, rather than weighting all the style textures, while considering the relationship between local and semantic regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.784 + ], + "angle": 0, + "content": "SSA first employs the transformed content features \\( F_{c} \\) as inputs for the query \\( (Q_{2}) \\) and the encoded style image features \\( F_{s} \\) as inputs for both the key \\( (K_{2}) \\) and value \\( (V_{2}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.796, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\nQ _ {2} = f _ {q} \\left(\\bar {F} _ {c}\\right), K _ {2} = f _ {k} \\left(\\bar {F} _ {s}\\right), V _ {2} = f _ {v} \\left(F _ {s}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.857 + ], + "angle": 0, + "content": "where \\(\\bar{F}_c\\) and \\(\\bar{F}_s\\) represent the normalized \\(F_{c}\\) and \\(F_{s}\\), respectively. \\(f_{q}, f_{k}\\), and \\(f_{v}\\) are the same as those in Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.857, + 0.483, + 0.872 + ], + "angle": 0, + "content": "Then, the initial attention map \\(\\mathcal{B}\\) of SSA can be obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.884, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} = Q _ {2} ^ {\\mathrm {T}} \\otimes K _ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes matrix multiplication. Since \\(F_{c}\\) and \\(F_{s}\\) capture the structures of the content and style images, \\(\\mathcal{B}\\) can consider the feature correlations based on local structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.273 + ], + "angle": 0, + "content": "We claim that the most accurate representation of the original stylistic texture resides in the encoded discrete feature point and the weighted style feature points struggle to preserve the original vivid stylistic textures. Thus, to get accurate and vivid textures, we adjust attention map \\(\\mathcal{B}\\) with operation \\(G_{2}\\) to target the discrete feature point within related regions, setting the only maximum attention weight of a content query point and all style key points in the same semantic category unchanged and others to negative infinity:" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.279, + 0.75, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\boldsymbol {B}}} = G _ {2} (\\mathcal {B}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.3, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\nG _ {2} \\left(\\mathcal {B} _ {q k}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {B} _ {q k}, \\text {i f} \\mathcal {B} _ {q k} = \\max \\left(\\mathcal {B} _ {q n}\\right), \\\\ n \\in \\{1, \\dots , \\operatorname {l e n} \\left(K _ {2}\\right) \\}, q, k, n \\in c \\\\ - \\infty , \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.439 + ], + "angle": 0, + "content": "where \\(q\\) represents the query point of \\(Q_{2}\\). \\(k\\) and \\(n\\) represent the key points of \\(K_{2}\\). \\(c\\) denotes the semantic category. \\(\\bar{\\mathcal{B}}\\) is the new attention map that can grasp the relationship between the content and style feature points, which share the most similar structure. This is visually shown in Fig. 3 (c)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.44, + 0.88, + 0.455 + ], + "angle": 0, + "content": "Eventually, we can obtain the stylized features \\( F_{ssa} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.462, + 0.905, + 0.478 + ], + "angle": 0, + "content": "\\[\nF _ {s s a} = f _ {o} \\left(\\text {s o f t m a x} (\\bar {\\mathcal {B}}) \\otimes V _ {2}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.651 + ], + "angle": 0, + "content": "where \\( f_{o} \\) is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of each content point and the most similar style point in the same semantic region are 1 and weights of this content point and other style points in the same semantic region are 0. Thus, each query point can match sparse key points in the same semantic region. SSA can intently concentrate on the specific stylistic texture of regions with the same semantics to transfer them to the matching semantic region in the stylized image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.652, + 0.906, + 0.681 + ], + "angle": 0, + "content": "Feature Fusion. With the \\(F_{sca}\\) and \\(F_{ssa}\\) of SCA and SSA, the feature fusion can be executed:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.69, + 0.905, + 0.706 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = \\alpha_ {1} \\times F _ {s c a} + \\alpha_ {2} \\times F _ {s s a} + F _ {c}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.906, + 0.759 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) denote the stylization degree of the overall style and vivid textures separately. \\(F_{c}\\) are the outputs of Eq. 8 for superior styling. \\(F_{cs}\\) are the final stylized features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.704, + 0.78 + ], + "angle": 0, + "content": "3.2. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.906, + 0.847 + ], + "angle": 0, + "content": "Revisit Attn-AST Approaches. Given a content image \\( I_{c} \\) and a style image \\( I_{s} \\), the Attn-AST model \\( M \\) could produce a stylized image \\( I_{cs} \\), which preserves the content of \\( I_{c} \\) and exhibit the style of \\( I_{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.848, + 0.905, + 0.877 + ], + "angle": 0, + "content": "Firstly, they achieve the encoded content features \\( F_{c} \\) and style features \\( F_{s} \\) through the encoder \\( E \\) of \\( M \\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\nF _ {c} = E \\left(I _ {c}\\right), F _ {s} = E \\left(I _ {s}\\right). \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13055" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.331, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.089, + 0.605, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.087, + 0.895, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 4. The overall frameworks of Attn-AST approaches-CNN-based, Transformer-based, and Diffusion-based methods with semantic continuous-sparse attention (SCSA). SCA depicts semantic continuous attention. SSA denotes semantic sparse attention. S-AdaIN is semantic adaptive instance normalization. UA expresses universal attention in the Attn-AST methods. The dashed line in (b) shows encoded content features \\( F_{c} \\) are only used in the first use of the feature transformation module, and the output features \\( F_{cs} \\) are used as new content features in subsequent transformation. The dashed line in (c) indicates S-AdaIN is used only when \\( t \\) is the maximum time step \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.373 + ], + "angle": 0, + "content": "Then, they use the feature transformation module \\( T \\) with universal attention (UA) to obtain the stylized features \\( F_{cs} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.381, + 0.483, + 0.399 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = T _ {U A} \\left(F _ {c}, F _ {s}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Eventually, they could acquire the stylized image with the decoder \\(D\\) of \\(M\\):" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.446, + 0.483, + 0.464 + ], + "angle": 0, + "content": "\\[\nI _ {c s} = D \\left(F _ {c s}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.517 + ], + "angle": 0, + "content": "where \\( I_{cs} \\) possesses the content of \\( I_{c} \\) and the style of \\( I_{s} \\), but the style of its corresponding semantic region is inconsistent with that of the style image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.652 + ], + "angle": 0, + "content": "Discussion. Despite the three different Attn-AST frameworks (CNN, Transformer, and Diffusion), their pipelines remain largely consistent with the above, differing only in model structures. The CNN-based method uses a simple encoder-decoder setup, while the Transformer-based approach employs two distinct encoders to process content and style separately. In contrast, the diffusion-based method also utilizes a single encoder-decoder structure but incorporates a time-step iteration mechanism for further refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Attn-AST Approaches with Our SCSA. SCSA is a plug-and-play method that can be incorporated into any Attn-AST method. Thus, given a content image \\( I_{c} \\) with its semantic map \\( I_{csem} \\), a style image \\( I_{s} \\) with its semantic map \\( I_{ssem} \\), and a specified Attn-AST model \\( M \\), our objective is that \\( M_{SCSA} \\), the approach obtained after incorporating SCSA into \\( M \\), can generate a stylized image \\( I_{cs} \\) using the quadruple \\( \\{I_{c}, I_{csem}, I_{s}, I_{ssem}\\} \\), exhibiting a consistent style with the style image within the same semantic regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.819 + ], + "angle": 0, + "content": "At first, we obtain the four encoded features of the quadruple through the encoder \\(E\\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.827, + 0.482, + 0.851 + ], + "angle": 0, + "content": "\\[\nF _ {c} = E \\left(I _ {c}\\right), F _ {s} = E \\left(I _ {s}\\right), \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.848, + 0.416, + 0.863 + ], + "angle": 0, + "content": "\\[\nF _ {c s e m} = E (I _ {c s e m}), F _ {s e m} = E (I _ {s e m}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then, we replace UA in the feature transformation module \\( T \\) of Eq. 15 with SCSA, which could perform semantic" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.373 + ], + "angle": 0, + "content": "feature conversion by enabling each query point to consider certain key points in the corresponding semantic region:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.38, + 0.905, + 0.397 + ], + "angle": 0, + "content": "\\[\nF _ {c s} = T _ {S C S A} \\left(F _ {c}, F _ {s}, F _ {c s e m}, F _ {s s e m}\\right), \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.404, + 0.901, + 0.419 + ], + "angle": 0, + "content": "where \\(F_{cs}\\) are stylized features that meet semantic criteria." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.42, + 0.888, + 0.435 + ], + "angle": 0, + "content": "Lastly, we get the stylized image with the decoder \\(D\\):" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.442, + 0.905, + 0.459 + ], + "angle": 0, + "content": "\\[\nI _ {c s} = D \\left(F _ {c s}\\right), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.466, + 0.906, + 0.51 + ], + "angle": 0, + "content": "where the style of the semantic region of the stylized image \\( I_{cs} \\) is consistent with that of the style image \\( I_{s} \\), while the content of \\( I_{cs} \\) is the same as that of the content image \\( I_{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.906, + 0.587 + ], + "angle": 0, + "content": "The frameworks of the CNN-based, Transformer-based, and Diffusion-based Attn-AST methods with our SCSA are shown in Fig. 4. They differ only in the layers where UA is replaced with SCSA and in the encoder and decoder structures, while the other operations remain largely consistent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.646, + 0.615 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.707, + 0.64 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.766 + ], + "angle": 0, + "content": "We select three proxy Attn-AST approaches (CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]), and five state-of-the-art (SOTA) semantic stylization methods (STROTSS [21], MAST [17], TR [40], DIA [25], and GLStyleNet [38]) as baselines. We obtain images from prior research, datasets, and the Internet, and construct their semantic maps to generate quadruple data. More details are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.736, + 0.79 + ], + "angle": 0, + "content": "4.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.825 + ], + "angle": 0, + "content": "We conduct extensive comparison experiments to evaluate the effectiveness and generalization of our SCSA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 5, Attn-AST approaches combined with our SCSA can generate high-quality and vivid stylized images that better meet semantic needs compared to traditional Attn-AST methods, showcasing rich textures and vibrant colors in corresponding regions, e.g., the background" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13056" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.088, + 0.895, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.294, + 0.807, + 0.309 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons among Attn-AST approaches, those with SCSA, and SOTA methods." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.321, + 0.88, + 0.38 + ], + "angle": 0, + "content": "
SANetSANet + SCSAStyTr2StyTr2 + SCSAStyleIDStyleID + SCSASTROTSSMASTTRDIAGLStyleNet
SSL ↓1.65830.87621.98261.22281.75381.24471.09811.73201.26311.93981.0305
FID ↓14.338513.078812.527312.396312.594412.449712.840016.516313.584620.794214.4880
CFSD ↓0.11030.08740.07520.07050.09160.11780.10080.07460.11490.11470.3991
Pref. ↑0.16850.83150.15760.84240.21920.78080.18670.11000.04500.01670.1183
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.39, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of Attn-AST approaches, those with SCSA, and several SOTA semantic style transfer methods. The best results for Attn-AST methods and Attn-AST methods with SCSA are in bold, while the best results among all methods are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.509 + ], + "angle": 0, + "content": "in the 1st row, along with the sky and ground in the 2nd row. Likewise, compared to SOTA semantic style transfer methods, the stylized images generated by SCSA demonstrate superior semantic accuracy and content preservation, e.g. the river in the 3rd row and the horses in the 4th row." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.571 + ], + "angle": 0, + "content": "In summary, SCSA is a plug-and-play solution that can be integrated into Attn-AST methods, addressing their semantic issues and offering a level of capability in semantic style transfer that surpasses that of existing SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.325, + 0.594 + ], + "angle": 0, + "content": "4.3. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.6, + 0.466, + 0.614 + ], + "angle": 0, + "content": "We also conduct quantitative analyses for full validation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Semantic Style Loss (SSL). Traditional style loss [16] is unsuitable for semantic style transfer. There are two main reasons: First, it focuses on the overall region, making it difficult to capture the semantic region style; second, the inconsistency in the sizes of the semantic regions between the style and stylized images affects the calculation of global style loss. Hence, we propose a semantic style loss to calculate each region style loss similar to Eq. 8 using the style loss of [16]. From the 1st row of Tab. 1, it is clear that the Attn-AST methods utilizing SCSA not only achieve the lowest SSL compared to traditional Attn-AST methods, but when SANet incorporates SCSA, it also reaches the best performance among all semantic style transfer methods. This indicates that SCSA can effectively facilitate semantic style transfer with its effectiveness and generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Frechet Inception Distance (FID). Similar to [8], we use FID [13] to evaluate the overall style fidelity between the stylized and style images. As indicated by the 2nd row of Tab. 1, the Attn-AST methods integrated with SCSA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.526 + ], + "angle": 0, + "content": "achieve the lowest FID value as opposed to typical Attn-AST methods. Particularly, \\(\\mathrm{StyTr}^2\\) with SCSA secures the optimal FID value among all the semantic style transfer methods. The above indicates that SCSA not only conducts the semantic style transfer but also markedly improves the overall style quality of the stylized images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.909, + 0.709 + ], + "angle": 0, + "content": "Content Feature Structural Distance (CFSD). As discussed in [8], content loss [16] and LPIPS [47] are affected by the stylization. Thus, we use the content feature structural distance [8] to assess content retention. As shown in the \\(3rd\\) row of Tab. 1, the SCSA achieves the lowest CFSD value on CNN-based and Transformer-based methods. Although the CFSD value is slightly higher on the Diffusion-based approach with SCSA, both quantitative and qualitative results indicate that SCSA greatly improves semantic stylization, making this trade-off worthwhile. Besides, StyTr² with SCSA has the lowest CFSD value in all SOTA methods, showcasing its content preservation advantage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.909, + 0.832 + ], + "angle": 0, + "content": "User Study. Our user study is divided into two parts: the first compares traditional Attn-AST methods with SCSA-enhanced ones, and the second evaluates Attn-AST methods with SCSA against existing semantic style transfer methods. As shown in the 4th row of Tab. 1, the stylized images of the Attn-AST methods with SCSA are more favored by the public, and their popularity surpasses that of the SOTA methods. These demonstrate the superiority of SCSA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.667, + 0.864 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct ablation experiments to verify the crucial roles of SCA and SSA and to affirm the potency of S-AdaIN." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13057" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.898, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.228, + 0.837, + 0.243 + ], + "angle": 0, + "content": "Figure 6. Qualitative and quantitative ablation results on SCA, SSA, and S-AdaIN. Zoom in for better comparison." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.258, + 0.475, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.366, + 0.465, + 0.38 + ], + "angle": 0, + "content": "Figure 7. The stylization of different content representation." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.387, + 0.47, + 0.434 + ], + "angle": 0, + "content": "
UASCASSASCSA
Runtime (milliseconds)1.2311.9172.0153.944
Memory Usage (bytes)4708479649726220
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.442, + 0.451, + 0.456 + ], + "angle": 0, + "content": "Table 2. Computation costs for feature fusion modules." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.638 + ], + "angle": 0, + "content": "Effect of SCA and SSA. We use SCA or SSA individually as the feature transformation module to assess their capabilities. As shown in Fig. 6, SCA alone facilitates overall color transfer in corresponding semantic regions but results in texture loss and reduced stylization, as reflected by the higher SSL and FID values. In contrast, SSA can effectively transfer specific textures but introduces color discrepancies between the semantic regions of the stylized and style images. This indicates that while SCA captures the overall style characteristics, SSA is crucial for transferring specific textures, highlighting the complementary roles of both." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Effect of S-AdaIN. To validate the effectiveness of S-AdaIN, we conducted ablation experiments by removing S-AdaIN. As shown in Fig. 6, without S-AdaIN, the stylized images exhibit less accurate global style transfer in semantic regions and reduced texture details, particularly evident for the StyleID method. Although S-AdaIN was not employed, CFSD values of StyTr² and StyleID with SCSA are lower, which comes at the expense of stylization quality. Therefore, after considering the pros and cons, S-AdaIN is essential and effective for semantic style transfer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.21, + 0.817 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.825, + 0.472, + 0.839 + ], + "angle": 0, + "content": "We perform more discussions about our task and method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Arbitrary Semantic Style Transfer. Our plug-and-play SCSA is built on the Attn-AST approaches, unlike methods that iteratively optimize the specific content and style images or only apply a predefined style to any content im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.905, + 0.352 + ], + "angle": 0, + "content": "age, so that it can stylize between any content and style images. Besides, it is important to note that SCSA leverages the user-provided semantic maps to guide the transfer process. Therefore, as long as the user's semantic maps mark the regions where the style should be applied, we can perform the relevant region stylization, as shown in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.535 + ], + "angle": 0, + "content": "Computation Costs. Since the costs vary with the feature size, we use the features of the \\(512 \\times 512\\) images encoded by ReLU4_1 layer of VGG-19 to discuss costs. The results are in Tab. 2. In the 1st row, SCA and SSA have slightly longer runtimes than UA due to the guidance operations. SCSA takes barely longer than the sum of SCA and SSA, as it merges both outputs. The 2nd row shows SCA and SSA use a bit more memory than UA, while SCSA uses less than SCA plus SSA. This is because some operations reduce memory usage by reusing existing variables. Despite the cost increases, this is acceptable and does not impact practical use. Cost refinements are left for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.549, + 0.634, + 0.564 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.756 + ], + "angle": 0, + "content": "In this paper, we highlight the challenge faced by existing arbitrary style transfer methods, with a particular emphasis on attention-based approaches, and provide an in-depth analysis of this issue, which is the failure to account for the relationship between local regions and semantic regions. To address this issue and generate high-quality stylized images that meet semantic needs, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA. It achieves precise semantic style transfer by allowing each query point to consider certain key points in the corresponding semantic region. Extensive experimental results demonstrate that our SCSA is effective and generalized." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.673, + 0.787 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was partly supported by the National Natural Science Foundation of China No. 62306277, the Fundamental Research Funds for the Zhejiang Provincial Universities Grant No. K20240090, the Research Fund for International Scientists of National Natural Science Foundation of China No. 72350710798, and the Natural Science Foundation of Zhejiang Province No. LDT23F02023F02. This" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13058" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "research was supported in part through funding from the following sources: the National Natural Science Foundation of China under Award No. 62306277, and the Fundamental Research Funds for Provincial Universities in Zhejiang Province through Grant No. K20240090." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.181, + 0.188, + 0.197 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.206, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.278, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[2] Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[3] Alex J Champandard. Semantic style transfer and turning two-bit doodles into fine artworks. arXiv preprint arXiv:1603.01768, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[4] Dar-Yen Chen. Artfusion: Arbitrary style transfer using dual conditional latent diffusion models. arXiv preprint arXiv:2306.09330, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.407, + 0.484, + 0.476 + ], + "angle": 0, + "content": "[5] Dar-Yen Chen, Hamish Tennent, and Ching-Wen Hsu. Art-adapter: Text-to-image style transfer using multi-level style encoder and explicit adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8619-8628, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.478, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[6] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[7] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.485, + 0.66 + ], + "angle": 0, + "content": "[8] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3730-3738. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Mark Hamazaspyan and Shant Navasardyan. Diffusion-enhanced patchmatch: A framework for arbitrary style transfer with diffusion models. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.206, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 797-805, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[15] Siyu Huang, Jie An, Donglai Wei, Jiebo Luo, and Hanspeter Pfister. Quantart: Quantizing image style transfer towards high visual fidelity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5947-5956, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[17] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14861-14869, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.908, + 0.545 + ], + "angle": 0, + "content": "[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence, pages 4369-4376, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[19] Nikolai Kalischek, Jan D Wegner, and Konrad Schindler. In the light of feature distributions: moment matching for neural style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9391, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Da-Hyun Kim, Dong-Hee Shin, and Tae-Eui Kam. Bridging the bci illiteracy gap: a subject-to-subject semantic style transfer for eeg-based motor imagery classification. Frontiers in Human Neuroscience, 17:1194751, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[21] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[22] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In International conference on machine learning, pages 957-966. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13059" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. arXiv preprint arXiv:1705.01088, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Yi-Sheng Liao and Chun-Rong Huang. Semantic context-aware image style transfer. IEEE Transactions on Image Processing, 31:1911-1923, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[28] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5951-5960. IEEE, .3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[30] Ming Lu, Hao Zhao, Anbang Yao, Feng Xu, Yurong Chen, and Li Zhang. Decoder network over lightweight reconstructed feature for fast semantic style transfer. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2488-2496. IEEE, .3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[31] Yingnan Ma, Chenqiu Zhao, Xudong Li, and Anup Basu. Rast: Restorable arbitrary style transfer via multi-restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 331-340, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[35] Onkar Susladkar, Gayatri Deshmukh, Sparsh Mittal, and Parth Shastri. D2styler: Advancing arbitrary style transfer with discrete diffusion methods. arXiv preprint arXiv:2408.03558, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[36] Gemma Canet Tarres, Dan Ruta, Tu Bui, and John Collosomasse. Parasol: Parametric style control for diffusion image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2432-2442, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[38] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Gliystylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: towards" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1095-1106, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2624-2632, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[42] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[43] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[44] Zhengtao Xiang, Xing Wan, Libo Xu, Xin Yu, and Yuhan Mao. A training-free latent diffusion style transfer method. Information, 15(10):588, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[45] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.907, + 0.557 + ], + "angle": 0, + "content": "[46] Chiyu Zhang, Xiaogang Xu, Lei Wang, Zaiyan Dai, and Jun Yang. S2wat: Image style transfer via hierarchical vision transformer using strips window attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7024-7032, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[48] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[49] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. A unified arbitrary style transfer framework via adaptive contrastive learning. ACM Trans. Graph., 42(5), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[50] Hui-Huang Zhao, Paul L Rosin, Yu-Kun Lai, and Yao-Nan Wang. Automatic semantic style transfer using deep convolutional neural networks and soft masks. The Visual Computer, 36(7):1307-1324, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13060" + } + ] +] \ No newline at end of file diff --git a/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_origin.pdf b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8eaf4a3366e446544e60c218b7062dbd32331e25 --- /dev/null +++ b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/fae97f07-7418-476a-be91-97b643e54a78_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6b859eb1ad3f5d1a4bfe52c7c1d90365b1f3840f3b22c458905126e37dfbcdd +size 6219308 diff --git a/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/full.md b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3719170456d1214920dc64256ec05ad5e191c421 --- /dev/null +++ b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/full.md @@ -0,0 +1,452 @@ +# SCSA: A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer + +Chunnan Shang + +Zhejiang University + +chunnan.22@intl.zju.edu.cn + +Hongwei Wang + +Zhejiang University + +hongweiwang@intl.zju.edu.cn + +Zhizhong Wang† + +Zhejiang University + +endywon@zju.edu.cn + +Xiangming Meng† + +Zhejiang University + +xiangmingmeng@intl.zju.edu.cn + +# Abstract + +Attention-based arbitrary style transfer methods, including CNN-based, Transformer-based, and Diffusion-based, have flourished and produced high-quality stylized images. However, they perform poorly on the content and style images with the same semantics, i.e., the style of the corresponding semantic region of the generated stylized image is inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. To address this issue, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer—each query point considers certain key points in the corresponding semantic region. Specifically, semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region that reflect the overall style characteristics of that region; Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region that exhibits the specific stylistic texture of that region. By combining the two modules, the resulting SCSA aligns the overall style of the corresponding semantic regions while transferring the vivid textures of these regions. Qualitative and quantitative results demonstrate that SCSA enables attention-based arbitrary style transfer methods to produce high-quality semantic stylized images. The code can be found in https://github.com/ecn-00/SCSA. + +# 1. Introduction + +Initially, style transfer [11] was proposed to preserve the content of a specified content image while transferring the style from a specified style image by iteratively optimizing + +Content Inputs + +![](images/29539f9393f59eaedf8e4f3df4fe7a2d31336c799a7ea941bd5a2d88c28e2bdf.jpg) + +Style Inputs + +![](images/1edea075c59643bdb6e64fc76bd85e63f217067de0a7c98b0dc932efea7ee423.jpg) + +SANet + +![](images/d3d5ff1194c5f66d5f494a69365fc6802c0f2c1a2c4904050a792ba40b25d14e.jpg) + +![](images/69520fff0fd62612fbfa196b714b03310c301051644920d01f06395aa510dd0e.jpg) +SANet + SCSA + +![](images/15e701c7f8fbe11b2a43211c375f3c3348b8c91b1ca4f31f4bffa759fc87c0db.jpg) +Content Inputs + +![](images/b07a3ebe2a8e54b28e9f8d93d9acfe4ef59b7447b6dc2499044183fefc17b97d.jpg) +Style Inputs + +![](images/596510bbb4c3721d65ef49ee5f6bd73a985709a969f13ca19a8534e36b7e10c8.jpg) +StyTr2 + +![](images/2db4d58977d0db3840ba12e85802fbb942e8b5673f201919e2d24d92f7bca1cc.jpg) +$\mathrm{StyTr}^2 + \mathrm{SCSA}$ + +![](images/09a63a75528b1fbb8509ed47479af42211873702bebafe06105d70d17228a1e7.jpg) +Content Inputs +Figure 1. Comparisons of the Attn-AST approaches-CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]-without and with our SCSA. The stylized images generated by the Attn-AST approaches exhibit style discontinuity in adjacent regions within the identical semantic regions (e.g., the background in the 1st row), style inconsistency between corresponding semantic (e.g., the cloth color in the 3rd row), and fewer textures (e.g., the block textures in the 3rd row). + +![](images/ff63f632d36177c00a54774ebecab33fc7a25d93f3a241aab2362609c72e2a84.jpg) +Style Inputs + +![](images/9531137a9b4cdacae2a503bbec2bdab64926ebc500a11531a6f7d7afe7a24a17.jpg) +StyleID + +![](images/6fede9470022d5f96840822978dd18f1779f9560b937c83c7137b440777a05c8.jpg) +StyleID + SCSA + +the generated stylized image. However, its excessive slowness prevents practical application. Given the need for efficiency, effectiveness, and generalization, Arbitrary Style Transfer (AST) [1, 4, 8, 23, 27] has become a highly prominent research focus in the style transfer field. AST seeks to generate a new image that replicates the content of an arbitrary content image while adopting the style of an arbitrary style image. In spite of many recent studies aiming to advance the AST effect by incorporating novel knowledge [15, 36, 39] and innovative constraints [41, 43, 49], the core of AST continues to reside in feature transformation modules, which remains indispensable for all approaches. Hence, the existing AST approaches can be classified into + +two categories, statistics-based methods [5, 16, 24, 48] and attention-based methods [6, 8, 9, 27, 31, 32]. + +Attention-based Arbitrary Style Transfer (Attn-AST) approaches, leveraging the ability to capture global information and focus on the related features of the attention mechanism [2], preserve the content of the content image while transferring the style that is relevant to the content of the content image from the style image. They encompass CNN-based [27, 31, 32, 39], Transformer-based [9, 14, 42, 46], and Diffusion-based methods [8, 35, 44]. Commonly, Attn-AST methods first establish a holistic relationship, an attention map, between each content query point of the content image features and all style key points of the style image features. Then, Attn-AST methods get the transferred style for each content point by weighting all the style points, where the weights are the softmax function outputs of a naturally calculated attention map of content and style points. + +Although Attn-AST methods have thrived and produced high-quality stylized images due to the capacity mentioned above for global capture and focus, they perform poorly on the content and style image with the same semantics. As shown in Fig. 1, the style of the corresponding semantic region of the stylized image generated by Attn-AST approaches exhibits a style inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. (1) During the formulation of the attention map, Attn-AST methods bluntly and simplistically consider the relationship between each query point and all key points, as shown in Fig. 2 (c), without adequately accounting for the semantic regions to which these points belong. This oversimplification may cause Attn-AST methods to disproportionately emphasize structurally similar points across different semantic regions, resulting in artifacts or inaccurate style [27, 32], e.g., the $2nd$ row of Fig. 1. (2) Attn-AST approaches fundamentally depend on structural principles for generating a global style, as they utilize content and style image features. The subtle disparity in content structure details across adjacent regions can cause different stylization effects. Thus, the stylized images often lack stylistic continuity across adjacent regions within the semantic regions, e.g., the background in the $1st$ row of Fig. 1. (3) Attn-AST approaches acquire the transferred style by weighting all the style points. However, the most accurate representation of the original stylistic texture resides in the encoded single points. The weighted style struggles to preserve the original stylistic texture, leading to a lack of vividness, e.g., the block textures in the $3rd$ row of Fig. 1. + +To maintain the benefits of Attn-AST methods while addressing their existing issues, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer. The key insight of SCSA is that, based on the interpretability of the attention mecha + +nism, each query point considers certain key points in the corresponding semantic region. Specifically, SCSA consists of two parts: Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA). (1) SCA: Semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region by using content and style semantic map features as query and key, respectively. Thus, SCA can achieve the overall style characteristics of the corresponding semantic regions of the style image to generate the stylized image with stylistic continuity across adjacent regions within the semantic regions. (2) SSA: Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region by using content and style image features as query and key, respectively. Hence, SSA can obtain the specific stylistic texture of the corresponding semantic regions of the style image to produce the stylized image with vivid stylistic textures. The intuitive comparison between our SCSA and universal attention can be found in Fig. 2. + +By seamlessly integrating SCSA into any Attn-AST approach without training, we enable arbitrary semantic style transfer while preserving and even enhancing the original stylization effect. To our knowledge, we are the first work to extend attention-based arbitrary style transfer to arbitrary semantic style transfer in a plug-and-play way. + +In brief, our contributions are divided into three parts: + +- We reveal that Attn-AST methods struggle with content and style images with the same semantics and identify the root causes for their subpar performance. +- We propose a semantic continuous-sparse attention, dubbed SCSA, that can extend Attn-AST to arbitrary semantic style transfer in a plug-and-play manner. +- We conduct extensive experiments and comparisons to demonstrate the effectiveness and generalization of SCSA, enabling Attn-AST methods (CNN-based, Transformer-based, and Diffusion-based approaches) to perform arbitrary semantic style transfer while preserving and even enhancing the original stylization. + +# 2. Related Work + +The existing AST approaches can be roughly divided into two groups: statistics-based and attention-based methods. + +Statistics-based Methods. Statistics-based AST methods [1, 18, 23, 43, 49] regard style as the statistical traits inherent in feature distribution. Specifically, they achieve stylized features by aligning the feature statistics of the content image with those of the style image, making them a global style transfer approach. The two most representative approaches are AdaIN [16] and WCT [24], which match the first-order and second-order statistics of features, respectively. To more comprehensively capture the stylistic characteristics of features, higher-order statistical methods, such as CMD [19] and EFDM [48], have been continually ex + +![](images/a8a070f469eff37ca8ddaf7419da0c7d5864bf20d620a954475bf76ed0ff2d75.jpg) +Figure 2. Comparison between Semantic Continuous-Sparse Attention (SCSA) and Universal Attention (UA). SCSA includes two parts: (a) Semantic Continuous Attention (SCA): The query point of the content semantic map features can match all continuous key points of the style semantic map features in the same semantic region. Therefore, SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of regions with the same semantics; (b) Semantic Sparse Attention (SSA): The query point of the content image features can match the most similar sparse key point of the style image features in the same semantic region. Hence, SSA can intently concentrate on the specific stylistic texture of regions with the same semantics. In contrast, (c) Universal Attention (UA) [32]: The query point of the content image features pays attention to all key points of the style image features. As a result, UA fails to accurately convey the intricate overall stylistic characteristics and specific textures of regions that share identical semantics. + +![](images/a91ff56b4b3596a3de4e931eaa36fb179c120abd4b24789adea2f56ea1cce33b.jpg) + +![](images/4c153bda1c4d810bec18fc672a835f943d801afa7446fb4cace1a6be8a3577a5.jpg) + +plored and refined. Expanding on the above foundation, semantic style transfer [10, 20, 26, 29, 30, 50] has been studied by aligning feature statistics within corresponding semantic regions. STROTSS [21] employs the Earth Movers Distance [22] to reduce the feature distribution difference within semantic regions and self-similarity to preserve the semantics and spatial structure. MAST [17] aligns the manifold distributions of semantic regions to facilitate semantic style transfer while preserving the content structure. + +Since feature statistics only provide a global representation, the stylized images they generate often lack fine local textures. To address this, we propose SCSA, which could yield a coherent overall style across adjacent regions by overlooking structure details, while producing vivid textures by emphasizing them via the attention mechanism. + +Attention-based Methods. Attention-based AST (Attn-AST) methods define style as the weighted convergence of all style features. Exactly, Attn-AST methods obtain the stylized feature of each content feature by weighting all the style features, where weights are derived from the correlation between the content feature and all style features obtained through the attention mechanism, making it a local structure style transfer. Attn-AST approaches mainly include CNN-based methods [27, 32], Transformer-based methods [9], and Diffusion-based methods [8]. SANet [32] is the pioneering work to integrate attention into AST, employing content features as query and style features as key and value to obtain stylized features via the attention mechanism. StyTr² [9] integrates the outputs of the content and style transformer [37] encoders into the transformer feature fusion module, obtaining stylized features. StyleID [8] injects the content query, style key, and style value, derived from DDIM inversion [34], into certain self-attention layers of U-Net [33] to merge style, achieving stylized features through multiple inversion steps. Even though they + +continually improve the quality of stylized images, their performance remains poor on content and style images with identical semantics. In contrast, patch-based methods [3, 7, 12, 38, 45] achieve semantic style transfer by swapping the styles of patches with the most similar content structures. When the patch size in patch-based methods is 1, they can be roughly considered attention-based methods. Deep-Image-Analogy (DIA) [25] introduces nearest-neighbor field search to tackle the challenge of patch matching for images, which are visually distinct yet semantically similar. GLStyleNet [38] uses multi-level pyramid features for semantic optimization via patch matching, iteratively creating stylized images that meet semantic criteria. TR [40] utilizes the patch match to achieve precise semantic-guided and structure-preserving texture transfer. + +As the above methods target local structural similarity, subtle changes in the content structure of adjacent regions can result in stylization discontinuity in the identical semantic region. To mitigate this concern, we introduce SCSA, which provides more stable and continuous stylization by fully comprehending semantics and transfers vivid stylistic textures by focusing on local structure. Most importantly, it can be easily integrated into existing frameworks, while other methods are either designed for specific networks or require multiple iterations for optimization. + +# 3. Method + +# 3.1. Semantic Continuous-Sparse Attention + +Revisit Universal Attention in Attn-AST. Attn-AST approaches feed the encoded content feature and style feature into Universal Attention (UA), constructing intricate connections between various positions within the two feature sequences by directly utilizing the self-attention [2] computational paradigm, to capture the long-range local textures + +![](images/2c3af095416a55cda99882831fec17124c534a0f2d7f8ae2ddea7d2cd03a5147.jpg) +Figure 3. Detailed Procedure of Semantic Continuous-Sparse Attention (SCSA). In (a), S-AdaIN denotes that AdaIN [16] is applied individually for each semantic region. It initializes $F_{c}$ , by matching the feature statistics of corresponding semantic regions in $F_{c}$ and $F_{s}$ , to ensure that $F_{c}$ has semantically aligned color style information to some extent. Subsequently, $F_{c}$ can provide a more accurate query $f_{q}(\bar{F}_{c})$ and more pure content structure features with less interference from the original color style. SCSA in (a) includes Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA) two parts. "G1" in SCA sets the value of the specific product point generated by multiplying two points from different semantic categories to negative infinity. As shown in (b), only the product of points with the same semantics is retained. Therefore, SCA can fully account for the overall stylistic characteristics of regions with the same semantics. "G2" in SSA retains only the maximum value of the product of a specified query point with all key points in the same semantic region, while all other values are set to negative infinity, as shown in (c). Hence, SSA can intently concentrate on the specific stylistic texture of the most similar structure in the regions with the same semantics. After SCSA, the generated $F_{cs}$ from (a) not only reflects the overall stylistic characteristics of the semantic region corresponding to $F_{s}$ but also captures the specific textures of that semantic region. Meanwhile, $F_{cs}$ also includes the content structure of $F_{c}$ . + +![](images/d186ca292dded83a9b23844c0e0538c4c6b7733c94a334f8aba382c3959e2180.jpg) + +relationship. Typically, given the encoded content features $F_{c}$ and style features $F_{s}$ , the query (Q), key (K), and value (V) of UA can be obtained as follows: + +$$ +Q = f _ {q} \left(\bar {F} _ {c}\right), K = f _ {k} \left(\bar {F} _ {s}\right), V = f _ {v} \left(F _ {s}\right), \tag {1} +$$ + +where $\bar{F}_c$ and $\bar{F}_s$ denote the normalized $F_{c}$ and $F_{s}$ , respectively. $f_{q}, f_{k}$ , and $f_{v}$ are the project networks for getting query, key, and value in UA of the given Attn-AST model. + +After that, the attention map $S$ , which represents a holistic relationship between each content query point of the content features and all style key points of the style features, can be obtained as follows: + +$$ +S = Q ^ {\mathsf {T}} \otimes K, \tag {2} +$$ + +where $\otimes$ denotes matrix multiplication. + +Finally, the stylized features $F_{cs}$ can be calculated as: + +$$ +F _ {c s} = f _ {o} (\text {s o f t m a x} (S) \otimes V) + F _ {c}, \tag {3} +$$ + +where $f_{o}$ is the output project networks in UA of the specified Attn-AST model. + +Although Attn-AST methods utilizing UA can achieve high-quality stylization results, they still underperform on content and style images with identical semantics due to inherent limitations, as detailed analysis in Sec. 1. + +Semantic Continuous Attention. To resolve the problem of the style discontinuity in adjacent regions within the identical semantic regions of the stylized image, as well as style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we propose semantic continuous attention, termed SCA, as shown in Fig. 2 (a) and detailed in Fig. 3. This method disregards the image structures while considering the relationship between local regions and semantic regions. + +In particular, SCA first takes the encoded content semantic map features $F_{csem}$ and style semantic map features $F_{ssem}$ as inputs for the query $(Q_1)$ and key $(K_1)$ , respectively, while the encoded style image features $F_s$ are used as inputs for the value $(V_1)$ : + +$$ +Q _ {1} = f _ {q} (\bar {F} _ {c s e m}), K _ {1} = f _ {k} (\bar {F} _ {s s e m}), V _ {1} = f _ {v} (F _ {s}), \tag {4} +$$ + +where $\bar{F}_{csem}$ and $\bar{F}_{ssem}$ represent the normalized form of $F_{csem}$ and $F_{ssem}$ , respectively. $f_{q}, f_{k}$ , and $f_{v}$ are the same as those in Eq. 1. + +Then, the preliminary attention map $\mathcal{A}$ can be obtained: + +$$ +\mathcal {A} = Q _ {1} ^ {\mathrm {T}} \otimes K _ {1}, \tag {5} +$$ + +where $\otimes$ denotes matrix multiplication. Since $F_{csem}$ and $F_{ssem}$ do not encompass the content structure of the content and style images, after attention computation, the correlations between content points within the same semantic region and all style points are identical. That is if the content query points $i$ and $j$ belong to the same semantic category, then $\mathcal{A}_{im} = \mathcal{A}_{jm}$ , where $m$ represents any style key point. Consequently, the issue of stylization inconsistency in adjacent areas of the same semantic region, stemming from structural differences, is effectively addressed. + +To make SCA focus on regions with the same semantics to transfer a coherent overall style, we modulate the attention map $\mathcal{A}$ with operation $G_{1}$ , setting attention weights of two points from different semantic categories to negative infinity and those within the same category unchanged: + +$$ +\bar {\mathcal {A}} = G _ {1} (\mathcal {A}), +$$ + +$$ +G _ {1} \left(\mathcal {A} _ {q k}\right) = \left\{ \begin{array}{l} \mathcal {A} _ {q k}, \text {i f} q \in c \text {a n d} k \in c \\ - \infty , \text {o t h e r w i s e} \end{array} , \right. \tag {6} +$$ + +where $q$ and $k$ represent the query and key point of $Q_{1}$ and $K_{1}$ , respectively. $c$ is the semantic category. $\bar{A}$ is the new attention map that captures the relationship exclusively between content query points and style key points of the same semantic category. This is visually shown in Fig. 3 (b). + +Ultimately, we can obtain the stylized features $F_{sca}$ : + +$$ +F _ {s c a} = f _ {o} \left(\text {s o f t m a x} (\bar {\mathcal {A}}) \otimes V _ {1}\right), \tag {7} +$$ + +where $f_{o}$ is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of different content points and all style points in the same semantic region are equal. Thus, each query point can match all continuous key points in the same semantic region. SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of the same semantic regions in the style image while maintaining style consistency in adjacent areas of the identical semantic region in the stylized image. + +Semantic Adaptive Instance Normalization. To provide more pure content structure features with less interference from the original color style and a more accurate query for computing the relationships of local structures in corresponding semantic regions of the content and style images, we apply AdaIN [16] to the encoded content image features $F_{c}$ and style image features $F_{s}$ for each semantic region: + +$$ +I _ {c s e m} \xrightarrow {K M} M _ {c} ^ {i}, I _ {s s e m} \xrightarrow {K M} M _ {s} ^ {i}, i = 1, \dots , N, +$$ + +$$ +F _ {c} ^ {i} = F _ {c} ^ {i} \odot M _ {c} ^ {i}, \quad F _ {s} ^ {i} = F _ {s} ^ {i} \odot M _ {s} ^ {i}, \tag {8} +$$ + +$$ +F _ {c} ^ {i} = \frac {F _ {c} ^ {i} - \mu (F _ {c} ^ {i})}{\sigma (F _ {c} ^ {i})} \times \sigma (F _ {s} ^ {i}) + \mu (F _ {s} ^ {i}), +$$ + +where $KM$ refers to K-Means clustering [28]. $M_c^i$ and $M_s^i$ are the masks containing only 0 and 1. $F_c^i$ and $F_s^i$ represent the features belong to corresponding semantic category $i$ . Then, we obtain the features $F_c = \sum_i F_c^i$ whose overall styles of relevant semantic regions align with those of $F_s$ . + +Semantic Sparse Attention. To preserve the original vivid stylistic texture of the style image and address the issue of style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we introduce semantic sparse attention, known as SSA, as shown in Fig. 2 (b) and Fig. 3. This method focuses solely on preserving the style texture with the highest weights, rather than weighting all the style textures, while considering the relationship between local and semantic regions. + +SSA first employs the transformed content features $F_{c}$ as inputs for the query $(Q_{2})$ and the encoded style image features $F_{s}$ as inputs for both the key $(K_{2})$ and value $(V_{2})$ : + +$$ +Q _ {2} = f _ {q} \left(\bar {F} _ {c}\right), K _ {2} = f _ {k} \left(\bar {F} _ {s}\right), V _ {2} = f _ {v} \left(F _ {s}\right), \tag {9} +$$ + +where $\bar{F}_c$ and $\bar{F}_s$ represent the normalized $F_{c}$ and $F_{s}$ , respectively. $f_{q}, f_{k}$ , and $f_{v}$ are the same as those in Eq. 1. + +Then, the initial attention map $\mathcal{B}$ of SSA can be obtained: + +$$ +\mathcal {B} = Q _ {2} ^ {\mathrm {T}} \otimes K _ {2}, \tag {10} +$$ + +where $\otimes$ denotes matrix multiplication. Since $F_{c}$ and $F_{s}$ capture the structures of the content and style images, $\mathcal{B}$ can consider the feature correlations based on local structures. + +We claim that the most accurate representation of the original stylistic texture resides in the encoded discrete feature point and the weighted style feature points struggle to preserve the original vivid stylistic textures. Thus, to get accurate and vivid textures, we adjust attention map $\mathcal{B}$ with operation $G_{2}$ to target the discrete feature point within related regions, setting the only maximum attention weight of a content query point and all style key points in the same semantic category unchanged and others to negative infinity: + +$$ +\overline {{\boldsymbol {B}}} = G _ {2} (\mathcal {B}), +$$ + +$$ +G _ {2} \left(\mathcal {B} _ {q k}\right) = \left\{ \begin{array}{l} \mathcal {B} _ {q k}, \text {i f} \mathcal {B} _ {q k} = \max \left(\mathcal {B} _ {q n}\right), \\ n \in \{1, \dots , \operatorname {l e n} \left(K _ {2}\right) \}, q, k, n \in c \\ - \infty , \text {o t h e r w i s e} \end{array} , \right. \tag {11} +$$ + +where $q$ represents the query point of $Q_{2}$ . $k$ and $n$ represent the key points of $K_{2}$ . $c$ denotes the semantic category. $\bar{\mathcal{B}}$ is the new attention map that can grasp the relationship between the content and style feature points, which share the most similar structure. This is visually shown in Fig. 3 (c). + +Eventually, we can obtain the stylized features $F_{ssa}$ : + +$$ +F _ {s s a} = f _ {o} \left(\text {s o f t m a x} (\bar {\mathcal {B}}) \otimes V _ {2}\right), \tag {12} +$$ + +where $f_{o}$ is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of each content point and the most similar style point in the same semantic region are 1 and weights of this content point and other style points in the same semantic region are 0. Thus, each query point can match sparse key points in the same semantic region. SSA can intently concentrate on the specific stylistic texture of regions with the same semantics to transfer them to the matching semantic region in the stylized image. + +Feature Fusion. With the $F_{sca}$ and $F_{ssa}$ of SCA and SSA, the feature fusion can be executed: + +$$ +F _ {c s} = \alpha_ {1} \times F _ {s c a} + \alpha_ {2} \times F _ {s s a} + F _ {c}, \tag {13} +$$ + +where $\alpha_{1}$ and $\alpha_{2}$ denote the stylization degree of the overall style and vivid textures separately. $F_{c}$ are the outputs of Eq. 8 for superior styling. $F_{cs}$ are the final stylized features. + +# 3.2. Overall Framework + +Revisit Attn-AST Approaches. Given a content image $I_{c}$ and a style image $I_{s}$ , the Attn-AST model $M$ could produce a stylized image $I_{cs}$ , which preserves the content of $I_{c}$ and exhibit the style of $I_{s}$ . + +Firstly, they achieve the encoded content features $F_{c}$ and style features $F_{s}$ through the encoder $E$ of $M$ : + +$$ +F _ {c} = E \left(I _ {c}\right), F _ {s} = E \left(I _ {s}\right). \tag {14} +$$ + +![](images/e375e3af126ce2b966761f4fba7e797de8b226be5718ef2bb3a72f1df267f0a2.jpg) +Figure 4. The overall frameworks of Attn-AST approaches-CNN-based, Transformer-based, and Diffusion-based methods with semantic continuous-sparse attention (SCSA). SCA depicts semantic continuous attention. SSA denotes semantic sparse attention. S-AdaIN is semantic adaptive instance normalization. UA expresses universal attention in the Attn-AST methods. The dashed line in (b) shows encoded content features $F_{c}$ are only used in the first use of the feature transformation module, and the output features $F_{cs}$ are used as new content features in subsequent transformation. The dashed line in (c) indicates S-AdaIN is used only when $t$ is the maximum time step $T$ . + +![](images/4d1a312e4e3a062108d6a8aa70b297b6d08fd5a0cb7faf5e88a1e23e60269064.jpg) + +![](images/9d5b85a6a727b3a716728cb0fe396c7792c9c2103ab3d4962fee85755ccc32b1.jpg) + +Then, they use the feature transformation module $T$ with universal attention (UA) to obtain the stylized features $F_{cs}$ : + +$$ +F _ {c s} = T _ {U A} \left(F _ {c}, F _ {s}\right). \tag {15} +$$ + +Eventually, they could acquire the stylized image with the decoder $D$ of $M$ : + +$$ +I _ {c s} = D \left(F _ {c s}\right), \tag {16} +$$ + +where $I_{cs}$ possesses the content of $I_{c}$ and the style of $I_{s}$ , but the style of its corresponding semantic region is inconsistent with that of the style image. + +Discussion. Despite the three different Attn-AST frameworks (CNN, Transformer, and Diffusion), their pipelines remain largely consistent with the above, differing only in model structures. The CNN-based method uses a simple encoder-decoder setup, while the Transformer-based approach employs two distinct encoders to process content and style separately. In contrast, the diffusion-based method also utilizes a single encoder-decoder structure but incorporates a time-step iteration mechanism for further refinement. + +Attn-AST Approaches with Our SCSA. SCSA is a plug-and-play method that can be incorporated into any Attn-AST method. Thus, given a content image $I_{c}$ with its semantic map $I_{csem}$ , a style image $I_{s}$ with its semantic map $I_{ssem}$ , and a specified Attn-AST model $M$ , our objective is that $M_{SCSA}$ , the approach obtained after incorporating SCSA into $M$ , can generate a stylized image $I_{cs}$ using the quadruple $\{I_{c}, I_{csem}, I_{s}, I_{ssem}\}$ , exhibiting a consistent style with the style image within the same semantic regions. + +At first, we obtain the four encoded features of the quadruple through the encoder $E$ : + +$$ +F _ {c} = E \left(I _ {c}\right), F _ {s} = E \left(I _ {s}\right), \tag {17} +$$ + +$$ +F _ {c s e m} = E (I _ {c s e m}), F _ {s e m} = E (I _ {s e m}). +$$ + +Then, we replace UA in the feature transformation module $T$ of Eq. 15 with SCSA, which could perform semantic + +feature conversion by enabling each query point to consider certain key points in the corresponding semantic region: + +$$ +F _ {c s} = T _ {S C S A} \left(F _ {c}, F _ {s}, F _ {c s e m}, F _ {s s e m}\right), \tag {18} +$$ + +where $F_{cs}$ are stylized features that meet semantic criteria. + +Lastly, we get the stylized image with the decoder $D$ : + +$$ +I _ {c s} = D \left(F _ {c s}\right), \tag {19} +$$ + +where the style of the semantic region of the stylized image $I_{cs}$ is consistent with that of the style image $I_{s}$ , while the content of $I_{cs}$ is the same as that of the content image $I_{c}$ . + +The frameworks of the CNN-based, Transformer-based, and Diffusion-based Attn-AST methods with our SCSA are shown in Fig. 4. They differ only in the layers where UA is replaced with SCSA and in the encoder and decoder structures, while the other operations remain largely consistent. + +# 4. Experiments + +# 4.1. Experiment Settings + +We select three proxy Attn-AST approaches (CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]), and five state-of-the-art (SOTA) semantic stylization methods (STROTSS [21], MAST [17], TR [40], DIA [25], and GLStyleNet [38]) as baselines. We obtain images from prior research, datasets, and the Internet, and construct their semantic maps to generate quadruple data. More details are in the supplementary materials. + +# 4.2. Qualitative Comparison + +We conduct extensive comparison experiments to evaluate the effectiveness and generalization of our SCSA. + +As shown in Fig. 5, Attn-AST approaches combined with our SCSA can generate high-quality and vivid stylized images that better meet semantic needs compared to traditional Attn-AST methods, showcasing rich textures and vibrant colors in corresponding regions, e.g., the background + +![](images/7992688d78e741a42f957464972fac140e56acb80298280e7d258b12c9dd4495.jpg) +Figure 5. Qualitative comparisons among Attn-AST approaches, those with SCSA, and SOTA methods. + +
SANetSANet + SCSAStyTr2StyTr2 + SCSAStyleIDStyleID + SCSASTROTSSMASTTRDIAGLStyleNet
SSL ↓1.65830.87621.98261.22281.75381.24471.09811.73201.26311.93981.0305
FID ↓14.338513.078812.527312.396312.594412.449712.840016.516313.584620.794214.4880
CFSD ↓0.11030.08740.07520.07050.09160.11780.10080.07460.11490.11470.3991
Pref. ↑0.16850.83150.15760.84240.21920.78080.18670.11000.04500.01670.1183
+ +Table 1. Quantitative comparisons of Attn-AST approaches, those with SCSA, and several SOTA semantic style transfer methods. The best results for Attn-AST methods and Attn-AST methods with SCSA are in bold, while the best results among all methods are underlined. + +in the 1st row, along with the sky and ground in the 2nd row. Likewise, compared to SOTA semantic style transfer methods, the stylized images generated by SCSA demonstrate superior semantic accuracy and content preservation, e.g. the river in the 3rd row and the horses in the 4th row. + +In summary, SCSA is a plug-and-play solution that can be integrated into Attn-AST methods, addressing their semantic issues and offering a level of capability in semantic style transfer that surpasses that of existing SOTA methods. + +# 4.3. Quantitative Comparison + +We also conduct quantitative analyses for full validation. + +Semantic Style Loss (SSL). Traditional style loss [16] is unsuitable for semantic style transfer. There are two main reasons: First, it focuses on the overall region, making it difficult to capture the semantic region style; second, the inconsistency in the sizes of the semantic regions between the style and stylized images affects the calculation of global style loss. Hence, we propose a semantic style loss to calculate each region style loss similar to Eq. 8 using the style loss of [16]. From the 1st row of Tab. 1, it is clear that the Attn-AST methods utilizing SCSA not only achieve the lowest SSL compared to traditional Attn-AST methods, but when SANet incorporates SCSA, it also reaches the best performance among all semantic style transfer methods. This indicates that SCSA can effectively facilitate semantic style transfer with its effectiveness and generalization. + +Frechet Inception Distance (FID). Similar to [8], we use FID [13] to evaluate the overall style fidelity between the stylized and style images. As indicated by the 2nd row of Tab. 1, the Attn-AST methods integrated with SCSA + +achieve the lowest FID value as opposed to typical Attn-AST methods. Particularly, $\mathrm{StyTr}^2$ with SCSA secures the optimal FID value among all the semantic style transfer methods. The above indicates that SCSA not only conducts the semantic style transfer but also markedly improves the overall style quality of the stylized images. + +Content Feature Structural Distance (CFSD). As discussed in [8], content loss [16] and LPIPS [47] are affected by the stylization. Thus, we use the content feature structural distance [8] to assess content retention. As shown in the $3rd$ row of Tab. 1, the SCSA achieves the lowest CFSD value on CNN-based and Transformer-based methods. Although the CFSD value is slightly higher on the Diffusion-based approach with SCSA, both quantitative and qualitative results indicate that SCSA greatly improves semantic stylization, making this trade-off worthwhile. Besides, StyTr² with SCSA has the lowest CFSD value in all SOTA methods, showcasing its content preservation advantage. + +User Study. Our user study is divided into two parts: the first compares traditional Attn-AST methods with SCSA-enhanced ones, and the second evaluates Attn-AST methods with SCSA against existing semantic style transfer methods. As shown in the 4th row of Tab. 1, the stylized images of the Attn-AST methods with SCSA are more favored by the public, and their popularity surpasses that of the SOTA methods. These demonstrate the superiority of SCSA. + +# 4.4. Ablation Study + +We conduct ablation experiments to verify the crucial roles of SCA and SSA and to affirm the potency of S-AdaIN. + +![](images/89f979d59d5b8f7f52558385078c333c94e608349fad7cadaf7fc7acb805a650.jpg) +Figure 6. Qualitative and quantitative ablation results on SCA, SSA, and S-AdaIN. Zoom in for better comparison. + +![](images/add8873588041f0364df3f3d211a37776aeb2194afdaf7a159209f12cc7a1f15.jpg) + +Figure 7. The stylization of different content representation. + +
UASCASSASCSA
Runtime (milliseconds)1.2311.9172.0153.944
Memory Usage (bytes)4708479649726220
+ +Table 2. Computation costs for feature fusion modules. + +Effect of SCA and SSA. We use SCA or SSA individually as the feature transformation module to assess their capabilities. As shown in Fig. 6, SCA alone facilitates overall color transfer in corresponding semantic regions but results in texture loss and reduced stylization, as reflected by the higher SSL and FID values. In contrast, SSA can effectively transfer specific textures but introduces color discrepancies between the semantic regions of the stylized and style images. This indicates that while SCA captures the overall style characteristics, SSA is crucial for transferring specific textures, highlighting the complementary roles of both. + +Effect of S-AdaIN. To validate the effectiveness of S-AdaIN, we conducted ablation experiments by removing S-AdaIN. As shown in Fig. 6, without S-AdaIN, the stylized images exhibit less accurate global style transfer in semantic regions and reduced texture details, particularly evident for the StyleID method. Although S-AdaIN was not employed, CFSD values of StyTr² and StyleID with SCSA are lower, which comes at the expense of stylization quality. Therefore, after considering the pros and cons, S-AdaIN is essential and effective for semantic style transfer. + +# 4.5. Discussion + +We perform more discussions about our task and method. + +Arbitrary Semantic Style Transfer. Our plug-and-play SCSA is built on the Attn-AST approaches, unlike methods that iteratively optimize the specific content and style images or only apply a predefined style to any content im + +age, so that it can stylize between any content and style images. Besides, it is important to note that SCSA leverages the user-provided semantic maps to guide the transfer process. Therefore, as long as the user's semantic maps mark the regions where the style should be applied, we can perform the relevant region stylization, as shown in Fig. 7. + +Computation Costs. Since the costs vary with the feature size, we use the features of the $512 \times 512$ images encoded by ReLU4_1 layer of VGG-19 to discuss costs. The results are in Tab. 2. In the 1st row, SCA and SSA have slightly longer runtimes than UA due to the guidance operations. SCSA takes barely longer than the sum of SCA and SSA, as it merges both outputs. The 2nd row shows SCA and SSA use a bit more memory than UA, while SCSA uses less than SCA plus SSA. This is because some operations reduce memory usage by reusing existing variables. Despite the cost increases, this is acceptable and does not impact practical use. Cost refinements are left for future work. + +# 5. Conclusion + +In this paper, we highlight the challenge faced by existing arbitrary style transfer methods, with a particular emphasis on attention-based approaches, and provide an in-depth analysis of this issue, which is the failure to account for the relationship between local regions and semantic regions. To address this issue and generate high-quality stylized images that meet semantic needs, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA. It achieves precise semantic style transfer by allowing each query point to consider certain key points in the corresponding semantic region. Extensive experimental results demonstrate that our SCSA is effective and generalized. + +# Acknowledgments + +This work was partly supported by the National Natural Science Foundation of China No. 62306277, the Fundamental Research Funds for the Zhejiang Provincial Universities Grant No. K20240090, the Research Fund for International Scientists of National Natural Science Foundation of China No. 72350710798, and the Natural Science Foundation of Zhejiang Province No. LDT23F02023F02. This + +research was supported in part through funding from the following sources: the National Natural Science Foundation of China under Award No. 62306277, and the Fundamental Research Funds for Provincial Universities in Zhejiang Province through Grant No. K20240090. + +# References + +[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 1, 2 +[2] Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3 +[3] Alex J Champandard. Semantic style transfer and turning two-bit doodles into fine artworks. arXiv preprint arXiv:1603.01768, 2016. 3 +[4] Dar-Yen Chen. Artfusion: Arbitrary style transfer using dual conditional latent diffusion models. arXiv preprint arXiv:2306.09330, 2023. 1 +[5] Dar-Yen Chen, Hamish Tennent, and Ching-Wen Hsu. Art-adapter: Text-to-image style transfer using multi-level style encoder and explicit adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8619-8628, 2024. 2 +[6] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2 +[7] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 3 +[8] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 1, 2, 3, 6, 7 +[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 1, 2, 3, 6 +[10] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3730-3738. IEEE. 3 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 1 +[12] Mark Hamazaspyan and Shant Navasardyan. Diffusion-enhanced patchmatch: A framework for arbitrary style transfer with diffusion models. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 797-805, 2023. 3 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[14] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 2 +[15] Siyu Huang, Jie An, Donglai Wei, Jiebo Luo, and Hanspeter Pfister. Quantart: Quantizing image style transfer towards high visual fidelity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5947-5956, 2023. 1 +[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4, 5, 7 +[17] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14861-14869, 2021. 3, 6 +[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence, pages 4369-4376, 2020. 2 +[19] Nikolai Kalischek, Jan D Wegner, and Konrad Schindler. In the light of feature distributions: moment matching for neural style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9391, 2021. 2 +[20] Da-Hyun Kim, Dong-Hee Shin, and Tae-Eui Kam. Bridging the bci illiteracy gap: a subject-to-subject semantic style transfer for eeg-based motor imagery classification. Frontiers in Human Neuroscience, 17:1194751, 2023. 3 +[21] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 3, 6 +[22] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In International conference on machine learning, pages 957-966. PMLR, 2015. 3 +[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer, 2018. 1, 2 +[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2 + +[25] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. arXiv preprint arXiv:1705.01088, 2017. 3, 6 +[26] Yi-Sheng Liao and Chun-Rong Huang. Semantic context-aware image style transfer. IEEE Transactions on Image Processing, 31:1911-1923, 2022. 3 +[27] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 1, 2, 3 +[28] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5 +[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5951-5960. IEEE, .3 +[30] Ming Lu, Hao Zhao, Anbang Yao, Feng Xu, Yurong Chen, and Li Zhang. Decoder network over lightweight reconstructed feature for fast semantic style transfer. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2488-2496. IEEE, .3 +[31] Yingnan Ma, Chenqiu Zhao, Xudong Li, and Anup Basu. Rast: Restorable arbitrary style transfer via multi-restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 331-340, 2023. 2 +[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 1, 2, 3, 6 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[35] Onkar Susladkar, Gayatri Deshmukh, Sparsh Mittal, and Parth Shastri. D2styler: Advancing arbitrary style transfer with discrete diffusion methods. arXiv preprint arXiv:2408.03558, 2024. 2 +[36] Gemma Canet Tarres, Dan Ruta, Tu Bui, and John Collosomasse. Parasol: Parametric style control for diffusion image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2432-2442, 2024. 1 +[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[38] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Gliystylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 3, 6 +[39] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: towards + +aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1095-1106, 2022. 1, 2 +[40] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2624-2632, 2022. 3, 6 +[41] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1 +[42] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2 +[43] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 1, 2 +[44] Zhengtao Xiang, Xing Wan, Libo Xu, Xin Yu, and Yuhan Mao. A training-free latent diffusion style transfer method. Information, 15(10):588, 2024. 2 +[45] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 3 +[46] Chiyu Zhang, Xiaogang Xu, Lei Wang, Zaiyan Dai, and Jun Yang. S2wat: Image style transfer via hierarchical vision transformer using strips window attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7024-7032, 2024. 2 +[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[48] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 2 +[49] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. A unified arbitrary style transfer framework via adaptive contrastive learning. ACM Trans. Graph., 42(5), 2023. 1, 2 +[50] Hui-Huang Zhao, Paul L Rosin, Yu-Kun Lai, and Yao-Nan Wang. Automatic semantic style transfer using deep convolutional neural networks and soft masks. The Visual Computer, 36(7):1307-1324, 2020. 3 \ No newline at end of file diff --git a/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/images.zip b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..271272eaf273c642734f1e63667e6d01179414de --- /dev/null +++ b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34b1c48dbfda1790ae72fed66e64546b722dfa1149c07fe50622667956a00aed +size 646095 diff --git a/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/layout.json b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b29e521dc690113149f72bd6c0ee28efba949cec --- /dev/null +++ b/2025/SCSA_ A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer/layout.json @@ -0,0 +1,11542 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "type": "text", + "content": "SCSA: A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 161, + 247, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 161, + 247, + 175 + ], + "spans": [ + { + "bbox": [ + 167, + 161, + 247, + 175 + ], + "type": "text", + "content": "Chunnan Shang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 158, + 176, + 257, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 176, + 257, + 190 + ], + "spans": [ + { + "bbox": [ + 158, + 176, + 257, + 190 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 192, + 279, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 192, + 279, + 203 + ], + "spans": [ + { + "bbox": [ + 137, + 192, + 279, + 203 + ], + "type": "text", + "content": "chunnan.22@intl.zju.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 204, + 246, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 204, + 246, + 217 + ], + "spans": [ + { + "bbox": [ + 169, + 204, + 246, + 217 + ], + "type": "text", + "content": "Hongwei Wang" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 158, + 218, + 257, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 218, + 257, + 232 + ], + "spans": [ + { + "bbox": [ + 158, + 218, + 257, + 232 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 133, + 234, + 282, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 234, + 282, + 245 + ], + "spans": [ + { + "bbox": [ + 133, + 234, + 282, + 245 + ], + "type": "text", + "content": "hongweiwang@intl.zju.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 161, + 440, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 161, + 440, + 176 + ], + "spans": [ + { + "bbox": [ + 356, + 161, + 440, + 176 + ], + "type": "text", + "content": "Zhizhong Wang†" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 346, + 176, + 444, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 176, + 444, + 190 + ], + "spans": [ + { + "bbox": [ + 346, + 176, + 444, + 190 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 192, + 443, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 192, + 443, + 203 + ], + "spans": [ + { + "bbox": [ + 343, + 192, + 443, + 203 + ], + "type": "text", + "content": "endywon@zju.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 203, + 444, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 203, + 444, + 217 + ], + "spans": [ + { + "bbox": [ + 352, + 203, + 444, + 217 + ], + "type": "text", + "content": "Xiangming Meng†" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 218, + 444, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 218, + 444, + 232 + ], + "spans": [ + { + "bbox": [ + 346, + 218, + 444, + 232 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 234, + 474, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 234, + 474, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 234, + 474, + 245 + ], + "type": "text", + "content": "xiangmingmeng@intl.zju.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 298, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 610 + ], + "type": "text", + "content": "Attention-based arbitrary style transfer methods, including CNN-based, Transformer-based, and Diffusion-based, have flourished and produced high-quality stylized images. However, they perform poorly on the content and style images with the same semantics, i.e., the style of the corresponding semantic region of the generated stylized image is inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. To address this issue, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer—each query point considers certain key points in the corresponding semantic region. Specifically, semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region that reflect the overall style characteristics of that region; Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region that exhibits the specific stylistic texture of that region. By combining the two modules, the resulting SCSA aligns the overall style of the corresponding semantic regions while transferring the vivid textures of these regions. Qualitative and quantitative results demonstrate that SCSA enables attention-based arbitrary style transfer methods to produce high-quality semantic stylized images. The code can be found in https://github.com/ecn-00/SCSA." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 693 + ], + "type": "text", + "content": "Initially, style transfer [11] was proposed to preserve the content of a specified content image while transferring the style from a specified style image by iteratively optimizing" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 336, + 274, + 383, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 274, + 383, + 282 + ], + "spans": [ + { + "bbox": [ + 336, + 274, + 383, + 282 + ], + "type": "text", + "content": "Content Inputs" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 335, + 282, + 383, + 331 + ], + "blocks": [ + { + "bbox": [ + 335, + 282, + 383, + 331 + ], + "lines": [ + { + "bbox": [ + 335, + 282, + 383, + 331 + ], + "spans": [ + { + "bbox": [ + 335, + 282, + 383, + 331 + ], + "type": "image", + "image_path": "29539f9393f59eaedf8e4f3df4fe7a2d31336c799a7ea941bd5a2d88c28e2bdf.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 390, + 274, + 428, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 274, + 428, + 282 + ], + "spans": [ + { + "bbox": [ + 390, + 274, + 428, + 282 + ], + "type": "text", + "content": "Style Inputs" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 385, + 282, + 433, + 331 + ], + "blocks": [ + { + "bbox": [ + 385, + 282, + 433, + 331 + ], + "lines": [ + { + "bbox": [ + 385, + 282, + 433, + 331 + ], + "spans": [ + { + "bbox": [ + 385, + 282, + 433, + 331 + ], + "type": "image", + "image_path": "1edea075c59643bdb6e64fc76bd85e63f217067de0a7c98b0dc932efea7ee423.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 448, + 274, + 470, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 274, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 448, + 274, + 470, + 281 + ], + "type": "text", + "content": "SANet" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 435, + 282, + 483, + 331 + ], + "blocks": [ + { + "bbox": [ + 435, + 282, + 483, + 331 + ], + "lines": [ + { + "bbox": [ + 435, + 282, + 483, + 331 + ], + "spans": [ + { + "bbox": [ + 435, + 282, + 483, + 331 + ], + "type": "image", + "image_path": "d3d5ff1194c5f66d5f494a69365fc6802c0f2c1a2c4904050a792ba40b25d14e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 485, + 282, + 533, + 331 + ], + "blocks": [ + { + "bbox": [ + 485, + 274, + 533, + 281 + ], + "lines": [ + { + "bbox": [ + 485, + 274, + 533, + 281 + ], + "spans": [ + { + "bbox": [ + 485, + 274, + 533, + 281 + ], + "type": "text", + "content": "SANet + SCSA" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 282, + 533, + 331 + ], + "lines": [ + { + "bbox": [ + 485, + 282, + 533, + 331 + ], + "spans": [ + { + "bbox": [ + 485, + 282, + 533, + 331 + ], + "type": "image", + "image_path": "69520fff0fd62612fbfa196b714b03310c301051644920d01f06395aa510dd0e.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 335, + 344, + 384, + 392 + ], + "blocks": [ + { + "bbox": [ + 337, + 335, + 382, + 343 + ], + "lines": [ + { + "bbox": [ + 337, + 335, + 382, + 343 + ], + "spans": [ + { + "bbox": [ + 337, + 335, + 382, + 343 + ], + "type": "text", + "content": "Content Inputs" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 344, + 384, + 392 + ], + "lines": [ + { + "bbox": [ + 335, + 344, + 384, + 392 + ], + "spans": [ + { + "bbox": [ + 335, + 344, + 384, + 392 + ], + "type": "image", + "image_path": "15e701c7f8fbe11b2a43211c375f3c3348b8c91b1ca4f31f4bffa759fc87c0db.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 385, + 344, + 433, + 392 + ], + "blocks": [ + { + "bbox": [ + 391, + 334, + 427, + 344 + ], + "lines": [ + { + "bbox": [ + 391, + 334, + 427, + 344 + ], + "spans": [ + { + "bbox": [ + 391, + 334, + 427, + 344 + ], + "type": "text", + "content": "Style Inputs" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 344, + 433, + 392 + ], + "lines": [ + { + "bbox": [ + 385, + 344, + 433, + 392 + ], + "spans": [ + { + "bbox": [ + 385, + 344, + 433, + 392 + ], + "type": "image", + "image_path": "b07a3ebe2a8e54b28e9f8d93d9acfe4ef59b7447b6dc2499044183fefc17b97d.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 435, + 344, + 483, + 392 + ], + "blocks": [ + { + "bbox": [ + 447, + 334, + 469, + 343 + ], + "lines": [ + { + "bbox": [ + 447, + 334, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 447, + 334, + 469, + 343 + ], + "type": "text", + "content": "StyTr2" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 344, + 483, + 392 + ], + "lines": [ + { + "bbox": [ + 435, + 344, + 483, + 392 + ], + "spans": [ + { + "bbox": [ + 435, + 344, + 483, + 392 + ], + "type": "image", + "image_path": "596510bbb4c3721d65ef49ee5f6bd73a985709a969f13ca19a8534e36b7e10c8.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 485, + 344, + 533, + 392 + ], + "blocks": [ + { + "bbox": [ + 485, + 334, + 532, + 343 + ], + "lines": [ + { + "bbox": [ + 485, + 334, + 532, + 343 + ], + "spans": [ + { + "bbox": [ + 485, + 334, + 532, + 343 + ], + "type": "inline_equation", + "content": "\\mathrm{StyTr}^2 + \\mathrm{SCSA}" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 344, + 533, + 392 + ], + "lines": [ + { + "bbox": [ + 485, + 344, + 533, + 392 + ], + "spans": [ + { + "bbox": [ + 485, + 344, + 533, + 392 + ], + "type": "image", + "image_path": "2db4d58977d0db3840ba12e85802fbb942e8b5673f201919e2d24d92f7bca1cc.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 335, + 404, + 384, + 453 + ], + "blocks": [ + { + "bbox": [ + 337, + 395, + 382, + 404 + ], + "lines": [ + { + "bbox": [ + 337, + 395, + 382, + 404 + ], + "spans": [ + { + "bbox": [ + 337, + 395, + 382, + 404 + ], + "type": "text", + "content": "Content Inputs" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 404, + 384, + 453 + ], + "lines": [ + { + "bbox": [ + 335, + 404, + 384, + 453 + ], + "spans": [ + { + "bbox": [ + 335, + 404, + 384, + 453 + ], + "type": "image", + "image_path": "09a63a75528b1fbb8509ed47479af42211873702bebafe06105d70d17228a1e7.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 457, + 555, + 546 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 546 + ], + "type": "text", + "content": "Figure 1. Comparisons of the Attn-AST approaches-CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]-without and with our SCSA. The stylized images generated by the Attn-AST approaches exhibit style discontinuity in adjacent regions within the identical semantic regions (e.g., the background in the 1st row), style inconsistency between corresponding semantic (e.g., the cloth color in the 3rd row), and fewer textures (e.g., the block textures in the 3rd row)." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 385, + 404, + 433, + 453 + ], + "blocks": [ + { + "bbox": [ + 391, + 395, + 427, + 404 + ], + "lines": [ + { + "bbox": [ + 391, + 395, + 427, + 404 + ], + "spans": [ + { + "bbox": [ + 391, + 395, + 427, + 404 + ], + "type": "text", + "content": "Style Inputs" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 404, + 433, + 453 + ], + "lines": [ + { + "bbox": [ + 385, + 404, + 433, + 453 + ], + "spans": [ + { + "bbox": [ + 385, + 404, + 433, + 453 + ], + "type": "image", + "image_path": "ff63f632d36177c00a54774ebecab33fc7a25d93f3a241aab2362609c72e2a84.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 435, + 404, + 483, + 453 + ], + "blocks": [ + { + "bbox": [ + 447, + 396, + 470, + 404 + ], + "lines": [ + { + "bbox": [ + 447, + 396, + 470, + 404 + ], + "spans": [ + { + "bbox": [ + 447, + 396, + 470, + 404 + ], + "type": "text", + "content": "StyleID" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 404, + 483, + 453 + ], + "lines": [ + { + "bbox": [ + 435, + 404, + 483, + 453 + ], + "spans": [ + { + "bbox": [ + 435, + 404, + 483, + 453 + ], + "type": "image", + "image_path": "9531137a9b4cdacae2a503bbec2bdab64926ebc500a11531a6f7d7afe7a24a17.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 485, + 404, + 533, + 453 + ], + "blocks": [ + { + "bbox": [ + 484, + 395, + 533, + 403 + ], + "lines": [ + { + "bbox": [ + 484, + 395, + 533, + 403 + ], + "spans": [ + { + "bbox": [ + 484, + 395, + 533, + 403 + ], + "type": "text", + "content": "StyleID + SCSA" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 485, + 404, + 533, + 453 + ], + "lines": [ + { + "bbox": [ + 485, + 404, + 533, + 453 + ], + "spans": [ + { + "bbox": [ + 485, + 404, + 533, + 453 + ], + "type": "image", + "image_path": "6fede9470022d5f96840822978dd18f1779f9560b937c83c7137b440777a05c8.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "the generated stylized image. However, its excessive slowness prevents practical application. Given the need for efficiency, effectiveness, and generalization, Arbitrary Style Transfer (AST) [1, 4, 8, 23, 27] has become a highly prominent research focus in the style transfer field. AST seeks to generate a new image that replicates the content of an arbitrary content image while adopting the style of an arbitrary style image. In spite of many recent studies aiming to advance the AST effect by incorporating novel knowledge [15, 36, 39] and innovative constraints [41, 43, 49], the core of AST continues to reside in feature transformation modules, which remains indispensable for all approaches. Hence, the existing AST approaches can be classified into" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 148, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13051" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "two categories, statistics-based methods [5, 16, 24, 48] and attention-based methods [6, 8, 9, 27, 31, 32]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 295, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 295, + 278 + ], + "type": "text", + "content": "Attention-based Arbitrary Style Transfer (Attn-AST) approaches, leveraging the ability to capture global information and focus on the related features of the attention mechanism [2], preserve the content of the content image while transferring the style that is relevant to the content of the content image from the style image. They encompass CNN-based [27, 31, 32, 39], Transformer-based [9, 14, 42, 46], and Diffusion-based methods [8, 35, 44]. Commonly, Attn-AST methods first establish a holistic relationship, an attention map, between each content query point of the content image features and all style key points of the style image features. Then, Attn-AST methods get the transferred style for each content point by weighting all the style points, where the weights are the softmax function outputs of a naturally calculated attention map of content and style points." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "text", + "content": "Although Attn-AST methods have thrived and produced high-quality stylized images due to the capacity mentioned above for global capture and focus, they perform poorly on the content and style image with the same semantics. As shown in Fig. 1, the style of the corresponding semantic region of the stylized image generated by Attn-AST approaches exhibits a style inconsistent with that of the style image. We argue that the root cause lies in their failure to consider the relationship between local regions and semantic regions. (1) During the formulation of the attention map, Attn-AST methods bluntly and simplistically consider the relationship between each query point and all key points, as shown in Fig. 2 (c), without adequately accounting for the semantic regions to which these points belong. This oversimplification may cause Attn-AST methods to disproportionately emphasize structurally similar points across different semantic regions, resulting in artifacts or inaccurate style [27, 32], e.g., the " + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "inline_equation", + "content": "2nd" + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "text", + "content": " row of Fig. 1. (2) Attn-AST approaches fundamentally depend on structural principles for generating a global style, as they utilize content and style image features. The subtle disparity in content structure details across adjacent regions can cause different stylization effects. Thus, the stylized images often lack stylistic continuity across adjacent regions within the semantic regions, e.g., the background in the " + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "inline_equation", + "content": "1st" + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "text", + "content": " row of Fig. 1. (3) Attn-AST approaches acquire the transferred style by weighting all the style points. However, the most accurate representation of the original stylistic texture resides in the encoded single points. The weighted style struggles to preserve the original stylistic texture, leading to a lack of vividness, e.g., the block textures in the " + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "inline_equation", + "content": "3rd" + }, + { + "bbox": [ + 55, + 280, + 295, + 651 + ], + "type": "text", + "content": " row of Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "To maintain the benefits of Attn-AST methods while addressing their existing issues, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA, for arbitrary semantic style transfer. The key insight of SCSA is that, based on the interpretability of the attention mecha" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "content": "nism, each query point considers certain key points in the corresponding semantic region. Specifically, SCSA consists of two parts: Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA). (1) SCA: Semantic continuous attention ensures each query point fully attends to all the continuous key points in the same semantic region by using content and style semantic map features as query and key, respectively. Thus, SCA can achieve the overall style characteristics of the corresponding semantic regions of the style image to generate the stylized image with stylistic continuity across adjacent regions within the semantic regions. (2) SSA: Semantic sparse attention allows each query point to focus on the most similar sparse key point in the same semantic region by using content and style image features as query and key, respectively. Hence, SSA can obtain the specific stylistic texture of the corresponding semantic regions of the style image to produce the stylized image with vivid stylistic textures. The intuitive comparison between our SCSA and universal attention can be found in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 300, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 371 + ], + "type": "text", + "content": "By seamlessly integrating SCSA into any Attn-AST approach without training, we enable arbitrary semantic style transfer while preserving and even enhancing the original stylization effect. To our knowledge, we are the first work to extend attention-based arbitrary style transfer to arbitrary semantic style transfer in a plug-and-play way." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 372, + 543, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 372, + 543, + 383 + ], + "spans": [ + { + "bbox": [ + 326, + 372, + 543, + 383 + ], + "type": "text", + "content": "In brief, our contributions are divided into three parts:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 384, + 553, + 526 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 326, + 384, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 384, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 326, + 384, + 553, + 418 + ], + "type": "text", + "content": "- We reveal that Attn-AST methods struggle with content and style images with the same semantics and identify the root causes for their subpar performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 419, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 419, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 326, + 419, + 553, + 455 + ], + "type": "text", + "content": "- We propose a semantic continuous-sparse attention, dubbed SCSA, that can extend Attn-AST to arbitrary semantic style transfer in a plug-and-play manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 456, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 456, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 326, + 456, + 553, + 526 + ], + "type": "text", + "content": "- We conduct extensive experiments and comparisons to demonstrate the effectiveness and generalization of SCSA, enabling Attn-AST methods (CNN-based, Transformer-based, and Diffusion-based approaches) to perform arbitrary semantic style transfer while preserving and even enhancing the original stylization." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 538, + 400, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 400, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 400, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": "The existing AST approaches can be roughly divided into two groups: statistics-based and attention-based methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Statistics-based Methods. Statistics-based AST methods [1, 18, 23, 43, 49] regard style as the statistical traits inherent in feature distribution. Specifically, they achieve stylized features by aligning the feature statistics of the content image with those of the style image, making them a global style transfer approach. The two most representative approaches are AdaIN [16] and WCT [24], which match the first-order and second-order statistics of features, respectively. To more comprehensively capture the stylistic characteristics of features, higher-order statistical methods, such as CMD [19] and EFDM [48], have been continually ex" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13052" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 217, + 180 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 217, + 180 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 217, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 217, + 180 + ], + "type": "image", + "image_path": "a8a070f469eff37ca8ddaf7419da0c7d5864bf20d620a954475bf76ed0ff2d75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 185, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 54, + 185, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 54, + 185, + 555, + 274 + ], + "type": "text", + "content": "Figure 2. Comparison between Semantic Continuous-Sparse Attention (SCSA) and Universal Attention (UA). SCSA includes two parts: (a) Semantic Continuous Attention (SCA): The query point of the content semantic map features can match all continuous key points of the style semantic map features in the same semantic region. Therefore, SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of regions with the same semantics; (b) Semantic Sparse Attention (SSA): The query point of the content image features can match the most similar sparse key point of the style image features in the same semantic region. Hence, SSA can intently concentrate on the specific stylistic texture of regions with the same semantics. In contrast, (c) Universal Attention (UA) [32]: The query point of the content image features pays attention to all key points of the style image features. As a result, UA fails to accurately convey the intricate overall stylistic characteristics and specific textures of regions that share identical semantics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 386, + 179 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 386, + 179 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 386, + 179 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 386, + 179 + ], + "type": "image", + "image_path": "a91ff56b4b3596a3de4e931eaa36fb179c120abd4b24789adea2f56ea1cce33b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 555, + 179 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 555, + 179 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 555, + 179 + ], + "type": "image", + "image_path": "4c153bda1c4d810bec18fc672a835f943d801afa7446fb4cace1a6be8a3577a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 287, + 294, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 294, + 395 + ], + "type": "text", + "content": "plored and refined. Expanding on the above foundation, semantic style transfer [10, 20, 26, 29, 30, 50] has been studied by aligning feature statistics within corresponding semantic regions. STROTSS [21] employs the Earth Movers Distance [22] to reduce the feature distribution difference within semantic regions and self-similarity to preserve the semantics and spatial structure. MAST [17] aligns the manifold distributions of semantic regions to facilitate semantic style transfer while preserving the content structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 399, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 295, + 471 + ], + "type": "text", + "content": "Since feature statistics only provide a global representation, the stylized images they generate often lack fine local textures. To address this, we propose SCSA, which could yield a coherent overall style across adjacent regions by overlooking structure details, while producing vivid textures by emphasizing them via the attention mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": "Attention-based Methods. Attention-based AST (Attn-AST) methods define style as the weighted convergence of all style features. Exactly, Attn-AST methods obtain the stylized feature of each content feature by weighting all the style features, where weights are derived from the correlation between the content feature and all style features obtained through the attention mechanism, making it a local structure style transfer. Attn-AST approaches mainly include CNN-based methods [27, 32], Transformer-based methods [9], and Diffusion-based methods [8]. SANet [32] is the pioneering work to integrate attention into AST, employing content features as query and style features as key and value to obtain stylized features via the attention mechanism. StyTr² [9] integrates the outputs of the content and style transformer [37] encoders into the transformer feature fusion module, obtaining stylized features. StyleID [8] injects the content query, style key, and style value, derived from DDIM inversion [34], into certain self-attention layers of U-Net [33] to merge style, achieving stylized features through multiple inversion steps. Even though they" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 287, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 467 + ], + "type": "text", + "content": "continually improve the quality of stylized images, their performance remains poor on content and style images with identical semantics. In contrast, patch-based methods [3, 7, 12, 38, 45] achieve semantic style transfer by swapping the styles of patches with the most similar content structures. When the patch size in patch-based methods is 1, they can be roughly considered attention-based methods. Deep-Image-Analogy (DIA) [25] introduces nearest-neighbor field search to tackle the challenge of patch matching for images, which are visually distinct yet semantically similar. GLStyleNet [38] uses multi-level pyramid features for semantic optimization via patch matching, iteratively creating stylized images that meet semantic criteria. TR [40] utilizes the patch match to achieve precise semantic-guided and structure-preserving texture transfer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 468, + 554, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 587 + ], + "type": "text", + "content": "As the above methods target local structural similarity, subtle changes in the content structure of adjacent regions can result in stylization discontinuity in the identical semantic region. To mitigate this concern, we introduce SCSA, which provides more stable and continuous stylization by fully comprehending semantics and transfers vivid stylistic textures by focusing on local structure. Most importantly, it can be easily integrated into existing frameworks, while other methods are either designed for specific networks or require multiple iterations for optimization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 623, + 519, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 519, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 519, + 635 + ], + "type": "text", + "content": "3.1. Semantic Continuous-Sparse Attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "Revisit Universal Attention in Attn-AST. Attn-AST approaches feed the encoded content feature and style feature into Universal Attention (UA), constructing intricate connections between various positions within the two feature sequences by directly utilizing the self-attention [2] computational paradigm, to capture the long-range local textures" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13053" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 157, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 157, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 157, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 157, + 202 + ], + "type": "image", + "image_path": "2c3af095416a55cda99882831fec17124c534a0f2d7f8ae2ddea7d2cd03a5147.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": "Figure 3. Detailed Procedure of Semantic Continuous-Sparse Attention (SCSA). In (a), S-AdaIN denotes that AdaIN [16] is applied individually for each semantic region. It initializes " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": ", by matching the feature statistics of corresponding semantic regions in " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": ", to ensure that " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " has semantically aligned color style information to some extent. Subsequently, " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " can provide a more accurate query " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "f_{q}(\\bar{F}_{c})" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " and more pure content structure features with less interference from the original color style. SCSA in (a) includes Semantic Continuous Attention (SCA) and Semantic Sparse Attention (SSA) two parts. \"G1\" in SCA sets the value of the specific product point generated by multiplying two points from different semantic categories to negative infinity. As shown in (b), only the product of points with the same semantics is retained. Therefore, SCA can fully account for the overall stylistic characteristics of regions with the same semantics. \"G2\" in SSA retains only the maximum value of the product of a specified query point with all key points in the same semantic region, while all other values are set to negative infinity, as shown in (c). Hence, SSA can intently concentrate on the specific stylistic texture of the most similar structure in the regions with the same semantics. After SCSA, the generated " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " from (a) not only reflects the overall stylistic characteristics of the semantic region corresponding to " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " but also captures the specific textures of that semantic region. Meanwhile, " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": " also includes the content structure of " + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 211, + 296, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 71, + 296, + 204 + ], + "blocks": [ + { + "bbox": [ + 160, + 71, + 296, + 204 + ], + "lines": [ + { + "bbox": [ + 160, + 71, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 160, + 71, + 296, + 204 + ], + "type": "image", + "image_path": "d186ca292dded83a9b23844c0e0538c4c6b7733c94a334f8aba382c3959e2180.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "type": "text", + "content": "relationship. Typically, given the encoded content features " + }, + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "type": "text", + "content": " and style features " + }, + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 488, + 296, + 525 + ], + "type": "text", + "content": ", the query (Q), key (K), and value (V) of UA can be obtained as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 533, + 295, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 533, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 90, + 533, + 295, + 547 + ], + "type": "interline_equation", + "content": "Q = f _ {q} \\left(\\bar {F} _ {c}\\right), K = f _ {k} \\left(\\bar {F} _ {s}\\right), V = f _ {v} \\left(F _ {s}\\right), \\tag {1}", + "image_path": "48e50ad3e7571df93b4645b58bdde2cb1bb6a64a07e36691c18869919377d455.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\bar{F}_c" + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\bar{F}_s" + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": " denote the normalized " + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "inline_equation", + "content": "f_{q}, f_{k}" + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 554, + 295, + 590 + ], + "type": "text", + "content": " are the project networks for getting query, key, and value in UA of the given Attn-AST model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 591, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 638 + ], + "type": "text", + "content": "After that, the attention map " + }, + { + "bbox": [ + 55, + 591, + 296, + 638 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 591, + 296, + 638 + ], + "type": "text", + "content": ", which represents a holistic relationship between each content query point of the content features and all style key points of the style features, can be obtained as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 145, + 645, + 295, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 645, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 145, + 645, + 295, + 660 + ], + "type": "interline_equation", + "content": "S = Q ^ {\\mathsf {T}} \\otimes K, \\tag {2}", + "image_path": "b4d48c4a5300e09438426b8c72a21d4060c34408f0427df34cc2ba16f98ac208.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 669, + 216, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 216, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 216, + 680 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 669, + 216, + 680 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 669, + 216, + 680 + ], + "type": "text", + "content": " denotes matrix multiplication." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 681, + 285, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 681, + 285, + 692 + ], + "spans": [ + { + "bbox": [ + 67, + 681, + 285, + 692 + ], + "type": "text", + "content": "Finally, the stylized features " + }, + { + "bbox": [ + 67, + 681, + 285, + 692 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 67, + 681, + 285, + 692 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 102, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "F _ {c s} = f _ {o} (\\text {s o f t m a x} (S) \\otimes V) + F _ {c}, \\tag {3}", + "image_path": "11ca20ecf8d01559501219d418cfffc430e1cec4479be4e62ca8a9be90050347.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " is the output project networks in UA of the specified Attn-AST model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "content": "Although Attn-AST methods utilizing UA can achieve high-quality stylization results, they still underperform on content and style images with identical semantics due to inherent limitations, as detailed analysis in Sec. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "content": "Semantic Continuous Attention. To resolve the problem of the style discontinuity in adjacent regions within the identical semantic regions of the stylized image, as well as style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we propose semantic continuous attention, termed SCA, as shown in Fig. 2 (a) and detailed in Fig. 3. This method disregards the image structures while considering the relationship between local regions and semantic regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": "In particular, SCA first takes the encoded content semantic map features " + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "inline_equation", + "content": "F_{csem}" + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": " and style semantic map features " + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "inline_equation", + "content": "F_{ssem}" + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": " as inputs for the query " + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "inline_equation", + "content": "(Q_1)" + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "inline_equation", + "content": "(K_1)" + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": ", respectively, while the encoded style image features " + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": " are used as inputs for the value " + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "inline_equation", + "content": "(V_1)" + }, + { + "bbox": [ + 313, + 252, + 554, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 318, + 553, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 318, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 323, + 318, + 553, + 333 + ], + "type": "interline_equation", + "content": "Q _ {1} = f _ {q} (\\bar {F} _ {c s e m}), K _ {1} = f _ {k} (\\bar {F} _ {s s e m}), V _ {1} = f _ {v} (F _ {s}), \\tag {4}", + "image_path": "3c3ed9d63c51b1ef6ef5c965e091cbc5ae57715b5cf29737e206082121269c4c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "inline_equation", + "content": "\\bar{F}_{csem}" + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "inline_equation", + "content": "\\bar{F}_{ssem}" + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": " represent the normalized form of " + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "inline_equation", + "content": "F_{csem}" + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "inline_equation", + "content": "F_{ssem}" + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "inline_equation", + "content": "f_{q}, f_{k}" + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 313, + 338, + 554, + 374 + ], + "type": "text", + "content": " are the same as those in Eq. 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 375, + 552, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 375, + 552, + 387 + ], + "spans": [ + { + "bbox": [ + 326, + 375, + 552, + 387 + ], + "type": "text", + "content": "Then, the preliminary attention map " + }, + { + "bbox": [ + 326, + 375, + 552, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 326, + 375, + 552, + 387 + ], + "type": "text", + "content": " can be obtained:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 400, + 392, + 553, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 392, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 400, + 392, + 553, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = Q _ {1} ^ {\\mathrm {T}} \\otimes K _ {1}, \\tag {5}", + "image_path": "bc28d0c0dcb31f6945ec3d2ecc570d5c5383cff743bb39b1464269c9f9addb82.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": " denotes matrix multiplication. Since " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "F_{csem}" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "F_{ssem}" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": " do not encompass the content structure of the content and style images, after attention computation, the correlations between content points within the same semantic region and all style points are identical. That is if the content query points " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": " belong to the same semantic category, then " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{im} = \\mathcal{A}_{jm}" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 414, + 554, + 533 + ], + "type": "text", + "content": " represents any style key point. Consequently, the issue of stylization inconsistency in adjacent areas of the same semantic region, stemming from structural differences, is effectively addressed." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "type": "text", + "content": "To make SCA focus on regions with the same semantics to transfer a coherent overall style, we modulate the attention map " + }, + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "type": "text", + "content": " with operation " + }, + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "type": "inline_equation", + "content": "G_{1}" + }, + { + "bbox": [ + 313, + 534, + 554, + 594 + ], + "type": "text", + "content": ", setting attention weights of two points from different semantic categories to negative infinity and those within the same category unchanged:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 405, + 600, + 460, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 600, + 460, + 613 + ], + "spans": [ + { + "bbox": [ + 405, + 600, + 460, + 613 + ], + "type": "interline_equation", + "content": "\\bar {\\mathcal {A}} = G _ {1} (\\mathcal {A}),", + "image_path": "672394bfe95d106b088be1945c7da19eb92ed863ae33f7e2f683c711afa12576.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 348, + 615, + 553, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 615, + 553, + 648 + ], + "spans": [ + { + "bbox": [ + 348, + 615, + 553, + 648 + ], + "type": "interline_equation", + "content": "G _ {1} \\left(\\mathcal {A} _ {q k}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {A} _ {q k}, \\text {i f} q \\in c \\text {a n d} k \\in c \\\\ - \\infty , \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {6}", + "image_path": "2b95aa79f308f3eb8e7b3d9db2bebf094a8ca8c764be458eaaceccde985cde18.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " represent the query and key point of " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "Q_{1}" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "K_{1}" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " is the semantic category. " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\bar{A}" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " is the new attention map that captures the relationship exclusively between content query points and style key points of the same semantic category. This is visually shown in Fig. 3 (b)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13054" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 279, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 279, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 279, + 84 + ], + "type": "text", + "content": "Ultimately, we can obtain the stylized features " + }, + { + "bbox": [ + 67, + 72, + 279, + 84 + ], + "type": "inline_equation", + "content": "F_{sca}" + }, + { + "bbox": [ + 67, + 72, + 279, + 84 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 94, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 94, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 110, + 94, + 295, + 107 + ], + "type": "interline_equation", + "content": "F _ {s c a} = f _ {o} \\left(\\text {s o f t m a x} (\\bar {\\mathcal {A}}) \\otimes V _ {1}\\right), \\tag {7}", + "image_path": "aa2216dd55ce07d9c6460f1daf51e810c2fe4e4ea7fc35bdae38a00579d2ed18.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 118, + 296, + 248 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 55, + 118, + 296, + 248 + ], + "type": "text", + "content": " is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of different content points and all style points in the same semantic region are equal. Thus, each query point can match all continuous key points in the same semantic region. SCA can fully account for the overall stylistic characteristics (e.g., color and texture) of the same semantic regions in the style image while maintaining style consistency in adjacent areas of the identical semantic region in the stylized image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "text", + "content": "Semantic Adaptive Instance Normalization. To provide more pure content structure features with less interference from the original color style and a more accurate query for computing the relationships of local structures in corresponding semantic regions of the content and style images, we apply AdaIN [16] to the encoded content image features " + }, + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "text", + "content": " and style image features " + }, + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "text", + "content": " for each semantic region:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 343, + 269, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 343, + 269, + 360 + ], + "spans": [ + { + "bbox": [ + 70, + 343, + 269, + 360 + ], + "type": "interline_equation", + "content": "I _ {c s e m} \\xrightarrow {K M} M _ {c} ^ {i}, I _ {s s e m} \\xrightarrow {K M} M _ {s} ^ {i}, i = 1, \\dots , N,", + "image_path": "31afc325c40b195c0a72867eed3937e4df395b351e6584444d7c3c48d7b3963f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 362, + 294, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 362, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 102, + 362, + 294, + 380 + ], + "type": "interline_equation", + "content": "F _ {c} ^ {i} = F _ {c} ^ {i} \\odot M _ {c} ^ {i}, \\quad F _ {s} ^ {i} = F _ {s} ^ {i} \\odot M _ {s} ^ {i}, \\tag {8}", + "image_path": "e59b1516c94d4e2c2a801ac0dd919967d80bc6dc620cc737d427178059d13e12.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 380, + 247, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 380, + 247, + 407 + ], + "spans": [ + { + "bbox": [ + 92, + 380, + 247, + 407 + ], + "type": "interline_equation", + "content": "F _ {c} ^ {i} = \\frac {F _ {c} ^ {i} - \\mu (F _ {c} ^ {i})}{\\sigma (F _ {c} ^ {i})} \\times \\sigma (F _ {s} ^ {i}) + \\mu (F _ {s} ^ {i}),", + "image_path": "10c25a318a9619881a1b4e657d16e53801aa743335b54668f0c8b7864a1295b3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "KM" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": " refers to K-Means clustering [28]. " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "M_c^i" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "M_s^i" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": " are the masks containing only 0 and 1. " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_c^i" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_s^i" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": " represent the features belong to corresponding semantic category " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": ". Then, we obtain the features " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_c = \\sum_i F_c^i" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": " whose overall styles of relevant semantic regions align with those of " + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 55, + 415, + 296, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 476, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 296, + 583 + ], + "type": "text", + "content": "Semantic Sparse Attention. To preserve the original vivid stylistic texture of the style image and address the issue of style inconsistency between corresponding semantic regions of stylized and style images, as analyzed in Sec. 1, we introduce semantic sparse attention, known as SSA, as shown in Fig. 2 (b) and Fig. 3. This method focuses solely on preserving the style texture with the highest weights, rather than weighting all the style textures, while considering the relationship between local and semantic regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "content": "SSA first employs the transformed content features " + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "content": " as inputs for the query " + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "inline_equation", + "content": "(Q_{2})" + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "content": " and the encoded style image features " + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "content": " as inputs for both the key " + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "inline_equation", + "content": "(K_{2})" + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "inline_equation", + "content": "(V_{2})" + }, + { + "bbox": [ + 55, + 584, + 295, + 620 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 630, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 630, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 85, + 630, + 295, + 644 + ], + "type": "interline_equation", + "content": "Q _ {2} = f _ {q} \\left(\\bar {F} _ {c}\\right), K _ {2} = f _ {k} \\left(\\bar {F} _ {s}\\right), V _ {2} = f _ {v} \\left(F _ {s}\\right), \\tag {9}", + "image_path": "da204e743422e2338a1c47c55baf95dbdfb803513e5d06a97bf3488c7dfcaa7d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "inline_equation", + "content": "\\bar{F}_c" + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "inline_equation", + "content": "\\bar{F}_s" + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": " represent the normalized " + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "inline_equation", + "content": "f_{q}, f_{k}" + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 653, + 295, + 678 + ], + "type": "text", + "content": " are the same as those in Eq. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 678, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 295, + 690 + ], + "type": "text", + "content": "Then, the initial attention map " + }, + { + "bbox": [ + 67, + 678, + 295, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 67, + 678, + 295, + 690 + ], + "type": "text", + "content": " of SSA can be obtained:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 143, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {B} = Q _ {2} ^ {\\mathrm {T}} \\otimes K _ {2}, \\tag {10}", + "image_path": "102f46de1e561bbe257306f421ecff99e40869216b270262bd0274d474d6d813.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " denotes matrix multiplication. Since " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " capture the structures of the content and style images, " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " can consider the feature correlations based on local structures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": "We claim that the most accurate representation of the original stylistic texture resides in the encoded discrete feature point and the weighted style feature points struggle to preserve the original vivid stylistic textures. Thus, to get accurate and vivid textures, we adjust attention map " + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": " with operation " + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "inline_equation", + "content": "G_{2}" + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": " to target the discrete feature point within related regions, setting the only maximum attention weight of a content query point and all style key points in the same semantic category unchanged and others to negative infinity:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 405, + 220, + 459, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 220, + 459, + 235 + ], + "spans": [ + { + "bbox": [ + 405, + 220, + 459, + 235 + ], + "type": "interline_equation", + "content": "\\overline {{\\boldsymbol {B}}} = G _ {2} (\\mathcal {B}),", + "image_path": "74d64936f759e1ba5639ffbd1fd56490e6e927263be6a401329f7b040e8c4c95.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 237, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 237, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 319, + 237, + 553, + 287 + ], + "type": "interline_equation", + "content": "G _ {2} \\left(\\mathcal {B} _ {q k}\\right) = \\left\\{ \\begin{array}{l} \\mathcal {B} _ {q k}, \\text {i f} \\mathcal {B} _ {q k} = \\max \\left(\\mathcal {B} _ {q n}\\right), \\\\ n \\in \\{1, \\dots , \\operatorname {l e n} \\left(K _ {2}\\right) \\}, q, k, n \\in c \\\\ - \\infty , \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {11}", + "image_path": "983c8f4625faddaa25613a856d396415a7da35eb53dd7366805faad7199ad1f9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": " represents the query point of " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "Q_{2}" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": " represent the key points of " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "K_{2}" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": " denotes the semantic category. " + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 288, + 554, + 347 + ], + "type": "text", + "content": " is the new attention map that can grasp the relationship between the content and style feature points, which share the most similar structure. This is visually shown in Fig. 3 (c)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 348, + 538, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 348, + 538, + 360 + ], + "spans": [ + { + "bbox": [ + 325, + 348, + 538, + 360 + ], + "type": "text", + "content": "Eventually, we can obtain the stylized features " + }, + { + "bbox": [ + 325, + 348, + 538, + 360 + ], + "type": "inline_equation", + "content": "F_{ssa}" + }, + { + "bbox": [ + 325, + 348, + 538, + 360 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 365, + 553, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 365, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 369, + 365, + 553, + 378 + ], + "type": "interline_equation", + "content": "F _ {s s a} = f _ {o} \\left(\\text {s o f t m a x} (\\bar {\\mathcal {B}}) \\otimes V _ {2}\\right), \\tag {12}", + "image_path": "f06e88a8b571fd61a6bea405ae07b9c71332f6a21b8d052b69ef7fe99987b3ac.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 384, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 384, + 554, + 515 + ], + "type": "inline_equation", + "content": "f_{o}" + }, + { + "bbox": [ + 313, + 384, + 554, + 515 + ], + "type": "text", + "content": " is the same as that in Eq. 1. Following the above operation, the softmax function could ensure that the weights of content and style points in different semantic regions become 0, while the weights of each content point and the most similar style point in the same semantic region are 1 and weights of this content point and other style points in the same semantic region are 0. Thus, each query point can match sparse key points in the same semantic region. SSA can intently concentrate on the specific stylistic texture of regions with the same semantics to transfer them to the matching semantic region in the stylized image." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "type": "text", + "content": "Feature Fusion. With the " + }, + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "type": "inline_equation", + "content": "F_{sca}" + }, + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "type": "inline_equation", + "content": "F_{ssa}" + }, + { + "bbox": [ + 313, + 516, + 554, + 539 + ], + "type": "text", + "content": " of SCA and SSA, the feature fusion can be executed:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 358, + 546, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 546, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 358, + 546, + 553, + 559 + ], + "type": "interline_equation", + "content": "F _ {c s} = \\alpha_ {1} \\times F _ {s c a} + \\alpha_ {2} \\times F _ {s s a} + F _ {c}, \\tag {13}", + "image_path": "632da3dded331c5162f76b87cd4a640a988c4f9b79c0b36299cda32d4990a5df.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "text", + "content": " denote the stylization degree of the overall style and vivid textures separately. " + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "text", + "content": " are the outputs of Eq. 8 for superior styling. " + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 313, + 563, + 554, + 601 + ], + "type": "text", + "content": " are the final stylized features." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 430, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 430, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 430, + 617 + ], + "type": "text", + "content": "3.2. Overall Framework" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": "Revisit Attn-AST Approaches. Given a content image " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": " and a style image " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": ", the Attn-AST model " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": " could produce a stylized image " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": ", which preserves the content of " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": " and exhibit the style of " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": "Firstly, they achieve the encoded content features " + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": " and style features " + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": " through the encoder " + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 380, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 380, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "F _ {c} = E \\left(I _ {c}\\right), F _ {s} = E \\left(I _ {s}\\right). \\tag {14}", + "image_path": "706587538ce3890d3c9bd6bcb17b9ad7b2acd8ba2fa79bb7a4eaeb88e94f4627.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13055" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 202, + 194 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 202, + 194 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 202, + 194 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 202, + 194 + ], + "type": "image", + "image_path": "e375e3af126ce2b966761f4fba7e797de8b226be5718ef2bb3a72f1df267f0a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. The overall frameworks of Attn-AST approaches-CNN-based, Transformer-based, and Diffusion-based methods with semantic continuous-sparse attention (SCSA). SCA depicts semantic continuous attention. SSA denotes semantic sparse attention. S-AdaIN is semantic adaptive instance normalization. UA expresses universal attention in the Attn-AST methods. The dashed line in (b) shows encoded content features " + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": " are only used in the first use of the feature transformation module, and the output features " + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": " are used as new content features in subsequent transformation. The dashed line in (c) indicates S-AdaIN is used only when " + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": " is the maximum time step " + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 70, + 370, + 194 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 370, + 194 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 370, + 194 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 370, + 194 + ], + "type": "image", + "image_path": "4d1a312e4e3a062108d6a8aa70b297b6d08fd5a0cb7faf5e88a1e23e60269064.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 68, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 378, + 68, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 378, + 68, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 378, + 68, + 547, + 194 + ], + "type": "image", + "image_path": "9d5b85a6a727b3a716728cb0fe396c7792c9c2103ab3d4962fee85755ccc32b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "text", + "content": "Then, they use the feature transformation module " + }, + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "text", + "content": " with universal attention (UA) to obtain the stylized features " + }, + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 55, + 270, + 295, + 295 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 301, + 295, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 301, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 133, + 301, + 295, + 316 + ], + "type": "interline_equation", + "content": "F _ {c s} = T _ {U A} \\left(F _ {c}, F _ {s}\\right). \\tag {15}", + "image_path": "29a02604051e3b4003316e72693c306d4227a391c69b7216d5bd585ef18e31b4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "type": "text", + "content": "Eventually, they could acquire the stylized image with the decoder " + }, + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 322, + 296, + 346 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 144, + 353, + 295, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 353, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 144, + 353, + 295, + 367 + ], + "type": "interline_equation", + "content": "I _ {c s} = D \\left(F _ {c s}\\right), \\tag {16}", + "image_path": "a701059d715c5438091227d9bd9a475e92ed19ed2dc3441ba9413282fad54544.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": " possesses the content of " + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": " and the style of " + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": ", but the style of its corresponding semantic region is inconsistent with that of the style image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 410, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 516 + ], + "type": "text", + "content": "Discussion. Despite the three different Attn-AST frameworks (CNN, Transformer, and Diffusion), their pipelines remain largely consistent with the above, differing only in model structures. The CNN-based method uses a simple encoder-decoder setup, while the Transformer-based approach employs two distinct encoders to process content and style separately. In contrast, the diffusion-based method also utilizes a single encoder-decoder structure but incorporates a time-step iteration mechanism for further refinement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": "Attn-AST Approaches with Our SCSA. SCSA is a plug-and-play method that can be incorporated into any Attn-AST method. Thus, given a content image " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": " with its semantic map " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "I_{csem}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": ", a style image " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": " with its semantic map " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "I_{ssem}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": ", and a specified Attn-AST model " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": ", our objective is that " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "M_{SCSA}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": ", the approach obtained after incorporating SCSA into " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": ", can generate a stylized image " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": " using the quadruple " + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\{I_{c}, I_{csem}, I_{s}, I_{ssem}\\}" + }, + { + "bbox": [ + 55, + 517, + 295, + 624 + ], + "type": "text", + "content": ", exhibiting a consistent style with the style image within the same semantic regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 624, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 648 + ], + "type": "text", + "content": "At first, we obtain the four encoded features of the quadruple through the encoder " + }, + { + "bbox": [ + 55, + 624, + 296, + 648 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 624, + 296, + 648 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 654, + 294, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 654, + 294, + 673 + ], + "spans": [ + { + "bbox": [ + 125, + 654, + 294, + 673 + ], + "type": "interline_equation", + "content": "F _ {c} = E \\left(I _ {c}\\right), F _ {s} = E \\left(I _ {s}\\right), \\tag {17}", + "image_path": "632ac920dc23a84319fa6fb469beea59ecae37145cbf429fc33ce9d04320f370.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 97, + 671, + 254, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 671, + 254, + 683 + ], + "spans": [ + { + "bbox": [ + 97, + 671, + 254, + 683 + ], + "type": "interline_equation", + "content": "F _ {c s e m} = E (I _ {c s e m}), F _ {s e m} = E (I _ {s e m}).", + "image_path": "eb07c41e09a2d64fec77e3a87cc4104d30c2e4a3ba411233fd44fb6751d14e11.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Then, we replace UA in the feature transformation module " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " of Eq. 15 with SCSA, which could perform semantic" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 270, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 295 + ], + "type": "text", + "content": "feature conversion by enabling each query point to consider certain key points in the corresponding semantic region:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 300, + 553, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 300, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 356, + 300, + 553, + 314 + ], + "type": "interline_equation", + "content": "F _ {c s} = T _ {S C S A} \\left(F _ {c}, F _ {s}, F _ {c s e m}, F _ {s s e m}\\right), \\tag {18}", + "image_path": "64e284f2e7fa4e7577e6d04ab6857a9a7df3c2d2f09d4eeedf6741d025d93a29.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 319, + 551, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 551, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 551, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 319, + 551, + 331 + ], + "type": "inline_equation", + "content": "F_{cs}" + }, + { + "bbox": [ + 313, + 319, + 551, + 331 + ], + "type": "text", + "content": " are stylized features that meet semantic criteria." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 332, + 543, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 332, + 543, + 344 + ], + "spans": [ + { + "bbox": [ + 325, + 332, + 543, + 344 + ], + "type": "text", + "content": "Lastly, we get the stylized image with the decoder " + }, + { + "bbox": [ + 325, + 332, + 543, + 344 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 325, + 332, + 543, + 344 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 403, + 350, + 553, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 350, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 403, + 350, + 553, + 363 + ], + "type": "interline_equation", + "content": "I _ {c s} = D \\left(F _ {c s}\\right), \\tag {19}", + "image_path": "f7712a3e76a5bf06c4b0a98f5c1282b965e0c1f4e2d29fc5f5082b784ebd5bc4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "text", + "content": "where the style of the semantic region of the stylized image " + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "text", + "content": " is consistent with that of the style image " + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "text", + "content": ", while the content of " + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "inline_equation", + "content": "I_{cs}" + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "text", + "content": " is the same as that of the content image " + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 369, + 554, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": "The frameworks of the CNN-based, Transformer-based, and Diffusion-based Attn-AST methods with our SCSA are shown in Fig. 4. They differ only in the layers where UA is replaced with SCSA and in the encoder and decoder structures, while the other operations remain largely consistent." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 473, + 395, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 395, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 395, + 487 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 493, + 432, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 432, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 432, + 506 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 510, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 606 + ], + "type": "text", + "content": "We select three proxy Attn-AST approaches (CNN-based SANet [32], Transformer-based StyTr² [9], and Diffusion-based StyleID [8]), and five state-of-the-art (SOTA) semantic stylization methods (STROTSS [21], MAST [17], TR [40], DIA [25], and GLStyleNet [38]) as baselines. We obtain images from prior research, datasets, and the Internet, and construct their semantic maps to generate quadruple data. More details are in the supplementary materials." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 613, + 450, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 450, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 450, + 625 + ], + "type": "text", + "content": "4.2. Qualitative Comparison" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 630, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 653 + ], + "type": "text", + "content": "We conduct extensive comparison experiments to evaluate the effectiveness and generalization of our SCSA." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "As shown in Fig. 5, Attn-AST approaches combined with our SCSA can generate high-quality and vivid stylized images that better meet semantic needs compared to traditional Attn-AST methods, showcasing rich textures and vibrant colors in corresponding regions, e.g., the background" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13056" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 69, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 547, + 224 + ], + "type": "image", + "image_path": "7992688d78e741a42f957464972fac140e56acb80298280e7d258b12c9dd4495.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 232, + 493, + 244 + ], + "lines": [ + { + "bbox": [ + 116, + 232, + 493, + 244 + ], + "spans": [ + { + "bbox": [ + 116, + 232, + 493, + 244 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons among Attn-AST approaches, those with SCSA, and SOTA methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 254, + 538, + 300 + ], + "blocks": [ + { + "bbox": [ + 72, + 254, + 538, + 300 + ], + "lines": [ + { + "bbox": [ + 72, + 254, + 538, + 300 + ], + "spans": [ + { + "bbox": [ + 72, + 254, + 538, + 300 + ], + "type": "table", + "html": "
SANetSANet + SCSAStyTr2StyTr2 + SCSAStyleIDStyleID + SCSASTROTSSMASTTRDIAGLStyleNet
SSL ↓1.65830.87621.98261.22281.75381.24471.09811.73201.26311.93981.0305
FID ↓14.338513.078812.527312.396312.594412.449712.840016.516313.584620.794214.4880
CFSD ↓0.11030.08740.07520.07050.09160.11780.10080.07460.11490.11470.3991
Pref. ↑0.16850.83150.15760.84240.21920.78080.18670.11000.04500.01670.1183
", + "image_path": "4ed1d6d2a8aa5156c8bee855b4c1ec214c358aa508fbe414f01dc636be6e343f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of Attn-AST approaches, those with SCSA, and several SOTA semantic style transfer methods. The best results for Attn-AST methods and Attn-AST methods with SCSA are in bold, while the best results among all methods are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 343, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 403 + ], + "type": "text", + "content": "in the 1st row, along with the sky and ground in the 2nd row. Likewise, compared to SOTA semantic style transfer methods, the stylized images generated by SCSA demonstrate superior semantic accuracy and content preservation, e.g. the river in the 3rd row and the horses in the 4th row." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 452 + ], + "type": "text", + "content": "In summary, SCSA is a plug-and-play solution that can be integrated into Attn-AST methods, addressing their semantic issues and offering a level of capability in semantic style transfer that surpasses that of existing SOTA methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 457, + 198, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 198, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 198, + 470 + ], + "type": "text", + "content": "4.3. Quantitative Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 285, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 285, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 285, + 486 + ], + "type": "text", + "content": "We also conduct quantitative analyses for full validation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 665 + ], + "type": "text", + "content": "Semantic Style Loss (SSL). Traditional style loss [16] is unsuitable for semantic style transfer. There are two main reasons: First, it focuses on the overall region, making it difficult to capture the semantic region style; second, the inconsistency in the sizes of the semantic regions between the style and stylized images affects the calculation of global style loss. Hence, we propose a semantic style loss to calculate each region style loss similar to Eq. 8 using the style loss of [16]. From the 1st row of Tab. 1, it is clear that the Attn-AST methods utilizing SCSA not only achieve the lowest SSL compared to traditional Attn-AST methods, but when SANet incorporates SCSA, it also reaches the best performance among all semantic style transfer methods. This indicates that SCSA can effectively facilitate semantic style transfer with its effectiveness and generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Frechet Inception Distance (FID). Similar to [8], we use FID [13] to evaluate the overall style fidelity between the stylized and style images. As indicated by the 2nd row of Tab. 1, the Attn-AST methods integrated with SCSA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 343, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 416 + ], + "type": "text", + "content": "achieve the lowest FID value as opposed to typical Attn-AST methods. Particularly, " + }, + { + "bbox": [ + 313, + 343, + 555, + 416 + ], + "type": "inline_equation", + "content": "\\mathrm{StyTr}^2" + }, + { + "bbox": [ + 313, + 343, + 555, + 416 + ], + "type": "text", + "content": " with SCSA secures the optimal FID value among all the semantic style transfer methods. The above indicates that SCSA not only conducts the semantic style transfer but also markedly improves the overall style quality of the stylized images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 417, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 556, + 561 + ], + "type": "text", + "content": "Content Feature Structural Distance (CFSD). As discussed in [8], content loss [16] and LPIPS [47] are affected by the stylization. Thus, we use the content feature structural distance [8] to assess content retention. As shown in the " + }, + { + "bbox": [ + 313, + 417, + 556, + 561 + ], + "type": "inline_equation", + "content": "3rd" + }, + { + "bbox": [ + 313, + 417, + 556, + 561 + ], + "type": "text", + "content": " row of Tab. 1, the SCSA achieves the lowest CFSD value on CNN-based and Transformer-based methods. Although the CFSD value is slightly higher on the Diffusion-based approach with SCSA, both quantitative and qualitative results indicate that SCSA greatly improves semantic stylization, making this trade-off worthwhile. Besides, StyTr² with SCSA has the lowest CFSD value in all SOTA methods, showcasing its content preservation advantage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 562, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 556, + 658 + ], + "type": "text", + "content": "User Study. Our user study is divided into two parts: the first compares traditional Attn-AST methods with SCSA-enhanced ones, and the second evaluates Attn-AST methods with SCSA against existing semantic style transfer methods. As shown in the 4th row of Tab. 1, the stylized images of the Attn-AST methods with SCSA are more favored by the public, and their popularity surpasses that of the SOTA methods. These demonstrate the superiority of SCSA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 670, + 408, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 408, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 408, + 684 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We conduct ablation experiments to verify the crucial roles of SCA and SSA and to affirm the potency of S-AdaIN." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13057" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 549, + 175 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 549, + 175 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 549, + 175 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 549, + 175 + ], + "type": "image", + "image_path": "89f979d59d5b8f7f52558385078c333c94e608349fad7cadaf7fc7acb805a650.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 180, + 512, + 192 + ], + "lines": [ + { + "bbox": [ + 96, + 180, + 512, + 192 + ], + "spans": [ + { + "bbox": [ + 96, + 180, + 512, + 192 + ], + "type": "text", + "content": "Figure 6. Qualitative and quantitative ablation results on SCA, SSA, and S-AdaIN. Zoom in for better comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 204, + 290, + 283 + ], + "blocks": [ + { + "bbox": [ + 63, + 204, + 290, + 283 + ], + "lines": [ + { + "bbox": [ + 63, + 204, + 290, + 283 + ], + "spans": [ + { + "bbox": [ + 63, + 204, + 290, + 283 + ], + "type": "image", + "image_path": "add8873588041f0364df3f3d211a37776aeb2194afdaf7a159209f12cc7a1f15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 306, + 287, + 343 + ], + "blocks": [ + { + "bbox": [ + 66, + 289, + 284, + 300 + ], + "lines": [ + { + "bbox": [ + 66, + 289, + 284, + 300 + ], + "spans": [ + { + "bbox": [ + 66, + 289, + 284, + 300 + ], + "type": "text", + "content": "Figure 7. The stylization of different content representation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 306, + 287, + 343 + ], + "lines": [ + { + "bbox": [ + 63, + 306, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 63, + 306, + 287, + 343 + ], + "type": "table", + "html": "
UASCASSASCSA
Runtime (milliseconds)1.2311.9172.0153.944
Memory Usage (bytes)4708479649726220
", + "image_path": "5edb33b049b1dc3d057d390562a26c3a8e38eefcfe032dd9c95bed0411116fe2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 350, + 276, + 361 + ], + "lines": [ + { + "bbox": [ + 74, + 350, + 276, + 361 + ], + "spans": [ + { + "bbox": [ + 74, + 350, + 276, + 361 + ], + "type": "text", + "content": "Table 2. Computation costs for feature fusion modules." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 373, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 505 + ], + "type": "text", + "content": "Effect of SCA and SSA. We use SCA or SSA individually as the feature transformation module to assess their capabilities. As shown in Fig. 6, SCA alone facilitates overall color transfer in corresponding semantic regions but results in texture loss and reduced stylization, as reflected by the higher SSL and FID values. In contrast, SSA can effectively transfer specific textures but introduces color discrepancies between the semantic regions of the stylized and style images. This indicates that while SCA captures the overall style characteristics, SSA is crucial for transferring specific textures, highlighting the complementary roles of both." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 625 + ], + "type": "text", + "content": "Effect of S-AdaIN. To validate the effectiveness of S-AdaIN, we conducted ablation experiments by removing S-AdaIN. As shown in Fig. 6, without S-AdaIN, the stylized images exhibit less accurate global style transfer in semantic regions and reduced texture details, particularly evident for the StyleID method. Although S-AdaIN was not employed, CFSD values of StyTr² and StyleID with SCSA are lower, which comes at the expense of stylization quality. Therefore, after considering the pros and cons, S-AdaIN is essential and effective for semantic style transfer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 288, + 664 + ], + "type": "text", + "content": "We perform more discussions about our task and method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Arbitrary Semantic Style Transfer. Our plug-and-play SCSA is built on the Attn-AST approaches, unlike methods that iteratively optimize the specific content and style images or only apply a predefined style to any content im" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 207, + 553, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 278 + ], + "type": "text", + "content": "age, so that it can stylize between any content and style images. Besides, it is important to note that SCSA leverages the user-provided semantic maps to guide the transfer process. Therefore, as long as the user's semantic maps mark the regions where the style should be applied, we can perform the relevant region stylization, as shown in Fig. 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 279, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 423 + ], + "type": "text", + "content": "Computation Costs. Since the costs vary with the feature size, we use the features of the " + }, + { + "bbox": [ + 313, + 279, + 555, + 423 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 279, + 555, + 423 + ], + "type": "text", + "content": " images encoded by ReLU4_1 layer of VGG-19 to discuss costs. The results are in Tab. 2. In the 1st row, SCA and SSA have slightly longer runtimes than UA due to the guidance operations. SCSA takes barely longer than the sum of SCA and SSA, as it merges both outputs. The 2nd row shows SCA and SSA use a bit more memory than UA, while SCSA uses less than SCA plus SSA. This is because some operations reduce memory usage by reusing existing variables. Despite the cost increases, this is acceptable and does not impact practical use. Cost refinements are left for future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 388, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 388, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 388, + 446 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 454, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 598 + ], + "type": "text", + "content": "In this paper, we highlight the challenge faced by existing arbitrary style transfer methods, with a particular emphasis on attention-based approaches, and provide an in-depth analysis of this issue, which is the failure to account for the relationship between local regions and semantic regions. To address this issue and generate high-quality stylized images that meet semantic needs, we propose a plug-and-play semantic continuous-sparse attention, dubbed SCSA. It achieves precise semantic style transfer by allowing each query point to consider certain key points in the corresponding semantic region. Extensive experimental results demonstrate that our SCSA is effective and generalized." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "This work was partly supported by the National Natural Science Foundation of China No. 62306277, the Fundamental Research Funds for the Zhejiang Provincial Universities Grant No. K20240090, the Research Fund for International Scientists of National Natural Science Foundation of China No. 72350710798, and the Natural Science Foundation of Zhejiang Province No. LDT23F02023F02. This" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13058" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "research was supported in part through funding from the following sources: the National Natural Science Foundation of China under Award No. 62306277, and the Fundamental Research Funds for Provincial Universities in Zhejiang Province through Grant No. K20240090." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 143, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 217 + ], + "type": "text", + "content": "[1] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 220, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 252 + ], + "type": "text", + "content": "[2] Dzmitry Bahdanau. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 254, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 296, + 286 + ], + "type": "text", + "content": "[3] Alex J Champandard. Semantic style transfer and turning two-bit doodles into fine artworks. arXiv preprint arXiv:1603.01768, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 288, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 295, + 319 + ], + "type": "text", + "content": "[4] Dar-Yen Chen. Artfusion: Arbitrary style transfer using dual conditional latent diffusion models. arXiv preprint arXiv:2306.09330, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 322, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 322, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 296, + 376 + ], + "type": "text", + "content": "[5] Dar-Yen Chen, Hamish Tennent, and Ching-Wen Hsu. Art-adapter: Text-to-image style transfer using multi-level style encoder and explicit adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8619-8628, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 378, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 378, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 61, + 378, + 296, + 432 + ], + "type": "text", + "content": "[6] Haibo Chen, Zhizhong Wang, Huiming Zhang, Zhiwen Zuo, Ailin Li, Wei Xing, Dongming Lu, et al. Artistic style transfer with internal-external learning and contrastive learning. Advances in Neural Information Processing Systems, 34:26561-26573, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "type": "text", + "content": "[7] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 296, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 522 + ], + "type": "text", + "content": "[8] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "text", + "content": "[9] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11326-11336, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 580, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 633 + ], + "type": "text", + "content": "[10] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3730-3738. IEEE. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 679 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Mark Hamazaspyan and Shant Navasardyan. Diffusion-enhanced patchmatch: A framework for arbitrary style transfer with diffusion models. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 797-805, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[14] Kibeom Hong, Seogkyu Jeon, Junsoo Lee, Namhyuk Ahn, Kunhee Kim, Pilhyeon Lee, Daesik Kim, Youngjung Uh, and Hyeran Byun. Aespa-net: Aesthetic pattern-aware style transfer networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22758-22767, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[15] Siyu Huang, Jie An, Donglai Wei, Jiebo Luo, and Hanspeter Pfister. Quantart: Quantizing image style transfer towards high visual fidelity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5947-5956, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 2, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "type": "text", + "content": "[17] Jing Huo, Shiyin Jin, Wenbin Li, Jing Wu, Yu-Kun Lai, Yinghuan Shi, and Yang Gao. Manifold alignment for semantically aligned style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14861-14869, 2021. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 555, + 431 + ], + "type": "text", + "content": "[18] Yongcheng Jing, Xiao Liu, Yukang Ding, Xinchao Wang, Errui Ding, Mingli Song, and Shilei Wen. Dynamic instance normalization for arbitrary style transfer. In Proceedings of the AAAI conference on artificial intelligence, pages 4369-4376, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[19] Nikolai Kalischek, Jan D Wegner, and Konrad Schindler. In the light of feature distributions: moment matching for neural style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9382-9391, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[20] Da-Hyun Kim, Dong-Hee Shin, and Tae-Eui Kam. Bridging the bci illiteracy gap: a subject-to-subject semantic style transfer for eeg-based motor imagery classification. Frontiers in Human Neuroscience, 17:1194751, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "type": "text", + "content": "[21] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[22] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In International conference on machine learning, pages 957-966. PMLR, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast arbitrary style transfer, 2018. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13059" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[25] Jing Liao, Yuan Yao, Lu Yuan, Gang Hua, and Sing Bing Kang. Visual attribute transfer through deep image analogy. arXiv preprint arXiv:1705.01088, 2017. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[26] Yi-Sheng Liao and Chun-Rong Huang. Semantic context-aware image style transfer. IEEE Transactions on Image Processing, 31:1911-1923, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "text", + "content": "[27] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6649-6658, 2021. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 229 + ], + "type": "text", + "content": "[28] Stuart Lloyd. Least squares quantization in lcm. IEEE transactions on information theory, 28(2):129-137, 1982. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 294, + 274 + ], + "type": "text", + "content": "[29] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5951-5960. IEEE, .3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[30] Ming Lu, Hao Zhao, Anbang Yao, Feng Xu, Yurong Chen, and Li Zhang. Decoder network over lightweight reconstructed feature for fast semantic style transfer. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2488-2496. IEEE, .3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "type": "text", + "content": "[31] Yingnan Ma, Chenqiu Zhao, Xudong Li, and Anup Basu. Rast: Restorable arbitrary style transfer via multi-restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 331-340, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 431 + ], + "type": "text", + "content": "[32] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5880-5888, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 565 + ], + "type": "text", + "content": "[35] Onkar Susladkar, Gayatri Deshmukh, Sparsh Mittal, and Parth Shastri. D2styler: Advancing arbitrary style transfer with discrete diffusion methods. arXiv preprint arXiv:2408.03558, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "text", + "content": "[36] Gemma Canet Tarres, Dan Ruta, Tu Bui, and John Collosomasse. Parasol: Parametric style control for diffusion image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2432-2442, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 645 + ], + "type": "text", + "content": "[37] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[38] Zhizhong Wang, Lei Zhao, Sihuan Lin, Qihang Mo, Huiming Zhang, Wei Xing, and Dongming Lu. Gliystylenet: exquisite style transfer combining global and local pyramid features. IET Computer Vision, 14(8):575-586, 2020. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[39] Zhizhong Wang, Zhanjie Zhang, Lei Zhao, Zhiwen Zuo, Ailin Li, Wei Xing, and Dongming Lu. Aesust: towards" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 643 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "aesthetic-enhanced universal style transfer. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1095-1106, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[40] Zhizhong Wang, Lei Zhao, Haibo Chen, Ailin Li, Zhiwen Zuo, Wei Xing, and Dongming Lu. Texture reformer: Towards fast and universal interactive texture transfer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2624-2632, 2022. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[41] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7677-7689, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "text", + "content": "[42] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14618-14627, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[43] Zijie Wu, Zhen Zhu, Junping Du, and Xiang Bai. Ccpl: Contrastive coherence preserving loss for versatile style transfer. In European Conference on Computer Vision, pages 189-206. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "type": "text", + "content": "[44] Zhengtao Xiang, Xing Wan, Libo Xu, Xin Yu, and Yuhan Mao. A training-free latent diffusion style transfer method. Information, 15(10):588, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[45] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22873-22882, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 441 + ], + "type": "text", + "content": "[46] Chiyu Zhang, Xiaogang Xu, Lei Wang, Zaiyan Dai, and Jun Yang. S2wat: Image style transfer via hierarchical vision transformer using strips window attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7024-7032, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 497 + ], + "type": "text", + "content": "[47] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 499, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 553 + ], + "type": "text", + "content": "[48] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8035-8045, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 598 + ], + "type": "text", + "content": "[49] Yuxin Zhang, Fan Tang, Weiming Dong, Haibin Huang, Chongyang Ma, Tong-Yee Lee, and Changsheng Xu. A unified arbitrary style transfer framework via adaptive contrastive learning. ACM Trans. Graph., 42(5), 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "text", + "content": "[50] Hui-Huang Zhao, Paul L Rosin, Yu-Kun Lai, and Yao-Nan Wang. Automatic semantic style transfer using deep convolutional neural networks and soft masks. The Visual Computer, 36(7):1307-1324, 2020. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13060" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_content_list.json b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1afb4da70031737fd1ee131f3947cb26407b2321 --- /dev/null +++ b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_content_list.json @@ -0,0 +1,1730 @@ +[ + { + "type": "text", + "text": "SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures", + "text_level": 1, + "bbox": [ + 151, + 128, + 844, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hui Liu $^{1,2,3}$ , Chen Jia $^{1,2,3,*}$ , Fan Shi $^{1,2,3}$ , Xu Cheng $^{1,2,3}$ , Shengyong Chen $^{1,2,3}$ $^{1}$ School of Computer Science and Engineering, Tianjin University of Technology \n $^{2}$ Engineering Research Center of Learning-Based Intelligent System (Ministry of Education) \n $^{3}$ Key Laboratory of Computer Vision and System (Ministry of Education)", + "bbox": [ + 133, + 202, + 862, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuhui1109@stud.tjut.edu.cn, {jiachen, shifan}@email.tjut.edu.cn, {xu.cheng, sy}@ieee.org", + "bbox": [ + 99, + 276, + 883, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pixel-level segmentation of structural cracks across various scenarios remains a considerable challenge. Current methods encounter challenges in effectively modeling crack morphology and texture, facing challenges in balancing segmentation quality with low computational resource usage. To overcome these limitations, we propose a lightweight Structure-Aware Vision Mamba Network (SCSegamba), capable of generating high-quality pixel-level segmentation maps by leveraging both the morphological information and texture cues of crack pixels with minimal computational cost. Specifically, we developed a Structure-Aware Visual State Space module (SAVSS), which incorporates a lightweight Gated Bottleneck Convolution (GBC) and a Structure-Aware Scanning Strategy (SASS). The key insight of GBC lies in its effectiveness in modeling the morphological information of cracks, while the SASS enhances the perception of crack topology and texture by strengthening the continuity of semantic information between crack pixels. Experiments on crack benchmark datasets demonstrate that our method outperforms other state-of-the-art (SOTA) methods, achieving the highest performance with only 2.8M parameters. On the multi-scenario dataset, our method reached 0.8390 in F1 score and 0.8479 in mIoU. The code is available at https://github.com/Karl1109/SCSegamba.", + "bbox": [ + 89, + 358, + 483, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Structures like bitumen pavement, concrete, and metal frequently develop cracks under shear stress, making regular health monitoring essential to avoid production issues [3, 4, 19, 23, 30]. Due to differences in material properties and environmental conditions, various materials exhibit significant variations in crack morphology and visual appearance [24]. As a result, achieving pixel-level crack seg-", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f7757cd470313dcf0ac3299ed67e5005cdc17c6eaa42a92a48434e38123b9a1.jpg", + "image_caption": [ + "(a) Comparison with SOTA methods.", + "(b) Different numbers of SAVSS layers." + ], + "image_footnote": [], + "bbox": [ + 514, + 325, + 908, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/10743ce2eebe51870d9739b4e9086550ce0b2e781a09400b0045c5896855bb95.jpg", + "image_caption": [ + "(c) Segmentation results in complex interference conditions." + ], + "image_footnote": [], + "bbox": [ + 537, + 444, + 704, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/caffed150fd9ea989e2dc81b4fb2817ecf722d1dff63a0fca4889904d9dd34ab.jpg", + "image_caption": [ + "Figure 1. Performance of SCSegamba on multi-scenario TUT [33] dataset. (a) Comparison with SOTA methods. (b) Impact of different SAVSS layer numbers on performance, with normalized metrics; FLOPs (G), Params (M), and Size (MB) decrease towards the edges. (c) Visual results under complex interference." + ], + "image_footnote": [], + "bbox": [ + 715, + 444, + 877, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mentation across diverse scenarios remains a complex challenge. Recently, Convolutional Neural Networks (CNNs), such as ECSNet [59] and SFIAN [5], have shown effective crack feature extraction capabilities in segmentation tasks due to their strong local inductive properties. However, their limited receptive field constrains their ability to model broad-scope irregular dependencies across the entire image, resulting in discontinuous segmentation and weak background noise suppression. Although dilated convolution [2] expand the receptive field, their inherent inductive bias still prevents them from fully addressing this issue [58], especially in complex crack patterns with heavy background interference.", + "bbox": [ + 511, + 580, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of Vision Transformer (ViT) [10, 45, 47] has demonstrated Transformer's effectiveness in capturing irregular pixel dependencies, which is crucial for recognizing complex crack textures, as seen in networks like DTRCNet [48], MFAFNet [9], and Crackmer [46]. However, the quadratic complexity of attention calculations with sequence length leads to high memory use and training challenges for high-resolution images, limiting deployment on", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29406", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resource-constrained edge devices and practical applications. As shown in Figure 1(a), Transformer-based methods like CTCrackseg [44] and DTrCNet [48] perform well, but their large parameter counts and high computational demands limit their deployability on resource-constrained devices. Although variants like Sparse Transformer [6] and Linear Transformer [21] reduce computational requirements by sparsifying or linearizing attention, they sacrifice the ability to model irregular dependencies and pixel textures, hindering pixel-level detection tasks. Consequently, Transformer-based methods struggle to balance segmentation performance with computational efficiency.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Selective State Space Models (SSMs) have attracted considerable interest due to Mamba's showing strong performance in sequence modeling while maintaining low computational demands [14, 15]. Vision Mamba (ViM) [62] and VMamba [36] have extended Mamba to the visual domain. The irregular extension of crack regions with numerous branches in low-contrast images, often affected by irrelevant areas and shadows, challenges existing Mamba VSS (Visual State Space Model) block structures and scanning strategies in capturing crack morphology and texture effectively. Most Mamba-based methods [37, 51, 52, 55] process feature maps through linear layers, limiting selective enhancement or suppression of crack features against irrelevant disturbances, thus reducing detailed morphological extraction. Additionally, common parallel or unidirectional diagonal scans [16] struggle to maintain semantic continuity when handling irregular, multi-directional pixel topologies, weakening their ability to manage complex textures and suppress noise. Consequently, current Mamba-based SSM frequently produce false or missed detections in multi-scenario crack images. Moreover, although these methods have the advantage of fewer parameters, there remains potential to further reduce their computational demands and enhance their deployability on edge devices. As shown in Figure 1(b), CSMamba [37], MambaIR [16], and PlainMamba [55] showed unsatisfactory performance on crack images, with room for improvement in parameter count and computational load.", + "bbox": [ + 91, + 276, + 483, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenge of balancing high segmentation quality with low computational demands, we propose the SCSegamba network that produces high-quality pixel-level crack segmentation maps with low computational resources. To improve the Mamba network's perception of irregular crack textures, we design a Structure-Aware Visual State Space block (SAVSS), employing a Structure-Aware Scanning Strategy (SASS) to enhance semantic continuity and strengthen crack morphology perception. For capturing crack shape cues while maintaining low parameter and computational costs, we designed a lightweight Gated Bottleneck Convolution (GBC) that dynamically adjusts weights for complex backgrounds and varying mor", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "phologies. Additionally, the Multi-scale Feature Segmentation head (MFS) integrates the GBC and Multi-layer Perceptron (MLP) to achieve high-quality segmentation maps with low computational requirements. As shown in Figure 1(c), optimal segmentation performance was achieved with four SAVSS layers, producing clear segmentation maps that effectively mask complex interference while maintaining model lightweightness.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In essence, the primary contributions of our work are outlined as follows:", + "bbox": [ + 511, + 212, + 903, + 239 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel lightweight vision Mamba network, SCSegamba, for crack segmentation. This model effectively captures morphological and irregular texture cues of crack pixels, using low computational resources to generate high-quality segmentation maps.", + "- We design the SAVSS with a lightweight GBC Convolution and a SASS scanning strategy to enhance the handling and perception of irregular texture cues in crack images. Additionally, a simple yet effective MFS is developed to generate segmentation maps with relatively low computational resources.", + "- We evaluate SCSegamba on benchmark datasets across diverse scenarios, with results demonstrating that our method outperforms other SOTA methods while maintaining the low parameter count." + ], + "bbox": [ + 513, + 244, + 903, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 479, + 661, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Crack Segmentation Network", + "text_level": 1, + "bbox": [ + 513, + 501, + 777, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early crack detection methods often relied on traditional feature extraction techniques, such as wavelet transform [61], percolation models [54], and the k-means algorithm [25]. Although straightforward to implement, these methods face challenges in suppressing background interference and achieving high segmentation accuracy. With advancements in deep learning, researchers have developed CNN-based crack segmentation networks that achieve SOTA performance [5, 7, 18, 26, 40, 49]. For instance, DeepCrack [35] enables end-to-end pixel-level segmentation, while FPHBN [56] demonstrates strong generalization capabilities. BARNet [17] detects crack boundaries by integrating image gradients with features, and ADDUNet [1] captures both fine and coarse crack characteristics across varied conditions. Although CNN-based methods show significant promise, the local operations and limited receptive fields of CNNs limit their ability to fully capture crack texture cues and effectively suppress background noise.", + "bbox": [ + 511, + 523, + 906, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers [45], with their self-attention mechanism, are well-suited for visual tasks that require long-range dependency modeling, making them increasingly popular in crack segmentation networks [9, 31, 32, 41, 53]. For example, VCVNet [39], based on Vision Transformer (ViT), is designed for bridge crack segmentation, addressing fine-grained segmentation challenges. SwT-CNN [27]", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1ff71d1dde6bdf425aceab4c1c2341b32a552eaa35d641fa32b9f38db4a7aa5f.jpg", + "image_caption": [ + "(a) Network Architecture of SCSegamba" + ], + "image_footnote": [], + "bbox": [ + 99, + 95, + 575, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/428c266e17810e2d4531dab486d923b5c6f34d4f7dcfc80d742206562c9b6a60.jpg", + "image_caption": [ + "(b) Architecture of SAVSS block", + "Figure 2. Overview of our proposed method. (a) illustrates the overall architecture of SCSegamba and the processing flow for crack images. (b) displays the structure of the SAVSS block. The input crack image undergoes comprehensive morphological and texture feature extraction through SAVSS, while MFS produces a high-quality pixel-level segmentation map." + ], + "image_footnote": [], + "bbox": [ + 581, + 95, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "combines Swin Transformer and CNN for automatic feature extraction, while TBUNet [60], a Transformer-based knowledge distillation model, achieves high-performance crack segmentation with a hybrid loss function. Although Transformer-based methods are highly effective at capturing crack texture cues and suppressing background noise, their self-attention mechanism introduces computational complexity that grows quadratically with sequence length. This results in a high parameter count and significant computational demands, which limit their deployment on resource-constrained edge devices.", + "bbox": [ + 89, + 347, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59e9419b25be7d1dd892e1d73ff2cc851473ebb1a5f259ecf6e68c1a751f2744.jpg", + "image_caption": [ + "Figure 3. Architecture of GBC. It employs bottleneck convolution to efficiently reduce the parameters and computational load, while the gating mechanism enhances the model's adaptability in processing diverse crack patterns and complex backgrounds. GN represents group normalization." + ], + "image_footnote": [], + "bbox": [ + 114, + 525, + 460, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Selective State Space Model", + "text_level": 1, + "bbox": [ + 89, + 787, + 339, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The introduction of the Selective State Space Models (S6) in the Mamba model [14] has highlighted the potential of SSM [12, 13]. Unlike the linear time-invariant S4 model, S6 efficiently captures complex long-distance dependencies while preserving computational efficiency, achieving strong performance in NLP, audio, and genomics. Consequently,", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "researchers have adapted Mamba to the visual domain, creating various VSS blocks. ViM [62] achieves comparable modeling to ViT [10] without attention mechanisms, using fewer computational resources, while VMamba [36] prioritizes efficient computation and high performance. Plain-Mamba [55] employs a fixed-width layer stacking approach, excelling in tasks such as instance segmentation and object detection. However, the VSS block and scanning strategy require specific optimizations for each visual task, as tasks differ in their reliance on long- and short-distance information, necessitating customized VSS block designs to ensure optimal performance.", + "bbox": [ + 511, + 347, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, no high-performing Mamba-based model exists for crack segmentation. Thus, designing an optimized VSS structure specifically for crack segmentation is essential to improve performance and efficiency. Given the intricate details and irregular textures of cracks, the VSS block requires strong shape extraction and directional awareness to effectively capture crack texture cues. Additionally, it should facilitate efficient crack segmentation while minimizing computational resource requirements.", + "bbox": [ + 511, + 529, + 908, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 672, + 648, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 696, + 642, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The complete architecture of our proposed SCSegamba is depicted in Figure 2. It includes two main components: the SAVSS for extracting crack shape and texture cues, and the MFS for efficient feature processing. To capture key crack region cues, we integrate the GBC at the initial stage of SAVSS and the final stage of MFS.", + "bbox": [ + 511, + 719, + 905, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a single RGB image $E \\in \\mathbb{R}^{3 \\times H \\times W}$ , spatial information is divided into $n$ patches, forming a sequence $\\{B_1, B_2, \\ldots, B_n\\}$ . This sequence is processed through the SAVSS block, embedding key crack pixel cues into multiscale feature maps $\\{F_1, F_2, F_3, F_4\\}$ . Finally, in the MFS, all information is consolidated into a single tensor, produc", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29408", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing a refined segmentation output $W\\in \\mathbb{R}^{1\\times H\\times W}$", + "bbox": [ + 89, + 89, + 424, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Lightweight Gated Bottleneck Convolution", + "text_level": 1, + "bbox": [ + 89, + 114, + 455, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The gating mechanism enables dynamic features for each spatial position and channel, enhancing the model's ability to capture details [8, 57]. To further reduce parameter count and computational cost, we embedded a bottleneck convolution (BottConv) with low-rank approximation [28], mapping matrices from high to low dimensional spaces and significantly lowering computational complexity.", + "bbox": [ + 89, + 137, + 483, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the convolution layer, assuming the spatial size of the filter is $p$ , the number of input channels is $d$ and the input is $s$ , the convolution response can be represented as:", + "bbox": [ + 89, + 243, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = Q s + c \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 294, + 480, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ is a matrix of size $f \\times (p^2 \\times d)$ , $f$ is the number of output channels, and $c$ is the original bias term. Assuming $z$ lies in a low-rank subspace of rank $f_0$ , it can be represented as $z = V(z - z_1) + z_1$ , where $z_1$ abstracts the mean vector of features, acting as an auxiliary variable to facilitate theoretical derivation and correct feature offsets, $V = LM^T$ ( $L \\in \\mathbb{R}^{f \\times f_0}$ , $M \\in \\mathbb{R}^{(p^2 d) \\times f_0}$ ) represents the low-rank projection matrix. The simplified response then becomes:", + "bbox": [ + 89, + 321, + 483, + 443 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = L M ^ {T} s + c ^ {\\prime} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 450, + 480, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $f_0 < f$ , the computational complexity reduces from $O(fp^2 d)$ to $O(f_0p^2 d) + O(ff_0)$ , where $O(ff_0) \\ll O(fp^2 d)$ , indicating that the computational complexity reduction is proportional to the original ratio of $f_0 / f$ .", + "bbox": [ + 89, + 479, + 483, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In BottConv, pointwise convolutions project features into or out of low-rank subspace, thus significantly reducing complexity, while depthwise convolution that perform spatial information-adequate extraction in subspace guarantees negligibly low complexity. As shown in Figure 5, BottConv in our GBC design significantly reduces parameter count and computational load compared to the original convolution, with minimal performance impact.", + "bbox": [ + 89, + 541, + 483, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, the input feature $x \\in \\mathbb{R}^{C \\times H \\times W}$ is retained as $x_{\\mathrm{residual}} = x$ to facilitate the residual connection. Subsequently, the feature $x$ is passed through the BottConv layer, followed by normalization and activation functions, resulting in the features $x_{1}$ and $g_{2}(x)$ as shown below:", + "bbox": [ + 89, + 659, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {1} (x) = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {1} \\left(f _ {1} (x)\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 737, + 480, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {1} = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {2} \\left(\\operatorname {B o t t C o n v} _ {2} \\left(g _ {1} (x)\\right)\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 760, + 480, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {2} (x) = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {3} \\left(\\operatorname {B o t t C o n v} _ {3} (x)\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 784, + 480, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate the gating feature map, $x_{1}$ and $g_{2}(x)$ are combined through the Hadamard product:", + "bbox": [ + 89, + 806, + 483, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm (x) = x _ {1} \\odot g _ {2} (x) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 843, + 480, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The gating feature map $m(x)$ is subsequently processed through BottConv once again to further refine fine-grained", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "details. After the residual connection is applied, the resulting output is:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {4} \\left(\\operatorname {B o t t C o n v} _ {4} (m (x)))\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 125, + 903, + 141 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO u t p u t = y + x _ {r e s i d u a l} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 148, + 903, + 164 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The design of BottConv and deeper gated branch enable the model to preserve basic crack features while dynamically refining the fine-grained feature characterization of the main branch, resulting in more accurate segmentation maps in detailed regions.", + "bbox": [ + 511, + 171, + 905, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Structure-Aware Visual State Space Module", + "text_level": 1, + "bbox": [ + 511, + 256, + 885, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our designed SAVSS features a two-dimensional selective scan (SS2D) tailored for visual tasks. Different scanning strategies impact the model's ability to capture continuous crack textures. As shown in Figure 4, current vision Mamba networks use various scanning directions, including parallel, snake, bidirectional, and diagonal scans [36, 55, 62]. Parallel and diagonal scans lack continuity across rows or diagonals, which limits their sensitivity to crack directions. Although bidirectional and snake scans maintain semantic continuity along horizontal or vertical paths, they struggle to capture diagonal or interwoven textures. To address this, our proposed diagonal snake scanning is designed to better capture complex crack texture cues.", + "bbox": [ + 511, + 277, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SASS consists of four paths: two parallel snake paths and two diagonal snake paths. This design enables the effective extraction of continuous semantic information in regular crack regions while preserving texture continuity in multiple directions, making it suitable for multi-scenario crack images with complex backgrounds.", + "bbox": [ + 511, + 474, + 905, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the RGB crack image undergoes Patch Embedding and Position Encoding, it is input as a sequence into the SAVSS block. To maintain a lightweight network, we use only 4 layers of SAVSS blocks. The processing equations are as follows:", + "bbox": [ + 511, + 565, + 905, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {P} = e ^ {\\Delta P} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 674, + 638, + 903, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {Q} = (\\Delta P) ^ {- 1} \\left(e ^ {\\Delta P} - I\\right) \\cdot \\Delta Q \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 662, + 903, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {k} = \\bar {P} z _ {k - 1} + \\bar {Q} w _ {k} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 686, + 903, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu _ {k} = R z _ {k} + S w _ {k} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 712, + 903, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In these equations, the input $w \\in \\mathbb{R}^{t \\times D}$ , $P \\in \\mathbb{R}^{G \\times D}$ controls the hidden spatial state, $S \\in \\mathbb{R}^{D \\times D}$ is used to initialize the skip connection for input, $z_{k}$ represents the specific hidden state at time step $k$ , and $Q \\in \\mathbb{R}^{G \\times D}$ and $R \\in \\mathbb{R}^{G \\times D}$ are matrices with hidden spatial dimensions $G$ and temporal dimensions $D$ , respectively, obtained through selective scanning SS2D. These are trainable parameters that are updated accordingly. $u_{k}$ represents the output at time step $k$ . SASS establishes multi-directional adjacency relationships, allowing the hidden state $z_{k}$ to capture more intricate topological and textural details, while enabling the", + "bbox": [ + 511, + 733, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29409", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0532b642a2f524458e624dc188b8391866f2c477d57dae05b9d9d7a24f52d8b4.jpg", + "image_caption": [ + "Figure 4. Illustration of our proposed SASS and other scanning strategies. The first row presents four commonly used single scanning paths, along with our proposed diagonal snake path. The second row illustrates the execution flow of our proposed SASS scanning strategy." + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 862, + 296 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df70699a78c38b72c085edb127b7778cac2d1f80aeb41ed1e5ebc64dec2abe67.jpg", + "image_caption": [ + "Figure 5. Performance comparison between using BottConv and raw convolution in GBC on the TUT [33] dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 348, + 483, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "output $u_{k}$ to more effectively integrate multi-directional features.", + "bbox": [ + 89, + 489, + 482, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively combine the initial sequence $x$ with the sequence processed through SS2D, we incorporate Pixel Attention-oriented Fusion (PAF) [33], enhancing SAVSS's ability to capture crack shape and texture details. Following selective scanning, a residual connection is applied to the fused information to preserve detail and facilitate feature flow. Furthermore, GBC refines the inter-layer output within SAVSS, strengthening crack information extraction and boosting performance in later stages.", + "bbox": [ + 89, + 520, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-scale Feature Segmentation Head", + "text_level": 1, + "bbox": [ + 89, + 666, + 429, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike convolutional layers, the MLP swiftly learns the mapping relationships between features and labels, thereby reducing model complexity. When the four feature maps $F_{1}, F_{2}, F_{3}, F_{4} \\in \\mathbb{R}^{C \\times H \\times W}$ produced by SAVSS are fed into MFS, they undergo individual processing through the efficient MLP operation and dynamic upsampling [34], restoring their resolution to the original size and yielding $F_{1}^{\\mathrm{up}}$ , $F_{2}^{\\mathrm{up}}$ , $F_{3}^{\\mathrm{up}}$ , $F_{4}^{\\mathrm{up}} \\in \\mathbb{R}^{C \\times H \\times W}$ . The formula is as follows:", + "bbox": [ + 89, + 686, + 482, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {u p} = D y S a m p l e \\left(M L P _ {i} \\left(F _ {i}\\right)\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 816, + 482, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i$ denotes the layer index.", + "bbox": [ + 89, + 840, + 303, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To integrate all multi-scale crack shape and texture representations, these feature maps are aggregated into a single tensor, obtaining a high-quality crack segmentation map", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$o\\in \\mathbb{R}^{1\\times H\\times W}$ as follows:", + "bbox": [ + 513, + 348, + 687, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\no _ {1} = G B C \\left(\\text {C o n c a t} \\left(F _ {1} ^ {u p}, F _ {2} ^ {u p}, F _ {3} ^ {u p}, F _ {4} ^ {u p}\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 377, + 906, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\no = M L P \\left(\\operatorname {C o n v} \\left(o _ {1}\\right)\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 407, + 905, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Objective Function", + "text_level": 1, + "bbox": [ + 513, + 431, + 696, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a blend of Binary Cross-Entropy loss (BCE) [29] and Dice loss [43] as the objective function, which helps improve the network's robustness to imbalanced pixel data. The overall loss function is expressed as follows:", + "bbox": [ + 511, + 454, + 905, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\alpha \\cdot L _ {D i c e} + \\beta \\cdot L _ {B C E} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 520, + 903, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the hyperparameters $\\alpha$ and $\\beta$ control the weights of the two loss components. The ratio of $\\alpha$ to $\\beta$ is set to 1:5.", + "bbox": [ + 511, + 544, + 906, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 580, + 645, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 606, + 614, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crack500 [56]. The images in this dataset were captured using a mobile phone. The original dataset consists of 500 bitumen crack images, which were expanded to 3368 images through data augmentation. Each image is paired with a corresponding pixel-level annotated binary image.", + "bbox": [ + 511, + 627, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DeepCrack [35]. The dataset comprises 537 RGB images of cement, bricks and bitumen cracks under various conditions, including fine, wide, stained, and fuzzy cracks, ensuring diversity and representativeness.", + "bbox": [ + 511, + 704, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**CrackMap [22].** The dataset was created for a road crack segmentation study and consists of 120 high-resolution RGB images capturing a variety of thin and complex bitumen road cracks.", + "bbox": [ + 511, + 763, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TUT [33]. In contrast to other datasets with simple backgrounds, this dataset includes a dense, cluttered background and features cracks with elaborate, intricate shapes. It contains 1408 RGB images across eight scenarios: bitumen, cement, bricks, runway, tiles, metal, blades, and pipes.", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29410", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/85a1d14f2f2da22f91a78b0e6f4bc308559fee5ca8eae5ab61884450ae63b721.jpg", + "image_caption": [ + "Figure 6. Visual comparison of typical cracks with 9 SOTA methods across four datasets. Red boxes highlight critical details, and green boxes mark misidentified regions." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 903, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During processing, all datasets were divided into training, validation, and test sets with a 7:1:2 ratio.", + "bbox": [ + 89, + 398, + 483, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details.", + "text_level": 1, + "bbox": [ + 89, + 441, + 312, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings. We built our SCSegamba network using PyTorch v1.13.1 and trained it on an Intel Xeon Platinum 8336C CPU with eight Nvidia GeForce RTX 4090 GPUs. The AdamW optimizer was used with an initial learning rate of 5e-4, PolyLR scheduling, a weight decay of 0.01, and a random seed of 42. The network was trained for 50 epochs, and the model with the best validation performance was selected for testing.", + "bbox": [ + 89, + 460, + 483, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods. To comprehensively evaluate our model, we compared SCSegamba with 9 SOTA methods. The CNN or Transformer-based models included RIND [38], SFIAN [5], CTCrackSeg [44], DTrCNet [48], Crackmer [46] and SimCrack [20]. Additionally, we compared it with other Mamba-based models, including CSMamba [37], PlainMamba [55], and MambaIR [16].", + "bbox": [ + 89, + 583, + 483, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We used six metrics to evaluate SCSegamba's performance: Precision (P), Recall (R), F1 Score $(F1 = \\frac{2RP}{R + P})$ , Optimal Dataset Scale (ODS), Optimal Image Scale (OIS), and mean Intersection over Union (mIoU). ODS measures the model's adaptability to datasets of varying scales at a fixed threshold m, while OIS evaluates adaptability across image scales at an optimal threshold n. The calculation formulas are as follows:", + "bbox": [ + 89, + 690, + 483, + 810 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nO D S = \\max _ {m} \\frac {2 \\cdot P _ {m} \\cdot R _ {m}}{P _ {m} + R _ {m}} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 821, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nO I S = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\max _ {n} \\frac {2 \\cdot P _ {n , i} \\cdot R _ {n , i}}{P _ {n , i} + R _ {n , i}} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 863, + 482, + 904 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mIoU is used to measure the mean proportion of the intersection over union between the ground truth and the predicted results. The calculation is given by the formula:", + "bbox": [ + 511, + 398, + 906, + 444 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm I o U = \\frac {1}{N + 1} \\sum_ {l = 0} ^ {N} \\frac {p u l}{\\sum_ {t = 0} ^ {N} p _ {t l} + \\sum_ {t = 0} ^ {N} p _ {t l} - p _ {u l}} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 455, + 906, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the number of classes, which we set as $N = 1$ ; $t$ represents the ground truth, $l$ represents the predicted value, and $p_{tl}$ represents the count of pixels classified as $l$ but belonging to $t$ .", + "bbox": [ + 511, + 508, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we evaluated our method's complexity using three metrics: FLOPs, Params, and Model Size, representing computational complexity, parameter complexity, and memory footprint.", + "bbox": [ + 511, + 570, + 905, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 638, + 805, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As listed in Table 1, compared with 9 other SOTA methods, our proposed SCSegamba achieves the best performance across four public datasets. Specifically, on the Crack500 [56] and DeepCrack [35] datasets, which contain larger and more complex crack regions, SCSegamba achieved the highest performance. Notably, on the DeepCrack dataset, it surpassed the next best method by $1.50\\%$ in F1 score and $1.09\\%$ in mIoU. This improvement is due to the robust ability of GBC to capture morphological clues in large crack areas, enhancing the model's representational power. On the CrackMap [22] dataset, which features thinner and more elongated cracks, our method surpasses all other SOTA methods in every metric, outperforming the next best method by $2.06\\%$ in F1 and $1.65\\%$ in mIoU. This demonstrates the effectiveness of SASS in capturing fine textures and elongated crack structures. As illustrated in", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3d318c16d3c9ff4fd57fad47b0fd78edd7c927dd95777463599ae524028ec9c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCrack500DeepCrack
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.64690.64830.69980.72450.71190.73810.80870.82670.78960.89200.83770.8391
SFIAN [5]0.69770.73480.69830.77420.73430.76040.86160.89280.85490.86920.86200.8776
CTCrackSeg [44]0.69410.70590.69400.77480.73220.75910.88190.89040.90110.88950.89520.8925
DTrCNet [48]0.70120.72410.65270.82800.73570.76270.84730.85120.89050.82510.85660.8661
Crackmer [46]0.69330.70970.69850.75720.72670.75910.87120.87850.89460.87830.88640.8844
SimCrack [20]0.71270.73080.70930.79840.75160.77150.85700.87220.89840.85490.87610.8744
CSMamba [37]0.69310.71620.68580.78230.73150.75920.87380.87660.90250.88630.89430.8863
PlainMamba [55]0.70350.71730.71700.75570.73580.76820.86460.86680.90500.86590.88500.8788
MambaIR [16]0.70430.71890.72040.76810.74350.76630.87960.88400.90560.88950.89750.8907
SCSegamba (Ours)0.72440.73700.72700.78590.75530.77780.89380.89900.90970.91240.91100.9022
MethodsCrackMapTUT
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.67450.69430.60230.75860.66990.74250.75310.78910.78720.76650.77670.8051
SFIAN [5]0.72000.74650.67150.76680.71600.77480.72900.75130.77150.73670.75370.7896
CTCrackSeg [44]0.72890.73730.69110.76690.72700.77850.79400.79960.82020.81950.81990.8301
DTrCNet [48]0.73280.74130.69120.76810.72760.78120.79870.80730.79720.84410.82020.8331
Crackmer [46]0.73950.74370.72290.74670.73460.78600.74290.76400.75010.76560.75780.7966
SimCrack [20]0.75590.76250.73800.76720.75230.79630.79840.80900.80510.83710.82080.8334
CSMamba [37]0.73710.74130.70530.76630.73460.78410.78790.79460.79470.83530.81460.8263
PlainMamba [55]0.71500.71890.66490.76160.70990.76990.78670.79670.77010.85230.81020.8253
MambaIR [16]0.73320.73470.75690.70130.72800.78340.78610.79300.78770.83870.81250.8249
SCSegamba (Ours)0.77410.77660.76290.77270.76780.80940.82040.82550.82410.85450.83900.8479
", + "bbox": [ + 91, + 90, + 908, + 460 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/107c194521607489837debbaf34564e439809c1cb730b5f2f3fa8f0fc37d12e6.jpg", + "table_caption": [ + "Table 1. Comparison with 9 SOTA methods across 4 datasets. Best results are in bold, and second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsYearFLOPs↓Params↓Size↓
RIND [38]2021695.77G59.39M453MB
SFIAN [5]202384.57G13.63M56MB
CTCrackSeg [44]202339.47G22.88M174MB
DTrCNet [48]2023123.20G63.45M317MB
Crackmer [46]202414.94G5.90M43MB
SimCrack [20]2024286.62G29.58M225MB
CSMamba [37]2024145.84G35.95M233MB
PlainMamba [55]202473.36G16.72M96MB
MambaIR [16]202447.32G10.34M79MB
SCSegamba (Ours)202418.16G2.80M37MB
", + "bbox": [ + 93, + 497, + 480, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of complexity with other methods. Best results are in bold, and second-best results are underlined.", + "bbox": [ + 89, + 684, + 482, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6, our method produces clearer and more precise feature maps, with superior detail capture in typical scenarios such as cement and bitumen, compared to other methods.", + "bbox": [ + 89, + 717, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the TUT dataset [33], which includes eight diverse scenarios, our method achieved the best performance, surpassing the next best method by $2.21\\%$ in F1 and $1.74\\%$ in mIoU. As shown in Figure 6, whether in the complex crack topology of plastic tracks, the noise-heavy backgrounds of metallic materials and turbine blades, or the low-contrast, dimly lit underground pipeline images, SCSegamba consistently produced high-quality segmentation maps while effectively suppressing irrelevant noise. This demonstrates", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that our method, with the enhanced crack morphology and texture perception from GBC and SASS, exhibits exceptional robustness and stability. Additionally, leveraging MFS for feature aggregation improves multi-scale perception, making our model particularly suited for diverse, interference-rich scenarios.", + "bbox": [ + 511, + 500, + 906, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Complexity Analysis", + "text_level": 1, + "bbox": [ + 511, + 594, + 709, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows a comparison of the complexity of our method with other SOTA methods when the input image size is uniformly set to 512. With only 2.80M parameters and a model size of 37MB, our method surpasses all others, being $52.54\\%$ and $13.95\\%$ lower than the next best result, respectively. Additionally, compared to Crackmer [46], which prioritizes computational efficiency, our method's FLOPs are only 3.22G higher. This demonstrates that the combination of lightweight SAVSS and MFS enables high-quality segmentation in noisy crack scenes with minimal parameters and low computational load, which is essential for resource-constrained devices.", + "bbox": [ + 511, + 609, + 906, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 792, + 679, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We performed ablation experiments on the representative multi-scenario dataset TUT [33].", + "bbox": [ + 511, + 810, + 903, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study of segmentation heads. As listed in Table 3, with our designed MFS, SCSegamba achieved the best results across all six metrics, with F1 and mIoU scores $1.57\\%$ and $1.21\\%$ higher than the second-best method. In", + "bbox": [ + 511, + 840, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29412", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fbc443af4a42095e11efae0256e5c2a748368d301fac0888a80ed0b53e485522.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Seg HeadODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
UNet [42]0.80550.81510.81480.83760.82600.83782.92M19.27G39MB
Ham [11]0.77030.77840.79620.78380.79090.81242.86M35.08G38MB
SegFormer [50]0.79470.79830.81700.81740.81720.83072.79M17.87G35MB
MFS0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
", + "bbox": [ + 109, + 88, + 887, + 175 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b92bd76e46e98ccf89a69a26aa92362b07614fd8437cda923486a7e8666f3f5c.jpg", + "table_caption": [ + "Table 3. Ablation study of different segmentation heads. UNet [42], Ham [11], and SegFormer [50] are high-performance heads." + ], + "table_footnote": [], + "table_body": "
GBCPAFResODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
✓XX0.81360.81960.82130.84610.83350.84342.49M16.75G34MB
X✓X0.79980.80840.79180.85240.82220.83432.28M14.91G33MB
XX✓0.79360.80690.79520.84380.81970.83132.48M15.65G35MB
✓✓X0.80470.81020.81740.83790.82750.83772.54M17.08G35MB
✓X✓0.81160.82000.81560.85220.83340.84252.75M17.82G37MB
X✓✓0.80230.80760.82190.83020.82600.83602.54M15.99G35MB
✓✓✓0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
", + "bbox": [ + 99, + 205, + 897, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "terms of complexity, although Params, FLOPs, and Model Size are only 0.01M, 0.29G, and 2MB larger than the Seg-Former head, our method surpasses it in F1 and mIoU by $2.67\\%$ and $2.07\\%$ , respectively. This demonstrates that MFS enhances SAVSS output integration, significantly improving performance while keeping the model lightweight.", + "bbox": [ + 89, + 366, + 482, + 455 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/03b094cd3e557221dd85e4322d0d5acc68935cc1b4362671fcf9edb2574eb9e9.jpg", + "table_caption": [ + "Table 4. Ablation study of components within the SAVSS block. Best results are in bold, and second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
ScanODSOISPRF1mIoU
Parallel0.81230.81840.81460.85230.83300.8427
Diag0.80910.81480.82250.84170.83200.8410
ParaSna0.81020.81620.82190.83650.82910.8408
DiagSna0.81530.82150.82370.84970.83650.8451
SASS0.82040.82550.82410.85450.83900.8479
", + "bbox": [ + 91, + 465, + 483, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies with different four-route scanning strategies in the SAVSS block, comparing parallel, diagonal, parallel snake, and diagonal snake scanning. The impact of different scanning strategies on complexity is negligible; thus, complexity analysis is omitted from this table. Best results are in bold, and second-best results are underlined.", + "bbox": [ + 89, + 578, + 482, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study of components. Table 4 shows the impact of each component in SAVSS on model performance. When fully utilizing GBC, PAF, and residual connections, our model achieved the best results across all metrics. Notably, adding GBC led to significant improvements in F1 and mIoU by $1.57\\%$ and $1.42\\%$ , respectively, highlighting its strength in capturing crack morphology cues. Similarly, residual connections boosted F1 and mIoU by $0.13\\%$ and $2.47\\%$ , indicating their role in focusing on essential crack features. Although using only PAF resulted in the lowest Params, FLOPs, and Model Size, it significantly reduced performance. These findings demonstrate that our fully integrated SAVSS effectively captures crack morphology and texture cues, achieving top pixel-level segmentation results while maintaining a lightweight model.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies of scanning strategies. As listed in Table 5, under the same conditions of using four different directional scanning paths, the model achieved the best performance with our designed SASS scanning strategy, improving F1 and mIoU by $0.30\\%$ and $0.33\\%$ over the diagonal snake strategy. This demonstrates SASS's ability to construct semantic and dependency information suited to crack topology, enhancing crack pixel perception in subsequent modules. More comprehensive experiments and real-world deployments are available in the Appendix.", + "bbox": [ + 511, + 366, + 903, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 520, + 633, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed SCSegamba, a lightweight structure-aware Vision Mamba for precise pixel-level crack segmentation. SCSegamba combines SAVSS and MFS to enhance crack shape and texture perception with a low parameter count. Equipped with GBC and SASS scanning, SAVSS, captures irregular crack textures across various structures. Experiments on four datasets show SCSegamba's exceptional performance, especially in complex, noisy scenarios. On the challenging multi-scenario dataset, it achieved an F1 score of 0.8390 and mIoU of 0.8479 with only 18.16G FLOPs and 2.8M parameters, demonstrating its effectiveness for real-world crack detection and suitability for edge devices. Future work will incorporate multimodal cues to enhance segmentation quality, while further optimizing VSS design and scan strategies to achieve high-quality results with low computational resources.", + "bbox": [ + 511, + 537, + 903, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 782, + 692, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62272342, 62020106004, 62306212, and T2422015; the Tianjin Natural Science Foundation under Grants 23JCJQJC00070 and 24PTLYHZ00320; and the Marie Skłodowska-Curie Actions (MSCA) under Project No. 101111188.", + "bbox": [ + 511, + 801, + 903, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29413", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zaid Al-Huda, Bo Peng, Riyadh Nazar Ali Algburi, Mugahed A Al-antari, AL-Jarazi Rabea, Omar Al-maqtari, and Donghai Zhai. Asymmetric dual-decoder-u-net for pavement crack semantic segmentation. Automation in Construction, 156:105138, 2023. 2", + "[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision, pages 801-818, 2018. 1", + "[3] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4703-4712, 2022. 1", + "[4] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 1", + "[5] Xu Cheng, Tian He, Fan Shi, Meng Zhao, Xiufeng Liu, and Shengyong Chen. Selective feature fusion and irregular-aware network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1, 2, 6, 7", + "[6] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2", + "[7] Wooram Choi and Young-Jin Cha. Sddnet: Real-time crack segmentation. IEEE Transactions on Industrial Electronics, 67(9):8016-8025, 2019. 2", + "[8] Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In International conference on machine learning, pages 933-941. PMLR, 2017. 4", + "[9] Jiaxiu Dong, Niannian Wang, Hongyuan Fang, Wentong Guo, Bin Li, and Kejie Zhai. Mfafnet: An innovative crack intelligent segmentation method based on multi-layer feature association fusion network. Advanced Engineering Informatics, 62:102584, 2024. 1, 2", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations, 2021. 1, 3", + "[11] Zhengyang Geng, Meng-Hao Guo, Hongxu Chen, Xia Li, Ke Wei, and Zhouchen Lin. Is attention better than matrix decomposition? In International Conference on Learning Representations, 2021. 8", + "[12] Karan Goel, Albert Gu, Chris Donahue, and Christopher Ré. It's raw! audio generation with state-space models. In International Conference on Machine Learning, pages 7616-7633. PMLR, 2022. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3", + "[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2, 3", + "[15] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2", + "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2024. 2, 6, 7", + "[17] Jing-Ming Guo, Herleeyandi Markoni, and Jiann-Der Lee. Barnet: Boundary aware refinement network for crack detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):7343-7358, 2021. 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2", + "[19] Yung-An Hsieh and Yichang James Tsai. Machine learning for crack detection: Review and model performance comparison. Journal of Computing in Civil Engineering, 34(5): 04020038, 2020. 1", + "[20] Achref Jaziri, Martin Mundt, Andres Fernandez, and Visvanathan Ramesh. Designing a hybrid neural system to learn real-world crack segmentation from fractal-based simulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8636-8646, 2024. 6, 7", + "[21] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 2", + "[22] Jason Katsamenis, Eftychios Protopapadakis, Nikolaos Bakalos, Andreas Varvarigos, Anastasios Doulamis, Nikolaos Doulamis, and Athanasios Voulodimos. A few-shot attention recurrent residual u-net for crack segmentation. In International Symposium on Visual Computing, pages 199-209. Springer, 2023. 5, 6", + "[23] Narges Kheradmandi and Vida Mehranfar. A critical review and comparative study on image segmentation-based techniques for pavement crack detection. Construction and Building Materials, 321:126162, 2022. 1", + "[24] Hong Lang, Ye Yuan, Jiang Chen, Shuo Ding, Jian John Lu, and Yong Zhang. Augmented concrete crack segmentation: Learning complete representation to defend background interference in concrete pavements. IEEE Transactions on Instrumentation and Measurement, 2024. 1", + "[25] David Lattanzi and Gregory R Miller. Robust automated concrete damage detection algorithms for field applications." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29414", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Journal of Computing in Civil Engineering, 28(2):253-262, 2014. 2", + "[26] Qin Lei, Jiang Zhong, and Chen Wang. Joint optimization of crack segmentation with an adaptive dynamic threshold module. IEEE Transactions on Intelligent Transportation Systems, 2024. 2", + "[27] Huaiyuan Li, Hui Li, Chuang Li, Baohai Wu, and Jinghuai Gao. Hybrid swim transformer-cnn model for pore-crack structure identification. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2", + "[28] Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, and Chengjie Wang. Lors: Low-rank residual structure for parameter-efficient network stacking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15866-15876, 2024. 4", + "[29] Qiufu Li, Xi Jia, Jiancan Zhou, Linlin Shen, and Jinming Duan. Rediscovering bce loss for uniform classification. arXiv preprint arXiv:2403.07289, 2024. 5", + "[30] Jianghai Liao, Yuanhao Yue, Dejin Zhang, Wei Tu, Rui Cao, Qin Zou, and Qingquan Li. Automatic tunnel crack inspection using an efficient mobile imaging module and a lightweight cnn. IEEE Transactions on Intelligent Transportation Systems, 23(9):15190-15203, 2022. 1", + "[31] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3783-3792, 2021. 2", + "[32] Huajun Liu, Jing Yang, Xiangyu Miao, Christoph Mertz, and Hui Kong. Crackformer network for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 24(9):9240-9252, 2023. 2", + "[33] Hui Liu, Chen Jia, Fan Shi, Xu Cheng, Mianzhao Wang, and Shengyong Chen. Staircase cascaded fusion of lightweight local pattern recognition and long-range dependencies for structural crack segmentation. arXiv preprint arXiv:2408.12815, 2024. 1, 5, 7", + "[34] Wenze Liu, Hao Lu, Hongtao Fu, and Zhiguo Cao. Learning to upsample by learning to sample. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6027-6037, 2023. 5", + "[35] Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li. Deepcrack: A deep hierarchical feature learning architecture for crack segmentation. Neurocomputing, 338:139-153, 2019. 2, 5, 6", + "[36] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4", + "[37] Liu Mushui, Jun Dan, Ziqian Lu, Yunlong Yu, Yingming Li, and Xi Li. Cm-unet: Hybrid cnn-mamba unet for remote sensing image semantic segmentation. arXiv preprint arXiv:2405.10530, 2024. 2, 6, 7", + "[38] Mengyang Pu, Yaping Huang, Qingji Guan, and Haibin Ling. Rindnet: Edge detection for discontinuity in reflectance, illumination, normal and depth. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6879-6888, 2021. 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Haochen Qi, Xiangwei Kong, Zhibo Jin, Jiqiang Zhang, and Zinan Wang. A vision-transformer-based convex variational network for bridge pavement defect segmentation. IEEE Transactions on Intelligent Transportation Systems, 2024. 2", + "[40] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 23(8):11710-11719, 2021. 2", + "[41] Jianing Quan, Baozhen Ge, and Min Wang. Crackvit: a unified cnn-transformer model for pixel-level crack extraction. Neural Computing and Applications, 35(15):10957-10973, 2023. 2", + "[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 8", + "[43] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, pages 240–248. Springer, 2017. 5", + "[44] Huaqi Tao, Bingxi Liu, Jinqiang Cui, and Hong Zhang. A convolutional-transformer network for crack segmentation with boundary awareness. In 2023 IEEE International Conference on Image Processing, pages 86-90. IEEE, 2023. 2, 6, 7", + "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2", + "[46] Jin Wang, Zhigao Zeng, Pradip Kumar Sharma, Osama Alfarraj, Amr Tolba, Jianming Zhang, and Lei Wang. Dualpath network combining cnn and transformer for pavement crack segmentation. Automation in Construction, 158: 105217, 2024. 1, 6, 7", + "[47] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5493-5502, 2024. 1", + "[48] Chao Xiang, Jingjing Guo, Ran Cao, and Lu Deng. A crack-segmentation algorithm fusing transformers and convolutional neural networks for complex detection scenarios. Automation in Construction, 152:104894, 2023. 1, 2, 6, 7", + "[49] Xiao Xiao, Shen Lian, Zhiming Luo, and Shaozi Li. Weighted res-unet for high-quality retina vessel segmentation. In 2018 9th international conference on information technology in medicine and education, pages 327-331. IEEE, 2018. 2", + "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29415", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2", + "[52] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2", + "[53] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2", + "[54] Tomoyuki Yamaguchi, Shingo Nakamura, Ryo Saegusa, and Shuji Hashimoto. Image-based crack detection for real concrete surfaces. IEEE Transactions on Electrical and Electronic Engineering, 3(1):128-135, 2008. 2", + "[55] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 2, 3, 4, 6, 7", + "[56] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 2, 5, 6", + "[57] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 4", + "[58] Hang Zhang, Allen A Zhang, Zishuo Dong, Anzheng He, Yang Liu, You Zhan, and Kelvin CP Wang. Robust semantic segmentation for automatic crack detection within pavement images using multi-mixing of global context and local image features. IEEE Transactions on Intelligent Transportation Systems, 2024. 1", + "[59] Tianjie Zhang, Donglei Wang, and Yang Lu. ECSnet: An accelerated real-time image segmentation cnn architecture for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1", + "[60] Xiaohu Zhang and Haifeng Huang. Distilling knowledge from a transformer-based crack segmentation model to a light-weighted symmetry model with mixed loss function for portable crack detection equipment. Symmetry, 16(5):520, 2024. 3", + "[61] Jian Zhou, Peisen S Huang, and Fu-Pen Chiang. Wavelet-based pavement distress detection and evaluation. Optical Engineering, 45(2):027007-027007, 2006. 2", + "[62] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In International Conference on Machine Learning, 2024. 2, 3, 4" + ], + "bbox": [ + 91, + 92, + 480, + 878 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29416", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_model.json b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f2a0600b15535d2bfc921af8d7f62428d24d8fa --- /dev/null +++ b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.13, + 0.845, + 0.175 + ], + "angle": 0, + "content": "SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.203, + 0.864, + 0.275 + ], + "angle": 0, + "content": "Hui Liu\\(^{1,2,3}\\), Chen Jia\\(^{1,2,3,*}\\), Fan Shi\\(^{1,2,3}\\), Xu Cheng\\(^{1,2,3}\\), Shengyong Chen\\(^{1,2,3}\\) \n\\(^{1}\\)School of Computer Science and Engineering, Tianjin University of Technology \n\\(^{2}\\)Engineering Research Center of Learning-Based Intelligent System (Ministry of Education) \n\\(^{3}\\)Key Laboratory of Computer Vision and System (Ministry of Education)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.277, + 0.885, + 0.293 + ], + "angle": 0, + "content": "liuhui1109@stud.tjut.edu.cn, {jiachen, shifan}@email.tjut.edu.cn, {xu.cheng, sy}@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.737 + ], + "angle": 0, + "content": "Pixel-level segmentation of structural cracks across various scenarios remains a considerable challenge. Current methods encounter challenges in effectively modeling crack morphology and texture, facing challenges in balancing segmentation quality with low computational resource usage. To overcome these limitations, we propose a lightweight Structure-Aware Vision Mamba Network (SCSegamba), capable of generating high-quality pixel-level segmentation maps by leveraging both the morphological information and texture cues of crack pixels with minimal computational cost. Specifically, we developed a Structure-Aware Visual State Space module (SAVSS), which incorporates a lightweight Gated Bottleneck Convolution (GBC) and a Structure-Aware Scanning Strategy (SASS). The key insight of GBC lies in its effectiveness in modeling the morphological information of cracks, while the SASS enhances the perception of crack topology and texture by strengthening the continuity of semantic information between crack pixels. Experiments on crack benchmark datasets demonstrate that our method outperforms other state-of-the-art (SOTA) methods, achieving the highest performance with only 2.8M parameters. On the multi-scenario dataset, our method reached 0.8390 in F1 score and 0.8479 in mIoU. The code is available at https://github.com/Karl1109/SCSegamba." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Structures like bitumen pavement, concrete, and metal frequently develop cracks under shear stress, making regular health monitoring essential to avoid production issues [3, 4, 19, 23, 30]. Due to differences in material properties and environmental conditions, various materials exhibit significant variations in crack morphology and visual appearance [24]. As a result, achieving pixel-level crack seg-" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.327, + 0.909, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.432, + 0.706, + 0.442 + ], + "angle": 0, + "content": "(a) Comparison with SOTA methods." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.445, + 0.705, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.477, + 0.851, + 0.488 + ], + "angle": 0, + "content": "(c) Segmentation results in complex interference conditions." + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.432, + 0.908, + 0.443 + ], + "angle": 0, + "content": "(b) Different numbers of SAVSS layers." + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.445, + 0.878, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Performance of SCSegamba on multi-scenario TUT [33] dataset. (a) Comparison with SOTA methods. (b) Impact of different SAVSS layer numbers on performance, with normalized metrics; FLOPs (G), Params (M), and Size (MB) decrease towards the edges. (c) Visual results under complex interference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.778 + ], + "angle": 0, + "content": "mentation across diverse scenarios remains a complex challenge. Recently, Convolutional Neural Networks (CNNs), such as ECSNet [59] and SFIAN [5], have shown effective crack feature extraction capabilities in segmentation tasks due to their strong local inductive properties. However, their limited receptive field constrains their ability to model broad-scope irregular dependencies across the entire image, resulting in discontinuous segmentation and weak background noise suppression. Although dilated convolution [2] expand the receptive field, their inherent inductive bias still prevents them from fully addressing this issue [58], especially in complex crack patterns with heavy background interference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The success of Vision Transformer (ViT) [10, 45, 47] has demonstrated Transformer's effectiveness in capturing irregular pixel dependencies, which is crucial for recognizing complex crack textures, as seen in networks like DTRCNet [48], MFAFNet [9], and Crackmer [46]. However, the quadratic complexity of attention calculations with sequence length leads to high memory use and training challenges for high-resolution images, limiting deployment on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "resource-constrained edge devices and practical applications. As shown in Figure 1(a), Transformer-based methods like CTCrackseg [44] and DTrCNet [48] perform well, but their large parameter counts and high computational demands limit their deployability on resource-constrained devices. Although variants like Sparse Transformer [6] and Linear Transformer [21] reduce computational requirements by sparsifying or linearizing attention, they sacrifice the ability to model irregular dependencies and pixel textures, hindering pixel-level detection tasks. Consequently, Transformer-based methods struggle to balance segmentation performance with computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.7 + ], + "angle": 0, + "content": "Recently, Selective State Space Models (SSMs) have attracted considerable interest due to Mamba's showing strong performance in sequence modeling while maintaining low computational demands [14, 15]. Vision Mamba (ViM) [62] and VMamba [36] have extended Mamba to the visual domain. The irregular extension of crack regions with numerous branches in low-contrast images, often affected by irrelevant areas and shadows, challenges existing Mamba VSS (Visual State Space Model) block structures and scanning strategies in capturing crack morphology and texture effectively. Most Mamba-based methods [37, 51, 52, 55] process feature maps through linear layers, limiting selective enhancement or suppression of crack features against irrelevant disturbances, thus reducing detailed morphological extraction. Additionally, common parallel or unidirectional diagonal scans [16] struggle to maintain semantic continuity when handling irregular, multi-directional pixel topologies, weakening their ability to manage complex textures and suppress noise. Consequently, current Mamba-based SSM frequently produce false or missed detections in multi-scenario crack images. Moreover, although these methods have the advantage of fewer parameters, there remains potential to further reduce their computational demands and enhance their deployability on edge devices. As shown in Figure 1(b), CSMamba [37], MambaIR [16], and PlainMamba [55] showed unsatisfactory performance on crack images, with room for improvement in parameter count and computational load." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To tackle the challenge of balancing high segmentation quality with low computational demands, we propose the SCSegamba network that produces high-quality pixel-level crack segmentation maps with low computational resources. To improve the Mamba network's perception of irregular crack textures, we design a Structure-Aware Visual State Space block (SAVSS), employing a Structure-Aware Scanning Strategy (SASS) to enhance semantic continuity and strengthen crack morphology perception. For capturing crack shape cues while maintaining low parameter and computational costs, we designed a lightweight Gated Bottleneck Convolution (GBC) that dynamically adjusts weights for complex backgrounds and varying mor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "phologies. Additionally, the Multi-scale Feature Segmentation head (MFS) integrates the GBC and Multi-layer Perceptron (MLP) to achieve high-quality segmentation maps with low computational requirements. As shown in Figure 1(c), optimal segmentation performance was achieved with four SAVSS layers, producing clear segmentation maps that effectively mask complex interference while maintaining model lightweightness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.241 + ], + "angle": 0, + "content": "In essence, the primary contributions of our work are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.245, + 0.905, + 0.32 + ], + "angle": 0, + "content": "- We propose a novel lightweight vision Mamba network, SCSegamba, for crack segmentation. This model effectively captures morphological and irregular texture cues of crack pixels, using low computational resources to generate high-quality segmentation maps." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.32, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- We design the SAVSS with a lightweight GBC Convolution and a SASS scanning strategy to enhance the handling and perception of irregular texture cues in crack images. Additionally, a simple yet effective MFS is developed to generate segmentation maps with relatively low computational resources." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.412, + 0.905, + 0.471 + ], + "angle": 0, + "content": "- We evaluate SCSegamba on benchmark datasets across diverse scenarios, with results demonstrating that our method outperforms other SOTA methods while maintaining the low parameter count." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.245, + 0.905, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.48, + 0.663, + 0.496 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.502, + 0.779, + 0.517 + ], + "angle": 0, + "content": "2.1. Crack Segmentation Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Early crack detection methods often relied on traditional feature extraction techniques, such as wavelet transform [61], percolation models [54], and the k-means algorithm [25]. Although straightforward to implement, these methods face challenges in suppressing background interference and achieving high segmentation accuracy. With advancements in deep learning, researchers have developed CNN-based crack segmentation networks that achieve SOTA performance [5, 7, 18, 26, 40, 49]. For instance, DeepCrack [35] enables end-to-end pixel-level segmentation, while FPHBN [56] demonstrates strong generalization capabilities. BARNet [17] detects crack boundaries by integrating image gradients with features, and ADDUNet [1] captures both fine and coarse crack characteristics across varied conditions. Although CNN-based methods show significant promise, the local operations and limited receptive fields of CNNs limit their ability to fully capture crack texture cues and effectively suppress background noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Transformers [45], with their self-attention mechanism, are well-suited for visual tasks that require long-range dependency modeling, making them increasingly popular in crack segmentation networks [9, 31, 32, 41, 53]. For example, VCVNet [39], based on Vision Transformer (ViT), is designed for bridge crack segmentation, addressing fine-grained segmentation challenges. SwT-CNN [27]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29407" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.096, + 0.576, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.273, + 0.416, + 0.285 + ], + "angle": 0, + "content": "(a) Network Architecture of SCSegamba" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.097, + 0.892, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.273, + 0.837, + 0.284 + ], + "angle": 0, + "content": "(b) Architecture of SAVSS block" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method. (a) illustrates the overall architecture of SCSegamba and the processing flow for crack images. (b) displays the structure of the SAVSS block. The input crack image undergoes comprehensive morphological and texture feature extraction through SAVSS, while MFS produces a high-quality pixel-level segmentation map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.514 + ], + "angle": 0, + "content": "combines Swin Transformer and CNN for automatic feature extraction, while TBUNet [60], a Transformer-based knowledge distillation model, achieves high-performance crack segmentation with a hybrid loss function. Although Transformer-based methods are highly effective at capturing crack texture cues and suppressing background noise, their self-attention mechanism introduces computational complexity that grows quadratically with sequence length. This results in a high parameter count and significant computational demands, which limit their deployment on resource-constrained edge devices." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.526, + 0.462, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.788 + ], + "angle": 0, + "content": "Figure 3. Architecture of GBC. It employs bottleneck convolution to efficiently reduce the parameters and computational load, while the gating mechanism enhances the model's adaptability in processing diverse crack patterns and complex backgrounds. GN represents group normalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.34, + 0.804 + ], + "angle": 0, + "content": "2.2. Selective State Space Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The introduction of the Selective State Space Models (S6) in the Mamba model [14] has highlighted the potential of SSM [12, 13]. Unlike the linear time-invariant S4 model, S6 efficiently captures complex long-distance dependencies while preserving computational efficiency, achieving strong performance in NLP, audio, and genomics. Consequently," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.528 + ], + "angle": 0, + "content": "researchers have adapted Mamba to the visual domain, creating various VSS blocks. ViM [62] achieves comparable modeling to ViT [10] without attention mechanisms, using fewer computational resources, while VMamba [36] prioritizes efficient computation and high performance. Plain-Mamba [55] employs a fixed-width layer stacking approach, excelling in tasks such as instance segmentation and object detection. However, the VSS block and scanning strategy require specific optimizations for each visual task, as tasks differ in their reliance on long- and short-distance information, necessitating customized VSS block designs to ensure optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Currently, no high-performing Mamba-based model exists for crack segmentation. Thus, designing an optimized VSS structure specifically for crack segmentation is essential to improve performance and efficiency. Given the intricate details and irregular textures of cracks, the VSS block requires strong shape extraction and directional awareness to effectively capture crack texture cues. Additionally, it should facilitate efficient crack segmentation while minimizing computational resource requirements." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.673, + 0.649, + 0.69 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.643, + 0.714 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.809 + ], + "angle": 0, + "content": "The complete architecture of our proposed SCSegamba is depicted in Figure 2. It includes two main components: the SAVSS for extracting crack shape and texture cues, and the MFS for efficient feature processing. To capture key crack region cues, we integrate the GBC at the initial stage of SAVSS and the final stage of MFS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For a single RGB image \\(E \\in \\mathbb{R}^{3 \\times H \\times W}\\), spatial information is divided into \\(n\\) patches, forming a sequence \\(\\{B_1, B_2, \\ldots, B_n\\}\\). This sequence is processed through the SAVSS block, embedding key crack pixel cues into multiscale feature maps \\(\\{F_1, F_2, F_3, F_4\\}\\). Finally, in the MFS, all information is consolidated into a single tensor, produc" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29408" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.09, + 0.425, + 0.107 + ], + "angle": 0, + "content": "ing a refined segmentation output \\(W\\in \\mathbb{R}^{1\\times H\\times W}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.116, + 0.457, + 0.132 + ], + "angle": 0, + "content": "3.2. Lightweight Gated Bottleneck Convolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.244 + ], + "angle": 0, + "content": "The gating mechanism enables dynamic features for each spatial position and channel, enhancing the model's ability to capture details [8, 57]. To further reduce parameter count and computational cost, we embedded a bottleneck convolution (BottConv) with low-rank approximation [28], mapping matrices from high to low dimensional spaces and significantly lowering computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.289 + ], + "angle": 0, + "content": "In the convolution layer, assuming the spatial size of the filter is \\( p \\), the number of input channels is \\( d \\) and the input is \\( s \\), the convolution response can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.295, + 0.482, + 0.31 + ], + "angle": 0, + "content": "\\[\nz = Q s + c \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.444 + ], + "angle": 0, + "content": "where \\(Q\\) is a matrix of size \\(f \\times (p^2 \\times d)\\), \\(f\\) is the number of output channels, and \\(c\\) is the original bias term. Assuming \\(z\\) lies in a low-rank subspace of rank \\(f_0\\), it can be represented as \\(z = V(z - z_1) + z_1\\), where \\(z_1\\) abstracts the mean vector of features, acting as an auxiliary variable to facilitate theoretical derivation and correct feature offsets, \\(V = LM^T\\) (\\(L \\in \\mathbb{R}^{f \\times f_0}\\), \\(M \\in \\mathbb{R}^{(p^2 d) \\times f_0}\\)) represents the low-rank projection matrix. The simplified response then becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.451, + 0.482, + 0.467 + ], + "angle": 0, + "content": "\\[\nz = L M ^ {T} s + c ^ {\\prime} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.54 + ], + "angle": 0, + "content": "Since \\( f_0 < f \\), the computational complexity reduces from \\( O(fp^2 d) \\) to \\( O(f_0p^2 d) + O(ff_0) \\), where \\( O(ff_0) \\ll O(fp^2 d) \\), indicating that the computational complexity reduction is proportional to the original ratio of \\( f_0 / f \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.66 + ], + "angle": 0, + "content": "In BottConv, pointwise convolutions project features into or out of low-rank subspace, thus significantly reducing complexity, while depthwise convolution that perform spatial information-adequate extraction in subspace guarantees negligibly low complexity. As shown in Figure 5, BottConv in our GBC design significantly reduces parameter count and computational load compared to the original convolution, with minimal performance impact." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.736 + ], + "angle": 0, + "content": "As shown in Figure 3, the input feature \\( x \\in \\mathbb{R}^{C \\times H \\times W} \\) is retained as \\( x_{\\mathrm{residual}} = x \\) to facilitate the residual connection. Subsequently, the feature \\( x \\) is passed through the BottConv layer, followed by normalization and activation functions, resulting in the features \\( x_{1} \\) and \\( g_{2}(x) \\) as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.738, + 0.482, + 0.754 + ], + "angle": 0, + "content": "\\[\ng _ {1} (x) = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {1} \\left(f _ {1} (x)\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.761, + 0.482, + 0.777 + ], + "angle": 0, + "content": "\\[\nx _ {1} = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {2} \\left(\\operatorname {B o t t C o n v} _ {2} \\left(g _ {1} (x)\\right)\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.785, + 0.482, + 0.801 + ], + "angle": 0, + "content": "\\[\ng _ {2} (x) = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {3} \\left(\\operatorname {B o t t C o n v} _ {3} (x)\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.838 + ], + "angle": 0, + "content": "To generate the gating feature map, \\( x_{1} \\) and \\( g_{2}(x) \\) are combined through the Hadamard product:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.844, + 0.482, + 0.859 + ], + "angle": 0, + "content": "\\[\nm (x) = x _ {1} \\odot g _ {2} (x) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The gating feature map \\( m(x) \\) is subsequently processed through BottConv once again to further refine fine-grained" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "details. After the residual connection is applied, the resulting output is:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.126, + 0.905, + 0.142 + ], + "angle": 0, + "content": "\\[\ny = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {4} \\left(\\operatorname {B o t t C o n v} _ {4} (m (x)))\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.15, + 0.905, + 0.165 + ], + "angle": 0, + "content": "\\[\nO u t p u t = y + x _ {r e s i d u a l} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.172, + 0.906, + 0.248 + ], + "angle": 0, + "content": "The design of BottConv and deeper gated branch enable the model to preserve basic crack features while dynamically refining the fine-grained feature characterization of the main branch, resulting in more accurate segmentation maps in detailed regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.257, + 0.887, + 0.273 + ], + "angle": 0, + "content": "3.3. Structure-Aware Visual State Space Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Our designed SAVSS features a two-dimensional selective scan (SS2D) tailored for visual tasks. Different scanning strategies impact the model's ability to capture continuous crack textures. As shown in Figure 4, current vision Mamba networks use various scanning directions, including parallel, snake, bidirectional, and diagonal scans [36, 55, 62]. Parallel and diagonal scans lack continuity across rows or diagonals, which limits their sensitivity to crack directions. Although bidirectional and snake scans maintain semantic continuity along horizontal or vertical paths, they struggle to capture diagonal or interwoven textures. To address this, our proposed diagonal snake scanning is designed to better capture complex crack texture cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.566 + ], + "angle": 0, + "content": "SASS consists of four paths: two parallel snake paths and two diagonal snake paths. This design enables the effective extraction of continuous semantic information in regular crack regions while preserving texture continuity in multiple directions, making it suitable for multi-scenario crack images with complex backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.639 + ], + "angle": 0, + "content": "After the RGB crack image undergoes Patch Embedding and Position Encoding, it is input as a sequence into the SAVSS block. To maintain a lightweight network, we use only 4 layers of SAVSS blocks. The processing equations are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.675, + 0.64, + 0.905, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\bar {P} = e ^ {\\Delta P} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.663, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\bar {Q} = (\\Delta P) ^ {- 1} \\left(e ^ {\\Delta P} - I\\right) \\cdot \\Delta Q \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.687, + 0.905, + 0.705 + ], + "angle": 0, + "content": "\\[\nz _ {k} = \\bar {P} z _ {k - 1} + \\bar {Q} w _ {k} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.713, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\nu _ {k} = R z _ {k} + S w _ {k} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In these equations, the input \\( w \\in \\mathbb{R}^{t \\times D} \\), \\( P \\in \\mathbb{R}^{G \\times D} \\) controls the hidden spatial state, \\( S \\in \\mathbb{R}^{D \\times D} \\) is used to initialize the skip connection for input, \\( z_{k} \\) represents the specific hidden state at time step \\( k \\), and \\( Q \\in \\mathbb{R}^{G \\times D} \\) and \\( R \\in \\mathbb{R}^{G \\times D} \\) are matrices with hidden spatial dimensions \\( G \\) and temporal dimensions \\( D \\), respectively, obtained through selective scanning SS2D. These are trainable parameters that are updated accordingly. \\( u_{k} \\) represents the output at time step \\( k \\). SASS establishes multi-directional adjacency relationships, allowing the hidden state \\( z_{k} \\) to capture more intricate topological and textural details, while enabling the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.863, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.311, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Illustration of our proposed SASS and other scanning strategies. The first row presents four commonly used single scanning paths, along with our proposed diagonal snake path. The second row illustrates the execution flow of our proposed SASS scanning strategy." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.349, + 0.485, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Figure 5. Performance comparison between using BottConv and raw convolution in GBC on the TUT [33] dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.519 + ], + "angle": 0, + "content": "output \\( u_{k} \\) to more effectively integrate multi-directional features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.658 + ], + "angle": 0, + "content": "To effectively combine the initial sequence \\( x \\) with the sequence processed through SS2D, we incorporate Pixel Attention-oriented Fusion (PAF) [33], enhancing SAVSS's ability to capture crack shape and texture details. Following selective scanning, a residual connection is applied to the fused information to preserve detail and facilitate feature flow. Furthermore, GBC refines the inter-layer output within SAVSS, strengthening crack information extraction and boosting performance in later stages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.431, + 0.683 + ], + "angle": 0, + "content": "3.4. Multi-scale Feature Segmentation Head" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Unlike convolutional layers, the MLP swiftly learns the mapping relationships between features and labels, thereby reducing model complexity. When the four feature maps \\( F_{1}, F_{2}, F_{3}, F_{4} \\in \\mathbb{R}^{C \\times H \\times W} \\) produced by SAVSS are fed into MFS, they undergo individual processing through the efficient MLP operation and dynamic upsampling [34], restoring their resolution to the original size and yielding \\( F_{1}^{\\mathrm{up}} \\), \\( F_{2}^{\\mathrm{up}} \\), \\( F_{3}^{\\mathrm{up}} \\), \\( F_{4}^{\\mathrm{up}} \\in \\mathbb{R}^{C \\times H \\times W} \\). The formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.817, + 0.483, + 0.835 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {u p} = D y S a m p l e \\left(M L P _ {i} \\left(F _ {i}\\right)\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.305, + 0.855 + ], + "angle": 0, + "content": "where \\(i\\) denotes the layer index." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To integrate all multi-scale crack shape and texture representations, these feature maps are aggregated into a single tensor, obtaining a high-quality crack segmentation map" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.349, + 0.688, + 0.365 + ], + "angle": 0, + "content": "\\(o\\in \\mathbb{R}^{1\\times H\\times W}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.378, + 0.907, + 0.396 + ], + "angle": 0, + "content": "\\[\no _ {1} = G B C \\left(\\text {C o n c a t} \\left(F _ {1} ^ {u p}, F _ {2} ^ {u p}, F _ {3} ^ {u p}, F _ {4} ^ {u p}\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.408, + 0.906, + 0.425 + ], + "angle": 0, + "content": "\\[\no = M L P \\left(\\operatorname {C o n v} \\left(o _ {1}\\right)\\right) \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.433, + 0.697, + 0.45 + ], + "angle": 0, + "content": "3.5. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.906, + 0.516 + ], + "angle": 0, + "content": "We use a blend of Binary Cross-Entropy loss (BCE) [29] and Dice loss [43] as the objective function, which helps improve the network's robustness to imbalanced pixel data. The overall loss function is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.521, + 0.905, + 0.537 + ], + "angle": 0, + "content": "\\[\nL = \\alpha \\cdot L _ {D i c e} + \\beta \\cdot L _ {B C E} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.575 + ], + "angle": 0, + "content": "where the hyperparameters \\(\\alpha\\) and \\(\\beta\\) control the weights of the two loss components. The ratio of \\(\\alpha\\) to \\(\\beta\\) is set to 1:5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.582, + 0.646, + 0.599 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.616, + 0.621 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Crack500 [56]. The images in this dataset were captured using a mobile phone. The original dataset consists of 500 bitumen crack images, which were expanded to 3368 images through data augmentation. Each image is paired with a corresponding pixel-level annotated binary image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.764 + ], + "angle": 0, + "content": "DeepCrack [35]. The dataset comprises 537 RGB images of cement, bricks and bitumen cracks under various conditions, including fine, wide, stained, and fuzzy cracks, ensuring diversity and representativeness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.824 + ], + "angle": 0, + "content": "**CrackMap [22].** The dataset was created for a road crack segmentation study and consists of 120 high-resolution RGB images capturing a variety of thin and complex bitumen road cracks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "TUT [33]. In contrast to other datasets with simple backgrounds, this dataset includes a dense, cluttered background and features cracks with elaborate, intricate shapes. It contains 1408 RGB images across eight scenarios: bitumen, cement, bricks, runway, tiles, metal, blades, and pipes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29410" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.905, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.361, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of typical cracks with 9 SOTA methods across four datasets. Red boxes highlight critical details, and green boxes mark misidentified regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.399, + 0.484, + 0.429 + ], + "angle": 0, + "content": "During processing, all datasets were divided into training, validation, and test sets with a 7:1:2 ratio." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.443, + 0.313, + 0.459 + ], + "angle": 0, + "content": "4.2. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.583 + ], + "angle": 0, + "content": "Experimental Settings. We built our SCSegamba network using PyTorch v1.13.1 and trained it on an Intel Xeon Platinum 8336C CPU with eight Nvidia GeForce RTX 4090 GPUs. The AdamW optimizer was used with an initial learning rate of 5e-4, PolyLR scheduling, a weight decay of 0.01, and a random seed of 42. The network was trained for 50 epochs, and the model with the best validation performance was selected for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Comparison Methods. To comprehensively evaluate our model, we compared SCSegamba with 9 SOTA methods. The CNN or Transformer-based models included RIND [38], SFIAN [5], CTCrackSeg [44], DTrCNet [48], Crackmer [46] and SimCrack [20]. Additionally, we compared it with other Mamba-based models, including CSMamba [37], PlainMamba [55], and MambaIR [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.484, + 0.811 + ], + "angle": 0, + "content": "Evaluation Metrics. We used six metrics to evaluate SCSegamba's performance: Precision (P), Recall (R), F1 Score \\((F1 = \\frac{2RP}{R + P})\\), Optimal Dataset Scale (ODS), Optimal Image Scale (OIS), and mean Intersection over Union (mIoU). ODS measures the model's adaptability to datasets of varying scales at a fixed threshold m, while OIS evaluates adaptability across image scales at an optimal threshold n. The calculation formulas are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.822, + 0.483, + 0.855 + ], + "angle": 0, + "content": "\\[\nO D S = \\max _ {m} \\frac {2 \\cdot P _ {m} \\cdot R _ {m}}{P _ {m} + R _ {m}} \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nO I S = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\max _ {n} \\frac {2 \\cdot P _ {n , i} \\cdot R _ {n , i}}{P _ {n , i} + R _ {n , i}} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.445 + ], + "angle": 0, + "content": "mIoU is used to measure the mean proportion of the intersection over union between the ground truth and the predicted results. The calculation is given by the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.456, + 0.907, + 0.499 + ], + "angle": 0, + "content": "\\[\nm I o U = \\frac {1}{N + 1} \\sum_ {l = 0} ^ {N} \\frac {p u l}{\\sum_ {t = 0} ^ {N} p _ {t l} + \\sum_ {t = 0} ^ {N} p _ {t l} - p _ {u l}} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.571 + ], + "angle": 0, + "content": "where \\(N\\) is the number of classes, which we set as \\(N = 1\\); \\(t\\) represents the ground truth, \\(l\\) represents the predicted value, and \\(p_{tl}\\) represents the count of pixels classified as \\(l\\) but belonging to \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.632 + ], + "angle": 0, + "content": "Additionally, we evaluated our method's complexity using three metrics: FLOPs, Params, and Model Size, representing computational complexity, parameter complexity, and memory footprint." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.64, + 0.807, + 0.657 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As listed in Table 1, compared with 9 other SOTA methods, our proposed SCSegamba achieves the best performance across four public datasets. Specifically, on the Crack500 [56] and DeepCrack [35] datasets, which contain larger and more complex crack regions, SCSegamba achieved the highest performance. Notably, on the DeepCrack dataset, it surpassed the next best method by \\(1.50\\%\\) in F1 score and \\(1.09\\%\\) in mIoU. This improvement is due to the robust ability of GBC to capture morphological clues in large crack areas, enhancing the model's representational power. On the CrackMap [22] dataset, which features thinner and more elongated cracks, our method surpasses all other SOTA methods in every metric, outperforming the next best method by \\(2.06\\%\\) in F1 and \\(1.65\\%\\) in mIoU. This demonstrates the effectiveness of SASS in capturing fine textures and elongated crack structures. As illustrated in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29411" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.091, + 0.909, + 0.462 + ], + "angle": 0, + "content": "
MethodsCrack500DeepCrack
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.64690.64830.69980.72450.71190.73810.80870.82670.78960.89200.83770.8391
SFIAN [5]0.69770.73480.69830.77420.73430.76040.86160.89280.85490.86920.86200.8776
CTCrackSeg [44]0.69410.70590.69400.77480.73220.75910.88190.89040.90110.88950.89520.8925
DTrCNet [48]0.70120.72410.65270.82800.73570.76270.84730.85120.89050.82510.85660.8661
Crackmer [46]0.69330.70970.69850.75720.72670.75910.87120.87850.89460.87830.88640.8844
SimCrack [20]0.71270.73080.70930.79840.75160.77150.85700.87220.89840.85490.87610.8744
CSMamba [37]0.69310.71620.68580.78230.73150.75920.87380.87660.90250.88630.89430.8863
PlainMamba [55]0.70350.71730.71700.75570.73580.76820.86460.86680.90500.86590.88500.8788
MambaIR [16]0.70430.71890.72040.76810.74350.76630.87960.88400.90560.88950.89750.8907
SCSegamba (Ours)0.72440.73700.72700.78590.75530.77780.89380.89900.90970.91240.91100.9022
MethodsCrackMapTUT
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.67450.69430.60230.75860.66990.74250.75310.78910.78720.76650.77670.8051
SFIAN [5]0.72000.74650.67150.76680.71600.77480.72900.75130.77150.73670.75370.7896
CTCrackSeg [44]0.72890.73730.69110.76690.72700.77850.79400.79960.82020.81950.81990.8301
DTrCNet [48]0.73280.74130.69120.76810.72760.78120.79870.80730.79720.84410.82020.8331
Crackmer [46]0.73950.74370.72290.74670.73460.78600.74290.76400.75010.76560.75780.7966
SimCrack [20]0.75590.76250.73800.76720.75230.79630.79840.80900.80510.83710.82080.8334
CSMamba [37]0.73710.74130.70530.76630.73460.78410.78790.79460.79470.83530.81460.8263
PlainMamba [55]0.71500.71890.66490.76160.70990.76990.78670.79670.77010.85230.81020.8253
MambaIR [16]0.73320.73470.75690.70130.72800.78340.78610.79300.78770.83870.81250.8249
SCSegamba (Ours)0.77410.77660.76290.77270.76780.80940.82040.82550.82410.85450.83900.8479
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.473, + 0.868, + 0.487 + ], + "angle": 0, + "content": "Table 1. Comparison with 9 SOTA methods across 4 datasets. Best results are in bold, and second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.498, + 0.481, + 0.674 + ], + "angle": 0, + "content": "
MethodsYearFLOPs↓Params↓Size↓
RIND [38]2021695.77G59.39M453MB
SFIAN [5]202384.57G13.63M56MB
CTCrackSeg [44]202339.47G22.88M174MB
DTrCNet [48]2023123.20G63.45M317MB
Crackmer [46]202414.94G5.90M43MB
SimCrack [20]2024286.62G29.58M225MB
CSMamba [37]2024145.84G35.95M233MB
PlainMamba [55]202473.36G16.72M96MB
MambaIR [16]202447.32G10.34M79MB
SCSegamba (Ours)202418.16G2.80M37MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Table 2. Comparison of complexity with other methods. Best results are in bold, and second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Figure 6, our method produces clearer and more precise feature maps, with superior detail capture in typical scenarios such as cement and bitumen, compared to other methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For the TUT dataset [33], which includes eight diverse scenarios, our method achieved the best performance, surpassing the next best method by \\(2.21\\%\\) in F1 and \\(1.74\\%\\) in mIoU. As shown in Figure 6, whether in the complex crack topology of plastic tracks, the noise-heavy backgrounds of metallic materials and turbine blades, or the low-contrast, dimly lit underground pipeline images, SCSegamba consistently produced high-quality segmentation maps while effectively suppressing irrelevant noise. This demonstrates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.59 + ], + "angle": 0, + "content": "that our method, with the enhanced crack morphology and texture perception from GBC and SASS, exhibits exceptional robustness and stability. Additionally, leveraging MFS for feature aggregation improves multi-scale perception, making our model particularly suited for diverse, interference-rich scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.595, + 0.71, + 0.61 + ], + "angle": 0, + "content": "4.4. Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Table 2 shows a comparison of the complexity of our method with other SOTA methods when the input image size is uniformly set to 512. With only 2.80M parameters and a model size of 37MB, our method surpasses all others, being \\(52.54\\%\\) and \\(13.95\\%\\) lower than the next best result, respectively. Additionally, compared to Crackmer [46], which prioritizes computational efficiency, our method's FLOPs are only 3.22G higher. This demonstrates that the combination of lightweight SAVSS and MFS enables high-quality segmentation in noisy crack scenes with minimal parameters and low computational load, which is essential for resource-constrained devices." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.68, + 0.808 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.841 + ], + "angle": 0, + "content": "We performed ablation experiments on the representative multi-scenario dataset TUT [33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Ablation study of segmentation heads. As listed in Table 3, with our designed MFS, SCSegamba achieved the best results across all six metrics, with F1 and mIoU scores \\(1.57\\%\\) and \\(1.21\\%\\) higher than the second-best method. In" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29412" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.888, + 0.176 + ], + "angle": 0, + "content": "
Seg HeadODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
UNet [42]0.80550.81510.81480.83760.82600.83782.92M19.27G39MB
Ham [11]0.77030.77840.79620.78380.79090.81242.86M35.08G38MB
SegFormer [50]0.79470.79830.81700.81740.81720.83072.79M17.87G35MB
MFS0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.186, + 0.877, + 0.2 + ], + "angle": 0, + "content": "Table 3. Ablation study of different segmentation heads. UNet [42], Ham [11], and SegFormer [50] are high-performance heads." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.207, + 0.898, + 0.339 + ], + "angle": 0, + "content": "
GBCPAFResODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
✓XX0.81360.81960.82130.84610.83350.84342.49M16.75G34MB
X✓X0.79980.80840.79180.85240.82220.83432.28M14.91G33MB
XX✓0.79360.80690.79520.84380.81970.83132.48M15.65G35MB
✓✓X0.80470.81020.81740.83790.82750.83772.54M17.08G35MB
✓X✓0.81160.82000.81560.85220.83340.84252.75M17.82G37MB
X✓✓0.80230.80760.82190.83020.82600.83602.54M15.99G35MB
✓✓✓0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.349, + 0.877, + 0.363 + ], + "angle": 0, + "content": "Table 4. Ablation study of components within the SAVSS block. Best results are in bold, and second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.457 + ], + "angle": 0, + "content": "terms of complexity, although Params, FLOPs, and Model Size are only 0.01M, 0.29G, and 2MB larger than the Seg-Former head, our method surpasses it in F1 and mIoU by \\(2.67\\%\\) and \\(2.07\\%\\), respectively. This demonstrates that MFS enhances SAVSS output integration, significantly improving performance while keeping the model lightweight." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.467, + 0.485, + 0.569 + ], + "angle": 0, + "content": "
ScanODSOISPRF1mIoU
Parallel0.81230.81840.81460.85230.83300.8427
Diag0.80910.81480.82250.84170.83200.8410
ParaSna0.81020.81620.82190.83650.82910.8408
DiagSna0.81530.82150.82370.84970.83650.8451
SASS0.82040.82550.82410.85450.83900.8479
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.662 + ], + "angle": 0, + "content": "Table 5. Ablation studies with different four-route scanning strategies in the SAVSS block, comparing parallel, diagonal, parallel snake, and diagonal snake scanning. The impact of different scanning strategies on complexity is negligible; thus, complexity analysis is omitted from this table. Best results are in bold, and second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation study of components. Table 4 shows the impact of each component in SAVSS on model performance. When fully utilizing GBC, PAF, and residual connections, our model achieved the best results across all metrics. Notably, adding GBC led to significant improvements in F1 and mIoU by \\(1.57\\%\\) and \\(1.42\\%\\), respectively, highlighting its strength in capturing crack morphology cues. Similarly, residual connections boosted F1 and mIoU by \\(0.13\\%\\) and \\(2.47\\%\\), indicating their role in focusing on essential crack features. Although using only PAF resulted in the lowest Params, FLOPs, and Model Size, it significantly reduced performance. These findings demonstrate that our fully integrated SAVSS effectively captures crack morphology and texture cues, achieving top pixel-level segmentation results while maintaining a lightweight model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.905, + 0.518 + ], + "angle": 0, + "content": "Ablation studies of scanning strategies. As listed in Table 5, under the same conditions of using four different directional scanning paths, the model achieved the best performance with our designed SASS scanning strategy, improving F1 and mIoU by \\(0.30\\%\\) and \\(0.33\\%\\) over the diagonal snake strategy. This demonstrates SASS's ability to construct semantic and dependency information suited to crack topology, enhancing crack pixel perception in subsequent modules. More comprehensive experiments and real-world deployments are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.521, + 0.634, + 0.536 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.78 + ], + "angle": 0, + "content": "In this paper, we proposed SCSegamba, a lightweight structure-aware Vision Mamba for precise pixel-level crack segmentation. SCSegamba combines SAVSS and MFS to enhance crack shape and texture perception with a low parameter count. Equipped with GBC and SASS scanning, SAVSS, captures irregular crack textures across various structures. Experiments on four datasets show SCSegamba's exceptional performance, especially in complex, noisy scenarios. On the challenging multi-scenario dataset, it achieved an F1 score of 0.8390 and mIoU of 0.8479 with only 18.16G FLOPs and 2.8M parameters, demonstrating its effectiveness for real-world crack detection and suitability for edge devices. Future work will incorporate multimodal cues to enhance segmentation quality, while further optimizing VSS design and scan strategies to achieve high-quality results with low computational resources." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.694, + 0.8 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.905, + 0.892 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62272342, 62020106004, 62306212, and T2422015; the Tianjin Natural Science Foundation under Grants 23JCJQJC00070 and 24PTLYHZ00320; and the Marie Skłodowska-Curie Actions (MSCA) under Project No. 101111188." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29413" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Zaid Al-Huda, Bo Peng, Riyadh Nazar Ali Algburi, Mugahed A Al-antari, AL-Jarazi Rabea, Omar Al-maqtari, and Donghai Zhai. Asymmetric dual-decoder-u-net for pavement crack semantic segmentation. Automation in Construction, 156:105138, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision, pages 801-818, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4703-4712, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[4] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[5] Xu Cheng, Tian He, Fan Shi, Meng Zhao, Xiufeng Liu, and Shengyong Chen. Selective feature fusion and irregular-aware network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[6] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[7] Wooram Choi and Young-Jin Cha. Sddnet: Real-time crack segmentation. IEEE Transactions on Industrial Electronics, 67(9):8016-8025, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[8] Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In International conference on machine learning, pages 933-941. PMLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[9] Jiaxiu Dong, Niannian Wang, Hongyuan Fang, Wentong Guo, Bin Li, and Kejie Zhai. Mfafnet: An innovative crack intelligent segmentation method based on multi-layer feature association fusion network. Advanced Engineering Informatics, 62:102584, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[11] Zhengyang Geng, Meng-Hao Guo, Hongxu Chen, Xia Li, Ke Wei, and Zhouchen Lin. Is attention better than matrix decomposition? In International Conference on Learning Representations, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] Karan Goel, Albert Gu, Chris Donahue, and Christopher Ré. It's raw! audio generation with state-space models. In International Conference on Machine Learning, pages 7616-7633. PMLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[15] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.904, + 0.389 + ], + "angle": 0, + "content": "[17] Jing-Ming Guo, Herleeyandi Markoni, and Jiann-Der Lee. Barnet: Boundary aware refinement network for crack detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):7343-7358, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] Yung-An Hsieh and Yichang James Tsai. Machine learning for crack detection: Review and model performance comparison. Journal of Computing in Civil Engineering, 34(5): 04020038, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[20] Achref Jaziri, Martin Mundt, Andres Fernandez, and Visvanathan Ramesh. Designing a hybrid neural system to learn real-world crack segmentation from fractal-based simulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8636-8646, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[21] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[22] Jason Katsamenis, Eftychios Protopapadakis, Nikolaos Bakalos, Andreas Varvarigos, Anastasios Doulamis, Nikolaos Doulamis, and Athanasios Voulodimos. A few-shot attention recurrent residual u-net for crack segmentation. In International Symposium on Visual Computing, pages 199-209. Springer, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Narges Kheradmandi and Vida Mehranfar. A critical review and comparative study on image segmentation-based techniques for pavement crack detection. Construction and Building Materials, 321:126162, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Hong Lang, Ye Yuan, Jiang Chen, Shuo Ding, Jian John Lu, and Yong Zhang. Augmented concrete crack segmentation: Learning complete representation to defend background interference in concrete pavements. IEEE Transactions on Instrumentation and Measurement, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] David Lattanzi and Gregory R Miller. Robust automated concrete damage detection algorithms for field applications." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29414" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Journal of Computing in Civil Engineering, 28(2):253-262, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[26] Qin Lei, Jiang Zhong, and Chen Wang. Joint optimization of crack segmentation with an adaptive dynamic threshold module. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Huaiyuan Li, Hui Li, Chuang Li, Baohai Wu, and Jinghuai Gao. Hybrid swim transformer-cnn model for pore-crack structure identification. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, and Chengjie Wang. Lors: Low-rank residual structure for parameter-efficient network stacking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15866-15876, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[29] Qiufu Li, Xi Jia, Jiancan Zhou, Linlin Shen, and Jinming Duan. Rediscovering bce loss for uniform classification. arXiv preprint arXiv:2403.07289, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.345, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[30] Jianghai Liao, Yuanhao Yue, Dejin Zhang, Wei Tu, Rui Cao, Qin Zou, and Qingquan Li. Automatic tunnel crack inspection using an efficient mobile imaging module and a lightweight cnn. IEEE Transactions on Intelligent Transportation Systems, 23(9):15190-15203, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[31] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3783-3792, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[32] Huajun Liu, Jing Yang, Xiangyu Miao, Christoph Mertz, and Hui Kong. Crackformer network for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 24(9):9240-9252, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[33] Hui Liu, Chen Jia, Fan Shi, Xu Cheng, Mianzhao Wang, and Shengyong Chen. Staircase cascaded fusion of lightweight local pattern recognition and long-range dependencies for structural crack segmentation. arXiv preprint arXiv:2408.12815, 2024. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[34] Wenze Liu, Hao Lu, Hongtao Fu, and Zhiguo Cao. Learning to upsample by learning to sample. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6027-6037, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[35] Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li. Deepcrack: A deep hierarchical feature learning architecture for crack segmentation. Neurocomputing, 338:139-153, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[36] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Liu Mushui, Jun Dan, Ziqian Lu, Yunlong Yu, Yingming Li, and Xi Li. Cm-unet: Hybrid cnn-mamba unet for remote sensing image semantic segmentation. arXiv preprint arXiv:2405.10530, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Mengyang Pu, Yaping Huang, Qingji Guan, and Haibin Ling. Rindnet: Edge detection for discontinuity in reflectance, illumination, normal and depth. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6879-6888, 2021. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[39] Haochen Qi, Xiangwei Kong, Zhibo Jin, Jiqiang Zhang, and Zinan Wang. A vision-transformer-based convex variational network for bridge pavement defect segmentation. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[40] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 23(8):11710-11719, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[41] Jianing Quan, Baozhen Ge, and Min Wang. Crackvit: a unified cnn-transformer model for pixel-level crack extraction. Neural Computing and Applications, 35(15):10957-10973, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[43] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, pages 240–248. Springer, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[44] Huaqi Tao, Bingxi Liu, Jinqiang Cui, and Hong Zhang. A convolutional-transformer network for crack segmentation with boundary awareness. In 2023 IEEE International Conference on Image Processing, pages 86-90. IEEE, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[46] Jin Wang, Zhigao Zeng, Pradip Kumar Sharma, Osama Alfarraj, Amr Tolba, Jianming Zhang, and Lei Wang. Dualpath network combining cnn and transformer for pavement crack segmentation. Automation in Construction, 158: 105217, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[47] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5493-5502, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[48] Chao Xiang, Jingjing Guo, Ran Cao, and Lu Deng. A crack-segmentation algorithm fusing transformers and convolutional neural networks for complex detection scenarios. Automation in Construction, 152:104894, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[49] Xiao Xiao, Shen Lian, Zhiming Luo, and Shaozi Li. Weighted res-unet for high-quality retina vessel segmentation. In 2018 9th international conference on information technology in medicine and education, pages 327-331. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29415" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[51] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[52] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[53] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[54] Tomoyuki Yamaguchi, Shingo Nakamura, Ryo Saegusa, and Shuji Hashimoto. Image-based crack detection for real concrete surfaces. IEEE Transactions on Electrical and Electronic Engineering, 3(1):128-135, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[55] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.499 + ], + "angle": 0, + "content": "[56] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.501, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[57] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[58] Hang Zhang, Allen A Zhang, Zishuo Dong, Anzheng He, Yang Liu, You Zhan, and Kelvin CP Wang. Robust semantic segmentation for automatic crack detection within pavement images using multi-mixing of global context and local image features. IEEE Transactions on Intelligent Transportation Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[59] Tianjie Zhang, Donglei Wang, and Yang Lu. ECSnet: An accelerated real-time image segmentation cnn architecture for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.766 + ], + "angle": 0, + "content": "[60] Xiaohu Zhang and Haifeng Huang. Distilling knowledge from a transformer-based crack segmentation model to a light-weighted symmetry model with mixed loss function for portable crack detection equipment. Symmetry, 16(5):520, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.77, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[61] Jian Zhou, Peisen S Huang, and Fu-Pen Chiang. Wavelet-based pavement distress detection and evaluation. Optical Engineering, 45(2):027007-027007, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.813, + 0.482, + 0.88 + ], + "angle": 0, + "content": "[62] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In International Conference on Machine Learning, 2024. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29416" + } + ] +] \ No newline at end of file diff --git a/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_origin.pdf b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5d898be9b2a3065f13199345f8d6889ad22ab9c3 --- /dev/null +++ b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/2870bb13-5ce1-4a70-8677-d49f0f5cde37_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20594d3f1beeceddc59cd86361b8e444b734d160c580b747ae29b823ee86a9ea +size 2175824 diff --git a/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/full.md b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c443c7f9e52355bb78bb8ce37da2863a545c36c --- /dev/null +++ b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/full.md @@ -0,0 +1,370 @@ +# SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures + +Hui Liu $^{1,2,3}$ , Chen Jia $^{1,2,3,*}$ , Fan Shi $^{1,2,3}$ , Xu Cheng $^{1,2,3}$ , Shengyong Chen $^{1,2,3}$ $^{1}$ School of Computer Science and Engineering, Tianjin University of Technology + $^{2}$ Engineering Research Center of Learning-Based Intelligent System (Ministry of Education) + $^{3}$ Key Laboratory of Computer Vision and System (Ministry of Education) + +liuhui1109@stud.tjut.edu.cn, {jiachen, shifan}@email.tjut.edu.cn, {xu.cheng, sy}@ieee.org + +# Abstract + +Pixel-level segmentation of structural cracks across various scenarios remains a considerable challenge. Current methods encounter challenges in effectively modeling crack morphology and texture, facing challenges in balancing segmentation quality with low computational resource usage. To overcome these limitations, we propose a lightweight Structure-Aware Vision Mamba Network (SCSegamba), capable of generating high-quality pixel-level segmentation maps by leveraging both the morphological information and texture cues of crack pixels with minimal computational cost. Specifically, we developed a Structure-Aware Visual State Space module (SAVSS), which incorporates a lightweight Gated Bottleneck Convolution (GBC) and a Structure-Aware Scanning Strategy (SASS). The key insight of GBC lies in its effectiveness in modeling the morphological information of cracks, while the SASS enhances the perception of crack topology and texture by strengthening the continuity of semantic information between crack pixels. Experiments on crack benchmark datasets demonstrate that our method outperforms other state-of-the-art (SOTA) methods, achieving the highest performance with only 2.8M parameters. On the multi-scenario dataset, our method reached 0.8390 in F1 score and 0.8479 in mIoU. The code is available at https://github.com/Karl1109/SCSegamba. + +# 1. Introduction + +Structures like bitumen pavement, concrete, and metal frequently develop cracks under shear stress, making regular health monitoring essential to avoid production issues [3, 4, 19, 23, 30]. Due to differences in material properties and environmental conditions, various materials exhibit significant variations in crack morphology and visual appearance [24]. As a result, achieving pixel-level crack seg- + +![](images/2f7757cd470313dcf0ac3299ed67e5005cdc17c6eaa42a92a48434e38123b9a1.jpg) +(a) Comparison with SOTA methods. +(b) Different numbers of SAVSS layers. + +![](images/10743ce2eebe51870d9739b4e9086550ce0b2e781a09400b0045c5896855bb95.jpg) +(c) Segmentation results in complex interference conditions. + +![](images/caffed150fd9ea989e2dc81b4fb2817ecf722d1dff63a0fca4889904d9dd34ab.jpg) +Figure 1. Performance of SCSegamba on multi-scenario TUT [33] dataset. (a) Comparison with SOTA methods. (b) Impact of different SAVSS layer numbers on performance, with normalized metrics; FLOPs (G), Params (M), and Size (MB) decrease towards the edges. (c) Visual results under complex interference. + +mentation across diverse scenarios remains a complex challenge. Recently, Convolutional Neural Networks (CNNs), such as ECSNet [59] and SFIAN [5], have shown effective crack feature extraction capabilities in segmentation tasks due to their strong local inductive properties. However, their limited receptive field constrains their ability to model broad-scope irregular dependencies across the entire image, resulting in discontinuous segmentation and weak background noise suppression. Although dilated convolution [2] expand the receptive field, their inherent inductive bias still prevents them from fully addressing this issue [58], especially in complex crack patterns with heavy background interference. + +The success of Vision Transformer (ViT) [10, 45, 47] has demonstrated Transformer's effectiveness in capturing irregular pixel dependencies, which is crucial for recognizing complex crack textures, as seen in networks like DTRCNet [48], MFAFNet [9], and Crackmer [46]. However, the quadratic complexity of attention calculations with sequence length leads to high memory use and training challenges for high-resolution images, limiting deployment on + +resource-constrained edge devices and practical applications. As shown in Figure 1(a), Transformer-based methods like CTCrackseg [44] and DTrCNet [48] perform well, but their large parameter counts and high computational demands limit their deployability on resource-constrained devices. Although variants like Sparse Transformer [6] and Linear Transformer [21] reduce computational requirements by sparsifying or linearizing attention, they sacrifice the ability to model irregular dependencies and pixel textures, hindering pixel-level detection tasks. Consequently, Transformer-based methods struggle to balance segmentation performance with computational efficiency. + +Recently, Selective State Space Models (SSMs) have attracted considerable interest due to Mamba's showing strong performance in sequence modeling while maintaining low computational demands [14, 15]. Vision Mamba (ViM) [62] and VMamba [36] have extended Mamba to the visual domain. The irregular extension of crack regions with numerous branches in low-contrast images, often affected by irrelevant areas and shadows, challenges existing Mamba VSS (Visual State Space Model) block structures and scanning strategies in capturing crack morphology and texture effectively. Most Mamba-based methods [37, 51, 52, 55] process feature maps through linear layers, limiting selective enhancement or suppression of crack features against irrelevant disturbances, thus reducing detailed morphological extraction. Additionally, common parallel or unidirectional diagonal scans [16] struggle to maintain semantic continuity when handling irregular, multi-directional pixel topologies, weakening their ability to manage complex textures and suppress noise. Consequently, current Mamba-based SSM frequently produce false or missed detections in multi-scenario crack images. Moreover, although these methods have the advantage of fewer parameters, there remains potential to further reduce their computational demands and enhance their deployability on edge devices. As shown in Figure 1(b), CSMamba [37], MambaIR [16], and PlainMamba [55] showed unsatisfactory performance on crack images, with room for improvement in parameter count and computational load. + +To tackle the challenge of balancing high segmentation quality with low computational demands, we propose the SCSegamba network that produces high-quality pixel-level crack segmentation maps with low computational resources. To improve the Mamba network's perception of irregular crack textures, we design a Structure-Aware Visual State Space block (SAVSS), employing a Structure-Aware Scanning Strategy (SASS) to enhance semantic continuity and strengthen crack morphology perception. For capturing crack shape cues while maintaining low parameter and computational costs, we designed a lightweight Gated Bottleneck Convolution (GBC) that dynamically adjusts weights for complex backgrounds and varying mor + +phologies. Additionally, the Multi-scale Feature Segmentation head (MFS) integrates the GBC and Multi-layer Perceptron (MLP) to achieve high-quality segmentation maps with low computational requirements. As shown in Figure 1(c), optimal segmentation performance was achieved with four SAVSS layers, producing clear segmentation maps that effectively mask complex interference while maintaining model lightweightness. + +In essence, the primary contributions of our work are outlined as follows: + +- We propose a novel lightweight vision Mamba network, SCSegamba, for crack segmentation. This model effectively captures morphological and irregular texture cues of crack pixels, using low computational resources to generate high-quality segmentation maps. +- We design the SAVSS with a lightweight GBC Convolution and a SASS scanning strategy to enhance the handling and perception of irregular texture cues in crack images. Additionally, a simple yet effective MFS is developed to generate segmentation maps with relatively low computational resources. +- We evaluate SCSegamba on benchmark datasets across diverse scenarios, with results demonstrating that our method outperforms other SOTA methods while maintaining the low parameter count. + +# 2. Related Works + +# 2.1. Crack Segmentation Network + +Early crack detection methods often relied on traditional feature extraction techniques, such as wavelet transform [61], percolation models [54], and the k-means algorithm [25]. Although straightforward to implement, these methods face challenges in suppressing background interference and achieving high segmentation accuracy. With advancements in deep learning, researchers have developed CNN-based crack segmentation networks that achieve SOTA performance [5, 7, 18, 26, 40, 49]. For instance, DeepCrack [35] enables end-to-end pixel-level segmentation, while FPHBN [56] demonstrates strong generalization capabilities. BARNet [17] detects crack boundaries by integrating image gradients with features, and ADDUNet [1] captures both fine and coarse crack characteristics across varied conditions. Although CNN-based methods show significant promise, the local operations and limited receptive fields of CNNs limit their ability to fully capture crack texture cues and effectively suppress background noise. + +Transformers [45], with their self-attention mechanism, are well-suited for visual tasks that require long-range dependency modeling, making them increasingly popular in crack segmentation networks [9, 31, 32, 41, 53]. For example, VCVNet [39], based on Vision Transformer (ViT), is designed for bridge crack segmentation, addressing fine-grained segmentation challenges. SwT-CNN [27] + +![](images/1ff71d1dde6bdf425aceab4c1c2341b32a552eaa35d641fa32b9f38db4a7aa5f.jpg) +(a) Network Architecture of SCSegamba + +![](images/428c266e17810e2d4531dab486d923b5c6f34d4f7dcfc80d742206562c9b6a60.jpg) +(b) Architecture of SAVSS block +Figure 2. Overview of our proposed method. (a) illustrates the overall architecture of SCSegamba and the processing flow for crack images. (b) displays the structure of the SAVSS block. The input crack image undergoes comprehensive morphological and texture feature extraction through SAVSS, while MFS produces a high-quality pixel-level segmentation map. + +combines Swin Transformer and CNN for automatic feature extraction, while TBUNet [60], a Transformer-based knowledge distillation model, achieves high-performance crack segmentation with a hybrid loss function. Although Transformer-based methods are highly effective at capturing crack texture cues and suppressing background noise, their self-attention mechanism introduces computational complexity that grows quadratically with sequence length. This results in a high parameter count and significant computational demands, which limit their deployment on resource-constrained edge devices. + +![](images/59e9419b25be7d1dd892e1d73ff2cc851473ebb1a5f259ecf6e68c1a751f2744.jpg) +Figure 3. Architecture of GBC. It employs bottleneck convolution to efficiently reduce the parameters and computational load, while the gating mechanism enhances the model's adaptability in processing diverse crack patterns and complex backgrounds. GN represents group normalization. + +# 2.2. Selective State Space Model + +The introduction of the Selective State Space Models (S6) in the Mamba model [14] has highlighted the potential of SSM [12, 13]. Unlike the linear time-invariant S4 model, S6 efficiently captures complex long-distance dependencies while preserving computational efficiency, achieving strong performance in NLP, audio, and genomics. Consequently, + +researchers have adapted Mamba to the visual domain, creating various VSS blocks. ViM [62] achieves comparable modeling to ViT [10] without attention mechanisms, using fewer computational resources, while VMamba [36] prioritizes efficient computation and high performance. Plain-Mamba [55] employs a fixed-width layer stacking approach, excelling in tasks such as instance segmentation and object detection. However, the VSS block and scanning strategy require specific optimizations for each visual task, as tasks differ in their reliance on long- and short-distance information, necessitating customized VSS block designs to ensure optimal performance. + +Currently, no high-performing Mamba-based model exists for crack segmentation. Thus, designing an optimized VSS structure specifically for crack segmentation is essential to improve performance and efficiency. Given the intricate details and irregular textures of cracks, the VSS block requires strong shape extraction and directional awareness to effectively capture crack texture cues. Additionally, it should facilitate efficient crack segmentation while minimizing computational resource requirements. + +# 3. Methodology + +# 3.1. Preliminary + +The complete architecture of our proposed SCSegamba is depicted in Figure 2. It includes two main components: the SAVSS for extracting crack shape and texture cues, and the MFS for efficient feature processing. To capture key crack region cues, we integrate the GBC at the initial stage of SAVSS and the final stage of MFS. + +For a single RGB image $E \in \mathbb{R}^{3 \times H \times W}$ , spatial information is divided into $n$ patches, forming a sequence $\{B_1, B_2, \ldots, B_n\}$ . This sequence is processed through the SAVSS block, embedding key crack pixel cues into multiscale feature maps $\{F_1, F_2, F_3, F_4\}$ . Finally, in the MFS, all information is consolidated into a single tensor, produc + +ing a refined segmentation output $W\in \mathbb{R}^{1\times H\times W}$ + +# 3.2. Lightweight Gated Bottleneck Convolution + +The gating mechanism enables dynamic features for each spatial position and channel, enhancing the model's ability to capture details [8, 57]. To further reduce parameter count and computational cost, we embedded a bottleneck convolution (BottConv) with low-rank approximation [28], mapping matrices from high to low dimensional spaces and significantly lowering computational complexity. + +In the convolution layer, assuming the spatial size of the filter is $p$ , the number of input channels is $d$ and the input is $s$ , the convolution response can be represented as: + +$$ +z = Q s + c \tag {1} +$$ + +where $Q$ is a matrix of size $f \times (p^2 \times d)$ , $f$ is the number of output channels, and $c$ is the original bias term. Assuming $z$ lies in a low-rank subspace of rank $f_0$ , it can be represented as $z = V(z - z_1) + z_1$ , where $z_1$ abstracts the mean vector of features, acting as an auxiliary variable to facilitate theoretical derivation and correct feature offsets, $V = LM^T$ ( $L \in \mathbb{R}^{f \times f_0}$ , $M \in \mathbb{R}^{(p^2 d) \times f_0}$ ) represents the low-rank projection matrix. The simplified response then becomes: + +$$ +z = L M ^ {T} s + c ^ {\prime} \tag {2} +$$ + +Since $f_0 < f$ , the computational complexity reduces from $O(fp^2 d)$ to $O(f_0p^2 d) + O(ff_0)$ , where $O(ff_0) \ll O(fp^2 d)$ , indicating that the computational complexity reduction is proportional to the original ratio of $f_0 / f$ . + +In BottConv, pointwise convolutions project features into or out of low-rank subspace, thus significantly reducing complexity, while depthwise convolution that perform spatial information-adequate extraction in subspace guarantees negligibly low complexity. As shown in Figure 5, BottConv in our GBC design significantly reduces parameter count and computational load compared to the original convolution, with minimal performance impact. + +As shown in Figure 3, the input feature $x \in \mathbb{R}^{C \times H \times W}$ is retained as $x_{\mathrm{residual}} = x$ to facilitate the residual connection. Subsequently, the feature $x$ is passed through the BottConv layer, followed by normalization and activation functions, resulting in the features $x_{1}$ and $g_{2}(x)$ as shown below: + +$$ +g _ {1} (x) = \operatorname {R e L U} \left(\operatorname {N o r m} _ {1} \left(f _ {1} (x)\right)\right) \tag {3} +$$ + +$$ +x _ {1} = \operatorname {R e L U} \left(\operatorname {N o r m} _ {2} \left(\operatorname {B o t t C o n v} _ {2} \left(g _ {1} (x)\right)\right)\right) \tag {4} +$$ + +$$ +g _ {2} (x) = \operatorname {R e L U} \left(\operatorname {N o r m} _ {3} \left(\operatorname {B o t t C o n v} _ {3} (x)\right)\right) \tag {5} +$$ + +To generate the gating feature map, $x_{1}$ and $g_{2}(x)$ are combined through the Hadamard product: + +$$ +m (x) = x _ {1} \odot g _ {2} (x) \tag {6} +$$ + +The gating feature map $m(x)$ is subsequently processed through BottConv once again to further refine fine-grained + +details. After the residual connection is applied, the resulting output is: + +$$ +y = \operatorname {R e L U} \left(\operatorname {N o r m} _ {4} \left(\operatorname {B o t t C o n v} _ {4} (m (x)))\right)\right) \tag {7} +$$ + +$$ +O u t p u t = y + x _ {r e s i d u a l} \tag {8} +$$ + +The design of BottConv and deeper gated branch enable the model to preserve basic crack features while dynamically refining the fine-grained feature characterization of the main branch, resulting in more accurate segmentation maps in detailed regions. + +# 3.3. Structure-Aware Visual State Space Module + +Our designed SAVSS features a two-dimensional selective scan (SS2D) tailored for visual tasks. Different scanning strategies impact the model's ability to capture continuous crack textures. As shown in Figure 4, current vision Mamba networks use various scanning directions, including parallel, snake, bidirectional, and diagonal scans [36, 55, 62]. Parallel and diagonal scans lack continuity across rows or diagonals, which limits their sensitivity to crack directions. Although bidirectional and snake scans maintain semantic continuity along horizontal or vertical paths, they struggle to capture diagonal or interwoven textures. To address this, our proposed diagonal snake scanning is designed to better capture complex crack texture cues. + +SASS consists of four paths: two parallel snake paths and two diagonal snake paths. This design enables the effective extraction of continuous semantic information in regular crack regions while preserving texture continuity in multiple directions, making it suitable for multi-scenario crack images with complex backgrounds. + +After the RGB crack image undergoes Patch Embedding and Position Encoding, it is input as a sequence into the SAVSS block. To maintain a lightweight network, we use only 4 layers of SAVSS blocks. The processing equations are as follows: + +$$ +\bar {P} = e ^ {\Delta P} \tag {9} +$$ + +$$ +\bar {Q} = (\Delta P) ^ {- 1} \left(e ^ {\Delta P} - I\right) \cdot \Delta Q \tag {10} +$$ + +$$ +z _ {k} = \bar {P} z _ {k - 1} + \bar {Q} w _ {k} \tag {11} +$$ + +$$ +u _ {k} = R z _ {k} + S w _ {k} \tag {12} +$$ + +In these equations, the input $w \in \mathbb{R}^{t \times D}$ , $P \in \mathbb{R}^{G \times D}$ controls the hidden spatial state, $S \in \mathbb{R}^{D \times D}$ is used to initialize the skip connection for input, $z_{k}$ represents the specific hidden state at time step $k$ , and $Q \in \mathbb{R}^{G \times D}$ and $R \in \mathbb{R}^{G \times D}$ are matrices with hidden spatial dimensions $G$ and temporal dimensions $D$ , respectively, obtained through selective scanning SS2D. These are trainable parameters that are updated accordingly. $u_{k}$ represents the output at time step $k$ . SASS establishes multi-directional adjacency relationships, allowing the hidden state $z_{k}$ to capture more intricate topological and textural details, while enabling the + +![](images/0532b642a2f524458e624dc188b8391866f2c477d57dae05b9d9d7a24f52d8b4.jpg) +Figure 4. Illustration of our proposed SASS and other scanning strategies. The first row presents four commonly used single scanning paths, along with our proposed diagonal snake path. The second row illustrates the execution flow of our proposed SASS scanning strategy. + +![](images/df70699a78c38b72c085edb127b7778cac2d1f80aeb41ed1e5ebc64dec2abe67.jpg) +Figure 5. Performance comparison between using BottConv and raw convolution in GBC on the TUT [33] dataset. + +output $u_{k}$ to more effectively integrate multi-directional features. + +To effectively combine the initial sequence $x$ with the sequence processed through SS2D, we incorporate Pixel Attention-oriented Fusion (PAF) [33], enhancing SAVSS's ability to capture crack shape and texture details. Following selective scanning, a residual connection is applied to the fused information to preserve detail and facilitate feature flow. Furthermore, GBC refines the inter-layer output within SAVSS, strengthening crack information extraction and boosting performance in later stages. + +# 3.4. Multi-scale Feature Segmentation Head + +Unlike convolutional layers, the MLP swiftly learns the mapping relationships between features and labels, thereby reducing model complexity. When the four feature maps $F_{1}, F_{2}, F_{3}, F_{4} \in \mathbb{R}^{C \times H \times W}$ produced by SAVSS are fed into MFS, they undergo individual processing through the efficient MLP operation and dynamic upsampling [34], restoring their resolution to the original size and yielding $F_{1}^{\mathrm{up}}$ , $F_{2}^{\mathrm{up}}$ , $F_{3}^{\mathrm{up}}$ , $F_{4}^{\mathrm{up}} \in \mathbb{R}^{C \times H \times W}$ . The formula is as follows: + +$$ +F _ {i} ^ {u p} = D y S a m p l e \left(M L P _ {i} \left(F _ {i}\right)\right) \tag {13} +$$ + +where $i$ denotes the layer index. + +To integrate all multi-scale crack shape and texture representations, these feature maps are aggregated into a single tensor, obtaining a high-quality crack segmentation map + +$o\in \mathbb{R}^{1\times H\times W}$ as follows: + +$$ +o _ {1} = G B C \left(\text {C o n c a t} \left(F _ {1} ^ {u p}, F _ {2} ^ {u p}, F _ {3} ^ {u p}, F _ {4} ^ {u p}\right)\right) \tag {14} +$$ + +$$ +o = M L P \left(\operatorname {C o n v} \left(o _ {1}\right)\right) \tag {15} +$$ + +# 3.5. Objective Function + +We use a blend of Binary Cross-Entropy loss (BCE) [29] and Dice loss [43] as the objective function, which helps improve the network's robustness to imbalanced pixel data. The overall loss function is expressed as follows: + +$$ +L = \alpha \cdot L _ {D i c e} + \beta \cdot L _ {B C E} \tag {16} +$$ + +where the hyperparameters $\alpha$ and $\beta$ control the weights of the two loss components. The ratio of $\alpha$ to $\beta$ is set to 1:5. + +# 4. Experiments + +# 4.1. Datasets + +Crack500 [56]. The images in this dataset were captured using a mobile phone. The original dataset consists of 500 bitumen crack images, which were expanded to 3368 images through data augmentation. Each image is paired with a corresponding pixel-level annotated binary image. + +DeepCrack [35]. The dataset comprises 537 RGB images of cement, bricks and bitumen cracks under various conditions, including fine, wide, stained, and fuzzy cracks, ensuring diversity and representativeness. + +**CrackMap [22].** The dataset was created for a road crack segmentation study and consists of 120 high-resolution RGB images capturing a variety of thin and complex bitumen road cracks. + +TUT [33]. In contrast to other datasets with simple backgrounds, this dataset includes a dense, cluttered background and features cracks with elaborate, intricate shapes. It contains 1408 RGB images across eight scenarios: bitumen, cement, bricks, runway, tiles, metal, blades, and pipes. + +![](images/85a1d14f2f2da22f91a78b0e6f4bc308559fee5ca8eae5ab61884450ae63b721.jpg) +Figure 6. Visual comparison of typical cracks with 9 SOTA methods across four datasets. Red boxes highlight critical details, and green boxes mark misidentified regions. + +During processing, all datasets were divided into training, validation, and test sets with a 7:1:2 ratio. + +# 4.2. Implementation Details. + +Experimental Settings. We built our SCSegamba network using PyTorch v1.13.1 and trained it on an Intel Xeon Platinum 8336C CPU with eight Nvidia GeForce RTX 4090 GPUs. The AdamW optimizer was used with an initial learning rate of 5e-4, PolyLR scheduling, a weight decay of 0.01, and a random seed of 42. The network was trained for 50 epochs, and the model with the best validation performance was selected for testing. + +Comparison Methods. To comprehensively evaluate our model, we compared SCSegamba with 9 SOTA methods. The CNN or Transformer-based models included RIND [38], SFIAN [5], CTCrackSeg [44], DTrCNet [48], Crackmer [46] and SimCrack [20]. Additionally, we compared it with other Mamba-based models, including CSMamba [37], PlainMamba [55], and MambaIR [16]. + +Evaluation Metrics. We used six metrics to evaluate SCSegamba's performance: Precision (P), Recall (R), F1 Score $(F1 = \frac{2RP}{R + P})$ , Optimal Dataset Scale (ODS), Optimal Image Scale (OIS), and mean Intersection over Union (mIoU). ODS measures the model's adaptability to datasets of varying scales at a fixed threshold m, while OIS evaluates adaptability across image scales at an optimal threshold n. The calculation formulas are as follows: + +$$ +O D S = \max _ {m} \frac {2 \cdot P _ {m} \cdot R _ {m}}{P _ {m} + R _ {m}} \tag {17} +$$ + +$$ +O I S = \frac {1}{N} \sum_ {i = 1} ^ {N} \max _ {n} \frac {2 \cdot P _ {n , i} \cdot R _ {n , i}}{P _ {n , i} + R _ {n , i}} \tag {18} +$$ + +mIoU is used to measure the mean proportion of the intersection over union between the ground truth and the predicted results. The calculation is given by the formula: + +$$ +m I o U = \frac {1}{N + 1} \sum_ {l = 0} ^ {N} \frac {p u l}{\sum_ {t = 0} ^ {N} p _ {t l} + \sum_ {t = 0} ^ {N} p _ {t l} - p _ {u l}} \tag {19} +$$ + +where $N$ is the number of classes, which we set as $N = 1$ ; $t$ represents the ground truth, $l$ represents the predicted value, and $p_{tl}$ represents the count of pixels classified as $l$ but belonging to $t$ . + +Additionally, we evaluated our method's complexity using three metrics: FLOPs, Params, and Model Size, representing computational complexity, parameter complexity, and memory footprint. + +# 4.3. Comparison with SOTA Methods + +As listed in Table 1, compared with 9 other SOTA methods, our proposed SCSegamba achieves the best performance across four public datasets. Specifically, on the Crack500 [56] and DeepCrack [35] datasets, which contain larger and more complex crack regions, SCSegamba achieved the highest performance. Notably, on the DeepCrack dataset, it surpassed the next best method by $1.50\%$ in F1 score and $1.09\%$ in mIoU. This improvement is due to the robust ability of GBC to capture morphological clues in large crack areas, enhancing the model's representational power. On the CrackMap [22] dataset, which features thinner and more elongated cracks, our method surpasses all other SOTA methods in every metric, outperforming the next best method by $2.06\%$ in F1 and $1.65\%$ in mIoU. This demonstrates the effectiveness of SASS in capturing fine textures and elongated crack structures. As illustrated in + +
MethodsCrack500DeepCrack
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.64690.64830.69980.72450.71190.73810.80870.82670.78960.89200.83770.8391
SFIAN [5]0.69770.73480.69830.77420.73430.76040.86160.89280.85490.86920.86200.8776
CTCrackSeg [44]0.69410.70590.69400.77480.73220.75910.88190.89040.90110.88950.89520.8925
DTrCNet [48]0.70120.72410.65270.82800.73570.76270.84730.85120.89050.82510.85660.8661
Crackmer [46]0.69330.70970.69850.75720.72670.75910.87120.87850.89460.87830.88640.8844
SimCrack [20]0.71270.73080.70930.79840.75160.77150.85700.87220.89840.85490.87610.8744
CSMamba [37]0.69310.71620.68580.78230.73150.75920.87380.87660.90250.88630.89430.8863
PlainMamba [55]0.70350.71730.71700.75570.73580.76820.86460.86680.90500.86590.88500.8788
MambaIR [16]0.70430.71890.72040.76810.74350.76630.87960.88400.90560.88950.89750.8907
SCSegamba (Ours)0.72440.73700.72700.78590.75530.77780.89380.89900.90970.91240.91100.9022
MethodsCrackMapTUT
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.67450.69430.60230.75860.66990.74250.75310.78910.78720.76650.77670.8051
SFIAN [5]0.72000.74650.67150.76680.71600.77480.72900.75130.77150.73670.75370.7896
CTCrackSeg [44]0.72890.73730.69110.76690.72700.77850.79400.79960.82020.81950.81990.8301
DTrCNet [48]0.73280.74130.69120.76810.72760.78120.79870.80730.79720.84410.82020.8331
Crackmer [46]0.73950.74370.72290.74670.73460.78600.74290.76400.75010.76560.75780.7966
SimCrack [20]0.75590.76250.73800.76720.75230.79630.79840.80900.80510.83710.82080.8334
CSMamba [37]0.73710.74130.70530.76630.73460.78410.78790.79460.79470.83530.81460.8263
PlainMamba [55]0.71500.71890.66490.76160.70990.76990.78670.79670.77010.85230.81020.8253
MambaIR [16]0.73320.73470.75690.70130.72800.78340.78610.79300.78770.83870.81250.8249
SCSegamba (Ours)0.77410.77660.76290.77270.76780.80940.82040.82550.82410.85450.83900.8479
+ +Table 1. Comparison with 9 SOTA methods across 4 datasets. Best results are in bold, and second-best results are underlined. + +
MethodsYearFLOPs↓Params↓Size↓
RIND [38]2021695.77G59.39M453MB
SFIAN [5]202384.57G13.63M56MB
CTCrackSeg [44]202339.47G22.88M174MB
DTrCNet [48]2023123.20G63.45M317MB
Crackmer [46]202414.94G5.90M43MB
SimCrack [20]2024286.62G29.58M225MB
CSMamba [37]2024145.84G35.95M233MB
PlainMamba [55]202473.36G16.72M96MB
MambaIR [16]202447.32G10.34M79MB
SCSegamba (Ours)202418.16G2.80M37MB
+ +Table 2. Comparison of complexity with other methods. Best results are in bold, and second-best results are underlined. + +Figure 6, our method produces clearer and more precise feature maps, with superior detail capture in typical scenarios such as cement and bitumen, compared to other methods. + +For the TUT dataset [33], which includes eight diverse scenarios, our method achieved the best performance, surpassing the next best method by $2.21\%$ in F1 and $1.74\%$ in mIoU. As shown in Figure 6, whether in the complex crack topology of plastic tracks, the noise-heavy backgrounds of metallic materials and turbine blades, or the low-contrast, dimly lit underground pipeline images, SCSegamba consistently produced high-quality segmentation maps while effectively suppressing irrelevant noise. This demonstrates + +that our method, with the enhanced crack morphology and texture perception from GBC and SASS, exhibits exceptional robustness and stability. Additionally, leveraging MFS for feature aggregation improves multi-scale perception, making our model particularly suited for diverse, interference-rich scenarios. + +# 4.4. Complexity Analysis + +Table 2 shows a comparison of the complexity of our method with other SOTA methods when the input image size is uniformly set to 512. With only 2.80M parameters and a model size of 37MB, our method surpasses all others, being $52.54\%$ and $13.95\%$ lower than the next best result, respectively. Additionally, compared to Crackmer [46], which prioritizes computational efficiency, our method's FLOPs are only 3.22G higher. This demonstrates that the combination of lightweight SAVSS and MFS enables high-quality segmentation in noisy crack scenes with minimal parameters and low computational load, which is essential for resource-constrained devices. + +# 4.5. Ablation Studies + +We performed ablation experiments on the representative multi-scenario dataset TUT [33]. + +Ablation study of segmentation heads. As listed in Table 3, with our designed MFS, SCSegamba achieved the best results across all six metrics, with F1 and mIoU scores $1.57\%$ and $1.21\%$ higher than the second-best method. In + +
Seg HeadODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
UNet [42]0.80550.81510.81480.83760.82600.83782.92M19.27G39MB
Ham [11]0.77030.77840.79620.78380.79090.81242.86M35.08G38MB
SegFormer [50]0.79470.79830.81700.81740.81720.83072.79M17.87G35MB
MFS0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
+ +Table 3. Ablation study of different segmentation heads. UNet [42], Ham [11], and SegFormer [50] are high-performance heads. + +
GBCPAFResODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
✓XX0.81360.81960.82130.84610.83350.84342.49M16.75G34MB
X✓X0.79980.80840.79180.85240.82220.83432.28M14.91G33MB
XX✓0.79360.80690.79520.84380.81970.83132.48M15.65G35MB
✓✓X0.80470.81020.81740.83790.82750.83772.54M17.08G35MB
✓X✓0.81160.82000.81560.85220.83340.84252.75M17.82G37MB
X✓✓0.80230.80760.82190.83020.82600.83602.54M15.99G35MB
✓✓✓0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
+ +terms of complexity, although Params, FLOPs, and Model Size are only 0.01M, 0.29G, and 2MB larger than the Seg-Former head, our method surpasses it in F1 and mIoU by $2.67\%$ and $2.07\%$ , respectively. This demonstrates that MFS enhances SAVSS output integration, significantly improving performance while keeping the model lightweight. + +Table 4. Ablation study of components within the SAVSS block. Best results are in bold, and second-best results are underlined. + +
ScanODSOISPRF1mIoU
Parallel0.81230.81840.81460.85230.83300.8427
Diag0.80910.81480.82250.84170.83200.8410
ParaSna0.81020.81620.82190.83650.82910.8408
DiagSna0.81530.82150.82370.84970.83650.8451
SASS0.82040.82550.82410.85450.83900.8479
+ +Table 5. Ablation studies with different four-route scanning strategies in the SAVSS block, comparing parallel, diagonal, parallel snake, and diagonal snake scanning. The impact of different scanning strategies on complexity is negligible; thus, complexity analysis is omitted from this table. Best results are in bold, and second-best results are underlined. + +Ablation study of components. Table 4 shows the impact of each component in SAVSS on model performance. When fully utilizing GBC, PAF, and residual connections, our model achieved the best results across all metrics. Notably, adding GBC led to significant improvements in F1 and mIoU by $1.57\%$ and $1.42\%$ , respectively, highlighting its strength in capturing crack morphology cues. Similarly, residual connections boosted F1 and mIoU by $0.13\%$ and $2.47\%$ , indicating their role in focusing on essential crack features. Although using only PAF resulted in the lowest Params, FLOPs, and Model Size, it significantly reduced performance. These findings demonstrate that our fully integrated SAVSS effectively captures crack morphology and texture cues, achieving top pixel-level segmentation results while maintaining a lightweight model. + +Ablation studies of scanning strategies. As listed in Table 5, under the same conditions of using four different directional scanning paths, the model achieved the best performance with our designed SASS scanning strategy, improving F1 and mIoU by $0.30\%$ and $0.33\%$ over the diagonal snake strategy. This demonstrates SASS's ability to construct semantic and dependency information suited to crack topology, enhancing crack pixel perception in subsequent modules. More comprehensive experiments and real-world deployments are available in the Appendix. + +# 5. Conclusion + +In this paper, we proposed SCSegamba, a lightweight structure-aware Vision Mamba for precise pixel-level crack segmentation. SCSegamba combines SAVSS and MFS to enhance crack shape and texture perception with a low parameter count. Equipped with GBC and SASS scanning, SAVSS, captures irregular crack textures across various structures. Experiments on four datasets show SCSegamba's exceptional performance, especially in complex, noisy scenarios. On the challenging multi-scenario dataset, it achieved an F1 score of 0.8390 and mIoU of 0.8479 with only 18.16G FLOPs and 2.8M parameters, demonstrating its effectiveness for real-world crack detection and suitability for edge devices. Future work will incorporate multimodal cues to enhance segmentation quality, while further optimizing VSS design and scan strategies to achieve high-quality results with low computational resources. + +# 6. Acknowledgement + +This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62272342, 62020106004, 62306212, and T2422015; the Tianjin Natural Science Foundation under Grants 23JCJQJC00070 and 24PTLYHZ00320; and the Marie Skłodowska-Curie Actions (MSCA) under Project No. 101111188. + +# References + +[1] Zaid Al-Huda, Bo Peng, Riyadh Nazar Ali Algburi, Mugahed A Al-antari, AL-Jarazi Rabea, Omar Al-maqtari, and Donghai Zhai. Asymmetric dual-decoder-u-net for pavement crack semantic segmentation. Automation in Construction, 156:105138, 2023. 2 +[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision, pages 801-818, 2018. 1 +[3] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4703-4712, 2022. 1 +[4] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 1 +[5] Xu Cheng, Tian He, Fan Shi, Meng Zhao, Xiufeng Liu, and Shengyong Chen. Selective feature fusion and irregular-aware network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1, 2, 6, 7 +[6] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2 +[7] Wooram Choi and Young-Jin Cha. Sddnet: Real-time crack segmentation. IEEE Transactions on Industrial Electronics, 67(9):8016-8025, 2019. 2 +[8] Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In International conference on machine learning, pages 933-941. PMLR, 2017. 4 +[9] Jiaxiu Dong, Niannian Wang, Hongyuan Fang, Wentong Guo, Bin Li, and Kejie Zhai. Mfafnet: An innovative crack intelligent segmentation method based on multi-layer feature association fusion network. Advanced Engineering Informatics, 62:102584, 2024. 1, 2 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations, 2021. 1, 3 +[11] Zhengyang Geng, Meng-Hao Guo, Hongxu Chen, Xia Li, Ke Wei, and Zhouchen Lin. Is attention better than matrix decomposition? In International Conference on Learning Representations, 2021. 8 +[12] Karan Goel, Albert Gu, Chris Donahue, and Christopher Ré. It's raw! audio generation with state-space models. In International Conference on Machine Learning, pages 7616-7633. PMLR, 2022. 3 + +[13] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3 +[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2, 3 +[15] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2 +[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2024. 2, 6, 7 +[17] Jing-Ming Guo, Herleeyandi Markoni, and Jiann-Der Lee. Barnet: Boundary aware refinement network for crack detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):7343-7358, 2021. 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2 +[19] Yung-An Hsieh and Yichang James Tsai. Machine learning for crack detection: Review and model performance comparison. Journal of Computing in Civil Engineering, 34(5): 04020038, 2020. 1 +[20] Achref Jaziri, Martin Mundt, Andres Fernandez, and Visvanathan Ramesh. Designing a hybrid neural system to learn real-world crack segmentation from fractal-based simulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8636-8646, 2024. 6, 7 +[21] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 2 +[22] Jason Katsamenis, Eftychios Protopapadakis, Nikolaos Bakalos, Andreas Varvarigos, Anastasios Doulamis, Nikolaos Doulamis, and Athanasios Voulodimos. A few-shot attention recurrent residual u-net for crack segmentation. In International Symposium on Visual Computing, pages 199-209. Springer, 2023. 5, 6 +[23] Narges Kheradmandi and Vida Mehranfar. A critical review and comparative study on image segmentation-based techniques for pavement crack detection. Construction and Building Materials, 321:126162, 2022. 1 +[24] Hong Lang, Ye Yuan, Jiang Chen, Shuo Ding, Jian John Lu, and Yong Zhang. Augmented concrete crack segmentation: Learning complete representation to defend background interference in concrete pavements. IEEE Transactions on Instrumentation and Measurement, 2024. 1 +[25] David Lattanzi and Gregory R Miller. Robust automated concrete damage detection algorithms for field applications. + +Journal of Computing in Civil Engineering, 28(2):253-262, 2014. 2 +[26] Qin Lei, Jiang Zhong, and Chen Wang. Joint optimization of crack segmentation with an adaptive dynamic threshold module. IEEE Transactions on Intelligent Transportation Systems, 2024. 2 +[27] Huaiyuan Li, Hui Li, Chuang Li, Baohai Wu, and Jinghuai Gao. Hybrid swim transformer-cnn model for pore-crack structure identification. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2 +[28] Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, and Chengjie Wang. Lors: Low-rank residual structure for parameter-efficient network stacking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15866-15876, 2024. 4 +[29] Qiufu Li, Xi Jia, Jiancan Zhou, Linlin Shen, and Jinming Duan. Rediscovering bce loss for uniform classification. arXiv preprint arXiv:2403.07289, 2024. 5 +[30] Jianghai Liao, Yuanhao Yue, Dejin Zhang, Wei Tu, Rui Cao, Qin Zou, and Qingquan Li. Automatic tunnel crack inspection using an efficient mobile imaging module and a lightweight cnn. IEEE Transactions on Intelligent Transportation Systems, 23(9):15190-15203, 2022. 1 +[31] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3783-3792, 2021. 2 +[32] Huajun Liu, Jing Yang, Xiangyu Miao, Christoph Mertz, and Hui Kong. Crackformer network for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 24(9):9240-9252, 2023. 2 +[33] Hui Liu, Chen Jia, Fan Shi, Xu Cheng, Mianzhao Wang, and Shengyong Chen. Staircase cascaded fusion of lightweight local pattern recognition and long-range dependencies for structural crack segmentation. arXiv preprint arXiv:2408.12815, 2024. 1, 5, 7 +[34] Wenze Liu, Hao Lu, Hongtao Fu, and Zhiguo Cao. Learning to upsample by learning to sample. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6027-6037, 2023. 5 +[35] Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li. Deepcrack: A deep hierarchical feature learning architecture for crack segmentation. Neurocomputing, 338:139-153, 2019. 2, 5, 6 +[36] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4 +[37] Liu Mushui, Jun Dan, Ziqian Lu, Yunlong Yu, Yingming Li, and Xi Li. Cm-unet: Hybrid cnn-mamba unet for remote sensing image semantic segmentation. arXiv preprint arXiv:2405.10530, 2024. 2, 6, 7 +[38] Mengyang Pu, Yaping Huang, Qingji Guan, and Haibin Ling. Rindnet: Edge detection for discontinuity in reflectance, illumination, normal and depth. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6879-6888, 2021. 6, 7 + +[39] Haochen Qi, Xiangwei Kong, Zhibo Jin, Jiqiang Zhang, and Zinan Wang. A vision-transformer-based convex variational network for bridge pavement defect segmentation. IEEE Transactions on Intelligent Transportation Systems, 2024. 2 +[40] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 23(8):11710-11719, 2021. 2 +[41] Jianing Quan, Baozhen Ge, and Min Wang. Crackvit: a unified cnn-transformer model for pixel-level crack extraction. Neural Computing and Applications, 35(15):10957-10973, 2023. 2 +[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 8 +[43] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, pages 240–248. Springer, 2017. 5 +[44] Huaqi Tao, Bingxi Liu, Jinqiang Cui, and Hong Zhang. A convolutional-transformer network for crack segmentation with boundary awareness. In 2023 IEEE International Conference on Image Processing, pages 86-90. IEEE, 2023. 2, 6, 7 +[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2 +[46] Jin Wang, Zhigao Zeng, Pradip Kumar Sharma, Osama Alfarraj, Amr Tolba, Jianming Zhang, and Lei Wang. Dualpath network combining cnn and transformer for pavement crack segmentation. Automation in Construction, 158: 105217, 2024. 1, 6, 7 +[47] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5493-5502, 2024. 1 +[48] Chao Xiang, Jingjing Guo, Ran Cao, and Lu Deng. A crack-segmentation algorithm fusing transformers and convolutional neural networks for complex detection scenarios. Automation in Construction, 152:104894, 2023. 1, 2, 6, 7 +[49] Xiao Xiao, Shen Lian, Zhiming Luo, and Shaozi Li. Weighted res-unet for high-quality retina vessel segmentation. In 2018 9th international conference on information technology in medicine and education, pages 327-331. IEEE, 2018. 2 +[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8 + +[51] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2 +[52] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2 +[53] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2 +[54] Tomoyuki Yamaguchi, Shingo Nakamura, Ryo Saegusa, and Shuji Hashimoto. Image-based crack detection for real concrete surfaces. IEEE Transactions on Electrical and Electronic Engineering, 3(1):128-135, 2008. 2 +[55] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 2, 3, 4, 6, 7 +[56] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 2, 5, 6 +[57] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 4 +[58] Hang Zhang, Allen A Zhang, Zishuo Dong, Anzheng He, Yang Liu, You Zhan, and Kelvin CP Wang. Robust semantic segmentation for automatic crack detection within pavement images using multi-mixing of global context and local image features. IEEE Transactions on Intelligent Transportation Systems, 2024. 1 +[59] Tianjie Zhang, Donglei Wang, and Yang Lu. ECSnet: An accelerated real-time image segmentation cnn architecture for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1 +[60] Xiaohu Zhang and Haifeng Huang. Distilling knowledge from a transformer-based crack segmentation model to a light-weighted symmetry model with mixed loss function for portable crack detection equipment. Symmetry, 16(5):520, 2024. 3 +[61] Jian Zhou, Peisen S Huang, and Fu-Pen Chiang. Wavelet-based pavement distress detection and evaluation. Optical Engineering, 45(2):027007-027007, 2006. 2 +[62] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In International Conference on Machine Learning, 2024. 2, 3, 4 \ No newline at end of file diff --git a/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/images.zip b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e03fcf1da12242c027f2bc5e77001ab7e90511e --- /dev/null +++ b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bcd349a198c5dc8736ee72adef3a12b790dca6b2e5caea4def0f05cbd442823 +size 964703 diff --git a/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/layout.json b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a1f0bae1434c3e4a59781ded43e0ea974f64aac --- /dev/null +++ b/2025/SCSegamba_ Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures/layout.json @@ -0,0 +1,9288 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 102, + 517, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 102, + 517, + 138 + ], + "spans": [ + { + "bbox": [ + 93, + 102, + 517, + 138 + ], + "type": "text", + "content": "SCSegamba: Lightweight Structure-Aware Vision Mamba for Crack Segmentation in Structures" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "spans": [ + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": "Hui Liu" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": ", Chen Jia" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": ", Fan Shi" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": ", Xu Cheng" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": ", Shengyong Chen" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Tianjin University of Technology \n" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": "Engineering Research Center of Learning-Based Intelligent System (Ministry of Education) \n" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 160, + 528, + 217 + ], + "type": "text", + "content": "Key Laboratory of Computer Vision and System (Ministry of Education)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 219, + 541, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 541, + 232 + ], + "type": "text", + "content": "liuhui1109@stud.tjut.edu.cn, {jiachen, shifan}@email.tjut.edu.cn, {xu.cheng, sy}@ieee.org" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 583 + ], + "type": "text", + "content": "Pixel-level segmentation of structural cracks across various scenarios remains a considerable challenge. Current methods encounter challenges in effectively modeling crack morphology and texture, facing challenges in balancing segmentation quality with low computational resource usage. To overcome these limitations, we propose a lightweight Structure-Aware Vision Mamba Network (SCSegamba), capable of generating high-quality pixel-level segmentation maps by leveraging both the morphological information and texture cues of crack pixels with minimal computational cost. Specifically, we developed a Structure-Aware Visual State Space module (SAVSS), which incorporates a lightweight Gated Bottleneck Convolution (GBC) and a Structure-Aware Scanning Strategy (SASS). The key insight of GBC lies in its effectiveness in modeling the morphological information of cracks, while the SASS enhances the perception of crack topology and texture by strengthening the continuity of semantic information between crack pixels. Experiments on crack benchmark datasets demonstrate that our method outperforms other state-of-the-art (SOTA) methods, achieving the highest performance with only 2.8M parameters. On the multi-scenario dataset, our method reached 0.8390 in F1 score and 0.8479 in mIoU. The code is available at https://github.com/Karl1109/SCSegamba." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "Structures like bitumen pavement, concrete, and metal frequently develop cracks under shear stress, making regular health monitoring essential to avoid production issues [3, 4, 19, 23, 30]. Due to differences in material properties and environmental conditions, various materials exhibit significant variations in crack morphology and visual appearance [24]. As a result, achieving pixel-level crack seg-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 258, + 556, + 340 + ], + "blocks": [ + { + "bbox": [ + 315, + 258, + 556, + 340 + ], + "lines": [ + { + "bbox": [ + 315, + 258, + 556, + 340 + ], + "spans": [ + { + "bbox": [ + 315, + 258, + 556, + 340 + ], + "type": "image", + "image_path": "2f7757cd470313dcf0ac3299ed67e5005cdc17c6eaa42a92a48434e38123b9a1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 342, + 432, + 350 + ], + "lines": [ + { + "bbox": [ + 323, + 342, + 432, + 350 + ], + "spans": [ + { + "bbox": [ + 323, + 342, + 432, + 350 + ], + "type": "text", + "content": "(a) Comparison with SOTA methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 342, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 438, + 342, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 438, + 342, + 555, + 350 + ], + "type": "text", + "content": "(b) Different numbers of SAVSS layers." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 352, + 431, + 376 + ], + "blocks": [ + { + "bbox": [ + 329, + 352, + 431, + 376 + ], + "lines": [ + { + "bbox": [ + 329, + 352, + 431, + 376 + ], + "spans": [ + { + "bbox": [ + 329, + 352, + 431, + 376 + ], + "type": "image", + "image_path": "10743ce2eebe51870d9739b4e9086550ce0b2e781a09400b0045c5896855bb95.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 377, + 520, + 386 + ], + "lines": [ + { + "bbox": [ + 345, + 377, + 520, + 386 + ], + "spans": [ + { + "bbox": [ + 345, + 377, + 520, + 386 + ], + "type": "text", + "content": "(c) Segmentation results in complex interference conditions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 438, + 352, + 537, + 376 + ], + "blocks": [ + { + "bbox": [ + 438, + 352, + 537, + 376 + ], + "lines": [ + { + "bbox": [ + 438, + 352, + 537, + 376 + ], + "spans": [ + { + "bbox": [ + 438, + 352, + 537, + 376 + ], + "type": "image", + "image_path": "caffed150fd9ea989e2dc81b4fb2817ecf722d1dff63a0fca4889904d9dd34ab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": "Figure 1. Performance of SCSegamba on multi-scenario TUT [33] dataset. (a) Comparison with SOTA methods. (b) Impact of different SAVSS layer numbers on performance, with normalized metrics; FLOPs (G), Params (M), and Size (MB) decrease towards the edges. (c) Visual results under complex interference." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": "mentation across diverse scenarios remains a complex challenge. Recently, Convolutional Neural Networks (CNNs), such as ECSNet [59] and SFIAN [5], have shown effective crack feature extraction capabilities in segmentation tasks due to their strong local inductive properties. However, their limited receptive field constrains their ability to model broad-scope irregular dependencies across the entire image, resulting in discontinuous segmentation and weak background noise suppression. Although dilated convolution [2] expand the receptive field, their inherent inductive bias still prevents them from fully addressing this issue [58], especially in complex crack patterns with heavy background interference." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "The success of Vision Transformer (ViT) [10, 45, 47] has demonstrated Transformer's effectiveness in capturing irregular pixel dependencies, which is crucial for recognizing complex crack textures, as seen in networks like DTRCNet [48], MFAFNet [9], and Crackmer [46]. However, the quadratic complexity of attention calculations with sequence length leads to high memory use and training challenges for high-resolution images, limiting deployment on" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29406" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "resource-constrained edge devices and practical applications. As shown in Figure 1(a), Transformer-based methods like CTCrackseg [44] and DTrCNet [48] perform well, but their large parameter counts and high computational demands limit their deployability on resource-constrained devices. Although variants like Sparse Transformer [6] and Linear Transformer [21] reduce computational requirements by sparsifying or linearizing attention, they sacrifice the ability to model irregular dependencies and pixel textures, hindering pixel-level detection tasks. Consequently, Transformer-based methods struggle to balance segmentation performance with computational efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 219, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 554 + ], + "type": "text", + "content": "Recently, Selective State Space Models (SSMs) have attracted considerable interest due to Mamba's showing strong performance in sequence modeling while maintaining low computational demands [14, 15]. Vision Mamba (ViM) [62] and VMamba [36] have extended Mamba to the visual domain. The irregular extension of crack regions with numerous branches in low-contrast images, often affected by irrelevant areas and shadows, challenges existing Mamba VSS (Visual State Space Model) block structures and scanning strategies in capturing crack morphology and texture effectively. Most Mamba-based methods [37, 51, 52, 55] process feature maps through linear layers, limiting selective enhancement or suppression of crack features against irrelevant disturbances, thus reducing detailed morphological extraction. Additionally, common parallel or unidirectional diagonal scans [16] struggle to maintain semantic continuity when handling irregular, multi-directional pixel topologies, weakening their ability to manage complex textures and suppress noise. Consequently, current Mamba-based SSM frequently produce false or missed detections in multi-scenario crack images. Moreover, although these methods have the advantage of fewer parameters, there remains potential to further reduce their computational demands and enhance their deployability on edge devices. As shown in Figure 1(b), CSMamba [37], MambaIR [16], and PlainMamba [55] showed unsatisfactory performance on crack images, with room for improvement in parameter count and computational load." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "To tackle the challenge of balancing high segmentation quality with low computational demands, we propose the SCSegamba network that produces high-quality pixel-level crack segmentation maps with low computational resources. To improve the Mamba network's perception of irregular crack textures, we design a Structure-Aware Visual State Space block (SAVSS), employing a Structure-Aware Scanning Strategy (SASS) to enhance semantic continuity and strengthen crack morphology perception. For capturing crack shape cues while maintaining low parameter and computational costs, we designed a lightweight Gated Bottleneck Convolution (GBC) that dynamically adjusts weights for complex backgrounds and varying mor" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "phologies. Additionally, the Multi-scale Feature Segmentation head (MFS) integrates the GBC and Multi-layer Perceptron (MLP) to achieve high-quality segmentation maps with low computational requirements. As shown in Figure 1(c), optimal segmentation performance was achieved with four SAVSS layers, producing clear segmentation maps that effectively mask complex interference while maintaining model lightweightness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 168, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 190 + ], + "type": "text", + "content": "In essence, the primary contributions of our work are outlined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 194, + 553, + 373 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 194, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 194, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 314, + 194, + 553, + 253 + ], + "type": "text", + "content": "- We propose a novel lightweight vision Mamba network, SCSegamba, for crack segmentation. This model effectively captures morphological and irregular texture cues of crack pixels, using low computational resources to generate high-quality segmentation maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 253, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 253, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 553, + 324 + ], + "type": "text", + "content": "- We design the SAVSS with a lightweight GBC Convolution and a SASS scanning strategy to enhance the handling and perception of irregular texture cues in crack images. Additionally, a simple yet effective MFS is developed to generate segmentation maps with relatively low computational resources." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 326, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 553, + 373 + ], + "type": "text", + "content": "- We evaluate SCSegamba on benchmark datasets across diverse scenarios, with results demonstrating that our method outperforms other SOTA methods while maintaining the low parameter count." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 380, + 405, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 405, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 405, + 392 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 397, + 476, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 476, + 409 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 476, + 409 + ], + "type": "text", + "content": "2.1. Crack Segmentation Network" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 629 + ], + "type": "text", + "content": "Early crack detection methods often relied on traditional feature extraction techniques, such as wavelet transform [61], percolation models [54], and the k-means algorithm [25]. Although straightforward to implement, these methods face challenges in suppressing background interference and achieving high segmentation accuracy. With advancements in deep learning, researchers have developed CNN-based crack segmentation networks that achieve SOTA performance [5, 7, 18, 26, 40, 49]. For instance, DeepCrack [35] enables end-to-end pixel-level segmentation, while FPHBN [56] demonstrates strong generalization capabilities. BARNet [17] detects crack boundaries by integrating image gradients with features, and ADDUNet [1] captures both fine and coarse crack characteristics across varied conditions. Although CNN-based methods show significant promise, the local operations and limited receptive fields of CNNs limit their ability to fully capture crack texture cues and effectively suppress background noise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Transformers [45], with their self-attention mechanism, are well-suited for visual tasks that require long-range dependency modeling, making them increasingly popular in crack segmentation networks [9, 31, 32, 41, 53]. For example, VCVNet [39], based on Vision Transformer (ViT), is designed for bridge crack segmentation, addressing fine-grained segmentation challenges. SwT-CNN [27]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29407" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 76, + 352, + 213 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 352, + 213 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 352, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 352, + 213 + ], + "type": "image", + "image_path": "1ff71d1dde6bdf425aceab4c1c2341b32a552eaa35d641fa32b9f38db4a7aa5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 216, + 254, + 225 + ], + "lines": [ + { + "bbox": [ + 127, + 216, + 254, + 225 + ], + "spans": [ + { + "bbox": [ + 127, + 216, + 254, + 225 + ], + "type": "text", + "content": "(a) Network Architecture of SCSegamba" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 356, + 76, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 356, + 76, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 356, + 76, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 356, + 76, + 545, + 213 + ], + "type": "image", + "image_path": "428c266e17810e2d4531dab486d923b5c6f34d4f7dcfc80d742206562c9b6a60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 216, + 512, + 224 + ], + "lines": [ + { + "bbox": [ + 410, + 216, + 512, + 224 + ], + "spans": [ + { + "bbox": [ + 410, + 216, + 512, + 224 + ], + "type": "text", + "content": "(b) Architecture of SAVSS block" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 235, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 555, + 269 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method. (a) illustrates the overall architecture of SCSegamba and the processing flow for crack images. (b) displays the structure of the SAVSS block. The input crack image undergoes comprehensive morphological and texture feature extraction through SAVSS, while MFS produces a high-quality pixel-level segmentation map." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 407 + ], + "type": "text", + "content": "combines Swin Transformer and CNN for automatic feature extraction, while TBUNet [60], a Transformer-based knowledge distillation model, achieves high-performance crack segmentation with a hybrid loss function. Although Transformer-based methods are highly effective at capturing crack texture cues and suppressing background noise, their self-attention mechanism introduces computational complexity that grows quadratically with sequence length. This results in a high parameter count and significant computational demands, which limit their deployment on resource-constrained edge devices." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 416, + 282, + 555 + ], + "blocks": [ + { + "bbox": [ + 70, + 416, + 282, + 555 + ], + "lines": [ + { + "bbox": [ + 70, + 416, + 282, + 555 + ], + "spans": [ + { + "bbox": [ + 70, + 416, + 282, + 555 + ], + "type": "image", + "image_path": "59e9419b25be7d1dd892e1d73ff2cc851473ebb1a5f259ecf6e68c1a751f2744.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 568, + 295, + 624 + ], + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 624 + ], + "type": "text", + "content": "Figure 3. Architecture of GBC. It employs bottleneck convolution to efficiently reduce the parameters and computational load, while the gating mechanism enhances the model's adaptability in processing diverse crack patterns and complex backgrounds. GN represents group normalization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 208, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 208, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 208, + 636 + ], + "type": "text", + "content": "2.2. Selective State Space Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "The introduction of the Selective State Space Models (S6) in the Mamba model [14] has highlighted the potential of SSM [12, 13]. Unlike the linear time-invariant S4 model, S6 efficiently captures complex long-distance dependencies while preserving computational efficiency, achieving strong performance in NLP, audio, and genomics. Consequently," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 275, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 418 + ], + "type": "text", + "content": "researchers have adapted Mamba to the visual domain, creating various VSS blocks. ViM [62] achieves comparable modeling to ViT [10] without attention mechanisms, using fewer computational resources, while VMamba [36] prioritizes efficient computation and high performance. Plain-Mamba [55] employs a fixed-width layer stacking approach, excelling in tasks such as instance segmentation and object detection. However, the VSS block and scanning strategy require specific optimizations for each visual task, as tasks differ in their reliance on long- and short-distance information, necessitating customized VSS block designs to ensure optimal performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 419, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 556, + 528 + ], + "type": "text", + "content": "Currently, no high-performing Mamba-based model exists for crack segmentation. Thus, designing an optimized VSS structure specifically for crack segmentation is essential to improve performance and efficiency. Given the intricate details and irregular textures of cracks, the VSS block requires strong shape extraction and directional awareness to effectively capture crack texture cues. Additionally, it should facilitate efficient crack segmentation while minimizing computational resource requirements." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 533, + 397, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 397, + 546 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 397, + 546 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 393, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 393, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 393, + 565 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 640 + ], + "type": "text", + "content": "The complete architecture of our proposed SCSegamba is depicted in Figure 2. It includes two main components: the SAVSS for extracting crack shape and texture cues, and the MFS for efficient feature processing. To capture key crack region cues, we integrate the GBC at the initial stage of SAVSS and the final stage of MFS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "For a single RGB image " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": ", spatial information is divided into " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " patches, forming a sequence " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{B_1, B_2, \\ldots, B_n\\}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": ". This sequence is processed through the SAVSS block, embedding key crack pixel cues into multiscale feature maps " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{F_1, F_2, F_3, F_4\\}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": ". Finally, in the MFS, all information is consolidated into a single tensor, produc" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29408" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 260, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 260, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 260, + 84 + ], + "type": "text", + "content": "ing a refined segmentation output " + }, + { + "bbox": [ + 55, + 71, + 260, + 84 + ], + "type": "inline_equation", + "content": "W\\in \\mathbb{R}^{1\\times H\\times W}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 279, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 279, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 279, + 104 + ], + "type": "text", + "content": "3.2. Lightweight Gated Bottleneck Convolution" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 193 + ], + "type": "text", + "content": "The gating mechanism enables dynamic features for each spatial position and channel, enhancing the model's ability to capture details [8, 57]. To further reduce parameter count and computational cost, we embedded a bottleneck convolution (BottConv) with low-rank approximation [28], mapping matrices from high to low dimensional spaces and significantly lowering computational complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "text", + "content": "In the convolution layer, assuming the spatial size of the filter is " + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "text", + "content": ", the number of input channels is " + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "text", + "content": " and the input is " + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 193, + 296, + 228 + ], + "type": "text", + "content": ", the convolution response can be represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 150, + 233, + 294, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 233, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 150, + 233, + 294, + 245 + ], + "type": "interline_equation", + "content": "z = Q s + c \\tag {1}", + "image_path": "5cb7b7f1bfaf1b45c529b08d8119abd7e738d95b4d32aac3deba658650866b40.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": " is a matrix of size " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "f \\times (p^2 \\times d)" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": " is the number of output channels, and " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": " is the original bias term. Assuming " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": " lies in a low-rank subspace of rank " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "f_0" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": ", it can be represented as " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "z = V(z - z_1) + z_1" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "z_1" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": " abstracts the mean vector of features, acting as an auxiliary variable to facilitate theoretical derivation and correct feature offsets, " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "V = LM^T" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{f \\times f_0}" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{(p^2 d) \\times f_0}" + }, + { + "bbox": [ + 55, + 255, + 296, + 351 + ], + "type": "text", + "content": ") represents the low-rank projection matrix. The simplified response then becomes:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 357, + 294, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 357, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 140, + 357, + 294, + 369 + ], + "type": "interline_equation", + "content": "z = L M ^ {T} s + c ^ {\\prime} \\tag {2}", + "image_path": "6304617a8a4577d1ca5b223e49f0b72cc1a2b1b85978ba5a2f37dde4c1c00ba9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "inline_equation", + "content": "f_0 < f" + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "content": ", the computational complexity reduces from " + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "inline_equation", + "content": "O(fp^2 d)" + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "inline_equation", + "content": "O(f_0p^2 d) + O(ff_0)" + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "inline_equation", + "content": "O(ff_0) \\ll O(fp^2 d)" + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "content": ", indicating that the computational complexity reduction is proportional to the original ratio of " + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "inline_equation", + "content": "f_0 / f" + }, + { + "bbox": [ + 55, + 380, + 296, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 429, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 522 + ], + "type": "text", + "content": "In BottConv, pointwise convolutions project features into or out of low-rank subspace, thus significantly reducing complexity, while depthwise convolution that perform spatial information-adequate extraction in subspace guarantees negligibly low complexity. As shown in Figure 5, BottConv in our GBC design significantly reduces parameter count and computational load compared to the original convolution, with minimal performance impact." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": "As shown in Figure 3, the input feature " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": " is retained as " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{residual}} = x" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": " to facilitate the residual connection. Subsequently, the feature " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": " is passed through the BottConv layer, followed by normalization and activation functions, resulting in the features " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "inline_equation", + "content": "g_{2}(x)" + }, + { + "bbox": [ + 55, + 522, + 296, + 582 + ], + "type": "text", + "content": " as shown below:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 584, + 294, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 584, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 107, + 584, + 294, + 597 + ], + "type": "interline_equation", + "content": "g _ {1} (x) = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {1} \\left(f _ {1} (x)\\right)\\right) \\tag {3}", + "image_path": "7c299e5ed0bcb4d445e5d89a0b05eb0baa2c7d3a6e49e04075ad38087fc4b75e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 602, + 294, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 602, + 294, + 615 + ], + "spans": [ + { + "bbox": [ + 85, + 602, + 294, + 615 + ], + "type": "interline_equation", + "content": "x _ {1} = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {2} \\left(\\operatorname {B o t t C o n v} _ {2} \\left(g _ {1} (x)\\right)\\right)\\right) \\tag {4}", + "image_path": "4e697a0f17028c00fd898031331eac562daf02df296f7474a28c12af13f6c121.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 88, + 621, + 294, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 621, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 88, + 621, + 294, + 634 + ], + "type": "interline_equation", + "content": "g _ {2} (x) = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {3} \\left(\\operatorname {B o t t C o n v} _ {3} (x)\\right)\\right) \\tag {5}", + "image_path": "88fb3f50a0712e485f0ad49fabcc5300f1efc0e383dffb91ffbe2f3bb30aad0a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "type": "text", + "content": "To generate the gating feature map, " + }, + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "type": "inline_equation", + "content": "g_{2}(x)" + }, + { + "bbox": [ + 55, + 639, + 296, + 663 + ], + "type": "text", + "content": " are combined through the Hadamard product:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 668, + 294, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 668, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 133, + 668, + 294, + 680 + ], + "type": "interline_equation", + "content": "m (x) = x _ {1} \\odot g _ {2} (x) \\tag {6}", + "image_path": "98b1b7ce3dbed9cd0f7c52ddbb3ca5faf2e256d79c668eaab6a77e6a08643556.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The gating feature map " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "m(x)" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is subsequently processed through BottConv once again to further refine fine-grained" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "details. After the residual connection is applied, the resulting output is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 99, + 553, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 99, + 553, + 112 + ], + "spans": [ + { + "bbox": [ + 347, + 99, + 553, + 112 + ], + "type": "interline_equation", + "content": "y = \\operatorname {R e L U} \\left(\\operatorname {N o r m} _ {4} \\left(\\operatorname {B o t t C o n v} _ {4} (m (x)))\\right)\\right) \\tag {7}", + "image_path": "c93e281f0996f03f02fb2c7de50f056a34f62f9c7980339044f27b19b54b04a3.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 118, + 553, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 118, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 383, + 118, + 553, + 130 + ], + "type": "interline_equation", + "content": "O u t p u t = y + x _ {r e s i d u a l} \\tag {8}", + "image_path": "10d4b3f4e4218ca7e68a4f61be2390ce52be20ab450581da9efe6fd27ef5df89.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 136, + 554, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 554, + 196 + ], + "type": "text", + "content": "The design of BottConv and deeper gated branch enable the model to preserve basic crack features while dynamically refining the fine-grained feature characterization of the main branch, resulting in more accurate segmentation maps in detailed regions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 203, + 542, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 542, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 542, + 216 + ], + "type": "text", + "content": "3.3. Structure-Aware Visual State Space Module" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "content": "Our designed SAVSS features a two-dimensional selective scan (SS2D) tailored for visual tasks. Different scanning strategies impact the model's ability to capture continuous crack textures. As shown in Figure 4, current vision Mamba networks use various scanning directions, including parallel, snake, bidirectional, and diagonal scans [36, 55, 62]. Parallel and diagonal scans lack continuity across rows or diagonals, which limits their sensitivity to crack directions. Although bidirectional and snake scans maintain semantic continuity along horizontal or vertical paths, they struggle to capture diagonal or interwoven textures. To address this, our proposed diagonal snake scanning is designed to better capture complex crack texture cues." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 376, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 448 + ], + "type": "text", + "content": "SASS consists of four paths: two parallel snake paths and two diagonal snake paths. This design enables the effective extraction of continuous semantic information in regular crack regions while preserving texture continuity in multiple directions, making it suitable for multi-scenario crack images with complex backgrounds." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 448, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 506 + ], + "type": "text", + "content": "After the RGB crack image undergoes Patch Embedding and Position Encoding, it is input as a sequence into the SAVSS block. To maintain a lightweight network, we use only 4 layers of SAVSS blocks. The processing equations are as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 413, + 506, + 553, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 506, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 413, + 506, + 553, + 520 + ], + "type": "interline_equation", + "content": "\\bar {P} = e ^ {\\Delta P} \\tag {9}", + "image_path": "db7787be8b3493bf5bd74095e4082c31ce8c0d4bfb6dad6f3b4fa4de56894db2.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 371, + 525, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 525, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 371, + 525, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\bar {Q} = (\\Delta P) ^ {- 1} \\left(e ^ {\\Delta P} - I\\right) \\cdot \\Delta Q \\tag {10}", + "image_path": "7f1bc6f8f988a2e3dab8bcdc202a157ec628034ee5f7c4960ba90986b07019ad.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 392, + 544, + 553, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 544, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 392, + 544, + 553, + 558 + ], + "type": "interline_equation", + "content": "z _ {k} = \\bar {P} z _ {k - 1} + \\bar {Q} w _ {k} \\tag {11}", + "image_path": "e3add92db3546c7d0c9316c5783074f41ba103ede9f901d8282670f4b289b921.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 397, + 564, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 564, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 397, + 564, + 553, + 576 + ], + "type": "interline_equation", + "content": "u _ {k} = R z _ {k} + S w _ {k} \\tag {12}", + "image_path": "4853c2de5d23602522e672916b17acd67ff7178bf0661e76fd57006824c5a32d.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": "In these equations, the input " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{R}^{t \\times D}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{G \\times D}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " controls the hidden spatial state, " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " is used to initialize the skip connection for input, " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " represents the specific hidden state at time step " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{G \\times D}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{G \\times D}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " are matrices with hidden spatial dimensions " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " and temporal dimensions " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ", respectively, obtained through selective scanning SS2D. These are trainable parameters that are updated accordingly. " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "u_{k}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " represents the output at time step " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": ". SASS establishes multi-directional adjacency relationships, allowing the hidden state " + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": " to capture more intricate topological and textural details, while enabling the" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29409" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 528, + 235 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 528, + 235 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 528, + 235 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 528, + 235 + ], + "type": "image", + "image_path": "0532b642a2f524458e624dc188b8391866f2c477d57dae05b9d9d7a24f52d8b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "type": "text", + "content": "Figure 4. Illustration of our proposed SASS and other scanning strategies. The first row presents four commonly used single scanning paths, along with our proposed diagonal snake path. The second row illustrates the execution flow of our proposed SASS scanning strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 276, + 296, + 355 + ], + "blocks": [ + { + "bbox": [ + 56, + 276, + 296, + 355 + ], + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 355 + ], + "type": "image", + "image_path": "df70699a78c38b72c085edb127b7778cac2d1f80aeb41ed1e5ebc64dec2abe67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "type": "text", + "content": "Figure 5. Performance comparison between using BottConv and raw convolution in GBC on the TUT [33] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "type": "text", + "content": "output " + }, + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "type": "inline_equation", + "content": "u_{k}" + }, + { + "bbox": [ + 55, + 388, + 295, + 411 + ], + "type": "text", + "content": " to more effectively integrate multi-directional features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": "To effectively combine the initial sequence " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " with the sequence processed through SS2D, we incorporate Pixel Attention-oriented Fusion (PAF) [33], enhancing SAVSS's ability to capture crack shape and texture details. Following selective scanning, a residual connection is applied to the fused information to preserve detail and facilitate feature flow. Furthermore, GBC refines the inter-layer output within SAVSS, strengthening crack information extraction and boosting performance in later stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 263, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 263, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 263, + 540 + ], + "type": "text", + "content": "3.4. Multi-scale Feature Segmentation Head" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "content": "Unlike convolutional layers, the MLP swiftly learns the mapping relationships between features and labels, thereby reducing model complexity. When the four feature maps " + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "inline_equation", + "content": "F_{1}, F_{2}, F_{3}, F_{4} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "content": " produced by SAVSS are fed into MFS, they undergo individual processing through the efficient MLP operation and dynamic upsampling [34], restoring their resolution to the original size and yielding " + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "inline_equation", + "content": "F_{1}^{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "inline_equation", + "content": "F_{2}^{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "inline_equation", + "content": "F_{3}^{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "inline_equation", + "content": "F_{4}^{\\mathrm{up}} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 544, + 295, + 643 + ], + "type": "text", + "content": ". The formula is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 647, + 295, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 647, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 108, + 647, + 295, + 661 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {u p} = D y S a m p l e \\left(M L P _ {i} \\left(F _ {i}\\right)\\right) \\tag {13}", + "image_path": "f8c47064ea2813b93f21ed106531364a5050bfe8e414aefaa4aec620580186c7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 186, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 186, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 186, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 186, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 666, + 186, + 677 + ], + "type": "text", + "content": " denotes the layer index." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To integrate all multi-scale crack shape and texture representations, these feature maps are aggregated into a single tensor, obtaining a high-quality crack segmentation map" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 276, + 421, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 421, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 421, + 289 + ], + "type": "inline_equation", + "content": "o\\in \\mathbb{R}^{1\\times H\\times W}" + }, + { + "bbox": [ + 314, + 276, + 421, + 289 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 299, + 555, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 299, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 335, + 299, + 555, + 313 + ], + "type": "interline_equation", + "content": "o _ {1} = G B C \\left(\\text {C o n c a t} \\left(F _ {1} ^ {u p}, F _ {2} ^ {u p}, F _ {3} ^ {u p}, F _ {4} ^ {u p}\\right)\\right) \\tag {14}", + "image_path": "8c9af0ddf72eb5845c34e523ab03b1a203fb2992b012c1704daeb29fe3f3796b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 323, + 554, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 323, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 386, + 323, + 554, + 336 + ], + "type": "interline_equation", + "content": "o = M L P \\left(\\operatorname {C o n v} \\left(o _ {1}\\right)\\right) \\tag {15}", + "image_path": "c2ae577f30470e625d45c9004bf228d483998f9f847428fd22d43620c5bdf3fd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 342, + 426, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 426, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 426, + 356 + ], + "type": "text", + "content": "3.5. Objective Function" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 360, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 408 + ], + "type": "text", + "content": "We use a blend of Binary Cross-Entropy loss (BCE) [29] and Dice loss [43] as the objective function, which helps improve the network's robustness to imbalanced pixel data. The overall loss function is expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 412, + 553, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 412, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 378, + 412, + 553, + 425 + ], + "type": "interline_equation", + "content": "L = \\alpha \\cdot L _ {D i c e} + \\beta \\cdot L _ {B C E} \\tag {16}", + "image_path": "205c35c9db38bccde5f670849e00f69bf5941ef60c17f356e37c77c1a9431a7a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "content": "where the hyperparameters " + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "content": " control the weights of the two loss components. The ratio of " + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "content": " is set to 1:5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 395, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 395, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 395, + 474 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 480, + 376, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 376, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 376, + 491 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "content": "Crack500 [56]. The images in this dataset were captured using a mobile phone. The original dataset consists of 500 bitumen crack images, which were expanded to 3368 images through data augmentation. Each image is paired with a corresponding pixel-level annotated binary image." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": "DeepCrack [35]. The dataset comprises 537 RGB images of cement, bricks and bitumen cracks under various conditions, including fine, wide, stained, and fuzzy cracks, ensuring diversity and representativeness." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 652 + ], + "type": "text", + "content": "**CrackMap [22].** The dataset was created for a road crack segmentation study and consists of 120 high-resolution RGB images capturing a variety of thin and complex bitumen road cracks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "TUT [33]. In contrast to other datasets with simple backgrounds, this dataset includes a dense, cluttered background and features cracks with elaborate, intricate shapes. It contains 1408 RGB images across eight scenarios: bitumen, cement, bricks, runway, tiles, metal, blades, and pipes." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29410" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 553, + 277 + ], + "type": "image", + "image_path": "85a1d14f2f2da22f91a78b0e6f4bc308559fee5ca8eae5ab61884450ae63b721.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "type": "text", + "content": "Figure 6. Visual comparison of typical cracks with 9 SOTA methods across four datasets. Red boxes highlight critical details, and green boxes mark misidentified regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 339 + ], + "type": "text", + "content": "During processing, all datasets were divided into training, validation, and test sets with a 7:1:2 ratio." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 191, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 191, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 191, + 363 + ], + "type": "text", + "content": "4.2. Implementation Details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 461 + ], + "type": "text", + "content": "Experimental Settings. We built our SCSegamba network using PyTorch v1.13.1 and trained it on an Intel Xeon Platinum 8336C CPU with eight Nvidia GeForce RTX 4090 GPUs. The AdamW optimizer was used with an initial learning rate of 5e-4, PolyLR scheduling, a weight decay of 0.01, and a random seed of 42. The network was trained for 50 epochs, and the model with the best validation performance was selected for testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 545 + ], + "type": "text", + "content": "Comparison Methods. To comprehensively evaluate our model, we compared SCSegamba with 9 SOTA methods. The CNN or Transformer-based models included RIND [38], SFIAN [5], CTCrackSeg [44], DTrCNet [48], Crackmer [46] and SimCrack [20]. Additionally, we compared it with other Mamba-based models, including CSMamba [37], PlainMamba [55], and MambaIR [16]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 547, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 296, + 642 + ], + "type": "text", + "content": "Evaluation Metrics. We used six metrics to evaluate SCSegamba's performance: Precision (P), Recall (R), F1 Score " + }, + { + "bbox": [ + 55, + 547, + 296, + 642 + ], + "type": "inline_equation", + "content": "(F1 = \\frac{2RP}{R + P})" + }, + { + "bbox": [ + 55, + 547, + 296, + 642 + ], + "type": "text", + "content": ", Optimal Dataset Scale (ODS), Optimal Image Scale (OIS), and mean Intersection over Union (mIoU). ODS measures the model's adaptability to datasets of varying scales at a fixed threshold m, while OIS evaluates adaptability across image scales at an optimal threshold n. The calculation formulas are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 651, + 295, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 651, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 119, + 651, + 295, + 677 + ], + "type": "interline_equation", + "content": "O D S = \\max _ {m} \\frac {2 \\cdot P _ {m} \\cdot R _ {m}}{P _ {m} + R _ {m}} \\tag {17}", + "image_path": "45b8cca22077f12873a6e0b1073d635359145366e3ce36ad4c474ae761706aa7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 104, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "O I S = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\max _ {n} \\frac {2 \\cdot P _ {n , i} \\cdot R _ {n , i}}{P _ {n , i} + R _ {n , i}} \\tag {18}", + "image_path": "df08a32e2d5cb6ab9c6dd18390aea9cebf9d9ed880019c3ac5455774940897b5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 316, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 352 + ], + "type": "text", + "content": "mIoU is used to measure the mean proportion of the intersection over union between the ground truth and the predicted results. The calculation is given by the formula:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 361, + 555, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 361, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 325, + 361, + 555, + 395 + ], + "type": "interline_equation", + "content": "m I o U = \\frac {1}{N + 1} \\sum_ {l = 0} ^ {N} \\frac {p u l}{\\sum_ {t = 0} ^ {N} p _ {t l} + \\sum_ {t = 0} ^ {N} p _ {t l} - p _ {u l}} \\tag {19}", + "image_path": "64aa6d270b78dbb915e00658413f09a55b7e81ce5289a03d18d55996e78e1006.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": " is the number of classes, which we set as " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": " represents the ground truth, " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": " represents the predicted value, and " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "p_{tl}" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": " represents the count of pixels classified as " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": " but belonging to " + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 403, + 555, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 500 + ], + "type": "text", + "content": "Additionally, we evaluated our method's complexity using three metrics: FLOPs, Params, and Model Size, representing computational complexity, parameter complexity, and memory footprint." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 493, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 493, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 493, + 520 + ], + "type": "text", + "content": "4.3. Comparison with SOTA Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "As listed in Table 1, compared with 9 other SOTA methods, our proposed SCSegamba achieves the best performance across four public datasets. Specifically, on the Crack500 [56] and DeepCrack [35] datasets, which contain larger and more complex crack regions, SCSegamba achieved the highest performance. Notably, on the DeepCrack dataset, it surpassed the next best method by " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.50\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " in F1 score and " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.09\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " in mIoU. This improvement is due to the robust ability of GBC to capture morphological clues in large crack areas, enhancing the model's representational power. On the CrackMap [22] dataset, which features thinner and more elongated cracks, our method surpasses all other SOTA methods in every metric, outperforming the next best method by " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.06\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " in F1 and " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.65\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " in mIoU. This demonstrates the effectiveness of SASS in capturing fine textures and elongated crack structures. As illustrated in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29411" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 72, + 556, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 556, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 556, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 556, + 365 + ], + "type": "table", + "html": "
MethodsCrack500DeepCrack
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.64690.64830.69980.72450.71190.73810.80870.82670.78960.89200.83770.8391
SFIAN [5]0.69770.73480.69830.77420.73430.76040.86160.89280.85490.86920.86200.8776
CTCrackSeg [44]0.69410.70590.69400.77480.73220.75910.88190.89040.90110.88950.89520.8925
DTrCNet [48]0.70120.72410.65270.82800.73570.76270.84730.85120.89050.82510.85660.8661
Crackmer [46]0.69330.70970.69850.75720.72670.75910.87120.87850.89460.87830.88640.8844
SimCrack [20]0.71270.73080.70930.79840.75160.77150.85700.87220.89840.85490.87610.8744
CSMamba [37]0.69310.71620.68580.78230.73150.75920.87380.87660.90250.88630.89430.8863
PlainMamba [55]0.70350.71730.71700.75570.73580.76820.86460.86680.90500.86590.88500.8788
MambaIR [16]0.70430.71890.72040.76810.74350.76630.87960.88400.90560.88950.89750.8907
SCSegamba (Ours)0.72440.73700.72700.78590.75530.77780.89380.89900.90970.91240.91100.9022
MethodsCrackMapTUT
ODSOISPRF1mIoUODSOISPRF1mIoU
RIND [38]0.67450.69430.60230.75860.66990.74250.75310.78910.78720.76650.77670.8051
SFIAN [5]0.72000.74650.67150.76680.71600.77480.72900.75130.77150.73670.75370.7896
CTCrackSeg [44]0.72890.73730.69110.76690.72700.77850.79400.79960.82020.81950.81990.8301
DTrCNet [48]0.73280.74130.69120.76810.72760.78120.79870.80730.79720.84410.82020.8331
Crackmer [46]0.73950.74370.72290.74670.73460.78600.74290.76400.75010.76560.75780.7966
SimCrack [20]0.75590.76250.73800.76720.75230.79630.79840.80900.80510.83710.82080.8334
CSMamba [37]0.73710.74130.70530.76630.73460.78410.78790.79460.79470.83530.81460.8263
PlainMamba [55]0.71500.71890.66490.76160.70990.76990.78670.79670.77010.85230.81020.8253
MambaIR [16]0.73320.73470.75690.70130.72800.78340.78610.79300.78770.83870.81250.8249
SCSegamba (Ours)0.77410.77660.76290.77270.76780.80940.82040.82550.82410.85450.83900.8479
", + "image_path": "3d318c16d3c9ff4fd57fad47b0fd78edd7c927dd95777463599ae524028ec9c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 394, + 294, + 533 + ], + "blocks": [ + { + "bbox": [ + 78, + 374, + 531, + 385 + ], + "lines": [ + { + "bbox": [ + 78, + 374, + 531, + 385 + ], + "spans": [ + { + "bbox": [ + 78, + 374, + 531, + 385 + ], + "type": "text", + "content": "Table 1. Comparison with 9 SOTA methods across 4 datasets. Best results are in bold, and second-best results are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 394, + 294, + 533 + ], + "lines": [ + { + "bbox": [ + 57, + 394, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 394, + 294, + 533 + ], + "type": "table", + "html": "
MethodsYearFLOPs↓Params↓Size↓
RIND [38]2021695.77G59.39M453MB
SFIAN [5]202384.57G13.63M56MB
CTCrackSeg [44]202339.47G22.88M174MB
DTrCNet [48]2023123.20G63.45M317MB
Crackmer [46]202414.94G5.90M43MB
SimCrack [20]2024286.62G29.58M225MB
CSMamba [37]2024145.84G35.95M233MB
PlainMamba [55]202473.36G16.72M96MB
MambaIR [16]202447.32G10.34M79MB
SCSegamba (Ours)202418.16G2.80M37MB
", + "image_path": "107c194521607489837debbaf34564e439809c1cb730b5f2f3fa8f0fc37d12e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 564 + ], + "type": "text", + "content": "Table 2. Comparison of complexity with other methods. Best results are in bold, and second-best results are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 568, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 604 + ], + "type": "text", + "content": "Figure 6, our method produces clearer and more precise feature maps, with superior detail capture in typical scenarios such as cement and bitumen, compared to other methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "For the TUT dataset [33], which includes eight diverse scenarios, our method achieved the best performance, surpassing the next best method by " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "2.21\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " in F1 and " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.74\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " in mIoU. As shown in Figure 6, whether in the complex crack topology of plastic tracks, the noise-heavy backgrounds of metallic materials and turbine blades, or the low-contrast, dimly lit underground pipeline images, SCSegamba consistently produced high-quality segmentation maps while effectively suppressing irrelevant noise. This demonstrates" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 396, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 467 + ], + "type": "text", + "content": "that our method, with the enhanced crack morphology and texture perception from GBC and SASS, exhibits exceptional robustness and stability. Additionally, leveraging MFS for feature aggregation improves multi-scale perception, making our model particularly suited for diverse, interference-rich scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 471, + 434, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 434, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 434, + 483 + ], + "type": "text", + "content": "4.4. Complexity Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "content": "Table 2 shows a comparison of the complexity of our method with other SOTA methods when the input image size is uniformly set to 512. With only 2.80M parameters and a model size of 37MB, our method surpasses all others, being " + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "inline_equation", + "content": "52.54\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "inline_equation", + "content": "13.95\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "content": " lower than the next best result, respectively. Additionally, compared to Crackmer [46], which prioritizes computational efficiency, our method's FLOPs are only 3.22G higher. This demonstrates that the combination of lightweight SAVSS and MFS enables high-quality segmentation in noisy crack scenes with minimal parameters and low computational load, which is essential for resource-constrained devices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 628, + 416, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 416, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 416, + 639 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": "We performed ablation experiments on the representative multi-scenario dataset TUT [33]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "text", + "content": "Ablation study of segmentation heads. As listed in Table 3, with our designed MFS, SCSegamba achieved the best results across all six metrics, with F1 and mIoU scores " + }, + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "inline_equation", + "content": "1.57\\%" + }, + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "text", + "content": " higher than the second-best method. In" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29412" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 543, + 139 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 543, + 139 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 543, + 139 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 543, + 139 + ], + "type": "table", + "html": "
Seg HeadODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
UNet [42]0.80550.81510.81480.83760.82600.83782.92M19.27G39MB
Ham [11]0.77030.77840.79620.78380.79090.81242.86M35.08G38MB
SegFormer [50]0.79470.79830.81700.81740.81720.83072.79M17.87G35MB
MFS0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
", + "image_path": "fbc443af4a42095e11efae0256e5c2a748368d301fac0888a80ed0b53e485522.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 163, + 549, + 268 + ], + "blocks": [ + { + "bbox": [ + 72, + 147, + 536, + 158 + ], + "lines": [ + { + "bbox": [ + 72, + 147, + 536, + 158 + ], + "spans": [ + { + "bbox": [ + 72, + 147, + 536, + 158 + ], + "type": "text", + "content": "Table 3. Ablation study of different segmentation heads. UNet [42], Ham [11], and SegFormer [50] are high-performance heads." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 163, + 549, + 268 + ], + "lines": [ + { + "bbox": [ + 61, + 163, + 549, + 268 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 549, + 268 + ], + "type": "table", + "html": "
GBCPAFResODSOISPRF1mIoUParams ↓FLOPs ↓Model Size ↓
✓XX0.81360.81960.82130.84610.83350.84342.49M16.75G34MB
X✓X0.79980.80840.79180.85240.82220.83432.28M14.91G33MB
XX✓0.79360.80690.79520.84380.81970.83132.48M15.65G35MB
✓✓X0.80470.81020.81740.83790.82750.83772.54M17.08G35MB
✓X✓0.81160.82000.81560.85220.83340.84252.75M17.82G37MB
X✓✓0.80230.80760.82190.83020.82600.83602.54M15.99G35MB
✓✓✓0.82040.82550.82410.85450.83900.84792.80M18.16G37MB
", + "image_path": "b92bd76e46e98ccf89a69a26aa92362b07614fd8437cda923486a7e8666f3f5c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "text", + "content": "terms of complexity, although Params, FLOPs, and Model Size are only 0.01M, 0.29G, and 2MB larger than the Seg-Former head, our method surpasses it in F1 and mIoU by " + }, + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "inline_equation", + "content": "2.67\\%" + }, + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "text", + "content": ", respectively. This demonstrates that MFS enhances SAVSS output integration, significantly improving performance while keeping the model lightweight." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 369, + 296, + 450 + ], + "blocks": [ + { + "bbox": [ + 73, + 276, + 536, + 287 + ], + "lines": [ + { + "bbox": [ + 73, + 276, + 536, + 287 + ], + "spans": [ + { + "bbox": [ + 73, + 276, + 536, + 287 + ], + "type": "text", + "content": "Table 4. Ablation study of components within the SAVSS block. Best results are in bold, and second-best results are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 369, + 296, + 450 + ], + "lines": [ + { + "bbox": [ + 56, + 369, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 296, + 450 + ], + "type": "table", + "html": "
ScanODSOISPRF1mIoU
Parallel0.81230.81840.81460.85230.83300.8427
Diag0.80910.81480.82250.84170.83200.8410
ParaSna0.81020.81620.82190.83650.82910.8408
DiagSna0.81530.82150.82370.84970.83650.8451
SASS0.82040.82550.82410.85450.83900.8479
", + "image_path": "03b094cd3e557221dd85e4322d0d5acc68935cc1b4362671fcf9edb2574eb9e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 458, + 295, + 524 + ], + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 524 + ], + "type": "text", + "content": "Table 5. Ablation studies with different four-route scanning strategies in the SAVSS block, comparing parallel, diagonal, parallel snake, and diagonal snake scanning. The impact of different scanning strategies on complexity is negligible; thus, complexity analysis is omitted from this table. Best results are in bold, and second-best results are underlined." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Ablation study of components. Table 4 shows the impact of each component in SAVSS on model performance. When fully utilizing GBC, PAF, and residual connections, our model achieved the best results across all metrics. Notably, adding GBC led to significant improvements in F1 and mIoU by " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.57\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.42\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", respectively, highlighting its strength in capturing crack morphology cues. Similarly, residual connections boosted F1 and mIoU by " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.13\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.47\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", indicating their role in focusing on essential crack features. Although using only PAF resulted in the lowest Params, FLOPs, and Model Size, it significantly reduced performance. These findings demonstrate that our fully integrated SAVSS effectively captures crack morphology and texture cues, achieving top pixel-level segmentation results while maintaining a lightweight model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "type": "text", + "content": "Ablation studies of scanning strategies. As listed in Table 5, under the same conditions of using four different directional scanning paths, the model achieved the best performance with our designed SASS scanning strategy, improving F1 and mIoU by " + }, + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "type": "inline_equation", + "content": "0.30\\%" + }, + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "type": "inline_equation", + "content": "0.33\\%" + }, + { + "bbox": [ + 313, + 290, + 553, + 410 + ], + "type": "text", + "content": " over the diagonal snake strategy. This demonstrates SASS's ability to construct semantic and dependency information suited to crack topology, enhancing crack pixel perception in subsequent modules. More comprehensive experiments and real-world deployments are available in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 412, + 388, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 388, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 388, + 424 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "type": "text", + "content": "In this paper, we proposed SCSegamba, a lightweight structure-aware Vision Mamba for precise pixel-level crack segmentation. SCSegamba combines SAVSS and MFS to enhance crack shape and texture perception with a low parameter count. Equipped with GBC and SASS scanning, SAVSS, captures irregular crack textures across various structures. Experiments on four datasets show SCSegamba's exceptional performance, especially in complex, noisy scenarios. On the challenging multi-scenario dataset, it achieved an F1 score of 0.8390 and mIoU of 0.8479 with only 18.16G FLOPs and 2.8M parameters, demonstrating its effectiveness for real-world crack detection and suitability for edge devices. Future work will incorporate multimodal cues to enhance segmentation quality, while further optimizing VSS design and scan strategies to achieve high-quality results with low computational resources." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 620, + 424, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 424, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 424, + 633 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 553, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 706 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62272342, 62020106004, 62306212, and T2422015; the Tianjin Natural Science Foundation under Grants 23JCJQJC00070 and 24PTLYHZ00320; and the Marie Skłodowska-Curie Actions (MSCA) under Project No. 101111188." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29413" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Zaid Al-Huda, Bo Peng, Riyadh Nazar Ali Algburi, Mugahed A Al-antari, AL-Jarazi Rabea, Omar Al-maqtari, and Donghai Zhai. Asymmetric dual-decoder-u-net for pavement crack semantic segmentation. Automation in Construction, 156:105138, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision, pages 801-818, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4703-4712, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 316 + ], + "type": "text", + "content": "[4] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 370 + ], + "type": "text", + "content": "[5] Xu Cheng, Tian He, Fan Shi, Meng Zhao, Xiufeng Liu, and Shengyong Chen. Selective feature fusion and irregular-aware network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "type": "text", + "content": "[6] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "text", + "content": "[7] Wooram Choi and Young-Jin Cha. Sddnet: Real-time crack segmentation. IEEE Transactions on Industrial Electronics, 67(9):8016-8025, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 294, + 487 + ], + "type": "text", + "content": "[8] Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In International conference on machine learning, pages 933-941. PMLR, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 489, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 294, + 543 + ], + "type": "text", + "content": "[9] Jiaxiu Dong, Niannian Wang, Hongyuan Fang, Wentong Guo, Bin Li, and Kejie Zhai. Mfafnet: An innovative crack intelligent segmentation method based on multi-layer feature association fusion network. Advanced Engineering Informatics, 62:102584, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 294, + 621 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "text", + "content": "[11] Zhengyang Geng, Meng-Hao Guo, Hongxu Chen, Xia Li, Ke Wei, and Zhouchen Lin. Is attention better than matrix decomposition? In International Conference on Learning Representations, 2021. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[12] Karan Goel, Albert Gu, Chris Donahue, and Christopher Ré. It's raw! audio generation with state-space models. In International Conference on Machine Learning, pages 7616-7633. PMLR, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 161 + ], + "type": "text", + "content": "[14] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "text", + "content": "[15] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. In The International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[17] Jing-Ming Guo, Herleeyandi Markoni, and Jiann-Der Lee. Barnet: Boundary aware refinement network for crack detection. IEEE Transactions on Intelligent Transportation Systems, 23(7):7343-7358, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 398 + ], + "type": "text", + "content": "[19] Yung-An Hsieh and Yichang James Tsai. Machine learning for crack detection: Review and model performance comparison. Journal of Computing in Civil Engineering, 34(5): 04020038, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 400, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 553, + 465 + ], + "type": "text", + "content": "[20] Achref Jaziri, Martin Mundt, Andres Fernandez, and Visvanathan Ramesh. Designing a hybrid neural system to learn real-world crack segmentation from fractal-based simulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8636-8646, 2024. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "text", + "content": "[21] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 589 + ], + "type": "text", + "content": "[22] Jason Katsamenis, Eftychios Protopapadakis, Nikolaos Bakalos, Andreas Varvarigos, Anastasios Doulamis, Nikolaos Doulamis, and Athanasios Voulodimos. A few-shot attention recurrent residual u-net for crack segmentation. In International Symposium on Visual Computing, pages 199-209. Springer, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "type": "text", + "content": "[23] Narges Kheradmandi and Vida Mehranfar. A critical review and comparative study on image segmentation-based techniques for pavement crack detection. Construction and Building Materials, 321:126162, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "text", + "content": "[24] Hong Lang, Ye Yuan, Jiang Chen, Shuo Ding, Jian John Lu, and Yong Zhang. Augmented concrete crack segmentation: Learning complete representation to defend background interference in concrete pavements. IEEE Transactions on Instrumentation and Measurement, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] David Lattanzi and Gregory R Miller. Robust automated concrete damage detection algorithms for field applications." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29414" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Journal of Computing in Civil Engineering, 28(2):253-262, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[26] Qin Lei, Jiang Zhong, and Chen Wang. Joint optimization of crack segmentation with an adaptive dynamic threshold module. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[27] Huaiyuan Li, Hui Li, Chuang Li, Baohai Wu, and Jinghuai Gao. Hybrid swim transformer-cnn model for pore-crack structure identification. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "text", + "content": "[28] Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, and Chengjie Wang. Lors: Low-rank residual structure for parameter-efficient network stacking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15866-15876, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "text", + "content": "[29] Qiufu Li, Xi Jia, Jiancan Zhou, Linlin Shen, and Jinming Duan. Rediscovering bce loss for uniform classification. arXiv preprint arXiv:2403.07289, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 273, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 294, + 327 + ], + "type": "text", + "content": "[30] Jianghai Liao, Yuanhao Yue, Dejin Zhang, Wei Tu, Rui Cao, Qin Zou, and Qingquan Li. Automatic tunnel crack inspection using an efficient mobile imaging module and a lightweight cnn. IEEE Transactions on Intelligent Transportation Systems, 23(9):15190-15203, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "text", + "content": "[31] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3783-3792, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[32] Huajun Liu, Jing Yang, Xiangyu Miao, Christoph Mertz, and Hui Kong. Crackformer network for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 24(9):9240-9252, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[33] Hui Liu, Chen Jia, Fan Shi, Xu Cheng, Mianzhao Wang, and Shengyong Chen. Staircase cascaded fusion of lightweight local pattern recognition and long-range dependencies for structural crack segmentation. arXiv preprint arXiv:2408.12815, 2024. 1, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[34] Wenze Liu, Hao Lu, Hongtao Fu, and Zhiguo Cao. Learning to upsample by learning to sample. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6027-6037, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[35] Yahui Liu, Jian Yao, Xiaohu Lu, Renping Xie, and Li Li. Deepcrack: A deep hierarchical feature learning architecture for crack segmentation. Neurocomputing, 338:139-153, 2019. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[36] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[37] Liu Mushui, Jun Dan, Ziqian Lu, Yunlong Yu, Yingming Li, and Xi Li. Cm-unet: Hybrid cnn-mamba unet for remote sensing image semantic segmentation. arXiv preprint arXiv:2405.10530, 2024. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[38] Mengyang Pu, Yaping Huang, Qingji Guan, and Haibin Ling. Rindnet: Edge detection for discontinuity in reflectance, illumination, normal and depth. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6879-6888, 2021. 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[39] Haochen Qi, Xiangwei Kong, Zhibo Jin, Jiqiang Zhang, and Zinan Wang. A vision-transformer-based convex variational network for bridge pavement defect segmentation. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[40] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 23(8):11710-11719, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "text", + "content": "[41] Jianing Quan, Baozhen Ge, and Min Wang. Crackvit: a unified cnn-transformer model for pixel-level crack extraction. Neural Computing and Applications, 35(15):10957-10973, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 282 + ], + "type": "text", + "content": "[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 371 + ], + "type": "text", + "content": "[43] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third International Workshop, DLMIA 2017, and 7th International Workshop, ML-CDS 2017, Held in Conjunction with MIC-CAI 2017, pages 240–248. Springer, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "text", + "content": "[44] Huaqi Tao, Bingxi Liu, Jinqiang Cui, and Hong Zhang. A convolutional-transformer network for crack segmentation with boundary awareness. In 2023 IEEE International Conference on Image Processing, pages 86-90. IEEE, 2023. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "text", + "content": "[45] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[46] Jin Wang, Zhigao Zeng, Pradip Kumar Sharma, Osama Alfarraj, Amr Tolba, Jianming Zhang, and Lei Wang. Dualpath network combining cnn and transformer for pavement crack segmentation. Automation in Construction, 158: 105217, 2024. 1, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "text", + "content": "[47] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5493-5502, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "text", + "content": "[48] Chao Xiang, Jingjing Guo, Ran Cao, and Lu Deng. A crack-segmentation algorithm fusing transformers and convolutional neural networks for complex detection scenarios. Automation in Construction, 152:104894, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "text", + "content": "[49] Xiao Xiao, Shen Lian, Zhiming Luo, and Shaozi Li. Weighted res-unet for high-quality retina vessel segmentation. In 2018 9th international conference on information technology in medicine and education, pages 327-331. IEEE, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29415" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 696 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[51] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[52] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 239 + ], + "type": "text", + "content": "[53] Kangmin Xu, Liang Liao, Jing Xiao, Chaofeng Chen, Haoning Wu, Qiong Yan, and Weisi Lin. Boosting image quality assessment through efficient transformer adaptation with local feature enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2662-2672, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 283 + ], + "type": "text", + "content": "[54] Tomoyuki Yamaguchi, Shingo Nakamura, Ryo Saegusa, and Shuji Hashimoto. Image-based crack detection for real concrete surfaces. IEEE Transactions on Electrical and Electronic Engineering, 3(1):128-135, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "text", + "content": "[55] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 395 + ], + "type": "text", + "content": "[56] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "text", + "content": "[57] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 507 + ], + "type": "text", + "content": "[58] Hang Zhang, Allen A Zhang, Zishuo Dong, Anzheng He, Yang Liu, You Zhan, and Kelvin CP Wang. Robust semantic segmentation for automatic crack detection within pavement images using multi-mixing of global context and local image features. IEEE Transactions on Intelligent Transportation Systems, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "text", + "content": "[59] Tianjie Zhang, Donglei Wang, and Yang Lu. ECSnet: An accelerated real-time image segmentation cnn architecture for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 554, + 294, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 606 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 606 + ], + "type": "text", + "content": "[60] Xiaohu Zhang and Haifeng Huang. Distilling knowledge from a transformer-based crack segmentation model to a light-weighted symmetry model with mixed loss function for portable crack detection equipment. Symmetry, 16(5):520, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 609, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 294, + 642 + ], + "type": "text", + "content": "[61] Jian Zhou, Peisen S Huang, and Fu-Pen Chiang. Wavelet-based pavement distress detection and evaluation. Optical Engineering, 45(2):027007-027007, 2006. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 643, + 294, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 294, + 696 + ], + "type": "text", + "content": "[62] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In International Conference on Machine Learning, 2024. 2, 3, 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29416" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_content_list.json b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d10bd54a1d8bb4c32e09c3262ad915c9fdedc65 --- /dev/null +++ b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_content_list.json @@ -0,0 +1,1913 @@ +[ + { + "type": "text", + "text": "SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models", + "text_level": 1, + "bbox": [ + 102, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaofan Bai $^{1}$ , Shixin Li $^{1}$ , Xiaojing Ma $^{1}$ , Bin Benjamin Zhu $^{2}$ , Dongmei Zhang $^{2}$ , Linchen Yu $^{1*}$ $^{1}$ Huazhong University of Science and Technology \n $^{2}$ Microsoft Corporation \n{xiaofanbai, shixinli, linda, linchenyu} @hust.edu.cn, {binzhu, dongmeiz} @microsoft.com", + "bbox": [ + 117, + 202, + 857, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cloud-based AI systems offer significant benefits but also introduce vulnerabilities, making deep neural network (DNN) models susceptible to malicious tampering. This tampering may involve harmful behavior injection or resource reduction, compromising model integrity and performance. To detect model tampering, hard-label fingerprinting techniques generate sensitive samples to probe and reveal tampering. Existing fingerprinting methods are mainly based on gradient-defined sensitivity or decision boundary, with the latter showing a manifest superior detection performance. However, all existing fingerprinting methods either suffer from insufficient sensitivity or incur high computational costs.", + "bbox": [ + 75, + 321, + 472, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we theoretically analyze the black-box co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity. Based on this, we further propose Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight approach for hard-label tampering detection that inherently and efficiently combines the strengths of existing fingerprinting techniques. SDBF places fingerprint samples near the steep decision boundary, where the outputs of samples are inherently highly sensitive to tampering. We also design a Max Boundary Coverage Strategy (MBCS), which enhances samples' diversity over the decision boundary. Theoretical analysis and extensive experimental results show that SDBF outperforms existing SOTA hard-label fingerprinting methods in both sensitivity and efficiency.", + "bbox": [ + 75, + 518, + 473, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have demonstrated excellent capabilities across various domains, such as computer vision [12, 18, 24]. An increasing number of applications have incorporated DNN models. Machine Learning as a Service (MLaaS) platforms, such as Google AutoML [9], Microsoft Azure ML [21], and Amazon SageMaker [1],", + "bbox": [ + 75, + 787, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cdb057ed36d004143488473c4d414e000e8237e180a99c43b1e8cf87d56554f.jpg", + "image_caption": [ + "Figure 1. Overview of potential model tampering and public tampering detection for models deployed in MLaaS." + ], + "image_footnote": [], + "bbox": [ + 503, + 290, + 890, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leverage powerful pre-trained DNN models, allowing users with limited computational resources to access cutting-edge technology.", + "bbox": [ + 496, + 595, + 892, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, models deployed in the cloud often operate beyond the direct control of their owners, leaving them vulnerable to unauthorized/illegal model tampering as shown in Fig. 1. For instance, unscrupulous service providers might compress a model—reducing its memory and computational footprint—as a cost-saving measure without the owner's consent, as demonstrated in prior work [3, 22]. Alternatively, adversaries could inject malicious behaviors into the model's parameters to manipulate the model's outputs [10, 14, 15, 27, 29]. Such tampering not only compromises the model's integrity, security, and overall performance but also erodes users' trust and the reliability of the system.", + "bbox": [ + 496, + 643, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many works [2, 6, 13, 17, 31, 33-35] have focused on detecting model tampering. Due to the black-box nature of the models deployed in the cloud, hard-label fingerprinting techniques are the most practical but also most challenging among existing tampering detection methods as their non", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version:", + "bbox": [ + 236, + 1, + 807, + 23 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 30, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29278", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "invasive and black-box nature aligns well with the MLaaS in practice. Existing hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]. The former enhances the sensitivity of fingerprint samples' output to model tampering by maximizing gradient-defined sensitivity, which is the $L_{p}$ norm of the gradient of the output with respect to the model parameters, and these fingerprinting techniques have been widely demonstrated to be much less effective than the latter. The latter improves the sensitivity of fingerprint samples to tampering by minimizing the distance to the decision boundary. In summary, both of these approaches are still not sensitive enough to detect certain types of model tampering, especially subtle ones. However, detecting subtle model tampering is crucial, as the end user may need to confirm if the model they are using is really the one they want to use.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, [11] introduced a meta-learning-based fingerprinting approach. However, due to the inherent complexity of meta-learning, this method incurs significantly higher overhead compared to the two aforementioned categories, making it impractical for resource-constrained model owners or users. We argue that, given the very purpose of black-box model tampering detection, a lightweight design for fingerprinting method is essential for its widespread application.", + "bbox": [ + 75, + 333, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, recognizing the importance of closing to decision boundary for tampering detection with fingerprint samples, we theoretically analyze the co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity in a black-box detection manner. Based on this, we further propose a novel lightweight hard-label tampering detection fingerprinting approach, Steep-Decision-Boundary Fingerprinting (SDBF) to improve the fingerprint samples' sensitivity with the inherent highly sensitive nature of samples near the steep decision boundary. Specifically, SDBF places fingerprint samples close to the steep decision boundary, where the outputs exhibit high sensitivity to the tampering-induced shift of decision boundary. To achieve this, SDBF perturbs normal input samples to maximize the $L_{2}$ norm of the logit output while simultaneously minimize its variance. The fingerprint samples are bound to be almost indistinguishable from the original normal samples. Moreover, to further enhance detection performance when using multiple fingerprint samples to detect, we propose a Max Boundary Coverage Strategy (MBCS). This strategy enables the fingerprint samples to explore a wide range of different regions around the decision boundary.", + "bbox": [ + 75, + 455, + 472, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our theoretical analysis and extensive experimental results demonstrate that the sensitivity of SDBF is superior to that of existing state-of-the-art (SOTA) hard-label fingerprinting methods while being time and memory-efficient. In summary, our contributions are as follows:", + "bbox": [ + 75, + 801, + 472, + 877 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We theoretically analyze the black-box co-optimal tam", + "bbox": [ + 94, + 885, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "pering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity.", + "- We introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight fingerprinting method that inherently and efficiently combines the strengths of fingerprinting techniques based on gradient-defined sensitivity and decision boundary.", + "- We propose a Max Boundary Coverage Strategy (MBCS) to ensure that the fingerprint samples widely explore the decision boundary, further enhancing the detection performance when using multiple fingerprint samples to detect model tampering.", + "- We present both theoretical analysis and extensive experimental results that demonstrate the superior performance and efficiency of SDBF compared to existing SOTA hard-label fingerprinting methods." + ], + "bbox": [ + 517, + 90, + 893, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 395, + 640, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Hard-Label Tampering Detection", + "text_level": 1, + "bbox": [ + 500, + 420, + 792, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tampering detection for DNN models aims to determine whether a model has been altered. Since MLaaS operates as a black-box to users, existing methods focus on hard-label tampering detection, where only the top-1 label is accessible. In this setting, watermarking and fingerprinting are widely used. Watermarking [35], an invasive method, embeds a fragile watermark during training that can later be queried with specific inputs to detect tampering. However, because modifying the model may impact performance, non-invasive fingerprinting techniques, the focus of this paper, have attracted more attention.", + "bbox": [ + 496, + 443, + 893, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Fingerprinting for Tampering Detection", + "text_level": 1, + "bbox": [ + 500, + 619, + 843, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing black-box hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34].", + "bbox": [ + 496, + 643, + 893, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient-defined sensitivity. Sensitive-Sample Fingerprinting (SSF) [13] introduces fingerprinting for detecting tampering in DNN models for the first time by gradient-defined sensitivity. SSF generates sensitive samples—natural-looking inputs that cause large output changes when the model's parameters are slightly altered. Following SSF, symbolic constraint solvers [6] and Bayesian optimization (BO) [17] have been used to optimize the generation of sensitive samples. However, such sensitive samples have been proven to be ineffective in extensive experiments for many kinds of model tampering.", + "bbox": [ + 496, + 689, + 893, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Decision boundary based sensitivity. PublicCheck [31] uses VQVAE models to generate fingerprint samples by carefully adding perturbations along with certain directions", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29279", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the latent space of VQVAE and stopping when it has just crossed the decision boundary of two different categories. Others like AID [2] and DBFW [34] generate fingerprint samples close to the decision boundary of all categories by minimizing output variance using mean square error (MSE) loss functions. Similar to AID and DBFW, IBSF [33] employs a Shannon entropy loss to generate diverse fingerprint samples near different $K(K \\geqslant 2)$ categories' decision boundary. While these fingerprinting techniques show significantly superior detection performance than the aforementioned fingerprinting techniques based on gradient-defined sensitivity, they are still not sensitive enough to detect certain types of model tampering, especially subtle ones. Notably, neither one has theoretically justified why samples that are close to the decision boundary are more effective than the faraway, nor has anyone theoretically analyzed the co-optimal tampering detection sensitivity of fingerprint samples in these two contexts.", + "bbox": [ + 75, + 90, + 472, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, [11] proposed a meta-learning-based sensitive fingerprinting method, MiSentry, but due to the inherent high overhead of meta-learning, the cost of MiSentry is significantly higher than the two aforementioned categories, making it unaffordable for those resources-limited model owner/users. We believe that due to the very purpose of black-box tampering detection, the lightweight of a fingerprinting method is crucial for its widespread application.", + "bbox": [ + 75, + 364, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Threat Model", + "text_level": 1, + "bbox": [ + 76, + 501, + 217, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aligned with existing works, we adopt a threat model that includes white-box generation and black-box detection. Specifically, during the fingerprint samples generation, we assume white-box access to the target DNN, using available normal samples as source data. This assumption is reasonable as fingerprint samples are generated by the model owner. In the detection phase, we assume there is only black-box access to the suspect model, where only an API-based query is allowed, and only the top-1 label of the input query will be returned. Moreover, following [31], we also assume a trustworthy third party that securely stores and distributes the fingerprint samples for public tampering detection, where the DNN service provider is assumed to be untrustworthy and may detect probing fingerprint samples, evading tampering detection. Additionally, adversaries might access some fingerprint samples through some illegal means, and use them to bypass tampering detection.", + "bbox": [ + 75, + 527, + 472, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Steep Decision Boundary and Sensitivity", + "text_level": 1, + "bbox": [ + 75, + 801, + 439, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first define the decision boundary in the context of the classification task. We then theoretically demonstrate why samples near the decision boundary are sensitive to shifts caused by model tampering. Next, we define the steep decision boundary and analyze how to place", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fingerprint samples near it to further enhance samples' sensitivity while improving optimization efficiency.", + "bbox": [ + 498, + 90, + 893, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Steep Decision Boundary", + "text_level": 1, + "bbox": [ + 498, + 128, + 728, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A DNN is a function $f$ , parameterized by $W$ , which maps an input $x$ to an output logits vector $\\mathbf{y} = f(W, x)$ . For classification tasks, let $N$ denotes the number of classes, the logits vector $\\mathbf{y}$ will then be transformed into a probability vector $Y$ by the following Softmax function:", + "bbox": [ + 496, + 152, + 893, + 228 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {i} = \\operatorname {S o f t m a x} \\left(\\mathbf {y} _ {i}\\right) = \\frac {e ^ {\\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {N} e ^ {\\mathbf {y} _ {j}}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 234, + 892, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Y_{i}$ is the confidence of class $i$ .", + "bbox": [ + 498, + 279, + 743, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4.1 (Decision Boundary). For a classification model, the decision boundary is the hypersurface that divides the feature space into regions, each corresponding to a different class label. Typically, the model assigns the same probabilities to the classes on the decision boundary, and the confidences of these categories will be much greater than the other categories. Specifically, for an $N$ -class classification model, the decision boundary of $K(K \\leqslant N)$ classes can be formulated as:", + "bbox": [ + 496, + 294, + 893, + 430 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {1} = Y _ {2} = \\dots = Y _ {K} \\gg \\sum \\tilde {Y} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 441, + 892, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sum \\tilde{Y} (0\\leqslant \\sum \\tilde{Y}\\leqslant 1)$ is the summation of the other $N - K$ categories' confidence and $\\begin{array}{r}\\sum_{i = 1}^{K}Y_{i} + \\sum \\tilde{Y} = 1. \\end{array}$", + "bbox": [ + 498, + 468, + 892, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For black-box tampering detection, the predicted hard label based on the Softmax output $Y$ is the only observable feedback from the suspect model. To maximize the sensitivity of fingerprint samples in this setting, we need to maximize the sensitivity of $Y$ with respect to $\\mathbf{y}$ to force the model to work in Softmax high-sensitivity region, making any tampering-induced $\\Delta \\mathbf{y} = f(W + \\Delta W, x) - f(W, x)$ will be nonlinearly amplified to $Y$ by Softmax, avoiding being suppressed by the saturation zone of Softmax when the fingerprint samples are faraway from the decision boundary.", + "bbox": [ + 496, + 510, + 893, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 4.2. Given a fingerprint sample $x$ , the sensitivity of its softmax output $\\mathbf{Y}$ with respect to $||\\Delta \\mathbf{y}||_2$ introduced by model tampering $\\Delta W$ will reach the maximum when $x$ close to the decision boundary, where $Y_i = \\frac{1}{N}$ .", + "bbox": [ + 496, + 665, + 893, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. To prove Prop.4.2, we need to analyze the Jacobian matrix $J$ of vector $\\mathbf{Y}$ with respect to vector $\\mathbf{y}$ and the maximum value of the Frobenius norm (equivalent to the $L_{2}$ norm) of $J$ . First, based on Eq.1, the element $J_{ij}$ of $J$ is the partial derivative of $Y_{i}$ with respect to $y_{j}$ :", + "bbox": [ + 496, + 732, + 893, + 809 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nJ _ {i j} = \\frac {\\partial Y _ {i}}{\\partial \\mathbf {y} _ {j}} = Y _ {i} \\left(\\delta_ {i j} - Y _ {j}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 816, + 892, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_{ij}$ is the Kronecker Delta function and $\\left\\{ \\begin{array}{l}\\delta_{ij} = 1,i = j\\\\ \\delta_{ij} = 0,i\\neq j \\end{array} \\right.$ . Thus, the squared Frobenius norm of $J$", + "bbox": [ + 500, + 856, + 893, + 904 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "29280", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can be formulated as:", + "bbox": [ + 76, + 90, + 223, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\| J \\| _ {F} ^ {2} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(J _ {i j}\\right) ^ {2} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(Y _ {i} \\left(\\delta_ {i j} - Y _ {j}\\right)\\right) ^ {2} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - 2 Y _ {i} + Y _ {i} ^ {2} + \\sum_ {j \\neq i} Y _ {j} ^ {2}\\right) \\tag {4} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - 2 Y _ {i} + Y _ {i} ^ {2}) + Y _ {i} ^ {2} \\sum_ {j \\neq i} Y _ {j} ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 113, + 468, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\sum_{i=1}^{N} Y_{i} = 1$ , thus $\\sum_{j \\neq i} Y_{j}^{2}$ can be replaced by $1 - Y_{i}^{2}$ :", + "bbox": [ + 76, + 256, + 468, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\| J \\| _ {F} ^ {2} = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - 2 Y _ {i} + Y _ {i} ^ {2}\\right) + Y _ {i} ^ {2} \\left(1 - Y _ {i} ^ {2}\\right) \\tag {5} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - Y _ {i}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 301, + 468, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we need to find the maximum of Eq.5. For such a constrained optimization problem, we use the Lagrange multiplier method to rigorously prove when the maximum of Eq.5 is achieved. Specifically, we introduce a Lagrange multiplier $\\lambda$ for the constraint $\\sum_{i=1}^{N} Y_i = 1$ , which is:", + "bbox": [ + 76, + 395, + 468, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(Y _ {1}, Y _ {2}, \\dots , Y _ {N}, \\lambda\\right) = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - Y _ {i}\\right) + \\lambda \\left(\\sum_ {i = 1} ^ {N} Y _ {i} - 1\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 482, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now compute the partial derivatives of $\\mathcal{L}$ with respect to each $Y_{i}$ . The partial derivative of $Y_{i}^{2}(1 - Y_{i})$ with respect to each $Y_{i}$ can be formulated as:", + "bbox": [ + 76, + 536, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial Y _ {i}} \\left(Y _ {i} ^ {2} \\left(1 - Y _ {i}\\right)\\right) = 2 Y _ {i} \\left(1 - Y _ {i}\\right) - Y _ {i} ^ {2} = 2 Y _ {i} - 3 Y _ {i} ^ {2} (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 603, + 468, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, for the constraint term, the partial derivative is:", + "bbox": [ + 96, + 643, + 468, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial Y _ {i}} \\left(\\sum_ {i = 1} ^ {N} Y _ {i} - 1\\right) = 1 \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 669, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let the total derivative of the Lagrangian equal to zero, we have:", + "bbox": [ + 76, + 720, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n2 Y _ {i} - 3 Y _ {i} ^ {2} + \\lambda = 0 \\Longrightarrow 2 Y _ {i} - 3 Y _ {i} ^ {2} = - \\lambda \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 760, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that we have the constraint $\\sum_{i=1}^{N} Y_i = 1$ , to find a feasible solution, since Eq.9 holds for each $Y_i$ , we can conclude that all $Y_i$ must be equal, meaning $Y_1 = Y_2 = \\dots = Y_N = Y$ . Otherwise, the right term can not be the same constant $-\\lambda$ . Then, we can obtain:", + "bbox": [ + 76, + 789, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN \\cdot Y = 1 \\quad \\Rightarrow \\quad Y = \\frac {1}{N} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 875, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, to verify that this solution is indeed the maximum, we can check the second order derivative value of $||J||_F^2$ at this solution. When all $Y_{i} = \\frac{1}{N}$ , the second order derivative of $||J||_F^2$ becomes:", + "bbox": [ + 498, + 90, + 893, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial^ {2} \\left| \\left| J \\right| \\right| _ {F} ^ {2}}{\\partial Y _ {i} ^ {2}} = \\sum_ {i = 1} ^ {N} \\frac {\\partial 2 Y _ {i} - 3 Y _ {i} ^ {2}}{\\partial Y _ {i}} = 2 - 6 \\sum_ {i = 1} ^ {N} Y _ {i} < 0. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 162, + 893, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, this solution is indeed the maximum. In summary, given a fingerprint sample, $\\| J\\| _F^2$ achieves maximum when it is close to the decision boundary defined in Def.4.1.", + "bbox": [ + 498, + 214, + 893, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, when $x$ is faraway from the decision boundary, some elements in $f(W,x)$ are much larger than others, leading to $\\text{softmax}(f(W,x))$ that are close to 0 or 1. In this case, the elements of the Jacobian matrix $J_{ij}$ will be significantly different, especially when $Y_{i} \\approx 1$ , causing the most of other $Y_{j} \\approx 0$ , that is:", + "bbox": [ + 498, + 258, + 893, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} Y _ {i} \\approx 1 \\\\ Y _ {j} \\approx 0 \\end{array} \\right. \\Longrightarrow \\left\\{ \\begin{array}{c} J _ {i i} \\approx Y _ {i} \\cdot (1 - Y _ {i}) \\approx 0 \\\\ J _ {i j} \\approx - Y _ {i} \\cdot Y _ {j} \\approx 0, i \\neq j \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 361, + 892, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, $||J||_F^2$ in this scenario will be significantly smaller as most elements are close to 0.", + "bbox": [ + 500, + 406, + 892, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 4.3 (Steep Decision Boundary). Given a fingerprint sample $x$ and a model tampering $\\Delta W$ , the steep decision boundary refers to the decision boundary where $f(W, x)$ shows a rapid change when $\\Delta W$ is applied to $W$ .", + "bbox": [ + 498, + 446, + 893, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 4.4. Give a sample $x$ and a DNN model $f(\\cdot)$ parameterized by $W$ , the sensitivity of output $f(W, x)$ with respect to $\\Delta W$ is positively correlated to the $||f(W, x)||_2$ .", + "bbox": [ + 498, + 518, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Inspired by [13], since all model tampering inevitably result in different performances in the last layer, the sensitivity of $f(W,x)$ to $W$ practically can be formulated as the following Frobenius norm of Jacobian matrix $\\hat{J}$ :", + "bbox": [ + 498, + 573, + 893, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\left| \\hat {J} \\right| \\right| _ {F} ^ {2} = \\left| \\left| \\frac {\\partial f (W , x)}{\\partial W _ {F C}} \\right| \\right| _ {2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 643, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{FC}$ is the weights of the last fully connected layer of $f(\\cdot)$ . Since all existing classification models do not use any nonlinear activation functions like $ReLu(\\cdot)$ in the last layer, the last layer of the model performs the following linear operations on the output of the previous layer:", + "bbox": [ + 498, + 686, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (W, x) = W _ {F C} \\times f _ {- 2} (W, x) + \\mathbf {b} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 773, + 890, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{-2}(W,x)$ and $\\mathbf{b}$ are the output and bias of the penultimate layer, respectively. Substituting Eq. 14 into Eq. 13, we can get:", + "bbox": [ + 498, + 801, + 893, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\left| \\hat {J} \\right| \\right| _ {F} ^ {2} = \\left| \\left| f _ {- 2} (W, x) \\right| \\right| _ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 845, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining Eq.15 and Eq.14, when we increase $||f(W,x)||_2$ , the sensitivity of $f(W,x)$ with respect to $\\Delta W$", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29281", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\| \\hat{J}\\| _F^2$ , will increase simultaneously. To valid Prop.4.4, we conduct an experiment by maximizing $\\| f(W,x)\\| _2$ while recording $\\| \\frac{\\partial f(W,x)}{\\partial\\mathbf{W}_{FC}}\\| _2$ for a ResNet20 trained on CIFAR10, as shown in Fig. 2. The results are highly consistent with Prop.4.4. Therefore, we can improve the sample's sensitivity of $f(W,x)$ to $\\Delta W$ by increasing $\\| f(W,x)\\| _2$", + "bbox": [ + 76, + 89, + 470, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8271a1f4520dad734185fc33a3aa57a632beb4854f741df230729e082b5f5d70.jpg", + "image_caption": [ + "Figure 2. Validation results on the relationship between $||f(W,x)||_2$ and $||\\frac{\\partial f(W,x)}{\\partial W_{FC}}||_2$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 207, + 467, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 4.5. Give a fingerprint sample $x$ and a DNN model $f(\\cdot)$ parameterized by $W$ , simultaneously increasing the sensitivity of softmax $(f(W, x))$ to corresponding $f(W, x)$ and the sensitivity of $f(W, x)$ to $\\Delta W$ can drive $x$ close to the steep decision boundary and thus further enhance the sensitivity of softmax $(f(W, x))$ to $\\Delta W$ .", + "bbox": [ + 76, + 443, + 470, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prop.4.5 can be easily proved with the linear translational invariance of $\\text{Softmax}(\\cdot)$ . More details are in the appendix. Based on Def.4.3 and Prop.4.5, we know how to efficiently place the fingerprint samples near steep decision boundary, where they are highly sensitive to model tampering $\\Delta W$ as it inherently combining the strength of fingerprinting methods based on gradient-defined sensitivity and decision boundary.", + "bbox": [ + 75, + 539, + 470, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Steep-Decision-Boundary Fingerprinting", + "text_level": 1, + "bbox": [ + 76, + 650, + 444, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce a novel sensitive fingerprinting method named Steep-Decision-Boundary Fingerprinting (SDBF). Our goal is to leverage the highly sensitive nature of the samples near the steep decision boundary to detect model tampering, as shown in Fig. 3.", + "bbox": [ + 75, + 675, + 470, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Fingerprint Samples Generation", + "text_level": 1, + "bbox": [ + 76, + 756, + 362, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from [13], we do not directly optimize Eq.13 and Eq.5 as their poor efficiency. We will further explain this in the appendix. Fortunately, based on Prop.4.2, Prop. 4.4 and Prop. 4.5, we can maximize the fingerprint samples' sensitivity by maximizing $||\\mathbf{y}||_2$ while minimizing the variance of $\\mathbf{y}$ , $\\sigma (\\mathbf{y})$ . Similar to IBSF [33], we can also freely choose to place fingerprint samples at $K(K\\leqslant N)$ categories' decision boundary, which will help increase the diversity of", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/74dfa27de7b4ea707353750cd5976de4a8e8f8b59c118f24140ecc0ca432ee6c.jpg", + "image_caption": [ + "Figure 3. Illustration of SDBF, gradient-defined fingerprinting SSF, and decision-boundary-based fingerprinting IBSF. SDBF shows an inherent high sensitivity to tampering-introduced boundary shift." + ], + "image_footnote": [], + "bbox": [ + 511, + 95, + 694, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/458c102fdccdb1ee207d49513250ee65516668a51bb2cdf3ad692f0819af0c7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 98, + 890, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the fingerprint samples. Specifically, let $\\mathbf{y}_{\\mathcal{C}_{sub}}$ denote the logits output of the chosen categories and $\\overline{\\mathbf{y}}_{\\mathcal{C}_{sub}} = \\frac{1}{K}\\sum_{i=1}^{k}\\mathbf{y}_i$ denote the averaged logits of the categories in $\\mathcal{C}_{sub}$ , our objective function for a sample $x$ can be formalized as:", + "bbox": [ + 496, + 292, + 893, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (x, \\mathcal {C} _ {s u b}) = \\sigma (\\mathbf {y} _ {\\mathcal {C} _ {s u b}}) - \\lambda \\cdot | | \\mathbf {y} _ {\\mathcal {C} _ {s u b}} | | _ {2} \\\\ = \\frac {1}{K} \\cdot \\sum_ {i = 1} ^ {K} \\left(\\mathbf {y} _ {i} - \\overline {{\\mathbf {y}}} _ {\\mathcal {C} _ {s u b}}\\right) ^ {2} - \\lambda \\cdot \\sum_ {i = 1} ^ {K} \\left\\| \\mathbf {y} _ {i} \\right\\| _ {2} \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 375, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the coefficient and $i\\in \\mathcal{C}_{sub}$ . Thus, given a normal sample $x$ , we can drive it close to the steep decision boundary by minimizing the objective function in Eq.16:", + "bbox": [ + 496, + 441, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\Delta x} \\mathcal {L} (x + \\Delta x, \\mathcal {C} _ {s u b}) \\tag {17} \\\\ s. t. | | \\Delta x | | _ {p} \\leqslant \\epsilon \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 498, + 890, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the input bound. The constraint $\\| \\Delta x\\| _p\\leqslant \\epsilon$ bounds the perturbation $\\Delta x$ in the input space, making the generated fingerprint sample perceptually similar to the original normal one. This naturalness makes it hard for an adversary to distinguish whether the input is a probe query by input anomaly detection. We solve Eq. 17 by using projected gradient descent (PGD) like [19]:", + "bbox": [ + 496, + 544, + 893, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i + 1} = \\operatorname {C l i p} _ {\\epsilon} \\left(x _ {i} + \\eta \\cdot \\nabla_ {x _ {i}} \\mathcal {L} \\left(x _ {i}, \\mathcal {C} _ {\\text {s u b}}\\right)\\right) \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 665, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{i} = x_{i - 1} + \\Delta x_{i}$ is the generated fingerprint sample at the $i$ -th iteration.", + "bbox": [ + 496, + 686, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization eventually yields a fingerprint sample positioned close to $\\mathcal{C}_{sub}$ steep decision boundary. Following [33], to reduce the perturbation effort, we select the top $|\\mathcal{C}_{sub}|$ confident categories in $\\mathbf{y}$ of original normal $x$ as $\\mathcal{C}_{sub}$ by default. This sample-dependent $\\mathcal{C}_{sub}$ will enhance the diversity of generated sensitive fingerprint samples, making it more robust to partial leakage.", + "bbox": [ + 496, + 715, + 890, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark. In the generation phase, except for fingerprint sample $x$ , we will also collect its corresponding top-1 label as the fingerprint label $\\mathcal{V}$ . In the de", + "bbox": [ + 524, + 839, + 870, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29282", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tection phase, $x$ will be used to query the suspect model to get the returned hard label $\\mathcal{V}_{predicted}$ and then we will determine whether the model has been tampered with by verifying whether $\\mathcal{V}_{predicted} = \\mathcal{V}$ .", + "bbox": [ + 101, + 99, + 444, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Max Boundary Coverage Strategy", + "text_level": 1, + "bbox": [ + 76, + 181, + 375, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When probing the tampered model with $M$ ( $M > 1$ ) fingerprint samples, we collect their returned hard labels. The model is considered untampered if all the labels match the corresponding fingerprint labels. To maximize the diversity of the $M$ fingerprint samples over the steep decision boundary, we propose the Max Boundary Coverage Strategy (MBCS) to make them widely explore the decision boundary. Specifically, we first randomly select a different fingerprint sample, denoted as $x_{1}$ . Then, let $m \\leqslant M$ and $x_{i}$ denote the number of selected fingerprint samples and the sample to select in $i$ -th iteration, we iteratively choose the remaining $M - 1$ samples to maximize the boundary diversity score:", + "bbox": [ + 75, + 204, + 470, + 387 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x _ {i}} \\mathcal {L} _ {d i v} = \\sum_ {i = 1} ^ {N} c _ {i} \\cdot \\log c _ {i} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 398, + 468, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the number of categories and", + "bbox": [ + 76, + 445, + 349, + 459 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} = \\frac {1}{m} \\cdot \\sum_ {j = 1} ^ {m} \\operatorname {s o f t m a x} _ {i} \\left(f \\left(x _ {j}\\right)\\right) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 470, + 468, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\text{softmax}_i(\\cdot)$ denotes softmax operation along with $i - th$ dimension.", + "bbox": [ + 75, + 523, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By repeating the $MBCS\\mathcal{R}$ times, we can get a $R$ -length and $M$ -width probe fingerprint sequence. For each probe query, we will send $M$ probe fingerprint samples to the suspect model until we go through all $\\mathcal{R} \\times M$ fingerprint samples. In our experiments, we use $MBCS$ as a post-processing technique to enhance the diversity of the fingerprint samples over the decision boundary.", + "bbox": [ + 75, + 553, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 76, + 670, + 310, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Datasets, Models, and Baselines", + "text_level": 1, + "bbox": [ + 76, + 695, + 356, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of SDBF, we use various common classification datasets, including CIFAR10 [16], GTSRB [28], and ImageNet [25]. To evaluate the generalizability of SDBF across different model architectures, we select three different popular architectures for the above three datasets. Specifically, we use ResNet20 [12] for CIFAR10, GoogLeNet [30] for GTSRB, and vision transformer ViT-B-32 [7] pretrained by Pytorch for ImageNet.", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the superiority of SDBF compared to existing SOTA fingerprinting methods, we select four SOTA fingerprinting methods, SSF [13], PublicCheck [31], IBSF [33], and MiSentry [11], as the baselines. We utilize the default", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "settings outlined in their respective papers unless explicitly specified otherwise. For each fingerprinting method, we independently generate 1000 fingerprint samples.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Tampering Settings and Detection Rate", + "text_level": 1, + "bbox": [ + 498, + 141, + 836, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tampering Settings. In our experiments, we evaluate SDBF and the baselines on the 9 types of model tampering selected from BIVBench [11], including 4 hardest-to-detect model tampering, targeted attack [15,27,29], finetuning the last layer with $10^{-5}$ learning rate, and Unlearning [8]/Online Learning (On-Learning) [26] makes a model forget/newly learn specific data, and 5 common model tampering consists of Backdoor [10], Bit-flipping Attack [23], Quantization [4], Pruning [20], and independently trained models with Different Architecture (Different Arch.) on the same dataset. All the tampering settings are aligned with BIVBench [11]. In this paper, the reported experimental results for each tampering type are averaged over 10 independently tampered models to mitigate randomness in evaluating tampering detection performance.", + "bbox": [ + 496, + 164, + 893, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tampering Detection Rate. For each tamper detection probe, $M$ ( $M \\geqslant 1$ ) fingerprint samples are selected from a pool of 1000 fingerprint samples to query the test model. The model's top-1 predicted labels are then compared with the corresponding fingerprint labels. If all predicted labels match the fingerprint labels, the model is classified as tampered; otherwise, it is marked as tampered. This procedure is repeated 100 times for each test model. Since 10 tampered models are generated for each tampering type, a total of 1000 tests are performed for each tampering type. The tampering detection rate is the proportion of the 1000 probe queries in which tampering is successfully detected. Note that there are no false positives (FP) in the tampering detection of classification model, as the classification model $f(\\cdot)$ provides a stable mapping from input $x$ to output $f(x)$", + "bbox": [ + 496, + 391, + 893, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Perceptual Quality of Fingerprint Samples", + "text_level": 1, + "bbox": [ + 498, + 619, + 862, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 shows a subset of fingerprint samples generated by SDBF and four baselines, all constrained by the same $L_{2}$ norm bound. Fingerprint samples generated by SDBF closely resemble those from SSF, IBSF, and MiSentry, with imperceptible pixel-level perturbations added to the original images. In contrast, fingerprint samples generated by PublicCheck exhibit noticeable artifacts, with semantic-level alterations to the original images, due to the use of generative models. For example, the distorted face in the 8th column of Fig. 4 highlights this effect.", + "bbox": [ + 496, + 642, + 893, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.4. Effectiveness of Tampering Detection", + "text_level": 1, + "bbox": [ + 498, + 801, + 821, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single Fingerprint Sample Detection. We first evaluate the detection performance in the single fingerprint sample detection setting, which only uses one fingerprint sample to query the suspect model. Specifically, following [33], we comprehensively evaluate SDBF and IBSF with $K$ ranging", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29283", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bda21067064b9de89cead5ab148cbcf00e5351d958b383072dfe13dc2df93c73.jpg", + "table_caption": [ + "Table 1. Tamper detection rate (%) of detecting tampering in a single fingerprint sample setting on CIFAR10 and GTSRB. For fine-tuning(FT), \"Last\" specifies only fine-tuning the last layer, with numbers indicating the ${10}^{-5}$ learning rate. The \"On-Learning\" refers to online learning." + ], + "table_footnote": [], + "table_body": "
DatasetMethodFTLast-5BackdoorT-AttackUnlearningQuantizationDifferent Arch.PruningOn-LearningBit-flipping
CIFAR10SSF0.064.911.43.266.379.766.76.784.4
MiSentry45.986.483.868.490.793.589.568.288.9
PublicCheck21.673.636.928.360.865.169.525.485.7
IBSFK=224.577.442.537.578.283.381.247.587.8
K=435.983.655.359.489.188.588.356.289.4
K=642.887.563.569.290.490.890.569.588.6
K=844.589.466.171.390.992.791.371.689.5
K=1049.490.973.272.891.393.891.976.490.1
SDBF(ours)K=238.181.456.456.586.685.384.158.188.4
K=447.587.368.669.889.791.591.266.989.2
K=653.189.979.773.491.592.892.373.588.3
K=856.291.483.175.191.993.792.975.890.3
K=1062.892.585.877.393.294.693.681.189.8
GTSRBSSF0.059.89.51.462.664.559.12.785.1
MiSentry35.286.475.159.487.979.583.759.889.3
PublicCheck18.467.837.321.769.266.570.428.387.3
IBSFK=222.776.339.232.879.872.177.542.688.5
K=431.979.952.955.986.179.783.257.390.4
K=638.485.859.861.889.183.385.563.490.1
K=841.588.563.470.191.285.988.465.189.6
K=1042.891.169.773.391.888.490.368.291.3
SDBF(ours)K=229.180.949.849.484.777.282.953.990.3
K=438.586.662.566.890.584.589.464.391.5
K=646.790.378.471.592.289.691.772.789.2
K=853.291.581.573.692.991.391.977.991.9
K=1057.192.784.678.293.693.892.880.190.7
", + "bbox": [ + 80, + 121, + 887, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from $K = 2$ to $K = N = 10$ and $K = 2$ to $K = 500$ on CIFAR10 and ImageNet, respectively. Tab.1 summarizes the tampering detection rates on CIFAR10 and GTSRB of existing SOTA fingerprinting methods and SDBF. The detailed results on ImageNet are in the appendix. From the results, we can conclude that:", + "bbox": [ + 75, + 426, + 470, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SDBF generally outperforms the four SOTA baselines, although exceptions arise for SDBF at $K = 2$ when using a single fingerprint sample to perform tampering detection, which can be attributed to MiSentry's inherent sensitivity to such tampering as some subtle tampering is in its model zoo of meta-learning. As $K$ increases, SDBF gradually outperforms MiSentry, with improvements of up to $16.9\\%$ on CIFAR10 at $K = 10$ for fine-tuning the last layer with $10^{-5}$ learning rate compared to MiSentry. The detection rate of IBSF shares a similar trend with our SDBF when $K$ increases, yet it is consistently performing worse than SDBF, evidencing our key intuition: the fingerprint samples close to the steep decision boundary are more sensitive.", + "bbox": [ + 75, + 518, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple Fingerprint Samples Detection. Next, we evaluate the detection performance in multiple fingerprint samples detection, which uses $M(M > 1)$ fingerprint samples to query the tampered model. If $\\mathcal{V}_{predicted}$ of any one fingerprint sample in the $M$ samples differ from the corresponding $\\mathcal{V}$ , the model will be deemed as a tampered one. Following the detection setting in IBSF, MiSentry, and PublicCheck, we repeat the detection probe 1000 times by selecting $M$ random fingerprint samples for them. As for SDBF and SSF, we use MBCS and Max Neuron Coverage Strategy (MNCS) [13], which is specifically designed for SSF to select $M$ fingerprint samples for each probe, respectively.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a94eef6b5d24eacf83c0a9211a3664726c2d49684c1bdcc258141ab445ca5f8b.jpg", + "image_caption": [ + "Figure 4. Source images (top) from ImageNet and fingerprint samples generated from them by SDBF (2nd row), IBSF (3rd row), SSF (4th row), MiSentry (5th row), and PublicCheck (bottom)." + ], + "image_footnote": [], + "bbox": [ + 503, + 422, + 743, + 648 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11df4dcc11f6c66a4a88e3188beb03ed301bad491c40837a65662220c55fc104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 422, + 888, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of detecting fine-tuning the last layer with $10^{-5}$ learning rate on CIFAR10 are shown in Fig.5. From the results, we can see that using multiple fingerprint samples for detection generally performs better than a single one. Among these methods, SDBF achieves perfect detection rates fastest as $M$ increases when $K = 2$ .", + "bbox": [ + 496, + 702, + 893, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we also conduct an ablation study on whether $MBCS$ improves detection performance when using multiple fingerprint samples (see SDBF-MBCS and SDBF Without $MBCS$ in Fig.5). The results demonstrate $MBCS$ indeed boosts the multiple fingerprint detection performance. For example, after being equipped with $MBCS$ , the detection rate of SDBF has been boosted from $68.3\\%$ to $84.5\\%$ and", + "bbox": [ + 496, + 795, + 893, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29284", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$79.1\\%$ to $97.5\\%$ when $M = 2$ and $M = 3$ at $K = 2$ , respectively, making the detection rate of SDBF faster grow up to perfect state than existing SOTA fingerprinting methods.", + "bbox": [ + 76, + 90, + 470, + 137 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0ee97b6628e0b0c975773bd3442a85d37347b2c7c5e5e0380a0c69501371f91.jpg", + "image_caption": [ + "(a) CIFAR10" + ], + "image_footnote": [], + "bbox": [ + 81, + 143, + 272, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9c78d4b5a9d992b8d2c28ac081c440c9e88d65bcb7ea7b1b36f079fb829f66d.jpg", + "image_caption": [ + "(b) GTSRB", + "Figure 5. Detection rates on CIFAR10 and GTSRB for fine-tuning the last layer with $10^{-5}$ learning rate in multiple fingerprint samples setting. Note that $K = 2$ for both IBSF and SDBF here." + ], + "image_footnote": [], + "bbox": [ + 274, + 143, + 467, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Post-Tampering Output Naturalness", + "text_level": 1, + "bbox": [ + 76, + 289, + 390, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When the model has been tampered with by adversaries, they might reject the low-confidence query to prevent detection by sensitive fingerprint samples based on the decision boundary. To evaluate the effectiveness of SDBF in responding to such output anomaly detection, we statistically analyzed the top-1 label probabilities of the backdoor-tampered model with SDBF. The results are shown in Fig.7. From the results, we can observe that SDBF achieves comparable covertness to MiSentry, which provides the highest output concealment. This result evidenced that the output probability of the fingerprint samples generated by SDBF is highly sensitive to model tampering and is consistent with our theoretical analysis. Unfortunately, IBSF is at risk of abnormal output rejection, especially on GTSRB.", + "bbox": [ + 75, + 311, + 470, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Robustness to Adaptive Attacks", + "text_level": 1, + "bbox": [ + 76, + 531, + 356, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adversaries may attempt to bypass SDBF by collecting generated fingerprint samples or generating new ones for adaptive attacks, aiming to manipulate the model while keeping the top-1 labels of the fingerprint samples unchanged. To evaluate SDBF's robustness against such attacks, we generate 1,000 additional fingerprint samples, referred to as leaked fingerprints, which are used for launching adaptive attacks. The set of original samples used for tampering detection is called the unleaked fingerprints.", + "bbox": [ + 75, + 553, + 468, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct adaptive backdoor attacks on CIFAR10, as shown in Fig. 6 (a). As training progresses, the detection rate for the leaked fingerprints drops to $0\\%$ , while the detection", + "bbox": [ + 75, + 690, + 468, + 736 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5eddbcf0af591f6238eff7b415587bbe1b7a8bbd4ebe0d786ad38fc685a2917.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 760, + 271, + 848 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a9880fedad863154345aeb7bdaafb3ac60575c6320a56e25678de76d45ff209.jpg", + "image_caption": [ + "(b)", + "Figure 6. (a) Tampering detection rates of detecting adaptive backdoor attacks on CIFAR10 at $K = 2$ in a single fingerprint sample setting. (b) Adaptive tampering detection rates when a larger number of fingerprints are leaked." + ], + "image_footnote": [], + "bbox": [ + 271, + 758, + 467, + 849 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e234e555a7b52c66bd7c1e1ec112542757ce1c9d6ddc7b33233d9086249876f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 82, + 629, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66456b913dae7c6cc6573418f1f1f1f76a3c6dc1966992b1dceae02476dd7c1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 82, + 756, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6931aab78375b367b04fe3be3b8ce7548f11074564e1b8e917515d62c86e3913.jpg", + "image_caption": [ + "Figure 7. Reuslts on post-tampering output naturalness. The black vertical dashed line shows the lowest $5\\%$ value of the confidence in the top-1 label of normal inputs." + ], + "image_footnote": [], + "bbox": [ + 763, + 82, + 887, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5bc4982cfd4ee41206862fd742b3abe7f2003a4c71c1ac819c16e8fc5ec8e595.jpg", + "table_caption": [ + "Table 2. Average generation computational cost (in seconds/GBs)." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-10GTSRBImageNet
MiSentry1641.43s/9.4GB1639.77s/11.3GB6112.89s/23.8GB
SSF1.39s/1.2GB1.42s/1.6GB184.13s/12.9GB
PublicCheck0.71s/1.1GB0.74s/1.3GB34.51s/6.7GB
SDBFK=20.20s/0.2GB0.21s/0.4GB8.48s/3.1GB
K=100.22s/0.2GB0.24s/0.4GB8.59s/3.1GB
IBSFK=20.19s/0.2GB0.21s/0.4GB8.44s/3.1GB
K=100.22s/0.2GB0.23s/0.4GB8.57s/3.1GB
", + "bbox": [ + 506, + 222, + 888, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rate for the unleaked fingerprints remains relatively high, with only a $12\\%$ decrease compared to the baseline. Such robustness of SDBF is attributed to the diversity of its source samples and sample-dependent $\\mathcal{C}_{sub}$ . We also conduct a stress test by gradually increasing the number of leaked fingerprint samples to 10,000, as shown in Fig. 6 (b). Despite the large number of leaked fingerprints, the detection rate of the unleaked fingerprints decreased but remained around $50\\%$ .", + "bbox": [ + 496, + 321, + 890, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we conduct an experiment to explore whether the steep decision boundary fingerprinting is applicable to the smoothed models such as adversarial training (AT) [32]. We find although AT slightly increased the optimization time cost of SDBF and slightly reduced its performance when $K = 2$ , SDBF remains efficient and effective when $K$ is larger. The detailed results are shown in the appendix.", + "bbox": [ + 496, + 443, + 892, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.7. Computational Cost", + "text_level": 1, + "bbox": [ + 500, + 555, + 691, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 compares the computational cost of SDBF with existing SOTA fingerprinting methods, evaluated on a single NVIDIA RTX 3090Ti GPU. We can observe that: (1) Increasing $K$ slightly raises the cost for both SDBF and IBSF; (2) SDBF incurs a computational overhead similar to IBSF, with both methods exhibiting minimal costs; and (3) Even with high-resolution ImageNet images and large-scale models like ViT, SDBF still remains highly efficient.", + "bbox": [ + 496, + 577, + 890, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel method for detecting hard-label tampering in DNN models. Our approach leverages the sensitive nature of steep decision boundary and maximizes fingerprint sample diversity over the decision boundary through the Max Boundary Coverage Strategy (MBCS). Through both theoretical analysis and extensive experiments, we have shown that SDBF outperforms existing SOTA hard-label fingerprinting techniques in both sensitivity and efficiency, making SDBF an effective and efficient solution for ensuring the integrity of cloud-deployed DNN models.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29285", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 78, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by National Natural Science Foundation of China under Grant No.62272175.", + "bbox": [ + 76, + 114, + 468, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 157, + 174, + 172 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amazon Web Services. Amazon sagemaker. https://aws.amazon.com/sagemaker, 2023. Last accessed March 20, 2023.", + "[2] Omid Aramoon, Pin-Yu Chen, and Gang Qu. Aid: Attesting the integrity of deep neural networks. In Proceedings of 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 19-24, 2021.", + "[3] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. A survey of model compression and acceleration for deep neural networks. arXiv preprint arXiv:1710.09282, 2017.", + "[4] PyTorch Contributors. Quantization. https://pytorch.org/docs/stable/quantization.html, 2022.", + "[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020.", + "[6] Amel Nestor Docena, Thomas Wahl, Trevor Pearce, and Yunsi Fei. Sensitive samples revisited: Detecting neural network attacks using constraint solvers. arXiv preprint arXiv:2109.03966, 2021.", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021.", + "[8] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304-9312, 2020.", + "[9] Google Cloud. AutoML. https://cloud.google.com/automl, 2023. Last accessed March 20, 2023.", + "[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017.", + "[11] Chaoxiang He, Xiaofan Bai, Xiaojing Ma, Bin Benjamin Zhu, Pingyi Hu, Jiayun Fu, Hai Jin, and Dongmei Zhang. Towards stricter black-box integrity verification of deep neural network models. In ACM Multimedia 2024, 2024.", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.", + "[13] Zecheng He, Tianwei Zhang, and Ruby Lee. Sensitive-sample fingerprinting of deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4729-4737, 2019.", + "[14] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In" + ], + "bbox": [ + 78, + 183, + 472, + 906 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 3104-3122, 2021.", + "[15] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of International Conference on Machine Learning, pages 1885-1894. PMLR, 2017.", + "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. M.S. thesis, Department of Computer Science, University of Toronto, 2009.", + "[17] Deepthi Praveenlal Kuttichira, Sunil Gupta, Dang Nguyen, Santu Rana, and Svetha Venkatesh. Verification of integrity of deployed deep learning models using bayesian optimization. Knowledge-Based Systems, 241:108238, 2022.", + "[18] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015.", + "[19] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018.", + "[20] Microsoft. Neural network intelligence. https://github.com/microsoft/nni, 2022.", + "[21] Microsoft Azure. Microsoft machine learning. https://azure.microsoft.com/en-us/products/machine-learning, 2023. Last accessed March 20, 2023.", + "[22] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11264-11272, 2019.", + "[23] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Bit-flip attack: Crushing neural network with progressive bit search. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1211-1220, 2019.", + "[24] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016.", + "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, C. Berg Alexander, and Fei-Fei Li. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015.", + "[26] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. Updates-leak: Data set inference and reconstruction attacks in online learning. In 29th USENIX Security Symposium, USENIX Security 2020, pages 1291-1308. USENIX Association, 2020.", + "[27] Ali Shafahi, W. Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information" + ], + "bbox": [ + 503, + 92, + 893, + 883 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29286", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing Systems 2018, NeurIPS 2018, pages 6106-6116, 2018.", + "[28] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural networks*, 32:323-332, 2012.", + "[29] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning fail? generalized transferability for evasion and poisoning attacks. In Proceedings of 27th USENIX Security Symposium (USENIX Security 18), pages 1299-1316, 2018.", + "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions, 2014.", + "[31] Shuo Wang, Sharif Abuadbba, Sidharth Agarwal, Kristen Moore, Ruoxi Sun, Minhui Xue, Surya Nepal, Seyit Camtepe, and Salil Kanhere. Publiccheck: Public integrity verification for services of run-time deep models. In Proceedings of 2023 IEEE Symposium on Security and Privacy (SP), pages 1348-1365. IEEE, 2023.", + "[32] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve" + ], + "bbox": [ + 78, + 90, + 470, + 416 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial training. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 36246-36263. PMLR, 23-29 Jul 2023.", + "[33] Bai Xiaofan, Chaoxiang He, Xiaojing Ma, Bin Benjamin Zhu, and Hai Jin. Intersecting-boundary-sensitive fingerprinting for tampering detection of DNN models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 54402-54413. PMLR, 21-27 Jul 2024.", + "[34] Zhaoxia Yin, Heng Yin, Hang Su, Xinpeng Zhang, and Zhenzhe Gao. Decision-based iterative fragile watermarking for model integrity verification. arXiv preprint arXiv:2305.09684, 2023.", + "[35] Renjie Zhu, Ping Wei, Sheng Li, Zhaoxia Yin, Xinpeng Zhang, and Zhenxing Qian. Fragile neural network watermarking with trigger image set. In Proceedings of International Conference on Knowledge Science, Engineering and Management, pages 280-293. Springer, 2021." + ], + "bbox": [ + 501, + 92, + 893, + 414 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29287", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_model.json b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..26d0076ef3fd105f7498465288d4ffee09e7b97f --- /dev/null +++ b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.002, + 0.808, + 0.025 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version:" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.031, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.203, + 0.858, + 0.258 + ], + "angle": 0, + "content": "Xiaofan Bai\\(^{1}\\), Shixin Li\\(^{1}\\), Xiaojing Ma\\(^{1}\\), Bin Benjamin Zhu\\(^{2}\\), Dongmei Zhang\\(^{2}\\), Linchen Yu\\(^{1*}\\) \n\\(^{1}\\)Huazhong University of Science and Technology \n\\(^{2}\\)Microsoft Corporation \n{xiaofanbai, shixinli, linda, linchenyu} @hust.edu.cn, {binzhu, dongmeiz} @microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.519 + ], + "angle": 0, + "content": "Cloud-based AI systems offer significant benefits but also introduce vulnerabilities, making deep neural network (DNN) models susceptible to malicious tampering. This tampering may involve harmful behavior injection or resource reduction, compromising model integrity and performance. To detect model tampering, hard-label fingerprinting techniques generate sensitive samples to probe and reveal tampering. Existing fingerprinting methods are mainly based on gradient-defined sensitivity or decision boundary, with the latter showing a manifest superior detection performance. However, all existing fingerprinting methods either suffer from insufficient sensitivity or incur high computational costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.474, + 0.749 + ], + "angle": 0, + "content": "In this paper, we theoretically analyze the black-box co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity. Based on this, we further propose Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight approach for hard-label tampering detection that inherently and efficiently combines the strengths of existing fingerprinting techniques. SDBF places fingerprint samples near the steep decision boundary, where the outputs of samples are inherently highly sensitive to tampering. We also design a Max Boundary Coverage Strategy (MBCS), which enhances samples' diversity over the decision boundary. Theoretical analysis and extensive experimental results show that SDBF outperforms existing SOTA hard-label fingerprinting methods in both sensitivity and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.763, + 0.21, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.472, + 0.88 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have demonstrated excellent capabilities across various domains, such as computer vision [12, 18, 24]. An increasing number of applications have incorporated DNN models. Machine Learning as a Service (MLaaS) platforms, such as Google AutoML [9], Microsoft Azure ML [21], and Amazon SageMaker [1]," + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.291, + 0.891, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.558, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Overview of potential model tampering and public tampering detection for models deployed in MLaaS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.642 + ], + "angle": 0, + "content": "leverage powerful pre-trained DNN models, allowing users with limited computational resources to access cutting-edge technology." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.825 + ], + "angle": 0, + "content": "However, models deployed in the cloud often operate beyond the direct control of their owners, leaving them vulnerable to unauthorized/illegal model tampering as shown in Fig. 1. For instance, unscrupulous service providers might compress a model—reducing its memory and computational footprint—as a cost-saving measure without the owner's consent, as demonstrated in prior work [3, 22]. Alternatively, adversaries could inject malicious behaviors into the model's parameters to manipulate the model's outputs [10, 14, 15, 27, 29]. Such tampering not only compromises the model's integrity, security, and overall performance but also erodes users' trust and the reliability of the system." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Many works [2, 6, 13, 17, 31, 33-35] have focused on detecting model tampering. Due to the black-box nature of the models deployed in the cloud, hard-label fingerprinting techniques are the most practical but also most challenging among existing tampering detection methods as their non" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29278" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "invasive and black-box nature aligns well with the MLaaS in practice. Existing hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]. The former enhances the sensitivity of fingerprint samples' output to model tampering by maximizing gradient-defined sensitivity, which is the \\( L_{p} \\) norm of the gradient of the output with respect to the model parameters, and these fingerprinting techniques have been widely demonstrated to be much less effective than the latter. The latter improves the sensitivity of fingerprint samples to tampering by minimizing the distance to the decision boundary. In summary, both of these approaches are still not sensitive enough to detect certain types of model tampering, especially subtle ones. However, detecting subtle model tampering is crucial, as the end user may need to confirm if the model they are using is really the one they want to use." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Additionally, [11] introduced a meta-learning-based fingerprinting approach. However, due to the inherent complexity of meta-learning, this method incurs significantly higher overhead compared to the two aforementioned categories, making it impractical for resource-constrained model owners or users. We argue that, given the very purpose of black-box model tampering detection, a lightweight design for fingerprinting method is essential for its widespread application." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.802 + ], + "angle": 0, + "content": "In this paper, recognizing the importance of closing to decision boundary for tampering detection with fingerprint samples, we theoretically analyze the co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity in a black-box detection manner. Based on this, we further propose a novel lightweight hard-label tampering detection fingerprinting approach, Steep-Decision-Boundary Fingerprinting (SDBF) to improve the fingerprint samples' sensitivity with the inherent highly sensitive nature of samples near the steep decision boundary. Specifically, SDBF places fingerprint samples close to the steep decision boundary, where the outputs exhibit high sensitivity to the tampering-induced shift of decision boundary. To achieve this, SDBF perturbs normal input samples to maximize the \\( L_{2} \\) norm of the logit output while simultaneously minimize its variance. The fingerprint samples are bound to be almost indistinguishable from the original normal samples. Moreover, to further enhance detection performance when using multiple fingerprint samples to detect, we propose a Max Boundary Coverage Strategy (MBCS). This strategy enables the fingerprint samples to explore a wide range of different regions around the decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.802, + 0.473, + 0.878 + ], + "angle": 0, + "content": "Our theoretical analysis and extensive experimental results demonstrate that the sensitivity of SDBF is superior to that of existing state-of-the-art (SOTA) hard-label fingerprinting methods while being time and memory-efficient. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.473, + 0.902 + ], + "angle": 0, + "content": "- We theoretically analyze the black-box co-optimal tam" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.894, + 0.138 + ], + "angle": 0, + "content": "pering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.148, + 0.895, + 0.224 + ], + "angle": 0, + "content": "- We introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight fingerprinting method that inherently and efficiently combines the strengths of fingerprinting techniques based on gradient-defined sensitivity and decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.235, + 0.895, + 0.311 + ], + "angle": 0, + "content": "- We propose a Max Boundary Coverage Strategy (MBCS) to ensure that the fingerprint samples widely explore the decision boundary, further enhancing the detection performance when using multiple fingerprint samples to detect model tampering." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.321, + 0.895, + 0.382 + ], + "angle": 0, + "content": "- We present both theoretical analysis and extensive experimental results that demonstrate the superior performance and efficiency of SDBF compared to existing SOTA hard-label fingerprinting methods." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.895, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.396, + 0.642, + 0.412 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.794, + 0.437 + ], + "angle": 0, + "content": "2.1. Hard-Label Tampering Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Tampering detection for DNN models aims to determine whether a model has been altered. Since MLaaS operates as a black-box to users, existing methods focus on hard-label tampering detection, where only the top-1 label is accessible. In this setting, watermarking and fingerprinting are widely used. Watermarking [35], an invasive method, embeds a fragile watermark during training that can later be queried with specific inputs to detect tampering. However, because modifying the model may impact performance, non-invasive fingerprinting techniques, the focus of this paper, have attracted more attention." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.844, + 0.637 + ], + "angle": 0, + "content": "2.2. Fingerprinting for Tampering Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.689 + ], + "angle": 0, + "content": "Existing black-box hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.856 + ], + "angle": 0, + "content": "Gradient-defined sensitivity. Sensitive-Sample Fingerprinting (SSF) [13] introduces fingerprinting for detecting tampering in DNN models for the first time by gradient-defined sensitivity. SSF generates sensitive samples—natural-looking inputs that cause large output changes when the model's parameters are slightly altered. Following SSF, symbolic constraint solvers [6] and Bayesian optimization (BO) [17] have been used to optimize the generation of sensitive samples. However, such sensitive samples have been proven to be ineffective in extensive experiments for many kinds of model tampering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Decision boundary based sensitivity. PublicCheck [31] uses VQVAE models to generate fingerprint samples by carefully adding perturbations along with certain directions" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29279" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.362 + ], + "angle": 0, + "content": "in the latent space of VQVAE and stopping when it has just crossed the decision boundary of two different categories. Others like AID [2] and DBFW [34] generate fingerprint samples close to the decision boundary of all categories by minimizing output variance using mean square error (MSE) loss functions. Similar to AID and DBFW, IBSF [33] employs a Shannon entropy loss to generate diverse fingerprint samples near different \\( K(K \\geqslant 2) \\) categories' decision boundary. While these fingerprinting techniques show significantly superior detection performance than the aforementioned fingerprinting techniques based on gradient-defined sensitivity, they are still not sensitive enough to detect certain types of model tampering, especially subtle ones. Notably, neither one has theoretically justified why samples that are close to the decision boundary are more effective than the faraway, nor has anyone theoretically analyzed the co-optimal tampering detection sensitivity of fingerprint samples in these two contexts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.486 + ], + "angle": 0, + "content": "Additionally, [11] proposed a meta-learning-based sensitive fingerprinting method, MiSentry, but due to the inherent high overhead of meta-learning, the cost of MiSentry is significantly higher than the two aforementioned categories, making it unaffordable for those resources-limited model owner/users. We believe that due to the very purpose of black-box tampering detection, the lightweight of a fingerprinting method is crucial for its widespread application." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.218, + 0.517 + ], + "angle": 0, + "content": "3. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.786 + ], + "angle": 0, + "content": "Aligned with existing works, we adopt a threat model that includes white-box generation and black-box detection. Specifically, during the fingerprint samples generation, we assume white-box access to the target DNN, using available normal samples as source data. This assumption is reasonable as fingerprint samples are generated by the model owner. In the detection phase, we assume there is only black-box access to the suspect model, where only an API-based query is allowed, and only the top-1 label of the input query will be returned. Moreover, following [31], we also assume a trustworthy third party that securely stores and distributes the fingerprint samples for public tampering detection, where the DNN service provider is assumed to be untrustworthy and may detect probing fingerprint samples, evading tampering detection. Additionally, adversaries might access some fingerprint samples through some illegal means, and use them to bypass tampering detection." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.803, + 0.44, + 0.82 + ], + "angle": 0, + "content": "4. Steep Decision Boundary and Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this section, we first define the decision boundary in the context of the classification task. We then theoretically demonstrate why samples near the decision boundary are sensitive to shifts caused by model tampering. Next, we define the steep decision boundary and analyze how to place" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.122 + ], + "angle": 0, + "content": "fingerprint samples near it to further enhance samples' sensitivity while improving optimization efficiency." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.13, + 0.73, + 0.146 + ], + "angle": 0, + "content": "4.1. Steep Decision Boundary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.894, + 0.229 + ], + "angle": 0, + "content": "A DNN is a function \\( f \\), parameterized by \\( W \\), which maps an input \\( x \\) to an output logits vector \\( \\mathbf{y} = f(W, x) \\). For classification tasks, let \\( N \\) denotes the number of classes, the logits vector \\( \\mathbf{y} \\) will then be transformed into a probability vector \\( Y \\) by the following Softmax function:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.236, + 0.893, + 0.273 + ], + "angle": 0, + "content": "\\[\nY _ {i} = \\operatorname {S o f t m a x} \\left(\\mathbf {y} _ {i}\\right) = \\frac {e ^ {\\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {N} e ^ {\\mathbf {y} _ {j}}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.28, + 0.744, + 0.294 + ], + "angle": 0, + "content": "where \\(Y_{i}\\) is the confidence of class \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.894, + 0.431 + ], + "angle": 0, + "content": "Definition 4.1 (Decision Boundary). For a classification model, the decision boundary is the hypersurface that divides the feature space into regions, each corresponding to a different class label. Typically, the model assigns the same probabilities to the classes on the decision boundary, and the confidences of these categories will be much greater than the other categories. Specifically, for an \\( N \\)-class classification model, the decision boundary of \\( K(K \\leqslant N) \\) classes can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.443, + 0.893, + 0.465 + ], + "angle": 0, + "content": "\\[\nY _ {1} = Y _ {2} = \\dots = Y _ {K} \\gg \\sum \\tilde {Y} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.893, + 0.506 + ], + "angle": 0, + "content": "where \\(\\sum \\tilde{Y} (0\\leqslant \\sum \\tilde{Y}\\leqslant 1)\\) is the summation of the other \\(N - K\\) categories' confidence and \\(\\begin{array}{r}\\sum_{i = 1}^{K}Y_{i} + \\sum \\tilde{Y} = 1. \\end{array}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.895, + 0.663 + ], + "angle": 0, + "content": "For black-box tampering detection, the predicted hard label based on the Softmax output \\( Y \\) is the only observable feedback from the suspect model. To maximize the sensitivity of fingerprint samples in this setting, we need to maximize the sensitivity of \\( Y \\) with respect to \\( \\mathbf{y} \\) to force the model to work in Softmax high-sensitivity region, making any tampering-induced \\( \\Delta \\mathbf{y} = f(W + \\Delta W, x) - f(W, x) \\) will be nonlinearly amplified to \\( Y \\) by Softmax, avoiding being suppressed by the saturation zone of Softmax when the fingerprint samples are faraway from the decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.894, + 0.729 + ], + "angle": 0, + "content": "Proposition 4.2. Given a fingerprint sample \\( x \\), the sensitivity of its softmax output \\( \\mathbf{Y} \\) with respect to \\( ||\\Delta \\mathbf{y}||_2 \\) introduced by model tampering \\( \\Delta W \\) will reach the maximum when \\( x \\) close to the decision boundary, where \\( Y_i = \\frac{1}{N} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.81 + ], + "angle": 0, + "content": "Proof. To prove Prop.4.2, we need to analyze the Jacobian matrix \\(J\\) of vector \\(\\mathbf{Y}\\) with respect to vector \\(\\mathbf{y}\\) and the maximum value of the Frobenius norm (equivalent to the \\(L_{2}\\) norm) of \\(J\\). First, based on Eq.1, the element \\(J_{ij}\\) of \\(J\\) is the partial derivative of \\(Y_{i}\\) with respect to \\(y_{j}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.817, + 0.893, + 0.85 + ], + "angle": 0, + "content": "\\[\nJ _ {i j} = \\frac {\\partial Y _ {i}}{\\partial \\mathbf {y} _ {j}} = Y _ {i} \\left(\\delta_ {i j} - Y _ {j}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.857, + 0.894, + 0.905 + ], + "angle": 0, + "content": "where \\(\\delta_{ij}\\) is the Kronecker Delta function and \\(\\left\\{ \\begin{array}{l}\\delta_{ij} = 1,i = j\\\\ \\delta_{ij} = 0,i\\neq j \\end{array} \\right.\\). Thus, the squared Frobenius norm of \\(J\\)" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29280" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.224, + 0.106 + ], + "angle": 0, + "content": "can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.114, + 0.47, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\| J \\| _ {F} ^ {2} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(J _ {i j}\\right) ^ {2} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(Y _ {i} \\left(\\delta_ {i j} - Y _ {j}\\right)\\right) ^ {2} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - 2 Y _ {i} + Y _ {i} ^ {2} + \\sum_ {j \\neq i} Y _ {j} ^ {2}\\right) \\tag {4} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - 2 Y _ {i} + Y _ {i} ^ {2}) + Y _ {i} ^ {2} \\sum_ {j \\neq i} Y _ {j} ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.257, + 0.47, + 0.293 + ], + "angle": 0, + "content": "Since \\(\\sum_{i=1}^{N} Y_{i} = 1\\), thus \\(\\sum_{j \\neq i} Y_{j}^{2}\\) can be replaced by \\(1 - Y_{i}^{2}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.303, + 0.47, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\| J \\| _ {F} ^ {2} = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - 2 Y _ {i} + Y _ {i} ^ {2}\\right) + Y _ {i} ^ {2} \\left(1 - Y _ {i} ^ {2}\\right) \\tag {5} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - Y _ {i}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.473 + ], + "angle": 0, + "content": "Next, we need to find the maximum of Eq.5. For such a constrained optimization problem, we use the Lagrange multiplier method to rigorously prove when the maximum of Eq.5 is achieved. Specifically, we introduce a Lagrange multiplier \\(\\lambda\\) for the constraint \\(\\sum_{i=1}^{N} Y_i = 1\\), which is:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.483, + 0.47, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(Y _ {1}, Y _ {2}, \\dots , Y _ {N}, \\lambda\\right) = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - Y _ {i}\\right) + \\lambda \\left(\\sum_ {i = 1} ^ {N} Y _ {i} - 1\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.537, + 0.47, + 0.582 + ], + "angle": 0, + "content": "We now compute the partial derivatives of \\(\\mathcal{L}\\) with respect to each \\(Y_{i}\\). The partial derivative of \\(Y_{i}^{2}(1 - Y_{i})\\) with respect to each \\(Y_{i}\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.604, + 0.47, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial Y _ {i}} \\left(Y _ {i} ^ {2} \\left(1 - Y _ {i}\\right)\\right) = 2 Y _ {i} \\left(1 - Y _ {i}\\right) - Y _ {i} ^ {2} = 2 Y _ {i} - 3 Y _ {i} ^ {2} (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.645, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Similarly, for the constraint term, the partial derivative is:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.67, + 0.47, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial Y _ {i}} \\left(\\sum_ {i = 1} ^ {N} Y _ {i} - 1\\right) = 1 \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.721, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Let the total derivative of the Lagrangian equal to zero, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.761, + 0.47, + 0.78 + ], + "angle": 0, + "content": "\\[\n2 Y _ {i} - 3 Y _ {i} ^ {2} + \\lambda = 0 \\Longrightarrow 2 Y _ {i} - 3 Y _ {i} ^ {2} = - \\lambda \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.79, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Recall that we have the constraint \\(\\sum_{i=1}^{N} Y_i = 1\\), to find a feasible solution, since Eq.9 holds for each \\(Y_i\\), we can conclude that all \\(Y_i\\) must be equal, meaning \\(Y_1 = Y_2 = \\dots = Y_N = Y\\). Otherwise, the right term can not be the same constant \\(-\\lambda\\). Then, we can obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.876, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\nN \\cdot Y = 1 \\quad \\Rightarrow \\quad Y = \\frac {1}{N} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.153 + ], + "angle": 0, + "content": "Finally, to verify that this solution is indeed the maximum, we can check the second order derivative value of \\(||J||_F^2\\) at this solution. When all \\(Y_{i} = \\frac{1}{N}\\), the second order derivative of \\(||J||_F^2\\) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.164, + 0.894, + 0.204 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial^ {2} \\left| \\left| J \\right| \\right| _ {F} ^ {2}}{\\partial Y _ {i} ^ {2}} = \\sum_ {i = 1} ^ {N} \\frac {\\partial 2 Y _ {i} - 3 Y _ {i} ^ {2}}{\\partial Y _ {i}} = 2 - 6 \\sum_ {i = 1} ^ {N} Y _ {i} < 0. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.215, + 0.894, + 0.26 + ], + "angle": 0, + "content": "Therefore, this solution is indeed the maximum. In summary, given a fingerprint sample, \\( \\| J\\| _F^2 \\) achieves maximum when it is close to the decision boundary defined in Def.4.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.26, + 0.894, + 0.351 + ], + "angle": 0, + "content": "However, when \\( x \\) is faraway from the decision boundary, some elements in \\( f(W,x) \\) are much larger than others, leading to \\( \\text{softmax}(f(W,x)) \\) that are close to 0 or 1. In this case, the elements of the Jacobian matrix \\( J_{ij} \\) will be significantly different, especially when \\( Y_{i} \\approx 1 \\), causing the most of other \\( Y_{j} \\approx 0 \\), that is:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.362, + 0.893, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} Y _ {i} \\approx 1 \\\\ Y _ {j} \\approx 0 \\end{array} \\right. \\Longrightarrow \\left\\{ \\begin{array}{c} J _ {i i} \\approx Y _ {i} \\cdot (1 - Y _ {i}) \\approx 0 \\\\ J _ {i j} \\approx - Y _ {i} \\cdot Y _ {j} \\approx 0, i \\neq j \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.407, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Hence, \\(||J||_F^2\\) in this scenario will be significantly smaller as most elements are close to 0." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.448, + 0.894, + 0.51 + ], + "angle": 0, + "content": "Definition 4.3 (Steep Decision Boundary). Given a fingerprint sample \\( x \\) and a model tampering \\( \\Delta W \\), the steep decision boundary refers to the decision boundary where \\( f(W, x) \\) shows a rapid change when \\( \\Delta W \\) is applied to \\( W \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.519, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Proposition 4.4. Give a sample \\( x \\) and a DNN model \\( f(\\cdot) \\) parameterized by \\( W \\), the sensitivity of output \\( f(W, x) \\) with respect to \\( \\Delta W \\) is positively correlated to the \\( ||f(W, x)||_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.574, + 0.894, + 0.635 + ], + "angle": 0, + "content": "Proof. Inspired by [13], since all model tampering inevitably result in different performances in the last layer, the sensitivity of \\( f(W,x) \\) to \\( W \\) practically can be formulated as the following Frobenius norm of Jacobian matrix \\( \\hat{J} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.645, + 0.892, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\left| \\left| \\hat {J} \\right| \\right| _ {F} ^ {2} = \\left| \\left| \\frac {\\partial f (W , x)}{\\partial W _ {F C}} \\right| \\right| _ {2} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where \\( W_{FC} \\) is the weights of the last fully connected layer of \\( f(\\cdot) \\). Since all existing classification models do not use any nonlinear activation functions like \\( ReLu(\\cdot) \\) in the last layer, the last layer of the model performs the following linear operations on the output of the previous layer:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.774, + 0.892, + 0.792 + ], + "angle": 0, + "content": "\\[\nf (W, x) = W _ {F C} \\times f _ {- 2} (W, x) + \\mathbf {b} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.803, + 0.894, + 0.847 + ], + "angle": 0, + "content": "where \\( f_{-2}(W,x) \\) and \\( \\mathbf{b} \\) are the output and bias of the penultimate layer, respectively. Substituting Eq. 14 into Eq. 13, we can get:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.847, + 0.892, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\left| \\left| \\hat {J} \\right| \\right| _ {F} ^ {2} = \\left| \\left| f _ {- 2} (W, x) \\right| \\right| _ {2} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Combining Eq.15 and Eq.14, when we increase \\(||f(W,x)||_2\\), the sensitivity of \\(f(W,x)\\) with respect to \\(\\Delta W\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29281" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.472, + 0.187 + ], + "angle": 0, + "content": "\\(\\| \\hat{J}\\| _F^2\\) , will increase simultaneously. To valid Prop.4.4, we conduct an experiment by maximizing \\(\\| f(W,x)\\| _2\\) while recording \\(\\| \\frac{\\partial f(W,x)}{\\partial\\mathbf{W}_{FC}}\\| _2\\) for a ResNet20 trained on CIFAR10, as shown in Fig. 2. The results are highly consistent with Prop.4.4. Therefore, we can improve the sample's sensitivity of \\(f(W,x)\\) to \\(\\Delta W\\) by increasing \\(\\| f(W,x)\\| _2\\)" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.208, + 0.468, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.392, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Validation results on the relationship between \\( ||f(W,x)||_2 \\) and \\( ||\\frac{\\partial f(W,x)}{\\partial W_{FC}}||_2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.472, + 0.537 + ], + "angle": 0, + "content": "Proposition 4.5. Give a fingerprint sample \\( x \\) and a DNN model \\( f(\\cdot) \\) parameterized by \\( W \\), simultaneously increasing the sensitivity of softmax \\( (f(W, x)) \\) to corresponding \\( f(W, x) \\) and the sensitivity of \\( f(W, x) \\) to \\( \\Delta W \\) can drive \\( x \\) close to the steep decision boundary and thus further enhance the sensitivity of softmax \\( (f(W, x)) \\) to \\( \\Delta W \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.472, + 0.646 + ], + "angle": 0, + "content": "Prop.4.5 can be easily proved with the linear translational invariance of \\( \\text{Softmax}(\\cdot) \\). More details are in the appendix. Based on Def.4.3 and Prop.4.5, we know how to efficiently place the fingerprint samples near steep decision boundary, where they are highly sensitive to model tampering \\( \\Delta W \\) as it inherently combining the strength of fingerprinting methods based on gradient-defined sensitivity and decision boundary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.445, + 0.669 + ], + "angle": 0, + "content": "5. Steep-Decision-Boundary Fingerprinting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.676, + 0.471, + 0.753 + ], + "angle": 0, + "content": "In this section, we introduce a novel sensitive fingerprinting method named Steep-Decision-Boundary Fingerprinting (SDBF). Our goal is to leverage the highly sensitive nature of the samples near the steep decision boundary to detect model tampering, as shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.364, + 0.773 + ], + "angle": 0, + "content": "5.1. Fingerprint Samples Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Different from [13], we do not directly optimize Eq.13 and Eq.5 as their poor efficiency. We will further explain this in the appendix. Fortunately, based on Prop.4.2, Prop. 4.4 and Prop. 4.5, we can maximize the fingerprint samples' sensitivity by maximizing \\(||\\mathbf{y}||_2\\) while minimizing the variance of \\(\\mathbf{y}\\), \\(\\sigma (\\mathbf{y})\\). Similar to IBSF [33], we can also freely choose to place fingerprint samples at \\(K(K\\leqslant N)\\) categories' decision boundary, which will help increase the diversity of" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.096, + 0.696, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.099, + 0.891, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.245, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Figure 3. Illustration of SDBF, gradient-defined fingerprinting SSF, and decision-boundary-based fingerprinting IBSF. SDBF shows an inherent high sensitivity to tampering-introduced boundary shift." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.895, + 0.356 + ], + "angle": 0, + "content": "the fingerprint samples. Specifically, let \\(\\mathbf{y}_{\\mathcal{C}_{sub}}\\) denote the logits output of the chosen categories and \\(\\overline{\\mathbf{y}}_{\\mathcal{C}_{sub}} = \\frac{1}{K}\\sum_{i=1}^{k}\\mathbf{y}_i\\) denote the averaged logits of the categories in \\(\\mathcal{C}_{sub}\\), our objective function for a sample \\(x\\) can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.375, + 0.892, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (x, \\mathcal {C} _ {s u b}) = \\sigma (\\mathbf {y} _ {\\mathcal {C} _ {s u b}}) - \\lambda \\cdot | | \\mathbf {y} _ {\\mathcal {C} _ {s u b}} | | _ {2} \\\\ = \\frac {1}{K} \\cdot \\sum_ {i = 1} ^ {K} \\left(\\mathbf {y} _ {i} - \\overline {{\\mathbf {y}}} _ {\\mathcal {C} _ {s u b}}\\right) ^ {2} - \\lambda \\cdot \\sum_ {i = 1} ^ {K} \\left\\| \\mathbf {y} _ {i} \\right\\| _ {2} \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.489 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the coefficient and \\(i\\in \\mathcal{C}_{sub}\\). Thus, given a normal sample \\(x\\), we can drive it close to the steep decision boundary by minimizing the objective function in Eq.16:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.499, + 0.892, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\Delta x} \\mathcal {L} (x + \\Delta x, \\mathcal {C} _ {s u b}) \\tag {17} \\\\ s. t. | | \\Delta x | | _ {p} \\leqslant \\epsilon \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.895, + 0.651 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the input bound. The constraint \\(\\| \\Delta x\\| _p\\leqslant \\epsilon\\) bounds the perturbation \\(\\Delta x\\) in the input space, making the generated fingerprint sample perceptually similar to the original normal one. This naturalness makes it hard for an adversary to distinguish whether the input is a probe query by input anomaly detection. We solve Eq. 17 by using projected gradient descent (PGD) like [19]:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.666, + 0.892, + 0.682 + ], + "angle": 0, + "content": "\\[\nx _ {i + 1} = \\operatorname {C l i p} _ {\\epsilon} \\left(x _ {i} + \\eta \\cdot \\nabla_ {x _ {i}} \\mathcal {L} \\left(x _ {i}, \\mathcal {C} _ {\\text {s u b}}\\right)\\right) \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.716 + ], + "angle": 0, + "content": "where \\( x_{i} = x_{i - 1} + \\Delta x_{i} \\) is the generated fingerprint sample at the \\( i \\)-th iteration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.822 + ], + "angle": 0, + "content": "The optimization eventually yields a fingerprint sample positioned close to \\( \\mathcal{C}_{sub} \\) steep decision boundary. Following [33], to reduce the perturbation effort, we select the top \\( |\\mathcal{C}_{sub}| \\) confident categories in \\( \\mathbf{y} \\) of original normal \\( x \\) as \\( \\mathcal{C}_{sub} \\) by default. This sample-dependent \\( \\mathcal{C}_{sub} \\) will enhance the diversity of generated sensitive fingerprint samples, making it more robust to partial leakage." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.84, + 0.871, + 0.887 + ], + "angle": 0, + "content": "Remark. In the generation phase, except for fingerprint sample \\( x \\), we will also collect its corresponding top-1 label as the fingerprint label \\( \\mathcal{V} \\). In the de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29282" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.102, + 0.1, + 0.445, + 0.162 + ], + "angle": 0, + "content": "tection phase, \\(x\\) will be used to query the suspect model to get the returned hard label \\(\\mathcal{V}_{predicted}\\) and then we will determine whether the model has been tampered with by verifying whether \\(\\mathcal{V}_{predicted} = \\mathcal{V}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.182, + 0.376, + 0.199 + ], + "angle": 0, + "content": "5.2. Max Boundary Coverage Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.205, + 0.471, + 0.388 + ], + "angle": 0, + "content": "When probing the tampered model with \\( M \\) (\\( M > 1 \\)) fingerprint samples, we collect their returned hard labels. The model is considered untampered if all the labels match the corresponding fingerprint labels. To maximize the diversity of the \\( M \\) fingerprint samples over the steep decision boundary, we propose the Max Boundary Coverage Strategy (MBCS) to make them widely explore the decision boundary. Specifically, we first randomly select a different fingerprint sample, denoted as \\( x_{1} \\). Then, let \\( m \\leqslant M \\) and \\( x_{i} \\) denote the number of selected fingerprint samples and the sample to select in \\( i \\)-th iteration, we iteratively choose the remaining \\( M - 1 \\) samples to maximize the boundary diversity score:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.399, + 0.47, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\max _ {x _ {i}} \\mathcal {L} _ {d i v} = \\sum_ {i = 1} ^ {N} c _ {i} \\cdot \\log c _ {i} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.351, + 0.46 + ], + "angle": 0, + "content": "where \\(N\\) is the number of categories and" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.472, + 0.47, + 0.512 + ], + "angle": 0, + "content": "\\[\nc _ {i} = \\frac {1}{m} \\cdot \\sum_ {j = 1} ^ {m} \\operatorname {s o f t m a x} _ {i} \\left(f \\left(x _ {j}\\right)\\right) \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.553 + ], + "angle": 0, + "content": "where \\( \\text{softmax}_i(\\cdot) \\) denotes softmax operation along with \\( i - th \\) dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.66 + ], + "angle": 0, + "content": "By repeating the \\(MBCS\\mathcal{R}\\) times, we can get a \\(R\\)-length and \\(M\\)-width probe fingerprint sequence. For each probe query, we will send \\(M\\) probe fingerprint samples to the suspect model until we go through all \\(\\mathcal{R} \\times M\\) fingerprint samples. In our experiments, we use \\(MBCS\\) as a post-processing technique to enhance the diversity of the fingerprint samples over the decision boundary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.312, + 0.687 + ], + "angle": 0, + "content": "6. Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.357, + 0.712 + ], + "angle": 0, + "content": "6.1. Datasets, Models, and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.84 + ], + "angle": 0, + "content": "To evaluate the performance of SDBF, we use various common classification datasets, including CIFAR10 [16], GTSRB [28], and ImageNet [25]. To evaluate the generalizability of SDBF across different model architectures, we select three different popular architectures for the above three datasets. Specifically, we use ResNet20 [12] for CIFAR10, GoogLeNet [30] for GTSRB, and vision transformer ViT-B-32 [7] pretrained by Pytorch for ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To assess the superiority of SDBF compared to existing SOTA fingerprinting methods, we select four SOTA fingerprinting methods, SSF [13], PublicCheck [31], IBSF [33], and MiSentry [11], as the baselines. We utilize the default" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "settings outlined in their respective papers unless explicitly specified otherwise. For each fingerprinting method, we independently generate 1000 fingerprint samples." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.142, + 0.837, + 0.159 + ], + "angle": 0, + "content": "6.2. Tampering Settings and Detection Rate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.165, + 0.894, + 0.391 + ], + "angle": 0, + "content": "Tampering Settings. In our experiments, we evaluate SDBF and the baselines on the 9 types of model tampering selected from BIVBench [11], including 4 hardest-to-detect model tampering, targeted attack [15,27,29], finetuning the last layer with \\(10^{-5}\\) learning rate, and Unlearning [8]/Online Learning (On-Learning) [26] makes a model forget/newly learn specific data, and 5 common model tampering consists of Backdoor [10], Bit-flipping Attack [23], Quantization [4], Pruning [20], and independently trained models with Different Architecture (Different Arch.) on the same dataset. All the tampering settings are aligned with BIVBench [11]. In this paper, the reported experimental results for each tampering type are averaged over 10 independently tampered models to mitigate randomness in evaluating tampering detection performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.392, + 0.894, + 0.619 + ], + "angle": 0, + "content": "Tampering Detection Rate. For each tamper detection probe, \\( M \\) (\\( M \\geqslant 1 \\)) fingerprint samples are selected from a pool of 1000 fingerprint samples to query the test model. The model's top-1 predicted labels are then compared with the corresponding fingerprint labels. If all predicted labels match the fingerprint labels, the model is classified as tampered; otherwise, it is marked as tampered. This procedure is repeated 100 times for each test model. Since 10 tampered models are generated for each tampering type, a total of 1000 tests are performed for each tampering type. The tampering detection rate is the proportion of the 1000 probe queries in which tampering is successfully detected. Note that there are no false positives (FP) in the tampering detection of classification model, as the classification model \\( f(\\cdot) \\) provides a stable mapping from input \\( x \\) to output \\( f(x) \\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.62, + 0.864, + 0.636 + ], + "angle": 0, + "content": "6.3. Perceptual Quality of Fingerprint Samples" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.794 + ], + "angle": 0, + "content": "Fig. 4 shows a subset of fingerprint samples generated by SDBF and four baselines, all constrained by the same \\( L_{2} \\) norm bound. Fingerprint samples generated by SDBF closely resemble those from SSF, IBSF, and MiSentry, with imperceptible pixel-level perturbations added to the original images. In contrast, fingerprint samples generated by PublicCheck exhibit noticeable artifacts, with semantic-level alterations to the original images, due to the use of generative models. For example, the distorted face in the 8th column of Fig. 4 highlights this effect." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.822, + 0.818 + ], + "angle": 0, + "content": "6.4. Effectiveness of Tampering Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Single Fingerprint Sample Detection. We first evaluate the detection performance in the single fingerprint sample detection setting, which only uses one fingerprint sample to query the suspect model. Specifically, following [33], we comprehensively evaluate SDBF and IBSF with \\( K \\) ranging" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29283" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 1. Tamper detection rate (%) of detecting tampering in a single fingerprint sample setting on CIFAR10 and GTSRB. For fine-tuning(FT), \"Last\" specifies only fine-tuning the last layer, with numbers indicating the \\( {10}^{-5} \\) learning rate. The \"On-Learning\" refers to online learning." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.122, + 0.888, + 0.415 + ], + "angle": 0, + "content": "
DatasetMethodFTLast-5BackdoorT-AttackUnlearningQuantizationDifferent Arch.PruningOn-LearningBit-flipping
CIFAR10SSF0.064.911.43.266.379.766.76.784.4
MiSentry45.986.483.868.490.793.589.568.288.9
PublicCheck21.673.636.928.360.865.169.525.485.7
IBSFK=224.577.442.537.578.283.381.247.587.8
K=435.983.655.359.489.188.588.356.289.4
K=642.887.563.569.290.490.890.569.588.6
K=844.589.466.171.390.992.791.371.689.5
K=1049.490.973.272.891.393.891.976.490.1
SDBF(ours)K=238.181.456.456.586.685.384.158.188.4
K=447.587.368.669.889.791.591.266.989.2
K=653.189.979.773.491.592.892.373.588.3
K=856.291.483.175.191.993.792.975.890.3
K=1062.892.585.877.393.294.693.681.189.8
GTSRBSSF0.059.89.51.462.664.559.12.785.1
MiSentry35.286.475.159.487.979.583.759.889.3
PublicCheck18.467.837.321.769.266.570.428.387.3
IBSFK=222.776.339.232.879.872.177.542.688.5
K=431.979.952.955.986.179.783.257.390.4
K=638.485.859.861.889.183.385.563.490.1
K=841.588.563.470.191.285.988.465.189.6
K=1042.891.169.773.391.888.490.368.291.3
SDBF(ours)K=229.180.949.849.484.777.282.953.990.3
K=438.586.662.566.890.584.589.464.391.5
K=646.790.378.471.592.289.691.772.789.2
K=853.291.581.573.692.991.391.977.991.9
K=1057.192.784.678.293.693.892.880.190.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.471, + 0.517 + ], + "angle": 0, + "content": "from \\( K = 2 \\) to \\( K = N = 10 \\) and \\( K = 2 \\) to \\( K = 500 \\) on CIFAR10 and ImageNet, respectively. Tab.1 summarizes the tampering detection rates on CIFAR10 and GTSRB of existing SOTA fingerprinting methods and SDBF. The detailed results on ImageNet are in the appendix. From the results, we can conclude that:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.717 + ], + "angle": 0, + "content": "SDBF generally outperforms the four SOTA baselines, although exceptions arise for SDBF at \\( K = 2 \\) when using a single fingerprint sample to perform tampering detection, which can be attributed to MiSentry's inherent sensitivity to such tampering as some subtle tampering is in its model zoo of meta-learning. As \\( K \\) increases, SDBF gradually outperforms MiSentry, with improvements of up to \\( 16.9\\% \\) on CIFAR10 at \\( K = 10 \\) for fine-tuning the last layer with \\( 10^{-5} \\) learning rate compared to MiSentry. The detection rate of IBSF shares a similar trend with our SDBF when \\( K \\) increases, yet it is consistently performing worse than SDBF, evidencing our key intuition: the fingerprint samples close to the steep decision boundary are more sensitive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Multiple Fingerprint Samples Detection. Next, we evaluate the detection performance in multiple fingerprint samples detection, which uses \\( M(M > 1) \\) fingerprint samples to query the tampered model. If \\( \\mathcal{V}_{predicted} \\) of any one fingerprint sample in the \\( M \\) samples differ from the corresponding \\( \\mathcal{V} \\), the model will be deemed as a tampered one. Following the detection setting in IBSF, MiSentry, and PublicCheck, we repeat the detection probe 1000 times by selecting \\( M \\) random fingerprint samples for them. As for SDBF and SSF, we use MBCS and Max Neuron Coverage Strategy (MNCS) [13], which is specifically designed for SSF to select \\( M \\) fingerprint samples for each probe, respectively." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.424, + 0.745, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.424, + 0.889, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.656, + 0.894, + 0.698 + ], + "angle": 0, + "content": "Figure 4. Source images (top) from ImageNet and fingerprint samples generated from them by SDBF (2nd row), IBSF (3rd row), SSF (4th row), MiSentry (5th row), and PublicCheck (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.894, + 0.793 + ], + "angle": 0, + "content": "The results of detecting fine-tuning the last layer with \\(10^{-5}\\) learning rate on CIFAR10 are shown in Fig.5. From the results, we can see that using multiple fingerprint samples for detection generally performs better than a single one. Among these methods, SDBF achieves perfect detection rates fastest as \\(M\\) increases when \\(K = 2\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Additionally, we also conduct an ablation study on whether \\( MBCS \\) improves detection performance when using multiple fingerprint samples (see SDBF-MBCS and SDBF Without \\( MBCS \\) in Fig.5). The results demonstrate \\( MBCS \\) indeed boosts the multiple fingerprint detection performance. For example, after being equipped with \\( MBCS \\), the detection rate of SDBF has been boosted from \\( 68.3\\% \\) to \\( 84.5\\% \\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.472, + 0.138 + ], + "angle": 0, + "content": "\\(79.1\\%\\) to \\(97.5\\%\\) when \\(M = 2\\) and \\(M = 3\\) at \\(K = 2\\), respectively, making the detection rate of SDBF faster grow up to perfect state than existing SOTA fingerprinting methods." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.144, + 0.273, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.231, + 0.215, + 0.24 + ], + "angle": 0, + "content": "(a) CIFAR10" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.144, + 0.468, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.231, + 0.403, + 0.24 + ], + "angle": 0, + "content": "(b) GTSRB" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.283 + ], + "angle": 0, + "content": "Figure 5. Detection rates on CIFAR10 and GTSRB for fine-tuning the last layer with \\(10^{-5}\\) learning rate in multiple fingerprint samples setting. Note that \\(K = 2\\) for both IBSF and SDBF here." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.29, + 0.392, + 0.306 + ], + "angle": 0, + "content": "6.5. Post-Tampering Output Naturalness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.471, + 0.525 + ], + "angle": 0, + "content": "When the model has been tampered with by adversaries, they might reject the low-confidence query to prevent detection by sensitive fingerprint samples based on the decision boundary. To evaluate the effectiveness of SDBF in responding to such output anomaly detection, we statistically analyzed the top-1 label probabilities of the backdoor-tampered model with SDBF. The results are shown in Fig.7. From the results, we can observe that SDBF achieves comparable covertness to MiSentry, which provides the highest output concealment. This result evidenced that the output probability of the fingerprint samples generated by SDBF is highly sensitive to model tampering and is consistent with our theoretical analysis. Unfortunately, IBSF is at risk of abnormal output rejection, especially on GTSRB." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.532, + 0.357, + 0.547 + ], + "angle": 0, + "content": "6.6. Robustness to Adaptive Attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Adversaries may attempt to bypass SDBF by collecting generated fingerprint samples or generating new ones for adaptive attacks, aiming to manipulate the model while keeping the top-1 labels of the fingerprint samples unchanged. To evaluate SDBF's robustness against such attacks, we generate 1,000 additional fingerprint samples, referred to as leaked fingerprints, which are used for launching adaptive attacks. The set of original samples used for tampering detection is called the unleaked fingerprints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.47, + 0.737 + ], + "angle": 0, + "content": "We conduct adaptive backdoor attacks on CIFAR10, as shown in Fig. 6 (a). As training progresses, the detection rate for the leaked fingerprints drops to \\(0\\%\\), while the detection" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.761, + 0.272, + 0.849 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.851, + 0.183, + 0.857 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.76, + 0.468, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.852, + 0.373, + 0.857 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.857, + 0.47, + 0.911 + ], + "angle": 0, + "content": "Figure 6. (a) Tampering detection rates of detecting adaptive backdoor attacks on CIFAR10 at \\( K = 2 \\) in a single fingerprint sample setting. (b) Adaptive tampering detection rates when a larger number of fingerprints are leaked." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.083, + 0.63, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.083, + 0.758, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.083, + 0.888, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.152, + 0.892, + 0.194 + ], + "angle": 0, + "content": "Figure 7. Reuslts on post-tampering output naturalness. The black vertical dashed line shows the lowest \\(5\\%\\) value of the confidence in the top-1 label of normal inputs." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.198, + 0.89, + 0.212 + ], + "angle": 0, + "content": "Table 2. Average generation computational cost (in seconds/GBs)." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.223, + 0.89, + 0.312 + ], + "angle": 0, + "content": "
MethodCIFAR-10GTSRBImageNet
MiSentry1641.43s/9.4GB1639.77s/11.3GB6112.89s/23.8GB
SSF1.39s/1.2GB1.42s/1.6GB184.13s/12.9GB
PublicCheck0.71s/1.1GB0.74s/1.3GB34.51s/6.7GB
SDBFK=20.20s/0.2GB0.21s/0.4GB8.48s/3.1GB
K=100.22s/0.2GB0.24s/0.4GB8.59s/3.1GB
IBSFK=20.19s/0.2GB0.21s/0.4GB8.44s/3.1GB
K=100.22s/0.2GB0.23s/0.4GB8.57s/3.1GB
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.444 + ], + "angle": 0, + "content": "rate for the unleaked fingerprints remains relatively high, with only a \\(12\\%\\) decrease compared to the baseline. Such robustness of SDBF is attributed to the diversity of its source samples and sample-dependent \\(\\mathcal{C}_{sub}\\). We also conduct a stress test by gradually increasing the number of leaked fingerprint samples to 10,000, as shown in Fig. 6 (b). Despite the large number of leaked fingerprints, the detection rate of the unleaked fingerprints decreased but remained around \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.893, + 0.55 + ], + "angle": 0, + "content": "Moreover, we conduct an experiment to explore whether the steep decision boundary fingerprinting is applicable to the smoothed models such as adversarial training (AT) [32]. We find although AT slightly increased the optimization time cost of SDBF and slightly reduced its performance when \\( K = 2 \\), SDBF remains efficient and effective when \\( K \\) is larger. The detailed results are shown in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.692, + 0.571 + ], + "angle": 0, + "content": "6.7. Computational Cost" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.7 + ], + "angle": 0, + "content": "Table 2 compares the computational cost of SDBF with existing SOTA fingerprinting methods, evaluated on a single NVIDIA RTX 3090Ti GPU. We can observe that: (1) Increasing \\( K \\) slightly raises the cost for both SDBF and IBSF; (2) SDBF incurs a computational overhead similar to IBSF, with both methods exhibiting minimal costs; and (3) Even with high-resolution ImageNet images and large-scale models like ViT, SDBF still remains highly efficient." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.724 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel method for detecting hard-label tampering in DNN models. Our approach leverages the sensitive nature of steep decision boundary and maximizes fingerprint sample diversity over the decision boundary through the Max Boundary Coverage Strategy (MBCS). Through both theoretical analysis and extensive experiments, we have shown that SDBF outperforms existing SOTA hard-label fingerprinting techniques in both sensitivity and efficiency, making SDBF an effective and efficient solution for ensuring the integrity of cloud-deployed DNN models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29285" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.115, + 0.47, + 0.145 + ], + "angle": 0, + "content": "This work was partially supported by National Natural Science Foundation of China under Grant No.62272175." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.158, + 0.175, + 0.173 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.184, + 0.472, + 0.224 + ], + "angle": 0, + "content": "[1] Amazon Web Services. Amazon sagemaker. https://aws.amazon.com/sagemaker, 2023. Last accessed March 20, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.226, + 0.473, + 0.28 + ], + "angle": 0, + "content": "[2] Omid Aramoon, Pin-Yu Chen, and Gang Qu. Aid: Attesting the integrity of deep neural networks. In Proceedings of 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 19-24, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.282, + 0.472, + 0.322 + ], + "angle": 0, + "content": "[3] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. A survey of model compression and acceleration for deep neural networks. arXiv preprint arXiv:1710.09282, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.324, + 0.472, + 0.351 + ], + "angle": 0, + "content": "[4] PyTorch Contributors. Quantization. https://pytorch.org/docs/stable/quantization.html, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.352, + 0.472, + 0.434 + ], + "angle": 0, + "content": "[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.436, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[6] Amel Nestor Docena, Thomas Wahl, Trevor Pearce, and Yunsi Fei. Sensitive samples revisited: Detecting neural network attacks using constraint solvers. arXiv preprint arXiv:2109.03966, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.491, + 0.472, + 0.572 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.575, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[8] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304-9312, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[9] Google Cloud. AutoML. https://cloud.google.com/automl, 2023. Last accessed March 20, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.472, + 0.714 + ], + "angle": 0, + "content": "[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.715, + 0.472, + 0.769 + ], + "angle": 0, + "content": "[11] Chaoxiang He, Xiaofan Bai, Xiaojing Ma, Bin Benjamin Zhu, Pingyi Hu, Jiayun Fu, Hai Jin, and Dongmei Zhang. Towards stricter black-box integrity verification of deep neural network models. In ACM Multimedia 2024, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.77, + 0.472, + 0.826 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.827, + 0.472, + 0.88 + ], + "angle": 0, + "content": "[13] Zecheng He, Tianwei Zhang, and Ruby Lee. Sensitive-sample fingerprinting of deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4729-4737, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.881, + 0.472, + 0.907 + ], + "angle": 0, + "content": "[14] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.184, + 0.473, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 3104-3122, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.895, + 0.176 + ], + "angle": 0, + "content": "[15] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of International Conference on Machine Learning, pages 1885-1894. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.895, + 0.22 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. M.S. thesis, Department of Computer Science, University of Toronto, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.895, + 0.276 + ], + "angle": 0, + "content": "[17] Deepthi Praveenlal Kuttichira, Sunil Gupta, Dang Nguyen, Santu Rana, and Svetha Venkatesh. Verification of integrity of deployed deep learning models using bayesian optimization. Knowledge-Based Systems, 241:108238, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.895, + 0.333 + ], + "angle": 0, + "content": "[18] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.895, + 0.389 + ], + "angle": 0, + "content": "[19] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.895, + 0.418 + ], + "angle": 0, + "content": "[20] Microsoft. Neural network intelligence. https://github.com/microsoft/nni, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.895, + 0.461 + ], + "angle": 0, + "content": "[21] Microsoft Azure. Microsoft machine learning. https://azure.microsoft.com/en-us/products/machine-learning, 2023. Last accessed March 20, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.895, + 0.531 + ], + "angle": 0, + "content": "[22] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11264-11272, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.895, + 0.588 + ], + "angle": 0, + "content": "[23] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Bit-flip attack: Crushing neural network with progressive bit search. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1211-1220, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.895, + 0.657 + ], + "angle": 0, + "content": "[24] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.895, + 0.742 + ], + "angle": 0, + "content": "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, C. Berg Alexander, and Fei-Fei Li. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.895, + 0.813 + ], + "angle": 0, + "content": "[26] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. Updates-leak: Data set inference and reconstruction attacks in online learning. In 29th USENIX Security Symposium, USENIX Security 2020, pages 1291-1308. USENIX Association, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.814, + 0.895, + 0.884 + ], + "angle": 0, + "content": "[27] Ali Shafahi, W. Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29286" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "Processing Systems 2018, NeurIPS 2018, pages 6106-6116, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[28] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural networks*, 32:323-332, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[29] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning fail? generalized transferability for evasion and poisoning attacks. In Proceedings of 27th USENIX Security Symposium (USENIX Security 18), pages 1299-1316, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[31] Shuo Wang, Sharif Abuadbba, Sidharth Agarwal, Kristen Moore, Ruoxi Sun, Minhui Xue, Surya Nepal, Seyit Camtepe, and Salil Kanhere. Publiccheck: Public integrity verification for services of run-time deep models. In Proceedings of 2023 IEEE Symposium on Security and Privacy (SP), pages 1348-1365. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[32] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.174 + ], + "angle": 0, + "content": "adversarial training. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 36246-36263. PMLR, 23-29 Jul 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[33] Bai Xiaofan, Chaoxiang He, Xiaojing Ma, Bin Benjamin Zhu, and Hai Jin. Intersecting-boundary-sensitive fingerprinting for tampering detection of DNN models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 54402-54413. PMLR, 21-27 Jul 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[34] Zhaoxia Yin, Heng Yin, Hang Su, Xinpeng Zhang, and Zhenzhe Gao. Decision-based iterative fragile watermarking for model integrity verification. arXiv preprint arXiv:2305.09684, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.346, + 0.894, + 0.415 + ], + "angle": 0, + "content": "[35] Renjie Zhu, Ping Wei, Sheng Li, Zhaoxia Yin, Xinpeng Zhang, and Zhenxing Qian. Fragile neural network watermarking with trigger image set. In Proceedings of International Conference on Knowledge Science, Engineering and Management, pages 280-293. Springer, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29287" + } + ] +] \ No newline at end of file diff --git a/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_origin.pdf b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ada239d327cdc8ee04f49bf04bbc3ce5046b25ec --- /dev/null +++ b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/6e3f109f-6d3c-482c-aac1-fe12d944b4cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edc1cb296fc662872884ef8fa31508999fea50516efb952c6b92c0c52e611e7d +size 2399305 diff --git a/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/full.md b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8bdab3356028fe45511bd9e9b69b77c6afb3341 --- /dev/null +++ b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/full.md @@ -0,0 +1,385 @@ +# SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models + +Xiaofan Bai $^{1}$ , Shixin Li $^{1}$ , Xiaojing Ma $^{1}$ , Bin Benjamin Zhu $^{2}$ , Dongmei Zhang $^{2}$ , Linchen Yu $^{1*}$ $^{1}$ Huazhong University of Science and Technology + $^{2}$ Microsoft Corporation +{xiaofanbai, shixinli, linda, linchenyu} @hust.edu.cn, {binzhu, dongmeiz} @microsoft.com + +# Abstract + +Cloud-based AI systems offer significant benefits but also introduce vulnerabilities, making deep neural network (DNN) models susceptible to malicious tampering. This tampering may involve harmful behavior injection or resource reduction, compromising model integrity and performance. To detect model tampering, hard-label fingerprinting techniques generate sensitive samples to probe and reveal tampering. Existing fingerprinting methods are mainly based on gradient-defined sensitivity or decision boundary, with the latter showing a manifest superior detection performance. However, all existing fingerprinting methods either suffer from insufficient sensitivity or incur high computational costs. + +In this paper, we theoretically analyze the black-box co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity. Based on this, we further propose Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight approach for hard-label tampering detection that inherently and efficiently combines the strengths of existing fingerprinting techniques. SDBF places fingerprint samples near the steep decision boundary, where the outputs of samples are inherently highly sensitive to tampering. We also design a Max Boundary Coverage Strategy (MBCS), which enhances samples' diversity over the decision boundary. Theoretical analysis and extensive experimental results show that SDBF outperforms existing SOTA hard-label fingerprinting methods in both sensitivity and efficiency. + +# 1. Introduction + +Deep neural networks (DNNs) have demonstrated excellent capabilities across various domains, such as computer vision [12, 18, 24]. An increasing number of applications have incorporated DNN models. Machine Learning as a Service (MLaaS) platforms, such as Google AutoML [9], Microsoft Azure ML [21], and Amazon SageMaker [1], + +![](images/0cdb057ed36d004143488473c4d414e000e8237e180a99c43b1e8cf87d56554f.jpg) +Figure 1. Overview of potential model tampering and public tampering detection for models deployed in MLaaS. + +leverage powerful pre-trained DNN models, allowing users with limited computational resources to access cutting-edge technology. + +However, models deployed in the cloud often operate beyond the direct control of their owners, leaving them vulnerable to unauthorized/illegal model tampering as shown in Fig. 1. For instance, unscrupulous service providers might compress a model—reducing its memory and computational footprint—as a cost-saving measure without the owner's consent, as demonstrated in prior work [3, 22]. Alternatively, adversaries could inject malicious behaviors into the model's parameters to manipulate the model's outputs [10, 14, 15, 27, 29]. Such tampering not only compromises the model's integrity, security, and overall performance but also erodes users' trust and the reliability of the system. + +Many works [2, 6, 13, 17, 31, 33-35] have focused on detecting model tampering. Due to the black-box nature of the models deployed in the cloud, hard-label fingerprinting techniques are the most practical but also most challenging among existing tampering detection methods as their non + +invasive and black-box nature aligns well with the MLaaS in practice. Existing hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]. The former enhances the sensitivity of fingerprint samples' output to model tampering by maximizing gradient-defined sensitivity, which is the $L_{p}$ norm of the gradient of the output with respect to the model parameters, and these fingerprinting techniques have been widely demonstrated to be much less effective than the latter. The latter improves the sensitivity of fingerprint samples to tampering by minimizing the distance to the decision boundary. In summary, both of these approaches are still not sensitive enough to detect certain types of model tampering, especially subtle ones. However, detecting subtle model tampering is crucial, as the end user may need to confirm if the model they are using is really the one they want to use. + +Additionally, [11] introduced a meta-learning-based fingerprinting approach. However, due to the inherent complexity of meta-learning, this method incurs significantly higher overhead compared to the two aforementioned categories, making it impractical for resource-constrained model owners or users. We argue that, given the very purpose of black-box model tampering detection, a lightweight design for fingerprinting method is essential for its widespread application. + +In this paper, recognizing the importance of closing to decision boundary for tampering detection with fingerprint samples, we theoretically analyze the co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity in a black-box detection manner. Based on this, we further propose a novel lightweight hard-label tampering detection fingerprinting approach, Steep-Decision-Boundary Fingerprinting (SDBF) to improve the fingerprint samples' sensitivity with the inherent highly sensitive nature of samples near the steep decision boundary. Specifically, SDBF places fingerprint samples close to the steep decision boundary, where the outputs exhibit high sensitivity to the tampering-induced shift of decision boundary. To achieve this, SDBF perturbs normal input samples to maximize the $L_{2}$ norm of the logit output while simultaneously minimize its variance. The fingerprint samples are bound to be almost indistinguishable from the original normal samples. Moreover, to further enhance detection performance when using multiple fingerprint samples to detect, we propose a Max Boundary Coverage Strategy (MBCS). This strategy enables the fingerprint samples to explore a wide range of different regions around the decision boundary. + +Our theoretical analysis and extensive experimental results demonstrate that the sensitivity of SDBF is superior to that of existing state-of-the-art (SOTA) hard-label fingerprinting methods while being time and memory-efficient. In summary, our contributions are as follows: + +- We theoretically analyze the black-box co-optimal tam + +pering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity. +- We introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight fingerprinting method that inherently and efficiently combines the strengths of fingerprinting techniques based on gradient-defined sensitivity and decision boundary. +- We propose a Max Boundary Coverage Strategy (MBCS) to ensure that the fingerprint samples widely explore the decision boundary, further enhancing the detection performance when using multiple fingerprint samples to detect model tampering. +- We present both theoretical analysis and extensive experimental results that demonstrate the superior performance and efficiency of SDBF compared to existing SOTA hard-label fingerprinting methods. + +# 2. Related Work + +# 2.1. Hard-Label Tampering Detection + +Tampering detection for DNN models aims to determine whether a model has been altered. Since MLaaS operates as a black-box to users, existing methods focus on hard-label tampering detection, where only the top-1 label is accessible. In this setting, watermarking and fingerprinting are widely used. Watermarking [35], an invasive method, embeds a fragile watermark during training that can later be queried with specific inputs to detect tampering. However, because modifying the model may impact performance, non-invasive fingerprinting techniques, the focus of this paper, have attracted more attention. + +# 2.2. Fingerprinting for Tampering Detection + +Existing black-box hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]. + +Gradient-defined sensitivity. Sensitive-Sample Fingerprinting (SSF) [13] introduces fingerprinting for detecting tampering in DNN models for the first time by gradient-defined sensitivity. SSF generates sensitive samples—natural-looking inputs that cause large output changes when the model's parameters are slightly altered. Following SSF, symbolic constraint solvers [6] and Bayesian optimization (BO) [17] have been used to optimize the generation of sensitive samples. However, such sensitive samples have been proven to be ineffective in extensive experiments for many kinds of model tampering. + +Decision boundary based sensitivity. PublicCheck [31] uses VQVAE models to generate fingerprint samples by carefully adding perturbations along with certain directions + +in the latent space of VQVAE and stopping when it has just crossed the decision boundary of two different categories. Others like AID [2] and DBFW [34] generate fingerprint samples close to the decision boundary of all categories by minimizing output variance using mean square error (MSE) loss functions. Similar to AID and DBFW, IBSF [33] employs a Shannon entropy loss to generate diverse fingerprint samples near different $K(K \geqslant 2)$ categories' decision boundary. While these fingerprinting techniques show significantly superior detection performance than the aforementioned fingerprinting techniques based on gradient-defined sensitivity, they are still not sensitive enough to detect certain types of model tampering, especially subtle ones. Notably, neither one has theoretically justified why samples that are close to the decision boundary are more effective than the faraway, nor has anyone theoretically analyzed the co-optimal tampering detection sensitivity of fingerprint samples in these two contexts. + +Additionally, [11] proposed a meta-learning-based sensitive fingerprinting method, MiSentry, but due to the inherent high overhead of meta-learning, the cost of MiSentry is significantly higher than the two aforementioned categories, making it unaffordable for those resources-limited model owner/users. We believe that due to the very purpose of black-box tampering detection, the lightweight of a fingerprinting method is crucial for its widespread application. + +# 3. Threat Model + +Aligned with existing works, we adopt a threat model that includes white-box generation and black-box detection. Specifically, during the fingerprint samples generation, we assume white-box access to the target DNN, using available normal samples as source data. This assumption is reasonable as fingerprint samples are generated by the model owner. In the detection phase, we assume there is only black-box access to the suspect model, where only an API-based query is allowed, and only the top-1 label of the input query will be returned. Moreover, following [31], we also assume a trustworthy third party that securely stores and distributes the fingerprint samples for public tampering detection, where the DNN service provider is assumed to be untrustworthy and may detect probing fingerprint samples, evading tampering detection. Additionally, adversaries might access some fingerprint samples through some illegal means, and use them to bypass tampering detection. + +# 4. Steep Decision Boundary and Sensitivity + +In this section, we first define the decision boundary in the context of the classification task. We then theoretically demonstrate why samples near the decision boundary are sensitive to shifts caused by model tampering. Next, we define the steep decision boundary and analyze how to place + +fingerprint samples near it to further enhance samples' sensitivity while improving optimization efficiency. + +# 4.1. Steep Decision Boundary + +A DNN is a function $f$ , parameterized by $W$ , which maps an input $x$ to an output logits vector $\mathbf{y} = f(W, x)$ . For classification tasks, let $N$ denotes the number of classes, the logits vector $\mathbf{y}$ will then be transformed into a probability vector $Y$ by the following Softmax function: + +$$ +Y _ {i} = \operatorname {S o f t m a x} \left(\mathbf {y} _ {i}\right) = \frac {e ^ {\mathbf {y} _ {i}}}{\sum_ {j = 1} ^ {N} e ^ {\mathbf {y} _ {j}}} \tag {1} +$$ + +where $Y_{i}$ is the confidence of class $i$ . + +Definition 4.1 (Decision Boundary). For a classification model, the decision boundary is the hypersurface that divides the feature space into regions, each corresponding to a different class label. Typically, the model assigns the same probabilities to the classes on the decision boundary, and the confidences of these categories will be much greater than the other categories. Specifically, for an $N$ -class classification model, the decision boundary of $K(K \leqslant N)$ classes can be formulated as: + +$$ +Y _ {1} = Y _ {2} = \dots = Y _ {K} \gg \sum \tilde {Y} \tag {2} +$$ + +where $\sum \tilde{Y} (0\leqslant \sum \tilde{Y}\leqslant 1)$ is the summation of the other $N - K$ categories' confidence and $\begin{array}{r}\sum_{i = 1}^{K}Y_{i} + \sum \tilde{Y} = 1. \end{array}$ + +For black-box tampering detection, the predicted hard label based on the Softmax output $Y$ is the only observable feedback from the suspect model. To maximize the sensitivity of fingerprint samples in this setting, we need to maximize the sensitivity of $Y$ with respect to $\mathbf{y}$ to force the model to work in Softmax high-sensitivity region, making any tampering-induced $\Delta \mathbf{y} = f(W + \Delta W, x) - f(W, x)$ will be nonlinearly amplified to $Y$ by Softmax, avoiding being suppressed by the saturation zone of Softmax when the fingerprint samples are faraway from the decision boundary. + +Proposition 4.2. Given a fingerprint sample $x$ , the sensitivity of its softmax output $\mathbf{Y}$ with respect to $||\Delta \mathbf{y}||_2$ introduced by model tampering $\Delta W$ will reach the maximum when $x$ close to the decision boundary, where $Y_i = \frac{1}{N}$ . + +Proof. To prove Prop.4.2, we need to analyze the Jacobian matrix $J$ of vector $\mathbf{Y}$ with respect to vector $\mathbf{y}$ and the maximum value of the Frobenius norm (equivalent to the $L_{2}$ norm) of $J$ . First, based on Eq.1, the element $J_{ij}$ of $J$ is the partial derivative of $Y_{i}$ with respect to $y_{j}$ : + +$$ +J _ {i j} = \frac {\partial Y _ {i}}{\partial \mathbf {y} _ {j}} = Y _ {i} \left(\delta_ {i j} - Y _ {j}\right) \tag {3} +$$ + +where $\delta_{ij}$ is the Kronecker Delta function and $\left\{ \begin{array}{l}\delta_{ij} = 1,i = j\\ \delta_{ij} = 0,i\neq j \end{array} \right.$ . Thus, the squared Frobenius norm of $J$ + +can be formulated as: + +$$ +\begin{array}{l} \| J \| _ {F} ^ {2} = \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \left(J _ {i j}\right) ^ {2} = \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \left(Y _ {i} \left(\delta_ {i j} - Y _ {j}\right)\right) ^ {2} \\ = \sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \left(1 - 2 Y _ {i} + Y _ {i} ^ {2} + \sum_ {j \neq i} Y _ {j} ^ {2}\right) \tag {4} \\ = \sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - 2 Y _ {i} + Y _ {i} ^ {2}) + Y _ {i} ^ {2} \sum_ {j \neq i} Y _ {j} ^ {2} \\ \end{array} +$$ + +Since $\sum_{i=1}^{N} Y_{i} = 1$ , thus $\sum_{j \neq i} Y_{j}^{2}$ can be replaced by $1 - Y_{i}^{2}$ : + +$$ +\begin{array}{l} \| J \| _ {F} ^ {2} = \sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \left(1 - 2 Y _ {i} + Y _ {i} ^ {2}\right) + Y _ {i} ^ {2} \left(1 - Y _ {i} ^ {2}\right) \tag {5} \\ = \sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - Y _ {i}) \\ \end{array} +$$ + +Next, we need to find the maximum of Eq.5. For such a constrained optimization problem, we use the Lagrange multiplier method to rigorously prove when the maximum of Eq.5 is achieved. Specifically, we introduce a Lagrange multiplier $\lambda$ for the constraint $\sum_{i=1}^{N} Y_i = 1$ , which is: + +$$ +\mathcal {L} \left(Y _ {1}, Y _ {2}, \dots , Y _ {N}, \lambda\right) = \sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \left(1 - Y _ {i}\right) + \lambda \left(\sum_ {i = 1} ^ {N} Y _ {i} - 1\right) \tag {6} +$$ + +We now compute the partial derivatives of $\mathcal{L}$ with respect to each $Y_{i}$ . The partial derivative of $Y_{i}^{2}(1 - Y_{i})$ with respect to each $Y_{i}$ can be formulated as: + +$$ +\frac {\partial}{\partial Y _ {i}} \left(Y _ {i} ^ {2} \left(1 - Y _ {i}\right)\right) = 2 Y _ {i} \left(1 - Y _ {i}\right) - Y _ {i} ^ {2} = 2 Y _ {i} - 3 Y _ {i} ^ {2} (7) +$$ + +Similarly, for the constraint term, the partial derivative is: + +$$ +\frac {\partial}{\partial Y _ {i}} \left(\sum_ {i = 1} ^ {N} Y _ {i} - 1\right) = 1 \tag {8} +$$ + +Let the total derivative of the Lagrangian equal to zero, we have: + +$$ +2 Y _ {i} - 3 Y _ {i} ^ {2} + \lambda = 0 \Longrightarrow 2 Y _ {i} - 3 Y _ {i} ^ {2} = - \lambda \tag {9} +$$ + +Recall that we have the constraint $\sum_{i=1}^{N} Y_i = 1$ , to find a feasible solution, since Eq.9 holds for each $Y_i$ , we can conclude that all $Y_i$ must be equal, meaning $Y_1 = Y_2 = \dots = Y_N = Y$ . Otherwise, the right term can not be the same constant $-\lambda$ . Then, we can obtain: + +$$ +N \cdot Y = 1 \quad \Rightarrow \quad Y = \frac {1}{N} \tag {10} +$$ + +Finally, to verify that this solution is indeed the maximum, we can check the second order derivative value of $||J||_F^2$ at this solution. When all $Y_{i} = \frac{1}{N}$ , the second order derivative of $||J||_F^2$ becomes: + +$$ +\frac {\partial^ {2} \left| \left| J \right| \right| _ {F} ^ {2}}{\partial Y _ {i} ^ {2}} = \sum_ {i = 1} ^ {N} \frac {\partial 2 Y _ {i} - 3 Y _ {i} ^ {2}}{\partial Y _ {i}} = 2 - 6 \sum_ {i = 1} ^ {N} Y _ {i} < 0. \tag {11} +$$ + +Therefore, this solution is indeed the maximum. In summary, given a fingerprint sample, $\| J\| _F^2$ achieves maximum when it is close to the decision boundary defined in Def.4.1. + +However, when $x$ is faraway from the decision boundary, some elements in $f(W,x)$ are much larger than others, leading to $\text{softmax}(f(W,x))$ that are close to 0 or 1. In this case, the elements of the Jacobian matrix $J_{ij}$ will be significantly different, especially when $Y_{i} \approx 1$ , causing the most of other $Y_{j} \approx 0$ , that is: + +$$ +\left\{ \begin{array}{l} Y _ {i} \approx 1 \\ Y _ {j} \approx 0 \end{array} \right. \Longrightarrow \left\{ \begin{array}{c} J _ {i i} \approx Y _ {i} \cdot (1 - Y _ {i}) \approx 0 \\ J _ {i j} \approx - Y _ {i} \cdot Y _ {j} \approx 0, i \neq j \end{array} \right. \tag {12} +$$ + +Hence, $||J||_F^2$ in this scenario will be significantly smaller as most elements are close to 0. + +Definition 4.3 (Steep Decision Boundary). Given a fingerprint sample $x$ and a model tampering $\Delta W$ , the steep decision boundary refers to the decision boundary where $f(W, x)$ shows a rapid change when $\Delta W$ is applied to $W$ . + +Proposition 4.4. Give a sample $x$ and a DNN model $f(\cdot)$ parameterized by $W$ , the sensitivity of output $f(W, x)$ with respect to $\Delta W$ is positively correlated to the $||f(W, x)||_2$ . + +Proof. Inspired by [13], since all model tampering inevitably result in different performances in the last layer, the sensitivity of $f(W,x)$ to $W$ practically can be formulated as the following Frobenius norm of Jacobian matrix $\hat{J}$ : + +$$ +\left| \left| \hat {J} \right| \right| _ {F} ^ {2} = \left| \left| \frac {\partial f (W , x)}{\partial W _ {F C}} \right| \right| _ {2} \tag {13} +$$ + +where $W_{FC}$ is the weights of the last fully connected layer of $f(\cdot)$ . Since all existing classification models do not use any nonlinear activation functions like $ReLu(\cdot)$ in the last layer, the last layer of the model performs the following linear operations on the output of the previous layer: + +$$ +f (W, x) = W _ {F C} \times f _ {- 2} (W, x) + \mathbf {b} \tag {14} +$$ + +where $f_{-2}(W,x)$ and $\mathbf{b}$ are the output and bias of the penultimate layer, respectively. Substituting Eq. 14 into Eq. 13, we can get: + +$$ +\left| \left| \hat {J} \right| \right| _ {F} ^ {2} = \left| \left| f _ {- 2} (W, x) \right| \right| _ {2} \tag {15} +$$ + +Combining Eq.15 and Eq.14, when we increase $||f(W,x)||_2$ , the sensitivity of $f(W,x)$ with respect to $\Delta W$ + +$\| \hat{J}\| _F^2$ , will increase simultaneously. To valid Prop.4.4, we conduct an experiment by maximizing $\| f(W,x)\| _2$ while recording $\| \frac{\partial f(W,x)}{\partial\mathbf{W}_{FC}}\| _2$ for a ResNet20 trained on CIFAR10, as shown in Fig. 2. The results are highly consistent with Prop.4.4. Therefore, we can improve the sample's sensitivity of $f(W,x)$ to $\Delta W$ by increasing $\| f(W,x)\| _2$ + +![](images/8271a1f4520dad734185fc33a3aa57a632beb4854f741df230729e082b5f5d70.jpg) +Figure 2. Validation results on the relationship between $||f(W,x)||_2$ and $||\frac{\partial f(W,x)}{\partial W_{FC}}||_2$ . + +Proposition 4.5. Give a fingerprint sample $x$ and a DNN model $f(\cdot)$ parameterized by $W$ , simultaneously increasing the sensitivity of softmax $(f(W, x))$ to corresponding $f(W, x)$ and the sensitivity of $f(W, x)$ to $\Delta W$ can drive $x$ close to the steep decision boundary and thus further enhance the sensitivity of softmax $(f(W, x))$ to $\Delta W$ . + +Prop.4.5 can be easily proved with the linear translational invariance of $\text{Softmax}(\cdot)$ . More details are in the appendix. Based on Def.4.3 and Prop.4.5, we know how to efficiently place the fingerprint samples near steep decision boundary, where they are highly sensitive to model tampering $\Delta W$ as it inherently combining the strength of fingerprinting methods based on gradient-defined sensitivity and decision boundary. + +# 5. Steep-Decision-Boundary Fingerprinting + +In this section, we introduce a novel sensitive fingerprinting method named Steep-Decision-Boundary Fingerprinting (SDBF). Our goal is to leverage the highly sensitive nature of the samples near the steep decision boundary to detect model tampering, as shown in Fig. 3. + +# 5.1. Fingerprint Samples Generation + +Different from [13], we do not directly optimize Eq.13 and Eq.5 as their poor efficiency. We will further explain this in the appendix. Fortunately, based on Prop.4.2, Prop. 4.4 and Prop. 4.5, we can maximize the fingerprint samples' sensitivity by maximizing $||\mathbf{y}||_2$ while minimizing the variance of $\mathbf{y}$ , $\sigma (\mathbf{y})$ . Similar to IBSF [33], we can also freely choose to place fingerprint samples at $K(K\leqslant N)$ categories' decision boundary, which will help increase the diversity of + +![](images/74dfa27de7b4ea707353750cd5976de4a8e8f8b59c118f24140ecc0ca432ee6c.jpg) +Figure 3. Illustration of SDBF, gradient-defined fingerprinting SSF, and decision-boundary-based fingerprinting IBSF. SDBF shows an inherent high sensitivity to tampering-introduced boundary shift. + +![](images/458c102fdccdb1ee207d49513250ee65516668a51bb2cdf3ad692f0819af0c7a.jpg) + +the fingerprint samples. Specifically, let $\mathbf{y}_{\mathcal{C}_{sub}}$ denote the logits output of the chosen categories and $\overline{\mathbf{y}}_{\mathcal{C}_{sub}} = \frac{1}{K}\sum_{i=1}^{k}\mathbf{y}_i$ denote the averaged logits of the categories in $\mathcal{C}_{sub}$ , our objective function for a sample $x$ can be formalized as: + +$$ +\begin{array}{l} \mathcal {L} (x, \mathcal {C} _ {s u b}) = \sigma (\mathbf {y} _ {\mathcal {C} _ {s u b}}) - \lambda \cdot | | \mathbf {y} _ {\mathcal {C} _ {s u b}} | | _ {2} \\ = \frac {1}{K} \cdot \sum_ {i = 1} ^ {K} \left(\mathbf {y} _ {i} - \overline {{\mathbf {y}}} _ {\mathcal {C} _ {s u b}}\right) ^ {2} - \lambda \cdot \sum_ {i = 1} ^ {K} \left\| \mathbf {y} _ {i} \right\| _ {2} \tag {16} \\ \end{array} +$$ + +where $\lambda$ is the coefficient and $i\in \mathcal{C}_{sub}$ . Thus, given a normal sample $x$ , we can drive it close to the steep decision boundary by minimizing the objective function in Eq.16: + +$$ +\begin{array}{l} \min _ {\Delta x} \mathcal {L} (x + \Delta x, \mathcal {C} _ {s u b}) \tag {17} \\ s. t. | | \Delta x | | _ {p} \leqslant \epsilon \\ \end{array} +$$ + +where $\epsilon$ is the input bound. The constraint $\| \Delta x\| _p\leqslant \epsilon$ bounds the perturbation $\Delta x$ in the input space, making the generated fingerprint sample perceptually similar to the original normal one. This naturalness makes it hard for an adversary to distinguish whether the input is a probe query by input anomaly detection. We solve Eq. 17 by using projected gradient descent (PGD) like [19]: + +$$ +x _ {i + 1} = \operatorname {C l i p} _ {\epsilon} \left(x _ {i} + \eta \cdot \nabla_ {x _ {i}} \mathcal {L} \left(x _ {i}, \mathcal {C} _ {\text {s u b}}\right)\right) \tag {18} +$$ + +where $x_{i} = x_{i - 1} + \Delta x_{i}$ is the generated fingerprint sample at the $i$ -th iteration. + +The optimization eventually yields a fingerprint sample positioned close to $\mathcal{C}_{sub}$ steep decision boundary. Following [33], to reduce the perturbation effort, we select the top $|\mathcal{C}_{sub}|$ confident categories in $\mathbf{y}$ of original normal $x$ as $\mathcal{C}_{sub}$ by default. This sample-dependent $\mathcal{C}_{sub}$ will enhance the diversity of generated sensitive fingerprint samples, making it more robust to partial leakage. + +Remark. In the generation phase, except for fingerprint sample $x$ , we will also collect its corresponding top-1 label as the fingerprint label $\mathcal{V}$ . In the de + +tection phase, $x$ will be used to query the suspect model to get the returned hard label $\mathcal{V}_{predicted}$ and then we will determine whether the model has been tampered with by verifying whether $\mathcal{V}_{predicted} = \mathcal{V}$ . + +# 5.2. Max Boundary Coverage Strategy + +When probing the tampered model with $M$ ( $M > 1$ ) fingerprint samples, we collect their returned hard labels. The model is considered untampered if all the labels match the corresponding fingerprint labels. To maximize the diversity of the $M$ fingerprint samples over the steep decision boundary, we propose the Max Boundary Coverage Strategy (MBCS) to make them widely explore the decision boundary. Specifically, we first randomly select a different fingerprint sample, denoted as $x_{1}$ . Then, let $m \leqslant M$ and $x_{i}$ denote the number of selected fingerprint samples and the sample to select in $i$ -th iteration, we iteratively choose the remaining $M - 1$ samples to maximize the boundary diversity score: + +$$ +\max _ {x _ {i}} \mathcal {L} _ {d i v} = \sum_ {i = 1} ^ {N} c _ {i} \cdot \log c _ {i} \tag {19} +$$ + +where $N$ is the number of categories and + +$$ +c _ {i} = \frac {1}{m} \cdot \sum_ {j = 1} ^ {m} \operatorname {s o f t m a x} _ {i} \left(f \left(x _ {j}\right)\right) \tag {20} +$$ + +where $\text{softmax}_i(\cdot)$ denotes softmax operation along with $i - th$ dimension. + +By repeating the $MBCS\mathcal{R}$ times, we can get a $R$ -length and $M$ -width probe fingerprint sequence. For each probe query, we will send $M$ probe fingerprint samples to the suspect model until we go through all $\mathcal{R} \times M$ fingerprint samples. In our experiments, we use $MBCS$ as a post-processing technique to enhance the diversity of the fingerprint samples over the decision boundary. + +# 6. Experimental Evaluation + +# 6.1. Datasets, Models, and Baselines + +To evaluate the performance of SDBF, we use various common classification datasets, including CIFAR10 [16], GTSRB [28], and ImageNet [25]. To evaluate the generalizability of SDBF across different model architectures, we select three different popular architectures for the above three datasets. Specifically, we use ResNet20 [12] for CIFAR10, GoogLeNet [30] for GTSRB, and vision transformer ViT-B-32 [7] pretrained by Pytorch for ImageNet. + +To assess the superiority of SDBF compared to existing SOTA fingerprinting methods, we select four SOTA fingerprinting methods, SSF [13], PublicCheck [31], IBSF [33], and MiSentry [11], as the baselines. We utilize the default + +settings outlined in their respective papers unless explicitly specified otherwise. For each fingerprinting method, we independently generate 1000 fingerprint samples. + +# 6.2. Tampering Settings and Detection Rate + +Tampering Settings. In our experiments, we evaluate SDBF and the baselines on the 9 types of model tampering selected from BIVBench [11], including 4 hardest-to-detect model tampering, targeted attack [15,27,29], finetuning the last layer with $10^{-5}$ learning rate, and Unlearning [8]/Online Learning (On-Learning) [26] makes a model forget/newly learn specific data, and 5 common model tampering consists of Backdoor [10], Bit-flipping Attack [23], Quantization [4], Pruning [20], and independently trained models with Different Architecture (Different Arch.) on the same dataset. All the tampering settings are aligned with BIVBench [11]. In this paper, the reported experimental results for each tampering type are averaged over 10 independently tampered models to mitigate randomness in evaluating tampering detection performance. + +Tampering Detection Rate. For each tamper detection probe, $M$ ( $M \geqslant 1$ ) fingerprint samples are selected from a pool of 1000 fingerprint samples to query the test model. The model's top-1 predicted labels are then compared with the corresponding fingerprint labels. If all predicted labels match the fingerprint labels, the model is classified as tampered; otherwise, it is marked as tampered. This procedure is repeated 100 times for each test model. Since 10 tampered models are generated for each tampering type, a total of 1000 tests are performed for each tampering type. The tampering detection rate is the proportion of the 1000 probe queries in which tampering is successfully detected. Note that there are no false positives (FP) in the tampering detection of classification model, as the classification model $f(\cdot)$ provides a stable mapping from input $x$ to output $f(x)$ + +# 6.3. Perceptual Quality of Fingerprint Samples + +Fig. 4 shows a subset of fingerprint samples generated by SDBF and four baselines, all constrained by the same $L_{2}$ norm bound. Fingerprint samples generated by SDBF closely resemble those from SSF, IBSF, and MiSentry, with imperceptible pixel-level perturbations added to the original images. In contrast, fingerprint samples generated by PublicCheck exhibit noticeable artifacts, with semantic-level alterations to the original images, due to the use of generative models. For example, the distorted face in the 8th column of Fig. 4 highlights this effect. + +# 6.4. Effectiveness of Tampering Detection + +Single Fingerprint Sample Detection. We first evaluate the detection performance in the single fingerprint sample detection setting, which only uses one fingerprint sample to query the suspect model. Specifically, following [33], we comprehensively evaluate SDBF and IBSF with $K$ ranging + +Table 1. Tamper detection rate (%) of detecting tampering in a single fingerprint sample setting on CIFAR10 and GTSRB. For fine-tuning(FT), "Last" specifies only fine-tuning the last layer, with numbers indicating the ${10}^{-5}$ learning rate. The "On-Learning" refers to online learning. + +
DatasetMethodFTLast-5BackdoorT-AttackUnlearningQuantizationDifferent Arch.PruningOn-LearningBit-flipping
CIFAR10SSF0.064.911.43.266.379.766.76.784.4
MiSentry45.986.483.868.490.793.589.568.288.9
PublicCheck21.673.636.928.360.865.169.525.485.7
IBSFK=224.577.442.537.578.283.381.247.587.8
K=435.983.655.359.489.188.588.356.289.4
K=642.887.563.569.290.490.890.569.588.6
K=844.589.466.171.390.992.791.371.689.5
K=1049.490.973.272.891.393.891.976.490.1
SDBF(ours)K=238.181.456.456.586.685.384.158.188.4
K=447.587.368.669.889.791.591.266.989.2
K=653.189.979.773.491.592.892.373.588.3
K=856.291.483.175.191.993.792.975.890.3
K=1062.892.585.877.393.294.693.681.189.8
GTSRBSSF0.059.89.51.462.664.559.12.785.1
MiSentry35.286.475.159.487.979.583.759.889.3
PublicCheck18.467.837.321.769.266.570.428.387.3
IBSFK=222.776.339.232.879.872.177.542.688.5
K=431.979.952.955.986.179.783.257.390.4
K=638.485.859.861.889.183.385.563.490.1
K=841.588.563.470.191.285.988.465.189.6
K=1042.891.169.773.391.888.490.368.291.3
SDBF(ours)K=229.180.949.849.484.777.282.953.990.3
K=438.586.662.566.890.584.589.464.391.5
K=646.790.378.471.592.289.691.772.789.2
K=853.291.581.573.692.991.391.977.991.9
K=1057.192.784.678.293.693.892.880.190.7
+ +from $K = 2$ to $K = N = 10$ and $K = 2$ to $K = 500$ on CIFAR10 and ImageNet, respectively. Tab.1 summarizes the tampering detection rates on CIFAR10 and GTSRB of existing SOTA fingerprinting methods and SDBF. The detailed results on ImageNet are in the appendix. From the results, we can conclude that: + +SDBF generally outperforms the four SOTA baselines, although exceptions arise for SDBF at $K = 2$ when using a single fingerprint sample to perform tampering detection, which can be attributed to MiSentry's inherent sensitivity to such tampering as some subtle tampering is in its model zoo of meta-learning. As $K$ increases, SDBF gradually outperforms MiSentry, with improvements of up to $16.9\%$ on CIFAR10 at $K = 10$ for fine-tuning the last layer with $10^{-5}$ learning rate compared to MiSentry. The detection rate of IBSF shares a similar trend with our SDBF when $K$ increases, yet it is consistently performing worse than SDBF, evidencing our key intuition: the fingerprint samples close to the steep decision boundary are more sensitive. + +Multiple Fingerprint Samples Detection. Next, we evaluate the detection performance in multiple fingerprint samples detection, which uses $M(M > 1)$ fingerprint samples to query the tampered model. If $\mathcal{V}_{predicted}$ of any one fingerprint sample in the $M$ samples differ from the corresponding $\mathcal{V}$ , the model will be deemed as a tampered one. Following the detection setting in IBSF, MiSentry, and PublicCheck, we repeat the detection probe 1000 times by selecting $M$ random fingerprint samples for them. As for SDBF and SSF, we use MBCS and Max Neuron Coverage Strategy (MNCS) [13], which is specifically designed for SSF to select $M$ fingerprint samples for each probe, respectively. + +![](images/a94eef6b5d24eacf83c0a9211a3664726c2d49684c1bdcc258141ab445ca5f8b.jpg) +Figure 4. Source images (top) from ImageNet and fingerprint samples generated from them by SDBF (2nd row), IBSF (3rd row), SSF (4th row), MiSentry (5th row), and PublicCheck (bottom). + +![](images/11df4dcc11f6c66a4a88e3188beb03ed301bad491c40837a65662220c55fc104.jpg) + +The results of detecting fine-tuning the last layer with $10^{-5}$ learning rate on CIFAR10 are shown in Fig.5. From the results, we can see that using multiple fingerprint samples for detection generally performs better than a single one. Among these methods, SDBF achieves perfect detection rates fastest as $M$ increases when $K = 2$ . + +Additionally, we also conduct an ablation study on whether $MBCS$ improves detection performance when using multiple fingerprint samples (see SDBF-MBCS and SDBF Without $MBCS$ in Fig.5). The results demonstrate $MBCS$ indeed boosts the multiple fingerprint detection performance. For example, after being equipped with $MBCS$ , the detection rate of SDBF has been boosted from $68.3\%$ to $84.5\%$ and + +$79.1\%$ to $97.5\%$ when $M = 2$ and $M = 3$ at $K = 2$ , respectively, making the detection rate of SDBF faster grow up to perfect state than existing SOTA fingerprinting methods. + +![](images/f0ee97b6628e0b0c975773bd3442a85d37347b2c7c5e5e0380a0c69501371f91.jpg) +(a) CIFAR10 + +![](images/d9c78d4b5a9d992b8d2c28ac081c440c9e88d65bcb7ea7b1b36f079fb829f66d.jpg) +(b) GTSRB +Figure 5. Detection rates on CIFAR10 and GTSRB for fine-tuning the last layer with $10^{-5}$ learning rate in multiple fingerprint samples setting. Note that $K = 2$ for both IBSF and SDBF here. + +# 6.5. Post-Tampering Output Naturalness + +When the model has been tampered with by adversaries, they might reject the low-confidence query to prevent detection by sensitive fingerprint samples based on the decision boundary. To evaluate the effectiveness of SDBF in responding to such output anomaly detection, we statistically analyzed the top-1 label probabilities of the backdoor-tampered model with SDBF. The results are shown in Fig.7. From the results, we can observe that SDBF achieves comparable covertness to MiSentry, which provides the highest output concealment. This result evidenced that the output probability of the fingerprint samples generated by SDBF is highly sensitive to model tampering and is consistent with our theoretical analysis. Unfortunately, IBSF is at risk of abnormal output rejection, especially on GTSRB. + +# 6.6. Robustness to Adaptive Attacks + +Adversaries may attempt to bypass SDBF by collecting generated fingerprint samples or generating new ones for adaptive attacks, aiming to manipulate the model while keeping the top-1 labels of the fingerprint samples unchanged. To evaluate SDBF's robustness against such attacks, we generate 1,000 additional fingerprint samples, referred to as leaked fingerprints, which are used for launching adaptive attacks. The set of original samples used for tampering detection is called the unleaked fingerprints. + +We conduct adaptive backdoor attacks on CIFAR10, as shown in Fig. 6 (a). As training progresses, the detection rate for the leaked fingerprints drops to $0\%$ , while the detection + +![](images/b5eddbcf0af591f6238eff7b415587bbe1b7a8bbd4ebe0d786ad38fc685a2917.jpg) +(a) + +![](images/0a9880fedad863154345aeb7bdaafb3ac60575c6320a56e25678de76d45ff209.jpg) +(b) +Figure 6. (a) Tampering detection rates of detecting adaptive backdoor attacks on CIFAR10 at $K = 2$ in a single fingerprint sample setting. (b) Adaptive tampering detection rates when a larger number of fingerprints are leaked. + +![](images/5e234e555a7b52c66bd7c1e1ec112542757ce1c9d6ddc7b33233d9086249876f.jpg) + +![](images/66456b913dae7c6cc6573418f1f1f1f76a3c6dc1966992b1dceae02476dd7c1b.jpg) + +![](images/6931aab78375b367b04fe3be3b8ce7548f11074564e1b8e917515d62c86e3913.jpg) +Figure 7. Reuslts on post-tampering output naturalness. The black vertical dashed line shows the lowest $5\%$ value of the confidence in the top-1 label of normal inputs. + +Table 2. Average generation computational cost (in seconds/GBs). + +
MethodCIFAR-10GTSRBImageNet
MiSentry1641.43s/9.4GB1639.77s/11.3GB6112.89s/23.8GB
SSF1.39s/1.2GB1.42s/1.6GB184.13s/12.9GB
PublicCheck0.71s/1.1GB0.74s/1.3GB34.51s/6.7GB
SDBFK=20.20s/0.2GB0.21s/0.4GB8.48s/3.1GB
K=100.22s/0.2GB0.24s/0.4GB8.59s/3.1GB
IBSFK=20.19s/0.2GB0.21s/0.4GB8.44s/3.1GB
K=100.22s/0.2GB0.23s/0.4GB8.57s/3.1GB
+ +rate for the unleaked fingerprints remains relatively high, with only a $12\%$ decrease compared to the baseline. Such robustness of SDBF is attributed to the diversity of its source samples and sample-dependent $\mathcal{C}_{sub}$ . We also conduct a stress test by gradually increasing the number of leaked fingerprint samples to 10,000, as shown in Fig. 6 (b). Despite the large number of leaked fingerprints, the detection rate of the unleaked fingerprints decreased but remained around $50\%$ . + +Moreover, we conduct an experiment to explore whether the steep decision boundary fingerprinting is applicable to the smoothed models such as adversarial training (AT) [32]. We find although AT slightly increased the optimization time cost of SDBF and slightly reduced its performance when $K = 2$ , SDBF remains efficient and effective when $K$ is larger. The detailed results are shown in the appendix. + +# 6.7. Computational Cost + +Table 2 compares the computational cost of SDBF with existing SOTA fingerprinting methods, evaluated on a single NVIDIA RTX 3090Ti GPU. We can observe that: (1) Increasing $K$ slightly raises the cost for both SDBF and IBSF; (2) SDBF incurs a computational overhead similar to IBSF, with both methods exhibiting minimal costs; and (3) Even with high-resolution ImageNet images and large-scale models like ViT, SDBF still remains highly efficient. + +# 7. Conclusion + +In this work, we introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel method for detecting hard-label tampering in DNN models. Our approach leverages the sensitive nature of steep decision boundary and maximizes fingerprint sample diversity over the decision boundary through the Max Boundary Coverage Strategy (MBCS). Through both theoretical analysis and extensive experiments, we have shown that SDBF outperforms existing SOTA hard-label fingerprinting techniques in both sensitivity and efficiency, making SDBF an effective and efficient solution for ensuring the integrity of cloud-deployed DNN models. + +# Acknowledgements + +This work was partially supported by National Natural Science Foundation of China under Grant No.62272175. + +# References + +[1] Amazon Web Services. Amazon sagemaker. https://aws.amazon.com/sagemaker, 2023. Last accessed March 20, 2023. +[2] Omid Aramoon, Pin-Yu Chen, and Gang Qu. Aid: Attesting the integrity of deep neural networks. In Proceedings of 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 19-24, 2021. +[3] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. A survey of model compression and acceleration for deep neural networks. arXiv preprint arXiv:1710.09282, 2017. +[4] PyTorch Contributors. Quantization. https://pytorch.org/docs/stable/quantization.html, 2022. +[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020. +[6] Amel Nestor Docena, Thomas Wahl, Trevor Pearce, and Yunsi Fei. Sensitive samples revisited: Detecting neural network attacks using constraint solvers. arXiv preprint arXiv:2109.03966, 2021. +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. +[8] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304-9312, 2020. +[9] Google Cloud. AutoML. https://cloud.google.com/automl, 2023. Last accessed March 20, 2023. +[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. +[11] Chaoxiang He, Xiaofan Bai, Xiaojing Ma, Bin Benjamin Zhu, Pingyi Hu, Jiayun Fu, Hai Jin, and Dongmei Zhang. Towards stricter black-box integrity verification of deep neural network models. In ACM Multimedia 2024, 2024. +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. +[13] Zecheng He, Tianwei Zhang, and Ruby Lee. Sensitive-sample fingerprinting of deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4729-4737, 2019. +[14] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In + +Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 3104-3122, 2021. +[15] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of International Conference on Machine Learning, pages 1885-1894. PMLR, 2017. +[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. M.S. thesis, Department of Computer Science, University of Toronto, 2009. +[17] Deepthi Praveenlal Kuttichira, Sunil Gupta, Dang Nguyen, Santu Rana, and Svetha Venkatesh. Verification of integrity of deployed deep learning models using bayesian optimization. Knowledge-Based Systems, 241:108238, 2022. +[18] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015. +[19] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. +[20] Microsoft. Neural network intelligence. https://github.com/microsoft/nni, 2022. +[21] Microsoft Azure. Microsoft machine learning. https://azure.microsoft.com/en-us/products/machine-learning, 2023. Last accessed March 20, 2023. +[22] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11264-11272, 2019. +[23] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Bit-flip attack: Crushing neural network with progressive bit search. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1211-1220, 2019. +[24] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016. +[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, C. Berg Alexander, and Fei-Fei Li. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. +[26] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. Updates-leak: Data set inference and reconstruction attacks in online learning. In 29th USENIX Security Symposium, USENIX Security 2020, pages 1291-1308. USENIX Association, 2020. +[27] Ali Shafahi, W. Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information + +Processing Systems 2018, NeurIPS 2018, pages 6106-6116, 2018. +[28] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural networks*, 32:323-332, 2012. +[29] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning fail? generalized transferability for evasion and poisoning attacks. In Proceedings of 27th USENIX Security Symposium (USENIX Security 18), pages 1299-1316, 2018. +[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions, 2014. +[31] Shuo Wang, Sharif Abuadbba, Sidharth Agarwal, Kristen Moore, Ruoxi Sun, Minhui Xue, Surya Nepal, Seyit Camtepe, and Salil Kanhere. Publiccheck: Public integrity verification for services of run-time deep models. In Proceedings of 2023 IEEE Symposium on Security and Privacy (SP), pages 1348-1365. IEEE, 2023. +[32] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve + +adversarial training. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 36246-36263. PMLR, 23-29 Jul 2023. +[33] Bai Xiaofan, Chaoxiang He, Xiaojing Ma, Bin Benjamin Zhu, and Hai Jin. Intersecting-boundary-sensitive fingerprinting for tampering detection of DNN models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 54402-54413. PMLR, 21-27 Jul 2024. +[34] Zhaoxia Yin, Heng Yin, Hang Su, Xinpeng Zhang, and Zhenzhe Gao. Decision-based iterative fragile watermarking for model integrity verification. arXiv preprint arXiv:2305.09684, 2023. +[35] Renjie Zhu, Ping Wei, Sheng Li, Zhaoxia Yin, Xinpeng Zhang, and Zhenxing Qian. Fragile neural network watermarking with trigger image set. In Proceedings of International Conference on Knowledge Science, Engineering and Management, pages 280-293. Springer, 2021. \ No newline at end of file diff --git a/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/images.zip b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..286cad70c15178511e2acc90e21343cb2f79409e --- /dev/null +++ b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3abbc21eca72f30cfaf8c5369d64156d2fde8d3e1affaa1003d00d3b2d3aca84 +size 568567 diff --git a/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/layout.json b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..67ea459107c9238cb61785e91159ad92033fc6d8 --- /dev/null +++ b/2025/SDBF_ Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models/layout.json @@ -0,0 +1,11418 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 531, + 138 + ], + "type": "text", + "content": "SDBF: Steep-Decision-Boundary Fingerprinting for Hard-Label Tampering Detection of DNN Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": "Xiaofan Bai" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": ", Shixin Li" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": ", Xiaojing Ma" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": ", Bin Benjamin Zhu" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": ", Dongmei Zhang" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": ", Linchen Yu" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": "Huazhong University of Science and Technology \n" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 525, + 204 + ], + "type": "text", + "content": "Microsoft Corporation \n{xiaofanbai, shixinli, linda, linchenyu} @hust.edu.cn, {binzhu, dongmeiz} @microsoft.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 255, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 411 + ], + "type": "text", + "content": "Cloud-based AI systems offer significant benefits but also introduce vulnerabilities, making deep neural network (DNN) models susceptible to malicious tampering. This tampering may involve harmful behavior injection or resource reduction, compromising model integrity and performance. To detect model tampering, hard-label fingerprinting techniques generate sensitive samples to probe and reveal tampering. Existing fingerprinting methods are mainly based on gradient-defined sensitivity or decision boundary, with the latter showing a manifest superior detection performance. However, all existing fingerprinting methods either suffer from insufficient sensitivity or incur high computational costs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 411, + 290, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 290, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 290, + 593 + ], + "type": "text", + "content": "In this paper, we theoretically analyze the black-box co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity. Based on this, we further propose Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight approach for hard-label tampering detection that inherently and efficiently combines the strengths of existing fingerprinting techniques. SDBF places fingerprint samples near the steep decision boundary, where the outputs of samples are inherently highly sensitive to tampering. We also design a Max Boundary Coverage Strategy (MBCS), which enhances samples' diversity over the decision boundary. Theoretical analysis and extensive experimental results show that SDBF outperforms existing SOTA hard-label fingerprinting methods in both sensitivity and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have demonstrated excellent capabilities across various domains, such as computer vision [12, 18, 24]. An increasing number of applications have incorporated DNN models. Machine Learning as a Service (MLaaS) platforms, such as Google AutoML [9], Microsoft Azure ML [21], and Amazon SageMaker [1]," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 230, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 308, + 230, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 434 + ], + "type": "image", + "image_path": "0cdb057ed36d004143488473c4d414e000e8237e180a99c43b1e8cf87d56554f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 441, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 441, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 547, + 464 + ], + "type": "text", + "content": "Figure 1. Overview of potential model tampering and public tampering detection for models deployed in MLaaS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 546, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 508 + ], + "type": "text", + "content": "leverage powerful pre-trained DNN models, allowing users with limited computational resources to access cutting-edge technology." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 653 + ], + "type": "text", + "content": "However, models deployed in the cloud often operate beyond the direct control of their owners, leaving them vulnerable to unauthorized/illegal model tampering as shown in Fig. 1. For instance, unscrupulous service providers might compress a model—reducing its memory and computational footprint—as a cost-saving measure without the owner's consent, as demonstrated in prior work [3, 22]. Alternatively, adversaries could inject malicious behaviors into the model's parameters to manipulate the model's outputs [10, 14, 15, 27, 29]. Such tampering not only compromises the model's integrity, security, and overall performance but also erodes users' trust and the reliability of the system." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Many works [2, 6, 13, 17, 31, 33-35] have focused on detecting model tampering. Due to the black-box nature of the models deployed in the cloud, hard-label fingerprinting techniques are the most practical but also most challenging among existing tampering detection methods as their non" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 1, + 494, + 19 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 1, + 494, + 19 + ], + "spans": [ + { + "bbox": [ + 145, + 1, + 494, + 19 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 24, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 24, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 24, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29278" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "invasive and black-box nature aligns well with the MLaaS in practice. Existing hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]. The former enhances the sensitivity of fingerprint samples' output to model tampering by maximizing gradient-defined sensitivity, which is the " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " norm of the gradient of the output with respect to the model parameters, and these fingerprinting techniques have been widely demonstrated to be much less effective than the latter. The latter improves the sensitivity of fingerprint samples to tampering by minimizing the distance to the decision boundary. In summary, both of these approaches are still not sensitive enough to detect certain types of model tampering, especially subtle ones. However, detecting subtle model tampering is crucial, as the end user may need to confirm if the model they are using is really the one they want to use." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 361 + ], + "type": "text", + "content": "Additionally, [11] introduced a meta-learning-based fingerprinting approach. However, due to the inherent complexity of meta-learning, this method incurs significantly higher overhead compared to the two aforementioned categories, making it impractical for resource-constrained model owners or users. We argue that, given the very purpose of black-box model tampering detection, a lightweight design for fingerprinting method is essential for its widespread application." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 635 + ], + "type": "text", + "content": "In this paper, recognizing the importance of closing to decision boundary for tampering detection with fingerprint samples, we theoretically analyze the co-optimal tampering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity in a black-box detection manner. Based on this, we further propose a novel lightweight hard-label tampering detection fingerprinting approach, Steep-Decision-Boundary Fingerprinting (SDBF) to improve the fingerprint samples' sensitivity with the inherent highly sensitive nature of samples near the steep decision boundary. Specifically, SDBF places fingerprint samples close to the steep decision boundary, where the outputs exhibit high sensitivity to the tampering-induced shift of decision boundary. To achieve this, SDBF perturbs normal input samples to maximize the " + }, + { + "bbox": [ + 46, + 361, + 289, + 635 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 361, + 289, + 635 + ], + "type": "text", + "content": " norm of the logit output while simultaneously minimize its variance. The fingerprint samples are bound to be almost indistinguishable from the original normal samples. Moreover, to further enhance detection performance when using multiple fingerprint samples to detect, we propose a Max Boundary Coverage Strategy (MBCS). This strategy enables the fingerprint samples to explore a wide range of different regions around the decision boundary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 635, + 289, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 289, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 289, + 695 + ], + "type": "text", + "content": "Our theoretical analysis and extensive experimental results demonstrate that the sensitivity of SDBF is superior to that of existing state-of-the-art (SOTA) hard-label fingerprinting methods while being time and memory-efficient. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 701, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 289, + 714 + ], + "type": "text", + "content": "- We theoretically analyze the black-box co-optimal tam" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 72, + 547, + 302 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 324, + 72, + 547, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 547, + 109 + ], + "type": "text", + "content": "pering detection sensitivity of fingerprint samples in the context of decision boundary and gradient-defined sensitivity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 117, + 547, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 117, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 117, + 547, + 177 + ], + "type": "text", + "content": "- We introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel lightweight fingerprinting method that inherently and efficiently combines the strengths of fingerprinting techniques based on gradient-defined sensitivity and decision boundary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 186, + 547, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 547, + 246 + ], + "type": "text", + "content": "- We propose a Max Boundary Coverage Strategy (MBCS) to ensure that the fingerprint samples widely explore the decision boundary, further enhancing the detection performance when using multiple fingerprint samples to detect model tampering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 254, + 547, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 547, + 302 + ], + "type": "text", + "content": "- We present both theoretical analysis and extensive experimental results that demonstrate the superior performance and efficiency of SDBF compared to existing SOTA hard-label fingerprinting methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 313, + 392, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 313, + 392, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 392, + 326 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 333, + 485, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 485, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 485, + 346 + ], + "type": "text", + "content": "2.1. Hard-Label Tampering Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 351, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 483 + ], + "type": "text", + "content": "Tampering detection for DNN models aims to determine whether a model has been altered. Since MLaaS operates as a black-box to users, existing methods focus on hard-label tampering detection, where only the top-1 label is accessible. In this setting, watermarking and fingerprinting are widely used. Watermarking [35], an invasive method, embeds a fragile watermark during training that can later be queried with specific inputs to detect tampering. However, because modifying the model may impact performance, non-invasive fingerprinting techniques, the focus of this paper, have attracted more attention." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 516, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 516, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 516, + 504 + ], + "type": "text", + "content": "2.2. Fingerprinting for Tampering Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 545 + ], + "type": "text", + "content": "Existing black-box hard-label fingerprinting techniques are mainly based on gradient-defined sensitivity [6, 13, 17] or decision boundary [2, 31, 33, 34]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 546, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 677 + ], + "type": "text", + "content": "Gradient-defined sensitivity. Sensitive-Sample Fingerprinting (SSF) [13] introduces fingerprinting for detecting tampering in DNN models for the first time by gradient-defined sensitivity. SSF generates sensitive samples—natural-looking inputs that cause large output changes when the model's parameters are slightly altered. Following SSF, symbolic constraint solvers [6] and Bayesian optimization (BO) [17] have been used to optimize the generation of sensitive samples. However, such sensitive samples have been proven to be ineffective in extensive experiments for many kinds of model tampering." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Decision boundary based sensitivity. PublicCheck [31] uses VQVAE models to generate fingerprint samples by carefully adding perturbations along with certain directions" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29279" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "content": "in the latent space of VQVAE and stopping when it has just crossed the decision boundary of two different categories. Others like AID [2] and DBFW [34] generate fingerprint samples close to the decision boundary of all categories by minimizing output variance using mean square error (MSE) loss functions. Similar to AID and DBFW, IBSF [33] employs a Shannon entropy loss to generate diverse fingerprint samples near different " + }, + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "K(K \\geqslant 2)" + }, + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "content": " categories' decision boundary. While these fingerprinting techniques show significantly superior detection performance than the aforementioned fingerprinting techniques based on gradient-defined sensitivity, they are still not sensitive enough to detect certain types of model tampering, especially subtle ones. Notably, neither one has theoretically justified why samples that are close to the decision boundary are more effective than the faraway, nor has anyone theoretically analyzed the co-optimal tampering detection sensitivity of fingerprint samples in these two contexts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 384 + ], + "type": "text", + "content": "Additionally, [11] proposed a meta-learning-based sensitive fingerprinting method, MiSentry, but due to the inherent high overhead of meta-learning, the cost of MiSentry is significantly higher than the two aforementioned categories, making it unaffordable for those resources-limited model owner/users. We believe that due to the very purpose of black-box tampering detection, the lightweight of a fingerprinting method is crucial for its widespread application." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 397, + 133, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 133, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 133, + 409 + ], + "type": "text", + "content": "3. Threat Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 622 + ], + "type": "text", + "content": "Aligned with existing works, we adopt a threat model that includes white-box generation and black-box detection. Specifically, during the fingerprint samples generation, we assume white-box access to the target DNN, using available normal samples as source data. This assumption is reasonable as fingerprint samples are generated by the model owner. In the detection phase, we assume there is only black-box access to the suspect model, where only an API-based query is allowed, and only the top-1 label of the input query will be returned. Moreover, following [31], we also assume a trustworthy third party that securely stores and distributes the fingerprint samples for public tampering detection, where the DNN service provider is assumed to be untrustworthy and may detect probing fingerprint samples, evading tampering detection. Additionally, adversaries might access some fingerprint samples through some illegal means, and use them to bypass tampering detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 635, + 269, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 269, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 269, + 649 + ], + "type": "text", + "content": "4. Steep Decision Boundary and Sensitivity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "In this section, we first define the decision boundary in the context of the classification task. We then theoretically demonstrate why samples near the decision boundary are sensitive to shifts caused by model tampering. Next, we define the steep decision boundary and analyze how to place" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "content": "fingerprint samples near it to further enhance samples' sensitivity while improving optimization efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 102, + 446, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 102, + 446, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 102, + 446, + 115 + ], + "type": "text", + "content": "4.1. Steep Decision Boundary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": "A DNN is a function " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": ", which maps an input " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": " to an output logits vector " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = f(W, x)" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": ". For classification tasks, let " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": " denotes the number of classes, the logits vector " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": " will then be transformed into a probability vector " + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 121, + 547, + 181 + ], + "type": "text", + "content": " by the following Softmax function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 186, + 546, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 186, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 356, + 186, + 546, + 216 + ], + "type": "interline_equation", + "content": "Y _ {i} = \\operatorname {S o f t m a x} \\left(\\mathbf {y} _ {i}\\right) = \\frac {e ^ {\\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {N} e ^ {\\mathbf {y} _ {j}}} \\tag {1}", + "image_path": "b95cbcf388eb0efb24c61f96120bf8e630afba4fc2b8f3e775286b8a595e456b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "type": "text", + "content": " is the confidence of class " + }, + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 221, + 455, + 232 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "type": "text", + "content": "Definition 4.1 (Decision Boundary). For a classification model, the decision boundary is the hypersurface that divides the feature space into regions, each corresponding to a different class label. Typically, the model assigns the same probabilities to the classes on the decision boundary, and the confidences of these categories will be much greater than the other categories. Specifically, for an " + }, + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "type": "text", + "content": "-class classification model, the decision boundary of " + }, + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "type": "inline_equation", + "content": "K(K \\leqslant N)" + }, + { + "bbox": [ + 304, + 233, + 547, + 341 + ], + "type": "text", + "content": " classes can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 350, + 546, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 350, + 546, + 368 + ], + "spans": [ + { + "bbox": [ + 364, + 350, + 546, + 368 + ], + "type": "interline_equation", + "content": "Y _ {1} = Y _ {2} = \\dots = Y _ {K} \\gg \\sum \\tilde {Y} \\tag {2}", + "image_path": "d3cd2916f681d30c365afbb7fb892b5744bc1f7518b3c7abbe8e2756b7146e60.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\sum \\tilde{Y} (0\\leqslant \\sum \\tilde{Y}\\leqslant 1)" + }, + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "text", + "content": " is the summation of the other " + }, + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "inline_equation", + "content": "N - K" + }, + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "text", + "content": " categories' confidence and " + }, + { + "bbox": [ + 305, + 371, + 546, + 400 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\sum_{i = 1}^{K}Y_{i} + \\sum \\tilde{Y} = 1. \\end{array}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "content": "For black-box tampering detection, the predicted hard label based on the Softmax output " + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "content": " is the only observable feedback from the suspect model. To maximize the sensitivity of fingerprint samples in this setting, we need to maximize the sensitivity of " + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "content": " to force the model to work in Softmax high-sensitivity region, making any tampering-induced " + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{y} = f(W + \\Delta W, x) - f(W, x)" + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "content": " will be nonlinearly amplified to " + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 404, + 547, + 525 + ], + "type": "text", + "content": " by Softmax, avoiding being suppressed by the saturation zone of Softmax when the fingerprint samples are faraway from the decision boundary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": "Proposition 4.2. Given a fingerprint sample " + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": ", the sensitivity of its softmax output " + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "inline_equation", + "content": "||\\Delta \\mathbf{y}||_2" + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": " introduced by model tampering " + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": " will reach the maximum when " + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": " close to the decision boundary, where " + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "inline_equation", + "content": "Y_i = \\frac{1}{N}" + }, + { + "bbox": [ + 304, + 527, + 547, + 577 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": "Proof. To prove Prop.4.2, we need to analyze the Jacobian matrix " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " of vector " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " with respect to vector " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " and the maximum value of the Frobenius norm (equivalent to the " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " norm) of " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": ". First, based on Eq.1, the element " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "J_{ij}" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " is the partial derivative of " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 647, + 546, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 647, + 546, + 673 + ], + "spans": [ + { + "bbox": [ + 370, + 647, + 546, + 673 + ], + "type": "interline_equation", + "content": "J _ {i j} = \\frac {\\partial Y _ {i}}{\\partial \\mathbf {y} _ {j}} = Y _ {i} \\left(\\delta_ {i j} - Y _ {j}\\right) \\tag {3}", + "image_path": "0b3a927eba3bb8842a9cafae7a6930f1aece0f3ba0c2837332f9724da36291c9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "spans": [ + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "text", + "content": " is the Kronecker Delta function and " + }, + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\left\\{ \\begin{array}{l}\\delta_{ij} = 1,i = j\\\\ \\delta_{ij} = 0,i\\neq j \\end{array} \\right." + }, + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "text", + "content": ". Thus, the squared Frobenius norm of " + }, + { + "bbox": [ + 306, + 678, + 547, + 716 + ], + "type": "inline_equation", + "content": "J" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29280" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 137, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 137, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 137, + 83 + ], + "type": "text", + "content": "can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 287, + 194 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\| J \\| _ {F} ^ {2} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(J _ {i j}\\right) ^ {2} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(Y _ {i} \\left(\\delta_ {i j} - Y _ {j}\\right)\\right) ^ {2} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - 2 Y _ {i} + Y _ {i} ^ {2} + \\sum_ {j \\neq i} Y _ {j} ^ {2}\\right) \\tag {4} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - 2 Y _ {i} + Y _ {i} ^ {2}) + Y _ {i} ^ {2} \\sum_ {j \\neq i} Y _ {j} ^ {2} \\\\ \\end{array}", + "image_path": "a0c03364842263c8ec416e6787e7d7b60fee9c2e893ec9aa1ee98ecbe8f2b02e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N} Y_{i} = 1" + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "text", + "content": ", thus " + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "inline_equation", + "content": "\\sum_{j \\neq i} Y_{j}^{2}" + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "text", + "content": " can be replaced by " + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "inline_equation", + "content": "1 - Y_{i}^{2}" + }, + { + "bbox": [ + 47, + 203, + 287, + 232 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 239, + 287, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 239, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 71, + 239, + 287, + 305 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\| J \\| _ {F} ^ {2} = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - 2 Y _ {i} + Y _ {i} ^ {2}\\right) + Y _ {i} ^ {2} \\left(1 - Y _ {i} ^ {2}\\right) \\tag {5} \\\\ = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} (1 - Y _ {i}) \\\\ \\end{array}", + "image_path": "32b1c1cfa6dd8a5e1f5e62fc0e86dc317290ca8466f5d7ccbf116c5ab908c919.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "type": "text", + "content": "Next, we need to find the maximum of Eq.5. For such a constrained optimization problem, we use the Lagrange multiplier method to rigorously prove when the maximum of Eq.5 is achieved. Specifically, we introduce a Lagrange multiplier " + }, + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "type": "text", + "content": " for the constraint " + }, + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N} Y_i = 1" + }, + { + "bbox": [ + 47, + 313, + 287, + 374 + ], + "type": "text", + "content": ", which is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 382, + 287, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 287, + 424 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(Y _ {1}, Y _ {2}, \\dots , Y _ {N}, \\lambda\\right) = \\sum_ {i = 1} ^ {N} Y _ {i} ^ {2} \\left(1 - Y _ {i}\\right) + \\lambda \\left(\\sum_ {i = 1} ^ {N} Y _ {i} - 1\\right) \\tag {6}", + "image_path": "be82eb905214ddc9da213e778808f9ffc4f7dacaff9786cc078b985c00b72176.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "text", + "content": "We now compute the partial derivatives of " + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "text", + "content": " with respect to each " + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "text", + "content": ". The partial derivative of " + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "Y_{i}^{2}(1 - Y_{i})" + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "text", + "content": " with respect to each " + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 47, + 425, + 287, + 460 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 478, + 287, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 503 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial Y _ {i}} \\left(Y _ {i} ^ {2} \\left(1 - Y _ {i}\\right)\\right) = 2 Y _ {i} \\left(1 - Y _ {i}\\right) - Y _ {i} ^ {2} = 2 Y _ {i} - 3 Y _ {i} ^ {2} (7)", + "image_path": "a499ab899a79f34a1a56bb361cf3b5f2439a5584ee022114e7642d161ab6d802.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 510, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 510, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 59, + 510, + 287, + 522 + ], + "type": "text", + "content": "Similarly, for the constraint term, the partial derivative is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 119, + 530, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 530, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 119, + 530, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial Y _ {i}} \\left(\\sum_ {i = 1} ^ {N} Y _ {i} - 1\\right) = 1 \\tag {8}", + "image_path": "5329de3a22a51e1b693de8f8c4f8fc1ed0ed9d173cc7a4f1fbb8d4a5c6b955ee.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 571, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 287, + 594 + ], + "type": "text", + "content": "Let the total derivative of the Lagrangian equal to zero, we have:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 602, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 602, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 76, + 602, + 287, + 617 + ], + "type": "interline_equation", + "content": "2 Y _ {i} - 3 Y _ {i} ^ {2} + \\lambda = 0 \\Longrightarrow 2 Y _ {i} - 3 Y _ {i} ^ {2} = - \\lambda \\tag {9}", + "image_path": "3ea41837dbb1918ad6d95229da21a9e24eade957e4c6c1230fc0a0d3ee26ede1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "content": "Recall that we have the constraint " + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N} Y_i = 1" + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "content": ", to find a feasible solution, since Eq.9 holds for each " + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "inline_equation", + "content": "Y_i" + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "content": ", we can conclude that all " + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "inline_equation", + "content": "Y_i" + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "content": " must be equal, meaning " + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "inline_equation", + "content": "Y_1 = Y_2 = \\dots = Y_N = Y" + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "content": ". Otherwise, the right term can not be the same constant " + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "inline_equation", + "content": "-\\lambda" + }, + { + "bbox": [ + 47, + 625, + 287, + 686 + ], + "type": "text", + "content": ". Then, we can obtain:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 693, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 693, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 110, + 693, + 287, + 717 + ], + "type": "interline_equation", + "content": "N \\cdot Y = 1 \\quad \\Rightarrow \\quad Y = \\frac {1}{N} \\tag {10}", + "image_path": "7cdc59ffaf844208c81c1951c72f67185f7f6089b37e8d942ebb3714b6e52f1c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "text", + "content": "Finally, to verify that this solution is indeed the maximum, we can check the second order derivative value of " + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "||J||_F^2" + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "text", + "content": " at this solution. When all " + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "Y_{i} = \\frac{1}{N}" + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "text", + "content": ", the second order derivative of " + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "||J||_F^2" + }, + { + "bbox": [ + 305, + 72, + 547, + 121 + ], + "type": "text", + "content": " becomes:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 129, + 547, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 547, + 161 + ], + "type": "interline_equation", + "content": "\\frac {\\partial^ {2} \\left| \\left| J \\right| \\right| _ {F} ^ {2}}{\\partial Y _ {i} ^ {2}} = \\sum_ {i = 1} ^ {N} \\frac {\\partial 2 Y _ {i} - 3 Y _ {i} ^ {2}}{\\partial Y _ {i}} = 2 - 6 \\sum_ {i = 1} ^ {N} Y _ {i} < 0. \\tag {11}", + "image_path": "a5289ff788e37a8b30d1cdfb35f72215cb89e018f6268a3e4915a583111354d8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 170, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 170, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 547, + 205 + ], + "type": "text", + "content": "Therefore, this solution is indeed the maximum. In summary, given a fingerprint sample, " + }, + { + "bbox": [ + 305, + 170, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\| J\\| _F^2" + }, + { + "bbox": [ + 305, + 170, + 547, + 205 + ], + "type": "text", + "content": " achieves maximum when it is close to the decision boundary defined in Def.4.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": "However, when " + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": " is faraway from the decision boundary, some elements in " + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "f(W,x)" + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": " are much larger than others, leading to " + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "\\text{softmax}(f(W,x))" + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": " that are close to 0 or 1. In this case, the elements of the Jacobian matrix " + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "J_{ij}" + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": " will be significantly different, especially when " + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "Y_{i} \\approx 1" + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": ", causing the most of other " + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "Y_{j} \\approx 0" + }, + { + "bbox": [ + 305, + 205, + 547, + 277 + ], + "type": "text", + "content": ", that is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 286, + 546, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 286, + 546, + 313 + ], + "spans": [ + { + "bbox": [ + 318, + 286, + 546, + 313 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} Y _ {i} \\approx 1 \\\\ Y _ {j} \\approx 0 \\end{array} \\right. \\Longrightarrow \\left\\{ \\begin{array}{c} J _ {i i} \\approx Y _ {i} \\cdot (1 - Y _ {i}) \\approx 0 \\\\ J _ {i j} \\approx - Y _ {i} \\cdot Y _ {j} \\approx 0, i \\neq j \\end{array} \\right. \\tag {12}", + "image_path": "adb7846e88ac4831c78674a2005286327d20bc8ee950e10f5e68751056c1d1ab.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 322, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 322, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 546, + 346 + ], + "type": "text", + "content": "Hence, " + }, + { + "bbox": [ + 306, + 322, + 546, + 346 + ], + "type": "inline_equation", + "content": "||J||_F^2" + }, + { + "bbox": [ + 306, + 322, + 546, + 346 + ], + "type": "text", + "content": " in this scenario will be significantly smaller as most elements are close to 0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "content": "Definition 4.3 (Steep Decision Boundary). Given a fingerprint sample " + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "content": " and a model tampering " + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "content": ", the steep decision boundary refers to the decision boundary where " + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "inline_equation", + "content": "f(W, x)" + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "content": " shows a rapid change when " + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 354, + 547, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": "Proposition 4.4. Give a sample " + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": " and a DNN model " + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": ", the sensitivity of output " + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "f(W, x)" + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": " is positively correlated to the " + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "inline_equation", + "content": "||f(W, x)||_2" + }, + { + "bbox": [ + 305, + 411, + 545, + 447 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "text", + "content": "Proof. Inspired by [13], since all model tampering inevitably result in different performances in the last layer, the sensitivity of " + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "inline_equation", + "content": "f(W,x)" + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "text", + "content": " practically can be formulated as the following Frobenius norm of Jacobian matrix " + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 305, + 454, + 547, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 376, + 510, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 510, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 376, + 510, + 545, + 536 + ], + "type": "interline_equation", + "content": "\\left| \\left| \\hat {J} \\right| \\right| _ {F} ^ {2} = \\left| \\left| \\frac {\\partial f (W , x)}{\\partial W _ {F C}} \\right| \\right| _ {2} \\tag {13}", + "image_path": "cd047403cbf47afd258baae6b2160786c38a325af84bdbf05c215b572ea481a6.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "W_{FC}" + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "content": " is the weights of the last fully connected layer of " + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "content": ". Since all existing classification models do not use any nonlinear activation functions like " + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "ReLu(\\cdot)" + }, + { + "bbox": [ + 305, + 544, + 545, + 604 + ], + "type": "text", + "content": " in the last layer, the last layer of the model performs the following linear operations on the output of the previous layer:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 353, + 613, + 545, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 613, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 353, + 613, + 545, + 627 + ], + "type": "interline_equation", + "content": "f (W, x) = W _ {F C} \\times f _ {- 2} (W, x) + \\mathbf {b} \\tag {14}", + "image_path": "9eed37605a62cb6261a2c5fcc76ee18d63dfbb8122683b61eeb0a30e48350a03.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "type": "inline_equation", + "content": "f_{-2}(W,x)" + }, + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 305, + 635, + 547, + 670 + ], + "type": "text", + "content": " are the output and bias of the penultimate layer, respectively. Substituting Eq. 14 into Eq. 13, we can get:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 376, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 376, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\left| \\left| \\hat {J} \\right| \\right| _ {F} ^ {2} = \\left| \\left| f _ {- 2} (W, x) \\right| \\right| _ {2} \\tag {15}", + "image_path": "3a2591c55fe48571e11412d1781e5b5a69982f191441c025bbe701dc097a3944.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Combining Eq.15 and Eq.14, when we increase " + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "||f(W,x)||_2" + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": ", the sensitivity of " + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "f(W,x)" + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Delta W" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29281" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "inline_equation", + "content": "\\| \\hat{J}\\| _F^2" + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "text", + "content": " , will increase simultaneously. To valid Prop.4.4, we conduct an experiment by maximizing " + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "inline_equation", + "content": "\\| f(W,x)\\| _2" + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "text", + "content": " while recording " + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "inline_equation", + "content": "\\| \\frac{\\partial f(W,x)}{\\partial\\mathbf{W}_{FC}}\\| _2" + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "text", + "content": " for a ResNet20 trained on CIFAR10, as shown in Fig. 2. The results are highly consistent with Prop.4.4. Therefore, we can improve the sample's sensitivity of " + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "inline_equation", + "content": "f(W,x)" + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "text", + "content": " by increasing " + }, + { + "bbox": [ + 47, + 71, + 288, + 148 + ], + "type": "inline_equation", + "content": "\\| f(W,x)\\| _2" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 164, + 286, + 301 + ], + "blocks": [ + { + "bbox": [ + 48, + 164, + 286, + 301 + ], + "lines": [ + { + "bbox": [ + 48, + 164, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 286, + 301 + ], + "type": "image", + "image_path": "8271a1f4520dad734185fc33a3aa57a632beb4854f741df230729e082b5f5d70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "type": "text", + "content": "Figure 2. Validation results on the relationship between " + }, + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "type": "inline_equation", + "content": "||f(W,x)||_2" + }, + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "type": "inline_equation", + "content": "||\\frac{\\partial f(W,x)}{\\partial W_{FC}}||_2" + }, + { + "bbox": [ + 47, + 310, + 287, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": "Proposition 4.5. Give a fingerprint sample " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " and a DNN model " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": ", simultaneously increasing the sensitivity of softmax " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "(f(W, x))" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " to corresponding " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "f(W, x)" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " and the sensitivity of " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "f(W, x)" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " can drive " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " close to the steep decision boundary and thus further enhance the sensitivity of softmax " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "(f(W, x))" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 47, + 351, + 288, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "type": "text", + "content": "Prop.4.5 can be easily proved with the linear translational invariance of " + }, + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "type": "inline_equation", + "content": "\\text{Softmax}(\\cdot)" + }, + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "type": "text", + "content": ". More details are in the appendix. Based on Def.4.3 and Prop.4.5, we know how to efficiently place the fingerprint samples near steep decision boundary, where they are highly sensitive to model tampering " + }, + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 46, + 427, + 288, + 511 + ], + "type": "text", + "content": " as it inherently combining the strength of fingerprinting methods based on gradient-defined sensitivity and decision boundary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 272, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 272, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 272, + 529 + ], + "type": "text", + "content": "5. Steep-Decision-Boundary Fingerprinting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 535, + 288, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 288, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 288, + 596 + ], + "type": "text", + "content": "In this section, we introduce a novel sensitive fingerprinting method named Steep-Decision-Boundary Fingerprinting (SDBF). Our goal is to leverage the highly sensitive nature of the samples near the steep decision boundary to detect model tampering, as shown in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 222, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 222, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 222, + 612 + ], + "type": "text", + "content": "5.1. Fingerprint Samples Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Different from [13], we do not directly optimize Eq.13 and Eq.5 as their poor efficiency. We will further explain this in the appendix. Fortunately, based on Prop.4.2, Prop. 4.4 and Prop. 4.5, we can maximize the fingerprint samples' sensitivity by maximizing " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "||\\mathbf{y}||_2" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " while minimizing the variance of " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{y})" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". Similar to IBSF [33], we can also freely choose to place fingerprint samples at " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "K(K\\leqslant N)" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " categories' decision boundary, which will help increase the diversity of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 76, + 425, + 180 + ], + "blocks": [ + { + "bbox": [ + 313, + 76, + 425, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 76, + 425, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 76, + 425, + 180 + ], + "type": "image", + "image_path": "74dfa27de7b4ea707353750cd5976de4a8e8f8b59c118f24140ecc0ca432ee6c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 194, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 228 + ], + "type": "text", + "content": "Figure 3. Illustration of SDBF, gradient-defined fingerprinting SSF, and decision-boundary-based fingerprinting IBSF. SDBF shows an inherent high sensitivity to tampering-introduced boundary shift." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 78, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 427, + 78, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 427, + 78, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 427, + 78, + 545, + 178 + ], + "type": "image", + "image_path": "458c102fdccdb1ee207d49513250ee65516668a51bb2cdf3ad692f0819af0c7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "content": "the fingerprint samples. Specifically, let " + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathcal{C}_{sub}}" + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "content": " denote the logits output of the chosen categories and " + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{y}}_{\\mathcal{C}_{sub}} = \\frac{1}{K}\\sum_{i=1}^{k}\\mathbf{y}_i" + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "content": " denote the averaged logits of the categories in " + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sub}" + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "content": ", our objective function for a sample " + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 232, + 547, + 281 + ], + "type": "text", + "content": " can be formalized as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 297, + 545, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 297, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 315, + 297, + 545, + 345 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (x, \\mathcal {C} _ {s u b}) = \\sigma (\\mathbf {y} _ {\\mathcal {C} _ {s u b}}) - \\lambda \\cdot | | \\mathbf {y} _ {\\mathcal {C} _ {s u b}} | | _ {2} \\\\ = \\frac {1}{K} \\cdot \\sum_ {i = 1} ^ {K} \\left(\\mathbf {y} _ {i} - \\overline {{\\mathbf {y}}} _ {\\mathcal {C} _ {s u b}}\\right) ^ {2} - \\lambda \\cdot \\sum_ {i = 1} ^ {K} \\left\\| \\mathbf {y} _ {i} \\right\\| _ {2} \\tag {16} \\\\ \\end{array}", + "image_path": "38b94a3ea019231b28e29fcc2c859eaeb05a3f4967935398c6912cb09fc16336.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": " is the coefficient and " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "i\\in \\mathcal{C}_{sub}" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": ". Thus, given a normal sample " + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 350, + 545, + 387 + ], + "type": "text", + "content": ", we can drive it close to the steep decision boundary by minimizing the objective function in Eq.16:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 395, + 545, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 395, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 381, + 395, + 545, + 428 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\Delta x} \\mathcal {L} (x + \\Delta x, \\mathcal {C} _ {s u b}) \\tag {17} \\\\ s. t. | | \\Delta x | | _ {p} \\leqslant \\epsilon \\\\ \\end{array}", + "image_path": "0175b55f36eec19076c82a1c7b0d9e60a138e9b9ee0c3e8888e5242325dad237.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "content": " is the input bound. The constraint " + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "inline_equation", + "content": "\\| \\Delta x\\| _p\\leqslant \\epsilon" + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "content": " bounds the perturbation " + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 304, + 431, + 547, + 515 + ], + "type": "text", + "content": " in the input space, making the generated fingerprint sample perceptually similar to the original normal one. This naturalness makes it hard for an adversary to distinguish whether the input is a probe query by input anomaly detection. We solve Eq. 17 by using projected gradient descent (PGD) like [19]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 527, + 545, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 527, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 345, + 527, + 545, + 540 + ], + "type": "interline_equation", + "content": "x _ {i + 1} = \\operatorname {C l i p} _ {\\epsilon} \\left(x _ {i} + \\eta \\cdot \\nabla_ {x _ {i}} \\mathcal {L} \\left(x _ {i}, \\mathcal {C} _ {\\text {s u b}}\\right)\\right) \\tag {18}", + "image_path": "d714199cf04a207df77bfd1a9ca47da5de09321231f25fdfc851e5efcadface1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "type": "inline_equation", + "content": "x_{i} = x_{i - 1} + \\Delta x_{i}" + }, + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "type": "text", + "content": " is the generated fingerprint sample at the " + }, + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 544, + 545, + 567 + ], + "type": "text", + "content": "-th iteration." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": "The optimization eventually yields a fingerprint sample positioned close to " + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sub}" + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": " steep decision boundary. Following [33], to reduce the perturbation effort, we select the top " + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_{sub}|" + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": " confident categories in " + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": " of original normal " + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sub}" + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": " by default. This sample-dependent " + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sub}" + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": " will enhance the diversity of generated sensitive fingerprint samples, making it more robust to partial leakage." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "spans": [ + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "type": "text", + "content": "Remark. In the generation phase, except for fingerprint sample " + }, + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "type": "text", + "content": ", we will also collect its corresponding top-1 label as the fingerprint label " + }, + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 321, + 665, + 533, + 702 + ], + "type": "text", + "content": ". In the de" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29282" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "spans": [ + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "text", + "content": "tection phase, " + }, + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "text", + "content": " will be used to query the suspect model to get the returned hard label " + }, + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{predicted}" + }, + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "text", + "content": " and then we will determine whether the model has been tampered with by verifying whether " + }, + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{predicted} = \\mathcal{V}" + }, + { + "bbox": [ + 62, + 79, + 272, + 128 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 230, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 230, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 230, + 157 + ], + "type": "text", + "content": "5.2. Max Boundary Coverage Strategy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": "When probing the tampered model with " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "M > 1" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": ") fingerprint samples, we collect their returned hard labels. The model is considered untampered if all the labels match the corresponding fingerprint labels. To maximize the diversity of the " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": " fingerprint samples over the steep decision boundary, we propose the Max Boundary Coverage Strategy (MBCS) to make them widely explore the decision boundary. Specifically, we first randomly select a different fingerprint sample, denoted as " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": ". Then, let " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "m \\leqslant M" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": " denote the number of selected fingerprint samples and the sample to select in " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": "-th iteration, we iteratively choose the remaining " + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 46, + 162, + 288, + 307 + ], + "type": "text", + "content": " samples to maximize the boundary diversity score:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 316, + 287, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 316, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 112, + 316, + 287, + 347 + ], + "type": "interline_equation", + "content": "\\max _ {x _ {i}} \\mathcal {L} _ {d i v} = \\sum_ {i = 1} ^ {N} c _ {i} \\cdot \\log c _ {i} \\tag {19}", + "image_path": "8083945a90ea9d11331256b5c7e20641f3ea0dee0c601e827c4a7a06e8a08cb4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 353, + 214, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 214, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 214, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 353, + 214, + 364 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 353, + 214, + 364 + ], + "type": "text", + "content": " is the number of categories and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 373, + 287, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 373, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 100, + 373, + 287, + 405 + ], + "type": "interline_equation", + "content": "c _ {i} = \\frac {1}{m} \\cdot \\sum_ {j = 1} ^ {m} \\operatorname {s o f t m a x} _ {i} \\left(f \\left(x _ {j}\\right)\\right) \\tag {20}", + "image_path": "d8aa14bd7ae586b7212698c5155bdf8e65c156bb3f4f24c65855ecde5e96ceaa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\text{softmax}_i(\\cdot)" + }, + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "text", + "content": " denotes softmax operation along with " + }, + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "inline_equation", + "content": "i - th" + }, + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "text", + "content": " dimension." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": "By repeating the " + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "inline_equation", + "content": "MBCS\\mathcal{R}" + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": " times, we can get a " + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": "-length and " + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": "-width probe fingerprint sequence. For each probe query, we will send " + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": " probe fingerprint samples to the suspect model until we go through all " + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\times M" + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": " fingerprint samples. In our experiments, we use " + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "inline_equation", + "content": "MBCS" + }, + { + "bbox": [ + 46, + 438, + 287, + 522 + ], + "type": "text", + "content": " as a post-processing technique to enhance the diversity of the fingerprint samples over the decision boundary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 531, + 190, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 190, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 190, + 544 + ], + "type": "text", + "content": "6. Experimental Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 551, + 218, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 218, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 218, + 563 + ], + "type": "text", + "content": "6.1. Datasets, Models, and Baselines" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "To evaluate the performance of SDBF, we use various common classification datasets, including CIFAR10 [16], GTSRB [28], and ImageNet [25]. To evaluate the generalizability of SDBF across different model architectures, we select three different popular architectures for the above three datasets. Specifically, we use ResNet20 [12] for CIFAR10, GoogLeNet [30] for GTSRB, and vision transformer ViT-B-32 [7] pretrained by Pytorch for ImageNet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "To assess the superiority of SDBF compared to existing SOTA fingerprinting methods, we select four SOTA fingerprinting methods, SSF [13], PublicCheck [31], IBSF [33], and MiSentry [11], as the baselines. We utilize the default" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "settings outlined in their respective papers unless explicitly specified otherwise. For each fingerprinting method, we independently generate 1000 fingerprint samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 112, + 512, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 112, + 512, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 112, + 512, + 125 + ], + "type": "text", + "content": "6.2. Tampering Settings and Detection Rate" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 130, + 547, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 130, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 304, + 130, + 547, + 309 + ], + "type": "text", + "content": "Tampering Settings. In our experiments, we evaluate SDBF and the baselines on the 9 types of model tampering selected from BIVBench [11], including 4 hardest-to-detect model tampering, targeted attack [15,27,29], finetuning the last layer with " + }, + { + "bbox": [ + 304, + 130, + 547, + 309 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 130, + 547, + 309 + ], + "type": "text", + "content": " learning rate, and Unlearning [8]/Online Learning (On-Learning) [26] makes a model forget/newly learn specific data, and 5 common model tampering consists of Backdoor [10], Bit-flipping Attack [23], Quantization [4], Pruning [20], and independently trained models with Different Architecture (Different Arch.) on the same dataset. All the tampering settings are aligned with BIVBench [11]. In this paper, the reported experimental results for each tampering type are averaged over 10 independently tampered models to mitigate randomness in evaluating tampering detection performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "text", + "content": "Tampering Detection Rate. For each tamper detection probe, " + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "inline_equation", + "content": "M \\geqslant 1" + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "text", + "content": ") fingerprint samples are selected from a pool of 1000 fingerprint samples to query the test model. The model's top-1 predicted labels are then compared with the corresponding fingerprint labels. If all predicted labels match the fingerprint labels, the model is classified as tampered; otherwise, it is marked as tampered. This procedure is repeated 100 times for each test model. Since 10 tampered models are generated for each tampering type, a total of 1000 tests are performed for each tampering type. The tampering detection rate is the proportion of the 1000 probe queries in which tampering is successfully detected. Note that there are no false positives (FP) in the tampering detection of classification model, as the classification model " + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "text", + "content": " provides a stable mapping from input " + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "text", + "content": " to output " + }, + { + "bbox": [ + 304, + 310, + 547, + 490 + ], + "type": "inline_equation", + "content": "f(x)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 491, + 528, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 528, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 528, + 503 + ], + "type": "text", + "content": "6.3. Perceptual Quality of Fingerprint Samples" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 509, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 628 + ], + "type": "text", + "content": "Fig. 4 shows a subset of fingerprint samples generated by SDBF and four baselines, all constrained by the same " + }, + { + "bbox": [ + 304, + 509, + 547, + 628 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 509, + 547, + 628 + ], + "type": "text", + "content": " norm bound. Fingerprint samples generated by SDBF closely resemble those from SSF, IBSF, and MiSentry, with imperceptible pixel-level perturbations added to the original images. In contrast, fingerprint samples generated by PublicCheck exhibit noticeable artifacts, with semantic-level alterations to the original images, due to the use of generative models. For example, the distorted face in the 8th column of Fig. 4 highlights this effect." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 635, + 503, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 503, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 503, + 647 + ], + "type": "text", + "content": "6.4. Effectiveness of Tampering Detection" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Single Fingerprint Sample Detection. We first evaluate the detection performance in the single fingerprint sample detection setting, which only uses one fingerprint sample to query the suspect model. Specifically, following [33], we comprehensively evaluate SDBF and IBSF with " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " ranging" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29283" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 96, + 543, + 328 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Tamper detection rate (%) of detecting tampering in a single fingerprint sample setting on CIFAR10 and GTSRB. For fine-tuning(FT), \"Last\" specifies only fine-tuning the last layer, with numbers indicating the " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "{10}^{-5}" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": " learning rate. The \"On-Learning\" refers to online learning." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 96, + 543, + 328 + ], + "lines": [ + { + "bbox": [ + 49, + 96, + 543, + 328 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 543, + 328 + ], + "type": "table", + "html": "
DatasetMethodFTLast-5BackdoorT-AttackUnlearningQuantizationDifferent Arch.PruningOn-LearningBit-flipping
CIFAR10SSF0.064.911.43.266.379.766.76.784.4
MiSentry45.986.483.868.490.793.589.568.288.9
PublicCheck21.673.636.928.360.865.169.525.485.7
IBSFK=224.577.442.537.578.283.381.247.587.8
K=435.983.655.359.489.188.588.356.289.4
K=642.887.563.569.290.490.890.569.588.6
K=844.589.466.171.390.992.791.371.689.5
K=1049.490.973.272.891.393.891.976.490.1
SDBF(ours)K=238.181.456.456.586.685.384.158.188.4
K=447.587.368.669.889.791.591.266.989.2
K=653.189.979.773.491.592.892.373.588.3
K=856.291.483.175.191.993.792.975.890.3
K=1062.892.585.877.393.294.693.681.189.8
GTSRBSSF0.059.89.51.462.664.559.12.785.1
MiSentry35.286.475.159.487.979.583.759.889.3
PublicCheck18.467.837.321.769.266.570.428.387.3
IBSFK=222.776.339.232.879.872.177.542.688.5
K=431.979.952.955.986.179.783.257.390.4
K=638.485.859.861.889.183.385.563.490.1
K=841.588.563.470.191.285.988.465.189.6
K=1042.891.169.773.391.888.490.368.291.3
SDBF(ours)K=229.180.949.849.484.777.282.953.990.3
K=438.586.662.566.890.584.589.464.391.5
K=646.790.378.471.592.289.691.772.789.2
K=853.291.581.573.692.991.391.977.991.9
K=1057.192.784.678.293.693.892.880.190.7
", + "image_path": "bda21067064b9de89cead5ab148cbcf00e5351d958b383072dfe13dc2df93c73.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "inline_equation", + "content": "K = N = 10" + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "inline_equation", + "content": "K = 500" + }, + { + "bbox": [ + 46, + 338, + 288, + 409 + ], + "type": "text", + "content": " on CIFAR10 and ImageNet, respectively. Tab.1 summarizes the tampering detection rates on CIFAR10 and GTSRB of existing SOTA fingerprinting methods and SDBF. The detailed results on ImageNet are in the appendix. From the results, we can conclude that:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": "SDBF generally outperforms the four SOTA baselines, although exceptions arise for SDBF at " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " when using a single fingerprint sample to perform tampering detection, which can be attributed to MiSentry's inherent sensitivity to such tampering as some subtle tampering is in its model zoo of meta-learning. As " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " increases, SDBF gradually outperforms MiSentry, with improvements of up to " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " on CIFAR10 at " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " for fine-tuning the last layer with " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " learning rate compared to MiSentry. The detection rate of IBSF shares a similar trend with our SDBF when " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " increases, yet it is consistently performing worse than SDBF, evidencing our key intuition: the fingerprint samples close to the steep decision boundary are more sensitive." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Multiple Fingerprint Samples Detection. Next, we evaluate the detection performance in multiple fingerprint samples detection, which uses " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "M(M > 1)" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " fingerprint samples to query the tampered model. If " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{predicted}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " of any one fingerprint sample in the " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " samples differ from the corresponding " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ", the model will be deemed as a tampered one. Following the detection setting in IBSF, MiSentry, and PublicCheck, we repeat the detection probe 1000 times by selecting " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " random fingerprint samples for them. As for SDBF and SSF, we use MBCS and Max Neuron Coverage Strategy (MNCS) [13], which is specifically designed for SSF to select " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " fingerprint samples for each probe, respectively." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 308, + 335, + 455, + 514 + ], + "blocks": [ + { + "bbox": [ + 308, + 335, + 455, + 514 + ], + "lines": [ + { + "bbox": [ + 308, + 335, + 455, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 455, + 514 + ], + "type": "image", + "image_path": "a94eef6b5d24eacf83c0a9211a3664726c2d49684c1bdcc258141ab445ca5f8b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 519, + 547, + 552 + ], + "lines": [ + { + "bbox": [ + 305, + 519, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 519, + 547, + 552 + ], + "type": "text", + "content": "Figure 4. Source images (top) from ImageNet and fingerprint samples generated from them by SDBF (2nd row), IBSF (3rd row), SSF (4th row), MiSentry (5th row), and PublicCheck (bottom)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 458, + 335, + 544, + 514 + ], + "blocks": [ + { + "bbox": [ + 458, + 335, + 544, + 514 + ], + "lines": [ + { + "bbox": [ + 458, + 335, + 544, + 514 + ], + "spans": [ + { + "bbox": [ + 458, + 335, + 544, + 514 + ], + "type": "image", + "image_path": "11df4dcc11f6c66a4a88e3188beb03ed301bad491c40837a65662220c55fc104.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": "The results of detecting fine-tuning the last layer with " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " learning rate on CIFAR10 are shown in Fig.5. From the results, we can see that using multiple fingerprint samples for detection generally performs better than a single one. Among these methods, SDBF achieves perfect detection rates fastest as " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " increases when " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": "Additionally, we also conduct an ablation study on whether " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "MBCS" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": " improves detection performance when using multiple fingerprint samples (see SDBF-MBCS and SDBF Without " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "MBCS" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": " in Fig.5). The results demonstrate " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "MBCS" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": " indeed boosts the multiple fingerprint detection performance. For example, after being equipped with " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "MBCS" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": ", the detection rate of SDBF has been boosted from " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "68.3\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "inline_equation", + "content": "84.5\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 712 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29284" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "inline_equation", + "content": "79.1\\%" + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "inline_equation", + "content": "97.5\\%" + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": ", respectively, making the detection rate of SDBF faster grow up to perfect state than existing SOTA fingerprinting methods." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 114, + 167, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 114, + 167, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 114, + 167, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 114, + 167, + 178 + ], + "type": "image", + "image_path": "f0ee97b6628e0b0c975773bd3442a85d37347b2c7c5e5e0380a0c69501371f91.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 182, + 131, + 190 + ], + "lines": [ + { + "bbox": [ + 97, + 182, + 131, + 190 + ], + "spans": [ + { + "bbox": [ + 97, + 182, + 131, + 190 + ], + "type": "text", + "content": "(a) CIFAR10" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 114, + 286, + 178 + ], + "blocks": [ + { + "bbox": [ + 168, + 114, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 168, + 114, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 168, + 114, + 286, + 178 + ], + "type": "image", + "image_path": "d9c78d4b5a9d992b8d2c28ac081c440c9e88d65bcb7ea7b1b36f079fb829f66d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 182, + 246, + 190 + ], + "lines": [ + { + "bbox": [ + 216, + 182, + 246, + 190 + ], + "spans": [ + { + "bbox": [ + 216, + 182, + 246, + 190 + ], + "type": "text", + "content": "(b) GTSRB" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "type": "text", + "content": "Figure 5. Detection rates on CIFAR10 and GTSRB for fine-tuning the last layer with " + }, + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "type": "text", + "content": " learning rate in multiple fingerprint samples setting. Note that " + }, + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 192, + 286, + 224 + ], + "type": "text", + "content": " for both IBSF and SDBF here." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 229, + 239, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 229, + 239, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 229, + 239, + 242 + ], + "type": "text", + "content": "6.5. Post-Tampering Output Naturalness" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 247, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 415 + ], + "type": "text", + "content": "When the model has been tampered with by adversaries, they might reject the low-confidence query to prevent detection by sensitive fingerprint samples based on the decision boundary. To evaluate the effectiveness of SDBF in responding to such output anomaly detection, we statistically analyzed the top-1 label probabilities of the backdoor-tampered model with SDBF. The results are shown in Fig.7. From the results, we can observe that SDBF achieves comparable covertness to MiSentry, which provides the highest output concealment. This result evidenced that the output probability of the fingerprint samples generated by SDBF is highly sensitive to model tampering and is consistent with our theoretical analysis. Unfortunately, IBSF is at risk of abnormal output rejection, especially on GTSRB." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 421, + 218, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 218, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 218, + 433 + ], + "type": "text", + "content": "6.6. Robustness to Adaptive Attacks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 438, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 546 + ], + "type": "text", + "content": "Adversaries may attempt to bypass SDBF by collecting generated fingerprint samples or generating new ones for adaptive attacks, aiming to manipulate the model while keeping the top-1 labels of the fingerprint samples unchanged. To evaluate SDBF's robustness against such attacks, we generate 1,000 additional fingerprint samples, referred to as leaked fingerprints, which are used for launching adaptive attacks. The set of original samples used for tampering detection is called the unleaked fingerprints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 547, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 583 + ], + "type": "text", + "content": "We conduct adaptive backdoor attacks on CIFAR10, as shown in Fig. 6 (a). As training progresses, the detection rate for the leaked fingerprints drops to " + }, + { + "bbox": [ + 46, + 547, + 287, + 583 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 46, + 547, + 287, + 583 + ], + "type": "text", + "content": ", while the detection" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 602, + 166, + 672 + ], + "blocks": [ + { + "bbox": [ + 50, + 602, + 166, + 672 + ], + "lines": [ + { + "bbox": [ + 50, + 602, + 166, + 672 + ], + "spans": [ + { + "bbox": [ + 50, + 602, + 166, + 672 + ], + "type": "image", + "image_path": "b5eddbcf0af591f6238eff7b415587bbe1b7a8bbd4ebe0d786ad38fc685a2917.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 673, + 111, + 678 + ], + "lines": [ + { + "bbox": [ + 99, + 673, + 111, + 678 + ], + "spans": [ + { + "bbox": [ + 99, + 673, + 111, + 678 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 166, + 601, + 286, + 673 + ], + "blocks": [ + { + "bbox": [ + 166, + 601, + 286, + 673 + ], + "lines": [ + { + "bbox": [ + 166, + 601, + 286, + 673 + ], + "spans": [ + { + "bbox": [ + 166, + 601, + 286, + 673 + ], + "type": "image", + "image_path": "0a9880fedad863154345aeb7bdaafb3ac60575c6320a56e25678de76d45ff209.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 674, + 228, + 678 + ], + "lines": [ + { + "bbox": [ + 217, + 674, + 228, + 678 + ], + "spans": [ + { + "bbox": [ + 217, + 674, + 228, + 678 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 678, + 287, + 721 + ], + "lines": [ + { + "bbox": [ + 46, + 678, + 287, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 287, + 721 + ], + "type": "text", + "content": "Figure 6. (a) Tampering detection rates of detecting adaptive backdoor attacks on CIFAR10 at " + }, + { + "bbox": [ + 46, + 678, + 287, + 721 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 678, + 287, + 721 + ], + "type": "text", + "content": " in a single fingerprint sample setting. (b) Adaptive tampering detection rates when a larger number of fingerprints are leaked." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 65, + 385, + 117 + ], + "blocks": [ + { + "bbox": [ + 307, + 65, + 385, + 117 + ], + "lines": [ + { + "bbox": [ + 307, + 65, + 385, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 65, + 385, + 117 + ], + "type": "image", + "image_path": "5e234e555a7b52c66bd7c1e1ec112542757ce1c9d6ddc7b33233d9086249876f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 388, + 65, + 463, + 117 + ], + "blocks": [ + { + "bbox": [ + 388, + 65, + 463, + 117 + ], + "lines": [ + { + "bbox": [ + 388, + 65, + 463, + 117 + ], + "spans": [ + { + "bbox": [ + 388, + 65, + 463, + 117 + ], + "type": "image", + "image_path": "66456b913dae7c6cc6573418f1f1f1f76a3c6dc1966992b1dceae02476dd7c1b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 467, + 65, + 543, + 118 + ], + "blocks": [ + { + "bbox": [ + 467, + 65, + 543, + 118 + ], + "lines": [ + { + "bbox": [ + 467, + 65, + 543, + 118 + ], + "spans": [ + { + "bbox": [ + 467, + 65, + 543, + 118 + ], + "type": "image", + "image_path": "6931aab78375b367b04fe3be3b8ce7548f11074564e1b8e917515d62c86e3913.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 120, + 545, + 153 + ], + "lines": [ + { + "bbox": [ + 305, + 120, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 545, + 153 + ], + "type": "text", + "content": "Figure 7. Reuslts on post-tampering output naturalness. The black vertical dashed line shows the lowest " + }, + { + "bbox": [ + 305, + 120, + 545, + 153 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 305, + 120, + 545, + 153 + ], + "type": "text", + "content": " value of the confidence in the top-1 label of normal inputs." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 310, + 176, + 544, + 247 + ], + "blocks": [ + { + "bbox": [ + 306, + 156, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 156, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 544, + 167 + ], + "type": "text", + "content": "Table 2. Average generation computational cost (in seconds/GBs)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 176, + 544, + 247 + ], + "lines": [ + { + "bbox": [ + 310, + 176, + 544, + 247 + ], + "spans": [ + { + "bbox": [ + 310, + 176, + 544, + 247 + ], + "type": "table", + "html": "
MethodCIFAR-10GTSRBImageNet
MiSentry1641.43s/9.4GB1639.77s/11.3GB6112.89s/23.8GB
SSF1.39s/1.2GB1.42s/1.6GB184.13s/12.9GB
PublicCheck0.71s/1.1GB0.74s/1.3GB34.51s/6.7GB
SDBFK=20.20s/0.2GB0.21s/0.4GB8.48s/3.1GB
K=100.22s/0.2GB0.24s/0.4GB8.59s/3.1GB
IBSFK=20.19s/0.2GB0.21s/0.4GB8.44s/3.1GB
K=100.22s/0.2GB0.23s/0.4GB8.57s/3.1GB
", + "image_path": "5bc4982cfd4ee41206862fd742b3abe7f2003a4c71c1ac819c16e8fc5ec8e595.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": "rate for the unleaked fingerprints remains relatively high, with only a " + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": " decrease compared to the baseline. Such robustness of SDBF is attributed to the diversity of its source samples and sample-dependent " + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{sub}" + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": ". We also conduct a stress test by gradually increasing the number of leaked fingerprint samples to 10,000, as shown in Fig. 6 (b). Despite the large number of leaked fingerprints, the detection rate of the unleaked fingerprints decreased but remained around " + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 255, + 545, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "text", + "content": "Moreover, we conduct an experiment to explore whether the steep decision boundary fingerprinting is applicable to the smoothed models such as adversarial training (AT) [32]. We find although AT slightly increased the optimization time cost of SDBF and slightly reduced its performance when " + }, + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "text", + "content": ", SDBF remains efficient and effective when " + }, + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 351, + 546, + 435 + ], + "type": "text", + "content": " is larger. The detailed results are shown in the appendix." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 440, + 423, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 423, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 423, + 452 + ], + "type": "text", + "content": "6.7. Computational Cost" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 457, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 554 + ], + "type": "text", + "content": "Table 2 compares the computational cost of SDBF with existing SOTA fingerprinting methods, evaluated on a single NVIDIA RTX 3090Ti GPU. We can observe that: (1) Increasing " + }, + { + "bbox": [ + 304, + 457, + 545, + 554 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 457, + 545, + 554 + ], + "type": "text", + "content": " slightly raises the cost for both SDBF and IBSF; (2) SDBF incurs a computational overhead similar to IBSF, with both methods exhibiting minimal costs; and (3) Even with high-resolution ImageNet images and large-scale models like ViT, SDBF still remains highly efficient." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 562, + 378, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 573 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "In this work, we introduce Steep-Decision-Boundary Fingerprinting (SDBF), a novel method for detecting hard-label tampering in DNN models. Our approach leverages the sensitive nature of steep decision boundary and maximizes fingerprint sample diversity over the decision boundary through the Max Boundary Coverage Strategy (MBCS). Through both theoretical analysis and extensive experiments, we have shown that SDBF outperforms existing SOTA hard-label fingerprinting techniques in both sensitivity and efficiency, making SDBF an effective and efficient solution for ensuring the integrity of cloud-deployed DNN models." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29285" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 114 + ], + "type": "text", + "content": "This work was partially supported by National Natural Science Foundation of China under Grant No.62272175." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 125, + 107, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 125, + 107, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 125, + 107, + 137 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 145, + 289, + 718 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 145, + 288, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 145, + 288, + 177 + ], + "spans": [ + { + "bbox": [ + 53, + 145, + 288, + 177 + ], + "type": "text", + "content": "[1] Amazon Web Services. Amazon sagemaker. https://aws.amazon.com/sagemaker, 2023. Last accessed March 20, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 178, + 289, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 289, + 221 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 289, + 221 + ], + "type": "text", + "content": "[2] Omid Aramoon, Pin-Yu Chen, and Gang Qu. Aid: Attesting the integrity of deep neural networks. In Proceedings of 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 19-24, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 223, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 223, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 223, + 288, + 255 + ], + "type": "text", + "content": "[3] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. A survey of model compression and acceleration for deep neural networks. arXiv preprint arXiv:1710.09282, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 256, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 288, + 277 + ], + "type": "text", + "content": "[4] PyTorch Contributors. Quantization. https://pytorch.org/docs/stable/quantization.html, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 278, + 288, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 278, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 53, + 278, + 288, + 343 + ], + "type": "text", + "content": "[5] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 2206-2216. PMLR, 13-18 Jul 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 345, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 345, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 345, + 288, + 388 + ], + "type": "text", + "content": "[6] Amel Nestor Docena, Thomas Wahl, Trevor Pearce, and Yunsi Fei. Sensitive samples revisited: Detecting neural network attacks using constraint solvers. arXiv preprint arXiv:2109.03966, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 388, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 453 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 455, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 455, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 455, + 288, + 509 + ], + "type": "text", + "content": "[8] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9304-9312, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 510, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 531 + ], + "type": "text", + "content": "[9] Google Cloud. AutoML. https://cloud.google.com/automl, 2023. Last accessed March 20, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 288, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 565 + ], + "type": "text", + "content": "[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 566, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 609 + ], + "type": "text", + "content": "[11] Chaoxiang He, Xiaofan Bai, Xiaojing Ma, Bin Benjamin Zhu, Pingyi Hu, Jiayun Fu, Hai Jin, and Dongmei Zhang. Towards stricter black-box integrity verification of deep neural network models. In ACM Multimedia 2024, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 609, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 288, + 654 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 654, + 288, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 288, + 696 + ], + "type": "text", + "content": "[13] Zecheng He, Tianwei Zhang, and Ruby Lee. Sensitive-sample fingerprinting of deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4729-4737, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 697, + 288, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 697, + 288, + 718 + ], + "spans": [ + { + "bbox": [ + 48, + 697, + 288, + 718 + ], + "type": "text", + "content": "[14] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 700 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "text", + "content": "Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 3104-3122, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "type": "text", + "content": "[15] Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of International Conference on Machine Learning, pages 1885-1894. PMLR, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 174 + ], + "type": "text", + "content": "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. M.S. thesis, Department of Computer Science, University of Toronto, 2009." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 547, + 218 + ], + "type": "text", + "content": "[17] Deepthi Praveenlal Kuttichira, Sunil Gupta, Dang Nguyen, Santu Rana, and Svetha Venkatesh. Verification of integrity of deployed deep learning models using bayesian optimization. Knowledge-Based Systems, 241:108238, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 220, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 263 + ], + "type": "text", + "content": "[18] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3431-3440, 2015." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "type": "text", + "content": "[19] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 309, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 331 + ], + "type": "text", + "content": "[20] Microsoft. Neural network intelligence. https://github.com/microsoft/nni, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 332, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 365 + ], + "type": "text", + "content": "[21] Microsoft Azure. Microsoft machine learning. https://azure.microsoft.com/en-us/products/machine-learning, 2023. Last accessed March 20, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 366, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 547, + 420 + ], + "type": "text", + "content": "[22] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11264-11272, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 422, + 547, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 465 + ], + "type": "text", + "content": "[23] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Bit-flip attack: Crushing neural network with progressive bit search. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1211-1220, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 467, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 547, + 520 + ], + "type": "text", + "content": "[24] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 779-788, 2016." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 522, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 547, + 587 + ], + "type": "text", + "content": "[25] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, C. Berg Alexander, and Fei-Fei Li. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 589, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 547, + 643 + ], + "type": "text", + "content": "[26] Ahmed Salem, Apratim Bhattacharya, Michael Backes, Mario Fritz, and Yang Zhang. Updates-leak: Data set inference and reconstruction attacks in online learning. In 29th USENIX Security Symposium, USENIX Security 2020, pages 1291-1308. USENIX Association, 2020." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 644, + 547, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 644, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 308, + 644, + 547, + 700 + ], + "type": "text", + "content": "[27] Ali Shafahi, W. Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29286" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 330 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "Processing Systems 2018, NeurIPS 2018, pages 6106-6116, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "text", + "content": "[28] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural networks*, 32:323-332, 2012." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[29] Octavian Suciu, Radu Marginean, Yigitcan Kaya, Hal Daume III, and Tudor Dumitras. When does machine learning fail? generalized transferability for evasion and poisoning attacks. In Proceedings of 27th USENIX Security Symposium (USENIX Security 18), pages 1299-1316, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 239 + ], + "type": "text", + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 306 + ], + "type": "text", + "content": "[31] Shuo Wang, Sharif Abuadbba, Sidharth Agarwal, Kristen Moore, Ruoxi Sun, Minhui Xue, Surya Nepal, Seyit Camtepe, and Salil Kanhere. Publiccheck: Public integrity verification for services of run-time deep models. In Proceedings of 2023 IEEE Symposium on Security and Privacy (SP), pages 1348-1365. IEEE, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 330 + ], + "type": "text", + "content": "[32] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 328 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 137 + ], + "type": "text", + "content": "adversarial training. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 36246-36263. PMLR, 23-29 Jul 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 140, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 228 + ], + "type": "text", + "content": "[33] Bai Xiaofan, Chaoxiang He, Xiaojing Ma, Bin Benjamin Zhu, and Hai Jin. Intersecting-boundary-sensitive fingerprinting for tampering detection of DNN models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 54402-54413. PMLR, 21-27 Jul 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 228, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 271 + ], + "type": "text", + "content": "[34] Zhaoxia Yin, Heng Yin, Hang Su, Xinpeng Zhang, and Zhenzhe Gao. Decision-based iterative fragile watermarking for model integrity verification. arXiv preprint arXiv:2305.09684, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 274, + 547, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 547, + 328 + ], + "type": "text", + "content": "[35] Renjie Zhu, Ping Wei, Sheng Li, Zhaoxia Yin, Xinpeng Zhang, and Zhenxing Qian. Fragile neural network watermarking with trigger image set. In Proceedings of International Conference on Knowledge Science, Engineering and Management, pages 280-293. Springer, 2021." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29287" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_content_list.json b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e6824fa21f940aa8f192fdd0072861171cf4b07 --- /dev/null +++ b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_content_list.json @@ -0,0 +1,1462 @@ +[ + { + "type": "text", + "text": "SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction", + "text_level": 1, + "bbox": [ + 81, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ZaiPeng Duan ChenXu Dang Xuzhong Hu Pei An Junfeng Ding Jie Zhan YunBiao Xu *Jie Ma Huazhong University of Science and Technology", + "bbox": [ + 86, + 203, + 883, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal 3D occupancy prediction has garnered significant attention for its potential in autonomous driving. However, most existing approaches are single-modality: camera-based methods lack depth information, while LiDAR-based methods struggle with occlusions. Current lightweight methods primarily rely on the Lift-Splat-Shoot (LSS) pipeline, which suffers from inaccurate depth estimation and fails to fully exploit the geometric and semantic information of 3D LiDAR points. Therefore, we propose a novel multimodal occupancy prediction network called SDG-OCC, which incorporates a joint semantic and depth-guided view transformation coupled with a fusion-to-occupancy-driven active distillation. The enhanced view transformation constructs accurate depth distributions by integrating pixel semantics and co-point depth through diffusion and bilinear discretization. The fusion-to-occupancy-driven active distillation extracts rich semantic information from multimodal data and selectively transfers knowledge to image features based on LiDAR-identified regions. Finally, for optimal performance, we introduce SDG-Fusion, which uses fusion alone, and SDG-KL, which integrates both fusion and distillation for faster inference. Our method achieves state-of-the-art (SOTA) performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes dataset, demonstrating its effectiveness and robustness. The code will be released at https://github.com/DzpLab/SDGOCC.", + "bbox": [ + 73, + 305, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate 3D perception of the surrounding environment forms the cornerstone of modern autonomous driving systems and robotics, ensuring efficient planning and safe control [7, 10]. In recent years, advancements in 3D object detection [14, 17, 20, 38, 41] and semantic segmentation [11, 12, 37, 39, 45] have significantly propelled the field of", + "bbox": [ + 75, + 784, + 470, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e92a5da0b3b2d37dee13a72f983719da60602930adfd30d7a502535c3e73f51.jpg", + "image_caption": [ + "Figure 1. Comparisons of the mIoU and inference speed (FPS) of various 3D occupancy prediction methods on the Occ3D-nuScenes validation set. SDG-OCC achieves higher accuracy and real-time inference speed." + ], + "image_footnote": [], + "bbox": [ + 501, + 273, + 890, + 457 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D perception. However, object detection relies on strict bounding boxes, making it difficult to recognize arbitrary shapes or unknown objects, while semantic segmentation struggles with fine-grained classification in complex scenes, especially under occlusion and overlap. In this context, 3D semantic occupancy prediction [30, 32] offers a more comprehensive approach to environment modeling. It simultaneously estimates the geometric structure and semantic categories of scene voxels, assigns labels to each 3D voxel, and provides a more complete perception, showing stronger robustness to arbitrary shapes and dynamic occlusions.", + "bbox": [ + 496, + 535, + 893, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leveraging the complementary strengths of LiDAR and camera data is crucial for various 3D perception tasks. However, due to the heterogeneity between modalities, fusing LiDAR and camera data for 3D occupancy prediction remains challenging. Specifically, cameras provide rich semantic information but lack precise depth details, while LiDAR offers accurate depth information but only captures sparse data, potentially missing comprehensive scene details such as occluded objects. Existing methods often suffer from significant computational burdens (see Fig. 1), with some approaches attempting to leverage the LSS [27] pipeline for real-time performance. Although LSS simulates the uncertainty of each pixel's depth through depth", + "bbox": [ + 496, + 703, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: majie@hust.edu.cn", + "bbox": [ + 94, + 886, + 328, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6751", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb93ae8a357d0683c4277efd18c53cb4e292e738d8ee380ba50f70ffb9b80e7a.jpg", + "image_caption": [ + "(a) Initial BEV features of LSS" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 269, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2597659a28c705aedf7287e49843ddc9c31c24af4908c7bd39831884d7c3413.jpg", + "image_caption": [ + "(b) Our initial BEV features" + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 454, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b94f120607103984eb1d4d92d9fc768bfb8eddc498e2ade016233ea57baa9dd2.jpg", + "image_caption": [ + "(c) Ground truth of BEV", + "Figure 2. (a) BEV feature map of LSS with a shape of $200 \\times 200$ . We can observe that LSS has an extremely low utilization rate for BEV space. (b) The corresponding BEV features in SDG-OCC. Using depth and semantic information, the 2D-to-3D view transformation achieves efficient occupancy and utilization of the BEV features. (c) The corresponding BEV features in Ground Truth. (d) The corresponding BEV features in the occupancy grid." + ], + "image_footnote": [], + "bbox": [ + 89, + 239, + 267, + 376 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac61a03477675b3caad4da7c0bdce97f3e8a1faa68c011e39438bfe40563c601.jpg", + "image_caption": [ + "(d) Occupancy grid of BEV" + ], + "image_footnote": [], + "bbox": [ + 274, + 239, + 452, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distribution (with depth intervals typically set to $0.5\\mathrm{m}$ ), its sparse BEV representation allows only $50\\%$ of the grids to receive valid image features [16] (see Fig. 2 (a)). While increasing the depth interval can improve depth estimation accuracy to mitigate sparsity, it significantly increases computational demands. Additionally, while LiDAR can provide valuable geometric priors, fusion-based methods that process both point clouds and images simultaneously impose heavy computational burdens, thereby increasing the strain on real-time applications.", + "bbox": [ + 75, + 518, + 467, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a multimodal 3D semantic occupancy prediction framework, named SDG-OCC, which aims to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. In this framework, we introduce a semantic and depth-guided view transformation to replace normal BEV feature generation. Specifically, after extracting features from camera data and obtaining semantic segmentation masks and depth distributions through a multi-task head, we use the semantic masks and depth maps provided by LiDAR to construct virtual points via local diffusion and bilinear discretization. Combined with the depth distribution, these points are then projected into the BEV space. The comparison between LSS and our generated BEV features is shown in Fig. 2. The SDG view transformation sig", + "bbox": [ + 75, + 674, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nificantly refines depth estimation accuracy and reduces redundant virtual point seeds, improving both the speed and accuracy of semantic occupancy.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, we introduce a fusion-to-occupancy-driven active distillation module. We first fuse LiDAR and camera features in the BEV space and then unidirectionally selectively transferred multimodal knowledge to image features based on LiDAR-identified regions. Our proposed SDG-Fusion, which includes only fusion, achieved SOTA performance on the Occ3D-nuscene [30] and SurroundOccnuScenes [35] validation dataset. In comparison, SDGKL, which combines fusion and unidirectional distillation, achieves real-time speed with a slight performance penalty.", + "bbox": [ + 498, + 137, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 517, + 289, + 841, + 303 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, aimed at achieving higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective.", + "- We propose a novel view transformation method that leverages the geometric and semantic information of point clouds to guide the 2D-3D view transformation. This significantly enhances the accuracy of depth estimation and improves both the speed and accuracy of semantic occupancy.", + "- We propose a fusion-to-occupancy-driven active distillation module that integrates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Building on this, we present SDG-Fusion for high performance and SDG-KL for faster inference.", + "- Our method achieves SOTA performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes validation dataset, demonstrating the effectiveness of our approach." + ], + "bbox": [ + 500, + 306, + 890, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 638, + 640, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision-Centric Occupancy Perception", + "text_level": 1, + "bbox": [ + 500, + 664, + 830, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by Tesla's autonomous driving perception system, vision-centric occupancy perception has garnered significant attention from both industry and academia. MonoScene [2] is a pioneering work that used only RGB inputs. TPVFormer [9] combines surround multi-camera inputs and uses transformer-based methods to lift features into a tri-perspective view space. SurroundOcc [35] extends high-dimensional BEV features into occupancy features and directly performs spatial cross-attention to generate geometric information. VoxFormer [13] introduces a two-stage transformer-based semantic scene completion framework, capable of outputting complete 3D volumetric semantics from 2D images alone. FlashOcc [42] transforms the channel to height, lifting BEV output to 3D space,", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6752", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "significantly improving operational efficiency. FBOcc [15] proposes a front-to-back view transformation module based on BEV features to address the limitations of different view transformations. Methods like UniOcc [26] and RenderOcc [25] use NeRF [34] to directly predict 3D semantic occupancy, but the rendering speed limits their efficiency. FastOcc [6] improves the occupancy prediction head to achieve a faster inference speed. COTR [21] builds compact 3D occupancy representations through explicit-implicit view transformation and coarse-to-fine semantic grouping. In this paper, we improve the speed and accuracy of 3D semantic occupancy prediction from the BEV space by incorporating the geometric and semantic information of the point cloud into the view transformation.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multi-Modal Occupancy Perception", + "text_level": 1, + "bbox": [ + 76, + 313, + 390, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal occupancy perception leverages the strengths of multiple modalities to overcome the limitations of unimodal perception. OpenOccupancy[32] introduced a benchmark for LiDAR-camera semantic occupancy prediction. Inspired by BEVFusion, OccFusion [22] concatenates 3D feature volumes from different modalities along the feature channels, followed by convolutional layers to combine them. CO-Occ [24] introduced the Geometric and Semantic Fusion (GSFusion) module, identifying voxels containing both point cloud and visual information using k-nearest neighbors (KNN) search. OccGen [31] employs an adaptive fusion module to dynamically integrate occupancy representations from camera and LiDAR branches, using 3D convolutions to determine fusion weights for aggregating LiDAR and camera features. HyDR [36] proposes to integrate multimodal information in both perspective view (PV) and bird's-eye view (BEV) representation spaces. In this paper, we enhance view transformation by incorporating semantic segmentation masks and LiDAR depth maps to achieve higher occupancy accuracy. Additionally, we fuse BEV features from multimodal data and unidirectionally distill them into camera features, improving the accuracy and inference speed of 3D semantic occupancy prediction.", + "bbox": [ + 75, + 338, + 472, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 700, + 210, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 724, + 207, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given joint input from multi-view images and LiDAR data, 3D occupancy prediction aims to estimate the occupancy state and semantic classification of 3D voxels surrounding the ego vehicle. Specifically, the input consists of a $T$ -frame consequent sequence of images $X_{C} \\in \\mathbb{R}^{N_{C} \\times H_{C} \\times W_{C} \\times 3}$ from $N_{C}$ surround-view cameras and point clouds $X_{L} \\in \\mathbb{R}^{N_{L} \\times (3 + d)}$ as multimodal input, represented as $X = \\{X_{C}, X_{L}\\}$ . Here $H_{C}$ , $W_{C}$ represent the height and width of the image, respectively, $N_{L}$ denotes the number of point clouds and $d$ denotes the initial addi", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tional features of the point cloud. Subsequently, we train a neural network to generate an occupancy voxel map $Y \\in \\mathbb{R}^{H \\times W \\times D \\times C_N}$ , where each voxel is assigned a label as unknown, occupied, or a semantic category from $\\{C_0$ to $C_N\\}$ . Here, $N$ denotes the total number of categories of interest, and $H, W, D$ represent the volume dimensions of the entire scene.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall Architecture", + "text_level": 1, + "bbox": [ + 498, + 205, + 696, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of SDGOCC is shown in Fig. 3. It mainly consists of four key modules: image feature encoder to extract image features, semantic and depth-guided view transformation to construct 2D-3D feature transformation, fusion-to-occupancy-driven active distillation for fusing multimodal features and selectively transferring knowledge to the image features, and the occupancy prediction head for final output.", + "bbox": [ + 496, + 229, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Image Encoder", + "text_level": 1, + "bbox": [ + 498, + 359, + 653, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The image feature encoder aims to capture multi-view features, providing a foundation for 2D-3D view transformation. Given RGB images from surround-view cameras, we first use a pre-trained image backbone network, such as classic ResNet [5] or strong Swin-Transformer [19], to extract multi-layer image features $F_{C} \\in \\mathbb{R}^{N_{C} \\times C \\times H \\times W}$ . These features are then aggregated using a feature pyramid network (FPN) [18], which combines fine-grained features and coarse-grained features and down-sampling them to a specific scale, typically 1/16.", + "bbox": [ + 496, + 383, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. SDG View Transformation", + "text_level": 1, + "bbox": [ + 498, + 544, + 741, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LSS pipeline is widely used for converting image features to BEV representations in 3D perception. It constructs virtual points based on a predefined depth range for each pixel and predicts the depth distribution weight $\\alpha$ and context feature $c$ . The feature representation at depth $d$ is given by $p_d = \\alpha_d c$ . All virtual points then are projected into BEV space, where features at each height $Z$ are aggregated to form BEV features. Although LSS handles depth uncertainties and ambiguities by modeling pixel depth using depth distributions, the number of per-pixel features remains large even with a 0.5-meter depth interval, an order of magnitude larger than point features. Meanwhile, BEV features are highly sparse, with less than $50\\%$ of the image features being effective, leading to suboptimal occupancy prediction performance. Reducing the depth interval improves accuracy but significantly increases computational burden and introduces irrelevant features, as most of the occupancy grid remains empty.", + "bbox": [ + 496, + 568, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, we propose a novel view transformation method that leverages the sparse depth information from LiDAR as a prior, diffuses within the same semantic class, followed by linear-increasing [28] and linear-decreasing dis", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6753", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7eaa41de797e1575753024e2a686ba37dba05c7a04dff0c8e0f1f54aeda7a000.jpg", + "image_caption": [ + "Figure 3. The overall architecture of SDG-OCC. $T$ -frame multi-view images and corresponding point clouds are fed into the image and LiDAR backbones to extract features. Image features are processed by a multi-task head to generate semantic masks and depth distributions, combined with LiDAR depth maps to create virtual points for image BEV features. These features are fused with point cloud BEV features and selectively transfer multimodal knowledge to image features based on LiDAR-identified regions. Finally, enhanced features are processed by the occupancy prediction head to generate the occupancy map." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 883, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f1267d7a84bf8fa93762361cf5fdbfced289b27ce922f8217a61af59253824d2.jpg", + "image_caption": [ + "Figure 4. The overall architecture of the SDG view transformation. The image features processed by DepthNet are divided into texture features, depth features, and semantic features. Semantic features generate masks via a segmentation head, which are combined with depth maps from LiDAR for local diffusion and bilinear discretization to create virtual points. These points and their features receive pillar features from texture and depth features via the outer product, and then through BEV pooling to generate the final image BEV features." + ], + "image_footnote": [], + "bbox": [ + 78, + 446, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cretization to generate high-precision virtual point seeds around each co-visible point, as shown in Fig. 4. First, we extract features from multi-camera images and generate semantic segmentation masks via a multi-task head, while simultaneously extracting image textual features and depth distribution weights, with the depth head and semantic head supplementing cross-task information through gated attention. To better utilize semantic information, we select 4x downsampled features for view transformation, as higher downsampling increases the semantic and depth ambiguity", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of pixels.", + "bbox": [ + 500, + 449, + 565, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the differences in sparsity between images and point clouds, we combine image semantic segmentation masks and sparse projected depth maps provided by LiDAR to diffuse depth values within the same semantic category masks, generating a semi-dense extended depth map. This process is as follows:", + "bbox": [ + 496, + 464, + 890, + 554 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\text {t e m p}} (i, j) = \\frac {\\sum_ {(p , q) \\in N (i , j)} D (p , q) \\cdot \\mathbb {I} [ M (p , q) = M (i , j) ]}{\\sum_ {(p , q) \\in N (i , j)} \\mathbb {I} [ M (p , q) = M (i , j) ]}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 560, + 890, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N(i,j)$ represents the circle area with radius $r$ around the current point, and $M(i,j)$ denotes the segmentation mask with $\\mathbf{N}$ category. And $\\mathbb{I}[M(p,q) = M(i,j)]$ checks if the semantic label at $(\\mathrm{p},\\mathrm{q})$ matches that at $(\\mathrm{i},\\mathrm{j})$ , as follows:", + "bbox": [ + 496, + 631, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {I} [ M (p, q) = M (i, j) ] = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} M (p, q) = M (i, j) \\neq 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 712, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "And the final extended depth map $D_{(i,j)}$ replaces the original co-points of the $D_{\\mathrm{temp}}(i,j)$ .", + "bbox": [ + 496, + 763, + 890, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the projection deviation from 2D pixels to 3D points, we apply bidirectional linear incremental discretization to the extended depth map to obtain discrete virtual points, enhancing the accuracy of depth estimation. These steps reduce the number of virtual points, thereby improving inference speed. Finally, the image textual features $F_{t}$ and depth distribution weights $D_{w}$ are calculated by the", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6754", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/52bb2163df066e884e468217a4f2be5104c99f3183ba6883e1cc8dbd79465a70.jpg", + "image_caption": [ + "Figure 5. (a) The pipeline of SDG-Fusion. (b) The pipeline of SDG-KL. (c) The overall architecture of fusion-to-occupancy-driven active distillation. The source feature extracts neighborhood information from the corresponding pixels of the cross feature, where the source serves as the query, and the cross feature acts as both key and value for feature interaction. This interaction is dynamically refined through a gated attention mechanism to produce the fused feature. Subsequently, an occupancy-driven active distillation is used to unidirectionally integrate multimodal information into the image feature." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 472, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "outer product $F_{t} \\otimes D_{w}$ to derive features for each virtual point and generate the BEV features $F_{bev}^{C}$ of the camera through BEV pooling. This method effectively integrates semantic information with sparse depth data, significantly enhancing the accuracy of pixel depth estimation and improving the speed of view transformations.", + "bbox": [ + 75, + 378, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Fusion-to-occupancy-driven Active Distillation", + "text_level": 1, + "bbox": [ + 76, + 481, + 468, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LiDAR branch encompasses point cloud feature extraction, multimodal fusion, and occupancy-driven active distillation (as shown in Fig. 5). Initially, the point cloud data undergoes voxelization and normalization to generate the initial features. We choose SPVCNN as our point-voxel feature encoder due to its efficiency in representing sparse point clouds while effectively preserving fine-grained details. Subsequently, we compress the voxel features at the corresponding scale to generate the BEV features $F_{\\text{bev}}^{L}$ .", + "bbox": [ + 75, + 506, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complementary information from LiDAR and cameras is critical for 3D perception. A naive fusion method typically concatenates LiDAR and image BEV features along the channel dimension to enhance performance. However, feature misalignment due to extrinsic conflicts [3] limits the effectiveness of the fusion. Therefore, we propose a dynamic neighborhood feature fusion module. This module unidirectionally extracts neighborhood features from cross features and dynamically adjusts their weights into the source features using a gating attention mechanism.", + "bbox": [ + 75, + 643, + 470, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "External projection deviations during the BEV feature construction process for LiDAR and images result in misalignment between LiDAR and camera BEV features [3, 29]. Therefore, we adopt neighborhood attention from [4] to extract local patch features corresponding to the pixel from the cross features, and dynamically adjust the weights through gated attention to selectively enhance the fused fea", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ture representation. Specifically, the image features $F_{bev}^{C}$ as the source feature, are represented as a feature vector sequence $F_{img} \\in \\mathbb{R}^{n \\times m}$ , which is projected through a linear layer to obtain the query features $Q_{s} \\in \\mathbb{R}^{n \\times q}$ . Similarly, the LiDAR features $F_{bev}^{L}$ as cross feature are projected to obtain the key $K_{c} \\in \\mathbb{R}^{n \\times q}$ and value $V_{c} \\in \\mathbb{R}^{n \\times v}$ features. The local neighborhood features $F_{\\mathrm{neighbor}}$ for a query point $i$ are computed by the following equation:", + "bbox": [ + 496, + 90, + 893, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {n e i g h b o r}} = \\sigma \\left(\\frac {Q _ {s} ^ {i} \\cdot \\left(K _ {c} ^ {n (i)}\\right) ^ {T} + B (i , n (i))}{\\sqrt {v}}\\right) \\cdot V _ {c} ^ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 220, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n(i)$ represent the neighborhood with the size of $k$ centered at the same position in the cross feature, $B(i,\\rho (i))$ denotes the relative positional biases and $\\sigma$ denotes to Softmax. For each pixel in the feature map, we calculate the local neighborhood features. The fused features $F_{\\text{bev}}^{\\text{fuse}}$ are then obtained from the local neighborhood features through gated attention, as follows:", + "bbox": [ + 496, + 270, + 893, + 377 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {b e v} ^ {f u s e} = \\left(\\sigma \\left(\\operatorname {C o n v} \\left(f _ {\\text {A v g}} \\left(F _ {\\text {n e i g h b o r}}\\right)\\right)\\right) \\cdot F _ {\\text {n e i g h b o r}}, \\quad (4) \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 385, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the sigmoid function and Conv denotes linear transform matrix (e.g., 1x1 convolution), $f_{\\mathrm{Avg}}$ refers Adaptive Average Pooling. The fused features $F_{bev}^{fuse}$ are processed by the occupancy prediction head to obtain the SDG-Fusion model.", + "bbox": [ + 496, + 414, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, to ensure real-time, we propose an occupancy-driven active distillation, where fused features are unidirectionally transferred to the image features. Specifically, LiDAR features are used as the source feature, while image features serve as the cross feature, resulting in LiDAR-dominant fused features. Inspired by [1], we then divide the space into two regions: the active region(AR), where both LiDAR and image features are occupied, and the inactive region(IR), where only LiDAR features are occupied. The details are as follows:", + "bbox": [ + 496, + 492, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA R = \\left(M _ {\\text {f u s e d}, i, j} = 1\\right) \\wedge \\left(M _ {\\text {i m g}, i, j} = 1\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 652, + 890, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI R = \\left(M _ {\\text {f u s e d}, i, j} = 1\\right) \\wedge \\left(M _ {\\text {i m g}, i, j} = 0\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 679, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where a value of 1 in $M_{\\mathrm{mode},i,j}$ indicates that the coordinate is occupied by the respective modality. Typically, the AR region is significantly larger than the IR region. To prevent the model from overemphasizing knowledge distillation in the AR region, we apply adaptive scaling based on the relative sizes of the AR and IR regions, as follows:", + "bbox": [ + 496, + 700, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {I, i, j} ^ {\\left(l _ {n}\\right)} = \\left\\{ \\begin{array}{l l} \\alpha , & \\text {i f} (i, j) \\in A R, \\\\ \\rho \\times \\beta , & \\text {i f} (i, j) \\in I R, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 801, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho = \\frac{N_{AR}}{N_{IR}}$ represents the relative importance of the $IR$ over $AR$ , $\\alpha$ and $\\beta$ are the intrinsic balancing parameters,", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6755", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $N_{AR}$ and $N_{IR}$ are the number of pixels in $AR$ and $IR$ , respectively.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The distillation loss between BEV feature from teacher $F^t$ and student $F^s$ are:", + "bbox": [ + 76, + 121, + 467, + 150 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i s t i l l}} = \\sum_ {c = 1} ^ {C} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} W _ {i, j} \\left(F _ {b e v} ^ {f u s e} - F _ {b e v} ^ {C}\\right) ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 157, + 467, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The network is trained with the sum of distillation and classification loss. The image features $F_{bev}^{C}$ are processed by the occupancy prediction head to obtain the SDG-KL model.", + "bbox": [ + 75, + 209, + 467, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Occupancy Prediction", + "text_level": 1, + "bbox": [ + 76, + 279, + 282, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain 3D prediction outputs from coarse BEV features generated by view transformation, we propose an occupancy prediction system consisting of a BEV feature encoder and an occupancy prediction head. The BEV encoder uses several residual blocks for multi-scale feature diffusion and integrates a feature pyramid to acquire BEV features at the target scale. The occupancy prediction head extracts global features with multiple $3 \\times 3$ convolutional layers and includes a channel-to-height transformation module. This module reshapes the BEV features from a $F_{out} \\in \\mathbb{R}^{B \\times C \\times H \\times W}$ to $F_{final} \\in \\mathbb{R}^{B \\times C_N \\times D \\times H \\times W}$ , where $B, C, W, H$ , and $D$ represent the batch size, channel number, class number, and the dimensions of the 3D space, respectively, with $C = C_N \\times D$ . Compared to traditional 3D encoders and occupancy prediction heads, this design significantly improves speed while maintaining comparable performance.", + "bbox": [ + 75, + 301, + 467, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 571, + 207, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on the large-scale benchmark dataset Occ3D-nuScenes to validate the efficacy of our proposed methods. Additionally, we conduct ablation experiments to verify the effectiveness of each component in our method.", + "bbox": [ + 75, + 595, + 467, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 681, + 178, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Occ3D-nuScenes [30] is a large-scale autonomous driving dataset, which includes 1,000 urban traffic scenes under various conditions, the data is split into 700 training, 150 validation, and 150 testing scenes. The occupancy grid is defined within a range of $-40\\mathrm{m}$ to $40\\mathrm{m}$ along the X and Y axes and $-1\\mathrm{m}$ to $5.4\\mathrm{m}$ along the Z axis. The voxel size for occupancy labeling is set to $0.4\\mathrm{m} \\times 0.4\\mathrm{m} \\times 0.4\\mathrm{m}$ . The semantic labels include 17 categories consisting of 16 known object classes with an additional 'empty' class. Compared to Occ3D-nuScenes, SurroundOcc [35] is also based on the nuScenes dataset but its prediction range is from $-50\\mathrm{m}$ to $50\\mathrm{m}$ for X and Y axes, and $-5\\mathrm{m}$ to $3\\mathrm{m}$ along the Z axis, with the voxel label size of $0.5\\mathrm{m} \\times 0.5\\mathrm{m} \\times 0.5\\mathrm{m}$ .", + "bbox": [ + 75, + 704, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 90, + 715, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use ResNet-50 as the default image backbone and SPVCNN as the LiDAR backbone. The model is trained on a GeForce RTX 4090 GPU using the AdamW optimizer with a learning rate of 1e-4 and gradient clipping. For semantic and depth-guided visual transformations, the bilinear incremental discretization range and the number of diffusion feature layers are set to 1 m and 8, respectively.", + "bbox": [ + 496, + 119, + 890, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparing with SOTA methods", + "text_level": 1, + "bbox": [ + 498, + 253, + 782, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Occ3D-nuScenes. As shown in Table 1, we report the quantitative comparison of existing state-of-the-art methods for 3D occupancy prediction tasks on Occ3D-nuScenes. Most existing approaches are predominantly based on camera-only algorithms, with relatively few focusing on multisensor fusion. Our method, which employs a compact backbone and a lightweight LiDAR branch, achieves state-of-the-art performance in terms of mIoU and the majority of class-wise IoUs. Additionally, our approach achieves the best inference speed, meeting the real-time requirements of autonomous driving scenarios. The visualization on Occ3D-nuScenes validation set is shown in Fig. 6. Compared to the baseline, our method effectively identifies categories that the baseline fails to correctly predict in both day and night scenarios. More visualizations can be found in supplementary material.", + "bbox": [ + 496, + 282, + 890, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SurroundOcc-nuScenes. Table 2 provides a quantitative comparison on the SurroundOcc validation set, highlighting the performance of our method, SDG-OCC, compared to other approaches. Using both LiDAR and camera inputs, our method achieves SOTA performance on the SurroundOcc validation set, even over larger distances. This success is attributed to our semantic and depth-guided view transformation, which enhances depth estimation accuracy and enables robust occupancy prediction across varying grid sizes and distances. Furthermore, we utilize visual masks generated from [32], achieving performance comparable to OCC3d-nuscene. Notably, our method uses only a lightweight ResNet50 backbone and a lower resolution of $256 \\times 704$ , underscoring its effectiveness and efficiency.", + "bbox": [ + 496, + 532, + 890, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of results within different ranges. We further evaluate different ranges surrounding the car to provide a comprehensive analysis. Fig. 7 clearly illustrates our mIoU and iou relative to the baseline FlashOcc. Short-range understanding is critical due to the limited reaction time for autonomous vehicles. Our method significantly outperforms the baseline in both mIoU and IoU. In long-range areas, where LiDAR data is sparse and few pixels define the depth of large regions, our approach still achieves superior IoU performance.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6756", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/50c45230dee196d1473535f2ae108df948735bcee0b698bbed7316c4f0c7dc38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputBackboneVisible MaskmIoUothersbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetationTime(ms)
■■■■■■■■■■■■■■■■■
TPVFormer [9]CR-50✓34.27.6844.0117.6640.8846.9815.0620.5424.6924.6624.2629.2879.2740.6548.4949.4432.6329.82289.85
SurroundOcc [35]CR-101✓37.18.9746.3317.0846.5452.0120.0521.4723.5218.6731.5137.5681.9141.6450.7653.9342.9137.16303.03
OccFormer [44]CR-50✓37.49.1545.8418.2042.8050.2724.0020.8022.8620.9831.9438.1380.1338.2450.8354.346.4140.15-
VoxFormer [13]CR-101✓40.7------------------
FBOcc [16]CR-50✓42.114.3049.7130.046.6251.5429.329.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58-
PanoOcc [33]CR-101-42.1311.6750.4829.6449.4455.5223.2933.2630.5530.9934.4342.5783.3144.2354.4056.0445.9440.40322.58
FastOcc [6]CR-101✓40.7512.8646.5829.9346.0754.0923.7431.1030.6828.5233.0839.6983.3344.6553.9055.4642.6136.50221.2
BEVDet4D [8]CSwin-B✓42.512.3750.1526.9751.8654.6528.3828.9629.0228.2837.0542.5282.5543.1554.8758.3348.7843.791000.0
FlashOcc [42]CSwin-B✓43.5213.3151.6228.0750.9155.6927.4631.0529.9829.2038.8643.6883.8745.6356.3359.0150.6344.56909.1
COTR [21]CSwin-B✓46.214.8553.2535.1950.8357.2535.3634.0633.5437.1438.9944.9784.4648.7357.6061.0851.6146.72840.34
HyDRa [36]C+RR-50-44.40---52.356.3-35.935.10--44.1-------
OCCFusion [23]C+LR-101-46.7911.6547.8132.0757.2757.5131.8040.1147.3533.7445.8150.3578.7937.1744.3653.3663.1863.20-
RadOcc-LC [43]C+LSwin-B✓49.3810.9358.2325.0157.8962.8534.0433.4550.0732.0548.8752.1182.9042.7355.2758.3468.6466.013333
SDG-KLC+LR-50✓50.1612.2657.1223.6958.7762.7434.5536.1950.132.0549.8951.2484.146.0557.261.4569.5665.7883
SDG-FusionC+LR-50✓51.6613.2157.7724.360.3364.2836.2139.4452.3635.8050.9153.6584.5647.4558.0061.6170.6767.65133
", + "bbox": [ + 83, + 90, + 892, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2b4e89ec1c3f1e57525ebddbb970582da549092f67bd4f2ca49336cccabdfae.jpg", + "table_caption": [ + "Table 1. 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. We present the mean IoU over categories and the IoUs for different classes. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. In the backbone, R represents ResNet, while Swin stands for Swin Transformer." + ], + "table_footnote": [], + "table_body": "
MethodInputBackbonemIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. suf.other flatsidewalkterrainmanmadevegetation
SurroundOcc [35]CR-10120.320.511.628.130.810.715.114.012.014.322.237.223.724.422.714.821.8
OccFormer [44]CR-10120.121.111.328.230.310.615.714.411.214.022.637.322.424.923.515.221.1
C-CONet [32]CR-10118.418.610.026.427.48.615.713.39.710.920.233.020.721.421.814.721.3
FB-Occ [16]CR-10119.620.611.326.929.810.413.613.711.411.520.638.221.524.622.714.821.6
RenderOcc [43]CR-10119.019.711.228.128.29.814.711.811.913.120.133.221.322.622.315.320.9
L-CONet [32]L-17.719.24.015.126.96.23.86.86.014.113.139.719.124.023.925.135.7
FlashOcc* [42]CR-5044.144.211.054.160.526.122.631.315.338.847.180.542.048.253.760.870.0
M-CONet [32]C+LR-10124.724.813.031.634.814.618.020.014.720.026.639.222.826.126.026.037.1
Co-Occ [24]C+LR-10127.128.116.134.037.217.021.620.815.921.928.742.325.429.128.628.238.0
OccFusion [23]C+L+RR-10127.327.119.633.736.221.724.825.316.321.830.039.519.924.926.528.940.4
DAOcc [40]C+LR-5030.530.819.534.038.825.027.729.922.523.231.641.025.929.429.935.243.5
SDG-FusionC+LR-5031.731.215.639.742.422.226.929.722.724.233.245.925.131.833.538.944.5
SDG-KL*C+LR-5050.451.420.959.665.834.133.149.023.445.153.776.544.053.558.265.272.9
SDG-Fusion*C+LR-5052.254.316.061.466.933.233.549.222.247.854.683.949.455.860.270.276.5
", + "bbox": [ + 84, + 364, + 890, + 626 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b514b4163d289cdc07a7d9b13f5a0ddf33716985cf50a739b8b98ea453b779cd.jpg", + "table_caption": [ + "Table 2. 3D Occupancy prediction performance on the SurroundOcc-nuScenes validation set. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. * means the performance is achieved through a visible mask similar to [30]." + ], + "table_footnote": [], + "table_body": "
BaselineSDGFOADiou(%)mIoU(%)
✓90.2737.84
✓✓94.6248.51
✓✓94.7644.92
✓✓✓95.3551.66
", + "bbox": [ + 109, + 686, + 436, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 827, + 227, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Effectiveness of Each Component. The results are shown in Table 3, we can observe that all components make their own performance contributions. The baseline achieves", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/515ade9e67a1ba6a6ee7591981a8b24c062276275c36c357ad33b24c50357db9.jpg", + "table_caption": [ + "Table 3. Ablation study on Occ3D-nuScenes dataset with SDG-Fusion. SDG: Semantic and Depth-Guided View Transformations. FOAD: Fusion-To-Occupancy-Driven Active Distillation." + ], + "table_footnote": [], + "table_body": "
rlIoU(%)mIoU (%)
1495.3451.15
1895.3551.66
11295.3851.28
2495.3550.4
2895.3951.03
21295.3051.12
", + "bbox": [ + 517, + 686, + 870, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of the hyperparameter used in SDG view transformation module on Occ3D-nuScenes.", + "bbox": [ + 500, + 801, + 890, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "90.27% of IoU and 37.84% of mIoU. We first integrated the Semantic and Depth-Guided (SDG) View Transformation into the baseline model, which brings 4.35% and 10.67% performance gain in IoU and mIoU. Fusion enhanced by in", + "bbox": [ + 500, + 839, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6757", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77149bf92e1d5104b39b390e8b152bee8a8780182e72f8b3f2238670780b6ec2.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 344, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d5cd69217362d6897d2ce6468fcb0c8cdef1112113a0e868a9632dcd15b4b97.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [], + "bbox": [ + 344, + 90, + 514, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bb46181bc752c2fe4ae54fec6549e27aeb0b8fda5bda0a95b46b14e0d4031ea.jpg", + "image_caption": [ + "Our" + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 683, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be27ec132b49c11470807cb630ea53f89d627881e63346cbe57617ab7d0fa0b2.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 684, + 90, + 854, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1798e593382ecc0aa9af069983464776dec4d2aaf94ceb94321970787a8d64aa.jpg", + "image_caption": [ + "Figure 6. Qualitative results of SDG on the validation set of Occ3D-nuScenes. Each pair of rows displays results from day and low-light scene, respectively. Within each row, images from left to right represent the input images, baseline, our results, and the ground truth.", + "Figure 7. Distance-based evaluation on Occ3D-nuScenes. As the distance increases, the point cloud becomes sparse." + ], + "image_footnote": [], + "bbox": [ + 102, + 444, + 437, + 573 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/56ac18d9c12a811c0a9ea3ae09ea7aa92c11f6b55b90e4742dc5d5308b38a110.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
k3579
IoU (%)95.3295.4095.3595.26
mIoU (%)51.2251.4051.6651.08
", + "bbox": [ + 80, + 601, + 460, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study of the hyperparameter used in feature fusion of FOAD on Occ3D-nuScenes. The K denotes the size of the pixel region corresponding to the neighborhood feature extraction.", + "bbox": [ + 75, + 656, + 467, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tegrating additional LiDAR information, the IoU and mIoU have been significantly improved by $4.49\\%$ and $7.08\\%$ , respectively. By using both SDG and Fusion, outperforming the baseline by $5.19\\%$ of IoU and $13.82\\%$ of mIoU.", + "bbox": [ + 75, + 703, + 467, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effectiveness of SDG View Transformation. To further demonstrate the effect of SDG View Transformation, we conducted hyperparameter analysis experiments. In the SDG view transformation, the range $r$ of bilinear growth discretization and the diffusion feature layers $l$ control the virtual point generation of SDG. As shown in Table 4, lower depth precision (e.g., $r = 2$ and $l = 4$ ) results in slightly reduced performance compared to other configurations. However, excessive depth precision does not lead to", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "additional gains, with the optimal performance observed at $r = 1$ and $l = 8$ .", + "bbox": [ + 498, + 445, + 890, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effectiveness of FOAD Module. We perform a hyperparameter analysis of the FOAD module. For neighborhood feature fusion, the parameter $K$ controls the fusion of features from neighboring pixels. As shown in Table 5, increasing $K$ does not consistently improve performance, with optimal results achieved at $K = 7$ .", + "bbox": [ + 496, + 479, + 890, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 598, + 617, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, designed to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. To address the inaccurate depth estimation in view transformations, we propose a semantic and depth-guided view transformation method. This approach integrates pixel semantics and corresponding point depth through diffusion and bilinear discretization, effectively reducing invalid image features and significantly enhancing the speed and accuracy of semantic occupancy. Meanwhile, We propose a fusion-to-occupancy-driven active distillation that incorporates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Our method achieves the SOTA performance with real-time processing on the Occ3D-nuScenes dataset and comparable performance on the more challenging SurroundOccnuScenes dataset, demonstrating its effectiveness.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6758", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Geonho Bang, Kwangjin Choi, Jisong Kim, Dongsuk Kum, and Jun Won Choi. Radardistill: Boosting radar-based object detection performance via knowledge distillation from lidar features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15491-15500, 2024. 5", + "[2] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2", + "[3] Jiahui Fu, Chen Gao, Zitian Wang, Lirong Yang, Xiaofei Wang, Beipeng Mu, and Si Liu. Eliminating cross-modal conflicts in bev space for lidar-camera 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16381-16387, 2024. 5", + "[4] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6185-6194, 2023. 5", + "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[6] Jiawei Hou, Xiaoyan Li, Wenhao Guan, Gang Zhang, Di Feng, Yuheng Du, Xiangyang Xue, and Jian Pu. Fastocc: Accelerating 3d occupancy prediction by fusing the 2d bird's-eye view and perspective view. arXiv preprint arXiv:2403.02710, 2024. 3, 7", + "[7] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7", + "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 7", + "[10] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 1", + "[11] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023. 1", + "[12] Jiale Li, Hang Dai, Hao Han, and Yong Ding. Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21694-21704, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 2, 7", + "[14] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1", + "[15] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 3", + "[16] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7", + "[17] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 1", + "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3", + "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3", + "[20] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1", + "[21] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 3, 7", + "[22] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: A straightforward and effective multi-sensor fusion framework for 3d occupancy prediction. arXiv preprint arXiv:2403.01644, 2024. 3", + "[23] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction. IEEE Transactions on Intelligent Vehicles, 2024. 7", + "[24] Jingyi Pan, Zipeng Wang, and Lin Wang. Co-occ: Coupling explicit feature fusion with volume rendering regularization for multi-modal 3d semantic occupancy prediction. IEEE Robotics and Automation Letters, 2024. 3, 7" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6759", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoji Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 3", + "[26] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghang Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 3", + "[27] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1", + "[28] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 3", + "[29] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 5", + "[30] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 7", + "[31] Guoqing Wang, Zhongdao Wang, Pin Tang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Occgen: Generative multi-modal 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2404.15014, 2024. 3", + "[32] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 3, 6, 7", + "[33] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17158-17168, 2024. 7", + "[34] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3", + "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2, 6, 7", + "[36] Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Anouar Laouichi, Martin Hofmann, and Gerhard Rigoll. Unleashing hydra: Hybrid fusion, depth consistency and radar for unified 3d perception. arXiv preprint arXiv:2403.07746, 2024. 3, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1", + "[38] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 1", + "[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 1", + "[40] Zhen Yang, Yanpeng Dong, and Heng Wang. Daocc: 3d object detection assisted multi-sensor fusion for 3d occupancy prediction. arXiv preprint arXiv:2409.19972, 2024. 7", + "[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 1", + "[42] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 2, 7", + "[43] Haiming Zhang, Xu Yan, Dongfeng Bai, Jiantao Gao, Pan Wang, Bingbing Liu, Shuguang Cui, and Zhen Li. Radocc: Learning cross-modality occupancy knowledge through rendering assisted distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7060-7068, 2024. 7", + "[44] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 7", + "[45] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9939-9948, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 737 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6760", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_model.json b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a28ffd59af3f067f303cf569ac08110854d4d2fb --- /dev/null +++ b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.13, + 0.89, + 0.178 + ], + "angle": 0, + "content": "SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.204, + 0.885, + 0.242 + ], + "angle": 0, + "content": "ZaiPeng Duan ChenXu Dang Xuzhong Hu Pei An Junfeng Ding Jie Zhan YunBiao Xu *Jie Ma Huazhong University of Science and Technology" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.291 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.306, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Multimodal 3D occupancy prediction has garnered significant attention for its potential in autonomous driving. However, most existing approaches are single-modality: camera-based methods lack depth information, while LiDAR-based methods struggle with occlusions. Current lightweight methods primarily rely on the Lift-Splat-Shoot (LSS) pipeline, which suffers from inaccurate depth estimation and fails to fully exploit the geometric and semantic information of 3D LiDAR points. Therefore, we propose a novel multimodal occupancy prediction network called SDG-OCC, which incorporates a joint semantic and depth-guided view transformation coupled with a fusion-to-occupancy-driven active distillation. The enhanced view transformation constructs accurate depth distributions by integrating pixel semantics and co-point depth through diffusion and bilinear discretization. The fusion-to-occupancy-driven active distillation extracts rich semantic information from multimodal data and selectively transfers knowledge to image features based on LiDAR-identified regions. Finally, for optimal performance, we introduce SDG-Fusion, which uses fusion alone, and SDG-KL, which integrates both fusion and distillation for faster inference. Our method achieves state-of-the-art (SOTA) performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes dataset, demonstrating its effectiveness and robustness. The code will be released at https://github.com/DzpLab/SDGOCC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.471, + 0.878 + ], + "angle": 0, + "content": "Accurate 3D perception of the surrounding environment forms the cornerstone of modern autonomous driving systems and robotics, ensuring efficient planning and safe control [7, 10]. In recent years, advancements in 3D object detection [14, 17, 20, 38, 41] and semantic segmentation [11, 12, 37, 39, 45] have significantly propelled the field of" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.274, + 0.891, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.464, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Figure 1. Comparisons of the mIoU and inference speed (FPS) of various 3D occupancy prediction methods on the Occ3D-nuScenes validation set. SDG-OCC achieves higher accuracy and real-time inference speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.703 + ], + "angle": 0, + "content": "3D perception. However, object detection relies on strict bounding boxes, making it difficult to recognize arbitrary shapes or unknown objects, while semantic segmentation struggles with fine-grained classification in complex scenes, especially under occlusion and overlap. In this context, 3D semantic occupancy prediction [30, 32] offers a more comprehensive approach to environment modeling. It simultaneously estimates the geometric structure and semantic categories of scene voxels, assigns labels to each 3D voxel, and provides a more complete perception, showing stronger robustness to arbitrary shapes and dynamic occlusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Leveraging the complementary strengths of LiDAR and camera data is crucial for various 3D perception tasks. However, due to the heterogeneity between modalities, fusing LiDAR and camera data for 3D occupancy prediction remains challenging. Specifically, cameras provide rich semantic information but lack precise depth details, while LiDAR offers accurate depth information but only captures sparse data, potentially missing comprehensive scene details such as occluded objects. Existing methods often suffer from significant computational burdens (see Fig. 1), with some approaches attempting to leverage the LSS [27] pipeline for real-time performance. Although LSS simulates the uncertainty of each pixel's depth through depth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.329, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author: majie@hust.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "6751" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.271, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.228, + 0.264, + 0.239 + ], + "angle": 0, + "content": "(a) Initial BEV features of LSS" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.455, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.228, + 0.441, + 0.239 + ], + "angle": 0, + "content": "(b) Our initial BEV features" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.24, + 0.269, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.379, + 0.245, + 0.391 + ], + "angle": 0, + "content": "(c) Ground truth of BEV" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.24, + 0.453, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.379, + 0.438, + 0.392 + ], + "angle": 0, + "content": "(d) Occupancy grid of BEV" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.499 + ], + "angle": 0, + "content": "Figure 2. (a) BEV feature map of LSS with a shape of \\(200 \\times 200\\). We can observe that LSS has an extremely low utilization rate for BEV space. (b) The corresponding BEV features in SDG-OCC. Using depth and semantic information, the 2D-to-3D view transformation achieves efficient occupancy and utilization of the BEV features. (c) The corresponding BEV features in Ground Truth. (d) The corresponding BEV features in the occupancy grid." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.671 + ], + "angle": 0, + "content": "distribution (with depth intervals typically set to \\(0.5\\mathrm{m}\\)), its sparse BEV representation allows only \\(50\\%\\) of the grids to receive valid image features [16] (see Fig. 2 (a)). While increasing the depth interval can improve depth estimation accuracy to mitigate sparsity, it significantly increases computational demands. Additionally, while LiDAR can provide valuable geometric priors, fusion-based methods that process both point clouds and images simultaneously impose heavy computational burdens, thereby increasing the strain on real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To address these issues, we propose a multimodal 3D semantic occupancy prediction framework, named SDG-OCC, which aims to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. In this framework, we introduce a semantic and depth-guided view transformation to replace normal BEV feature generation. Specifically, after extracting features from camera data and obtaining semantic segmentation masks and depth distributions through a multi-task head, we use the semantic masks and depth maps provided by LiDAR to construct virtual points via local diffusion and bilinear discretization. Combined with the depth distribution, these points are then projected into the BEV space. The comparison between LSS and our generated BEV features is shown in Fig. 2. The SDG view transformation sig" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "nificantly refines depth estimation accuracy and reduces redundant virtual point seeds, improving both the speed and accuracy of semantic occupancy." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Secondly, we introduce a fusion-to-occupancy-driven active distillation module. We first fuse LiDAR and camera features in the BEV space and then unidirectionally selectively transferred multimodal knowledge to image features based on LiDAR-identified regions. Our proposed SDG-Fusion, which includes only fusion, achieved SOTA performance on the Occ3D-nuscene [30] and SurroundOccnuScenes [35] validation dataset. In comparison, SDGKL, which combines fusion and unidirectional distillation, achieves real-time speed with a slight performance penalty." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.29, + 0.843, + 0.304 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.307, + 0.891, + 0.367 + ], + "angle": 0, + "content": "- We introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, aimed at achieving higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.457 + ], + "angle": 0, + "content": "- We propose a novel view transformation method that leverages the geometric and semantic information of point clouds to guide the 2D-3D view transformation. This significantly enhances the accuracy of depth estimation and improves both the speed and accuracy of semantic occupancy." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.457, + 0.891, + 0.546 + ], + "angle": 0, + "content": "- We propose a fusion-to-occupancy-driven active distillation module that integrates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Building on this, we present SDG-Fusion for high performance and SDG-KL for faster inference." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.548, + 0.891, + 0.623 + ], + "angle": 0, + "content": "- Our method achieves SOTA performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes validation dataset, demonstrating the effectiveness of our approach." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.307, + 0.892, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.639, + 0.641, + 0.655 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.831, + 0.682 + ], + "angle": 0, + "content": "2.1. Vision-Centric Occupancy Perception" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inspired by Tesla's autonomous driving perception system, vision-centric occupancy perception has garnered significant attention from both industry and academia. MonoScene [2] is a pioneering work that used only RGB inputs. TPVFormer [9] combines surround multi-camera inputs and uses transformer-based methods to lift features into a tri-perspective view space. SurroundOcc [35] extends high-dimensional BEV features into occupancy features and directly performs spatial cross-attention to generate geometric information. VoxFormer [13] introduces a two-stage transformer-based semantic scene completion framework, capable of outputting complete 3D volumetric semantics from 2D images alone. FlashOcc [42] transforms the channel to height, lifting BEV output to 3D space," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6752" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "significantly improving operational efficiency. FBOcc [15] proposes a front-to-back view transformation module based on BEV features to address the limitations of different view transformations. Methods like UniOcc [26] and RenderOcc [25] use NeRF [34] to directly predict 3D semantic occupancy, but the rendering speed limits their efficiency. FastOcc [6] improves the occupancy prediction head to achieve a faster inference speed. COTR [21] builds compact 3D occupancy representations through explicit-implicit view transformation and coarse-to-fine semantic grouping. In this paper, we improve the speed and accuracy of 3D semantic occupancy prediction from the BEV space by incorporating the geometric and semantic information of the point cloud into the view transformation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.314, + 0.391, + 0.331 + ], + "angle": 0, + "content": "2.2. Multi-Modal Occupancy Perception" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.473, + 0.686 + ], + "angle": 0, + "content": "Multimodal occupancy perception leverages the strengths of multiple modalities to overcome the limitations of unimodal perception. OpenOccupancy[32] introduced a benchmark for LiDAR-camera semantic occupancy prediction. Inspired by BEVFusion, OccFusion [22] concatenates 3D feature volumes from different modalities along the feature channels, followed by convolutional layers to combine them. CO-Occ [24] introduced the Geometric and Semantic Fusion (GSFusion) module, identifying voxels containing both point cloud and visual information using k-nearest neighbors (KNN) search. OccGen [31] employs an adaptive fusion module to dynamically integrate occupancy representations from camera and LiDAR branches, using 3D convolutions to determine fusion weights for aggregating LiDAR and camera features. HyDR [36] proposes to integrate multimodal information in both perspective view (PV) and bird's-eye view (BEV) representation spaces. In this paper, we enhance view transformation by incorporating semantic segmentation masks and LiDAR depth maps to achieve higher occupancy accuracy. Additionally, we fuse BEV features from multimodal data and unidirectionally distill them into camera features, improving the accuracy and inference speed of 3D semantic occupancy prediction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.212, + 0.718 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.726, + 0.208, + 0.743 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given joint input from multi-view images and LiDAR data, 3D occupancy prediction aims to estimate the occupancy state and semantic classification of 3D voxels surrounding the ego vehicle. Specifically, the input consists of a \\(T\\)-frame consequent sequence of images \\(X_{C} \\in \\mathbb{R}^{N_{C} \\times H_{C} \\times W_{C} \\times 3}\\) from \\(N_{C}\\) surround-view cameras and point clouds \\(X_{L} \\in \\mathbb{R}^{N_{L} \\times (3 + d)}\\) as multimodal input, represented as \\(X = \\{X_{C}, X_{L}\\}\\). Here \\(H_{C}\\), \\(W_{C}\\) represent the height and width of the image, respectively, \\(N_{L}\\) denotes the number of point clouds and \\(d\\) denotes the initial addi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "tional features of the point cloud. Subsequently, we train a neural network to generate an occupancy voxel map \\( Y \\in \\mathbb{R}^{H \\times W \\times D \\times C_N} \\), where each voxel is assigned a label as unknown, occupied, or a semantic category from \\( \\{C_0 \\) to \\( C_N\\} \\). Here, \\( N \\) denotes the total number of categories of interest, and \\( H, W, D \\) represent the volume dimensions of the entire scene." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.207, + 0.697, + 0.223 + ], + "angle": 0, + "content": "3.2. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.352 + ], + "angle": 0, + "content": "An overview of SDGOCC is shown in Fig. 3. It mainly consists of four key modules: image feature encoder to extract image features, semantic and depth-guided view transformation to construct 2D-3D feature transformation, fusion-to-occupancy-driven active distillation for fusing multimodal features and selectively transferring knowledge to the image features, and the occupancy prediction head for final output." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.361, + 0.655, + 0.377 + ], + "angle": 0, + "content": "3.3. Image Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.537 + ], + "angle": 0, + "content": "The image feature encoder aims to capture multi-view features, providing a foundation for 2D-3D view transformation. Given RGB images from surround-view cameras, we first use a pre-trained image backbone network, such as classic ResNet [5] or strong Swin-Transformer [19], to extract multi-layer image features \\( F_{C} \\in \\mathbb{R}^{N_{C} \\times C \\times H \\times W} \\). These features are then aggregated using a feature pyramid network (FPN) [18], which combines fine-grained features and coarse-grained features and down-sampling them to a specific scale, typically 1/16." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.742, + 0.56 + ], + "angle": 0, + "content": "3.4. SDG View Transformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.84 + ], + "angle": 0, + "content": "The LSS pipeline is widely used for converting image features to BEV representations in 3D perception. It constructs virtual points based on a predefined depth range for each pixel and predicts the depth distribution weight \\(\\alpha\\) and context feature \\(c\\). The feature representation at depth \\(d\\) is given by \\(p_d = \\alpha_d c\\). All virtual points then are projected into BEV space, where features at each height \\(Z\\) are aggregated to form BEV features. Although LSS handles depth uncertainties and ambiguities by modeling pixel depth using depth distributions, the number of per-pixel features remains large even with a 0.5-meter depth interval, an order of magnitude larger than point features. Meanwhile, BEV features are highly sparse, with less than \\(50\\%\\) of the image features being effective, leading to suboptimal occupancy prediction performance. Reducing the depth interval improves accuracy but significantly increases computational burden and introduces irrelevant features, as most of the occupancy grid remains empty." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To address this, we propose a novel view transformation method that leverages the sparse depth information from LiDAR as a prior, diffuses within the same semantic class, followed by linear-increasing [28] and linear-decreasing dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.885, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Figure 3. The overall architecture of SDG-OCC. \\(T\\)-frame multi-view images and corresponding point clouds are fed into the image and LiDAR backbones to extract features. Image features are processed by a multi-task head to generate semantic masks and depth distributions, combined with LiDAR depth maps to create virtual points for image BEV features. These features are fused with point cloud BEV features and selectively transfer multimodal knowledge to image features based on LiDAR-identified regions. Finally, enhanced features are processed by the occupancy prediction head to generate the occupancy map." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.607, + 0.471, + 0.732 + ], + "angle": 0, + "content": "Figure 4. The overall architecture of the SDG view transformation. The image features processed by DepthNet are divided into texture features, depth features, and semantic features. Semantic features generate masks via a segmentation head, which are combined with depth maps from LiDAR for local diffusion and bilinear discretization to create virtual points. These points and their features receive pillar features from texture and depth features via the outer product, and then through BEV pooling to generate the final image BEV features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "cretization to generate high-precision virtual point seeds around each co-visible point, as shown in Fig. 4. First, we extract features from multi-camera images and generate semantic segmentation masks via a multi-task head, while simultaneously extracting image textual features and depth distribution weights, with the depth head and semantic head supplementing cross-task information through gated attention. To better utilize semantic information, we select 4x downsampled features for view transformation, as higher downsampling increases the semantic and depth ambiguity" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.45, + 0.566, + 0.463 + ], + "angle": 0, + "content": "of pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Given the differences in sparsity between images and point clouds, we combine image semantic segmentation masks and sparse projected depth maps provided by LiDAR to diffuse depth values within the same semantic category masks, generating a semi-dense extended depth map. This process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.631 + ], + "angle": 0, + "content": "\\[\nD _ {\\text {t e m p}} (i, j) = \\frac {\\sum_ {(p , q) \\in N (i , j)} D (p , q) \\cdot \\mathbb {I} [ M (p , q) = M (i , j) ]}{\\sum_ {(p , q) \\in N (i , j)} \\mathbb {I} [ M (p , q) = M (i , j) ]}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.707 + ], + "angle": 0, + "content": "where \\(N(i,j)\\) represents the circle area with radius \\(r\\) around the current point, and \\(M(i,j)\\) denotes the segmentation mask with \\(\\mathbf{N}\\) category. And \\(\\mathbb{I}[M(p,q) = M(i,j)]\\) checks if the semantic label at \\((\\mathrm{p},\\mathrm{q})\\) matches that at \\((\\mathrm{i},\\mathrm{j})\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.713, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathbb {I} [ M (p, q) = M (i, j) ] = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} M (p, q) = M (i, j) \\neq 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.891, + 0.796 + ], + "angle": 0, + "content": "And the final extended depth map \\( D_{(i,j)} \\) replaces the original co-points of the \\( D_{\\mathrm{temp}}(i,j) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Due to the projection deviation from 2D pixels to 3D points, we apply bidirectional linear incremental discretization to the extended depth map to obtain discrete virtual points, enhancing the accuracy of depth estimation. These steps reduce the number of virtual points, thereby improving inference speed. Finally, the image textual features \\(F_{t}\\) and depth distribution weights \\(D_{w}\\) are calculated by the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6754" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.473, + 0.368 + ], + "angle": 0, + "content": "Figure 5. (a) The pipeline of SDG-Fusion. (b) The pipeline of SDG-KL. (c) The overall architecture of fusion-to-occupancy-driven active distillation. The source feature extracts neighborhood information from the corresponding pixels of the cross feature, where the source serves as the query, and the cross feature acts as both key and value for feature interaction. This interaction is dynamically refined through a gated attention mechanism to produce the fused feature. Subsequently, an occupancy-driven active distillation is used to unidirectionally integrate multimodal information into the image feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.471, + 0.471 + ], + "angle": 0, + "content": "outer product \\( F_{t} \\otimes D_{w} \\) to derive features for each virtual point and generate the BEV features \\( F_{bev}^{C} \\) of the camera through BEV pooling. This method effectively integrates semantic information with sparse depth data, significantly enhancing the accuracy of pixel depth estimation and improving the speed of view transformations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.47, + 0.499 + ], + "angle": 0, + "content": "3.5. Fusion-to-occupancy-driven Active Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.644 + ], + "angle": 0, + "content": "The LiDAR branch encompasses point cloud feature extraction, multimodal fusion, and occupancy-driven active distillation (as shown in Fig. 5). Initially, the point cloud data undergoes voxelization and normalization to generate the initial features. We choose SPVCNN as our point-voxel feature encoder due to its efficiency in representing sparse point clouds while effectively preserving fine-grained details. Subsequently, we compress the voxel features at the corresponding scale to generate the BEV features \\( F_{\\text{bev}}^{L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.794 + ], + "angle": 0, + "content": "The complementary information from LiDAR and cameras is critical for 3D perception. A naive fusion method typically concatenates LiDAR and image BEV features along the channel dimension to enhance performance. However, feature misalignment due to extrinsic conflicts [3] limits the effectiveness of the fusion. Therefore, we propose a dynamic neighborhood feature fusion module. This module unidirectionally extracts neighborhood features from cross features and dynamically adjusts their weights into the source features using a gating attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "External projection deviations during the BEV feature construction process for LiDAR and images result in misalignment between LiDAR and camera BEV features [3, 29]. Therefore, we adopt neighborhood attention from [4] to extract local patch features corresponding to the pixel from the cross features, and dynamically adjust the weights through gated attention to selectively enhance the fused fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.214 + ], + "angle": 0, + "content": "ture representation. Specifically, the image features \\(F_{bev}^{C}\\) as the source feature, are represented as a feature vector sequence \\(F_{img} \\in \\mathbb{R}^{n \\times m}\\), which is projected through a linear layer to obtain the query features \\(Q_{s} \\in \\mathbb{R}^{n \\times q}\\). Similarly, the LiDAR features \\(F_{bev}^{L}\\) as cross feature are projected to obtain the key \\(K_{c} \\in \\mathbb{R}^{n \\times q}\\) and value \\(V_{c} \\in \\mathbb{R}^{n \\times v}\\) features. The local neighborhood features \\(F_{\\mathrm{neighbor}}\\) for a query point \\(i\\) are computed by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.222, + 0.892, + 0.263 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {n e i g h b o r}} = \\sigma \\left(\\frac {Q _ {s} ^ {i} \\cdot \\left(K _ {c} ^ {n (i)}\\right) ^ {T} + B (i , n (i))}{\\sqrt {v}}\\right) \\cdot V _ {c} ^ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.895, + 0.378 + ], + "angle": 0, + "content": "where \\( n(i) \\) represent the neighborhood with the size of \\( k \\) centered at the same position in the cross feature, \\( B(i,\\rho (i)) \\) denotes the relative positional biases and \\( \\sigma \\) denotes to Softmax. For each pixel in the feature map, we calculate the local neighborhood features. The fused features \\( F_{\\text{bev}}^{\\text{fuse}} \\) are then obtained from the local neighborhood features through gated attention, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.386, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\nF _ {b e v} ^ {f u s e} = \\left(\\sigma \\left(\\operatorname {C o n v} \\left(f _ {\\text {A v g}} \\left(F _ {\\text {n e i g h b o r}}\\right)\\right)\\right) \\cdot F _ {\\text {n e i g h b o r}}, \\quad (4) \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.492 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the sigmoid function and Conv denotes linear transform matrix (e.g., 1x1 convolution), \\(f_{\\mathrm{Avg}}\\) refers Adaptive Average Pooling. The fused features \\(F_{bev}^{fuse}\\) are processed by the occupancy prediction head to obtain the SDG-Fusion model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Additionally, to ensure real-time, we propose an occupancy-driven active distillation, where fused features are unidirectionally transferred to the image features. Specifically, LiDAR features are used as the source feature, while image features serve as the cross feature, resulting in LiDAR-dominant fused features. Inspired by [1], we then divide the space into two regions: the active region(AR), where both LiDAR and image features are occupied, and the inactive region(IR), where only LiDAR features are occupied. The details are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.654, + 0.892, + 0.672 + ], + "angle": 0, + "content": "\\[\nA R = \\left(M _ {\\text {f u s e d}, i, j} = 1\\right) \\wedge \\left(M _ {\\text {i m g}, i, j} = 1\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.68, + 0.892, + 0.698 + ], + "angle": 0, + "content": "\\[\nI R = \\left(M _ {\\text {f u s e d}, i, j} = 1\\right) \\wedge \\left(M _ {\\text {i m g}, i, j} = 0\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.793 + ], + "angle": 0, + "content": "where a value of 1 in \\( M_{\\mathrm{mode},i,j} \\) indicates that the coordinate is occupied by the respective modality. Typically, the AR region is significantly larger than the IR region. To prevent the model from overemphasizing knowledge distillation in the AR region, we apply adaptive scaling based on the relative sizes of the AR and IR regions, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.803, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\nW _ {I, i, j} ^ {\\left(l _ {n}\\right)} = \\left\\{ \\begin{array}{l l} \\alpha , & \\text {i f} (i, j) \\in A R, \\\\ \\rho \\times \\beta , & \\text {i f} (i, j) \\in I R, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\rho = \\frac{N_{AR}}{N_{IR}}\\) represents the relative importance of the \\(IR\\) over \\(AR\\), \\(\\alpha\\) and \\(\\beta\\) are the intrinsic balancing parameters," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "and \\(N_{AR}\\) and \\(N_{IR}\\) are the number of pixels in \\(AR\\) and \\(IR\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.122, + 0.468, + 0.151 + ], + "angle": 0, + "content": "The distillation loss between BEV feature from teacher \\( F^t \\) and student \\( F^s \\) are:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.159, + 0.468, + 0.202 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i s t i l l}} = \\sum_ {c = 1} ^ {C} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} W _ {i, j} \\left(F _ {b e v} ^ {f u s e} - F _ {b e v} ^ {C}\\right) ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.21, + 0.468, + 0.269 + ], + "angle": 0, + "content": "The network is trained with the sum of distillation and classification loss. The image features \\( F_{bev}^{C} \\) are processed by the occupancy prediction head to obtain the SDG-KL model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.28, + 0.284, + 0.295 + ], + "angle": 0, + "content": "3.6. Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.468, + 0.558 + ], + "angle": 0, + "content": "To obtain 3D prediction outputs from coarse BEV features generated by view transformation, we propose an occupancy prediction system consisting of a BEV feature encoder and an occupancy prediction head. The BEV encoder uses several residual blocks for multi-scale feature diffusion and integrates a feature pyramid to acquire BEV features at the target scale. The occupancy prediction head extracts global features with multiple \\(3 \\times 3\\) convolutional layers and includes a channel-to-height transformation module. This module reshapes the BEV features from a \\(F_{out} \\in \\mathbb{R}^{B \\times C \\times H \\times W}\\) to \\(F_{final} \\in \\mathbb{R}^{B \\times C_N \\times D \\times H \\times W}\\), where \\(B, C, W, H\\), and \\(D\\) represent the batch size, channel number, class number, and the dimensions of the 3D space, respectively, with \\(C = C_N \\times D\\). Compared to traditional 3D encoders and occupancy prediction heads, this design significantly improves speed while maintaining comparable performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.209, + 0.589 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.468, + 0.673 + ], + "angle": 0, + "content": "We conduct experiments on the large-scale benchmark dataset Occ3D-nuScenes to validate the efficacy of our proposed methods. Additionally, we conduct ablation experiments to verify the effectiveness of each component in our method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.179, + 0.696 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Occ3D-nuScenes [30] is a large-scale autonomous driving dataset, which includes 1,000 urban traffic scenes under various conditions, the data is split into 700 training, 150 validation, and 150 testing scenes. The occupancy grid is defined within a range of \\(-40\\mathrm{m}\\) to \\(40\\mathrm{m}\\) along the X and Y axes and \\(-1\\mathrm{m}\\) to \\(5.4\\mathrm{m}\\) along the Z axis. The voxel size for occupancy labeling is set to \\(0.4\\mathrm{m} \\times 0.4\\mathrm{m} \\times 0.4\\mathrm{m}\\). The semantic labels include 17 categories consisting of 16 known object classes with an additional 'empty' class. Compared to Occ3D-nuScenes, SurroundOcc [35] is also based on the nuScenes dataset but its prediction range is from \\(-50\\mathrm{m}\\) to \\(50\\mathrm{m}\\) for X and Y axes, and \\(-5\\mathrm{m}\\) to \\(3\\mathrm{m}\\) along the Z axis, with the voxel label size of \\(0.5\\mathrm{m} \\times 0.5\\mathrm{m} \\times 0.5\\mathrm{m}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.717, + 0.107 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.12, + 0.892, + 0.227 + ], + "angle": 0, + "content": "We use ResNet-50 as the default image backbone and SPVCNN as the LiDAR backbone. The model is trained on a GeForce RTX 4090 GPU using the AdamW optimizer with a learning rate of 1e-4 and gradient clipping. For semantic and depth-guided visual transformations, the bilinear incremental discretization range and the number of diffusion feature layers are set to 1 m and 8, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.255, + 0.783, + 0.271 + ], + "angle": 0, + "content": "4.3. Comparing with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Occ3D-nuScenes. As shown in Table 1, we report the quantitative comparison of existing state-of-the-art methods for 3D occupancy prediction tasks on Occ3D-nuScenes. Most existing approaches are predominantly based on camera-only algorithms, with relatively few focusing on multisensor fusion. Our method, which employs a compact backbone and a lightweight LiDAR branch, achieves state-of-the-art performance in terms of mIoU and the majority of class-wise IoUs. Additionally, our approach achieves the best inference speed, meeting the real-time requirements of autonomous driving scenarios. The visualization on Occ3D-nuScenes validation set is shown in Fig. 6. Compared to the baseline, our method effectively identifies categories that the baseline fails to correctly predict in both day and night scenarios. More visualizations can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.744 + ], + "angle": 0, + "content": "SurroundOcc-nuScenes. Table 2 provides a quantitative comparison on the SurroundOcc validation set, highlighting the performance of our method, SDG-OCC, compared to other approaches. Using both LiDAR and camera inputs, our method achieves SOTA performance on the SurroundOcc validation set, even over larger distances. This success is attributed to our semantic and depth-guided view transformation, which enhances depth estimation accuracy and enables robust occupancy prediction across varying grid sizes and distances. Furthermore, we utilize visual masks generated from [32], achieving performance comparable to OCC3d-nuscene. Notably, our method uses only a lightweight ResNet50 backbone and a lower resolution of \\(256 \\times 704\\), underscoring its effectiveness and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Analysis of results within different ranges. We further evaluate different ranges surrounding the car to provide a comprehensive analysis. Fig. 7 clearly illustrates our mIoU and iou relative to the baseline FlashOcc. Short-range understanding is critical due to the limited reaction time for autonomous vehicles. Our method significantly outperforms the baseline in both mIoU and IoU. In long-range areas, where LiDAR data is sparse and few pixels define the depth of large regions, our approach still achieves superior IoU performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6756" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.091, + 0.893, + 0.307 + ], + "angle": 0, + "content": "
MethodInputBackboneVisible MaskmIoUothersbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetationTime(ms)
■■■■■■■■■■■■■■■■■
TPVFormer [9]CR-50✓34.27.6844.0117.6640.8846.9815.0620.5424.6924.6624.2629.2879.2740.6548.4949.4432.6329.82289.85
SurroundOcc [35]CR-101✓37.18.9746.3317.0846.5452.0120.0521.4723.5218.6731.5137.5681.9141.6450.7653.9342.9137.16303.03
OccFormer [44]CR-50✓37.49.1545.8418.2042.8050.2724.0020.8022.8620.9831.9438.1380.1338.2450.8354.346.4140.15-
VoxFormer [13]CR-101✓40.7------------------
FBOcc [16]CR-50✓42.114.3049.7130.046.6251.5429.329.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58-
PanoOcc [33]CR-101-42.1311.6750.4829.6449.4455.5223.2933.2630.5530.9934.4342.5783.3144.2354.4056.0445.9440.40322.58
FastOcc [6]CR-101✓40.7512.8646.5829.9346.0754.0923.7431.1030.6828.5233.0839.6983.3344.6553.9055.4642.6136.50221.2
BEVDet4D [8]CSwin-B✓42.512.3750.1526.9751.8654.6528.3828.9629.0228.2837.0542.5282.5543.1554.8758.3348.7843.791000.0
FlashOcc [42]CSwin-B✓43.5213.3151.6228.0750.9155.6927.4631.0529.9829.2038.8643.6883.8745.6356.3359.0150.6344.56909.1
COTR [21]CSwin-B✓46.214.8553.2535.1950.8357.2535.3634.0633.5437.1438.9944.9784.4648.7357.6061.0851.6146.72840.34
HyDRa [36]C+RR-50-44.40---52.356.3-35.935.10--44.1-------
OCCFusion [23]C+LR-101-46.7911.6547.8132.0757.2757.5131.8040.1147.3533.7445.8150.3578.7937.1744.3653.3663.1863.20-
RadOcc-LC [43]C+LSwin-B✓49.3810.9358.2325.0157.8962.8534.0433.4550.0732.0548.8752.1182.9042.7355.2758.3468.6466.013333
SDG-KLC+LR-50✓50.1612.2657.1223.6958.7762.7434.5536.1950.132.0549.8951.2484.146.0557.261.4569.5665.7883
SDG-FusionC+LR-50✓51.6613.2157.7724.360.3364.2836.2139.4452.3635.8050.9153.6584.5647.4558.0061.6170.6767.65133
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.311, + 0.893, + 0.355 + ], + "angle": 0, + "content": "Table 1. 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. We present the mean IoU over categories and the IoUs for different classes. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. In the backbone, R represents ResNet, while Swin stands for Swin Transformer." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.366, + 0.891, + 0.627 + ], + "angle": 0, + "content": "
MethodInputBackbonemIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. suf.other flatsidewalkterrainmanmadevegetation
SurroundOcc [35]CR-10120.320.511.628.130.810.715.114.012.014.322.237.223.724.422.714.821.8
OccFormer [44]CR-10120.121.111.328.230.310.615.714.411.214.022.637.322.424.923.515.221.1
C-CONet [32]CR-10118.418.610.026.427.48.615.713.39.710.920.233.020.721.421.814.721.3
FB-Occ [16]CR-10119.620.611.326.929.810.413.613.711.411.520.638.221.524.622.714.821.6
RenderOcc [43]CR-10119.019.711.228.128.29.814.711.811.913.120.133.221.322.622.315.320.9
L-CONet [32]L-17.719.24.015.126.96.23.86.86.014.113.139.719.124.023.925.135.7
FlashOcc* [42]CR-5044.144.211.054.160.526.122.631.315.338.847.180.542.048.253.760.870.0
M-CONet [32]C+LR-10124.724.813.031.634.814.618.020.014.720.026.639.222.826.126.026.037.1
Co-Occ [24]C+LR-10127.128.116.134.037.217.021.620.815.921.928.742.325.429.128.628.238.0
OccFusion [23]C+L+RR-10127.327.119.633.736.221.724.825.316.321.830.039.519.924.926.528.940.4
DAOcc [40]C+LR-5030.530.819.534.038.825.027.729.922.523.231.641.025.929.429.935.243.5
SDG-FusionC+LR-5031.731.215.639.742.422.226.929.722.724.233.245.925.131.833.538.944.5
SDG-KL*C+LR-5050.451.420.959.665.834.133.149.023.445.153.776.544.053.558.265.272.9
SDG-Fusion*C+LR-5052.254.316.061.466.933.233.549.222.247.854.683.949.455.860.270.276.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.63, + 0.893, + 0.675 + ], + "angle": 0, + "content": "Table 2. 3D Occupancy prediction performance on the SurroundOcc-nuScenes validation set. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. * means the performance is achieved through a visible mask similar to [30]." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.688, + 0.437, + 0.766 + ], + "angle": 0, + "content": "
BaselineSDGFOADiou(%)mIoU(%)
✓90.2737.84
✓✓94.6248.51
✓✓94.7644.92
✓✓✓95.3551.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.771, + 0.468, + 0.814 + ], + "angle": 0, + "content": "Table 3. Ablation study on Occ3D-nuScenes dataset with SDG-Fusion. SDG: Semantic and Depth-Guided View Transformations. FOAD: Fusion-To-Occupancy-Driven Active Distillation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.828, + 0.228, + 0.845 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The Effectiveness of Each Component. The results are shown in Table 3, we can observe that all components make their own performance contributions. The baseline achieves" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.688, + 0.872, + 0.798 + ], + "angle": 0, + "content": "
rlIoU(%)mIoU (%)
1495.3451.15
1895.3551.66
11295.3851.28
2495.3550.4
2895.3951.03
21295.3051.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.802, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Table 4. Ablation study of the hyperparameter used in SDG view transformation module on Occ3D-nuScenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.893, + 0.903 + ], + "angle": 0, + "content": "90.27% of IoU and 37.84% of mIoU. We first integrated the Semantic and Depth-Guided (SDG) View Transformation into the baseline model, which brings 4.35% and 10.67% performance gain in IoU and mIoU. Fusion enhanced by in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.346, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.39, + 0.267, + 0.404 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.091, + 0.515, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.39, + 0.459, + 0.403 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.091, + 0.684, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.39, + 0.614, + 0.402 + ], + "angle": 0, + "content": "Our" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.091, + 0.855, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.39, + 0.814, + 0.402 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.892, + 0.434 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of SDG on the validation set of Occ3D-nuScenes. Each pair of rows displays results from day and low-light scene, respectively. Within each row, images from left to right represent the input images, baseline, our results, and the ground truth." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.445, + 0.439, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.575, + 0.468, + 0.601 + ], + "angle": 0, + "content": "Figure 7. Distance-based evaluation on Occ3D-nuScenes. As the distance increases, the point cloud becomes sparse." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.602, + 0.462, + 0.651 + ], + "angle": 0, + "content": "
k3579
IoU (%)95.3295.4095.3595.26
mIoU (%)51.2251.4051.6651.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.657, + 0.468, + 0.699 + ], + "angle": 0, + "content": "Table 5. Ablation study of the hyperparameter used in feature fusion of FOAD on Occ3D-nuScenes. The K denotes the size of the pixel region corresponding to the neighborhood feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.764 + ], + "angle": 0, + "content": "tegrating additional LiDAR information, the IoU and mIoU have been significantly improved by \\(4.49\\%\\) and \\(7.08\\%\\), respectively. By using both SDG and Fusion, outperforming the baseline by \\(5.19\\%\\) of IoU and \\(13.82\\%\\) of mIoU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The Effectiveness of SDG View Transformation. To further demonstrate the effect of SDG View Transformation, we conducted hyperparameter analysis experiments. In the SDG view transformation, the range \\( r \\) of bilinear growth discretization and the diffusion feature layers \\( l \\) control the virtual point generation of SDG. As shown in Table 4, lower depth precision (e.g., \\( r = 2 \\) and \\( l = 4 \\)) results in slightly reduced performance compared to other configurations. However, excessive depth precision does not lead to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.446, + 0.892, + 0.474 + ], + "angle": 0, + "content": "additional gains, with the optimal performance observed at \\( r = 1 \\) and \\( l = 8 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.571 + ], + "angle": 0, + "content": "The Effectiveness of FOAD Module. We perform a hyperparameter analysis of the FOAD module. For neighborhood feature fusion, the parameter \\( K \\) controls the fusion of features from neighboring pixels. As shown in Table 5, increasing \\( K \\) does not consistently improve performance, with optimal results achieved at \\( K = 7 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.599, + 0.619, + 0.615 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, designed to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. To address the inaccurate depth estimation in view transformations, we propose a semantic and depth-guided view transformation method. This approach integrates pixel semantics and corresponding point depth through diffusion and bilinear discretization, effectively reducing invalid image features and significantly enhancing the speed and accuracy of semantic occupancy. Meanwhile, We propose a fusion-to-occupancy-driven active distillation that incorporates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Our method achieves the SOTA performance with real-time processing on the Occ3D-nuScenes dataset and comparable performance on the more challenging SurroundOccnuScenes dataset, demonstrating its effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6758" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Geonho Bang, Kwangjin Choi, Jisong Kim, Dongsuk Kum, and Jun Won Choi. Radardistill: Boosting radar-based object detection performance via knowledge distillation from lidar features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15491-15500, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Jiahui Fu, Chen Gao, Zitian Wang, Lirong Yang, Xiaofei Wang, Beipeng Mu, and Si Liu. Eliminating cross-modal conflicts in bev space for lidar-camera 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16381-16387, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[4] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6185-6194, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[6] Jiawei Hou, Xiaoyan Li, Wenhao Guan, Gang Zhang, Di Feng, Yuheng Du, Xiangyang Xue, and Jian Pu. Fastocc: Accelerating 3d occupancy prediction by fusing the 2d bird's-eye view and perspective view. arXiv preprint arXiv:2403.02710, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[7] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.622, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[10] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[11] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[12] Jiale Li, Hang Dai, Hao Han, and Yong Ding. Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21694-21704, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[13] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[14] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[15] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[16] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[17] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[20] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.894, + 0.73 + ], + "angle": 0, + "content": "[21] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[22] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: A straightforward and effective multi-sensor fusion framework for 3d occupancy prediction. arXiv preprint arXiv:2403.01644, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[23] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction. IEEE Transactions on Intelligent Vehicles, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Jingyi Pan, Zipeng Wang, and Lin Wang. Co-occ: Coupling explicit feature fusion with volume rendering regularization for multi-modal 3d semantic occupancy prediction. IEEE Robotics and Automation Letters, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6759" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[25] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoji Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[26] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghang Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[27] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[28] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[29] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[30] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[31] Guoqing Wang, Zhongdao Wang, Pin Tang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Occgen: Generative multi-modal 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2404.15014, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[32] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[33] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17158-17168, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[34] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Anouar Laouichi, Martin Hofmann, and Gerhard Rigoll. Unleashing hydra: Hybrid fusion, depth consistency and radar for unified 3d perception. arXiv preprint arXiv:2403.07746, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[38] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.318, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[40] Zhen Yang, Yanpeng Dong, and Heng Wang. Daocc: 3d object detection assisted multi-sensor fusion for 3d occupancy prediction. arXiv preprint arXiv:2409.19972, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[42] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.516, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[43] Haiming Zhang, Xu Yan, Dongfeng Bai, Jiantao Gao, Pan Wang, Bingbing Liu, Shuguang Cui, and Zhen Li. Radocc: Learning cross-modality occupancy knowledge through rendering assisted distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7060-7068, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.586, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[44] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[45] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9939-9948, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6760" + } + ] +] \ No newline at end of file diff --git a/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_origin.pdf b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0215a7cf70b1213488a03de21a71bf4c0b18341e --- /dev/null +++ b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/14986d14-2105-4389-9a81-6b64509e2379_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88ad3668fa98c2e5614e2d8c53ce20c86660a0619d97d8303fe115880ee3334e +size 2787467 diff --git a/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/full.md b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac5fd7cf19563a405382e55e8eaad9145490833d --- /dev/null +++ b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/full.md @@ -0,0 +1,290 @@ +# SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction + +ZaiPeng Duan ChenXu Dang Xuzhong Hu Pei An Junfeng Ding Jie Zhan YunBiao Xu *Jie Ma Huazhong University of Science and Technology + +# Abstract + +Multimodal 3D occupancy prediction has garnered significant attention for its potential in autonomous driving. However, most existing approaches are single-modality: camera-based methods lack depth information, while LiDAR-based methods struggle with occlusions. Current lightweight methods primarily rely on the Lift-Splat-Shoot (LSS) pipeline, which suffers from inaccurate depth estimation and fails to fully exploit the geometric and semantic information of 3D LiDAR points. Therefore, we propose a novel multimodal occupancy prediction network called SDG-OCC, which incorporates a joint semantic and depth-guided view transformation coupled with a fusion-to-occupancy-driven active distillation. The enhanced view transformation constructs accurate depth distributions by integrating pixel semantics and co-point depth through diffusion and bilinear discretization. The fusion-to-occupancy-driven active distillation extracts rich semantic information from multimodal data and selectively transfers knowledge to image features based on LiDAR-identified regions. Finally, for optimal performance, we introduce SDG-Fusion, which uses fusion alone, and SDG-KL, which integrates both fusion and distillation for faster inference. Our method achieves state-of-the-art (SOTA) performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes dataset, demonstrating its effectiveness and robustness. The code will be released at https://github.com/DzpLab/SDGOCC. + +# 1. Introduction + +Accurate 3D perception of the surrounding environment forms the cornerstone of modern autonomous driving systems and robotics, ensuring efficient planning and safe control [7, 10]. In recent years, advancements in 3D object detection [14, 17, 20, 38, 41] and semantic segmentation [11, 12, 37, 39, 45] have significantly propelled the field of + +![](images/1e92a5da0b3b2d37dee13a72f983719da60602930adfd30d7a502535c3e73f51.jpg) +Figure 1. Comparisons of the mIoU and inference speed (FPS) of various 3D occupancy prediction methods on the Occ3D-nuScenes validation set. SDG-OCC achieves higher accuracy and real-time inference speed. + +3D perception. However, object detection relies on strict bounding boxes, making it difficult to recognize arbitrary shapes or unknown objects, while semantic segmentation struggles with fine-grained classification in complex scenes, especially under occlusion and overlap. In this context, 3D semantic occupancy prediction [30, 32] offers a more comprehensive approach to environment modeling. It simultaneously estimates the geometric structure and semantic categories of scene voxels, assigns labels to each 3D voxel, and provides a more complete perception, showing stronger robustness to arbitrary shapes and dynamic occlusions. + +Leveraging the complementary strengths of LiDAR and camera data is crucial for various 3D perception tasks. However, due to the heterogeneity between modalities, fusing LiDAR and camera data for 3D occupancy prediction remains challenging. Specifically, cameras provide rich semantic information but lack precise depth details, while LiDAR offers accurate depth information but only captures sparse data, potentially missing comprehensive scene details such as occluded objects. Existing methods often suffer from significant computational burdens (see Fig. 1), with some approaches attempting to leverage the LSS [27] pipeline for real-time performance. Although LSS simulates the uncertainty of each pixel's depth through depth + +![](images/fb93ae8a357d0683c4277efd18c53cb4e292e738d8ee380ba50f70ffb9b80e7a.jpg) +(a) Initial BEV features of LSS + +![](images/b2597659a28c705aedf7287e49843ddc9c31c24af4908c7bd39831884d7c3413.jpg) +(b) Our initial BEV features + +![](images/b94f120607103984eb1d4d92d9fc768bfb8eddc498e2ade016233ea57baa9dd2.jpg) +(c) Ground truth of BEV +Figure 2. (a) BEV feature map of LSS with a shape of $200 \times 200$ . We can observe that LSS has an extremely low utilization rate for BEV space. (b) The corresponding BEV features in SDG-OCC. Using depth and semantic information, the 2D-to-3D view transformation achieves efficient occupancy and utilization of the BEV features. (c) The corresponding BEV features in Ground Truth. (d) The corresponding BEV features in the occupancy grid. + +![](images/ac61a03477675b3caad4da7c0bdce97f3e8a1faa68c011e39438bfe40563c601.jpg) +(d) Occupancy grid of BEV + +distribution (with depth intervals typically set to $0.5\mathrm{m}$ ), its sparse BEV representation allows only $50\%$ of the grids to receive valid image features [16] (see Fig. 2 (a)). While increasing the depth interval can improve depth estimation accuracy to mitigate sparsity, it significantly increases computational demands. Additionally, while LiDAR can provide valuable geometric priors, fusion-based methods that process both point clouds and images simultaneously impose heavy computational burdens, thereby increasing the strain on real-time applications. + +To address these issues, we propose a multimodal 3D semantic occupancy prediction framework, named SDG-OCC, which aims to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. In this framework, we introduce a semantic and depth-guided view transformation to replace normal BEV feature generation. Specifically, after extracting features from camera data and obtaining semantic segmentation masks and depth distributions through a multi-task head, we use the semantic masks and depth maps provided by LiDAR to construct virtual points via local diffusion and bilinear discretization. Combined with the depth distribution, these points are then projected into the BEV space. The comparison between LSS and our generated BEV features is shown in Fig. 2. The SDG view transformation sig + +nificantly refines depth estimation accuracy and reduces redundant virtual point seeds, improving both the speed and accuracy of semantic occupancy. + +Secondly, we introduce a fusion-to-occupancy-driven active distillation module. We first fuse LiDAR and camera features in the BEV space and then unidirectionally selectively transferred multimodal knowledge to image features based on LiDAR-identified regions. Our proposed SDG-Fusion, which includes only fusion, achieved SOTA performance on the Occ3D-nuscene [30] and SurroundOccnuScenes [35] validation dataset. In comparison, SDGKL, which combines fusion and unidirectional distillation, achieves real-time speed with a slight performance penalty. + +Our contributions can be summarized as follows: + +- We introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, aimed at achieving higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. +- We propose a novel view transformation method that leverages the geometric and semantic information of point clouds to guide the 2D-3D view transformation. This significantly enhances the accuracy of depth estimation and improves both the speed and accuracy of semantic occupancy. +- We propose a fusion-to-occupancy-driven active distillation module that integrates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Building on this, we present SDG-Fusion for high performance and SDG-KL for faster inference. +- Our method achieves SOTA performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes validation dataset, demonstrating the effectiveness of our approach. + +# 2. Related Work + +# 2.1. Vision-Centric Occupancy Perception + +Inspired by Tesla's autonomous driving perception system, vision-centric occupancy perception has garnered significant attention from both industry and academia. MonoScene [2] is a pioneering work that used only RGB inputs. TPVFormer [9] combines surround multi-camera inputs and uses transformer-based methods to lift features into a tri-perspective view space. SurroundOcc [35] extends high-dimensional BEV features into occupancy features and directly performs spatial cross-attention to generate geometric information. VoxFormer [13] introduces a two-stage transformer-based semantic scene completion framework, capable of outputting complete 3D volumetric semantics from 2D images alone. FlashOcc [42] transforms the channel to height, lifting BEV output to 3D space, + +significantly improving operational efficiency. FBOcc [15] proposes a front-to-back view transformation module based on BEV features to address the limitations of different view transformations. Methods like UniOcc [26] and RenderOcc [25] use NeRF [34] to directly predict 3D semantic occupancy, but the rendering speed limits their efficiency. FastOcc [6] improves the occupancy prediction head to achieve a faster inference speed. COTR [21] builds compact 3D occupancy representations through explicit-implicit view transformation and coarse-to-fine semantic grouping. In this paper, we improve the speed and accuracy of 3D semantic occupancy prediction from the BEV space by incorporating the geometric and semantic information of the point cloud into the view transformation. + +# 2.2. Multi-Modal Occupancy Perception + +Multimodal occupancy perception leverages the strengths of multiple modalities to overcome the limitations of unimodal perception. OpenOccupancy[32] introduced a benchmark for LiDAR-camera semantic occupancy prediction. Inspired by BEVFusion, OccFusion [22] concatenates 3D feature volumes from different modalities along the feature channels, followed by convolutional layers to combine them. CO-Occ [24] introduced the Geometric and Semantic Fusion (GSFusion) module, identifying voxels containing both point cloud and visual information using k-nearest neighbors (KNN) search. OccGen [31] employs an adaptive fusion module to dynamically integrate occupancy representations from camera and LiDAR branches, using 3D convolutions to determine fusion weights for aggregating LiDAR and camera features. HyDR [36] proposes to integrate multimodal information in both perspective view (PV) and bird's-eye view (BEV) representation spaces. In this paper, we enhance view transformation by incorporating semantic segmentation masks and LiDAR depth maps to achieve higher occupancy accuracy. Additionally, we fuse BEV features from multimodal data and unidirectionally distill them into camera features, improving the accuracy and inference speed of 3D semantic occupancy prediction. + +# 3. Methodology + +# 3.1. Preliminary + +Given joint input from multi-view images and LiDAR data, 3D occupancy prediction aims to estimate the occupancy state and semantic classification of 3D voxels surrounding the ego vehicle. Specifically, the input consists of a $T$ -frame consequent sequence of images $X_{C} \in \mathbb{R}^{N_{C} \times H_{C} \times W_{C} \times 3}$ from $N_{C}$ surround-view cameras and point clouds $X_{L} \in \mathbb{R}^{N_{L} \times (3 + d)}$ as multimodal input, represented as $X = \{X_{C}, X_{L}\}$ . Here $H_{C}$ , $W_{C}$ represent the height and width of the image, respectively, $N_{L}$ denotes the number of point clouds and $d$ denotes the initial addi + +tional features of the point cloud. Subsequently, we train a neural network to generate an occupancy voxel map $Y \in \mathbb{R}^{H \times W \times D \times C_N}$ , where each voxel is assigned a label as unknown, occupied, or a semantic category from $\{C_0$ to $C_N\}$ . Here, $N$ denotes the total number of categories of interest, and $H, W, D$ represent the volume dimensions of the entire scene. + +# 3.2. Overall Architecture + +An overview of SDGOCC is shown in Fig. 3. It mainly consists of four key modules: image feature encoder to extract image features, semantic and depth-guided view transformation to construct 2D-3D feature transformation, fusion-to-occupancy-driven active distillation for fusing multimodal features and selectively transferring knowledge to the image features, and the occupancy prediction head for final output. + +# 3.3. Image Encoder + +The image feature encoder aims to capture multi-view features, providing a foundation for 2D-3D view transformation. Given RGB images from surround-view cameras, we first use a pre-trained image backbone network, such as classic ResNet [5] or strong Swin-Transformer [19], to extract multi-layer image features $F_{C} \in \mathbb{R}^{N_{C} \times C \times H \times W}$ . These features are then aggregated using a feature pyramid network (FPN) [18], which combines fine-grained features and coarse-grained features and down-sampling them to a specific scale, typically 1/16. + +# 3.4. SDG View Transformation + +The LSS pipeline is widely used for converting image features to BEV representations in 3D perception. It constructs virtual points based on a predefined depth range for each pixel and predicts the depth distribution weight $\alpha$ and context feature $c$ . The feature representation at depth $d$ is given by $p_d = \alpha_d c$ . All virtual points then are projected into BEV space, where features at each height $Z$ are aggregated to form BEV features. Although LSS handles depth uncertainties and ambiguities by modeling pixel depth using depth distributions, the number of per-pixel features remains large even with a 0.5-meter depth interval, an order of magnitude larger than point features. Meanwhile, BEV features are highly sparse, with less than $50\%$ of the image features being effective, leading to suboptimal occupancy prediction performance. Reducing the depth interval improves accuracy but significantly increases computational burden and introduces irrelevant features, as most of the occupancy grid remains empty. + +To address this, we propose a novel view transformation method that leverages the sparse depth information from LiDAR as a prior, diffuses within the same semantic class, followed by linear-increasing [28] and linear-decreasing dis + +![](images/7eaa41de797e1575753024e2a686ba37dba05c7a04dff0c8e0f1f54aeda7a000.jpg) +Figure 3. The overall architecture of SDG-OCC. $T$ -frame multi-view images and corresponding point clouds are fed into the image and LiDAR backbones to extract features. Image features are processed by a multi-task head to generate semantic masks and depth distributions, combined with LiDAR depth maps to create virtual points for image BEV features. These features are fused with point cloud BEV features and selectively transfer multimodal knowledge to image features based on LiDAR-identified regions. Finally, enhanced features are processed by the occupancy prediction head to generate the occupancy map. + +![](images/f1267d7a84bf8fa93762361cf5fdbfced289b27ce922f8217a61af59253824d2.jpg) +Figure 4. The overall architecture of the SDG view transformation. The image features processed by DepthNet are divided into texture features, depth features, and semantic features. Semantic features generate masks via a segmentation head, which are combined with depth maps from LiDAR for local diffusion and bilinear discretization to create virtual points. These points and their features receive pillar features from texture and depth features via the outer product, and then through BEV pooling to generate the final image BEV features. + +cretization to generate high-precision virtual point seeds around each co-visible point, as shown in Fig. 4. First, we extract features from multi-camera images and generate semantic segmentation masks via a multi-task head, while simultaneously extracting image textual features and depth distribution weights, with the depth head and semantic head supplementing cross-task information through gated attention. To better utilize semantic information, we select 4x downsampled features for view transformation, as higher downsampling increases the semantic and depth ambiguity + +of pixels. + +Given the differences in sparsity between images and point clouds, we combine image semantic segmentation masks and sparse projected depth maps provided by LiDAR to diffuse depth values within the same semantic category masks, generating a semi-dense extended depth map. This process is as follows: + +$$ +D _ {\text {t e m p}} (i, j) = \frac {\sum_ {(p , q) \in N (i , j)} D (p , q) \cdot \mathbb {I} [ M (p , q) = M (i , j) ]}{\sum_ {(p , q) \in N (i , j)} \mathbb {I} [ M (p , q) = M (i , j) ]}, \tag {1} +$$ + +where $N(i,j)$ represents the circle area with radius $r$ around the current point, and $M(i,j)$ denotes the segmentation mask with $\mathbf{N}$ category. And $\mathbb{I}[M(p,q) = M(i,j)]$ checks if the semantic label at $(\mathrm{p},\mathrm{q})$ matches that at $(\mathrm{i},\mathrm{j})$ , as follows: + +$$ +\mathbb {I} [ M (p, q) = M (i, j) ] = \left\{ \begin{array}{l l} 1, & \text {i f} M (p, q) = M (i, j) \neq 0 \\ 0, & \text {o t h e r w i s e} \end{array} . \right. \tag {2} +$$ + +And the final extended depth map $D_{(i,j)}$ replaces the original co-points of the $D_{\mathrm{temp}}(i,j)$ . + +Due to the projection deviation from 2D pixels to 3D points, we apply bidirectional linear incremental discretization to the extended depth map to obtain discrete virtual points, enhancing the accuracy of depth estimation. These steps reduce the number of virtual points, thereby improving inference speed. Finally, the image textual features $F_{t}$ and depth distribution weights $D_{w}$ are calculated by the + +![](images/52bb2163df066e884e468217a4f2be5104c99f3183ba6883e1cc8dbd79465a70.jpg) +Figure 5. (a) The pipeline of SDG-Fusion. (b) The pipeline of SDG-KL. (c) The overall architecture of fusion-to-occupancy-driven active distillation. The source feature extracts neighborhood information from the corresponding pixels of the cross feature, where the source serves as the query, and the cross feature acts as both key and value for feature interaction. This interaction is dynamically refined through a gated attention mechanism to produce the fused feature. Subsequently, an occupancy-driven active distillation is used to unidirectionally integrate multimodal information into the image feature. + +outer product $F_{t} \otimes D_{w}$ to derive features for each virtual point and generate the BEV features $F_{bev}^{C}$ of the camera through BEV pooling. This method effectively integrates semantic information with sparse depth data, significantly enhancing the accuracy of pixel depth estimation and improving the speed of view transformations. + +# 3.5. Fusion-to-occupancy-driven Active Distillation + +The LiDAR branch encompasses point cloud feature extraction, multimodal fusion, and occupancy-driven active distillation (as shown in Fig. 5). Initially, the point cloud data undergoes voxelization and normalization to generate the initial features. We choose SPVCNN as our point-voxel feature encoder due to its efficiency in representing sparse point clouds while effectively preserving fine-grained details. Subsequently, we compress the voxel features at the corresponding scale to generate the BEV features $F_{\text{bev}}^{L}$ . + +The complementary information from LiDAR and cameras is critical for 3D perception. A naive fusion method typically concatenates LiDAR and image BEV features along the channel dimension to enhance performance. However, feature misalignment due to extrinsic conflicts [3] limits the effectiveness of the fusion. Therefore, we propose a dynamic neighborhood feature fusion module. This module unidirectionally extracts neighborhood features from cross features and dynamically adjusts their weights into the source features using a gating attention mechanism. + +External projection deviations during the BEV feature construction process for LiDAR and images result in misalignment between LiDAR and camera BEV features [3, 29]. Therefore, we adopt neighborhood attention from [4] to extract local patch features corresponding to the pixel from the cross features, and dynamically adjust the weights through gated attention to selectively enhance the fused fea + +ture representation. Specifically, the image features $F_{bev}^{C}$ as the source feature, are represented as a feature vector sequence $F_{img} \in \mathbb{R}^{n \times m}$ , which is projected through a linear layer to obtain the query features $Q_{s} \in \mathbb{R}^{n \times q}$ . Similarly, the LiDAR features $F_{bev}^{L}$ as cross feature are projected to obtain the key $K_{c} \in \mathbb{R}^{n \times q}$ and value $V_{c} \in \mathbb{R}^{n \times v}$ features. The local neighborhood features $F_{\mathrm{neighbor}}$ for a query point $i$ are computed by the following equation: + +$$ +F _ {\text {n e i g h b o r}} = \sigma \left(\frac {Q _ {s} ^ {i} \cdot \left(K _ {c} ^ {n (i)}\right) ^ {T} + B (i , n (i))}{\sqrt {v}}\right) \cdot V _ {c} ^ {i}, \tag {3} +$$ + +where $n(i)$ represent the neighborhood with the size of $k$ centered at the same position in the cross feature, $B(i,\rho (i))$ denotes the relative positional biases and $\sigma$ denotes to Softmax. For each pixel in the feature map, we calculate the local neighborhood features. The fused features $F_{\text{bev}}^{\text{fuse}}$ are then obtained from the local neighborhood features through gated attention, as follows: + +$$ +F _ {b e v} ^ {f u s e} = \left(\sigma \left(\operatorname {C o n v} \left(f _ {\text {A v g}} \left(F _ {\text {n e i g h b o r}}\right)\right)\right) \cdot F _ {\text {n e i g h b o r}}, \quad (4) \right. +$$ + +where $\sigma$ denotes the sigmoid function and Conv denotes linear transform matrix (e.g., 1x1 convolution), $f_{\mathrm{Avg}}$ refers Adaptive Average Pooling. The fused features $F_{bev}^{fuse}$ are processed by the occupancy prediction head to obtain the SDG-Fusion model. + +Additionally, to ensure real-time, we propose an occupancy-driven active distillation, where fused features are unidirectionally transferred to the image features. Specifically, LiDAR features are used as the source feature, while image features serve as the cross feature, resulting in LiDAR-dominant fused features. Inspired by [1], we then divide the space into two regions: the active region(AR), where both LiDAR and image features are occupied, and the inactive region(IR), where only LiDAR features are occupied. The details are as follows: + +$$ +A R = \left(M _ {\text {f u s e d}, i, j} = 1\right) \wedge \left(M _ {\text {i m g}, i, j} = 1\right), \tag {5} +$$ + +$$ +I R = \left(M _ {\text {f u s e d}, i, j} = 1\right) \wedge \left(M _ {\text {i m g}, i, j} = 0\right), \tag {6} +$$ + +where a value of 1 in $M_{\mathrm{mode},i,j}$ indicates that the coordinate is occupied by the respective modality. Typically, the AR region is significantly larger than the IR region. To prevent the model from overemphasizing knowledge distillation in the AR region, we apply adaptive scaling based on the relative sizes of the AR and IR regions, as follows: + +$$ +W _ {I, i, j} ^ {\left(l _ {n}\right)} = \left\{ \begin{array}{l l} \alpha , & \text {i f} (i, j) \in A R, \\ \rho \times \beta , & \text {i f} (i, j) \in I R, \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {7} +$$ + +where $\rho = \frac{N_{AR}}{N_{IR}}$ represents the relative importance of the $IR$ over $AR$ , $\alpha$ and $\beta$ are the intrinsic balancing parameters, + +and $N_{AR}$ and $N_{IR}$ are the number of pixels in $AR$ and $IR$ , respectively. + +The distillation loss between BEV feature from teacher $F^t$ and student $F^s$ are: + +$$ +L _ {\text {d i s t i l l}} = \sum_ {c = 1} ^ {C} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} W _ {i, j} \left(F _ {b e v} ^ {f u s e} - F _ {b e v} ^ {C}\right) ^ {2}. \tag {8} +$$ + +The network is trained with the sum of distillation and classification loss. The image features $F_{bev}^{C}$ are processed by the occupancy prediction head to obtain the SDG-KL model. + +# 3.6. Occupancy Prediction + +To obtain 3D prediction outputs from coarse BEV features generated by view transformation, we propose an occupancy prediction system consisting of a BEV feature encoder and an occupancy prediction head. The BEV encoder uses several residual blocks for multi-scale feature diffusion and integrates a feature pyramid to acquire BEV features at the target scale. The occupancy prediction head extracts global features with multiple $3 \times 3$ convolutional layers and includes a channel-to-height transformation module. This module reshapes the BEV features from a $F_{out} \in \mathbb{R}^{B \times C \times H \times W}$ to $F_{final} \in \mathbb{R}^{B \times C_N \times D \times H \times W}$ , where $B, C, W, H$ , and $D$ represent the batch size, channel number, class number, and the dimensions of the 3D space, respectively, with $C = C_N \times D$ . Compared to traditional 3D encoders and occupancy prediction heads, this design significantly improves speed while maintaining comparable performance. + +# 4. Experiments + +We conduct experiments on the large-scale benchmark dataset Occ3D-nuScenes to validate the efficacy of our proposed methods. Additionally, we conduct ablation experiments to verify the effectiveness of each component in our method. + +# 4.1. Datasets + +Occ3D-nuScenes [30] is a large-scale autonomous driving dataset, which includes 1,000 urban traffic scenes under various conditions, the data is split into 700 training, 150 validation, and 150 testing scenes. The occupancy grid is defined within a range of $-40\mathrm{m}$ to $40\mathrm{m}$ along the X and Y axes and $-1\mathrm{m}$ to $5.4\mathrm{m}$ along the Z axis. The voxel size for occupancy labeling is set to $0.4\mathrm{m} \times 0.4\mathrm{m} \times 0.4\mathrm{m}$ . The semantic labels include 17 categories consisting of 16 known object classes with an additional 'empty' class. Compared to Occ3D-nuScenes, SurroundOcc [35] is also based on the nuScenes dataset but its prediction range is from $-50\mathrm{m}$ to $50\mathrm{m}$ for X and Y axes, and $-5\mathrm{m}$ to $3\mathrm{m}$ along the Z axis, with the voxel label size of $0.5\mathrm{m} \times 0.5\mathrm{m} \times 0.5\mathrm{m}$ . + +# 4.2. Implementation Details + +We use ResNet-50 as the default image backbone and SPVCNN as the LiDAR backbone. The model is trained on a GeForce RTX 4090 GPU using the AdamW optimizer with a learning rate of 1e-4 and gradient clipping. For semantic and depth-guided visual transformations, the bilinear incremental discretization range and the number of diffusion feature layers are set to 1 m and 8, respectively. + +# 4.3. Comparing with SOTA methods + +Occ3D-nuScenes. As shown in Table 1, we report the quantitative comparison of existing state-of-the-art methods for 3D occupancy prediction tasks on Occ3D-nuScenes. Most existing approaches are predominantly based on camera-only algorithms, with relatively few focusing on multisensor fusion. Our method, which employs a compact backbone and a lightweight LiDAR branch, achieves state-of-the-art performance in terms of mIoU and the majority of class-wise IoUs. Additionally, our approach achieves the best inference speed, meeting the real-time requirements of autonomous driving scenarios. The visualization on Occ3D-nuScenes validation set is shown in Fig. 6. Compared to the baseline, our method effectively identifies categories that the baseline fails to correctly predict in both day and night scenarios. More visualizations can be found in supplementary material. + +SurroundOcc-nuScenes. Table 2 provides a quantitative comparison on the SurroundOcc validation set, highlighting the performance of our method, SDG-OCC, compared to other approaches. Using both LiDAR and camera inputs, our method achieves SOTA performance on the SurroundOcc validation set, even over larger distances. This success is attributed to our semantic and depth-guided view transformation, which enhances depth estimation accuracy and enables robust occupancy prediction across varying grid sizes and distances. Furthermore, we utilize visual masks generated from [32], achieving performance comparable to OCC3d-nuscene. Notably, our method uses only a lightweight ResNet50 backbone and a lower resolution of $256 \times 704$ , underscoring its effectiveness and efficiency. + +Analysis of results within different ranges. We further evaluate different ranges surrounding the car to provide a comprehensive analysis. Fig. 7 clearly illustrates our mIoU and iou relative to the baseline FlashOcc. Short-range understanding is critical due to the limited reaction time for autonomous vehicles. Our method significantly outperforms the baseline in both mIoU and IoU. In long-range areas, where LiDAR data is sparse and few pixels define the depth of large regions, our approach still achieves superior IoU performance. + +
MethodInputBackboneVisible MaskmIoUothersbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetationTime(ms)
■■■■■■■■■■■■■■■■■
TPVFormer [9]CR-50✓34.27.6844.0117.6640.8846.9815.0620.5424.6924.6624.2629.2879.2740.6548.4949.4432.6329.82289.85
SurroundOcc [35]CR-101✓37.18.9746.3317.0846.5452.0120.0521.4723.5218.6731.5137.5681.9141.6450.7653.9342.9137.16303.03
OccFormer [44]CR-50✓37.49.1545.8418.2042.8050.2724.0020.8022.8620.9831.9438.1380.1338.2450.8354.346.4140.15-
VoxFormer [13]CR-101✓40.7------------------
FBOcc [16]CR-50✓42.114.3049.7130.046.6251.5429.329.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58-
PanoOcc [33]CR-101-42.1311.6750.4829.6449.4455.5223.2933.2630.5530.9934.4342.5783.3144.2354.4056.0445.9440.40322.58
FastOcc [6]CR-101✓40.7512.8646.5829.9346.0754.0923.7431.1030.6828.5233.0839.6983.3344.6553.9055.4642.6136.50221.2
BEVDet4D [8]CSwin-B✓42.512.3750.1526.9751.8654.6528.3828.9629.0228.2837.0542.5282.5543.1554.8758.3348.7843.791000.0
FlashOcc [42]CSwin-B✓43.5213.3151.6228.0750.9155.6927.4631.0529.9829.2038.8643.6883.8745.6356.3359.0150.6344.56909.1
COTR [21]CSwin-B✓46.214.8553.2535.1950.8357.2535.3634.0633.5437.1438.9944.9784.4648.7357.6061.0851.6146.72840.34
HyDRa [36]C+RR-50-44.40---52.356.3-35.935.10--44.1-------
OCCFusion [23]C+LR-101-46.7911.6547.8132.0757.2757.5131.8040.1147.3533.7445.8150.3578.7937.1744.3653.3663.1863.20-
RadOcc-LC [43]C+LSwin-B✓49.3810.9358.2325.0157.8962.8534.0433.4550.0732.0548.8752.1182.9042.7355.2758.3468.6466.013333
SDG-KLC+LR-50✓50.1612.2657.1223.6958.7762.7434.5536.1950.132.0549.8951.2484.146.0557.261.4569.5665.7883
SDG-FusionC+LR-50✓51.6613.2157.7724.360.3364.2836.2139.4452.3635.8050.9153.6584.5647.4558.0061.6170.6767.65133
+ +Table 1. 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. We present the mean IoU over categories and the IoUs for different classes. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. In the backbone, R represents ResNet, while Swin stands for Swin Transformer. + +
MethodInputBackbonemIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. suf.other flatsidewalkterrainmanmadevegetation
SurroundOcc [35]CR-10120.320.511.628.130.810.715.114.012.014.322.237.223.724.422.714.821.8
OccFormer [44]CR-10120.121.111.328.230.310.615.714.411.214.022.637.322.424.923.515.221.1
C-CONet [32]CR-10118.418.610.026.427.48.615.713.39.710.920.233.020.721.421.814.721.3
FB-Occ [16]CR-10119.620.611.326.929.810.413.613.711.411.520.638.221.524.622.714.821.6
RenderOcc [43]CR-10119.019.711.228.128.29.814.711.811.913.120.133.221.322.622.315.320.9
L-CONet [32]L-17.719.24.015.126.96.23.86.86.014.113.139.719.124.023.925.135.7
FlashOcc* [42]CR-5044.144.211.054.160.526.122.631.315.338.847.180.542.048.253.760.870.0
M-CONet [32]C+LR-10124.724.813.031.634.814.618.020.014.720.026.639.222.826.126.026.037.1
Co-Occ [24]C+LR-10127.128.116.134.037.217.021.620.815.921.928.742.325.429.128.628.238.0
OccFusion [23]C+L+RR-10127.327.119.633.736.221.724.825.316.321.830.039.519.924.926.528.940.4
DAOcc [40]C+LR-5030.530.819.534.038.825.027.729.922.523.231.641.025.929.429.935.243.5
SDG-FusionC+LR-5031.731.215.639.742.422.226.929.722.724.233.245.925.131.833.538.944.5
SDG-KL*C+LR-5050.451.420.959.665.834.133.149.023.445.153.776.544.053.558.265.272.9
SDG-Fusion*C+LR-5052.254.316.061.466.933.233.549.222.247.854.683.949.455.860.270.276.5
+ +Table 2. 3D Occupancy prediction performance on the SurroundOcc-nuScenes validation set. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. * means the performance is achieved through a visible mask similar to [30]. + +
BaselineSDGFOADiou(%)mIoU(%)
✓90.2737.84
✓✓94.6248.51
✓✓94.7644.92
✓✓✓95.3551.66
+ +# 4.4. Ablation study + +The Effectiveness of Each Component. The results are shown in Table 3, we can observe that all components make their own performance contributions. The baseline achieves + +Table 3. Ablation study on Occ3D-nuScenes dataset with SDG-Fusion. SDG: Semantic and Depth-Guided View Transformations. FOAD: Fusion-To-Occupancy-Driven Active Distillation. + +
rlIoU(%)mIoU (%)
1495.3451.15
1895.3551.66
11295.3851.28
2495.3550.4
2895.3951.03
21295.3051.12
+ +Table 4. Ablation study of the hyperparameter used in SDG view transformation module on Occ3D-nuScenes. + +90.27% of IoU and 37.84% of mIoU. We first integrated the Semantic and Depth-Guided (SDG) View Transformation into the baseline model, which brings 4.35% and 10.67% performance gain in IoU and mIoU. Fusion enhanced by in + +![](images/77149bf92e1d5104b39b390e8b152bee8a8780182e72f8b3f2238670780b6ec2.jpg) +Input Image + +![](images/2d5cd69217362d6897d2ce6468fcb0c8cdef1112113a0e868a9632dcd15b4b97.jpg) +Baseline + +![](images/7bb46181bc752c2fe4ae54fec6549e27aeb0b8fda5bda0a95b46b14e0d4031ea.jpg) +Our + +![](images/be27ec132b49c11470807cb630ea53f89d627881e63346cbe57617ab7d0fa0b2.jpg) +Ground Truth + +![](images/1798e593382ecc0aa9af069983464776dec4d2aaf94ceb94321970787a8d64aa.jpg) +Figure 6. Qualitative results of SDG on the validation set of Occ3D-nuScenes. Each pair of rows displays results from day and low-light scene, respectively. Within each row, images from left to right represent the input images, baseline, our results, and the ground truth. +Figure 7. Distance-based evaluation on Occ3D-nuScenes. As the distance increases, the point cloud becomes sparse. + +
k3579
IoU (%)95.3295.4095.3595.26
mIoU (%)51.2251.4051.6651.08
+ +Table 5. Ablation study of the hyperparameter used in feature fusion of FOAD on Occ3D-nuScenes. The K denotes the size of the pixel region corresponding to the neighborhood feature extraction. + +tegrating additional LiDAR information, the IoU and mIoU have been significantly improved by $4.49\%$ and $7.08\%$ , respectively. By using both SDG and Fusion, outperforming the baseline by $5.19\%$ of IoU and $13.82\%$ of mIoU. + +The Effectiveness of SDG View Transformation. To further demonstrate the effect of SDG View Transformation, we conducted hyperparameter analysis experiments. In the SDG view transformation, the range $r$ of bilinear growth discretization and the diffusion feature layers $l$ control the virtual point generation of SDG. As shown in Table 4, lower depth precision (e.g., $r = 2$ and $l = 4$ ) results in slightly reduced performance compared to other configurations. However, excessive depth precision does not lead to + +additional gains, with the optimal performance observed at $r = 1$ and $l = 8$ . + +The Effectiveness of FOAD Module. We perform a hyperparameter analysis of the FOAD module. For neighborhood feature fusion, the parameter $K$ controls the fusion of features from neighboring pixels. As shown in Table 5, increasing $K$ does not consistently improve performance, with optimal results achieved at $K = 7$ . + +# 5. Conclusion + +In this paper, we introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, designed to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. To address the inaccurate depth estimation in view transformations, we propose a semantic and depth-guided view transformation method. This approach integrates pixel semantics and corresponding point depth through diffusion and bilinear discretization, effectively reducing invalid image features and significantly enhancing the speed and accuracy of semantic occupancy. Meanwhile, We propose a fusion-to-occupancy-driven active distillation that incorporates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Our method achieves the SOTA performance with real-time processing on the Occ3D-nuScenes dataset and comparable performance on the more challenging SurroundOccnuScenes dataset, demonstrating its effectiveness. + +# References + +[1] Geonho Bang, Kwangjin Choi, Jisong Kim, Dongsuk Kum, and Jun Won Choi. Radardistill: Boosting radar-based object detection performance via knowledge distillation from lidar features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15491-15500, 2024. 5 +[2] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2 +[3] Jiahui Fu, Chen Gao, Zitian Wang, Lirong Yang, Xiaofei Wang, Beipeng Mu, and Si Liu. Eliminating cross-modal conflicts in bev space for lidar-camera 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16381-16387, 2024. 5 +[4] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6185-6194, 2023. 5 +[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[6] Jiawei Hou, Xiaoyan Li, Wenhao Guan, Gang Zhang, Di Feng, Yuheng Du, Xiangyang Xue, and Jian Pu. Fastocc: Accelerating 3d occupancy prediction by fusing the 2d bird's-eye view and perspective view. arXiv preprint arXiv:2403.02710, 2024. 3, 7 +[7] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7 +[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 7 +[10] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 1 +[11] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023. 1 +[12] Jiale Li, Hang Dai, Hao Han, and Yong Ding. Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21694-21704, 2023. 1 + +[13] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 2, 7 +[14] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 +[15] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 3 +[16] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7 +[17] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 1 +[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3 +[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3 +[20] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1 +[21] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 3, 7 +[22] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: A straightforward and effective multi-sensor fusion framework for 3d occupancy prediction. arXiv preprint arXiv:2403.01644, 2024. 3 +[23] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction. IEEE Transactions on Intelligent Vehicles, 2024. 7 +[24] Jingyi Pan, Zipeng Wang, and Lin Wang. Co-occ: Coupling explicit feature fusion with volume rendering regularization for multi-modal 3d semantic occupancy prediction. IEEE Robotics and Automation Letters, 2024. 3, 7 + +[25] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoji Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 3 +[26] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghang Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 3 +[27] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1 +[28] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 3 +[29] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 5 +[30] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 7 +[31] Guoqing Wang, Zhongdao Wang, Pin Tang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Occgen: Generative multi-modal 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2404.15014, 2024. 3 +[32] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 3, 6, 7 +[33] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17158-17168, 2024. 7 +[34] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3 +[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2, 6, 7 +[36] Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Anouar Laouichi, Martin Hofmann, and Gerhard Rigoll. Unleashing hydra: Hybrid fusion, depth consistency and radar for unified 3d perception. arXiv preprint arXiv:2403.07746, 2024. 3, 7 + +[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1 +[38] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 1 +[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 1 +[40] Zhen Yang, Yanpeng Dong, and Heng Wang. Daocc: 3d object detection assisted multi-sensor fusion for 3d occupancy prediction. arXiv preprint arXiv:2409.19972, 2024. 7 +[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 1 +[42] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 2, 7 +[43] Haiming Zhang, Xu Yan, Dongfeng Bai, Jiantao Gao, Pan Wang, Bingbing Liu, Shuguang Cui, and Zhen Li. Radocc: Learning cross-modality occupancy knowledge through rendering assisted distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7060-7068, 2024. 7 +[44] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 7 +[45] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9939-9948, 2021. 1 \ No newline at end of file diff --git a/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/images.zip b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..82aa1e6287da29f55e3eff0728eefe34ae89e7b6 --- /dev/null +++ b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbfb753e4791aa3fbc1238a02a78de11a92c5b9088e8e490f79535a51be60aeb +size 811540 diff --git a/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/layout.json b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aec3037f306ae702a9737665ce6da9480f4508fd --- /dev/null +++ b/2025/SDGOCC_ Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction/layout.json @@ -0,0 +1,8315 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 140 + ], + "type": "text", + "content": "SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 161, + 541, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 541, + 191 + ], + "type": "text", + "content": "ZaiPeng Duan ChenXu Dang Xuzhong Hu Pei An Junfeng Ding Jie Zhan YunBiao Xu *Jie Ma Huazhong University of Science and Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 192, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 230 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 230 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 242, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 242, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 45, + 242, + 290, + 579 + ], + "type": "text", + "content": "Multimodal 3D occupancy prediction has garnered significant attention for its potential in autonomous driving. However, most existing approaches are single-modality: camera-based methods lack depth information, while LiDAR-based methods struggle with occlusions. Current lightweight methods primarily rely on the Lift-Splat-Shoot (LSS) pipeline, which suffers from inaccurate depth estimation and fails to fully exploit the geometric and semantic information of 3D LiDAR points. Therefore, we propose a novel multimodal occupancy prediction network called SDG-OCC, which incorporates a joint semantic and depth-guided view transformation coupled with a fusion-to-occupancy-driven active distillation. The enhanced view transformation constructs accurate depth distributions by integrating pixel semantics and co-point depth through diffusion and bilinear discretization. The fusion-to-occupancy-driven active distillation extracts rich semantic information from multimodal data and selectively transfers knowledge to image features based on LiDAR-identified regions. Finally, for optimal performance, we introduce SDG-Fusion, which uses fusion alone, and SDG-KL, which integrates both fusion and distillation for faster inference. Our method achieves state-of-the-art (SOTA) performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes dataset, demonstrating its effectiveness and robustness. The code will be released at https://github.com/DzpLab/SDGOCC." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 621, + 288, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 288, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 288, + 695 + ], + "type": "text", + "content": "Accurate 3D perception of the surrounding environment forms the cornerstone of modern autonomous driving systems and robotics, ensuring efficient planning and safe control [7, 10]. In recent years, advancements in 3D object detection [14, 17, 20, 38, 41] and semantic segmentation [11, 12, 37, 39, 45] have significantly propelled the field of" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 217, + 545, + 362 + ], + "blocks": [ + { + "bbox": [ + 307, + 217, + 545, + 362 + ], + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 362 + ], + "type": "image", + "image_path": "1e92a5da0b3b2d37dee13a72f983719da60602930adfd30d7a502535c3e73f51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 367, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 547, + 413 + ], + "type": "text", + "content": "Figure 1. Comparisons of the mIoU and inference speed (FPS) of various 3D occupancy prediction methods on the Occ3D-nuScenes validation set. SDG-OCC achieves higher accuracy and real-time inference speed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 424, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 556 + ], + "type": "text", + "content": "3D perception. However, object detection relies on strict bounding boxes, making it difficult to recognize arbitrary shapes or unknown objects, while semantic segmentation struggles with fine-grained classification in complex scenes, especially under occlusion and overlap. In this context, 3D semantic occupancy prediction [30, 32] offers a more comprehensive approach to environment modeling. It simultaneously estimates the geometric structure and semantic categories of scene voxels, assigns labels to each 3D voxel, and provides a more complete perception, showing stronger robustness to arbitrary shapes and dynamic occlusions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 548, + 715 + ], + "type": "text", + "content": "Leveraging the complementary strengths of LiDAR and camera data is crucial for various 3D perception tasks. However, due to the heterogeneity between modalities, fusing LiDAR and camera data for 3D occupancy prediction remains challenging. Specifically, cameras provide rich semantic information but lack precise depth details, while LiDAR offers accurate depth information but only captures sparse data, potentially missing comprehensive scene details such as occluded objects. Existing methods often suffer from significant computational burdens (see Fig. 1), with some approaches attempting to leverage the LSS [27] pipeline for real-time performance. Although LSS simulates the uncertainty of each pixel's depth through depth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 201, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 201, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 201, + 714 + ], + "type": "text", + "content": "*Corresponding author: majie@hust.edu.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "6751" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 165, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 165, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 165, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 165, + 178 + ], + "type": "image", + "image_path": "fb93ae8a357d0683c4277efd18c53cb4e292e738d8ee380ba50f70ffb9b80e7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 180, + 161, + 189 + ], + "lines": [ + { + "bbox": [ + 60, + 180, + 161, + 189 + ], + "spans": [ + { + "bbox": [ + 60, + 180, + 161, + 189 + ], + "type": "text", + "content": "(a) Initial BEV features of LSS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 278, + 178 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 278, + 178 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 278, + 178 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 278, + 178 + ], + "type": "image", + "image_path": "b2597659a28c705aedf7287e49843ddc9c31c24af4908c7bd39831884d7c3413.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 180, + 269, + 189 + ], + "lines": [ + { + "bbox": [ + 179, + 180, + 269, + 189 + ], + "spans": [ + { + "bbox": [ + 179, + 180, + 269, + 189 + ], + "type": "text", + "content": "(b) Our initial BEV features" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 190, + 164, + 298 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 164, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 164, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 164, + 298 + ], + "type": "image", + "image_path": "b94f120607103984eb1d4d92d9fc768bfb8eddc498e2ade016233ea57baa9dd2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 300, + 149, + 309 + ], + "lines": [ + { + "bbox": [ + 69, + 300, + 149, + 309 + ], + "spans": [ + { + "bbox": [ + 69, + 300, + 149, + 309 + ], + "type": "text", + "content": "(c) Ground truth of BEV" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 316, + 287, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 395 + ], + "type": "text", + "content": "Figure 2. (a) BEV feature map of LSS with a shape of " + }, + { + "bbox": [ + 46, + 316, + 287, + 395 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 46, + 316, + 287, + 395 + ], + "type": "text", + "content": ". We can observe that LSS has an extremely low utilization rate for BEV space. (b) The corresponding BEV features in SDG-OCC. Using depth and semantic information, the 2D-to-3D view transformation achieves efficient occupancy and utilization of the BEV features. (c) The corresponding BEV features in Ground Truth. (d) The corresponding BEV features in the occupancy grid." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 190, + 277, + 298 + ], + "blocks": [ + { + "bbox": [ + 168, + 190, + 277, + 298 + ], + "lines": [ + { + "bbox": [ + 168, + 190, + 277, + 298 + ], + "spans": [ + { + "bbox": [ + 168, + 190, + 277, + 298 + ], + "type": "image", + "image_path": "ac61a03477675b3caad4da7c0bdce97f3e8a1faa68c011e39438bfe40563c601.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 300, + 268, + 310 + ], + "lines": [ + { + "bbox": [ + 178, + 300, + 268, + 310 + ], + "spans": [ + { + "bbox": [ + 178, + 300, + 268, + 310 + ], + "type": "text", + "content": "(d) Occupancy grid of BEV" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "type": "text", + "content": "distribution (with depth intervals typically set to " + }, + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "type": "text", + "content": "), its sparse BEV representation allows only " + }, + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 411, + 286, + 531 + ], + "type": "text", + "content": " of the grids to receive valid image features [16] (see Fig. 2 (a)). While increasing the depth interval can improve depth estimation accuracy to mitigate sparsity, it significantly increases computational demands. Additionally, while LiDAR can provide valuable geometric priors, fusion-based methods that process both point clouds and images simultaneously impose heavy computational burdens, thereby increasing the strain on real-time applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 714 + ], + "type": "text", + "content": "To address these issues, we propose a multimodal 3D semantic occupancy prediction framework, named SDG-OCC, which aims to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. In this framework, we introduce a semantic and depth-guided view transformation to replace normal BEV feature generation. Specifically, after extracting features from camera data and obtaining semantic segmentation masks and depth distributions through a multi-task head, we use the semantic masks and depth maps provided by LiDAR to construct virtual points via local diffusion and bilinear discretization. Combined with the depth distribution, these points are then projected into the BEV space. The comparison between LSS and our generated BEV features is shown in Fig. 2. The SDG view transformation sig" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "nificantly refines depth estimation accuracy and reduces redundant virtual point seeds, improving both the speed and accuracy of semantic occupancy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 109, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 545, + 228 + ], + "type": "text", + "content": "Secondly, we introduce a fusion-to-occupancy-driven active distillation module. We first fuse LiDAR and camera features in the BEV space and then unidirectionally selectively transferred multimodal knowledge to image features based on LiDAR-identified regions. Our proposed SDG-Fusion, which includes only fusion, achieved SOTA performance on the Occ3D-nuscene [30] and SurroundOccnuScenes [35] validation dataset. In comparison, SDGKL, which combines fusion and unidirectional distillation, achieves real-time speed with a slight performance penalty." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 229, + 515, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 229, + 515, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 229, + 515, + 240 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 243, + 545, + 493 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 306, + 243, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 545, + 290 + ], + "type": "text", + "content": "- We introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, aimed at achieving higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 545, + 361 + ], + "type": "text", + "content": "- We propose a novel view transformation method that leverages the geometric and semantic information of point clouds to guide the 2D-3D view transformation. This significantly enhances the accuracy of depth estimation and improves both the speed and accuracy of semantic occupancy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 361, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 432 + ], + "type": "text", + "content": "- We propose a fusion-to-occupancy-driven active distillation module that integrates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Building on this, we present SDG-Fusion for high performance and SDG-KL for faster inference." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 434, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 434, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 434, + 545, + 493 + ], + "type": "text", + "content": "- Our method achieves SOTA performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes validation dataset, demonstrating the effectiveness of our approach." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 506, + 392, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 392, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 392, + 518 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 526, + 508, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 508, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 508, + 540 + ], + "type": "text", + "content": "2.1. Vision-Centric Occupancy Perception" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Inspired by Tesla's autonomous driving perception system, vision-centric occupancy perception has garnered significant attention from both industry and academia. MonoScene [2] is a pioneering work that used only RGB inputs. TPVFormer [9] combines surround multi-camera inputs and uses transformer-based methods to lift features into a tri-perspective view space. SurroundOcc [35] extends high-dimensional BEV features into occupancy features and directly performs spatial cross-attention to generate geometric information. VoxFormer [13] introduces a two-stage transformer-based semantic scene completion framework, capable of outputting complete 3D volumetric semantics from 2D images alone. FlashOcc [42] transforms the channel to height, lifting BEV output to 3D space," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6752" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "significantly improving operational efficiency. FBOcc [15] proposes a front-to-back view transformation module based on BEV features to address the limitations of different view transformations. Methods like UniOcc [26] and RenderOcc [25] use NeRF [34] to directly predict 3D semantic occupancy, but the rendering speed limits their efficiency. FastOcc [6] improves the occupancy prediction head to achieve a faster inference speed. COTR [21] builds compact 3D occupancy representations through explicit-implicit view transformation and coarse-to-fine semantic grouping. In this paper, we improve the speed and accuracy of 3D semantic occupancy prediction from the BEV space by incorporating the geometric and semantic information of the point cloud into the view transformation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 248, + 239, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 239, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 239, + 262 + ], + "type": "text", + "content": "2.2. Multi-Modal Occupancy Perception" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 268, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 289, + 543 + ], + "type": "text", + "content": "Multimodal occupancy perception leverages the strengths of multiple modalities to overcome the limitations of unimodal perception. OpenOccupancy[32] introduced a benchmark for LiDAR-camera semantic occupancy prediction. Inspired by BEVFusion, OccFusion [22] concatenates 3D feature volumes from different modalities along the feature channels, followed by convolutional layers to combine them. CO-Occ [24] introduced the Geometric and Semantic Fusion (GSFusion) module, identifying voxels containing both point cloud and visual information using k-nearest neighbors (KNN) search. OccGen [31] employs an adaptive fusion module to dynamically integrate occupancy representations from camera and LiDAR branches, using 3D convolutions to determine fusion weights for aggregating LiDAR and camera features. HyDR [36] proposes to integrate multimodal information in both perspective view (PV) and bird's-eye view (BEV) representation spaces. In this paper, we enhance view transformation by incorporating semantic segmentation masks and LiDAR depth maps to achieve higher occupancy accuracy. Additionally, we fuse BEV features from multimodal data and unidirectionally distill them into camera features, improving the accuracy and inference speed of 3D semantic occupancy prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 555, + 129, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 129, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 129, + 568 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 574, + 127, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 127, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 127, + 588 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Given joint input from multi-view images and LiDAR data, 3D occupancy prediction aims to estimate the occupancy state and semantic classification of 3D voxels surrounding the ego vehicle. Specifically, the input consists of a " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "-frame consequent sequence of images " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "X_{C} \\in \\mathbb{R}^{N_{C} \\times H_{C} \\times W_{C} \\times 3}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{C}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " surround-view cameras and point clouds " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "X_{L} \\in \\mathbb{R}^{N_{L} \\times (3 + d)}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " as multimodal input, represented as " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "X = \\{X_{C}, X_{L}\\}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "H_{C}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "W_{C}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " represent the height and width of the image, respectively, " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "N_{L}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " denotes the number of point clouds and " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " denotes the initial addi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "tional features of the point cloud. Subsequently, we train a neural network to generate an occupancy voxel map " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{H \\times W \\times D \\times C_N}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", where each voxel is assigned a label as unknown, occupied, or a semantic category from " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\{C_0" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "C_N\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " denotes the total number of categories of interest, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "H, W, D" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " represent the volume dimensions of the entire scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 163, + 426, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 163, + 426, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 163, + 426, + 176 + ], + "type": "text", + "content": "3.2. Overall Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 278 + ], + "type": "text", + "content": "An overview of SDGOCC is shown in Fig. 3. It mainly consists of four key modules: image feature encoder to extract image features, semantic and depth-guided view transformation to construct 2D-3D feature transformation, fusion-to-occupancy-driven active distillation for fusing multimodal features and selectively transferring knowledge to the image features, and the occupancy prediction head for final output." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 285, + 400, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 400, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 400, + 298 + ], + "type": "text", + "content": "3.3. Image Encoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 425 + ], + "type": "text", + "content": "The image feature encoder aims to capture multi-view features, providing a foundation for 2D-3D view transformation. Given RGB images from surround-view cameras, we first use a pre-trained image backbone network, such as classic ResNet [5] or strong Swin-Transformer [19], to extract multi-layer image features " + }, + { + "bbox": [ + 304, + 304, + 545, + 425 + ], + "type": "inline_equation", + "content": "F_{C} \\in \\mathbb{R}^{N_{C} \\times C \\times H \\times W}" + }, + { + "bbox": [ + 304, + 304, + 545, + 425 + ], + "type": "text", + "content": ". These features are then aggregated using a feature pyramid network (FPN) [18], which combines fine-grained features and coarse-grained features and down-sampling them to a specific scale, typically 1/16." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 431, + 454, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 454, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 454, + 443 + ], + "type": "text", + "content": "3.4. SDG View Transformation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": "The LSS pipeline is widely used for converting image features to BEV representations in 3D perception. It constructs virtual points based on a predefined depth range for each pixel and predicts the depth distribution weight " + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": " and context feature " + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": ". The feature representation at depth " + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "inline_equation", + "content": "p_d = \\alpha_d c" + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": ". All virtual points then are projected into BEV space, where features at each height " + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": " are aggregated to form BEV features. Although LSS handles depth uncertainties and ambiguities by modeling pixel depth using depth distributions, the number of per-pixel features remains large even with a 0.5-meter depth interval, an order of magnitude larger than point features. Meanwhile, BEV features are highly sparse, with less than " + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 665 + ], + "type": "text", + "content": " of the image features being effective, leading to suboptimal occupancy prediction performance. Reducing the depth interval improves accuracy but significantly increases computational burden and introduces irrelevant features, as most of the occupancy grid remains empty." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "To address this, we propose a novel view transformation method that leverages the sparse depth information from LiDAR as a prior, diffuses within the same semantic class, followed by linear-increasing [28] and linear-decreasing dis" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6753" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 541, + 289 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 541, + 289 + ], + "type": "image", + "image_path": "7eaa41de797e1575753024e2a686ba37dba05c7a04dff0c8e0f1f54aeda7a000.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "type": "text", + "content": "Figure 3. The overall architecture of SDG-OCC. " + }, + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "type": "text", + "content": "-frame multi-view images and corresponding point clouds are fed into the image and LiDAR backbones to extract features. Image features are processed by a multi-task head to generate semantic masks and depth distributions, combined with LiDAR depth maps to create virtual points for image BEV features. These features are fused with point cloud BEV features and selectively transfer multimodal knowledge to image features based on LiDAR-identified regions. Finally, enhanced features are processed by the occupancy prediction head to generate the occupancy map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 354, + 287, + 467 + ], + "blocks": [ + { + "bbox": [ + 48, + 354, + 287, + 467 + ], + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 467 + ], + "type": "image", + "image_path": "f1267d7a84bf8fa93762361cf5fdbfced289b27ce922f8217a61af59253824d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 480, + 288, + 579 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 288, + 579 + ], + "type": "text", + "content": "Figure 4. The overall architecture of the SDG view transformation. The image features processed by DepthNet are divided into texture features, depth features, and semantic features. Semantic features generate masks via a segmentation head, which are combined with depth maps from LiDAR for local diffusion and bilinear discretization to create virtual points. These points and their features receive pillar features from texture and depth features via the outer product, and then through BEV pooling to generate the final image BEV features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "cretization to generate high-precision virtual point seeds around each co-visible point, as shown in Fig. 4. First, we extract features from multi-camera images and generate semantic segmentation masks via a multi-task head, while simultaneously extracting image textual features and depth distribution weights, with the depth head and semantic head supplementing cross-task information through gated attention. To better utilize semantic information, we select 4x downsampled features for view transformation, as higher downsampling increases the semantic and depth ambiguity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 356, + 346, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 346, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 346, + 366 + ], + "type": "text", + "content": "of pixels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 368, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 439 + ], + "type": "text", + "content": "Given the differences in sparsity between images and point clouds, we combine image semantic segmentation masks and sparse projected depth maps provided by LiDAR to diffuse depth values within the same semantic category masks, generating a semi-dense extended depth map. This process is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 444, + 545, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 499 + ], + "type": "interline_equation", + "content": "D _ {\\text {t e m p}} (i, j) = \\frac {\\sum_ {(p , q) \\in N (i , j)} D (p , q) \\cdot \\mathbb {I} [ M (p , q) = M (i , j) ]}{\\sum_ {(p , q) \\in N (i , j)} \\mathbb {I} [ M (p , q) = M (i , j) ]}, \\tag {1}", + "image_path": "e4b3f0e7a665db08559751bf105a1702fde9b558b528330b39a76788652bb459.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "N(i,j)" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": " represents the circle area with radius " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": " around the current point, and " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "M(i,j)" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": " denotes the segmentation mask with " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": " category. And " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[M(p,q) = M(i,j)]" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": " checks if the semantic label at " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "(\\mathrm{p},\\mathrm{q})" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": " matches that at " + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "inline_equation", + "content": "(\\mathrm{i},\\mathrm{j})" + }, + { + "bbox": [ + 304, + 500, + 545, + 559 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 564, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 564, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\mathbb {I} [ M (p, q) = M (i, j) ] = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} M (p, q) = M (i, j) \\neq 0 \\\\ 0, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}", + "image_path": "5d55d5e385907498ca53531fb77b1827459cef86af09c1332112282c232a13a4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "type": "text", + "content": "And the final extended depth map " + }, + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "type": "inline_equation", + "content": "D_{(i,j)}" + }, + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "type": "text", + "content": " replaces the original co-points of the " + }, + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{temp}}(i,j)" + }, + { + "bbox": [ + 304, + 605, + 545, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "Due to the projection deviation from 2D pixels to 3D points, we apply bidirectional linear incremental discretization to the extended depth map to obtain discrete virtual points, enhancing the accuracy of depth estimation. These steps reduce the number of virtual points, thereby improving inference speed. Finally, the image textual features " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": " and depth distribution weights " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "D_{w}" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": " are calculated by the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6754" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 289, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 171 + ], + "type": "image", + "image_path": "52bb2163df066e884e468217a4f2be5104c99f3183ba6883e1cc8dbd79465a70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 179, + 289, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 289, + 291 + ], + "type": "text", + "content": "Figure 5. (a) The pipeline of SDG-Fusion. (b) The pipeline of SDG-KL. (c) The overall architecture of fusion-to-occupancy-driven active distillation. The source feature extracts neighborhood information from the corresponding pixels of the cross feature, where the source serves as the query, and the cross feature acts as both key and value for feature interaction. This interaction is dynamically refined through a gated attention mechanism to produce the fused feature. Subsequently, an occupancy-driven active distillation is used to unidirectionally integrate multimodal information into the image feature." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "content": "outer product " + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "inline_equation", + "content": "F_{t} \\otimes D_{w}" + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "content": " to derive features for each virtual point and generate the BEV features " + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "inline_equation", + "content": "F_{bev}^{C}" + }, + { + "bbox": [ + 46, + 300, + 288, + 373 + ], + "type": "text", + "content": " of the camera through BEV pooling. This method effectively integrates semantic information with sparse depth data, significantly enhancing the accuracy of pixel depth estimation and improving the speed of view transformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 381, + 287, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 287, + 395 + ], + "type": "text", + "content": "3.5. Fusion-to-occupancy-driven Active Distillation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 510 + ], + "type": "text", + "content": "The LiDAR branch encompasses point cloud feature extraction, multimodal fusion, and occupancy-driven active distillation (as shown in Fig. 5). Initially, the point cloud data undergoes voxelization and normalization to generate the initial features. We choose SPVCNN as our point-voxel feature encoder due to its efficiency in representing sparse point clouds while effectively preserving fine-grained details. Subsequently, we compress the voxel features at the corresponding scale to generate the BEV features " + }, + { + "bbox": [ + 46, + 401, + 287, + 510 + ], + "type": "inline_equation", + "content": "F_{\\text{bev}}^{L}" + }, + { + "bbox": [ + 46, + 401, + 287, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 628 + ], + "type": "text", + "content": "The complementary information from LiDAR and cameras is critical for 3D perception. A naive fusion method typically concatenates LiDAR and image BEV features along the channel dimension to enhance performance. However, feature misalignment due to extrinsic conflicts [3] limits the effectiveness of the fusion. Therefore, we propose a dynamic neighborhood feature fusion module. This module unidirectionally extracts neighborhood features from cross features and dynamically adjusts their weights into the source features using a gating attention mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "External projection deviations during the BEV feature construction process for LiDAR and images result in misalignment between LiDAR and camera BEV features [3, 29]. Therefore, we adopt neighborhood attention from [4] to extract local patch features corresponding to the pixel from the cross features, and dynamically adjust the weights through gated attention to selectively enhance the fused fea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": "ture representation. Specifically, the image features " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "F_{bev}^{C}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " as the source feature, are represented as a feature vector sequence " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "F_{img} \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": ", which is projected through a linear layer to obtain the query features " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "Q_{s} \\in \\mathbb{R}^{n \\times q}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": ". Similarly, the LiDAR features " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "F_{bev}^{L}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " as cross feature are projected to obtain the key " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "K_{c} \\in \\mathbb{R}^{n \\times q}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "V_{c} \\in \\mathbb{R}^{n \\times v}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " features. The local neighborhood features " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{neighbor}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " for a query point " + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 547, + 169 + ], + "type": "text", + "content": " are computed by the following equation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 175, + 545, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 545, + 208 + ], + "type": "interline_equation", + "content": "F _ {\\text {n e i g h b o r}} = \\sigma \\left(\\frac {Q _ {s} ^ {i} \\cdot \\left(K _ {c} ^ {n (i)}\\right) ^ {T} + B (i , n (i))}{\\sqrt {v}}\\right) \\cdot V _ {c} ^ {i}, \\tag {3}", + "image_path": "f4b666cd8191694c8606a608f1f0d2936159d3fc3612481830fb8a2bb81899b4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "inline_equation", + "content": "n(i)" + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "content": " represent the neighborhood with the size of " + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "content": " centered at the same position in the cross feature, " + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "inline_equation", + "content": "B(i,\\rho (i))" + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "content": " denotes the relative positional biases and " + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "content": " denotes to Softmax. For each pixel in the feature map, we calculate the local neighborhood features. The fused features " + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "inline_equation", + "content": "F_{\\text{bev}}^{\\text{fuse}}" + }, + { + "bbox": [ + 304, + 214, + 547, + 299 + ], + "type": "text", + "content": " are then obtained from the local neighborhood features through gated attention, as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 332, + 305, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 305, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 332, + 305, + 545, + 322 + ], + "type": "interline_equation", + "content": "F _ {b e v} ^ {f u s e} = \\left(\\sigma \\left(\\operatorname {C o n v} \\left(f _ {\\text {A v g}} \\left(F _ {\\text {n e i g h b o r}}\\right)\\right)\\right) \\cdot F _ {\\text {n e i g h b o r}}, \\quad (4) \\right.", + "image_path": "e9249d1e3e303ef92f2e83293a7816f3b986ebd28500774ae3b50b1d3a74e8df.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "content": " denotes the sigmoid function and Conv denotes linear transform matrix (e.g., 1x1 convolution), " + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Avg}}" + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "content": " refers Adaptive Average Pooling. The fused features " + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "inline_equation", + "content": "F_{bev}^{fuse}" + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "content": " are processed by the occupancy prediction head to obtain the SDG-Fusion model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "Additionally, to ensure real-time, we propose an occupancy-driven active distillation, where fused features are unidirectionally transferred to the image features. Specifically, LiDAR features are used as the source feature, while image features serve as the cross feature, resulting in LiDAR-dominant fused features. Inspired by [1], we then divide the space into two regions: the active region(AR), where both LiDAR and image features are occupied, and the inactive region(IR), where only LiDAR features are occupied. The details are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 517, + 545, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 517, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 341, + 517, + 545, + 532 + ], + "type": "interline_equation", + "content": "A R = \\left(M _ {\\text {f u s e d}, i, j} = 1\\right) \\wedge \\left(M _ {\\text {i m g}, i, j} = 1\\right), \\tag {5}", + "image_path": "db17e76f4c6a557f6fe30f235f245c721bbf736211d1db8f9116fc1e91f08d17.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 538, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 538, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 343, + 538, + 545, + 552 + ], + "type": "interline_equation", + "content": "I R = \\left(M _ {\\text {f u s e d}, i, j} = 1\\right) \\wedge \\left(M _ {\\text {i m g}, i, j} = 0\\right), \\tag {6}", + "image_path": "b6d39faab9907b6c50d6fe4c1ed2681cad4a6d13d4158f1a201a6d69ea9c1ad7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": "where a value of 1 in " + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{mode},i,j}" + }, + { + "bbox": [ + 304, + 555, + 545, + 628 + ], + "type": "text", + "content": " indicates that the coordinate is occupied by the respective modality. Typically, the AR region is significantly larger than the IR region. To prevent the model from overemphasizing knowledge distillation in the AR region, we apply adaptive scaling based on the relative sizes of the AR and IR regions, as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 635, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 635, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 351, + 635, + 545, + 681 + ], + "type": "interline_equation", + "content": "W _ {I, i, j} ^ {\\left(l _ {n}\\right)} = \\left\\{ \\begin{array}{l l} \\alpha , & \\text {i f} (i, j) \\in A R, \\\\ \\rho \\times \\beta , & \\text {i f} (i, j) \\in I R, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {7}", + "image_path": "1eb0978ea1fd12cb1c7a01fe6fa96db7faaffc1dedf18dbd4f926eeabfa97c6f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\rho = \\frac{N_{AR}}{N_{IR}}" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " represents the relative importance of the " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "IR" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " are the intrinsic balancing parameters," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6755" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "N_{AR}" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "N_{IR}" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": " are the number of pixels in " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "AR" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "IR" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "type": "text", + "content": "The distillation loss between BEV feature from teacher " + }, + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "type": "inline_equation", + "content": "F^t" + }, + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "type": "text", + "content": " and student " + }, + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "type": "inline_equation", + "content": "F^s" + }, + { + "bbox": [ + 47, + 96, + 286, + 119 + ], + "type": "text", + "content": " are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 125, + 286, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 125, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 74, + 125, + 286, + 159 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i s t i l l}} = \\sum_ {c = 1} ^ {C} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} W _ {i, j} \\left(F _ {b e v} ^ {f u s e} - F _ {b e v} ^ {C}\\right) ^ {2}. \\tag {8}", + "image_path": "bf5a054ef6e5f07a1f390a1446a81350d3e0de35025996060c462e0ac4f58d15.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 166, + 286, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 166, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 286, + 213 + ], + "type": "text", + "content": "The network is trained with the sum of distillation and classification loss. The image features " + }, + { + "bbox": [ + 46, + 166, + 286, + 213 + ], + "type": "inline_equation", + "content": "F_{bev}^{C}" + }, + { + "bbox": [ + 46, + 166, + 286, + 213 + ], + "type": "text", + "content": " are processed by the occupancy prediction head to obtain the SDG-KL model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 221, + 173, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 173, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 173, + 233 + ], + "type": "text", + "content": "3.6. Occupancy Prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": "To obtain 3D prediction outputs from coarse BEV features generated by view transformation, we propose an occupancy prediction system consisting of a BEV feature encoder and an occupancy prediction head. The BEV encoder uses several residual blocks for multi-scale feature diffusion and integrates a feature pyramid to acquire BEV features at the target scale. The occupancy prediction head extracts global features with multiple " + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": " convolutional layers and includes a channel-to-height transformation module. This module reshapes the BEV features from a " + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "inline_equation", + "content": "F_{out} \\in \\mathbb{R}^{B \\times C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "inline_equation", + "content": "F_{final} \\in \\mathbb{R}^{B \\times C_N \\times D \\times H \\times W}" + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "inline_equation", + "content": "B, C, W, H" + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": " represent the batch size, channel number, class number, and the dimensions of the 3D space, respectively, with " + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "inline_equation", + "content": "C = C_N \\times D" + }, + { + "bbox": [ + 46, + 239, + 286, + 441 + ], + "type": "text", + "content": ". Compared to traditional 3D encoders and occupancy prediction heads, this design significantly improves speed while maintaining comparable performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 127, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 127, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 127, + 466 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 533 + ], + "type": "text", + "content": "We conduct experiments on the large-scale benchmark dataset Occ3D-nuScenes to validate the efficacy of our proposed methods. Additionally, we conduct ablation experiments to verify the effectiveness of each component in our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 540, + 109, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 109, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 109, + 551 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": "Occ3D-nuScenes [30] is a large-scale autonomous driving dataset, which includes 1,000 urban traffic scenes under various conditions, the data is split into 700 training, 150 validation, and 150 testing scenes. The occupancy grid is defined within a range of " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "-40\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "40\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " along the X and Y axes and " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "-1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "5.4\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " along the Z axis. The voxel size for occupancy labeling is set to " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m} \\times 0.4\\mathrm{m} \\times 0.4\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": ". The semantic labels include 17 categories consisting of 16 known object classes with an additional 'empty' class. Compared to Occ3D-nuScenes, SurroundOcc [35] is also based on the nuScenes dataset but its prediction range is from " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "-50\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "50\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " for X and Y axes, and " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "-5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "3\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": " along the Z axis, with the voxel label size of " + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m} \\times 0.5\\mathrm{m} \\times 0.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 558, + 286, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 438, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 438, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 438, + 84 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 95, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 545, + 179 + ], + "type": "text", + "content": "We use ResNet-50 as the default image backbone and SPVCNN as the LiDAR backbone. The model is trained on a GeForce RTX 4090 GPU using the AdamW optimizer with a learning rate of 1e-4 and gradient clipping. For semantic and depth-guided visual transformations, the bilinear incremental discretization range and the number of diffusion feature layers are set to 1 m and 8, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 201, + 479, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 479, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 479, + 214 + ], + "type": "text", + "content": "4.3. Comparing with SOTA methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 224, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 416 + ], + "type": "text", + "content": "Occ3D-nuScenes. As shown in Table 1, we report the quantitative comparison of existing state-of-the-art methods for 3D occupancy prediction tasks on Occ3D-nuScenes. Most existing approaches are predominantly based on camera-only algorithms, with relatively few focusing on multisensor fusion. Our method, which employs a compact backbone and a lightweight LiDAR branch, achieves state-of-the-art performance in terms of mIoU and the majority of class-wise IoUs. Additionally, our approach achieves the best inference speed, meeting the real-time requirements of autonomous driving scenarios. The visualization on Occ3D-nuScenes validation set is shown in Fig. 6. Compared to the baseline, our method effectively identifies categories that the baseline fails to correctly predict in both day and night scenarios. More visualizations can be found in supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "content": "SurroundOcc-nuScenes. Table 2 provides a quantitative comparison on the SurroundOcc validation set, highlighting the performance of our method, SDG-OCC, compared to other approaches. Using both LiDAR and camera inputs, our method achieves SOTA performance on the SurroundOcc validation set, even over larger distances. This success is attributed to our semantic and depth-guided view transformation, which enhances depth estimation accuracy and enables robust occupancy prediction across varying grid sizes and distances. Furthermore, we utilize visual masks generated from [32], achieving performance comparable to OCC3d-nuscene. Notably, our method uses only a lightweight ResNet50 backbone and a lower resolution of " + }, + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "inline_equation", + "content": "256 \\times 704" + }, + { + "bbox": [ + 304, + 422, + 545, + 589 + ], + "type": "text", + "content": ", underscoring its effectiveness and efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Analysis of results within different ranges. We further evaluate different ranges surrounding the car to provide a comprehensive analysis. Fig. 7 clearly illustrates our mIoU and iou relative to the baseline FlashOcc. Short-range understanding is critical due to the limited reaction time for autonomous vehicles. Our method significantly outperforms the baseline in both mIoU and IoU. In long-range areas, where LiDAR data is sparse and few pixels define the depth of large regions, our approach still achieves superior IoU performance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6756" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 72, + 546, + 243 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 546, + 243 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 546, + 243 + ], + "type": "table", + "html": "
MethodInputBackboneVisible MaskmIoUothersbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. surf.other flatsidewalkterrainmannadevegetationTime(ms)
■■■■■■■■■■■■■■■■■
TPVFormer [9]CR-50✓34.27.6844.0117.6640.8846.9815.0620.5424.6924.6624.2629.2879.2740.6548.4949.4432.6329.82289.85
SurroundOcc [35]CR-101✓37.18.9746.3317.0846.5452.0120.0521.4723.5218.6731.5137.5681.9141.6450.7653.9342.9137.16303.03
OccFormer [44]CR-50✓37.49.1545.8418.2042.8050.2724.0020.8022.8620.9831.9438.1380.1338.2450.8354.346.4140.15-
VoxFormer [13]CR-101✓40.7------------------
FBOcc [16]CR-50✓42.114.3049.7130.046.6251.5429.329.1329.3530.4834.9739.3683.0747.1655.6259.8844.8939.58-
PanoOcc [33]CR-101-42.1311.6750.4829.6449.4455.5223.2933.2630.5530.9934.4342.5783.3144.2354.4056.0445.9440.40322.58
FastOcc [6]CR-101✓40.7512.8646.5829.9346.0754.0923.7431.1030.6828.5233.0839.6983.3344.6553.9055.4642.6136.50221.2
BEVDet4D [8]CSwin-B✓42.512.3750.1526.9751.8654.6528.3828.9629.0228.2837.0542.5282.5543.1554.8758.3348.7843.791000.0
FlashOcc [42]CSwin-B✓43.5213.3151.6228.0750.9155.6927.4631.0529.9829.2038.8643.6883.8745.6356.3359.0150.6344.56909.1
COTR [21]CSwin-B✓46.214.8553.2535.1950.8357.2535.3634.0633.5437.1438.9944.9784.4648.7357.6061.0851.6146.72840.34
HyDRa [36]C+RR-50-44.40---52.356.3-35.935.10--44.1-------
OCCFusion [23]C+LR-101-46.7911.6547.8132.0757.2757.5131.8040.1147.3533.7445.8150.3578.7937.1744.3653.3663.1863.20-
RadOcc-LC [43]C+LSwin-B✓49.3810.9358.2325.0157.8962.8534.0433.4550.0732.0548.8752.1182.9042.7355.2758.3468.6466.013333
SDG-KLC+LR-50✓50.1612.2657.1223.6958.7762.7434.5536.1950.132.0549.8951.2484.146.0557.261.4569.5665.7883
SDG-FusionC+LR-50✓51.6613.2157.7724.360.3364.2836.2139.4452.3635.8050.9153.6584.5647.4558.0061.6170.6767.65133
", + "image_path": "50c45230dee196d1473535f2ae108df948735bcee0b698bbed7316c4f0c7dc38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 289, + 545, + 496 + ], + "blocks": [ + { + "bbox": [ + 47, + 246, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 246, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 546, + 281 + ], + "type": "text", + "content": "Table 1. 3D Occupancy prediction performance on the Occ3D-nuScenes dataset. We present the mean IoU over categories and the IoUs for different classes. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. In the backbone, R represents ResNet, while Swin stands for Swin Transformer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 289, + 545, + 496 + ], + "lines": [ + { + "bbox": [ + 52, + 289, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 52, + 289, + 545, + 496 + ], + "type": "table", + "html": "
MethodInputBackbonemIoUbarrierbicyclebuscarconst. veh.motorcyclepedestriantraffic conetrailertruckdrive. suf.other flatsidewalkterrainmanmadevegetation
SurroundOcc [35]CR-10120.320.511.628.130.810.715.114.012.014.322.237.223.724.422.714.821.8
OccFormer [44]CR-10120.121.111.328.230.310.615.714.411.214.022.637.322.424.923.515.221.1
C-CONet [32]CR-10118.418.610.026.427.48.615.713.39.710.920.233.020.721.421.814.721.3
FB-Occ [16]CR-10119.620.611.326.929.810.413.613.711.411.520.638.221.524.622.714.821.6
RenderOcc [43]CR-10119.019.711.228.128.29.814.711.811.913.120.133.221.322.622.315.320.9
L-CONet [32]L-17.719.24.015.126.96.23.86.86.014.113.139.719.124.023.925.135.7
FlashOcc* [42]CR-5044.144.211.054.160.526.122.631.315.338.847.180.542.048.253.760.870.0
M-CONet [32]C+LR-10124.724.813.031.634.814.618.020.014.720.026.639.222.826.126.026.037.1
Co-Occ [24]C+LR-10127.128.116.134.037.217.021.620.815.921.928.742.325.429.128.628.238.0
OccFusion [23]C+L+RR-10127.327.119.633.736.221.724.825.316.321.830.039.519.924.926.528.940.4
DAOcc [40]C+LR-5030.530.819.534.038.825.027.729.922.523.231.641.025.929.429.935.243.5
SDG-FusionC+LR-5031.731.215.639.742.422.226.929.722.724.233.245.925.131.833.538.944.5
SDG-KL*C+LR-5050.451.420.959.665.834.133.149.023.445.153.776.544.053.558.265.272.9
SDG-Fusion*C+LR-5052.254.316.061.466.933.233.549.222.247.854.683.949.455.860.270.276.5
", + "image_path": "e2b4e89ec1c3f1e57525ebddbb970582da549092f67bd4f2ca49336cccabdfae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 544, + 267, + 606 + ], + "blocks": [ + { + "bbox": [ + 47, + 498, + 546, + 534 + ], + "lines": [ + { + "bbox": [ + 47, + 498, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 546, + 534 + ], + "type": "text", + "content": "Table 2. 3D Occupancy prediction performance on the SurroundOcc-nuScenes validation set. The best scores for each class are highlighted in bold. In the Input, the C, L, and R denote camera, LiDAR, and radar, respectively. * means the performance is achieved through a visible mask similar to [30]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 544, + 267, + 606 + ], + "lines": [ + { + "bbox": [ + 67, + 544, + 267, + 606 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 267, + 606 + ], + "type": "table", + "html": "
BaselineSDGFOADiou(%)mIoU(%)
✓90.2737.84
✓✓94.6248.51
✓✓94.7644.92
✓✓✓95.3551.66
", + "image_path": "b514b4163d289cdc07a7d9b13f5a0ddf33716985cf50a739b8b98ea453b779cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 655, + 139, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 139, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 139, + 669 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "The Effectiveness of Each Component. The results are shown in Table 3, we can observe that all components make their own performance contributions. The baseline achieves" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 544, + 533, + 632 + ], + "blocks": [ + { + "bbox": [ + 47, + 610, + 286, + 644 + ], + "lines": [ + { + "bbox": [ + 47, + 610, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 286, + 644 + ], + "type": "text", + "content": "Table 3. Ablation study on Occ3D-nuScenes dataset with SDG-Fusion. SDG: Semantic and Depth-Guided View Transformations. FOAD: Fusion-To-Occupancy-Driven Active Distillation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 544, + 533, + 632 + ], + "lines": [ + { + "bbox": [ + 317, + 544, + 533, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 533, + 632 + ], + "type": "table", + "html": "
rlIoU(%)mIoU (%)
1495.3451.15
1895.3551.66
11295.3851.28
2495.3550.4
2895.3951.03
21295.3051.12
", + "image_path": "515ade9e67a1ba6a6ee7591981a8b24c062276275c36c357ad33b24c50357db9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 635, + 545, + 657 + ], + "lines": [ + { + "bbox": [ + 306, + 635, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 545, + 657 + ], + "type": "text", + "content": "Table 4. Ablation study of the hyperparameter used in SDG view transformation module on Occ3D-nuScenes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 665, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 546, + 715 + ], + "type": "text", + "content": "90.27% of IoU and 37.84% of mIoU. We first integrated the Semantic and Depth-Guided (SDG) View Transformation into the baseline model, which brings 4.35% and 10.67% performance gain in IoU and mIoU. Fusion enhanced by in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6757" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 211, + 305 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 211, + 305 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 211, + 305 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 211, + 305 + ], + "type": "image", + "image_path": "77149bf92e1d5104b39b390e8b152bee8a8780182e72f8b3f2238670780b6ec2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 308, + 163, + 319 + ], + "lines": [ + { + "bbox": [ + 115, + 308, + 163, + 319 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 163, + 319 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 72, + 315, + 304 + ], + "blocks": [ + { + "bbox": [ + 211, + 72, + 315, + 304 + ], + "lines": [ + { + "bbox": [ + 211, + 72, + 315, + 304 + ], + "spans": [ + { + "bbox": [ + 211, + 72, + 315, + 304 + ], + "type": "image", + "image_path": "2d5cd69217362d6897d2ce6468fcb0c8cdef1112113a0e868a9632dcd15b4b97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 308, + 280, + 319 + ], + "lines": [ + { + "bbox": [ + 246, + 308, + 280, + 319 + ], + "spans": [ + { + "bbox": [ + 246, + 308, + 280, + 319 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 418, + 304 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 418, + 304 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 418, + 304 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 418, + 304 + ], + "type": "image", + "image_path": "7bb46181bc752c2fe4ae54fec6549e27aeb0b8fda5bda0a95b46b14e0d4031ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 308, + 375, + 318 + ], + "lines": [ + { + "bbox": [ + 359, + 308, + 375, + 318 + ], + "spans": [ + { + "bbox": [ + 359, + 308, + 375, + 318 + ], + "type": "text", + "content": "Our" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 419, + 72, + 523, + 304 + ], + "blocks": [ + { + "bbox": [ + 419, + 72, + 523, + 304 + ], + "lines": [ + { + "bbox": [ + 419, + 72, + 523, + 304 + ], + "spans": [ + { + "bbox": [ + 419, + 72, + 523, + 304 + ], + "type": "image", + "image_path": "be27ec132b49c11470807cb630ea53f89d627881e63346cbe57617ab7d0fa0b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 308, + 498, + 318 + ], + "lines": [ + { + "bbox": [ + 444, + 308, + 498, + 318 + ], + "spans": [ + { + "bbox": [ + 444, + 308, + 498, + 318 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 352, + 268, + 454 + ], + "blocks": [ + { + "bbox": [ + 46, + 321, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 545, + 343 + ], + "type": "text", + "content": "Figure 6. Qualitative results of SDG on the validation set of Occ3D-nuScenes. Each pair of rows displays results from day and low-light scene, respectively. Within each row, images from left to right represent the input images, baseline, our results, and the ground truth." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 352, + 268, + 454 + ], + "lines": [ + { + "bbox": [ + 63, + 352, + 268, + 454 + ], + "spans": [ + { + "bbox": [ + 63, + 352, + 268, + 454 + ], + "type": "image", + "image_path": "1798e593382ecc0aa9af069983464776dec4d2aaf94ceb94321970787a8d64aa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 455, + 286, + 475 + ], + "lines": [ + { + "bbox": [ + 47, + 455, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 286, + 475 + ], + "type": "text", + "content": "Figure 7. Distance-based evaluation on Occ3D-nuScenes. As the distance increases, the point cloud becomes sparse." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 49, + 476, + 282, + 515 + ], + "blocks": [ + { + "bbox": [ + 49, + 476, + 282, + 515 + ], + "lines": [ + { + "bbox": [ + 49, + 476, + 282, + 515 + ], + "spans": [ + { + "bbox": [ + 49, + 476, + 282, + 515 + ], + "type": "table", + "html": "
k3579
IoU (%)95.3295.4095.3595.26
mIoU (%)51.2251.4051.6651.08
", + "image_path": "56ac18d9c12a811c0a9ea3ae09ea7aa92c11f6b55b90e4742dc5d5308b38a110.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 520, + 286, + 553 + ], + "lines": [ + { + "bbox": [ + 46, + 520, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 286, + 553 + ], + "type": "text", + "content": "Table 5. Ablation study of the hyperparameter used in feature fusion of FOAD on Occ3D-nuScenes. The K denotes the size of the pixel region corresponding to the neighborhood feature extraction." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "text", + "content": "tegrating additional LiDAR information, the IoU and mIoU have been significantly improved by " + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "inline_equation", + "content": "4.49\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "inline_equation", + "content": "7.08\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "text", + "content": ", respectively. By using both SDG and Fusion, outperforming the baseline by " + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "inline_equation", + "content": "5.19\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "text", + "content": " of IoU and " + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "inline_equation", + "content": "13.82\\%" + }, + { + "bbox": [ + 46, + 557, + 286, + 605 + ], + "type": "text", + "content": " of mIoU." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "The Effectiveness of SDG View Transformation. To further demonstrate the effect of SDG View Transformation, we conducted hyperparameter analysis experiments. In the SDG view transformation, the range " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " of bilinear growth discretization and the diffusion feature layers " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " control the virtual point generation of SDG. As shown in Table 4, lower depth precision (e.g., " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "r = 2" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "l = 4" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": ") results in slightly reduced performance compared to other configurations. However, excessive depth precision does not lead to" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "content": "additional gains, with the optimal performance observed at " + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "inline_equation", + "content": "l = 8" + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "content": "The Effectiveness of FOAD Module. We perform a hyperparameter analysis of the FOAD module. For neighborhood feature fusion, the parameter " + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "content": " controls the fusion of features from neighboring pixels. As shown in Table 5, increasing " + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "content": " does not consistently improve performance, with optimal results achieved at " + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "inline_equation", + "content": "K = 7" + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 474, + 378, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a multimodal 3D semantic occupancy prediction framework, termed SDG-OCC, designed to achieve higher accuracy and competitive inference speed by fusing LiDAR information in the BEV perspective. To address the inaccurate depth estimation in view transformations, we propose a semantic and depth-guided view transformation method. This approach integrates pixel semantics and corresponding point depth through diffusion and bilinear discretization, effectively reducing invalid image features and significantly enhancing the speed and accuracy of semantic occupancy. Meanwhile, We propose a fusion-to-occupancy-driven active distillation that incorporates multimodal features and selectively transfers multimodal knowledge to image features based on LiDAR-identified regions. Our method achieves the SOTA performance with real-time processing on the Occ3D-nuScenes dataset and comparable performance on the more challenging SurroundOccnuScenes dataset, demonstrating its effectiveness." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6758" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Geonho Bang, Kwangjin Choi, Jisong Kim, Dongsuk Kum, and Jun Won Choi. Radardistill: Boosting radar-based object detection performance via knowledge distillation from lidar features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15491-15500, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 201 + ], + "type": "text", + "content": "[2] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "text", + "content": "[3] Jiahui Fu, Chen Gao, Zitian Wang, Lirong Yang, Xiaofei Wang, Beipeng Mu, and Si Liu. Eliminating cross-modal conflicts in bev space for lidar-camera 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16381-16387, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 302 + ], + "type": "text", + "content": "[4] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6185-6194, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 303, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 346 + ], + "type": "text", + "content": "[5] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 401 + ], + "type": "text", + "content": "[6] Jiawei Hou, Xiaoyan Li, Wenhao Guan, Gang Zhang, Di Feng, Yuheng Du, Xiangyang Xue, and Jian Pu. Fastocc: Accelerating 3d occupancy prediction by fusing the 2d bird's-eye view and perspective view. arXiv preprint arXiv:2403.02710, 2024. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 457 + ], + "type": "text", + "content": "[7] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 491 + ], + "type": "text", + "content": "[8] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 492, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 288, + 546 + ], + "type": "text", + "content": "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[10] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[11] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17545-17555, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Jiale Li, Hang Dai, Hao Han, and Yong Ding. Mseg3d: Multi-modal 3d semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21694-21704, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[13] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 546, + 195 + ], + "type": "text", + "content": "[14] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[15] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "text", + "content": "[16] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "text", + "content": "[17] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. Advances in Neural Information Processing Systems, 35:10421-10434, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 354, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 409 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[19] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "text", + "content": "[20] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 547, + 578 + ], + "type": "text", + "content": "[21] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[22] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: A straightforward and effective multi-sensor fusion framework for 3d occupancy prediction. arXiv preprint arXiv:2403.01644, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[23] Zhenxing Ming, Julie Stephany Berrio, Mao Shan, and Stewart Worrall. Occfusion: Multi-sensor fusion framework for 3d semantic occupancy prediction. IEEE Transactions on Intelligent Vehicles, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[24] Jingyi Pan, Zipeng Wang, and Lin Wang. Co-occ: Coupling explicit feature fusion with volume rendering regularization for multi-modal 3d semantic occupancy prediction. IEEE Robotics and Automation Letters, 2024. 3, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6759" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[25] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoji Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[26] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghang Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[27] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[28] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "text", + "content": "[29] Ziying Song, Lei Yang, Shaoqing Xu, Lin Liu, Dongyang Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Graphbev: Towards robust bev feature alignment for multi-modal 3d object detection. arXiv preprint arXiv:2403.11848, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 387 + ], + "type": "text", + "content": "[30] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[31] Guoqing Wang, Zhongdao Wang, Pin Tang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Occgen: Generative multi-modal 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2404.15014, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 499 + ], + "type": "text", + "content": "[32] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[33] Yuqi Wang, Yuntao Chen, Xingyu Liao, Lue Fan, and Zhaoxiang Zhang. Panoocc: Unified occupancy representation for camera-based 3d panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17158-17168, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "text", + "content": "[34] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[36] Philipp Wolters, Johannes Gilg, Torben Teepe, Fabian Herzog, Anouar Laouichi, Martin Hofmann, and Gerhard Rigoll. Unleashing hydra: Hybrid fusion, depth consistency and radar for unified 3d perception. arXiv preprint arXiv:2403.07746, 2024. 3, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 584 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[37] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "text", + "content": "[38] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 196, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 250 + ], + "type": "text", + "content": "[39] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European Conference on Computer Vision, pages 677-695. Springer, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 251, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 284 + ], + "type": "text", + "content": "[40] Zhen Yang, Yanpeng Dong, and Heng Wang. Daocc: 3d object detection assisted multi-sensor fusion for 3d occupancy prediction. arXiv preprint arXiv:2409.19972, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 285, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 350 + ], + "type": "text", + "content": "[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 353, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 406 + ], + "type": "text", + "content": "[42] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 408, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 408, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 408, + 545, + 462 + ], + "type": "text", + "content": "[43] Haiming Zhang, Xu Yan, Dongfeng Bai, Jiantao Gao, Pan Wang, Bingbing Liu, Shuguang Cui, and Zhen Li. Radocc: Learning cross-modality occupancy knowledge through rendering assisted distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7060-7068, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 464, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 517 + ], + "type": "text", + "content": "[44] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 519, + 545, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 584 + ], + "type": "text", + "content": "[45] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9939-9948, 2021. 1" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6760" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_content_list.json b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..28506c2893e182029e620b3574ae27dc31d3a054 --- /dev/null +++ b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_content_list.json @@ -0,0 +1,1559 @@ +[ + { + "type": "image", + "img_path": "images/6c54110e18f1ff4f83f3abe5fb0be8f6f1a2943486504e70db095fcd940d7e26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 126, + 176, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SEAL: Semiclastic Attention Learning for Long Video Representation", + "text_level": 1, + "bbox": [ + 173, + 137, + 864, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lan Wang $^{1,2*}$ Yujia Chen $^{2}$ Du Tran $^{2}$ Vishnu Naresh Boddeti $^{1}$ Wen-Sheng Chu $^{2}$ $^{1}$ Michigan State University \n $^{2}$ Google", + "bbox": [ + 127, + 185, + 867, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wanglan3, vishnu}@msu.edu {yujiachen, tranldu, wschu}@google.com", + "bbox": [ + 196, + 224, + 794, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long video understanding presents challenges due to the inherent high computational complexity and redundant temporal information. An effective representation for long videos must efficiently process such redundancy while preserving essential contents for downstream tasks. This paper introduces SEmatic Attention Learning (SEAL), a novel unified representation for long videos. To reduce computational complexity, long videos are decomposed into three distinct types of semantic entities: scenes, objects, and actions, allowing models to operate on a compact set of entities rather than a large number of frames or pixels. To further address redundancy, we propose an attention learning module that balances token relevance with diversity, formulated as a subset selection optimization problem. Our representation is versatile and applicable across various long video understanding tasks. Extensive experiments demonstrate that SEAL significantly outperforms state-of-the-art methods in video question answering and temporal grounding tasks across diverse benchmarks, including LVBench, MovieChat-1K, and Ego4D.", + "bbox": [ + 89, + 306, + 485, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 627, + 220, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art video understanding models excel at short video tasks such as video classification [7, 13], temporal grounding [38] and action detection [10, 12], which involve videos lasting from a few seconds to minutes. However, their performance declines on hour-long videos [40, 41]. In contrast, a 10-year-old child can watch a full-length movie (one to two hours) and effortlessly answer questions at various levels of detail. This disparity between humans and machines emphasizes the foundational challenges in long video understanding for AI models, including: (1) Increased complexity: Long videos require more computation and memory than current hardware can support for training or inference, (2) Temporal redundancy: Slow-changing scenes and objects introduce significant redundancy, and (3) Cross-task generalization: A robust repre", + "bbox": [ + 89, + 650, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sentation must adapt to various tasks, from fine-grained fact retrieval to high-level reasoning. These challenges, which appear trivial for a child, remain challenging for AI models.", + "bbox": [ + 511, + 276, + 903, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How does the human brain process long videos, particularly in addressing the above-mentioned challenges? First, rather than processing every pixel or frame, the brain selectively attends to new information to efficiently manage temporal redundancy [4, 22]. Second, humans process videos in an online fashion, continuously updating their understanding and memories as they watch, rather than deferring reasoning until the end. This continuous knowledge update, combined with selective attention, allows the brain to efficiently handle the complexity of long videos. Finally, attention dynamically shifts based on context. Without specific guidance, a child may focus on naturally engaging or memorable moments. However, when given specific questions in advance, attention becomes goal-oriented to seek relevant details while maintaining a broad understanding. This suggests effective representations should balance between task-specific focus and holistic understanding of the video to enable cross-task generalization.", + "bbox": [ + 511, + 324, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by how humans process long videos, we introduce SEmantic Attention Learning (SEAL), a novel unified representation designed to tackle the three key challenges in long video understanding. SEAL consists of two main steps: Semantic Decomposition and Attention Learning. In Semantic Decomposition, long videos are decomposed into three semantic entities such as scenes, objects, and actions, which are then treated as tokens. While the scene and object tokens represent static content assumed not to change rapidly, the action tokens are designed to capture the dynamic, fast-changing moments of the video. We note that these semantic tokens efficiently encode the essential information needed to answer \"where\", \"what\", or \"how\" questions about the videos. This decomposition significantly reduces complexity by allowing AI models to operate on a compact set of tokens instead of raw pixels or frames. Figure 1(a) illustrates a conventional uniformly sampled video $\\mathcal{V}$ , where redundant frames create cluttered visual information that hinders effective analysis for both models and humans. Figure 1(b) shows our semantic decomposition,", + "bbox": [ + 511, + 599, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done as Lan Wang's internship project at Google.", + "bbox": [ + 107, + 886, + 457, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26192", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e376a8c5c37aa7672a656904d522374fcf96f305a9d4174a4544970c9b6bf9ab.jpg", + "image_caption": [ + "(a) Uniform sampling of a long video $\\mathcal{V}$", + "(b) Semantic decomposition of the same video $\\mathcal{V}$" + ], + "image_footnote": [], + "bbox": [ + 151, + 107, + 843, + 199 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb45a9f5db8e5a1015b0c03a9cae93ffbaad42b71a359c6424a3f3af6da7c8ba.jpg", + "image_caption": [ + "Scenes (static)", + "S1" + ], + "image_footnote": [], + "bbox": [ + 133, + 229, + 217, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4430aff250ab55c574d0e11b3d01d8a4104e847aecbc8d82fcf97ebc3acd2dc9.jpg", + "image_caption": [ + "S2" + ], + "image_footnote": [], + "bbox": [ + 225, + 229, + 310, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3b368e06436a23df4c216b91bdbf69267f17f41a19a6de9876ab9a4776386c63.jpg", + "image_caption": [ + "S3" + ], + "image_footnote": [], + "bbox": [ + 318, + 229, + 401, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a11ac0bd5adae8ff7092b5f35c482185a9f1173f3cd6c23cd7e9e4eb6b40d44.jpg", + "image_caption": [ + "S4" + ], + "image_footnote": [], + "bbox": [ + 410, + 229, + 493, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62df9a031fe037ddbe4e763f15cc213a54b1063b2cde4366e5a724f5c8e68a92.jpg", + "image_caption": [ + "Objects (static)", + "01" + ], + "image_footnote": [], + "bbox": [ + 535, + 234, + 568, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cf5411fcd1ed2b2307eadaa0915ffab70e28c25da68192c14fbd5a3e8fb1082e.jpg", + "image_caption": [ + "02" + ], + "image_footnote": [], + "bbox": [ + 576, + 238, + 651, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8bef0394388c1efaf48b92b127da581f826a7d0b0aa4d72843ae52fd29b3dc2a.jpg", + "image_caption": [ + "03" + ], + "image_footnote": [], + "bbox": [ + 656, + 238, + 700, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e89a7bf3469a622daec24e880f4ef50506dd7660a06a8c27a2bdccb92b1765a0.jpg", + "image_caption": [ + "04" + ], + "image_footnote": [], + "bbox": [ + 718, + 234, + 759, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/df85f832aef5638952ac25ce0b3af698934f391557cf6943f0816dc43b776630.jpg", + "image_caption": [ + "05" + ], + "image_footnote": [], + "bbox": [ + 764, + 234, + 797, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6e3b0724e15b656553c253e615f66f79a1e6fb1c9c97a13c256e2ed280c8aad5.jpg", + "image_caption": [ + "06" + ], + "image_footnote": [], + "bbox": [ + 805, + 229, + 870, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0de4353d795382f2a4127a8c6d82f141adc2277343ce3df2743e8009dcda911d.jpg", + "image_caption": [ + "(c) Query-aware attention learning", + "Figure 1. Long Video Representation with Semantic Attention Learning (SEAL): (a) Conventional uniform sampling results in redundant and cluttered visual information, making it difficult for both AI models and human brains to process efficiently. (b) Decomposing long videos into semantic entities such as scenes, objects, and actions reduces temporal redundancy, thus making model training and inference more efficient. In this example, the long video $\\mathcal{V}$ is decomposed into 4 scene tokens (S1-S4), 6 object tokens (O1-O6), and 4 action/event tokens (A1-A4). (c) Query-aware attention learning module improves downstream task performance by focusing on relevant information rather than processing everything. Queries (Q1-Q4) are shown with their most relevant tokens. (best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 117, + 280, + 267, + 353 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/82b5b95f132b0db4daee7c99b1dbd1f6a1e04a55ef4093c071c954bef278b790.jpg", + "image_caption": [ + "A2" + ], + "image_footnote": [], + "bbox": [ + 276, + 294, + 447, + 339 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f326d91f32ce2dc7c6a6a24008362b8a25066b80c776c161e5667f12aa40c04c.jpg", + "image_caption": [ + "A3" + ], + "image_footnote": [], + "bbox": [ + 460, + 284, + 653, + 340 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a375c373de1ff5ddebd5a3d778ffdeec1026a813758fdd46adafd3f8c23f3ed.jpg", + "image_caption": [ + "A4" + ], + "image_footnote": [], + "bbox": [ + 689, + 285, + 866, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q1: What does the vlogger family find in visitor center?", + "bbox": [ + 116, + 371, + 295, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "S101", + "bbox": [ + 117, + 396, + 151, + 405 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q2: What does the vlogger family do after buying fishing supplies?", + "bbox": [ + 305, + 371, + 500, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "S2 S3 O2 A1", + "bbox": [ + 305, + 396, + 377, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q3: What foods do the vlogger family eat the most here?", + "bbox": [ + 504, + 371, + 702, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "O3 O4 O5 A3", + "bbox": [ + 504, + 396, + 589, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q4: What is the most common form of transportation there?", + "bbox": [ + 709, + 371, + 870, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "S4 O6 A2 A4", + "bbox": [ + 709, + 396, + 790, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query set", + "bbox": [ + 818, + 400, + 880, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene token Object token Action token", + "bbox": [ + 116, + 415, + 379, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "breaking down the video into distinct scenes, static objects, and dynamic actions. In Attention Learning, we formulate a subset selection problem that maximizes the query relevance while ensuring token diversity. This step not only mitigates redundancy but also enhances cross-task generalization by prioritizing the most informative tokens. Figure 1(c) illustrates our attention learning module with four different queries and their selected tokens that capture relevance and diverse video content. Finally, SEAL is designed to work for both global and streaming modes, enabling it to process arbitrarily long videos. Extensive ablations and experiments confirm SEAL's superior performance over existing methods. Our key contributions include:", + "bbox": [ + 89, + 531, + 483, + 727 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SEAL, a novel unified representation for long videos by decomposing them into three semantic tokens, namely scenes, objects, and actions.", + "- Our attention learning module reduces temporal redundancy while supporting strong cross-task generalization. We show SEAL works in both global and streaming modes, making it adaptable to arbitrarily long videos.", + "- SEALoutperforms state-of-the-art methods on various long video understanding tasks and benchmarks including: video QA (MovieChat-1K [33], LVBench [40]), and egocentric video grounding (Ego4D [8])." + ], + "bbox": [ + 91, + 731, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 530, + 653, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent approaches have unified various video understanding tasks by framing them as video QA tasks [40], leveraging the capabilities of LLMs. However, fine-tuning with task-specific vision heads continues to offer advantages in memory efficiency and task-specific performance [29], particularly for temporal grounding. In this section, we review advancements in Video Question Answering (QA) and temporal grounding for long video understanding.", + "bbox": [ + 511, + 553, + 906, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video QA for long videos. The main challenge for long video QA is the memory constraint. He et al. [9] introduced a sequential framework that uses a memory bank to enhance long-term comprehension. Song et al. [33, 34] integrated video foundation models with LLMs through a memory mechanism inspired by the Atkinson-Shiffrin model, reducing computational complexity while preserving long-term memory. Another line of work improves efficiency by decomposing video content. Rui et al. [30] employed Memory-Propagated Streaming Encoding to segment videos into short clips, with Adaptive Memory Selection enhancing response accuracy by identifying question-relevant memories. Min et al. [24] introduced a multi-stage, training-free framework, emphasizing task decomposition into parsing, grounding, and reasoning stages. More re", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26193", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cently, Weng et al. [41] proposed a hierarchical framework that encodes local features and integrates global semantics for detailed comprehension of extended video content. To further reduce model's hallucination in QA, Sun et al. [35] proposed a question-guided pipeline by focusing on relevant frames and controlled answer generation.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal localization for long videos. Recent research in temporal localization for long videos has explored two primary directions: LLM-based and non-LLM-based approaches. With the advances in LLMs, researchers have expanded its use beyond traditional VQA, leveraging the capabilities for temporal grounding tasks, with a primary focus on enhancing localization accuracy. Ren et al. [32] proposed a timestamp-aware model that aligns visual content with temporal cues, enabling adaptive processing of sequential events for tasks like localization. Similarly, Fan et al. [5] introduced a memory-enhanced framework that captures contextual information across video segments, improving temporal and spatial reasoning. Korbar et al. [15] developed a text-guided resampling mechanism that dynamically selects video segments, focusing on relevant scenes to enhance temporal and spatial comprehension.", + "bbox": [ + 89, + 181, + 482, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, non-LLM-based approaches typically rely on training a regression layer or decoder for temporal localization, with efforts concentrated on refining visual features and multimodal fusion to improve alignment. For instance, Hou et al. [11] introduced a hierarchical framework that combines coarse scanning with fine-grained alignment to optimize both precision and efficiency in localizing target moments. Pan et al. [28] applied a coarse-to-fine pipeline for single-pass temporal grounding that improves both efficiency and alignment. Additionally, Mu et al. [26] proposed a cost-effective late fusion approach paired with a video-centric sampling scheme to improve scalability.", + "bbox": [ + 89, + 424, + 482, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike prior works that focus on specific tasks, SEAL proposes a unified and generic framework for long video understanding, capable of adapting to various prediction heads and tasks. Additionally, SEAL supports both global and streaming modes, making it adaptable to arbitrarily long videos.", + "bbox": [ + 89, + 604, + 482, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Semantic Decomposition and Attention Learning", + "text_level": 1, + "bbox": [ + 89, + 705, + 482, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{V} = \\{v_{i}\\}_{i = 1}^{T_{V}}$ be an arbitrarily long untrimmed video, where $v_{1},\\ldots ,v_{T_{V}}$ denote the sequence of $T_{V}$ frames forming the video. Let $q$ be a query from long video understanding tasks, comprising a sequence of $l_{q}$ tokens. The query $q$ may take different forms depending on the task, such as natural language text for video question and answering (e.g., MovieChatQA [33], LVBench [40]) or visual/text template or action label for episodic memory tasks in egocentric video understanding (e.g., Ego4D [8]). Our method establishes a unified video representation to gener", + "bbox": [ + 89, + 748, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "alize across these diverse long video understanding tasks. An overview of our approach is presented in Figure 2.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Semantic Decomposition of Long Videos", + "text_level": 1, + "bbox": [ + 511, + 133, + 857, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main challenge in long video representation lies in capturing diverse content within limited memory. Conventional methods resort to frame sampling [42, 44] or maintaining a memory bank that merges similar frames [33, 34]. However, these approaches can vary greatly across tasks. We propose a novel decomposition approach that structures long videos into three distinct token types representing different type of semantic entities: (1) Scene tokens $\\mathbf{T}_{\\mathrm{scene}}$ capture background context, providing essential cues about the environment. (2) Action tokens $\\mathbf{T}_{\\mathrm{action}}$ represent moving elements, focusing on temporal information such as motions, activities, or events. (3) Object tokens $\\mathbf{T}_{\\mathrm{object}}$ , highlight key static elements relevant to specific tasks. With this structured tokenization, we create a unified, compact, task-agnostic representation that minimizes the need for redundant and dense video storage while preserving a comprehensive understanding of long video content.", + "bbox": [ + 511, + 157, + 903, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Tokens. Any frame in a video can serve as a scene token because it captures the environment where it was recorded, e.g., indoor gym, outdoor mountain, etc. Although shot boundary detection could split a long video into shots with one scene token per shot, this approach has two drawbacks: (1) Shot detection algorithms are often imperfect, with no opportunity to correct once shot boundary is determined. (2) Shot detection often ignores the specific query and downstream tasks, and thus could cause suboptimal performance. To overcome these drawbacks, we propose a two-step approach. First, we over-sample scene tokens, and then later perform an attention learning step to maximize query relevance and token diversity. Specifically, we uniformly sample $N_{\\mathrm{scene}}$ frames to capture a diverse background. These pre-sampled scene tokens, denoted as $\\mathbf{T}_{\\mathrm{scene}} = [t_i^{\\mathrm{scene}}]_{i=1}^{N_{\\mathrm{scene}}}$ , will undergo another round of attention learning, detailed in Section 3.2.", + "bbox": [ + 511, + 415, + 903, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action Tokens. The purpose of the action token is to capture temporal information of moving objects such as low-level motions, activities, and events. We begin by using a class-agnostic object tracker, e.g., SAM-2 [31], to extract multiple initial dynamic tracklets. Tracklets shorter than $L_{\\mathrm{min}}$ are discarded, while those longer than $L_{\\mathrm{max}}$ are split into multiple tracklets with length of $L_{\\mathrm{max}}$ . For each tracklet, we take the spatial union of bounding boxes across frames, allowing the dynamic token to capture not only motion information but also the spatial movement of people and objects. This process yields $N_{\\mathrm{tracklet}}$ tracklets, denoted as $\\mathbf{T}_{\\mathrm{action}} = \\{\\tau_{\\mathrm{dynamic}}^{i}\\}_{i=1}^{N_{\\mathrm{tracklet}}}$ .", + "bbox": [ + 511, + 672, + 903, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Tokens. For the object tokens, we utilize a class-agnostic grouping method such as SAM [14] to generate masks for all objects in each frame. This class-agnostic", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26194", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f743335994fba06e3aff28542a42bd82535cbf3c548349218c6fe930b9e2ef49.jpg", + "image_caption": [ + "Figure 2. SEAL Overview. During semantic decomposition, a long video $\\mathcal{V}$ is decomposed into semantic tokens representing scenes, objects, and actions. Then, during attention learning, these tokens and the query q, are optimized for query relevance $R(\\cdot)$ and token diversity $S(\\cdot)$ . The resulting attended token subset is then passed to a vision or an MLLM head for predictions." + ], + "image_footnote": [], + "bbox": [ + 130, + 95, + 851, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "segmentation approach enables comprehensive object information capture. Specifically, we apply this process on $N_{\\mathrm{key}}$ key frames. On each of the $N_{\\mathrm{key}}$ frames, we apply SAM to obtain in total $N_{\\mathrm{object}}$ object masks, which we denote as $\\mathbf{T}_{\\mathrm{object}} = \\{\\mathbf{M}_{\\mathrm{object}}^i\\}_{i=1}^{N_{\\mathrm{object}}}$ . Key frame selection varies by task and is detailed in the following sections.", + "bbox": [ + 89, + 449, + 483, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Attention Learning", + "text_level": 1, + "bbox": [ + 89, + 551, + 276, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the long video is decomposed into different token representations, the resulting tokens are still redundant. To address this, we propose a sampling approach that balances between query-relevance and token-diversity, denoted as Attention Learning. Specifically, we formulate our sampling as an optimization problem with two main objectives: query-relevance and token-diversity.", + "bbox": [ + 89, + 574, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T_{s}^{*} = \\operatorname *{arg max}_{T_{s}\\subset T_{G}}F_{s}(T_{s}|T_{G},q) \\\\ = \\underset {T _ {s} \\subset T _ {G}} {\\arg \\max} \\alpha \\sum_ {t _ {s} \\in T _ {s}} R (t _ {s}, q) + (1 - \\alpha) \\sum_ {t _ {i}, t _ {j} \\in T _ {s}, i \\neq j} \\frac {1}{S (t _ {i} , t _ {j})}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 688, + 478, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to solve a subset selection problem: find a fixed-size subset $T_{s} \\subset T_{G}$ that maximizes the objective function $F_{s}(T_{s}|T_{G},q)$ . Here, $T_{G}$ represents the set of all tokens used, which can be any set of tokens such as scene $\\mathbf{T}_{\\mathrm{scene}}$ , object $\\mathbf{T}_{\\mathrm{object}}$ , action $\\mathbf{T}_{\\mathrm{action}}$ , or a combination of all $(\\mathbf{T}_{\\mathrm{scene}} \\cup \\mathbf{T}_{\\mathrm{object}} \\cup \\mathbf{T}_{\\mathrm{action}})$ . $q$ denotes the query being asked by the downstream understanding task. And, $T_{s}^{*}$ denotes the optimal subset that maximizes $F_{s}$ . $F_{s}$ is decomposed into two terms: $R(\\cdot)$ measures the relevance between a vi", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sual token and the query which we compute by encoding them with the BLIP-2 model [17] and calculating their cosine similarity; and $S(\\cdot)$ calculates the cosine similarity of the paired tokens. We note that the first term maximizes the relevance between selected tokens and the query while the second term enforces token diversity (via minimizing the token pairwise similarity). Finally, $\\alpha$ is a hyper-parameter to balance between query-relevance and token-diversity.", + "bbox": [ + 511, + 449, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Streaming and Global Mode", + "text_level": 1, + "bbox": [ + 511, + 575, + 769, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed method offers two ways of representing long videos: streaming and global. In the global mode, the model fully \"watches\" (or processes) the entire video and then provides a single representation. In contrast, in the streaming mode, the model processes the video buffer by buffer and provides an updated representation at any given time step. The partly-observed video representation can change over time as the video progresses. This setting simulates the situation when you watch a movie together with a child and have interactions with her or him as the movie is still going on.", + "bbox": [ + 511, + 598, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the global model, all tokens undergo a single optimization through Attention Leaning, resulting in $T_{\\mathrm{sub}}$ , which includes $k$ tokens. This $T_{\\mathrm{sub}}$ is regarded as the representation for the entire video. We note that the global mode may not scale well with arbitrary long videos because all tokens cannot be fit into a limited memory for sampling. One can opt to use more aggressive uniform temporal presampling to reduce the numbers of tokens before Attention Learning. This workaround can bypass the memory lim", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26195", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "itation, but may also lead to sub-optimal solutions due to missing important tokens due to uniform sampling.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alternatively, we propose an online streaming approach for representing partly-observed videos. Specifically, we use a fixed-size sliding window with a size set to the maximum number of tokens $l$ allowed by memory capacity, denoted as $T$ . At each step $t$ , we apply Attention Learning to the union set of the tokens in the current window $T_{t}$ and the previous selected subset of tokens $T_{\\mathrm{sub}}^{t - 1}$ and obtain the representation of the video at time $T_{\\mathrm{sub}}^{t}$ . At the beginning, the selected subset is set to empty $(T_{\\mathrm{sub}}^{0} = \\emptyset)$ .", + "bbox": [ + 89, + 121, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {\\text {s u b}} ^ {t} = \\text {A t t e n t i o n L e a r n i n g} \\left(T _ {t} \\cup T _ {\\text {s u b}} ^ {t - 1}\\right) \\forall t > 0. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 268, + 482, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This streaming mode allows us to use $T_{\\mathrm{sub}}^{t}$ as the partly-observed representation of the video and can be fed into any prediction head for video understanding tasks. As an immediate benefit of the streaming mode, our proposed representation now can handle arbitrary long videos.", + "bbox": [ + 89, + 300, + 483, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prediction Heads", + "text_level": 1, + "bbox": [ + 89, + 386, + 261, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our unified representation is adaptable to most long video understanding tasks using different prediction heads. In this paper, we demonstrate two specific use cases of our representation: one is used with the traditional vision head for video temporal grounding and the other one is with the Multimodal LLMs (MLLM) head for video QA.", + "bbox": [ + 89, + 409, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Grounding with Vision Heads. Given a query $q$ , the task is to locate the start and end times, $t_{\\mathrm{start}}$ and $t_{\\mathrm{end}}$ , where the answers could be deduced. We first encode the sampled tokens $T_{\\mathrm{sub}}$ and the query using encoders $\\mathbb{E}_V$ and $\\mathbb{E}_q$ to obtain embeddings for each video token $z_v^i = \\mathbb{E}_V(T_{\\mathrm{sub}^i})$ and query $z_q = \\mathbb{E}_q(q)$ . The cross-modal fusion is then performed to obtain the fused representation:", + "bbox": [ + 89, + 500, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {j o i n t}} ^ {i} = \\operatorname {C r o s s M o d a l F u s i o n} \\left(z _ {v} ^ {i}, z _ {q}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 621, + 482, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, a moment decoder is applied to predict the start and end time for the query $q$ . The moment decoder includes a classification head and a regression head. The classification head is used to predict the score $p_{score}^{i}$ of each token, while the regression head predicts the normalized distances $(d_{\\mathrm{start}}^{i}, d_{\\mathrm{end}}^{i})$ from each token to the moment boundaries:", + "bbox": [ + 89, + 646, + 483, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\text {s c o r e}} ^ {i}, \\left(d _ {\\text {s t a r t}} ^ {i}, d _ {\\text {e n d}} ^ {i}\\right) = \\text {M o m e n t D e c o d e r} \\left(z _ {\\text {j o i n t}} ^ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 763, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The regression and classification heads are optimized using an IoU distance and a focal loss as used in [25]. The final moment is calculated as $t_{start}^{i}, t_{end}^{i} = (t_{i} - d_{start}^{i}) \\times L_{V}, (t_{i} + d_{end}^{i}) \\times L_{V}$ , where $L_{V}$ is the length of the input video. The proposed unified representation enables efficient handling of long video sequences and allows for localization related tasks.", + "bbox": [ + 89, + 794, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video QA using MLLM Heads. The proposed representation can also be connected to an MLLM head, making it applicable to various video QA-related tasks, such as reasoning, understanding, and summarization. Moreover, some grounding tasks can also be addressed in a QA format. Specifically, $z_v$ is projected through an MLP to a visual token $T_v^{\\mathrm{MLLM}}$ that the MLLM can interpret, which is then input into the MLLM along with text tokens $T_{text}$ . Based on different benchmarks, the MLLM performs multiple-choice or open-ended answering.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 258, + 645, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 282, + 730, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. We evaluate SEAL on three datasets, each selected for its relevance to long video understanding on different capabilities. LVBench [40] contains 1,549 QA pairs across six tasks, with videos averaging 4,101 seconds (approximate 1 hour 8 minutes). Each question presents a single-choice format with four options. Accuracy serves as the evaluation metric for individual tasks as well as overall performance across all tasks. We primarily focus on this dataset due to its emphasis on hour-long videos. Moviechat-1K [33] includes 1,000 video clips with dense captions spanning 15 categories, averaging 564 seconds (about 10 minutes). The benchmark employs LLMs, specifically GPT-3.5 [27], to evaluate the quality of generated answers. A rating ranging from 0 to 5 is used to compute the overall score, while a binary preference from the LLM is used to calculate the accuracy. Ego4D-NLQ [8] is a part of the Ego4D Episodic Memory challenge for Natural Language Queries (NLQ) task. This dataset requires localizing a temporal window where the answers can be deduced from untrimmed egocentric videos. It contains 1,259 videos, averaging 10 minutes each. Our experiments apply memory constraints to simulate long video scenarios. Metrics include Top-1 and Top-5 recall at various thresholds.", + "bbox": [ + 511, + 309, + 906, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment setup. We fine-tune the projection layers and Q-former [17] to adapt to different types of tokens for 20 epochs using the training split of MovieChat-1K. For Ego4D-NLQ, we use the training split to finetune the vision heads for 7 epochs. The AdamW optimizer [21] is employed with default beta values of (0.9, 0.999) and a weight decay of 0.05. For token extraction, we use SAM2 [31] to obtain object tokens, YOLOv10-X [37] with BoT-SORT [1] for action tokens. Scene, action, and object tokens are extracted at 8, 10, 1 FPS for MovieChat-1K, Ego4D-NLQ and LVbench. For the LVBench, we follow the settings of InternVL2 and utilize Yi-34B [43]. For MovieChat-1K, we adopt the same settings as [33] and use Vicuna-7B as [33]. For Ego4D-NLQ, we adhere to the settings specified in SnAG [25] and use EgoVLP [19] as vision encoder. We use 0.9 as the default value for the only hyper-parameter $\\alpha$ .", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26196", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8943efc0c357d83b21c5a8a9d9fa5ac61b8eb189b6da2e871bb197e1117c7f65.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLLM SizeOverall (%)KIR (%)EU (%)Sum (%)ER (%)Rea (%)TG (%)
Qwen2-VL [39]72B41.338.341.146.638.046.541.4
InternVL2 [3]34B39.643.439.741.437.442.531.4
LLaVA-NeXT [45]34B32.234.131.227.630.135.031.4
Oryx [20]34B30.432.129.227.630.134.029.1
PLLaVA [42]34B26.126.224.925.925.030.021.4
SEAL (Ours)34B45.951.541.339.747.943.332.3
", + "bbox": [ + 125, + 88, + 867, + 213 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/68923e8b7db31f54e20cca12735062c24b57276770f4354cddb0b7895d1895a4.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art on LVBench. Our method achieves the highest overall score, with notable gains in Key Information Retrieval (KIR) and Entity Recognition (ER), demonstrating the effectiveness of our representation in locating key information and entities by eliminating redundancy. The best methods are highlighted in bold, and the second-best are underlined." + ], + "table_footnote": [], + "table_body": "
Model#TokenR@1R@5
0.30.5Avg0.30.5Avg
SnAGall15.7210.7813.2538.3927.4432.92
SnAG45013.449.2311.3434.0223.0428.53
SEAL45013.789.2611.5234.7923.1028.95
SnAG20010.036.358.1926.5616.9021.73
SEAL20010.837.068.9527.3917.4122.40
", + "bbox": [ + 91, + 271, + 485, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 482, + 390, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct quantitative analysis on LVBench dataset which contains videos on lengths averaged more than 1 hour. Additionally, we demonstrate the generalizability of the proposed unified representation using MovieChat-1K, which has open-ended QA questions on a variety of scenarios, and Ego4D-NLQ, where we use traditional vision decoders (LLM-independent) for the temporal grounding task. LVBench. Table 1 shows the accuracy of our method on different categories of LVBench dataset [40], demonstrating that SEAL achieves the highest overall score, notably outperforming even larger models like Qwen2-VL72B [39] by $4.6\\%$ . Our method excels in Key Information Retrieval (KIR) and Entity Recognition (ER), outperforming the strongest alternatives by $8.1\\%$ and $5.1\\%$ respectively. These results highlight that our action and object tokens effectively locate key information and entities by removing redundancies. While larger LLMs often exhibit better performance, this analysis suggests that it is not the only determinant, and a unified representation like ours can achieve state-of-the-art results with fewer parameters.", + "bbox": [ + 88, + 505, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 presents a qualitative analysis of SEAL on LVBench, demonstrating that our approach is able to pay attention to relevant semantic tokens and make correct answers to different types of questions. SEAL accurately locates relevant tokens, e.g., identifying the royal family's stool color (Q1.a), counting meals eaten (Q2.a), or deter", + "bbox": [ + 88, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/59040e26d7681e787c1d683a1514124ce6ef53b77813f7eee398254ca278d899.jpg", + "table_caption": [ + "Table 2. Comparisons with SoTA methods on Ego4D-NLQ. Quantitative results for temporal grounding on Ego4D episodic memory Natural Language Queries (NLQ) task show our method, SEAL, consistently outperforms SnAG [25] in all metrics under memory constraints with varying number of tokens." + ], + "table_footnote": [], + "table_body": "
MethodsAccuracy(%)Score
Zero-shotVideoChat [18]61.03.34
VideoLLaMA [44]51.43.10
Video-ChatGPT [23]44.22.71
MovieChat [33]67.83.81
SupervisedTimeChat-Hal [35]73.83.58
HERMES [6]84.94.40
SEAL (Ours)86.84.35
", + "bbox": [ + 514, + 271, + 908, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with SoTA methods on MovieChat-1K. SEAL outperforms the second best method, HERMES, by $1.9\\%$ on accuracy while being comparable on the score metric.", + "bbox": [ + 511, + 406, + 906, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mining scene or location (Q2.b). We also show a failure case (Q2.c), where the nuanced \"why\" question requires complicated causal relations of different scenes.", + "bbox": [ + 511, + 462, + 903, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moviechat-1K. Table 3 compares our method with state-of-the-art methods on MovieChat-1K. We focus on evaluating performance in the global mode, which assesses the model's ability to comprehend information from the entire video, rather than the breakpoint mode, which primarily evaluates its ability to answer questions related to specific timestamps. Our method surpasses all the alternatives, achieving the highest accuracy and a strong score. Notably, unlike LVBench, which employs multi-choice questions, this dataset evaluates generated answers using an off-the-shelf LLM. This highlights the ability of our learned representations to effectively transfer to downstream generative tasks, enabling generating accurate and detailed responses.", + "bbox": [ + 511, + 508, + 903, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ego4D-NLQ. We also evaluate our approach on the Ego4D-NLQ task, which differs from previous datasets as it uses a decoder to locate events for temporal grounding, rather than relying on an LLM to generate text-based answers. Table 2 shows the results under memory-constrained conditions, where we limit token numbers to simulate real-world memory limitations in long video understanding. Our method consistently achieves the highest recalls across various thresholds. With further memory constraints (e.g., reducing tokens to 200), SEAL outperforms the current SoTA method, SnAG [25], by a even larger margin. This demonstrates the robustness of our unified representation on various downstream applications for long video understanding.", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26197", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc3d656346e45dbdda0e6e2f7bdef1786a9aa941c14e18f539ecc60ab7a3dff9.jpg", + "image_caption": [ + "Figure 3. Qualitative results on LVBench. Two long videos visualized with questions, multiple choice options, and SEAL predicted answers. SEAL attends to relevant entities such as \"royal family\" and \"stool\" (Q1.a), different \"meals\" and \"drinks\" (Q2.a), \"scene\" and \"location\" (Q2.b) and correctly answers these questions. Although attending to relevant \"push-up\" activity (Q2.c), SEAL fails to predict the right answer due to the challenging in the causal reasoning question." + ], + "image_footnote": [], + "bbox": [ + 156, + 89, + 831, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 633, + 243, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic decomposition. We analyze the impact of the three proposed token types on LVBench (Table 4). To prevent LLMs from leveraging prior knowledge to \"guess\" answers using only the text query, we present the baseline that uses random token inputs, isolating the LLM's performance. Results indicate that using each token type individually significantly improves performance, highlighting the value of each token's contribution. Notably, scene tokens provide the most substantial improvement, aligning with prior research [2, 36, 41] that emphasizes frame-based scene tokens. Furthermore, different token combinations yield additional gains: adding action tokens enhances Key Information Retrieval (KIR), Reasoning (Rea) and Temporal Grounding (TG) tasks by capturing temporal dynamics in the queries, while object tokens boost Entity Recognition (ER) performance by retaining detailed object-specific", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "information. We observe high variance in the Summarization (Sum) accuracy due to the small number of questions in this category and the randomness of LLM. Similar improvements have been observed on the MovieChat-1K in Table 5, where the accuracy improved by $10.45\\%$ with action and object tokens. Ultimately, using all three token types together achieves the highest performance, demonstrating the complementary strengths of scene, action, and object tokens as a unified representation in long video understanding.", + "bbox": [ + 511, + 633, + 906, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that the latency bottleneck mostly comes from the tracker when processing multiple objects. Table 6 presents an analysis of different trackers for extracting action tokens. While YOLO+BoT-SORT operates in a much higher frame rate (14 FPS), it lacks open-set detection capabilities. On the other hand, SAM2 can detect all objects within a scene but operates significantly slower (8FPS). Our findings indicate that the proposed action tokens does not", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26198", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ed12d805212c17721298f170ccb13037fc1f6fb084a0b6defb7cab7185b24a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelOverallKIREUSumERReaTG
Random tokens24.824.725.234.523.526.420.9
Action only34.033.331.736.234.335.833.6
Object only33.934.432.631.033.440.829.5
Scene only42.850.540.625.944.341.327.2
Scene+Object43.447.440.024.147.642.335.5
Scene+Action44.453.340.024.146.243.338.2
SEAL (Ours)45.951.541.339.747.943.332.3
", + "bbox": [ + 91, + 88, + 480, + 208 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa64c16d4673f39a1bc1e062087a2fb963986f5e4aed9b0b349d6bdee0b434ec.jpg", + "table_caption": [ + "Table 4. The effects of different types of tokens on LVbench. Using any type of tokens outperforms the random baseline, while applying all three types of tokens brings the best performance." + ], + "table_footnote": [], + "table_body": "
MethodSceneScene+ObjectScene+ActionSEAL
Accuracy76.3379.4981.4686.78
Score4.154.264.284.35
", + "bbox": [ + 91, + 263, + 478, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "require capturing all the objects in the scene. In practice, YOLO-X pretrained on COCO, with BoT-SORT, effectively captures essential information for action token extraction, maintaining a balance between performance and efficiency. Attention learning. Table 6 highlights the effectiveness of query relevance and token diversity terms in Section 3.2. When $\\alpha = 0$ , the optimization focuses on the diversity term to capture extensive contextual information for long-video tasks such as Summarization (Sum) and Reasoning (Rea), achieving reasonable results comparable to other state-of-the-art approaches in Table 1. With $\\alpha = 1$ , the optimization prioritizes the relevance term to reduce redundancy by sampling tokens that closely align with the query, so performance can be improved for tasks like Key Information Retrieval (KIR). However, this approach loses global context and results in weaker summarization performance. Balancing both terms yields the best overall results.", + "bbox": [ + 89, + 383, + 482, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Streaming mode. We compare the global mode and streaming mode of our method on LVBench. As discussed in Section §3, the streaming mode better simulates real-world scenarios by sequentially processing partial observations to aggregate information over arbitrarily long videos. Table 6 demonstrates that streaming mode performs worse on globally-dependent tasks such as Event Understanding (EU) and Key Information Retrieval (KIR). However, it excels in temporally-intensive tasks like Temporal Grounding (TG) and Reasoning (Rea). Notably, streaming mode still surpasses the most competitive baseline, even when the baseline uses a significantly larger LLM.", + "bbox": [ + 89, + 642, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Complexity. We provide an accuracy-efficiency trade-off comparison in Fig 4 where both SEAL and InternVL2 are with a varying number of tokens. FPS is calculated during inference that includes all the processing time from the raw video frames. SEAL reduces the re", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8f9f01ea6434155313f54c5aa0f7edf3e851a86379b7813417e5a3741e1697b.jpg", + "table_caption": [ + "Table 5. The effects of different types of tokens on MovieChat1K. Results are on the test set with our global inference mode. Using all three types of tokens provides the best accuracy." + ], + "table_footnote": [], + "table_body": "
ModelOverallKIREUSumERReaTG
SEAL (Ours)45.951.541.339.747.943.332.3
rep Yolo w/ SAM243.351.939.929.346.440.332.7
⇒ Streaming44.250.939.737.945.044.334.1
Diversity only38.740.438.840.038.740.933.7
Relevance only42.048.539.927.644.037.331.8
", + "bbox": [ + 514, + 88, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies on LVBench. SEAL defaults to global mode with YOLO for efficiency, while SAM2 performs similarly. Global mode excels in long-context tasks, and streaming mode in temporally-intensive tasks. Combining Relevance and Diversity in attention learning achieves the best performance.", + "bbox": [ + 511, + 183, + 903, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2de949438a1e2de40a862c4683a57e79b78dac320b3a53ac7e48bc57c04eacd1.jpg", + "image_caption": [ + "Figure 4. Accuracy vs. efficiency trade-off on LVBench. SEAL runs $2 - 3\\mathrm{x}$ faster than InternVL2 at the same accuracy, and is more accurate when compared at the same FPS." + ], + "image_footnote": [], + "bbox": [ + 519, + 263, + 895, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "liaence on densely sampled tokens and uses subject-level and motion aware representations/tokens to improve efficiency while maintain good accuracy. From Fig 4, at 10 FPS, SEAL is about $5\\%$ more accurate than InternVL2. When comparing at the same accuracy of $41.5\\%$ and $40\\%$ , SEAL is about 3x and 2x faster than InternVL2, respectively.", + "bbox": [ + 511, + 494, + 905, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 601, + 782, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose SEAL, a novel unified representation for long video understanding that addresses computational complexity, temporal redundancy, and cross-task generalization. SEAL leverages semantic decomposition to break videos into scene, object, and action tokens, reducing redundancy and enabling efficient processing. It incorporates attention learning to balance query relevance and token diversity, enhancing performance across diverse tasks. SEAL achieves state-of-the-art results on benchmarks like MovieChat-1K, LVBench, and Ego4D-NLQ, demonstrating its versatility and effectiveness for long video understanding.", + "bbox": [ + 511, + 627, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. The Attention Learning module is bounded by the memory constraint for the QP solver, making it not fully end-to-end trainable. For future work, we plan to integrate Attention Learning in our streaming mode for full end-to-end learning. Developing special prediction heads to solve causal reasoning [16], where MLLM heads showed limitations, is also an interesting area for future exploration.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26199", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by the Office of Naval Research (award #N00014-23-1-2417). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of ONR.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 179, + 187, + 194 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. ArXiv, abs/2206.14651, 2022. 5", + "[2] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In CVPR, 2022. 7", + "[3] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6", + "[4] Meghomala Das, David M Bennett, and Gordon N Dutton. Visual attention as an important visual function: an outline of manifestations, diagnosis and management of impaired visual attention. British Journal of Ophthalmology, 91(11): 1556-1560, 2007. 1", + "[5] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In ECCV, pages 75-92. Springer, 2024. 3", + "[6] Gueter Josmy Faure, Jia-Fong Yeh, Min-Hung Chen, Hung-Ting Su, Shang-Hong Lai, and Winston H. Hsu. Hermes: temporal-coherent long-form understanding with episodes and semantics. arXiv preprint arXiv:2408.17443, 2024. 6", + "[7] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 1", + "[8] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonio Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawa" + ], + "bbox": [ + 99, + 204, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "har, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 5", + "[9] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, 2024. 2", + "[10] Fabian Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1", + "[11] Zhijian Hou, Wanjun Zhong, Lei Ji, Difei Gao, Kun Yan, Wing-Kwong Chan, Chong-Wah Ngo, Zheng Shou, and Nan Duan. Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918, 2022.3", + "[12] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos \"in the wild\". CoRR, abs/1604.06182, 2016. 1", + "[13] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. CoRR, abs/1705.06950, 2017. 1", + "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 3", + "[15] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In ECCV, pages 271-288. Springer, 2024. 3", + "[16] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 8", + "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML. PMLR, 2023. 4, 5", + "[18] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6", + "[19] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022.5", + "[20] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 6", + "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26200", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Steven J. Luck and Michelle Ford. On the role of selective attention in visual perception. Proceedings of the National Academy of Sciences of the United States of America, 95(3): 825-830, 1998. 1", + "[23] Muhammad Maaz, Hanoona Abdul Rasheed, Salman H. Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In AMACL, 2023. 6", + "[24] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In CVPR, pages 13235-13245, 2024. 2", + "[25] Fangzhou Mu, Sicheng Mo, and Yin Li. SnAG: Scalable and accurate video grounding. In CVPR, 2024. 5, 6", + "[26] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 3", + "[27] OpenAI. Gpt-3.5, 2023. 5", + "[28] Yulin Pan, Xiangteng He, Biao Gong, Yiliang Lv, Yujun Shen, Yuxin Peng, and Deli Zhao. Scanning only once: An end-to-end framework for fast temporal grounding in long videos. In ICCV, pages 13767-13777, 2023. 3", + "[29] Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. Egovieo: Exploring egocentric foundation model and downstream adaptation. arXiv preprint arXiv:2406.18070, 2024. 2", + "[30] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024. 2", + "[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos, 2024. 3, 5", + "[32] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 3", + "[33] E. Song, W. Chai, G. Wang, Y. Zhang, H. Zhou, F. Wu, X. Guo, T. Ye, Y. Lu, JN. Hwang, and G. Wang. MovieChat: From dense token to sparse memory for long video understanding. In CVPR, 2024. 2, 3, 5, 6", + "[34] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 2, 3", + "[35] Yiwei Sun, Zhihang Liu, Chuanbin Liu, Bowei Pu, Zhihan Zhang, and Hongtao Xie. Hallucination mitigation prompts long-term video understanding. arXiv preprint arXiv:2406.11333, 2024. 3, 6", + "[36] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In CVPR, 2024. 7" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection. ArXiv, abs/2405.14458, 2024. 5", + "[38] Lan Wang, Gaurav Mittal, Sandra Sajeev, Ye Yu, Matthew Hall, Vishnu Naresh Boddeti, and Mei Chen. Protege: Untrimmed pretraining for video temporal grounding by video temporal grounding. In CVPR, 2023. 1", + "[39] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6", + "[40] W. Wang, Z. He, W. Hong, Y. Cheng, X. Zhang, J. Qi, S. Huang, B. Xu, Y. Dong, M. Ding, and J. Tang. LVBench: An extreme long video understanding benchmark, 2024. 1, 2, 3, 5, 6", + "[41] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 1, 3, 7", + "[42] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6", + "[43] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, Xiaohui Hu, Xiaoyi Ren, Xinyao Niu, Pengcheng Nie, Yuchi Xu, Yudong Liu, Yue Wang, Yuxuan Cai, Zhenyu Gu, Zhiyuan Liu, and Zonghong Dai. Yi: Open foundation models by 01.ai. ArXiv, abs/2403.04652, 2024. 5", + "[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 3, 6", + "[45] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26201", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_model.json b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32fe86b335842e0b011e029666a51d2b8396b07f --- /dev/null +++ b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.127, + 0.177, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.138, + 0.865, + 0.16 + ], + "angle": 0, + "content": "SEAL: Semiclastic Attention Learning for Long Video Representation" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.186, + 0.868, + 0.224 + ], + "angle": 0, + "content": "Lan Wang\\(^{1,2*}\\) Yujia Chen\\(^{2}\\) Du Tran\\(^{2}\\) Vishnu Naresh Boddeti\\(^{1}\\) Wen-Sheng Chu\\(^{2}\\) \n\\(^{1}\\) Michigan State University \n\\(^{2}\\) Google" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.226, + 0.795, + 0.24 + ], + "angle": 0, + "content": "{wanglan3, vishnu}@msu.edu {yujiachen, tranldu, wschu}@google.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.275, + 0.327, + 0.291 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.486, + 0.61 + ], + "angle": 0, + "content": "Long video understanding presents challenges due to the inherent high computational complexity and redundant temporal information. An effective representation for long videos must efficiently process such redundancy while preserving essential contents for downstream tasks. This paper introduces SEmatic Attention Learning (SEAL), a novel unified representation for long videos. To reduce computational complexity, long videos are decomposed into three distinct types of semantic entities: scenes, objects, and actions, allowing models to operate on a compact set of entities rather than a large number of frames or pixels. To further address redundancy, we propose an attention learning module that balances token relevance with diversity, formulated as a subset selection optimization problem. Our representation is versatile and applicable across various long video understanding tasks. Extensive experiments demonstrate that SEAL significantly outperforms state-of-the-art methods in video question answering and temporal grounding tasks across diverse benchmarks, including LVBench, MovieChat-1K, and Ego4D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.222, + 0.644 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.879 + ], + "angle": 0, + "content": "State-of-the-art video understanding models excel at short video tasks such as video classification [7, 13], temporal grounding [38] and action detection [10, 12], which involve videos lasting from a few seconds to minutes. However, their performance declines on hour-long videos [40, 41]. In contrast, a 10-year-old child can watch a full-length movie (one to two hours) and effortlessly answer questions at various levels of detail. This disparity between humans and machines emphasizes the foundational challenges in long video understanding for AI models, including: (1) Increased complexity: Long videos require more computation and memory than current hardware can support for training or inference, (2) Temporal redundancy: Slow-changing scenes and objects introduce significant redundancy, and (3) Cross-task generalization: A robust repre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.323 + ], + "angle": 0, + "content": "sentation must adapt to various tasks, from fine-grained fact retrieval to high-level reasoning. These challenges, which appear trivial for a child, remain challenging for AI models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.597 + ], + "angle": 0, + "content": "How does the human brain process long videos, particularly in addressing the above-mentioned challenges? First, rather than processing every pixel or frame, the brain selectively attends to new information to efficiently manage temporal redundancy [4, 22]. Second, humans process videos in an online fashion, continuously updating their understanding and memories as they watch, rather than deferring reasoning until the end. This continuous knowledge update, combined with selective attention, allows the brain to efficiently handle the complexity of long videos. Finally, attention dynamically shifts based on context. Without specific guidance, a child may focus on naturally engaging or memorable moments. However, when given specific questions in advance, attention becomes goal-oriented to seek relevant details while maintaining a broad understanding. This suggests effective representations should balance between task-specific focus and holistic understanding of the video to enable cross-task generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Inspired by how humans process long videos, we introduce SEmantic Attention Learning (SEAL), a novel unified representation designed to tackle the three key challenges in long video understanding. SEAL consists of two main steps: Semantic Decomposition and Attention Learning. In Semantic Decomposition, long videos are decomposed into three semantic entities such as scenes, objects, and actions, which are then treated as tokens. While the scene and object tokens represent static content assumed not to change rapidly, the action tokens are designed to capture the dynamic, fast-changing moments of the video. We note that these semantic tokens efficiently encode the essential information needed to answer \"where\", \"what\", or \"how\" questions about the videos. This decomposition significantly reduces complexity by allowing AI models to operate on a compact set of tokens instead of raw pixels or frames. Figure 1(a) illustrates a conventional uniformly sampled video \\(\\mathcal{V}\\), where redundant frames create cluttered visual information that hinders effective analysis for both models and humans. Figure 1(b) shows our semantic decomposition," + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.458, + 0.902 + ], + "angle": 0, + "content": "*This work was done as Lan Wang's internship project at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26192" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.091, + 0.35, + 0.105 + ], + "angle": 0, + "content": "(a) Uniform sampling of a long video \\(\\mathcal{V}\\)" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.108, + 0.844, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.203, + 0.402, + 0.216 + ], + "angle": 0, + "content": "(b) Semantic decomposition of the same video \\(\\mathcal{V}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.217, + 0.2, + 0.23 + ], + "angle": 0, + "content": "Scenes (static)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.23, + 0.218, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.267, + 0.184, + 0.276 + ], + "angle": 0, + "content": "S1" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.23, + 0.311, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.267, + 0.276, + 0.275 + ], + "angle": 0, + "content": "S2" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.23, + 0.403, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.267, + 0.367, + 0.275 + ], + "angle": 0, + "content": "S3" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.23, + 0.495, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.267, + 0.468, + 0.275 + ], + "angle": 0, + "content": "S4" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.217, + 0.609, + 0.23 + ], + "angle": 0, + "content": "Objects (static)" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.235, + 0.57, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.269, + 0.558, + 0.277 + ], + "angle": 0, + "content": "01" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.239, + 0.652, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.267, + 0.622, + 0.276 + ], + "angle": 0, + "content": "02" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.239, + 0.701, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.267, + 0.691, + 0.276 + ], + "angle": 0, + "content": "03" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.235, + 0.761, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.266, + 0.748, + 0.276 + ], + "angle": 0, + "content": "04" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.236, + 0.798, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.266, + 0.799, + 0.276 + ], + "angle": 0, + "content": "05" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.23, + 0.872, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.266, + 0.859, + 0.276 + ], + "angle": 0, + "content": "06" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.281, + 0.269, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.295, + 0.449, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.343, + 0.364, + 0.352 + ], + "angle": 0, + "content": "A2" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.285, + 0.655, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.343, + 0.571, + 0.352 + ], + "angle": 0, + "content": "A3" + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.286, + 0.867, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.343, + 0.788, + 0.352 + ], + "angle": 0, + "content": "A4" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.357, + 0.318, + 0.37 + ], + "angle": 0, + "content": "(c) Query-aware attention learning" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.372, + 0.296, + 0.395 + ], + "angle": 0, + "content": "Q1: What does the vlogger family find in visitor center?" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.397, + 0.152, + 0.406 + ], + "angle": 0, + "content": "S101" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.372, + 0.5, + 0.396 + ], + "angle": 0, + "content": "Q2: What does the vlogger family do after buying fishing supplies?" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.397, + 0.378, + 0.407 + ], + "angle": 0, + "content": "S2 S3 O2 A1" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.372, + 0.704, + 0.395 + ], + "angle": 0, + "content": "Q3: What foods do the vlogger family eat the most here?" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.397, + 0.59, + 0.407 + ], + "angle": 0, + "content": "O3 O4 O5 A3" + }, + { + "type": "text", + "bbox": [ + 0.71, + 0.372, + 0.871, + 0.395 + ], + "angle": 0, + "content": "Q4: What is the most common form of transportation there?" + }, + { + "type": "text", + "bbox": [ + 0.71, + 0.397, + 0.792, + 0.407 + ], + "angle": 0, + "content": "S4 O6 A2 A4" + }, + { + "type": "text", + "bbox": [ + 0.819, + 0.401, + 0.881, + 0.414 + ], + "angle": 0, + "content": "Query set" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.416, + 0.38, + 0.431 + ], + "angle": 0, + "content": "Scene token Object token Action token" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.433, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Long Video Representation with Semantic Attention Learning (SEAL): (a) Conventional uniform sampling results in redundant and cluttered visual information, making it difficult for both AI models and human brains to process efficiently. (b) Decomposing long videos into semantic entities such as scenes, objects, and actions reduces temporal redundancy, thus making model training and inference more efficient. In this example, the long video \\(\\mathcal{V}\\) is decomposed into 4 scene tokens (S1-S4), 6 object tokens (O1-O6), and 4 action/event tokens (A1-A4). (c) Query-aware attention learning module improves downstream task performance by focusing on relevant information rather than processing everything. Queries (Q1-Q4) are shown with their most relevant tokens. (best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.728 + ], + "angle": 0, + "content": "breaking down the video into distinct scenes, static objects, and dynamic actions. In Attention Learning, we formulate a subset selection problem that maximizes the query relevance while ensuring token diversity. This step not only mitigates redundancy but also enhances cross-task generalization by prioritizing the most informative tokens. Figure 1(c) illustrates our attention learning module with four different queries and their selected tokens that capture relevance and diverse video content. Finally, SEAL is designed to work for both global and streaming modes, enabling it to process arbitrarily long videos. Extensive ablations and experiments confirm SEAL's superior performance over existing methods. Our key contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.732, + 0.483, + 0.776 + ], + "angle": 0, + "content": "- We introduce SEAL, a novel unified representation for long videos by decomposing them into three semantic tokens, namely scenes, objects, and actions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.778, + 0.482, + 0.837 + ], + "angle": 0, + "content": "- Our attention learning module reduces temporal redundancy while supporting strong cross-task generalization. We show SEAL works in both global and streaming modes, making it adaptable to arbitrarily long videos." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.839, + 0.482, + 0.899 + ], + "angle": 0, + "content": "- SEALoutperforms state-of-the-art methods on various long video understanding tasks and benchmarks including: video QA (MovieChat-1K [33], LVBench [40]), and egocentric video grounding (Ego4D [8])." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.732, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.531, + 0.655, + 0.545 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Recent approaches have unified various video understanding tasks by framing them as video QA tasks [40], leveraging the capabilities of LLMs. However, fine-tuning with task-specific vision heads continues to offer advantages in memory efficiency and task-specific performance [29], particularly for temporal grounding. In this section, we review advancements in Video Question Answering (QA) and temporal grounding for long video understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Video QA for long videos. The main challenge for long video QA is the memory constraint. He et al. [9] introduced a sequential framework that uses a memory bank to enhance long-term comprehension. Song et al. [33, 34] integrated video foundation models with LLMs through a memory mechanism inspired by the Atkinson-Shiffrin model, reducing computational complexity while preserving long-term memory. Another line of work improves efficiency by decomposing video content. Rui et al. [30] employed Memory-Propagated Streaming Encoding to segment videos into short clips, with Adaptive Memory Selection enhancing response accuracy by identifying question-relevant memories. Min et al. [24] introduced a multi-stage, training-free framework, emphasizing task decomposition into parsing, grounding, and reasoning stages. More re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26193" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "cently, Weng et al. [41] proposed a hierarchical framework that encodes local features and integrates global semantics for detailed comprehension of extended video content. To further reduce model's hallucination in QA, Sun et al. [35] proposed a question-guided pipeline by focusing on relevant frames and controlled answer generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Temporal localization for long videos. Recent research in temporal localization for long videos has explored two primary directions: LLM-based and non-LLM-based approaches. With the advances in LLMs, researchers have expanded its use beyond traditional VQA, leveraging the capabilities for temporal grounding tasks, with a primary focus on enhancing localization accuracy. Ren et al. [32] proposed a timestamp-aware model that aligns visual content with temporal cues, enabling adaptive processing of sequential events for tasks like localization. Similarly, Fan et al. [5] introduced a memory-enhanced framework that captures contextual information across video segments, improving temporal and spatial reasoning. Korbar et al. [15] developed a text-guided resampling mechanism that dynamically selects video segments, focusing on relevant scenes to enhance temporal and spatial comprehension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.605 + ], + "angle": 0, + "content": "In contrast, non-LLM-based approaches typically rely on training a regression layer or decoder for temporal localization, with efforts concentrated on refining visual features and multimodal fusion to improve alignment. For instance, Hou et al. [11] introduced a hierarchical framework that combines coarse scanning with fine-grained alignment to optimize both precision and efficiency in localizing target moments. Pan et al. [28] applied a coarse-to-fine pipeline for single-pass temporal grounding that improves both efficiency and alignment. Additionally, Mu et al. [26] proposed a cost-effective late fusion approach paired with a video-centric sampling scheme to improve scalability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Unlike prior works that focus on specific tasks, SEAL proposes a unified and generic framework for long video understanding, capable of adapting to various prediction heads and tasks. Additionally, SEAL supports both global and streaming modes, making it adaptable to arbitrarily long videos." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.743 + ], + "angle": 0, + "content": "3. Semantic Decomposition and Attention Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Let \\(\\mathcal{V} = \\{v_{i}\\}_{i = 1}^{T_{V}}\\) be an arbitrarily long untrimmed video, where \\(v_{1},\\ldots ,v_{T_{V}}\\) denote the sequence of \\(T_{V}\\) frames forming the video. Let \\(q\\) be a query from long video understanding tasks, comprising a sequence of \\(l_{q}\\) tokens. The query \\(q\\) may take different forms depending on the task, such as natural language text for video question and answering (e.g., MovieChatQA [33], LVBench [40]) or visual/text template or action label for episodic memory tasks in egocentric video understanding (e.g., Ego4D [8]). Our method establishes a unified video representation to gener" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "alize across these diverse long video understanding tasks. An overview of our approach is presented in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.135, + 0.859, + 0.151 + ], + "angle": 0, + "content": "3.1. Semantic Decomposition of Long Videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.158, + 0.905, + 0.415 + ], + "angle": 0, + "content": "The main challenge in long video representation lies in capturing diverse content within limited memory. Conventional methods resort to frame sampling [42, 44] or maintaining a memory bank that merges similar frames [33, 34]. However, these approaches can vary greatly across tasks. We propose a novel decomposition approach that structures long videos into three distinct token types representing different type of semantic entities: (1) Scene tokens \\(\\mathbf{T}_{\\mathrm{scene}}\\) capture background context, providing essential cues about the environment. (2) Action tokens \\(\\mathbf{T}_{\\mathrm{action}}\\) represent moving elements, focusing on temporal information such as motions, activities, or events. (3) Object tokens \\(\\mathbf{T}_{\\mathrm{object}}\\), highlight key static elements relevant to specific tasks. With this structured tokenization, we create a unified, compact, task-agnostic representation that minimizes the need for redundant and dense video storage while preserving a comprehensive understanding of long video content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.672 + ], + "angle": 0, + "content": "Scene Tokens. Any frame in a video can serve as a scene token because it captures the environment where it was recorded, e.g., indoor gym, outdoor mountain, etc. Although shot boundary detection could split a long video into shots with one scene token per shot, this approach has two drawbacks: (1) Shot detection algorithms are often imperfect, with no opportunity to correct once shot boundary is determined. (2) Shot detection often ignores the specific query and downstream tasks, and thus could cause suboptimal performance. To overcome these drawbacks, we propose a two-step approach. First, we over-sample scene tokens, and then later perform an attention learning step to maximize query relevance and token diversity. Specifically, we uniformly sample \\( N_{\\mathrm{scene}} \\) frames to capture a diverse background. These pre-sampled scene tokens, denoted as \\( \\mathbf{T}_{\\mathrm{scene}} = [t_i^{\\mathrm{scene}}]_{i=1}^{N_{\\mathrm{scene}}} \\), will undergo another round of attention learning, detailed in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.905, + 0.856 + ], + "angle": 0, + "content": "Action Tokens. The purpose of the action token is to capture temporal information of moving objects such as low-level motions, activities, and events. We begin by using a class-agnostic object tracker, e.g., SAM-2 [31], to extract multiple initial dynamic tracklets. Tracklets shorter than \\( L_{\\mathrm{min}} \\) are discarded, while those longer than \\( L_{\\mathrm{max}} \\) are split into multiple tracklets with length of \\( L_{\\mathrm{max}} \\). For each tracklet, we take the spatial union of bounding boxes across frames, allowing the dynamic token to capture not only motion information but also the spatial movement of people and objects. This process yields \\( N_{\\mathrm{tracklet}} \\) tracklets, denoted as \\( \\mathbf{T}_{\\mathrm{action}} = \\{\\tau_{\\mathrm{dynamic}}^{i}\\}_{i=1}^{N_{\\mathrm{tracklet}}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Object Tokens. For the object tokens, we utilize a class-agnostic grouping method such as SAM [14] to generate masks for all objects in each frame. This class-agnostic" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26194" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.096, + 0.852, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.387, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. SEAL Overview. During semantic decomposition, a long video \\(\\mathcal{V}\\) is decomposed into semantic tokens representing scenes, objects, and actions. Then, during attention learning, these tokens and the query q, are optimized for query relevance \\(R(\\cdot)\\) and token diversity \\(S(\\cdot)\\). The resulting attended token subset is then passed to a vision or an MLLM head for predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.545 + ], + "angle": 0, + "content": "segmentation approach enables comprehensive object information capture. Specifically, we apply this process on \\( N_{\\mathrm{key}} \\) key frames. On each of the \\( N_{\\mathrm{key}} \\) frames, we apply SAM to obtain in total \\( N_{\\mathrm{object}} \\) object masks, which we denote as \\( \\mathbf{T}_{\\mathrm{object}} = \\{\\mathbf{M}_{\\mathrm{object}}^i\\}_{i=1}^{N_{\\mathrm{object}}} \\). Key frame selection varies by task and is detailed in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.553, + 0.277, + 0.569 + ], + "angle": 0, + "content": "3.2. Attention Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Although the long video is decomposed into different token representations, the resulting tokens are still redundant. To address this, we propose a sampling approach that balances between query-relevance and token-diversity, denoted as Attention Learning. Specifically, we formulate our sampling as an optimization problem with two main objectives: query-relevance and token-diversity." + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.689, + 0.479, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T_{s}^{*} = \\operatorname *{arg max}_{T_{s}\\subset T_{G}}F_{s}(T_{s}|T_{G},q) \\\\ = \\underset {T _ {s} \\subset T _ {G}} {\\arg \\max} \\alpha \\sum_ {t _ {s} \\in T _ {s}} R (t _ {s}, q) + (1 - \\alpha) \\sum_ {t _ {i}, t _ {j} \\in T _ {s}, i \\neq j} \\frac {1}{S (t _ {i} , t _ {j})}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We aim to solve a subset selection problem: find a fixed-size subset \\( T_{s} \\subset T_{G} \\) that maximizes the objective function \\( F_{s}(T_{s}|T_{G},q) \\). Here, \\( T_{G} \\) represents the set of all tokens used, which can be any set of tokens such as scene \\( \\mathbf{T}_{\\mathrm{scene}} \\), object \\( \\mathbf{T}_{\\mathrm{object}} \\), action \\( \\mathbf{T}_{\\mathrm{action}} \\), or a combination of all \\( (\\mathbf{T}_{\\mathrm{scene}} \\cup \\mathbf{T}_{\\mathrm{object}} \\cup \\mathbf{T}_{\\mathrm{action}}) \\). \\( q \\) denotes the query being asked by the downstream understanding task. And, \\( T_{s}^{*} \\) denotes the optimal subset that maximizes \\( F_{s} \\). \\( F_{s} \\) is decomposed into two terms: \\( R(\\cdot) \\) measures the relevance between a vi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.572 + ], + "angle": 0, + "content": "sual token and the query which we compute by encoding them with the BLIP-2 model [17] and calculating their cosine similarity; and \\( S(\\cdot) \\) calculates the cosine similarity of the paired tokens. We note that the first term maximizes the relevance between selected tokens and the query while the second term enforces token diversity (via minimizing the token pairwise similarity). Finally, \\( \\alpha \\) is a hyper-parameter to balance between query-relevance and token-diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.771, + 0.593 + ], + "angle": 0, + "content": "3.3. Streaming and Global Mode" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Our proposed method offers two ways of representing long videos: streaming and global. In the global mode, the model fully \"watches\" (or processes) the entire video and then provides a single representation. In contrast, in the streaming mode, the model processes the video buffer by buffer and provides an updated representation at any given time step. The partly-observed video representation can change over time as the video progresses. This setting simulates the situation when you watch a movie together with a child and have interactions with her or him as the movie is still going on." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the global model, all tokens undergo a single optimization through Attention Leaning, resulting in \\( T_{\\mathrm{sub}} \\), which includes \\( k \\) tokens. This \\( T_{\\mathrm{sub}} \\) is regarded as the representation for the entire video. We note that the global mode may not scale well with arbitrary long videos because all tokens cannot be fit into a limited memory for sampling. One can opt to use more aggressive uniform temporal presampling to reduce the numbers of tokens before Attention Learning. This workaround can bypass the memory lim" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26195" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "itation, but may also lead to sub-optimal solutions due to missing important tokens due to uniform sampling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.26 + ], + "angle": 0, + "content": "Alternatively, we propose an online streaming approach for representing partly-observed videos. Specifically, we use a fixed-size sliding window with a size set to the maximum number of tokens \\( l \\) allowed by memory capacity, denoted as \\( T \\). At each step \\( t \\), we apply Attention Learning to the union set of the tokens in the current window \\( T_{t} \\) and the previous selected subset of tokens \\( T_{\\mathrm{sub}}^{t - 1} \\) and obtain the representation of the video at time \\( T_{\\mathrm{sub}}^{t} \\). At the beginning, the selected subset is set to empty \\( (T_{\\mathrm{sub}}^{0} = \\emptyset) \\)." + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.27, + 0.483, + 0.29 + ], + "angle": 0, + "content": "\\[\nT _ {\\text {s u b}} ^ {t} = \\text {A t t e n t i o n L e a r n i n g} \\left(T _ {t} \\cup T _ {\\text {s u b}} ^ {t - 1}\\right) \\forall t > 0. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.378 + ], + "angle": 0, + "content": "This streaming mode allows us to use \\( T_{\\mathrm{sub}}^{t} \\) as the partly-observed representation of the video and can be fed into any prediction head for video understanding tasks. As an immediate benefit of the streaming mode, our proposed representation now can handle arbitrary long videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.262, + 0.403 + ], + "angle": 0, + "content": "3.4. Prediction Heads" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.5 + ], + "angle": 0, + "content": "Our unified representation is adaptable to most long video understanding tasks using different prediction heads. In this paper, we demonstrate two specific use cases of our representation: one is used with the traditional vision head for video temporal grounding and the other one is with the Multimodal LLMs (MLLM) head for video QA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Temporal Grounding with Vision Heads. Given a query \\( q \\), the task is to locate the start and end times, \\( t_{\\mathrm{start}} \\) and \\( t_{\\mathrm{end}} \\), where the answers could be deduced. We first encode the sampled tokens \\( T_{\\mathrm{sub}} \\) and the query using encoders \\( \\mathbb{E}_V \\) and \\( \\mathbb{E}_q \\) to obtain embeddings for each video token \\( z_v^i = \\mathbb{E}_V(T_{\\mathrm{sub}^i}) \\) and query \\( z_q = \\mathbb{E}_q(q) \\). The cross-modal fusion is then performed to obtain the fused representation:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.622, + 0.483, + 0.643 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {j o i n t}} ^ {i} = \\operatorname {C r o s s M o d a l F u s i o n} \\left(z _ {v} ^ {i}, z _ {q}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.74 + ], + "angle": 0, + "content": "Finally, a moment decoder is applied to predict the start and end time for the query \\( q \\). The moment decoder includes a classification head and a regression head. The classification head is used to predict the score \\( p_{score}^{i} \\) of each token, while the regression head predicts the normalized distances \\( (d_{\\mathrm{start}}^{i}, d_{\\mathrm{end}}^{i}) \\) from each token to the moment boundaries:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.765, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\np _ {\\text {s c o r e}} ^ {i}, \\left(d _ {\\text {s t a r t}} ^ {i}, d _ {\\text {e n d}} ^ {i}\\right) = \\text {M o m e n t D e c o d e r} \\left(z _ {\\text {j o i n t}} ^ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The regression and classification heads are optimized using an IoU distance and a focal loss as used in [25]. The final moment is calculated as \\( t_{start}^{i}, t_{end}^{i} = (t_{i} - d_{start}^{i}) \\times L_{V}, (t_{i} + d_{end}^{i}) \\times L_{V} \\), where \\( L_{V} \\) is the length of the input video. The proposed unified representation enables efficient handling of long video sequences and allows for localization related tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "Video QA using MLLM Heads. The proposed representation can also be connected to an MLLM head, making it applicable to various video QA-related tasks, such as reasoning, understanding, and summarization. Moreover, some grounding tasks can also be addressed in a QA format. Specifically, \\( z_v \\) is projected through an MLP to a visual token \\( T_v^{\\mathrm{MLLM}} \\) that the MLLM can interpret, which is then input into the MLLM along with text tokens \\( T_{text} \\). Based on different benchmarks, the MLLM performs multiple-choice or open-ended answering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.26, + 0.646, + 0.276 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.731, + 0.3 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Datasets and metrics. We evaluate SEAL on three datasets, each selected for its relevance to long video understanding on different capabilities. LVBench [40] contains 1,549 QA pairs across six tasks, with videos averaging 4,101 seconds (approximate 1 hour 8 minutes). Each question presents a single-choice format with four options. Accuracy serves as the evaluation metric for individual tasks as well as overall performance across all tasks. We primarily focus on this dataset due to its emphasis on hour-long videos. Moviechat-1K [33] includes 1,000 video clips with dense captions spanning 15 categories, averaging 564 seconds (about 10 minutes). The benchmark employs LLMs, specifically GPT-3.5 [27], to evaluate the quality of generated answers. A rating ranging from 0 to 5 is used to compute the overall score, while a binary preference from the LLM is used to calculate the accuracy. Ego4D-NLQ [8] is a part of the Ego4D Episodic Memory challenge for Natural Language Queries (NLQ) task. This dataset requires localizing a temporal window where the answers can be deduced from untrimmed egocentric videos. It contains 1,259 videos, averaging 10 minutes each. Our experiments apply memory constraints to simulate long video scenarios. Metrics include Top-1 and Top-5 recall at various thresholds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Experiment setup. We fine-tune the projection layers and Q-former [17] to adapt to different types of tokens for 20 epochs using the training split of MovieChat-1K. For Ego4D-NLQ, we use the training split to finetune the vision heads for 7 epochs. The AdamW optimizer [21] is employed with default beta values of (0.9, 0.999) and a weight decay of 0.05. For token extraction, we use SAM2 [31] to obtain object tokens, YOLOv10-X [37] with BoT-SORT [1] for action tokens. Scene, action, and object tokens are extracted at 8, 10, 1 FPS for MovieChat-1K, Ego4D-NLQ and LVbench. For the LVBench, we follow the settings of InternVL2 and utilize Yi-34B [43]. For MovieChat-1K, we adopt the same settings as [33] and use Vicuna-7B as [33]. For Ego4D-NLQ, we adhere to the settings specified in SnAG [25] and use EgoVLP [19] as vision encoder. We use 0.9 as the default value for the only hyper-parameter \\(\\alpha\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26196" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.868, + 0.214 + ], + "angle": 0, + "content": "
ModelLLM SizeOverall (%)KIR (%)EU (%)Sum (%)ER (%)Rea (%)TG (%)
Qwen2-VL [39]72B41.338.341.146.638.046.541.4
InternVL2 [3]34B39.643.439.741.437.442.531.4
LLaVA-NeXT [45]34B32.234.131.227.630.135.031.4
Oryx [20]34B30.432.129.227.630.134.029.1
PLLaVA [42]34B26.126.224.925.925.030.021.4
SEAL (Ours)34B45.951.541.339.747.943.332.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.216, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art on LVBench. Our method achieves the highest overall score, with notable gains in Key Information Retrieval (KIR) and Entity Recognition (ER), demonstrating the effectiveness of our representation in locating key information and entities by eliminating redundancy. The best methods are highlighted in bold, and the second-best are underlined." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.272, + 0.486, + 0.389 + ], + "angle": 0, + "content": "
Model#TokenR@1R@5
0.30.5Avg0.30.5Avg
SnAGall15.7210.7813.2538.3927.4432.92
SnAG45013.449.2311.3434.0223.0428.53
SEAL45013.789.2611.5234.7923.1028.95
SnAG20010.036.358.1926.5616.9021.73
SEAL20010.837.068.9527.3917.4122.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Table 2. Comparisons with SoTA methods on Ego4D-NLQ. Quantitative results for temporal grounding on Ego4D episodic memory Natural Language Queries (NLQ) task show our method, SEAL, consistently outperforms SnAG [25] in all metrics under memory constraints with varying number of tokens." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.483, + 0.392, + 0.499 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We conduct quantitative analysis on LVBench dataset which contains videos on lengths averaged more than 1 hour. Additionally, we demonstrate the generalizability of the proposed unified representation using MovieChat-1K, which has open-ended QA questions on a variety of scenarios, and Ego4D-NLQ, where we use traditional vision decoders (LLM-independent) for the temporal grounding task. LVBench. Table 1 shows the accuracy of our method on different categories of LVBench dataset [40], demonstrating that SEAL achieves the highest overall score, notably outperforming even larger models like Qwen2-VL72B [39] by \\(4.6\\%\\). Our method excels in Key Information Retrieval (KIR) and Entity Recognition (ER), outperforming the strongest alternatives by \\(8.1\\%\\) and \\(5.1\\%\\) respectively. These results highlight that our action and object tokens effectively locate key information and entities by removing redundancies. While larger LLMs often exhibit better performance, this analysis suggests that it is not the only determinant, and a unified representation like ours can achieve state-of-the-art results with fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Figure 3 presents a qualitative analysis of SEAL on LVBench, demonstrating that our approach is able to pay attention to relevant semantic tokens and make correct answers to different types of questions. SEAL accurately locates relevant tokens, e.g., identifying the royal family's stool color (Q1.a), counting meals eaten (Q2.a), or deter" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.272, + 0.909, + 0.405 + ], + "angle": 0, + "content": "
MethodsAccuracy(%)Score
Zero-shotVideoChat [18]61.03.34
VideoLLaMA [44]51.43.10
Video-ChatGPT [23]44.22.71
MovieChat [33]67.83.81
SupervisedTimeChat-Hal [35]73.83.58
HERMES [6]84.94.40
SEAL (Ours)86.84.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.407, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Table 3. Comparison with SoTA methods on MovieChat-1K. SEAL outperforms the second best method, HERMES, by \\(1.9\\%\\) on accuracy while being comparable on the score metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.905, + 0.508 + ], + "angle": 0, + "content": "mining scene or location (Q2.b). We also show a failure case (Q2.c), where the nuanced \"why\" question requires complicated causal relations of different scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.704 + ], + "angle": 0, + "content": "Moviechat-1K. Table 3 compares our method with state-of-the-art methods on MovieChat-1K. We focus on evaluating performance in the global mode, which assesses the model's ability to comprehend information from the entire video, rather than the breakpoint mode, which primarily evaluates its ability to answer questions related to specific timestamps. Our method surpasses all the alternatives, achieving the highest accuracy and a strong score. Notably, unlike LVBench, which employs multi-choice questions, this dataset evaluates generated answers using an off-the-shelf LLM. This highlights the ability of our learned representations to effectively transfer to downstream generative tasks, enabling generating accurate and detailed responses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ego4D-NLQ. We also evaluate our approach on the Ego4D-NLQ task, which differs from previous datasets as it uses a decoder to locate events for temporal grounding, rather than relying on an LLM to generate text-based answers. Table 2 shows the results under memory-constrained conditions, where we limit token numbers to simulate real-world memory limitations in long video understanding. Our method consistently achieves the highest recalls across various thresholds. With further memory constraints (e.g., reducing tokens to 200), SEAL outperforms the current SoTA method, SnAG [25], by a even larger margin. This demonstrates the robustness of our unified representation on various downstream applications for long video understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26197" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.09, + 0.833, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.561, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on LVBench. Two long videos visualized with questions, multiple choice options, and SEAL predicted answers. SEAL attends to relevant entities such as \"royal family\" and \"stool\" (Q1.a), different \"meals\" and \"drinks\" (Q2.a), \"scene\" and \"location\" (Q2.b) and correctly answers these questions. Although attending to relevant \"push-up\" activity (Q2.c), SEAL fails to predict the right answer due to the challenging in the causal reasoning question." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.245, + 0.65 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Semantic decomposition. We analyze the impact of the three proposed token types on LVBench (Table 4). To prevent LLMs from leveraging prior knowledge to \"guess\" answers using only the text query, we present the baseline that uses random token inputs, isolating the LLM's performance. Results indicate that using each token type individually significantly improves performance, highlighting the value of each token's contribution. Notably, scene tokens provide the most substantial improvement, aligning with prior research [2, 36, 41] that emphasizes frame-based scene tokens. Furthermore, different token combinations yield additional gains: adding action tokens enhances Key Information Retrieval (KIR), Reasoning (Rea) and Temporal Grounding (TG) tasks by capturing temporal dynamics in the queries, while object tokens boost Entity Recognition (ER) performance by retaining detailed object-specific" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.771 + ], + "angle": 0, + "content": "information. We observe high variance in the Summarization (Sum) accuracy due to the small number of questions in this category and the randomness of LLM. Similar improvements have been observed on the MovieChat-1K in Table 5, where the accuracy improved by \\(10.45\\%\\) with action and object tokens. Ultimately, using all three token types together achieves the highest performance, demonstrating the complementary strengths of scene, action, and object tokens as a unified representation in long video understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We observe that the latency bottleneck mostly comes from the tracker when processing multiple objects. Table 6 presents an analysis of different trackers for extracting action tokens. While YOLO+BoT-SORT operates in a much higher frame rate (14 FPS), it lacks open-set detection capabilities. On the other hand, SAM2 can detect all objects within a scene but operates significantly slower (8FPS). Our findings indicate that the proposed action tokens does not" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26198" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.209 + ], + "angle": 0, + "content": "
ModelOverallKIREUSumERReaTG
Random tokens24.824.725.234.523.526.420.9
Action only34.033.331.736.234.335.833.6
Object only33.934.432.631.033.440.829.5
Scene only42.850.540.625.944.341.327.2
Scene+Object43.447.440.024.147.642.335.5
Scene+Action44.453.340.024.146.243.338.2
SEAL (Ours)45.951.541.339.747.943.332.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.211, + 0.482, + 0.253 + ], + "angle": 0, + "content": "Table 4. The effects of different types of tokens on LVbench. Using any type of tokens outperforms the random baseline, while applying all three types of tokens brings the best performance." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.265, + 0.48, + 0.317 + ], + "angle": 0, + "content": "
MethodSceneScene+ObjectScene+ActionSEAL
Accuracy76.3379.4981.4686.78
Score4.154.264.284.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.36 + ], + "angle": 0, + "content": "Table 5. The effects of different types of tokens on MovieChat1K. Results are on the test set with our global inference mode. Using all three types of tokens provides the best accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.641 + ], + "angle": 0, + "content": "require capturing all the objects in the scene. In practice, YOLO-X pretrained on COCO, with BoT-SORT, effectively captures essential information for action token extraction, maintaining a balance between performance and efficiency. Attention learning. Table 6 highlights the effectiveness of query relevance and token diversity terms in Section 3.2. When \\(\\alpha = 0\\), the optimization focuses on the diversity term to capture extensive contextual information for long-video tasks such as Summarization (Sum) and Reasoning (Rea), achieving reasonable results comparable to other state-of-the-art approaches in Table 1. With \\(\\alpha = 1\\), the optimization prioritizes the relevance term to reduce redundancy by sampling tokens that closely align with the query, so performance can be improved for tasks like Key Information Retrieval (KIR). However, this approach loses global context and results in weaker summarization performance. Balancing both terms yields the best overall results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Streaming mode. We compare the global mode and streaming mode of our method on LVBench. As discussed in Section §3, the streaming mode better simulates real-world scenarios by sequentially processing partial observations to aggregate information over arbitrarily long videos. Table 6 demonstrates that streaming mode performs worse on globally-dependent tasks such as Event Understanding (EU) and Key Information Retrieval (KIR). However, it excels in temporally-intensive tasks like Temporal Grounding (TG) and Reasoning (Rea). Notably, streaming mode still surpasses the most competitive baseline, even when the baseline uses a significantly larger LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Computational Complexity. We provide an accuracy-efficiency trade-off comparison in Fig 4 where both SEAL and InternVL2 are with a varying number of tokens. FPS is calculated during inference that includes all the processing time from the raw video frames. SEAL reduces the re" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.183 + ], + "angle": 0, + "content": "
ModelOverallKIREUSumERReaTG
SEAL (Ours)45.951.541.339.747.943.332.3
rep Yolo w/ SAM243.351.939.929.346.440.332.7
⇒ Streaming44.250.939.737.945.044.334.1
Diversity only38.740.438.840.038.740.933.7
Relevance only42.048.539.927.644.037.331.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.184, + 0.905, + 0.254 + ], + "angle": 0, + "content": "Table 6. Ablation studies on LVBench. SEAL defaults to global mode with YOLO for efficiency, while SAM2 performs similarly. Global mode excels in long-context tasks, and streaming mode in temporally-intensive tasks. Combining Relevance and Diversity in attention learning achieves the best performance." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.264, + 0.897, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.43, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Figure 4. Accuracy vs. efficiency trade-off on LVBench. SEAL runs \\(2 - 3\\mathrm{x}\\) faster than InternVL2 at the same accuracy, and is more accurate when compared at the same FPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.587 + ], + "angle": 0, + "content": "liaence on densely sampled tokens and uses subject-level and motion aware representations/tokens to improve efficiency while maintain good accuracy. From Fig 4, at 10 FPS, SEAL is about \\(5\\%\\) more accurate than InternVL2. When comparing at the same accuracy of \\(41.5\\%\\) and \\(40\\%\\), SEAL is about 3x and 2x faster than InternVL2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.783, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.794 + ], + "angle": 0, + "content": "We propose SEAL, a novel unified representation for long video understanding that addresses computational complexity, temporal redundancy, and cross-task generalization. SEAL leverages semantic decomposition to break videos into scene, object, and action tokens, reducing redundancy and enabling efficient processing. It incorporates attention learning to balance query relevance and token diversity, enhancing performance across diverse tasks. SEAL achieves state-of-the-art results on benchmarks like MovieChat-1K, LVBench, and Ego4D-NLQ, demonstrating its versatility and effectiveness for long video understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limitation. The Attention Learning module is bounded by the memory constraint for the QP solver, making it not fully end-to-end trainable. For future work, we plan to integrate Attention Learning in our streaming mode for full end-to-end learning. Developing special prediction heads to solve causal reasoning [16], where MLLM heads showed limitations, is also an interesting area for future exploration." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26199" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by the Office of Naval Research (award #N00014-23-1-2417). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of ONR." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.18, + 0.188, + 0.195 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. ArXiv, abs/2206.14651, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[2] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[3] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[4] Meghomala Das, David M Bennett, and Gordon N Dutton. Visual attention as an important visual function: an outline of manifestations, diagnosis and management of impaired visual attention. British Journal of Ophthalmology, 91(11): 1556-1560, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[5] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In ECCV, pages 75-92. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.485, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[6] Gueter Josmy Faure, Jia-Fong Yeh, Min-Hung Chen, Hung-Ting Su, Shang-Hong Lai, and Winston H. Hsu. Hermes: temporal-coherent long-form understanding with episodes and semantics. arXiv preprint arXiv:2408.17443, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[7] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.625, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[8] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonio Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawa" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.205, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "har, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.177, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[9] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.285 + ], + "angle": 0, + "content": "[10] Fabian Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[11] Zhijian Hou, Wanjun Zhong, Lei Ji, Difei Gao, Kun Yan, Wing-Kwong Chan, Chong-Wah Ngo, Zheng Shou, and Nan Duan. Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.412 + ], + "angle": 0, + "content": "[12] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos \"in the wild\". CoRR, abs/1604.06182, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.904, + 0.481 + ], + "angle": 0, + "content": "[13] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. CoRR, abs/1705.06950, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.904, + 0.537 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[15] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In ECCV, pages 271-288. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.904, + 0.635 + ], + "angle": 0, + "content": "[16] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.904, + 0.69 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML. PMLR, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[18] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[19] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[20] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.904, + 0.9 + ], + "angle": 0, + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26200" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[22] Steven J. Luck and Michelle Ford. On the role of selective attention in visual perception. Proceedings of the National Academy of Sciences of the United States of America, 95(3): 825-830, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[23] Muhammad Maaz, Hanoona Abdul Rasheed, Salman H. Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In AMACL, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[24] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In CVPR, pages 13235-13245, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[25] Fangzhou Mu, Sicheng Mo, and Yin Li. SnAG: Scalable and accurate video grounding. In CVPR, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.289, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[26] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.331, + 0.283, + 0.344 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-3.5, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.345, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[28] Yulin Pan, Xiangteng He, Biao Gong, Yiliang Lv, Yujun Shen, Yuxin Peng, and Deli Zhao. Scanning only once: An end-to-end framework for fast temporal grounding in long videos. In ICCV, pages 13767-13777, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.401, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[29] Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. Egovieo: Exploring egocentric foundation model and downstream adaptation. arXiv preprint arXiv:2406.18070, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[30] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[32] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[33] E. Song, W. Chai, G. Wang, Y. Zhang, H. Zhou, F. Wu, X. Guo, T. Ye, Y. Lu, JN. Hwang, and G. Wang. MovieChat: From dense token to sparse memory for long video understanding. In CVPR, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[34] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[35] Yiwei Sun, Zhihang Liu, Chuanbin Liu, Bowei Pu, Zhihan Zhang, and Hongtao Xie. Hallucination mitigation prompts long-term video understanding. arXiv preprint arXiv:2406.11333, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In CVPR, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection. ArXiv, abs/2405.14458, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[38] Lan Wang, Gaurav Mittal, Sandra Sajeev, Ye Yu, Matthew Hall, Vishnu Naresh Boddeti, and Mei Chen. Protege: Untrimmed pretraining for video temporal grounding by video temporal grounding. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[39] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[40] W. Wang, Z. He, W. Hong, Y. Cheng, X. Zhang, J. Qi, S. Huang, B. Xu, Y. Dong, M. Ding, and J. Tang. LVBench: An extreme long video understanding benchmark, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[41] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.461, + 0.905, + 0.57 + ], + "angle": 0, + "content": "[43] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, Xiaohui Hu, Xiaoyi Ren, Xinyao Niu, Pengcheng Nie, Yuchi Xu, Yudong Liu, Yue Wang, Yuxuan Cai, Zhenyu Gu, Zhiyuan Liu, and Zonghong Dai. Yi: Open foundation models by 01.ai. ArXiv, abs/2403.04652, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[45] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26201" + } + ] +] \ No newline at end of file diff --git a/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_origin.pdf b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e714dd266df9a90fe192d5b7b3b7a8c7a8ce551 --- /dev/null +++ b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/382a607d-24da-4509-8433-748d58e2f990_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41fc08c704032e15434bd9e7a899f6204170d387934796d90bba313212e3c65b +size 3215829 diff --git a/2025/SEAL_ Semantic Attention Learning for Long Video Representation/full.md b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee2dcbb4e891b440e545cffa6c9362f54f38562b --- /dev/null +++ b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/full.md @@ -0,0 +1,306 @@ +![](images/6c54110e18f1ff4f83f3abe5fb0be8f6f1a2943486504e70db095fcd940d7e26.jpg) + +# SEAL: Semiclastic Attention Learning for Long Video Representation + +Lan Wang $^{1,2*}$ Yujia Chen $^{2}$ Du Tran $^{2}$ Vishnu Naresh Boddeti $^{1}$ Wen-Sheng Chu $^{2}$ $^{1}$ Michigan State University + $^{2}$ Google + +{wanglan3, vishnu}@msu.edu {yujiachen, tranldu, wschu}@google.com + +# Abstract + +Long video understanding presents challenges due to the inherent high computational complexity and redundant temporal information. An effective representation for long videos must efficiently process such redundancy while preserving essential contents for downstream tasks. This paper introduces SEmatic Attention Learning (SEAL), a novel unified representation for long videos. To reduce computational complexity, long videos are decomposed into three distinct types of semantic entities: scenes, objects, and actions, allowing models to operate on a compact set of entities rather than a large number of frames or pixels. To further address redundancy, we propose an attention learning module that balances token relevance with diversity, formulated as a subset selection optimization problem. Our representation is versatile and applicable across various long video understanding tasks. Extensive experiments demonstrate that SEAL significantly outperforms state-of-the-art methods in video question answering and temporal grounding tasks across diverse benchmarks, including LVBench, MovieChat-1K, and Ego4D. + +# 1. Introduction + +State-of-the-art video understanding models excel at short video tasks such as video classification [7, 13], temporal grounding [38] and action detection [10, 12], which involve videos lasting from a few seconds to minutes. However, their performance declines on hour-long videos [40, 41]. In contrast, a 10-year-old child can watch a full-length movie (one to two hours) and effortlessly answer questions at various levels of detail. This disparity between humans and machines emphasizes the foundational challenges in long video understanding for AI models, including: (1) Increased complexity: Long videos require more computation and memory than current hardware can support for training or inference, (2) Temporal redundancy: Slow-changing scenes and objects introduce significant redundancy, and (3) Cross-task generalization: A robust repre + +sentation must adapt to various tasks, from fine-grained fact retrieval to high-level reasoning. These challenges, which appear trivial for a child, remain challenging for AI models. + +How does the human brain process long videos, particularly in addressing the above-mentioned challenges? First, rather than processing every pixel or frame, the brain selectively attends to new information to efficiently manage temporal redundancy [4, 22]. Second, humans process videos in an online fashion, continuously updating their understanding and memories as they watch, rather than deferring reasoning until the end. This continuous knowledge update, combined with selective attention, allows the brain to efficiently handle the complexity of long videos. Finally, attention dynamically shifts based on context. Without specific guidance, a child may focus on naturally engaging or memorable moments. However, when given specific questions in advance, attention becomes goal-oriented to seek relevant details while maintaining a broad understanding. This suggests effective representations should balance between task-specific focus and holistic understanding of the video to enable cross-task generalization. + +Inspired by how humans process long videos, we introduce SEmantic Attention Learning (SEAL), a novel unified representation designed to tackle the three key challenges in long video understanding. SEAL consists of two main steps: Semantic Decomposition and Attention Learning. In Semantic Decomposition, long videos are decomposed into three semantic entities such as scenes, objects, and actions, which are then treated as tokens. While the scene and object tokens represent static content assumed not to change rapidly, the action tokens are designed to capture the dynamic, fast-changing moments of the video. We note that these semantic tokens efficiently encode the essential information needed to answer "where", "what", or "how" questions about the videos. This decomposition significantly reduces complexity by allowing AI models to operate on a compact set of tokens instead of raw pixels or frames. Figure 1(a) illustrates a conventional uniformly sampled video $\mathcal{V}$ , where redundant frames create cluttered visual information that hinders effective analysis for both models and humans. Figure 1(b) shows our semantic decomposition, + +![](images/e376a8c5c37aa7672a656904d522374fcf96f305a9d4174a4544970c9b6bf9ab.jpg) +(a) Uniform sampling of a long video $\mathcal{V}$ +(b) Semantic decomposition of the same video $\mathcal{V}$ + +![](images/bb45a9f5db8e5a1015b0c03a9cae93ffbaad42b71a359c6424a3f3af6da7c8ba.jpg) +Scenes (static) +S1 + +![](images/4430aff250ab55c574d0e11b3d01d8a4104e847aecbc8d82fcf97ebc3acd2dc9.jpg) +S2 + +![](images/3b368e06436a23df4c216b91bdbf69267f17f41a19a6de9876ab9a4776386c63.jpg) +S3 + +![](images/7a11ac0bd5adae8ff7092b5f35c482185a9f1173f3cd6c23cd7e9e4eb6b40d44.jpg) +S4 + +![](images/62df9a031fe037ddbe4e763f15cc213a54b1063b2cde4366e5a724f5c8e68a92.jpg) +Objects (static) +01 + +![](images/cf5411fcd1ed2b2307eadaa0915ffab70e28c25da68192c14fbd5a3e8fb1082e.jpg) +02 + +![](images/8bef0394388c1efaf48b92b127da581f826a7d0b0aa4d72843ae52fd29b3dc2a.jpg) +03 + +![](images/e89a7bf3469a622daec24e880f4ef50506dd7660a06a8c27a2bdccb92b1765a0.jpg) +04 + +![](images/df85f832aef5638952ac25ce0b3af698934f391557cf6943f0816dc43b776630.jpg) +05 + +![](images/6e3b0724e15b656553c253e615f66f79a1e6fb1c9c97a13c256e2ed280c8aad5.jpg) +06 + +![](images/0de4353d795382f2a4127a8c6d82f141adc2277343ce3df2743e8009dcda911d.jpg) +(c) Query-aware attention learning +Figure 1. Long Video Representation with Semantic Attention Learning (SEAL): (a) Conventional uniform sampling results in redundant and cluttered visual information, making it difficult for both AI models and human brains to process efficiently. (b) Decomposing long videos into semantic entities such as scenes, objects, and actions reduces temporal redundancy, thus making model training and inference more efficient. In this example, the long video $\mathcal{V}$ is decomposed into 4 scene tokens (S1-S4), 6 object tokens (O1-O6), and 4 action/event tokens (A1-A4). (c) Query-aware attention learning module improves downstream task performance by focusing on relevant information rather than processing everything. Queries (Q1-Q4) are shown with their most relevant tokens. (best viewed in color) + +![](images/82b5b95f132b0db4daee7c99b1dbd1f6a1e04a55ef4093c071c954bef278b790.jpg) +A2 + +![](images/f326d91f32ce2dc7c6a6a24008362b8a25066b80c776c161e5667f12aa40c04c.jpg) +A3 + +![](images/6a375c373de1ff5ddebd5a3d778ffdeec1026a813758fdd46adafd3f8c23f3ed.jpg) +A4 + +Q1: What does the vlogger family find in visitor center? + +S101 + +Q2: What does the vlogger family do after buying fishing supplies? + +S2 S3 O2 A1 + +Q3: What foods do the vlogger family eat the most here? + +O3 O4 O5 A3 + +Q4: What is the most common form of transportation there? + +S4 O6 A2 A4 + +Query set + +Scene token Object token Action token + +breaking down the video into distinct scenes, static objects, and dynamic actions. In Attention Learning, we formulate a subset selection problem that maximizes the query relevance while ensuring token diversity. This step not only mitigates redundancy but also enhances cross-task generalization by prioritizing the most informative tokens. Figure 1(c) illustrates our attention learning module with four different queries and their selected tokens that capture relevance and diverse video content. Finally, SEAL is designed to work for both global and streaming modes, enabling it to process arbitrarily long videos. Extensive ablations and experiments confirm SEAL's superior performance over existing methods. Our key contributions include: + +- We introduce SEAL, a novel unified representation for long videos by decomposing them into three semantic tokens, namely scenes, objects, and actions. +- Our attention learning module reduces temporal redundancy while supporting strong cross-task generalization. We show SEAL works in both global and streaming modes, making it adaptable to arbitrarily long videos. +- SEALoutperforms state-of-the-art methods on various long video understanding tasks and benchmarks including: video QA (MovieChat-1K [33], LVBench [40]), and egocentric video grounding (Ego4D [8]). + +# 2. Related Work + +Recent approaches have unified various video understanding tasks by framing them as video QA tasks [40], leveraging the capabilities of LLMs. However, fine-tuning with task-specific vision heads continues to offer advantages in memory efficiency and task-specific performance [29], particularly for temporal grounding. In this section, we review advancements in Video Question Answering (QA) and temporal grounding for long video understanding. + +Video QA for long videos. The main challenge for long video QA is the memory constraint. He et al. [9] introduced a sequential framework that uses a memory bank to enhance long-term comprehension. Song et al. [33, 34] integrated video foundation models with LLMs through a memory mechanism inspired by the Atkinson-Shiffrin model, reducing computational complexity while preserving long-term memory. Another line of work improves efficiency by decomposing video content. Rui et al. [30] employed Memory-Propagated Streaming Encoding to segment videos into short clips, with Adaptive Memory Selection enhancing response accuracy by identifying question-relevant memories. Min et al. [24] introduced a multi-stage, training-free framework, emphasizing task decomposition into parsing, grounding, and reasoning stages. More re + +cently, Weng et al. [41] proposed a hierarchical framework that encodes local features and integrates global semantics for detailed comprehension of extended video content. To further reduce model's hallucination in QA, Sun et al. [35] proposed a question-guided pipeline by focusing on relevant frames and controlled answer generation. + +Temporal localization for long videos. Recent research in temporal localization for long videos has explored two primary directions: LLM-based and non-LLM-based approaches. With the advances in LLMs, researchers have expanded its use beyond traditional VQA, leveraging the capabilities for temporal grounding tasks, with a primary focus on enhancing localization accuracy. Ren et al. [32] proposed a timestamp-aware model that aligns visual content with temporal cues, enabling adaptive processing of sequential events for tasks like localization. Similarly, Fan et al. [5] introduced a memory-enhanced framework that captures contextual information across video segments, improving temporal and spatial reasoning. Korbar et al. [15] developed a text-guided resampling mechanism that dynamically selects video segments, focusing on relevant scenes to enhance temporal and spatial comprehension. + +In contrast, non-LLM-based approaches typically rely on training a regression layer or decoder for temporal localization, with efforts concentrated on refining visual features and multimodal fusion to improve alignment. For instance, Hou et al. [11] introduced a hierarchical framework that combines coarse scanning with fine-grained alignment to optimize both precision and efficiency in localizing target moments. Pan et al. [28] applied a coarse-to-fine pipeline for single-pass temporal grounding that improves both efficiency and alignment. Additionally, Mu et al. [26] proposed a cost-effective late fusion approach paired with a video-centric sampling scheme to improve scalability. + +Unlike prior works that focus on specific tasks, SEAL proposes a unified and generic framework for long video understanding, capable of adapting to various prediction heads and tasks. Additionally, SEAL supports both global and streaming modes, making it adaptable to arbitrarily long videos. + +# 3. Semantic Decomposition and Attention Learning + +Let $\mathcal{V} = \{v_{i}\}_{i = 1}^{T_{V}}$ be an arbitrarily long untrimmed video, where $v_{1},\ldots ,v_{T_{V}}$ denote the sequence of $T_{V}$ frames forming the video. Let $q$ be a query from long video understanding tasks, comprising a sequence of $l_{q}$ tokens. The query $q$ may take different forms depending on the task, such as natural language text for video question and answering (e.g., MovieChatQA [33], LVBench [40]) or visual/text template or action label for episodic memory tasks in egocentric video understanding (e.g., Ego4D [8]). Our method establishes a unified video representation to gener + +alize across these diverse long video understanding tasks. An overview of our approach is presented in Figure 2. + +# 3.1. Semantic Decomposition of Long Videos + +The main challenge in long video representation lies in capturing diverse content within limited memory. Conventional methods resort to frame sampling [42, 44] or maintaining a memory bank that merges similar frames [33, 34]. However, these approaches can vary greatly across tasks. We propose a novel decomposition approach that structures long videos into three distinct token types representing different type of semantic entities: (1) Scene tokens $\mathbf{T}_{\mathrm{scene}}$ capture background context, providing essential cues about the environment. (2) Action tokens $\mathbf{T}_{\mathrm{action}}$ represent moving elements, focusing on temporal information such as motions, activities, or events. (3) Object tokens $\mathbf{T}_{\mathrm{object}}$ , highlight key static elements relevant to specific tasks. With this structured tokenization, we create a unified, compact, task-agnostic representation that minimizes the need for redundant and dense video storage while preserving a comprehensive understanding of long video content. + +Scene Tokens. Any frame in a video can serve as a scene token because it captures the environment where it was recorded, e.g., indoor gym, outdoor mountain, etc. Although shot boundary detection could split a long video into shots with one scene token per shot, this approach has two drawbacks: (1) Shot detection algorithms are often imperfect, with no opportunity to correct once shot boundary is determined. (2) Shot detection often ignores the specific query and downstream tasks, and thus could cause suboptimal performance. To overcome these drawbacks, we propose a two-step approach. First, we over-sample scene tokens, and then later perform an attention learning step to maximize query relevance and token diversity. Specifically, we uniformly sample $N_{\mathrm{scene}}$ frames to capture a diverse background. These pre-sampled scene tokens, denoted as $\mathbf{T}_{\mathrm{scene}} = [t_i^{\mathrm{scene}}]_{i=1}^{N_{\mathrm{scene}}}$ , will undergo another round of attention learning, detailed in Section 3.2. + +Action Tokens. The purpose of the action token is to capture temporal information of moving objects such as low-level motions, activities, and events. We begin by using a class-agnostic object tracker, e.g., SAM-2 [31], to extract multiple initial dynamic tracklets. Tracklets shorter than $L_{\mathrm{min}}$ are discarded, while those longer than $L_{\mathrm{max}}$ are split into multiple tracklets with length of $L_{\mathrm{max}}$ . For each tracklet, we take the spatial union of bounding boxes across frames, allowing the dynamic token to capture not only motion information but also the spatial movement of people and objects. This process yields $N_{\mathrm{tracklet}}$ tracklets, denoted as $\mathbf{T}_{\mathrm{action}} = \{\tau_{\mathrm{dynamic}}^{i}\}_{i=1}^{N_{\mathrm{tracklet}}}$ . + +Object Tokens. For the object tokens, we utilize a class-agnostic grouping method such as SAM [14] to generate masks for all objects in each frame. This class-agnostic + +![](images/f743335994fba06e3aff28542a42bd82535cbf3c548349218c6fe930b9e2ef49.jpg) +Figure 2. SEAL Overview. During semantic decomposition, a long video $\mathcal{V}$ is decomposed into semantic tokens representing scenes, objects, and actions. Then, during attention learning, these tokens and the query q, are optimized for query relevance $R(\cdot)$ and token diversity $S(\cdot)$ . The resulting attended token subset is then passed to a vision or an MLLM head for predictions. + +segmentation approach enables comprehensive object information capture. Specifically, we apply this process on $N_{\mathrm{key}}$ key frames. On each of the $N_{\mathrm{key}}$ frames, we apply SAM to obtain in total $N_{\mathrm{object}}$ object masks, which we denote as $\mathbf{T}_{\mathrm{object}} = \{\mathbf{M}_{\mathrm{object}}^i\}_{i=1}^{N_{\mathrm{object}}}$ . Key frame selection varies by task and is detailed in the following sections. + +# 3.2. Attention Learning + +Although the long video is decomposed into different token representations, the resulting tokens are still redundant. To address this, we propose a sampling approach that balances between query-relevance and token-diversity, denoted as Attention Learning. Specifically, we formulate our sampling as an optimization problem with two main objectives: query-relevance and token-diversity. + +$$ +\begin{array}{l} T_{s}^{*} = \operatorname *{arg max}_{T_{s}\subset T_{G}}F_{s}(T_{s}|T_{G},q) \\ = \underset {T _ {s} \subset T _ {G}} {\arg \max} \alpha \sum_ {t _ {s} \in T _ {s}} R (t _ {s}, q) + (1 - \alpha) \sum_ {t _ {i}, t _ {j} \in T _ {s}, i \neq j} \frac {1}{S (t _ {i} , t _ {j})}. \\ \end{array} +$$ + +We aim to solve a subset selection problem: find a fixed-size subset $T_{s} \subset T_{G}$ that maximizes the objective function $F_{s}(T_{s}|T_{G},q)$ . Here, $T_{G}$ represents the set of all tokens used, which can be any set of tokens such as scene $\mathbf{T}_{\mathrm{scene}}$ , object $\mathbf{T}_{\mathrm{object}}$ , action $\mathbf{T}_{\mathrm{action}}$ , or a combination of all $(\mathbf{T}_{\mathrm{scene}} \cup \mathbf{T}_{\mathrm{object}} \cup \mathbf{T}_{\mathrm{action}})$ . $q$ denotes the query being asked by the downstream understanding task. And, $T_{s}^{*}$ denotes the optimal subset that maximizes $F_{s}$ . $F_{s}$ is decomposed into two terms: $R(\cdot)$ measures the relevance between a vi + +sual token and the query which we compute by encoding them with the BLIP-2 model [17] and calculating their cosine similarity; and $S(\cdot)$ calculates the cosine similarity of the paired tokens. We note that the first term maximizes the relevance between selected tokens and the query while the second term enforces token diversity (via minimizing the token pairwise similarity). Finally, $\alpha$ is a hyper-parameter to balance between query-relevance and token-diversity. + +# 3.3. Streaming and Global Mode + +Our proposed method offers two ways of representing long videos: streaming and global. In the global mode, the model fully "watches" (or processes) the entire video and then provides a single representation. In contrast, in the streaming mode, the model processes the video buffer by buffer and provides an updated representation at any given time step. The partly-observed video representation can change over time as the video progresses. This setting simulates the situation when you watch a movie together with a child and have interactions with her or him as the movie is still going on. + +In the global model, all tokens undergo a single optimization through Attention Leaning, resulting in $T_{\mathrm{sub}}$ , which includes $k$ tokens. This $T_{\mathrm{sub}}$ is regarded as the representation for the entire video. We note that the global mode may not scale well with arbitrary long videos because all tokens cannot be fit into a limited memory for sampling. One can opt to use more aggressive uniform temporal presampling to reduce the numbers of tokens before Attention Learning. This workaround can bypass the memory lim + +itation, but may also lead to sub-optimal solutions due to missing important tokens due to uniform sampling. + +Alternatively, we propose an online streaming approach for representing partly-observed videos. Specifically, we use a fixed-size sliding window with a size set to the maximum number of tokens $l$ allowed by memory capacity, denoted as $T$ . At each step $t$ , we apply Attention Learning to the union set of the tokens in the current window $T_{t}$ and the previous selected subset of tokens $T_{\mathrm{sub}}^{t - 1}$ and obtain the representation of the video at time $T_{\mathrm{sub}}^{t}$ . At the beginning, the selected subset is set to empty $(T_{\mathrm{sub}}^{0} = \emptyset)$ . + +$$ +T _ {\text {s u b}} ^ {t} = \text {A t t e n t i o n L e a r n i n g} \left(T _ {t} \cup T _ {\text {s u b}} ^ {t - 1}\right) \forall t > 0. \tag {1} +$$ + +This streaming mode allows us to use $T_{\mathrm{sub}}^{t}$ as the partly-observed representation of the video and can be fed into any prediction head for video understanding tasks. As an immediate benefit of the streaming mode, our proposed representation now can handle arbitrary long videos. + +# 3.4. Prediction Heads + +Our unified representation is adaptable to most long video understanding tasks using different prediction heads. In this paper, we demonstrate two specific use cases of our representation: one is used with the traditional vision head for video temporal grounding and the other one is with the Multimodal LLMs (MLLM) head for video QA. + +Temporal Grounding with Vision Heads. Given a query $q$ , the task is to locate the start and end times, $t_{\mathrm{start}}$ and $t_{\mathrm{end}}$ , where the answers could be deduced. We first encode the sampled tokens $T_{\mathrm{sub}}$ and the query using encoders $\mathbb{E}_V$ and $\mathbb{E}_q$ to obtain embeddings for each video token $z_v^i = \mathbb{E}_V(T_{\mathrm{sub}^i})$ and query $z_q = \mathbb{E}_q(q)$ . The cross-modal fusion is then performed to obtain the fused representation: + +$$ +z _ {\text {j o i n t}} ^ {i} = \operatorname {C r o s s M o d a l F u s i o n} \left(z _ {v} ^ {i}, z _ {q}\right) \tag {2} +$$ + +Finally, a moment decoder is applied to predict the start and end time for the query $q$ . The moment decoder includes a classification head and a regression head. The classification head is used to predict the score $p_{score}^{i}$ of each token, while the regression head predicts the normalized distances $(d_{\mathrm{start}}^{i}, d_{\mathrm{end}}^{i})$ from each token to the moment boundaries: + +$$ +p _ {\text {s c o r e}} ^ {i}, \left(d _ {\text {s t a r t}} ^ {i}, d _ {\text {e n d}} ^ {i}\right) = \text {M o m e n t D e c o d e r} \left(z _ {\text {j o i n t}} ^ {i}\right). \tag {3} +$$ + +The regression and classification heads are optimized using an IoU distance and a focal loss as used in [25]. The final moment is calculated as $t_{start}^{i}, t_{end}^{i} = (t_{i} - d_{start}^{i}) \times L_{V}, (t_{i} + d_{end}^{i}) \times L_{V}$ , where $L_{V}$ is the length of the input video. The proposed unified representation enables efficient handling of long video sequences and allows for localization related tasks. + +Video QA using MLLM Heads. The proposed representation can also be connected to an MLLM head, making it applicable to various video QA-related tasks, such as reasoning, understanding, and summarization. Moreover, some grounding tasks can also be addressed in a QA format. Specifically, $z_v$ is projected through an MLP to a visual token $T_v^{\mathrm{MLLM}}$ that the MLLM can interpret, which is then input into the MLLM along with text tokens $T_{text}$ . Based on different benchmarks, the MLLM performs multiple-choice or open-ended answering. + +# 4. Experiments + +# 4.1. Implementation Details + +Datasets and metrics. We evaluate SEAL on three datasets, each selected for its relevance to long video understanding on different capabilities. LVBench [40] contains 1,549 QA pairs across six tasks, with videos averaging 4,101 seconds (approximate 1 hour 8 minutes). Each question presents a single-choice format with four options. Accuracy serves as the evaluation metric for individual tasks as well as overall performance across all tasks. We primarily focus on this dataset due to its emphasis on hour-long videos. Moviechat-1K [33] includes 1,000 video clips with dense captions spanning 15 categories, averaging 564 seconds (about 10 minutes). The benchmark employs LLMs, specifically GPT-3.5 [27], to evaluate the quality of generated answers. A rating ranging from 0 to 5 is used to compute the overall score, while a binary preference from the LLM is used to calculate the accuracy. Ego4D-NLQ [8] is a part of the Ego4D Episodic Memory challenge for Natural Language Queries (NLQ) task. This dataset requires localizing a temporal window where the answers can be deduced from untrimmed egocentric videos. It contains 1,259 videos, averaging 10 minutes each. Our experiments apply memory constraints to simulate long video scenarios. Metrics include Top-1 and Top-5 recall at various thresholds. + +Experiment setup. We fine-tune the projection layers and Q-former [17] to adapt to different types of tokens for 20 epochs using the training split of MovieChat-1K. For Ego4D-NLQ, we use the training split to finetune the vision heads for 7 epochs. The AdamW optimizer [21] is employed with default beta values of (0.9, 0.999) and a weight decay of 0.05. For token extraction, we use SAM2 [31] to obtain object tokens, YOLOv10-X [37] with BoT-SORT [1] for action tokens. Scene, action, and object tokens are extracted at 8, 10, 1 FPS for MovieChat-1K, Ego4D-NLQ and LVbench. For the LVBench, we follow the settings of InternVL2 and utilize Yi-34B [43]. For MovieChat-1K, we adopt the same settings as [33] and use Vicuna-7B as [33]. For Ego4D-NLQ, we adhere to the settings specified in SnAG [25] and use EgoVLP [19] as vision encoder. We use 0.9 as the default value for the only hyper-parameter $\alpha$ . + +
ModelLLM SizeOverall (%)KIR (%)EU (%)Sum (%)ER (%)Rea (%)TG (%)
Qwen2-VL [39]72B41.338.341.146.638.046.541.4
InternVL2 [3]34B39.643.439.741.437.442.531.4
LLaVA-NeXT [45]34B32.234.131.227.630.135.031.4
Oryx [20]34B30.432.129.227.630.134.029.1
PLLaVA [42]34B26.126.224.925.925.030.021.4
SEAL (Ours)34B45.951.541.339.747.943.332.3
+ +Table 1. Comparison with state-of-the-art on LVBench. Our method achieves the highest overall score, with notable gains in Key Information Retrieval (KIR) and Entity Recognition (ER), demonstrating the effectiveness of our representation in locating key information and entities by eliminating redundancy. The best methods are highlighted in bold, and the second-best are underlined. + +
Model#TokenR@1R@5
0.30.5Avg0.30.5Avg
SnAGall15.7210.7813.2538.3927.4432.92
SnAG45013.449.2311.3434.0223.0428.53
SEAL45013.789.2611.5234.7923.1028.95
SnAG20010.036.358.1926.5616.9021.73
SEAL20010.837.068.9527.3917.4122.40
+ +# 4.2. Comparison with State-of-the-arts + +We conduct quantitative analysis on LVBench dataset which contains videos on lengths averaged more than 1 hour. Additionally, we demonstrate the generalizability of the proposed unified representation using MovieChat-1K, which has open-ended QA questions on a variety of scenarios, and Ego4D-NLQ, where we use traditional vision decoders (LLM-independent) for the temporal grounding task. LVBench. Table 1 shows the accuracy of our method on different categories of LVBench dataset [40], demonstrating that SEAL achieves the highest overall score, notably outperforming even larger models like Qwen2-VL72B [39] by $4.6\%$ . Our method excels in Key Information Retrieval (KIR) and Entity Recognition (ER), outperforming the strongest alternatives by $8.1\%$ and $5.1\%$ respectively. These results highlight that our action and object tokens effectively locate key information and entities by removing redundancies. While larger LLMs often exhibit better performance, this analysis suggests that it is not the only determinant, and a unified representation like ours can achieve state-of-the-art results with fewer parameters. + +Figure 3 presents a qualitative analysis of SEAL on LVBench, demonstrating that our approach is able to pay attention to relevant semantic tokens and make correct answers to different types of questions. SEAL accurately locates relevant tokens, e.g., identifying the royal family's stool color (Q1.a), counting meals eaten (Q2.a), or deter + +Table 2. Comparisons with SoTA methods on Ego4D-NLQ. Quantitative results for temporal grounding on Ego4D episodic memory Natural Language Queries (NLQ) task show our method, SEAL, consistently outperforms SnAG [25] in all metrics under memory constraints with varying number of tokens. + +
MethodsAccuracy(%)Score
Zero-shotVideoChat [18]61.03.34
VideoLLaMA [44]51.43.10
Video-ChatGPT [23]44.22.71
MovieChat [33]67.83.81
SupervisedTimeChat-Hal [35]73.83.58
HERMES [6]84.94.40
SEAL (Ours)86.84.35
+ +Table 3. Comparison with SoTA methods on MovieChat-1K. SEAL outperforms the second best method, HERMES, by $1.9\%$ on accuracy while being comparable on the score metric. + +mining scene or location (Q2.b). We also show a failure case (Q2.c), where the nuanced "why" question requires complicated causal relations of different scenes. + +Moviechat-1K. Table 3 compares our method with state-of-the-art methods on MovieChat-1K. We focus on evaluating performance in the global mode, which assesses the model's ability to comprehend information from the entire video, rather than the breakpoint mode, which primarily evaluates its ability to answer questions related to specific timestamps. Our method surpasses all the alternatives, achieving the highest accuracy and a strong score. Notably, unlike LVBench, which employs multi-choice questions, this dataset evaluates generated answers using an off-the-shelf LLM. This highlights the ability of our learned representations to effectively transfer to downstream generative tasks, enabling generating accurate and detailed responses. + +Ego4D-NLQ. We also evaluate our approach on the Ego4D-NLQ task, which differs from previous datasets as it uses a decoder to locate events for temporal grounding, rather than relying on an LLM to generate text-based answers. Table 2 shows the results under memory-constrained conditions, where we limit token numbers to simulate real-world memory limitations in long video understanding. Our method consistently achieves the highest recalls across various thresholds. With further memory constraints (e.g., reducing tokens to 200), SEAL outperforms the current SoTA method, SnAG [25], by a even larger margin. This demonstrates the robustness of our unified representation on various downstream applications for long video understanding. + +![](images/fc3d656346e45dbdda0e6e2f7bdef1786a9aa941c14e18f539ecc60ab7a3dff9.jpg) +Figure 3. Qualitative results on LVBench. Two long videos visualized with questions, multiple choice options, and SEAL predicted answers. SEAL attends to relevant entities such as "royal family" and "stool" (Q1.a), different "meals" and "drinks" (Q2.a), "scene" and "location" (Q2.b) and correctly answers these questions. Although attending to relevant "push-up" activity (Q2.c), SEAL fails to predict the right answer due to the challenging in the causal reasoning question. + +# 4.3. Ablation Study + +Semantic decomposition. We analyze the impact of the three proposed token types on LVBench (Table 4). To prevent LLMs from leveraging prior knowledge to "guess" answers using only the text query, we present the baseline that uses random token inputs, isolating the LLM's performance. Results indicate that using each token type individually significantly improves performance, highlighting the value of each token's contribution. Notably, scene tokens provide the most substantial improvement, aligning with prior research [2, 36, 41] that emphasizes frame-based scene tokens. Furthermore, different token combinations yield additional gains: adding action tokens enhances Key Information Retrieval (KIR), Reasoning (Rea) and Temporal Grounding (TG) tasks by capturing temporal dynamics in the queries, while object tokens boost Entity Recognition (ER) performance by retaining detailed object-specific + +information. We observe high variance in the Summarization (Sum) accuracy due to the small number of questions in this category and the randomness of LLM. Similar improvements have been observed on the MovieChat-1K in Table 5, where the accuracy improved by $10.45\%$ with action and object tokens. Ultimately, using all three token types together achieves the highest performance, demonstrating the complementary strengths of scene, action, and object tokens as a unified representation in long video understanding. + +We observe that the latency bottleneck mostly comes from the tracker when processing multiple objects. Table 6 presents an analysis of different trackers for extracting action tokens. While YOLO+BoT-SORT operates in a much higher frame rate (14 FPS), it lacks open-set detection capabilities. On the other hand, SAM2 can detect all objects within a scene but operates significantly slower (8FPS). Our findings indicate that the proposed action tokens does not + +
ModelOverallKIREUSumERReaTG
Random tokens24.824.725.234.523.526.420.9
Action only34.033.331.736.234.335.833.6
Object only33.934.432.631.033.440.829.5
Scene only42.850.540.625.944.341.327.2
Scene+Object43.447.440.024.147.642.335.5
Scene+Action44.453.340.024.146.243.338.2
SEAL (Ours)45.951.541.339.747.943.332.3
+ +Table 4. The effects of different types of tokens on LVbench. Using any type of tokens outperforms the random baseline, while applying all three types of tokens brings the best performance. + +
MethodSceneScene+ObjectScene+ActionSEAL
Accuracy76.3379.4981.4686.78
Score4.154.264.284.35
+ +require capturing all the objects in the scene. In practice, YOLO-X pretrained on COCO, with BoT-SORT, effectively captures essential information for action token extraction, maintaining a balance between performance and efficiency. Attention learning. Table 6 highlights the effectiveness of query relevance and token diversity terms in Section 3.2. When $\alpha = 0$ , the optimization focuses on the diversity term to capture extensive contextual information for long-video tasks such as Summarization (Sum) and Reasoning (Rea), achieving reasonable results comparable to other state-of-the-art approaches in Table 1. With $\alpha = 1$ , the optimization prioritizes the relevance term to reduce redundancy by sampling tokens that closely align with the query, so performance can be improved for tasks like Key Information Retrieval (KIR). However, this approach loses global context and results in weaker summarization performance. Balancing both terms yields the best overall results. + +Streaming mode. We compare the global mode and streaming mode of our method on LVBench. As discussed in Section §3, the streaming mode better simulates real-world scenarios by sequentially processing partial observations to aggregate information over arbitrarily long videos. Table 6 demonstrates that streaming mode performs worse on globally-dependent tasks such as Event Understanding (EU) and Key Information Retrieval (KIR). However, it excels in temporally-intensive tasks like Temporal Grounding (TG) and Reasoning (Rea). Notably, streaming mode still surpasses the most competitive baseline, even when the baseline uses a significantly larger LLM. + +Computational Complexity. We provide an accuracy-efficiency trade-off comparison in Fig 4 where both SEAL and InternVL2 are with a varying number of tokens. FPS is calculated during inference that includes all the processing time from the raw video frames. SEAL reduces the re + +Table 5. The effects of different types of tokens on MovieChat1K. Results are on the test set with our global inference mode. Using all three types of tokens provides the best accuracy. + +
ModelOverallKIREUSumERReaTG
SEAL (Ours)45.951.541.339.747.943.332.3
rep Yolo w/ SAM243.351.939.929.346.440.332.7
⇒ Streaming44.250.939.737.945.044.334.1
Diversity only38.740.438.840.038.740.933.7
Relevance only42.048.539.927.644.037.331.8
+ +Table 6. Ablation studies on LVBench. SEAL defaults to global mode with YOLO for efficiency, while SAM2 performs similarly. Global mode excels in long-context tasks, and streaming mode in temporally-intensive tasks. Combining Relevance and Diversity in attention learning achieves the best performance. + +![](images/2de949438a1e2de40a862c4683a57e79b78dac320b3a53ac7e48bc57c04eacd1.jpg) +Figure 4. Accuracy vs. efficiency trade-off on LVBench. SEAL runs $2 - 3\mathrm{x}$ faster than InternVL2 at the same accuracy, and is more accurate when compared at the same FPS. + +liaence on densely sampled tokens and uses subject-level and motion aware representations/tokens to improve efficiency while maintain good accuracy. From Fig 4, at 10 FPS, SEAL is about $5\%$ more accurate than InternVL2. When comparing at the same accuracy of $41.5\%$ and $40\%$ , SEAL is about 3x and 2x faster than InternVL2, respectively. + +# 5. Conclusion and Future Work + +We propose SEAL, a novel unified representation for long video understanding that addresses computational complexity, temporal redundancy, and cross-task generalization. SEAL leverages semantic decomposition to break videos into scene, object, and action tokens, reducing redundancy and enabling efficient processing. It incorporates attention learning to balance query relevance and token diversity, enhancing performance across diverse tasks. SEAL achieves state-of-the-art results on benchmarks like MovieChat-1K, LVBench, and Ego4D-NLQ, demonstrating its versatility and effectiveness for long video understanding. + +Limitation. The Attention Learning module is bounded by the memory constraint for the QP solver, making it not fully end-to-end trainable. For future work, we plan to integrate Attention Learning in our streaming mode for full end-to-end learning. Developing special prediction heads to solve causal reasoning [16], where MLLM heads showed limitations, is also an interesting area for future exploration. + +Acknowledgements. This work was partially supported by the Office of Naval Research (award #N00014-23-1-2417). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of ONR. + +# References + +[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. ArXiv, abs/2206.14651, 2022. 5 +[2] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In CVPR, 2022. 7 +[3] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6 +[4] Meghomala Das, David M Bennett, and Gordon N Dutton. Visual attention as an important visual function: an outline of manifestations, diagnosis and management of impaired visual attention. British Journal of Ophthalmology, 91(11): 1556-1560, 2007. 1 +[5] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In ECCV, pages 75-92. Springer, 2024. 3 +[6] Gueter Josmy Faure, Jia-Fong Yeh, Min-Hung Chen, Hung-Ting Su, Shang-Hong Lai, and Winston H. Hsu. Hermes: temporal-coherent long-form understanding with episodes and semantics. arXiv preprint arXiv:2408.17443, 2024. 6 +[7] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In ICCV, 2017. 1 +[8] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonio Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawa + +har, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 5 +[9] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, 2024. 2 +[10] Fabian Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1 +[11] Zhijian Hou, Wanjun Zhong, Lei Ji, Difei Gao, Kun Yan, Wing-Kwong Chan, Chong-Wah Ngo, Zheng Shou, and Nan Duan. Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918, 2022.3 +[12] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos "in the wild". CoRR, abs/1604.06182, 2016. 1 +[13] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. CoRR, abs/1705.06950, 2017. 1 +[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 3 +[15] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In ECCV, pages 271-288. Springer, 2024. 3 +[16] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 8 +[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML. PMLR, 2023. 4, 5 +[18] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6 +[19] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022.5 +[20] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 6 +[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 + +[22] Steven J. Luck and Michelle Ford. On the role of selective attention in visual perception. Proceedings of the National Academy of Sciences of the United States of America, 95(3): 825-830, 1998. 1 +[23] Muhammad Maaz, Hanoona Abdul Rasheed, Salman H. Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In AMACL, 2023. 6 +[24] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In CVPR, pages 13235-13245, 2024. 2 +[25] Fangzhou Mu, Sicheng Mo, and Yin Li. SnAG: Scalable and accurate video grounding. In CVPR, 2024. 5, 6 +[26] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 3 +[27] OpenAI. Gpt-3.5, 2023. 5 +[28] Yulin Pan, Xiangteng He, Biao Gong, Yiliang Lv, Yujun Shen, Yuxin Peng, and Deli Zhao. Scanning only once: An end-to-end framework for fast temporal grounding in long videos. In ICCV, pages 13767-13777, 2023. 3 +[29] Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. Egovieo: Exploring egocentric foundation model and downstream adaptation. arXiv preprint arXiv:2406.18070, 2024. 2 +[30] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024. 2 +[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos, 2024. 3, 5 +[32] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 3 +[33] E. Song, W. Chai, G. Wang, Y. Zhang, H. Zhou, F. Wu, X. Guo, T. Ye, Y. Lu, JN. Hwang, and G. Wang. MovieChat: From dense token to sparse memory for long video understanding. In CVPR, 2024. 2, 3, 5, 6 +[34] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 2, 3 +[35] Yiwei Sun, Zhihang Liu, Chuanbin Liu, Bowei Pu, Zhihan Zhang, and Hongtao Xie. Hallucination mitigation prompts long-term video understanding. arXiv preprint arXiv:2406.11333, 2024. 3, 6 +[36] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In CVPR, 2024. 7 + +[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection. ArXiv, abs/2405.14458, 2024. 5 +[38] Lan Wang, Gaurav Mittal, Sandra Sajeev, Ye Yu, Matthew Hall, Vishnu Naresh Boddeti, and Mei Chen. Protege: Untrimmed pretraining for video temporal grounding by video temporal grounding. In CVPR, 2023. 1 +[39] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6 +[40] W. Wang, Z. He, W. Hong, Y. Cheng, X. Zhang, J. Qi, S. Huang, B. Xu, Y. Dong, M. Ding, and J. Tang. LVBench: An extreme long video understanding benchmark, 2024. 1, 2, 3, 5, 6 +[41] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 1, 3, 7 +[42] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6 +[43] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, Xiaohui Hu, Xiaoyi Ren, Xinyao Niu, Pengcheng Nie, Yuchi Xu, Yudong Liu, Yue Wang, Yuxuan Cai, Zhenyu Gu, Zhiyuan Liu, and Zonghong Dai. Yi: Open foundation models by 01.ai. ArXiv, abs/2403.04652, 2024. 5 +[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 3, 6 +[45] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6 \ No newline at end of file diff --git a/2025/SEAL_ Semantic Attention Learning for Long Video Representation/images.zip b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d164aba3366179c11955c6deed2b1ea7192a1640 --- /dev/null +++ b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84e977cd0c2621dadb5f8209b67c4a9c1630647e592dcfd1c29c7b883d4be6f9 +size 599282 diff --git a/2025/SEAL_ Semantic Attention Learning for Long Video Representation/layout.json b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..17bf047ade87cb3bf5d413a61a597f9e95de7a60 --- /dev/null +++ b/2025/SEAL_ Semantic Attention Learning for Long Video Representation/layout.json @@ -0,0 +1,8641 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 100, + 108, + 122 + ], + "blocks": [ + { + "bbox": [ + 85, + 100, + 108, + 122 + ], + "lines": [ + { + "bbox": [ + 85, + 100, + 108, + 122 + ], + "spans": [ + { + "bbox": [ + 85, + 100, + 108, + 122 + ], + "type": "image", + "image_path": "6c54110e18f1ff4f83f3abe5fb0be8f6f1a2943486504e70db095fcd940d7e26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 109, + 529, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 109, + 529, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 529, + 126 + ], + "type": "text", + "content": "SEAL: Semiclastic Attention Learning for Long Video Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "spans": [ + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": "Lan Wang" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": " Yujia Chen" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": " Du Tran" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": " Vishnu Naresh Boddeti" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": " Wen-Sheng Chu" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": " Michigan State University \n" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 147, + 531, + 177 + ], + "type": "text", + "content": " Google" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 178, + 486, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 178, + 486, + 190 + ], + "spans": [ + { + "bbox": [ + 120, + 178, + 486, + 190 + ], + "type": "text", + "content": "{wanglan3, vishnu}@msu.edu {yujiachen, tranldu, wschu}@google.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 217, + 200, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 230 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 230 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 243, + 297, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 297, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 297, + 483 + ], + "type": "text", + "content": "Long video understanding presents challenges due to the inherent high computational complexity and redundant temporal information. An effective representation for long videos must efficiently process such redundancy while preserving essential contents for downstream tasks. This paper introduces SEmatic Attention Learning (SEAL), a novel unified representation for long videos. To reduce computational complexity, long videos are decomposed into three distinct types of semantic entities: scenes, objects, and actions, allowing models to operate on a compact set of entities rather than a large number of frames or pixels. To further address redundancy, we propose an attention learning module that balances token relevance with diversity, formulated as a subset selection optimization problem. Our representation is versatile and applicable across various long video understanding tasks. Extensive experiments demonstrate that SEAL significantly outperforms state-of-the-art methods in video question answering and temporal grounding tasks across diverse benchmarks, including LVBench, MovieChat-1K, and Ego4D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 515, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 696 + ], + "type": "text", + "content": "State-of-the-art video understanding models excel at short video tasks such as video classification [7, 13], temporal grounding [38] and action detection [10, 12], which involve videos lasting from a few seconds to minutes. However, their performance declines on hour-long videos [40, 41]. In contrast, a 10-year-old child can watch a full-length movie (one to two hours) and effortlessly answer questions at various levels of detail. This disparity between humans and machines emphasizes the foundational challenges in long video understanding for AI models, including: (1) Increased complexity: Long videos require more computation and memory than current hardware can support for training or inference, (2) Temporal redundancy: Slow-changing scenes and objects introduce significant redundancy, and (3) Cross-task generalization: A robust repre" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 219, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 255 + ], + "type": "text", + "content": "sentation must adapt to various tasks, from fine-grained fact retrieval to high-level reasoning. These challenges, which appear trivial for a child, remain challenging for AI models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 257, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 472 + ], + "type": "text", + "content": "How does the human brain process long videos, particularly in addressing the above-mentioned challenges? First, rather than processing every pixel or frame, the brain selectively attends to new information to efficiently manage temporal redundancy [4, 22]. Second, humans process videos in an online fashion, continuously updating their understanding and memories as they watch, rather than deferring reasoning until the end. This continuous knowledge update, combined with selective attention, allows the brain to efficiently handle the complexity of long videos. Finally, attention dynamically shifts based on context. Without specific guidance, a child may focus on naturally engaging or memorable moments. However, when given specific questions in advance, attention becomes goal-oriented to seek relevant details while maintaining a broad understanding. This suggests effective representations should balance between task-specific focus and holistic understanding of the video to enable cross-task generalization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": "Inspired by how humans process long videos, we introduce SEmantic Attention Learning (SEAL), a novel unified representation designed to tackle the three key challenges in long video understanding. SEAL consists of two main steps: Semantic Decomposition and Attention Learning. In Semantic Decomposition, long videos are decomposed into three semantic entities such as scenes, objects, and actions, which are then treated as tokens. While the scene and object tokens represent static content assumed not to change rapidly, the action tokens are designed to capture the dynamic, fast-changing moments of the video. We note that these semantic tokens efficiently encode the essential information needed to answer \"where\", \"what\", or \"how\" questions about the videos. This decomposition significantly reduces complexity by allowing AI models to operate on a compact set of tokens instead of raw pixels or frames. Figure 1(a) illustrates a conventional uniformly sampled video " + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": ", where redundant frames create cluttered visual information that hinders effective analysis for both models and humans. Figure 1(b) shows our semantic decomposition," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 280, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 280, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 280, + 714 + ], + "type": "text", + "content": "*This work was done as Lan Wang's internship project at Google." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26192" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 85, + 516, + 158 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "type": "text", + "content": "(a) Uniform sampling of a long video " + }, + { + "bbox": [ + 67, + 72, + 214, + 83 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 85, + 516, + 158 + ], + "lines": [ + { + "bbox": [ + 93, + 85, + 516, + 158 + ], + "spans": [ + { + "bbox": [ + 93, + 85, + 516, + 158 + ], + "type": "image", + "image_path": "e376a8c5c37aa7672a656904d522374fcf96f305a9d4174a4544970c9b6bf9ab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 160, + 246, + 171 + ], + "lines": [ + { + "bbox": [ + 69, + 160, + 246, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 246, + 171 + ], + "type": "text", + "content": "(b) Semantic decomposition of the same video " + }, + { + "bbox": [ + 69, + 160, + 246, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 182, + 133, + 210 + ], + "blocks": [ + { + "bbox": [ + 74, + 171, + 122, + 182 + ], + "lines": [ + { + "bbox": [ + 74, + 171, + 122, + 182 + ], + "spans": [ + { + "bbox": [ + 74, + 171, + 122, + 182 + ], + "type": "text", + "content": "Scenes (static)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 182, + 133, + 210 + ], + "lines": [ + { + "bbox": [ + 82, + 182, + 133, + 210 + ], + "spans": [ + { + "bbox": [ + 82, + 182, + 133, + 210 + ], + "type": "image", + "image_path": "bb45a9f5db8e5a1015b0c03a9cae93ffbaad42b71a359c6424a3f3af6da7c8ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 211, + 112, + 218 + ], + "lines": [ + { + "bbox": [ + 103, + 211, + 112, + 218 + ], + "spans": [ + { + "bbox": [ + 103, + 211, + 112, + 218 + ], + "type": "text", + "content": "S1" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 182, + 190, + 210 + ], + "blocks": [ + { + "bbox": [ + 138, + 182, + 190, + 210 + ], + "lines": [ + { + "bbox": [ + 138, + 182, + 190, + 210 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 190, + 210 + ], + "type": "image", + "image_path": "4430aff250ab55c574d0e11b3d01d8a4104e847aecbc8d82fcf97ebc3acd2dc9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 211, + 168, + 217 + ], + "lines": [ + { + "bbox": [ + 159, + 211, + 168, + 217 + ], + "spans": [ + { + "bbox": [ + 159, + 211, + 168, + 217 + ], + "type": "text", + "content": "S2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 195, + 182, + 246, + 210 + ], + "blocks": [ + { + "bbox": [ + 195, + 182, + 246, + 210 + ], + "lines": [ + { + "bbox": [ + 195, + 182, + 246, + 210 + ], + "spans": [ + { + "bbox": [ + 195, + 182, + 246, + 210 + ], + "type": "image", + "image_path": "3b368e06436a23df4c216b91bdbf69267f17f41a19a6de9876ab9a4776386c63.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 211, + 224, + 217 + ], + "lines": [ + { + "bbox": [ + 215, + 211, + 224, + 217 + ], + "spans": [ + { + "bbox": [ + 215, + 211, + 224, + 217 + ], + "type": "text", + "content": "S3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 251, + 182, + 302, + 210 + ], + "blocks": [ + { + "bbox": [ + 251, + 182, + 302, + 210 + ], + "lines": [ + { + "bbox": [ + 251, + 182, + 302, + 210 + ], + "spans": [ + { + "bbox": [ + 251, + 182, + 302, + 210 + ], + "type": "image", + "image_path": "7a11ac0bd5adae8ff7092b5f35c482185a9f1173f3cd6c23cd7e9e4eb6b40d44.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 211, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 276, + 211, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 276, + 211, + 286, + 217 + ], + "type": "text", + "content": "S4" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 186, + 348, + 213 + ], + "blocks": [ + { + "bbox": [ + 316, + 171, + 372, + 182 + ], + "lines": [ + { + "bbox": [ + 316, + 171, + 372, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 171, + 372, + 182 + ], + "type": "text", + "content": "Objects (static)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 186, + 348, + 213 + ], + "lines": [ + { + "bbox": [ + 328, + 186, + 348, + 213 + ], + "spans": [ + { + "bbox": [ + 328, + 186, + 348, + 213 + ], + "type": "image", + "image_path": "62df9a031fe037ddbe4e763f15cc213a54b1063b2cde4366e5a724f5c8e68a92.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 213, + 341, + 219 + ], + "lines": [ + { + "bbox": [ + 327, + 213, + 341, + 219 + ], + "spans": [ + { + "bbox": [ + 327, + 213, + 341, + 219 + ], + "type": "text", + "content": "01" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 353, + 189, + 399, + 209 + ], + "blocks": [ + { + "bbox": [ + 353, + 189, + 399, + 209 + ], + "lines": [ + { + "bbox": [ + 353, + 189, + 399, + 209 + ], + "spans": [ + { + "bbox": [ + 353, + 189, + 399, + 209 + ], + "type": "image", + "image_path": "cf5411fcd1ed2b2307eadaa0915ffab70e28c25da68192c14fbd5a3e8fb1082e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 211, + 380, + 218 + ], + "lines": [ + { + "bbox": [ + 369, + 211, + 380, + 218 + ], + "spans": [ + { + "bbox": [ + 369, + 211, + 380, + 218 + ], + "type": "text", + "content": "02" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 402, + 189, + 429, + 209 + ], + "blocks": [ + { + "bbox": [ + 402, + 189, + 429, + 209 + ], + "lines": [ + { + "bbox": [ + 402, + 189, + 429, + 209 + ], + "spans": [ + { + "bbox": [ + 402, + 189, + 429, + 209 + ], + "type": "image", + "image_path": "8bef0394388c1efaf48b92b127da581f826a7d0b0aa4d72843ae52fd29b3dc2a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 211, + 422, + 218 + ], + "lines": [ + { + "bbox": [ + 410, + 211, + 422, + 218 + ], + "spans": [ + { + "bbox": [ + 410, + 211, + 422, + 218 + ], + "type": "text", + "content": "03" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 440, + 186, + 465, + 209 + ], + "blocks": [ + { + "bbox": [ + 440, + 186, + 465, + 209 + ], + "lines": [ + { + "bbox": [ + 440, + 186, + 465, + 209 + ], + "spans": [ + { + "bbox": [ + 440, + 186, + 465, + 209 + ], + "type": "image", + "image_path": "e89a7bf3469a622daec24e880f4ef50506dd7660a06a8c27a2bdccb92b1765a0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 210, + 457, + 218 + ], + "lines": [ + { + "bbox": [ + 447, + 210, + 457, + 218 + ], + "spans": [ + { + "bbox": [ + 447, + 210, + 457, + 218 + ], + "type": "text", + "content": "04" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 468, + 186, + 488, + 208 + ], + "blocks": [ + { + "bbox": [ + 468, + 186, + 488, + 208 + ], + "lines": [ + { + "bbox": [ + 468, + 186, + 488, + 208 + ], + "spans": [ + { + "bbox": [ + 468, + 186, + 488, + 208 + ], + "type": "image", + "image_path": "df85f832aef5638952ac25ce0b3af698934f391557cf6943f0816dc43b776630.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 210, + 488, + 218 + ], + "lines": [ + { + "bbox": [ + 477, + 210, + 488, + 218 + ], + "spans": [ + { + "bbox": [ + 477, + 210, + 488, + 218 + ], + "type": "text", + "content": "05" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 493, + 182, + 533, + 209 + ], + "blocks": [ + { + "bbox": [ + 493, + 182, + 533, + 209 + ], + "lines": [ + { + "bbox": [ + 493, + 182, + 533, + 209 + ], + "spans": [ + { + "bbox": [ + 493, + 182, + 533, + 209 + ], + "type": "image", + "image_path": "6e3b0724e15b656553c253e615f66f79a1e6fb1c9c97a13c256e2ed280c8aad5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 210, + 525, + 218 + ], + "lines": [ + { + "bbox": [ + 514, + 210, + 525, + 218 + ], + "spans": [ + { + "bbox": [ + 514, + 210, + 525, + 218 + ], + "type": "text", + "content": "06" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 72, + 222, + 164, + 280 + ], + "blocks": [ + { + "bbox": [ + 72, + 222, + 164, + 280 + ], + "lines": [ + { + "bbox": [ + 72, + 222, + 164, + 280 + ], + "spans": [ + { + "bbox": [ + 72, + 222, + 164, + 280 + ], + "type": "image", + "image_path": "0de4353d795382f2a4127a8c6d82f141adc2277343ce3df2743e8009dcda911d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 282, + 194, + 293 + ], + "lines": [ + { + "bbox": [ + 69, + 282, + 194, + 293 + ], + "spans": [ + { + "bbox": [ + 69, + 282, + 194, + 293 + ], + "type": "text", + "content": "(c) Query-aware attention learning" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "type": "text", + "content": "Figure 1. Long Video Representation with Semantic Attention Learning (SEAL): (a) Conventional uniform sampling results in redundant and cluttered visual information, making it difficult for both AI models and human brains to process efficiently. (b) Decomposing long videos into semantic entities such as scenes, objects, and actions reduces temporal redundancy, thus making model training and inference more efficient. In this example, the long video " + }, + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "type": "text", + "content": " is decomposed into 4 scene tokens (S1-S4), 6 object tokens (O1-O6), and 4 action/event tokens (A1-A4). (c) Query-aware attention learning module improves downstream task performance by focusing on relevant information rather than processing everything. Queries (Q1-Q4) are shown with their most relevant tokens. (best viewed in color)" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 169, + 233, + 274, + 269 + ], + "blocks": [ + { + "bbox": [ + 169, + 233, + 274, + 269 + ], + "lines": [ + { + "bbox": [ + 169, + 233, + 274, + 269 + ], + "spans": [ + { + "bbox": [ + 169, + 233, + 274, + 269 + ], + "type": "image", + "image_path": "82b5b95f132b0db4daee7c99b1dbd1f6a1e04a55ef4093c071c954bef278b790.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 271, + 222, + 278 + ], + "lines": [ + { + "bbox": [ + 212, + 271, + 222, + 278 + ], + "spans": [ + { + "bbox": [ + 212, + 271, + 222, + 278 + ], + "type": "text", + "content": "A2" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 282, + 225, + 400, + 270 + ], + "blocks": [ + { + "bbox": [ + 282, + 225, + 400, + 270 + ], + "lines": [ + { + "bbox": [ + 282, + 225, + 400, + 270 + ], + "spans": [ + { + "bbox": [ + 282, + 225, + 400, + 270 + ], + "type": "image", + "image_path": "f326d91f32ce2dc7c6a6a24008362b8a25066b80c776c161e5667f12aa40c04c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 271, + 349, + 278 + ], + "lines": [ + { + "bbox": [ + 338, + 271, + 349, + 278 + ], + "spans": [ + { + "bbox": [ + 338, + 271, + 349, + 278 + ], + "type": "text", + "content": "A3" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 422, + 226, + 530, + 267 + ], + "blocks": [ + { + "bbox": [ + 422, + 226, + 530, + 267 + ], + "lines": [ + { + "bbox": [ + 422, + 226, + 530, + 267 + ], + "spans": [ + { + "bbox": [ + 422, + 226, + 530, + 267 + ], + "type": "image", + "image_path": "6a375c373de1ff5ddebd5a3d778ffdeec1026a813758fdd46adafd3f8c23f3ed.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 271, + 482, + 278 + ], + "lines": [ + { + "bbox": [ + 471, + 271, + 482, + 278 + ], + "spans": [ + { + "bbox": [ + 471, + 271, + 482, + 278 + ], + "type": "text", + "content": "A4" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 71, + 294, + 181, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 294, + 181, + 312 + ], + "spans": [ + { + "bbox": [ + 71, + 294, + 181, + 312 + ], + "type": "text", + "content": "Q1: What does the vlogger family find in visitor center?" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 72, + 314, + 93, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 314, + 93, + 321 + ], + "spans": [ + { + "bbox": [ + 72, + 314, + 93, + 321 + ], + "type": "text", + "content": "S101" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 187, + 294, + 306, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 294, + 306, + 313 + ], + "spans": [ + { + "bbox": [ + 187, + 294, + 306, + 313 + ], + "type": "text", + "content": "Q2: What does the vlogger family do after buying fishing supplies?" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 187, + 314, + 231, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 314, + 231, + 322 + ], + "spans": [ + { + "bbox": [ + 187, + 314, + 231, + 322 + ], + "type": "text", + "content": "S2 S3 O2 A1" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 309, + 294, + 430, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 294, + 430, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 294, + 430, + 312 + ], + "type": "text", + "content": "Q3: What foods do the vlogger family eat the most here?" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 309, + 314, + 361, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 314, + 361, + 322 + ], + "spans": [ + { + "bbox": [ + 309, + 314, + 361, + 322 + ], + "type": "text", + "content": "O3 O4 O5 A3" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 434, + 294, + 533, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 434, + 294, + 533, + 312 + ], + "spans": [ + { + "bbox": [ + 434, + 294, + 533, + 312 + ], + "type": "text", + "content": "Q4: What is the most common form of transportation there?" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 434, + 314, + 484, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 434, + 314, + 484, + 322 + ], + "spans": [ + { + "bbox": [ + 434, + 314, + 484, + 322 + ], + "type": "text", + "content": "S4 O6 A2 A4" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 501, + 317, + 539, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 501, + 317, + 539, + 327 + ], + "spans": [ + { + "bbox": [ + 501, + 317, + 539, + 327 + ], + "type": "text", + "content": "Query set" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 71, + 329, + 232, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 329, + 232, + 341 + ], + "spans": [ + { + "bbox": [ + 71, + 329, + 232, + 341 + ], + "type": "text", + "content": "Scene token Object token Action token" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 55, + 421, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 576 + ], + "type": "text", + "content": "breaking down the video into distinct scenes, static objects, and dynamic actions. In Attention Learning, we formulate a subset selection problem that maximizes the query relevance while ensuring token diversity. This step not only mitigates redundancy but also enhances cross-task generalization by prioritizing the most informative tokens. Figure 1(c) illustrates our attention learning module with four different queries and their selected tokens that capture relevance and diverse video content. Finally, SEAL is designed to work for both global and streaming modes, enabling it to process arbitrarily long videos. Extensive ablations and experiments confirm SEAL's superior performance over existing methods. Our key contributions include:" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 56, + 579, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 48, + "blocks": [ + { + "bbox": [ + 56, + 579, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 614 + ], + "type": "text", + "content": "- We introduce SEAL, a novel unified representation for long videos by decomposing them into three semantic tokens, namely scenes, objects, and actions." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 56, + 616, + 294, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 294, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 294, + 662 + ], + "type": "text", + "content": "- Our attention learning module reduces temporal redundancy while supporting strong cross-task generalization. We show SEAL works in both global and streaming modes, making it adaptable to arbitrarily long videos." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 56, + 664, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 664, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 664, + 294, + 712 + ], + "type": "text", + "content": "- SEALoutperforms state-of-the-art methods on various long video understanding tasks and benchmarks including: video QA (MovieChat-1K [33], LVBench [40]), and egocentric video grounding (Ego4D [8])." + } + ] + } + ], + "index": 47 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 420, + 400, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 400, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 400, + 431 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 438, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 533 + ], + "type": "text", + "content": "Recent approaches have unified various video understanding tasks by framing them as video QA tasks [40], leveraging the capabilities of LLMs. However, fine-tuning with task-specific vision heads continues to offer advantages in memory efficiency and task-specific performance [29], particularly for temporal grounding. In this section, we review advancements in Video Question Answering (QA) and temporal grounding for long video understanding." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "Video QA for long videos. The main challenge for long video QA is the memory constraint. He et al. [9] introduced a sequential framework that uses a memory bank to enhance long-term comprehension. Song et al. [33, 34] integrated video foundation models with LLMs through a memory mechanism inspired by the Atkinson-Shiffrin model, reducing computational complexity while preserving long-term memory. Another line of work improves efficiency by decomposing video content. Rui et al. [30] employed Memory-Propagated Streaming Encoding to segment videos into short clips, with Adaptive Memory Selection enhancing response accuracy by identifying question-relevant memories. Min et al. [24] introduced a multi-stage, training-free framework, emphasizing task decomposition into parsing, grounding, and reasoning stages. More re" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26193" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "cently, Weng et al. [41] proposed a hierarchical framework that encodes local features and integrates global semantics for detailed comprehension of extended video content. To further reduce model's hallucination in QA, Sun et al. [35] proposed a question-guided pipeline by focusing on relevant frames and controlled answer generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 335 + ], + "type": "text", + "content": "Temporal localization for long videos. Recent research in temporal localization for long videos has explored two primary directions: LLM-based and non-LLM-based approaches. With the advances in LLMs, researchers have expanded its use beyond traditional VQA, leveraging the capabilities for temporal grounding tasks, with a primary focus on enhancing localization accuracy. Ren et al. [32] proposed a timestamp-aware model that aligns visual content with temporal cues, enabling adaptive processing of sequential events for tasks like localization. Similarly, Fan et al. [5] introduced a memory-enhanced framework that captures contextual information across video segments, improving temporal and spatial reasoning. Korbar et al. [15] developed a text-guided resampling mechanism that dynamically selects video segments, focusing on relevant scenes to enhance temporal and spatial comprehension." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 336, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 479 + ], + "type": "text", + "content": "In contrast, non-LLM-based approaches typically rely on training a regression layer or decoder for temporal localization, with efforts concentrated on refining visual features and multimodal fusion to improve alignment. For instance, Hou et al. [11] introduced a hierarchical framework that combines coarse scanning with fine-grained alignment to optimize both precision and efficiency in localizing target moments. Pan et al. [28] applied a coarse-to-fine pipeline for single-pass temporal grounding that improves both efficiency and alignment. Additionally, Mu et al. [26] proposed a cost-effective late fusion approach paired with a video-centric sampling scheme to improve scalability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 551 + ], + "type": "text", + "content": "Unlike prior works that focus on specific tasks, SEAL proposes a unified and generic framework for long video understanding, capable of adapting to various prediction heads and tasks. Additionally, SEAL supports both global and streaming modes, making it adaptable to arbitrarily long videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 295, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 588 + ], + "type": "text", + "content": "3. Semantic Decomposition and Attention Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{i}\\}_{i = 1}^{T_{V}}" + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": " be an arbitrarily long untrimmed video, where " + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "inline_equation", + "content": "v_{1},\\ldots ,v_{T_{V}}" + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": " denote the sequence of " + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "inline_equation", + "content": "T_{V}" + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": " frames forming the video. Let " + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": " be a query from long video understanding tasks, comprising a sequence of " + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "inline_equation", + "content": "l_{q}" + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": " tokens. The query " + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 593, + 295, + 714 + ], + "type": "text", + "content": " may take different forms depending on the task, such as natural language text for video question and answering (e.g., MovieChatQA [33], LVBench [40]) or visual/text template or action label for episodic memory tasks in egocentric video understanding (e.g., Ego4D [8]). Our method establishes a unified video representation to gener" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "alize across these diverse long video understanding tasks. An overview of our approach is presented in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 106, + 525, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 106, + 525, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 106, + 525, + 119 + ], + "type": "text", + "content": "3.1. Semantic Decomposition of Long Videos" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "text", + "content": "The main challenge in long video representation lies in capturing diverse content within limited memory. Conventional methods resort to frame sampling [42, 44] or maintaining a memory bank that merges similar frames [33, 34]. However, these approaches can vary greatly across tasks. We propose a novel decomposition approach that structures long videos into three distinct token types representing different type of semantic entities: (1) Scene tokens " + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{scene}}" + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "text", + "content": " capture background context, providing essential cues about the environment. (2) Action tokens " + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{action}}" + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "text", + "content": " represent moving elements, focusing on temporal information such as motions, activities, or events. (3) Object tokens " + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{object}}" + }, + { + "bbox": [ + 313, + 125, + 553, + 328 + ], + "type": "text", + "content": ", highlight key static elements relevant to specific tasks. With this structured tokenization, we create a unified, compact, task-agnostic representation that minimizes the need for redundant and dense video storage while preserving a comprehensive understanding of long video content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "type": "text", + "content": "Scene Tokens. Any frame in a video can serve as a scene token because it captures the environment where it was recorded, e.g., indoor gym, outdoor mountain, etc. Although shot boundary detection could split a long video into shots with one scene token per shot, this approach has two drawbacks: (1) Shot detection algorithms are often imperfect, with no opportunity to correct once shot boundary is determined. (2) Shot detection often ignores the specific query and downstream tasks, and thus could cause suboptimal performance. To overcome these drawbacks, we propose a two-step approach. First, we over-sample scene tokens, and then later perform an attention learning step to maximize query relevance and token diversity. Specifically, we uniformly sample " + }, + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{scene}}" + }, + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "type": "text", + "content": " frames to capture a diverse background. These pre-sampled scene tokens, denoted as " + }, + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{scene}} = [t_i^{\\mathrm{scene}}]_{i=1}^{N_{\\mathrm{scene}}}" + }, + { + "bbox": [ + 313, + 329, + 553, + 532 + ], + "type": "text", + "content": ", will undergo another round of attention learning, detailed in Section 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "content": "Action Tokens. The purpose of the action token is to capture temporal information of moving objects such as low-level motions, activities, and events. We begin by using a class-agnostic object tracker, e.g., SAM-2 [31], to extract multiple initial dynamic tracklets. Tracklets shorter than " + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "content": " are discarded, while those longer than " + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "content": " are split into multiple tracklets with length of " + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "content": ". For each tracklet, we take the spatial union of bounding boxes across frames, allowing the dynamic token to capture not only motion information but also the spatial movement of people and objects. This process yields " + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{tracklet}}" + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "content": " tracklets, denoted as " + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{action}} = \\{\\tau_{\\mathrm{dynamic}}^{i}\\}_{i=1}^{N_{\\mathrm{tracklet}}}" + }, + { + "bbox": [ + 313, + 533, + 553, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Object Tokens. For the object tokens, we utilize a class-agnostic grouping method such as SAM [14] to generate masks for all objects in each frame. This class-agnostic" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26194" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 76, + 521, + 305 + ], + "blocks": [ + { + "bbox": [ + 80, + 76, + 521, + 305 + ], + "lines": [ + { + "bbox": [ + 80, + 76, + 521, + 305 + ], + "spans": [ + { + "bbox": [ + 80, + 76, + 521, + 305 + ], + "type": "image", + "image_path": "f743335994fba06e3aff28542a42bd82535cbf3c548349218c6fe930b9e2ef49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. SEAL Overview. During semantic decomposition, a long video " + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": " is decomposed into semantic tokens representing scenes, objects, and actions. Then, during attention learning, these tokens and the query q, are optimized for query relevance " + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": " and token diversity " + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": ". The resulting attended token subset is then passed to a vision or an MLLM head for predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "text", + "content": "segmentation approach enables comprehensive object information capture. Specifically, we apply this process on " + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{key}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "text", + "content": " key frames. On each of the " + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{key}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "text", + "content": " frames, we apply SAM to obtain in total " + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{object}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "text", + "content": " object masks, which we denote as " + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{object}} = \\{\\mathbf{M}_{\\mathrm{object}}^i\\}_{i=1}^{N_{\\mathrm{object}}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 431 + ], + "type": "text", + "content": ". Key frame selection varies by task and is detailed in the following sections." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 169, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 169, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 169, + 450 + ], + "type": "text", + "content": "3.2. Attention Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 539 + ], + "type": "text", + "content": "Although the long video is decomposed into different token representations, the resulting tokens are still redundant. To address this, we propose a sampling approach that balances between query-relevance and token-diversity, denoted as Attention Learning. Specifically, we formulate our sampling as an optimization problem with two main objectives: query-relevance and token-diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 545, + 293, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 293, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 293, + 598 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T_{s}^{*} = \\operatorname *{arg max}_{T_{s}\\subset T_{G}}F_{s}(T_{s}|T_{G},q) \\\\ = \\underset {T _ {s} \\subset T _ {G}} {\\arg \\max} \\alpha \\sum_ {t _ {s} \\in T _ {s}} R (t _ {s}, q) + (1 - \\alpha) \\sum_ {t _ {i}, t _ {j} \\in T _ {s}, i \\neq j} \\frac {1}{S (t _ {i} , t _ {j})}. \\\\ \\end{array}", + "image_path": "a8b25ed04327821080f49e89f54c0ced1e579534327f4f02e682efd82aaba286.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We aim to solve a subset selection problem: find a fixed-size subset " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "T_{s} \\subset T_{G}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " that maximizes the objective function " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{s}(T_{s}|T_{G},q)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "T_{G}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " represents the set of all tokens used, which can be any set of tokens such as scene " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{scene}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", object " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{object}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", action " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{action}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", or a combination of all " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{T}_{\\mathrm{scene}} \\cup \\mathbf{T}_{\\mathrm{object}} \\cup \\mathbf{T}_{\\mathrm{action}})" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " denotes the query being asked by the downstream understanding task. And, " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "T_{s}^{*}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " denotes the optimal subset that maximizes " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is decomposed into two terms: " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " measures the relevance between a vi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "content": "sual token and the query which we compute by encoding them with the BLIP-2 model [17] and calculating their cosine similarity; and " + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "content": " calculates the cosine similarity of the paired tokens. We note that the first term maximizes the relevance between selected tokens and the query while the second term enforces token diversity (via minimizing the token pairwise similarity). Finally, " + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "content": " is a hyper-parameter to balance between query-relevance and token-diversity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 471, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 471, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 471, + 469 + ], + "type": "text", + "content": "3.3. Streaming and Global Mode" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 474, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 605 + ], + "type": "text", + "content": "Our proposed method offers two ways of representing long videos: streaming and global. In the global mode, the model fully \"watches\" (or processes) the entire video and then provides a single representation. In contrast, in the streaming mode, the model processes the video buffer by buffer and provides an updated representation at any given time step. The partly-observed video representation can change over time as the video progresses. This setting simulates the situation when you watch a movie together with a child and have interactions with her or him as the movie is still going on." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "In the global model, all tokens undergo a single optimization through Attention Leaning, resulting in " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sub}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", which includes " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " tokens. This " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sub}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " is regarded as the representation for the entire video. We note that the global mode may not scale well with arbitrary long videos because all tokens cannot be fit into a limited memory for sampling. One can opt to use more aggressive uniform temporal presampling to reduce the numbers of tokens before Attention Learning. This workaround can bypass the memory lim" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26195" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "itation, but may also lead to sub-optimal solutions due to missing important tokens due to uniform sampling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": "Alternatively, we propose an online streaming approach for representing partly-observed videos. Specifically, we use a fixed-size sliding window with a size set to the maximum number of tokens " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": " allowed by memory capacity, denoted as " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": ". At each step " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": ", we apply Attention Learning to the union set of the tokens in the current window " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": " and the previous selected subset of tokens " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sub}}^{t - 1}" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": " and obtain the representation of the video at time " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sub}}^{t}" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": ". At the beginning, the selected subset is set to empty " + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "inline_equation", + "content": "(T_{\\mathrm{sub}}^{0} = \\emptyset)" + }, + { + "bbox": [ + 55, + 96, + 296, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 213, + 295, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 213, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 79, + 213, + 295, + 229 + ], + "type": "interline_equation", + "content": "T _ {\\text {s u b}} ^ {t} = \\text {A t t e n t i o n L e a r n i n g} \\left(T _ {t} \\cup T _ {\\text {s u b}} ^ {t - 1}\\right) \\forall t > 0. \\tag {1}", + "image_path": "7a02b3884e96143138a5cb8633672768c88c70dcfb2b9d961592780f1f4283c9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 299 + ], + "type": "text", + "content": "This streaming mode allows us to use " + }, + { + "bbox": [ + 55, + 238, + 296, + 299 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sub}}^{t}" + }, + { + "bbox": [ + 55, + 238, + 296, + 299 + ], + "type": "text", + "content": " as the partly-observed representation of the video and can be fed into any prediction head for video understanding tasks. As an immediate benefit of the streaming mode, our proposed representation now can handle arbitrary long videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 160, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 160, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 160, + 319 + ], + "type": "text", + "content": "3.4. Prediction Heads" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 324, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 396 + ], + "type": "text", + "content": "Our unified representation is adaptable to most long video understanding tasks using different prediction heads. In this paper, we demonstrate two specific use cases of our representation: one is used with the traditional vision head for video temporal grounding and the other one is with the Multimodal LLMs (MLLM) head for video QA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": "Temporal Grounding with Vision Heads. Given a query " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": ", the task is to locate the start and end times, " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{start}}" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{end}}" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": ", where the answers could be deduced. We first encode the sampled tokens " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{sub}}" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": " and the query using encoders " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_V" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_q" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": " to obtain embeddings for each video token " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "z_v^i = \\mathbb{E}_V(T_{\\mathrm{sub}^i})" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "inline_equation", + "content": "z_q = \\mathbb{E}_q(q)" + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": ". The cross-modal fusion is then performed to obtain the fused representation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 492, + 295, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 492, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 295, + 509 + ], + "type": "interline_equation", + "content": "z _ {\\text {j o i n t}} ^ {i} = \\operatorname {C r o s s M o d a l F u s i o n} \\left(z _ {v} ^ {i}, z _ {q}\\right) \\tag {2}", + "image_path": "6a042ef160bba8f47705774e121553b26dbcc3dbe56c5400910301a32f795874.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "content": "Finally, a moment decoder is applied to predict the start and end time for the query " + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "content": ". The moment decoder includes a classification head and a regression head. The classification head is used to predict the score " + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "inline_equation", + "content": "p_{score}^{i}" + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "content": " of each token, while the regression head predicts the normalized distances " + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{start}}^{i}, d_{\\mathrm{end}}^{i})" + }, + { + "bbox": [ + 55, + 512, + 296, + 586 + ], + "type": "text", + "content": " from each token to the moment boundaries:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 81, + 605, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 605, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 81, + 605, + 295, + 622 + ], + "type": "interline_equation", + "content": "p _ {\\text {s c o r e}} ^ {i}, \\left(d _ {\\text {s t a r t}} ^ {i}, d _ {\\text {e n d}} ^ {i}\\right) = \\text {M o m e n t D e c o d e r} \\left(z _ {\\text {j o i n t}} ^ {i}\\right). \\tag {3}", + "image_path": "ebbd57e5eafd5757ba002268d3ca6cee4774112c7a20c0428f660487e51a67ee.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": "The regression and classification heads are optimized using an IoU distance and a focal loss as used in [25]. The final moment is calculated as " + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "inline_equation", + "content": "t_{start}^{i}, t_{end}^{i} = (t_{i} - d_{start}^{i}) \\times L_{V}, (t_{i} + d_{end}^{i}) \\times L_{V}" + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "inline_equation", + "content": "L_{V}" + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": " is the length of the input video. The proposed unified representation enables efficient handling of long video sequences and allows for localization related tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "Video QA using MLLM Heads. The proposed representation can also be connected to an MLLM head, making it applicable to various video QA-related tasks, such as reasoning, understanding, and summarization. Moreover, some grounding tasks can also be addressed in a QA format. Specifically, " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "z_v" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " is projected through an MLP to a visual token " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "T_v^{\\mathrm{MLLM}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " that the MLLM can interpret, which is then input into the MLLM along with text tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "T_{text}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". Based on different benchmarks, the MLLM performs multiple-choice or open-ended answering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 395, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 395, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 395, + 218 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 224, + 447, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 447, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 447, + 237 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 245, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 519 + ], + "type": "text", + "content": "Datasets and metrics. We evaluate SEAL on three datasets, each selected for its relevance to long video understanding on different capabilities. LVBench [40] contains 1,549 QA pairs across six tasks, with videos averaging 4,101 seconds (approximate 1 hour 8 minutes). Each question presents a single-choice format with four options. Accuracy serves as the evaluation metric for individual tasks as well as overall performance across all tasks. We primarily focus on this dataset due to its emphasis on hour-long videos. Moviechat-1K [33] includes 1,000 video clips with dense captions spanning 15 categories, averaging 564 seconds (about 10 minutes). The benchmark employs LLMs, specifically GPT-3.5 [27], to evaluate the quality of generated answers. A rating ranging from 0 to 5 is used to compute the overall score, while a binary preference from the LLM is used to calculate the accuracy. Ego4D-NLQ [8] is a part of the Ego4D Episodic Memory challenge for Natural Language Queries (NLQ) task. This dataset requires localizing a temporal window where the answers can be deduced from untrimmed egocentric videos. It contains 1,259 videos, averaging 10 minutes each. Our experiments apply memory constraints to simulate long video scenarios. Metrics include Top-1 and Top-5 recall at various thresholds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "Experiment setup. We fine-tune the projection layers and Q-former [17] to adapt to different types of tokens for 20 epochs using the training split of MovieChat-1K. For Ego4D-NLQ, we use the training split to finetune the vision heads for 7 epochs. The AdamW optimizer [21] is employed with default beta values of (0.9, 0.999) and a weight decay of 0.05. For token extraction, we use SAM2 [31] to obtain object tokens, YOLOv10-X [37] with BoT-SORT [1] for action tokens. Scene, action, and object tokens are extracted at 8, 10, 1 FPS for MovieChat-1K, Ego4D-NLQ and LVbench. For the LVBench, we follow the settings of InternVL2 and utilize Yi-34B [43]. For MovieChat-1K, we adopt the same settings as [33] and use Vicuna-7B as [33]. For Ego4D-NLQ, we adhere to the settings specified in SnAG [25] and use EgoVLP [19] as vision encoder. We use 0.9 as the default value for the only hyper-parameter " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26196" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 531, + 169 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 531, + 169 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 531, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 531, + 169 + ], + "type": "table", + "html": "
ModelLLM SizeOverall (%)KIR (%)EU (%)Sum (%)ER (%)Rea (%)TG (%)
Qwen2-VL [39]72B41.338.341.146.638.046.541.4
InternVL2 [3]34B39.643.439.741.437.442.531.4
LLaVA-NeXT [45]34B32.234.131.227.630.135.031.4
Oryx [20]34B30.432.129.227.630.134.029.1
PLLaVA [42]34B26.126.224.925.925.030.021.4
SEAL (Ours)34B45.951.541.339.747.943.332.3
", + "image_path": "8943efc0c357d83b21c5a8a9d9fa5ac61b8eb189b6da2e871bb197e1117c7f65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 215, + 297, + 308 + ], + "blocks": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art on LVBench. Our method achieves the highest overall score, with notable gains in Key Information Retrieval (KIR) and Entity Recognition (ER), demonstrating the effectiveness of our representation in locating key information and entities by eliminating redundancy. The best methods are highlighted in bold, and the second-best are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 215, + 297, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 215, + 297, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 215, + 297, + 308 + ], + "type": "table", + "html": "
Model#TokenR@1R@5
0.30.5Avg0.30.5Avg
SnAGall15.7210.7813.2538.3927.4432.92
SnAG45013.449.2311.3434.0223.0428.53
SEAL45013.789.2611.5234.7923.1028.95
SnAG20010.036.358.1926.5616.9021.73
SEAL20010.837.068.9527.3917.4122.40
", + "image_path": "68923e8b7db31f54e20cca12735062c24b57276770f4354cddb0b7895d1895a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 382, + 239, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 239, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 239, + 395 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "text", + "content": "We conduct quantitative analysis on LVBench dataset which contains videos on lengths averaged more than 1 hour. Additionally, we demonstrate the generalizability of the proposed unified representation using MovieChat-1K, which has open-ended QA questions on a variety of scenarios, and Ego4D-NLQ, where we use traditional vision decoders (LLM-independent) for the temporal grounding task. LVBench. Table 1 shows the accuracy of our method on different categories of LVBench dataset [40], demonstrating that SEAL achieves the highest overall score, notably outperforming even larger models like Qwen2-VL72B [39] by " + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "text", + "content": ". Our method excels in Key Information Retrieval (KIR) and Entity Recognition (ER), outperforming the strongest alternatives by " + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 54, + 400, + 295, + 640 + ], + "type": "text", + "content": " respectively. These results highlight that our action and object tokens effectively locate key information and entities by removing redundancies. While larger LLMs often exhibit better performance, this analysis suggests that it is not the only determinant, and a unified representation like ours can achieve state-of-the-art results with fewer parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "content": "Figure 3 presents a qualitative analysis of SEAL on LVBench, demonstrating that our approach is able to pay attention to relevant semantic tokens and make correct answers to different types of questions. SEAL accurately locates relevant tokens, e.g., identifying the royal family's stool color (Q1.a), counting meals eaten (Q2.a), or deter" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 215, + 556, + 320 + ], + "blocks": [ + { + "bbox": [ + 55, + 310, + 296, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 365 + ], + "type": "text", + "content": "Table 2. Comparisons with SoTA methods on Ego4D-NLQ. Quantitative results for temporal grounding on Ego4D episodic memory Natural Language Queries (NLQ) task show our method, SEAL, consistently outperforms SnAG [25] in all metrics under memory constraints with varying number of tokens." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 215, + 556, + 320 + ], + "lines": [ + { + "bbox": [ + 315, + 215, + 556, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 215, + 556, + 320 + ], + "type": "table", + "html": "
MethodsAccuracy(%)Score
Zero-shotVideoChat [18]61.03.34
VideoLLaMA [44]51.43.10
Video-ChatGPT [23]44.22.71
MovieChat [33]67.83.81
SupervisedTimeChat-Hal [35]73.83.58
HERMES [6]84.94.40
SEAL (Ours)86.84.35
", + "image_path": "59040e26d7681e787c1d683a1514124ce6ef53b77813f7eee398254ca278d899.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 322, + 555, + 356 + ], + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 356 + ], + "type": "text", + "content": "Table 3. Comparison with SoTA methods on MovieChat-1K. SEAL outperforms the second best method, HERMES, by " + }, + { + "bbox": [ + 313, + 322, + 555, + 356 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 322, + 555, + 356 + ], + "type": "text", + "content": " on accuracy while being comparable on the score metric." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 366, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 402 + ], + "type": "text", + "content": "mining scene or location (Q2.b). We also show a failure case (Q2.c), where the nuanced \"why\" question requires complicated causal relations of different scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 557 + ], + "type": "text", + "content": "Moviechat-1K. Table 3 compares our method with state-of-the-art methods on MovieChat-1K. We focus on evaluating performance in the global mode, which assesses the model's ability to comprehend information from the entire video, rather than the breakpoint mode, which primarily evaluates its ability to answer questions related to specific timestamps. Our method surpasses all the alternatives, achieving the highest accuracy and a strong score. Notably, unlike LVBench, which employs multi-choice questions, this dataset evaluates generated answers using an off-the-shelf LLM. This highlights the ability of our learned representations to effectively transfer to downstream generative tasks, enabling generating accurate and detailed responses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Ego4D-NLQ. We also evaluate our approach on the Ego4D-NLQ task, which differs from previous datasets as it uses a decoder to locate events for temporal grounding, rather than relying on an LLM to generate text-based answers. Table 2 shows the results under memory-constrained conditions, where we limit token numbers to simulate real-world memory limitations in long video understanding. Our method consistently achieves the highest recalls across various thresholds. With further memory constraints (e.g., reducing tokens to 200), SEAL outperforms the current SoTA method, SnAG [25], by a even larger margin. This demonstrates the robustness of our unified representation on various downstream applications for long video understanding." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26197" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 71, + 509, + 441 + ], + "blocks": [ + { + "bbox": [ + 96, + 71, + 509, + 441 + ], + "lines": [ + { + "bbox": [ + 96, + 71, + 509, + 441 + ], + "spans": [ + { + "bbox": [ + 96, + 71, + 509, + 441 + ], + "type": "image", + "image_path": "fc3d656346e45dbdda0e6e2f7bdef1786a9aa941c14e18f539ecc60ab7a3dff9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 444, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 555, + 489 + ], + "type": "text", + "content": "Figure 3. Qualitative results on LVBench. Two long videos visualized with questions, multiple choice options, and SEAL predicted answers. SEAL attends to relevant entities such as \"royal family\" and \"stool\" (Q1.a), different \"meals\" and \"drinks\" (Q2.a), \"scene\" and \"location\" (Q2.b) and correctly answers these questions. Although attending to relevant \"push-up\" activity (Q2.c), SEAL fails to predict the right answer due to the challenging in the causal reasoning question." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 502, + 149, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 149, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 149, + 514 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "Semantic decomposition. We analyze the impact of the three proposed token types on LVBench (Table 4). To prevent LLMs from leveraging prior knowledge to \"guess\" answers using only the text query, we present the baseline that uses random token inputs, isolating the LLM's performance. Results indicate that using each token type individually significantly improves performance, highlighting the value of each token's contribution. Notably, scene tokens provide the most substantial improvement, aligning with prior research [2, 36, 41] that emphasizes frame-based scene tokens. Furthermore, different token combinations yield additional gains: adding action tokens enhances Key Information Retrieval (KIR), Reasoning (Rea) and Temporal Grounding (TG) tasks by capturing temporal dynamics in the queries, while object tokens boost Entity Recognition (ER) performance by retaining detailed object-specific" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 502, + 555, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 610 + ], + "type": "text", + "content": "information. We observe high variance in the Summarization (Sum) accuracy due to the small number of questions in this category and the randomness of LLM. Similar improvements have been observed on the MovieChat-1K in Table 5, where the accuracy improved by " + }, + { + "bbox": [ + 313, + 502, + 555, + 610 + ], + "type": "inline_equation", + "content": "10.45\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 610 + ], + "type": "text", + "content": " with action and object tokens. Ultimately, using all three token types together achieves the highest performance, demonstrating the complementary strengths of scene, action, and object tokens as a unified representation in long video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "We observe that the latency bottleneck mostly comes from the tracker when processing multiple objects. Table 6 presents an analysis of different trackers for extracting action tokens. While YOLO+BoT-SORT operates in a much higher frame rate (14 FPS), it lacks open-set detection capabilities. On the other hand, SAM2 can detect all objects within a scene but operates significantly slower (8FPS). Our findings indicate that the proposed action tokens does not" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26198" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 165 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 165 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 165 + ], + "type": "table", + "html": "
ModelOverallKIREUSumERReaTG
Random tokens24.824.725.234.523.526.420.9
Action only34.033.331.736.234.335.833.6
Object only33.934.432.631.033.440.829.5
Scene only42.850.540.625.944.341.327.2
Scene+Object43.447.440.024.147.642.335.5
Scene+Action44.453.340.024.146.243.338.2
SEAL (Ours)45.951.541.339.747.943.332.3
", + "image_path": "7ed12d805212c17721298f170ccb13037fc1f6fb084a0b6defb7cab7185b24a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 209, + 293, + 251 + ], + "blocks": [ + { + "bbox": [ + 56, + 167, + 294, + 200 + ], + "lines": [ + { + "bbox": [ + 56, + 167, + 294, + 200 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 294, + 200 + ], + "type": "text", + "content": "Table 4. The effects of different types of tokens on LVbench. Using any type of tokens outperforms the random baseline, while applying all three types of tokens brings the best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 209, + 293, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 293, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 293, + 251 + ], + "type": "table", + "html": "
MethodSceneScene+ObjectScene+ActionSEAL
Accuracy76.3379.4981.4686.78
Score4.154.264.284.35
", + "image_path": "aa64c16d4673f39a1bc1e062087a2fb963986f5e4aed9b0b349d6bdee0b434ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "type": "text", + "content": "require capturing all the objects in the scene. In practice, YOLO-X pretrained on COCO, with BoT-SORT, effectively captures essential information for action token extraction, maintaining a balance between performance and efficiency. Attention learning. Table 6 highlights the effectiveness of query relevance and token diversity terms in Section 3.2. When " + }, + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "type": "text", + "content": ", the optimization focuses on the diversity term to capture extensive contextual information for long-video tasks such as Summarization (Sum) and Reasoning (Rea), achieving reasonable results comparable to other state-of-the-art approaches in Table 1. With " + }, + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 55, + 304, + 295, + 507 + ], + "type": "text", + "content": ", the optimization prioritizes the relevance term to reduce redundancy by sampling tokens that closely align with the query, so performance can be improved for tasks like Key Information Retrieval (KIR). However, this approach loses global context and results in weaker summarization performance. Balancing both terms yields the best overall results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 652 + ], + "type": "text", + "content": "Streaming mode. We compare the global mode and streaming mode of our method on LVBench. As discussed in Section §3, the streaming mode better simulates real-world scenarios by sequentially processing partial observations to aggregate information over arbitrarily long videos. Table 6 demonstrates that streaming mode performs worse on globally-dependent tasks such as Event Understanding (EU) and Key Information Retrieval (KIR). However, it excels in temporally-intensive tasks like Temporal Grounding (TG) and Reasoning (Rea). Notably, streaming mode still surpasses the most competitive baseline, even when the baseline uses a significantly larger LLM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Computational Complexity. We provide an accuracy-efficiency trade-off comparison in Fig 4 where both SEAL and InternVL2 are with a varying number of tokens. FPS is calculated during inference that includes all the processing time from the raw video frames. SEAL reduces the re" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 56, + 251, + 294, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 285 + ], + "type": "text", + "content": "Table 5. The effects of different types of tokens on MovieChat1K. Results are on the test set with our global inference mode. Using all three types of tokens provides the best accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 144 + ], + "type": "table", + "html": "
ModelOverallKIREUSumERReaTG
SEAL (Ours)45.951.541.339.747.943.332.3
rep Yolo w/ SAM243.351.939.929.346.440.332.7
⇒ Streaming44.250.939.737.945.044.334.1
Diversity only38.740.438.840.038.740.933.7
Relevance only42.048.539.927.644.037.331.8
", + "image_path": "a8f9f01ea6434155313f54c5aa0f7edf3e851a86379b7813417e5a3741e1697b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 145, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 201 + ], + "type": "text", + "content": "Table 6. Ablation studies on LVBench. SEAL defaults to global mode with YOLO for efficiency, while SAM2 performs similarly. Global mode excels in long-context tasks, and streaming mode in temporally-intensive tasks. Combining Relevance and Diversity in attention learning achieves the best performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 209, + 548, + 337 + ], + "blocks": [ + { + "bbox": [ + 318, + 209, + 548, + 337 + ], + "lines": [ + { + "bbox": [ + 318, + 209, + 548, + 337 + ], + "spans": [ + { + "bbox": [ + 318, + 209, + 548, + 337 + ], + "type": "image", + "image_path": "2de949438a1e2de40a862c4683a57e79b78dac320b3a53ac7e48bc57c04eacd1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 340, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 373 + ], + "type": "text", + "content": "Figure 4. Accuracy vs. efficiency trade-off on LVBench. SEAL runs " + }, + { + "bbox": [ + 313, + 340, + 555, + 373 + ], + "type": "inline_equation", + "content": "2 - 3\\mathrm{x}" + }, + { + "bbox": [ + 313, + 340, + 555, + 373 + ], + "type": "text", + "content": " faster than InternVL2 at the same accuracy, and is more accurate when compared at the same FPS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": "liaence on densely sampled tokens and uses subject-level and motion aware representations/tokens to improve efficiency while maintain good accuracy. From Fig 4, at 10 FPS, SEAL is about " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " more accurate than InternVL2. When comparing at the same accuracy of " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "41.5\\%" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": ", SEAL is about 3x and 2x faster than InternVL2, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 476, + 479, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 479, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 479, + 489 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 497, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 628 + ], + "type": "text", + "content": "We propose SEAL, a novel unified representation for long video understanding that addresses computational complexity, temporal redundancy, and cross-task generalization. SEAL leverages semantic decomposition to break videos into scene, object, and action tokens, reducing redundancy and enabling efficient processing. It incorporates attention learning to balance query relevance and token diversity, enhancing performance across diverse tasks. SEAL achieves state-of-the-art results on benchmarks like MovieChat-1K, LVBench, and Ego4D-NLQ, demonstrating its versatility and effectiveness for long video understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Limitation. The Attention Learning module is bounded by the memory constraint for the QP solver, making it not fully end-to-end trainable. For future work, we plan to integrate Attention Learning in our streaming mode for full end-to-end learning. Developing special prediction heads to solve causal reasoning [16], where MLLM heads showed limitations, is also an interesting area for future exploration." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26199" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by the Office of Naval Research (award #N00014-23-1-2417). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of ONR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 142, + 115, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 115, + 154 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 115, + 154 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 162, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 294, + 194 + ], + "type": "text", + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. ArXiv, abs/2206.14651, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 196, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 295, + 227 + ], + "type": "text", + "content": "[2] Minghao Chen, Fangyun Wei, Chong Li, and Deng Cai. Frame-wise action representations for long videos via sequence contrastive learning. In CVPR, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 282 + ], + "type": "text", + "content": "[3] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 337 + ], + "type": "text", + "content": "[4] Meghomala Das, David M Bennett, and Gordon N Dutton. Visual attention as an important visual function: an outline of manifestations, diagnosis and management of impaired visual attention. British Journal of Ophthalmology, 91(11): 1556-1560, 2007. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 339, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 295, + 382 + ], + "type": "text", + "content": "[5] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In ECCV, pages 75-92. Springer, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 384, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 294, + 426 + ], + "type": "text", + "content": "[6] Gueter Josmy Faure, Jia-Fong Yeh, Min-Hung Chen, Hung-Ting Su, Shang-Hong Lai, and Winston H. Hsu. Hermes: temporal-coherent long-form understanding with episodes and semantics. arXiv preprint arXiv:2408.17443, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 491 + ], + "type": "text", + "content": "[7] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 495, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 495, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 495, + 295, + 713 + ], + "type": "text", + "content": "[8] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonio Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawa" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 137 + ], + "type": "text", + "content": "har, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 553, + 183 + ], + "type": "text", + "content": "[9] Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xuefei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. In CVPR, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 184, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 225 + ], + "type": "text", + "content": "[10] Fabian Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. ActivityNet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[11] Zhijian Hou, Wanjun Zhong, Lei Ji, Difei Gao, Kun Yan, Wing-Kwong Chan, Chong-Wah Ngo, Zheng Shou, and Nan Duan. Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. arXiv preprint arXiv:2209.10918, 2022.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "text", + "content": "[12] Haroon Idrees, Amir R. Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The THUMOS challenge on action recognition for videos \"in the wild\". CoRR, abs/1604.06182, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "text", + "content": "[13] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. CoRR, abs/1705.06950, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In ICCV, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 469 + ], + "type": "text", + "content": "[15] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In ECCV, pages 271-288. Springer, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 472, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 502 + ], + "type": "text", + "content": "[16] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In CVPR, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 504, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 546 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML. PMLR, 2023. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "text", + "content": "[18] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wen Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "text", + "content": "[19] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022.5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "text", + "content": "[20] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26200" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[22] Steven J. Luck and Michelle Ford. On the role of selective attention in visual perception. Proceedings of the National Academy of Sciences of the United States of America, 95(3): 825-830, 1998. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[23] Muhammad Maaz, Hanoona Abdul Rasheed, Salman H. Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In AMACL, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "type": "text", + "content": "[24] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In CVPR, pages 13235-13245, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 294, + 227 + ], + "type": "text", + "content": "[25] Fangzhou Mu, Sicheng Mo, and Yin Li. SnAG: Scalable and accurate video grounding. In CVPR, 2024. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 294, + 259 + ], + "type": "text", + "content": "[26] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 262, + 173, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 262, + 173, + 272 + ], + "spans": [ + { + "bbox": [ + 57, + 262, + 173, + 272 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-3.5, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 273, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 273, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 273, + 294, + 316 + ], + "type": "text", + "content": "[28] Yulin Pan, Xiangteng He, Biao Gong, Yiliang Lv, Yujun Shen, Yuxin Peng, and Deli Zhao. Scanning only once: An end-to-end framework for fast temporal grounding in long videos. In ICCV, pages 13767-13777, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 317, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 317, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 317, + 294, + 370 + ], + "type": "text", + "content": "[29] Baoqi Pei, Guo Chen, Jilan Xu, Yuping He, Yicheng Liu, Kanghua Pan, Yifei Huang, Yali Wang, Tong Lu, Limin Wang, et al. Egovieo: Exploring egocentric foundation model and downstream adaptation. arXiv preprint arXiv:2406.18070, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "text", + "content": "[30] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. arXiv preprint arXiv:2405.16009, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 416, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 491 + ], + "type": "text", + "content": "[31] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos, 2024. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "text", + "content": "[32] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[33] E. Song, W. Chai, G. Wang, Y. Zhang, H. Zhou, F. Wu, X. Guo, T. Ye, Y. Lu, JN. Hwang, and G. Wang. MovieChat: From dense token to sparse memory for long video understanding. In CVPR, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[34] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[35] Yiwei Sun, Zhihang Liu, Chuanbin Liu, Bowei Pu, Zhihan Zhang, and Hongtao Xie. Hallucination mitigation prompts long-term video understanding. arXiv preprint arXiv:2406.11333, 2024. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[36] Reuben Tan, Xineng Sun, Ping Hu, Jui-hsien Wang, Hanieh Deilamsalehy, Bryan A Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-llm. In CVPR, 2024. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 529 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Yolov10: Real-time end-to-end object detection. ArXiv, abs/2405.14458, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[38] Lan Wang, Gaurav Mittal, Sandra Sajeev, Ye Yu, Matthew Hall, Vishnu Naresh Boddeti, and Mei Chen. Protege: Untrimmed pretraining for video temporal grounding by video temporal grounding. In CVPR, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 228 + ], + "type": "text", + "content": "[39] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 230, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 272 + ], + "type": "text", + "content": "[40] W. Wang, Z. He, W. Hong, Y. Cheng, X. Zhang, J. Qi, S. Huang, B. Xu, Y. Dong, M. Ding, and J. Tang. LVBench: An extreme long video understanding benchmark, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[41] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In ECCV, pages 453-470. Springer, 2024. 1, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[42] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 365, + 553, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 451 + ], + "type": "text", + "content": "[43] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, Kaidong Yu, Peng Liu, Qiang Liu, Shawn Yue, Senbin Yang, Shiming Yang, Tao Yu, Wen Xie, Wenhao Huang, Xiaohui Hu, Xiaoyi Ren, Xinyao Niu, Pengcheng Nie, Yuchi Xu, Yudong Liu, Yue Wang, Yuxuan Cai, Zhenyu Gu, Zhiyuan Liu, and Zonghong Dai. Yi: Open foundation models by 01.ai. ArXiv, abs/2403.04652, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "type": "text", + "content": "[44] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "type": "text", + "content": "[45] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26201" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_content_list.json b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..693b6fe4639b2f7b7fbdabea4c20b3e4d512da33 --- /dev/null +++ b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_content_list.json @@ -0,0 +1,1971 @@ +[ + { + "type": "text", + "text": "SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 178, + 138, + 818, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ye Liu $^{1,2}$ Meng Yang $^{1,2*}$", + "bbox": [ + 408, + 210, + 617, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou, Guangdong, China", + "bbox": [ + 112, + 229, + 908, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, Shaanxi, China", + "bbox": [ + 137, + 246, + 883, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuy2629@mail2.sysu.edu.cn, yangm6@mail.sysu.edu.cn", + "bbox": [ + 287, + 265, + 728, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 314, + 330, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot class-incremental learning (FSCIL) presents a significant challenge in machine learning, requiring models to integrate new classes from limited examples while preserving performance on previously learned classes. Recently, prompt-based CIL approaches leverage ample data to train prompts, effectively mitigating catastrophic forgetting. However, these methods do not account for the semantic features embedded in prompts, exacerbating the plasticity-stability dilemma in few-shot incremental learning. In this paper, we propose a novel and simple framework named SEmantic Complementary Prompt(SEC-Prompt), which learns two sets of semantically complementary prompts based on an adaptive query: discriminative prompts(D-Prompt) and non-discriminative prompts(ND-Prompt). D-Prompt enhances the separation of class-specific feature distributions by strengthening key discriminative features, while ND-Prompt balances nondiscriminative information to promote generalization to novel classes. To efficiently learn high-quality knowledge from limited samples, we leverage ND-Prompt for data augmentation to increase sample diversity and introduce Prompt Clustering Loss to prevent noise contamination in D-Prompt, ensuring robust discriminative feature learning and improved generalization. Our experimental results showcase state-of-the-art performance across three benchmark datasets, including CIFAR100, ImageNet-R and CUB datasets.", + "bbox": [ + 99, + 344, + 485, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 102, + 768, + 230, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AI applications, once deployed, frequently encounter novel classes within continuous data streams [11], necessitating the addition of these new classes to recognition tasks. This scenario drives the study of class incremental learning (CIL) [32], which incrementally learns new class con", + "bbox": [ + 99, + 792, + 483, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cepts post-deployment while preserving knowledge of previously learned classes. In more practical scenarios, new classes arise sporadically and usually lack sufficient data, for example, industrial defect detection and medical image recognition, which motivates the development of few-shot class incremental learning (FSCIL) [36] with only a limited number of samples for new classes.", + "bbox": [ + 509, + 315, + 895, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Both CIL and FSCIL face a central challenge: catastrophic forgetting [28]. As models are incrementally trained to recognize new classes, fine-tuning or expanding parameters often overwrites previously acquired knowledge, leading to significant performance drops on earlier classes. Moreover, research in FSCIL has shown that learning from limited samples can lead to overfitting [7]: the model quickly memorizes small sample sets but fails to generalize, which compromises its robustness. Recent efforts have focused on transferring knowledge from base classes to enhance generalization or on adjusting new class distributions to stabilize learning.", + "bbox": [ + 509, + 419, + 895, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent prompt-based continual learning methods [34, 43, 44] fix the backbone parameters while training additional prompt parameters using a key-query mechanism to learn task-specific knowledge. These methods store the learned knowledge in a prompt pool, effectively preventing forgetting. However, when only a few samples are available, they struggle to effectively learn the knowledge of new classes. In contrast, humans can easily learn to distinguish new categories without disrupting previously learned knowledge. This raises a key question: How do humans achieve FSCIL capabilities? The key to classification tasks lies in learning discriminative features. Humans can identify the key features of new categories while ignoring irrelevant information, which allows them to learn accurate patterns and avoid the influence of noise. Thus, we hypothesize that a selective attention mechanism plays a crucial role in this process.", + "bbox": [ + 509, + 595, + 895, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework called Semantic Complementary Prompting (SEC-Prompt) to enhance FSCIL by emulating the human selective attenuate the semantic representation of the human sentence.", + "bbox": [ + 511, + 845, + 893, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 119, + 876, + 243, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25643", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion mechanism. By designining two novel attentions, we propose to leverage extensive prior knowledge for generalization and introduce two prompt components: Discriminative Prompt (D-Prompt) and Non-Discriminative Prompt (ND-Prompt). Using distinct matching mechanisms to learn complementary knowledge, D-Prompt and ND-Prompt separately learn task-specific discriminative information and task-agnostic nondiscriminative information, addressing catastrophic forgetting and overfitting in FSCIL.", + "bbox": [ + 99, + 102, + 480, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SEC-Prompt incorporates both querying and matching mechanisms. Firstly, we use the class token from the input of each Transformer layer as an adaptive query to capture multi-level feature information, thereby enhancing the model's generalization ability and improving its capability to learn new classes. Then, the D-Prompt utilizes an attention mechanism as a matching mechanism to learn and store discriminative features for each class. Specially, D-Prompt strengthens the discriminative features matching the instance, thereby facilitating class separation. To improve the quality of discriminative features, we propose a Prompt Clustering Loss that minimizes the distance between D-Prompts corresponding to the same class, thereby mitigating the potential introduction of noise. Finally, ND-prompt employs a reverse attention mechanism to capture task-agnostic non-discriminative features, Furthermore, a data augmentation strategy based on ND-Prompt is used to generate diverse samples by introducing perturbations to the ND-Prompt, thereby aiding in the mitigation of overfitting. We demonstrate robust performance across three datasets, outperforming previous state-of-the-art methods. In summary, our contributions are as follows:", + "bbox": [ + 99, + 234, + 482, + 556 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the SEC-Prompt framework, which leverages Adaptive Hierarchical Query to capture feature semantics and designs distinct matching mechanisms to learn complementary semantic knowledge, thereby facilitating new class learning in FSCIL.", + "- We propose D-Prompt, which leverages an attention mechanism to learn discriminative features, and introduce Prompt Clustering Loss to enhance the feature discrimination.", + "- We design a reverse attention mechanism as the matching strategy for ND-Prompt to learn non-discriminative features, and propose an ND-Prompt-based data augmentation strategy to enhance sample diversity." + ], + "bbox": [ + 101, + 561, + 482, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 101, + 768, + 238, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Class-Incremental Learning (CIL)", + "text_level": 1, + "bbox": [ + 101, + 792, + 393, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-prompt-based approaches. To address the challenges of CIL, existing methods are primarily classified into two categories. Replay-based methods transfer old knowledge using knowledge distillation, where old exemplars are stored in a memory buffer [4, 13, 19, 26, 46], or a gen", + "bbox": [ + 99, + 815, + 482, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erator [17, 18, 29, 38] is trained to synthesize past data. Regularization-based methods [2, 3, 5, 37, 48] constrain changes to critical parameters to maintain a balance between stability and plasticity.", + "bbox": [ + 511, + 102, + 893, + 160 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt-based approaches. Prompt-based methods [23, 24, 45] were initially introduced in Natural Language Processing (NLP) to guide attention toward specific downstream tasks using additional prompt information. Recently, prompt-based CIL approaches have emerged as a promising solution to mitigate catastrophic forgetting in incremental learning without relying on rehearsal strategies. L2P [44] was the first to propose using a key-query mechanism to select task-specific prompts from a prompt pool. DualP [43] added task-invariant prompts to share information across tasks. CodaP [34] introduced an attention-based key-query mechanism and trained the prompt components end-to-end. DAP [16] designed an adaptive prompt generator that creates instance-specific fine-grained instructions for each input. Recently, HideP [41] decomposed the CIL task into hierarchical components and optimized them individually. These methods encode new knowledge without modifying the backbone encoder's parameters, offering a significant advantage by avoiding exemplar storage. However, they face challenges in few-shot learning scenarios due to their reliance on sufficient training data.", + "bbox": [ + 509, + 162, + 893, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Few-Shot Class-Incremental Learning (FSCIL)", + "text_level": 1, + "bbox": [ + 511, + 483, + 893, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TOPIC [36] introduced Few-Shot Class-Incremental Learning to address Class-Incremental Learning challenges in scenarios with limited data availability. According to [51], existing FSCIL methods can be divided into two main approaches. Backward-compatible methods [8, 9, 20, 27, 33, 36, 49, 50, 54] focus on mitigating catastrophic forgetting and overfitting caused by parameter updates during new class learning. Forward-compatible methods [1, 35, 42, 47, 51-53], emphasize learning high-quality knowledge from base classes to facilitate generalization and improve new class learning. The former suffers from catastrophic forgetting due to fine-tuning the backbone, while the latter, by freezing the backbone, has limited capacity for learning new classes, leading to the stability-plasticity dilemma.", + "bbox": [ + 509, + 506, + 893, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt-based approaches. Recently, prompt-based methods have facilitated more efficient knowledge transfer and retention. For example, PKT [30] employs a ViT backbone and fine-tunes selected layers to balance generalization and discrimination, while leveraging prompts to encode domain-specific and semantic knowledge. ASP [25] uses a frozen ViT backbone, with attention-aware task-invariant prompts and self-adaptive task-specific prompts to transfer knowledge from base to new classes, thereby mitigating both forgetting and overfitting. These methods focus solely on task-specific discriminative information, while overlooking the importance of non-discriminative informa", + "bbox": [ + 509, + 712, + 893, + 888 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25644", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion for both old knowledge transfer and new knowledge learning. To address this, our approach decomposes base-class knowledge into discriminative and non-discriminative information.", + "bbox": [ + 99, + 102, + 485, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 101, + 176, + 225, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 101, + 200, + 259, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FSCIL aims to incrementally learn through a sequence of disjoint classes with a few samples in each class while retaining knowledge of previously learned classes. Let $\\{\\mathcal{D}_{\\mathrm{train}}^0,\\mathcal{D}_{\\mathrm{train}}^1,\\dots ,\\mathcal{D}_{\\mathrm{train}}^T\\}$ represent the sequence of labeled training datasets across sessions, with $\\mathcal{C}_t$ denoting the corresponding label set at each session $t$ $(t = 0,1,\\ldots ,T)$ . The label sets are disjoint among different sessions, i.e., $\\mathcal{C}^i\\cap \\mathcal{C}^j = \\emptyset$ for any $i\\neq j$ , while the label sets of the test set $\\mathcal{D}_{test}^{t}$ includes samples from base session to session $t$ , that is, the cumulative label sets $\\mathcal{C}^0\\cup \\mathcal{C}^1\\cup \\dots \\cup \\mathcal{C}^n$ . Generally, the training set in the 0-th session (also known as the base session), $\\mathcal{D}_{\\mathrm{train}}^0$ , is a relatively large dataset with sufficient data for training. In each subsequent few-shot incremental session $(t > 0)$ , the training set $\\mathcal{D}_{\\mathrm{train}}^t$ comprises $N$ classes, each containing $K$ samples, typically referred to as an $N$ -way $K$ -shot training set.", + "bbox": [ + 99, + 222, + 483, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual Prompt Tuning", + "text_level": 1, + "bbox": [ + 101, + 468, + 303, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual prompt tuning(VPT) [15] introduces a small amount of task-specific learnable parameters to adapt transformer models for downstream vision tasks. In this paper, we do not alter the technical foundations of prompt tuning. Specifically, prompts are prepended to the inputs of the selected Transformer layer. For the $i$ -th layer, let the input be $h_i \\in \\mathbb{R}^{L \\times D}$ . The operation of VPT is defined as:", + "bbox": [ + 99, + 489, + 483, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {P r o T}} \\left(p _ {i}, h _ {i}\\right) = \\operatorname {L a y e r} _ {i} \\left(\\left[ p _ {i}; h _ {i} \\right]\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 607, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_i \\in \\mathbb{R}^{L_p \\times D}$ represents the prompt, $L_{p}$ is its length (a hyperparameter), and $[\\cdot; \\cdot]$ denotes concatenation along the sequence length dimension. As a result, we can fix the Transformer encoder and rely on a small number of parameters (the prompts) to learn session-specific information, effectively preventing forgetting. The key remaining challenge is how to design and update prompts to avoid interference between sessions.", + "bbox": [ + 99, + 633, + 483, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Attention-based Prompting", + "text_level": 1, + "bbox": [ + 101, + 763, + 341, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The prompt pool method typically trains additional prompt parameters using a key-query mechanism and stores the prompts from each session in the pool to preserve knowledge. This design consists of two key components: a query function $q(x)$ , which maps an input $x$ into a feature space, and a matching function $m(q(x), k)$ , which selects the most appropriate prompt from the pool based on the query $q(x)$ .", + "bbox": [ + 99, + 786, + 483, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query function. Existing prompt-based methods [15, 25, 34] define $q(x)$ by utilizing the feature vector corresponding to the [class] token from the entire ViT as the query:", + "bbox": [ + 511, + 102, + 893, + 160 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (x) = \\theta (x) [ 0 ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 161, + 893, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ is the pretrained ViT encoder. Although it has achieved success in incremental learning, it struggles to learn new class knowledge in FSCIL due to its inability to account for feature abstraction levels and variations.", + "bbox": [ + 511, + 181, + 893, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Matching function. Coda-Prompt [34] introduces attention components that compute a weighted sum to generate a composite prompt. Specifically, it incorporates three sets of parameters: prompts $P \\in \\mathbb{R}^{M \\times L_p \\times D}$ , keys $K \\in \\mathbb{R}^{M \\times D}$ and attention $A \\in \\mathbb{R}^{M \\times D}$ . The attention parameters $A$ are used to select features of the query, facilitating effective matching with the keys $K$ . Each prompt $P^j \\in \\mathbb{R}^{L_p \\times D}$ in the prompt pool is associated with a corresponding key $K^j \\in \\mathbb{R}^D$ and attention parameter $A^j \\in \\mathbb{R}^D$ . Given a query $q(x)$ , the attention parameters $A^j$ and keys $K^j$ are utilized to compute the weighting scores between the query and each key:", + "bbox": [ + 511, + 241, + 893, + 417 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {j} = \\gamma (q (x) \\odot A ^ {j}, K ^ {j}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 426, + 893, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is the element-wise product, $\\gamma$ is the cosine similarity. The similarity scores $\\alpha_{i}$ are then used to weight the prompts, resulting in the final prompt $p\\in \\mathbb{R}^{L_p\\times D}$ :", + "bbox": [ + 511, + 454, + 893, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {p} (\\mathrm {x}) = \\sum_ {j = 1} ^ {M} \\alpha^ {j} \\cdot P ^ {j} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 507, + 893, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, Coda-Prompt[34] introduces an attention mechanism as a prompt matching mechanism, assigning positive weights to prompts that match the query to enhance their discriminative features, and assigning negative weights to prompts that do not match the query to suppress their nondiscriminative features. However, the simultaneous learning of discriminative and non-discriminative features introduces conflicts, undermining training stability in incremental sessions, particularly within FSCIL settings.", + "bbox": [ + 511, + 558, + 893, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 703, + 643, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by human selective attention mechanisms, we categorize task information into discriminative and nondiscriminative features and propose the SEC-Prompt framework to learn complementary features. By enhancing sparse discriminative features, we promote inter-class separation without mutual interference, while balancing dense nondiscriminative features reduces intra-class distance, thereby facilitating the learning of discriminative features.", + "bbox": [ + 511, + 728, + 893, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to recent prompt-based approaches [25], we fix the pre-trained ViT as the backbone and train the prompt to learn task-specific knowledge. The prompt tokens are", + "bbox": [ + 511, + 845, + 893, + 890 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25645", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/24f438f3cbc76b7841adc1d0e398861b5f69f995dafbe2053483ebc9fc7b9202.jpg", + "image_caption": [ + "Figure 1. An overview of our SEC-Prompt approach to few-shot incremental learning. For an input query, the weighting scores are computed using attention parameters and keys. D-Prompt applies the softmax function to enhance the most relevant features, focusing on learning sparse discriminative features critical for class separation. ND-Prompt utilizes a reverse operation followed by the ReLU function to amplify mismatched features, enabling the learning of dense non-discriminative features that contribute to a more comprehensive representation. During incremental sessions, D-Prompt and ND-Prompt freeze the existing parameters. D-Prompt introduces new trainable parameters to the current session, while ND-Prompt extends the dataset by generating diverse training samples. This is achieved by adding perturbations to the similarity score vector, thereby increasing the diversity of samples and enhancing the model's robustness." + ], + "image_footnote": [], + "bbox": [ + 106, + 104, + 890, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "concatenated with the input tokens of the current layer, and discarded from the output. Figure 1 presents an overview of our proposed method. First, we propose an Adaptive Hierarchical Query function to adaptively capture hierarchical feature information (Section 4.1). Next, we introduce Semantic Complementary Prompt (SEC-Prompt) to learn transferable knowledge from the base session. SEC-Prompt consists of the D-Prompt and ND-Prompt modules, which use attention and reverse-attention mechanisms, respectively, to capture discriminative and non-discriminative information from the base session. In the incremental session, we further propose a Prompt Cluster Loss to enhance the quality of D-Prompt and a data augmentation method based on ND-Prompt to increase the diversity of new class samples (Section 4.2). Finally, we employ a cosine classifier and optimize the prompt components using Prompt Cluster Loss and cross-entropy classification loss (Section 4.3).", + "bbox": [ + 99, + 450, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Adaptive Hierarchical Query", + "text_level": 1, + "bbox": [ + 101, + 718, + 354, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing prompt-based methods rely on the fixed output of the ViT encoder as the query, which restricts the prompt modules at each ViT layer from adapting dynamically to feature variations. On one hand, the level of feature abstraction varies across Transformer layers. Shallow layers tend to capture more non-discriminative information, such as textures and edge details, while deeper layers emphasize abstract, class-specific features. Relying exclusively on the [class] token from the final layer as the query overlooks critical shallow-layer feature information, thereby constrain-", + "bbox": [ + 99, + 742, + 483, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing knowledge transfer. On the other hand, since the Vision Transformer (ViT) model is frozen, the queries remain static and fail to adapt to the evolving feature spaces. The prompt module depends entirely on the feature space of the pre-trained ViT to establish reliable mappings, which poses significant challenges for new-class learning. This difficulty arises because the model struggles to construct accurate mappings with limited sample availability.", + "bbox": [ + 511, + 450, + 893, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we propose an Adaptive Hierarchical Query function to address the limitations of static queries. Specifically, for the prompt of the $(i + 1)$ -th Transformer layer, the [class] token from the input of the current layer is used as the query:", + "bbox": [ + 511, + 568, + 893, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i + 1} (x) = \\operatorname {L a y e r} _ {i} ([ p _ {i}, h _ {i} ]) [ 0 ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 654, + 893, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_j$ and $h_j$ is the prompt and input token of the j-th layer. This enables the query to dynamically align with the evolving feature representations across different layers.", + "bbox": [ + 511, + 681, + 893, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Semantic Complementary Prompt", + "text_level": 1, + "bbox": [ + 511, + 734, + 803, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on previous work [15, 23], prompts act as task-specific extensions that guide the model to attend to relevant features and relationships, enabling adaptation to new tasks while preserving the base model's parameters.", + "bbox": [ + 511, + 756, + 893, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that although Coda-Prompt [34] achieves strong adaptability, it also results in more severe forgetting. We believe this is due to the model learning unstable discriminative features. Because discriminative and non-discriminative features exhibit fundamentally different", + "bbox": [ + 511, + 816, + 893, + 888 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25646", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "characteristics, simultaneously learning both types of features is challenging. Specifically, discriminative features are distinctive attributes shared by specific classes, contributing to positive weights exclusively for queries within those classes. In contrast, non-discriminative features are more sporadic, appearing in various instances without class specificity. These features can yield either positive or negative weights for queries belonging to the same class. Thus, using only the positive or negative nature of weights is insufficient to distinguish between discriminative and non-discriminative features. In fact, assigning weights to prompts based on negative similarity scores may introduce additional noise. To address this, we propose a combined approach using both attention and reverse attention mechanisms to separately model discriminative and non-discriminative features.", + "bbox": [ + 99, + 102, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discriminative Prompt. Firstly, we introduce attention mechanism as the D-Prompt matching strategy to enhance focus on each instance's discriminative features. As illustrated in Figure 1, D-Prompt computes an attended query and calculates the cosine similarity between this attended query and the keys within Equation 3 to obtain similarity scores. The difference is that we apply a softmax operation to the similarity scores:", + "bbox": [ + 99, + 337, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\mathrm {D}} ^ {j} = \\frac {e ^ {\\alpha_ {\\mathrm {D}} ^ {j}}}{\\sum_ {k} e ^ {\\alpha_ {\\mathrm {D}} ^ {k}}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 460, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{\\mathrm{D}}^{j}$ is the similarity scores of $j$ -th D-Prompt. This allows the model to amplify the discriminative features of the current class while minimizing the influence of other information, thereby focusing on enhancing the discriminative characteristics. These scores are then used in Equation 4 to perform a weighted sum of the prompts.", + "bbox": [ + 99, + 503, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discriminative Prompt Clustering Loss. To achieve classification performance, we optimize our model using cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ to separate features of different classes and aggregate features of the same class. However, the cross-entropy loss can easily lead D-Prompt to fit nondiscriminative features for better separation, which introduces noise into the discriminative information and harms the model's generalization ability. Therefore, we introduce the prompt clustering loss:", + "bbox": [ + 99, + 594, + 483, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P C}} (x, y) = \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(1 - \\gamma \\left(p _ {\\mathrm {D}} \\left(x _ {i}\\right), p _ {\\mathrm {D}} \\left(x _ {j}\\right)\\right)\\right) \\cdot \\mathbb {I} \\left(y _ {i} = y _ {j}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 732, + 486, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{D}}(x_i)$ represents the result of the D-Prompt calculation for input $x_i$ , as described in Equation 4. $N$ is the batch size, and $\\gamma$ denotes the cosine similarity used to measure the alignment between representations. By aligning discriminative features within similar samples, we encourage shared discriminative traits, avoiding noise introduction for individual sample separation. Notably, we also attempted to", + "bbox": [ + 99, + 787, + 483, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "include negative samples to construct contrastive loss, but this led to a performance degradation. We believe this is due to shared discriminative information across categories, i.e., horses and zebras share similar body shapes.", + "bbox": [ + 511, + 102, + 893, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Non-Discriminative Prompt. D-Prompt can learn discriminative information to achieve a separated feature space. However, due to discrepancies between the training set distribution and the true distribution, the model often struggles to distinguish between discriminative and non-discriminative features. This issue becomes more pronounced when only limited data is available, leading to severe overfitting. Non-discriminative features, often regarded as noise in classification tasks, can increase the sparsity of effective training samples and hinder the capture of discriminative features. Eliminating such noise is inherently challenging. To address this, we propose explicitly modeling non-discriminative features to balance their representation by emphasizing mismatched non-discriminative features. For example, with images of different styles, the underlying logic is to eliminate the impact of a particular style by introducing various styles.", + "bbox": [ + 511, + 161, + 893, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reverse attention mechanism [6, 14] has been employed in previous image detection tasks to model background information and reduce its interference. Recently, [22] utilized reverse attention to suppress style-related features during text generation, thereby minimizing their interference with content representation. Inspired by this, we adopt reverse attention as the matching mechanism for ND-Prompt. Specifically, ND-Prompt assigns high weights to dissimilar queries and keys while allocating low weights to similar queries and keys. Subsequently, we apply the ReLU function instead of softmax to truncate negative weights, preventing the relationships among non-discriminative features from being disrupted:", + "bbox": [ + 511, + 411, + 893, + 601 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {\\mathrm {N D}} ^ {j} = \\max (0, - \\gamma (q (x) \\odot A _ {\\mathrm {N D}} ^ {j}, K _ {\\mathrm {N D}} ^ {j})) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 611, + 893, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By balancing the representation of non-discriminative features, this approach effectively reduces intra-class distance and helps D-Prompt to learn more precise discriminative information. It's worth noting that we avoid using negative weights to generate prompts, as shifts in the feature space during incremental learning can make negative weights prone to introducing noise, thereby destabilizing the model.", + "bbox": [ + 511, + 638, + 893, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ND-Prompt-Based Data Augmentation. The limited samples still make it challenging for the model to accurately capture the unique discriminative features of new classes. A naive approach to alleviating the overfitting problem is to increase the number of training samples. Our ND-Prompt naturally incorporates non-discriminative information that does not interfere with the learnable D-Prompt, which focuses on capturing the distinctive discriminative features of each class during the incremental phase. Therefore, we", + "bbox": [ + 511, + 757, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25647", + "bbox": [ + 478, + 939, + 517, + 950 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "propose using ND-Prompt for data augmentation. We introduce spherical Gaussian perturbation into the weighting scores $\\alpha$ :", + "bbox": [ + 99, + 102, + 485, + 145 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\alpha} _ {\\mathrm {N D}} = \\alpha_ {\\mathrm {N D}} + \\mathcal {N} (\\mu , \\sigma^ {2}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 150, + 483, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{N}$ stands for spherical Gaussian distribution, $\\mu$ is the mean, $\\sigma$ is the variance. Since $\\alpha$ is a normalized vector, we set $\\mu = 0$ and $\\sigma = 1$ to sample perturbations from the standard spherical Gaussian distribution. These vectors $\\tilde{\\alpha}$ are then used to generate $n$ sets of perturbed ND-Prompts with varying non-discriminative information for each original sample, where $n$ is a hyperparameter. We treat the concatenated instances of different ND-Prompts as samples containing distinct discriminative information, which can increase sample diversity without affecting the discriminative information, thereby alleviating the model's overfitting.", + "bbox": [ + 99, + 179, + 483, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Optimization", + "text_level": 1, + "bbox": [ + 101, + 354, + 236, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous work, we use a cosine classifier to classify these features into class predictions, as the cosine classifier effectively drives the feature extractor to learn discriminative features. Additionally, it learns an embedding space where features of the same class form tighter clusters, which aids in learning non-discriminative features as well. After training, we replace the classifier parameters with class prototypes to achieve better generalization:", + "bbox": [ + 99, + 378, + 483, + 497 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {k} = \\frac {1}{N _ {k}} \\sum_ {(x, y) \\in \\mathcal {D} _ {\\text {t r a i n}} ^ {t}} I [ y = c ] f _ {\\theta , N, D} (x) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 508, + 483, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $f_{\\theta ,N,D}$ represents the ViT encoder with the inserted D-Prompt and ND-Prompt. $I$ denotes the indicator function, and $N_{k}$ is the sample size of class $c$ in $\\mathcal{D}_{train}^{t}$ .", + "bbox": [ + 99, + 561, + 482, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-discriminative information is shared across all classes, and we assume the base classes contain sufficient non-discriminative information. Therefore, we freeze ND-Prompt after the base session. In contrast, each class possesses unique features, and relying solely on previous knowledge can be insufficient for separating new classes and may even lead to confusion. Consequently, in the incremental session, we freeze D-Prompt to prevent forgetting and extend it to learn the discriminative features of new classes. Specifically, we expand the prompt pool of D-Prompt based on the number of categories in the current task, assigning one prompt component for every two categories. We optimize using a common cross-entropy loss and a Discriminative Prompt Clustering Loss across all sessions. The overall training objective can be expressed as:", + "bbox": [ + 99, + 607, + 485, + 829 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} (x, y) + \\lambda \\mathcal {L} _ {\\mathrm {P C}} (x, y) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 843, + 483, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter.", + "bbox": [ + 101, + 875, + 292, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 101, + 640, + 118 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 126, + 697, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate our method on three datasets: CIFAR100 [21], CUB200-2011 [40], and ImageNet-R [12], following the setup in [25]. For details, CIFAR100 are organized into 60 base classes and 40 incremental classes structured in a 5-way 5-shot FSCIL scenario for a total of 8 sessions. CUB200 and ImageNet-R are divided into 100 classes for the base task, and the left 100 classes are divided into ten 10-way 5-shot incremental tasks.", + "bbox": [ + 511, + 147, + 895, + 265 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. As adopted in [25], we use the ViT-B/16-1K [10], pre-trained on ImageNet1K, as the backbone. For all datasets, input images are resized to $224 \\times 224$ . The model is trained for 30 epochs during the base session and for 10 epochs during each incremental session using the SGD optimizer. In the base session, we set the learning rate to 0.1 and the batch size to 64 for ImageNet-R, and 32 for CIFAR-100 and CUB-200-2011. In the incremental session, we use a batch size of 16 and set the learning rate to 0.0005 for ImageNet-R, and 0.001 for CIFAR100 and CUB200-2011. The prompt token length $L_{D} = L_{ND}$ is set to 5 for ImageNet-R and 3 for CIFAR100 and CUB200. We assign one D-Prompt and one ND-Prompt component for every pair of classes. The size of the prompt pool is set to half the number of classes. Source code is available at https://github.com/yeyeyeye33/SEC-Prompt.", + "bbox": [ + 511, + 266, + 895, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. Following previous works [51, 54], we denote the Top-1 accuracy after the $i$ -th session as $A_{i}$ . And the forgetting phenomenon is measured by the performance dropping rate (PD), i.e., $\\mathrm{PD} = A_0 - A_T$ , where $A_0$ stands for the accuracy after the base session, and $A_T$ is the accuracy after the last session. Additionally, since it does not reflect the relationship between old and new class accuracies, we use Harmonic Accuracy (HAcc) [31] to evaluate balanced performance across both base and new classes after task $T: A_h = \\frac{2 \\times A_o \\times A_n}{A_o + A_n}$ , where $A_o$ denotes the accuracy of base classes in task 0, and $A_n$ is the average accuracy across all classes in sessions $t > 0$ .", + "bbox": [ + 511, + 502, + 895, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare recent state-of-the-art (SOTA) prompt-based CIL approaches, including L2P [44], DualP [43], and CodaP [34]. Additionally, we compare SOTA FSCIL algorithms such as TEEN [42], PriViLege [30], and ASP [25].", + "bbox": [ + 511, + 679, + 895, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Benchmark Comparison", + "text_level": 1, + "bbox": [ + 511, + 763, + 732, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we report the performance of baselines and SEC-Prompt under FSCIL setting. We also evaluate SECPrompt-F (Ours-F), where no additional trainable parameters are introduced during the incremental sessions. Table 1 presents the evaluation results on the ImageNet-R dataset. Figure 2 illustrates the evaluation results on the CUB200 and CIFAR100 datasets.", + "bbox": [ + 511, + 786, + 895, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25648", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1354e8d09e6b3e74c0e940622ea5f7b41ba6a1ea826824dc69a7f5aec4431f35.jpg", + "table_caption": [ + "Table 1. Comparison with SOTA methods on ImageNet-R dataset for FSCIL. More results on other datasets can be found in the appendix." + ], + "table_footnote": [], + "table_body": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]84.676.768.867.664.360.658.356.156.154.754.929.745.4
PriViLege [30]80.077.377.073.872.872.371.369.769.567.766.513.564.7
ASP [25]83.380.479.677.075.674.773.072.171.970.969.713.567.0
L2P [44]73.970.969.365.964.062.660.159.559.058.256.817.252.0
DualP [43]71.569.069.067.466.665.964.164.064.063.462.59.061.6
Coda-P [34]82.880.178.675.373.171.870.069.268.867.466.416.861.7
Ours-F82.980.579.877.576.776.374.874.674.173.873.09.972.1
Ours82.980.680.178.277.877.676.776.175.975.074.58.374.2
", + "bbox": [ + 125, + 125, + 867, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f85e03df315860667b5fc93f0790e626547aa288d271e755d0c081b55a5cb1f.jpg", + "table_caption": [ + "Table 2. Ablation studies on ImageNet-R benchmark. D, ND, PC, DA denote D-Prompt, ND-Prompt, Discriminative Prompt Clustering Loss, Non-Discriminative Prompt-Based Data Augmentation, respectively." + ], + "table_footnote": [], + "table_body": "
DNDPCDAAcc. in each session ↑PD ↓HAcc ↑
012345678910
82.880.178.675.373.171.870.069.268.867.466.416.861.7
✓82.680.580.177.977.376.875.675.274.974.173.19.572.8
✓✓82.980.580.178.177.777.476.676.075.674.974.18.873.8
✓✓✓82.980.580.178.277.777.476.776.175.774.974.38.673.7
✓✓✓✓82.980.680.178.277.877.676.776.175.975.074.58.374.2
", + "bbox": [ + 107, + 357, + 883, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the experimental results on CIFAR100, CUB200, and ImageNet-R, SEC-Prompt outperforms other methods in every incremental task, except for the base session. In the base session, it is reasonable that the state-of-the-art (SOTA) FSCIL methods perform better, as they fine-tune the ViT parameters to fit the base class distribution. Furthermore, SEC-Prompt-F achieves the best performance among all methods, except SEC-Prompt itself, validating that our approach has strong generalization ability. Finally, SEC-Prompt achieves the best performance in both PD and HAcc, particularly on ImageNet-R, surpassing the most recent SOTA method ASP by $5.2\\%$ and $7.2\\%$ , respectively. This demonstrates that our method exhibits stronger stability and plasticity.", + "bbox": [ + 99, + 507, + 483, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 101, + 720, + 250, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted an ablation study to demonstrate the significance of the components we proposed. As shown in Figure 3, our Adaptive Hierarchical Query mechanism significantly improves new class accuracy compared to the Fixed Query, demonstrating its ability to help the model learn more robust transferable knowledge and enhance the plasticity of new classes. To further validate our approach, we used Coda-Prompt as a baseline on the ImageNet-R dataset and progressively added components. As shown in Table 2, we observed a significant performance improvement", + "bbox": [ + 99, + 742, + 483, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36365daa0289fd30c6f7f59d49f2f1bff9710732f27024879d8dc3fcda8fb6d1.jpg", + "image_caption": [ + "Figure 2. Comparison with SOTA methods on Cifar100 and CUB200 dataset for FSCIL." + ], + "image_footnote": [ + "Ours TEEN ASP DualP Ours-F PriViLega L2P Coda-P" + ], + "bbox": [ + 516, + 507, + 702, + 633 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef792cd235bb8433a3033c75b32d4d77cadb1d19fa2d2768e0223a6adbb75600.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 507, + 890, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ASP DualP ege L2P Coda-P", + "bbox": [ + 704, + 638, + 861, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "when replacing Coda-Prompt with D-Prompt combined with Adaptive Hierarchical Query, supporting the rationale behind separating discriminative and non-discriminative information. Adding ND-Prompt further improved Harmonic Accuracy (HAcc), demonstrating that non-discriminative features contribute to better balancing stability and adaptability. Finally, the introduction of Discriminative Prompt Clustering Loss and ND-Prompt-Based Data Augmentation resulted in additional performance gains.", + "bbox": [ + 509, + 756, + 893, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25649", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9794be7c1410b12fea1186aaa1ca64b55fad1e9c701574a8a12ff120274a1114.jpg", + "image_caption": [ + "Figure 3. The impact of different query functions on new class accuracy in incremental sessions. \"FQ\" stands for Fixed Query, and \"AHQ\" represents Adaptive Hierarchical Query. The \"-F\" suffix indicates that the prompt pool is not expanded during the incremental session." + ], + "image_footnote": [], + "bbox": [ + 169, + 108, + 405, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e12eb53c056cb4899ef33bfb20cbe129462c9a27110c12752ebc7c0d83549cf.jpg", + "image_caption": [ + "Figure 4. The t-SNE [39] visualization on the ImageNet-R dataset illustrates the embeddings learned by D-Prompt and SEC-Prompt. Class 0-4 represent the base classes while class 110 represent the novel classes." + ], + "image_footnote": [], + "bbox": [ + 106, + 349, + 478, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Further Analysis", + "text_level": 1, + "bbox": [ + 101, + 585, + 264, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of ND-Prompt. To verify whether ND-Prompt promotes intra-class compactness by balancing non-discriminative features, we visualized the embeddings learned by D-Prompt and SEC-Prompt. As shown in Figure 4, incorporating ND-Prompt significantly reduced the intra-class distance of the base classes, with features clustering closer to the center. Although ND-Prompt does not update during the incremental sessions, the intra-class features of novel classes also became more compact. This suggests that the model has learned the distribution of nondiscriminative features from the base classes, which aids in the learning of new classes. These observations support the idea that ND-Prompt helps the model achieve a better balance between stability and plasticity during incremental learning.", + "bbox": [ + 99, + 608, + 483, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of hyper-parameter. We conducted a sensitivity analysis of $n$ and $\\lambda$ across three benchmark datasets. Here, $n$ represents the number of groups of augmented samples generated by ND-Prompt-Based Data Augmentation,", + "bbox": [ + 99, + 830, + 483, + 891 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5b959bb2de93f312c4517feeb8ed9e523b171e0a581115f7d1b22caa75d806a.jpg", + "image_caption": [ + "Figure 5. Sensitive analysis of $\\lambda$ and n" + ], + "image_footnote": [], + "bbox": [ + 514, + 99, + 692, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81061937159791e3149112f3d33ca499a27cad4fa97c0fc0268bf331060a8da9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 99, + 893, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c168a291ad4448c661b93dd1bc0a4f7f399a2e9e41c38895585c4b613d13910.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]5.9105.9
Ours-F2.9102.9
Ours4.6104.6
", + "bbox": [ + 542, + 247, + 864, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model.", + "bbox": [ + 511, + 362, + 895, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with each group containing the same number of samples as the batch size. The hyperparameter $\\lambda$ is the coefficient of the Discriminative Prompt Clustering Loss. The results of the analysis are shown in Figure 5. To achieve the highest accuracy in the last session $(A_T)$ , we selected $\\lambda = 0.5$ and $n = 2$ for all benchmark datasets. Further details on the selection of prompt-related hyperparameters are provided in the appendix.", + "bbox": [ + 511, + 450, + 895, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Parameter Efficiency. Since the use of prompts introduces additional trainable parameters, we compared the parameter sizes of our method with other SOTA FSCIL approaches. As shown in Table 3, our method achieves higher performance with fewer parameters. This is primarily due to the separation of discriminative and nondiscriminative information, which allows the model to learn feature representations more efficiently.", + "bbox": [ + 511, + 571, + 895, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 714, + 627, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper addresses the FSCIL problem with the SEC-Prompt framework. Inspired by human selective attention, we separate task knowledge into discriminative and non-discriminative components and design complementary prompts for each. The attention mechanism captures discriminative features, while reverse attention balances nondiscriminative ones, creating a compact, robust feature space. Our model achieves SOTA performance on three FSCIL benchmarks, demonstrating strong stability and adaptability.", + "bbox": [ + 509, + 742, + 895, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25650", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 102, + 101, + 256, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No. 62176271) and Basic and Applied Basic Research Foundation of Guangdong Province (Grant No. 2024A1515011692). The work is also partially supported by Key Laboratory of Cognitive Intelligence and Content Security, Ministry of Education (Grant no. RZZN202407, Harbin Institute of Technology).", + "bbox": [ + 99, + 125, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 243, + 194, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2", + "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 844-853, 2021. 2", + "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 2", + "[4] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in neural information processing systems, 32, 2019. 2", + "[5] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2", + "[6] Shuhan Chen, Xiuli Tan, Ben Wang, Huchuan Lu, Xuelong Hu, and Yun Fu. Reverse attention-based residual network for salient object detection. IEEE Transactions on Image Processing, 29:3763-3776, 2020. 5", + "[7] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 1", + "[8] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2", + "[9] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + ], + "bbox": [ + 102, + 268, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6", + "[11] Heitor Murilo Gomes, Jean Paul Barddal, Fabricio Enembreck, and Albert Bifet. A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR), 50(2):1-36, 2017. 1", + "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 6, 1", + "[13] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 2", + "[14] Qin Huang, Chunyang Xia, Chihao Wu, Siyang Li, Ye Wang, Yuhang Song, and C-C Jay Kuo. Semantic segmentation with reverse attention. arXiv preprint arXiv:1707.06426, 2017.5", + "[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4, 2", + "[16] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2", + "[17] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2", + "[18] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In International Conference on Learning Representations, 2018. 2", + "[19] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2", + "[20] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1", + "[22] Dongkyu Lee, Zhiliang Tian, Lanqing Xue, and Nevin L. Zhang. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 93-102. Association for Computational Linguistics, 2021. 5", + "[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 4" + ], + "bbox": [ + 514, + 103, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25651", + "bbox": [ + 478, + 940, + 517, + 950 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2", + "[25] Chenxi Liu, Zhenyi Wang, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. Few-shot class incremental learning with attention-aware self-adaptive prompt. arXiv preprint arXiv:2403.09857, 2024. 2, 3, 6, 7, 8, 1", + "[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2", + "[27] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2", + "[28] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 1", + "[29] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11321-11329, 2019. 2", + "[30] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 2, 6, 7, 8, 1", + "[31] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 6", + "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1", + "[33] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 2", + "[34] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 1, 2, 3, 4, 6, 7", + "[35] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24183-24192, 2023. 2", + "[36] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF con" + ], + "bbox": [ + 102, + 103, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on computer vision and pattern recognition, pages 12183-12192, 2020. 1, 2", + "[37] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. arXiv preprint arXiv:1901.11356, 2019. 2", + "[38] Gido M Van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):4069, 2020. 2", + "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8", + "[40] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 1", + "[41] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[42] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7, 8, 1", + "[43] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 1, 2, 6, 7", + "[44] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 1, 2, 6, 7", + "[45] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2", + "[46] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 374-382, 2019. 2", + "[47] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2", + "[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 2", + "[49] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 516, + 104, + 893, + 888 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25652", + "bbox": [ + 478, + 940, + 519, + 950 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2", + "[50] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11838-11847, 2023. 2", + "[51] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 2, 6", + "[52] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021.", + "[53] Mingli Zhu, Zihao Zhu, Sihong Chen, Chen Chen, and Baoyuan Wu. Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208, 2024. 2", + "[54] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7746-7755, 2023. 2, 6" + ], + "bbox": [ + 104, + 103, + 482, + 459 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25653", + "bbox": [ + 478, + 940, + 517, + 950 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 179, + 97, + 818, + 140 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Supplementary Material", + "bbox": [ + 385, + 151, + 609, + 170 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A. Additional Experimental Details", + "text_level": 1, + "bbox": [ + 101, + 199, + 392, + 217 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.1. Dataset Details", + "text_level": 1, + "bbox": [ + 101, + 224, + 253, + 238 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We follow [25] and use the CIFAR100 [21], ImageNet-R [12], and CUB200-2011 [40] datasets. CIFAR100 consists of 100 classes, each with 500 training images and 100 testing images. The CUB200 dataset contains 11,788 high-resolution bird images categorized into 200 species, with 5,994 samples for training and 5,794 samples for testing. The ImageNet-R dataset is a subset of ImageNet that includes 30,000 images across 200 object categories, focusing on artistic renditions such as paintings, cartoons, sculptures, and sketches.", + "bbox": [ + 99, + 244, + 893, + 319 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.2. Additional Results", + "text_level": 1, + "bbox": [ + 101, + 328, + 279, + 342 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In the main paper, we presented the Top-1 accuracy of CUB200 and CIFAR100 in Figure 2. Here, we provide a detailed comparison on CUB200 in Table 4 and CIFAR100 in Table 5.", + "bbox": [ + 99, + 349, + 893, + 378 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/138356d2174484575412a339a311d4ee00fb214d6a34d9a51fa59bd00ca2e718.jpg", + "table_caption": [ + "Table 4. Comparison with SOTA methods on CUB200 dataset for FSCIL." + ], + "table_footnote": [], + "table_body": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]88.886.285.583.083.081.781.579.779.979.580.08.880.2
PriViLege [30]82.381.380.577.877.876.075.776.075.275.275.17.272.3
ASP [25]87.186.084.983.483.682.482.683.082.683.083.53.683.4
L2P [44]82.481.279.076.876.274.774.174.172.773.073.68.773.6
DualP [43]83.582.280.979.578.677.076.377.075.776.176.57.076.3
Coda-P [34]79.678.176.475.675.073.172.572.872.072.472.96.772.5
Ours-F87.586.685.584.684.983.883.383.883.783.884.43.284.3
Ours87.586.685.584.684.984.083.784.283.884.084.82.684.7
", + "bbox": [ + 125, + 415, + 867, + 598 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/d406d9ecb2e2a9872e3a54bd9e31af13d3a294443707b52ff8e6fbea3697896a.jpg", + "table_caption": [ + "Table 5. Comparison with SOTA methods on CIFAR100 dataset for FSCIL." + ], + "table_footnote": [], + "table_body": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678
TEEN [42]92.990.288.486.886.486.085.885.184.08.981.2
PriViLego [30]90.989.489.087.687.987.487.687.286.14.884.7
ASP [25]92.290.790.088.788.788.288.287.886.75.585.3
L2P [44]84.782.380.177.577.076.075.674.172.312.468.0
DualP [43]86.083.682.980.280.680.280.579.077.48.575.3
Coda-P [34]86.083.681.679.279.178.578.377.075.410.672.2
Ours-F92.090.590.288.888.988.488.488.086.85.285.7
Ours92.090.890.689.189.289.189.288.787.54.587.1
", + "bbox": [ + 171, + 648, + 821, + 830 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Our SEC-Prompt framework achieves the best PD and HAcc on both datasets. Specifically, on CUB200, SEC-Prompt achieves a PD of $2.6\\%$ and an HAcc of $24.7\\%$ , surpassing the second-best method by $1.0\\%$ and $1.3\\%$ , respectively. On CIFAR100, SEC-Prompt achieves a PD of $4.5\\%$ and an HAcc of $87.1\\%$ , outperforming the second-best method by $0.3\\%$ and", + "bbox": [ + 99, + 845, + 893, + 890 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "25654", + "bbox": [ + 478, + 939, + 519, + 952 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "$1.8\\%$ , respectively. The performance advantage of SEC-Prompt on CUB200 and CIFAR100 is less pronounced compared to ImageNet-R. This is likely due to the closer distributional similarity of these datasets to ImageNet, which enables the pretrained ViT generalization capability to sufficiently balance stability and plasticity. Conversely, ImageNet-R, with its diverse styles such as cartoons, sculptures, and paintings, poses greater learning challenges and highlights the significance of non-discriminative information.", + "bbox": [ + 99, + 102, + 895, + 176 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/e3613c9c7b3706f6865ecd37947e5ff3e57e5d1f4b98df25d3b6d0cb65a3ee1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.1101.1
Ours1.5101.5
", + "bbox": [ + 130, + 183, + 450, + 289 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/631b36936d668d8b481c47585cd3a857e8248cc5b52b9f532bb23e0f76430c37.jpg", + "table_caption": [ + "(a) CIFAR100" + ], + "table_footnote": [], + "table_body": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLege [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.7101.7
Ours2.8102.8
", + "bbox": [ + 545, + 183, + 867, + 289 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(b) CUB200", + "bbox": [ + 674, + 291, + 740, + 301 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model.", + "bbox": [ + 99, + 314, + 895, + 343 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Additionally, as shown in Table 6b and Table 6a, we report the parameter efficiency analysis on CUB200 and CIFAR100. Notably, the parameter gap between SEC-Prompt and SOTA methods narrows on CUB200, leading to a reduced difference in base class accuracy as well.", + "bbox": [ + 99, + 359, + 895, + 405 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B. Additional Implementation Details", + "text_level": 1, + "bbox": [ + 101, + 416, + 413, + 434 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "In the main paper, we introduced the key experimental details. In this section, we provide an analysis of the selection of parameters, including the prompt insertion depth of the network, prompt length, and prompt pool size. In Figure 6, we present the results of our parameter selection experiments conducted on ImageNet-R, where LAcc represents the accuracy of the final session, and HAcc denotes the harmonic accuracy.", + "bbox": [ + 99, + 441, + 895, + 501 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/4fbd107ae6333e47710cdb05f0778376b520ec2d7789438d4570541eeedb5b8d.jpg", + "image_caption": [ + "(a) Depth of D-Prompt" + ], + "image_footnote": [], + "bbox": [ + 101, + 510, + 282, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/c9c3d911ff35a09d1b63d309fddd00aa0e8d1cde2dd25138b642970bb8d75e40.jpg", + "image_caption": [ + "(b) Depth of ND-Prompt", + "Figure 6. Selection of prompt parameters, including Layer Depth, prompt length, and prompt pool size. LAcc represents the accuracy of the last session, while HAcc stands for Harmonic Accuracy." + ], + "image_footnote": [], + "bbox": [ + 305, + 511, + 485, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5d56d3a5098364449cdd0d96cd7820d83b58fa49a7554fe5a7c926cf924c8e33.jpg", + "image_caption": [ + "(c) Prompt Lenth" + ], + "image_footnote": [], + "bbox": [ + 511, + 511, + 689, + 614 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/3eae44799103e5997d6677480848abb5fb3e3946d2445402796aed1accbad1b9.jpg", + "image_caption": [ + "(d) Prompt Pool Size" + ], + "image_footnote": [], + "bbox": [ + 715, + 511, + 893, + 614 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.1. Decoupled prompt positions", + "text_level": 1, + "bbox": [ + 101, + 691, + 349, + 708 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Visual Prompt Tuning [15] involves concatenating the prompt with the input of a specified layer in the backbone to fine-tune the model. Since different layers of the backbone exhibit varying levels of feature abstraction, we conduct experiments to determine the optimal layer for prompt insertion. Specifically, as in [34, 43], we begin by inserting the prompt at $\\text{Layer}_1$ and search for the optimal final layer for prompt insertion. Discriminative information is the dominant feature in classification tasks, while non-discriminative information assists the classification task by complementing the discriminative features. Therefore, we first select the insertion layer for D-Prompt independently, and then, based on this, search for the optimal insertion layer for ND-Prompt. As shown in Figures 6a and 6b, D-Prompt achieves the highest accuracy at Layer 8, while ND-Prompt achieves the highest accuracy at Layer 5, aligning with the motivation behind the design of SEC-Prompt. D-Prompt captures high-level discriminative features, whereas ND-Prompt captures low-level non-discriminative features. Similar to findings in [43], we observe that model performance decreases with increasing depth. Specifically, while base class accuracy improves, new class accuracy declines. We hypothesize that this is due to the semantic information in deep features causing overfitting, which adversely affects the model's generalization ability.", + "bbox": [ + 99, + 712, + 895, + 890 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "25655", + "bbox": [ + 478, + 939, + 517, + 950 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.2. Prompt Lenth", + "text_level": 1, + "bbox": [ + 104, + 102, + 245, + 117 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Prompt Length refers to the number of prompt tokens selected from the prompt pool for each instance query. As shown in 6c, increasing the prompt length initially improves accuracy but no longer contributes once the accuracy peaks, consistent with the findings in [34]. To achieve the optimal balance between parameter efficiency and model performance, we set the prompt length to 5.", + "bbox": [ + 104, + 123, + 893, + 180 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B.3. Prompt Pool Size", + "text_level": 1, + "bbox": [ + 104, + 191, + 267, + 205 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Coda-Prompt [34] learns task-specific prompts added to the prompt pool, selecting prompt pool sizes arbitrarily to validate model performance and parameter efficiency. In contrast, our SEC-Prompt framework divides prompts into D-Prompt and ND-Prompt to learn semantic feature knowledge, making it class-dependent. Thus, we select prompt pool sizes proportional to the number of classes. As shown in Figure 6d, for ImageNet-R with 200 classes, we choose [50, 100, 200, 400, 1000] as prompt pool sizes and set the prompt pool size to 100, allocating one prompt component for every two classes. Unlike Coda-Prompt, our model does not continuously benefit from increasing the prompt pool size, instead, performance declines after reaching its peak. We observe that a larger prompt pool size results in more severe forgetting. We hypothesize that while our method enhances learning capability in few-shot scenarios, the quality of learned knowledge for new classes remains suboptimal. A larger prompt pool size amplifies the impact on old knowledge, leading to greater forgetting. This raises an important future research question: how to further improve the quality of features learned in few-shot scenarios.", + "bbox": [ + 104, + 212, + 893, + 358 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "25656", + "bbox": [ + 478, + 940, + 517, + 950 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_model.json b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b0cb688799b15d811a56f309dbce36e1341a39b --- /dev/null +++ b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_model.json @@ -0,0 +1,2648 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.14, + 0.82, + 0.185 + ], + "angle": 0, + "content": "SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.41, + 0.211, + 0.619, + 0.23 + ], + "angle": 0, + "content": "Ye Liu\\(^{1,2}\\) Meng Yang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.23, + 0.91, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou, Guangdong, China" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.247, + 0.885, + 0.264 + ], + "angle": 0, + "content": "\\(^{2}\\)State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, Shaanxi, China" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.266, + 0.729, + 0.281 + ], + "angle": 0, + "content": "liuy2629@mail2.sysu.edu.cn, yangm6@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.255, + 0.315, + 0.331, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.345, + 0.486, + 0.742 + ], + "angle": 0, + "content": "Few-shot class-incremental learning (FSCIL) presents a significant challenge in machine learning, requiring models to integrate new classes from limited examples while preserving performance on previously learned classes. Recently, prompt-based CIL approaches leverage ample data to train prompts, effectively mitigating catastrophic forgetting. However, these methods do not account for the semantic features embedded in prompts, exacerbating the plasticity-stability dilemma in few-shot incremental learning. In this paper, we propose a novel and simple framework named SEmantic Complementary Prompt(SEC-Prompt), which learns two sets of semantically complementary prompts based on an adaptive query: discriminative prompts(D-Prompt) and non-discriminative prompts(ND-Prompt). D-Prompt enhances the separation of class-specific feature distributions by strengthening key discriminative features, while ND-Prompt balances nondiscriminative information to promote generalization to novel classes. To efficiently learn high-quality knowledge from limited samples, we leverage ND-Prompt for data augmentation to increase sample diversity and introduce Prompt Clustering Loss to prevent noise contamination in D-Prompt, ensuring robust discriminative feature learning and improved generalization. Our experimental results showcase state-of-the-art performance across three benchmark datasets, including CIFAR100, ImageNet-R and CUB datasets." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.769, + 0.231, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.794, + 0.485, + 0.869 + ], + "angle": 0, + "content": "AI applications, once deployed, frequently encounter novel classes within continuous data streams [11], necessitating the addition of these new classes to recognition tasks. This scenario drives the study of class incremental learning (CIL) [32], which incrementally learns new class con" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.316, + 0.896, + 0.419 + ], + "angle": 0, + "content": "cepts post-deployment while preserving knowledge of previously learned classes. In more practical scenarios, new classes arise sporadically and usually lack sufficient data, for example, industrial defect detection and medical image recognition, which motivates the development of few-shot class incremental learning (FSCIL) [36] with only a limited number of samples for new classes." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.42, + 0.897, + 0.595 + ], + "angle": 0, + "content": "Both CIL and FSCIL face a central challenge: catastrophic forgetting [28]. As models are incrementally trained to recognize new classes, fine-tuning or expanding parameters often overwrites previously acquired knowledge, leading to significant performance drops on earlier classes. Moreover, research in FSCIL has shown that learning from limited samples can lead to overfitting [7]: the model quickly memorizes small sample sets but fails to generalize, which compromises its robustness. Recent efforts have focused on transferring knowledge from base classes to enhance generalization or on adjusting new class distributions to stabilize learning." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.897, + 0.845 + ], + "angle": 0, + "content": "Recent prompt-based continual learning methods [34, 43, 44] fix the backbone parameters while training additional prompt parameters using a key-query mechanism to learn task-specific knowledge. These methods store the learned knowledge in a prompt pool, effectively preventing forgetting. However, when only a few samples are available, they struggle to effectively learn the knowledge of new classes. In contrast, humans can easily learn to distinguish new categories without disrupting previously learned knowledge. This raises a key question: How do humans achieve FSCIL capabilities? The key to classification tasks lies in learning discriminative features. Humans can identify the key features of new categories while ignoring irrelevant information, which allows them to learn accurate patterns and avoid the influence of noise. Thus, we hypothesize that a selective attention mechanism plays a crucial role in this process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.846, + 0.895, + 0.89 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework called Semantic Complementary Prompting (SEC-Prompt) to enhance FSCIL by emulating the human selective attenuate the semantic representation of the human sentence." + }, + { + "type": "page_footnote", + "bbox": [ + 0.12, + 0.877, + 0.244, + 0.89 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25643" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.482, + 0.235 + ], + "angle": 0, + "content": "tion mechanism. By designining two novel attentions, we propose to leverage extensive prior knowledge for generalization and introduce two prompt components: Discriminative Prompt (D-Prompt) and Non-Discriminative Prompt (ND-Prompt). Using distinct matching mechanisms to learn complementary knowledge, D-Prompt and ND-Prompt separately learn task-specific discriminative information and task-agnostic nondiscriminative information, addressing catastrophic forgetting and overfitting in FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.236, + 0.483, + 0.558 + ], + "angle": 0, + "content": "SEC-Prompt incorporates both querying and matching mechanisms. Firstly, we use the class token from the input of each Transformer layer as an adaptive query to capture multi-level feature information, thereby enhancing the model's generalization ability and improving its capability to learn new classes. Then, the D-Prompt utilizes an attention mechanism as a matching mechanism to learn and store discriminative features for each class. Specially, D-Prompt strengthens the discriminative features matching the instance, thereby facilitating class separation. To improve the quality of discriminative features, we propose a Prompt Clustering Loss that minimizes the distance between D-Prompts corresponding to the same class, thereby mitigating the potential introduction of noise. Finally, ND-prompt employs a reverse attention mechanism to capture task-agnostic non-discriminative features, Furthermore, a data augmentation strategy based on ND-Prompt is used to generate diverse samples by introducing perturbations to the ND-Prompt, thereby aiding in the mitigation of overfitting. We demonstrate robust performance across three datasets, outperforming previous state-of-the-art methods. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.563, + 0.483, + 0.636 + ], + "angle": 0, + "content": "- We introduce the SEC-Prompt framework, which leverages Adaptive Hierarchical Query to capture feature semantics and designs distinct matching mechanisms to learn complementary semantic knowledge, thereby facilitating new class learning in FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.637, + 0.483, + 0.693 + ], + "angle": 0, + "content": "- We propose D-Prompt, which leverages an attention mechanism to learn discriminative features, and introduce Prompt Clustering Loss to enhance the feature discrimination." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.695, + 0.483, + 0.754 + ], + "angle": 0, + "content": "- We design a reverse attention mechanism as the matching strategy for ND-Prompt to learn non-discriminative features, and propose an ND-Prompt-based data augmentation strategy to enhance sample diversity." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.563, + 0.483, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.769, + 0.24, + 0.784 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.794, + 0.394, + 0.81 + ], + "angle": 0, + "content": "2.1. Class-Incremental Learning (CIL)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.816, + 0.483, + 0.89 + ], + "angle": 0, + "content": "Non-prompt-based approaches. To address the challenges of CIL, existing methods are primarily classified into two categories. Replay-based methods transfer old knowledge using knowledge distillation, where old exemplars are stored in a memory buffer [4, 13, 19, 26, 46], or a gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.894, + 0.161 + ], + "angle": 0, + "content": "erator [17, 18, 29, 38] is trained to synthesize past data. Regularization-based methods [2, 3, 5, 37, 48] constrain changes to critical parameters to maintain a balance between stability and plasticity." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.163, + 0.895, + 0.472 + ], + "angle": 0, + "content": "Prompt-based approaches. Prompt-based methods [23, 24, 45] were initially introduced in Natural Language Processing (NLP) to guide attention toward specific downstream tasks using additional prompt information. Recently, prompt-based CIL approaches have emerged as a promising solution to mitigate catastrophic forgetting in incremental learning without relying on rehearsal strategies. L2P [44] was the first to propose using a key-query mechanism to select task-specific prompts from a prompt pool. DualP [43] added task-invariant prompts to share information across tasks. CodaP [34] introduced an attention-based key-query mechanism and trained the prompt components end-to-end. DAP [16] designed an adaptive prompt generator that creates instance-specific fine-grained instructions for each input. Recently, HideP [41] decomposed the CIL task into hierarchical components and optimized them individually. These methods encode new knowledge without modifying the backbone encoder's parameters, offering a significant advantage by avoiding exemplar storage. However, they face challenges in few-shot learning scenarios due to their reliance on sufficient training data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.894, + 0.5 + ], + "angle": 0, + "content": "2.2. Few-Shot Class-Incremental Learning (FSCIL)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.507, + 0.894, + 0.713 + ], + "angle": 0, + "content": "TOPIC [36] introduced Few-Shot Class-Incremental Learning to address Class-Incremental Learning challenges in scenarios with limited data availability. According to [51], existing FSCIL methods can be divided into two main approaches. Backward-compatible methods [8, 9, 20, 27, 33, 36, 49, 50, 54] focus on mitigating catastrophic forgetting and overfitting caused by parameter updates during new class learning. Forward-compatible methods [1, 35, 42, 47, 51-53], emphasize learning high-quality knowledge from base classes to facilitate generalization and improve new class learning. The former suffers from catastrophic forgetting due to fine-tuning the backbone, while the latter, by freezing the backbone, has limited capacity for learning new classes, leading to the stability-plasticity dilemma." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.713, + 0.894, + 0.89 + ], + "angle": 0, + "content": "Prompt-based approaches. Recently, prompt-based methods have facilitated more efficient knowledge transfer and retention. For example, PKT [30] employs a ViT backbone and fine-tunes selected layers to balance generalization and discrimination, while leveraging prompts to encode domain-specific and semantic knowledge. ASP [25] uses a frozen ViT backbone, with attention-aware task-invariant prompts and self-adaptive task-specific prompts to transfer knowledge from base to new classes, thereby mitigating both forgetting and overfitting. These methods focus solely on task-specific discriminative information, while overlooking the importance of non-discriminative informa" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25644" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.162 + ], + "angle": 0, + "content": "tion for both old knowledge transfer and new knowledge learning. To address this, our approach decomposes base-class knowledge into discriminative and non-discriminative information." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.177, + 0.226, + 0.194 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.201, + 0.26, + 0.218 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.223, + 0.485, + 0.46 + ], + "angle": 0, + "content": "FSCIL aims to incrementally learn through a sequence of disjoint classes with a few samples in each class while retaining knowledge of previously learned classes. Let \\(\\{\\mathcal{D}_{\\mathrm{train}}^0,\\mathcal{D}_{\\mathrm{train}}^1,\\dots ,\\mathcal{D}_{\\mathrm{train}}^T\\}\\) represent the sequence of labeled training datasets across sessions, with \\(\\mathcal{C}_t\\) denoting the corresponding label set at each session \\(t\\) \\((t = 0,1,\\ldots ,T)\\). The label sets are disjoint among different sessions, i.e., \\(\\mathcal{C}^i\\cap \\mathcal{C}^j = \\emptyset\\) for any \\(i\\neq j\\), while the label sets of the test set \\(\\mathcal{D}_{test}^{t}\\) includes samples from base session to session \\(t\\), that is, the cumulative label sets \\(\\mathcal{C}^0\\cup \\mathcal{C}^1\\cup \\dots \\cup \\mathcal{C}^n\\). Generally, the training set in the 0-th session (also known as the base session), \\(\\mathcal{D}_{\\mathrm{train}}^0\\), is a relatively large dataset with sufficient data for training. In each subsequent few-shot incremental session \\((t > 0)\\), the training set \\(\\mathcal{D}_{\\mathrm{train}}^t\\) comprises \\(N\\) classes, each containing \\(K\\) samples, typically referred to as an \\(N\\)-way \\(K\\)-shot training set." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.469, + 0.305, + 0.485 + ], + "angle": 0, + "content": "3.2. Visual Prompt Tuning" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.491, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Visual prompt tuning(VPT) [15] introduces a small amount of task-specific learnable parameters to adapt transformer models for downstream vision tasks. In this paper, we do not alter the technical foundations of prompt tuning. Specifically, prompts are prepended to the inputs of the selected Transformer layer. For the \\(i\\)-th layer, let the input be \\(h_i \\in \\mathbb{R}^{L \\times D}\\). The operation of VPT is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.608, + 0.483, + 0.625 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {P r o T}} \\left(p _ {i}, h _ {i}\\right) = \\operatorname {L a y e r} _ {i} \\left(\\left[ p _ {i}; h _ {i} \\right]\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.635, + 0.485, + 0.753 + ], + "angle": 0, + "content": "where \\( p_i \\in \\mathbb{R}^{L_p \\times D} \\) represents the prompt, \\( L_{p} \\) is its length (a hyperparameter), and \\( [\\cdot; \\cdot] \\) denotes concatenation along the sequence length dimension. As a result, we can fix the Transformer encoder and rely on a small number of parameters (the prompts) to learn session-specific information, effectively preventing forgetting. The key remaining challenge is how to design and update prompts to avoid interference between sessions." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.765, + 0.342, + 0.782 + ], + "angle": 0, + "content": "3.3. Attention-based Prompting" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.787, + 0.484, + 0.892 + ], + "angle": 0, + "content": "The prompt pool method typically trains additional prompt parameters using a key-query mechanism and stores the prompts from each session in the pool to preserve knowledge. This design consists of two key components: a query function \\( q(x) \\), which maps an input \\( x \\) into a feature space, and a matching function \\( m(q(x), k) \\), which selects the most appropriate prompt from the pool based on the query \\( q(x) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Query function. Existing prompt-based methods [15, 25, 34] define \\( q(x) \\) by utilizing the feature vector corresponding to the [class] token from the entire ViT as the query:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.162, + 0.894, + 0.178 + ], + "angle": 0, + "content": "\\[\nq (x) = \\theta (x) [ 0 ] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.895, + 0.241 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the pretrained ViT encoder. Although it has achieved success in incremental learning, it struggles to learn new class knowledge in FSCIL due to its inability to account for feature abstraction levels and variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.895, + 0.418 + ], + "angle": 0, + "content": "Matching function. Coda-Prompt [34] introduces attention components that compute a weighted sum to generate a composite prompt. Specifically, it incorporates three sets of parameters: prompts \\( P \\in \\mathbb{R}^{M \\times L_p \\times D} \\), keys \\( K \\in \\mathbb{R}^{M \\times D} \\) and attention \\( A \\in \\mathbb{R}^{M \\times D} \\). The attention parameters \\( A \\) are used to select features of the query, facilitating effective matching with the keys \\( K \\). Each prompt \\( P^j \\in \\mathbb{R}^{L_p \\times D} \\) in the prompt pool is associated with a corresponding key \\( K^j \\in \\mathbb{R}^D \\) and attention parameter \\( A^j \\in \\mathbb{R}^D \\). Given a query \\( q(x) \\), the attention parameters \\( A^j \\) and keys \\( K^j \\) are utilized to compute the weighting scores between the query and each key:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.427, + 0.894, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {j} = \\gamma (q (x) \\odot A ^ {j}, K ^ {j}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.895, + 0.5 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the element-wise product, \\(\\gamma\\) is the cosine similarity. The similarity scores \\(\\alpha_{i}\\) are then used to weight the prompts, resulting in the final prompt \\(p\\in \\mathbb{R}^{L_p\\times D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.508, + 0.894, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathrm {p} (\\mathrm {x}) = \\sum_ {j = 1} ^ {M} \\alpha^ {j} \\cdot P ^ {j} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.895, + 0.692 + ], + "angle": 0, + "content": "Intuitively, Coda-Prompt[34] introduces an attention mechanism as a prompt matching mechanism, assigning positive weights to prompts that match the query to enhance their discriminative features, and assigning negative weights to prompts that do not match the query to suppress their nondiscriminative features. However, the simultaneous learning of discriminative and non-discriminative features introduces conflicts, undermining training stability in incremental sessions, particularly within FSCIL settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.704, + 0.644, + 0.721 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.895, + 0.845 + ], + "angle": 0, + "content": "Inspired by human selective attention mechanisms, we categorize task information into discriminative and nondiscriminative features and propose the SEC-Prompt framework to learn complementary features. By enhancing sparse discriminative features, we promote inter-class separation without mutual interference, while balancing dense nondiscriminative features reduces intra-class distance, thereby facilitating the learning of discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.846, + 0.895, + 0.891 + ], + "angle": 0, + "content": "Similar to recent prompt-based approaches [25], we fix the pre-trained ViT as the backbone and train the prompt to learn task-specific knowledge. The prompt tokens are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25645" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.106, + 0.891, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.331, + 0.897, + 0.427 + ], + "angle": 0, + "content": "Figure 1. An overview of our SEC-Prompt approach to few-shot incremental learning. For an input query, the weighting scores are computed using attention parameters and keys. D-Prompt applies the softmax function to enhance the most relevant features, focusing on learning sparse discriminative features critical for class separation. ND-Prompt utilizes a reverse operation followed by the ReLU function to amplify mismatched features, enabling the learning of dense non-discriminative features that contribute to a more comprehensive representation. During incremental sessions, D-Prompt and ND-Prompt freeze the existing parameters. D-Prompt introduces new trainable parameters to the current session, while ND-Prompt extends the dataset by generating diverse training samples. This is achieved by adding perturbations to the similarity score vector, thereby increasing the diversity of samples and enhancing the model's robustness." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.451, + 0.484, + 0.701 + ], + "angle": 0, + "content": "concatenated with the input tokens of the current layer, and discarded from the output. Figure 1 presents an overview of our proposed method. First, we propose an Adaptive Hierarchical Query function to adaptively capture hierarchical feature information (Section 4.1). Next, we introduce Semantic Complementary Prompt (SEC-Prompt) to learn transferable knowledge from the base session. SEC-Prompt consists of the D-Prompt and ND-Prompt modules, which use attention and reverse-attention mechanisms, respectively, to capture discriminative and non-discriminative information from the base session. In the incremental session, we further propose a Prompt Cluster Loss to enhance the quality of D-Prompt and a data augmentation method based on ND-Prompt to increase the diversity of new class samples (Section 4.2). Finally, we employ a cosine classifier and optimize the prompt components using Prompt Cluster Loss and cross-entropy classification loss (Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.719, + 0.355, + 0.735 + ], + "angle": 0, + "content": "4.1. Adaptive Hierarchical Query" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.484, + 0.89 + ], + "angle": 0, + "content": "Existing prompt-based methods rely on the fixed output of the ViT encoder as the query, which restricts the prompt modules at each ViT layer from adapting dynamically to feature variations. On one hand, the level of feature abstraction varies across Transformer layers. Shallow layers tend to capture more non-discriminative information, such as textures and edge details, while deeper layers emphasize abstract, class-specific features. Relying exclusively on the [class] token from the final layer as the query overlooks critical shallow-layer feature information, thereby constrain-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.895, + 0.569 + ], + "angle": 0, + "content": "ing knowledge transfer. On the other hand, since the Vision Transformer (ViT) model is frozen, the queries remain static and fail to adapt to the evolving feature spaces. The prompt module depends entirely on the feature space of the pre-trained ViT to establish reliable mappings, which poses significant challenges for new-class learning. This difficulty arises because the model struggles to construct accurate mappings with limited sample availability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.895, + 0.643 + ], + "angle": 0, + "content": "Therefore, we propose an Adaptive Hierarchical Query function to address the limitations of static queries. Specifically, for the prompt of the \\((i + 1)\\)-th Transformer layer, the [class] token from the input of the current layer is used as the query:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.655, + 0.894, + 0.672 + ], + "angle": 0, + "content": "\\[\nq _ {i + 1} (x) = \\operatorname {L a y e r} _ {i} ([ p _ {i}, h _ {i} ]) [ 0 ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.682, + 0.894, + 0.726 + ], + "angle": 0, + "content": "where \\( p_j \\) and \\( h_j \\) is the prompt and input token of the j-th layer. This enables the query to dynamically align with the evolving feature representations across different layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.805, + 0.752 + ], + "angle": 0, + "content": "4.2. Semantic Complementary Prompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.895, + 0.816 + ], + "angle": 0, + "content": "Based on previous work [15, 23], prompts act as task-specific extensions that guide the model to attend to relevant features and relationships, enabling adaptation to new tasks while preserving the base model's parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.895, + 0.89 + ], + "angle": 0, + "content": "We observe that although Coda-Prompt [34] achieves strong adaptability, it also results in more severe forgetting. We believe this is due to the model learning unstable discriminative features. Because discriminative and non-discriminative features exhibit fundamentally different" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.484, + 0.336 + ], + "angle": 0, + "content": "characteristics, simultaneously learning both types of features is challenging. Specifically, discriminative features are distinctive attributes shared by specific classes, contributing to positive weights exclusively for queries within those classes. In contrast, non-discriminative features are more sporadic, appearing in various instances without class specificity. These features can yield either positive or negative weights for queries belonging to the same class. Thus, using only the positive or negative nature of weights is insufficient to distinguish between discriminative and non-discriminative features. In fact, assigning weights to prompts based on negative similarity scores may introduce additional noise. To address this, we propose a combined approach using both attention and reverse attention mechanisms to separately model discriminative and non-discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.338, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Discriminative Prompt. Firstly, we introduce attention mechanism as the D-Prompt matching strategy to enhance focus on each instance's discriminative features. As illustrated in Figure 1, D-Prompt computes an attended query and calculates the cosine similarity between this attended query and the keys within Equation 3 to obtain similarity scores. The difference is that we apply a softmax operation to the similarity scores:" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.462, + 0.484, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\mathrm {D}} ^ {j} = \\frac {e ^ {\\alpha_ {\\mathrm {D}} ^ {j}}}{\\sum_ {k} e ^ {\\alpha_ {\\mathrm {D}} ^ {k}}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.505, + 0.484, + 0.594 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\mathrm{D}}^{j}\\) is the similarity scores of \\(j\\)-th D-Prompt. This allows the model to amplify the discriminative features of the current class while minimizing the influence of other information, thereby focusing on enhancing the discriminative characteristics. These scores are then used in Equation 4 to perform a weighted sum of the prompts." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.595, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Discriminative Prompt Clustering Loss. To achieve classification performance, we optimize our model using cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) to separate features of different classes and aggregate features of the same class. However, the cross-entropy loss can easily lead D-Prompt to fit nondiscriminative features for better separation, which introduces noise into the discriminative information and harms the model's generalization ability. Therefore, we introduce the prompt clustering loss:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.733, + 0.487, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P C}} (x, y) = \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(1 - \\gamma \\left(p _ {\\mathrm {D}} \\left(x _ {i}\\right), p _ {\\mathrm {D}} \\left(x _ {j}\\right)\\right)\\right) \\cdot \\mathbb {I} \\left(y _ {i} = y _ {j}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.788, + 0.484, + 0.89 + ], + "angle": 0, + "content": "where \\( p_{\\mathrm{D}}(x_i) \\) represents the result of the D-Prompt calculation for input \\( x_i \\), as described in Equation 4. \\( N \\) is the batch size, and \\( \\gamma \\) denotes the cosine similarity used to measure the alignment between representations. By aligning discriminative features within similar samples, we encourage shared discriminative traits, avoiding noise introduction for individual sample separation. Notably, we also attempted to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.894, + 0.161 + ], + "angle": 0, + "content": "include negative samples to construct contrastive loss, but this led to a performance degradation. We believe this is due to shared discriminative information across categories, i.e., horses and zebras share similar body shapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.162, + 0.895, + 0.412 + ], + "angle": 0, + "content": "Non-Discriminative Prompt. D-Prompt can learn discriminative information to achieve a separated feature space. However, due to discrepancies between the training set distribution and the true distribution, the model often struggles to distinguish between discriminative and non-discriminative features. This issue becomes more pronounced when only limited data is available, leading to severe overfitting. Non-discriminative features, often regarded as noise in classification tasks, can increase the sparsity of effective training samples and hinder the capture of discriminative features. Eliminating such noise is inherently challenging. To address this, we propose explicitly modeling non-discriminative features to balance their representation by emphasizing mismatched non-discriminative features. For example, with images of different styles, the underlying logic is to eliminate the impact of a particular style by introducing various styles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.894, + 0.602 + ], + "angle": 0, + "content": "The reverse attention mechanism [6, 14] has been employed in previous image detection tasks to model background information and reduce its interference. Recently, [22] utilized reverse attention to suppress style-related features during text generation, thereby minimizing their interference with content representation. Inspired by this, we adopt reverse attention as the matching mechanism for ND-Prompt. Specifically, ND-Prompt assigns high weights to dissimilar queries and keys while allocating low weights to similar queries and keys. Subsequently, we apply the ReLU function instead of softmax to truncate negative weights, preventing the relationships among non-discriminative features from being disrupted:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.612, + 0.894, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {\\mathrm {N D}} ^ {j} = \\max (0, - \\gamma (q (x) \\odot A _ {\\mathrm {N D}} ^ {j}, K _ {\\mathrm {N D}} ^ {j})) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.894, + 0.756 + ], + "angle": 0, + "content": "By balancing the representation of non-discriminative features, this approach effectively reduces intra-class distance and helps D-Prompt to learn more precise discriminative information. It's worth noting that we avoid using negative weights to generate prompts, as shifts in the feature space during incremental learning can make negative weights prone to introducing noise, thereby destabilizing the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.894, + 0.89 + ], + "angle": 0, + "content": "ND-Prompt-Based Data Augmentation. The limited samples still make it challenging for the model to accurately capture the unique discriminative features of new classes. A naive approach to alleviating the overfitting problem is to increase the number of training samples. Our ND-Prompt naturally incorporates non-discriminative information that does not interfere with the learnable D-Prompt, which focuses on capturing the distinctive discriminative features of each class during the incremental phase. Therefore, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.519, + 0.952 + ], + "angle": 0, + "content": "25647" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.146 + ], + "angle": 0, + "content": "propose using ND-Prompt for data augmentation. We introduce spherical Gaussian perturbation into the weighting scores \\(\\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.151, + 0.484, + 0.17 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\alpha} _ {\\mathrm {N D}} = \\alpha_ {\\mathrm {N D}} + \\mathcal {N} (\\mu , \\sigma^ {2}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.485, + 0.342 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}\\) stands for spherical Gaussian distribution, \\(\\mu\\) is the mean, \\(\\sigma\\) is the variance. Since \\(\\alpha\\) is a normalized vector, we set \\(\\mu = 0\\) and \\(\\sigma = 1\\) to sample perturbations from the standard spherical Gaussian distribution. These vectors \\(\\tilde{\\alpha}\\) are then used to generate \\(n\\) sets of perturbed ND-Prompts with varying non-discriminative information for each original sample, where \\(n\\) is a hyperparameter. We treat the concatenated instances of different ND-Prompts as samples containing distinct discriminative information, which can increase sample diversity without affecting the discriminative information, thereby alleviating the model's overfitting." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.356, + 0.238, + 0.372 + ], + "angle": 0, + "content": "4.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.379, + 0.485, + 0.498 + ], + "angle": 0, + "content": "Following previous work, we use a cosine classifier to classify these features into class predictions, as the cosine classifier effectively drives the feature extractor to learn discriminative features. Additionally, it learns an embedding space where features of the same class form tighter clusters, which aids in learning non-discriminative features as well. After training, we replace the classifier parameters with class prototypes to achieve better generalization:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.51, + 0.484, + 0.55 + ], + "angle": 0, + "content": "\\[\nc _ {k} = \\frac {1}{N _ {k}} \\sum_ {(x, y) \\in \\mathcal {D} _ {\\text {t r a i n}} ^ {t}} I [ y = c ] f _ {\\theta , N, D} (x) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.562, + 0.483, + 0.607 + ], + "angle": 0, + "content": "Where \\(f_{\\theta ,N,D}\\) represents the ViT encoder with the inserted D-Prompt and ND-Prompt. \\(I\\) denotes the indicator function, and \\(N_{k}\\) is the sample size of class \\(c\\) in \\(\\mathcal{D}_{train}^{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.608, + 0.486, + 0.83 + ], + "angle": 0, + "content": "Non-discriminative information is shared across all classes, and we assume the base classes contain sufficient non-discriminative information. Therefore, we freeze ND-Prompt after the base session. In contrast, each class possesses unique features, and relying solely on previous knowledge can be insufficient for separating new classes and may even lead to confusion. Consequently, in the incremental session, we freeze D-Prompt to prevent forgetting and extend it to learn the discriminative features of new classes. Specifically, we expand the prompt pool of D-Prompt based on the number of categories in the current task, assigning one prompt component for every two categories. We optimize using a common cross-entropy loss and a Discriminative Prompt Clustering Loss across all sessions. The overall training objective can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.844, + 0.484, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} (x, y) + \\lambda \\mathcal {L} _ {\\mathrm {P C}} (x, y) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.875, + 0.294, + 0.891 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.102, + 0.642, + 0.119 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.127, + 0.698, + 0.143 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.148, + 0.896, + 0.266 + ], + "angle": 0, + "content": "Dataset. We evaluate our method on three datasets: CIFAR100 [21], CUB200-2011 [40], and ImageNet-R [12], following the setup in [25]. For details, CIFAR100 are organized into 60 base classes and 40 incremental classes structured in a 5-way 5-shot FSCIL scenario for a total of 8 sessions. CUB200 and ImageNet-R are divided into 100 classes for the base task, and the left 100 classes are divided into ten 10-way 5-shot incremental tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.897, + 0.502 + ], + "angle": 0, + "content": "Training details. As adopted in [25], we use the ViT-B/16-1K [10], pre-trained on ImageNet1K, as the backbone. For all datasets, input images are resized to \\(224 \\times 224\\). The model is trained for 30 epochs during the base session and for 10 epochs during each incremental session using the SGD optimizer. In the base session, we set the learning rate to 0.1 and the batch size to 64 for ImageNet-R, and 32 for CIFAR-100 and CUB-200-2011. In the incremental session, we use a batch size of 16 and set the learning rate to 0.0005 for ImageNet-R, and 0.001 for CIFAR100 and CUB200-2011. The prompt token length \\(L_{D} = L_{ND}\\) is set to 5 for ImageNet-R and 3 for CIFAR100 and CUB200. We assign one D-Prompt and one ND-Prompt component for every pair of classes. The size of the prompt pool is set to half the number of classes. Source code is available at https://github.com/yeyeyeye33/SEC-Prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.896, + 0.679 + ], + "angle": 0, + "content": "Evaluation protocol. Following previous works [51, 54], we denote the Top-1 accuracy after the \\(i\\)-th session as \\(A_{i}\\). And the forgetting phenomenon is measured by the performance dropping rate (PD), i.e., \\(\\mathrm{PD} = A_0 - A_T\\), where \\(A_0\\) stands for the accuracy after the base session, and \\(A_T\\) is the accuracy after the last session. Additionally, since it does not reflect the relationship between old and new class accuracies, we use Harmonic Accuracy (HAcc) [31] to evaluate balanced performance across both base and new classes after task \\(T: A_h = \\frac{2 \\times A_o \\times A_n}{A_o + A_n}\\), where \\(A_o\\) denotes the accuracy of base classes in task 0, and \\(A_n\\) is the average accuracy across all classes in sessions \\(t > 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.896, + 0.754 + ], + "angle": 0, + "content": "Baselines. We compare recent state-of-the-art (SOTA) prompt-based CIL approaches, including L2P [44], DualP [43], and CodaP [34]. Additionally, we compare SOTA FSCIL algorithms such as TEEN [42], PriViLege [30], and ASP [25]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.733, + 0.781 + ], + "angle": 0, + "content": "5.2. Benchmark Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.896, + 0.89 + ], + "angle": 0, + "content": "In this section, we report the performance of baselines and SEC-Prompt under FSCIL setting. We also evaluate SECPrompt-F (Ours-F), where no additional trainable parameters are introduced during the incremental sessions. Table 1 presents the evaluation results on the ImageNet-R dataset. Figure 2 illustrates the evaluation results on the CUB200 and CIFAR100 datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25648" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.101, + 0.894, + 0.115 + ], + "angle": 0, + "content": "Table 1. Comparison with SOTA methods on ImageNet-R dataset for FSCIL. More results on other datasets can be found in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.126, + 0.869, + 0.307 + ], + "angle": 0, + "content": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]84.676.768.867.664.360.658.356.156.154.754.929.745.4
PriViLege [30]80.077.377.073.872.872.371.369.769.567.766.513.564.7
ASP [25]83.380.479.677.075.674.773.072.171.970.969.713.567.0
L2P [44]73.970.969.365.964.062.660.159.559.058.256.817.252.0
DualP [43]71.569.069.067.466.665.964.164.064.063.462.59.061.6
Coda-P [34]82.880.178.675.373.171.870.069.268.867.466.416.861.7
Ours-F82.980.579.877.576.776.374.874.674.173.873.09.972.1
Ours82.980.680.178.277.877.676.776.175.975.074.58.374.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.32, + 0.897, + 0.348 + ], + "angle": 0, + "content": "Table 2. Ablation studies on ImageNet-R benchmark. D, ND, PC, DA denote D-Prompt, ND-Prompt, Discriminative Prompt Clustering Loss, Non-Discriminative Prompt-Based Data Augmentation, respectively." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.358, + 0.885, + 0.484 + ], + "angle": 0, + "content": "
DNDPCDAAcc. in each session ↑PD ↓HAcc ↑
012345678910
82.880.178.675.373.171.870.069.268.867.466.416.861.7
✓82.680.580.177.977.376.875.675.274.974.173.19.572.8
✓✓82.980.580.178.177.777.476.676.075.674.974.18.873.8
✓✓✓82.980.580.178.277.777.476.776.175.774.974.38.673.7
✓✓✓✓82.980.680.178.277.877.676.776.175.975.074.58.374.2
" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.508, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Based on the experimental results on CIFAR100, CUB200, and ImageNet-R, SEC-Prompt outperforms other methods in every incremental task, except for the base session. In the base session, it is reasonable that the state-of-the-art (SOTA) FSCIL methods perform better, as they fine-tune the ViT parameters to fit the base class distribution. Furthermore, SEC-Prompt-F achieves the best performance among all methods, except SEC-Prompt itself, validating that our approach has strong generalization ability. Finally, SEC-Prompt achieves the best performance in both PD and HAcc, particularly on ImageNet-R, surpassing the most recent SOTA method ASP by \\(5.2\\%\\) and \\(7.2\\%\\), respectively. This demonstrates that our method exhibits stronger stability and plasticity." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.721, + 0.251, + 0.737 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.484, + 0.89 + ], + "angle": 0, + "content": "We conducted an ablation study to demonstrate the significance of the components we proposed. As shown in Figure 3, our Adaptive Hierarchical Query mechanism significantly improves new class accuracy compared to the Fixed Query, demonstrating its ability to help the model learn more robust transferable knowledge and enhance the plasticity of new classes. To further validate our approach, we used Coda-Prompt as a baseline on the ImageNet-R dataset and progressively added components. As shown in Table 2, we observed a significant performance improvement" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.508, + 0.703, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.568, + 0.639, + 0.862, + 0.661 + ], + "angle": 0, + "content": "Ours TEEN ASP DualP Ours-F PriViLega L2P Coda-P" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.508, + 0.892, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.705, + 0.639, + 0.862, + 0.661 + ], + "angle": 0, + "content": "ASP DualP ege L2P Coda-P" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.674, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Figure 2. Comparison with SOTA methods on Cifar100 and CUB200 dataset for FSCIL." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.757, + 0.895, + 0.89 + ], + "angle": 0, + "content": "when replacing Coda-Prompt with D-Prompt combined with Adaptive Hierarchical Query, supporting the rationale behind separating discriminative and non-discriminative information. Adding ND-Prompt further improved Harmonic Accuracy (HAcc), demonstrating that non-discriminative features contribute to better balancing stability and adaptability. Finally, the introduction of Discriminative Prompt Clustering Loss and ND-Prompt-Based Data Augmentation resulted in additional performance gains." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25649" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.109, + 0.406, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.265, + 0.486, + 0.334 + ], + "angle": 0, + "content": "Figure 3. The impact of different query functions on new class accuracy in incremental sessions. \"FQ\" stands for Fixed Query, and \"AHQ\" represents Adaptive Hierarchical Query. The \"-F\" suffix indicates that the prompt pool is not expanded during the incremental session." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.351, + 0.48, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.502, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Figure 4. The t-SNE [39] visualization on the ImageNet-R dataset illustrates the embeddings learned by D-Prompt and SEC-Prompt. Class 0-4 represent the base classes while class 110 represent the novel classes." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.586, + 0.266, + 0.601 + ], + "angle": 0, + "content": "5.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.609, + 0.484, + 0.829 + ], + "angle": 0, + "content": "Analysis of ND-Prompt. To verify whether ND-Prompt promotes intra-class compactness by balancing non-discriminative features, we visualized the embeddings learned by D-Prompt and SEC-Prompt. As shown in Figure 4, incorporating ND-Prompt significantly reduced the intra-class distance of the base classes, with features clustering closer to the center. Although ND-Prompt does not update during the incremental sessions, the intra-class features of novel classes also became more compact. This suggests that the model has learned the distribution of nondiscriminative features from the base classes, which aids in the learning of new classes. These observations support the idea that ND-Prompt helps the model achieve a better balance between stability and plasticity during incremental learning." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.831, + 0.484, + 0.892 + ], + "angle": 0, + "content": "Analysis of hyper-parameter. We conducted a sensitivity analysis of \\( n \\) and \\( \\lambda \\) across three benchmark datasets. Here, \\( n \\) represents the number of groups of augmented samples generated by ND-Prompt-Based Data Augmentation," + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.1, + 0.693, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.1, + 0.895, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.213, + 0.818, + 0.227 + ], + "angle": 0, + "content": "Figure 5. Sensitive analysis of \\(\\lambda\\) and n" + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.248, + 0.865, + 0.353 + ], + "angle": 0, + "content": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]5.9105.9
Ours-F2.9102.9
Ours4.6104.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.363, + 0.896, + 0.417 + ], + "angle": 0, + "content": "Table 3. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.896, + 0.569 + ], + "angle": 0, + "content": "with each group containing the same number of samples as the batch size. The hyperparameter \\(\\lambda\\) is the coefficient of the Discriminative Prompt Clustering Loss. The results of the analysis are shown in Figure 5. To achieve the highest accuracy in the last session \\((A_T)\\), we selected \\(\\lambda = 0.5\\) and \\(n = 2\\) for all benchmark datasets. Further details on the selection of prompt-related hyperparameters are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.897, + 0.69 + ], + "angle": 0, + "content": "Analysis of Parameter Efficiency. Since the use of prompts introduces additional trainable parameters, we compared the parameter sizes of our method with other SOTA FSCIL approaches. As shown in Table 3, our method achieves higher performance with fewer parameters. This is primarily due to the separation of discriminative and nondiscriminative information, which allows the model to learn feature representations more efficiently." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.629, + 0.73 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.743, + 0.896, + 0.891 + ], + "angle": 0, + "content": "This paper addresses the FSCIL problem with the SEC-Prompt framework. Inspired by human selective attention, we separate task knowledge into discriminative and non-discriminative components and design complementary prompts for each. The attention mechanism captures discriminative features, while reverse attention balances nondiscriminative ones, creating a compact, robust feature space. Our model achieves SOTA performance on three FSCIL benchmarks, demonstrating strong stability and adaptability." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25650" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.102, + 0.257, + 0.118 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.125, + 0.484, + 0.221 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. 62176271) and Basic and Applied Basic Research Foundation of Guangdong Province (Grant No. 2024A1515011692). The work is also partially supported by Key Laboratory of Cognitive Intelligence and Content Security, Ministry of Education (Grant no. RZZN202407, Harbin Institute of Technology)." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.244, + 0.196, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.269, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.338, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 844-853, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.404, + 0.484, + 0.47 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.472, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[4] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.539, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[5] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.594, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[6] Shuhan Chen, Xiuli Tan, Ben Wang, Huchuan Lu, Xuelong Hu, and Yun Fu. Reverse attention-based residual network for salient object detection. IEEE Transactions on Image Processing, 29:3763-3776, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.647, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.688, + 0.484, + 0.768 + ], + "angle": 0, + "content": "[8] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.769, + 0.484, + 0.835 + ], + "angle": 0, + "content": "[9] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.836, + 0.484, + 0.89 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.269, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.104, + 0.895, + 0.13 + ], + "angle": 0, + "content": "worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.132, + 0.895, + 0.186 + ], + "angle": 0, + "content": "[11] Heitor Murilo Gomes, Jean Paul Barddal, Fabricio Enembreck, and Albert Bifet. A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR), 50(2):1-36, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.188, + 0.895, + 0.268 + ], + "angle": 0, + "content": "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.27, + 0.895, + 0.336 + ], + "angle": 0, + "content": "[13] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.339, + 0.895, + 0.391 + ], + "angle": 0, + "content": "[14] Qin Huang, Chunyang Xia, Chihao Wu, Siyang Li, Ye Wang, Yuhang Song, and C-C Jay Kuo. Semantic segmentation with reverse attention. arXiv preprint arXiv:1707.06426, 2017.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.394, + 0.895, + 0.448 + ], + "angle": 0, + "content": "[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.449, + 0.895, + 0.515 + ], + "angle": 0, + "content": "[16] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.518, + 0.895, + 0.558 + ], + "angle": 0, + "content": "[17] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.56, + 0.895, + 0.599 + ], + "angle": 0, + "content": "[18] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.601, + 0.895, + 0.655 + ], + "angle": 0, + "content": "[19] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.657, + 0.895, + 0.71 + ], + "angle": 0, + "content": "[20] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.712, + 0.895, + 0.739 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.741, + 0.895, + 0.847 + ], + "angle": 0, + "content": "[22] Dongkyu Lee, Zhiliang Tian, Lanqing Xue, and Nevin L. Zhang. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 93-102. Association for Computational Linguistics, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.849, + 0.895, + 0.889 + ], + "angle": 0, + "content": "[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.895, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.941, + 0.518, + 0.952 + ], + "angle": 0, + "content": "25651" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.143 + ], + "angle": 0, + "content": "[24] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.146, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[25] Chenxi Liu, Zhenyi Wang, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. Few-shot class incremental learning with attention-aware self-adaptive prompt. arXiv preprint arXiv:2403.09857, 2024. 2, 3, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.202, + 0.483, + 0.242 + ], + "angle": 0, + "content": "[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.244, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[27] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.285, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[28] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.341, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[29] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11321-11329, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.409, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[30] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 2, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.479, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[31] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.534, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.602, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[33] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.672, + 0.483, + 0.764 + ], + "angle": 0, + "content": "[34] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.767, + 0.483, + 0.846 + ], + "angle": 0, + "content": "[35] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24183-24192, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.849, + 0.483, + 0.89 + ], + "angle": 0, + "content": "[36] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF con" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.105, + 0.894, + 0.13 + ], + "angle": 0, + "content": "ference on computer vision and pattern recognition, pages 12183-12192, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.132, + 0.894, + 0.185 + ], + "angle": 0, + "content": "[37] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. arXiv preprint arXiv:1901.11356, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.186, + 0.894, + 0.237 + ], + "angle": 0, + "content": "[38] Gido M Van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):4069, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.24, + 0.894, + 0.279 + ], + "angle": 0, + "content": "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.893, + 0.32 + ], + "angle": 0, + "content": "[40] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[41] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.894, + 0.442 + ], + "angle": 0, + "content": "[42] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.444, + 0.894, + 0.523 + ], + "angle": 0, + "content": "[43] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.894, + 0.604 + ], + "angle": 0, + "content": "[44] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[45] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.894, + 0.739 + ], + "angle": 0, + "content": "[46] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 374-382, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.741, + 0.894, + 0.794 + ], + "angle": 0, + "content": "[47] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.796, + 0.894, + 0.847 + ], + "angle": 0, + "content": "[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.849, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[49] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.105, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.941, + 0.52, + 0.952 + ], + "angle": 0, + "content": "25652" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.104, + 0.482, + 0.131 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.132, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[50] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11838-11847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.201, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[51] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.269, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[52] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.337, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[53] Mingli Zhu, Zihao Zhu, Sihong Chen, Chen Chen, and Baoyuan Wu. Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.392, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[54] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7746-7755, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.104, + 0.483, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.941, + 0.519, + 0.952 + ], + "angle": 0, + "content": "25653" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.18, + 0.098, + 0.82, + 0.141 + ], + "angle": 0, + "content": "SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.152, + 0.611, + 0.171 + ], + "angle": 0, + "content": "Supplementary Material" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.2, + 0.393, + 0.218 + ], + "angle": 0, + "content": "A. Additional Experimental Details" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.225, + 0.254, + 0.239 + ], + "angle": 0, + "content": "A.1. Dataset Details" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.245, + 0.895, + 0.32 + ], + "angle": 0, + "content": "We follow [25] and use the CIFAR100 [21], ImageNet-R [12], and CUB200-2011 [40] datasets. CIFAR100 consists of 100 classes, each with 500 training images and 100 testing images. The CUB200 dataset contains 11,788 high-resolution bird images categorized into 200 species, with 5,994 samples for training and 5,794 samples for testing. The ImageNet-R dataset is a subset of ImageNet that includes 30,000 images across 200 object categories, focusing on artistic renditions such as paintings, cartoons, sculptures, and sketches." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.329, + 0.28, + 0.343 + ], + "angle": 0, + "content": "A.2. Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.35, + 0.895, + 0.38 + ], + "angle": 0, + "content": "In the main paper, we presented the Top-1 accuracy of CUB200 and CIFAR100 in Figure 2. Here, we provide a detailed comparison on CUB200 in Table 4 and CIFAR100 in Table 5." + }, + { + "type": "table_caption", + "bbox": [ + 0.285, + 0.392, + 0.71, + 0.406 + ], + "angle": 0, + "content": "Table 4. Comparison with SOTA methods on CUB200 dataset for FSCIL." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.416, + 0.869, + 0.599 + ], + "angle": 0, + "content": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]88.886.285.583.083.081.781.579.779.979.580.08.880.2
PriViLege [30]82.381.380.577.877.876.075.776.075.275.275.17.272.3
ASP [25]87.186.084.983.483.682.482.683.082.683.083.53.683.4
L2P [44]82.481.279.076.876.274.774.174.172.773.073.68.773.6
DualP [43]83.582.280.979.578.677.076.377.075.776.176.57.076.3
Coda-P [34]79.678.176.475.675.073.172.572.872.072.472.96.772.5
Ours-F87.586.685.584.684.983.883.383.883.783.884.43.284.3
Ours87.586.685.584.684.984.083.784.283.884.084.82.684.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.625, + 0.715, + 0.639 + ], + "angle": 0, + "content": "Table 5. Comparison with SOTA methods on CIFAR100 dataset for FSCIL." + }, + { + "type": "table", + "bbox": [ + 0.173, + 0.649, + 0.822, + 0.831 + ], + "angle": 0, + "content": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678
TEEN [42]92.990.288.486.886.486.085.885.184.08.981.2
PriViLego [30]90.989.489.087.687.987.487.687.286.14.884.7
ASP [25]92.290.790.088.788.788.288.287.886.75.585.3
L2P [44]84.782.380.177.577.076.075.674.172.312.468.0
DualP [43]86.083.682.980.280.680.280.579.077.48.575.3
Coda-P [34]86.083.681.679.279.178.578.377.075.410.672.2
Ours-F92.090.590.288.888.988.488.488.086.85.285.7
Ours92.090.890.689.189.289.189.288.787.54.587.1
" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.846, + 0.895, + 0.891 + ], + "angle": 0, + "content": "Our SEC-Prompt framework achieves the best PD and HAcc on both datasets. Specifically, on CUB200, SEC-Prompt achieves a PD of \\(2.6\\%\\) and an HAcc of \\(24.7\\%\\), surpassing the second-best method by \\(1.0\\%\\) and \\(1.3\\%\\), respectively. On CIFAR100, SEC-Prompt achieves a PD of \\(4.5\\%\\) and an HAcc of \\(87.1\\%\\), outperforming the second-best method by \\(0.3\\%\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.521, + 0.953 + ], + "angle": 0, + "content": "25654" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.897, + 0.178 + ], + "angle": 0, + "content": "\\(1.8\\%\\) , respectively. The performance advantage of SEC-Prompt on CUB200 and CIFAR100 is less pronounced compared to ImageNet-R. This is likely due to the closer distributional similarity of these datasets to ImageNet, which enables the pretrained ViT generalization capability to sufficiently balance stability and plasticity. Conversely, ImageNet-R, with its diverse styles such as cartoons, sculptures, and paintings, poses greater learning challenges and highlights the significance of non-discriminative information." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.184, + 0.452, + 0.29 + ], + "angle": 0, + "content": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.1101.1
Ours1.5101.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.253, + 0.292, + 0.329, + 0.303 + ], + "angle": 0, + "content": "(a) CIFAR100" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.184, + 0.868, + 0.29 + ], + "angle": 0, + "content": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLege [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.7101.7
Ours2.8102.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.675, + 0.292, + 0.741, + 0.303 + ], + "angle": 0, + "content": "(b) CUB200" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.315, + 0.897, + 0.344 + ], + "angle": 0, + "content": "Table 6. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.361, + 0.897, + 0.406 + ], + "angle": 0, + "content": "Additionally, as shown in Table 6b and Table 6a, we report the parameter efficiency analysis on CUB200 and CIFAR100. Notably, the parameter gap between SEC-Prompt and SOTA methods narrows on CUB200, leading to a reduced difference in base class accuracy as well." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.417, + 0.414, + 0.435 + ], + "angle": 0, + "content": "B. Additional Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.442, + 0.897, + 0.502 + ], + "angle": 0, + "content": "In the main paper, we introduced the key experimental details. In this section, we provide an analysis of the selection of parameters, including the prompt insertion depth of the network, prompt length, and prompt pool size. In Figure 6, we present the results of our parameter selection experiments conducted on ImageNet-R, where LAcc represents the accuracy of the final session, and HAcc denotes the harmonic accuracy." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.511, + 0.283, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.617, + 0.252, + 0.63 + ], + "angle": 0, + "content": "(a) Depth of D-Prompt" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.512, + 0.486, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.617, + 0.461, + 0.63 + ], + "angle": 0, + "content": "(b) Depth of ND-Prompt" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.512, + 0.69, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.617, + 0.646, + 0.629 + ], + "angle": 0, + "content": "(c) Prompt Lenth" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.512, + 0.895, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.617, + 0.861, + 0.629 + ], + "angle": 0, + "content": "(d) Prompt Pool Size" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.641, + 0.897, + 0.669 + ], + "angle": 0, + "content": "Figure 6. Selection of prompt parameters, including Layer Depth, prompt length, and prompt pool size. LAcc represents the accuracy of the last session, while HAcc stands for Harmonic Accuracy." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.692, + 0.351, + 0.709 + ], + "angle": 0, + "content": "B.1. Decoupled prompt positions" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.713, + 0.897, + 0.891 + ], + "angle": 0, + "content": "Visual Prompt Tuning [15] involves concatenating the prompt with the input of a specified layer in the backbone to fine-tune the model. Since different layers of the backbone exhibit varying levels of feature abstraction, we conduct experiments to determine the optimal layer for prompt insertion. Specifically, as in [34, 43], we begin by inserting the prompt at \\( \\text{Layer}_1 \\) and search for the optimal final layer for prompt insertion. Discriminative information is the dominant feature in classification tasks, while non-discriminative information assists the classification task by complementing the discriminative features. Therefore, we first select the insertion layer for D-Prompt independently, and then, based on this, search for the optimal insertion layer for ND-Prompt. As shown in Figures 6a and 6b, D-Prompt achieves the highest accuracy at Layer 8, while ND-Prompt achieves the highest accuracy at Layer 5, aligning with the motivation behind the design of SEC-Prompt. D-Prompt captures high-level discriminative features, whereas ND-Prompt captures low-level non-discriminative features. Similar to findings in [43], we observe that model performance decreases with increasing depth. Specifically, while base class accuracy improves, new class accuracy declines. We hypothesize that this is due to the semantic information in deep features causing overfitting, which adversely affects the model's generalization ability." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.94, + 0.519, + 0.952 + ], + "angle": 0, + "content": "25655" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.106, + 0.103, + 0.246, + 0.118 + ], + "angle": 0, + "content": "B.2. Prompt Lenth" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.124, + 0.895, + 0.181 + ], + "angle": 0, + "content": "Prompt Length refers to the number of prompt tokens selected from the prompt pool for each instance query. As shown in 6c, increasing the prompt length initially improves accuracy but no longer contributes once the accuracy peaks, consistent with the findings in [34]. To achieve the optimal balance between parameter efficiency and model performance, we set the prompt length to 5." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.192, + 0.268, + 0.207 + ], + "angle": 0, + "content": "B.3. Prompt Pool Size" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.213, + 0.895, + 0.359 + ], + "angle": 0, + "content": "Coda-Prompt [34] learns task-specific prompts added to the prompt pool, selecting prompt pool sizes arbitrarily to validate model performance and parameter efficiency. In contrast, our SEC-Prompt framework divides prompts into D-Prompt and ND-Prompt to learn semantic feature knowledge, making it class-dependent. Thus, we select prompt pool sizes proportional to the number of classes. As shown in Figure 6d, for ImageNet-R with 200 classes, we choose [50, 100, 200, 400, 1000] as prompt pool sizes and set the prompt pool size to 100, allocating one prompt component for every two classes. Unlike Coda-Prompt, our model does not continuously benefit from increasing the prompt pool size, instead, performance declines after reaching its peak. We observe that a larger prompt pool size results in more severe forgetting. We hypothesize that while our method enhances learning capability in few-shot scenarios, the quality of learned knowledge for new classes remains suboptimal. A larger prompt pool size amplifies the impact on old knowledge, leading to greater forgetting. This raises an important future research question: how to further improve the quality of features learned in few-shot scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.941, + 0.519, + 0.952 + ], + "angle": 0, + "content": "25656" + } + ] +] \ No newline at end of file diff --git a/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_origin.pdf b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b514badc951f0974c5fa09bc159e1699c324ed1 --- /dev/null +++ b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/fea93fa7-25dd-4a35-b985-6366f8b59366_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00d2f1026f6217ed9cfc4a6d6bd6f1b2a36723410cd01bd8a87db4c9db3e0fc8 +size 394323 diff --git a/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/full.md b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a0685ca2cc3dca54ebcd0036a27730c4176c9a8b --- /dev/null +++ b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/full.md @@ -0,0 +1,386 @@ +# SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning + +Ye Liu $^{1,2}$ Meng Yang $^{1,2*}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou, Guangdong, China + +$^{2}$ State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, Shaanxi, China + +liuy2629@mail2.sysu.edu.cn, yangm6@mail.sysu.edu.cn + +# Abstract + +Few-shot class-incremental learning (FSCIL) presents a significant challenge in machine learning, requiring models to integrate new classes from limited examples while preserving performance on previously learned classes. Recently, prompt-based CIL approaches leverage ample data to train prompts, effectively mitigating catastrophic forgetting. However, these methods do not account for the semantic features embedded in prompts, exacerbating the plasticity-stability dilemma in few-shot incremental learning. In this paper, we propose a novel and simple framework named SEmantic Complementary Prompt(SEC-Prompt), which learns two sets of semantically complementary prompts based on an adaptive query: discriminative prompts(D-Prompt) and non-discriminative prompts(ND-Prompt). D-Prompt enhances the separation of class-specific feature distributions by strengthening key discriminative features, while ND-Prompt balances nondiscriminative information to promote generalization to novel classes. To efficiently learn high-quality knowledge from limited samples, we leverage ND-Prompt for data augmentation to increase sample diversity and introduce Prompt Clustering Loss to prevent noise contamination in D-Prompt, ensuring robust discriminative feature learning and improved generalization. Our experimental results showcase state-of-the-art performance across three benchmark datasets, including CIFAR100, ImageNet-R and CUB datasets. + +# 1. Introduction + +AI applications, once deployed, frequently encounter novel classes within continuous data streams [11], necessitating the addition of these new classes to recognition tasks. This scenario drives the study of class incremental learning (CIL) [32], which incrementally learns new class con + +cepts post-deployment while preserving knowledge of previously learned classes. In more practical scenarios, new classes arise sporadically and usually lack sufficient data, for example, industrial defect detection and medical image recognition, which motivates the development of few-shot class incremental learning (FSCIL) [36] with only a limited number of samples for new classes. + +Both CIL and FSCIL face a central challenge: catastrophic forgetting [28]. As models are incrementally trained to recognize new classes, fine-tuning or expanding parameters often overwrites previously acquired knowledge, leading to significant performance drops on earlier classes. Moreover, research in FSCIL has shown that learning from limited samples can lead to overfitting [7]: the model quickly memorizes small sample sets but fails to generalize, which compromises its robustness. Recent efforts have focused on transferring knowledge from base classes to enhance generalization or on adjusting new class distributions to stabilize learning. + +Recent prompt-based continual learning methods [34, 43, 44] fix the backbone parameters while training additional prompt parameters using a key-query mechanism to learn task-specific knowledge. These methods store the learned knowledge in a prompt pool, effectively preventing forgetting. However, when only a few samples are available, they struggle to effectively learn the knowledge of new classes. In contrast, humans can easily learn to distinguish new categories without disrupting previously learned knowledge. This raises a key question: How do humans achieve FSCIL capabilities? The key to classification tasks lies in learning discriminative features. Humans can identify the key features of new categories while ignoring irrelevant information, which allows them to learn accurate patterns and avoid the influence of noise. Thus, we hypothesize that a selective attention mechanism plays a crucial role in this process. + +In this paper, we propose a novel framework called Semantic Complementary Prompting (SEC-Prompt) to enhance FSCIL by emulating the human selective attenuate the semantic representation of the human sentence. + +tion mechanism. By designining two novel attentions, we propose to leverage extensive prior knowledge for generalization and introduce two prompt components: Discriminative Prompt (D-Prompt) and Non-Discriminative Prompt (ND-Prompt). Using distinct matching mechanisms to learn complementary knowledge, D-Prompt and ND-Prompt separately learn task-specific discriminative information and task-agnostic nondiscriminative information, addressing catastrophic forgetting and overfitting in FSCIL. + +SEC-Prompt incorporates both querying and matching mechanisms. Firstly, we use the class token from the input of each Transformer layer as an adaptive query to capture multi-level feature information, thereby enhancing the model's generalization ability and improving its capability to learn new classes. Then, the D-Prompt utilizes an attention mechanism as a matching mechanism to learn and store discriminative features for each class. Specially, D-Prompt strengthens the discriminative features matching the instance, thereby facilitating class separation. To improve the quality of discriminative features, we propose a Prompt Clustering Loss that minimizes the distance between D-Prompts corresponding to the same class, thereby mitigating the potential introduction of noise. Finally, ND-prompt employs a reverse attention mechanism to capture task-agnostic non-discriminative features, Furthermore, a data augmentation strategy based on ND-Prompt is used to generate diverse samples by introducing perturbations to the ND-Prompt, thereby aiding in the mitigation of overfitting. We demonstrate robust performance across three datasets, outperforming previous state-of-the-art methods. In summary, our contributions are as follows: + +- We introduce the SEC-Prompt framework, which leverages Adaptive Hierarchical Query to capture feature semantics and designs distinct matching mechanisms to learn complementary semantic knowledge, thereby facilitating new class learning in FSCIL. +- We propose D-Prompt, which leverages an attention mechanism to learn discriminative features, and introduce Prompt Clustering Loss to enhance the feature discrimination. +- We design a reverse attention mechanism as the matching strategy for ND-Prompt to learn non-discriminative features, and propose an ND-Prompt-based data augmentation strategy to enhance sample diversity. + +# 2. Related Work + +# 2.1. Class-Incremental Learning (CIL) + +Non-prompt-based approaches. To address the challenges of CIL, existing methods are primarily classified into two categories. Replay-based methods transfer old knowledge using knowledge distillation, where old exemplars are stored in a memory buffer [4, 13, 19, 26, 46], or a gen + +erator [17, 18, 29, 38] is trained to synthesize past data. Regularization-based methods [2, 3, 5, 37, 48] constrain changes to critical parameters to maintain a balance between stability and plasticity. + +Prompt-based approaches. Prompt-based methods [23, 24, 45] were initially introduced in Natural Language Processing (NLP) to guide attention toward specific downstream tasks using additional prompt information. Recently, prompt-based CIL approaches have emerged as a promising solution to mitigate catastrophic forgetting in incremental learning without relying on rehearsal strategies. L2P [44] was the first to propose using a key-query mechanism to select task-specific prompts from a prompt pool. DualP [43] added task-invariant prompts to share information across tasks. CodaP [34] introduced an attention-based key-query mechanism and trained the prompt components end-to-end. DAP [16] designed an adaptive prompt generator that creates instance-specific fine-grained instructions for each input. Recently, HideP [41] decomposed the CIL task into hierarchical components and optimized them individually. These methods encode new knowledge without modifying the backbone encoder's parameters, offering a significant advantage by avoiding exemplar storage. However, they face challenges in few-shot learning scenarios due to their reliance on sufficient training data. + +# 2.2. Few-Shot Class-Incremental Learning (FSCIL) + +TOPIC [36] introduced Few-Shot Class-Incremental Learning to address Class-Incremental Learning challenges in scenarios with limited data availability. According to [51], existing FSCIL methods can be divided into two main approaches. Backward-compatible methods [8, 9, 20, 27, 33, 36, 49, 50, 54] focus on mitigating catastrophic forgetting and overfitting caused by parameter updates during new class learning. Forward-compatible methods [1, 35, 42, 47, 51-53], emphasize learning high-quality knowledge from base classes to facilitate generalization and improve new class learning. The former suffers from catastrophic forgetting due to fine-tuning the backbone, while the latter, by freezing the backbone, has limited capacity for learning new classes, leading to the stability-plasticity dilemma. + +Prompt-based approaches. Recently, prompt-based methods have facilitated more efficient knowledge transfer and retention. For example, PKT [30] employs a ViT backbone and fine-tunes selected layers to balance generalization and discrimination, while leveraging prompts to encode domain-specific and semantic knowledge. ASP [25] uses a frozen ViT backbone, with attention-aware task-invariant prompts and self-adaptive task-specific prompts to transfer knowledge from base to new classes, thereby mitigating both forgetting and overfitting. These methods focus solely on task-specific discriminative information, while overlooking the importance of non-discriminative informa + +tion for both old knowledge transfer and new knowledge learning. To address this, our approach decomposes base-class knowledge into discriminative and non-discriminative information. + +# 3. Preliminary + +# 3.1. Problem Setting + +FSCIL aims to incrementally learn through a sequence of disjoint classes with a few samples in each class while retaining knowledge of previously learned classes. Let $\{\mathcal{D}_{\mathrm{train}}^0,\mathcal{D}_{\mathrm{train}}^1,\dots ,\mathcal{D}_{\mathrm{train}}^T\}$ represent the sequence of labeled training datasets across sessions, with $\mathcal{C}_t$ denoting the corresponding label set at each session $t$ $(t = 0,1,\ldots ,T)$ . The label sets are disjoint among different sessions, i.e., $\mathcal{C}^i\cap \mathcal{C}^j = \emptyset$ for any $i\neq j$ , while the label sets of the test set $\mathcal{D}_{test}^{t}$ includes samples from base session to session $t$ , that is, the cumulative label sets $\mathcal{C}^0\cup \mathcal{C}^1\cup \dots \cup \mathcal{C}^n$ . Generally, the training set in the 0-th session (also known as the base session), $\mathcal{D}_{\mathrm{train}}^0$ , is a relatively large dataset with sufficient data for training. In each subsequent few-shot incremental session $(t > 0)$ , the training set $\mathcal{D}_{\mathrm{train}}^t$ comprises $N$ classes, each containing $K$ samples, typically referred to as an $N$ -way $K$ -shot training set. + +# 3.2. Visual Prompt Tuning + +Visual prompt tuning(VPT) [15] introduces a small amount of task-specific learnable parameters to adapt transformer models for downstream vision tasks. In this paper, we do not alter the technical foundations of prompt tuning. Specifically, prompts are prepended to the inputs of the selected Transformer layer. For the $i$ -th layer, let the input be $h_i \in \mathbb{R}^{L \times D}$ . The operation of VPT is defined as: + +$$ +f _ {\text {P r o T}} \left(p _ {i}, h _ {i}\right) = \operatorname {L a y e r} _ {i} \left(\left[ p _ {i}; h _ {i} \right]\right), \tag {1} +$$ + +where $p_i \in \mathbb{R}^{L_p \times D}$ represents the prompt, $L_{p}$ is its length (a hyperparameter), and $[\cdot; \cdot]$ denotes concatenation along the sequence length dimension. As a result, we can fix the Transformer encoder and rely on a small number of parameters (the prompts) to learn session-specific information, effectively preventing forgetting. The key remaining challenge is how to design and update prompts to avoid interference between sessions. + +# 3.3. Attention-based Prompting + +The prompt pool method typically trains additional prompt parameters using a key-query mechanism and stores the prompts from each session in the pool to preserve knowledge. This design consists of two key components: a query function $q(x)$ , which maps an input $x$ into a feature space, and a matching function $m(q(x), k)$ , which selects the most appropriate prompt from the pool based on the query $q(x)$ . + +Query function. Existing prompt-based methods [15, 25, 34] define $q(x)$ by utilizing the feature vector corresponding to the [class] token from the entire ViT as the query: + +$$ +q (x) = \theta (x) [ 0 ] \tag {2} +$$ + +where $\theta$ is the pretrained ViT encoder. Although it has achieved success in incremental learning, it struggles to learn new class knowledge in FSCIL due to its inability to account for feature abstraction levels and variations. + +Matching function. Coda-Prompt [34] introduces attention components that compute a weighted sum to generate a composite prompt. Specifically, it incorporates three sets of parameters: prompts $P \in \mathbb{R}^{M \times L_p \times D}$ , keys $K \in \mathbb{R}^{M \times D}$ and attention $A \in \mathbb{R}^{M \times D}$ . The attention parameters $A$ are used to select features of the query, facilitating effective matching with the keys $K$ . Each prompt $P^j \in \mathbb{R}^{L_p \times D}$ in the prompt pool is associated with a corresponding key $K^j \in \mathbb{R}^D$ and attention parameter $A^j \in \mathbb{R}^D$ . Given a query $q(x)$ , the attention parameters $A^j$ and keys $K^j$ are utilized to compute the weighting scores between the query and each key: + +$$ +\alpha^ {j} = \gamma (q (x) \odot A ^ {j}, K ^ {j}) \tag {3} +$$ + +where $\odot$ is the element-wise product, $\gamma$ is the cosine similarity. The similarity scores $\alpha_{i}$ are then used to weight the prompts, resulting in the final prompt $p\in \mathbb{R}^{L_p\times D}$ : + +$$ +\mathrm {p} (\mathrm {x}) = \sum_ {j = 1} ^ {M} \alpha^ {j} \cdot P ^ {j} \tag {4} +$$ + +Intuitively, Coda-Prompt[34] introduces an attention mechanism as a prompt matching mechanism, assigning positive weights to prompts that match the query to enhance their discriminative features, and assigning negative weights to prompts that do not match the query to suppress their nondiscriminative features. However, the simultaneous learning of discriminative and non-discriminative features introduces conflicts, undermining training stability in incremental sessions, particularly within FSCIL settings. + +# 4. Methodology + +Inspired by human selective attention mechanisms, we categorize task information into discriminative and nondiscriminative features and propose the SEC-Prompt framework to learn complementary features. By enhancing sparse discriminative features, we promote inter-class separation without mutual interference, while balancing dense nondiscriminative features reduces intra-class distance, thereby facilitating the learning of discriminative features. + +Similar to recent prompt-based approaches [25], we fix the pre-trained ViT as the backbone and train the prompt to learn task-specific knowledge. The prompt tokens are + +![](images/24f438f3cbc76b7841adc1d0e398861b5f69f995dafbe2053483ebc9fc7b9202.jpg) +Figure 1. An overview of our SEC-Prompt approach to few-shot incremental learning. For an input query, the weighting scores are computed using attention parameters and keys. D-Prompt applies the softmax function to enhance the most relevant features, focusing on learning sparse discriminative features critical for class separation. ND-Prompt utilizes a reverse operation followed by the ReLU function to amplify mismatched features, enabling the learning of dense non-discriminative features that contribute to a more comprehensive representation. During incremental sessions, D-Prompt and ND-Prompt freeze the existing parameters. D-Prompt introduces new trainable parameters to the current session, while ND-Prompt extends the dataset by generating diverse training samples. This is achieved by adding perturbations to the similarity score vector, thereby increasing the diversity of samples and enhancing the model's robustness. + +concatenated with the input tokens of the current layer, and discarded from the output. Figure 1 presents an overview of our proposed method. First, we propose an Adaptive Hierarchical Query function to adaptively capture hierarchical feature information (Section 4.1). Next, we introduce Semantic Complementary Prompt (SEC-Prompt) to learn transferable knowledge from the base session. SEC-Prompt consists of the D-Prompt and ND-Prompt modules, which use attention and reverse-attention mechanisms, respectively, to capture discriminative and non-discriminative information from the base session. In the incremental session, we further propose a Prompt Cluster Loss to enhance the quality of D-Prompt and a data augmentation method based on ND-Prompt to increase the diversity of new class samples (Section 4.2). Finally, we employ a cosine classifier and optimize the prompt components using Prompt Cluster Loss and cross-entropy classification loss (Section 4.3). + +# 4.1. Adaptive Hierarchical Query + +Existing prompt-based methods rely on the fixed output of the ViT encoder as the query, which restricts the prompt modules at each ViT layer from adapting dynamically to feature variations. On one hand, the level of feature abstraction varies across Transformer layers. Shallow layers tend to capture more non-discriminative information, such as textures and edge details, while deeper layers emphasize abstract, class-specific features. Relying exclusively on the [class] token from the final layer as the query overlooks critical shallow-layer feature information, thereby constrain- + +ing knowledge transfer. On the other hand, since the Vision Transformer (ViT) model is frozen, the queries remain static and fail to adapt to the evolving feature spaces. The prompt module depends entirely on the feature space of the pre-trained ViT to establish reliable mappings, which poses significant challenges for new-class learning. This difficulty arises because the model struggles to construct accurate mappings with limited sample availability. + +Therefore, we propose an Adaptive Hierarchical Query function to address the limitations of static queries. Specifically, for the prompt of the $(i + 1)$ -th Transformer layer, the [class] token from the input of the current layer is used as the query: + +$$ +q _ {i + 1} (x) = \operatorname {L a y e r} _ {i} ([ p _ {i}, h _ {i} ]) [ 0 ] \tag {5} +$$ + +where $p_j$ and $h_j$ is the prompt and input token of the j-th layer. This enables the query to dynamically align with the evolving feature representations across different layers. + +# 4.2. Semantic Complementary Prompt + +Based on previous work [15, 23], prompts act as task-specific extensions that guide the model to attend to relevant features and relationships, enabling adaptation to new tasks while preserving the base model's parameters. + +We observe that although Coda-Prompt [34] achieves strong adaptability, it also results in more severe forgetting. We believe this is due to the model learning unstable discriminative features. Because discriminative and non-discriminative features exhibit fundamentally different + +characteristics, simultaneously learning both types of features is challenging. Specifically, discriminative features are distinctive attributes shared by specific classes, contributing to positive weights exclusively for queries within those classes. In contrast, non-discriminative features are more sporadic, appearing in various instances without class specificity. These features can yield either positive or negative weights for queries belonging to the same class. Thus, using only the positive or negative nature of weights is insufficient to distinguish between discriminative and non-discriminative features. In fact, assigning weights to prompts based on negative similarity scores may introduce additional noise. To address this, we propose a combined approach using both attention and reverse attention mechanisms to separately model discriminative and non-discriminative features. + +Discriminative Prompt. Firstly, we introduce attention mechanism as the D-Prompt matching strategy to enhance focus on each instance's discriminative features. As illustrated in Figure 1, D-Prompt computes an attended query and calculates the cosine similarity between this attended query and the keys within Equation 3 to obtain similarity scores. The difference is that we apply a softmax operation to the similarity scores: + +$$ +\alpha_ {\mathrm {D}} ^ {j} = \frac {e ^ {\alpha_ {\mathrm {D}} ^ {j}}}{\sum_ {k} e ^ {\alpha_ {\mathrm {D}} ^ {k}}} \tag {6} +$$ + +where $\alpha_{\mathrm{D}}^{j}$ is the similarity scores of $j$ -th D-Prompt. This allows the model to amplify the discriminative features of the current class while minimizing the influence of other information, thereby focusing on enhancing the discriminative characteristics. These scores are then used in Equation 4 to perform a weighted sum of the prompts. + +Discriminative Prompt Clustering Loss. To achieve classification performance, we optimize our model using cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ to separate features of different classes and aggregate features of the same class. However, the cross-entropy loss can easily lead D-Prompt to fit nondiscriminative features for better separation, which introduces noise into the discriminative information and harms the model's generalization ability. Therefore, we introduce the prompt clustering loss: + +$$ +\mathcal {L} _ {\mathrm {P C}} (x, y) = \frac {1}{2 N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \left(1 - \gamma \left(p _ {\mathrm {D}} \left(x _ {i}\right), p _ {\mathrm {D}} \left(x _ {j}\right)\right)\right) \cdot \mathbb {I} \left(y _ {i} = y _ {j}\right) \tag {7} +$$ + +where $p_{\mathrm{D}}(x_i)$ represents the result of the D-Prompt calculation for input $x_i$ , as described in Equation 4. $N$ is the batch size, and $\gamma$ denotes the cosine similarity used to measure the alignment between representations. By aligning discriminative features within similar samples, we encourage shared discriminative traits, avoiding noise introduction for individual sample separation. Notably, we also attempted to + +include negative samples to construct contrastive loss, but this led to a performance degradation. We believe this is due to shared discriminative information across categories, i.e., horses and zebras share similar body shapes. + +Non-Discriminative Prompt. D-Prompt can learn discriminative information to achieve a separated feature space. However, due to discrepancies between the training set distribution and the true distribution, the model often struggles to distinguish between discriminative and non-discriminative features. This issue becomes more pronounced when only limited data is available, leading to severe overfitting. Non-discriminative features, often regarded as noise in classification tasks, can increase the sparsity of effective training samples and hinder the capture of discriminative features. Eliminating such noise is inherently challenging. To address this, we propose explicitly modeling non-discriminative features to balance their representation by emphasizing mismatched non-discriminative features. For example, with images of different styles, the underlying logic is to eliminate the impact of a particular style by introducing various styles. + +The reverse attention mechanism [6, 14] has been employed in previous image detection tasks to model background information and reduce its interference. Recently, [22] utilized reverse attention to suppress style-related features during text generation, thereby minimizing their interference with content representation. Inspired by this, we adopt reverse attention as the matching mechanism for ND-Prompt. Specifically, ND-Prompt assigns high weights to dissimilar queries and keys while allocating low weights to similar queries and keys. Subsequently, we apply the ReLU function instead of softmax to truncate negative weights, preventing the relationships among non-discriminative features from being disrupted: + +$$ +\alpha_ {\mathrm {N D}} ^ {j} = \max (0, - \gamma (q (x) \odot A _ {\mathrm {N D}} ^ {j}, K _ {\mathrm {N D}} ^ {j})) \tag {8} +$$ + +By balancing the representation of non-discriminative features, this approach effectively reduces intra-class distance and helps D-Prompt to learn more precise discriminative information. It's worth noting that we avoid using negative weights to generate prompts, as shifts in the feature space during incremental learning can make negative weights prone to introducing noise, thereby destabilizing the model. + +ND-Prompt-Based Data Augmentation. The limited samples still make it challenging for the model to accurately capture the unique discriminative features of new classes. A naive approach to alleviating the overfitting problem is to increase the number of training samples. Our ND-Prompt naturally incorporates non-discriminative information that does not interfere with the learnable D-Prompt, which focuses on capturing the distinctive discriminative features of each class during the incremental phase. Therefore, we + +propose using ND-Prompt for data augmentation. We introduce spherical Gaussian perturbation into the weighting scores $\alpha$ : + +$$ +\tilde {\alpha} _ {\mathrm {N D}} = \alpha_ {\mathrm {N D}} + \mathcal {N} (\mu , \sigma^ {2}) \tag {9} +$$ + +where $\mathcal{N}$ stands for spherical Gaussian distribution, $\mu$ is the mean, $\sigma$ is the variance. Since $\alpha$ is a normalized vector, we set $\mu = 0$ and $\sigma = 1$ to sample perturbations from the standard spherical Gaussian distribution. These vectors $\tilde{\alpha}$ are then used to generate $n$ sets of perturbed ND-Prompts with varying non-discriminative information for each original sample, where $n$ is a hyperparameter. We treat the concatenated instances of different ND-Prompts as samples containing distinct discriminative information, which can increase sample diversity without affecting the discriminative information, thereby alleviating the model's overfitting. + +# 4.3. Optimization + +Following previous work, we use a cosine classifier to classify these features into class predictions, as the cosine classifier effectively drives the feature extractor to learn discriminative features. Additionally, it learns an embedding space where features of the same class form tighter clusters, which aids in learning non-discriminative features as well. After training, we replace the classifier parameters with class prototypes to achieve better generalization: + +$$ +c _ {k} = \frac {1}{N _ {k}} \sum_ {(x, y) \in \mathcal {D} _ {\text {t r a i n}} ^ {t}} I [ y = c ] f _ {\theta , N, D} (x) \tag {10} +$$ + +Where $f_{\theta ,N,D}$ represents the ViT encoder with the inserted D-Prompt and ND-Prompt. $I$ denotes the indicator function, and $N_{k}$ is the sample size of class $c$ in $\mathcal{D}_{train}^{t}$ . + +Non-discriminative information is shared across all classes, and we assume the base classes contain sufficient non-discriminative information. Therefore, we freeze ND-Prompt after the base session. In contrast, each class possesses unique features, and relying solely on previous knowledge can be insufficient for separating new classes and may even lead to confusion. Consequently, in the incremental session, we freeze D-Prompt to prevent forgetting and extend it to learn the discriminative features of new classes. Specifically, we expand the prompt pool of D-Prompt based on the number of categories in the current task, assigning one prompt component for every two categories. We optimize using a common cross-entropy loss and a Discriminative Prompt Clustering Loss across all sessions. The overall training objective can be expressed as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C E}} (x, y) + \lambda \mathcal {L} _ {\mathrm {P C}} (x, y) \tag {11} +$$ + +where $\lambda$ is a hyperparameter. + +# 5. Experiments + +# 5.1. Experimental Setup + +Dataset. We evaluate our method on three datasets: CIFAR100 [21], CUB200-2011 [40], and ImageNet-R [12], following the setup in [25]. For details, CIFAR100 are organized into 60 base classes and 40 incremental classes structured in a 5-way 5-shot FSCIL scenario for a total of 8 sessions. CUB200 and ImageNet-R are divided into 100 classes for the base task, and the left 100 classes are divided into ten 10-way 5-shot incremental tasks. + +Training details. As adopted in [25], we use the ViT-B/16-1K [10], pre-trained on ImageNet1K, as the backbone. For all datasets, input images are resized to $224 \times 224$ . The model is trained for 30 epochs during the base session and for 10 epochs during each incremental session using the SGD optimizer. In the base session, we set the learning rate to 0.1 and the batch size to 64 for ImageNet-R, and 32 for CIFAR-100 and CUB-200-2011. In the incremental session, we use a batch size of 16 and set the learning rate to 0.0005 for ImageNet-R, and 0.001 for CIFAR100 and CUB200-2011. The prompt token length $L_{D} = L_{ND}$ is set to 5 for ImageNet-R and 3 for CIFAR100 and CUB200. We assign one D-Prompt and one ND-Prompt component for every pair of classes. The size of the prompt pool is set to half the number of classes. Source code is available at https://github.com/yeyeyeye33/SEC-Prompt. + +Evaluation protocol. Following previous works [51, 54], we denote the Top-1 accuracy after the $i$ -th session as $A_{i}$ . And the forgetting phenomenon is measured by the performance dropping rate (PD), i.e., $\mathrm{PD} = A_0 - A_T$ , where $A_0$ stands for the accuracy after the base session, and $A_T$ is the accuracy after the last session. Additionally, since it does not reflect the relationship between old and new class accuracies, we use Harmonic Accuracy (HAcc) [31] to evaluate balanced performance across both base and new classes after task $T: A_h = \frac{2 \times A_o \times A_n}{A_o + A_n}$ , where $A_o$ denotes the accuracy of base classes in task 0, and $A_n$ is the average accuracy across all classes in sessions $t > 0$ . + +Baselines. We compare recent state-of-the-art (SOTA) prompt-based CIL approaches, including L2P [44], DualP [43], and CodaP [34]. Additionally, we compare SOTA FSCIL algorithms such as TEEN [42], PriViLege [30], and ASP [25]. + +# 5.2. Benchmark Comparison + +In this section, we report the performance of baselines and SEC-Prompt under FSCIL setting. We also evaluate SECPrompt-F (Ours-F), where no additional trainable parameters are introduced during the incremental sessions. Table 1 presents the evaluation results on the ImageNet-R dataset. Figure 2 illustrates the evaluation results on the CUB200 and CIFAR100 datasets. + +Table 1. Comparison with SOTA methods on ImageNet-R dataset for FSCIL. More results on other datasets can be found in the appendix. + +
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]84.676.768.867.664.360.658.356.156.154.754.929.745.4
PriViLege [30]80.077.377.073.872.872.371.369.769.567.766.513.564.7
ASP [25]83.380.479.677.075.674.773.072.171.970.969.713.567.0
L2P [44]73.970.969.365.964.062.660.159.559.058.256.817.252.0
DualP [43]71.569.069.067.466.665.964.164.064.063.462.59.061.6
Coda-P [34]82.880.178.675.373.171.870.069.268.867.466.416.861.7
Ours-F82.980.579.877.576.776.374.874.674.173.873.09.972.1
Ours82.980.680.178.277.877.676.776.175.975.074.58.374.2
+ +Table 2. Ablation studies on ImageNet-R benchmark. D, ND, PC, DA denote D-Prompt, ND-Prompt, Discriminative Prompt Clustering Loss, Non-Discriminative Prompt-Based Data Augmentation, respectively. + +
DNDPCDAAcc. in each session ↑PD ↓HAcc ↑
012345678910
82.880.178.675.373.171.870.069.268.867.466.416.861.7
✓82.680.580.177.977.376.875.675.274.974.173.19.572.8
✓✓82.980.580.178.177.777.476.676.075.674.974.18.873.8
✓✓✓82.980.580.178.277.777.476.776.175.774.974.38.673.7
✓✓✓✓82.980.680.178.277.877.676.776.175.975.074.58.374.2
+ +Based on the experimental results on CIFAR100, CUB200, and ImageNet-R, SEC-Prompt outperforms other methods in every incremental task, except for the base session. In the base session, it is reasonable that the state-of-the-art (SOTA) FSCIL methods perform better, as they fine-tune the ViT parameters to fit the base class distribution. Furthermore, SEC-Prompt-F achieves the best performance among all methods, except SEC-Prompt itself, validating that our approach has strong generalization ability. Finally, SEC-Prompt achieves the best performance in both PD and HAcc, particularly on ImageNet-R, surpassing the most recent SOTA method ASP by $5.2\%$ and $7.2\%$ , respectively. This demonstrates that our method exhibits stronger stability and plasticity. + +# 5.3. Ablation Study + +We conducted an ablation study to demonstrate the significance of the components we proposed. As shown in Figure 3, our Adaptive Hierarchical Query mechanism significantly improves new class accuracy compared to the Fixed Query, demonstrating its ability to help the model learn more robust transferable knowledge and enhance the plasticity of new classes. To further validate our approach, we used Coda-Prompt as a baseline on the ImageNet-R dataset and progressively added components. As shown in Table 2, we observed a significant performance improvement + +Figure 2. Comparison with SOTA methods on Cifar100 and CUB200 dataset for FSCIL. +![](images/36365daa0289fd30c6f7f59d49f2f1bff9710732f27024879d8dc3fcda8fb6d1.jpg) +Ours TEEN ASP DualP Ours-F PriViLega L2P Coda-P + +![](images/ef792cd235bb8433a3033c75b32d4d77cadb1d19fa2d2768e0223a6adbb75600.jpg) + +ASP DualP ege L2P Coda-P + +when replacing Coda-Prompt with D-Prompt combined with Adaptive Hierarchical Query, supporting the rationale behind separating discriminative and non-discriminative information. Adding ND-Prompt further improved Harmonic Accuracy (HAcc), demonstrating that non-discriminative features contribute to better balancing stability and adaptability. Finally, the introduction of Discriminative Prompt Clustering Loss and ND-Prompt-Based Data Augmentation resulted in additional performance gains. + +![](images/9794be7c1410b12fea1186aaa1ca64b55fad1e9c701574a8a12ff120274a1114.jpg) +Figure 3. The impact of different query functions on new class accuracy in incremental sessions. "FQ" stands for Fixed Query, and "AHQ" represents Adaptive Hierarchical Query. The "-F" suffix indicates that the prompt pool is not expanded during the incremental session. + +![](images/6e12eb53c056cb4899ef33bfb20cbe129462c9a27110c12752ebc7c0d83549cf.jpg) +Figure 4. The t-SNE [39] visualization on the ImageNet-R dataset illustrates the embeddings learned by D-Prompt and SEC-Prompt. Class 0-4 represent the base classes while class 110 represent the novel classes. + +# 5.4. Further Analysis + +Analysis of ND-Prompt. To verify whether ND-Prompt promotes intra-class compactness by balancing non-discriminative features, we visualized the embeddings learned by D-Prompt and SEC-Prompt. As shown in Figure 4, incorporating ND-Prompt significantly reduced the intra-class distance of the base classes, with features clustering closer to the center. Although ND-Prompt does not update during the incremental sessions, the intra-class features of novel classes also became more compact. This suggests that the model has learned the distribution of nondiscriminative features from the base classes, which aids in the learning of new classes. These observations support the idea that ND-Prompt helps the model achieve a better balance between stability and plasticity during incremental learning. + +Analysis of hyper-parameter. We conducted a sensitivity analysis of $n$ and $\lambda$ across three benchmark datasets. Here, $n$ represents the number of groups of augmented samples generated by ND-Prompt-Based Data Augmentation, + +![](images/d5b959bb2de93f312c4517feeb8ed9e523b171e0a581115f7d1b22caa75d806a.jpg) +Figure 5. Sensitive analysis of $\lambda$ and n + +![](images/81061937159791e3149112f3d33ca499a27cad4fa97c0fc0268bf331060a8da9.jpg) + +
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]5.9105.9
Ours-F2.9102.9
Ours4.6104.6
+ +Table 3. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model. + +with each group containing the same number of samples as the batch size. The hyperparameter $\lambda$ is the coefficient of the Discriminative Prompt Clustering Loss. The results of the analysis are shown in Figure 5. To achieve the highest accuracy in the last session $(A_T)$ , we selected $\lambda = 0.5$ and $n = 2$ for all benchmark datasets. Further details on the selection of prompt-related hyperparameters are provided in the appendix. + +Analysis of Parameter Efficiency. Since the use of prompts introduces additional trainable parameters, we compared the parameter sizes of our method with other SOTA FSCIL approaches. As shown in Table 3, our method achieves higher performance with fewer parameters. This is primarily due to the separation of discriminative and nondiscriminative information, which allows the model to learn feature representations more efficiently. + +# 6. Conclusion + +This paper addresses the FSCIL problem with the SEC-Prompt framework. Inspired by human selective attention, we separate task knowledge into discriminative and non-discriminative components and design complementary prompts for each. The attention mechanism captures discriminative features, while reverse attention balances nondiscriminative ones, creating a compact, robust feature space. Our model achieves SOTA performance on three FSCIL benchmarks, demonstrating strong stability and adaptability. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (No. 62176271) and Basic and Applied Basic Research Foundation of Guangdong Province (Grant No. 2024A1515011692). The work is also partially supported by Key Laboratory of Cognitive Intelligence and Content Security, Ministry of Education (Grant no. RZZN202407, Harbin Institute of Technology). + +# References + +[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2 +[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 844-853, 2021. 2 +[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 2 +[4] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in neural information processing systems, 32, 2019. 2 +[5] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2 +[6] Shuhan Chen, Xiuli Tan, Ben Wang, Huchuan Lu, Xuelong Hu, and Yun Fu. Reverse attention-based residual network for salient object detection. IEEE Transactions on Image Processing, 29:3763-3776, 2020. 5 +[7] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 1 +[8] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2 +[9] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is + +worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6 +[11] Heitor Murilo Gomes, Jean Paul Barddal, Fabricio Enembreck, and Albert Bifet. A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR), 50(2):1-36, 2017. 1 +[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 6, 1 +[13] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 2 +[14] Qin Huang, Chunyang Xia, Chihao Wu, Siyang Li, Ye Wang, Yuhang Song, and C-C Jay Kuo. Semantic segmentation with reverse attention. arXiv preprint arXiv:1707.06426, 2017.5 +[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4, 2 +[16] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2 +[17] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2 +[18] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In International Conference on Learning Representations, 2018. 2 +[19] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2 +[20] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 2 +[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1 +[22] Dongkyu Lee, Zhiliang Tian, Lanqing Xue, and Nevin L. Zhang. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 93-102. Association for Computational Linguistics, 2021. 5 +[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 4 + +[24] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2 +[25] Chenxi Liu, Zhenyi Wang, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. Few-shot class incremental learning with attention-aware self-adaptive prompt. arXiv preprint arXiv:2403.09857, 2024. 2, 3, 6, 7, 8, 1 +[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2 +[27] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2 +[28] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 1 +[29] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11321-11329, 2019. 2 +[30] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 2, 6, 7, 8, 1 +[31] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 6 +[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1 +[33] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 2 +[34] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 1, 2, 3, 4, 6, 7 +[35] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24183-24192, 2023. 2 +[36] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF con + +ference on computer vision and pattern recognition, pages 12183-12192, 2020. 1, 2 +[37] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. arXiv preprint arXiv:1901.11356, 2019. 2 +[38] Gido M Van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):4069, 2020. 2 +[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8 +[40] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 1 +[41] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2 +[42] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7, 8, 1 +[43] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 1, 2, 6, 7 +[44] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 1, 2, 6, 7 +[45] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2 +[46] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 374-382, 2019. 2 +[47] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2 +[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 2 +[49] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF + +conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2 +[50] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11838-11847, 2023. 2 +[51] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 2, 6 +[52] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021. +[53] Mingli Zhu, Zihao Zhu, Sihong Chen, Chen Chen, and Baoyuan Wu. Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208, 2024. 2 +[54] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7746-7755, 2023. 2, 6 + +# SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning + +Supplementary Material + +# A. Additional Experimental Details + +# A.1. Dataset Details + +We follow [25] and use the CIFAR100 [21], ImageNet-R [12], and CUB200-2011 [40] datasets. CIFAR100 consists of 100 classes, each with 500 training images and 100 testing images. The CUB200 dataset contains 11,788 high-resolution bird images categorized into 200 species, with 5,994 samples for training and 5,794 samples for testing. The ImageNet-R dataset is a subset of ImageNet that includes 30,000 images across 200 object categories, focusing on artistic renditions such as paintings, cartoons, sculptures, and sketches. + +# A.2. Additional Results + +In the main paper, we presented the Top-1 accuracy of CUB200 and CIFAR100 in Figure 2. Here, we provide a detailed comparison on CUB200 in Table 4 and CIFAR100 in Table 5. + +Table 4. Comparison with SOTA methods on CUB200 dataset for FSCIL. + +
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]88.886.285.583.083.081.781.579.779.979.580.08.880.2
PriViLege [30]82.381.380.577.877.876.075.776.075.275.275.17.272.3
ASP [25]87.186.084.983.483.682.482.683.082.683.083.53.683.4
L2P [44]82.481.279.076.876.274.774.174.172.773.073.68.773.6
DualP [43]83.582.280.979.578.677.076.377.075.776.176.57.076.3
Coda-P [34]79.678.176.475.675.073.172.572.872.072.472.96.772.5
Ours-F87.586.685.584.684.983.883.383.883.783.884.43.284.3
Ours87.586.685.584.684.984.083.784.283.884.084.82.684.7
+ +Table 5. Comparison with SOTA methods on CIFAR100 dataset for FSCIL. + +
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678
TEEN [42]92.990.288.486.886.486.085.885.184.08.981.2
PriViLego [30]90.989.489.087.687.987.487.687.286.14.884.7
ASP [25]92.290.790.088.788.788.288.287.886.75.585.3
L2P [44]84.782.380.177.577.076.075.674.172.312.468.0
DualP [43]86.083.682.980.280.680.280.579.077.48.575.3
Coda-P [34]86.083.681.679.279.178.578.377.075.410.672.2
Ours-F92.090.590.288.888.988.488.488.086.85.285.7
Ours92.090.890.689.189.289.189.288.787.54.587.1
+ +Our SEC-Prompt framework achieves the best PD and HAcc on both datasets. Specifically, on CUB200, SEC-Prompt achieves a PD of $2.6\%$ and an HAcc of $24.7\%$ , surpassing the second-best method by $1.0\%$ and $1.3\%$ , respectively. On CIFAR100, SEC-Prompt achieves a PD of $4.5\%$ and an HAcc of $87.1\%$ , outperforming the second-best method by $0.3\%$ and + +$1.8\%$ , respectively. The performance advantage of SEC-Prompt on CUB200 and CIFAR100 is less pronounced compared to ImageNet-R. This is likely due to the closer distributional similarity of these datasets to ImageNet, which enables the pretrained ViT generalization capability to sufficiently balance stability and plasticity. Conversely, ImageNet-R, with its diverse styles such as cartoons, sculptures, and paintings, poses greater learning challenges and highlights the significance of non-discriminative information. + +
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.1101.1
Ours1.5101.5
+ +(a) CIFAR100 + +
MethodTrainable paramFinal param
TEEN [42]100100
PriViLege [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.7101.7
Ours2.8102.8
+ +(b) CUB200 + +Table 6. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model. + +Additionally, as shown in Table 6b and Table 6a, we report the parameter efficiency analysis on CUB200 and CIFAR100. Notably, the parameter gap between SEC-Prompt and SOTA methods narrows on CUB200, leading to a reduced difference in base class accuracy as well. + +# B. Additional Implementation Details + +In the main paper, we introduced the key experimental details. In this section, we provide an analysis of the selection of parameters, including the prompt insertion depth of the network, prompt length, and prompt pool size. In Figure 6, we present the results of our parameter selection experiments conducted on ImageNet-R, where LAcc represents the accuracy of the final session, and HAcc denotes the harmonic accuracy. + +![](images/4fbd107ae6333e47710cdb05f0778376b520ec2d7789438d4570541eeedb5b8d.jpg) +(a) Depth of D-Prompt + +![](images/c9c3d911ff35a09d1b63d309fddd00aa0e8d1cde2dd25138b642970bb8d75e40.jpg) +(b) Depth of ND-Prompt +Figure 6. Selection of prompt parameters, including Layer Depth, prompt length, and prompt pool size. LAcc represents the accuracy of the last session, while HAcc stands for Harmonic Accuracy. + +![](images/5d56d3a5098364449cdd0d96cd7820d83b58fa49a7554fe5a7c926cf924c8e33.jpg) +(c) Prompt Lenth + +![](images/3eae44799103e5997d6677480848abb5fb3e3946d2445402796aed1accbad1b9.jpg) +(d) Prompt Pool Size + +# B.1. Decoupled prompt positions + +Visual Prompt Tuning [15] involves concatenating the prompt with the input of a specified layer in the backbone to fine-tune the model. Since different layers of the backbone exhibit varying levels of feature abstraction, we conduct experiments to determine the optimal layer for prompt insertion. Specifically, as in [34, 43], we begin by inserting the prompt at $\text{Layer}_1$ and search for the optimal final layer for prompt insertion. Discriminative information is the dominant feature in classification tasks, while non-discriminative information assists the classification task by complementing the discriminative features. Therefore, we first select the insertion layer for D-Prompt independently, and then, based on this, search for the optimal insertion layer for ND-Prompt. As shown in Figures 6a and 6b, D-Prompt achieves the highest accuracy at Layer 8, while ND-Prompt achieves the highest accuracy at Layer 5, aligning with the motivation behind the design of SEC-Prompt. D-Prompt captures high-level discriminative features, whereas ND-Prompt captures low-level non-discriminative features. Similar to findings in [43], we observe that model performance decreases with increasing depth. Specifically, while base class accuracy improves, new class accuracy declines. We hypothesize that this is due to the semantic information in deep features causing overfitting, which adversely affects the model's generalization ability. + +# B.2. Prompt Lenth + +Prompt Length refers to the number of prompt tokens selected from the prompt pool for each instance query. As shown in 6c, increasing the prompt length initially improves accuracy but no longer contributes once the accuracy peaks, consistent with the findings in [34]. To achieve the optimal balance between parameter efficiency and model performance, we set the prompt length to 5. + +# B.3. Prompt Pool Size + +Coda-Prompt [34] learns task-specific prompts added to the prompt pool, selecting prompt pool sizes arbitrarily to validate model performance and parameter efficiency. In contrast, our SEC-Prompt framework divides prompts into D-Prompt and ND-Prompt to learn semantic feature knowledge, making it class-dependent. Thus, we select prompt pool sizes proportional to the number of classes. As shown in Figure 6d, for ImageNet-R with 200 classes, we choose [50, 100, 200, 400, 1000] as prompt pool sizes and set the prompt pool size to 100, allocating one prompt component for every two classes. Unlike Coda-Prompt, our model does not continuously benefit from increasing the prompt pool size, instead, performance declines after reaching its peak. We observe that a larger prompt pool size results in more severe forgetting. We hypothesize that while our method enhances learning capability in few-shot scenarios, the quality of learned knowledge for new classes remains suboptimal. A larger prompt pool size amplifies the impact on old knowledge, leading to greater forgetting. This raises an important future research question: how to further improve the quality of features learned in few-shot scenarios. \ No newline at end of file diff --git a/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/images.zip b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..422f88a9eeff26d8df5fff4aa7066c4f5dc0ccf6 --- /dev/null +++ b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd2c55b0a4e6a817be0da4c7416052ba381e32c35bef008120d0bf3b72d9e912 +size 521309 diff --git a/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/layout.json b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24698fbebff9c6348b72a8e5568ffceffa9f0d07 --- /dev/null +++ b/2025/SEC-Prompt_SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning/layout.json @@ -0,0 +1,10124 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 110, + 501, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 110, + 501, + 146 + ], + "spans": [ + { + "bbox": [ + 109, + 110, + 501, + 146 + ], + "type": "text", + "content": "SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 250, + 167, + 378, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 167, + 378, + 182 + ], + "spans": [ + { + "bbox": [ + 250, + 167, + 378, + 182 + ], + "type": "text", + "content": "Ye Liu" + }, + { + "bbox": [ + 250, + 167, + 378, + 182 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 250, + 167, + 378, + 182 + ], + "type": "text", + "content": " Meng Yang" + }, + { + "bbox": [ + 250, + 167, + 378, + 182 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 182, + 556, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 182, + 556, + 195 + ], + "spans": [ + { + "bbox": [ + 69, + 182, + 556, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 182, + 556, + 195 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou, Guangdong, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 195, + 541, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 195, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 84, + 195, + 541, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 195, + 541, + 209 + ], + "type": "text", + "content": "State Key Laboratory of Integrated Services Networks, Xidian University, Xi'an, Shaanxi, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 210, + 446, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 210, + 446, + 222 + ], + "spans": [ + { + "bbox": [ + 176, + 210, + 446, + 222 + ], + "type": "text", + "content": "liuy2629@mail2.sysu.edu.cn, yangm6@mail.sysu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 156, + 249, + 202, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 249, + 202, + 261 + ], + "spans": [ + { + "bbox": [ + 156, + 249, + 202, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 273, + 297, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 273, + 297, + 587 + ], + "spans": [ + { + "bbox": [ + 61, + 273, + 297, + 587 + ], + "type": "text", + "content": "Few-shot class-incremental learning (FSCIL) presents a significant challenge in machine learning, requiring models to integrate new classes from limited examples while preserving performance on previously learned classes. Recently, prompt-based CIL approaches leverage ample data to train prompts, effectively mitigating catastrophic forgetting. However, these methods do not account for the semantic features embedded in prompts, exacerbating the plasticity-stability dilemma in few-shot incremental learning. In this paper, we propose a novel and simple framework named SEmantic Complementary Prompt(SEC-Prompt), which learns two sets of semantically complementary prompts based on an adaptive query: discriminative prompts(D-Prompt) and non-discriminative prompts(ND-Prompt). D-Prompt enhances the separation of class-specific feature distributions by strengthening key discriminative features, while ND-Prompt balances nondiscriminative information to promote generalization to novel classes. To efficiently learn high-quality knowledge from limited samples, we leverage ND-Prompt for data augmentation to increase sample diversity and introduce Prompt Clustering Loss to prevent noise contamination in D-Prompt, ensuring robust discriminative feature learning and improved generalization. Our experimental results showcase state-of-the-art performance across three benchmark datasets, including CIFAR100, ImageNet-R and CUB datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 609, + 141, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 609, + 141, + 621 + ], + "spans": [ + { + "bbox": [ + 63, + 609, + 141, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 628, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 628, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 61, + 628, + 296, + 688 + ], + "type": "text", + "content": "AI applications, once deployed, frequently encounter novel classes within continuous data streams [11], necessitating the addition of these new classes to recognition tasks. This scenario drives the study of class incremental learning (CIL) [32], which incrementally learns new class con" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 250, + 548, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 548, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 548, + 331 + ], + "type": "text", + "content": "cepts post-deployment while preserving knowledge of previously learned classes. In more practical scenarios, new classes arise sporadically and usually lack sufficient data, for example, industrial defect detection and medical image recognition, which motivates the development of few-shot class incremental learning (FSCIL) [36] with only a limited number of samples for new classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 332, + 548, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 548, + 471 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 548, + 471 + ], + "type": "text", + "content": "Both CIL and FSCIL face a central challenge: catastrophic forgetting [28]. As models are incrementally trained to recognize new classes, fine-tuning or expanding parameters often overwrites previously acquired knowledge, leading to significant performance drops on earlier classes. Moreover, research in FSCIL has shown that learning from limited samples can lead to overfitting [7]: the model quickly memorizes small sample sets but fails to generalize, which compromises its robustness. Recent efforts have focused on transferring knowledge from base classes to enhance generalization or on adjusting new class distributions to stabilize learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 472, + 548, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 548, + 669 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 548, + 669 + ], + "type": "text", + "content": "Recent prompt-based continual learning methods [34, 43, 44] fix the backbone parameters while training additional prompt parameters using a key-query mechanism to learn task-specific knowledge. These methods store the learned knowledge in a prompt pool, effectively preventing forgetting. However, when only a few samples are available, they struggle to effectively learn the knowledge of new classes. In contrast, humans can easily learn to distinguish new categories without disrupting previously learned knowledge. This raises a key question: How do humans achieve FSCIL capabilities? The key to classification tasks lies in learning discriminative features. Humans can identify the key features of new categories while ignoring irrelevant information, which allows them to learn accurate patterns and avoid the influence of noise. Thus, we hypothesize that a selective attention mechanism plays a crucial role in this process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 670, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 547, + 704 + ], + "type": "text", + "content": "In this paper, we propose a novel framework called Semantic Complementary Prompting (SEC-Prompt) to enhance FSCIL by emulating the human selective attenuate the semantic representation of the human sentence." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 694, + 149, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 694, + 149, + 704 + ], + "spans": [ + { + "bbox": [ + 73, + 694, + 149, + 704 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25643" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 294, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 294, + 186 + ], + "type": "text", + "content": "tion mechanism. By designining two novel attentions, we propose to leverage extensive prior knowledge for generalization and introduce two prompt components: Discriminative Prompt (D-Prompt) and Non-Discriminative Prompt (ND-Prompt). Using distinct matching mechanisms to learn complementary knowledge, D-Prompt and ND-Prompt separately learn task-specific discriminative information and task-agnostic nondiscriminative information, addressing catastrophic forgetting and overfitting in FSCIL." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 186, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 295, + 441 + ], + "type": "text", + "content": "SEC-Prompt incorporates both querying and matching mechanisms. Firstly, we use the class token from the input of each Transformer layer as an adaptive query to capture multi-level feature information, thereby enhancing the model's generalization ability and improving its capability to learn new classes. Then, the D-Prompt utilizes an attention mechanism as a matching mechanism to learn and store discriminative features for each class. Specially, D-Prompt strengthens the discriminative features matching the instance, thereby facilitating class separation. To improve the quality of discriminative features, we propose a Prompt Clustering Loss that minimizes the distance between D-Prompts corresponding to the same class, thereby mitigating the potential introduction of noise. Finally, ND-prompt employs a reverse attention mechanism to capture task-agnostic non-discriminative features, Furthermore, a data augmentation strategy based on ND-Prompt is used to generate diverse samples by introducing perturbations to the ND-Prompt, thereby aiding in the mitigation of overfitting. We demonstrate robust performance across three datasets, outperforming previous state-of-the-art methods. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 445, + 295, + 597 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 62, + 445, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 503 + ], + "type": "text", + "content": "- We introduce the SEC-Prompt framework, which leverages Adaptive Hierarchical Query to capture feature semantics and designs distinct matching mechanisms to learn complementary semantic knowledge, thereby facilitating new class learning in FSCIL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 504, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 295, + 548 + ], + "type": "text", + "content": "- We propose D-Prompt, which leverages an attention mechanism to learn discriminative features, and introduce Prompt Clustering Loss to enhance the feature discrimination." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 550, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 550, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 550, + 295, + 597 + ], + "type": "text", + "content": "- We design a reverse attention mechanism as the matching strategy for ND-Prompt to learn non-discriminative features, and propose an ND-Prompt-based data augmentation strategy to enhance sample diversity." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 62, + 609, + 146, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 609, + 146, + 620 + ], + "spans": [ + { + "bbox": [ + 62, + 609, + 146, + 620 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 628, + 241, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 628, + 241, + 641 + ], + "spans": [ + { + "bbox": [ + 62, + 628, + 241, + 641 + ], + "type": "text", + "content": "2.1. Class-Incremental Learning (CIL)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 646, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 646, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 646, + 295, + 704 + ], + "type": "text", + "content": "Non-prompt-based approaches. To address the challenges of CIL, existing methods are primarily classified into two categories. Replay-based methods transfer old knowledge using knowledge distillation, where old exemplars are stored in a memory buffer [4, 13, 19, 26, 46], or a gen" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "content": "erator [17, 18, 29, 38] is trained to synthesize past data. Regularization-based methods [2, 3, 5, 37, 48] constrain changes to critical parameters to maintain a balance between stability and plasticity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 129, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 547, + 373 + ], + "type": "text", + "content": "Prompt-based approaches. Prompt-based methods [23, 24, 45] were initially introduced in Natural Language Processing (NLP) to guide attention toward specific downstream tasks using additional prompt information. Recently, prompt-based CIL approaches have emerged as a promising solution to mitigate catastrophic forgetting in incremental learning without relying on rehearsal strategies. L2P [44] was the first to propose using a key-query mechanism to select task-specific prompts from a prompt pool. DualP [43] added task-invariant prompts to share information across tasks. CodaP [34] introduced an attention-based key-query mechanism and trained the prompt components end-to-end. DAP [16] designed an adaptive prompt generator that creates instance-specific fine-grained instructions for each input. Recently, HideP [41] decomposed the CIL task into hierarchical components and optimized them individually. These methods encode new knowledge without modifying the backbone encoder's parameters, offering a significant advantage by avoiding exemplar storage. However, they face challenges in few-shot learning scenarios due to their reliance on sufficient training data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 383, + 547, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 547, + 396 + ], + "type": "text", + "content": "2.2. Few-Shot Class-Incremental Learning (FSCIL)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 401, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 401, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 312, + 401, + 547, + 564 + ], + "type": "text", + "content": "TOPIC [36] introduced Few-Shot Class-Incremental Learning to address Class-Incremental Learning challenges in scenarios with limited data availability. According to [51], existing FSCIL methods can be divided into two main approaches. Backward-compatible methods [8, 9, 20, 27, 33, 36, 49, 50, 54] focus on mitigating catastrophic forgetting and overfitting caused by parameter updates during new class learning. Forward-compatible methods [1, 35, 42, 47, 51-53], emphasize learning high-quality knowledge from base classes to facilitate generalization and improve new class learning. The former suffers from catastrophic forgetting due to fine-tuning the backbone, while the latter, by freezing the backbone, has limited capacity for learning new classes, leading to the stability-plasticity dilemma." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 564, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 564, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 564, + 547, + 704 + ], + "type": "text", + "content": "Prompt-based approaches. Recently, prompt-based methods have facilitated more efficient knowledge transfer and retention. For example, PKT [30] employs a ViT backbone and fine-tunes selected layers to balance generalization and discrimination, while leveraging prompts to encode domain-specific and semantic knowledge. ASP [25] uses a frozen ViT backbone, with attention-aware task-invariant prompts and self-adaptive task-specific prompts to transfer knowledge from base to new classes, thereby mitigating both forgetting and overfitting. These methods focus solely on task-specific discriminative information, while overlooking the importance of non-discriminative informa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25644" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 128 + ], + "type": "text", + "content": "tion for both old knowledge transfer and new knowledge learning. To address this, our approach decomposes base-class knowledge into discriminative and non-discriminative information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 140, + 138, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 138, + 153 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 138, + 153 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 159, + 159, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 159, + 172 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 159, + 172 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": "FSCIL aims to incrementally learn through a sequence of disjoint classes with a few samples in each class while retaining knowledge of previously learned classes. Let " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{\\mathrm{train}}^0,\\mathcal{D}_{\\mathrm{train}}^1,\\dots ,\\mathcal{D}_{\\mathrm{train}}^T\\}" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " represent the sequence of labeled training datasets across sessions, with " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " denoting the corresponding label set at each session " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "(t = 0,1,\\ldots ,T)" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": ". The label sets are disjoint among different sessions, i.e., " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^i\\cap \\mathcal{C}^j = \\emptyset" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "i\\neq j" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": ", while the label sets of the test set " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}^{t}" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " includes samples from base session to session " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": ", that is, the cumulative label sets " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^0\\cup \\mathcal{C}^1\\cup \\dots \\cup \\mathcal{C}^n" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": ". Generally, the training set in the 0-th session (also known as the base session), " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}^0" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": ", is a relatively large dataset with sufficient data for training. In each subsequent few-shot incremental session " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "(t > 0)" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": ", the training set " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}^t" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " classes, each containing " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": " samples, typically referred to as an " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 61, + 176, + 296, + 364 + ], + "type": "text", + "content": "-shot training set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 371, + 186, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 186, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 186, + 384 + ], + "type": "text", + "content": "3.2. Visual Prompt Tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "type": "text", + "content": "Visual prompt tuning(VPT) [15] introduces a small amount of task-specific learnable parameters to adapt transformer models for downstream vision tasks. In this paper, we do not alter the technical foundations of prompt tuning. Specifically, prompts are prepended to the inputs of the selected Transformer layer. For the " + }, + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "type": "text", + "content": "-th layer, let the input be " + }, + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "type": "inline_equation", + "content": "h_i \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 61, + 388, + 296, + 471 + ], + "type": "text", + "content": ". The operation of VPT is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 481, + 295, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 481, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 481, + 295, + 495 + ], + "type": "interline_equation", + "content": "f _ {\\text {P r o T}} \\left(p _ {i}, h _ {i}\\right) = \\operatorname {L a y e r} _ {i} \\left(\\left[ p _ {i}; h _ {i} \\right]\\right), \\tag {1}", + "image_path": "f2c11f19da45223358e5d1f1cd074c900300cba824f257fa3faaaf79d5a67c50.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathbb{R}^{L_p \\times D}" + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "text", + "content": " represents the prompt, " + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "text", + "content": " is its length (a hyperparameter), and " + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "inline_equation", + "content": "[\\cdot; \\cdot]" + }, + { + "bbox": [ + 61, + 502, + 296, + 596 + ], + "type": "text", + "content": " denotes concatenation along the sequence length dimension. As a result, we can fix the Transformer encoder and rely on a small number of parameters (the prompts) to learn session-specific information, effectively preventing forgetting. The key remaining challenge is how to design and update prompts to avoid interference between sessions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 605, + 209, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 605, + 209, + 619 + ], + "spans": [ + { + "bbox": [ + 62, + 605, + 209, + 619 + ], + "type": "text", + "content": "3.3. Attention-based Prompting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "text", + "content": "The prompt pool method typically trains additional prompt parameters using a key-query mechanism and stores the prompts from each session in the pool to preserve knowledge. This design consists of two key components: a query function " + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "text", + "content": ", which maps an input " + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "text", + "content": " into a feature space, and a matching function " + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "inline_equation", + "content": "m(q(x), k)" + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "text", + "content": ", which selects the most appropriate prompt from the pool based on the query " + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 61, + 623, + 296, + 706 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "content": "Query function. Existing prompt-based methods [15, 25, 34] define " + }, + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "content": " by utilizing the feature vector corresponding to the [class] token from the entire ViT as the query:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 399, + 128, + 547, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 128, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 399, + 128, + 547, + 140 + ], + "type": "interline_equation", + "content": "q (x) = \\theta (x) [ 0 ] \\tag {2}", + "image_path": "085d691737620812ae12e4e540d38ba61c9068c2f84a9c7206a63f9e2bffea73.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 547, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 547, + 190 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 144, + 547, + 190 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 144, + 547, + 190 + ], + "type": "text", + "content": " is the pretrained ViT encoder. Although it has achieved success in incremental learning, it struggles to learn new class knowledge in FSCIL due to its inability to account for feature abstraction levels and variations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": "Matching function. Coda-Prompt [34] introduces attention components that compute a weighted sum to generate a composite prompt. Specifically, it incorporates three sets of parameters: prompts " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{M \\times L_p \\times D}" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": " and attention " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": ". The attention parameters " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": " are used to select features of the query, facilitating effective matching with the keys " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": ". Each prompt " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "P^j \\in \\mathbb{R}^{L_p \\times D}" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": " in the prompt pool is associated with a corresponding key " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "K^j \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": " and attention parameter " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "A^j \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": ". Given a query " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": ", the attention parameters " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "A^j" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "inline_equation", + "content": "K^j" + }, + { + "bbox": [ + 313, + 191, + 547, + 331 + ], + "type": "text", + "content": " are utilized to compute the weighting scores between the query and each key:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 338, + 547, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 338, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 381, + 338, + 547, + 352 + ], + "type": "interline_equation", + "content": "\\alpha^ {j} = \\gamma (q (x) \\odot A ^ {j}, K ^ {j}) \\tag {3}", + "image_path": "fb5aa4825e394ea069682bdf40022d1226a91fec52698b1e96445de11fa67db2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "text", + "content": " is the element-wise product, " + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "text", + "content": " is the cosine similarity. The similarity scores " + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "text", + "content": " are then used to weight the prompts, resulting in the final prompt " + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "inline_equation", + "content": "p\\in \\mathbb{R}^{L_p\\times D}" + }, + { + "bbox": [ + 313, + 360, + 547, + 396 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 391, + 402, + 547, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 402, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 391, + 402, + 547, + 435 + ], + "type": "interline_equation", + "content": "\\mathrm {p} (\\mathrm {x}) = \\sum_ {j = 1} ^ {M} \\alpha^ {j} \\cdot P ^ {j} \\tag {4}", + "image_path": "8d07f220d4ad8c23cede85de690f11f7cb35f22f0ec98e0438ef8133a7676b9b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 442, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 547, + 548 + ], + "type": "text", + "content": "Intuitively, Coda-Prompt[34] introduces an attention mechanism as a prompt matching mechanism, assigning positive weights to prompts that match the query to enhance their discriminative features, and assigning negative weights to prompts that do not match the query to suppress their nondiscriminative features. However, the simultaneous learning of discriminative and non-discriminative features introduces conflicts, undermining training stability in incremental sessions, particularly within FSCIL settings." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 557, + 394, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 394, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 394, + 571 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 577, + 547, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 547, + 669 + ], + "type": "text", + "content": "Inspired by human selective attention mechanisms, we categorize task information into discriminative and nondiscriminative features and propose the SEC-Prompt framework to learn complementary features. By enhancing sparse discriminative features, we promote inter-class separation without mutual interference, while balancing dense nondiscriminative features reduces intra-class distance, thereby facilitating the learning of discriminative features." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 670, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 547, + 705 + ], + "type": "text", + "content": "Similar to recent prompt-based approaches [25], we fix the pre-trained ViT as the backbone and train the prompt to learn task-specific knowledge. The prompt tokens are" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25645" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 83, + 545, + 254 + ], + "blocks": [ + { + "bbox": [ + 65, + 83, + 545, + 254 + ], + "lines": [ + { + "bbox": [ + 65, + 83, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 65, + 83, + 545, + 254 + ], + "type": "image", + "image_path": "24f438f3cbc76b7841adc1d0e398861b5f69f995dafbe2053483ebc9fc7b9202.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 262, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 61, + 262, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 548, + 338 + ], + "type": "text", + "content": "Figure 1. An overview of our SEC-Prompt approach to few-shot incremental learning. For an input query, the weighting scores are computed using attention parameters and keys. D-Prompt applies the softmax function to enhance the most relevant features, focusing on learning sparse discriminative features critical for class separation. ND-Prompt utilizes a reverse operation followed by the ReLU function to amplify mismatched features, enabling the learning of dense non-discriminative features that contribute to a more comprehensive representation. During incremental sessions, D-Prompt and ND-Prompt freeze the existing parameters. D-Prompt introduces new trainable parameters to the current session, while ND-Prompt extends the dataset by generating diverse training samples. This is achieved by adding perturbations to the similarity score vector, thereby increasing the diversity of samples and enhancing the model's robustness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 357, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 357, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 61, + 357, + 296, + 555 + ], + "type": "text", + "content": "concatenated with the input tokens of the current layer, and discarded from the output. Figure 1 presents an overview of our proposed method. First, we propose an Adaptive Hierarchical Query function to adaptively capture hierarchical feature information (Section 4.1). Next, we introduce Semantic Complementary Prompt (SEC-Prompt) to learn transferable knowledge from the base session. SEC-Prompt consists of the D-Prompt and ND-Prompt modules, which use attention and reverse-attention mechanisms, respectively, to capture discriminative and non-discriminative information from the base session. In the incremental session, we further propose a Prompt Cluster Loss to enhance the quality of D-Prompt and a data augmentation method based on ND-Prompt to increase the diversity of new class samples (Section 4.2). Finally, we employ a cosine classifier and optimize the prompt components using Prompt Cluster Loss and cross-entropy classification loss (Section 4.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 569, + 217, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 217, + 582 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 217, + 582 + ], + "type": "text", + "content": "4.1. Adaptive Hierarchical Query" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": "Existing prompt-based methods rely on the fixed output of the ViT encoder as the query, which restricts the prompt modules at each ViT layer from adapting dynamically to feature variations. On one hand, the level of feature abstraction varies across Transformer layers. Shallow layers tend to capture more non-discriminative information, such as textures and edge details, while deeper layers emphasize abstract, class-specific features. Relying exclusively on the [class] token from the final layer as the query overlooks critical shallow-layer feature information, thereby constrain-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 357, + 547, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 547, + 450 + ], + "type": "text", + "content": "ing knowledge transfer. On the other hand, since the Vision Transformer (ViT) model is frozen, the queries remain static and fail to adapt to the evolving feature spaces. The prompt module depends entirely on the feature space of the pre-trained ViT to establish reliable mappings, which poses significant challenges for new-class learning. This difficulty arises because the model struggles to construct accurate mappings with limited sample availability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 450, + 547, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 547, + 509 + ], + "type": "text", + "content": "Therefore, we propose an Adaptive Hierarchical Query function to address the limitations of static queries. Specifically, for the prompt of the " + }, + { + "bbox": [ + 313, + 450, + 547, + 509 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 313, + 450, + 547, + 509 + ], + "type": "text", + "content": "-th Transformer layer, the [class] token from the input of the current layer is used as the query:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 372, + 518, + 547, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 518, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 372, + 518, + 547, + 532 + ], + "type": "interline_equation", + "content": "q _ {i + 1} (x) = \\operatorname {L a y e r} _ {i} ([ p _ {i}, h _ {i} ]) [ 0 ] \\tag {5}", + "image_path": "64a0b6d9bc3eed496a0b9110c1c45e3f7456d074d536776b27a9fe1b4ea458d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 313, + 540, + 547, + 574 + ], + "type": "text", + "content": " is the prompt and input token of the j-th layer. This enables the query to dynamically align with the evolving feature representations across different layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 492, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 492, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 492, + 595 + ], + "type": "text", + "content": "4.2. Semantic Complementary Prompt" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 599, + 547, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 547, + 646 + ], + "type": "text", + "content": "Based on previous work [15, 23], prompts act as task-specific extensions that guide the model to attend to relevant features and relationships, enabling adaptation to new tasks while preserving the base model's parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": "We observe that although Coda-Prompt [34] achieves strong adaptability, it also results in more severe forgetting. We believe this is due to the model learning unstable discriminative features. Because discriminative and non-discriminative features exhibit fundamentally different" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25646" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 296, + 266 + ], + "type": "text", + "content": "characteristics, simultaneously learning both types of features is challenging. Specifically, discriminative features are distinctive attributes shared by specific classes, contributing to positive weights exclusively for queries within those classes. In contrast, non-discriminative features are more sporadic, appearing in various instances without class specificity. These features can yield either positive or negative weights for queries belonging to the same class. Thus, using only the positive or negative nature of weights is insufficient to distinguish between discriminative and non-discriminative features. In fact, assigning weights to prompts based on negative similarity scores may introduce additional noise. To address this, we propose a combined approach using both attention and reverse attention mechanisms to separately model discriminative and non-discriminative features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 267, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 267, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 267, + 296, + 361 + ], + "type": "text", + "content": "Discriminative Prompt. Firstly, we introduce attention mechanism as the D-Prompt matching strategy to enhance focus on each instance's discriminative features. As illustrated in Figure 1, D-Prompt computes an attended query and calculates the cosine similarity between this attended query and the keys within Equation 3 to obtain similarity scores. The difference is that we apply a softmax operation to the similarity scores:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 148, + 365, + 296, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 365, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 148, + 365, + 296, + 396 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\mathrm {D}} ^ {j} = \\frac {e ^ {\\alpha_ {\\mathrm {D}} ^ {j}}}{\\sum_ {k} e ^ {\\alpha_ {\\mathrm {D}} ^ {k}}} \\tag {6}", + "image_path": "c20065322ea1a87c232dcaf08943a7e15a03af869ba8187190272100e12fbfff.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{D}}^{j}" + }, + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "type": "text", + "content": " is the similarity scores of " + }, + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 61, + 399, + 296, + 470 + ], + "type": "text", + "content": "-th D-Prompt. This allows the model to amplify the discriminative features of the current class while minimizing the influence of other information, thereby focusing on enhancing the discriminative characteristics. These scores are then used in Equation 4 to perform a weighted sum of the prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 471, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 471, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 61, + 471, + 296, + 575 + ], + "type": "text", + "content": "Discriminative Prompt Clustering Loss. To achieve classification performance, we optimize our model using cross-entropy loss " + }, + { + "bbox": [ + 61, + 471, + 296, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 61, + 471, + 296, + 575 + ], + "type": "text", + "content": " to separate features of different classes and aggregate features of the same class. However, the cross-entropy loss can easily lead D-Prompt to fit nondiscriminative features for better separation, which introduces noise into the discriminative information and harms the model's generalization ability. Therefore, we introduce the prompt clustering loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 580, + 298, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 298, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 298, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P C}} (x, y) = \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\left(1 - \\gamma \\left(p _ {\\mathrm {D}} \\left(x _ {i}\\right), p _ {\\mathrm {D}} \\left(x _ {j}\\right)\\right)\\right) \\cdot \\mathbb {I} \\left(y _ {i} = y _ {j}\\right) \\tag {7}", + "image_path": "4b82c4f4c5d16a1ba5cd1f2a982b8ab8ecb87a7dc55823d0d577b9b5c102d0f5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{D}}(x_i)" + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "text", + "content": " represents the result of the D-Prompt calculation for input " + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "text", + "content": ", as described in Equation 4. " + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "text", + "content": " is the batch size, and " + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 61, + 624, + 296, + 704 + ], + "type": "text", + "content": " denotes the cosine similarity used to measure the alignment between representations. By aligning discriminative features within similar samples, we encourage shared discriminative traits, avoiding noise introduction for individual sample separation. Notably, we also attempted to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 127 + ], + "type": "text", + "content": "include negative samples to construct contrastive loss, but this led to a performance degradation. We believe this is due to shared discriminative information across categories, i.e., horses and zebras share similar body shapes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 128, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 547, + 326 + ], + "type": "text", + "content": "Non-Discriminative Prompt. D-Prompt can learn discriminative information to achieve a separated feature space. However, due to discrepancies between the training set distribution and the true distribution, the model often struggles to distinguish between discriminative and non-discriminative features. This issue becomes more pronounced when only limited data is available, leading to severe overfitting. Non-discriminative features, often regarded as noise in classification tasks, can increase the sparsity of effective training samples and hinder the capture of discriminative features. Eliminating such noise is inherently challenging. To address this, we propose explicitly modeling non-discriminative features to balance their representation by emphasizing mismatched non-discriminative features. For example, with images of different styles, the underlying logic is to eliminate the impact of a particular style by introducing various styles." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 326, + 547, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 547, + 476 + ], + "type": "text", + "content": "The reverse attention mechanism [6, 14] has been employed in previous image detection tasks to model background information and reduce its interference. Recently, [22] utilized reverse attention to suppress style-related features during text generation, thereby minimizing their interference with content representation. Inspired by this, we adopt reverse attention as the matching mechanism for ND-Prompt. Specifically, ND-Prompt assigns high weights to dissimilar queries and keys while allocating low weights to similar queries and keys. Subsequently, we apply the ReLU function instead of softmax to truncate negative weights, preventing the relationships among non-discriminative features from being disrupted:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 484, + 547, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 484, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 351, + 484, + 547, + 499 + ], + "type": "interline_equation", + "content": "\\alpha_ {\\mathrm {N D}} ^ {j} = \\max (0, - \\gamma (q (x) \\odot A _ {\\mathrm {N D}} ^ {j}, K _ {\\mathrm {N D}} ^ {j})) \\tag {8}", + "image_path": "9d820b358eb638bace432d9674952a701f03e6e9ab30292a1ba138a8e5b0f759.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 547, + 598 + ], + "type": "text", + "content": "By balancing the representation of non-discriminative features, this approach effectively reduces intra-class distance and helps D-Prompt to learn more precise discriminative information. It's worth noting that we avoid using negative weights to generate prompts, as shifts in the feature space during incremental learning can make negative weights prone to introducing noise, thereby destabilizing the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 547, + 704 + ], + "type": "text", + "content": "ND-Prompt-Based Data Augmentation. The limited samples still make it challenging for the model to accurately capture the unique discriminative features of new classes. A naive approach to alleviating the overfitting problem is to increase the number of training samples. Our ND-Prompt naturally incorporates non-discriminative information that does not interfere with the learnable D-Prompt, which focuses on capturing the distinctive discriminative features of each class during the incremental phase. Therefore, we" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 317, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 317, + 753 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 317, + 753 + ], + "type": "text", + "content": "25647" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 115 + ], + "type": "text", + "content": "propose using ND-Prompt for data augmentation. We introduce spherical Gaussian perturbation into the weighting scores " + }, + { + "bbox": [ + 61, + 81, + 297, + 115 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 61, + 81, + 297, + 115 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 119, + 296, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 119, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 129, + 119, + 296, + 134 + ], + "type": "interline_equation", + "content": "\\tilde {\\alpha} _ {\\mathrm {N D}} = \\alpha_ {\\mathrm {N D}} + \\mathcal {N} (\\mu , \\sigma^ {2}) \\tag {9}", + "image_path": "d8726dd99f7c9fab8be6bdb77af542db47a420986181370f18e6ad57a85bf752.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " stands for spherical Gaussian distribution, " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " is the mean, " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " is the variance. Since " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " is a normalized vector, we set " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\mu = 0" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\sigma = 1" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " to sample perturbations from the standard spherical Gaussian distribution. These vectors " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " are then used to generate " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " sets of perturbed ND-Prompts with varying non-discriminative information for each original sample, where " + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 142, + 296, + 270 + ], + "type": "text", + "content": " is a hyperparameter. We treat the concatenated instances of different ND-Prompts as samples containing distinct discriminative information, which can increase sample diversity without affecting the discriminative information, thereby alleviating the model's overfitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 281, + 145, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 145, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 145, + 294 + ], + "type": "text", + "content": "4.3. Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 300, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 296, + 394 + ], + "type": "text", + "content": "Following previous work, we use a cosine classifier to classify these features into class predictions, as the cosine classifier effectively drives the feature extractor to learn discriminative features. Additionally, it learns an embedding space where features of the same class form tighter clusters, which aids in learning non-discriminative features as well. After training, we replace the classifier parameters with class prototypes to achieve better generalization:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 403, + 296, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 403, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 100, + 403, + 296, + 435 + ], + "type": "interline_equation", + "content": "c _ {k} = \\frac {1}{N _ {k}} \\sum_ {(x, y) \\in \\mathcal {D} _ {\\text {t r a i n}} ^ {t}} I [ y = c ] f _ {\\theta , N, D} (x) \\tag {10}", + "image_path": "13cbeec57ed9a033532de07a0dcc39e16db5836820dd3c6e6f5bdf0e89ed52f6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "inline_equation", + "content": "f_{\\theta ,N,D}" + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "content": " represents the ViT encoder with the inserted D-Prompt and ND-Prompt. " + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "content": " denotes the indicator function, and " + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "content": " is the sample size of class " + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}^{t}" + }, + { + "bbox": [ + 61, + 445, + 295, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 481, + 297, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 481, + 297, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 481, + 297, + 657 + ], + "type": "text", + "content": "Non-discriminative information is shared across all classes, and we assume the base classes contain sufficient non-discriminative information. Therefore, we freeze ND-Prompt after the base session. In contrast, each class possesses unique features, and relying solely on previous knowledge can be insufficient for separating new classes and may even lead to confusion. Consequently, in the incremental session, we freeze D-Prompt to prevent forgetting and extend it to learn the discriminative features of new classes. Specifically, we expand the prompt pool of D-Prompt based on the number of categories in the current task, assigning one prompt component for every two categories. We optimize using a common cross-entropy loss and a Discriminative Prompt Clustering Loss across all sessions. The overall training objective can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 668, + 296, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 668, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 120, + 668, + 296, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} (x, y) + \\lambda \\mathcal {L} _ {\\mathrm {P C}} (x, y) \\tag {11}", + "image_path": "625219c2560bbf678e4ab9daaa8634fefc29d95eeeb851970ec4048f557f126d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 693, + 179, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 693, + 179, + 705 + ], + "spans": [ + { + "bbox": [ + 62, + 693, + 179, + 705 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 62, + 693, + 179, + 705 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 62, + 693, + 179, + 705 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 80, + 392, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 80, + 392, + 94 + ], + "spans": [ + { + "bbox": [ + 313, + 80, + 392, + 94 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 100, + 427, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 100, + 427, + 113 + ], + "spans": [ + { + "bbox": [ + 313, + 100, + 427, + 113 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 117, + 548, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 548, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 548, + 210 + ], + "type": "text", + "content": "Dataset. We evaluate our method on three datasets: CIFAR100 [21], CUB200-2011 [40], and ImageNet-R [12], following the setup in [25]. For details, CIFAR100 are organized into 60 base classes and 40 incremental classes structured in a 5-way 5-shot FSCIL scenario for a total of 8 sessions. CUB200 and ImageNet-R are divided into 100 classes for the base task, and the left 100 classes are divided into ten 10-way 5-shot incremental tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "type": "text", + "content": "Training details. As adopted in [25], we use the ViT-B/16-1K [10], pre-trained on ImageNet1K, as the backbone. For all datasets, input images are resized to " + }, + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "type": "text", + "content": ". The model is trained for 30 epochs during the base session and for 10 epochs during each incremental session using the SGD optimizer. In the base session, we set the learning rate to 0.1 and the batch size to 64 for ImageNet-R, and 32 for CIFAR-100 and CUB-200-2011. In the incremental session, we use a batch size of 16 and set the learning rate to 0.0005 for ImageNet-R, and 0.001 for CIFAR100 and CUB200-2011. The prompt token length " + }, + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "type": "inline_equation", + "content": "L_{D} = L_{ND}" + }, + { + "bbox": [ + 313, + 211, + 548, + 397 + ], + "type": "text", + "content": " is set to 5 for ImageNet-R and 3 for CIFAR100 and CUB200. We assign one D-Prompt and one ND-Prompt component for every pair of classes. The size of the prompt pool is set to half the number of classes. Source code is available at https://github.com/yeyeyeye33/SEC-Prompt." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": "Evaluation protocol. Following previous works [51, 54], we denote the Top-1 accuracy after the " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": "-th session as " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": ". And the forgetting phenomenon is measured by the performance dropping rate (PD), i.e., " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "\\mathrm{PD} = A_0 - A_T" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "A_0" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": " stands for the accuracy after the base session, and " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": " is the accuracy after the last session. Additionally, since it does not reflect the relationship between old and new class accuracies, we use Harmonic Accuracy (HAcc) [31] to evaluate balanced performance across both base and new classes after task " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "T: A_h = \\frac{2 \\times A_o \\times A_n}{A_o + A_n}" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "A_o" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": " denotes the accuracy of base classes in task 0, and " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "A_n" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": " is the average accuracy across all classes in sessions " + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 313, + 398, + 548, + 537 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 538, + 548, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 548, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 548, + 597 + ], + "type": "text", + "content": "Baselines. We compare recent state-of-the-art (SOTA) prompt-based CIL approaches, including L2P [44], DualP [43], and CodaP [34]. Additionally, we compare SOTA FSCIL algorithms such as TEEN [42], PriViLege [30], and ASP [25]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 448, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 448, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 448, + 618 + ], + "type": "text", + "content": "5.2. Benchmark Comparison" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": "In this section, we report the performance of baselines and SEC-Prompt under FSCIL setting. We also evaluate SECPrompt-F (Ours-F), where no additional trainable parameters are introduced during the incremental sessions. Table 1 presents the evaluation results on the ImageNet-R dataset. Figure 2 illustrates the evaluation results on the CUB200 and CIFAR100 datasets." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25648" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 99, + 531, + 243 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 547, + 91 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 547, + 91 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 547, + 91 + ], + "type": "text", + "content": "Table 1. Comparison with SOTA methods on ImageNet-R dataset for FSCIL. More results on other datasets can be found in the appendix." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 99, + 531, + 243 + ], + "lines": [ + { + "bbox": [ + 77, + 99, + 531, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 99, + 531, + 243 + ], + "type": "table", + "html": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]84.676.768.867.664.360.658.356.156.154.754.929.745.4
PriViLege [30]80.077.377.073.872.872.371.369.769.567.766.513.564.7
ASP [25]83.380.479.677.075.674.773.072.171.970.969.713.567.0
L2P [44]73.970.969.365.964.062.660.159.559.058.256.817.252.0
DualP [43]71.569.069.067.466.665.964.164.064.063.462.59.061.6
Coda-P [34]82.880.178.675.373.171.870.069.268.867.466.416.861.7
Ours-F82.980.579.877.576.776.374.874.674.173.873.09.972.1
Ours82.980.680.178.277.877.676.776.175.975.074.58.374.2
", + "image_path": "1354e8d09e6b3e74c0e940622ea5f7b41ba6a1ea826824dc69a7f5aec4431f35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 283, + 541, + 383 + ], + "blocks": [ + { + "bbox": [ + 61, + 253, + 548, + 275 + ], + "lines": [ + { + "bbox": [ + 61, + 253, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 253, + 548, + 275 + ], + "type": "text", + "content": "Table 2. Ablation studies on ImageNet-R benchmark. D, ND, PC, DA denote D-Prompt, ND-Prompt, Discriminative Prompt Clustering Loss, Non-Discriminative Prompt-Based Data Augmentation, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 283, + 541, + 383 + ], + "lines": [ + { + "bbox": [ + 66, + 283, + 541, + 383 + ], + "spans": [ + { + "bbox": [ + 66, + 283, + 541, + 383 + ], + "type": "table", + "html": "
DNDPCDAAcc. in each session ↑PD ↓HAcc ↑
012345678910
82.880.178.675.373.171.870.069.268.867.466.416.861.7
✓82.680.580.177.977.376.875.675.274.974.173.19.572.8
✓✓82.980.580.178.177.777.476.676.075.674.974.18.873.8
✓✓✓82.980.580.178.277.777.476.776.175.774.974.38.673.7
✓✓✓✓82.980.680.178.277.877.676.776.175.975.074.58.374.2
", + "image_path": "5f85e03df315860667b5fc93f0790e626547aa288d271e755d0c081b55a5cb1f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "type": "text", + "content": "Based on the experimental results on CIFAR100, CUB200, and ImageNet-R, SEC-Prompt outperforms other methods in every incremental task, except for the base session. In the base session, it is reasonable that the state-of-the-art (SOTA) FSCIL methods perform better, as they fine-tune the ViT parameters to fit the base class distribution. Furthermore, SEC-Prompt-F achieves the best performance among all methods, except SEC-Prompt itself, validating that our approach has strong generalization ability. Finally, SEC-Prompt achieves the best performance in both PD and HAcc, particularly on ImageNet-R, surpassing the most recent SOTA method ASP by " + }, + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 61, + 402, + 296, + 564 + ], + "type": "text", + "content": ", respectively. This demonstrates that our method exhibits stronger stability and plasticity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 571, + 153, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 153, + 583 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 153, + 583 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": "We conducted an ablation study to demonstrate the significance of the components we proposed. As shown in Figure 3, our Adaptive Hierarchical Query mechanism significantly improves new class accuracy compared to the Fixed Query, demonstrating its ability to help the model learn more robust transferable knowledge and enhance the plasticity of new classes. To further validate our approach, we used Coda-Prompt as a baseline on the ImageNet-R dataset and progressively added components. As shown in Table 2, we observed a significant performance improvement" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 402, + 430, + 502 + ], + "blocks": [ + { + "bbox": [ + 316, + 402, + 430, + 502 + ], + "lines": [ + { + "bbox": [ + 316, + 402, + 430, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 402, + 430, + 502 + ], + "type": "image", + "image_path": "36365daa0289fd30c6f7f59d49f2f1bff9710732f27024879d8dc3fcda8fb6d1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 506, + 527, + 523 + ], + "lines": [ + { + "bbox": [ + 347, + 506, + 527, + 523 + ], + "spans": [ + { + "bbox": [ + 347, + 506, + 527, + 523 + ], + "type": "text", + "content": "Ours TEEN ASP DualP Ours-F PriViLega L2P Coda-P" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 533, + 547, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 533, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 547, + 555 + ], + "type": "text", + "content": "Figure 2. Comparison with SOTA methods on Cifar100 and CUB200 dataset for FSCIL." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 430, + 402, + 545, + 502 + ], + "blocks": [ + { + "bbox": [ + 430, + 402, + 545, + 502 + ], + "lines": [ + { + "bbox": [ + 430, + 402, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 430, + 402, + 545, + 502 + ], + "type": "image", + "image_path": "ef792cd235bb8433a3033c75b32d4d77cadb1d19fa2d2768e0223a6adbb75600.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 431, + 506, + 527, + 523 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 431, + 506, + 527, + 523 + ], + "spans": [ + { + "bbox": [ + 431, + 506, + 527, + 523 + ], + "type": "text", + "content": "ASP DualP ege L2P Coda-P" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 312, + 599, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 599, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 599, + 547, + 704 + ], + "type": "text", + "content": "when replacing Coda-Prompt with D-Prompt combined with Adaptive Hierarchical Query, supporting the rationale behind separating discriminative and non-discriminative information. Adding ND-Prompt further improved Harmonic Accuracy (HAcc), demonstrating that non-discriminative features contribute to better balancing stability and adaptability. Finally, the introduction of Discriminative Prompt Clustering Loss and ND-Prompt-Based Data Augmentation resulted in additional performance gains." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25649" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 86, + 248, + 201 + ], + "blocks": [ + { + "bbox": [ + 104, + 86, + 248, + 201 + ], + "lines": [ + { + "bbox": [ + 104, + 86, + 248, + 201 + ], + "spans": [ + { + "bbox": [ + 104, + 86, + 248, + 201 + ], + "type": "image", + "image_path": "9794be7c1410b12fea1186aaa1ca64b55fad1e9c701574a8a12ff120274a1114.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 209, + 297, + 264 + ], + "lines": [ + { + "bbox": [ + 61, + 209, + 297, + 264 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 297, + 264 + ], + "type": "text", + "content": "Figure 3. The impact of different query functions on new class accuracy in incremental sessions. \"FQ\" stands for Fixed Query, and \"AHQ\" represents Adaptive Hierarchical Query. The \"-F\" suffix indicates that the prompt pool is not expanded during the incremental session." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 277, + 293, + 388 + ], + "blocks": [ + { + "bbox": [ + 65, + 277, + 293, + 388 + ], + "lines": [ + { + "bbox": [ + 65, + 277, + 293, + 388 + ], + "spans": [ + { + "bbox": [ + 65, + 277, + 293, + 388 + ], + "type": "image", + "image_path": "6e12eb53c056cb4899ef33bfb20cbe129462c9a27110c12752ebc7c0d83549cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 397, + 296, + 440 + ], + "lines": [ + { + "bbox": [ + 61, + 397, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 61, + 397, + 296, + 440 + ], + "type": "text", + "content": "Figure 4. The t-SNE [39] visualization on the ImageNet-R dataset illustrates the embeddings learned by D-Prompt and SEC-Prompt. Class 0-4 represent the base classes while class 110 represent the novel classes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 464, + 162, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 162, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 162, + 475 + ], + "type": "text", + "content": "5.4. Further Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 482, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 482, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 61, + 482, + 296, + 656 + ], + "type": "text", + "content": "Analysis of ND-Prompt. To verify whether ND-Prompt promotes intra-class compactness by balancing non-discriminative features, we visualized the embeddings learned by D-Prompt and SEC-Prompt. As shown in Figure 4, incorporating ND-Prompt significantly reduced the intra-class distance of the base classes, with features clustering closer to the center. Although ND-Prompt does not update during the incremental sessions, the intra-class features of novel classes also became more compact. This suggests that the model has learned the distribution of nondiscriminative features from the base classes, which aids in the learning of new classes. These observations support the idea that ND-Prompt helps the model achieve a better balance between stability and plasticity during incremental learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "content": "Analysis of hyper-parameter. We conducted a sensitivity analysis of " + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "content": " across three benchmark datasets. Here, " + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "content": " represents the number of groups of augmented samples generated by ND-Prompt-Based Data Augmentation," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 79, + 424, + 159 + ], + "blocks": [ + { + "bbox": [ + 315, + 79, + 424, + 159 + ], + "lines": [ + { + "bbox": [ + 315, + 79, + 424, + 159 + ], + "spans": [ + { + "bbox": [ + 315, + 79, + 424, + 159 + ], + "type": "image", + "image_path": "d5b959bb2de93f312c4517feeb8ed9e523b171e0a581115f7d1b22caa75d806a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 168, + 500, + 179 + ], + "lines": [ + { + "bbox": [ + 361, + 168, + 500, + 179 + ], + "spans": [ + { + "bbox": [ + 361, + 168, + 500, + 179 + ], + "type": "text", + "content": "Figure 5. Sensitive analysis of " + }, + { + "bbox": [ + 361, + 168, + 500, + 179 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 361, + 168, + 500, + 179 + ], + "type": "text", + "content": " and n" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 79, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 437, + 79, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 437, + 79, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 437, + 79, + 547, + 159 + ], + "type": "image", + "image_path": "81061937159791e3149112f3d33ca499a27cad4fa97c0fc0268bf331060a8da9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 196, + 529, + 279 + ], + "blocks": [ + { + "bbox": [ + 332, + 196, + 529, + 279 + ], + "lines": [ + { + "bbox": [ + 332, + 196, + 529, + 279 + ], + "spans": [ + { + "bbox": [ + 332, + 196, + 529, + 279 + ], + "type": "table", + "html": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]5.9105.9
Ours-F2.9102.9
Ours4.6104.6
", + "image_path": "9c168a291ad4448c661b93dd1bc0a4f7f399a2e9e41c38895585c4b613d13910.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 287, + 548, + 330 + ], + "lines": [ + { + "bbox": [ + 313, + 287, + 548, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 548, + 330 + ], + "type": "text", + "content": "Table 3. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "text", + "content": "with each group containing the same number of samples as the batch size. The hyperparameter " + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "text", + "content": " is the coefficient of the Discriminative Prompt Clustering Loss. The results of the analysis are shown in Figure 5. To achieve the highest accuracy in the last session " + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "inline_equation", + "content": "(A_T)" + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "text", + "content": ", we selected " + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 313, + 357, + 548, + 450 + ], + "type": "text", + "content": " for all benchmark datasets. Further details on the selection of prompt-related hyperparameters are provided in the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 548, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 548, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 548, + 546 + ], + "type": "text", + "content": "Analysis of Parameter Efficiency. Since the use of prompts introduces additional trainable parameters, we compared the parameter sizes of our method with other SOTA FSCIL approaches. As shown in Table 3, our method achieves higher performance with fewer parameters. This is primarily due to the separation of discriminative and nondiscriminative information, which allows the model to learn feature representations more efficiently." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 566, + 384, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 384, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 384, + 578 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 588, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 588, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 548, + 705 + ], + "type": "text", + "content": "This paper addresses the FSCIL problem with the SEC-Prompt framework. Inspired by human selective attention, we separate task knowledge into discriminative and non-discriminative components and design complementary prompts for each. The attention mechanism captures discriminative features, while reverse attention balances nondiscriminative ones, creating a compact, robust feature space. Our model achieves SOTA performance on three FSCIL benchmarks, demonstrating strong stability and adaptability." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25650" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 157, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 157, + 93 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 157, + 93 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 99, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 99, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 99, + 296, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. 62176271) and Basic and Applied Basic Research Foundation of Guangdong Province (Grant No. 2024A1515011692). The work is also partially supported by Key Laboratory of Cognitive Intelligence and Content Security, Ministry of Education (Grant no. RZZN202407, Harbin Institute of Technology)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 193, + 119, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 193, + 119, + 205 + ], + "spans": [ + { + "bbox": [ + 63, + 193, + 119, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 213, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 213, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 213, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 296, + 266 + ], + "type": "text", + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28762-28771, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 267, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 267, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 296, + 319 + ], + "type": "text", + "content": "[2] Hongjoon Ahn, Jihwan Kwak, Subin Lim, Hyeonsu Bang, Hyojun Kim, and Taesup Moon. Ss-il: Separated softmax for incremental learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 844-853, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 319, + 296, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 319, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 67, + 319, + 296, + 372 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision (ECCV), pages 139-154, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 373, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 373, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 67, + 373, + 296, + 426 + ], + "type": "text", + "content": "[4] Rahaf Aljundi, Eugene Belilovsky, Tinne Tuytelaars, Laurent Charlin, Massimo Caccia, Min Lin, and Lucas Page-Caccia. Online continual learning with maximal interfered retrieval. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 426, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 426, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 67, + 426, + 296, + 469 + ], + "type": "text", + "content": "[5] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 470, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 296, + 512 + ], + "type": "text", + "content": "[6] Shuhan Chen, Xiuli Tan, Ben Wang, Huchuan Lu, Xuelong Hu, and Yun Fu. Reverse attention-based residual network for salient object detection. IEEE Transactions on Image Processing, 29:3763-3776, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 512, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 512, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 512, + 296, + 544 + ], + "type": "text", + "content": "[7] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. arXiv preprint arXiv:1904.04232, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 544, + 296, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 296, + 608 + ], + "type": "text", + "content": "[8] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2534-2543, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 609, + 296, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 609, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 67, + 609, + 296, + 661 + ], + "type": "text", + "content": "[9] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1255-1263, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 662, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 662, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 662, + 296, + 704 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 82, + 547, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 82, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 82, + 547, + 102 + ], + "type": "text", + "content": "worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 104, + 547, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 104, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 315, + 104, + 547, + 147 + ], + "type": "text", + "content": "[11] Heitor Murilo Gomes, Jean Paul Barddal, Fabricio Enembreck, and Albert Bifet. A survey on ensemble learning for data stream classification. ACM Computing Surveys (CSUR), 50(2):1-36, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 148, + 547, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 148, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 315, + 148, + 547, + 212 + ], + "type": "text", + "content": "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 6, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 213, + 547, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 213, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 547, + 266 + ], + "type": "text", + "content": "[13] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 268, + 547, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 268, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 547, + 309 + ], + "type": "text", + "content": "[14] Qin Huang, Chunyang Xia, Chihao Wu, Siyang Li, Ye Wang, Yuhang Song, and C-C Jay Kuo. Semantic segmentation with reverse attention. arXiv preprint arXiv:1707.06426, 2017.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 312, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 312, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 312, + 547, + 354 + ], + "type": "text", + "content": "[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 355, + 547, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 355, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 355, + 547, + 407 + ], + "type": "text", + "content": "[16] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11847-11857, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 410, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 547, + 441 + ], + "type": "text", + "content": "[17] Nitin Kamra, Umang Gupta, and Yan Liu. Deep generative dual memory network for continual learning. arXiv preprint arXiv:1710.10368, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 443, + 547, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 443, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 315, + 443, + 547, + 474 + ], + "type": "text", + "content": "[18] Ronald Kemker and Christopher Kanan. Fearnet: Brain-inspired model for incremental learning. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 475, + 547, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 547, + 518 + ], + "type": "text", + "content": "[19] Ronald Kemker, Marc McClure, Angelina Abitino, Tyler Hayes, and Christopher Kanan. Measuring catastrophic forgetting in neural networks. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 520, + 547, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 520, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 547, + 562 + ], + "type": "text", + "content": "[20] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 563, + 547, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 563, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 315, + 563, + 547, + 585 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 586, + 547, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 586, + 547, + 670 + ], + "spans": [ + { + "bbox": [ + 315, + 586, + 547, + 670 + ], + "type": "text", + "content": "[22] Dongkyu Lee, Zhiliang Tian, Lanqing Xue, and Nevin L. Zhang. Enhancing content preservation in text style transfer using reverse attention and conditional layer normalization. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 93-102. Association for Computational Linguistics, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "type": "text", + "content": "[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 2, 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "type": "text", + "content": "25651" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 82, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 296, + 113 + ], + "type": "text", + "content": "[24] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 115, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 115, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 64, + 115, + 295, + 158 + ], + "type": "text", + "content": "[25] Chenxi Liu, Zhenyi Wang, Tianyi Xiong, Ruibo Chen, Yihan Wu, Junfeng Guo, and Heng Huang. Few-shot class incremental learning with attention-aware self-adaptive prompt. arXiv preprint arXiv:2403.09857, 2024. 2, 3, 6, 7, 8, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 159, + 295, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 295, + 191 + ], + "type": "text", + "content": "[26] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 193, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 193, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 64, + 193, + 295, + 224 + ], + "type": "text", + "content": "[27] Pratik Mazumder, Pravendra Singh, and Piyush Rai. Few-shot lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2337-2345, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 225, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 225, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 64, + 225, + 295, + 267 + ], + "type": "text", + "content": "[28] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 270, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 270, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 64, + 270, + 295, + 322 + ], + "type": "text", + "content": "[29] Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jahnichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11321-11329, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 323, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 323, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 64, + 323, + 295, + 376 + ], + "type": "text", + "content": "[30] Keon-Hee Park, Kyungwoo Song, and Gyeong-Moon Park. Pre-trained vision and language transformers are few-shot incremental learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23881-23890, 2024. 2, 6, 7, 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 379, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 379, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 64, + 379, + 295, + 421 + ], + "type": "text", + "content": "[31] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In European Conference on Computer Vision, pages 382-397. Springer, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 422, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 422, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 64, + 422, + 295, + 474 + ], + "type": "text", + "content": "[32] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 476, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 476, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 64, + 476, + 295, + 529 + ], + "type": "text", + "content": "[33] Guangyuan Shi, Jiaxin Chen, Wenlong Zhang, Li-Ming Zhan, and Xiao-Ming Wu. Overcoming catastrophic forgetting in incremental few-shot learning by finding flat minima. Advances in neural information processing systems, 34: 6747-6761, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 532, + 295, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 532, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 64, + 532, + 295, + 605 + ], + "type": "text", + "content": "[34] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11909-11919, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 607, + 295, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 607, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 64, + 607, + 295, + 670 + ], + "type": "text", + "content": "[35] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24183-24192, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 672, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 672, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 64, + 672, + 295, + 704 + ], + "type": "text", + "content": "[36] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class incremental learning. In Proceedings of the IEEE/CVF con" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 83, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 83, + 547, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 83, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 333, + 83, + 547, + 102 + ], + "type": "text", + "content": "ference on computer vision and pattern recognition, pages 12183-12192, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 104, + 547, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 104, + 547, + 146 + ], + "spans": [ + { + "bbox": [ + 316, + 104, + 547, + 146 + ], + "type": "text", + "content": "[37] Michalis K Titsias, Jonathan Schwarz, Alexander G de G Matthews, Razvan Pascanu, and Yee Whye Teh. Functional regularisation for continual learning with gaussian processes. arXiv preprint arXiv:1901.11356, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 147, + 547, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 147, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 147, + 547, + 187 + ], + "type": "text", + "content": "[38] Gido M Van de Ven, Hava T Siegelmann, and Andreas S Tolias. Brain-inspired replay for continual learning with artificial neural networks. Nature communications, 11(1):4069, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 190, + 547, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 190, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 190, + 547, + 220 + ], + "type": "text", + "content": "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 222, + 546, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 546, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 546, + 253 + ], + "type": "text", + "content": "[40] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 255, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 547, + 306 + ], + "type": "text", + "content": "[41] Liyuan Wang, Jingyi Xie, Xingxing Zhang, Mingyi Huang, Hang Su, and Jun Zhu. Hierarchical decomposition of prompt-based continual learning: Rethinking obscured suboptimality. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 547, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 547, + 350 + ], + "type": "text", + "content": "[42] Qi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan, and Han-Jia Ye. Few-shot class-incremental learning via training-free prototype calibration. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7, 8, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 351, + 547, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 351, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 351, + 547, + 414 + ], + "type": "text", + "content": "[43] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 547, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 547, + 478 + ], + "type": "text", + "content": "[44] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 139-149, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 547, + 532 + ], + "type": "text", + "content": "[45] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 547, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 547, + 585 + ], + "type": "text", + "content": "[46] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 374-382, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 586, + 547, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 547, + 628 + ], + "type": "text", + "content": "[47] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. arXiv preprint arXiv:2302.03004, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 630, + 547, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 630, + 547, + 670 + ], + "spans": [ + { + "bbox": [ + 316, + 630, + 547, + 670 + ], + "type": "text", + "content": "[48] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In International conference on machine learning, pages 3987-3995. PMLR, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 672, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 672, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 316, + 672, + 547, + 704 + ], + "type": "text", + "content": "[49] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 318, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 318, + 753 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 318, + 753 + ], + "type": "text", + "content": "25652" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 82, + 295, + 364 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 82, + 82, + 294, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 82, + 294, + 103 + ], + "spans": [ + { + "bbox": [ + 82, + 82, + 294, + 103 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 12455-12464, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 104, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 104, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 64, + 104, + 295, + 158 + ], + "type": "text", + "content": "[50] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11838-11847, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 159, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 295, + 212 + ], + "type": "text", + "content": "[51] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9046–9056, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 213, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 213, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 64, + 213, + 294, + 265 + ], + "type": "text", + "content": "[52] Kai Zhu, Yang Cao, Wei Zhai, Jie Cheng, and Zheng-Jun Zha. Self-promoted prototype refinement for few-shot class incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6801-6810, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 266, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 266, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 64, + 266, + 294, + 308 + ], + "type": "text", + "content": "[53] Mingli Zhu, Zihao Zhu, Sihong Chen, Chen Chen, and Baoyuan Wu. Enhanced few-shot class-incremental learning via ensemble models. arXiv preprint arXiv:2401.07208, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 310, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 310, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 64, + 310, + 294, + 364 + ], + "type": "text", + "content": "[54] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7746-7755, 2023. 2, 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "type": "text", + "content": "25653" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 110, + 77, + 501, + 111 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 77, + 501, + 111 + ], + "spans": [ + { + "bbox": [ + 110, + 77, + 501, + 111 + ], + "type": "text", + "content": "SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 236, + 120, + 373, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 120, + 373, + 135 + ], + "spans": [ + { + "bbox": [ + 236, + 120, + 373, + 135 + ], + "type": "text", + "content": "Supplementary Material" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 158, + 240, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 240, + 172 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 240, + 172 + ], + "type": "text", + "content": "A. Additional Experimental Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 178, + 155, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 155, + 189 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 155, + 189 + ], + "type": "text", + "content": "A.1. Dataset Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 547, + 253 + ], + "type": "text", + "content": "We follow [25] and use the CIFAR100 [21], ImageNet-R [12], and CUB200-2011 [40] datasets. CIFAR100 consists of 100 classes, each with 500 training images and 100 testing images. The CUB200 dataset contains 11,788 high-resolution bird images categorized into 200 species, with 5,994 samples for training and 5,794 samples for testing. The ImageNet-R dataset is a subset of ImageNet that includes 30,000 images across 200 object categories, focusing on artistic renditions such as paintings, cartoons, sculptures, and sketches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 260, + 171, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 171, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 171, + 271 + ], + "type": "text", + "content": "A.2. Additional Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 277, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 547, + 300 + ], + "type": "text", + "content": "In the main paper, we presented the Top-1 accuracy of CUB200 and CIFAR100 in Figure 2. Here, we provide a detailed comparison on CUB200 in Table 4 and CIFAR100 in Table 5." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 77, + 329, + 531, + 474 + ], + "blocks": [ + { + "bbox": [ + 174, + 310, + 434, + 321 + ], + "lines": [ + { + "bbox": [ + 174, + 310, + 434, + 321 + ], + "spans": [ + { + "bbox": [ + 174, + 310, + 434, + 321 + ], + "type": "text", + "content": "Table 4. Comparison with SOTA methods on CUB200 dataset for FSCIL." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 329, + 531, + 474 + ], + "lines": [ + { + "bbox": [ + 77, + 329, + 531, + 474 + ], + "spans": [ + { + "bbox": [ + 77, + 329, + 531, + 474 + ], + "type": "table", + "html": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678910
TEEN [42]88.886.285.583.083.081.781.579.779.979.580.08.880.2
PriViLege [30]82.381.380.577.877.876.075.776.075.275.275.17.272.3
ASP [25]87.186.084.983.483.682.482.683.082.683.083.53.683.4
L2P [44]82.481.279.076.876.274.774.174.172.773.073.68.773.6
DualP [43]83.582.280.979.578.677.076.377.075.776.176.57.076.3
Coda-P [34]79.678.176.475.675.073.172.572.872.072.472.96.772.5
Ours-F87.586.685.584.684.983.883.383.883.783.884.43.284.3
Ours87.586.685.584.684.984.083.784.283.884.084.82.684.7
", + "image_path": "138356d2174484575412a339a311d4ee00fb214d6a34d9a51fa59bd00ca2e718.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 105, + 514, + 503, + 658 + ], + "blocks": [ + { + "bbox": [ + 171, + 495, + 437, + 506 + ], + "lines": [ + { + "bbox": [ + 171, + 495, + 437, + 506 + ], + "spans": [ + { + "bbox": [ + 171, + 495, + 437, + 506 + ], + "type": "text", + "content": "Table 5. Comparison with SOTA methods on CIFAR100 dataset for FSCIL." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 514, + 503, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 503, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 503, + 658 + ], + "type": "table", + "html": "
MethodAcc. in each session ↑PD ↓HAcc ↑
012345678
TEEN [42]92.990.288.486.886.486.085.885.184.08.981.2
PriViLego [30]90.989.489.087.687.987.487.687.286.14.884.7
ASP [25]92.290.790.088.788.788.288.287.886.75.585.3
L2P [44]84.782.380.177.577.076.075.674.172.312.468.0
DualP [43]86.083.682.980.280.680.280.579.077.48.575.3
Coda-P [34]86.083.681.679.279.178.578.377.075.410.672.2
Ours-F92.090.590.288.888.988.488.488.086.85.285.7
Ours92.090.890.689.189.289.189.288.787.54.587.1
", + "image_path": "d406d9ecb2e2a9872e3a54bd9e31af13d3a294443707b52ff8e6fbea3697896a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": "Our SEC-Prompt framework achieves the best PD and HAcc on both datasets. Specifically, on CUB200, SEC-Prompt achieves a PD of " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": " and an HAcc of " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "24.7\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": ", surpassing the second-best method by " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": ", respectively. On CIFAR100, SEC-Prompt achieves a PD of " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": " and an HAcc of " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "87.1\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": ", outperforming the second-best method by " + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 61, + 670, + 547, + 705 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 318, + 754 + ], + "type": "text", + "content": "25654" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 548, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 548, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 548, + 140 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 61, + 81, + 548, + 140 + ], + "type": "text", + "content": " , respectively. The performance advantage of SEC-Prompt on CUB200 and CIFAR100 is less pronounced compared to ImageNet-R. This is likely due to the closer distributional similarity of these datasets to ImageNet, which enables the pretrained ViT generalization capability to sufficiently balance stability and plasticity. Conversely, ImageNet-R, with its diverse styles such as cartoons, sculptures, and paintings, poses greater learning challenges and highlights the significance of non-discriminative information." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 145, + 276, + 229 + ], + "blocks": [ + { + "bbox": [ + 80, + 145, + 276, + 229 + ], + "lines": [ + { + "bbox": [ + 80, + 145, + 276, + 229 + ], + "spans": [ + { + "bbox": [ + 80, + 145, + 276, + 229 + ], + "type": "table", + "html": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLge [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.1101.1
Ours1.5101.5
", + "image_path": "e3613c9c7b3706f6865ecd37947e5ff3e57e5d1f4b98df25d3b6d0cb65a3ee1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 334, + 145, + 531, + 229 + ], + "blocks": [ + { + "bbox": [ + 154, + 231, + 201, + 239 + ], + "lines": [ + { + "bbox": [ + 154, + 231, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 154, + 231, + 201, + 239 + ], + "type": "text", + "content": "(a) CIFAR100" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 145, + 531, + 229 + ], + "lines": [ + { + "bbox": [ + 334, + 145, + 531, + 229 + ], + "spans": [ + { + "bbox": [ + 334, + 145, + 531, + 229 + ], + "type": "table", + "html": "
MethodTrainable paramFinal param
TEEN [42]100100
PriViLege [30]16.3102.3
ASP [25]3.0103.0
Ours-F1.7101.7
Ours2.8102.8
", + "image_path": "631b36936d668d8b481c47585cd3a857e8248cc5b52b9f532bb23e0f76430c37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 413, + 231, + 453, + 239 + ], + "lines": [ + { + "bbox": [ + 413, + 231, + 453, + 239 + ], + "spans": [ + { + "bbox": [ + 413, + 231, + 453, + 239 + ], + "type": "text", + "content": "(b) CUB200" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 249, + 548, + 272 + ], + "lines": [ + { + "bbox": [ + 61, + 249, + 548, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 548, + 272 + ], + "type": "text", + "content": "Table 6. Comparison of parameter effectiveness with SOTA FSCIL method. Trainable parameters and Final parameters represent the percentages of trainable and total parameters relative to the base ViT pre-trained model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 285, + 548, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 548, + 321 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 548, + 321 + ], + "type": "text", + "content": "Additionally, as shown in Table 6b and Table 6a, we report the parameter efficiency analysis on CUB200 and CIFAR100. Notably, the parameter gap between SEC-Prompt and SOTA methods narrows on CUB200, leading to a reduced difference in base class accuracy as well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 253, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 253, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 253, + 344 + ], + "type": "text", + "content": "B. Additional Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 350, + 548, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 548, + 397 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 548, + 397 + ], + "type": "text", + "content": "In the main paper, we introduced the key experimental details. In this section, we provide an analysis of the selection of parameters, including the prompt insertion depth of the network, prompt length, and prompt pool size. In Figure 6, we present the results of our parameter selection experiments conducted on ImageNet-R, where LAcc represents the accuracy of the final session, and HAcc denotes the harmonic accuracy." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 62, + 404, + 173, + 487 + ], + "blocks": [ + { + "bbox": [ + 62, + 404, + 173, + 487 + ], + "lines": [ + { + "bbox": [ + 62, + 404, + 173, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 173, + 487 + ], + "type": "image", + "image_path": "4fbd107ae6333e47710cdb05f0778376b520ec2d7789438d4570541eeedb5b8d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 488, + 154, + 498 + ], + "lines": [ + { + "bbox": [ + 82, + 488, + 154, + 498 + ], + "spans": [ + { + "bbox": [ + 82, + 488, + 154, + 498 + ], + "type": "text", + "content": "(a) Depth of D-Prompt" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 187, + 405, + 297, + 487 + ], + "blocks": [ + { + "bbox": [ + 187, + 405, + 297, + 487 + ], + "lines": [ + { + "bbox": [ + 187, + 405, + 297, + 487 + ], + "spans": [ + { + "bbox": [ + 187, + 405, + 297, + 487 + ], + "type": "image", + "image_path": "c9c3d911ff35a09d1b63d309fddd00aa0e8d1cde2dd25138b642970bb8d75e40.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 488, + 282, + 498 + ], + "lines": [ + { + "bbox": [ + 203, + 488, + 282, + 498 + ], + "spans": [ + { + "bbox": [ + 203, + 488, + 282, + 498 + ], + "type": "text", + "content": "(b) Depth of ND-Prompt" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 507, + 548, + 529 + ], + "lines": [ + { + "bbox": [ + 61, + 507, + 548, + 529 + ], + "spans": [ + { + "bbox": [ + 61, + 507, + 548, + 529 + ], + "type": "text", + "content": "Figure 6. Selection of prompt parameters, including Layer Depth, prompt length, and prompt pool size. LAcc represents the accuracy of the last session, while HAcc stands for Harmonic Accuracy." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 405, + 422, + 487 + ], + "blocks": [ + { + "bbox": [ + 313, + 405, + 422, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 405, + 422, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 422, + 487 + ], + "type": "image", + "image_path": "5d56d3a5098364449cdd0d96cd7820d83b58fa49a7554fe5a7c926cf924c8e33.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 488, + 395, + 498 + ], + "lines": [ + { + "bbox": [ + 340, + 488, + 395, + 498 + ], + "spans": [ + { + "bbox": [ + 340, + 488, + 395, + 498 + ], + "type": "text", + "content": "(c) Prompt Lenth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 438, + 405, + 547, + 487 + ], + "blocks": [ + { + "bbox": [ + 438, + 405, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 438, + 405, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 438, + 405, + 547, + 487 + ], + "type": "image", + "image_path": "3eae44799103e5997d6677480848abb5fb3e3946d2445402796aed1accbad1b9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 488, + 526, + 498 + ], + "lines": [ + { + "bbox": [ + 460, + 488, + 526, + 498 + ], + "spans": [ + { + "bbox": [ + 460, + 488, + 526, + 498 + ], + "type": "text", + "content": "(d) Prompt Pool Size" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 548, + 214, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 214, + 561 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 214, + 561 + ], + "type": "text", + "content": "B.1. Decoupled prompt positions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 61, + 564, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 564, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 564, + 548, + 705 + ], + "type": "text", + "content": "Visual Prompt Tuning [15] involves concatenating the prompt with the input of a specified layer in the backbone to fine-tune the model. Since different layers of the backbone exhibit varying levels of feature abstraction, we conduct experiments to determine the optimal layer for prompt insertion. Specifically, as in [34, 43], we begin by inserting the prompt at " + }, + { + "bbox": [ + 61, + 564, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\text{Layer}_1" + }, + { + "bbox": [ + 61, + 564, + 548, + 705 + ], + "type": "text", + "content": " and search for the optimal final layer for prompt insertion. Discriminative information is the dominant feature in classification tasks, while non-discriminative information assists the classification task by complementing the discriminative features. Therefore, we first select the insertion layer for D-Prompt independently, and then, based on this, search for the optimal insertion layer for ND-Prompt. As shown in Figures 6a and 6b, D-Prompt achieves the highest accuracy at Layer 8, while ND-Prompt achieves the highest accuracy at Layer 5, aligning with the motivation behind the design of SEC-Prompt. D-Prompt captures high-level discriminative features, whereas ND-Prompt captures low-level non-discriminative features. Similar to findings in [43], we observe that model performance decreases with increasing depth. Specifically, while base class accuracy improves, new class accuracy declines. We hypothesize that this is due to the semantic information in deep features causing overfitting, which adversely affects the model's generalization ability." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 744, + 317, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 744, + 317, + 753 + ], + "spans": [ + { + "bbox": [ + 293, + 744, + 317, + 753 + ], + "type": "text", + "content": "25655" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 81, + 150, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 81, + 150, + 93 + ], + "spans": [ + { + "bbox": [ + 64, + 81, + 150, + 93 + ], + "type": "text", + "content": "B.2. Prompt Lenth" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 98, + 547, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 98, + 547, + 143 + ], + "spans": [ + { + "bbox": [ + 64, + 98, + 547, + 143 + ], + "type": "text", + "content": "Prompt Length refers to the number of prompt tokens selected from the prompt pool for each instance query. As shown in 6c, increasing the prompt length initially improves accuracy but no longer contributes once the accuracy peaks, consistent with the findings in [34]. To achieve the optimal balance between parameter efficiency and model performance, we set the prompt length to 5." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 152, + 164, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 152, + 164, + 163 + ], + "spans": [ + { + "bbox": [ + 64, + 152, + 164, + 163 + ], + "type": "text", + "content": "B.3. Prompt Pool Size" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 168, + 547, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 168, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 64, + 168, + 547, + 284 + ], + "type": "text", + "content": "Coda-Prompt [34] learns task-specific prompts added to the prompt pool, selecting prompt pool sizes arbitrarily to validate model performance and parameter efficiency. In contrast, our SEC-Prompt framework divides prompts into D-Prompt and ND-Prompt to learn semantic feature knowledge, making it class-dependent. Thus, we select prompt pool sizes proportional to the number of classes. As shown in Figure 6d, for ImageNet-R with 200 classes, we choose [50, 100, 200, 400, 1000] as prompt pool sizes and set the prompt pool size to 100, allocating one prompt component for every two classes. Unlike Coda-Prompt, our model does not continuously benefit from increasing the prompt pool size, instead, performance declines after reaching its peak. We observe that a larger prompt pool size results in more severe forgetting. We hypothesize that while our method enhances learning capability in few-shot scenarios, the quality of learned knowledge for new classes remains suboptimal. A larger prompt pool size amplifies the impact on old knowledge, leading to greater forgetting. This raises an important future research question: how to further improve the quality of features learned in few-shot scenarios." + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "spans": [ + { + "bbox": [ + 293, + 745, + 317, + 753 + ], + "type": "text", + "content": "25656" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_content_list.json b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3806cf07adb16a9090cc06240ecfae88f8a73e54 --- /dev/null +++ b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_content_list.json @@ -0,0 +1,1808 @@ +[ + { + "type": "text", + "text": "SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haochen Li $^{1,4}$ Rui Zhang $^{2*}$ Hantao Yao $^{3}$ Xin Zhang $^{2}$ Yifan Hao $^{2}$ Xinkai Song $^{2}$", + "bbox": [ + 138, + 202, + 828, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaohui Peng $^{1}$ Yongwei Zhao $^{2}$ Chen Zhao $^{5}$ Yanjun Wu $^{5}$ Ling Li $^{1,4*}$", + "bbox": [ + 187, + 220, + 787, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Intelligent Software Research Center, Institute of Software, CAS", + "bbox": [ + 227, + 238, + 743, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Lab of Processors, Institute of Computing Technology, CAS", + "bbox": [ + 202, + 256, + 766, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Information Science and Technology, University of Science and Technology of China", + "bbox": [ + 102, + 273, + 867, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ University of Chinese Academy of Sciences $^{5}$ Institute of Software, CAS", + "bbox": [ + 181, + 292, + 787, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{haochen2021, pengshaohui, zhaochen, wuyanjun, liling}@iscas.ac.cn", + "bbox": [ + 192, + 311, + 774, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangrui, zhangxin, haoyifan, songxinkai, zhaoyongwei}@ict.ac.cn, yaohantao@ustc.edu.cn", + "bbox": [ + 96, + 330, + 870, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 359, + 313, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain to an unlabelled target domain. Traditional works focus on aligning visual features between domains to extract domain-invariant knowledge, and recent VLM-based DAOD methods leverage semantic information provided by the textual encoder to supplement domain-specific features for each domain. However, they overlook the role of semantic information in guiding the learning of visual features that are beneficial for adaptation. To solve the problem, we propose semantic entropy to quantify the semantic information contained in visual features, and design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to adaptively refine visual features with the semantic information of two domains. Semantic entropy reflects the importance of features based on semantic information, which can serve as attention to select discriminative visual features and suppress semantically irrelevant redundant information. Guided by semantic entropy, we introduce domain-aware attention modules into the visual encoder in SEEN-DA. It utilizes an inter-domain attention branch to extract domain-invariant features and eliminate redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. Comprehensive experiments validate the effectiveness of SEEN-DA, demonstrating significant improvements in cross-domain object detection performance.", + "bbox": [ + 76, + 391, + 472, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 819, + 209, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection [8, 33, 44, 45] have been proven to attain encouraging performance assisted by high-quality im", + "bbox": [ + 76, + 845, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ages with accurate annotations. However, due to the noticeable discrepancy between domains, they suffer severe performance degradation when dealing with unseen data in distinct scenes. To alleviate this problem, domain adaptive object detection (DAOD) [3] aims to transfer a detector trained on the labelled source domain to a unlabelled target domain.", + "bbox": [ + 500, + 359, + 890, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional DAOD works [3, 30, 47] focus on reducing domain discrepancy by aligning visual features between domains to extract domain-invariant knowledge. Additionally, [24,32,49,58] decouple the unique features of each domain, i.e.domain-specific features, to prevent its interference during feature alignment. Although effective, they train the detector with domain-shared and semantically agnostic category labels (such as one-hot encoding), which ignores the linguistic properties of categories and inter-domain differences and commonality, i.esemantic information. Recently, visual-language models (VLMs) [26,43,65] are proposed to ground visual with textual modalities. Through the textual encoder, they align visual features with the textual prompt, which is constructed from natural language and contains rich semantic information. Motivated by this, recent VLM-based DAOD methods, such as DA-Pro [23], utilize the textual encoder as detection head and tune the textual prompt to learn domain-invariant semantics while supplementing domain-specific semantics for each domain. However, they freeze the visual encoder and leverage semantic information solely in the detection head, limiting the discriminability of visual features.", + "bbox": [ + 496, + 455, + 893, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Summarily, despite promising performance, existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. They typically optimize the visual encoder with semantically agnostic category labels and adversarial loss, or directly freeze it, resulting in two limitations. First, they fail to filter out redundant information from visual fea", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25465", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a212bc74c96c4ce6cb01b6d07cad6562cc92ff3bc4ba1abe480c28abb9cf1fe1.jpg", + "image_caption": [ + "(a.1) Source Image and Bounding Box" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 339, + 145 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/908a4ff3771a5e61c585974b055e63d1c23ab40ee7f3fd10a995a26ff50b58b7.jpg", + "image_caption": [ + "(b.1) Detection Results on Source Image (DA-Pro[23])" + ], + "image_footnote": [], + "bbox": [ + 83, + 156, + 339, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/293a8e9c282983095cb06ea627fd148bb8b416015d3311a80b0c4e1f5ddea2be.jpg", + "image_caption": [ + "(c.1) Detection Results on Source Image (SEEN-DA (Ours))", + "Figure 1. Visualization of the detection results and attention map. (a) Source and Target domain image and bounding boxes. (b) Existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. Infected by the redundant information and lack of domain-specific semantic information, some objects that are correctly detected on the source image are misclassified or missed on the target image. (c) We propose semantic entropy to evaluate the importance of visual features based on domain and category semantic information. By integrating it as domain-aware attention in SEEN-DA, we eliminate redundant information and enhance domain-specific features, thereby improving the detector's discriminability for DAOD." + ], + "image_footnote": [], + "bbox": [ + 83, + 223, + 339, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a3d9277509ec7243cc9772c15b82d663d5bb3274b05a7ec3d00338cf4973a01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 88, + 483, + 145 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c6518125155277825551b384999203dbf6fb5a77b2d0e08bdbb2758378558d1.jpg", + "image_caption": [ + "(a.2) Target Image and Bounding Box", + "(b.2) Detection Results on Target Image (DA-Pro[23])" + ], + "image_footnote": [], + "bbox": [ + 354, + 156, + 483, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7e417aaf183812e4524c94166600cf237e0240a1354ae410791fb85d5a44ee6e.jpg", + "image_caption": [ + "(c.2) Detection Results on Target Image (SEEN-DA (Ours))" + ], + "image_footnote": [], + "bbox": [ + 354, + 223, + 483, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/978473a08a87f9f2bbcfd8871dfced3d83d740b65892bf964691f2687470e98c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 614, + 145 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5865e700f55eaf2b81d9307f6e69af825ad5b1cf9e6393192a387bf7bc6169b3.jpg", + "image_caption": [ + "Target Image (DA-Pro[23])" + ], + "image_footnote": [], + "bbox": [ + 486, + 146, + 614, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/38303ba9edf02acf570da08b7dc804221edaff239e9f735f2b004943376f7f1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 223, + 614, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Target Image (SEEN-DA (Ours))", + "bbox": [ + 486, + 280, + 620, + 290 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f06f523db07280dc62c5650973862f875db4fa1a8aeff94094a2539d2f5a9e0e.jpg", + "image_caption": [ + "(d) Domain-aware Attention Map in SEEN-DA (Ours)" + ], + "image_footnote": [], + "bbox": [ + 629, + 89, + 890, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures, which is semantically irrelevant and can lead to incorrect feature alignment. Redundant information includes attributes unrelated to object identification, such as color, texture, etc. Although they do not contain domain and category semantic information, they are often extracted by adversarially trained visual encoder since they exist in both domains. Therefore, such semantically irrelevant information may interfere with the alignment of domain-invariant features. In Fig. 1(b.1), existing methods detect the bicycle covered with a white cloth on the source image (clear day). However, since target images (foggy day) tend to be white overall, they misclassify the white cloth as a rider in Fig. 1(b.2). Second, existing methods discard domain-specific features in the visual encoder, which contain rich domain semantic information and are important for the discriminability on the target domain. For example, they correctly detect the bicycle and rider on the source image Fig. 1(b.1), but miss them in Fig. 1(b.2) due to insufficient feature learning of domain-specific properties of target domain, e.g.fog.", + "bbox": [ + 75, + 390, + 472, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve these problems, we propose semantic entropy which evaluates the importance of features based on domain and category semantic information, and accordingly design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to eliminate redundant information and supplement domain-specific features for DAOD. The semantic entropy of visual features refers to the amount of semantic information related to a certain domain. As the textual encoder of VLM provides rich domain and category semantic information in textual embedding, the semantic entropy can be expressed as the information entropy of the classification probability between the visual and textual embedding. Therefore, semantic entropy reflects the importance of visual features in adaptation. Low semantic entropy indicates", + "bbox": [ + 75, + 689, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that visual features match the textual embedding of a certain category and are likely to be transferable foreground objects. On the contrary, high semantic entropy means that they are unpredictable and may be redundant or even harmful for detection.", + "bbox": [ + 496, + 390, + 892, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Applying semantic entropy as attention, we design the SEEN-DA framework, introducing domain-aware attention modules into the visual encoder to select visual features with low semantic entropy and suppress regions with high ones. Specifically, the domain-aware attention module utilizes two parallel branches to learn inter- and intra-domain attention with semantic entropy. The inter-domain attention branch is shared between domains and evaluates the semantic entropy on domain-shared textual embedding, aiming to learn domain-invariant features while eliminating redundant information. The intra-domain attention branch is private in each domain to evaluate the semantic entropy on source/target textual embedding respectively, supplementing domain-specific features for each domain. The SEEN-DA is able to adaptively adjust visual features with the semantic commonalities and disparities between domains. As shown in Fig. 1(c), domain-aware attention map highlights features with rich semantic information, i.eforeground regions, and hinders the negative impact of background regions. Thereby, the proposed SEEN-DA correctly detects objects misclassified by existing methods, improving the discriminability of the detector on the target domain.", + "bbox": [ + 496, + 479, + 892, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the proposed method on four benchmarks. The experimental results highlight the effectiveness of our SEEN-DA, e.g., it reaches $57.5\\%$ mAP on Cityscapes $\\rightarrow$ Foggy Cityscapes, surpassing the state-of-the-art DA-Pro [23] by $1.6\\%$ .", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25466", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual-Language Models Large-scale visual-language models (VLMs) [25, 26, 43] embed visual and text modalities into a shared space, enabling cross-modal alignment. Pre-trained with an astonishing scale of image-text pairs, they demonstrate comprehensive visual understanding. As a pioneering work, CLIP [43] trains a visual encoder and a textual encoder with 400 million image-text pairs in a contrastive manner, showing promising classification ability for the seen and unseen classes. Following this spirit, [10, 16, 52, 65] freeze the textual encoder and adapt the visual encoder into Faster-RCNN [45] as a two-stage detector. In this work, we apply the textual encoder of CLIP [43] to infer textual embedding that contains semantic information, and the visual encoder of RegionCLIP [65] as the backbone.", + "bbox": [ + 76, + 114, + 468, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Adaptive Object Detection (DAOD) aims to adapt the object detector [45] trained on the labelled source domain to the unlabelled target domain. Previous approaches can be broadly divided into three categories: feature alignment, semi-supervised learning and VLM-based domain alignment. As a fundamental direction, feature alignment [12, 37, 39] aims to align the feature distributions of the two domains to generate domain-invariant features [3, 30, 47]. Typically, the domain discriminator is proposed to align feature distribution adversarially in image level [3], instance level [47, 63] and category level [29, 53]. To reduce the negative impact of aligning unique features of each domain, [24, 32, 34, 49, 54, 58] decouple the domain-invariant and domain-specific knowledge between two domains. In parallel, semi-supervised learning strives to generate unbiased training data with style transfer [6, 19, 35] and pseudo label [2, 7, 31, 50]. However, they train the detector with semantically agnostic category labels (such as one-hot encoding), ignoring the semantic information inherently contained in the category name. To overcome this, recent VLM-based domain alignment methods exploit the semantic information provided by the textual encoder of VLMs to learn semantic commonalities and disparities. Pioneeringly, DA-Pro [23] tuning the textual encoder into a domain-aware detection head, learning domain-invariant semantics while supplementing domain-specific semantics for each domain. Despite the promising performance, existing methods freeze the visual encoder and ignore using semantic information to guide the learning of visual features, which leads to two limitations. First, they fail to filter out redundant information from visual features, which is semantically irrelevant and can lead to incorrect feature alignment. Second, they discard domain-specific features in the visual encoder, which contain rich domain semantic information and are vital for DAOD.", + "bbox": [ + 76, + 325, + 472, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Entropy Optimization Information entropy was first utilized as entropy minimization in [15] for semi-supervised learning to learn decision rules for unlabelled data. Con-", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sidering its effectiveness, recent works introduce entropy optimization into DAOD. [40, 41, 46, 62] propose minimizing entropy on unlabelled data to make decisions consistent in the source and target domains. [55] apply entropy as metrics to measure the transferability of data samples. [38] employs entropy as attention to control the uncertainty of the classifier predictions, alleviating the effect of negative transfer. Instead of directly minimizing entropy, [60] further introduces diversity maximization to explore close-to-ideal domain adaptation solutions. However, the computation of entropy in existing methods is inherently semantically agnostic. Based on the output probabilities from a classifier, they treat categories merely as identifiers and convert the category labels into one-hot encoding. Therefore, they ignore the rich semantic information embedded within the category names, which is vital for DAOD. In this paper, we propose a novel semantic entropy, evaluating the transferability of visual features through the semantic commonalities and disparities between domains, improving the discrimination of the visual encoder.", + "bbox": [ + 496, + 90, + 893, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 405, + 635, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present a novel SEmantic entropy guided Domain-aware Attention (SEEN-DA) for DAOD, which introduces domain-aware attention module to eliminate redundant information and supplement domain-specific features, guided by the proposed semantic entropy. Semantic entropy quantifies the amount of semantic information within visual features, providing a measure of their importance in domain adaptation. Utilizing semantic entropy as attention, the domain-aware attention module adopts an inter-domain attention branch to extract domain-invariant features and eliminate semantically irrelevant redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. As shown in Fig. 2(a), given input image $x$ , the visual encoder of SEEN-DA is divided into $N$ frozen visual block $\\{\\mathcal{V}_i\\}_{i=1}^N$ , and each domain-aware attention module $\\{\\mathcal{A}_i\\}_{i=1}^N$ is attached to $\\mathcal{V}_i$ as a lightweight learnable module:", + "bbox": [ + 496, + 430, + 893, + 702 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {0} = \\mathcal {E} (x); \\mathbf {h} _ {i} = \\mathcal {A} _ {i} \\left(\\mathcal {V} _ {i} \\left(\\mathbf {h} _ {i - 1}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 705, + 890, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is the stem layer.", + "bbox": [ + 500, + 724, + 671, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Semantic Entropy (SEEN)", + "text_level": 1, + "bbox": [ + 500, + 750, + 738, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally speaking, given feature $\\mathbf{f}$ , the probability that this feature belongs to a specific object category is:", + "bbox": [ + 498, + 772, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (c, \\mathbf {f}) = \\frac {\\exp \\left(v _ {c} (\\mathbf {f})\\right)}{\\sum_ {c} \\exp \\left(v _ {c} (\\mathbf {f})\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 804, + 890, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c, v_{c}(\\cdot)$ represents an object category and its corresponding classifier. $p(\\cdot, \\mathbf{f})$ is the probability distribution function, expected to output 1 on the corresponding category and 0 on other category, i.e. one-hot vector.", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25467", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/141f1c0e6581a5a6849f274af53ffb0551d495f1a154bdb2cde23f68b4e371b2.jpg", + "image_caption": [ + "(a) Overview of SEEN-DA" + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 519, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef1b5a8e7dbefcf9007886a15b64d1c7b7196cebe73860b814f709b42e582c8c.jpg", + "image_caption": [ + "(b) Domain-aware Attention Module", + "Figure 2. (a) Overview of the proposed SEEN-DA for DAOD, where the semantic entropy is utilized as attention in domain-aware attention module. (b) The architecture of domain-aware attention module, consisting of an inter-domain and an intra-domain attention branch." + ], + "image_footnote": [], + "bbox": [ + 531, + 87, + 879, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tradition DAOD approaches focus on reducing domain discrepancy by aligning $\\mathbf{f}$ between the source and target domain. Although effective, as shown in Eq. 2, they treat categories merely as one-hot vectors, which is inherently semantically agnostic. Therefore, they ignore the rich semantic information inherently embedded within the category names. As feature alignment relies heavily on a profound understanding of the semantic disparities between the source and target domains, this limitation hinders the learning of visual features for adaptation.", + "bbox": [ + 75, + 286, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we propose semantic entropy, which quantifies the semantic information embedded in visual features by leveraging textual embeddings from Visual-language models (VLMs). VLMs utilize a textual encoder to encode textual prompts that describe category attributes into textual embeddings, aligning them with visual features. As a result, textual embeddings inherit rich semantic information from natural language, making them effective representations of the semantics of visual features. Motivated by this, we first infer textual embedding $\\mathbb{T} = \\{t_c\\}_c^K$ for $K$ categories by feeding prompts like \"A photo of [Class] in [Domain]\" into a textual encoder, where each $t_c$ contains semantic information of category $c$ and its domain. Thus, the probability distribution that $\\mathbf{f}$ corresponds to certain semantics can be estimated by the similarity between the visual feature $\\mathbf{f}$ and textual embedding $\\mathbb{T}$ :", + "bbox": [ + 75, + 438, + 468, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(t _ {c}, \\mathbf {f}\\right) = \\frac {\\exp \\left(\\sin \\left(t _ {c} , \\mathbf {f}\\right) / \\tau\\right)}{\\sum_ {c} \\exp \\left(\\sin \\left(t _ {c} , \\mathbf {f}\\right) / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 680, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot ,\\cdot)$ is the cosine similarity, and $\\tau$ is a learnable temperature parameter. Though $p(t_{c},\\mathbf{f})$ provides a measure of semantic relevance, it does not directly quantify the amount of semantic information embedded in $\\mathbf{f}$ . Therefore, we draw inspiration from the concept of information entropy, which is widely used to measure the uncertainty and informativeness of a probability distribution. Formally, the proposed semantic entropy is defined as the information entropy of the probability distribution $p(t_{c},\\mathbf{f})$ to evaluate the semantic uncertainty of $\\mathbf{f}$ :", + "bbox": [ + 75, + 714, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S E} (\\mathbb {T}, \\mathbf {f}) = - \\sum_ {c} p \\left(t _ {c}, \\mathbf {f}\\right) \\log \\left(p \\left(t _ {c}, \\mathbf {f}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 872, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the framework of DAOD, we apply semantic entropy as attention to select high discriminative visual features. According to Eq. 4, features with low semantic entropy exhibit clear and definite semantic information, expected to be foreground features suitable for transfer. Conversely, features with high semantic entropy indicate high semantic uncertainty, suggesting they may contain semantically irrelevant redundant information, which should be deemphasized during adaptation. To achieve this, we regulate Eq. 4 to propose semantic entropy-guided attention:", + "bbox": [ + 496, + 286, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S E A t t e n t i o n} (\\mathbb {T}, \\mathbf {f}) = \\sum_ {c} p \\left(t _ {c}, \\mathbf {f}\\right) \\log \\left(p \\left(t _ {c}, \\mathbf {f}\\right)\\right) + \\log K, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 444, + 890, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of categories. Based on semantic entropy, Eq. 5 effectively adjusts the weight of visual features, focusing on features with rich semantic information while minimizing the influence of redundant information.", + "bbox": [ + 496, + 488, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Domain-aware Attention Module", + "text_level": 1, + "bbox": [ + 500, + 556, + 790, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The domain-aware attention module is designed to adaptively refine visual features of each domain by utilizing the semantic entropy. As shown in Fig. 2(b), the module incorporates two complementary branches: an inter-domain attention branch to learn domain-invariant features and eliminate semantically irrelevant information, and an intra-domain attention branch to supplement the domain-specific features unique to each domain.", + "bbox": [ + 496, + 579, + 890, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inter-domain Attention The inter-domain attention is designed to modify domain-invariant features of the frozen visual encoder with semantic information. Specifically, for the output visual features $\\mathbf{f}_d = \\mathcal{V}_i(\\mathbf{h}_{i - 1})$ of the i-th visual block, we first forward it to a domain-shared convolution module $\\mathcal{C}$ to capture task-related knowledge:", + "bbox": [ + 496, + 704, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} = \\mathcal {C} \\left(\\mathbf {f} _ {d}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 801, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d \\in \\{\\mathrm{s},\\mathrm{t}\\}$ denotes whether $\\mathbf{f}_d$ belongs to source domain s or target domain t.", + "bbox": [ + 496, + 824, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously, the visual feature $\\mathbf{f}_d$ is forwarded to the SEEN module $S$ to obtain attention map $\\mathbf{w}$ guided by semantic entropy. With the help of VLM, we can represent the", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25468", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domain-invariant semantic information through domain-shared prompts. Technically, we input textual prompts such as \"A photo of [Class]\" to the textual encoder to retrieve domain-invariant textual embedding $\\mathbb{T}$ . As described in Eq. 4, the similarity of visual features and textual embeddings is needed to calculate semantic entropy attention. To facilitate this, a projection layer $\\mathcal{P}$ is firstly introduced to map $\\mathbf{f}_d$ from the visual feature space to the textual embedding space:", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathrm {p}} = \\mathcal {P} \\left(\\mathbf {f} _ {d}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 228, + 468, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we calculate the semantic entropy-guided attention of the visual feature $\\mathbf{f}_d$ over text embedding $\\mathbb{T}$ :", + "bbox": [ + 76, + 252, + 468, + 285 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = \\mathcal {S} (\\mathbb {T}, \\mathbf {f} _ {d}) = \\operatorname {S E A t t e n t i o n} (\\mathbb {T}, \\mathbf {f} _ {\\mathrm {p}}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 296, + 468, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the attention map $\\mathbf{w}$ , we can focus on semantically informative features and suppress redundant information in domain-invariant features. Typically, $\\mathbf{w}$ is applied to $\\mathbf{f}$ via an element-wise product, and the weighted feature is summed with the original visual feature:", + "bbox": [ + 76, + 325, + 468, + 401 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {\\prime} = \\mathbf {w} \\cdot \\mathbf {f} + \\mathbf {f} _ {d}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 414, + 468, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expecting to be domain-invariant, the extracted $\\mathbf{f}'$ is forwarded to a domain discriminator $\\mathcal{D}$ , which aligns the source and target domain adversarily.", + "bbox": [ + 76, + 443, + 468, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intra-domain Attention With the inter-domain attention branch, the extracted domain-invariant features contain sufficient semantic information shared between domains. However, it ignores the semantic discrepancy between the two domains, resulting in limited discriminability in the target domain. Therefore, we design the intra-domain attention branch to enrich visual features with domain-specific semantic information. Similar to the inter-domain attention branch, we first introduce a learnable layer $\\mathcal{C}^c$ to adapt visual features for to both domains:", + "bbox": [ + 75, + 510, + 468, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {d} ^ {\\mathrm {c}} = \\mathcal {C} ^ {\\mathrm {c}} \\left(\\mathbf {f} _ {d}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 672, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we generate domain-aware textual embedding that captures the semantics of each domains. Motivated by DA-Pro [23], we tune learnable domain-adaptive prompts, i.e. $\\left[[v^{\\mathrm{c}}][v^{\\mathrm{s}}]\\right]$ [Class] for the source domain and $\\left[[v^{\\mathrm{c}}][v^{\\mathrm{t}}]\\right]$ [Class] for the target domain, to explicitly model the domain and category semantic commonalities and disparities. These prompts are forwarded to the textual encoder to derive domain-adaptive textual embedding $\\mathbb{T}_{\\mathrm{s}},\\mathbb{T}_{\\mathrm{t}}$ for each domain.", + "bbox": [ + 75, + 703, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the visual and semantic differences between the source and target domain, a domain-shared SEEN module is inadequate. To calculate the semantic entropy for each domain, visual features must be independently mapped into", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "their respective textual spaces. Therefore, we apply separate SEEN modules $S^{\\mathrm{s}}$ and $S^{\\mathrm{t}}$ with independent projection layers $\\mathcal{P}^{\\mathrm{s}},\\mathcal{P}^{\\mathrm{t}}$ and textual embedding $\\mathbb{T}_{\\mathrm{s}},\\mathbb{T}_{\\mathrm{t}}$ :", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {\\mathrm {s}} = \\mathcal {S} ^ {\\mathrm {s}} \\left(\\mathbb {T} _ {\\mathrm {s}}, \\mathbf {f} _ {\\mathrm {s}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 143, + 890, + 160 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {\\mathrm {t}} = \\mathcal {S} ^ {\\mathrm {t}} \\left(\\mathbb {T} _ {\\mathrm {t}}, \\mathbf {f} _ {\\mathrm {t}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 176, + 890, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\mathbf{w}_{\\mathrm{s}}$ and $\\mathbf{w}_{\\mathrm{t}}$ denotes the source- and target-specific attention maps.", + "bbox": [ + 496, + 196, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The inter-domain attention branch extracts domain-invariant features while eliminating redundant information, whereas the intra-domain attention branch enriches visual features with domain-specific semantics. Overall, the output for the i-th domain-aware attention module $\\mathcal{A}_i$ can be expressed as follows:", + "bbox": [ + 496, + 228, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathrm {s}} ^ {\\prime} = \\mathcal {A} _ {i} \\left(\\mathbf {f} _ {\\mathrm {s}}\\right) = \\mathbf {f} ^ {\\prime} + \\mathbf {w} _ {\\mathrm {s}} \\cdot \\mathbf {f} _ {\\mathrm {s}} ^ {\\mathrm {c}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 325, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\mathrm {t}} ^ {\\prime} = \\mathcal {A} _ {i} \\left(\\mathbf {f} _ {\\mathrm {t}}\\right) = \\mathbf {f} ^ {\\prime} + \\mathbf {w} _ {\\mathrm {t}} \\cdot \\mathbf {f} _ {\\mathrm {t}} ^ {\\mathrm {c}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 359, + 890, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{\\mathrm{s}}^{\\prime}$ and $\\mathbf{f}_{\\mathrm{t}}^{\\prime}$ are output features of $\\mathcal{A}_i$ for the source and target domain. Since $\\mathbf{w}_{\\mathrm{s}}$ and $\\mathbf{w}_{\\mathrm{t}}$ are involved in the computation of $\\mathbf{f}_{\\mathrm{s}}^{\\prime}$ and $\\mathbf{f}_{\\mathrm{t}}^{\\prime}$ , SEEN modules $\\mathcal{S}^{\\mathrm{s}}$ and $\\mathcal{S}^{\\mathrm{t}}$ will be optimized by the classification losses from the two domains respectively, thereby learning domain-specific semantic information.", + "bbox": [ + 496, + 378, + 890, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimization Objective", + "text_level": 1, + "bbox": [ + 500, + 479, + 718, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce multiple domain-aware attention modules into the frozen visual encoder as learnable attachments. To this end, we expect them to capture visual features suitable for cross-domain adaptation and supplement domain-specific semantic information. First, the domain discriminator $\\mathcal{D}$ is applied to the inter-domain attention branch to learn domain-invariant features. Formally, we minimize the adversarial loss over the output of the inter-domain attention branch $\\mathbf{f}'$ for the source image $x_{\\mathrm{s}}$ and the target image $x_{\\mathrm{t}}$ :", + "bbox": [ + 496, + 503, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} = - \\sum_ {i = 1} ^ {N} \\left[ \\mathbb {E} _ {x _ {\\mathrm {s}}} \\left\\| \\mathcal {D} _ {i} \\left(\\mathbf {f} ^ {\\prime}\\right) \\right\\| _ {2} ^ {2} + \\mathbb {E} _ {x _ {\\mathrm {t}}} \\left\\| \\mathcal {D} _ {i} \\left(\\mathbf {f} ^ {\\prime}\\right) - \\mathbf {1} \\right\\| _ {2} ^ {2} \\right], \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 667, + 890, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, the detection loss is calculated over both domains to learn the semantic commonalities and disparities. For the source domain, we minimize the cross-entropy $\\mathcal{L}_{\\mathrm{s}}$ between the logits and ground truth. For the target domain, we use the prompt \"A photo of [Class] in foggy day\" to generate high-confidence pseudo labels and minimize the cross-entropy $\\mathcal{L}_{\\mathrm{t}}$ . Overall, the optimization objective is:", + "bbox": [ + 496, + 720, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {t}} \\mathcal {L} _ {\\mathrm {t}} + \\lambda_ {\\mathrm {a d v}} \\mathcal {L} _ {\\mathrm {a d v}} + \\mathcal {L} _ {\\mathrm {r e g}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 840, + 890, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{reg}}$ is the regression loss, and $\\lambda_{\\mathrm{t}}, \\lambda_{\\mathrm{adv}}$ are balance ratios.", + "bbox": [ + 496, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4ebfcaa68475b041233126fbe52572fd9a9d638665cc7e2c97274a7227be7e72.jpg", + "table_caption": [ + "Table 1. Comparison (%) with existing methods on Cross-Weather adaptation Cityscapes $\\rightarrow$ Foggy Cityscapes (C→F), Cross-Fov adaptation KITTI $\\rightarrow$ Cityscapes (K→C) and Sim-to-Real adaptation SIM10K $\\rightarrow$ Cityscapes (S→C)." + ], + "table_footnote": [], + "table_body": "
MethodsVenueC→FK→CS→C
PersonRiderCarTruckBusTrainMotorBicyclemAPmAPmAP
DA-Faster [3]CVPR 201829.240.443.419.738.328.523.732.732.041.938.2
VDD [59]ICCV 202133.444.051.733.952.034.734.236.840.0--
SCAN [28]AAAI 202241.743.957.328.748.648.731.037.342.145.852.6
TIA [64]CVPR 202252.138.149.737.734.846.348.631.142.344.0-
LRA [42]TNNLS 202445.647.159.731.252.444.628.139.543.549.455.7
SIGMA++ [30]TPAMI 202346.445.161.032.152.244.634.839.944.549.557.7
CIGAR [36]CVPR 202346.147.362.127.856.644.333.741.344.948.558.5
OADA [61]ECCV 202247.846.562.932.148.550.934.339.845.447.859.2
MTM [57]AAAI 202451.053.467.237.254.441.638.447.748.9-58.1
AT [31]CVPR 202256.351.964.238.545.555.154.335.050.9--
SOCKER [5]ACMMM 202451.757.768.638.251.647.541.651.751.1-63.8
DSD-DA [11]ICML 202449.059.665.335.761.046.543.957.352.349.352.5
CAT [22]CVPR 202444.657.163.740.866.049.744.953.052.5--
NSA-UDA [66]ICCV 202350.360.167.737.457.446.947.354.352.755.656.3
REACT [24]TIP 202451.457.967.437.758.452.844.654.653.159.558.6
DA-Pro [23]NeurIPS 202355.462.970.940.363.454.042.358.055.961.462.9
RegionCLIP [65]CVPR 202249.655.063.234.155.648.336.047.048.659.158.9
SEEN-DA (Ours)-58.564.571.742.061.254.847.159.957.567.166.8
", + "bbox": [ + 119, + 122, + 848, + 393 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e68593b47a0619b95958ebd7a5efb13e8674adbf7f0bd8e568a2a995176deb9.jpg", + "table_caption": [ + "Table 2. Comparison (%) with existing methods on Cross-Style adaptation task Pascal VOC→Clipart" + ], + "table_footnote": [], + "table_body": "
MethodsAeroBikeBirdBoatBottleBusCarCatChairCowTableDogHorseMotorPersonPlantSheepSofaTrainTvmAP
UaDAN [17]35.073.741.024.421.369.853.52.334.261.231.029.547.963.662.261.313.97.648.623.940.2
TFD [54]27.964.828.429.525.764.247.713.547.550.950.821.333.960.265.642.515.140.545.548.641.2
FGRR [1]30.852.135.132.442.262.842.621.442.858.633.520.837.281.466.250.321.529.358.247.043.3
UMT [6]39.659.132.435.045.161.948.47.546.067.621.429.548.275.970.556.725.928.939.443.644.1
SIGMA [29]40.155.437.431.154.954.346.623.044.765.623.022.042.855.667.255.232.940.845.058.644.5
ATMT [27]37.563.437.929.845.162.741.219.543.757.422.925.339.687.170.950.629.132.258.450.545.2
CIGAR [36]35.255.039.230.760.158.146.931.847.061.021.826.744.652.468.554.431.338.856.563.546.2
TIA [64]42.266.036.937.343.771.849.718.244.958.918.229.140.787.867.449.727.427.857.150.646.3
SIGMA++ [30]36.354.640.131.658.060.446.233.644.466.225.725.344.458.864.855.436.238.654.159.346.7
CMT [51]39.856.338.739.760.435.056.07.160.160.435.828.167.884.580.155.520.332.842.338.247.0
RegionCLIP [65]38.170.448.837.344.855.843.514.448.247.814.318.358.378.467.922.230.116.948.450.242.7
SEEN-DA(Ours)44.173.454.747.145.176.051.620.451.753.018.517.361.886.872.222.837.721.158.952.747.9
", + "bbox": [ + 81, + 419, + 890, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 626, + 200, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on mainstream DAOD scenarios, including the Cross-Weather, Cross-Fov, Sim-to-Real and Cross-Style. Further ablation studies are conducted to validate the effectiveness of the proposed domain-aware attention module.", + "bbox": [ + 75, + 652, + 470, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 739, + 181, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Weather Cityscapes [4] contains diverse street scenes in daylight, consisting of 2,975 training and 500 validation images annotated with eight classes. Foggy Cityscapes [48] simulates three distinct densities of fog on Cityscapes, containing 8,925 training images and 1,500 validation images. A standard configuration for cross-weather adaptation is to take the training set of Cityscapes as the source domain and the training set of foggy Cityscapes as the target domain, evaluating cross-weather adaptation per", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formance on the 1500-sized validation set in all eight categories.", + "bbox": [ + 496, + 627, + 890, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-FoV KITTI [14] is a crucial dataset for self-driving that includes 7,481 photos annotated with cars. Collected by driving in rural areas and on highways, it provides data with a different Field of View (FoV). To fairly compare with other methods, we migrate KITTI to Cityscapes solely on the car category.", + "bbox": [ + 496, + 657, + 890, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sim-to-Real The synthetic dataset SIM10k [21] has 10,000 photos from the GTA V video game with bounding boxes of the class car. We follow existing works to perform sim-to-real adaptation and report the performance on class car.", + "bbox": [ + 496, + 750, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Style Pascal VOC [9] dataset consists of 16551 images from the real world and has 20 categories. Clipart [20] contains 1000 comical images and has the same 20 categories as Pascal VOC. Following the mainstream splitting, we use Pascal VOC as the source domain and all Clipart images as the target domain.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25470", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/65e66d299679b8646f7b11d6a178d412a8d1b13fda48d7b5dc8f85563e735d78.jpg", + "table_caption": [ + "Table 3. Comparison $(\\%)$ of standard attention with domain-aware attention." + ], + "table_footnote": [], + "table_body": "
BenchmarkSelf-AttentionCross-AttentionDomain-aware Attention
Cross-Weather54.855.557.5
Cross-FoV61.963.567.1
Sim-to-Real62.363.766.8
Cross-Style43.145.247.9
", + "bbox": [ + 80, + 122, + 467, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e750792bdeff5dd26eb109a301bf0d795e79460883e6080fb95a025545a379cc.jpg", + "table_caption": [ + "Table 4. Ablation studies (%) of domain-aware attention on Cross-Weather adaptation." + ], + "table_footnote": [], + "table_body": "
Inter-domain AttentionLadvIntra-domain AttentionmAPGains
52.6-
✓54.9+2.3
✓✓55.8+0.9
✓✓✓57.5+1.7
", + "bbox": [ + 80, + 238, + 467, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 324, + 294, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [23], we adapt RegionCLIP(ResNet-50 [18]) with the Faster-RCNN [45] as the detector. In each iteration, one batch of source images with ground truth and one batch of target domain images with pseudo labels are forwarded to calculate the classification, adversarial and regression loss. The hyperparameter $\\lambda_{adv}$ , $\\lambda_t$ is set to 0.1 and 1.0. We set the batch size of each domain to 8 and use the SGD optimizer with a warm-up learning rate. Mean Average Precision (mAP) with a threshold of 0.5 is taken as the evaluation metric. All experiments are deployed on 8 Tesla V100 GPUs.", + "bbox": [ + 75, + 348, + 468, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison to SOTA methods", + "text_level": 1, + "bbox": [ + 76, + 525, + 349, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present representative state-of-the-art DAOD approaches for comparison, including feature alignment, semi-supervised learning and VLM-based domain alignment methods.", + "bbox": [ + 75, + 550, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Weather Adaptation Scenario As shown in Table 1 (C→F), the proposed SEEN-DA outperforms all compared methods in terms of mAP and advances SOTA by $1.6\\%$ , from $55.9\\%$ to $57.5\\%$ . Specifically, our method improves performance over six categories (i.e. person, rider, car, truck, train, and bicycle) ranging from $0.8\\%$ to $3.1\\%$ .", + "bbox": [ + 75, + 611, + 468, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-FOV Adaptation Scenario Table $1(\\mathrm{K}\\rightarrow \\mathrm{C})$ reports result for KITTI $\\rightarrow$ Cityscapes. SEEN-DA achieves SOTA performance of $67.1\\%$ mAP, gaining an improvement of $5.7\\%$ .", + "bbox": [ + 75, + 702, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sim-to-Real Adaptation Scenario Table 1 (S→C) shows that the proposed method achieves the best results of $66.8\\%$ mAP, outperforming the previous best entry SOCCER [5] $63.8\\%$ with $3.0\\%$ .", + "bbox": [ + 75, + 763, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Style Adaptation Scenario Additionally, we assess SEEN-DA on the more challenging Cross-Style adaptation, where the semantic hierarchy has broader discrepancies. In Table 2, SEEN-DA peaks with $47.9\\%$ mAP and improves six categories (aeroplane, bird, boat, bus, sheep and train).", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/233f24130efaf042b57e1f6fd7539ce97cc1ccdd32abd4c144d0c100dbeeef2a.jpg", + "table_caption": [ + "Table 5. Comparison $(\\%)$ of computational efficiency on Cross-Weather adaptation" + ], + "table_footnote": [], + "table_body": "
MethodBackbone Parm(M) LearnableParm(M) mAPAbs. Gains
DSS [56]26.83426.83440.9+4.2
CSDA [13]33.64533.64545.3+6.9
REACT [24]29.81229.81248.8+7.9
AT [31]39.22518.72350.9+7.9
DA-Pro [23]34.8340.00855.9+3.3
SEEN-DA (Ours)36.7011.87557.5+4.9
", + "bbox": [ + 504, + 121, + 888, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 222, + 663, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report detailed ablation studies (Table 3, 4, 5, Figure 4, 5) to validate the effectiveness of proposed SEEN.", + "bbox": [ + 496, + 244, + 890, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Standard Attention vs. Domain-aware Attention We compare the performance of domain-aware attention with existing attention methods, including self-attention and cross-attention, as shown in Table 3. With the guidance of textual information, the cross-attention improves the self-attention by $0.7\\% \\sim 2.1\\%$ . Furthermore, the proposed domain-aware attention gains improvements of $2.0\\% \\sim 33.6\\%$ over cross-attention, showing a superior ability to select features with high transferability.", + "bbox": [ + 496, + 277, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation for Domain-aware Attention Module We conduct ablation experiments to further analyze the significance of proposed domain-aware attention module in Table 4. Introducing the inter-domain attention branch attains an mAP of $54.9\\%$ , and optimizing with adversarial loss $\\mathcal{L}_{adv}$ gains $0.9\\%$ . Meanwhile, applying the intra-domain attention branch improves $1.7\\%$ on the target domain. These results showing the effectiveness of the domain-aware attention module to eliminate redundant information and supplement domain-specific semantic information.", + "bbox": [ + 496, + 414, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational Efficiency We discussed the computational efficiency of the proposed SEEN-DA in Table 5. Our method yields the same parameters scale as existing methods and achieves the highest mAP. In addition, SEEN-DA only needs to train very few parameters (1.875M), showing remarkable efficiency.", + "bbox": [ + 496, + 566, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Numbers of the Attention Module We explore the impact of numbers of domain-aware attention modules in Figure 4. Starting from the last visual block, gradually introducing domain-aware attention modules to each visual block improves performance and peaks when using 3 modules. Considering efficiency, we apply 3 domain-aware attention modules into the visual encoder.", + "bbox": [ + 496, + 657, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Projection layer in SEEN module We explore the projection layer in the SEEN module in Figure 5. Projecting textual embedding into visual space, denoted as T2V, only achieves $55.1\\%$ . Projecting textual and visual embedding into a latent space, denoted as T&V, improves performance by $2.1\\%$ . And projecting visual embedding into textual space achieves the best mAP of $57.5\\%$ . We analyze the reason for these results as follows. Visual embedding contains more sample noise and redundant information than textual", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc7431c25c8a57c1d07ad9d0f616a3cc011ca3de1ff79b778e19ea897bdcd09c.jpg", + "image_caption": [ + "(a) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 354, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1833aa968d6a00cad5eb5261add9b84402964fb83850fe591836b5a4946de3e6.jpg", + "image_caption": [ + "(b) DA-Pro[23]" + ], + "image_footnote": [], + "bbox": [ + 354, + 85, + 609, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/506621036038f9f185e7139f7d1642ddf01c16005aa4194f907503865a83c744.jpg", + "image_caption": [ + "(c) SEEN-DA (Ours)" + ], + "image_footnote": [], + "bbox": [ + 609, + 85, + 870, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b3c8b2a75c6de307c143227ffe30e431bf2ba93b8eaffb23228d66f5128065b.jpg", + "image_caption": [ + "Figure 4. Ablation $(\\%)$ on the number of attention modules on Cross-Weather adaptation." + ], + "image_footnote": [], + "bbox": [ + 117, + 345, + 413, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "embedding. And learning to project visual embedding to textual space can reduce the interference of these factors. Therefore, we adopt the V2T projection in SEEN module.", + "bbox": [ + 75, + 547, + 468, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 595, + 215, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The detection results of ground truth (a), DA-Pro [23] (b) and the proposed SEEN-DA (c) on the target domain of Cross-Weather adaptation are displayed in Figure 3. The figure shows that SEEN-DA outperforms the SOTA method in terms of the ability to detect objects. Specifically, SEEN-DA detects the objects missed (e.g. the rider on the right of figures in 1st row and the person on the left of figures in 2nd row) and misclassified (e.g. the bicycle on the middle of figures in 1st row and the person on the left of figures in 2nd row), revealing the effectiveness of domain-aware attention. By utilizing semantic entropy to reduce redundant information and supplement domain-specific semantic information, the SEEN-DA improves the discriminability of the detector on the target domain.", + "bbox": [ + 75, + 619, + 468, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel SEMantic ENtropy guided Domain-aware Attention (SEEN-DA) for DAOD.", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f500264c46a79bd25ef6898c1d2cfd1ca571d241efcfbded1da13f5c4fff49ea.jpg", + "image_caption": [ + "Figure 3. Detection comparison on the Cross-Weather adaptation scenario. We visualize (a) the ground truth, the detection boxes of (b) SOTA DA-Pro [23] and (c) our methods SEEN-DA.", + "Figure 5. Ablation $(\\%)$ on projection layer of the domain-aware attention module on Cross-Weather adaptation." + ], + "image_footnote": [], + "bbox": [ + 531, + 344, + 830, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We design the semantic entropy to quantify the semantic information contained in visual features. Guided by semantic entropy, the domain-aware attention module is introduced to adaptively refine visual features with the semantic commonalities and disparities between domains. Precisely, it consists of an inter-domain attention branch for extracting domain-invariant features and eliminating redundant information between domains, and an intra-domain attention branch for supplementing the domain-specific semantic information for each domain. Extensive experiments over multiple benchmarks validate the effectiveness of SEENDA in inferring a discriminative detector.", + "bbox": [ + 496, + 550, + 890, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This work is partially supported by the National Key R&D Program of China(under Grant 2023YFB4502200), the NSF of China(under Grants 92364202, U22A2028, 62102398, 62102399, 62302478, 62302482, 62376268), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660300, XDB0660301, XDB0660302), Major Program of ISCAS(Grant No. ISCAS-ZD-202402), Beijing Natural Science Foundation(4222039), CAS Project for Young Scientists in Basic Research(YSBR-029) and Youth Innovation Promotion Association CAS.", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25472", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaoqi Chen, Jiongcheng Li, Hong-Yu Zhou, Xiaoguang Han, Yue Huang, Xinghao Ding, and Yizhou Yu. Relation matters: foreground-aware graph-based relational reasoning for domain adaptive object detection. TPAMI, 45(03):3677-3694, 2023. 6", + "[2] Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xin chao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, et al. Learning domain adaptive object detection with probabilistic teacher. In ICML, pages 3040-3055, 2022. 3", + "[3] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, pages 3339-3348, 2018. 1, 3, 6", + "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6", + "[5] Yiming Cui, Liang Li, Jiehua Zhang, Chenggang Yan, Hongkui Wang, Shuai Wang, Heng Jin, and Li Wu. Stochastic context consistency reasoning for domain adaptive object detection. In ACMMM, pages 1331-1340, 2024. 6, 7", + "[6] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In CVPR, pages 4091-4101, 2021. 3, 6", + "[7] Jinhong Deng, Dongli Xu, Wen Li, and Lixin Duan. Harmonious teacher for cross-domain object detection. In CVPR, pages 23829-23838, 2023. 3", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[9] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6", + "[10] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul De Charette. Poda: Prompt-driven zero-shot domain adaptation. In ICCV, pages 18623-18633, 2023. 3", + "[11] Yongchao Feng, Shiwei Li, Yingjie Gao, Ziyue Huang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Dsd-da: Distillation-based source debiasing for domain adaptive object detection. In ICML, 2024. 6", + "[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, 17(1):2096-2030, 2016. 3", + "[13] Changlong Gao, Chengxu Liu, Yujie Dun, and Xueming Qian. Csda: Learning category-scale joint feature for domain adaptive object detection. In ICCV, pages 11421-11430, 2023. 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6", + "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3", + "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3", + "[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu, and Yanpeng Cao. Uncertainty-aware unsupervised domain adaptation in object detection. TMM, 24:2502-2514, 2021. 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7", + "[19] Mengzhe He, Yali Wang, Jiaxi Wu, Yiru Wang, Hanqing Li, Bo Li, Weihao Gan, Wei Wu, and Yu Qiao. Cross domain object detection by target-perceived dual branch distillation. In CVPR, pages 9570-9580, 2022. 3", + "[20] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, pages 5001-5009, 2018. 6", + "[21] Matthew Johnson-Roberson, Charles Barto, Rounak Mehta, Sharath Nittur Sridhar, Karl Rosaen, and Ram Vasudevan. Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks? In ICRA 2017, pages 746-753. IEEE, 2017. 6", + "[22] Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T Tan. Cat: Exploiting inter-class dynamics for domain adaptive object detection. In CVPR, pages 16541-16550, 2024. 6", + "[23] Haochen Li, Rui Zhang, Hantao Yao, Xinkai Song, Yifan Hao, Yongwei Zhao, Ling Li, and Yunji Chen. Learning domain-aware detection head with prompt tuning. NeurIPS, 36, 2024. 1, 2, 3, 5, 6, 7, 8", + "[24] Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, and Ling Li. React: Remainder adaptive compensation for domain adaptive object detection. TIP, 33:3735-3748, 2024. 1, 3, 6, 7", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3", + "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 3", + "[27] Kai Li, Curtis Wigington, Chris Tensmeyer, Vlad I. Morariu, Handong Zhao, Varun Manjunatha, Nikolaos Barmpalios, and Yun Fu. Improving cross-domain detection with self-supervised learning. In CVPR, pages 4745-4754, 2023. 6", + "[28] Wuyang Li, Xinyu Liu, Xiwen Yao, and Yixuan Yuan. Scan: Cross domain object detection with semantic conditioned adaptation. In AAAI, volume 6, page 7, 2022. 6", + "[29] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object detection. In CVPR, pages 5291-5300, 2022. 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI, 45(07):9022-9040, 2023. 1, 3, 6", + "[31] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Va-jda. Cross-domain adaptive teacher for object detection. In CVPR, pages 7581-7590, 2022. 3, 6, 7", + "[32] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1, 3", + "[33] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1", + "[34] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. TMM, 25:1333-1344, 2022. 3", + "[35] Rui Liu, Yahong Han, Yaowei Wang, and Qi Tian. Frequency spectrum augmentation consistency for domain adaptive object detection. arXiv preprint arXiv:2112.08605, 2021. 3", + "[36] Yabo Liu, Jinghua Wang, Chao Huang, Yaowei Wang, and Yong Xu. Cigar: Cross-modality graph reasoning for domain adaptive object detection. In CVPR, pages 23776-23786, 2023. 6", + "[37] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 3", + "[38] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. NeurIPS, 31, 2018. 3", + "[39] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, pages 2208-2217, 2017. 3", + "[40] Fabio Maria Carlucci, Lorenzo Porzi, Barbara Caputo, Elisa Ricci, and Samuel Rota Bulo. Autodial: Automatic domain alignment layers. In ICCV, pages 5067-5075, 2017. 3", + "[41] Pietro Morerio, Jacopo Cavazza, and Vittorio Murino. Minimal-entropy correlation alignment for unsupervised deep domain adaptation. In ICLR, 2018. 3", + "[42] Zhengquan Piao, Linbo Tang, and Baojun Zhao. Unsupervised domain-adaptive object detection via localization regression alignment. TNNLS, 35(11):15170-15181, 2024. 6", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3", + "[44] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1", + "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 3, 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In ICCV, pages 8050-8058, 2019. 3", + "[47] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In CVPR, pages 6956–6965, 2019. 1, 3", + "[48] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. IJCV, 126(9):973-992, Sep 2018. 6", + "[49] Sunandini Sanyal, Ashish Ramayee Asokan, Suvaansh Bhambri, Akshay Kulkarni, Jogendra Nath Kundu, and R Venkatesh Babu. Domain-specificity inducing transformers for source-free domain adaptation. In ICCV, pages 18928-18937, 2023. 1, 3", + "[50] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NPIS, 30, 2017. 3", + "[51] VS Vibashan, Poojan Oza, and Vishal M Patel. Instance relation graph guided source-free domain adaptive object detection. In CVPR, pages 3520-3530, 2023. 6", + "[52] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 3", + "[53] Vibashan Vs, Vikram Gupta, Poojan Oza, Vishwanath A Sindagi, and Vishal M Patel. Mega-cda: Memory guided attention for category-aware unsupervised domain adaptive object detection. In CVPR, pages 4516-4526, 2021. 3", + "[54] Haoan Wang, Shilong Jia, Tieyong Zeng, Guixu Zhang, and Zhi Li. Triple feature disentanglement for one-stage adaptive object detection. In AAAI, pages 5401-5409, 2024. 3, 6", + "[55] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, volume 33, pages 5345-5352, 2019. 3", + "[56] Yu Wang, Rui Zhang, Shuo Zhang, Miao Li, Yangyang Xia, XiShan Zhang, and ShaoLi Liu. Domain-specific suppression for adaptive object detection. In CVPR, pages 9603-9612, 2021. 7", + "[57] Weixi Weng and Chun Yuan. Mean teacher detr with masked feature alignment: A robust domain adaptive detection transformer framework. In AAAI, pages 5912-5920, 2024. 6", + "[58] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. TPAMI, 44(8):4178-4193, 2021. 1, 3", + "[59] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, pages 9342-9351, 2021. 6", + "[60] Xiaofu Wu, Suofei Zhang, Quan Zhou, Zhen Yang, Chunming Zhao, and Longin Jan Latecki. Entropy minimization versus diversity maximization for domain adaptation. TNNLS, 34(6):2896-2907, 2021. 3", + "[61] Jayeon Yoo, Inseop Chung, and Nojun Kwak. Unsupervised domain adaptation for one-stage object detector using offsets to bounding box. In ECCV, pages 691-708, 2022. 6", + "[62] Yifan Zhang, Hanbo Chen, Ying Wei, Peilin Zhao, Jiezhang Cao, Xinjuan Fan, Xiaoying Lou, Hailing Liu, Jinlong Hou," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25474", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Xiao Han, et al. From whole slide imaging to microscopy: Deep microscopy adaptation network for histopathology cancer image classification. In MICCAI, pages 360-368, 2019. 3", + "[63] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, pages 12425-12434, 2021. 3", + "[64] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In CVPR, pages 14217-14226, 2022. 6", + "[65] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 1, 3, 6", + "[66] Wenzhang Zhou, Heng Fan, Tiejian Luo, and Libo Zhang. Unsupervised domain adaptive detection with network stability analysis. In ICCV, pages 6986-6995, 2023. 6" + ], + "bbox": [ + 78, + 90, + 467, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_model.json b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fae16220edc882fea2330d2c687f98d2025ad71a --- /dev/null +++ b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.178 + ], + "angle": 0, + "content": "SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.203, + 0.83, + 0.222 + ], + "angle": 0, + "content": "Haochen Li\\(^{1,4}\\) Rui Zhang\\(^{2*}\\) Hantao Yao\\(^{3}\\) Xin Zhang\\(^{2}\\) Yifan Hao\\(^{2}\\) Xinkai Song\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.222, + 0.788, + 0.239 + ], + "angle": 0, + "content": "Shaohui Peng\\(^{1}\\) Yongwei Zhao\\(^{2}\\) Chen Zhao\\(^{5}\\) Yanjun Wu\\(^{5}\\) Ling Li\\(^{1,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.239, + 0.744, + 0.256 + ], + "angle": 0, + "content": "1Intelligent Software Research Center, Institute of Software, CAS" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.257, + 0.767, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)State Key Lab of Processors, Institute of Computing Technology, CAS" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.274, + 0.868, + 0.292 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Information Science and Technology, University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.293, + 0.788, + 0.309 + ], + "angle": 0, + "content": "\\(^{4}\\) University of Chinese Academy of Sciences \\(^{5}\\) Institute of Software, CAS" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.313, + 0.776, + 0.327 + ], + "angle": 0, + "content": "{haochen2021, pengshaohui, zhaochen, wuyanjun, liling}@iscas.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.331, + 0.872, + 0.345 + ], + "angle": 0, + "content": "{zhangrui, zhangxin, haoyifan, songxinkai, zhaoyongwei}@ict.ac.cn, yaohantao@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.36, + 0.314, + 0.377 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.392, + 0.473, + 0.801 + ], + "angle": 0, + "content": "Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain to an unlabelled target domain. Traditional works focus on aligning visual features between domains to extract domain-invariant knowledge, and recent VLM-based DAOD methods leverage semantic information provided by the textual encoder to supplement domain-specific features for each domain. However, they overlook the role of semantic information in guiding the learning of visual features that are beneficial for adaptation. To solve the problem, we propose semantic entropy to quantify the semantic information contained in visual features, and design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to adaptively refine visual features with the semantic information of two domains. Semantic entropy reflects the importance of features based on semantic information, which can serve as attention to select discriminative visual features and suppress semantically irrelevant redundant information. Guided by semantic entropy, we introduce domain-aware attention modules into the visual encoder in SEEN-DA. It utilizes an inter-domain attention branch to extract domain-invariant features and eliminate redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. Comprehensive experiments validate the effectiveness of SEEN-DA, demonstrating significant improvements in cross-domain object detection performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.82, + 0.21, + 0.836 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.846, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Object detection [8, 33, 44, 45] have been proven to attain encouraging performance assisted by high-quality im" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.361, + 0.892, + 0.452 + ], + "angle": 0, + "content": "ages with accurate annotations. However, due to the noticeable discrepancy between domains, they suffer severe performance degradation when dealing with unseen data in distinct scenes. To alleviate this problem, domain adaptive object detection (DAOD) [3] aims to transfer a detector trained on the labelled source domain to a unlabelled target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.895, + 0.789 + ], + "angle": 0, + "content": "Traditional DAOD works [3, 30, 47] focus on reducing domain discrepancy by aligning visual features between domains to extract domain-invariant knowledge. Additionally, [24,32,49,58] decouple the unique features of each domain, i.e.domain-specific features, to prevent its interference during feature alignment. Although effective, they train the detector with domain-shared and semantically agnostic category labels (such as one-hot encoding), which ignores the linguistic properties of categories and inter-domain differences and commonality, i.esemantic information. Recently, visual-language models (VLMs) [26,43,65] are proposed to ground visual with textual modalities. Through the textual encoder, they align visual features with the textual prompt, which is constructed from natural language and contains rich semantic information. Motivated by this, recent VLM-based DAOD methods, such as DA-Pro [23], utilize the textual encoder as detection head and tune the textual prompt to learn domain-invariant semantics while supplementing domain-specific semantics for each domain. However, they freeze the visual encoder and leverage semantic information solely in the detection head, limiting the discriminability of visual features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Summarily, despite promising performance, existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. They typically optimize the visual encoder with semantically agnostic category labels and adversarial loss, or directly freeze it, resulting in two limitations. First, they fail to filter out redundant information from visual fea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25465" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.341, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.147, + 0.294, + 0.156 + ], + "angle": 0, + "content": "(a.1) Source Image and Bounding Box" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.157, + 0.341, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.213, + 0.335, + 0.223 + ], + "angle": 0, + "content": "(b.1) Detection Results on Source Image (DA-Pro[23])" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.224, + 0.34, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.281, + 0.349, + 0.291 + ], + "angle": 0, + "content": "(c.1) Detection Results on Source Image (SEEN-DA (Ours))" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.089, + 0.484, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.147, + 0.57, + 0.157 + ], + "angle": 0, + "content": "(a.2) Target Image and Bounding Box" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.157, + 0.485, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.213, + 0.608, + 0.223 + ], + "angle": 0, + "content": "(b.2) Detection Results on Target Image (DA-Pro[23])" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.224, + 0.485, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.281, + 0.622, + 0.291 + ], + "angle": 0, + "content": "(c.2) Detection Results on Target Image (SEEN-DA (Ours))" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.615, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.147, + 0.615, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.213, + 0.615, + 0.223 + ], + "angle": 0, + "content": "Target Image (DA-Pro[23])" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.224, + 0.615, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.487, + 0.281, + 0.622, + 0.291 + ], + "angle": 0, + "content": "Target Image (SEEN-DA (Ours))" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.09, + 0.891, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.281, + 0.88, + 0.291 + ], + "angle": 0, + "content": "(d) Domain-aware Attention Map in SEEN-DA (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.3, + 0.894, + 0.384 + ], + "angle": 0, + "content": "Figure 1. Visualization of the detection results and attention map. (a) Source and Target domain image and bounding boxes. (b) Existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. Infected by the redundant information and lack of domain-specific semantic information, some objects that are correctly detected on the source image are misclassified or missed on the target image. (c) We propose semantic entropy to evaluate the importance of visual features based on domain and category semantic information. By integrating it as domain-aware attention in SEEN-DA, we eliminate redundant information and enhance domain-specific features, thereby improving the detector's discriminability for DAOD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.473, + 0.68 + ], + "angle": 0, + "content": "tures, which is semantically irrelevant and can lead to incorrect feature alignment. Redundant information includes attributes unrelated to object identification, such as color, texture, etc. Although they do not contain domain and category semantic information, they are often extracted by adversarially trained visual encoder since they exist in both domains. Therefore, such semantically irrelevant information may interfere with the alignment of domain-invariant features. In Fig. 1(b.1), existing methods detect the bicycle covered with a white cloth on the source image (clear day). However, since target images (foggy day) tend to be white overall, they misclassify the white cloth as a rider in Fig. 1(b.2). Second, existing methods discard domain-specific features in the visual encoder, which contain rich domain semantic information and are important for the discriminability on the target domain. For example, they correctly detect the bicycle and rider on the source image Fig. 1(b.1), but miss them in Fig. 1(b.2) due to insufficient feature learning of domain-specific properties of target domain, e.g.fog." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To solve these problems, we propose semantic entropy which evaluates the importance of features based on domain and category semantic information, and accordingly design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to eliminate redundant information and supplement domain-specific features for DAOD. The semantic entropy of visual features refers to the amount of semantic information related to a certain domain. As the textual encoder of VLM provides rich domain and category semantic information in textual embedding, the semantic entropy can be expressed as the information entropy of the classification probability between the visual and textual embedding. Therefore, semantic entropy reflects the importance of visual features in adaptation. Low semantic entropy indicates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.893, + 0.466 + ], + "angle": 0, + "content": "that visual features match the textual embedding of a certain category and are likely to be transferable foreground objects. On the contrary, high semantic entropy means that they are unpredictable and may be redundant or even harmful for detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.893, + 0.813 + ], + "angle": 0, + "content": "Applying semantic entropy as attention, we design the SEEN-DA framework, introducing domain-aware attention modules into the visual encoder to select visual features with low semantic entropy and suppress regions with high ones. Specifically, the domain-aware attention module utilizes two parallel branches to learn inter- and intra-domain attention with semantic entropy. The inter-domain attention branch is shared between domains and evaluates the semantic entropy on domain-shared textual embedding, aiming to learn domain-invariant features while eliminating redundant information. The intra-domain attention branch is private in each domain to evaluate the semantic entropy on source/target textual embedding respectively, supplementing domain-specific features for each domain. The SEEN-DA is able to adaptively adjust visual features with the semantic commonalities and disparities between domains. As shown in Fig. 1(c), domain-aware attention map highlights features with rich semantic information, i.eforeground regions, and hinders the negative impact of background regions. Thereby, the proposed SEEN-DA correctly detects objects misclassified by existing methods, improving the discriminability of the detector on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed method on four benchmarks. The experimental results highlight the effectiveness of our SEEN-DA, e.g., it reaches \\(57.5\\%\\) mAP on Cityscapes \\(\\rightarrow\\) Foggy Cityscapes, surpassing the state-of-the-art DA-Pro [23] by \\(1.6\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25466" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.47, + 0.327 + ], + "angle": 0, + "content": "Visual-Language Models Large-scale visual-language models (VLMs) [25, 26, 43] embed visual and text modalities into a shared space, enabling cross-modal alignment. Pre-trained with an astonishing scale of image-text pairs, they demonstrate comprehensive visual understanding. As a pioneering work, CLIP [43] trains a visual encoder and a textual encoder with 400 million image-text pairs in a contrastive manner, showing promising classification ability for the seen and unseen classes. Following this spirit, [10, 16, 52, 65] freeze the textual encoder and adapt the visual encoder into Faster-RCNN [45] as a two-stage detector. In this work, we apply the textual encoder of CLIP [43] to infer textual embedding that contains semantic information, and the visual encoder of RegionCLIP [65] as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.327, + 0.473, + 0.856 + ], + "angle": 0, + "content": "Domain Adaptive Object Detection (DAOD) aims to adapt the object detector [45] trained on the labelled source domain to the unlabelled target domain. Previous approaches can be broadly divided into three categories: feature alignment, semi-supervised learning and VLM-based domain alignment. As a fundamental direction, feature alignment [12, 37, 39] aims to align the feature distributions of the two domains to generate domain-invariant features [3, 30, 47]. Typically, the domain discriminator is proposed to align feature distribution adversarially in image level [3], instance level [47, 63] and category level [29, 53]. To reduce the negative impact of aligning unique features of each domain, [24, 32, 34, 49, 54, 58] decouple the domain-invariant and domain-specific knowledge between two domains. In parallel, semi-supervised learning strives to generate unbiased training data with style transfer [6, 19, 35] and pseudo label [2, 7, 31, 50]. However, they train the detector with semantically agnostic category labels (such as one-hot encoding), ignoring the semantic information inherently contained in the category name. To overcome this, recent VLM-based domain alignment methods exploit the semantic information provided by the textual encoder of VLMs to learn semantic commonalities and disparities. Pioneeringly, DA-Pro [23] tuning the textual encoder into a domain-aware detection head, learning domain-invariant semantics while supplementing domain-specific semantics for each domain. Despite the promising performance, existing methods freeze the visual encoder and ignore using semantic information to guide the learning of visual features, which leads to two limitations. First, they fail to filter out redundant information from visual features, which is semantically irrelevant and can lead to incorrect feature alignment. Second, they discard domain-specific features in the visual encoder, which contain rich domain semantic information and are vital for DAOD." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Entropy Optimization Information entropy was first utilized as entropy minimization in [15] for semi-supervised learning to learn decision rules for unlabelled data. Con-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.394 + ], + "angle": 0, + "content": "sidering its effectiveness, recent works introduce entropy optimization into DAOD. [40, 41, 46, 62] propose minimizing entropy on unlabelled data to make decisions consistent in the source and target domains. [55] apply entropy as metrics to measure the transferability of data samples. [38] employs entropy as attention to control the uncertainty of the classifier predictions, alleviating the effect of negative transfer. Instead of directly minimizing entropy, [60] further introduces diversity maximization to explore close-to-ideal domain adaptation solutions. However, the computation of entropy in existing methods is inherently semantically agnostic. Based on the output probabilities from a classifier, they treat categories merely as identifiers and convert the category labels into one-hot encoding. Therefore, they ignore the rich semantic information embedded within the category names, which is vital for DAOD. In this paper, we propose a novel semantic entropy, evaluating the transferability of visual features through the semantic commonalities and disparities between domains, improving the discrimination of the visual encoder." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.406, + 0.636, + 0.424 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.895, + 0.703 + ], + "angle": 0, + "content": "In this section, we present a novel SEmantic entropy guided Domain-aware Attention (SEEN-DA) for DAOD, which introduces domain-aware attention module to eliminate redundant information and supplement domain-specific features, guided by the proposed semantic entropy. Semantic entropy quantifies the amount of semantic information within visual features, providing a measure of their importance in domain adaptation. Utilizing semantic entropy as attention, the domain-aware attention module adopts an inter-domain attention branch to extract domain-invariant features and eliminate semantically irrelevant redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. As shown in Fig. 2(a), given input image \\( x \\), the visual encoder of SEEN-DA is divided into \\( N \\) frozen visual block \\( \\{\\mathcal{V}_i\\}_{i=1}^N \\), and each domain-aware attention module \\( \\{\\mathcal{A}_i\\}_{i=1}^N \\) is attached to \\( \\mathcal{V}_i \\) as a lightweight learnable module:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.707, + 0.892, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {0} = \\mathcal {E} (x); \\mathbf {h} _ {i} = \\mathcal {A} _ {i} \\left(\\mathcal {V} _ {i} \\left(\\mathbf {h} _ {i - 1}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.726, + 0.673, + 0.741 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is the stem layer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.75, + 0.74, + 0.766 + ], + "angle": 0, + "content": "3.1. Semantic Entropy (SEEN)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.773, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Generally speaking, given feature \\( \\mathbf{f} \\), the probability that this feature belongs to a specific object category is:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.805, + 0.892, + 0.839 + ], + "angle": 0, + "content": "\\[\np (c, \\mathbf {f}) = \\frac {\\exp \\left(v _ {c} (\\mathbf {f})\\right)}{\\sum_ {c} \\exp \\left(v _ {c} (\\mathbf {f})\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "where \\(c, v_{c}(\\cdot)\\) represents an object category and its corresponding classifier. \\(p(\\cdot, \\mathbf{f})\\) is the probability distribution function, expected to output 1 on the corresponding category and 0 on other category, i.e. one-hot vector." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25467" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.52, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.242, + 0.368, + 0.253 + ], + "angle": 0, + "content": "(a) Overview of SEEN-DA" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.088, + 0.88, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.242, + 0.812, + 0.253 + ], + "angle": 0, + "content": "(b) Domain-aware Attention Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Figure 2. (a) Overview of the proposed SEEN-DA for DAOD, where the semantic entropy is utilized as attention in domain-aware attention module. (b) The architecture of domain-aware attention module, consisting of an inter-domain and an intra-domain attention branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.469, + 0.437 + ], + "angle": 0, + "content": "Tradition DAOD approaches focus on reducing domain discrepancy by aligning \\( \\mathbf{f} \\) between the source and target domain. Although effective, as shown in Eq. 2, they treat categories merely as one-hot vectors, which is inherently semantically agnostic. Therefore, they ignore the rich semantic information inherently embedded within the category names. As feature alignment relies heavily on a profound understanding of the semantic disparities between the source and target domains, this limitation hinders the learning of visual features for adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.469, + 0.679 + ], + "angle": 0, + "content": "To address this, we propose semantic entropy, which quantifies the semantic information embedded in visual features by leveraging textual embeddings from Visual-language models (VLMs). VLMs utilize a textual encoder to encode textual prompts that describe category attributes into textual embeddings, aligning them with visual features. As a result, textual embeddings inherit rich semantic information from natural language, making them effective representations of the semantics of visual features. Motivated by this, we first infer textual embedding \\(\\mathbb{T} = \\{t_c\\}_c^K\\) for \\(K\\) categories by feeding prompts like \"A photo of [Class] in [Domain]\" into a textual encoder, where each \\(t_c\\) contains semantic information of category \\(c\\) and its domain. Thus, the probability distribution that \\(\\mathbf{f}\\) corresponds to certain semantics can be estimated by the similarity between the visual feature \\(\\mathbf{f}\\) and textual embedding \\(\\mathbb{T}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.681, + 0.469, + 0.713 + ], + "angle": 0, + "content": "\\[\np \\left(t _ {c}, \\mathbf {f}\\right) = \\frac {\\exp \\left(\\sin \\left(t _ {c} , \\mathbf {f}\\right) / \\tau\\right)}{\\sum_ {c} \\exp \\left(\\sin \\left(t _ {c} , \\mathbf {f}\\right) / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.715, + 0.469, + 0.865 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot ,\\cdot)\\) is the cosine similarity, and \\(\\tau\\) is a learnable temperature parameter. Though \\(p(t_{c},\\mathbf{f})\\) provides a measure of semantic relevance, it does not directly quantify the amount of semantic information embedded in \\(\\mathbf{f}\\). Therefore, we draw inspiration from the concept of information entropy, which is widely used to measure the uncertainty and informativeness of a probability distribution. Formally, the proposed semantic entropy is defined as the information entropy of the probability distribution \\(p(t_{c},\\mathbf{f})\\) to evaluate the semantic uncertainty of \\(\\mathbf{f}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.873, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S E} (\\mathbb {T}, \\mathbf {f}) = - \\sum_ {c} p \\left(t _ {c}, \\mathbf {f}\\right) \\log \\left(p \\left(t _ {c}, \\mathbf {f}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.438 + ], + "angle": 0, + "content": "In the framework of DAOD, we apply semantic entropy as attention to select high discriminative visual features. According to Eq. 4, features with low semantic entropy exhibit clear and definite semantic information, expected to be foreground features suitable for transfer. Conversely, features with high semantic entropy indicate high semantic uncertainty, suggesting they may contain semantically irrelevant redundant information, which should be deemphasized during adaptation. To achieve this, we regulate Eq. 4 to propose semantic entropy-guided attention:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.445, + 0.892, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S E A t t e n t i o n} (\\mathbb {T}, \\mathbf {f}) = \\sum_ {c} p \\left(t _ {c}, \\mathbf {f}\\right) \\log \\left(p \\left(t _ {c}, \\mathbf {f}\\right)\\right) + \\log K, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.549 + ], + "angle": 0, + "content": "where \\(K\\) is the number of categories. Based on semantic entropy, Eq. 5 effectively adjusts the weight of visual features, focusing on features with rich semantic information while minimizing the influence of redundant information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.791, + 0.57 + ], + "angle": 0, + "content": "3.2. Domain-aware Attention Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.891, + 0.7 + ], + "angle": 0, + "content": "The domain-aware attention module is designed to adaptively refine visual features of each domain by utilizing the semantic entropy. As shown in Fig. 2(b), the module incorporates two complementary branches: an inter-domain attention branch to learn domain-invariant features and eliminate semantically irrelevant information, and an intra-domain attention branch to supplement the domain-specific features unique to each domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Inter-domain Attention The inter-domain attention is designed to modify domain-invariant features of the frozen visual encoder with semantic information. Specifically, for the output visual features \\(\\mathbf{f}_d = \\mathcal{V}_i(\\mathbf{h}_{i - 1})\\) of the i-th visual block, we first forward it to a domain-shared convolution module \\(\\mathcal{C}\\) to capture task-related knowledge:" + }, + { + "type": "equation", + "bbox": [ + 0.658, + 0.803, + 0.891, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} = \\mathcal {C} \\left(\\mathbf {f} _ {d}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.891, + 0.855 + ], + "angle": 0, + "content": "where \\(d \\in \\{\\mathrm{s},\\mathrm{t}\\}\\) denotes whether \\(\\mathbf{f}_d\\) belongs to source domain s or target domain t." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Simultaneously, the visual feature \\(\\mathbf{f}_d\\) is forwarded to the SEEN module \\(S\\) to obtain attention map \\(\\mathbf{w}\\) guided by semantic entropy. With the help of VLM, we can represent the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "domain-invariant semantic information through domain-shared prompts. Technically, we input textual prompts such as \"A photo of [Class]\" to the textual encoder to retrieve domain-invariant textual embedding \\(\\mathbb{T}\\). As described in Eq. 4, the similarity of visual features and textual embeddings is needed to calculate semantic entropy attention. To facilitate this, a projection layer \\(\\mathcal{P}\\) is firstly introduced to map \\(\\mathbf{f}_d\\) from the visual feature space to the textual embedding space:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.229, + 0.47, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathrm {p}} = \\mathcal {P} \\left(\\mathbf {f} _ {d}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.253, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Then we calculate the semantic entropy-guided attention of the visual feature \\(\\mathbf{f}_d\\) over text embedding \\(\\mathbb{T}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.297, + 0.47, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = \\mathcal {S} (\\mathbb {T}, \\mathbf {f} _ {d}) = \\operatorname {S E A t t e n t i o n} (\\mathbb {T}, \\mathbf {f} _ {\\mathrm {p}}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.47, + 0.402 + ], + "angle": 0, + "content": "With the attention map \\(\\mathbf{w}\\), we can focus on semantically informative features and suppress redundant information in domain-invariant features. Typically, \\(\\mathbf{w}\\) is applied to \\(\\mathbf{f}\\) via an element-wise product, and the weighted feature is summed with the original visual feature:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.415, + 0.469, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {\\prime} = \\mathbf {w} \\cdot \\mathbf {f} + \\mathbf {f} _ {d}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.444, + 0.469, + 0.49 + ], + "angle": 0, + "content": "Expecting to be domain-invariant, the extracted \\(\\mathbf{f}'\\) is forwarded to a domain discriminator \\(\\mathcal{D}\\), which aligns the source and target domain adversarily." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.661 + ], + "angle": 0, + "content": "Intra-domain Attention With the inter-domain attention branch, the extracted domain-invariant features contain sufficient semantic information shared between domains. However, it ignores the semantic discrepancy between the two domains, resulting in limited discriminability in the target domain. Therefore, we design the intra-domain attention branch to enrich visual features with domain-specific semantic information. Similar to the inter-domain attention branch, we first introduce a learnable layer \\(\\mathcal{C}^c\\) to adapt visual features for to both domains:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.674, + 0.469, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {d} ^ {\\mathrm {c}} = \\mathcal {C} ^ {\\mathrm {c}} \\left(\\mathbf {f} _ {d}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Subsequently, we generate domain-aware textual embedding that captures the semantics of each domains. Motivated by DA-Pro [23], we tune learnable domain-adaptive prompts, i.e. \\(\\left[[v^{\\mathrm{c}}][v^{\\mathrm{s}}]\\right]\\) [Class] for the source domain and \\(\\left[[v^{\\mathrm{c}}][v^{\\mathrm{t}}]\\right]\\) [Class] for the target domain, to explicitly model the domain and category semantic commonalities and disparities. These prompts are forwarded to the textual encoder to derive domain-adaptive textual embedding \\(\\mathbb{T}_{\\mathrm{s}},\\mathbb{T}_{\\mathrm{t}}\\) for each domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Due to the visual and semantic differences between the source and target domain, a domain-shared SEEN module is inadequate. To calculate the semantic entropy for each domain, visual features must be independently mapped into" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "their respective textual spaces. Therefore, we apply separate SEEN modules \\(S^{\\mathrm{s}}\\) and \\(S^{\\mathrm{t}}\\) with independent projection layers \\(\\mathcal{P}^{\\mathrm{s}},\\mathcal{P}^{\\mathrm{t}}\\) and textual embedding \\(\\mathbb{T}_{\\mathrm{s}},\\mathbb{T}_{\\mathrm{t}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.144, + 0.892, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {\\mathrm {s}} = \\mathcal {S} ^ {\\mathrm {s}} \\left(\\mathbb {T} _ {\\mathrm {s}}, \\mathbf {f} _ {\\mathrm {s}}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.178, + 0.892, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {\\mathrm {t}} = \\mathcal {S} ^ {\\mathrm {t}} \\left(\\mathbb {T} _ {\\mathrm {t}}, \\mathbf {f} _ {\\mathrm {t}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.228 + ], + "angle": 0, + "content": "where the \\(\\mathbf{w}_{\\mathrm{s}}\\) and \\(\\mathbf{w}_{\\mathrm{t}}\\) denotes the source- and target-specific attention maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.318 + ], + "angle": 0, + "content": "The inter-domain attention branch extracts domain-invariant features while eliminating redundant information, whereas the intra-domain attention branch enriches visual features with domain-specific semantics. Overall, the output for the i-th domain-aware attention module \\(\\mathcal{A}_i\\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.326, + 0.892, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathrm {s}} ^ {\\prime} = \\mathcal {A} _ {i} \\left(\\mathbf {f} _ {\\mathrm {s}}\\right) = \\mathbf {f} ^ {\\prime} + \\mathbf {w} _ {\\mathrm {s}} \\cdot \\mathbf {f} _ {\\mathrm {s}} ^ {\\mathrm {c}}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.36, + 0.892, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\mathrm {t}} ^ {\\prime} = \\mathcal {A} _ {i} \\left(\\mathbf {f} _ {\\mathrm {t}}\\right) = \\mathbf {f} ^ {\\prime} + \\mathbf {w} _ {\\mathrm {t}} \\cdot \\mathbf {f} _ {\\mathrm {t}} ^ {\\mathrm {c}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.469 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{\\mathrm{s}}^{\\prime}\\) and \\(\\mathbf{f}_{\\mathrm{t}}^{\\prime}\\) are output features of \\(\\mathcal{A}_i\\) for the source and target domain. Since \\(\\mathbf{w}_{\\mathrm{s}}\\) and \\(\\mathbf{w}_{\\mathrm{t}}\\) are involved in the computation of \\(\\mathbf{f}_{\\mathrm{s}}^{\\prime}\\) and \\(\\mathbf{f}_{\\mathrm{t}}^{\\prime}\\), SEEN modules \\(\\mathcal{S}^{\\mathrm{s}}\\) and \\(\\mathcal{S}^{\\mathrm{t}}\\) will be optimized by the classification losses from the two domains respectively, thereby learning domain-specific semantic information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.719, + 0.498 + ], + "angle": 0, + "content": "3.3. Optimization Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.655 + ], + "angle": 0, + "content": "We introduce multiple domain-aware attention modules into the frozen visual encoder as learnable attachments. To this end, we expect them to capture visual features suitable for cross-domain adaptation and supplement domain-specific semantic information. First, the domain discriminator \\(\\mathcal{D}\\) is applied to the inter-domain attention branch to learn domain-invariant features. Formally, we minimize the adversarial loss over the output of the inter-domain attention branch \\(\\mathbf{f}'\\) for the source image \\(x_{\\mathrm{s}}\\) and the target image \\(x_{\\mathrm{t}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.668, + 0.892, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} = - \\sum_ {i = 1} ^ {N} \\left[ \\mathbb {E} _ {x _ {\\mathrm {s}}} \\left\\| \\mathcal {D} _ {i} \\left(\\mathbf {f} ^ {\\prime}\\right) \\right\\| _ {2} ^ {2} + \\mathbb {E} _ {x _ {\\mathrm {t}}} \\left\\| \\mathcal {D} _ {i} \\left(\\mathbf {f} ^ {\\prime}\\right) - \\mathbf {1} \\right\\| _ {2} ^ {2} \\right], \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.722, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Second, the detection loss is calculated over both domains to learn the semantic commonalities and disparities. For the source domain, we minimize the cross-entropy \\(\\mathcal{L}_{\\mathrm{s}}\\) between the logits and ground truth. For the target domain, we use the prompt \"A photo of [Class] in foggy day\" to generate high-confidence pseudo labels and minimize the cross-entropy \\(\\mathcal{L}_{\\mathrm{t}}\\). Overall, the optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.842, + 0.892, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {t}} \\mathcal {L} _ {\\mathrm {t}} + \\lambda_ {\\mathrm {a d v}} \\mathcal {L} _ {\\mathrm {a d v}} + \\mathcal {L} _ {\\mathrm {r e g}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{reg}}\\) is the regression loss, and \\(\\lambda_{\\mathrm{t}}, \\lambda_{\\mathrm{adv}}\\) are balance ratios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25469" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison (%) with existing methods on Cross-Weather adaptation Cityscapes \\(\\rightarrow\\) Foggy Cityscapes (C→F), Cross-Fov adaptation KITTI \\(\\rightarrow\\) Cityscapes (K→C) and Sim-to-Real adaptation SIM10K \\(\\rightarrow\\) Cityscapes (S→C)." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.123, + 0.849, + 0.394 + ], + "angle": 0, + "content": "
MethodsVenueC→FK→CS→C
PersonRiderCarTruckBusTrainMotorBicyclemAPmAPmAP
DA-Faster [3]CVPR 201829.240.443.419.738.328.523.732.732.041.938.2
VDD [59]ICCV 202133.444.051.733.952.034.734.236.840.0--
SCAN [28]AAAI 202241.743.957.328.748.648.731.037.342.145.852.6
TIA [64]CVPR 202252.138.149.737.734.846.348.631.142.344.0-
LRA [42]TNNLS 202445.647.159.731.252.444.628.139.543.549.455.7
SIGMA++ [30]TPAMI 202346.445.161.032.152.244.634.839.944.549.557.7
CIGAR [36]CVPR 202346.147.362.127.856.644.333.741.344.948.558.5
OADA [61]ECCV 202247.846.562.932.148.550.934.339.845.447.859.2
MTM [57]AAAI 202451.053.467.237.254.441.638.447.748.9-58.1
AT [31]CVPR 202256.351.964.238.545.555.154.335.050.9--
SOCKER [5]ACMMM 202451.757.768.638.251.647.541.651.751.1-63.8
DSD-DA [11]ICML 202449.059.665.335.761.046.543.957.352.349.352.5
CAT [22]CVPR 202444.657.163.740.866.049.744.953.052.5--
NSA-UDA [66]ICCV 202350.360.167.737.457.446.947.354.352.755.656.3
REACT [24]TIP 202451.457.967.437.758.452.844.654.653.159.558.6
DA-Pro [23]NeurIPS 202355.462.970.940.363.454.042.358.055.961.462.9
RegionCLIP [65]CVPR 202249.655.063.234.155.648.336.047.048.659.158.9
SEEN-DA (Ours)-58.564.571.742.061.254.847.159.957.567.166.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.184, + 0.4, + 0.786, + 0.415 + ], + "angle": 0, + "content": "Table 2. Comparison (%) with existing methods on Cross-Style adaptation task Pascal VOC→Clipart" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.42, + 0.891, + 0.616 + ], + "angle": 0, + "content": "
MethodsAeroBikeBirdBoatBottleBusCarCatChairCowTableDogHorseMotorPersonPlantSheepSofaTrainTvmAP
UaDAN [17]35.073.741.024.421.369.853.52.334.261.231.029.547.963.662.261.313.97.648.623.940.2
TFD [54]27.964.828.429.525.764.247.713.547.550.950.821.333.960.265.642.515.140.545.548.641.2
FGRR [1]30.852.135.132.442.262.842.621.442.858.633.520.837.281.466.250.321.529.358.247.043.3
UMT [6]39.659.132.435.045.161.948.47.546.067.621.429.548.275.970.556.725.928.939.443.644.1
SIGMA [29]40.155.437.431.154.954.346.623.044.765.623.022.042.855.667.255.232.940.845.058.644.5
ATMT [27]37.563.437.929.845.162.741.219.543.757.422.925.339.687.170.950.629.132.258.450.545.2
CIGAR [36]35.255.039.230.760.158.146.931.847.061.021.826.744.652.468.554.431.338.856.563.546.2
TIA [64]42.266.036.937.343.771.849.718.244.958.918.229.140.787.867.449.727.427.857.150.646.3
SIGMA++ [30]36.354.640.131.658.060.446.233.644.466.225.725.344.458.864.855.436.238.654.159.346.7
CMT [51]39.856.338.739.760.435.056.07.160.160.435.828.167.884.580.155.520.332.842.338.247.0
RegionCLIP [65]38.170.448.837.344.855.843.514.448.247.814.318.358.378.467.922.230.116.948.450.242.7
SEEN-DA(Ours)44.173.454.747.145.176.051.620.451.753.018.517.361.886.872.222.837.721.158.952.747.9
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.627, + 0.202, + 0.645 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.471, + 0.729 + ], + "angle": 0, + "content": "We evaluate our method on mainstream DAOD scenarios, including the Cross-Weather, Cross-Fov, Sim-to-Real and Cross-Style. Further ablation studies are conducted to validate the effectiveness of the proposed domain-aware attention module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.182, + 0.755 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Cross-Weather Cityscapes [4] contains diverse street scenes in daylight, consisting of 2,975 training and 500 validation images annotated with eight classes. Foggy Cityscapes [48] simulates three distinct densities of fog on Cityscapes, containing 8,925 training images and 1,500 validation images. A standard configuration for cross-weather adaptation is to take the training set of Cityscapes as the source domain and the training set of foggy Cityscapes as the target domain, evaluating cross-weather adaptation per" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.657 + ], + "angle": 0, + "content": "formance on the 1500-sized validation set in all eight categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Cross-FoV KITTI [14] is a crucial dataset for self-driving that includes 7,481 photos annotated with cars. Collected by driving in rural areas and on highways, it provides data with a different Field of View (FoV). To fairly compare with other methods, we migrate KITTI to Cityscapes solely on the car category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Sim-to-Real The synthetic dataset SIM10k [21] has 10,000 photos from the GTA V video game with bounding boxes of the class car. We follow existing works to perform sim-to-real adaptation and report the performance on class car." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Cross-Style Pascal VOC [9] dataset consists of 16551 images from the real world and has 20 categories. Clipart [20] contains 1000 comical images and has the same 20 categories as Pascal VOC. Following the mainstream splitting, we use Pascal VOC as the source domain and all Clipart images as the target domain." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25470" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 3. Comparison \\((\\%)\\) of standard attention with domain-aware attention." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.123, + 0.468, + 0.197 + ], + "angle": 0, + "content": "
BenchmarkSelf-AttentionCross-AttentionDomain-aware Attention
Cross-Weather54.855.557.5
Cross-FoV61.963.567.1
Sim-to-Real62.363.766.8
Cross-Style43.145.247.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.206, + 0.468, + 0.234 + ], + "angle": 0, + "content": "Table 4. Ablation studies (%) of domain-aware attention on Cross-Weather adaptation." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.239, + 0.468, + 0.317 + ], + "angle": 0, + "content": "
Inter-domain AttentionLadvIntra-domain AttentionmAPGains
52.6-
✓54.9+2.3
✓✓55.8+0.9
✓✓✓57.5+1.7
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.295, + 0.341 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.514 + ], + "angle": 0, + "content": "Following [23], we adapt RegionCLIP(ResNet-50 [18]) with the Faster-RCNN [45] as the detector. In each iteration, one batch of source images with ground truth and one batch of target domain images with pseudo labels are forwarded to calculate the classification, adversarial and regression loss. The hyperparameter \\(\\lambda_{adv}\\), \\(\\lambda_t\\) is set to 0.1 and 1.0. We set the batch size of each domain to 8 and use the SGD optimizer with a warm-up learning rate. Mean Average Precision (mAP) with a threshold of 0.5 is taken as the evaluation metric. All experiments are deployed on 8 Tesla V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.35, + 0.543 + ], + "angle": 0, + "content": "4.3. Comparison to SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.469, + 0.61 + ], + "angle": 0, + "content": "We present representative state-of-the-art DAOD approaches for comparison, including feature alignment, semi-supervised learning and VLM-based domain alignment methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.702 + ], + "angle": 0, + "content": "Cross-Weather Adaptation Scenario As shown in Table 1 (C→F), the proposed SEEN-DA outperforms all compared methods in terms of mAP and advances SOTA by \\(1.6\\%\\), from \\(55.9\\%\\) to \\(57.5\\%\\). Specifically, our method improves performance over six categories (i.e. person, rider, car, truck, train, and bicycle) ranging from \\(0.8\\%\\) to \\(3.1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Cross-FOV Adaptation Scenario Table \\(1(\\mathrm{K}\\rightarrow \\mathrm{C})\\) reports result for KITTI \\(\\rightarrow\\) Cityscapes. SEEN-DA achieves SOTA performance of \\(67.1\\%\\) mAP, gaining an improvement of \\(5.7\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Sim-to-Real Adaptation Scenario Table 1 (S→C) shows that the proposed method achieves the best results of \\(66.8\\%\\) mAP, outperforming the previous best entry SOCCER [5] \\(63.8\\%\\) with \\(3.0\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Cross-Style Adaptation Scenario Additionally, we assess SEEN-DA on the more challenging Cross-Style adaptation, where the semantic hierarchy has broader discrepancies. In Table 2, SEEN-DA peaks with \\(47.9\\%\\) mAP and improves six categories (aeroplane, bird, boat, bus, sheep and train)." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 5. Comparison \\((\\%)\\) of computational efficiency on Cross-Weather adaptation" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.122, + 0.89, + 0.214 + ], + "angle": 0, + "content": "
MethodBackbone Parm(M) LearnableParm(M) mAPAbs. Gains
DSS [56]26.83426.83440.9+4.2
CSDA [13]33.64533.64545.3+6.9
REACT [24]29.81229.81248.8+7.9
AT [31]39.22518.72350.9+7.9
DA-Pro [23]34.8340.00855.9+3.3
SEEN-DA (Ours)36.7011.87557.5+4.9
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.223, + 0.665, + 0.238 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.277 + ], + "angle": 0, + "content": "We report detailed ablation studies (Table 3, 4, 5, Figure 4, 5) to validate the effectiveness of proposed SEEN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Standard Attention vs. Domain-aware Attention We compare the performance of domain-aware attention with existing attention methods, including self-attention and cross-attention, as shown in Table 3. With the guidance of textual information, the cross-attention improves the self-attention by \\(0.7\\% \\sim 2.1\\%\\). Furthermore, the proposed domain-aware attention gains improvements of \\(2.0\\% \\sim 33.6\\%\\) over cross-attention, showing a superior ability to select features with high transferability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Ablation for Domain-aware Attention Module We conduct ablation experiments to further analyze the significance of proposed domain-aware attention module in Table 4. Introducing the inter-domain attention branch attains an mAP of \\(54.9\\%\\), and optimizing with adversarial loss \\(\\mathcal{L}_{adv}\\) gains \\(0.9\\%\\). Meanwhile, applying the intra-domain attention branch improves \\(1.7\\%\\) on the target domain. These results showing the effectiveness of the domain-aware attention module to eliminate redundant information and supplement domain-specific semantic information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Computational Efficiency We discussed the computational efficiency of the proposed SEEN-DA in Table 5. Our method yields the same parameters scale as existing methods and achieves the highest mAP. In addition, SEEN-DA only needs to train very few parameters (1.875M), showing remarkable efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Numbers of the Attention Module We explore the impact of numbers of domain-aware attention modules in Figure 4. Starting from the last visual block, gradually introducing domain-aware attention modules to each visual block improves performance and peaks when using 3 modules. Considering efficiency, we apply 3 domain-aware attention modules into the visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Projection layer in SEEN module We explore the projection layer in the SEEN module in Figure 5. Projecting textual embedding into visual space, denoted as T2V, only achieves \\(55.1\\%\\). Projecting textual and visual embedding into a latent space, denoted as T&V, improves performance by \\(2.1\\%\\). And projecting visual embedding into textual space achieves the best mAP of \\(57.5\\%\\). We analyze the reason for these results as follows. Visual embedding contains more sample noise and redundant information than textual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.087, + 0.355, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.291, + 0.272, + 0.303 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.087, + 0.611, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.291, + 0.534, + 0.304 + ], + "angle": 0, + "content": "(b) DA-Pro[23]" + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.087, + 0.871, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.291, + 0.8, + 0.304 + ], + "angle": 0, + "content": "(c) SEEN-DA (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.309, + 0.893, + 0.339 + ], + "angle": 0, + "content": "Figure 3. Detection comparison on the Cross-Weather adaptation scenario. We visualize (a) the ground truth, the detection boxes of (b) SOTA DA-Pro [23] and (c) our methods SEEN-DA." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.347, + 0.415, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.536 + ], + "angle": 0, + "content": "Figure 4. Ablation \\((\\%)\\) on the number of attention modules on Cross-Weather adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.469, + 0.595 + ], + "angle": 0, + "content": "embedding. And learning to project visual embedding to textual space can reduce the interference of these factors. Therefore, we adopt the V2T projection in SEEN module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.597, + 0.216, + 0.612 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.833 + ], + "angle": 0, + "content": "The detection results of ground truth (a), DA-Pro [23] (b) and the proposed SEEN-DA (c) on the target domain of Cross-Weather adaptation are displayed in Figure 3. The figure shows that SEEN-DA outperforms the SOTA method in terms of the ability to detect objects. Specifically, SEEN-DA detects the objects missed (e.g. the rider on the right of figures in 1st row and the person on the left of figures in 2nd row) and misclassified (e.g. the bicycle on the middle of figures in 1st row and the person on the left of figures in 2nd row), revealing the effectiveness of domain-aware attention. By utilizing semantic entropy to reduce redundant information and supplement domain-specific semantic information, the SEEN-DA improves the discriminability of the detector on the target domain." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel SEMantic ENtropy guided Domain-aware Attention (SEEN-DA) for DAOD." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.345, + 0.831, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.507, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Figure 5. Ablation \\((\\%)\\) on projection layer of the domain-aware attention module on Cross-Weather adaptation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.732 + ], + "angle": 0, + "content": "We design the semantic entropy to quantify the semantic information contained in visual features. Guided by semantic entropy, the domain-aware attention module is introduced to adaptively refine visual features with the semantic commonalities and disparities between domains. Precisely, it consists of an inter-domain attention branch for extracting domain-invariant features and eliminating redundant information between domains, and an intra-domain attention branch for supplementing the domain-specific semantic information for each domain. Extensive experiments over multiple benchmarks validate the effectiveness of SEENDA in inferring a discriminative detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment This work is partially supported by the National Key R&D Program of China(under Grant 2023YFB4502200), the NSF of China(under Grants 92364202, U22A2028, 62102398, 62102399, 62302478, 62302482, 62376268), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660300, XDB0660301, XDB0660302), Major Program of ISCAS(Grant No. ISCAS-ZD-202402), Beijing Natural Science Foundation(4222039), CAS Project for Young Scientists in Basic Research(YSBR-029) and Youth Innovation Promotion Association CAS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25472" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Chaoqi Chen, Jiongcheng Li, Hong-Yu Zhou, Xiaoguang Han, Yue Huang, Xinghao Ding, and Yizhou Yu. Relation matters: foreground-aware graph-based relational reasoning for domain adaptive object detection. TPAMI, 45(03):3677-3694, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xin chao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, et al. Learning domain adaptive object detection with probabilistic teacher. In ICML, pages 3040-3055, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, pages 3339-3348, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Yiming Cui, Liang Li, Jiehua Zhang, Chenggang Yan, Hongkui Wang, Shuai Wang, Heng Jin, and Li Wu. Stochastic context consistency reasoning for domain adaptive object detection. In ACMMM, pages 1331-1340, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[6] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In CVPR, pages 4091-4101, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[7] Jinhong Deng, Dongli Xu, Wen Li, and Lixin Duan. Harmonious teacher for cross-domain object detection. In CVPR, pages 23829-23838, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[9] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[10] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul De Charette. Poda: Prompt-driven zero-shot domain adaptation. In ICCV, pages 18623-18633, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[11] Yongchao Feng, Shiwei Li, Yingjie Gao, Ziyue Huang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Dsd-da: Distillation-based source debiasing for domain adaptive object detection. In ICML, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, 17(1):2096-2030, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[13] Changlong Gao, Chengxu Liu, Yujie Dun, and Xueming Qian. Csda: Learning category-scale joint feature for domain adaptive object detection. In ICCV, pages 11421-11430, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu, and Yanpeng Cao. Uncertainty-aware unsupervised domain adaptation in object detection. TMM, 24:2502-2514, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[19] Mengzhe He, Yali Wang, Jiaxi Wu, Yiru Wang, Hanqing Li, Bo Li, Weihao Gan, Wei Wu, and Yu Qiao. Cross domain object detection by target-perceived dual branch distillation. In CVPR, pages 9570-9580, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[20] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, pages 5001-5009, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Matthew Johnson-Roberson, Charles Barto, Rounak Mehta, Sharath Nittur Sridhar, Karl Rosaen, and Ram Vasudevan. Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks? In ICRA 2017, pages 746-753. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[22] Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T Tan. Cat: Exploiting inter-class dynamics for domain adaptive object detection. In CVPR, pages 16541-16550, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[23] Haochen Li, Rui Zhang, Hantao Yao, Xinkai Song, Yifan Hao, Yongwei Zhao, Ling Li, and Yunji Chen. Learning domain-aware detection head with prompt tuning. NeurIPS, 36, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[24] Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, and Ling Li. React: Remainder adaptive compensation for domain adaptive object detection. TIP, 33:3735-3748, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[27] Kai Li, Curtis Wigington, Chris Tensmeyer, Vlad I. Morariu, Handong Zhao, Varun Manjunatha, Nikolaos Barmpalios, and Yun Fu. Improving cross-domain detection with self-supervised learning. In CVPR, pages 4745-4754, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[28] Wuyang Li, Xinyu Liu, Xiwen Yao, and Yixuan Yuan. Scan: Cross domain object detection with semantic conditioned adaptation. In AAAI, volume 6, page 7, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object detection. In CVPR, pages 5291-5300, 2022. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25473" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[30] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI, 45(07):9022-9040, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[31] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Va-jda. Cross-domain adaptive teacher for object detection. In CVPR, pages 7581-7590, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.209, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[32] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.266, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.323, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[34] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. TMM, 25:1333-1344, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.381, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[35] Rui Liu, Yahong Han, Yaowei Wang, and Qi Tian. Frequency spectrum augmentation consistency for domain adaptive object detection. arXiv preprint arXiv:2112.08605, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.424, + 0.469, + 0.478 + ], + "angle": 0, + "content": "[36] Yabo Liu, Jinghua Wang, Chao Huang, Yaowei Wang, and Yong Xu. Cigar: Cross-modality graph reasoning for domain adaptive object detection. In CVPR, pages 23776-23786, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.482, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[37] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.526, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[38] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. NeurIPS, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.569, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[39] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, pages 2208-2217, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.613, + 0.469, + 0.654 + ], + "angle": 0, + "content": "[40] Fabio Maria Carlucci, Lorenzo Porzi, Barbara Caputo, Elisa Ricci, and Samuel Rota Bulo. Autodial: Automatic domain alignment layers. In ICCV, pages 5067-5075, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.657, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[41] Pietro Morerio, Jacopo Cavazza, and Vittorio Murino. Minimal-entropy correlation alignment for unsupervised deep domain adaptation. In ICLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.701, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[42] Zhengquan Piao, Linbo Tang, and Baojun Zhao. Unsupervised domain-adaptive object detection via localization regression alignment. TNNLS, 35(11):15170-15181, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.744, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[44] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[46] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In ICCV, pages 8050-8058, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[47] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In CVPR, pages 6956–6965, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[48] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. IJCV, 126(9):973-992, Sep 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[49] Sunandini Sanyal, Ashish Ramayee Asokan, Suvaansh Bhambri, Akshay Kulkarni, Jogendra Nath Kundu, and R Venkatesh Babu. Domain-specificity inducing transformers for source-free domain adaptation. In ICCV, pages 18928-18937, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[50] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NPIS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[51] VS Vibashan, Poojan Oza, and Vishal M Patel. Instance relation graph guided source-free domain adaptive object detection. In CVPR, pages 3520-3530, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[52] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[53] Vibashan Vs, Vikram Gupta, Poojan Oza, Vishwanath A Sindagi, and Vishal M Patel. Mega-cda: Memory guided attention for category-aware unsupervised domain adaptive object detection. In CVPR, pages 4516-4526, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[54] Haoan Wang, Shilong Jia, Tieyong Zeng, Guixu Zhang, and Zhi Li. Triple feature disentanglement for one-stage adaptive object detection. In AAAI, pages 5401-5409, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[55] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, volume 33, pages 5345-5352, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[56] Yu Wang, Rui Zhang, Shuo Zhang, Miao Li, Yangyang Xia, XiShan Zhang, and ShaoLi Liu. Domain-specific suppression for adaptive object detection. In CVPR, pages 9603-9612, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[57] Weixi Weng and Chun Yuan. Mean teacher detr with masked feature alignment: A robust domain adaptive detection transformer framework. In AAAI, pages 5912-5920, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[58] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. TPAMI, 44(8):4178-4193, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[59] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, pages 9342-9351, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[60] Xiaofu Wu, Suofei Zhang, Quan Zhou, Zhen Yang, Chunming Zhao, and Longin Jan Latecki. Entropy minimization versus diversity maximization for domain adaptation. TNNLS, 34(6):2896-2907, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[61] Jayeon Yoo, Inseop Chung, and Nojun Kwak. Unsupervised domain adaptation for one-stage object detector using offsets to bounding box. In ECCV, pages 691-708, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[62] Yifan Zhang, Hanbo Chen, Ying Wei, Peilin Zhao, Jiezhang Cao, Xinjuan Fan, Xiaoying Lou, Hailing Liu, Jinlong Hou," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25474" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.146 + ], + "angle": 0, + "content": "Xiao Han, et al. From whole slide imaging to microscopy: Deep microscopy adaptation network for histopathology cancer image classification. In MICCAI, pages 360-368, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[63] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, pages 12425-12434, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[64] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In CVPR, pages 14217-14226, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.302 + ], + "angle": 0, + "content": "[65] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[66] Wenzhang Zhou, Heng Fan, Tiejian Luo, and Libo Zhang. Unsupervised domain adaptive detection with network stability analysis. In ICCV, pages 6986-6995, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25475" + } + ] +] \ No newline at end of file diff --git a/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_origin.pdf b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1f0247f0472784bd94a98043e780e4889ad223a2 --- /dev/null +++ b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/48c4d107-49aa-405e-ad25-28dd7ace0585_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02ea9c5b3bcf11f5f49aee66341f987a65548d641683788a455a546d129211a0 +size 8726486 diff --git a/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/full.md b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7858aae6b2bfeba1ef7658d2c7bed084834b1b68 --- /dev/null +++ b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/full.md @@ -0,0 +1,385 @@ +# SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection + +Haochen Li $^{1,4}$ Rui Zhang $^{2*}$ Hantao Yao $^{3}$ Xin Zhang $^{2}$ Yifan Hao $^{2}$ Xinkai Song $^{2}$ + +Shaohui Peng $^{1}$ Yongwei Zhao $^{2}$ Chen Zhao $^{5}$ Yanjun Wu $^{5}$ Ling Li $^{1,4*}$ + +1Intelligent Software Research Center, Institute of Software, CAS + +$^{2}$ State Key Lab of Processors, Institute of Computing Technology, CAS + +$^{3}$ School of Information Science and Technology, University of Science and Technology of China + +$^{4}$ University of Chinese Academy of Sciences $^{5}$ Institute of Software, CAS + +{haochen2021, pengshaohui, zhaochen, wuyanjun, liling}@iscas.ac.cn + +{zhangrui, zhangxin, haoyifan, songxinkai, zhaoyongwei}@ict.ac.cn, yaohantao@ustc.edu.cn + +# Abstract + +Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain to an unlabelled target domain. Traditional works focus on aligning visual features between domains to extract domain-invariant knowledge, and recent VLM-based DAOD methods leverage semantic information provided by the textual encoder to supplement domain-specific features for each domain. However, they overlook the role of semantic information in guiding the learning of visual features that are beneficial for adaptation. To solve the problem, we propose semantic entropy to quantify the semantic information contained in visual features, and design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to adaptively refine visual features with the semantic information of two domains. Semantic entropy reflects the importance of features based on semantic information, which can serve as attention to select discriminative visual features and suppress semantically irrelevant redundant information. Guided by semantic entropy, we introduce domain-aware attention modules into the visual encoder in SEEN-DA. It utilizes an inter-domain attention branch to extract domain-invariant features and eliminate redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. Comprehensive experiments validate the effectiveness of SEEN-DA, demonstrating significant improvements in cross-domain object detection performance. + +# 1. Introduction + +Object detection [8, 33, 44, 45] have been proven to attain encouraging performance assisted by high-quality im + +ages with accurate annotations. However, due to the noticeable discrepancy between domains, they suffer severe performance degradation when dealing with unseen data in distinct scenes. To alleviate this problem, domain adaptive object detection (DAOD) [3] aims to transfer a detector trained on the labelled source domain to a unlabelled target domain. + +Traditional DAOD works [3, 30, 47] focus on reducing domain discrepancy by aligning visual features between domains to extract domain-invariant knowledge. Additionally, [24,32,49,58] decouple the unique features of each domain, i.e.domain-specific features, to prevent its interference during feature alignment. Although effective, they train the detector with domain-shared and semantically agnostic category labels (such as one-hot encoding), which ignores the linguistic properties of categories and inter-domain differences and commonality, i.esemantic information. Recently, visual-language models (VLMs) [26,43,65] are proposed to ground visual with textual modalities. Through the textual encoder, they align visual features with the textual prompt, which is constructed from natural language and contains rich semantic information. Motivated by this, recent VLM-based DAOD methods, such as DA-Pro [23], utilize the textual encoder as detection head and tune the textual prompt to learn domain-invariant semantics while supplementing domain-specific semantics for each domain. However, they freeze the visual encoder and leverage semantic information solely in the detection head, limiting the discriminability of visual features. + +Summarily, despite promising performance, existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. They typically optimize the visual encoder with semantically agnostic category labels and adversarial loss, or directly freeze it, resulting in two limitations. First, they fail to filter out redundant information from visual fea + +![](images/a212bc74c96c4ce6cb01b6d07cad6562cc92ff3bc4ba1abe480c28abb9cf1fe1.jpg) +(a.1) Source Image and Bounding Box + +![](images/908a4ff3771a5e61c585974b055e63d1c23ab40ee7f3fd10a995a26ff50b58b7.jpg) +(b.1) Detection Results on Source Image (DA-Pro[23]) + +![](images/293a8e9c282983095cb06ea627fd148bb8b416015d3311a80b0c4e1f5ddea2be.jpg) +(c.1) Detection Results on Source Image (SEEN-DA (Ours)) +Figure 1. Visualization of the detection results and attention map. (a) Source and Target domain image and bounding boxes. (b) Existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. Infected by the redundant information and lack of domain-specific semantic information, some objects that are correctly detected on the source image are misclassified or missed on the target image. (c) We propose semantic entropy to evaluate the importance of visual features based on domain and category semantic information. By integrating it as domain-aware attention in SEEN-DA, we eliminate redundant information and enhance domain-specific features, thereby improving the detector's discriminability for DAOD. + +![](images/4a3d9277509ec7243cc9772c15b82d663d5bb3274b05a7ec3d00338cf4973a01.jpg) + +![](images/9c6518125155277825551b384999203dbf6fb5a77b2d0e08bdbb2758378558d1.jpg) +(a.2) Target Image and Bounding Box +(b.2) Detection Results on Target Image (DA-Pro[23]) + +![](images/7e417aaf183812e4524c94166600cf237e0240a1354ae410791fb85d5a44ee6e.jpg) +(c.2) Detection Results on Target Image (SEEN-DA (Ours)) + +![](images/978473a08a87f9f2bbcfd8871dfced3d83d740b65892bf964691f2687470e98c.jpg) + +![](images/5865e700f55eaf2b81d9307f6e69af825ad5b1cf9e6393192a387bf7bc6169b3.jpg) +Target Image (DA-Pro[23]) + +![](images/38303ba9edf02acf570da08b7dc804221edaff239e9f735f2b004943376f7f1e.jpg) + +Target Image (SEEN-DA (Ours)) + +![](images/f06f523db07280dc62c5650973862f875db4fa1a8aeff94094a2539d2f5a9e0e.jpg) +(d) Domain-aware Attention Map in SEEN-DA (Ours) + +tures, which is semantically irrelevant and can lead to incorrect feature alignment. Redundant information includes attributes unrelated to object identification, such as color, texture, etc. Although they do not contain domain and category semantic information, they are often extracted by adversarially trained visual encoder since they exist in both domains. Therefore, such semantically irrelevant information may interfere with the alignment of domain-invariant features. In Fig. 1(b.1), existing methods detect the bicycle covered with a white cloth on the source image (clear day). However, since target images (foggy day) tend to be white overall, they misclassify the white cloth as a rider in Fig. 1(b.2). Second, existing methods discard domain-specific features in the visual encoder, which contain rich domain semantic information and are important for the discriminability on the target domain. For example, they correctly detect the bicycle and rider on the source image Fig. 1(b.1), but miss them in Fig. 1(b.2) due to insufficient feature learning of domain-specific properties of target domain, e.g.fog. + +To solve these problems, we propose semantic entropy which evaluates the importance of features based on domain and category semantic information, and accordingly design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to eliminate redundant information and supplement domain-specific features for DAOD. The semantic entropy of visual features refers to the amount of semantic information related to a certain domain. As the textual encoder of VLM provides rich domain and category semantic information in textual embedding, the semantic entropy can be expressed as the information entropy of the classification probability between the visual and textual embedding. Therefore, semantic entropy reflects the importance of visual features in adaptation. Low semantic entropy indicates + +that visual features match the textual embedding of a certain category and are likely to be transferable foreground objects. On the contrary, high semantic entropy means that they are unpredictable and may be redundant or even harmful for detection. + +Applying semantic entropy as attention, we design the SEEN-DA framework, introducing domain-aware attention modules into the visual encoder to select visual features with low semantic entropy and suppress regions with high ones. Specifically, the domain-aware attention module utilizes two parallel branches to learn inter- and intra-domain attention with semantic entropy. The inter-domain attention branch is shared between domains and evaluates the semantic entropy on domain-shared textual embedding, aiming to learn domain-invariant features while eliminating redundant information. The intra-domain attention branch is private in each domain to evaluate the semantic entropy on source/target textual embedding respectively, supplementing domain-specific features for each domain. The SEEN-DA is able to adaptively adjust visual features with the semantic commonalities and disparities between domains. As shown in Fig. 1(c), domain-aware attention map highlights features with rich semantic information, i.eforeground regions, and hinders the negative impact of background regions. Thereby, the proposed SEEN-DA correctly detects objects misclassified by existing methods, improving the discriminability of the detector on the target domain. + +We evaluate the proposed method on four benchmarks. The experimental results highlight the effectiveness of our SEEN-DA, e.g., it reaches $57.5\%$ mAP on Cityscapes $\rightarrow$ Foggy Cityscapes, surpassing the state-of-the-art DA-Pro [23] by $1.6\%$ . + +# 2. Related Work + +Visual-Language Models Large-scale visual-language models (VLMs) [25, 26, 43] embed visual and text modalities into a shared space, enabling cross-modal alignment. Pre-trained with an astonishing scale of image-text pairs, they demonstrate comprehensive visual understanding. As a pioneering work, CLIP [43] trains a visual encoder and a textual encoder with 400 million image-text pairs in a contrastive manner, showing promising classification ability for the seen and unseen classes. Following this spirit, [10, 16, 52, 65] freeze the textual encoder and adapt the visual encoder into Faster-RCNN [45] as a two-stage detector. In this work, we apply the textual encoder of CLIP [43] to infer textual embedding that contains semantic information, and the visual encoder of RegionCLIP [65] as the backbone. + +Domain Adaptive Object Detection (DAOD) aims to adapt the object detector [45] trained on the labelled source domain to the unlabelled target domain. Previous approaches can be broadly divided into three categories: feature alignment, semi-supervised learning and VLM-based domain alignment. As a fundamental direction, feature alignment [12, 37, 39] aims to align the feature distributions of the two domains to generate domain-invariant features [3, 30, 47]. Typically, the domain discriminator is proposed to align feature distribution adversarially in image level [3], instance level [47, 63] and category level [29, 53]. To reduce the negative impact of aligning unique features of each domain, [24, 32, 34, 49, 54, 58] decouple the domain-invariant and domain-specific knowledge between two domains. In parallel, semi-supervised learning strives to generate unbiased training data with style transfer [6, 19, 35] and pseudo label [2, 7, 31, 50]. However, they train the detector with semantically agnostic category labels (such as one-hot encoding), ignoring the semantic information inherently contained in the category name. To overcome this, recent VLM-based domain alignment methods exploit the semantic information provided by the textual encoder of VLMs to learn semantic commonalities and disparities. Pioneeringly, DA-Pro [23] tuning the textual encoder into a domain-aware detection head, learning domain-invariant semantics while supplementing domain-specific semantics for each domain. Despite the promising performance, existing methods freeze the visual encoder and ignore using semantic information to guide the learning of visual features, which leads to two limitations. First, they fail to filter out redundant information from visual features, which is semantically irrelevant and can lead to incorrect feature alignment. Second, they discard domain-specific features in the visual encoder, which contain rich domain semantic information and are vital for DAOD. + +Entropy Optimization Information entropy was first utilized as entropy minimization in [15] for semi-supervised learning to learn decision rules for unlabelled data. Con- + +sidering its effectiveness, recent works introduce entropy optimization into DAOD. [40, 41, 46, 62] propose minimizing entropy on unlabelled data to make decisions consistent in the source and target domains. [55] apply entropy as metrics to measure the transferability of data samples. [38] employs entropy as attention to control the uncertainty of the classifier predictions, alleviating the effect of negative transfer. Instead of directly minimizing entropy, [60] further introduces diversity maximization to explore close-to-ideal domain adaptation solutions. However, the computation of entropy in existing methods is inherently semantically agnostic. Based on the output probabilities from a classifier, they treat categories merely as identifiers and convert the category labels into one-hot encoding. Therefore, they ignore the rich semantic information embedded within the category names, which is vital for DAOD. In this paper, we propose a novel semantic entropy, evaluating the transferability of visual features through the semantic commonalities and disparities between domains, improving the discrimination of the visual encoder. + +# 3. Methodology + +In this section, we present a novel SEmantic entropy guided Domain-aware Attention (SEEN-DA) for DAOD, which introduces domain-aware attention module to eliminate redundant information and supplement domain-specific features, guided by the proposed semantic entropy. Semantic entropy quantifies the amount of semantic information within visual features, providing a measure of their importance in domain adaptation. Utilizing semantic entropy as attention, the domain-aware attention module adopts an inter-domain attention branch to extract domain-invariant features and eliminate semantically irrelevant redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. As shown in Fig. 2(a), given input image $x$ , the visual encoder of SEEN-DA is divided into $N$ frozen visual block $\{\mathcal{V}_i\}_{i=1}^N$ , and each domain-aware attention module $\{\mathcal{A}_i\}_{i=1}^N$ is attached to $\mathcal{V}_i$ as a lightweight learnable module: + +$$ +\mathbf {h} _ {0} = \mathcal {E} (x); \mathbf {h} _ {i} = \mathcal {A} _ {i} \left(\mathcal {V} _ {i} \left(\mathbf {h} _ {i - 1}\right)\right), \tag {1} +$$ + +where $\mathcal{E}$ is the stem layer. + +# 3.1. Semantic Entropy (SEEN) + +Generally speaking, given feature $\mathbf{f}$ , the probability that this feature belongs to a specific object category is: + +$$ +p (c, \mathbf {f}) = \frac {\exp \left(v _ {c} (\mathbf {f})\right)}{\sum_ {c} \exp \left(v _ {c} (\mathbf {f})\right)}, \tag {2} +$$ + +where $c, v_{c}(\cdot)$ represents an object category and its corresponding classifier. $p(\cdot, \mathbf{f})$ is the probability distribution function, expected to output 1 on the corresponding category and 0 on other category, i.e. one-hot vector. + +![](images/141f1c0e6581a5a6849f274af53ffb0551d495f1a154bdb2cde23f68b4e371b2.jpg) +(a) Overview of SEEN-DA + +![](images/ef1b5a8e7dbefcf9007886a15b64d1c7b7196cebe73860b814f709b42e582c8c.jpg) +(b) Domain-aware Attention Module +Figure 2. (a) Overview of the proposed SEEN-DA for DAOD, where the semantic entropy is utilized as attention in domain-aware attention module. (b) The architecture of domain-aware attention module, consisting of an inter-domain and an intra-domain attention branch. + +Tradition DAOD approaches focus on reducing domain discrepancy by aligning $\mathbf{f}$ between the source and target domain. Although effective, as shown in Eq. 2, they treat categories merely as one-hot vectors, which is inherently semantically agnostic. Therefore, they ignore the rich semantic information inherently embedded within the category names. As feature alignment relies heavily on a profound understanding of the semantic disparities between the source and target domains, this limitation hinders the learning of visual features for adaptation. + +To address this, we propose semantic entropy, which quantifies the semantic information embedded in visual features by leveraging textual embeddings from Visual-language models (VLMs). VLMs utilize a textual encoder to encode textual prompts that describe category attributes into textual embeddings, aligning them with visual features. As a result, textual embeddings inherit rich semantic information from natural language, making them effective representations of the semantics of visual features. Motivated by this, we first infer textual embedding $\mathbb{T} = \{t_c\}_c^K$ for $K$ categories by feeding prompts like "A photo of [Class] in [Domain]" into a textual encoder, where each $t_c$ contains semantic information of category $c$ and its domain. Thus, the probability distribution that $\mathbf{f}$ corresponds to certain semantics can be estimated by the similarity between the visual feature $\mathbf{f}$ and textual embedding $\mathbb{T}$ : + +$$ +p \left(t _ {c}, \mathbf {f}\right) = \frac {\exp \left(\sin \left(t _ {c} , \mathbf {f}\right) / \tau\right)}{\sum_ {c} \exp \left(\sin \left(t _ {c} , \mathbf {f}\right) / \tau\right)}, \tag {3} +$$ + +where $\mathrm{sim}(\cdot ,\cdot)$ is the cosine similarity, and $\tau$ is a learnable temperature parameter. Though $p(t_{c},\mathbf{f})$ provides a measure of semantic relevance, it does not directly quantify the amount of semantic information embedded in $\mathbf{f}$ . Therefore, we draw inspiration from the concept of information entropy, which is widely used to measure the uncertainty and informativeness of a probability distribution. Formally, the proposed semantic entropy is defined as the information entropy of the probability distribution $p(t_{c},\mathbf{f})$ to evaluate the semantic uncertainty of $\mathbf{f}$ : + +$$ +\operatorname {S E} (\mathbb {T}, \mathbf {f}) = - \sum_ {c} p \left(t _ {c}, \mathbf {f}\right) \log \left(p \left(t _ {c}, \mathbf {f}\right)\right), \tag {4} +$$ + +In the framework of DAOD, we apply semantic entropy as attention to select high discriminative visual features. According to Eq. 4, features with low semantic entropy exhibit clear and definite semantic information, expected to be foreground features suitable for transfer. Conversely, features with high semantic entropy indicate high semantic uncertainty, suggesting they may contain semantically irrelevant redundant information, which should be deemphasized during adaptation. To achieve this, we regulate Eq. 4 to propose semantic entropy-guided attention: + +$$ +\operatorname {S E A t t e n t i o n} (\mathbb {T}, \mathbf {f}) = \sum_ {c} p \left(t _ {c}, \mathbf {f}\right) \log \left(p \left(t _ {c}, \mathbf {f}\right)\right) + \log K, \tag {5} +$$ + +where $K$ is the number of categories. Based on semantic entropy, Eq. 5 effectively adjusts the weight of visual features, focusing on features with rich semantic information while minimizing the influence of redundant information. + +# 3.2. Domain-aware Attention Module + +The domain-aware attention module is designed to adaptively refine visual features of each domain by utilizing the semantic entropy. As shown in Fig. 2(b), the module incorporates two complementary branches: an inter-domain attention branch to learn domain-invariant features and eliminate semantically irrelevant information, and an intra-domain attention branch to supplement the domain-specific features unique to each domain. + +Inter-domain Attention The inter-domain attention is designed to modify domain-invariant features of the frozen visual encoder with semantic information. Specifically, for the output visual features $\mathbf{f}_d = \mathcal{V}_i(\mathbf{h}_{i - 1})$ of the i-th visual block, we first forward it to a domain-shared convolution module $\mathcal{C}$ to capture task-related knowledge: + +$$ +\mathbf {f} = \mathcal {C} \left(\mathbf {f} _ {d}\right), \tag {6} +$$ + +where $d \in \{\mathrm{s},\mathrm{t}\}$ denotes whether $\mathbf{f}_d$ belongs to source domain s or target domain t. + +Simultaneously, the visual feature $\mathbf{f}_d$ is forwarded to the SEEN module $S$ to obtain attention map $\mathbf{w}$ guided by semantic entropy. With the help of VLM, we can represent the + +domain-invariant semantic information through domain-shared prompts. Technically, we input textual prompts such as "A photo of [Class]" to the textual encoder to retrieve domain-invariant textual embedding $\mathbb{T}$ . As described in Eq. 4, the similarity of visual features and textual embeddings is needed to calculate semantic entropy attention. To facilitate this, a projection layer $\mathcal{P}$ is firstly introduced to map $\mathbf{f}_d$ from the visual feature space to the textual embedding space: + +$$ +\mathbf {f} _ {\mathrm {p}} = \mathcal {P} \left(\mathbf {f} _ {d}\right), \tag {7} +$$ + +Then we calculate the semantic entropy-guided attention of the visual feature $\mathbf{f}_d$ over text embedding $\mathbb{T}$ : + +$$ +\mathbf {w} = \mathcal {S} (\mathbb {T}, \mathbf {f} _ {d}) = \operatorname {S E A t t e n t i o n} (\mathbb {T}, \mathbf {f} _ {\mathrm {p}}), \tag {8} +$$ + +With the attention map $\mathbf{w}$ , we can focus on semantically informative features and suppress redundant information in domain-invariant features. Typically, $\mathbf{w}$ is applied to $\mathbf{f}$ via an element-wise product, and the weighted feature is summed with the original visual feature: + +$$ +\mathbf {f} ^ {\prime} = \mathbf {w} \cdot \mathbf {f} + \mathbf {f} _ {d}, \tag {9} +$$ + +Expecting to be domain-invariant, the extracted $\mathbf{f}'$ is forwarded to a domain discriminator $\mathcal{D}$ , which aligns the source and target domain adversarily. + +Intra-domain Attention With the inter-domain attention branch, the extracted domain-invariant features contain sufficient semantic information shared between domains. However, it ignores the semantic discrepancy between the two domains, resulting in limited discriminability in the target domain. Therefore, we design the intra-domain attention branch to enrich visual features with domain-specific semantic information. Similar to the inter-domain attention branch, we first introduce a learnable layer $\mathcal{C}^c$ to adapt visual features for to both domains: + +$$ +\mathbf {f} _ {d} ^ {\mathrm {c}} = \mathcal {C} ^ {\mathrm {c}} \left(\mathbf {f} _ {d}\right), \tag {10} +$$ + +Subsequently, we generate domain-aware textual embedding that captures the semantics of each domains. Motivated by DA-Pro [23], we tune learnable domain-adaptive prompts, i.e. $\left[[v^{\mathrm{c}}][v^{\mathrm{s}}]\right]$ [Class] for the source domain and $\left[[v^{\mathrm{c}}][v^{\mathrm{t}}]\right]$ [Class] for the target domain, to explicitly model the domain and category semantic commonalities and disparities. These prompts are forwarded to the textual encoder to derive domain-adaptive textual embedding $\mathbb{T}_{\mathrm{s}},\mathbb{T}_{\mathrm{t}}$ for each domain. + +Due to the visual and semantic differences between the source and target domain, a domain-shared SEEN module is inadequate. To calculate the semantic entropy for each domain, visual features must be independently mapped into + +their respective textual spaces. Therefore, we apply separate SEEN modules $S^{\mathrm{s}}$ and $S^{\mathrm{t}}$ with independent projection layers $\mathcal{P}^{\mathrm{s}},\mathcal{P}^{\mathrm{t}}$ and textual embedding $\mathbb{T}_{\mathrm{s}},\mathbb{T}_{\mathrm{t}}$ : + +$$ +\mathbf {w} _ {\mathrm {s}} = \mathcal {S} ^ {\mathrm {s}} \left(\mathbb {T} _ {\mathrm {s}}, \mathbf {f} _ {\mathrm {s}}\right), \tag {11} +$$ + +$$ +\mathbf {w} _ {\mathrm {t}} = \mathcal {S} ^ {\mathrm {t}} \left(\mathbb {T} _ {\mathrm {t}}, \mathbf {f} _ {\mathrm {t}}\right), \tag {12} +$$ + +where the $\mathbf{w}_{\mathrm{s}}$ and $\mathbf{w}_{\mathrm{t}}$ denotes the source- and target-specific attention maps. + +The inter-domain attention branch extracts domain-invariant features while eliminating redundant information, whereas the intra-domain attention branch enriches visual features with domain-specific semantics. Overall, the output for the i-th domain-aware attention module $\mathcal{A}_i$ can be expressed as follows: + +$$ +\mathbf {f} _ {\mathrm {s}} ^ {\prime} = \mathcal {A} _ {i} \left(\mathbf {f} _ {\mathrm {s}}\right) = \mathbf {f} ^ {\prime} + \mathbf {w} _ {\mathrm {s}} \cdot \mathbf {f} _ {\mathrm {s}} ^ {\mathrm {c}}, \tag {13} +$$ + +$$ +\mathbf {f} _ {\mathrm {t}} ^ {\prime} = \mathcal {A} _ {i} \left(\mathbf {f} _ {\mathrm {t}}\right) = \mathbf {f} ^ {\prime} + \mathbf {w} _ {\mathrm {t}} \cdot \mathbf {f} _ {\mathrm {t}} ^ {\mathrm {c}}, \tag {14} +$$ + +where $\mathbf{f}_{\mathrm{s}}^{\prime}$ and $\mathbf{f}_{\mathrm{t}}^{\prime}$ are output features of $\mathcal{A}_i$ for the source and target domain. Since $\mathbf{w}_{\mathrm{s}}$ and $\mathbf{w}_{\mathrm{t}}$ are involved in the computation of $\mathbf{f}_{\mathrm{s}}^{\prime}$ and $\mathbf{f}_{\mathrm{t}}^{\prime}$ , SEEN modules $\mathcal{S}^{\mathrm{s}}$ and $\mathcal{S}^{\mathrm{t}}$ will be optimized by the classification losses from the two domains respectively, thereby learning domain-specific semantic information. + +# 3.3. Optimization Objective + +We introduce multiple domain-aware attention modules into the frozen visual encoder as learnable attachments. To this end, we expect them to capture visual features suitable for cross-domain adaptation and supplement domain-specific semantic information. First, the domain discriminator $\mathcal{D}$ is applied to the inter-domain attention branch to learn domain-invariant features. Formally, we minimize the adversarial loss over the output of the inter-domain attention branch $\mathbf{f}'$ for the source image $x_{\mathrm{s}}$ and the target image $x_{\mathrm{t}}$ : + +$$ +\mathcal {L} _ {\mathrm {a d v}} = - \sum_ {i = 1} ^ {N} \left[ \mathbb {E} _ {x _ {\mathrm {s}}} \left\| \mathcal {D} _ {i} \left(\mathbf {f} ^ {\prime}\right) \right\| _ {2} ^ {2} + \mathbb {E} _ {x _ {\mathrm {t}}} \left\| \mathcal {D} _ {i} \left(\mathbf {f} ^ {\prime}\right) - \mathbf {1} \right\| _ {2} ^ {2} \right], \tag {15} +$$ + +Second, the detection loss is calculated over both domains to learn the semantic commonalities and disparities. For the source domain, we minimize the cross-entropy $\mathcal{L}_{\mathrm{s}}$ between the logits and ground truth. For the target domain, we use the prompt "A photo of [Class] in foggy day" to generate high-confidence pseudo labels and minimize the cross-entropy $\mathcal{L}_{\mathrm{t}}$ . Overall, the optimization objective is: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {s}} + \lambda_ {\mathrm {t}} \mathcal {L} _ {\mathrm {t}} + \lambda_ {\mathrm {a d v}} \mathcal {L} _ {\mathrm {a d v}} + \mathcal {L} _ {\mathrm {r e g}}, \tag {16} +$$ + +where $\mathcal{L}_{\mathrm{reg}}$ is the regression loss, and $\lambda_{\mathrm{t}}, \lambda_{\mathrm{adv}}$ are balance ratios. + +Table 1. Comparison (%) with existing methods on Cross-Weather adaptation Cityscapes $\rightarrow$ Foggy Cityscapes (C→F), Cross-Fov adaptation KITTI $\rightarrow$ Cityscapes (K→C) and Sim-to-Real adaptation SIM10K $\rightarrow$ Cityscapes (S→C). + +
MethodsVenueC→FK→CS→C
PersonRiderCarTruckBusTrainMotorBicyclemAPmAPmAP
DA-Faster [3]CVPR 201829.240.443.419.738.328.523.732.732.041.938.2
VDD [59]ICCV 202133.444.051.733.952.034.734.236.840.0--
SCAN [28]AAAI 202241.743.957.328.748.648.731.037.342.145.852.6
TIA [64]CVPR 202252.138.149.737.734.846.348.631.142.344.0-
LRA [42]TNNLS 202445.647.159.731.252.444.628.139.543.549.455.7
SIGMA++ [30]TPAMI 202346.445.161.032.152.244.634.839.944.549.557.7
CIGAR [36]CVPR 202346.147.362.127.856.644.333.741.344.948.558.5
OADA [61]ECCV 202247.846.562.932.148.550.934.339.845.447.859.2
MTM [57]AAAI 202451.053.467.237.254.441.638.447.748.9-58.1
AT [31]CVPR 202256.351.964.238.545.555.154.335.050.9--
SOCKER [5]ACMMM 202451.757.768.638.251.647.541.651.751.1-63.8
DSD-DA [11]ICML 202449.059.665.335.761.046.543.957.352.349.352.5
CAT [22]CVPR 202444.657.163.740.866.049.744.953.052.5--
NSA-UDA [66]ICCV 202350.360.167.737.457.446.947.354.352.755.656.3
REACT [24]TIP 202451.457.967.437.758.452.844.654.653.159.558.6
DA-Pro [23]NeurIPS 202355.462.970.940.363.454.042.358.055.961.462.9
RegionCLIP [65]CVPR 202249.655.063.234.155.648.336.047.048.659.158.9
SEEN-DA (Ours)-58.564.571.742.061.254.847.159.957.567.166.8
+ +Table 2. Comparison (%) with existing methods on Cross-Style adaptation task Pascal VOC→Clipart + +
MethodsAeroBikeBirdBoatBottleBusCarCatChairCowTableDogHorseMotorPersonPlantSheepSofaTrainTvmAP
UaDAN [17]35.073.741.024.421.369.853.52.334.261.231.029.547.963.662.261.313.97.648.623.940.2
TFD [54]27.964.828.429.525.764.247.713.547.550.950.821.333.960.265.642.515.140.545.548.641.2
FGRR [1]30.852.135.132.442.262.842.621.442.858.633.520.837.281.466.250.321.529.358.247.043.3
UMT [6]39.659.132.435.045.161.948.47.546.067.621.429.548.275.970.556.725.928.939.443.644.1
SIGMA [29]40.155.437.431.154.954.346.623.044.765.623.022.042.855.667.255.232.940.845.058.644.5
ATMT [27]37.563.437.929.845.162.741.219.543.757.422.925.339.687.170.950.629.132.258.450.545.2
CIGAR [36]35.255.039.230.760.158.146.931.847.061.021.826.744.652.468.554.431.338.856.563.546.2
TIA [64]42.266.036.937.343.771.849.718.244.958.918.229.140.787.867.449.727.427.857.150.646.3
SIGMA++ [30]36.354.640.131.658.060.446.233.644.466.225.725.344.458.864.855.436.238.654.159.346.7
CMT [51]39.856.338.739.760.435.056.07.160.160.435.828.167.884.580.155.520.332.842.338.247.0
RegionCLIP [65]38.170.448.837.344.855.843.514.448.247.814.318.358.378.467.922.230.116.948.450.242.7
SEEN-DA(Ours)44.173.454.747.145.176.051.620.451.753.018.517.361.886.872.222.837.721.158.952.747.9
+ +# 4. Experiment + +We evaluate our method on mainstream DAOD scenarios, including the Cross-Weather, Cross-Fov, Sim-to-Real and Cross-Style. Further ablation studies are conducted to validate the effectiveness of the proposed domain-aware attention module. + +# 4.1. Datasets + +Cross-Weather Cityscapes [4] contains diverse street scenes in daylight, consisting of 2,975 training and 500 validation images annotated with eight classes. Foggy Cityscapes [48] simulates three distinct densities of fog on Cityscapes, containing 8,925 training images and 1,500 validation images. A standard configuration for cross-weather adaptation is to take the training set of Cityscapes as the source domain and the training set of foggy Cityscapes as the target domain, evaluating cross-weather adaptation per + +formance on the 1500-sized validation set in all eight categories. + +Cross-FoV KITTI [14] is a crucial dataset for self-driving that includes 7,481 photos annotated with cars. Collected by driving in rural areas and on highways, it provides data with a different Field of View (FoV). To fairly compare with other methods, we migrate KITTI to Cityscapes solely on the car category. + +Sim-to-Real The synthetic dataset SIM10k [21] has 10,000 photos from the GTA V video game with bounding boxes of the class car. We follow existing works to perform sim-to-real adaptation and report the performance on class car. + +Cross-Style Pascal VOC [9] dataset consists of 16551 images from the real world and has 20 categories. Clipart [20] contains 1000 comical images and has the same 20 categories as Pascal VOC. Following the mainstream splitting, we use Pascal VOC as the source domain and all Clipart images as the target domain. + +Table 3. Comparison $(\%)$ of standard attention with domain-aware attention. + +
BenchmarkSelf-AttentionCross-AttentionDomain-aware Attention
Cross-Weather54.855.557.5
Cross-FoV61.963.567.1
Sim-to-Real62.363.766.8
Cross-Style43.145.247.9
+ +Table 4. Ablation studies (%) of domain-aware attention on Cross-Weather adaptation. + +
Inter-domain AttentionLadvIntra-domain AttentionmAPGains
52.6-
✓54.9+2.3
✓✓55.8+0.9
✓✓✓57.5+1.7
+ +# 4.2. Implementation Details + +Following [23], we adapt RegionCLIP(ResNet-50 [18]) with the Faster-RCNN [45] as the detector. In each iteration, one batch of source images with ground truth and one batch of target domain images with pseudo labels are forwarded to calculate the classification, adversarial and regression loss. The hyperparameter $\lambda_{adv}$ , $\lambda_t$ is set to 0.1 and 1.0. We set the batch size of each domain to 8 and use the SGD optimizer with a warm-up learning rate. Mean Average Precision (mAP) with a threshold of 0.5 is taken as the evaluation metric. All experiments are deployed on 8 Tesla V100 GPUs. + +# 4.3. Comparison to SOTA methods + +We present representative state-of-the-art DAOD approaches for comparison, including feature alignment, semi-supervised learning and VLM-based domain alignment methods. + +Cross-Weather Adaptation Scenario As shown in Table 1 (C→F), the proposed SEEN-DA outperforms all compared methods in terms of mAP and advances SOTA by $1.6\%$ , from $55.9\%$ to $57.5\%$ . Specifically, our method improves performance over six categories (i.e. person, rider, car, truck, train, and bicycle) ranging from $0.8\%$ to $3.1\%$ . + +Cross-FOV Adaptation Scenario Table $1(\mathrm{K}\rightarrow \mathrm{C})$ reports result for KITTI $\rightarrow$ Cityscapes. SEEN-DA achieves SOTA performance of $67.1\%$ mAP, gaining an improvement of $5.7\%$ . + +Sim-to-Real Adaptation Scenario Table 1 (S→C) shows that the proposed method achieves the best results of $66.8\%$ mAP, outperforming the previous best entry SOCCER [5] $63.8\%$ with $3.0\%$ . + +Cross-Style Adaptation Scenario Additionally, we assess SEEN-DA on the more challenging Cross-Style adaptation, where the semantic hierarchy has broader discrepancies. In Table 2, SEEN-DA peaks with $47.9\%$ mAP and improves six categories (aeroplane, bird, boat, bus, sheep and train). + +Table 5. Comparison $(\%)$ of computational efficiency on Cross-Weather adaptation + +
MethodBackbone Parm(M) LearnableParm(M) mAPAbs. Gains
DSS [56]26.83426.83440.9+4.2
CSDA [13]33.64533.64545.3+6.9
REACT [24]29.81229.81248.8+7.9
AT [31]39.22518.72350.9+7.9
DA-Pro [23]34.8340.00855.9+3.3
SEEN-DA (Ours)36.7011.87557.5+4.9
+ +# 4.4. Ablation Studies + +We report detailed ablation studies (Table 3, 4, 5, Figure 4, 5) to validate the effectiveness of proposed SEEN. + +Standard Attention vs. Domain-aware Attention We compare the performance of domain-aware attention with existing attention methods, including self-attention and cross-attention, as shown in Table 3. With the guidance of textual information, the cross-attention improves the self-attention by $0.7\% \sim 2.1\%$ . Furthermore, the proposed domain-aware attention gains improvements of $2.0\% \sim 33.6\%$ over cross-attention, showing a superior ability to select features with high transferability. + +Ablation for Domain-aware Attention Module We conduct ablation experiments to further analyze the significance of proposed domain-aware attention module in Table 4. Introducing the inter-domain attention branch attains an mAP of $54.9\%$ , and optimizing with adversarial loss $\mathcal{L}_{adv}$ gains $0.9\%$ . Meanwhile, applying the intra-domain attention branch improves $1.7\%$ on the target domain. These results showing the effectiveness of the domain-aware attention module to eliminate redundant information and supplement domain-specific semantic information. + +Computational Efficiency We discussed the computational efficiency of the proposed SEEN-DA in Table 5. Our method yields the same parameters scale as existing methods and achieves the highest mAP. In addition, SEEN-DA only needs to train very few parameters (1.875M), showing remarkable efficiency. + +Numbers of the Attention Module We explore the impact of numbers of domain-aware attention modules in Figure 4. Starting from the last visual block, gradually introducing domain-aware attention modules to each visual block improves performance and peaks when using 3 modules. Considering efficiency, we apply 3 domain-aware attention modules into the visual encoder. + +Projection layer in SEEN module We explore the projection layer in the SEEN module in Figure 5. Projecting textual embedding into visual space, denoted as T2V, only achieves $55.1\%$ . Projecting textual and visual embedding into a latent space, denoted as T&V, improves performance by $2.1\%$ . And projecting visual embedding into textual space achieves the best mAP of $57.5\%$ . We analyze the reason for these results as follows. Visual embedding contains more sample noise and redundant information than textual + +![](images/bc7431c25c8a57c1d07ad9d0f616a3cc011ca3de1ff79b778e19ea897bdcd09c.jpg) +(a) Ground Truth + +![](images/1833aa968d6a00cad5eb5261add9b84402964fb83850fe591836b5a4946de3e6.jpg) +(b) DA-Pro[23] + +![](images/506621036038f9f185e7139f7d1642ddf01c16005aa4194f907503865a83c744.jpg) +(c) SEEN-DA (Ours) + +![](images/4b3c8b2a75c6de307c143227ffe30e431bf2ba93b8eaffb23228d66f5128065b.jpg) +Figure 4. Ablation $(\%)$ on the number of attention modules on Cross-Weather adaptation. + +embedding. And learning to project visual embedding to textual space can reduce the interference of these factors. Therefore, we adopt the V2T projection in SEEN module. + +# 4.5. Visualization + +The detection results of ground truth (a), DA-Pro [23] (b) and the proposed SEEN-DA (c) on the target domain of Cross-Weather adaptation are displayed in Figure 3. The figure shows that SEEN-DA outperforms the SOTA method in terms of the ability to detect objects. Specifically, SEEN-DA detects the objects missed (e.g. the rider on the right of figures in 1st row and the person on the left of figures in 2nd row) and misclassified (e.g. the bicycle on the middle of figures in 1st row and the person on the left of figures in 2nd row), revealing the effectiveness of domain-aware attention. By utilizing semantic entropy to reduce redundant information and supplement domain-specific semantic information, the SEEN-DA improves the discriminability of the detector on the target domain. + +# 5. Conclusion + +In this paper, we propose a novel SEMantic ENtropy guided Domain-aware Attention (SEEN-DA) for DAOD. + +![](images/f500264c46a79bd25ef6898c1d2cfd1ca571d241efcfbded1da13f5c4fff49ea.jpg) +Figure 3. Detection comparison on the Cross-Weather adaptation scenario. We visualize (a) the ground truth, the detection boxes of (b) SOTA DA-Pro [23] and (c) our methods SEEN-DA. +Figure 5. Ablation $(\%)$ on projection layer of the domain-aware attention module on Cross-Weather adaptation. + +We design the semantic entropy to quantify the semantic information contained in visual features. Guided by semantic entropy, the domain-aware attention module is introduced to adaptively refine visual features with the semantic commonalities and disparities between domains. Precisely, it consists of an inter-domain attention branch for extracting domain-invariant features and eliminating redundant information between domains, and an intra-domain attention branch for supplementing the domain-specific semantic information for each domain. Extensive experiments over multiple benchmarks validate the effectiveness of SEENDA in inferring a discriminative detector. + +Acknowledgment This work is partially supported by the National Key R&D Program of China(under Grant 2023YFB4502200), the NSF of China(under Grants 92364202, U22A2028, 62102398, 62102399, 62302478, 62302482, 62376268), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660300, XDB0660301, XDB0660302), Major Program of ISCAS(Grant No. ISCAS-ZD-202402), Beijing Natural Science Foundation(4222039), CAS Project for Young Scientists in Basic Research(YSBR-029) and Youth Innovation Promotion Association CAS. + +# References + +[1] Chaoqi Chen, Jiongcheng Li, Hong-Yu Zhou, Xiaoguang Han, Yue Huang, Xinghao Ding, and Yizhou Yu. Relation matters: foreground-aware graph-based relational reasoning for domain adaptive object detection. TPAMI, 45(03):3677-3694, 2023. 6 +[2] Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xin chao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, et al. Learning domain adaptive object detection with probabilistic teacher. In ICML, pages 3040-3055, 2022. 3 +[3] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, pages 3339-3348, 2018. 1, 3, 6 +[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6 +[5] Yiming Cui, Liang Li, Jiehua Zhang, Chenggang Yan, Hongkui Wang, Shuai Wang, Heng Jin, and Li Wu. Stochastic context consistency reasoning for domain adaptive object detection. In ACMMM, pages 1331-1340, 2024. 6, 7 +[6] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In CVPR, pages 4091-4101, 2021. 3, 6 +[7] Jinhong Deng, Dongli Xu, Wen Li, and Lixin Duan. Harmonious teacher for cross-domain object detection. In CVPR, pages 23829-23838, 2023. 3 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[9] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6 +[10] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul De Charette. Poda: Prompt-driven zero-shot domain adaptation. In ICCV, pages 18623-18633, 2023. 3 +[11] Yongchao Feng, Shiwei Li, Yingjie Gao, Ziyue Huang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Dsd-da: Distillation-based source debiasing for domain adaptive object detection. In ICML, 2024. 6 +[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, 17(1):2096-2030, 2016. 3 +[13] Changlong Gao, Chengxu Liu, Yujie Dun, and Xueming Qian. Csda: Learning category-scale joint feature for domain adaptive object detection. In ICCV, pages 11421-11430, 2023. 7 + +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6 +[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3 +[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3 +[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu, and Yanpeng Cao. Uncertainty-aware unsupervised domain adaptation in object detection. TMM, 24:2502-2514, 2021. 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7 +[19] Mengzhe He, Yali Wang, Jiaxi Wu, Yiru Wang, Hanqing Li, Bo Li, Weihao Gan, Wei Wu, and Yu Qiao. Cross domain object detection by target-perceived dual branch distillation. In CVPR, pages 9570-9580, 2022. 3 +[20] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, pages 5001-5009, 2018. 6 +[21] Matthew Johnson-Roberson, Charles Barto, Rounak Mehta, Sharath Nittur Sridhar, Karl Rosaen, and Ram Vasudevan. Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks? In ICRA 2017, pages 746-753. IEEE, 2017. 6 +[22] Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T Tan. Cat: Exploiting inter-class dynamics for domain adaptive object detection. In CVPR, pages 16541-16550, 2024. 6 +[23] Haochen Li, Rui Zhang, Hantao Yao, Xinkai Song, Yifan Hao, Yongwei Zhao, Ling Li, and Yunji Chen. Learning domain-aware detection head with prompt tuning. NeurIPS, 36, 2024. 1, 2, 3, 5, 6, 7, 8 +[24] Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, and Ling Li. React: Remainder adaptive compensation for domain adaptive object detection. TIP, 33:3735-3748, 2024. 1, 3, 6, 7 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3 +[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 3 +[27] Kai Li, Curtis Wigington, Chris Tensmeyer, Vlad I. Morariu, Handong Zhao, Varun Manjunatha, Nikolaos Barmpalios, and Yun Fu. Improving cross-domain detection with self-supervised learning. In CVPR, pages 4745-4754, 2023. 6 +[28] Wuyang Li, Xinyu Liu, Xiwen Yao, and Yixuan Yuan. Scan: Cross domain object detection with semantic conditioned adaptation. In AAAI, volume 6, page 7, 2022. 6 +[29] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object detection. In CVPR, pages 5291-5300, 2022. 3, 6 + +[30] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI, 45(07):9022-9040, 2023. 1, 3, 6 +[31] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Va-jda. Cross-domain adaptive teacher for object detection. In CVPR, pages 7581-7590, 2022. 3, 6, 7 +[32] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1, 3 +[33] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1 +[34] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. TMM, 25:1333-1344, 2022. 3 +[35] Rui Liu, Yahong Han, Yaowei Wang, and Qi Tian. Frequency spectrum augmentation consistency for domain adaptive object detection. arXiv preprint arXiv:2112.08605, 2021. 3 +[36] Yabo Liu, Jinghua Wang, Chao Huang, Yaowei Wang, and Yong Xu. Cigar: Cross-modality graph reasoning for domain adaptive object detection. In CVPR, pages 23776-23786, 2023. 6 +[37] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 3 +[38] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. NeurIPS, 31, 2018. 3 +[39] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, pages 2208-2217, 2017. 3 +[40] Fabio Maria Carlucci, Lorenzo Porzi, Barbara Caputo, Elisa Ricci, and Samuel Rota Bulo. Autodial: Automatic domain alignment layers. In ICCV, pages 5067-5075, 2017. 3 +[41] Pietro Morerio, Jacopo Cavazza, and Vittorio Murino. Minimal-entropy correlation alignment for unsupervised deep domain adaptation. In ICLR, 2018. 3 +[42] Zhengquan Piao, Linbo Tang, and Baojun Zhao. Unsupervised domain-adaptive object detection via localization regression alignment. TNNLS, 35(11):15170-15181, 2024. 6 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3 +[44] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1 +[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 3, 7 + +[46] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In ICCV, pages 8050-8058, 2019. 3 +[47] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In CVPR, pages 6956–6965, 2019. 1, 3 +[48] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. IJCV, 126(9):973-992, Sep 2018. 6 +[49] Sunandini Sanyal, Ashish Ramayee Asokan, Suvaansh Bhambri, Akshay Kulkarni, Jogendra Nath Kundu, and R Venkatesh Babu. Domain-specificity inducing transformers for source-free domain adaptation. In ICCV, pages 18928-18937, 2023. 1, 3 +[50] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NPIS, 30, 2017. 3 +[51] VS Vibashan, Poojan Oza, and Vishal M Patel. Instance relation graph guided source-free domain adaptive object detection. In CVPR, pages 3520-3530, 2023. 6 +[52] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 3 +[53] Vibashan Vs, Vikram Gupta, Poojan Oza, Vishwanath A Sindagi, and Vishal M Patel. Mega-cda: Memory guided attention for category-aware unsupervised domain adaptive object detection. In CVPR, pages 4516-4526, 2021. 3 +[54] Haoan Wang, Shilong Jia, Tieyong Zeng, Guixu Zhang, and Zhi Li. Triple feature disentanglement for one-stage adaptive object detection. In AAAI, pages 5401-5409, 2024. 3, 6 +[55] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, volume 33, pages 5345-5352, 2019. 3 +[56] Yu Wang, Rui Zhang, Shuo Zhang, Miao Li, Yangyang Xia, XiShan Zhang, and ShaoLi Liu. Domain-specific suppression for adaptive object detection. In CVPR, pages 9603-9612, 2021. 7 +[57] Weixi Weng and Chun Yuan. Mean teacher detr with masked feature alignment: A robust domain adaptive detection transformer framework. In AAAI, pages 5912-5920, 2024. 6 +[58] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. TPAMI, 44(8):4178-4193, 2021. 1, 3 +[59] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, pages 9342-9351, 2021. 6 +[60] Xiaofu Wu, Suofei Zhang, Quan Zhou, Zhen Yang, Chunming Zhao, and Longin Jan Latecki. Entropy minimization versus diversity maximization for domain adaptation. TNNLS, 34(6):2896-2907, 2021. 3 +[61] Jayeon Yoo, Inseop Chung, and Nojun Kwak. Unsupervised domain adaptation for one-stage object detector using offsets to bounding box. In ECCV, pages 691-708, 2022. 6 +[62] Yifan Zhang, Hanbo Chen, Ying Wei, Peilin Zhao, Jiezhang Cao, Xinjuan Fan, Xiaoying Lou, Hailing Liu, Jinlong Hou, + +Xiao Han, et al. From whole slide imaging to microscopy: Deep microscopy adaptation network for histopathology cancer image classification. In MICCAI, pages 360-368, 2019. 3 +[63] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, pages 12425-12434, 2021. 3 +[64] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In CVPR, pages 14217-14226, 2022. 6 +[65] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 1, 3, 6 +[66] Wenzhang Zhou, Heng Fan, Tiejian Luo, and Libo Zhang. Unsupervised domain adaptive detection with network stability analysis. In ICCV, pages 6986-6995, 2023. 6 \ No newline at end of file diff --git a/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/images.zip b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..29bbf577fe725721070a563600fd1ed2b0e9ca0f --- /dev/null +++ b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6b4ea81ff958aed84f9cb65aa79cf515ca41eaa4d7b919c0f81e165d04e97a6 +size 624769 diff --git a/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/layout.json b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e6d8245ef588fd86fa5e859e82e7a5174d27a4a --- /dev/null +++ b/2025/SEEN-DA_ SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection/layout.json @@ -0,0 +1,10415 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "text", + "content": "SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "content": "Haochen Li" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "content": " Rui Zhang" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "content": " Hantao Yao" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "content": " Xin Zhang" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "content": " Yifan Hao" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "text", + "content": " Xinkai Song" + }, + { + "bbox": [ + 85, + 160, + 507, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "text", + "content": "Shaohui Peng" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "text", + "content": " Yongwei Zhao" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "text", + "content": " Chen Zhao" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "text", + "content": " Yanjun Wu" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "text", + "content": " Ling Li" + }, + { + "bbox": [ + 115, + 175, + 482, + 189 + ], + "type": "inline_equation", + "content": "^{1,4*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 189, + 455, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 189, + 455, + 202 + ], + "spans": [ + { + "bbox": [ + 139, + 189, + 455, + 202 + ], + "type": "text", + "content": "1Intelligent Software Research Center, Institute of Software, CAS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 203, + 469, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 203, + 469, + 217 + ], + "spans": [ + { + "bbox": [ + 124, + 203, + 469, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 203, + 469, + 217 + ], + "type": "text", + "content": "State Key Lab of Processors, Institute of Computing Technology, CAS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 217, + 531, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 217, + 531, + 231 + ], + "spans": [ + { + "bbox": [ + 63, + 217, + 531, + 231 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 217, + 531, + 231 + ], + "type": "text", + "content": " School of Information Science and Technology, University of Science and Technology of China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 232, + 482, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 232, + 482, + 244 + ], + "spans": [ + { + "bbox": [ + 111, + 232, + 482, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 232, + 482, + 244 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 111, + 232, + 482, + 244 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 111, + 232, + 482, + 244 + ], + "type": "text", + "content": " Institute of Software, CAS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 118, + 247, + 474, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 247, + 474, + 258 + ], + "spans": [ + { + "bbox": [ + 118, + 247, + 474, + 258 + ], + "type": "text", + "content": "{haochen2021, pengshaohui, zhaochen, wuyanjun, liling}@iscas.ac.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 262, + 533, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 262, + 533, + 273 + ], + "spans": [ + { + "bbox": [ + 59, + 262, + 533, + 273 + ], + "type": "text", + "content": "{zhangrui, zhangxin, haoyifan, songxinkai, zhaoyongwei}@ict.ac.cn, yaohantao@ustc.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 285, + 192, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 285, + 192, + 298 + ], + "spans": [ + { + "bbox": [ + 143, + 285, + 192, + 298 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 310, + 289, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 289, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 289, + 634 + ], + "type": "text", + "content": "Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain to an unlabelled target domain. Traditional works focus on aligning visual features between domains to extract domain-invariant knowledge, and recent VLM-based DAOD methods leverage semantic information provided by the textual encoder to supplement domain-specific features for each domain. However, they overlook the role of semantic information in guiding the learning of visual features that are beneficial for adaptation. To solve the problem, we propose semantic entropy to quantify the semantic information contained in visual features, and design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to adaptively refine visual features with the semantic information of two domains. Semantic entropy reflects the importance of features based on semantic information, which can serve as attention to select discriminative visual features and suppress semantically irrelevant redundant information. Guided by semantic entropy, we introduce domain-aware attention modules into the visual encoder in SEEN-DA. It utilizes an inter-domain attention branch to extract domain-invariant features and eliminate redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. Comprehensive experiments validate the effectiveness of SEEN-DA, demonstrating significant improvements in cross-domain object detection performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 649, + 128, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 128, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 128, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "content": "Object detection [8, 33, 44, 45] have been proven to attain encouraging performance assisted by high-quality im" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 285, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 545, + 357 + ], + "type": "text", + "content": "ages with accurate annotations. However, due to the noticeable discrepancy between domains, they suffer severe performance degradation when dealing with unseen data in distinct scenes. To alleviate this problem, domain adaptive object detection (DAOD) [3] aims to transfer a detector trained on the labelled source domain to a unlabelled target domain." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 361, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 624 + ], + "type": "text", + "content": "Traditional DAOD works [3, 30, 47] focus on reducing domain discrepancy by aligning visual features between domains to extract domain-invariant knowledge. Additionally, [24,32,49,58] decouple the unique features of each domain, i.e.domain-specific features, to prevent its interference during feature alignment. Although effective, they train the detector with domain-shared and semantically agnostic category labels (such as one-hot encoding), which ignores the linguistic properties of categories and inter-domain differences and commonality, i.esemantic information. Recently, visual-language models (VLMs) [26,43,65] are proposed to ground visual with textual modalities. Through the textual encoder, they align visual features with the textual prompt, which is constructed from natural language and contains rich semantic information. Motivated by this, recent VLM-based DAOD methods, such as DA-Pro [23], utilize the textual encoder as detection head and tune the textual prompt to learn domain-invariant semantics while supplementing domain-specific semantics for each domain. However, they freeze the visual encoder and leverage semantic information solely in the detection head, limiting the discriminability of visual features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Summarily, despite promising performance, existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. They typically optimize the visual encoder with semantically agnostic category labels and adversarial loss, or directly freeze it, resulting in two limitations. First, they fail to filter out redundant information from visual fea" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25465" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 208, + 115 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 208, + 115 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 208, + 115 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 208, + 115 + ], + "type": "image", + "image_path": "a212bc74c96c4ce6cb01b6d07cad6562cc92ff3bc4ba1abe480c28abb9cf1fe1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 116, + 179, + 123 + ], + "lines": [ + { + "bbox": [ + 74, + 116, + 179, + 123 + ], + "spans": [ + { + "bbox": [ + 74, + 116, + 179, + 123 + ], + "type": "text", + "content": "(a.1) Source Image and Bounding Box" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 124, + 208, + 168 + ], + "blocks": [ + { + "bbox": [ + 51, + 124, + 208, + 168 + ], + "lines": [ + { + "bbox": [ + 51, + 124, + 208, + 168 + ], + "spans": [ + { + "bbox": [ + 51, + 124, + 208, + 168 + ], + "type": "image", + "image_path": "908a4ff3771a5e61c585974b055e63d1c23ab40ee7f3fd10a995a26ff50b58b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 168, + 205, + 176 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 205, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 205, + 176 + ], + "type": "text", + "content": "(b.1) Detection Results on Source Image (DA-Pro[23])" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 177, + 208, + 221 + ], + "blocks": [ + { + "bbox": [ + 51, + 177, + 208, + 221 + ], + "lines": [ + { + "bbox": [ + 51, + 177, + 208, + 221 + ], + "spans": [ + { + "bbox": [ + 51, + 177, + 208, + 221 + ], + "type": "image", + "image_path": "293a8e9c282983095cb06ea627fd148bb8b416015d3311a80b0c4e1f5ddea2be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 222, + 213, + 230 + ], + "lines": [ + { + "bbox": [ + 48, + 222, + 213, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 213, + 230 + ], + "type": "text", + "content": "(c.1) Detection Results on Source Image (SEEN-DA (Ours))" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 237, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 547, + 304 + ], + "type": "text", + "content": "Figure 1. Visualization of the detection results and attention map. (a) Source and Target domain image and bounding boxes. (b) Existing methods ignore the incorporation of semantic information to guide the learning of visual features that are beneficial for adaptation. Infected by the redundant information and lack of domain-specific semantic information, some objects that are correctly detected on the source image are misclassified or missed on the target image. (c) We propose semantic entropy to evaluate the importance of visual features based on domain and category semantic information. By integrating it as domain-aware attention in SEEN-DA, we eliminate redundant information and enhance domain-specific features, thereby improving the detector's discriminability for DAOD." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 296, + 115 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 296, + 115 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 296, + 115 + ], + "type": "image", + "image_path": "4a3d9277509ec7243cc9772c15b82d663d5bb3274b05a7ec3d00338cf4973a01.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 124, + 296, + 168 + ], + "blocks": [ + { + "bbox": [ + 242, + 116, + 348, + 124 + ], + "lines": [ + { + "bbox": [ + 242, + 116, + 348, + 124 + ], + "spans": [ + { + "bbox": [ + 242, + 116, + 348, + 124 + ], + "type": "text", + "content": "(a.2) Target Image and Bounding Box" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 124, + 296, + 168 + ], + "lines": [ + { + "bbox": [ + 217, + 124, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 217, + 124, + 296, + 168 + ], + "type": "image", + "image_path": "9c6518125155277825551b384999203dbf6fb5a77b2d0e08bdbb2758378558d1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 168, + 372, + 176 + ], + "lines": [ + { + "bbox": [ + 223, + 168, + 372, + 176 + ], + "spans": [ + { + "bbox": [ + 223, + 168, + 372, + 176 + ], + "type": "text", + "content": "(b.2) Detection Results on Target Image (DA-Pro[23])" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 217, + 177, + 296, + 221 + ], + "blocks": [ + { + "bbox": [ + 217, + 177, + 296, + 221 + ], + "lines": [ + { + "bbox": [ + 217, + 177, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 217, + 177, + 296, + 221 + ], + "type": "image", + "image_path": "7e417aaf183812e4524c94166600cf237e0240a1354ae410791fb85d5a44ee6e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 222, + 380, + 230 + ], + "lines": [ + { + "bbox": [ + 217, + 222, + 380, + 230 + ], + "spans": [ + { + "bbox": [ + 217, + 222, + 380, + 230 + ], + "type": "text", + "content": "(c.2) Detection Results on Target Image (SEEN-DA (Ours))" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 376, + 115 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 376, + 115 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 376, + 115 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 376, + 115 + ], + "type": "image", + "image_path": "978473a08a87f9f2bbcfd8871dfced3d83d740b65892bf964691f2687470e98c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 298, + 116, + 376, + 168 + ], + "blocks": [ + { + "bbox": [ + 298, + 116, + 376, + 168 + ], + "lines": [ + { + "bbox": [ + 298, + 116, + 376, + 168 + ], + "spans": [ + { + "bbox": [ + 298, + 116, + 376, + 168 + ], + "type": "image", + "image_path": "5865e700f55eaf2b81d9307f6e69af825ad5b1cf9e6393192a387bf7bc6169b3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 168, + 376, + 176 + ], + "lines": [ + { + "bbox": [ + 298, + 168, + 376, + 176 + ], + "spans": [ + { + "bbox": [ + 298, + 168, + 376, + 176 + ], + "type": "text", + "content": "Target Image (DA-Pro[23])" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 298, + 177, + 376, + 221 + ], + "blocks": [ + { + "bbox": [ + 298, + 177, + 376, + 221 + ], + "lines": [ + { + "bbox": [ + 298, + 177, + 376, + 221 + ], + "spans": [ + { + "bbox": [ + 298, + 177, + 376, + 221 + ], + "type": "image", + "image_path": "38303ba9edf02acf570da08b7dc804221edaff239e9f735f2b004943376f7f1e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 298, + 222, + 380, + 230 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 222, + 380, + 230 + ], + "spans": [ + { + "bbox": [ + 298, + 222, + 380, + 230 + ], + "type": "text", + "content": "Target Image (SEEN-DA (Ours))" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 385, + 71, + 545, + 221 + ], + "blocks": [ + { + "bbox": [ + 385, + 71, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 385, + 71, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 385, + 71, + 545, + 221 + ], + "type": "image", + "image_path": "f06f523db07280dc62c5650973862f875db4fa1a8aeff94094a2539d2f5a9e0e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 222, + 538, + 230 + ], + "lines": [ + { + "bbox": [ + 388, + 222, + 538, + 230 + ], + "spans": [ + { + "bbox": [ + 388, + 222, + 538, + 230 + ], + "type": "text", + "content": "(d) Domain-aware Attention Map in SEEN-DA (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 309, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 289, + 538 + ], + "type": "text", + "content": "tures, which is semantically irrelevant and can lead to incorrect feature alignment. Redundant information includes attributes unrelated to object identification, such as color, texture, etc. Although they do not contain domain and category semantic information, they are often extracted by adversarially trained visual encoder since they exist in both domains. Therefore, such semantically irrelevant information may interfere with the alignment of domain-invariant features. In Fig. 1(b.1), existing methods detect the bicycle covered with a white cloth on the source image (clear day). However, since target images (foggy day) tend to be white overall, they misclassify the white cloth as a rider in Fig. 1(b.2). Second, existing methods discard domain-specific features in the visual encoder, which contain rich domain semantic information and are important for the discriminability on the target domain. For example, they correctly detect the bicycle and rider on the source image Fig. 1(b.1), but miss them in Fig. 1(b.2) due to insufficient feature learning of domain-specific properties of target domain, e.g.fog." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": "To solve these problems, we propose semantic entropy which evaluates the importance of features based on domain and category semantic information, and accordingly design SEmantic ENtropy guided Domain-aware Attention (SEEN-DA) to eliminate redundant information and supplement domain-specific features for DAOD. The semantic entropy of visual features refers to the amount of semantic information related to a certain domain. As the textual encoder of VLM provides rich domain and category semantic information in textual embedding, the semantic entropy can be expressed as the information entropy of the classification probability between the visual and textual embedding. Therefore, semantic entropy reflects the importance of visual features in adaptation. Low semantic entropy indicates" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 309, + 546, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 546, + 369 + ], + "type": "text", + "content": "that visual features match the textual embedding of a certain category and are likely to be transferable foreground objects. On the contrary, high semantic entropy means that they are unpredictable and may be redundant or even harmful for detection." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 380, + 546, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 643 + ], + "type": "text", + "content": "Applying semantic entropy as attention, we design the SEEN-DA framework, introducing domain-aware attention modules into the visual encoder to select visual features with low semantic entropy and suppress regions with high ones. Specifically, the domain-aware attention module utilizes two parallel branches to learn inter- and intra-domain attention with semantic entropy. The inter-domain attention branch is shared between domains and evaluates the semantic entropy on domain-shared textual embedding, aiming to learn domain-invariant features while eliminating redundant information. The intra-domain attention branch is private in each domain to evaluate the semantic entropy on source/target textual embedding respectively, supplementing domain-specific features for each domain. The SEEN-DA is able to adaptively adjust visual features with the semantic commonalities and disparities between domains. As shown in Fig. 1(c), domain-aware attention map highlights features with rich semantic information, i.eforeground regions, and hinders the negative impact of background regions. Thereby, the proposed SEEN-DA correctly detects objects misclassified by existing methods, improving the discriminability of the detector on the target domain." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We evaluate the proposed method on four benchmarks. The experimental results highlight the effectiveness of our SEEN-DA, e.g., it reaches " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "57.5\\%" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " mAP on Cityscapes " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " Foggy Cityscapes, surpassing the state-of-the-art DA-Pro [23] by " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25466" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 258 + ], + "type": "text", + "content": "Visual-Language Models Large-scale visual-language models (VLMs) [25, 26, 43] embed visual and text modalities into a shared space, enabling cross-modal alignment. Pre-trained with an astonishing scale of image-text pairs, they demonstrate comprehensive visual understanding. As a pioneering work, CLIP [43] trains a visual encoder and a textual encoder with 400 million image-text pairs in a contrastive manner, showing promising classification ability for the seen and unseen classes. Following this spirit, [10, 16, 52, 65] freeze the textual encoder and adapt the visual encoder into Faster-RCNN [45] as a two-stage detector. In this work, we apply the textual encoder of CLIP [43] to infer textual embedding that contains semantic information, and the visual encoder of RegionCLIP [65] as the backbone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 289, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 289, + 677 + ], + "type": "text", + "content": "Domain Adaptive Object Detection (DAOD) aims to adapt the object detector [45] trained on the labelled source domain to the unlabelled target domain. Previous approaches can be broadly divided into three categories: feature alignment, semi-supervised learning and VLM-based domain alignment. As a fundamental direction, feature alignment [12, 37, 39] aims to align the feature distributions of the two domains to generate domain-invariant features [3, 30, 47]. Typically, the domain discriminator is proposed to align feature distribution adversarially in image level [3], instance level [47, 63] and category level [29, 53]. To reduce the negative impact of aligning unique features of each domain, [24, 32, 34, 49, 54, 58] decouple the domain-invariant and domain-specific knowledge between two domains. In parallel, semi-supervised learning strives to generate unbiased training data with style transfer [6, 19, 35] and pseudo label [2, 7, 31, 50]. However, they train the detector with semantically agnostic category labels (such as one-hot encoding), ignoring the semantic information inherently contained in the category name. To overcome this, recent VLM-based domain alignment methods exploit the semantic information provided by the textual encoder of VLMs to learn semantic commonalities and disparities. Pioneeringly, DA-Pro [23] tuning the textual encoder into a domain-aware detection head, learning domain-invariant semantics while supplementing domain-specific semantics for each domain. Despite the promising performance, existing methods freeze the visual encoder and ignore using semantic information to guide the learning of visual features, which leads to two limitations. First, they fail to filter out redundant information from visual features, which is semantically irrelevant and can lead to incorrect feature alignment. Second, they discard domain-specific features in the visual encoder, which contain rich domain semantic information and are vital for DAOD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Entropy Optimization Information entropy was first utilized as entropy minimization in [15] for semi-supervised learning to learn decision rules for unlabelled data. Con-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 312 + ], + "type": "text", + "content": "sidering its effectiveness, recent works introduce entropy optimization into DAOD. [40, 41, 46, 62] propose minimizing entropy on unlabelled data to make decisions consistent in the source and target domains. [55] apply entropy as metrics to measure the transferability of data samples. [38] employs entropy as attention to control the uncertainty of the classifier predictions, alleviating the effect of negative transfer. Instead of directly minimizing entropy, [60] further introduces diversity maximization to explore close-to-ideal domain adaptation solutions. However, the computation of entropy in existing methods is inherently semantically agnostic. Based on the output probabilities from a classifier, they treat categories merely as identifiers and convert the category labels into one-hot encoding. Therefore, they ignore the rich semantic information embedded within the category names, which is vital for DAOD. In this paper, we propose a novel semantic entropy, evaluating the transferability of visual features through the semantic commonalities and disparities between domains, improving the discrimination of the visual encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 321, + 389, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 321, + 389, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 389, + 335 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "content": "In this section, we present a novel SEmantic entropy guided Domain-aware Attention (SEEN-DA) for DAOD, which introduces domain-aware attention module to eliminate redundant information and supplement domain-specific features, guided by the proposed semantic entropy. Semantic entropy quantifies the amount of semantic information within visual features, providing a measure of their importance in domain adaptation. Utilizing semantic entropy as attention, the domain-aware attention module adopts an inter-domain attention branch to extract domain-invariant features and eliminate semantically irrelevant redundant information, and an intra-domain attention branch to supplement the domain-specific semantic information discriminative on each domain. As shown in Fig. 2(a), given input image " + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "content": ", the visual encoder of SEEN-DA is divided into " + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "content": " frozen visual block " + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "content": ", and each domain-aware attention module " + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{A}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "content": " is attached to " + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_i" + }, + { + "bbox": [ + 304, + 341, + 547, + 556 + ], + "type": "text", + "content": " as a lightweight learnable module:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 358, + 559, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 559, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 358, + 559, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {0} = \\mathcal {E} (x); \\mathbf {h} _ {i} = \\mathcal {A} _ {i} \\left(\\mathcal {V} _ {i} \\left(\\mathbf {h} _ {i - 1}\\right)\\right), \\tag {1}", + "image_path": "9dd8f3aa8016588790109b2c932f8574806706e6ed36b98402ca543e5ec43cad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 574, + 411, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 411, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 411, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 574, + 411, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 306, + 574, + 411, + 586 + ], + "type": "text", + "content": " is the stem layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 594, + 452, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 452, + 606 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 452, + 606 + ], + "type": "text", + "content": "3.1. Semantic Entropy (SEEN)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 612, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 612, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 612, + 545, + 635 + ], + "type": "text", + "content": "Generally speaking, given feature " + }, + { + "bbox": [ + 305, + 612, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 305, + 612, + 545, + 635 + ], + "type": "text", + "content": ", the probability that this feature belongs to a specific object category is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 637, + 545, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 637, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 370, + 637, + 545, + 664 + ], + "type": "interline_equation", + "content": "p (c, \\mathbf {f}) = \\frac {\\exp \\left(v _ {c} (\\mathbf {f})\\right)}{\\sum_ {c} \\exp \\left(v _ {c} (\\mathbf {f})\\right)}, \\tag {2}", + "image_path": "148e34554f05d579d5b7e4bc1c1063e0c41fc2495a9993219241c82fba4e0b3b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "c, v_{c}(\\cdot)" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " represents an object category and its corresponding classifier. " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "p(\\cdot, \\mathbf{f})" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": " is the probability distribution function, expected to output 1 on the corresponding category and 0 on other category, i.e. one-hot vector." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25467" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 318, + 188 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 318, + 188 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 318, + 188 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 318, + 188 + ], + "type": "image", + "image_path": "141f1c0e6581a5a6849f274af53ffb0551d495f1a154bdb2cde23f68b4e371b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 191, + 225, + 200 + ], + "lines": [ + { + "bbox": [ + 132, + 191, + 225, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 191, + 225, + 200 + ], + "type": "text", + "content": "(a) Overview of SEEN-DA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 69, + 538, + 189 + ], + "blocks": [ + { + "bbox": [ + 325, + 69, + 538, + 189 + ], + "lines": [ + { + "bbox": [ + 325, + 69, + 538, + 189 + ], + "spans": [ + { + "bbox": [ + 325, + 69, + 538, + 189 + ], + "type": "image", + "image_path": "ef1b5a8e7dbefcf9007886a15b64d1c7b7196cebe73860b814f709b42e582c8c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 191, + 496, + 200 + ], + "lines": [ + { + "bbox": [ + 368, + 191, + 496, + 200 + ], + "spans": [ + { + "bbox": [ + 368, + 191, + 496, + 200 + ], + "type": "text", + "content": "(b) Domain-aware Attention Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 201, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 545, + 224 + ], + "type": "text", + "content": "Figure 2. (a) Overview of the proposed SEEN-DA for DAOD, where the semantic entropy is utilized as attention in domain-aware attention module. (b) The architecture of domain-aware attention module, consisting of an inter-domain and an intra-domain attention branch." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 227, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 346 + ], + "type": "text", + "content": "Tradition DAOD approaches focus on reducing domain discrepancy by aligning " + }, + { + "bbox": [ + 46, + 227, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 227, + 287, + 346 + ], + "type": "text", + "content": " between the source and target domain. Although effective, as shown in Eq. 2, they treat categories merely as one-hot vectors, which is inherently semantically agnostic. Therefore, they ignore the rich semantic information inherently embedded within the category names. As feature alignment relies heavily on a profound understanding of the semantic disparities between the source and target domains, this limitation hinders the learning of visual features for adaptation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": "To address this, we propose semantic entropy, which quantifies the semantic information embedded in visual features by leveraging textual embeddings from Visual-language models (VLMs). VLMs utilize a textual encoder to encode textual prompts that describe category attributes into textual embeddings, aligning them with visual features. As a result, textual embeddings inherit rich semantic information from natural language, making them effective representations of the semantics of visual features. Motivated by this, we first infer textual embedding " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbb{T} = \\{t_c\\}_c^K" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": " categories by feeding prompts like \"A photo of [Class] in [Domain]\" into a textual encoder, where each " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": " contains semantic information of category " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": " and its domain. Thus, the probability distribution that " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": " corresponds to certain semantics can be estimated by the similarity between the visual feature " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": " and textual embedding " + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 46, + 347, + 287, + 537 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 539, + 287, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 539, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 96, + 539, + 287, + 564 + ], + "type": "interline_equation", + "content": "p \\left(t _ {c}, \\mathbf {f}\\right) = \\frac {\\exp \\left(\\sin \\left(t _ {c} , \\mathbf {f}\\right) / \\tau\\right)}{\\sum_ {c} \\exp \\left(\\sin \\left(t _ {c} , \\mathbf {f}\\right) / \\tau\\right)}, \\tag {3}", + "image_path": "66c6494c7c5918fc12e4d8c100db13048782ab24780552c2ca8446347bc71ced.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": " is the cosine similarity, and " + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": " is a learnable temperature parameter. Though " + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "p(t_{c},\\mathbf{f})" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": " provides a measure of semantic relevance, it does not directly quantify the amount of semantic information embedded in " + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": ". Therefore, we draw inspiration from the concept of information entropy, which is widely used to measure the uncertainty and informativeness of a probability distribution. Formally, the proposed semantic entropy is defined as the information entropy of the probability distribution " + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "p(t_{c},\\mathbf{f})" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": " to evaluate the semantic uncertainty of " + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 566, + 287, + 685 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 691, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 85, + 691, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\operatorname {S E} (\\mathbb {T}, \\mathbf {f}) = - \\sum_ {c} p \\left(t _ {c}, \\mathbf {f}\\right) \\log \\left(p \\left(t _ {c}, \\mathbf {f}\\right)\\right), \\tag {4}", + "image_path": "b85157ed8225d3fcbcfd6a2838029ae262f4813d915e4dfbdea4d320d687e15d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 227, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 346 + ], + "type": "text", + "content": "In the framework of DAOD, we apply semantic entropy as attention to select high discriminative visual features. According to Eq. 4, features with low semantic entropy exhibit clear and definite semantic information, expected to be foreground features suitable for transfer. Conversely, features with high semantic entropy indicate high semantic uncertainty, suggesting they may contain semantically irrelevant redundant information, which should be deemphasized during adaptation. To achieve this, we regulate Eq. 4 to propose semantic entropy-guided attention:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 352, + 545, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 352, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 310, + 352, + 545, + 386 + ], + "type": "interline_equation", + "content": "\\operatorname {S E A t t e n t i o n} (\\mathbb {T}, \\mathbf {f}) = \\sum_ {c} p \\left(t _ {c}, \\mathbf {f}\\right) \\log \\left(p \\left(t _ {c}, \\mathbf {f}\\right)\\right) + \\log K, \\tag {5}", + "image_path": "1bace0a8ac86445936ac232411474a7e9312ce692aac152d26cc3e5e83b274ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 387, + 545, + 434 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 387, + 545, + 434 + ], + "type": "text", + "content": " is the number of categories. Based on semantic entropy, Eq. 5 effectively adjusts the weight of visual features, focusing on features with rich semantic information while minimizing the influence of redundant information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 441, + 484, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 484, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 484, + 451 + ], + "type": "text", + "content": "3.2. Domain-aware Attention Module" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "content": "The domain-aware attention module is designed to adaptively refine visual features of each domain by utilizing the semantic entropy. As shown in Fig. 2(b), the module incorporates two complementary branches: an inter-domain attention branch to learn domain-invariant features and eliminate semantically irrelevant information, and an intra-domain attention branch to supplement the domain-specific features unique to each domain." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": "Inter-domain Attention The inter-domain attention is designed to modify domain-invariant features of the frozen visual encoder with semantic information. Specifically, for the output visual features " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d = \\mathcal{V}_i(\\mathbf{h}_{i - 1})" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " of the i-th visual block, we first forward it to a domain-shared convolution module " + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 558, + 545, + 629 + ], + "type": "text", + "content": " to capture task-related knowledge:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 402, + 635, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 635, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 402, + 635, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {f} = \\mathcal {C} \\left(\\mathbf {f} _ {d}\\right), \\tag {6}", + "image_path": "83c94a39df90a1db9b62c0343092d707d089924cc9b2e129ed3cc53295373a91.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "d \\in \\{\\mathrm{s},\\mathrm{t}\\}" + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "content": " denotes whether " + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 304, + 653, + 545, + 677 + ], + "type": "text", + "content": " belongs to source domain s or target domain t." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Simultaneously, the visual feature " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is forwarded to the SEEN module " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to obtain attention map " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " guided by semantic entropy. With the help of VLM, we can represent the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25468" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "domain-invariant semantic information through domain-shared prompts. Technically, we input textual prompts such as \"A photo of [Class]\" to the textual encoder to retrieve domain-invariant textual embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". As described in Eq. 4, the similarity of visual features and textual embeddings is needed to calculate semantic entropy attention. To facilitate this, a projection layer " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " is firstly introduced to map " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " from the visual feature space to the textual embedding space:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 140, + 181, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 181, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 140, + 181, + 287, + 194 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathrm {p}} = \\mathcal {P} \\left(\\mathbf {f} _ {d}\\right), \\tag {7}", + "image_path": "d44461350a2760bb75f90b9efe7a0efe35e5b34874c9f43586479245441ac176.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "type": "text", + "content": "Then we calculate the semantic entropy-guided attention of the visual feature " + }, + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "type": "text", + "content": " over text embedding " + }, + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 47, + 200, + 287, + 226 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 235, + 287, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 235, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 88, + 235, + 287, + 248 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = \\mathcal {S} (\\mathbb {T}, \\mathbf {f} _ {d}) = \\operatorname {S E A t t e n t i o n} (\\mathbb {T}, \\mathbf {f} _ {\\mathrm {p}}), \\tag {8}", + "image_path": "8361e8ecaf4676069946f9dace1ec15a3e676a4afa43aefb238584f19472380b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "text", + "content": "With the attention map " + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "text", + "content": ", we can focus on semantically informative features and suppress redundant information in domain-invariant features. Typically, " + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 47, + 258, + 287, + 318 + ], + "type": "text", + "content": " via an element-wise product, and the weighted feature is summed with the original visual feature:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 328, + 287, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 328, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 133, + 328, + 287, + 341 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {\\prime} = \\mathbf {w} \\cdot \\mathbf {f} + \\mathbf {f} _ {d}, \\tag {9}", + "image_path": "d9895ade3475917a29f2a2187d8e58c913a6d3d6847678ea0595090f057b2565.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "type": "text", + "content": "Expecting to be domain-invariant, the extracted " + }, + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "type": "text", + "content": " is forwarded to a domain discriminator " + }, + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 351, + 287, + 388 + ], + "type": "text", + "content": ", which aligns the source and target domain adversarily." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "content": "Intra-domain Attention With the inter-domain attention branch, the extracted domain-invariant features contain sufficient semantic information shared between domains. However, it ignores the semantic discrepancy between the two domains, resulting in limited discriminability in the target domain. Therefore, we design the intra-domain attention branch to enrich visual features with domain-specific semantic information. Similar to the inter-domain attention branch, we first introduce a learnable layer " + }, + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^c" + }, + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "content": " to adapt visual features for to both domains:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 533, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 533, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 140, + 533, + 287, + 548 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {d} ^ {\\mathrm {c}} = \\mathcal {C} ^ {\\mathrm {c}} \\left(\\mathbf {f} _ {d}\\right), \\tag {10}", + "image_path": "a12deba9c98c1f35a1248ca4b00d8fb738ee5a3ae4d3c6b772bd58410e2bcc77.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "content": "Subsequently, we generate domain-aware textual embedding that captures the semantics of each domains. Motivated by DA-Pro [23], we tune learnable domain-adaptive prompts, i.e. " + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\left[[v^{\\mathrm{c}}][v^{\\mathrm{s}}]\\right]" + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "content": " [Class] for the source domain and " + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\left[[v^{\\mathrm{c}}][v^{\\mathrm{t}}]\\right]" + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "content": " [Class] for the target domain, to explicitly model the domain and category semantic commonalities and disparities. These prompts are forwarded to the textual encoder to derive domain-adaptive textual embedding " + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_{\\mathrm{s}},\\mathbb{T}_{\\mathrm{t}}" + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "content": " for each domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Due to the visual and semantic differences between the source and target domain, a domain-shared SEEN module is inadequate. To calculate the semantic entropy for each domain, visual features must be independently mapped into" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "their respective textual spaces. Therefore, we apply separate SEEN modules " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": " with independent projection layers " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{s}},\\mathcal{P}^{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": " and textual embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_{\\mathrm{s}},\\mathbb{T}_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 390, + 114, + 545, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 114, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 390, + 114, + 545, + 127 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {\\mathrm {s}} = \\mathcal {S} ^ {\\mathrm {s}} \\left(\\mathbb {T} _ {\\mathrm {s}}, \\mathbf {f} _ {\\mathrm {s}}\\right), \\tag {11}", + "image_path": "5a26a7fb7ef82cd277b851f9ace7fd62de38c201c4213108d1739bfb2466561b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 390, + 140, + 545, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 140, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 390, + 140, + 545, + 153 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {\\mathrm {t}} = \\mathcal {S} ^ {\\mathrm {t}} \\left(\\mathbb {T} _ {\\mathrm {t}}, \\mathbf {f} _ {\\mathrm {t}}\\right), \\tag {12}", + "image_path": "bc1a6774b9f0dcb277403d82fbc855737126686166ecac896e7e6dc318a44d66.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 156, + 545, + 180 + ], + "type": "text", + "content": " denotes the source- and target-specific attention maps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 181, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 251 + ], + "type": "text", + "content": "The inter-domain attention branch extracts domain-invariant features while eliminating redundant information, whereas the intra-domain attention branch enriches visual features with domain-specific semantics. Overall, the output for the i-th domain-aware attention module " + }, + { + "bbox": [ + 304, + 181, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 304, + 181, + 545, + 251 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 368, + 258, + 545, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 258, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 368, + 258, + 545, + 271 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathrm {s}} ^ {\\prime} = \\mathcal {A} _ {i} \\left(\\mathbf {f} _ {\\mathrm {s}}\\right) = \\mathbf {f} ^ {\\prime} + \\mathbf {w} _ {\\mathrm {s}} \\cdot \\mathbf {f} _ {\\mathrm {s}} ^ {\\mathrm {c}}, \\tag {13}", + "image_path": "ac46652659a13a9f08f3a586031c82d69d281a1a528281a79692cb481d1129e3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 285, + 545, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 285, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 369, + 285, + 545, + 298 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\mathrm {t}} ^ {\\prime} = \\mathcal {A} _ {i} \\left(\\mathbf {f} _ {\\mathrm {t}}\\right) = \\mathbf {f} ^ {\\prime} + \\mathbf {w} _ {\\mathrm {t}} \\cdot \\mathbf {f} _ {\\mathrm {t}} ^ {\\mathrm {c}}, \\tag {14}", + "image_path": "01b8cc25ff7f88b772f7ac5d8eff679c26b9366716d1bbb9c8c70029aea7ddd8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{s}}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{t}}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " are output features of " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " for the source and target domain. Since " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " are involved in the computation of " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{s}}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{t}}^{\\prime}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": ", SEEN modules " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": " will be optimized by the classification losses from the two domains respectively, thereby learning domain-specific semantic information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 380, + 440, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 440, + 394 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 440, + 394 + ], + "type": "text", + "content": "3.3. Optimization Objective" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": "We introduce multiple domain-aware attention modules into the frozen visual encoder as learnable attachments. To this end, we expect them to capture visual features suitable for cross-domain adaptation and supplement domain-specific semantic information. First, the domain discriminator " + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": " is applied to the inter-domain attention branch to learn domain-invariant features. Formally, we minimize the adversarial loss over the output of the inter-domain attention branch " + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{f}'" + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": " for the source image " + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": " and the target image " + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 529, + 545, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 529, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 311, + 529, + 545, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} = - \\sum_ {i = 1} ^ {N} \\left[ \\mathbb {E} _ {x _ {\\mathrm {s}}} \\left\\| \\mathcal {D} _ {i} \\left(\\mathbf {f} ^ {\\prime}\\right) \\right\\| _ {2} ^ {2} + \\mathbb {E} _ {x _ {\\mathrm {t}}} \\left\\| \\mathcal {D} _ {i} \\left(\\mathbf {f} ^ {\\prime}\\right) - \\mathbf {1} \\right\\| _ {2} ^ {2} \\right], \\tag {15}", + "image_path": "29526d829ddcd1ba371ad05610557edd213a41d3b95517478e7746cf11493885.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "type": "text", + "content": "Second, the detection loss is calculated over both domains to learn the semantic commonalities and disparities. For the source domain, we minimize the cross-entropy " + }, + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "type": "text", + "content": " between the logits and ground truth. For the target domain, we use the prompt \"A photo of [Class] in foggy day\" to generate high-confidence pseudo labels and minimize the cross-entropy " + }, + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{t}}" + }, + { + "bbox": [ + 304, + 571, + 545, + 656 + ], + "type": "text", + "content": ". Overall, the optimization objective is:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 350, + 666, + 545, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 666, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 350, + 666, + 545, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {t}} \\mathcal {L} _ {\\mathrm {t}} + \\lambda_ {\\mathrm {a d v}} \\mathcal {L} _ {\\mathrm {a d v}} + \\mathcal {L} _ {\\mathrm {r e g}}, \\tag {16}", + "image_path": "4280b6dafde6a79f0689a9326c30eed951a4610b506a7f2ccd733a3bc35c7b61.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": " is the regression loss, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{t}}, \\lambda_{\\mathrm{adv}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": " are balance ratios." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25469" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 97, + 519, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Comparison (%) with existing methods on Cross-Weather adaptation Cityscapes " + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": " Foggy Cityscapes (C→F), Cross-Fov adaptation KITTI " + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": " Cityscapes (K→C) and Sim-to-Real adaptation SIM10K " + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": " Cityscapes (S→C)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 97, + 519, + 312 + ], + "lines": [ + { + "bbox": [ + 73, + 97, + 519, + 312 + ], + "spans": [ + { + "bbox": [ + 73, + 97, + 519, + 312 + ], + "type": "table", + "html": "
MethodsVenueC→FK→CS→C
PersonRiderCarTruckBusTrainMotorBicyclemAPmAPmAP
DA-Faster [3]CVPR 201829.240.443.419.738.328.523.732.732.041.938.2
VDD [59]ICCV 202133.444.051.733.952.034.734.236.840.0--
SCAN [28]AAAI 202241.743.957.328.748.648.731.037.342.145.852.6
TIA [64]CVPR 202252.138.149.737.734.846.348.631.142.344.0-
LRA [42]TNNLS 202445.647.159.731.252.444.628.139.543.549.455.7
SIGMA++ [30]TPAMI 202346.445.161.032.152.244.634.839.944.549.557.7
CIGAR [36]CVPR 202346.147.362.127.856.644.333.741.344.948.558.5
OADA [61]ECCV 202247.846.562.932.148.550.934.339.845.447.859.2
MTM [57]AAAI 202451.053.467.237.254.441.638.447.748.9-58.1
AT [31]CVPR 202256.351.964.238.545.555.154.335.050.9--
SOCKER [5]ACMMM 202451.757.768.638.251.647.541.651.751.1-63.8
DSD-DA [11]ICML 202449.059.665.335.761.046.543.957.352.349.352.5
CAT [22]CVPR 202444.657.163.740.866.049.744.953.052.5--
NSA-UDA [66]ICCV 202350.360.167.737.457.446.947.354.352.755.656.3
REACT [24]TIP 202451.457.967.437.758.452.844.654.653.159.558.6
DA-Pro [23]NeurIPS 202355.462.970.940.363.454.042.358.055.961.462.9
RegionCLIP [65]CVPR 202249.655.063.234.155.648.336.047.048.659.158.9
SEEN-DA (Ours)-58.564.571.742.061.254.847.159.957.567.166.8
", + "image_path": "4ebfcaa68475b041233126fbe52572fd9a9d638665cc7e2c97274a7227be7e72.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 332, + 545, + 487 + ], + "blocks": [ + { + "bbox": [ + 112, + 316, + 481, + 328 + ], + "lines": [ + { + "bbox": [ + 112, + 316, + 481, + 328 + ], + "spans": [ + { + "bbox": [ + 112, + 316, + 481, + 328 + ], + "type": "text", + "content": "Table 2. Comparison (%) with existing methods on Cross-Style adaptation task Pascal VOC→Clipart" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 332, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 50, + 332, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 50, + 332, + 545, + 487 + ], + "type": "table", + "html": "
MethodsAeroBikeBirdBoatBottleBusCarCatChairCowTableDogHorseMotorPersonPlantSheepSofaTrainTvmAP
UaDAN [17]35.073.741.024.421.369.853.52.334.261.231.029.547.963.662.261.313.97.648.623.940.2
TFD [54]27.964.828.429.525.764.247.713.547.550.950.821.333.960.265.642.515.140.545.548.641.2
FGRR [1]30.852.135.132.442.262.842.621.442.858.633.520.837.281.466.250.321.529.358.247.043.3
UMT [6]39.659.132.435.045.161.948.47.546.067.621.429.548.275.970.556.725.928.939.443.644.1
SIGMA [29]40.155.437.431.154.954.346.623.044.765.623.022.042.855.667.255.232.940.845.058.644.5
ATMT [27]37.563.437.929.845.162.741.219.543.757.422.925.339.687.170.950.629.132.258.450.545.2
CIGAR [36]35.255.039.230.760.158.146.931.847.061.021.826.744.652.468.554.431.338.856.563.546.2
TIA [64]42.266.036.937.343.771.849.718.244.958.918.229.140.787.867.449.727.427.857.150.646.3
SIGMA++ [30]36.354.640.131.658.060.446.233.644.466.225.725.344.458.864.855.436.238.654.159.346.7
CMT [51]39.856.338.739.760.435.056.07.160.160.435.828.167.884.580.155.520.332.842.338.247.0
RegionCLIP [65]38.170.448.837.344.855.843.514.448.247.814.318.358.378.467.922.230.116.948.450.242.7
SEEN-DA(Ours)44.173.454.747.145.176.051.620.451.753.018.517.361.886.872.222.837.721.158.952.747.9
", + "image_path": "8e68593b47a0619b95958ebd7a5efb13e8674adbf7f0bd8e568a2a995176deb9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 496, + 123, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 123, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 123, + 510 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 517, + 288, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 577 + ], + "type": "text", + "content": "We evaluate our method on mainstream DAOD scenarios, including the Cross-Weather, Cross-Fov, Sim-to-Real and Cross-Style. Further ablation studies are conducted to validate the effectiveness of the proposed domain-aware attention module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 111, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 111, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 111, + 597 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Cross-Weather Cityscapes [4] contains diverse street scenes in daylight, consisting of 2,975 training and 500 validation images annotated with eight classes. Foggy Cityscapes [48] simulates three distinct densities of fog on Cityscapes, containing 8,925 training images and 1,500 validation images. A standard configuration for cross-weather adaptation is to take the training set of Cityscapes as the source domain and the training set of foggy Cityscapes as the target domain, evaluating cross-weather adaptation per" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 497, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 520 + ], + "type": "text", + "content": "formance on the 1500-sized validation set in all eight categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 593 + ], + "type": "text", + "content": "Cross-FoV KITTI [14] is a crucial dataset for self-driving that includes 7,481 photos annotated with cars. Collected by driving in rural areas and on highways, it provides data with a different Field of View (FoV). To fairly compare with other methods, we migrate KITTI to Cityscapes solely on the car category." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 641 + ], + "type": "text", + "content": "Sim-to-Real The synthetic dataset SIM10k [21] has 10,000 photos from the GTA V video game with bounding boxes of the class car. We follow existing works to perform sim-to-real adaptation and report the performance on class car." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Cross-Style Pascal VOC [9] dataset consists of 16551 images from the real world and has 20 categories. Clipart [20] contains 1000 comical images and has the same 20 categories as Pascal VOC. Following the mainstream splitting, we use Pascal VOC as the source domain and all Clipart images as the target domain." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25470" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 97, + 286, + 156 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": "Table 3. Comparison " + }, + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 47, + 70, + 287, + 92 + ], + "type": "text", + "content": " of standard attention with domain-aware attention." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 97, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 49, + 97, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 49, + 97, + 286, + 156 + ], + "type": "table", + "html": "
BenchmarkSelf-AttentionCross-AttentionDomain-aware Attention
Cross-Weather54.855.557.5
Cross-FoV61.963.567.1
Sim-to-Real62.363.766.8
Cross-Style43.145.247.9
", + "image_path": "65e66d299679b8646f7b11d6a178d412a8d1b13fda48d7b5dc8f85563e735d78.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 189, + 286, + 251 + ], + "blocks": [ + { + "bbox": [ + 47, + 163, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 47, + 163, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 286, + 185 + ], + "type": "text", + "content": "Table 4. Ablation studies (%) of domain-aware attention on Cross-Weather adaptation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 189, + 286, + 251 + ], + "lines": [ + { + "bbox": [ + 49, + 189, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 189, + 286, + 251 + ], + "type": "table", + "html": "
Inter-domain AttentionLadvIntra-domain AttentionmAPGains
52.6-
✓54.9+2.3
✓✓55.8+0.9
✓✓✓57.5+1.7
", + "image_path": "e750792bdeff5dd26eb109a301bf0d795e79460883e6080fb95a025545a379cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 257, + 180, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 180, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 180, + 270 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "type": "text", + "content": "Following [23], we adapt RegionCLIP(ResNet-50 [18]) with the Faster-RCNN [45] as the detector. In each iteration, one batch of source images with ground truth and one batch of target domain images with pseudo labels are forwarded to calculate the classification, adversarial and regression loss. The hyperparameter " + }, + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\lambda_{adv}" + }, + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 46, + 276, + 287, + 407 + ], + "type": "text", + "content": " is set to 0.1 and 1.0. We set the batch size of each domain to 8 and use the SGD optimizer with a warm-up learning rate. Mean Average Precision (mAP) with a threshold of 0.5 is taken as the evaluation metric. All experiments are deployed on 8 Tesla V100 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 416, + 214, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 214, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 214, + 430 + ], + "type": "text", + "content": "4.3. Comparison to SOTA methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 436, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 483 + ], + "type": "text", + "content": "We present representative state-of-the-art DAOD approaches for comparison, including feature alignment, semi-supervised learning and VLM-based domain alignment methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "content": "Cross-Weather Adaptation Scenario As shown in Table 1 (C→F), the proposed SEEN-DA outperforms all compared methods in terms of mAP and advances SOTA by " + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "content": ", from " + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "inline_equation", + "content": "55.9\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "inline_equation", + "content": "57.5\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "content": ". Specifically, our method improves performance over six categories (i.e. person, rider, car, truck, train, and bicycle) ranging from " + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 484, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "content": "Cross-FOV Adaptation Scenario Table " + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "inline_equation", + "content": "1(\\mathrm{K}\\rightarrow \\mathrm{C})" + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "content": " reports result for KITTI " + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "content": " Cityscapes. SEEN-DA achieves SOTA performance of " + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "inline_equation", + "content": "67.1\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "content": " mAP, gaining an improvement of " + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "content": "Sim-to-Real Adaptation Scenario Table 1 (S→C) shows that the proposed method achieves the best results of " + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "inline_equation", + "content": "66.8\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "content": " mAP, outperforming the previous best entry SOCCER [5] " + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "inline_equation", + "content": "63.8\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Cross-Style Adaptation Scenario Additionally, we assess SEEN-DA on the more challenging Cross-Style adaptation, where the semantic hierarchy has broader discrepancies. In Table 2, SEEN-DA peaks with " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "47.9\\%" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " mAP and improves six categories (aeroplane, bird, boat, bus, sheep and train)." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 96, + 544, + 169 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 5. Comparison " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " of computational efficiency on Cross-Weather adaptation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 96, + 544, + 169 + ], + "lines": [ + { + "bbox": [ + 309, + 96, + 544, + 169 + ], + "spans": [ + { + "bbox": [ + 309, + 96, + 544, + 169 + ], + "type": "table", + "html": "
MethodBackbone Parm(M) LearnableParm(M) mAPAbs. Gains
DSS [56]26.83426.83440.9+4.2
CSDA [13]33.64533.64545.3+6.9
REACT [24]29.81229.81248.8+7.9
AT [31]39.22518.72350.9+7.9
DA-Pro [23]34.8340.00855.9+3.3
SEEN-DA (Ours)36.7011.87557.5+4.9
", + "image_path": "233f24130efaf042b57e1f6fd7539ce97cc1ccdd32abd4c144d0c100dbeeef2a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 176, + 406, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 406, + 188 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 406, + 188 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 194, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 219 + ], + "type": "text", + "content": "We report detailed ablation studies (Table 3, 4, 5, Figure 4, 5) to validate the effectiveness of proposed SEEN." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "content": "Standard Attention vs. Domain-aware Attention We compare the performance of domain-aware attention with existing attention methods, including self-attention and cross-attention, as shown in Table 3. With the guidance of textual information, the cross-attention improves the self-attention by " + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "inline_equation", + "content": "0.7\\% \\sim 2.1\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "content": ". Furthermore, the proposed domain-aware attention gains improvements of " + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "inline_equation", + "content": "2.0\\% \\sim 33.6\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 327 + ], + "type": "text", + "content": " over cross-attention, showing a superior ability to select features with high transferability." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "text", + "content": "Ablation for Domain-aware Attention Module We conduct ablation experiments to further analyze the significance of proposed domain-aware attention module in Table 4. Introducing the inter-domain attention branch attains an mAP of " + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "inline_equation", + "content": "54.9\\%" + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "text", + "content": ", and optimizing with adversarial loss " + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}" + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "text", + "content": " gains " + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "text", + "content": ". Meanwhile, applying the intra-domain attention branch improves " + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 304, + 328, + 545, + 448 + ], + "type": "text", + "content": " on the target domain. These results showing the effectiveness of the domain-aware attention module to eliminate redundant information and supplement domain-specific semantic information." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 520 + ], + "type": "text", + "content": "Computational Efficiency We discussed the computational efficiency of the proposed SEEN-DA in Table 5. Our method yields the same parameters scale as existing methods and achieves the highest mAP. In addition, SEEN-DA only needs to train very few parameters (1.875M), showing remarkable efficiency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 604 + ], + "type": "text", + "content": "Numbers of the Attention Module We explore the impact of numbers of domain-aware attention modules in Figure 4. Starting from the last visual block, gradually introducing domain-aware attention modules to each visual block improves performance and peaks when using 3 modules. Considering efficiency, we apply 3 domain-aware attention modules into the visual encoder." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Projection layer in SEEN module We explore the projection layer in the SEEN module in Figure 5. Projecting textual embedding into visual space, denoted as T2V, only achieves " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "55.1\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". Projecting textual and visual embedding into a latent space, denoted as T&V, improves performance by " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". And projecting visual embedding into textual space achieves the best mAP of " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "57.5\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". We analyze the reason for these results as follows. Visual embedding contains more sample noise and redundant information than textual" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25471" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 217, + 228 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 217, + 228 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 217, + 228 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 217, + 228 + ], + "type": "image", + "image_path": "bc7431c25c8a57c1d07ad9d0f616a3cc011ca3de1ff79b778e19ea897bdcd09c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 230, + 166, + 239 + ], + "lines": [ + { + "bbox": [ + 96, + 230, + 166, + 239 + ], + "spans": [ + { + "bbox": [ + 96, + 230, + 166, + 239 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 68, + 373, + 228 + ], + "blocks": [ + { + "bbox": [ + 217, + 68, + 373, + 228 + ], + "lines": [ + { + "bbox": [ + 217, + 68, + 373, + 228 + ], + "spans": [ + { + "bbox": [ + 217, + 68, + 373, + 228 + ], + "type": "image", + "image_path": "1833aa968d6a00cad5eb5261add9b84402964fb83850fe591836b5a4946de3e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 230, + 326, + 240 + ], + "lines": [ + { + "bbox": [ + 263, + 230, + 326, + 240 + ], + "spans": [ + { + "bbox": [ + 263, + 230, + 326, + 240 + ], + "type": "text", + "content": "(b) DA-Pro[23]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 373, + 68, + 533, + 228 + ], + "blocks": [ + { + "bbox": [ + 373, + 68, + 533, + 228 + ], + "lines": [ + { + "bbox": [ + 373, + 68, + 533, + 228 + ], + "spans": [ + { + "bbox": [ + 373, + 68, + 533, + 228 + ], + "type": "image", + "image_path": "506621036038f9f185e7139f7d1642ddf01c16005aa4194f907503865a83c744.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 230, + 489, + 240 + ], + "lines": [ + { + "bbox": [ + 410, + 230, + 489, + 240 + ], + "spans": [ + { + "bbox": [ + 410, + 230, + 489, + 240 + ], + "type": "text", + "content": "(c) SEEN-DA (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 274, + 253, + 399 + ], + "blocks": [ + { + "bbox": [ + 72, + 274, + 253, + 399 + ], + "lines": [ + { + "bbox": [ + 72, + 274, + 253, + 399 + ], + "spans": [ + { + "bbox": [ + 72, + 274, + 253, + 399 + ], + "type": "image", + "image_path": "4b3c8b2a75c6de307c143227ffe30e431bf2ba93b8eaffb23228d66f5128065b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 288, + 424 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 424 + ], + "type": "text", + "content": "Figure 4. Ablation " + }, + { + "bbox": [ + 46, + 402, + 288, + 424 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 402, + 288, + 424 + ], + "type": "text", + "content": " on the number of attention modules on Cross-Weather adaptation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 434, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 471 + ], + "type": "text", + "content": "embedding. And learning to project visual embedding to textual space can reduce the interference of these factors. Therefore, we adopt the V2T projection in SEEN module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 472, + 132, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 132, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 132, + 484 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 491, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 659 + ], + "type": "text", + "content": "The detection results of ground truth (a), DA-Pro [23] (b) and the proposed SEEN-DA (c) on the target domain of Cross-Weather adaptation are displayed in Figure 3. The figure shows that SEEN-DA outperforms the SOTA method in terms of the ability to detect objects. Specifically, SEEN-DA detects the objects missed (e.g. the rider on the right of figures in 1st row and the person on the left of figures in 2nd row) and misclassified (e.g. the bicycle on the middle of figures in 1st row and the person on the left of figures in 2nd row), revealing the effectiveness of domain-aware attention. By utilizing semantic entropy to reduce redundant information and supplement domain-specific semantic information, the SEEN-DA improves the discriminability of the detector on the target domain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel SEMantic ENtropy guided Domain-aware Attention (SEEN-DA) for DAOD." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 273, + 508, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 244, + 546, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 244, + 546, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 546, + 268 + ], + "type": "text", + "content": "Figure 3. Detection comparison on the Cross-Weather adaptation scenario. We visualize (a) the ground truth, the detection boxes of (b) SOTA DA-Pro [23] and (c) our methods SEEN-DA." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 273, + 508, + 399 + ], + "lines": [ + { + "bbox": [ + 325, + 273, + 508, + 399 + ], + "spans": [ + { + "bbox": [ + 325, + 273, + 508, + 399 + ], + "type": "image", + "image_path": "f500264c46a79bd25ef6898c1d2cfd1ca571d241efcfbded1da13f5c4fff49ea.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 401, + 545, + 424 + ], + "lines": [ + { + "bbox": [ + 306, + 401, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 545, + 424 + ], + "type": "text", + "content": "Figure 5. Ablation " + }, + { + "bbox": [ + 306, + 401, + 545, + 424 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 306, + 401, + 545, + 424 + ], + "type": "text", + "content": " on projection layer of the domain-aware attention module on Cross-Weather adaptation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 579 + ], + "type": "text", + "content": "We design the semantic entropy to quantify the semantic information contained in visual features. Guided by semantic entropy, the domain-aware attention module is introduced to adaptively refine visual features with the semantic commonalities and disparities between domains. Precisely, it consists of an inter-domain attention branch for extracting domain-invariant features and eliminating redundant information between domains, and an intra-domain attention branch for supplementing the domain-specific semantic information for each domain. Extensive experiments over multiple benchmarks validate the effectiveness of SEENDA in inferring a discriminative detector." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgment This work is partially supported by the National Key R&D Program of China(under Grant 2023YFB4502200), the NSF of China(under Grants 92364202, U22A2028, 62102398, 62102399, 62302478, 62302482, 62376268), Strategic Priority Research Program of the Chinese Academy of Sciences, (Grant No.XDB0660300, XDB0660301, XDB0660302), Major Program of ISCAS(Grant No. ISCAS-ZD-202402), Beijing Natural Science Foundation(4222039), CAS Project for Young Scientists in Basic Research(YSBR-029) and Youth Innovation Promotion Association CAS." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25472" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Chaoqi Chen, Jiongcheng Li, Hong-Yu Zhou, Xiaoguang Han, Yue Huang, Xinghao Ding, and Yizhou Yu. Relation matters: foreground-aware graph-based relational reasoning for domain adaptive object detection. TPAMI, 45(03):3677-3694, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xin chao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, et al. Learning domain adaptive object detection with probabilistic teacher. In ICML, pages 3040-3055, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "text", + "content": "[3] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, pages 3339-3348, 2018. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 303 + ], + "type": "text", + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[5] Yiming Cui, Liang Li, Jiehua Zhang, Chenggang Yan, Hongkui Wang, Shuai Wang, Heng Jin, and Li Wu. Stochastic context consistency reasoning for domain adaptive object detection. In ACMMM, pages 1331-1340, 2024. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 384 + ], + "type": "text", + "content": "[6] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In CVPR, pages 4091-4101, 2021. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 418 + ], + "type": "text", + "content": "[7] Jinhong Deng, Dongli Xu, Wen Li, and Lixin Duan. Harmonious teacher for cross-domain object detection. In CVPR, pages 23829-23838, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 485 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 531 + ], + "type": "text", + "content": "[9] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 575 + ], + "type": "text", + "content": "[10] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul De Charette. Poda: Prompt-driven zero-shot domain adaptation. In ICCV, pages 18623-18633, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[11] Yongchao Feng, Shiwei Li, Yingjie Gao, Ziyue Huang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Dsd-da: Distillation-based source debiasing for domain adaptive object detection. In ICML, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, 17(1):2096-2030, 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[13] Changlong Gao, Chengxu Liu, Yujie Dun, and Xueming Qian. Csda: Learning category-scale joint feature for domain adaptive object detection. In ICCV, pages 11421-11430, 2023. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 128 + ], + "type": "text", + "content": "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. NeurIPS, 17, 2004. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "text", + "content": "[16] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "text", + "content": "[17] Dayan Guan, Jiaxing Huang, Aoran Xiao, Shijian Lu, and Yanpeng Cao. Uncertainty-aware unsupervised domain adaptation in object detection. TMM, 24:2502-2514, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 239 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[19] Mengzhe He, Yali Wang, Jiaxi Wu, Yiru Wang, Hanqing Li, Bo Li, Weihao Gan, Wei Wu, and Yu Qiao. Cross domain object detection by target-perceived dual branch distillation. In CVPR, pages 9570-9580, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 327 + ], + "type": "text", + "content": "[20] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In CVPR, pages 5001-5009, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 381 + ], + "type": "text", + "content": "[21] Matthew Johnson-Roberson, Charles Barto, Rounak Mehta, Sharath Nittur Sridhar, Karl Rosaen, and Ram Vasudevan. Driving in the matrix: Can virtual worlds replace human-generated annotations for real world tasks? In ICRA 2017, pages 746-753. IEEE, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[22] Mikhail Kennerley, Jian-Gang Wang, Bharadwaj Veeravalli, and Robby T Tan. Cat: Exploiting inter-class dynamics for domain adaptive object detection. In CVPR, pages 16541-16550, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[23] Haochen Li, Rui Zhang, Hantao Yao, Xinkai Song, Yifan Hao, Yongwei Zhao, Ling Li, and Yunji Chen. Learning domain-aware detection head with prompt tuning. NeurIPS, 36, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[24] Haochen Li, Rui Zhang, Hantao Yao, Xin Zhang, Yifan Hao, Xinkai Song, and Ling Li. React: Remainder adaptive compensation for domain adaptive object detection. TIP, 33:3735-3748, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 514, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 559 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 647 + ], + "type": "text", + "content": "[27] Kai Li, Curtis Wigington, Chris Tensmeyer, Vlad I. Morariu, Handong Zhao, Varun Manjunatha, Nikolaos Barmpalios, and Yun Fu. Improving cross-domain detection with self-supervised learning. In CVPR, pages 4745-4754, 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[28] Wuyang Li, Xinyu Liu, Xiwen Yao, and Yixuan Yuan. Scan: Cross domain object detection with semantic conditioned adaptation. In AAAI, volume 6, page 7, 2022. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[29] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object detection. In CVPR, pages 5291-5300, 2022. 3, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25473" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[30] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma++: Improved semantic-complete graph matching for domain adaptive object detection. TPAMI, 45(07):9022-9040, 2023. 1, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 163 + ], + "type": "text", + "content": "[31] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Va-jda. Cross-domain adaptive teacher for object detection. In CVPR, pages 7581-7590, 2022. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 165, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 165, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 165, + 287, + 208 + ], + "type": "text", + "content": "[32] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 210, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 210, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 210, + 287, + 253 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 255, + 287, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 255, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 49, + 255, + 287, + 299 + ], + "type": "text", + "content": "[34] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. TMM, 25:1333-1344, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 301, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 301, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 49, + 301, + 287, + 334 + ], + "type": "text", + "content": "[35] Rui Liu, Yahong Han, Yaowei Wang, and Qi Tian. Frequency spectrum augmentation consistency for domain adaptive object detection. arXiv preprint arXiv:2112.08605, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 335, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 335, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 49, + 335, + 287, + 378 + ], + "type": "text", + "content": "[36] Yabo Liu, Jinghua Wang, Chao Huang, Yaowei Wang, and Yong Xu. Cigar: Cross-modality graph reasoning for domain adaptive object detection. In CVPR, pages 23776-23786, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 381, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 381, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 381, + 287, + 414 + ], + "type": "text", + "content": "[37] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 416, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 416, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 416, + 287, + 448 + ], + "type": "text", + "content": "[38] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. NeurIPS, 31, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 450, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 450, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 49, + 450, + 287, + 483 + ], + "type": "text", + "content": "[39] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, pages 2208-2217, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 485, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 485, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 485, + 287, + 517 + ], + "type": "text", + "content": "[40] Fabio Maria Carlucci, Lorenzo Porzi, Barbara Caputo, Elisa Ricci, and Samuel Rota Bulo. Autodial: Automatic domain alignment layers. In ICCV, pages 5067-5075, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 520, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 520, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 49, + 520, + 287, + 552 + ], + "type": "text", + "content": "[41] Pietro Morerio, Jacopo Cavazza, and Vittorio Murino. Minimal-entropy correlation alignment for unsupervised deep domain adaptation. In ICLR, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 49, + 555, + 287, + 587 + ], + "type": "text", + "content": "[42] Zhengquan Piao, Linbo Tang, and Baojun Zhao. Unsupervised domain-adaptive object detection via localization regression alignment. TNNLS, 35(11):15170-15181, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 589, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 589, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 589, + 287, + 644 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 287, + 678 + ], + "type": "text", + "content": "[44] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "text", + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 28, 2015. 1, 3, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[46] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In ICCV, pages 8050-8058, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[47] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In CVPR, pages 6956–6965, 2019. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[48] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Semantic foggy scene understanding with synthetic data. IJCV, 126(9):973-992, Sep 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[49] Sunandini Sanyal, Ashish Ramayee Asokan, Suvaansh Bhambri, Akshay Kulkarni, Jogendra Nath Kundu, and R Venkatesh Babu. Domain-specificity inducing transformers for source-free domain adaptation. In ICCV, pages 18928-18937, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[50] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NPIS, 30, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[51] VS Vibashan, Poojan Oza, and Vishal M Patel. Instance relation graph guided source-free domain adaptive object detection. In CVPR, pages 3520-3530, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "text", + "content": "[52] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 376 + ], + "type": "text", + "content": "[53] Vibashan Vs, Vikram Gupta, Poojan Oza, Vishwanath A Sindagi, and Vishal M Patel. Mega-cda: Memory guided attention for category-aware unsupervised domain adaptive object detection. In CVPR, pages 4516-4526, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 377, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 410 + ], + "type": "text", + "content": "[54] Haoan Wang, Shilong Jia, Tieyong Zeng, Guixu Zhang, and Zhi Li. Triple feature disentanglement for one-stage adaptive object detection. In AAAI, pages 5401-5409, 2024. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 443 + ], + "type": "text", + "content": "[55] Ximei Wang, Liang Li, Weirui Ye, Mingsheng Long, and Jianmin Wang. Transferable attention for domain adaptation. In AAAI, volume 33, pages 5345-5352, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[56] Yu Wang, Rui Zhang, Shuo Zhang, Miao Li, Yangyang Xia, XiShan Zhang, and ShaoLi Liu. Domain-specific suppression for adaptive object detection. In CVPR, pages 9603-9612, 2021. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[57] Weixi Weng and Chun Yuan. Mean teacher detr with masked feature alignment: A robust domain adaptive detection transformer framework. In AAAI, pages 5912-5920, 2024. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[58] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. TPAMI, 44(8):4178-4193, 2021. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "type": "text", + "content": "[59] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. In ICCV, pages 9342-9351, 2021. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 655 + ], + "type": "text", + "content": "[60] Xiaofu Wu, Suofei Zhang, Quan Zhou, Zhen Yang, Chunming Zhao, and Longin Jan Latecki. Entropy minimization versus diversity maximization for domain adaptation. TNNLS, 34(6):2896-2907, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[61] Jayeon Yoo, Inseop Chung, and Nojun Kwak. Unsupervised domain adaptation for one-stage object detector using offsets to bounding box. In ECCV, pages 691-708, 2022. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[62] Yifan Zhang, Hanbo Chen, Ying Wei, Peilin Zhao, Jiezhang Cao, Xinjuan Fan, Xiaoying Lou, Hailing Liu, Jinlong Hou," + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25474" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 115 + ], + "type": "text", + "content": "Xiao Han, et al. From whole slide imaging to microscopy: Deep microscopy adaptation network for histopathology cancer image classification. In MICCAI, pages 360-368, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 286, + 150 + ], + "type": "text", + "content": "[63] Yixin Zhang, Zilei Wang, and Yushi Mao. Rpn prototype alignment for domain adaptive object detector. In CVPR, pages 12425-12434, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "text", + "content": "[64] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In CVPR, pages 14217-14226, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "type": "text", + "content": "[65] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "text", + "content": "[66] Wenzhang Zhou, Heng Fan, Tiejian Luo, and Libo Zhang. Unsupervised domain adaptive detection with network stability analysis. In ICCV, pages 6986-6995, 2023. 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25475" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_content_list.json b/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..92027b7ce9d5c7b8ba43204d82bcb8cfb3c589a5 --- /dev/null +++ b/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_content_list.json @@ -0,0 +1,1510 @@ +[ + { + "type": "text", + "text": "SET: Spectral Enhancement for Tiny Object Detection", + "text_level": 1, + "bbox": [ + 222, + 130, + 774, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huixin Sun $^{1}$ , Runqi Wang $^{2}$ , Yanjing Li $^{1}$ , Linlin Yang $^{5}$ , Shaohui Lin $^{6}$ , Xianbin Cao $^{1*}$ , Baochang Zhang $^{3,4*}$", + "bbox": [ + 230, + 179, + 769, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electronic Information Engineering, Beihang University, Beijing, China", + "bbox": [ + 171, + 215, + 825, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computer Science and Technology, Beijing Jiaotong University", + "bbox": [ + 205, + 233, + 792, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Artificial Intelligence, Beihang University, Beijing, China", + "bbox": [ + 227, + 251, + 769, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Zhongguancun Laboratory, Beijing, China", + "bbox": [ + 328, + 268, + 669, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "${}^{5}$ State Key Laboratory of Media Convergence and Communication,", + "bbox": [ + 233, + 286, + 763, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Communication University of China, Beijing, China", + "bbox": [ + 292, + 305, + 705, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ School of Computer Science and Technology, East China Normal University, Shanghai, China", + "bbox": [ + 120, + 321, + 870, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 375, + 326, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has significantly advanced the object detection field. However, tiny object detection (TOD) remains a challenging problem. We provide a new analysis method to examine the TOD challenge through occlusion-based attribution analysis in the frequency domain. We observe that tiny objects become less distinct after feature encoding and can benefit from the removal of high-frequency information. In this paper, we propose a novel approach named Spectral Enhancement for Tiny object detection (SET), which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate the effectiveness of our method. Especially, SET boosts the prior art RFLA by $3.2\\%$ AP on the AI-TOD dataset.", + "bbox": [ + 89, + 407, + 485, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 727, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in Deep Neural Networks (DNNs) [14, 60] have significantly improved the object detection field [11, 22]. Despite the progress, Tiny Object Detection (TOD) remains a challenging problem. Tiny objects, characterized by their very limited pixel input, occupy areas equal to or less than $16 \\times 16$ pixels [44]. Compared to traditional object detection [10, 21], generic object detectors often fail to manifest effectiveness on the TOD task. For instance,", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e0b8c54b258c0a80bd94592a0ff244e9d8820ecbd3957f12f3786a192ad2c7e.jpg", + "image_caption": [ + "Figure 1. A feature-level occlusion-based attribution analysis in the frequency domain. Mid-frequency features ranges from $40\\%$ to $60\\%$ of the total energy, high-frequency features are above $90\\%$ of the total energy. The model is trained on the AI-TOD [44] trainval and validated on the AI-TOD test. Detection performance is evaluated by inferencing a $1\\times$ FCOS detector [39] with the masked features. Results exhibit that tiny objects are more likely to be affected by high-frequency noise and can benefit from the direct removal of high-frequency information." + ], + "image_footnote": [], + "bbox": [ + 514, + 373, + 906, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DINO [60], one of the most representative transformer-based object detector, achieves $37.6\\%$ AP on medium-sized objects but only $9.9\\%$ AP on very tiny objects on the AITOD [44] benchmark. Moreover, the vanilla FCOS [39] achieves only $2.5\\%$ AP on very tiny objects on the AITOD [44], which is far from sufficient to meet the demands of real-world applications, such as autonomous driving, maritime rescue, and traffic management.", + "bbox": [ + 511, + 625, + 906, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Constrained by the inherently low resolution, one key challenge in TOD is to extract discriminative foreground features [7], especially in the high-level features of detection architecture after down-samplings. In the meantime, it is difficult to learn about tiny objects from the noisy and dominant background clutter [54]. Most existing works [1, 3, 30, 54] focus on enhancing feature representations of tiny objects in the spatial domain. We provide a new analysis method and address the challenge from a frequency spectrum perspective.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors. Email: sunhuixin@buaa.edu.cn", + "bbox": [ + 107, + 886, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4713", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/484f014fad10e1ba08980fd642514dbdb723c4613adf3cecc7d6b49c3d76c51c.jpg", + "image_caption": [ + "Figure 2. The analysis examines different objects after feature encoding in the frequency domain. The frequency spectrum is calculated using the Fast Fourier Transform (FFT) [31] on the cropped $P_{3}$ feature and selecting the channel with the highest energy. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by occlusion-based attribution analysis [33, 45], we perform an occlusion-based analysis on the intermediate Feature Pyramid Network (FPN) features to examine the impact of masking specific frequency bands on the detection performance across object scales. Specifically, we transform the features into the frequency domain using the n-dimensional Fast Fourier Transform (FFT) [31] and apply static filters to the amplitude spectrums, which are then fed to Inverse Fast Fourier Transform (IFFT) with phase spectrums to generate the masked spatial features. High-frequency mask results in Fig. 1 reveal that directly removing high-frequency features has a positive effect on very tiny and tiny objects, while negatively affecting the detection of larger objects. The disparity can be attributed to that tiny objects become less distinct after feature encoding, exhibiting weak high-frequency components (example (b), Fig. 2) and extremely vague frequency signatures in the cluttered background (example (c), Fig. 2). Therefore, they are more likely to be affected by high-frequency noise and benefit from the direct removal of high-frequency information. In contrast, larger objects exhibit distinct high-frequency signatures (example (a), Fig. 2) and have a higher dependency on high-frequency information. Based on the observations, we investigate an explicit measure to enhance the frequency signatures for tiny objects by removing the high-frequency information in the background. Background (BG) high-frequency mask results show that the approach enhances detection performance across all three object scales, resulting in improvements of $15\\%$ , $8\\%$ , and $4\\%$ , bringing significant gains for very tiny and tiny objects without compromising larger objects.", + "bbox": [ + 91, + 356, + 483, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of the analysis above, we propose a generic Spectral Enhancement (SET) method to enhance the frequency signatures of Tiny objects using a heterogeneous architecture for foreground and background feature refinement. The method includes two modules. As illustrated in", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 3, the first Hierarchical Background Smoothing (HBS) module suppresses the high-frequency information in the background through adaptive smoothing operations while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. The second Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. In addition, API facilitates robust feature representations through adversarial training. SET is simple yet effective and can be easily mounted on existing detectors during the training process while bringing no extra burden to the inference procedure.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our major contributions in this paper are summarized as:", + "bbox": [ + 511, + 287, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A feature-level occlusion-based attribution analysis from the frequency spectrum perspective is conducted to investigate the Tiny Object Detection (TOD) challenge, which shows that tiny objects are more likely to be affected by high-frequency noise. We introduce a spectral enhancement scheme for tiny object detection (SET) by designing a heterogeneous architecture for foreground and background feature refinement.", + "- Two new modules are designed for the TOD task. The HBS module is used to suppress the high-frequency noise in the background through adaptive smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training.", + "- Extensive results reveal that our SET boosts the well-known baselines and the prior art by a large margin." + ], + "bbox": [ + 511, + 303, + 905, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 558, + 656, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tiny Object Detection. Most current approaches for tiny object detection (TOD) can be grouped into four main categories: data augmentation, multi-scale feature learning, label assignment, and feature enhancement strategies for tiny objects. Several data augmentation strategies [6, 37] target tiny objects by prompting the detector to focus on instances of specific scales during training. Conventional approaches also utilize multi-scale feature learning [23, 26, 62]. Another method line focuses on improving tiny object detection from the label assignment and proposal refinement perspectives [52, 58]. Various works aim to enhance the feature representations of tiny objects. Yuan et al. [58] propose a Feature Imitation (FI) mechanism, aligning RoI features with their counterparts in an exemplar feature set within the embedding space. Recent state-of-the-art TOD methods have proposed auxiliary self-reconstruction branches to enhance the weak representations of tiny objects [3, 54]. Differently, our proposed method enhances the discrimination of tiny objects from a frequency perspective, prompting the refinement of object features during training without adding any extra burden to the inference process.", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4714", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bdc484cd6bab9053a4ed6b8c44e788405c0eef8fc2501b90cf20c24b057b10ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 504, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/94f5b3567065d32c04c37661336e95e48327e29f292a22c2ef6a866435461421.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 903, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a8391d7369fcd36840858097f8a68f58c6802a9589f046895e7fd1ab20670af.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed SET method with FCOS framework. The red dashed line denotes generating the adversarial perturbations $\\epsilon_{i}^{\\mathrm{adv}}$ . The HBS module can suppress the high-frequency noise in the background through scale-wise smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and facilitate robust feature representations through adversarial training." + ], + "image_footnote": [], + "bbox": [ + 94, + 296, + 903, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Training for Object Detection. Many effective attacks crafted for object detectors have been proposed, with most generating adversarial examples at the image level [28, 38, 48]. The most common method to defend against the attacks is adversarial training [20, 29, 40], which involves augmenting training datasets with adversarial examples. For object detection, [59] extend adversarial training to the scenario of object detection by leveraging the attacks sourced from both classification and localization domains. Different from the aforementioned methods, we introduce feature-level adversarial perturbations that balance the detection semantics across branches for enhanced tiny object detection, resulting in more salient foreground features and improved model robustness.", + "bbox": [ + 88, + 515, + 485, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 750, + 191, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents the Spectral Enhancement (SET) method for Tiny object detection, depicted in Fig. 3. SET employs a heterogeneous architecture. This includes the Hierarchical Background Smoothing (HBS) module, which suppresses high-frequency information in the background, and the Adversarial Perturbation Injection (API) module, which enhances the feature saliency of critical regions and prompts the refinement of object features during training.", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Hierarchical Background Smoothing", + "text_level": 1, + "bbox": [ + 511, + 515, + 834, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the analysis in Sec. 1, we design a Hierarchical Background Smoothing (HBS) module to suppress the high-frequency information in the background feature while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. This is achieved through background feature decoupling and adaptive smoothing, as described below.", + "bbox": [ + 511, + 537, + 906, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $X\\in \\mathbb{R}^{3\\times W\\times H}$ , a binary mask $M$ is generated according to the ground truth box $B$ :", + "bbox": [ + 511, + 641, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} = \\mathbf {1} [ (i, j) \\in B ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 676, + 903, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M \\in \\{0,1\\}^{W \\times H}$ . The indicative function 1 denotes the value of location $(i,j)$ , which is 1 if it belongs to an object and 0 otherwise. Then we utilize the generated binary mask to decouple image features and perform background smoothing, as shown in Fig. 3. In a Feature Pyramid Network (FPN) comprising $N$ layers, the feature of the $i$ -th layer is denoted as $\\mathbf{P}_i$ , $(i \\in [1,N])$ . The process of HBS on the $i$ -th layer is as follows:", + "bbox": [ + 511, + 698, + 906, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {P} _ {i} ^ {E} = \\mathbf {P} _ {i} ^ {f g} + \\phi_ {i} \\left(\\mathbf {P} _ {i} ^ {b g}, r\\right) \\tag {2} \\\\ = \\mathbf {P} _ {i} * M + \\phi_ {i} (\\mathbf {P} _ {i} * \\neg M, r), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 823, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_i^E$ , $\\mathbf{P}_i^{fg}$ and $\\mathbf{P}_i^{bg}$ denote the enhanced feature, foreground feature and background feature, $\\neg M$ represents the", + "bbox": [ + 511, + 868, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4715", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e06a9f340b904b2f945069a37ed92043781a1654d730b865e62da90ad1aff6e6.jpg", + "image_caption": [ + "Figure 4. We visualize (a) the input image with ground-truths shown in green boxes, (b) $P_{3}$ feature in the vanilla FCOS after Principal Component Analysis (PCA) [46], and (c) $P_{3}$ feature in FCOS enhanced with HBS after PCA. Darker colors in the PCA maps indicate more critical components in the feature. As shown, HBS smoothes the background regions and significantly enhances the contrast between the foreground and background." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 460, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "complement of the binary mask $M$ , and $\\circledast$ is the Hadamard product with broadcasting to accommodate the dimensions. Previous works [18, 41] reveal that feature dimensionality reduction methods can lead to smoother and more efficient feature spaces. Based on this idea, we design an adaptive smoothing operation $\\phi_i(\\cdot ,r)$ with channel reduction rate $r$ , formulated as:", + "bbox": [ + 89, + 565, + 482, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} \\left(\\mathbf {P} _ {i} ^ {\\mathrm {b g}}, r\\right) = \\sigma \\left(\\mathbf {w} _ {i} ^ {\\mathrm {e}} \\otimes \\sigma \\left(\\mathbf {w} _ {i} ^ {\\mathrm {r}} \\otimes \\mathbf {P} _ {i} ^ {\\mathrm {b g}}\\right)\\right) + \\mathbf {P} _ {i} ^ {\\mathrm {b g}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 675, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ represents the convolution operation and $\\mathbf{w}_i^{\\mathrm{r}}\\in$ $\\mathbb{R}^{C_i\\times C_{i / r}\\times K_i\\times K_i}$ and $\\mathbf{w}_i^{\\mathrm{e}}\\in \\mathbb{R}^{C_{i / r}\\times C_i\\times K_i\\times K_i}$ are the weights of the channel reducing and expanding kernels, respectively. $\\sigma$ denotes the non-linear function ReLU. The reduction step compresses information while retaining salient features. When features are expanded, the high-frequency information are difficult to be recreated, resulting in a smoother background. The effectiveness of smoothing is exhibited through statistical analysis in Sec. 4.7. The smoothing effect is further visualized through PCA analysis. As shown in Fig. 4, feature maps enhanced with HBS exhibit a reduced intensity range and smoother transitions in the background regions compared to the vanilla FCOS,", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "effectively suppressing noise without losing essential background context.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kernel size $K_{i}$ at each scale level is determined based on the FPN stride $S_{i}$ , where smaller kernels are utilized to remove the fine-grained details in lower feature layers and larger kernels for more coarse-grained noise in higher layers, determined as:", + "bbox": [ + 511, + 119, + 903, + 195 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {i} = g \\left(S _ {i}\\right) = \\left(\\left\\lfloor \\frac {\\log_ {2} \\left(S _ {i}\\right)}{2} \\right\\rfloor \\times 2\\right) + 1, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 198, + 903, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rounding up $S_{i}$ to an odd integer and ensuring central pixel alignment in the convolutional kernel. The padding is determined as $(K_{i} - 1) / 2$ to preserve the spatial dimensions of the feature map. In addition, we also tested other $g(S_{i})$ , and the results are shown in the Tab. 5. The HBS architecture allows for the suppression of high-frequency information in the background while preserving foreground details, significantly enhancing the frequency signatures of tiny objects amid high-frequency noise.", + "bbox": [ + 511, + 238, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adversarial Perturbation Injection", + "text_level": 1, + "bbox": [ + 511, + 383, + 816, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis in Sec. 1 reveals that tiny objects exhibit weak high-frequency signatures after feature encoding and low discrimination against the cluttered background. To address this, we design an Adversarial Perturbation Injection (API) module, which leverages adversarial perturbations to enhance the saliency in critical regions. In addition, API promotes robust feature representations for tiny objects through adversarial training.", + "bbox": [ + 511, + 406, + 905, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by prior works in adversarial training [16, 17, 42], we exploit feature-level perturbations that adversarially change the model output to enhance feature representations. The objective of API for each FPN layer is formulated as follows:", + "bbox": [ + 511, + 527, + 905, + 601 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {P} _ {i}, \\theta_ {i}} \\left(\\max _ {\\| \\epsilon_ {i, \\mathrm {c l s}} \\| \\leq \\rho} \\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i} + \\epsilon_ {i, \\mathrm {c l s}}\\right) + \\gamma \\| \\mathbf {P} _ {i} \\| _ {2} ^ {2}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 606, + 903, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{cls}}$ denotes the classification loss. The inner optimization injects adversarial perturbation $\\epsilon_{i,\\mathrm{cls}}$ into the feature space of $\\mathbf{P}_i$ , where $\\rho$ defines the perturbation size, and $\\gamma$ is a hyperparameter that controls the strength of regularization. $\\theta_{i}$ denotes the model parameters at the $i$ -th layer. We derive a closed-form solution for the adversarial perturbation under the $L_{2}$ norm following [12]:", + "bbox": [ + 511, + 647, + 905, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {i, \\mathrm {c l s}} ^ {*} \\approx \\rho \\cdot \\frac {\\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {P} _ {i})}{\\| \\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {P} _ {i}) \\| _ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 758, + 903, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The perturbation can elevate the gradient of critical regions, thereby increasing their saliency during training. The gradient change induced by the perturbation on the classification branch can be calculated as:", + "bbox": [ + 511, + 799, + 903, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\left(\\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}}\\right) = \\left\\| \\nabla_ {\\mathbf {P} _ {i}} \\left(\\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i} + \\epsilon_ {i, \\mathrm {c l s}} ^ {*}\\right) - \\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i}\\right)\\right) \\right\\|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 869, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4716", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c3dafd05e366d1c8e20b3717800f36da73f88a0378f437de00087f3d2819032.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 89, + 344, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd9bab4cf0fb41c54e42f3c5664afb6c6c8c404d3cb32ec213b66db81e78e0e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 89, + 589, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/728dfac29a8eb039e8dcfd1078460a60c4f1c2e3c32caf5542c37b5deade5283.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2d4c907d5a5ae2e7f9a631618afbee807ac66f77009eb57766cdfdfba328d8c.jpg", + "image_caption": [ + "Figure 5. Average saliency visualization in (a) FCOS and (b) FCOS w/ API. The saliency maps are first computed based on the $\\ell_2$ -norm of the gradient [13] at neck feature $P_{3}$ . The saliency of each object is then summed and annotated within their respective bounding box, which is different between (a) and (b). The average saliency of each object scale is calculated as the mean saliency across objects in the corresponding size category, displayed in the bottom-left corner of each image." + ], + "image_footnote": [], + "bbox": [ + 98, + 198, + 344, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc9290d4ac9dce2f6cf1e326f26c5559de9e43770ea7cf5ca17e9ba3a9275240.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 198, + 589, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b083ab31446257913304a21d7a3a130ee3d51019e0e4a13152f32b021df4cc4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 198, + 903, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High $\\Delta (\\nabla_{\\mathbf{P}_i}\\mathcal{L}_{\\mathrm{cls}})$ corresponds to regions where model outputs are highly sensitive to minor changes in activations, indicating the presence of semantic information about the image [49]. This elevated gradient enhances the feature saliency of these critical regions and prompts the refinement of object features during training. The effect particularly benefits tiny objects as they exhibit weak high-frequency signatures and face a training deficiency problem. The enhanced object saliency phenomenon is discussed in Sec. 4.6 through Fig. 5. Furthermore, the adversarial training facilitates robust feature representations and reduces the sensitivity to high-frequency noise [63], thereby enhancing the discrimination of tiny objects.", + "bbox": [ + 88, + 391, + 482, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We formulate the adversarial perturbation $\\epsilon_{i}^{\\mathrm{adv}}$ for each FPN layer as an expectation over $M$ object detection branches (e.g., the classification, regression and center-ness branch in a FCOS detector [39]), each scaled by a corresponding balancing parameter $\\lambda_{m}$ . This is generated as follows:", + "bbox": [ + 89, + 589, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {i} ^ {\\mathrm {a d v}} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} \\cdot \\epsilon_ {i, m} ^ {*}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 676, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{i,m}^{*}$ represents the adversarial perturbation for the $m$ -th branch. The design allows for a balanced incorporation of detection semantics across multiple branches, ensuring that the perturbations are distributed for more effective training.", + "bbox": [ + 89, + 724, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Auxiliary Optimization", + "text_level": 1, + "bbox": [ + 89, + 816, + 305, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization leverages a total loss function that integrates both the original detection loss $\\mathcal{L}_{\\mathrm{detection}}$ and an auxiliary loss derived from the HBS and API enhanced features. This joint loss function is defined as:", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\underbrace {\\lambda \\cdot \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {i} ^ {\\mathrm {a u x}} \\left(\\mathbf {P} _ {i} ^ {E} + \\epsilon_ {i} ^ {\\mathrm {a d v}}\\right)} _ {\\text {a u x i l i a r y l o s s}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 388, + 906, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_i^{\\mathrm{aux}}$ denotes the auxiliary loss component for the $i$ -th layer, which is summed and scaled by a balancing hyperparameter $\\lambda$ . The perturbation $\\epsilon_i^{\\mathrm{adv}}$ is applied to the HBS-enhanced feature $\\mathbf{P}_i^E$ . The auxiliary optimization process updates the model's parameters in concert with the original loss, leading to more robust and discriminative features for tiny objects. We visualize the effectiveness of the method in Fig. 2. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS, benefiting their detection amid high-frequency noise and in the presence of larger objects.", + "bbox": [ + 511, + 450, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 646, + 645, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 670, + 833, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed method on AI-TOD [44], VisDrone2019 [65], DOTA-v2.0 [47] and COCO 2017 [22]. The main experiments were conducted using the AI-TOD dataset, with an average object size of 12.8 pixels, substantially smaller than other object detection datasets such as PASCAL VOC (156.6 pixels) [10] and MS COCO (99.5 pixels) [21]. Furthermore, we test our method on the VisDrone2019 [65] and DOTA-v2.0 [47], which comprises high-resolution drone-shot images with a substantial proportion of tiny instances. We also conduct experiments on the general object detection benchmark COCO 2017 [22], incorporating $\\mathrm{AP}_{vt}$ , $\\mathrm{AP}_t$ , $\\mathrm{AP}_s$ metrics to evaluate objects of very tiny (2-8 pixels), tiny (8-16 pixels) and small (16-32 pixels), following [44].", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4717", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7d7e3a25a001472d7b50707dbc7936aaaede3092b825e6e37131a8f43965f4d6.jpg", + "table_caption": [ + "Table 1. Main results with various frameworks on AI-TOD [44]. Models are trained on the AI-TOD trainval and validated on the AI-TOD test. We report APs (\\%) with different IoU thresholds and APs (\\%) for objects of various sizes based on the AI-TOD criterion. RFLA is based on Cascade R-CNN. The * denotes using P2~P6 FPN features. The **bold** indicates the best result." + ], + "table_footnote": [], + "table_body": "
FrameworkAP\\( AP_{0.5} \\)\\( AP_{0.75} \\)\\( AP_{vt} \\)\\( AP_t \\)\\( AP_s \\)
PAA [19]10.026.56.73.510.513.1
ATSS [61]11.628.57.62.511.915.9
Centernet [9]13.439.25.03.812.117.7
DetectoRS [34]14.832.811.50.010.818.3
DotD [50]16.139.210.68.317.618.1
DetectoRS w/NWD [43]20.849.314.36.419.729.6
DetectoRS w/ SR-TOD [3]24.054.617.110.124.829.3
One-stage
RetinaNet [24]7.220.53.62.56.811.0
RetinaNet w/ SET9.124.65.22.98.212.8
RetinaNet* [24]13.331.78.63.914.417.7
RetinaNet* w/ QueryDet [55]12.229.37.32.410.518.5
RetinaNet* w/ SET16.538.610.95.015.721.8
FCOS [39]12.029.08.02.511.917.1
FCOS w/ SET14.234.99.82.913.020.2
FCOS* [39]15.135.810.25.916.618.8
FCOS* w/ SET18.042.712.47.218.523.0
Multi-stage
Faster R-CNN [35]11.126.38.10.07.223.3
Faster R-CNN w/ SET12.228.29.70.19.124.3
Cascade R-CNN [2]13.630.310.60.09.925.5
Cascade R-CNN w/ SET14.832.411.20.410.826.7
RFLA [52]21.750.515.38.321.824.5
RFLA w/ SR-TOD [3]21.850.815.49.721.827.4
RFLA w/ SET24.955.617.99.825.231.0
Transformer-based
DINO-5scale [60]23.256.615.49.923.129.3
DINO-5scale w/ DNTR [25]26.256.720.212.826.431.0
DINO-5scale w/ SET26.657.120.813.227.131.5
", + "bbox": [ + 127, + 133, + 861, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We conducted the experiments on a computer with an NVIDIA RTX 3090 GPU. The experiments are implemented with PyTorch [32], with core codes built upon MMdetection [5]. The ImageNet [36] pre-trained models are used as the backbones. All CNN-based models utilize the ResNet-50 [15] backbone, trained using the Stochastic Gradient Descent (SGD) optimizer for 12 epochs with 0.9 momentum, 0.0001 weight decay, and a batch size 2. The initial learning rate is 0.005, decaying at the 8th and 11th epochs. The data processing adheres to the default configurations of each dataset (e.g., fixed at $800 \\times 800$ for AI-TOD). We also train a transformer-based detector, DINO [60], with 5-scale feature maps for 36 epochs as a baseline. The training uses an Adam optimizer with a weight decay of 0.0001, following the random crop and scale augmentation strategies of DETR [4].", + "bbox": [ + 88, + 574, + 480, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on AI-TOD", + "text_level": 1, + "bbox": [ + 89, + 830, + 274, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed SET method on various detectors and compare it with state-of-the-art TOD methods on the AI-TOD benchmark [51]. Tab. 1 shows that SET enhances", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "all baselines by about $2\\%$ AP, which is significant. Specifically, SET improves the one-stage detector FCOS [39] and RetinaNet [24] by $2.2\\% / 1.9\\%$ AP and $1.1\\% / 1.4\\%$ $\\mathrm{AP}_t$ , respectively. We further revise the detector architecture and use P2~P6 FPN features, a representative approach for TOD tasks where the high-resolution P2 layer benefits tiny objects. With the adjustment, SET boosts the RetinaNet and FCOS baselines by $3.2\\% / 2.9\\%$ AP and $1.1\\% / 1.3\\% \\mathrm{AP}_{vt}$ a significant margin. In contrast, the state-of-the-art SOD method QueryDet [55] decreases the baseline performance, showing limitations in addressing the TOD challenge. SET also generalizes to multi-stage detectors, enhancing the Faster R-CNN [35] and Cascade R-CNN [2] baselines by $1.1\\% / 1.2\\%$ AP and $1.9\\% / 1.1\\%$ $\\mathrm{AP}_t$ respectively.", + "bbox": [ + 511, + 574, + 906, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SET works orthogonally with the state-of-the-art TOD method RFLA [52], achieving a $3.2\\%$ AP increment over the RFLA baseline and surpassing the comparison SR-TOD [3] by $3.1\\%$ AP and $3.4\\%$ $\\mathrm{AP}_t$ . We also investigate SET's compatibility with transformer-based detectors. As shown, SET results in $3.4\\%$ AP increase over the DINO-5scale [60] baseline, outperforming the prior", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4718", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89d823cdf0677a5fb80331337e89b5f40a314d92f1b5a2e120a154d497a6c079.jpg", + "table_caption": [ + "Table 2. Detection performance on the VisDrone2019 [65] validation set. RFLA is based on Cascade R-CNN. The **bold** denotes the best result." + ], + "table_footnote": [], + "table_body": "
FrameworkAPAPvtAPtAPs
RetinaNet [24]15.20.41.69.3
Faster R-CNN [35]23.10.15.119.6
FCOS [39]19.90.74.515.7
FCOS w/ SET21.91.65.717.9
Cascade R-CNN [2]23.00.15.019.4
Cascade R-CNN w/ SET24.31.26.321.0
RFLA [52]27.24.513.023.6
RFLA w/ SR-TOD [3]27.84.813.124.5
RFLA w/ SET28.55.213.524.6
", + "bbox": [ + 91, + 141, + 485, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "art DNTR [25]. Notably, DINO-5scale w/SET derives $26.6\\%$ AP, outperforming competitors including DotD [50], NWD [43], and SR-TOD [3]. The results demonstrate SET as a robust solution for the tiny object detection task.", + "bbox": [ + 89, + 304, + 483, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on VisDrone2019 and DOTA-V2.0", + "text_level": 1, + "bbox": [ + 89, + 378, + 449, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate SET on the drone-shot image benchmark VisDrone2019 [65], which features tiny and small objects. As shown in Tab. 2, the proposed SET effectively improves the performance of both one-stage and multi-stage detectors and exhibits consistent $\\mathrm{AP}_{vt}$ and $\\mathrm{AP}_t$ improvements. Specifically, SET improved the the one-stage detector FCOS by $2.0\\% /0.9\\% /1.2\\%$ AP, $\\mathrm{AP}_{vt}$ and $\\mathrm{AP}_t$ , respectively. SET also improves multi-stage detector Cascade RCNN [2] by $1.1\\% /1.3\\%$ $\\mathrm{AP}_{vt}$ and $\\mathrm{AP}_t$ , respectively. Notably, SET improves the prior art RFLA, achieving a $28.5\\%$ AP with $0.7\\%$ increment in $\\mathrm{AP}_{vt}$ and $0.5\\%$ in $\\mathrm{AP}_t$", + "bbox": [ + 88, + 402, + 482, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 demonstrates the DOTA-v2.0 [8] results. As shown, SET obtains a substantial performance gain of $3.1\\%$ and $1.3\\%$ in $\\mathrm{AP}_t$ with the FCOS and AutoAssign baseline, underscoring the efficacy of our method in detecting tiny objects. Compared to the state-of-the-art comparisons RFLA [52] and DCFL [53], SET achieves $2.6\\%$ and $0.9\\%$ AP gain, respectively. More experimental results on the TinyPerson [56] and SeaPerson [57] dataset are provided in the supplementary material.", + "bbox": [ + 89, + 569, + 482, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on COCO", + "text_level": 1, + "bbox": [ + 89, + 719, + 266, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further conduct experiments on the general object detection benchmark MS COCO [21]. As shown in Tab. 9, SET achieves $1.0\\%$ AP improvement over the FCOS baseline, with a notable $1.7\\%$ improvement in $\\mathrm{AP}_t$ , demonstrating that SET can enhance the detection of tiny objects in general detection tasks.", + "bbox": [ + 89, + 742, + 482, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 847, + 243, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following experiments, we explore the best structure of SET by ablating and tuning its parts using AI-TOD.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3e418ef43ecbf75785e1b4915036554f29684eeaed0a316b14a174d55c300e33.jpg", + "table_caption": [ + "Table 3. Detection performance on DOTA-v2.0 [8]. Note that models are trained on the DOTA-v2.0 train and validated on the DOTA-v2.0 val. The bold denotes the best result." + ], + "table_footnote": [], + "table_body": "
FrameworkAPAPvtAPtAPs
ATSS [61]32.70.76.923.4
Faster R-CNN [35]35.60.07.128.9
FCOS [39]31.80.34.019.4
FCOS w/ RFLA [52]32.10.76.823.6
FCOS w/ DCFL [53]33.80.56.924.2
FCOS w/ SET34.70.67.124.8
AutoAssign [64]33.80.97.322.4
AutoAssign w/ SET35.21.18.623.6
", + "bbox": [ + 514, + 141, + 908, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f863f6561d5f03b74ae21e79ce4ca16a5b854d31fbca21d6243119a2db8b6c92.jpg", + "table_caption": [ + "Table 4. Detection performance on MS COCO [21]. Note that models are trained on COCO train2017 and validated on COCO val2017. $(\\cdot)^{+}$ denotes performance increment." + ], + "table_footnote": [], + "table_body": "
FrameworkAPAPvtAPtAPsAPm
FCOS [39]36.47.919.627.243.6
FCOS w/ SET37.4+1.08.3+0.423.3+1.727.6+0.444.8+1.2
", + "bbox": [ + 514, + 333, + 916, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness and Cost of Components. We present quantitative improvements of SET components in Tab. 5. As shown, applying the adaptive smoothing operations to full features $(+\\phi_{i}(\\mathbf{P}_{i},r))$ achieves a $0.4\\%$ increase in AP and $0.3\\%$ in $\\mathrm{AP}_{vt}$ , surpassing the performance of static high-frequency filtering in Fig. 1. With decoupling, the HBS module boosts the overall AP by $1.9\\%$ AP and $0.4\\%$ $\\mathrm{AP}_{vt}$ . The API module alone achieves a $1.2\\%$ AP improvement. Combining HBS and API results in a total AP increase of $2.2\\%$ . Results demonstrate the synergy between the heterogeneous operations, where HBS effectively smooths the clutter background and API prompts the learning of object features, cumulatively improving the detection accuracy of tiny objects. Tab. 5 also presents the cost of components. Results show that the HBS module leads to $17\\%$ , $0.9\\%$ , and $0.9\\%$ increase in training time, computational and memory cost. The API module results in a more significant increase in training time due to adversarial training but adds no forward computational cost and minimal increase in memory usage. The overall minimal impact indicates that SET can be implemented with negligible additional overhead.", + "bbox": [ + 511, + 401, + 908, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyper-parameter Selection. We first select the hyperparameters channel reduction rate $r$ and perturbation size $\\rho$ using the FCOS [39] framework trained three epochs in Fig. 6 (a). Model performance (AP) with different hyperparameter combinations $\\{r, \\rho\\}$ is presented. Results indicate that performance initially improves and then declines as $r$ varies from 2 to 16, with $r = 16$ perform worse than the baseline. Additionally, SET with adversarial perturbations $(\\rho > 0)$ exhibits stronger performance than without $(\\rho = 0)$ , but large-sized perturbations $(\\rho = 5)$ performs worse than all alternatives. Exploring the setups, we identify $\\{r, \\rho\\} = \\{4, 1\\}$ as the combination for SET. Fig. 6 (b) explores the", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4719", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8717e3c301cb756d30cf517a0e9b905696461d547d18febfe946195f015fd44a.jpg", + "table_caption": [ + "Table 5. The impact of different components and various channel reduction operations in SET. The hyper-parameters are fixed with channel reduction rate $r$ being 4, perturbation size $\\rho$ being 1, and the balancing hyper-parameters $\\lambda$ of auxiliary optimization being 1." + ], + "table_footnote": [], + "table_body": "
Frameworkφig(Si)APAPvtAPtTime(s/batch)GFLOPsParams (M)
FCOS [39]--12.02.511.90.17126.1337.04
+ φi(Pi, r)--12.72.712.90.17127.3637.41
Eqn. 3313.42.712.6
Eqn. 3log2(Si)/2×2+113.92.712.9
+ HBSEqn. 3513.12.312.50.20127.3237.41
Eqn. 3log2(Si+1)/2×2-113.42.812.4
Eqn. 3712.12.512.2
+ API--13.22.612.40.39126.1337.16
+ Both (SET)Eqn. 3Eqn. 414.22.913.00.42127.3237.85
", + "bbox": [ + 91, + 127, + 906, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c206ef2afd100a6fd3aeb89ceef8ffd5959e034570a8b9a49dfd9ccb2f45068f.jpg", + "image_caption": [ + "Figure 6. We select hyper-parameters channel reduction rate $r$ with perturbation size $\\rho$ in (a), and the balancing hyper-parameter $\\lambda$ of Eqn. 9 in (b)." + ], + "image_footnote": [], + "bbox": [ + 91, + 301, + 287, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68468b3b4f4089924a4fda40e525e9ca399c0f1a7bbe44ed024ea4e6851b68e4.jpg", + "image_caption": [ + "Figure 7. The analysis examines the adaptive smoothing operation, obtained from 1000 randomly selected images from the Vis-Drone2019 [65] train set." + ], + "image_footnote": [], + "bbox": [ + 292, + 301, + 480, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "balancing hyper-parameter $\\lambda$ in Eqn. 9 that controls the proportion of the auxiliary optimization, using the FCOS [39], RetinaNet [24], and Faster R-CNN [35] framework. As observed, the performances increase first and then decrease with the increase of $\\lambda$ , with $\\lambda = 1$ yielding the best performance across all frameworks. However, the efficacy of SET diminishes when $\\lambda$ exceeds the 1, as larger-sized $\\lambda$ ( $\\lambda > 1$ ) can impose too strong regularization to the detector, reducing the detection accuracy despite increasing robustness. Therefore, we set the hyper-parameter $\\lambda$ to 1 for experiments in this paper.", + "bbox": [ + 89, + 515, + 483, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Enhanced Object Saliency", + "text_level": 1, + "bbox": [ + 89, + 695, + 328, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate effectiveness of API in increasing the saliency of object features in Fig. 5. Comparing the average saliency of each object scale, tiny objects are the least salient due to their inherently limited pixel input. API leads to a notable increase in the average saliency of tiny objects (e.g., from 53.99 to 69.93), benefited from the injection of perturbations (Eqn. 7). Fig. 5 also demonstrates that API can enhance the feature saliency of small and medium objects, though the degree of improvement is less. The comparison underscores that API could benefit general object detection, with more significant and essential improvements for tiny objects.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a6f021a8eb905f53dc3e1a29e203e9461e4e8afaf5e2cfff66bb1bd9946ca44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 306, + 864, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Effectiveness in Smoothing", + "text_level": 1, + "bbox": [ + 511, + 506, + 756, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we perform a statistic analysis on $P_{i}^{bg}$ to investigate the adaptive smoothing operation. We convert feature $P_{i}^{bg}$ and $\\phi_{i}(P_{i}^{bg}, r)$ into the frequency space using FFT [31] and separate the amplitude spectrums of these features into three parts (low-, mid-, and high-frequency) using static thresholds as Fig. 1. The amplitudes in each band are then summed and averaged to measure the signal strength. Results verify that the channel reducing and expand design effects in smoothing. Comparing $A(P_{i})(u, v)$ and $A(\\phi_{i}(P_{i}^{bg}, r))(u, v)$ , the low-frequency signals are enhanced (from 0.3787 to 0.5205), while mid- and high-frequency signals are attenuated.", + "bbox": [ + 511, + 527, + 906, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes the Spectral Enhancement for Tiny objects (SET) method, which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate its effectiveness.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4720", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. The work was supported by the National Natural Science Foundation of China (No. 61827901, 623B2016, and 62406298). This research was also supported by the National Key Research and Development Program of China (Grant No. 2023YFC3306401), Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F020007, Beijing Natural Science Foundation L223024, L244043 and Z241100001324017, \"One Thousand Plan\" projects in Jiangxi Province Jxsq2023102268, and the Beijing Jiaotong University \"Jingying Plan\" No. K24XKRC00130.", + "bbox": [ + 89, + 90, + 485, + 247 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 273, + 187, + 289 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, pages 206-221, 2018. 1", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 6, 7", + "[3] Bing Cao, Haiyu Yao, Pengfei Zhu, and Qinghua Hu. Visible and clear: Finding tiny objects in difference map. ECCV, 2024. 1, 2, 6, 7", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6", + "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6", + "[6] Yukang Chen, Peizhen Zhang, Zeming Li, Yanwei Li, Xiangyu Zhang, Lu Qi, Jian Sun, and Jiaya Jia. Dynamic scale training for object detection. arXiv preprint arXiv:2004.12432, 2020. 2", + "[7] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE TPAMI, 2023. 1", + "[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE TPAMI, 44(11):7778-7796, 2021. 7", + "[9] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, pages 6569-6578, 2019. 6", + "[10] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. ICCV, 88:303-308, 2009. 1, 5", + "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 1" + ], + "bbox": [ + 93, + 299, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 4, 1", + "[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021. 5, 1", + "[14] Kaiming He and Jian Sun. Convolutional neural networks at constrained time cost. In CVPR, pages 5353-5360, 2015. 1", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[16] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 4", + "[17] Masato Ishii and Atsushi Sato. Training deep neural networks with adversarially augmented features for small-scale training datasets. In IJCNN, pages 1-8. IEEE, 2019. 4", + "[18] Weikuan Jia, Meili Sun, Jian Lian, and Sujuan Hou. Feature dimensionality reduction: a review. Complex & Intelligent Systems, 8(3):2663-2693, 2022. 4", + "[19] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371, 2020. 6", + "[20] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR*, 2016. 3", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 1, 5, 7, 3", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 1, 5", + "[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2", + "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 6, 7, 8, 2, 3", + "[25] Hou-I Liu, Yu-Wen Tseng, Kai-Cheng Chang, Pin-Jyun Wang, Hong-Han Shuai, and Wen-Huang Cheng. A denoising fpn with transformer r-cnn for tiny object detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6, 7", + "[26] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 2", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3", + "[28] Jiajun Lu, Hussein Sibai, and Evan Fabry. Adversarial examples that fool detectors. arXiv preprint arXiv:1712.02494, 2017.3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4721", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] T. Na, J. H. Ko, and S. Mukhopadhyay. Cascade adversarial machine learning regularized with a unified embedding. In ICLR, 2017. 3", + "[30] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In ICCV, pages 9725-9734, 2019. 1", + "[31] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 2, 8", + "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS Workshops, pages 1-12, 2019. 6", + "[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In CVPR, pages 11443-11452, 2021. 2", + "[34] Siyuan Qiao, Liang-Chieh Chen, and Alan Yuille. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In CVPR, pages 10213-10224, 2021. 6", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(06):1137-1149, 2017. 6, 7, 8", + "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. ICCV, 115:211-252, 2015. 6", + "[37] Bharat Singh, Mahyar Najibi, Abhishek Sharma, and Larry S Davis. Scale normalized image pyramids with autofocus for object detection. IEEE TPAMI, 44(7):3749-3766, 2021. 2", + "[38] Dawn Song, Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Florian Tramer, Atul Prakash, and Tadayoshi Kohno. Physical adversarial examples for object detectors. In 12th USENIX workshop on offensive technologies (WOOT '18), 2018. 3", + "[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 5, 6, 7, 8, 2, 3", + "[40] F Tramér, D Boneh, A Kurakin, I Goodfellow, N Papernot, and P McDaniel. Ensemble adversarial training: Attacks and defenses. In *ICLR*, 2018. 3", + "[41] S Velliangiri, SJPCS Alagumuthukrishnan, et al. A review of dimensionality reduction techniques for efficient computation. Procedia Computer Science, 165:104-111, 2019. 4", + "[42] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, pages 5495-5504, 2018. 4", + "[43] Jinwang Wang, Chang Xu, Wen Yang, and Lei Yu. A normalized gaussian Wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389, 2021. 6, 7", + "[44] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In ICPR, pages 3791-3798, 2021. 1, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Kunyu Wang, Xueyang Fu, Yukun Huang, Chengzhi Cao, Gege Shi, and Zheng-Jun Zha. Generalized uav object detection via frequency domain disentanglement. In CVPR, pages 1064-1073, 2023. 2", + "[46] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 4, 2", + "[47] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, pages 3974-3983, 2018. 5", + "[48] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In ICCV, pages 1369-1378, 2017. 3", + "[49] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In CVPR, pages 501-509, 2019. 5", + "[50] Chang Xu, Jinwang Wang, Wen Yang, and Lei Yu. Dot distance for tiny object detection in aerial images. In ICCV, pages 1192-1201, 2021. 6, 7", + "[51] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Detecting tiny objects in aerial images: A normalized wasserstein distance and a new benchmark. IS-PRS J. Photo. and Rem. Sen., 190:79-93, 2022. 6, 3", + "[52] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543, 2022. 2, 6, 7, 3", + "[53] Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Dynamic coarse-to-fine learning for oriented tiny object detection. In CVPR, pages 7318-7328, 2023. 7", + "[54] Sheng Xu, Mingze Wang, Yanjing Li, Mingbao Lin, Baochang Zhang, David Doermann, and Xiao Sun. Learning 1-bit tiny object detector with discriminative feature refinement. In ICML, 2024. 1, 2", + "[55] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 6", + "[56] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In WACV, pages 1257–1265, 2020. 7, 2", + "[57] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. Cpr++: Object localization via single coarse point supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4908-4925, 2024. 7, 3", + "[58] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small object detection via coarse-to-fine proposal generation and imitation learning. In ICCV, pages 6317-6327, 2023. 2", + "[59] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, pages 421-430, 2019. 3", + "[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4722", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1, 6", + "[61] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 6, 7, 3", + "[62] Qijie Zhao, Tao Sheng, Yongtao Wang, Zhi Tang, Ying Chen, Ling Cai, and Haibin Ling. M2det: A single-shot object detector based on multi-level feature pyramid network. In AAAI, pages 9259-9266, 2019. 2", + "[63] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, pages 452-467, 2018. 5", + "[64] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020. 7, 2, 3", + "[65] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Haibin Ling, Qinghua Hu, Qinqin Nie, Hao Cheng, Chenfeng Liu, Xiaoyu Liu, et al. Visdrone-det2018: The vision meets drone object detection in image challenge results. In ECCV Workshops, pages 0–0, 2018. 5, 7, 8" + ], + "bbox": [ + 91, + 90, + 483, + 402 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4723", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_model.json b/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..371fdb159ad9c2bbac1b1316b9721ad47102f31b --- /dev/null +++ b/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.776, + 0.154 + ], + "angle": 0, + "content": "SET: Spectral Enhancement for Tiny Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.77, + 0.218 + ], + "angle": 0, + "content": "Huixin Sun\\(^{1}\\), Runqi Wang\\(^{2}\\), Yanjing Li\\(^{1}\\), Linlin Yang\\(^{5}\\), Shaohui Lin\\(^{6}\\), Xianbin Cao\\(^{1*}\\), Baochang Zhang\\(^{3,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.217, + 0.826, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electronic Information Engineering, Beihang University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.234, + 0.794, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computer Science and Technology, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.252, + 0.77, + 0.269 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Artificial Intelligence, Beihang University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.27, + 0.671, + 0.288 + ], + "angle": 0, + "content": "\\(^{4}\\)Zhongguancun Laboratory, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.287, + 0.764, + 0.305 + ], + "angle": 0, + "content": "\\( {}^{5} \\) State Key Laboratory of Media Convergence and Communication," + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.306, + 0.706, + 0.322 + ], + "angle": 0, + "content": "Communication University of China, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.322, + 0.871, + 0.341 + ], + "angle": 0, + "content": "\\(^{6}\\)School of Computer Science and Technology, East China Normal University, Shanghai, China" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.375, + 0.327, + 0.391 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.486, + 0.709 + ], + "angle": 0, + "content": "Deep learning has significantly advanced the object detection field. However, tiny object detection (TOD) remains a challenging problem. We provide a new analysis method to examine the TOD challenge through occlusion-based attribution analysis in the frequency domain. We observe that tiny objects become less distinct after feature encoding and can benefit from the removal of high-frequency information. In this paper, we propose a novel approach named Spectral Enhancement for Tiny object detection (SET), which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate the effectiveness of our method. Especially, SET boosts the prior art RFLA by \\(3.2\\%\\) AP on the AI-TOD dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Recent advances in Deep Neural Networks (DNNs) [14, 60] have significantly improved the object detection field [11, 22]. Despite the progress, Tiny Object Detection (TOD) remains a challenging problem. Tiny objects, characterized by their very limited pixel input, occupy areas equal to or less than \\(16 \\times 16\\) pixels [44]. Compared to traditional object detection [10, 21], generic object detectors often fail to manifest effectiveness on the TOD task. For instance," + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.374, + 0.907, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Figure 1. A feature-level occlusion-based attribution analysis in the frequency domain. Mid-frequency features ranges from \\(40\\%\\) to \\(60\\%\\) of the total energy, high-frequency features are above \\(90\\%\\) of the total energy. The model is trained on the AI-TOD [44] trainval and validated on the AI-TOD test. Detection performance is evaluated by inferencing a \\(1\\times\\) FCOS detector [39] with the masked features. Results exhibit that tiny objects are more likely to be affected by high-frequency noise and can benefit from the direct removal of high-frequency information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.747 + ], + "angle": 0, + "content": "DINO [60], one of the most representative transformer-based object detector, achieves \\(37.6\\%\\) AP on medium-sized objects but only \\(9.9\\%\\) AP on very tiny objects on the AITOD [44] benchmark. Moreover, the vanilla FCOS [39] achieves only \\(2.5\\%\\) AP on very tiny objects on the AITOD [44], which is far from sufficient to meet the demands of real-world applications, such as autonomous driving, maritime rescue, and traffic management." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Constrained by the inherently low resolution, one key challenge in TOD is to extract discriminative foreground features [7], especially in the high-level features of detection architecture after down-samplings. In the meantime, it is difficult to learn about tiny objects from the noisy and dominant background clutter [54]. Most existing works [1, 3, 30, 54] focus on enhancing feature representations of tiny objects in the spatial domain. We provide a new analysis method and address the challenge from a frequency spectrum perspective." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.471, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors. Email: sunhuixin@buaa.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.481, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.249, + 0.485, + 0.348 + ], + "angle": 0, + "content": "Figure 2. The analysis examines different objects after feature encoding in the frequency domain. The frequency spectrum is calculated using the Fast Fourier Transform (FFT) [31] on the cropped \\(P_{3}\\) feature and selecting the channel with the highest energy. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.357, + 0.485, + 0.825 + ], + "angle": 0, + "content": "Motivated by occlusion-based attribution analysis [33, 45], we perform an occlusion-based analysis on the intermediate Feature Pyramid Network (FPN) features to examine the impact of masking specific frequency bands on the detection performance across object scales. Specifically, we transform the features into the frequency domain using the n-dimensional Fast Fourier Transform (FFT) [31] and apply static filters to the amplitude spectrums, which are then fed to Inverse Fast Fourier Transform (IFFT) with phase spectrums to generate the masked spatial features. High-frequency mask results in Fig. 1 reveal that directly removing high-frequency features has a positive effect on very tiny and tiny objects, while negatively affecting the detection of larger objects. The disparity can be attributed to that tiny objects become less distinct after feature encoding, exhibiting weak high-frequency components (example (b), Fig. 2) and extremely vague frequency signatures in the cluttered background (example (c), Fig. 2). Therefore, they are more likely to be affected by high-frequency noise and benefit from the direct removal of high-frequency information. In contrast, larger objects exhibit distinct high-frequency signatures (example (a), Fig. 2) and have a higher dependency on high-frequency information. Based on the observations, we investigate an explicit measure to enhance the frequency signatures for tiny objects by removing the high-frequency information in the background. Background (BG) high-frequency mask results show that the approach enhances detection performance across all three object scales, resulting in improvements of \\(15\\%\\), \\(8\\%\\), and \\(4\\%\\), bringing significant gains for very tiny and tiny objects without compromising larger objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.901 + ], + "angle": 0, + "content": "In light of the analysis above, we propose a generic Spectral Enhancement (SET) method to enhance the frequency signatures of Tiny objects using a heterogeneous architecture for foreground and background feature refinement. The method includes two modules. As illustrated in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.288 + ], + "angle": 0, + "content": "Fig. 3, the first Hierarchical Background Smoothing (HBS) module suppresses the high-frequency information in the background through adaptive smoothing operations while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. The second Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. In addition, API facilitates robust feature representations through adversarial training. SET is simple yet effective and can be easily mounted on existing detectors during the training process while bringing no extra burden to the inference procedure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.303 + ], + "angle": 0, + "content": "Our major contributions in this paper are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.304, + 0.906, + 0.424 + ], + "angle": 0, + "content": "- A feature-level occlusion-based attribution analysis from the frequency spectrum perspective is conducted to investigate the Tiny Object Detection (TOD) challenge, which shows that tiny objects are more likely to be affected by high-frequency noise. We introduce a spectral enhancement scheme for tiny object detection (SET) by designing a heterogeneous architecture for foreground and background feature refinement." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.906, + 0.515 + ], + "angle": 0, + "content": "- Two new modules are designed for the TOD task. The HBS module is used to suppress the high-frequency noise in the background through adaptive smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.905, + 0.545 + ], + "angle": 0, + "content": "- Extensive results reveal that our SET boosts the well-known baselines and the prior art by a large margin." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.304, + 0.906, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.657, + 0.575 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Tiny Object Detection. Most current approaches for tiny object detection (TOD) can be grouped into four main categories: data augmentation, multi-scale feature learning, label assignment, and feature enhancement strategies for tiny objects. Several data augmentation strategies [6, 37] target tiny objects by prompting the detector to focus on instances of specific scales during training. Conventional approaches also utilize multi-scale feature learning [23, 26, 62]. Another method line focuses on improving tiny object detection from the label assignment and proposal refinement perspectives [52, 58]. Various works aim to enhance the feature representations of tiny objects. Yuan et al. [58] propose a Feature Imitation (FI) mechanism, aligning RoI features with their counterparts in an exemplar feature set within the embedding space. Recent state-of-the-art TOD methods have proposed auxiliary self-reconstruction branches to enhance the weak representations of tiny objects [3, 54]. Differently, our proposed method enhances the discrimination of tiny objects from a frequency perspective, prompting the refinement of object features during training without adding any extra burden to the inference process." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.506, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.091, + 0.905, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.297, + 0.905, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.444, + 0.908, + 0.502 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed SET method with FCOS framework. The red dashed line denotes generating the adversarial perturbations \\(\\epsilon_{i}^{\\mathrm{adv}}\\). The HBS module can suppress the high-frequency noise in the background through scale-wise smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and facilitate robust feature representations through adversarial training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.486, + 0.729 + ], + "angle": 0, + "content": "Adversarial Training for Object Detection. Many effective attacks crafted for object detectors have been proposed, with most generating adversarial examples at the image level [28, 38, 48]. The most common method to defend against the attacks is adversarial training [20, 29, 40], which involves augmenting training datasets with adversarial examples. For object detection, [59] extend adversarial training to the scenario of object detection by leveraging the attacks sourced from both classification and localization domains. Different from the aforementioned methods, we introduce feature-level adversarial perturbations that balance the detection semantics across branches for enhanced tiny object detection, resulting in more salient foreground features and improved model robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.192, + 0.767 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.903 + ], + "angle": 0, + "content": "This section presents the Spectral Enhancement (SET) method for Tiny object detection, depicted in Fig. 3. SET employs a heterogeneous architecture. This includes the Hierarchical Background Smoothing (HBS) module, which suppresses high-frequency information in the background, and the Adversarial Perturbation Injection (API) module, which enhances the feature saliency of critical regions and prompts the refinement of object features during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.836, + 0.532 + ], + "angle": 0, + "content": "3.1. Hierarchical Background Smoothing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Based on the analysis in Sec. 1, we design a Hierarchical Background Smoothing (HBS) module to suppress the high-frequency information in the background feature while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. This is achieved through background feature decoupling and adaptive smoothing, as described below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Given an input image \\(X\\in \\mathbb{R}^{3\\times W\\times H}\\), a binary mask \\(M\\) is generated according to the ground truth box \\(B\\):" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.677, + 0.905, + 0.695 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} = \\mathbf {1} [ (i, j) \\in B ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.907, + 0.821 + ], + "angle": 0, + "content": "where \\( M \\in \\{0,1\\}^{W \\times H} \\). The indicative function 1 denotes the value of location \\((i,j)\\), which is 1 if it belongs to an object and 0 otherwise. Then we utilize the generated binary mask to decouple image features and perform background smoothing, as shown in Fig. 3. In a Feature Pyramid Network (FPN) comprising \\( N \\) layers, the feature of the \\( i \\)-th layer is denoted as \\(\\mathbf{P}_i\\), \\((i \\in [1,N])\\). The process of HBS on the \\( i \\)-th layer is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.824, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {P} _ {i} ^ {E} = \\mathbf {P} _ {i} ^ {f g} + \\phi_ {i} \\left(\\mathbf {P} _ {i} ^ {b g}, r\\right) \\tag {2} \\\\ = \\mathbf {P} _ {i} * M + \\phi_ {i} (\\mathbf {P} _ {i} * \\neg M, r), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.869, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_i^E\\), \\(\\mathbf{P}_i^{fg}\\) and \\(\\mathbf{P}_i^{bg}\\) denote the enhanced feature, foreground feature and background feature, \\(\\neg M\\) represents the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4715" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.462, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.545 + ], + "angle": 0, + "content": "Figure 4. We visualize (a) the input image with ground-truths shown in green boxes, (b) \\(P_{3}\\) feature in the vanilla FCOS after Principal Component Analysis (PCA) [46], and (c) \\(P_{3}\\) feature in FCOS enhanced with HBS after PCA. Darker colors in the PCA maps indicate more critical components in the feature. As shown, HBS smoothes the background regions and significantly enhances the contrast between the foreground and background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.671 + ], + "angle": 0, + "content": "complement of the binary mask \\(M\\), and \\(\\circledast\\) is the Hadamard product with broadcasting to accommodate the dimensions. Previous works [18, 41] reveal that feature dimensionality reduction methods can lead to smoother and more efficient feature spaces. Based on this idea, we design an adaptive smoothing operation \\(\\phi_i(\\cdot ,r)\\) with channel reduction rate \\(r\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.676, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} \\left(\\mathbf {P} _ {i} ^ {\\mathrm {b g}}, r\\right) = \\sigma \\left(\\mathbf {w} _ {i} ^ {\\mathrm {e}} \\otimes \\sigma \\left(\\mathbf {w} _ {i} ^ {\\mathrm {r}} \\otimes \\mathbf {P} _ {i} ^ {\\mathrm {b g}}\\right)\\right) + \\mathbf {P} _ {i} ^ {\\mathrm {b g}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\otimes\\) represents the convolution operation and \\(\\mathbf{w}_i^{\\mathrm{r}}\\in\\) \\(\\mathbb{R}^{C_i\\times C_{i / r}\\times K_i\\times K_i}\\) and \\(\\mathbf{w}_i^{\\mathrm{e}}\\in \\mathbb{R}^{C_{i / r}\\times C_i\\times K_i\\times K_i}\\) are the weights of the channel reducing and expanding kernels, respectively. \\(\\sigma\\) denotes the non-linear function ReLU. The reduction step compresses information while retaining salient features. When features are expanded, the high-frequency information are difficult to be recreated, resulting in a smoother background. The effectiveness of smoothing is exhibited through statistical analysis in Sec. 4.7. The smoothing effect is further visualized through PCA analysis. As shown in Fig. 4, feature maps enhanced with HBS exhibit a reduced intensity range and smoother transitions in the background regions compared to the vanilla FCOS," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "effectively suppressing noise without losing essential background context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.121, + 0.905, + 0.196 + ], + "angle": 0, + "content": "Kernel size \\( K_{i} \\) at each scale level is determined based on the FPN stride \\( S_{i} \\), where smaller kernels are utilized to remove the fine-grained details in lower feature layers and larger kernels for more coarse-grained noise in higher layers, determined as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.199, + 0.905, + 0.234 + ], + "angle": 0, + "content": "\\[\nK _ {i} = g \\left(S _ {i}\\right) = \\left(\\left\\lfloor \\frac {\\log_ {2} \\left(S _ {i}\\right)}{2} \\right\\rfloor \\times 2\\right) + 1, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.907, + 0.376 + ], + "angle": 0, + "content": "rounding up \\( S_{i} \\) to an odd integer and ensuring central pixel alignment in the convolutional kernel. The padding is determined as \\( (K_{i} - 1) / 2 \\) to preserve the spatial dimensions of the feature map. In addition, we also tested other \\( g(S_{i}) \\), and the results are shown in the Tab. 5. The HBS architecture allows for the suppression of high-frequency information in the background while preserving foreground details, significantly enhancing the frequency signatures of tiny objects amid high-frequency noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.818, + 0.401 + ], + "angle": 0, + "content": "3.2. Adversarial Perturbation Injection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.906, + 0.527 + ], + "angle": 0, + "content": "Analysis in Sec. 1 reveals that tiny objects exhibit weak high-frequency signatures after feature encoding and low discrimination against the cluttered background. To address this, we design an Adversarial Perturbation Injection (API) module, which leverages adversarial perturbations to enhance the saliency in critical regions. In addition, API promotes robust feature representations for tiny objects through adversarial training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.602 + ], + "angle": 0, + "content": "Inspired by prior works in adversarial training [16, 17, 42], we exploit feature-level perturbations that adversarially change the model output to enhance feature representations. The objective of API for each FPN layer is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.607, + 0.905, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {P} _ {i}, \\theta_ {i}} \\left(\\max _ {\\| \\epsilon_ {i, \\mathrm {c l s}} \\| \\leq \\rho} \\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i} + \\epsilon_ {i, \\mathrm {c l s}}\\right) + \\gamma \\| \\mathbf {P} _ {i} \\| _ {2} ^ {2}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.906, + 0.754 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{cls}}\\) denotes the classification loss. The inner optimization injects adversarial perturbation \\(\\epsilon_{i,\\mathrm{cls}}\\) into the feature space of \\(\\mathbf{P}_i\\), where \\(\\rho\\) defines the perturbation size, and \\(\\gamma\\) is a hyperparameter that controls the strength of regularization. \\(\\theta_{i}\\) denotes the model parameters at the \\(i\\)-th layer. We derive a closed-form solution for the adversarial perturbation under the \\(L_{2}\\) norm following [12]:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.76, + 0.905, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {i, \\mathrm {c l s}} ^ {*} \\approx \\rho \\cdot \\frac {\\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {P} _ {i})}{\\| \\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {P} _ {i}) \\| _ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.8, + 0.905, + 0.859 + ], + "angle": 0, + "content": "The perturbation can elevate the gradient of critical regions, thereby increasing their saliency during training. The gradient change induced by the perturbation on the classification branch can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.87, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\Delta \\left(\\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}}\\right) = \\left\\| \\nabla_ {\\mathbf {P} _ {i}} \\left(\\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i} + \\epsilon_ {i, \\mathrm {c l s}} ^ {*}\\right) - \\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i}\\right)\\right) \\right\\|. \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4716" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.346, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.09, + 0.59, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.091, + 0.905, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.199, + 0.346, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.199, + 0.59, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.199, + 0.905, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 5. Average saliency visualization in (a) FCOS and (b) FCOS w/ API. The saliency maps are first computed based on the \\(\\ell_2\\)-norm of the gradient [13] at neck feature \\(P_{3}\\). The saliency of each object is then summed and annotated within their respective bounding box, which is different between (a) and (b). The average saliency of each object scale is calculated as the mean saliency across objects in the corresponding size category, displayed in the bottom-left corner of each image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.392, + 0.483, + 0.589 + ], + "angle": 0, + "content": "High \\(\\Delta (\\nabla_{\\mathbf{P}_i}\\mathcal{L}_{\\mathrm{cls}})\\) corresponds to regions where model outputs are highly sensitive to minor changes in activations, indicating the presence of semantic information about the image [49]. This elevated gradient enhances the feature saliency of these critical regions and prompts the refinement of object features during training. The effect particularly benefits tiny objects as they exhibit weak high-frequency signatures and face a training deficiency problem. The enhanced object saliency phenomenon is discussed in Sec. 4.6 through Fig. 5. Furthermore, the adversarial training facilitates robust feature representations and reduces the sensitivity to high-frequency noise [63], thereby enhancing the discrimination of tiny objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.679 + ], + "angle": 0, + "content": "We formulate the adversarial perturbation \\(\\epsilon_{i}^{\\mathrm{adv}}\\) for each FPN layer as an expectation over \\(M\\) object detection branches (e.g., the classification, regression and center-ness branch in a FCOS detector [39]), each scaled by a corresponding balancing parameter \\(\\lambda_{m}\\). This is generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.677, + 0.483, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {i} ^ {\\mathrm {a d v}} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} \\cdot \\epsilon_ {i, m} ^ {*}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.802 + ], + "angle": 0, + "content": "where \\(\\epsilon_{i,m}^{*}\\) represents the adversarial perturbation for the \\(m\\)-th branch. The design allows for a balanced incorporation of detection semantics across multiple branches, ensuring that the perturbations are distributed for more effective training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.307, + 0.833 + ], + "angle": 0, + "content": "3.3. Auxiliary Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The optimization leverages a total loss function that integrates both the original detection loss \\(\\mathcal{L}_{\\mathrm{detection}}\\) and an auxiliary loss derived from the HBS and API enhanced features. This joint loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.389, + 0.907, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\underbrace {\\lambda \\cdot \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {i} ^ {\\mathrm {a u x}} \\left(\\mathbf {P} _ {i} ^ {E} + \\epsilon_ {i} ^ {\\mathrm {a d v}}\\right)} _ {\\text {a u x i l i a r y l o s s}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.635 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_i^{\\mathrm{aux}}\\) denotes the auxiliary loss component for the \\(i\\)-th layer, which is summed and scaled by a balancing hyperparameter \\(\\lambda\\). The perturbation \\(\\epsilon_i^{\\mathrm{adv}}\\) is applied to the HBS-enhanced feature \\(\\mathbf{P}_i^E\\). The auxiliary optimization process updates the model's parameters in concert with the original loss, leading to more robust and discriminative features for tiny objects. We visualize the effectiveness of the method in Fig. 2. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS, benefiting their detection amid high-frequency noise and in the presence of larger objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.647, + 0.646, + 0.664 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.834, + 0.687 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed method on AI-TOD [44], VisDrone2019 [65], DOTA-v2.0 [47] and COCO 2017 [22]. The main experiments were conducted using the AI-TOD dataset, with an average object size of 12.8 pixels, substantially smaller than other object detection datasets such as PASCAL VOC (156.6 pixels) [10] and MS COCO (99.5 pixels) [21]. Furthermore, we test our method on the VisDrone2019 [65] and DOTA-v2.0 [47], which comprises high-resolution drone-shot images with a substantial proportion of tiny instances. We also conduct experiments on the general object detection benchmark COCO 2017 [22], incorporating \\(\\mathrm{AP}_{vt}\\), \\(\\mathrm{AP}_t\\), \\(\\mathrm{AP}_s\\) metrics to evaluate objects of very tiny (2-8 pixels), tiny (8-16 pixels) and small (16-32 pixels), following [44]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4717" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Main results with various frameworks on AI-TOD [44]. Models are trained on the AI-TOD trainval and validated on the AI-TOD test. We report APs (\\%) with different IoU thresholds and APs (\\%) for objects of various sizes based on the AI-TOD criterion. RFLA is based on Cascade R-CNN. The * denotes using P2~P6 FPN features. The **bold** indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.135, + 0.862, + 0.562 + ], + "angle": 0, + "content": "
FrameworkAP\\( AP_{0.5} \\)\\( AP_{0.75} \\)\\( AP_{vt} \\)\\( AP_t \\)\\( AP_s \\)
PAA [19]10.026.56.73.510.513.1
ATSS [61]11.628.57.62.511.915.9
Centernet [9]13.439.25.03.812.117.7
DetectoRS [34]14.832.811.50.010.818.3
DotD [50]16.139.210.68.317.618.1
DetectoRS w/NWD [43]20.849.314.36.419.729.6
DetectoRS w/ SR-TOD [3]24.054.617.110.124.829.3
One-stage
RetinaNet [24]7.220.53.62.56.811.0
RetinaNet w/ SET9.124.65.22.98.212.8
RetinaNet* [24]13.331.78.63.914.417.7
RetinaNet* w/ QueryDet [55]12.229.37.32.410.518.5
RetinaNet* w/ SET16.538.610.95.015.721.8
FCOS [39]12.029.08.02.511.917.1
FCOS w/ SET14.234.99.82.913.020.2
FCOS* [39]15.135.810.25.916.618.8
FCOS* w/ SET18.042.712.47.218.523.0
Multi-stage
Faster R-CNN [35]11.126.38.10.07.223.3
Faster R-CNN w/ SET12.228.29.70.19.124.3
Cascade R-CNN [2]13.630.310.60.09.925.5
Cascade R-CNN w/ SET14.832.411.20.410.826.7
RFLA [52]21.750.515.38.321.824.5
RFLA w/ SR-TOD [3]21.850.815.49.721.827.4
RFLA w/ SET24.955.617.99.825.231.0
Transformer-based
DINO-5scale [60]23.256.615.49.923.129.3
DINO-5scale w/ DNTR [25]26.256.720.212.826.431.0
DINO-5scale w/ SET26.657.120.813.227.131.5
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.575, + 0.482, + 0.818 + ], + "angle": 0, + "content": "Implementation Details. We conducted the experiments on a computer with an NVIDIA RTX 3090 GPU. The experiments are implemented with PyTorch [32], with core codes built upon MMdetection [5]. The ImageNet [36] pre-trained models are used as the backbones. All CNN-based models utilize the ResNet-50 [15] backbone, trained using the Stochastic Gradient Descent (SGD) optimizer for 12 epochs with 0.9 momentum, 0.0001 weight decay, and a batch size 2. The initial learning rate is 0.005, decaying at the 8th and 11th epochs. The data processing adheres to the default configurations of each dataset (e.g., fixed at \\(800 \\times 800\\) for AI-TOD). We also train a transformer-based detector, DINO [60], with 5-scale feature maps for 36 epochs as a baseline. The training uses an Adam optimizer with a weight decay of 0.0001, following the random crop and scale augmentation strategies of DETR [4]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.276, + 0.846 + ], + "angle": 0, + "content": "4.2. Results on AI-TOD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed SET method on various detectors and compare it with state-of-the-art TOD methods on the AI-TOD benchmark [51]. Tab. 1 shows that SET enhances" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.788 + ], + "angle": 0, + "content": "all baselines by about \\(2\\%\\) AP, which is significant. Specifically, SET improves the one-stage detector FCOS [39] and RetinaNet [24] by \\(2.2\\% / 1.9\\%\\) AP and \\(1.1\\% / 1.4\\%\\) \\(\\mathrm{AP}_t\\), respectively. We further revise the detector architecture and use P2~P6 FPN features, a representative approach for TOD tasks where the high-resolution P2 layer benefits tiny objects. With the adjustment, SET boosts the RetinaNet and FCOS baselines by \\(3.2\\% / 2.9\\%\\) AP and \\(1.1\\% / 1.3\\% \\mathrm{AP}_{vt}\\) a significant margin. In contrast, the state-of-the-art SOD method QueryDet [55] decreases the baseline performance, showing limitations in addressing the TOD challenge. SET also generalizes to multi-stage detectors, enhancing the Faster R-CNN [35] and Cascade R-CNN [2] baselines by \\(1.1\\% / 1.2\\%\\) AP and \\(1.9\\% / 1.1\\%\\) \\(\\mathrm{AP}_t\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "SET works orthogonally with the state-of-the-art TOD method RFLA [52], achieving a \\(3.2\\%\\) AP increment over the RFLA baseline and surpassing the comparison SR-TOD [3] by \\(3.1\\%\\) AP and \\(3.4\\%\\) \\(\\mathrm{AP}_t\\). We also investigate SET's compatibility with transformer-based detectors. As shown, SET results in \\(3.4\\%\\) AP increase over the DINO-5scale [60] baseline, outperforming the prior" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4718" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 2. Detection performance on the VisDrone2019 [65] validation set. RFLA is based on Cascade R-CNN. The **bold** denotes the best result." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.142, + 0.486, + 0.288 + ], + "angle": 0, + "content": "
FrameworkAPAPvtAPtAPs
RetinaNet [24]15.20.41.69.3
Faster R-CNN [35]23.10.15.119.6
FCOS [39]19.90.74.515.7
FCOS w/ SET21.91.65.717.9
Cascade R-CNN [2]23.00.15.019.4
Cascade R-CNN w/ SET24.31.26.321.0
RFLA [52]27.24.513.023.6
RFLA w/ SR-TOD [3]27.84.813.124.5
RFLA w/ SET28.55.213.524.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.485, + 0.368 + ], + "angle": 0, + "content": "art DNTR [25]. Notably, DINO-5scale w/SET derives \\(26.6\\%\\) AP, outperforming competitors including DotD [50], NWD [43], and SR-TOD [3]. The results demonstrate SET as a robust solution for the tiny object detection task." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.379, + 0.45, + 0.395 + ], + "angle": 0, + "content": "4.3. Results on VisDrone2019 and DOTA-V2.0" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.483, + 0.569 + ], + "angle": 0, + "content": "We further evaluate SET on the drone-shot image benchmark VisDrone2019 [65], which features tiny and small objects. As shown in Tab. 2, the proposed SET effectively improves the performance of both one-stage and multi-stage detectors and exhibits consistent \\(\\mathrm{AP}_{vt}\\) and \\(\\mathrm{AP}_t\\) improvements. Specifically, SET improved the the one-stage detector FCOS by \\(2.0\\% /0.9\\% /1.2\\%\\) AP, \\(\\mathrm{AP}_{vt}\\) and \\(\\mathrm{AP}_t\\) , respectively. SET also improves multi-stage detector Cascade RCNN [2] by \\(1.1\\% /1.3\\%\\) \\(\\mathrm{AP}_{vt}\\) and \\(\\mathrm{AP}_t\\) , respectively. Notably, SET improves the prior art RFLA, achieving a \\(28.5\\%\\) AP with \\(0.7\\%\\) increment in \\(\\mathrm{AP}_{vt}\\) and \\(0.5\\%\\) in \\(\\mathrm{AP}_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.707 + ], + "angle": 0, + "content": "Tab. 3 demonstrates the DOTA-v2.0 [8] results. As shown, SET obtains a substantial performance gain of \\(3.1\\%\\) and \\(1.3\\%\\) in \\(\\mathrm{AP}_t\\) with the FCOS and AutoAssign baseline, underscoring the efficacy of our method in detecting tiny objects. Compared to the state-of-the-art comparisons RFLA [52] and DCFL [53], SET achieves \\(2.6\\%\\) and \\(0.9\\%\\) AP gain, respectively. More experimental results on the TinyPerson [56] and SeaPerson [57] dataset are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.267, + 0.735 + ], + "angle": 0, + "content": "4.4. Results on COCO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.835 + ], + "angle": 0, + "content": "We further conduct experiments on the general object detection benchmark MS COCO [21]. As shown in Tab. 9, SET achieves \\(1.0\\%\\) AP improvement over the FCOS baseline, with a notable \\(1.7\\%\\) improvement in \\(\\mathrm{AP}_t\\), demonstrating that SET can enhance the detection of tiny objects in general detection tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.245, + 0.864 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In the following experiments, we explore the best structure of SET by ablating and tuning its parts using AI-TOD." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 3. Detection performance on DOTA-v2.0 [8]. Note that models are trained on the DOTA-v2.0 train and validated on the DOTA-v2.0 val. The bold denotes the best result." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.142, + 0.909, + 0.273 + ], + "angle": 0, + "content": "
FrameworkAPAPvtAPtAPs
ATSS [61]32.70.76.923.4
Faster R-CNN [35]35.60.07.128.9
FCOS [39]31.80.34.019.4
FCOS w/ RFLA [52]32.10.76.823.6
FCOS w/ DCFL [53]33.80.56.924.2
FCOS w/ SET34.70.67.124.8
AutoAssign [64]33.80.97.322.4
AutoAssign w/ SET35.21.18.623.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.282, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Table 4. Detection performance on MS COCO [21]. Note that models are trained on COCO train2017 and validated on COCO val2017. \\((\\cdot)^{+}\\) denotes performance increment." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.334, + 0.917, + 0.381 + ], + "angle": 0, + "content": "
FrameworkAPAPvtAPtAPsAPm
FCOS [39]36.47.919.627.243.6
FCOS w/ SET37.4+1.08.3+0.423.3+1.727.6+0.444.8+1.2
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.402, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Effectiveness and Cost of Components. We present quantitative improvements of SET components in Tab. 5. As shown, applying the adaptive smoothing operations to full features \\((+\\phi_{i}(\\mathbf{P}_{i},r))\\) achieves a \\(0.4\\%\\) increase in AP and \\(0.3\\%\\) in \\(\\mathrm{AP}_{vt}\\), surpassing the performance of static high-frequency filtering in Fig. 1. With decoupling, the HBS module boosts the overall AP by \\(1.9\\%\\) AP and \\(0.4\\%\\) \\(\\mathrm{AP}_{vt}\\). The API module alone achieves a \\(1.2\\%\\) AP improvement. Combining HBS and API results in a total AP increase of \\(2.2\\%\\). Results demonstrate the synergy between the heterogeneous operations, where HBS effectively smooths the clutter background and API prompts the learning of object features, cumulatively improving the detection accuracy of tiny objects. Tab. 5 also presents the cost of components. Results show that the HBS module leads to \\(17\\%\\), \\(0.9\\%\\), and \\(0.9\\%\\) increase in training time, computational and memory cost. The API module results in a more significant increase in training time due to adversarial training but adds no forward computational cost and minimal increase in memory usage. The overall minimal impact indicates that SET can be implemented with negligible additional overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Hyper-parameter Selection. We first select the hyperparameters channel reduction rate \\( r \\) and perturbation size \\( \\rho \\) using the FCOS [39] framework trained three epochs in Fig. 6 (a). Model performance (AP) with different hyperparameter combinations \\( \\{r, \\rho\\} \\) is presented. Results indicate that performance initially improves and then declines as \\( r \\) varies from 2 to 16, with \\( r = 16 \\) perform worse than the baseline. Additionally, SET with adversarial perturbations \\( (\\rho > 0) \\) exhibits stronger performance than without \\( (\\rho = 0) \\), but large-sized perturbations \\( (\\rho = 5) \\) performs worse than all alternatives. Exploring the setups, we identify \\( \\{r, \\rho\\} = \\{4, 1\\} \\) as the combination for SET. Fig. 6 (b) explores the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4719" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 5. The impact of different components and various channel reduction operations in SET. The hyper-parameters are fixed with channel reduction rate \\( r \\) being 4, perturbation size \\( \\rho \\) being 1, and the balancing hyper-parameters \\( \\lambda \\) of auxiliary optimization being 1." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.128, + 0.907, + 0.288 + ], + "angle": 0, + "content": "
Frameworkφig(Si)APAPvtAPtTime(s/batch)GFLOPsParams (M)
FCOS [39]--12.02.511.90.17126.1337.04
+ φi(Pi, r)--12.72.712.90.17127.3637.41
Eqn. 3313.42.712.6
Eqn. 3log2(Si)/2×2+113.92.712.9
+ HBSEqn. 3513.12.312.50.20127.3237.41
Eqn. 3log2(Si+1)/2×2-113.42.812.4
Eqn. 3712.12.512.2
+ API--13.22.612.40.39126.1337.16
+ Both (SET)Eqn. 3Eqn. 414.22.913.00.42127.3237.85
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.303, + 0.289, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.303, + 0.482, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.505 + ], + "angle": 0, + "content": "Figure 6. We select hyper-parameters channel reduction rate \\( r \\) with perturbation size \\( \\rho \\) in (a), and the balancing hyper-parameter \\( \\lambda \\) of Eqn. 9 in (b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.683 + ], + "angle": 0, + "content": "balancing hyper-parameter \\(\\lambda\\) in Eqn. 9 that controls the proportion of the auxiliary optimization, using the FCOS [39], RetinaNet [24], and Faster R-CNN [35] framework. As observed, the performances increase first and then decrease with the increase of \\(\\lambda\\), with \\(\\lambda = 1\\) yielding the best performance across all frameworks. However, the efficacy of SET diminishes when \\(\\lambda\\) exceeds the 1, as larger-sized \\(\\lambda\\) (\\(\\lambda > 1\\)) can impose too strong regularization to the detector, reducing the detection accuracy despite increasing robustness. Therefore, we set the hyper-parameter \\(\\lambda\\) to 1 for experiments in this paper." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.33, + 0.714 + ], + "angle": 0, + "content": "4.6. Enhanced Object Saliency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We demonstrate effectiveness of API in increasing the saliency of object features in Fig. 5. Comparing the average saliency of each object scale, tiny objects are the least salient due to their inherently limited pixel input. API leads to a notable increase in the average saliency of tiny objects (e.g., from 53.99 to 69.93), benefited from the injection of perturbations (Eqn. 7). Fig. 5 also demonstrates that API can enhance the feature saliency of small and medium objects, though the degree of improvement is less. The comparison underscores that API could benefit general object detection, with more significant and essential improvements for tiny objects." + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.307, + 0.866, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.453, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Figure 7. The analysis examines the adaptive smoothing operation, obtained from 1000 randomly selected images from the Vis-Drone2019 [65] train set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.757, + 0.523 + ], + "angle": 0, + "content": "4.7. Effectiveness in Smoothing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.712 + ], + "angle": 0, + "content": "In Fig. 7, we perform a statistic analysis on \\( P_{i}^{bg} \\) to investigate the adaptive smoothing operation. We convert feature \\( P_{i}^{bg} \\) and \\( \\phi_{i}(P_{i}^{bg}, r) \\) into the frequency space using FFT [31] and separate the amplitude spectrums of these features into three parts (low-, mid-, and high-frequency) using static thresholds as Fig. 1. The amplitudes in each band are then summed and averaged to measure the signal strength. Results verify that the channel reducing and expand design effects in smoothing. Comparing \\( A(P_{i})(u, v) \\) and \\( A(\\phi_{i}(P_{i}^{bg}, r))(u, v) \\), the low-frequency signals are enhanced (from 0.3787 to 0.5205), while mid- and high-frequency signals are attenuated." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This paper proposes the Spectral Enhancement for Tiny objects (SET) method, which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate its effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4720" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.248 + ], + "angle": 0, + "content": "Acknowledgments. The work was supported by the National Natural Science Foundation of China (No. 61827901, 623B2016, and 62406298). This research was also supported by the National Key Research and Development Program of China (Grant No. 2023YFC3306401), Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F020007, Beijing Natural Science Foundation L223024, L244043 and Z241100001324017, \"One Thousand Plan\" projects in Jiangxi Province Jxsq2023102268, and the Beijing Jiaotong University \"Jingying Plan\" No. K24XKRC00130." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.275, + 0.188, + 0.29 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.3, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, pages 206-221, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[3] Bing Cao, Haiyu Yao, Pengfei Zhu, and Qinghua Hu. Visible and clear: Finding tiny objects in difference map. ECCV, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.498 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[6] Yukang Chen, Peizhen Zhang, Zeming Li, Yanwei Li, Xiangyu Zhang, Lu Qi, Jian Sun, and Jiaya Jia. Dynamic scale training for object detection. arXiv preprint arXiv:2004.12432, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[7] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.756 + ], + "angle": 0, + "content": "[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE TPAMI, 44(11):7778-7796, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.759, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[9] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, pages 6569-6578, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[10] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. ICCV, 88:303-308, 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.3, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Kaiming He and Jian Sun. Convolutional neural networks at constrained time cost. In CVPR, pages 5353-5360, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.904, + 0.319 + ], + "angle": 0, + "content": "[16] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[17] Masato Ishii and Atsushi Sato. Training deep neural networks with adversarially augmented features for small-scale training datasets. In IJCNN, pages 1-8. IEEE, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[18] Weikuan Jia, Meili Sun, Jian Lian, and Sujuan Hou. Feature dimensionality reduction: a review. Complex & Intelligent Systems, 8(3):2663-2693, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[19] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.904, + 0.476 + ], + "angle": 0, + "content": "[20] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR*, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 1, 5, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.535, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 6, 7, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[25] Hou-I Liu, Yu-Wen Tseng, Kai-Cheng Chang, Pin-Jyun Wang, Hong-Han Shuai, and Wen-Huang Cheng. A denoising fpn with transformer r-cnn for tiny object detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[26] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[28] Jiajun Lu, Hussein Sibai, and Evan Fabry. Adversarial examples that fool detectors. arXiv preprint arXiv:1712.02494, 2017.3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4721" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] T. Na, J. H. Ko, and S. Mukhopadhyay. Cascade adversarial machine learning regularized with a unified embedding. In ICLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[30] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In ICCV, pages 9725-9734, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[31] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS Workshops, pages 1-12, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In CVPR, pages 11443-11452, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[34] Siyuan Qiao, Liang-Chieh Chen, and Alan Yuille. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In CVPR, pages 10213-10224, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(06):1137-1149, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. ICCV, 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[37] Bharat Singh, Mahyar Najibi, Abhishek Sharma, and Larry S Davis. Scale normalized image pyramids with autofocus for object detection. IEEE TPAMI, 44(7):3749-3766, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[38] Dawn Song, Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Florian Tramer, Atul Prakash, and Tadayoshi Kohno. Physical adversarial examples for object detectors. In 12th USENIX workshop on offensive technologies (WOOT '18), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 5, 6, 7, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[40] F Tramér, D Boneh, A Kurakin, I Goodfellow, N Papernot, and P McDaniel. Ensemble adversarial training: Attacks and defenses. In *ICLR*, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[41] S Velliangiri, SJPCS Alagumuthukrishnan, et al. A review of dimensionality reduction techniques for efficient computation. Procedia Computer Science, 165:104-111, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[42] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, pages 5495-5504, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[43] Jinwang Wang, Chang Xu, Wen Yang, and Lei Yu. A normalized gaussian Wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[44] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In ICPR, pages 3791-3798, 2021. 1, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[45] Kunyu Wang, Xueyang Fu, Yukun Huang, Chengzhi Cao, Gege Shi, and Zheng-Jun Zha. Generalized uav object detection via frequency domain disentanglement. In CVPR, pages 1064-1073, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[47] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, pages 3974-3983, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[48] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In ICCV, pages 1369-1378, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[49] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In CVPR, pages 501-509, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[50] Chang Xu, Jinwang Wang, Wen Yang, and Lei Yu. Dot distance for tiny object detection in aerial images. In ICCV, pages 1192-1201, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[51] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Detecting tiny objects in aerial images: A normalized wasserstein distance and a new benchmark. IS-PRS J. Photo. and Rem. Sen., 190:79-93, 2022. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[52] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543, 2022. 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[53] Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Dynamic coarse-to-fine learning for oriented tiny object detection. In CVPR, pages 7318-7328, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.562, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[54] Sheng Xu, Mingze Wang, Yanjing Li, Mingbao Lin, Baochang Zhang, David Doermann, and Xiao Sun. Learning 1-bit tiny object detector with discriminative feature refinement. In ICML, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.617, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[55] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[56] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In WACV, pages 1257–1265, 2020. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.717, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[57] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. Cpr++: Object localization via single coarse point supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4908-4925, 2024. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.788, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[58] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small object detection via coarse-to-fine proposal generation and imitation learning. In ICCV, pages 6317-6327, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.845, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[59] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, pages 421-430, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4722" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[61] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[62] Qijie Zhao, Tao Sheng, Yongtao Wang, Zhi Tang, Ying Chen, Ling Cai, and Haibin Ling. M2det: A single-shot object detector based on multi-level feature pyramid network. In AAAI, pages 9259-9266, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[63] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, pages 452-467, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[64] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020. 7, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[65] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Haibin Ling, Qinghua Hu, Qinqin Nie, Hao Cheng, Chenfeng Liu, Xiaoyu Liu, et al. Visdrone-det2018: The vision meets drone object detection in image challenge results. In ECCV Workshops, pages 0–0, 2018. 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4723" + } + ] +] \ No newline at end of file diff --git a/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_origin.pdf b/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c1050aa322466109cc3174d5bd49ccdbca30100 --- /dev/null +++ b/2025/SET_ Spectral Enhancement for Tiny Object Detection/70fd0953-a894-43a1-b9c6-1b75dbe9ffb2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e754e58152998af68701916e74936e45758aeabfa2c55e72f2b77c263e33c744 +size 946260 diff --git a/2025/SET_ Spectral Enhancement for Tiny Object Detection/full.md b/2025/SET_ Spectral Enhancement for Tiny Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..041f9115b84f1878cfbd32dcc91ede7899dc5909 --- /dev/null +++ b/2025/SET_ Spectral Enhancement for Tiny Object Detection/full.md @@ -0,0 +1,311 @@ +# SET: Spectral Enhancement for Tiny Object Detection + +Huixin Sun $^{1}$ , Runqi Wang $^{2}$ , Yanjing Li $^{1}$ , Linlin Yang $^{5}$ , Shaohui Lin $^{6}$ , Xianbin Cao $^{1*}$ , Baochang Zhang $^{3,4*}$ + +$^{1}$ School of Electronic Information Engineering, Beihang University, Beijing, China + +$^{2}$ School of Computer Science and Technology, Beijing Jiaotong University + +$^{3}$ School of Artificial Intelligence, Beihang University, Beijing, China + +$^{4}$ Zhongguancun Laboratory, Beijing, China + +${}^{5}$ State Key Laboratory of Media Convergence and Communication, + +Communication University of China, Beijing, China + +$^{6}$ School of Computer Science and Technology, East China Normal University, Shanghai, China + +# Abstract + +Deep learning has significantly advanced the object detection field. However, tiny object detection (TOD) remains a challenging problem. We provide a new analysis method to examine the TOD challenge through occlusion-based attribution analysis in the frequency domain. We observe that tiny objects become less distinct after feature encoding and can benefit from the removal of high-frequency information. In this paper, we propose a novel approach named Spectral Enhancement for Tiny object detection (SET), which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate the effectiveness of our method. Especially, SET boosts the prior art RFLA by $3.2\%$ AP on the AI-TOD dataset. + +# 1. Introduction + +Recent advances in Deep Neural Networks (DNNs) [14, 60] have significantly improved the object detection field [11, 22]. Despite the progress, Tiny Object Detection (TOD) remains a challenging problem. Tiny objects, characterized by their very limited pixel input, occupy areas equal to or less than $16 \times 16$ pixels [44]. Compared to traditional object detection [10, 21], generic object detectors often fail to manifest effectiveness on the TOD task. For instance, + +![](images/1e0b8c54b258c0a80bd94592a0ff244e9d8820ecbd3957f12f3786a192ad2c7e.jpg) +Figure 1. A feature-level occlusion-based attribution analysis in the frequency domain. Mid-frequency features ranges from $40\%$ to $60\%$ of the total energy, high-frequency features are above $90\%$ of the total energy. The model is trained on the AI-TOD [44] trainval and validated on the AI-TOD test. Detection performance is evaluated by inferencing a $1\times$ FCOS detector [39] with the masked features. Results exhibit that tiny objects are more likely to be affected by high-frequency noise and can benefit from the direct removal of high-frequency information. + +DINO [60], one of the most representative transformer-based object detector, achieves $37.6\%$ AP on medium-sized objects but only $9.9\%$ AP on very tiny objects on the AITOD [44] benchmark. Moreover, the vanilla FCOS [39] achieves only $2.5\%$ AP on very tiny objects on the AITOD [44], which is far from sufficient to meet the demands of real-world applications, such as autonomous driving, maritime rescue, and traffic management. + +Constrained by the inherently low resolution, one key challenge in TOD is to extract discriminative foreground features [7], especially in the high-level features of detection architecture after down-samplings. In the meantime, it is difficult to learn about tiny objects from the noisy and dominant background clutter [54]. Most existing works [1, 3, 30, 54] focus on enhancing feature representations of tiny objects in the spatial domain. We provide a new analysis method and address the challenge from a frequency spectrum perspective. + +![](images/484f014fad10e1ba08980fd642514dbdb723c4613adf3cecc7d6b49c3d76c51c.jpg) +Figure 2. The analysis examines different objects after feature encoding in the frequency domain. The frequency spectrum is calculated using the Fast Fourier Transform (FFT) [31] on the cropped $P_{3}$ feature and selecting the channel with the highest energy. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS. + +Motivated by occlusion-based attribution analysis [33, 45], we perform an occlusion-based analysis on the intermediate Feature Pyramid Network (FPN) features to examine the impact of masking specific frequency bands on the detection performance across object scales. Specifically, we transform the features into the frequency domain using the n-dimensional Fast Fourier Transform (FFT) [31] and apply static filters to the amplitude spectrums, which are then fed to Inverse Fast Fourier Transform (IFFT) with phase spectrums to generate the masked spatial features. High-frequency mask results in Fig. 1 reveal that directly removing high-frequency features has a positive effect on very tiny and tiny objects, while negatively affecting the detection of larger objects. The disparity can be attributed to that tiny objects become less distinct after feature encoding, exhibiting weak high-frequency components (example (b), Fig. 2) and extremely vague frequency signatures in the cluttered background (example (c), Fig. 2). Therefore, they are more likely to be affected by high-frequency noise and benefit from the direct removal of high-frequency information. In contrast, larger objects exhibit distinct high-frequency signatures (example (a), Fig. 2) and have a higher dependency on high-frequency information. Based on the observations, we investigate an explicit measure to enhance the frequency signatures for tiny objects by removing the high-frequency information in the background. Background (BG) high-frequency mask results show that the approach enhances detection performance across all three object scales, resulting in improvements of $15\%$ , $8\%$ , and $4\%$ , bringing significant gains for very tiny and tiny objects without compromising larger objects. + +In light of the analysis above, we propose a generic Spectral Enhancement (SET) method to enhance the frequency signatures of Tiny objects using a heterogeneous architecture for foreground and background feature refinement. The method includes two modules. As illustrated in + +Fig. 3, the first Hierarchical Background Smoothing (HBS) module suppresses the high-frequency information in the background through adaptive smoothing operations while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. The second Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. In addition, API facilitates robust feature representations through adversarial training. SET is simple yet effective and can be easily mounted on existing detectors during the training process while bringing no extra burden to the inference procedure. + +Our major contributions in this paper are summarized as: + +- A feature-level occlusion-based attribution analysis from the frequency spectrum perspective is conducted to investigate the Tiny Object Detection (TOD) challenge, which shows that tiny objects are more likely to be affected by high-frequency noise. We introduce a spectral enhancement scheme for tiny object detection (SET) by designing a heterogeneous architecture for foreground and background feature refinement. +- Two new modules are designed for the TOD task. The HBS module is used to suppress the high-frequency noise in the background through adaptive smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. +- Extensive results reveal that our SET boosts the well-known baselines and the prior art by a large margin. + +# 2. Related Work + +Tiny Object Detection. Most current approaches for tiny object detection (TOD) can be grouped into four main categories: data augmentation, multi-scale feature learning, label assignment, and feature enhancement strategies for tiny objects. Several data augmentation strategies [6, 37] target tiny objects by prompting the detector to focus on instances of specific scales during training. Conventional approaches also utilize multi-scale feature learning [23, 26, 62]. Another method line focuses on improving tiny object detection from the label assignment and proposal refinement perspectives [52, 58]. Various works aim to enhance the feature representations of tiny objects. Yuan et al. [58] propose a Feature Imitation (FI) mechanism, aligning RoI features with their counterparts in an exemplar feature set within the embedding space. Recent state-of-the-art TOD methods have proposed auxiliary self-reconstruction branches to enhance the weak representations of tiny objects [3, 54]. Differently, our proposed method enhances the discrimination of tiny objects from a frequency perspective, prompting the refinement of object features during training without adding any extra burden to the inference process. + +![](images/bdc484cd6bab9053a4ed6b8c44e788405c0eef8fc2501b90cf20c24b057b10ff.jpg) + +![](images/94f5b3567065d32c04c37661336e95e48327e29f292a22c2ef6a866435461421.jpg) + +![](images/1a8391d7369fcd36840858097f8a68f58c6802a9589f046895e7fd1ab20670af.jpg) +Figure 3. Overview of the proposed SET method with FCOS framework. The red dashed line denotes generating the adversarial perturbations $\epsilon_{i}^{\mathrm{adv}}$ . The HBS module can suppress the high-frequency noise in the background through scale-wise smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and facilitate robust feature representations through adversarial training. + +Adversarial Training for Object Detection. Many effective attacks crafted for object detectors have been proposed, with most generating adversarial examples at the image level [28, 38, 48]. The most common method to defend against the attacks is adversarial training [20, 29, 40], which involves augmenting training datasets with adversarial examples. For object detection, [59] extend adversarial training to the scenario of object detection by leveraging the attacks sourced from both classification and localization domains. Different from the aforementioned methods, we introduce feature-level adversarial perturbations that balance the detection semantics across branches for enhanced tiny object detection, resulting in more salient foreground features and improved model robustness. + +# 3. Methods + +This section presents the Spectral Enhancement (SET) method for Tiny object detection, depicted in Fig. 3. SET employs a heterogeneous architecture. This includes the Hierarchical Background Smoothing (HBS) module, which suppresses high-frequency information in the background, and the Adversarial Perturbation Injection (API) module, which enhances the feature saliency of critical regions and prompts the refinement of object features during training. + +# 3.1. Hierarchical Background Smoothing + +Based on the analysis in Sec. 1, we design a Hierarchical Background Smoothing (HBS) module to suppress the high-frequency information in the background feature while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. This is achieved through background feature decoupling and adaptive smoothing, as described below. + +Given an input image $X\in \mathbb{R}^{3\times W\times H}$ , a binary mask $M$ is generated according to the ground truth box $B$ : + +$$ +M _ {i, j} = \mathbf {1} [ (i, j) \in B ], \tag {1} +$$ + +where $M \in \{0,1\}^{W \times H}$ . The indicative function 1 denotes the value of location $(i,j)$ , which is 1 if it belongs to an object and 0 otherwise. Then we utilize the generated binary mask to decouple image features and perform background smoothing, as shown in Fig. 3. In a Feature Pyramid Network (FPN) comprising $N$ layers, the feature of the $i$ -th layer is denoted as $\mathbf{P}_i$ , $(i \in [1,N])$ . The process of HBS on the $i$ -th layer is as follows: + +$$ +\begin{array}{l} \mathbf {P} _ {i} ^ {E} = \mathbf {P} _ {i} ^ {f g} + \phi_ {i} \left(\mathbf {P} _ {i} ^ {b g}, r\right) \tag {2} \\ = \mathbf {P} _ {i} * M + \phi_ {i} (\mathbf {P} _ {i} * \neg M, r), \\ \end{array} +$$ + +where $\mathbf{P}_i^E$ , $\mathbf{P}_i^{fg}$ and $\mathbf{P}_i^{bg}$ denote the enhanced feature, foreground feature and background feature, $\neg M$ represents the + +![](images/e06a9f340b904b2f945069a37ed92043781a1654d730b865e62da90ad1aff6e6.jpg) +Figure 4. We visualize (a) the input image with ground-truths shown in green boxes, (b) $P_{3}$ feature in the vanilla FCOS after Principal Component Analysis (PCA) [46], and (c) $P_{3}$ feature in FCOS enhanced with HBS after PCA. Darker colors in the PCA maps indicate more critical components in the feature. As shown, HBS smoothes the background regions and significantly enhances the contrast between the foreground and background. + +complement of the binary mask $M$ , and $\circledast$ is the Hadamard product with broadcasting to accommodate the dimensions. Previous works [18, 41] reveal that feature dimensionality reduction methods can lead to smoother and more efficient feature spaces. Based on this idea, we design an adaptive smoothing operation $\phi_i(\cdot ,r)$ with channel reduction rate $r$ , formulated as: + +$$ +\phi_ {i} \left(\mathbf {P} _ {i} ^ {\mathrm {b g}}, r\right) = \sigma \left(\mathbf {w} _ {i} ^ {\mathrm {e}} \otimes \sigma \left(\mathbf {w} _ {i} ^ {\mathrm {r}} \otimes \mathbf {P} _ {i} ^ {\mathrm {b g}}\right)\right) + \mathbf {P} _ {i} ^ {\mathrm {b g}}, \tag {3} +$$ + +where $\otimes$ represents the convolution operation and $\mathbf{w}_i^{\mathrm{r}}\in$ $\mathbb{R}^{C_i\times C_{i / r}\times K_i\times K_i}$ and $\mathbf{w}_i^{\mathrm{e}}\in \mathbb{R}^{C_{i / r}\times C_i\times K_i\times K_i}$ are the weights of the channel reducing and expanding kernels, respectively. $\sigma$ denotes the non-linear function ReLU. The reduction step compresses information while retaining salient features. When features are expanded, the high-frequency information are difficult to be recreated, resulting in a smoother background. The effectiveness of smoothing is exhibited through statistical analysis in Sec. 4.7. The smoothing effect is further visualized through PCA analysis. As shown in Fig. 4, feature maps enhanced with HBS exhibit a reduced intensity range and smoother transitions in the background regions compared to the vanilla FCOS, + +effectively suppressing noise without losing essential background context. + +Kernel size $K_{i}$ at each scale level is determined based on the FPN stride $S_{i}$ , where smaller kernels are utilized to remove the fine-grained details in lower feature layers and larger kernels for more coarse-grained noise in higher layers, determined as: + +$$ +K _ {i} = g \left(S _ {i}\right) = \left(\left\lfloor \frac {\log_ {2} \left(S _ {i}\right)}{2} \right\rfloor \times 2\right) + 1, \tag {4} +$$ + +rounding up $S_{i}$ to an odd integer and ensuring central pixel alignment in the convolutional kernel. The padding is determined as $(K_{i} - 1) / 2$ to preserve the spatial dimensions of the feature map. In addition, we also tested other $g(S_{i})$ , and the results are shown in the Tab. 5. The HBS architecture allows for the suppression of high-frequency information in the background while preserving foreground details, significantly enhancing the frequency signatures of tiny objects amid high-frequency noise. + +# 3.2. Adversarial Perturbation Injection + +Analysis in Sec. 1 reveals that tiny objects exhibit weak high-frequency signatures after feature encoding and low discrimination against the cluttered background. To address this, we design an Adversarial Perturbation Injection (API) module, which leverages adversarial perturbations to enhance the saliency in critical regions. In addition, API promotes robust feature representations for tiny objects through adversarial training. + +Inspired by prior works in adversarial training [16, 17, 42], we exploit feature-level perturbations that adversarially change the model output to enhance feature representations. The objective of API for each FPN layer is formulated as follows: + +$$ +\min _ {\mathbf {P} _ {i}, \theta_ {i}} \left(\max _ {\| \epsilon_ {i, \mathrm {c l s}} \| \leq \rho} \mathcal {L} _ {\mathrm {c l s}} \left(\mathbf {P} _ {i} + \epsilon_ {i, \mathrm {c l s}}\right) + \gamma \| \mathbf {P} _ {i} \| _ {2} ^ {2}\right), \tag {5} +$$ + +where $\mathcal{L}_{\mathrm{cls}}$ denotes the classification loss. The inner optimization injects adversarial perturbation $\epsilon_{i,\mathrm{cls}}$ into the feature space of $\mathbf{P}_i$ , where $\rho$ defines the perturbation size, and $\gamma$ is a hyperparameter that controls the strength of regularization. $\theta_{i}$ denotes the model parameters at the $i$ -th layer. We derive a closed-form solution for the adversarial perturbation under the $L_{2}$ norm following [12]: + +$$ +\epsilon_ {i, \mathrm {c l s}} ^ {*} \approx \rho \cdot \frac {\nabla_ {\mathbf {P} _ {i}} \mathcal {L} _ {\mathrm {c l s}} (\mathbf {P} _ {i})}{\| \nabla_ {\mathbf {P} _ {i}} \mathcal {L} _ {\mathrm {c l s}} (\mathbf {P} _ {i}) \| _ {2}}. \tag {6} +$$ + +The perturbation can elevate the gradient of critical regions, thereby increasing their saliency during training. The gradient change induced by the perturbation on the classification branch can be calculated as: + +$$ +\Delta \left(\nabla_ {\mathbf {P} _ {i}} \mathcal {L} _ {\mathrm {c l s}}\right) = \left\| \nabla_ {\mathbf {P} _ {i}} \left(\mathcal {L} _ {\mathrm {c l s}} \left(\mathbf {P} _ {i} + \epsilon_ {i, \mathrm {c l s}} ^ {*}\right) - \mathcal {L} _ {\mathrm {c l s}} \left(\mathbf {P} _ {i}\right)\right) \right\|. \tag {7} +$$ + +![](images/6c3dafd05e366d1c8e20b3717800f36da73f88a0378f437de00087f3d2819032.jpg) + +![](images/dd9bab4cf0fb41c54e42f3c5664afb6c6c8c404d3cb32ec213b66db81e78e0e6.jpg) + +![](images/728dfac29a8eb039e8dcfd1078460a60c4f1c2e3c32caf5542c37b5deade5283.jpg) + +![](images/a2d4c907d5a5ae2e7f9a631618afbee807ac66f77009eb57766cdfdfba328d8c.jpg) +Figure 5. Average saliency visualization in (a) FCOS and (b) FCOS w/ API. The saliency maps are first computed based on the $\ell_2$ -norm of the gradient [13] at neck feature $P_{3}$ . The saliency of each object is then summed and annotated within their respective bounding box, which is different between (a) and (b). The average saliency of each object scale is calculated as the mean saliency across objects in the corresponding size category, displayed in the bottom-left corner of each image. + +![](images/dc9290d4ac9dce2f6cf1e326f26c5559de9e43770ea7cf5ca17e9ba3a9275240.jpg) + +![](images/b083ab31446257913304a21d7a3a130ee3d51019e0e4a13152f32b021df4cc4b.jpg) + +High $\Delta (\nabla_{\mathbf{P}_i}\mathcal{L}_{\mathrm{cls}})$ corresponds to regions where model outputs are highly sensitive to minor changes in activations, indicating the presence of semantic information about the image [49]. This elevated gradient enhances the feature saliency of these critical regions and prompts the refinement of object features during training. The effect particularly benefits tiny objects as they exhibit weak high-frequency signatures and face a training deficiency problem. The enhanced object saliency phenomenon is discussed in Sec. 4.6 through Fig. 5. Furthermore, the adversarial training facilitates robust feature representations and reduces the sensitivity to high-frequency noise [63], thereby enhancing the discrimination of tiny objects. + +We formulate the adversarial perturbation $\epsilon_{i}^{\mathrm{adv}}$ for each FPN layer as an expectation over $M$ object detection branches (e.g., the classification, regression and center-ness branch in a FCOS detector [39]), each scaled by a corresponding balancing parameter $\lambda_{m}$ . This is generated as follows: + +$$ +\epsilon_ {i} ^ {\mathrm {a d v}} = \sum_ {m = 1} ^ {M} \lambda_ {m} \cdot \epsilon_ {i, m} ^ {*}, \tag {8} +$$ + +where $\epsilon_{i,m}^{*}$ represents the adversarial perturbation for the $m$ -th branch. The design allows for a balanced incorporation of detection semantics across multiple branches, ensuring that the perturbations are distributed for more effective training. + +# 3.3. Auxiliary Optimization + +The optimization leverages a total loss function that integrates both the original detection loss $\mathcal{L}_{\mathrm{detection}}$ and an auxiliary loss derived from the HBS and API enhanced features. This joint loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e t e c t i o n}} + \underbrace {\lambda \cdot \sum_ {i = 1} ^ {N} \mathcal {L} _ {i} ^ {\mathrm {a u x}} \left(\mathbf {P} _ {i} ^ {E} + \epsilon_ {i} ^ {\mathrm {a d v}}\right)} _ {\text {a u x i l i a r y l o s s}}, \tag {9} +$$ + +where $\mathcal{L}_i^{\mathrm{aux}}$ denotes the auxiliary loss component for the $i$ -th layer, which is summed and scaled by a balancing hyperparameter $\lambda$ . The perturbation $\epsilon_i^{\mathrm{adv}}$ is applied to the HBS-enhanced feature $\mathbf{P}_i^E$ . The auxiliary optimization process updates the model's parameters in concert with the original loss, leading to more robust and discriminative features for tiny objects. We visualize the effectiveness of the method in Fig. 2. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS, benefiting their detection amid high-frequency noise and in the presence of larger objects. + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +Datasets. We evaluate the proposed method on AI-TOD [44], VisDrone2019 [65], DOTA-v2.0 [47] and COCO 2017 [22]. The main experiments were conducted using the AI-TOD dataset, with an average object size of 12.8 pixels, substantially smaller than other object detection datasets such as PASCAL VOC (156.6 pixels) [10] and MS COCO (99.5 pixels) [21]. Furthermore, we test our method on the VisDrone2019 [65] and DOTA-v2.0 [47], which comprises high-resolution drone-shot images with a substantial proportion of tiny instances. We also conduct experiments on the general object detection benchmark COCO 2017 [22], incorporating $\mathrm{AP}_{vt}$ , $\mathrm{AP}_t$ , $\mathrm{AP}_s$ metrics to evaluate objects of very tiny (2-8 pixels), tiny (8-16 pixels) and small (16-32 pixels), following [44]. + +Table 1. Main results with various frameworks on AI-TOD [44]. Models are trained on the AI-TOD trainval and validated on the AI-TOD test. We report APs (\%) with different IoU thresholds and APs (\%) for objects of various sizes based on the AI-TOD criterion. RFLA is based on Cascade R-CNN. The * denotes using P2~P6 FPN features. The **bold** indicates the best result. + +
FrameworkAP\( AP_{0.5} \)\( AP_{0.75} \)\( AP_{vt} \)\( AP_t \)\( AP_s \)
PAA [19]10.026.56.73.510.513.1
ATSS [61]11.628.57.62.511.915.9
Centernet [9]13.439.25.03.812.117.7
DetectoRS [34]14.832.811.50.010.818.3
DotD [50]16.139.210.68.317.618.1
DetectoRS w/NWD [43]20.849.314.36.419.729.6
DetectoRS w/ SR-TOD [3]24.054.617.110.124.829.3
One-stage
RetinaNet [24]7.220.53.62.56.811.0
RetinaNet w/ SET9.124.65.22.98.212.8
RetinaNet* [24]13.331.78.63.914.417.7
RetinaNet* w/ QueryDet [55]12.229.37.32.410.518.5
RetinaNet* w/ SET16.538.610.95.015.721.8
FCOS [39]12.029.08.02.511.917.1
FCOS w/ SET14.234.99.82.913.020.2
FCOS* [39]15.135.810.25.916.618.8
FCOS* w/ SET18.042.712.47.218.523.0
Multi-stage
Faster R-CNN [35]11.126.38.10.07.223.3
Faster R-CNN w/ SET12.228.29.70.19.124.3
Cascade R-CNN [2]13.630.310.60.09.925.5
Cascade R-CNN w/ SET14.832.411.20.410.826.7
RFLA [52]21.750.515.38.321.824.5
RFLA w/ SR-TOD [3]21.850.815.49.721.827.4
RFLA w/ SET24.955.617.99.825.231.0
Transformer-based
DINO-5scale [60]23.256.615.49.923.129.3
DINO-5scale w/ DNTR [25]26.256.720.212.826.431.0
DINO-5scale w/ SET26.657.120.813.227.131.5
+ +Implementation Details. We conducted the experiments on a computer with an NVIDIA RTX 3090 GPU. The experiments are implemented with PyTorch [32], with core codes built upon MMdetection [5]. The ImageNet [36] pre-trained models are used as the backbones. All CNN-based models utilize the ResNet-50 [15] backbone, trained using the Stochastic Gradient Descent (SGD) optimizer for 12 epochs with 0.9 momentum, 0.0001 weight decay, and a batch size 2. The initial learning rate is 0.005, decaying at the 8th and 11th epochs. The data processing adheres to the default configurations of each dataset (e.g., fixed at $800 \times 800$ for AI-TOD). We also train a transformer-based detector, DINO [60], with 5-scale feature maps for 36 epochs as a baseline. The training uses an Adam optimizer with a weight decay of 0.0001, following the random crop and scale augmentation strategies of DETR [4]. + +# 4.2. Results on AI-TOD + +We evaluate the proposed SET method on various detectors and compare it with state-of-the-art TOD methods on the AI-TOD benchmark [51]. Tab. 1 shows that SET enhances + +all baselines by about $2\%$ AP, which is significant. Specifically, SET improves the one-stage detector FCOS [39] and RetinaNet [24] by $2.2\% / 1.9\%$ AP and $1.1\% / 1.4\%$ $\mathrm{AP}_t$ , respectively. We further revise the detector architecture and use P2~P6 FPN features, a representative approach for TOD tasks where the high-resolution P2 layer benefits tiny objects. With the adjustment, SET boosts the RetinaNet and FCOS baselines by $3.2\% / 2.9\%$ AP and $1.1\% / 1.3\% \mathrm{AP}_{vt}$ a significant margin. In contrast, the state-of-the-art SOD method QueryDet [55] decreases the baseline performance, showing limitations in addressing the TOD challenge. SET also generalizes to multi-stage detectors, enhancing the Faster R-CNN [35] and Cascade R-CNN [2] baselines by $1.1\% / 1.2\%$ AP and $1.9\% / 1.1\%$ $\mathrm{AP}_t$ respectively. + +SET works orthogonally with the state-of-the-art TOD method RFLA [52], achieving a $3.2\%$ AP increment over the RFLA baseline and surpassing the comparison SR-TOD [3] by $3.1\%$ AP and $3.4\%$ $\mathrm{AP}_t$ . We also investigate SET's compatibility with transformer-based detectors. As shown, SET results in $3.4\%$ AP increase over the DINO-5scale [60] baseline, outperforming the prior + +Table 2. Detection performance on the VisDrone2019 [65] validation set. RFLA is based on Cascade R-CNN. The **bold** denotes the best result. + +
FrameworkAPAPvtAPtAPs
RetinaNet [24]15.20.41.69.3
Faster R-CNN [35]23.10.15.119.6
FCOS [39]19.90.74.515.7
FCOS w/ SET21.91.65.717.9
Cascade R-CNN [2]23.00.15.019.4
Cascade R-CNN w/ SET24.31.26.321.0
RFLA [52]27.24.513.023.6
RFLA w/ SR-TOD [3]27.84.813.124.5
RFLA w/ SET28.55.213.524.6
+ +art DNTR [25]. Notably, DINO-5scale w/SET derives $26.6\%$ AP, outperforming competitors including DotD [50], NWD [43], and SR-TOD [3]. The results demonstrate SET as a robust solution for the tiny object detection task. + +# 4.3. Results on VisDrone2019 and DOTA-V2.0 + +We further evaluate SET on the drone-shot image benchmark VisDrone2019 [65], which features tiny and small objects. As shown in Tab. 2, the proposed SET effectively improves the performance of both one-stage and multi-stage detectors and exhibits consistent $\mathrm{AP}_{vt}$ and $\mathrm{AP}_t$ improvements. Specifically, SET improved the the one-stage detector FCOS by $2.0\% /0.9\% /1.2\%$ AP, $\mathrm{AP}_{vt}$ and $\mathrm{AP}_t$ , respectively. SET also improves multi-stage detector Cascade RCNN [2] by $1.1\% /1.3\%$ $\mathrm{AP}_{vt}$ and $\mathrm{AP}_t$ , respectively. Notably, SET improves the prior art RFLA, achieving a $28.5\%$ AP with $0.7\%$ increment in $\mathrm{AP}_{vt}$ and $0.5\%$ in $\mathrm{AP}_t$ + +Tab. 3 demonstrates the DOTA-v2.0 [8] results. As shown, SET obtains a substantial performance gain of $3.1\%$ and $1.3\%$ in $\mathrm{AP}_t$ with the FCOS and AutoAssign baseline, underscoring the efficacy of our method in detecting tiny objects. Compared to the state-of-the-art comparisons RFLA [52] and DCFL [53], SET achieves $2.6\%$ and $0.9\%$ AP gain, respectively. More experimental results on the TinyPerson [56] and SeaPerson [57] dataset are provided in the supplementary material. + +# 4.4. Results on COCO + +We further conduct experiments on the general object detection benchmark MS COCO [21]. As shown in Tab. 9, SET achieves $1.0\%$ AP improvement over the FCOS baseline, with a notable $1.7\%$ improvement in $\mathrm{AP}_t$ , demonstrating that SET can enhance the detection of tiny objects in general detection tasks. + +# 4.5. Ablation Study + +In the following experiments, we explore the best structure of SET by ablating and tuning its parts using AI-TOD. + +Table 3. Detection performance on DOTA-v2.0 [8]. Note that models are trained on the DOTA-v2.0 train and validated on the DOTA-v2.0 val. The bold denotes the best result. + +
FrameworkAPAPvtAPtAPs
ATSS [61]32.70.76.923.4
Faster R-CNN [35]35.60.07.128.9
FCOS [39]31.80.34.019.4
FCOS w/ RFLA [52]32.10.76.823.6
FCOS w/ DCFL [53]33.80.56.924.2
FCOS w/ SET34.70.67.124.8
AutoAssign [64]33.80.97.322.4
AutoAssign w/ SET35.21.18.623.6
+ +Table 4. Detection performance on MS COCO [21]. Note that models are trained on COCO train2017 and validated on COCO val2017. $(\cdot)^{+}$ denotes performance increment. + +
FrameworkAPAPvtAPtAPsAPm
FCOS [39]36.47.919.627.243.6
FCOS w/ SET37.4+1.08.3+0.423.3+1.727.6+0.444.8+1.2
+ +Effectiveness and Cost of Components. We present quantitative improvements of SET components in Tab. 5. As shown, applying the adaptive smoothing operations to full features $(+\phi_{i}(\mathbf{P}_{i},r))$ achieves a $0.4\%$ increase in AP and $0.3\%$ in $\mathrm{AP}_{vt}$ , surpassing the performance of static high-frequency filtering in Fig. 1. With decoupling, the HBS module boosts the overall AP by $1.9\%$ AP and $0.4\%$ $\mathrm{AP}_{vt}$ . The API module alone achieves a $1.2\%$ AP improvement. Combining HBS and API results in a total AP increase of $2.2\%$ . Results demonstrate the synergy between the heterogeneous operations, where HBS effectively smooths the clutter background and API prompts the learning of object features, cumulatively improving the detection accuracy of tiny objects. Tab. 5 also presents the cost of components. Results show that the HBS module leads to $17\%$ , $0.9\%$ , and $0.9\%$ increase in training time, computational and memory cost. The API module results in a more significant increase in training time due to adversarial training but adds no forward computational cost and minimal increase in memory usage. The overall minimal impact indicates that SET can be implemented with negligible additional overhead. + +Hyper-parameter Selection. We first select the hyperparameters channel reduction rate $r$ and perturbation size $\rho$ using the FCOS [39] framework trained three epochs in Fig. 6 (a). Model performance (AP) with different hyperparameter combinations $\{r, \rho\}$ is presented. Results indicate that performance initially improves and then declines as $r$ varies from 2 to 16, with $r = 16$ perform worse than the baseline. Additionally, SET with adversarial perturbations $(\rho > 0)$ exhibits stronger performance than without $(\rho = 0)$ , but large-sized perturbations $(\rho = 5)$ performs worse than all alternatives. Exploring the setups, we identify $\{r, \rho\} = \{4, 1\}$ as the combination for SET. Fig. 6 (b) explores the + +Table 5. The impact of different components and various channel reduction operations in SET. The hyper-parameters are fixed with channel reduction rate $r$ being 4, perturbation size $\rho$ being 1, and the balancing hyper-parameters $\lambda$ of auxiliary optimization being 1. + +
Frameworkφig(Si)APAPvtAPtTime(s/batch)GFLOPsParams (M)
FCOS [39]--12.02.511.90.17126.1337.04
+ φi(Pi, r)--12.72.712.90.17127.3637.41
Eqn. 3313.42.712.6
Eqn. 3log2(Si)/2×2+113.92.712.9
+ HBSEqn. 3513.12.312.50.20127.3237.41
Eqn. 3log2(Si+1)/2×2-113.42.812.4
Eqn. 3712.12.512.2
+ API--13.22.612.40.39126.1337.16
+ Both (SET)Eqn. 3Eqn. 414.22.913.00.42127.3237.85
+ +![](images/c206ef2afd100a6fd3aeb89ceef8ffd5959e034570a8b9a49dfd9ccb2f45068f.jpg) +Figure 6. We select hyper-parameters channel reduction rate $r$ with perturbation size $\rho$ in (a), and the balancing hyper-parameter $\lambda$ of Eqn. 9 in (b). + +![](images/68468b3b4f4089924a4fda40e525e9ca399c0f1a7bbe44ed024ea4e6851b68e4.jpg) +Figure 7. The analysis examines the adaptive smoothing operation, obtained from 1000 randomly selected images from the Vis-Drone2019 [65] train set. + +balancing hyper-parameter $\lambda$ in Eqn. 9 that controls the proportion of the auxiliary optimization, using the FCOS [39], RetinaNet [24], and Faster R-CNN [35] framework. As observed, the performances increase first and then decrease with the increase of $\lambda$ , with $\lambda = 1$ yielding the best performance across all frameworks. However, the efficacy of SET diminishes when $\lambda$ exceeds the 1, as larger-sized $\lambda$ ( $\lambda > 1$ ) can impose too strong regularization to the detector, reducing the detection accuracy despite increasing robustness. Therefore, we set the hyper-parameter $\lambda$ to 1 for experiments in this paper. + +# 4.6. Enhanced Object Saliency + +We demonstrate effectiveness of API in increasing the saliency of object features in Fig. 5. Comparing the average saliency of each object scale, tiny objects are the least salient due to their inherently limited pixel input. API leads to a notable increase in the average saliency of tiny objects (e.g., from 53.99 to 69.93), benefited from the injection of perturbations (Eqn. 7). Fig. 5 also demonstrates that API can enhance the feature saliency of small and medium objects, though the degree of improvement is less. The comparison underscores that API could benefit general object detection, with more significant and essential improvements for tiny objects. + +![](images/5a6f021a8eb905f53dc3e1a29e203e9461e4e8afaf5e2cfff66bb1bd9946ca44.jpg) + +# 4.7. Effectiveness in Smoothing + +In Fig. 7, we perform a statistic analysis on $P_{i}^{bg}$ to investigate the adaptive smoothing operation. We convert feature $P_{i}^{bg}$ and $\phi_{i}(P_{i}^{bg}, r)$ into the frequency space using FFT [31] and separate the amplitude spectrums of these features into three parts (low-, mid-, and high-frequency) using static thresholds as Fig. 1. The amplitudes in each band are then summed and averaged to measure the signal strength. Results verify that the channel reducing and expand design effects in smoothing. Comparing $A(P_{i})(u, v)$ and $A(\phi_{i}(P_{i}^{bg}, r))(u, v)$ , the low-frequency signals are enhanced (from 0.3787 to 0.5205), while mid- and high-frequency signals are attenuated. + +# 5. Conclusion + +This paper proposes the Spectral Enhancement for Tiny objects (SET) method, which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate its effectiveness. + +Acknowledgments. The work was supported by the National Natural Science Foundation of China (No. 61827901, 623B2016, and 62406298). This research was also supported by the National Key Research and Development Program of China (Grant No. 2023YFC3306401), Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F020007, Beijing Natural Science Foundation L223024, L244043 and Z241100001324017, "One Thousand Plan" projects in Jiangxi Province Jxsq2023102268, and the Beijing Jiaotong University "Jingying Plan" No. K24XKRC00130. + +# References + +[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, pages 206-221, 2018. 1 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 6, 7 +[3] Bing Cao, Haiyu Yao, Pengfei Zhu, and Qinghua Hu. Visible and clear: Finding tiny objects in difference map. ECCV, 2024. 1, 2, 6, 7 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6 +[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6 +[6] Yukang Chen, Peizhen Zhang, Zeming Li, Yanwei Li, Xiangyu Zhang, Lu Qi, Jian Sun, and Jiaya Jia. Dynamic scale training for object detection. arXiv preprint arXiv:2004.12432, 2020. 2 +[7] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE TPAMI, 2023. 1 +[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE TPAMI, 44(11):7778-7796, 2021. 7 +[9] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, pages 6569-6578, 2019. 6 +[10] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. ICCV, 88:303-308, 2009. 1, 5 +[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 1 + +[12] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 4, 1 +[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021. 5, 1 +[14] Kaiming He and Jian Sun. Convolutional neural networks at constrained time cost. In CVPR, pages 5353-5360, 2015. 1 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[16] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 4 +[17] Masato Ishii and Atsushi Sato. Training deep neural networks with adversarially augmented features for small-scale training datasets. In IJCNN, pages 1-8. IEEE, 2019. 4 +[18] Weikuan Jia, Meili Sun, Jian Lian, and Sujuan Hou. Feature dimensionality reduction: a review. Complex & Intelligent Systems, 8(3):2663-2693, 2022. 4 +[19] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371, 2020. 6 +[20] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR*, 2016. 3 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 1, 5, 7, 3 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 1, 5 +[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2 +[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 6, 7, 8, 2, 3 +[25] Hou-I Liu, Yu-Wen Tseng, Kai-Cheng Chang, Pin-Jyun Wang, Hong-Han Shuai, and Wen-Huang Cheng. A denoising fpn with transformer r-cnn for tiny object detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6, 7 +[26] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 2 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3 +[28] Jiajun Lu, Hussein Sibai, and Evan Fabry. Adversarial examples that fool detectors. arXiv preprint arXiv:1712.02494, 2017.3 + +[29] T. Na, J. H. Ko, and S. Mukhopadhyay. Cascade adversarial machine learning regularized with a unified embedding. In ICLR, 2017. 3 +[30] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In ICCV, pages 9725-9734, 2019. 1 +[31] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 2, 8 +[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS Workshops, pages 1-12, 2019. 6 +[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In CVPR, pages 11443-11452, 2021. 2 +[34] Siyuan Qiao, Liang-Chieh Chen, and Alan Yuille. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In CVPR, pages 10213-10224, 2021. 6 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(06):1137-1149, 2017. 6, 7, 8 +[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. ICCV, 115:211-252, 2015. 6 +[37] Bharat Singh, Mahyar Najibi, Abhishek Sharma, and Larry S Davis. Scale normalized image pyramids with autofocus for object detection. IEEE TPAMI, 44(7):3749-3766, 2021. 2 +[38] Dawn Song, Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Florian Tramer, Atul Prakash, and Tadayoshi Kohno. Physical adversarial examples for object detectors. In 12th USENIX workshop on offensive technologies (WOOT '18), 2018. 3 +[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 5, 6, 7, 8, 2, 3 +[40] F Tramér, D Boneh, A Kurakin, I Goodfellow, N Papernot, and P McDaniel. Ensemble adversarial training: Attacks and defenses. In *ICLR*, 2018. 3 +[41] S Velliangiri, SJPCS Alagumuthukrishnan, et al. A review of dimensionality reduction techniques for efficient computation. Procedia Computer Science, 165:104-111, 2019. 4 +[42] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, pages 5495-5504, 2018. 4 +[43] Jinwang Wang, Chang Xu, Wen Yang, and Lei Yu. A normalized gaussian Wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389, 2021. 6, 7 +[44] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In ICPR, pages 3791-3798, 2021. 1, 5, 6 + +[45] Kunyu Wang, Xueyang Fu, Yukun Huang, Chengzhi Cao, Gege Shi, and Zheng-Jun Zha. Generalized uav object detection via frequency domain disentanglement. In CVPR, pages 1064-1073, 2023. 2 +[46] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 4, 2 +[47] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, pages 3974-3983, 2018. 5 +[48] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In ICCV, pages 1369-1378, 2017. 3 +[49] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In CVPR, pages 501-509, 2019. 5 +[50] Chang Xu, Jinwang Wang, Wen Yang, and Lei Yu. Dot distance for tiny object detection in aerial images. In ICCV, pages 1192-1201, 2021. 6, 7 +[51] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Detecting tiny objects in aerial images: A normalized wasserstein distance and a new benchmark. IS-PRS J. Photo. and Rem. Sen., 190:79-93, 2022. 6, 3 +[52] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543, 2022. 2, 6, 7, 3 +[53] Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Dynamic coarse-to-fine learning for oriented tiny object detection. In CVPR, pages 7318-7328, 2023. 7 +[54] Sheng Xu, Mingze Wang, Yanjing Li, Mingbao Lin, Baochang Zhang, David Doermann, and Xiao Sun. Learning 1-bit tiny object detector with discriminative feature refinement. In ICML, 2024. 1, 2 +[55] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 6 +[56] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In WACV, pages 1257–1265, 2020. 7, 2 +[57] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. Cpr++: Object localization via single coarse point supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4908-4925, 2024. 7, 3 +[58] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small object detection via coarse-to-fine proposal generation and imitation learning. In ICCV, pages 6317-6327, 2023. 2 +[59] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, pages 421-430, 2019. 3 +[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr + +with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1, 6 +[61] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 6, 7, 3 +[62] Qijie Zhao, Tao Sheng, Yongtao Wang, Zhi Tang, Ying Chen, Ling Cai, and Haibin Ling. M2det: A single-shot object detector based on multi-level feature pyramid network. In AAAI, pages 9259-9266, 2019. 2 +[63] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, pages 452-467, 2018. 5 +[64] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020. 7, 2, 3 +[65] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Haibin Ling, Qinghua Hu, Qinqin Nie, Hao Cheng, Chenfeng Liu, Xiaoyu Liu, et al. Visdrone-det2018: The vision meets drone object detection in image challenge results. In ECCV Workshops, pages 0–0, 2018. 5, 7, 8 \ No newline at end of file diff --git a/2025/SET_ Spectral Enhancement for Tiny Object Detection/images.zip b/2025/SET_ Spectral Enhancement for Tiny Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa8c14b20b7fc6b5f2254d0ea41dc058cbdd0f80 --- /dev/null +++ b/2025/SET_ Spectral Enhancement for Tiny Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e0998419edaa46b4ea53dbaddfbc4f95fa3dc432ad0ce7a1d7dbf1d8feada90 +size 731910 diff --git a/2025/SET_ Spectral Enhancement for Tiny Object Detection/layout.json b/2025/SET_ Spectral Enhancement for Tiny Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4afa4824f82dd46151bde7df51594d4b6474686c --- /dev/null +++ b/2025/SET_ Spectral Enhancement for Tiny Object Detection/layout.json @@ -0,0 +1,9973 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 474, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 474, + 121 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 474, + 121 + ], + "type": "text", + "content": "SET: Spectral Enhancement for Tiny Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": "Huixin Sun" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": ", Runqi Wang" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": ", Yanjing Li" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": ", Linlin Yang" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": ", Shaohui Lin" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": ", Xianbin Cao" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "text", + "content": ", Baochang Zhang" + }, + { + "bbox": [ + 141, + 142, + 471, + 172 + ], + "type": "inline_equation", + "content": "^{3,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "type": "text", + "content": "School of Electronic Information Engineering, Beihang University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 185, + 485, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 185, + 485, + 199 + ], + "spans": [ + { + "bbox": [ + 126, + 185, + 485, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 185, + 485, + 199 + ], + "type": "text", + "content": "School of Computer Science and Technology, Beijing Jiaotong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 199, + 471, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 199, + 471, + 213 + ], + "spans": [ + { + "bbox": [ + 139, + 199, + 471, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 199, + 471, + 213 + ], + "type": "text", + "content": "School of Artificial Intelligence, Beihang University, Beijing, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 201, + 213, + 410, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 213, + 410, + 228 + ], + "spans": [ + { + "bbox": [ + 201, + 213, + 410, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 201, + 213, + 410, + 228 + ], + "type": "text", + "content": "Zhongguancun Laboratory, Beijing, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 227, + 467, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 467, + 241 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 467, + 241 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 143, + 227, + 467, + 241 + ], + "type": "text", + "content": " State Key Laboratory of Media Convergence and Communication," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 179, + 242, + 432, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 242, + 432, + 255 + ], + "spans": [ + { + "bbox": [ + 179, + 242, + 432, + 255 + ], + "type": "text", + "content": "Communication University of China, Beijing, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 255, + 533, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 255, + 533, + 270 + ], + "spans": [ + { + "bbox": [ + 74, + 255, + 533, + 270 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 74, + 255, + 533, + 270 + ], + "type": "text", + "content": "School of Computer Science and Technology, East China Normal University, Shanghai, China" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 152, + 297, + 200, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 297, + 200, + 309 + ], + "spans": [ + { + "bbox": [ + 152, + 297, + 200, + 309 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 323, + 297, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 297, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 297, + 561 + ], + "type": "text", + "content": "Deep learning has significantly advanced the object detection field. However, tiny object detection (TOD) remains a challenging problem. We provide a new analysis method to examine the TOD challenge through occlusion-based attribution analysis in the frequency domain. We observe that tiny objects become less distinct after feature encoding and can benefit from the removal of high-frequency information. In this paper, we propose a novel approach named Spectral Enhancement for Tiny object detection (SET), which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate the effectiveness of our method. Especially, SET boosts the prior art RFLA by " + }, + { + "bbox": [ + 55, + 323, + 297, + 561 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 55, + 323, + 297, + 561 + ], + "type": "text", + "content": " AP on the AI-TOD dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 576, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "Recent advances in Deep Neural Networks (DNNs) [14, 60] have significantly improved the object detection field [11, 22]. Despite the progress, Tiny Object Detection (TOD) remains a challenging problem. Tiny objects, characterized by their very limited pixel input, occupy areas equal to or less than " + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": " pixels [44]. Compared to traditional object detection [10, 21], generic object detectors often fail to manifest effectiveness on the TOD task. For instance," + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 296, + 555, + 380 + ], + "blocks": [ + { + "bbox": [ + 315, + 296, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 315, + 296, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 296, + 555, + 380 + ], + "type": "image", + "image_path": "1e0b8c54b258c0a80bd94592a0ff244e9d8820ecbd3957f12f3786a192ad2c7e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "text", + "content": "Figure 1. A feature-level occlusion-based attribution analysis in the frequency domain. Mid-frequency features ranges from " + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "text", + "content": " of the total energy, high-frequency features are above " + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "text", + "content": " of the total energy. The model is trained on the AI-TOD [44] trainval and validated on the AI-TOD test. Detection performance is evaluated by inferencing a " + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 313, + 384, + 555, + 484 + ], + "type": "text", + "content": " FCOS detector [39] with the masked features. Results exhibit that tiny objects are more likely to be affected by high-frequency noise and can benefit from the direct removal of high-frequency information." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "text", + "content": "DINO [60], one of the most representative transformer-based object detector, achieves " + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "inline_equation", + "content": "37.6\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "text", + "content": " AP on medium-sized objects but only " + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "text", + "content": " AP on very tiny objects on the AITOD [44] benchmark. Moreover, the vanilla FCOS [39] achieves only " + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 495, + 555, + 591 + ], + "type": "text", + "content": " AP on very tiny objects on the AITOD [44], which is far from sufficient to meet the demands of real-world applications, such as autonomous driving, maritime rescue, and traffic management." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Constrained by the inherently low resolution, one key challenge in TOD is to extract discriminative foreground features [7], especially in the high-level features of detection architecture after down-samplings. In the meantime, it is difficult to learn about tiny objects from the noisy and dominant background clutter [54]. Most existing works [1, 3, 30, 54] focus on enhancing feature representations of tiny objects in the spatial domain. We provide a new analysis method and address the challenge from a frequency spectrum perspective." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 288, + 713 + ], + "type": "text", + "content": "*Corresponding authors. Email: sunhuixin@buaa.edu.cn" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4713" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 190 + ], + "type": "image", + "image_path": "484f014fad10e1ba08980fd642514dbdb723c4613adf3cecc7d6b49c3d76c51c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 197, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 275 + ], + "type": "text", + "content": "Figure 2. The analysis examines different objects after feature encoding in the frequency domain. The frequency spectrum is calculated using the Fast Fourier Transform (FFT) [31] on the cropped " + }, + { + "bbox": [ + 55, + 197, + 296, + 275 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 55, + 197, + 296, + 275 + ], + "type": "text", + "content": " feature and selecting the channel with the highest energy. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "text", + "content": "Motivated by occlusion-based attribution analysis [33, 45], we perform an occlusion-based analysis on the intermediate Feature Pyramid Network (FPN) features to examine the impact of masking specific frequency bands on the detection performance across object scales. Specifically, we transform the features into the frequency domain using the n-dimensional Fast Fourier Transform (FFT) [31] and apply static filters to the amplitude spectrums, which are then fed to Inverse Fast Fourier Transform (IFFT) with phase spectrums to generate the masked spatial features. High-frequency mask results in Fig. 1 reveal that directly removing high-frequency features has a positive effect on very tiny and tiny objects, while negatively affecting the detection of larger objects. The disparity can be attributed to that tiny objects become less distinct after feature encoding, exhibiting weak high-frequency components (example (b), Fig. 2) and extremely vague frequency signatures in the cluttered background (example (c), Fig. 2). Therefore, they are more likely to be affected by high-frequency noise and benefit from the direct removal of high-frequency information. In contrast, larger objects exhibit distinct high-frequency signatures (example (a), Fig. 2) and have a higher dependency on high-frequency information. Based on the observations, we investigate an explicit measure to enhance the frequency signatures for tiny objects by removing the high-frequency information in the background. Background (BG) high-frequency mask results show that the approach enhances detection performance across all three object scales, resulting in improvements of " + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 56, + 282, + 296, + 653 + ], + "type": "text", + "content": ", bringing significant gains for very tiny and tiny objects without compromising larger objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "In light of the analysis above, we propose a generic Spectral Enhancement (SET) method to enhance the frequency signatures of Tiny objects using a heterogeneous architecture for foreground and background feature refinement. The method includes two modules. As illustrated in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "Fig. 3, the first Hierarchical Background Smoothing (HBS) module suppresses the high-frequency information in the background through adaptive smoothing operations while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. The second Adversarial Perturbation Injection (API) module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training. In addition, API facilitates robust feature representations through adversarial training. SET is simple yet effective and can be easily mounted on existing detectors during the training process while bringing no extra burden to the inference procedure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 228, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 239 + ], + "type": "text", + "content": "Our major contributions in this paper are summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 240, + 554, + 431 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 240, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 554, + 335 + ], + "type": "text", + "content": "- A feature-level occlusion-based attribution analysis from the frequency spectrum perspective is conducted to investigate the Tiny Object Detection (TOD) challenge, which shows that tiny objects are more likely to be affected by high-frequency noise. We introduce a spectral enhancement scheme for tiny object detection (SET) by designing a heterogeneous architecture for foreground and background feature refinement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 407 + ], + "type": "text", + "content": "- Two new modules are designed for the TOD task. The HBS module is used to suppress the high-frequency noise in the background through adaptive smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and prompt the refinement of object features during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 431 + ], + "type": "text", + "content": "- Extensive results reveal that our SET boosts the well-known baselines and the prior art by a large margin." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 442, + 402, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 402, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 402, + 455 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "Tiny Object Detection. Most current approaches for tiny object detection (TOD) can be grouped into four main categories: data augmentation, multi-scale feature learning, label assignment, and feature enhancement strategies for tiny objects. Several data augmentation strategies [6, 37] target tiny objects by prompting the detector to focus on instances of specific scales during training. Conventional approaches also utilize multi-scale feature learning [23, 26, 62]. Another method line focuses on improving tiny object detection from the label assignment and proposal refinement perspectives [52, 58]. Various works aim to enhance the feature representations of tiny objects. Yuan et al. [58] propose a Feature Imitation (FI) mechanism, aligning RoI features with their counterparts in an exemplar feature set within the embedding space. Recent state-of-the-art TOD methods have proposed auxiliary self-reconstruction branches to enhance the weak representations of tiny objects [3, 54]. Differently, our proposed method enhances the discrimination of tiny objects from a frequency perspective, prompting the refinement of object features during training without adding any extra burden to the inference process." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4714" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 309, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 309, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 309, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 309, + 232 + ], + "type": "image", + "image_path": "bdc484cd6bab9053a4ed6b8c44e788405c0eef8fc2501b90cf20c24b057b10ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 553, + 232 + ], + "type": "image", + "image_path": "94f5b3567065d32c04c37661336e95e48327e29f292a22c2ef6a866435461421.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 235, + 553, + 343 + ], + "blocks": [ + { + "bbox": [ + 58, + 235, + 553, + 343 + ], + "lines": [ + { + "bbox": [ + 58, + 235, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 235, + 553, + 343 + ], + "type": "image", + "image_path": "1a8391d7369fcd36840858097f8a68f58c6802a9589f046895e7fd1ab20670af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 351, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 54, + 351, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 555, + 397 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed SET method with FCOS framework. The red dashed line denotes generating the adversarial perturbations " + }, + { + "bbox": [ + 54, + 351, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}^{\\mathrm{adv}}" + }, + { + "bbox": [ + 54, + 351, + 555, + 397 + ], + "type": "text", + "content": ". The HBS module can suppress the high-frequency noise in the background through scale-wise smoothing operations. The API module leverages adversarial perturbations to increase feature saliency in critical regions and facilitate robust feature representations through adversarial training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 408, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 297, + 577 + ], + "type": "text", + "content": "Adversarial Training for Object Detection. Many effective attacks crafted for object detectors have been proposed, with most generating adversarial examples at the image level [28, 38, 48]. The most common method to defend against the attacks is adversarial training [20, 29, 40], which involves augmenting training datasets with adversarial examples. For object detection, [59] extend adversarial training to the scenario of object detection by leveraging the attacks sourced from both classification and localization domains. Different from the aforementioned methods, we introduce feature-level adversarial perturbations that balance the detection semantics across branches for enhanced tiny object detection, resulting in more salient foreground features and improved model robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 117, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 117, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 117, + 607 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "This section presents the Spectral Enhancement (SET) method for Tiny object detection, depicted in Fig. 3. SET employs a heterogeneous architecture. This includes the Hierarchical Background Smoothing (HBS) module, which suppresses high-frequency information in the background, and the Adversarial Perturbation Injection (API) module, which enhances the feature saliency of critical regions and prompts the refinement of object features during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 408, + 511, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 511, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 511, + 421 + ], + "type": "text", + "content": "3.1. Hierarchical Background Smoothing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 508 + ], + "type": "text", + "content": "Based on the analysis in Sec. 1, we design a Hierarchical Background Smoothing (HBS) module to suppress the high-frequency information in the background feature while preserving foreground details, thereby accentuating the frequency signatures of tiny objects. This is achieved through background feature decoupling and adaptive smoothing, as described below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{3\\times W\\times H}" + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "content": ", a binary mask " + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "content": " is generated according to the ground truth box " + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 508, + 554, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 389, + 536, + 553, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 536, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 389, + 536, + 553, + 550 + ], + "type": "interline_equation", + "content": "M _ {i, j} = \\mathbf {1} [ (i, j) \\in B ], \\tag {1}", + "image_path": "d0f7e125da278f5807d66b967239c507191dbbf9f141e42d0dc069cda9769da8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "M \\in \\{0,1\\}^{W \\times H}" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": ". The indicative function 1 denotes the value of location " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": ", which is 1 if it belongs to an object and 0 otherwise. Then we utilize the generated binary mask to decouple image features and perform background smoothing, as shown in Fig. 3. In a Feature Pyramid Network (FPN) comprising " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": " layers, the feature of the " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": "-th layer is denoted as " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "(i \\in [1,N])" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": ". The process of HBS on the " + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 553, + 555, + 650 + ], + "type": "text", + "content": "-th layer is as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 652, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 652, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 359, + 652, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {P} _ {i} ^ {E} = \\mathbf {P} _ {i} ^ {f g} + \\phi_ {i} \\left(\\mathbf {P} _ {i} ^ {b g}, r\\right) \\tag {2} \\\\ = \\mathbf {P} _ {i} * M + \\phi_ {i} (\\mathbf {P} _ {i} * \\neg M, r), \\\\ \\end{array}", + "image_path": "2126f3834d484d0d41448841cc6d9595bdce69b58c0f63da05daff8b29fa6fbe.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^E" + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^{fg}" + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^{bg}" + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": " denote the enhanced feature, foreground feature and background feature, " + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\neg M" + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": " represents the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4715" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 282, + 345 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 282, + 345 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 282, + 345 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 282, + 345 + ], + "type": "image", + "image_path": "e06a9f340b904b2f945069a37ed92043781a1654d730b865e62da90ad1aff6e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "type": "text", + "content": "Figure 4. We visualize (a) the input image with ground-truths shown in green boxes, (b) " + }, + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "type": "text", + "content": " feature in the vanilla FCOS after Principal Component Analysis (PCA) [46], and (c) " + }, + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 431 + ], + "type": "text", + "content": " feature in FCOS enhanced with HBS after PCA. Darker colors in the PCA maps indicate more critical components in the feature. As shown, HBS smoothes the background regions and significantly enhances the contrast between the foreground and background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "text", + "content": "complement of the binary mask " + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\circledast" + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "text", + "content": " is the Hadamard product with broadcasting to accommodate the dimensions. Previous works [18, 41] reveal that feature dimensionality reduction methods can lead to smoother and more efficient feature spaces. Based on this idea, we design an adaptive smoothing operation " + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\phi_i(\\cdot ,r)" + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "text", + "content": " with channel reduction rate " + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 448, + 295, + 531 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 535, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 535, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 85, + 535, + 295, + 552 + ], + "type": "interline_equation", + "content": "\\phi_ {i} \\left(\\mathbf {P} _ {i} ^ {\\mathrm {b g}}, r\\right) = \\sigma \\left(\\mathbf {w} _ {i} ^ {\\mathrm {e}} \\otimes \\sigma \\left(\\mathbf {w} _ {i} ^ {\\mathrm {r}} \\otimes \\mathbf {P} _ {i} ^ {\\mathrm {b g}}\\right)\\right) + \\mathbf {P} _ {i} ^ {\\mathrm {b g}}, \\tag {3}", + "image_path": "0f6c171cd3aee4a0ec81e410451c90c7a83ddbac420468d3a4e48903baa7b7f5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " represents the convolution operation and " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{\\mathrm{r}}\\in" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C_i\\times C_{i / r}\\times K_i\\times K_i}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{\\mathrm{e}}\\in \\mathbb{R}^{C_{i / r}\\times C_i\\times K_i\\times K_i}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " are the weights of the channel reducing and expanding kernels, respectively. " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " denotes the non-linear function ReLU. The reduction step compresses information while retaining salient features. When features are expanded, the high-frequency information are difficult to be recreated, resulting in a smoother background. The effectiveness of smoothing is exhibited through statistical analysis in Sec. 4.7. The smoothing effect is further visualized through PCA analysis. As shown in Fig. 4, feature maps enhanced with HBS exhibit a reduced intensity range and smoother transitions in the background regions compared to the vanilla FCOS," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "effectively suppressing noise without losing essential background context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "type": "text", + "content": "Kernel size " + }, + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "type": "text", + "content": " at each scale level is determined based on the FPN stride " + }, + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 95, + 553, + 155 + ], + "type": "text", + "content": ", where smaller kernels are utilized to remove the fine-grained details in lower feature layers and larger kernels for more coarse-grained noise in higher layers, determined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 351, + 157, + 553, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 157, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 351, + 157, + 553, + 185 + ], + "type": "interline_equation", + "content": "K _ {i} = g \\left(S _ {i}\\right) = \\left(\\left\\lfloor \\frac {\\log_ {2} \\left(S _ {i}\\right)}{2} \\right\\rfloor \\times 2\\right) + 1, \\tag {4}", + "image_path": "ca165cf367d1055259c8b02f21857a54c1bde66a524b9b75eb3990b506aeac7a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "text", + "content": "rounding up " + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "text", + "content": " to an odd integer and ensuring central pixel alignment in the convolutional kernel. The padding is determined as " + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "inline_equation", + "content": "(K_{i} - 1) / 2" + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "text", + "content": " to preserve the spatial dimensions of the feature map. In addition, we also tested other " + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "inline_equation", + "content": "g(S_{i})" + }, + { + "bbox": [ + 313, + 189, + 555, + 297 + ], + "type": "text", + "content": ", and the results are shown in the Tab. 5. The HBS architecture allows for the suppression of high-frequency information in the background while preserving foreground details, significantly enhancing the frequency signatures of tiny objects amid high-frequency noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 500, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 500, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 500, + 317 + ], + "type": "text", + "content": "3.2. Adversarial Perturbation Injection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "type": "text", + "content": "Analysis in Sec. 1 reveals that tiny objects exhibit weak high-frequency signatures after feature encoding and low discrimination against the cluttered background. To address this, we design an Adversarial Perturbation Injection (API) module, which leverages adversarial perturbations to enhance the saliency in critical regions. In addition, API promotes robust feature representations for tiny objects through adversarial training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 418, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 476 + ], + "type": "text", + "content": "Inspired by prior works in adversarial training [16, 17, 42], we exploit feature-level perturbations that adversarially change the model output to enhance feature representations. The objective of API for each FPN layer is formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 480, + 553, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 480, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 341, + 480, + 553, + 508 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {P} _ {i}, \\theta_ {i}} \\left(\\max _ {\\| \\epsilon_ {i, \\mathrm {c l s}} \\| \\leq \\rho} \\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i} + \\epsilon_ {i, \\mathrm {c l s}}\\right) + \\gamma \\| \\mathbf {P} _ {i} \\| _ {2} ^ {2}\\right), \\tag {5}", + "image_path": "886f11d28fe78805b65430addc61bcf09a87d381de77dc13d1675713ec730050.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": " denotes the classification loss. The inner optimization injects adversarial perturbation " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": " into the feature space of " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": " defines the perturbation size, and " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": " is a hyperparameter that controls the strength of regularization. " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": " denotes the model parameters at the " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": "-th layer. We derive a closed-form solution for the adversarial perturbation under the " + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 513, + 554, + 597 + ], + "type": "text", + "content": " norm following [12]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 601, + 553, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 601, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 375, + 601, + 553, + 628 + ], + "type": "interline_equation", + "content": "\\epsilon_ {i, \\mathrm {c l s}} ^ {*} \\approx \\rho \\cdot \\frac {\\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {P} _ {i})}{\\| \\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}} (\\mathbf {P} _ {i}) \\| _ {2}}. \\tag {6}", + "image_path": "83d968d27b140d193acce38595dfc54e461f7fa3b9fe2156bf27f49123d8d26a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 633, + 553, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 553, + 680 + ], + "type": "text", + "content": "The perturbation can elevate the gradient of critical regions, thereby increasing their saliency during training. The gradient change induced by the perturbation on the classification branch can be calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 689, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 689, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\Delta \\left(\\nabla_ {\\mathbf {P} _ {i}} \\mathcal {L} _ {\\mathrm {c l s}}\\right) = \\left\\| \\nabla_ {\\mathbf {P} _ {i}} \\left(\\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i} + \\epsilon_ {i, \\mathrm {c l s}} ^ {*}\\right) - \\mathcal {L} _ {\\mathrm {c l s}} \\left(\\mathbf {P} _ {i}\\right)\\right) \\right\\|. \\tag {7}", + "image_path": "9fc01a52fdcd7d9aabe91d233124e36f1bfac083d94ae9e28e886c6c04083c1a.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4716" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 211, + 156 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 211, + 156 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 211, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 211, + 156 + ], + "type": "image", + "image_path": "6c3dafd05e366d1c8e20b3717800f36da73f88a0378f437de00087f3d2819032.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 71, + 361, + 156 + ], + "blocks": [ + { + "bbox": [ + 211, + 71, + 361, + 156 + ], + "lines": [ + { + "bbox": [ + 211, + 71, + 361, + 156 + ], + "spans": [ + { + "bbox": [ + 211, + 71, + 361, + 156 + ], + "type": "image", + "image_path": "dd9bab4cf0fb41c54e42f3c5664afb6c6c8c404d3cb32ec213b66db81e78e0e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 361, + 72, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 361, + 72, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 361, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 361, + 72, + 553, + 156 + ], + "type": "image", + "image_path": "728dfac29a8eb039e8dcfd1078460a60c4f1c2e3c32caf5542c37b5deade5283.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 157, + 211, + 242 + ], + "blocks": [ + { + "bbox": [ + 60, + 157, + 211, + 242 + ], + "lines": [ + { + "bbox": [ + 60, + 157, + 211, + 242 + ], + "spans": [ + { + "bbox": [ + 60, + 157, + 211, + 242 + ], + "type": "image", + "image_path": "a2d4c907d5a5ae2e7f9a631618afbee807ac66f77009eb57766cdfdfba328d8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "Figure 5. Average saliency visualization in (a) FCOS and (b) FCOS w/ API. The saliency maps are first computed based on the " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "-norm of the gradient [13] at neck feature " + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": ". The saliency of each object is then summed and annotated within their respective bounding box, which is different between (a) and (b). The average saliency of each object scale is calculated as the mean saliency across objects in the corresponding size category, displayed in the bottom-left corner of each image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 211, + 157, + 361, + 242 + ], + "blocks": [ + { + "bbox": [ + 211, + 157, + 361, + 242 + ], + "lines": [ + { + "bbox": [ + 211, + 157, + 361, + 242 + ], + "spans": [ + { + "bbox": [ + 211, + 157, + 361, + 242 + ], + "type": "image", + "image_path": "dc9290d4ac9dce2f6cf1e326f26c5559de9e43770ea7cf5ca17e9ba3a9275240.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 157, + 553, + 242 + ], + "blocks": [ + { + "bbox": [ + 361, + 157, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 361, + 157, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 361, + 157, + 553, + 242 + ], + "type": "image", + "image_path": "b083ab31446257913304a21d7a3a130ee3d51019e0e4a13152f32b021df4cc4b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 310, + 295, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 310, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 295, + 466 + ], + "type": "text", + "content": "High " + }, + { + "bbox": [ + 54, + 310, + 295, + 466 + ], + "type": "inline_equation", + "content": "\\Delta (\\nabla_{\\mathbf{P}_i}\\mathcal{L}_{\\mathrm{cls}})" + }, + { + "bbox": [ + 54, + 310, + 295, + 466 + ], + "type": "text", + "content": " corresponds to regions where model outputs are highly sensitive to minor changes in activations, indicating the presence of semantic information about the image [49]. This elevated gradient enhances the feature saliency of these critical regions and prompts the refinement of object features during training. The effect particularly benefits tiny objects as they exhibit weak high-frequency signatures and face a training deficiency problem. The enhanced object saliency phenomenon is discussed in Sec. 4.6 through Fig. 5. Furthermore, the adversarial training facilitates robust feature representations and reduces the sensitivity to high-frequency noise [63], thereby enhancing the discrimination of tiny objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "text", + "content": "We formulate the adversarial perturbation " + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}^{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "text", + "content": " for each FPN layer as an expectation over " + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "text", + "content": " object detection branches (e.g., the classification, regression and center-ness branch in a FCOS detector [39]), each scaled by a corresponding balancing parameter " + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 55, + 467, + 295, + 537 + ], + "type": "text", + "content": ". This is generated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 536, + 295, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 536, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 129, + 536, + 295, + 569 + ], + "type": "interline_equation", + "content": "\\epsilon_ {i} ^ {\\mathrm {a d v}} = \\sum_ {m = 1} ^ {M} \\lambda_ {m} \\cdot \\epsilon_ {i, m} ^ {*}, \\tag {8}", + "image_path": "61d56aed8322d9a5dab0dd1928d23c3157d5430571d66a2ff008ccb33c9e4eec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,m}^{*}" + }, + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "type": "text", + "content": " represents the adversarial perturbation for the " + }, + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 574, + 295, + 635 + ], + "type": "text", + "content": "-th branch. The design allows for a balanced incorporation of detection semantics across multiple branches, ensuring that the perturbations are distributed for more effective training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 187, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 187, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 187, + 659 + ], + "type": "text", + "content": "3.3. Auxiliary Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "The optimization leverages a total loss function that integrates both the original detection loss " + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{detection}}" + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": " and an auxiliary loss derived from the HBS and API enhanced features. This joint loss function is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 308, + 555, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 308, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 350, + 308, + 555, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\underbrace {\\lambda \\cdot \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {i} ^ {\\mathrm {a u x}} \\left(\\mathbf {P} _ {i} ^ {E} + \\epsilon_ {i} ^ {\\mathrm {a d v}}\\right)} _ {\\text {a u x i l i a r y l o s s}}, \\tag {9}", + "image_path": "b55162b1b1fe858f3ef52df5de58adcb9210c6b378a168d9ca799af9d115dcd7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i^{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "content": " denotes the auxiliary loss component for the " + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "content": "-th layer, which is summed and scaled by a balancing hyperparameter " + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "content": ". The perturbation " + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_i^{\\mathrm{adv}}" + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "content": " is applied to the HBS-enhanced feature " + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i^E" + }, + { + "bbox": [ + 313, + 357, + 555, + 502 + ], + "type": "text", + "content": ". The auxiliary optimization process updates the model's parameters in concert with the original loss, leading to more robust and discriminative features for tiny objects. We visualize the effectiveness of the method in Fig. 2. The frequency signatures of tiny objects become more pronounced after applying SET, manifesting distinct frequency response on the edges of the spectrum compared to vanilla FCOS, benefiting their detection amid high-frequency noise and in the presence of larger objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 512, + 395, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 395, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 395, + 525 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 531, + 510, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 510, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 510, + 544 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed method on AI-TOD [44], VisDrone2019 [65], DOTA-v2.0 [47] and COCO 2017 [22]. The main experiments were conducted using the AI-TOD dataset, with an average object size of 12.8 pixels, substantially smaller than other object detection datasets such as PASCAL VOC (156.6 pixels) [10] and MS COCO (99.5 pixels) [21]. Furthermore, we test our method on the VisDrone2019 [65] and DOTA-v2.0 [47], which comprises high-resolution drone-shot images with a substantial proportion of tiny instances. We also conduct experiments on the general object detection benchmark COCO 2017 [22], incorporating " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_s" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " metrics to evaluate objects of very tiny (2-8 pixels), tiny (8-16 pixels) and small (16-32 pixels), following [44]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4717" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 106, + 527, + 445 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Main results with various frameworks on AI-TOD [44]. Models are trained on the AI-TOD trainval and validated on the AI-TOD test. We report APs (\\%) with different IoU thresholds and APs (\\%) for objects of various sizes based on the AI-TOD criterion. RFLA is based on Cascade R-CNN. The * denotes using P2~P6 FPN features. The **bold** indicates the best result." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 106, + 527, + 445 + ], + "lines": [ + { + "bbox": [ + 78, + 106, + 527, + 445 + ], + "spans": [ + { + "bbox": [ + 78, + 106, + 527, + 445 + ], + "type": "table", + "html": "
FrameworkAP\\( AP_{0.5} \\)\\( AP_{0.75} \\)\\( AP_{vt} \\)\\( AP_t \\)\\( AP_s \\)
PAA [19]10.026.56.73.510.513.1
ATSS [61]11.628.57.62.511.915.9
Centernet [9]13.439.25.03.812.117.7
DetectoRS [34]14.832.811.50.010.818.3
DotD [50]16.139.210.68.317.618.1
DetectoRS w/NWD [43]20.849.314.36.419.729.6
DetectoRS w/ SR-TOD [3]24.054.617.110.124.829.3
One-stage
RetinaNet [24]7.220.53.62.56.811.0
RetinaNet w/ SET9.124.65.22.98.212.8
RetinaNet* [24]13.331.78.63.914.417.7
RetinaNet* w/ QueryDet [55]12.229.37.32.410.518.5
RetinaNet* w/ SET16.538.610.95.015.721.8
FCOS [39]12.029.08.02.511.917.1
FCOS w/ SET14.234.99.82.913.020.2
FCOS* [39]15.135.810.25.916.618.8
FCOS* w/ SET18.042.712.47.218.523.0
Multi-stage
Faster R-CNN [35]11.126.38.10.07.223.3
Faster R-CNN w/ SET12.228.29.70.19.124.3
Cascade R-CNN [2]13.630.310.60.09.925.5
Cascade R-CNN w/ SET14.832.411.20.410.826.7
RFLA [52]21.750.515.38.321.824.5
RFLA w/ SR-TOD [3]21.850.815.49.721.827.4
RFLA w/ SET24.955.617.99.825.231.0
Transformer-based
DINO-5scale [60]23.256.615.49.923.129.3
DINO-5scale w/ DNTR [25]26.256.720.212.826.431.0
DINO-5scale w/ SET26.657.120.813.227.131.5
", + "image_path": "7d7e3a25a001472d7b50707dbc7936aaaede3092b825e6e37131a8f43965f4d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 455, + 294, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 455, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 455, + 294, + 647 + ], + "type": "text", + "content": "Implementation Details. We conducted the experiments on a computer with an NVIDIA RTX 3090 GPU. The experiments are implemented with PyTorch [32], with core codes built upon MMdetection [5]. The ImageNet [36] pre-trained models are used as the backbones. All CNN-based models utilize the ResNet-50 [15] backbone, trained using the Stochastic Gradient Descent (SGD) optimizer for 12 epochs with 0.9 momentum, 0.0001 weight decay, and a batch size 2. The initial learning rate is 0.005, decaying at the 8th and 11th epochs. The data processing adheres to the default configurations of each dataset (e.g., fixed at " + }, + { + "bbox": [ + 54, + 455, + 294, + 647 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 54, + 455, + 294, + 647 + ], + "type": "text", + "content": " for AI-TOD). We also train a transformer-based detector, DINO [60], with 5-scale feature maps for 36 epochs as a baseline. The training uses an Adam optimizer with a weight decay of 0.0001, following the random crop and scale augmentation strategies of DETR [4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 658, + 168, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 168, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 168, + 670 + ], + "type": "text", + "content": "4.2. Results on AI-TOD" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the proposed SET method on various detectors and compare it with state-of-the-art TOD methods on the AI-TOD benchmark [51]. Tab. 1 shows that SET enhances" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": "all baselines by about " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": " AP, which is significant. Specifically, SET improves the one-stage detector FCOS [39] and RetinaNet [24] by " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "2.2\\% / 1.9\\%" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "1.1\\% / 1.4\\%" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": ", respectively. We further revise the detector architecture and use P2~P6 FPN features, a representative approach for TOD tasks where the high-resolution P2 layer benefits tiny objects. With the adjustment, SET boosts the RetinaNet and FCOS baselines by " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "3.2\\% / 2.9\\%" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "1.1\\% / 1.3\\% \\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": " a significant margin. In contrast, the state-of-the-art SOD method QueryDet [55] decreases the baseline performance, showing limitations in addressing the TOD challenge. SET also generalizes to multi-stage detectors, enhancing the Faster R-CNN [35] and Cascade R-CNN [2] baselines by " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "1.1\\% / 1.2\\%" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "1.9\\% / 1.1\\%" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 313, + 455, + 555, + 624 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "SET works orthogonally with the state-of-the-art TOD method RFLA [52], achieving a " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " AP increment over the RFLA baseline and surpassing the comparison SR-TOD [3] by " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". We also investigate SET's compatibility with transformer-based detectors. As shown, SET results in " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " AP increase over the DINO-5scale [60] baseline, outperforming the prior" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "4718" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 112, + 297, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 2. Detection performance on the VisDrone2019 [65] validation set. RFLA is based on Cascade R-CNN. The **bold** denotes the best result." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 112, + 297, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 112, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 297, + 228 + ], + "type": "table", + "html": "
FrameworkAPAPvtAPtAPs
RetinaNet [24]15.20.41.69.3
Faster R-CNN [35]23.10.15.119.6
FCOS [39]19.90.74.515.7
FCOS w/ SET21.91.65.717.9
Cascade R-CNN [2]23.00.15.019.4
Cascade R-CNN w/ SET24.31.26.321.0
RFLA [52]27.24.513.023.6
RFLA w/ SR-TOD [3]27.84.813.124.5
RFLA w/ SET28.55.213.524.6
", + "image_path": "89d823cdf0677a5fb80331337e89b5f40a314d92f1b5a2e120a154d497a6c079.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 291 + ], + "type": "text", + "content": "art DNTR [25]. Notably, DINO-5scale w/SET derives " + }, + { + "bbox": [ + 55, + 241, + 296, + 291 + ], + "type": "inline_equation", + "content": "26.6\\%" + }, + { + "bbox": [ + 55, + 241, + 296, + 291 + ], + "type": "text", + "content": " AP, outperforming competitors including DotD [50], NWD [43], and SR-TOD [3]. The results demonstrate SET as a robust solution for the tiny object detection task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 275, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 275, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 275, + 312 + ], + "type": "text", + "content": "4.3. Results on VisDrone2019 and DOTA-V2.0" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": "We further evaluate SET on the drone-shot image benchmark VisDrone2019 [65], which features tiny and small objects. As shown in Tab. 2, the proposed SET effectively improves the performance of both one-stage and multi-stage detectors and exhibits consistent " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " improvements. Specifically, SET improved the the one-stage detector FCOS by " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "2.0\\% /0.9\\% /1.2\\%" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " AP, " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " , respectively. SET also improves multi-stage detector Cascade RCNN [2] by " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "1.1\\% /1.3\\%" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " , respectively. Notably, SET improves the prior art RFLA, achieving a " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " AP with " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " increment in " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 319, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": "Tab. 3 demonstrates the DOTA-v2.0 [8] results. As shown, SET obtains a substantial performance gain of " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": " with the FCOS and AutoAssign baseline, underscoring the efficacy of our method in detecting tiny objects. Compared to the state-of-the-art comparisons RFLA [52] and DCFL [53], SET achieves " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": " AP gain, respectively. More experimental results on the TinyPerson [56] and SeaPerson [57] dataset are provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 163, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 163, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 163, + 582 + ], + "type": "text", + "content": "4.4. Results on COCO" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "text", + "content": "We further conduct experiments on the general object detection benchmark MS COCO [21]. As shown in Tab. 9, SET achieves " + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "text", + "content": " AP improvement over the FCOS baseline, with a notable " + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "text", + "content": " improvement in " + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_t" + }, + { + "bbox": [ + 55, + 588, + 295, + 661 + ], + "type": "text", + "content": ", demonstrating that SET can enhance the detection of tiny objects in general detection tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 149, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 149, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 149, + 684 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In the following experiments, we explore the best structure of SET by ablating and tuning its parts using AI-TOD." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 112, + 556, + 216 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 3. Detection performance on DOTA-v2.0 [8]. Note that models are trained on the DOTA-v2.0 train and validated on the DOTA-v2.0 val. The bold denotes the best result." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 112, + 556, + 216 + ], + "lines": [ + { + "bbox": [ + 315, + 112, + 556, + 216 + ], + "spans": [ + { + "bbox": [ + 315, + 112, + 556, + 216 + ], + "type": "table", + "html": "
FrameworkAPAPvtAPtAPs
ATSS [61]32.70.76.923.4
Faster R-CNN [35]35.60.07.128.9
FCOS [39]31.80.34.019.4
FCOS w/ RFLA [52]32.10.76.823.6
FCOS w/ DCFL [53]33.80.56.924.2
FCOS w/ SET34.70.67.124.8
AutoAssign [64]33.80.97.322.4
AutoAssign w/ SET35.21.18.623.6
", + "image_path": "3e418ef43ecbf75785e1b4915036554f29684eeaed0a316b14a174d55c300e33.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 264, + 561, + 301 + ], + "blocks": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "type": "text", + "content": "Table 4. Detection performance on MS COCO [21]. Note that models are trained on COCO train2017 and validated on COCO val2017. " + }, + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "type": "inline_equation", + "content": "(\\cdot)^{+}" + }, + { + "bbox": [ + 313, + 223, + 555, + 257 + ], + "type": "text", + "content": " denotes performance increment." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 264, + 561, + 301 + ], + "lines": [ + { + "bbox": [ + 315, + 264, + 561, + 301 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 561, + 301 + ], + "type": "table", + "html": "
FrameworkAPAPvtAPtAPsAPm
FCOS [39]36.47.919.627.243.6
FCOS w/ SET37.4+1.08.3+0.423.3+1.727.6+0.444.8+1.2
", + "image_path": "f863f6561d5f03b74ae21e79ce4ca16a5b854d31fbca21d6243119a2db8b6c92.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": "Effectiveness and Cost of Components. We present quantitative improvements of SET components in Tab. 5. As shown, applying the adaptive smoothing operations to full features " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "(+\\phi_{i}(\\mathbf{P}_{i},r))" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": " achieves a " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": " increase in AP and " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": ", surpassing the performance of static high-frequency filtering in Fig. 1. With decoupling, the HBS module boosts the overall AP by " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": " AP and " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{vt}" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": ". The API module alone achieves a " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": " AP improvement. Combining HBS and API results in a total AP increase of " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": ". Results demonstrate the synergy between the heterogeneous operations, where HBS effectively smooths the clutter background and API prompts the learning of object features, cumulatively improving the detection accuracy of tiny objects. Tab. 5 also presents the cost of components. Results show that the HBS module leads to " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 313, + 318, + 556, + 570 + ], + "type": "text", + "content": " increase in training time, computational and memory cost. The API module results in a more significant increase in training time due to adversarial training but adds no forward computational cost and minimal increase in memory usage. The overall minimal impact indicates that SET can be implemented with negligible additional overhead." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Hyper-parameter Selection. We first select the hyperparameters channel reduction rate " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " and perturbation size " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " using the FCOS [39] framework trained three epochs in Fig. 6 (a). Model performance (AP) with different hyperparameter combinations " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{r, \\rho\\}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " is presented. Results indicate that performance initially improves and then declines as " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " varies from 2 to 16, with " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " perform worse than the baseline. Additionally, SET with adversarial perturbations " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\rho > 0)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " exhibits stronger performance than without " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\rho = 0)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", but large-sized perturbations " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\rho = 5)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " performs worse than all alternatives. Exploring the setups, we identify " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{r, \\rho\\} = \\{4, 1\\}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " as the combination for SET. Fig. 6 (b) explores the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4719" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 555, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 5. The impact of different components and various channel reduction operations in SET. The hyper-parameters are fixed with channel reduction rate " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " being 4, perturbation size " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " being 1, and the balancing hyper-parameters " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " of auxiliary optimization being 1." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 555, + 228 + ], + "type": "table", + "html": "
Frameworkφig(Si)APAPvtAPtTime(s/batch)GFLOPsParams (M)
FCOS [39]--12.02.511.90.17126.1337.04
+ φi(Pi, r)--12.72.712.90.17127.3637.41
Eqn. 3313.42.712.6
Eqn. 3log2(Si)/2×2+113.92.712.9
+ HBSEqn. 3513.12.312.50.20127.3237.41
Eqn. 3log2(Si+1)/2×2-113.42.812.4
Eqn. 3712.12.512.2
+ API--13.22.612.40.39126.1337.16
+ Both (SET)Eqn. 3Eqn. 414.22.913.00.42127.3237.85
", + "image_path": "8717e3c301cb756d30cf517a0e9b905696461d547d18febfe946195f015fd44a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 239, + 176, + 357 + ], + "blocks": [ + { + "bbox": [ + 56, + 239, + 176, + 357 + ], + "lines": [ + { + "bbox": [ + 56, + 239, + 176, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 176, + 357 + ], + "type": "image", + "image_path": "c206ef2afd100a6fd3aeb89ceef8ffd5959e034570a8b9a49dfd9ccb2f45068f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": "Figure 6. We select hyper-parameters channel reduction rate " + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": " with perturbation size " + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": " in (a), and the balancing hyper-parameter " + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 365, + 295, + 399 + ], + "type": "text", + "content": " of Eqn. 9 in (b)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 239, + 294, + 357 + ], + "blocks": [ + { + "bbox": [ + 179, + 239, + 294, + 357 + ], + "lines": [ + { + "bbox": [ + 179, + 239, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 179, + 239, + 294, + 357 + ], + "type": "image", + "image_path": "68468b3b4f4089924a4fda40e525e9ca399c0f1a7bbe44ed024ea4e6851b68e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 392 + ], + "type": "text", + "content": "Figure 7. The analysis examines the adaptive smoothing operation, obtained from 1000 randomly selected images from the Vis-Drone2019 [65] train set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": "balancing hyper-parameter " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " in Eqn. 9 that controls the proportion of the auxiliary optimization, using the FCOS [39], RetinaNet [24], and Faster R-CNN [35] framework. As observed, the performances increase first and then decrease with the increase of " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " yielding the best performance across all frameworks. However, the efficacy of SET diminishes when " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " exceeds the 1, as larger-sized " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda > 1" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": ") can impose too strong regularization to the detector, reducing the detection accuracy despite increasing robustness. Therefore, we set the hyper-parameter " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " to 1 for experiments in this paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 551, + 201, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 201, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 201, + 565 + ], + "type": "text", + "content": "4.6. Enhanced Object Saliency" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "We demonstrate effectiveness of API in increasing the saliency of object features in Fig. 5. Comparing the average saliency of each object scale, tiny objects are the least salient due to their inherently limited pixel input. API leads to a notable increase in the average saliency of tiny objects (e.g., from 53.99 to 69.93), benefited from the injection of perturbations (Eqn. 7). Fig. 5 also demonstrates that API can enhance the feature saliency of small and medium objects, though the degree of improvement is less. The comparison underscores that API could benefit general object detection, with more significant and essential improvements for tiny objects." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 243, + 529, + 350 + ], + "blocks": [ + { + "bbox": [ + 341, + 243, + 529, + 350 + ], + "lines": [ + { + "bbox": [ + 341, + 243, + 529, + 350 + ], + "spans": [ + { + "bbox": [ + 341, + 243, + 529, + 350 + ], + "type": "image", + "image_path": "5a6f021a8eb905f53dc3e1a29e203e9461e4e8afaf5e2cfff66bb1bd9946ca44.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 401, + 463, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 463, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 463, + 414 + ], + "type": "text", + "content": "4.7. Effectiveness in Smoothing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "content": "In Fig. 7, we perform a statistic analysis on " + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "inline_equation", + "content": "P_{i}^{bg}" + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "content": " to investigate the adaptive smoothing operation. We convert feature " + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "inline_equation", + "content": "P_{i}^{bg}" + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\phi_{i}(P_{i}^{bg}, r)" + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "content": " into the frequency space using FFT [31] and separate the amplitude spectrums of these features into three parts (low-, mid-, and high-frequency) using static thresholds as Fig. 1. The amplitudes in each band are then summed and averaged to measure the signal strength. Results verify that the channel reducing and expand design effects in smoothing. Comparing " + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "inline_equation", + "content": "A(P_{i})(u, v)" + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "inline_equation", + "content": "A(\\phi_{i}(P_{i}^{bg}, r))(u, v)" + }, + { + "bbox": [ + 313, + 418, + 555, + 563 + ], + "type": "text", + "content": ", the low-frequency signals are enhanced (from 0.3787 to 0.5205), while mid- and high-frequency signals are attenuated." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "This paper proposes the Spectral Enhancement for Tiny objects (SET) method, which amplifies the frequency signatures of tiny objects in a heterogeneous architecture. SET includes two modules. The Hierarchical Background Smoothing (HBS) module suppresses high-frequency noise in the background through adaptive smoothing operations. The Adversarial Perturbation Injection (API) module leverages adversarial perturbations to prompt the refinement of object features during training. Extensive experiments on four datasets demonstrate its effectiveness." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4720" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 196 + ], + "type": "text", + "content": "Acknowledgments. The work was supported by the National Natural Science Foundation of China (No. 61827901, 623B2016, and 62406298). This research was also supported by the National Key Research and Development Program of China (Grant No. 2023YFC3306401), Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F020007, Beijing Natural Science Foundation L223024, L244043 and Z241100001324017, \"One Thousand Plan\" projects in Jiangxi Province Jxsq2023102268, and the Beijing Jiaotong University \"Jingying Plan\" No. K24XKRC00130." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 217, + 115, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 115, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 115, + 229 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 237, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 237, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 294, + 280 + ], + "type": "text", + "content": "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, pages 206-221, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 282, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 295, + 315 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 317, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 349 + ], + "type": "text", + "content": "[3] Bing Cao, Haiyu Yao, Pengfei Zhu, and Qinghua Hu. Visible and clear: Finding tiny objects in difference map. ECCV, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 352, + 294, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 394 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 397, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 294, + 451 + ], + "type": "text", + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 454, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 496 + ], + "type": "text", + "content": "[6] Yukang Chen, Peizhen Zhang, Zeming Li, Yanwei Li, Xiangyu Zhang, Lu Qi, Jian Sun, and Jiaya Jia. Dynamic scale training for object detection. arXiv preprint arXiv:2004.12432, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 499, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 541 + ], + "type": "text", + "content": "[7] Gong Cheng, Xiang Yuan, Xiwen Yao, Kebing Yan, Qinghua Zeng, Xingxing Xie, and Junwei Han. Towards large-scale small object detection: Survey and benchmarks. IEEE TPAMI, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 294, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 598 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 598 + ], + "type": "text", + "content": "[8] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE TPAMI, 44(11):7778-7796, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 633 + ], + "type": "text", + "content": "[9] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In ICCV, pages 6569-6578, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "text", + "content": "[10] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. ICCV, 88:303-308, 2009. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 4, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu. Distilling object detectors via decoupled features. In CVPR, pages 2154-2164, 2021. 5, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 185 + ], + "type": "text", + "content": "[14] Kaiming He and Jian Sun. Convolutional neural networks at constrained time cost. In CVPR, pages 5353-5360, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[16] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37. Springer, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[17] Masato Ishii and Atsushi Sato. Training deep neural networks with adversarially augmented features for small-scale training datasets. In IJCNN, pages 1-8. IEEE, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[18] Weikuan Jia, Meili Sun, Jian Lian, and Sujuan Hou. Feature dimensionality reduction: a review. Complex & Intelligent Systems, 8(3):2663-2693, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "type": "text", + "content": "[19] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, pages 355-371, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "text", + "content": "[20] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. In *ICLR*, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 378, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 553, + 422 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 1, 5, 7, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 423, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 466 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 1, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 6, 7, 8, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "text", + "content": "[25] Hou-I Liu, Yu-Wen Tseng, Kai-Cheng Chang, Pin-Jyun Wang, Hong-Han Shuai, and Wen-Huang Cheng. A denoising fpn with transformer r-cnn for tiny object detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 633 + ], + "type": "text", + "content": "[26] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[28] Jiajun Lu, Hussein Sibai, and Evan Fabry. Adversarial examples that fool detectors. arXiv preprint arXiv:1712.02494, 2017.3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4721" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] T. Na, J. H. Ko, and S. Mukhopadhyay. Cascade adversarial machine learning regularized with a unified embedding. In ICLR, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[30] Junhyug Noh, Wonho Bae, Wonhee Lee, Jinhwan Seo, and Gunhee Kim. Better to follow, follow to be better: Towards precise supervision of feature super-resolution for small object detection. In ICCV, pages 9725-9734, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 295, + 174 + ], + "type": "text", + "content": "[31] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[32] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS Workshops, pages 1-12, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "text", + "content": "[33] Vitali Petsiuk, Rajiv Jain, Varun Manjunatha, Vlad I Morariu, Ashutosh Mehra, Vicente Ordonez, and Kate Saenko. Black-box explanation of object detectors via saliency maps. In CVPR, pages 11443-11452, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[34] Siyuan Qiao, Liang-Chieh Chen, and Alan Yuille. Detectors: Detecting objects with recursive feature pyramid and switchable atrous convolution. In CVPR, pages 10213-10224, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 363 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(06):1137-1149, 2017. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 419 + ], + "type": "text", + "content": "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. ICCV, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 295, + 454 + ], + "type": "text", + "content": "[37] Bharat Singh, Mahyar Najibi, Abhishek Sharma, and Larry S Davis. Scale normalized image pyramids with autofocus for object detection. IEEE TPAMI, 44(7):3749-3766, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "text", + "content": "[38] Dawn Song, Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Florian Tramer, Atul Prakash, and Tadayoshi Kohno. Physical adversarial examples for object detectors. In 12th USENIX workshop on offensive technologies (WOOT '18), 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 544 + ], + "type": "text", + "content": "[39] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, pages 9627-9636, 2019. 1, 5, 6, 7, 8, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 577 + ], + "type": "text", + "content": "[40] F Tramér, D Boneh, A Kurakin, I Goodfellow, N Papernot, and P McDaniel. Ensemble adversarial training: Attacks and defenses. In *ICLR*, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 611 + ], + "type": "text", + "content": "[41] S Velliangiri, SJPCS Alagumuthukrishnan, et al. A review of dimensionality reduction techniques for efficient computation. Procedia Computer Science, 165:104-111, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "type": "text", + "content": "[42] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, pages 5495-5504, 2018. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "text", + "content": "[43] Jinwang Wang, Chang Xu, Wen Yang, and Lei Yu. A normalized gaussian Wasserstein distance for tiny object detection. arXiv preprint arXiv:2110.13389, 2021. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[44] Jinwang Wang, Wen Yang, Haowen Guo, Ruixiang Zhang, and Gui-Song Xia. Tiny object detection in aerial images. In ICPR, pages 3791-3798, 2021. 1, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[45] Kunyu Wang, Xueyang Fu, Yukun Huang, Chengzhi Cao, Gege Shi, and Zheng-Jun Zha. Generalized uav object detection via frequency domain disentanglement. In CVPR, pages 1064-1073, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[46] Svante Wold, Kim Esbensen, and Paul Geladi. Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1-3):37-52, 1987. 4, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "text", + "content": "[47] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In CVPR, pages 3974-3983, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 197, + 553, + 239 + ], + "type": "text", + "content": "[48] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection. In ICCV, pages 1369-1378, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "type": "text", + "content": "[49] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In CVPR, pages 501-509, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 308 + ], + "type": "text", + "content": "[50] Chang Xu, Jinwang Wang, Wen Yang, and Lei Yu. Dot distance for tiny object detection in aerial images. In ICCV, pages 1192-1201, 2021. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 315, + 310, + 553, + 353 + ], + "type": "text", + "content": "[51] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Detecting tiny objects in aerial images: A normalized wasserstein distance and a new benchmark. IS-PRS J. Photo. and Rem. Sen., 190:79-93, 2022. 6, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "type": "text", + "content": "[52] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543, 2022. 2, 6, 7, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 442 + ], + "type": "text", + "content": "[53] Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Dynamic coarse-to-fine learning for oriented tiny object detection. In CVPR, pages 7318-7328, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 553, + 487 + ], + "type": "text", + "content": "[54] Sheng Xu, Mingze Wang, Yanjing Li, Mingbao Lin, Baochang Zhang, David Doermann, and Xiao Sun. Learning 1-bit tiny object detector with discriminative feature refinement. In ICML, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 488, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 488, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 315, + 488, + 553, + 531 + ], + "type": "text", + "content": "[55] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 315, + 533, + 553, + 566 + ], + "type": "text", + "content": "[56] Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, and Zhenjun Han. Scale match for tiny person detection. In WACV, pages 1257–1265, 2020. 7, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 567, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 567, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 315, + 567, + 553, + 622 + ], + "type": "text", + "content": "[57] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. Cpr++: Object localization via single coarse point supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4908-4925, 2024. 7, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 315, + 624, + 553, + 666 + ], + "type": "text", + "content": "[58] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small object detection via coarse-to-fine proposal generation and imitation learning. In ICCV, pages 6317-6327, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 669, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 669, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 315, + 669, + 553, + 690 + ], + "type": "text", + "content": "[59] Haichao Zhang and Jianyu Wang. Towards adversarially robust object detection. In ICCV, pages 421-430, 2019. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 692, + 553, + 713 + ], + "type": "text", + "content": "[60] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4722" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 319 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[61] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In CVPR, pages 9759-9768, 2020. 6, 7, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 184 + ], + "type": "text", + "content": "[62] Qijie Zhao, Tao Sheng, Yongtao Wang, Zhi Tang, Ying Chen, Ling Cai, and Haibin Ling. M2det: A single-shot object detector based on multi-level feature pyramid network. In AAAI, pages 9259-9266, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 218 + ], + "type": "text", + "content": "[63] Wen Zhou, Xin Hou, Yongjun Chen, Mengyun Tang, Xiangqi Huang, Xiang Gan, and Yong Yang. Transferable adversarial perturbations. In ECCV, pages 452-467, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[64] Benjin Zhu, Jianfeng Wang, Zhengkai Jiang, Fuhang Zong, Songtao Liu, Zeming Li, and Jian Sun. Autoassign: Differentiable label assignment for dense object detection. arXiv preprint arXiv:2007.03496, 2020. 7, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[65] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Haibin Ling, Qinghua Hu, Qinqin Nie, Hao Cheng, Chenfeng Liu, Xiaoyu Liu, et al. Visdrone-det2018: The vision meets drone object detection in image challenge results. In ECCV Workshops, pages 0–0, 2018. 5, 7, 8" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4723" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_content_list.json b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..197d092ff8fba16a6f054b2256473edb79c8c007 --- /dev/null +++ b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_content_list.json @@ -0,0 +1,1611 @@ +[ + { + "type": "text", + "text": "$\\mathbf{SF}^{2}\\mathbf{T}$ : Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding", + "text_level": 1, + "bbox": [ + 109, + 128, + 887, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yangliu Hu $^{1}$ , Zikai Song $^{1\\dagger}$ , Na Feng $^{1}$ , Yawei Luo $^{2}$ , Junqing Yu $^{1}$ , Yi-Ping Phoebe Chen $^{3}$ , Wei Yang $^{1\\dagger}$ $^{1}$ Huazhong University of Science and Technology $^{2}$ Zhejiang University $^{3}$ La Trobe University", + "bbox": [ + 101, + 202, + 892, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huyangliu,skyesong,fengna,yjqing,weiyangcs}@hust.edu.cn", + "bbox": [ + 250, + 244, + 746, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yaweiluo@zju.edu.cn phoebe.chen@latrobe.edu.au", + "bbox": [ + 285, + 263, + 707, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 311, + 326, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-based Large Language Models (Video-LLMs) have witnessed substantial advancements in recent years, propelled by the advancement in multi-modal LLMs. Although these models have demonstrated proficiency in providing the overall description of videos, they struggle with fine-grained understanding, particularly in aspects such as visual dynamics and video details inquiries. To tackle these shortcomings, we find that fine-tuning Video-LLMs on self-supervised fragment tasks, greatly improve their fine-grained video understanding abilities. Hence we propose two key contributions: (1) Self-Supervised Fragment Fine-Tuning $(SF^2 T)$ , a novel effortless fine-tuning method, employs the rich inherent characteristics of videos for training, while unlocking more fine-grained understanding ability of Video-LLMs. Moreover, it relieves researchers from labor-intensive annotations and smartly circumvents the limitations of natural language, which often fails to capture the complex spatiotemporal variations in videos; (2) A novel benchmark dataset, namely FineVidBench, for rigorously assessing Video-LLMs' performance at both the scene and fragment levels, offering a comprehensive evaluation of their capabilities. We assessed multiple models and validated the effectiveness of $SF^2 T$ on them. Experimental results reveal that our approach improves their ability to capture and interpret spatiotemporal details.", + "bbox": [ + 88, + 359, + 485, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 763, + 220, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have showcased significant emergent capabilities, such as in-context learning [19], instruction-following [23], and chain-of-thought reasoning [30], driven by expansive datasets and advanced model architectures. Extending these advancements, Video-LLMs through mechanisms like pooling or query aggregation", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b7eb9014d7abb5b36913b7f76bf78d0aca112be4c48c98f53b28ffaa2069a16.jpg", + "image_caption": [ + "Figure 1. Performance w/ and w/o $\\mathbf{SF}^2\\mathbf{T}$ . We evaluated four advanced Video-LLMs w/ and w/o $\\mathrm{SF}^2\\mathrm{T}$ on our proposed FineVidBench with two baselines: (1) Base: performance without any fine-tuning (blue dashed), and (2) Base (SFT): performance with supervised fine-tuning (red dashed). After applying $\\mathrm{SF}^2\\mathrm{T}$ , all models showed significant improvements (solid blue and red), underscoring its broad effectiveness." + ], + "image_footnote": [], + "bbox": [ + 535, + 311, + 883, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "across numerous visual tokens, have broadened the scope of LLMs to encompass video information processing [11, 14, 35]. This evolution markedly advances their potential for in-depth real-world comprehension, opening applications in intelligent surveillance, virtual reality, and autonomous driving, further enriching the landscape of video analytics and interpretation.", + "bbox": [ + 511, + 684, + 906, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various Video-LLMs, exemplified by GPT4-V, VideoLaMA 2 [4], MiniCPM-V [34], and Qwen2-VL [28], have been crafted by leading corporations and research institutions, demonstrating proficiency in capturing the overarching content of videos. When adapting to new videos and tasks, they predominantly rely on Supervised FineTuning (SFT) [26] or Reinforcement Learning from Hu", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 120, + 887, + 254, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29108", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "man Feedback (RLHF) [39], both of which are heavily contingent upon extensive manual annotation. This dependence poses several key problems: (1) it necessitates substantial human resources, particularly highly trained annotators; (2) the inherent complexity of video content and task demands frequently introduces inconsistencies and subjectivity, rendering the maintenance of high-quality annotations particularly arduous; and (3) subtle temporal variations across video frames are challenging to articulate with precision, often yielding generalized descriptions that constrain the Video-LLMs' potential. Consequently, existing Video-LLMs struggle with fine-grained video understanding tasks, particularly in aspects such as visual dynamics (e.g., motion patterns, object interactions) and video details inquiries (e.g., positional changes, detail variations).", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we observe that finetuning Video-LLMs with self-supervised fragment tasks, by \"fragment\" we mean temporal frame level specifications of the video, could improve the model's sensitivity to spatiotemporal scene-level details (related to video contents). Driven by this, we introduce the Self-supervised Fragment Fine-Tuning $(\\mathrm{SF}^2\\mathrm{T})$ , a effortless fine-tuning strategy for Video-LLMs that help to improve the fine-grained video understanding. $\\mathrm{SF}^2\\mathrm{T}$ consists of five fragment-level tasks—Counting, Consistency Verification, Localization, Disorder Detection and Rearrangement—that automatically generate labels from various spatiotemporal perspectives. This approach maximizes the use of frame-level information while minimizing reliance on complex human instructions and annotations.", + "bbox": [ + 89, + 316, + 482, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, to evaluate the fine-grained visual dynamic perception of Video-LLMs and fully demonstrate the effectiveness of our $\\mathrm{SF}^2\\mathrm{T}$ , we present the FineVidBench, a novel benchmark. FineVidBench comprises 910 videos and 22,718 question-answer pairs, with videos sourced from diverse public datasets, including Something-Something V2 (SSv2) [6], Moments in Time (MiT) [21], etc. The question-answer pairs are auto-generated in single-choice format, incorporating distractors to increase testing difficulty. We evaluated several notable Video-LLMs developed in recent years, and find they generally fail to understand the execution sequence of actions and struggling to grasp fine-grained spatiotemporal information. While after fine-tuning with $\\mathrm{SF}^2\\mathrm{T}$ , the Video-LLMs better recognize spatiotemporal details, leading to a holistic and marked improvement in fine-grained understanding.", + "bbox": [ + 89, + 544, + 482, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 799, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video-LLMs Finetuning Video-LLMs are primarily finetuned by adjusting the parameters of small, trainable adapters for task adaptation, without changing the entire model, saving resources and enhancing efficiency. The connective adapter (e.g., MLP/Linear Layer [15], Q", + "bbox": [ + 89, + 824, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "former [10]) links the Video Embedder and LLM, aligning video embeddings with LLM input tokens, while insertive adapters (e.g., LoRA [8]) are directly integrated into the LLM to modify its behavior. Most Video-LLMs combine both types of adapters and typically use multi-stage finetuning [4, 11, 13, 24, 35]. First, the model learns to establish relationships between images, videos, and text using large-scale multimodal datasets [1, 2, 29, 31]. In the second stage, the model is fine-tuned with an curated instruction-following dataset [11, 17, 18]. Besides, there are full finetuning, which updates all LLM parameters with a lower learning rate [25, 33], and zero-shot models, which transforms the video task into a text task, typically relying on a powerful LLM [32]. However, annotating video data remains a labor-intensive and time-consuming task, particularly for long videos or those involving complex actions.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benchmarks on Video-LLMs Currently, many studies [3, 5, 38] focus on evaluating the temporal perception capabilities of Video-LLMs. MVBench [12] designs 20 tasks from temporal and spatial perspectives, and Tempcompass [16] introduces 5 temporal aspects and 4 task formats. VN-Bench [36] decouples video content from the QA pairs by inserting irrelevant images or text \"needles\" into the original video. Moment-10M [22] has constructed a large-scale dataset on temporal localization tasks. However, as illustrated in Table 1, these studies often focus on gathering diverse videos or evaluating the models' performance with long videos, while somewhat neglecting the models' ability to perform fine-grained perception of temporal details. To address this gap, FineVidBench breaks videos into multiple sets of frames and generates annotations from diverse spatiotemporal perspectives, introducing novel evaluation methods for fine-grained understanding.", + "bbox": [ + 511, + 333, + 903, + 589 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/8fc059292041f5e55d6b484db9afa9eca18e426fc41d85f00a5152e809657be8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarksVideo num.QA num.Input ChangeTemporal DiversityFine-Grained EvaluationHierarchical Test
Video-MME9002700XXXX
TempCompass4107540X✓✓X
VN bench-1350X✓✓X
Moment-10M64.9k10.4MXXXX
AutoEval-Video327327XXXX
MV bench36414000XX✓X
MLVU13342593XXXX
FineVidBench91022,718✓✓✓✓
", + "bbox": [ + 516, + 599, + 903, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison with related benchmarks. Our approach offers significant advantages in input formats, evaluation methods, granularity, and temporal diversity.", + "bbox": [ + 511, + 762, + 903, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. FineVidBench Benchmark", + "text_level": 1, + "bbox": [ + 511, + 829, + 759, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is broadly recognized that Video-LLMs struggle with fine-grained video understanding tasks, yet no comprehensive benchmarks exist to thoroughly investigate this issue.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29109", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this gap, we introduce FineVidBench, a multidimensional, fine-grained evaluation framework specifically designed to assess and improve the overall capabilities of Video-LLMs.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Construction", + "text_level": 1, + "bbox": [ + 89, + 162, + 230, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data collection We selected videos from various public datasets, including SS-v2 [6], MiT [21], and Ego4D [7], with a particular emphasis on temporally-sensitive content, to focus the model on the entire video sequence rather than individual frames.", + "bbox": [ + 89, + 185, + 482, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action categorization As shown in Figure 2, we compiled 52 actions, categorizing them into 3 types based on intraclass variance. The distribution varies significantly: \"Distinctive Actions\" $(39\\%)$ are easily recognizable, encompassing a total of 36 actions. \"Non-typical Actions\" $(57\\%)$ refer to flexible actions with no clear defining characteristics, spanning 14 types. The broad diversity and complexity in this category require more extensive video coverage to adequately capture the range of expressions and variations. \"Slight Movements\" $(4\\%)$ represent subtle actions, such as \"hold\" and \"show\", which are difficult to detect with the naked eye and constitute a small proportion.", + "bbox": [ + 89, + 261, + 482, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data augmentation The original videos were augmented using frame interpolation and skipping techniques for speed transformation, along with a motion-salient area sampling algorithm to capture dynamic motion. This process generated speed-varied versions and multiple sets of keyframes for each video.", + "bbox": [ + 89, + 443, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Statistics With our augmentation strategy, FineVidBench includes 910 videos, 1,820 speed-variant videos, and 2,670 sets of keyframes enriched with dynamic visual information. Building on this, we generated 22,718 QA pairs from the video content through a combination of automated processes and manual review. The quality assurance process involved rigorous cross-verification, where reviewers checked each QA pair for accuracy and contextual relevance, making corrections to ensure high quality.", + "bbox": [ + 89, + 534, + 482, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Benchmarking Dimensions", + "text_level": 1, + "bbox": [ + 89, + 681, + 334, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 3, FineVidBench encompasses both scene-level and fragment-level evaluations. The scene-level evaluation assesses both original and speed-adjusted videos across three dimensions: (1) Action, which evaluates the model's holistic understanding of video content. To increase difficulty, \"Visual Synonyms\" are added as distractors, requiring VideoLLM to distinguish visually similar actions with subtle differences, a challenge common in real-world scenarios. (2) Effect, which focuses on the model's comprehension of the visual changes resulting from actions. This understanding is essential for revealing object properties and interpreting complex dynamic scenes, and could significantly enhance the reasoning capabilities of Video", + "bbox": [ + 89, + 704, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8e219e19edad16fd218dca29a3cb48b7d80101a471ef88867ad2f65e384007a1.jpg", + "image_caption": [ + "Figure 2. We show the action semantics and their respective proportions in FineVidBench. Distinctive Action: easily recognizable actions. Non-typical Action: flexible actions with no clear characteristics, like \"put\" and \"move.\" Slight Movement: subtle actions, such as \"hold\" and \"show,\" difficult to detect with the naked eye." + ], + "image_footnote": [], + "bbox": [ + 547, + 94, + 875, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs and LLM-aided agents. (3) Speed, which tests the model's sensitivity to changes in video speed and its capability to maintain consistent understanding across varying speeds, with slow motion revealing hidden details and fast motion obscuring them. This capability is crucial for optimizing the model's performance across diverse scenarios.", + "bbox": [ + 511, + 462, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For fragment-level evaluation, We've designed a structured evaluation format for video dynamic keyframes, employing a step-by-step inquiry framework: (1) Frame Count: Models are queried on the number of frames in sequences using dynamically refined keyframes to assess counting accuracy. (2) Meaning of Order: Understanding of sequence order is tested by asking about the first or last frames the targets appear in, or the frames they are present. e.g., \"At which frame does the target object first appear?\". (3) Frame Comparison: Two frames are randomly selected from the sequence for visual comparison, with differences varying in size but generally staying within human visual comfort limits. (4) Adjust-or-Not and Rearrangement: These two tasks involve a shuffled sequence of keyframes, and the model is asked to determine whether the order needs adjustment and, if so, how to correct it. They evaluate the model's ability to understand and restore the video's temporal sequence.", + "bbox": [ + 511, + 553, + 908, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Benchmark Results", + "text_level": 1, + "bbox": [ + 513, + 833, + 702, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluated six of the most advanced open-source models: LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4], Qwen2-VL[28], ShareGPT4Video [2] and", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29110", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f2c27bd209bc48a9f381966094271a4758a614629c12b08a213350398427d324.jpg", + "image_caption": [ + "Figure 3. FineVidBench evaluates videos augmented with speed variations and fragments. Scene-level tests include the following: Action: Tests recognition accuracy amidst distractors like \"Visual Synonyms\". Effect: Assesses the model's ability to identify pre- and post-action changes. Speed: Measures the model's sensitivity to changes in video speed. Fragment-level tests, employing a step-by-step inquiry framework, focus on challenges such as Frame Count, Meaning of Order, Frame Comparison, Adjust-or-Not and Rearrangement." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 426, + 303 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9726bd792966151b9e71d0c4fb9e8145f434a5d37572a9bba433c9ff69e6d8e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 88, + 658, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/56d0a2df8e3d8c352b70564101947f32b8297d02ddf5616cb21cf61904f974f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 112, + 885, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video-CCAM [27], each employing different architectures and training strategies. Table 3 summarizes the results across the eight tasks. We discuss the results from scene-level and fragment-level.", + "bbox": [ + 89, + 396, + 482, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Scene-level Results and Analysis", + "text_level": 1, + "bbox": [ + 89, + 469, + 333, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action The scores for this task varied significantly, with models trained in relevant video data—such as Video-CCAM, Qwen2-VL, and VideoLLaMA 2.1—achieving notably higher performance. However, as shown on the left side of Table 2, interference from \"Visual Synonyms\" prevented these models from achieving their full potential, resulting in declines of varying degrees and indicating difficulties in distinguishing visually similar actions.", + "bbox": [ + 89, + 489, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Effect All models exhibited average performance on this task, indicating a superficial understanding of aspects such as object attributes, object relationships, and action properties. This task tests the model's ability to grasp how actions affect objects, focusing on causal relationships and temporal reasoning—particularly for actions like \"push\" and \"pull\", which share similar execution flows. The model must distinguish them based on dynamic effects, such as changes in direction and speed, but most models perform moderately in this regard.", + "bbox": [ + 89, + 612, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Speed The results show that all models are insensitive to speed variations, likely because they were not adequately exposed to speed changes during training. Figure 4 shows that models are more sensitive to slow motion than fast playback, and struggled with identifying \"normal speed\" and \"no speed\", except for VideoLLaMA 2.1. This may be due to the loss of coherence in fast-moving video content, while slow-motion videos highlight more distinct details, aiding the model in making accurate judgments.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/000c48ff5f7a82a3b563edc36bee7663f09f35fa0b71d19da0b0c003be44159c.jpg", + "image_caption": [ + "Figure 4. Accuracy across different video speeds. All models are more sensitive to slow-speed videos and struggle to understand \"normal speed\" and \"no speed\", except for VideoLLaMA 2.1." + ], + "image_footnote": [], + "bbox": [ + 545, + 396, + 875, + 583 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8b9fbb96f8d43c7c9f39a2aae7856e3c87004d890d6039e0136f60e1d2523d08.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Video-LLMsActionFrame Number
w/o VSw/ VSAvg.345
LLaVA-NeXT-Video37.3135.0419.3720.3319.7717.98
MiniCPM-V 2.643.3740.1590.3293.8290.6686.44
Video-LLaMA 2.163.2653.9830.1742.8639.897.45
Qwen2-VL68.1856.6296.6597.2596.6396.05
ShareGPT4Video46.9030.8426.3360.9916.780.00
Video-CCAM73.1060.2323.4514.188.9647.61
", + "bbox": [ + 514, + 646, + 916, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Left: Accuracy of the Action task with or without \"Visual Synonyms\". It is obvious that the \"Visual Synonyms\" have significantly impacted the model's judgment. Right: Accuracy of the counting task across different frame counts. Except for Video-CCAM, all other models exhibited a decline in performance as the number of frames increased.", + "bbox": [ + 511, + 811, + 906, + 893 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29111", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a26750f6f0df52394cea209edd778f9d22347560d0ea6525b5d50ed1e0b86fbb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Video-LLMsParams.Scene-LevelFragment-LevelS-Avg.FG-Avg.A-Avg.
ActionEffectSpeedFCntMoOFCmpAoNRearr
(Random)-25.0025.0025.0025.0025.0033.3333.3325.0025.0028.3327.08
LLaVA-NeXT-Video7B37.3142.6722.3519.3724.0253.7575.4520.6734.1138.6536.95
MiniCPM-V 2.68B43.3752.5619.1390.3256.4275.6676.4918.0938.3563.4054.01
Video-LLaMA 2.17B63.2650.9219.8930.1742.2776.0189.9226.8744.6953.0549.91
Qwen2-VL7B68.1857.1424.6296.6533.3374.5390.7022.4849.9863.5458.45
ShareGPT4Video8B46.9043.8831.7626.3361.0588.4484.8023.3640.8557.1150.82
Video-CCAM9B73.1055.9031.6523.4545.6664.9590.2722.7253.5548.4750.96
", + "bbox": [ + 106, + 88, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. The overall performances of notable Video-LLMs on FineVidBench. FCnt: Frame Count. MoO: Meaning of Order. Fcmp: Frame Comparison. AoN: Adjust or Not. Rearr: Rearrangement. S-Avg.: the average performance of scene-level tasks; FG-Avg.: the average performance of fragment-level tasks. A-Avg.: the average performance of all tasks.", + "bbox": [ + 89, + 313, + 906, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Fragment-level Results and Analysis", + "text_level": 1, + "bbox": [ + 89, + 382, + 362, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(1) Frame-count accuracy varied significantly across models, with the lower-performing models likely lacking targeted training. The trend shown in the right side of Table 2, where accuracy decreases as frame count increases, highlights the models' insufficient temporal reasoning on longer sequences. (2) ShareGPT4Video and MiniCPM-V 2.6 showed better comprehension in the Meaning-of-Order task, while other models lagged, suggesting a lack of explicit focus on \"order\". (3) Most models excelled in frame comparison due to image-text alignment training. ShareGPT4Video achieved the best performance, owing to its Differential Sliding-Window Captioning (DiffSW) strategy, which emphasizes capturing the changes between frames when generating video descriptions. This also improved its Meaning-of-Order performance. (4) In the sorting task, models generally succeeded in the \"Adjust or Not\" response but performed poorly in the more complex \"Rearrangement\" task, indicating they can detect, but not correct, sequence errors.", + "bbox": [ + 88, + 402, + 485, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Self-supervised Fragment Finetuning", + "text_level": 1, + "bbox": [ + 89, + 707, + 426, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above benchmark results show the existing Video-LLMs generally fail to tackle fine-grained video understanding tasks. Videos often contain subtle, complex changes that natural language alone fails to fully capture. The core component of Video-LLMs, LLMs, as generalized pattern recognizers, offers a promising solution. LLMs have the potential to detect and interpret intricate spatiotemporal dynamics that were previously difficult to represent. Given that these changes cannot be directly annotated, using self-supervised learning naturally becomes the solution, bypassing the bottleneck of manual annotation and significantly re", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ducing labeling costs. Given these factors, we propose the $\\mathrm{SF^2T}$ to fine-tune Video-LLMs. While we do not expect $\\mathrm{SF^2T}$ to replace the supervised fine-tuning, instead it's an effortless complementary to SFT. Comparing $\\mathrm{SF^2T}$ with SFT, they primarily differ in data construction and content focus level, with each method aligned with distinct training objectives as shown in Figure 5.", + "bbox": [ + 511, + 382, + 906, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. SFT Tasks", + "text_level": 1, + "bbox": [ + 511, + 498, + 632, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first review the common SFT tasks to set a baseline for comparing our $\\mathrm{SF^2T}$", + "bbox": [ + 511, + 520, + 906, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "General QA on Video Content This method focuses on understanding the main events and context of a video by directly asking questions about its content. While effective for grasping the video's key moments, it lacks finer spatiotemporal details and requires significant human effort to create standardized but constrained answers.", + "bbox": [ + 511, + 551, + 905, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frame Description Integration This method typically samples video frames evenly, generates detailed descriptions for each, and integrates them into a cohesive but lengthy summary. While it enhances the model's understanding of continuity and micro-dynamics, it often proves incapable of capturing complex or subtle details that are beyond natural language's scope. Moreover, although frame descriptions can be generated using powerful multi-model LLMs like GPT-4o, significant human effort is still required to review the quality of the generated responses.", + "bbox": [ + 511, + 642, + 908, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Fragment-level Tasks of $\\mathbf{SF}^2\\mathbf{T}$", + "text_level": 1, + "bbox": [ + 511, + 801, + 779, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SFT tasks require manual annotations, and even automation annotation is labor-intensive and error-prone. To address, we introduce $\\mathrm{SF}^2\\mathrm{T}$ which generates accurate fragment-level labels accurately. $\\mathrm{SF}^2\\mathrm{T}$ comprises five tasks—Counting, Consistency Verification, Localization, Disorder Detection", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29112", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/205a408f47737ab3e7ada0948b109312f60fa3951eebb56720eff570ee330e62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 93, + 477, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What is the main content of the video?", + "bbox": [ + 120, + 161, + 383, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video shows a person bowling, including their four-step approach, the smooth release of the ball down the lane, its path toward the pins, and...", + "bbox": [ + 147, + 180, + 423, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d73b992c7705e58db518f61c8e87a993d684148c83ad06bf6ee75bc240b2dd85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 229, + 475, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "What is the main content of the video?", + "bbox": [ + 120, + 305, + 385, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video shows a person bowling: (Frame 1) The scene shows a bowling alley... (Frame 2) The player swing the bowling ball... (Frame 4) The bowling ball approaches the pins... (Frame 6) The bowling ball strikes the pins... (Frame 8) All the pins are down.", + "bbox": [ + 148, + 324, + 424, + 377 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67c18e2e5d3b5ddf1ade5d3208662e77bc4273c7f4af0846b550bfcea83f9e48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 319, + 447, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2809af5779cd4055ae1b311000647b93fc824f660e342bc23cbfbcf66bbcf07.jpg", + "image_caption": [ + "Scene-Level Tasks" + ], + "image_footnote": [], + "bbox": [ + 125, + 419, + 444, + 492 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2db6c38cc0fd2073da14990dcbf1438642ff33fea8f737419c9f49191f9fe9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 505, + 143, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How many frames?", + "bbox": [ + 151, + 510, + 267, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On which frames?", + "bbox": [ + 150, + 530, + 263, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Same frames?", + "bbox": [ + 150, + 549, + 241, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adjust or not?", + "bbox": [ + 150, + 580, + 236, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rearrange it.", + "bbox": [ + 150, + 599, + 230, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80f3bdf9d832b46ec11dfa8a148226f2604d3174b3dd3e36d58a372ea997442d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 507, + 370, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/191add99094c9cf437d91ba98b8f1d1c1e3d0af2e8b4bb85365ec85bac2ba6ce.jpg", + "image_caption": [ + "Figure 5. Comparison between $\\mathrm{SF}^2\\mathrm{T}$ and SFT. SFT depends on manual and model-driven design to generate QA pairs for scene-level video understanding, $\\mathrm{SF}^2\\mathrm{T}$ , in contrast, automatically constructs training data based on pre-defined rules that cover various temporal and spatial aspects of the video. $\\mathrm{SF}^2\\mathrm{T}$ enables the model to focus on a fine-grained content analysis, and offering insights that supervised labels cannot achieve." + ], + "image_footnote": [], + "bbox": [ + 287, + 529, + 366, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b29d67543db69e81306b3d01d8a56e2955a4bfc3bfa673551b6bfd0bcdb6a23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 547, + 366, + 560 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a37214fdf9867e16e0bccbe439bebc013a75c6b18bd07cb3822a9ba0d8ff0dd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 568, + 366, + 595 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7895528dc17bddd0cdfd12a2b6f192d8f17af7af0313cc10e30ea23288f819c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 601, + 366, + 618 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/139eef899517a8522b5168019c8b5d3c8b9299d5d7c4b909bfa4288d42b9e477.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 508, + 447, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2nd", + "bbox": [ + 395, + 530, + 419, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "No", + "bbox": [ + 400, + 550, + 419, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Yes", + "bbox": [ + 388, + 580, + 419, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3412", + "bbox": [ + 387, + 599, + 419, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fragment-Level Tasks", + "text_level": 1, + "bbox": [ + 200, + 638, + 372, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and Rearrangement—designed to train the model to rearrange a set of out-of-order frames into their original sequence. This is a robust indicator of a modal's mastery over the visual dynamics of an action, requiring the model to detect subtle frame changes and understand the overall coherence and temporal trends. Mastery of these tasks enables the model to recognize frames and their temporal re", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "relationships, enhancing its ability to predict and reconstruct action sequences and improving performance on more complex video tasks. Our method first extracts multiple sets of dynamic keyframes from each video. These fragments capture the key dynamic information from multiple temporal perspectives, offering a more efficient representation of redundant video data. It then applies pseudo-labeling, distinguishing it from traditional video-level labeling. By designing proxy tasks that leverage intrinsic information rather than predefined prior knowledge, it smartly circumvents the annotation bottleneck, enabling a deeper temporal understanding and offering insights that traditional video-level labeling cannot achieve.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Counting We input N frames into the Video-LLM and ask it to count them. Although this task seems straightforward, it proves challenging for current Video-LLMs, particularly as the number of frames increases, revealing a decline in accuracy. The model's inability to perform basic quantitative tasks points to a broader limitations in understanding the overall sequence integrity.", + "bbox": [ + 511, + 289, + 903, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistency Verification Video-LLMs are tasked with identifying two frames sampled from the same video, which may show subtle differences. This task sharpens the model's sensitivity to visual details by encouraging a thorough analysis and comparison of the images, countering its tendency to focus on primary subjects while neglecting the background and other subtle features.", + "bbox": [ + 511, + 396, + 903, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Localization Video-LLMs must accurately locate a specified target (from video metadata) within a sequence of frames, identifying the frames in which it appears, disappears, or persists. This naturally human ability is a significant challenge for these models, as they often struggle to perceive sequential relationships between frames and face additional obstacles, such as occlusion, interference from similar objects, lighting variations, and memory limitations.", + "bbox": [ + 511, + 503, + 905, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Disorder Detection and Rearrangement Video-LLMs must determine whether and how to adjust the order of a given frame sequence. When frames are randomized, the loss of spatiotemporal coherence and logical continuity makes it exceptionally challenging to reconstruct their original sequence, especially as interactions within frames become more complex [20]. This task is evaluated in two ways: the yes/no task tests the model's sensitivity to temporal consistency, while the sorting task, which leverages capabilities from the other four tasks, requires advanced reasoning and adjustments.", + "bbox": [ + 511, + 627, + 905, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 813, + 643, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we fine-tuned four of the most advanced open-source Video-LLMs using the $\\mathrm{SF}^2\\mathrm{T}$ method to evaluate its effectiveness, alongside ablation studies and interpretability analyses to explore the underlying mechanisms.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29113", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78f02b4b1443fcbd31810d2cb4c2be8b80bd146e6b513946e6cd4650dd9034d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLLaVA-NEXT-VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2-VL
ActionEffectSpeedActionEffectSpeedActionEffectSpeedActionEffectSpeed
Base37.3142.6722.3543.3752.5619.1363.2650.9219.8968.1857.1424.62
Base+SF2T48.6743.7724.8365.9160.6228.6067.4257.3331.6373.8663.3731.92
Base(SFT)62.6944.6322.3577.6575.0970.8377.6565.9429.7378.6066.3030.87
Base(SFT)+SF2T63.0745.2432.0181.6376.9286.7479.7368.6831.8281.2573.2632.38
", + "bbox": [ + 133, + 88, + 867, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43d552bf5399c00a0569cb06ec2e9d002d3107cd4d82b287dea20b52b8ce81dc.jpg", + "table_caption": [ + "Table 4. Performance on FineVidBench. We tested on two baselines: (1) Base: Results without any fine-tuning. (2) Base(SFT): Results after fine-tuning in supervised way. After $\\mathrm{SF}^2\\mathrm{T}$ , all models improved in all three tasks, highlighting its broad effectiveness and the value of fragment-level tasks in enhancing scene-level comprehension. Notably, $\\mathrm{SF}^2\\mathrm{T}$ outperformed SFT in the Speed task (except MiniCPM-V 2.6), highlighting the key role of fine-grained temporal understanding in distinguishing video speeds." + ], + "table_footnote": [], + "table_body": "
MethodsLLaVA-NeXT -VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2 -VL
MVBench
Base36.8440.2354.1855.97
Base+SF2T42.9256.0257.9763.76
Video-MME(no subtitle)
Base29.7643.1749.0243.77
Base+SF2T34.8453.1951.8853.60
MLVU
Base36.3241.5852.3242.81
Base+SF2T41.9155.3256.1154.67
", + "bbox": [ + 102, + 318, + 473, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 584, + 307, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To ensure fairness, experiments were conducted on LoRA-compatible models, including LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4] and Qwen2-VL[28], using their default or recommended settings, with all models trained for one epoch. All experiments were performed under identical hardware conditions, utilizing NVIDIA A100 40GB GPU for computation. It should be emphasized that our goal is to validate the effectiveness of $\\mathrm{SF}^2\\mathrm{T}$ , not to optimize models for maximum performance.", + "bbox": [ + 89, + 609, + 483, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We randomly sampled videos from SSv2 and MiT for training, ensuring no overlap with the FineVidBench dataset. MGSampler [37] was used to extract N sets of M-frame sequences from each video, capturing dynamic changes while preserving overall characteristics. M is chosen based on the video's characteristics to capture content flow, while N is determined by content complexity, with more complex content requiring a larger N to cover more temporal perspectives. In this study, we set $\\mathrm{N} = 3$ and M between 3 and 5, though these values may vary for other", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da236d8ee06ac6dc6f73c71bbc3751866fd6aa2481f436ae04e6af2b84ce1c1f.jpg", + "table_caption": [ + "Table 5. Performance on public benchmarks. $\\mathrm{SF}^2\\mathrm{T}$ consistently enhances performance across all three benchmarks, reaffirming its effectiveness as a spatiotemporal enhancer." + ], + "table_footnote": [], + "table_body": "
Methodsrandomuniformkeyframemotion-salient
SF2T70.3171.6772.1173.86
", + "bbox": [ + 540, + 323, + 877, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c46016b2b6ceb85609d527febf51144299c88c854a5f41617de9296fff7c372d.jpg", + "table_caption": [ + "Table 6. Impact of sampling. As shown, motion-salient area sampling outperforms others by better capturing motion fluidity and temporal details, while the other methods fail to fully utilize their potential, leading to suboptimal performance." + ], + "table_footnote": [], + "table_body": "
Methodslongshortrandom
SF2T69.3871.4073.86
", + "bbox": [ + 604, + 449, + 815, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Impact of temporal span. Both long- and short-range temporal modeling reduced $\\mathrm{SF}^2\\mathrm{T}$ 's performance, emphasizing the importance of multi-scale temporal modeling.", + "bbox": [ + 509, + 512, + 906, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "datasets. We then generated QA pairs for each frame sequence based on the five tasks defined in $\\mathrm{SF^2T}$ for training. Evaluations were performed on FineVidBench's scene-level tasks, including Action, Effect, and Speed. To compare with traditional SFT, we also generated and manually reviewed QA pairs for these videos in a supervised setting.", + "bbox": [ + 511, + 584, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 694, + 653, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 summarizes the results of the scene-level tasks. After $\\mathrm{SF}^2\\mathrm{T}$ training, all models showed significant improvement, emphasizing that fragment-level tasks can notably enhance scene-level comprehension. Integrating $\\mathrm{SF}^2\\mathrm{T}$ with SFT is also leads to performance gains, demonstrating that fragment-level training positively impacts SFT and enhances its effectiveness. Surprisingly, in the Speed task, many base models outperformed SFT after applying $\\mathrm{SF}^2\\mathrm{T}$ , highlighting the importance of fine-grained temporal understanding in distinguishing video speeds. This improvement likely stems from $\\mathrm{SF}^2\\mathrm{T}$ 's ability to enhance the model's sensitivity to temporal cues, such as the loss or enhancement of", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29114", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "information during acceleration or deceleration, as well as content coherence—all crucial for speed judgment. As expected, $\\mathrm{SF}^2\\mathrm{T}$ currently lags behind SFT, since its training objective is not fully aligned with scene-level tasks. However, we do not expect $\\mathrm{SF}^2\\mathrm{T}$ to replace supervised fine-tuning; rather, our experiments suggest that it can serve as an effortless and effective complement to SFT.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to FineVidBench, we evaluated $\\mathrm{SF}^{2}\\mathrm{T}$ on three public video understanding benchmarks (Table 5). The results demonstrate consistent improvements across various video tasks, validating $\\mathrm{SF}^{2}\\mathrm{T}$ as an effective spatiotemporal enhancer for a wide range of video understanding tasks. All models were tested with an 8-frame input.", + "bbox": [ + 89, + 198, + 480, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation and Interpretability Analyses", + "text_level": 1, + "bbox": [ + 89, + 301, + 421, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluated the impact of frame sampling strategies on $\\mathrm{SF}^2\\mathrm{T}$ , as each method provides a unique \"temporal information perspective\" that influencing video understanding performance. As shown in Table 6, we assessed four strategies on Qwen2-VL in the Action task: random, uniform interval, keyframe, and motion-salient area sampling [37]. Motion-salient area sampling performed best, likely due to its ability to capture continuous motion dynamics, thereby enhancing the model's understanding of action fluidity and temporal detail. In comparison, the other methods had limitations: keyframe sampling misses intermediate action phases, fixed-interval sampling may overlook critical moments, and random sampling lacks temporal consistency. Notably, different datasets may favor different strategies. For example, some datasets may perform better with uniform interval sampling, or their motion features may align better with the model's specific capabilities.", + "bbox": [ + 89, + 324, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examined the effects of long- and short-range temporal modeling on $\\mathrm{SF}^2\\mathrm{T}$ . In the Consistency Verification task, we constrained the random selection of frame pairs to adjacent frames for local continuity or non-adjacent frames to capture long-range dependencies. As shown in Table 7, both settings decreased $\\mathrm{SF}^2\\mathrm{T}$ 's performance on the Action task of Qwen2-VL, indicating that an overemphasis on either long- or short-range information leads to temporal imbalance and incomplete dynamics. This underscores the importance of combining both approaches to leverage their broader temporal span and frame variations for a more comprehensive feature representation.", + "bbox": [ + 89, + 583, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyzed the attention map of Qwen2-VL on the Action task, particularly in cases where the model's predictions were corrected after $\\mathrm{SF}^2\\mathrm{T}$ . As shown in Figure 6, we found that $\\mathrm{SF}^2\\mathrm{T}$ enhances the model's ability to capture fine-grained spatial changes and temporal dynamics. (1) Spatial Aspects. After $\\mathrm{SF}^2\\mathrm{T}$ , the model shows increased attention to action execution areas, particularly the hands and objects they interact with. It shows better sensitivity to small targets, likely due to the Consistency Verification", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9553a2acfb87e33c118766225e4eb5707c664e1fa8e5864d9f59bea5dffc0cda.jpg", + "image_caption": [ + "Figure 6. Two exemplary visualizations of the attention map on Qwen2-VL. For each example: top - Original frames; middle - Base (SFT); bottom - $\\mathrm{SF^2T}$ applied. As shown by the red boxes, after applying $\\mathrm{SF^2T}$ , the model better focuses on action execution areas and interacting objects. The $\\mathrm{SF^2T}$ fine-tuned model has the ability to predict the direction of motion, as seen in the trajectories of the red bottle and Cheerios." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 883, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "task, which enhances spatial perception by refining sensitivity to subtle image differences. (2) Temporal Aspects. After $\\mathrm{SF}^2\\mathrm{T}$ , we observed that the model can predict object movement trajectories in certain actions, indicating an advanced level of temporal understanding. This ability likely stems from the sorting task, which strengthens the model's comprehension of action flows and movement patterns.", + "bbox": [ + 511, + 595, + 906, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 733, + 633, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose $\\mathrm{SF}^2\\mathrm{T}$ to overcome the limitations of Video-LLMs in fine-grained video understanding. $\\mathrm{SF}^2\\mathrm{T}$ is an innovative fine-tuning method that eliminates the need for labor-intensive annotations and effectively bypasses the constraints of natural language descriptions. Additionally, we introduce FineVidBench, a benchmark for evaluating Video-LLMs at both scene and fragment levels. In the future, we plan to expand our dataset with larger videos and more tasks to increase its impact.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29115", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Program of China (No.2020YBF2901202), National Natural Science Foundation of China (NSFC No. 62272184 and No. 62402189), the China Postdoctoral Science Foundation under Grant Number GZC20230894, the China Postdoctoral Science Foundation (Certificate Number: 2024M751012), and the Postdoctor Project of Hubei Province under Grant Number 2024HBBHCXB014, and the \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (No. 2024C01161). The computation is completed in the HPC Platform of Huazhong University of Science and Technology.", + "bbox": [ + 89, + 114, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 310, + 187, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2", + "[2] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 3", + "[3] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 2", + "[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 2, 3, 7", + "[5] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2", + "[6] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3", + "[7] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3", + "[8] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[9] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave:" + ], + "bbox": [ + 99, + 335, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3, 7", + "[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2", + "[11] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2", + "[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2", + "[13] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2", + "[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1", + "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[16] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 2", + "[17] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2", + "[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2", + "[19] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1", + "[20] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 527-544. Springer, 2016. 6", + "[21] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019. 2, 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29116", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2", + "[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1", + "[24] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2", + "[25] Fangxun Shu, Lei Zhang, Hao Jiang, and Cihang Xie. Audio-visual llm for video understanding. arXiv preprint arXiv:2312.06720, 2023. 2", + "[26] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 1", + "[27] TencentQQ Multimedia Research Team. Video-cam: Advancing video-language understanding with causal cross-attention masks. https://github.com/QQ-MM/Video-CCAM, 2024.4", + "[28] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 7", + "[29] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2", + "[30] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1", + "[31] Haiyang Xu, Qinghao Ye, Xuan Wu, Ming Yan, Yuan Miao, Jiabo Ye, Guohai Xu, Anwen Hu, Yaya Shi, Guangwei Xu, et al. Youku-mplug: A 10 million large-scale chinese video-language dataset for pre-training and benchmarks. arXiv preprint arXiv:2306.04362, 2023. 2", + "[32] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 2", + "[33] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 3, 7", + "[35] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2", + "[36] Zijia Zhao, Haoyu Lu, Yuqi Huo, Yifan Du, Tongtian Yue, Longteng Guo, Bingning Wang, Weipeng Chen, and Jing Liu. Needle in a video haystack: A scalable synthetic framework for benchmarking video mllms. arXiv preprint arXiv:2406.09367, 2024. 2", + "[37] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 7, 8", + "[38] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2", + "[39] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2" + ], + "bbox": [ + 516, + 92, + 903, + 458 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29117", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_model.json b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d1dee02d52794ab0941f1a6569593b077e43c3b5 --- /dev/null +++ b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.888, + 0.177 + ], + "angle": 0, + "content": "\\(\\mathbf{SF}^{2}\\mathbf{T}\\): Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.893, + 0.244 + ], + "angle": 0, + "content": "Yangliu Hu\\(^{1}\\), Zikai Song\\(^{1\\dagger}\\), Na Feng\\(^{1}\\), Yawei Luo\\(^{2}\\), Junqing Yu\\(^{1}\\), Yi-Ping Phoebe Chen\\(^{3}\\), Wei Yang\\(^{1\\dagger}\\) \n\\(^{1}\\)Huazhong University of Science and Technology \\(^{2}\\)Zhejiang University \\(^{3}\\)La Trobe University" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.246, + 0.747, + 0.261 + ], + "angle": 0, + "content": "{huyangliu,skyesong,fengna,yjqing,weiyangcs}@hust.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.264, + 0.708, + 0.278 + ], + "angle": 0, + "content": "yaweiluo@zju.edu.cn phoebe.chen@latrobe.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.313, + 0.327, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.486, + 0.737 + ], + "angle": 0, + "content": "Video-based Large Language Models (Video-LLMs) have witnessed substantial advancements in recent years, propelled by the advancement in multi-modal LLMs. Although these models have demonstrated proficiency in providing the overall description of videos, they struggle with fine-grained understanding, particularly in aspects such as visual dynamics and video details inquiries. To tackle these shortcomings, we find that fine-tuning Video-LLMs on self-supervised fragment tasks, greatly improve their fine-grained video understanding abilities. Hence we propose two key contributions: (1) Self-Supervised Fragment Fine-Tuning \\((SF^2 T)\\), a novel effortless fine-tuning method, employs the rich inherent characteristics of videos for training, while unlocking more fine-grained understanding ability of Video-LLMs. Moreover, it relieves researchers from labor-intensive annotations and smartly circumvents the limitations of natural language, which often fails to capture the complex spatiotemporal variations in videos; (2) A novel benchmark dataset, namely FineVidBench, for rigorously assessing Video-LLMs' performance at both the scene and fragment levels, offering a comprehensive evaluation of their capabilities. We assessed multiple models and validated the effectiveness of \\(SF^2 T\\) on them. Experimental results reveal that our approach improves their ability to capture and interpret spatiotemporal details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.222, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have showcased significant emergent capabilities, such as in-context learning [19], instruction-following [23], and chain-of-thought reasoning [30], driven by expansive datasets and advanced model architectures. Extending these advancements, Video-LLMs through mechanisms like pooling or query aggregation" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.313, + 0.885, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.553, + 0.908, + 0.651 + ], + "angle": 0, + "content": "Figure 1. Performance w/ and w/o \\(\\mathbf{SF}^2\\mathbf{T}\\). We evaluated four advanced Video-LLMs w/ and w/o \\(\\mathrm{SF}^2\\mathrm{T}\\) on our proposed FineVidBench with two baselines: (1) Base: performance without any fine-tuning (blue dashed), and (2) Base (SFT): performance with supervised fine-tuning (red dashed). After applying \\(\\mathrm{SF}^2\\mathrm{T}\\), all models showed significant improvements (solid blue and red), underscoring its broad effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.791 + ], + "angle": 0, + "content": "across numerous visual tokens, have broadened the scope of LLMs to encompass video information processing [11, 14, 35]. This evolution markedly advances their potential for in-depth real-world comprehension, opening applications in intelligent surveillance, virtual reality, and autonomous driving, further enriching the landscape of video analytics and interpretation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Various Video-LLMs, exemplified by GPT4-V, VideoLaMA 2 [4], MiniCPM-V [34], and Qwen2-VL [28], have been crafted by leading corporations and research institutions, demonstrating proficiency in capturing the overarching content of videos. When adapting to new videos and tasks, they predominantly rely on Supervised FineTuning (SFT) [26] or Reinforcement Learning from Hu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.121, + 0.888, + 0.255, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29108" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "man Feedback (RLHF) [39], both of which are heavily contingent upon extensive manual annotation. This dependence poses several key problems: (1) it necessitates substantial human resources, particularly highly trained annotators; (2) the inherent complexity of video content and task demands frequently introduces inconsistencies and subjectivity, rendering the maintenance of high-quality annotations particularly arduous; and (3) subtle temporal variations across video frames are challenging to articulate with precision, often yielding generalized descriptions that constrain the Video-LLMs' potential. Consequently, existing Video-LLMs struggle with fine-grained video understanding tasks, particularly in aspects such as visual dynamics (e.g., motion patterns, object interactions) and video details inquiries (e.g., positional changes, detail variations)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.544 + ], + "angle": 0, + "content": "To address these challenges, we observe that finetuning Video-LLMs with self-supervised fragment tasks, by \"fragment\" we mean temporal frame level specifications of the video, could improve the model's sensitivity to spatiotemporal scene-level details (related to video contents). Driven by this, we introduce the Self-supervised Fragment Fine-Tuning \\((\\mathrm{SF}^2\\mathrm{T})\\), a effortless fine-tuning strategy for Video-LLMs that help to improve the fine-grained video understanding. \\(\\mathrm{SF}^2\\mathrm{T}\\) consists of five fragment-level tasks—Counting, Consistency Verification, Localization, Disorder Detection and Rearrangement—that automatically generate labels from various spatiotemporal perspectives. This approach maximizes the use of frame-level information while minimizing reliance on complex human instructions and annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Moreover, to evaluate the fine-grained visual dynamic perception of Video-LLMs and fully demonstrate the effectiveness of our \\(\\mathrm{SF}^2\\mathrm{T}\\), we present the FineVidBench, a novel benchmark. FineVidBench comprises 910 videos and 22,718 question-answer pairs, with videos sourced from diverse public datasets, including Something-Something V2 (SSv2) [6], Moments in Time (MiT) [21], etc. The question-answer pairs are auto-generated in single-choice format, incorporating distractors to increase testing difficulty. We evaluated several notable Video-LLMs developed in recent years, and find they generally fail to understand the execution sequence of actions and struggling to grasp fine-grained spatiotemporal information. While after fine-tuning with \\(\\mathrm{SF}^2\\mathrm{T}\\), the Video-LLMs better recognize spatiotemporal details, leading to a holistic and marked improvement in fine-grained understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Video-LLMs Finetuning Video-LLMs are primarily finetuned by adjusting the parameters of small, trainable adapters for task adaptation, without changing the entire model, saving resources and enhancing efficiency. The connective adapter (e.g., MLP/Linear Layer [15], Q" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "former [10]) links the Video Embedder and LLM, aligning video embeddings with LLM input tokens, while insertive adapters (e.g., LoRA [8]) are directly integrated into the LLM to modify its behavior. Most Video-LLMs combine both types of adapters and typically use multi-stage finetuning [4, 11, 13, 24, 35]. First, the model learns to establish relationships between images, videos, and text using large-scale multimodal datasets [1, 2, 29, 31]. In the second stage, the model is fine-tuned with an curated instruction-following dataset [11, 17, 18]. Besides, there are full finetuning, which updates all LLM parameters with a lower learning rate [25, 33], and zero-shot models, which transforms the video task into a text task, typically relying on a powerful LLM [32]. However, annotating video data remains a labor-intensive and time-consuming task, particularly for long videos or those involving complex actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.59 + ], + "angle": 0, + "content": "Benchmarks on Video-LLMs Currently, many studies [3, 5, 38] focus on evaluating the temporal perception capabilities of Video-LLMs. MVBench [12] designs 20 tasks from temporal and spatial perspectives, and Tempcompass [16] introduces 5 temporal aspects and 4 task formats. VN-Bench [36] decouples video content from the QA pairs by inserting irrelevant images or text \"needles\" into the original video. Moment-10M [22] has constructed a large-scale dataset on temporal localization tasks. However, as illustrated in Table 1, these studies often focus on gathering diverse videos or evaluating the models' performance with long videos, while somewhat neglecting the models' ability to perform fine-grained perception of temporal details. To address this gap, FineVidBench breaks videos into multiple sets of frames and generates annotations from diverse spatiotemporal perspectives, introducing novel evaluation methods for fine-grained understanding." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.753 + ], + "angle": 0, + "content": "
BenchmarksVideo num.QA num.Input ChangeTemporal DiversityFine-Grained EvaluationHierarchical Test
Video-MME9002700XXXX
TempCompass4107540X✓✓X
VN bench-1350X✓✓X
Moment-10M64.9k10.4MXXXX
AutoEval-Video327327XXXX
MV bench36414000XX✓X
MLVU13342593XXXX
FineVidBench91022,718✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.763, + 0.905, + 0.806 + ], + "angle": 0, + "content": "Table 1. Comparison with related benchmarks. Our approach offers significant advantages in input formats, evaluation methods, granularity, and temporal diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.761, + 0.846 + ], + "angle": 0, + "content": "3. FineVidBench Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "It is broadly recognized that Video-LLMs struggle with fine-grained video understanding tasks, yet no comprehensive benchmarks exist to thoroughly investigate this issue." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29109" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "To address this gap, we introduce FineVidBench, a multidimensional, fine-grained evaluation framework specifically designed to assess and improve the overall capabilities of Video-LLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.164, + 0.231, + 0.179 + ], + "angle": 0, + "content": "3.1. Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.261 + ], + "angle": 0, + "content": "Data collection We selected videos from various public datasets, including SS-v2 [6], MiT [21], and Ego4D [7], with a particular emphasis on temporally-sensitive content, to focus the model on the entire video sequence rather than individual frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.444 + ], + "angle": 0, + "content": "Action categorization As shown in Figure 2, we compiled 52 actions, categorizing them into 3 types based on intraclass variance. The distribution varies significantly: \"Distinctive Actions\" \\((39\\%)\\) are easily recognizable, encompassing a total of 36 actions. \"Non-typical Actions\" \\((57\\%)\\) refer to flexible actions with no clear defining characteristics, spanning 14 types. The broad diversity and complexity in this category require more extensive video coverage to adequately capture the range of expressions and variations. \"Slight Movements\" \\((4\\%)\\) represent subtle actions, such as \"hold\" and \"show\", which are difficult to detect with the naked eye and constitute a small proportion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Data augmentation The original videos were augmented using frame interpolation and skipping techniques for speed transformation, along with a motion-salient area sampling algorithm to capture dynamic motion. This process generated speed-varied versions and multiple sets of keyframes for each video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Statistics With our augmentation strategy, FineVidBench includes 910 videos, 1,820 speed-variant videos, and 2,670 sets of keyframes enriched with dynamic visual information. Building on this, we generated 22,718 QA pairs from the video content through a combination of automated processes and manual review. The quality assurance process involved rigorous cross-verification, where reviewers checked each QA pair for accuracy and contextual relevance, making corrections to ensure high quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.336, + 0.699 + ], + "angle": 0, + "content": "3.2. Benchmarking Dimensions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.903 + ], + "angle": 0, + "content": "As shown in Figure 3, FineVidBench encompasses both scene-level and fragment-level evaluations. The scene-level evaluation assesses both original and speed-adjusted videos across three dimensions: (1) Action, which evaluates the model's holistic understanding of video content. To increase difficulty, \"Visual Synonyms\" are added as distractors, requiring VideoLLM to distinguish visually similar actions with subtle differences, a challenge common in real-world scenarios. (2) Effect, which focuses on the model's comprehension of the visual changes resulting from actions. This understanding is essential for revealing object properties and interpreting complex dynamic scenes, and could significantly enhance the reasoning capabilities of Video" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.095, + 0.876, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.368, + 0.908, + 0.44 + ], + "angle": 0, + "content": "Figure 2. We show the action semantics and their respective proportions in FineVidBench. Distinctive Action: easily recognizable actions. Non-typical Action: flexible actions with no clear characteristics, like \"put\" and \"move.\" Slight Movement: subtle actions, such as \"hold\" and \"show,\" difficult to detect with the naked eye." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.553 + ], + "angle": 0, + "content": "LLMs and LLM-aided agents. (3) Speed, which tests the model's sensitivity to changes in video speed and its capability to maintain consistent understanding across varying speeds, with slow motion revealing hidden details and fast motion obscuring them. This capability is crucial for optimizing the model's performance across diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.827 + ], + "angle": 0, + "content": "For fragment-level evaluation, We've designed a structured evaluation format for video dynamic keyframes, employing a step-by-step inquiry framework: (1) Frame Count: Models are queried on the number of frames in sequences using dynamically refined keyframes to assess counting accuracy. (2) Meaning of Order: Understanding of sequence order is tested by asking about the first or last frames the targets appear in, or the frames they are present. e.g., \"At which frame does the target object first appear?\". (3) Frame Comparison: Two frames are randomly selected from the sequence for visual comparison, with differences varying in size but generally staying within human visual comfort limits. (4) Adjust-or-Not and Rearrangement: These two tasks involve a shuffled sequence of keyframes, and the model is asked to determine whether the order needs adjustment and, if so, how to correct it. They evaluate the model's ability to understand and restore the video's temporal sequence." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.834, + 0.704, + 0.849 + ], + "angle": 0, + "content": "3.3. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We evaluated six of the most advanced open-source models: LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4], Qwen2-VL[28], ShareGPT4Video [2] and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.427, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.089, + 0.66, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.113, + 0.887, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.314, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Figure 3. FineVidBench evaluates videos augmented with speed variations and fragments. Scene-level tests include the following: Action: Tests recognition accuracy amidst distractors like \"Visual Synonyms\". Effect: Assesses the model's ability to identify pre- and post-action changes. Speed: Measures the model's sensitivity to changes in video speed. Fragment-level tests, employing a step-by-step inquiry framework, focus on challenges such as Frame Count, Meaning of Order, Frame Comparison, Adjust-or-Not and Rearrangement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.458 + ], + "angle": 0, + "content": "Video-CCAM [27], each employing different architectures and training strategies. Table 3 summarizes the results across the eight tasks. We discuss the results from scene-level and fragment-level." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.334, + 0.485 + ], + "angle": 0, + "content": "- Scene-level Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Action The scores for this task varied significantly, with models trained in relevant video data—such as Video-CCAM, Qwen2-VL, and VideoLLaMA 2.1—achieving notably higher performance. However, as shown on the left side of Table 2, interference from \"Visual Synonyms\" prevented these models from achieving their full potential, resulting in declines of varying degrees and indicating difficulties in distinguishing visually similar actions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Effect All models exhibited average performance on this task, indicating a superficial understanding of aspects such as object attributes, object relationships, and action properties. This task tests the model's ability to grasp how actions affect objects, focusing on causal relationships and temporal reasoning—particularly for actions like \"push\" and \"pull\", which share similar execution flows. The model must distinguish them based on dynamic effects, such as changes in direction and speed, but most models perform moderately in this regard." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Speed The results show that all models are insensitive to speed variations, likely because they were not adequately exposed to speed changes during training. Figure 4 shows that models are more sensitive to slow motion than fast playback, and struggled with identifying \"normal speed\" and \"no speed\", except for VideoLLaMA 2.1. This may be due to the loss of coherence in fast-moving video content, while slow-motion videos highlight more distinct details, aiding the model in making accurate judgments." + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.397, + 0.876, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.594, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Figure 4. Accuracy across different video speeds. All models are more sensitive to slow-speed videos and struggle to understand \"normal speed\" and \"no speed\", except for VideoLLaMA 2.1." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.647, + 0.917, + 0.802 + ], + "angle": 0, + "content": "
Video-LLMsActionFrame Number
w/o VSw/ VSAvg.345
LLaVA-NeXT-Video37.3135.0419.3720.3319.7717.98
MiniCPM-V 2.643.3740.1590.3293.8290.6686.44
Video-LLaMA 2.163.2653.9830.1742.8639.897.45
Qwen2-VL68.1856.6296.6597.2596.6396.05
ShareGPT4Video46.9030.8426.3360.9916.780.00
Video-CCAM73.1060.2323.4514.188.9647.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.812, + 0.907, + 0.895 + ], + "angle": 0, + "content": "Table 2. Left: Accuracy of the Action task with or without \"Visual Synonyms\". It is obvious that the \"Visual Synonyms\" have significantly impacted the model's judgment. Right: Accuracy of the counting task across different frame counts. Except for Video-CCAM, all other models exhibited a decline in performance as the number of frames increased." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29111" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.304 + ], + "angle": 0, + "content": "
Video-LLMsParams.Scene-LevelFragment-LevelS-Avg.FG-Avg.A-Avg.
ActionEffectSpeedFCntMoOFCmpAoNRearr
(Random)-25.0025.0025.0025.0025.0033.3333.3325.0025.0028.3327.08
LLaVA-NeXT-Video7B37.3142.6722.3519.3724.0253.7575.4520.6734.1138.6536.95
MiniCPM-V 2.68B43.3752.5619.1390.3256.4275.6676.4918.0938.3563.4054.01
Video-LLaMA 2.17B63.2650.9219.8930.1742.2776.0189.9226.8744.6953.0549.91
Qwen2-VL7B68.1857.1424.6296.6533.3374.5390.7022.4849.9863.5458.45
ShareGPT4Video8B46.9043.8831.7626.3361.0588.4484.8023.3640.8557.1150.82
Video-CCAM9B73.1055.9031.6523.4545.6664.9590.2722.7253.5548.4750.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.314, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Table 3. The overall performances of notable Video-LLMs on FineVidBench. FCnt: Frame Count. MoO: Meaning of Order. Fcmp: Frame Comparison. AoN: Adjust or Not. Rearr: Rearrangement. S-Avg.: the average performance of scene-level tasks; FG-Avg.: the average performance of fragment-level tasks. A-Avg.: the average performance of all tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.383, + 0.363, + 0.399 + ], + "angle": 0, + "content": "- Fragment-level Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.486, + 0.692 + ], + "angle": 0, + "content": "(1) Frame-count accuracy varied significantly across models, with the lower-performing models likely lacking targeted training. The trend shown in the right side of Table 2, where accuracy decreases as frame count increases, highlights the models' insufficient temporal reasoning on longer sequences. (2) ShareGPT4Video and MiniCPM-V 2.6 showed better comprehension in the Meaning-of-Order task, while other models lagged, suggesting a lack of explicit focus on \"order\". (3) Most models excelled in frame comparison due to image-text alignment training. ShareGPT4Video achieved the best performance, owing to its Differential Sliding-Window Captioning (DiffSW) strategy, which emphasizes capturing the changes between frames when generating video descriptions. This also improved its Meaning-of-Order performance. (4) In the sorting task, models generally succeeded in the \"Adjust or Not\" response but performed poorly in the more complex \"Rearrangement\" task, indicating they can detect, but not correct, sequence errors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.708, + 0.427, + 0.726 + ], + "angle": 0, + "content": "4. Self-supervised Fragment Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "The above benchmark results show the existing Video-LLMs generally fail to tackle fine-grained video understanding tasks. Videos often contain subtle, complex changes that natural language alone fails to fully capture. The core component of Video-LLMs, LLMs, as generalized pattern recognizers, offers a promising solution. LLMs have the potential to detect and interpret intricate spatiotemporal dynamics that were previously difficult to represent. Given that these changes cannot be directly annotated, using self-supervised learning naturally becomes the solution, bypassing the bottleneck of manual annotation and significantly re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.908, + 0.492 + ], + "angle": 0, + "content": "ducing labeling costs. Given these factors, we propose the \\(\\mathrm{SF^2T}\\) to fine-tune Video-LLMs. While we do not expect \\(\\mathrm{SF^2T}\\) to replace the supervised fine-tuning, instead it's an effortless complementary to SFT. Comparing \\(\\mathrm{SF^2T}\\) with SFT, they primarily differ in data construction and content focus level, with each method aligned with distinct training objectives as shown in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.633, + 0.513 + ], + "angle": 0, + "content": "4.1. SFT Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.551 + ], + "angle": 0, + "content": "We first review the common SFT tasks to set a baseline for comparing our \\(\\mathrm{SF^2T}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.642 + ], + "angle": 0, + "content": "General QA on Video Content This method focuses on understanding the main events and context of a video by directly asking questions about its content. While effective for grasping the video's key moments, it lacks finer spatiotemporal details and requires significant human effort to create standardized but constrained answers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Frame Description Integration This method typically samples video frames evenly, generates detailed descriptions for each, and integrates them into a cohesive but lengthy summary. While it enhances the model's understanding of continuity and micro-dynamics, it often proves incapable of capturing complex or subtle details that are beyond natural language's scope. Moreover, although frame descriptions can be generated using powerful multi-model LLMs like GPT-4o, significant human effort is still required to review the quality of the generated responses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.78, + 0.82 + ], + "angle": 0, + "content": "4.2. Fragment-level Tasks of \\(\\mathbf{SF}^2\\mathbf{T}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.903 + ], + "angle": 0, + "content": "SFT tasks require manual annotations, and even automation annotation is labor-intensive and error-prone. To address, we introduce \\(\\mathrm{SF}^2\\mathrm{T}\\) which generates accurate fragment-level labels accurately. \\(\\mathrm{SF}^2\\mathrm{T}\\) comprises five tasks—Counting, Consistency Verification, Localization, Disorder Detection" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29112" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.094, + 0.478, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.162, + 0.385, + 0.176 + ], + "angle": 0, + "content": "What is the main content of the video?" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.181, + 0.424, + 0.213 + ], + "angle": 0, + "content": "The video shows a person bowling, including their four-step approach, the smooth release of the ball down the lane, its path toward the pins, and..." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.231, + 0.476, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.306, + 0.387, + 0.32 + ], + "angle": 0, + "content": "What is the main content of the video?" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.325, + 0.425, + 0.378 + ], + "angle": 0, + "content": "The video shows a person bowling: (Frame 1) The scene shows a bowling alley... (Frame 2) The player swing the bowling ball... (Frame 4) The bowling ball approaches the pins... (Frame 6) The bowling ball strikes the pins... (Frame 8) All the pins are down." + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.32, + 0.448, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.395, + 0.361, + 0.408 + ], + "angle": 0, + "content": "Scene-Level Tasks" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.42, + 0.445, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.506, + 0.144, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.511, + 0.268, + 0.524 + ], + "angle": 0, + "content": "How many frames?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.531, + 0.264, + 0.542 + ], + "angle": 0, + "content": "On which frames?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.55, + 0.242, + 0.561 + ], + "angle": 0, + "content": "Same frames?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.581, + 0.237, + 0.593 + ], + "angle": 0, + "content": "Adjust or not?" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.6, + 0.231, + 0.611 + ], + "angle": 0, + "content": "Rearrange it." + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.508, + 0.371, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.53, + 0.367, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.549, + 0.367, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.569, + 0.367, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.602, + 0.367, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.51, + 0.449, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.531, + 0.421, + 0.542 + ], + "angle": 0, + "content": "2nd" + }, + { + "type": "text", + "bbox": [ + 0.401, + 0.551, + 0.421, + 0.561 + ], + "angle": 0, + "content": "No" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.581, + 0.421, + 0.592 + ], + "angle": 0, + "content": "Yes" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.6, + 0.421, + 0.61 + ], + "angle": 0, + "content": "3412" + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.639, + 0.374, + 0.654 + ], + "angle": 0, + "content": "Fragment-Level Tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Figure 5. Comparison between \\(\\mathrm{SF}^2\\mathrm{T}\\) and SFT. SFT depends on manual and model-driven design to generate QA pairs for scene-level video understanding, \\(\\mathrm{SF}^2\\mathrm{T}\\), in contrast, automatically constructs training data based on pre-defined rules that cover various temporal and spatial aspects of the video. \\(\\mathrm{SF}^2\\mathrm{T}\\) enables the model to focus on a fine-grained content analysis, and offering insights that supervised labels cannot achieve." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "and Rearrangement—designed to train the model to rearrange a set of out-of-order frames into their original sequence. This is a robust indicator of a modal's mastery over the visual dynamics of an action, requiring the model to detect subtle frame changes and understand the overall coherence and temporal trends. Mastery of these tasks enables the model to recognize frames and their temporal re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "relationships, enhancing its ability to predict and reconstruct action sequences and improving performance on more complex video tasks. Our method first extracts multiple sets of dynamic keyframes from each video. These fragments capture the key dynamic information from multiple temporal perspectives, offering a more efficient representation of redundant video data. It then applies pseudo-labeling, distinguishing it from traditional video-level labeling. By designing proxy tasks that leverage intrinsic information rather than predefined prior knowledge, it smartly circumvents the annotation bottleneck, enabling a deeper temporal understanding and offering insights that traditional video-level labeling cannot achieve." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- Counting We input N frames into the Video-LLM and ask it to count them. Although this task seems straightforward, it proves challenging for current Video-LLMs, particularly as the number of frames increases, revealing a decline in accuracy. The model's inability to perform basic quantitative tasks points to a broader limitations in understanding the overall sequence integrity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.502 + ], + "angle": 0, + "content": "Consistency Verification Video-LLMs are tasked with identifying two frames sampled from the same video, which may show subtle differences. This task sharpens the model's sensitivity to visual details by encouraging a thorough analysis and comparison of the images, countering its tendency to focus on primary subjects while neglecting the background and other subtle features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Localization Video-LLMs must accurately locate a specified target (from video metadata) within a sequence of frames, identifying the frames in which it appears, disappears, or persists. This naturally human ability is a significant challenge for these models, as they often struggle to perceive sequential relationships between frames and face additional obstacles, such as occlusion, interference from similar objects, lighting variations, and memory limitations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Disorder Detection and Rearrangement Video-LLMs must determine whether and how to adjust the order of a given frame sequence. When frames are randomized, the loss of spatiotemporal coherence and logical continuity makes it exceptionally challenging to reconstruct their original sequence, especially as interactions within frames become more complex [20]. This task is evaluated in two ways: the yes/no task tests the model's sensitivity to temporal consistency, while the sorting task, which leverages capabilities from the other four tasks, requires advanced reasoning and adjustments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.645, + 0.83 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we fine-tuned four of the most advanced open-source Video-LLMs using the \\(\\mathrm{SF}^2\\mathrm{T}\\) method to evaluate its effectiveness, alongside ablation studies and interpretability analyses to explore the underlying mechanisms." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29113" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.24 + ], + "angle": 0, + "content": "
MethodsLLaVA-NEXT-VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2-VL
ActionEffectSpeedActionEffectSpeedActionEffectSpeedActionEffectSpeed
Base37.3142.6722.3543.3752.5619.1363.2650.9219.8968.1857.1424.62
Base+SF2T48.6743.7724.8365.9160.6228.6067.4257.3331.6373.8663.3731.92
Base(SFT)62.6944.6322.3577.6575.0970.8377.6565.9429.7378.6066.3030.87
Base(SFT)+SF2T63.0745.2432.0181.6376.9286.7479.7368.6831.8281.2573.2632.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.25, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Table 4. Performance on FineVidBench. We tested on two baselines: (1) Base: Results without any fine-tuning. (2) Base(SFT): Results after fine-tuning in supervised way. After \\(\\mathrm{SF}^2\\mathrm{T}\\), all models improved in all three tasks, highlighting its broad effectiveness and the value of fragment-level tasks in enhancing scene-level comprehension. Notably, \\(\\mathrm{SF}^2\\mathrm{T}\\) outperformed SFT in the Speed task (except MiniCPM-V 2.6), highlighting the key role of fine-grained temporal understanding in distinguishing video speeds." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.319, + 0.475, + 0.507 + ], + "angle": 0, + "content": "
MethodsLLaVA-NeXT -VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2 -VL
MVBench
Base36.8440.2354.1855.97
Base+SF2T42.9256.0257.9763.76
Video-MME(no subtitle)
Base29.7643.1749.0243.77
Base+SF2T34.8453.1951.8853.60
MLVU
Base36.3241.5852.3242.81
Base+SF2T41.9155.3256.1154.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.516, + 0.486, + 0.56 + ], + "angle": 0, + "content": "Table 5. Performance on public benchmarks. \\(\\mathrm{SF}^2\\mathrm{T}\\) consistently enhances performance across all three benchmarks, reaffirming its effectiveness as a spatiotemporal enhancer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.308, + 0.601 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.746 + ], + "angle": 0, + "content": "To ensure fairness, experiments were conducted on LoRA-compatible models, including LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4] and Qwen2-VL[28], using their default or recommended settings, with all models trained for one epoch. All experiments were performed under identical hardware conditions, utilizing NVIDIA A100 40GB GPU for computation. It should be emphasized that our goal is to validate the effectiveness of \\(\\mathrm{SF}^2\\mathrm{T}\\), not to optimize models for maximum performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We randomly sampled videos from SSv2 and MiT for training, ensuring no overlap with the FineVidBench dataset. MGSampler [37] was used to extract N sets of M-frame sequences from each video, capturing dynamic changes while preserving overall characteristics. M is chosen based on the video's characteristics to capture content flow, while N is determined by content complexity, with more complex content requiring a larger N to cover more temporal perspectives. In this study, we set \\( \\mathrm{N} = 3 \\) and M between 3 and 5, though these values may vary for other" + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.324, + 0.878, + 0.376 + ], + "angle": 0, + "content": "
Methodsrandomuniformkeyframemotion-salient
SF2T70.3171.6772.1173.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.387, + 0.907, + 0.443 + ], + "angle": 0, + "content": "Table 6. Impact of sampling. As shown, motion-salient area sampling outperforms others by better capturing motion fluidity and temporal details, while the other methods fail to fully utilize their potential, leading to suboptimal performance." + }, + { + "type": "table", + "bbox": [ + 0.606, + 0.45, + 0.816, + 0.502 + ], + "angle": 0, + "content": "
Methodslongshortrandom
SF2T69.3871.4073.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.513, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Table 7. Impact of temporal span. Both long- and short-range temporal modeling reduced \\(\\mathrm{SF}^2\\mathrm{T}\\) 's performance, emphasizing the importance of multi-scale temporal modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.678 + ], + "angle": 0, + "content": "datasets. We then generated QA pairs for each frame sequence based on the five tasks defined in \\(\\mathrm{SF^2T}\\) for training. Evaluations were performed on FineVidBench's scene-level tasks, including Action, Effect, and Speed. To compare with traditional SFT, we also generated and manually reviewed QA pairs for these videos in a supervised setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.655, + 0.712 + ], + "angle": 0, + "content": "5.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Table 4 summarizes the results of the scene-level tasks. After \\(\\mathrm{SF}^2\\mathrm{T}\\) training, all models showed significant improvement, emphasizing that fragment-level tasks can notably enhance scene-level comprehension. Integrating \\(\\mathrm{SF}^2\\mathrm{T}\\) with SFT is also leads to performance gains, demonstrating that fragment-level training positively impacts SFT and enhances its effectiveness. Surprisingly, in the Speed task, many base models outperformed SFT after applying \\(\\mathrm{SF}^2\\mathrm{T}\\), highlighting the importance of fine-grained temporal understanding in distinguishing video speeds. This improvement likely stems from \\(\\mathrm{SF}^2\\mathrm{T}\\)'s ability to enhance the model's sensitivity to temporal cues, such as the loss or enhancement of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29114" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "information during acceleration or deceleration, as well as content coherence—all crucial for speed judgment. As expected, \\(\\mathrm{SF}^2\\mathrm{T}\\) currently lags behind SFT, since its training objective is not fully aligned with scene-level tasks. However, we do not expect \\(\\mathrm{SF}^2\\mathrm{T}\\) to replace supervised fine-tuning; rather, our experiments suggest that it can serve as an effortless and effective complement to SFT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.29 + ], + "angle": 0, + "content": "In addition to FineVidBench, we evaluated \\(\\mathrm{SF}^{2}\\mathrm{T}\\) on three public video understanding benchmarks (Table 5). The results demonstrate consistent improvements across various video tasks, validating \\(\\mathrm{SF}^{2}\\mathrm{T}\\) as an effective spatiotemporal enhancer for a wide range of video understanding tasks. All models were tested with an 8-frame input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.302, + 0.423, + 0.318 + ], + "angle": 0, + "content": "5.3. Ablation and Interpretability Analyses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.483, + 0.581 + ], + "angle": 0, + "content": "We evaluated the impact of frame sampling strategies on \\(\\mathrm{SF}^2\\mathrm{T}\\), as each method provides a unique \"temporal information perspective\" that influencing video understanding performance. As shown in Table 6, we assessed four strategies on Qwen2-VL in the Action task: random, uniform interval, keyframe, and motion-salient area sampling [37]. Motion-salient area sampling performed best, likely due to its ability to capture continuous motion dynamics, thereby enhancing the model's understanding of action fluidity and temporal detail. In comparison, the other methods had limitations: keyframe sampling misses intermediate action phases, fixed-interval sampling may overlook critical moments, and random sampling lacks temporal consistency. Notably, different datasets may favor different strategies. For example, some datasets may perform better with uniform interval sampling, or their motion features may align better with the model's specific capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We examined the effects of long- and short-range temporal modeling on \\(\\mathrm{SF}^2\\mathrm{T}\\). In the Consistency Verification task, we constrained the random selection of frame pairs to adjacent frames for local continuity or non-adjacent frames to capture long-range dependencies. As shown in Table 7, both settings decreased \\(\\mathrm{SF}^2\\mathrm{T}\\)'s performance on the Action task of Qwen2-VL, indicating that an overemphasis on either long- or short-range information leads to temporal imbalance and incomplete dynamics. This underscores the importance of combining both approaches to leverage their broader temporal span and frame variations for a more comprehensive feature representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We analyzed the attention map of Qwen2-VL on the Action task, particularly in cases where the model's predictions were corrected after \\(\\mathrm{SF}^2\\mathrm{T}\\). As shown in Figure 6, we found that \\(\\mathrm{SF}^2\\mathrm{T}\\) enhances the model's ability to capture fine-grained spatial changes and temporal dynamics. (1) Spatial Aspects. After \\(\\mathrm{SF}^2\\mathrm{T}\\), the model shows increased attention to action execution areas, particularly the hands and objects they interact with. It shows better sensitivity to small targets, likely due to the Consistency Verification" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.885, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.461, + 0.907, + 0.557 + ], + "angle": 0, + "content": "Figure 6. Two exemplary visualizations of the attention map on Qwen2-VL. For each example: top - Original frames; middle - Base (SFT); bottom - \\(\\mathrm{SF^2T}\\) applied. As shown by the red boxes, after applying \\(\\mathrm{SF^2T}\\), the model better focuses on action execution areas and interacting objects. The \\(\\mathrm{SF^2T}\\) fine-tuned model has the ability to predict the direction of motion, as seen in the trajectories of the red bottle and Cheerios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.704 + ], + "angle": 0, + "content": "task, which enhances spatial perception by refining sensitivity to subtle image differences. (2) Temporal Aspects. After \\(\\mathrm{SF}^2\\mathrm{T}\\), we observed that the model can predict object movement trajectories in certain actions, indicating an advanced level of temporal understanding. This ability likely stems from the sorting task, which strengthens the model's comprehension of action flows and movement patterns." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.734, + 0.634, + 0.75 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose \\(\\mathrm{SF}^2\\mathrm{T}\\) to overcome the limitations of Video-LLMs in fine-grained video understanding. \\(\\mathrm{SF}^2\\mathrm{T}\\) is an innovative fine-tuning method that eliminates the need for labor-intensive annotations and effectively bypasses the constraints of natural language descriptions. Additionally, we introduce FineVidBench, a benchmark for evaluating Video-LLMs at both scene and fragment levels. In the future, we plan to expand our dataset with larger videos and more tasks to increase its impact." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29115" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.298 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China (No.2020YBF2901202), National Natural Science Foundation of China (NSFC No. 62272184 and No. 62402189), the China Postdoctoral Science Foundation under Grant Number GZC20230894, the China Postdoctoral Science Foundation (Certificate Number: 2024M751012), and the Postdoctor Project of Hubei Province under Grant Number 2024HBBHCXB014, and the \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (No. 2024C01161). The computation is completed in the HPC Platform of Huazhong University of Science and Technology." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.311, + 0.188, + 0.327 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[1] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.366, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[2] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[3] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.484, + 0.561 + ], + "angle": 0, + "content": "[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[5] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[6] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[7] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[8] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave:" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.336, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.906, + 0.249 + ], + "angle": 0, + "content": "[11] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.906, + 0.335 + ], + "angle": 0, + "content": "[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.906, + 0.38 + ], + "angle": 0, + "content": "[13] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.906, + 0.438 + ], + "angle": 0, + "content": "[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.483 + ], + "angle": 0, + "content": "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.486, + 0.906, + 0.54 + ], + "angle": 0, + "content": "[16] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.906, + 0.612 + ], + "angle": 0, + "content": "[17] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.906, + 0.729 + ], + "angle": 0, + "content": "[19] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[20] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 527-544. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29116" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[22] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[24] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[25] Fangxun Shu, Lei Zhang, Hao Jiang, and Cihang Xie. Audio-visual llm for video understanding. arXiv preprint arXiv:2312.06720, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[26] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[27] TencentQQ Multimedia Research Team. Video-cam: Advancing video-language understanding with causal cross-attention masks. https://github.com/QQ-MM/Video-CCAM, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[28] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[29] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[30] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[31] Haiyang Xu, Qinghao Ye, Xuan Wu, Ming Yan, Yuan Miao, Jiabo Ye, Guohai Xu, Anwen Hu, Yaya Shi, Guangwei Xu, et al. Youku-mplug: A 10 million large-scale chinese video-language dataset for pre-training and benchmarks. arXiv preprint arXiv:2306.04362, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[32] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[33] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[34] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[35] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[36] Zijia Zhao, Haoyu Lu, Yuqi Huo, Yifan Du, Tongtian Yue, Longteng Guo, Bingning Wang, Weipeng Chen, and Jing Liu. Needle in a video haystack: A scalable synthetic framework for benchmarking video mllms. arXiv preprint arXiv:2406.09367, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[37] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[38] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[39] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29117" + } + ] +] \ No newline at end of file diff --git a/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_origin.pdf b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c22887e6588185b0c2ba52fdbd6371c8bc4fa1f3 --- /dev/null +++ b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/d4e22b54-75f3-40d9-971f-0e4ee1868232_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e3dfa9e709d1aa5f0ee327ce805e97568d9b5b7e88a106e66482f3618d979fa +size 1490669 diff --git a/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/full.md b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4cdfab98a145ba9adb43c83fd20fad8982f6b0b7 --- /dev/null +++ b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/full.md @@ -0,0 +1,288 @@ +# $\mathbf{SF}^{2}\mathbf{T}$ : Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding + +Yangliu Hu $^{1}$ , Zikai Song $^{1\dagger}$ , Na Feng $^{1}$ , Yawei Luo $^{2}$ , Junqing Yu $^{1}$ , Yi-Ping Phoebe Chen $^{3}$ , Wei Yang $^{1\dagger}$ $^{1}$ Huazhong University of Science and Technology $^{2}$ Zhejiang University $^{3}$ La Trobe University + +{huyangliu,skyesong,fengna,yjqing,weiyangcs}@hust.edu.cn + +yaweiluo@zju.edu.cn phoebe.chen@latrobe.edu.au + +# Abstract + +Video-based Large Language Models (Video-LLMs) have witnessed substantial advancements in recent years, propelled by the advancement in multi-modal LLMs. Although these models have demonstrated proficiency in providing the overall description of videos, they struggle with fine-grained understanding, particularly in aspects such as visual dynamics and video details inquiries. To tackle these shortcomings, we find that fine-tuning Video-LLMs on self-supervised fragment tasks, greatly improve their fine-grained video understanding abilities. Hence we propose two key contributions: (1) Self-Supervised Fragment Fine-Tuning $(SF^2 T)$ , a novel effortless fine-tuning method, employs the rich inherent characteristics of videos for training, while unlocking more fine-grained understanding ability of Video-LLMs. Moreover, it relieves researchers from labor-intensive annotations and smartly circumvents the limitations of natural language, which often fails to capture the complex spatiotemporal variations in videos; (2) A novel benchmark dataset, namely FineVidBench, for rigorously assessing Video-LLMs' performance at both the scene and fragment levels, offering a comprehensive evaluation of their capabilities. We assessed multiple models and validated the effectiveness of $SF^2 T$ on them. Experimental results reveal that our approach improves their ability to capture and interpret spatiotemporal details. + +# 1. Introduction + +Large Language Models (LLMs) have showcased significant emergent capabilities, such as in-context learning [19], instruction-following [23], and chain-of-thought reasoning [30], driven by expansive datasets and advanced model architectures. Extending these advancements, Video-LLMs through mechanisms like pooling or query aggregation + +![](images/8b7eb9014d7abb5b36913b7f76bf78d0aca112be4c48c98f53b28ffaa2069a16.jpg) +Figure 1. Performance w/ and w/o $\mathbf{SF}^2\mathbf{T}$ . We evaluated four advanced Video-LLMs w/ and w/o $\mathrm{SF}^2\mathrm{T}$ on our proposed FineVidBench with two baselines: (1) Base: performance without any fine-tuning (blue dashed), and (2) Base (SFT): performance with supervised fine-tuning (red dashed). After applying $\mathrm{SF}^2\mathrm{T}$ , all models showed significant improvements (solid blue and red), underscoring its broad effectiveness. + +across numerous visual tokens, have broadened the scope of LLMs to encompass video information processing [11, 14, 35]. This evolution markedly advances their potential for in-depth real-world comprehension, opening applications in intelligent surveillance, virtual reality, and autonomous driving, further enriching the landscape of video analytics and interpretation. + +Various Video-LLMs, exemplified by GPT4-V, VideoLaMA 2 [4], MiniCPM-V [34], and Qwen2-VL [28], have been crafted by leading corporations and research institutions, demonstrating proficiency in capturing the overarching content of videos. When adapting to new videos and tasks, they predominantly rely on Supervised FineTuning (SFT) [26] or Reinforcement Learning from Hu + +man Feedback (RLHF) [39], both of which are heavily contingent upon extensive manual annotation. This dependence poses several key problems: (1) it necessitates substantial human resources, particularly highly trained annotators; (2) the inherent complexity of video content and task demands frequently introduces inconsistencies and subjectivity, rendering the maintenance of high-quality annotations particularly arduous; and (3) subtle temporal variations across video frames are challenging to articulate with precision, often yielding generalized descriptions that constrain the Video-LLMs' potential. Consequently, existing Video-LLMs struggle with fine-grained video understanding tasks, particularly in aspects such as visual dynamics (e.g., motion patterns, object interactions) and video details inquiries (e.g., positional changes, detail variations). + +To address these challenges, we observe that finetuning Video-LLMs with self-supervised fragment tasks, by "fragment" we mean temporal frame level specifications of the video, could improve the model's sensitivity to spatiotemporal scene-level details (related to video contents). Driven by this, we introduce the Self-supervised Fragment Fine-Tuning $(\mathrm{SF}^2\mathrm{T})$ , a effortless fine-tuning strategy for Video-LLMs that help to improve the fine-grained video understanding. $\mathrm{SF}^2\mathrm{T}$ consists of five fragment-level tasks—Counting, Consistency Verification, Localization, Disorder Detection and Rearrangement—that automatically generate labels from various spatiotemporal perspectives. This approach maximizes the use of frame-level information while minimizing reliance on complex human instructions and annotations. + +Moreover, to evaluate the fine-grained visual dynamic perception of Video-LLMs and fully demonstrate the effectiveness of our $\mathrm{SF}^2\mathrm{T}$ , we present the FineVidBench, a novel benchmark. FineVidBench comprises 910 videos and 22,718 question-answer pairs, with videos sourced from diverse public datasets, including Something-Something V2 (SSv2) [6], Moments in Time (MiT) [21], etc. The question-answer pairs are auto-generated in single-choice format, incorporating distractors to increase testing difficulty. We evaluated several notable Video-LLMs developed in recent years, and find they generally fail to understand the execution sequence of actions and struggling to grasp fine-grained spatiotemporal information. While after fine-tuning with $\mathrm{SF}^2\mathrm{T}$ , the Video-LLMs better recognize spatiotemporal details, leading to a holistic and marked improvement in fine-grained understanding. + +# 2. Related Work + +Video-LLMs Finetuning Video-LLMs are primarily finetuned by adjusting the parameters of small, trainable adapters for task adaptation, without changing the entire model, saving resources and enhancing efficiency. The connective adapter (e.g., MLP/Linear Layer [15], Q + +former [10]) links the Video Embedder and LLM, aligning video embeddings with LLM input tokens, while insertive adapters (e.g., LoRA [8]) are directly integrated into the LLM to modify its behavior. Most Video-LLMs combine both types of adapters and typically use multi-stage finetuning [4, 11, 13, 24, 35]. First, the model learns to establish relationships between images, videos, and text using large-scale multimodal datasets [1, 2, 29, 31]. In the second stage, the model is fine-tuned with an curated instruction-following dataset [11, 17, 18]. Besides, there are full finetuning, which updates all LLM parameters with a lower learning rate [25, 33], and zero-shot models, which transforms the video task into a text task, typically relying on a powerful LLM [32]. However, annotating video data remains a labor-intensive and time-consuming task, particularly for long videos or those involving complex actions. + +Benchmarks on Video-LLMs Currently, many studies [3, 5, 38] focus on evaluating the temporal perception capabilities of Video-LLMs. MVBench [12] designs 20 tasks from temporal and spatial perspectives, and Tempcompass [16] introduces 5 temporal aspects and 4 task formats. VN-Bench [36] decouples video content from the QA pairs by inserting irrelevant images or text "needles" into the original video. Moment-10M [22] has constructed a large-scale dataset on temporal localization tasks. However, as illustrated in Table 1, these studies often focus on gathering diverse videos or evaluating the models' performance with long videos, while somewhat neglecting the models' ability to perform fine-grained perception of temporal details. To address this gap, FineVidBench breaks videos into multiple sets of frames and generates annotations from diverse spatiotemporal perspectives, introducing novel evaluation methods for fine-grained understanding. + +
BenchmarksVideo num.QA num.Input ChangeTemporal DiversityFine-Grained EvaluationHierarchical Test
Video-MME9002700XXXX
TempCompass4107540X✓✓X
VN bench-1350X✓✓X
Moment-10M64.9k10.4MXXXX
AutoEval-Video327327XXXX
MV bench36414000XX✓X
MLVU13342593XXXX
FineVidBench91022,718✓✓✓✓
+ +Table 1. Comparison with related benchmarks. Our approach offers significant advantages in input formats, evaluation methods, granularity, and temporal diversity. + +# 3. FineVidBench Benchmark + +It is broadly recognized that Video-LLMs struggle with fine-grained video understanding tasks, yet no comprehensive benchmarks exist to thoroughly investigate this issue. + +To address this gap, we introduce FineVidBench, a multidimensional, fine-grained evaluation framework specifically designed to assess and improve the overall capabilities of Video-LLMs. + +# 3.1. Construction + +Data collection We selected videos from various public datasets, including SS-v2 [6], MiT [21], and Ego4D [7], with a particular emphasis on temporally-sensitive content, to focus the model on the entire video sequence rather than individual frames. + +Action categorization As shown in Figure 2, we compiled 52 actions, categorizing them into 3 types based on intraclass variance. The distribution varies significantly: "Distinctive Actions" $(39\%)$ are easily recognizable, encompassing a total of 36 actions. "Non-typical Actions" $(57\%)$ refer to flexible actions with no clear defining characteristics, spanning 14 types. The broad diversity and complexity in this category require more extensive video coverage to adequately capture the range of expressions and variations. "Slight Movements" $(4\%)$ represent subtle actions, such as "hold" and "show", which are difficult to detect with the naked eye and constitute a small proportion. + +Data augmentation The original videos were augmented using frame interpolation and skipping techniques for speed transformation, along with a motion-salient area sampling algorithm to capture dynamic motion. This process generated speed-varied versions and multiple sets of keyframes for each video. + +Statistics With our augmentation strategy, FineVidBench includes 910 videos, 1,820 speed-variant videos, and 2,670 sets of keyframes enriched with dynamic visual information. Building on this, we generated 22,718 QA pairs from the video content through a combination of automated processes and manual review. The quality assurance process involved rigorous cross-verification, where reviewers checked each QA pair for accuracy and contextual relevance, making corrections to ensure high quality. + +# 3.2. Benchmarking Dimensions + +As shown in Figure 3, FineVidBench encompasses both scene-level and fragment-level evaluations. The scene-level evaluation assesses both original and speed-adjusted videos across three dimensions: (1) Action, which evaluates the model's holistic understanding of video content. To increase difficulty, "Visual Synonyms" are added as distractors, requiring VideoLLM to distinguish visually similar actions with subtle differences, a challenge common in real-world scenarios. (2) Effect, which focuses on the model's comprehension of the visual changes resulting from actions. This understanding is essential for revealing object properties and interpreting complex dynamic scenes, and could significantly enhance the reasoning capabilities of Video + +![](images/8e219e19edad16fd218dca29a3cb48b7d80101a471ef88867ad2f65e384007a1.jpg) +Figure 2. We show the action semantics and their respective proportions in FineVidBench. Distinctive Action: easily recognizable actions. Non-typical Action: flexible actions with no clear characteristics, like "put" and "move." Slight Movement: subtle actions, such as "hold" and "show," difficult to detect with the naked eye. + +LLMs and LLM-aided agents. (3) Speed, which tests the model's sensitivity to changes in video speed and its capability to maintain consistent understanding across varying speeds, with slow motion revealing hidden details and fast motion obscuring them. This capability is crucial for optimizing the model's performance across diverse scenarios. + +For fragment-level evaluation, We've designed a structured evaluation format for video dynamic keyframes, employing a step-by-step inquiry framework: (1) Frame Count: Models are queried on the number of frames in sequences using dynamically refined keyframes to assess counting accuracy. (2) Meaning of Order: Understanding of sequence order is tested by asking about the first or last frames the targets appear in, or the frames they are present. e.g., "At which frame does the target object first appear?". (3) Frame Comparison: Two frames are randomly selected from the sequence for visual comparison, with differences varying in size but generally staying within human visual comfort limits. (4) Adjust-or-Not and Rearrangement: These two tasks involve a shuffled sequence of keyframes, and the model is asked to determine whether the order needs adjustment and, if so, how to correct it. They evaluate the model's ability to understand and restore the video's temporal sequence. + +# 3.3. Benchmark Results + +We evaluated six of the most advanced open-source models: LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4], Qwen2-VL[28], ShareGPT4Video [2] and + +![](images/f2c27bd209bc48a9f381966094271a4758a614629c12b08a213350398427d324.jpg) +Figure 3. FineVidBench evaluates videos augmented with speed variations and fragments. Scene-level tests include the following: Action: Tests recognition accuracy amidst distractors like "Visual Synonyms". Effect: Assesses the model's ability to identify pre- and post-action changes. Speed: Measures the model's sensitivity to changes in video speed. Fragment-level tests, employing a step-by-step inquiry framework, focus on challenges such as Frame Count, Meaning of Order, Frame Comparison, Adjust-or-Not and Rearrangement. + +![](images/9726bd792966151b9e71d0c4fb9e8145f434a5d37572a9bba433c9ff69e6d8e5.jpg) + +![](images/56d0a2df8e3d8c352b70564101947f32b8297d02ddf5616cb21cf61904f974f5.jpg) + +Video-CCAM [27], each employing different architectures and training strategies. Table 3 summarizes the results across the eight tasks. We discuss the results from scene-level and fragment-level. + +# - Scene-level Results and Analysis + +Action The scores for this task varied significantly, with models trained in relevant video data—such as Video-CCAM, Qwen2-VL, and VideoLLaMA 2.1—achieving notably higher performance. However, as shown on the left side of Table 2, interference from "Visual Synonyms" prevented these models from achieving their full potential, resulting in declines of varying degrees and indicating difficulties in distinguishing visually similar actions. + +Effect All models exhibited average performance on this task, indicating a superficial understanding of aspects such as object attributes, object relationships, and action properties. This task tests the model's ability to grasp how actions affect objects, focusing on causal relationships and temporal reasoning—particularly for actions like "push" and "pull", which share similar execution flows. The model must distinguish them based on dynamic effects, such as changes in direction and speed, but most models perform moderately in this regard. + +Speed The results show that all models are insensitive to speed variations, likely because they were not adequately exposed to speed changes during training. Figure 4 shows that models are more sensitive to slow motion than fast playback, and struggled with identifying "normal speed" and "no speed", except for VideoLLaMA 2.1. This may be due to the loss of coherence in fast-moving video content, while slow-motion videos highlight more distinct details, aiding the model in making accurate judgments. + +![](images/000c48ff5f7a82a3b563edc36bee7663f09f35fa0b71d19da0b0c003be44159c.jpg) +Figure 4. Accuracy across different video speeds. All models are more sensitive to slow-speed videos and struggle to understand "normal speed" and "no speed", except for VideoLLaMA 2.1. + +
Video-LLMsActionFrame Number
w/o VSw/ VSAvg.345
LLaVA-NeXT-Video37.3135.0419.3720.3319.7717.98
MiniCPM-V 2.643.3740.1590.3293.8290.6686.44
Video-LLaMA 2.163.2653.9830.1742.8639.897.45
Qwen2-VL68.1856.6296.6597.2596.6396.05
ShareGPT4Video46.9030.8426.3360.9916.780.00
Video-CCAM73.1060.2323.4514.188.9647.61
+ +Table 2. Left: Accuracy of the Action task with or without "Visual Synonyms". It is obvious that the "Visual Synonyms" have significantly impacted the model's judgment. Right: Accuracy of the counting task across different frame counts. Except for Video-CCAM, all other models exhibited a decline in performance as the number of frames increased. + +
Video-LLMsParams.Scene-LevelFragment-LevelS-Avg.FG-Avg.A-Avg.
ActionEffectSpeedFCntMoOFCmpAoNRearr
(Random)-25.0025.0025.0025.0025.0033.3333.3325.0025.0028.3327.08
LLaVA-NeXT-Video7B37.3142.6722.3519.3724.0253.7575.4520.6734.1138.6536.95
MiniCPM-V 2.68B43.3752.5619.1390.3256.4275.6676.4918.0938.3563.4054.01
Video-LLaMA 2.17B63.2650.9219.8930.1742.2776.0189.9226.8744.6953.0549.91
Qwen2-VL7B68.1857.1424.6296.6533.3374.5390.7022.4849.9863.5458.45
ShareGPT4Video8B46.9043.8831.7626.3361.0588.4484.8023.3640.8557.1150.82
Video-CCAM9B73.1055.9031.6523.4545.6664.9590.2722.7253.5548.4750.96
+ +Table 3. The overall performances of notable Video-LLMs on FineVidBench. FCnt: Frame Count. MoO: Meaning of Order. Fcmp: Frame Comparison. AoN: Adjust or Not. Rearr: Rearrangement. S-Avg.: the average performance of scene-level tasks; FG-Avg.: the average performance of fragment-level tasks. A-Avg.: the average performance of all tasks. + +# - Fragment-level Results and Analysis + +(1) Frame-count accuracy varied significantly across models, with the lower-performing models likely lacking targeted training. The trend shown in the right side of Table 2, where accuracy decreases as frame count increases, highlights the models' insufficient temporal reasoning on longer sequences. (2) ShareGPT4Video and MiniCPM-V 2.6 showed better comprehension in the Meaning-of-Order task, while other models lagged, suggesting a lack of explicit focus on "order". (3) Most models excelled in frame comparison due to image-text alignment training. ShareGPT4Video achieved the best performance, owing to its Differential Sliding-Window Captioning (DiffSW) strategy, which emphasizes capturing the changes between frames when generating video descriptions. This also improved its Meaning-of-Order performance. (4) In the sorting task, models generally succeeded in the "Adjust or Not" response but performed poorly in the more complex "Rearrangement" task, indicating they can detect, but not correct, sequence errors. + +# 4. Self-supervised Fragment Finetuning + +The above benchmark results show the existing Video-LLMs generally fail to tackle fine-grained video understanding tasks. Videos often contain subtle, complex changes that natural language alone fails to fully capture. The core component of Video-LLMs, LLMs, as generalized pattern recognizers, offers a promising solution. LLMs have the potential to detect and interpret intricate spatiotemporal dynamics that were previously difficult to represent. Given that these changes cannot be directly annotated, using self-supervised learning naturally becomes the solution, bypassing the bottleneck of manual annotation and significantly re + +ducing labeling costs. Given these factors, we propose the $\mathrm{SF^2T}$ to fine-tune Video-LLMs. While we do not expect $\mathrm{SF^2T}$ to replace the supervised fine-tuning, instead it's an effortless complementary to SFT. Comparing $\mathrm{SF^2T}$ with SFT, they primarily differ in data construction and content focus level, with each method aligned with distinct training objectives as shown in Figure 5. + +# 4.1. SFT Tasks + +We first review the common SFT tasks to set a baseline for comparing our $\mathrm{SF^2T}$ + +General QA on Video Content This method focuses on understanding the main events and context of a video by directly asking questions about its content. While effective for grasping the video's key moments, it lacks finer spatiotemporal details and requires significant human effort to create standardized but constrained answers. + +Frame Description Integration This method typically samples video frames evenly, generates detailed descriptions for each, and integrates them into a cohesive but lengthy summary. While it enhances the model's understanding of continuity and micro-dynamics, it often proves incapable of capturing complex or subtle details that are beyond natural language's scope. Moreover, although frame descriptions can be generated using powerful multi-model LLMs like GPT-4o, significant human effort is still required to review the quality of the generated responses. + +# 4.2. Fragment-level Tasks of $\mathbf{SF}^2\mathbf{T}$ + +SFT tasks require manual annotations, and even automation annotation is labor-intensive and error-prone. To address, we introduce $\mathrm{SF}^2\mathrm{T}$ which generates accurate fragment-level labels accurately. $\mathrm{SF}^2\mathrm{T}$ comprises five tasks—Counting, Consistency Verification, Localization, Disorder Detection + +![](images/205a408f47737ab3e7ada0948b109312f60fa3951eebb56720eff570ee330e62.jpg) + +What is the main content of the video? + +The video shows a person bowling, including their four-step approach, the smooth release of the ball down the lane, its path toward the pins, and... + +![](images/d73b992c7705e58db518f61c8e87a993d684148c83ad06bf6ee75bc240b2dd85.jpg) + +What is the main content of the video? + +The video shows a person bowling: (Frame 1) The scene shows a bowling alley... (Frame 2) The player swing the bowling ball... (Frame 4) The bowling ball approaches the pins... (Frame 6) The bowling ball strikes the pins... (Frame 8) All the pins are down. + +![](images/67c18e2e5d3b5ddf1ade5d3208662e77bc4273c7f4af0846b550bfcea83f9e48.jpg) + +![](images/a2809af5779cd4055ae1b311000647b93fc824f660e342bc23cbfbcf66bbcf07.jpg) +Scene-Level Tasks + +![](images/c2db6c38cc0fd2073da14990dcbf1438642ff33fea8f737419c9f49191f9fe9e.jpg) + +How many frames? + +On which frames? + +Same frames? + +Adjust or not? + +Rearrange it. + +![](images/80f3bdf9d832b46ec11dfa8a148226f2604d3174b3dd3e36d58a372ea997442d.jpg) + +![](images/191add99094c9cf437d91ba98b8f1d1c1e3d0af2e8b4bb85365ec85bac2ba6ce.jpg) +Figure 5. Comparison between $\mathrm{SF}^2\mathrm{T}$ and SFT. SFT depends on manual and model-driven design to generate QA pairs for scene-level video understanding, $\mathrm{SF}^2\mathrm{T}$ , in contrast, automatically constructs training data based on pre-defined rules that cover various temporal and spatial aspects of the video. $\mathrm{SF}^2\mathrm{T}$ enables the model to focus on a fine-grained content analysis, and offering insights that supervised labels cannot achieve. + +![](images/6b29d67543db69e81306b3d01d8a56e2955a4bfc3bfa673551b6bfd0bcdb6a23.jpg) + +![](images/a37214fdf9867e16e0bccbe439bebc013a75c6b18bd07cb3822a9ba0d8ff0dd9.jpg) + +![](images/7895528dc17bddd0cdfd12a2b6f192d8f17af7af0313cc10e30ea23288f819c2.jpg) + +![](images/139eef899517a8522b5168019c8b5d3c8b9299d5d7c4b909bfa4288d42b9e477.jpg) + +2nd + +No + +Yes + +3412 + +# Fragment-Level Tasks + +and Rearrangement—designed to train the model to rearrange a set of out-of-order frames into their original sequence. This is a robust indicator of a modal's mastery over the visual dynamics of an action, requiring the model to detect subtle frame changes and understand the overall coherence and temporal trends. Mastery of these tasks enables the model to recognize frames and their temporal re + +relationships, enhancing its ability to predict and reconstruct action sequences and improving performance on more complex video tasks. Our method first extracts multiple sets of dynamic keyframes from each video. These fragments capture the key dynamic information from multiple temporal perspectives, offering a more efficient representation of redundant video data. It then applies pseudo-labeling, distinguishing it from traditional video-level labeling. By designing proxy tasks that leverage intrinsic information rather than predefined prior knowledge, it smartly circumvents the annotation bottleneck, enabling a deeper temporal understanding and offering insights that traditional video-level labeling cannot achieve. + +- Counting We input N frames into the Video-LLM and ask it to count them. Although this task seems straightforward, it proves challenging for current Video-LLMs, particularly as the number of frames increases, revealing a decline in accuracy. The model's inability to perform basic quantitative tasks points to a broader limitations in understanding the overall sequence integrity. + +Consistency Verification Video-LLMs are tasked with identifying two frames sampled from the same video, which may show subtle differences. This task sharpens the model's sensitivity to visual details by encouraging a thorough analysis and comparison of the images, countering its tendency to focus on primary subjects while neglecting the background and other subtle features. + +Localization Video-LLMs must accurately locate a specified target (from video metadata) within a sequence of frames, identifying the frames in which it appears, disappears, or persists. This naturally human ability is a significant challenge for these models, as they often struggle to perceive sequential relationships between frames and face additional obstacles, such as occlusion, interference from similar objects, lighting variations, and memory limitations. + +Disorder Detection and Rearrangement Video-LLMs must determine whether and how to adjust the order of a given frame sequence. When frames are randomized, the loss of spatiotemporal coherence and logical continuity makes it exceptionally challenging to reconstruct their original sequence, especially as interactions within frames become more complex [20]. This task is evaluated in two ways: the yes/no task tests the model's sensitivity to temporal consistency, while the sorting task, which leverages capabilities from the other four tasks, requires advanced reasoning and adjustments. + +# 5. Experiments + +In this section, we fine-tuned four of the most advanced open-source Video-LLMs using the $\mathrm{SF}^2\mathrm{T}$ method to evaluate its effectiveness, alongside ablation studies and interpretability analyses to explore the underlying mechanisms. + +
MethodsLLaVA-NEXT-VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2-VL
ActionEffectSpeedActionEffectSpeedActionEffectSpeedActionEffectSpeed
Base37.3142.6722.3543.3752.5619.1363.2650.9219.8968.1857.1424.62
Base+SF2T48.6743.7724.8365.9160.6228.6067.4257.3331.6373.8663.3731.92
Base(SFT)62.6944.6322.3577.6575.0970.8377.6565.9429.7378.6066.3030.87
Base(SFT)+SF2T63.0745.2432.0181.6376.9286.7479.7368.6831.8281.2573.2632.38
+ +Table 4. Performance on FineVidBench. We tested on two baselines: (1) Base: Results without any fine-tuning. (2) Base(SFT): Results after fine-tuning in supervised way. After $\mathrm{SF}^2\mathrm{T}$ , all models improved in all three tasks, highlighting its broad effectiveness and the value of fragment-level tasks in enhancing scene-level comprehension. Notably, $\mathrm{SF}^2\mathrm{T}$ outperformed SFT in the Speed task (except MiniCPM-V 2.6), highlighting the key role of fine-grained temporal understanding in distinguishing video speeds. + +
MethodsLLaVA-NeXT -VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2 -VL
MVBench
Base36.8440.2354.1855.97
Base+SF2T42.9256.0257.9763.76
Video-MME(no subtitle)
Base29.7643.1749.0243.77
Base+SF2T34.8453.1951.8853.60
MLVU
Base36.3241.5852.3242.81
Base+SF2T41.9155.3256.1154.67
+ +# 5.1. Implementation Details + +To ensure fairness, experiments were conducted on LoRA-compatible models, including LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4] and Qwen2-VL[28], using their default or recommended settings, with all models trained for one epoch. All experiments were performed under identical hardware conditions, utilizing NVIDIA A100 40GB GPU for computation. It should be emphasized that our goal is to validate the effectiveness of $\mathrm{SF}^2\mathrm{T}$ , not to optimize models for maximum performance. + +We randomly sampled videos from SSv2 and MiT for training, ensuring no overlap with the FineVidBench dataset. MGSampler [37] was used to extract N sets of M-frame sequences from each video, capturing dynamic changes while preserving overall characteristics. M is chosen based on the video's characteristics to capture content flow, while N is determined by content complexity, with more complex content requiring a larger N to cover more temporal perspectives. In this study, we set $\mathrm{N} = 3$ and M between 3 and 5, though these values may vary for other + +Table 5. Performance on public benchmarks. $\mathrm{SF}^2\mathrm{T}$ consistently enhances performance across all three benchmarks, reaffirming its effectiveness as a spatiotemporal enhancer. + +
Methodsrandomuniformkeyframemotion-salient
SF2T70.3171.6772.1173.86
+ +Table 6. Impact of sampling. As shown, motion-salient area sampling outperforms others by better capturing motion fluidity and temporal details, while the other methods fail to fully utilize their potential, leading to suboptimal performance. + +
Methodslongshortrandom
SF2T69.3871.4073.86
+ +Table 7. Impact of temporal span. Both long- and short-range temporal modeling reduced $\mathrm{SF}^2\mathrm{T}$ 's performance, emphasizing the importance of multi-scale temporal modeling. + +datasets. We then generated QA pairs for each frame sequence based on the five tasks defined in $\mathrm{SF^2T}$ for training. Evaluations were performed on FineVidBench's scene-level tasks, including Action, Effect, and Speed. To compare with traditional SFT, we also generated and manually reviewed QA pairs for these videos in a supervised setting. + +# 5.2. Comparisons + +Table 4 summarizes the results of the scene-level tasks. After $\mathrm{SF}^2\mathrm{T}$ training, all models showed significant improvement, emphasizing that fragment-level tasks can notably enhance scene-level comprehension. Integrating $\mathrm{SF}^2\mathrm{T}$ with SFT is also leads to performance gains, demonstrating that fragment-level training positively impacts SFT and enhances its effectiveness. Surprisingly, in the Speed task, many base models outperformed SFT after applying $\mathrm{SF}^2\mathrm{T}$ , highlighting the importance of fine-grained temporal understanding in distinguishing video speeds. This improvement likely stems from $\mathrm{SF}^2\mathrm{T}$ 's ability to enhance the model's sensitivity to temporal cues, such as the loss or enhancement of + +information during acceleration or deceleration, as well as content coherence—all crucial for speed judgment. As expected, $\mathrm{SF}^2\mathrm{T}$ currently lags behind SFT, since its training objective is not fully aligned with scene-level tasks. However, we do not expect $\mathrm{SF}^2\mathrm{T}$ to replace supervised fine-tuning; rather, our experiments suggest that it can serve as an effortless and effective complement to SFT. + +In addition to FineVidBench, we evaluated $\mathrm{SF}^{2}\mathrm{T}$ on three public video understanding benchmarks (Table 5). The results demonstrate consistent improvements across various video tasks, validating $\mathrm{SF}^{2}\mathrm{T}$ as an effective spatiotemporal enhancer for a wide range of video understanding tasks. All models were tested with an 8-frame input. + +# 5.3. Ablation and Interpretability Analyses + +We evaluated the impact of frame sampling strategies on $\mathrm{SF}^2\mathrm{T}$ , as each method provides a unique "temporal information perspective" that influencing video understanding performance. As shown in Table 6, we assessed four strategies on Qwen2-VL in the Action task: random, uniform interval, keyframe, and motion-salient area sampling [37]. Motion-salient area sampling performed best, likely due to its ability to capture continuous motion dynamics, thereby enhancing the model's understanding of action fluidity and temporal detail. In comparison, the other methods had limitations: keyframe sampling misses intermediate action phases, fixed-interval sampling may overlook critical moments, and random sampling lacks temporal consistency. Notably, different datasets may favor different strategies. For example, some datasets may perform better with uniform interval sampling, or their motion features may align better with the model's specific capabilities. + +We examined the effects of long- and short-range temporal modeling on $\mathrm{SF}^2\mathrm{T}$ . In the Consistency Verification task, we constrained the random selection of frame pairs to adjacent frames for local continuity or non-adjacent frames to capture long-range dependencies. As shown in Table 7, both settings decreased $\mathrm{SF}^2\mathrm{T}$ 's performance on the Action task of Qwen2-VL, indicating that an overemphasis on either long- or short-range information leads to temporal imbalance and incomplete dynamics. This underscores the importance of combining both approaches to leverage their broader temporal span and frame variations for a more comprehensive feature representation. + +We analyzed the attention map of Qwen2-VL on the Action task, particularly in cases where the model's predictions were corrected after $\mathrm{SF}^2\mathrm{T}$ . As shown in Figure 6, we found that $\mathrm{SF}^2\mathrm{T}$ enhances the model's ability to capture fine-grained spatial changes and temporal dynamics. (1) Spatial Aspects. After $\mathrm{SF}^2\mathrm{T}$ , the model shows increased attention to action execution areas, particularly the hands and objects they interact with. It shows better sensitivity to small targets, likely due to the Consistency Verification + +![](images/9553a2acfb87e33c118766225e4eb5707c664e1fa8e5864d9f59bea5dffc0cda.jpg) +Figure 6. Two exemplary visualizations of the attention map on Qwen2-VL. For each example: top - Original frames; middle - Base (SFT); bottom - $\mathrm{SF^2T}$ applied. As shown by the red boxes, after applying $\mathrm{SF^2T}$ , the model better focuses on action execution areas and interacting objects. The $\mathrm{SF^2T}$ fine-tuned model has the ability to predict the direction of motion, as seen in the trajectories of the red bottle and Cheerios. + +task, which enhances spatial perception by refining sensitivity to subtle image differences. (2) Temporal Aspects. After $\mathrm{SF}^2\mathrm{T}$ , we observed that the model can predict object movement trajectories in certain actions, indicating an advanced level of temporal understanding. This ability likely stems from the sorting task, which strengthens the model's comprehension of action flows and movement patterns. + +# 6. Conclusion + +In this work, we propose $\mathrm{SF}^2\mathrm{T}$ to overcome the limitations of Video-LLMs in fine-grained video understanding. $\mathrm{SF}^2\mathrm{T}$ is an innovative fine-tuning method that eliminates the need for labor-intensive annotations and effectively bypasses the constraints of natural language descriptions. Additionally, we introduce FineVidBench, a benchmark for evaluating Video-LLMs at both scene and fragment levels. In the future, we plan to expand our dataset with larger videos and more tasks to increase its impact. + +# Acknowledgments + +This work is supported by the National Key Research and Development Program of China (No.2020YBF2901202), National Natural Science Foundation of China (NSFC No. 62272184 and No. 62402189), the China Postdoctoral Science Foundation under Grant Number GZC20230894, the China Postdoctoral Science Foundation (Certificate Number: 2024M751012), and the Postdoctor Project of Hubei Province under Grant Number 2024HBBHCXB014, and the "Pioneer" and "Leading Goose" R&D Program of Zhejiang (No. 2024C01161). The computation is completed in the HPC Platform of Huazhong University of Science and Technology. + +# References + +[1] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2 +[2] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 3 +[3] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 2 +[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 2, 3, 7 +[5] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2 +[6] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3 +[7] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3 +[8] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[9] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: + +Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3, 7 +[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[11] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2 +[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2 +[13] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2 +[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1 +[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[16] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 2 +[17] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2 +[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2 +[19] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1 +[20] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 527-544. Springer, 2016. 6 +[21] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019. 2, 3 + +[22] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2 +[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1 +[24] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2 +[25] Fangxun Shu, Lei Zhang, Hao Jiang, and Cihang Xie. Audio-visual llm for video understanding. arXiv preprint arXiv:2312.06720, 2023. 2 +[26] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 1 +[27] TencentQQ Multimedia Research Team. Video-cam: Advancing video-language understanding with causal cross-attention masks. https://github.com/QQ-MM/Video-CCAM, 2024.4 +[28] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 7 +[29] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2 +[30] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1 +[31] Haiyang Xu, Qinghao Ye, Xuan Wu, Ming Yan, Yuan Miao, Jiabo Ye, Guohai Xu, Anwen Hu, Yaya Shi, Guangwei Xu, et al. Youku-mplug: A 10 million large-scale chinese video-language dataset for pre-training and benchmarks. arXiv preprint arXiv:2306.04362, 2023. 2 +[32] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 2 +[33] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2 + +[34] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 3, 7 +[35] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2 +[36] Zijia Zhao, Haoyu Lu, Yuqi Huo, Yifan Du, Tongtian Yue, Longteng Guo, Bingning Wang, Weipeng Chen, and Jing Liu. Needle in a video haystack: A scalable synthetic framework for benchmarking video mllms. arXiv preprint arXiv:2406.09367, 2024. 2 +[37] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 7, 8 +[38] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2 +[39] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2 \ No newline at end of file diff --git a/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/images.zip b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8e84750a7c7ac9f92a44e33e2dff57499989302f --- /dev/null +++ b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b99077d4e6086a40759befba95ad20c491f9c474864827b5fcf2cc8810f07e55 +size 573738 diff --git a/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/layout.json b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fa612eb5efe0572308493ccfda3c56c2826450a --- /dev/null +++ b/2025/SF2T_ Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding/layout.json @@ -0,0 +1,7566 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{SF}^{2}\\mathbf{T}" + }, + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "text", + "content": ": Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": "Yangliu Hu" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": ", Zikai Song" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": ", Na Feng" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": ", Yawei Luo" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": ", Junqing Yu" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": ", Yi-Ping Phoebe Chen" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": ", Wei Yang" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": "Huazhong University of Science and Technology " + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 160, + 546, + 193 + ], + "type": "text", + "content": "La Trobe University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 194, + 457, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 194, + 457, + 206 + ], + "spans": [ + { + "bbox": [ + 153, + 194, + 457, + 206 + ], + "type": "text", + "content": "{huyangliu,skyesong,fengna,yjqing,weiyangcs}@hust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 209, + 433, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 209, + 433, + 220 + ], + "spans": [ + { + "bbox": [ + 175, + 209, + 433, + 220 + ], + "type": "text", + "content": "yaweiluo@zju.edu.cn phoebe.chen@latrobe.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "type": "text", + "content": "Video-based Large Language Models (Video-LLMs) have witnessed substantial advancements in recent years, propelled by the advancement in multi-modal LLMs. Although these models have demonstrated proficiency in providing the overall description of videos, they struggle with fine-grained understanding, particularly in aspects such as visual dynamics and video details inquiries. To tackle these shortcomings, we find that fine-tuning Video-LLMs on self-supervised fragment tasks, greatly improve their fine-grained video understanding abilities. Hence we propose two key contributions: (1) Self-Supervised Fragment Fine-Tuning " + }, + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "type": "inline_equation", + "content": "(SF^2 T)" + }, + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "type": "text", + "content": ", a novel effortless fine-tuning method, employs the rich inherent characteristics of videos for training, while unlocking more fine-grained understanding ability of Video-LLMs. Moreover, it relieves researchers from labor-intensive annotations and smartly circumvents the limitations of natural language, which often fails to capture the complex spatiotemporal variations in videos; (2) A novel benchmark dataset, namely FineVidBench, for rigorously assessing Video-LLMs' performance at both the scene and fragment levels, offering a comprehensive evaluation of their capabilities. We assessed multiple models and validated the effectiveness of " + }, + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "type": "inline_equation", + "content": "SF^2 T" + }, + { + "bbox": [ + 54, + 285, + 297, + 583 + ], + "type": "text", + "content": " on them. Experimental results reveal that our approach improves their ability to capture and interpret spatiotemporal details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 135, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 135, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 135, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "Large Language Models (LLMs) have showcased significant emergent capabilities, such as in-context learning [19], instruction-following [23], and chain-of-thought reasoning [30], driven by expansive datasets and advanced model architectures. Extending these advancements, Video-LLMs through mechanisms like pooling or query aggregation" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 247, + 541, + 430 + ], + "blocks": [ + { + "bbox": [ + 328, + 247, + 541, + 430 + ], + "lines": [ + { + "bbox": [ + 328, + 247, + 541, + 430 + ], + "spans": [ + { + "bbox": [ + 328, + 247, + 541, + 430 + ], + "type": "image", + "image_path": "8b7eb9014d7abb5b36913b7f76bf78d0aca112be4c48c98f53b28ffaa2069a16.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "text", + "content": "Figure 1. Performance w/ and w/o " + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{SF}^2\\mathbf{T}" + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "text", + "content": ". We evaluated four advanced Video-LLMs w/ and w/o " + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "text", + "content": " on our proposed FineVidBench with two baselines: (1) Base: performance without any fine-tuning (blue dashed), and (2) Base (SFT): performance with supervised fine-tuning (red dashed). After applying " + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "text", + "content": ", all models showed significant improvements (solid blue and red), underscoring its broad effectiveness." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 542, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 626 + ], + "type": "text", + "content": "across numerous visual tokens, have broadened the scope of LLMs to encompass video information processing [11, 14, 35]. This evolution markedly advances their potential for in-depth real-world comprehension, opening applications in intelligent surveillance, virtual reality, and autonomous driving, further enriching the landscape of video analytics and interpretation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Various Video-LLMs, exemplified by GPT4-V, VideoLaMA 2 [4], MiniCPM-V [34], and Qwen2-VL [28], have been crafted by leading corporations and research institutions, demonstrating proficiency in capturing the overarching content of videos. When adapting to new videos and tasks, they predominantly rely on Supervised FineTuning (SFT) [26] or Reinforcement Learning from Hu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 703, + 156, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 703, + 156, + 713 + ], + "spans": [ + { + "bbox": [ + 74, + 703, + 156, + 713 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "man Feedback (RLHF) [39], both of which are heavily contingent upon extensive manual annotation. This dependence poses several key problems: (1) it necessitates substantial human resources, particularly highly trained annotators; (2) the inherent complexity of video content and task demands frequently introduces inconsistencies and subjectivity, rendering the maintenance of high-quality annotations particularly arduous; and (3) subtle temporal variations across video frames are challenging to articulate with precision, often yielding generalized descriptions that constrain the Video-LLMs' potential. Consequently, existing Video-LLMs struggle with fine-grained video understanding tasks, particularly in aspects such as visual dynamics (e.g., motion patterns, object interactions) and video details inquiries (e.g., positional changes, detail variations)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "type": "text", + "content": "To address these challenges, we observe that finetuning Video-LLMs with self-supervised fragment tasks, by \"fragment\" we mean temporal frame level specifications of the video, could improve the model's sensitivity to spatiotemporal scene-level details (related to video contents). Driven by this, we introduce the Self-supervised Fragment Fine-Tuning " + }, + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "type": "inline_equation", + "content": "(\\mathrm{SF}^2\\mathrm{T})" + }, + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "type": "text", + "content": ", a effortless fine-tuning strategy for Video-LLMs that help to improve the fine-grained video understanding. " + }, + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 251, + 295, + 430 + ], + "type": "text", + "content": " consists of five fragment-level tasks—Counting, Consistency Verification, Localization, Disorder Detection and Rearrangement—that automatically generate labels from various spatiotemporal perspectives. This approach maximizes the use of frame-level information while minimizing reliance on complex human instructions and annotations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "text", + "content": "Moreover, to evaluate the fine-grained visual dynamic perception of Video-LLMs and fully demonstrate the effectiveness of our " + }, + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "text", + "content": ", we present the FineVidBench, a novel benchmark. FineVidBench comprises 910 videos and 22,718 question-answer pairs, with videos sourced from diverse public datasets, including Something-Something V2 (SSv2) [6], Moments in Time (MiT) [21], etc. The question-answer pairs are auto-generated in single-choice format, incorporating distractors to increase testing difficulty. We evaluated several notable Video-LLMs developed in recent years, and find they generally fail to understand the execution sequence of actions and struggling to grasp fine-grained spatiotemporal information. While after fine-tuning with " + }, + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "text", + "content": ", the Video-LLMs better recognize spatiotemporal details, leading to a holistic and marked improvement in fine-grained understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 714 + ], + "type": "text", + "content": "Video-LLMs Finetuning Video-LLMs are primarily finetuned by adjusting the parameters of small, trainable adapters for task adaptation, without changing the entire model, saving resources and enhancing efficiency. The connective adapter (e.g., MLP/Linear Layer [15], Q" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "former [10]) links the Video Embedder and LLM, aligning video embeddings with LLM input tokens, while insertive adapters (e.g., LoRA [8]) are directly integrated into the LLM to modify its behavior. Most Video-LLMs combine both types of adapters and typically use multi-stage finetuning [4, 11, 13, 24, 35]. First, the model learns to establish relationships between images, videos, and text using large-scale multimodal datasets [1, 2, 29, 31]. In the second stage, the model is fine-tuned with an curated instruction-following dataset [11, 17, 18]. Besides, there are full finetuning, which updates all LLM parameters with a lower learning rate [25, 33], and zero-shot models, which transforms the video task into a text task, typically relying on a powerful LLM [32]. However, annotating video data remains a labor-intensive and time-consuming task, particularly for long videos or those involving complex actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 264, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 467 + ], + "type": "text", + "content": "Benchmarks on Video-LLMs Currently, many studies [3, 5, 38] focus on evaluating the temporal perception capabilities of Video-LLMs. MVBench [12] designs 20 tasks from temporal and spatial perspectives, and Tempcompass [16] introduces 5 temporal aspects and 4 task formats. VN-Bench [36] decouples video content from the QA pairs by inserting irrelevant images or text \"needles\" into the original video. Moment-10M [22] has constructed a large-scale dataset on temporal localization tasks. However, as illustrated in Table 1, these studies often focus on gathering diverse videos or evaluating the models' performance with long videos, while somewhat neglecting the models' ability to perform fine-grained perception of temporal details. To address this gap, FineVidBench breaks videos into multiple sets of frames and generates annotations from diverse spatiotemporal perspectives, introducing novel evaluation methods for fine-grained understanding." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 475, + 553, + 596 + ], + "blocks": [ + { + "bbox": [ + 316, + 475, + 553, + 596 + ], + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 596 + ], + "type": "table", + "html": "
BenchmarksVideo num.QA num.Input ChangeTemporal DiversityFine-Grained EvaluationHierarchical Test
Video-MME9002700XXXX
TempCompass4107540X✓✓X
VN bench-1350X✓✓X
Moment-10M64.9k10.4MXXXX
AutoEval-Video327327XXXX
MV bench36414000XX✓X
MLVU13342593XXXX
FineVidBench91022,718✓✓✓✓
", + "image_path": "8fc059292041f5e55d6b484db9afa9eca18e426fc41d85f00a5152e809657be8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 604, + 553, + 638 + ], + "lines": [ + { + "bbox": [ + 313, + 604, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 553, + 638 + ], + "type": "text", + "content": "Table 1. Comparison with related benchmarks. Our approach offers significant advantages in input formats, evaluation methods, granularity, and temporal diversity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 657, + 465, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 465, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 465, + 670 + ], + "type": "text", + "content": "3. FineVidBench Benchmark" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "It is broadly recognized that Video-LLMs struggle with fine-grained video understanding tasks, yet no comprehensive benchmarks exist to thoroughly investigate this issue." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29109" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "To address this gap, we introduce FineVidBench, a multidimensional, fine-grained evaluation framework specifically designed to assess and improve the overall capabilities of Video-LLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 141, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 141, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 141, + 141 + ], + "type": "text", + "content": "3.1. Construction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 295, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 206 + ], + "type": "text", + "content": "Data collection We selected videos from various public datasets, including SS-v2 [6], MiT [21], and Ego4D [7], with a particular emphasis on temporally-sensitive content, to focus the model on the entire video sequence rather than individual frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "text", + "content": "Action categorization As shown in Figure 2, we compiled 52 actions, categorizing them into 3 types based on intraclass variance. The distribution varies significantly: \"Distinctive Actions\" " + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "inline_equation", + "content": "(39\\%)" + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "text", + "content": " are easily recognizable, encompassing a total of 36 actions. \"Non-typical Actions\" " + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "inline_equation", + "content": "(57\\%)" + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "text", + "content": " refer to flexible actions with no clear defining characteristics, spanning 14 types. The broad diversity and complexity in this category require more extensive video coverage to adequately capture the range of expressions and variations. \"Slight Movements\" " + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "inline_equation", + "content": "(4\\%)" + }, + { + "bbox": [ + 55, + 207, + 295, + 351 + ], + "type": "text", + "content": " represent subtle actions, such as \"hold\" and \"show\", which are difficult to detect with the naked eye and constitute a small proportion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": "Data augmentation The original videos were augmented using frame interpolation and skipping techniques for speed transformation, along with a motion-salient area sampling algorithm to capture dynamic motion. This process generated speed-varied versions and multiple sets of keyframes for each video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 533 + ], + "type": "text", + "content": "Statistics With our augmentation strategy, FineVidBench includes 910 videos, 1,820 speed-variant videos, and 2,670 sets of keyframes enriched with dynamic visual information. Building on this, we generated 22,718 QA pairs from the video content through a combination of automated processes and manual review. The quality assurance process involved rigorous cross-verification, where reviewers checked each QA pair for accuracy and contextual relevance, making corrections to ensure high quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 205, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 205, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 205, + 553 + ], + "type": "text", + "content": "3.2. Benchmarking Dimensions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": "As shown in Figure 3, FineVidBench encompasses both scene-level and fragment-level evaluations. The scene-level evaluation assesses both original and speed-adjusted videos across three dimensions: (1) Action, which evaluates the model's holistic understanding of video content. To increase difficulty, \"Visual Synonyms\" are added as distractors, requiring VideoLLM to distinguish visually similar actions with subtle differences, a challenge common in real-world scenarios. (2) Effect, which focuses on the model's comprehension of the visual changes resulting from actions. This understanding is essential for revealing object properties and interpreting complex dynamic scenes, and could significantly enhance the reasoning capabilities of Video" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 75, + 536, + 281 + ], + "blocks": [ + { + "bbox": [ + 335, + 75, + 536, + 281 + ], + "lines": [ + { + "bbox": [ + 335, + 75, + 536, + 281 + ], + "spans": [ + { + "bbox": [ + 335, + 75, + 536, + 281 + ], + "type": "image", + "image_path": "8e219e19edad16fd218dca29a3cb48b7d80101a471ef88867ad2f65e384007a1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 291, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 348 + ], + "type": "text", + "content": "Figure 2. We show the action semantics and their respective proportions in FineVidBench. Distinctive Action: easily recognizable actions. Non-typical Action: flexible actions with no clear characteristics, like \"put\" and \"move.\" Slight Movement: subtle actions, such as \"hold\" and \"show,\" difficult to detect with the naked eye." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 366, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 437 + ], + "type": "text", + "content": "LLMs and LLM-aided agents. (3) Speed, which tests the model's sensitivity to changes in video speed and its capability to maintain consistent understanding across varying speeds, with slow motion revealing hidden details and fast motion obscuring them. This capability is crucial for optimizing the model's performance across diverse scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 654 + ], + "type": "text", + "content": "For fragment-level evaluation, We've designed a structured evaluation format for video dynamic keyframes, employing a step-by-step inquiry framework: (1) Frame Count: Models are queried on the number of frames in sequences using dynamically refined keyframes to assess counting accuracy. (2) Meaning of Order: Understanding of sequence order is tested by asking about the first or last frames the targets appear in, or the frames they are present. e.g., \"At which frame does the target object first appear?\". (3) Frame Comparison: Two frames are randomly selected from the sequence for visual comparison, with differences varying in size but generally staying within human visual comfort limits. (4) Adjust-or-Not and Rearrangement: These two tasks involve a shuffled sequence of keyframes, and the model is asked to determine whether the order needs adjustment and, if so, how to correct it. They evaluate the model's ability to understand and restore the video's temporal sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 660, + 430, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 660, + 430, + 672 + ], + "spans": [ + { + "bbox": [ + 314, + 660, + 430, + 672 + ], + "type": "text", + "content": "3.3. Benchmark Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We evaluated six of the most advanced open-source models: LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4], Qwen2-VL[28], ShareGPT4Video [2] and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29110" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 261, + 240 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 261, + 240 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 261, + 240 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 261, + 240 + ], + "type": "image", + "image_path": "f2c27bd209bc48a9f381966094271a4758a614629c12b08a213350398427d324.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 248, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. FineVidBench evaluates videos augmented with speed variations and fragments. Scene-level tests include the following: Action: Tests recognition accuracy amidst distractors like \"Visual Synonyms\". Effect: Assesses the model's ability to identify pre- and post-action changes. Speed: Measures the model's sensitivity to changes in video speed. Fragment-level tests, employing a step-by-step inquiry framework, focus on challenges such as Frame Count, Meaning of Order, Frame Comparison, Adjust-or-Not and Rearrangement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 267, + 70, + 403, + 239 + ], + "blocks": [ + { + "bbox": [ + 267, + 70, + 403, + 239 + ], + "lines": [ + { + "bbox": [ + 267, + 70, + 403, + 239 + ], + "spans": [ + { + "bbox": [ + 267, + 70, + 403, + 239 + ], + "type": "image", + "image_path": "9726bd792966151b9e71d0c4fb9e8145f434a5d37572a9bba433c9ff69e6d8e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 408, + 89, + 542, + 238 + ], + "blocks": [ + { + "bbox": [ + 408, + 89, + 542, + 238 + ], + "lines": [ + { + "bbox": [ + 408, + 89, + 542, + 238 + ], + "spans": [ + { + "bbox": [ + 408, + 89, + 542, + 238 + ], + "type": "image", + "image_path": "56d0a2df8e3d8c352b70564101947f32b8297d02ddf5616cb21cf61904f974f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 314, + 295, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 362 + ], + "type": "text", + "content": "Video-CCAM [27], each employing different architectures and training strategies. Table 3 summarizes the results across the eight tasks. We discuss the results from scene-level and fragment-level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 204, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 204, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 204, + 384 + ], + "type": "text", + "content": "- Scene-level Results and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "content": "Action The scores for this task varied significantly, with models trained in relevant video data—such as Video-CCAM, Qwen2-VL, and VideoLLaMA 2.1—achieving notably higher performance. However, as shown on the left side of Table 2, interference from \"Visual Synonyms\" prevented these models from achieving their full potential, resulting in declines of varying degrees and indicating difficulties in distinguishing visually similar actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 485, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 605 + ], + "type": "text", + "content": "Effect All models exhibited average performance on this task, indicating a superficial understanding of aspects such as object attributes, object relationships, and action properties. This task tests the model's ability to grasp how actions affect objects, focusing on causal relationships and temporal reasoning—particularly for actions like \"push\" and \"pull\", which share similar execution flows. The model must distinguish them based on dynamic effects, such as changes in direction and speed, but most models perform moderately in this regard." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Speed The results show that all models are insensitive to speed variations, likely because they were not adequately exposed to speed changes during training. Figure 4 shows that models are more sensitive to slow motion than fast playback, and struggled with identifying \"normal speed\" and \"no speed\", except for VideoLLaMA 2.1. This may be due to the loss of coherence in fast-moving video content, while slow-motion videos highlight more distinct details, aiding the model in making accurate judgments." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 334, + 314, + 536, + 462 + ], + "blocks": [ + { + "bbox": [ + 334, + 314, + 536, + 462 + ], + "lines": [ + { + "bbox": [ + 334, + 314, + 536, + 462 + ], + "spans": [ + { + "bbox": [ + 334, + 314, + 536, + 462 + ], + "type": "image", + "image_path": "000c48ff5f7a82a3b563edc36bee7663f09f35fa0b71d19da0b0c003be44159c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 470, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 504 + ], + "type": "text", + "content": "Figure 4. Accuracy across different video speeds. All models are more sensitive to slow-speed videos and struggle to understand \"normal speed\" and \"no speed\", except for VideoLLaMA 2.1." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 512, + 561, + 635 + ], + "blocks": [ + { + "bbox": [ + 315, + 512, + 561, + 635 + ], + "lines": [ + { + "bbox": [ + 315, + 512, + 561, + 635 + ], + "spans": [ + { + "bbox": [ + 315, + 512, + 561, + 635 + ], + "type": "table", + "html": "
Video-LLMsActionFrame Number
w/o VSw/ VSAvg.345
LLaVA-NeXT-Video37.3135.0419.3720.3319.7717.98
MiniCPM-V 2.643.3740.1590.3293.8290.6686.44
Video-LLaMA 2.163.2653.9830.1742.8639.897.45
Qwen2-VL68.1856.6296.6597.2596.6396.05
ShareGPT4Video46.9030.8426.3360.9916.780.00
Video-CCAM73.1060.2323.4514.188.9647.61
", + "image_path": "8b9fbb96f8d43c7c9f39a2aae7856e3c87004d890d6039e0136f60e1d2523d08.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 643, + 555, + 708 + ], + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 708 + ], + "type": "text", + "content": "Table 2. Left: Accuracy of the Action task with or without \"Visual Synonyms\". It is obvious that the \"Visual Synonyms\" have significantly impacted the model's judgment. Right: Accuracy of the counting task across different frame counts. Except for Video-CCAM, all other models exhibited a decline in performance as the number of frames increased." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29111" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 240 + ], + "type": "table", + "html": "
Video-LLMsParams.Scene-LevelFragment-LevelS-Avg.FG-Avg.A-Avg.
ActionEffectSpeedFCntMoOFCmpAoNRearr
(Random)-25.0025.0025.0025.0025.0033.3333.3325.0025.0028.3327.08
LLaVA-NeXT-Video7B37.3142.6722.3519.3724.0253.7575.4520.6734.1138.6536.95
MiniCPM-V 2.68B43.3752.5619.1390.3256.4275.6676.4918.0938.3563.4054.01
Video-LLaMA 2.17B63.2650.9219.8930.1742.2776.0189.9226.8744.6953.0549.91
Qwen2-VL7B68.1857.1424.6296.6533.3374.5390.7022.4849.9863.5458.45
ShareGPT4Video8B46.9043.8831.7626.3361.0588.4484.8023.3640.8557.1150.82
Video-CCAM9B73.1055.9031.6523.4545.6664.9590.2722.7253.5548.4750.96
", + "image_path": "a26750f6f0df52394cea209edd778f9d22347560d0ea6525b5d50ed1e0b86fbb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 248, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 555, + 283 + ], + "type": "text", + "content": "Table 3. The overall performances of notable Video-LLMs on FineVidBench. FCnt: Frame Count. MoO: Meaning of Order. Fcmp: Frame Comparison. AoN: Adjust or Not. Rearr: Rearrangement. S-Avg.: the average performance of scene-level tasks; FG-Avg.: the average performance of fragment-level tasks. A-Avg.: the average performance of all tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 303, + 222, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 222, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 222, + 316 + ], + "type": "text", + "content": "- Fragment-level Results and Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 319, + 297, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 297, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 297, + 548 + ], + "type": "text", + "content": "(1) Frame-count accuracy varied significantly across models, with the lower-performing models likely lacking targeted training. The trend shown in the right side of Table 2, where accuracy decreases as frame count increases, highlights the models' insufficient temporal reasoning on longer sequences. (2) ShareGPT4Video and MiniCPM-V 2.6 showed better comprehension in the Meaning-of-Order task, while other models lagged, suggesting a lack of explicit focus on \"order\". (3) Most models excelled in frame comparison due to image-text alignment training. ShareGPT4Video achieved the best performance, owing to its Differential Sliding-Window Captioning (DiffSW) strategy, which emphasizes capturing the changes between frames when generating video descriptions. This also improved its Meaning-of-Order performance. (4) In the sorting task, models generally succeeded in the \"Adjust or Not\" response but performed poorly in the more complex \"Rearrangement\" task, indicating they can detect, but not correct, sequence errors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 560, + 261, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 261, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 261, + 574 + ], + "type": "text", + "content": "4. Self-supervised Fragment Finetuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "The above benchmark results show the existing Video-LLMs generally fail to tackle fine-grained video understanding tasks. Videos often contain subtle, complex changes that natural language alone fails to fully capture. The core component of Video-LLMs, LLMs, as generalized pattern recognizers, offers a promising solution. LLMs have the potential to detect and interpret intricate spatiotemporal dynamics that were previously difficult to represent. Given that these changes cannot be directly annotated, using self-supervised learning naturally becomes the solution, bypassing the bottleneck of manual annotation and significantly re" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "text", + "content": "ducing labeling costs. Given these factors, we propose the " + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "text", + "content": " to fine-tune Video-LLMs. While we do not expect " + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "text", + "content": " to replace the supervised fine-tuning, instead it's an effortless complementary to SFT. Comparing " + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 303, + 555, + 389 + ], + "type": "text", + "content": " with SFT, they primarily differ in data construction and content focus level, with each method aligned with distinct training objectives as shown in Figure 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 395, + 387, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 387, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 387, + 406 + ], + "type": "text", + "content": "4.1. SFT Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 412, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 436 + ], + "type": "text", + "content": "We first review the common SFT tasks to set a baseline for comparing our " + }, + { + "bbox": [ + 313, + 412, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 508 + ], + "type": "text", + "content": "General QA on Video Content This method focuses on understanding the main events and context of a video by directly asking questions about its content. While effective for grasping the video's key moments, it lacks finer spatiotemporal details and requires significant human effort to create standardized but constrained answers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 629 + ], + "type": "text", + "content": "Frame Description Integration This method typically samples video frames evenly, generates detailed descriptions for each, and integrates them into a cohesive but lengthy summary. While it enhances the model's understanding of continuity and micro-dynamics, it often proves incapable of capturing complex or subtle details that are beyond natural language's scope. Moreover, although frame descriptions can be generated using powerful multi-model LLMs like GPT-4o, significant human effort is still required to review the quality of the generated responses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 477, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 477, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 477, + 649 + ], + "type": "text", + "content": "4.2. Fragment-level Tasks of " + }, + { + "bbox": [ + 313, + 635, + 477, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{SF}^2\\mathbf{T}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "SFT tasks require manual annotations, and even automation annotation is labor-intensive and error-prone. To address, we introduce " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " which generates accurate fragment-level labels accurately. " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " comprises five tasks—Counting, Consistency Verification, Localization, Disorder Detection" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29112" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 292, + 118 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 292, + 118 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 292, + 118 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 292, + 118 + ], + "type": "image", + "image_path": "205a408f47737ab3e7ada0948b109312f60fa3951eebb56720eff570ee330e62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 128, + 235, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 128, + 235, + 139 + ], + "spans": [ + { + "bbox": [ + 74, + 128, + 235, + 139 + ], + "type": "text", + "content": "What is the main content of the video?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 90, + 143, + 259, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 143, + 259, + 168 + ], + "spans": [ + { + "bbox": [ + 90, + 143, + 259, + 168 + ], + "type": "text", + "content": "The video shows a person bowling, including their four-step approach, the smooth release of the ball down the lane, its path toward the pins, and..." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 182, + 291, + 234 + ], + "blocks": [ + { + "bbox": [ + 61, + 182, + 291, + 234 + ], + "lines": [ + { + "bbox": [ + 61, + 182, + 291, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 291, + 234 + ], + "type": "image", + "image_path": "d73b992c7705e58db518f61c8e87a993d684148c83ad06bf6ee75bc240b2dd85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 242, + 236, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 242, + 236, + 253 + ], + "spans": [ + { + "bbox": [ + 74, + 242, + 236, + 253 + ], + "type": "text", + "content": "What is the main content of the video?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 257, + 260, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 257, + 260, + 299 + ], + "spans": [ + { + "bbox": [ + 91, + 257, + 260, + 299 + ], + "type": "text", + "content": "The video shows a person bowling: (Frame 1) The scene shows a bowling alley... (Frame 2) The player swing the bowling ball... (Frame 4) The bowling ball approaches the pins... (Frame 6) The bowling ball strikes the pins... (Frame 8) All the pins are down." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 263, + 253, + 274, + 266 + ], + "blocks": [ + { + "bbox": [ + 263, + 253, + 274, + 266 + ], + "lines": [ + { + "bbox": [ + 263, + 253, + 274, + 266 + ], + "spans": [ + { + "bbox": [ + 263, + 253, + 274, + 266 + ], + "type": "image", + "image_path": "67c18e2e5d3b5ddf1ade5d3208662e77bc4273c7f4af0846b550bfcea83f9e48.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 77, + 332, + 272, + 390 + ], + "blocks": [ + { + "bbox": [ + 131, + 312, + 220, + 323 + ], + "lines": [ + { + "bbox": [ + 131, + 312, + 220, + 323 + ], + "spans": [ + { + "bbox": [ + 131, + 312, + 220, + 323 + ], + "type": "text", + "content": "Scene-Level Tasks" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 332, + 272, + 390 + ], + "lines": [ + { + "bbox": [ + 77, + 332, + 272, + 390 + ], + "spans": [ + { + "bbox": [ + 77, + 332, + 272, + 390 + ], + "type": "image", + "image_path": "a2809af5779cd4055ae1b311000647b93fc824f660e342bc23cbfbcf66bbcf07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 75, + 400, + 88, + 413 + ], + "blocks": [ + { + "bbox": [ + 75, + 400, + 88, + 413 + ], + "lines": [ + { + "bbox": [ + 75, + 400, + 88, + 413 + ], + "spans": [ + { + "bbox": [ + 75, + 400, + 88, + 413 + ], + "type": "image", + "image_path": "c2db6c38cc0fd2073da14990dcbf1438642ff33fea8f737419c9f49191f9fe9e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 93, + 404, + 164, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 404, + 164, + 415 + ], + "spans": [ + { + "bbox": [ + 93, + 404, + 164, + 415 + ], + "type": "text", + "content": "How many frames?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 92, + 420, + 161, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 420, + 161, + 429 + ], + "spans": [ + { + "bbox": [ + 92, + 420, + 161, + 429 + ], + "type": "text", + "content": "On which frames?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 92, + 435, + 148, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 435, + 148, + 444 + ], + "spans": [ + { + "bbox": [ + 92, + 435, + 148, + 444 + ], + "type": "text", + "content": "Same frames?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 92, + 460, + 145, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 460, + 145, + 469 + ], + "spans": [ + { + "bbox": [ + 92, + 460, + 145, + 469 + ], + "type": "text", + "content": "Adjust or not?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 92, + 475, + 141, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 475, + 141, + 483 + ], + "spans": [ + { + "bbox": [ + 92, + 475, + 141, + 483 + ], + "type": "text", + "content": "Rearrange it." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 175, + 402, + 227, + 415 + ], + "blocks": [ + { + "bbox": [ + 175, + 402, + 227, + 415 + ], + "lines": [ + { + "bbox": [ + 175, + 402, + 227, + 415 + ], + "spans": [ + { + "bbox": [ + 175, + 402, + 227, + 415 + ], + "type": "image", + "image_path": "80f3bdf9d832b46ec11dfa8a148226f2604d3174b3dd3e36d58a372ea997442d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 176, + 419, + 224, + 429 + ], + "blocks": [ + { + "bbox": [ + 176, + 419, + 224, + 429 + ], + "lines": [ + { + "bbox": [ + 176, + 419, + 224, + 429 + ], + "spans": [ + { + "bbox": [ + 176, + 419, + 224, + 429 + ], + "type": "image", + "image_path": "191add99094c9cf437d91ba98b8f1d1c1e3d0af2e8b4bb85365ec85bac2ba6ce.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "text", + "content": "Figure 5. Comparison between " + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "text", + "content": " and SFT. SFT depends on manual and model-driven design to generate QA pairs for scene-level video understanding, " + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "text", + "content": ", in contrast, automatically constructs training data based on pre-defined rules that cover various temporal and spatial aspects of the video. " + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 525, + 295, + 602 + ], + "type": "text", + "content": " enables the model to focus on a fine-grained content analysis, and offering insights that supervised labels cannot achieve." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 176, + 434, + 224, + 444 + ], + "blocks": [ + { + "bbox": [ + 176, + 434, + 224, + 444 + ], + "lines": [ + { + "bbox": [ + 176, + 434, + 224, + 444 + ], + "spans": [ + { + "bbox": [ + 176, + 434, + 224, + 444 + ], + "type": "image", + "image_path": "6b29d67543db69e81306b3d01d8a56e2955a4bfc3bfa673551b6bfd0bcdb6a23.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 176, + 450, + 224, + 472 + ], + "blocks": [ + { + "bbox": [ + 176, + 450, + 224, + 472 + ], + "lines": [ + { + "bbox": [ + 176, + 450, + 224, + 472 + ], + "spans": [ + { + "bbox": [ + 176, + 450, + 224, + 472 + ], + "type": "image", + "image_path": "a37214fdf9867e16e0bccbe439bebc013a75c6b18bd07cb3822a9ba0d8ff0dd9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 177, + 476, + 224, + 490 + ], + "blocks": [ + { + "bbox": [ + 177, + 476, + 224, + 490 + ], + "lines": [ + { + "bbox": [ + 177, + 476, + 224, + 490 + ], + "spans": [ + { + "bbox": [ + 177, + 476, + 224, + 490 + ], + "type": "image", + "image_path": "7895528dc17bddd0cdfd12a2b6f192d8f17af7af0313cc10e30ea23288f819c2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 251, + 403, + 274, + 415 + ], + "blocks": [ + { + "bbox": [ + 251, + 403, + 274, + 415 + ], + "lines": [ + { + "bbox": [ + 251, + 403, + 274, + 415 + ], + "spans": [ + { + "bbox": [ + 251, + 403, + 274, + 415 + ], + "type": "image", + "image_path": "139eef899517a8522b5168019c8b5d3c8b9299d5d7c4b909bfa4288d42b9e477.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 242, + 420, + 257, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 420, + 257, + 429 + ], + "spans": [ + { + "bbox": [ + 242, + 420, + 257, + 429 + ], + "type": "text", + "content": "2nd" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 245, + 436, + 257, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 436, + 257, + 444 + ], + "spans": [ + { + "bbox": [ + 245, + 436, + 257, + 444 + ], + "type": "text", + "content": "No" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 238, + 460, + 257, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 460, + 257, + 468 + ], + "spans": [ + { + "bbox": [ + 238, + 460, + 257, + 468 + ], + "type": "text", + "content": "Yes" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 237, + 475, + 257, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 475, + 257, + 483 + ], + "spans": [ + { + "bbox": [ + 237, + 475, + 257, + 483 + ], + "type": "text", + "content": "3412" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 123, + 506, + 228, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 506, + 228, + 517 + ], + "spans": [ + { + "bbox": [ + 123, + 506, + 228, + 517 + ], + "type": "text", + "content": "Fragment-Level Tasks" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "and Rearrangement—designed to train the model to rearrange a set of out-of-order frames into their original sequence. This is a robust indicator of a modal's mastery over the visual dynamics of an action, requiring the model to detect subtle frame changes and understand the overall coherence and temporal trends. Mastery of these tasks enables the model to recognize frames and their temporal re" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "relationships, enhancing its ability to predict and reconstruct action sequences and improving performance on more complex video tasks. Our method first extracts multiple sets of dynamic keyframes from each video. These fragments capture the key dynamic information from multiple temporal perspectives, offering a more efficient representation of redundant video data. It then applies pseudo-labeling, distinguishing it from traditional video-level labeling. By designing proxy tasks that leverage intrinsic information rather than predefined prior knowledge, it smartly circumvents the annotation bottleneck, enabling a deeper temporal understanding and offering insights that traditional video-level labeling cannot achieve." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 229, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 312 + ], + "type": "text", + "content": "- Counting We input N frames into the Video-LLM and ask it to count them. Although this task seems straightforward, it proves challenging for current Video-LLMs, particularly as the number of frames increases, revealing a decline in accuracy. The model's inability to perform basic quantitative tasks points to a broader limitations in understanding the overall sequence integrity." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 314, + 553, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 397 + ], + "type": "text", + "content": "Consistency Verification Video-LLMs are tasked with identifying two frames sampled from the same video, which may show subtle differences. This task sharpens the model's sensitivity to visual details by encouraging a thorough analysis and comparison of the images, countering its tendency to focus on primary subjects while neglecting the background and other subtle features." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 399, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 495 + ], + "type": "text", + "content": "Localization Video-LLMs must accurately locate a specified target (from video metadata) within a sequence of frames, identifying the frames in which it appears, disappears, or persists. This naturally human ability is a significant challenge for these models, as they often struggle to perceive sequential relationships between frames and face additional obstacles, such as occlusion, interference from similar objects, lighting variations, and memory limitations." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 497, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 628 + ], + "type": "text", + "content": "Disorder Detection and Rearrangement Video-LLMs must determine whether and how to adjust the order of a given frame sequence. When frames are randomized, the loss of spatiotemporal coherence and logical continuity makes it exceptionally challenging to reconstruct their original sequence, especially as interactions within frames become more complex [20]. This task is evaluated in two ways: the yes/no task tests the model's sensitivity to temporal consistency, while the sorting task, which leverages capabilities from the other four tasks, requires advanced reasoning and adjustments." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 644, + 394, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 394, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 394, + 657 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "In this section, we fine-tuned four of the most advanced open-source Video-LLMs using the " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " method to evaluate its effectiveness, alongside ablation studies and interpretability analyses to explore the underlying mechanisms." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29113" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 531, + 190 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 190 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 190 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 190 + ], + "type": "table", + "html": "
MethodsLLaVA-NEXT-VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2-VL
ActionEffectSpeedActionEffectSpeedActionEffectSpeedActionEffectSpeed
Base37.3142.6722.3543.3752.5619.1363.2650.9219.8968.1857.1424.62
Base+SF2T48.6743.7724.8365.9160.6228.6067.4257.3331.6373.8663.3731.92
Base(SFT)62.6944.6322.3577.6575.0970.8377.6565.9429.7378.6066.3030.87
Base(SFT)+SF2T63.0745.2432.0181.6376.9286.7479.7368.6831.8281.2573.2632.38
", + "image_path": "78f02b4b1443fcbd31810d2cb4c2be8b80bd146e6b513946e6cd4650dd9034d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 252, + 290, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "text", + "content": "Table 4. Performance on FineVidBench. We tested on two baselines: (1) Base: Results without any fine-tuning. (2) Base(SFT): Results after fine-tuning in supervised way. After " + }, + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "text", + "content": ", all models improved in all three tasks, highlighting its broad effectiveness and the value of fragment-level tasks in enhancing scene-level comprehension. Notably, " + }, + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "text", + "content": " outperformed SFT in the Speed task (except MiniCPM-V 2.6), highlighting the key role of fine-grained temporal understanding in distinguishing video speeds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 252, + 290, + 401 + ], + "lines": [ + { + "bbox": [ + 63, + 252, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 63, + 252, + 290, + 401 + ], + "type": "table", + "html": "
MethodsLLaVA-NeXT -VideoMiniCPM-V 2.6VideoLLaMA 2.1Qwen2 -VL
MVBench
Base36.8440.2354.1855.97
Base+SF2T42.9256.0257.9763.76
Video-MME(no subtitle)
Base29.7643.1749.0243.77
Base+SF2T34.8453.1951.8853.60
MLVU
Base36.3241.5852.3242.81
Base+SF2T41.9155.3256.1154.67
", + "image_path": "43d552bf5399c00a0569cb06ec2e9d002d3107cd4d82b287dea20b52b8ce81dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 188, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 188, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 188, + 475 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": "To ensure fairness, experiments were conducted on LoRA-compatible models, including LLaVA-NeXT-Video[9], MiniCPM-V 2.6[34], VideoLLaMA 2.1[4] and Qwen2-VL[28], using their default or recommended settings, with all models trained for one epoch. All experiments were performed under identical hardware conditions, utilizing NVIDIA A100 40GB GPU for computation. It should be emphasized that our goal is to validate the effectiveness of " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": ", not to optimize models for maximum performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "We randomly sampled videos from SSv2 and MiT for training, ensuring no overlap with the FineVidBench dataset. MGSampler [37] was used to extract N sets of M-frame sequences from each video, capturing dynamic changes while preserving overall characteristics. M is chosen based on the video's characteristics to capture content flow, while N is determined by content complexity, with more complex content requiring a larger N to cover more temporal perspectives. In this study, we set " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{N} = 3" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and M between 3 and 5, though these values may vary for other" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 331, + 256, + 537, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 408, + 297, + 443 + ], + "lines": [ + { + "bbox": [ + 55, + 408, + 297, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 297, + 443 + ], + "type": "text", + "content": "Table 5. Performance on public benchmarks. " + }, + { + "bbox": [ + 55, + 408, + 297, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 408, + 297, + 443 + ], + "type": "text", + "content": " consistently enhances performance across all three benchmarks, reaffirming its effectiveness as a spatiotemporal enhancer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 256, + 537, + 297 + ], + "lines": [ + { + "bbox": [ + 331, + 256, + 537, + 297 + ], + "spans": [ + { + "bbox": [ + 331, + 256, + 537, + 297 + ], + "type": "table", + "html": "
Methodsrandomuniformkeyframemotion-salient
SF2T70.3171.6772.1173.86
", + "image_path": "da236d8ee06ac6dc6f73c71bbc3751866fd6aa2481f436ae04e6af2b84ce1c1f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 370, + 356, + 499, + 397 + ], + "blocks": [ + { + "bbox": [ + 312, + 306, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 312, + 306, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 312, + 306, + 555, + 350 + ], + "type": "text", + "content": "Table 6. Impact of sampling. As shown, motion-salient area sampling outperforms others by better capturing motion fluidity and temporal details, while the other methods fail to fully utilize their potential, leading to suboptimal performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 370, + 356, + 499, + 397 + ], + "lines": [ + { + "bbox": [ + 370, + 356, + 499, + 397 + ], + "spans": [ + { + "bbox": [ + 370, + 356, + 499, + 397 + ], + "type": "table", + "html": "
Methodslongshortrandom
SF2T69.3871.4073.86
", + "image_path": "c46016b2b6ceb85609d527febf51144299c88c854a5f41617de9296fff7c372d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 406, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 312, + 406, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 312, + 406, + 555, + 440 + ], + "type": "text", + "content": "Table 7. Impact of temporal span. Both long- and short-range temporal modeling reduced " + }, + { + "bbox": [ + 312, + 406, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 312, + 406, + 555, + 440 + ], + "type": "text", + "content": " 's performance, emphasizing the importance of multi-scale temporal modeling." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "content": "datasets. We then generated QA pairs for each frame sequence based on the five tasks defined in " + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "content": " for training. Evaluations were performed on FineVidBench's scene-level tasks, including Action, Effect, and Speed. To compare with traditional SFT, we also generated and manually reviewed QA pairs for these videos in a supervised setting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 550, + 400, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 400, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 400, + 563 + ], + "type": "text", + "content": "5.2. Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Table 4 summarizes the results of the scene-level tasks. After " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " training, all models showed significant improvement, emphasizing that fragment-level tasks can notably enhance scene-level comprehension. Integrating " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " with SFT is also leads to performance gains, demonstrating that fragment-level training positively impacts SFT and enhances its effectiveness. Surprisingly, in the Speed task, many base models outperformed SFT after applying " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", highlighting the importance of fine-grained temporal understanding in distinguishing video speeds. This improvement likely stems from " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "'s ability to enhance the model's sensitivity to temporal cues, such as the loss or enhancement of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29114" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "information during acceleration or deceleration, as well as content coherence—all crucial for speed judgment. As expected, " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " currently lags behind SFT, since its training objective is not fully aligned with scene-level tasks. However, we do not expect " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " to replace supervised fine-tuning; rather, our experiments suggest that it can serve as an effortless and effective complement to SFT." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "type": "text", + "content": "In addition to FineVidBench, we evaluated " + }, + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^{2}\\mathrm{T}" + }, + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "type": "text", + "content": " on three public video understanding benchmarks (Table 5). The results demonstrate consistent improvements across various video tasks, validating " + }, + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^{2}\\mathrm{T}" + }, + { + "bbox": [ + 55, + 157, + 294, + 229 + ], + "type": "text", + "content": " as an effective spatiotemporal enhancer for a wide range of video understanding tasks. All models were tested with an 8-frame input." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 258, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 258, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 258, + 251 + ], + "type": "text", + "content": "5.3. Ablation and Interpretability Analyses" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 460 + ], + "type": "text", + "content": "We evaluated the impact of frame sampling strategies on " + }, + { + "bbox": [ + 55, + 257, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 257, + 295, + 460 + ], + "type": "text", + "content": ", as each method provides a unique \"temporal information perspective\" that influencing video understanding performance. As shown in Table 6, we assessed four strategies on Qwen2-VL in the Action task: random, uniform interval, keyframe, and motion-salient area sampling [37]. Motion-salient area sampling performed best, likely due to its ability to capture continuous motion dynamics, thereby enhancing the model's understanding of action fluidity and temporal detail. In comparison, the other methods had limitations: keyframe sampling misses intermediate action phases, fixed-interval sampling may overlook critical moments, and random sampling lacks temporal consistency. Notably, different datasets may favor different strategies. For example, some datasets may perform better with uniform interval sampling, or their motion features may align better with the model's specific capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": "We examined the effects of long- and short-range temporal modeling on " + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": ". In the Consistency Verification task, we constrained the random selection of frame pairs to adjacent frames for local continuity or non-adjacent frames to capture long-range dependencies. As shown in Table 7, both settings decreased " + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 462, + 295, + 605 + ], + "type": "text", + "content": "'s performance on the Action task of Qwen2-VL, indicating that an overemphasis on either long- or short-range information leads to temporal imbalance and incomplete dynamics. This underscores the importance of combining both approaches to leverage their broader temporal span and frame variations for a more comprehensive feature representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "We analyzed the attention map of Qwen2-VL on the Action task, particularly in cases where the model's predictions were corrected after " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ". As shown in Figure 6, we found that " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " enhances the model's ability to capture fine-grained spatial changes and temporal dynamics. (1) Spatial Aspects. After " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ", the model shows increased attention to action execution areas, particularly the hands and objects they interact with. It shows better sensitivity to small targets, likely due to the Consistency Verification" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 541, + 357 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 541, + 357 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 541, + 357 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 541, + 357 + ], + "type": "image", + "image_path": "9553a2acfb87e33c118766225e4eb5707c664e1fa8e5864d9f59bea5dffc0cda.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "text", + "content": "Figure 6. Two exemplary visualizations of the attention map on Qwen2-VL. For each example: top - Original frames; middle - Base (SFT); bottom - " + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "text", + "content": " applied. As shown by the red boxes, after applying " + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "text", + "content": ", the model better focuses on action execution areas and interacting objects. The " + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{SF^2T}" + }, + { + "bbox": [ + 313, + 365, + 555, + 441 + ], + "type": "text", + "content": " fine-tuned model has the ability to predict the direction of motion, as seen in the trajectories of the red bottle and Cheerios." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 472, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 557 + ], + "type": "text", + "content": "task, which enhances spatial perception by refining sensitivity to subtle image differences. (2) Temporal Aspects. After " + }, + { + "bbox": [ + 313, + 472, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 472, + 555, + 557 + ], + "type": "text", + "content": ", we observed that the model can predict object movement trajectories in certain actions, indicating an advanced level of temporal understanding. This ability likely stems from the sorting task, which strengthens the model's comprehension of action flows and movement patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 581, + 388, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 581, + 388, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 581, + 388, + 594 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " to overcome the limitations of Video-LLMs in fine-grained video understanding. " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SF}^2\\mathrm{T}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " is an innovative fine-tuning method that eliminates the need for labor-intensive annotations and effectively bypasses the constraints of natural language descriptions. Additionally, we introduce FineVidBench, a benchmark for evaluating Video-LLMs at both scene and fragment levels. In the future, we plan to expand our dataset with larger videos and more tasks to increase its impact." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29115" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China (No.2020YBF2901202), National Natural Science Foundation of China (NSFC No. 62272184 and No. 62402189), the China Postdoctoral Science Foundation under Grant Number GZC20230894, the China Postdoctoral Science Foundation (Certificate Number: 2024M751012), and the Postdoctor Project of Hubei Province under Grant Number 2024HBBHCXB014, and the \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (No. 2024C01161). The computation is completed in the HPC Platform of Huazhong University of Science and Technology." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 266, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 266, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 296, + 287 + ], + "type": "text", + "content": "[1] FirstName Alpher. Frobnication. IEEE TPAMI, 12(1):234-778, 2002. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 289, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 296, + 343 + ], + "type": "text", + "content": "[2] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 345, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 296, + 389 + ], + "type": "text", + "content": "[3] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 390, + 296, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 296, + 444 + ], + "type": "text", + "content": "[4] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 446, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 296, + 500 + ], + "type": "text", + "content": "[5] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 578 + ], + "type": "text", + "content": "[6] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "text", + "content": "[7] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 689 + ], + "type": "text", + "content": "[8] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "type": "text", + "content": "[9] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave:" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 98, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 554, + 151 + ], + "type": "text", + "content": "[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 154, + 554, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 554, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 554, + 197 + ], + "type": "text", + "content": "[11] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 201, + 554, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 554, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 554, + 265 + ], + "type": "text", + "content": "[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 268, + 554, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 554, + 300 + ], + "type": "text", + "content": "[13] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 554, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 346 + ], + "type": "text", + "content": "[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 554, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 382 + ], + "type": "text", + "content": "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 384, + 554, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 554, + 427 + ], + "type": "text", + "content": "[16] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 430, + 554, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 554, + 484 + ], + "type": "text", + "content": "[17] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 487, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 554, + 531 + ], + "type": "text", + "content": "[18] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 533, + 554, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 577 + ], + "type": "text", + "content": "[19] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 580, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 554, + 643 + ], + "type": "text", + "content": "[20] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 527-544. Springer, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[21] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. IEEE transactions on pattern analysis and machine intelligence, 42(2):502-508, 2019. 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29116" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[22] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[23] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "text", + "content": "[24] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 270 + ], + "type": "text", + "content": "[25] Fangxun Shu, Lei Zhang, Hao Jiang, and Cihang Xie. Audio-visual llm for video understanding. arXiv preprint arXiv:2312.06720, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "type": "text", + "content": "[26] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7464-7473, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[27] TencentQQ Multimedia Research Team. Video-cam: Advancing video-language understanding with causal cross-attention masks. https://github.com/QQ-MM/Video-CCAM, 2024.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "text", + "content": "[28] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 427, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 480 + ], + "type": "text", + "content": "[29] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[30] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "type": "text", + "content": "[31] Haiyang Xu, Qinghao Ye, Xuan Wu, Ming Yan, Yuan Miao, Jiabo Ye, Guohai Xu, Anwen Hu, Yaya Shi, Guangwei Xu, et al. Youku-mplug: A 10 million large-scale chinese video-language dataset for pre-training and benchmarks. arXiv preprint arXiv:2306.04362, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "text", + "content": "[32] Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[33] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 363 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "text", + "content": "[34] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 119, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 150 + ], + "type": "text", + "content": "[35] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[36] Zijia Zhao, Haoyu Lu, Yuqi Huo, Yifan Du, Tongtian Yue, Longteng Guo, Bingning Wang, Weipeng Chen, and Jing Liu. Needle in a video haystack: A scalable synthetic framework for benchmarking video mllms. arXiv preprint arXiv:2406.09367, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "text", + "content": "[37] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[38] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[39] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29117" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_content_list.json b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e12396f36c8e283b16e0eff0b3dc814c26cfc75c --- /dev/null +++ b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_content_list.json @@ -0,0 +1,1317 @@ +[ + { + "type": "text", + "text": "SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement", + "text_level": 1, + "bbox": [ + 133, + 128, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mark Boss1 Zixuan Huang1,2† Aaryaman Vasishta1 Varun Jampani1 \n1Stability AI ²UIUC", + "bbox": [ + 189, + 209, + 776, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d5377ee19947da1d6960e6092985f4fd109b3d9d83b7984b065abfdab93e9da.jpg", + "image_caption": [ + "Figure 1. SF3D Creates High-quality Object Meshes from Single Images with Materials, Delighting and UV-unwrapped textured meshes in $0.3\\mathrm{s}$ . Here, we show sample SF3D results from different input images. SF3D handles both realistic and non-realistic styles well." + ], + "image_footnote": [], + "bbox": [ + 140, + 301, + 831, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 656, + 313, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SF3D, a novel method for rapid and high-quality textured object mesh reconstruction from a single image in just 0.3 seconds. Unlike most existing approaches, SF3D is explicitly trained for mesh generation, incorporating a fast UV unwrapping technique that enables swift texture generation rather than relying on vertex colors. The method also learns to predict material parameters and normal maps to enhance the visual quality of the reconstructed 3D meshes. Furthermore, SF3D integrates a delightting step to effectively remove low-frequency illumination effects, ensuring that the reconstructed meshes can be easily used in novel illumination conditions. Experiments demonstrate the su", + "bbox": [ + 75, + 688, + 472, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perior performance of SF3D over the existing techniques.", + "bbox": [ + 500, + 657, + 880, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page with code and model: https://stable-fast-3d.github.io", + "bbox": [ + 500, + 674, + 893, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 720, + 632, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality object meshes are essential for various use cases in movies, gaming, e-commerce, and AR/VR. In this work, we tackle the problem of generating high-quality 3D object mesh from a single image. This is an ill-posed and challenging problem as this requires reasoning about the object's 3D shape and texture from only a single 2D projection (image) of that object. Single-image object generation can simplify the tedious and manual object creation process.", + "bbox": [ + 496, + 748, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The quality of object mesh generation from a single image has dramatically improved in the past couple of years", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done during internship at Stability AI.", + "bbox": [ + 94, + 886, + 334, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16240", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb55ed3e8a5331b0c70279fd98590c49c5c51bad7e934db3bbdd48f0b3f10fa4.jpg", + "image_caption": [ + "Figure 2. SF3D Improvements on Different Prevalent Issues. Here, we compare our results with TripoSR [55]. The top shows the effect of light bake-in when relighting the asset. SF3D produces a more plausible relighting (see input in inset). By not using vertex colors, our method is capable of encoding finer details while also having a lower polygon count. Our vertex displacement enables estimating smooth shapes, which do not introduce stair-stepping artifacts from marching cubes. Lastly, our material property prediction allows us to express a variety of different surface types." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 450, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with the advances in transformer models [19, 21, 55], large synthetic datasets [12] and 3D-aware image/video generative models [33, 52, 58, 68]. Especially, the transformer-based reconstruction models [19, 21, 55] demonstrate phenomenal generalization capabilities on real-world images despite being trained only on synthetic datasets while also generating 3D assets from a single image in under 1 s.", + "bbox": [ + 75, + 575, + 468, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite this rapid progress, several issues remain in these feed-forward fast 3D reconstruction models [19, 21, 55]. These techniques often produce 3D assets that are not usable for downstream applications or require laborious manual post-processing. We identify several key issues in these techniques and propose a fast generation technique called 'Stable Fast 3D' (SF3D) that generates higher quality and more usable 3D assets from single images, while also retaining the fast generation speed within 0.3 seconds on a H100 GPU. Next, we briefly introduce these issues and how we tackle those in SF3D.", + "bbox": [ + 75, + 685, + 468, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Light Bake-in. Having shadows or other illumination effects in a given input image is common. Most existing works bake these effects into textures, making the resulting", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D assets less usable. Having consistent lighting helps in easy integration into graphics pipelines. In SF3D, we propose decomposing the illumination and reflective properties by incorporating explicit illumination and a differentiable shading model using Spherical Gaussians (SG). Fig. 2 (top row) shows a sample result of SF3D, where the light bake-in is considerably reduced compared to the prior art.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vertex Coloring. Another issue we found in most 3D generation models is that they produce meshes with a high vertex count, using vertex coloring to represent object texture. This makes the resulting 3D assets inefficient to use in applications such as games. A key issue is the additional processing time of UV unwrapping, which can take longer than the entire object generation. For example, xatlas [73] and geogram [29], can take up to $30\\mathrm{s}$ or $10\\mathrm{s}$ for a single asset, respectively. To tackle this, we propose a highly parallelizable fast box projection-based UV unwrapping technique to achieve a $0.5\\mathrm{s}$ generation time. The effect of relying on vertex coloring vs. UV Unwrapping can be seen in Fig. 2 (middle row), where TripoSR captures fewer details than SF3D despite requiring $10\\times$ higher polygon count.", + "bbox": [ + 496, + 199, + 892, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Marching Cubes Artifacts. The feed-forward networks often create volumetric representations such as Triplane-NeRFs [9] which are converted to meshes using the Marching Cubes (MC) [37] algorithm. MC can cause 'stair-stepping' artifacts, which can be somewhat reduced by increasing the volume resolution. However, this comes at the cost of large computational overhead. In contrast, SF3D uses more efficient architecture for higher resolution triplanes and also produces meshes using DMTet [48] with learned vertex displacements and normal maps, resulting in smoother mesh surfaces. Fig. 2 demonstrates the smoothness of the SF3D mesh compared to that of TripoSR.", + "bbox": [ + 496, + 412, + 892, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lack of Material Properties. The generations from previous feed-forward techniques often look dull when they are rendered using different illuminations. This is mainly due to the lack of explicit material properties in the output generations, which can influence the light reflection. To tackle this, we predict non-spatially varying material properties. This addition is apparent when rendering different generated objects in Fig. 2 (bottom row).", + "bbox": [ + 496, + 597, + 890, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these advances, SF3D can generate high-quality 3D meshes from a single image with several desirable properties for downstream applications on both shapes (low-polygon yet smooth) and textures (Illumination disentangled UV maps with material properties). The 3D assets are small in size (under $1\\mathrm{MB}$ ) and can be created in $0.5\\mathrm{s}$ . For text-to-3D mesh generation, a fast Text-to-Image (T2I) model [47] can be combined with SF3D to produce meshes in about $1\\mathrm{s}$ . Experimental results demonstrate higher quality results with SF3D compared to the existing works. In short, SF3D provides a comprehensive technique for fast, high-quality 3D object generation from single images, ad", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16241", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dressing both speed and usability in practical applications.", + "bbox": [ + 76, + 90, + 462, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 125, + 218, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Reconstruction using Image Generative Priors. Diffusion models [20, 51] have proven to be powerful generative models for various tasks [2, 3, 16, 45, 46, 57]. Several works such as Zero123 [33] and others [12, 26, 52, 81] leverage the object priors in these diffusion models by adapting the generative models for 3D generation. Score Distillation Sampling (SDS) [43] is often used to optimize 3D representation using the 2D diffusion models. However, it was found that relying solely on the image prior does not always produce consistent multi-view results. This issue is improved in follow-up works [34, 39, 49, 50] by simultaneously generating multiple views of an object. Another approach is to explicitly introduce 3D awareness [31, 32, 72] or use a multi-view diffusion process [2, 15, 27, 35, 36, 38, 50, 58, 63, 80] to generate 3D objects. While diffusion models can generate videos or multi-view images relatively quickly, they require a 3D reconstruction step to create 3D mesh from a single image. Even with fast techniques, generating an object can still take several minutes. Our work focuses on fast generation speeds in $0.3\\mathrm{s}$ for image-to-3D.", + "bbox": [ + 76, + 154, + 472, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feed-Forward 3D Reconstruction. Recently, LRM [21] and the follow-up works [19, 55] demonstrated that the fast and reliable 3D generation is possible using feed-forward networks trained on large synthetic datasets. These works use large transformer models to directly predict triplanes as the volume representations, which can be ray-marched using NeRF [40]. This allows these models to train on multi-view datasets, as only the rendering loss is required for training. A set of follow-up works address to remedy the reliance on multi-view datasets [24, 59, 66]. This can produce meshes from images in seconds compared to the generative-prior based approaches.", + "bbox": [ + 76, + 460, + 472, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several follow-up works emerged which use Gaussian Splatting [25] as the representation [53, 54, 71, 77, 83] or directly integrate a mesh prediction [61, 64, 67, 69, 74, 82]. Another group of methods integrates diffusion models with the feed-forward models by either directly generating the triplanes [28, 60] or 3D [11], using LRM as the denoiser [70], or using it as the conditioning [62]. As single-image reconstruction is a challenging task, several models leverage the extensive prior of multi-view diffusion models to generate multiple views of an object which the feed-forward model then uses to produce a 3D output [30, 54, 61, 64, 67, 69, 71, 74, 77, 82]. Generally, these models learn the scene's radiance, meaning lighting information is baked into the objects. LDM [67] tries to remedy this by learning an additional shading color to capture the shading information. However, training requires having access to the ground-truth albedo color as a supervision", + "bbox": [ + 76, + 643, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "signal. Compared to LDM, our method also learns an illumination model, enabling training on regular multi-view datasets, and we predict more material properties. We also tackle fast UV unwrapping, so unlike previous works, we do not have to rely on vertex colors while keeping the 3D generation time short.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material Decomposition. Predicting only the radiance of an object has the downside that relighting does not produce convincing results. Current works in single scene optimization are often based on NeRF [40] or Gaussian Splating [25] and decompose multiple input images ( $>50$ ) into light and materials [4-6, 14, 18, 41, 76]. These methods often predict materials properties of a Physically Based Rendering (PBR) [7] shading model. A few recent works tackle joint 3D shape and material generation, such as UniDream [35] or Fantasia3D [10] by optimizing the material properties using a SDS loss. However, this optimization takes several hours to converge. Another set of works aims to texture existing objects [56, 79] using diffusion models. Our work generates a textured object with homogeneous material properties from a single image under natural illumination at fast generation speeds.", + "bbox": [ + 496, + 184, + 893, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 444, + 591, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose SF3D, which converts a single object image into a textured and UV-unwrapped 3D model with de-lit albedo and material properties. As explained in the introduction section, with SF3D, we aim to fix the issues shown in Fig. 2 and introduce additional quality enhancements.", + "bbox": [ + 496, + 470, + 893, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. Our method is based on TripoSR [21, 55], which trains a large transformer-based network that outputs a Triplane [9] based 3D representation from a single image. TripoSR is trained with multi-view image datasets without explicit 3D supervision. In TripoSR, the image is encoded using DINO [8] and passed through a Transformer network to generate a 3D triplane at a resolution of $64 \\times 64$ . The triplane features are then decoded into RGB colors and rendered using standard NeRF [40] rendering into multiple views for training. TripoSR only learns the view-independent colors and cannot model reflective objects. Several issues of the TripoSR (and other similar networks) are explained in the introduction (Fig. 2).", + "bbox": [ + 496, + 549, + 893, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SF3D Overview. With SF3D, we propose several improvements to TripoSR [55] to improve the output quality in different aspects. As illustrated in Fig. 3, SF3D has 5 main components: 1. An enhanced transformer network that predicts higher resolution triplanes, which helps in reducing aliasing artifacts (top left in the figure); 2. A material estimation network (bottom left) predicts material properties, which helps handle an object's reflective properties. 3. Illumination prediction (bottom right) to tackle illumination disentangling, which helps in outputting homogeneous", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16242", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8e457eeb0d8dda08547ac97ef40a15a5594e2a94da8d22d76bddadea36e23eb5.jpg", + "image_caption": [ + "Figure 3. SF3D Overview. SF3D improves on TripoSR by addressing the issues in Fig. 2 with 5 novel modules: 1. An enhanced transformer for higher resolution triplanes (top left); 2. Material estimation with Material Net (bottom left); 3. Explicit illumination estimation using Light Net (bottom right); 4. Smooth mesh extraction with the estimation of vertex offsets and normals (top right); and finally 5. an export pipeline with fast UV-unwrapping (right)." + ], + "image_footnote": [], + "bbox": [ + 117, + 95, + 830, + 300 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b51713e89043b01dcb251614b23dcaad55a142fbb1d850fd567c320081e1ef17.jpg", + "image_caption": [ + "Figure 4. Triplane Resolution Aliasing. We found that low-resolution triplanes struggle with high-frequency and high-contrast textures and produce grid-like aliasing artifacts. Our method increases the triplane size from $64^{2}$ to $384^{2}$ allowing our method to reproduce these textures with fewer artifacts." + ], + "image_footnote": [], + "bbox": [ + 106, + 383, + 444, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "objects without shadows; 4. Mesh extraction and refinement with the prediction of vertex offsets and surface normals (top right), which helps in smoother output shapes with fewer mesh extraction artifacts; and 5. A fast UV-unwrapping and export module (right) that helps produce low-poly meshes and high-resolution textures. Next, we explain each of these modules in detail.", + "bbox": [ + 75, + 595, + 468, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Enhanced Transformer", + "text_level": 1, + "bbox": [ + 76, + 712, + 292, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, as illustrated in Fig. 3, we transitioned from DINO [8] used in TripoSR to the improved DINOv2 [42] network to obtain image tokens for the transformer. We observed that low-resolution $(64\\times 64)$ triplanes used in TripoSR and other networks [19, 21] introduce noticeable artifacts, especially in scenarios with high-frequency and high-contrast texture patterns. Fig. 4 (Middle) illustrates these aliasing artifacts. The triplane resolution is directly correlated with the presence of these artifacts, which we identify as an aliasing issue that can be mitigated by increasing the resolution. The increased capacity also improves the geometry.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Naively increasing the triplane resolution quadratically increases the transformer complexity. We take inspiration from the recent PointInfinity [23] work and propose an enhanced Transformer network that outputs higher-resolution triplanes. PointInfinity proposes an architecture where the complexity remains linear concerning the input size by avoiding the self-attention on the higher resolution triplane tokens. With this addition, we produce $96 \\times 96$ resolution triplanes with 1024 channels. We further increased the triplane resolution by shuffling the output features across dimensions, resulting in 40-channel features at $384 \\times 384$ resolution. Further details about the architecture are available in the supplements. Fig. 4 (Right) shows fewer aliasing artifacts with our higher resolution triplanes.", + "bbox": [ + 496, + 388, + 893, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Material Estimation", + "text_level": 1, + "bbox": [ + 498, + 607, + 692, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the output mesh appearance for the reflective objects, SF3D also outputs the material properties of metallic and roughness parameters. Ideally, one would like to estimate the spatially varying material properties at 3D output locations, but this is an inherently challenging and ill-posed learning problem that requires many high-quality 3D data with spatially varying materials. To overcome these challenges, we propose simplifying the material estimation problem by estimating a single metallic and roughness value for the entire object. Although this non-spatially varying material mainly applies to homogeneous objects, we find that it significantly improves the visual quality of our mesh predictions. Specifically, as illustrated in Fig. 3, we propose 'Material Net' that predicts the metallic and roughness values from the input image.", + "bbox": [ + 496, + 628, + 890, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For pre-training the Material Net, we selected a subset of 3D objects with PBR material properties from the synthetic training dataset and rendered them under different illumi", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16243", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nations and viewpoints. We observe that directly regressing the material values often leads to training collapse, where the network always predicts a roughness value of 0.5 and a metallic value of 0. As a remedy, we propose a probabilistic prediction approach, where we predict the parameters of a Beta distribution and minimize the log-likelihood during training. This stabilizes the training by allowing for uncertainty in this ambiguous material estimation task and prevents the collapse observed with direct regression. During inference and training of SF3D, we do not sample the distribution but calculate the mode of the distribution.", + "bbox": [ + 75, + 90, + 472, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement the Material Net by first passing the image through the frozen CLIP image encoder [44] to extract semantically meaningful latents and pass them through 2 separate MLPs with 3 hidden layers and 512 width to output the parameters for the distributions.", + "bbox": [ + 75, + 257, + 470, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Illumination Modeling", + "text_level": 1, + "bbox": [ + 76, + 339, + 287, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose explicitly estimating the illumination in the input image to account for varying shadings (e.g., shadows). Otherwise, the 3D outputs would have baked-in illumination effects into their RGB colors, as illustrated in Fig. 2. To this end, we propose a Light Net (Fig. 3 bottom right) that predicts the spherical Gaussian (SG) illumination map from the estimated triplanes. The rationale here is that the triplanes encode the global structure and appearance of the input object and should account for the 3D spatial relationships and changes in illumination over the object surface. We use the $96 \\times 96$ resolution triplanes from the transformer and pass them through 2 CNN layers, followed by a max pool and final MLP with three hidden layers and a feature dimension 512 for all layers. Light Net outputs the grayscale amplitude values for 24 SGs with a Softplus activation to ensure positive values. The axis and sharpness values for these SGs remain fixed and are set up to cover the entire sphere. These amplitude values allow us to implement a deferred physically based rendering approach similar to that used in NeRD [4].", + "bbox": [ + 75, + 359, + 472, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method also incorporates a lighting demodulation loss $\\mathcal{L}_{\\mathrm{Demod}}$ during the training phase, inspired by the works of Hasselgren et al. [18] and Voleti et al. [58]. This loss function ensures that the lighting on an object with an entirely white albedo closely matches the luminance of the input image. The demodulation loss enforces consistency between the learned illumination and the lighting conditions observed in the training data. This can be seen as a bias to resolve the ambiguity between appearance and shading [1].", + "bbox": [ + 75, + 662, + 470, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mesh Extraction and Refinement", + "text_level": 1, + "bbox": [ + 76, + 805, + 369, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 (top right) illustrates this module. We convert the estimated triplanes into a mesh using a differentiable Marching Tetrahedron (DMTet) [48] technique. As explained in the introduction (Fig. 2), Marching Cubes (MC) usually results in several staircase artifacts on the resulting meshes.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f75041f11cccf95da2243a4605439b01227e21f35790fd6850b4fb2c31de017.jpg", + "image_caption": [ + "Figure 5. Export Pipeline. Our export process starts with the mesh and is followed by UV unwrapping, occupancy and world position baking, material querying, and UV island margins." + ], + "image_footnote": [], + "bbox": [ + 532, + 87, + 866, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose two new MLP heads to refine the meshes as a remedy. One predicts vertex offsets $\\boldsymbol{v}_o \\in \\mathbb{R}^3$ , and another one predicts the world space vertex normals $\\hat{\\boldsymbol{n}} \\in \\mathbb{R}^3$ . Inspired by MeshLRM [61], we also implemented small split decoder MLPs for these two networks, which proved beneficial for performance and efficiency. We found that the vertex offsets can reduce artifacts from the tetrahedral grids, and the world space normals can add details to the flat mesh triangles. Given that the normal predictions are initially unreliable, we stabilize the training by using spherical linear interpolation (slerp) between the geometry normals $\\boldsymbol{n} \\in \\mathbb{R}^3$ and our predictions. This slerp is used during the initial 5K training steps.", + "bbox": [ + 496, + 252, + 893, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To regularize the mesh estimation, we use several training losses: a normal consistency loss $\\mathcal{L}_{\\mathrm{Nrm~consistency}}$ , a Laplacian smoothness loss $\\mathcal{L}_{\\mathrm{Laplacian}}$ as implemented by threestudio [17], and a vertex offset regularization $\\mathcal{L}_{\\mathrm{Offset}} = \\pmb{v}_o^2$ . For supervising the normal prediction, we use a geometry normal replication loss $\\mathcal{L}_{\\mathrm{Nrm~repl}} = 1 - \\pmb{n} \\cdot \\hat{\\pmb{n}}$ , where $\\cdot$ is the dot product and a normal smoothness loss to ensure the smoothness of normal predictions in 3D. This loss is achieved by adding a small offset $\\epsilon \\in \\mathbb{R}^3$ around a query location $\\pmb{x} \\in \\mathbb{R}^3$ . The loss is then defined as $\\mathcal{L}_{\\mathrm{Nrm~smooth}} = (\\hat{n}(\\pmb{x}) - \\hat{n}(\\pmb{x} + \\epsilon))^2$ .", + "bbox": [ + 496, + 450, + 893, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Fast UV-Unwrapping and Export", + "text_level": 1, + "bbox": [ + 500, + 623, + 794, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final stage of SF3D is an export pipeline that outputs the final 3D mesh along with the corresponding UV atlas. Our export pipeline follows multiple stages to ensure efficient and effective handling of 3D models. An overview of these stages is provided in Fig. 5, where we first do fast UV-unwrapping. We then bake the world positions and occupancy to the UV atlas, which we use for querying the albedo and normal. This results in the final textured 3D mesh. The entire export process only takes $10\\mathrm{ms}$ .", + "bbox": [ + 496, + 643, + 893, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UV unwrapping is traditionally a computationally intensive process. Existing methods require several seconds for UV unwrapping, which is impractical when we are aiming for sub-second generation speeds. To address this inefficiency, we propose a Cube projection-based unwrapping method. The key advantage of this approach is it is parallelizable: each face of the mesh can independently decide which cube face to project onto, based on its surface normal.", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16244", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15ef6c79262304687f563ffec52e7cc444a03f990d3a8c9a49b3b125a6386e9c.jpg", + "image_caption": [ + "Figure 6. UV Unwrapping. Our UV unwrapping technique uses projection mapping, allowing each face to independently select a projection, enabling easy parallelization. A naive approach could lead to the same UV coordinates being assigned to different vertices due to occlusions. We identify potential overlaps from occlusions on the 2D mapped surfaces and relocate them to different areas within the UV atlas. Any remaining areas are placed in the bottom right of the UV atlas. This method minimizes distortion and ensures most surfaces are preserved in a connected area." + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 444, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our UV unwrapping process is illustrated in Fig. 6. We initially align the output mesh based on the most dominant axes with the cube projection coordinate system. After each mesh face selects the appropriate cube direction, we address potential occlusions. Without managing occlusions, different faces could share the same UV coordinates, leading to artifacts in the texture. We detect occlusions in the UV atlas by performing 2D triangle-triangle intersection tests. We filter triangles by their proximity to the triangle centers to make the process efficient. If an intersection is detected, we sort the intersecting triangles based on their depth in the plane, keeping the first intersection and marking the others for reassignment to different UV atlas areas. The first intersection is placed in the top third of the UV atlas and the second intersection is placed in the bottom left area. The remaining triangles are organized into a grid in the bottom right section of the atlas. We also rotate each island to minimize the shading seams by following radial z tangent orientation. We then assign each face to a position in the UV atlas, as illustrated in Fig. 6.", + "bbox": [ + 75, + 397, + 472, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we bake the world positions and the occupancy data with UV-unwrapping into the final UV atlas. This allows us to query the world positions within the chart from our triplanes and decode the albedo and surface normals into additional textures. We transform the world-space normal map into a tangent-space normal map using the tangent and bitangent vectors. We add margins to the UV atlases to prevent visible seams at UV island borders. This is achieved through an iterative process: in each iteration, we perform a $3 \\times 3$ partial convolution based on the occupied areas, using the valid regions of the kernel. We then use a $3 \\times 3$ max pooling operation to expand the occupied regions of the UV atlas, placing the mean values in the newly expanded areas", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while preserving the original regions. This iterative extension ensures that the textures smoothly blend outwards.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We incorporate our image estimator's metallic and roughness values and pack everything into a GLB file, ready for efficient rendering and use in various applications.", + "bbox": [ + 498, + 122, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Overall Training and Loss Functions", + "text_level": 1, + "bbox": [ + 498, + 175, + 818, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Directly training our method with mesh rendering yielded unsatisfactory results. Hence, we pre-trained it on the NeRF task. Following this pre-training, we transitioned to mesh training, replacing the NeRF rendering with differentiable mesh rendering and SG-based shading. Given the introduction of light estimation, we found that using larger batch sizes aids convergence. We initiate training with a batch size of 192 and a rendering resolution of $128 \\times 128$ , training for 10K steps. In the subsequent stage, we reduce the batch size to 128 and increase the resolution to $256 \\times 256$ , continuing for 20K steps. The final stage involves 80K steps at a $512 \\times 512$ resolution with a batch size of 96.", + "bbox": [ + 496, + 196, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The loss functions remain consistent across all mesh training stages. We primarily use image-based metrics to compare our rendered and shaded reconstructions $\\hat{I}$ with the GT image $I$ . These include MSE $\\mathcal{L}_{\\mathrm{MSE}}$ and LPIPS [78] $\\mathcal{L}_{\\mathrm{LPIPS}}$ losses. We also incorporate a mask loss $\\mathcal{L}_{\\mathrm{Mask}}$ between the GT mask $M$ and the predicted opacity $\\hat{M}$ , defined as an MSE loss. We then define three loss formulations for the rendering, mesh regularization, and shading:", + "bbox": [ + 496, + 378, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n d e r}} = \\underbrace {\\lambda_ {\\mathrm {M S E}}} _ {1 0} \\mathcal {L} _ {\\mathrm {M S E}} + \\underbrace {\\lambda_ {\\mathrm {L P I P S}}} _ {2} \\mathcal {L} _ {\\mathrm {L P I P S}} + \\underbrace {\\lambda_ {\\mathrm {M a s k}}} _ {1 0} \\mathcal {L} _ {\\mathrm {M a s k}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 516, + 890, + 542 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m e s h}} = \\underbrace {\\lambda_ {\\text {L a p l a c i a n}}} _ {0. 0 1} \\mathcal {L} _ {\\text {L a p l a c i a n}} + \\underbrace {\\lambda_ {\\text {N r m C o n s i s t e n c y}}} _ {0. 0 0 1} \\mathcal {L} _ {\\text {N r m c o n s i s t e n c y}} + \\underbrace {\\lambda_ {\\text {O f f s e t}}} _ {0. 1} \\mathcal {L} _ {\\text {O f f s e t}} ^ {(2)}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 545, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s h a d i n g}} = \\underbrace {\\lambda_ {\\mathrm {N r m} \\text {r e p l}}} _ {0. 2} \\mathcal {L} _ {\\mathrm {N r m} \\text {r e p l}} \\underbrace {\\lambda_ {\\mathrm {N r m} \\text {s m o o t h}}} _ {0. 0 2} \\mathcal {L} _ {\\mathrm {N r m} \\text {s m o o t h}} + \\underbrace {\\lambda_ {\\mathrm {D e m o d}}} _ {0. 0 1} \\mathcal {L} _ {\\mathrm {D e m o d}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 573, + 890, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The total loss is defined as:", + "bbox": [ + 500, + 611, + 681, + 625 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e n d e r}} + \\mathcal {L} _ {\\text {m e s h}} + \\mathcal {L} _ {\\text {s h a d i n g}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 633, + 890, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 674, + 586, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For comparisons, we select GSO [13] and OmniObject3D [65] as our primary datasets for evaluation. We select 278 random scenes from GSO and 308 scenes from OmniObject3D for comparison. We render 16 views around the object and select a frontal as the conditioning view.", + "bbox": [ + 496, + 702, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with several recent methods for fast 3D object reconstruction from a single image. We mainly focus on fast reconstruction models to maintain a consistent evaluation protocol across different techniques. We also mainly consider meshes as outputs and perform the evaluations on the mesh. We use the official implementation for all the baselines, and we evaluate all the methods under the same protocol. We selected several recent and concurrent", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16245", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b03ab237c238fb37ae07d73bb4f24653335b01c0ddde8b1007f4ee16df720c8.jpg", + "image_caption": [ + "Figure 7. Comparison on GSO and OmniObject3D. Notice how our reconstructions produce consistent results with detailed textures and smooth shading." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 880, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a36058dbad440bc448572095f7bb30abce16b718e51769259579f5a1cde4b1a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTime [s]↓GSOOmniObject
3DImage3DImage
CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓
ZeroShape [22]0.90.1600.4890.7590.952———0.1440.5070.7860.969———
OpenLRM [19]2.00.1600.4720.7510.95415.6890.7870.2060.1390.5210.7980.97113.9750.7600.229
TripoSR [55]0.30.1110.6450.8690.98016.4450.7890.1940.1030.6720.8890.98614.3310.7550.224
LGM [54]64.60.1950.3760.6540.92814.3770.7620.2480.2050.3440.6310.92112.6620.7320.273
CRM [60]10.20.1790.4110.6990.94515.0540.7780.2280.1580.4690.7520.96013.4620.7550.245
InstantMesh [69]32.40.1380.5490.8010.96715.4340.7850.2030.1380.5600.8110.96413.5310.7570.235
LN3Diff [28]5.00.1840.4220.6920.93915.4470.7790.2080.1650.4700.7350.95213.2560.7500.243
3DTopia-XL [11]38.30.2010.4100.6700.91614.9730.7800.2230.1970.3970.6700.92612.7100.7530.259
SF3D (Ours)0.30.0960.7100.8970.98817.6190.8210.1520.0890.7310.9220.99015.1360.7850.209
", + "bbox": [ + 96, + 521, + 872, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "works with source code releases for comparisons. Specifically, we compare against OpenLRM [19], TripoSR [55], LGM [54], CRM [60], InstantMesh [69], ZeroShape [22], 3DTopia-XL [11] and LN3Diff [28]. For OpenLRM, we selected the large Objaverse 1.1 model. We use H100 GPU for comparisons.", + "bbox": [ + 75, + 742, + 468, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Evaluations.** For runtime comparisons, we consider the meshes as the final output and calculate the entire run time required to go from the input image to the final mesh. We then perform separate evaluations for the shape qual", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8600cb094d6d5391714649ea65554d03a2af9319c6de4e0df81f3b7dd8348da.jpg", + "table_caption": [ + "Table 1. Comparison on 3D and Image Metrics Demonstrating State-of-the-art Performance of SF3D. It is worth pointing out that all other methods produce meshes with drastically higher polygon counts. This allows them to follow the surfaces more closely than our low polygon meshes. Our meshes can outperform the others by using our learned vertex offsets and high resolution supervision. Our method is also one of fastest to generate a mesh from an image due to our efficient extraction pipeline." + ], + "table_footnote": [], + "table_body": "
MethodGSO
CD↓FS@0.1↑FS@0.2↑FS@0.5↑
TripoSR [55]0.1110.6450.8690.980
SF3D w/o Enh. Transformer0.1080.6600.8720.982
SF3D (Ours)0.0980.7010.8940.988
", + "bbox": [ + 532, + 739, + 859, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation. As our model builds upon TripoSR, we use it as our baseline. Without our enhanced transformer, our mesh training has already improved upon it. With our architectural improvements, we outperform the baselines significantly.", + "bbox": [ + 496, + 813, + 893, + 869 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16246", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29602994d6a88c1e75c40b92cc233387b377e42ed86e7afbd074974d65134e2e.jpg", + "image_caption": [ + "Figure 8. Decomposition Results. Here, we use high-quality objects from Polyhaven [75] and render them under natural illumination. These illuminations are highly challenging for material estimation. Still our model estimates sensible material properties, which allow for a convincing relighting." + ], + "image_footnote": [], + "bbox": [ + 102, + 85, + 436, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10cf8b321af06033b59c6198557b8d460ff516348cee8009eccf8810236bb8e5.jpg", + "image_caption": [ + "Figure 9. Image-To-Mesh Time vs. Reconstruction Quality. Our method is not only one of the fastest reconstruction methods, it is also capable of producing highly accurate geometry." + ], + "image_footnote": [ + "SF3D OpenLRM LGM InstantMesh ZeroShape TripoSR LN3Diff CRM 3DTopia-XL" + ], + "bbox": [ + 81, + 337, + 421, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ity. Several models cannot be conditioned on intrinsic and extrinsic camera parameters, so we propose performing an alignment step. We normalize the mesh, perform brute-force rotation alignment tests, and select the rotation with the lowest Chamfer Distance (CD). We then run another alignment step using Iterative Closest Point (ICP), where we further optimize the rotation and translation. We then calculate the standard shape metrics of CD and F-score (FS). For image metrics we further scale the normalized objects back to the actual GT scale and run another finer ICP optimization to adjust the fine scale. This alignment might still misalign some symmetric objects for re-rendering.", + "bbox": [ + 75, + 625, + 468, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. In Table 1, we compare our method with all the baselines. Here, our method outperforms all current and concurrent baselines on both 3D metrics such as CD and F-scores and image metrics as PSNR, SSIM, and LPIPS. This indicates that our method can reconstruct accurate shapes and textures, even if our models have fewer polygons than", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the other baselines. From the visual comparisons in Fig. 7, it can be seen that the accurate shape reconstruction of SF3D also translates well to the visual quality of the 3D assets. Here, it is worth noting that SF3D also handles fine geometry like the glasses well and produces consistent shapes with more detailed textures than the SOTA methods. Moreover, the results also show sensible material properties and albedo as seen in Fig. 8. This is highly challenging as the objects are only rendered under natural illumination. Estimating material properties without any knowledge about the illumination is a highly ambiguous problem.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference Speed vs. Reconstruction Quality. In Fig. 9, we plot the inference speed vs. reconstruction quality for different techniques. The best-performing methods should be located in the top left corner. While our method is slightly slower than TripoSR, the reconstruction accuracy is considerably better for SF3D. It is also worth noting that our reconstruction has smoother shading with less pronounced marching cube artifacts, as seen in Fig. 7. The final apparent asset quality is thus even higher for our method.", + "bbox": [ + 496, + 258, + 892, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations. We evaluate our additions in Table 2 against the baseline models. Our method is built upon TripoSR, so we use it as our initial baseline. If we add mesh training and relighting as a fine-tuning step, we can see that 'SF3D w/o Enhanced Transformer' already improves upon TripoSR, demonstrating the use of our mesh-based training. This improvement is mainly due to the efficient rendering, enabling higher resolution supervision during training and smoother meshes from vertex offsets. If we add our high-resolution triplanes using our enhanced transformer, SF3D further outperforms the baseline considerably.", + "bbox": [ + 496, + 396, + 892, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Outlook. As seen in Fig. 7 (top row), the cup's albedo is not perfectly matched. This is related to the LDR input, where the dark spots do not contain any useful information. Furthermore, our roughness and metallic properties are homogeneous, which limits their usefulness for objects containing multiple drastically different materials that are spatially-varying. In addition, our method also introduces material prediction and delighting without explicit supervision. As we do not require explicit supervision of these parameters, our method could be extended to train on real-world datasets, which we leave for future work. Similarly, the UV unwrapping could leverage existing datasets for further improvements.", + "bbox": [ + 496, + 565, + 892, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 768, + 619, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SF3D, a fast single view to uv-unwrapped textured object reconstruction method. In addition to our fast extraction pipeline, we introduce several architectural improvements to feed-forward-based 3D reconstruction methods, which help our model produce highly detailed geometry and texture. In our extensive evaluation, we show that our method outperforms existing and concurrent baselines in both speed and quality.", + "bbox": [ + 496, + 787, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 173, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] E.H. Adelson and A.P. Pentland. The perception of shading and reflectance, page 409-424. Cambridge University Press, 1996. 5", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv, 2023. 3", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv, 2023. 3", + "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 3, 5", + "[5] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021.", + "[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. NeurIPS, 2022. 3", + "[7] Brent Burley. Physically based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 3", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. ICCV, 2021. 3, 4", + "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3", + "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 3", + "[11] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia-xl: High-quality 3d pbr asset generation via primitive diffusion. arXiv, 2024. 3, 7", + "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of $10\\mathrm{m} + 3\\mathrm{D}$ objects. arXiv, 2023. 2, 3", + "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6" + ], + "bbox": [ + 78, + 143, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Andreas Engelhardt, Amit Raj, Mark Boss, Yunzhi Zhang, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SHINOBI: Shape and Illumination using Neural Object decomposition via Brdf optimization In-the-wild. In CVPR, 2024. 3", + "[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create anything in 3D with multi-view diffusion models. arXiv, 2024. 3", + "[16] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning, 2023. 3", + "[17] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5", + "[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light & Material Decomposition from Images using Monte Carlo Rendering and Denoising. NeurIPS, 20222. 3, 5", + "[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2, 3, 4, 7", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[21] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. ICLR, 2024. 2, 3, 4", + "[22] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M. Rehg. ZeroShape: Regression-based zero-shot shape reconstruction. arXiv, 2023. 7", + "[23] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In CVPR, 2024. 4", + "[24] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3D: Scaling up large reconstruction models with real-world images. arXiv, 2024. 3", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4), 2023. 3", + "[26] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. EscherNet: A generative model for scalable view synthesis. arXiv, 2024. 3", + "[27] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. ViVid-1-to-3: Novel view synthesis with video diffusion models. CVPR, 2024. 3", + "[28] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 3, 7", + "[29] Bruno Levy. geogram. https://github.com/BrunoLevy/geogram, 2024.2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16248", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. arXiv, 2023. 3", + "[31] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv, 2023. 3", + "[32] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. NeurIPS, 2023. 3", + "[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. ICCV, 2023. 2, 3", + "[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv, 2023. 3", + "[35] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for re-1ightable text-to-3D generation. arXiv, 2023. 3", + "[36] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. arXiv, 2023. 3", + "[37] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM Transactions on Graphics (SIGGRAPH), 1987. 2", + "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv, 2024. 3", + "[39] Antoine Mercier, Ramin Nakhli, Mahesh Reddy, Rajeev Yasarla, Hong Cai, Fatih Porikli, and Guillaume Berger. HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv, 2024. 3", + "[40] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. ECCV, 2020. 3", + "[41] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 3", + "[42] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2d diffusion. arXiv, 2022. 3", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3", + "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv, 2022. 3", + "[47] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv, 2023. 2", + "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 5", + "[49] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv, 2023. 3", + "[50] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv, 2023. 3", + "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv, 2020. 3", + "[52] StabilityAI. Stable Zero123, 2023. 2, 3", + "[53] Stanislaw Szymanowicz, C. Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3D reconstruction. CVPR, 2024. 3", + "[54] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. arXiv, 2024. 3, 7", + "[55] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. TripoSR: Fast 3D object reconstruction from a single image. arXiv, 2024. 2, 3, 7", + "[56] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned PBR image generation. arXiv, 2024. 3", + "[57] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 3", + "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16249", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. arXiv, 2024. 2, 3, 5", + "[59] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. PF-LRM: Pose-free large reconstruction model for joint pose and shape prediction. arXiv, 2023. 3", + "[60] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. CRM: Single image to 3D textured mesh with convolutional reconstruction model. arXiv, 2024. 3, 7", + "[61] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv, 2024. 3, 5", + "[62] Hao Wen, Zehuan Huang, Yaohui Wang, Xinyuan Chen, Yu Qiao, and Lu Sheng. Ouroboros3D: Image-to-3D generation via 3D-aware recursive diffusion. arXiv, 2024. 3", + "[63] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3D object synthesis. arXiv, 2023. 3", + "[64] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3D: High-quality and efficient 3D mesh generation from a single image. arXiv, 2024. 3", + "[65] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[66] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie E. Kaufman, Xin Sun, and Hao Tan. LRM-Zero: Training large reconstruction models with synthesized data. arXiv, 2024. 3", + "[67] Rengan Xie, Wenting Zheng, Kai Huang, Yizheng Chen, Qi Wang, Qi Ye, Wei Chen, and Yuchi Huo. LDM: Large tensorial SDF model for textured mesh generation. arXiv, 2024. 3", + "[68] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. 2", + "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: Efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2024. 3, 7", + "[70] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multi-view diffusion using 3D large reconstruction model. arXiv, 2023. 3", + "[71] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2024. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In 3DV, 2024. 3", + "[73] Jonathan Young. xatlas. https://github.com/jpcy/xatlas, 2024. 2", + "[74] Meng yue Li, Xiaoxiao Long, Yixun Liang, Weiyu Li, Yuan Liu, Peng Li, Xiaowei Chi, Xingqun Qi, Wei Xue, Wenhan Luo, Qi fei Liu, and Yike Guo. M-LRM: Multi-view large reconstruction model. arXiv, 2024. 3", + "[75] Greg Zaal. Poly haven, 2024. https://polyhaven.com/. 8", + "[76] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. CVPR, 2021. 3", + "[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv, 2024. 3", + "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6", + "[79] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. DreamMat: High-quality PBR material generation with geometry- and light-aware diffusion models. arXiv, 2024. 3", + "[80] Ruowen Zhao, Zhengyi Wang, Yikai Wang, Zihan Zhou, and Jun Zhu. FlexiDreamer: Single image-to-3D generation with flexicubes. arXiv, 2024. 3", + "[81] Chuanxia Zheng and Andrea Vedaldi. Free3D: Consistent novel view synthesis without 3D representation. arXiv, 2023. 3", + "[82] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, S. Tulyakov, and Hsin-Ying Lee. GTR: Improving large 3D reconstruction models through geometry and texture refinement. arXiv, 2024. 3", + "[83] Zi-Xin Zou, Zhipeng Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 672 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16250", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_model.json b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d5bbbc3519e7476fc85e8b1348826dce4a5866d --- /dev/null +++ b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.13, + 0.838, + 0.177 + ], + "angle": 0, + "content": "SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.21, + 0.777, + 0.269 + ], + "angle": 0, + "content": "Mark Boss1 Zixuan Huang1,2† Aaryaman Vasishta1 Varun Jampani1 \n1Stability AI ²UIUC" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.303, + 0.832, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.615, + 0.895, + 0.645 + ], + "angle": 0, + "content": "Figure 1. SF3D Creates High-quality Object Meshes from Single Images with Materials, Delighting and UV-unwrapped textured meshes in \\(0.3\\mathrm{s}\\). Here, we show sample SF3D results from different input images. SF3D handles both realistic and non-realistic styles well." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.657, + 0.314, + 0.673 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.871 + ], + "angle": 0, + "content": "We present SF3D, a novel method for rapid and high-quality textured object mesh reconstruction from a single image in just 0.3 seconds. Unlike most existing approaches, SF3D is explicitly trained for mesh generation, incorporating a fast UV unwrapping technique that enables swift texture generation rather than relying on vertex colors. The method also learns to predict material parameters and normal maps to enhance the visual quality of the reconstructed 3D meshes. Furthermore, SF3D integrates a delightting step to effectively remove low-frequency illumination effects, ensuring that the reconstructed meshes can be easily used in novel illumination conditions. Experiments demonstrate the su" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.658, + 0.881, + 0.674 + ], + "angle": 0, + "content": "perior performance of SF3D over the existing techniques." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.675, + 0.894, + 0.704 + ], + "angle": 0, + "content": "Project page with code and model: https://stable-fast-3d.github.io" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.722, + 0.633, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.87 + ], + "angle": 0, + "content": "High-quality object meshes are essential for various use cases in movies, gaming, e-commerce, and AR/VR. In this work, we tackle the problem of generating high-quality 3D object mesh from a single image. This is an ill-posed and challenging problem as this requires reasoning about the object's 3D shape and texture from only a single 2D projection (image) of that object. Single-image object generation can simplify the tedious and manual object creation process." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The quality of object mesh generation from a single image has dramatically improved in the past couple of years" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.336, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done during internship at Stability AI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16240" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.451, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.546 + ], + "angle": 0, + "content": "Figure 2. SF3D Improvements on Different Prevalent Issues. Here, we compare our results with TripoSR [55]. The top shows the effect of light bake-in when relighting the asset. SF3D produces a more plausible relighting (see input in inset). By not using vertex colors, our method is capable of encoding finer details while also having a lower polygon count. Our vertex displacement enables estimating smooth shapes, which do not introduce stair-stepping artifacts from marching cubes. Lastly, our material property prediction allows us to express a variety of different surface types." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.469, + 0.682 + ], + "angle": 0, + "content": "with the advances in transformer models [19, 21, 55], large synthetic datasets [12] and 3D-aware image/video generative models [33, 52, 58, 68]. Especially, the transformer-based reconstruction models [19, 21, 55] demonstrate phenomenal generalization capabilities on real-world images despite being trained only on synthetic datasets while also generating 3D assets from a single image in under 1 s." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.469, + 0.851 + ], + "angle": 0, + "content": "Despite this rapid progress, several issues remain in these feed-forward fast 3D reconstruction models [19, 21, 55]. These techniques often produce 3D assets that are not usable for downstream applications or require laborious manual post-processing. We identify several key issues in these techniques and propose a fast generation technique called 'Stable Fast 3D' (SF3D) that generates higher quality and more usable 3D assets from single images, while also retaining the fast generation speed within 0.3 seconds on a H100 GPU. Next, we briefly introduce these issues and how we tackle those in SF3D." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Light Bake-in. Having shadows or other illumination effects in a given input image is common. Most existing works bake these effects into textures, making the resulting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "3D assets less usable. Having consistent lighting helps in easy integration into graphics pipelines. In SF3D, we propose decomposing the illumination and reflective properties by incorporating explicit illumination and a differentiable shading model using Spherical Gaussians (SG). Fig. 2 (top row) shows a sample result of SF3D, where the light bake-in is considerably reduced compared to the prior art." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.893, + 0.412 + ], + "angle": 0, + "content": "Vertex Coloring. Another issue we found in most 3D generation models is that they produce meshes with a high vertex count, using vertex coloring to represent object texture. This makes the resulting 3D assets inefficient to use in applications such as games. A key issue is the additional processing time of UV unwrapping, which can take longer than the entire object generation. For example, xatlas [73] and geogram [29], can take up to \\(30\\mathrm{s}\\) or \\(10\\mathrm{s}\\) for a single asset, respectively. To tackle this, we propose a highly parallelizable fast box projection-based UV unwrapping technique to achieve a \\(0.5\\mathrm{s}\\) generation time. The effect of relying on vertex coloring vs. UV Unwrapping can be seen in Fig. 2 (middle row), where TripoSR captures fewer details than SF3D despite requiring \\(10\\times\\) higher polygon count." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.893, + 0.595 + ], + "angle": 0, + "content": "Marching Cubes Artifacts. The feed-forward networks often create volumetric representations such as Triplane-NeRFs [9] which are converted to meshes using the Marching Cubes (MC) [37] algorithm. MC can cause 'stair-stepping' artifacts, which can be somewhat reduced by increasing the volume resolution. However, this comes at the cost of large computational overhead. In contrast, SF3D uses more efficient architecture for higher resolution triplanes and also produces meshes using DMTet [48] with learned vertex displacements and normal maps, resulting in smoother mesh surfaces. Fig. 2 demonstrates the smoothness of the SF3D mesh compared to that of TripoSR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Lack of Material Properties. The generations from previous feed-forward techniques often look dull when they are rendered using different illuminations. This is mainly due to the lack of explicit material properties in the output generations, which can influence the light reflection. To tackle this, we predict non-spatially varying material properties. This addition is apparent when rendering different generated objects in Fig. 2 (bottom row)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "With these advances, SF3D can generate high-quality 3D meshes from a single image with several desirable properties for downstream applications on both shapes (low-polygon yet smooth) and textures (Illumination disentangled UV maps with material properties). The 3D assets are small in size (under \\(1\\mathrm{MB}\\)) and can be created in \\(0.5\\mathrm{s}\\). For text-to-3D mesh generation, a fast Text-to-Image (T2I) model [47] can be combined with SF3D to produce meshes in about \\(1\\mathrm{s}\\). Experimental results demonstrate higher quality results with SF3D compared to the existing works. In short, SF3D provides a comprehensive technique for fast, high-quality 3D object generation from single images, ad" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16241" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.463, + 0.108 + ], + "angle": 0, + "content": "dressing both speed and usability in practical applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.126, + 0.22, + 0.143 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.155, + 0.473, + 0.458 + ], + "angle": 0, + "content": "3D Reconstruction using Image Generative Priors. Diffusion models [20, 51] have proven to be powerful generative models for various tasks [2, 3, 16, 45, 46, 57]. Several works such as Zero123 [33] and others [12, 26, 52, 81] leverage the object priors in these diffusion models by adapting the generative models for 3D generation. Score Distillation Sampling (SDS) [43] is often used to optimize 3D representation using the 2D diffusion models. However, it was found that relying solely on the image prior does not always produce consistent multi-view results. This issue is improved in follow-up works [34, 39, 49, 50] by simultaneously generating multiple views of an object. Another approach is to explicitly introduce 3D awareness [31, 32, 72] or use a multi-view diffusion process [2, 15, 27, 35, 36, 38, 50, 58, 63, 80] to generate 3D objects. While diffusion models can generate videos or multi-view images relatively quickly, they require a 3D reconstruction step to create 3D mesh from a single image. Even with fast techniques, generating an object can still take several minutes. Our work focuses on fast generation speeds in \\(0.3\\mathrm{s}\\) for image-to-3D." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.461, + 0.473, + 0.644 + ], + "angle": 0, + "content": "Feed-Forward 3D Reconstruction. Recently, LRM [21] and the follow-up works [19, 55] demonstrated that the fast and reliable 3D generation is possible using feed-forward networks trained on large synthetic datasets. These works use large transformer models to directly predict triplanes as the volume representations, which can be ray-marched using NeRF [40]. This allows these models to train on multi-view datasets, as only the rendering loss is required for training. A set of follow-up works address to remedy the reliance on multi-view datasets [24, 59, 66]. This can produce meshes from images in seconds compared to the generative-prior based approaches." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Several follow-up works emerged which use Gaussian Splatting [25] as the representation [53, 54, 71, 77, 83] or directly integrate a mesh prediction [61, 64, 67, 69, 74, 82]. Another group of methods integrates diffusion models with the feed-forward models by either directly generating the triplanes [28, 60] or 3D [11], using LRM as the denoiser [70], or using it as the conditioning [62]. As single-image reconstruction is a challenging task, several models leverage the extensive prior of multi-view diffusion models to generate multiple views of an object which the feed-forward model then uses to produce a 3D output [30, 54, 61, 64, 67, 69, 71, 74, 77, 82]. Generally, these models learn the scene's radiance, meaning lighting information is baked into the objects. LDM [67] tries to remedy this by learning an additional shading color to capture the shading information. However, training requires having access to the ground-truth albedo color as a supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.183 + ], + "angle": 0, + "content": "signal. Compared to LDM, our method also learns an illumination model, enabling training on regular multi-view datasets, and we predict more material properties. We also tackle fast UV unwrapping, so unlike previous works, we do not have to rely on vertex colors while keeping the 3D generation time short." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.895, + 0.429 + ], + "angle": 0, + "content": "Material Decomposition. Predicting only the radiance of an object has the downside that relighting does not produce convincing results. Current works in single scene optimization are often based on NeRF [40] or Gaussian Splating [25] and decompose multiple input images (\\(>50\\)) into light and materials [4-6, 14, 18, 41, 76]. These methods often predict materials properties of a Physically Based Rendering (PBR) [7] shading model. A few recent works tackle joint 3D shape and material generation, such as UniDream [35] or Fantasia3D [10] by optimizing the material properties using a SDS loss. However, this optimization takes several hours to converge. Another set of works aims to texture existing objects [56, 79] using diffusion models. Our work generates a textured object with homogeneous material properties from a single image under natural illumination at fast generation speeds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.445, + 0.593, + 0.46 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.895, + 0.548 + ], + "angle": 0, + "content": "We propose SF3D, which converts a single object image into a textured and UV-unwrapped 3D model with de-lit albedo and material properties. As explained in the introduction section, with SF3D, we aim to fix the issues shown in Fig. 2 and introduce additional quality enhancements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.748 + ], + "angle": 0, + "content": "Preliminaries. Our method is based on TripoSR [21, 55], which trains a large transformer-based network that outputs a Triplane [9] based 3D representation from a single image. TripoSR is trained with multi-view image datasets without explicit 3D supervision. In TripoSR, the image is encoded using DINO [8] and passed through a Transformer network to generate a 3D triplane at a resolution of \\(64 \\times 64\\). The triplane features are then decoded into RGB colors and rendered using standard NeRF [40] rendering into multiple views for training. TripoSR only learns the view-independent colors and cannot model reflective objects. Several issues of the TripoSR (and other similar networks) are explained in the introduction (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "SF3D Overview. With SF3D, we propose several improvements to TripoSR [55] to improve the output quality in different aspects. As illustrated in Fig. 3, SF3D has 5 main components: 1. An enhanced transformer network that predicts higher resolution triplanes, which helps in reducing aliasing artifacts (top left in the figure); 2. A material estimation network (bottom left) predicts material properties, which helps handle an object's reflective properties. 3. Illumination prediction (bottom right) to tackle illumination disentangling, which helps in outputting homogeneous" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16242" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.096, + 0.831, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.895, + 0.365 + ], + "angle": 0, + "content": "Figure 3. SF3D Overview. SF3D improves on TripoSR by addressing the issues in Fig. 2 with 5 novel modules: 1. An enhanced transformer for higher resolution triplanes (top left); 2. Material estimation with Material Net (bottom left); 3. Explicit illumination estimation using Light Net (bottom right); 4. Smooth mesh extraction with the estimation of vertex offsets and normals (top right); and finally 5. an export pipeline with fast UV-unwrapping (right)." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.384, + 0.446, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.495, + 0.471, + 0.566 + ], + "angle": 0, + "content": "Figure 4. Triplane Resolution Aliasing. We found that low-resolution triplanes struggle with high-frequency and high-contrast textures and produce grid-like aliasing artifacts. Our method increases the triplane size from \\(64^{2}\\) to \\(384^{2}\\) allowing our method to reproduce these textures with fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.701 + ], + "angle": 0, + "content": "objects without shadows; 4. Mesh extraction and refinement with the prediction of vertex offsets and surface normals (top right), which helps in smoother output shapes with fewer mesh extraction artifacts; and 5. A fast UV-unwrapping and export module (right) that helps produce low-poly meshes and high-resolution textures. Next, we explain each of these modules in detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.294, + 0.728 + ], + "angle": 0, + "content": "3.1. Enhanced Transformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "First, as illustrated in Fig. 3, we transitioned from DINO [8] used in TripoSR to the improved DINOv2 [42] network to obtain image tokens for the transformer. We observed that low-resolution \\((64\\times 64)\\) triplanes used in TripoSR and other networks [19, 21] introduce noticeable artifacts, especially in scenarios with high-frequency and high-contrast texture patterns. Fig. 4 (Middle) illustrates these aliasing artifacts. The triplane resolution is directly correlated with the presence of these artifacts, which we identify as an aliasing issue that can be mitigated by increasing the resolution. The increased capacity also improves the geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.602 + ], + "angle": 0, + "content": "Naively increasing the triplane resolution quadratically increases the transformer complexity. We take inspiration from the recent PointInfinity [23] work and propose an enhanced Transformer network that outputs higher-resolution triplanes. PointInfinity proposes an architecture where the complexity remains linear concerning the input size by avoiding the self-attention on the higher resolution triplane tokens. With this addition, we produce \\(96 \\times 96\\) resolution triplanes with 1024 channels. We further increased the triplane resolution by shuffling the output features across dimensions, resulting in 40-channel features at \\(384 \\times 384\\) resolution. Further details about the architecture are available in the supplements. Fig. 4 (Right) shows fewer aliasing artifacts with our higher resolution triplanes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.608, + 0.693, + 0.623 + ], + "angle": 0, + "content": "3.2. Material Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.854 + ], + "angle": 0, + "content": "To enhance the output mesh appearance for the reflective objects, SF3D also outputs the material properties of metallic and roughness parameters. Ideally, one would like to estimate the spatially varying material properties at 3D output locations, but this is an inherently challenging and ill-posed learning problem that requires many high-quality 3D data with spatially varying materials. To overcome these challenges, we propose simplifying the material estimation problem by estimating a single metallic and roughness value for the entire object. Although this non-spatially varying material mainly applies to homogeneous objects, we find that it significantly improves the visual quality of our mesh predictions. Specifically, as illustrated in Fig. 3, we propose 'Material Net' that predicts the metallic and roughness values from the input image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "For pre-training the Material Net, we selected a subset of 3D objects with PBR material properties from the synthetic training dataset and rendered them under different illumi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16243" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.256 + ], + "angle": 0, + "content": "nations and viewpoints. We observe that directly regressing the material values often leads to training collapse, where the network always predicts a roughness value of 0.5 and a metallic value of 0. As a remedy, we propose a probabilistic prediction approach, where we predict the parameters of a Beta distribution and minimize the log-likelihood during training. This stabilizes the training by allowing for uncertainty in this ambiguous material estimation task and prevents the collapse observed with direct regression. During inference and training of SF3D, we do not sample the distribution but calculate the mode of the distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.472, + 0.334 + ], + "angle": 0, + "content": "We implement the Material Net by first passing the image through the frozen CLIP image encoder [44] to extract semantically meaningful latents and pass them through 2 separate MLPs with 3 hidden layers and 512 width to output the parameters for the distributions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.288, + 0.357 + ], + "angle": 0, + "content": "3.3. Illumination Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.473, + 0.662 + ], + "angle": 0, + "content": "We propose explicitly estimating the illumination in the input image to account for varying shadings (e.g., shadows). Otherwise, the 3D outputs would have baked-in illumination effects into their RGB colors, as illustrated in Fig. 2. To this end, we propose a Light Net (Fig. 3 bottom right) that predicts the spherical Gaussian (SG) illumination map from the estimated triplanes. The rationale here is that the triplanes encode the global structure and appearance of the input object and should account for the 3D spatial relationships and changes in illumination over the object surface. We use the \\(96 \\times 96\\) resolution triplanes from the transformer and pass them through 2 CNN layers, followed by a max pool and final MLP with three hidden layers and a feature dimension 512 for all layers. Light Net outputs the grayscale amplitude values for 24 SGs with a Softplus activation to ensure positive values. The axis and sharpness values for these SGs remain fixed and are set up to cover the entire sphere. These amplitude values allow us to implement a deferred physically based rendering approach similar to that used in NeRD [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.472, + 0.8 + ], + "angle": 0, + "content": "Our method also incorporates a lighting demodulation loss \\(\\mathcal{L}_{\\mathrm{Demod}}\\) during the training phase, inspired by the works of Hasselgren et al. [18] and Voleti et al. [58]. This loss function ensures that the lighting on an object with an entirely white albedo closely matches the luminance of the input image. The demodulation loss enforces consistency between the learned illumination and the lighting conditions observed in the training data. This can be seen as a bias to resolve the ambiguity between appearance and shading [1]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.806, + 0.37, + 0.82 + ], + "angle": 0, + "content": "3.4. Mesh Extraction and Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Fig. 3 (top right) illustrates this module. We convert the estimated triplanes into a mesh using a differentiable Marching Tetrahedron (DMTet) [48] technique. As explained in the introduction (Fig. 2), Marching Cubes (MC) usually results in several staircase artifacts on the resulting meshes." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.088, + 0.867, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.184, + 0.895, + 0.227 + ], + "angle": 0, + "content": "Figure 5. Export Pipeline. Our export process starts with the mesh and is followed by UV unwrapping, occupancy and world position baking, material querying, and UV island margins." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.895, + 0.45 + ], + "angle": 0, + "content": "We propose two new MLP heads to refine the meshes as a remedy. One predicts vertex offsets \\( \\boldsymbol{v}_o \\in \\mathbb{R}^3 \\), and another one predicts the world space vertex normals \\( \\hat{\\boldsymbol{n}} \\in \\mathbb{R}^3 \\). Inspired by MeshLRM [61], we also implemented small split decoder MLPs for these two networks, which proved beneficial for performance and efficiency. We found that the vertex offsets can reduce artifacts from the tetrahedral grids, and the world space normals can add details to the flat mesh triangles. Given that the normal predictions are initially unreliable, we stabilize the training by using spherical linear interpolation (slerp) between the geometry normals \\( \\boldsymbol{n} \\in \\mathbb{R}^3 \\) and our predictions. This slerp is used during the initial 5K training steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.617 + ], + "angle": 0, + "content": "To regularize the mesh estimation, we use several training losses: a normal consistency loss \\(\\mathcal{L}_{\\mathrm{Nrm~consistency}}\\), a Laplacian smoothness loss \\(\\mathcal{L}_{\\mathrm{Laplacian}}\\) as implemented by threestudio [17], and a vertex offset regularization \\(\\mathcal{L}_{\\mathrm{Offset}} = \\pmb{v}_o^2\\). For supervising the normal prediction, we use a geometry normal replication loss \\(\\mathcal{L}_{\\mathrm{Nrm~repl}} = 1 - \\pmb{n} \\cdot \\hat{\\pmb{n}}\\), where \\(\\cdot\\) is the dot product and a normal smoothness loss to ensure the smoothness of normal predictions in 3D. This loss is achieved by adding a small offset \\(\\epsilon \\in \\mathbb{R}^3\\) around a query location \\(\\pmb{x} \\in \\mathbb{R}^3\\). The loss is then defined as \\(\\mathcal{L}_{\\mathrm{Nrm~smooth}} = (\\hat{n}(\\pmb{x}) - \\hat{n}(\\pmb{x} + \\epsilon))^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.795, + 0.64 + ], + "angle": 0, + "content": "3.5. Fast UV-Unwrapping and Export" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.894, + 0.78 + ], + "angle": 0, + "content": "The final stage of SF3D is an export pipeline that outputs the final 3D mesh along with the corresponding UV atlas. Our export pipeline follows multiple stages to ensure efficient and effective handling of 3D models. An overview of these stages is provided in Fig. 5, where we first do fast UV-unwrapping. We then bake the world positions and occupancy to the UV atlas, which we use for querying the albedo and normal. This results in the final textured 3D mesh. The entire export process only takes \\(10\\mathrm{ms}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "UV unwrapping is traditionally a computationally intensive process. Existing methods require several seconds for UV unwrapping, which is impractical when we are aiming for sub-second generation speeds. To address this inefficiency, we propose a Cube projection-based unwrapping method. The key advantage of this approach is it is parallelizable: each face of the mesh can independently decide which cube face to project onto, based on its surface normal." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16244" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.091, + 0.445, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Figure 6. UV Unwrapping. Our UV unwrapping technique uses projection mapping, allowing each face to independently select a projection, enabling easy parallelization. A naive approach could lead to the same UV coordinates being assigned to different vertices due to occlusions. We identify potential overlaps from occlusions on the 2D mapped surfaces and relocate them to different areas within the UV atlas. Any remaining areas are placed in the bottom right of the UV atlas. This method minimizes distortion and ensures most surfaces are preserved in a connected area." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.701 + ], + "angle": 0, + "content": "Our UV unwrapping process is illustrated in Fig. 6. We initially align the output mesh based on the most dominant axes with the cube projection coordinate system. After each mesh face selects the appropriate cube direction, we address potential occlusions. Without managing occlusions, different faces could share the same UV coordinates, leading to artifacts in the texture. We detect occlusions in the UV atlas by performing 2D triangle-triangle intersection tests. We filter triangles by their proximity to the triangle centers to make the process efficient. If an intersection is detected, we sort the intersecting triangles based on their depth in the plane, keeping the first intersection and marking the others for reassignment to different UV atlas areas. The first intersection is placed in the top third of the UV atlas and the second intersection is placed in the bottom left area. The remaining triangles are organized into a grid in the bottom right section of the atlas. We also rotate each island to minimize the shading seams by following radial z tangent orientation. We then assign each face to a position in the UV atlas, as illustrated in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Next, we bake the world positions and the occupancy data with UV-unwrapping into the final UV atlas. This allows us to query the world positions within the chart from our triplanes and decode the albedo and surface normals into additional textures. We transform the world-space normal map into a tangent-space normal map using the tangent and bitangent vectors. We add margins to the UV atlases to prevent visible seams at UV island borders. This is achieved through an iterative process: in each iteration, we perform a \\(3 \\times 3\\) partial convolution based on the occupied areas, using the valid regions of the kernel. We then use a \\(3 \\times 3\\) max pooling operation to expand the occupied regions of the UV atlas, placing the mean values in the newly expanded areas" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "while preserving the original regions. This iterative extension ensures that the textures smoothly blend outwards." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.123, + 0.892, + 0.169 + ], + "angle": 0, + "content": "We incorporate our image estimator's metallic and roughness values and pack everything into a GLB file, ready for efficient rendering and use in various applications." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.176, + 0.819, + 0.193 + ], + "angle": 0, + "content": "3.6. Overall Training and Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Directly training our method with mesh rendering yielded unsatisfactory results. Hence, we pre-trained it on the NeRF task. Following this pre-training, we transitioned to mesh training, replacing the NeRF rendering with differentiable mesh rendering and SG-based shading. Given the introduction of light estimation, we found that using larger batch sizes aids convergence. We initiate training with a batch size of 192 and a rendering resolution of \\(128 \\times 128\\), training for 10K steps. In the subsequent stage, we reduce the batch size to 128 and increase the resolution to \\(256 \\times 256\\), continuing for 20K steps. The final stage involves 80K steps at a \\(512 \\times 512\\) resolution with a batch size of 96." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.501 + ], + "angle": 0, + "content": "The loss functions remain consistent across all mesh training stages. We primarily use image-based metrics to compare our rendered and shaded reconstructions \\(\\hat{I}\\) with the GT image \\(I\\). These include MSE \\(\\mathcal{L}_{\\mathrm{MSE}}\\) and LPIPS [78] \\(\\mathcal{L}_{\\mathrm{LPIPS}}\\) losses. We also incorporate a mask loss \\(\\mathcal{L}_{\\mathrm{Mask}}\\) between the GT mask \\(M\\) and the predicted opacity \\(\\hat{M}\\), defined as an MSE loss. We then define three loss formulations for the rendering, mesh regularization, and shading:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.517, + 0.892, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n d e r}} = \\underbrace {\\lambda_ {\\mathrm {M S E}}} _ {1 0} \\mathcal {L} _ {\\mathrm {M S E}} + \\underbrace {\\lambda_ {\\mathrm {L P I P S}}} _ {2} \\mathcal {L} _ {\\mathrm {L P I P S}} + \\underbrace {\\lambda_ {\\mathrm {M a s k}}} _ {1 0} \\mathcal {L} _ {\\mathrm {M a s k}} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.546, + 0.892, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m e s h}} = \\underbrace {\\lambda_ {\\text {L a p l a c i a n}}} _ {0. 0 1} \\mathcal {L} _ {\\text {L a p l a c i a n}} + \\underbrace {\\lambda_ {\\text {N r m C o n s i s t e n c y}}} _ {0. 0 0 1} \\mathcal {L} _ {\\text {N r m c o n s i s t e n c y}} + \\underbrace {\\lambda_ {\\text {O f f s e t}}} _ {0. 1} \\mathcal {L} _ {\\text {O f f s e t}} ^ {(2)}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.574, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s h a d i n g}} = \\underbrace {\\lambda_ {\\mathrm {N r m} \\text {r e p l}}} _ {0. 2} \\mathcal {L} _ {\\mathrm {N r m} \\text {r e p l}} \\underbrace {\\lambda_ {\\mathrm {N r m} \\text {s m o o t h}}} _ {0. 0 2} \\mathcal {L} _ {\\mathrm {N r m} \\text {s m o o t h}} + \\underbrace {\\lambda_ {\\mathrm {D e m o d}}} _ {0. 0 1} \\mathcal {L} _ {\\mathrm {D e m o d}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.612, + 0.682, + 0.626 + ], + "angle": 0, + "content": "The total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.634, + 0.892, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e n d e r}} + \\mathcal {L} _ {\\text {m e s h}} + \\mathcal {L} _ {\\text {s h a d i n g}} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.675, + 0.587, + 0.69 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Datasets. For comparisons, we select GSO [13] and OmniObject3D [65] as our primary datasets for evaluation. We select 278 random scenes from GSO and 308 scenes from OmniObject3D for comparison. We render 16 views around the object and select a frontal as the conditioning view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare with several recent methods for fast 3D object reconstruction from a single image. We mainly focus on fast reconstruction models to maintain a consistent evaluation protocol across different techniques. We also mainly consider meshes as outputs and perform the evaluations on the mesh. We use the official implementation for all the baselines, and we evaluate all the methods under the same protocol. We selected several recent and concurrent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16245" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.086, + 0.882, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.48, + 0.893, + 0.509 + ], + "angle": 0, + "content": "Figure 7. Comparison on GSO and OmniObject3D. Notice how our reconstructions produce consistent results with detailed textures and smooth shading." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.522, + 0.873, + 0.655 + ], + "angle": 0, + "content": "
MethodTime [s]↓GSOOmniObject
3DImage3DImage
CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓
ZeroShape [22]0.90.1600.4890.7590.952———0.1440.5070.7860.969———
OpenLRM [19]2.00.1600.4720.7510.95415.6890.7870.2060.1390.5210.7980.97113.9750.7600.229
TripoSR [55]0.30.1110.6450.8690.98016.4450.7890.1940.1030.6720.8890.98614.3310.7550.224
LGM [54]64.60.1950.3760.6540.92814.3770.7620.2480.2050.3440.6310.92112.6620.7320.273
CRM [60]10.20.1790.4110.6990.94515.0540.7780.2280.1580.4690.7520.96013.4620.7550.245
InstantMesh [69]32.40.1380.5490.8010.96715.4340.7850.2030.1380.5600.8110.96413.5310.7570.235
LN3Diff [28]5.00.1840.4220.6920.93915.4470.7790.2080.1650.4700.7350.95213.2560.7500.243
3DTopia-XL [11]38.30.2010.4100.6700.91614.9730.7800.2230.1970.3970.6700.92612.7100.7530.259
SF3D (Ours)0.30.0960.7100.8970.98817.6190.8210.1520.0890.7310.9220.99015.1360.7850.209
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.66, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Table 1. Comparison on 3D and Image Metrics Demonstrating State-of-the-art Performance of SF3D. It is worth pointing out that all other methods produce meshes with drastically higher polygon counts. This allows them to follow the surfaces more closely than our low polygon meshes. Our meshes can outperform the others by using our learned vertex offsets and high resolution supervision. Our method is also one of fastest to generate a mesh from an image due to our efficient extraction pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.834 + ], + "angle": 0, + "content": "works with source code releases for comparisons. Specifically, we compare against OpenLRM [19], TripoSR [55], LGM [54], CRM [60], InstantMesh [69], ZeroShape [22], 3DTopia-XL [11] and LN3Diff [28]. For OpenLRM, we selected the large Objaverse 1.1 model. We use H100 GPU for comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "**Evaluations.** For runtime comparisons, we consider the meshes as the final output and calculate the entire run time required to go from the input image to the final mesh. We then perform separate evaluations for the shape qual" + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.74, + 0.861, + 0.804 + ], + "angle": 0, + "content": "
MethodGSO
CD↓FS@0.1↑FS@0.2↑FS@0.5↑
TripoSR [55]0.1110.6450.8690.980
SF3D w/o Enh. Transformer0.1080.6600.8720.982
SF3D (Ours)0.0980.7010.8940.988
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.814, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Table 2. Ablation. As our model builds upon TripoSR, we use it as our baseline. Without our enhanced transformer, our mesh training has already improved upon it. With our architectural improvements, we outperform the baselines significantly." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16246" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.086, + 0.437, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.473, + 0.323 + ], + "angle": 0, + "content": "Figure 8. Decomposition Results. Here, we use high-quality objects from Polyhaven [75] and render them under natural illumination. These illuminations are highly challenging for material estimation. Still our model estimates sensible material properties, which allow for a convincing relighting." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.338, + 0.422, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.085, + 0.507, + 0.468, + 0.539 + ], + "angle": 0, + "content": "SF3D OpenLRM LGM InstantMesh ZeroShape TripoSR LN3Diff CRM 3DTopia-XL" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.596 + ], + "angle": 0, + "content": "Figure 9. Image-To-Mesh Time vs. Reconstruction Quality. Our method is not only one of the fastest reconstruction methods, it is also capable of producing highly accurate geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.807 + ], + "angle": 0, + "content": "ity. Several models cannot be conditioned on intrinsic and extrinsic camera parameters, so we propose performing an alignment step. We normalize the mesh, perform brute-force rotation alignment tests, and select the rotation with the lowest Chamfer Distance (CD). We then run another alignment step using Iterative Closest Point (ICP), where we further optimize the rotation and translation. We then calculate the standard shape metrics of CD and F-score (FS). For image metrics we further scale the normalized objects back to the actual GT scale and run another finer ICP optimization to adjust the fine scale. This alignment might still misalign some symmetric objects for re-rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results. In Table 1, we compare our method with all the baselines. Here, our method outperforms all current and concurrent baselines on both 3D metrics such as CD and F-scores and image metrics as PSNR, SSIM, and LPIPS. This indicates that our method can reconstruct accurate shapes and textures, even if our models have fewer polygons than" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "the other baselines. From the visual comparisons in Fig. 7, it can be seen that the accurate shape reconstruction of SF3D also translates well to the visual quality of the 3D assets. Here, it is worth noting that SF3D also handles fine geometry like the glasses well and produces consistent shapes with more detailed textures than the SOTA methods. Moreover, the results also show sensible material properties and albedo as seen in Fig. 8. This is highly challenging as the objects are only rendered under natural illumination. Estimating material properties without any knowledge about the illumination is a highly ambiguous problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Inference Speed vs. Reconstruction Quality. In Fig. 9, we plot the inference speed vs. reconstruction quality for different techniques. The best-performing methods should be located in the top left corner. While our method is slightly slower than TripoSR, the reconstruction accuracy is considerably better for SF3D. It is also worth noting that our reconstruction has smoother shading with less pronounced marching cube artifacts, as seen in Fig. 7. The final apparent asset quality is thus even higher for our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.893, + 0.563 + ], + "angle": 0, + "content": "Ablations. We evaluate our additions in Table 2 against the baseline models. Our method is built upon TripoSR, so we use it as our initial baseline. If we add mesh training and relighting as a fine-tuning step, we can see that 'SF3D w/o Enhanced Transformer' already improves upon TripoSR, demonstrating the use of our mesh-based training. This improvement is mainly due to the efficient rendering, enabling higher resolution supervision during training and smoother meshes from vertex offsets. If we add our high-resolution triplanes using our enhanced transformer, SF3D further outperforms the baseline considerably." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.762 + ], + "angle": 0, + "content": "Limitations and Outlook. As seen in Fig. 7 (top row), the cup's albedo is not perfectly matched. This is related to the LDR input, where the dark spots do not contain any useful information. Furthermore, our roughness and metallic properties are homogeneous, which limits their usefulness for objects containing multiple drastically different materials that are spatially-varying. In addition, our method also introduces material prediction and delighting without explicit supervision. As we do not require explicit supervision of these parameters, our method could be extended to train on real-world datasets, which we leave for future work. Similarly, the UV unwrapping could leverage existing datasets for further improvements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.62, + 0.784 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We present SF3D, a fast single view to uv-unwrapped textured object reconstruction method. In addition to our fast extraction pipeline, we introduce several architectural improvements to feed-forward-based 3D reconstruction methods, which help our model produce highly detailed geometry and texture. In our extensive evaluation, we show that our method outperforms existing and concurrent baselines in both speed and quality." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16247" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.119, + 0.174, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.185 + ], + "angle": 0, + "content": "[1] E.H. Adelson and A.P. Pentland. The perception of shading and reflectance, page 409-424. Cambridge University Press, 1996. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[5] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[7] Brent Burley. Physically based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.468, + 0.578 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. ICCV, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.468, + 0.648 + ], + "angle": 0, + "content": "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[11] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia-xl: High-quality 3d pbr asset generation via primitive diffusion. arXiv, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of \\(10\\mathrm{m} + 3\\mathrm{D}\\) objects. arXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.145, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Andreas Engelhardt, Amit Raj, Mark Boss, Yunzhi Zhang, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SHINOBI: Shape and Illumination using Neural Object decomposition via Brdf optimization In-the-wild. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create anything in 3D with multi-view diffusion models. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[16] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[17] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light & Material Decomposition from Images using Monte Carlo Rendering and Denoising. NeurIPS, 20222. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[21] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. ICLR, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M. Rehg. ZeroShape: Regression-based zero-shot shape reconstruction. arXiv, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[23] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[24] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3D: Scaling up large reconstruction models with real-world images. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[26] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. EscherNet: A generative model for scalable view synthesis. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[27] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. ViVid-1-to-3: Novel view synthesis with video diffusion models. CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Bruno Levy. geogram. https://github.com/BrunoLevy/geogram, 2024.2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16248" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[31] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[32] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[35] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for re-1ightable text-to-3D generation. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[36] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[37] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM Transactions on Graphics (SIGGRAPH), 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[39] Antoine Mercier, Ramin Nakhli, Mahesh Reddy, Rajeev Yasarla, Hong Cai, Fatih Porikli, and Guillaume Berger. HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.72 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.776 + ], + "angle": 0, + "content": "[41] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2d diffusion. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[47] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.432, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[49] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[50] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.771, + 0.603 + ], + "angle": 0, + "content": "[52] StabilityAI. Stable Zero123, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[53] Stanislaw Szymanowicz, C. Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3D reconstruction. CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[54] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. arXiv, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[55] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. TripoSR: Fast 3D object reconstruction from a single image. arXiv, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[56] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned PBR image generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[57] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16249" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. arXiv, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[59] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. PF-LRM: Pose-free large reconstruction model for joint pose and shape prediction. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[60] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. CRM: Single image to 3D textured mesh with convolutional reconstruction model. arXiv, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[61] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[62] Hao Wen, Zehuan Huang, Yaohui Wang, Xinyuan Chen, Yu Qiao, and Lu Sheng. Ouroboros3D: Image-to-3D generation via 3D-aware recursive diffusion. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[63] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3D object synthesis. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[64] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3D: High-quality and efficient 3D mesh generation from a single image. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[65] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[66] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie E. Kaufman, Xin Sun, and Hao Tan. LRM-Zero: Training large reconstruction models with synthesized data. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[67] Rengan Xie, Wenting Zheng, Kai Huang, Yizheng Chen, Qi Wang, Qi Ye, Wei Chen, and Yuchi Huo. LDM: Large tensorial SDF model for textured mesh generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[68] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: Efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[70] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multi-view diffusion using 3D large reconstruction model. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[71] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[72] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In 3DV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[73] Jonathan Young. xatlas. https://github.com/jpcy/xatlas, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[74] Meng yue Li, Xiaoxiao Long, Yixun Liang, Weiyu Li, Yuan Liu, Peng Li, Xiaowei Chi, Xingqun Qi, Wei Xue, Wenhan Luo, Qi fei Liu, and Yike Guo. M-LRM: Multi-view large reconstruction model. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.867, + 0.248 + ], + "angle": 0, + "content": "[75] Greg Zaal. Poly haven, 2024. https://polyhaven.com/. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[76] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[79] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. DreamMat: High-quality PBR material generation with geometry- and light-aware diffusion models. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[80] Ruowen Zhao, Zhengyi Wang, Yikai Wang, Zihan Zhou, and Jun Zhu. FlexiDreamer: Single image-to-3D generation with flexicubes. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[81] Chuanxia Zheng and Andrea Vedaldi. Free3D: Consistent novel view synthesis without 3D representation. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[82] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, S. Tulyakov, and Hsin-Ying Lee. GTR: Improving large 3D reconstruction models through geometry and texture refinement. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[83] Zi-Xin Zou, Zhipeng Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16250" + } + ] +] \ No newline at end of file diff --git a/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_origin.pdf b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f15bddaffac60fe31bf526a7b226d1b634b6fc20 --- /dev/null +++ b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/3b32e463-bad2-49dc-a4f1-d9f41d2a1f1d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cec8ed962a69261e654d23e050792c7be35baed843411c0d7ba08bae6ef5f3e9 +size 4626881 diff --git a/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/full.md b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8438d42f3d70700e8e0640456e7a1de1361f8e87 --- /dev/null +++ b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/full.md @@ -0,0 +1,280 @@ +# SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement + +Mark Boss1 Zixuan Huang1,2† Aaryaman Vasishta1 Varun Jampani1 +1Stability AI ²UIUC + +![](images/3d5377ee19947da1d6960e6092985f4fd109b3d9d83b7984b065abfdab93e9da.jpg) +Figure 1. SF3D Creates High-quality Object Meshes from Single Images with Materials, Delighting and UV-unwrapped textured meshes in $0.3\mathrm{s}$ . Here, we show sample SF3D results from different input images. SF3D handles both realistic and non-realistic styles well. + +# Abstract + +We present SF3D, a novel method for rapid and high-quality textured object mesh reconstruction from a single image in just 0.3 seconds. Unlike most existing approaches, SF3D is explicitly trained for mesh generation, incorporating a fast UV unwrapping technique that enables swift texture generation rather than relying on vertex colors. The method also learns to predict material parameters and normal maps to enhance the visual quality of the reconstructed 3D meshes. Furthermore, SF3D integrates a delightting step to effectively remove low-frequency illumination effects, ensuring that the reconstructed meshes can be easily used in novel illumination conditions. Experiments demonstrate the su + +perior performance of SF3D over the existing techniques. + +Project page with code and model: https://stable-fast-3d.github.io + +# 1. Introduction + +High-quality object meshes are essential for various use cases in movies, gaming, e-commerce, and AR/VR. In this work, we tackle the problem of generating high-quality 3D object mesh from a single image. This is an ill-posed and challenging problem as this requires reasoning about the object's 3D shape and texture from only a single 2D projection (image) of that object. Single-image object generation can simplify the tedious and manual object creation process. + +The quality of object mesh generation from a single image has dramatically improved in the past couple of years + +![](images/bb55ed3e8a5331b0c70279fd98590c49c5c51bad7e934db3bbdd48f0b3f10fa4.jpg) +Figure 2. SF3D Improvements on Different Prevalent Issues. Here, we compare our results with TripoSR [55]. The top shows the effect of light bake-in when relighting the asset. SF3D produces a more plausible relighting (see input in inset). By not using vertex colors, our method is capable of encoding finer details while also having a lower polygon count. Our vertex displacement enables estimating smooth shapes, which do not introduce stair-stepping artifacts from marching cubes. Lastly, our material property prediction allows us to express a variety of different surface types. + +with the advances in transformer models [19, 21, 55], large synthetic datasets [12] and 3D-aware image/video generative models [33, 52, 58, 68]. Especially, the transformer-based reconstruction models [19, 21, 55] demonstrate phenomenal generalization capabilities on real-world images despite being trained only on synthetic datasets while also generating 3D assets from a single image in under 1 s. + +Despite this rapid progress, several issues remain in these feed-forward fast 3D reconstruction models [19, 21, 55]. These techniques often produce 3D assets that are not usable for downstream applications or require laborious manual post-processing. We identify several key issues in these techniques and propose a fast generation technique called 'Stable Fast 3D' (SF3D) that generates higher quality and more usable 3D assets from single images, while also retaining the fast generation speed within 0.3 seconds on a H100 GPU. Next, we briefly introduce these issues and how we tackle those in SF3D. + +Light Bake-in. Having shadows or other illumination effects in a given input image is common. Most existing works bake these effects into textures, making the resulting + +3D assets less usable. Having consistent lighting helps in easy integration into graphics pipelines. In SF3D, we propose decomposing the illumination and reflective properties by incorporating explicit illumination and a differentiable shading model using Spherical Gaussians (SG). Fig. 2 (top row) shows a sample result of SF3D, where the light bake-in is considerably reduced compared to the prior art. + +Vertex Coloring. Another issue we found in most 3D generation models is that they produce meshes with a high vertex count, using vertex coloring to represent object texture. This makes the resulting 3D assets inefficient to use in applications such as games. A key issue is the additional processing time of UV unwrapping, which can take longer than the entire object generation. For example, xatlas [73] and geogram [29], can take up to $30\mathrm{s}$ or $10\mathrm{s}$ for a single asset, respectively. To tackle this, we propose a highly parallelizable fast box projection-based UV unwrapping technique to achieve a $0.5\mathrm{s}$ generation time. The effect of relying on vertex coloring vs. UV Unwrapping can be seen in Fig. 2 (middle row), where TripoSR captures fewer details than SF3D despite requiring $10\times$ higher polygon count. + +Marching Cubes Artifacts. The feed-forward networks often create volumetric representations such as Triplane-NeRFs [9] which are converted to meshes using the Marching Cubes (MC) [37] algorithm. MC can cause 'stair-stepping' artifacts, which can be somewhat reduced by increasing the volume resolution. However, this comes at the cost of large computational overhead. In contrast, SF3D uses more efficient architecture for higher resolution triplanes and also produces meshes using DMTet [48] with learned vertex displacements and normal maps, resulting in smoother mesh surfaces. Fig. 2 demonstrates the smoothness of the SF3D mesh compared to that of TripoSR. + +Lack of Material Properties. The generations from previous feed-forward techniques often look dull when they are rendered using different illuminations. This is mainly due to the lack of explicit material properties in the output generations, which can influence the light reflection. To tackle this, we predict non-spatially varying material properties. This addition is apparent when rendering different generated objects in Fig. 2 (bottom row). + +With these advances, SF3D can generate high-quality 3D meshes from a single image with several desirable properties for downstream applications on both shapes (low-polygon yet smooth) and textures (Illumination disentangled UV maps with material properties). The 3D assets are small in size (under $1\mathrm{MB}$ ) and can be created in $0.5\mathrm{s}$ . For text-to-3D mesh generation, a fast Text-to-Image (T2I) model [47] can be combined with SF3D to produce meshes in about $1\mathrm{s}$ . Experimental results demonstrate higher quality results with SF3D compared to the existing works. In short, SF3D provides a comprehensive technique for fast, high-quality 3D object generation from single images, ad + +dressing both speed and usability in practical applications. + +# 2. Related Work + +3D Reconstruction using Image Generative Priors. Diffusion models [20, 51] have proven to be powerful generative models for various tasks [2, 3, 16, 45, 46, 57]. Several works such as Zero123 [33] and others [12, 26, 52, 81] leverage the object priors in these diffusion models by adapting the generative models for 3D generation. Score Distillation Sampling (SDS) [43] is often used to optimize 3D representation using the 2D diffusion models. However, it was found that relying solely on the image prior does not always produce consistent multi-view results. This issue is improved in follow-up works [34, 39, 49, 50] by simultaneously generating multiple views of an object. Another approach is to explicitly introduce 3D awareness [31, 32, 72] or use a multi-view diffusion process [2, 15, 27, 35, 36, 38, 50, 58, 63, 80] to generate 3D objects. While diffusion models can generate videos or multi-view images relatively quickly, they require a 3D reconstruction step to create 3D mesh from a single image. Even with fast techniques, generating an object can still take several minutes. Our work focuses on fast generation speeds in $0.3\mathrm{s}$ for image-to-3D. + +Feed-Forward 3D Reconstruction. Recently, LRM [21] and the follow-up works [19, 55] demonstrated that the fast and reliable 3D generation is possible using feed-forward networks trained on large synthetic datasets. These works use large transformer models to directly predict triplanes as the volume representations, which can be ray-marched using NeRF [40]. This allows these models to train on multi-view datasets, as only the rendering loss is required for training. A set of follow-up works address to remedy the reliance on multi-view datasets [24, 59, 66]. This can produce meshes from images in seconds compared to the generative-prior based approaches. + +Several follow-up works emerged which use Gaussian Splatting [25] as the representation [53, 54, 71, 77, 83] or directly integrate a mesh prediction [61, 64, 67, 69, 74, 82]. Another group of methods integrates diffusion models with the feed-forward models by either directly generating the triplanes [28, 60] or 3D [11], using LRM as the denoiser [70], or using it as the conditioning [62]. As single-image reconstruction is a challenging task, several models leverage the extensive prior of multi-view diffusion models to generate multiple views of an object which the feed-forward model then uses to produce a 3D output [30, 54, 61, 64, 67, 69, 71, 74, 77, 82]. Generally, these models learn the scene's radiance, meaning lighting information is baked into the objects. LDM [67] tries to remedy this by learning an additional shading color to capture the shading information. However, training requires having access to the ground-truth albedo color as a supervision + +signal. Compared to LDM, our method also learns an illumination model, enabling training on regular multi-view datasets, and we predict more material properties. We also tackle fast UV unwrapping, so unlike previous works, we do not have to rely on vertex colors while keeping the 3D generation time short. + +Material Decomposition. Predicting only the radiance of an object has the downside that relighting does not produce convincing results. Current works in single scene optimization are often based on NeRF [40] or Gaussian Splating [25] and decompose multiple input images ( $>50$ ) into light and materials [4-6, 14, 18, 41, 76]. These methods often predict materials properties of a Physically Based Rendering (PBR) [7] shading model. A few recent works tackle joint 3D shape and material generation, such as UniDream [35] or Fantasia3D [10] by optimizing the material properties using a SDS loss. However, this optimization takes several hours to converge. Another set of works aims to texture existing objects [56, 79] using diffusion models. Our work generates a textured object with homogeneous material properties from a single image under natural illumination at fast generation speeds. + +# 3. Method + +We propose SF3D, which converts a single object image into a textured and UV-unwrapped 3D model with de-lit albedo and material properties. As explained in the introduction section, with SF3D, we aim to fix the issues shown in Fig. 2 and introduce additional quality enhancements. + +Preliminaries. Our method is based on TripoSR [21, 55], which trains a large transformer-based network that outputs a Triplane [9] based 3D representation from a single image. TripoSR is trained with multi-view image datasets without explicit 3D supervision. In TripoSR, the image is encoded using DINO [8] and passed through a Transformer network to generate a 3D triplane at a resolution of $64 \times 64$ . The triplane features are then decoded into RGB colors and rendered using standard NeRF [40] rendering into multiple views for training. TripoSR only learns the view-independent colors and cannot model reflective objects. Several issues of the TripoSR (and other similar networks) are explained in the introduction (Fig. 2). + +SF3D Overview. With SF3D, we propose several improvements to TripoSR [55] to improve the output quality in different aspects. As illustrated in Fig. 3, SF3D has 5 main components: 1. An enhanced transformer network that predicts higher resolution triplanes, which helps in reducing aliasing artifacts (top left in the figure); 2. A material estimation network (bottom left) predicts material properties, which helps handle an object's reflective properties. 3. Illumination prediction (bottom right) to tackle illumination disentangling, which helps in outputting homogeneous + +![](images/8e457eeb0d8dda08547ac97ef40a15a5594e2a94da8d22d76bddadea36e23eb5.jpg) +Figure 3. SF3D Overview. SF3D improves on TripoSR by addressing the issues in Fig. 2 with 5 novel modules: 1. An enhanced transformer for higher resolution triplanes (top left); 2. Material estimation with Material Net (bottom left); 3. Explicit illumination estimation using Light Net (bottom right); 4. Smooth mesh extraction with the estimation of vertex offsets and normals (top right); and finally 5. an export pipeline with fast UV-unwrapping (right). + +![](images/b51713e89043b01dcb251614b23dcaad55a142fbb1d850fd567c320081e1ef17.jpg) +Figure 4. Triplane Resolution Aliasing. We found that low-resolution triplanes struggle with high-frequency and high-contrast textures and produce grid-like aliasing artifacts. Our method increases the triplane size from $64^{2}$ to $384^{2}$ allowing our method to reproduce these textures with fewer artifacts. + +objects without shadows; 4. Mesh extraction and refinement with the prediction of vertex offsets and surface normals (top right), which helps in smoother output shapes with fewer mesh extraction artifacts; and 5. A fast UV-unwrapping and export module (right) that helps produce low-poly meshes and high-resolution textures. Next, we explain each of these modules in detail. + +# 3.1. Enhanced Transformer + +First, as illustrated in Fig. 3, we transitioned from DINO [8] used in TripoSR to the improved DINOv2 [42] network to obtain image tokens for the transformer. We observed that low-resolution $(64\times 64)$ triplanes used in TripoSR and other networks [19, 21] introduce noticeable artifacts, especially in scenarios with high-frequency and high-contrast texture patterns. Fig. 4 (Middle) illustrates these aliasing artifacts. The triplane resolution is directly correlated with the presence of these artifacts, which we identify as an aliasing issue that can be mitigated by increasing the resolution. The increased capacity also improves the geometry. + +Naively increasing the triplane resolution quadratically increases the transformer complexity. We take inspiration from the recent PointInfinity [23] work and propose an enhanced Transformer network that outputs higher-resolution triplanes. PointInfinity proposes an architecture where the complexity remains linear concerning the input size by avoiding the self-attention on the higher resolution triplane tokens. With this addition, we produce $96 \times 96$ resolution triplanes with 1024 channels. We further increased the triplane resolution by shuffling the output features across dimensions, resulting in 40-channel features at $384 \times 384$ resolution. Further details about the architecture are available in the supplements. Fig. 4 (Right) shows fewer aliasing artifacts with our higher resolution triplanes. + +# 3.2. Material Estimation + +To enhance the output mesh appearance for the reflective objects, SF3D also outputs the material properties of metallic and roughness parameters. Ideally, one would like to estimate the spatially varying material properties at 3D output locations, but this is an inherently challenging and ill-posed learning problem that requires many high-quality 3D data with spatially varying materials. To overcome these challenges, we propose simplifying the material estimation problem by estimating a single metallic and roughness value for the entire object. Although this non-spatially varying material mainly applies to homogeneous objects, we find that it significantly improves the visual quality of our mesh predictions. Specifically, as illustrated in Fig. 3, we propose 'Material Net' that predicts the metallic and roughness values from the input image. + +For pre-training the Material Net, we selected a subset of 3D objects with PBR material properties from the synthetic training dataset and rendered them under different illumi + +nations and viewpoints. We observe that directly regressing the material values often leads to training collapse, where the network always predicts a roughness value of 0.5 and a metallic value of 0. As a remedy, we propose a probabilistic prediction approach, where we predict the parameters of a Beta distribution and minimize the log-likelihood during training. This stabilizes the training by allowing for uncertainty in this ambiguous material estimation task and prevents the collapse observed with direct regression. During inference and training of SF3D, we do not sample the distribution but calculate the mode of the distribution. + +We implement the Material Net by first passing the image through the frozen CLIP image encoder [44] to extract semantically meaningful latents and pass them through 2 separate MLPs with 3 hidden layers and 512 width to output the parameters for the distributions. + +# 3.3. Illumination Modeling + +We propose explicitly estimating the illumination in the input image to account for varying shadings (e.g., shadows). Otherwise, the 3D outputs would have baked-in illumination effects into their RGB colors, as illustrated in Fig. 2. To this end, we propose a Light Net (Fig. 3 bottom right) that predicts the spherical Gaussian (SG) illumination map from the estimated triplanes. The rationale here is that the triplanes encode the global structure and appearance of the input object and should account for the 3D spatial relationships and changes in illumination over the object surface. We use the $96 \times 96$ resolution triplanes from the transformer and pass them through 2 CNN layers, followed by a max pool and final MLP with three hidden layers and a feature dimension 512 for all layers. Light Net outputs the grayscale amplitude values for 24 SGs with a Softplus activation to ensure positive values. The axis and sharpness values for these SGs remain fixed and are set up to cover the entire sphere. These amplitude values allow us to implement a deferred physically based rendering approach similar to that used in NeRD [4]. + +Our method also incorporates a lighting demodulation loss $\mathcal{L}_{\mathrm{Demod}}$ during the training phase, inspired by the works of Hasselgren et al. [18] and Voleti et al. [58]. This loss function ensures that the lighting on an object with an entirely white albedo closely matches the luminance of the input image. The demodulation loss enforces consistency between the learned illumination and the lighting conditions observed in the training data. This can be seen as a bias to resolve the ambiguity between appearance and shading [1]. + +# 3.4. Mesh Extraction and Refinement + +Fig. 3 (top right) illustrates this module. We convert the estimated triplanes into a mesh using a differentiable Marching Tetrahedron (DMTet) [48] technique. As explained in the introduction (Fig. 2), Marching Cubes (MC) usually results in several staircase artifacts on the resulting meshes. + +![](images/8f75041f11cccf95da2243a4605439b01227e21f35790fd6850b4fb2c31de017.jpg) +Figure 5. Export Pipeline. Our export process starts with the mesh and is followed by UV unwrapping, occupancy and world position baking, material querying, and UV island margins. + +We propose two new MLP heads to refine the meshes as a remedy. One predicts vertex offsets $\boldsymbol{v}_o \in \mathbb{R}^3$ , and another one predicts the world space vertex normals $\hat{\boldsymbol{n}} \in \mathbb{R}^3$ . Inspired by MeshLRM [61], we also implemented small split decoder MLPs for these two networks, which proved beneficial for performance and efficiency. We found that the vertex offsets can reduce artifacts from the tetrahedral grids, and the world space normals can add details to the flat mesh triangles. Given that the normal predictions are initially unreliable, we stabilize the training by using spherical linear interpolation (slerp) between the geometry normals $\boldsymbol{n} \in \mathbb{R}^3$ and our predictions. This slerp is used during the initial 5K training steps. + +To regularize the mesh estimation, we use several training losses: a normal consistency loss $\mathcal{L}_{\mathrm{Nrm~consistency}}$ , a Laplacian smoothness loss $\mathcal{L}_{\mathrm{Laplacian}}$ as implemented by threestudio [17], and a vertex offset regularization $\mathcal{L}_{\mathrm{Offset}} = \pmb{v}_o^2$ . For supervising the normal prediction, we use a geometry normal replication loss $\mathcal{L}_{\mathrm{Nrm~repl}} = 1 - \pmb{n} \cdot \hat{\pmb{n}}$ , where $\cdot$ is the dot product and a normal smoothness loss to ensure the smoothness of normal predictions in 3D. This loss is achieved by adding a small offset $\epsilon \in \mathbb{R}^3$ around a query location $\pmb{x} \in \mathbb{R}^3$ . The loss is then defined as $\mathcal{L}_{\mathrm{Nrm~smooth}} = (\hat{n}(\pmb{x}) - \hat{n}(\pmb{x} + \epsilon))^2$ . + +# 3.5. Fast UV-Unwrapping and Export + +The final stage of SF3D is an export pipeline that outputs the final 3D mesh along with the corresponding UV atlas. Our export pipeline follows multiple stages to ensure efficient and effective handling of 3D models. An overview of these stages is provided in Fig. 5, where we first do fast UV-unwrapping. We then bake the world positions and occupancy to the UV atlas, which we use for querying the albedo and normal. This results in the final textured 3D mesh. The entire export process only takes $10\mathrm{ms}$ . + +UV unwrapping is traditionally a computationally intensive process. Existing methods require several seconds for UV unwrapping, which is impractical when we are aiming for sub-second generation speeds. To address this inefficiency, we propose a Cube projection-based unwrapping method. The key advantage of this approach is it is parallelizable: each face of the mesh can independently decide which cube face to project onto, based on its surface normal. + +![](images/15ef6c79262304687f563ffec52e7cc444a03f990d3a8c9a49b3b125a6386e9c.jpg) +Figure 6. UV Unwrapping. Our UV unwrapping technique uses projection mapping, allowing each face to independently select a projection, enabling easy parallelization. A naive approach could lead to the same UV coordinates being assigned to different vertices due to occlusions. We identify potential overlaps from occlusions on the 2D mapped surfaces and relocate them to different areas within the UV atlas. Any remaining areas are placed in the bottom right of the UV atlas. This method minimizes distortion and ensures most surfaces are preserved in a connected area. + +Our UV unwrapping process is illustrated in Fig. 6. We initially align the output mesh based on the most dominant axes with the cube projection coordinate system. After each mesh face selects the appropriate cube direction, we address potential occlusions. Without managing occlusions, different faces could share the same UV coordinates, leading to artifacts in the texture. We detect occlusions in the UV atlas by performing 2D triangle-triangle intersection tests. We filter triangles by their proximity to the triangle centers to make the process efficient. If an intersection is detected, we sort the intersecting triangles based on their depth in the plane, keeping the first intersection and marking the others for reassignment to different UV atlas areas. The first intersection is placed in the top third of the UV atlas and the second intersection is placed in the bottom left area. The remaining triangles are organized into a grid in the bottom right section of the atlas. We also rotate each island to minimize the shading seams by following radial z tangent orientation. We then assign each face to a position in the UV atlas, as illustrated in Fig. 6. + +Next, we bake the world positions and the occupancy data with UV-unwrapping into the final UV atlas. This allows us to query the world positions within the chart from our triplanes and decode the albedo and surface normals into additional textures. We transform the world-space normal map into a tangent-space normal map using the tangent and bitangent vectors. We add margins to the UV atlases to prevent visible seams at UV island borders. This is achieved through an iterative process: in each iteration, we perform a $3 \times 3$ partial convolution based on the occupied areas, using the valid regions of the kernel. We then use a $3 \times 3$ max pooling operation to expand the occupied regions of the UV atlas, placing the mean values in the newly expanded areas + +while preserving the original regions. This iterative extension ensures that the textures smoothly blend outwards. + +We incorporate our image estimator's metallic and roughness values and pack everything into a GLB file, ready for efficient rendering and use in various applications. + +# 3.6. Overall Training and Loss Functions + +Directly training our method with mesh rendering yielded unsatisfactory results. Hence, we pre-trained it on the NeRF task. Following this pre-training, we transitioned to mesh training, replacing the NeRF rendering with differentiable mesh rendering and SG-based shading. Given the introduction of light estimation, we found that using larger batch sizes aids convergence. We initiate training with a batch size of 192 and a rendering resolution of $128 \times 128$ , training for 10K steps. In the subsequent stage, we reduce the batch size to 128 and increase the resolution to $256 \times 256$ , continuing for 20K steps. The final stage involves 80K steps at a $512 \times 512$ resolution with a batch size of 96. + +The loss functions remain consistent across all mesh training stages. We primarily use image-based metrics to compare our rendered and shaded reconstructions $\hat{I}$ with the GT image $I$ . These include MSE $\mathcal{L}_{\mathrm{MSE}}$ and LPIPS [78] $\mathcal{L}_{\mathrm{LPIPS}}$ losses. We also incorporate a mask loss $\mathcal{L}_{\mathrm{Mask}}$ between the GT mask $M$ and the predicted opacity $\hat{M}$ , defined as an MSE loss. We then define three loss formulations for the rendering, mesh regularization, and shading: + +$$ +\mathcal {L} _ {\text {r e n d e r}} = \underbrace {\lambda_ {\mathrm {M S E}}} _ {1 0} \mathcal {L} _ {\mathrm {M S E}} + \underbrace {\lambda_ {\mathrm {L P I P S}}} _ {2} \mathcal {L} _ {\mathrm {L P I P S}} + \underbrace {\lambda_ {\mathrm {M a s k}}} _ {1 0} \mathcal {L} _ {\mathrm {M a s k}} \tag {1} +$$ + +$$ +\mathcal {L} _ {\text {m e s h}} = \underbrace {\lambda_ {\text {L a p l a c i a n}}} _ {0. 0 1} \mathcal {L} _ {\text {L a p l a c i a n}} + \underbrace {\lambda_ {\text {N r m C o n s i s t e n c y}}} _ {0. 0 0 1} \mathcal {L} _ {\text {N r m c o n s i s t e n c y}} + \underbrace {\lambda_ {\text {O f f s e t}}} _ {0. 1} \mathcal {L} _ {\text {O f f s e t}} ^ {(2)} +$$ + +$$ +\mathcal {L} _ {\text {s h a d i n g}} = \underbrace {\lambda_ {\mathrm {N r m} \text {r e p l}}} _ {0. 2} \mathcal {L} _ {\mathrm {N r m} \text {r e p l}} \underbrace {\lambda_ {\mathrm {N r m} \text {s m o o t h}}} _ {0. 0 2} \mathcal {L} _ {\mathrm {N r m} \text {s m o o t h}} + \underbrace {\lambda_ {\mathrm {D e m o d}}} _ {0. 0 1} \mathcal {L} _ {\mathrm {D e m o d}} \tag {3} +$$ + +The total loss is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e n d e r}} + \mathcal {L} _ {\text {m e s h}} + \mathcal {L} _ {\text {s h a d i n g}} \tag {4} +$$ + +# 4. Results + +Datasets. For comparisons, we select GSO [13] and OmniObject3D [65] as our primary datasets for evaluation. We select 278 random scenes from GSO and 308 scenes from OmniObject3D for comparison. We render 16 views around the object and select a frontal as the conditioning view. + +Baselines. We compare with several recent methods for fast 3D object reconstruction from a single image. We mainly focus on fast reconstruction models to maintain a consistent evaluation protocol across different techniques. We also mainly consider meshes as outputs and perform the evaluations on the mesh. We use the official implementation for all the baselines, and we evaluate all the methods under the same protocol. We selected several recent and concurrent + +![](images/4b03ab237c238fb37ae07d73bb4f24653335b01c0ddde8b1007f4ee16df720c8.jpg) +Figure 7. Comparison on GSO and OmniObject3D. Notice how our reconstructions produce consistent results with detailed textures and smooth shading. + +
MethodTime [s]↓GSOOmniObject
3DImage3DImage
CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓
ZeroShape [22]0.90.1600.4890.7590.952———0.1440.5070.7860.969———
OpenLRM [19]2.00.1600.4720.7510.95415.6890.7870.2060.1390.5210.7980.97113.9750.7600.229
TripoSR [55]0.30.1110.6450.8690.98016.4450.7890.1940.1030.6720.8890.98614.3310.7550.224
LGM [54]64.60.1950.3760.6540.92814.3770.7620.2480.2050.3440.6310.92112.6620.7320.273
CRM [60]10.20.1790.4110.6990.94515.0540.7780.2280.1580.4690.7520.96013.4620.7550.245
InstantMesh [69]32.40.1380.5490.8010.96715.4340.7850.2030.1380.5600.8110.96413.5310.7570.235
LN3Diff [28]5.00.1840.4220.6920.93915.4470.7790.2080.1650.4700.7350.95213.2560.7500.243
3DTopia-XL [11]38.30.2010.4100.6700.91614.9730.7800.2230.1970.3970.6700.92612.7100.7530.259
SF3D (Ours)0.30.0960.7100.8970.98817.6190.8210.1520.0890.7310.9220.99015.1360.7850.209
+ +works with source code releases for comparisons. Specifically, we compare against OpenLRM [19], TripoSR [55], LGM [54], CRM [60], InstantMesh [69], ZeroShape [22], 3DTopia-XL [11] and LN3Diff [28]. For OpenLRM, we selected the large Objaverse 1.1 model. We use H100 GPU for comparisons. + +**Evaluations.** For runtime comparisons, we consider the meshes as the final output and calculate the entire run time required to go from the input image to the final mesh. We then perform separate evaluations for the shape qual + +Table 1. Comparison on 3D and Image Metrics Demonstrating State-of-the-art Performance of SF3D. It is worth pointing out that all other methods produce meshes with drastically higher polygon counts. This allows them to follow the surfaces more closely than our low polygon meshes. Our meshes can outperform the others by using our learned vertex offsets and high resolution supervision. Our method is also one of fastest to generate a mesh from an image due to our efficient extraction pipeline. + +
MethodGSO
CD↓FS@0.1↑FS@0.2↑FS@0.5↑
TripoSR [55]0.1110.6450.8690.980
SF3D w/o Enh. Transformer0.1080.6600.8720.982
SF3D (Ours)0.0980.7010.8940.988
+ +Table 2. Ablation. As our model builds upon TripoSR, we use it as our baseline. Without our enhanced transformer, our mesh training has already improved upon it. With our architectural improvements, we outperform the baselines significantly. + +![](images/29602994d6a88c1e75c40b92cc233387b377e42ed86e7afbd074974d65134e2e.jpg) +Figure 8. Decomposition Results. Here, we use high-quality objects from Polyhaven [75] and render them under natural illumination. These illuminations are highly challenging for material estimation. Still our model estimates sensible material properties, which allow for a convincing relighting. + +Figure 9. Image-To-Mesh Time vs. Reconstruction Quality. Our method is not only one of the fastest reconstruction methods, it is also capable of producing highly accurate geometry. +![](images/10cf8b321af06033b59c6198557b8d460ff516348cee8009eccf8810236bb8e5.jpg) +SF3D OpenLRM LGM InstantMesh ZeroShape TripoSR LN3Diff CRM 3DTopia-XL + +ity. Several models cannot be conditioned on intrinsic and extrinsic camera parameters, so we propose performing an alignment step. We normalize the mesh, perform brute-force rotation alignment tests, and select the rotation with the lowest Chamfer Distance (CD). We then run another alignment step using Iterative Closest Point (ICP), where we further optimize the rotation and translation. We then calculate the standard shape metrics of CD and F-score (FS). For image metrics we further scale the normalized objects back to the actual GT scale and run another finer ICP optimization to adjust the fine scale. This alignment might still misalign some symmetric objects for re-rendering. + +Results. In Table 1, we compare our method with all the baselines. Here, our method outperforms all current and concurrent baselines on both 3D metrics such as CD and F-scores and image metrics as PSNR, SSIM, and LPIPS. This indicates that our method can reconstruct accurate shapes and textures, even if our models have fewer polygons than + +the other baselines. From the visual comparisons in Fig. 7, it can be seen that the accurate shape reconstruction of SF3D also translates well to the visual quality of the 3D assets. Here, it is worth noting that SF3D also handles fine geometry like the glasses well and produces consistent shapes with more detailed textures than the SOTA methods. Moreover, the results also show sensible material properties and albedo as seen in Fig. 8. This is highly challenging as the objects are only rendered under natural illumination. Estimating material properties without any knowledge about the illumination is a highly ambiguous problem. + +Inference Speed vs. Reconstruction Quality. In Fig. 9, we plot the inference speed vs. reconstruction quality for different techniques. The best-performing methods should be located in the top left corner. While our method is slightly slower than TripoSR, the reconstruction accuracy is considerably better for SF3D. It is also worth noting that our reconstruction has smoother shading with less pronounced marching cube artifacts, as seen in Fig. 7. The final apparent asset quality is thus even higher for our method. + +Ablations. We evaluate our additions in Table 2 against the baseline models. Our method is built upon TripoSR, so we use it as our initial baseline. If we add mesh training and relighting as a fine-tuning step, we can see that 'SF3D w/o Enhanced Transformer' already improves upon TripoSR, demonstrating the use of our mesh-based training. This improvement is mainly due to the efficient rendering, enabling higher resolution supervision during training and smoother meshes from vertex offsets. If we add our high-resolution triplanes using our enhanced transformer, SF3D further outperforms the baseline considerably. + +Limitations and Outlook. As seen in Fig. 7 (top row), the cup's albedo is not perfectly matched. This is related to the LDR input, where the dark spots do not contain any useful information. Furthermore, our roughness and metallic properties are homogeneous, which limits their usefulness for objects containing multiple drastically different materials that are spatially-varying. In addition, our method also introduces material prediction and delighting without explicit supervision. As we do not require explicit supervision of these parameters, our method could be extended to train on real-world datasets, which we leave for future work. Similarly, the UV unwrapping could leverage existing datasets for further improvements. + +# 5. Conclusion + +We present SF3D, a fast single view to uv-unwrapped textured object reconstruction method. In addition to our fast extraction pipeline, we introduce several architectural improvements to feed-forward-based 3D reconstruction methods, which help our model produce highly detailed geometry and texture. In our extensive evaluation, we show that our method outperforms existing and concurrent baselines in both speed and quality. + +# References + +[1] E.H. Adelson and A.P. Pentland. The perception of shading and reflectance, page 409-424. Cambridge University Press, 1996. 5 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv, 2023. 3 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv, 2023. 3 +[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 3, 5 +[5] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021. +[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. NeurIPS, 2022. 3 +[7] Brent Burley. Physically based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 3 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. ICCV, 2021. 3, 4 +[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3 +[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 3 +[11] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia-xl: High-quality 3d pbr asset generation via primitive diffusion. arXiv, 2024. 3, 7 +[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of $10\mathrm{m} + 3\mathrm{D}$ objects. arXiv, 2023. 2, 3 +[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6 + +[14] Andreas Engelhardt, Amit Raj, Mark Boss, Yunzhi Zhang, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SHINOBI: Shape and Illumination using Neural Object decomposition via Brdf optimization In-the-wild. In CVPR, 2024. 3 +[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create anything in 3D with multi-view diffusion models. arXiv, 2024. 3 +[16] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning, 2023. 3 +[17] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5 +[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light & Material Decomposition from Images using Monte Carlo Rendering and Denoising. NeurIPS, 20222. 3, 5 +[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2, 3, 4, 7 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[21] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. ICLR, 2024. 2, 3, 4 +[22] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M. Rehg. ZeroShape: Regression-based zero-shot shape reconstruction. arXiv, 2023. 7 +[23] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In CVPR, 2024. 4 +[24] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3D: Scaling up large reconstruction models with real-world images. arXiv, 2024. 3 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4), 2023. 3 +[26] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. EscherNet: A generative model for scalable view synthesis. arXiv, 2024. 3 +[27] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. ViVid-1-to-3: Novel view synthesis with video diffusion models. CVPR, 2024. 3 +[28] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 3, 7 +[29] Bruno Levy. geogram. https://github.com/BrunoLevy/geogram, 2024.2 + +[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. arXiv, 2023. 3 +[31] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv, 2023. 3 +[32] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. NeurIPS, 2023. 3 +[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. ICCV, 2023. 2, 3 +[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv, 2023. 3 +[35] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for re-1ightable text-to-3D generation. arXiv, 2023. 3 +[36] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. arXiv, 2023. 3 +[37] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM Transactions on Graphics (SIGGRAPH), 1987. 2 +[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv, 2024. 3 +[39] Antoine Mercier, Ramin Nakhli, Mahesh Reddy, Rajeev Yasarla, Hong Cai, Fatih Porikli, and Guillaume Berger. HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv, 2024. 3 +[40] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. ECCV, 2020. 3 +[41] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 3 +[42] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4 + +[43] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2d diffusion. arXiv, 2022. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3 +[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv, 2022. 3 +[47] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv, 2023. 2 +[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 5 +[49] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv, 2023. 3 +[50] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv, 2023. 3 +[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv, 2020. 3 +[52] StabilityAI. Stable Zero123, 2023. 2, 3 +[53] Stanislaw Szymanowicz, C. Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3D reconstruction. CVPR, 2024. 3 +[54] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. arXiv, 2024. 3, 7 +[55] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. TripoSR: Fast 3D object reconstruction from a single image. arXiv, 2024. 2, 3, 7 +[56] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned PBR image generation. arXiv, 2024. 3 +[57] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 3 +[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin + +Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. arXiv, 2024. 2, 3, 5 +[59] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. PF-LRM: Pose-free large reconstruction model for joint pose and shape prediction. arXiv, 2023. 3 +[60] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. CRM: Single image to 3D textured mesh with convolutional reconstruction model. arXiv, 2024. 3, 7 +[61] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv, 2024. 3, 5 +[62] Hao Wen, Zehuan Huang, Yaohui Wang, Xinyuan Chen, Yu Qiao, and Lu Sheng. Ouroboros3D: Image-to-3D generation via 3D-aware recursive diffusion. arXiv, 2024. 3 +[63] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3D object synthesis. arXiv, 2023. 3 +[64] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3D: High-quality and efficient 3D mesh generation from a single image. arXiv, 2024. 3 +[65] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[66] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie E. Kaufman, Xin Sun, and Hao Tan. LRM-Zero: Training large reconstruction models with synthesized data. arXiv, 2024. 3 +[67] Rengan Xie, Wenting Zheng, Kai Huang, Yizheng Chen, Qi Wang, Qi Ye, Wei Chen, and Yuchi Huo. LDM: Large tensorial SDF model for textured mesh generation. arXiv, 2024. 3 +[68] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. 2 +[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: Efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2024. 3, 7 +[70] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multi-view diffusion using 3D large reconstruction model. arXiv, 2023. 3 +[71] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2024. 3 + +[72] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In 3DV, 2024. 3 +[73] Jonathan Young. xatlas. https://github.com/jpcy/xatlas, 2024. 2 +[74] Meng yue Li, Xiaoxiao Long, Yixun Liang, Weiyu Li, Yuan Liu, Peng Li, Xiaowei Chi, Xingqun Qi, Wei Xue, Wenhan Luo, Qi fei Liu, and Yike Guo. M-LRM: Multi-view large reconstruction model. arXiv, 2024. 3 +[75] Greg Zaal. Poly haven, 2024. https://polyhaven.com/. 8 +[76] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. CVPR, 2021. 3 +[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv, 2024. 3 +[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6 +[79] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. DreamMat: High-quality PBR material generation with geometry- and light-aware diffusion models. arXiv, 2024. 3 +[80] Ruowen Zhao, Zhengyi Wang, Yikai Wang, Zihan Zhou, and Jun Zhu. FlexiDreamer: Single image-to-3D generation with flexicubes. arXiv, 2024. 3 +[81] Chuanxia Zheng and Andrea Vedaldi. Free3D: Consistent novel view synthesis without 3D representation. arXiv, 2023. 3 +[82] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, S. Tulyakov, and Hsin-Ying Lee. GTR: Improving large 3D reconstruction models through geometry and texture refinement. arXiv, 2024. 3 +[83] Zi-Xin Zou, Zhipeng Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv, 2023. 3 \ No newline at end of file diff --git a/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/images.zip b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..06f62e71f37632817506d9c4271040ec080c8e5b --- /dev/null +++ b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:675101b36ac8529e47880da1063d04c22254cb84ada17fd3e06ceaf137fe8dc7 +size 528398 diff --git a/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/layout.json b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..63b2670a06341daff5c308b9aaab57e22d1d0466 --- /dev/null +++ b/2025/SF3D_ Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement/layout.json @@ -0,0 +1,7621 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 512, + 140 + ], + "type": "text", + "content": "SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 166, + 475, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 166, + 475, + 213 + ], + "spans": [ + { + "bbox": [ + 116, + 166, + 475, + 213 + ], + "type": "text", + "content": "Mark Boss1 Zixuan Huang1,2† Aaryaman Vasishta1 Varun Jampani1 \n1Stability AI ²UIUC" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 86, + 239, + 509, + 478 + ], + "blocks": [ + { + "bbox": [ + 86, + 239, + 509, + 478 + ], + "lines": [ + { + "bbox": [ + 86, + 239, + 509, + 478 + ], + "spans": [ + { + "bbox": [ + 86, + 239, + 509, + 478 + ], + "type": "image", + "image_path": "3d5377ee19947da1d6960e6092985f4fd109b3d9d83b7984b065abfdab93e9da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 487, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 45, + 487, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 45, + 487, + 547, + 510 + ], + "type": "text", + "content": "Figure 1. SF3D Creates High-quality Object Meshes from Single Images with Materials, Delighting and UV-unwrapped textured meshes in " + }, + { + "bbox": [ + 45, + 487, + 547, + 510 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{s}" + }, + { + "bbox": [ + 45, + 487, + 547, + 510 + ], + "type": "text", + "content": ". Here, we show sample SF3D results from different input images. SF3D handles both realistic and non-realistic styles well." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "spans": [ + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 689 + ], + "type": "text", + "content": "We present SF3D, a novel method for rapid and high-quality textured object mesh reconstruction from a single image in just 0.3 seconds. Unlike most existing approaches, SF3D is explicitly trained for mesh generation, incorporating a fast UV unwrapping technique that enables swift texture generation rather than relying on vertex colors. The method also learns to predict material parameters and normal maps to enhance the visual quality of the reconstructed 3D meshes. Furthermore, SF3D integrates a delightting step to effectively remove low-frequency illumination effects, ensuring that the reconstructed meshes can be easily used in novel illumination conditions. Experiments demonstrate the su" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 521, + 539, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 539, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 539, + 533 + ], + "type": "text", + "content": "perior performance of SF3D over the existing techniques." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 534, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 547, + 557 + ], + "type": "text", + "content": "Project page with code and model: https://stable-fast-3d.github.io" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 571, + 387, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 387, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 387, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 689 + ], + "type": "text", + "content": "High-quality object meshes are essential for various use cases in movies, gaming, e-commerce, and AR/VR. In this work, we tackle the problem of generating high-quality 3D object mesh from a single image. This is an ill-posed and challenging problem as this requires reasoning about the object's 3D shape and texture from only a single 2D projection (image) of that object. Single-image object generation can simplify the tedious and manual object creation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "The quality of object mesh generation from a single image has dramatically improved in the past couple of years" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 205, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 205, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 205, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 702, + 205, + 713 + ], + "type": "text", + "content": " Work done during internship at Stability AI." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16240" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 276, + 315 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 276, + 315 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 276, + 315 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 276, + 315 + ], + "type": "image", + "image_path": "bb55ed3e8a5331b0c70279fd98590c49c5c51bad7e934db3bbdd48f0b3f10fa4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 287, + 432 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 432 + ], + "type": "text", + "content": "Figure 2. SF3D Improvements on Different Prevalent Issues. Here, we compare our results with TripoSR [55]. The top shows the effect of light bake-in when relighting the asset. SF3D produces a more plausible relighting (see input in inset). By not using vertex colors, our method is capable of encoding finer details while also having a lower polygon count. Our vertex displacement enables estimating smooth shapes, which do not introduce stair-stepping artifacts from marching cubes. Lastly, our material property prediction allows us to express a variety of different surface types." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": "with the advances in transformer models [19, 21, 55], large synthetic datasets [12] and 3D-aware image/video generative models [33, 52, 58, 68]. Especially, the transformer-based reconstruction models [19, 21, 55] demonstrate phenomenal generalization capabilities on real-world images despite being trained only on synthetic datasets while also generating 3D assets from a single image in under 1 s." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 543, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 673 + ], + "type": "text", + "content": "Despite this rapid progress, several issues remain in these feed-forward fast 3D reconstruction models [19, 21, 55]. These techniques often produce 3D assets that are not usable for downstream applications or require laborious manual post-processing. We identify several key issues in these techniques and propose a fast generation technique called 'Stable Fast 3D' (SF3D) that generates higher quality and more usable 3D assets from single images, while also retaining the fast generation speed within 0.3 seconds on a H100 GPU. Next, we briefly introduce these issues and how we tackle those in SF3D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "Light Bake-in. Having shadows or other illumination effects in a given input image is common. Most existing works bake these effects into textures, making the resulting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "3D assets less usable. Having consistent lighting helps in easy integration into graphics pipelines. In SF3D, we propose decomposing the illumination and reflective properties by incorporating explicit illumination and a differentiable shading model using Spherical Gaussians (SG). Fig. 2 (top row) shows a sample result of SF3D, where the light bake-in is considerably reduced compared to the prior art." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "text", + "content": "Vertex Coloring. Another issue we found in most 3D generation models is that they produce meshes with a high vertex count, using vertex coloring to represent object texture. This makes the resulting 3D assets inefficient to use in applications such as games. A key issue is the additional processing time of UV unwrapping, which can take longer than the entire object generation. For example, xatlas [73] and geogram [29], can take up to " + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "inline_equation", + "content": "30\\mathrm{s}" + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "inline_equation", + "content": "10\\mathrm{s}" + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "text", + "content": " for a single asset, respectively. To tackle this, we propose a highly parallelizable fast box projection-based UV unwrapping technique to achieve a " + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{s}" + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "text", + "content": " generation time. The effect of relying on vertex coloring vs. UV Unwrapping can be seen in Fig. 2 (middle row), where TripoSR captures fewer details than SF3D despite requiring " + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 304, + 158, + 546, + 326 + ], + "type": "text", + "content": " higher polygon count." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 327, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 471 + ], + "type": "text", + "content": "Marching Cubes Artifacts. The feed-forward networks often create volumetric representations such as Triplane-NeRFs [9] which are converted to meshes using the Marching Cubes (MC) [37] algorithm. MC can cause 'stair-stepping' artifacts, which can be somewhat reduced by increasing the volume resolution. However, this comes at the cost of large computational overhead. In contrast, SF3D uses more efficient architecture for higher resolution triplanes and also produces meshes using DMTet [48] with learned vertex displacements and normal maps, resulting in smoother mesh surfaces. Fig. 2 demonstrates the smoothness of the SF3D mesh compared to that of TripoSR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 473, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 568 + ], + "type": "text", + "content": "Lack of Material Properties. The generations from previous feed-forward techniques often look dull when they are rendered using different illuminations. This is mainly due to the lack of explicit material properties in the output generations, which can influence the light reflection. To tackle this, we predict non-spatially varying material properties. This addition is apparent when rendering different generated objects in Fig. 2 (bottom row)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "With these advances, SF3D can generate high-quality 3D meshes from a single image with several desirable properties for downstream applications on both shapes (low-polygon yet smooth) and textures (Illumination disentangled UV maps with material properties). The 3D assets are small in size (under " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{MB}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ") and can be created in " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{s}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". For text-to-3D mesh generation, a fast Text-to-Image (T2I) model [47] can be combined with SF3D to produce meshes in about " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{s}" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ". Experimental results demonstrate higher quality results with SF3D compared to the existing works. In short, SF3D provides a comprehensive technique for fast, high-quality 3D object generation from single images, ad" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16241" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 283, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 283, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 283, + 85 + ], + "type": "text", + "content": "dressing both speed and usability in practical applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 134, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 134, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 134, + 113 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 122, + 289, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 122, + 289, + 362 + ], + "spans": [ + { + "bbox": [ + 47, + 122, + 289, + 362 + ], + "type": "text", + "content": "3D Reconstruction using Image Generative Priors. Diffusion models [20, 51] have proven to be powerful generative models for various tasks [2, 3, 16, 45, 46, 57]. Several works such as Zero123 [33] and others [12, 26, 52, 81] leverage the object priors in these diffusion models by adapting the generative models for 3D generation. Score Distillation Sampling (SDS) [43] is often used to optimize 3D representation using the 2D diffusion models. However, it was found that relying solely on the image prior does not always produce consistent multi-view results. This issue is improved in follow-up works [34, 39, 49, 50] by simultaneously generating multiple views of an object. Another approach is to explicitly introduce 3D awareness [31, 32, 72] or use a multi-view diffusion process [2, 15, 27, 35, 36, 38, 50, 58, 63, 80] to generate 3D objects. While diffusion models can generate videos or multi-view images relatively quickly, they require a 3D reconstruction step to create 3D mesh from a single image. Even with fast techniques, generating an object can still take several minutes. Our work focuses on fast generation speeds in " + }, + { + "bbox": [ + 47, + 122, + 289, + 362 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{s}" + }, + { + "bbox": [ + 47, + 122, + 289, + 362 + ], + "type": "text", + "content": " for image-to-3D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 289, + 510 + ], + "type": "text", + "content": "Feed-Forward 3D Reconstruction. Recently, LRM [21] and the follow-up works [19, 55] demonstrated that the fast and reliable 3D generation is possible using feed-forward networks trained on large synthetic datasets. These works use large transformer models to directly predict triplanes as the volume representations, which can be ray-marched using NeRF [40]. This allows these models to train on multi-view datasets, as only the rendering loss is required for training. A set of follow-up works address to remedy the reliance on multi-view datasets [24, 59, 66]. This can produce meshes from images in seconds compared to the generative-prior based approaches." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 289, + 715 + ], + "type": "text", + "content": "Several follow-up works emerged which use Gaussian Splatting [25] as the representation [53, 54, 71, 77, 83] or directly integrate a mesh prediction [61, 64, 67, 69, 74, 82]. Another group of methods integrates diffusion models with the feed-forward models by either directly generating the triplanes [28, 60] or 3D [11], using LRM as the denoiser [70], or using it as the conditioning [62]. As single-image reconstruction is a challenging task, several models leverage the extensive prior of multi-view diffusion models to generate multiple views of an object which the feed-forward model then uses to produce a 3D output [30, 54, 61, 64, 67, 69, 71, 74, 77, 82]. Generally, these models learn the scene's radiance, meaning lighting information is baked into the objects. LDM [67] tries to remedy this by learning an additional shading color to capture the shading information. However, training requires having access to the ground-truth albedo color as a supervision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "signal. Compared to LDM, our method also learns an illumination model, enabling training on regular multi-view datasets, and we predict more material properties. We also tackle fast UV unwrapping, so unlike previous works, we do not have to rely on vertex colors while keeping the 3D generation time short." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 146, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 547, + 339 + ], + "type": "text", + "content": "Material Decomposition. Predicting only the radiance of an object has the downside that relighting does not produce convincing results. Current works in single scene optimization are often based on NeRF [40] or Gaussian Splating [25] and decompose multiple input images (" + }, + { + "bbox": [ + 304, + 146, + 547, + 339 + ], + "type": "inline_equation", + "content": ">50" + }, + { + "bbox": [ + 304, + 146, + 547, + 339 + ], + "type": "text", + "content": ") into light and materials [4-6, 14, 18, 41, 76]. These methods often predict materials properties of a Physically Based Rendering (PBR) [7] shading model. A few recent works tackle joint 3D shape and material generation, such as UniDream [35] or Fantasia3D [10] by optimizing the material properties using a SDS loss. However, this optimization takes several hours to converge. Another set of works aims to texture existing objects [56, 79] using diffusion models. Our work generates a textured object with homogeneous material properties from a single image under natural illumination at fast generation speeds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 352, + 362, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 362, + 364 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 362, + 364 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 373, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 434 + ], + "type": "text", + "content": "We propose SF3D, which converts a single object image into a textured and UV-unwrapped 3D model with de-lit albedo and material properties. As explained in the introduction section, with SF3D, we aim to fix the issues shown in Fig. 2 and introduce additional quality enhancements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 435, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 592 + ], + "type": "text", + "content": "Preliminaries. Our method is based on TripoSR [21, 55], which trains a large transformer-based network that outputs a Triplane [9] based 3D representation from a single image. TripoSR is trained with multi-view image datasets without explicit 3D supervision. In TripoSR, the image is encoded using DINO [8] and passed through a Transformer network to generate a 3D triplane at a resolution of " + }, + { + "bbox": [ + 304, + 435, + 547, + 592 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 435, + 547, + 592 + ], + "type": "text", + "content": ". The triplane features are then decoded into RGB colors and rendered using standard NeRF [40] rendering into multiple views for training. TripoSR only learns the view-independent colors and cannot model reflective objects. Several issues of the TripoSR (and other similar networks) are explained in the introduction (Fig. 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "SF3D Overview. With SF3D, we propose several improvements to TripoSR [55] to improve the output quality in different aspects. As illustrated in Fig. 3, SF3D has 5 main components: 1. An enhanced transformer network that predicts higher resolution triplanes, which helps in reducing aliasing artifacts (top left in the figure); 2. A material estimation network (bottom left) predicts material properties, which helps handle an object's reflective properties. 3. Illumination prediction (bottom right) to tackle illumination disentangling, which helps in outputting homogeneous" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16242" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 76, + 508, + 238 + ], + "blocks": [ + { + "bbox": [ + 72, + 76, + 508, + 238 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 508, + 238 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 508, + 238 + ], + "type": "image", + "image_path": "8e457eeb0d8dda08547ac97ef40a15a5594e2a94da8d22d76bddadea36e23eb5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 289 + ], + "type": "text", + "content": "Figure 3. SF3D Overview. SF3D improves on TripoSR by addressing the issues in Fig. 2 with 5 novel modules: 1. An enhanced transformer for higher resolution triplanes (top left); 2. Material estimation with Material Net (bottom left); 3. Explicit illumination estimation using Light Net (bottom right); 4. Smooth mesh extraction with the estimation of vertex offsets and normals (top right); and finally 5. an export pipeline with fast UV-unwrapping (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 304, + 272, + 381 + ], + "blocks": [ + { + "bbox": [ + 65, + 304, + 272, + 381 + ], + "lines": [ + { + "bbox": [ + 65, + 304, + 272, + 381 + ], + "spans": [ + { + "bbox": [ + 65, + 304, + 272, + 381 + ], + "type": "image", + "image_path": "b51713e89043b01dcb251614b23dcaad55a142fbb1d850fd567c320081e1ef17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "type": "text", + "content": "Figure 4. Triplane Resolution Aliasing. We found that low-resolution triplanes struggle with high-frequency and high-contrast textures and produce grid-like aliasing artifacts. Our method increases the triplane size from " + }, + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "type": "inline_equation", + "content": "64^{2}" + }, + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "type": "inline_equation", + "content": "384^{2}" + }, + { + "bbox": [ + 46, + 392, + 288, + 448 + ], + "type": "text", + "content": " allowing our method to reproduce these textures with fewer artifacts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 555 + ], + "type": "text", + "content": "objects without shadows; 4. Mesh extraction and refinement with the prediction of vertex offsets and surface normals (top right), which helps in smoother output shapes with fewer mesh extraction artifacts; and 5. A fast UV-unwrapping and export module (right) that helps produce low-poly meshes and high-resolution textures. Next, we explain each of these modules in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 564, + 179, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 179, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 179, + 576 + ], + "type": "text", + "content": "3.1. Enhanced Transformer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "First, as illustrated in Fig. 3, we transitioned from DINO [8] used in TripoSR to the improved DINOv2 [42] network to obtain image tokens for the transformer. We observed that low-resolution " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "(64\\times 64)" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " triplanes used in TripoSR and other networks [19, 21] introduce noticeable artifacts, especially in scenarios with high-frequency and high-contrast texture patterns. Fig. 4 (Middle) illustrates these aliasing artifacts. The triplane resolution is directly correlated with the presence of these artifacts, which we identify as an aliasing issue that can be mitigated by increasing the resolution. The increased capacity also improves the geometry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "type": "text", + "content": "Naively increasing the triplane resolution quadratically increases the transformer complexity. We take inspiration from the recent PointInfinity [23] work and propose an enhanced Transformer network that outputs higher-resolution triplanes. PointInfinity proposes an architecture where the complexity remains linear concerning the input size by avoiding the self-attention on the higher resolution triplane tokens. With this addition, we produce " + }, + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "type": "text", + "content": " resolution triplanes with 1024 channels. We further increased the triplane resolution by shuffling the output features across dimensions, resulting in 40-channel features at " + }, + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 308, + 547, + 476 + ], + "type": "text", + "content": " resolution. Further details about the architecture are available in the supplements. Fig. 4 (Right) shows fewer aliasing artifacts with our higher resolution triplanes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 481, + 424, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 424, + 493 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 424, + 493 + ], + "type": "text", + "content": "3.2. Material Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 676 + ], + "type": "text", + "content": "To enhance the output mesh appearance for the reflective objects, SF3D also outputs the material properties of metallic and roughness parameters. Ideally, one would like to estimate the spatially varying material properties at 3D output locations, but this is an inherently challenging and ill-posed learning problem that requires many high-quality 3D data with spatially varying materials. To overcome these challenges, we propose simplifying the material estimation problem by estimating a single metallic and roughness value for the entire object. Although this non-spatially varying material mainly applies to homogeneous objects, we find that it significantly improves the visual quality of our mesh predictions. Specifically, as illustrated in Fig. 3, we propose 'Material Net' that predicts the metallic and roughness values from the input image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "For pre-training the Material Net, we selected a subset of 3D objects with PBR material properties from the synthetic training dataset and rendered them under different illumi" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16243" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 202 + ], + "type": "text", + "content": "nations and viewpoints. We observe that directly regressing the material values often leads to training collapse, where the network always predicts a roughness value of 0.5 and a metallic value of 0. As a remedy, we propose a probabilistic prediction approach, where we predict the parameters of a Beta distribution and minimize the log-likelihood during training. This stabilizes the training by allowing for uncertainty in this ambiguous material estimation task and prevents the collapse observed with direct regression. During inference and training of SF3D, we do not sample the distribution but calculate the mode of the distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 264 + ], + "type": "text", + "content": "We implement the Material Net by first passing the image through the frozen CLIP image encoder [44] to extract semantically meaningful latents and pass them through 2 separate MLPs with 3 hidden layers and 512 width to output the parameters for the distributions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 269, + 176, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 176, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 176, + 282 + ], + "type": "text", + "content": "3.3. Illumination Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 285, + 289, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 524 + ], + "type": "text", + "content": "We propose explicitly estimating the illumination in the input image to account for varying shadings (e.g., shadows). Otherwise, the 3D outputs would have baked-in illumination effects into their RGB colors, as illustrated in Fig. 2. To this end, we propose a Light Net (Fig. 3 bottom right) that predicts the spherical Gaussian (SG) illumination map from the estimated triplanes. The rationale here is that the triplanes encode the global structure and appearance of the input object and should account for the 3D spatial relationships and changes in illumination over the object surface. We use the " + }, + { + "bbox": [ + 46, + 285, + 289, + 524 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 46, + 285, + 289, + 524 + ], + "type": "text", + "content": " resolution triplanes from the transformer and pass them through 2 CNN layers, followed by a max pool and final MLP with three hidden layers and a feature dimension 512 for all layers. Light Net outputs the grayscale amplitude values for 24 SGs with a Softplus activation to ensure positive values. The axis and sharpness values for these SGs remain fixed and are set up to cover the entire sphere. These amplitude values allow us to implement a deferred physically based rendering approach similar to that used in NeRD [4]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 525, + 288, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 633 + ], + "type": "text", + "content": "Our method also incorporates a lighting demodulation loss " + }, + { + "bbox": [ + 46, + 525, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Demod}}" + }, + { + "bbox": [ + 46, + 525, + 288, + 633 + ], + "type": "text", + "content": " during the training phase, inspired by the works of Hasselgren et al. [18] and Voleti et al. [58]. This loss function ensures that the lighting on an object with an entirely white albedo closely matches the luminance of the input image. The demodulation loss enforces consistency between the learned illumination and the lighting conditions observed in the training data. This can be seen as a bias to resolve the ambiguity between appearance and shading [1]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 638, + 226, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 226, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 226, + 649 + ], + "type": "text", + "content": "3.4. Mesh Extraction and Refinement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Fig. 3 (top right) illustrates this module. We convert the estimated triplanes into a mesh using a differentiable Marching Tetrahedron (DMTet) [48] technique. As explained in the introduction (Fig. 2), Marching Cubes (MC) usually results in several staircase artifacts on the resulting meshes." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 530, + 141 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 530, + 141 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 530, + 141 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 530, + 141 + ], + "type": "image", + "image_path": "8f75041f11cccf95da2243a4605439b01227e21f35790fd6850b4fb2c31de017.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 145, + 547, + 179 + ], + "lines": [ + { + "bbox": [ + 305, + 145, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 547, + 179 + ], + "type": "text", + "content": "Figure 5. Export Pipeline. Our export process starts with the mesh and is followed by UV unwrapping, occupancy and world position baking, material querying, and UV island margins." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "text", + "content": "We propose two new MLP heads to refine the meshes as a remedy. One predicts vertex offsets " + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_o \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "text", + "content": ", and another one predicts the world space vertex normals " + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{n}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "text", + "content": ". Inspired by MeshLRM [61], we also implemented small split decoder MLPs for these two networks, which proved beneficial for performance and efficiency. We found that the vertex offsets can reduce artifacts from the tetrahedral grids, and the world space normals can add details to the flat mesh triangles. Given that the normal predictions are initially unreliable, we stabilize the training by using spherical linear interpolation (slerp) between the geometry normals " + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "inline_equation", + "content": "\\boldsymbol{n} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 200, + 547, + 356 + ], + "type": "text", + "content": " and our predictions. This slerp is used during the initial 5K training steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": "To regularize the mesh estimation, we use several training losses: a normal consistency loss " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Nrm~consistency}}" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": ", a Laplacian smoothness loss " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Laplacian}}" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": " as implemented by threestudio [17], and a vertex offset regularization " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Offset}} = \\pmb{v}_o^2" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": ". For supervising the normal prediction, we use a geometry normal replication loss " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Nrm~repl}} = 1 - \\pmb{n} \\cdot \\hat{\\pmb{n}}" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": " is the dot product and a normal smoothness loss to ensure the smoothness of normal predictions in 3D. This loss is achieved by adding a small offset " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": " around a query location " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": ". The loss is then defined as " + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Nrm~smooth}} = (\\hat{n}(\\pmb{x}) - \\hat{n}(\\pmb{x} + \\epsilon))^2" + }, + { + "bbox": [ + 304, + 357, + 547, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 494, + 486, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 486, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 486, + 506 + ], + "type": "text", + "content": "3.5. Fast UV-Unwrapping and Export" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 617 + ], + "type": "text", + "content": "The final stage of SF3D is an export pipeline that outputs the final 3D mesh along with the corresponding UV atlas. Our export pipeline follows multiple stages to ensure efficient and effective handling of 3D models. An overview of these stages is provided in Fig. 5, where we first do fast UV-unwrapping. We then bake the world positions and occupancy to the UV atlas, which we use for querying the albedo and normal. This results in the final textured 3D mesh. The entire export process only takes " + }, + { + "bbox": [ + 304, + 510, + 547, + 617 + ], + "type": "inline_equation", + "content": "10\\mathrm{ms}" + }, + { + "bbox": [ + 304, + 510, + 547, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "UV unwrapping is traditionally a computationally intensive process. Existing methods require several seconds for UV unwrapping, which is impractical when we are aiming for sub-second generation speeds. To address this inefficiency, we propose a Cube projection-based unwrapping method. The key advantage of this approach is it is parallelizable: each face of the mesh can independently decide which cube face to project onto, based on its surface normal." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16244" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 272, + 183 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 272, + 183 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 272, + 183 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 272, + 183 + ], + "type": "image", + "image_path": "15ef6c79262304687f563ffec52e7cc444a03f990d3a8c9a49b3b125a6386e9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 289, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 289, + 288 + ], + "type": "text", + "content": "Figure 6. UV Unwrapping. Our UV unwrapping technique uses projection mapping, allowing each face to independently select a projection, enabling easy parallelization. A naive approach could lead to the same UV coordinates being assigned to different vertices due to occlusions. We identify potential overlaps from occlusions on the 2D mapped surfaces and relocate them to different areas within the UV atlas. Any remaining areas are placed in the bottom right of the UV atlas. This method minimizes distortion and ensures most surfaces are preserved in a connected area." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 555 + ], + "type": "text", + "content": "Our UV unwrapping process is illustrated in Fig. 6. We initially align the output mesh based on the most dominant axes with the cube projection coordinate system. After each mesh face selects the appropriate cube direction, we address potential occlusions. Without managing occlusions, different faces could share the same UV coordinates, leading to artifacts in the texture. We detect occlusions in the UV atlas by performing 2D triangle-triangle intersection tests. We filter triangles by their proximity to the triangle centers to make the process efficient. If an intersection is detected, we sort the intersecting triangles based on their depth in the plane, keeping the first intersection and marking the others for reassignment to different UV atlas areas. The first intersection is placed in the top third of the UV atlas and the second intersection is placed in the bottom left area. The remaining triangles are organized into a grid in the bottom right section of the atlas. We also rotate each island to minimize the shading seams by following radial z tangent orientation. We then assign each face to a position in the UV atlas, as illustrated in Fig. 6." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Next, we bake the world positions and the occupancy data with UV-unwrapping into the final UV atlas. This allows us to query the world positions within the chart from our triplanes and decode the albedo and surface normals into additional textures. We transform the world-space normal map into a tangent-space normal map using the tangent and bitangent vectors. We add margins to the UV atlases to prevent visible seams at UV island borders. This is achieved through an iterative process: in each iteration, we perform a " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " partial convolution based on the occupied areas, using the valid regions of the kernel. We then use a " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": " max pooling operation to expand the occupied regions of the UV atlas, placing the mean values in the newly expanded areas" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "while preserving the original regions. This iterative extension ensures that the textures smoothly blend outwards." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 545, + 133 + ], + "type": "text", + "content": "We incorporate our image estimator's metallic and roughness values and pack everything into a GLB file, ready for efficient rendering and use in various applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 139, + 501, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 139, + 501, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 139, + 501, + 152 + ], + "type": "text", + "content": "3.6. Overall Training and Loss Functions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "text", + "content": "Directly training our method with mesh rendering yielded unsatisfactory results. Hence, we pre-trained it on the NeRF task. Following this pre-training, we transitioned to mesh training, replacing the NeRF rendering with differentiable mesh rendering and SG-based shading. Given the introduction of light estimation, we found that using larger batch sizes aids convergence. We initiate training with a batch size of 192 and a rendering resolution of " + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "text", + "content": ", training for 10K steps. In the subsequent stage, we reduce the batch size to 128 and increase the resolution to " + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "text", + "content": ", continuing for 20K steps. The final stage involves 80K steps at a " + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 156, + 545, + 299 + ], + "type": "text", + "content": " resolution with a batch size of 96." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": "The loss functions remain consistent across all mesh training stages. We primarily use image-based metrics to compare our rendered and shaded reconstructions " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": " with the GT image " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": ". These include MSE " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": " and LPIPS [78] " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": " losses. We also incorporate a mask loss " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Mask}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": " between the GT mask " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": " and the predicted opacity " + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": ", defined as an MSE loss. We then define three loss formulations for the rendering, mesh regularization, and shading:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 409, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 409, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 319, + 409, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n d e r}} = \\underbrace {\\lambda_ {\\mathrm {M S E}}} _ {1 0} \\mathcal {L} _ {\\mathrm {M S E}} + \\underbrace {\\lambda_ {\\mathrm {L P I P S}}} _ {2} \\mathcal {L} _ {\\mathrm {L P I P S}} + \\underbrace {\\lambda_ {\\mathrm {M a s k}}} _ {1 0} \\mathcal {L} _ {\\mathrm {M a s k}} \\tag {1}", + "image_path": "425fcf200551d433990bbd0f41d42b0f37b063e764a71d7be1fc26b1f08797ab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 432, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 432, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 323, + 432, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m e s h}} = \\underbrace {\\lambda_ {\\text {L a p l a c i a n}}} _ {0. 0 1} \\mathcal {L} _ {\\text {L a p l a c i a n}} + \\underbrace {\\lambda_ {\\text {N r m C o n s i s t e n c y}}} _ {0. 0 0 1} \\mathcal {L} _ {\\text {N r m c o n s i s t e n c y}} + \\underbrace {\\lambda_ {\\text {O f f s e t}}} _ {0. 1} \\mathcal {L} _ {\\text {O f f s e t}} ^ {(2)}", + "image_path": "44ccd397133a0e8839a3d30526a85b6b89ce1515e5ac8c09c24264575bcaafca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 454, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 454, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 319, + 454, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s h a d i n g}} = \\underbrace {\\lambda_ {\\mathrm {N r m} \\text {r e p l}}} _ {0. 2} \\mathcal {L} _ {\\mathrm {N r m} \\text {r e p l}} \\underbrace {\\lambda_ {\\mathrm {N r m} \\text {s m o o t h}}} _ {0. 0 2} \\mathcal {L} _ {\\mathrm {N r m} \\text {s m o o t h}} + \\underbrace {\\lambda_ {\\mathrm {D e m o d}}} _ {0. 0 1} \\mathcal {L} _ {\\mathrm {D e m o d}} \\tag {3}", + "image_path": "6f2f63c071e5d32bab3f333c163173f830e27416d861faee23f5f453e1cadaa3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 484, + 417, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 484, + 417, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 417, + 495 + ], + "type": "text", + "content": "The total loss is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 502, + 545, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 502, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 381, + 502, + 545, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e n d e r}} + \\mathcal {L} _ {\\text {m e s h}} + \\mathcal {L} _ {\\text {s h a d i n g}} \\tag {4}", + "image_path": "080dca6918ab0bdc5a11412be5a0dfcdc8fa3228702e1287b13d183d2077d032.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 534, + 359, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 359, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 359, + 546 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 616 + ], + "type": "text", + "content": "Datasets. For comparisons, we select GSO [13] and OmniObject3D [65] as our primary datasets for evaluation. We select 278 random scenes from GSO and 308 scenes from OmniObject3D for comparison. We render 16 views around the object and select a frontal as the conditioning view." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Baselines. We compare with several recent methods for fast 3D object reconstruction from a single image. We mainly focus on fast reconstruction models to maintain a consistent evaluation protocol across different techniques. We also mainly consider meshes as outputs and perform the evaluations on the mesh. We use the official implementation for all the baselines, and we evaluate all the methods under the same protocol. We selected several recent and concurrent" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16245" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 539, + 376 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 539, + 376 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 539, + 376 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 539, + 376 + ], + "type": "image", + "image_path": "4b03ab237c238fb37ae07d73bb4f24653335b01c0ddde8b1007f4ee16df720c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 546, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 546, + 403 + ], + "type": "text", + "content": "Figure 7. Comparison on GSO and OmniObject3D. Notice how our reconstructions produce consistent results with detailed textures and smooth shading." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 413, + 534, + 518 + ], + "blocks": [ + { + "bbox": [ + 59, + 413, + 534, + 518 + ], + "lines": [ + { + "bbox": [ + 59, + 413, + 534, + 518 + ], + "spans": [ + { + "bbox": [ + 59, + 413, + 534, + 518 + ], + "type": "table", + "html": "
MethodTime [s]↓GSOOmniObject
3DImage3DImage
CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓CD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM ↑LPIPS↓
ZeroShape [22]0.90.1600.4890.7590.952———0.1440.5070.7860.969———
OpenLRM [19]2.00.1600.4720.7510.95415.6890.7870.2060.1390.5210.7980.97113.9750.7600.229
TripoSR [55]0.30.1110.6450.8690.98016.4450.7890.1940.1030.6720.8890.98614.3310.7550.224
LGM [54]64.60.1950.3760.6540.92814.3770.7620.2480.2050.3440.6310.92112.6620.7320.273
CRM [60]10.20.1790.4110.6990.94515.0540.7780.2280.1580.4690.7520.96013.4620.7550.245
InstantMesh [69]32.40.1380.5490.8010.96715.4340.7850.2030.1380.5600.8110.96413.5310.7570.235
LN3Diff [28]5.00.1840.4220.6920.93915.4470.7790.2080.1650.4700.7350.95213.2560.7500.243
3DTopia-XL [11]38.30.2010.4100.6700.91614.9730.7800.2230.1970.3970.6700.92612.7100.7530.259
SF3D (Ours)0.30.0960.7100.8970.98817.6190.8210.1520.0890.7310.9220.99015.1360.7850.209
", + "image_path": "a36058dbad440bc448572095f7bb30abce16b718e51769259579f5a1cde4b1a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 588, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 660 + ], + "type": "text", + "content": "works with source code releases for comparisons. Specifically, we compare against OpenLRM [19], TripoSR [55], LGM [54], CRM [60], InstantMesh [69], ZeroShape [22], 3DTopia-XL [11] and LN3Diff [28]. For OpenLRM, we selected the large Objaverse 1.1 model. We use H100 GPU for comparisons." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "**Evaluations.** For runtime comparisons, we consider the meshes as the final output and calculate the entire run time required to go from the input image to the final mesh. We then perform separate evaluations for the shape qual" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 326, + 586, + 526, + 636 + ], + "blocks": [ + { + "bbox": [ + 46, + 522, + 547, + 567 + ], + "lines": [ + { + "bbox": [ + 46, + 522, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 547, + 567 + ], + "type": "text", + "content": "Table 1. Comparison on 3D and Image Metrics Demonstrating State-of-the-art Performance of SF3D. It is worth pointing out that all other methods produce meshes with drastically higher polygon counts. This allows them to follow the surfaces more closely than our low polygon meshes. Our meshes can outperform the others by using our learned vertex offsets and high resolution supervision. Our method is also one of fastest to generate a mesh from an image due to our efficient extraction pipeline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 586, + 526, + 636 + ], + "lines": [ + { + "bbox": [ + 326, + 586, + 526, + 636 + ], + "spans": [ + { + "bbox": [ + 326, + 586, + 526, + 636 + ], + "type": "table", + "html": "
MethodGSO
CD↓FS@0.1↑FS@0.2↑FS@0.5↑
TripoSR [55]0.1110.6450.8690.980
SF3D w/o Enh. Transformer0.1080.6600.8720.982
SF3D (Ours)0.0980.7010.8940.988
", + "image_path": "f8600cb094d6d5391714649ea65554d03a2af9319c6de4e0df81f3b7dd8348da.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 644, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 644, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 644, + 547, + 689 + ], + "type": "text", + "content": "Table 2. Ablation. As our model builds upon TripoSR, we use it as our baseline. Without our enhanced transformer, our mesh training has already improved upon it. With our architectural improvements, we outperform the baselines significantly." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16246" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 68, + 267, + 194 + ], + "blocks": [ + { + "bbox": [ + 63, + 68, + 267, + 194 + ], + "lines": [ + { + "bbox": [ + 63, + 68, + 267, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 68, + 267, + 194 + ], + "type": "image", + "image_path": "29602994d6a88c1e75c40b92cc233387b377e42ed86e7afbd074974d65134e2e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 289, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 289, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 289, + 255 + ], + "type": "text", + "content": "Figure 8. Decomposition Results. Here, we use high-quality objects from Polyhaven [75] and render them under natural illumination. These illuminations are highly challenging for material estimation. Still our model estimates sensible material properties, which allow for a convincing relighting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 267, + 258, + 399 + ], + "blocks": [ + { + "bbox": [ + 50, + 267, + 258, + 399 + ], + "lines": [ + { + "bbox": [ + 50, + 267, + 258, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 267, + 258, + 399 + ], + "type": "image", + "image_path": "10cf8b321af06033b59c6198557b8d460ff516348cee8009eccf8810236bb8e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 401, + 286, + 426 + ], + "lines": [ + { + "bbox": [ + 52, + 401, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 52, + 401, + 286, + 426 + ], + "type": "text", + "content": "SF3D OpenLRM LGM InstantMesh ZeroShape TripoSR LN3Diff CRM 3DTopia-XL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 437, + 287, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 472 + ], + "type": "text", + "content": "Figure 9. Image-To-Mesh Time vs. Reconstruction Quality. Our method is not only one of the fastest reconstruction methods, it is also capable of producing highly accurate geometry." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 495, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 639 + ], + "type": "text", + "content": "ity. Several models cannot be conditioned on intrinsic and extrinsic camera parameters, so we propose performing an alignment step. We normalize the mesh, perform brute-force rotation alignment tests, and select the rotation with the lowest Chamfer Distance (CD). We then run another alignment step using Iterative Closest Point (ICP), where we further optimize the rotation and translation. We then calculate the standard shape metrics of CD and F-score (FS). For image metrics we further scale the normalized objects back to the actual GT scale and run another finer ICP optimization to adjust the fine scale. This alignment might still misalign some symmetric objects for re-rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Results. In Table 1, we compare our method with all the baselines. Here, our method outperforms all current and concurrent baselines on both 3D metrics such as CD and F-scores and image metrics as PSNR, SSIM, and LPIPS. This indicates that our method can reconstruct accurate shapes and textures, even if our models have fewer polygons than" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "the other baselines. From the visual comparisons in Fig. 7, it can be seen that the accurate shape reconstruction of SF3D also translates well to the visual quality of the 3D assets. Here, it is worth noting that SF3D also handles fine geometry like the glasses well and produces consistent shapes with more detailed textures than the SOTA methods. Moreover, the results also show sensible material properties and albedo as seen in Fig. 8. This is highly challenging as the objects are only rendered under natural illumination. Estimating material properties without any knowledge about the illumination is a highly ambiguous problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 205, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 312 + ], + "type": "text", + "content": "Inference Speed vs. Reconstruction Quality. In Fig. 9, we plot the inference speed vs. reconstruction quality for different techniques. The best-performing methods should be located in the top left corner. While our method is slightly slower than TripoSR, the reconstruction accuracy is considerably better for SF3D. It is also worth noting that our reconstruction has smoother shading with less pronounced marching cube artifacts, as seen in Fig. 7. The final apparent asset quality is thus even higher for our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "type": "text", + "content": "Ablations. We evaluate our additions in Table 2 against the baseline models. Our method is built upon TripoSR, so we use it as our initial baseline. If we add mesh training and relighting as a fine-tuning step, we can see that 'SF3D w/o Enhanced Transformer' already improves upon TripoSR, demonstrating the use of our mesh-based training. This improvement is mainly due to the efficient rendering, enabling higher resolution supervision during training and smoother meshes from vertex offsets. If we add our high-resolution triplanes using our enhanced transformer, SF3D further outperforms the baseline considerably." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 546, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 603 + ], + "type": "text", + "content": "Limitations and Outlook. As seen in Fig. 7 (top row), the cup's albedo is not perfectly matched. This is related to the LDR input, where the dark spots do not contain any useful information. Furthermore, our roughness and metallic properties are homogeneous, which limits their usefulness for objects containing multiple drastically different materials that are spatially-varying. In addition, our method also introduces material prediction and delighting without explicit supervision. As we do not require explicit supervision of these parameters, our method could be extended to train on real-world datasets, which we leave for future work. Similarly, the UV unwrapping could leverage existing datasets for further improvements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 609, + 379, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 379, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 379, + 620 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 624, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 714 + ], + "type": "text", + "content": "We present SF3D, a fast single view to uv-unwrapped textured object reconstruction method. In addition to our fast extraction pipeline, we introduce several architectural improvements to feed-forward-based 3D reconstruction methods, which help our model produce highly detailed geometry and texture. In our extensive evaluation, we show that our method outperforms existing and concurrent baselines in both speed and quality." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16247" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 106, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 106, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 106, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 146 + ], + "type": "text", + "content": "[1] E.H. Adelson and A.P. Pentland. The perception of shading and reflectance, page 409-424. Cambridge University Press, 1996. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 201 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable Video Diffusion: Scaling latent video diffusion models to large datasets. arXiv, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. arXiv, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "text", + "content": "[4] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. NeRD: Neural reflectance decomposition from image collections. ICCV, 2021. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 286, + 335 + ], + "type": "text", + "content": "[5] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 286, + 390 + ], + "type": "text", + "content": "[6] Mark Boss, Andreas Engelhardt, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 413 + ], + "type": "text", + "content": "[7] Brent Burley. Physically based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 286, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 286, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 286, + 457 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. ICCV, 2021. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 286, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 286, + 513 + ], + "type": "text", + "content": "[9] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 546 + ], + "type": "text", + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling geometry and appearance for high-quality text-to-3D content creation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 601 + ], + "type": "text", + "content": "[11] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia-xl: High-quality 3d pbr asset generation via primitive diffusion. arXiv, 2024. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "text", + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-XL: A universe of " + }, + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{D}" + }, + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "text", + "content": " objects. arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "text", + "content": "[13] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Andreas Engelhardt, Amit Raj, Mark Boss, Yunzhi Zhang, Abhishek Kar, Yuanzhen Li, Deqing Sun, Jonathan T. Barron, Hendrik P.A. Lensch, and Varun Jampani. SHINOBI: Shape and Illumination using Neural Object decomposition via Brdf optimization In-the-wild. In CVPR, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 181 + ], + "type": "text", + "content": "[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create anything in 3D with multi-view diffusion models. arXiv, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 183, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 237 + ], + "type": "text", + "content": "[16] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. EMU VIDEO: Factorizing Text-to-Video Generation by Explicit Image Conditioning, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 303 + ], + "type": "text", + "content": "[17] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 346 + ], + "type": "text", + "content": "[18] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light & Material Decomposition from Images using Monte Carlo Rendering and Denoising. NeurIPS, 20222. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 349, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 380 + ], + "type": "text", + "content": "[19] Zexin He and Tengfei Wang. OpenLRM: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 403 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 446 + ], + "type": "text", + "content": "[21] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. ICLR, 2024. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "text", + "content": "[22] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M. Rehg. ZeroShape: Regression-based zero-shot shape reconstruction. arXiv, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 514 + ], + "type": "text", + "content": "[23] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In CVPR, 2024. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 547 + ], + "type": "text", + "content": "[24] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3D: Scaling up large reconstruction models with real-world images. arXiv, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 580 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 42(4), 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[26] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J Davison. EscherNet: A generative model for scalable view synthesis. arXiv, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "text", + "content": "[27] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. ViVid-1-to-3: Novel view synthesis with video diffusion models. CVPR, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[28] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, 2024. 3, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 712 + ], + "type": "text", + "content": "[29] Bruno Levy. geogram. https://github.com/BrunoLevy/geogram, 2024.2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16248" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. arXiv, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[31] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3D objects with consistent multi-view generation and 3D diffusion. arXiv, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 227 + ], + "type": "text", + "content": "[32] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3D mesh in 45 seconds without per-shape optimization. NeurIPS, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 261 + ], + "type": "text", + "content": "[33] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[34] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Generating multiview-consistent images from a single-view image. arXiv, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 350 + ], + "type": "text", + "content": "[35] Zexiang Liu, Yangguang Li, Youtian Lin, Xin Yu, Sida Peng, Yan-Pei Cao, Xiaojuan Qi, Xiaoshui Huang, Ding Liang, and Wanli Ouyang. Unidream: Unifying diffusion priors for re-1ightable text-to-3D generation. arXiv, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[36] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3D: Single image to 3D using cross-domain diffusion. arXiv, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "text", + "content": "[37] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM Transactions on Graphics (SIGGRAPH), 1987. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 482 + ], + "type": "text", + "content": "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "text", + "content": "[39] Antoine Mercier, Ramin Nakhli, Mahesh Reddy, Rajeev Yasarla, Hong Cai, Fatih Porikli, and Guillaume Berger. HexaGen3D: Stablediffusion is just one step away from fast and diverse Text-to-3D generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 570 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. ECCV, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 614 + ], + "type": "text", + "content": "[41] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Mueller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. CVPR, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 713 + ], + "type": "text", + "content": "[42] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3D using 2d diffusion. arXiv, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 173 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 262 + ], + "type": "text", + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 262, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 545, + 286 + ], + "type": "text", + "content": "[47] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a hybrid representation for high-resolution 3D shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "text", + "content": "[49] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 420 + ], + "type": "text", + "content": "[50] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3d generation. arXiv, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "text", + "content": "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 465, + 471, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 471, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 471, + 477 + ], + "type": "text", + "content": "[52] StabilityAI. Stable Zero123, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 477, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 510 + ], + "type": "text", + "content": "[53] Stanislaw Szymanowicz, C. Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3D reconstruction. CVPR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": "[54] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. arXiv, 2024. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 555, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 611 + ], + "type": "text", + "content": "[55] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. TripoSR: Fast 3D object reconstruction from a single image. arXiv, 2024. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 611, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 657 + ], + "type": "text", + "content": "[56] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned PBR image generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "type": "text", + "content": "[57] Vikram Voleti, Alexia Jolicoeur-Martineau, and Christopher Pal. MCVD: Masked conditional video diffusion for prediction, generation, and interpolation. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 545, + 713 + ], + "type": "text", + "content": "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16249" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. arXiv, 2024. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "text", + "content": "[59] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. PF-LRM: Pose-free large reconstruction model for joint pose and shape prediction. arXiv, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[60] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. CRM: Single image to 3D textured mesh with convolutional reconstruction model. arXiv, 2024. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "text", + "content": "[61] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: Large reconstruction model for high-quality mesh. arXiv, 2024. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 274 + ], + "type": "text", + "content": "[62] Hao Wen, Zehuan Huang, Yaohui Wang, Xinyuan Chen, Yu Qiao, and Lu Sheng. Ouroboros3D: Image-to-3D generation via 3D-aware recursive diffusion. arXiv, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[63] Haohan Weng, Tianyu Yang, Jianan Wang, Yu Li, Tong Zhang, C. L. Philip Chen, and Lei Zhang. Consistent123: Improve consistency for one image to 3D object synthesis. arXiv, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 365 + ], + "type": "text", + "content": "[64] Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3D: High-quality and efficient 3D mesh generation from a single image. arXiv, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 432 + ], + "type": "text", + "content": "[65] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 476 + ], + "type": "text", + "content": "[66] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie E. Kaufman, Xin Sun, and Hao Tan. LRM-Zero: Training large reconstruction models with synthesized data. arXiv, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "type": "text", + "content": "[67] Rengan Xie, Wenting Zheng, Kai Huang, Yizheng Chen, Qi Wang, Qi Ye, Wei Chen, and Yuchi Huo. LDM: Large tensorial SDF model for textured mesh generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[68] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. Sv4d: Dynamic 3d content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: Efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2024. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[70] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multi-view diffusion using 3D large reconstruction model. arXiv, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[71] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 533 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[72] Jianglong Ye, Peng Wang, Kejie Li, Yichun Shi, and Heng Wang. Consistent-1-to-3: Consistent image to 3D view synthesis via geometry-aware diffusion models. In 3DV, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 139 + ], + "type": "text", + "content": "[73] Jonathan Young. xatlas. https://github.com/jpcy/xatlas, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[74] Meng yue Li, Xiaoxiao Long, Yixun Liang, Weiyu Li, Yuan Liu, Peng Li, Xiaowei Chi, Xingqun Qi, Wei Xue, Wenhan Luo, Qi fei Liu, and Yike Guo. M-LRM: Multi-view large reconstruction model. arXiv, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 530, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 530, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 530, + 196 + ], + "type": "text", + "content": "[75] Greg Zaal. Poly haven, 2024. https://polyhaven.com/. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 241 + ], + "type": "text", + "content": "[76] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical Gaussians for physics-based material editing and relighting. CVPR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 243, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 276 + ], + "type": "text", + "content": "[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D gaussian splatting. arXiv, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 277, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 309 + ], + "type": "text", + "content": "[78] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. CVPR, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "text", + "content": "[79] Yuqing Zhang, Yuan Liu, Zhiyu Xie, Lei Yang, Zhongyuan Liu, Mengzhou Yang, Runze Zhang, Qilong Kou, Cheng Lin, Wenping Wang, and Xiaogang Jin. DreamMat: High-quality PBR material generation with geometry- and light-aware diffusion models. arXiv, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[80] Ruowen Zhao, Zhengyi Wang, Yikai Wang, Zihan Zhou, and Jun Zhu. FlexiDreamer: Single image-to-3D generation with flexicubes. arXiv, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[81] Chuanxia Zheng and Andrea Vedaldi. Free3D: Consistent novel view synthesis without 3D representation. arXiv, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "text", + "content": "[82] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, S. Tulyakov, and Hsin-Ying Lee. GTR: Improving large 3D reconstruction models through geometry and texture refinement. arXiv, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 533 + ], + "type": "text", + "content": "[83] Zi-Xin Zou, Zhipeng Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16250" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_content_list.json b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..898b9b28ea93b02a39b232618ecd6e7abdbd1b2b --- /dev/null +++ b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_content_list.json @@ -0,0 +1,1470 @@ +[ + { + "type": "text", + "text": "SFDM: Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images", + "text_level": 1, + "bbox": [ + 106, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daisheng Jin $^{1}$ Jiangbei Hu $^{1,2}$ Baixin Xu $^{1}$ Yuxin Dai $^{1}$ Chen Qian $^{3}$ Ying He $^{1*}$ $^{1}$ S-Lab, Nanyang Technological University, $^{2}$ Dalian University of Technology, $^{3}$ SenseTime Research", + "bbox": [ + 99, + 202, + 895, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this study, we introduce a novel two-stage technique for decomposing and reconstructing facial features from sparse-view images, a task made challenging by the unique geometry and complex skin reflectance of each individual. To synthesize 3D facial models more realistically, we endeavor to decouple key facial attributes from the RGB color, including geometry, diffuse reflectance, and specular reflectance. Specifically, we design a Sparse-view Face Decomposition Model (SFDM): 1) In the first stage, we create a general facial template from a wide array of individual faces, encapsulating essential geometric and reflectance characteristics. 2) Guided by this template, we refine a specific facial model for each individual in the second stage, considering the interaction between geometry and reflectance, as well as the effects of subsurface scattering on the skin. With these advances, our method can reconstruct high-quality facial representations from as few as three images. The comprehensive evaluation and comparison reveal that our approach outperforms existing methods by effectively disentangling geometric and reflectance components, significantly enhancing the quality of synthesized novel views, and paving the way for applications in facial relighting and reflectance editing. Visit our project page for more details https://kingjg.github.io/SFDM.github.io/.", + "bbox": [ + 86, + 305, + 483, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 685, + 220, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D human face reconstruction finds widespread applications in various fields, including game design and film production. Developments in neural rendering techniques have shown promising performance and advantages in synthesizing photo-realistic images for complex scenes [3, 32, 33, 43]. Concurrently, a surge of research focusing on head and facial attribute reconstruction from images via implicit neural representations [18, 36, 37, 59] has emerged. These innovative approaches address the limitations associated with traditional explicit methods, such as those based on the 3D Morphable Model (3DMM) [6], which struggled with is", + "bbox": [ + 89, + 710, + 483, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/151c96cf6e0f022bf42d81202e47e1e251ebe55d38faa536ed2a990aa5ccef4b.jpg", + "image_caption": [ + "Figure 1. Our method learns a general facial template with distinct key facial attributes from multiple individuals. This template enables photorealistic novel-view synthesis for each individual, as well as other applications." + ], + "image_footnote": [], + "bbox": [ + 521, + 276, + 905, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sues like dependency on mesh resolution and difficulties in managing topological variations. Furthermore, implicit neural representation methods demonstrate their effectiveness in 3D face reconstruction from low-view images [47], expression transfer [1, 23, 30], and animations [19, 59, 60].", + "bbox": [ + 511, + 520, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In 3D face model rendering, both accurate shape attributes and high-quality reflectance properties are essential for synthesizing photorealistic faces. Thus, robustly decomposing key facial attributes including geometry, diffuse and specular reflectance, and environment light is a crucial task, which not only enhances the accuracy of face reconstruction but also opens up possibilities for more applications like relighting [7, 11, 22, 29, 55]. However, most previous studies only use neural networks to learn radiance and geometric information from images, neglecting the effect of rendering-related physical information such as facial skin characteristics and environmental lighting conditions. NeRV [42], PhySG [53], and following methods [8, 9, 13, 21, 31, 34] incorporate the Bidirectional Reflectance Distribution Functions (BRDFs) into the neural rendering framework, facilitating the decomposition and joint reconstruction of shape, reflectance, and illumination. Nonetheless, these methods primarily concentrate on non-biological materials, which are not suitable for decomposing face models that possess complex attributes. NeuFace [58] introduces an approxi", + "bbox": [ + 511, + 599, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Y. He (Email: yhe@ntu.edu.sg)", + "bbox": [ + 107, + 886, + 406, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26409", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mated BRDF integration and learns the physically meaningful underlying representations for 3D face reconstruction. Nevertheless, due to the ill-posedness of decomposition, NeuFace often converges to a suboptimal solution under sparse-view settings, in which specular reflectance is absent, and all radiance is derived from diffuse attributes. In addition, the accuracy of geometric reconstruction significantly diminishes as the number of views decreases. Therefore, it still poses a significant challenge to decompose shape, reflectance, and environment light attributes from sparse view images in 3D face model reconstruction.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we introduce a novel two-stage face decomposition framework, termed Sparse-view Face Decomposition Model (SFDM), aimed at reconstructing 3D faces from sparse-view images. The pipeline of our proposed algorithm is depicted in Fig. 2. Based on the observation that human faces share similar geometric features and reflectance properties, we first derive a novel general facial template from images of a set of varied individual faces in Stage 1. In Stage 2, we employ the learned facial template with both geometry and reflectance attributes to refine the decomposition of each individual face model. In both stages, we decompose the shape and material attributes using a geometry module and a reflectance module, respectively.", + "bbox": [ + 89, + 260, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the geometry module consists of a template network and a deformation network. In the first stage, we can get a template model with common geometry features based on the training of multi-view images from varied faces, which can be further transformed to a specific individual face model by deformation. Building on the facial geometry template, we innovatively propose a reflectance module that integrates reflectance information into the facial template. Specifically, we introduce the BRDF as a coarse representation of reflectance for facial skin. We employ a BRDF template network to learn the shared characteristics of diffuse and specular reflectance properties, and a BRDF offset network is utilized to reconstruct identity-specific reflectance details. Furthermore, we propose a novel albedo gradient predictor designed to capture more general facial reflectance attributes. In the second stage, based on the learned template with geometry and reflectance attributes, we refine the attribute details from sparse-view images of each individual face model. To this end, we incorporate a displacement network and a reflectance offset network on the learned template, employing a design that harmonizes geometry and reflectance information to capture high-frequency details of facial attributes. In addition, we introduce a diffuse offset network to take into account the effect of subsurface scattering, thus providing a more realistic simulation of the facial skin. We summarize the main contributions of this study as follows:", + "bbox": [ + 89, + 459, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel decomposition framework for human facial representations to learn a template with various at", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tributes, including geometric features, diffuse reflectance, and specular reflectance.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Based on the learned template, our tailored refinement stage enables robust decomposition of attributes for each individual face under as few as three images, attaining superior geometric accuracy and reflectance detail.", + "- Extensive experiments demonstrate the face decomposition results obtained through our framework can facilitate more realistic synthesis in novel views, as well as applications such as specular editing and relighting." + ], + "bbox": [ + 513, + 121, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 256, + 650, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physically based rendering. To achieve a photorealistic rendering effect, physically based rendering (PBR) [39] is proposed to simulate the light and surfaces using real-world optical principles. Reversely, physically based inverse rendering seeks a way to decompose light and surface factors from the images [4, 5, 8, 34]. IDR [50] proposes a surface reconstruction method that simultaneously disentangles the geometry and appearance. Subsequent works incorporate inverse rendering into the neural rendering pipeline, developing various methods for general objects [9, 53], and enhancing decomposition effects for objects with specialized materials [10, 31]. NeuFace [58] proposes a decomposition pipeline for human faces, taking into account the sophisticated properties of facial skin reflectance. However, without ground truth for reflectance components, face decomposition is an ill-posed problem, causing NeuFace to often struggle with accurately decomposing reflectance. Additionally, NeuFace does not consider the effect of subsurface scattering on facial skin, resulting in less realistic PBR [49]. In this work, with our novel facial template and the single-person refinement module, which further considers subsurface scattering effects, we can robustly decompose geometry and reflectance in a more realistic manner from sparse-view images.", + "bbox": [ + 511, + 281, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face reconstruction. Given the inherent similarities among human faces, it is intuitive to develop a template that serves as a prior to enhance facial reconstruction. Based on a large amount of facial data, 3DMM [6] builds a 3D morphable face to describe the transformation of shape and expression, which is widely used in face recognition [56], 3D reconstruction, face editing [17, 36], etc. Using 3D morphable faces also enables face reconstruction from monocular images [14, 16, 27, 35, 61]. For instance, Han et al. [20] build a 3D morphable face with reflectance on BFM09 [38] for material editing. However, in face reconstruction, the use of 3D morphable faces often leads to a loss of detail during the compression and fitting process, resulting in overly coarse reconstructions. Moreover, given limited view information, monocular methods typically require reflectance ground truth for training [41]. As implicit neural representation shows extraordinary efficacy for 3D representation,", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26410", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "numerous studies [47, 52, 57, 59] construct 3D parametric face or head models based on neural networks, yielding improved geometric details for faces or heads. Besides, some facial capturing methods utilize more informative inputs such as videos [46] and polarized images [2], enabling high-quality reconstruction across a wider range of scenarios. Our goal is to achieve face reconstruction solely from sparse-view images, without any reflectance ground truth. Our Stage 1 employs an implicit representation to construct a facial template, incorporating prior knowledge of facial geometry and reflectance. This approach enables robust decomposition and realistic face reconstruction during the single-person refinement process under sparse views.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 300, + 217, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The effectiveness of PBR has been validated by numerous studies [25, 45, 58] for photo-realistic rendering in scenarios with complex reflection properties. Inspired by these works and as illustrated in Fig. 3, we formulate the output radiance $L_{\\mathrm{o}} \\in \\mathbb{R}^{3}$ at surface position $\\mathbf{x} \\in \\mathbb{R}^{3}$ along output direction $\\omega_{\\mathrm{o}} \\in \\mathbb{R}^{3}$ as: $L_{\\mathrm{o}}(\\mathbf{x}, \\omega_{\\mathrm{o}}) = \\int_{\\Omega} f(\\mathbf{x}, \\omega_{\\mathrm{i}}, \\omega_{\\mathrm{o}}) L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})(\\omega_{\\mathrm{i}} \\cdot \\mathbf{n}) \\, \\mathrm{d}\\omega_{\\mathrm{i}}$ , where $f$ represents the BRDF evaluation, $L_{\\mathrm{i}} \\in \\mathbb{R}^{3}$ is the incident light intensity from input direction $\\omega_{\\mathrm{i}} \\in \\mathbb{R}^{3}$ , $\\mathbf{n} \\in \\mathbb{R}^{3}$ is the surface normal. To decouple the view-dependent and view-independent radiance, we decompose $L_{\\mathrm{o}}$ into the diffuse term and the specular term by:", + "bbox": [ + 89, + 324, + 483, + 507 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {0} (\\mathbf {x}, \\boldsymbol {\\omega} _ {0}) = \\underbrace {\\frac {b _ {\\mathrm {a}} (\\mathbf {x})}{\\pi} \\int_ {\\Omega} L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}}) (\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}}} _ {\\text {d i f f u s e}} \\\\ + \\underbrace {\\int_ {\\Omega} f _ {s} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}} , \\boldsymbol {\\omega} _ {\\mathrm {o}} ; b _ {\\mathrm {s}} (\\mathbf {x}) , b _ {\\mathrm {r}} (\\mathbf {x})) L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}}) (\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}} ,} _ {\\text {s p e c u l a r}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 513, + 467, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $b_{\\mathrm{a}}(\\mathbf{x}) \\in \\mathbb{R}^3$ , $b_{\\mathrm{s}}(\\mathbf{x}) \\in \\mathbb{R}^4$ , $b_{\\mathrm{r}}(\\mathbf{x}) \\in \\mathbb{R}$ are diffuse albedo, specular, and roughness fields obtained from BRDF, respectively. $f_{s}$ represents the BRDF evaluation that only accounts for the specular property. Furthermore, we improve the efficiency of integration computation by using the approximated pre-computed method [9] as:", + "bbox": [ + 89, + 625, + 483, + 718 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {o}} (\\mathbf {x}, \\boldsymbol {\\omega} _ {\\mathrm {o}}) \\approx \\underbrace {\\frac {b _ {\\mathrm {a}} (\\mathbf {x})}{\\pi} \\hat {L} _ {\\mathrm {i}} (\\mathbf {n} , 1)} _ {\\text {d i f f u s e}} + \\underbrace {\\hat {f} _ {\\mathrm {s}} (\\mathbf {x} ; b _ {\\mathrm {s}} (\\mathbf {x}) , b _ {\\mathrm {r}} (\\mathbf {x})) \\hat {L} _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {\\mathrm {r}} , b _ {\\mathrm {r}} (\\mathbf {x})\\right)} _ {\\text {s p e c u l a r}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 724, + 506, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega_{\\mathrm{r}} \\in \\mathbb{R}^{3}$ represent the mirrored view direction, $\\hat{L}_i$ is the integration of input light, $\\hat{f}_s$ is the integration of specular reflectance. Specifically, following [26], we employ an MLP to simulate two look-up-textures (LUT) $B_0$ and $B_1$ to calculate the BRDF integration as:", + "bbox": [ + 89, + 785, + 483, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {\\mathrm {s}} (\\mathbf {x}) = b _ {\\mathrm {s}} (\\mathbf {x}) \\left[ F _ {0} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}, \\mathbf {n}\\right) B _ {0} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}} \\cdot \\mathbf {n}, b _ {\\mathrm {r}} (\\mathbf {x})\\right) + B _ {1} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}} \\cdot \\mathbf {n}, b _ {\\mathrm {r}} (\\mathbf {x})\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 867, + 495, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{0}$ is the Fresnel term at normal incidence as [9]. For light representation, we use spherical harmonics [44] to approximate environment light by:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {i}} \\left(\\omega_ {\\mathrm {i}}\\right) \\approx \\sum_ {\\ell = 0} ^ {k} \\sum_ {m = - \\ell} ^ {\\ell} c _ {\\ell} ^ {m} Y _ {\\ell} ^ {m} \\left(\\omega_ {\\mathrm {i}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 148, + 906, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Y$ represents spherical basis functions and $c$ denotes corresponding coefficients. Considering that the facial material is not of high specular reflectance, we set the order as $k = 10$ [12]. To integrate incident light rays, the light for the specular term can be integrated to the mirrored direction of the observation view, denoted as $\\hat{L}_{\\mathrm{i}}(\\omega_{\\mathrm{r}},b_{\\mathrm{r}})$ , and light for the diffuse term (Lambertian) can be approximated by $\\hat{L}_{\\mathrm{i}}(\\mathbf{n},1)$ [9], which assumes light in the normal direction hits a completely rough surface. For our SH light representation, we integrate $\\hat{L}_{\\mathrm{i}}(\\omega_{\\mathrm{r}},b_{\\mathrm{r}})$ for the specular term and $\\hat{L}_{\\mathrm{i}}(\\mathbf{n},1)$ for the diffuse term through an approximation using the first ten orders of spherical harmonics as [58].", + "bbox": [ + 511, + 204, + 906, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 401, + 604, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typical human faces share common features, such as two eyes and one nose, with variations in characteristics such as shape, skin color, and skin texture. To better utilize the shared features of human faces, we divide the decomposition process into two stages to learn the neural fields of human faces from multi-view images. Fig. 2 presents the architecture of the Sparse-view Face Decomposition Model (SFDM). In Stage 1, we create a general facial template with geometry and reflectance attributes from a collection of portraits of multiple individuals. In Stage 2, we refine the specific attributes for each individual face model based on the learned facial template under sparse-view settings.", + "bbox": [ + 511, + 428, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Generating facial template", + "text_level": 1, + "bbox": [ + 511, + 622, + 754, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a collection of portraits of multiple individuals $\\{\\mathcal{I}_i\\}$ , we first learn a general facial template with both geometry and reflectance attributes.", + "bbox": [ + 511, + 643, + 905, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Geometry template. Following the previous work [47, 57], we jointly employ a geometry template network $T_{\\mathrm{geo}}$ and a deformation network $F_{\\mathrm{dfm}}$ to reconstruct the geometry (represented by signed distance field $S$ ) from given images. As illustrated in Fig. 2, we formulate the geometry network as:", + "bbox": [ + 511, + 691, + 905, + 767 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (\\mathbf {x}) = T _ {\\text {g e o}} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\text {g e o}})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 784, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x} \\in \\mathbb{R}^3$ is a query point, $S(\\mathbf{x}) \\in \\mathbb{R}$ returns the signed distance value. We take $\\mathbf{z}_{\\mathrm{geo}} \\in \\mathbb{R}^{128}$ as an identity-related latent code that encodes geometry, so we can get the unique SDF for each individual. Similar to VolSDF [51], we convert the SDF value to density $\\sigma$ and then feed it into the volume rendering block.", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/607b10f409ce9cfb1ecc289feb583462d367727e0e2a1b8d20ba61cf2c25920a.jpg", + "image_caption": [ + "Figure 2. Pipeline of SFDM. We propose a geometry and reflectance decomposition framework for 3D face reconstruction from multiview images. We first learn a general facial template from multi-view images of multiple individuals in Stage 1. Then the learned facial template guides the further attribute refinement for each individual face model in Stage 2." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 906, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reflectance template. Apart from geometry attributes, we also consider reflectance attributes, which are equally crucial and share similar features across human faces. Therefore, we aim to create a coarse but general template that captures the underlying common reflectance characteristics of human faces in Stage 1.", + "bbox": [ + 88, + 378, + 480, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the reflectance attributes, we approximate the Bidirectional Reflectance Distribution Function (BRDF) by the network that consists of a reflectance template $T_{\\mathrm{rfl}}$ and a BRDF offset network $F_{\\mathrm{ofs}}$ . Similar to the geometry template, we leverage $T_{\\mathrm{rfl}}$ to fit the general facial BRDF for multiple individuals and then offset the BRDF value of the query points on each face model through $F_{\\mathrm{ofs}}$ . We formulate the calculation of BRDF $\\mathcal{B}_{\\mathrm{st1}}(\\mathbf{x})$ as follows:", + "bbox": [ + 88, + 469, + 482, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {\\mathrm {s t l}} (\\mathbf {x}) = T _ {\\mathrm {r f l}} \\left[ F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {g e o}}) \\right] + w F _ {\\mathrm {o f s}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {r f l}}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 601, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the query points $\\mathbf{x}$ in the observation space are deformed to the surface of the facial template through $F_{\\mathrm{dfm}}$ , and $T_{\\mathrm{rfl}}$ return the corresponding BRDF attributes $\\mathcal{B}(\\mathbf{x}) = [b_{\\mathrm{a}}(\\mathbf{x}), b_{\\mathrm{s}}(\\mathbf{x}), b_{\\mathrm{r}}(\\mathbf{x})] \\in \\mathbb{R}^8$ . Subsequently, the specific reflectance attributes for each individual are determined by learning the BRDF offsets with $F_{\\mathrm{ofs}}$ . To capture the unique reflectance characteristics associated with individual identities, we incorporate a reflectance latent code $\\mathbf{z}_{\\mathrm{rfl}} \\in \\mathbb{R}^{128}$ . Moreover, we adjust the offset using a weighting parameter $w$ , setting it to a low value of $w = 0.05$ . This choice reflects our preference for the template network to primarily capture reflectance properties, minimizing dependence on the BRDF offset network. To simulate environment light, we apply the spherical harmonics as in Eq. (3) and integrate the lights as in [58].", + "bbox": [ + 89, + 628, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the geometry, reflectance, and environment light information, we calculate the final color (output radiance) as Eq. (1-3) and integrate it by volume rendering.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1, a general facial template can be obtained and robustly decomposed into geometry, diffuse, and specular attributes.", + "bbox": [ + 511, + 378, + 906, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Albedo gradient predictor. Considering the wide range of skin color differences among various ethnic groups, simply relying on the absolute values of albedo on the facial template does not adequately convey the prior knowledge of faces. We observe that the continuity of facial albedo variations across 3D positions exhibits certain regularities, such as drastic changes around the edges of the lips and the eyes. Thus, we introduce the concept of albedo gradient as an additional attribute for the facial template. Specifically, during the training phase of Stage 1, we calculate the gradient of the albedo values $\\nabla \\mathbf{b}_{\\mathrm{a}}(\\mathbf{x}) \\in \\mathbb{R}^{3}$ from the fitted neural field with respect to changes in 3D coordinates $\\mathbf{x} \\in \\mathbb{R}^{3}$ . This gradient plot effectively outlines facial contours like facial sketches. Following this, we use it to supervise the training of an additional albedo gradient predictor branch $\\mathcal{G}(\\mathbf{x})$ within the reflectance template, guiding the model to pay attention to the patterns of albedo changes. During training, once the point $F_{\\mathrm{dfm}}(\\mathbf{x})$ in the template space is obtained, it serves as the input for $\\mathcal{G}$ and the output target is $\\nabla \\mathbf{b}_{\\mathrm{a}}(\\mathbf{x})$ in the observation space. In the end, the well-trained albedo gradient predictor provides more comprehensive information and precise guidance for single-person refinement in Stage 2.", + "bbox": [ + 511, + 426, + 906, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Individual decomposition under sparse views", + "text_level": 1, + "bbox": [ + 511, + 785, + 895, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking the facial template learned from Stage 1 as prior, we perform a fine-detailed decomposition from sparse-view images of a single face in Stage 2.", + "bbox": [ + 511, + 808, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry displacement and BRDF offset. The geometry template $T_{\\mathrm{geo}}$ only represents a rough shape with shared features of multiple human faces. Additionally, the defor", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26412", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mation net $F_{\\mathrm{dfm}}$ transforms the points on the surface of the facial template to the surface of a specific individual's face. However, the transformation is constrained, making it challenging to create topological features beyond the template domain and capture high-frequency details. Consequently, we incorporate a displacement network $F_{\\mathrm{dis}}$ in Stage 2 to modify the SDF obtained from the geometry template for a more accurate reconstruction. Building on Eq. (4), we incorporate the offset term as follows: $\\mathcal{S}(\\mathbf{x}) = T_{\\mathrm{geo}}(F_{\\mathrm{dfm}}(\\mathbf{x},\\mathbf{z}_{\\mathrm{geo}})) + F_{\\mathrm{dis}}(\\mathbf{x})$", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the reflectance template $T_{\\mathrm{rfl}}$ learned from Stage 1, we calculate the offset value of BRDF through $F_{\\mathrm{ofs}}$ , similar to Stage 1, to obtain the specific BRDF attributes for each individual. To address the distinct diversity between diffuse albedo $b_{a}$ , specular $b_{s}$ and roughness $b_{r}$ , we add a learnable weight term $\\mathbf{W}(\\mathbf{x}) = [\\mathbf{w}_{\\mathrm{a}}, \\mathbf{w}_{\\mathrm{s}}, \\mathbf{w}_{\\mathrm{r}}]$ to control the contribution of each BRDF attribute from the reflectance template $T_{\\mathrm{rfl}}$ , where $\\mathbf{w}_{\\mathrm{a}} \\in \\mathbb{R}^{3}, \\mathbf{w}_{\\mathrm{s}} \\in \\mathbb{R}^{4}, \\mathbf{w}_{\\mathrm{r}} \\in \\mathbb{R}$ are the weights for diffuse albedo, specular, and roughness, respectively. Thus, we obtain the BRDF attributes as:", + "bbox": [ + 89, + 243, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {\\mathrm {s t 2}} (\\mathbf {x}) = \\mathbf {W} (\\mathbf {x}) \\cdot T _ {\\mathrm {r f l}} \\left[ F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {g e o}}) \\right] + F _ {\\mathrm {o f s}} (\\mathbf {x}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 407, + 480, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering that the diffuse albedo attribute has a large variation between different people, we initialize $\\mathbf{w}_{\\mathrm{a}}$ with relatively low values.", + "bbox": [ + 89, + 436, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribute synergy. In our framework, we decompose facial attributes into two components: geometry and reflectance. These components are not isolated but can enhance each other cooperatively. Sometimes, distinguishing between geometry and reflectance can be challenging, leading to inaccuracies in the decomposition process. For example, a dark region in a facial image might result from sunken areas or pigmentation spots on the face. Conventional neural rendering techniques rely solely on color information from images, making it hard to distinguish such ambiguous scenes, often leading to incorrect reconstruction. Since our method can decouple rough geometry and reflectance attributes in Stage 1, we leverage the prior from the reflectance template to facilitate geometry inference, and vice versa. Specifically, we obtain the displacement $\\mathcal{D}(\\mathbf{x})$ as follows:", + "bbox": [ + 89, + 482, + 483, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} (\\mathbf {x}) = F _ {\\text {d i s}} (\\mathbf {x}, \\operatorname {C o n c a t} \\left(\\mathbf {f} _ {\\mathrm {d e f}}, \\mathbf {f} _ {\\mathrm {g e o}}, \\mathbf {f} _ {\\mathrm {r f l}}, \\mathcal {G} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x})\\right)\\right)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 720, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{\\mathrm{dfm}} \\in \\mathbb{R}^{128}$ , $\\mathbf{f}_{\\mathrm{geo}} \\in \\mathbb{R}^{128}$ , and $\\mathbf{f}_{\\mathrm{rfl}} \\in \\mathbb{R}^{128}$ are extracted features from the output of the final hidden layer of $F_{\\mathrm{dfm}}$ , $T_{\\mathrm{geo}}$ , and $T_{\\mathrm{rfl}}$ , respectively. $\\mathcal{G}(F_{\\mathrm{dfm}}(\\mathbf{x}))$ is the output from the albedo gradient predictor. We also extract another feature vector $\\mathbf{f}_{\\mathrm{dis}} \\in \\mathbb{R}^{64}$ from $F_{\\mathrm{dis}}$ to provide more geometry information to the BRDF offset network $F_{\\mathrm{ofs}}$ . Similarly, we calculate the BRDF offset $\\mathcal{B}_{\\mathrm{ofs}}(\\mathbf{x})$ as:", + "bbox": [ + 89, + 750, + 483, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {\\mathrm {o f s}} (\\mathbf {x}) = F _ {\\mathrm {o f s}} (\\mathbf {x}, \\operatorname {C o n c a t} \\left(\\mathbf {f} _ {\\mathrm {d e f}}, \\mathbf {f} _ {\\mathrm {g e o}}, \\mathbf {f} _ {\\mathrm {d i s}}, \\mathbf {f} _ {\\mathrm {r f l}}, \\mathcal {G} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x})\\right)\\right)). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our ablation study validates the effectiveness of this attribute synergy module in the accuracy of both geometry reconstruction and rendering results.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0b20076120f680c41ce96bbb0585b4066006fbfe8f3debdd746d76bd763748da.jpg", + "image_caption": [ + "Figure 3. Left: BRDF rendering, where light reflects at the incident point. Right: Subsurface scattering on facial skin, showing the offset between the incident and outgoing points." + ], + "image_footnote": [], + "bbox": [ + 532, + 148, + 870, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsurface scattering offset. Human facial skin exhibits a complexity far beyond that of non-biological materials like metal and plastic. Building on insights from previous research [15], rendering realistic faces necessitates accounting for subsurface scattering effects. As illustrated in Fig. 3, this phenomenon can be modeled using a three-layer approach, where specular reflections occur directly at the point of light incidence on the surface, while diffuse reflections involve light penetrating the skin, traversing three layers, undergoing multiple subsurface reflections, and then emerging from the skin surface slightly away from the point of incidence. This effect, known as subsurface scattering, is crucial for realistic skin rendering and can be mathematically represented as $L_{\\mathrm{o}}(\\mathbf{x}_{\\mathrm{o}}, \\omega_{\\mathrm{o}}) = \\int_{A} \\int_{\\Omega} S(\\mathbf{x}_{\\mathrm{o}}, \\omega_{\\mathrm{o}}, \\mathbf{x}_{\\mathrm{i}}, \\omega_{\\mathrm{i}}) L_{\\mathrm{i}}(\\mathbf{x}_{\\mathrm{i}}, \\omega_{\\mathrm{i}})(\\boldsymbol{\\omega}_{\\mathrm{i}} \\cdot \\mathbf{n}) \\, \\mathrm{d}\\omega_{\\mathrm{i}} \\, \\mathrm{d}A$ , where $\\mathbf{x}_{\\mathrm{i}}$ and $\\mathbf{x}_{\\mathrm{o}}$ are the positions of the light incident and output points, respectively, $A$ is a small area around $\\mathbf{x}_{\\mathrm{o}}$ , and $S$ is the scattering evaluation. In practice, the BRDF mainly accounts for light that scatters back at the point where it hits a surface, but simulating surface scattering accurately is computationally expensive. To improve face reconstruction quality, we integrate subsurface scattering into the BRDF's diffuse component as an offset. We simplify the process by assuming uniform subsurface scattering properties within a small area and utilize scattering profiles to approximate this effect. As illustrated in Fig. 3, these profiles show how the reflectance due to subsurface scattering changes with the distance between where light enters and exits the skin.", + "bbox": [ + 511, + 328, + 906, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As [28] suggests, one can approximate each RGB channel of the scattering profiles using a combination of six Gaussian functions. We employ the weights of these Gaussian functions to form a feature vector $\\mathbf{f}_{\\mathrm{sp}}^{\\mathbf{x}} \\in \\mathbb{R}^{18}$ , which represents the scattering profile of the point $\\mathbf{x}$ . Consequently, we calculate the offset of the diffuse term as:", + "bbox": [ + 511, + 751, + 905, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\mathrm {s s}} (\\mathbf {x}) = v \\hat {F} _ {\\mathrm {s p}} (\\mathbf {x}, \\mathbf {f} _ {\\mathrm {s p}} ^ {\\mathbf {x}}) \\hat {L} _ {\\mathrm {i}} (\\mathbf {x}, \\mathcal {E} (\\mathbf {x})), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 852, + 903, + 872 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v$ is the weight of scattering offset, $\\hat{F}_{\\mathrm{sp}}$ is a network", + "bbox": [ + 511, + 883, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26413", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to integrate scattering reflectance, $\\hat{L}_{\\mathrm{i}}$ is a network to integrate the incoming light in a small region around $\\mathbf{x}$ , and $\\mathcal{E}(\\mathbf{x}) \\in \\mathbb{R}^3$ is a learnable vector that defines the size of the integration domain. The values of $\\mathbf{f}_{\\mathrm{sp}}^{\\mathbf{x}}$ and $\\mathcal{E}(\\mathbf{x})$ are obtained through a parameter network. We refer readers to the supplementary for more details of the network architecture.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training losses", + "text_level": 1, + "bbox": [ + 89, + 194, + 241, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For both stages, we consider reconstruction losses including the pixel-wise color loss $\\mathcal{L}_{\\mathrm{col}} = \\| \\mathbf{C} - \\mathbf{C}_{\\mathrm{gt}}\\| _1$ and the eikonal loss $\\mathcal{L}_{\\mathrm{eik}} = (\\| \\nabla S\\| _2 - 1)^2$ . Similar to NeuFace [58], we incorporate a regularization term $\\mathcal{L}_{\\mathrm{light}}$ to assume a nearly white environment light and penalize the specular energy by $\\mathcal{L}_{\\mathrm{spec}}$ . Additionally, we add $\\mathcal{L}_{\\mathrm{code}} = \\| \\mathbf{z}_{\\mathrm{geo}}\\| _2 + \\| \\mathbf{z}_{\\mathrm{rfl}}\\| _2$ to regularize the latent codes. To avoid a large variation between the predict reflectance attributes and the template during training, we apply a constraint to BRDF offset as $\\mathcal{L}_{\\mathrm{ofs}} = \\| \\mathcal{B}_{\\mathrm{ofs}}(\\mathbf{x})\\| _1$ . Additionally, the albedo gradient predictor employs $\\mathcal{L}_{\\mathrm{g}}$ , an L1 loss, which is supervised by the computed albedo gradient. As a result, we formulate the loss function in Stage 1 as follows:", + "bbox": [ + 89, + 215, + 483, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s t 1}} = \\mathcal {L} _ {\\mathrm {c o l}} + \\mathcal {L} _ {\\mathrm {e i k}} + \\mathcal {L} _ {\\mathrm {l i g h t}} + \\mathcal {L} _ {\\mathrm {s p e c}} + \\mathcal {L} _ {\\mathrm {o f s}} + \\mathcal {L} _ {\\mathrm {c o d e}} + \\mathcal {L} _ {\\mathrm {g}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 426, + 480, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the second stage, we add additional displacement loss $\\mathcal{L}_{\\mathrm{dis}}$ and scatter offset loss $\\mathcal{L}_{\\mathrm{ss}}$ : $\\mathcal{L}_{\\mathrm{st2}} = \\mathcal{L}_{\\mathrm{st1}} + \\mathcal{L}_{\\mathrm{dis}} + \\mathcal{L}_{\\mathrm{ss}}$ , where $\\mathcal{L}_{\\mathrm{dis}} = \\| \\mathcal{D}(\\mathbf{x})\\|_1$ and $\\mathcal{L}_{\\mathrm{ss}} = \\| F_{\\mathrm{ss}}(\\mathbf{x})\\|_1$ are regularization losses aimed at constraining high values of the displacement and the subsurface scattering offset.", + "bbox": [ + 89, + 457, + 482, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental results", + "text_level": 1, + "bbox": [ + 89, + 549, + 290, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 574, + 277, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and metrics. We evaluate our method on a large human face dataset Facescape [48]. Specifically, we utilize authorized data from approximately 30 individuals, 4 of whom are from the publishable list. To protect the privacy of other identities that are not on the publishable list, we employed mosaic technology in the eyes region. To validate the generalization performance of our method, we additionally use dataset H3DS [40] in the supplementary material. We assess the quality of image rendering using the PSNR, SSIM, and LPIPS [54] metrics, and the Chamfer distance (CD) for evaluating the accuracy of geometry reconstruction. We compare our method with two 3D reconstruction methods VolSDF [51], DeformHead [47] and three BRDF decomposition methods PhySG [53], TensoIR [24], and NeuFace [58].", + "bbox": [ + 89, + 597, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Facescape has over 50 scans of human heads along with camera parameters and meshes. We preprocess the images to obtain face masks using ImFace [57] tools. For Stage 1 training, we select 10 subjects, each with 10 views. For Stage 2, we use an additional", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "20 subjects and choose 3, 5, and 10 views from the available $50+$ views as the training set, with another 10 views reserved for testing. The 10-view experiments specifically evaluate performance in a low-view setting. All test data is independent of the training set and is used consistently across all methods for comparison. We employ the Adam optimizer with an initial learning rate $1 \\times 10^{-4}$ to train both stages for 3000 epochs. Additionally, we add a calibration module as [58] to mitigate the exposure differences between various views and apply this calibration to all the methods for a fair comparison. For more implementation details, please refer to the supplementary material.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation & comparison", + "text_level": 1, + "bbox": [ + 511, + 286, + 746, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To corroborate the effectiveness of our proposed framework in the reconstruction of appearance and geometry, we conducted a comprehensive comparison with state-of-the-art methods. These methods are categorized into two groups: geometry-oriented methods, which include VolSDF [51] and DeformHead [47]; and decomposition-based methods, including PhySG [53], TensoIR [24], NeuFace [58], and our method SFDM. As shown in Tab. 1, we enumerate the quantitative performance of different methods in the evaluation metrics. To assess the performance capability under sparse-view images, we conducted tests on 20 subjects with 3, 5, and 10 views, respectively. The visual results of the reconstruction from 3 views are presented in Fig. 4, showcasing the synthesized rendering images from a novel viewpoint, the geometric reconstruction, and the decomposition of diffuse and specular attributes.", + "bbox": [ + 511, + 309, + 903, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decomposition-based methods. Our method exhibits superior performance in novel view rendering and geometry reconstruction compared to other reflectance decomposition methods. In contrast to PhySG and TensoIR, designed for common objects, our method significantly outperforms them across all metrics. This is due to the complexity of facial material reflectance, which requires specialized reflectance modeling. NeuFace, which is customized for facial skin, poses a challenge due to its non-robust decomposition of geometry and reflectance attributes. As shown in Fig. 4, it often happens in NeuFace that all reflectance is baked into albedo, but the specular term is entirely absent. This degenerate Lambertian model may reduce the realism of the synthesized rendering output, especially for elements such as teeth, eyes, or the oily layer on facial skin, which often exhibit specular reflection effects. In contrast, we are able to stably decompose the diffuse and specular properties, resulting in more realistic rendering effects. Additionally, we demonstrate higher accuracy in geometric reconstruction, whereas the quality of NeuFace's geometry deteriorates significantly as the number of views decreases.", + "bbox": [ + 511, + 551, + 903, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry-oriented methods. Our method demonstrates a comparable level of accuracy in geometry reconstruc", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26414", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1c1692e091e7d80015824f9b2e53b1f0d012ac0bad1df64e423bbcfc52f6d8c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRelight3 views5 views10 views
PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓
VolSDF [51]✗22.950.8686.9218.2825.450.8976.097.4027.340.9105.185.25
DeformHead [47]✗26.020.8865.769.6026.410.8794.838.4427.350.8915.205.42
PhySG [53]✓22.030.8637.8170.8722.690.8727.0563.2023.110.8767.1255.7
TensoIR [24]✓22.350.6457.1428.6024.610.7045.8617.1026.140.7694.7310.10
NeuFace [58]✓23.540.8765.7815.7425.810.9044.758.6227.190.9144.036.40
Ours✓25.540.9024.469.2726.730.9113.837.0027.640.9203.865.53
", + "bbox": [ + 114, + 88, + 885, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61a0061f77082847219d80bf365a60c6c19520fdcf4f2a7e5fc11bbee735fc0e.jpg", + "image_caption": [ + "Figure 4. Visualization of 3-view reconstruction results. VolSDF and DeformHead do not incorporate reflectance attributes, resulting in the absence of corresponding outcomes." + ], + "image_footnote": [], + "bbox": [ + 106, + 257, + 467, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion compared to VolSDF and DeformHead. VolSDF exhibits poor 3D reconstruction results in sparse view settings. While DeformHead achieves a more detailed 3D reconstruction utilizing a geometry template, it is overly sensitive to geometry deformation, leading to noticeable indenta", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions. Although the PSNR of our method is lower than that of DeformHead under 3 views, our LPIPS outperforms it, indicating more photorealistic synthesis results. Failure instances in reconstruction have been observed with VolSDF and DeformHead, with failure rates of $21.7\\%$ and $6.7\\%$ , respectively. We exclude these failures from the comparison.", + "bbox": [ + 511, + 262, + 906, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f2e285b1bd6d72dfc1dd4f3864d12669744213c5b62ea795d03505a926fa858.jpg", + "table_caption": [ + "Table 1. Evaluation. The LPIPS and CD metrics are measured on scales of ${10}^{-2}$ and ${10}^{-3}$ ,respectively. The best and second-best results are highlighted in red and orange, respectively." + ], + "table_footnote": [], + "table_body": "
ExpMethodPSNR↑SSIM↑LPIPS↓CD↓
1w/o facial template24.270.8575.5917.22
2w/o albedo gradient25.130.8774.668.83
3w/o attribute synergy24.790.8764.908.80
4w/o scattering offset24.600.8745.738.61
5Ours25.650.8824.457.94
", + "bbox": [ + 517, + 364, + 906, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparative evaluation metrics for ablation studies.", + "bbox": [ + 527, + 479, + 890, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/067950823dea8c9180dbf1f0ab6bed6a216d70d8b7027c411f2f339ce3ef6151.jpg", + "image_caption": [ + "Figure 5. Ablation visualization. Top: Rendering. Bottom: Geometric errors." + ], + "image_footnote": [], + "bbox": [ + 532, + 508, + 888, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 696, + 674, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select a subset of 10 subjects from the dataset to test the effectiveness of the major components of SFDM. In Tab. 2, Exp.1-2 are results for the components in Stage 1, and Exp.3-4 cover the components in Stage 2. Exp.1-2 indicate that the facial template significantly influences the robustness of face decomposition, with the albedo gradient predictor further enriching the reconstruction quality by introducing additional facial prior knowledge. From Exp.3, we observe that the attribute synergy module appears to contribute to geometry decomposition, and its absence leads to a noticeable decrease in geometric accuracy. Based on Exp.4, the subsurface scatter offset module can enhance", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26415", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the realism of the synthesized images from novel views, which is suggested by improvements in the LPIPS metric. In Fig. 5, columns 3-6 show the visualization results when omitting the facial template $T_{*}$ , albedo gradient predictor $\\mathcal{G}$ , attribute synergy module $\\mathbf{f}_{*}$ , and subsurface scattering $F_{\\mathrm{ss}}$ , respectively. The bottom row highlights the albedo gradient predictor's more significant impact on individuals who exhibit greater deviations from the facial template. In Stage 2, the subsurface scattering offset introduces additional details in light reflection on the skin, resulting in more photorealistic rendering results. Regarding the attribute synergy module, the exchange of information between geometry and reflectance enables a more accurate decomposition, as is evident in the details of geometry reconstruction.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Reflectance decomposition analysis", + "text_level": 1, + "bbox": [ + 89, + 311, + 397, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method focuses on reflectance decomposition from multi-view facial images without any ground truth reflectance data, enabling realistic reconstruction of facial texture and highlights for novel view rendering. While our goal is not a physically accurate decomposition of skin reflectance, we aim to evaluate our decomposition results both visually and quantitatively. To this end, we use the single-view face reconstruction model, RefMM [20], to fit and decompose images in the test set, generating a coarse diffuse and specular result as pseudo ground truth for evaluation. Tab. 3 shows that our decomposition accuracy generally outperforms that of NeuFace. The overall low SSIM values are primarily due to the coarse nature of the pseudo ground truth, as illustrated in Fig. 6. This reflects certain advantages and limitations of single-view decomposition methods. These methods are typically trained on extensive facial image datasets with specialized lighting conditions or pseudo reflectance ground truth, allowing them to robustly decompose faces from monocular images. However, because they rely heavily on a parameterized facial model and limited input with less 3D information, the reconstructed results lack many facial details, making the rendered images appear less realistic. As shown in Fig. 6 and Fig. 7, we can observe these differences visually in the rendered results.", + "bbox": [ + 91, + 334, + 483, + 696 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9d8855a34049f9e2ba93d54dc262eb6ddc38c15311dab06a6ee324a34509d642.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDiffuseSpecular
NeuFace0.7260.552
Ours0.7610.692
", + "bbox": [ + 171, + 709, + 405, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Quantitative diffuse and specular results (SSIM↑).", + "bbox": [ + 111, + 780, + 460, + 795 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a36b91b247728a7e3d087dc5ffbf506322a7d0e1b02808ffbf17d381a0bb785a.jpg", + "image_caption": [ + "Figure 6. Decomposition results of RefMM [20]." + ], + "image_footnote": [], + "bbox": [ + 169, + 805, + 406, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Applications", + "text_level": 1, + "bbox": [ + 513, + 90, + 648, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specular editing. Our robust decomposition of diffuse and specular reflectances allows us to easily adjust the specular reflectance, resulting in the ability to make faces appear shinier or dimmer. The specular editing outcomes are showcased in Fig. 7. Specifically, we increase the specular weight by 1.5 times, while maintaining the overall specular distribution unchanged. As a result, the edited images can retain their photorealistic appearance.", + "bbox": [ + 511, + 114, + 906, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3effd274954fa97bb2e5d16afce37fceb07ec51d9b5c43d41813f34dc7122fcc.jpg", + "image_caption": [ + "Figure 7. Applications. (a) Specular editing by increasing the shininess. (b) Relighting under different environments." + ], + "image_footnote": [], + "bbox": [ + 514, + 250, + 901, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relighting. Decomposing reflectance and environment light facilitates the relighting of human faces without altering their inherent properties. As shown in Fig. 7, we conducted a comparison of the relighting effects between our method and NeuFace in sparse view settings. Our method excels in decomposing specular reflectance, leading to more vibrant and realistic relighting outcomes under varying environmental lighting conditions [12].", + "bbox": [ + 511, + 439, + 906, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 582, + 633, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel face decomposition method capable of rendering and reconstructing 3D human faces from sparse-view images. Our framework first learns a facial template with both decomposed geometry and reflectance attributes from multi-view images of different individuals. Then we utilize the prior knowledge of the facial template and further consider the interaction between geometry and reflectance, as well as subsurface scattering effects, which enable our method to achieve robust face decomposition from sparse-view images. Based on decomposed facial attributes, our method achieves superior results in the rendering and geometric reconstruction of 3D human faces and can be utilized for relighting and specular editing.", + "bbox": [ + 511, + 609, + 906, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & future work. As we focus on face decomposition, it is hard to extend to general objects directly with our method. In the future, we could involve techniques such as hash encoding [33] to accelerate our method and leverage fine-detailed 3D prior or reflectance ground truth for a better decomposition effect.", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26416", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22) and the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). J.Hu was supported by the National Natural Science Foundation of China under Grants 62402083.", + "bbox": [ + 89, + 114, + 483, + 234 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 251, + 187, + 267 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1", + "[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 3", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1", + "[4] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hašan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2", + "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2", + "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164, 2023. 1, 2", + "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 1", + "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2", + "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2, 3" + ], + "bbox": [ + 99, + 276, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Chengqian Che, Fujun Luan, Shuang Zhao, Kavita Bala, and Ioannis Gkioulekas. Towards learning-based inverse subsurface scattering. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2", + "[11] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 1", + "[12] Zhang Chen, Anpei Chen, Guli Zhang, Chengyuan Wang, Yu Ji, Kiriakos N Kutulakos, and Jingyi Yu. A neural rendering framework for free-viewpoint relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5610, 2020. 3, 8", + "[13] Ziang Cheng, Hongdong Li, Yuta Asano, Yinqiang Zheng, and Imari Sato. Multi-view 3d reconstruction of a textureless smooth surface of unknown generic reflectance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16226-16235, 2021. 1", + "[14] Abdallah Dib, Luiz Gustavo Hafemann, Emeline Got, Trevor Anderson, Amin Fadainejad, Rafael MO Cruz, and Marc-Andre Carbonneau. Mosar: Monocular semi-supervised model for avatar reconstruction using differentiable shading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1770–1780, 2024. 2", + "[15] Craig Donner and Henrik Wann Jensen. Light diffusion in multi-layered translucent materials. ACM Transactions on Graphics (ToG), 24(3):1032-1039, 2005. 5", + "[16] Haiwen Feng, Timo Bolkart, Joachim Tesch, Michael J Black, and Victoria Abrevaya. Towards racially unbiased skin tone estimation via scene disambiguation. In European Conference on Computer Vision, pages 72-90. Springer, 2022. 2", + "[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 2", + "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1", + "[19] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 1", + "[20] Yuxuan Han, Zhibo Wang, and Feng Xu. Learning a 3d morphable face reflectance model from low-cost data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8598-8608, 2023. 2, 8", + "[21] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 1", + "[22] Mingming He, Pascal Clausen, Ahmet Levent Taşel, Li Ma, Oliver Pilarski, Wenqi Xian, Laszlo Rikker, Xueming Yu, Ryan Burgert, Ning Yu, et al. Diffrelight: Diffusion-based" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26417", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "facial performance relighting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 1", + "[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 1", + "[24] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 6, 7", + "[25] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3", + "[26] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3", + "[27] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8629-8640, 2023. 2", + "[28] Andrew Lauritzen. Gpu gems 3, 2007. 5", + "[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 1", + "[30] Connor Lin, Koki Nagano, Jan Kautz, Eric Chan, Umar Iqbal, Leonidas Guibas, Gordon Wetzstein, and Sameh Khamis. Single-shot implicit morphable faces with consistent texture parameterization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1", + "[31] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Trans. Graph., 42(4): 114:1-114:22, 2023. 1, 2", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 8", + "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2", + "[36] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2", + "[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 1", + "[38] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2", + "[39] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2", + "[40] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 6", + "[41] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 2", + "[42] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1", + "[43] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1", + "[44] Borom Tunwattanapong, Graham Fyffe, Paul Graham, Jay Busch, Xueming Yu, Abhijeet Ghosh, and Paul Debevec. Acquiring reflectance and shape from continuous spherical harmonic illumination. ACM Transactions on graphics (TOG), 32(4):1-12, 2013. 3", + "[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3", + "[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and re" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26418", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 3", + "[47] Baixin Xu, Jiarui Zhang, Kwan-Yee Lin, Chen Qian, and Ying He. Deformable model-driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17924-17934, 2023. 1, 3, 6, 7", + "[48] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 6", + "[49] Jing Yang, Hanyuan Xiao, Wenbin Teng, Yunxuan Cai, and Yajie Zhao. Light sampling field and BRDF representation for physically-based neural rendering. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2", + "[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2", + "[51] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 6, 7", + "[52] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 3", + "[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 1, 2, 6, 7", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[55] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1", + "[56] Wenyi Zhao, Rama Chellappa, P Jonathon Phillips, and Azriel Rosenfeld. Face recognition: A literature survey. ACM computing surveys (CSUR), 35(4):399-458, 2003. 2", + "[57] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 3, 6", + "[58] Mingwu Zheng, Haiyu Zhang, Hongyu Yang, and Di Huang. Neuface: Realistic 3d neural face rendering from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16868-16877, 2023. 1, 2, 3, 4, 6, 7", + "[59] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 3", + "[60] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 1", + "[61] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum, pages 523-550. Wiley Online Library, 2018. 2" + ], + "bbox": [ + 516, + 92, + 903, + 429 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26419", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_model.json b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_model.json new file mode 100644 index 0000000000000000000000000000000000000000..df595a1a204f82786e477be6d67c7da145aed23f --- /dev/null +++ b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_model.json @@ -0,0 +1,2158 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.892, + 0.177 + ], + "angle": 0, + "content": "SFDM: Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.897, + 0.242 + ], + "angle": 0, + "content": "Daisheng Jin\\(^{1}\\) Jiangbei Hu\\(^{1,2}\\) Baixin Xu\\(^{1}\\) Yuxin Dai\\(^{1}\\) Chen Qian\\(^{3}\\) Ying He\\(^{1*}\\) \n\\(^{1}\\)S-Lab, Nanyang Technological University, \\(^{2}\\)Dalian University of Technology, \\(^{3}\\)SenseTime Research" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.306, + 0.485, + 0.671 + ], + "angle": 0, + "content": "In this study, we introduce a novel two-stage technique for decomposing and reconstructing facial features from sparse-view images, a task made challenging by the unique geometry and complex skin reflectance of each individual. To synthesize 3D facial models more realistically, we endeavor to decouple key facial attributes from the RGB color, including geometry, diffuse reflectance, and specular reflectance. Specifically, we design a Sparse-view Face Decomposition Model (SFDM): 1) In the first stage, we create a general facial template from a wide array of individual faces, encapsulating essential geometric and reflectance characteristics. 2) Guided by this template, we refine a specific facial model for each individual in the second stage, considering the interaction between geometry and reflectance, as well as the effects of subsurface scattering on the skin. With these advances, our method can reconstruct high-quality facial representations from as few as three images. The comprehensive evaluation and comparison reveal that our approach outperforms existing methods by effectively disentangling geometric and reflectance components, significantly enhancing the quality of synthesized novel views, and paving the way for applications in facial relighting and reflectance editing. Visit our project page for more details https://kingjg.github.io/SFDM.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.686, + 0.222, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.878 + ], + "angle": 0, + "content": "3D human face reconstruction finds widespread applications in various fields, including game design and film production. Developments in neural rendering techniques have shown promising performance and advantages in synthesizing photo-realistic images for complex scenes [3, 32, 33, 43]. Concurrently, a surge of research focusing on head and facial attribute reconstruction from images via implicit neural representations [18, 36, 37, 59] has emerged. These innovative approaches address the limitations associated with traditional explicit methods, such as those based on the 3D Morphable Model (3DMM) [6], which struggled with is" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.277, + 0.906, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Figure 1. Our method learns a general facial template with distinct key facial attributes from multiple individuals. This template enables photorealistic novel-view synthesis for each individual, as well as other applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.597 + ], + "angle": 0, + "content": "sues like dependency on mesh resolution and difficulties in managing topological variations. Furthermore, implicit neural representation methods demonstrate their effectiveness in 3D face reconstruction from low-view images [47], expression transfer [1, 23, 30], and animations [19, 59, 60]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In 3D face model rendering, both accurate shape attributes and high-quality reflectance properties are essential for synthesizing photorealistic faces. Thus, robustly decomposing key facial attributes including geometry, diffuse and specular reflectance, and environment light is a crucial task, which not only enhances the accuracy of face reconstruction but also opens up possibilities for more applications like relighting [7, 11, 22, 29, 55]. However, most previous studies only use neural networks to learn radiance and geometric information from images, neglecting the effect of rendering-related physical information such as facial skin characteristics and environmental lighting conditions. NeRV [42], PhySG [53], and following methods [8, 9, 13, 21, 31, 34] incorporate the Bidirectional Reflectance Distribution Functions (BRDFs) into the neural rendering framework, facilitating the decomposition and joint reconstruction of shape, reflectance, and illumination. Nonetheless, these methods primarily concentrate on non-biological materials, which are not suitable for decomposing face models that possess complex attributes. NeuFace [58] introduces an approxi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.407, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Y. He (Email: yhe@ntu.edu.sg)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26409" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "mated BRDF integration and learns the physically meaningful underlying representations for 3D face reconstruction. Nevertheless, due to the ill-posedness of decomposition, NeuFace often converges to a suboptimal solution under sparse-view settings, in which specular reflectance is absent, and all radiance is derived from diffuse attributes. In addition, the accuracy of geometric reconstruction significantly diminishes as the number of views decreases. Therefore, it still poses a significant challenge to decompose shape, reflectance, and environment light attributes from sparse view images in 3D face model reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.457 + ], + "angle": 0, + "content": "In this study, we introduce a novel two-stage face decomposition framework, termed Sparse-view Face Decomposition Model (SFDM), aimed at reconstructing 3D faces from sparse-view images. The pipeline of our proposed algorithm is depicted in Fig. 2. Based on the observation that human faces share similar geometric features and reflectance properties, we first derive a novel general facial template from images of a set of varied individual faces in Stage 1. In Stage 2, we employ the learned facial template with both geometry and reflectance attributes to refine the decomposition of each individual face model. In both stages, we decompose the shape and material attributes using a geometry module and a reflectance module, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Specifically, the geometry module consists of a template network and a deformation network. In the first stage, we can get a template model with common geometry features based on the training of multi-view images from varied faces, which can be further transformed to a specific individual face model by deformation. Building on the facial geometry template, we innovatively propose a reflectance module that integrates reflectance information into the facial template. Specifically, we introduce the BRDF as a coarse representation of reflectance for facial skin. We employ a BRDF template network to learn the shared characteristics of diffuse and specular reflectance properties, and a BRDF offset network is utilized to reconstruct identity-specific reflectance details. Furthermore, we propose a novel albedo gradient predictor designed to capture more general facial reflectance attributes. In the second stage, based on the learned template with geometry and reflectance attributes, we refine the attribute details from sparse-view images of each individual face model. To this end, we incorporate a displacement network and a reflectance offset network on the learned template, employing a design that harmonizes geometry and reflectance information to capture high-frequency details of facial attributes. In addition, we introduce a diffuse offset network to take into account the effect of subsurface scattering, thus providing a more realistic simulation of the facial skin. We summarize the main contributions of this study as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel decomposition framework for human facial representations to learn a template with various at" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "tributes, including geometric features, diffuse reflectance, and specular reflectance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Based on the learned template, our tailored refinement stage enables robust decomposition of attributes for each individual face under as few as three images, attaining superior geometric accuracy and reflectance detail." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.243 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the face decomposition results obtained through our framework can facilitate more realistic synthesis in novel views, as well as applications such as specular editing and relighting." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.257, + 0.651, + 0.273 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.644 + ], + "angle": 0, + "content": "Physically based rendering. To achieve a photorealistic rendering effect, physically based rendering (PBR) [39] is proposed to simulate the light and surfaces using real-world optical principles. Reversely, physically based inverse rendering seeks a way to decompose light and surface factors from the images [4, 5, 8, 34]. IDR [50] proposes a surface reconstruction method that simultaneously disentangles the geometry and appearance. Subsequent works incorporate inverse rendering into the neural rendering pipeline, developing various methods for general objects [9, 53], and enhancing decomposition effects for objects with specialized materials [10, 31]. NeuFace [58] proposes a decomposition pipeline for human faces, taking into account the sophisticated properties of facial skin reflectance. However, without ground truth for reflectance components, face decomposition is an ill-posed problem, causing NeuFace to often struggle with accurately decomposing reflectance. Additionally, NeuFace does not consider the effect of subsurface scattering on facial skin, resulting in less realistic PBR [49]. In this work, with our novel facial template and the single-person refinement module, which further considers subsurface scattering effects, we can robustly decompose geometry and reflectance in a more realistic manner from sparse-view images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Face reconstruction. Given the inherent similarities among human faces, it is intuitive to develop a template that serves as a prior to enhance facial reconstruction. Based on a large amount of facial data, 3DMM [6] builds a 3D morphable face to describe the transformation of shape and expression, which is widely used in face recognition [56], 3D reconstruction, face editing [17, 36], etc. Using 3D morphable faces also enables face reconstruction from monocular images [14, 16, 27, 35, 61]. For instance, Han et al. [20] build a 3D morphable face with reflectance on BFM09 [38] for material editing. However, in face reconstruction, the use of 3D morphable faces often leads to a loss of detail during the compression and fitting process, resulting in overly coarse reconstructions. Moreover, given limited view information, monocular methods typically require reflectance ground truth for training [41]. As implicit neural representation shows extraordinary efficacy for 3D representation," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26410" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "numerous studies [47, 52, 57, 59] construct 3D parametric face or head models based on neural networks, yielding improved geometric details for faces or heads. Besides, some facial capturing methods utilize more informative inputs such as videos [46] and polarized images [2], enabling high-quality reconstruction across a wider range of scenarios. Our goal is to achieve face reconstruction solely from sparse-view images, without any reflectance ground truth. Our Stage 1 employs an implicit representation to construct a facial template, incorporating prior knowledge of facial geometry and reflectance. This approach enables robust decomposition and realistic face reconstruction during the single-person refinement process under sparse views." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.218, + 0.318 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.508 + ], + "angle": 0, + "content": "The effectiveness of PBR has been validated by numerous studies [25, 45, 58] for photo-realistic rendering in scenarios with complex reflection properties. Inspired by these works and as illustrated in Fig. 3, we formulate the output radiance \\( L_{\\mathrm{o}} \\in \\mathbb{R}^{3} \\) at surface position \\( \\mathbf{x} \\in \\mathbb{R}^{3} \\) along output direction \\( \\omega_{\\mathrm{o}} \\in \\mathbb{R}^{3} \\) as: \\( L_{\\mathrm{o}}(\\mathbf{x}, \\omega_{\\mathrm{o}}) = \\int_{\\Omega} f(\\mathbf{x}, \\omega_{\\mathrm{i}}, \\omega_{\\mathrm{o}}) L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})(\\omega_{\\mathrm{i}} \\cdot \\mathbf{n}) \\, \\mathrm{d}\\omega_{\\mathrm{i}} \\), where \\( f \\) represents the BRDF evaluation, \\( L_{\\mathrm{i}} \\in \\mathbb{R}^{3} \\) is the incident light intensity from input direction \\( \\omega_{\\mathrm{i}} \\in \\mathbb{R}^{3} \\), \\( \\mathbf{n} \\in \\mathbb{R}^{3} \\) is the surface normal. To decouple the view-dependent and view-independent radiance, we decompose \\( L_{\\mathrm{o}} \\) into the diffuse term and the specular term by:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.515, + 0.468, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {0} (\\mathbf {x}, \\boldsymbol {\\omega} _ {0}) = \\underbrace {\\frac {b _ {\\mathrm {a}} (\\mathbf {x})}{\\pi} \\int_ {\\Omega} L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}}) (\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}}} _ {\\text {d i f f u s e}} \\\\ + \\underbrace {\\int_ {\\Omega} f _ {s} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}} , \\boldsymbol {\\omega} _ {\\mathrm {o}} ; b _ {\\mathrm {s}} (\\mathbf {x}) , b _ {\\mathrm {r}} (\\mathbf {x})) L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}}) (\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}} ,} _ {\\text {s p e c u l a r}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.719 + ], + "angle": 0, + "content": "where \\( b_{\\mathrm{a}}(\\mathbf{x}) \\in \\mathbb{R}^3 \\), \\( b_{\\mathrm{s}}(\\mathbf{x}) \\in \\mathbb{R}^4 \\), \\( b_{\\mathrm{r}}(\\mathbf{x}) \\in \\mathbb{R} \\) are diffuse albedo, specular, and roughness fields obtained from BRDF, respectively. \\( f_{s} \\) represents the BRDF evaluation that only accounts for the specular property. Furthermore, we improve the efficiency of integration computation by using the approximated pre-computed method [9] as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.726, + 0.507, + 0.786 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {o}} (\\mathbf {x}, \\boldsymbol {\\omega} _ {\\mathrm {o}}) \\approx \\underbrace {\\frac {b _ {\\mathrm {a}} (\\mathbf {x})}{\\pi} \\hat {L} _ {\\mathrm {i}} (\\mathbf {n} , 1)} _ {\\text {d i f f u s e}} + \\underbrace {\\hat {f} _ {\\mathrm {s}} (\\mathbf {x} ; b _ {\\mathrm {s}} (\\mathbf {x}) , b _ {\\mathrm {r}} (\\mathbf {x})) \\hat {L} _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {\\mathrm {r}} , b _ {\\mathrm {r}} (\\mathbf {x})\\right)} _ {\\text {s p e c u l a r}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.861 + ], + "angle": 0, + "content": "where \\(\\omega_{\\mathrm{r}} \\in \\mathbb{R}^{3}\\) represent the mirrored view direction, \\(\\hat{L}_i\\) is the integration of input light, \\(\\hat{f}_s\\) is the integration of specular reflectance. Specifically, following [26], we employ an MLP to simulate two look-up-textures (LUT) \\(B_0\\) and \\(B_1\\) to calculate the BRDF integration as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.868, + 0.496, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {\\mathrm {s}} (\\mathbf {x}) = b _ {\\mathrm {s}} (\\mathbf {x}) \\left[ F _ {0} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}, \\mathbf {n}\\right) B _ {0} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}} \\cdot \\mathbf {n}, b _ {\\mathrm {r}} (\\mathbf {x})\\right) + B _ {1} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}} \\cdot \\mathbf {n}, b _ {\\mathrm {r}} (\\mathbf {x})\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "where \\(F_{0}\\) is the Fresnel term at normal incidence as [9]. For light representation, we use spherical harmonics [44] to approximate environment light by:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.15, + 0.907, + 0.192 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {i}} \\left(\\omega_ {\\mathrm {i}}\\right) \\approx \\sum_ {\\ell = 0} ^ {k} \\sum_ {m = - \\ell} ^ {\\ell} c _ {\\ell} ^ {m} Y _ {\\ell} ^ {m} \\left(\\omega_ {\\mathrm {i}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.907, + 0.388 + ], + "angle": 0, + "content": "where \\( Y \\) represents spherical basis functions and \\( c \\) denotes corresponding coefficients. Considering that the facial material is not of high specular reflectance, we set the order as \\( k = 10 \\) [12]. To integrate incident light rays, the light for the specular term can be integrated to the mirrored direction of the observation view, denoted as \\( \\hat{L}_{\\mathrm{i}}(\\omega_{\\mathrm{r}},b_{\\mathrm{r}}) \\), and light for the diffuse term (Lambertian) can be approximated by \\( \\hat{L}_{\\mathrm{i}}(\\mathbf{n},1) \\) [9], which assumes light in the normal direction hits a completely rough surface. For our SH light representation, we integrate \\( \\hat{L}_{\\mathrm{i}}(\\omega_{\\mathrm{r}},b_{\\mathrm{r}}) \\) for the specular term and \\( \\hat{L}_{\\mathrm{i}}(\\mathbf{n},1) \\) for the diffuse term through an approximation using the first ten orders of spherical harmonics as [58]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.402, + 0.605, + 0.418 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.611 + ], + "angle": 0, + "content": "Typical human faces share common features, such as two eyes and one nose, with variations in characteristics such as shape, skin color, and skin texture. To better utilize the shared features of human faces, we divide the decomposition process into two stages to learn the neural fields of human faces from multi-view images. Fig. 2 presents the architecture of the Sparse-view Face Decomposition Model (SFDM). In Stage 1, we create a general facial template with geometry and reflectance attributes from a collection of portraits of multiple individuals. In Stage 2, we refine the specific attributes for each individual face model based on the learned facial template under sparse-view settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.756, + 0.64 + ], + "angle": 0, + "content": "4.1. Generating facial template" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.69 + ], + "angle": 0, + "content": "Given a collection of portraits of multiple individuals \\(\\{\\mathcal{I}_i\\}\\), we first learn a general facial template with both geometry and reflectance attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.768 + ], + "angle": 0, + "content": "Geometry template. Following the previous work [47, 57], we jointly employ a geometry template network \\( T_{\\mathrm{geo}} \\) and a deformation network \\( F_{\\mathrm{dfm}} \\) to reconstruct the geometry (represented by signed distance field \\( S \\)) from given images. As illustrated in Fig. 2, we formulate the geometry network as:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.785, + 0.907, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (\\mathbf {x}) = T _ {\\text {g e o}} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\text {g e o}})\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} \\in \\mathbb{R}^3\\) is a query point, \\(S(\\mathbf{x}) \\in \\mathbb{R}\\) returns the signed distance value. We take \\(\\mathbf{z}_{\\mathrm{geo}} \\in \\mathbb{R}^{128}\\) as an identity-related latent code that encodes geometry, so we can get the unique SDF for each individual. Similar to VolSDF [51], we convert the SDF value to density \\(\\sigma\\) and then feed it into the volume rendering block." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.323, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 2. Pipeline of SFDM. We propose a geometry and reflectance decomposition framework for 3D face reconstruction from multiview images. We first learn a general facial template from multi-view images of multiple individuals in Stage 1. Then the learned facial template guides the further attribute refinement for each individual face model in Stage 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.482, + 0.47 + ], + "angle": 0, + "content": "Reflectance template. Apart from geometry attributes, we also consider reflectance attributes, which are equally crucial and share similar features across human faces. Therefore, we aim to create a coarse but general template that captures the underlying common reflectance characteristics of human faces in Stage 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.483, + 0.592 + ], + "angle": 0, + "content": "To obtain the reflectance attributes, we approximate the Bidirectional Reflectance Distribution Function (BRDF) by the network that consists of a reflectance template \\( T_{\\mathrm{rfl}} \\) and a BRDF offset network \\( F_{\\mathrm{ofs}} \\). Similar to the geometry template, we leverage \\( T_{\\mathrm{rfl}} \\) to fit the general facial BRDF for multiple individuals and then offset the BRDF value of the query points on each face model through \\( F_{\\mathrm{ofs}} \\). We formulate the calculation of BRDF \\( \\mathcal{B}_{\\mathrm{st1}}(\\mathbf{x}) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.602, + 0.483, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {\\mathrm {s t l}} (\\mathbf {x}) = T _ {\\mathrm {r f l}} \\left[ F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {g e o}}) \\right] + w F _ {\\mathrm {o f s}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {r f l}}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where the query points \\(\\mathbf{x}\\) in the observation space are deformed to the surface of the facial template through \\(F_{\\mathrm{dfm}}\\), and \\(T_{\\mathrm{rfl}}\\) return the corresponding BRDF attributes \\(\\mathcal{B}(\\mathbf{x}) = [b_{\\mathrm{a}}(\\mathbf{x}), b_{\\mathrm{s}}(\\mathbf{x}), b_{\\mathrm{r}}(\\mathbf{x})] \\in \\mathbb{R}^8\\). Subsequently, the specific reflectance attributes for each individual are determined by learning the BRDF offsets with \\(F_{\\mathrm{ofs}}\\). To capture the unique reflectance characteristics associated with individual identities, we incorporate a reflectance latent code \\(\\mathbf{z}_{\\mathrm{rfl}} \\in \\mathbb{R}^{128}\\). Moreover, we adjust the offset using a weighting parameter \\(w\\), setting it to a low value of \\(w = 0.05\\). This choice reflects our preference for the template network to primarily capture reflectance properties, minimizing dependence on the BRDF offset network. To simulate environment light, we apply the spherical harmonics as in Eq. (3) and integrate the lights as in [58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After obtaining the geometry, reflectance, and environment light information, we calculate the final color (output radiance) as Eq. (1-3) and integrate it by volume rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.425 + ], + "angle": 0, + "content": "As shown in Fig. 1, a general facial template can be obtained and robustly decomposed into geometry, diffuse, and specular attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.908, + 0.775 + ], + "angle": 0, + "content": "Albedo gradient predictor. Considering the wide range of skin color differences among various ethnic groups, simply relying on the absolute values of albedo on the facial template does not adequately convey the prior knowledge of faces. We observe that the continuity of facial albedo variations across 3D positions exhibits certain regularities, such as drastic changes around the edges of the lips and the eyes. Thus, we introduce the concept of albedo gradient as an additional attribute for the facial template. Specifically, during the training phase of Stage 1, we calculate the gradient of the albedo values \\(\\nabla \\mathbf{b}_{\\mathrm{a}}(\\mathbf{x}) \\in \\mathbb{R}^{3}\\) from the fitted neural field with respect to changes in 3D coordinates \\(\\mathbf{x} \\in \\mathbb{R}^{3}\\). This gradient plot effectively outlines facial contours like facial sketches. Following this, we use it to supervise the training of an additional albedo gradient predictor branch \\(\\mathcal{G}(\\mathbf{x})\\) within the reflectance template, guiding the model to pay attention to the patterns of albedo changes. During training, once the point \\(F_{\\mathrm{dfm}}(\\mathbf{x})\\) in the template space is obtained, it serves as the input for \\(\\mathcal{G}\\) and the output target is \\(\\nabla \\mathbf{b}_{\\mathrm{a}}(\\mathbf{x})\\) in the observation space. In the end, the well-trained albedo gradient predictor provides more comprehensive information and precise guidance for single-person refinement in Stage 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.896, + 0.803 + ], + "angle": 0, + "content": "4.2. Individual decomposition under sparse views" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.809, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Taking the facial template learned from Stage 1 as prior, we perform a fine-detailed decomposition from sparse-view images of a single face in Stage 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Geometry displacement and BRDF offset. The geometry template \\( T_{\\mathrm{geo}} \\) only represents a rough shape with shared features of multiple human faces. Additionally, the defor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26412" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.243 + ], + "angle": 0, + "content": "mation net \\(F_{\\mathrm{dfm}}\\) transforms the points on the surface of the facial template to the surface of a specific individual's face. However, the transformation is constrained, making it challenging to create topological features beyond the template domain and capture high-frequency details. Consequently, we incorporate a displacement network \\(F_{\\mathrm{dis}}\\) in Stage 2 to modify the SDF obtained from the geometry template for a more accurate reconstruction. Building on Eq. (4), we incorporate the offset term as follows: \\(\\mathcal{S}(\\mathbf{x}) = T_{\\mathrm{geo}}(F_{\\mathrm{dfm}}(\\mathbf{x},\\mathbf{z}_{\\mathrm{geo}})) + F_{\\mathrm{dis}}(\\mathbf{x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Based on the reflectance template \\( T_{\\mathrm{rfl}} \\) learned from Stage 1, we calculate the offset value of BRDF through \\( F_{\\mathrm{ofs}} \\), similar to Stage 1, to obtain the specific BRDF attributes for each individual. To address the distinct diversity between diffuse albedo \\( b_{a} \\), specular \\( b_{s} \\) and roughness \\( b_{r} \\), we add a learnable weight term \\( \\mathbf{W}(\\mathbf{x}) = [\\mathbf{w}_{\\mathrm{a}}, \\mathbf{w}_{\\mathrm{s}}, \\mathbf{w}_{\\mathrm{r}}] \\) to control the contribution of each BRDF attribute from the reflectance template \\( T_{\\mathrm{rfl}} \\), where \\( \\mathbf{w}_{\\mathrm{a}} \\in \\mathbb{R}^{3}, \\mathbf{w}_{\\mathrm{s}} \\in \\mathbb{R}^{4}, \\mathbf{w}_{\\mathrm{r}} \\in \\mathbb{R} \\) are the weights for diffuse albedo, specular, and roughness, respectively. Thus, we obtain the BRDF attributes as:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.408, + 0.482, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {\\mathrm {s t 2}} (\\mathbf {x}) = \\mathbf {W} (\\mathbf {x}) \\cdot T _ {\\mathrm {r f l}} \\left[ F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {g e o}}) \\right] + F _ {\\mathrm {o f s}} (\\mathbf {x}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.481 + ], + "angle": 0, + "content": "Considering that the diffuse albedo attribute has a large variation between different people, we initialize \\(\\mathbf{w}_{\\mathrm{a}}\\) with relatively low values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.484, + 0.709 + ], + "angle": 0, + "content": "Attribute synergy. In our framework, we decompose facial attributes into two components: geometry and reflectance. These components are not isolated but can enhance each other cooperatively. Sometimes, distinguishing between geometry and reflectance can be challenging, leading to inaccuracies in the decomposition process. For example, a dark region in a facial image might result from sunken areas or pigmentation spots on the face. Conventional neural rendering techniques rely solely on color information from images, making it hard to distinguish such ambiguous scenes, often leading to incorrect reconstruction. Since our method can decouple rough geometry and reflectance attributes in Stage 1, we leverage the prior from the reflectance template to facilitate geometry inference, and vice versa. Specifically, we obtain the displacement \\(\\mathcal{D}(\\mathbf{x})\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.722, + 0.483, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} (\\mathbf {x}) = F _ {\\text {d i s}} (\\mathbf {x}, \\operatorname {C o n c a t} \\left(\\mathbf {f} _ {\\mathrm {d e f}}, \\mathbf {f} _ {\\mathrm {g e o}}, \\mathbf {f} _ {\\mathrm {r f l}}, \\mathcal {G} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x})\\right)\\right)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.858 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{\\mathrm{dfm}} \\in \\mathbb{R}^{128}\\), \\(\\mathbf{f}_{\\mathrm{geo}} \\in \\mathbb{R}^{128}\\), and \\(\\mathbf{f}_{\\mathrm{rfl}} \\in \\mathbb{R}^{128}\\) are extracted features from the output of the final hidden layer of \\(F_{\\mathrm{dfm}}\\), \\(T_{\\mathrm{geo}}\\), and \\(T_{\\mathrm{rfl}}\\), respectively. \\(\\mathcal{G}(F_{\\mathrm{dfm}}(\\mathbf{x}))\\) is the output from the albedo gradient predictor. We also extract another feature vector \\(\\mathbf{f}_{\\mathrm{dis}} \\in \\mathbb{R}^{64}\\) from \\(F_{\\mathrm{dis}}\\) to provide more geometry information to the BRDF offset network \\(F_{\\mathrm{ofs}}\\). Similarly, we calculate the BRDF offset \\(\\mathcal{B}_{\\mathrm{ofs}}(\\mathbf{x})\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {\\mathrm {o f s}} (\\mathbf {x}) = F _ {\\mathrm {o f s}} (\\mathbf {x}, \\operatorname {C o n c a t} \\left(\\mathbf {f} _ {\\mathrm {d e f}}, \\mathbf {f} _ {\\mathrm {g e o}}, \\mathbf {f} _ {\\mathrm {d i s}}, \\mathbf {f} _ {\\mathrm {r f l}}, \\mathcal {G} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x})\\right)\\right)). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "Our ablation study validates the effectiveness of this attribute synergy module in the accuracy of both geometry reconstruction and rendering results." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.15, + 0.871, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.27, + 0.906, + 0.313 + ], + "angle": 0, + "content": "Figure 3. Left: BRDF rendering, where light reflects at the incident point. Right: Subsurface scattering on facial skin, showing the offset between the incident and outgoing points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.751 + ], + "angle": 0, + "content": "Subsurface scattering offset. Human facial skin exhibits a complexity far beyond that of non-biological materials like metal and plastic. Building on insights from previous research [15], rendering realistic faces necessitates accounting for subsurface scattering effects. As illustrated in Fig. 3, this phenomenon can be modeled using a three-layer approach, where specular reflections occur directly at the point of light incidence on the surface, while diffuse reflections involve light penetrating the skin, traversing three layers, undergoing multiple subsurface reflections, and then emerging from the skin surface slightly away from the point of incidence. This effect, known as subsurface scattering, is crucial for realistic skin rendering and can be mathematically represented as \\( L_{\\mathrm{o}}(\\mathbf{x}_{\\mathrm{o}}, \\omega_{\\mathrm{o}}) = \\int_{A} \\int_{\\Omega} S(\\mathbf{x}_{\\mathrm{o}}, \\omega_{\\mathrm{o}}, \\mathbf{x}_{\\mathrm{i}}, \\omega_{\\mathrm{i}}) L_{\\mathrm{i}}(\\mathbf{x}_{\\mathrm{i}}, \\omega_{\\mathrm{i}})(\\boldsymbol{\\omega}_{\\mathrm{i}} \\cdot \\mathbf{n}) \\, \\mathrm{d}\\omega_{\\mathrm{i}} \\, \\mathrm{d}A \\), where \\( \\mathbf{x}_{\\mathrm{i}} \\) and \\( \\mathbf{x}_{\\mathrm{o}} \\) are the positions of the light incident and output points, respectively, \\( A \\) is a small area around \\( \\mathbf{x}_{\\mathrm{o}} \\), and \\( S \\) is the scattering evaluation. In practice, the BRDF mainly accounts for light that scatters back at the point where it hits a surface, but simulating surface scattering accurately is computationally expensive. To improve face reconstruction quality, we integrate subsurface scattering into the BRDF's diffuse component as an offset. We simplify the process by assuming uniform subsurface scattering properties within a small area and utilize scattering profiles to approximate this effect. As illustrated in Fig. 3, these profiles show how the reflectance due to subsurface scattering changes with the distance between where light enters and exits the skin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.752, + 0.906, + 0.842 + ], + "angle": 0, + "content": "As [28] suggests, one can approximate each RGB channel of the scattering profiles using a combination of six Gaussian functions. We employ the weights of these Gaussian functions to form a feature vector \\(\\mathbf{f}_{\\mathrm{sp}}^{\\mathbf{x}} \\in \\mathbb{R}^{18}\\), which represents the scattering profile of the point \\(\\mathbf{x}\\). Consequently, we calculate the offset of the diffuse term as:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.853, + 0.905, + 0.873 + ], + "angle": 0, + "content": "\\[\nF _ {\\mathrm {s s}} (\\mathbf {x}) = v \\hat {F} _ {\\mathrm {s p}} (\\mathbf {x}, \\mathbf {f} _ {\\mathrm {s p}} ^ {\\mathbf {x}}) \\hat {L} _ {\\mathrm {i}} (\\mathbf {x}, \\mathcal {E} (\\mathbf {x})), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.885, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(v\\) is the weight of scattering offset, \\(\\hat{F}_{\\mathrm{sp}}\\) is a network" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26413" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.182 + ], + "angle": 0, + "content": "to integrate scattering reflectance, \\(\\hat{L}_{\\mathrm{i}}\\) is a network to integrate the incoming light in a small region around \\(\\mathbf{x}\\), and \\(\\mathcal{E}(\\mathbf{x}) \\in \\mathbb{R}^3\\) is a learnable vector that defines the size of the integration domain. The values of \\(\\mathbf{f}_{\\mathrm{sp}}^{\\mathbf{x}}\\) and \\(\\mathcal{E}(\\mathbf{x})\\) are obtained through a parameter network. We refer readers to the supplementary for more details of the network architecture." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.195, + 0.242, + 0.21 + ], + "angle": 0, + "content": "4.3. Training losses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.414 + ], + "angle": 0, + "content": "For both stages, we consider reconstruction losses including the pixel-wise color loss \\(\\mathcal{L}_{\\mathrm{col}} = \\| \\mathbf{C} - \\mathbf{C}_{\\mathrm{gt}}\\| _1\\) and the eikonal loss \\(\\mathcal{L}_{\\mathrm{eik}} = (\\| \\nabla S\\| _2 - 1)^2\\). Similar to NeuFace [58], we incorporate a regularization term \\(\\mathcal{L}_{\\mathrm{light}}\\) to assume a nearly white environment light and penalize the specular energy by \\(\\mathcal{L}_{\\mathrm{spec}}\\). Additionally, we add \\(\\mathcal{L}_{\\mathrm{code}} = \\| \\mathbf{z}_{\\mathrm{geo}}\\| _2 + \\| \\mathbf{z}_{\\mathrm{rfl}}\\| _2\\) to regularize the latent codes. To avoid a large variation between the predict reflectance attributes and the template during training, we apply a constraint to BRDF offset as \\(\\mathcal{L}_{\\mathrm{ofs}} = \\| \\mathcal{B}_{\\mathrm{ofs}}(\\mathbf{x})\\| _1\\). Additionally, the albedo gradient predictor employs \\(\\mathcal{L}_{\\mathrm{g}}\\), an L1 loss, which is supervised by the computed albedo gradient. As a result, we formulate the loss function in Stage 1 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.428, + 0.482, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s t 1}} = \\mathcal {L} _ {\\mathrm {c o l}} + \\mathcal {L} _ {\\mathrm {e i k}} + \\mathcal {L} _ {\\mathrm {l i g h t}} + \\mathcal {L} _ {\\mathrm {s p e c}} + \\mathcal {L} _ {\\mathrm {o f s}} + \\mathcal {L} _ {\\mathrm {c o d e}} + \\mathcal {L} _ {\\mathrm {g}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.534 + ], + "angle": 0, + "content": "For the second stage, we add additional displacement loss \\(\\mathcal{L}_{\\mathrm{dis}}\\) and scatter offset loss \\(\\mathcal{L}_{\\mathrm{ss}}\\): \\(\\mathcal{L}_{\\mathrm{st2}} = \\mathcal{L}_{\\mathrm{st1}} + \\mathcal{L}_{\\mathrm{dis}} + \\mathcal{L}_{\\mathrm{ss}}\\), where \\(\\mathcal{L}_{\\mathrm{dis}} = \\| \\mathcal{D}(\\mathbf{x})\\|_1\\) and \\(\\mathcal{L}_{\\mathrm{ss}} = \\| F_{\\mathrm{ss}}(\\mathbf{x})\\|_1\\) are regularization losses aimed at constraining high values of the displacement and the subsurface scattering offset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.291, + 0.567 + ], + "angle": 0, + "content": "5. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.279, + 0.592 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Dataset and metrics. We evaluate our method on a large human face dataset Facescape [48]. Specifically, we utilize authorized data from approximately 30 individuals, 4 of whom are from the publishable list. To protect the privacy of other identities that are not on the publishable list, we employed mosaic technology in the eyes region. To validate the generalization performance of our method, we additionally use dataset H3DS [40] in the supplementary material. We assess the quality of image rendering using the PSNR, SSIM, and LPIPS [54] metrics, and the Chamfer distance (CD) for evaluating the accuracy of geometry reconstruction. We compare our method with two 3D reconstruction methods VolSDF [51], DeformHead [47] and three BRDF decomposition methods PhySG [53], TensoIR [24], and NeuFace [58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation details. Facescape has over 50 scans of human heads along with camera parameters and meshes. We preprocess the images to obtain face masks using ImFace [57] tools. For Stage 1 training, we select 10 subjects, each with 10 views. For Stage 2, we use an additional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "20 subjects and choose 3, 5, and 10 views from the available \\(50+\\) views as the training set, with another 10 views reserved for testing. The 10-view experiments specifically evaluate performance in a low-view setting. All test data is independent of the training set and is used consistently across all methods for comparison. We employ the Adam optimizer with an initial learning rate \\(1 \\times 10^{-4}\\) to train both stages for 3000 epochs. Additionally, we add a calibration module as [58] to mitigate the exposure differences between various views and apply this calibration to all the methods for a fair comparison. For more implementation details, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.748, + 0.303 + ], + "angle": 0, + "content": "5.2. Evaluation & comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.55 + ], + "angle": 0, + "content": "To corroborate the effectiveness of our proposed framework in the reconstruction of appearance and geometry, we conducted a comprehensive comparison with state-of-the-art methods. These methods are categorized into two groups: geometry-oriented methods, which include VolSDF [51] and DeformHead [47]; and decomposition-based methods, including PhySG [53], TensoIR [24], NeuFace [58], and our method SFDM. As shown in Tab. 1, we enumerate the quantitative performance of different methods in the evaluation metrics. To assess the performance capability under sparse-view images, we conducted tests on 20 subjects with 3, 5, and 10 views, respectively. The visual results of the reconstruction from 3 views are presented in Fig. 4, showcasing the synthesized rendering images from a novel viewpoint, the geometric reconstruction, and the decomposition of diffuse and specular attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.905, + 0.869 + ], + "angle": 0, + "content": "Decomposition-based methods. Our method exhibits superior performance in novel view rendering and geometry reconstruction compared to other reflectance decomposition methods. In contrast to PhySG and TensoIR, designed for common objects, our method significantly outperforms them across all metrics. This is due to the complexity of facial material reflectance, which requires specialized reflectance modeling. NeuFace, which is customized for facial skin, poses a challenge due to its non-robust decomposition of geometry and reflectance attributes. As shown in Fig. 4, it often happens in NeuFace that all reflectance is baked into albedo, but the specular term is entirely absent. This degenerate Lambertian model may reduce the realism of the synthesized rendering output, especially for elements such as teeth, eyes, or the oily layer on facial skin, which often exhibit specular reflection effects. In contrast, we are able to stably decompose the diffuse and specular properties, resulting in more realistic rendering effects. Additionally, we demonstrate higher accuracy in geometric reconstruction, whereas the quality of NeuFace's geometry deteriorates significantly as the number of views decreases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Geometry-oriented methods. Our method demonstrates a comparable level of accuracy in geometry reconstruc" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26414" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.887, + 0.204 + ], + "angle": 0, + "content": "
MethodRelight3 views5 views10 views
PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓
VolSDF [51]✗22.950.8686.9218.2825.450.8976.097.4027.340.9105.185.25
DeformHead [47]✗26.020.8865.769.6026.410.8794.838.4427.350.8915.205.42
PhySG [53]✓22.030.8637.8170.8722.690.8727.0563.2023.110.8767.1255.7
TensoIR [24]✓22.350.6457.1428.6024.610.7045.8617.1026.140.7694.7310.10
NeuFace [58]✓23.540.8765.7815.7425.810.9044.758.6227.190.9144.036.40
Ours✓25.540.9024.469.2726.730.9113.837.0027.640.9203.865.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.213, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 1. Evaluation. The LPIPS and CD metrics are measured on scales of \\( {10}^{-2} \\) and \\( {10}^{-3} \\) ,respectively. The best and second-best results are highlighted in red and orange, respectively." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.258, + 0.468, + 0.749 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.8 + ], + "angle": 0, + "content": "Figure 4. Visualization of 3-view reconstruction results. VolSDF and DeformHead do not incorporate reflectance attributes, resulting in the absence of corresponding outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "tion compared to VolSDF and DeformHead. VolSDF exhibits poor 3D reconstruction results in sparse view settings. While DeformHead achieves a more detailed 3D reconstruction utilizing a geometry template, it is overly sensitive to geometry deformation, leading to noticeable indenta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.907, + 0.355 + ], + "angle": 0, + "content": "tions. Although the PSNR of our method is lower than that of DeformHead under 3 views, our LPIPS outperforms it, indicating more photorealistic synthesis results. Failure instances in reconstruction have been observed with VolSDF and DeformHead, with failure rates of \\(21.7\\%\\) and \\(6.7\\%\\), respectively. We exclude these failures from the comparison." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.366, + 0.907, + 0.472 + ], + "angle": 0, + "content": "
ExpMethodPSNR↑SSIM↑LPIPS↓CD↓
1w/o facial template24.270.8575.5917.22
2w/o albedo gradient25.130.8774.668.83
3w/o attribute synergy24.790.8764.908.80
4w/o scattering offset24.600.8745.738.61
5Ours25.650.8824.457.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.48, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Table 2. Comparative evaluation metrics for ablation studies." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.51, + 0.889, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.643, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Figure 5. Ablation visualization. Top: Rendering. Bottom: Geometric errors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.676, + 0.713 + ], + "angle": 0, + "content": "5.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We select a subset of 10 subjects from the dataset to test the effectiveness of the major components of SFDM. In Tab. 2, Exp.1-2 are results for the components in Stage 1, and Exp.3-4 cover the components in Stage 2. Exp.1-2 indicate that the facial template significantly influences the robustness of face decomposition, with the albedo gradient predictor further enriching the reconstruction quality by introducing additional facial prior knowledge. From Exp.3, we observe that the attribute synergy module appears to contribute to geometry decomposition, and its absence leads to a noticeable decrease in geometric accuracy. Based on Exp.4, the subsurface scatter offset module can enhance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26415" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "the realism of the synthesized images from novel views, which is suggested by improvements in the LPIPS metric. In Fig. 5, columns 3-6 show the visualization results when omitting the facial template \\( T_{*} \\), albedo gradient predictor \\( \\mathcal{G} \\), attribute synergy module \\( \\mathbf{f}_{*} \\), and subsurface scattering \\( F_{\\mathrm{ss}} \\), respectively. The bottom row highlights the albedo gradient predictor's more significant impact on individuals who exhibit greater deviations from the facial template. In Stage 2, the subsurface scattering offset introduces additional details in light reflection on the skin, resulting in more photorealistic rendering results. Regarding the attribute synergy module, the exchange of information between geometry and reflectance enables a more accurate decomposition, as is evident in the details of geometry reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.313, + 0.398, + 0.329 + ], + "angle": 0, + "content": "5.4. Reflectance decomposition analysis" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.335, + 0.485, + 0.697 + ], + "angle": 0, + "content": "Our method focuses on reflectance decomposition from multi-view facial images without any ground truth reflectance data, enabling realistic reconstruction of facial texture and highlights for novel view rendering. While our goal is not a physically accurate decomposition of skin reflectance, we aim to evaluate our decomposition results both visually and quantitatively. To this end, we use the single-view face reconstruction model, RefMM [20], to fit and decompose images in the test set, generating a coarse diffuse and specular result as pseudo ground truth for evaluation. Tab. 3 shows that our decomposition accuracy generally outperforms that of NeuFace. The overall low SSIM values are primarily due to the coarse nature of the pseudo ground truth, as illustrated in Fig. 6. This reflects certain advantages and limitations of single-view decomposition methods. These methods are typically trained on extensive facial image datasets with specialized lighting conditions or pseudo reflectance ground truth, allowing them to robustly decompose faces from monocular images. However, because they rely heavily on a parameterized facial model and limited input with less 3D information, the reconstructed results lack many facial details, making the rendered images appear less realistic. As shown in Fig. 6 and Fig. 7, we can observe these differences visually in the rendered results." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.71, + 0.406, + 0.776 + ], + "angle": 0, + "content": "
MethodDiffuseSpecular
NeuFace0.7260.552
Ours0.7610.692
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.781, + 0.462, + 0.796 + ], + "angle": 0, + "content": "Table 3. Quantitative diffuse and specular results (SSIM↑)." + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.806, + 0.408, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.883, + 0.433, + 0.897 + ], + "angle": 0, + "content": "Figure 6. Decomposition results of RefMM [20]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.649, + 0.108 + ], + "angle": 0, + "content": "5.5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.238 + ], + "angle": 0, + "content": "Specular editing. Our robust decomposition of diffuse and specular reflectances allows us to easily adjust the specular reflectance, resulting in the ability to make faces appear shinier or dimmer. The specular editing outcomes are showcased in Fig. 7. Specifically, we increase the specular weight by 1.5 times, while maintaining the overall specular distribution unchanged. As a result, the edited images can retain their photorealistic appearance." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.25, + 0.903, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.396, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 7. Applications. (a) Specular editing by increasing the shininess. (b) Relighting under different environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Relighting. Decomposing reflectance and environment light facilitates the relighting of human faces without altering their inherent properties. As shown in Fig. 7, we conducted a comparison of the relighting effects between our method and NeuFace in sparse view settings. Our method excels in decomposing specular reflectance, leading to more vibrant and realistic relighting outcomes under varying environmental lighting conditions [12]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.634, + 0.599 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.807 + ], + "angle": 0, + "content": "We propose a novel face decomposition method capable of rendering and reconstructing 3D human faces from sparse-view images. Our framework first learns a facial template with both decomposed geometry and reflectance attributes from multi-view images of different individuals. Then we utilize the prior knowledge of the facial template and further consider the interaction between geometry and reflectance, as well as subsurface scattering effects, which enable our method to achieve robust face decomposition from sparse-view images. Based on decomposed facial attributes, our method achieves superior results in the rendering and geometric reconstruction of 3D human faces and can be utilized for relighting and specular editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations & future work. As we focus on face decomposition, it is hard to extend to general objects directly with our method. In the future, we could involve techniques such as hash encoding [33] to accelerate our method and leverage fine-detailed 3D prior or reflectance ground truth for a better decomposition effect." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26416" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.236 + ], + "angle": 0, + "content": "This work was supported in part by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22) and the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). J.Hu was supported by the National Natural Science Foundation of China under Grants 62402083." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.252, + 0.188, + 0.268 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.349, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.484, + 0.501 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[4] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hašan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[10] Chengqian Che, Fujun Luan, Shuang Zhao, Kavita Bala, and Ioannis Gkioulekas. Towards learning-based inverse subsurface scattering. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[11] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[12] Zhang Chen, Anpei Chen, Guli Zhang, Chengyuan Wang, Yu Ji, Kiriakos N Kutulakos, and Jingyi Yu. A neural rendering framework for free-viewpoint relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5610, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[13] Ziang Cheng, Hongdong Li, Yuta Asano, Yinqiang Zheng, and Imari Sato. Multi-view 3d reconstruction of a textureless smooth surface of unknown generic reflectance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16226-16235, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[14] Abdallah Dib, Luiz Gustavo Hafemann, Emeline Got, Trevor Anderson, Amin Fadainejad, Rafael MO Cruz, and Marc-Andre Carbonneau. Mosar: Monocular semi-supervised model for avatar reconstruction using differentiable shading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1770–1780, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[15] Craig Donner and Henrik Wann Jensen. Light diffusion in multi-layered translucent materials. ACM Transactions on Graphics (ToG), 24(3):1032-1039, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[16] Haiwen Feng, Timo Bolkart, Joachim Tesch, Michael J Black, and Victoria Abrevaya. Towards racially unbiased skin tone estimation via scene disambiguation. In European Conference on Computer Vision, pages 72-90. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[19] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[20] Yuxuan Han, Zhibo Wang, and Feng Xu. Learning a 3d morphable face reflectance model from low-cost data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8598-8608, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[21] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Mingming He, Pascal Clausen, Ahmet Levent Taşel, Li Ma, Oliver Pilarski, Wenqi Xian, Laszlo Rikker, Xueming Yu, Ryan Burgert, Ning Yu, et al. Diffrelight: Diffusion-based" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26417" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "facial performance relighting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.124, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.483, + 0.265 + ], + "angle": 0, + "content": "[24] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.483, + 0.309 + ], + "angle": 0, + "content": "[25] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[26] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.356, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[27] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8629-8640, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.369, + 0.456 + ], + "angle": 0, + "content": "[28] Andrew Lauritzen. Gpu gems 3, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.458, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[30] Connor Lin, Koki Nagano, Jan Kautz, Eric Chan, Umar Iqbal, Leonidas Guibas, Gordon Wetzstein, and Sameh Khamis. Single-shot implicit morphable faces with consistent texture parameterization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[31] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Trans. Graph., 42(4): 114:1-114:22, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[35] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[36] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[38] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[39] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[40] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[41] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[42] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[43] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[44] Borom Tunwattanapong, Graham Fyffe, Paul Graham, Jay Busch, Xueming Yu, Abhijeet Ghosh, and Paul Debevec. Acquiring reflectance and shape from continuous spherical harmonic illumination. ACM Transactions on graphics (TOG), 32(4):1-12, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and re" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26418" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "lighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[47] Baixin Xu, Jiarui Zhang, Kwan-Yee Lin, Chen Qian, and Ying He. Deformable model-driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17924-17934, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[48] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[49] Jing Yang, Hanyuan Xiao, Wenbin Teng, Yunxuan Cai, and Yajie Zhao. Light sampling field and BRDF representation for physically-based neural rendering. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[51] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[52] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[55] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[56] Wenyi Zhao, Rama Chellappa, P Jonathon Phillips, and Azriel Rosenfeld. Face recognition: A literature survey. ACM computing surveys (CSUR), 35(4):399-458, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[57] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[58] Mingwu Zheng, Haiyu Zhang, Hongyu Yang, and Di Huang. Neuface: Realistic 3d neural face rendering from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16868-16877, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[59] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[60] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[61] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum, pages 523-550. Wiley Online Library, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26419" + } + ] +] \ No newline at end of file diff --git a/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_origin.pdf b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c725c34efc7bb10b1c3e9694776a7b763db4e53c --- /dev/null +++ b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/54e9e069-f5ec-481a-a9c4-0b86f22aae79_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c0a60ec862645e06c3155f501b14a77694c02eda4466b4491fda7b1eb7dc8c4 +size 3951755 diff --git a/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/full.md b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76bea356c703d777ff765b61c3b42ffb83c14d62 --- /dev/null +++ b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/full.md @@ -0,0 +1,301 @@ +# SFDM: Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images + +Daisheng Jin $^{1}$ Jiangbei Hu $^{1,2}$ Baixin Xu $^{1}$ Yuxin Dai $^{1}$ Chen Qian $^{3}$ Ying He $^{1*}$ $^{1}$ S-Lab, Nanyang Technological University, $^{2}$ Dalian University of Technology, $^{3}$ SenseTime Research + +# Abstract + +In this study, we introduce a novel two-stage technique for decomposing and reconstructing facial features from sparse-view images, a task made challenging by the unique geometry and complex skin reflectance of each individual. To synthesize 3D facial models more realistically, we endeavor to decouple key facial attributes from the RGB color, including geometry, diffuse reflectance, and specular reflectance. Specifically, we design a Sparse-view Face Decomposition Model (SFDM): 1) In the first stage, we create a general facial template from a wide array of individual faces, encapsulating essential geometric and reflectance characteristics. 2) Guided by this template, we refine a specific facial model for each individual in the second stage, considering the interaction between geometry and reflectance, as well as the effects of subsurface scattering on the skin. With these advances, our method can reconstruct high-quality facial representations from as few as three images. The comprehensive evaluation and comparison reveal that our approach outperforms existing methods by effectively disentangling geometric and reflectance components, significantly enhancing the quality of synthesized novel views, and paving the way for applications in facial relighting and reflectance editing. Visit our project page for more details https://kingjg.github.io/SFDM.github.io/. + +# 1. Introduction + +3D human face reconstruction finds widespread applications in various fields, including game design and film production. Developments in neural rendering techniques have shown promising performance and advantages in synthesizing photo-realistic images for complex scenes [3, 32, 33, 43]. Concurrently, a surge of research focusing on head and facial attribute reconstruction from images via implicit neural representations [18, 36, 37, 59] has emerged. These innovative approaches address the limitations associated with traditional explicit methods, such as those based on the 3D Morphable Model (3DMM) [6], which struggled with is + +![](images/151c96cf6e0f022bf42d81202e47e1e251ebe55d38faa536ed2a990aa5ccef4b.jpg) +Figure 1. Our method learns a general facial template with distinct key facial attributes from multiple individuals. This template enables photorealistic novel-view synthesis for each individual, as well as other applications. + +sues like dependency on mesh resolution and difficulties in managing topological variations. Furthermore, implicit neural representation methods demonstrate their effectiveness in 3D face reconstruction from low-view images [47], expression transfer [1, 23, 30], and animations [19, 59, 60]. + +In 3D face model rendering, both accurate shape attributes and high-quality reflectance properties are essential for synthesizing photorealistic faces. Thus, robustly decomposing key facial attributes including geometry, diffuse and specular reflectance, and environment light is a crucial task, which not only enhances the accuracy of face reconstruction but also opens up possibilities for more applications like relighting [7, 11, 22, 29, 55]. However, most previous studies only use neural networks to learn radiance and geometric information from images, neglecting the effect of rendering-related physical information such as facial skin characteristics and environmental lighting conditions. NeRV [42], PhySG [53], and following methods [8, 9, 13, 21, 31, 34] incorporate the Bidirectional Reflectance Distribution Functions (BRDFs) into the neural rendering framework, facilitating the decomposition and joint reconstruction of shape, reflectance, and illumination. Nonetheless, these methods primarily concentrate on non-biological materials, which are not suitable for decomposing face models that possess complex attributes. NeuFace [58] introduces an approxi + +mated BRDF integration and learns the physically meaningful underlying representations for 3D face reconstruction. Nevertheless, due to the ill-posedness of decomposition, NeuFace often converges to a suboptimal solution under sparse-view settings, in which specular reflectance is absent, and all radiance is derived from diffuse attributes. In addition, the accuracy of geometric reconstruction significantly diminishes as the number of views decreases. Therefore, it still poses a significant challenge to decompose shape, reflectance, and environment light attributes from sparse view images in 3D face model reconstruction. + +In this study, we introduce a novel two-stage face decomposition framework, termed Sparse-view Face Decomposition Model (SFDM), aimed at reconstructing 3D faces from sparse-view images. The pipeline of our proposed algorithm is depicted in Fig. 2. Based on the observation that human faces share similar geometric features and reflectance properties, we first derive a novel general facial template from images of a set of varied individual faces in Stage 1. In Stage 2, we employ the learned facial template with both geometry and reflectance attributes to refine the decomposition of each individual face model. In both stages, we decompose the shape and material attributes using a geometry module and a reflectance module, respectively. + +Specifically, the geometry module consists of a template network and a deformation network. In the first stage, we can get a template model with common geometry features based on the training of multi-view images from varied faces, which can be further transformed to a specific individual face model by deformation. Building on the facial geometry template, we innovatively propose a reflectance module that integrates reflectance information into the facial template. Specifically, we introduce the BRDF as a coarse representation of reflectance for facial skin. We employ a BRDF template network to learn the shared characteristics of diffuse and specular reflectance properties, and a BRDF offset network is utilized to reconstruct identity-specific reflectance details. Furthermore, we propose a novel albedo gradient predictor designed to capture more general facial reflectance attributes. In the second stage, based on the learned template with geometry and reflectance attributes, we refine the attribute details from sparse-view images of each individual face model. To this end, we incorporate a displacement network and a reflectance offset network on the learned template, employing a design that harmonizes geometry and reflectance information to capture high-frequency details of facial attributes. In addition, we introduce a diffuse offset network to take into account the effect of subsurface scattering, thus providing a more realistic simulation of the facial skin. We summarize the main contributions of this study as follows: + +- We propose a novel decomposition framework for human facial representations to learn a template with various at + +tributes, including geometric features, diffuse reflectance, and specular reflectance. + +- Based on the learned template, our tailored refinement stage enables robust decomposition of attributes for each individual face under as few as three images, attaining superior geometric accuracy and reflectance detail. +- Extensive experiments demonstrate the face decomposition results obtained through our framework can facilitate more realistic synthesis in novel views, as well as applications such as specular editing and relighting. + +# 2. Related work + +Physically based rendering. To achieve a photorealistic rendering effect, physically based rendering (PBR) [39] is proposed to simulate the light and surfaces using real-world optical principles. Reversely, physically based inverse rendering seeks a way to decompose light and surface factors from the images [4, 5, 8, 34]. IDR [50] proposes a surface reconstruction method that simultaneously disentangles the geometry and appearance. Subsequent works incorporate inverse rendering into the neural rendering pipeline, developing various methods for general objects [9, 53], and enhancing decomposition effects for objects with specialized materials [10, 31]. NeuFace [58] proposes a decomposition pipeline for human faces, taking into account the sophisticated properties of facial skin reflectance. However, without ground truth for reflectance components, face decomposition is an ill-posed problem, causing NeuFace to often struggle with accurately decomposing reflectance. Additionally, NeuFace does not consider the effect of subsurface scattering on facial skin, resulting in less realistic PBR [49]. In this work, with our novel facial template and the single-person refinement module, which further considers subsurface scattering effects, we can robustly decompose geometry and reflectance in a more realistic manner from sparse-view images. + +Face reconstruction. Given the inherent similarities among human faces, it is intuitive to develop a template that serves as a prior to enhance facial reconstruction. Based on a large amount of facial data, 3DMM [6] builds a 3D morphable face to describe the transformation of shape and expression, which is widely used in face recognition [56], 3D reconstruction, face editing [17, 36], etc. Using 3D morphable faces also enables face reconstruction from monocular images [14, 16, 27, 35, 61]. For instance, Han et al. [20] build a 3D morphable face with reflectance on BFM09 [38] for material editing. However, in face reconstruction, the use of 3D morphable faces often leads to a loss of detail during the compression and fitting process, resulting in overly coarse reconstructions. Moreover, given limited view information, monocular methods typically require reflectance ground truth for training [41]. As implicit neural representation shows extraordinary efficacy for 3D representation, + +numerous studies [47, 52, 57, 59] construct 3D parametric face or head models based on neural networks, yielding improved geometric details for faces or heads. Besides, some facial capturing methods utilize more informative inputs such as videos [46] and polarized images [2], enabling high-quality reconstruction across a wider range of scenarios. Our goal is to achieve face reconstruction solely from sparse-view images, without any reflectance ground truth. Our Stage 1 employs an implicit representation to construct a facial template, incorporating prior knowledge of facial geometry and reflectance. This approach enables robust decomposition and realistic face reconstruction during the single-person refinement process under sparse views. + +# 3. Preliminary + +The effectiveness of PBR has been validated by numerous studies [25, 45, 58] for photo-realistic rendering in scenarios with complex reflection properties. Inspired by these works and as illustrated in Fig. 3, we formulate the output radiance $L_{\mathrm{o}} \in \mathbb{R}^{3}$ at surface position $\mathbf{x} \in \mathbb{R}^{3}$ along output direction $\omega_{\mathrm{o}} \in \mathbb{R}^{3}$ as: $L_{\mathrm{o}}(\mathbf{x}, \omega_{\mathrm{o}}) = \int_{\Omega} f(\mathbf{x}, \omega_{\mathrm{i}}, \omega_{\mathrm{o}}) L_{\mathrm{i}}(\mathbf{x}, \omega_{\mathrm{i}})(\omega_{\mathrm{i}} \cdot \mathbf{n}) \, \mathrm{d}\omega_{\mathrm{i}}$ , where $f$ represents the BRDF evaluation, $L_{\mathrm{i}} \in \mathbb{R}^{3}$ is the incident light intensity from input direction $\omega_{\mathrm{i}} \in \mathbb{R}^{3}$ , $\mathbf{n} \in \mathbb{R}^{3}$ is the surface normal. To decouple the view-dependent and view-independent radiance, we decompose $L_{\mathrm{o}}$ into the diffuse term and the specular term by: + +$$ +\begin{array}{l} L _ {0} (\mathbf {x}, \boldsymbol {\omega} _ {0}) = \underbrace {\frac {b _ {\mathrm {a}} (\mathbf {x})}{\pi} \int_ {\Omega} L _ {\mathrm {i}} (\mathbf {x} , \boldsymbol {\omega} _ {\mathrm {i}}) (\boldsymbol {\omega} _ {\mathrm {i}} \cdot \mathbf {n}) \mathrm {d} \boldsymbol {\omega} _ {\mathrm {i}}} _ {\text {d i f f u s e}} \\ + \underbrace {\int_ {\Omega} f _ {s} (\mathbf {x} , \boldsymbol {\omega} _ {\mathrm {i}} , \boldsymbol {\omega} _ {\mathrm {o}} ; b _ {\mathrm {s}} (\mathbf {x}) , b _ {\mathrm {r}} (\mathbf {x})) L _ {\mathrm {i}} (\mathbf {x} , \boldsymbol {\omega} _ {\mathrm {i}}) (\boldsymbol {\omega} _ {\mathrm {i}} \cdot \mathbf {n}) \mathrm {d} \boldsymbol {\omega} _ {\mathrm {i}} ,} _ {\text {s p e c u l a r}} \\ \end{array} +$$ + +where $b_{\mathrm{a}}(\mathbf{x}) \in \mathbb{R}^3$ , $b_{\mathrm{s}}(\mathbf{x}) \in \mathbb{R}^4$ , $b_{\mathrm{r}}(\mathbf{x}) \in \mathbb{R}$ are diffuse albedo, specular, and roughness fields obtained from BRDF, respectively. $f_{s}$ represents the BRDF evaluation that only accounts for the specular property. Furthermore, we improve the efficiency of integration computation by using the approximated pre-computed method [9] as: + +$$ +L _ {\mathrm {o}} (\mathbf {x}, \boldsymbol {\omega} _ {\mathrm {o}}) \approx \underbrace {\frac {b _ {\mathrm {a}} (\mathbf {x})}{\pi} \hat {L} _ {\mathrm {i}} (\mathbf {n} , 1)} _ {\text {d i f f u s e}} + \underbrace {\hat {f} _ {\mathrm {s}} (\mathbf {x} ; b _ {\mathrm {s}} (\mathbf {x}) , b _ {\mathrm {r}} (\mathbf {x})) \hat {L} _ {\mathrm {i}} \left(\boldsymbol {\omega} _ {\mathrm {r}} , b _ {\mathrm {r}} (\mathbf {x})\right)} _ {\text {s p e c u l a r}}, \tag {1} +$$ + +where $\omega_{\mathrm{r}} \in \mathbb{R}^{3}$ represent the mirrored view direction, $\hat{L}_i$ is the integration of input light, $\hat{f}_s$ is the integration of specular reflectance. Specifically, following [26], we employ an MLP to simulate two look-up-textures (LUT) $B_0$ and $B_1$ to calculate the BRDF integration as: + +$$ +\hat {f} _ {\mathrm {s}} (\mathbf {x}) = b _ {\mathrm {s}} (\mathbf {x}) \left[ F _ {0} \left(\boldsymbol {\omega} _ {\mathrm {o}}, \mathbf {n}\right) B _ {0} \left(\boldsymbol {\omega} _ {\mathrm {o}} \cdot \mathbf {n}, b _ {\mathrm {r}} (\mathbf {x})\right) + B _ {1} \left(\boldsymbol {\omega} _ {\mathrm {o}} \cdot \mathbf {n}, b _ {\mathrm {r}} (\mathbf {x})\right) \right], \tag {2} +$$ + +where $F_{0}$ is the Fresnel term at normal incidence as [9]. For light representation, we use spherical harmonics [44] to approximate environment light by: + +$$ +L _ {\mathrm {i}} \left(\omega_ {\mathrm {i}}\right) \approx \sum_ {\ell = 0} ^ {k} \sum_ {m = - \ell} ^ {\ell} c _ {\ell} ^ {m} Y _ {\ell} ^ {m} \left(\omega_ {\mathrm {i}}\right), \tag {3} +$$ + +where $Y$ represents spherical basis functions and $c$ denotes corresponding coefficients. Considering that the facial material is not of high specular reflectance, we set the order as $k = 10$ [12]. To integrate incident light rays, the light for the specular term can be integrated to the mirrored direction of the observation view, denoted as $\hat{L}_{\mathrm{i}}(\omega_{\mathrm{r}},b_{\mathrm{r}})$ , and light for the diffuse term (Lambertian) can be approximated by $\hat{L}_{\mathrm{i}}(\mathbf{n},1)$ [9], which assumes light in the normal direction hits a completely rough surface. For our SH light representation, we integrate $\hat{L}_{\mathrm{i}}(\omega_{\mathrm{r}},b_{\mathrm{r}})$ for the specular term and $\hat{L}_{\mathrm{i}}(\mathbf{n},1)$ for the diffuse term through an approximation using the first ten orders of spherical harmonics as [58]. + +# 4. Method + +Typical human faces share common features, such as two eyes and one nose, with variations in characteristics such as shape, skin color, and skin texture. To better utilize the shared features of human faces, we divide the decomposition process into two stages to learn the neural fields of human faces from multi-view images. Fig. 2 presents the architecture of the Sparse-view Face Decomposition Model (SFDM). In Stage 1, we create a general facial template with geometry and reflectance attributes from a collection of portraits of multiple individuals. In Stage 2, we refine the specific attributes for each individual face model based on the learned facial template under sparse-view settings. + +# 4.1. Generating facial template + +Given a collection of portraits of multiple individuals $\{\mathcal{I}_i\}$ , we first learn a general facial template with both geometry and reflectance attributes. + +Geometry template. Following the previous work [47, 57], we jointly employ a geometry template network $T_{\mathrm{geo}}$ and a deformation network $F_{\mathrm{dfm}}$ to reconstruct the geometry (represented by signed distance field $S$ ) from given images. As illustrated in Fig. 2, we formulate the geometry network as: + +$$ +\mathcal {S} (\mathbf {x}) = T _ {\text {g e o}} \left(F _ {\mathrm {d f m}} (\mathbf {x}, \mathbf {z} _ {\text {g e o}})\right), \tag {4} +$$ + +where $\mathbf{x} \in \mathbb{R}^3$ is a query point, $S(\mathbf{x}) \in \mathbb{R}$ returns the signed distance value. We take $\mathbf{z}_{\mathrm{geo}} \in \mathbb{R}^{128}$ as an identity-related latent code that encodes geometry, so we can get the unique SDF for each individual. Similar to VolSDF [51], we convert the SDF value to density $\sigma$ and then feed it into the volume rendering block. + +![](images/607b10f409ce9cfb1ecc289feb583462d367727e0e2a1b8d20ba61cf2c25920a.jpg) +Figure 2. Pipeline of SFDM. We propose a geometry and reflectance decomposition framework for 3D face reconstruction from multiview images. We first learn a general facial template from multi-view images of multiple individuals in Stage 1. Then the learned facial template guides the further attribute refinement for each individual face model in Stage 2. + +Reflectance template. Apart from geometry attributes, we also consider reflectance attributes, which are equally crucial and share similar features across human faces. Therefore, we aim to create a coarse but general template that captures the underlying common reflectance characteristics of human faces in Stage 1. + +To obtain the reflectance attributes, we approximate the Bidirectional Reflectance Distribution Function (BRDF) by the network that consists of a reflectance template $T_{\mathrm{rfl}}$ and a BRDF offset network $F_{\mathrm{ofs}}$ . Similar to the geometry template, we leverage $T_{\mathrm{rfl}}$ to fit the general facial BRDF for multiple individuals and then offset the BRDF value of the query points on each face model through $F_{\mathrm{ofs}}$ . We formulate the calculation of BRDF $\mathcal{B}_{\mathrm{st1}}(\mathbf{x})$ as follows: + +$$ +\mathcal {B} _ {\mathrm {s t l}} (\mathbf {x}) = T _ {\mathrm {r f l}} \left[ F _ {\mathrm {d f m}} (\mathbf {x}, \mathbf {z} _ {\mathrm {g e o}}) \right] + w F _ {\mathrm {o f s}} (\mathbf {x}, \mathbf {z} _ {\mathrm {r f l}}), \tag {5} +$$ + +where the query points $\mathbf{x}$ in the observation space are deformed to the surface of the facial template through $F_{\mathrm{dfm}}$ , and $T_{\mathrm{rfl}}$ return the corresponding BRDF attributes $\mathcal{B}(\mathbf{x}) = [b_{\mathrm{a}}(\mathbf{x}), b_{\mathrm{s}}(\mathbf{x}), b_{\mathrm{r}}(\mathbf{x})] \in \mathbb{R}^8$ . Subsequently, the specific reflectance attributes for each individual are determined by learning the BRDF offsets with $F_{\mathrm{ofs}}$ . To capture the unique reflectance characteristics associated with individual identities, we incorporate a reflectance latent code $\mathbf{z}_{\mathrm{rfl}} \in \mathbb{R}^{128}$ . Moreover, we adjust the offset using a weighting parameter $w$ , setting it to a low value of $w = 0.05$ . This choice reflects our preference for the template network to primarily capture reflectance properties, minimizing dependence on the BRDF offset network. To simulate environment light, we apply the spherical harmonics as in Eq. (3) and integrate the lights as in [58]. + +After obtaining the geometry, reflectance, and environment light information, we calculate the final color (output radiance) as Eq. (1-3) and integrate it by volume rendering. + +As shown in Fig. 1, a general facial template can be obtained and robustly decomposed into geometry, diffuse, and specular attributes. + +Albedo gradient predictor. Considering the wide range of skin color differences among various ethnic groups, simply relying on the absolute values of albedo on the facial template does not adequately convey the prior knowledge of faces. We observe that the continuity of facial albedo variations across 3D positions exhibits certain regularities, such as drastic changes around the edges of the lips and the eyes. Thus, we introduce the concept of albedo gradient as an additional attribute for the facial template. Specifically, during the training phase of Stage 1, we calculate the gradient of the albedo values $\nabla \mathbf{b}_{\mathrm{a}}(\mathbf{x}) \in \mathbb{R}^{3}$ from the fitted neural field with respect to changes in 3D coordinates $\mathbf{x} \in \mathbb{R}^{3}$ . This gradient plot effectively outlines facial contours like facial sketches. Following this, we use it to supervise the training of an additional albedo gradient predictor branch $\mathcal{G}(\mathbf{x})$ within the reflectance template, guiding the model to pay attention to the patterns of albedo changes. During training, once the point $F_{\mathrm{dfm}}(\mathbf{x})$ in the template space is obtained, it serves as the input for $\mathcal{G}$ and the output target is $\nabla \mathbf{b}_{\mathrm{a}}(\mathbf{x})$ in the observation space. In the end, the well-trained albedo gradient predictor provides more comprehensive information and precise guidance for single-person refinement in Stage 2. + +# 4.2. Individual decomposition under sparse views + +Taking the facial template learned from Stage 1 as prior, we perform a fine-detailed decomposition from sparse-view images of a single face in Stage 2. + +Geometry displacement and BRDF offset. The geometry template $T_{\mathrm{geo}}$ only represents a rough shape with shared features of multiple human faces. Additionally, the defor + +mation net $F_{\mathrm{dfm}}$ transforms the points on the surface of the facial template to the surface of a specific individual's face. However, the transformation is constrained, making it challenging to create topological features beyond the template domain and capture high-frequency details. Consequently, we incorporate a displacement network $F_{\mathrm{dis}}$ in Stage 2 to modify the SDF obtained from the geometry template for a more accurate reconstruction. Building on Eq. (4), we incorporate the offset term as follows: $\mathcal{S}(\mathbf{x}) = T_{\mathrm{geo}}(F_{\mathrm{dfm}}(\mathbf{x},\mathbf{z}_{\mathrm{geo}})) + F_{\mathrm{dis}}(\mathbf{x})$ + +Based on the reflectance template $T_{\mathrm{rfl}}$ learned from Stage 1, we calculate the offset value of BRDF through $F_{\mathrm{ofs}}$ , similar to Stage 1, to obtain the specific BRDF attributes for each individual. To address the distinct diversity between diffuse albedo $b_{a}$ , specular $b_{s}$ and roughness $b_{r}$ , we add a learnable weight term $\mathbf{W}(\mathbf{x}) = [\mathbf{w}_{\mathrm{a}}, \mathbf{w}_{\mathrm{s}}, \mathbf{w}_{\mathrm{r}}]$ to control the contribution of each BRDF attribute from the reflectance template $T_{\mathrm{rfl}}$ , where $\mathbf{w}_{\mathrm{a}} \in \mathbb{R}^{3}, \mathbf{w}_{\mathrm{s}} \in \mathbb{R}^{4}, \mathbf{w}_{\mathrm{r}} \in \mathbb{R}$ are the weights for diffuse albedo, specular, and roughness, respectively. Thus, we obtain the BRDF attributes as: + +$$ +\mathcal {B} _ {\mathrm {s t 2}} (\mathbf {x}) = \mathbf {W} (\mathbf {x}) \cdot T _ {\mathrm {r f l}} \left[ F _ {\mathrm {d f m}} (\mathbf {x}, \mathbf {z} _ {\mathrm {g e o}}) \right] + F _ {\mathrm {o f s}} (\mathbf {x}). \tag {6} +$$ + +Considering that the diffuse albedo attribute has a large variation between different people, we initialize $\mathbf{w}_{\mathrm{a}}$ with relatively low values. + +Attribute synergy. In our framework, we decompose facial attributes into two components: geometry and reflectance. These components are not isolated but can enhance each other cooperatively. Sometimes, distinguishing between geometry and reflectance can be challenging, leading to inaccuracies in the decomposition process. For example, a dark region in a facial image might result from sunken areas or pigmentation spots on the face. Conventional neural rendering techniques rely solely on color information from images, making it hard to distinguish such ambiguous scenes, often leading to incorrect reconstruction. Since our method can decouple rough geometry and reflectance attributes in Stage 1, we leverage the prior from the reflectance template to facilitate geometry inference, and vice versa. Specifically, we obtain the displacement $\mathcal{D}(\mathbf{x})$ as follows: + +$$ +\mathcal {D} (\mathbf {x}) = F _ {\text {d i s}} (\mathbf {x}, \operatorname {C o n c a t} \left(\mathbf {f} _ {\mathrm {d e f}}, \mathbf {f} _ {\mathrm {g e o}}, \mathbf {f} _ {\mathrm {r f l}}, \mathcal {G} \left(F _ {\mathrm {d f m}} (\mathbf {x})\right)\right)), \tag {7} +$$ + +where $\mathbf{f}_{\mathrm{dfm}} \in \mathbb{R}^{128}$ , $\mathbf{f}_{\mathrm{geo}} \in \mathbb{R}^{128}$ , and $\mathbf{f}_{\mathrm{rfl}} \in \mathbb{R}^{128}$ are extracted features from the output of the final hidden layer of $F_{\mathrm{dfm}}$ , $T_{\mathrm{geo}}$ , and $T_{\mathrm{rfl}}$ , respectively. $\mathcal{G}(F_{\mathrm{dfm}}(\mathbf{x}))$ is the output from the albedo gradient predictor. We also extract another feature vector $\mathbf{f}_{\mathrm{dis}} \in \mathbb{R}^{64}$ from $F_{\mathrm{dis}}$ to provide more geometry information to the BRDF offset network $F_{\mathrm{ofs}}$ . Similarly, we calculate the BRDF offset $\mathcal{B}_{\mathrm{ofs}}(\mathbf{x})$ as: + +$$ +\mathcal {B} _ {\mathrm {o f s}} (\mathbf {x}) = F _ {\mathrm {o f s}} (\mathbf {x}, \operatorname {C o n c a t} \left(\mathbf {f} _ {\mathrm {d e f}}, \mathbf {f} _ {\mathrm {g e o}}, \mathbf {f} _ {\mathrm {d i s}}, \mathbf {f} _ {\mathrm {r f l}}, \mathcal {G} \left(F _ {\mathrm {d f m}} (\mathbf {x})\right)\right)). \tag {8} +$$ + +Our ablation study validates the effectiveness of this attribute synergy module in the accuracy of both geometry reconstruction and rendering results. + +![](images/0b20076120f680c41ce96bbb0585b4066006fbfe8f3debdd746d76bd763748da.jpg) +Figure 3. Left: BRDF rendering, where light reflects at the incident point. Right: Subsurface scattering on facial skin, showing the offset between the incident and outgoing points. + +Subsurface scattering offset. Human facial skin exhibits a complexity far beyond that of non-biological materials like metal and plastic. Building on insights from previous research [15], rendering realistic faces necessitates accounting for subsurface scattering effects. As illustrated in Fig. 3, this phenomenon can be modeled using a three-layer approach, where specular reflections occur directly at the point of light incidence on the surface, while diffuse reflections involve light penetrating the skin, traversing three layers, undergoing multiple subsurface reflections, and then emerging from the skin surface slightly away from the point of incidence. This effect, known as subsurface scattering, is crucial for realistic skin rendering and can be mathematically represented as $L_{\mathrm{o}}(\mathbf{x}_{\mathrm{o}}, \omega_{\mathrm{o}}) = \int_{A} \int_{\Omega} S(\mathbf{x}_{\mathrm{o}}, \omega_{\mathrm{o}}, \mathbf{x}_{\mathrm{i}}, \omega_{\mathrm{i}}) L_{\mathrm{i}}(\mathbf{x}_{\mathrm{i}}, \omega_{\mathrm{i}})(\boldsymbol{\omega}_{\mathrm{i}} \cdot \mathbf{n}) \, \mathrm{d}\omega_{\mathrm{i}} \, \mathrm{d}A$ , where $\mathbf{x}_{\mathrm{i}}$ and $\mathbf{x}_{\mathrm{o}}$ are the positions of the light incident and output points, respectively, $A$ is a small area around $\mathbf{x}_{\mathrm{o}}$ , and $S$ is the scattering evaluation. In practice, the BRDF mainly accounts for light that scatters back at the point where it hits a surface, but simulating surface scattering accurately is computationally expensive. To improve face reconstruction quality, we integrate subsurface scattering into the BRDF's diffuse component as an offset. We simplify the process by assuming uniform subsurface scattering properties within a small area and utilize scattering profiles to approximate this effect. As illustrated in Fig. 3, these profiles show how the reflectance due to subsurface scattering changes with the distance between where light enters and exits the skin. + +As [28] suggests, one can approximate each RGB channel of the scattering profiles using a combination of six Gaussian functions. We employ the weights of these Gaussian functions to form a feature vector $\mathbf{f}_{\mathrm{sp}}^{\mathbf{x}} \in \mathbb{R}^{18}$ , which represents the scattering profile of the point $\mathbf{x}$ . Consequently, we calculate the offset of the diffuse term as: + +$$ +F _ {\mathrm {s s}} (\mathbf {x}) = v \hat {F} _ {\mathrm {s p}} (\mathbf {x}, \mathbf {f} _ {\mathrm {s p}} ^ {\mathbf {x}}) \hat {L} _ {\mathrm {i}} (\mathbf {x}, \mathcal {E} (\mathbf {x})), \tag {9} +$$ + +where $v$ is the weight of scattering offset, $\hat{F}_{\mathrm{sp}}$ is a network + +to integrate scattering reflectance, $\hat{L}_{\mathrm{i}}$ is a network to integrate the incoming light in a small region around $\mathbf{x}$ , and $\mathcal{E}(\mathbf{x}) \in \mathbb{R}^3$ is a learnable vector that defines the size of the integration domain. The values of $\mathbf{f}_{\mathrm{sp}}^{\mathbf{x}}$ and $\mathcal{E}(\mathbf{x})$ are obtained through a parameter network. We refer readers to the supplementary for more details of the network architecture. + +# 4.3. Training losses + +For both stages, we consider reconstruction losses including the pixel-wise color loss $\mathcal{L}_{\mathrm{col}} = \| \mathbf{C} - \mathbf{C}_{\mathrm{gt}}\| _1$ and the eikonal loss $\mathcal{L}_{\mathrm{eik}} = (\| \nabla S\| _2 - 1)^2$ . Similar to NeuFace [58], we incorporate a regularization term $\mathcal{L}_{\mathrm{light}}$ to assume a nearly white environment light and penalize the specular energy by $\mathcal{L}_{\mathrm{spec}}$ . Additionally, we add $\mathcal{L}_{\mathrm{code}} = \| \mathbf{z}_{\mathrm{geo}}\| _2 + \| \mathbf{z}_{\mathrm{rfl}}\| _2$ to regularize the latent codes. To avoid a large variation between the predict reflectance attributes and the template during training, we apply a constraint to BRDF offset as $\mathcal{L}_{\mathrm{ofs}} = \| \mathcal{B}_{\mathrm{ofs}}(\mathbf{x})\| _1$ . Additionally, the albedo gradient predictor employs $\mathcal{L}_{\mathrm{g}}$ , an L1 loss, which is supervised by the computed albedo gradient. As a result, we formulate the loss function in Stage 1 as follows: + +$$ +\mathcal {L} _ {\mathrm {s t 1}} = \mathcal {L} _ {\mathrm {c o l}} + \mathcal {L} _ {\mathrm {e i k}} + \mathcal {L} _ {\mathrm {l i g h t}} + \mathcal {L} _ {\mathrm {s p e c}} + \mathcal {L} _ {\mathrm {o f s}} + \mathcal {L} _ {\mathrm {c o d e}} + \mathcal {L} _ {\mathrm {g}}. \tag {10} +$$ + +For the second stage, we add additional displacement loss $\mathcal{L}_{\mathrm{dis}}$ and scatter offset loss $\mathcal{L}_{\mathrm{ss}}$ : $\mathcal{L}_{\mathrm{st2}} = \mathcal{L}_{\mathrm{st1}} + \mathcal{L}_{\mathrm{dis}} + \mathcal{L}_{\mathrm{ss}}$ , where $\mathcal{L}_{\mathrm{dis}} = \| \mathcal{D}(\mathbf{x})\|_1$ and $\mathcal{L}_{\mathrm{ss}} = \| F_{\mathrm{ss}}(\mathbf{x})\|_1$ are regularization losses aimed at constraining high values of the displacement and the subsurface scattering offset. + +# 5. Experimental results + +# 5.1. Experimental setup + +Dataset and metrics. We evaluate our method on a large human face dataset Facescape [48]. Specifically, we utilize authorized data from approximately 30 individuals, 4 of whom are from the publishable list. To protect the privacy of other identities that are not on the publishable list, we employed mosaic technology in the eyes region. To validate the generalization performance of our method, we additionally use dataset H3DS [40] in the supplementary material. We assess the quality of image rendering using the PSNR, SSIM, and LPIPS [54] metrics, and the Chamfer distance (CD) for evaluating the accuracy of geometry reconstruction. We compare our method with two 3D reconstruction methods VolSDF [51], DeformHead [47] and three BRDF decomposition methods PhySG [53], TensoIR [24], and NeuFace [58]. + +Implementation details. Facescape has over 50 scans of human heads along with camera parameters and meshes. We preprocess the images to obtain face masks using ImFace [57] tools. For Stage 1 training, we select 10 subjects, each with 10 views. For Stage 2, we use an additional + +20 subjects and choose 3, 5, and 10 views from the available $50+$ views as the training set, with another 10 views reserved for testing. The 10-view experiments specifically evaluate performance in a low-view setting. All test data is independent of the training set and is used consistently across all methods for comparison. We employ the Adam optimizer with an initial learning rate $1 \times 10^{-4}$ to train both stages for 3000 epochs. Additionally, we add a calibration module as [58] to mitigate the exposure differences between various views and apply this calibration to all the methods for a fair comparison. For more implementation details, please refer to the supplementary material. + +# 5.2. Evaluation & comparison + +To corroborate the effectiveness of our proposed framework in the reconstruction of appearance and geometry, we conducted a comprehensive comparison with state-of-the-art methods. These methods are categorized into two groups: geometry-oriented methods, which include VolSDF [51] and DeformHead [47]; and decomposition-based methods, including PhySG [53], TensoIR [24], NeuFace [58], and our method SFDM. As shown in Tab. 1, we enumerate the quantitative performance of different methods in the evaluation metrics. To assess the performance capability under sparse-view images, we conducted tests on 20 subjects with 3, 5, and 10 views, respectively. The visual results of the reconstruction from 3 views are presented in Fig. 4, showcasing the synthesized rendering images from a novel viewpoint, the geometric reconstruction, and the decomposition of diffuse and specular attributes. + +Decomposition-based methods. Our method exhibits superior performance in novel view rendering and geometry reconstruction compared to other reflectance decomposition methods. In contrast to PhySG and TensoIR, designed for common objects, our method significantly outperforms them across all metrics. This is due to the complexity of facial material reflectance, which requires specialized reflectance modeling. NeuFace, which is customized for facial skin, poses a challenge due to its non-robust decomposition of geometry and reflectance attributes. As shown in Fig. 4, it often happens in NeuFace that all reflectance is baked into albedo, but the specular term is entirely absent. This degenerate Lambertian model may reduce the realism of the synthesized rendering output, especially for elements such as teeth, eyes, or the oily layer on facial skin, which often exhibit specular reflection effects. In contrast, we are able to stably decompose the diffuse and specular properties, resulting in more realistic rendering effects. Additionally, we demonstrate higher accuracy in geometric reconstruction, whereas the quality of NeuFace's geometry deteriorates significantly as the number of views decreases. + +Geometry-oriented methods. Our method demonstrates a comparable level of accuracy in geometry reconstruc + +
MethodRelight3 views5 views10 views
PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓
VolSDF [51]✗22.950.8686.9218.2825.450.8976.097.4027.340.9105.185.25
DeformHead [47]✗26.020.8865.769.6026.410.8794.838.4427.350.8915.205.42
PhySG [53]✓22.030.8637.8170.8722.690.8727.0563.2023.110.8767.1255.7
TensoIR [24]✓22.350.6457.1428.6024.610.7045.8617.1026.140.7694.7310.10
NeuFace [58]✓23.540.8765.7815.7425.810.9044.758.6227.190.9144.036.40
Ours✓25.540.9024.469.2726.730.9113.837.0027.640.9203.865.53
+ +![](images/61a0061f77082847219d80bf365a60c6c19520fdcf4f2a7e5fc11bbee735fc0e.jpg) +Figure 4. Visualization of 3-view reconstruction results. VolSDF and DeformHead do not incorporate reflectance attributes, resulting in the absence of corresponding outcomes. + +tion compared to VolSDF and DeformHead. VolSDF exhibits poor 3D reconstruction results in sparse view settings. While DeformHead achieves a more detailed 3D reconstruction utilizing a geometry template, it is overly sensitive to geometry deformation, leading to noticeable indenta + +tions. Although the PSNR of our method is lower than that of DeformHead under 3 views, our LPIPS outperforms it, indicating more photorealistic synthesis results. Failure instances in reconstruction have been observed with VolSDF and DeformHead, with failure rates of $21.7\%$ and $6.7\%$ , respectively. We exclude these failures from the comparison. + +Table 1. Evaluation. The LPIPS and CD metrics are measured on scales of ${10}^{-2}$ and ${10}^{-3}$ ,respectively. The best and second-best results are highlighted in red and orange, respectively. + +
ExpMethodPSNR↑SSIM↑LPIPS↓CD↓
1w/o facial template24.270.8575.5917.22
2w/o albedo gradient25.130.8774.668.83
3w/o attribute synergy24.790.8764.908.80
4w/o scattering offset24.600.8745.738.61
5Ours25.650.8824.457.94
+ +Table 2. Comparative evaluation metrics for ablation studies. + +![](images/067950823dea8c9180dbf1f0ab6bed6a216d70d8b7027c411f2f339ce3ef6151.jpg) +Figure 5. Ablation visualization. Top: Rendering. Bottom: Geometric errors. + +# 5.3. Ablation studies + +We select a subset of 10 subjects from the dataset to test the effectiveness of the major components of SFDM. In Tab. 2, Exp.1-2 are results for the components in Stage 1, and Exp.3-4 cover the components in Stage 2. Exp.1-2 indicate that the facial template significantly influences the robustness of face decomposition, with the albedo gradient predictor further enriching the reconstruction quality by introducing additional facial prior knowledge. From Exp.3, we observe that the attribute synergy module appears to contribute to geometry decomposition, and its absence leads to a noticeable decrease in geometric accuracy. Based on Exp.4, the subsurface scatter offset module can enhance + +the realism of the synthesized images from novel views, which is suggested by improvements in the LPIPS metric. In Fig. 5, columns 3-6 show the visualization results when omitting the facial template $T_{*}$ , albedo gradient predictor $\mathcal{G}$ , attribute synergy module $\mathbf{f}_{*}$ , and subsurface scattering $F_{\mathrm{ss}}$ , respectively. The bottom row highlights the albedo gradient predictor's more significant impact on individuals who exhibit greater deviations from the facial template. In Stage 2, the subsurface scattering offset introduces additional details in light reflection on the skin, resulting in more photorealistic rendering results. Regarding the attribute synergy module, the exchange of information between geometry and reflectance enables a more accurate decomposition, as is evident in the details of geometry reconstruction. + +# 5.4. Reflectance decomposition analysis + +Our method focuses on reflectance decomposition from multi-view facial images without any ground truth reflectance data, enabling realistic reconstruction of facial texture and highlights for novel view rendering. While our goal is not a physically accurate decomposition of skin reflectance, we aim to evaluate our decomposition results both visually and quantitatively. To this end, we use the single-view face reconstruction model, RefMM [20], to fit and decompose images in the test set, generating a coarse diffuse and specular result as pseudo ground truth for evaluation. Tab. 3 shows that our decomposition accuracy generally outperforms that of NeuFace. The overall low SSIM values are primarily due to the coarse nature of the pseudo ground truth, as illustrated in Fig. 6. This reflects certain advantages and limitations of single-view decomposition methods. These methods are typically trained on extensive facial image datasets with specialized lighting conditions or pseudo reflectance ground truth, allowing them to robustly decompose faces from monocular images. However, because they rely heavily on a parameterized facial model and limited input with less 3D information, the reconstructed results lack many facial details, making the rendered images appear less realistic. As shown in Fig. 6 and Fig. 7, we can observe these differences visually in the rendered results. + +
MethodDiffuseSpecular
NeuFace0.7260.552
Ours0.7610.692
+ +Table 3. Quantitative diffuse and specular results (SSIM↑). + +![](images/a36b91b247728a7e3d087dc5ffbf506322a7d0e1b02808ffbf17d381a0bb785a.jpg) +Figure 6. Decomposition results of RefMM [20]. + +# 5.5. Applications + +Specular editing. Our robust decomposition of diffuse and specular reflectances allows us to easily adjust the specular reflectance, resulting in the ability to make faces appear shinier or dimmer. The specular editing outcomes are showcased in Fig. 7. Specifically, we increase the specular weight by 1.5 times, while maintaining the overall specular distribution unchanged. As a result, the edited images can retain their photorealistic appearance. + +![](images/3effd274954fa97bb2e5d16afce37fceb07ec51d9b5c43d41813f34dc7122fcc.jpg) +Figure 7. Applications. (a) Specular editing by increasing the shininess. (b) Relighting under different environments. + +Relighting. Decomposing reflectance and environment light facilitates the relighting of human faces without altering their inherent properties. As shown in Fig. 7, we conducted a comparison of the relighting effects between our method and NeuFace in sparse view settings. Our method excels in decomposing specular reflectance, leading to more vibrant and realistic relighting outcomes under varying environmental lighting conditions [12]. + +# 6. Conclusion + +We propose a novel face decomposition method capable of rendering and reconstructing 3D human faces from sparse-view images. Our framework first learns a facial template with both decomposed geometry and reflectance attributes from multi-view images of different individuals. Then we utilize the prior knowledge of the facial template and further consider the interaction between geometry and reflectance, as well as subsurface scattering effects, which enable our method to achieve robust face decomposition from sparse-view images. Based on decomposed facial attributes, our method achieves superior results in the rendering and geometric reconstruction of 3D human faces and can be utilized for relighting and specular editing. + +Limitations & future work. As we focus on face decomposition, it is hard to extend to general objects directly with our method. In the future, we could involve techniques such as hash encoding [33] to accelerate our method and leverage fine-detailed 3D prior or reflectance ground truth for a better decomposition effect. + +# Acknowledgment + +This work was supported in part by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22) and the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). J.Hu was supported by the National Natural Science Foundation of China under Grants 62402083. + +# References + +[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1 +[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 3 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1 +[4] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hašan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2 +[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2 +[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164, 2023. 1, 2 +[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 1 +[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2 +[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2, 3 + +[10] Chengqian Che, Fujun Luan, Shuang Zhao, Kavita Bala, and Ioannis Gkioulekas. Towards learning-based inverse subsurface scattering. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2 +[11] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 1 +[12] Zhang Chen, Anpei Chen, Guli Zhang, Chengyuan Wang, Yu Ji, Kiriakos N Kutulakos, and Jingyi Yu. A neural rendering framework for free-viewpoint relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5610, 2020. 3, 8 +[13] Ziang Cheng, Hongdong Li, Yuta Asano, Yinqiang Zheng, and Imari Sato. Multi-view 3d reconstruction of a textureless smooth surface of unknown generic reflectance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16226-16235, 2021. 1 +[14] Abdallah Dib, Luiz Gustavo Hafemann, Emeline Got, Trevor Anderson, Amin Fadainejad, Rafael MO Cruz, and Marc-Andre Carbonneau. Mosar: Monocular semi-supervised model for avatar reconstruction using differentiable shading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1770–1780, 2024. 2 +[15] Craig Donner and Henrik Wann Jensen. Light diffusion in multi-layered translucent materials. ACM Transactions on Graphics (ToG), 24(3):1032-1039, 2005. 5 +[16] Haiwen Feng, Timo Bolkart, Joachim Tesch, Michael J Black, and Victoria Abrevaya. Towards racially unbiased skin tone estimation via scene disambiguation. In European Conference on Computer Vision, pages 72-90. Springer, 2022. 2 +[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 2 +[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1 +[19] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 1 +[20] Yuxuan Han, Zhibo Wang, and Feng Xu. Learning a 3d morphable face reflectance model from low-cost data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8598-8608, 2023. 2, 8 +[21] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 1 +[22] Mingming He, Pascal Clausen, Ahmet Levent Taşel, Li Ma, Oliver Pilarski, Wenqi Xian, Laszlo Rikker, Xueming Yu, Ryan Burgert, Ning Yu, et al. Diffrelight: Diffusion-based + +facial performance relighting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 1 +[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 1 +[24] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 6, 7 +[25] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3 +[26] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3 +[27] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8629-8640, 2023. 2 +[28] Andrew Lauritzen. Gpu gems 3, 2007. 5 +[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 1 +[30] Connor Lin, Koki Nagano, Jan Kautz, Eric Chan, Umar Iqbal, Leonidas Guibas, Gordon Wetzstein, and Sameh Khamis. Single-shot implicit morphable faces with consistent texture parameterization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1 +[31] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Trans. Graph., 42(4): 114:1-114:22, 2023. 1, 2 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 8 +[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 1, 2 + +[35] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2 +[36] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2 +[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 1 +[38] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2 +[39] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2 +[40] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 6 +[41] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 2 +[42] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1 +[43] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1 +[44] Borom Tunwattanapong, Graham Fyffe, Paul Graham, Jay Busch, Xueming Yu, Abhijeet Ghosh, and Paul Debevec. Acquiring reflectance and shape from continuous spherical harmonic illumination. ACM Transactions on graphics (TOG), 32(4):1-12, 2013. 3 +[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3 +[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and re + +lighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 3 +[47] Baixin Xu, Jiarui Zhang, Kwan-Yee Lin, Chen Qian, and Ying He. Deformable model-driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17924-17934, 2023. 1, 3, 6, 7 +[48] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 6 +[49] Jing Yang, Hanyuan Xiao, Wenbin Teng, Yunxuan Cai, and Yajie Zhao. Light sampling field and BRDF representation for physically-based neural rendering. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2 +[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2 +[51] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 6, 7 +[52] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 3 +[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 1, 2, 6, 7 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[55] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1 +[56] Wenyi Zhao, Rama Chellappa, P Jonathon Phillips, and Azriel Rosenfeld. Face recognition: A literature survey. ACM computing surveys (CSUR), 35(4):399-458, 2003. 2 +[57] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit + +neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 3, 6 +[58] Mingwu Zheng, Haiyu Zhang, Hongyu Yang, and Di Huang. Neuface: Realistic 3d neural face rendering from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16868-16877, 2023. 1, 2, 3, 4, 6, 7 +[59] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 3 +[60] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 1 +[61] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum, pages 523-550. Wiley Online Library, 2018. 2 \ No newline at end of file diff --git a/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/images.zip b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e384e3485170d2c79cb3d3a0143314d00f3a96b --- /dev/null +++ b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af987bb53643532416e887694d42a1039b443cbfd8d7db321477b4545c1a3bed +size 465797 diff --git a/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/layout.json b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0a36dffcde0c7a16be0a459f313af02e84da9012 --- /dev/null +++ b/2025/SFDM_ Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images/layout.json @@ -0,0 +1,8990 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 545, + 140 + ], + "type": "text", + "content": "SFDM: Robust Decomposition of Geometry and Reflectance for Realistic Face Rendering from Sparse-view Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": "Daisheng Jin" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": " Jiangbei Hu" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": " Baixin Xu" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": " Yuxin Dai" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": " Chen Qian" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": " Ying He" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University, " + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": "Dalian University of Technology, " + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 548, + 191 + ], + "type": "text", + "content": "SenseTime Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 242, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 242, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 242, + 296, + 531 + ], + "type": "text", + "content": "In this study, we introduce a novel two-stage technique for decomposing and reconstructing facial features from sparse-view images, a task made challenging by the unique geometry and complex skin reflectance of each individual. To synthesize 3D facial models more realistically, we endeavor to decouple key facial attributes from the RGB color, including geometry, diffuse reflectance, and specular reflectance. Specifically, we design a Sparse-view Face Decomposition Model (SFDM): 1) In the first stage, we create a general facial template from a wide array of individual faces, encapsulating essential geometric and reflectance characteristics. 2) Guided by this template, we refine a specific facial model for each individual in the second stage, considering the interaction between geometry and reflectance, as well as the effects of subsurface scattering on the skin. With these advances, our method can reconstruct high-quality facial representations from as few as three images. The comprehensive evaluation and comparison reveal that our approach outperforms existing methods by effectively disentangling geometric and reflectance components, significantly enhancing the quality of synthesized novel views, and paving the way for applications in facial relighting and reflectance editing. Visit our project page for more details https://kingjg.github.io/SFDM.github.io/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 543, + 135, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 135, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 135, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 695 + ], + "type": "text", + "content": "3D human face reconstruction finds widespread applications in various fields, including game design and film production. Developments in neural rendering techniques have shown promising performance and advantages in synthesizing photo-realistic images for complex scenes [3, 32, 33, 43]. Concurrently, a surge of research focusing on head and facial attribute reconstruction from images via implicit neural representations [18, 36, 37, 59] has emerged. These innovative approaches address the limitations associated with traditional explicit methods, such as those based on the 3D Morphable Model (3DMM) [6], which struggled with is" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 219, + 554, + 346 + ], + "blocks": [ + { + "bbox": [ + 319, + 219, + 554, + 346 + ], + "lines": [ + { + "bbox": [ + 319, + 219, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 319, + 219, + 554, + 346 + ], + "type": "image", + "image_path": "151c96cf6e0f022bf42d81202e47e1e251ebe55d38faa536ed2a990aa5ccef4b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 357, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Our method learns a general facial template with distinct key facial attributes from multiple individuals. This template enables photorealistic novel-view synthesis for each individual, as well as other applications." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 472 + ], + "type": "text", + "content": "sues like dependency on mesh resolution and difficulties in managing topological variations. Furthermore, implicit neural representation methods demonstrate their effectiveness in 3D face reconstruction from low-view images [47], expression transfer [1, 23, 30], and animations [19, 59, 60]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 715 + ], + "type": "text", + "content": "In 3D face model rendering, both accurate shape attributes and high-quality reflectance properties are essential for synthesizing photorealistic faces. Thus, robustly decomposing key facial attributes including geometry, diffuse and specular reflectance, and environment light is a crucial task, which not only enhances the accuracy of face reconstruction but also opens up possibilities for more applications like relighting [7, 11, 22, 29, 55]. However, most previous studies only use neural networks to learn radiance and geometric information from images, neglecting the effect of rendering-related physical information such as facial skin characteristics and environmental lighting conditions. NeRV [42], PhySG [53], and following methods [8, 9, 13, 21, 31, 34] incorporate the Bidirectional Reflectance Distribution Functions (BRDFs) into the neural rendering framework, facilitating the decomposition and joint reconstruction of shape, reflectance, and illumination. Nonetheless, these methods primarily concentrate on non-biological materials, which are not suitable for decomposing face models that possess complex attributes. NeuFace [58] introduces an approxi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 249, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 249, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 249, + 713 + ], + "type": "text", + "content": "*Corresponding author: Y. He (Email: yhe@ntu.edu.sg)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26409" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "mated BRDF integration and learns the physically meaningful underlying representations for 3D face reconstruction. Nevertheless, due to the ill-posedness of decomposition, NeuFace often converges to a suboptimal solution under sparse-view settings, in which specular reflectance is absent, and all radiance is derived from diffuse attributes. In addition, the accuracy of geometric reconstruction significantly diminishes as the number of views decreases. Therefore, it still poses a significant challenge to decompose shape, reflectance, and environment light attributes from sparse view images in 3D face model reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 361 + ], + "type": "text", + "content": "In this study, we introduce a novel two-stage face decomposition framework, termed Sparse-view Face Decomposition Model (SFDM), aimed at reconstructing 3D faces from sparse-view images. The pipeline of our proposed algorithm is depicted in Fig. 2. Based on the observation that human faces share similar geometric features and reflectance properties, we first derive a novel general facial template from images of a set of varied individual faces in Stage 1. In Stage 2, we employ the learned facial template with both geometry and reflectance attributes to refine the decomposition of each individual face model. In both stages, we decompose the shape and material attributes using a geometry module and a reflectance module, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 687 + ], + "type": "text", + "content": "Specifically, the geometry module consists of a template network and a deformation network. In the first stage, we can get a template model with common geometry features based on the training of multi-view images from varied faces, which can be further transformed to a specific individual face model by deformation. Building on the facial geometry template, we innovatively propose a reflectance module that integrates reflectance information into the facial template. Specifically, we introduce the BRDF as a coarse representation of reflectance for facial skin. We employ a BRDF template network to learn the shared characteristics of diffuse and specular reflectance properties, and a BRDF offset network is utilized to reconstruct identity-specific reflectance details. Furthermore, we propose a novel albedo gradient predictor designed to capture more general facial reflectance attributes. In the second stage, based on the learned template with geometry and reflectance attributes, we refine the attribute details from sparse-view images of each individual face model. To this end, we incorporate a displacement network and a reflectance offset network on the learned template, employing a design that harmonizes geometry and reflectance information to capture high-frequency details of facial attributes. In addition, we introduce a diffuse offset network to take into account the effect of subsurface scattering, thus providing a more realistic simulation of the facial skin. We summarize the main contributions of this study as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose a novel decomposition framework for human facial representations to learn a template with various at" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "tributes, including geometric features, diffuse reflectance, and specular reflectance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 96, + 553, + 192 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "type": "text", + "content": "- Based on the learned template, our tailored refinement stage enables robust decomposition of attributes for each individual face under as few as three images, attaining superior geometric accuracy and reflectance detail." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the face decomposition results obtained through our framework can facilitate more realistic synthesis in novel views, as well as applications such as specular editing and relighting." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 203, + 398, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 203, + 398, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 203, + 398, + 216 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 223, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 510 + ], + "type": "text", + "content": "Physically based rendering. To achieve a photorealistic rendering effect, physically based rendering (PBR) [39] is proposed to simulate the light and surfaces using real-world optical principles. Reversely, physically based inverse rendering seeks a way to decompose light and surface factors from the images [4, 5, 8, 34]. IDR [50] proposes a surface reconstruction method that simultaneously disentangles the geometry and appearance. Subsequent works incorporate inverse rendering into the neural rendering pipeline, developing various methods for general objects [9, 53], and enhancing decomposition effects for objects with specialized materials [10, 31]. NeuFace [58] proposes a decomposition pipeline for human faces, taking into account the sophisticated properties of facial skin reflectance. However, without ground truth for reflectance components, face decomposition is an ill-posed problem, causing NeuFace to often struggle with accurately decomposing reflectance. Additionally, NeuFace does not consider the effect of subsurface scattering on facial skin, resulting in less realistic PBR [49]. In this work, with our novel facial template and the single-person refinement module, which further considers subsurface scattering effects, we can robustly decompose geometry and reflectance in a more realistic manner from sparse-view images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Face reconstruction. Given the inherent similarities among human faces, it is intuitive to develop a template that serves as a prior to enhance facial reconstruction. Based on a large amount of facial data, 3DMM [6] builds a 3D morphable face to describe the transformation of shape and expression, which is widely used in face recognition [56], 3D reconstruction, face editing [17, 36], etc. Using 3D morphable faces also enables face reconstruction from monocular images [14, 16, 27, 35, 61]. For instance, Han et al. [20] build a 3D morphable face with reflectance on BFM09 [38] for material editing. However, in face reconstruction, the use of 3D morphable faces often leads to a loss of detail during the compression and fitting process, resulting in overly coarse reconstructions. Moreover, given limited view information, monocular methods typically require reflectance ground truth for training [41]. As implicit neural representation shows extraordinary efficacy for 3D representation," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26410" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "numerous studies [47, 52, 57, 59] construct 3D parametric face or head models based on neural networks, yielding improved geometric details for faces or heads. Besides, some facial capturing methods utilize more informative inputs such as videos [46] and polarized images [2], enabling high-quality reconstruction across a wider range of scenarios. Our goal is to achieve face reconstruction solely from sparse-view images, without any reflectance ground truth. Our Stage 1 employs an implicit representation to construct a facial template, incorporating prior knowledge of facial geometry and reflectance. This approach enables robust decomposition and realistic face reconstruction during the single-person refinement process under sparse views." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 133, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 133, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 133, + 251 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": "The effectiveness of PBR has been validated by numerous studies [25, 45, 58] for photo-realistic rendering in scenarios with complex reflection properties. Inspired by these works and as illustrated in Fig. 3, we formulate the output radiance " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{o}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " at surface position " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " along output direction " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{o}}(\\mathbf{x}, \\omega_{\\mathrm{o}}) = \\int_{\\Omega} f(\\mathbf{x}, \\omega_{\\mathrm{i}}, \\omega_{\\mathrm{o}}) L_{\\mathrm{i}}(\\mathbf{x}, \\omega_{\\mathrm{i}})(\\omega_{\\mathrm{i}} \\cdot \\mathbf{n}) \\, \\mathrm{d}\\omega_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " represents the BRDF evaluation, " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{i}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " is the incident light intensity from input direction " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{i}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{n} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " is the surface normal. To decouple the view-dependent and view-independent radiance, we decompose " + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 257, + 296, + 402 + ], + "type": "text", + "content": " into the diffuse term and the specular term by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 407, + 286, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 407, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 66, + 407, + 286, + 490 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {0} (\\mathbf {x}, \\boldsymbol {\\omega} _ {0}) = \\underbrace {\\frac {b _ {\\mathrm {a}} (\\mathbf {x})}{\\pi} \\int_ {\\Omega} L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}}) (\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}}} _ {\\text {d i f f u s e}} \\\\ + \\underbrace {\\int_ {\\Omega} f _ {s} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}} , \\boldsymbol {\\omega} _ {\\mathrm {o}} ; b _ {\\mathrm {s}} (\\mathbf {x}) , b _ {\\mathrm {r}} (\\mathbf {x})) L _ {\\mathrm {i}} (\\mathbf {x} , \\boldsymbol {\\omega} _ {\\mathrm {i}}) (\\boldsymbol {\\omega} _ {\\mathrm {i}} \\cdot \\mathbf {n}) \\mathrm {d} \\boldsymbol {\\omega} _ {\\mathrm {i}} ,} _ {\\text {s p e c u l a r}} \\\\ \\end{array}", + "image_path": "b65911f1fe29fab831f25c329e06f6e684e9be9dab7549798394ead8b9d4d9a4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{a}}(\\mathbf{x}) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{s}}(\\mathbf{x}) \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "inline_equation", + "content": "b_{\\mathrm{r}}(\\mathbf{x}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "text", + "content": " are diffuse albedo, specular, and roughness fields obtained from BRDF, respectively. " + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 55, + 495, + 296, + 569 + ], + "type": "text", + "content": " represents the BRDF evaluation that only accounts for the specular property. Furthermore, we improve the efficiency of integration computation by using the approximated pre-computed method [9] as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 310, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 310, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 310, + 622 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {o}} (\\mathbf {x}, \\boldsymbol {\\omega} _ {\\mathrm {o}}) \\approx \\underbrace {\\frac {b _ {\\mathrm {a}} (\\mathbf {x})}{\\pi} \\hat {L} _ {\\mathrm {i}} (\\mathbf {n} , 1)} _ {\\text {d i f f u s e}} + \\underbrace {\\hat {f} _ {\\mathrm {s}} (\\mathbf {x} ; b _ {\\mathrm {s}} (\\mathbf {x}) , b _ {\\mathrm {r}} (\\mathbf {x})) \\hat {L} _ {\\mathrm {i}} \\left(\\boldsymbol {\\omega} _ {\\mathrm {r}} , b _ {\\mathrm {r}} (\\mathbf {x})\\right)} _ {\\text {s p e c u l a r}}, \\tag {1}", + "image_path": "cfc4b41e5492fe1f32172447f522b13c4bbe6ea86e16a2095b10a35a2ff399fa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{r}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "content": " represent the mirrored view direction, " + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{L}_i" + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "content": " is the integration of input light, " + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{f}_s" + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "content": " is the integration of specular reflectance. Specifically, following [26], we employ an MLP to simulate two look-up-textures (LUT) " + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "inline_equation", + "content": "B_0" + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "inline_equation", + "content": "B_1" + }, + { + "bbox": [ + 55, + 622, + 296, + 681 + ], + "type": "text", + "content": " to calculate the BRDF integration as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 687, + 303, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 303, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 303, + 714 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {\\mathrm {s}} (\\mathbf {x}) = b _ {\\mathrm {s}} (\\mathbf {x}) \\left[ F _ {0} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}, \\mathbf {n}\\right) B _ {0} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}} \\cdot \\mathbf {n}, b _ {\\mathrm {r}} (\\mathbf {x})\\right) + B _ {1} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}} \\cdot \\mathbf {n}, b _ {\\mathrm {r}} (\\mathbf {x})\\right) \\right], \\tag {2}", + "image_path": "014976597227dcb39eda09df63f2fa1da6c32705b1124011b0adc4adc21bb951.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " is the Fresnel term at normal incidence as [9]. For light representation, we use spherical harmonics [44] to approximate environment light by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 118, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 118, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 367, + 118, + 555, + 152 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {i}} \\left(\\omega_ {\\mathrm {i}}\\right) \\approx \\sum_ {\\ell = 0} ^ {k} \\sum_ {m = - \\ell} ^ {\\ell} c _ {\\ell} ^ {m} Y _ {\\ell} ^ {m} \\left(\\omega_ {\\mathrm {i}}\\right), \\tag {3}", + "image_path": "58fc3f8bb40445650df32a5f8fe23245dd18d04ff2a21ebf872dd9a12de5e722.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": " represents spherical basis functions and " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": " denotes corresponding coefficients. Considering that the facial material is not of high specular reflectance, we set the order as " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": " [12]. To integrate incident light rays, the light for the specular term can be integrated to the mirrored direction of the observation view, denoted as " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{i}}(\\omega_{\\mathrm{r}},b_{\\mathrm{r}})" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": ", and light for the diffuse term (Lambertian) can be approximated by " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{i}}(\\mathbf{n},1)" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": " [9], which assumes light in the normal direction hits a completely rough surface. For our SH light representation, we integrate " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{i}}(\\omega_{\\mathrm{r}},b_{\\mathrm{r}})" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": " for the specular term and " + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{i}}(\\mathbf{n},1)" + }, + { + "bbox": [ + 313, + 162, + 555, + 307 + ], + "type": "text", + "content": " for the diffuse term through an approximation using the first ten orders of spherical harmonics as [58]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 318, + 370, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 318, + 370, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 318, + 370, + 331 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 483 + ], + "type": "text", + "content": "Typical human faces share common features, such as two eyes and one nose, with variations in characteristics such as shape, skin color, and skin texture. To better utilize the shared features of human faces, we divide the decomposition process into two stages to learn the neural fields of human faces from multi-view images. Fig. 2 presents the architecture of the Sparse-view Face Decomposition Model (SFDM). In Stage 1, we create a general facial template with geometry and reflectance attributes from a collection of portraits of multiple individuals. In Stage 2, we refine the specific attributes for each individual face model based on the learned facial template under sparse-view settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 493, + 462, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 462, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 462, + 506 + ], + "type": "text", + "content": "4.1. Generating facial template" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "type": "text", + "content": "Given a collection of portraits of multiple individuals " + }, + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_i\\}" + }, + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "type": "text", + "content": ", we first learn a general facial template with both geometry and reflectance attributes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": "Geometry template. Following the previous work [47, 57], we jointly employ a geometry template network " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{geo}}" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": " and a deformation network " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dfm}}" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": " to reconstruct the geometry (represented by signed distance field " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": ") from given images. As illustrated in Fig. 2, we formulate the geometry network as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 621, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 621, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 376, + 621, + 555, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (\\mathbf {x}) = T _ {\\text {g e o}} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\text {g e o}})\\right), \\tag {4}", + "image_path": "09f8803b266c2ef658da2c4d8b1c85540275034fb4ed143b7038b13e9744628d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " is a query point, " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " returns the signed distance value. We take " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{geo}} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " as an identity-related latent code that encodes geometry, so we can get the unique SDF for each individual. Similar to VolSDF [51], we convert the SDF value to density " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and then feed it into the volume rendering block." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26411" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 242 + ], + "type": "image", + "image_path": "607b10f409ce9cfb1ecc289feb583462d367727e0e2a1b8d20ba61cf2c25920a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 291 + ], + "type": "text", + "content": "Figure 2. Pipeline of SFDM. We propose a geometry and reflectance decomposition framework for 3D face reconstruction from multiview images. We first learn a general facial template from multi-view images of multiple individuals in Stage 1. Then the learned facial template guides the further attribute refinement for each individual face model in Stage 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 294, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 294, + 372 + ], + "type": "text", + "content": "Reflectance template. Apart from geometry attributes, we also consider reflectance attributes, which are equally crucial and share similar features across human faces. Therefore, we aim to create a coarse but general template that captures the underlying common reflectance characteristics of human faces in Stage 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "content": "To obtain the reflectance attributes, we approximate the Bidirectional Reflectance Distribution Function (BRDF) by the network that consists of a reflectance template " + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rfl}}" + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "content": " and a BRDF offset network " + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ofs}}" + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "content": ". Similar to the geometry template, we leverage " + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rfl}}" + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "content": " to fit the general facial BRDF for multiple individuals and then offset the BRDF value of the query points on each face model through " + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ofs}}" + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "content": ". We formulate the calculation of BRDF " + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{st1}}(\\mathbf{x})" + }, + { + "bbox": [ + 54, + 372, + 295, + 468 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 476, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 476, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 83, + 476, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {\\mathrm {s t l}} (\\mathbf {x}) = T _ {\\mathrm {r f l}} \\left[ F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {g e o}}) \\right] + w F _ {\\mathrm {o f s}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {r f l}}), \\tag {5}", + "image_path": "30c330b91df4b3f4f950eebf2b80b12f22dde813db3518d8ca43981122890893.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": "where the query points " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": " in the observation space are deformed to the surface of the facial template through " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dfm}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rfl}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": " return the corresponding BRDF attributes " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathbf{x}) = [b_{\\mathrm{a}}(\\mathbf{x}), b_{\\mathrm{s}}(\\mathbf{x}), b_{\\mathrm{r}}(\\mathbf{x})] \\in \\mathbb{R}^8" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ". Subsequently, the specific reflectance attributes for each individual are determined by learning the BRDF offsets with " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ofs}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ". To capture the unique reflectance characteristics associated with individual identities, we incorporate a reflectance latent code " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{rfl}} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ". Moreover, we adjust the offset using a weighting parameter " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ", setting it to a low value of " + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "inline_equation", + "content": "w = 0.05" + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": ". This choice reflects our preference for the template network to primarily capture reflectance properties, minimizing dependence on the BRDF offset network. To simulate environment light, we apply the spherical harmonics as in Eq. (3) and integrate the lights as in [58]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "After obtaining the geometry, reflectance, and environment light information, we calculate the final color (output radiance) as Eq. (1-3) and integrate it by volume rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 336 + ], + "type": "text", + "content": "As shown in Fig. 1, a general facial template can be obtained and robustly decomposed into geometry, diffuse, and specular attributes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": "Albedo gradient predictor. Considering the wide range of skin color differences among various ethnic groups, simply relying on the absolute values of albedo on the facial template does not adequately convey the prior knowledge of faces. We observe that the continuity of facial albedo variations across 3D positions exhibits certain regularities, such as drastic changes around the edges of the lips and the eyes. Thus, we introduce the concept of albedo gradient as an additional attribute for the facial template. Specifically, during the training phase of Stage 1, we calculate the gradient of the albedo values " + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{b}_{\\mathrm{a}}(\\mathbf{x}) \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": " from the fitted neural field with respect to changes in 3D coordinates " + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": ". This gradient plot effectively outlines facial contours like facial sketches. Following this, we use it to supervise the training of an additional albedo gradient predictor branch " + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": " within the reflectance template, guiding the model to pay attention to the patterns of albedo changes. During training, once the point " + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dfm}}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": " in the template space is obtained, it serves as the input for " + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": " and the output target is " + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{b}_{\\mathrm{a}}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 338, + 555, + 613 + ], + "type": "text", + "content": " in the observation space. In the end, the well-trained albedo gradient predictor provides more comprehensive information and precise guidance for single-person refinement in Stage 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 622, + 548, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 548, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 548, + 635 + ], + "type": "text", + "content": "4.2. Individual decomposition under sparse views" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 640, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 554, + 676 + ], + "type": "text", + "content": "Taking the facial template learned from Stage 1 as prior, we perform a fine-detailed decomposition from sparse-view images of a single face in Stage 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Geometry displacement and BRDF offset. The geometry template " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{geo}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " only represents a rough shape with shared features of multiple human faces. Additionally, the defor" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26412" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "mation net " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dfm}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " transforms the points on the surface of the facial template to the surface of a specific individual's face. However, the transformation is constrained, making it challenging to create topological features beyond the template domain and capture high-frequency details. Consequently, we incorporate a displacement network " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dis}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " in Stage 2 to modify the SDF obtained from the geometry template for a more accurate reconstruction. Building on Eq. (4), we incorporate the offset term as follows: " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(\\mathbf{x}) = T_{\\mathrm{geo}}(F_{\\mathrm{dfm}}(\\mathbf{x},\\mathbf{z}_{\\mathrm{geo}})) + F_{\\mathrm{dis}}(\\mathbf{x})" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": "Based on the reflectance template " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rfl}}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": " learned from Stage 1, we calculate the offset value of BRDF through " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ofs}}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": ", similar to Stage 1, to obtain the specific BRDF attributes for each individual. To address the distinct diversity between diffuse albedo " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "b_{a}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": ", specular " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "b_{s}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": " and roughness " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "b_{r}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": ", we add a learnable weight term " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{W}(\\mathbf{x}) = [\\mathbf{w}_{\\mathrm{a}}, \\mathbf{w}_{\\mathrm{s}}, \\mathbf{w}_{\\mathrm{r}}]" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": " to control the contribution of each BRDF attribute from the reflectance template " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rfl}}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{a}} \\in \\mathbb{R}^{3}, \\mathbf{w}_{\\mathrm{s}} \\in \\mathbb{R}^{4}, \\mathbf{w}_{\\mathrm{r}} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 193, + 296, + 312 + ], + "type": "text", + "content": " are the weights for diffuse albedo, specular, and roughness, respectively. Thus, we obtain the BRDF attributes as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 323, + 294, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 323, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 78, + 323, + 294, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {\\mathrm {s t 2}} (\\mathbf {x}) = \\mathbf {W} (\\mathbf {x}) \\cdot T _ {\\mathrm {r f l}} \\left[ F _ {\\mathrm {d f m}} (\\mathbf {x}, \\mathbf {z} _ {\\mathrm {g e o}}) \\right] + F _ {\\mathrm {o f s}} (\\mathbf {x}). \\tag {6}", + "image_path": "308360e44fe2e51823f313302c32c660a34d39bb935400ccbe8a476236fb8355.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 380 + ], + "type": "text", + "content": "Considering that the diffuse albedo attribute has a large variation between different people, we initialize " + }, + { + "bbox": [ + 55, + 346, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 380 + ], + "type": "text", + "content": " with relatively low values." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 296, + 561 + ], + "type": "text", + "content": "Attribute synergy. In our framework, we decompose facial attributes into two components: geometry and reflectance. These components are not isolated but can enhance each other cooperatively. Sometimes, distinguishing between geometry and reflectance can be challenging, leading to inaccuracies in the decomposition process. For example, a dark region in a facial image might result from sunken areas or pigmentation spots on the face. Conventional neural rendering techniques rely solely on color information from images, making it hard to distinguish such ambiguous scenes, often leading to incorrect reconstruction. Since our method can decouple rough geometry and reflectance attributes in Stage 1, we leverage the prior from the reflectance template to facilitate geometry inference, and vice versa. Specifically, we obtain the displacement " + }, + { + "bbox": [ + 55, + 382, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 382, + 296, + 561 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 571, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 571, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 66, + 571, + 295, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {D} (\\mathbf {x}) = F _ {\\text {d i s}} (\\mathbf {x}, \\operatorname {C o n c a t} \\left(\\mathbf {f} _ {\\mathrm {d e f}}, \\mathbf {f} _ {\\mathrm {g e o}}, \\mathbf {f} _ {\\mathrm {r f l}}, \\mathcal {G} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x})\\right)\\right)), \\tag {7}", + "image_path": "a15da6ce7c72a9d7f964259886cd8902fcdd3f92bbc8a63e40f72722421ddae1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{dfm}} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{geo}} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{rfl}} \\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": " are extracted features from the output of the final hidden layer of " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dfm}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{geo}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rfl}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(F_{\\mathrm{dfm}}(\\mathbf{x}))" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": " is the output from the albedo gradient predictor. We also extract another feature vector " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{dis}} \\in \\mathbb{R}^{64}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{dis}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": " to provide more geometry information to the BRDF offset network " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ofs}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": ". Similarly, we calculate the BRDF offset " + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{ofs}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 594, + 296, + 679 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 689, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 689, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {\\mathrm {o f s}} (\\mathbf {x}) = F _ {\\mathrm {o f s}} (\\mathbf {x}, \\operatorname {C o n c a t} \\left(\\mathbf {f} _ {\\mathrm {d e f}}, \\mathbf {f} _ {\\mathrm {g e o}}, \\mathbf {f} _ {\\mathrm {d i s}}, \\mathbf {f} _ {\\mathrm {r f l}}, \\mathcal {G} \\left(F _ {\\mathrm {d f m}} (\\mathbf {x})\\right)\\right)). \\tag {8}", + "image_path": "73494b0f83d25b40dffc831b15b1d62b9f36108a4a5d2fa0442cc4aa1f3186ad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "Our ablation study validates the effectiveness of this attribute synergy module in the accuracy of both geometry reconstruction and rendering results." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 118, + 533, + 206 + ], + "blocks": [ + { + "bbox": [ + 326, + 118, + 533, + 206 + ], + "lines": [ + { + "bbox": [ + 326, + 118, + 533, + 206 + ], + "spans": [ + { + "bbox": [ + 326, + 118, + 533, + 206 + ], + "type": "image", + "image_path": "0b20076120f680c41ce96bbb0585b4066006fbfe8f3debdd746d76bd763748da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 554, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 247 + ], + "type": "text", + "content": "Figure 3. Left: BRDF rendering, where light reflects at the incident point. Right: Subsurface scattering on facial skin, showing the offset between the incident and outgoing points." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": "Subsurface scattering offset. Human facial skin exhibits a complexity far beyond that of non-biological materials like metal and plastic. Building on insights from previous research [15], rendering realistic faces necessitates accounting for subsurface scattering effects. As illustrated in Fig. 3, this phenomenon can be modeled using a three-layer approach, where specular reflections occur directly at the point of light incidence on the surface, while diffuse reflections involve light penetrating the skin, traversing three layers, undergoing multiple subsurface reflections, and then emerging from the skin surface slightly away from the point of incidence. This effect, known as subsurface scattering, is crucial for realistic skin rendering and can be mathematically represented as " + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{o}}(\\mathbf{x}_{\\mathrm{o}}, \\omega_{\\mathrm{o}}) = \\int_{A} \\int_{\\Omega} S(\\mathbf{x}_{\\mathrm{o}}, \\omega_{\\mathrm{o}}, \\mathbf{x}_{\\mathrm{i}}, \\omega_{\\mathrm{i}}) L_{\\mathrm{i}}(\\mathbf{x}_{\\mathrm{i}}, \\omega_{\\mathrm{i}})(\\boldsymbol{\\omega}_{\\mathrm{i}} \\cdot \\mathbf{n}) \\, \\mathrm{d}\\omega_{\\mathrm{i}} \\, \\mathrm{d}A" + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{i}}" + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": " are the positions of the light incident and output points, respectively, " + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": " is a small area around " + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 260, + 555, + 594 + ], + "type": "text", + "content": " is the scattering evaluation. In practice, the BRDF mainly accounts for light that scatters back at the point where it hits a surface, but simulating surface scattering accurately is computationally expensive. To improve face reconstruction quality, we integrate subsurface scattering into the BRDF's diffuse component as an offset. We simplify the process by assuming uniform subsurface scattering properties within a small area and utilize scattering profiles to approximate this effect. As illustrated in Fig. 3, these profiles show how the reflectance due to subsurface scattering changes with the distance between where light enters and exits the skin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "type": "text", + "content": "As [28] suggests, one can approximate each RGB channel of the scattering profiles using a combination of six Gaussian functions. We employ the weights of these Gaussian functions to form a feature vector " + }, + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{sp}}^{\\mathbf{x}} \\in \\mathbb{R}^{18}" + }, + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "type": "text", + "content": ", which represents the scattering profile of the point " + }, + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 595, + 554, + 666 + ], + "type": "text", + "content": ". Consequently, we calculate the offset of the diffuse term as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 675, + 553, + 691 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 675, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 364, + 675, + 553, + 691 + ], + "type": "interline_equation", + "content": "F _ {\\mathrm {s s}} (\\mathbf {x}) = v \\hat {F} _ {\\mathrm {s p}} (\\mathbf {x}, \\mathbf {f} _ {\\mathrm {s p}} ^ {\\mathbf {x}}) \\hat {L} _ {\\mathrm {i}} (\\mathbf {x}, \\mathcal {E} (\\mathbf {x})), \\tag {9}", + "image_path": "bc15ba7c36b9e53157140794976af00b27f4dcc007497005fb87d4b737034be6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "type": "text", + "content": " is the weight of scattering offset, " + }, + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\hat{F}_{\\mathrm{sp}}" + }, + { + "bbox": [ + 313, + 700, + 554, + 714 + ], + "type": "text", + "content": " is a network" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26413" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "to integrate scattering reflectance, " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\hat{L}_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is a network to integrate the incoming light in a small region around " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathbf{x}) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is a learnable vector that defines the size of the integration domain. The values of " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{sp}}^{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " are obtained through a parameter network. We refer readers to the supplementary for more details of the network architecture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 154, + 148, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 148, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 148, + 166 + ], + "type": "text", + "content": "4.3. Training losses" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": "For both stages, we consider reconstruction losses including the pixel-wise color loss " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{col}} = \\| \\mathbf{C} - \\mathbf{C}_{\\mathrm{gt}}\\| _1" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": " and the eikonal loss " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eik}} = (\\| \\nabla S\\| _2 - 1)^2" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": ". Similar to NeuFace [58], we incorporate a regularization term " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{light}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": " to assume a nearly white environment light and penalize the specular energy by " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{spec}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": ". Additionally, we add " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{code}} = \\| \\mathbf{z}_{\\mathrm{geo}}\\| _2 + \\| \\mathbf{z}_{\\mathrm{rfl}}\\| _2" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": " to regularize the latent codes. To avoid a large variation between the predict reflectance attributes and the template during training, we apply a constraint to BRDF offset as " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ofs}} = \\| \\mathcal{B}_{\\mathrm{ofs}}(\\mathbf{x})\\| _1" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": ". Additionally, the albedo gradient predictor employs " + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{g}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 327 + ], + "type": "text", + "content": ", an L1 loss, which is supervised by the computed albedo gradient. As a result, we formulate the loss function in Stage 1 as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 338, + 294, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 294, + 353 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s t 1}} = \\mathcal {L} _ {\\mathrm {c o l}} + \\mathcal {L} _ {\\mathrm {e i k}} + \\mathcal {L} _ {\\mathrm {l i g h t}} + \\mathcal {L} _ {\\mathrm {s p e c}} + \\mathcal {L} _ {\\mathrm {o f s}} + \\mathcal {L} _ {\\mathrm {c o d e}} + \\mathcal {L} _ {\\mathrm {g}}. \\tag {10}", + "image_path": "19b78d6090861f3b1a4a5a0a9fd04ed276d4072bc22441b081852a0c86b51cbe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "content": "For the second stage, we add additional displacement loss " + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "content": " and scatter offset loss " + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ss}}" + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{st2}} = \\mathcal{L}_{\\mathrm{st1}} + \\mathcal{L}_{\\mathrm{dis}} + \\mathcal{L}_{\\mathrm{ss}}" + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}} = \\| \\mathcal{D}(\\mathbf{x})\\|_1" + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ss}} = \\| F_{\\mathrm{ss}}(\\mathbf{x})\\|_1" + }, + { + "bbox": [ + 55, + 362, + 295, + 422 + ], + "type": "text", + "content": " are regularization losses aimed at constraining high values of the displacement and the subsurface scattering offset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 178, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 178, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 178, + 449 + ], + "type": "text", + "content": "5. Experimental results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 455, + 170, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 170, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 170, + 468 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 653 + ], + "type": "text", + "content": "Dataset and metrics. We evaluate our method on a large human face dataset Facescape [48]. Specifically, we utilize authorized data from approximately 30 individuals, 4 of whom are from the publishable list. To protect the privacy of other identities that are not on the publishable list, we employed mosaic technology in the eyes region. To validate the generalization performance of our method, we additionally use dataset H3DS [40] in the supplementary material. We assess the quality of image rendering using the PSNR, SSIM, and LPIPS [54] metrics, and the Chamfer distance (CD) for evaluating the accuracy of geometry reconstruction. We compare our method with two 3D reconstruction methods VolSDF [51], DeformHead [47] and three BRDF decomposition methods PhySG [53], TensoIR [24], and NeuFace [58]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Implementation details. Facescape has over 50 scans of human heads along with camera parameters and meshes. We preprocess the images to obtain face masks using ImFace [57] tools. For Stage 1 training, we select 10 subjects, each with 10 views. For Stage 2, we use an additional" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "20 subjects and choose 3, 5, and 10 views from the available " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "50+" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " views as the training set, with another 10 views reserved for testing. The 10-view experiments specifically evaluate performance in a low-view setting. All test data is independent of the training set and is used consistently across all methods for comparison. We employ the Adam optimizer with an initial learning rate " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " to train both stages for 3000 epochs. Additionally, we add a calibration module as [58] to mitigate the exposure differences between various views and apply this calibration to all the methods for a fair comparison. For more implementation details, please refer to the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 227, + 457, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 457, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 457, + 239 + ], + "type": "text", + "content": "5.2. Evaluation & comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 553, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 435 + ], + "type": "text", + "content": "To corroborate the effectiveness of our proposed framework in the reconstruction of appearance and geometry, we conducted a comprehensive comparison with state-of-the-art methods. These methods are categorized into two groups: geometry-oriented methods, which include VolSDF [51] and DeformHead [47]; and decomposition-based methods, including PhySG [53], TensoIR [24], NeuFace [58], and our method SFDM. As shown in Tab. 1, we enumerate the quantitative performance of different methods in the evaluation metrics. To assess the performance capability under sparse-view images, we conducted tests on 20 subjects with 3, 5, and 10 views, respectively. The visual results of the reconstruction from 3 views are presented in Fig. 4, showcasing the synthesized rendering images from a novel viewpoint, the geometric reconstruction, and the decomposition of diffuse and specular attributes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 688 + ], + "type": "text", + "content": "Decomposition-based methods. Our method exhibits superior performance in novel view rendering and geometry reconstruction compared to other reflectance decomposition methods. In contrast to PhySG and TensoIR, designed for common objects, our method significantly outperforms them across all metrics. This is due to the complexity of facial material reflectance, which requires specialized reflectance modeling. NeuFace, which is customized for facial skin, poses a challenge due to its non-robust decomposition of geometry and reflectance attributes. As shown in Fig. 4, it often happens in NeuFace that all reflectance is baked into albedo, but the specular term is entirely absent. This degenerate Lambertian model may reduce the realism of the synthesized rendering output, especially for elements such as teeth, eyes, or the oily layer on facial skin, which often exhibit specular reflection effects. In contrast, we are able to stably decompose the diffuse and specular properties, resulting in more realistic rendering effects. Additionally, we demonstrate higher accuracy in geometric reconstruction, whereas the quality of NeuFace's geometry deteriorates significantly as the number of views decreases." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Geometry-oriented methods. Our method demonstrates a comparable level of accuracy in geometry reconstruc" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26414" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 542, + 161 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 542, + 161 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 542, + 161 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 542, + 161 + ], + "type": "table", + "html": "
MethodRelight3 views5 views10 views
PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓PSNR↑SSIM↑LPIPS↓CD↓
VolSDF [51]✗22.950.8686.9218.2825.450.8976.097.4027.340.9105.185.25
DeformHead [47]✗26.020.8865.769.6026.410.8794.838.4427.350.8915.205.42
PhySG [53]✓22.030.8637.8170.8722.690.8727.0563.2023.110.8767.1255.7
TensoIR [24]✓22.350.6457.1428.6024.610.7045.8617.1026.140.7694.7310.10
NeuFace [58]✓23.540.8765.7815.7425.810.9044.758.6227.190.9144.036.40
Ours✓25.540.9024.469.2726.730.9113.837.0027.640.9203.865.53
", + "image_path": "1c1692e091e7d80015824f9b2e53b1f0d012ac0bad1df64e423bbcfc52f6d8c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 204, + 286, + 593 + ], + "blocks": [ + { + "bbox": [ + 65, + 204, + 286, + 593 + ], + "lines": [ + { + "bbox": [ + 65, + 204, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 65, + 204, + 286, + 593 + ], + "type": "image", + "image_path": "61a0061f77082847219d80bf365a60c6c19520fdcf4f2a7e5fc11bbee735fc0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 598, + 295, + 633 + ], + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 633 + ], + "type": "text", + "content": "Figure 4. Visualization of 3-view reconstruction results. VolSDF and DeformHead do not incorporate reflectance attributes, resulting in the absence of corresponding outcomes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "tion compared to VolSDF and DeformHead. VolSDF exhibits poor 3D reconstruction results in sparse view settings. While DeformHead achieves a more detailed 3D reconstruction utilizing a geometry template, it is overly sensitive to geometry deformation, leading to noticeable indenta" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "type": "text", + "content": "tions. Although the PSNR of our method is lower than that of DeformHead under 3 views, our LPIPS outperforms it, indicating more photorealistic synthesis results. Failure instances in reconstruction have been observed with VolSDF and DeformHead, with failure rates of " + }, + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "type": "inline_equation", + "content": "21.7\\%" + }, + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 208, + 555, + 281 + ], + "type": "text", + "content": ", respectively. We exclude these failures from the comparison." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 289, + 555, + 373 + ], + "blocks": [ + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "type": "text", + "content": "Table 1. Evaluation. The LPIPS and CD metrics are measured on scales of " + }, + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "type": "inline_equation", + "content": "{10}^{-2}" + }, + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "type": "inline_equation", + "content": "{10}^{-3}" + }, + { + "bbox": [ + 55, + 168, + 555, + 194 + ], + "type": "text", + "content": " ,respectively. The best and second-best results are highlighted in red and orange, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 289, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 289, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 555, + 373 + ], + "type": "table", + "html": "
ExpMethodPSNR↑SSIM↑LPIPS↓CD↓
1w/o facial template24.270.8575.5917.22
2w/o albedo gradient25.130.8774.668.83
3w/o attribute synergy24.790.8764.908.80
4w/o scattering offset24.600.8745.738.61
5Ours25.650.8824.457.94
", + "image_path": "9f2e285b1bd6d72dfc1dd4f3864d12669744213c5b62ea795d03505a926fa858.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 380, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 323, + 380, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 323, + 380, + 545, + 392 + ], + "type": "text", + "content": "Table 2. Comparative evaluation metrics for ablation studies." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 326, + 403, + 544, + 509 + ], + "blocks": [ + { + "bbox": [ + 326, + 403, + 544, + 509 + ], + "lines": [ + { + "bbox": [ + 326, + 403, + 544, + 509 + ], + "spans": [ + { + "bbox": [ + 326, + 403, + 544, + 509 + ], + "type": "image", + "image_path": "067950823dea8c9180dbf1f0ab6bed6a216d70d8b7027c411f2f339ce3ef6151.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 509, + 555, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 531 + ], + "type": "text", + "content": "Figure 5. Ablation visualization. Top: Rendering. Bottom: Geometric errors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 552, + 413, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 413, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 413, + 564 + ], + "type": "text", + "content": "5.3. Ablation studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We select a subset of 10 subjects from the dataset to test the effectiveness of the major components of SFDM. In Tab. 2, Exp.1-2 are results for the components in Stage 1, and Exp.3-4 cover the components in Stage 2. Exp.1-2 indicate that the facial template significantly influences the robustness of face decomposition, with the albedo gradient predictor further enriching the reconstruction quality by introducing additional facial prior knowledge. From Exp.3, we observe that the attribute synergy module appears to contribute to geometry decomposition, and its absence leads to a noticeable decrease in geometric accuracy. Based on Exp.4, the subsurface scatter offset module can enhance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26415" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "the realism of the synthesized images from novel views, which is suggested by improvements in the LPIPS metric. In Fig. 5, columns 3-6 show the visualization results when omitting the facial template " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "T_{*}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", albedo gradient predictor " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", attribute synergy module " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{*}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", and subsurface scattering " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ss}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", respectively. The bottom row highlights the albedo gradient predictor's more significant impact on individuals who exhibit greater deviations from the facial template. In Stage 2, the subsurface scattering offset introduces additional details in light reflection on the skin, resulting in more photorealistic rendering results. Regarding the attribute synergy module, the exchange of information between geometry and reflectance enables a more accurate decomposition, as is evident in the details of geometry reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 243, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 243, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 243, + 260 + ], + "type": "text", + "content": "5.4. Reflectance decomposition analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 265, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 552 + ], + "type": "text", + "content": "Our method focuses on reflectance decomposition from multi-view facial images without any ground truth reflectance data, enabling realistic reconstruction of facial texture and highlights for novel view rendering. While our goal is not a physically accurate decomposition of skin reflectance, we aim to evaluate our decomposition results both visually and quantitatively. To this end, we use the single-view face reconstruction model, RefMM [20], to fit and decompose images in the test set, generating a coarse diffuse and specular result as pseudo ground truth for evaluation. Tab. 3 shows that our decomposition accuracy generally outperforms that of NeuFace. The overall low SSIM values are primarily due to the coarse nature of the pseudo ground truth, as illustrated in Fig. 6. This reflects certain advantages and limitations of single-view decomposition methods. These methods are typically trained on extensive facial image datasets with specialized lighting conditions or pseudo reflectance ground truth, allowing them to robustly decompose faces from monocular images. However, because they rely heavily on a parameterized facial model and limited input with less 3D information, the reconstructed results lack many facial details, making the rendered images appear less realistic. As shown in Fig. 6 and Fig. 7, we can observe these differences visually in the rendered results." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 105, + 562, + 248, + 614 + ], + "blocks": [ + { + "bbox": [ + 105, + 562, + 248, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 562, + 248, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 248, + 614 + ], + "type": "table", + "html": "
MethodDiffuseSpecular
NeuFace0.7260.552
Ours0.7610.692
", + "image_path": "9d8855a34049f9e2ba93d54dc262eb6ddc38c15311dab06a6ee324a34509d642.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 618, + 282, + 630 + ], + "lines": [ + { + "bbox": [ + 68, + 618, + 282, + 630 + ], + "spans": [ + { + "bbox": [ + 68, + 618, + 282, + 630 + ], + "type": "text", + "content": "Table 3. Quantitative diffuse and specular results (SSIM↑)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 104, + 638, + 249, + 696 + ], + "blocks": [ + { + "bbox": [ + 104, + 638, + 249, + 696 + ], + "lines": [ + { + "bbox": [ + 104, + 638, + 249, + 696 + ], + "spans": [ + { + "bbox": [ + 104, + 638, + 249, + 696 + ], + "type": "image", + "image_path": "a36b91b247728a7e3d087dc5ffbf506322a7d0e1b02808ffbf17d381a0bb785a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 699, + 264, + 710 + ], + "lines": [ + { + "bbox": [ + 86, + 699, + 264, + 710 + ], + "spans": [ + { + "bbox": [ + 86, + 699, + 264, + 710 + ], + "type": "text", + "content": "Figure 6. Decomposition results of RefMM [20]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 397, + 85 + ], + "type": "text", + "content": "5.5. Applications" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 188 + ], + "type": "text", + "content": "Specular editing. Our robust decomposition of diffuse and specular reflectances allows us to easily adjust the specular reflectance, resulting in the ability to make faces appear shinier or dimmer. The specular editing outcomes are showcased in Fig. 7. Specifically, we increase the specular weight by 1.5 times, while maintaining the overall specular distribution unchanged. As a result, the edited images can retain their photorealistic appearance." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 198, + 552, + 310 + ], + "blocks": [ + { + "bbox": [ + 315, + 198, + 552, + 310 + ], + "lines": [ + { + "bbox": [ + 315, + 198, + 552, + 310 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 552, + 310 + ], + "type": "image", + "image_path": "3effd274954fa97bb2e5d16afce37fceb07ec51d9b5c43d41813f34dc7122fcc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 313, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 335 + ], + "type": "text", + "content": "Figure 7. Applications. (a) Specular editing by increasing the shininess. (b) Relighting under different environments." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 444 + ], + "type": "text", + "content": "Relighting. Decomposing reflectance and environment light facilitates the relighting of human faces without altering their inherent properties. As shown in Fig. 7, we conducted a comparison of the relighting effects between our method and NeuFace in sparse view settings. Our method excels in decomposing specular reflectance, leading to more vibrant and realistic relighting outcomes under varying environmental lighting conditions [12]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 388, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 388, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 388, + 474 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 483, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 639 + ], + "type": "text", + "content": "We propose a novel face decomposition method capable of rendering and reconstructing 3D human faces from sparse-view images. Our framework first learns a facial template with both decomposed geometry and reflectance attributes from multi-view images of different individuals. Then we utilize the prior knowledge of the facial template and further consider the interaction between geometry and reflectance, as well as subsurface scattering effects, which enable our method to achieve robust face decomposition from sparse-view images. Based on decomposed facial attributes, our method achieves superior results in the rendering and geometric reconstruction of 3D human faces and can be utilized for relighting and specular editing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "Limitations & future work. As we focus on face decomposition, it is hard to extend to general objects directly with our method. In the future, we could involve techniques such as hash encoding [33] to accelerate our method and leverage fine-detailed 3D prior or reflectance ground truth for a better decomposition effect." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26416" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 186 + ], + "type": "text", + "content": "This work was supported in part by the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RT19/22) and the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). J.Hu was supported by the National Natural Science Foundation of China under Grants 62402083." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 199, + 115, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 115, + 212 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 115, + 212 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 219, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 219, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 274 + ], + "type": "text", + "content": "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 61, + 276, + 296, + 330 + ], + "type": "text", + "content": "[2] Dejan Azinović, Olivier Maury, Christophe Hery, Matthias Nießner, and Justus Thies. High-res facial appearance capture from polarized smartphone images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16836-16846, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 332, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 296, + 396 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "text", + "content": "[4] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Milos Hašan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv preprint arXiv:2008.03824, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "type": "text", + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 512, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 544 + ], + "type": "text", + "content": "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 296, + 600 + ], + "type": "text", + "content": "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik Lensch, and Jan Kautz. Two-shot spatially-varying brdf and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3982-3991, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "type": "text", + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34: 10691-10704, 2021. 1, 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[10] Chengqian Che, Fujun Luan, Shuang Zhao, Kavita Bala, and Ioannis Gkioulekas. Towards learning-based inverse subsurface scattering. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "text", + "content": "[11] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 164, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 555, + 218 + ], + "type": "text", + "content": "[12] Zhang Chen, Anpei Chen, Guli Zhang, Chengyuan Wang, Yu Ji, Kiriakos N Kutulakos, and Jingyi Yu. A neural rendering framework for free-viewpoint relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5599-5610, 2020. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[13] Ziang Cheng, Hongdong Li, Yuta Asano, Yinqiang Zheng, and Imari Sato. Multi-view 3d reconstruction of a textureless smooth surface of unknown generic reflectance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16226-16235, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "type": "text", + "content": "[14] Abdallah Dib, Luiz Gustavo Hafemann, Emeline Got, Trevor Anderson, Amin Fadainejad, Rafael MO Cruz, and Marc-Andre Carbonneau. Mosar: Monocular semi-supervised model for avatar reconstruction using differentiable shading. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1770–1780, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 343, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 374 + ], + "type": "text", + "content": "[15] Craig Donner and Henrik Wann Jensen. Light diffusion in multi-layered translucent materials. ACM Transactions on Graphics (ToG), 24(3):1032-1039, 2005. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 377, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 554, + 430 + ], + "type": "text", + "content": "[16] Haiwen Feng, Timo Bolkart, Joachim Tesch, Michael J Black, and Victoria Abrevaya. Towards racially unbiased skin tone estimation via scene disambiguation. In European Conference on Computer Vision, pages 72-90. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 433, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 476 + ], + "type": "text", + "content": "[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 478, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 554, + 533 + ], + "type": "text", + "content": "[18] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 589 + ], + "type": "text", + "content": "[19] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 634 + ], + "type": "text", + "content": "[20] Yuxuan Han, Zhibo Wang, and Feng Xu. Learning a 3d morphable face reflectance model from low-cost data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8598-8608, 2023. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 635, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 679 + ], + "type": "text", + "content": "[21] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[22] Mingming He, Pascal Clausen, Ahmet Levent Taşel, Li Ma, Oliver Pilarski, Wenqi Xian, Laszlo Rikker, Xueming Yu, Ryan Burgert, Ning Yu, et al. Diffrelight: Diffusion-based" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26417" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "text", + "content": "facial performance relighting. In SIGGRAPH Asia 2024 Conference Papers, pages 1-12, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 295, + 151 + ], + "type": "text", + "content": "[23] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 295, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 295, + 209 + ], + "type": "text", + "content": "[24] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 212, + 295, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 295, + 244 + ], + "type": "text", + "content": "[25] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 295, + 278 + ], + "type": "text", + "content": "[26] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 281, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 295, + 346 + ], + "type": "text", + "content": "[27] Alexandros Lattas, Stylianos Moschoglou, Stylianos Ploumpis, Baris Gecer, Jiankang Deng, and Stefanos Zafeiriou. Fitme: Deep photorealistic 3d morphable model avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8629-8640, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 225, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 225, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 225, + 361 + ], + "type": "text", + "content": "[28] Andrew Lauritzen. Gpu gems 3, 2007. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 362, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 362, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 362, + 295, + 429 + ], + "type": "text", + "content": "[29] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 430, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 485 + ], + "type": "text", + "content": "[30] Connor Lin, Koki Nagano, Jan Kautz, Eric Chan, Umar Iqbal, Leonidas Guibas, Gordon Wetzstein, and Sameh Khamis. Single-shot implicit morphable faces with consistent texture parameterization. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 295, + 542 + ], + "type": "text", + "content": "[31] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Trans. Graph., 42(4): 114:1-114:22, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "text", + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 1, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "text", + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[35] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[36] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "text", + "content": "[37] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: a higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40 (6):1-12, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[38] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "text", + "content": "[39] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. MIT Press, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "text", + "content": "[40] Eduard Ramon, Gil Triginer, Janna Escur, Albert Pumarola, Jaime Garcia, Xavier Giro-i Nieto, and Francesc Moreno-Noguer. H3d-net: Few-shot high-fidelity 3d head reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5620-5629, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "type": "text", + "content": "[41] William AP Smith, Alassane Seck, Hannah Dee, Bernard Tiddeman, Joshua B Tenenbaum, and Bernhard Egger. A morphable face albedo model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5011-5020, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "type": "text", + "content": "[42] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "type": "text", + "content": "[43] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "text", + "content": "[44] Borom Tunwattanapong, Graham Fyffe, Paul Graham, Jay Busch, Xueming Yu, Abhijeet Ghosh, and Paul Debevec. Acquiring reflectance and shape from continuous spherical harmonic illumination. ACM Transactions on graphics (TOG), 32(4):1-12, 2013. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "type": "text", + "content": "[45] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[46] Yifan Wang, Aleksander Holynski, Xiuming Zhang, and Xuaner Zhang. Sunstage: Portrait reconstruction and re" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26418" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "type": "text", + "content": "lighting using the sun as a light stage. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20792-20802, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "text", + "content": "[47] Baixin Xu, Jiarui Zhang, Kwan-Yee Lin, Chen Qian, and Ying He. Deformable model-driven neural rendering for high-fidelity 3d reconstruction of human heads under low-view settings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17924-17934, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "type": "text", + "content": "[48] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 295, + 306 + ], + "type": "text", + "content": "[49] Jing Yang, Hanyuan Xiao, Wenbin Teng, Yunxuan Cai, and Yajie Zhao. Light sampling field and BRDF representation for physically-based neural rendering. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "text", + "content": "[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "type": "text", + "content": "[51] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "type": "text", + "content": "[52] Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12803-12813, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 542 + ], + "type": "text", + "content": "[53] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 599 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "type": "text", + "content": "[55] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Ner-factor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 295, + 689 + ], + "type": "text", + "content": "[56] Wenyi Zhao, Rama Chellappa, P Jonathon Phillips, and Azriel Rosenfeld. Face recognition: A literature survey. ACM computing surveys (CSUR), 35(4):399-458, 2003. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[57] Mingwu Zheng, Hongyu Yang, Di Huang, and Liming Chen. Imface: A nonlinear 3d morphable face model with implicit" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 340 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20343-20352, 2022. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[58] Mingwu Zheng, Haiyu Zhang, Hongyu Yang, and Di Huang. Neuface: Realistic 3d neural face rendering from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16868-16877, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[59] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[60] Zerong Zheng, Han Huang, Tao Yu, Hongwen Zhang, Yandong Guo, and Yebin Liu. Structured local radiance fields for human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15893-15903, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 275, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 340 + ], + "type": "text", + "content": "[61] Michael Zollhöfer, Justus Thies, Pablo Garrido, Derek Bradley, Thabo Beeler, Patrick Pérez, Marc Stamminger, Matthias Nießner, and Christian Theobalt. State of the art on monocular 3d face reconstruction, tracking, and applications. In Computer graphics forum, pages 523-550. Wiley Online Library, 2018. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26419" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_content_list.json b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf4381e679b59453320086f96113e41e0a439a1 --- /dev/null +++ b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "SGC-Net: Stratified Granular Comparison Network for Open-Vocabulary HOI Detection", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Lin Chong Shi Zuopeng Yang* Haojin Tang* Zhili Zhou Guangzhou University", + "bbox": [ + 222, + 203, + 776, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "linxin94@gzhu.edu.cn, shichong@e.gzhu.edu.cn, yzpeng@gzhu.edu.cn,", + "bbox": [ + 197, + 241, + 781, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tanghaojin@gzhu.edu.cn, zhou_zhili@163.com", + "bbox": [ + 305, + 258, + 683, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent open-vocabulary human-object interaction (OV-HOI) detection methods primarily rely on large language model (LLM) for generating auxiliary descriptions and leverage knowledge distilled from CLIP to detect unseen interaction categories. Despite their effectiveness, these methods face two challenges: (1) feature granularity deficiency, due to reliance on last layer visual features for text alignment, leading to the neglect of crucial object-level details from intermediate layers; (2) semantic similarity confusion, resulting from CLIP's inherent biases toward certain classes, while LLM-generated descriptions based solely on labels fail to adequately capture inter-class similarities. To address these challenges, we propose a stratified granular comparison network. First, we introduce a granularity sensing alignment module that aggregates global semantic features with local details, refining interaction representations and ensuring robust alignment between intermediate visual features and text embeddings. Second, we develop a hierarchical group comparison module that recursively compares and groups classes using LLMs, generating fine-grained and discriminative descriptions for each interaction category. Experimental results on two widely-used benchmark datasets, SWIG-HOI and HICO-DET, demonstrate that our method achieves state-of-the-art results in OV-HOI detection. Codes is available at GitHub.", + "bbox": [ + 86, + 340, + 483, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 746, + 223, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-Object Interaction (HOI) detection aims to localize human-object pairs and recognize their interactions, providing an efficient way for human-centric scene understanding. It plays a crucial role in various computer vision tasks, such as assistive robotics [36, 37] and video analysis [23, 51]. Recently, the emerging field of open-vocabulary HOI (OV-HOI) detection, which broadens the scope of HOI detection", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4ca9d7b9568999a5a5b647785ff2e13d338e679aef1f0b132c036eff939a3c9.jpg", + "image_caption": [ + "(a) Feature granularity deficiency" + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 898, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1142b082fad16054671e9ff7fe0c7d6b0c2565d355c28f2564464214bb83905b.jpg", + "image_caption": [ + "(b) Semantic similarity confusion", + "Figure 1. (a) The last layer captures high-level global semantics but contains fewer low-level local details compared to intermediate layers. The red dot marks the selected patch. (b) Category-level and part-level descriptions overlook inter-class similarity, leading to difficulty in distinguishing semantically similar classes." + ], + "image_footnote": [], + "bbox": [ + 517, + 421, + 898, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by recognizing and associating objects beyond predefined categories, has gained increasing attention.", + "bbox": [ + 511, + 611, + 905, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OV-HOI detection has made notable progress, largely due to the success of vision-language models (VLMs) like CLIP [42]. Unlike zero-shot learning, which prohibits the access of any unannotated unseen classes during training, open-vocabulary learning leverages class name embeddings from VLMs for auxiliary supervision [50, 62]. Therefore, existing CLIP-based HOI detection methods naturally fall under OV-HOI detection. However, these methods face challenges in open-world scenarios because they rely on pretrained object detectors, and using category names as classifiers struggles to capture the variability of HOIs. To overcome these challenges, recent works [22, 48] have removed the need for pretrained detectors by extracting HOI features directly through image encoders while incorporating language priors from LLMs via text encoders.", + "bbox": [ + 509, + 642, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, current OV-HOI detection methods without pretrained object detectors still suffer from", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4539", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "feature granularity deficiency. This issue arises because CLIP, trained with image-level alignment, produces globally aligned image-text features lacking the local detail needed for effective OV-HOI detection. As shown in Figure 1(a), the last layer focuses on high-level semantic features but neglects low-level local details like arm and face postures, compared to intermediate layers. Additionally, feature dissimilarity between shallow and deep layers increases with network depth, posing significant challenges in aligning intermediate visual features with text embeddings. Moreover, OV-HOI methods that rely on category-level or part-level text classifiers are prone to semantic confusion due to several inherent limitations. First, CLIP's training on large-scale, long-tail datasets tends to introduce biases towards certain classes [49]. For example, as shown on the left side of Figure 1(b), CLIP often confuses \"hug cat\" with \"hold cat.\" Second, descriptions generated by LLMs based solely on labels may fail to distinguish between semantically similar classes. As illustrated on the right side of Figure 1(b), LLMs produce similar part-level descriptions like \"arm extension\" for both \"hold cat\" and \"chase cat\".", + "bbox": [ + 89, + 90, + 480, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose an end-to-end OV-HOI detection network named SGC-Net. First, we design the granularity sensing alignment (GSA) module, leveraging multi-granularity visual features from CLIP to improve the transferability for novel classes. The primary challenge in this process is effectively aggregating global semantic features with local details to refine coarse interaction representations while ensuring alignment between intermediate visual features and text embeddings. Accordingly, we introduce a block partitioning strategy that groups layers based on their relative distances. Specifically, we divide the visual encoder of CLIP into several blocks, ensuring a little variation in the features within each block. To align local details more precisely, we perform a fusion of multi-granularity features within each block, assigning distinct and trainable Gaussian weights to each layer. Once these intra-block features are fused, we proceed to aggregate the fused features across all blocks to maintain global semantic consistency. Furthermore, to preserve the pretrained alignment between text and visual features in CLIP, we employ visual prompt tuning [7, 25, 61] by adding learnable tokens to the visual features of each layer.", + "bbox": [ + 91, + 412, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, we propose the hierarchical group comparison (HGC) module to effectively differentiate semantically similar categories by recursively grouping and comparing classes via LLMs. Specifically, we utilize clustering algorithms (e.g., $K$ -means [34]) for grouping. Even with extensive prior knowledge of LLMs, comparing numerous categories for OV-HOI detection remains challenging due to the quadratic growth of the description matrix with the number of categories. To address this challenge, we adopt group-specific comparison strategies to improve efficiency. For", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "smaller groups, we directly query LLMs for comparative descriptions, whereas for larger groups, we summarize the group's characteristics using LLMs and incorporate these summaries into new queries to highlight distinctive features of each class within the group. By recursively construct a class hierarchy, we can classify HOI by descending from the top to the bottom of the hierarchy, comparing HOI and text embeddings at each level.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the innovation of the proposed SGC-Net is three-fold: (1) The GSA module effectively aggregates multi-granular features while aligning the local details and global semantic features of OV-HOI; (2) The HGC module iteratively generates discriminative descriptions to refine the classification boundaries of labels; (3) The efficacy of the proposed SGC-Net is evaluated on two widely-used OV-HOI detection benchmark datasets, i.e., HICO-DET [5] and SWIG-HOI [46]. Experimental results show that our SGC-Net consistently outperforms state-of-the-art methods.", + "bbox": [ + 511, + 212, + 906, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 378, + 653, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. HOI Detection", + "text_level": 1, + "bbox": [ + 513, + 402, + 661, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing HOI detection works can be divided into two categories, two-stage methods [3, 9, 21, 26, 32, 41, 43, 55, 56] and one-stage methods [6, 11, 18, 19, 28, 44, 58]. Two-stage methods typically use an object detector to recognize humans and objects, followed by specialized modules to associate humans with objects and identify their interactions, e.g., multi-steams [11, 12, 27], graphs [9, 30, 31, 54] or compositional learning [13, 14]. In contrast, the one-stage approach detects human-object pairs and their interactions simultaneously, without requiring stepwise processing. In particular, RLIP [60] proposes a pre-training strategy for HOI detection based on image captions. PPDM [28] reformulates the HOI detection task as a point detection and matching problem and achieves simultaneous object and interaction detection. However, existing HOI detection approaches are constrained by a closed-set assumption, which restricts their ability to recognize only predefined interaction categories. In contrast, our work aims to detect and recognize HOIs in the open-vocabulary scenario.", + "bbox": [ + 511, + 426, + 903, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. CLIP-based Open Vocabulary HOI Detection", + "text_level": 1, + "bbox": [ + 511, + 724, + 897, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the closed-set limitation, many studies [10, 22, 29, 35, 40, 47, 52] have been devoted to OV-HOI detection, aiming to identify both base and novel categories of HOIs while only base categories are available during training. Specifically, they transfer knowledge from the large-scale visual-linguistic pre-trained model CLIP to enhance interaction understanding. For instance, GEN-VLKTs [29] and HOICLIP [40] convert HOI labels into phrase descriptions to initialize the classifier and extract visual features from the CLIP image encoder to guide the interaction visual feature", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4540", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning. MP-HOI [52] incorporates the visual prompts into language-guided-only HOI detectors to enhance its generalization capability. THID [48] proposes a HOI sequence parser to detect multiple interactions. CMD-SE [22] detects HOIs at different distances using distinct feature maps. However, the adopted loss function in [22] involves minimizing differences between continuous and discrete variables therefore hard to optimize. Besides, aligning intermediate features with text embeddings may break the alignment in pre-trained CLIP due to the substantial differences. The proposed approach is a method free from pretrained object detectors but has the following advantages compared with existing works. First, it aggregates multi-granularity features from the visual encoder more appropriately. Second, it is easy to optimize and deploy to existing models.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Leverage LLM for visual classification", + "text_level": 1, + "bbox": [ + 89, + 325, + 421, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, LLMs [1, 2, 8, 16] have made remarkable progress, revolutionizing natural language processing tasks with hundreds of billions of parameters and techniques such as reinforcement learning. Existing works [4, 17, 45, 59] have shown their effectiveness in generating comprehensive descriptions, particularly in classification and detection tasks. Specifically, Menon and Vondrick [38] generate textual descriptions directly from labels to assist VLMs in image classification. I2MVFormer [39] leverages LLM to generate multi-view document supervision for zero-shot image classification. ContextDET [53] utilizes contextual LLM tokens as conditional object queries to enhance the visual decoder for object detection. RECODE [24] leverages LLMs to generate descriptions for different components of relation categories. However, descriptions generated solely by LLMs from labels often tend to be generic and lack sufficient discriminability among semantically similar classes. To address this issue, our work leverages LLM to generate text descriptions by comparing different classes to reduce inter-class similarity and make the decision boundary of each class more compact.", + "bbox": [ + 89, + 349, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 679, + 181, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides details of the proposed stratified granular comparison network (SGC-Net). Specifically, we first describe the network architecture, followed by an explanation of the training and inference pipeline. As shown in Figure 2, SGC-Net comprises a granularity sensing alignment (GSA) module and a hierarchical group comparison (HGC) module. The GSA module enhances coarse HOI representations by utilizing distance-Aware Gaussian weighting (DGW) and visual prompt tuning to combine multi-granularity features from CLIP's image encoder. The HGC module refines decision boundaries for semantically similar classes by recursively comparing and grouping them using LLM-derived knowledge.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Granularity Sensing Alignment", + "text_level": 1, + "bbox": [ + 513, + 90, + 790, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing OV-HOI methods [22, 29, 40] typically use features from the last layer of CLIP's image encoder to model HOIs. While these deep features effectively capture high-level semantics, they often lack the local details necessary for HOI detection. One approach to address this issue is to aggregate feature maps from different levels and feed them into an interaction decoder. However, the significant differences between shallow and deep features can disrupt the pre-trained vision-language alignment, ultimately reducing CLIP's zero-shot capability for OV-HOI tasks.", + "bbox": [ + 511, + 112, + 903, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we propose the GSA module, which effectively captures fine-grained details in human-object interactions while preserving the vision-language alignment in the pre-trained CLIP model. Specifically, we first divide the CLIP's image encoder into $S$ blocks, ensuring a little variation in the features within each block. Then, within each block, distinct distance-aware Gaussian weights are assigned to the transformer layers based on their relative neighborhood distances. These weights are trainable, enabling the model to adaptively learn layer- and block-specific information from the training data. Finally, we aggregate the features across different blocks. For each block containing $d$ transformer layers, the aggregated feature $Z$ can be represented as follows:", + "bbox": [ + 511, + 265, + 906, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\alpha_ {l} ^ {s} = \\exp \\left(- \\frac {1}{2} \\frac {(d - l) ^ {2}}{\\sigma^ {2}}\\right), \\quad l \\in [ 1, d ], \\\\ \\boldsymbol {Z} = \\sum_ {s = 1} ^ {S} \\alpha_ {s} \\left(\\sum_ {l = 1} ^ {d} \\alpha_ {l} ^ {s} \\boldsymbol {F} _ {l}\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 486, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $l$ represents the layer index. $\\pmb{F}_l$ is a feature map from the $l$ -th layer of CLIP image encoder. $\\alpha_l^s$ and $\\alpha_s$ denote the distance-aware Gaussian weights between layers within the same block and across different blocks, respectively. By adjusting variance parameter $\\sigma$ , the Eq. (1) can assign a higher weight to nearby layers and a lower weight to distant ones, facilitating more effective and flexible integration of features across different depth levels. Notably, the last layer of CLIP is treated as a separate block with a large weight. This design aims to keep the original visual-language correlations established in the pre-trained model. Different blocks can capture varying levels of granularity. Specifically, the early blocks focus more on fine-grained local details, while the later block emphasizes coarse-grained global features.", + "bbox": [ + 511, + 568, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better align the intermediate visual features with the text embeddings, we employ visual prompt tuning [59] by introducing learnable tokens onto the visual features of each layer in the frozen encoder. During visual prompt tuning, the GSA module enables the gradients to be directly backpropagated to the middle layers of the visual encoder. It can promote the alignment of mid-layer features and text embeddings, enhancing the similarity across different layers.", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4541", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/741a6c5728444953f1fa0a0b475ab9f1e19b064463d0935e07a47f15dcf36b33.jpg", + "image_caption": [ + "Figure 2. The framework of SGC-Net. It eliminates the need for a pretrained object detector and includes two new modules for OV-HOI detection: (1) The GSA module partitions the CLIP visual encoder into blocks, aggregates features via DGW, and integrates an HOI decoder for fine-grained HOI representations. (2) The HGC module uses LLM to recursively construct class hierarchy, enabling HOI classification by traversing the hierarchy from top to bottom and comparing HOI representations with text embeddings at each level." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with CMD-SE [22], we effectively aggregate multi-granularity features from the CLIP image encoder using trainable Gaussian weights, which have the following advantages. First, it allows intermediate layer visual features to complement last layer features while preserving the visual-language associations of the pre-trained CLIP. This approach addresses the issue in CMD-SE [22], where relying solely on discrete single-layer features for HOI modeling results in a loss of fine-grained information. Second, our method is easy to optimize. We eliminate the need to use loss functions that align discrete transformer layer indices with continuous human-object interaction distances, thereby avoiding the complex optimization processes.", + "bbox": [ + 88, + 500, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, following previous works [22, 29, 40], the final interaction representation $X$ is formulated as follows:", + "bbox": [ + 89, + 696, + 483, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} = \\operatorname {D e c} (\\boldsymbol {Q}, \\boldsymbol {Z}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 741, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ refers to HOI queries, while $Z$ is treated as both the key and value input to the HOI Decoder. The output $X$ is later passed to the bounding box regressor and classifier to predict the bounding boxes of human-object pairs and their interaction categories.", + "bbox": [ + 89, + 763, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Hierarchical Group Comparison", + "text_level": 1, + "bbox": [ + 89, + 847, + 377, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the fine-grained interaction representation by the GSA module, designing a discriminative classifier re", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mains a challenging task. Existing methods [22, 29, 40] typically design classifiers with the manual prompt \"a photo of a person [verb] [object]\". However, these classifiers primarily rely on category names, often neglecting the contextual information provided by language. Recent research [22] has introduced the category descriptions produced by LLMs to enhance the model's generalization capability. However, these descriptions directly generated by LLMs from labels are often generic and semantically similar.", + "bbox": [ + 511, + 500, + 906, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we propose the hierarchical group comparison (HGC) module. Inspired by the coarse-to-fine approach humans use to recognize objects, HGC employs three strategies: Grouping, Comparison, and Hierarchical Classification. It identifies semantically similar descriptions and refines decision boundaries from generic to more discriminative. Specifically, we first utilize LLMs to generate initial descriptions for each HOI category.", + "bbox": [ + 511, + 637, + 908, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Q: What features are useful to distinguish {HOI category} in a photo?", + "bbox": [ + 511, + 762, + 888, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A:-", + "bbox": [ + 514, + 792, + 553, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grouping: Following [38], we employ the pretrained CLIP text encoder to map the LLM-generated descriptions into the latent space. Utilizing these features, we apply a clustering algorithm, such as $K$ -means [34], to perform grouping and identify semantic neighbors. Specifically, the value of $K$ is calculated as the number of categories divided by a", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4542", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predefined grouping threshold $N$ . Even with extensive prior knowledge of LLMs, comparing a large number of categories presents a formidable challenge, as it needs to generate a comprehensive description matrix that grows quadratically with the number of categories. To address this issue, we employ tailored strategies for generating descriptions based on the group size, allowing for more efficient and effective comparisons. Specifically, we adopt summary-based comparison for larger groups and direct comparison for smaller ones.", + "bbox": [ + 89, + 90, + 480, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing: For groups involving a large number of categories (i.e., exceeding half of the grouping threshold), we leverage the LLM to summarize their overall characteristics. By incorporating these summarized characteristics into a new query, we can generate a comparative description that captures the distinctive features and relationships among the categories. This approach enhances our ability to effectively compare and contrast the elements within a group.", + "bbox": [ + 89, + 242, + 480, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Summarize the following interactions", + "bbox": [ + 89, + 368, + 437, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with one sentence: {category list}?", + "bbox": [ + 91, + 383, + 408, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: {subset description}", + "bbox": [ + 91, + 400, + 295, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: What features are useful to distinguish {HOI category} from {subset description}?", + "bbox": [ + 89, + 422, + 465, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:-", + "bbox": [ + 91, + 455, + 130, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the number of elements within a group is relatively small, a direct query to the LLM enables us to obtain an exceptionally comparative description. This approach capitalizes on the LLM's capabilities to provide a comprehensive and insightful analysis, enhancing the comparative understanding of the elements within the group.", + "bbox": [ + 89, + 470, + 480, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: What features are useful to distinguish {target category} from {other categories} in a photo?", + "bbox": [ + 89, + 566, + 478, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:-", + "bbox": [ + 91, + 614, + 130, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Utilizing these detailed and comparative descriptions, we can achieve more nuanced text embedding. Subsequently, we can continue to perform clustering, followed by additional comparisons to derive new descriptors.", + "bbox": [ + 89, + 631, + 480, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical Classification: After constructing the class hierarchy, we can obtain text features at multiple hierarchical levels. More specific, the hierarchical text features can be represented as $\\mathcal{T} = \\{D_1, D_2, \\dots, D_T\\}$ , where $T$ is the number of test classes. $D_i \\in \\mathbb{R}^{M_j \\times C}$ represents the hierarchical text embeddings for category $i$ . $C$ is the embedding dimension. $M_j$ is the number of comparative descriptions for category $i$ . Accordingly, the cosine similarity score between the HOI feature $\\pmb{x}$ and the $j$ -th description of the $i$ -th HOI category, denoted as $D_i^j$ , can be expressed as follows:", + "bbox": [ + 89, + 691, + 482, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} ^ {j} = \\boldsymbol {D} _ {i} ^ {j} \\cdot \\boldsymbol {x} ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 858, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, unreliable low-level descriptions can introduce", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "errors and redundancies, degrading the quality of high-level descriptions and ultimately impairing the model's classification performance. To address this issue, we define an iterative evaluator $\\pmb{u}_i^k$ to evaluate the acceptability of the current $k$ -th level description for the $i$ -th category as follows:", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {i} ^ {k} = \\mathbb {I} \\left(p _ {i} ^ {k + 1} > p _ {i} ^ {k} + \\tau\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 181, + 903, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ denotes the tolerance parameter, $\\mathbb{I}$ represents indicator function. Eq. (4) ensures that a new score is merged only if a subsequent discriminative description yields a higher score than the current one. Subsequently, we obtain the running average of the longest sequence of monotonically increasing $p$ values as follows:", + "bbox": [ + 511, + 209, + 903, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (\\boldsymbol {x}, i) = \\frac {p _ {i} ^ {1} + \\sum_ {j = 2} ^ {M _ {i}} p _ {i} ^ {j} \\prod_ {k = 1} ^ {j - 1} \\boldsymbol {u} _ {i} ^ {k}}{1 + \\sum_ {j = 2} ^ {M _ {i}} \\prod_ {k = 1} ^ {j - 1} \\boldsymbol {u} _ {i} ^ {k}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 311, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the similarity score $s(\\pmb{x}, i)$ between HOI feature $\\pmb{x}$ and the $i$ -th interaction label embedding is computed using a weighted average function, as follows:", + "bbox": [ + 511, + 367, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns (\\boldsymbol {x}, i) = (1 - \\lambda) \\left(\\boldsymbol {p} _ {i} ^ {1} + \\boldsymbol {t} \\cdot \\boldsymbol {x} ^ {T}\\right) + \\lambda (r (\\boldsymbol {x}, i)). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 425, + 903, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we apply prompt tuning [48] with learnable text tokens $t$ to learn the sentence format rather than using manually defined context words. Besides, our fusion method takes into account the score of the initial description $p_i^1$ as an offset, and we introduce a constant hyperparameter $\\lambda \\in [0, 1]$ to balance the two terms.", + "bbox": [ + 511, + 455, + 905, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 513, + 558, + 725, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we elaborate on the processes of training and inference of our model.", + "bbox": [ + 511, + 582, + 903, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. During the training stage, we follow the query-based methods [22, 29, 48] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. The matching cost $\\mathcal{L}$ for the matching process and the targeting cost for the training back-propagation share the same strategy, which is formulated as follows:", + "bbox": [ + 511, + 613, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {b} \\sum_ {i \\in \\{h, o \\}} \\mathcal {L} _ {b} ^ {i} + \\lambda_ {i o u} \\sum_ {i \\in \\{h, o \\}} \\mathcal {L} _ {i o u} ^ {i} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 729, + 903, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_b$ , $\\mathcal{L}_{iou}$ , and $\\mathcal{L}_{cls}$ denote the box regression, intersection over union, and classification losses, respectively. During the training stage, we follow the query-based methods [22, 29, 40] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. $\\lambda_b$ , $\\lambda_{iou}$ , and $\\lambda_{cls}$ are the hyper-parameter weight.", + "bbox": [ + 511, + 776, + 905, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. For each HOI prediction, including the bounding-box pair $(\\hat{b}_h^i,\\hat{b}_o^i)$ , the bounding box score $\\hat{c}_i$ from", + "bbox": [ + 511, + 868, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4543", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the box regressor, and the interaction score $\\hat{s}_i$ from the interaction classifier, the final score $\\hat{s}_i'$ is computed as:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {i} ^ {\\prime} = \\hat {s} _ {i} \\cdot \\hat {c} _ {i} ^ {\\gamma} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 133, + 482, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma > 1$ is a constant used during inference to suppress overconfident objects [54, 55].", + "bbox": [ + 89, + 164, + 482, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 209, + 215, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 234, + 290, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our experiments are conducted on two datasets: namely, SWIG-HOI [46] and HICO-DET [5]. The SWIG-HOI dataset provides diverse human interactions with large-vocabulary objects, comprising 400 human actions and 1,000 object categories. During the testing, we utilize approximately 5,500 interactions, including around 1,800 interactions that are not found in the training set. The annotations of HICO-DET include 600 combinations of 117 human actions and 80 objects. Among the 600 interactions, follow [22, 48], we simulate the open-vocabulary detection setting by holding out 120 rare interactions.", + "bbox": [ + 89, + 258, + 482, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Following in [5, 22, 29, 33, 48], we use the mean Average Precision (mAP) for evaluation. An HOI triplet prediction is classified as a true-positive example when the following criteria are satisfied: 1) The intersection over union of the human bounding box and object bounding box are larger than 0.5 w.r.t. the GT bounding boxes; 2) the predicted interaction category is accurate.", + "bbox": [ + 89, + 426, + 482, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We follow the settings of previous work [22, 48] to build our model upon the pretrained CLIP. Specifically, for the visual encoder, we employ the ViT-B/16 version as our visual encoder and apply 12 learnable tokens in each layer to detect human-object interactions. For the text encoder, we introduce 8 prefix tokens and 4 conjunctive tokens to connect the words of human actions and objects for adaptively learning categories' information. Our model is optimized using AdamW with an initial learning rate of $10^{-4}$ , using 64 as the batch size for SWIG-HOI dataset and 32 for HICO-DET. We set the cost weights $\\lambda_{b}$ , $\\lambda_{cls}$ and $\\lambda_{iou}$ to 5, 2, and 5, respectively. The LLM we utilize is GPT-3.5. We set the hyperparameter $\\lambda$ in Eq. (6) as 0.5. In all experiments, the variance parameter $\\sigma$ is set to 1, and the tolerance parameter $\\tau$ is set to 0.", + "bbox": [ + 89, + 532, + 482, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 770, + 465, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HICO-DET. To facilitate fair comparison, we first compare our model with state-of-the-art methods without pretrained detectors on the HICO-DET dataset. As shown in Table 1, SGC-Net outperforms CMD-SE [22] by $6.57\\%$ , $4.39\\%$ , and $4.87\\%$ in mAP on the Unseen, Seen, and Full categories, respectively. Furthermore, even when compared to methods using pretrained detectors, SGC-Net achieves", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0274f524d1678b71172bdb00bcddc1de0e633de2904acba9465713458b9441e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPretrained DetectorUnseenSeenFull
SCL [15]✓19.0730.3928.08
GEN-VLKT [29]✓21.3632.9130.56
OpenCat [57]✓21.4633.8631.38
HOICLIP [40]✓23.4834.4732.26
THID [48]✗15.5324.3222.38
CMD-SE [22]✗16.7023.9522.35
SGC-Net✗23.2728.3427.22
", + "bbox": [ + 516, + 89, + 903, + 252 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df89ab6a7eccf5fe45af0940af81e2e5e1f20545a220168f02480529bc00e3f1.jpg", + "table_caption": [ + "Table 1. Comparison of our proposed SGC-Net with state-of-the-art methods on the HICO-DET dataset." + ], + "table_footnote": [], + "table_body": "
MethodNon-rareRareUnseenFull
CHOID [46]10.936.632.646.64
QPIC [43]16.9510.846.2111.12
GEN-VLKT [29]20.9110.41-10.87
MP-HOI [52]20.2814.78-12.61
THID [48]17.6712.8210.0413.26
CMD-SE [22]21.4614.6410.7015.26
SGC-Net23.6716.5512.4617.20
", + "bbox": [ + 519, + 311, + 898, + 468 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/21604ab7af9ec9bf1880b2560c78edfcf0393fc244ab44c0ed7cb231e4adc571.jpg", + "table_caption": [ + "Table 2. Comparison of our proposed SGC-Net with state-of-the-art methods on the SWIG-HOI dataset." + ], + "table_footnote": [], + "table_body": "
MethodNon-rareRareUnseenFull
Base15.6911.537.3211.45
+ GSA22.7416.0011.6416.49
+ HGC21.1814.1910.6914.81
SGC-Net23.6716.5512.4617.20
", + "bbox": [ + 545, + 527, + 874, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation studies of the proposed method.", + "bbox": [ + 558, + 646, + 857, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "competitive performance. Although recent OV-HOI methods (e.g., GEN-VLKT [29], OpenCat [57], HOICLIP [40]) leverage CLIP text embeddings for interaction classification, they typically rely on a DETR architecture with pretrained weights. It is important to note that comparing our method with these approaches on HICO-DET dataset is not entirely fair, as the COCO dataset used for DETR pretraining shares the same object label space as HICO-DET.", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SWIG-HOI. Table 2 shows that SGC-Net outperforms all state-of-the-art methods on various setups. Specifically, SGC-Net outperforms CMD-SE [22] by $1.76\\%$ and $1.91\\%$ in mAP on the Rare and Unseen categories, respectively. Furthermore, SGC-Net outperforms the best OV-HOI method with a pretrained object detector by $4.59\\%$ in mAP on the Full category. These results indicate that", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4544", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/246720302836e8ecc43fa04265533b241d977e29f14036001f6f0a39d40ab6ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Layer splitting mannerNon-rareRareUnseenFull
{4-6}, {7-9}, {10-12}23.0315.7211.3516.35
{6-8}, {9-10}, {11-12}23.6116.5211.6417.01
{6-8}, {9-11}, {12}23.6716.5512.4617.20
", + "bbox": [ + 91, + 89, + 482, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78a6eba360e6bfbd2e042806eda08829caf5372b2a4798324a47ee7c0767ad6d.jpg", + "table_caption": [ + "Table 4. Effectiveness of different layer splitting manners. The elements in $\\{\\}$ means the layer numbers that are used." + ], + "table_footnote": [], + "table_body": "
Number of blocksNon-rareRareUnseenFull
{12}21.1814.1910.0914.81
{9-11}, {12}21.9114.4810.0114.78
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{3-5}, {6-8}, {9-11}, {12}22.6115.2512.0116.53
Number of layers
{8-9}, {10-11}, {12}22.4515.1711.5515.98
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{4-7}, {8-11}, {12}23.5016.2112.1316.75
", + "bbox": [ + 91, + 212, + 483, + 358 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea7ca0e49a9a028124b1c348c9e4e6814f6956b96175f5845f736e4cb6488ad4.jpg", + "table_caption": [ + "Table 5. Evaluation on the number of blocks and layers." + ], + "table_footnote": [], + "table_body": "
AggregationNon-rareRareUnseenFull
Self-attention21.3713.437.8813.90
Concat21.8415.5210.2215.68
Sum22.0914.899.5915.26
DGW23.4016.1411.6016.71
Sum*22.2115.1610.4515.64
DGW*23.6716.5512.4617.20
", + "bbox": [ + 109, + 390, + 464, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Comparison of different aggregation methods. * denotes that the weights are trainable.", + "bbox": [ + 89, + 536, + 482, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods such as GEN-VLKT [29] and MP-HOI [52], which rely on pretrained detectors, perform suboptimally on the SWIG-HOI dataset. This is primarily because these methods struggle to scale effectively with vocabulary size, ultimately limiting their applicability in open-world scenarios. In contrast, SGC-Net overcomes this constraint by not relying on any detection pretraining, demonstrating superior capabilities in detecting and recognizing OV-HOI.", + "bbox": [ + 88, + 580, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 243, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To prove the effectiveness of our proposed methods, we conduct five ablation studies on the SWIG-HOI dataset. Results of the ablation studies are summarized in Table 3-6, and Figure 3, respectively.", + "bbox": [ + 89, + 734, + 482, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the proposed modules. We first perform an ablation study to validate the effectiveness of the GSA and HGC modules. Results are summarized in Table 3. Our baseline model follows CMD-SE [22], which utilizes CLIP and removes the need for pretrained object detectors in OV-HOI detection. First, we apply the GSA module on the baseline model to improve its transferability for novel", + "bbox": [ + 88, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/557e00ab2c038bac9807248f7eb225e1320931244c580e5b4861354f5540343f.jpg", + "image_caption": [ + "Figure 3. Evaluation on the value of grouping threshold $N$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 859, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classes, denoted as $+GSA$ . This modification achieves mAP improvements of $7.05\\%$ for Non-Rare, $2.66\\%$ for Rare, $4.32\\%$ for Unseen, and $5.04\\%$ for Full categories. Moreover, we integrate the HGC module with the baseline to enhance its ability to distinguish semantically similar interactions, denoted as $+HGC$ . The results indicate that the HGC module enables the baseline model to achieve $14.81\\%$ in mAP for the Full categories. Finally, combining both modules results in the best performance, with an mAP of $12.46\\%$ on the Unseen categories. These results demonstrate the significant potential of SGC-Net in enhancing interaction understanding within an open-vocabulary setting.", + "bbox": [ + 511, + 309, + 906, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the proposed layer splitting manner. The block partitioning strategy plays a crucial role in the GSA module due to its potential to preserve visual-language associations. Experimental results in Table 4 indicate that separating the last layer into an independent block is more effective than other combinations of partitioning strategies. This is because the last layer of CLIP's image encoder has the strongest association with text embeddings, and partitioning it into a separate block helps preserve this correlation while minimizing potential disruptions.", + "bbox": [ + 511, + 491, + 905, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of blocks and number of layers in each block. To show the importance of multi-granularity feature fusion across different layers, we compare the performance of SGC-Net with varying numbers of blocks and layers per block. As shown in Table 5, increasing the number of blocks from 1 to 3 gradually improves the performance, particularly for Unseen interactions. The best performance is achieved with three blocks, each containing three layers. Reducing the number of layers to 2 degrades performance, primarily due to neglecting mid-layer features. It should be noted that the beginning layers are excluded, as they primarily encode features with limited semantics.", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of different aggregation methods. Due to the substantial differences between shallow and deep features, directly integrating multi-granularity features may disrupt the alignment of vision-language embeddings in the pre-trained CLIP model. To address this issue, we intro-", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4545", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca73b4b25118e362dbaabcea7a673b7bc42c770fc5ceea5f056d477730280d57.jpg", + "image_caption": [ + "Figure 4. Qualitative results of SGC-Net. On the left-hand side, we visualize feature correspondence through cosine similarity calculations from both deep and shallow layers, and compare the predicted HOI categories. On the right-hand side, we show two inference examples and the absolute score gap between two images at each hierarchical level by querying the description with extracted HOI representations." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 496, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/197b8058686b77ad67f4140cc3f469d85d2d0d80b70ee5ae9823dafd9a044837.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 90, + 903, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "duce the DGW function, which mitigates these disruptions by accounting for the distances between layers. As shown in Table 6, DGW consistently outperforms direct aggregation strategies such as Sum, Concat. Additionally, setting the aggregation weights as learnable further enhances the performance of DGW. These results demonstrate that our method, by assigning smaller weights to features from more distant layers during multi-granularity fusion, effectively improves model performance in open-vocabulary tasks.", + "bbox": [ + 88, + 362, + 480, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Grouping threshold $N$ . A large grouping threshold, such as $N = 8$ , tends to produce rudimentary hierarchies and generalized descriptions. In contrast, smaller grouping thresholds increase the time required to construct the class hierarchy due to the more detailed recursive comparisons and groupings involved. As shown in Figure 3, the performance of SGC-Net improves as the grouping threshold increases. To balance computational efficiency and performance gains, we chose to set $N = 6$ .", + "bbox": [ + 88, + 500, + 482, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 648, + 272, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the proposed GSA module in leveraging CLIP's multi-granularity features, we present cosine similarity maps of features and qualitative OV-HOI detection results. At the left hand side of Figure 4, we show the patch similarity of Unseen classes not included in the training process. We observe that the intermediate layers of our method contain detailed information on local objects, including interaction patterns. Moreover, by leveraging these distinctive features, our SGC-Net improves HOI detection performance for both seen and unseen classes compared to using only last-layer features.", + "bbox": [ + 88, + 672, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At the right hand side of Figure 4, we show two groups of examples that demonstrate the inference and absolute score gap calculation via the proposed HGC module. Specifically, we input two Unseen class images (② and ④) and two", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Rare class images (① and ③) to evaluate their similarity with the hierarchical descriptions. The absolute score gap is computed as the absolute difference in similarity scores between paired images: i.e., ① vs. ② and ③ vs. ④. Due to the weak comparative information in the early descriptions, similar images tend to yield similar scores at the beginning. As we progress towards a more comparative description, the disparity between the scores of the two images gradually increases. Besides, this also highlights the interpretability of our method, as it provides the user insights regarding the attributes that elicit stronger responses to the input. Additionally, it indicates the specific layer of description at which the scores of different images diverge, widening the gap.", + "bbox": [ + 511, + 362, + 906, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 579, + 633, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose the SGC-Net which enhances CLIP's transferability and discriminability to handle two critical issues in OV-HOI: feature granularity deficiency and semantic similarity confusion. Specifically, the GSA module enhances CLIP's image encoder by improving its ability to capture fine-grained HOI features while maintaining alignment between multi-granularity visual features and text embeddings. The HGC module strengthens CLIP's text encoder in distinguishing semantically related HOI categories by incorporating a grouping and comparison strategy into the LLM. Through extensive comparative experiments, we validate the effectiveness of SGC-Net on two datasets.", + "bbox": [ + 511, + 606, + 906, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work is supported by the Guangzhou University Research Project (No: RQ2021013, 2025A03J3128), the Guangdong Natural Science Funds for Distinguished Young Scholars (No. 2023B1515020041), the Guangdong Basic and Applied Basic Research Foundation (No. 2023A1515110077), the NSF of China (No. 62372125).", + "bbox": [ + 511, + 789, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4546", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 3", + "[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[3] Yichao Cao, Qingfei Tang, Feng Yang, Xiu Su, Shan You, Xiaobo Lu, and Chang Xu. Re-min, learn and reason: Exploring the cross-modal semantic correlations for language-guided hoi detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23492-23503, 2023. 2", + "[4] Yichao Cao, Qingfei Tang, Xiu Su, Song Chen, Shan You, Xiaobo Lu, and Chang Xu. Detecting any human-object interaction relationship: Universal hoi detector with spatial prompt learning on foundation models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[5] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6", + "[6] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004-9013, 2021. 2", + "[7] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, pages 11583-11592, 2022. 2", + "[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3", + "[9] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 696-712. Springer, 2020. 2", + "[10] Jianjun Gao, Kim-Hui Yap, Kejun Wu, Duc Tri Phan, Kratika Garg, and Boon Siew Han. Contextual human object interaction understanding from pre-trained large language model. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13436-13440. IEEE, 2024. 2", + "[11] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2", + "[12] Tanmay Gupta, Alexander Schwing, and Derek Hoiem. Nofrills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2", + "[13] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 2", + "[14] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 2", + "[15] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 6", + "[16] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-imiL: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 3", + "[17] Sheng Jin, Xueying Jiang, Jiaxing Huang, Lewei Lu, and Shijian Lu. Llms meet vlms: Boost open vocabulary object detection with fine-grained descriptors. arXiv preprint arXiv:2402.04630, 2024. 3", + "[18] Bumsoo Kim, Taeho Choi, Jaewoo Kang, and Hyunwoo J Kim. Uniondet: Union-level detector towards real-time human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 498-514. Springer, 2020. 2", + "[19] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 2", + "[20] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[21] Ting Lei, Fabian Caba, Qingchao Chen, Hailin Jin, Yuxin Peng, and Yang Liu. Efficient adaptive human-object interaction detection with concept-guided memory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6480-6490, 2023. 2", + "[22] Ting Lei, Shaofeng Yin, and Yang Liu. Exploring the potential of large foundation models for open-vocabulary hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16657-16667, 2024. 1, 2, 3, 4, 5, 6, 7", + "[23] Huadong Li, Ying Wei, Shuailei Ma, Mingyu Chen, and Ge Li. Ripple transformer: A human-object interaction backbone and a new prediction strategy for smart surveillance devices. IEEE Transactions on Consumer Electronics, 2024. 1", + "[24] Lin Li, Jun Xiao, Guikun Chen, Jian Shao, Yueting Zhuang, and Long Chen. Zero-shot visual relation detection via com" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4547", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pose visual cues from large language models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[25] Yunheng Li, ZhongYu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. arXiv preprint arXiv:2406.00670, 2024. 2", + "[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2", + "[27] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2", + "[28] Yue Liao, Si Liu, Fei Wang, Yanjie Chen, Chen Qian, and Ji-ashi Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482–490, 2020. 2", + "[29] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2, 3, 4, 5, 6, 7", + "[30] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2", + "[31] Xin Lin, Changxing Ding, Yibing Zhan, Zijian Li, and Dacheng Tao. Hl-net: Heterophily learning network for scene graph generation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19476-19485, 2022. 2", + "[32] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 2", + "[33] Xinpeng Liu, Yong-Lu Li, Xiaogian Wu, Yu-Wing Tai, Cewu Lu, and Chi-Keung Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113–20122, 2022. 6", + "[34] J Macqueen. Some methods for classification and analysis of multivariate observations. In Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press, 1967. 2, 4", + "[35] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. Advances in Neural Information Processing Systems, 36:45895-45906, 2023. 2" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Esteve Valls Mascaro, Daniel Sliwowski, and Dongheui Lee. Hoi4abot: Human-object interaction anticipation for human intention reading collaborative robots. arXiv preprint arXiv:2309.16524, 2023. 1", + "[37] Jean Massardi, Mathieu Gravel, and Éric Beaudry. Parc: A plan and activity recognition component for assistive robots. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3025-3031. IEEE, 2020. 1", + "[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3, 4", + "[39] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. 12mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3", + "[40] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23507-23517, 2023. 2, 3, 4, 5, 6", + "[41] Jeeseung Park, Jin-Woo Park, and Jong-Seok Lee. Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17152-17162, 2023. 2", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[43] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 2, 6", + "[44] Danyang Tu, Wei Sun, Guangtao Zhai, and Wei Shen. Agglomerative transformer for human-object interaction detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21614-21624, 2023. 2", + "[45] Mesut Erhan Unal and Adriana Kovashka. Weakly-supervised hoi detection from interaction labels only and language/vision-language priors. arXiv preprint arXiv:2303.05546, 2023. 3", + "[46] Suchen Wang, Kim-Hui Yap, Henghui Ding, Jiyan Wu, Jun-song Yuan, and Yap-Peng Tan. Discovering human interactions with large-vocabulary objects via query and multi-scale detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13475-13484, 2021. 2, 6", + "[47] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4548", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognexploitation (CVPR), pages 939-948, 2022. 2", + "[48] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 1, 3, 5, 6", + "[49] Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14647-14657, 2022. 2", + "[50] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[51] Nan Xi, Jingjing Meng, and Junsong Yuan. Open set video hoi detection from action-centric chain-of-look prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3079–3089, 2023. 1", + "[52] Jie Yang, Bingliang Li, Ailing Zeng, Lei Zhang, and Ruimao Zhang. Open-world human-object interaction detection via multi-modal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2024. 2, 3, 6, 7", + "[53] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. International Journal of Computer Vision, pages 1-19, 2024. 3", + "[54] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 6", + "[55] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 2, 6", + "[56] Frederic Z Zhang, Yuhui Yuan, Dylan Campbell, Zhuoyao Zhong, and Stephen Gould. Exploring predicate visual context in detecting of human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10411-10421, 2023. 2", + "[57] Sipeng Zheng, Boshen Xu, and Qin Jin. Open-category human-object interaction pre-training via language modeling framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19392-19402, 2023. 6", + "[58] Xubin Zhong, Xian Qu, Changxing Ding, and Dacheng Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 2", + "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In-" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 3", + "[60] Penghao Zhou and Mingmin Chi. Relation parsing neural network for human-object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 843-851, 2019. 2", + "[61] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2", + "[62] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 304 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4549", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_model.json b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32d51d6d76704bdf77cf7acadca9b9725a12a0a6 --- /dev/null +++ b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "SGC-Net: Stratified Granular Comparison Network for Open-Vocabulary HOI Detection" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.204, + 0.777, + 0.24 + ], + "angle": 0, + "content": "Xin Lin Chong Shi Zuopeng Yang* Haojin Tang* Zhili Zhou Guangzhou University" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.242, + 0.782, + 0.258 + ], + "angle": 0, + "content": "linxin94@gzhu.edu.cn, shichong@e.gzhu.edu.cn, yzpeng@gzhu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.26, + 0.684, + 0.275 + ], + "angle": 0, + "content": "tanghaojin@gzhu.edu.cn, zhou_zhili@163.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.342, + 0.485, + 0.719 + ], + "angle": 0, + "content": "Recent open-vocabulary human-object interaction (OV-HOI) detection methods primarily rely on large language model (LLM) for generating auxiliary descriptions and leverage knowledge distilled from CLIP to detect unseen interaction categories. Despite their effectiveness, these methods face two challenges: (1) feature granularity deficiency, due to reliance on last layer visual features for text alignment, leading to the neglect of crucial object-level details from intermediate layers; (2) semantic similarity confusion, resulting from CLIP's inherent biases toward certain classes, while LLM-generated descriptions based solely on labels fail to adequately capture inter-class similarities. To address these challenges, we propose a stratified granular comparison network. First, we introduce a granularity sensing alignment module that aggregates global semantic features with local details, refining interaction representations and ensuring robust alignment between intermediate visual features and text embeddings. Second, we develop a hierarchical group comparison module that recursively compares and groups classes using LLMs, generating fine-grained and discriminative descriptions for each interaction category. Experimental results on two widely-used benchmark datasets, SWIG-HOI and HICO-DET, demonstrate that our method achieves state-of-the-art results in OV-HOI detection. Codes is available at GitHub." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.747, + 0.224, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Human-Object Interaction (HOI) detection aims to localize human-object pairs and recognize their interactions, providing an efficient way for human-centric scene understanding. It plays a crucial role in various computer vision tasks, such as assistive robotics [36, 37] and video analysis [23, 51]. Recently, the emerging field of open-vocabulary HOI (OV-HOI) detection, which broadens the scope of HOI detection" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.309, + 0.9, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.409, + 0.805, + 0.421 + ], + "angle": 0, + "content": "(a) Feature granularity deficiency" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.422, + 0.9, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.501, + 0.805, + 0.513 + ], + "angle": 0, + "content": "(b) Semantic similarity confusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. (a) The last layer captures high-level global semantics but contains fewer low-level local details compared to intermediate layers. The red dot marks the selected patch. (b) Category-level and part-level descriptions overlook inter-class similarity, leading to difficulty in distinguishing semantically similar classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.642 + ], + "angle": 0, + "content": "by recognizing and associating objects beyond predefined categories, has gained increasing attention." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.869 + ], + "angle": 0, + "content": "OV-HOI detection has made notable progress, largely due to the success of vision-language models (VLMs) like CLIP [42]. Unlike zero-shot learning, which prohibits the access of any unannotated unseen classes during training, open-vocabulary learning leverages class name embeddings from VLMs for auxiliary supervision [50, 62]. Therefore, existing CLIP-based HOI detection methods naturally fall under OV-HOI detection. However, these methods face challenges in open-world scenarios because they rely on pretrained object detectors, and using category names as classifiers struggles to capture the variability of HOIs. To overcome these challenges, recent works [22, 48] have removed the need for pretrained detectors by extracting HOI features directly through image encoders while incorporating language priors from LLMs via text encoders." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Despite these advancements, current OV-HOI detection methods without pretrained object detectors still suffer from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4539" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.408 + ], + "angle": 0, + "content": "feature granularity deficiency. This issue arises because CLIP, trained with image-level alignment, produces globally aligned image-text features lacking the local detail needed for effective OV-HOI detection. As shown in Figure 1(a), the last layer focuses on high-level semantic features but neglects low-level local details like arm and face postures, compared to intermediate layers. Additionally, feature dissimilarity between shallow and deep layers increases with network depth, posing significant challenges in aligning intermediate visual features with text embeddings. Moreover, OV-HOI methods that rely on category-level or part-level text classifiers are prone to semantic confusion due to several inherent limitations. First, CLIP's training on large-scale, long-tail datasets tends to introduce biases towards certain classes [49]. For example, as shown on the left side of Figure 1(b), CLIP often confuses \"hug cat\" with \"hold cat.\" Second, descriptions generated by LLMs based solely on labels may fail to distinguish between semantically similar classes. As illustrated on the right side of Figure 1(b), LLMs produce similar part-level descriptions like \"arm extension\" for both \"hold cat\" and \"chase cat\"." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.745 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose an end-to-end OV-HOI detection network named SGC-Net. First, we design the granularity sensing alignment (GSA) module, leveraging multi-granularity visual features from CLIP to improve the transferability for novel classes. The primary challenge in this process is effectively aggregating global semantic features with local details to refine coarse interaction representations while ensuring alignment between intermediate visual features and text embeddings. Accordingly, we introduce a block partitioning strategy that groups layers based on their relative distances. Specifically, we divide the visual encoder of CLIP into several blocks, ensuring a little variation in the features within each block. To align local details more precisely, we perform a fusion of multi-granularity features within each block, assigning distinct and trainable Gaussian weights to each layer. Once these intra-block features are fused, we proceed to aggregate the fused features across all blocks to maintain global semantic consistency. Furthermore, to preserve the pretrained alignment between text and visual features in CLIP, we employ visual prompt tuning [7, 25, 61] by adding learnable tokens to the visual features of each layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Secondly, we propose the hierarchical group comparison (HGC) module to effectively differentiate semantically similar categories by recursively grouping and comparing classes via LLMs. Specifically, we utilize clustering algorithms (e.g., \\(K\\)-means [34]) for grouping. Even with extensive prior knowledge of LLMs, comparing numerous categories for OV-HOI detection remains challenging due to the quadratic growth of the description matrix with the number of categories. To address this challenge, we adopt group-specific comparison strategies to improve efficiency. For" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "smaller groups, we directly query LLMs for comparative descriptions, whereas for larger groups, we summarize the group's characteristics using LLMs and incorporate these summaries into new queries to highlight distinctive features of each class within the group. By recursively construct a class hierarchy, we can classify HOI by descending from the top to the bottom of the hierarchy, comparing HOI and text embeddings at each level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.364 + ], + "angle": 0, + "content": "In summary, the innovation of the proposed SGC-Net is three-fold: (1) The GSA module effectively aggregates multi-granular features while aligning the local details and global semantic features of OV-HOI; (2) The HGC module iteratively generates discriminative descriptions to refine the classification boundaries of labels; (3) The efficacy of the proposed SGC-Net is evaluated on two widely-used OV-HOI detection benchmark datasets, i.e., HICO-DET [5] and SWIG-HOI [46]. Experimental results show that our SGC-Net consistently outperforms state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.38, + 0.655, + 0.395 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.404, + 0.662, + 0.419 + ], + "angle": 0, + "content": "2.1. HOI Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.905, + 0.716 + ], + "angle": 0, + "content": "Existing HOI detection works can be divided into two categories, two-stage methods [3, 9, 21, 26, 32, 41, 43, 55, 56] and one-stage methods [6, 11, 18, 19, 28, 44, 58]. Two-stage methods typically use an object detector to recognize humans and objects, followed by specialized modules to associate humans with objects and identify their interactions, e.g., multi-steams [11, 12, 27], graphs [9, 30, 31, 54] or compositional learning [13, 14]. In contrast, the one-stage approach detects human-object pairs and their interactions simultaneously, without requiring stepwise processing. In particular, RLIP [60] proposes a pre-training strategy for HOI detection based on image captions. PPDM [28] reformulates the HOI detection task as a point detection and matching problem and achieves simultaneous object and interaction detection. However, existing HOI detection approaches are constrained by a closed-set assumption, which restricts their ability to recognize only predefined interaction categories. In contrast, our work aims to detect and recognize HOIs in the open-vocabulary scenario." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.898, + 0.742 + ], + "angle": 0, + "content": "2.2. CLIP-based Open Vocabulary HOI Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To alleviate the closed-set limitation, many studies [10, 22, 29, 35, 40, 47, 52] have been devoted to OV-HOI detection, aiming to identify both base and novel categories of HOIs while only base categories are available during training. Specifically, they transfer knowledge from the large-scale visual-linguistic pre-trained model CLIP to enhance interaction understanding. For instance, GEN-VLKTs [29] and HOICLIP [40] convert HOI labels into phrase descriptions to initialize the classifier and extract visual features from the CLIP image encoder to guide the interaction visual feature" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "learning. MP-HOI [52] incorporates the visual prompts into language-guided-only HOI detectors to enhance its generalization capability. THID [48] proposes a HOI sequence parser to detect multiple interactions. CMD-SE [22] detects HOIs at different distances using distinct feature maps. However, the adopted loss function in [22] involves minimizing differences between continuous and discrete variables therefore hard to optimize. Besides, aligning intermediate features with text embeddings may break the alignment in pre-trained CLIP due to the substantial differences. The proposed approach is a method free from pretrained object detectors but has the following advantages compared with existing works. First, it aggregates multi-granularity features from the visual encoder more appropriately. Second, it is easy to optimize and deploy to existing models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.423, + 0.342 + ], + "angle": 0, + "content": "2.3. Leverage LLM for visual classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.485, + 0.668 + ], + "angle": 0, + "content": "Recently, LLMs [1, 2, 8, 16] have made remarkable progress, revolutionizing natural language processing tasks with hundreds of billions of parameters and techniques such as reinforcement learning. Existing works [4, 17, 45, 59] have shown their effectiveness in generating comprehensive descriptions, particularly in classification and detection tasks. Specifically, Menon and Vondrick [38] generate textual descriptions directly from labels to assist VLMs in image classification. I2MVFormer [39] leverages LLM to generate multi-view document supervision for zero-shot image classification. ContextDET [53] utilizes contextual LLM tokens as conditional object queries to enhance the visual decoder for object detection. RECODE [24] leverages LLMs to generate descriptions for different components of relation categories. However, descriptions generated solely by LLMs from labels often tend to be generic and lack sufficient discriminability among semantically similar classes. To address this issue, our work leverages LLM to generate text descriptions by comparing different classes to reduce inter-class similarity and make the decision boundary of each class more compact." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.182, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This section provides details of the proposed stratified granular comparison network (SGC-Net). Specifically, we first describe the network architecture, followed by an explanation of the training and inference pipeline. As shown in Figure 2, SGC-Net comprises a granularity sensing alignment (GSA) module and a hierarchical group comparison (HGC) module. The GSA module enhances coarse HOI representations by utilizing distance-Aware Gaussian weighting (DGW) and visual prompt tuning to combine multi-granularity features from CLIP's image encoder. The HGC module refines decision boundaries for semantically similar classes by recursively comparing and grouping them using LLM-derived knowledge." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.791, + 0.108 + ], + "angle": 0, + "content": "3.1. Granularity Sensing Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.265 + ], + "angle": 0, + "content": "Existing OV-HOI methods [22, 29, 40] typically use features from the last layer of CLIP's image encoder to model HOIs. While these deep features effectively capture high-level semantics, they often lack the local details necessary for HOI detection. One approach to address this issue is to aggregate feature maps from different levels and feed them into an interaction decoder. However, the significant differences between shallow and deep features can disrupt the pre-trained vision-language alignment, ultimately reducing CLIP's zero-shot capability for OV-HOI tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.476 + ], + "angle": 0, + "content": "To address this issue, we propose the GSA module, which effectively captures fine-grained details in human-object interactions while preserving the vision-language alignment in the pre-trained CLIP model. Specifically, we first divide the CLIP's image encoder into \\(S\\) blocks, ensuring a little variation in the features within each block. Then, within each block, distinct distance-aware Gaussian weights are assigned to the transformer layers based on their relative neighborhood distances. These weights are trainable, enabling the model to adaptively learn layer- and block-specific information from the training data. Finally, we aggregate the features across different blocks. For each block containing \\(d\\) transformer layers, the aggregated feature \\(Z\\) can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.487, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\alpha_ {l} ^ {s} = \\exp \\left(- \\frac {1}{2} \\frac {(d - l) ^ {2}}{\\sigma^ {2}}\\right), \\quad l \\in [ 1, d ], \\\\ \\boldsymbol {Z} = \\sum_ {s = 1} ^ {S} \\alpha_ {s} \\left(\\sum_ {l = 1} ^ {d} \\alpha_ {l} ^ {s} \\boldsymbol {F} _ {l}\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Here, \\(l\\) represents the layer index. \\(\\pmb{F}_l\\) is a feature map from the \\(l\\)-th layer of CLIP image encoder. \\(\\alpha_l^s\\) and \\(\\alpha_s\\) denote the distance-aware Gaussian weights between layers within the same block and across different blocks, respectively. By adjusting variance parameter \\(\\sigma\\), the Eq. (1) can assign a higher weight to nearby layers and a lower weight to distant ones, facilitating more effective and flexible integration of features across different depth levels. Notably, the last layer of CLIP is treated as a separate block with a large weight. This design aims to keep the original visual-language correlations established in the pre-trained model. Different blocks can capture varying levels of granularity. Specifically, the early blocks focus more on fine-grained local details, while the later block emphasizes coarse-grained global features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To better align the intermediate visual features with the text embeddings, we employ visual prompt tuning [59] by introducing learnable tokens onto the visual features of each layer in the frozen encoder. During visual prompt tuning, the GSA module enables the gradients to be directly backpropagated to the middle layers of the visual encoder. It can promote the alignment of mid-layer features and text embeddings, enhancing the similarity across different layers." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Figure 2. The framework of SGC-Net. It eliminates the need for a pretrained object detector and includes two new modules for OV-HOI detection: (1) The GSA module partitions the CLIP visual encoder into blocks, aggregates features via DGW, and integrates an HOI decoder for fine-grained HOI representations. (2) The HGC module uses LLM to recursively construct class hierarchy, enabling HOI classification by traversing the hierarchy from top to bottom and comparing HOI representations with text embeddings at each level." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.501, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Compared with CMD-SE [22], we effectively aggregate multi-granularity features from the CLIP image encoder using trainable Gaussian weights, which have the following advantages. First, it allows intermediate layer visual features to complement last layer features while preserving the visual-language associations of the pre-trained CLIP. This approach addresses the issue in CMD-SE [22], where relying solely on discrete single-layer features for HOI modeling results in a loss of fine-grained information. Second, our method is easy to optimize. We eliminate the need to use loss functions that align discrete transformer layer indices with continuous human-object interaction distances, thereby avoiding the complex optimization processes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.697, + 0.484, + 0.729 + ], + "angle": 0, + "content": "Subsequently, following previous works [22, 29, 40], the final interaction representation \\(X\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.742, + 0.483, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} = \\operatorname {D e c} (\\boldsymbol {Q}, \\boldsymbol {Z}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.484, + 0.84 + ], + "angle": 0, + "content": "where \\( Q \\) refers to HOI queries, while \\( Z \\) is treated as both the key and value input to the HOI Decoder. The output \\( X \\) is later passed to the bounding box regressor and classifier to predict the bounding boxes of human-object pairs and their interaction categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.379, + 0.865 + ], + "angle": 0, + "content": "3.2. Hierarchical Group Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After obtaining the fine-grained interaction representation by the GSA module, designing a discriminative classifier re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.637 + ], + "angle": 0, + "content": "mains a challenging task. Existing methods [22, 29, 40] typically design classifiers with the manual prompt \"a photo of a person [verb] [object]\". However, these classifiers primarily rely on category names, often neglecting the contextual information provided by language. Recent research [22] has introduced the category descriptions produced by LLMs to enhance the model's generalization capability. However, these descriptions directly generated by LLMs from labels are often generic and semantically similar." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.909, + 0.759 + ], + "angle": 0, + "content": "To address this issue, we propose the hierarchical group comparison (HGC) module. Inspired by the coarse-to-fine approach humans use to recognize objects, HGC employs three strategies: Grouping, Comparison, and Hierarchical Classification. It identifies semantically similar descriptions and refines decision boundaries from generic to more discriminative. Specifically, we first utilize LLMs to generate initial descriptions for each HOI category." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.763, + 0.889, + 0.793 + ], + "angle": 0, + "content": "Q: What features are useful to distinguish {HOI category} in a photo?" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.794, + 0.555, + 0.805 + ], + "angle": 0, + "content": "A:-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Grouping: Following [38], we employ the pretrained CLIP text encoder to map the LLM-generated descriptions into the latent space. Utilizing these features, we apply a clustering algorithm, such as \\( K \\)-means [34], to perform grouping and identify semantic neighbors. Specifically, the value of \\( K \\) is calculated as the number of categories divided by a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.241 + ], + "angle": 0, + "content": "predefined grouping threshold \\( N \\). Even with extensive prior knowledge of LLMs, comparing a large number of categories presents a formidable challenge, as it needs to generate a comprehensive description matrix that grows quadratically with the number of categories. To address this issue, we employ tailored strategies for generating descriptions based on the group size, allowing for more efficient and effective comparisons. Specifically, we adopt summary-based comparison for larger groups and direct comparison for smaller ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.364 + ], + "angle": 0, + "content": "Comparing: For groups involving a large number of categories (i.e., exceeding half of the grouping threshold), we leverage the LLM to summarize their overall characteristics. By incorporating these summarized characteristics into a new query, we can generate a comparative description that captures the distinctive features and relationships among the categories. This approach enhances our ability to effectively compare and contrast the elements within a group." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.439, + 0.383 + ], + "angle": 0, + "content": "Q: Summarize the following interactions" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.385, + 0.41, + 0.4 + ], + "angle": 0, + "content": "with one sentence: {category list}?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.401, + 0.297, + 0.416 + ], + "angle": 0, + "content": "A: {subset description}" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.466, + 0.455 + ], + "angle": 0, + "content": "Q: What features are useful to distinguish {HOI category} from {subset description}?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.456, + 0.131, + 0.466 + ], + "angle": 0, + "content": "A:-" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.482, + 0.563 + ], + "angle": 0, + "content": "When the number of elements within a group is relatively small, a direct query to the LLM enables us to obtain an exceptionally comparative description. This approach capitalizes on the LLM's capabilities to provide a comprehensive and insightful analysis, enhancing the comparative understanding of the elements within the group." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.48, + 0.613 + ], + "angle": 0, + "content": "Q: What features are useful to distinguish {target category} from {other categories} in a photo?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.616, + 0.131, + 0.626 + ], + "angle": 0, + "content": "A:-" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.482, + 0.692 + ], + "angle": 0, + "content": "Utilizing these detailed and comparative descriptions, we can achieve more nuanced text embedding. Subsequently, we can continue to perform clustering, followed by additional comparisons to derive new descriptors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.844 + ], + "angle": 0, + "content": "Hierarchical Classification: After constructing the class hierarchy, we can obtain text features at multiple hierarchical levels. More specific, the hierarchical text features can be represented as \\(\\mathcal{T} = \\{D_1, D_2, \\dots, D_T\\}\\), where \\(T\\) is the number of test classes. \\(D_i \\in \\mathbb{R}^{M_j \\times C}\\) represents the hierarchical text embeddings for category \\(i\\). \\(C\\) is the embedding dimension. \\(M_j\\) is the number of comparative descriptions for category \\(i\\). Accordingly, the cosine similarity score between the HOI feature \\(\\pmb{x}\\) and the \\(j\\)-th description of the \\(i\\)-th HOI category, denoted as \\(D_i^j\\), can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.859, + 0.483, + 0.879 + ], + "angle": 0, + "content": "\\[\np _ {i} ^ {j} = \\boldsymbol {D} _ {i} ^ {j} \\cdot \\boldsymbol {x} ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, unreliable low-level descriptions can introduce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "errors and redundancies, degrading the quality of high-level descriptions and ultimately impairing the model's classification performance. To address this issue, we define an iterative evaluator \\( \\pmb{u}_i^k \\) to evaluate the acceptability of the current \\( k \\)-th level description for the \\( i \\)-th category as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.183, + 0.905, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {i} ^ {k} = \\mathbb {I} \\left(p _ {i} ^ {k + 1} > p _ {i} ^ {k} + \\tau\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.905, + 0.3 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes the tolerance parameter, \\(\\mathbb{I}\\) represents indicator function. Eq. (4) ensures that a new score is merged only if a subsequent discriminative description yields a higher score than the current one. Subsequently, we obtain the running average of the longest sequence of monotonically increasing \\(p\\) values as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.313, + 0.905, + 0.356 + ], + "angle": 0, + "content": "\\[\nr (\\boldsymbol {x}, i) = \\frac {p _ {i} ^ {1} + \\sum_ {j = 2} ^ {M _ {i}} p _ {i} ^ {j} \\prod_ {k = 1} ^ {j - 1} \\boldsymbol {u} _ {i} ^ {k}}{1 + \\sum_ {j = 2} ^ {M _ {i}} \\prod_ {k = 1} ^ {j - 1} \\boldsymbol {u} _ {i} ^ {k}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.905, + 0.413 + ], + "angle": 0, + "content": "Finally, the similarity score \\( s(\\pmb{x}, i) \\) between HOI feature \\( \\pmb{x} \\) and the \\( i \\)-th interaction label embedding is computed using a weighted average function, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.426, + 0.905, + 0.444 + ], + "angle": 0, + "content": "\\[\ns (\\boldsymbol {x}, i) = (1 - \\lambda) \\left(\\boldsymbol {p} _ {i} ^ {1} + \\boldsymbol {t} \\cdot \\boldsymbol {x} ^ {T}\\right) + \\lambda (r (\\boldsymbol {x}, i)). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.547 + ], + "angle": 0, + "content": "Here, we apply prompt tuning [48] with learnable text tokens \\( t \\) to learn the sentence format rather than using manually defined context words. Besides, our fusion method takes into account the score of the initial description \\( p_i^1 \\) as an offset, and we introduce a constant hyperparameter \\( \\lambda \\in [0, 1] \\) to balance the two terms." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.727, + 0.574 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.612 + ], + "angle": 0, + "content": "In this subsection, we elaborate on the processes of training and inference of our model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Training. During the training stage, we follow the query-based methods [22, 29, 48] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. The matching cost \\(\\mathcal{L}\\) for the matching process and the targeting cost for the training back-propagation share the same strategy, which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.731, + 0.905, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {b} \\sum_ {i \\in \\{h, o \\}} \\mathcal {L} _ {b} ^ {i} + \\lambda_ {i o u} \\sum_ {i \\in \\{h, o \\}} \\mathcal {L} _ {i o u} ^ {i} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.906, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_b\\), \\(\\mathcal{L}_{iou}\\), and \\(\\mathcal{L}_{cls}\\) denote the box regression, intersection over union, and classification losses, respectively. During the training stage, we follow the query-based methods [22, 29, 40] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. \\(\\lambda_b\\), \\(\\lambda_{iou}\\), and \\(\\lambda_{cls}\\) are the hyper-parameter weight." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.869, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Inference. For each HOI prediction, including the bounding-box pair \\((\\hat{b}_h^i,\\hat{b}_o^i)\\), the bounding box score \\(\\hat{c}_i\\) from" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4543" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "the box regressor, and the interaction score \\(\\hat{s}_i\\) from the interaction classifier, the final score \\(\\hat{s}_i'\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.134, + 0.483, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {i} ^ {\\prime} = \\hat {s} _ {i} \\cdot \\hat {c} _ {i} ^ {\\gamma} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.165, + 0.483, + 0.196 + ], + "angle": 0, + "content": "where \\(\\gamma > 1\\) is a constant used during inference to suppress overconfident objects [54, 55]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.21, + 0.216, + 0.228 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.236, + 0.292, + 0.254 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Datasets. Our experiments are conducted on two datasets: namely, SWIG-HOI [46] and HICO-DET [5]. The SWIG-HOI dataset provides diverse human interactions with large-vocabulary objects, comprising 400 human actions and 1,000 object categories. During the testing, we utilize approximately 5,500 interactions, including around 1,800 interactions that are not found in the training set. The annotations of HICO-DET include 600 combinations of 117 human actions and 80 objects. Among the 600 interactions, follow [22, 48], we simulate the open-vocabulary detection setting by holding out 120 rare interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.533 + ], + "angle": 0, + "content": "Evaluation Metric. Following in [5, 22, 29, 33, 48], we use the mean Average Precision (mAP) for evaluation. An HOI triplet prediction is classified as a true-positive example when the following criteria are satisfied: 1) The intersection over union of the human bounding box and object bounding box are larger than 0.5 w.r.t. the GT bounding boxes; 2) the predicted interaction category is accurate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Implementation Details. We follow the settings of previous work [22, 48] to build our model upon the pretrained CLIP. Specifically, for the visual encoder, we employ the ViT-B/16 version as our visual encoder and apply 12 learnable tokens in each layer to detect human-object interactions. For the text encoder, we introduce 8 prefix tokens and 4 conjunctive tokens to connect the words of human actions and objects for adaptively learning categories' information. Our model is optimized using AdamW with an initial learning rate of \\(10^{-4}\\), using 64 as the batch size for SWIG-HOI dataset and 32 for HICO-DET. We set the cost weights \\(\\lambda_{b}\\), \\(\\lambda_{cls}\\) and \\(\\lambda_{iou}\\) to 5, 2, and 5, respectively. The LLM we utilize is GPT-3.5. We set the hyperparameter \\(\\lambda\\) in Eq. (6) as 0.5. In all experiments, the variance parameter \\(\\sigma\\) is set to 1, and the tolerance parameter \\(\\tau\\) is set to 0." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.466, + 0.788 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "HICO-DET. To facilitate fair comparison, we first compare our model with state-of-the-art methods without pretrained detectors on the HICO-DET dataset. As shown in Table 1, SGC-Net outperforms CMD-SE [22] by \\(6.57\\%\\), \\(4.39\\%\\), and \\(4.87\\%\\) in mAP on the Unseen, Seen, and Full categories, respectively. Furthermore, even when compared to methods using pretrained detectors, SGC-Net achieves" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.253 + ], + "angle": 0, + "content": "
MethodPretrained DetectorUnseenSeenFull
SCL [15]✓19.0730.3928.08
GEN-VLKT [29]✓21.3632.9130.56
OpenCat [57]✓21.4633.8631.38
HOICLIP [40]✓23.4834.4732.26
THID [48]✗15.5324.3222.38
CMD-SE [22]✗16.7023.9522.35
SGC-Net✗23.2728.3427.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Table 1. Comparison of our proposed SGC-Net with state-of-the-art methods on the HICO-DET dataset." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.312, + 0.9, + 0.469 + ], + "angle": 0, + "content": "
MethodNon-rareRareUnseenFull
CHOID [46]10.936.632.646.64
QPIC [43]16.9510.846.2111.12
GEN-VLKT [29]20.9110.41-10.87
MP-HOI [52]20.2814.78-12.61
THID [48]17.6712.8210.0413.26
CMD-SE [22]21.4614.6410.7015.26
SGC-Net23.6716.5512.4617.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.483, + 0.905, + 0.51 + ], + "angle": 0, + "content": "Table 2. Comparison of our proposed SGC-Net with state-of-the-art methods on the SWIG-HOI dataset." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.528, + 0.875, + 0.634 + ], + "angle": 0, + "content": "
MethodNon-rareRareUnseenFull
Base15.6911.537.3211.45
+ GSA22.7416.0011.6416.49
+ HGC21.1814.1910.6914.81
SGC-Net23.6716.5512.4617.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.647, + 0.858, + 0.662 + ], + "angle": 0, + "content": "Table 3. Ablation studies of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.794 + ], + "angle": 0, + "content": "competitive performance. Although recent OV-HOI methods (e.g., GEN-VLKT [29], OpenCat [57], HOICLIP [40]) leverage CLIP text embeddings for interaction classification, they typically rely on a DETR architecture with pretrained weights. It is important to note that comparing our method with these approaches on HICO-DET dataset is not entirely fair, as the COCO dataset used for DETR pretraining shares the same object label space as HICO-DET." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "SWIG-HOI. Table 2 shows that SGC-Net outperforms all state-of-the-art methods on various setups. Specifically, SGC-Net outperforms CMD-SE [22] by \\(1.76\\%\\) and \\(1.91\\%\\) in mAP on the Rare and Unseen categories, respectively. Furthermore, SGC-Net outperforms the best OV-HOI method with a pretrained object detector by \\(4.59\\%\\) in mAP on the Full category. These results indicate that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4544" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.171 + ], + "angle": 0, + "content": "
Layer splitting mannerNon-rareRareUnseenFull
{4-6}, {7-9}, {10-12}23.0315.7211.3516.35
{6-8}, {9-10}, {11-12}23.6116.5211.6417.01
{6-8}, {9-11}, {12}23.6716.5512.4617.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.175, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Table 4. Effectiveness of different layer splitting manners. The elements in \\(\\{\\}\\) means the layer numbers that are used." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.213, + 0.485, + 0.359 + ], + "angle": 0, + "content": "
Number of blocksNon-rareRareUnseenFull
{12}21.1814.1910.0914.81
{9-11}, {12}21.9114.4810.0114.78
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{3-5}, {6-8}, {9-11}, {12}22.6115.2512.0116.53
Number of layers
{8-9}, {10-11}, {12}22.4515.1711.5515.98
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{4-7}, {8-11}, {12}23.5016.2112.1316.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.363, + 0.454, + 0.377 + ], + "angle": 0, + "content": "Table 5. Evaluation on the number of blocks and layers." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.391, + 0.465, + 0.53 + ], + "angle": 0, + "content": "
AggregationNon-rareRareUnseenFull
Self-attention21.3713.437.8813.90
Concat21.8415.5210.2215.68
Sum22.0914.899.5915.26
DGW23.4016.1411.6016.71
Sum*22.2115.1610.4515.64
DGW*23.6716.5512.4617.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Table 6. Comparison of different aggregation methods. * denotes that the weights are trainable." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.703 + ], + "angle": 0, + "content": "methods such as GEN-VLKT [29] and MP-HOI [52], which rely on pretrained detectors, perform suboptimally on the SWIG-HOI dataset. This is primarily because these methods struggle to scale effectively with vocabulary size, ultimately limiting their applicability in open-world scenarios. In contrast, SGC-Net overcomes this constraint by not relying on any detection pretraining, demonstrating superior capabilities in detecting and recognizing OV-HOI." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.245, + 0.728 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "To prove the effectiveness of our proposed methods, we conduct five ablation studies on the SWIG-HOI dataset. Results of the ablation studies are summarized in Table 3-6, and Figure 3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of the proposed modules. We first perform an ablation study to validate the effectiveness of the GSA and HGC modules. Results are summarized in Table 3. Our baseline model follows CMD-SE [22], which utilizes CLIP and removes the need for pretrained object detectors in OV-HOI detection. First, we apply the GSA module on the baseline model to improve its transferability for novel" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.861, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.277, + 0.886, + 0.292 + ], + "angle": 0, + "content": "Figure 3. Evaluation on the value of grouping threshold \\(N\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.492 + ], + "angle": 0, + "content": "classes, denoted as \\(+GSA\\). This modification achieves mAP improvements of \\(7.05\\%\\) for Non-Rare, \\(2.66\\%\\) for Rare, \\(4.32\\%\\) for Unseen, and \\(5.04\\%\\) for Full categories. Moreover, we integrate the HGC module with the baseline to enhance its ability to distinguish semantically similar interactions, denoted as \\(+HGC\\). The results indicate that the HGC module enables the baseline model to achieve \\(14.81\\%\\) in mAP for the Full categories. Finally, combining both modules results in the best performance, with an mAP of \\(12.46\\%\\) on the Unseen categories. These results demonstrate the significant potential of SGC-Net in enhancing interaction understanding within an open-vocabulary setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Effectiveness of the proposed layer splitting manner. The block partitioning strategy plays a crucial role in the GSA module due to its potential to preserve visual-language associations. Experimental results in Table 4 indicate that separating the last layer into an independent block is more effective than other combinations of partitioning strategies. This is because the last layer of CLIP's image encoder has the strongest association with text embeddings, and partitioning it into a separate block helps preserve this correlation while minimizing potential disruptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Number of blocks and number of layers in each block. To show the importance of multi-granularity feature fusion across different layers, we compare the performance of SGC-Net with varying numbers of blocks and layers per block. As shown in Table 5, increasing the number of blocks from 1 to 3 gradually improves the performance, particularly for Unseen interactions. The best performance is achieved with three blocks, each containing three layers. Reducing the number of layers to 2 degrades performance, primarily due to neglecting mid-layer features. It should be noted that the beginning layers are excluded, as they primarily encode features with limited semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Comparison of different aggregation methods. Due to the substantial differences between shallow and deep features, directly integrating multi-granularity features may disrupt the alignment of vision-language embeddings in the pre-trained CLIP model. To address this issue, we intro-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.498, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.091, + 0.904, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.305, + 0.908, + 0.349 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of SGC-Net. On the left-hand side, we visualize feature correspondence through cosine similarity calculations from both deep and shallow layers, and compare the predicted HOI categories. On the right-hand side, we show two inference examples and the absolute score gap between two images at each hierarchical level by querying the description with extracted HOI representations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.363, + 0.482, + 0.499 + ], + "angle": 0, + "content": "duce the DGW function, which mitigates these disruptions by accounting for the distances between layers. As shown in Table 6, DGW consistently outperforms direct aggregation strategies such as Sum, Concat. Additionally, setting the aggregation weights as learnable further enhances the performance of DGW. These results demonstrate that our method, by assigning smaller weights to features from more distant layers during multi-granularity fusion, effectively improves model performance in open-vocabulary tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.5, + 0.483, + 0.635 + ], + "angle": 0, + "content": "Grouping threshold \\( N \\). A large grouping threshold, such as \\( N = 8 \\), tends to produce rudimentary hierarchies and generalized descriptions. In contrast, smaller grouping thresholds increase the time required to construct the class hierarchy due to the more detailed recursive comparisons and groupings involved. As shown in Figure 3, the performance of SGC-Net improves as the grouping threshold increases. To balance computational efficiency and performance gains, we chose to set \\( N = 6 \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.274, + 0.665 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the proposed GSA module in leveraging CLIP's multi-granularity features, we present cosine similarity maps of features and qualitative OV-HOI detection results. At the left hand side of Figure 4, we show the patch similarity of Unseen classes not included in the training process. We observe that the intermediate layers of our method contain detailed information on local objects, including interaction patterns. Moreover, by leveraging these distinctive features, our SGC-Net improves HOI detection performance for both seen and unseen classes compared to using only last-layer features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "At the right hand side of Figure 4, we show two groups of examples that demonstrate the inference and absolute score gap calculation via the proposed HGC module. Specifically, we input two Unseen class images (② and ④) and two" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.56 + ], + "angle": 0, + "content": "Rare class images (① and ③) to evaluate their similarity with the hierarchical descriptions. The absolute score gap is computed as the absolute difference in similarity scores between paired images: i.e., ① vs. ② and ③ vs. ④. Due to the weak comparative information in the early descriptions, similar images tend to yield similar scores at the beginning. As we progress towards a more comparative description, the disparity between the scores of the two images gradually increases. Besides, this also highlights the interpretability of our method, as it provides the user insights regarding the attributes that elicit stronger responses to the input. Additionally, it indicates the specific layer of description at which the scores of different images diverge, widening the gap." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.58, + 0.634, + 0.595 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.788 + ], + "angle": 0, + "content": "In this paper, we propose the SGC-Net which enhances CLIP's transferability and discriminability to handle two critical issues in OV-HOI: feature granularity deficiency and semantic similarity confusion. Specifically, the GSA module enhances CLIP's image encoder by improving its ability to capture fine-grained HOI features while maintaining alignment between multi-granularity visual features and text embeddings. The HGC module strengthens CLIP's text encoder in distinguishing semantically related HOI categories by incorporating a grouping and comparison strategy into the LLM. Through extensive comparative experiments, we validate the effectiveness of SGC-Net on two datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.79, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Acknowledgment. This work is supported by the Guangzhou University Research Project (No: RQ2021013, 2025A03J3128), the Guangdong Natural Science Funds for Distinguished Young Scholars (No. 2023B1515020041), the Guangdong Basic and Applied Basic Research Foundation (No. 2023A1515110077), the NSF of China (No. 62372125)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4546" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.211 + ], + "angle": 0, + "content": "[1] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.215, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.245, + 0.484, + 0.326 + ], + "angle": 0, + "content": "[3] Yichao Cao, Qingfei Tang, Feng Yang, Xiu Su, Shan You, Xiaobo Lu, and Chang Xu. Re-min, learn and reason: Exploring the cross-modal semantic correlations for language-guided hoi detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23492-23503, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[4] Yichao Cao, Qingfei Tang, Xiu Su, Song Chen, Shan You, Xiaobo Lu, and Chang Xu. Detecting any human-object interaction relationship: Universal hoi detector with spatial prompt learning on foundation models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[5] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.527 + ], + "angle": 0, + "content": "[6] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004-9013, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[7] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, pages 11583-11592, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[9] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 696-712. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[10] Jianjun Gao, Kim-Hui Yap, Kejun Wu, Duc Tri Phan, Kratika Garg, and Boon Siew Han. Contextual human object interaction understanding from pre-trained large language model. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13436-13440. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[11] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Tanmay Gupta, Alexander Schwing, and Derek Hoiem. Nofrills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[13] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[14] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[16] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-imiL: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[17] Sheng Jin, Xueying Jiang, Jiaxing Huang, Lewei Lu, and Shijian Lu. Llms meet vlms: Boost open vocabulary object detection with fine-grained descriptors. arXiv preprint arXiv:2402.04630, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.449, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Bumsoo Kim, Taeho Choi, Jaewoo Kang, and Hyunwoo J Kim. Uniondet: Union-level detector towards real-time human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 498-514. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[19] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[20] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[21] Ting Lei, Fabian Caba, Qingchao Chen, Hailin Jin, Yuxin Peng, and Yang Liu. Efficient adaptive human-object interaction detection with concept-guided memory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6480-6490, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[22] Ting Lei, Shaofeng Yin, and Yang Liu. Exploring the potential of large foundation models for open-vocabulary hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16657-16667, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.869 + ], + "angle": 0, + "content": "[23] Huadong Li, Ying Wei, Shuailei Ma, Mingyu Chen, and Ge Li. Ripple transformer: A human-object interaction backbone and a new prediction strategy for smart surveillance devices. IEEE Transactions on Consumer Electronics, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Lin Li, Jun Xiao, Guikun Chen, Jian Shao, Yueting Zhuang, and Long Chen. Zero-shot visual relation detection via com" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4547" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "pose visual cues from large language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] Yunheng Li, ZhongYu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. arXiv preprint arXiv:2406.00670, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.345 + ], + "angle": 0, + "content": "[27] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[28] Yue Liao, Si Liu, Fei Wang, Yanjie Chen, Chen Qian, and Ji-ashi Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482–490, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[29] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[30] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[31] Xin Lin, Changxing Ding, Yibing Zhan, Zijian Li, and Dacheng Tao. Hl-net: Heterophily learning network for scene graph generation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19476-19485, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[32] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[33] Xinpeng Liu, Yong-Lu Li, Xiaogian Wu, Yu-Wing Tai, Cewu Lu, and Chi-Keung Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113–20122, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[34] J Macqueen. Some methods for classification and analysis of multivariate observations. In Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press, 1967. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. Advances in Neural Information Processing Systems, 36:45895-45906, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[36] Esteve Valls Mascaro, Daniel Sliwowski, and Dongheui Lee. Hoi4abot: Human-object interaction anticipation for human intention reading collaborative robots. arXiv preprint arXiv:2309.16524, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.206 + ], + "angle": 0, + "content": "[37] Jean Massardi, Mathieu Gravel, and Éric Beaudry. Parc: A plan and activity recognition component for assistive robots. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3025-3031. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[39] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. 12mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[40] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23507-23517, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.907, + 0.49 + ], + "angle": 0, + "content": "[41] Jeeseung Park, Jin-Woo Park, and Jong-Seok Lee. Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17152-17162, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[43] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[44] Danyang Tu, Wei Sun, Guangtao Zhai, and Wei Shen. Agglomerative transformer for human-object interaction detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21614-21624, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[45] Mesut Erhan Unal and Adriana Kovashka. Weakly-supervised hoi detection from interaction labels only and language/vision-language priors. arXiv preprint arXiv:2303.05546, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[46] Suchen Wang, Kim-Hui Yap, Henghui Ding, Jiyan Wu, Jun-song Yuan, and Yap-Peng Tan. Discovering human interactions with large-vocabulary objects via query and multi-scale detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13475-13484, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[47] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4548" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognexploitation (CVPR), pages 939-948, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[48] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[49] Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14647-14657, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[50] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[51] Nan Xi, Jingjing Meng, and Junsong Yuan. Open set video hoi detection from action-centric chain-of-look prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3079–3089, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[52] Jie Yang, Bingliang Li, Ailing Zeng, Lei Zhang, and Ruimao Zhang. Open-world human-object interaction detection via multi-modal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[53] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. International Journal of Computer Vision, pages 1-19, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.593 + ], + "angle": 0, + "content": "[54] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[55] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[56] Frederic Z Zhang, Yuhui Yuan, Dylan Campbell, Zhuoyao Zhong, and Stephen Gould. Exploring predicate visual context in detecting of human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10411-10421, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[57] Sipeng Zheng, Boshen Xu, and Qin Jin. Open-category human-object interaction pre-training via language modeling framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19392-19402, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[58] Xubin Zhong, Xian Qu, Changxing Ding, and Dacheng Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[60] Penghao Zhou and Mingmin Chi. Relation parsing neural network for human-object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 843-851, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[61] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[62] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4549" + } + ] +] \ No newline at end of file diff --git a/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_origin.pdf b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7431b70c1e2d6c54518ccf0fa5fc3375bfdd272e --- /dev/null +++ b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/de53cd58-a64d-437d-bffd-56e874a1fc12_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82094a7af6abd5a5c4fc4ab873c7a8d7ca16575e25789ad7f6097daabaf56204 +size 1422559 diff --git a/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/full.md b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d8de098788586607f1c74a3edb056789b578ea7 --- /dev/null +++ b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/full.md @@ -0,0 +1,329 @@ +# SGC-Net: Stratified Granular Comparison Network for Open-Vocabulary HOI Detection + +Xin Lin Chong Shi Zuopeng Yang* Haojin Tang* Zhili Zhou Guangzhou University + +linxin94@gzhu.edu.cn, shichong@e.gzhu.edu.cn, yzpeng@gzhu.edu.cn, + +tanghaojin@gzhu.edu.cn, zhou_zhili@163.com + +# Abstract + +Recent open-vocabulary human-object interaction (OV-HOI) detection methods primarily rely on large language model (LLM) for generating auxiliary descriptions and leverage knowledge distilled from CLIP to detect unseen interaction categories. Despite their effectiveness, these methods face two challenges: (1) feature granularity deficiency, due to reliance on last layer visual features for text alignment, leading to the neglect of crucial object-level details from intermediate layers; (2) semantic similarity confusion, resulting from CLIP's inherent biases toward certain classes, while LLM-generated descriptions based solely on labels fail to adequately capture inter-class similarities. To address these challenges, we propose a stratified granular comparison network. First, we introduce a granularity sensing alignment module that aggregates global semantic features with local details, refining interaction representations and ensuring robust alignment between intermediate visual features and text embeddings. Second, we develop a hierarchical group comparison module that recursively compares and groups classes using LLMs, generating fine-grained and discriminative descriptions for each interaction category. Experimental results on two widely-used benchmark datasets, SWIG-HOI and HICO-DET, demonstrate that our method achieves state-of-the-art results in OV-HOI detection. Codes is available at GitHub. + +# 1. Introduction + +Human-Object Interaction (HOI) detection aims to localize human-object pairs and recognize their interactions, providing an efficient way for human-centric scene understanding. It plays a crucial role in various computer vision tasks, such as assistive robotics [36, 37] and video analysis [23, 51]. Recently, the emerging field of open-vocabulary HOI (OV-HOI) detection, which broadens the scope of HOI detection + +![](images/f4ca9d7b9568999a5a5b647785ff2e13d338e679aef1f0b132c036eff939a3c9.jpg) +(a) Feature granularity deficiency + +![](images/1142b082fad16054671e9ff7fe0c7d6b0c2565d355c28f2564464214bb83905b.jpg) +(b) Semantic similarity confusion +Figure 1. (a) The last layer captures high-level global semantics but contains fewer low-level local details compared to intermediate layers. The red dot marks the selected patch. (b) Category-level and part-level descriptions overlook inter-class similarity, leading to difficulty in distinguishing semantically similar classes. + +by recognizing and associating objects beyond predefined categories, has gained increasing attention. + +OV-HOI detection has made notable progress, largely due to the success of vision-language models (VLMs) like CLIP [42]. Unlike zero-shot learning, which prohibits the access of any unannotated unseen classes during training, open-vocabulary learning leverages class name embeddings from VLMs for auxiliary supervision [50, 62]. Therefore, existing CLIP-based HOI detection methods naturally fall under OV-HOI detection. However, these methods face challenges in open-world scenarios because they rely on pretrained object detectors, and using category names as classifiers struggles to capture the variability of HOIs. To overcome these challenges, recent works [22, 48] have removed the need for pretrained detectors by extracting HOI features directly through image encoders while incorporating language priors from LLMs via text encoders. + +Despite these advancements, current OV-HOI detection methods without pretrained object detectors still suffer from + +feature granularity deficiency. This issue arises because CLIP, trained with image-level alignment, produces globally aligned image-text features lacking the local detail needed for effective OV-HOI detection. As shown in Figure 1(a), the last layer focuses on high-level semantic features but neglects low-level local details like arm and face postures, compared to intermediate layers. Additionally, feature dissimilarity between shallow and deep layers increases with network depth, posing significant challenges in aligning intermediate visual features with text embeddings. Moreover, OV-HOI methods that rely on category-level or part-level text classifiers are prone to semantic confusion due to several inherent limitations. First, CLIP's training on large-scale, long-tail datasets tends to introduce biases towards certain classes [49]. For example, as shown on the left side of Figure 1(b), CLIP often confuses "hug cat" with "hold cat." Second, descriptions generated by LLMs based solely on labels may fail to distinguish between semantically similar classes. As illustrated on the right side of Figure 1(b), LLMs produce similar part-level descriptions like "arm extension" for both "hold cat" and "chase cat". + +To address the aforementioned challenges, we propose an end-to-end OV-HOI detection network named SGC-Net. First, we design the granularity sensing alignment (GSA) module, leveraging multi-granularity visual features from CLIP to improve the transferability for novel classes. The primary challenge in this process is effectively aggregating global semantic features with local details to refine coarse interaction representations while ensuring alignment between intermediate visual features and text embeddings. Accordingly, we introduce a block partitioning strategy that groups layers based on their relative distances. Specifically, we divide the visual encoder of CLIP into several blocks, ensuring a little variation in the features within each block. To align local details more precisely, we perform a fusion of multi-granularity features within each block, assigning distinct and trainable Gaussian weights to each layer. Once these intra-block features are fused, we proceed to aggregate the fused features across all blocks to maintain global semantic consistency. Furthermore, to preserve the pretrained alignment between text and visual features in CLIP, we employ visual prompt tuning [7, 25, 61] by adding learnable tokens to the visual features of each layer. + +Secondly, we propose the hierarchical group comparison (HGC) module to effectively differentiate semantically similar categories by recursively grouping and comparing classes via LLMs. Specifically, we utilize clustering algorithms (e.g., $K$ -means [34]) for grouping. Even with extensive prior knowledge of LLMs, comparing numerous categories for OV-HOI detection remains challenging due to the quadratic growth of the description matrix with the number of categories. To address this challenge, we adopt group-specific comparison strategies to improve efficiency. For + +smaller groups, we directly query LLMs for comparative descriptions, whereas for larger groups, we summarize the group's characteristics using LLMs and incorporate these summaries into new queries to highlight distinctive features of each class within the group. By recursively construct a class hierarchy, we can classify HOI by descending from the top to the bottom of the hierarchy, comparing HOI and text embeddings at each level. + +In summary, the innovation of the proposed SGC-Net is three-fold: (1) The GSA module effectively aggregates multi-granular features while aligning the local details and global semantic features of OV-HOI; (2) The HGC module iteratively generates discriminative descriptions to refine the classification boundaries of labels; (3) The efficacy of the proposed SGC-Net is evaluated on two widely-used OV-HOI detection benchmark datasets, i.e., HICO-DET [5] and SWIG-HOI [46]. Experimental results show that our SGC-Net consistently outperforms state-of-the-art methods. + +# 2. Related Work + +# 2.1. HOI Detection + +Existing HOI detection works can be divided into two categories, two-stage methods [3, 9, 21, 26, 32, 41, 43, 55, 56] and one-stage methods [6, 11, 18, 19, 28, 44, 58]. Two-stage methods typically use an object detector to recognize humans and objects, followed by specialized modules to associate humans with objects and identify their interactions, e.g., multi-steams [11, 12, 27], graphs [9, 30, 31, 54] or compositional learning [13, 14]. In contrast, the one-stage approach detects human-object pairs and their interactions simultaneously, without requiring stepwise processing. In particular, RLIP [60] proposes a pre-training strategy for HOI detection based on image captions. PPDM [28] reformulates the HOI detection task as a point detection and matching problem and achieves simultaneous object and interaction detection. However, existing HOI detection approaches are constrained by a closed-set assumption, which restricts their ability to recognize only predefined interaction categories. In contrast, our work aims to detect and recognize HOIs in the open-vocabulary scenario. + +# 2.2. CLIP-based Open Vocabulary HOI Detection + +To alleviate the closed-set limitation, many studies [10, 22, 29, 35, 40, 47, 52] have been devoted to OV-HOI detection, aiming to identify both base and novel categories of HOIs while only base categories are available during training. Specifically, they transfer knowledge from the large-scale visual-linguistic pre-trained model CLIP to enhance interaction understanding. For instance, GEN-VLKTs [29] and HOICLIP [40] convert HOI labels into phrase descriptions to initialize the classifier and extract visual features from the CLIP image encoder to guide the interaction visual feature + +learning. MP-HOI [52] incorporates the visual prompts into language-guided-only HOI detectors to enhance its generalization capability. THID [48] proposes a HOI sequence parser to detect multiple interactions. CMD-SE [22] detects HOIs at different distances using distinct feature maps. However, the adopted loss function in [22] involves minimizing differences between continuous and discrete variables therefore hard to optimize. Besides, aligning intermediate features with text embeddings may break the alignment in pre-trained CLIP due to the substantial differences. The proposed approach is a method free from pretrained object detectors but has the following advantages compared with existing works. First, it aggregates multi-granularity features from the visual encoder more appropriately. Second, it is easy to optimize and deploy to existing models. + +# 2.3. Leverage LLM for visual classification + +Recently, LLMs [1, 2, 8, 16] have made remarkable progress, revolutionizing natural language processing tasks with hundreds of billions of parameters and techniques such as reinforcement learning. Existing works [4, 17, 45, 59] have shown their effectiveness in generating comprehensive descriptions, particularly in classification and detection tasks. Specifically, Menon and Vondrick [38] generate textual descriptions directly from labels to assist VLMs in image classification. I2MVFormer [39] leverages LLM to generate multi-view document supervision for zero-shot image classification. ContextDET [53] utilizes contextual LLM tokens as conditional object queries to enhance the visual decoder for object detection. RECODE [24] leverages LLMs to generate descriptions for different components of relation categories. However, descriptions generated solely by LLMs from labels often tend to be generic and lack sufficient discriminability among semantically similar classes. To address this issue, our work leverages LLM to generate text descriptions by comparing different classes to reduce inter-class similarity and make the decision boundary of each class more compact. + +# 3. Method + +This section provides details of the proposed stratified granular comparison network (SGC-Net). Specifically, we first describe the network architecture, followed by an explanation of the training and inference pipeline. As shown in Figure 2, SGC-Net comprises a granularity sensing alignment (GSA) module and a hierarchical group comparison (HGC) module. The GSA module enhances coarse HOI representations by utilizing distance-Aware Gaussian weighting (DGW) and visual prompt tuning to combine multi-granularity features from CLIP's image encoder. The HGC module refines decision boundaries for semantically similar classes by recursively comparing and grouping them using LLM-derived knowledge. + +# 3.1. Granularity Sensing Alignment + +Existing OV-HOI methods [22, 29, 40] typically use features from the last layer of CLIP's image encoder to model HOIs. While these deep features effectively capture high-level semantics, they often lack the local details necessary for HOI detection. One approach to address this issue is to aggregate feature maps from different levels and feed them into an interaction decoder. However, the significant differences between shallow and deep features can disrupt the pre-trained vision-language alignment, ultimately reducing CLIP's zero-shot capability for OV-HOI tasks. + +To address this issue, we propose the GSA module, which effectively captures fine-grained details in human-object interactions while preserving the vision-language alignment in the pre-trained CLIP model. Specifically, we first divide the CLIP's image encoder into $S$ blocks, ensuring a little variation in the features within each block. Then, within each block, distinct distance-aware Gaussian weights are assigned to the transformer layers based on their relative neighborhood distances. These weights are trainable, enabling the model to adaptively learn layer- and block-specific information from the training data. Finally, we aggregate the features across different blocks. For each block containing $d$ transformer layers, the aggregated feature $Z$ can be represented as follows: + +$$ +\begin{array}{l} \alpha_ {l} ^ {s} = \exp \left(- \frac {1}{2} \frac {(d - l) ^ {2}}{\sigma^ {2}}\right), \quad l \in [ 1, d ], \\ \boldsymbol {Z} = \sum_ {s = 1} ^ {S} \alpha_ {s} \left(\sum_ {l = 1} ^ {d} \alpha_ {l} ^ {s} \boldsymbol {F} _ {l}\right) \tag {1} \\ \end{array} +$$ + +Here, $l$ represents the layer index. $\pmb{F}_l$ is a feature map from the $l$ -th layer of CLIP image encoder. $\alpha_l^s$ and $\alpha_s$ denote the distance-aware Gaussian weights between layers within the same block and across different blocks, respectively. By adjusting variance parameter $\sigma$ , the Eq. (1) can assign a higher weight to nearby layers and a lower weight to distant ones, facilitating more effective and flexible integration of features across different depth levels. Notably, the last layer of CLIP is treated as a separate block with a large weight. This design aims to keep the original visual-language correlations established in the pre-trained model. Different blocks can capture varying levels of granularity. Specifically, the early blocks focus more on fine-grained local details, while the later block emphasizes coarse-grained global features. + +To better align the intermediate visual features with the text embeddings, we employ visual prompt tuning [59] by introducing learnable tokens onto the visual features of each layer in the frozen encoder. During visual prompt tuning, the GSA module enables the gradients to be directly backpropagated to the middle layers of the visual encoder. It can promote the alignment of mid-layer features and text embeddings, enhancing the similarity across different layers. + +![](images/741a6c5728444953f1fa0a0b475ab9f1e19b064463d0935e07a47f15dcf36b33.jpg) +Figure 2. The framework of SGC-Net. It eliminates the need for a pretrained object detector and includes two new modules for OV-HOI detection: (1) The GSA module partitions the CLIP visual encoder into blocks, aggregates features via DGW, and integrates an HOI decoder for fine-grained HOI representations. (2) The HGC module uses LLM to recursively construct class hierarchy, enabling HOI classification by traversing the hierarchy from top to bottom and comparing HOI representations with text embeddings at each level. + +Compared with CMD-SE [22], we effectively aggregate multi-granularity features from the CLIP image encoder using trainable Gaussian weights, which have the following advantages. First, it allows intermediate layer visual features to complement last layer features while preserving the visual-language associations of the pre-trained CLIP. This approach addresses the issue in CMD-SE [22], where relying solely on discrete single-layer features for HOI modeling results in a loss of fine-grained information. Second, our method is easy to optimize. We eliminate the need to use loss functions that align discrete transformer layer indices with continuous human-object interaction distances, thereby avoiding the complex optimization processes. + +Subsequently, following previous works [22, 29, 40], the final interaction representation $X$ is formulated as follows: + +$$ +\boldsymbol {X} = \operatorname {D e c} (\boldsymbol {Q}, \boldsymbol {Z}), \tag {2} +$$ + +where $Q$ refers to HOI queries, while $Z$ is treated as both the key and value input to the HOI Decoder. The output $X$ is later passed to the bounding box regressor and classifier to predict the bounding boxes of human-object pairs and their interaction categories. + +# 3.2. Hierarchical Group Comparison + +After obtaining the fine-grained interaction representation by the GSA module, designing a discriminative classifier re + +mains a challenging task. Existing methods [22, 29, 40] typically design classifiers with the manual prompt "a photo of a person [verb] [object]". However, these classifiers primarily rely on category names, often neglecting the contextual information provided by language. Recent research [22] has introduced the category descriptions produced by LLMs to enhance the model's generalization capability. However, these descriptions directly generated by LLMs from labels are often generic and semantically similar. + +To address this issue, we propose the hierarchical group comparison (HGC) module. Inspired by the coarse-to-fine approach humans use to recognize objects, HGC employs three strategies: Grouping, Comparison, and Hierarchical Classification. It identifies semantically similar descriptions and refines decision boundaries from generic to more discriminative. Specifically, we first utilize LLMs to generate initial descriptions for each HOI category. + +Q: What features are useful to distinguish {HOI category} in a photo? + +A:- + +Grouping: Following [38], we employ the pretrained CLIP text encoder to map the LLM-generated descriptions into the latent space. Utilizing these features, we apply a clustering algorithm, such as $K$ -means [34], to perform grouping and identify semantic neighbors. Specifically, the value of $K$ is calculated as the number of categories divided by a + +predefined grouping threshold $N$ . Even with extensive prior knowledge of LLMs, comparing a large number of categories presents a formidable challenge, as it needs to generate a comprehensive description matrix that grows quadratically with the number of categories. To address this issue, we employ tailored strategies for generating descriptions based on the group size, allowing for more efficient and effective comparisons. Specifically, we adopt summary-based comparison for larger groups and direct comparison for smaller ones. + +Comparing: For groups involving a large number of categories (i.e., exceeding half of the grouping threshold), we leverage the LLM to summarize their overall characteristics. By incorporating these summarized characteristics into a new query, we can generate a comparative description that captures the distinctive features and relationships among the categories. This approach enhances our ability to effectively compare and contrast the elements within a group. + +Q: Summarize the following interactions + +with one sentence: {category list}? + +A: {subset description} + +Q: What features are useful to distinguish {HOI category} from {subset description}? + +A:- + +When the number of elements within a group is relatively small, a direct query to the LLM enables us to obtain an exceptionally comparative description. This approach capitalizes on the LLM's capabilities to provide a comprehensive and insightful analysis, enhancing the comparative understanding of the elements within the group. + +Q: What features are useful to distinguish {target category} from {other categories} in a photo? + +A:- + +Utilizing these detailed and comparative descriptions, we can achieve more nuanced text embedding. Subsequently, we can continue to perform clustering, followed by additional comparisons to derive new descriptors. + +Hierarchical Classification: After constructing the class hierarchy, we can obtain text features at multiple hierarchical levels. More specific, the hierarchical text features can be represented as $\mathcal{T} = \{D_1, D_2, \dots, D_T\}$ , where $T$ is the number of test classes. $D_i \in \mathbb{R}^{M_j \times C}$ represents the hierarchical text embeddings for category $i$ . $C$ is the embedding dimension. $M_j$ is the number of comparative descriptions for category $i$ . Accordingly, the cosine similarity score between the HOI feature $\pmb{x}$ and the $j$ -th description of the $i$ -th HOI category, denoted as $D_i^j$ , can be expressed as follows: + +$$ +p _ {i} ^ {j} = \boldsymbol {D} _ {i} ^ {j} \cdot \boldsymbol {x} ^ {T}. \tag {3} +$$ + +However, unreliable low-level descriptions can introduce + +errors and redundancies, degrading the quality of high-level descriptions and ultimately impairing the model's classification performance. To address this issue, we define an iterative evaluator $\pmb{u}_i^k$ to evaluate the acceptability of the current $k$ -th level description for the $i$ -th category as follows: + +$$ +\boldsymbol {u} _ {i} ^ {k} = \mathbb {I} \left(p _ {i} ^ {k + 1} > p _ {i} ^ {k} + \tau\right), \tag {4} +$$ + +where $\tau$ denotes the tolerance parameter, $\mathbb{I}$ represents indicator function. Eq. (4) ensures that a new score is merged only if a subsequent discriminative description yields a higher score than the current one. Subsequently, we obtain the running average of the longest sequence of monotonically increasing $p$ values as follows: + +$$ +r (\boldsymbol {x}, i) = \frac {p _ {i} ^ {1} + \sum_ {j = 2} ^ {M _ {i}} p _ {i} ^ {j} \prod_ {k = 1} ^ {j - 1} \boldsymbol {u} _ {i} ^ {k}}{1 + \sum_ {j = 2} ^ {M _ {i}} \prod_ {k = 1} ^ {j - 1} \boldsymbol {u} _ {i} ^ {k}}. \tag {5} +$$ + +Finally, the similarity score $s(\pmb{x}, i)$ between HOI feature $\pmb{x}$ and the $i$ -th interaction label embedding is computed using a weighted average function, as follows: + +$$ +s (\boldsymbol {x}, i) = (1 - \lambda) \left(\boldsymbol {p} _ {i} ^ {1} + \boldsymbol {t} \cdot \boldsymbol {x} ^ {T}\right) + \lambda (r (\boldsymbol {x}, i)). \tag {6} +$$ + +Here, we apply prompt tuning [48] with learnable text tokens $t$ to learn the sentence format rather than using manually defined context words. Besides, our fusion method takes into account the score of the initial description $p_i^1$ as an offset, and we introduce a constant hyperparameter $\lambda \in [0, 1]$ to balance the two terms. + +# 3.3. Training and Inference + +In this subsection, we elaborate on the processes of training and inference of our model. + +Training. During the training stage, we follow the query-based methods [22, 29, 48] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. The matching cost $\mathcal{L}$ for the matching process and the targeting cost for the training back-propagation share the same strategy, which is formulated as follows: + +$$ +\mathcal {L} = \lambda_ {b} \sum_ {i \in \{h, o \}} \mathcal {L} _ {b} ^ {i} + \lambda_ {i o u} \sum_ {i \in \{h, o \}} \mathcal {L} _ {i o u} ^ {i} + \lambda_ {c l s} \mathcal {L} _ {c l s}, \tag {7} +$$ + +where $\mathcal{L}_b$ , $\mathcal{L}_{iou}$ , and $\mathcal{L}_{cls}$ denote the box regression, intersection over union, and classification losses, respectively. During the training stage, we follow the query-based methods [22, 29, 40] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. $\lambda_b$ , $\lambda_{iou}$ , and $\lambda_{cls}$ are the hyper-parameter weight. + +Inference. For each HOI prediction, including the bounding-box pair $(\hat{b}_h^i,\hat{b}_o^i)$ , the bounding box score $\hat{c}_i$ from + +the box regressor, and the interaction score $\hat{s}_i$ from the interaction classifier, the final score $\hat{s}_i'$ is computed as: + +$$ +\hat {s} _ {i} ^ {\prime} = \hat {s} _ {i} \cdot \hat {c} _ {i} ^ {\gamma} \tag {8} +$$ + +where $\gamma > 1$ is a constant used during inference to suppress overconfident objects [54, 55]. + +# 4. Experiment + +# 4.1. Experimental Setting + +Datasets. Our experiments are conducted on two datasets: namely, SWIG-HOI [46] and HICO-DET [5]. The SWIG-HOI dataset provides diverse human interactions with large-vocabulary objects, comprising 400 human actions and 1,000 object categories. During the testing, we utilize approximately 5,500 interactions, including around 1,800 interactions that are not found in the training set. The annotations of HICO-DET include 600 combinations of 117 human actions and 80 objects. Among the 600 interactions, follow [22, 48], we simulate the open-vocabulary detection setting by holding out 120 rare interactions. + +Evaluation Metric. Following in [5, 22, 29, 33, 48], we use the mean Average Precision (mAP) for evaluation. An HOI triplet prediction is classified as a true-positive example when the following criteria are satisfied: 1) The intersection over union of the human bounding box and object bounding box are larger than 0.5 w.r.t. the GT bounding boxes; 2) the predicted interaction category is accurate. + +Implementation Details. We follow the settings of previous work [22, 48] to build our model upon the pretrained CLIP. Specifically, for the visual encoder, we employ the ViT-B/16 version as our visual encoder and apply 12 learnable tokens in each layer to detect human-object interactions. For the text encoder, we introduce 8 prefix tokens and 4 conjunctive tokens to connect the words of human actions and objects for adaptively learning categories' information. Our model is optimized using AdamW with an initial learning rate of $10^{-4}$ , using 64 as the batch size for SWIG-HOI dataset and 32 for HICO-DET. We set the cost weights $\lambda_{b}$ , $\lambda_{cls}$ and $\lambda_{iou}$ to 5, 2, and 5, respectively. The LLM we utilize is GPT-3.5. We set the hyperparameter $\lambda$ in Eq. (6) as 0.5. In all experiments, the variance parameter $\sigma$ is set to 1, and the tolerance parameter $\tau$ is set to 0. + +# 4.2. Comparisons with State-of-the-Art Methods + +HICO-DET. To facilitate fair comparison, we first compare our model with state-of-the-art methods without pretrained detectors on the HICO-DET dataset. As shown in Table 1, SGC-Net outperforms CMD-SE [22] by $6.57\%$ , $4.39\%$ , and $4.87\%$ in mAP on the Unseen, Seen, and Full categories, respectively. Furthermore, even when compared to methods using pretrained detectors, SGC-Net achieves + +
MethodPretrained DetectorUnseenSeenFull
SCL [15]✓19.0730.3928.08
GEN-VLKT [29]✓21.3632.9130.56
OpenCat [57]✓21.4633.8631.38
HOICLIP [40]✓23.4834.4732.26
THID [48]✗15.5324.3222.38
CMD-SE [22]✗16.7023.9522.35
SGC-Net✗23.2728.3427.22
+ +Table 1. Comparison of our proposed SGC-Net with state-of-the-art methods on the HICO-DET dataset. + +
MethodNon-rareRareUnseenFull
CHOID [46]10.936.632.646.64
QPIC [43]16.9510.846.2111.12
GEN-VLKT [29]20.9110.41-10.87
MP-HOI [52]20.2814.78-12.61
THID [48]17.6712.8210.0413.26
CMD-SE [22]21.4614.6410.7015.26
SGC-Net23.6716.5512.4617.20
+ +Table 2. Comparison of our proposed SGC-Net with state-of-the-art methods on the SWIG-HOI dataset. + +
MethodNon-rareRareUnseenFull
Base15.6911.537.3211.45
+ GSA22.7416.0011.6416.49
+ HGC21.1814.1910.6914.81
SGC-Net23.6716.5512.4617.20
+ +Table 3. Ablation studies of the proposed method. + +competitive performance. Although recent OV-HOI methods (e.g., GEN-VLKT [29], OpenCat [57], HOICLIP [40]) leverage CLIP text embeddings for interaction classification, they typically rely on a DETR architecture with pretrained weights. It is important to note that comparing our method with these approaches on HICO-DET dataset is not entirely fair, as the COCO dataset used for DETR pretraining shares the same object label space as HICO-DET. + +SWIG-HOI. Table 2 shows that SGC-Net outperforms all state-of-the-art methods on various setups. Specifically, SGC-Net outperforms CMD-SE [22] by $1.76\%$ and $1.91\%$ in mAP on the Rare and Unseen categories, respectively. Furthermore, SGC-Net outperforms the best OV-HOI method with a pretrained object detector by $4.59\%$ in mAP on the Full category. These results indicate that + +
Layer splitting mannerNon-rareRareUnseenFull
{4-6}, {7-9}, {10-12}23.0315.7211.3516.35
{6-8}, {9-10}, {11-12}23.6116.5211.6417.01
{6-8}, {9-11}, {12}23.6716.5512.4617.20
+ +Table 4. Effectiveness of different layer splitting manners. The elements in $\{\}$ means the layer numbers that are used. + +
Number of blocksNon-rareRareUnseenFull
{12}21.1814.1910.0914.81
{9-11}, {12}21.9114.4810.0114.78
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{3-5}, {6-8}, {9-11}, {12}22.6115.2512.0116.53
Number of layers
{8-9}, {10-11}, {12}22.4515.1711.5515.98
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{4-7}, {8-11}, {12}23.5016.2112.1316.75
+ +Table 5. Evaluation on the number of blocks and layers. + +
AggregationNon-rareRareUnseenFull
Self-attention21.3713.437.8813.90
Concat21.8415.5210.2215.68
Sum22.0914.899.5915.26
DGW23.4016.1411.6016.71
Sum*22.2115.1610.4515.64
DGW*23.6716.5512.4617.20
+ +Table 6. Comparison of different aggregation methods. * denotes that the weights are trainable. + +methods such as GEN-VLKT [29] and MP-HOI [52], which rely on pretrained detectors, perform suboptimally on the SWIG-HOI dataset. This is primarily because these methods struggle to scale effectively with vocabulary size, ultimately limiting their applicability in open-world scenarios. In contrast, SGC-Net overcomes this constraint by not relying on any detection pretraining, demonstrating superior capabilities in detecting and recognizing OV-HOI. + +# 4.3. Ablation Study + +To prove the effectiveness of our proposed methods, we conduct five ablation studies on the SWIG-HOI dataset. Results of the ablation studies are summarized in Table 3-6, and Figure 3, respectively. + +Effectiveness of the proposed modules. We first perform an ablation study to validate the effectiveness of the GSA and HGC modules. Results are summarized in Table 3. Our baseline model follows CMD-SE [22], which utilizes CLIP and removes the need for pretrained object detectors in OV-HOI detection. First, we apply the GSA module on the baseline model to improve its transferability for novel + +![](images/557e00ab2c038bac9807248f7eb225e1320931244c580e5b4861354f5540343f.jpg) +Figure 3. Evaluation on the value of grouping threshold $N$ . + +classes, denoted as $+GSA$ . This modification achieves mAP improvements of $7.05\%$ for Non-Rare, $2.66\%$ for Rare, $4.32\%$ for Unseen, and $5.04\%$ for Full categories. Moreover, we integrate the HGC module with the baseline to enhance its ability to distinguish semantically similar interactions, denoted as $+HGC$ . The results indicate that the HGC module enables the baseline model to achieve $14.81\%$ in mAP for the Full categories. Finally, combining both modules results in the best performance, with an mAP of $12.46\%$ on the Unseen categories. These results demonstrate the significant potential of SGC-Net in enhancing interaction understanding within an open-vocabulary setting. + +Effectiveness of the proposed layer splitting manner. The block partitioning strategy plays a crucial role in the GSA module due to its potential to preserve visual-language associations. Experimental results in Table 4 indicate that separating the last layer into an independent block is more effective than other combinations of partitioning strategies. This is because the last layer of CLIP's image encoder has the strongest association with text embeddings, and partitioning it into a separate block helps preserve this correlation while minimizing potential disruptions. + +Number of blocks and number of layers in each block. To show the importance of multi-granularity feature fusion across different layers, we compare the performance of SGC-Net with varying numbers of blocks and layers per block. As shown in Table 5, increasing the number of blocks from 1 to 3 gradually improves the performance, particularly for Unseen interactions. The best performance is achieved with three blocks, each containing three layers. Reducing the number of layers to 2 degrades performance, primarily due to neglecting mid-layer features. It should be noted that the beginning layers are excluded, as they primarily encode features with limited semantics. + +Comparison of different aggregation methods. Due to the substantial differences between shallow and deep features, directly integrating multi-granularity features may disrupt the alignment of vision-language embeddings in the pre-trained CLIP model. To address this issue, we intro- + +![](images/ca73b4b25118e362dbaabcea7a673b7bc42c770fc5ceea5f056d477730280d57.jpg) +Figure 4. Qualitative results of SGC-Net. On the left-hand side, we visualize feature correspondence through cosine similarity calculations from both deep and shallow layers, and compare the predicted HOI categories. On the right-hand side, we show two inference examples and the absolute score gap between two images at each hierarchical level by querying the description with extracted HOI representations. + +![](images/197b8058686b77ad67f4140cc3f469d85d2d0d80b70ee5ae9823dafd9a044837.jpg) + +duce the DGW function, which mitigates these disruptions by accounting for the distances between layers. As shown in Table 6, DGW consistently outperforms direct aggregation strategies such as Sum, Concat. Additionally, setting the aggregation weights as learnable further enhances the performance of DGW. These results demonstrate that our method, by assigning smaller weights to features from more distant layers during multi-granularity fusion, effectively improves model performance in open-vocabulary tasks. + +Grouping threshold $N$ . A large grouping threshold, such as $N = 8$ , tends to produce rudimentary hierarchies and generalized descriptions. In contrast, smaller grouping thresholds increase the time required to construct the class hierarchy due to the more detailed recursive comparisons and groupings involved. As shown in Figure 3, the performance of SGC-Net improves as the grouping threshold increases. To balance computational efficiency and performance gains, we chose to set $N = 6$ . + +# 4.4. Qualitative Results + +To demonstrate the effectiveness of the proposed GSA module in leveraging CLIP's multi-granularity features, we present cosine similarity maps of features and qualitative OV-HOI detection results. At the left hand side of Figure 4, we show the patch similarity of Unseen classes not included in the training process. We observe that the intermediate layers of our method contain detailed information on local objects, including interaction patterns. Moreover, by leveraging these distinctive features, our SGC-Net improves HOI detection performance for both seen and unseen classes compared to using only last-layer features. + +At the right hand side of Figure 4, we show two groups of examples that demonstrate the inference and absolute score gap calculation via the proposed HGC module. Specifically, we input two Unseen class images (② and ④) and two + +Rare class images (① and ③) to evaluate their similarity with the hierarchical descriptions. The absolute score gap is computed as the absolute difference in similarity scores between paired images: i.e., ① vs. ② and ③ vs. ④. Due to the weak comparative information in the early descriptions, similar images tend to yield similar scores at the beginning. As we progress towards a more comparative description, the disparity between the scores of the two images gradually increases. Besides, this also highlights the interpretability of our method, as it provides the user insights regarding the attributes that elicit stronger responses to the input. Additionally, it indicates the specific layer of description at which the scores of different images diverge, widening the gap. + +# 5. Conclusion + +In this paper, we propose the SGC-Net which enhances CLIP's transferability and discriminability to handle two critical issues in OV-HOI: feature granularity deficiency and semantic similarity confusion. Specifically, the GSA module enhances CLIP's image encoder by improving its ability to capture fine-grained HOI features while maintaining alignment between multi-granularity visual features and text embeddings. The HGC module strengthens CLIP's text encoder in distinguishing semantically related HOI categories by incorporating a grouping and comparison strategy into the LLM. Through extensive comparative experiments, we validate the effectiveness of SGC-Net on two datasets. + +Acknowledgment. This work is supported by the Guangzhou University Research Project (No: RQ2021013, 2025A03J3128), the Guangdong Natural Science Funds for Distinguished Young Scholars (No. 2023B1515020041), the Guangdong Basic and Applied Basic Research Foundation (No. 2023A1515110077), the NSF of China (No. 62372125). + +# References + +[1] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 3 +[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[3] Yichao Cao, Qingfei Tang, Feng Yang, Xiu Su, Shan You, Xiaobo Lu, and Chang Xu. Re-min, learn and reason: Exploring the cross-modal semantic correlations for language-guided hoi detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23492-23503, 2023. 2 +[4] Yichao Cao, Qingfei Tang, Xiu Su, Song Chen, Shan You, Xiaobo Lu, and Chang Xu. Detecting any human-object interaction relationship: Universal hoi detector with spatial prompt learning on foundation models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[5] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6 +[6] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004-9013, 2021. 2 +[7] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, pages 11583-11592, 2022. 2 +[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3 +[9] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 696-712. Springer, 2020. 2 +[10] Jianjun Gao, Kim-Hui Yap, Kejun Wu, Duc Tri Phan, Kratika Garg, and Boon Siew Han. Contextual human object interaction understanding from pre-trained large language model. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13436-13440. IEEE, 2024. 2 +[11] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2 +[12] Tanmay Gupta, Alexander Schwing, and Derek Hoiem. Nofrills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of + +the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2 +[13] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 2 +[14] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 2 +[15] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 6 +[16] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-imiL: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 3 +[17] Sheng Jin, Xueying Jiang, Jiaxing Huang, Lewei Lu, and Shijian Lu. Llms meet vlms: Boost open vocabulary object detection with fine-grained descriptors. arXiv preprint arXiv:2402.04630, 2024. 3 +[18] Bumsoo Kim, Taeho Choi, Jaewoo Kang, and Hyunwoo J Kim. Uniondet: Union-level detector towards real-time human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 498-514. Springer, 2020. 2 +[19] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 2 +[20] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[21] Ting Lei, Fabian Caba, Qingchao Chen, Hailin Jin, Yuxin Peng, and Yang Liu. Efficient adaptive human-object interaction detection with concept-guided memory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6480-6490, 2023. 2 +[22] Ting Lei, Shaofeng Yin, and Yang Liu. Exploring the potential of large foundation models for open-vocabulary hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16657-16667, 2024. 1, 2, 3, 4, 5, 6, 7 +[23] Huadong Li, Ying Wei, Shuailei Ma, Mingyu Chen, and Ge Li. Ripple transformer: A human-object interaction backbone and a new prediction strategy for smart surveillance devices. IEEE Transactions on Consumer Electronics, 2024. 1 +[24] Lin Li, Jun Xiao, Guikun Chen, Jian Shao, Yueting Zhuang, and Long Chen. Zero-shot visual relation detection via com + +pose visual cues from large language models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[25] Yunheng Li, ZhongYu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. arXiv preprint arXiv:2406.00670, 2024. 2 +[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2 +[27] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2 +[28] Yue Liao, Si Liu, Fei Wang, Yanjie Chen, Chen Qian, and Ji-ashi Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482–490, 2020. 2 +[29] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2, 3, 4, 5, 6, 7 +[30] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2 +[31] Xin Lin, Changxing Ding, Yibing Zhan, Zijian Li, and Dacheng Tao. Hl-net: Heterophily learning network for scene graph generation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19476-19485, 2022. 2 +[32] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 2 +[33] Xinpeng Liu, Yong-Lu Li, Xiaogian Wu, Yu-Wing Tai, Cewu Lu, and Chi-Keung Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113–20122, 2022. 6 +[34] J Macqueen. Some methods for classification and analysis of multivariate observations. In Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press, 1967. 2, 4 +[35] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. Advances in Neural Information Processing Systems, 36:45895-45906, 2023. 2 + +[36] Esteve Valls Mascaro, Daniel Sliwowski, and Dongheui Lee. Hoi4abot: Human-object interaction anticipation for human intention reading collaborative robots. arXiv preprint arXiv:2309.16524, 2023. 1 +[37] Jean Massardi, Mathieu Gravel, and Éric Beaudry. Parc: A plan and activity recognition component for assistive robots. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3025-3031. IEEE, 2020. 1 +[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3, 4 +[39] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. 12mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3 +[40] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23507-23517, 2023. 2, 3, 4, 5, 6 +[41] Jeeseung Park, Jin-Woo Park, and Jong-Seok Lee. Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17152-17162, 2023. 2 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[43] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 2, 6 +[44] Danyang Tu, Wei Sun, Guangtao Zhai, and Wei Shen. Agglomerative transformer for human-object interaction detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21614-21624, 2023. 2 +[45] Mesut Erhan Unal and Adriana Kovashka. Weakly-supervised hoi detection from interaction labels only and language/vision-language priors. arXiv preprint arXiv:2303.05546, 2023. 3 +[46] Suchen Wang, Kim-Hui Yap, Henghui Ding, Jiyan Wu, Jun-song Yuan, and Yap-Peng Tan. Discovering human interactions with large-vocabulary objects via query and multi-scale detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13475-13484, 2021. 2, 6 +[47] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognexploitation (CVPR), pages 939-948, 2022. 2 +[48] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 1, 3, 5, 6 +[49] Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14647-14657, 2022. 2 +[50] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[51] Nan Xi, Jingjing Meng, and Junsong Yuan. Open set video hoi detection from action-centric chain-of-look prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3079–3089, 2023. 1 +[52] Jie Yang, Bingliang Li, Ailing Zeng, Lei Zhang, and Ruimao Zhang. Open-world human-object interaction detection via multi-modal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2024. 2, 3, 6, 7 +[53] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. International Journal of Computer Vision, pages 1-19, 2024. 3 +[54] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 6 +[55] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 2, 6 +[56] Frederic Z Zhang, Yuhui Yuan, Dylan Campbell, Zhuoyao Zhong, and Stephen Gould. Exploring predicate visual context in detecting of human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10411-10421, 2023. 2 +[57] Sipeng Zheng, Boshen Xu, and Qin Jin. Open-category human-object interaction pre-training via language modeling framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19392-19402, 2023. 6 +[58] Xubin Zhong, Xian Qu, Changxing Ding, and Dacheng Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 2 +[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In- + +ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 3 +[60] Penghao Zhou and Mingmin Chi. Relation parsing neural network for human-object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 843-851, 2019. 2 +[61] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2 +[62] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 \ No newline at end of file diff --git a/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/images.zip b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4729227167db615da703e1b0b06d6e043c0cabf7 --- /dev/null +++ b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1eafadf43623c8cfd307c1a3a875d1da8cd565a33397d2867b1aebf1e63576b +size 591331 diff --git a/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/layout.json b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..50f0cacae2372f15117e73ca3f06fbdc2ccda0d1 --- /dev/null +++ b/2025/SGC-Net_ Stratified Granular Comparison Network for Open-Vocabulary HOI Detection/layout.json @@ -0,0 +1,8661 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 138 + ], + "type": "text", + "content": "SGC-Net: Stratified Granular Comparison Network for Open-Vocabulary HOI Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 161, + 475, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 161, + 475, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 161, + 475, + 190 + ], + "type": "text", + "content": "Xin Lin Chong Shi Zuopeng Yang* Haojin Tang* Zhili Zhou Guangzhou University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 191, + 478, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 191, + 478, + 204 + ], + "spans": [ + { + "bbox": [ + 121, + 191, + 478, + 204 + ], + "type": "text", + "content": "linxin94@gzhu.edu.cn, shichong@e.gzhu.edu.cn, yzpeng@gzhu.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 187, + 205, + 418, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 205, + 418, + 217 + ], + "spans": [ + { + "bbox": [ + 187, + 205, + 418, + 217 + ], + "type": "text", + "content": "tanghaojin@gzhu.edu.cn, zhou_zhili@163.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 270, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 569 + ], + "type": "text", + "content": "Recent open-vocabulary human-object interaction (OV-HOI) detection methods primarily rely on large language model (LLM) for generating auxiliary descriptions and leverage knowledge distilled from CLIP to detect unseen interaction categories. Despite their effectiveness, these methods face two challenges: (1) feature granularity deficiency, due to reliance on last layer visual features for text alignment, leading to the neglect of crucial object-level details from intermediate layers; (2) semantic similarity confusion, resulting from CLIP's inherent biases toward certain classes, while LLM-generated descriptions based solely on labels fail to adequately capture inter-class similarities. To address these challenges, we propose a stratified granular comparison network. First, we introduce a granularity sensing alignment module that aggregates global semantic features with local details, refining interaction representations and ensuring robust alignment between intermediate visual features and text embeddings. Second, we develop a hierarchical group comparison module that recursively compares and groups classes using LLMs, generating fine-grained and discriminative descriptions for each interaction category. Experimental results on two widely-used benchmark datasets, SWIG-HOI and HICO-DET, demonstrate that our method achieves state-of-the-art results in OV-HOI detection. Codes is available at GitHub." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 591, + 137, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 137, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 137, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "Human-Object Interaction (HOI) detection aims to localize human-object pairs and recognize their interactions, providing an efficient way for human-centric scene understanding. It plays a crucial role in various computer vision tasks, such as assistive robotics [36, 37] and video analysis [23, 51]. Recently, the emerging field of open-vocabulary HOI (OV-HOI) detection, which broadens the scope of HOI detection" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 550, + 323 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 550, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 550, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 550, + 323 + ], + "type": "image", + "image_path": "f4ca9d7b9568999a5a5b647785ff2e13d338e679aef1f0b132c036eff939a3c9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 323, + 492, + 333 + ], + "lines": [ + { + "bbox": [ + 373, + 323, + 492, + 333 + ], + "spans": [ + { + "bbox": [ + 373, + 323, + 492, + 333 + ], + "type": "text", + "content": "(a) Feature granularity deficiency" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 334, + 550, + 396 + ], + "blocks": [ + { + "bbox": [ + 317, + 334, + 550, + 396 + ], + "lines": [ + { + "bbox": [ + 317, + 334, + 550, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 334, + 550, + 396 + ], + "type": "image", + "image_path": "1142b082fad16054671e9ff7fe0c7d6b0c2565d355c28f2564464214bb83905b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 396, + 492, + 406 + ], + "lines": [ + { + "bbox": [ + 374, + 396, + 492, + 406 + ], + "spans": [ + { + "bbox": [ + 374, + 396, + 492, + 406 + ], + "type": "text", + "content": "(b) Semantic similarity confusion" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. (a) The last layer captures high-level global semantics but contains fewer low-level local details compared to intermediate layers. The red dot marks the selected patch. (b) Category-level and part-level descriptions overlook inter-class similarity, leading to difficulty in distinguishing semantically similar classes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 484, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 508 + ], + "type": "text", + "content": "by recognizing and associating objects beyond predefined categories, has gained increasing attention." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 509, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 688 + ], + "type": "text", + "content": "OV-HOI detection has made notable progress, largely due to the success of vision-language models (VLMs) like CLIP [42]. Unlike zero-shot learning, which prohibits the access of any unannotated unseen classes during training, open-vocabulary learning leverages class name embeddings from VLMs for auxiliary supervision [50, 62]. Therefore, existing CLIP-based HOI detection methods naturally fall under OV-HOI detection. However, these methods face challenges in open-world scenarios because they rely on pretrained object detectors, and using category names as classifiers struggles to capture the variability of HOIs. To overcome these challenges, recent works [22, 48] have removed the need for pretrained detectors by extracting HOI features directly through image encoders while incorporating language priors from LLMs via text encoders." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Despite these advancements, current OV-HOI detection methods without pretrained object detectors still suffer from" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4539" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": "feature granularity deficiency. This issue arises because CLIP, trained with image-level alignment, produces globally aligned image-text features lacking the local detail needed for effective OV-HOI detection. As shown in Figure 1(a), the last layer focuses on high-level semantic features but neglects low-level local details like arm and face postures, compared to intermediate layers. Additionally, feature dissimilarity between shallow and deep layers increases with network depth, posing significant challenges in aligning intermediate visual features with text embeddings. Moreover, OV-HOI methods that rely on category-level or part-level text classifiers are prone to semantic confusion due to several inherent limitations. First, CLIP's training on large-scale, long-tail datasets tends to introduce biases towards certain classes [49]. For example, as shown on the left side of Figure 1(b), CLIP often confuses \"hug cat\" with \"hold cat.\" Second, descriptions generated by LLMs based solely on labels may fail to distinguish between semantically similar classes. As illustrated on the right side of Figure 1(b), LLMs produce similar part-level descriptions like \"arm extension\" for both \"hold cat\" and \"chase cat\"." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 327, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 590 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose an end-to-end OV-HOI detection network named SGC-Net. First, we design the granularity sensing alignment (GSA) module, leveraging multi-granularity visual features from CLIP to improve the transferability for novel classes. The primary challenge in this process is effectively aggregating global semantic features with local details to refine coarse interaction representations while ensuring alignment between intermediate visual features and text embeddings. Accordingly, we introduce a block partitioning strategy that groups layers based on their relative distances. Specifically, we divide the visual encoder of CLIP into several blocks, ensuring a little variation in the features within each block. To align local details more precisely, we perform a fusion of multi-granularity features within each block, assigning distinct and trainable Gaussian weights to each layer. Once these intra-block features are fused, we proceed to aggregate the fused features across all blocks to maintain global semantic consistency. Furthermore, to preserve the pretrained alignment between text and visual features in CLIP, we employ visual prompt tuning [7, 25, 61] by adding learnable tokens to the visual features of each layer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Secondly, we propose the hierarchical group comparison (HGC) module to effectively differentiate semantically similar categories by recursively grouping and comparing classes via LLMs. Specifically, we utilize clustering algorithms (e.g., " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "-means [34]) for grouping. Even with extensive prior knowledge of LLMs, comparing numerous categories for OV-HOI detection remains challenging due to the quadratic growth of the description matrix with the number of categories. To address this challenge, we adopt group-specific comparison strategies to improve efficiency. For" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "smaller groups, we directly query LLMs for comparative descriptions, whereas for larger groups, we summarize the group's characteristics using LLMs and incorporate these summaries into new queries to highlight distinctive features of each class within the group. By recursively construct a class hierarchy, we can classify HOI by descending from the top to the bottom of the hierarchy, comparing HOI and text embeddings at each level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 168, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 288 + ], + "type": "text", + "content": "In summary, the innovation of the proposed SGC-Net is three-fold: (1) The GSA module effectively aggregates multi-granular features while aligning the local details and global semantic features of OV-HOI; (2) The HGC module iteratively generates discriminative descriptions to refine the classification boundaries of labels; (3) The efficacy of the proposed SGC-Net is evaluated on two widely-used OV-HOI detection benchmark datasets, i.e., HICO-DET [5] and SWIG-HOI [46]. Experimental results show that our SGC-Net consistently outperforms state-of-the-art methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 300, + 400, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 400, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 400, + 312 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 319, + 405, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 405, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 405, + 331 + ], + "type": "text", + "content": "2.1. HOI Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 338, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 567 + ], + "type": "text", + "content": "Existing HOI detection works can be divided into two categories, two-stage methods [3, 9, 21, 26, 32, 41, 43, 55, 56] and one-stage methods [6, 11, 18, 19, 28, 44, 58]. Two-stage methods typically use an object detector to recognize humans and objects, followed by specialized modules to associate humans with objects and identify their interactions, e.g., multi-steams [11, 12, 27], graphs [9, 30, 31, 54] or compositional learning [13, 14]. In contrast, the one-stage approach detects human-object pairs and their interactions simultaneously, without requiring stepwise processing. In particular, RLIP [60] proposes a pre-training strategy for HOI detection based on image captions. PPDM [28] reformulates the HOI detection task as a point detection and matching problem and achieves simultaneous object and interaction detection. However, existing HOI detection approaches are constrained by a closed-set assumption, which restricts their ability to recognize only predefined interaction categories. In contrast, our work aims to detect and recognize HOIs in the open-vocabulary scenario." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 574, + 549, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 549, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 549, + 587 + ], + "type": "text", + "content": "2.2. CLIP-based Open Vocabulary HOI Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "To alleviate the closed-set limitation, many studies [10, 22, 29, 35, 40, 47, 52] have been devoted to OV-HOI detection, aiming to identify both base and novel categories of HOIs while only base categories are available during training. Specifically, they transfer knowledge from the large-scale visual-linguistic pre-trained model CLIP to enhance interaction understanding. For instance, GEN-VLKTs [29] and HOICLIP [40] convert HOI labels into phrase descriptions to initialize the classifier and extract visual features from the CLIP image encoder to guide the interaction visual feature" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4540" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "learning. MP-HOI [52] incorporates the visual prompts into language-guided-only HOI detectors to enhance its generalization capability. THID [48] proposes a HOI sequence parser to detect multiple interactions. CMD-SE [22] detects HOIs at different distances using distinct feature maps. However, the adopted loss function in [22] involves minimizing differences between continuous and discrete variables therefore hard to optimize. Besides, aligning intermediate features with text embeddings may break the alignment in pre-trained CLIP due to the substantial differences. The proposed approach is a method free from pretrained object detectors but has the following advantages compared with existing works. First, it aggregates multi-granularity features from the visual encoder more appropriately. Second, it is easy to optimize and deploy to existing models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 258, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 258, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 258, + 270 + ], + "type": "text", + "content": "2.3. Leverage LLM for visual classification" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 529 + ], + "type": "text", + "content": "Recently, LLMs [1, 2, 8, 16] have made remarkable progress, revolutionizing natural language processing tasks with hundreds of billions of parameters and techniques such as reinforcement learning. Existing works [4, 17, 45, 59] have shown their effectiveness in generating comprehensive descriptions, particularly in classification and detection tasks. Specifically, Menon and Vondrick [38] generate textual descriptions directly from labels to assist VLMs in image classification. I2MVFormer [39] leverages LLM to generate multi-view document supervision for zero-shot image classification. ContextDET [53] utilizes contextual LLM tokens as conditional object queries to enhance the visual decoder for object detection. RECODE [24] leverages LLMs to generate descriptions for different components of relation categories. However, descriptions generated solely by LLMs from labels often tend to be generic and lack sufficient discriminability among semantically similar classes. To address this issue, our work leverages LLM to generate text descriptions by comparing different classes to reduce inter-class similarity and make the decision boundary of each class more compact." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 538, + 111, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 111, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 111, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "This section provides details of the proposed stratified granular comparison network (SGC-Net). Specifically, we first describe the network architecture, followed by an explanation of the training and inference pipeline. As shown in Figure 2, SGC-Net comprises a granularity sensing alignment (GSA) module and a hierarchical group comparison (HGC) module. The GSA module enhances coarse HOI representations by utilizing distance-Aware Gaussian weighting (DGW) and visual prompt tuning to combine multi-granularity features from CLIP's image encoder. The HGC module refines decision boundaries for semantically similar classes by recursively comparing and grouping them using LLM-derived knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 484, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 484, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 484, + 85 + ], + "type": "text", + "content": "3.1. Granularity Sensing Alignment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 209 + ], + "type": "text", + "content": "Existing OV-HOI methods [22, 29, 40] typically use features from the last layer of CLIP's image encoder to model HOIs. While these deep features effectively capture high-level semantics, they often lack the local details necessary for HOI detection. One approach to address this issue is to aggregate feature maps from different levels and feed them into an interaction decoder. However, the significant differences between shallow and deep features can disrupt the pre-trained vision-language alignment, ultimately reducing CLIP's zero-shot capability for OV-HOI tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "text", + "content": "To address this issue, we propose the GSA module, which effectively captures fine-grained details in human-object interactions while preserving the vision-language alignment in the pre-trained CLIP model. Specifically, we first divide the CLIP's image encoder into " + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "text", + "content": " blocks, ensuring a little variation in the features within each block. Then, within each block, distinct distance-aware Gaussian weights are assigned to the transformer layers based on their relative neighborhood distances. These weights are trainable, enabling the model to adaptively learn layer- and block-specific information from the training data. Finally, we aggregate the features across different blocks. For each block containing " + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "text", + "content": " transformer layers, the aggregated feature " + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 210, + 555, + 376 + ], + "type": "text", + "content": " can be represented as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 385, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 385, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 351, + 385, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\alpha_ {l} ^ {s} = \\exp \\left(- \\frac {1}{2} \\frac {(d - l) ^ {2}}{\\sigma^ {2}}\\right), \\quad l \\in [ 1, d ], \\\\ \\boldsymbol {Z} = \\sum_ {s = 1} ^ {S} \\alpha_ {s} \\left(\\sum_ {l = 1} ^ {d} \\alpha_ {l} ^ {s} \\boldsymbol {F} _ {l}\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "f4371a0e0ea92f0418c8c02bbe47a83872a1c6827924dcb5d281a198e62660db.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": " represents the layer index. " + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{F}_l" + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": " is a feature map from the " + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": "-th layer of CLIP image encoder. " + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\alpha_l^s" + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\alpha_s" + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": " denote the distance-aware Gaussian weights between layers within the same block and across different blocks, respectively. By adjusting variance parameter " + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 450, + 554, + 617 + ], + "type": "text", + "content": ", the Eq. (1) can assign a higher weight to nearby layers and a lower weight to distant ones, facilitating more effective and flexible integration of features across different depth levels. Notably, the last layer of CLIP is treated as a separate block with a large weight. This design aims to keep the original visual-language correlations established in the pre-trained model. Different blocks can capture varying levels of granularity. Specifically, the early blocks focus more on fine-grained local details, while the later block emphasizes coarse-grained global features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "To better align the intermediate visual features with the text embeddings, we employ visual prompt tuning [59] by introducing learnable tokens onto the visual features of each layer in the frozen encoder. During visual prompt tuning, the GSA module enables the gradients to be directly backpropagated to the middle layers of the visual encoder. It can promote the alignment of mid-layer features and text embeddings, enhancing the similarity across different layers." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4541" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 333 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 333 + ], + "type": "image", + "image_path": "741a6c5728444953f1fa0a0b475ab9f1e19b064463d0935e07a47f15dcf36b33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 384 + ], + "type": "text", + "content": "Figure 2. The framework of SGC-Net. It eliminates the need for a pretrained object detector and includes two new modules for OV-HOI detection: (1) The GSA module partitions the CLIP visual encoder into blocks, aggregates features via DGW, and integrates an HOI decoder for fine-grained HOI representations. (2) The HGC module uses LLM to recursively construct class hierarchy, enabling HOI classification by traversing the hierarchy from top to bottom and comparing HOI representations with text embeddings at each level." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 396, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 295, + 552 + ], + "type": "text", + "content": "Compared with CMD-SE [22], we effectively aggregate multi-granularity features from the CLIP image encoder using trainable Gaussian weights, which have the following advantages. First, it allows intermediate layer visual features to complement last layer features while preserving the visual-language associations of the pre-trained CLIP. This approach addresses the issue in CMD-SE [22], where relying solely on discrete single-layer features for HOI modeling results in a loss of fine-grained information. Second, our method is easy to optimize. We eliminate the need to use loss functions that align discrete transformer layer indices with continuous human-object interaction distances, thereby avoiding the complex optimization processes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "content": "Subsequently, following previous works [22, 29, 40], the final interaction representation " + }, + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 587, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 587, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 135, + 587, + 295, + 601 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} = \\operatorname {D e c} (\\boldsymbol {Q}, \\boldsymbol {Z}), \\tag {2}", + "image_path": "9f965f3023658fedf6ad20de069b92ed9f26fe019a56bf548d3ef3553fd5a1e9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": " refers to HOI queries, while " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": " is treated as both the key and value input to the HOI Decoder. The output " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": " is later passed to the bounding box regressor and classifier to predict the bounding boxes of human-object pairs and their interaction categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 231, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 231, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 231, + 685 + ], + "type": "text", + "content": "3.2. Hierarchical Group Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "After obtaining the fine-grained interaction representation by the GSA module, designing a discriminative classifier re" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 504 + ], + "type": "text", + "content": "mains a challenging task. Existing methods [22, 29, 40] typically design classifiers with the manual prompt \"a photo of a person [verb] [object]\". However, these classifiers primarily rely on category names, often neglecting the contextual information provided by language. Recent research [22] has introduced the category descriptions produced by LLMs to enhance the model's generalization capability. However, these descriptions directly generated by LLMs from labels are often generic and semantically similar." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 505, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 556, + 601 + ], + "type": "text", + "content": "To address this issue, we propose the hierarchical group comparison (HGC) module. Inspired by the coarse-to-fine approach humans use to recognize objects, HGC employs three strategies: Grouping, Comparison, and Hierarchical Classification. It identifies semantically similar descriptions and refines decision boundaries from generic to more discriminative. Specifically, we first utilize LLMs to generate initial descriptions for each HOI category." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 604, + 544, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 544, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 544, + 628 + ], + "type": "text", + "content": "Q: What features are useful to distinguish {HOI category} in a photo?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 628, + 339, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 628, + 339, + 637 + ], + "spans": [ + { + "bbox": [ + 315, + 628, + 339, + 637 + ], + "type": "text", + "content": "A:-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "Grouping: Following [38], we employ the pretrained CLIP text encoder to map the LLM-generated descriptions into the latent space. Utilizing these features, we apply a clustering algorithm, such as " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "-means [34], to perform grouping and identify semantic neighbors. Specifically, the value of " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": " is calculated as the number of categories divided by a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4542" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "content": "predefined grouping threshold " + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 190 + ], + "type": "text", + "content": ". Even with extensive prior knowledge of LLMs, comparing a large number of categories presents a formidable challenge, as it needs to generate a comprehensive description matrix that grows quadratically with the number of categories. To address this issue, we employ tailored strategies for generating descriptions based on the group size, allowing for more efficient and effective comparisons. Specifically, we adopt summary-based comparison for larger groups and direct comparison for smaller ones." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 288 + ], + "type": "text", + "content": "Comparing: For groups involving a large number of categories (i.e., exceeding half of the grouping threshold), we leverage the LLM to summarize their overall characteristics. By incorporating these summarized characteristics into a new query, we can generate a comparative description that captures the distinctive features and relationships among the categories. This approach enhances our ability to effectively compare and contrast the elements within a group." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 292, + 268, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 268, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 268, + 303 + ], + "type": "text", + "content": "Q: Summarize the following interactions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 304, + 250, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 250, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 250, + 316 + ], + "type": "text", + "content": "with one sentence: {category list}?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 317, + 181, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 181, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 181, + 329 + ], + "type": "text", + "content": "A: {subset description}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 285, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 285, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 285, + 360 + ], + "type": "text", + "content": "Q: What features are useful to distinguish {HOI category} from {subset description}?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 80, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 80, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 80, + 369 + ], + "type": "text", + "content": "A:-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 373, + 294, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 294, + 445 + ], + "type": "text", + "content": "When the number of elements within a group is relatively small, a direct query to the LLM enables us to obtain an exceptionally comparative description. This approach capitalizes on the LLM's capabilities to provide a comprehensive and insightful analysis, enhancing the comparative understanding of the elements within the group." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 449, + 293, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 293, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 293, + 485 + ], + "type": "text", + "content": "Q: What features are useful to distinguish {target category} from {other categories} in a photo?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 80, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 80, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 80, + 495 + ], + "type": "text", + "content": "A:-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 500, + 294, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 294, + 548 + ], + "type": "text", + "content": "Utilizing these detailed and comparative descriptions, we can achieve more nuanced text embedding. Subsequently, we can continue to perform clustering, followed by additional comparisons to derive new descriptors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": "Hierarchical Classification: After constructing the class hierarchy, we can obtain text features at multiple hierarchical levels. More specific, the hierarchical text features can be represented as " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{D_1, D_2, \\dots, D_T\\}" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " is the number of test classes. " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "D_i \\in \\mathbb{R}^{M_j \\times C}" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " represents the hierarchical text embeddings for category " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " is the embedding dimension. " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "M_j" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " is the number of comparative descriptions for category " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": ". Accordingly, the cosine similarity score between the HOI feature " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": "-th description of the " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": "-th HOI category, denoted as " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "D_i^j" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": ", can be expressed as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 680, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 680, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 143, + 680, + 295, + 696 + ], + "type": "interline_equation", + "content": "p _ {i} ^ {j} = \\boldsymbol {D} _ {i} ^ {j} \\cdot \\boldsymbol {x} ^ {T}. \\tag {3}", + "image_path": "1dd78e1f7f4bebed66b6f2e8bebb3dfa93c27db5adc01b02256b1bc40b50a6b3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "However, unreliable low-level descriptions can introduce" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "errors and redundancies, degrading the quality of high-level descriptions and ultimately impairing the model's classification performance. To address this issue, we define an iterative evaluator " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{u}_i^k" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " to evaluate the acceptability of the current " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "-th level description for the " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "-th category as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 379, + 144, + 553, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 144, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 379, + 144, + 553, + 159 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {i} ^ {k} = \\mathbb {I} \\left(p _ {i} ^ {k + 1} > p _ {i} ^ {k} + \\tau\\right), \\tag {4}", + "image_path": "a1cbc9ada89cc71041a507de3a32edc2101d7d63ced1983b8306616f7052ed53.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "text", + "content": " denotes the tolerance parameter, " + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "text", + "content": " represents indicator function. Eq. (4) ensures that a new score is merged only if a subsequent discriminative description yields a higher score than the current one. Subsequently, we obtain the running average of the longest sequence of monotonically increasing " + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 166, + 553, + 237 + ], + "type": "text", + "content": " values as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 247, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 247, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 361, + 247, + 553, + 281 + ], + "type": "interline_equation", + "content": "r (\\boldsymbol {x}, i) = \\frac {p _ {i} ^ {1} + \\sum_ {j = 2} ^ {M _ {i}} p _ {i} ^ {j} \\prod_ {k = 1} ^ {j - 1} \\boldsymbol {u} _ {i} ^ {k}}{1 + \\sum_ {j = 2} ^ {M _ {i}} \\prod_ {k = 1} ^ {j - 1} \\boldsymbol {u} _ {i} ^ {k}}. \\tag {5}", + "image_path": "e12db4165966f8e40cb0078a7c9ac551f571882cd0b783677e7d9c48695b25d4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "content": "Finally, the similarity score " + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "inline_equation", + "content": "s(\\pmb{x}, i)" + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "content": " between HOI feature " + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "content": "-th interaction label embedding is computed using a weighted average function, as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 340, + 337, + 553, + 351 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 337, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 340, + 337, + 553, + 351 + ], + "type": "interline_equation", + "content": "s (\\boldsymbol {x}, i) = (1 - \\lambda) \\left(\\boldsymbol {p} _ {i} ^ {1} + \\boldsymbol {t} \\cdot \\boldsymbol {x} ^ {T}\\right) + \\lambda (r (\\boldsymbol {x}, i)). \\tag {6}", + "image_path": "8f4d29a811e729ee40d23fc05b76a616986d661f36837145dcece21db25ee887.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "content": "Here, we apply prompt tuning [48] with learnable text tokens " + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "content": " to learn the sentence format rather than using manually defined context words. Besides, our fusion method takes into account the score of the initial description " + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "inline_equation", + "content": "p_i^1" + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "content": " as an offset, and we introduce a constant hyperparameter " + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0, 1]" + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "content": " to balance the two terms." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 442, + 444, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 444, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 444, + 454 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 461, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 484 + ], + "type": "text", + "content": "In this subsection, we elaborate on the processes of training and inference of our model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "Training. During the training stage, we follow the query-based methods [22, 29, 48] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. The matching cost " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " for the matching process and the targeting cost for the training back-propagation share the same strategy, which is formulated as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 326, + 578, + 553, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 578, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 326, + 578, + 553, + 606 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {b} \\sum_ {i \\in \\{h, o \\}} \\mathcal {L} _ {b} ^ {i} + \\lambda_ {i o u} \\sum_ {i \\in \\{h, o \\}} \\mathcal {L} _ {i o u} ^ {i} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {7}", + "image_path": "c06346012a61854066045bce26f554cd2c322fa6ca9b20dcf851ba58155fadf8.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_b" + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{iou}" + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": " denote the box regression, intersection over union, and classification losses, respectively. During the training stage, we follow the query-based methods [22, 29, 40] to assign a bipartite matching prediction with each ground-truth using the Hungarian algorithm [20]. " + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\lambda_b" + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\lambda_{iou}" + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": " are the hyper-parameter weight." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "type": "text", + "content": "Inference. For each HOI prediction, including the bounding-box pair " + }, + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "type": "inline_equation", + "content": "(\\hat{b}_h^i,\\hat{b}_o^i)" + }, + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "type": "text", + "content": ", the bounding box score " + }, + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 313, + 688, + 553, + 714 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4543" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "the box regressor, and the interaction score " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\hat{s}_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " from the interaction classifier, the final score " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\hat{s}_i'" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 148, + 106, + 295, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 106, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 106, + 295, + 121 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {i} ^ {\\prime} = \\hat {s} _ {i} \\cdot \\hat {c} _ {i} ^ {\\gamma} \\tag {8}", + "image_path": "5e7d8dc4fdf94aae9ce04469af99a958f12c1f4efde1c4ebf11a03a620486268.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 130, + 295, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 295, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 130, + 295, + 155 + ], + "type": "inline_equation", + "content": "\\gamma > 1" + }, + { + "bbox": [ + 55, + 130, + 295, + 155 + ], + "type": "text", + "content": " is a constant used during inference to suppress overconfident objects [54, 55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 166, + 132, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 132, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 132, + 180 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 186, + 178, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 178, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 178, + 201 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 205, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 337 + ], + "type": "text", + "content": "Datasets. Our experiments are conducted on two datasets: namely, SWIG-HOI [46] and HICO-DET [5]. The SWIG-HOI dataset provides diverse human interactions with large-vocabulary objects, comprising 400 human actions and 1,000 object categories. During the testing, we utilize approximately 5,500 interactions, including around 1,800 interactions that are not found in the training set. The annotations of HICO-DET include 600 combinations of 117 human actions and 80 objects. Among the 600 interactions, follow [22, 48], we simulate the open-vocabulary detection setting by holding out 120 rare interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 422 + ], + "type": "text", + "content": "Evaluation Metric. Following in [5, 22, 29, 33, 48], we use the mean Average Precision (mAP) for evaluation. An HOI triplet prediction is classified as a true-positive example when the following criteria are satisfied: 1) The intersection over union of the human bounding box and object bounding box are larger than 0.5 w.r.t. the GT bounding boxes; 2) the predicted interaction category is accurate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": "Implementation Details. We follow the settings of previous work [22, 48] to build our model upon the pretrained CLIP. Specifically, for the visual encoder, we employ the ViT-B/16 version as our visual encoder and apply 12 learnable tokens in each layer to detect human-object interactions. For the text encoder, we introduce 8 prefix tokens and 4 conjunctive tokens to connect the words of human actions and objects for adaptively learning categories' information. Our model is optimized using AdamW with an initial learning rate of " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": ", using 64 as the batch size for SWIG-HOI dataset and 32 for HICO-DET. We set the cost weights " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_{b}" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\lambda_{iou}" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": " to 5, 2, and 5, respectively. The LLM we utilize is GPT-3.5. We set the hyperparameter " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": " in Eq. (6) as 0.5. In all experiments, the variance parameter " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": " is set to 1, and the tolerance parameter " + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 422, + 295, + 601 + ], + "type": "text", + "content": " is set to 0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 610, + 285, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 285, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 285, + 624 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "HICO-DET. To facilitate fair comparison, we first compare our model with state-of-the-art methods without pretrained detectors on the HICO-DET dataset. As shown in Table 1, SGC-Net outperforms CMD-SE [22] by " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "6.57\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "4.39\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "4.87\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " in mAP on the Unseen, Seen, and Full categories, respectively. Furthermore, even when compared to methods using pretrained detectors, SGC-Net achieves" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 200 + ], + "type": "table", + "html": "
MethodPretrained DetectorUnseenSeenFull
SCL [15]✓19.0730.3928.08
GEN-VLKT [29]✓21.3632.9130.56
OpenCat [57]✓21.4633.8631.38
HOICLIP [40]✓23.4834.4732.26
THID [48]✗15.5324.3222.38
CMD-SE [22]✗16.7023.9522.35
SGC-Net✗23.2728.3427.22
", + "image_path": "0274f524d1678b71172bdb00bcddc1de0e633de2904acba9465713458b9441e0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 247, + 550, + 371 + ], + "blocks": [ + { + "bbox": [ + 313, + 211, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 232 + ], + "type": "text", + "content": "Table 1. Comparison of our proposed SGC-Net with state-of-the-art methods on the HICO-DET dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 247, + 550, + 371 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 550, + 371 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 550, + 371 + ], + "type": "table", + "html": "
MethodNon-rareRareUnseenFull
CHOID [46]10.936.632.646.64
QPIC [43]16.9510.846.2111.12
GEN-VLKT [29]20.9110.41-10.87
MP-HOI [52]20.2814.78-12.61
THID [48]17.6712.8210.0413.26
CMD-SE [22]21.4614.6410.7015.26
SGC-Net23.6716.5512.4617.20
", + "image_path": "df89ab6a7eccf5fe45af0940af81e2e5e1f20545a220168f02480529bc00e3f1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 334, + 418, + 535, + 502 + ], + "blocks": [ + { + "bbox": [ + 313, + 382, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 553, + 403 + ], + "type": "text", + "content": "Table 2. Comparison of our proposed SGC-Net with state-of-the-art methods on the SWIG-HOI dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 418, + 535, + 502 + ], + "lines": [ + { + "bbox": [ + 334, + 418, + 535, + 502 + ], + "spans": [ + { + "bbox": [ + 334, + 418, + 535, + 502 + ], + "type": "table", + "html": "
MethodNon-rareRareUnseenFull
Base15.6911.537.3211.45
+ GSA22.7416.0011.6416.49
+ HGC21.1814.1910.6914.81
SGC-Net23.6716.5512.4617.20
", + "image_path": "21604ab7af9ec9bf1880b2560c78edfcf0393fc244ab44c0ed7cb231e4adc571.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 342, + 512, + 525, + 524 + ], + "lines": [ + { + "bbox": [ + 342, + 512, + 525, + 524 + ], + "spans": [ + { + "bbox": [ + 342, + 512, + 525, + 524 + ], + "type": "text", + "content": "Table 3. Ablation studies of the proposed method." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "competitive performance. Although recent OV-HOI methods (e.g., GEN-VLKT [29], OpenCat [57], HOICLIP [40]) leverage CLIP text embeddings for interaction classification, they typically rely on a DETR architecture with pretrained weights. It is important to note that comparing our method with these approaches on HICO-DET dataset is not entirely fair, as the COCO dataset used for DETR pretraining shares the same object label space as HICO-DET." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "SWIG-HOI. Table 2 shows that SGC-Net outperforms all state-of-the-art methods on various setups. Specifically, SGC-Net outperforms CMD-SE [22] by " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.76\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.91\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " in mAP on the Rare and Unseen categories, respectively. Furthermore, SGC-Net outperforms the best OV-HOI method with a pretrained object detector by " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "4.59\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " in mAP on the Full category. These results indicate that" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4544" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 295, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 135 + ], + "type": "table", + "html": "
Layer splitting mannerNon-rareRareUnseenFull
{4-6}, {7-9}, {10-12}23.0315.7211.3516.35
{6-8}, {9-10}, {11-12}23.6116.5211.6417.01
{6-8}, {9-11}, {12}23.6716.5512.4617.20
", + "image_path": "246720302836e8ecc43fa04265533b241d977e29f14036001f6f0a39d40ab6ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 168, + 296, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "type": "text", + "content": "Table 4. Effectiveness of different layer splitting manners. The elements in " + }, + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "type": "text", + "content": " means the layer numbers that are used." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 168, + 296, + 284 + ], + "lines": [ + { + "bbox": [ + 56, + 168, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 296, + 284 + ], + "type": "table", + "html": "
Number of blocksNon-rareRareUnseenFull
{12}21.1814.1910.0914.81
{9-11}, {12}21.9114.4810.0114.78
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{3-5}, {6-8}, {9-11}, {12}22.6115.2512.0116.53
Number of layers
{8-9}, {10-11}, {12}22.4515.1711.5515.98
{6-8}, {9-11}, {12}23.6716.5512.4617.20
{4-7}, {8-11}, {12}23.5016.2112.1316.75
", + "image_path": "78a6eba360e6bfbd2e042806eda08829caf5372b2a4798324a47ee7c0767ad6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 309, + 284, + 419 + ], + "blocks": [ + { + "bbox": [ + 73, + 287, + 277, + 298 + ], + "lines": [ + { + "bbox": [ + 73, + 287, + 277, + 298 + ], + "spans": [ + { + "bbox": [ + 73, + 287, + 277, + 298 + ], + "type": "text", + "content": "Table 5. Evaluation on the number of blocks and layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 309, + 284, + 419 + ], + "lines": [ + { + "bbox": [ + 67, + 309, + 284, + 419 + ], + "spans": [ + { + "bbox": [ + 67, + 309, + 284, + 419 + ], + "type": "table", + "html": "
AggregationNon-rareRareUnseenFull
Self-attention21.3713.437.8813.90
Concat21.8415.5210.2215.68
Sum22.0914.899.5915.26
DGW23.4016.1411.6016.71
Sum*22.2115.1610.4515.64
DGW*23.6716.5512.4617.20
", + "image_path": "ea7ca0e49a9a028124b1c348c9e4e6814f6956b96175f5845f736e4cb6488ad4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 295, + 448 + ], + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 448 + ], + "type": "text", + "content": "Table 6. Comparison of different aggregation methods. * denotes that the weights are trainable." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 556 + ], + "type": "text", + "content": "methods such as GEN-VLKT [29] and MP-HOI [52], which rely on pretrained detectors, perform suboptimally on the SWIG-HOI dataset. This is primarily because these methods struggle to scale effectively with vocabulary size, ultimately limiting their applicability in open-world scenarios. In contrast, SGC-Net overcomes this constraint by not relying on any detection pretraining, demonstrating superior capabilities in detecting and recognizing OV-HOI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 149, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 149, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 149, + 576 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "To prove the effectiveness of our proposed methods, we conduct five ablation studies on the SWIG-HOI dataset. Results of the ablation studies are summarized in Table 3-6, and Figure 3, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": "Effectiveness of the proposed modules. We first perform an ablation study to validate the effectiveness of the GSA and HGC modules. Results are summarized in Table 3. Our baseline model follows CMD-SE [22], which utilizes CLIP and removes the need for pretrained object detectors in OV-HOI detection. First, we apply the GSA module on the baseline model to improve its transferability for novel" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 526, + 217 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 526, + 217 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 526, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 526, + 217 + ], + "type": "image", + "image_path": "557e00ab2c038bac9807248f7eb225e1320931244c580e5b4861354f5540343f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 219, + 542, + 231 + ], + "lines": [ + { + "bbox": [ + 326, + 219, + 542, + 231 + ], + "spans": [ + { + "bbox": [ + 326, + 219, + 542, + 231 + ], + "type": "text", + "content": "Figure 3. Evaluation on the value of grouping threshold " + }, + { + "bbox": [ + 326, + 219, + 542, + 231 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 326, + 219, + 542, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": "classes, denoted as " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "+GSA" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": ". This modification achieves mAP improvements of " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "7.05\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": " for Non-Rare, " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "2.66\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": " for Rare, " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "4.32\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": " for Unseen, and " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "5.04\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": " for Full categories. Moreover, we integrate the HGC module with the baseline to enhance its ability to distinguish semantically similar interactions, denoted as " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "+HGC" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": ". The results indicate that the HGC module enables the baseline model to achieve " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "14.81\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": " in mAP for the Full categories. Finally, combining both modules results in the best performance, with an mAP of " + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "inline_equation", + "content": "12.46\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": " on the Unseen categories. These results demonstrate the significant potential of SGC-Net in enhancing interaction understanding within an open-vocabulary setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 510 + ], + "type": "text", + "content": "Effectiveness of the proposed layer splitting manner. The block partitioning strategy plays a crucial role in the GSA module due to its potential to preserve visual-language associations. Experimental results in Table 4 indicate that separating the last layer into an independent block is more effective than other combinations of partitioning strategies. This is because the last layer of CLIP's image encoder has the strongest association with text embeddings, and partitioning it into a separate block helps preserve this correlation while minimizing potential disruptions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "Number of blocks and number of layers in each block. To show the importance of multi-granularity feature fusion across different layers, we compare the performance of SGC-Net with varying numbers of blocks and layers per block. As shown in Table 5, increasing the number of blocks from 1 to 3 gradually improves the performance, particularly for Unseen interactions. The best performance is achieved with three blocks, each containing three layers. Reducing the number of layers to 2 degrades performance, primarily due to neglecting mid-layer features. It should be noted that the beginning layers are excluded, as they primarily encode features with limited semantics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Comparison of different aggregation methods. Due to the substantial differences between shallow and deep features, directly integrating multi-granularity features may disrupt the alignment of vision-language embeddings in the pre-trained CLIP model. To address this issue, we intro-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4545" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 304, + 239 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 304, + 239 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 304, + 239 + ], + "type": "image", + "image_path": "ca73b4b25118e362dbaabcea7a673b7bc42c770fc5ceea5f056d477730280d57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 241, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 555, + 276 + ], + "type": "text", + "content": "Figure 4. Qualitative results of SGC-Net. On the left-hand side, we visualize feature correspondence through cosine similarity calculations from both deep and shallow layers, and compare the predicted HOI categories. On the right-hand side, we show two inference examples and the absolute score gap between two images at each hierarchical level by querying the description with extracted HOI representations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 291, + 72, + 553, + 239 + ], + "blocks": [ + { + "bbox": [ + 291, + 72, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 291, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 291, + 72, + 553, + 239 + ], + "type": "image", + "image_path": "197b8058686b77ad67f4140cc3f469d85d2d0d80b70ee5ae9823dafd9a044837.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 287, + 294, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 294, + 395 + ], + "type": "text", + "content": "duce the DGW function, which mitigates these disruptions by accounting for the distances between layers. As shown in Table 6, DGW consistently outperforms direct aggregation strategies such as Sum, Concat. Additionally, setting the aggregation weights as learnable further enhances the performance of DGW. These results demonstrate that our method, by assigning smaller weights to features from more distant layers during multi-granularity fusion, effectively improves model performance in open-vocabulary tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "content": "Grouping threshold " + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "content": ". A large grouping threshold, such as " + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "content": ", tends to produce rudimentary hierarchies and generalized descriptions. In contrast, smaller grouping thresholds increase the time required to construct the class hierarchy due to the more detailed recursive comparisons and groupings involved. As shown in Figure 3, the performance of SGC-Net improves as the grouping threshold increases. To balance computational efficiency and performance gains, we chose to set " + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 167, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 167, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 167, + 526 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the proposed GSA module in leveraging CLIP's multi-granularity features, we present cosine similarity maps of features and qualitative OV-HOI detection results. At the left hand side of Figure 4, we show the patch similarity of Unseen classes not included in the training process. We observe that the intermediate layers of our method contain detailed information on local objects, including interaction patterns. Moreover, by leveraging these distinctive features, our SGC-Net improves HOI detection performance for both seen and unseen classes compared to using only last-layer features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "At the right hand side of Figure 4, we show two groups of examples that demonstrate the inference and absolute score gap calculation via the proposed HGC module. Specifically, we input two Unseen class images (② and ④) and two" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 287, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 443 + ], + "type": "text", + "content": "Rare class images (① and ③) to evaluate their similarity with the hierarchical descriptions. The absolute score gap is computed as the absolute difference in similarity scores between paired images: i.e., ① vs. ② and ③ vs. ④. Due to the weak comparative information in the early descriptions, similar images tend to yield similar scores at the beginning. As we progress towards a more comparative description, the disparity between the scores of the two images gradually increases. Besides, this also highlights the interpretability of our method, as it provides the user insights regarding the attributes that elicit stronger responses to the input. Additionally, it indicates the specific layer of description at which the scores of different images diverge, widening the gap." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 459, + 388, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 459, + 388, + 471 + ], + "spans": [ + { + "bbox": [ + 314, + 459, + 388, + 471 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 480, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 624 + ], + "type": "text", + "content": "In this paper, we propose the SGC-Net which enhances CLIP's transferability and discriminability to handle two critical issues in OV-HOI: feature granularity deficiency and semantic similarity confusion. Specifically, the GSA module enhances CLIP's image encoder by improving its ability to capture fine-grained HOI features while maintaining alignment between multi-granularity visual features and text embeddings. The HGC module strengthens CLIP's text encoder in distinguishing semantically related HOI categories by incorporating a grouping and comparison strategy into the LLM. Through extensive comparative experiments, we validate the effectiveness of SGC-Net on two datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 625, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 555, + 703 + ], + "type": "text", + "content": "Acknowledgment. This work is supported by the Guangzhou University Research Project (No: RQ2021013, 2025A03J3128), the Guangdong Natural Science Funds for Distinguished Young Scholars (No. 2023B1515020041), the Guangdong Basic and Applied Basic Research Foundation (No. 2023A1515110077), the NSF of China (No. 62372125)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4546" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "type": "text", + "content": "[1] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 294, + 191 + ], + "type": "text", + "content": "[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 194, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 258 + ], + "type": "text", + "content": "[3] Yichao Cao, Qingfei Tang, Feng Yang, Xiu Su, Shan You, Xiaobo Lu, and Chang Xu. Re-min, learn and reason: Exploring the cross-modal semantic correlations for language-guided hoi detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23492-23503, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 296, + 316 + ], + "type": "text", + "content": "[4] Yichao Cao, Qingfei Tang, Xiu Su, Song Chen, Shan You, Xiaobo Lu, and Chang Xu. Detecting any human-object interaction relationship: Universal hoi detector with spatial prompt learning on foundation models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "type": "text", + "content": "[5] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 296, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 417 + ], + "type": "text", + "content": "[6] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9004-9013, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "type": "text", + "content": "[7] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In CVPR, pages 11583-11592, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "type": "text", + "content": "[8] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 566 + ], + "type": "text", + "content": "[9] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 696-712. Springer, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 296, + 633 + ], + "type": "text", + "content": "[10] Jianjun Gao, Kim-Hui Yap, Kejun Wu, Duc Tri Phan, Kratika Garg, and Boon Siew Han. Contextual human object interaction understanding from pre-trained large language model. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13436-13440. IEEE, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "text", + "content": "[11] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[12] Tanmay Gupta, Alexander Schwing, and Derek Hoiem. Nofrills human-object interaction detection: Factorization, layout encodings, and training techniques. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF International Conference on Computer Vision, pages 9677-9685, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 152 + ], + "type": "text", + "content": "[13] Zhi Hou, Xiaojiang Peng, Yu Qiao, and Dacheng Tao. Visual compositional learning for human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 584-600. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 207 + ], + "type": "text", + "content": "[14] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14646-14655, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "text", + "content": "[15] Zhi Hou, Baosheng Yu, and Dacheng Tao. Discovering human-object interaction concepts via self-compositional learning. In European Conference on Computer Vision, pages 461-478. Springer, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "text", + "content": "[16] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-imiL: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "type": "text", + "content": "[17] Sheng Jin, Xueying Jiang, Jiaxing Huang, Lewei Lu, and Shijian Lu. Llms meet vlms: Boost open vocabulary object detection with fine-grained descriptors. arXiv preprint arXiv:2402.04630, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 420 + ], + "type": "text", + "content": "[18] Bumsoo Kim, Taeho Choi, Jaewoo Kang, and Hyunwoo J Kim. Uniondet: Union-level detector towards real-time human-object interaction detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 498-514. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 486 + ], + "type": "text", + "content": "[19] Bumsoo Kim, Jonghwan Mun, Kyoung-Woon On, Minchul Shin, Junhyun Lee, and Eun-Sol Kim. Mstr: Multi-scale transformer for end-to-end human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19578-19587, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 488, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 520 + ], + "type": "text", + "content": "[20] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[21] Ting Lei, Fabian Caba, Qingchao Chen, Hailin Jin, Yuxin Peng, and Yang Liu. Efficient adaptive human-object interaction detection with concept-guided memory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6480-6490, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "type": "text", + "content": "[22] Ting Lei, Shaofeng Yin, and Yang Liu. Exploring the potential of large foundation models for open-vocabulary hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16657-16667, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 553, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 688 + ], + "type": "text", + "content": "[23] Huadong Li, Ying Wei, Shuailei Ma, Mingyu Chen, and Ge Li. Ripple transformer: A human-object interaction backbone and a new prediction strategy for smart surveillance devices. IEEE Transactions on Consumer Electronics, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Lin Li, Jun Xiao, Guikun Chen, Jian Shao, Yueting Zhuang, and Long Chen. Zero-shot visual relation detection via com" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4547" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "pose visual cues from large language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Yunheng Li, ZhongYu Li, Quansheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-clip: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. arXiv preprint arXiv:2406.00670, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 295, + 207 + ], + "type": "text", + "content": "[26] Yong-Lu Li, Siyuan Zhou, Xijie Huang, Liang Xu, Ze Ma, Hao-Shu Fang, Yanfeng Wang, and Cewu Lu. Transferable interactiveness knowledge for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3585-3594, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 273 + ], + "type": "text", + "content": "[27] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 331 + ], + "type": "text", + "content": "[28] Yue Liao, Si Liu, Fei Wang, Yanjie Chen, Chen Qian, and Ji-ashi Feng. Ppdm: Parallel point detection and matching for real-time human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 482–490, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 333, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 397 + ], + "type": "text", + "content": "[29] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "text", + "content": "[30] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3746-3753, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "text", + "content": "[31] Xin Lin, Changxing Ding, Yibing Zhan, Zijian Li, and Dacheng Tao. Hl-net: Heterophily learning network for scene graph generation. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19476-19485, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[32] Xin Lin, Changxing Ding, Jing Zhang, Yibing Zhan, and Dacheng Tao. Ru-net: Regularized unrolling network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19466, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "text", + "content": "[33] Xinpeng Liu, Yong-Lu Li, Xiaogian Wu, Yu-Wing Tai, Cewu Lu, and Chi-Keung Tang. Interactiveness field in human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20113–20122, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[34] J Macqueen. Some methods for classification and analysis of multivariate observations. In Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press, 1967. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[35] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: towards adapting clip for practical zero-shot hoi detection. Advances in Neural Information Processing Systems, 36:45895-45906, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[36] Esteve Valls Mascaro, Daniel Sliwowski, and Dongheui Lee. Hoi4abot: Human-object interaction anticipation for human intention reading collaborative robots. arXiv preprint arXiv:2309.16524, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 163 + ], + "type": "text", + "content": "[37] Jean Massardi, Mathieu Gravel, and Éric Beaudry. Parc: A plan and activity recognition component for assistive robots. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3025-3031. IEEE, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 164, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 196 + ], + "type": "text", + "content": "[38] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 274 + ], + "type": "text", + "content": "[39] Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. 12mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15169-15179, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 331 + ], + "type": "text", + "content": "[40] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23507-23517, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 333, + 555, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 388 + ], + "type": "text", + "content": "[41] Jeeseung Park, Jin-Woo Park, and Jong-Seok Lee. Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17152-17162, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 453 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 456, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 510 + ], + "type": "text", + "content": "[43] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10410-10419, 2021. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "text", + "content": "[44] Danyang Tu, Wei Sun, Guangtao Zhai, and Wei Shen. Agglomerative transformer for human-object interaction detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21614-21624, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 557, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 601 + ], + "type": "text", + "content": "[45] Mesut Erhan Unal and Adriana Kovashka. Weakly-supervised hoi detection from interaction labels only and language/vision-language priors. arXiv preprint arXiv:2303.05546, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 602, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 667 + ], + "type": "text", + "content": "[46] Suchen Wang, Kim-Hui Yap, Henghui Ding, Jiyan Wu, Jun-song Yuan, and Yap-Peng Tan. Discovering human interactions with large-vocabulary objects via query and multi-scale detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13475-13484, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[47] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4548" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognexploitation (CVPR), pages 939-948, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[48] Suchen Wang, Yueqi Duan, Henghui Ding, Yap-Peng Tan, Kim-Hui Yap, and Junsong Yuan. Learning transferable human-object interaction detector with natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939-948, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "type": "text", + "content": "[49] Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14647-14657, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[50] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, et al. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 316 + ], + "type": "text", + "content": "[51] Nan Xi, Jingjing Meng, and Junsong Yuan. Open set video hoi detection from action-centric chain-of-look prompting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3079–3089, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "type": "text", + "content": "[52] Jie Yang, Bingliang Li, Ailing Zeng, Lei Zhang, and Ruimao Zhang. Open-world human-object interaction detection via multi-modal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "text", + "content": "[53] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. International Journal of Computer Vision, pages 1-19, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 295, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 469 + ], + "type": "text", + "content": "[54] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13319-13327, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[55] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20104-20112, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[56] Frederic Z Zhang, Yuhui Yuan, Dylan Campbell, Zhuoyao Zhong, and Stephen Gould. Exploring predicate visual context in detecting of human-object interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10411-10421, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "text", + "content": "[57] Sipeng Zheng, Boshen Xu, and Qin Jin. Open-category human-object interaction pre-training via language modeling framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19392-19402, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 691 + ], + "type": "text", + "content": "[58] Xubin Zhong, Xian Qu, Changxing Ding, and Dacheng Tao. Glance and gaze: Inferring action-aware points for one-stage human-object interaction detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13234-13243, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. In-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 241 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "ternational Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[60] Penghao Zhou and Mingmin Chi. Relation parsing neural network for human-object interaction detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 843-851, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[61] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175-11185, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[62] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4549" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_content_list.json b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d2676d85ee6247b4688d5cb8ee4e31a5578a1761 --- /dev/null +++ b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_content_list.json @@ -0,0 +1,1719 @@ +[ + { + "type": "text", + "text": "SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction", + "text_level": 1, + "bbox": [ + 161, + 130, + 834, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinran Yang $^{1}$ , Donghao Ji $^{1}$ , Yuanqi Li $^{1}$ , Jie Guo $^{1}$ , Yanwen Guo $^{1*}$ , Junyuan Xie $^{1}$ $^{1}$ Nanjing University, Nanjing, China", + "bbox": [ + 179, + 179, + 815, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xryang,donghaoji}@smail.nju.edu.cn", + "bbox": [ + 344, + 219, + 653, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuanqili,guojie,ywguo,jyxie}@nju.edu.cn", + "bbox": [ + 320, + 236, + 671, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec9ce4293209621d31fd801112fd0f36b864b75d1209a8aae9b181ffcc5198cf.jpg", + "image_caption": [ + "Figure 1. We present a novel method using 2D images to generate Spherical Gaussians, enabling the accurate representation of 3D edges. Spherical Gaussians combine the advantages from 3D Gaussian Splatting and the ability to preserve excellent 3D structures. With a corresponding curve extraction algorithm, Spherical Gaussians efficiently reconstruct 3D edge curves by only 2D supervision." + ], + "image_footnote": [], + "bbox": [ + 93, + 280, + 906, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 573, + 326, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting technique has achieved high quality novel view synthesis as well as fast rendering speed. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is due to the fact that Gaussian's attributes are primarily tailored and fine-tuned for rendering diverse 2D images by their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries, from which we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians is optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and then compared to the ground-truth edge map extracted from the corresponding image, without the need for any 3D guidance or su", + "bbox": [ + 88, + 604, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pervision. Given Spherical Gaussians serve as intermedi- dia for the robust edge representation, we further introduce a novel optimization-based algorithm called SGCR to directly extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying great efficiency. Code is available at https://github.com/Martinyxr/SGCR.", + "bbox": [ + 511, + 575, + 908, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 709, + 645, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In computer graphics and vision, photorealistic novel view synthesis (NVS) and accurate geometry reconstruction are two significant objectives. 3D Gaussian Splatting [19] offers an appealing explicit representation—3D Gaussians, resulting in high-quality synthesis of novel views and real-time rendering speed. However, it falls short in capturing accurate 3D geometric structures since the anisotropic 3D Gaussians, which are intended for efficient rasterization and sufficient expression ability, do not align with the geometric features (surfaces, edges and corners) in a faithful manner. A question naturally arises: Can we utilize 3D Gaussians", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 114, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5793", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/127d104e1879a4c7f9cdbeea0d6311b729694c564043572e8308ed1bbc7b9fea.jpg", + "image_caption": [ + "(a) 3D Gaussian Splatting" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 272, + 207 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1143525946b7625596f3e50a6476fc618efe0c54f9a3b38986126672a16b54a.jpg", + "image_caption": [ + "(b) Spherical Gaussians", + "Figure 2. Visualization for optimized Gaussians. (a) Gaussians from original 3D Gaussian Splatting. (b) Our Spherical Gaussians." + ], + "image_footnote": [], + "bbox": [ + 303, + 87, + 480, + 207 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for the direct representation of intricate geometric structures while retaining their benefits in 2D rendering?", + "bbox": [ + 89, + 277, + 482, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several recent works attempt to directly address the task of sdf generation or mesh reconstruction [6, 12, 17, 52] with 3D Gaussians. In this paper, we try to answer this question from a new perspective—reconstructing 3D parametric curves with 'Spherical Gaussians'. Following the settings of classic rendering works [19, 29], we use a set of calibrated multi-view images as input, but our goal substitutes novel view synthesis for 3D edge/curve reconstruction.", + "bbox": [ + 89, + 310, + 483, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature curves serve as critical geometric cues in characterizing the structure of 3D shapes, benefiting surface reconstruction, shape perception or other downstream tasks. Traditional methods [9, 13, 18, 21] usually work with the direct representation of 3D shapes, such as polygonal meshes or point clouds. However, the presence of sharp edges may be compromised or entirely overlooked on account of imperfect 3D scanning and acquisition, especially when dealing with sparse and noisy data. In recent years, learning based-methods [26, 28, 41, 54] are proposed to mitigate this, but obtaining precise ground-truth labels for 3D edge supervision involves laborious manual annotation, and inevitably limiting their generality. On the contrary, 2D images is easily accessible and allow for straightforward identification of feature edges through mature edge detection methods, making 2D supervision more attractive for edge processing tasks. Despite advantages in detection and self-supervision, there still remains a huge obstacle to overcome—establishing correspondence between edges across different views when working with multi-view images, in other words, how to combine information from multiple views to construct a complete 3D structure? This is exactly the factor for our design of Spherical Gaussians (as well as SGCR), which serve as a robust bridge to connect 3D depiction with 2D guidance.", + "bbox": [ + 89, + 431, + 483, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method consists of two main stages: Spherical Gaussians generation and parametric curves extraction. During the first stage, we utilize 2D edge maps to refine Gaussians using a view-based rendering loss compared to the ground-truth edge map derived from the image of that view. Optimizing Gaussians directly as vanilla 3D Gaus", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/001a975b016e292d31f6e58d6ea667e93a57fab3111ecdc60028ce89abf4d888.jpg", + "image_caption": [ + "Figure 3. Illustration for fitting Bézier curve from Spherical Gaussians. $P_{1}$ , $P_{2}$ , $P_{3}$ and $P_{4}$ are weighted control points of the curve." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 179 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sian Splatting [19] is problematic due to the sparse distribution of edge pixels and the diversity of Gaussian attributes. The outcome Gaussians may recover 2D images to some extent but exhibit redundant irregular ellipsoids (see Fig. 2 (a)) and lack solid 3D geometric meanings, making following reconstruction tasks quite difficult. Therefore, we implement several constraints to transform standard 3D Gaussians into Spherical Gaussians. We introduce grid initialization, fixing-covariance training and a two-phase pruning strategy to guide the optimization process. Our final distribution of gaussians lie faithfully on sharp edges of objects in 3D space, appearing the same shape like spheres (see Fig. 2 (b)) and holding attributes in opacity and gray-scale color.", + "bbox": [ + 511, + 228, + 906, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second stage, we utilize these discrete Spherical Gaussians to extract smooth 3D parametric curves for final reconstruction of edges. We present a novel optimization-based algorithm called SGCR, which involves fitting edges represented by Spherical Gaussians onto rational Bezier curves. The process begins with line fitting and then gradually refines the curves by global optimization. The final results are stored by all control points and weights in these curves. An illustration for curve extraction from Spherical Gaussians is shown in Fig. 3.", + "bbox": [ + 511, + 425, + 905, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method using a variety of object categories with complex models and scenes from ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. Extensive experiments show that SGCR, which utilizes Spherical Gaussians as bridge to reconstruct 3D parametric curves by 2D supervisions, outperforms existing state-of-the-art methods on all metrics and achieves great efficiency. In summary, our contributions include:", + "bbox": [ + 511, + 575, + 905, + 696 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new 3D edge extraction method from multiview 2D images using Spherical Gaussians (SG).", + "- We introduce a new optimization-based algorithm (SGCR) tailored for directly reconstructing 3D parametric curves from our generated Spherical Gaussians.", + "- Our approach achieves the state-of-the-art performance as well as high efficiency on curve reconstruction tasks." + ], + "bbox": [ + 511, + 700, + 903, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 821, + 653, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-based 3D Representation and Rendering", + "text_level": 1, + "bbox": [ + 511, + 847, + 903, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based rendering (IBR) techniques use a collection of 2D images of a scene to generate the visual represent-", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5794", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tation of the scene and render novel views. Significant advancements have been achieved by Neural Radiance Fields (NeRF) [29], which depicts a scene's geometry and visual attributes as a radiance field, enabling the retrieval of color and volume density at any point in space and from any viewing angle for the purpose of rendering. Recently, the unstructured GPU-optimized 3D Gaussian Splitting (3DGS) [19] presents the explicit Gaussian primitive that achieves enhanced rendering speeds and improved quality without the reliance on neural components. It has also been quickly extended to multiple domains [12, 27, 45, 53, 55]. We convert standard 3D Gaussians into Spherical Gaussians, enabling significant advantages in both view rendering and the representation of geometric structures.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3D Curve Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 314, + 321, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The convention of 3D curve reconstruction is based on point cloud. Traditional methods [3, 9, 21, 22, 40] focus on local geometric properties of point clouds, such as normals [7, 42], curvatures [49] and hierarchical clustering [10]. Data-driven methods [15, 26, 28, 39] often adopt edge detection as a binary classification for point clouds. As network architectures advance, the classifier used for edge detection has become more complex models [5, 33, 41, 51]. RFEPS [47] presents a multistage optimization to augment the point set so that sufficient points are generated on potential edges of geometry. NerVE [54] presents a neural volumetric edge representation learned through a volumetric learning framework.", + "bbox": [ + 89, + 337, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extraction curves from multi-view images was pioneered by [4] which introduces a full SfM system using line segments. With the developments of line detections [30, 31, 48] and matching [2, 30, 32], several works [14, 16, 24, 25, 43] have attempted to revisit the line mapping problem, but they are restricted to straight lines and frequently generate fragmented small lines when dealing with curves. Lifting 2D sketches/curves to 3D curves [11, 34, 35, 46] has also been exploited, but the ability to generate parametric curves is still limited. Recently, NEF [50] learns a neural implicit field to capture the distribution of edge densities, and EMAP [23] encodes 3D edge distance and direction in Unsigned Distance Functions (UDF) from multi-view edge maps. However, they all require large training time for even a simple scene. In contrast, our method employs Spherical Gaussians as the direct representation for edges, achieving 3D parametric curve reconstruction both accurately and efficiently.", + "bbox": [ + 89, + 534, + 482, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 821, + 181, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 847, + 218, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. 3DGS [19] is a recently popular method for representing 3D scenes with 3D Gaussian", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "primitives, achieving both explicit representation and high-quality real-time rendering. Each Gaussian primitive is defined by a covariance matrix $\\Sigma$ , and a center point $p$ which is the mean of the Gaussian. The 3D distribution can be represented as: $G(x) = e^{-\\frac{1}{2} (x - p)^{\\mathrm{T}}\\Sigma^{-1}(x - p)}$ .", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A complete 3D Gaussian is given by its mean/center $p \\in \\mathbb{R}^3$ , color represented with spherical harmonics coefficients $\\mathbf{H} \\in \\mathbb{R}^k$ , opacity $\\alpha \\in \\mathbb{R}$ , quaternion $\\mathbf{r} \\in \\mathbb{R}^4$ , and scaling factor $\\mathbf{s} \\in \\mathbb{R}^3$ . For a given pixel, the combined color and opacity from multiple Gaussians are weighted by Eq. 3. The color blending for overlapping points is: $\\mathbf{C} = \\sum_{i \\in N} \\mathbf{c}_i \\alpha_i \\prod_{j=1}^{i-1} (1 - \\alpha_j)$ where $\\mathbf{c}_i, \\alpha_i$ denote the color and density of a point.", + "bbox": [ + 511, + 167, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spherical Gaussians. Our Spherical Gaussians enjoy the same nature of 3D Gaussians, but make the following changes for geometry priority and simplicity: (1) Remove the covariance matrix $\\pmb{\\Sigma}$ of Gaussians (attributes of scaling $\\mathbf{S}$ and rotation $\\mathbf{R}$ ) and replace it with a fixed radius $r_0$ , i.e. convert Gaussian primitives from various ellipsoids into spheres of uniform size. (2) Remove spherical harmonics coefficients $\\mathbf{H}$ and modify color to one-dimensional gray value (just for edge map rendering). According to our experiments, these changes do not hamper the backward propagation of gradients from 2D edge-map supervision, but impose great regularization on geometry distribution of 3D isotropic Gaussians, facilitating reconstruction tasks.", + "bbox": [ + 511, + 287, + 903, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anisotropic Gaussians v.s. Isotropic Gaussians. Traditional 3DGS [19] and its subsequent works mainly center on the technique of 'splatting'—utilizing different anisotropic Gaussian splats to accurately represent the entire scene. However, our work focuses on each 'Gaussian' itself, essentially displaying a reverse process: decomposing the splating results into small 'atoms' and assigning clear geometric meanings to every individual Gaussian. A thin and elongated Gaussian seems to be more efficient in representing edges, but it can be split into smaller 'basis-Gaussians' with few sacrifices but for better interpretation and convenience, especially when following pointcloud-related works. That's why we favour Spherical Gaussians and more discussions can be found in supplementary materials.", + "bbox": [ + 511, + 484, + 905, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spherical Gaussians Generation", + "text_level": 1, + "bbox": [ + 511, + 707, + 795, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Initialization", + "text_level": 1, + "bbox": [ + 511, + 729, + 648, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our full pipeline is illustrated in Fig. 4. Given a set of calibrated multi-view images, we first adopt edge detection method [38] to generate 2D edge maps for each view. 3DGS uses structure-from-Motion (SfM) [36] points as initialization for Gaussians, but this approach is not ideal for handling edge maps. Thus, we design grid initialization for Spherical Gaussians. We create a $50 \\times 50 \\times 50$ grid with points evenly distributed in space of $[0,1]^3$ (see Fig. 4—Grid initialization) as the initial centers for our Spherical Gaussians. This setting is better than random initialization", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5795", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/567532b9367e791888bd11304d8fcac8d67e532a51c3e2b2a3d0c39c65b8838e.jpg", + "image_caption": [ + "Figure 4. Pipeline illustration across 2D and 3D modality. Stage1: Spherical Gaussians generation. Stage2: Parametric curves extraction." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 903, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in producing neat structures. And we choose a constant radius $r_0$ to replace the covariance matrix with Spherical Gaussians of the same size. In our experiments, we use $r_0 = 0.005$ .", + "bbox": [ + 89, + 388, + 482, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Training Losses", + "text_level": 1, + "bbox": [ + 89, + 455, + 246, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edge Loss. The original loss functions in 3DGS [19] to train 3D Gaussians for each view is defined as:", + "bbox": [ + 89, + 474, + 482, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {D - S S I M} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 515, + 482, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_1$ and $\\mathcal{L}_{D-SSIM}$ are supervised by the ground-truth view. However, in edge maps, edge pixels are often sparsely distributed. Relying solely on this loss function will easily make the training process stuck in local optima, causing the opacities and colors for all Gaussians to converge towards zero and resulting in all rendered images becoming black. Hence, we opt for a new edge-aware loss $\\mathcal{L}_{edge}$ instead of $\\mathcal{L}_1$ in order to balance the impact of both edge and non-edge pixels within each edge map. The edge-aware loss is defined as:", + "bbox": [ + 89, + 541, + 483, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e d g e}} = \\frac {N _ {I} - | E _ {I} |}{N _ {I}} \\sum_ {i \\in E _ {I}} \\| I _ {i} - \\hat {I} _ {i} \\| _ {2} ^ {2} + \\frac {| E _ {I} |}{N _ {I}} \\sum_ {i \\notin E _ {I}} \\| I _ {i} - \\hat {I} _ {i} \\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 696, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{I}$ is the image resolution and $E_{I}$ refers to all edge pixels in current edge map (gray scale larger than threshold $\\eta = 0.3$ ), $I_{i}$ and $\\hat{I}_i$ refers to the pixel in Gaussian-rendering image and ground-truth edge-map image respectively. This edge-aware loss enhances the ability of Spherical Gaussians to identify edges from edge pixels.", + "bbox": [ + 89, + 750, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Opacity-Color Loss. Another issue during edge training is about occluded edges. Some 2D edge maps may lose true edge pixels due to occlusions from these image views (see the difference between Fig. 4-Edge maps and Rendering", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "views), while from other views the occlusion may not happen. This inconsistency among views will confuse Spherical Gaussians during training, potentially leading to unstable optimization. During training, color attribute is directly supervised by $\\mathcal{L}_{edge}$ loss, while opacity is indirectly optimized and reset at certain iterations. This difference sets them apart, and equating them impedes smooth optimization. Therefore, we introduce opacity-color loss $\\mathcal{L}_{oc}$ for all Spherical Gaussians to make the value of Gaussian opacity more consistent. This opacity-color loss is defined as:", + "bbox": [ + 511, + 388, + 906, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o c} = \\sum_ {i = 1} ^ {N _ {S G}} \\| o _ {i} - c _ {i} \\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 553, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{SG}$ is the number of all Spherical Gaussians during current training step, $o_i$ and $c_i$ refer to opacity and color attribute of each Spherical Gaussian respectively. $\\mathcal{L}_{oc}$ helps to keep consistency between edge density and color intensity during multi-view supervision. This mechanism ensures that occluded edges—represented by Gaussians with low opacity values—are less susceptible to premature pruning during optimization, resulting in more complete 3D reconstructions. It also helps rectify edge predictions missed by 2D detectors. In our experiment, after implementing $\\mathcal{L}_{oc}$ , Gaussian-rendered images can depict complete edge structures with no occlusion (Fig. 4—Rendering Views).", + "bbox": [ + 511, + 608, + 906, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularization Loss. To restrict total number of Spherical Gaussians and accelerate convergence, we add an extra regularization term $\\mathcal{L}_{\\text {regular }}$ on Gaussian's opacity. The regularization loss is defined as:", + "bbox": [ + 511, + 790, + 906, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\sum_ {i = 1} ^ {N _ {S G}} \\log \\left(1 + \\frac {o _ {i} ^ {2}}{0 . 5}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 862, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5796", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e6d6f64c05812fb532543fe202cf803d0d513edbae22feaa1a90f80d353cadd7.jpg", + "image_caption": [ + "Figure 5. The change on number of Gaussians during training." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{SG}$ is the number of current Spherical Gaussians and $o_i$ is each Gaussian's opacity. During the training process, this loss function works implicitly because Spherical Gaussians with opacity below a certain threshold will be pruned (as in 3DGS [19]), allowing for a comfortable number of total Gaussian primitives for next stage.", + "bbox": [ + 89, + 270, + 483, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, our final loss function for training Spherical Gaussians is presented as:", + "bbox": [ + 89, + 361, + 483, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left(1 - \\lambda_ {1}\\right) \\mathcal {L} _ {\\text {e d g e}} + \\lambda_ {1} \\mathcal {L} _ {D - S S I M} + \\lambda_ {2} \\mathcal {L} _ {o c} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e g}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 402, + 482, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the balancing parameters $\\lambda_1, \\lambda_2, \\lambda_3$ are set to 0.2, 2 and 0.01 in our experiments respectively.", + "bbox": [ + 89, + 431, + 483, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 470, + 261, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The effective training strategy is essential for creating properly structured Spherical Gaussians. We adopt a two-phase training strategy, each phase consists of 3,000 iterations. During the first phase, we conduct densification (split and clone) of Spherical Gaussians every 200 iterations. And we reset the opacity attribute to 0.1 every 1,000 iterations. At the end of first phase, we design a final-prune—removing all Spherical Gaussians whose opacity $o_i < 0.5$ and color $c_i < 0.1$ . This operation largely reduces the total number of Gaussians, leading to substantial reductions in both training time and noise. During the second phase, we no longer densify Spherical Gaussians but focus on refining their positions and attributes. In the end, we perform another final-prune to retain all significant Spherical Gaussians (at quantity of 1k to 5k) with their attributes. The variation of numbers for Spherical Gaussians during all iterations is depicted in Fig. 5. The whole training process of Spherical Gaussians for single scene takes only about 1 minute.", + "bbox": [ + 89, + 489, + 483, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Parametric Curve Extraction from Spherical Gaussians", + "text_level": 1, + "bbox": [ + 89, + 772, + 483, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the first stage, we get the generated Spherical Gaussians (denoted as $G$ ) aligned with object's edges, resembling a 3D wireframe constructed by many small spheres. In the second stage (Fig. 4—Stage 2), we present of Spherical-Gaussians Curve Reconstruction (SGCR) to accomplish the final 3D edge reconstruction task. SGCR is an", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Line Fitting", + "text_level": 1, + "bbox": [ + 514, + 90, + 684, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: a set of Spherical Gaussians $G$ with radius $r_0$ .", + "bbox": [ + 514, + 108, + 870, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: the set of line endpoints $L = \\{(p_i, q_i)\\}_{i=1}^{|L|}$ for the coarse line-fitting of 3D edges.", + "bbox": [ + 513, + 123, + 903, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denote: $P(G)$ is the centers of $G$ , and $N(0,1)$ is the Standard Normal Distribution.", + "bbox": [ + 513, + 154, + 903, + 184 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $L = \\emptyset$", + "2: while $|G| > N_0$ do", + "3: $S_{min} =$ MaxIntValue", + "4: for $i = 1$ to $n$ do", + "5: $p, q = \\text{RandomChoice}(P(G))$", + "6: $L_{p,q} = \\text{interpolate}(p, q, N_s) + r_0 * N(0,1)$", + "7: $G_{i} = \\{g\\in G|\\min_{x\\in L_{p,q}}||P(g) - x||_{2}^{2} < \\delta_{1}\\}$", + "8: $p', q' = \\operatorname*{arg min}_{p,q} \\mathcal{L}_{CD}(L_{p,q}, P(G_i))$", + "9: $S_{i} = \\min_{p,q}\\mathcal{L}_{CD}(L_{p,q},P(G_{i}))$", + "10: if $S_{i} \\leq S_{min}$ then", + "11: $S_{min} = S_i, G' = G_i, (\\hat{p}, \\hat{q}) = (p', q')$", + "12: end if", + "13: end for", + "14: $G = G - G^{\\prime},L = L\\cup \\{(\\hat{p},\\hat{q})\\} ;$", + "15: end while", + "16: return $L$ ;" + ], + "bbox": [ + 519, + 188, + 872, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "optimization-based algorithm guided by $G$ which includes two parts: line fitting and global optimization.", + "bbox": [ + 511, + 479, + 903, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Line Fitting", + "text_level": 1, + "bbox": [ + 511, + 518, + 643, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first fit a set of line segments to $G$ 's centers as the coarse fitting. The pseudo code of this part is shown in Algorithm 1. The whole process consists of several line-fitting turns and $G$ will be gradually fitted and partly removed by one line each turn. Line fitting will stop when the number of remaining $G$ is less than $N_0 = 5$ .", + "bbox": [ + 511, + 537, + 905, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In each turn, we conduct $n$ random searches and record the best result. At the beginning of $i$ -th search, we randomly choose two nearest Gaussian centers as initial endpoints $p, q$ of the fitting line, and interpolate $N_{s}$ points along the line. To simulate the shape of Spherical Gaussians, we dilate the sampled points up by duplicating and adding Gaussian noise with standard deviation equals $G$ 's radius $r_0$ . Then we apply Chamfer Distance to compute the fitting loss between dilated line points and $G$ . This Chamfer loss is defined as:", + "bbox": [ + 511, + 628, + 906, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C D} \\left(L _ {p, q}, P \\left(G _ {i}\\right)\\right) = \\frac {\\gamma_ {1}}{\\left| L _ {p , q} \\right|} \\sum_ {x \\in L _ {p, q}} \\min _ {y \\in P \\left(G _ {i}\\right)} \\| x - y \\| _ {2} ^ {2} \\\\ + \\frac {1}{| P (G _ {i}) |} \\sum_ {y \\in P (G _ {i})} \\min _ {x \\in L _ {p, q}} \\| x - y \\| _ {2} ^ {2} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 785, + 903, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {i} = \\left\\{g \\in G \\mid \\min _ {x \\in L _ {p, q}} \\left| | P (g) - x \\| _ {2} ^ {2} < \\delta_ {1} \\right. \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 877, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5797", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{p,q}$ denotes the dilated line points between $p, q$ in $i$ -th search, $P(G)$ represents Gaussian centers of all current $G$ and $\\gamma_1 = 2$ is a balance parameter. $G_i$ is a subset of $G$ whose centers are close to $L_{p,q}$ within a threshold $\\delta_1 = 0.02$ . We optimize $p, q$ by minimizing Eq. (6) with certain iterations and use the value of $\\mathcal{L}_{CD}$ after optimization as the final fitting score $S_i$ for $i$ -th search.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each turn, suppose the best fitting is in $k$ -th search, i.e. $S_{k} = \\min S_{i}$ , we record the optimized endpoints $p^{\\prime}, q^{\\prime}$ in $k$ -th search and delete $G_{k}$ to update the current $G$ . The next turn will then proceed accordingly. Once all the turns are completed, we will get a set of paired line endpoints $L = \\{(p_{i}, q_{i})\\}_{i=1}^{|L|}$ which together form an approximate line-fitting result (Fig. 4—Line Fitting).", + "bbox": [ + 89, + 196, + 480, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2. Global Optimization", + "text_level": 1, + "bbox": [ + 89, + 311, + 279, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After line fitting, we resume all $G$ deleted in previous part and take opacity attribute into account as the fine fitting. We interpolate another two control points between each line endpoint pair $(p_i, q_i)$ and introduce point weight to initialize 3rd order rational Bézier curves, which is defined as:", + "bbox": [ + 89, + 330, + 480, + 405 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nB (u) _ {p, w} = \\frac {\\sum_ {i = 0} ^ {3} B _ {3 , i} (u) p _ {i} w _ {i}}{\\sum_ {i = 0} ^ {3} B _ {3 , i} (u) w _ {i}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 414, + 480, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B_{3,i}$ is basis functions of Bézier curves, $p_i$ is the $i$ -th control points of curve and $w_i$ is the weight of control point $p_i$ ( $w_i \\equiv 1$ for simple Bézier curves). We opt for the rational Bézier curves over simple Bézier curves because simple Bézier curves cannot perfectly fit circles, which are frequently seen in object's edges.", + "bbox": [ + 89, + 460, + 482, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also sample $N_{s}$ points for each Bézier curve and dilate them up as in line fitting. Since $G$ with higher opacity values are more significant during fine fitting, we apply the weighted Chamfer Distance to compute the fine fitting loss between global curve points and all $G$ :", + "bbox": [ + 89, + 551, + 482, + 626 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {W C D} (L _ {a l l}, P (G)) = \\frac {\\gamma_ {2}}{| L _ {a l l} |} \\sum_ {x \\in L _ {a l l}} o _ {y _ {m i n}} \\cdot \\min _ {y \\in P (G)} \\| x - y \\| _ {2} ^ {2} \\\\ + \\frac {1}{| P (G) |} \\sum_ {y \\in P (G)} o _ {y} \\cdot \\min _ {x \\in L _ {\\text {a l l}}} \\| x - y \\| _ {2} ^ {2} \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 635, + 493, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{all}$ refers to global dilated points from all curves and $P(G)$ refers to Gaussian centers of $G$ , $o_y$ is the opacity attribute in $y$ -th $G$ and $\\gamma_2 = 2$ is another balance parameter. We also adopt endpoints loss to encourage connections between curves, which is defined as:", + "bbox": [ + 89, + 719, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n d p o i n t s}} = \\sum_ {x, y \\in P _ {E}} \\min \\left(\\| x - y \\| _ {2} ^ {2} - \\delta_ {2}, 0\\right) \\frac {\\| x - y \\| _ {2} ^ {2}}{\\| x - y \\| _ {2} ^ {2} - \\delta_ {2}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 803, + 491, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{E}$ is the set of endpoints (the first and last control points) and $\\delta_{2}$ (we choose 0.01) is a threshold to only regularize endpoints which are close to each other. The final", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "objective function to globally optimize all curves is defined as:", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\arg \\min \\quad \\left(\\mathcal {L} _ {W C D} + \\lambda \\mathcal {L} _ {\\text {e n d p o i n t s}}\\right) \\tag {11} \\\\ \\{\\{z _ {i} ^ {j}, w _ {i} ^ {j} \\} _ {j = 1} ^ {4} \\} _ {i = 1} ^ {| L |} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 123, + 903, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z_{i}^{j}$ and $w_{i}^{j}$ are coordinates and weights of four control points in each Bézier curve, $\\lambda$ is a balancing parameter which is set to 0.005 in our experiments. The pseudo code of this part can be found in the supplementary material.", + "bbox": [ + 511, + 164, + 903, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After global optimization, we finally extract 3D parametric curves from Spherical Gaussians in form of 3rd order rational Bezier curves (Fig. 4—Rational Bezier Curves).", + "bbox": [ + 511, + 227, + 903, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 287, + 643, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 314, + 607, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on the ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. ABC [20] consists of massive CAD models with edge annotations, and we use a portion of it—ABC-NEF, provided by [50]—which consisting of 115 distinct and challenging CAD models. ModelNet [44] includes 3D models of everyday objects such as chairs and tables, but it does not have ground-truth edge annotations. So we manually select and annotate edges on ModelNet ground-truth meshes. Our edge dataset consists of 120 furniture models with annotated edges and is used for a more comprehensive evaluation. We utilize BlenderProc [8] to generate rendered images with these objects. We generate 50 views of $800 \\times 800$ images for each object and adopt PiDiNet [38] to predict 2D edge maps. DTU [1] and Replica [37] are real-world datasets with multi-view images as input and we use it for evaluation on real-world reconstruction.", + "bbox": [ + 511, + 335, + 906, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 590, + 651, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. We compare our method against the state-of-the-art methods in 3D curve reconstruction, including RFEPS [47], NerVE [54], NEF [50] and EMAP [23]. The first two requires point clouds as input, while the last two and ours use only 2D images in the same input frames and resolution. We all adopt their pre-trained models and corresponding reconstruction approaches to extract curves for evaluation. More experiments, discussions and details can be found in the supplementary material.", + "bbox": [ + 511, + 613, + 903, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We first normalize and align all reconstructed curves and ground truth edges into the range of $[0, 1]$ . Then, we sample points on all reconstructed curves and measure the distances between these points and ground-truth edge points. When assessing IoU, precision, recall, and F-score, points are considered matched if there is at least one ground-truth edge point with an $L_{2}$ distance smaller than 0.02. We also adopt the Chamfer Distance (CD) between point clouds to measure the geometric accuracy of reconstructed parametric curves. As reported in", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5798", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f3a7fb84705df8b4e48c578c3f2c6b64c2de11bcc9fcbf23801346cd2c82565.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons. The first two objects come from ABC-NEF dataset and the last one comes from ModelNet dataset. Ours(SG) refers to our generated Spherical Gaussian in the first stage, GT refers to the ground-truth edge." + ], + "image_footnote": [], + "bbox": [ + 104, + 88, + 893, + 333 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ef5b7ab525c6cc83530aa91db90419b7422e211529514c7c50026a38a6e7812.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on ABC-NEF Dataset [50] and ModelNet Dataset[44]." + ], + "table_footnote": [], + "table_body": "
MethodABC-NEF DatasetModelNet DatasetInputTrainingReconstruction
CD↓Precision↑Recall↑F-score↑IoU↑CD↓Precision↑Recall↑F-score↑IoU↑ModalityTime↓Time↓
RFEPS [47]0.03230.89620.85600.86730.81900.07310.70320.59650.64120.5621Point Cloud-48s
NerVE [54]0.03890.95180.73070.82670.68450.02940.95680.80170.87240.7931Point Cloud>10h1s
NEF [50]0.03530.93870.88380.90440.82830.03210.90720.81580.85590.7475Images>1h119s
EMAP [23]0.02910.95130.89260.92100.84730.02900.91150.84880.87840.7751Images>2h40s
Ours0.02800.95460.90520.92600.86150.02640.96270.84150.89520.8164Images87s32s
", + "bbox": [ + 94, + 412, + 903, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1, our method outperforms other methods in all metrics on ABC-NEF [50] and most metrics on ModelNet [44].", + "bbox": [ + 89, + 545, + 482, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. We illustrate the visual comparisons in Fig. 6. It can be observed that RFEPS [47] and NerVE [54] may miss edges with weak curvature, leading to incomplete reconstruction. NEF [50] and EMAP [23] struggle with handling thin structures and make errors in curve connection. On the contrary, our method benefits from aligned Spherical Gaussians (see Ours(SG) in Fig. 6) and maintains a high level of robustness on 3D curve reconstruction.", + "bbox": [ + 89, + 579, + 483, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Comparison. We also make comparisons on training time and average reconstruction time (see the last two columns in Tab. 1). NerVE [54] is a learning-based methods that rely on large point cloud datasets, requiring extensive training time. RFEPS [47] is learning-free but takes time in point cloud optimization. NEF [50] and EMAP [23] are supervised on images, but they all involve a time-consuming training process (several hours for single scene) for predicting edge densities or unsigned distance functions. Our SGCR generates and optimizes only a small number of Spherical Gaussians (with two stages completing within minutes, $50 \\times$ faster than NEF and EMAP), making 3D", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "curve reconstruction quite efficient.", + "bbox": [ + 513, + 545, + 750, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Real-world Scene", + "text_level": 1, + "bbox": [ + 511, + 573, + 687, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further test our performance in real-world scenes on DTU [1] and Replica [37] dataset. They present challenging objects and complex scenes for edge reconstruction respectively. The visual comparisons against NEF [50] and EMAP [23] on DTU [1] dataset are presented in Fig. 7. NEF misses", + "bbox": [ + 511, + 595, + 908, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8806b33bc763145fe29e1f923c1fedbe9e139e270d54a08c2d8c28599efc70e7.jpg", + "image_caption": [ + "Figure 7. Comparisons on real-world scenes from DTU [1]." + ], + "image_footnote": [], + "bbox": [ + 514, + 700, + 901, + 872 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5799", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/906342259482a6d534a982ef1a34972564f054bd3e4a6ad66bf4e81166c34346.jpg", + "image_caption": [ + "Figure 8. Comparisons on real-world scenes from Replica [37]." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 228, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2e97f069488d28b0bfe006109edaa907b4a7c39c51aa3ad78e6bdb640854565.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 85, + 357, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bcfd47ae5971cffc3fb5547026ec96d05837a42ea07318e4df5ca4222c32e81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 85, + 483, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "details when fitting complex parts, and EMAP produces too many chaotic and redundant line segments, while our method reconstructs edges accurately and preserves details clearly. Another visual comparison on Replica [37] dataset are presented in Fig. 8. It can be seen that our method still works well on room-level edge reconstruction (NEF fails, and EMAP's training time for single real-world scene is $100 \\times$ slower than ours), demonstrating our method's great generalization ability and broad applicability.", + "bbox": [ + 89, + 248, + 483, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 393, + 254, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pipeline. We conduct ablation studies on ABC-NEF dataset [50] to evaluate the effectiveness of core designs in our method, including Spherical Gaussians, loss designs and two-phase training strategy in the first stage, as well as line fitting and global optimization in the second stage. Tab. 2 lists the quantitative results. w/o Spherical Gaussians stands for using original 3D Gaussians [19] trained by edge maps to directly represent edges and this result is quite poor, highlighting the significance of Spherical Gaussians. Statistics in Tab. 2 reveal that removing any component from our pipeline negatively impacts the final result.", + "bbox": [ + 89, + 416, + 483, + 583 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90b12519afb42447ad4f7aa43b738639f45d4950106d7082b91ed545ba6fec2e.jpg", + "table_caption": [ + "Table 2. Quantitative ablation studies on our pipeline." + ], + "table_footnote": [], + "table_body": "
PipelineCD↓F-score↑IoU↑
w/o Spherical Gaussians0.1420.3300.194
w/o grid initialization0.0320.9010.822
w/o edge loss Ledge0.0520.8530.714
w/o opacity-color loss Loc0.0510.8590.723
w/o regularization loss Regular0.0290.9130.846
w/o two-phase training strategy0.0300.9200.860
w/o line fitting0.0300.9110.822
w/o global optimization0.0320.8760.757
Full Model0.0280.9260.862
", + "bbox": [ + 91, + 621, + 483, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Radius. We analyze the effect of radius in Spherical Gaussians. We test $r_0 = 0.002$ , 0.005, 0.01 and variable. Experimental results are presented in Tab. 3 (setting $r_0$ as a variable does not perform well). As the radius increases, the total number of Spherical Gaussians decreases, with more", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a34a52c6e66d251b49160250155ef64627bccc30bd0136807541e11cc67f8b7b.jpg", + "table_caption": [ + "Table 3. Comparison on different radius choices for Spherical Gaussians. $N_{SG}$ : Number of Spherical Gaussians. T1: Time for Gaussian generation. T2: Time for curve reconstruction." + ], + "table_footnote": [], + "table_body": "
Radius r0CD↓F-score↑IoU ↑NSGT1T2
0.0020.0320.9060.8495,12484s56s
0.0050.0280.9260.8622,36190s36s
0.010.0350.8810.810592101s30s
Variable0.0400.8630.7853,076110s44s
", + "bbox": [ + 521, + 142, + 895, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cebe1863b9a84ef3dfd1c7eecddfae1ebb5c5ce504a3e369d7a4cd81659ceba.jpg", + "image_caption": [ + "Figure 9. Visualization for Gaussians with different radii." + ], + "image_footnote": [], + "bbox": [ + 514, + 253, + 906, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93135909d991fdb2acf1ab31d25db43f5bc162bd2f7a902fb52ee4ac17d8fa84.jpg", + "image_caption": [ + "Figure 10. From left to right, we present 2D images, rendered Spherical Gaussians and final curves in 5, 10, 30 and 50 views." + ], + "image_footnote": [], + "bbox": [ + 517, + 371, + 903, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "time spent in the first stage and less time in the second stage. The visual comparisons can be found in Fig. 9. An inappropriate radius size may lead to noise or underfitting. Using $r_0 = 0.005$ leads to the optimal reconstruction metrics as well as a reasonable running time.", + "bbox": [ + 511, + 539, + 905, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Views. We also study the effect of the number of input views. We train Spherical Gaussians with 5, 10, 30 and 50 views respectively (all evenly distributed). The visual results are illustrated in Fig. 10. For simple cases, results with 10 views are close to satisfactory, while for complex cases, more views are required for better performance.", + "bbox": [ + 511, + 619, + 906, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 734, + 638, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented SGCR—a novel method using Spherical Gaussians for efficient 3D parametric curve reconstruction from multi-view images. The main concept involves generating Spherical Gaussians by 2D supervision and extract 3D parametric curves from them. Our method outperforms previous methods quantitatively and qualitatively, while enjoying great efficiency. SGCR also demonstrates the potential in integrating multi-modal information efficiently, bridging the gap between 2D rendering and 3D reconstruction.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5800", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 62032011.", + "bbox": [ + 91, + 114, + 482, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 172 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2, 6, 7", + "[2] Hichem Abdellali, Robert Frohlich, Viktor Vilagos, and Zoltan Kato. L2d2: Learnable line detector and descriptor. In 2021 International Conference on 3D Vision (3DV), pages 442-452. IEEE, 2021. 3", + "[3] Nina Amenta, Marshall Bern, and David Eppstein. The crust and the $\\beta$ -skeleton: Combinatorial curve reconstruction. Graphical models and image processing, 60(2):125-135, 1998. 3", + "[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 3", + "[5] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 3", + "[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2", + "[7] Kris Demarsin, Denis Vanderstraeten, Tim Volodine, and Dirk Roose. Detection of closed sharp edges in point clouds using normal estimation and graph theory. Computer-Aided Design, 39(4):276-283, 2007. 3", + "[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6", + "[9] Tamal K Dey, Kurt Mehlhorn, and Edgar A Ramos. Curve reconstruction: Connecting dots with good reason. In Proceedings of the fifteenth annual symposium on Computational geometry, pages 197-206, 1999. 2, 3", + "[10] Chen Feng, Yuichi Taguchi, and Vineet R Kamat. Fast plane extraction in organized point clouds using agglomerative hierarchical clustering. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 6218-6225. IEEE, 2014. 3", + "[11] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 3", + "[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3", + "[13] Stefan Gumhold, Xinlong Wang, Rob S MacLeod, et al. Feature extraction from point clouds. In IMR, pages 293-305, 2001. 2" + ], + "bbox": [ + 93, + 183, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yijia He, Ji Zhao, Yue Guo, Wenhao He, and Kui Yuan. Plvio: Tightly-coupled monocular visual-inertial odometry using point and line features. Sensors, 18(4):1159, 2018. 3", + "[15] Chems-Eddine Himeur, Thibault Lejemble, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 3", + "[16] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3", + "[17] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. arXiv preprint arXiv:2403.17888, 2024. 2", + "[18] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2", + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 8", + "[20] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6", + "[21] G Saravana Kumar, Prem Kumar Kalra, and Sanjay G Dhande. Curve and surface reconstruction from points: an approach based on self-organizing maps. Applied Soft Computing, 5(1):55-66, 2004. 2, 3", + "[22] In-Kwon Lee. Curve reconstruction from unorganized points. Computer aided geometric design, 17(2):161-177, 2000. 3", + "[23] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6, 7", + "[24] Hyunjun Lim, Jinwoo Jeon, and Hyun Myung. Uv-slam: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters, 7(2):1518-1525, 2022. 3", + "[25] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3", + "[26] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and JanDirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. Cornell University - arXiv, Cornell University - arXiv, 2021. 2, 3", + "[27] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5801", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2, 3", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3", + "[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 3", + "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 3", + "[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 3", + "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[34] Alec Rivers, Frédo Durand, and Takeo Igarashi. 3d modeling with silhouettes. In Acm siggraph 2010 papers, pages 1-8. 2010. 3", + "[35] Ryan Schmidt, Azam Khan, Karan Singh, and Gord Kurtenbach. Analytic drawing of 3d scaffolds. In ACM SIGGRAPH Asia 2009 papers, pages 1-10. 2009. 3", + "[36] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism. ACM Transactions on Graphics, page 835-846, 2006. 3", + "[37] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8", + "[38] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3, 6", + "[39] Shuangbu Wang, Yu Xia, Lihua You, and Jianjun Zhang. Reconstruction of curve networks from unorganized spatial points. Journal of Universal Computer Science (J. UCS), 26 (9):1265-1280, 2020. 3", + "[40] Wenping Wang, Helmut Pottmann, and Yang Liu. Fitting b-spline curves to point clouds by curvature-based squared distance minimization. ACM Transactions on Graphics (ToG), 25(2):214-238, 2006. 3", + "[41] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net:" + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3", + "[42] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 3", + "[43] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 3", + "[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2, 6, 7", + "[45] Tianyi Xie, Zeshun Zong, Yuxin Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 3", + "[46] Baoxuan Xu, William Chang, Alla Sheffer, Adrien Bousseau, James McCrae, and Karan Singh. True2form: 3d curve networks from 2d sketches via selective regularization. ACM Transactions on Graphics, 33(4), 2014. 3", + "[47] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3, 6, 7", + "[48] Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing: From supervised to self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[49] Bisheng Yang and Yufu Zang. Automated registration of dense terrestrial laser-scanning point clouds using curves. ISPRS journal of photogrammetry and remote sensing, 95: 109-121, 2014. 3", + "[50] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8486-8495, 2023. 3, 6, 7, 8", + "[51] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 3", + "[52] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2", + "[53] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5802", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 7", + "[55] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + ], + "bbox": [ + 91, + 90, + 480, + 229 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5803", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_model.json b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a976c71661563b998f761eb47bc6299c7ebe7b6 --- /dev/null +++ b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.131, + 0.836, + 0.152 + ], + "angle": 0, + "content": "SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.18, + 0.816, + 0.218 + ], + "angle": 0, + "content": "Xinran Yang\\(^{1}\\), Donghao Ji\\(^{1}\\), Yuanqi Li\\(^{1}\\), Jie Guo\\(^{1}\\), Yanwen Guo\\(^{1*}\\), Junyuan Xie\\(^{1}\\) \n\\(^{1}\\)Nanjing University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.22, + 0.654, + 0.234 + ], + "angle": 0, + "content": "{xryang,donghaoji}@smail.nju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.237, + 0.673, + 0.253 + ], + "angle": 0, + "content": "{yuanqili,guojie,ywguo,jyxie}@nju.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.281, + 0.907, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.519, + 0.908, + 0.562 + ], + "angle": 0, + "content": "Figure 1. We present a novel method using 2D images to generate Spherical Gaussians, enabling the accurate representation of 3D edges. Spherical Gaussians combine the advantages from 3D Gaussian Splatting and the ability to preserve excellent 3D structures. With a corresponding curve extraction algorithm, Spherical Gaussians efficiently reconstruct 3D edge curves by only 2D supervision." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.574, + 0.327, + 0.59 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.606, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting technique has achieved high quality novel view synthesis as well as fast rendering speed. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is due to the fact that Gaussian's attributes are primarily tailored and fine-tuned for rendering diverse 2D images by their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries, from which we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians is optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and then compared to the ground-truth edge map extracted from the corresponding image, without the need for any 3D guidance or su" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.909, + 0.698 + ], + "angle": 0, + "content": "pervision. Given Spherical Gaussians serve as intermedi- dia for the robust edge representation, we further introduce a novel optimization-based algorithm called SGCR to directly extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying great efficiency. Code is available at https://github.com/Martinyxr/SGCR." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.646, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In computer graphics and vision, photorealistic novel view synthesis (NVS) and accurate geometry reconstruction are two significant objectives. 3D Gaussian Splatting [19] offers an appealing explicit representation—3D Gaussians, resulting in high-quality synthesis of novel views and real-time rendering speed. However, it falls short in capturing accurate 3D geometric structures since the anisotropic 3D Gaussians, which are intended for efficient rasterization and sufficient expression ability, do not align with the geometric features (surfaces, edges and corners) in a faithful manner. A question naturally arises: Can we utilize 3D Gaussians" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.245, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.274, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.212, + 0.274, + 0.227 + ], + "angle": 0, + "content": "(a) 3D Gaussian Splatting" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.088, + 0.482, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.213, + 0.479, + 0.226 + ], + "angle": 0, + "content": "(b) Spherical Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.24, + 0.483, + 0.269 + ], + "angle": 0, + "content": "Figure 2. Visualization for optimized Gaussians. (a) Gaussians from original 3D Gaussian Splatting. (b) Our Spherical Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.483, + 0.31 + ], + "angle": 0, + "content": "for the direct representation of intricate geometric structures while retaining their benefits in 2D rendering?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.431 + ], + "angle": 0, + "content": "Several recent works attempt to directly address the task of sdf generation or mesh reconstruction [6, 12, 17, 52] with 3D Gaussians. In this paper, we try to answer this question from a new perspective—reconstructing 3D parametric curves with 'Spherical Gaussians'. Following the settings of classic rendering works [19, 29], we use a set of calibrated multi-view images as input, but our goal substitutes novel view synthesis for 3D edge/curve reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Feature curves serve as critical geometric cues in characterizing the structure of 3D shapes, benefiting surface reconstruction, shape perception or other downstream tasks. Traditional methods [9, 13, 18, 21] usually work with the direct representation of 3D shapes, such as polygonal meshes or point clouds. However, the presence of sharp edges may be compromised or entirely overlooked on account of imperfect 3D scanning and acquisition, especially when dealing with sparse and noisy data. In recent years, learning based-methods [26, 28, 41, 54] are proposed to mitigate this, but obtaining precise ground-truth labels for 3D edge supervision involves laborious manual annotation, and inevitably limiting their generality. On the contrary, 2D images is easily accessible and allow for straightforward identification of feature edges through mature edge detection methods, making 2D supervision more attractive for edge processing tasks. Despite advantages in detection and self-supervision, there still remains a huge obstacle to overcome—establishing correspondence between edges across different views when working with multi-view images, in other words, how to combine information from multiple views to construct a complete 3D structure? This is exactly the factor for our design of Spherical Gaussians (as well as SGCR), which serve as a robust bridge to connect 3D depiction with 2D guidance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our method consists of two main stages: Spherical Gaussians generation and parametric curves extraction. During the first stage, we utilize 2D edge maps to refine Gaussians using a view-based rendering loss compared to the ground-truth edge map derived from the image of that view. Optimizing Gaussians directly as vanilla 3D Gaus" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.19, + 0.907, + 0.219 + ], + "angle": 0, + "content": "Figure 3. Illustration for fitting Bézier curve from Spherical Gaussians. \\( P_{1} \\), \\( P_{2} \\), \\( P_{3} \\) and \\( P_{4} \\) are weighted control points of the curve." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.426 + ], + "angle": 0, + "content": "sian Splatting [19] is problematic due to the sparse distribution of edge pixels and the diversity of Gaussian attributes. The outcome Gaussians may recover 2D images to some extent but exhibit redundant irregular ellipsoids (see Fig. 2 (a)) and lack solid 3D geometric meanings, making following reconstruction tasks quite difficult. Therefore, we implement several constraints to transform standard 3D Gaussians into Spherical Gaussians. We introduce grid initialization, fixing-covariance training and a two-phase pruning strategy to guide the optimization process. Our final distribution of gaussians lie faithfully on sharp edges of objects in 3D space, appearing the same shape like spheres (see Fig. 2 (b)) and holding attributes in opacity and gray-scale color." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.576 + ], + "angle": 0, + "content": "In the second stage, we utilize these discrete Spherical Gaussians to extract smooth 3D parametric curves for final reconstruction of edges. We present a novel optimization-based algorithm called SGCR, which involves fitting edges represented by Spherical Gaussians onto rational Bezier curves. The process begins with line fitting and then gradually refines the curves by global optimization. The final results are stored by all control points and weights in these curves. An illustration for curve extraction from Spherical Gaussians is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.698 + ], + "angle": 0, + "content": "We evaluate our method using a variety of object categories with complex models and scenes from ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. Extensive experiments show that SGCR, which utilizes Spherical Gaussians as bridge to reconstruct 3D parametric curves by 2D supervisions, outperforms existing state-of-the-art methods on all metrics and achieves great efficiency. In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.701, + 0.905, + 0.731 + ], + "angle": 0, + "content": "- We propose a new 3D edge extraction method from multiview 2D images using Spherical Gaussians (SG)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.732, + 0.905, + 0.776 + ], + "angle": 0, + "content": "- We introduce a new optimization-based algorithm (SGCR) tailored for directly reconstructing 3D parametric curves from our generated Spherical Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.778, + 0.905, + 0.807 + ], + "angle": 0, + "content": "- Our approach achieves the state-of-the-art performance as well as high efficiency on curve reconstruction tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.701, + 0.905, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.655, + 0.838 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.905, + 0.865 + ], + "angle": 0, + "content": "2.1. Image-based 3D Representation and Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Image-based rendering (IBR) techniques use a collection of 2D images of a scene to generate the visual represent-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "tation of the scene and render novel views. Significant advancements have been achieved by Neural Radiance Fields (NeRF) [29], which depicts a scene's geometry and visual attributes as a radiance field, enabling the retrieval of color and volume density at any point in space and from any viewing angle for the purpose of rendering. Recently, the unstructured GPU-optimized 3D Gaussian Splitting (3DGS) [19] presents the explicit Gaussian primitive that achieves enhanced rendering speeds and improved quality without the reliance on neural components. It has also been quickly extended to multiple domains [12, 27, 45, 53, 55]. We convert standard 3D Gaussians into Spherical Gaussians, enabling significant advantages in both view rendering and the representation of geometric structures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.315, + 0.323, + 0.33 + ], + "angle": 0, + "content": "2.2.3D Curve Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.534 + ], + "angle": 0, + "content": "The convention of 3D curve reconstruction is based on point cloud. Traditional methods [3, 9, 21, 22, 40] focus on local geometric properties of point clouds, such as normals [7, 42], curvatures [49] and hierarchical clustering [10]. Data-driven methods [15, 26, 28, 39] often adopt edge detection as a binary classification for point clouds. As network architectures advance, the classifier used for edge detection has become more complex models [5, 33, 41, 51]. RFEPS [47] presents a multistage optimization to augment the point set so that sufficient points are generated on potential edges of geometry. NerVE [54] presents a neural volumetric edge representation learned through a volumetric learning framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Extraction curves from multi-view images was pioneered by [4] which introduces a full SfM system using line segments. With the developments of line detections [30, 31, 48] and matching [2, 30, 32], several works [14, 16, 24, 25, 43] have attempted to revisit the line mapping problem, but they are restricted to straight lines and frequently generate fragmented small lines when dealing with curves. Lifting 2D sketches/curves to 3D curves [11, 34, 35, 46] has also been exploited, but the ability to generate parametric curves is still limited. Recently, NEF [50] learns a neural implicit field to capture the distribution of edge densities, and EMAP [23] encodes 3D edge distance and direction in Unsigned Distance Functions (UDF) from multi-view edge maps. However, they all require large training time for even a simple scene. In contrast, our method employs Spherical Gaussians as the direct representation for edges, achieving 3D parametric curve reconstruction both accurately and efficiently." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.182, + 0.837 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.22, + 0.864 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3D Gaussian Splatting. 3DGS [19] is a recently popular method for representing 3D scenes with 3D Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "primitives, achieving both explicit representation and high-quality real-time rendering. Each Gaussian primitive is defined by a covariance matrix \\(\\Sigma\\), and a center point \\(p\\) which is the mean of the Gaussian. The 3D distribution can be represented as: \\(G(x) = e^{-\\frac{1}{2} (x - p)^{\\mathrm{T}}\\Sigma^{-1}(x - p)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.289 + ], + "angle": 0, + "content": "A complete 3D Gaussian is given by its mean/center \\(p \\in \\mathbb{R}^3\\), color represented with spherical harmonics coefficients \\(\\mathbf{H} \\in \\mathbb{R}^k\\), opacity \\(\\alpha \\in \\mathbb{R}\\), quaternion \\(\\mathbf{r} \\in \\mathbb{R}^4\\), and scaling factor \\(\\mathbf{s} \\in \\mathbb{R}^3\\). For a given pixel, the combined color and opacity from multiple Gaussians are weighted by Eq. 3. The color blending for overlapping points is: \\(\\mathbf{C} = \\sum_{i \\in N} \\mathbf{c}_i \\alpha_i \\prod_{j=1}^{i-1} (1 - \\alpha_j)\\) where \\(\\mathbf{c}_i, \\alpha_i\\) denote the color and density of a point." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.485 + ], + "angle": 0, + "content": "Spherical Gaussians. Our Spherical Gaussians enjoy the same nature of 3D Gaussians, but make the following changes for geometry priority and simplicity: (1) Remove the covariance matrix \\(\\pmb{\\Sigma}\\) of Gaussians (attributes of scaling \\(\\mathbf{S}\\) and rotation \\(\\mathbf{R}\\)) and replace it with a fixed radius \\(r_0\\), i.e. convert Gaussian primitives from various ellipsoids into spheres of uniform size. (2) Remove spherical harmonics coefficients \\(\\mathbf{H}\\) and modify color to one-dimensional gray value (just for edge map rendering). According to our experiments, these changes do not hamper the backward propagation of gradients from 2D edge-map supervision, but impose great regularization on geometry distribution of 3D isotropic Gaussians, facilitating reconstruction tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.698 + ], + "angle": 0, + "content": "Anisotropic Gaussians v.s. Isotropic Gaussians. Traditional 3DGS [19] and its subsequent works mainly center on the technique of 'splatting'—utilizing different anisotropic Gaussian splats to accurately represent the entire scene. However, our work focuses on each 'Gaussian' itself, essentially displaying a reverse process: decomposing the splating results into small 'atoms' and assigning clear geometric meanings to every individual Gaussian. A thin and elongated Gaussian seems to be more efficient in representing edges, but it can be split into smaller 'basis-Gaussians' with few sacrifices but for better interpretation and convenience, especially when following pointcloud-related works. That's why we favour Spherical Gaussians and more discussions can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.796, + 0.724 + ], + "angle": 0, + "content": "3.2. Spherical Gaussians Generation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.65, + 0.744 + ], + "angle": 0, + "content": "3.2.1. Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our full pipeline is illustrated in Fig. 4. Given a set of calibrated multi-view images, we first adopt edge detection method [38] to generate 2D edge maps for each view. 3DGS uses structure-from-Motion (SfM) [36] points as initialization for Gaussians, but this approach is not ideal for handling edge maps. Thus, we design grid initialization for Spherical Gaussians. We create a \\(50 \\times 50 \\times 50\\) grid with points evenly distributed in space of \\([0,1]^3\\) (see Fig. 4—Grid initialization) as the initial centers for our Spherical Gaussians. This setting is better than random initialization" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.356, + 0.905, + 0.371 + ], + "angle": 0, + "content": "Figure 4. Pipeline illustration across 2D and 3D modality. Stage1: Spherical Gaussians generation. Stage2: Parametric curves extraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.39, + 0.483, + 0.451 + ], + "angle": 0, + "content": "in producing neat structures. And we choose a constant radius \\( r_0 \\) to replace the covariance matrix with Spherical Gaussians of the same size. In our experiments, we use \\( r_0 = 0.005 \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.457, + 0.248, + 0.472 + ], + "angle": 0, + "content": "3.2.2. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.506 + ], + "angle": 0, + "content": "Edge Loss. The original loss functions in 3DGS [19] to train 3D Gaussians for each view is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.516, + 0.483, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {D - S S I M} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.691 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_1\\) and \\(\\mathcal{L}_{D-SSIM}\\) are supervised by the ground-truth view. However, in edge maps, edge pixels are often sparsely distributed. Relying solely on this loss function will easily make the training process stuck in local optima, causing the opacities and colors for all Gaussians to converge towards zero and resulting in all rendered images becoming black. Hence, we opt for a new edge-aware loss \\(\\mathcal{L}_{edge}\\) instead of \\(\\mathcal{L}_1\\) in order to balance the impact of both edge and non-edge pixels within each edge map. The edge-aware loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e d g e}} = \\frac {N _ {I} - | E _ {I} |}{N _ {I}} \\sum_ {i \\in E _ {I}} \\| I _ {i} - \\hat {I} _ {i} \\| _ {2} ^ {2} + \\frac {| E _ {I} |}{N _ {I}} \\sum_ {i \\notin E _ {I}} \\| I _ {i} - \\hat {I} _ {i} \\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.841 + ], + "angle": 0, + "content": "where \\( N_{I} \\) is the image resolution and \\( E_{I} \\) refers to all edge pixels in current edge map (gray scale larger than threshold \\( \\eta = 0.3 \\)), \\( I_{i} \\) and \\( \\hat{I}_i \\) refers to the pixel in Gaussian-rendering image and ground-truth edge-map image respectively. This edge-aware loss enhances the ability of Spherical Gaussians to identify edges from edge pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Opacity-Color Loss. Another issue during edge training is about occluded edges. Some 2D edge maps may lose true edge pixels due to occlusions from these image views (see the difference between Fig. 4-Edge maps and Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.542 + ], + "angle": 0, + "content": "views), while from other views the occlusion may not happen. This inconsistency among views will confuse Spherical Gaussians during training, potentially leading to unstable optimization. During training, color attribute is directly supervised by \\(\\mathcal{L}_{edge}\\) loss, while opacity is indirectly optimized and reset at certain iterations. This difference sets them apart, and equating them impedes smooth optimization. Therefore, we introduce opacity-color loss \\(\\mathcal{L}_{oc}\\) for all Spherical Gaussians to make the value of Gaussian opacity more consistent. This opacity-color loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.554, + 0.905, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o c} = \\sum_ {i = 1} ^ {N _ {S G}} \\| o _ {i} - c _ {i} \\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.79 + ], + "angle": 0, + "content": "where \\(N_{SG}\\) is the number of all Spherical Gaussians during current training step, \\(o_i\\) and \\(c_i\\) refer to opacity and color attribute of each Spherical Gaussian respectively. \\(\\mathcal{L}_{oc}\\) helps to keep consistency between edge density and color intensity during multi-view supervision. This mechanism ensures that occluded edges—represented by Gaussians with low opacity values—are less susceptible to premature pruning during optimization, resulting in more complete 3D reconstructions. It also helps rectify edge predictions missed by 2D detectors. In our experiment, after implementing \\(\\mathcal{L}_{oc}\\), Gaussian-rendered images can depict complete edge structures with no occlusion (Fig. 4—Rendering Views)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.907, + 0.85 + ], + "angle": 0, + "content": "Regularization Loss. To restrict total number of Spherical Gaussians and accelerate convergence, we add an extra regularization term \\(\\mathcal{L}_{\\text {regular }}\\) on Gaussian's opacity. The regularization loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.863, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\sum_ {i = 1} ^ {N _ {S G}} \\log \\left(1 + \\frac {o _ {i} ^ {2}}{0 . 5}\\right) \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.247, + 0.473, + 0.262 + ], + "angle": 0, + "content": "Figure 5. The change on number of Gaussians during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.484, + 0.361 + ], + "angle": 0, + "content": "where \\( N_{SG} \\) is the number of current Spherical Gaussians and \\( o_i \\) is each Gaussian's opacity. During the training process, this loss function works implicitly because Spherical Gaussians with opacity below a certain threshold will be pruned (as in 3DGS [19]), allowing for a comfortable number of total Gaussian primitives for next stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.392 + ], + "angle": 0, + "content": "To summarize, our final loss function for training Spherical Gaussians is presented as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.404, + 0.483, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left(1 - \\lambda_ {1}\\right) \\mathcal {L} _ {\\text {e d g e}} + \\lambda_ {1} \\mathcal {L} _ {D - S S I M} + \\lambda_ {2} \\mathcal {L} _ {o c} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e g}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.464 + ], + "angle": 0, + "content": "where the balancing parameters \\(\\lambda_1, \\lambda_2, \\lambda_3\\) are set to 0.2, 2 and 0.01 in our experiments respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.472, + 0.262, + 0.487 + ], + "angle": 0, + "content": "3.2.3. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.763 + ], + "angle": 0, + "content": "The effective training strategy is essential for creating properly structured Spherical Gaussians. We adopt a two-phase training strategy, each phase consists of 3,000 iterations. During the first phase, we conduct densification (split and clone) of Spherical Gaussians every 200 iterations. And we reset the opacity attribute to 0.1 every 1,000 iterations. At the end of first phase, we design a final-prune—removing all Spherical Gaussians whose opacity \\( o_i < 0.5 \\) and color \\( c_i < 0.1 \\). This operation largely reduces the total number of Gaussians, leading to substantial reductions in both training time and noise. During the second phase, we no longer densify Spherical Gaussians but focus on refining their positions and attributes. In the end, we perform another final-prune to retain all significant Spherical Gaussians (at quantity of 1k to 5k) with their attributes. The variation of numbers for Spherical Gaussians during all iterations is depicted in Fig. 5. The whole training process of Spherical Gaussians for single scene takes only about 1 minute." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.484, + 0.803 + ], + "angle": 0, + "content": "3.3. Parametric Curve Extraction from Spherical Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "After the first stage, we get the generated Spherical Gaussians (denoted as \\( G \\)) aligned with object's edges, resembling a 3D wireframe constructed by many small spheres. In the second stage (Fig. 4—Stage 2), we present of Spherical-Gaussians Curve Reconstruction (SGCR) to accomplish the final 3D edge reconstruction task. SGCR is an" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.092, + 0.686, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Line Fitting" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.109, + 0.871, + 0.123 + ], + "angle": 0, + "content": "Input: a set of Spherical Gaussians \\(G\\) with radius \\(r_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.124, + 0.905, + 0.155 + ], + "angle": 0, + "content": "Output: the set of line endpoints \\( L = \\{(p_i, q_i)\\}_{i=1}^{|L|} \\) for the coarse line-fitting of 3D edges." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.155, + 0.905, + 0.185 + ], + "angle": 0, + "content": "Denote: \\( P(G) \\) is the centers of \\( G \\), and \\( N(0,1) \\) is the Standard Normal Distribution." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.189, + 0.589, + 0.202 + ], + "angle": 0, + "content": "1: \\(L = \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.204, + 0.674, + 0.218 + ], + "angle": 0, + "content": "2: while \\( |G| > N_0 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.219, + 0.7, + 0.234 + ], + "angle": 0, + "content": "3: \\( S_{min} = \\) MaxIntValue" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.235, + 0.676, + 0.247 + ], + "angle": 0, + "content": "4: for \\( i = 1 \\) to \\( n \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.249, + 0.786, + 0.264 + ], + "angle": 0, + "content": "5: \\(p, q = \\text{RandomChoice}(P(G))\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.264, + 0.873, + 0.279 + ], + "angle": 0, + "content": "6: \\(L_{p,q} = \\text{interpolate}(p, q, N_s) + r_0 * N(0,1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.279, + 0.846, + 0.302 + ], + "angle": 0, + "content": "7: \\(G_{i} = \\{g\\in G|\\min_{x\\in L_{p,q}}||P(g) - x||_{2}^{2} < \\delta_{1}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.302, + 0.81, + 0.326 + ], + "angle": 0, + "content": "8: \\(p', q' = \\operatorname*{arg min}_{p,q} \\mathcal{L}_{CD}(L_{p,q}, P(G_i))\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.326, + 0.768, + 0.347 + ], + "angle": 0, + "content": "9: \\(S_{i} = \\min_{p,q}\\mathcal{L}_{CD}(L_{p,q},P(G_{i}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.347, + 0.7, + 0.36 + ], + "angle": 0, + "content": "10: if \\( S_{i} \\leq S_{min} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.361, + 0.843, + 0.376 + ], + "angle": 0, + "content": "11: \\(S_{min} = S_i, G' = G_i, (\\hat{p}, \\hat{q}) = (p', q')\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.377, + 0.62, + 0.389 + ], + "angle": 0, + "content": "12: end if" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.392, + 0.613, + 0.404 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.407, + 0.774, + 0.422 + ], + "angle": 0, + "content": "14: \\(G = G - G^{\\prime},L = L\\cup \\{(\\hat{p},\\hat{q})\\} ;\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.423, + 0.613, + 0.434 + ], + "angle": 0, + "content": "15: end while" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.437, + 0.614, + 0.45 + ], + "angle": 0, + "content": "16: return \\( L \\);" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.189, + 0.873, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.905, + 0.512 + ], + "angle": 0, + "content": "optimization-based algorithm guided by \\(G\\) which includes two parts: line fitting and global optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.644, + 0.535 + ], + "angle": 0, + "content": "3.3.1. Line Fitting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.628 + ], + "angle": 0, + "content": "We first fit a set of line segments to \\( G \\)'s centers as the coarse fitting. The pseudo code of this part is shown in Algorithm 1. The whole process consists of several line-fitting turns and \\( G \\) will be gradually fitted and partly removed by one line each turn. Line fitting will stop when the number of remaining \\( G \\) is less than \\( N_0 = 5 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.765 + ], + "angle": 0, + "content": "In each turn, we conduct \\( n \\) random searches and record the best result. At the beginning of \\( i \\)-th search, we randomly choose two nearest Gaussian centers as initial endpoints \\( p, q \\) of the fitting line, and interpolate \\( N_{s} \\) points along the line. To simulate the shape of Spherical Gaussians, we dilate the sampled points up by duplicating and adding Gaussian noise with standard deviation equals \\( G \\)'s radius \\( r_0 \\). Then we apply Chamfer Distance to compute the fitting loss between dilated line points and \\( G \\). This Chamfer loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.786, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C D} \\left(L _ {p, q}, P \\left(G _ {i}\\right)\\right) = \\frac {\\gamma_ {1}}{\\left| L _ {p , q} \\right|} \\sum_ {x \\in L _ {p, q}} \\min _ {y \\in P \\left(G _ {i}\\right)} \\| x - y \\| _ {2} ^ {2} \\\\ + \\frac {1}{| P (G _ {i}) |} \\sum_ {y \\in P (G _ {i})} \\min _ {x \\in L _ {p, q}} \\| x - y \\| _ {2} ^ {2} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.878, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nG _ {i} = \\left\\{g \\in G \\mid \\min _ {x \\in L _ {p, q}} \\left| | P (g) - x \\| _ {2} ^ {2} < \\delta_ {1} \\right. \\right\\} \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5797" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "where \\(L_{p,q}\\) denotes the dilated line points between \\(p, q\\) in \\(i\\)-th search, \\(P(G)\\) represents Gaussian centers of all current \\(G\\) and \\(\\gamma_1 = 2\\) is a balance parameter. \\(G_i\\) is a subset of \\(G\\) whose centers are close to \\(L_{p,q}\\) within a threshold \\(\\delta_1 = 0.02\\). We optimize \\(p, q\\) by minimizing Eq. (6) with certain iterations and use the value of \\(\\mathcal{L}_{CD}\\) after optimization as the final fitting score \\(S_i\\) for \\(i\\)-th search." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.306 + ], + "angle": 0, + "content": "For each turn, suppose the best fitting is in \\(k\\)-th search, i.e. \\(S_{k} = \\min S_{i}\\), we record the optimized endpoints \\(p^{\\prime}, q^{\\prime}\\) in \\(k\\)-th search and delete \\(G_{k}\\) to update the current \\(G\\). The next turn will then proceed accordingly. Once all the turns are completed, we will get a set of paired line endpoints \\(L = \\{(p_{i}, q_{i})\\}_{i=1}^{|L|}\\) which together form an approximate line-fitting result (Fig. 4—Line Fitting)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.281, + 0.326 + ], + "angle": 0, + "content": "3.3.2. Global Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.482, + 0.406 + ], + "angle": 0, + "content": "After line fitting, we resume all \\( G \\) deleted in previous part and take opacity attribute into account as the fine fitting. We interpolate another two control points between each line endpoint pair \\( (p_i, q_i) \\) and introduce point weight to initialize 3rd order rational Bézier curves, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.415, + 0.482, + 0.455 + ], + "angle": 0, + "content": "\\[\nB (u) _ {p, w} = \\frac {\\sum_ {i = 0} ^ {3} B _ {3 , i} (u) p _ {i} w _ {i}}{\\sum_ {i = 0} ^ {3} B _ {3 , i} (u) w _ {i}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.551 + ], + "angle": 0, + "content": "where \\( B_{3,i} \\) is basis functions of Bézier curves, \\( p_i \\) is the \\( i \\)-th control points of curve and \\( w_i \\) is the weight of control point \\( p_i \\) (\\( w_i \\equiv 1 \\) for simple Bézier curves). We opt for the rational Bézier curves over simple Bézier curves because simple Bézier curves cannot perfectly fit circles, which are frequently seen in object's edges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.627 + ], + "angle": 0, + "content": "We also sample \\( N_{s} \\) points for each Bézier curve and dilate them up as in line fitting. Since \\( G \\) with higher opacity values are more significant during fine fitting, we apply the weighted Chamfer Distance to compute the fine fitting loss between global curve points and all \\( G \\):" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.636, + 0.495, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {W C D} (L _ {a l l}, P (G)) = \\frac {\\gamma_ {2}}{| L _ {a l l} |} \\sum_ {x \\in L _ {a l l}} o _ {y _ {m i n}} \\cdot \\min _ {y \\in P (G)} \\| x - y \\| _ {2} ^ {2} \\\\ + \\frac {1}{| P (G) |} \\sum_ {y \\in P (G)} o _ {y} \\cdot \\min _ {x \\in L _ {\\text {a l l}}} \\| x - y \\| _ {2} ^ {2} \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where \\( L_{all} \\) refers to global dilated points from all curves and \\( P(G) \\) refers to Gaussian centers of \\( G \\), \\( o_y \\) is the opacity attribute in \\( y \\)-th \\( G \\) and \\( \\gamma_2 = 2 \\) is another balance parameter. We also adopt endpoints loss to encourage connections between curves, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.804, + 0.493, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n d p o i n t s}} = \\sum_ {x, y \\in P _ {E}} \\min \\left(\\| x - y \\| _ {2} ^ {2} - \\delta_ {2}, 0\\right) \\frac {\\| x - y \\| _ {2} ^ {2}}{\\| x - y \\| _ {2} ^ {2} - \\delta_ {2}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(P_{E}\\) is the set of endpoints (the first and last control points) and \\(\\delta_{2}\\) (we choose 0.01) is a threshold to only regularize endpoints which are close to each other. The final" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "objective function to globally optimize all curves is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.124, + 0.905, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\arg \\min \\quad \\left(\\mathcal {L} _ {W C D} + \\lambda \\mathcal {L} _ {\\text {e n d p o i n t s}}\\right) \\tag {11} \\\\ \\{\\{z _ {i} ^ {j}, w _ {i} ^ {j} \\} _ {j = 1} ^ {4} \\} _ {i = 1} ^ {| L |} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.165, + 0.905, + 0.227 + ], + "angle": 0, + "content": "where \\( z_{i}^{j} \\) and \\( w_{i}^{j} \\) are coordinates and weights of four control points in each Bézier curve, \\( \\lambda \\) is a balancing parameter which is set to 0.005 in our experiments. The pseudo code of this part can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.273 + ], + "angle": 0, + "content": "After global optimization, we finally extract 3D parametric curves from Spherical Gaussians in form of 3rd order rational Bezier curves (Fig. 4—Rational Bezier Curves)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.645, + 0.306 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.609, + 0.329 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.579 + ], + "angle": 0, + "content": "We conduct experiments on the ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. ABC [20] consists of massive CAD models with edge annotations, and we use a portion of it—ABC-NEF, provided by [50]—which consisting of 115 distinct and challenging CAD models. ModelNet [44] includes 3D models of everyday objects such as chairs and tables, but it does not have ground-truth edge annotations. So we manually select and annotate edges on ModelNet ground-truth meshes. Our edge dataset consists of 120 furniture models with annotated edges and is used for a more comprehensive evaluation. We utilize BlenderProc [8] to generate rendered images with these objects. We generate 50 views of \\(800 \\times 800\\) images for each object and adopt PiDiNet [38] to predict 2D edge maps. DTU [1] and Replica [37] are real-world datasets with multi-view images as input and we use it for evaluation on real-world reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.653, + 0.607 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.749 + ], + "angle": 0, + "content": "Settings. We compare our method against the state-of-the-art methods in 3D curve reconstruction, including RFEPS [47], NerVE [54], NEF [50] and EMAP [23]. The first two requires point clouds as input, while the last two and ours use only 2D images in the same input frames and resolution. We all adopt their pre-trained models and corresponding reconstruction approaches to extract curves for evaluation. More experiments, discussions and details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. We first normalize and align all reconstructed curves and ground truth edges into the range of \\([0, 1]\\). Then, we sample points on all reconstructed curves and measure the distances between these points and ground-truth edge points. When assessing IoU, precision, recall, and F-score, points are considered matched if there is at least one ground-truth edge point with an \\(L_{2}\\) distance smaller than 0.02. We also adopt the Chamfer Distance (CD) between point clouds to measure the geometric accuracy of reconstructed parametric curves. As reported in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.106, + 0.089, + 0.895, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons. The first two objects come from ABC-NEF dataset and the last one comes from ModelNet dataset. Ours(SG) refers to our generated Spherical Gaussian in the first stage, GT refers to the ground-truth edge." + }, + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.388, + 0.773, + 0.406 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on ABC-NEF Dataset [50] and ModelNet Dataset[44]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.414, + 0.905, + 0.528 + ], + "angle": 0, + "content": "
MethodABC-NEF DatasetModelNet DatasetInputTrainingReconstruction
CD↓Precision↑Recall↑F-score↑IoU↑CD↓Precision↑Recall↑F-score↑IoU↑ModalityTime↓Time↓
RFEPS [47]0.03230.89620.85600.86730.81900.07310.70320.59650.64120.5621Point Cloud-48s
NerVE [54]0.03890.95180.73070.82670.68450.02940.95680.80170.87240.7931Point Cloud>10h1s
NEF [50]0.03530.93870.88380.90440.82830.03210.90720.81580.85590.7475Images>1h119s
EMAP [23]0.02910.95130.89260.92100.84730.02900.91150.84880.87840.7751Images>2h40s
Ours0.02800.95460.90520.92600.86150.02640.96270.84150.89520.8164Images87s32s
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Table 1, our method outperforms other methods in all metrics on ABC-NEF [50] and most metrics on ModelNet [44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Qualitative Comparison. We illustrate the visual comparisons in Fig. 6. It can be observed that RFEPS [47] and NerVE [54] may miss edges with weak curvature, leading to incomplete reconstruction. NEF [50] and EMAP [23] struggle with handling thin structures and make errors in curve connection. On the contrary, our method benefits from aligned Spherical Gaussians (see Ours(SG) in Fig. 6) and maintains a high level of robustness on 3D curve reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Time Comparison. We also make comparisons on training time and average reconstruction time (see the last two columns in Tab. 1). NerVE [54] is a learning-based methods that rely on large point cloud datasets, requiring extensive training time. RFEPS [47] is learning-free but takes time in point cloud optimization. NEF [50] and EMAP [23] are supervised on images, but they all involve a time-consuming training process (several hours for single scene) for predicting edge densities or unsigned distance functions. Our SGCR generates and optimizes only a small number of Spherical Gaussians (with two stages completing within minutes, \\(50 \\times\\) faster than NEF and EMAP), making 3D" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.546, + 0.751, + 0.562 + ], + "angle": 0, + "content": "curve reconstruction quite efficient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.688, + 0.589 + ], + "angle": 0, + "content": "4.3. Real-world Scene" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.675 + ], + "angle": 0, + "content": "We further test our performance in real-world scenes on DTU [1] and Replica [37] dataset. They present challenging objects and complex scenes for edge reconstruction respectively. The visual comparisons against NEF [50] and EMAP [23] on DTU [1] dataset are presented in Fig. 7. NEF misses" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.701, + 0.902, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.882, + 0.896, + 0.898 + ], + "angle": 0, + "content": "Figure 7. Comparisons on real-world scenes from DTU [1]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5799" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.229, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.087, + 0.359, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.087, + 0.484, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.225, + 0.483, + 0.24 + ], + "angle": 0, + "content": "Figure 8. Comparisons on real-world scenes from Replica [37]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.385 + ], + "angle": 0, + "content": "details when fitting complex parts, and EMAP produces too many chaotic and redundant line segments, while our method reconstructs edges accurately and preserves details clearly. Another visual comparison on Replica [37] dataset are presented in Fig. 8. It can be seen that our method still works well on room-level edge reconstruction (NEF fails, and EMAP's training time for single real-world scene is \\(100 \\times\\) slower than ours), demonstrating our method's great generalization ability and broad applicability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.395, + 0.255, + 0.41 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Pipeline. We conduct ablation studies on ABC-NEF dataset [50] to evaluate the effectiveness of core designs in our method, including Spherical Gaussians, loss designs and two-phase training strategy in the first stage, as well as line fitting and global optimization in the second stage. Tab. 2 lists the quantitative results. w/o Spherical Gaussians stands for using original 3D Gaussians [19] trained by edge maps to directly represent edges and this result is quite poor, highlighting the significance of Spherical Gaussians. Statistics in Tab. 2 reveal that removing any component from our pipeline negatively impacts the final result." + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.597, + 0.446, + 0.612 + ], + "angle": 0, + "content": "Table 2. Quantitative ablation studies on our pipeline." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.622, + 0.485, + 0.81 + ], + "angle": 0, + "content": "
PipelineCD↓F-score↑IoU↑
w/o Spherical Gaussians0.1420.3300.194
w/o grid initialization0.0320.9010.822
w/o edge loss Ledge0.0520.8530.714
w/o opacity-color loss Loc0.0510.8590.723
w/o regularization loss Regular0.0290.9130.846
w/o two-phase training strategy0.0300.9200.860
w/o line fitting0.0300.9110.822
w/o global optimization0.0320.8760.757
Full Model0.0280.9260.862
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Radius. We analyze the effect of radius in Spherical Gaussians. We test \\( r_0 = 0.002 \\), 0.005, 0.01 and variable. Experimental results are presented in Tab. 3 (setting \\( r_0 \\) as a variable does not perform well). As the radius increases, the total number of Spherical Gaussians decreases, with more" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 3. Comparison on different radius choices for Spherical Gaussians. \\( N_{SG} \\): Number of Spherical Gaussians. T1: Time for Gaussian generation. T2: Time for curve reconstruction." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.143, + 0.897, + 0.235 + ], + "angle": 0, + "content": "
Radius r0CD↓F-score↑IoU ↑NSGT1T2
0.0020.0320.9060.8495,12484s56s
0.0050.0280.9260.8622,36190s36s
0.010.0350.8810.810592101s30s
Variable0.0400.8630.7853,076110s44s
" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.255, + 0.907, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.336, + 0.88, + 0.351 + ], + "angle": 0, + "content": "Figure 9. Visualization for Gaussians with different radii." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.488, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 10. From left to right, we present 2D images, rendered Spherical Gaussians and final curves in 5, 10, 30 and 50 views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.906, + 0.616 + ], + "angle": 0, + "content": "time spent in the first stage and less time in the second stage. The visual comparisons can be found in Fig. 9. An inappropriate radius size may lead to noise or underfitting. Using \\( r_0 = 0.005 \\) leads to the optimal reconstruction metrics as well as a reasonable running time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.71 + ], + "angle": 0, + "content": "Views. We also study the effect of the number of input views. We train Spherical Gaussians with 5, 10, 30 and 50 views respectively (all evenly distributed). The visual results are illustrated in Fig. 10. For simple cases, results with 10 views are close to satisfactory, while for complex cases, more views are required for better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.64, + 0.751 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We have presented SGCR—a novel method using Spherical Gaussians for efficient 3D parametric curve reconstruction from multi-view images. The main concept involves generating Spherical Gaussians by 2D supervision and extract 3D parametric curves from them. Our method outperforms previous methods quantitatively and qualitatively, while enjoying great efficiency. SGCR also demonstrates the potential in integrating multi-modal information efficiently, bridging the gap between 2D rendering and 3D reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5800" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.115, + 0.483, + 0.145 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62032011." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.188, + 0.174 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.483, + 0.239 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.241, + 0.483, + 0.294 + ], + "angle": 0, + "content": "[2] Hichem Abdellali, Robert Frohlich, Viktor Vilagos, and Zoltan Kato. L2d2: Learnable line detector and descriptor. In 2021 International Conference on 3D Vision (3DV), pages 442-452. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[3] Nina Amenta, Marshall Bern, and David Eppstein. The crust and the \\(\\beta\\)-skeleton: Combinatorial curve reconstruction. Graphical models and image processing, 60(2):125-135, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[7] Kris Demarsin, Denis Vanderstraeten, Tim Volodine, and Dirk Roose. Detection of closed sharp edges in point clouds using normal estimation and graph theory. Computer-Aided Design, 39(4):276-283, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[9] Tamal K Dey, Kurt Mehlhorn, and Edgar A Ramos. Curve reconstruction: Connecting dots with good reason. In Proceedings of the fifteenth annual symposium on Computational geometry, pages 197-206, 1999. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[10] Chen Feng, Yuichi Taguchi, and Vineet R Kamat. Fast plane extraction in organized point clouds using agglomerative hierarchical clustering. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 6218-6225. IEEE, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[11] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[13] Stefan Gumhold, Xinlong Wang, Rob S MacLeod, et al. Feature extraction from point clouds. In IMR, pages 293-305, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[14] Yijia He, Ji Zhao, Yue Guo, Wenhao He, and Kui Yuan. Plvio: Tightly-coupled monocular visual-inertial odometry using point and line features. Sensors, 18(4):1159, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Chems-Eddine Himeur, Thibault Lejemble, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[17] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. arXiv preprint arXiv:2403.17888, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[18] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[20] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] G Saravana Kumar, Prem Kumar Kalra, and Sanjay G Dhande. Curve and surface reconstruction from points: an approach based on self-organizing maps. Applied Soft Computing, 5(1):55-66, 2004. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[22] In-Kwon Lee. Curve reconstruction from unorganized points. Computer aided geometric design, 17(2):161-177, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[23] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Hyunjun Lim, Jinwoo Jeon, and Hyun Myung. Uv-slam: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters, 7(2):1518-1525, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[25] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[26] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and JanDirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. Cornell University - arXiv, Cornell University - arXiv, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[27] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5801" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[28] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[34] Alec Rivers, Frédo Durand, and Takeo Igarashi. 3d modeling with silhouettes. In Acm siggraph 2010 papers, pages 1-8. 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[35] Ryan Schmidt, Azam Khan, Karan Singh, and Gord Kurtenbach. Analytic drawing of 3d scaffolds. In ACM SIGGRAPH Asia 2009 papers, pages 1-10. 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[36] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism. ACM Transactions on Graphics, page 835-846, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[37] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[38] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[39] Shuangbu Wang, Yu Xia, Lihua You, and Jianjun Zhang. Reconstruction of curve networks from unorganized spatial points. Journal of Universal Computer Science (J. UCS), 26 (9):1265-1280, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[40] Wenping Wang, Helmut Pottmann, and Yang Liu. Fitting b-spline curves to point clouds by curvature-based squared distance minimization. ACM Transactions on Graphics (ToG), 25(2):214-238, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[41] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[42] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[43] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[45] Tianyi Xie, Zeshun Zong, Yuxin Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[46] Baoxuan Xu, William Chang, Alla Sheffer, Adrien Bousseau, James McCrae, and Karan Singh. True2form: 3d curve networks from 2d sketches via selective regularization. ACM Transactions on Graphics, 33(4), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.504 + ], + "angle": 0, + "content": "[47] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.575 + ], + "angle": 0, + "content": "[48] Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing: From supervised to self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[49] Bisheng Yang and Yufu Zang. Automated registration of dense terrestrial laser-scanning point clouds using curves. ISPRS journal of photogrammetry and remote sensing, 95: 109-121, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[50] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8486-8495, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[51] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[52] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[53] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[54] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[55] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5803" + } + ] +] \ No newline at end of file diff --git a/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_origin.pdf b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..759268f29b813d7caf092d18a8bd2a840d91a1ec --- /dev/null +++ b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/b3444c69-4274-4c13-82cc-b440a9e78c57_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b4aae427db7ab5cac7e5f49b6dbfcc9b23c9c1c3a0bfa8fe290542a81dc0d66 +size 9444046 diff --git a/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/full.md b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3671e4a42c1a262635b26b3bb94fc2e472eca22b --- /dev/null +++ b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/full.md @@ -0,0 +1,357 @@ +# SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction + +Xinran Yang $^{1}$ , Donghao Ji $^{1}$ , Yuanqi Li $^{1}$ , Jie Guo $^{1}$ , Yanwen Guo $^{1*}$ , Junyuan Xie $^{1}$ $^{1}$ Nanjing University, Nanjing, China + +{xryang,donghaoji}@smail.nju.edu.cn + +{yuanqili,guojie,ywguo,jyxie}@nju.edu.cn + +![](images/ec9ce4293209621d31fd801112fd0f36b864b75d1209a8aae9b181ffcc5198cf.jpg) +Figure 1. We present a novel method using 2D images to generate Spherical Gaussians, enabling the accurate representation of 3D edges. Spherical Gaussians combine the advantages from 3D Gaussian Splatting and the ability to preserve excellent 3D structures. With a corresponding curve extraction algorithm, Spherical Gaussians efficiently reconstruct 3D edge curves by only 2D supervision. + +# Abstract + +Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting technique has achieved high quality novel view synthesis as well as fast rendering speed. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is due to the fact that Gaussian's attributes are primarily tailored and fine-tuned for rendering diverse 2D images by their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries, from which we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians is optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and then compared to the ground-truth edge map extracted from the corresponding image, without the need for any 3D guidance or su + +pervision. Given Spherical Gaussians serve as intermedi- dia for the robust edge representation, we further introduce a novel optimization-based algorithm called SGCR to directly extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying great efficiency. Code is available at https://github.com/Martinyxr/SGCR. + +# 1. Introduction + +In computer graphics and vision, photorealistic novel view synthesis (NVS) and accurate geometry reconstruction are two significant objectives. 3D Gaussian Splatting [19] offers an appealing explicit representation—3D Gaussians, resulting in high-quality synthesis of novel views and real-time rendering speed. However, it falls short in capturing accurate 3D geometric structures since the anisotropic 3D Gaussians, which are intended for efficient rasterization and sufficient expression ability, do not align with the geometric features (surfaces, edges and corners) in a faithful manner. A question naturally arises: Can we utilize 3D Gaussians + +![](images/127d104e1879a4c7f9cdbeea0d6311b729694c564043572e8308ed1bbc7b9fea.jpg) +(a) 3D Gaussian Splatting + +![](images/c1143525946b7625596f3e50a6476fc618efe0c54f9a3b38986126672a16b54a.jpg) +(b) Spherical Gaussians +Figure 2. Visualization for optimized Gaussians. (a) Gaussians from original 3D Gaussian Splatting. (b) Our Spherical Gaussians. + +for the direct representation of intricate geometric structures while retaining their benefits in 2D rendering? + +Several recent works attempt to directly address the task of sdf generation or mesh reconstruction [6, 12, 17, 52] with 3D Gaussians. In this paper, we try to answer this question from a new perspective—reconstructing 3D parametric curves with 'Spherical Gaussians'. Following the settings of classic rendering works [19, 29], we use a set of calibrated multi-view images as input, but our goal substitutes novel view synthesis for 3D edge/curve reconstruction. + +Feature curves serve as critical geometric cues in characterizing the structure of 3D shapes, benefiting surface reconstruction, shape perception or other downstream tasks. Traditional methods [9, 13, 18, 21] usually work with the direct representation of 3D shapes, such as polygonal meshes or point clouds. However, the presence of sharp edges may be compromised or entirely overlooked on account of imperfect 3D scanning and acquisition, especially when dealing with sparse and noisy data. In recent years, learning based-methods [26, 28, 41, 54] are proposed to mitigate this, but obtaining precise ground-truth labels for 3D edge supervision involves laborious manual annotation, and inevitably limiting their generality. On the contrary, 2D images is easily accessible and allow for straightforward identification of feature edges through mature edge detection methods, making 2D supervision more attractive for edge processing tasks. Despite advantages in detection and self-supervision, there still remains a huge obstacle to overcome—establishing correspondence between edges across different views when working with multi-view images, in other words, how to combine information from multiple views to construct a complete 3D structure? This is exactly the factor for our design of Spherical Gaussians (as well as SGCR), which serve as a robust bridge to connect 3D depiction with 2D guidance. + +Our method consists of two main stages: Spherical Gaussians generation and parametric curves extraction. During the first stage, we utilize 2D edge maps to refine Gaussians using a view-based rendering loss compared to the ground-truth edge map derived from the image of that view. Optimizing Gaussians directly as vanilla 3D Gaus + +![](images/001a975b016e292d31f6e58d6ea667e93a57fab3111ecdc60028ce89abf4d888.jpg) +Figure 3. Illustration for fitting Bézier curve from Spherical Gaussians. $P_{1}$ , $P_{2}$ , $P_{3}$ and $P_{4}$ are weighted control points of the curve. + +sian Splatting [19] is problematic due to the sparse distribution of edge pixels and the diversity of Gaussian attributes. The outcome Gaussians may recover 2D images to some extent but exhibit redundant irregular ellipsoids (see Fig. 2 (a)) and lack solid 3D geometric meanings, making following reconstruction tasks quite difficult. Therefore, we implement several constraints to transform standard 3D Gaussians into Spherical Gaussians. We introduce grid initialization, fixing-covariance training and a two-phase pruning strategy to guide the optimization process. Our final distribution of gaussians lie faithfully on sharp edges of objects in 3D space, appearing the same shape like spheres (see Fig. 2 (b)) and holding attributes in opacity and gray-scale color. + +In the second stage, we utilize these discrete Spherical Gaussians to extract smooth 3D parametric curves for final reconstruction of edges. We present a novel optimization-based algorithm called SGCR, which involves fitting edges represented by Spherical Gaussians onto rational Bezier curves. The process begins with line fitting and then gradually refines the curves by global optimization. The final results are stored by all control points and weights in these curves. An illustration for curve extraction from Spherical Gaussians is shown in Fig. 3. + +We evaluate our method using a variety of object categories with complex models and scenes from ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. Extensive experiments show that SGCR, which utilizes Spherical Gaussians as bridge to reconstruct 3D parametric curves by 2D supervisions, outperforms existing state-of-the-art methods on all metrics and achieves great efficiency. In summary, our contributions include: + +- We propose a new 3D edge extraction method from multiview 2D images using Spherical Gaussians (SG). +- We introduce a new optimization-based algorithm (SGCR) tailored for directly reconstructing 3D parametric curves from our generated Spherical Gaussians. +- Our approach achieves the state-of-the-art performance as well as high efficiency on curve reconstruction tasks. + +# 2. Related Work + +# 2.1. Image-based 3D Representation and Rendering + +Image-based rendering (IBR) techniques use a collection of 2D images of a scene to generate the visual represent- + +tation of the scene and render novel views. Significant advancements have been achieved by Neural Radiance Fields (NeRF) [29], which depicts a scene's geometry and visual attributes as a radiance field, enabling the retrieval of color and volume density at any point in space and from any viewing angle for the purpose of rendering. Recently, the unstructured GPU-optimized 3D Gaussian Splitting (3DGS) [19] presents the explicit Gaussian primitive that achieves enhanced rendering speeds and improved quality without the reliance on neural components. It has also been quickly extended to multiple domains [12, 27, 45, 53, 55]. We convert standard 3D Gaussians into Spherical Gaussians, enabling significant advantages in both view rendering and the representation of geometric structures. + +# 2.2.3D Curve Reconstruction + +The convention of 3D curve reconstruction is based on point cloud. Traditional methods [3, 9, 21, 22, 40] focus on local geometric properties of point clouds, such as normals [7, 42], curvatures [49] and hierarchical clustering [10]. Data-driven methods [15, 26, 28, 39] often adopt edge detection as a binary classification for point clouds. As network architectures advance, the classifier used for edge detection has become more complex models [5, 33, 41, 51]. RFEPS [47] presents a multistage optimization to augment the point set so that sufficient points are generated on potential edges of geometry. NerVE [54] presents a neural volumetric edge representation learned through a volumetric learning framework. + +Extraction curves from multi-view images was pioneered by [4] which introduces a full SfM system using line segments. With the developments of line detections [30, 31, 48] and matching [2, 30, 32], several works [14, 16, 24, 25, 43] have attempted to revisit the line mapping problem, but they are restricted to straight lines and frequently generate fragmented small lines when dealing with curves. Lifting 2D sketches/curves to 3D curves [11, 34, 35, 46] has also been exploited, but the ability to generate parametric curves is still limited. Recently, NEF [50] learns a neural implicit field to capture the distribution of edge densities, and EMAP [23] encodes 3D edge distance and direction in Unsigned Distance Functions (UDF) from multi-view edge maps. However, they all require large training time for even a simple scene. In contrast, our method employs Spherical Gaussians as the direct representation for edges, achieving 3D parametric curve reconstruction both accurately and efficiently. + +# 3. Method + +# 3.1. Preliminary + +3D Gaussian Splatting. 3DGS [19] is a recently popular method for representing 3D scenes with 3D Gaussian + +primitives, achieving both explicit representation and high-quality real-time rendering. Each Gaussian primitive is defined by a covariance matrix $\Sigma$ , and a center point $p$ which is the mean of the Gaussian. The 3D distribution can be represented as: $G(x) = e^{-\frac{1}{2} (x - p)^{\mathrm{T}}\Sigma^{-1}(x - p)}$ . + +A complete 3D Gaussian is given by its mean/center $p \in \mathbb{R}^3$ , color represented with spherical harmonics coefficients $\mathbf{H} \in \mathbb{R}^k$ , opacity $\alpha \in \mathbb{R}$ , quaternion $\mathbf{r} \in \mathbb{R}^4$ , and scaling factor $\mathbf{s} \in \mathbb{R}^3$ . For a given pixel, the combined color and opacity from multiple Gaussians are weighted by Eq. 3. The color blending for overlapping points is: $\mathbf{C} = \sum_{i \in N} \mathbf{c}_i \alpha_i \prod_{j=1}^{i-1} (1 - \alpha_j)$ where $\mathbf{c}_i, \alpha_i$ denote the color and density of a point. + +Spherical Gaussians. Our Spherical Gaussians enjoy the same nature of 3D Gaussians, but make the following changes for geometry priority and simplicity: (1) Remove the covariance matrix $\pmb{\Sigma}$ of Gaussians (attributes of scaling $\mathbf{S}$ and rotation $\mathbf{R}$ ) and replace it with a fixed radius $r_0$ , i.e. convert Gaussian primitives from various ellipsoids into spheres of uniform size. (2) Remove spherical harmonics coefficients $\mathbf{H}$ and modify color to one-dimensional gray value (just for edge map rendering). According to our experiments, these changes do not hamper the backward propagation of gradients from 2D edge-map supervision, but impose great regularization on geometry distribution of 3D isotropic Gaussians, facilitating reconstruction tasks. + +Anisotropic Gaussians v.s. Isotropic Gaussians. Traditional 3DGS [19] and its subsequent works mainly center on the technique of 'splatting'—utilizing different anisotropic Gaussian splats to accurately represent the entire scene. However, our work focuses on each 'Gaussian' itself, essentially displaying a reverse process: decomposing the splating results into small 'atoms' and assigning clear geometric meanings to every individual Gaussian. A thin and elongated Gaussian seems to be more efficient in representing edges, but it can be split into smaller 'basis-Gaussians' with few sacrifices but for better interpretation and convenience, especially when following pointcloud-related works. That's why we favour Spherical Gaussians and more discussions can be found in supplementary materials. + +# 3.2. Spherical Gaussians Generation + +# 3.2.1. Initialization + +Our full pipeline is illustrated in Fig. 4. Given a set of calibrated multi-view images, we first adopt edge detection method [38] to generate 2D edge maps for each view. 3DGS uses structure-from-Motion (SfM) [36] points as initialization for Gaussians, but this approach is not ideal for handling edge maps. Thus, we design grid initialization for Spherical Gaussians. We create a $50 \times 50 \times 50$ grid with points evenly distributed in space of $[0,1]^3$ (see Fig. 4—Grid initialization) as the initial centers for our Spherical Gaussians. This setting is better than random initialization + +![](images/567532b9367e791888bd11304d8fcac8d67e532a51c3e2b2a3d0c39c65b8838e.jpg) +Figure 4. Pipeline illustration across 2D and 3D modality. Stage1: Spherical Gaussians generation. Stage2: Parametric curves extraction. + +in producing neat structures. And we choose a constant radius $r_0$ to replace the covariance matrix with Spherical Gaussians of the same size. In our experiments, we use $r_0 = 0.005$ . + +# 3.2.2. Training Losses + +Edge Loss. The original loss functions in 3DGS [19] to train 3D Gaussians for each view is defined as: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {D - S S I M} \tag {1} +$$ + +where $\mathcal{L}_1$ and $\mathcal{L}_{D-SSIM}$ are supervised by the ground-truth view. However, in edge maps, edge pixels are often sparsely distributed. Relying solely on this loss function will easily make the training process stuck in local optima, causing the opacities and colors for all Gaussians to converge towards zero and resulting in all rendered images becoming black. Hence, we opt for a new edge-aware loss $\mathcal{L}_{edge}$ instead of $\mathcal{L}_1$ in order to balance the impact of both edge and non-edge pixels within each edge map. The edge-aware loss is defined as: + +$$ +\mathcal {L} _ {\text {e d g e}} = \frac {N _ {I} - | E _ {I} |}{N _ {I}} \sum_ {i \in E _ {I}} \| I _ {i} - \hat {I} _ {i} \| _ {2} ^ {2} + \frac {| E _ {I} |}{N _ {I}} \sum_ {i \notin E _ {I}} \| I _ {i} - \hat {I} _ {i} \| _ {2} ^ {2} \tag {2} +$$ + +where $N_{I}$ is the image resolution and $E_{I}$ refers to all edge pixels in current edge map (gray scale larger than threshold $\eta = 0.3$ ), $I_{i}$ and $\hat{I}_i$ refers to the pixel in Gaussian-rendering image and ground-truth edge-map image respectively. This edge-aware loss enhances the ability of Spherical Gaussians to identify edges from edge pixels. + +Opacity-Color Loss. Another issue during edge training is about occluded edges. Some 2D edge maps may lose true edge pixels due to occlusions from these image views (see the difference between Fig. 4-Edge maps and Rendering + +views), while from other views the occlusion may not happen. This inconsistency among views will confuse Spherical Gaussians during training, potentially leading to unstable optimization. During training, color attribute is directly supervised by $\mathcal{L}_{edge}$ loss, while opacity is indirectly optimized and reset at certain iterations. This difference sets them apart, and equating them impedes smooth optimization. Therefore, we introduce opacity-color loss $\mathcal{L}_{oc}$ for all Spherical Gaussians to make the value of Gaussian opacity more consistent. This opacity-color loss is defined as: + +$$ +\mathcal {L} _ {o c} = \sum_ {i = 1} ^ {N _ {S G}} \| o _ {i} - c _ {i} \| _ {2} ^ {2} \tag {3} +$$ + +where $N_{SG}$ is the number of all Spherical Gaussians during current training step, $o_i$ and $c_i$ refer to opacity and color attribute of each Spherical Gaussian respectively. $\mathcal{L}_{oc}$ helps to keep consistency between edge density and color intensity during multi-view supervision. This mechanism ensures that occluded edges—represented by Gaussians with low opacity values—are less susceptible to premature pruning during optimization, resulting in more complete 3D reconstructions. It also helps rectify edge predictions missed by 2D detectors. In our experiment, after implementing $\mathcal{L}_{oc}$ , Gaussian-rendered images can depict complete edge structures with no occlusion (Fig. 4—Rendering Views). + +Regularization Loss. To restrict total number of Spherical Gaussians and accelerate convergence, we add an extra regularization term $\mathcal{L}_{\text {regular }}$ on Gaussian's opacity. The regularization loss is defined as: + +$$ +\mathcal {L} _ {r e g} = \sum_ {i = 1} ^ {N _ {S G}} \log \left(1 + \frac {o _ {i} ^ {2}}{0 . 5}\right) \tag {4} +$$ + +![](images/e6d6f64c05812fb532543fe202cf803d0d513edbae22feaa1a90f80d353cadd7.jpg) +Figure 5. The change on number of Gaussians during training. + +where $N_{SG}$ is the number of current Spherical Gaussians and $o_i$ is each Gaussian's opacity. During the training process, this loss function works implicitly because Spherical Gaussians with opacity below a certain threshold will be pruned (as in 3DGS [19]), allowing for a comfortable number of total Gaussian primitives for next stage. + +To summarize, our final loss function for training Spherical Gaussians is presented as: + +$$ +\mathcal {L} = \left(1 - \lambda_ {1}\right) \mathcal {L} _ {\text {e d g e}} + \lambda_ {1} \mathcal {L} _ {D - S S I M} + \lambda_ {2} \mathcal {L} _ {o c} + \lambda_ {3} \mathcal {L} _ {\text {r e g}} \tag {5} +$$ + +where the balancing parameters $\lambda_1, \lambda_2, \lambda_3$ are set to 0.2, 2 and 0.01 in our experiments respectively. + +# 3.2.3. Training Strategy + +The effective training strategy is essential for creating properly structured Spherical Gaussians. We adopt a two-phase training strategy, each phase consists of 3,000 iterations. During the first phase, we conduct densification (split and clone) of Spherical Gaussians every 200 iterations. And we reset the opacity attribute to 0.1 every 1,000 iterations. At the end of first phase, we design a final-prune—removing all Spherical Gaussians whose opacity $o_i < 0.5$ and color $c_i < 0.1$ . This operation largely reduces the total number of Gaussians, leading to substantial reductions in both training time and noise. During the second phase, we no longer densify Spherical Gaussians but focus on refining their positions and attributes. In the end, we perform another final-prune to retain all significant Spherical Gaussians (at quantity of 1k to 5k) with their attributes. The variation of numbers for Spherical Gaussians during all iterations is depicted in Fig. 5. The whole training process of Spherical Gaussians for single scene takes only about 1 minute. + +# 3.3. Parametric Curve Extraction from Spherical Gaussians + +After the first stage, we get the generated Spherical Gaussians (denoted as $G$ ) aligned with object's edges, resembling a 3D wireframe constructed by many small spheres. In the second stage (Fig. 4—Stage 2), we present of Spherical-Gaussians Curve Reconstruction (SGCR) to accomplish the final 3D edge reconstruction task. SGCR is an + +# Algorithm 1 Line Fitting + +Input: a set of Spherical Gaussians $G$ with radius $r_0$ . + +Output: the set of line endpoints $L = \{(p_i, q_i)\}_{i=1}^{|L|}$ for the coarse line-fitting of 3D edges. + +Denote: $P(G)$ is the centers of $G$ , and $N(0,1)$ is the Standard Normal Distribution. + +1: $L = \emptyset$ +2: while $|G| > N_0$ do +3: $S_{min} =$ MaxIntValue +4: for $i = 1$ to $n$ do +5: $p, q = \text{RandomChoice}(P(G))$ +6: $L_{p,q} = \text{interpolate}(p, q, N_s) + r_0 * N(0,1)$ +7: $G_{i} = \{g\in G|\min_{x\in L_{p,q}}||P(g) - x||_{2}^{2} < \delta_{1}\}$ +8: $p', q' = \operatorname*{arg min}_{p,q} \mathcal{L}_{CD}(L_{p,q}, P(G_i))$ +9: $S_{i} = \min_{p,q}\mathcal{L}_{CD}(L_{p,q},P(G_{i}))$ +10: if $S_{i} \leq S_{min}$ then +11: $S_{min} = S_i, G' = G_i, (\hat{p}, \hat{q}) = (p', q')$ +12: end if +13: end for +14: $G = G - G^{\prime},L = L\cup \{(\hat{p},\hat{q})\} ;$ +15: end while +16: return $L$ ; + +optimization-based algorithm guided by $G$ which includes two parts: line fitting and global optimization. + +# 3.3.1. Line Fitting + +We first fit a set of line segments to $G$ 's centers as the coarse fitting. The pseudo code of this part is shown in Algorithm 1. The whole process consists of several line-fitting turns and $G$ will be gradually fitted and partly removed by one line each turn. Line fitting will stop when the number of remaining $G$ is less than $N_0 = 5$ . + +In each turn, we conduct $n$ random searches and record the best result. At the beginning of $i$ -th search, we randomly choose two nearest Gaussian centers as initial endpoints $p, q$ of the fitting line, and interpolate $N_{s}$ points along the line. To simulate the shape of Spherical Gaussians, we dilate the sampled points up by duplicating and adding Gaussian noise with standard deviation equals $G$ 's radius $r_0$ . Then we apply Chamfer Distance to compute the fitting loss between dilated line points and $G$ . This Chamfer loss is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {C D} \left(L _ {p, q}, P \left(G _ {i}\right)\right) = \frac {\gamma_ {1}}{\left| L _ {p , q} \right|} \sum_ {x \in L _ {p, q}} \min _ {y \in P \left(G _ {i}\right)} \| x - y \| _ {2} ^ {2} \\ + \frac {1}{| P (G _ {i}) |} \sum_ {y \in P (G _ {i})} \min _ {x \in L _ {p, q}} \| x - y \| _ {2} ^ {2} \tag {6} \\ \end{array} +$$ + +$$ +G _ {i} = \left\{g \in G \mid \min _ {x \in L _ {p, q}} \left| | P (g) - x \| _ {2} ^ {2} < \delta_ {1} \right. \right\} \tag {7} +$$ + +where $L_{p,q}$ denotes the dilated line points between $p, q$ in $i$ -th search, $P(G)$ represents Gaussian centers of all current $G$ and $\gamma_1 = 2$ is a balance parameter. $G_i$ is a subset of $G$ whose centers are close to $L_{p,q}$ within a threshold $\delta_1 = 0.02$ . We optimize $p, q$ by minimizing Eq. (6) with certain iterations and use the value of $\mathcal{L}_{CD}$ after optimization as the final fitting score $S_i$ for $i$ -th search. + +For each turn, suppose the best fitting is in $k$ -th search, i.e. $S_{k} = \min S_{i}$ , we record the optimized endpoints $p^{\prime}, q^{\prime}$ in $k$ -th search and delete $G_{k}$ to update the current $G$ . The next turn will then proceed accordingly. Once all the turns are completed, we will get a set of paired line endpoints $L = \{(p_{i}, q_{i})\}_{i=1}^{|L|}$ which together form an approximate line-fitting result (Fig. 4—Line Fitting). + +# 3.3.2. Global Optimization + +After line fitting, we resume all $G$ deleted in previous part and take opacity attribute into account as the fine fitting. We interpolate another two control points between each line endpoint pair $(p_i, q_i)$ and introduce point weight to initialize 3rd order rational Bézier curves, which is defined as: + +$$ +B (u) _ {p, w} = \frac {\sum_ {i = 0} ^ {3} B _ {3 , i} (u) p _ {i} w _ {i}}{\sum_ {i = 0} ^ {3} B _ {3 , i} (u) w _ {i}} \tag {8} +$$ + +where $B_{3,i}$ is basis functions of Bézier curves, $p_i$ is the $i$ -th control points of curve and $w_i$ is the weight of control point $p_i$ ( $w_i \equiv 1$ for simple Bézier curves). We opt for the rational Bézier curves over simple Bézier curves because simple Bézier curves cannot perfectly fit circles, which are frequently seen in object's edges. + +We also sample $N_{s}$ points for each Bézier curve and dilate them up as in line fitting. Since $G$ with higher opacity values are more significant during fine fitting, we apply the weighted Chamfer Distance to compute the fine fitting loss between global curve points and all $G$ : + +$$ +\begin{array}{l} \mathcal {L} _ {W C D} (L _ {a l l}, P (G)) = \frac {\gamma_ {2}}{| L _ {a l l} |} \sum_ {x \in L _ {a l l}} o _ {y _ {m i n}} \cdot \min _ {y \in P (G)} \| x - y \| _ {2} ^ {2} \\ + \frac {1}{| P (G) |} \sum_ {y \in P (G)} o _ {y} \cdot \min _ {x \in L _ {\text {a l l}}} \| x - y \| _ {2} ^ {2} \tag {9} \\ \end{array} +$$ + +where $L_{all}$ refers to global dilated points from all curves and $P(G)$ refers to Gaussian centers of $G$ , $o_y$ is the opacity attribute in $y$ -th $G$ and $\gamma_2 = 2$ is another balance parameter. We also adopt endpoints loss to encourage connections between curves, which is defined as: + +$$ +\mathcal {L} _ {\text {e n d p o i n t s}} = \sum_ {x, y \in P _ {E}} \min \left(\| x - y \| _ {2} ^ {2} - \delta_ {2}, 0\right) \frac {\| x - y \| _ {2} ^ {2}}{\| x - y \| _ {2} ^ {2} - \delta_ {2}} \tag {10} +$$ + +where $P_{E}$ is the set of endpoints (the first and last control points) and $\delta_{2}$ (we choose 0.01) is a threshold to only regularize endpoints which are close to each other. The final + +objective function to globally optimize all curves is defined as: + +$$ +\begin{array}{l} \arg \min \quad \left(\mathcal {L} _ {W C D} + \lambda \mathcal {L} _ {\text {e n d p o i n t s}}\right) \tag {11} \\ \{\{z _ {i} ^ {j}, w _ {i} ^ {j} \} _ {j = 1} ^ {4} \} _ {i = 1} ^ {| L |} \\ \end{array} +$$ + +where $z_{i}^{j}$ and $w_{i}^{j}$ are coordinates and weights of four control points in each Bézier curve, $\lambda$ is a balancing parameter which is set to 0.005 in our experiments. The pseudo code of this part can be found in the supplementary material. + +After global optimization, we finally extract 3D parametric curves from Spherical Gaussians in form of 3rd order rational Bezier curves (Fig. 4—Rational Bezier Curves). + +# 4. Experiments + +# 4.1. Dataset + +We conduct experiments on the ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. ABC [20] consists of massive CAD models with edge annotations, and we use a portion of it—ABC-NEF, provided by [50]—which consisting of 115 distinct and challenging CAD models. ModelNet [44] includes 3D models of everyday objects such as chairs and tables, but it does not have ground-truth edge annotations. So we manually select and annotate edges on ModelNet ground-truth meshes. Our edge dataset consists of 120 furniture models with annotated edges and is used for a more comprehensive evaluation. We utilize BlenderProc [8] to generate rendered images with these objects. We generate 50 views of $800 \times 800$ images for each object and adopt PiDiNet [38] to predict 2D edge maps. DTU [1] and Replica [37] are real-world datasets with multi-view images as input and we use it for evaluation on real-world reconstruction. + +# 4.2. Comparisons + +Settings. We compare our method against the state-of-the-art methods in 3D curve reconstruction, including RFEPS [47], NerVE [54], NEF [50] and EMAP [23]. The first two requires point clouds as input, while the last two and ours use only 2D images in the same input frames and resolution. We all adopt their pre-trained models and corresponding reconstruction approaches to extract curves for evaluation. More experiments, discussions and details can be found in the supplementary material. + +Quantitative Comparison. We first normalize and align all reconstructed curves and ground truth edges into the range of $[0, 1]$ . Then, we sample points on all reconstructed curves and measure the distances between these points and ground-truth edge points. When assessing IoU, precision, recall, and F-score, points are considered matched if there is at least one ground-truth edge point with an $L_{2}$ distance smaller than 0.02. We also adopt the Chamfer Distance (CD) between point clouds to measure the geometric accuracy of reconstructed parametric curves. As reported in + +![](images/7f3a7fb84705df8b4e48c578c3f2c6b64c2de11bcc9fcbf23801346cd2c82565.jpg) +Figure 6. Qualitative comparisons. The first two objects come from ABC-NEF dataset and the last one comes from ModelNet dataset. Ours(SG) refers to our generated Spherical Gaussian in the first stage, GT refers to the ground-truth edge. + +Table 1. Quantitative comparisons on ABC-NEF Dataset [50] and ModelNet Dataset[44]. + +
MethodABC-NEF DatasetModelNet DatasetInputTrainingReconstruction
CD↓Precision↑Recall↑F-score↑IoU↑CD↓Precision↑Recall↑F-score↑IoU↑ModalityTime↓Time↓
RFEPS [47]0.03230.89620.85600.86730.81900.07310.70320.59650.64120.5621Point Cloud-48s
NerVE [54]0.03890.95180.73070.82670.68450.02940.95680.80170.87240.7931Point Cloud>10h1s
NEF [50]0.03530.93870.88380.90440.82830.03210.90720.81580.85590.7475Images>1h119s
EMAP [23]0.02910.95130.89260.92100.84730.02900.91150.84880.87840.7751Images>2h40s
Ours0.02800.95460.90520.92600.86150.02640.96270.84150.89520.8164Images87s32s
+ +Table 1, our method outperforms other methods in all metrics on ABC-NEF [50] and most metrics on ModelNet [44]. + +Qualitative Comparison. We illustrate the visual comparisons in Fig. 6. It can be observed that RFEPS [47] and NerVE [54] may miss edges with weak curvature, leading to incomplete reconstruction. NEF [50] and EMAP [23] struggle with handling thin structures and make errors in curve connection. On the contrary, our method benefits from aligned Spherical Gaussians (see Ours(SG) in Fig. 6) and maintains a high level of robustness on 3D curve reconstruction. + +Time Comparison. We also make comparisons on training time and average reconstruction time (see the last two columns in Tab. 1). NerVE [54] is a learning-based methods that rely on large point cloud datasets, requiring extensive training time. RFEPS [47] is learning-free but takes time in point cloud optimization. NEF [50] and EMAP [23] are supervised on images, but they all involve a time-consuming training process (several hours for single scene) for predicting edge densities or unsigned distance functions. Our SGCR generates and optimizes only a small number of Spherical Gaussians (with two stages completing within minutes, $50 \times$ faster than NEF and EMAP), making 3D + +curve reconstruction quite efficient. + +# 4.3. Real-world Scene + +We further test our performance in real-world scenes on DTU [1] and Replica [37] dataset. They present challenging objects and complex scenes for edge reconstruction respectively. The visual comparisons against NEF [50] and EMAP [23] on DTU [1] dataset are presented in Fig. 7. NEF misses + +![](images/8806b33bc763145fe29e1f923c1fedbe9e139e270d54a08c2d8c28599efc70e7.jpg) +Figure 7. Comparisons on real-world scenes from DTU [1]. + +![](images/906342259482a6d534a982ef1a34972564f054bd3e4a6ad66bf4e81166c34346.jpg) +Figure 8. Comparisons on real-world scenes from Replica [37]. + +![](images/a2e97f069488d28b0bfe006109edaa907b4a7c39c51aa3ad78e6bdb640854565.jpg) + +![](images/7bcfd47ae5971cffc3fb5547026ec96d05837a42ea07318e4df5ca4222c32e81.jpg) + +details when fitting complex parts, and EMAP produces too many chaotic and redundant line segments, while our method reconstructs edges accurately and preserves details clearly. Another visual comparison on Replica [37] dataset are presented in Fig. 8. It can be seen that our method still works well on room-level edge reconstruction (NEF fails, and EMAP's training time for single real-world scene is $100 \times$ slower than ours), demonstrating our method's great generalization ability and broad applicability. + +# 4.4. Ablation Studies + +Pipeline. We conduct ablation studies on ABC-NEF dataset [50] to evaluate the effectiveness of core designs in our method, including Spherical Gaussians, loss designs and two-phase training strategy in the first stage, as well as line fitting and global optimization in the second stage. Tab. 2 lists the quantitative results. w/o Spherical Gaussians stands for using original 3D Gaussians [19] trained by edge maps to directly represent edges and this result is quite poor, highlighting the significance of Spherical Gaussians. Statistics in Tab. 2 reveal that removing any component from our pipeline negatively impacts the final result. + +Table 2. Quantitative ablation studies on our pipeline. + +
PipelineCD↓F-score↑IoU↑
w/o Spherical Gaussians0.1420.3300.194
w/o grid initialization0.0320.9010.822
w/o edge loss Ledge0.0520.8530.714
w/o opacity-color loss Loc0.0510.8590.723
w/o regularization loss Regular0.0290.9130.846
w/o two-phase training strategy0.0300.9200.860
w/o line fitting0.0300.9110.822
w/o global optimization0.0320.8760.757
Full Model0.0280.9260.862
+ +Radius. We analyze the effect of radius in Spherical Gaussians. We test $r_0 = 0.002$ , 0.005, 0.01 and variable. Experimental results are presented in Tab. 3 (setting $r_0$ as a variable does not perform well). As the radius increases, the total number of Spherical Gaussians decreases, with more + +Table 3. Comparison on different radius choices for Spherical Gaussians. $N_{SG}$ : Number of Spherical Gaussians. T1: Time for Gaussian generation. T2: Time for curve reconstruction. + +
Radius r0CD↓F-score↑IoU ↑NSGT1T2
0.0020.0320.9060.8495,12484s56s
0.0050.0280.9260.8622,36190s36s
0.010.0350.8810.810592101s30s
Variable0.0400.8630.7853,076110s44s
+ +![](images/2cebe1863b9a84ef3dfd1c7eecddfae1ebb5c5ce504a3e369d7a4cd81659ceba.jpg) +Figure 9. Visualization for Gaussians with different radii. + +![](images/93135909d991fdb2acf1ab31d25db43f5bc162bd2f7a902fb52ee4ac17d8fa84.jpg) +Figure 10. From left to right, we present 2D images, rendered Spherical Gaussians and final curves in 5, 10, 30 and 50 views. + +time spent in the first stage and less time in the second stage. The visual comparisons can be found in Fig. 9. An inappropriate radius size may lead to noise or underfitting. Using $r_0 = 0.005$ leads to the optimal reconstruction metrics as well as a reasonable running time. + +Views. We also study the effect of the number of input views. We train Spherical Gaussians with 5, 10, 30 and 50 views respectively (all evenly distributed). The visual results are illustrated in Fig. 10. For simple cases, results with 10 views are close to satisfactory, while for complex cases, more views are required for better performance. + +# 5. Conclusions + +We have presented SGCR—a novel method using Spherical Gaussians for efficient 3D parametric curve reconstruction from multi-view images. The main concept involves generating Spherical Gaussians by 2D supervision and extract 3D parametric curves from them. Our method outperforms previous methods quantitatively and qualitatively, while enjoying great efficiency. SGCR also demonstrates the potential in integrating multi-modal information efficiently, bridging the gap between 2D rendering and 3D reconstruction. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China under Grant 62032011. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2, 6, 7 +[2] Hichem Abdellali, Robert Frohlich, Viktor Vilagos, and Zoltan Kato. L2d2: Learnable line detector and descriptor. In 2021 International Conference on 3D Vision (3DV), pages 442-452. IEEE, 2021. 3 +[3] Nina Amenta, Marshall Bern, and David Eppstein. The crust and the $\beta$ -skeleton: Combinatorial curve reconstruction. Graphical models and image processing, 60(2):125-135, 1998. 3 +[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 3 +[5] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 3 +[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2 +[7] Kris Demarsin, Denis Vanderstraeten, Tim Volodine, and Dirk Roose. Detection of closed sharp edges in point clouds using normal estimation and graph theory. Computer-Aided Design, 39(4):276-283, 2007. 3 +[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6 +[9] Tamal K Dey, Kurt Mehlhorn, and Edgar A Ramos. Curve reconstruction: Connecting dots with good reason. In Proceedings of the fifteenth annual symposium on Computational geometry, pages 197-206, 1999. 2, 3 +[10] Chen Feng, Yuichi Taguchi, and Vineet R Kamat. Fast plane extraction in organized point clouds using agglomerative hierarchical clustering. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 6218-6225. IEEE, 2014. 3 +[11] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 3 +[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3 +[13] Stefan Gumhold, Xinlong Wang, Rob S MacLeod, et al. Feature extraction from point clouds. In IMR, pages 293-305, 2001. 2 + +[14] Yijia He, Ji Zhao, Yue Guo, Wenhao He, and Kui Yuan. Plvio: Tightly-coupled monocular visual-inertial odometry using point and line features. Sensors, 18(4):1159, 2018. 3 +[15] Chems-Eddine Himeur, Thibault Lejemble, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 3 +[16] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3 +[17] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. arXiv preprint arXiv:2403.17888, 2024. 2 +[18] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 8 +[20] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6 +[21] G Saravana Kumar, Prem Kumar Kalra, and Sanjay G Dhande. Curve and surface reconstruction from points: an approach based on self-organizing maps. Applied Soft Computing, 5(1):55-66, 2004. 2, 3 +[22] In-Kwon Lee. Curve reconstruction from unorganized points. Computer aided geometric design, 17(2):161-177, 2000. 3 +[23] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6, 7 +[24] Hyunjun Lim, Jinwoo Jeon, and Hyun Myung. Uv-slam: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters, 7(2):1518-1525, 2022. 3 +[25] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3 +[26] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and JanDirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. Cornell University - arXiv, Cornell University - arXiv, 2021. 2, 3 +[27] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3 + +[28] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2, 3 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3 +[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 3 +[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 3 +[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 3 +[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[34] Alec Rivers, Frédo Durand, and Takeo Igarashi. 3d modeling with silhouettes. In Acm siggraph 2010 papers, pages 1-8. 2010. 3 +[35] Ryan Schmidt, Azam Khan, Karan Singh, and Gord Kurtenbach. Analytic drawing of 3d scaffolds. In ACM SIGGRAPH Asia 2009 papers, pages 1-10. 2009. 3 +[36] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism. ACM Transactions on Graphics, page 835-846, 2006. 3 +[37] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8 +[38] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3, 6 +[39] Shuangbu Wang, Yu Xia, Lihua You, and Jianjun Zhang. Reconstruction of curve networks from unorganized spatial points. Journal of Universal Computer Science (J. UCS), 26 (9):1265-1280, 2020. 3 +[40] Wenping Wang, Helmut Pottmann, and Yang Liu. Fitting b-spline curves to point clouds by curvature-based squared distance minimization. ACM Transactions on Graphics (ToG), 25(2):214-238, 2006. 3 +[41] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: + +Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3 +[42] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 3 +[43] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 3 +[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2, 6, 7 +[45] Tianyi Xie, Zeshun Zong, Yuxin Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 3 +[46] Baoxuan Xu, William Chang, Alla Sheffer, Adrien Bousseau, James McCrae, and Karan Singh. True2form: 3d curve networks from 2d sketches via selective regularization. ACM Transactions on Graphics, 33(4), 2014. 3 +[47] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3, 6, 7 +[48] Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing: From supervised to self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[49] Bisheng Yang and Yufu Zang. Automated registration of dense terrestrial laser-scanning point clouds using curves. ISPRS journal of photogrammetry and remote sensing, 95: 109-121, 2014. 3 +[50] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8486-8495, 2023. 3, 6, 7, 8 +[51] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 3 +[52] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2 +[53] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 3 + +[54] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 7 +[55] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3 \ No newline at end of file diff --git a/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/images.zip b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed3a434d60fe623a6e32de51313e45b7be1e9951 --- /dev/null +++ b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:284bdf94fbfa3aaca90541d9f2fa8acbb54c72fa4642c5cbf4ee5388e92dcae6 +size 656392 diff --git a/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/layout.json b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..27362305010d2969f899887e520e805bc97d0174 --- /dev/null +++ b/2025/SGCR_ Spherical Gaussians for Efficient 3D Curve Reconstruction/layout.json @@ -0,0 +1,10447 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 511, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 511, + 120 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 511, + 120 + ], + "type": "text", + "content": "SGCR: Spherical Gaussians for Efficient 3D Curve Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": "Xinran Yang" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Donghao Ji" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Yuanqi Li" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Jie Guo" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Yanwen Guo" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": ", Junyuan Xie" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 142, + 499, + 172 + ], + "type": "text", + "content": "Nanjing University, Nanjing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 174, + 400, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 174, + 400, + 185 + ], + "spans": [ + { + "bbox": [ + 211, + 174, + 400, + 185 + ], + "type": "text", + "content": "{xryang,donghaoji}@smail.nju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 196, + 187, + 411, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 187, + 411, + 200 + ], + "spans": [ + { + "bbox": [ + 196, + 187, + 411, + 200 + ], + "type": "text", + "content": "{yuanqili,guojie,ywguo,jyxie}@nju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 222, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 57, + 222, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 57, + 222, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 57, + 222, + 555, + 401 + ], + "type": "image", + "image_path": "ec9ce4293209621d31fd801112fd0f36b864b75d1209a8aae9b181ffcc5198cf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 411, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 54, + 411, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. We present a novel method using 2D images to generate Spherical Gaussians, enabling the accurate representation of 3D edges. Spherical Gaussians combine the advantages from 3D Gaussian Splatting and the ability to preserve excellent 3D structures. With a corresponding curve extraction algorithm, Spherical Gaussians efficiently reconstruct 3D edge curves by only 2D supervision." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 454, + 200, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 454, + 200, + 467 + ], + "spans": [ + { + "bbox": [ + 151, + 454, + 200, + 467 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 479, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 297, + 696 + ], + "type": "text", + "content": "Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting technique has achieved high quality novel view synthesis as well as fast rendering speed. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is due to the fact that Gaussian's attributes are primarily tailored and fine-tuned for rendering diverse 2D images by their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries, from which we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians is optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and then compared to the ground-truth edge map extracted from the corresponding image, without the need for any 3D guidance or su" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 456, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 552 + ], + "type": "text", + "content": "pervision. Given Spherical Gaussians serve as intermedi- dia for the robust edge representation, we further introduce a novel optimization-based algorithm called SGCR to directly extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying great efficiency. Code is available at https://github.com/Martinyxr/SGCR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In computer graphics and vision, photorealistic novel view synthesis (NVS) and accurate geometry reconstruction are two significant objectives. 3D Gaussian Splatting [19] offers an appealing explicit representation—3D Gaussians, resulting in high-quality synthesis of novel views and real-time rendering speed. However, it falls short in capturing accurate 3D geometric structures since the anisotropic 3D Gaussians, which are intended for efficient rasterization and sufficient expression ability, do not align with the geometric features (surfaces, edges and corners) in a faithful manner. A question naturally arises: Can we utilize 3D Gaussians" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5793" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 167, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 167, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 167, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 167, + 164 + ], + "type": "image", + "image_path": "127d104e1879a4c7f9cdbeea0d6311b729694c564043572e8308ed1bbc7b9fea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 167, + 167, + 179 + ], + "lines": [ + { + "bbox": [ + 57, + 167, + 167, + 179 + ], + "spans": [ + { + "bbox": [ + 57, + 167, + 167, + 179 + ], + "type": "text", + "content": "(a) 3D Gaussian Splatting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 69, + 294, + 164 + ], + "blocks": [ + { + "bbox": [ + 186, + 69, + 294, + 164 + ], + "lines": [ + { + "bbox": [ + 186, + 69, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 186, + 69, + 294, + 164 + ], + "type": "image", + "image_path": "c1143525946b7625596f3e50a6476fc618efe0c54f9a3b38986126672a16b54a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 168, + 293, + 178 + ], + "lines": [ + { + "bbox": [ + 189, + 168, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 189, + 168, + 293, + 178 + ], + "type": "text", + "content": "(b) Spherical Gaussians" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "content": "Figure 2. Visualization for optimized Gaussians. (a) Gaussians from original 3D Gaussian Splatting. (b) Our Spherical Gaussians." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 295, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 245 + ], + "type": "text", + "content": "for the direct representation of intricate geometric structures while retaining their benefits in 2D rendering?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 341 + ], + "type": "text", + "content": "Several recent works attempt to directly address the task of sdf generation or mesh reconstruction [6, 12, 17, 52] with 3D Gaussians. In this paper, we try to answer this question from a new perspective—reconstructing 3D parametric curves with 'Spherical Gaussians'. Following the settings of classic rendering works [19, 29], we use a set of calibrated multi-view images as input, but our goal substitutes novel view synthesis for 3D edge/curve reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 640 + ], + "type": "text", + "content": "Feature curves serve as critical geometric cues in characterizing the structure of 3D shapes, benefiting surface reconstruction, shape perception or other downstream tasks. Traditional methods [9, 13, 18, 21] usually work with the direct representation of 3D shapes, such as polygonal meshes or point clouds. However, the presence of sharp edges may be compromised or entirely overlooked on account of imperfect 3D scanning and acquisition, especially when dealing with sparse and noisy data. In recent years, learning based-methods [26, 28, 41, 54] are proposed to mitigate this, but obtaining precise ground-truth labels for 3D edge supervision involves laborious manual annotation, and inevitably limiting their generality. On the contrary, 2D images is easily accessible and allow for straightforward identification of feature edges through mature edge detection methods, making 2D supervision more attractive for edge processing tasks. Despite advantages in detection and self-supervision, there still remains a huge obstacle to overcome—establishing correspondence between edges across different views when working with multi-view images, in other words, how to combine information from multiple views to construct a complete 3D structure? This is exactly the factor for our design of Spherical Gaussians (as well as SGCR), which serve as a robust bridge to connect 3D depiction with 2D guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Our method consists of two main stages: Spherical Gaussians generation and parametric curves extraction. During the first stage, we utilize 2D edge maps to refine Gaussians using a view-based rendering loss compared to the ground-truth edge map derived from the image of that view. Optimizing Gaussians directly as vanilla 3D Gaus" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 142 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 142 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 142 + ], + "type": "image", + "image_path": "001a975b016e292d31f6e58d6ea667e93a57fab3111ecdc60028ce89abf4d888.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "text", + "content": "Figure 3. Illustration for fitting Bézier curve from Spherical Gaussians. " + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "inline_equation", + "content": "P_{4}" + }, + { + "bbox": [ + 313, + 150, + 555, + 173 + ], + "type": "text", + "content": " are weighted control points of the curve." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 181, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 337 + ], + "type": "text", + "content": "sian Splatting [19] is problematic due to the sparse distribution of edge pixels and the diversity of Gaussian attributes. The outcome Gaussians may recover 2D images to some extent but exhibit redundant irregular ellipsoids (see Fig. 2 (a)) and lack solid 3D geometric meanings, making following reconstruction tasks quite difficult. Therefore, we implement several constraints to transform standard 3D Gaussians into Spherical Gaussians. We introduce grid initialization, fixing-covariance training and a two-phase pruning strategy to guide the optimization process. Our final distribution of gaussians lie faithfully on sharp edges of objects in 3D space, appearing the same shape like spheres (see Fig. 2 (b)) and holding attributes in opacity and gray-scale color." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 456 + ], + "type": "text", + "content": "In the second stage, we utilize these discrete Spherical Gaussians to extract smooth 3D parametric curves for final reconstruction of edges. We present a novel optimization-based algorithm called SGCR, which involves fitting edges represented by Spherical Gaussians onto rational Bezier curves. The process begins with line fitting and then gradually refines the curves by global optimization. The final results are stored by all control points and weights in these curves. An illustration for curve extraction from Spherical Gaussians is shown in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "content": "We evaluate our method using a variety of object categories with complex models and scenes from ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. Extensive experiments show that SGCR, which utilizes Spherical Gaussians as bridge to reconstruct 3D parametric curves by 2D supervisions, outperforms existing state-of-the-art methods on all metrics and achieves great efficiency. In summary, our contributions include:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 555, + 553, + 639 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 555, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 578 + ], + "type": "text", + "content": "- We propose a new 3D edge extraction method from multiview 2D images using Spherical Gaussians (SG)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 579, + 553, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 553, + 614 + ], + "type": "text", + "content": "- We introduce a new optimization-based algorithm (SGCR) tailored for directly reconstructing 3D parametric curves from our generated Spherical Gaussians." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 639 + ], + "type": "text", + "content": "- Our approach achieves the state-of-the-art performance as well as high efficiency on curve reconstruction tasks." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 651, + 400, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 400, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 400, + 663 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 553, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 553, + 685 + ], + "type": "text", + "content": "2.1. Image-based 3D Representation and Rendering" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Image-based rendering (IBR) techniques use a collection of 2D images of a scene to generate the visual represent-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5794" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "tation of the scene and render novel views. Significant advancements have been achieved by Neural Radiance Fields (NeRF) [29], which depicts a scene's geometry and visual attributes as a radiance field, enabling the retrieval of color and volume density at any point in space and from any viewing angle for the purpose of rendering. Recently, the unstructured GPU-optimized 3D Gaussian Splitting (3DGS) [19] presents the explicit Gaussian primitive that achieves enhanced rendering speeds and improved quality without the reliance on neural components. It has also been quickly extended to multiple domains [12, 27, 45, 53, 55]. We convert standard 3D Gaussians into Spherical Gaussians, enabling significant advantages in both view rendering and the representation of geometric structures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 197, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 197, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 197, + 261 + ], + "type": "text", + "content": "2.2.3D Curve Reconstruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 267, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 422 + ], + "type": "text", + "content": "The convention of 3D curve reconstruction is based on point cloud. Traditional methods [3, 9, 21, 22, 40] focus on local geometric properties of point clouds, such as normals [7, 42], curvatures [49] and hierarchical clustering [10]. Data-driven methods [15, 26, 28, 39] often adopt edge detection as a binary classification for point clouds. As network architectures advance, the classifier used for edge detection has become more complex models [5, 33, 41, 51]. RFEPS [47] presents a multistage optimization to augment the point set so that sufficient points are generated on potential edges of geometry. NerVE [54] presents a neural volumetric edge representation learned through a volumetric learning framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 639 + ], + "type": "text", + "content": "Extraction curves from multi-view images was pioneered by [4] which introduces a full SfM system using line segments. With the developments of line detections [30, 31, 48] and matching [2, 30, 32], several works [14, 16, 24, 25, 43] have attempted to revisit the line mapping problem, but they are restricted to straight lines and frequently generate fragmented small lines when dealing with curves. Lifting 2D sketches/curves to 3D curves [11, 34, 35, 46] has also been exploited, but the ability to generate parametric curves is still limited. Recently, NEF [50] learns a neural implicit field to capture the distribution of edge densities, and EMAP [23] encodes 3D edge distance and direction in Unsigned Distance Functions (UDF) from multi-view edge maps. However, they all require large training time for even a simple scene. In contrast, our method employs Spherical Gaussians as the direct representation for edges, achieving 3D parametric curve reconstruction both accurately and efficiently." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 651, + 111, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 111, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 111, + 662 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 134, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 134, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 134, + 684 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "3D Gaussian Splatting. 3DGS [19] is a recently popular method for representing 3D scenes with 3D Gaussian" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "primitives, achieving both explicit representation and high-quality real-time rendering. Each Gaussian primitive is defined by a covariance matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", and a center point " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " which is the mean of the Gaussian. The 3D distribution can be represented as: " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "G(x) = e^{-\\frac{1}{2} (x - p)^{\\mathrm{T}}\\Sigma^{-1}(x - p)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": "A complete 3D Gaussian is given by its mean/center " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": ", color represented with spherical harmonics coefficients " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": ", quaternion " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": ", and scaling factor " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": ". For a given pixel, the combined color and opacity from multiple Gaussians are weighted by Eq. 3. The color blending for overlapping points is: " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\sum_{i \\in N} \\mathbf{c}_i \\alpha_i \\prod_{j=1}^{i-1} (1 - \\alpha_j)" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i, \\alpha_i" + }, + { + "bbox": [ + 313, + 133, + 553, + 228 + ], + "type": "text", + "content": " denote the color and density of a point." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "content": "Spherical Gaussians. Our Spherical Gaussians enjoy the same nature of 3D Gaussians, but make the following changes for geometry priority and simplicity: (1) Remove the covariance matrix " + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "content": " of Gaussians (attributes of scaling " + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "content": ") and replace it with a fixed radius " + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "content": ", i.e. convert Gaussian primitives from various ellipsoids into spheres of uniform size. (2) Remove spherical harmonics coefficients " + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 228, + 553, + 384 + ], + "type": "text", + "content": " and modify color to one-dimensional gray value (just for edge map rendering). According to our experiments, these changes do not hamper the backward propagation of gradients from 2D edge-map supervision, but impose great regularization on geometry distribution of 3D isotropic Gaussians, facilitating reconstruction tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 552 + ], + "type": "text", + "content": "Anisotropic Gaussians v.s. Isotropic Gaussians. Traditional 3DGS [19] and its subsequent works mainly center on the technique of 'splatting'—utilizing different anisotropic Gaussian splats to accurately represent the entire scene. However, our work focuses on each 'Gaussian' itself, essentially displaying a reverse process: decomposing the splating results into small 'atoms' and assigning clear geometric meanings to every individual Gaussian. A thin and elongated Gaussian seems to be more efficient in representing edges, but it can be split into smaller 'basis-Gaussians' with few sacrifices but for better interpretation and convenience, especially when following pointcloud-related works. That's why we favour Spherical Gaussians and more discussions can be found in supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 560, + 487, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 487, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 487, + 573 + ], + "type": "text", + "content": "3.2. Spherical Gaussians Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 578, + 397, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 397, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 397, + 589 + ], + "type": "text", + "content": "3.2.1. Initialization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "Our full pipeline is illustrated in Fig. 4. Given a set of calibrated multi-view images, we first adopt edge detection method [38] to generate 2D edge maps for each view. 3DGS uses structure-from-Motion (SfM) [36] points as initialization for Gaussians, but this approach is not ideal for handling edge maps. Thus, we design grid initialization for Spherical Gaussians. We create a " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "50 \\times 50 \\times 50" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": " grid with points evenly distributed in space of " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "[0,1]^3" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": " (see Fig. 4—Grid initialization) as the initial centers for our Spherical Gaussians. This setting is better than random initialization" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5795" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 274 + ], + "type": "image", + "image_path": "567532b9367e791888bd11304d8fcac8d67e532a51c3e2b2a3d0c39c65b8838e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 553, + 293 + ], + "type": "text", + "content": "Figure 4. Pipeline illustration across 2D and 3D modality. Stage1: Spherical Gaussians generation. Stage2: Parametric curves extraction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "type": "text", + "content": "in producing neat structures. And we choose a constant radius " + }, + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "type": "text", + "content": " to replace the covariance matrix with Spherical Gaussians of the same size. In our experiments, we use " + }, + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "type": "inline_equation", + "content": "r_0 = 0.005" + }, + { + "bbox": [ + 55, + 308, + 295, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 151, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 151, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 151, + 373 + ], + "type": "text", + "content": "3.2.2. Training Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 400 + ], + "type": "text", + "content": "Edge Loss. The original loss functions in 3DGS [19] to train 3D Gaussians for each view is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 408, + 295, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 408, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 111, + 408, + 295, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {D - S S I M} \\tag {1}", + "image_path": "f69e1c4477eb6051bb763dce025cee1c6766dcfbb3f6f0e7f7cabd8f0eac8f68.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{D-SSIM}" + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "text", + "content": " are supervised by the ground-truth view. However, in edge maps, edge pixels are often sparsely distributed. Relying solely on this loss function will easily make the training process stuck in local optima, causing the opacities and colors for all Gaussians to converge towards zero and resulting in all rendered images becoming black. Hence, we opt for a new edge-aware loss " + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{edge}" + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 429, + 296, + 547 + ], + "type": "text", + "content": " in order to balance the impact of both edge and non-edge pixels within each edge map. The edge-aware loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 295, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 593 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e d g e}} = \\frac {N _ {I} - | E _ {I} |}{N _ {I}} \\sum_ {i \\in E _ {I}} \\| I _ {i} - \\hat {I} _ {i} \\| _ {2} ^ {2} + \\frac {| E _ {I} |}{N _ {I}} \\sum_ {i \\notin E _ {I}} \\| I _ {i} - \\hat {I} _ {i} \\| _ {2} ^ {2} \\tag {2}", + "image_path": "9c70fca1ba3a4957c95bca32eace5c38a4516a36c695d7667d5f27f3ad2f15e4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "N_{I}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " is the image resolution and " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " refers to all edge pixels in current edge map (gray scale larger than threshold " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\eta = 0.3" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " refers to the pixel in Gaussian-rendering image and ground-truth edge-map image respectively. This edge-aware loss enhances the ability of Spherical Gaussians to identify edges from edge pixels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Opacity-Color Loss. Another issue during edge training is about occluded edges. Some 2D edge maps may lose true edge pixels due to occlusions from these image views (see the difference between Fig. 4-Edge maps and Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "type": "text", + "content": "views), while from other views the occlusion may not happen. This inconsistency among views will confuse Spherical Gaussians during training, potentially leading to unstable optimization. During training, color attribute is directly supervised by " + }, + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{edge}" + }, + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "type": "text", + "content": " loss, while opacity is indirectly optimized and reset at certain iterations. This difference sets them apart, and equating them impedes smooth optimization. Therefore, we introduce opacity-color loss " + }, + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{oc}" + }, + { + "bbox": [ + 313, + 308, + 555, + 429 + ], + "type": "text", + "content": " for all Spherical Gaussians to make the value of Gaussian opacity more consistent. This opacity-color loss is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 389, + 438, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 438, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 389, + 438, + 553, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o c} = \\sum_ {i = 1} ^ {N _ {S G}} \\| o _ {i} - c _ {i} \\| _ {2} ^ {2} \\tag {3}", + "image_path": "393787b561fdc776b5c0db84a1f22747a11c987a6b876a2d1748fa3b839a8c30.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "inline_equation", + "content": "N_{SG}" + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "content": " is the number of all Spherical Gaussians during current training step, " + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "content": " refer to opacity and color attribute of each Spherical Gaussian respectively. " + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{oc}" + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "content": " helps to keep consistency between edge density and color intensity during multi-view supervision. This mechanism ensures that occluded edges—represented by Gaussians with low opacity values—are less susceptible to premature pruning during optimization, resulting in more complete 3D reconstructions. It also helps rectify edge predictions missed by 2D detectors. In our experiment, after implementing " + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{oc}" + }, + { + "bbox": [ + 313, + 482, + 555, + 625 + ], + "type": "text", + "content": ", Gaussian-rendered images can depict complete edge structures with no occlusion (Fig. 4—Rendering Views)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 626, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 555, + 673 + ], + "type": "text", + "content": "Regularization Loss. To restrict total number of Spherical Gaussians and accelerate convergence, we add an extra regularization term " + }, + { + "bbox": [ + 313, + 626, + 555, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {regular }}" + }, + { + "bbox": [ + 313, + 626, + 555, + 673 + ], + "type": "text", + "content": " on Gaussian's opacity. The regularization loss is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 683, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 683, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 380, + 683, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\sum_ {i = 1} ^ {N _ {S G}} \\log \\left(1 + \\frac {o _ {i} ^ {2}}{0 . 5}\\right) \\tag {4}", + "image_path": "3647cc6c6b1d678816fdc2258e1c71c44c0813468f3ce89fc8bd4d55322ac689.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5796" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 190 + ], + "type": "image", + "image_path": "e6d6f64c05812fb532543fe202cf803d0d513edbae22feaa1a90f80d353cadd7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 195, + 289, + 207 + ], + "lines": [ + { + "bbox": [ + 61, + 195, + 289, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 289, + 207 + ], + "type": "text", + "content": "Figure 5. The change on number of Gaussians during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "type": "inline_equation", + "content": "N_{SG}" + }, + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "type": "text", + "content": " is the number of current Spherical Gaussians and " + }, + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 55, + 214, + 296, + 285 + ], + "type": "text", + "content": " is each Gaussian's opacity. During the training process, this loss function works implicitly because Spherical Gaussians with opacity below a certain threshold will be pruned (as in 3DGS [19]), allowing for a comfortable number of total Gaussian primitives for next stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 310 + ], + "type": "text", + "content": "To summarize, our final loss function for training Spherical Gaussians is presented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 319, + 295, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 295, + 334 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left(1 - \\lambda_ {1}\\right) \\mathcal {L} _ {\\text {e d g e}} + \\lambda_ {1} \\mathcal {L} _ {D - S S I M} + \\lambda_ {2} \\mathcal {L} _ {o c} + \\lambda_ {3} \\mathcal {L} _ {\\text {r e g}} \\tag {5}", + "image_path": "cced898e5ca40593c89430f00b9dd9f388a812c2485145bf1a0cfeb16a385ac0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "type": "text", + "content": "where the balancing parameters " + }, + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "type": "text", + "content": " are set to 0.2, 2 and 0.01 in our experiments respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 160, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 160, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 160, + 385 + ], + "type": "text", + "content": "3.2.3. Training Strategy" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "type": "text", + "content": "The effective training strategy is essential for creating properly structured Spherical Gaussians. We adopt a two-phase training strategy, each phase consists of 3,000 iterations. During the first phase, we conduct densification (split and clone) of Spherical Gaussians every 200 iterations. And we reset the opacity attribute to 0.1 every 1,000 iterations. At the end of first phase, we design a final-prune—removing all Spherical Gaussians whose opacity " + }, + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "type": "inline_equation", + "content": "o_i < 0.5" + }, + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "type": "inline_equation", + "content": "c_i < 0.1" + }, + { + "bbox": [ + 55, + 388, + 296, + 604 + ], + "type": "text", + "content": ". This operation largely reduces the total number of Gaussians, leading to substantial reductions in both training time and noise. During the second phase, we no longer densify Spherical Gaussians but focus on refining their positions and attributes. In the end, we perform another final-prune to retain all significant Spherical Gaussians (at quantity of 1k to 5k) with their attributes. The variation of numbers for Spherical Gaussians during all iterations is depicted in Fig. 5. The whole training process of Spherical Gaussians for single scene takes only about 1 minute." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 296, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 635 + ], + "type": "text", + "content": "3.3. Parametric Curve Extraction from Spherical Gaussians" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "After the first stage, we get the generated Spherical Gaussians (denoted as " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ") aligned with object's edges, resembling a 3D wireframe constructed by many small spheres. In the second stage (Fig. 4—Stage 2), we present of Spherical-Gaussians Curve Reconstruction (SGCR) to accomplish the final 3D edge reconstruction task. SGCR is an" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 72, + 419, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 419, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 419, + 84 + ], + "type": "text", + "content": "Algorithm 1 Line Fitting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "spans": [ + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "type": "text", + "content": "Input: a set of Spherical Gaussians " + }, + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "type": "text", + "content": " with radius " + }, + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 315, + 86, + 533, + 97 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 98, + 553, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 98, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 314, + 98, + 553, + 122 + ], + "type": "text", + "content": "Output: the set of line endpoints " + }, + { + "bbox": [ + 314, + 98, + 553, + 122 + ], + "type": "inline_equation", + "content": "L = \\{(p_i, q_i)\\}_{i=1}^{|L|}" + }, + { + "bbox": [ + 314, + 98, + 553, + 122 + ], + "type": "text", + "content": " for the coarse line-fitting of 3D edges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "text", + "content": "Denote: " + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "inline_equation", + "content": "P(G)" + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "text", + "content": " is the centers of " + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "inline_equation", + "content": "N(0,1)" + }, + { + "bbox": [ + 314, + 122, + 553, + 146 + ], + "type": "text", + "content": " is the Standard Normal Distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 149, + 534, + 356 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 321, + 149, + 360, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 149, + 360, + 159 + ], + "spans": [ + { + "bbox": [ + 321, + 149, + 360, + 159 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 321, + 149, + 360, + 159 + ], + "type": "inline_equation", + "content": "L = \\emptyset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 161, + 412, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 161, + 412, + 172 + ], + "spans": [ + { + "bbox": [ + 321, + 161, + 412, + 172 + ], + "type": "text", + "content": "2: while " + }, + { + "bbox": [ + 321, + 161, + 412, + 172 + ], + "type": "inline_equation", + "content": "|G| > N_0" + }, + { + "bbox": [ + 321, + 161, + 412, + 172 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 173, + 428, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 173, + 428, + 185 + ], + "spans": [ + { + "bbox": [ + 321, + 173, + 428, + 185 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 173, + 428, + 185 + ], + "type": "inline_equation", + "content": "S_{min} =" + }, + { + "bbox": [ + 321, + 173, + 428, + 185 + ], + "type": "text", + "content": " MaxIntValue" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 321, + 186, + 413, + 195 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 197, + 481, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 197, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 197, + 481, + 209 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 197, + 481, + 209 + ], + "type": "inline_equation", + "content": "p, q = \\text{RandomChoice}(P(G))" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 209, + 534, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 209, + 534, + 220 + ], + "spans": [ + { + "bbox": [ + 321, + 209, + 534, + 220 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 209, + 534, + 220 + ], + "type": "inline_equation", + "content": "L_{p,q} = \\text{interpolate}(p, q, N_s) + r_0 * N(0,1)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 220, + 517, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 517, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 517, + 239 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 220, + 517, + 239 + ], + "type": "inline_equation", + "content": "G_{i} = \\{g\\in G|\\min_{x\\in L_{p,q}}||P(g) - x||_{2}^{2} < \\delta_{1}\\}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 239, + 495, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 239, + 495, + 258 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 495, + 258 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 239, + 495, + 258 + ], + "type": "inline_equation", + "content": "p', q' = \\operatorname*{arg min}_{p,q} \\mathcal{L}_{CD}(L_{p,q}, P(G_i))" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 258, + 470, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 258, + 470, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 470, + 274 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 258, + 470, + 274 + ], + "type": "inline_equation", + "content": "S_{i} = \\min_{p,q}\\mathcal{L}_{CD}(L_{p,q},P(G_{i}))" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 274, + 428, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 274, + 428, + 285 + ], + "spans": [ + { + "bbox": [ + 318, + 274, + 428, + 285 + ], + "type": "text", + "content": "10: if " + }, + { + "bbox": [ + 318, + 274, + 428, + 285 + ], + "type": "inline_equation", + "content": "S_{i} \\leq S_{min}" + }, + { + "bbox": [ + 318, + 274, + 428, + 285 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 285, + 515, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 285, + 515, + 297 + ], + "spans": [ + { + "bbox": [ + 318, + 285, + 515, + 297 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 318, + 285, + 515, + 297 + ], + "type": "inline_equation", + "content": "S_{min} = S_i, G' = G_i, (\\hat{p}, \\hat{q}) = (p', q')" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 298, + 379, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 298, + 379, + 308 + ], + "spans": [ + { + "bbox": [ + 318, + 298, + 379, + 308 + ], + "type": "text", + "content": "12: end if" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 310, + 375, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 310, + 375, + 319 + ], + "spans": [ + { + "bbox": [ + 318, + 310, + 375, + 319 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 322, + 473, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 322, + 473, + 334 + ], + "spans": [ + { + "bbox": [ + 318, + 322, + 473, + 334 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 318, + 322, + 473, + 334 + ], + "type": "inline_equation", + "content": "G = G - G^{\\prime},L = L\\cup \\{(\\hat{p},\\hat{q})\\} ;" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 335, + 375, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 335, + 375, + 343 + ], + "spans": [ + { + "bbox": [ + 318, + 335, + 375, + 343 + ], + "type": "text", + "content": "15: end while" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "spans": [ + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "type": "text", + "content": "16: return " + }, + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 318, + 346, + 375, + 356 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 380, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 405 + ], + "type": "text", + "content": "optimization-based algorithm guided by " + }, + { + "bbox": [ + 313, + 380, + 553, + 405 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 380, + 553, + 405 + ], + "type": "text", + "content": " which includes two parts: line fitting and global optimization." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 411, + 394, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 394, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 394, + 423 + ], + "type": "text", + "content": "3.3.1. Line Fitting" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "text", + "content": "We first fit a set of line segments to " + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "text", + "content": "'s centers as the coarse fitting. The pseudo code of this part is shown in Algorithm 1. The whole process consists of several line-fitting turns and " + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "text", + "content": " will be gradually fitted and partly removed by one line each turn. Line fitting will stop when the number of remaining " + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "text", + "content": " is less than " + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "inline_equation", + "content": "N_0 = 5" + }, + { + "bbox": [ + 313, + 426, + 554, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": "In each turn, we conduct " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": " random searches and record the best result. At the beginning of " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": "-th search, we randomly choose two nearest Gaussian centers as initial endpoints " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "p, q" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": " of the fitting line, and interpolate " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": " points along the line. To simulate the shape of Spherical Gaussians, we dilate the sampled points up by duplicating and adding Gaussian noise with standard deviation equals " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": "'s radius " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": ". Then we apply Chamfer Distance to compute the fitting loss between dilated line points and " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": ". This Chamfer loss is defined as:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 319, + 622, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 622, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 319, + 622, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {C D} \\left(L _ {p, q}, P \\left(G _ {i}\\right)\\right) = \\frac {\\gamma_ {1}}{\\left| L _ {p , q} \\right|} \\sum_ {x \\in L _ {p, q}} \\min _ {y \\in P \\left(G _ {i}\\right)} \\| x - y \\| _ {2} ^ {2} \\\\ + \\frac {1}{| P (G _ {i}) |} \\sum_ {y \\in P (G _ {i})} \\min _ {x \\in L _ {p, q}} \\| x - y \\| _ {2} ^ {2} \\tag {6} \\\\ \\end{array}", + "image_path": "ed115dd989ecd8dc0b786b293d16c3a35618c07964d3f7f7a880648731f19767.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 350, + 695, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 695, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 695, + 553, + 715 + ], + "type": "interline_equation", + "content": "G _ {i} = \\left\\{g \\in G \\mid \\min _ {x \\in L _ {p, q}} \\left| | P (g) - x \\| _ {2} ^ {2} < \\delta_ {1} \\right. \\right\\} \\tag {7}", + "image_path": "c5f5aa921d9ae71aa8278dffc5f5164bb7dcfeecabdfd82691efdf32ea2c4096.jpg" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5797" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "L_{p,q}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " denotes the dilated line points between " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "p, q" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "-th search, " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "P(G)" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " represents Gaussian centers of all current " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\gamma_1 = 2" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " is a balance parameter. " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " is a subset of " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " whose centers are close to " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "L_{p,q}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " within a threshold " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\delta_1 = 0.02" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". We optimize " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "p, q" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " by minimizing Eq. (6) with certain iterations and use the value of " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CD}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " after optimization as the final fitting score " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "-th search." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": "For each turn, suppose the best fitting is in " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": "-th search, i.e. " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "S_{k} = \\min S_{i}" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": ", we record the optimized endpoints " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "p^{\\prime}, q^{\\prime}" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": "-th search and delete " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "G_{k}" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": " to update the current " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": ". The next turn will then proceed accordingly. Once all the turns are completed, we will get a set of paired line endpoints " + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "inline_equation", + "content": "L = \\{(p_{i}, q_{i})\\}_{i=1}^{|L|}" + }, + { + "bbox": [ + 55, + 156, + 294, + 242 + ], + "type": "text", + "content": " which together form an approximate line-fitting result (Fig. 4—Line Fitting)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 171, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 171, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 171, + 258 + ], + "type": "text", + "content": "3.3.2. Global Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "type": "text", + "content": "After line fitting, we resume all " + }, + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "type": "text", + "content": " deleted in previous part and take opacity attribute into account as the fine fitting. We interpolate another two control points between each line endpoint pair " + }, + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "type": "inline_equation", + "content": "(p_i, q_i)" + }, + { + "bbox": [ + 55, + 262, + 294, + 321 + ], + "type": "text", + "content": " and introduce point weight to initialize 3rd order rational Bézier curves, which is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 328, + 294, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 328, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 111, + 328, + 294, + 360 + ], + "type": "interline_equation", + "content": "B (u) _ {p, w} = \\frac {\\sum_ {i = 0} ^ {3} B _ {3 , i} (u) p _ {i} w _ {i}}{\\sum_ {i = 0} ^ {3} B _ {3 , i} (u) w _ {i}} \\tag {8}", + "image_path": "14e1e038308833b384710d1fd782f4e2163393b09cd144629b23ea94cb2ae54c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "inline_equation", + "content": "B_{3,i}" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": " is basis functions of Bézier curves, " + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": "-th control points of curve and " + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": " is the weight of control point " + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "inline_equation", + "content": "w_i \\equiv 1" + }, + { + "bbox": [ + 55, + 365, + 295, + 436 + ], + "type": "text", + "content": " for simple Bézier curves). We opt for the rational Bézier curves over simple Bézier curves because simple Bézier curves cannot perfectly fit circles, which are frequently seen in object's edges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "text", + "content": "We also sample " + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "text", + "content": " points for each Bézier curve and dilate them up as in line fitting. Since " + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "text", + "content": " with higher opacity values are more significant during fine fitting, we apply the weighted Chamfer Distance to compute the fine fitting loss between global curve points and all " + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 437, + 295, + 496 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 503, + 302, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 302, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 302, + 563 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {W C D} (L _ {a l l}, P (G)) = \\frac {\\gamma_ {2}}{| L _ {a l l} |} \\sum_ {x \\in L _ {a l l}} o _ {y _ {m i n}} \\cdot \\min _ {y \\in P (G)} \\| x - y \\| _ {2} ^ {2} \\\\ + \\frac {1}{| P (G) |} \\sum_ {y \\in P (G)} o _ {y} \\cdot \\min _ {x \\in L _ {\\text {a l l}}} \\| x - y \\| _ {2} ^ {2} \\tag {9} \\\\ \\end{array}", + "image_path": "dab1ff0ab52cd747590c971b55b7d36e73f634f712e1f3efb652925895f9639d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "L_{all}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " refers to global dilated points from all curves and " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "P(G)" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " refers to Gaussian centers of " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "o_y" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " is the opacity attribute in " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "-th " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\gamma_2 = 2" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " is another balance parameter. We also adopt endpoints loss to encourage connections between curves, which is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 301, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 301, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 301, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n d p o i n t s}} = \\sum_ {x, y \\in P _ {E}} \\min \\left(\\| x - y \\| _ {2} ^ {2} - \\delta_ {2}, 0\\right) \\frac {\\| x - y \\| _ {2} ^ {2}}{\\| x - y \\| _ {2} ^ {2} - \\delta_ {2}} \\tag {10}", + "image_path": "975b9aed4729a8556ca86761a1718bd548dcdcb8432e5c157817b40f60c38ee5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "P_{E}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " is the set of endpoints (the first and last control points) and " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{2}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " (we choose 0.01) is a threshold to only regularize endpoints which are close to each other. The final" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "objective function to globally optimize all curves is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 98, + 553, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 98, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 354, + 98, + 553, + 122 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\arg \\min \\quad \\left(\\mathcal {L} _ {W C D} + \\lambda \\mathcal {L} _ {\\text {e n d p o i n t s}}\\right) \\tag {11} \\\\ \\{\\{z _ {i} ^ {j}, w _ {i} ^ {j} \\} _ {j = 1} ^ {4} \\} _ {i = 1} ^ {| L |} \\\\ \\end{array}", + "image_path": "a621b34f192a140c46dbb23b90e1c41e99cebb8e78d523a4fae3b5e8ad19d595.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "inline_equation", + "content": "z_{i}^{j}" + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "inline_equation", + "content": "w_{i}^{j}" + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "text", + "content": " are coordinates and weights of four control points in each Bézier curve, " + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 130, + 553, + 179 + ], + "type": "text", + "content": " is a balancing parameter which is set to 0.005 in our experiments. The pseudo code of this part can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": "After global optimization, we finally extract 3D parametric curves from Spherical Gaussians in form of 3rd order rational Bezier curves (Fig. 4—Rational Bezier Curves)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 228, + 394, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 394, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 394, + 242 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 249, + 372, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 372, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 372, + 260 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "type": "text", + "content": "We conduct experiments on the ABC [20], ModelNet [44], DTU [1] and Replica [37] datasets. ABC [20] consists of massive CAD models with edge annotations, and we use a portion of it—ABC-NEF, provided by [50]—which consisting of 115 distinct and challenging CAD models. ModelNet [44] includes 3D models of everyday objects such as chairs and tables, but it does not have ground-truth edge annotations. So we manually select and annotate edges on ModelNet ground-truth meshes. Our edge dataset consists of 120 furniture models with annotated edges and is used for a more comprehensive evaluation. We utilize BlenderProc [8] to generate rendered images with these objects. We generate 50 views of " + }, + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "type": "text", + "content": " images for each object and adopt PiDiNet [38] to predict 2D edge maps. DTU [1] and Replica [37] are real-world datasets with multi-view images as input and we use it for evaluation on real-world reconstruction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 468, + 399, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 399, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 399, + 480 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 593 + ], + "type": "text", + "content": "Settings. We compare our method against the state-of-the-art methods in 3D curve reconstruction, including RFEPS [47], NerVE [54], NEF [50] and EMAP [23]. The first two requires point clouds as input, while the last two and ours use only 2D images in the same input frames and resolution. We all adopt their pre-trained models and corresponding reconstruction approaches to extract curves for evaluation. More experiments, discussions and details can be found in the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. We first normalize and align all reconstructed curves and ground truth edges into the range of " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". Then, we sample points on all reconstructed curves and measure the distances between these points and ground-truth edge points. When assessing IoU, precision, recall, and F-score, points are considered matched if there is at least one ground-truth edge point with an " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " distance smaller than 0.02. We also adopt the Chamfer Distance (CD) between point clouds to measure the geometric accuracy of reconstructed parametric curves. As reported in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5798" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 70, + 547, + 264 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 547, + 264 + ], + "type": "image", + "image_path": "7f3a7fb84705df8b4e48c578c3f2c6b64c2de11bcc9fcbf23801346cd2c82565.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 298 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons. The first two objects come from ABC-NEF dataset and the last one comes from ModelNet dataset. Ours(SG) refers to our generated Spherical Gaussian in the first stage, GT refers to the ground-truth edge." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 327, + 553, + 418 + ], + "blocks": [ + { + "bbox": [ + 138, + 307, + 473, + 321 + ], + "lines": [ + { + "bbox": [ + 138, + 307, + 473, + 321 + ], + "spans": [ + { + "bbox": [ + 138, + 307, + 473, + 321 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on ABC-NEF Dataset [50] and ModelNet Dataset[44]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 327, + 553, + 418 + ], + "lines": [ + { + "bbox": [ + 58, + 327, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 327, + 553, + 418 + ], + "type": "table", + "html": "
MethodABC-NEF DatasetModelNet DatasetInputTrainingReconstruction
CD↓Precision↑Recall↑F-score↑IoU↑CD↓Precision↑Recall↑F-score↑IoU↑ModalityTime↓Time↓
RFEPS [47]0.03230.89620.85600.86730.81900.07310.70320.59650.64120.5621Point Cloud-48s
NerVE [54]0.03890.95180.73070.82670.68450.02940.95680.80170.87240.7931Point Cloud>10h1s
NEF [50]0.03530.93870.88380.90440.82830.03210.90720.81580.85590.7475Images>1h119s
EMAP [23]0.02910.95130.89260.92100.84730.02900.91150.84880.87840.7751Images>2h40s
Ours0.02800.95460.90520.92600.86150.02640.96270.84150.89520.8164Images87s32s
", + "image_path": "2ef5b7ab525c6cc83530aa91db90419b7422e211529514c7c50026a38a6e7812.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 432, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 456 + ], + "type": "text", + "content": "Table 1, our method outperforms other methods in all metrics on ABC-NEF [50] and most metrics on ModelNet [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 567 + ], + "type": "text", + "content": "Qualitative Comparison. We illustrate the visual comparisons in Fig. 6. It can be observed that RFEPS [47] and NerVE [54] may miss edges with weak curvature, leading to incomplete reconstruction. NEF [50] and EMAP [23] struggle with handling thin structures and make errors in curve connection. On the contrary, our method benefits from aligned Spherical Gaussians (see Ours(SG) in Fig. 6) and maintains a high level of robustness on 3D curve reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Time Comparison. We also make comparisons on training time and average reconstruction time (see the last two columns in Tab. 1). NerVE [54] is a learning-based methods that rely on large point cloud datasets, requiring extensive training time. RFEPS [47] is learning-free but takes time in point cloud optimization. NEF [50] and EMAP [23] are supervised on images, but they all involve a time-consuming training process (several hours for single scene) for predicting edge densities or unsigned distance functions. Our SGCR generates and optimizes only a small number of Spherical Gaussians (with two stages completing within minutes, " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " faster than NEF and EMAP), making 3D" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 432, + 459, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 432, + 459, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 432, + 459, + 445 + ], + "type": "text", + "content": "curve reconstruction quite efficient." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 454, + 421, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 421, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 421, + 466 + ], + "type": "text", + "content": "4.3. Real-world Scene" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 534 + ], + "type": "text", + "content": "We further test our performance in real-world scenes on DTU [1] and Replica [37] dataset. They present challenging objects and complex scenes for edge reconstruction respectively. The visual comparisons against NEF [50] and EMAP [23] on DTU [1] dataset are presented in Fig. 7. NEF misses" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 555, + 552, + 691 + ], + "blocks": [ + { + "bbox": [ + 315, + 555, + 552, + 691 + ], + "lines": [ + { + "bbox": [ + 315, + 555, + 552, + 691 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 552, + 691 + ], + "type": "image", + "image_path": "8806b33bc763145fe29e1f923c1fedbe9e139e270d54a08c2d8c28599efc70e7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 698, + 548, + 711 + ], + "lines": [ + { + "bbox": [ + 320, + 698, + 548, + 711 + ], + "spans": [ + { + "bbox": [ + 320, + 698, + 548, + 711 + ], + "type": "text", + "content": "Figure 7. Comparisons on real-world scenes from DTU [1]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5799" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 140, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 140, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 140, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 140, + 171 + ], + "type": "image", + "image_path": "906342259482a6d534a982ef1a34972564f054bd3e4a6ad66bf4e81166c34346.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 295, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 295, + 190 + ], + "type": "text", + "content": "Figure 8. Comparisons on real-world scenes from Replica [37]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 68, + 219, + 170 + ], + "blocks": [ + { + "bbox": [ + 143, + 68, + 219, + 170 + ], + "lines": [ + { + "bbox": [ + 143, + 68, + 219, + 170 + ], + "spans": [ + { + "bbox": [ + 143, + 68, + 219, + 170 + ], + "type": "image", + "image_path": "a2e97f069488d28b0bfe006109edaa907b4a7c39c51aa3ad78e6bdb640854565.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 220, + 68, + 296, + 170 + ], + "blocks": [ + { + "bbox": [ + 220, + 68, + 296, + 170 + ], + "lines": [ + { + "bbox": [ + 220, + 68, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 220, + 68, + 296, + 170 + ], + "type": "image", + "image_path": "7bcfd47ae5971cffc3fb5547026ec96d05837a42ea07318e4df5ca4222c32e81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "type": "text", + "content": "details when fitting complex parts, and EMAP produces too many chaotic and redundant line segments, while our method reconstructs edges accurately and preserves details clearly. Another visual comparison on Replica [37] dataset are presented in Fig. 8. It can be seen that our method still works well on room-level edge reconstruction (NEF fails, and EMAP's training time for single real-world scene is " + }, + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 55, + 197, + 296, + 304 + ], + "type": "text", + "content": " slower than ours), demonstrating our method's great generalization ability and broad applicability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 156, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 156, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 156, + 324 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 330, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 462 + ], + "type": "text", + "content": "Pipeline. We conduct ablation studies on ABC-NEF dataset [50] to evaluate the effectiveness of core designs in our method, including Spherical Gaussians, loss designs and two-phase training strategy in the first stage, as well as line fitting and global optimization in the second stage. Tab. 2 lists the quantitative results. w/o Spherical Gaussians stands for using original 3D Gaussians [19] trained by edge maps to directly represent edges and this result is quite poor, highlighting the significance of Spherical Gaussians. Statistics in Tab. 2 reveal that removing any component from our pipeline negatively impacts the final result." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 56, + 492, + 296, + 641 + ], + "blocks": [ + { + "bbox": [ + 77, + 472, + 272, + 484 + ], + "lines": [ + { + "bbox": [ + 77, + 472, + 272, + 484 + ], + "spans": [ + { + "bbox": [ + 77, + 472, + 272, + 484 + ], + "type": "text", + "content": "Table 2. Quantitative ablation studies on our pipeline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 492, + 296, + 641 + ], + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 641 + ], + "type": "table", + "html": "
PipelineCD↓F-score↑IoU↑
w/o Spherical Gaussians0.1420.3300.194
w/o grid initialization0.0320.9010.822
w/o edge loss Ledge0.0520.8530.714
w/o opacity-color loss Loc0.0510.8590.723
w/o regularization loss Regular0.0290.9130.846
w/o two-phase training strategy0.0300.9200.860
w/o line fitting0.0300.9110.822
w/o global optimization0.0320.8760.757
Full Model0.0280.9260.862
", + "image_path": "90b12519afb42447ad4f7aa43b738639f45d4950106d7082b91ed545ba6fec2e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Radius. We analyze the effect of radius in Spherical Gaussians. We test " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "r_0 = 0.002" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", 0.005, 0.01 and variable. Experimental results are presented in Tab. 3 (setting " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " as a variable does not perform well). As the radius increases, the total number of Spherical Gaussians decreases, with more" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 113, + 548, + 186 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 3. Comparison on different radius choices for Spherical Gaussians. " + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "N_{SG}" + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": ": Number of Spherical Gaussians. T1: Time for Gaussian generation. T2: Time for curve reconstruction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 113, + 548, + 186 + ], + "lines": [ + { + "bbox": [ + 319, + 113, + 548, + 186 + ], + "spans": [ + { + "bbox": [ + 319, + 113, + 548, + 186 + ], + "type": "table", + "html": "
Radius r0CD↓F-score↑IoU ↑NSGT1T2
0.0020.0320.9060.8495,12484s56s
0.0050.0280.9260.8622,36190s36s
0.010.0350.8810.810592101s30s
Variable0.0400.8630.7853,076110s44s
", + "image_path": "a34a52c6e66d251b49160250155ef64627bccc30bd0136807541e11cc67f8b7b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 201, + 555, + 262 + ], + "blocks": [ + { + "bbox": [ + 315, + 201, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 315, + 201, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 201, + 555, + 262 + ], + "type": "image", + "image_path": "2cebe1863b9a84ef3dfd1c7eecddfae1ebb5c5ce504a3e369d7a4cd81659ceba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 266, + 538, + 277 + ], + "lines": [ + { + "bbox": [ + 329, + 266, + 538, + 277 + ], + "spans": [ + { + "bbox": [ + 329, + 266, + 538, + 277 + ], + "type": "text", + "content": "Figure 9. Visualization for Gaussians with different radii." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 294, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 294, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 384 + ], + "type": "image", + "image_path": "93135909d991fdb2acf1ab31d25db43f5bc162bd2f7a902fb52ee4ac17d8fa84.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 386, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 409 + ], + "type": "text", + "content": "Figure 10. From left to right, we present 2D images, rendered Spherical Gaussians and final curves in 5, 10, 30 and 50 views." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 427, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 487 + ], + "type": "text", + "content": "time spent in the first stage and less time in the second stage. The visual comparisons can be found in Fig. 9. An inappropriate radius size may lead to noise or underfitting. Using " + }, + { + "bbox": [ + 313, + 427, + 554, + 487 + ], + "type": "inline_equation", + "content": "r_0 = 0.005" + }, + { + "bbox": [ + 313, + 427, + 554, + 487 + ], + "type": "text", + "content": " leads to the optimal reconstruction metrics as well as a reasonable running time." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 491, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 562 + ], + "type": "text", + "content": "Views. We also study the effect of the number of input views. We train Spherical Gaussians with 5, 10, 30 and 50 views respectively (all evenly distributed). The visual results are illustrated in Fig. 10. For simple cases, results with 10 views are close to satisfactory, while for complex cases, more views are required for better performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 391, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 391, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 391, + 594 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We have presented SGCR—a novel method using Spherical Gaussians for efficient 3D parametric curve reconstruction from multi-view images. The main concept involves generating Spherical Gaussians by 2D supervision and extract 3D parametric curves from them. Our method outperforms previous methods quantitatively and qualitatively, while enjoying great efficiency. SGCR also demonstrates the potential in integrating multi-modal information efficiently, bridging the gap between 2D rendering and 3D reconstruction." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5800" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 295, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 295, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 295, + 114 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62032011." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 137 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 145, + 295, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 295, + 189 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 190, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 190, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 295, + 232 + ], + "type": "text", + "content": "[2] Hichem Abdellali, Robert Frohlich, Viktor Vilagos, and Zoltan Kato. L2d2: Learnable line detector and descriptor. In 2021 International Conference on 3D Vision (3DV), pages 442-452. IEEE, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 235, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 294, + 277 + ], + "type": "text", + "content": "[3] Nina Amenta, Marshall Bern, and David Eppstein. The crust and the " + }, + { + "bbox": [ + 62, + 235, + 294, + 277 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 62, + 235, + 294, + 277 + ], + "type": "text", + "content": "-skeleton: Combinatorial curve reconstruction. Graphical models and image processing, 60(2):125-135, 1998. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 322 + ], + "type": "text", + "content": "[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 324, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 356 + ], + "type": "text", + "content": "[5] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 357, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 294, + 390 + ], + "type": "text", + "content": "[6] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 434 + ], + "type": "text", + "content": "[7] Kris Demarsin, Denis Vanderstraeten, Tim Volodine, and Dirk Roose. Detection of closed sharp edges in point clouds using normal estimation and graph theory. Computer-Aided Design, 39(4):276-283, 2007. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 490 + ], + "type": "text", + "content": "[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "text", + "content": "[9] Tamal K Dey, Kurt Mehlhorn, and Edgar A Ramos. Curve reconstruction: Connecting dots with good reason. In Proceedings of the fifteenth annual symposium on Computational geometry, pages 197-206, 1999. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[10] Chen Feng, Yuichi Taguchi, and Vineet R Kamat. Fast plane extraction in organized point clouds using agglomerative hierarchical clustering. In 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 6218-6225. IEEE, 2014. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 634 + ], + "type": "text", + "content": "[11] Yulia Gryaditskaya, Felix Hahnlein, Chenxi Liu, Alla Sheffer, and Adrien Bousseau. Lifting freehand concept sketches into 3d. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 678 + ], + "type": "text", + "content": "[12] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[13] Stefan Gumhold, Xinlong Wang, Rob S MacLeod, et al. Feature extraction from point clouds. In IMR, pages 293-305, 2001. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Yijia He, Ji Zhao, Yue Guo, Wenhao He, and Kui Yuan. Plvio: Tightly-coupled monocular visual-inertial odometry using point and line features. Sensors, 18(4):1159, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[15] Chems-Eddine Himeur, Thibault Lejemble, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "text", + "content": "[16] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Computer Vision and Image Understanding, 157:167-178, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 239 + ], + "type": "text", + "content": "[17] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. arXiv preprint arXiv:2403.17888, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[18] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 398 + ], + "type": "text", + "content": "[20] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[21] G Saravana Kumar, Prem Kumar Kalra, and Sanjay G Dhande. Curve and surface reconstruction from points: an approach based on self-organizing maps. Applied Soft Computing, 5(1):55-66, 2004. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "text", + "content": "[22] In-Kwon Lee. Curve reconstruction from unorganized points. Computer aided geometric design, 17(2):161-177, 2000. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "text", + "content": "[23] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "type": "text", + "content": "[24] Hyunjun Lim, Jinwoo Jeon, and Hyun Myung. Uv-slam: Unconstrained line-based slam using vanishing points for structural mapping. IEEE Robotics and Automation Letters, 7(2):1518-1525, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 569, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 612 + ], + "type": "text", + "content": "[25] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "type": "text", + "content": "[26] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and JanDirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. Cornell University - arXiv, Cornell University - arXiv, 2021. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[27] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5801" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[28] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 239 + ], + "type": "text", + "content": "[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "text", + "content": "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "text", + "content": "[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "type": "text", + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 425 + ], + "type": "text", + "content": "[34] Alec Rivers, Frédo Durand, and Takeo Igarashi. 3d modeling with silhouettes. In Acm siggraph 2010 papers, pages 1-8. 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "text", + "content": "[35] Ryan Schmidt, Azam Khan, Karan Singh, and Gord Kurtenbach. Analytic drawing of 3d scaffolds. In ACM SIGGRAPH Asia 2009 papers, pages 1-10. 2009. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 491 + ], + "type": "text", + "content": "[36] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism. ACM Transactions on Graphics, page 835-846, 2006. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 547 + ], + "type": "text", + "content": "[37] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 602 + ], + "type": "text", + "content": "[38] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 646 + ], + "type": "text", + "content": "[39] Shuangbu Wang, Yu Xia, Lihua You, and Jianjun Zhang. Reconstruction of curve networks from unorganized spatial points. Journal of Universal Computer Science (J. UCS), 26 (9):1265-1280, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "text", + "content": "[40] Wenping Wang, Helmut Pottmann, and Yang Liu. Fitting b-spline curves to point clouds by curvature-based squared distance minimization. ACM Transactions on Graphics (ToG), 25(2):214-238, 2006. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[41] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[42] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "text", + "content": "[43] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 263 + ], + "type": "text", + "content": "[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[45] Tianyi Xie, Zeshun Zong, Yuxin Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. arXiv preprint arXiv:2311.12198, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "text", + "content": "[46] Baoxuan Xu, William Chang, Alla Sheffer, Adrien Bousseau, James McCrae, and Karan Singh. True2form: 3d curve networks from 2d sketches via selective regularization. ACM Transactions on Graphics, 33(4), 2014. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 399 + ], + "type": "text", + "content": "[47] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 554, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 455 + ], + "type": "text", + "content": "[48] Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang, and Philip HS Torr. Holistically-attracted wireframe parsing: From supervised to self-supervised learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 499 + ], + "type": "text", + "content": "[49] Bisheng Yang and Yufu Zang. Automated registration of dense terrestrial laser-scanning point clouds using curves. ISPRS journal of photogrammetry and remote sensing, 95: 109-121, 2014. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 567 + ], + "type": "text", + "content": "[50] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8486-8495, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "text", + "content": "[51] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 656 + ], + "type": "text", + "content": "[52] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[53] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5802" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 182 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[54] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 182 + ], + "type": "text", + "content": "[55] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5803" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_content_list.json b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6712229d02e400a668b2fe3611802bdc3d368db1 --- /dev/null +++ b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 135, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiyue Guo $^{1}$ * Jiarui Hu $^{1}$ * Junjie Hu $^{2}$ Hujun Bao $^{1}$ Guofeng Zhang $^{1\\dagger}$ $^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Chinese University of Hong Kong, Shenzhen", + "bbox": [ + 102, + 178, + 888, + 219 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db473754cb9e8d28756643520cf73d3a40023057c727074de02f405e4a8a105b.jpg", + "image_caption": [ + "Figure 1. SGFormer, which adopts satellite-ground cooperative fusion, can achieve state-of-the-art performance in scene completion and semantic prediction. Benefiting from informative satellite images and a well-designed dual-branch pipeline, SGFormer can effectively improve semantic prediction accuracy and solve the long-standing visual occlusion bottleneck suffered by purely ground-view methods." + ], + "image_footnote": [], + "bbox": [ + 91, + 227, + 320, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/044e60761fc3c4be2eed91b50a9021513e0a7724c98e5fbc2bac7db6bf924523.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 253, + 513, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0d667f6b5b23b8b531e39ba28d4f40f90e247d8013f9da403c3dfe469f78096.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 257, + 712, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/366326938b3df96b851a53ada3c80f7b812df22266bc8abeaa19330a36cd8a46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 253, + 905, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 431, + 326, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, camera-based solutions have been extensively explored for scene semantic completion (SSC). Despite their success in visible areas, existing methods struggle to capture complete scene semantics due to frequent visual occlusions. To address this limitation, this paper presents the first satellite-ground cooperative SSC framework, i.e., SGFormer, exploring the potential of satellite-ground image pairs in the SSC task. Specifically, we propose a dual-branch architecture that encodes orthogonal satellite and ground views in parallel, unifying them into a common domain. Additionally, we design a ground-view guidance strategy that corrects satellite image biases during feature encoding, addressing misalignment between satellite and ground views. Moreover, we develop an adaptive weighting strategy that balances contributions from satellite and ground views. Experiments demonstrate that SGFormer outperforms the state of the art on SemanticKITTI and SSCBench-KITTI-360 datasets. Our code is available on https://github.com/gxytcrc/SGFormer.", + "bbox": [ + 86, + 463, + 485, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 779, + 220, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Urban scene semantic completion (SSC) has been an increasingly prominent problem in 3D computer vision over recent decades, which targets predicting 3D semantic and geometric occupancy of immediately observed surround-", + "bbox": [ + 89, + 804, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ings with a variety of downstream applications such as autonomous driving, robot navigation, and augmented reality (AR). Lidar-based methods have achieved remarkable progress and decent performance [26, 31, 38], while the underlying point cloud representation inherently suffers from weak semantic context derived only from geometry shapes. In contrast, cost-effective camera-based methods can lift rich 2D cues into the 3D world, demonstrating potential in scene reconstruction and understanding.", + "bbox": [ + 511, + 433, + 906, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing camera-based methods have shown promising performance on the SSC task [9, 13, 18, 35, 40]. Nevertheless, even with depth information for visible areas, they still face the challenge of non-unique correspondences between 3D volumes and 2D pixels. Specifically, multiple 3D volumes correspond to significantly overlapping regions within the 2D image plane, which causes semantic ambiguity and radial artifacts in the final reconstruction. For occluded areas, these methods generally lack a long-range global perspective, which makes them struggle to restore a complete scene and serve the following planning and decision steps.", + "bbox": [ + 511, + 575, + 908, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, for the first time, we propose to incorporate satellite imagery into the SSC task. Alongside the development of remote sensing technology, satellite imagery has emerged as a low-cost and widely available reference information. A satellite image, covering major traffic flows in a city, typically requires a light-weight storage footprint, which makes it a compact and memory-efficient representation. Meanwhile, the bird's-eye view (BEV) observation is highly suitable for the horizontal layout of urban scenes. It provides a broad perspective of surrounding obvious ob", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. Email: zhangguofeng@zju.edu.cn", + "bbox": [ + 112, + 875, + 424, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 112, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11929", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jects to effectively enhance semantic certainty. These two orthogonal views, the satellite and ground view, can optimally compensate for each other's blind fields and provide an ideal solution to the previous occlusion bottleneck.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there are two difficulties in incorporating satellite imagery into the SSC task. The first is the misalignment issue. A local satellite image is captured as a fixed-size 2D segment centered around a specific location. An unpredictable deviation from noisy localization and top-down occlusions in the satellite image are the root causes of this issue. Misaligned satellite features tend to bring feature-level inconsistency in fusion and have a negative impact on convergence efficiency. Second, satellite images are typically pre-captured and mainly focus on the scene layout, while inevitably discarding details and long-term changes, such as traffic signs and temporarily parked vehicles, and degrading SSC performance in such low-occupancy but important regions.", + "bbox": [ + 89, + 152, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above challenges, we propose SGFormer, a tightly coupled satellite-assisted framework for the SSC task. In order to perform satellite-ground cross-view fusion, SGFormer innovatively proposes a tailored dual-branch framework to synchronously encode ground and satellite images, aligning them within a unified feature domain. To overcome the misalignment issue, we design a feature-level satellite correction strategy regarding vertically squeezed ground-view features. Specifically, we initialize the learnable BEV parameters and iteratively query squeezed ground-view features in the deformable self-attention layer, where ground-view guidance plays a crucial role in warming up BEV queries in advance to enable coordinated fusion. In addition, we propose an adaptive fusion module with a dual-path weight generator to achieve a reasonable trade-off between two orthogonal views. Within this module, the concatenated satellite-ground features are decoded into channel-wise weight vectors for both views in each voxel. This module allows our fusion pipeline to handle temporal updates and objects of various sizes.", + "bbox": [ + 89, + 364, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method on two benchmark semantic datasets, including SemanticKITTI [1] and SSCBenchKITTI-360 [17]. Extensive experimental results demonstrate that our SGFormer outperforms previous approaches, yielding superior performance in terms of scene completion and semantic prediction, as shown in Figure 1. Overall, our contributions can be summarized as follows:", + "bbox": [ + 89, + 669, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first tightly coupled satellite-assisted SSC framework with a dual-branch design tailored for satellite observations.", + "- We propose a feature-level satellite correction strategy based on deformable self-attention to address the misalignment between satellite and ground views.", + "- We develop an adaptive weighting strategy to balance contributions from satellite and ground view, enabling" + ], + "bbox": [ + 109, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "effective perception of dynamic updates and objects of different sizes.", + "bbox": [ + 544, + 90, + 906, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Experiments on various datasets illustrate that our method can achieve better scene semantic completion results compared to baselines.", + "bbox": [ + 532, + 121, + 906, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 191, + 653, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 513, + 219, + 790, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Semantic Scene Completion (SSC) was first proposed by SSCNet [31], aiming to predict the occupancy and semantic information of each voxel in 3D spaces [5, 26, 37, 38]. Subsequent methods can be categorized into two types: one relies on depth sensors, such as LiDAR, to directly compute spatial features for estimating 3D semantics, while the other is camera-based and involves lifting image features to the 3D spaces before estimating semantic information. Earlier approaches mainly focused on the first type, while in recent years, camera-based solutions have gained increasing attention due to their high cost-efficiency [3, 13, 18, 35].", + "bbox": [ + 511, + 244, + 903, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MonoScene [3] presents the first pure vision solution, leveraging 2D and 3D U-Nets bridged by a sight projection module. OccFormer [41] follows the strategy of LSS [25], estimates the depth distribution of the image, and then projects the features into 3D spaces through depth guidance.", + "bbox": [ + 511, + 412, + 903, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some works are inspired by BEVFormer [19], which utilizes the Transformer framework to estimate occupancy. These methods use spatial relationships from 3D to 2D to query the information from image features via spatial deformable attention [43]. Among them, TPVFormer [13] proposes a Tri-perspective view representation. It first acquires the features of these planes and recovers them into 3D accordingly. SurrunodOcc [35] introduces a coarse-to-fine strategy, generating 3D features at multiple scales and progressively integrating occupancy grid predictions. These transformer-based methods show significant progress in the performance of 3D semantic prediction. However, due to the non-one-to-one nature of the 3D-2D projection relationship and the lack of geometric constraints, these methods struggle to reconstruct the semantic distribution of 3D scenes accurately.", + "bbox": [ + 511, + 489, + 905, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problem, some works leverage depth input and try to incorporate geometric priors into occupancy predictions. OccDepth [23] projects the image features by stereo depth. Voxformer [18] initializes the sparse proposal from pre-trained depth. Subsequent studies [15, 33, 34] have further improved upon this by integrating in-depth information, thereby enhancing the performance of SSC. However, due to inherent issues arising from limitations of ground camera observation range and occlusions make predictive semantic ambiguities and geometric distribution errors almost inevitable.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11930", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b68b3ba5d8943e25ebd867582e8c5cb0cdd8bdf9560a5367e96774c1982ebfd7.jpg", + "image_caption": [ + "Figure 2. Overview of SGFormer. Overall, SGFormer feeds a satellite-ground image pair into similar backbone networks in different branches to extract multi-level feature maps respectively (Left Part). Then, leveraging deformable attention, it transforms satellite and ground features into volume and BEV spaces (Middle Part) for following feature fusion and decoding (Right Part). Specifically, in the ground branch, we use a depth estimator to produce voxel proposals for targeted querying on non-empty feature volumes. In the satellite branch, we fuse vertically squeezed ground-view features into BEV queries to warm up satellite features (Red Line). Before final fusion, encoded features from both branches are enhanced through 2D/3D convolution networks. Our proposed fusion module, which is detailed in Figure 3, is able to adaptively fuse satellite and ground features, followed by a seg head to output semantic reconstruction results." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 906, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Satellite Assist Perception", + "text_level": 1, + "bbox": [ + 89, + 391, + 325, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, integrating satellite images with ground images has gained increasing attention. This is mainly due to the low economic and storage costs of satellite images, as well as the wealth of information they contain.", + "bbox": [ + 89, + 422, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most of these works mainly focus on the localization problem, aiming at estimating the position of ground vehicles in real-world coordinates by matching ground and satellite images. Some methods divide satellite maps into many small patches, aiming to find the patch that is most similar to the ground image through image retrieval methods [2, 4, 12, 21]. Other approaches try integrating satellite and ground features into the common coordinate system to achieve higher localization accuracy, such as applying homography to project satellite features into the ground perspective and then determining precise localization results [30] or projecting ground images into a bird's-eye view (BEV) and aligning them with satellite features [6, 8, 27]. On this basis, some works have attempted to incorporate the results of such cross-view localization into traditional SLAM systems to enhance SLAM localization performance [8, 24, 42].", + "bbox": [ + 89, + 491, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to localization efforts, some works have incorporated satellite images into map construction tasks. SG-BEV [39] is the first to combine ground and satellite features to accomplish fine-grained building attribute segmentation tasks, while SNAP [28] included satellite images in the construction of 2D neural maps.", + "bbox": [ + 89, + 756, + 483, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Against this backdrop, our approach is the first to introduce satellite images into the SSC task, exploring the potential of cooperative satellite and ground perception.", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 390, + 604, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 416, + 624, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework of our proposed SGFormer is illustrated in Figure 2. Our work takes both ground and satellite images as inputs, processed through two distinct branches: the ground branch and the satellite branch. Each branch consists of two-step operations, where the first is the feature extraction to generate multi-scale features from ground and satellite images and the second is the feature transformation to covert the features into voxel or BEV spaces, followed by feature diffusion and enhancement. Then, the voxel and BEV features are fused through a fusion module.", + "bbox": [ + 511, + 440, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Within SGFormer, we make following main technical contributions: 1) Satellite-view feature correction: the satellite correction method employs compressed features from the ground branch to guide and correct the feature learning of the satellite branch (detailed in Sec. 3.3.2). 2) Adaptive fusion: The adaptive fusion module merges BEV features from the satellite branch with voxel features from the ground branch through the attention mechanism. After fusing features from both branches, we further refine them and pass the output into a segmentation head, which up-samples these features and produces voxel-wise class predictions (detailed in Sec. 3.4).", + "bbox": [ + 511, + 592, + 908, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Extraction.", + "text_level": 1, + "bbox": [ + 513, + 785, + 700, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature extraction of the ground branch aims to extract 2D features $\\mathbf{F}_g^{2D} \\in \\mathbb{R}^{H_g' \\times W_g' \\times D}$ from the ground image, where $H_g' \\times W_g'$ represents the feature resolution, and $D$ is the feature dimension. Similarly, in the satellite branch, we obtain 2D satellite features $\\mathbf{F}_s^{2D} \\in \\mathbb{R}^{H_s' \\times W_s' \\times D}$ . EfficientNet-B7 [32] serves as the backbone for ground images, while", + "bbox": [ + 511, + 808, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11931", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/937e85317ecaf6ea4e7a2ad6ca34ddfe6661dda284f80279dfdc444a781b4779.jpg", + "image_caption": [ + "Figure 3. Fusion Module. Our fusion module (Top) mainly consists of two parts: the weight module and the probability network. This weight module (Bottom Left) can facilitate mutual perception between the BEV feature $\\mathbf{F}_s^{\\prime BEV}$ and voxel feature $\\mathbf{F}_g^{\\prime 3D}$ in both 2D and 3D spaces to produce channel-wise weight vectors. This probability network (Bottom Right) takes the weighted average 3D voxel feature as input to infer the approximate occupancy probability per voxel to achieve better geometry reconstruction." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 480, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ResNet-50 [10] is used for satellite images, with both backbones followed by a Feature Pyramid Network (FPN) [20]. Depth Estimation. Before encoding the features of ground view to 3D spaces, we estimate the visible voxels using depth estimation, employing the pre-trained MobileStereoNet [29] for depth estimation, aligning to VoxFormer [18]. However, instead of using an additional stage to refine the binary occupancy like VoxFormer [18], we directly employ the voxelized depth as our query proposals for feature encoding. Specifically, depth points are projected into the voxel map, with a voxel set to 1 if the number of occupied points exceeds a threshold; otherwise, it is set to 0.", + "bbox": [ + 89, + 433, + 483, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature Transformation", + "text_level": 1, + "bbox": [ + 89, + 625, + 312, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To efficiently transform the image features to real-world coordinates, we use deformable attention [43], represented as DA, to aggregate query features $q$ from target image features $v$ , as described by the equation:", + "bbox": [ + 89, + 647, + 483, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D A} (q, p, v) = \\sum_ {k = 1} ^ {K} A _ {k} W v (p + \\Delta p), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 718, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ is the reference point, $\\Delta p$ is the learnable sampling offset, $A_{k}$ represents the attention weight, and $W$ denotes the projection weight.", + "bbox": [ + 89, + 768, + 482, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Ground-view Feature Transformation", + "text_level": 1, + "bbox": [ + 89, + 830, + 408, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the ground branch, we initialize the voxel queries $\\mathbf{Q}_v\\in$ $\\mathbb{R}^{H\\times W\\times Z\\times D}$ with learnable embeddings, where $H\\times W\\times$ $Z$ is regarded as the voxel dimension. Guided by the query", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proposals, we query the voxel features from image features $\\mathbf{F}_g^{2D}$ by employing the deformable cross-attention:", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {p} = \\mathrm {D A} \\left(\\mathbf {Q} _ {p}, \\mathcal {P} (\\mathbf {p}, g), \\mathbf {F} _ {g} ^ {2 D}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 133, + 906, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\mathbf{Q}_p$ consists of the visible voxels selected by the query proposals. $\\mathcal{P}(\\mathbf{p},g)$ denotes the corresponding pixels in $F_{q}^{2D}$ generated by camera projection function.", + "bbox": [ + 511, + 164, + 905, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After several layers of cross-attention, we merge the query features $Q_{p}$ with other invisible voxels in $Q_{v}$ to get the voxel features $\\mathbf{F}_{g}^{3D}$ , and then diffuse the features to all voxels through deformable self-attention, as formulated as:", + "bbox": [ + 511, + 209, + 906, + 268 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {g} ^ {3 D} = \\mathrm {D A} \\left(\\mathbf {F} _ {g} ^ {3 D}, \\mathbf {p _ {v}}, \\mathbf {F} _ {g} ^ {3 D}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 279, + 906, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_v$ is the reference point in 3D space.", + "bbox": [ + 511, + 310, + 807, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Satellite-view Feature Transformation", + "text_level": 1, + "bbox": [ + 511, + 342, + 833, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the satellite branch, we encode the satellite features to BEV space under the guidance of ground observations. We first initialize the BEV queries $\\mathbf{Q}_{bev} \\in \\mathbb{R}^{H \\times W \\times D}$ . Simultaneously, we compress the voxel features $\\mathbf{F}_g^{3D}$ from the ground branch into BEV form $\\mathbf{F}_g^{BEV}$ through max pooling. We then combine $\\mathbf{Q}_{bev}$ and $\\mathbf{F}_g^{BEV}$ to form a hybrid feature $v_{hybrid}$ , which is fed into a self-attention module:", + "bbox": [ + 511, + 364, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {b e v} = \\mathrm {D A} \\left(\\mathbf {Q} _ {b e v}, \\mathbf {p} _ {b e v}, v _ {\\text {h y b r i d}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 487, + 906, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{\\mathrm{bev}}$ is the reference point in BEV space. Through self-attention, the BEV queries are efficiently fused from ground-view information. The output queries are then passed into the cross-attention module to query the information from the satellite features $\\mathbf{F}_s^{2D}$ :", + "bbox": [ + 511, + 515, + 905, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {b e v} = \\mathrm {D A} \\left(\\mathbf {Q} _ {b e v}, \\mathcal {P} (\\mathbf {p}, s), \\mathbf {F} _ {s} ^ {2 D}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 601, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{P}(\\mathbf{p}, s)$ represents the corresponding pixels of BEV grids in the satellite image. With the ground-view observation, the offset layers predict more suitable offsets during cross-attention. We iterate the self and cross attention several times, resulting in the BEV features $\\mathbf{F}_s^{BEV}$ .", + "bbox": [ + 511, + 630, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3 Convolutional Enhancement", + "text_level": 1, + "bbox": [ + 511, + 723, + 767, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After encoding the features into voxel and BEV spaces, we feed them into convolutional layers to further enhance feature representations through neighborhood interactions. We use 3D-UNet for feature aggregation in the ground branch and 2D-UNet in the satellite branch.", + "bbox": [ + 511, + 747, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Feature Fusion", + "text_level": 1, + "bbox": [ + 511, + 832, + 666, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our adaptive fusion module is shown in Figure 3. The BEV features $\\mathbf{F}_s^{\\prime BEV}$ of the satellite branch can cover a wider range, allowing for a more complete representation of scene", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11932", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layouts such as roads and buildings. However, BEV features lack occupancy geometry (empty or occupied) and do not perform well on small objects. On the other hand, voxel features $\\mathbf{F}_{g}^{r3D}$ based on ground images have significant advantages in handling small, dynamic objects and detailed elements, as well as spatial occupancy, but the effective information they contain is limited to the unoccluded field of view. Therefore, it is important to keep the advantages of each view while minimizing the impact of their limitations.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the challenges, we propose a fusion module that dynamically predicts weights for both the feature channel and spatial domains. Specifically, we process the features from two branches: a channel-attention network [11] and a spatial-attention network to balance the contributions at the object level and perception filed level.", + "bbox": [ + 89, + 226, + 480, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the channel-attention path, we first lift the BEV features $\\mathbf{F}_{s}^{\\prime BEV}$ to 3D volumes, denoted as $\\mathbf{F}_{s}^{\\prime 3D}$ . Then, we concatenate the features from the ground and satellite branches in 3D volume to form the combined features and pass them into the channel-attention network, resulting in channel-domain weights $\\mathbf{W}_c\\in \\mathbb{R}^{D\\times 1\\times 1\\times 1}$ .", + "bbox": [ + 89, + 316, + 480, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the spatial-attention path, we compress the voxel features $\\mathbf{F}_{g}^{\\prime 3D}$ along the $z$ -axis into BEV space, resulting in $\\mathbf{F}_{g}^{\\prime BEV}$ . Similarly to the channel-attention path, we concatenate the BEV features and pass them into the spatial-attention network, generating spatial-domain weights $\\mathbf{W}_s \\in \\mathbb{R}^{1 \\times H \\times W}$ in BEV space. The two attention weights are summed together and then combined with the result from an additional MLP to obtain the fused weight $\\mathbf{W}_a \\in \\mathbb{R}^{D \\times H \\times W \\times Z}$ , expressed as follows:", + "bbox": [ + 89, + 407, + 480, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {a} = \\operatorname {M L P} \\left(\\mathbf {F} _ {c} ^ {\\prime 3 D}\\right) \\oplus \\mathrm {C} \\left(\\mathbf {F} _ {c} ^ {\\prime 3 D}\\right) \\oplus \\mathrm {S} \\left(\\mathbf {F} _ {c} ^ {\\prime B E V}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 555, + 482, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where C and S represent the channel and spatial attention networks, respectively. $\\mathbf{F}_{c}^{\\prime 3D}$ and $\\mathbf{F}_{c}^{\\prime BEV}$ are the concatenated features in 3D and BEV spaces. MLP refers to the MLP operation applied to $\\mathbf{F}_{c}^{\\prime 3D}$ . After obtaining the attention weight, we fuse the two branches' features by equation:", + "bbox": [ + 89, + 584, + 480, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {f} ^ {3 D} = \\mathbf {W} _ {a} \\cdot \\mathbf {F} _ {g} ^ {\\prime 3 D} + (1 - \\mathbf {W} _ {a}) \\cdot \\mathbf {F} _ {s} ^ {\\prime 3 D}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 669, + 480, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_f^{3D}$ is the fused voxel features.", + "bbox": [ + 89, + 700, + 346, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, to minimize the negative impact of empty voxels, we get inspiration from [16] and [36], applying a probability network along with spatial attention to identify valuable voxels. Those valuable voxels will get a higher weight through the network to enhance learning efficiency.", + "bbox": [ + 89, + 717, + 480, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Refinement. We further refine the fused features before we output the final voxel prediction. Instead of using dense operations to refine all grids, we only focus on the grids with high uncertainty. Therefore, we first project the features $\\mathbf{F}_f^{3D}$ into semantic classes $\\mathbf{L}_{\\text{coarse}} \\in \\mathbb{R}^{N \\times H \\times W \\times Z}$ , where $N$ is the classes number. We then compute the entropy of each grid and select top-k grids with", + "bbox": [ + 89, + 792, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the highest entropy scores. These high-uncertainty voxels resample features from ground features $\\mathbf{F}_g^{2D}$ through deformable cross-attention. The refined features are then upsampled and output through our semantic head.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Loss", + "text_level": 1, + "bbox": [ + 511, + 159, + 656, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [3] and [18], we train our model with weighted cross-entropy loss $\\mathcal{L}_{ce}$ , as well as scene class affinity loss $\\mathcal{L}_{scal}^{geo}$ and $\\mathcal{L}_{scal}^{sem}$ . Moreover, in order to enhance the supervision, we additionally apply the BEV loss $\\mathcal{L}_{bev}$ from the satellite branch and coarse loss $\\mathcal{L}_{co}$ from the uncertainty refinement module. Specifically, $\\mathcal{L}_{bev}$ is the cross-entropy between the BEV semantic estimation $\\mathbf{L}_{BEV} \\in \\mathbb{R}^{N \\times H \\times W}$ from the satellite branch and the squeezed ground truth, while $\\mathcal{L}_{co}$ is the cross-entropy between $\\mathbf{L}_{coarse}$ and downsampled ground truth. These two losses are weighted by scale factor $\\lambda_{bev}$ and $\\lambda_{co}$ , respectively. The total loss function can be expressed as follows:", + "bbox": [ + 511, + 181, + 903, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} + \\mathcal {L} _ {c e} + \\lambda_ {b e v} \\mathcal {L} _ {b e v} + \\lambda_ {c o} \\mathcal {L} _ {c o}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 369, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the scale factor $L_{bev}$ and $\\lambda_{co}$ are set to 1 and 0.25. In addition, we employ the class weight refer to [18].", + "bbox": [ + 511, + 393, + 903, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 436, + 643, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fairly evaluate the performance of SGFormer, we conduct experiments on the SemanticKITTI [1] and SSCBench-KITTI-360 [17] datasets. In Sec 4.3, we compare our method against existing approaches on both datasets. We mark the top-3 results of camera-based methods in red, green, and blue. In Sec 4.4, we present ablation studies to demonstrate the effectiveness of each module and selection. Finally, in Sec 4.5, we show our visualization results.", + "bbox": [ + 511, + 462, + 903, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 606, + 614, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SemanticKITTI dataset contains 22 sequences. Among them, 11 sequences are used for training, 1 sequence is for validation, and 10 sequences are for testing. It provides ground images with the shape of $1226 \\times 370$ . Meanwhile, corresponding satellite image data are provided by [30]. Each satellite image has the size of $512 \\times 512$ , with a scale factor of 0.2 meters per pixel. In this paper, due to the lack of GPS information, we only use 10 sequences to train the model (without sequence 03) and evaluate the performance on the validation set.", + "bbox": [ + 511, + 628, + 903, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SSCBench-KITTI-360 dataset contains 9 sequences, 7 sequences are used for training, 1 sequence for validation, and 1 for test. The input size of the ground image is $1480 \\times 376$ . Since the dataset does not include satellite data, we obtained corresponding satellite images from Google Maps [7] using the ground-truth poses provided. The satellite image settings are consistent with those used in SemanticKITTI.", + "bbox": [ + 511, + 779, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11933", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5eaf0b2e10ffe8fa2742fe2ce7b3306a934109fb976216a4ebe4ef0cbe365198.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIoUmIoUroad* (15.30%)sidewalk* (11.13%)parking* (11.2%)other-grnd.* (0.56%)building* (0.41%)car (0.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation* (0.93%)trunk (0.51%)terrain* (0.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (0.90%)pole (0.29%)traff-sign (0.08%)
LiDAR-based methods
LMSCNet [26]28.616.7040.6818.224.380.0010.3118.330.000.000.000.0013.660.0220.540.000.001.210.000.00
JS3C-Net [38]38.9810.3150.4923.7411.940.0715.0324.654.410.000.006.1518.114.3326.860.670.270.203.943.771.45
Camera-based methods
MonoScene [3]36.8611.0856.5226.7214.270.4614.0923.266.980.610.451.4817.892.8129.641.861.200.005.844.142.25
TPVFormer [13]35.6111.3656.5025.8720.600.8513.8823.818.080.360.054.3516.922.2630.380.510.890.005.943.141.52
VoxFormer [18]44.0512.3054.4825.9516.210.6817.6125.825.480.610.503.8624.224.8829.241.883.130.007.887.014.20
OccFormer [41]36.5013.4658.8526.8819.610.3114.4025.0925.530.811.198.5219.633.9332.622.782.820.005.614.262.86
SurroundOcc [35]37.2412.7059.2028.2421.421.6714.9426.2614.751.672.377.7319.093.5131.043.602.740.006.654.532.73
Symphonies [15]41.9214.8956.3527.5815.280.9521.6428.6820.442.542.8213.8925.726.6030.873.522.240.008.409.575.76
Ours45.0116.6860.8531.5319.322.1426.0532.1720.302.953.1011.5527.118.2838.473.661.490.009.2811.587.22
", + "bbox": [ + 94, + 88, + 906, + 329 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4b66156924c013f2d5d94eefe4a09c6d078ea5b4721d53b417f7888ee5a8b81d.jpg", + "table_caption": [ + "Table 1. Quantitative results on SemanticKITTI val set. $\\star$ denotes the scene layout structures." + ], + "table_footnote": [], + "table_body": "
MethodIoUmIoUcar(2.85%)bicycle(0.01%)motorcycle(0.00%)truck(0.16%)other-veh.(5.75%)person(0.02%)road*(14.98%)parking*(2.31%)sidewalk*(6.43%)other-grmd.(2.05%)building*(15.67%)fence(0.96%)vegetation*(41.99%)pole(0.22%)traff.-sign(0.06%)other-struct.(4.33%)other-obj.(0.28%)
LiDAR-based methods
SSCNet [31]53.5816.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [26]47.3513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8712.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [13]40.2213.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [18]38.7611.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [41]40.2713.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies [15]43.4117.8226.864.214.9014.207.766.5757.3013.5835.244.5739.207.9534.2319.1914.0416.788.236.04
GaussianFormer [14]35.3812.9218.931.024.6218.077.593.3545.4710.8925.035.3228.445.6829.548.622.993.329.515.14
Ours46.3518.3027.800.912.5510.735.674.2861.0413.2137.005.0743.057.4638.9824.8715.7516.908.855.33
", + "bbox": [ + 98, + 369, + 898, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on SSCBench-KITTI-360 test set. $\\star$ denotes the scene layout structures.", + "bbox": [ + 191, + 641, + 800, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both SemanticKITTI and SSCBench-KITTI-360 datasets provide the voxelized point clouds with labels as ground truth, measuring the 3D volume with a range of $51.2m \\times 51.2m \\times 6.4m$ . The dimension of voxel grids is $256 \\times 256 \\times 32$ , making each grid cell a 0.2-meter cube. In our evaluation, we report the intersection over union (IoU) for geometry performance and mean IoU (mIoU) metrics for semantic performance, aligning with mainstream works.", + "bbox": [ + 89, + 671, + 482, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 801, + 307, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our SGFormer for 25 epochs on 3 NVIDIA 3090 GPUs, with a batch size of 3. We employ AdamW optimizer [22] with an initial learning rate of 4e-4, with a weight decay of 0.01. We employ the cosine learning rate strategy to reduce the learning rate during the training. Furthermore,", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the feature dimension is set to 128, and each batch consumes around 20GB of GPU memory.", + "bbox": [ + 511, + 671, + 903, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 710, + 651, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with state-of-the-art on both semanticKITTI and SSCBench-KITTI-360 datasets. The baseline methods include camera-based methods, as well as LiDAR-based methods. Since SSC performance is sensitive to depth quality, we re-evaluated the depth-based method Voxformer [18] and Symphonize [15] using our generated depth data (the depth estimation method is the same). The results for other methods are obtained from [15, 17]. Table 1 and Table 2 show the quantitative results. It is demonstrated that our method achieves superior performance on both geometry (IoU) and semantics predictions (mIoU). Regarding", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11934", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c87e375af3fbf81e45a4f39ac6995eba9fe1c32f4b44633c6d5bede716cb16e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sat.-branchSat.-corr.FusionIoUmIoUGlobalDetailParams (M)Memory (M)
XXX44.5314.8026.138.2193.4915424
✓XX44.0015.0126.548.26126.4318263
✓✓X43.9015.5928.438.10126.5318967
✓X✓44.7415.8127.509.01126.8919123
✓✓✓45.0116.6829.319.29126.9919865
", + "bbox": [ + 192, + 88, + 803, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/57bd5918e7a86d34f581019c80ab73517d81475d5a1cd1c76271ebc2b6e5ec5f.jpg", + "table_caption": [ + "Table 3. Ablation on main components of SGFormer." + ], + "table_footnote": [], + "table_body": "
IoUmIoUGlobalDetail
w/o Sat.-corr.
w/o noise44.7415.8127.509.01
±5m44.8315.1026.088.70
with Sat.-corr.
w/o noise45.0116.6829.319.29
±5m45.2415.9628.758.53
", + "bbox": [ + 150, + 212, + 426, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation on localization noise.", + "bbox": [ + 168, + 344, + 403, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "semantic prediction, our method achieved $16.68\\mathrm{mIoU}$ on the SemanticKITTI dataset and $18.30\\mathrm{mIoU}$ on SSCBenchKITTI-360, significantly outperforming all other methods. For occupancy prediction, our method achieved 45.01 IoU on the SemanticKITTI dataset, surpassing all previous approaches, including VoxFormer, which uses additional steps to estimate occupancy, and all LiDAR-based methods. On the SSCBench-KITTI-360 dataset, our method achieved 46.55 IoU, outperforming all other camera-based methods and achieving results that are also quite comparable to LiDAR-based methods. For a detailed analysis of specific labels, our method performs significantly better than other camera-based approaches in predicting scene layouts. On the SSCBench-KITTI-360 dataset, SGFormer is the only camera-based method to achieve performance comparable to LiDAR-based methods in scene layout estimation. This improvement is due to the additional satellite input, which provides a more complete view of these objects through aerial observations. Moreover, when evaluating the performance of small or dynamic objects, we find that our method also achieves excellent performance. The adaptive fusion module effectively mitigates the negative effects of satellite observations on those small objects. A more detailed analysis of this part is provided in the ablation study section 4.4.", + "bbox": [ + 89, + 369, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 741, + 243, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct the ablations on the SemanticKITTI dataset, mainly analyzing the effectiveness of our core designs and the impact of localization noise. In addition to the IoU and mIoU, we also report the mean prediction accuracy (mIoU) of different semantic categories for further analysis. Specifically, we set those scene layout structures as the \"global\" category, while other dynamic, small objects are regarded as the \"detail\" category, reporting the performance of detail estimation.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Effectiveness on Core Designs", + "text_level": 1, + "bbox": [ + 511, + 215, + 774, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 presents the ablation on three core designs: the satellite branch (sat.-branch. in the table), satellite correction strategy (sat.-corr. in the table), and adaptive fusion module (fusion in the table). A single ground-branch approach is set as the baseline.", + "bbox": [ + 511, + 238, + 903, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Satellite branch. As shown, adding a new branch with satellite inputs can improve the overall performance. However, the improvement is slight in both categories, and it leads to a degradation in IoU performance. These observations indicate two key points. First, adding satellite observations does enhance the prediction of scene distribution, but misalignment issues such as satellite image localization errors and vertical occlusions can limit this improvement. Second, the lack of observations on dynamic objects, small objects, and occupancy information in satellite images leads to negative effects on predictions for these \"detailed\" objects, as well as the geometry prediction.", + "bbox": [ + 511, + 314, + 905, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Satellite correction strategy. As seen in the table, employing our satellite correction strategy can significantly improve the prediction accuracy of scene layout structures in the \"global\" category, increasing from 26.54 to 28.43. This improvement demonstrates that our satellite correction strategy can address the misalignment issues. The groundview observation assists the satellite branch in completing the scene layout.", + "bbox": [ + 511, + 496, + 905, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive fusion module. The ablation also indicates that using our adaptive fusion module can significantly enhance the accuracy for small and dynamic objects. As shown in the table, with the fusion module, the mIoU of \"detail\" category is increased from 8.26 to 9.01. Furthermore, adding the fusion module will also lead to an improvement of IoU.", + "bbox": [ + 511, + 617, + 905, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is noted that combining all these components yields a substantial performance improvement, as shown in the last row of the table.", + "bbox": [ + 511, + 707, + 905, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Impact of Localization Noise", + "text_level": 1, + "bbox": [ + 511, + 771, + 767, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the experiment to explore the impact of the localization noise, as shown in table 4. We add a random noise of $\\pm 5$ meters along both latitude and longitude to align the possible localization errors in the real world. Overall, our method is sensitive to localization noise: adding a 5-meter noise results in a decrease of $4.3\\%$ and $4.5\\%$ mIoU for SG-Former with and without the satellite correction strategy,", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11935", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42300d31d1c827ae3e3e64808482aca1925a0cdb59928f4cdff74b91fa8e5c58.jpg", + "image_caption": [ + "Figure 4. Scene Semantic Completion on SemanticKITTI [1]. Left: we show satellite-ground image pairs and indicate the vehicle's travel direction as a yellow arrow. Right: we qualitatively compare scene semantic completion results from SGFormer and other baselines, where SGFormer can produce more complete and accurate semantic reconstruction relying on the satellite-ground fusion." + ], + "image_footnote": [], + "bbox": [ + 116, + 103, + 851, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "respectively. Our satellite correction does indeed reduce the negative impact caused by localization noise, which is particularly evident in the scene layout performance. Without satellite correction, the mIoU for the \"global\" category dropped to 26.08, even lower than the baseline value in table 3, indicating that satellite observations almost no longer provide a positive effect. However, when using our correction strategy, the performance decline in scene layout prediction is significantly mitigated.", + "bbox": [ + 89, + 569, + 482, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 89, + 723, + 318, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 shows the qualitative comparisons between our SGFormer and two baseline methods, VoxFormer [18] and Symphonies[15] on SemanticKITTI. As shown, the results predicted by our method are significantly better than the other two methods. Overall, our output shows a more accurate scene layout, particularly in complex scenes such as cross-road. The other two methods struggle to reconstruct these complex scenes completely, whereas our method handles this situation effectively. Considering the detail part, VoxFormer produces the scene with radial artifacts, while", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Symphonies performs somewhat better in this regard but still has many cluttered areas due to occlusions. Our reconstructed results not only avoid these issues but also produce clearer boundaries between different semantics. These comparisons demonstrate the superiority of our method both global and local scale.", + "bbox": [ + 511, + 569, + 906, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SGFormer, the first satellite-ground cooperative approach with a tightly coupled dual-branch framework design for urban semantic scene completion. SGFormer effectively fuses observations from groundview and satellite-view images, enabling joint consideration of global scene layout and local details for occupancy grid prediction. Experimental results show that our method surpasses state-of-the-art camera-based approaches and achieves performance on par with LiDAR-based methods across many metrics. These findings suggest that integrating satellite images into the SSC task offers a cost-effective yet highly promising solution. We hope our work inspires further research in this community.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11936", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 243, + 187, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 8", + "[2] Sudong Cai, Yulan Guo, Salman Khan, Jiwei Hu, and Gongjian Wen. Ground-to-aerial image geo-localization with a hard exemplar reweighting triplet loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8391-8400, 2019. 3", + "[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 5, 6", + "[4] Yue Cao, Mingsheng Long, Jianmin Wang, and Shichen Liu. Deep visual-semantic quantization for efficient image retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1328–1337, 2017. 3", + "[5] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In Conference on Robot Learning, pages 2148-2161. PMLR, 2021. 2", + "[6] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 3", + "[7] Google. Google maps api, 2024. Accessed: 2024-11-14. 5", + "[8] Xiyue Guo, Haocheng Peng, Junjie Hu, Hujun Bao, and Guofeng Zhang. From satellite to ground: Satellite assisted visual localization with cross-view semantic matching. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 3977-3983. IEEE, 2024. 3", + "[9] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Hongdong Li. Daniel Cremers. S4c: Self-supervised semantic scene completion with neural fields. In 2024 International Conference on 3D Vision (3DV), pages 409-420. IEEE, 2024. 1", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + ], + "bbox": [ + 93, + 268, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was partially supported by NSF of China (No. 62425209). The author Junjie Hu acknowledges support from the Guangdong Natural Science Fund under Grant 2024A1515010252. The authors thank Shang Liu for her help in drawing Figures 1, 2, and 3. Special thanks also go to Bo Wen, Xingyuan Yu, Han Zhang, and Ran Zuo for their invaluable support.", + "[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 5", + "[12] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 3", + "[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 9223-9232, 2023. 1, 2, 6", + "[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2405.17429, 2024. 6", + "[15] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20258-20267, 2024. 2, 6, 8", + "[16] Peidong Li, Wancheng Shen, Qihao Huang, and Dixiao Cui. Dualbev: Cnn is all you need in view transformation. arXiv preprint arXiv:2403.05402, 2024. 5", + "[17] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6", + "[18] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 4, 5, 6, 8", + "[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 2", + "[20] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 4", + "[21] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5624-5633, 2019. 3", + "[22] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[23] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 2" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11937", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Ian D Miller, Anthony Cowley, Ravi Konkimalla, Shreyas S Shivakumar, Ty Nguyen, Trey Smith, Camillo Jose Taylor, and Vijay Kumar. Any way you look at it: Semantic crossview localization and mapping with lidar. IEEE Robotics and Automation Letters, 6(2):2397-2404, 2021. 3", + "[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2", + "[26] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 1, 2, 6", + "[27] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 3", + "[28] Paul-Edouard Sarlin, Eduard Trulls, Marc Pollefeys, Jan Hosang, and Simon Lynen. Snap: Self-supervised neural maps for visual positioning and semantic understanding. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[29] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In Proceedings of the IEEE/cvf Winter Conference on Applications of Computer Vision, pages 2417-2426, 2022. 4", + "[30] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17010-17020, 2022. 3, 5", + "[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1746-1754, 2017. 1, 2, 6", + "[32] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 3", + "[33] Xin Tan, Wenbin Wu, Zhiwei Zhang, Chaojie Fan, Yong Peng, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision. arXiv preprint arXiv:2405.10591, 2024. 2", + "[34] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2", + "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occu" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 6", + "[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3–19, 2018. 5", + "[37] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 2", + "[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 1, 2, 6", + "[39] Junyan Ye, Qiyan Luo, Jinhua Yu, Huaping Zhong, Zhimeng Zheng, Conghui He, and Weijia Li. Sg-bev: Satellite-guided bev fusion for cross-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27748-27757, 2024. 3", + "[40] Zhu Yu, Runmin Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Si-Yuan Cao, and Hui-liang Shen. Context and geometry aware voxel transformer for semantic scene completion. In Advances in Neural Information Processing Systems, pages 1531-1555, 2024. 1", + "[41] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 2, 6", + "[42] Yanhao Zhang, Yujiao Shi, Shan Wang, Ankit Vora, Akhil Perincherry, Yongbo Chen, and Hongdong Li. Increasing slam pose accuracy by ground-to-satellite image registration. arXiv preprint arXiv:2404.09169, 2024. 3", + "[43] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 4" + ], + "bbox": [ + 516, + 92, + 906, + 655 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11938", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_model.json b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd81cc1b4183ac34c74f123f1f4c2bfcb2fbbbf9 --- /dev/null +++ b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.179, + 0.89, + 0.22 + ], + "angle": 0, + "content": "Xiyue Guo\\(^{1}\\) * Jiarui Hu\\(^{1}\\) * Junjie Hu\\(^{2}\\) Hujun Bao\\(^{1}\\) Guofeng Zhang\\(^{1\\dagger}\\) \n\\(^{1}\\)State Key Lab of CAD&CG, Zhejiang University \\(^{2}\\)Chinese University of Hong Kong, Shenzhen" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.228, + 0.321, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.255, + 0.514, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.258, + 0.714, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.255, + 0.906, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.388, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 1. SGFormer, which adopts satellite-ground cooperative fusion, can achieve state-of-the-art performance in scene completion and semantic prediction. Benefiting from informative satellite images and a well-designed dual-branch pipeline, SGFormer can effectively improve semantic prediction accuracy and solve the long-standing visual occlusion bottleneck suffered by purely ground-view methods." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.433, + 0.327, + 0.448 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.464, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Recently, camera-based solutions have been extensively explored for scene semantic completion (SSC). Despite their success in visible areas, existing methods struggle to capture complete scene semantics due to frequent visual occlusions. To address this limitation, this paper presents the first satellite-ground cooperative SSC framework, i.e., SGFormer, exploring the potential of satellite-ground image pairs in the SSC task. Specifically, we propose a dual-branch architecture that encodes orthogonal satellite and ground views in parallel, unifying them into a common domain. Additionally, we design a ground-view guidance strategy that corrects satellite image biases during feature encoding, addressing misalignment between satellite and ground views. Moreover, we develop an adaptive weighting strategy that balances contributions from satellite and ground views. Experiments demonstrate that SGFormer outperforms the state of the art on SemanticKITTI and SSCBench-KITTI-360 datasets. Our code is available on https://github.com/gxytcrc/SGFormer." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.78, + 0.222, + 0.795 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Urban scene semantic completion (SSC) has been an increasingly prominent problem in 3D computer vision over recent decades, which targets predicting 3D semantic and geometric occupancy of immediately observed surround-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.571 + ], + "angle": 0, + "content": "ings with a variety of downstream applications such as autonomous driving, robot navigation, and augmented reality (AR). Lidar-based methods have achieved remarkable progress and decent performance [26, 31, 38], while the underlying point cloud representation inherently suffers from weak semantic context derived only from geometry shapes. In contrast, cost-effective camera-based methods can lift rich 2D cues into the 3D world, demonstrating potential in scene reconstruction and understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.744 + ], + "angle": 0, + "content": "Existing camera-based methods have shown promising performance on the SSC task [9, 13, 18, 35, 40]. Nevertheless, even with depth information for visible areas, they still face the challenge of non-unique correspondences between 3D volumes and 2D pixels. Specifically, multiple 3D volumes correspond to significantly overlapping regions within the 2D image plane, which causes semantic ambiguity and radial artifacts in the final reconstruction. For occluded areas, these methods generally lack a long-range global perspective, which makes them struggle to restore a complete scene and serve the following planning and decision steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, for the first time, we propose to incorporate satellite imagery into the SSC task. Alongside the development of remote sensing technology, satellite imagery has emerged as a low-cost and widely available reference information. A satellite image, covering major traffic flows in a city, typically requires a light-weight storage footprint, which makes it a compact and memory-efficient representation. Meanwhile, the bird's-eye view (BEV) observation is highly suitable for the horizontal layout of urban scenes. It provides a broad perspective of surrounding obvious ob" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.425, + 0.888 + ], + "angle": 0, + "content": "†Corresponding author. Email: zhangguofeng@zju.edu.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.425, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11929" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "jects to effectively enhance semantic certainty. These two orthogonal views, the satellite and ground view, can optimally compensate for each other's blind fields and provide an ideal solution to the previous occlusion bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.365 + ], + "angle": 0, + "content": "However, there are two difficulties in incorporating satellite imagery into the SSC task. The first is the misalignment issue. A local satellite image is captured as a fixed-size 2D segment centered around a specific location. An unpredictable deviation from noisy localization and top-down occlusions in the satellite image are the root causes of this issue. Misaligned satellite features tend to bring feature-level inconsistency in fusion and have a negative impact on convergence efficiency. Second, satellite images are typically pre-captured and mainly focus on the scene layout, while inevitably discarding details and long-term changes, such as traffic signs and temporarily parked vehicles, and degrading SSC performance in such low-occupancy but important regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.668 + ], + "angle": 0, + "content": "To solve the above challenges, we propose SGFormer, a tightly coupled satellite-assisted framework for the SSC task. In order to perform satellite-ground cross-view fusion, SGFormer innovatively proposes a tailored dual-branch framework to synchronously encode ground and satellite images, aligning them within a unified feature domain. To overcome the misalignment issue, we design a feature-level satellite correction strategy regarding vertically squeezed ground-view features. Specifically, we initialize the learnable BEV parameters and iteratively query squeezed ground-view features in the deformable self-attention layer, where ground-view guidance plays a crucial role in warming up BEV queries in advance to enable coordinated fusion. In addition, we propose an adaptive fusion module with a dual-path weight generator to achieve a reasonable trade-off between two orthogonal views. Within this module, the concatenated satellite-ground features are decoded into channel-wise weight vectors for both views in each voxel. This module allows our fusion pipeline to handle temporal updates and objects of various sizes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.775 + ], + "angle": 0, + "content": "We evaluate our method on two benchmark semantic datasets, including SemanticKITTI [1] and SSCBenchKITTI-360 [17]. Extensive experimental results demonstrate that our SGFormer outperforms previous approaches, yielding superior performance in terms of scene completion and semantic prediction, as shown in Figure 1. Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- We present the first tightly coupled satellite-assisted SSC framework with a dual-branch design tailored for satellite observations." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- We propose a feature-level satellite correction strategy based on deformable self-attention to address the misalignment between satellite and ground views." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We develop an adaptive weighting strategy to balance contributions from satellite and ground view, enabling" + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "effective perception of dynamic updates and objects of different sizes." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.122, + 0.907, + 0.168 + ], + "angle": 0, + "content": "- Experiments on various datasets illustrate that our method can achieve better scene semantic completion results compared to baselines." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.193, + 0.655, + 0.209 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.22, + 0.791, + 0.236 + ], + "angle": 0, + "content": "2.1.3D Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.412 + ], + "angle": 0, + "content": "3D Semantic Scene Completion (SSC) was first proposed by SSCNet [31], aiming to predict the occupancy and semantic information of each voxel in 3D spaces [5, 26, 37, 38]. Subsequent methods can be categorized into two types: one relies on depth sensors, such as LiDAR, to directly compute spatial features for estimating 3D semantics, while the other is camera-based and involves lifting image features to the 3D spaces before estimating semantic information. Earlier approaches mainly focused on the first type, while in recent years, camera-based solutions have gained increasing attention due to their high cost-efficiency [3, 13, 18, 35]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.413, + 0.905, + 0.49 + ], + "angle": 0, + "content": "MonoScene [3] presents the first pure vision solution, leveraging 2D and 3D U-Nets bridged by a sight projection module. OccFormer [41] follows the strategy of LSS [25], estimates the depth distribution of the image, and then projects the features into 3D spaces through depth guidance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Some works are inspired by BEVFormer [19], which utilizes the Transformer framework to estimate occupancy. These methods use spatial relationships from 3D to 2D to query the information from image features via spatial deformable attention [43]. Among them, TPVFormer [13] proposes a Tri-perspective view representation. It first acquires the features of these planes and recovers them into 3D accordingly. SurrunodOcc [35] introduces a coarse-to-fine strategy, generating 3D features at multiple scales and progressively integrating occupancy grid predictions. These transformer-based methods show significant progress in the performance of 3D semantic prediction. However, due to the non-one-to-one nature of the 3D-2D projection relationship and the lack of geometric constraints, these methods struggle to reconstruct the semantic distribution of 3D scenes accurately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To address the problem, some works leverage depth input and try to incorporate geometric priors into occupancy predictions. OccDepth [23] projects the image features by stereo depth. Voxformer [18] initializes the sparse proposal from pre-trained depth. Subsequent studies [15, 33, 34] have further improved upon this by integrating in-depth information, thereby enhancing the performance of SSC. However, due to inherent issues arising from limitations of ground camera observation range and occlusions make predictive semantic ambiguities and geometric distribution errors almost inevitable." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11930" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.907, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.283, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 2. Overview of SGFormer. Overall, SGFormer feeds a satellite-ground image pair into similar backbone networks in different branches to extract multi-level feature maps respectively (Left Part). Then, leveraging deformable attention, it transforms satellite and ground features into volume and BEV spaces (Middle Part) for following feature fusion and decoding (Right Part). Specifically, in the ground branch, we use a depth estimator to produce voxel proposals for targeted querying on non-empty feature volumes. In the satellite branch, we fuse vertically squeezed ground-view features into BEV queries to warm up satellite features (Red Line). Before final fusion, encoded features from both branches are enhanced through 2D/3D convolution networks. Our proposed fusion module, which is detailed in Figure 3, is able to adaptively fuse satellite and ground features, followed by a seg head to output semantic reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.392, + 0.326, + 0.409 + ], + "angle": 0, + "content": "2.2. Satellite Assist Perception" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Recently, integrating satellite images with ground images has gained increasing attention. This is mainly due to the low economic and storage costs of satellite images, as well as the wealth of information they contain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Most of these works mainly focus on the localization problem, aiming at estimating the position of ground vehicles in real-world coordinates by matching ground and satellite images. Some methods divide satellite maps into many small patches, aiming to find the patch that is most similar to the ground image through image retrieval methods [2, 4, 12, 21]. Other approaches try integrating satellite and ground features into the common coordinate system to achieve higher localization accuracy, such as applying homography to project satellite features into the ground perspective and then determining precise localization results [30] or projecting ground images into a bird's-eye view (BEV) and aligning them with satellite features [6, 8, 27]. On this basis, some works have attempted to incorporate the results of such cross-view localization into traditional SLAM systems to enhance SLAM localization performance [8, 24, 42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.484, + 0.849 + ], + "angle": 0, + "content": "In addition to localization efforts, some works have incorporated satellite images into map construction tasks. SG-BEV [39] is the first to combine ground and satellite features to accomplish fine-grained building attribute segmentation tasks, while SNAP [28] included satellite images in the construction of 2D neural maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Against this backdrop, our approach is the first to introduce satellite images into the SSC task, exploring the potential of cooperative satellite and ground perception." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.391, + 0.606, + 0.407 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.417, + 0.625, + 0.432 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.592 + ], + "angle": 0, + "content": "The framework of our proposed SGFormer is illustrated in Figure 2. Our work takes both ground and satellite images as inputs, processed through two distinct branches: the ground branch and the satellite branch. Each branch consists of two-step operations, where the first is the feature extraction to generate multi-scale features from ground and satellite images and the second is the feature transformation to covert the features into voxel or BEV spaces, followed by feature diffusion and enhancement. Then, the voxel and BEV features are fused through a fusion module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.775 + ], + "angle": 0, + "content": "Within SGFormer, we make following main technical contributions: 1) Satellite-view feature correction: the satellite correction method employs compressed features from the ground branch to guide and correct the feature learning of the satellite branch (detailed in Sec. 3.3.2). 2) Adaptive fusion: The adaptive fusion module merges BEV features from the satellite branch with voxel features from the ground branch through the attention mechanism. After fusing features from both branches, we further refine them and pass the output into a segmentation head, which up-samples these features and produces voxel-wise class predictions (detailed in Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.786, + 0.702, + 0.8 + ], + "angle": 0, + "content": "3.2. Feature Extraction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.809, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Feature extraction of the ground branch aims to extract 2D features \\(\\mathbf{F}_g^{2D} \\in \\mathbb{R}^{H_g' \\times W_g' \\times D}\\) from the ground image, where \\(H_g' \\times W_g'\\) represents the feature resolution, and \\(D\\) is the feature dimension. Similarly, in the satellite branch, we obtain 2D satellite features \\(\\mathbf{F}_s^{2D} \\in \\mathbb{R}^{H_s' \\times W_s' \\times D}\\). EfficientNet-B7 [32] serves as the backbone for ground images, while" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11931" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.482, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Figure 3. Fusion Module. Our fusion module (Top) mainly consists of two parts: the weight module and the probability network. This weight module (Bottom Left) can facilitate mutual perception between the BEV feature \\(\\mathbf{F}_s^{\\prime BEV}\\) and voxel feature \\(\\mathbf{F}_g^{\\prime 3D}\\) in both 2D and 3D spaces to produce channel-wise weight vectors. This probability network (Bottom Right) takes the weighted average 3D voxel feature as input to infer the approximate occupancy probability per voxel to achieve better geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.616 + ], + "angle": 0, + "content": "ResNet-50 [10] is used for satellite images, with both backbones followed by a Feature Pyramid Network (FPN) [20]. Depth Estimation. Before encoding the features of ground view to 3D spaces, we estimate the visible voxels using depth estimation, employing the pre-trained MobileStereoNet [29] for depth estimation, aligning to VoxFormer [18]. However, instead of using an additional stage to refine the binary occupancy like VoxFormer [18], we directly employ the voxelized depth as our query proposals for feature encoding. Specifically, depth points are projected into the voxel map, with a voxel set to 1 if the number of occupied points exceeds a threshold; otherwise, it is set to 0." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.313, + 0.64 + ], + "angle": 0, + "content": "3.3. Feature Transformation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.484, + 0.709 + ], + "angle": 0, + "content": "To efficiently transform the image features to real-world coordinates, we use deformable attention [43], represented as DA, to aggregate query features \\( q \\) from target image features \\( v \\), as described by the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D A} (q, p, v) = \\sum_ {k = 1} ^ {K} A _ {k} W v (p + \\Delta p), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.816 + ], + "angle": 0, + "content": "where \\( p \\) is the reference point, \\( \\Delta p \\) is the learnable sampling offset, \\( A_{k} \\) represents the attention weight, and \\( W \\) denotes the projection weight." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.41, + 0.847 + ], + "angle": 0, + "content": "3.3.1 Ground-view Feature Transformation" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In the ground branch, we initialize the voxel queries \\(\\mathbf{Q}_v\\in\\) \\(\\mathbb{R}^{H\\times W\\times Z\\times D}\\) with learnable embeddings, where \\(H\\times W\\times\\) \\(Z\\) is regarded as the voxel dimension. Guided by the query" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "proposals, we query the voxel features from image features \\(\\mathbf{F}_g^{2D}\\) by employing the deformable cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.135, + 0.907, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {p} = \\mathrm {D A} \\left(\\mathbf {Q} _ {p}, \\mathcal {P} (\\mathbf {p}, g), \\mathbf {F} _ {g} ^ {2 D}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.165, + 0.906, + 0.21 + ], + "angle": 0, + "content": "where the \\(\\mathbf{Q}_p\\) consists of the visible voxels selected by the query proposals. \\(\\mathcal{P}(\\mathbf{p},g)\\) denotes the corresponding pixels in \\(F_{q}^{2D}\\) generated by camera projection function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.907, + 0.27 + ], + "angle": 0, + "content": "After several layers of cross-attention, we merge the query features \\( Q_{p} \\) with other invisible voxels in \\( Q_{v} \\) to get the voxel features \\( \\mathbf{F}_{g}^{3D} \\), and then diffuse the features to all voxels through deformable self-attention, as formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.28, + 0.907, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {g} ^ {3 D} = \\mathrm {D A} \\left(\\mathbf {F} _ {g} ^ {3 D}, \\mathbf {p _ {v}}, \\mathbf {F} _ {g} ^ {3 D}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.311, + 0.808, + 0.326 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_v\\) is the reference point in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.343, + 0.834, + 0.358 + ], + "angle": 0, + "content": "3.3.2 Satellite-view Feature Transformation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.476 + ], + "angle": 0, + "content": "In the satellite branch, we encode the satellite features to BEV space under the guidance of ground observations. We first initialize the BEV queries \\(\\mathbf{Q}_{bev} \\in \\mathbb{R}^{H \\times W \\times D}\\). Simultaneously, we compress the voxel features \\(\\mathbf{F}_g^{3D}\\) from the ground branch into BEV form \\(\\mathbf{F}_g^{BEV}\\) through max pooling. We then combine \\(\\mathbf{Q}_{bev}\\) and \\(\\mathbf{F}_g^{BEV}\\) to form a hybrid feature \\(v_{hybrid}\\), which is fed into a self-attention module:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.488, + 0.907, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {b e v} = \\mathrm {D A} \\left(\\mathbf {Q} _ {b e v}, \\mathbf {p} _ {b e v}, v _ {\\text {h y b r i d}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.593 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{\\mathrm{bev}}\\) is the reference point in BEV space. Through self-attention, the BEV queries are efficiently fused from ground-view information. The output queries are then passed into the cross-attention module to query the information from the satellite features \\(\\mathbf{F}_s^{2D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.602, + 0.906, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {b e v} = \\mathrm {D A} \\left(\\mathbf {Q} _ {b e v}, \\mathcal {P} (\\mathbf {p}, s), \\mathbf {F} _ {s} ^ {2 D}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.906, + 0.707 + ], + "angle": 0, + "content": "\\(\\mathcal{P}(\\mathbf{p}, s)\\) represents the corresponding pixels of BEV grids in the satellite image. With the ground-view observation, the offset layers predict more suitable offsets during cross-attention. We iterate the self and cross attention several times, resulting in the BEV features \\(\\mathbf{F}_s^{BEV}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.768, + 0.739 + ], + "angle": 0, + "content": "3.3.3 Convolutional Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.822 + ], + "angle": 0, + "content": "After encoding the features into voxel and BEV spaces, we feed them into convolutional layers to further enhance feature representations through neighborhood interactions. We use 3D-UNet for feature aggregation in the ground branch and 2D-UNet in the satellite branch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.667, + 0.847 + ], + "angle": 0, + "content": "3.4. Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our adaptive fusion module is shown in Figure 3. The BEV features \\(\\mathbf{F}_s^{\\prime BEV}\\) of the satellite branch can cover a wider range, allowing for a more complete representation of scene" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "Layouts such as roads and buildings. However, BEV features lack occupancy geometry (empty or occupied) and do not perform well on small objects. On the other hand, voxel features \\(\\mathbf{F}_{g}^{r3D}\\) based on ground images have significant advantages in handling small, dynamic objects and detailed elements, as well as spatial occupancy, but the effective information they contain is limited to the unoccluded field of view. Therefore, it is important to keep the advantages of each view while minimizing the impact of their limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.482, + 0.318 + ], + "angle": 0, + "content": "To address the challenges, we propose a fusion module that dynamically predicts weights for both the feature channel and spatial domains. Specifically, we process the features from two branches: a channel-attention network [11] and a spatial-attention network to balance the contributions at the object level and perception filed level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.482, + 0.408 + ], + "angle": 0, + "content": "In the channel-attention path, we first lift the BEV features \\(\\mathbf{F}_{s}^{\\prime BEV}\\) to 3D volumes, denoted as \\(\\mathbf{F}_{s}^{\\prime 3D}\\). Then, we concatenate the features from the ground and satellite branches in 3D volume to form the combined features and pass them into the channel-attention network, resulting in channel-domain weights \\(\\mathbf{W}_c\\in \\mathbb{R}^{D\\times 1\\times 1\\times 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.482, + 0.548 + ], + "angle": 0, + "content": "In the spatial-attention path, we compress the voxel features \\(\\mathbf{F}_{g}^{\\prime 3D}\\) along the \\(z\\)-axis into BEV space, resulting in \\(\\mathbf{F}_{g}^{\\prime BEV}\\). Similarly to the channel-attention path, we concatenate the BEV features and pass them into the spatial-attention network, generating spatial-domain weights \\(\\mathbf{W}_s \\in \\mathbb{R}^{1 \\times H \\times W}\\) in BEV space. The two attention weights are summed together and then combined with the result from an additional MLP to obtain the fused weight \\(\\mathbf{W}_a \\in \\mathbb{R}^{D \\times H \\times W \\times Z}\\), expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.556, + 0.483, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {a} = \\operatorname {M L P} \\left(\\mathbf {F} _ {c} ^ {\\prime 3 D}\\right) \\oplus \\mathrm {C} \\left(\\mathbf {F} _ {c} ^ {\\prime 3 D}\\right) \\oplus \\mathrm {S} \\left(\\mathbf {F} _ {c} ^ {\\prime B E V}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.482, + 0.66 + ], + "angle": 0, + "content": "where C and S represent the channel and spatial attention networks, respectively. \\(\\mathbf{F}_{c}^{\\prime 3D}\\) and \\(\\mathbf{F}_{c}^{\\prime BEV}\\) are the concatenated features in 3D and BEV spaces. MLP refers to the MLP operation applied to \\(\\mathbf{F}_{c}^{\\prime 3D}\\). After obtaining the attention weight, we fuse the two branches' features by equation:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.67, + 0.482, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {f} ^ {3 D} = \\mathbf {W} _ {a} \\cdot \\mathbf {F} _ {g} ^ {\\prime 3 D} + (1 - \\mathbf {W} _ {a}) \\cdot \\mathbf {F} _ {s} ^ {\\prime 3 D}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.702, + 0.347, + 0.717 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_f^{3D}\\) is the fused voxel features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.482, + 0.793 + ], + "angle": 0, + "content": "Additionally, to minimize the negative impact of empty voxels, we get inspiration from [16] and [36], applying a probability network along with spatial attention to identify valuable voxels. Those valuable voxels will get a higher weight through the network to enhance learning efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Feature Refinement. We further refine the fused features before we output the final voxel prediction. Instead of using dense operations to refine all grids, we only focus on the grids with high uncertainty. Therefore, we first project the features \\(\\mathbf{F}_f^{3D}\\) into semantic classes \\(\\mathbf{L}_{\\text{coarse}} \\in \\mathbb{R}^{N \\times H \\times W \\times Z}\\), where \\(N\\) is the classes number. We then compute the entropy of each grid and select top-k grids with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "the highest entropy scores. These high-uncertainty voxels resample features from ground features \\(\\mathbf{F}_g^{2D}\\) through deformable cross-attention. The refined features are then upsampled and output through our semantic head." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.16, + 0.657, + 0.175 + ], + "angle": 0, + "content": "3.5. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Following [3] and [18], we train our model with weighted cross-entropy loss \\(\\mathcal{L}_{ce}\\), as well as scene class affinity loss \\(\\mathcal{L}_{scal}^{geo}\\) and \\(\\mathcal{L}_{scal}^{sem}\\). Moreover, in order to enhance the supervision, we additionally apply the BEV loss \\(\\mathcal{L}_{bev}\\) from the satellite branch and coarse loss \\(\\mathcal{L}_{co}\\) from the uncertainty refinement module. Specifically, \\(\\mathcal{L}_{bev}\\) is the cross-entropy between the BEV semantic estimation \\(\\mathbf{L}_{BEV} \\in \\mathbb{R}^{N \\times H \\times W}\\) from the satellite branch and the squeezed ground truth, while \\(\\mathcal{L}_{co}\\) is the cross-entropy between \\(\\mathbf{L}_{coarse}\\) and downsampled ground truth. These two losses are weighted by scale factor \\(\\lambda_{bev}\\) and \\(\\lambda_{co}\\), respectively. The total loss function can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.371, + 0.905, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} + \\mathcal {L} _ {c e} + \\lambda_ {b e v} \\mathcal {L} _ {b e v} + \\lambda_ {c o} \\mathcal {L} _ {c o}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.426 + ], + "angle": 0, + "content": "where the scale factor \\( L_{bev} \\) and \\( \\lambda_{co} \\) are set to 1 and 0.25. In addition, we employ the class weight refer to [18]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.645, + 0.455 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.905, + 0.597 + ], + "angle": 0, + "content": "To fairly evaluate the performance of SGFormer, we conduct experiments on the SemanticKITTI [1] and SSCBench-KITTI-360 [17] datasets. In Sec 4.3, we compare our method against existing approaches on both datasets. We mark the top-3 results of camera-based methods in red, green, and blue. In Sec 4.4, we present ablation studies to demonstrate the effectiveness of each module and selection. Finally, in Sec 4.5, we show our visualization results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.615, + 0.62 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.779 + ], + "angle": 0, + "content": "The SemanticKITTI dataset contains 22 sequences. Among them, 11 sequences are used for training, 1 sequence is for validation, and 10 sequences are for testing. It provides ground images with the shape of \\(1226 \\times 370\\). Meanwhile, corresponding satellite image data are provided by [30]. Each satellite image has the size of \\(512 \\times 512\\), with a scale factor of 0.2 meters per pixel. In this paper, due to the lack of GPS information, we only use 10 sequences to train the model (without sequence 03) and evaluate the performance on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.9 + ], + "angle": 0, + "content": "SSCBench-KITTI-360 dataset contains 9 sequences, 7 sequences are used for training, 1 sequence for validation, and 1 for test. The input size of the ground image is \\(1480 \\times 376\\). Since the dataset does not include satellite data, we obtained corresponding satellite images from Google Maps [7] using the ground-truth poses provided. The satellite image settings are consistent with those used in SemanticKITTI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11933" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.33 + ], + "angle": 0, + "content": "
MethodIoUmIoUroad* (15.30%)sidewalk* (11.13%)parking* (11.2%)other-grnd.* (0.56%)building* (0.41%)car (0.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation* (0.93%)trunk (0.51%)terrain* (0.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (0.90%)pole (0.29%)traff-sign (0.08%)
LiDAR-based methods
LMSCNet [26]28.616.7040.6818.224.380.0010.3118.330.000.000.000.0013.660.0220.540.000.001.210.000.00
JS3C-Net [38]38.9810.3150.4923.7411.940.0715.0324.654.410.000.006.1518.114.3326.860.670.270.203.943.771.45
Camera-based methods
MonoScene [3]36.8611.0856.5226.7214.270.4614.0923.266.980.610.451.4817.892.8129.641.861.200.005.844.142.25
TPVFormer [13]35.6111.3656.5025.8720.600.8513.8823.818.080.360.054.3516.922.2630.380.510.890.005.943.141.52
VoxFormer [18]44.0512.3054.4825.9516.210.6817.6125.825.480.610.503.8624.224.8829.241.883.130.007.887.014.20
OccFormer [41]36.5013.4658.8526.8819.610.3114.4025.0925.530.811.198.5219.633.9332.622.782.820.005.614.262.86
SurroundOcc [35]37.2412.7059.2028.2421.421.6714.9426.2614.751.672.377.7319.093.5131.043.602.740.006.654.532.73
Symphonies [15]41.9214.8956.3527.5815.280.9521.6428.6820.442.542.8213.8925.726.6030.873.522.240.008.409.575.76
Ours45.0116.6860.8531.5319.322.1426.0532.1720.302.953.1011.5527.118.2838.473.661.490.009.2811.587.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.342, + 0.778, + 0.356 + ], + "angle": 0, + "content": "Table 1. Quantitative results on SemanticKITTI val set. \\(\\star\\) denotes the scene layout structures." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.37, + 0.9, + 0.632 + ], + "angle": 0, + "content": "
MethodIoUmIoUcar(2.85%)bicycle(0.01%)motorcycle(0.00%)truck(0.16%)other-veh.(5.75%)person(0.02%)road*(14.98%)parking*(2.31%)sidewalk*(6.43%)other-grmd.(2.05%)building*(15.67%)fence(0.96%)vegetation*(41.99%)pole(0.22%)traff.-sign(0.06%)other-struct.(4.33%)other-obj.(0.28%)
LiDAR-based methods
SSCNet [31]53.5816.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [26]47.3513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8712.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [13]40.2213.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [18]38.7611.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [41]40.2713.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies [15]43.4117.8226.864.214.9014.207.766.5757.3013.5835.244.5739.207.9534.2319.1914.0416.788.236.04
GaussianFormer [14]35.3812.9218.931.024.6218.077.593.3545.4710.8925.035.3228.445.6829.548.622.993.329.515.14
Ours46.3518.3027.800.912.5510.735.674.2861.0413.2137.005.0743.057.4638.9824.8715.7516.908.855.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.192, + 0.642, + 0.802, + 0.655 + ], + "angle": 0, + "content": "Table 2. Quantitative results on SSCBench-KITTI-360 test set. \\(\\star\\) denotes the scene layout structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.792 + ], + "angle": 0, + "content": "Both SemanticKITTI and SSCBench-KITTI-360 datasets provide the voxelized point clouds with labels as ground truth, measuring the 3D volume with a range of \\(51.2m \\times 51.2m \\times 6.4m\\). The dimension of voxel grids is \\(256 \\times 256 \\times 32\\), making each grid cell a 0.2-meter cube. In our evaluation, we report the intersection over union (IoU) for geometry performance and mean IoU (mIoU) metrics for semantic performance, aligning with mainstream works." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.308, + 0.818 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We train our SGFormer for 25 epochs on 3 NVIDIA 3090 GPUs, with a batch size of 3. We employ AdamW optimizer [22] with an initial learning rate of 4e-4, with a weight decay of 0.01. We employ the cosine learning rate strategy to reduce the learning rate during the training. Furthermore," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.905, + 0.703 + ], + "angle": 0, + "content": "the feature dimension is set to 128, and each batch consumes around 20GB of GPU memory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.652, + 0.726 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We compare our method with state-of-the-art on both semanticKITTI and SSCBench-KITTI-360 datasets. The baseline methods include camera-based methods, as well as LiDAR-based methods. Since SSC performance is sensitive to depth quality, we re-evaluated the depth-based method Voxformer [18] and Symphonize [15] using our generated depth data (the depth estimation method is the same). The results for other methods are obtained from [15, 17]. Table 1 and Table 2 show the quantitative results. It is demonstrated that our method achieves superior performance on both geometry (IoU) and semantics predictions (mIoU). Regarding" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11934" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.193, + 0.089, + 0.805, + 0.187 + ], + "angle": 0, + "content": "
Sat.-branchSat.-corr.FusionIoUmIoUGlobalDetailParams (M)Memory (M)
XXX44.5314.8026.138.2193.4915424
✓XX44.0015.0126.548.26126.4318263
✓✓X43.9015.5928.438.10126.5318967
✓X✓44.7415.8127.509.01126.8919123
✓✓✓45.0116.6829.319.29126.9919865
" + }, + { + "type": "table_caption", + "bbox": [ + 0.339, + 0.193, + 0.658, + 0.207 + ], + "angle": 0, + "content": "Table 3. Ablation on main components of SGFormer." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.213, + 0.428, + 0.334 + ], + "angle": 0, + "content": "
IoUmIoUGlobalDetail
w/o Sat.-corr.
w/o noise44.7415.8127.509.01
±5m44.8315.1026.088.70
with Sat.-corr.
w/o noise45.0116.6829.319.29
±5m45.2415.9628.758.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.169, + 0.345, + 0.405, + 0.358 + ], + "angle": 0, + "content": "Table 4. Ablation on localization noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.734 + ], + "angle": 0, + "content": "semantic prediction, our method achieved \\(16.68\\mathrm{mIoU}\\) on the SemanticKITTI dataset and \\(18.30\\mathrm{mIoU}\\) on SSCBenchKITTI-360, significantly outperforming all other methods. For occupancy prediction, our method achieved 45.01 IoU on the SemanticKITTI dataset, surpassing all previous approaches, including VoxFormer, which uses additional steps to estimate occupancy, and all LiDAR-based methods. On the SSCBench-KITTI-360 dataset, our method achieved 46.55 IoU, outperforming all other camera-based methods and achieving results that are also quite comparable to LiDAR-based methods. For a detailed analysis of specific labels, our method performs significantly better than other camera-based approaches in predicting scene layouts. On the SSCBench-KITTI-360 dataset, SGFormer is the only camera-based method to achieve performance comparable to LiDAR-based methods in scene layout estimation. This improvement is due to the additional satellite input, which provides a more complete view of these objects through aerial observations. Moreover, when evaluating the performance of small or dynamic objects, we find that our method also achieves excellent performance. The adaptive fusion module effectively mitigates the negative effects of satellite observations on those small objects. A more detailed analysis of this part is provided in the ablation study section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.245, + 0.759 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct the ablations on the SemanticKITTI dataset, mainly analyzing the effectiveness of our core designs and the impact of localization noise. In addition to the IoU and mIoU, we also report the mean prediction accuracy (mIoU) of different semantic categories for further analysis. Specifically, we set those scene layout structures as the \"global\" category, while other dynamic, small objects are regarded as the \"detail\" category, reporting the performance of detail estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.216, + 0.776, + 0.232 + ], + "angle": 0, + "content": "4.4.1 Effectiveness on Core Designs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Table 3 presents the ablation on three core designs: the satellite branch (sat.-branch. in the table), satellite correction strategy (sat.-corr. in the table), and adaptive fusion module (fusion in the table). A single ground-branch approach is set as the baseline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.906, + 0.496 + ], + "angle": 0, + "content": "Satellite branch. As shown, adding a new branch with satellite inputs can improve the overall performance. However, the improvement is slight in both categories, and it leads to a degradation in IoU performance. These observations indicate two key points. First, adding satellite observations does enhance the prediction of scene distribution, but misalignment issues such as satellite image localization errors and vertical occlusions can limit this improvement. Second, the lack of observations on dynamic objects, small objects, and occupancy information in satellite images leads to negative effects on predictions for these \"detailed\" objects, as well as the geometry prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.617 + ], + "angle": 0, + "content": "Satellite correction strategy. As seen in the table, employing our satellite correction strategy can significantly improve the prediction accuracy of scene layout structures in the \"global\" category, increasing from 26.54 to 28.43. This improvement demonstrates that our satellite correction strategy can address the misalignment issues. The groundview observation assists the satellite branch in completing the scene layout." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.906, + 0.707 + ], + "angle": 0, + "content": "Adaptive fusion module. The ablation also indicates that using our adaptive fusion module can significantly enhance the accuracy for small and dynamic objects. As shown in the table, with the fusion module, the mIoU of \"detail\" category is increased from 8.26 to 9.01. Furthermore, adding the fusion module will also lead to an improvement of IoU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.752 + ], + "angle": 0, + "content": "It is noted that combining all these components yields a substantial performance improvement, as shown in the last row of the table." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.768, + 0.787 + ], + "angle": 0, + "content": "4.4.2 Impact of Localization Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct the experiment to explore the impact of the localization noise, as shown in table 4. We add a random noise of \\(\\pm 5\\) meters along both latitude and longitude to align the possible localization errors in the real world. Overall, our method is sensitive to localization noise: adding a 5-meter noise results in a decrease of \\(4.3\\%\\) and \\(4.5\\%\\) mIoU for SG-Former with and without the satellite correction strategy," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11935" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.104, + 0.852, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.519, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Figure 4. Scene Semantic Completion on SemanticKITTI [1]. Left: we show satellite-ground image pairs and indicate the vehicle's travel direction as a yellow arrow. Right: we qualitatively compare scene semantic completion results from SGFormer and other baselines, where SGFormer can produce more complete and accurate semantic reconstruction relying on the satellite-ground fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.707 + ], + "angle": 0, + "content": "respectively. Our satellite correction does indeed reduce the negative impact caused by localization noise, which is particularly evident in the scene layout performance. Without satellite correction, the mIoU for the \"global\" category dropped to 26.08, even lower than the baseline value in table 3, indicating that satellite observations almost no longer provide a positive effect. However, when using our correction strategy, the performance decline in scene layout prediction is significantly mitigated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.724, + 0.32, + 0.74 + ], + "angle": 0, + "content": "4.5. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Figure 4 shows the qualitative comparisons between our SGFormer and two baseline methods, VoxFormer [18] and Symphonies[15] on SemanticKITTI. As shown, the results predicted by our method are significantly better than the other two methods. Overall, our output shows a more accurate scene layout, particularly in complex scenes such as cross-road. The other two methods struggle to reconstruct these complex scenes completely, whereas our method handles this situation effectively. Considering the detail part, VoxFormer produces the scene with radial artifacts, while" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Symphonies performs somewhat better in this regard but still has many cluttered areas due to occlusions. Our reconstructed results not only avoid these issues but also produce clearer boundaries between different semantics. These comparisons demonstrate the superiority of our method both global and local scale." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce SGFormer, the first satellite-ground cooperative approach with a tightly coupled dual-branch framework design for urban semantic scene completion. SGFormer effectively fuses observations from groundview and satellite-view images, enabling joint consideration of global scene layout and local details for occupancy grid prediction. Experimental results show that our method surpasses state-of-the-art camera-based approaches and achieves performance on par with LiDAR-based methods across many metrics. These findings suggest that integrating satellite images into the SSC task offers a cost-effective yet highly promising solution. We hope our work inspires further research in this community." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11936" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.214 + ], + "angle": 0, + "content": "This work was partially supported by NSF of China (No. 62425209). The author Junjie Hu acknowledges support from the Guangdong Natural Science Fund under Grant 2024A1515010252. The authors thank Shang Liu for her help in drawing Figures 1, 2, and 3. Special thanks also go to Bo Wen, Xingyuan Yu, Han Zhang, and Ran Zuo for their invaluable support." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.244, + 0.188, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.269, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.355, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[2] Sudong Cai, Yulan Guo, Salman Khan, Jiwei Hu, and Gongjian Wen. Ground-to-aerial image geo-localization with a hard exemplar reweighting triplet loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8391-8400, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.426, + 0.485, + 0.482 + ], + "angle": 0, + "content": "[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.485, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[4] Yue Cao, Mingsheng Long, Jianmin Wang, and Shichen Liu. Deep visual-semantic quantization for efficient image retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1328–1337, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.542, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[5] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In Conference on Robot Learning, pages 2148-2161. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.484, + 0.683 + ], + "angle": 0, + "content": "[6] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.686, + 0.476, + 0.7 + ], + "angle": 0, + "content": "[7] Google. Google maps api, 2024. Accessed: 2024-11-14. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[8] Xiyue Guo, Haocheng Peng, Junjie Hu, Hujun Bao, and Guofeng Zhang. From satellite to ground: Satellite assisted visual localization with cross-view semantic matching. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 3977-3983. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[9] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Hongdong Li. Daniel Cremers. S4c: Self-supervised semantic scene completion with neural fields. In 2024 International Conference on 3D Vision (3DV), pages 409-420. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.269, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[12] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 9223-9232, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2405.17429, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[15] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20258-20267, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[16] Peidong Li, Wancheng Shen, Qihao Huang, and Dixiao Cui. Dualbev: Cnn is all you need in view transformation. arXiv preprint arXiv:2403.05402, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[17] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.908, + 0.616 + ], + "angle": 0, + "content": "[18] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[21] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5624-5633, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[22] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11937" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Ian D Miller, Anthony Cowley, Ravi Konkimalla, Shreyas S Shivakumar, Ty Nguyen, Trey Smith, Camillo Jose Taylor, and Vijay Kumar. Any way you look at it: Semantic crossview localization and mapping with lidar. IEEE Robotics and Automation Letters, 6(2):2397-2404, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[26] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[27] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[28] Paul-Edouard Sarlin, Eduard Trulls, Marc Pollefeys, Jan Hosang, and Simon Lynen. Snap: Self-supervised neural maps for visual positioning and semantic understanding. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[29] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In Proceedings of the IEEE/cvf Winter Conference on Applications of Computer Vision, pages 2417-2426, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[30] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17010-17020, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1746-1754, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[32] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[33] Xin Tan, Wenbin Wu, Zhiwei Zhang, Chaojie Fan, Yong Peng, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision. arXiv preprint arXiv:2405.10591, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[34] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occu" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "pancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3–19, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[37] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[39] Junyan Ye, Qiyan Luo, Jinhua Yu, Huaping Zhong, Zhimeng Zheng, Conghui He, and Weijia Li. Sg-bev: Satellite-guided bev fusion for cross-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27748-27757, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[40] Zhu Yu, Runmin Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Si-Yuan Cao, and Hui-liang Shen. Context and geometry aware voxel transformer for semantic scene completion. In Advances in Neural Information Processing Systems, pages 1531-1555, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[41] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[42] Yanhao Zhang, Yujiao Shi, Shan Wang, Ankit Vora, Akhil Perincherry, Yongbo Chen, and Hongdong Li. Increasing slam pose accuracy by ground-to-satellite image registration. arXiv preprint arXiv:2404.09169, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[43] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11938" + } + ] +] \ No newline at end of file diff --git a/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_origin.pdf b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c93648c219c9330c32c59fd0546bc2d1c3bdf170 --- /dev/null +++ b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/fa8f1162-46b6-4b8d-ab34-ea04e929ede5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1783f41b3ff1559df192ea4708835b340392b1a7b510cf8d463be27325c5b1d8 +size 5622949 diff --git a/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/full.md b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2532aa1dd93b76673e77298dcf31c340727d35a0 --- /dev/null +++ b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/full.md @@ -0,0 +1,302 @@ +# SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion + +Xiyue Guo $^{1}$ * Jiarui Hu $^{1}$ * Junjie Hu $^{2}$ Hujun Bao $^{1}$ Guofeng Zhang $^{1\dagger}$ $^{1}$ State Key Lab of CAD&CG, Zhejiang University $^{2}$ Chinese University of Hong Kong, Shenzhen + +![](images/db473754cb9e8d28756643520cf73d3a40023057c727074de02f405e4a8a105b.jpg) +Figure 1. SGFormer, which adopts satellite-ground cooperative fusion, can achieve state-of-the-art performance in scene completion and semantic prediction. Benefiting from informative satellite images and a well-designed dual-branch pipeline, SGFormer can effectively improve semantic prediction accuracy and solve the long-standing visual occlusion bottleneck suffered by purely ground-view methods. + +![](images/044e60761fc3c4be2eed91b50a9021513e0a7724c98e5fbc2bac7db6bf924523.jpg) + +![](images/c0d667f6b5b23b8b531e39ba28d4f40f90e247d8013f9da403c3dfe469f78096.jpg) + +![](images/366326938b3df96b851a53ada3c80f7b812df22266bc8abeaa19330a36cd8a46.jpg) + +# Abstract + +Recently, camera-based solutions have been extensively explored for scene semantic completion (SSC). Despite their success in visible areas, existing methods struggle to capture complete scene semantics due to frequent visual occlusions. To address this limitation, this paper presents the first satellite-ground cooperative SSC framework, i.e., SGFormer, exploring the potential of satellite-ground image pairs in the SSC task. Specifically, we propose a dual-branch architecture that encodes orthogonal satellite and ground views in parallel, unifying them into a common domain. Additionally, we design a ground-view guidance strategy that corrects satellite image biases during feature encoding, addressing misalignment between satellite and ground views. Moreover, we develop an adaptive weighting strategy that balances contributions from satellite and ground views. Experiments demonstrate that SGFormer outperforms the state of the art on SemanticKITTI and SSCBench-KITTI-360 datasets. Our code is available on https://github.com/gxytcrc/SGFormer. + +# 1. Introduction + +Urban scene semantic completion (SSC) has been an increasingly prominent problem in 3D computer vision over recent decades, which targets predicting 3D semantic and geometric occupancy of immediately observed surround- + +ings with a variety of downstream applications such as autonomous driving, robot navigation, and augmented reality (AR). Lidar-based methods have achieved remarkable progress and decent performance [26, 31, 38], while the underlying point cloud representation inherently suffers from weak semantic context derived only from geometry shapes. In contrast, cost-effective camera-based methods can lift rich 2D cues into the 3D world, demonstrating potential in scene reconstruction and understanding. + +Existing camera-based methods have shown promising performance on the SSC task [9, 13, 18, 35, 40]. Nevertheless, even with depth information for visible areas, they still face the challenge of non-unique correspondences between 3D volumes and 2D pixels. Specifically, multiple 3D volumes correspond to significantly overlapping regions within the 2D image plane, which causes semantic ambiguity and radial artifacts in the final reconstruction. For occluded areas, these methods generally lack a long-range global perspective, which makes them struggle to restore a complete scene and serve the following planning and decision steps. + +In this paper, for the first time, we propose to incorporate satellite imagery into the SSC task. Alongside the development of remote sensing technology, satellite imagery has emerged as a low-cost and widely available reference information. A satellite image, covering major traffic flows in a city, typically requires a light-weight storage footprint, which makes it a compact and memory-efficient representation. Meanwhile, the bird's-eye view (BEV) observation is highly suitable for the horizontal layout of urban scenes. It provides a broad perspective of surrounding obvious ob + +jects to effectively enhance semantic certainty. These two orthogonal views, the satellite and ground view, can optimally compensate for each other's blind fields and provide an ideal solution to the previous occlusion bottleneck. + +However, there are two difficulties in incorporating satellite imagery into the SSC task. The first is the misalignment issue. A local satellite image is captured as a fixed-size 2D segment centered around a specific location. An unpredictable deviation from noisy localization and top-down occlusions in the satellite image are the root causes of this issue. Misaligned satellite features tend to bring feature-level inconsistency in fusion and have a negative impact on convergence efficiency. Second, satellite images are typically pre-captured and mainly focus on the scene layout, while inevitably discarding details and long-term changes, such as traffic signs and temporarily parked vehicles, and degrading SSC performance in such low-occupancy but important regions. + +To solve the above challenges, we propose SGFormer, a tightly coupled satellite-assisted framework for the SSC task. In order to perform satellite-ground cross-view fusion, SGFormer innovatively proposes a tailored dual-branch framework to synchronously encode ground and satellite images, aligning them within a unified feature domain. To overcome the misalignment issue, we design a feature-level satellite correction strategy regarding vertically squeezed ground-view features. Specifically, we initialize the learnable BEV parameters and iteratively query squeezed ground-view features in the deformable self-attention layer, where ground-view guidance plays a crucial role in warming up BEV queries in advance to enable coordinated fusion. In addition, we propose an adaptive fusion module with a dual-path weight generator to achieve a reasonable trade-off between two orthogonal views. Within this module, the concatenated satellite-ground features are decoded into channel-wise weight vectors for both views in each voxel. This module allows our fusion pipeline to handle temporal updates and objects of various sizes. + +We evaluate our method on two benchmark semantic datasets, including SemanticKITTI [1] and SSCBenchKITTI-360 [17]. Extensive experimental results demonstrate that our SGFormer outperforms previous approaches, yielding superior performance in terms of scene completion and semantic prediction, as shown in Figure 1. Overall, our contributions can be summarized as follows: + +- We present the first tightly coupled satellite-assisted SSC framework with a dual-branch design tailored for satellite observations. +- We propose a feature-level satellite correction strategy based on deformable self-attention to address the misalignment between satellite and ground views. +- We develop an adaptive weighting strategy to balance contributions from satellite and ground view, enabling + +effective perception of dynamic updates and objects of different sizes. + +- Experiments on various datasets illustrate that our method can achieve better scene semantic completion results compared to baselines. + +# 2. Related Work + +# 2.1.3D Semantic Scene Completion + +3D Semantic Scene Completion (SSC) was first proposed by SSCNet [31], aiming to predict the occupancy and semantic information of each voxel in 3D spaces [5, 26, 37, 38]. Subsequent methods can be categorized into two types: one relies on depth sensors, such as LiDAR, to directly compute spatial features for estimating 3D semantics, while the other is camera-based and involves lifting image features to the 3D spaces before estimating semantic information. Earlier approaches mainly focused on the first type, while in recent years, camera-based solutions have gained increasing attention due to their high cost-efficiency [3, 13, 18, 35]. + +MonoScene [3] presents the first pure vision solution, leveraging 2D and 3D U-Nets bridged by a sight projection module. OccFormer [41] follows the strategy of LSS [25], estimates the depth distribution of the image, and then projects the features into 3D spaces through depth guidance. + +Some works are inspired by BEVFormer [19], which utilizes the Transformer framework to estimate occupancy. These methods use spatial relationships from 3D to 2D to query the information from image features via spatial deformable attention [43]. Among them, TPVFormer [13] proposes a Tri-perspective view representation. It first acquires the features of these planes and recovers them into 3D accordingly. SurrunodOcc [35] introduces a coarse-to-fine strategy, generating 3D features at multiple scales and progressively integrating occupancy grid predictions. These transformer-based methods show significant progress in the performance of 3D semantic prediction. However, due to the non-one-to-one nature of the 3D-2D projection relationship and the lack of geometric constraints, these methods struggle to reconstruct the semantic distribution of 3D scenes accurately. + +To address the problem, some works leverage depth input and try to incorporate geometric priors into occupancy predictions. OccDepth [23] projects the image features by stereo depth. Voxformer [18] initializes the sparse proposal from pre-trained depth. Subsequent studies [15, 33, 34] have further improved upon this by integrating in-depth information, thereby enhancing the performance of SSC. However, due to inherent issues arising from limitations of ground camera observation range and occlusions make predictive semantic ambiguities and geometric distribution errors almost inevitable. + +![](images/b68b3ba5d8943e25ebd867582e8c5cb0cdd8bdf9560a5367e96774c1982ebfd7.jpg) +Figure 2. Overview of SGFormer. Overall, SGFormer feeds a satellite-ground image pair into similar backbone networks in different branches to extract multi-level feature maps respectively (Left Part). Then, leveraging deformable attention, it transforms satellite and ground features into volume and BEV spaces (Middle Part) for following feature fusion and decoding (Right Part). Specifically, in the ground branch, we use a depth estimator to produce voxel proposals for targeted querying on non-empty feature volumes. In the satellite branch, we fuse vertically squeezed ground-view features into BEV queries to warm up satellite features (Red Line). Before final fusion, encoded features from both branches are enhanced through 2D/3D convolution networks. Our proposed fusion module, which is detailed in Figure 3, is able to adaptively fuse satellite and ground features, followed by a seg head to output semantic reconstruction results. + +# 2.2. Satellite Assist Perception + +Recently, integrating satellite images with ground images has gained increasing attention. This is mainly due to the low economic and storage costs of satellite images, as well as the wealth of information they contain. + +Most of these works mainly focus on the localization problem, aiming at estimating the position of ground vehicles in real-world coordinates by matching ground and satellite images. Some methods divide satellite maps into many small patches, aiming to find the patch that is most similar to the ground image through image retrieval methods [2, 4, 12, 21]. Other approaches try integrating satellite and ground features into the common coordinate system to achieve higher localization accuracy, such as applying homography to project satellite features into the ground perspective and then determining precise localization results [30] or projecting ground images into a bird's-eye view (BEV) and aligning them with satellite features [6, 8, 27]. On this basis, some works have attempted to incorporate the results of such cross-view localization into traditional SLAM systems to enhance SLAM localization performance [8, 24, 42]. + +In addition to localization efforts, some works have incorporated satellite images into map construction tasks. SG-BEV [39] is the first to combine ground and satellite features to accomplish fine-grained building attribute segmentation tasks, while SNAP [28] included satellite images in the construction of 2D neural maps. + +Against this backdrop, our approach is the first to introduce satellite images into the SSC task, exploring the potential of cooperative satellite and ground perception. + +# 3. Method + +# 3.1. Overview + +The framework of our proposed SGFormer is illustrated in Figure 2. Our work takes both ground and satellite images as inputs, processed through two distinct branches: the ground branch and the satellite branch. Each branch consists of two-step operations, where the first is the feature extraction to generate multi-scale features from ground and satellite images and the second is the feature transformation to covert the features into voxel or BEV spaces, followed by feature diffusion and enhancement. Then, the voxel and BEV features are fused through a fusion module. + +Within SGFormer, we make following main technical contributions: 1) Satellite-view feature correction: the satellite correction method employs compressed features from the ground branch to guide and correct the feature learning of the satellite branch (detailed in Sec. 3.3.2). 2) Adaptive fusion: The adaptive fusion module merges BEV features from the satellite branch with voxel features from the ground branch through the attention mechanism. After fusing features from both branches, we further refine them and pass the output into a segmentation head, which up-samples these features and produces voxel-wise class predictions (detailed in Sec. 3.4). + +# 3.2. Feature Extraction. + +Feature extraction of the ground branch aims to extract 2D features $\mathbf{F}_g^{2D} \in \mathbb{R}^{H_g' \times W_g' \times D}$ from the ground image, where $H_g' \times W_g'$ represents the feature resolution, and $D$ is the feature dimension. Similarly, in the satellite branch, we obtain 2D satellite features $\mathbf{F}_s^{2D} \in \mathbb{R}^{H_s' \times W_s' \times D}$ . EfficientNet-B7 [32] serves as the backbone for ground images, while + +![](images/937e85317ecaf6ea4e7a2ad6ca34ddfe6661dda284f80279dfdc444a781b4779.jpg) +Figure 3. Fusion Module. Our fusion module (Top) mainly consists of two parts: the weight module and the probability network. This weight module (Bottom Left) can facilitate mutual perception between the BEV feature $\mathbf{F}_s^{\prime BEV}$ and voxel feature $\mathbf{F}_g^{\prime 3D}$ in both 2D and 3D spaces to produce channel-wise weight vectors. This probability network (Bottom Right) takes the weighted average 3D voxel feature as input to infer the approximate occupancy probability per voxel to achieve better geometry reconstruction. + +ResNet-50 [10] is used for satellite images, with both backbones followed by a Feature Pyramid Network (FPN) [20]. Depth Estimation. Before encoding the features of ground view to 3D spaces, we estimate the visible voxels using depth estimation, employing the pre-trained MobileStereoNet [29] for depth estimation, aligning to VoxFormer [18]. However, instead of using an additional stage to refine the binary occupancy like VoxFormer [18], we directly employ the voxelized depth as our query proposals for feature encoding. Specifically, depth points are projected into the voxel map, with a voxel set to 1 if the number of occupied points exceeds a threshold; otherwise, it is set to 0. + +# 3.3. Feature Transformation + +To efficiently transform the image features to real-world coordinates, we use deformable attention [43], represented as DA, to aggregate query features $q$ from target image features $v$ , as described by the equation: + +$$ +\mathrm {D A} (q, p, v) = \sum_ {k = 1} ^ {K} A _ {k} W v (p + \Delta p), \tag {1} +$$ + +where $p$ is the reference point, $\Delta p$ is the learnable sampling offset, $A_{k}$ represents the attention weight, and $W$ denotes the projection weight. + +# 3.3.1 Ground-view Feature Transformation + +In the ground branch, we initialize the voxel queries $\mathbf{Q}_v\in$ $\mathbb{R}^{H\times W\times Z\times D}$ with learnable embeddings, where $H\times W\times$ $Z$ is regarded as the voxel dimension. Guided by the query + +proposals, we query the voxel features from image features $\mathbf{F}_g^{2D}$ by employing the deformable cross-attention: + +$$ +\mathbf {Q} _ {p} = \mathrm {D A} \left(\mathbf {Q} _ {p}, \mathcal {P} (\mathbf {p}, g), \mathbf {F} _ {g} ^ {2 D}\right), \tag {2} +$$ + +where the $\mathbf{Q}_p$ consists of the visible voxels selected by the query proposals. $\mathcal{P}(\mathbf{p},g)$ denotes the corresponding pixels in $F_{q}^{2D}$ generated by camera projection function. + +After several layers of cross-attention, we merge the query features $Q_{p}$ with other invisible voxels in $Q_{v}$ to get the voxel features $\mathbf{F}_{g}^{3D}$ , and then diffuse the features to all voxels through deformable self-attention, as formulated as: + +$$ +\mathbf {F} _ {g} ^ {3 D} = \mathrm {D A} \left(\mathbf {F} _ {g} ^ {3 D}, \mathbf {p _ {v}}, \mathbf {F} _ {g} ^ {3 D}\right), \tag {3} +$$ + +where $\mathbf{p}_v$ is the reference point in 3D space. + +# 3.3.2 Satellite-view Feature Transformation + +In the satellite branch, we encode the satellite features to BEV space under the guidance of ground observations. We first initialize the BEV queries $\mathbf{Q}_{bev} \in \mathbb{R}^{H \times W \times D}$ . Simultaneously, we compress the voxel features $\mathbf{F}_g^{3D}$ from the ground branch into BEV form $\mathbf{F}_g^{BEV}$ through max pooling. We then combine $\mathbf{Q}_{bev}$ and $\mathbf{F}_g^{BEV}$ to form a hybrid feature $v_{hybrid}$ , which is fed into a self-attention module: + +$$ +\mathbf {Q} _ {b e v} = \mathrm {D A} \left(\mathbf {Q} _ {b e v}, \mathbf {p} _ {b e v}, v _ {\text {h y b r i d}}\right), \tag {4} +$$ + +where $\mathbf{p}_{\mathrm{bev}}$ is the reference point in BEV space. Through self-attention, the BEV queries are efficiently fused from ground-view information. The output queries are then passed into the cross-attention module to query the information from the satellite features $\mathbf{F}_s^{2D}$ : + +$$ +\mathbf {Q} _ {b e v} = \mathrm {D A} \left(\mathbf {Q} _ {b e v}, \mathcal {P} (\mathbf {p}, s), \mathbf {F} _ {s} ^ {2 D}\right), \tag {5} +$$ + +$\mathcal{P}(\mathbf{p}, s)$ represents the corresponding pixels of BEV grids in the satellite image. With the ground-view observation, the offset layers predict more suitable offsets during cross-attention. We iterate the self and cross attention several times, resulting in the BEV features $\mathbf{F}_s^{BEV}$ . + +# 3.3.3 Convolutional Enhancement + +After encoding the features into voxel and BEV spaces, we feed them into convolutional layers to further enhance feature representations through neighborhood interactions. We use 3D-UNet for feature aggregation in the ground branch and 2D-UNet in the satellite branch. + +# 3.4. Feature Fusion + +Our adaptive fusion module is shown in Figure 3. The BEV features $\mathbf{F}_s^{\prime BEV}$ of the satellite branch can cover a wider range, allowing for a more complete representation of scene + +Layouts such as roads and buildings. However, BEV features lack occupancy geometry (empty or occupied) and do not perform well on small objects. On the other hand, voxel features $\mathbf{F}_{g}^{r3D}$ based on ground images have significant advantages in handling small, dynamic objects and detailed elements, as well as spatial occupancy, but the effective information they contain is limited to the unoccluded field of view. Therefore, it is important to keep the advantages of each view while minimizing the impact of their limitations. + +To address the challenges, we propose a fusion module that dynamically predicts weights for both the feature channel and spatial domains. Specifically, we process the features from two branches: a channel-attention network [11] and a spatial-attention network to balance the contributions at the object level and perception filed level. + +In the channel-attention path, we first lift the BEV features $\mathbf{F}_{s}^{\prime BEV}$ to 3D volumes, denoted as $\mathbf{F}_{s}^{\prime 3D}$ . Then, we concatenate the features from the ground and satellite branches in 3D volume to form the combined features and pass them into the channel-attention network, resulting in channel-domain weights $\mathbf{W}_c\in \mathbb{R}^{D\times 1\times 1\times 1}$ . + +In the spatial-attention path, we compress the voxel features $\mathbf{F}_{g}^{\prime 3D}$ along the $z$ -axis into BEV space, resulting in $\mathbf{F}_{g}^{\prime BEV}$ . Similarly to the channel-attention path, we concatenate the BEV features and pass them into the spatial-attention network, generating spatial-domain weights $\mathbf{W}_s \in \mathbb{R}^{1 \times H \times W}$ in BEV space. The two attention weights are summed together and then combined with the result from an additional MLP to obtain the fused weight $\mathbf{W}_a \in \mathbb{R}^{D \times H \times W \times Z}$ , expressed as follows: + +$$ +\mathbf {W} _ {a} = \operatorname {M L P} \left(\mathbf {F} _ {c} ^ {\prime 3 D}\right) \oplus \mathrm {C} \left(\mathbf {F} _ {c} ^ {\prime 3 D}\right) \oplus \mathrm {S} \left(\mathbf {F} _ {c} ^ {\prime B E V}\right), \tag {6} +$$ + +where C and S represent the channel and spatial attention networks, respectively. $\mathbf{F}_{c}^{\prime 3D}$ and $\mathbf{F}_{c}^{\prime BEV}$ are the concatenated features in 3D and BEV spaces. MLP refers to the MLP operation applied to $\mathbf{F}_{c}^{\prime 3D}$ . After obtaining the attention weight, we fuse the two branches' features by equation: + +$$ +\mathbf {F} _ {f} ^ {3 D} = \mathbf {W} _ {a} \cdot \mathbf {F} _ {g} ^ {\prime 3 D} + (1 - \mathbf {W} _ {a}) \cdot \mathbf {F} _ {s} ^ {\prime 3 D}, \tag {7} +$$ + +where $\mathbf{F}_f^{3D}$ is the fused voxel features. + +Additionally, to minimize the negative impact of empty voxels, we get inspiration from [16] and [36], applying a probability network along with spatial attention to identify valuable voxels. Those valuable voxels will get a higher weight through the network to enhance learning efficiency. + +Feature Refinement. We further refine the fused features before we output the final voxel prediction. Instead of using dense operations to refine all grids, we only focus on the grids with high uncertainty. Therefore, we first project the features $\mathbf{F}_f^{3D}$ into semantic classes $\mathbf{L}_{\text{coarse}} \in \mathbb{R}^{N \times H \times W \times Z}$ , where $N$ is the classes number. We then compute the entropy of each grid and select top-k grids with + +the highest entropy scores. These high-uncertainty voxels resample features from ground features $\mathbf{F}_g^{2D}$ through deformable cross-attention. The refined features are then upsampled and output through our semantic head. + +# 3.5. Training Loss + +Following [3] and [18], we train our model with weighted cross-entropy loss $\mathcal{L}_{ce}$ , as well as scene class affinity loss $\mathcal{L}_{scal}^{geo}$ and $\mathcal{L}_{scal}^{sem}$ . Moreover, in order to enhance the supervision, we additionally apply the BEV loss $\mathcal{L}_{bev}$ from the satellite branch and coarse loss $\mathcal{L}_{co}$ from the uncertainty refinement module. Specifically, $\mathcal{L}_{bev}$ is the cross-entropy between the BEV semantic estimation $\mathbf{L}_{BEV} \in \mathbb{R}^{N \times H \times W}$ from the satellite branch and the squeezed ground truth, while $\mathcal{L}_{co}$ is the cross-entropy between $\mathbf{L}_{coarse}$ and downsampled ground truth. These two losses are weighted by scale factor $\lambda_{bev}$ and $\lambda_{co}$ , respectively. The total loss function can be expressed as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s c a l}} ^ {\text {g e o}} + \mathcal {L} _ {\text {s c a l}} ^ {\text {s e m}} + \mathcal {L} _ {c e} + \lambda_ {b e v} \mathcal {L} _ {b e v} + \lambda_ {c o} \mathcal {L} _ {c o}, \tag {8} +$$ + +where the scale factor $L_{bev}$ and $\lambda_{co}$ are set to 1 and 0.25. In addition, we employ the class weight refer to [18]. + +# 4. Experiments + +To fairly evaluate the performance of SGFormer, we conduct experiments on the SemanticKITTI [1] and SSCBench-KITTI-360 [17] datasets. In Sec 4.3, we compare our method against existing approaches on both datasets. We mark the top-3 results of camera-based methods in red, green, and blue. In Sec 4.4, we present ablation studies to demonstrate the effectiveness of each module and selection. Finally, in Sec 4.5, we show our visualization results. + +# 4.1. Datasets + +The SemanticKITTI dataset contains 22 sequences. Among them, 11 sequences are used for training, 1 sequence is for validation, and 10 sequences are for testing. It provides ground images with the shape of $1226 \times 370$ . Meanwhile, corresponding satellite image data are provided by [30]. Each satellite image has the size of $512 \times 512$ , with a scale factor of 0.2 meters per pixel. In this paper, due to the lack of GPS information, we only use 10 sequences to train the model (without sequence 03) and evaluate the performance on the validation set. + +SSCBench-KITTI-360 dataset contains 9 sequences, 7 sequences are used for training, 1 sequence for validation, and 1 for test. The input size of the ground image is $1480 \times 376$ . Since the dataset does not include satellite data, we obtained corresponding satellite images from Google Maps [7] using the ground-truth poses provided. The satellite image settings are consistent with those used in SemanticKITTI. + +
MethodIoUmIoUroad* (15.30%)sidewalk* (11.13%)parking* (11.2%)other-grnd.* (0.56%)building* (0.41%)car (0.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation* (0.93%)trunk (0.51%)terrain* (0.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (0.90%)pole (0.29%)traff-sign (0.08%)
LiDAR-based methods
LMSCNet [26]28.616.7040.6818.224.380.0010.3118.330.000.000.000.0013.660.0220.540.000.001.210.000.00
JS3C-Net [38]38.9810.3150.4923.7411.940.0715.0324.654.410.000.006.1518.114.3326.860.670.270.203.943.771.45
Camera-based methods
MonoScene [3]36.8611.0856.5226.7214.270.4614.0923.266.980.610.451.4817.892.8129.641.861.200.005.844.142.25
TPVFormer [13]35.6111.3656.5025.8720.600.8513.8823.818.080.360.054.3516.922.2630.380.510.890.005.943.141.52
VoxFormer [18]44.0512.3054.4825.9516.210.6817.6125.825.480.610.503.8624.224.8829.241.883.130.007.887.014.20
OccFormer [41]36.5013.4658.8526.8819.610.3114.4025.0925.530.811.198.5219.633.9332.622.782.820.005.614.262.86
SurroundOcc [35]37.2412.7059.2028.2421.421.6714.9426.2614.751.672.377.7319.093.5131.043.602.740.006.654.532.73
Symphonies [15]41.9214.8956.3527.5815.280.9521.6428.6820.442.542.8213.8925.726.6030.873.522.240.008.409.575.76
Ours45.0116.6860.8531.5319.322.1426.0532.1720.302.953.1011.5527.118.2838.473.661.490.009.2811.587.22
+ +Table 1. Quantitative results on SemanticKITTI val set. $\star$ denotes the scene layout structures. + +
MethodIoUmIoUcar(2.85%)bicycle(0.01%)motorcycle(0.00%)truck(0.16%)other-veh.(5.75%)person(0.02%)road*(14.98%)parking*(2.31%)sidewalk*(6.43%)other-grmd.(2.05%)building*(15.67%)fence(0.96%)vegetation*(41.99%)pole(0.22%)traff.-sign(0.06%)other-struct.(4.33%)other-obj.(0.28%)
LiDAR-based methods
SSCNet [31]53.5816.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [26]47.3513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8712.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [13]40.2213.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [18]38.7611.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [41]40.2713.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies [15]43.4117.8226.864.214.9014.207.766.5757.3013.5835.244.5739.207.9534.2319.1914.0416.788.236.04
GaussianFormer [14]35.3812.9218.931.024.6218.077.593.3545.4710.8925.035.3228.445.6829.548.622.993.329.515.14
Ours46.3518.3027.800.912.5510.735.674.2861.0413.2137.005.0743.057.4638.9824.8715.7516.908.855.33
+ +Table 2. Quantitative results on SSCBench-KITTI-360 test set. $\star$ denotes the scene layout structures. + +Both SemanticKITTI and SSCBench-KITTI-360 datasets provide the voxelized point clouds with labels as ground truth, measuring the 3D volume with a range of $51.2m \times 51.2m \times 6.4m$ . The dimension of voxel grids is $256 \times 256 \times 32$ , making each grid cell a 0.2-meter cube. In our evaluation, we report the intersection over union (IoU) for geometry performance and mean IoU (mIoU) metrics for semantic performance, aligning with mainstream works. + +# 4.2. Implementation Details + +We train our SGFormer for 25 epochs on 3 NVIDIA 3090 GPUs, with a batch size of 3. We employ AdamW optimizer [22] with an initial learning rate of 4e-4, with a weight decay of 0.01. We employ the cosine learning rate strategy to reduce the learning rate during the training. Furthermore, + +the feature dimension is set to 128, and each batch consumes around 20GB of GPU memory. + +# 4.3. Main Results + +We compare our method with state-of-the-art on both semanticKITTI and SSCBench-KITTI-360 datasets. The baseline methods include camera-based methods, as well as LiDAR-based methods. Since SSC performance is sensitive to depth quality, we re-evaluated the depth-based method Voxformer [18] and Symphonize [15] using our generated depth data (the depth estimation method is the same). The results for other methods are obtained from [15, 17]. Table 1 and Table 2 show the quantitative results. It is demonstrated that our method achieves superior performance on both geometry (IoU) and semantics predictions (mIoU). Regarding + +
Sat.-branchSat.-corr.FusionIoUmIoUGlobalDetailParams (M)Memory (M)
XXX44.5314.8026.138.2193.4915424
✓XX44.0015.0126.548.26126.4318263
✓✓X43.9015.5928.438.10126.5318967
✓X✓44.7415.8127.509.01126.8919123
✓✓✓45.0116.6829.319.29126.9919865
+ +Table 3. Ablation on main components of SGFormer. + +
IoUmIoUGlobalDetail
w/o Sat.-corr.
w/o noise44.7415.8127.509.01
±5m44.8315.1026.088.70
with Sat.-corr.
w/o noise45.0116.6829.319.29
±5m45.2415.9628.758.53
+ +Table 4. Ablation on localization noise. + +semantic prediction, our method achieved $16.68\mathrm{mIoU}$ on the SemanticKITTI dataset and $18.30\mathrm{mIoU}$ on SSCBenchKITTI-360, significantly outperforming all other methods. For occupancy prediction, our method achieved 45.01 IoU on the SemanticKITTI dataset, surpassing all previous approaches, including VoxFormer, which uses additional steps to estimate occupancy, and all LiDAR-based methods. On the SSCBench-KITTI-360 dataset, our method achieved 46.55 IoU, outperforming all other camera-based methods and achieving results that are also quite comparable to LiDAR-based methods. For a detailed analysis of specific labels, our method performs significantly better than other camera-based approaches in predicting scene layouts. On the SSCBench-KITTI-360 dataset, SGFormer is the only camera-based method to achieve performance comparable to LiDAR-based methods in scene layout estimation. This improvement is due to the additional satellite input, which provides a more complete view of these objects through aerial observations. Moreover, when evaluating the performance of small or dynamic objects, we find that our method also achieves excellent performance. The adaptive fusion module effectively mitigates the negative effects of satellite observations on those small objects. A more detailed analysis of this part is provided in the ablation study section 4.4. + +# 4.4. Ablation Study + +In this section, we conduct the ablations on the SemanticKITTI dataset, mainly analyzing the effectiveness of our core designs and the impact of localization noise. In addition to the IoU and mIoU, we also report the mean prediction accuracy (mIoU) of different semantic categories for further analysis. Specifically, we set those scene layout structures as the "global" category, while other dynamic, small objects are regarded as the "detail" category, reporting the performance of detail estimation. + +# 4.4.1 Effectiveness on Core Designs + +Table 3 presents the ablation on three core designs: the satellite branch (sat.-branch. in the table), satellite correction strategy (sat.-corr. in the table), and adaptive fusion module (fusion in the table). A single ground-branch approach is set as the baseline. + +Satellite branch. As shown, adding a new branch with satellite inputs can improve the overall performance. However, the improvement is slight in both categories, and it leads to a degradation in IoU performance. These observations indicate two key points. First, adding satellite observations does enhance the prediction of scene distribution, but misalignment issues such as satellite image localization errors and vertical occlusions can limit this improvement. Second, the lack of observations on dynamic objects, small objects, and occupancy information in satellite images leads to negative effects on predictions for these "detailed" objects, as well as the geometry prediction. + +Satellite correction strategy. As seen in the table, employing our satellite correction strategy can significantly improve the prediction accuracy of scene layout structures in the "global" category, increasing from 26.54 to 28.43. This improvement demonstrates that our satellite correction strategy can address the misalignment issues. The groundview observation assists the satellite branch in completing the scene layout. + +Adaptive fusion module. The ablation also indicates that using our adaptive fusion module can significantly enhance the accuracy for small and dynamic objects. As shown in the table, with the fusion module, the mIoU of "detail" category is increased from 8.26 to 9.01. Furthermore, adding the fusion module will also lead to an improvement of IoU. + +It is noted that combining all these components yields a substantial performance improvement, as shown in the last row of the table. + +# 4.4.2 Impact of Localization Noise + +We conduct the experiment to explore the impact of the localization noise, as shown in table 4. We add a random noise of $\pm 5$ meters along both latitude and longitude to align the possible localization errors in the real world. Overall, our method is sensitive to localization noise: adding a 5-meter noise results in a decrease of $4.3\%$ and $4.5\%$ mIoU for SG-Former with and without the satellite correction strategy, + +![](images/42300d31d1c827ae3e3e64808482aca1925a0cdb59928f4cdff74b91fa8e5c58.jpg) +Figure 4. Scene Semantic Completion on SemanticKITTI [1]. Left: we show satellite-ground image pairs and indicate the vehicle's travel direction as a yellow arrow. Right: we qualitatively compare scene semantic completion results from SGFormer and other baselines, where SGFormer can produce more complete and accurate semantic reconstruction relying on the satellite-ground fusion. + +respectively. Our satellite correction does indeed reduce the negative impact caused by localization noise, which is particularly evident in the scene layout performance. Without satellite correction, the mIoU for the "global" category dropped to 26.08, even lower than the baseline value in table 3, indicating that satellite observations almost no longer provide a positive effect. However, when using our correction strategy, the performance decline in scene layout prediction is significantly mitigated. + +# 4.5. Qualitative Comparisons + +Figure 4 shows the qualitative comparisons between our SGFormer and two baseline methods, VoxFormer [18] and Symphonies[15] on SemanticKITTI. As shown, the results predicted by our method are significantly better than the other two methods. Overall, our output shows a more accurate scene layout, particularly in complex scenes such as cross-road. The other two methods struggle to reconstruct these complex scenes completely, whereas our method handles this situation effectively. Considering the detail part, VoxFormer produces the scene with radial artifacts, while + +Symphonies performs somewhat better in this regard but still has many cluttered areas due to occlusions. Our reconstructed results not only avoid these issues but also produce clearer boundaries between different semantics. These comparisons demonstrate the superiority of our method both global and local scale. + +# 5. Conclusion + +In this paper, we introduce SGFormer, the first satellite-ground cooperative approach with a tightly coupled dual-branch framework design for urban semantic scene completion. SGFormer effectively fuses observations from groundview and satellite-view images, enabling joint consideration of global scene layout and local details for occupancy grid prediction. Experimental results show that our method surpasses state-of-the-art camera-based approaches and achieves performance on par with LiDAR-based methods across many metrics. These findings suggest that integrating satellite images into the SSC task offers a cost-effective yet highly promising solution. We hope our work inspires further research in this community. + +# Acknowledgment + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 8 +[2] Sudong Cai, Yulan Guo, Salman Khan, Jiwei Hu, and Gongjian Wen. Ground-to-aerial image geo-localization with a hard exemplar reweighting triplet loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8391-8400, 2019. 3 +[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 5, 6 +[4] Yue Cao, Mingsheng Long, Jianmin Wang, and Shichen Liu. Deep visual-semantic quantization for efficient image retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1328–1337, 2017. 3 +[5] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In Conference on Robot Learning, pages 2148-2161. PMLR, 2021. 2 +[6] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 3 +[7] Google. Google maps api, 2024. Accessed: 2024-11-14. 5 +[8] Xiyue Guo, Haocheng Peng, Junjie Hu, Hujun Bao, and Guofeng Zhang. From satellite to ground: Satellite assisted visual localization with cross-view semantic matching. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 3977-3983. IEEE, 2024. 3 +[9] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Hongdong Li. Daniel Cremers. S4c: Self-supervised semantic scene completion with neural fields. In 2024 International Conference on 3D Vision (3DV), pages 409-420. IEEE, 2024. 1 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4 + +This work was partially supported by NSF of China (No. 62425209). The author Junjie Hu acknowledges support from the Guangdong Natural Science Fund under Grant 2024A1515010252. The authors thank Shang Liu for her help in drawing Figures 1, 2, and 3. Special thanks also go to Bo Wen, Xingyuan Yu, Han Zhang, and Ran Zuo for their invaluable support. +[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 5 +[12] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 3 +[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 9223-9232, 2023. 1, 2, 6 +[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2405.17429, 2024. 6 +[15] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20258-20267, 2024. 2, 6, 8 +[16] Peidong Li, Wancheng Shen, Qihao Huang, and Dixiao Cui. Dualbev: Cnn is all you need in view transformation. arXiv preprint arXiv:2403.05402, 2024. 5 +[17] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6 +[18] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 4, 5, 6, 8 +[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 2 +[20] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 4 +[21] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5624-5633, 2019. 3 +[22] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[23] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 2 + +[24] Ian D Miller, Anthony Cowley, Ravi Konkimalla, Shreyas S Shivakumar, Ty Nguyen, Trey Smith, Camillo Jose Taylor, and Vijay Kumar. Any way you look at it: Semantic crossview localization and mapping with lidar. IEEE Robotics and Automation Letters, 6(2):2397-2404, 2021. 3 +[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2 +[26] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 1, 2, 6 +[27] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 3 +[28] Paul-Edouard Sarlin, Eduard Trulls, Marc Pollefeys, Jan Hosang, and Simon Lynen. Snap: Self-supervised neural maps for visual positioning and semantic understanding. Advances in Neural Information Processing Systems, 36, 2024. 3 +[29] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In Proceedings of the IEEE/cvf Winter Conference on Applications of Computer Vision, pages 2417-2426, 2022. 4 +[30] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17010-17020, 2022. 3, 5 +[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1746-1754, 2017. 1, 2, 6 +[32] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 3 +[33] Xin Tan, Wenbin Wu, Zhiwei Zhang, Chaojie Fan, Yong Peng, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision. arXiv preprint arXiv:2405.10591, 2024. 2 +[34] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2 +[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occu + +pancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 6 +[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3–19, 2018. 5 +[37] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 2 +[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 1, 2, 6 +[39] Junyan Ye, Qiyan Luo, Jinhua Yu, Huaping Zhong, Zhimeng Zheng, Conghui He, and Weijia Li. Sg-bev: Satellite-guided bev fusion for cross-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27748-27757, 2024. 3 +[40] Zhu Yu, Runmin Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Si-Yuan Cao, and Hui-liang Shen. Context and geometry aware voxel transformer for semantic scene completion. In Advances in Neural Information Processing Systems, pages 1531-1555, 2024. 1 +[41] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 2, 6 +[42] Yanhao Zhang, Yujiao Shi, Shan Wang, Ankit Vora, Akhil Perincherry, Yongbo Chen, and Hongdong Li. Increasing slam pose accuracy by ground-to-satellite image registration. arXiv preprint arXiv:2404.09169, 2024. 3 +[43] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 4 \ No newline at end of file diff --git a/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/images.zip b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..503aca51b597e03f9add651b9f4f7d3a42e87cd4 --- /dev/null +++ b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99474f0ce0e0382be0271d03fa07c3d00b18bb0a2fcc4a5828e9cecdbf2c3500 +size 758312 diff --git a/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/layout.json b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..218d3a983ace3ee23a4657c7803af97498aaff37 --- /dev/null +++ b/2025/SGFormer_ Satellite-Ground Fusion for 3D Semantic Scene Completion/layout.json @@ -0,0 +1,7943 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "type": "text", + "content": "SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "spans": [ + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": "Xiyue Guo" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": " * Jiarui Hu" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": " * Junjie Hu" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": " Hujun Bao" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": " Guofeng Zhang" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": "State Key Lab of CAD&CG, Zhejiang University " + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 141, + 544, + 174 + ], + "type": "text", + "content": "Chinese University of Hong Kong, Shenzhen" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 180, + 196, + 305 + ], + "blocks": [ + { + "bbox": [ + 56, + 180, + 196, + 305 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 196, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 196, + 305 + ], + "type": "image", + "image_path": "db473754cb9e8d28756643520cf73d3a40023057c727074de02f405e4a8a105b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 307, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 555, + 340 + ], + "type": "text", + "content": "Figure 1. SGFormer, which adopts satellite-ground cooperative fusion, can achieve state-of-the-art performance in scene completion and semantic prediction. Benefiting from informative satellite images and a well-designed dual-branch pipeline, SGFormer can effectively improve semantic prediction accuracy and solve the long-standing visual occlusion bottleneck suffered by purely ground-view methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 198, + 201, + 314, + 304 + ], + "blocks": [ + { + "bbox": [ + 198, + 201, + 314, + 304 + ], + "lines": [ + { + "bbox": [ + 198, + 201, + 314, + 304 + ], + "spans": [ + { + "bbox": [ + 198, + 201, + 314, + 304 + ], + "type": "image", + "image_path": "044e60761fc3c4be2eed91b50a9021513e0a7724c98e5fbc2bac7db6bf924523.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 204, + 436, + 304 + ], + "blocks": [ + { + "bbox": [ + 318, + 204, + 436, + 304 + ], + "lines": [ + { + "bbox": [ + 318, + 204, + 436, + 304 + ], + "spans": [ + { + "bbox": [ + 318, + 204, + 436, + 304 + ], + "type": "image", + "image_path": "c0d667f6b5b23b8b531e39ba28d4f40f90e247d8013f9da403c3dfe469f78096.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 439, + 201, + 554, + 304 + ], + "blocks": [ + { + "bbox": [ + 439, + 201, + 554, + 304 + ], + "lines": [ + { + "bbox": [ + 439, + 201, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 439, + 201, + 554, + 304 + ], + "type": "image", + "image_path": "366326938b3df96b851a53ada3c80f7b812df22266bc8abeaa19330a36cd8a46.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 342, + 200, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 342, + 200, + 354 + ], + "spans": [ + { + "bbox": [ + 152, + 342, + 200, + 354 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 367, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 367, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 297, + 594 + ], + "type": "text", + "content": "Recently, camera-based solutions have been extensively explored for scene semantic completion (SSC). Despite their success in visible areas, existing methods struggle to capture complete scene semantics due to frequent visual occlusions. To address this limitation, this paper presents the first satellite-ground cooperative SSC framework, i.e., SGFormer, exploring the potential of satellite-ground image pairs in the SSC task. Specifically, we propose a dual-branch architecture that encodes orthogonal satellite and ground views in parallel, unifying them into a common domain. Additionally, we design a ground-view guidance strategy that corrects satellite image biases during feature encoding, addressing misalignment between satellite and ground views. Moreover, we develop an adaptive weighting strategy that balances contributions from satellite and ground views. Experiments demonstrate that SGFormer outperforms the state of the art on SemanticKITTI and SSCBench-KITTI-360 datasets. Our code is available on https://github.com/gxytcrc/SGFormer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 617, + 135, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 617, + 135, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 617, + 135, + 629 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 637, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 685 + ], + "type": "text", + "content": "Urban scene semantic completion (SSC) has been an increasingly prominent problem in 3D computer vision over recent decades, which targets predicting 3D semantic and geometric occupancy of immediately observed surround-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 343, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 452 + ], + "type": "text", + "content": "ings with a variety of downstream applications such as autonomous driving, robot navigation, and augmented reality (AR). Lidar-based methods have achieved remarkable progress and decent performance [26, 31, 38], while the underlying point cloud representation inherently suffers from weak semantic context derived only from geometry shapes. In contrast, cost-effective camera-based methods can lift rich 2D cues into the 3D world, demonstrating potential in scene reconstruction and understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "type": "text", + "content": "Existing camera-based methods have shown promising performance on the SSC task [9, 13, 18, 35, 40]. Nevertheless, even with depth information for visible areas, they still face the challenge of non-unique correspondences between 3D volumes and 2D pixels. Specifically, multiple 3D volumes correspond to significantly overlapping regions within the 2D image plane, which causes semantic ambiguity and radial artifacts in the final reconstruction. For occluded areas, these methods generally lack a long-range global perspective, which makes them struggle to restore a complete scene and serve the following planning and decision steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this paper, for the first time, we propose to incorporate satellite imagery into the SSC task. Alongside the development of remote sensing technology, satellite imagery has emerged as a low-cost and widely available reference information. A satellite image, covering major traffic flows in a city, typically requires a light-weight storage footprint, which makes it a compact and memory-efficient representation. Meanwhile, the bird's-eye view (BEV) observation is highly suitable for the horizontal layout of urban scenes. It provides a broad perspective of surrounding obvious ob" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 260, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 260, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 260, + 703 + ], + "type": "text", + "content": "†Corresponding author. Email: zhangguofeng@zju.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 143, + 712 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11929" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "jects to effectively enhance semantic certainty. These two orthogonal views, the satellite and ground view, can optimally compensate for each other's blind fields and provide an ideal solution to the previous occlusion bottleneck." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 289 + ], + "type": "text", + "content": "However, there are two difficulties in incorporating satellite imagery into the SSC task. The first is the misalignment issue. A local satellite image is captured as a fixed-size 2D segment centered around a specific location. An unpredictable deviation from noisy localization and top-down occlusions in the satellite image are the root causes of this issue. Misaligned satellite features tend to bring feature-level inconsistency in fusion and have a negative impact on convergence efficiency. Second, satellite images are typically pre-captured and mainly focus on the scene layout, while inevitably discarding details and long-term changes, such as traffic signs and temporarily parked vehicles, and degrading SSC performance in such low-occupancy but important regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 529 + ], + "type": "text", + "content": "To solve the above challenges, we propose SGFormer, a tightly coupled satellite-assisted framework for the SSC task. In order to perform satellite-ground cross-view fusion, SGFormer innovatively proposes a tailored dual-branch framework to synchronously encode ground and satellite images, aligning them within a unified feature domain. To overcome the misalignment issue, we design a feature-level satellite correction strategy regarding vertically squeezed ground-view features. Specifically, we initialize the learnable BEV parameters and iteratively query squeezed ground-view features in the deformable self-attention layer, where ground-view guidance plays a crucial role in warming up BEV queries in advance to enable coordinated fusion. In addition, we propose an adaptive fusion module with a dual-path weight generator to achieve a reasonable trade-off between two orthogonal views. Within this module, the concatenated satellite-ground features are decoded into channel-wise weight vectors for both views in each voxel. This module allows our fusion pipeline to handle temporal updates and objects of various sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 530, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 613 + ], + "type": "text", + "content": "We evaluate our method on two benchmark semantic datasets, including SemanticKITTI [1] and SSCBenchKITTI-360 [17]. Extensive experimental results demonstrate that our SGFormer outperforms previous approaches, yielding superior performance in terms of scene completion and semantic prediction, as shown in Figure 1. Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 618, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 67, + 618, + 295, + 653 + ], + "type": "text", + "content": "- We present the first tightly coupled satellite-assisted SSC framework with a dual-branch design tailored for satellite observations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 295, + 689 + ], + "type": "text", + "content": "- We propose a feature-level satellite correction strategy based on deformable self-attention to address the misalignment between satellite and ground views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We develop an adaptive weighting strategy to balance contributions from satellite and ground view, enabling" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "effective perception of dynamic updates and objects of different sizes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 96, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 96, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 326, + 96, + 555, + 133 + ], + "type": "text", + "content": "- Experiments on various datasets illustrate that our method can achieve better scene semantic completion results compared to baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 152, + 400, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 152, + 400, + 165 + ], + "spans": [ + { + "bbox": [ + 314, + 152, + 400, + 165 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 174, + 484, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 174, + 484, + 186 + ], + "spans": [ + { + "bbox": [ + 314, + 174, + 484, + 186 + ], + "type": "text", + "content": "2.1.3D Semantic Scene Completion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 194, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 326 + ], + "type": "text", + "content": "3D Semantic Scene Completion (SSC) was first proposed by SSCNet [31], aiming to predict the occupancy and semantic information of each voxel in 3D spaces [5, 26, 37, 38]. Subsequent methods can be categorized into two types: one relies on depth sensors, such as LiDAR, to directly compute spatial features for estimating 3D semantics, while the other is camera-based and involves lifting image features to the 3D spaces before estimating semantic information. Earlier approaches mainly focused on the first type, while in recent years, camera-based solutions have gained increasing attention due to their high cost-efficiency [3, 13, 18, 35]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 388 + ], + "type": "text", + "content": "MonoScene [3] presents the first pure vision solution, leveraging 2D and 3D U-Nets bridged by a sight projection module. OccFormer [41] follows the strategy of LSS [25], estimates the depth distribution of the image, and then projects the features into 3D spaces through depth guidance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 388, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 581 + ], + "type": "text", + "content": "Some works are inspired by BEVFormer [19], which utilizes the Transformer framework to estimate occupancy. These methods use spatial relationships from 3D to 2D to query the information from image features via spatial deformable attention [43]. Among them, TPVFormer [13] proposes a Tri-perspective view representation. It first acquires the features of these planes and recovers them into 3D accordingly. SurrunodOcc [35] introduces a coarse-to-fine strategy, generating 3D features at multiple scales and progressively integrating occupancy grid predictions. These transformer-based methods show significant progress in the performance of 3D semantic prediction. However, due to the non-one-to-one nature of the 3D-2D projection relationship and the lack of geometric constraints, these methods struggle to reconstruct the semantic distribution of 3D scenes accurately." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "To address the problem, some works leverage depth input and try to incorporate geometric priors into occupancy predictions. OccDepth [23] projects the image features by stereo depth. Voxformer [18] initializes the sparse proposal from pre-trained depth. Subsequent studies [15, 33, 34] have further improved upon this by integrating in-depth information, thereby enhancing the performance of SSC. However, due to inherent issues arising from limitations of ground camera observation range and occlusions make predictive semantic ambiguities and geometric distribution errors almost inevitable." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11930" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 555, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 220 + ], + "type": "image", + "image_path": "b68b3ba5d8943e25ebd867582e8c5cb0cdd8bdf9560a5367e96774c1982ebfd7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 224, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 555, + 302 + ], + "type": "text", + "content": "Figure 2. Overview of SGFormer. Overall, SGFormer feeds a satellite-ground image pair into similar backbone networks in different branches to extract multi-level feature maps respectively (Left Part). Then, leveraging deformable attention, it transforms satellite and ground features into volume and BEV spaces (Middle Part) for following feature fusion and decoding (Right Part). Specifically, in the ground branch, we use a depth estimator to produce voxel proposals for targeted querying on non-empty feature volumes. In the satellite branch, we fuse vertically squeezed ground-view features into BEV queries to warm up satellite features (Red Line). Before final fusion, encoded features from both branches are enhanced through 2D/3D convolution networks. Our proposed fusion module, which is detailed in Figure 3, is able to adaptively fuse satellite and ground features, followed by a seg head to output semantic reconstruction results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 310, + 199, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 199, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 199, + 323 + ], + "type": "text", + "content": "2.2. Satellite Assist Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": "Recently, integrating satellite images with ground images has gained increasing attention. This is mainly due to the low economic and storage costs of satellite images, as well as the wealth of information they contain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 594 + ], + "type": "text", + "content": "Most of these works mainly focus on the localization problem, aiming at estimating the position of ground vehicles in real-world coordinates by matching ground and satellite images. Some methods divide satellite maps into many small patches, aiming to find the patch that is most similar to the ground image through image retrieval methods [2, 4, 12, 21]. Other approaches try integrating satellite and ground features into the common coordinate system to achieve higher localization accuracy, such as applying homography to project satellite features into the ground perspective and then determining precise localization results [30] or projecting ground images into a bird's-eye view (BEV) and aligning them with satellite features [6, 8, 27]. On this basis, some works have attempted to incorporate the results of such cross-view localization into traditional SLAM systems to enhance SLAM localization performance [8, 24, 42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 296, + 672 + ], + "type": "text", + "content": "In addition to localization efforts, some works have incorporated satellite images into map construction tasks. SG-BEV [39] is the first to combine ground and satellite features to accomplish fine-grained building attribute segmentation tasks, while SNAP [28] included satellite images in the construction of 2D neural maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Against this backdrop, our approach is the first to introduce satellite images into the SSC task, exploring the potential of cooperative satellite and ground perception." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 309, + 370, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 309, + 370, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 309, + 370, + 322 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 330, + 382, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 330, + 382, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 330, + 382, + 342 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 349, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 468 + ], + "type": "text", + "content": "The framework of our proposed SGFormer is illustrated in Figure 2. Our work takes both ground and satellite images as inputs, processed through two distinct branches: the ground branch and the satellite branch. Each branch consists of two-step operations, where the first is the feature extraction to generate multi-scale features from ground and satellite images and the second is the feature transformation to covert the features into voxel or BEV spaces, followed by feature diffusion and enhancement. Then, the voxel and BEV features are fused through a fusion module." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 469, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 613 + ], + "type": "text", + "content": "Within SGFormer, we make following main technical contributions: 1) Satellite-view feature correction: the satellite correction method employs compressed features from the ground branch to guide and correct the feature learning of the satellite branch (detailed in Sec. 3.3.2). 2) Adaptive fusion: The adaptive fusion module merges BEV features from the satellite branch with voxel features from the ground branch through the attention mechanism. After fusing features from both branches, we further refine them and pass the output into a segmentation head, which up-samples these features and produces voxel-wise class predictions (detailed in Sec. 3.4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 622, + 429, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 622, + 429, + 633 + ], + "spans": [ + { + "bbox": [ + 314, + 622, + 429, + 633 + ], + "type": "text", + "content": "3.2. Feature Extraction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "text", + "content": "Feature extraction of the ground branch aims to extract 2D features " + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{2D} \\in \\mathbb{R}^{H_g' \\times W_g' \\times D}" + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "text", + "content": " from the ground image, where " + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "inline_equation", + "content": "H_g' \\times W_g'" + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "text", + "content": " represents the feature resolution, and " + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "text", + "content": " is the feature dimension. Similarly, in the satellite branch, we obtain 2D satellite features " + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^{2D} \\in \\mathbb{R}^{H_s' \\times W_s' \\times D}" + }, + { + "bbox": [ + 313, + 640, + 556, + 715 + ], + "type": "text", + "content": ". EfficientNet-B7 [32] serves as the backbone for ground images, while" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11931" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 294, + 238 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 294, + 238 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 294, + 238 + ], + "type": "image", + "image_path": "937e85317ecaf6ea4e7a2ad6ca34ddfe6661dda284f80279dfdc444a781b4779.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "type": "text", + "content": "Figure 3. Fusion Module. Our fusion module (Top) mainly consists of two parts: the weight module and the probability network. This weight module (Bottom Left) can facilitate mutual perception between the BEV feature " + }, + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^{\\prime BEV}" + }, + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "type": "text", + "content": " and voxel feature " + }, + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{\\prime 3D}" + }, + { + "bbox": [ + 55, + 247, + 296, + 335 + ], + "type": "text", + "content": " in both 2D and 3D spaces to produce channel-wise weight vectors. This probability network (Bottom Right) takes the weighted average 3D voxel feature as input to infer the approximate occupancy probability per voxel to achieve better geometry reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 343, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 487 + ], + "type": "text", + "content": "ResNet-50 [10] is used for satellite images, with both backbones followed by a Feature Pyramid Network (FPN) [20]. Depth Estimation. Before encoding the features of ground view to 3D spaces, we estimate the visible voxels using depth estimation, employing the pre-trained MobileStereoNet [29] for depth estimation, aligning to VoxFormer [18]. However, instead of using an additional stage to refine the binary occupancy like VoxFormer [18], we directly employ the voxelized depth as our query proposals for feature encoding. Specifically, depth points are projected into the voxel map, with a voxel set to 1 if the number of occupied points exceeds a threshold; otherwise, it is set to 0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 495, + 191, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 191, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 191, + 506 + ], + "type": "text", + "content": "3.3. Feature Transformation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "text", + "content": "To efficiently transform the image features to real-world coordinates, we use deformable attention [43], represented as DA, to aggregate query features " + }, + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "text", + "content": " from target image features " + }, + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "text", + "content": ", as described by the equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 569, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 97, + 569, + 295, + 602 + ], + "type": "interline_equation", + "content": "\\mathrm {D A} (q, p, v) = \\sum_ {k = 1} ^ {K} A _ {k} W v (p + \\Delta p), \\tag {1}", + "image_path": "f035cba749b970e6dff1e535274d1ca7137bcc517dc5c0caab645ab66e919df1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "text", + "content": " is the reference point, " + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "inline_equation", + "content": "\\Delta p" + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "text", + "content": " is the learnable sampling offset, " + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "text", + "content": " represents the attention weight, and " + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 609, + 295, + 646 + ], + "type": "text", + "content": " denotes the projection weight." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 250, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 250, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 250, + 670 + ], + "type": "text", + "content": "3.3.1 Ground-view Feature Transformation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "In the ground branch, we initialize the voxel queries " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_v\\in" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times Z\\times D}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " with learnable embeddings, where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "H\\times W\\times" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " is regarded as the voxel dimension. Guided by the query" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "proposals, we query the voxel features from image features " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{2D}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " by employing the deformable cross-attention:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 368, + 106, + 555, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 106, + 555, + 124 + ], + "spans": [ + { + "bbox": [ + 368, + 106, + 555, + 124 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {p} = \\mathrm {D A} \\left(\\mathbf {Q} _ {p}, \\mathcal {P} (\\mathbf {p}, g), \\mathbf {F} _ {g} ^ {2 D}\\right), \\tag {2}", + "image_path": "9484a26da1a4c81f8723466de1a897f886b9727eb31b47348fa18c37c88a4e7c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "content": " consists of the visible voxels selected by the query proposals. " + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathbf{p},g)" + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "content": " denotes the corresponding pixels in " + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "inline_equation", + "content": "F_{q}^{2D}" + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "content": " generated by camera projection function." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "text", + "content": "After several layers of cross-attention, we merge the query features " + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "inline_equation", + "content": "Q_{p}" + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "text", + "content": " with other invisible voxels in " + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "inline_equation", + "content": "Q_{v}" + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "text", + "content": " to get the voxel features " + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}^{3D}" + }, + { + "bbox": [ + 313, + 166, + 555, + 213 + ], + "type": "text", + "content": ", and then diffuse the features to all voxels through deformable self-attention, as formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 221, + 555, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 221, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 372, + 221, + 555, + 239 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {g} ^ {3 D} = \\mathrm {D A} \\left(\\mathbf {F} _ {g} ^ {3 D}, \\mathbf {p _ {v}}, \\mathbf {F} _ {g} ^ {3 D}\\right), \\tag {3}", + "image_path": "d670aa17bc61ecd9044d256fa3c0901af55f68dfb8147d930918901e86796fa8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 246, + 494, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 494, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 494, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 246, + 494, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_v" + }, + { + "bbox": [ + 313, + 246, + 494, + 258 + ], + "type": "text", + "content": " is the reference point in 3D space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 271, + 510, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 510, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 510, + 283 + ], + "type": "text", + "content": "3.3.2 Satellite-view Feature Transformation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": "In the satellite branch, we encode the satellite features to BEV space under the guidance of ground observations. We first initialize the BEV queries " + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{bev} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": ". Simultaneously, we compress the voxel features " + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{3D}" + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": " from the ground branch into BEV form " + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{BEV}" + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": " through max pooling. We then combine " + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{bev}" + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{BEV}" + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": " to form a hybrid feature " + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "inline_equation", + "content": "v_{hybrid}" + }, + { + "bbox": [ + 313, + 289, + 555, + 376 + ], + "type": "text", + "content": ", which is fed into a self-attention module:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 386, + 555, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 386, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 362, + 386, + 555, + 399 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {b e v} = \\mathrm {D A} \\left(\\mathbf {Q} _ {b e v}, \\mathbf {p} _ {b e v}, v _ {\\text {h y b r i d}}\\right), \\tag {4}", + "image_path": "027d14a95c86616620b54b5980a09dea1e0c291ef885c42a69f234a077931876.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "type": "text", + "content": " is the reference point in BEV space. Through self-attention, the BEV queries are efficiently fused from ground-view information. The output queries are then passed into the cross-attention module to query the information from the satellite features " + }, + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^{2D}" + }, + { + "bbox": [ + 313, + 408, + 554, + 469 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 476, + 554, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 476, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 361, + 476, + 554, + 491 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {b e v} = \\mathrm {D A} \\left(\\mathbf {Q} _ {b e v}, \\mathcal {P} (\\mathbf {p}, s), \\mathbf {F} _ {s} ^ {2 D}\\right), \\tag {5}", + "image_path": "c4ede0e70fac64006544c48d4ee0d5c42fb4721f388fa9ed08ef026072308405.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathbf{p}, s)" + }, + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "text", + "content": " represents the corresponding pixels of BEV grids in the satellite image. With the ground-view observation, the offset layers predict more suitable offsets during cross-attention. We iterate the self and cross attention several times, resulting in the BEV features " + }, + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^{BEV}" + }, + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 573, + 470, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 470, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 470, + 585 + ], + "type": "text", + "content": "3.3.3 Convolutional Enhancement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 592, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 651 + ], + "type": "text", + "content": "After encoding the features into voxel and BEV spaces, we feed them into convolutional layers to further enhance feature representations through neighborhood interactions. We use 3D-UNet for feature aggregation in the ground branch and 2D-UNet in the satellite branch." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 659, + 408, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 408, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 408, + 670 + ], + "type": "text", + "content": "3.4. Feature Fusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Our adaptive fusion module is shown in Figure 3. The BEV features " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_s^{\\prime BEV}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " of the satellite branch can cover a wider range, allowing for a more complete representation of scene" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11932" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "Layouts such as roads and buildings. However, BEV features lack occupancy geometry (empty or occupied) and do not perform well on small objects. On the other hand, voxel features " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}^{r3D}" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " based on ground images have significant advantages in handling small, dynamic objects and detailed elements, as well as spatial occupancy, but the effective information they contain is limited to the unoccluded field of view. Therefore, it is important to keep the advantages of each view while minimizing the impact of their limitations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 179, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 294, + 251 + ], + "type": "text", + "content": "To address the challenges, we propose a fusion module that dynamically predicts weights for both the feature channel and spatial domains. Specifically, we process the features from two branches: a channel-attention network [11] and a spatial-attention network to balance the contributions at the object level and perception filed level." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "text", + "content": "In the channel-attention path, we first lift the BEV features " + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{s}^{\\prime BEV}" + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "text", + "content": " to 3D volumes, denoted as " + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{s}^{\\prime 3D}" + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "text", + "content": ". Then, we concatenate the features from the ground and satellite branches in 3D volume to form the combined features and pass them into the channel-attention network, resulting in channel-domain weights " + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_c\\in \\mathbb{R}^{D\\times 1\\times 1\\times 1}" + }, + { + "bbox": [ + 55, + 251, + 294, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "content": "In the spatial-attention path, we compress the voxel features " + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}^{\\prime 3D}" + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "content": "-axis into BEV space, resulting in " + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{g}^{\\prime BEV}" + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "content": ". Similarly to the channel-attention path, we concatenate the BEV features and pass them into the spatial-attention network, generating spatial-domain weights " + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "content": " in BEV space. The two attention weights are summed together and then combined with the result from an additional MLP to obtain the fused weight " + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_a \\in \\mathbb{R}^{D \\times H \\times W \\times Z}" + }, + { + "bbox": [ + 55, + 323, + 294, + 434 + ], + "type": "text", + "content": ", expressed as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 440, + 295, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 440, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 86, + 440, + 295, + 455 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {a} = \\operatorname {M L P} \\left(\\mathbf {F} _ {c} ^ {\\prime 3 D}\\right) \\oplus \\mathrm {C} \\left(\\mathbf {F} _ {c} ^ {\\prime 3 D}\\right) \\oplus \\mathrm {S} \\left(\\mathbf {F} _ {c} ^ {\\prime B E V}\\right), \\tag {6}", + "image_path": "7057b61a84758799e1e693c9402f3cc0b7ebe93fe768ff95a2cac073879083a0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "text", + "content": "where C and S represent the channel and spatial attention networks, respectively. " + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{c}^{\\prime 3D}" + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{c}^{\\prime BEV}" + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "text", + "content": " are the concatenated features in 3D and BEV spaces. MLP refers to the MLP operation applied to " + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{c}^{\\prime 3D}" + }, + { + "bbox": [ + 55, + 463, + 294, + 522 + ], + "type": "text", + "content": ". After obtaining the attention weight, we fuse the two branches' features by equation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 530, + 294, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 530, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 93, + 530, + 294, + 547 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {f} ^ {3 D} = \\mathbf {W} _ {a} \\cdot \\mathbf {F} _ {g} ^ {\\prime 3 D} + (1 - \\mathbf {W} _ {a}) \\cdot \\mathbf {F} _ {s} ^ {\\prime 3 D}, \\tag {7}", + "image_path": "c8ff1b31c849911bc9b7f932a40985da50f2ccdbf46c014300e01cc228dd4504.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 212, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 212, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 212, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 555, + 212, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f^{3D}" + }, + { + "bbox": [ + 55, + 555, + 212, + 567 + ], + "type": "text", + "content": " is the fused voxel features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 568, + 294, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 294, + 628 + ], + "type": "text", + "content": "Additionally, to minimize the negative impact of empty voxels, we get inspiration from [16] and [36], applying a probability network along with spatial attention to identify valuable voxels. Those valuable voxels will get a higher weight through the network to enhance learning efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "text", + "content": "Feature Refinement. We further refine the fused features before we output the final voxel prediction. Instead of using dense operations to refine all grids, we only focus on the grids with high uncertainty. Therefore, we first project the features " + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_f^{3D}" + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "text", + "content": " into semantic classes " + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\text{coarse}} \\in \\mathbb{R}^{N \\times H \\times W \\times Z}" + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "text", + "content": " is the classes number. We then compute the entropy of each grid and select top-k grids with" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "the highest entropy scores. These high-uncertainty voxels resample features from ground features " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_g^{2D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " through deformable cross-attention. The refined features are then upsampled and output through our semantic head." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 126, + 402, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 402, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 402, + 138 + ], + "type": "text", + "content": "3.5. Training Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": "Following [3] and [18], we train our model with weighted cross-entropy loss " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": ", as well as scene class affinity loss " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{scal}^{geo}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{scal}^{sem}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": ". Moreover, in order to enhance the supervision, we additionally apply the BEV loss " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bev}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " from the satellite branch and coarse loss " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " from the uncertainty refinement module. Specifically, " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bev}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " is the cross-entropy between the BEV semantic estimation " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{BEV} \\in \\mathbb{R}^{N \\times H \\times W}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " from the satellite branch and the squeezed ground truth, while " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{co}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " is the cross-entropy between " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{coarse}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " and downsampled ground truth. These two losses are weighted by scale factor " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\lambda_{bev}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "inline_equation", + "content": "\\lambda_{co}" + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": ", respectively. The total loss function can be expressed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 293, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 293, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 329, + 293, + 553, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} + \\mathcal {L} _ {c e} + \\lambda_ {b e v} \\mathcal {L} _ {b e v} + \\lambda_ {c o} \\mathcal {L} _ {c o}, \\tag {8}", + "image_path": "504d8828fb48ff4059a435a81e367835cf804d51c24830e7977909e0191ee608.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "type": "text", + "content": "where the scale factor " + }, + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "type": "inline_equation", + "content": "L_{bev}" + }, + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "type": "inline_equation", + "content": "\\lambda_{co}" + }, + { + "bbox": [ + 313, + 312, + 553, + 337 + ], + "type": "text", + "content": " are set to 1 and 0.25. In addition, we employ the class weight refer to [18]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 346, + 394, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 394, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 394, + 360 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 366, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 472 + ], + "type": "text", + "content": "To fairly evaluate the performance of SGFormer, we conduct experiments on the SemanticKITTI [1] and SSCBench-KITTI-360 [17] datasets. In Sec 4.3, we compare our method against existing approaches on both datasets. We mark the top-3 results of camera-based methods in red, green, and blue. In Sec 4.4, we present ablation studies to demonstrate the effectiveness of each module and selection. Finally, in Sec 4.5, we show our visualization results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 376, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 376, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 376, + 491 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "text", + "content": "The SemanticKITTI dataset contains 22 sequences. Among them, 11 sequences are used for training, 1 sequence is for validation, and 10 sequences are for testing. It provides ground images with the shape of " + }, + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "inline_equation", + "content": "1226 \\times 370" + }, + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "text", + "content": ". Meanwhile, corresponding satellite image data are provided by [30]. Each satellite image has the size of " + }, + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "text", + "content": ", with a scale factor of 0.2 meters per pixel. In this paper, due to the lack of GPS information, we only use 10 sequences to train the model (without sequence 03) and evaluate the performance on the validation set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 712 + ], + "type": "text", + "content": "SSCBench-KITTI-360 dataset contains 9 sequences, 7 sequences are used for training, 1 sequence for validation, and 1 for test. The input size of the ground image is " + }, + { + "bbox": [ + 313, + 617, + 553, + 712 + ], + "type": "inline_equation", + "content": "1480 \\times 376" + }, + { + "bbox": [ + 313, + 617, + 553, + 712 + ], + "type": "text", + "content": ". Since the dataset does not include satellite data, we obtained corresponding satellite images from Google Maps [7] using the ground-truth poses provided. The satellite image settings are consistent with those used in SemanticKITTI." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11933" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 261 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 261 + ], + "type": "table", + "html": "
MethodIoUmIoUroad* (15.30%)sidewalk* (11.13%)parking* (11.2%)other-grnd.* (0.56%)building* (0.41%)car (0.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation* (0.93%)trunk (0.51%)terrain* (0.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (0.90%)pole (0.29%)traff-sign (0.08%)
LiDAR-based methods
LMSCNet [26]28.616.7040.6818.224.380.0010.3118.330.000.000.000.0013.660.0220.540.000.001.210.000.00
JS3C-Net [38]38.9810.3150.4923.7411.940.0715.0324.654.410.000.006.1518.114.3326.860.670.270.203.943.771.45
Camera-based methods
MonoScene [3]36.8611.0856.5226.7214.270.4614.0923.266.980.610.451.4817.892.8129.641.861.200.005.844.142.25
TPVFormer [13]35.6111.3656.5025.8720.600.8513.8823.818.080.360.054.3516.922.2630.380.510.890.005.943.141.52
VoxFormer [18]44.0512.3054.4825.9516.210.6817.6125.825.480.610.503.8624.224.8829.241.883.130.007.887.014.20
OccFormer [41]36.5013.4658.8526.8819.610.3114.4025.0925.530.811.198.5219.633.9332.622.782.820.005.614.262.86
SurroundOcc [35]37.2412.7059.2028.2421.421.6714.9426.2614.751.672.377.7319.093.5131.043.602.740.006.654.532.73
Symphonies [15]41.9214.8956.3527.5815.280.9521.6428.6820.442.542.8213.8925.726.6030.873.522.240.008.409.575.76
Ours45.0116.6860.8531.5319.322.1426.0532.1720.302.953.1011.5527.118.2838.473.661.490.009.2811.587.22
", + "image_path": "5eaf0b2e10ffe8fa2742fe2ce7b3306a934109fb976216a4ebe4ef0cbe365198.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 293, + 550, + 500 + ], + "blocks": [ + { + "bbox": [ + 133, + 270, + 476, + 281 + ], + "lines": [ + { + "bbox": [ + 133, + 270, + 476, + 281 + ], + "spans": [ + { + "bbox": [ + 133, + 270, + 476, + 281 + ], + "type": "text", + "content": "Table 1. Quantitative results on SemanticKITTI val set. " + }, + { + "bbox": [ + 133, + 270, + 476, + 281 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 133, + 270, + 476, + 281 + ], + "type": "text", + "content": " denotes the scene layout structures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 293, + 550, + 500 + ], + "lines": [ + { + "bbox": [ + 60, + 293, + 550, + 500 + ], + "spans": [ + { + "bbox": [ + 60, + 293, + 550, + 500 + ], + "type": "table", + "html": "
MethodIoUmIoUcar(2.85%)bicycle(0.01%)motorcycle(0.00%)truck(0.16%)other-veh.(5.75%)person(0.02%)road*(14.98%)parking*(2.31%)sidewalk*(6.43%)other-grmd.(2.05%)building*(15.67%)fence(0.96%)vegetation*(41.99%)pole(0.22%)traff.-sign(0.06%)other-struct.(4.33%)other-obj.(0.28%)
LiDAR-based methods
SSCNet [31]53.5816.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [26]47.3513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Camera-based methods
MonoScene [3]37.8712.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [13]40.2213.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
VoxFormer [18]38.7611.9117.841.160.894.562.061.6347.019.6727.212.8931.184.9728.9914.696.516.923.792.43
OccFormer [41]40.2713.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
Symphonies [15]43.4117.8226.864.214.9014.207.766.5757.3013.5835.244.5739.207.9534.2319.1914.0416.788.236.04
GaussianFormer [14]35.3812.9218.931.024.6218.077.593.3545.4710.8925.035.3228.445.6829.548.622.993.329.515.14
Ours46.3518.3027.800.912.5510.735.674.2861.0413.2137.005.0743.057.4638.9824.8715.7516.908.855.33
", + "image_path": "4b66156924c013f2d5d94eefe4a09c6d078ea5b4721d53b417f7888ee5a8b81d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 508, + 490, + 518 + ], + "lines": [ + { + "bbox": [ + 117, + 508, + 490, + 518 + ], + "spans": [ + { + "bbox": [ + 117, + 508, + 490, + 518 + ], + "type": "text", + "content": "Table 2. Quantitative results on SSCBench-KITTI-360 test set. " + }, + { + "bbox": [ + 117, + 508, + 490, + 518 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 117, + 508, + 490, + 518 + ], + "type": "text", + "content": " denotes the scene layout structures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": "Both SemanticKITTI and SSCBench-KITTI-360 datasets provide the voxelized point clouds with labels as ground truth, measuring the 3D volume with a range of " + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "inline_equation", + "content": "51.2m \\times 51.2m \\times 6.4m" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": ". The dimension of voxel grids is " + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": ", making each grid cell a 0.2-meter cube. In our evaluation, we report the intersection over union (IoU) for geometry performance and mean IoU (mIoU) metrics for semantic performance, aligning with mainstream works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 188, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 188, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 188, + 647 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "We train our SGFormer for 25 epochs on 3 NVIDIA 3090 GPUs, with a batch size of 3. We employ AdamW optimizer [22] with an initial learning rate of 4e-4, with a weight decay of 0.01. We employ the cosine learning rate strategy to reduce the learning rate during the training. Furthermore," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 532, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 553, + 556 + ], + "type": "text", + "content": "the feature dimension is set to 128, and each batch consumes around 20GB of GPU memory." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 563, + 399, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 399, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 399, + 574 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We compare our method with state-of-the-art on both semanticKITTI and SSCBench-KITTI-360 datasets. The baseline methods include camera-based methods, as well as LiDAR-based methods. Since SSC performance is sensitive to depth quality, we re-evaluated the depth-based method Voxformer [18] and Symphonize [15] using our generated depth data (the depth estimation method is the same). The results for other methods are obtained from [15, 17]. Table 1 and Table 2 show the quantitative results. It is demonstrated that our method achieves superior performance on both geometry (IoU) and semantics predictions (mIoU). Regarding" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11934" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 70, + 492, + 148 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 492, + 148 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 492, + 148 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 492, + 148 + ], + "type": "table", + "html": "
Sat.-branchSat.-corr.FusionIoUmIoUGlobalDetailParams (M)Memory (M)
XXX44.5314.8026.138.2193.4915424
✓XX44.0015.0126.548.26126.4318263
✓✓X43.9015.5928.438.10126.5318967
✓X✓44.7415.8127.509.01126.8919123
✓✓✓45.0116.6829.319.29126.9919865
", + "image_path": "c87e375af3fbf81e45a4f39ac6995eba9fe1c32f4b44633c6d5bede716cb16e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 92, + 168, + 261, + 264 + ], + "blocks": [ + { + "bbox": [ + 207, + 152, + 402, + 163 + ], + "lines": [ + { + "bbox": [ + 207, + 152, + 402, + 163 + ], + "spans": [ + { + "bbox": [ + 207, + 152, + 402, + 163 + ], + "type": "text", + "content": "Table 3. Ablation on main components of SGFormer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 168, + 261, + 264 + ], + "lines": [ + { + "bbox": [ + 92, + 168, + 261, + 264 + ], + "spans": [ + { + "bbox": [ + 92, + 168, + 261, + 264 + ], + "type": "table", + "html": "
IoUmIoUGlobalDetail
w/o Sat.-corr.
w/o noise44.7415.8127.509.01
±5m44.8315.1026.088.70
with Sat.-corr.
w/o noise45.0116.6829.319.29
±5m45.2415.9628.758.53
", + "image_path": "57bd5918e7a86d34f581019c80ab73517d81475d5a1cd1c76271ebc2b6e5ec5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 273, + 247, + 283 + ], + "lines": [ + { + "bbox": [ + 103, + 273, + 247, + 283 + ], + "spans": [ + { + "bbox": [ + 103, + 273, + 247, + 283 + ], + "type": "text", + "content": "Table 4. Ablation on localization noise." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "type": "text", + "content": "semantic prediction, our method achieved " + }, + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "type": "inline_equation", + "content": "16.68\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "type": "text", + "content": " on the SemanticKITTI dataset and " + }, + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "type": "inline_equation", + "content": "18.30\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 293, + 295, + 581 + ], + "type": "text", + "content": " on SSCBenchKITTI-360, significantly outperforming all other methods. For occupancy prediction, our method achieved 45.01 IoU on the SemanticKITTI dataset, surpassing all previous approaches, including VoxFormer, which uses additional steps to estimate occupancy, and all LiDAR-based methods. On the SSCBench-KITTI-360 dataset, our method achieved 46.55 IoU, outperforming all other camera-based methods and achieving results that are also quite comparable to LiDAR-based methods. For a detailed analysis of specific labels, our method performs significantly better than other camera-based approaches in predicting scene layouts. On the SSCBench-KITTI-360 dataset, SGFormer is the only camera-based method to achieve performance comparable to LiDAR-based methods in scene layout estimation. This improvement is due to the additional satellite input, which provides a more complete view of these objects through aerial observations. Moreover, when evaluating the performance of small or dynamic objects, we find that our method also achieves excellent performance. The adaptive fusion module effectively mitigates the negative effects of satellite observations on those small objects. A more detailed analysis of this part is provided in the ablation study section 4.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 149, + 601 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "In this section, we conduct the ablations on the SemanticKITTI dataset, mainly analyzing the effectiveness of our core designs and the impact of localization noise. In addition to the IoU and mIoU, we also report the mean prediction accuracy (mIoU) of different semantic categories for further analysis. Specifically, we set those scene layout structures as the \"global\" category, while other dynamic, small objects are regarded as the \"detail\" category, reporting the performance of detail estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 171, + 474, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 474, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 474, + 183 + ], + "type": "text", + "content": "4.4.1 Effectiveness on Core Designs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 189, + 553, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 553, + 249 + ], + "type": "text", + "content": "Table 3 presents the ablation on three core designs: the satellite branch (sat.-branch. in the table), satellite correction strategy (sat.-corr. in the table), and adaptive fusion module (fusion in the table). A single ground-branch approach is set as the baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 249, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 554, + 392 + ], + "type": "text", + "content": "Satellite branch. As shown, adding a new branch with satellite inputs can improve the overall performance. However, the improvement is slight in both categories, and it leads to a degradation in IoU performance. These observations indicate two key points. First, adding satellite observations does enhance the prediction of scene distribution, but misalignment issues such as satellite image localization errors and vertical occlusions can limit this improvement. Second, the lack of observations on dynamic objects, small objects, and occupancy information in satellite images leads to negative effects on predictions for these \"detailed\" objects, as well as the geometry prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 488 + ], + "type": "text", + "content": "Satellite correction strategy. As seen in the table, employing our satellite correction strategy can significantly improve the prediction accuracy of scene layout structures in the \"global\" category, increasing from 26.54 to 28.43. This improvement demonstrates that our satellite correction strategy can address the misalignment issues. The groundview observation assists the satellite branch in completing the scene layout." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 489, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 559 + ], + "type": "text", + "content": "Adaptive fusion module. The ablation also indicates that using our adaptive fusion module can significantly enhance the accuracy for small and dynamic objects. As shown in the table, with the fusion module, the mIoU of \"detail\" category is increased from 8.26 to 9.01. Furthermore, adding the fusion module will also lead to an improvement of IoU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 560, + 554, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 595 + ], + "type": "text", + "content": "It is noted that combining all these components yields a substantial performance improvement, as shown in the last row of the table." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 470, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 470, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 470, + 623 + ], + "type": "text", + "content": "4.4.2 Impact of Localization Noise" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "We conduct the experiment to explore the impact of the localization noise, as shown in table 4. We add a random noise of " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\pm 5" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " meters along both latitude and longitude to align the possible localization errors in the real world. Overall, our method is sensitive to localization noise: adding a 5-meter noise results in a decrease of " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " mIoU for SG-Former with and without the satellite correction strategy," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11935" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 82, + 521, + 403 + ], + "blocks": [ + { + "bbox": [ + 71, + 82, + 521, + 403 + ], + "lines": [ + { + "bbox": [ + 71, + 82, + 521, + 403 + ], + "spans": [ + { + "bbox": [ + 71, + 82, + 521, + 403 + ], + "type": "image", + "image_path": "42300d31d1c827ae3e3e64808482aca1925a0cdb59928f4cdff74b91fa8e5c58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 555, + 445 + ], + "type": "text", + "content": "Figure 4. Scene Semantic Completion on SemanticKITTI [1]. Left: we show satellite-ground image pairs and indicate the vehicle's travel direction as a yellow arrow. Right: we qualitatively compare scene semantic completion results from SGFormer and other baselines, where SGFormer can produce more complete and accurate semantic reconstruction relying on the satellite-ground fusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": "respectively. Our satellite correction does indeed reduce the negative impact caused by localization noise, which is particularly evident in the scene layout performance. Without satellite correction, the mIoU for the \"global\" category dropped to 26.08, even lower than the baseline value in table 3, indicating that satellite observations almost no longer provide a positive effect. However, when using our correction strategy, the performance decline in scene layout prediction is significantly mitigated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 573, + 195, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 195, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 195, + 586 + ], + "type": "text", + "content": "4.5. Qualitative Comparisons" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Figure 4 shows the qualitative comparisons between our SGFormer and two baseline methods, VoxFormer [18] and Symphonies[15] on SemanticKITTI. As shown, the results predicted by our method are significantly better than the other two methods. Overall, our output shows a more accurate scene layout, particularly in complex scenes such as cross-road. The other two methods struggle to reconstruct these complex scenes completely, whereas our method handles this situation effectively. Considering the detail part, VoxFormer produces the scene with radial artifacts, while" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 451, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 523 + ], + "type": "text", + "content": "Symphonies performs somewhat better in this regard but still has many cluttered areas due to occlusions. Our reconstructed results not only avoid these issues but also produce clearer boundaries between different semantics. These comparisons demonstrate the superiority of our method both global and local scale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce SGFormer, the first satellite-ground cooperative approach with a tightly coupled dual-branch framework design for urban semantic scene completion. SGFormer effectively fuses observations from groundview and satellite-view images, enabling joint consideration of global scene layout and local details for occupancy grid prediction. Experimental results show that our method surpasses state-of-the-art camera-based approaches and achieves performance on par with LiDAR-based methods across many metrics. These findings suggest that integrating satellite images into the SSC task offers a cost-effective yet highly promising solution. We hope our work inspires further research in this community." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11936" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 213, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 213, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 278 + ], + "type": "text", + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 2, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 281, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 296, + 335 + ], + "type": "text", + "content": "[2] Sudong Cai, Yulan Guo, Salman Khan, Jiwei Hu, and Gongjian Wen. Ground-to-aerial image geo-localization with a hard exemplar reweighting triplet loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8391-8400, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 337, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 337, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 337, + 296, + 381 + ], + "type": "text", + "content": "[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "type": "text", + "content": "[4] Yue Cao, Mingsheng Long, Jianmin Wang, and Shichen Liu. Deep visual-semantic quantization for efficient image retrieval. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1328–1337, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 429, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 429, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 295, + 473 + ], + "type": "text", + "content": "[5] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In Conference on Robot Learning, pages 2148-2161. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 475, + 296, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 540 + ], + "type": "text", + "content": "[6] Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621-21631, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 543, + 291, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 543, + 291, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 543, + 291, + 554 + ], + "type": "text", + "content": "[7] Google. Google maps api, 2024. Accessed: 2024-11-14. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 555, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 610 + ], + "type": "text", + "content": "[8] Xiyue Guo, Haocheng Peng, Junjie Hu, Hujun Bao, and Guofeng Zhang. From satellite to ground: Satellite assisted visual localization with cross-view semantic matching. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 3977-3983. IEEE, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "text", + "content": "[9] Adrian Hayler, Felix Wimbauer, Dominik Muhle, Christian Rupprecht, and Hongdong Li. Daniel Cremers. S4c: Self-supervised semantic scene completion with neural fields. In 2024 International Conference on 3D Vision (3DV), pages 409-420. IEEE, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 169 + ], + "type": "text", + "content": "This work was partially supported by NSF of China (No. 62425209). The author Junjie Hu acknowledges support from the Guangdong Natural Science Fund under Grant 2024A1515010252. The authors thank Shang Liu for her help in drawing Figures 1, 2, and 3. Special thanks also go to Bo Wen, Xingyuan Yu, Han Zhang, and Ran Zuo for their invaluable support." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[11] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7132-7141, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 554, + 162 + ], + "type": "text", + "content": "[12] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "text", + "content": "[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 9223-9232, 2023. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2405.17429, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "text", + "content": "[15] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20258-20267, 2024. 2, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[16] Peidong Li, Wancheng Shen, Qihao Huang, and Dixiao Cui. Dualbev: Cnn is all you need in view transformation. arXiv preprint arXiv:2403.05402, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "text", + "content": "[17] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, et al. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 487 + ], + "type": "text", + "content": "[18] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 554, + 544 + ], + "type": "text", + "content": "[19] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 545, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 554, + 599 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "type": "text", + "content": "[21] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition, pages 5624-5633, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "type": "text", + "content": "[22] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11937" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Ian D Miller, Anthony Cowley, Ravi Konkimalla, Shreyas S Shivakumar, Ty Nguyen, Trey Smith, Camillo Jose Taylor, and Vijay Kumar. Any way you look at it: Semantic crossview localization and mapping with lidar. IEEE Robotics and Automation Letters, 6(2):2397-2404, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[26] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 303 + ], + "type": "text", + "content": "[27] Paul-Edouard Sarlin, Daniel DeTone, Tsun-Yi Yang, Armen Avetisyan, Julian Straub, Tomasz Malisiewicz, Samuel Rota Bulo, Richard Newcombe, Peter Kontschieder, and Vasileios Balntas. Orienternet: Visual localization in 2d public maps with neural matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21632–21642, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "text", + "content": "[28] Paul-Edouard Sarlin, Eduard Trulls, Marc Pollefeys, Jan Hosang, and Simon Lynen. Snap: Self-supervised neural maps for visual positioning and semantic understanding. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[29] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In Proceedings of the IEEE/cvf Winter Conference on Applications of Computer Vision, pages 2417-2426, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "text", + "content": "[30] Yujiao Shi and Hongdong Li. Beyond cross-view image retrieval: Highly accurate vehicle localization using satellite image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17010-17020, 2022. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1746-1754, 2017. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[32] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114. PMLR, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[33] Xin Tan, Wenbin Wu, Zhiwei Zhang, Chaojie Fan, Yong Peng, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Geocc: Geometrically enhanced 3d occupancy network with implicit-explicit depth fusion and contextual self-supervision. arXiv preprint arXiv:2405.10591, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "text", + "content": "[34] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occu" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 519 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "pancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[36] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3–19, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[37] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17642-17651, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 207, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 262 + ], + "type": "text", + "content": "[38] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101-3109, 2021. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "text", + "content": "[39] Junyan Ye, Qiyan Luo, Jinhua Yu, Huaping Zhong, Zhimeng Zheng, Conghui He, and Weijia Li. Sg-bev: Satellite-guided bev fusion for cross-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27748-27757, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[40] Zhu Yu, Runmin Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Si-Yuan Cao, and Hui-liang Shen. Context and geometry aware voxel transformer for semantic scene completion. In Advances in Neural Information Processing Systems, pages 1531-1555, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "text", + "content": "[41] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "type": "text", + "content": "[42] Yanhao Zhang, Yujiao Shi, Shan Wang, Ankit Vora, Akhil Perincherry, Yongbo Chen, and Hongdong Li. Increasing slam pose accuracy by ground-to-satellite image registration. arXiv preprint arXiv:2404.09169, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 475, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 519 + ], + "type": "text", + "content": "[43] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 4" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11938" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_content_list.json b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..49d7842e675581c9118b43286aa4df31a7986180 --- /dev/null +++ b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_content_list.json @@ -0,0 +1,1646 @@ +[ + { + "type": "text", + "text": "SGSST: Scaling Gaussian Splatting Style Transfer", + "text_level": 1, + "bbox": [ + 243, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bruno Galerne $^{1,2}$ , Jianling Wang $^{1}$ , Lara Raad $^{3}$ , and Jean-Michel Morel $^{4}$ $^{1}$ Université d'Orléans, Université de Tours, CNRS, IDP, UMR 7013, Orléans, France \n $^{2}$ Institut Universitaire de France (IUF)", + "bbox": [ + 225, + 180, + 792, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Instituto de Ingeniería Eléctrica, Facultad de Ingeniería, Universidad de la Republica $^{4}$ City University of Hong Kong", + "bbox": [ + 112, + 232, + 901, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de06ada5b3d524820e24de71826a3553f25248cccaea80e5dfcf6e9589b636e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 282, + 362, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b71d6ab9ed5269de7ecfc03ccfe4dad6f25f38d035b64203aeac34e6151ba19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 417, + 362, + 553 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95e7be46152d6fe6611224de0ffcde9fc03c40238f151c3470f6ced37493b867.jpg", + "image_caption": [ + "Figure 1. Various ultra-high definition style transfers of a Gaussian splattering 3D scene. SGSST transfers a very large set of global style statistics of an image to a 3DGS scene by minimizing a tailored multiscale SOS loss, yielding 3D style transfer of superior quality and at unprecedented high resolution (images have size $5187 \\times 3361$ )." + ], + "image_footnote": [], + "bbox": [ + 94, + 554, + 362, + 688 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d77b5a44a8876ebe874e29b631e9e3fd81680a9beda2be069e017c7ac6ed352c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 282, + 632, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47abad2cd38069a40dbe926a18e5d9ffa25d9d29c54c101edd22daeed9d03e8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 417, + 632, + 553 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b801c471712eeb801a57ac683844d4835f174f5d9811af696b5dae513c602e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 554, + 632, + 688 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53dfb53fae69cbf3dbfaab5f83bde197be0e5b7b0c57c566d8f53c338af013d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 282, + 901, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/529a8b274e630287f10f7bff4e84d7e7e3594158eac76e78ddc1220fc5e522b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 417, + 901, + 553 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66774d176e9ea7e04e597358a655248b6e4e5281adf444020159f9f62c775fab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 554, + 901, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 755, + 326, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Applying style transfer to a full 3D environment is a challenging task that has seen many developments since the advent of neural rendering. 3D Gaussian splatting (3DGS) has recently pushed further many limits of neural rendering in terms of training speed and reconstruction quality. This work introduces SGSST: Scaling Gaussian Splatting Style Transfer, an optimization-based method to apply style", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "transfer to pretrained 3DGS scenes. We demonstrate that a new multiscale loss based on global neural statistics, that we name SOS for Simultaneously Optimized Scales, enables style transfer to ultra-high resolution 3D scenes. Not only SGSST pioneers 3D scene style transfer at such high image resolutions, it also produces superior visual quality as assessed by thorough qualitative, quantitative and perceptual comparisons.", + "bbox": [ + 511, + 757, + 906, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26535", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dealing with ultra-high resolution (UHR) rendering is capital for AR/VR applications. Indeed, when navigating in a 3D environment the user only sees a partial field of view of the environment. This adds challenging issues for applying style transfer to 3D environments, that is, transferring the visual characteristics of an image such as a painting to a 3D scene. Indeed, while the general scene should convey the global color palette of the style painting, when getting closer to objects in a stylized environment the user should expect to see fine painting patterns such as brushstrokes. However, in the current state of the art, the user quickly encounters resolution-limited content in the form of blurry interpolated features.", + "bbox": [ + 89, + 116, + 480, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ever since the seminal work of Mildenhall et al. [28], neural radiance fields (NeRF) have seen many improvements. Several 3D scene representations have been proposed for improving the quality and resolution of the reconstructed scenes as well as easing the training, e.g. [2, 3, 9, 29]. The recent 3D Gaussian splatting (3DGS) [20] has introduced an efficient high-resolution (HR) scene representation which has stirred much interest [45]. Both frameworks have stirred attempts to 3D style transfer algorithms [6, 16, 22, 25, 26, 33, 46, 48], but these methods have so far produced medium resolution outputs. They do not faithfully transport high resolution multiscale textures such as those present in paintings. Motivated by a recent neural style transfer (NST) [11] contribution tailored for UHR images [10], i.e. with resolution larger than $4\\mathrm{k}$ images, we show that 3DGS can be leveraged for UHR style transfer.", + "bbox": [ + 89, + 314, + 482, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are the following:", + "bbox": [ + 109, + 556, + 434, + 571 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SOS, a Simultaneously Optimized Scales loss expressed in a single parameterless and explainable formula.", + "- By solely optimizing the SOS loss, we reach UHR for 3DGS style transfer and we scale Gaussian Splatting Style Transfer by a four times resolution gain.", + "- Superior quality transfer: By transferring a very large set of global style statistics, we obtain superior style transfer quality even at HR resolution, as confirmed by a comparative perceptual study." + ], + "bbox": [ + 91, + 577, + 482, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In short, our approach is the first method that allows UHR style transfer directly to 3DGS. It produces high visual quality results and relies on optimizing a single multiscale loss. The simplicity of our approach ensures its reproducibility. Being optimization-based, SGSST's main limitation is a fairly large training time that is two to eight times longer than the initial 3DGS training depending on the image resolution. Yet, considering the high quality of the results and their reproducibility, this contribution is valuable for AR/VR applications that require high visual quality for their user experience.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 89, + 651, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural style transfer In the seminal work of Gatys et al. [11] NST is formulated as an optimization problem minimizing the distances between Gram matrices of VGG [37] features. Even though other VGG statistics have been considered, almost all subsequent style transfer and texture synthesis methods rely on VGG [8, 13, 14, 27, 32, 35, 42]. To accelerate style transfer, several methods [18, 40, 41] have attempted to train feed-forward networks approximately minimizing the Gram loss [11]. However, these approximate methods require learning a new model for each style type. This latter limitation has been addressed by fast Universal Style Transfer (UST) approaches [4, 7, 15, 23, 24, 30, 36] that use VGG feature decoders.", + "bbox": [ + 511, + 114, + 903, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For HR images, coarse-to-fine multiscale strategies [12, 13, 38] have proved effective, but still face limitations due the high GPU memory usage of VGG statistics. Fast HR alternatives [1, 5, 39, 43, 44] do exist but generally suffer from artifacts and struggle to capture the full style complexity. Recently, SPST [10] proposed an implementation of the Gatys et al. method adapted to UHR (larger that $4\\mathrm{k}$ ) images. The visual quality of SPST's results is superior, at the cost of a long optimization time.", + "bbox": [ + 511, + 311, + 903, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Style transfer for neural radiance fields NeRFs [28] have completely redefined the field of 3D scene modeling and novel view synthesis. Editing the visual aspect of NeRFs via style transfer has quickly been addressed [6, 16, 25, 48], usually by fine tuning a pretrained NeRF representation using a style transfer loss, or training an additional fast style transfer module. ARF [48] is a notable exception: It uses Nearest Neighbor Feature Matching (NNFM) for fine tuning a plenoxel radiance field [9], produces high-quality results at moderate resolution, and is the base model for other methods [19, 47]. While these works paved the way for radiance field style transfer, they are all limited in input and output image resolution.", + "bbox": [ + 511, + 470, + 903, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Style transfer for Gaussian splatting A few recent contributions show that 3DGS is a promising framework for 3D scene style transfer. Saroha et al [33] propose a solution for universal style transfer of a given 3DGS scene. The method processes the colors of Gaussians with a tiny MLP trained using fast AdaIN [15] and relying on a multiresolution hash grid representation [29]. StyleGaussian [26] is a concurrent approach that relies on transferring encoding of VGG features to each Gaussian and applying AdaIN to these features. The new Gaussian features are then decoded into an RGB color by processing the K-nearest neighbors of the Gaussians. After training, both methods allow for instantaneous stylization with any style image, but the visual quality of the results is quite low, since high-resolution", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26536", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "details such as brushstroke patterns are not transferred.", + "bbox": [ + 89, + 90, + 454, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StylizedGS [46] is an optimization-based method that extends ARF [48] to 3DGS. It uses a training loss made of six terms combined with a color transfer preprocessing, the loss being designed to minimize changes in the 3DGS geometry while letting the style evolve via VGG NN matching. $\\mathcal{G}$ -Style [22] also is an optimization based approach that further uses a CLIP-based loss [31]. While these approaches produce slightly better results than ARF [48], both methods are limited in content and style resolution due to the use of nearest-neighbor matching of VGG features.", + "bbox": [ + 89, + 106, + 483, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "None of the current state of the art deals with HR style transfer. The stylization is either fast but very approximate due to AdaIn [26, 33] or unable to use HR style images due to NNFM [22, 46, 48]. To the best of our knowledge, SGSST is the first procedure allowing high-quality transfer for 3DGS that is trained and rendered at UHR.", + "bbox": [ + 89, + 257, + 483, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 361, + 230, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Gaussian splatting representation", + "text_level": 1, + "bbox": [ + 89, + 386, + 385, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Starting from a multiview training set of images $\\{u_i\\}_{i=1}^{N_{\\mathrm{views}}}$ of a static scene accompanied with corresponding calibrated cameras $\\{\\mathcal{C}_i\\}_{i=1}^{N_{\\mathrm{views}}}$ computed by structure from motion [34], the 3DGS algorithm [20] trains a set of colored 3D Gaussians $\\{\\mathcal{G}_j\\}_{j=1}^{N_{\\mathrm{Gaussians}}}$ so that they represent the 3D scene from any camera position. Each Gaussian $\\mathcal{G}_j$ is represented by a finite set of features: a center position $\\mu_j$ , a covariance matrix $\\Sigma_j$ (encoded by a scaling diagonal matrix and a rotation matrix), an opacity $\\alpha_j$ and a view-dependent color function $c_j$ . These parameters are used in a volumetric renderer that determines the color by summing the contribution of each Gaussian that intersects a ray with direction $(\\theta, \\varphi)$ via $\\alpha$ -blending (see e.g. [28]). The contribution of a Gaussian is its color $c_j(\\theta, \\varphi)$ multiplied by an opacity $\\sigma_j$ defined as the maximal opacity $\\alpha_j$ times the unnormalized Gaussian density at the ray position [20]. Thus, the resulting color $C$ for the ray is", + "bbox": [ + 89, + 407, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {j = 1} ^ {N} c _ {j} (\\theta , \\varphi) \\sigma_ {j} \\prod_ {k = 1} ^ {j - 1} (1 - \\sigma_ {k}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 676, + 483, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the sum is over all Gaussians intersecting the ray. The color function $c_{j}(\\theta ,\\varphi)$ associated with a Gaussian depends on the spherical direction $(\\theta ,\\varphi)$ through an order 3 spherical harmonics polynomial function given by", + "bbox": [ + 89, + 729, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc _ {j} (\\theta , \\varphi) = c _ {j, 0} + \\sum_ {\\ell = 1} ^ {3} \\sum_ {m = 1} ^ {2 \\ell + 1} c _ {j, \\ell , m} Y _ {\\ell , m} (\\theta , \\varphi) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 801, + 483, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the vectors $c_{j,0}$ and $c_{j,\\ell,m}$ are in $\\mathbb{R}^3$ and the $Y_{\\ell,m}$ form a basis of the spherical harmonics polynomials of degree $\\ell$ . In short, $c_{j,0} \\in \\mathbb{R}^3$ is the main color and the ad", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ditional coefficients $c_{j,\\ell,m}$ encode smooth variations of this color when the viewing angle changes.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Like for NeRF, the key ingredient of the 3DGS parametrization is the differentiable rendering function $\\mathcal{R}(\\mathcal{C};\\Theta)$ that renders a view of the scene for any camera $\\mathcal{C}$ given the scene parameters $\\Theta = \\{(\\mu_j,\\Sigma_j,\\alpha_j,c_{j,0},(c_{j,\\ell ,m})_{\\ell ,m})\\}_{j = 1}^{N_{\\mathrm{Gaussians}}}$ . This differentiable rendering function allows to train the Gaussian parameters to minimize the reconstruction error", + "bbox": [ + 511, + 121, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\Theta} \\frac {1}{N _ {\\mathrm {v i e w s}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v i e w s}}} E _ {\\mathrm {r e c o n s t r u c t i o n}} (\\mathcal {R} (\\mathcal {C} _ {i}; \\Theta); u _ {i}) \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 237, + 906, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{reconstruction}}$ is a 2D image comparison error (such as a combination of mean square error and SSIM [20]). The minimization is conducted using Adam [21] for 30k iterations by randomly picking one view at each iteration.", + "bbox": [ + 511, + 289, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Style transfer loss for UHR images", + "text_level": 1, + "bbox": [ + 511, + 359, + 815, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach relies on optimizing VGG19 [37] feature statistics as initially proposed by Gatys et al. [11]. Content consistency is loosely monitored by preservation of the feature layer $L_{\\mathrm{c}} = \\mathrm{ReLU\\_4\\_2}$ while style transfer is imposed by matching spatial statistics of five VGG19 layers, namely the set $\\mathcal{L}_{\\mathrm{s}} = \\{\\mathrm{ReLU\\_k\\_1}, k \\in \\{1, 2, 3, 4, 5\\}\\}$ . The statistics of interest of the feature response $V^{L}(w)$ of an image $w$ at some layer $L \\in \\mathcal{L}_{\\mathrm{s}}$ having $n_{\\mathrm{c}}^{L}$ feature channels are its Gram matrix $\\operatorname{Gram}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L} \\times n_{\\mathrm{c}}^{L}}$ , its spatial mean vector $\\operatorname{mean}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L}}$ , and its standard deviation vector $\\operatorname{std}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L}}$ . Given a content image $u$ and a style image $v$ , we consider the loss function", + "bbox": [ + 511, + 382, + 908, + 570 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {t r a n s f e r}} (x; u, v) = E _ {\\text {c o n t e n t}} (x; u) + E _ {\\text {s t y l e}} (x; v) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 582, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{content}}(x;u) = \\lambda_{\\mathrm{c}}\\left\\| V^{L_{\\mathrm{c}}}(\\boldsymbol {x}) - V^{L_{\\mathrm{c}}}(\\boldsymbol {u})\\right\\|^{2}$ , with $\\lambda_{\\mathrm{c}} > 0$ and the style loss is defined by", + "bbox": [ + 511, + 611, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {s t y l e}} (x; v) = \\sum_ {L \\in \\mathcal {L} _ {\\mathrm {s}}} E _ {\\text {s t y l e}} ^ {L} (x; v). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 656, + 906, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{style}}^{L}(x;v)$ is a linear combination of the mean square error between the VGG19 statistics of $V^{L}(x)$ and the one of the style features $V^{L}(v)$ [10]. While only the Gram matrices were originally used [11], it has been shown that adding control for the mean and standard deviation corrects some style transfer color artefacts [10] previously identified in the literature [14, 32, 35].", + "bbox": [ + 511, + 700, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain high-quality style transfer for HR images one needs to optimize the style transfer loss using several scales and a coarse-to-fine approach [12]. Indeed, if one applies style transfer on the highest resolution only, the changes within the content image are limited to local texture and the results does not convey a painting aspect. Due to the use of", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26537", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/413a883f11bd3798e04372d1cb863202ca3a20ae935a9c76d1f4592d6658bb0b.jpg", + "image_caption": [ + "Figure 2. Overview of SGSST. Starting from a pretrained realistic 3DGS scene [20], we optimize the colors of each Gaussian using the new multiscale SOS loss (involving $n_{\\mathrm{s}} = 3$ scales in the illustration). Computing the gradient w.r.t. the loss is feasible for UHR images thanks to the SPST partition-based implementation [10]. Multiscale gradient stacking is used at the node of the rendered image to perform only one backpropagation per iteration through the 3DGS rendering pipeline." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 908, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VGG19 features, computing the loss $E_{\\mathrm{transfer}}(x;u,v)$ and its gradient with respect to (w.r.t.) $x$ via backpropagation is memory prohibiting for UHR images. However, using a grid partition and local loss backpropagation based on precomputed global statistics, SPST [10] allows for the exact evaluation of this gradient.", + "bbox": [ + 88, + 473, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Scaling Gaussian splatting style transfer", + "text_level": 1, + "bbox": [ + 89, + 580, + 452, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A complete overview of our SGSST algorithm is given in Figure 2. Starting from a realistic 3DGS representation [20], we optimize the colors of each Gaussian using a multiscale style transfer loss.", + "bbox": [ + 89, + 606, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Stylizing the 3DGS representation", + "text_level": 1, + "bbox": [ + 89, + 679, + 390, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By changing the reconstruction loss of Equation (3) with a style transfer loss for the input style image $v$ , one can hope to stylize a realistic 3DGS. However, given the complexity of the 3DGS representation and the many interacting parameters, it is not such an easy task to alter the 3DGS aspects without loosing the content geometry. Our solution is to only optimize for the constant color components $\\Theta_{\\mathrm{color}} = \\{c_{j,0}\\}_{j=1}^{N_{\\mathrm{Gaussians}}}$ of the Gaussians and simply freeze all the other parameters $\\Theta_{\\mathrm{init}}$ given by the initial realistic 3DGS training.", + "bbox": [ + 89, + 703, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We experimentally found that this robust approach ensures a rich style transfer and fully preserves the 3D geometry. Indeed, fixing all the Gaussian parameters except for", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the main color component $c_{j,0}$ preserves the scene geometry, as the location and size of the Gaussians are being kept (see Section 5.3 for ablation experiments).", + "bbox": [ + 511, + 474, + 906, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Multiscale Style Transfer Loss", + "text_level": 1, + "bbox": [ + 511, + 527, + 782, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the Simultaneously Optimized Scales (SOS) loss defined as", + "bbox": [ + 511, + 550, + 905, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {S O S}} (x; u, v) = \\frac {1}{n _ {\\mathrm {s}}} \\sum_ {s = 0} ^ {n _ {\\mathrm {s}} - 1} E _ {\\text {t r a n s f e r}} \\left(x ^ {\\downarrow 2 ^ {s}}; u ^ {\\downarrow 2 ^ {s}}, v ^ {\\downarrow 2 ^ {s}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 585, + 905, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n_{\\mathrm{s}} \\geq 1$ is the number of considered scales and $u^{\\downarrow 2^{s}}$ denotes an image $u$ downscaled by a $2^{s}$ factor. This SOS loss (6) enables style transfer simultaneously at all scales. A somewhat similar approach was proposed for 2D texture synthesis [38] but, as already mentioned, multiscale 2D style transfer is generally conducted using a coarse-to-fine strategy [12]. However, for 3DGS we observed that the initial configuration had only little influence on the final result, making the coarse-to-fine strategy ineffective for multiscale style transfer (See Section 5.3).", + "bbox": [ + 511, + 633, + 905, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the end, as illustrated by Figure 2, the stylization of UHR 3DGS is conducted by solving for", + "bbox": [ + 511, + 786, + 906, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\Theta_ {\\text {c o l o r}}} \\frac {1}{N _ {\\text {v i e w s}}} \\sum_ {i = 1} ^ {N _ {\\text {v i e w s}}} E _ {\\text {S O S}} \\left(\\mathcal {R} \\left(\\mathcal {C} _ {i}; \\Theta\\right); u _ {i}, v\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 823, + 905, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Theta$ stands for the 3DGS parameters obtained by replacing the color components of $\\Theta_{\\mathrm{init}}$ by the values of the", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26538", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eeec0944ae3ccc579a39d2848e9d071444014c7099aefb6325676e47d2aa5859.jpg", + "image_caption": [ + "Figure 3. UHR 3DGS style transfer. SGSST allows for the multiscale style transfer of 3DGS scenes at UHR. From left to right: Style image, one UHR stylized view, three magnified details, and evolution of the SOS loss and each style transfer loss that contributes to it. We first optimize the transfer loss for the coarsest scale (yellow curve) for $10\\mathrm{k}$ iterations and then optimize for another $10\\mathrm{k}$ iterations the SOS loss (light blue curve), namely the mean of the four transfer losses. Images sizes are $5187\\times 3361$ for content and $4230\\times 3361$ for style." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 294, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/da26b8c99258f4ebc42db88da241f81055755c6edb98f152dc0383612e7c17ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 88, + 537, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a763e2db65789129d44dea624b26750219869325e16f115b9491821249ee1fd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 88, + 684, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9d3b5195b95b89d4bafac744b4c4b51800ba3bfe578d03a566bc494ed431d284.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 88, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "optimization variable $\\Theta_{\\mathrm{color}}$", + "bbox": [ + 89, + 306, + 282, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 330, + 305, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multiscale gradient stacking The 2D SPST method [10] provides the gradient of each term $E_{\\mathrm{transfer}}(x^{\\downarrow 2^{s}};u^{\\downarrow 2^{s}},v^{\\downarrow 2^{s}})$ w.r.t. to the input $x^{\\downarrow 2^{s}} = \\mathcal{R}(\\mathcal{C}_{i};\\Theta)^{\\downarrow 2^{s}}$ . We first backpropagate each of these gradients through the downscaling operator and stack them at the level of the rendered image $\\mathcal{R}(\\mathcal{C}_i;\\Theta)$ . When each of the $n_{\\mathrm{s}}$ scales has been treated, the stacked gradient is equal to the gradient of the full loss $E_{\\mathrm{SOS}}(\\mathcal{R}(\\mathcal{C}_i;\\Theta);u_i,v)$ w.r.t. to the rendered image $\\mathcal{R}(\\mathcal{C}_i;\\Theta)$ . This gradient is then backpropagated through the Gaussian rendering pipeline to obtain the gradient w.r.t. the Gaussian colors $\\Theta_{\\mathrm{color}}$ (see both orange arrows in Figure 2). Using this strategy we only backpropagate one time per iteration through the 3DGS rendering pipeline instead of $n_{\\mathrm{s}}$ times.", + "bbox": [ + 89, + 354, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Color transfer via style transfer at the coarsest scale To speed up the color transfer, we first optimize for 10k Adam iterations with the loss restricted to the coarsest scale $E_{\\mathrm{transfer}}(\\mathcal{R}(\\mathcal{C}_i;\\Theta)^{\\downarrow 2^s};u_i^{\\downarrow 2^s},v^{\\downarrow 2^s})$ with $s = n_{\\mathrm{s}} - 1$ , then optimize the SOS loss for another 10k Adam iterations.", + "bbox": [ + 89, + 584, + 483, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Number of scales $n_{\\mathrm{s}}$ is set automatically to use all available scales, the coarsest resolution having sides larger than 256 for VGG19 statistics to be reliable.", + "bbox": [ + 89, + 676, + 483, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reproducibility All our experiments have been conducted using the same SOS loss and training procedure, making our approach parameterless and fully reproducible. Our public PyTorch implementation is based on the public source codes1 for 3DGS [20] training and SPST [10]. Our code and videos of our results are available online2.", + "bbox": [ + 89, + 741, + 482, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1https://github.com/graphdeco-inria/gaussian-splatting; https://github.com/bgalerne/scaling_PA painting_style_transfer 2Code: https://github.com/JianlingWANG2021/SGSST; Videos: https://www.idpoisson.fr/galerne/sgsst/", + "bbox": [ + 89, + 838, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 304, + 645, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Ultra-high resolution results", + "text_level": 1, + "bbox": [ + 511, + 330, + 767, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our multiscale stylization algorithm is able to transfer style details at UHR for both the content image resolution and the style image resolution. This results in unprecedentedly rich style transfer, as illustrated by Figure 3. As can be observed, minimizing the SOS loss indeed allows to decrease the style transfer loss for all scales (Figure 3 right). The approach is especially relevant when transferring the style of an UHR painting presenting style features at several scales, ranging from a specific color palette to a main curve style and to fine brushstrokes or canvas texture (see close-up views of Figure 3 and the first two lines of Figure 1). To obtain such results, combining style transfer at the largest possible number of scales is critical (see ablation in supp. mat. Figure 9). In addition, the approach is also efficient to transfer the style of a medium resolution style image to an UHR scene, as shown in the last row of Figure 1.", + "bbox": [ + 511, + 354, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To the best of our knowledge, our approach is the first to work at UHR resolution for neural style transfer of neurally rendered 3D scenes. One of the main advantages of the Gram-based loss is that it does not depend on the style resolution and, when both the content and style images grow in $O(N)$ , its complexity scales in $O(N)$ while NNFM used in ARF [48] scales in $O(N^2)$ .", + "bbox": [ + 511, + 598, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Yet, applying style transfer at such resolutions remains computationally heavy: for the garden image of size $5187 \\times 3361$ (Figures 1 and 3) the style transfer takes 25.5 hours (VS 3 hours to train the initial 3DGS model), that is an 8.5 overhead factor. For an image of moderate size (Figure 4 left and middle) the style transfer and the initial 3DGS training take respectively 22 and 10 minutes, that is, the overhead factor is only 2.2. These time values were obtained using a single A100 GPU with 80 GB of memory and could be accelerated by adapting the SOS loss implementation to a multi-GPU setting. Note that high computation times are inherent to UHR style transfer: Running the 2D SPST method for the 185 garden training images takes 27", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26539", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/45042e9c71b475d242850656464a4cd131be0725f8932bc821865ce62cc5c9b8.jpg", + "image_caption": [ + "Figure 4. Comparison of SGSST (ours, top) with StyleGaussian [26] (middle) and ARF [48] (bottom). From left to right the content resolutions are $980 \\times 545$ (train), $979 \\times 546$ (truck), and $3115 \\times 2076$ (counter). For the first two examples, the various outputs keep the resolution of the content, but for the HR counter scene, the output sizes are $3115 \\times 2076$ for SGSST, $1600 \\times 1066$ for StyleGaussian and $779 \\times 519$ for ARF (see supp. mat. Figure 28 for ARF results without downscaling). Thanks to its multiscale global VGG statistics, SGSST is the most faithful method regarding style consistency." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hours (8.9 min. per image).", + "bbox": [ + 89, + 441, + 274, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison", + "text_level": 1, + "bbox": [ + 89, + 468, + 222, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform a thorough comparative study using 40 3D style transfer experiments using 9 different scenes and various style images (see supp. mat.). We compare our results with the NeRF-based ARF [48] and the 3DGS-based StyleGaussian [26] algorithms, described in Section 2, using their public implementations3.", + "bbox": [ + 89, + 492, + 482, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing style transfer methods is challenging because each algorithm treats the style image differently. Following high quality 2D style transfer [10, 12], our loss uses up to four scales and each scale uses five VGG layers. The UHR style images were downscaled so that the style has the same size as the content images (no upscale was applied if the style image is smaller than the content image). In contrast, ARF uses a single VGG layer from a single scale and the style image is downscaled to be twice smaller than the content input, resulting in smaller local texture. StyleGaussian reduces the style images so that it has the size $256 \\times 256$ , a scale that is hardly sufficient to represent HR style images such as paintings.", + "bbox": [ + 89, + 584, + 482, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison Figure 4 shows three different comparative experiments. As one can observe, the results of StyleGaussian [26] are generally not satisfying since the method fails to transfer local texture or to preserve the style", + "bbox": [ + 89, + 801, + 482, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image's color palette. ARF results better reproduce brushstroke textures, but the transfer is limited, as the method only involves a single VGG layer at a single scale. Also, NNFM does not ensure the preservation of a global color palette. In comparison, SGSST preserves the style at all considered scales. This results in color palette preservation, as well as a verifiable transfer of features of any size, from large brushstrokes (Figure 4 middle) to local grain transfer (Figure 4 right). In addition, SGSST is the only method that performs style transfer at the original resolution of the HR example of Figure 4 right.", + "bbox": [ + 511, + 441, + 906, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison Even though there is no consensus for the quantitative evaluation of NST [17], following previous works, we report two different metrics for style transfer quality and texture consistency across views. Style transfer quality can be measured by the Gram loss [11]. A second metric proper to NeRF style transfer [25, 26, 33] is to check the short-term and long-term consistency of the radiance fields in terms of LPIPS [49] and RMSE between wrapped views. Since SGSST is the only method working with UHR images, when necessary we forcefully downgraded all the results to the resolution of ARF for comparison. The average of these two metrics over our 40 experiments is reported in Table 1. The Gram loss is the best for SGSST and, surprisingly, StyleGaussian achieves a lower Gram loss than ARF, which is not consistent with the qualitative evaluation, probably explained by ARF using a single VGG layer compared to five for the two other methods. In terms of consistency metrics, StyleGaussian reports to be", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://github.com/Kai-46/ARF-svox2; https://github.com/Kunhao-Liu/StyleGaussian", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26540", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d20ea9173a183d6fc974d97b87f78cbb93ccd8810f13469632468312b4f9df56.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTransfer qualityShort-range consistencyLong-range consistency
Gram↓LPIPS↓RMSE↓LPIPS↓RMSE↓
SGSST2.59e80.0300.0320.0550.063
StyleGaussian4.61e80.0330.0290.0500.056
ARF5.77e80.0400.0370.0720.066
", + "bbox": [ + 91, + 88, + 486, + 191 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ddcb1c10da5b49da43f47dd85d16c5a339e78e50f81cad2e316316005c2cfae.jpg", + "table_caption": [ + "Table 1. Quantitative comparison. Style transfer quality and texture consistency metrics averaged over 40 experiments for SGSST, StyleGaussian, and ARF. Best results in bold, second best underlined." + ], + "table_footnote": [], + "table_body": "
SGSSTARFStyleGaussian
Voting results (%)66.331.62.1
", + "bbox": [ + 93, + 271, + 480, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Perceptual study. Summary of the 680 votes for the most style consistent result.", + "bbox": [ + 89, + 325, + 483, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the most stable approach followed by SGSST but note that this metric favors the lack of texture.", + "bbox": [ + 89, + 381, + 483, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perceptual study To further support our results, we conducted a perceptual study comparing the 40 3D style transfer experiments. For each example, four images were displayed (at a resolution of $1280 \\times 720$ ): the style image and, in a blind random order, the results of the three methods (SGSST, ARF and StyleGaussian) shown at a common viewpoint (also randomly selected). Each participant was shown ten random instances and was asked to select the result that was the most faithful to the style image. The study was conducted on the web with volunteer participants.", + "bbox": [ + 88, + 429, + 482, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A total of 68 participants took part in the test, resulting in 680 votes summarized in Table 2. This perceptual study shows that the fast style transfer performed by Style-Gaussian is consistently judged inferior in quality over ARF and SGSST. It also confirms that SGSST is far superior to ARF in terms of visual quality since $66\\%$ of the participants ranked our method first for its painting style transfer quality, even when presented with results downscaled in resolution.", + "bbox": [ + 89, + 580, + 483, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 710, + 241, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of optimization parameters Our SGSST algorithm stylizes a realistic 3DGS scene by tuning the constant color components of the 3DGS Gaussians using a single 2D style transfer loss function at multiple scales (Equation (6)). In contrast, Zhang et al. [46] optimize all 3DGS parameters. This, however, necessitates a complex loss made of six different terms to avoid artefacts: a loss term enforces consistency with the original geometry via depth preservation while a preprocessing of Gaussians floaters is necessary after color transfer.", + "bbox": [ + 89, + 734, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 shows that optimizing more parameters of the", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f107ed0c047f8a35e16ef9fde4dab2e0a14903d4717e44bf630c4099c18ff780.jpg", + "image_caption": [ + "Figure 5. Influence of optimization parameters: Allowing more 3DGS parameters to be optimized when minimizing the SOS loss does not improve the stylization quality and can dramatically impact the geometry. From left to right: Style image, content, SGSST default (optimization of colors), results when optimizing all spherical harmonics, results when optimizing all parameters." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3DGS for SGSST brings no benefit. Optimizing all the spherical harmonic coefficients does not improve the result, and letting all the parameters free like in [22, 46] leads to a strong degradation of the geometry. Note that two instantaneous 3DGS style transfer methods [26, 33] are based on modifying the color features via some neural networks, but their results are not comparable to optimization-based approaches in terms of visual quality.", + "bbox": [ + 511, + 531, + 906, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Failure of coarse-to-fine strategy Our approach minimizes a style transfer loss simultaneously at all scales. This is different from the coarse-to-fine style transfer strategy that has proven successful for HR style transfer [12]. Figure 6 illustrates that this coarse-to-fine strategy fails in the context of 3DGS. Indeed, the gain obtained by optimizing at a given scale is quickly lost when optimizing the next one leading to the disappearance of large scale features. This can be explained by the fact that the 3DGS representation is a more constrained representation than pixel grids due to its sparsity.", + "bbox": [ + 511, + 670, + 908, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Color transfer via style transfer at the coarsest scale As described in Section 4.3, we first optimize the style transfer loss at the lowest resolution for $10\\mathrm{k}$ iterations and", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/062e384cd99470f7558e5c2c0ef740b56f6bd1c788610ba0d90d2e86f6e9e061.jpg", + "image_caption": [ + "Figure 6. Failure of coarse-to-fine strategy. Results of coarse-to-fine style transfer for the example of Figure 3. Each scale is initialized with the output of the previous scale. As shown by the evolution of the losses (right), when training at a new scale, the loss of the previous scales increases quickly. This explains why the large scale painting features disappear progressively and are absent after training the target UHR (see close-up details)." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09bc5f84e0d52196052696820f1705d77c8d398ab2f4fff84605218fbd5032cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 316, + 220, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29067caafabea0c53c0defeaad7eccb2353db0f6e9ed76927f4d43e387edbcc5.jpg", + "image_caption": [ + "Figure 7. Color transfer via style transfer at the coarsest scale. The first 10k iterations at the coarsest scale allow for quick color changes in the 3DGS scene. Removing this step may lead to color artifacts in the result (top) compared to our default two-step optimization (bottom)." + ], + "image_footnote": [], + "bbox": [ + 96, + 392, + 220, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c0ca8096b488dd408aadf4bf8d7f2defb40ed52ccc8ae0c17d2da4bfeeaa775.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 316, + 349, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5c13c406e9db9d6c6c87f37ab1c0315ca5d7cf4a018be6dd2c0be85364eccee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 392, + 351, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5b3ec75dc83ffdc37180a102b13a153c1c804b5350d76fe19efd335a245b060.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 316, + 478, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2670d0503eb4b5450b93caa17d0bdc62f0503ba8dab3d775b77e2f4b8269779d.jpg", + "image_caption": [ + "Figure 8. Limited texture representation due to low Gaussian density of the initial 3DGS. From left to right: Original scene, 3DGS reconstruction, stylization of the 3DGS scene." + ], + "image_footnote": [], + "bbox": [ + 351, + 392, + 480, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "then optimize the SOS loss for another 10k iterations. Figure 7 illustrates that these first iterations are necessary for a faithful style color palette reproduction.", + "bbox": [ + 89, + 575, + 483, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and limitations", + "text_level": 1, + "bbox": [ + 89, + 637, + 334, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Texture representation The texture representation within a 3DGS scene depends on the density of Gaussians and may be limited in low density areas. Isolated Gaussians can sometimes be spotted as illustrated by Figure 8.", + "bbox": [ + 89, + 662, + 483, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Large computation time Depending on the resolution, SGSST requires from several minutes to several hours of computation. On the other hand, fast 3DGS stylization approaches [26, 33] do not reach a satisfactory visual quality.", + "bbox": [ + 89, + 743, + 483, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Content-style mismatch As said earlier, the Gram loss has the advantage of being independent of the style's image resolution. Also, it enables a faithful transfer of global statistics of the style image, such as its color palette. This important feature is, nevertheless, counterproductive when", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3643e4be91fb4f374d2ed4bfa5dc376d53a768566e6d90a7b81562192ea5d68b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 316, + 903, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the style and content images strongly mismatch, leading to color bleeding or texturing of flat areas. Other controls can be added to mitigate these artefacts [12] and it was shown that these controls are effective for 3DGS scenes [46].", + "bbox": [ + 511, + 541, + 906, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 614, + 633, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work we presented SGSST, a method that, for the first time, enables UHR 3DGS style transfer. To that aim, among other innovations, we introduced the simultaneously optimized scales (SOS) loss. Our qualitative, quantitative and perceptual studies show that SGSST obtains superior style transfer quality than state of the art, even after reducing our results' resolution for a fair comparison with methods that do not reach UHR. Such high quality UHR results necessitate a large computation time that, nevertheless, remains comparable with that of UHR 3DGS training.", + "bbox": [ + 511, + 640, + 906, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work opens the way to several research directions. A first challenge is to produce equally high quality style transfer with a faster algorithm based on UST. A second more exploratory direction is to investigate geometry style transfer for 3DGS by designing adapted regularization to avoid the caveats depicted by Figure 5.", + "bbox": [ + 511, + 792, + 908, + 882 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26542", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: B. Galerne and L. Raad acknowledge the support of the project MISTIC (ANR-19-CE40-005).", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 138, + 187, + 155 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jie An, Tao Li, Haozhi Huang, Li Shen, Xuan Wang, Yongyi Tang, Jinwen Ma, Wei Liu, and Jiebo Luo. Real-time universal style transfer on high-resolution images via zero-channel pruning. arXiv preprint arXiv:2006.09029, 2020. 2", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470–5479, 2022. 2", + "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022, pages 333-350, Cham, 2022. Springer Nature Switzerland. 2", + "[4] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2", + "[5] Zhe Chen, Wenhai Wang, Enze Xie, Tong Lu, and Ping Luo. Towards ultra-resolution neural style transfer via thumbnail instance normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 393-400, 2022. 2", + "[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1475-1484, 2022. 2", + "[7] Tai-Yin Chiu and Danna Gurari. Iterative feature transformation for fast and versatile universal style transfer. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2", + "[8] Valentin De Bortoli, Agnès Desolneux, Alain Durmus, Bruno Galerne, and Arthur Leclaire. Maximum entropy methods for texture synthesis: Theory and practice. SIAM Journal on Mathematics of Data Science, 3(1):52-82, 2021. 2", + "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2", + "[10] Bruno Galerne, Lara Raad, José Lezama, and Jean-Michel Morel. Scaling Painting Style Transfer. Computer Graphics Forum, 2024. 2, 3, 4, 5, 6", + "[11] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 2, 3, 6", + "[12] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3, 4, 6, 7, 8" + ], + "bbox": [ + 93, + 164, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Nicolas Gonthier, Yann Gousseau, and Saïd Ladjal. High-resolution neural texture synthesis with long-range constraints. Journal of Mathematical Imaging and Vision, 64 (5):478-492, 2022. 2", + "[14] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9412-9420, 2021. 2, 3", + "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[16] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18342–18352, 2022. 2", + "[17] Eleftherios Ioannou and Steve Maddock. Evaluation in neural style transfer: A review. Computer Graphics Forum, 43 (6):e15165, 2024. 6", + "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711, 2016. 2", + "[19] Hyunyoung Jung, Seonghyeon Nam, Nikolaos Sarafianos, Sungwoo Yoo, Alexander Sorkine-Hornung, and Rakesh Ranjan. Geometry transfer for stylizing radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8565-8575, 2024. 2", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. 2, 3, 4, 5", + "[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 3", + "[22] Aron Samuel Kovács, Pedro Hermosilla, and Renata G. Rraidou. $\\mathcal{G}$ -style: Stylized gaussian splatting. Computer Graphics Forum, n/a(n/a):e15259, 2024. 2, 3, 7", + "[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3809-3817, 2019. 2", + "[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30:386-396, 2017. 2", + "[25] Kunhao Liu, Fangneng Zhan, Yiwen Chen, Jiahui Zhang, Yingchen Yu, Abdulmotaleb El Saddik, Shijian Lu, and Eric P. Xing. Stylerf: Zero-shot 3d style transfer of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8338-8348, 2023. 2, 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26543", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. StyleGaussian: Instant 3d style transfer with Gaussian splatting, 2024. 2, 3, 6, 7, 8", + "[27] Yang Lu, Song-chun Zhu, and Ying Nian Wu. Learning frame models using cnn filters. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 2", + "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 2, 3", + "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), 2022. 2", + "[30] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[32] Eric Risser, Pierre Wilmot, and Connelly Barnes. Stable and controllable neural texture synthesis and style transfer using histogram losses, 2017. 2, 3", + "[33] Abhishek Saroha, Marija Gladkova, Cecilia Curreli, Tarun Yenamandra, and Daniel Cremers. Gaussian splatting in style, 2024. 2, 3, 6, 7, 8", + "[34] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3", + "[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. Graph., 36(5), 2017. 2, 3", + "[36] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2", + "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2, 3", + "[38] Xavier Snelgrove. High-resolution multi-scale neural texture synthesis. In SIGGRAPH Asia 2017 Technical Briefs, New York, NY, USA, 2017. Association for Computing Machinery. 2, 4", + "[39] Ondrej Texler, David Futschik, Jakub Fišer, Michal Lukáč, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Arbitrary style transfer using neurally-guided patch-based synthesis. Computers & Graphics, 87:62-71, 2020. 2", + "[40] D. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In ICML, pages 1349-1357, 2016. 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[42] Jonathan Vacher, Aida Davila, Adam Kohn, and Ruben Coen-Cagli. Texture interpolation for probing visual perception. In Advances in Neural Information Processing Systems, pages 22146-22157. Curran Associates, Inc., 2020. 2", + "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[44] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. Proceedings of the AAAI Conference on Artificial Intelligence, 37(3): 2742-2750, 2023. 2", + "[45] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2", + "[46] Dingxi Zhang, Zhuoxun Chen, Yu-Jie Yuan, Fang-Lue Zhang, Zhenliang He, Shiguang Shan, and Lin Gao. StylizedGS: Controllable stylization for 3d Gaussian splatting, 2024. 2, 3, 7, 8", + "[47] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. CoARF: Controllable 3D Artistic Style Transfer for Radiance Fields. In 2024 International Conference on 3D Vision (3DV), pages 612-622, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2", + "[48] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. ARF: Artistic radiance fields. In ECCV 2022, 2022. 2, 3, 5, 6", + "[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + ], + "bbox": [ + 516, + 92, + 903, + 641 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26544", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_model.json b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59d7f12d76837b869acabdeaef5283aaf44f6e9c --- /dev/null +++ b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.245, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "SGSST: Scaling Gaussian Splatting Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.181, + 0.793, + 0.232 + ], + "angle": 0, + "content": "Bruno Galerne\\(^{1,2}\\), Jianling Wang\\(^{1}\\), Lara Raad\\(^{3}\\), and Jean-Michel Morel\\(^{4}\\) \n\\(^{1}\\)Université d'Orléans, Université de Tours, CNRS, IDP, UMR 7013, Orléans, France \n\\(^{2}\\)Institut Universitaire de France (IUF)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.233, + 0.903, + 0.25 + ], + "angle": 0, + "content": "\\(^{3}\\)Instituto de Ingeniería Eléctrica, Facultad de Ingeniería, Universidad de la Republica \\(^{4}\\)City University of Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.284, + 0.364, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.419, + 0.364, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.555, + 0.364, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.283, + 0.633, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.419, + 0.633, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.555, + 0.633, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.283, + 0.903, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.419, + 0.903, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.555, + 0.903, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.701, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Figure 1. Various ultra-high definition style transfers of a Gaussian splattering 3D scene. SGSST transfers a very large set of global style statistics of an image to a 3DGS scene by minimizing a tailored multiscale SOS loss, yielding 3D style transfer of superior quality and at unprecedented high resolution (images have size \\(5187 \\times 3361\\))." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.756, + 0.327, + 0.772 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Applying style transfer to a full 3D environment is a challenging task that has seen many developments since the advent of neural rendering. 3D Gaussian splatting (3DGS) has recently pushed further many limits of neural rendering in terms of training speed and reconstruction quality. This work introduces SGSST: Scaling Gaussian Splatting Style Transfer, an optimization-based method to apply style" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.879 + ], + "angle": 0, + "content": "transfer to pretrained 3DGS scenes. We demonstrate that a new multiscale loss based on global neural statistics, that we name SOS for Simultaneously Optimized Scales, enables style transfer to ultra-high resolution 3D scenes. Not only SGSST pioneers 3D scene style transfer at such high image resolutions, it also produces superior visual quality as assessed by thorough qualitative, quantitative and perceptual comparisons." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26535" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.482, + 0.314 + ], + "angle": 0, + "content": "Dealing with ultra-high resolution (UHR) rendering is capital for AR/VR applications. Indeed, when navigating in a 3D environment the user only sees a partial field of view of the environment. This adds challenging issues for applying style transfer to 3D environments, that is, transferring the visual characteristics of an image such as a painting to a 3D scene. Indeed, while the general scene should convey the global color palette of the style painting, when getting closer to objects in a stylized environment the user should expect to see fine painting patterns such as brushstrokes. However, in the current state of the art, the user quickly encounters resolution-limited content in the form of blurry interpolated features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.483, + 0.557 + ], + "angle": 0, + "content": "Ever since the seminal work of Mildenhall et al. [28], neural radiance fields (NeRF) have seen many improvements. Several 3D scene representations have been proposed for improving the quality and resolution of the reconstructed scenes as well as easing the training, e.g. [2, 3, 9, 29]. The recent 3D Gaussian splatting (3DGS) [20] has introduced an efficient high-resolution (HR) scene representation which has stirred much interest [45]. Both frameworks have stirred attempts to 3D style transfer algorithms [6, 16, 22, 25, 26, 33, 46, 48], but these methods have so far produced medium resolution outputs. They do not faithfully transport high resolution multiscale textures such as those present in paintings. Motivated by a recent neural style transfer (NST) [11] contribution tailored for UHR images [10], i.e. with resolution larger than \\(4\\mathrm{k}\\) images, we show that 3DGS can be leveraged for UHR style transfer." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.558, + 0.435, + 0.573 + ], + "angle": 0, + "content": "The contributions of this work are the following:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.578, + 0.483, + 0.623 + ], + "angle": 0, + "content": "- We introduce SOS, a Simultaneously Optimized Scales loss expressed in a single parameterless and explainable formula." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.624, + 0.483, + 0.669 + ], + "angle": 0, + "content": "- By solely optimizing the SOS loss, we reach UHR for 3DGS style transfer and we scale Gaussian Splatting Style Transfer by a four times resolution gain." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.669, + 0.483, + 0.731 + ], + "angle": 0, + "content": "- Superior quality transfer: By transferring a very large set of global style statistics, we obtain superior style transfer quality even at HR resolution, as confirmed by a comparative perceptual study." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.578, + 0.483, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In short, our approach is the first method that allows UHR style transfer directly to 3DGS. It produces high visual quality results and relies on optimizing a single multiscale loss. The simplicity of our approach ensures its reproducibility. Being optimization-based, SGSST's main limitation is a fairly large training time that is two to eight times longer than the initial 3DGS training depending on the image resolution. Yet, considering the high quality of the results and their reproducibility, this contribution is valuable for AR/VR applications that require high visual quality for their user experience." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.652, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.116, + 0.905, + 0.312 + ], + "angle": 0, + "content": "Neural style transfer In the seminal work of Gatys et al. [11] NST is formulated as an optimization problem minimizing the distances between Gram matrices of VGG [37] features. Even though other VGG statistics have been considered, almost all subsequent style transfer and texture synthesis methods rely on VGG [8, 13, 14, 27, 32, 35, 42]. To accelerate style transfer, several methods [18, 40, 41] have attempted to train feed-forward networks approximately minimizing the Gram loss [11]. However, these approximate methods require learning a new model for each style type. This latter limitation has been addressed by fast Universal Style Transfer (UST) approaches [4, 7, 15, 23, 24, 30, 36] that use VGG feature decoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.905, + 0.45 + ], + "angle": 0, + "content": "For HR images, coarse-to-fine multiscale strategies [12, 13, 38] have proved effective, but still face limitations due the high GPU memory usage of VGG statistics. Fast HR alternatives [1, 5, 39, 43, 44] do exist but generally suffer from artifacts and struggle to capture the full style complexity. Recently, SPST [10] proposed an implementation of the Gatys et al. method adapted to UHR (larger that \\(4\\mathrm{k}\\)) images. The visual quality of SPST's results is superior, at the cost of a long optimization time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.905, + 0.669 + ], + "angle": 0, + "content": "Style transfer for neural radiance fields NeRFs [28] have completely redefined the field of 3D scene modeling and novel view synthesis. Editing the visual aspect of NeRFs via style transfer has quickly been addressed [6, 16, 25, 48], usually by fine tuning a pretrained NeRF representation using a style transfer loss, or training an additional fast style transfer module. ARF [48] is a notable exception: It uses Nearest Neighbor Feature Matching (NNFM) for fine tuning a plenoxel radiance field [9], produces high-quality results at moderate resolution, and is the base model for other methods [19, 47]. While these works paved the way for radiance field style transfer, they are all limited in input and output image resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Style transfer for Gaussian splatting A few recent contributions show that 3DGS is a promising framework for 3D scene style transfer. Saroha et al [33] propose a solution for universal style transfer of a given 3DGS scene. The method processes the colors of Gaussians with a tiny MLP trained using fast AdaIN [15] and relying on a multiresolution hash grid representation [29]. StyleGaussian [26] is a concurrent approach that relies on transferring encoding of VGG features to each Gaussian and applying AdaIN to these features. The new Gaussian features are then decoded into an RGB color by processing the K-nearest neighbors of the Gaussians. After training, both methods allow for instantaneous stylization with any style image, but the visual quality of the results is quite low, since high-resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26536" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.455, + 0.106 + ], + "angle": 0, + "content": "details such as brushstroke patterns are not transferred." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.257 + ], + "angle": 0, + "content": "StylizedGS [46] is an optimization-based method that extends ARF [48] to 3DGS. It uses a training loss made of six terms combined with a color transfer preprocessing, the loss being designed to minimize changes in the 3DGS geometry while letting the style evolve via VGG NN matching. \\(\\mathcal{G}\\) -Style [22] also is an optimization based approach that further uses a CLIP-based loss [31]. While these approaches produce slightly better results than ARF [48], both methods are limited in content and style resolution due to the use of nearest-neighbor matching of VGG features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.348 + ], + "angle": 0, + "content": "None of the current state of the art deals with HR style transfer. The stylization is either fast but very approximate due to AdaIn [26, 33] or unable to use HR style images due to NNFM [22, 46, 48]. To the best of our knowledge, SGSST is the first procedure allowing high-quality transfer for 3DGS that is trained and rendered at UHR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.231, + 0.377 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.387, + 0.403 + ], + "angle": 0, + "content": "3.1. Gaussian splatting representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.485, + 0.668 + ], + "angle": 0, + "content": "Starting from a multiview training set of images \\(\\{u_i\\}_{i=1}^{N_{\\mathrm{views}}}\\) of a static scene accompanied with corresponding calibrated cameras \\(\\{\\mathcal{C}_i\\}_{i=1}^{N_{\\mathrm{views}}}\\) computed by structure from motion [34], the 3DGS algorithm [20] trains a set of colored 3D Gaussians \\(\\{\\mathcal{G}_j\\}_{j=1}^{N_{\\mathrm{Gaussians}}}\\) so that they represent the 3D scene from any camera position. Each Gaussian \\(\\mathcal{G}_j\\) is represented by a finite set of features: a center position \\(\\mu_j\\), a covariance matrix \\(\\Sigma_j\\) (encoded by a scaling diagonal matrix and a rotation matrix), an opacity \\(\\alpha_j\\) and a view-dependent color function \\(c_j\\). These parameters are used in a volumetric renderer that determines the color by summing the contribution of each Gaussian that intersects a ray with direction \\((\\theta, \\varphi)\\) via \\(\\alpha\\)-blending (see e.g. [28]). The contribution of a Gaussian is its color \\(c_j(\\theta, \\varphi)\\) multiplied by an opacity \\(\\sigma_j\\) defined as the maximal opacity \\(\\alpha_j\\) times the unnormalized Gaussian density at the ray position [20]. Thus, the resulting color \\(C\\) for the ray is" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.678, + 0.484, + 0.721 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {j = 1} ^ {N} c _ {j} (\\theta , \\varphi) \\sigma_ {j} \\prod_ {k = 1} ^ {j - 1} (1 - \\sigma_ {k}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.485, + 0.793 + ], + "angle": 0, + "content": "where the sum is over all Gaussians intersecting the ray. The color function \\(c_{j}(\\theta ,\\varphi)\\) associated with a Gaussian depends on the spherical direction \\((\\theta ,\\varphi)\\) through an order 3 spherical harmonics polynomial function given by" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.803, + 0.484, + 0.845 + ], + "angle": 0, + "content": "\\[\nc _ {j} (\\theta , \\varphi) = c _ {j, 0} + \\sum_ {\\ell = 1} ^ {3} \\sum_ {m = 1} ^ {2 \\ell + 1} c _ {j, \\ell , m} Y _ {\\ell , m} (\\theta , \\varphi) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "where the vectors \\(c_{j,0}\\) and \\(c_{j,\\ell,m}\\) are in \\(\\mathbb{R}^3\\) and the \\(Y_{\\ell,m}\\) form a basis of the spherical harmonics polynomials of degree \\(\\ell\\). In short, \\(c_{j,0} \\in \\mathbb{R}^3\\) is the main color and the ad" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "ditional coefficients \\(c_{j,\\ell,m}\\) encode smooth variations of this color when the viewing angle changes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Like for NeRF, the key ingredient of the 3DGS parametrization is the differentiable rendering function \\(\\mathcal{R}(\\mathcal{C};\\Theta)\\) that renders a view of the scene for any camera \\(\\mathcal{C}\\) given the scene parameters \\(\\Theta = \\{(\\mu_j,\\Sigma_j,\\alpha_j,c_{j,0},(c_{j,\\ell ,m})_{\\ell ,m})\\}_{j = 1}^{N_{\\mathrm{Gaussians}}}\\). This differentiable rendering function allows to train the Gaussian parameters to minimize the reconstruction error" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.238, + 0.907, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\Theta} \\frac {1}{N _ {\\mathrm {v i e w s}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v i e w s}}} E _ {\\mathrm {r e c o n s t r u c t i o n}} (\\mathcal {R} (\\mathcal {C} _ {i}; \\Theta); u _ {i}) \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.351 + ], + "angle": 0, + "content": "where \\( E_{\\mathrm{reconstruction}} \\) is a 2D image comparison error (such as a combination of mean square error and SSIM [20]). The minimization is conducted using Adam [21] for 30k iterations by randomly picking one view at each iteration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.816, + 0.376 + ], + "angle": 0, + "content": "3.2. Style transfer loss for UHR images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.571 + ], + "angle": 0, + "content": "Our approach relies on optimizing VGG19 [37] feature statistics as initially proposed by Gatys et al. [11]. Content consistency is loosely monitored by preservation of the feature layer \\( L_{\\mathrm{c}} = \\mathrm{ReLU\\_4\\_2} \\) while style transfer is imposed by matching spatial statistics of five VGG19 layers, namely the set \\( \\mathcal{L}_{\\mathrm{s}} = \\{\\mathrm{ReLU\\_k\\_1}, k \\in \\{1, 2, 3, 4, 5\\}\\} \\). The statistics of interest of the feature response \\( V^{L}(w) \\) of an image \\( w \\) at some layer \\( L \\in \\mathcal{L}_{\\mathrm{s}} \\) having \\( n_{\\mathrm{c}}^{L} \\) feature channels are its Gram matrix \\( \\operatorname{Gram}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L} \\times n_{\\mathrm{c}}^{L}} \\), its spatial mean vector \\( \\operatorname{mean}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L}} \\), and its standard deviation vector \\( \\operatorname{std}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L}} \\). Given a content image \\( u \\) and a style image \\( v \\), we consider the loss function" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.583, + 0.907, + 0.6 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {t r a n s f e r}} (x; u, v) = E _ {\\text {c o n t e n t}} (x; u) + E _ {\\text {s t y l e}} (x; v) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.906, + 0.645 + ], + "angle": 0, + "content": "where \\( E_{\\mathrm{content}}(x;u) = \\lambda_{\\mathrm{c}}\\left\\| V^{L_{\\mathrm{c}}}(\\boldsymbol {x}) - V^{L_{\\mathrm{c}}}(\\boldsymbol {u})\\right\\|^{2} \\), with \\( \\lambda_{\\mathrm{c}} > 0 \\) and the style loss is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.657, + 0.907, + 0.69 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {s t y l e}} (x; v) = \\sum_ {L \\in \\mathcal {L} _ {\\mathrm {s}}} E _ {\\text {s t y l e}} ^ {L} (x; v). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.809 + ], + "angle": 0, + "content": "where \\( E_{\\mathrm{style}}^{L}(x;v) \\) is a linear combination of the mean square error between the VGG19 statistics of \\( V^{L}(x) \\) and the one of the style features \\( V^{L}(v) \\) [10]. While only the Gram matrices were originally used [11], it has been shown that adding control for the mean and standard deviation corrects some style transfer color artefacts [10] previously identified in the literature [14, 32, 35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To obtain high-quality style transfer for HR images one needs to optimize the style transfer loss using several scales and a coarse-to-fine approach [12]. Indeed, if one applies style transfer on the highest resolution only, the changes within the content image are limited to local texture and the results does not convey a painting aspect. Due to the use of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26537" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.087, + 0.91, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.392, + 0.908, + 0.45 + ], + "angle": 0, + "content": "Figure 2. Overview of SGSST. Starting from a pretrained realistic 3DGS scene [20], we optimize the colors of each Gaussian using the new multiscale SOS loss (involving \\( n_{\\mathrm{s}} = 3 \\) scales in the illustration). Computing the gradient w.r.t. the loss is feasible for UHR images thanks to the SPST partition-based implementation [10]. Multiscale gradient stacking is used at the node of the rendered image to perform only one backpropagation per iteration through the 3DGS rendering pipeline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.484, + 0.567 + ], + "angle": 0, + "content": "VGG19 features, computing the loss \\( E_{\\mathrm{transfer}}(x;u,v) \\) and its gradient with respect to (w.r.t.) \\( x \\) via backpropagation is memory prohibiting for UHR images. However, using a grid partition and local loss backpropagation based on precomputed global statistics, SPST [10] allows for the exact evaluation of this gradient." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.581, + 0.453, + 0.6 + ], + "angle": 0, + "content": "4. Scaling Gaussian splatting style transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.669 + ], + "angle": 0, + "content": "A complete overview of our SGSST algorithm is given in Figure 2. Starting from a realistic 3DGS representation [20], we optimize the colors of each Gaussian using a multiscale style transfer loss." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.68, + 0.391, + 0.696 + ], + "angle": 0, + "content": "4.1. Stylizing the 3DGS representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.854 + ], + "angle": 0, + "content": "By changing the reconstruction loss of Equation (3) with a style transfer loss for the input style image \\( v \\), one can hope to stylize a realistic 3DGS. However, given the complexity of the 3DGS representation and the many interacting parameters, it is not such an easy task to alter the 3DGS aspects without loosing the content geometry. Our solution is to only optimize for the constant color components \\( \\Theta_{\\mathrm{color}} = \\{c_{j,0}\\}_{j=1}^{N_{\\mathrm{Gaussians}}} \\) of the Gaussians and simply freeze all the other parameters \\( \\Theta_{\\mathrm{init}} \\) given by the initial realistic 3DGS training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We experimentally found that this robust approach ensures a rich style transfer and fully preserves the 3D geometry. Indeed, fixing all the Gaussian parameters except for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.521 + ], + "angle": 0, + "content": "the main color component \\( c_{j,0} \\) preserves the scene geometry, as the location and size of the Gaussians are being kept (see Section 5.3 for ablation experiments)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.783, + 0.544 + ], + "angle": 0, + "content": "4.2. Multiscale Style Transfer Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.58 + ], + "angle": 0, + "content": "We introduce the Simultaneously Optimized Scales (SOS) loss defined as" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.587, + 0.906, + 0.627 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {S O S}} (x; u, v) = \\frac {1}{n _ {\\mathrm {s}}} \\sum_ {s = 0} ^ {n _ {\\mathrm {s}} - 1} E _ {\\text {t r a n s f e r}} \\left(x ^ {\\downarrow 2 ^ {s}}; u ^ {\\downarrow 2 ^ {s}}, v ^ {\\downarrow 2 ^ {s}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.786 + ], + "angle": 0, + "content": "where \\(n_{\\mathrm{s}} \\geq 1\\) is the number of considered scales and \\(u^{\\downarrow 2^{s}}\\) denotes an image \\(u\\) downscaled by a \\(2^{s}\\) factor. This SOS loss (6) enables style transfer simultaneously at all scales. A somewhat similar approach was proposed for 2D texture synthesis [38] but, as already mentioned, multiscale 2D style transfer is generally conducted using a coarse-to-fine strategy [12]. However, for 3DGS we observed that the initial configuration had only little influence on the final result, making the coarse-to-fine strategy ineffective for multiscale style transfer (See Section 5.3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.787, + 0.907, + 0.817 + ], + "angle": 0, + "content": "In the end, as illustrated by Figure 2, the stylization of UHR 3DGS is conducted by solving for" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.824, + 0.906, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\Theta_ {\\text {c o l o r}}} \\frac {1}{N _ {\\text {v i e w s}}} \\sum_ {i = 1} ^ {N _ {\\text {v i e w s}}} E _ {\\text {S O S}} \\left(\\mathcal {R} \\left(\\mathcal {C} _ {i}; \\Theta\\right); u _ {i}, v\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Theta\\) stands for the 3DGS parameters obtained by replacing the color components of \\(\\Theta_{\\mathrm{init}}\\) by the values of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26538" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.295, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.089, + 0.539, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.685, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.089, + 0.904, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.224, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Figure 3. UHR 3DGS style transfer. SGSST allows for the multiscale style transfer of 3DGS scenes at UHR. From left to right: Style image, one UHR stylized view, three magnified details, and evolution of the SOS loss and each style transfer loss that contributes to it. We first optimize the transfer loss for the coarsest scale (yellow curve) for \\(10\\mathrm{k}\\) iterations and then optimize for another \\(10\\mathrm{k}\\) iterations the SOS loss (light blue curve), namely the mean of the four transfer losses. Images sizes are \\(5187\\times 3361\\) for content and \\(4230\\times 3361\\) for style." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.307, + 0.284, + 0.322 + ], + "angle": 0, + "content": "optimization variable \\(\\Theta_{\\mathrm{color}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.331, + 0.306, + 0.347 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Multiscale gradient stacking The 2D SPST method [10] provides the gradient of each term \\( E_{\\mathrm{transfer}}(x^{\\downarrow 2^{s}};u^{\\downarrow 2^{s}},v^{\\downarrow 2^{s}}) \\) w.r.t. to the input \\( x^{\\downarrow 2^{s}} = \\mathcal{R}(\\mathcal{C}_{i};\\Theta)^{\\downarrow 2^{s}} \\). We first backpropagate each of these gradients through the downscaling operator and stack them at the level of the rendered image \\( \\mathcal{R}(\\mathcal{C}_i;\\Theta) \\). When each of the \\( n_{\\mathrm{s}} \\) scales has been treated, the stacked gradient is equal to the gradient of the full loss \\( E_{\\mathrm{SOS}}(\\mathcal{R}(\\mathcal{C}_i;\\Theta);u_i,v) \\) w.r.t. to the rendered image \\( \\mathcal{R}(\\mathcal{C}_i;\\Theta) \\). This gradient is then backpropagated through the Gaussian rendering pipeline to obtain the gradient w.r.t. the Gaussian colors \\( \\Theta_{\\mathrm{color}} \\) (see both orange arrows in Figure 2). Using this strategy we only backpropagate one time per iteration through the 3DGS rendering pipeline instead of \\( n_{\\mathrm{s}} \\) times." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Color transfer via style transfer at the coarsest scale To speed up the color transfer, we first optimize for 10k Adam iterations with the loss restricted to the coarsest scale \\(E_{\\mathrm{transfer}}(\\mathcal{R}(\\mathcal{C}_i;\\Theta)^{\\downarrow 2^s};u_i^{\\downarrow 2^s},v^{\\downarrow 2^s})\\) with \\(s = n_{\\mathrm{s}} - 1\\), then optimize the SOS loss for another 10k Adam iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.484, + 0.723 + ], + "angle": 0, + "content": "Number of scales \\( n_{\\mathrm{s}} \\) is set automatically to use all available scales, the coarsest resolution having sides larger than 256 for VGG19 statistics to be reliable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Reproducibility All our experiments have been conducted using the same SOS loss and training procedure, making our approach parameterless and fully reproducible. Our public PyTorch implementation is based on the public source codes1 for 3DGS [20] training and SPST [10]. Our code and videos of our results are available online2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.839, + 0.484, + 0.902 + ], + "angle": 0, + "content": "1https://github.com/graphdeco-inria/gaussian-splatting; https://github.com/bgalerne/scaling_PA painting_style_transfer 2Code: https://github.com/JianlingWANG2021/SGSST; Videos: https://www.idpoisson.fr/galerne/sgsst/" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.646, + 0.323 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.331, + 0.768, + 0.347 + ], + "angle": 0, + "content": "5.1. Ultra-high resolution results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Our multiscale stylization algorithm is able to transfer style details at UHR for both the content image resolution and the style image resolution. This results in unprecedentedly rich style transfer, as illustrated by Figure 3. As can be observed, minimizing the SOS loss indeed allows to decrease the style transfer loss for all scales (Figure 3 right). The approach is especially relevant when transferring the style of an UHR painting presenting style features at several scales, ranging from a specific color palette to a main curve style and to fine brushstrokes or canvas texture (see close-up views of Figure 3 and the first two lines of Figure 1). To obtain such results, combining style transfer at the largest possible number of scales is critical (see ablation in supp. mat. Figure 9). In addition, the approach is also efficient to transfer the style of a medium resolution style image to an UHR scene, as shown in the last row of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.704 + ], + "angle": 0, + "content": "To the best of our knowledge, our approach is the first to work at UHR resolution for neural style transfer of neurally rendered 3D scenes. One of the main advantages of the Gram-based loss is that it does not depend on the style resolution and, when both the content and style images grow in \\( O(N) \\), its complexity scales in \\( O(N) \\) while NNFM used in ARF [48] scales in \\( O(N^2) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Yet, applying style transfer at such resolutions remains computationally heavy: for the garden image of size \\(5187 \\times 3361\\) (Figures 1 and 3) the style transfer takes 25.5 hours (VS 3 hours to train the initial 3DGS model), that is an 8.5 overhead factor. For an image of moderate size (Figure 4 left and middle) the style transfer and the initial 3DGS training take respectively 22 and 10 minutes, that is, the overhead factor is only 2.2. These time values were obtained using a single A100 GPU with 80 GB of memory and could be accelerated by adapting the SOS loss implementation to a multi-GPU setting. Note that high computation times are inherent to UHR style transfer: Running the 2D SPST method for the 185 garden training images takes 27" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26539" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.904, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.908, + 0.417 + ], + "angle": 0, + "content": "Figure 4. Comparison of SGSST (ours, top) with StyleGaussian [26] (middle) and ARF [48] (bottom). From left to right the content resolutions are \\(980 \\times 545\\) (train), \\(979 \\times 546\\) (truck), and \\(3115 \\times 2076\\) (counter). For the first two examples, the various outputs keep the resolution of the content, but for the HR counter scene, the output sizes are \\(3115 \\times 2076\\) for SGSST, \\(1600 \\times 1066\\) for StyleGaussian and \\(779 \\times 519\\) for ARF (see supp. mat. Figure 28 for ARF results without downscaling). Thanks to its multiscale global VGG statistics, SGSST is the most faithful method regarding style consistency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.443, + 0.275, + 0.459 + ], + "angle": 0, + "content": "hours (8.9 min. per image)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.223, + 0.485 + ], + "angle": 0, + "content": "5.2. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.584 + ], + "angle": 0, + "content": "We perform a thorough comparative study using 40 3D style transfer experiments using 9 different scenes and various style images (see supp. mat.). We compare our results with the NeRF-based ARF [48] and the 3DGS-based StyleGaussian [26] algorithms, described in Section 2, using their public implementations3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.782 + ], + "angle": 0, + "content": "Comparing style transfer methods is challenging because each algorithm treats the style image differently. Following high quality 2D style transfer [10, 12], our loss uses up to four scales and each scale uses five VGG layers. The UHR style images were downscaled so that the style has the same size as the content images (no upscale was applied if the style image is smaller than the content image). In contrast, ARF uses a single VGG layer from a single scale and the style image is downscaled to be twice smaller than the content input, resulting in smaller local texture. StyleGaussian reduces the style images so that it has the size \\(256 \\times 256\\), a scale that is hardly sufficient to represent HR style images such as paintings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Qualitative comparison Figure 4 shows three different comparative experiments. As one can observe, the results of StyleGaussian [26] are generally not satisfying since the method fails to transfer local texture or to preserve the style" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.61 + ], + "angle": 0, + "content": "image's color palette. ARF results better reproduce brushstroke textures, but the transfer is limited, as the method only involves a single VGG layer at a single scale. Also, NNFM does not ensure the preservation of a global color palette. In comparison, SGSST preserves the style at all considered scales. This results in color palette preservation, as well as a verifiable transfer of features of any size, from large brushstrokes (Figure 4 middle) to local grain transfer (Figure 4 right). In addition, SGSST is the only method that performs style transfer at the original resolution of the HR example of Figure 4 right." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative comparison Even though there is no consensus for the quantitative evaluation of NST [17], following previous works, we report two different metrics for style transfer quality and texture consistency across views. Style transfer quality can be measured by the Gram loss [11]. A second metric proper to NeRF style transfer [25, 26, 33] is to check the short-term and long-term consistency of the radiance fields in terms of LPIPS [49] and RMSE between wrapped views. Since SGSST is the only method working with UHR images, when necessary we forcefully downgraded all the results to the resolution of ARF for comparison. The average of these two metrics over our 40 experiments is reported in Table 1. The Gram loss is the best for SGSST and, surprisingly, StyleGaussian achieves a lower Gram loss than ARF, which is not consistent with the qualitative evaluation, probably explained by ARF using a single VGG layer compared to five for the two other methods. In terms of consistency metrics, StyleGaussian reports to be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3https://github.com/Kai-46/ARF-svox2; https://github.com/Kunhao-Liu/StyleGaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26540" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.487, + 0.193 + ], + "angle": 0, + "content": "
MethodTransfer qualityShort-range consistencyLong-range consistency
Gram↓LPIPS↓RMSE↓LPIPS↓RMSE↓
SGSST2.59e80.0300.0320.0550.063
StyleGaussian4.61e80.0330.0290.0500.056
ARF5.77e80.0400.0370.0720.066
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.202, + 0.485, + 0.258 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison. Style transfer quality and texture consistency metrics averaged over 40 experiments for SGSST, StyleGaussian, and ARF. Best results in bold, second best underlined." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.272, + 0.482, + 0.318 + ], + "angle": 0, + "content": "
SGSSTARFStyleGaussian
Voting results (%)66.331.62.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.357 + ], + "angle": 0, + "content": "Table 2. Perceptual study. Summary of the 680 votes for the most style consistent result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.413 + ], + "angle": 0, + "content": "the most stable approach followed by SGSST but note that this metric favors the lack of texture." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Perceptual study To further support our results, we conducted a perceptual study comparing the 40 3D style transfer experiments. For each example, four images were displayed (at a resolution of \\(1280 \\times 720\\)): the style image and, in a blind random order, the results of the three methods (SGSST, ARF and StyleGaussian) shown at a common viewpoint (also randomly selected). Each participant was shown ten random instances and was asked to select the result that was the most faithful to the style image. The study was conducted on the web with volunteer participants." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.703 + ], + "angle": 0, + "content": "A total of 68 participants took part in the test, resulting in 680 votes summarized in Table 2. This perceptual study shows that the fast style transfer performed by Style-Gaussian is consistently judged inferior in quality over ARF and SGSST. It also confirms that SGSST is far superior to ARF in terms of visual quality since \\(66\\%\\) of the participants ranked our method first for its painting style transfer quality, even when presented with results downscaled in resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.242, + 0.728 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Influence of optimization parameters Our SGSST algorithm stylizes a realistic 3DGS scene by tuning the constant color components of the 3DGS Gaussians using a single 2D style transfer loss function at multiple scales (Equation (6)). In contrast, Zhang et al. [46] optimize all 3DGS parameters. This, however, necessitates a complex loss made of six different terms to avoid artefacts: a loss term enforces consistency with the original geometry via depth preservation while a preprocessing of Gaussians floaters is necessary after color transfer." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Figure 5 shows that optimizing more parameters of the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.422, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Figure 5. Influence of optimization parameters: Allowing more 3DGS parameters to be optimized when minimizing the SOS loss does not improve the stylization quality and can dramatically impact the geometry. From left to right: Style image, content, SGSST default (optimization of colors), results when optimizing all spherical harmonics, results when optimizing all parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.653 + ], + "angle": 0, + "content": "3DGS for SGSST brings no benefit. Optimizing all the spherical harmonic coefficients does not improve the result, and letting all the parameters free like in [22, 46] leads to a strong degradation of the geometry. Note that two instantaneous 3DGS style transfer methods [26, 33] are based on modifying the color features via some neural networks, but their results are not comparable to optimization-based approaches in terms of visual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Failure of coarse-to-fine strategy Our approach minimizes a style transfer loss simultaneously at all scales. This is different from the coarse-to-fine style transfer strategy that has proven successful for HR style transfer [12]. Figure 6 illustrates that this coarse-to-fine strategy fails in the context of 3DGS. Indeed, the gain obtained by optimizing at a given scale is quickly lost when optimizing the next one leading to the disappearance of large scale features. This can be explained by the fact that the 3DGS representation is a more constrained representation than pixel grids due to its sparsity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Color transfer via style transfer at the coarsest scale As described in Section 4.3, we first optimize the style transfer loss at the lowest resolution for \\(10\\mathrm{k}\\) iterations and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 6. Failure of coarse-to-fine strategy. Results of coarse-to-fine style transfer for the example of Figure 3. Each scale is initialized with the output of the previous scale. As shown by the evolution of the losses (right), when training at a new scale, the loss of the previous scales increases quickly. This explains why the large scale painting features disappear progressively and are absent after training the target UHR (see close-up details)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.317, + 0.222, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.393, + 0.221, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.317, + 0.351, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.393, + 0.352, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.317, + 0.48, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.393, + 0.481, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Figure 7. Color transfer via style transfer at the coarsest scale. The first 10k iterations at the coarsest scale allow for quick color changes in the 3DGS scene. Removing this step may lead to color artifacts in the result (top) compared to our default two-step optimization (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.622 + ], + "angle": 0, + "content": "then optimize the SOS loss for another 10k iterations. Figure 7 illustrates that these first iterations are necessary for a faithful style color palette reproduction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.336, + 0.652 + ], + "angle": 0, + "content": "6. Discussion and limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.724 + ], + "angle": 0, + "content": "Texture representation The texture representation within a 3DGS scene depends on the density of Gaussians and may be limited in low density areas. Isolated Gaussians can sometimes be spotted as illustrated by Figure 8." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Large computation time Depending on the resolution, SGSST requires from several minutes to several hours of computation. On the other hand, fast 3DGS stylization approaches [26, 33] do not reach a satisfactory visual quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Content-style mismatch As said earlier, the Gram loss has the advantage of being independent of the style's image resolution. Also, it enables a faithful transfer of global statistics of the style image, such as its color palette. This important feature is, nevertheless, counterproductive when" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Figure 8. Limited texture representation due to low Gaussian density of the initial 3DGS. From left to right: Original scene, 3DGS reconstruction, stylization of the 3DGS scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.603 + ], + "angle": 0, + "content": "the style and content images strongly mismatch, leading to color bleeding or texturing of flat areas. Other controls can be added to mitigate these artefacts [12] and it was shown that these controls are effective for 3DGS scenes [46]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.616, + 0.634, + 0.631 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.791 + ], + "angle": 0, + "content": "In this work we presented SGSST, a method that, for the first time, enables UHR 3DGS style transfer. To that aim, among other innovations, we introduced the simultaneously optimized scales (SOS) loss. Our qualitative, quantitative and perceptual studies show that SGSST obtains superior style transfer quality than state of the art, even after reducing our results' resolution for a fair comparison with methods that do not reach UHR. Such high quality UHR results necessitate a large computation time that, nevertheless, remains comparable with that of UHR 3DGS training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.909, + 0.883 + ], + "angle": 0, + "content": "This work opens the way to several research directions. A first challenge is to produce equally high quality style transfer with a faster algorithm based on UST. A second more exploratory direction is to investigate geometry style transfer for 3DGS by designing adapted regularization to avoid the caveats depicted by Figure 5." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "Acknowledgements: B. Galerne and L. Raad acknowledge the support of the project MISTIC (ANR-19-CE40-005)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.14, + 0.188, + 0.156 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.165, + 0.483, + 0.221 + ], + "angle": 0, + "content": "[1] Jie An, Tao Li, Haozhi Huang, Li Shen, Xuan Wang, Yongyi Tang, Jinwen Ma, Wei Liu, and Jiebo Luo. Real-time universal style transfer on high-resolution images via zero-channel pruning. arXiv preprint arXiv:2006.09029, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.223, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470–5479, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.296, + 0.484, + 0.35 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022, pages 333-350, Cham, 2022. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.483, + 0.393 + ], + "angle": 0, + "content": "[4] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.398, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[5] Zhe Chen, Wenhai Wang, Enze Xie, Tong Lu, and Ping Luo. Towards ultra-resolution neural style transfer via thumbnail instance normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 393-400, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.456, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1475-1484, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.528, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[7] Tai-Yin Chiu and Danna Gurari. Iterative feature transformation for fast and versatile universal style transfer. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[8] Valentin De Bortoli, Agnès Desolneux, Alain Durmus, Bruno Galerne, and Arthur Leclaire. Maximum entropy methods for texture synthesis: Theory and practice. SIAM Journal on Mathematics of Data Science, 3(1):52-82, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.729, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[10] Bruno Galerne, Lara Raad, José Lezama, and Jean-Michel Morel. Scaling Painting Style Transfer. Computer Graphics Forum, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[11] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.165, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Nicolas Gonthier, Yann Gousseau, and Saïd Ladjal. High-resolution neural texture synthesis with long-range constraints. Journal of Mathematical Imaging and Vision, 64 (5):478-492, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9412-9420, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.907, + 0.35 + ], + "angle": 0, + "content": "[16] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18342–18352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[17] Eleftherios Ioannou and Steve Maddock. Evaluation in neural style transfer: A review. Computer Graphics Forum, 43 (6):e15165, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[19] Hyunyoung Jung, Seonghyeon Nam, Nikolaos Sarafianos, Sungwoo Yoo, Alexander Sorkine-Hornung, and Rakesh Ranjan. Geometry transfer for stylizing radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8565-8575, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[22] Aron Samuel Kovács, Pedro Hermosilla, and Renata G. Rraidou. \\(\\mathcal{G}\\)-style: Stylized gaussian splatting. Computer Graphics Forum, n/a(n/a):e15259, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3809-3817, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30:386-396, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Kunhao Liu, Fangneng Zhan, Yiwen Chen, Jiahui Zhang, Yingchen Yu, Abdulmotaleb El Saddik, Shijian Lu, and Eric P. Xing. Stylerf: Zero-shot 3d style transfer of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8338-8348, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26543" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[26] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. StyleGaussian: Instant 3d style transfer with Gaussian splatting, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[27] Yang Lu, Song-chun Zhu, and Ying Nian Wu. Learning frame models using cnn filters. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[30] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[32] Eric Risser, Pierre Wilmot, and Connelly Barnes. Stable and controllable neural texture synthesis and style transfer using histogram losses, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[33] Abhishek Saroha, Marija Gladkova, Cecilia Curreli, Tarun Yenamandra, and Daniel Cremers. Gaussian splatting in style, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[34] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. Graph., 36(5), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[36] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[38] Xavier Snelgrove. High-resolution multi-scale neural texture synthesis. In SIGGRAPH Asia 2017 Technical Briefs, New York, NY, USA, 2017. Association for Computing Machinery. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[39] Ondrej Texler, David Futschik, Jakub Fišer, Michal Lukáč, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Arbitrary style transfer using neurally-guided patch-based synthesis. Computers & Graphics, 87:62-71, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] D. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In ICML, pages 1349-1357, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[42] Jonathan Vacher, Aida Davila, Adam Kohn, and Ruben Coen-Cagli. Texture interpolation for probing visual perception. In Advances in Neural Information Processing Systems, pages 22146-22157. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[44] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. Proceedings of the AAAI Conference on Artificial Intelligence, 37(3): 2742-2750, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[45] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[46] Dingxi Zhang, Zhuoxun Chen, Yu-Jie Yuan, Fang-Lue Zhang, Zhenliang He, Shiguang Shan, and Lin Gao. StylizedGS: Controllable stylization for 3d Gaussian splatting, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[47] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. CoARF: Controllable 3D Artistic Style Transfer for Radiance Fields. In 2024 International Conference on 3D Vision (3DV), pages 612-622, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[48] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. ARF: Artistic radiance fields. In ECCV 2022, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26544" + } + ] +] \ No newline at end of file diff --git a/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_origin.pdf b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..797fc308ece74d54d3f2154c49238ee5345f407c --- /dev/null +++ b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/5b7056b2-fc5b-410d-b257-8d1a879ecc29_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dbb1125e4e71acb723ffc09fee2ef390a7b4eb17ddd5b6d5c79788abf559985 +size 9116522 diff --git a/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/full.md b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0212520d6ee291044751c441648329757ec17937 --- /dev/null +++ b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/full.md @@ -0,0 +1,315 @@ +# SGSST: Scaling Gaussian Splatting Style Transfer + +Bruno Galerne $^{1,2}$ , Jianling Wang $^{1}$ , Lara Raad $^{3}$ , and Jean-Michel Morel $^{4}$ $^{1}$ Université d'Orléans, Université de Tours, CNRS, IDP, UMR 7013, Orléans, France + $^{2}$ Institut Universitaire de France (IUF) + +$^{3}$ Instituto de Ingeniería Eléctrica, Facultad de Ingeniería, Universidad de la Republica $^{4}$ City University of Hong Kong + +![](images/de06ada5b3d524820e24de71826a3553f25248cccaea80e5dfcf6e9589b636e1.jpg) + +![](images/2b71d6ab9ed5269de7ecfc03ccfe4dad6f25f38d035b64203aeac34e6151ba19.jpg) + +![](images/95e7be46152d6fe6611224de0ffcde9fc03c40238f151c3470f6ced37493b867.jpg) +Figure 1. Various ultra-high definition style transfers of a Gaussian splattering 3D scene. SGSST transfers a very large set of global style statistics of an image to a 3DGS scene by minimizing a tailored multiscale SOS loss, yielding 3D style transfer of superior quality and at unprecedented high resolution (images have size $5187 \times 3361$ ). + +![](images/d77b5a44a8876ebe874e29b631e9e3fd81680a9beda2be069e017c7ac6ed352c.jpg) + +![](images/47abad2cd38069a40dbe926a18e5d9ffa25d9d29c54c101edd22daeed9d03e8d.jpg) + +![](images/b801c471712eeb801a57ac683844d4835f174f5d9811af696b5dae513c602e05.jpg) + +![](images/53dfb53fae69cbf3dbfaab5f83bde197be0e5b7b0c57c566d8f53c338af013d9.jpg) + +![](images/529a8b274e630287f10f7bff4e84d7e7e3594158eac76e78ddc1220fc5e522b1.jpg) + +![](images/66774d176e9ea7e04e597358a655248b6e4e5281adf444020159f9f62c775fab.jpg) + +# Abstract + +Applying style transfer to a full 3D environment is a challenging task that has seen many developments since the advent of neural rendering. 3D Gaussian splatting (3DGS) has recently pushed further many limits of neural rendering in terms of training speed and reconstruction quality. This work introduces SGSST: Scaling Gaussian Splatting Style Transfer, an optimization-based method to apply style + +transfer to pretrained 3DGS scenes. We demonstrate that a new multiscale loss based on global neural statistics, that we name SOS for Simultaneously Optimized Scales, enables style transfer to ultra-high resolution 3D scenes. Not only SGSST pioneers 3D scene style transfer at such high image resolutions, it also produces superior visual quality as assessed by thorough qualitative, quantitative and perceptual comparisons. + +# 1. Introduction + +Dealing with ultra-high resolution (UHR) rendering is capital for AR/VR applications. Indeed, when navigating in a 3D environment the user only sees a partial field of view of the environment. This adds challenging issues for applying style transfer to 3D environments, that is, transferring the visual characteristics of an image such as a painting to a 3D scene. Indeed, while the general scene should convey the global color palette of the style painting, when getting closer to objects in a stylized environment the user should expect to see fine painting patterns such as brushstrokes. However, in the current state of the art, the user quickly encounters resolution-limited content in the form of blurry interpolated features. + +Ever since the seminal work of Mildenhall et al. [28], neural radiance fields (NeRF) have seen many improvements. Several 3D scene representations have been proposed for improving the quality and resolution of the reconstructed scenes as well as easing the training, e.g. [2, 3, 9, 29]. The recent 3D Gaussian splatting (3DGS) [20] has introduced an efficient high-resolution (HR) scene representation which has stirred much interest [45]. Both frameworks have stirred attempts to 3D style transfer algorithms [6, 16, 22, 25, 26, 33, 46, 48], but these methods have so far produced medium resolution outputs. They do not faithfully transport high resolution multiscale textures such as those present in paintings. Motivated by a recent neural style transfer (NST) [11] contribution tailored for UHR images [10], i.e. with resolution larger than $4\mathrm{k}$ images, we show that 3DGS can be leveraged for UHR style transfer. + +The contributions of this work are the following: + +- We introduce SOS, a Simultaneously Optimized Scales loss expressed in a single parameterless and explainable formula. +- By solely optimizing the SOS loss, we reach UHR for 3DGS style transfer and we scale Gaussian Splatting Style Transfer by a four times resolution gain. +- Superior quality transfer: By transferring a very large set of global style statistics, we obtain superior style transfer quality even at HR resolution, as confirmed by a comparative perceptual study. + +In short, our approach is the first method that allows UHR style transfer directly to 3DGS. It produces high visual quality results and relies on optimizing a single multiscale loss. The simplicity of our approach ensures its reproducibility. Being optimization-based, SGSST's main limitation is a fairly large training time that is two to eight times longer than the initial 3DGS training depending on the image resolution. Yet, considering the high quality of the results and their reproducibility, this contribution is valuable for AR/VR applications that require high visual quality for their user experience. + +# 2. Related work + +Neural style transfer In the seminal work of Gatys et al. [11] NST is formulated as an optimization problem minimizing the distances between Gram matrices of VGG [37] features. Even though other VGG statistics have been considered, almost all subsequent style transfer and texture synthesis methods rely on VGG [8, 13, 14, 27, 32, 35, 42]. To accelerate style transfer, several methods [18, 40, 41] have attempted to train feed-forward networks approximately minimizing the Gram loss [11]. However, these approximate methods require learning a new model for each style type. This latter limitation has been addressed by fast Universal Style Transfer (UST) approaches [4, 7, 15, 23, 24, 30, 36] that use VGG feature decoders. + +For HR images, coarse-to-fine multiscale strategies [12, 13, 38] have proved effective, but still face limitations due the high GPU memory usage of VGG statistics. Fast HR alternatives [1, 5, 39, 43, 44] do exist but generally suffer from artifacts and struggle to capture the full style complexity. Recently, SPST [10] proposed an implementation of the Gatys et al. method adapted to UHR (larger that $4\mathrm{k}$ ) images. The visual quality of SPST's results is superior, at the cost of a long optimization time. + +Style transfer for neural radiance fields NeRFs [28] have completely redefined the field of 3D scene modeling and novel view synthesis. Editing the visual aspect of NeRFs via style transfer has quickly been addressed [6, 16, 25, 48], usually by fine tuning a pretrained NeRF representation using a style transfer loss, or training an additional fast style transfer module. ARF [48] is a notable exception: It uses Nearest Neighbor Feature Matching (NNFM) for fine tuning a plenoxel radiance field [9], produces high-quality results at moderate resolution, and is the base model for other methods [19, 47]. While these works paved the way for radiance field style transfer, they are all limited in input and output image resolution. + +Style transfer for Gaussian splatting A few recent contributions show that 3DGS is a promising framework for 3D scene style transfer. Saroha et al [33] propose a solution for universal style transfer of a given 3DGS scene. The method processes the colors of Gaussians with a tiny MLP trained using fast AdaIN [15] and relying on a multiresolution hash grid representation [29]. StyleGaussian [26] is a concurrent approach that relies on transferring encoding of VGG features to each Gaussian and applying AdaIN to these features. The new Gaussian features are then decoded into an RGB color by processing the K-nearest neighbors of the Gaussians. After training, both methods allow for instantaneous stylization with any style image, but the visual quality of the results is quite low, since high-resolution + +details such as brushstroke patterns are not transferred. + +StylizedGS [46] is an optimization-based method that extends ARF [48] to 3DGS. It uses a training loss made of six terms combined with a color transfer preprocessing, the loss being designed to minimize changes in the 3DGS geometry while letting the style evolve via VGG NN matching. $\mathcal{G}$ -Style [22] also is an optimization based approach that further uses a CLIP-based loss [31]. While these approaches produce slightly better results than ARF [48], both methods are limited in content and style resolution due to the use of nearest-neighbor matching of VGG features. + +None of the current state of the art deals with HR style transfer. The stylization is either fast but very approximate due to AdaIn [26, 33] or unable to use HR style images due to NNFM [22, 46, 48]. To the best of our knowledge, SGSST is the first procedure allowing high-quality transfer for 3DGS that is trained and rendered at UHR. + +# 3. Preliminaries + +# 3.1. Gaussian splatting representation + +Starting from a multiview training set of images $\{u_i\}_{i=1}^{N_{\mathrm{views}}}$ of a static scene accompanied with corresponding calibrated cameras $\{\mathcal{C}_i\}_{i=1}^{N_{\mathrm{views}}}$ computed by structure from motion [34], the 3DGS algorithm [20] trains a set of colored 3D Gaussians $\{\mathcal{G}_j\}_{j=1}^{N_{\mathrm{Gaussians}}}$ so that they represent the 3D scene from any camera position. Each Gaussian $\mathcal{G}_j$ is represented by a finite set of features: a center position $\mu_j$ , a covariance matrix $\Sigma_j$ (encoded by a scaling diagonal matrix and a rotation matrix), an opacity $\alpha_j$ and a view-dependent color function $c_j$ . These parameters are used in a volumetric renderer that determines the color by summing the contribution of each Gaussian that intersects a ray with direction $(\theta, \varphi)$ via $\alpha$ -blending (see e.g. [28]). The contribution of a Gaussian is its color $c_j(\theta, \varphi)$ multiplied by an opacity $\sigma_j$ defined as the maximal opacity $\alpha_j$ times the unnormalized Gaussian density at the ray position [20]. Thus, the resulting color $C$ for the ray is + +$$ +C = \sum_ {j = 1} ^ {N} c _ {j} (\theta , \varphi) \sigma_ {j} \prod_ {k = 1} ^ {j - 1} (1 - \sigma_ {k}) \tag {1} +$$ + +where the sum is over all Gaussians intersecting the ray. The color function $c_{j}(\theta ,\varphi)$ associated with a Gaussian depends on the spherical direction $(\theta ,\varphi)$ through an order 3 spherical harmonics polynomial function given by + +$$ +c _ {j} (\theta , \varphi) = c _ {j, 0} + \sum_ {\ell = 1} ^ {3} \sum_ {m = 1} ^ {2 \ell + 1} c _ {j, \ell , m} Y _ {\ell , m} (\theta , \varphi) \tag {2} +$$ + +where the vectors $c_{j,0}$ and $c_{j,\ell,m}$ are in $\mathbb{R}^3$ and the $Y_{\ell,m}$ form a basis of the spherical harmonics polynomials of degree $\ell$ . In short, $c_{j,0} \in \mathbb{R}^3$ is the main color and the ad + +ditional coefficients $c_{j,\ell,m}$ encode smooth variations of this color when the viewing angle changes. + +Like for NeRF, the key ingredient of the 3DGS parametrization is the differentiable rendering function $\mathcal{R}(\mathcal{C};\Theta)$ that renders a view of the scene for any camera $\mathcal{C}$ given the scene parameters $\Theta = \{(\mu_j,\Sigma_j,\alpha_j,c_{j,0},(c_{j,\ell ,m})_{\ell ,m})\}_{j = 1}^{N_{\mathrm{Gaussians}}}$ . This differentiable rendering function allows to train the Gaussian parameters to minimize the reconstruction error + +$$ +\min _ {\Theta} \frac {1}{N _ {\mathrm {v i e w s}}} \sum_ {i = 1} ^ {N _ {\mathrm {v i e w s}}} E _ {\mathrm {r e c o n s t r u c t i o n}} (\mathcal {R} (\mathcal {C} _ {i}; \Theta); u _ {i}) \qquad (3) +$$ + +where $E_{\mathrm{reconstruction}}$ is a 2D image comparison error (such as a combination of mean square error and SSIM [20]). The minimization is conducted using Adam [21] for 30k iterations by randomly picking one view at each iteration. + +# 3.2. Style transfer loss for UHR images + +Our approach relies on optimizing VGG19 [37] feature statistics as initially proposed by Gatys et al. [11]. Content consistency is loosely monitored by preservation of the feature layer $L_{\mathrm{c}} = \mathrm{ReLU\_4\_2}$ while style transfer is imposed by matching spatial statistics of five VGG19 layers, namely the set $\mathcal{L}_{\mathrm{s}} = \{\mathrm{ReLU\_k\_1}, k \in \{1, 2, 3, 4, 5\}\}$ . The statistics of interest of the feature response $V^{L}(w)$ of an image $w$ at some layer $L \in \mathcal{L}_{\mathrm{s}}$ having $n_{\mathrm{c}}^{L}$ feature channels are its Gram matrix $\operatorname{Gram}(V^{L}(w)) \in \mathbb{R}^{n_{\mathrm{c}}^{L} \times n_{\mathrm{c}}^{L}}$ , its spatial mean vector $\operatorname{mean}(V^{L}(w)) \in \mathbb{R}^{n_{\mathrm{c}}^{L}}$ , and its standard deviation vector $\operatorname{std}(V^{L}(w)) \in \mathbb{R}^{n_{\mathrm{c}}^{L}}$ . Given a content image $u$ and a style image $v$ , we consider the loss function + +$$ +E _ {\text {t r a n s f e r}} (x; u, v) = E _ {\text {c o n t e n t}} (x; u) + E _ {\text {s t y l e}} (x; v) \tag {4} +$$ + +where $E_{\mathrm{content}}(x;u) = \lambda_{\mathrm{c}}\left\| V^{L_{\mathrm{c}}}(\boldsymbol {x}) - V^{L_{\mathrm{c}}}(\boldsymbol {u})\right\|^{2}$ , with $\lambda_{\mathrm{c}} > 0$ and the style loss is defined by + +$$ +E _ {\text {s t y l e}} (x; v) = \sum_ {L \in \mathcal {L} _ {\mathrm {s}}} E _ {\text {s t y l e}} ^ {L} (x; v). \tag {5} +$$ + +where $E_{\mathrm{style}}^{L}(x;v)$ is a linear combination of the mean square error between the VGG19 statistics of $V^{L}(x)$ and the one of the style features $V^{L}(v)$ [10]. While only the Gram matrices were originally used [11], it has been shown that adding control for the mean and standard deviation corrects some style transfer color artefacts [10] previously identified in the literature [14, 32, 35]. + +To obtain high-quality style transfer for HR images one needs to optimize the style transfer loss using several scales and a coarse-to-fine approach [12]. Indeed, if one applies style transfer on the highest resolution only, the changes within the content image are limited to local texture and the results does not convey a painting aspect. Due to the use of + +![](images/413a883f11bd3798e04372d1cb863202ca3a20ae935a9c76d1f4592d6658bb0b.jpg) +Figure 2. Overview of SGSST. Starting from a pretrained realistic 3DGS scene [20], we optimize the colors of each Gaussian using the new multiscale SOS loss (involving $n_{\mathrm{s}} = 3$ scales in the illustration). Computing the gradient w.r.t. the loss is feasible for UHR images thanks to the SPST partition-based implementation [10]. Multiscale gradient stacking is used at the node of the rendered image to perform only one backpropagation per iteration through the 3DGS rendering pipeline. + +VGG19 features, computing the loss $E_{\mathrm{transfer}}(x;u,v)$ and its gradient with respect to (w.r.t.) $x$ via backpropagation is memory prohibiting for UHR images. However, using a grid partition and local loss backpropagation based on precomputed global statistics, SPST [10] allows for the exact evaluation of this gradient. + +# 4. Scaling Gaussian splatting style transfer + +A complete overview of our SGSST algorithm is given in Figure 2. Starting from a realistic 3DGS representation [20], we optimize the colors of each Gaussian using a multiscale style transfer loss. + +# 4.1. Stylizing the 3DGS representation + +By changing the reconstruction loss of Equation (3) with a style transfer loss for the input style image $v$ , one can hope to stylize a realistic 3DGS. However, given the complexity of the 3DGS representation and the many interacting parameters, it is not such an easy task to alter the 3DGS aspects without loosing the content geometry. Our solution is to only optimize for the constant color components $\Theta_{\mathrm{color}} = \{c_{j,0}\}_{j=1}^{N_{\mathrm{Gaussians}}}$ of the Gaussians and simply freeze all the other parameters $\Theta_{\mathrm{init}}$ given by the initial realistic 3DGS training. + +We experimentally found that this robust approach ensures a rich style transfer and fully preserves the 3D geometry. Indeed, fixing all the Gaussian parameters except for + +the main color component $c_{j,0}$ preserves the scene geometry, as the location and size of the Gaussians are being kept (see Section 5.3 for ablation experiments). + +# 4.2. Multiscale Style Transfer Loss + +We introduce the Simultaneously Optimized Scales (SOS) loss defined as + +$$ +E _ {\mathrm {S O S}} (x; u, v) = \frac {1}{n _ {\mathrm {s}}} \sum_ {s = 0} ^ {n _ {\mathrm {s}} - 1} E _ {\text {t r a n s f e r}} \left(x ^ {\downarrow 2 ^ {s}}; u ^ {\downarrow 2 ^ {s}}, v ^ {\downarrow 2 ^ {s}}\right) \tag {6} +$$ + +where $n_{\mathrm{s}} \geq 1$ is the number of considered scales and $u^{\downarrow 2^{s}}$ denotes an image $u$ downscaled by a $2^{s}$ factor. This SOS loss (6) enables style transfer simultaneously at all scales. A somewhat similar approach was proposed for 2D texture synthesis [38] but, as already mentioned, multiscale 2D style transfer is generally conducted using a coarse-to-fine strategy [12]. However, for 3DGS we observed that the initial configuration had only little influence on the final result, making the coarse-to-fine strategy ineffective for multiscale style transfer (See Section 5.3). + +In the end, as illustrated by Figure 2, the stylization of UHR 3DGS is conducted by solving for + +$$ +\min _ {\Theta_ {\text {c o l o r}}} \frac {1}{N _ {\text {v i e w s}}} \sum_ {i = 1} ^ {N _ {\text {v i e w s}}} E _ {\text {S O S}} \left(\mathcal {R} \left(\mathcal {C} _ {i}; \Theta\right); u _ {i}, v\right), \tag {7} +$$ + +where $\Theta$ stands for the 3DGS parameters obtained by replacing the color components of $\Theta_{\mathrm{init}}$ by the values of the + +![](images/eeec0944ae3ccc579a39d2848e9d071444014c7099aefb6325676e47d2aa5859.jpg) +Figure 3. UHR 3DGS style transfer. SGSST allows for the multiscale style transfer of 3DGS scenes at UHR. From left to right: Style image, one UHR stylized view, three magnified details, and evolution of the SOS loss and each style transfer loss that contributes to it. We first optimize the transfer loss for the coarsest scale (yellow curve) for $10\mathrm{k}$ iterations and then optimize for another $10\mathrm{k}$ iterations the SOS loss (light blue curve), namely the mean of the four transfer losses. Images sizes are $5187\times 3361$ for content and $4230\times 3361$ for style. + +![](images/da26b8c99258f4ebc42db88da241f81055755c6edb98f152dc0383612e7c17ee.jpg) + +![](images/a763e2db65789129d44dea624b26750219869325e16f115b9491821249ee1fd0.jpg) + +![](images/9d3b5195b95b89d4bafac744b4c4b51800ba3bfe578d03a566bc494ed431d284.jpg) + +optimization variable $\Theta_{\mathrm{color}}$ + +# 4.3. Implementation details + +Multiscale gradient stacking The 2D SPST method [10] provides the gradient of each term $E_{\mathrm{transfer}}(x^{\downarrow 2^{s}};u^{\downarrow 2^{s}},v^{\downarrow 2^{s}})$ w.r.t. to the input $x^{\downarrow 2^{s}} = \mathcal{R}(\mathcal{C}_{i};\Theta)^{\downarrow 2^{s}}$ . We first backpropagate each of these gradients through the downscaling operator and stack them at the level of the rendered image $\mathcal{R}(\mathcal{C}_i;\Theta)$ . When each of the $n_{\mathrm{s}}$ scales has been treated, the stacked gradient is equal to the gradient of the full loss $E_{\mathrm{SOS}}(\mathcal{R}(\mathcal{C}_i;\Theta);u_i,v)$ w.r.t. to the rendered image $\mathcal{R}(\mathcal{C}_i;\Theta)$ . This gradient is then backpropagated through the Gaussian rendering pipeline to obtain the gradient w.r.t. the Gaussian colors $\Theta_{\mathrm{color}}$ (see both orange arrows in Figure 2). Using this strategy we only backpropagate one time per iteration through the 3DGS rendering pipeline instead of $n_{\mathrm{s}}$ times. + +Color transfer via style transfer at the coarsest scale To speed up the color transfer, we first optimize for 10k Adam iterations with the loss restricted to the coarsest scale $E_{\mathrm{transfer}}(\mathcal{R}(\mathcal{C}_i;\Theta)^{\downarrow 2^s};u_i^{\downarrow 2^s},v^{\downarrow 2^s})$ with $s = n_{\mathrm{s}} - 1$ , then optimize the SOS loss for another 10k Adam iterations. + +Number of scales $n_{\mathrm{s}}$ is set automatically to use all available scales, the coarsest resolution having sides larger than 256 for VGG19 statistics to be reliable. + +Reproducibility All our experiments have been conducted using the same SOS loss and training procedure, making our approach parameterless and fully reproducible. Our public PyTorch implementation is based on the public source codes1 for 3DGS [20] training and SPST [10]. Our code and videos of our results are available online2. + +1https://github.com/graphdeco-inria/gaussian-splatting; https://github.com/bgalerne/scaling_PA painting_style_transfer 2Code: https://github.com/JianlingWANG2021/SGSST; Videos: https://www.idpoisson.fr/galerne/sgsst/ + +# 5. Experiments + +# 5.1. Ultra-high resolution results + +Our multiscale stylization algorithm is able to transfer style details at UHR for both the content image resolution and the style image resolution. This results in unprecedentedly rich style transfer, as illustrated by Figure 3. As can be observed, minimizing the SOS loss indeed allows to decrease the style transfer loss for all scales (Figure 3 right). The approach is especially relevant when transferring the style of an UHR painting presenting style features at several scales, ranging from a specific color palette to a main curve style and to fine brushstrokes or canvas texture (see close-up views of Figure 3 and the first two lines of Figure 1). To obtain such results, combining style transfer at the largest possible number of scales is critical (see ablation in supp. mat. Figure 9). In addition, the approach is also efficient to transfer the style of a medium resolution style image to an UHR scene, as shown in the last row of Figure 1. + +To the best of our knowledge, our approach is the first to work at UHR resolution for neural style transfer of neurally rendered 3D scenes. One of the main advantages of the Gram-based loss is that it does not depend on the style resolution and, when both the content and style images grow in $O(N)$ , its complexity scales in $O(N)$ while NNFM used in ARF [48] scales in $O(N^2)$ . + +Yet, applying style transfer at such resolutions remains computationally heavy: for the garden image of size $5187 \times 3361$ (Figures 1 and 3) the style transfer takes 25.5 hours (VS 3 hours to train the initial 3DGS model), that is an 8.5 overhead factor. For an image of moderate size (Figure 4 left and middle) the style transfer and the initial 3DGS training take respectively 22 and 10 minutes, that is, the overhead factor is only 2.2. These time values were obtained using a single A100 GPU with 80 GB of memory and could be accelerated by adapting the SOS loss implementation to a multi-GPU setting. Note that high computation times are inherent to UHR style transfer: Running the 2D SPST method for the 185 garden training images takes 27 + +![](images/45042e9c71b475d242850656464a4cd131be0725f8932bc821865ce62cc5c9b8.jpg) +Figure 4. Comparison of SGSST (ours, top) with StyleGaussian [26] (middle) and ARF [48] (bottom). From left to right the content resolutions are $980 \times 545$ (train), $979 \times 546$ (truck), and $3115 \times 2076$ (counter). For the first two examples, the various outputs keep the resolution of the content, but for the HR counter scene, the output sizes are $3115 \times 2076$ for SGSST, $1600 \times 1066$ for StyleGaussian and $779 \times 519$ for ARF (see supp. mat. Figure 28 for ARF results without downscaling). Thanks to its multiscale global VGG statistics, SGSST is the most faithful method regarding style consistency. + +hours (8.9 min. per image). + +# 5.2. Comparison + +We perform a thorough comparative study using 40 3D style transfer experiments using 9 different scenes and various style images (see supp. mat.). We compare our results with the NeRF-based ARF [48] and the 3DGS-based StyleGaussian [26] algorithms, described in Section 2, using their public implementations3. + +Comparing style transfer methods is challenging because each algorithm treats the style image differently. Following high quality 2D style transfer [10, 12], our loss uses up to four scales and each scale uses five VGG layers. The UHR style images were downscaled so that the style has the same size as the content images (no upscale was applied if the style image is smaller than the content image). In contrast, ARF uses a single VGG layer from a single scale and the style image is downscaled to be twice smaller than the content input, resulting in smaller local texture. StyleGaussian reduces the style images so that it has the size $256 \times 256$ , a scale that is hardly sufficient to represent HR style images such as paintings. + +Qualitative comparison Figure 4 shows three different comparative experiments. As one can observe, the results of StyleGaussian [26] are generally not satisfying since the method fails to transfer local texture or to preserve the style + +image's color palette. ARF results better reproduce brushstroke textures, but the transfer is limited, as the method only involves a single VGG layer at a single scale. Also, NNFM does not ensure the preservation of a global color palette. In comparison, SGSST preserves the style at all considered scales. This results in color palette preservation, as well as a verifiable transfer of features of any size, from large brushstrokes (Figure 4 middle) to local grain transfer (Figure 4 right). In addition, SGSST is the only method that performs style transfer at the original resolution of the HR example of Figure 4 right. + +Quantitative comparison Even though there is no consensus for the quantitative evaluation of NST [17], following previous works, we report two different metrics for style transfer quality and texture consistency across views. Style transfer quality can be measured by the Gram loss [11]. A second metric proper to NeRF style transfer [25, 26, 33] is to check the short-term and long-term consistency of the radiance fields in terms of LPIPS [49] and RMSE between wrapped views. Since SGSST is the only method working with UHR images, when necessary we forcefully downgraded all the results to the resolution of ARF for comparison. The average of these two metrics over our 40 experiments is reported in Table 1. The Gram loss is the best for SGSST and, surprisingly, StyleGaussian achieves a lower Gram loss than ARF, which is not consistent with the qualitative evaluation, probably explained by ARF using a single VGG layer compared to five for the two other methods. In terms of consistency metrics, StyleGaussian reports to be + +
MethodTransfer qualityShort-range consistencyLong-range consistency
Gram↓LPIPS↓RMSE↓LPIPS↓RMSE↓
SGSST2.59e80.0300.0320.0550.063
StyleGaussian4.61e80.0330.0290.0500.056
ARF5.77e80.0400.0370.0720.066
+ +Table 1. Quantitative comparison. Style transfer quality and texture consistency metrics averaged over 40 experiments for SGSST, StyleGaussian, and ARF. Best results in bold, second best underlined. + +
SGSSTARFStyleGaussian
Voting results (%)66.331.62.1
+ +Table 2. Perceptual study. Summary of the 680 votes for the most style consistent result. + +the most stable approach followed by SGSST but note that this metric favors the lack of texture. + +Perceptual study To further support our results, we conducted a perceptual study comparing the 40 3D style transfer experiments. For each example, four images were displayed (at a resolution of $1280 \times 720$ ): the style image and, in a blind random order, the results of the three methods (SGSST, ARF and StyleGaussian) shown at a common viewpoint (also randomly selected). Each participant was shown ten random instances and was asked to select the result that was the most faithful to the style image. The study was conducted on the web with volunteer participants. + +A total of 68 participants took part in the test, resulting in 680 votes summarized in Table 2. This perceptual study shows that the fast style transfer performed by Style-Gaussian is consistently judged inferior in quality over ARF and SGSST. It also confirms that SGSST is far superior to ARF in terms of visual quality since $66\%$ of the participants ranked our method first for its painting style transfer quality, even when presented with results downscaled in resolution. + +# 5.3. Ablation study + +Influence of optimization parameters Our SGSST algorithm stylizes a realistic 3DGS scene by tuning the constant color components of the 3DGS Gaussians using a single 2D style transfer loss function at multiple scales (Equation (6)). In contrast, Zhang et al. [46] optimize all 3DGS parameters. This, however, necessitates a complex loss made of six different terms to avoid artefacts: a loss term enforces consistency with the original geometry via depth preservation while a preprocessing of Gaussians floaters is necessary after color transfer. + +Figure 5 shows that optimizing more parameters of the + +![](images/f107ed0c047f8a35e16ef9fde4dab2e0a14903d4717e44bf630c4099c18ff780.jpg) +Figure 5. Influence of optimization parameters: Allowing more 3DGS parameters to be optimized when minimizing the SOS loss does not improve the stylization quality and can dramatically impact the geometry. From left to right: Style image, content, SGSST default (optimization of colors), results when optimizing all spherical harmonics, results when optimizing all parameters. + +3DGS for SGSST brings no benefit. Optimizing all the spherical harmonic coefficients does not improve the result, and letting all the parameters free like in [22, 46] leads to a strong degradation of the geometry. Note that two instantaneous 3DGS style transfer methods [26, 33] are based on modifying the color features via some neural networks, but their results are not comparable to optimization-based approaches in terms of visual quality. + +Failure of coarse-to-fine strategy Our approach minimizes a style transfer loss simultaneously at all scales. This is different from the coarse-to-fine style transfer strategy that has proven successful for HR style transfer [12]. Figure 6 illustrates that this coarse-to-fine strategy fails in the context of 3DGS. Indeed, the gain obtained by optimizing at a given scale is quickly lost when optimizing the next one leading to the disappearance of large scale features. This can be explained by the fact that the 3DGS representation is a more constrained representation than pixel grids due to its sparsity. + +Color transfer via style transfer at the coarsest scale As described in Section 4.3, we first optimize the style transfer loss at the lowest resolution for $10\mathrm{k}$ iterations and + +![](images/062e384cd99470f7558e5c2c0ef740b56f6bd1c788610ba0d90d2e86f6e9e061.jpg) +Figure 6. Failure of coarse-to-fine strategy. Results of coarse-to-fine style transfer for the example of Figure 3. Each scale is initialized with the output of the previous scale. As shown by the evolution of the losses (right), when training at a new scale, the loss of the previous scales increases quickly. This explains why the large scale painting features disappear progressively and are absent after training the target UHR (see close-up details). + +![](images/09bc5f84e0d52196052696820f1705d77c8d398ab2f4fff84605218fbd5032cd.jpg) + +![](images/29067caafabea0c53c0defeaad7eccb2353db0f6e9ed76927f4d43e387edbcc5.jpg) +Figure 7. Color transfer via style transfer at the coarsest scale. The first 10k iterations at the coarsest scale allow for quick color changes in the 3DGS scene. Removing this step may lead to color artifacts in the result (top) compared to our default two-step optimization (bottom). + +![](images/1c0ca8096b488dd408aadf4bf8d7f2defb40ed52ccc8ae0c17d2da4bfeeaa775.jpg) + +![](images/d5c13c406e9db9d6c6c87f37ab1c0315ca5d7cf4a018be6dd2c0be85364eccee.jpg) + +![](images/f5b3ec75dc83ffdc37180a102b13a153c1c804b5350d76fe19efd335a245b060.jpg) + +![](images/2670d0503eb4b5450b93caa17d0bdc62f0503ba8dab3d775b77e2f4b8269779d.jpg) +Figure 8. Limited texture representation due to low Gaussian density of the initial 3DGS. From left to right: Original scene, 3DGS reconstruction, stylization of the 3DGS scene. + +then optimize the SOS loss for another 10k iterations. Figure 7 illustrates that these first iterations are necessary for a faithful style color palette reproduction. + +# 6. Discussion and limitations + +Texture representation The texture representation within a 3DGS scene depends on the density of Gaussians and may be limited in low density areas. Isolated Gaussians can sometimes be spotted as illustrated by Figure 8. + +Large computation time Depending on the resolution, SGSST requires from several minutes to several hours of computation. On the other hand, fast 3DGS stylization approaches [26, 33] do not reach a satisfactory visual quality. + +Content-style mismatch As said earlier, the Gram loss has the advantage of being independent of the style's image resolution. Also, it enables a faithful transfer of global statistics of the style image, such as its color palette. This important feature is, nevertheless, counterproductive when + +![](images/3643e4be91fb4f374d2ed4bfa5dc376d53a768566e6d90a7b81562192ea5d68b.jpg) + +the style and content images strongly mismatch, leading to color bleeding or texturing of flat areas. Other controls can be added to mitigate these artefacts [12] and it was shown that these controls are effective for 3DGS scenes [46]. + +# 7. Conclusion + +In this work we presented SGSST, a method that, for the first time, enables UHR 3DGS style transfer. To that aim, among other innovations, we introduced the simultaneously optimized scales (SOS) loss. Our qualitative, quantitative and perceptual studies show that SGSST obtains superior style transfer quality than state of the art, even after reducing our results' resolution for a fair comparison with methods that do not reach UHR. Such high quality UHR results necessitate a large computation time that, nevertheless, remains comparable with that of UHR 3DGS training. + +This work opens the way to several research directions. A first challenge is to produce equally high quality style transfer with a faster algorithm based on UST. A second more exploratory direction is to investigate geometry style transfer for 3DGS by designing adapted regularization to avoid the caveats depicted by Figure 5. + +Acknowledgements: B. Galerne and L. Raad acknowledge the support of the project MISTIC (ANR-19-CE40-005). + +# References + +[1] Jie An, Tao Li, Haozhi Huang, Li Shen, Xuan Wang, Yongyi Tang, Jinwen Ma, Wei Liu, and Jiebo Luo. Real-time universal style transfer on high-resolution images via zero-channel pruning. arXiv preprint arXiv:2006.09029, 2020. 2 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470–5479, 2022. 2 +[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022, pages 333-350, Cham, 2022. Springer Nature Switzerland. 2 +[4] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2 +[5] Zhe Chen, Wenhai Wang, Enze Xie, Tong Lu, and Ping Luo. Towards ultra-resolution neural style transfer via thumbnail instance normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 393-400, 2022. 2 +[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1475-1484, 2022. 2 +[7] Tai-Yin Chiu and Danna Gurari. Iterative feature transformation for fast and versatile universal style transfer. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2 +[8] Valentin De Bortoli, Agnès Desolneux, Alain Durmus, Bruno Galerne, and Arthur Leclaire. Maximum entropy methods for texture synthesis: Theory and practice. SIAM Journal on Mathematics of Data Science, 3(1):52-82, 2021. 2 +[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2 +[10] Bruno Galerne, Lara Raad, José Lezama, and Jean-Michel Morel. Scaling Painting Style Transfer. Computer Graphics Forum, 2024. 2, 3, 4, 5, 6 +[11] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 2, 3, 6 +[12] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3, 4, 6, 7, 8 + +[13] Nicolas Gonthier, Yann Gousseau, and Saïd Ladjal. High-resolution neural texture synthesis with long-range constraints. Journal of Mathematical Imaging and Vision, 64 (5):478-492, 2022. 2 +[14] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9412-9420, 2021. 2, 3 +[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[16] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18342–18352, 2022. 2 +[17] Eleftherios Ioannou and Steve Maddock. Evaluation in neural style transfer: A review. Computer Graphics Forum, 43 (6):e15165, 2024. 6 +[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711, 2016. 2 +[19] Hyunyoung Jung, Seonghyeon Nam, Nikolaos Sarafianos, Sungwoo Yoo, Alexander Sorkine-Hornung, and Rakesh Ranjan. Geometry transfer for stylizing radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8565-8575, 2024. 2 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. 2, 3, 4, 5 +[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 3 +[22] Aron Samuel Kovács, Pedro Hermosilla, and Renata G. Rraidou. $\mathcal{G}$ -style: Stylized gaussian splatting. Computer Graphics Forum, n/a(n/a):e15259, 2024. 2, 3, 7 +[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3809-3817, 2019. 2 +[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30:386-396, 2017. 2 +[25] Kunhao Liu, Fangneng Zhan, Yiwen Chen, Jiahui Zhang, Yingchen Yu, Abdulmotaleb El Saddik, Shijian Lu, and Eric P. Xing. Stylerf: Zero-shot 3d style transfer of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8338-8348, 2023. 2, 6 + +[26] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. StyleGaussian: Instant 3d style transfer with Gaussian splatting, 2024. 2, 3, 6, 7, 8 +[27] Yang Lu, Song-chun Zhu, and Ying Nian Wu. Learning frame models using cnn filters. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 2 +[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 2, 3 +[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), 2022. 2 +[30] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[32] Eric Risser, Pierre Wilmot, and Connelly Barnes. Stable and controllable neural texture synthesis and style transfer using histogram losses, 2017. 2, 3 +[33] Abhishek Saroha, Marija Gladkova, Cecilia Curreli, Tarun Yenamandra, and Daniel Cremers. Gaussian splatting in style, 2024. 2, 3, 6, 7, 8 +[34] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. Graph., 36(5), 2017. 2, 3 +[36] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2 +[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2, 3 +[38] Xavier Snelgrove. High-resolution multi-scale neural texture synthesis. In SIGGRAPH Asia 2017 Technical Briefs, New York, NY, USA, 2017. Association for Computing Machinery. 2, 4 +[39] Ondrej Texler, David Futschik, Jakub Fišer, Michal Lukáč, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Arbitrary style transfer using neurally-guided patch-based synthesis. Computers & Graphics, 87:62-71, 2020. 2 +[40] D. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In ICML, pages 1349-1357, 2016. 2 + +[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[42] Jonathan Vacher, Aida Davila, Adam Kohn, and Ruben Coen-Cagli. Texture interpolation for probing visual perception. In Advances in Neural Information Processing Systems, pages 22146-22157. Curran Associates, Inc., 2020. 2 +[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[44] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. Proceedings of the AAAI Conference on Artificial Intelligence, 37(3): 2742-2750, 2023. 2 +[45] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2 +[46] Dingxi Zhang, Zhuoxun Chen, Yu-Jie Yuan, Fang-Lue Zhang, Zhenliang He, Shiguang Shan, and Lin Gao. StylizedGS: Controllable stylization for 3d Gaussian splatting, 2024. 2, 3, 7, 8 +[47] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. CoARF: Controllable 3D Artistic Style Transfer for Radiance Fields. In 2024 International Conference on 3D Vision (3DV), pages 612-622, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2 +[48] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. ARF: Artistic radiance fields. In ECCV 2022, 2022. 2, 3, 5, 6 +[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 \ No newline at end of file diff --git a/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/images.zip b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..adb891ff81f9628a66635cc298b143dee4020765 --- /dev/null +++ b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3df4f52d126b273fecf6002fc9459375d065eea9f1e22d58ce18c67acd41246 +size 1100989 diff --git a/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/layout.json b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ca11930bab4634a1bc9d09d67d466caf6ab42824 --- /dev/null +++ b/2025/SGSST_ Scaling Gaussian Splatting Style Transfer/layout.json @@ -0,0 +1,8760 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 460, + 121 + ], + "type": "text", + "content": "SGSST: Scaling Gaussian Splatting Style Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "spans": [ + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "content": "Bruno Galerne" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "content": ", Jianling Wang" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "content": ", Lara Raad" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "content": ", and Jean-Michel Morel" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "content": "Université d'Orléans, Université de Tours, CNRS, IDP, UMR 7013, Orléans, France \n" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 143, + 485, + 183 + ], + "type": "text", + "content": "Institut Universitaire de France (IUF)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 184, + 552, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 184, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 184, + 552, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 184, + 552, + 198 + ], + "type": "text", + "content": "Instituto de Ingeniería Eléctrica, Facultad de Ingeniería, Universidad de la Republica " + }, + { + "bbox": [ + 69, + 184, + 552, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 184, + 552, + 198 + ], + "type": "text", + "content": "City University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 222, + 331 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 222, + 331 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 222, + 331 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 222, + 331 + ], + "type": "image", + "image_path": "de06ada5b3d524820e24de71826a3553f25248cccaea80e5dfcf6e9589b636e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 331, + 222, + 438 + ], + "blocks": [ + { + "bbox": [ + 58, + 331, + 222, + 438 + ], + "lines": [ + { + "bbox": [ + 58, + 331, + 222, + 438 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 222, + 438 + ], + "type": "image", + "image_path": "2b71d6ab9ed5269de7ecfc03ccfe4dad6f25f38d035b64203aeac34e6151ba19.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 439, + 222, + 545 + ], + "blocks": [ + { + "bbox": [ + 58, + 439, + 222, + 545 + ], + "lines": [ + { + "bbox": [ + 58, + 439, + 222, + 545 + ], + "spans": [ + { + "bbox": [ + 58, + 439, + 222, + 545 + ], + "type": "image", + "image_path": "95e7be46152d6fe6611224de0ffcde9fc03c40238f151c3470f6ced37493b867.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 555, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 555, + 589 + ], + "type": "text", + "content": "Figure 1. Various ultra-high definition style transfers of a Gaussian splattering 3D scene. SGSST transfers a very large set of global style statistics of an image to a 3DGS scene by minimizing a tailored multiscale SOS loss, yielding 3D style transfer of superior quality and at unprecedented high resolution (images have size " + }, + { + "bbox": [ + 55, + 555, + 555, + 589 + ], + "type": "inline_equation", + "content": "5187 \\times 3361" + }, + { + "bbox": [ + 55, + 555, + 555, + 589 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 223, + 224, + 387, + 331 + ], + "blocks": [ + { + "bbox": [ + 223, + 224, + 387, + 331 + ], + "lines": [ + { + "bbox": [ + 223, + 224, + 387, + 331 + ], + "spans": [ + { + "bbox": [ + 223, + 224, + 387, + 331 + ], + "type": "image", + "image_path": "d77b5a44a8876ebe874e29b631e9e3fd81680a9beda2be069e017c7ac6ed352c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 331, + 387, + 438 + ], + "blocks": [ + { + "bbox": [ + 223, + 331, + 387, + 438 + ], + "lines": [ + { + "bbox": [ + 223, + 331, + 387, + 438 + ], + "spans": [ + { + "bbox": [ + 223, + 331, + 387, + 438 + ], + "type": "image", + "image_path": "47abad2cd38069a40dbe926a18e5d9ffa25d9d29c54c101edd22daeed9d03e8d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 439, + 387, + 545 + ], + "blocks": [ + { + "bbox": [ + 223, + 439, + 387, + 545 + ], + "lines": [ + { + "bbox": [ + 223, + 439, + 387, + 545 + ], + "spans": [ + { + "bbox": [ + 223, + 439, + 387, + 545 + ], + "type": "image", + "image_path": "b801c471712eeb801a57ac683844d4835f174f5d9811af696b5dae513c602e05.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 388, + 224, + 552, + 331 + ], + "blocks": [ + { + "bbox": [ + 388, + 224, + 552, + 331 + ], + "lines": [ + { + "bbox": [ + 388, + 224, + 552, + 331 + ], + "spans": [ + { + "bbox": [ + 388, + 224, + 552, + 331 + ], + "type": "image", + "image_path": "53dfb53fae69cbf3dbfaab5f83bde197be0e5b7b0c57c566d8f53c338af013d9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 388, + 331, + 552, + 438 + ], + "blocks": [ + { + "bbox": [ + 388, + 331, + 552, + 438 + ], + "lines": [ + { + "bbox": [ + 388, + 331, + 552, + 438 + ], + "spans": [ + { + "bbox": [ + 388, + 331, + 552, + 438 + ], + "type": "image", + "image_path": "529a8b274e630287f10f7bff4e84d7e7e3594158eac76e78ddc1220fc5e522b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 388, + 439, + 552, + 545 + ], + "blocks": [ + { + "bbox": [ + 388, + 439, + 552, + 545 + ], + "lines": [ + { + "bbox": [ + 388, + 439, + 552, + 545 + ], + "spans": [ + { + "bbox": [ + 388, + 439, + 552, + 545 + ], + "type": "image", + "image_path": "66774d176e9ea7e04e597358a655248b6e4e5281adf444020159f9f62c775fab.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 152, + 598, + 200, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 598, + 200, + 611 + ], + "spans": [ + { + "bbox": [ + 152, + 598, + 200, + 611 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Applying style transfer to a full 3D environment is a challenging task that has seen many developments since the advent of neural rendering. 3D Gaussian splatting (3DGS) has recently pushed further many limits of neural rendering in terms of training speed and reconstruction quality. This work introduces SGSST: Scaling Gaussian Splatting Style Transfer, an optimization-based method to apply style" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": "transfer to pretrained 3DGS scenes. We demonstrate that a new multiscale loss based on global neural statistics, that we name SOS for Simultaneously Optimized Scales, enables style transfer to ultra-high resolution 3D scenes. Not only SGSST pioneers 3D scene style transfer at such high image resolutions, it also produces superior visual quality as assessed by thorough qualitative, quantitative and perceptual comparisons." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26535" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 294, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 294, + 248 + ], + "type": "text", + "content": "Dealing with ultra-high resolution (UHR) rendering is capital for AR/VR applications. Indeed, when navigating in a 3D environment the user only sees a partial field of view of the environment. This adds challenging issues for applying style transfer to 3D environments, that is, transferring the visual characteristics of an image such as a painting to a 3D scene. Indeed, while the general scene should convey the global color palette of the style painting, when getting closer to objects in a stylized environment the user should expect to see fine painting patterns such as brushstrokes. However, in the current state of the art, the user quickly encounters resolution-limited content in the form of blurry interpolated features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 249, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 295, + 441 + ], + "type": "text", + "content": "Ever since the seminal work of Mildenhall et al. [28], neural radiance fields (NeRF) have seen many improvements. Several 3D scene representations have been proposed for improving the quality and resolution of the reconstructed scenes as well as easing the training, e.g. [2, 3, 9, 29]. The recent 3D Gaussian splatting (3DGS) [20] has introduced an efficient high-resolution (HR) scene representation which has stirred much interest [45]. Both frameworks have stirred attempts to 3D style transfer algorithms [6, 16, 22, 25, 26, 33, 46, 48], but these methods have so far produced medium resolution outputs. They do not faithfully transport high resolution multiscale textures such as those present in paintings. Motivated by a recent neural style transfer (NST) [11] contribution tailored for UHR images [10], i.e. with resolution larger than " + }, + { + "bbox": [ + 55, + 249, + 295, + 441 + ], + "type": "inline_equation", + "content": "4\\mathrm{k}" + }, + { + "bbox": [ + 55, + 249, + 295, + 441 + ], + "type": "text", + "content": " images, we show that 3DGS can be leveraged for UHR style transfer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 441, + 266, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 441, + 266, + 453 + ], + "spans": [ + { + "bbox": [ + 67, + 441, + 266, + 453 + ], + "type": "text", + "content": "The contributions of this work are the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 457, + 295, + 578 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 457, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 295, + 493 + ], + "type": "text", + "content": "- We introduce SOS, a Simultaneously Optimized Scales loss expressed in a single parameterless and explainable formula." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 494, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 494, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 494, + 295, + 529 + ], + "type": "text", + "content": "- By solely optimizing the SOS loss, we reach UHR for 3DGS style transfer and we scale Gaussian Splatting Style Transfer by a four times resolution gain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 529, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 295, + 578 + ], + "type": "text", + "content": "- Superior quality transfer: By transferring a very large set of global style statistics, we obtain superior style transfer quality even at HR resolution, as confirmed by a comparative perceptual study." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "In short, our approach is the first method that allows UHR style transfer directly to 3DGS. It produces high visual quality results and relies on optimizing a single multiscale loss. The simplicity of our approach ensures its reproducibility. Being optimization-based, SGSST's main limitation is a fairly large training time that is two to eight times longer than the initial 3DGS training depending on the image resolution. Yet, considering the high quality of the results and their reproducibility, this contribution is valuable for AR/VR applications that require high visual quality for their user experience." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 247 + ], + "type": "text", + "content": "Neural style transfer In the seminal work of Gatys et al. [11] NST is formulated as an optimization problem minimizing the distances between Gram matrices of VGG [37] features. Even though other VGG statistics have been considered, almost all subsequent style transfer and texture synthesis methods rely on VGG [8, 13, 14, 27, 32, 35, 42]. To accelerate style transfer, several methods [18, 40, 41] have attempted to train feed-forward networks approximately minimizing the Gram loss [11]. However, these approximate methods require learning a new model for each style type. This latter limitation has been addressed by fast Universal Style Transfer (UST) approaches [4, 7, 15, 23, 24, 30, 36] that use VGG feature decoders." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 553, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 356 + ], + "type": "text", + "content": "For HR images, coarse-to-fine multiscale strategies [12, 13, 38] have proved effective, but still face limitations due the high GPU memory usage of VGG statistics. Fast HR alternatives [1, 5, 39, 43, 44] do exist but generally suffer from artifacts and struggle to capture the full style complexity. Recently, SPST [10] proposed an implementation of the Gatys et al. method adapted to UHR (larger that " + }, + { + "bbox": [ + 313, + 247, + 553, + 356 + ], + "type": "inline_equation", + "content": "4\\mathrm{k}" + }, + { + "bbox": [ + 313, + 247, + 553, + 356 + ], + "type": "text", + "content": ") images. The visual quality of SPST's results is superior, at the cost of a long optimization time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 373, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 553, + 529 + ], + "type": "text", + "content": "Style transfer for neural radiance fields NeRFs [28] have completely redefined the field of 3D scene modeling and novel view synthesis. Editing the visual aspect of NeRFs via style transfer has quickly been addressed [6, 16, 25, 48], usually by fine tuning a pretrained NeRF representation using a style transfer loss, or training an additional fast style transfer module. ARF [48] is a notable exception: It uses Nearest Neighbor Feature Matching (NNFM) for fine tuning a plenoxel radiance field [9], produces high-quality results at moderate resolution, and is the base model for other methods [19, 47]. While these works paved the way for radiance field style transfer, they are all limited in input and output image resolution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "Style transfer for Gaussian splatting A few recent contributions show that 3DGS is a promising framework for 3D scene style transfer. Saroha et al [33] propose a solution for universal style transfer of a given 3DGS scene. The method processes the colors of Gaussians with a tiny MLP trained using fast AdaIN [15] and relying on a multiresolution hash grid representation [29]. StyleGaussian [26] is a concurrent approach that relies on transferring encoding of VGG features to each Gaussian and applying AdaIN to these features. The new Gaussian features are then decoded into an RGB color by processing the K-nearest neighbors of the Gaussians. After training, both methods allow for instantaneous stylization with any style image, but the visual quality of the results is quite low, since high-resolution" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26536" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 278, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 278, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 278, + 83 + ], + "type": "text", + "content": "details such as brushstroke patterns are not transferred." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "type": "text", + "content": "StylizedGS [46] is an optimization-based method that extends ARF [48] to 3DGS. It uses a training loss made of six terms combined with a color transfer preprocessing, the loss being designed to minimize changes in the 3DGS geometry while letting the style evolve via VGG NN matching. " + }, + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "type": "text", + "content": " -Style [22] also is an optimization based approach that further uses a CLIP-based loss [31]. While these approaches produce slightly better results than ARF [48], both methods are limited in content and style resolution due to the use of nearest-neighbor matching of VGG features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 275 + ], + "type": "text", + "content": "None of the current state of the art deals with HR style transfer. The stylization is either fast but very approximate due to AdaIn [26, 33] or unable to use HR style images due to NNFM [22, 46, 48]. To the best of our knowledge, SGSST is the first procedure allowing high-quality transfer for 3DGS that is trained and rendered at UHR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 141, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 141, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 141, + 298 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 236, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 236, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 236, + 319 + ], + "type": "text", + "content": "3.1. Gaussian splatting representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": "Starting from a multiview training set of images " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{u_i\\}_{i=1}^{N_{\\mathrm{views}}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " of a static scene accompanied with corresponding calibrated cameras " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_i\\}_{i=1}^{N_{\\mathrm{views}}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " computed by structure from motion [34], the 3DGS algorithm [20] trains a set of colored 3D Gaussians " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_j\\}_{j=1}^{N_{\\mathrm{Gaussians}}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " so that they represent the 3D scene from any camera position. Each Gaussian " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " is represented by a finite set of features: a center position " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mu_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ", a covariance matrix " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\Sigma_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " (encoded by a scaling diagonal matrix and a rotation matrix), an opacity " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\alpha_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " and a view-dependent color function " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": ". These parameters are used in a volumetric renderer that determines the color by summing the contribution of each Gaussian that intersects a ray with direction " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "(\\theta, \\varphi)" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": "-blending (see e.g. [28]). The contribution of a Gaussian is its color " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "c_j(\\theta, \\varphi)" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " multiplied by an opacity " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\sigma_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " defined as the maximal opacity " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\alpha_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " times the unnormalized Gaussian density at the ray position [20]. Thus, the resulting color " + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 323, + 296, + 529 + ], + "type": "text", + "content": " for the ray is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 536, + 296, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 536, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 107, + 536, + 296, + 571 + ], + "type": "interline_equation", + "content": "C = \\sum_ {j = 1} ^ {N} c _ {j} (\\theta , \\varphi) \\sigma_ {j} \\prod_ {k = 1} ^ {j - 1} (1 - \\sigma_ {k}) \\tag {1}", + "image_path": "ba2c6fac7a927f5a027d096b7cb158808e0f64ca114512f23f79988adfe5033e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": "where the sum is over all Gaussians intersecting the ray. The color function " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "c_{j}(\\theta ,\\varphi)" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " associated with a Gaussian depends on the spherical direction " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "(\\theta ,\\varphi)" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " through an order 3 spherical harmonics polynomial function given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 635, + 296, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 86, + 635, + 296, + 669 + ], + "type": "interline_equation", + "content": "c _ {j} (\\theta , \\varphi) = c _ {j, 0} + \\sum_ {\\ell = 1} ^ {3} \\sum_ {m = 1} ^ {2 \\ell + 1} c _ {j, \\ell , m} Y _ {\\ell , m} (\\theta , \\varphi) \\tag {2}", + "image_path": "b079cd8b90355a832cafa9ca6259cfe3dad73f8ccebb40f833a4c24b4f4e2275.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "where the vectors " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{j,0}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{j,\\ell,m}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " are in " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "Y_{\\ell,m}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " form a basis of the spherical harmonics polynomials of degree " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". In short, " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{j,0} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " is the main color and the ad" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "ditional coefficients " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "c_{j,\\ell,m}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " encode smooth variations of this color when the viewing angle changes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "text", + "content": "Like for NeRF, the key ingredient of the 3DGS parametrization is the differentiable rendering function " + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathcal{C};\\Theta)" + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "text", + "content": " that renders a view of the scene for any camera " + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "text", + "content": " given the scene parameters " + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\Theta = \\{(\\mu_j,\\Sigma_j,\\alpha_j,c_{j,0},(c_{j,\\ell ,m})_{\\ell ,m})\\}_{j = 1}^{N_{\\mathrm{Gaussians}}}" + }, + { + "bbox": [ + 313, + 96, + 555, + 180 + ], + "type": "text", + "content": ". This differentiable rendering function allows to train the Gaussian parameters to minimize the reconstruction error" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 188, + 555, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 188, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 331, + 188, + 555, + 220 + ], + "type": "interline_equation", + "content": "\\min _ {\\Theta} \\frac {1}{N _ {\\mathrm {v i e w s}}} \\sum_ {i = 1} ^ {N _ {\\mathrm {v i e w s}}} E _ {\\mathrm {r e c o n s t r u c t i o n}} (\\mathcal {R} (\\mathcal {C} _ {i}; \\Theta); u _ {i}) \\qquad (3)", + "image_path": "61fb467023ae202c9a5ba6c4e5b1078f16517e2097b5075a637dbbfff1345a01.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{reconstruction}}" + }, + { + "bbox": [ + 313, + 229, + 555, + 277 + ], + "type": "text", + "content": " is a 2D image comparison error (such as a combination of mean square error and SSIM [20]). The minimization is conducted using Adam [21] for 30k iterations by randomly picking one view at each iteration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 285, + 499, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 499, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 499, + 297 + ], + "type": "text", + "content": "3.2. Style transfer loss for UHR images" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": "Our approach relies on optimizing VGG19 [37] feature statistics as initially proposed by Gatys et al. [11]. Content consistency is loosely monitored by preservation of the feature layer " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{c}} = \\mathrm{ReLU\\_4\\_2}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": " while style transfer is imposed by matching spatial statistics of five VGG19 layers, namely the set " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{s}} = \\{\\mathrm{ReLU\\_k\\_1}, k \\in \\{1, 2, 3, 4, 5\\}\\}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": ". The statistics of interest of the feature response " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "V^{L}(w)" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": " of an image " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": " at some layer " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "L \\in \\mathcal{L}_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": " having " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{c}}^{L}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": " feature channels are its Gram matrix " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\operatorname{Gram}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L} \\times n_{\\mathrm{c}}^{L}}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": ", its spatial mean vector " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\operatorname{mean}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L}}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": ", and its standard deviation vector " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\operatorname{std}(V^{L}(w)) \\in \\mathbb{R}^{n_{\\mathrm{c}}^{L}}" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": ". Given a content image " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": " and a style image " + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 303, + 556, + 452 + ], + "type": "text", + "content": ", we consider the loss function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 461, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 461, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 328, + 461, + 555, + 475 + ], + "type": "interline_equation", + "content": "E _ {\\text {t r a n s f e r}} (x; u, v) = E _ {\\text {c o n t e n t}} (x; u) + E _ {\\text {s t y l e}} (x; v) \\tag {4}", + "image_path": "0cea58838b3b4ec70a38b6b237b9e5e0fe075e401b6a11b34bbbefa543f7eb21.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{content}}(x;u) = \\lambda_{\\mathrm{c}}\\left\\| V^{L_{\\mathrm{c}}}(\\boldsymbol {x}) - V^{L_{\\mathrm{c}}}(\\boldsymbol {u})\\right\\|^{2}" + }, + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{c}} > 0" + }, + { + "bbox": [ + 313, + 484, + 554, + 510 + ], + "type": "text", + "content": " and the style loss is defined by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 367, + 520, + 555, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 520, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 367, + 520, + 555, + 546 + ], + "type": "interline_equation", + "content": "E _ {\\text {s t y l e}} (x; v) = \\sum_ {L \\in \\mathcal {L} _ {\\mathrm {s}}} E _ {\\text {s t y l e}} ^ {L} (x; v). \\tag {5}", + "image_path": "915d7f960644cb3d34f7dbd6f091b4d87cea7dbc0470a5e36c784000a4d9f774.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{style}}^{L}(x;v)" + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "text", + "content": " is a linear combination of the mean square error between the VGG19 statistics of " + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "inline_equation", + "content": "V^{L}(x)" + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "text", + "content": " and the one of the style features " + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "inline_equation", + "content": "V^{L}(v)" + }, + { + "bbox": [ + 313, + 555, + 555, + 640 + ], + "type": "text", + "content": " [10]. While only the Gram matrices were originally used [11], it has been shown that adding control for the mean and standard deviation corrects some style transfer color artefacts [10] previously identified in the literature [14, 32, 35]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "To obtain high-quality style transfer for HR images one needs to optimize the style transfer loss using several scales and a coarse-to-fine approach [12]. Indeed, if one applies style transfer on the highest resolution only, the changes within the content image are limited to local texture and the results does not convey a painting aspect. Due to the use of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26537" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 556, + 302 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 556, + 302 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 556, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 556, + 302 + ], + "type": "image", + "image_path": "413a883f11bd3798e04372d1cb863202ca3a20ae935a9c76d1f4592d6658bb0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 310, + 555, + 356 + ], + "lines": [ + { + "bbox": [ + 54, + 310, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 555, + 356 + ], + "type": "text", + "content": "Figure 2. Overview of SGSST. Starting from a pretrained realistic 3DGS scene [20], we optimize the colors of each Gaussian using the new multiscale SOS loss (involving " + }, + { + "bbox": [ + 54, + 310, + 555, + 356 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{s}} = 3" + }, + { + "bbox": [ + 54, + 310, + 555, + 356 + ], + "type": "text", + "content": " scales in the illustration). Computing the gradient w.r.t. the loss is feasible for UHR images thanks to the SPST partition-based implementation [10]. Multiscale gradient stacking is used at the node of the rendered image to perform only one backpropagation per iteration through the 3DGS rendering pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "type": "text", + "content": "VGG19 features, computing the loss " + }, + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{transfer}}(x;u,v)" + }, + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "type": "text", + "content": " and its gradient with respect to (w.r.t.) " + }, + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 375, + 296, + 449 + ], + "type": "text", + "content": " via backpropagation is memory prohibiting for UHR images. However, using a grid partition and local loss backpropagation based on precomputed global statistics, SPST [10] allows for the exact evaluation of this gradient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 277, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 277, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 277, + 475 + ], + "type": "text", + "content": "4. Scaling Gaussian splatting style transfer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "type": "text", + "content": "A complete overview of our SGSST algorithm is given in Figure 2. Starting from a realistic 3DGS representation [20], we optimize the colors of each Gaussian using a multiscale style transfer loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 538, + 239, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 239, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 239, + 551 + ], + "type": "text", + "content": "4.1. Stylizing the 3DGS representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": "By changing the reconstruction loss of Equation (3) with a style transfer loss for the input style image " + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": ", one can hope to stylize a realistic 3DGS. However, given the complexity of the 3DGS representation and the many interacting parameters, it is not such an easy task to alter the 3DGS aspects without loosing the content geometry. Our solution is to only optimize for the constant color components " + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{color}} = \\{c_{j,0}\\}_{j=1}^{N_{\\mathrm{Gaussians}}}" + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": " of the Gaussians and simply freeze all the other parameters " + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{init}}" + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": " given by the initial realistic 3DGS training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We experimentally found that this robust approach ensures a rich style transfer and fully preserves the 3D geometry. Indeed, fixing all the Gaussian parameters except for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 412 + ], + "type": "text", + "content": "the main color component " + }, + { + "bbox": [ + 313, + 376, + 555, + 412 + ], + "type": "inline_equation", + "content": "c_{j,0}" + }, + { + "bbox": [ + 313, + 376, + 555, + 412 + ], + "type": "text", + "content": " preserves the scene geometry, as the location and size of the Gaussians are being kept (see Section 5.3 for ablation experiments)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 479, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 479, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 479, + 430 + ], + "type": "text", + "content": "4.2. Multiscale Style Transfer Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 436, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 459 + ], + "type": "text", + "content": "We introduce the Simultaneously Optimized Scales (SOS) loss defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 464, + 554, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 464, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 323, + 464, + 554, + 496 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {S O S}} (x; u, v) = \\frac {1}{n _ {\\mathrm {s}}} \\sum_ {s = 0} ^ {n _ {\\mathrm {s}} - 1} E _ {\\text {t r a n s f e r}} \\left(x ^ {\\downarrow 2 ^ {s}}; u ^ {\\downarrow 2 ^ {s}}, v ^ {\\downarrow 2 ^ {s}}\\right) \\tag {6}", + "image_path": "f61e220b6d99e78e0447b1ba0d457d546a06533b52807e06c8673de6c394f14e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{s}} \\geq 1" + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "text", + "content": " is the number of considered scales and " + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "inline_equation", + "content": "u^{\\downarrow 2^{s}}" + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "text", + "content": " denotes an image " + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "text", + "content": " downscaled by a " + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "inline_equation", + "content": "2^{s}" + }, + { + "bbox": [ + 313, + 502, + 554, + 622 + ], + "type": "text", + "content": " factor. This SOS loss (6) enables style transfer simultaneously at all scales. A somewhat similar approach was proposed for 2D texture synthesis [38] but, as already mentioned, multiscale 2D style transfer is generally conducted using a coarse-to-fine strategy [12]. However, for 3DGS we observed that the initial configuration had only little influence on the final result, making the coarse-to-fine strategy ineffective for multiscale style transfer (See Section 5.3)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "content": "In the end, as illustrated by Figure 2, the stylization of UHR 3DGS is conducted by solving for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 652, + 554, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 652, + 554, + 684 + ], + "spans": [ + { + "bbox": [ + 345, + 652, + 554, + 684 + ], + "type": "interline_equation", + "content": "\\min _ {\\Theta_ {\\text {c o l o r}}} \\frac {1}{N _ {\\text {v i e w s}}} \\sum_ {i = 1} ^ {N _ {\\text {v i e w s}}} E _ {\\text {S O S}} \\left(\\mathcal {R} \\left(\\mathcal {C} _ {i}; \\Theta\\right); u _ {i}, v\\right), \\tag {7}", + "image_path": "1443635311bf8cd4166693d9342f6c79eadcaee76a53b30da78e98b604af31de.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " stands for the 3DGS parameters obtained by replacing the color components of " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{init}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " by the values of the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26538" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 180, + 168 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 180, + 168 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 180, + 168 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 180, + 168 + ], + "type": "image", + "image_path": "eeec0944ae3ccc579a39d2848e9d071444014c7099aefb6325676e47d2aa5859.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "text", + "content": "Figure 3. UHR 3DGS style transfer. SGSST allows for the multiscale style transfer of 3DGS scenes at UHR. From left to right: Style image, one UHR stylized view, three magnified details, and evolution of the SOS loss and each style transfer loss that contributes to it. We first optimize the transfer loss for the coarsest scale (yellow curve) for " + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "text", + "content": " iterations and then optimize for another " + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "text", + "content": " iterations the SOS loss (light blue curve), namely the mean of the four transfer losses. Images sizes are " + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "inline_equation", + "content": "5187\\times 3361" + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "text", + "content": " for content and " + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "inline_equation", + "content": "4230\\times 3361" + }, + { + "bbox": [ + 54, + 177, + 555, + 222 + ], + "type": "text", + "content": " for style." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 180, + 70, + 329, + 168 + ], + "blocks": [ + { + "bbox": [ + 180, + 70, + 329, + 168 + ], + "lines": [ + { + "bbox": [ + 180, + 70, + 329, + 168 + ], + "spans": [ + { + "bbox": [ + 180, + 70, + 329, + 168 + ], + "type": "image", + "image_path": "da26b8c99258f4ebc42db88da241f81055755c6edb98f152dc0383612e7c17ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 419, + 168 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 419, + 168 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 419, + 168 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 419, + 168 + ], + "type": "image", + "image_path": "a763e2db65789129d44dea624b26750219869325e16f115b9491821249ee1fd0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 421, + 70, + 553, + 168 + ], + "blocks": [ + { + "bbox": [ + 421, + 70, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 421, + 70, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 421, + 70, + 553, + 168 + ], + "type": "image", + "image_path": "9d3b5195b95b89d4bafac744b4c4b51800ba3bfe578d03a566bc494ed431d284.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 243, + 173, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 173, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 173, + 255 + ], + "type": "text", + "content": "optimization variable " + }, + { + "bbox": [ + 55, + 243, + 173, + 255 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{color}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 262, + 187, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 187, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 187, + 274 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": "Multiscale gradient stacking The 2D SPST method [10] provides the gradient of each term " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{transfer}}(x^{\\downarrow 2^{s}};u^{\\downarrow 2^{s}},v^{\\downarrow 2^{s}})" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": " w.r.t. to the input " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "x^{\\downarrow 2^{s}} = \\mathcal{R}(\\mathcal{C}_{i};\\Theta)^{\\downarrow 2^{s}}" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": ". We first backpropagate each of these gradients through the downscaling operator and stack them at the level of the rendered image " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathcal{C}_i;\\Theta)" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": ". When each of the " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": " scales has been treated, the stacked gradient is equal to the gradient of the full loss " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{SOS}}(\\mathcal{R}(\\mathcal{C}_i;\\Theta);u_i,v)" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": " w.r.t. to the rendered image " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathcal{C}_i;\\Theta)" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": ". This gradient is then backpropagated through the Gaussian rendering pipeline to obtain the gradient w.r.t. the Gaussian colors " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{color}}" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": " (see both orange arrows in Figure 2). Using this strategy we only backpropagate one time per iteration through the 3DGS rendering pipeline instead of " + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 281, + 296, + 448 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "type": "text", + "content": "Color transfer via style transfer at the coarsest scale To speed up the color transfer, we first optimize for 10k Adam iterations with the loss restricted to the coarsest scale " + }, + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{transfer}}(\\mathcal{R}(\\mathcal{C}_i;\\Theta)^{\\downarrow 2^s};u_i^{\\downarrow 2^s},v^{\\downarrow 2^s})" + }, + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "type": "inline_equation", + "content": "s = n_{\\mathrm{s}} - 1" + }, + { + "bbox": [ + 55, + 463, + 296, + 523 + ], + "type": "text", + "content": ", then optimize the SOS loss for another 10k Adam iterations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "type": "text", + "content": "Number of scales " + }, + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 536, + 296, + 572 + ], + "type": "text", + "content": " is set automatically to use all available scales, the coarsest resolution having sides larger than 256 for VGG19 statistics to be reliable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": "Reproducibility All our experiments have been conducted using the same SOS loss and training procedure, making our approach parameterless and fully reproducible. Our public PyTorch implementation is based on the public source codes1 for 3DGS [20] training and SPST [10]. Our code and videos of our results are available online2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": "1https://github.com/graphdeco-inria/gaussian-splatting; https://github.com/bgalerne/scaling_PA painting_style_transfer 2Code: https://github.com/JianlingWANG2021/SGSST; Videos: https://www.idpoisson.fr/galerne/sgsst/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 241, + 395, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 395, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 395, + 255 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 262, + 470, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 470, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 470, + 274 + ], + "type": "text", + "content": "5.1. Ultra-high resolution results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 281, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 472 + ], + "type": "text", + "content": "Our multiscale stylization algorithm is able to transfer style details at UHR for both the content image resolution and the style image resolution. This results in unprecedentedly rich style transfer, as illustrated by Figure 3. As can be observed, minimizing the SOS loss indeed allows to decrease the style transfer loss for all scales (Figure 3 right). The approach is especially relevant when transferring the style of an UHR painting presenting style features at several scales, ranging from a specific color palette to a main curve style and to fine brushstrokes or canvas texture (see close-up views of Figure 3 and the first two lines of Figure 1). To obtain such results, combining style transfer at the largest possible number of scales is critical (see ablation in supp. mat. Figure 9). In addition, the approach is also efficient to transfer the style of a medium resolution style image to an UHR scene, as shown in the last row of Figure 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": "To the best of our knowledge, our approach is the first to work at UHR resolution for neural style transfer of neurally rendered 3D scenes. One of the main advantages of the Gram-based loss is that it does not depend on the style resolution and, when both the content and style images grow in " + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": ", its complexity scales in " + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "inline_equation", + "content": "O(N)" + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": " while NNFM used in ARF [48] scales in " + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Yet, applying style transfer at such resolutions remains computationally heavy: for the garden image of size " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "5187 \\times 3361" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " (Figures 1 and 3) the style transfer takes 25.5 hours (VS 3 hours to train the initial 3DGS model), that is an 8.5 overhead factor. For an image of moderate size (Figure 4 left and middle) the style transfer and the initial 3DGS training take respectively 22 and 10 minutes, that is, the overhead factor is only 2.2. These time values were obtained using a single A100 GPU with 80 GB of memory and could be accelerated by adapting the SOS loss implementation to a multi-GPU setting. Note that high computation times are inherent to UHR style transfer: Running the 2D SPST method for the 185 garden training images takes 27" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26539" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 265 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 265 + ], + "type": "image", + "image_path": "45042e9c71b475d242850656464a4cd131be0725f8932bc821865ce62cc5c9b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": "Figure 4. Comparison of SGSST (ours, top) with StyleGaussian [26] (middle) and ARF [48] (bottom). From left to right the content resolutions are " + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "inline_equation", + "content": "980 \\times 545" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": " (train), " + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "inline_equation", + "content": "979 \\times 546" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": " (truck), and " + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "inline_equation", + "content": "3115 \\times 2076" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": " (counter). For the first two examples, the various outputs keep the resolution of the content, but for the HR counter scene, the output sizes are " + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "inline_equation", + "content": "3115 \\times 2076" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": " for SGSST, " + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "inline_equation", + "content": "1600 \\times 1066" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": " for StyleGaussian and " + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "inline_equation", + "content": "779 \\times 519" + }, + { + "bbox": [ + 55, + 274, + 555, + 330 + ], + "type": "text", + "content": " for ARF (see supp. mat. Figure 28 for ARF results without downscaling). Thanks to its multiscale global VGG statistics, SGSST is the most faithful method regarding style consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 350, + 168, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 168, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 168, + 363 + ], + "type": "text", + "content": "hours (8.9 min. per image)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 136, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 136, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 136, + 384 + ], + "type": "text", + "content": "5.2. Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 462 + ], + "type": "text", + "content": "We perform a thorough comparative study using 40 3D style transfer experiments using 9 different scenes and various style images (see supp. mat.). We compare our results with the NeRF-based ARF [48] and the 3DGS-based StyleGaussian [26] algorithms, described in Section 2, using their public implementations3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 619 + ], + "type": "text", + "content": "Comparing style transfer methods is challenging because each algorithm treats the style image differently. Following high quality 2D style transfer [10, 12], our loss uses up to four scales and each scale uses five VGG layers. The UHR style images were downscaled so that the style has the same size as the content images (no upscale was applied if the style image is smaller than the content image). In contrast, ARF uses a single VGG layer from a single scale and the style image is downscaled to be twice smaller than the content input, resulting in smaller local texture. StyleGaussian reduces the style images so that it has the size " + }, + { + "bbox": [ + 55, + 463, + 295, + 619 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 463, + 295, + 619 + ], + "type": "text", + "content": ", a scale that is hardly sufficient to represent HR style images such as paintings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "text", + "content": "Qualitative comparison Figure 4 shows three different comparative experiments. As one can observe, the results of StyleGaussian [26] are generally not satisfying since the method fails to transfer local texture or to preserve the style" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 483 + ], + "type": "text", + "content": "image's color palette. ARF results better reproduce brushstroke textures, but the transfer is limited, as the method only involves a single VGG layer at a single scale. Also, NNFM does not ensure the preservation of a global color palette. In comparison, SGSST preserves the style at all considered scales. This results in color palette preservation, as well as a verifiable transfer of features of any size, from large brushstrokes (Figure 4 middle) to local grain transfer (Figure 4 right). In addition, SGSST is the only method that performs style transfer at the original resolution of the HR example of Figure 4 right." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "Quantitative comparison Even though there is no consensus for the quantitative evaluation of NST [17], following previous works, we report two different metrics for style transfer quality and texture consistency across views. Style transfer quality can be measured by the Gram loss [11]. A second metric proper to NeRF style transfer [25, 26, 33] is to check the short-term and long-term consistency of the radiance fields in terms of LPIPS [49] and RMSE between wrapped views. Since SGSST is the only method working with UHR images, when necessary we forcefully downgraded all the results to the resolution of ARF for comparison. The average of these two metrics over our 40 experiments is reported in Table 1. The Gram loss is the best for SGSST and, surprisingly, StyleGaussian achieves a lower Gram loss than ARF, which is not consistent with the qualitative evaluation, probably explained by ARF using a single VGG layer compared to five for the two other methods. In terms of consistency metrics, StyleGaussian reports to be" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "3https://github.com/Kai-46/ARF-svox2; https://github.com/Kunhao-Liu/StyleGaussian" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26540" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 152 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 152 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 152 + ], + "type": "table", + "html": "
MethodTransfer qualityShort-range consistencyLong-range consistency
Gram↓LPIPS↓RMSE↓LPIPS↓RMSE↓
SGSST2.59e80.0300.0320.0550.063
StyleGaussian4.61e80.0330.0290.0500.056
ARF5.77e80.0400.0370.0720.066
", + "image_path": "d20ea9173a183d6fc974d97b87f78cbb93ccd8810f13469632468312b4f9df56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 215, + 294, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 296, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 204 + ], + "type": "text", + "content": "Table 1. Quantitative comparison. Style transfer quality and texture consistency metrics averaged over 40 experiments for SGSST, StyleGaussian, and ARF. Best results in bold, second best underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 215, + 294, + 251 + ], + "lines": [ + { + "bbox": [ + 57, + 215, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 215, + 294, + 251 + ], + "type": "table", + "html": "
SGSSTARFStyleGaussian
Voting results (%)66.331.62.1
", + "image_path": "7ddcb1c10da5b49da43f47dd85d16c5a339e78e50f81cad2e316316005c2cfae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 296, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 282 + ], + "type": "text", + "content": "Table 2. Perceptual study. Summary of the 680 votes for the most style consistent result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 302, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 327 + ], + "type": "text", + "content": "the most stable approach followed by SGSST but note that this metric favors the lack of texture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 340, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 295, + 460 + ], + "type": "text", + "content": "Perceptual study To further support our results, we conducted a perceptual study comparing the 40 3D style transfer experiments. For each example, four images were displayed (at a resolution of " + }, + { + "bbox": [ + 54, + 340, + 295, + 460 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 54, + 340, + 295, + 460 + ], + "type": "text", + "content": "): the style image and, in a blind random order, the results of the three methods (SGSST, ARF and StyleGaussian) shown at a common viewpoint (also randomly selected). Each participant was shown ten random instances and was asked to select the result that was the most faithful to the style image. The study was conducted on the web with volunteer participants." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": "A total of 68 participants took part in the test, resulting in 680 votes summarized in Table 2. This perceptual study shows that the fast style transfer performed by Style-Gaussian is consistently judged inferior in quality over ARF and SGSST. It also confirms that SGSST is far superior to ARF in terms of visual quality since " + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": " of the participants ranked our method first for its painting style transfer quality, even when presented with results downscaled in resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 148, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 148, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 148, + 576 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "content": "Influence of optimization parameters Our SGSST algorithm stylizes a realistic 3DGS scene by tuning the constant color components of the 3DGS Gaussians using a single 2D style transfer loss function at multiple scales (Equation (6)). In contrast, Zhang et al. [46] optimize all 3DGS parameters. This, however, necessitates a complex loss made of six different terms to avoid artefacts: a loss term enforces consistency with the original geometry via depth preservation while a preprocessing of Gaussians floaters is necessary after color transfer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Figure 5 shows that optimizing more parameters of the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 326 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 326 + ], + "type": "image", + "image_path": "f107ed0c047f8a35e16ef9fde4dab2e0a14903d4717e44bf630c4099c18ff780.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 334, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 401 + ], + "type": "text", + "content": "Figure 5. Influence of optimization parameters: Allowing more 3DGS parameters to be optimized when minimizing the SOS loss does not improve the stylization quality and can dramatically impact the geometry. From left to right: Style image, content, SGSST default (optimization of colors), results when optimizing all spherical harmonics, results when optimizing all parameters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "type": "text", + "content": "3DGS for SGSST brings no benefit. Optimizing all the spherical harmonic coefficients does not improve the result, and letting all the parameters free like in [22, 46] leads to a strong degradation of the geometry. Note that two instantaneous 3DGS style transfer methods [26, 33] are based on modifying the color features via some neural networks, but their results are not comparable to optimization-based approaches in terms of visual quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "content": "Failure of coarse-to-fine strategy Our approach minimizes a style transfer loss simultaneously at all scales. This is different from the coarse-to-fine style transfer strategy that has proven successful for HR style transfer [12]. Figure 6 illustrates that this coarse-to-fine strategy fails in the context of 3DGS. Indeed, the gain obtained by optimizing at a given scale is quickly lost when optimizing the next one leading to the disappearance of large scale features. This can be explained by the fact that the 3DGS representation is a more constrained representation than pixel grids due to its sparsity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Color transfer via style transfer at the coarsest scale As described in Section 4.3, we first optimize the style transfer loss at the lowest resolution for " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": " iterations and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26541" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 553, + 178 + ], + "type": "image", + "image_path": "062e384cd99470f7558e5c2c0ef740b56f6bd1c788610ba0d90d2e86f6e9e061.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 187, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 555, + 232 + ], + "type": "text", + "content": "Figure 6. Failure of coarse-to-fine strategy. Results of coarse-to-fine style transfer for the example of Figure 3. Each scale is initialized with the output of the previous scale. As shown by the evolution of the losses (right), when training at a new scale, the loss of the previous scales increases quickly. This explains why the large scale painting features disappear progressively and are absent after training the target UHR (see close-up details)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 251, + 135, + 310 + ], + "blocks": [ + { + "bbox": [ + 58, + 251, + 135, + 310 + ], + "lines": [ + { + "bbox": [ + 58, + 251, + 135, + 310 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 135, + 310 + ], + "type": "image", + "image_path": "09bc5f84e0d52196052696820f1705d77c8d398ab2f4fff84605218fbd5032cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 311, + 135, + 369 + ], + "blocks": [ + { + "bbox": [ + 59, + 311, + 135, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 311, + 135, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 311, + 135, + 369 + ], + "type": "image", + "image_path": "29067caafabea0c53c0defeaad7eccb2353db0f6e9ed76927f4d43e387edbcc5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 379, + 295, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 434 + ], + "type": "text", + "content": "Figure 7. Color transfer via style transfer at the coarsest scale. The first 10k iterations at the coarsest scale allow for quick color changes in the 3DGS scene. Removing this step may lead to color artifacts in the result (top) compared to our default two-step optimization (bottom)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 136, + 251, + 214, + 310 + ], + "blocks": [ + { + "bbox": [ + 136, + 251, + 214, + 310 + ], + "lines": [ + { + "bbox": [ + 136, + 251, + 214, + 310 + ], + "spans": [ + { + "bbox": [ + 136, + 251, + 214, + 310 + ], + "type": "image", + "image_path": "1c0ca8096b488dd408aadf4bf8d7f2defb40ed52ccc8ae0c17d2da4bfeeaa775.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 311, + 215, + 369 + ], + "blocks": [ + { + "bbox": [ + 136, + 311, + 215, + 369 + ], + "lines": [ + { + "bbox": [ + 136, + 311, + 215, + 369 + ], + "spans": [ + { + "bbox": [ + 136, + 311, + 215, + 369 + ], + "type": "image", + "image_path": "d5c13c406e9db9d6c6c87f37ab1c0315ca5d7cf4a018be6dd2c0be85364eccee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 215, + 251, + 293, + 310 + ], + "blocks": [ + { + "bbox": [ + 215, + 251, + 293, + 310 + ], + "lines": [ + { + "bbox": [ + 215, + 251, + 293, + 310 + ], + "spans": [ + { + "bbox": [ + 215, + 251, + 293, + 310 + ], + "type": "image", + "image_path": "f5b3ec75dc83ffdc37180a102b13a153c1c804b5350d76fe19efd335a245b060.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 215, + 311, + 294, + 369 + ], + "blocks": [ + { + "bbox": [ + 215, + 311, + 294, + 369 + ], + "lines": [ + { + "bbox": [ + 215, + 311, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 215, + 311, + 294, + 369 + ], + "type": "image", + "image_path": "2670d0503eb4b5450b93caa17d0bdc62f0503ba8dab3d775b77e2f4b8269779d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 374, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 407 + ], + "type": "text", + "content": "Figure 8. Limited texture representation due to low Gaussian density of the initial 3DGS. From left to right: Original scene, 3DGS reconstruction, stylization of the 3DGS scene." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 492 + ], + "type": "text", + "content": "then optimize the SOS loss for another 10k iterations. Figure 7 illustrates that these first iterations are necessary for a faithful style color palette reproduction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 505, + 205, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 205, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 205, + 516 + ], + "type": "text", + "content": "6. Discussion and limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 525, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 573 + ], + "type": "text", + "content": "Texture representation The texture representation within a 3DGS scene depends on the density of Gaussians and may be limited in low density areas. Isolated Gaussians can sometimes be spotted as illustrated by Figure 8." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 589, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 638 + ], + "type": "text", + "content": "Large computation time Depending on the resolution, SGSST requires from several minutes to several hours of computation. On the other hand, fast 3DGS stylization approaches [26, 33] do not reach a satisfactory visual quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Content-style mismatch As said earlier, the Gram loss has the advantage of being independent of the style's image resolution. Also, it enables a faithful transfer of global statistics of the style image, such as its color palette. This important feature is, nevertheless, counterproductive when" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 251, + 553, + 366 + ], + "blocks": [ + { + "bbox": [ + 316, + 251, + 553, + 366 + ], + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 366 + ], + "type": "image", + "image_path": "3643e4be91fb4f374d2ed4bfa5dc376d53a768566e6d90a7b81562192ea5d68b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 429, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 477 + ], + "type": "text", + "content": "the style and content images strongly mismatch, leading to color bleeding or texturing of flat areas. Other controls can be added to mitigate these artefacts [12] and it was shown that these controls are effective for 3DGS scenes [46]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 507, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 626 + ], + "type": "text", + "content": "In this work we presented SGSST, a method that, for the first time, enables UHR 3DGS style transfer. To that aim, among other innovations, we introduced the simultaneously optimized scales (SOS) loss. Our qualitative, quantitative and perceptual studies show that SGSST obtains superior style transfer quality than state of the art, even after reducing our results' resolution for a fair comparison with methods that do not reach UHR. Such high quality UHR results necessitate a large computation time that, nevertheless, remains comparable with that of UHR 3DGS training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 699 + ], + "type": "text", + "content": "This work opens the way to several research directions. A first challenge is to produce equally high quality style transfer with a faster algorithm based on UST. A second more exploratory direction is to investigate geometry style transfer for 3DGS by designing adapted regularization to avoid the caveats depicted by Figure 5." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26542" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "Acknowledgements: B. Galerne and L. Raad acknowledge the support of the project MISTIC (ANR-19-CE40-005)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 115, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 115, + 123 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 115, + 123 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 130, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 130, + 295, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 130, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 295, + 175 + ], + "type": "text", + "content": "[1] Jie An, Tao Li, Haozhi Huang, Li Shen, Xuan Wang, Yongyi Tang, Jinwen Ma, Wei Liu, and Jiebo Luo. Real-time universal style transfer on high-resolution images via zero-channel pruning. arXiv preprint arXiv:2006.09029, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 176, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 232 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470–5479, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 234, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 234, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 234, + 296, + 277 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022, pages 333-350, Cham, 2022. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 280, + 295, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 295, + 311 + ], + "type": "text", + "content": "[4] Tian Qi Chen and Mark Schmidt. Fast patch-based style transfer of arbitrary style. arXiv preprint arXiv:1612.04337, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 315, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 294, + 358 + ], + "type": "text", + "content": "[5] Zhe Chen, Wenhai Wang, Enze Xie, Tong Lu, and Ping Luo. Towards ultra-resolution neural style transfer via thumbnail instance normalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 393-400, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 361, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 295, + 415 + ], + "type": "text", + "content": "[6] Pei-Ze Chiang, Meng-Shiun Tsai, Hung-Yu Tseng, Wei-Sheng Lai, and Wei-Chen Chiu. Stylizing 3d scene via implicit representation and hypernetwork. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1475-1484, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 460 + ], + "type": "text", + "content": "[7] Tai-Yin Chiu and Danna Gurari. Iterative feature transformation for fast and versatile universal style transfer. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 517 + ], + "type": "text", + "content": "[8] Valentin De Bortoli, Agnès Desolneux, Alain Durmus, Bruno Galerne, and Arthur Leclaire. Maximum entropy methods for texture synthesis: Theory and practice. SIAM Journal on Mathematics of Data Science, 3(1):52-82, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 520, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 295, + 575 + ], + "type": "text", + "content": "[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 577, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 577, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 577, + 295, + 609 + ], + "type": "text", + "content": "[10] Bruno Galerne, Lara Raad, José Lezama, and Jean-Michel Morel. Scaling Painting Style Transfer. Computer Graphics Forum, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "text", + "content": "[11] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[12] Leon A. Gatys, Alexander S. Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Nicolas Gonthier, Yann Gousseau, and Saïd Ladjal. High-resolution neural texture synthesis with long-range constraints. Journal of Mathematical Imaging and Vision, 64 (5):478-492, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "text", + "content": "[14] Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour. A sliced wasserstein loss for neural texture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9412-9420, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "type": "text", + "content": "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 222, + 555, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 555, + 277 + ], + "type": "text", + "content": "[16] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: Consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18342–18352, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "text", + "content": "[17] Eleftherios Ioannou and Steve Maddock. Evaluation in neural style transfer: A review. Computer Graphics Forum, 43 (6):e15165, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "type": "text", + "content": "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[19] Hyunyoung Jung, Seonghyeon Nam, Nikolaos Sarafianos, Sungwoo Yoo, Alexander Sorkine-Hornung, and Rakesh Ranjan. Geometry transfer for stylizing radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8565-8575, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 418, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 460 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "type": "text", + "content": "[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "type": "text", + "content": "[22] Aron Samuel Kovács, Pedro Hermosilla, and Renata G. Rraidou. " + }, + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 316, + 510, + 553, + 542 + ], + "type": "text", + "content": "-style: Stylized gaussian splatting. Computer Graphics Forum, n/a(n/a):e15259, 2024. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "type": "text", + "content": "[23] Xueting Li, Sifei Liu, Jan Kautz, and Ming-Hsuan Yang. Learning linear transformations for fast image and video style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3809-3817, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[24] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30:386-396, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[25] Kunhao Liu, Fangneng Zhan, Yiwen Chen, Jiahui Zhang, Yingchen Yu, Abdulmotaleb El Saddik, Shijian Lu, and Eric P. Xing. Stylerf: Zero-shot 3d style transfer of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8338-8348, 2023. 2, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26543" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[26] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. StyleGaussian: Instant 3d style transfer with Gaussian splatting, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[27] Yang Lu, Song-chun Zhu, and Ying Nian Wu. Learning frame models using cnn filters. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision - ECCV 2020, pages 405-421, Cham, 2020. Springer International Publishing. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 230 + ], + "type": "text", + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[30] Dae Young Park and Kwang Hee Lee. Arbitrary style transfer with style-attentional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5880–5888, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 352 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "text", + "content": "[32] Eric Risser, Pierre Wilmot, and Connelly Barnes. Stable and controllable neural texture synthesis and style transfer using histogram losses, 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[33] Abhishek Saroha, Marija Gladkova, Cecilia Curreli, Tarun Yenamandra, and Daniel Cremers. Gaussian splatting in style, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[34] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 488 + ], + "type": "text", + "content": "[35] Omry Sendik and Daniel Cohen-Or. Deep correlations for texture synthesis. ACM Trans. Graph., 36(5), 2017. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[36] Lu Sheng, Ziyi Lin, Jing Shao, and Xiaogang Wang. Avatarnet: Multi-scale zero-shot style transfer by feature decoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8242-8250, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "text", + "content": "[37] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "text", + "content": "[38] Xavier Snelgrove. High-resolution multi-scale neural texture synthesis. In SIGGRAPH Asia 2017 Technical Briefs, New York, NY, USA, 2017. Association for Computing Machinery. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[39] Ondrej Texler, David Futschik, Jakub Fišer, Michal Lukáč, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Arbitrary style transfer using neurally-guided patch-based synthesis. Computers & Graphics, 87:62-71, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[40] D. Ulyanov, V. Lebedev, A. Vedaldi, and V. Lempitsky. Texture networks: Feed-forward synthesis of textures and stylized images. In ICML, pages 1349-1357, 2016. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 508 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[41] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[42] Jonathan Vacher, Aida Davila, Adam Kohn, and Ruben Coen-Cagli. Texture interpolation for probing visual perception. In Advances in Neural Information Processing Systems, pages 22146-22157. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[43] Huan Wang, Yijun Li, Yuehai Wang, Haoji Hu, and Ming-Hsuan Yang. Collaborative distillation for ultra-resolution universal style transfer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "text", + "content": "[44] Zhizhong Wang, Lei Zhao, Zhiwen Zuo, Ailin Li, Haibo Chen, Wei Xing, and Dongming Lu. Microast: Towards super-fast ultra-resolution arbitrary style transfer. Proceedings of the AAAI Conference on Artificial Intelligence, 37(3): 2742-2750, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "text", + "content": "[45] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 319, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 362 + ], + "type": "text", + "content": "[46] Dingxi Zhang, Zhuoxun Chen, Yu-Jie Yuan, Fang-Lue Zhang, Zhenliang He, Shiguang Shan, and Lin Gao. StylizedGS: Controllable stylization for 3d Gaussian splatting, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "text", + "content": "[47] Deheng Zhang, Clara Fernandez-Labrador, and Christopher Schroers. CoARF: Controllable 3D Artistic Style Transfer for Radiance Fields. In 2024 International Conference on 3D Vision (3DV), pages 612-622, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "text", + "content": "[48] Kai Zhang, Nick Kolkin, Sai Bi, Fujun Luan, Zexiang Xu, Eli Shechtman, and Noah Snavely. ARF: Artistic radiance fields. In ECCV 2022, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26544" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_content_list.json b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98dcdc5f12ebaaf1c90efd211ebca31a5819a23b --- /dev/null +++ b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_content_list.json @@ -0,0 +1,2254 @@ +[ + { + "type": "text", + "text": "SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model", + "text_level": 1, + "bbox": [ + 107, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenglin Huang $^{1}$ Jinwei Hu $^{1}$ Xiangtai Li $^{2\\dagger}$ Yiwei He $^{1}$ \nXingyu Zhao $^{3}$ Bei Peng $^{1}$ Baoyuan Wu $^{4}$ Xiaowei Huang $^{1}$ Guangliang Cheng $^{1\\dagger}$", + "bbox": [ + 143, + 202, + 844, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Liverpool, UK 2Nanyang Technological University, SG", + "bbox": [ + 256, + 243, + 738, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 WMG, University of Warwick 4The Chinese University of Hong Kong, Shenzhen, Guangdong, China", + "bbox": [ + 138, + 262, + 836, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: https://hzlsaber.github.io/projects/SIDA/", + "bbox": [ + 250, + 280, + 745, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Corresponding author. E-mail: guangliang.cheng@liverpool.ac.uk xiangtai94@gmail.com", + "bbox": [ + 184, + 297, + 805, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 349, + 326, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of generative models in creating highly realistic images poses substantial risks for misinformation dissemination. For instance, a synthetic image, when shared on social media, can mislead extensive audiences and erode trust in digital content, resulting in severe repercussions. Despite some progress, academia has not yet created a large and diversified deepfake detection dataset for social media, nor has it devised an effective solution to address this issue. In this paper, we introduce the Social media Image Detection dataSet (SID-Set), which offers three key advantages: (1) extensive volume, featuring 300K AI-generated/tampered and authentic images with comprehensive annotations, (2) broad diversity, encompassing fully synthetic and tampered images across various classes, and (3) elevated realism, with images that are predominantly indistinguishable from genuine ones through mere visual inspection. Furthermore, leveraging the exceptional capabilities of large multimodal models, we propose a new image deepfake detection, localization, and explanation framework, named SIDA (Social media Image Detection, localization, and explanation Assistant). SIDA not only discerns the authenticity of images, but also delineates tampered regions through mask prediction and provides textual explanations of the model's judgment criteria. Compared with state-of-the-art deepfake detection models on SID-Set and other benchmarks, extensive experiments demonstrate that SIDA achieves superior performance among diversified settings. The code, model, and dataset will be released.", + "bbox": [ + 88, + 380, + 483, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 829, + 220, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in generative AI [9, 57, 75] have significantly improved the ability to generate highly realistic images, making it easier to create content that closely resem", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd1bf6fb12dc6a8e113de6aab54d4e307954fc8979f8f8bee3e359a79aed1bbb.jpg", + "image_caption": [ + "Figure 1. The framework comparisons. Existing deepfake methods (a-b) are limited to detection, localization, or both. In contrast, SIDA (c) offers a more comprehensive solution, capable of handling detection, localization, and explanation tasks." + ], + "image_footnote": [], + "bbox": [ + 532, + 347, + 903, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bles real-world events. However, these advancements also bring new risks of malicious misuse, particularly in creating deceptive content aimed at misleading public opinion or distorting historical records. Such concerns have motivated the computer vision community to develop more sophisticated deepfake detection techniques. Contemporary methods [32, 70] primarily focus on assessing the authenticity of facial images (i.e., real or fake), while an emerging subset aims to detect and localize facial manipulations [21, 22]. These methods are typically trained on datasets containing real and fake images, aiming to detect images as real or fake, or to localize the tampered regions. Consequently, the quality and diversity of the datasets used for training and evaluation play a crucial role in achieving high accuracy in deepfake detection and localization. A well-curated dataset can enable models to learn nuanced features, improving robustness and generalization in real-world scenarios.", + "bbox": [ + 509, + 566, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing deepfake detection and localization datasets face two main challenges: 1) Insufficient Diversity. The majority of existing datasets for deepfake detection focus mainly on facial imagery [7, 24]. However, given the growing capabilities of generative AI, the issue of non", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28831", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "facial image falsification on social media cannot be overlooked. While researchers [66, 86] have developed relatively large datasets based on ImageNet for image deepfake detection, these datasets typically consist of images from simple scenarios that do not specifically focus on social media. Additionally, they often utilize somewhat outdated image-generation techniques, which can result in less convincing forgeries that are easier for both humans and models to detect. Currently, there is a substantial lack of large-scale image deepfake datasets specifically designed for social media that leverage the latest generative methods. 2) Limited Comprehensiveness. Existing datasets are typically suited either for deepfake detection or for tampered region localization [21-23, 73], focusing on specific types of generative methods or image manipulations. However, an ideal deep-fake dataset should encompass a wide range of scenarios to reflect the complexity of real social media content, where fake images may be fully generated or manipulated through image editing strategies [75]. Furthermore, most existing datasets primarily focus on binary real/fake classification or tampered region localization, with limited emphasis on explaining the cues that models use to make these decisions.", + "bbox": [ + 89, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce the Social media Image Detection dataSet (SID-Set), which consists of 300K images (i.e., 100K real, 100K synthetic, and 100K tampered images), providing a comprehensive resource for the deepfake detection community. Additionally, we include textual descriptions explaining the model's judgment basis. As shown in Figures 2 and 4, synthetic and tampered images are indistinguishable to humans. In particular, challenges for the SID-Set include: 1) subtle alterations of just dozens of pixels; 2) natural-looking local manipulations; 3) complex scenes in datasets. To our knowledge, SID-Set is the first dataset of its scale with extensive annotations, making it the largest and most comprehensive dataset for social media deepfake detection to date. Compared to existing datasets in Table 1, SID-Set addresses the challenges of limited diversity and outdated generative techniques by providing a more comprehensive set of high-quality and diverse images. Accordingly, we propose a new VLMs-based deepfake detection framework, named the Social media Image Detection, localization, and explanation Assistant (SIDA), which achieves the state-of-the-art (SOTA) performance on SID-Set and generalizes effectively across other benchmarks. SIDA can serve as a baseline model on SID-Set, offering a new framework for tackling social media image deepfake detection and localization.", + "bbox": [ + 91, + 426, + 482, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows:", + "bbox": [ + 109, + 806, + 450, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We establish SID-Set, a comprehensive benchmark for detecting, localizing, and explaining deepfakes in social media images, featuring multiple image types and extensive annotations. SID-Set holds the potential for advancing the field of deepfake detection and ensuring", + "bbox": [ + 99, + 824, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/c9b5686711e62ba70430b96c272e31cbcae4f6d32a99a874f48a26c67de818a3.jpg", + "table_caption": [ + "Table 1. Comparison with existing image deepfake datasets." + ], + "table_footnote": [], + "table_body": "
DatasetContentData SourceGenerator YearMulticlasslesMasksExplanation
OHImg [39]OverheadGoogle Map2023✘✘✘
FakeSpotter [63]FaceCelebA, FFHQ2020✘✓✘
ForgeryNet [24]FaceCREMA-D2021✘✓✘
DCFace [27]FaceFFHQ2023✘✘✘
DFF [58]FaceIMDB-WIKI2023✘✘✘
RealFaces [48]FacePrompts2023✘✘✘
M3Dsynth [87]BiologyDDPM, CycleGAN2023✘✓✘
CNNSpot [64]ObjectSDXL2020✓✘✘
CIFAKE [3]ObjectCIFAR2023✓✘✘
CASIA 2.0 [12]GeneralCorel2022✘✓✘
Artifact [53]GeneralCOCO, FFHQ, LSUN2023✘✘✘
IMD2020 [45]GeneralPlaces22020✘✓✘
AIGCD [84]GeneralLSUN, COCO, FFHQ2023✘✘✘
GenImage [86]GeneralImageNet2023✘✘✘
SID-SetGeneralCOCO, Flickr30k, MagicBrush2024✓✓✓
", + "bbox": [ + 516, + 114, + 897, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "robust model performance in complex real-world scenarios.", + "bbox": [ + 535, + 250, + 903, + 279 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SIDA, a new image deepfake detection, localization, and explanation framework that not only detects images with high accuracy but also localizes and explains potential manipulations, enhancing the transparency and utility of deepfake detection technologies.", + "- Extensive experiments demonstrate that SIDA effectively identifies and delineates tampered areas within images, supporting the development of more robust and interpretable deepfake detection systems. Notably, SIDA demonstrates superior or equivalent performance on the SID-Set and other benchmarks." + ], + "bbox": [ + 522, + 281, + 903, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 460, + 653, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Deepfake Datasets", + "text_level": 1, + "bbox": [ + 511, + 484, + 741, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the realm of deepfake detection, the primary focus has historically centered on the identification of facial deepfakes. Renowned datasets such as ForgeryNet [24], DeepFakeFace [58], and DFFD [7] have been pivotal in this area. As the field evolves, there is a growing shift among researchers towards exploring non-facial deepfake data. Advanced methodologies involving text-to-image or imaged-to-image generation techniques [76], utilizing GANs or the stable diffusion series, have facilitated the creation of expansive deepfake datasets like GenImage [86], HiFi-IFDL [21], and DiffForensics [66]. These datasets are characterized by their enlarged data volumes, diversified generation methodologies, and enriched annotation details. Furthermore, beyond the conventional real/fake annotations, certain datasets [21, 22, 73] now include more granular annotations. Table 1 delineates a detailed comparison among various deepfake datasets, highlighting key differences in generation scenarios and annotation types supported. It shows that SID-Set is particularly tailored towards social media data, incorporating the latest SOTA generation models, emphasizing high-quality production, and providing much more comprehensive and diverse annotations.", + "bbox": [ + 511, + 506, + 906, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Deepfake Detection and Localization", + "text_level": 1, + "bbox": [ + 511, + 848, + 883, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deepfake detection methods [4, 6, 38, 49, 62] are typically approached as classification tasks within the data-driven", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28832", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/900e06e414fc30c68fb51eafa1819f86407d0770c2ebab240b89e4d32f372746.jpg", + "image_caption": [ + "Figure 2. SID-Set examples. The 1st row is the synthetic images, while the 2nd row shows tampered images. (Zoom in to view)" + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 478, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "paradigm. These strategies primarily leverage diverse architectures [38, 49], including Convolutional Neural Networks (CNNs) and Transformers, to detect distinctive artifacts. Some scholars have attempted to achieve relatively high precision and generality by employing strategies such as employing techniques such as data augmentation [62], adversarial training [6], reconstruction [4], etc. On the other hand, some researchers [25, 60] have explored extracting features from the frequency domain for deepfake identification. Efforts [13, 65] have also been made to fuse features from both spatial and frequency domains to obtain a more comprehensive set of discriminative features for deepfake detection. Although these methods have shown some progress, they still struggle with issues of generalization. Furthermore, some scholars [20, 40, 44, 61, 80, 83] have gone beyond the basic classification between real and fake by gradually constructing datasets annotated with masks of locally tampered areas, thereby addressing both image deepfake detection and localization tasks. However, these datasets are concentrated mainly on facial data, with fewer datasets available for non-facial deepfake detection and localization, and even fewer large and public datasets for realistic social media data. As a result, our work aims to address these critical gaps by providing a new, comprehensive dataset that includes diverse manipulations beyond facial data, particularly focusing on social media images.", + "bbox": [ + 89, + 248, + 485, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Large Multimodal Models", + "text_level": 1, + "bbox": [ + 89, + 651, + 328, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The progress in large language models (LLMs) [14-16, 41] and vision-language models (VLMs) [14-16, 29, 71, 78] has notably improved multimodal comprehension, seamlessly integrating visual and textual data. The LLaMA series [14-16] optimize language understanding with a compact, high-performance design using fewer parameters than prior models. LLaVA series [34, 35] enhance visual question answering by synchronizing visual features with textual data. Additionally, LISA series [29, 71] employ LLM for accurate image segmentation, merging visual perception with linguistic insights to precisely segment the targeted areas. Several grounding large multimodal models [50, 54, 55, 67, 68, 72, 74, 77, 81] have been proposed to localize the contents based on linguistic information.", + "bbox": [ + 89, + 672, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Advancements [5, 10, 19, 69, 82] in integrating multi", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modal data such as visual and linguistic information have also significantly improved the performance of models in detecting and pinpointing deepfakes. For instance, AntifakePrompt [5] approaches deepfake detection by formulating it as a visual question answering problem, which adjusts soft prompts for InstructBLIP [10], enabling it to determine whether a query image is real or fake. ForgeryGPT [31] enhances image forgery detection and localization by integrating advanced forensic knowledge with a mask-aware forgery extractor that targets pixel-level fraud detection. As a concurrent work, FakeShield [69] leverages the capabilities of LLaVA to identify and localize altered regions while providing interpretable insights into the findings. Our method diverges significantly from existing approaches by creating the most extensive deepfake dataset tailored specifically for social media images with comprehensive annotations. Additionally, we also set a new standard for social media image deepfake detection by utilizing the visual interpretation strengths of VLMs to boost detection accuracy, pinpoint forgeries, and provide clearer explanations within a comprehensive framework.", + "bbox": [ + 511, + 90, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Benchmark", + "text_level": 1, + "bbox": [ + 513, + 426, + 635, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 513, + 452, + 635, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the realm of deepfake detection and localization, research has predominantly focused on facial deepfakes due to their substantial societal impact. However, with the advancements in generative technology, the scope of deepfakes has extended beyond facial content to include non-facial manipulations. Historically, non-facial deepfakes were less prevalent, largely limited by technological constraints that produced low-quality, easily detectable forgeries. Although some datasets, such as GenImage [86] and AIGCD [84], have been constructed, they suffer from several limitations: 1) They often utilize relatively outdated generative technologies, resulting in lower quality data easily distinguished by humans. 2) They primarily focus on text-to-image or image-to-image generation, neglecting the need for data involving manipulations of specific regions, objects, or parts. Such tampered manipulations can be especially insidious as they introduce subtle misinformation, making existing SOTA deepfake detection methods less effective. 3) They lack well-defined criteria for content authenticity, limiting their effectiveness in providing interpretative insights and educating the public on distinguishing synthetic content.", + "bbox": [ + 511, + 476, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Addressing these limitations is crucial for improving the transparency and utility of deepfake detection systems. Furthermore, most existing datasets emphasize either fully synthetic or tampered images, as shown in Table 1, but in real applications, we don't know the image deepfake type in advance. Effective deepfake detection and localization methods should be capable of addressing both scenarios, as", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28833", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7d4e80538d0cedc06c6ff327ed89b1c9acbdf47bbcd410eed18cad201fc0b3a.jpg", + "image_caption": [ + "Figure 3. Tampered image generation pipeline: It consists of four stages—extracting objects from captions using GPT-4o, obtaining object masks with Language-SAM, setting up replacement dictionaries for generating tampered images, and generating new images using Latent Diffusion. This figure illustrates an example of object replacement (e.g., “cat” to “dog”) and attribute modification." + ], + "image_footnote": [], + "bbox": [ + 138, + 103, + 313, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8344ea02b8ed04918e727bb17423363240ffa9770bdf7037f949469b4fe2bf08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 103, + 496, + 284 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0516d556f88e5dce042f78bc0252f097d6fd95bc81476383774671076022833a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 103, + 676, + 284 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0671702a1689addff8324475dc29763e71b7ab4ac300c689fed73c51c0cc48d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 103, + 859, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "social media images often involve complex combinations of synthetic and tampered content. Therefore, developing a comprehensive benchmark for detecting and localizing deepfakes in social media images is essential. We propose SID-Set, which encompasses comprehensive, high-quality annotations for detection and localization, along with detailed textual explanations of the judgment criteria.", + "bbox": [ + 89, + 367, + 483, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Benchmark Construction", + "text_level": 1, + "bbox": [ + 89, + 482, + 323, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Details. To develop an effective benchmark for detecting and localizing images on social media, we created SID-Set, a dataset with real, synthetic, and tampered images reflecting diverse real-world scenarios. Our benchmark assesses whether models can differentiate among real, synthetic, and tampered images, as well as accurately identify altered regions in tampered images.", + "bbox": [ + 89, + 503, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real Images: 100K images from OpenImages $\\mathrm{V}7^{1}$ , with a wide range of scenarios reflecting real-world diversity.", + "bbox": [ + 89, + 609, + 482, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthetic Images: 100K images generated through FLUX [43], specifically designed to challenge identification due to their high-quality, highly realistic appearance.", + "bbox": [ + 89, + 640, + 482, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tampered Images: 100K tampered images, with specific objects or regions replaced or altered; the detailed generation process is shown in Figure 4.", + "bbox": [ + 89, + 686, + 482, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Generation. To generate highly realistic synthetic images, we experimented with several open-source SOTA generative models, such as FLUX [43], Kandinsky 3.0 [1], SDXL [52], AbsoluteReality [42], and others. Following a review by five human experts of 1,000 images from each generative model, FLUX emerged as the top performer, producing highly convincing images that were indistinguishable from real ones to human experts. Consequently, we employed FLUX to create 100K synthetic images based on", + "bbox": [ + 89, + 731, + 482, + 867 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de1d4cf5dd0cc21a7e9e57da614f084929da6b2922307435435695494833426b.jpg", + "image_caption": [ + "bird" + ], + "image_footnote": [], + "bbox": [ + 522, + 364, + 578, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48ca8d52b993b4983b20c04513176cc78bea2e92700bd98cc2c3195281b72154.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 364, + 635, + 407 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26044953c5c9b46fa493d87a54404aedd5ddf2de562b40abba2e99558414eda2.jpg", + "image_caption": [ + "eagle" + ], + "image_footnote": [], + "bbox": [ + 637, + 366, + 692, + 407 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4095ae0660da8175d8a17be01a0d65d498f0c9a885816e8eba455e5b41d68b5b.jpg", + "image_caption": [ + "cat" + ], + "image_footnote": [], + "bbox": [ + 524, + 417, + 578, + 463 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a49b555a043a0427611175aa13d4d3221685b254e3b1cf58c6317979564ac29e.jpg", + "image_caption": [ + "dog" + ], + "image_footnote": [], + "bbox": [ + 581, + 417, + 637, + 463 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96e032ec2989046286f6412203fee5791be076ddb2334d39659de7a84bbdca30.jpg", + "image_caption": [ + "bodyboard" + ], + "image_footnote": [], + "bbox": [ + 714, + 364, + 771, + 407 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49f9a0dbdf350f25cd71c1376b5bcb405809ef7defae0f6c092eb534b022d5c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 364, + 830, + 407 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db7f9096a2b1372de8d9d344772e2f39c0c3ddbced67d20ca7d1e02339b80ca3.jpg", + "image_caption": [ + "coffee table", + "Figure 4. Examples of tampered images. (Zoom in to view)" + ], + "image_footnote": [], + "bbox": [ + 712, + 417, + 769, + 462 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09257dd71993a5510489016414d1f4cc3b8aca492b0f3ab46b1ba410289d1c92.jpg", + "image_caption": [ + "bodyboard", + "coffee table" + ], + "image_footnote": [], + "bbox": [ + 774, + 417, + 830, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the Flickr30k [51] and COCO [33]. The image tampering process depicted in Figure 4 follows four distinct stages, utilizing the COCO image as an example.", + "bbox": [ + 511, + 513, + 903, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: We extract objects from an image's caption using GPT-4o [47]. For instance, from the caption \"A large fluffy cat laying on top of a wooden table\", GPT-4o identifies relevant COCO class objects or retains nouns if no match exists. This extraction is documented in an \"Image-Caption-Object\" JSON file.", + "bbox": [ + 511, + 561, + 905, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2: Employing Language-SAM [30], we generate masks for identified objects as training ground truth.", + "bbox": [ + 511, + 654, + 903, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 3: We establish dictionaries for full and partial image tampering using COCO classes for object replacement and attribute modifications, respectively. For example, replacing \"dog\" with animals like \"cat\" or adding attributes such as \"happy\" or \"angry\" to the \"dog\" class. For more details, please refer to the Appendix.", + "bbox": [ + 511, + 686, + 905, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 4: Utilizing Latent Diffusion [56], we modify captions and regenerate images, either replacing or retaining original objects based on availability. An example modification is altering \"cat\" to \"dog\" in the image caption.", + "bbox": [ + 511, + 777, + 905, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In total, we generated 80,000 object-tampered images and 20,000 partially tampered images. To demonstrate the explainability of SIDA, we used GPT-4o to generate textual descriptions for the judgment basis of 3,000 images from", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://storage.googleapis.com/openimages/web/index.html", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28834", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42197889730a533eac850997b77f30edbf2e7d1e393280ceebb589a8e13b85bd.jpg", + "image_caption": [ + "Figure 5. The pipeline of SIDA: Given an image $x_{i}$ and the corresponding text input $x_{t}$ , the last hidden layer for the token provides the detection result. If the detection result indicates a tampered image, SIDA extracts the token to generate masks for the tampered regions. This figure shows an example where the man's face has been manipulated." + ], + "image_footnote": [], + "bbox": [ + 158, + 92, + 836, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the SID-Set, divided equally among three types. Additionally, to ensure the realism of synthetic images, tampered images, and their textual descriptions, we engaged 5 annotation experts for quality control and adjustments. Detailed prompts and generation pipelines used to create descriptions for each type of image are provided in the Appendix.", + "bbox": [ + 88, + 445, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 551, + 183, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first present the model architecture of SIDA in Section 4.1, followed by an introduction to the training process of our method in Section 4.2.", + "bbox": [ + 89, + 578, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Architecture", + "text_level": 1, + "bbox": [ + 89, + 635, + 225, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Large vision-language models have demonstrated remarkable capabilities in understanding the alignment between textual and visual information. For instance, LLaVA [34] leverages language alone to achieve a comprehensive understanding of both visual and linguistic content. Building on LLaVA, LISA [29] extends this capability by providing fine-grained segmentation masks along with corresponding textual descriptions. However, to effectively detect and localize synthetic images, VLMs must not only be capable of multimodal understanding but also possess the ability to identify and segment manipulated regions, providing detailed explanations for both synthetic and tampered images.", + "bbox": [ + 88, + 657, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we propose SIDA to tackle the task of synthetic image detection and tampered region localization. The pipeline of our method is illustrated in Figure 5. Inspired by previous approaches [28, 29, 85], we expand the", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "original vocabulary of VLMs by adding two new tokens, $<\\mathrm{DET}>$ and $<\\mathrm{SEG}>$ , to enable the model to extract detection and segmentation information. Given an image $x_{i}$ and a text prompt $x_{t}$ , such as \"Can you identify if this image is real, fully synthetic, or tampered? Please mask the tampered object/part if it is tampered.\" We feed them into the VLM. The VLM then outputs a text description $\\hat{y}_{\\mathrm{des}}$ , while the last hidden layer $h_{\\mathrm{hid}}$ contains the $<\\mathrm{DET}>$ and $<\\mathrm{SEG}>$ tokens. This process can be formulated as follows:", + "bbox": [ + 511, + 445, + 906, + 582 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {\\mathrm {d e s}} = \\operatorname {V L M} \\left(x _ {i}, x _ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 597, + 903, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we extract the $<\\mathrm{DET}>$ token from the last hidden layer $h_{\\mathrm{hid}}$ to obtain $h_{\\mathrm{det}}$ . The representation $h_{\\mathrm{det}}$ is then passed through a detection head $F_{\\mathrm{det}}$ to determine whether the image is real, fully synthetic, or tampered. We denote the final detection result by $\\hat{\\mathsf{D}}$ :", + "bbox": [ + 511, + 619, + 906, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathrm {D}} = F _ {\\text {d e t}} \\left(h _ {\\text {d e t}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 710, + 903, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$F_{\\mathrm{det}}$ is the detection head that processes the extracted $\\langle \\mathrm{DET}\\rangle$ representation $h_\\mathrm{det}$ to produce the detection output.", + "bbox": [ + 511, + 734, + 903, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the detection result indicates that the image has been tampered with, SIDA will then predict a mask for the tampered regions. The $h_{\\mathrm{seg}}$ feature is extracted from the hidden layer $h_{\\mathrm{hid}}$ , similar to the extraction process for the token. Given that the token encapsulates crucial information that can aid in generating fine-grained segmentation masks, the representation $h_{\\mathrm{det}}$ is transformed using a fully connected layer $F$ to align with the dimensions of $h_{\\mathrm{seg}}$ . To further capture the relationship between the $h_{\\mathrm{det}}$ and $h_{\\mathrm{seg}}$", + "bbox": [ + 511, + 765, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28835", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features, we apply a single-layer Multihead Attention, facilitating effective feature interaction and enhancing mask quality. The process can be formulated as follows:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {h} _ {\\mathrm {d e t}} = F (h _ {\\mathrm {d e t}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 148, + 315, + 164 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {h} _ {\\text {s e g}} = \\operatorname {M S A} \\left(\\tilde {h} _ {\\det }, h _ {\\text {s e g}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 167, + 480, + 186 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {h} _ {\\mathrm {s e g}} = \\tilde {h} _ {\\mathrm {s e g}} + h _ {\\mathrm {s e g}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 189, + 334, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this formulation, we treat the detection features as the query, and the segmentation features as the key and value. A residual connection is employed to combine both the original and transformed information, thereby enhancing the representation for precise segmentation.", + "bbox": [ + 89, + 219, + 483, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we employ a frozen image encoder $F_{\\mathrm{enc}}$ to extract visual features from the input image $x_{i}$ , resulting in visual features $f$ . The segmentation embedding $\\tilde{h}_{\\mathrm{seg}}$ is then combined with the visual features $f$ and fed into a decoder to produce the final mask $\\hat{M}$ . This can be formulated as:", + "bbox": [ + 89, + 295, + 483, + 369 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf = F _ {\\mathrm {e n c}} (x _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 382, + 321, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathrm {M}} = F _ {\\mathrm {d e c}} \\left(\\tilde {h} _ {\\mathrm {s e g}}, f\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 396, + 480, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Training", + "text_level": 1, + "bbox": [ + 89, + 426, + 194, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Objectives. The training loss, $\\mathcal{L}$ , for the SIDA consists of three components: the detection loss $\\mathcal{L}_{det}$ , the text generation loss $\\mathcal{L}_{txt}$ , and the segmentation mask loss $\\mathcal{L}_{mask}$ . Initially, SIDA is trained in an end-to-end manner by employing the detection loss and the segmentation loss. For detection, we use CrossEntropy loss, while for the segmentation task, we use a weighted combination of binary cross-entropy (BCE) and DICE loss, with respective loss weights $\\lambda_{\\mathrm{bce}}$ and $\\lambda_{\\mathrm{dice}}$ . This can be formulated as:", + "bbox": [ + 89, + 449, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}},\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 598, + 351, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e t} = \\mathcal {L} _ {\\mathrm {C E}} (\\hat {\\mathrm {D}}, \\mathrm {D}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 616, + 480, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {B C E} (\\hat {\\mathrm {M}}, \\mathrm {M}) + \\lambda_ {d i c e} \\mathcal {L} _ {D I C E} (\\hat {\\mathrm {M}}, \\mathrm {M}).\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 633, + 439, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After completing the training phase, we proceed to fine-tune the SIDA model by utilizing detailed textural descriptions from 3,000 images as the ground truth, represented by $y_{des}$ . This phase focuses on optimizing the text generation component, $\\mathcal{L}_{txt}$ , to improve its ability in textural interpretability. The final loss function is as follows:", + "bbox": [ + 89, + 662, + 483, + 752 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t x t} = \\mathcal {L} _ {\\mathrm {C E}} (y _ {\\hat {d e s}}, y _ {d e s}),\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 765, + 302, + 779 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {t x t}} \\mathcal {L} _ {\\text {t x t}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 777, + 480, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{det}$ , $\\lambda_{mask}$ , and $\\lambda_{txt}$ are the weighting factors that balance the contributions of the detection, segmentation, and text generation losses, respectively.", + "bbox": [ + 89, + 810, + 483, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Data. We utilize the SID-Set, consisting of 300k images, to train SIDA. To further enhance diversity, we incorporate the MagicBrush dataset [79] after filtering out", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "low-quality images. The combined dataset supports robust training for both the detection and localization of synthetic content. Additionally, we generate descriptions for 3,000 randomly selected images using LLMs.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 167, + 645, + 185 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We choose LISA as the base large vision language model due to its strong capability for reasoning-based localization. We fine-tuned both LISA-7B-v1 and LISA-13B-v1 on the SID-Set using LoRA, setting $\\alpha$ at 16 and the dropout rate at 0.05. The input images are resized to $1024 \\times 1024$ . The loss weights in Eq. (6) for the detection $(\\lambda_{\\mathrm{det}})$ , the text generation $(\\lambda_{\\mathrm{txt}})$ , and the localization $(\\lambda_{\\mathrm{mask}})$ are set to 1.0, respectively. The localization loss weights in Eq. (5) for the $\\lambda_{\\mathrm{bce}}$ and $\\lambda_{\\mathrm{dice}}$ are set to 2.0 and 0.5, respectively. To determine the optimal weight configuration, we perform ablation studies as detailed in Section 5.5. During the detection and localization training stage, the image encoder is frozen, and all other modules are trainable. For the text generation stage, only vision-language models are fine-tuned using the LoRA strategy. The initial learning rate is set to $1 \\times 10^{-4}$ , with a batch size of 2 per device and a gradient accumulation step of 10. We use two NVIDIA A100 GPUs (40GB each). Training for SIDA-7B and SIDA-13B took 48 hours and 72 hours, respectively.", + "bbox": [ + 511, + 194, + 906, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate detection using image-level accuracy and F1 scores. For forgery localization, our metrics include Area Under the Curve (AUC), F1 scores, and Intersection over Union (IoU).", + "bbox": [ + 511, + 482, + 905, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Detection Evaluation", + "text_level": 1, + "bbox": [ + 511, + 555, + 710, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SIDA against other SOTA deepfake detection methods on SID-Set, including CnnSpot [17], AntifakePrompt [5], FreDect [18], Fusing [26], Gram-Net [37], UniFDP [46], LGrad [59], and LNP [2]. To ensure a fair comparison, we first evaluate these models on our dataset using their original pre-trained weights, then retrain them with the SID-Set to assess performance improvements. Table 2 demonstrates that SIDA achieves better or comparable results among all the evaluated methods. Notably, LGrad [59] achieves the highest accuracy and F1 score on tampered images after retraining, but this comes at the expense of lower performance in other metrics. Our analysis indicates that LGrad's high recall and false positive rates stem from its propensity to misclassify other types as tampered. The training details are provided in the Appendix.", + "bbox": [ + 511, + 578, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Localization Results", + "text_level": 1, + "bbox": [ + 511, + 816, + 705, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents the forgery localization performance on the SID-Set. We selected PSCC-Net [36], MVSS-Net [11], and HIFI-Net [21] as representative IFDL methods. Additionally, we chose LISA [29] as a representative LLM due to", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28836", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/10c8d3229089fe4ea36be612686c6f04d0109b7308d4ce10d64cd68570678b68.jpg", + "table_caption": [ + "Table 2. Comparison of SIDA with other deepfake detection methods. Values outside parentheses are evaluated with open-source models directly on the SID-Set; values in parentheses indicate performance changes after training with the SID-Set. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsYearRealFully syntheticTamperedOverall
AccF1AccF1AccF1AccF1
AntifakePrompt [5]202464.8(↑24.1)78.6(↑10.5)93.8(↑3.7)96.8(↑1.1)30.8(↑60.1)47.2(↑33.2)63.1(↑29.1)69.3(↑23.4)
CnnSpott [17]202179.8(↑9.2)88.7(↑2.1)39.5(↑51.2)56.6(↑31.5)6.9(↑61.2)12.9(↑51.1)42.1(↑39.3)69.6(↑20.7)
FreDect [18]202083.7(↓37.7)91.1(↓43.5)16.8(↑44.1)28.8(↑37.2)11.9(↑25.2)21.3(↑31.7)37.4(↑33.6)23.4(↑47.2)
Fusing [26]202285.1(↑4.1)92.0(↑0.7)34.0(↑54.1)50.7(↑38.4)2.7(↑24.3)5.3(↑26.1)40.1(↑33.1)29.1(↑40.3)
Gram-Net [37]202070.1(↑19.1)82.4(↑9.3)93.5(↑4.4)96.6(↑2.0)0.8(↑89.1)1.6(↑85.3)55.0(↑37.1)58.0(↑35.1)
UnivFD [46]202368.0(↑0.3)67.4(↑1.1)62.1(↑24.3)87.5(↑10.5)64.0(↑28.5)85.3(↑4.7)64.0(↑21.7)85.3(↑4.5)
LGrad [59]202364.8(↓2.8)78.6(↓2.5)83.5(↓25.5)91.0(↓23.7)6.8(↑92.3)12.7(↑86.1)51.8(↑20.2)55.5(↑23.9)
LNP [2]202371.2(↓56.8)83.2(↓60.2)91.8(↓55.6)95.7(↓60.1)2.9(↑90.4)5.7(↑88.9)55.2(↓7.6)58.2(↑4.1)
SIDA-7B202489.191.098.798.691.291.093.593.5
SIDA-13B202489.691.198.598.792.991.293.693.5
", + "bbox": [ + 102, + 128, + 893, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "its segmentation reasoning capabilities. We used LISA-7Bv1 and fine-tuned it on SID-Set. The results indicate that SIDA achieves the best performance. We suppose that while LISA possesses strong general segmentation capabilities, it lacks the specific specialized features required to detect subtle manipulations, ultimately limiting the effectiveness of fine-tuning for precise forgery localization.", + "bbox": [ + 89, + 338, + 483, + 445 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2d5a4ba70606e8844e510a108f4d0c12034dc04f6d6f7ec2b6f3edabd48bf9b.jpg", + "table_caption": [ + "Table 3. Comparison between SIDA and other IFDL approaches. * indicates the use of the pre-trained model from the original paper due to unavailable training code." + ], + "table_footnote": [], + "table_body": "
MethodsYearsTampered
AUCF1IOU
MVSS-Net* [11]202348.931.623.7
HIFI-Net* [21]202364.045.921.1
PSCC-Net [36]202282.171.335.7
LISA-7B-v1 [29]202478.469.132.5
SIDA-7B202487.373.943.8
", + "bbox": [ + 112, + 503, + 455, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Robustness Study", + "text_level": 1, + "bbox": [ + 89, + 651, + 264, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate the robustness of SIDA against common image perturbations found in social media, such as JPEG compression, resizing, and Gaussian noise. Table 4 shows our model's performance on the SID-Set under six degradation scenarios: JPEG compression (with quality levels of 70 and 80), resizing (with scaling factors of 0.5 and 0.75), and Gaussian noise (with variances of 5 and 10). Despite not being explicitly trained on degraded data, SIDA demonstrates resilience to these low-level distortions. The model's stable performance against common social media perturbations highlights its robustness and practical applicability.", + "bbox": [ + 89, + 672, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Test on Other Benchmark", + "text_level": 1, + "bbox": [ + 89, + 848, + 328, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this stage, we evaluate SIDA on DMImage [8] dataset to assess its generalization capabilities. We compare SIDA", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb1049f7235e88308195a8ced4c1d1ab278c39bdf17805d429c2f8875f5126ae.jpg", + "table_caption": [ + "Table 4. Performance of SIDA under different perturbations." + ], + "table_footnote": [], + "table_body": "
DetectionLocalization
ACCF1AUCF1IOU
JPEG 7089.490.186.271.842.3
JPEG 8088.789.585.871.141.7
Resize 0.589.391.186.872.543.2
Resize 0.7589.991.687.173.043.5
Gaussian 1086.989.384.170.241.0
Gaussian 588.489.985.371.041.5
SIDA-7B93.593.587.373.943.8
", + "bbox": [ + 531, + 354, + 885, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with CNNSpot [17], Fusing [26], Gram-Net [37], LNP [2], UnivFD [46], and AntifakePrompt [5]. For these methods, we use the original hyperparameter settings and pre-trained weights provided by the authors. The results in Table 5 show that SIDA achieves superior performance, demonstrating its strong adaptability.", + "bbox": [ + 511, + 507, + 906, + 599 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a95dd370faa02191c41db8e2dce6fa3e249fd87b03f4db9379422553e6e04633.jpg", + "table_caption": [ + "Table 5. Comparison with other deepfake detection methods on DMImage [8]. We used original weights for each method." + ], + "table_footnote": [], + "table_body": "
MethodsRealFakeOverall
AccF1AccF1AccF1
CNNSpot [17]87.888.428.444.240.643.3
Gram-Net [37]62.854.178.888.167.479.4
Fusing [26]87.786.115.527.240.436.5
LNP [2]63.167.456.972.558.268.3
UnivFD [46]89.488.344.961.253.960.7
AntifakePrompt [5]91.392.589.391.290.691.2
SIDA-7B92.993.190.791.091.892.4
", + "bbox": [ + 522, + 648, + 897, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 801, + 666, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attention Module. We conducted ablation experiments to assess the importance of the attention module. Variants included removing the attention module entirely and replacing it with fully connected (FC) layers. Results in Table 6 show that removing the attention module or replacing", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28837", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3be4828393098d6cd896c94fcd57e334a9d7dcb1983fbe7350ec3da5a626892d.jpg", + "table_caption": [ + "Table 6. Ablation study results for attention module in SIDA." + ], + "table_footnote": [], + "table_body": "
DetectionLocalization
ACCF1AUCF1IOU
FC91.190.384.371.638.9
w/o Attention90.389.984.171.338.8
SIDA93.593.587.373.943.8
", + "bbox": [ + 101, + 102, + 465, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "it with FC significantly reduces performance, underscoring the critical role of attention in enhancing feature interaction and improving detection and localization accuracy.", + "bbox": [ + 89, + 196, + 482, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Weights. SIDA training utilizes weighted losses to balance task contributions. In the detection and localization stages, detection loss is adjusted by a weight $\\lambda_{\\mathrm{det}}$ , and localization loss by binary cross-entropy (BCE) and DICE losses, with weights $\\lambda_{\\mathrm{bce}}$ and $\\lambda_{\\mathrm{dice}}$ respectively. For our experiments, we set $\\lambda_{\\mathrm{det}}$ to 1, $\\lambda_{\\mathrm{bce}}$ to 2.0, and $\\lambda_{\\mathrm{dice}}$ to 0.5 to maintain balance between detection and localization, enhancing model stability and performance. We summarize the outcomes of various weight configurations in Table 7.", + "bbox": [ + 89, + 242, + 483, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f205179618c5c520474973b0d3cec0dfe0232ca1fa530d5ab1d858fe57eb92ec.jpg", + "table_caption": [ + "Table 7. Different weight configurations in SIDA training." + ], + "table_footnote": [], + "table_body": "
λdetλbceλdiceAccF1 Score
1.02.00.593.5691.01
1.04.01.093.4990.86
", + "bbox": [ + 143, + 404, + 423, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 477, + 272, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present examples of SIDA's output for tampered images, showcasing its detection, localization, and explanation capabilities. SIDA accurately identifies tampered regions and provides explanations for its decisions. We also include some challenging failure cases where SIDA was unable to accurately detect the tampered regions, highlighting areas for future improvement. Additional visualization results are available in the Appendix.", + "bbox": [ + 89, + 500, + 483, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussions", + "text_level": 1, + "bbox": [ + 89, + 633, + 346, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present SID-Set for social image deepfake detection, localization, and explanation tasks, consisting of 100k real images, 100k fully synthetic images, and 100k tampered images. Furthermore, we propose a new VLMs-based deepfake detection framework, SIDA, to address these tasks. SIDA demonstrates its ability to detect fake types, localize tampered regions, and provide explanations for its decisions. We believe that the integration of VLMs into deepfake detection tasks, as demonstrated by SIDA, offers promising new avenues for future research.", + "bbox": [ + 89, + 659, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the development of the SID-Set and the introduction of the SIDA framework have yielded favorable outcomes in deepfake detection tasks, we recognize some potential limitations and will optimize them in future research. Dataset Size. While SID-Set includes 100k fully synthetic and 100k tampered images, the complexity of real social", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can you identify if this image is real, full synthetic, or tampered? Please mask the tampered object/part if it is tampered.", + "bbox": [ + 535, + 107, + 795, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cea45cf8fea55b4926330f894c18d05f6deb08fff83a21fd7e5076479c50d5ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 127, + 619, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c200b6ce03b3931f3d454cab54e2684f47613a32ea7ecb693f80664fd17f538d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 128, + 702, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5378186f9d0b3ca33adfc77b147ccfa0659099504b64c2a085bb99be9287f1c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 127, + 781, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/079219d1ce457548ba22ecf335969b56e524b9638d61fcaf3b92efeb35f77ca0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 126, + 867, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce7e752e97820e7a8226687c15abd7008aceaf0b6bd19c156788361551a45a9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 186, + 568, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object ampered", + "text_level": 1, + "bbox": [ + 568, + 210, + 602, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb062c8e102d27df3f642a0a9898045943daa8236db0862d47fb0aafada622e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 231, + 619, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Partial tampered", + "text_level": 1, + "bbox": [ + 630, + 210, + 692, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e319df19ca84f1e3b75245fd10a4f9810d97beb316ed10ad146bd76cc6d2a7f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 231, + 705, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Partial tampered", + "text_level": 1, + "bbox": [ + 709, + 210, + 785, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62013276a45753571f0ff721eabd0b1b48e45e5abfbc92bb27c42901d3dc32ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 232, + 785, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object tampered", + "text_level": 1, + "bbox": [ + 813, + 210, + 849, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2429f968347ef108e363714fc4ae8c9e58a74641c8ceef03c045d5a6284092b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 232, + 870, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Type of Image: object tampered The tampered areas are specifically located in the h The tampered. Absolute Positino The location is", + "bbox": [ + 544, + 289, + 593, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Type of Image: part tampered Tampering Type and Location... Relative Position: The tampered part located in the face of the women... Absolute Position...", + "bbox": [ + 635, + 289, + 700, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Type of Image: object \ntampered Tampering \nType and Location is the face of the man... \nAbsolute Position \nTampered Content: Modification... \nInconsistencies: Lighting...", + "bbox": [ + 714, + 290, + 781, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Type of Image: part tampered Tampered Content: Modification... Relative Position: Inconsistencies... Colour in the background looks ...", + "bbox": [ + 797, + 290, + 867, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11cdcee45f455b539c81086e3fcd9513808fc335d8ccea4096713c27a0c98e59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 354, + 568, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a) Correct examples", + "text_level": 1, + "bbox": [ + 638, + 357, + 746, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14d6f60b7b0334a52063389610d2bea7f39c4c3b83e067466feb6f3c10cf8e27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 383, + 617, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb7b756972ab1bf586fc87727112f928b3ae74a50723f70cdf684013e6c7afc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 383, + 697, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2fe3771f02eab9bbb4d1c8c29db5f187e33548a73c30bd7e7ce0133cfda18a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 383, + 782, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1a5e589bf67da6b38cec5820377fc7b1981fcf416524b65def7ad4e2ba1e3d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 383, + 867, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/512d5490c30048e1b356504f7698c6ba0cd23d4f5ef762b7f836ba62bd9771aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 439, + 617, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e87ad72ff1ac4415100190a6a8af837624093e789d0f5cf7c3696f8f9a1bede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 439, + 697, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0140ded12571ec68e5754f4aee42949c691cc306aa0071d125340cac24d60006.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 439, + 782, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e600bd8b9b358bc9f427a2fc421aa5ef57d09b29234c191ab40f76c238401334.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 439, + 867, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36ab3eb1f014cf103ac1139689ef532fbdead74790683bd375209dd21c2fbe90.jpg", + "image_caption": [ + "b) Failure examples" + ], + "image_footnote": [], + "bbox": [ + 544, + 498, + 617, + 554 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad4bcb55a045272f48a59cacd607cd3f77c9142b8494f6f864dbff97e04911c2.jpg", + "image_caption": [ + "Figure 6. Visual results of SIDA on tampered images." + ], + "image_footnote": [], + "bbox": [ + 625, + 498, + 697, + 555 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cc53ac93d82242eb6072fe6b0402e0cdb5f7dfcc25c56274a934a5e4e8464da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 500, + 782, + 554 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a4d2a53420c420336ae94a1141e2184c8d71d248cdb2cd33a130a3d20594551.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 500, + 867, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "media environments demands a larger dataset. Therefore, expanding the dataset with additional images is a crucial objective for future research.", + "bbox": [ + 511, + 632, + 903, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Domain. Other methods often generate social media images that lack authenticity or are easily identifiable. Consequently, we used only FLUX to generate fake images due to its superior quality. However, depending exclusively on one method may lead to issues of data skew, although this was not significantly evident in our experiments. Such skew could potentially impair performance on diverse datasets. Moving forward, we plan to explore additional generative methods and integrate various other generation techniques to produce a more varied and higher-quality set of images.", + "bbox": [ + 511, + 678, + 906, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Localization Results. Although SIDA demonstrates relatively strong performance on our dataset, there is still room for improvement. Certain tampered regions are not reliably detected, underscoring the need for further advancements.", + "bbox": [ + 511, + 829, + 903, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28838", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by The Alan Turing Institute (UK) through the project 'Turing-DSO Labs Singapore Collaboration' (SDCfP2\\100009).", + "bbox": [ + 89, + 113, + 483, + 156 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 183, + 187, + 199 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. Arxiv, 2023. 4", + "[2] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. Detecting generated images by real images only. Arxiv, 2023. 6, 7", + "[3] Jordan J. Bird and Ahmad Lotfi. CIFAKE: image classification and explainable identification of ai-generated synthetic images. IEEE Access, 2024. 2", + "[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In CVPR, 2022. 2, 3", + "[5] You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors. Arxiv, 2023. 3, 6, 7", + "[6] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 2, 3", + "[7] Harry Cheng, Yangyang Guo, Tianyi Wang, Liqiang Nie, and Mohan S. Kankanhalli. Diffusion facial forgery detection. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 1, 2", + "[8] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, 2023. 7", + "[9] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 2023. 1", + "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 3", + "[11] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE TPAMI, 2023. 6, 7", + "[12] Jing Dong, Wei Wang, and Tieniu Tan. CASIA image tampering detection evaluation database. In ISIP, 2013. 2", + "[13] Junxian Duan, Yang Ai, Jipeng Liu, Shenyuan Huang, Huaibo Huang, Jie Cao, and Ran He. Test-time forgery detection with spatial-frequency prompt learning. IJCV, 2024. 3", + "[14] Abhimanyu Dubey et al. The llama 3 herd of models. Arxiv, 2024. 3" + ], + "bbox": [ + 93, + 209, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Hugo Touvron et al. Llama: Open and efficient foundation language models. Arxiv, 2023.", + "[16] Hugo Touvron et al. Llama 2: Open foundation and finetuned chat models. Arxiv, 2023. 3", + "[17] Joel Frank and Thorsten Holz. Cnn-generated images are surprisingly easy to spot...for now. Arxiv, 2021. 6, 7", + "[18] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 6, 7", + "[19] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In AAAI, 2024. 3", + "[20] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In CVPR, 2023. 3", + "[21] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 1, 2, 6, 7", + "[22] Xiao Guo, Xiaohong Liu, Iacopo Masi, and Xiaoming Liu. Language-guided hierarchical fine-grained image forgery detection and localization. Arxiv, 2024. 1, 2", + "[23] Ruidong Han, Xiaofeng Wang, Ningning Bai, Yihang Wang, Jianpeng Hou, and Jianru Xue. HDF-net: Capturing homogeneity difference features to localize the tampered image. IEEE TPAMI, 2024. 2", + "[24] Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. Forgerynet: A versatile benchmark for comprehensive forgery analysis. In CVPR, 2021. 1, 2", + "[25] Yonghyun Jeong, Doyeon Kim, Youngmin Ro, and Jongwon Choi. Frepgan: Robust deepfake detection using frequency-level perturbations. In AAAI, 2022. 3", + "[26] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In ICIP, 2022. 6, 7", + "[27] Minchul Kim, Feng Liu, Anil K. Jain, and Xiaoming Liu. Dcface: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 2", + "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021. 5", + "[29] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: reasoning segmentation via large language model. In CVPR, 2024. 3, 5, 6, 7", + "[30] lang-sam team. Lang-segment-anything. https://github.com/luca-medeiros/lang-segment-anything, 2024. Accessed: 2024-11-15. 4", + "[31] Jiawei Li, Fanrui Zhang, Jiaying Zhu, Esther Sun, Qiang Zhang, and Zheng-Jun Zha. Forgerygpt: Multimodal large language model for explainable image forgery detection and localization. Arxiv, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28839", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. Detecting multimedia generated by large AI models: A survey. Arxiv, 2024. 1", + "[33] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 5", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 3", + "[36] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for image manipulation detection and localization. IEEE TCSVT, 2022, 6, 7", + "[37] Zhengzhe Liu, Xiaojuan Qi, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. In CVPR, 2020. 6, 7", + "[38] Momina Masood, Marriam Nawaz, Khalid Mahmood Malik, Ali Javed, Ain Irtaza, and Hafiz Malik. Deepfakes generation and detection: state-of-the-art, open challenges, countermeasures, and way forward. Appl. Intell., 2023. 2, 3", + "[39] Brandon B. May, Kirill Trapeznikov, Shengbang Fang, and Matthew C. Stamm. Comprehensive dataset of synthetic and manipulated overhead imagery for development and evaluation of forensic tools. In IH&MMSec, 2023. 2", + "[40] Changtao Miao, Qi Chu, Zhentao Tan, Zhenchao Jin, Wanyi Zhuang, Yue Wu, Bin Liu, Honggang Hu, and Nenghai Yu. Multi-spectral class center network for face manipulation detection and localization. Arxiv, 2023. 3", + "[41] Shervin Minaee, Tomás Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. Arxiv, 2024. 3", + "[42] Absolutereality model team. Absolutereality model repository. https://huggingface.co/jochemstoel/absolutereality-model-repository, 2024. Accessed: 2024-11-15. 4", + "[43] Flux model team. Flux model. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. Accessed: 2024-11-15. 4", + "[44] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In CVPR, 2024. 3", + "[45] Adam Novozámsky, Babak Mahdian, and Stanislav Saic. IMD2020: A large-scale annotated dataset tailored for detecting manipulated images. In WACvw, 2020. 2", + "[46] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, 2023. 6, 7", + "[47] OpenAI. GPT-4 technical report. Arxiv, 2023. 4", + "[48] Lorenzo Papa, Lorenzo Faiella, Luca Corvitto, Luca Maiano, and Irene Amerini. On the use of stable diffusion for creating realistic faces: from generation to detection. In IWBF, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Gan Pei, Jiangning Zhang, Menghan Hu, Guangtao Zhai, Chengjie Wang, Zhenyu Zhang, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey. Arxiv, 2024. 2, 3", + "[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. Arxiv, 2023. 3", + "[51] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 2017. 4", + "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 4", + "[53] Md Awsafur Rahman, Bishmoy Paul, Najibul Haque Sarker, Zaber Ibn Abdul Hakim, and Shaikh Anowarul Fattah. Artifact: A large-scale dataset with artificial and factual images for generalizable and robust synthetic image detection. In ICIP, 2023. 2", + "[54] Hanoona Abdul Rasheed, Muhammad Maaz, Sahal Shaji Mullappilly, Abdelrahman M. Shaker, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Eric P. Xing, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Glamm: Pixel grounding large multimodal model. In CVPR, 2024. 3", + "[55] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In CVPR, 2024. 3", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4", + "[57] Xincheng Shuai, Henghui Ding, Xingjun Ma, Rongcheng Tu, Yu-Gang Jiang, and Dacheng Tao. A survey of multimodal-guided image editing with text-to-image diffusion models. Arxiv, 2024. 1", + "[58] Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. Arxiv, 2023. 2", + "[59] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In CVPR, 2023. 6, 7", + "[60] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In AAAI, 2024. 3", + "[61] Dragos-Constantin Tântaru, Elisabeta Oneata, and Dan Oneata. Weakly-supervised deepfake localization in diffusion-generated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 3", + "[62] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, 2021. 2, 3", + "[63] Run Wang, Felix Juefei-Xu, Lei Ma, Xiaofei Xie, Yihao Huang, Jian Wang, and Yang Liu. Fakespotter: A simple yet robust baseline for spotting ai-synthesized fake faces. In IJCAI, 2020. 2" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28840", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2", + "[65] Yuan Wang, Kun Yu, Chen Chen, Xiyuan Hu, and Silong Peng. Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In CVPR, 2023. 3", + "[66] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In ICCV, 2023. 2", + "[67] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. Arxiv, 2024. 3", + "[68] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. GSVA: generalized segmentation via multimodal large language models. In CVPR, 2024. 3", + "[69] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. Arxiv, 2024. 3", + "[70] Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A comprehensive benchmark of deepfake detection. In NeurIPS, 2023. 1", + "[71] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. *Arxiv*, 2023. 3", + "[72] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In ICLR, 2024. 3", + "[73] Zeqin Yu, Jiangqun Ni, Yuzhen Lin, Haoyi Deng, and Bin Li. Diffforensics: Leveraging diffusion prior to image forgery detection and localization. In CVPR, 2024. 2", + "[74] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Arxiv, 2023. 3", + "[75] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, Lingjie Liu, Adam Kortylewski, Christian Theobalt, and Eric P. Xing. Multimodal image synthesis and editing: The generative AI era. IEEE TPAMI, 2023. 1, 2", + "[76] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion models in generative AI: A survey. Arxiv, 2023. 2", + "[77] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, and Jainwei Yang. Llava-grounding: Grounded visual chat with large multimodal models. In ECCV, 2024.3", + "[78] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE TPAMI, 2024. 3", + "[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2023. 6", + "[80] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In ICCV, 2023. 3", + "[81] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Change Loy Chen, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. In NeurIPS, 2024. 3", + "[82] Yue Zhang, Ben Colman, Ali Shahriyari, and Gaurav Bharaj. Common sense reasoning for deep fake detection. *Arxiv*, 2024. 3", + "[83] Yi Zhang, Changtao Miao, Man Luo, Jianshu Li, Wenzhong Deng, Weibin Yao, Zhe Li, Bingyu Hu, Weiwei Feng, Tao Gong, and Qi Chu. MFMS: learning modality-fused and modality-specific features for deepfake detection and localization tasks. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 3", + "[84] Nan Zhong, Yiran Xu, Zhenxing Qian, and Xinpeng Zhang. Rich and poor texture contrast: A simple yet effective approach for ai-generated image detection. Arxiv, 2023. 2, 3", + "[85] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and VQA. In AAAI, 2020. 5", + "[86] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In NeurIPS, 2023. 2, 3", + "[87] Giada Zingarini, Davide Cozzolino, Riccardo Corvi, Giovanni Poggi, and Luisa Verdoliva. M3DSYNTH: A dataset of medical 3d images with ai-generated local manipulations. In ICASSP, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 501 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_model.json b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e263047794d72f921f66265d4a5917c44aefdfe1 --- /dev/null +++ b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_model.json @@ -0,0 +1,3225 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.846, + 0.242 + ], + "angle": 0, + "content": "Zhenglin Huang\\(^{1}\\) Jinwei Hu\\(^{1}\\) Xiangtai Li\\(^{2\\dagger}\\) Yiwei He\\(^{1}\\) \nXingyu Zhao\\(^{3}\\) Bei Peng\\(^{1}\\) Baoyuan Wu\\(^{4}\\) Xiaowei Huang\\(^{1}\\) Guangliang Cheng\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.244, + 0.74, + 0.262 + ], + "angle": 0, + "content": "1University of Liverpool, UK 2Nanyang Technological University, SG" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.263, + 0.837, + 0.279 + ], + "angle": 0, + "content": "3 WMG, University of Warwick 4The Chinese University of Hong Kong, Shenzhen, Guangdong, China" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.281, + 0.746, + 0.297 + ], + "angle": 0, + "content": "Project Page: https://hzlsaber.github.io/projects/SIDA/" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.298, + 0.807, + 0.315 + ], + "angle": 0, + "content": "† Corresponding author. E-mail: guangliang.cheng@liverpool.ac.uk xiangtai94@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.35, + 0.327, + 0.365 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.485, + 0.804 + ], + "angle": 0, + "content": "The rapid advancement of generative models in creating highly realistic images poses substantial risks for misinformation dissemination. For instance, a synthetic image, when shared on social media, can mislead extensive audiences and erode trust in digital content, resulting in severe repercussions. Despite some progress, academia has not yet created a large and diversified deepfake detection dataset for social media, nor has it devised an effective solution to address this issue. In this paper, we introduce the Social media Image Detection dataSet (SID-Set), which offers three key advantages: (1) extensive volume, featuring 300K AI-generated/tampered and authentic images with comprehensive annotations, (2) broad diversity, encompassing fully synthetic and tampered images across various classes, and (3) elevated realism, with images that are predominantly indistinguishable from genuine ones through mere visual inspection. Furthermore, leveraging the exceptional capabilities of large multimodal models, we propose a new image deepfake detection, localization, and explanation framework, named SIDA (Social media Image Detection, localization, and explanation Assistant). SIDA not only discerns the authenticity of images, but also delineates tampered regions through mask prediction and provides textual explanations of the model's judgment criteria. Compared with state-of-the-art deepfake detection models on SID-Set and other benchmarks, extensive experiments demonstrate that SIDA achieves superior performance among diversified settings. The code, model, and dataset will be released." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.222, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recent advances in generative AI [9, 57, 75] have significantly improved the ability to generate highly realistic images, making it easier to create content that closely resem" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.348, + 0.905, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Figure 1. The framework comparisons. Existing deepfake methods (a-b) are limited to detection, localization, or both. In contrast, SIDA (c) offers a more comprehensive solution, capable of handling detection, localization, and explanation tasks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.907, + 0.824 + ], + "angle": 0, + "content": "bles real-world events. However, these advancements also bring new risks of malicious misuse, particularly in creating deceptive content aimed at misleading public opinion or distorting historical records. Such concerns have motivated the computer vision community to develop more sophisticated deepfake detection techniques. Contemporary methods [32, 70] primarily focus on assessing the authenticity of facial images (i.e., real or fake), while an emerging subset aims to detect and localize facial manipulations [21, 22]. These methods are typically trained on datasets containing real and fake images, aiming to detect images as real or fake, or to localize the tampered regions. Consequently, the quality and diversity of the datasets used for training and evaluation play a crucial role in achieving high accuracy in deepfake detection and localization. A well-curated dataset can enable models to learn nuanced features, improving robustness and generalization in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, existing deepfake detection and localization datasets face two main challenges: 1) Insufficient Diversity. The majority of existing datasets for deepfake detection focus mainly on facial imagery [7, 24]. However, given the growing capabilities of generative AI, the issue of non" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.424 + ], + "angle": 0, + "content": "facial image falsification on social media cannot be overlooked. While researchers [66, 86] have developed relatively large datasets based on ImageNet for image deepfake detection, these datasets typically consist of images from simple scenarios that do not specifically focus on social media. Additionally, they often utilize somewhat outdated image-generation techniques, which can result in less convincing forgeries that are easier for both humans and models to detect. Currently, there is a substantial lack of large-scale image deepfake datasets specifically designed for social media that leverage the latest generative methods. 2) Limited Comprehensiveness. Existing datasets are typically suited either for deepfake detection or for tampered region localization [21-23, 73], focusing on specific types of generative methods or image manipulations. However, an ideal deep-fake dataset should encompass a wide range of scenarios to reflect the complexity of real social media content, where fake images may be fully generated or manipulated through image editing strategies [75]. Furthermore, most existing datasets primarily focus on binary real/fake classification or tampered region localization, with limited emphasis on explaining the cues that models use to make these decisions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.805 + ], + "angle": 0, + "content": "To address these challenges, we introduce the Social media Image Detection dataSet (SID-Set), which consists of 300K images (i.e., 100K real, 100K synthetic, and 100K tampered images), providing a comprehensive resource for the deepfake detection community. Additionally, we include textual descriptions explaining the model's judgment basis. As shown in Figures 2 and 4, synthetic and tampered images are indistinguishable to humans. In particular, challenges for the SID-Set include: 1) subtle alterations of just dozens of pixels; 2) natural-looking local manipulations; 3) complex scenes in datasets. To our knowledge, SID-Set is the first dataset of its scale with extensive annotations, making it the largest and most comprehensive dataset for social media deepfake detection to date. Compared to existing datasets in Table 1, SID-Set addresses the challenges of limited diversity and outdated generative techniques by providing a more comprehensive set of high-quality and diverse images. Accordingly, we propose a new VLMs-based deepfake detection framework, named the Social media Image Detection, localization, and explanation Assistant (SIDA), which achieves the state-of-the-art (SOTA) performance on SID-Set and generalizes effectively across other benchmarks. SIDA can serve as a baseline model on SID-Set, offering a new framework for tackling social media image deepfake detection and localization." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.808, + 0.451, + 0.822 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We establish SID-Set, a comprehensive benchmark for detecting, localizing, and explaining deepfakes in social media images, featuring multiple image types and extensive annotations. SID-Set holds the potential for advancing the field of deepfake detection and ensuring" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.09, + 0.887, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison with existing image deepfake datasets." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.115, + 0.898, + 0.239 + ], + "angle": 0, + "content": "
DatasetContentData SourceGenerator YearMulticlasslesMasksExplanation
OHImg [39]OverheadGoogle Map2023✘✘✘
FakeSpotter [63]FaceCelebA, FFHQ2020✘✓✘
ForgeryNet [24]FaceCREMA-D2021✘✓✘
DCFace [27]FaceFFHQ2023✘✘✘
DFF [58]FaceIMDB-WIKI2023✘✘✘
RealFaces [48]FacePrompts2023✘✘✘
M3Dsynth [87]BiologyDDPM, CycleGAN2023✘✓✘
CNNSpot [64]ObjectSDXL2020✓✘✘
CIFAKE [3]ObjectCIFAR2023✓✘✘
CASIA 2.0 [12]GeneralCorel2022✘✓✘
Artifact [53]GeneralCOCO, FFHQ, LSUN2023✘✘✘
IMD2020 [45]GeneralPlaces22020✘✓✘
AIGCD [84]GeneralLSUN, COCO, FFHQ2023✘✘✘
GenImage [86]GeneralImageNet2023✘✘✘
SID-SetGeneralCOCO, Flickr30k, MagicBrush2024✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.251, + 0.904, + 0.28 + ], + "angle": 0, + "content": "robust model performance in complex real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.282, + 0.905, + 0.356 + ], + "angle": 0, + "content": "- We propose SIDA, a new image deepfake detection, localization, and explanation framework that not only detects images with high accuracy but also localizes and explains potential manipulations, enhancing the transparency and utility of deepfake detection technologies." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.357, + 0.905, + 0.446 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that SIDA effectively identifies and delineates tampered areas within images, supporting the development of more robust and interpretable deepfake detection systems. Notably, SIDA demonstrates superior or equivalent performance on the SID-Set and other benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.282, + 0.905, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.461, + 0.655, + 0.476 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.743, + 0.501 + ], + "angle": 0, + "content": "2.1. Image Deepfake Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In the realm of deepfake detection, the primary focus has historically centered on the identification of facial deepfakes. Renowned datasets such as ForgeryNet [24], DeepFakeFace [58], and DFFD [7] have been pivotal in this area. As the field evolves, there is a growing shift among researchers towards exploring non-facial deepfake data. Advanced methodologies involving text-to-image or imaged-to-image generation techniques [76], utilizing GANs or the stable diffusion series, have facilitated the creation of expansive deepfake datasets like GenImage [86], HiFi-IFDL [21], and DiffForensics [66]. These datasets are characterized by their enlarged data volumes, diversified generation methodologies, and enriched annotation details. Furthermore, beyond the conventional real/fake annotations, certain datasets [21, 22, 73] now include more granular annotations. Table 1 delineates a detailed comparison among various deepfake datasets, highlighting key differences in generation scenarios and annotation types supported. It shows that SID-Set is particularly tailored towards social media data, incorporating the latest SOTA generation models, emphasizing high-quality production, and providing much more comprehensive and diverse annotations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.884, + 0.865 + ], + "angle": 0, + "content": "2.2. Image Deepfake Detection and Localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Deepfake detection methods [4, 6, 38, 49, 62] are typically approached as classification tasks within the data-driven" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.096, + 0.48, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.195, + 0.483, + 0.224 + ], + "angle": 0, + "content": "Figure 2. SID-Set examples. The 1st row is the synthetic images, while the 2nd row shows tampered images. (Zoom in to view)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.486, + 0.642 + ], + "angle": 0, + "content": "paradigm. These strategies primarily leverage diverse architectures [38, 49], including Convolutional Neural Networks (CNNs) and Transformers, to detect distinctive artifacts. Some scholars have attempted to achieve relatively high precision and generality by employing strategies such as employing techniques such as data augmentation [62], adversarial training [6], reconstruction [4], etc. On the other hand, some researchers [25, 60] have explored extracting features from the frequency domain for deepfake identification. Efforts [13, 65] have also been made to fuse features from both spatial and frequency domains to obtain a more comprehensive set of discriminative features for deepfake detection. Although these methods have shown some progress, they still struggle with issues of generalization. Furthermore, some scholars [20, 40, 44, 61, 80, 83] have gone beyond the basic classification between real and fake by gradually constructing datasets annotated with masks of locally tampered areas, thereby addressing both image deepfake detection and localization tasks. However, these datasets are concentrated mainly on facial data, with fewer datasets available for non-facial deepfake detection and localization, and even fewer large and public datasets for realistic social media data. As a result, our work aims to address these critical gaps by providing a new, comprehensive dataset that includes diverse manipulations beyond facial data, particularly focusing on social media images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.33, + 0.669 + ], + "angle": 0, + "content": "2.3. Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.884 + ], + "angle": 0, + "content": "The progress in large language models (LLMs) [14-16, 41] and vision-language models (VLMs) [14-16, 29, 71, 78] has notably improved multimodal comprehension, seamlessly integrating visual and textual data. The LLaMA series [14-16] optimize language understanding with a compact, high-performance design using fewer parameters than prior models. LLaVA series [34, 35] enhance visual question answering by synchronizing visual features with textual data. Additionally, LISA series [29, 71] employ LLM for accurate image segmentation, merging visual perception with linguistic insights to precisely segment the targeted areas. Several grounding large multimodal models [50, 54, 55, 67, 68, 72, 74, 77, 81] have been proposed to localize the contents based on linguistic information." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Advancements [5, 10, 19, 69, 82] in integrating multi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.41 + ], + "angle": 0, + "content": "modal data such as visual and linguistic information have also significantly improved the performance of models in detecting and pinpointing deepfakes. For instance, AntifakePrompt [5] approaches deepfake detection by formulating it as a visual question answering problem, which adjusts soft prompts for InstructBLIP [10], enabling it to determine whether a query image is real or fake. ForgeryGPT [31] enhances image forgery detection and localization by integrating advanced forensic knowledge with a mask-aware forgery extractor that targets pixel-level fraud detection. As a concurrent work, FakeShield [69] leverages the capabilities of LLaVA to identify and localize altered regions while providing interpretable insights into the findings. Our method diverges significantly from existing approaches by creating the most extensive deepfake dataset tailored specifically for social media images with comprehensive annotations. Additionally, we also set a new standard for social media image deepfake detection by utilizing the visual interpretation strengths of VLMs to boost detection accuracy, pinpoint forgeries, and provide clearer explanations within a comprehensive framework." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.427, + 0.637, + 0.442 + ], + "angle": 0, + "content": "3. Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.453, + 0.636, + 0.468 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.793 + ], + "angle": 0, + "content": "In the realm of deepfake detection and localization, research has predominantly focused on facial deepfakes due to their substantial societal impact. However, with the advancements in generative technology, the scope of deepfakes has extended beyond facial content to include non-facial manipulations. Historically, non-facial deepfakes were less prevalent, largely limited by technological constraints that produced low-quality, easily detectable forgeries. Although some datasets, such as GenImage [86] and AIGCD [84], have been constructed, they suffer from several limitations: 1) They often utilize relatively outdated generative technologies, resulting in lower quality data easily distinguished by humans. 2) They primarily focus on text-to-image or image-to-image generation, neglecting the need for data involving manipulations of specific regions, objects, or parts. Such tampered manipulations can be especially insidious as they introduce subtle misinformation, making existing SOTA deepfake detection methods less effective. 3) They lack well-defined criteria for content authenticity, limiting their effectiveness in providing interpretative insights and educating the public on distinguishing synthetic content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Addressing these limitations is crucial for improving the transparency and utility of deepfake detection systems. Furthermore, most existing datasets emphasize either fully synthetic or tampered images, as shown in Table 1, but in real applications, we don't know the image deepfake type in advance. Effective deepfake detection and localization methods should be capable of addressing both scenarios, as" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.104, + 0.315, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.104, + 0.498, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.104, + 0.677, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.104, + 0.861, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 3. Tampered image generation pipeline: It consists of four stages—extracting objects from captions using GPT-4o, obtaining object masks with Language-SAM, setting up replacement dictionaries for generating tampered images, and generating new images using Latent Diffusion. This figure illustrates an example of object replacement (e.g., “cat” to “dog”) and attribute modification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.474 + ], + "angle": 0, + "content": "social media images often involve complex combinations of synthetic and tampered content. Therefore, developing a comprehensive benchmark for detecting and localizing deepfakes in social media images is essential. We propose SID-Set, which encompasses comprehensive, high-quality annotations for detection and localization, along with detailed textual explanations of the judgment criteria." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.325, + 0.497 + ], + "angle": 0, + "content": "3.2. Benchmark Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Data Details. To develop an effective benchmark for detecting and localizing images on social media, we created SID-Set, a dataset with real, synthetic, and tampered images reflecting diverse real-world scenarios. Our benchmark assesses whether models can differentiate among real, synthetic, and tampered images, as well as accurately identify altered regions in tampered images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.611, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Real Images: 100K images from OpenImages \\(\\mathrm{V}7^{1}\\), with a wide range of scenarios reflecting real-world diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Synthetic Images: 100K images generated through FLUX [43], specifically designed to challenge identification due to their high-quality, highly realistic appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Tampered Images: 100K tampered images, with specific objects or regions replaced or altered; the detailed generation process is shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Data Generation. To generate highly realistic synthetic images, we experimented with several open-source SOTA generative models, such as FLUX [43], Kandinsky 3.0 [1], SDXL [52], AbsoluteReality [42], and others. Following a review by five human experts of 1,000 images from each generative model, FLUX emerged as the top performer, producing highly convincing images that were indistinguishable from real ones to human experts. Consequently, we employed FLUX to create 100K synthetic images based on" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.366, + 0.58, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.409, + 0.562, + 0.417 + ], + "angle": 0, + "content": "bird" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.366, + 0.637, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.367, + 0.694, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.41, + 0.683, + 0.418 + ], + "angle": 0, + "content": "eagle" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.419, + 0.58, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.465, + 0.559, + 0.473 + ], + "angle": 0, + "content": "cat" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.419, + 0.638, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.465, + 0.681, + 0.474 + ], + "angle": 0, + "content": "dog" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.366, + 0.772, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.409, + 0.77, + 0.418 + ], + "angle": 0, + "content": "bodyboard" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.366, + 0.831, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.41, + 0.893, + 0.419 + ], + "angle": 0, + "content": "bodyboard" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.419, + 0.771, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.464, + 0.772, + 0.472 + ], + "angle": 0, + "content": "coffee table" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.419, + 0.831, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.464, + 0.894, + 0.472 + ], + "angle": 0, + "content": "coffee table" + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.485, + 0.887, + 0.5 + ], + "angle": 0, + "content": "Figure 4. Examples of tampered images. (Zoom in to view)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.905, + 0.561 + ], + "angle": 0, + "content": "the Flickr30k [51] and COCO [33]. The image tampering process depicted in Figure 4 follows four distinct stages, utilizing the COCO image as an example." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.906, + 0.652 + ], + "angle": 0, + "content": "Stage 1: We extract objects from an image's caption using GPT-4o [47]. For instance, from the caption \"A large fluffy cat laying on top of a wooden table\", GPT-4o identifies relevant COCO class objects or retains nouns if no match exists. This extraction is documented in an \"Image-Caption-Object\" JSON file." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.905, + 0.684 + ], + "angle": 0, + "content": "Stage 2: Employing Language-SAM [30], we generate masks for identified objects as training ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Stage 3: We establish dictionaries for full and partial image tampering using COCO classes for object replacement and attribute modifications, respectively. For example, replacing \"dog\" with animals like \"cat\" or adding attributes such as \"happy\" or \"angry\" to the \"dog\" class. For more details, please refer to the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Stage 4: Utilizing Latent Diffusion [56], we modify captions and regenerate images, either replacing or retaining original objects based on availability. An example modification is altering \"cat\" to \"dog\" in the image caption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In total, we generated 80,000 object-tampered images and 20,000 partially tampered images. To demonstrate the explainability of SIDA, we used GPT-4o to generate textual descriptions for the judgment basis of 3,000 images from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "1https://storage.googleapis.com/openimages/web/index.html" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28834" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.093, + 0.837, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.379, + 0.908, + 0.422 + ], + "angle": 0, + "content": "Figure 5. The pipeline of SIDA: Given an image \\( x_{i} \\) and the corresponding text input \\( x_{t} \\), the last hidden layer for the token provides the detection result. If the detection result indicates a tampered image, SIDA extracts the token to generate masks for the tampered regions. This figure shows an example where the man's face has been manipulated." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.484, + 0.538 + ], + "angle": 0, + "content": "the SID-Set, divided equally among three types. Additionally, to ensure the realism of synthetic images, tampered images, and their textual descriptions, we engaged 5 annotation experts for quality control and adjustments. Detailed prompts and generation pipelines used to create descriptions for each type of image are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.184, + 0.568 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.625 + ], + "angle": 0, + "content": "In this section, we first present the model architecture of SIDA in Section 4.1, followed by an introduction to the training process of our method in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.636, + 0.226, + 0.65 + ], + "angle": 0, + "content": "4.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Large vision-language models have demonstrated remarkable capabilities in understanding the alignment between textual and visual information. For instance, LLaVA [34] leverages language alone to achieve a comprehensive understanding of both visual and linguistic content. Building on LLaVA, LISA [29] extends this capability by providing fine-grained segmentation masks along with corresponding textual descriptions. However, to effectively detect and localize synthetic images, VLMs must not only be capable of multimodal understanding but also possess the ability to identify and segment manipulated regions, providing detailed explanations for both synthetic and tampered images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To this end, we propose SIDA to tackle the task of synthetic image detection and tampered region localization. The pipeline of our method is illustrated in Figure 5. Inspired by previous approaches [28, 29, 85], we expand the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.583 + ], + "angle": 0, + "content": "original vocabulary of VLMs by adding two new tokens, \\(<\\mathrm{DET}>\\) and \\(<\\mathrm{SEG}>\\), to enable the model to extract detection and segmentation information. Given an image \\(x_{i}\\) and a text prompt \\(x_{t}\\), such as \"Can you identify if this image is real, fully synthetic, or tampered? Please mask the tampered object/part if it is tampered.\" We feed them into the VLM. The VLM then outputs a text description \\(\\hat{y}_{\\mathrm{des}}\\), while the last hidden layer \\(h_{\\mathrm{hid}}\\) contains the \\(<\\mathrm{DET}>\\) and \\(<\\mathrm{SEG}>\\) tokens. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.598, + 0.905, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {\\mathrm {d e s}} = \\operatorname {V L M} \\left(x _ {i}, x _ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Next, we extract the \\(<\\mathrm{DET}>\\) token from the last hidden layer \\(h_{\\mathrm{hid}}\\) to obtain \\(h_{\\mathrm{det}}\\). The representation \\(h_{\\mathrm{det}}\\) is then passed through a detection head \\(F_{\\mathrm{det}}\\) to determine whether the image is real, fully synthetic, or tampered. We denote the final detection result by \\(\\hat{\\mathsf{D}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.711, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathrm {D}} = F _ {\\text {d e t}} \\left(h _ {\\text {d e t}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.765 + ], + "angle": 0, + "content": "\\(F_{\\mathrm{det}}\\) is the detection head that processes the extracted \\(\\langle \\mathrm{DET}\\rangle\\) representation \\(h_\\mathrm{det}\\) to produce the detection output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.903 + ], + "angle": 0, + "content": "If the detection result indicates that the image has been tampered with, SIDA will then predict a mask for the tampered regions. The \\( h_{\\mathrm{seg}} \\) feature is extracted from the hidden layer \\( h_{\\mathrm{hid}} \\), similar to the extraction process for the token. Given that the token encapsulates crucial information that can aid in generating fine-grained segmentation masks, the representation \\( h_{\\mathrm{det}} \\) is transformed using a fully connected layer \\( F \\) to align with the dimensions of \\( h_{\\mathrm{seg}} \\). To further capture the relationship between the \\( h_{\\mathrm{det}} \\) and \\( h_{\\mathrm{seg}} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28835" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "features, we apply a single-layer Multihead Attention, facilitating effective feature interaction and enhancing mask quality. The process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.149, + 0.316, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\tilde {h} _ {\\mathrm {d e t}} = F (h _ {\\mathrm {d e t}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.169, + 0.482, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\tilde {h} _ {\\text {s e g}} = \\operatorname {M S A} \\left(\\tilde {h} _ {\\det }, h _ {\\text {s e g}}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.19, + 0.336, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\tilde {h} _ {\\mathrm {s e g}} = \\tilde {h} _ {\\mathrm {s e g}} + h _ {\\mathrm {s e g}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.295 + ], + "angle": 0, + "content": "In this formulation, we treat the detection features as the query, and the segmentation features as the key and value. A residual connection is employed to combine both the original and transformed information, thereby enhancing the representation for precise segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.37 + ], + "angle": 0, + "content": "Finally, we employ a frozen image encoder \\( F_{\\mathrm{enc}} \\) to extract visual features from the input image \\( x_{i} \\), resulting in visual features \\( f \\). The segmentation embedding \\( \\tilde{h}_{\\mathrm{seg}} \\) is then combined with the visual features \\( f \\) and fed into a decoder to produce the final mask \\( \\hat{M} \\). This can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.383, + 0.323, + 0.399 + ], + "angle": 0, + "content": "\\[\nf = F _ {\\mathrm {e n c}} (x _ {i}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.397, + 0.482, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathrm {M}} = F _ {\\mathrm {d e c}} \\left(\\tilde {h} _ {\\mathrm {s e g}}, f\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.428, + 0.196, + 0.445 + ], + "angle": 0, + "content": "4.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Training Objectives. The training loss, \\(\\mathcal{L}\\), for the SIDA consists of three components: the detection loss \\(\\mathcal{L}_{det}\\), the text generation loss \\(\\mathcal{L}_{txt}\\), and the segmentation mask loss \\(\\mathcal{L}_{mask}\\). Initially, SIDA is trained in an end-to-end manner by employing the detection loss and the segmentation loss. For detection, we use CrossEntropy loss, while for the segmentation task, we use a weighted combination of binary cross-entropy (BCE) and DICE loss, with respective loss weights \\(\\lambda_{\\mathrm{bce}}\\) and \\(\\lambda_{\\mathrm{dice}}\\). This can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.599, + 0.352, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.617, + 0.482, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e t} = \\mathcal {L} _ {\\mathrm {C E}} (\\hat {\\mathrm {D}}, \\mathrm {D}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.635, + 0.441, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {B C E} (\\hat {\\mathrm {M}}, \\mathrm {M}) + \\lambda_ {d i c e} \\mathcal {L} _ {D I C E} (\\hat {\\mathrm {M}}, \\mathrm {M}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.753 + ], + "angle": 0, + "content": "After completing the training phase, we proceed to fine-tune the SIDA model by utilizing detailed textural descriptions from 3,000 images as the ground truth, represented by \\( y_{des} \\). This phase focuses on optimizing the text generation component, \\( \\mathcal{L}_{txt} \\), to improve its ability in textural interpretability. The final loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.766, + 0.303, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t x t} = \\mathcal {L} _ {\\mathrm {C E}} (y _ {\\hat {d e s}}, y _ {d e s}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.778, + 0.482, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {t x t}} \\mathcal {L} _ {\\text {t x t}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.856 + ], + "angle": 0, + "content": "where \\(\\lambda_{det}\\), \\(\\lambda_{mask}\\), and \\(\\lambda_{txt}\\) are the weighting factors that balance the contributions of the detection, segmentation, and text generation losses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training Data. We utilize the SID-Set, consisting of 300k images, to train SIDA. To further enhance diversity, we incorporate the MagicBrush dataset [79] after filtering out" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "low-quality images. The combined dataset supports robust training for both the detection and localization of synthetic content. Additionally, we generate descriptions for 3,000 randomly selected images using LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.169, + 0.646, + 0.186 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Implementation Details. We choose LISA as the base large vision language model due to its strong capability for reasoning-based localization. We fine-tuned both LISA-7B-v1 and LISA-13B-v1 on the SID-Set using LoRA, setting \\(\\alpha\\) at 16 and the dropout rate at 0.05. The input images are resized to \\(1024 \\times 1024\\). The loss weights in Eq. (6) for the detection \\((\\lambda_{\\mathrm{det}})\\), the text generation \\((\\lambda_{\\mathrm{txt}})\\), and the localization \\((\\lambda_{\\mathrm{mask}})\\) are set to 1.0, respectively. The localization loss weights in Eq. (5) for the \\(\\lambda_{\\mathrm{bce}}\\) and \\(\\lambda_{\\mathrm{dice}}\\) are set to 2.0 and 0.5, respectively. To determine the optimal weight configuration, we perform ablation studies as detailed in Section 5.5. During the detection and localization training stage, the image encoder is frozen, and all other modules are trainable. For the text generation stage, only vision-language models are fine-tuned using the LoRA strategy. The initial learning rate is set to \\(1 \\times 10^{-4}\\), with a batch size of 2 per device and a gradient accumulation step of 10. We use two NVIDIA A100 GPUs (40GB each). Training for SIDA-7B and SIDA-13B took 48 hours and 72 hours, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate detection using image-level accuracy and F1 scores. For forgery localization, our metrics include Area Under the Curve (AUC), F1 scores, and Intersection over Union (IoU)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.556, + 0.712, + 0.57 + ], + "angle": 0, + "content": "5.1. Detection Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.806 + ], + "angle": 0, + "content": "We compare SIDA against other SOTA deepfake detection methods on SID-Set, including CnnSpot [17], AntifakePrompt [5], FreDect [18], Fusing [26], Gram-Net [37], UniFDP [46], LGrad [59], and LNP [2]. To ensure a fair comparison, we first evaluate these models on our dataset using their original pre-trained weights, then retrain them with the SID-Set to assess performance improvements. Table 2 demonstrates that SIDA achieves better or comparable results among all the evaluated methods. Notably, LGrad [59] achieves the highest accuracy and F1 score on tampered images after retraining, but this comes at the expense of lower performance in other metrics. Our analysis indicates that LGrad's high recall and false positive rates stem from its propensity to misclassify other types as tampered. The training details are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.706, + 0.832 + ], + "angle": 0, + "content": "5.2. Localization Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Table 3 presents the forgery localization performance on the SID-Set. We selected PSCC-Net [36], MVSS-Net [11], and HIFI-Net [21] as representative IFDL methods. Additionally, we chose LISA [29] as a representative LLM due to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28836" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison of SIDA with other deepfake detection methods. Values outside parentheses are evaluated with open-source models directly on the SID-Set; values in parentheses indicate performance changes after training with the SID-Set. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.129, + 0.894, + 0.314 + ], + "angle": 0, + "content": "
MethodsYearRealFully syntheticTamperedOverall
AccF1AccF1AccF1AccF1
AntifakePrompt [5]202464.8(↑24.1)78.6(↑10.5)93.8(↑3.7)96.8(↑1.1)30.8(↑60.1)47.2(↑33.2)63.1(↑29.1)69.3(↑23.4)
CnnSpott [17]202179.8(↑9.2)88.7(↑2.1)39.5(↑51.2)56.6(↑31.5)6.9(↑61.2)12.9(↑51.1)42.1(↑39.3)69.6(↑20.7)
FreDect [18]202083.7(↓37.7)91.1(↓43.5)16.8(↑44.1)28.8(↑37.2)11.9(↑25.2)21.3(↑31.7)37.4(↑33.6)23.4(↑47.2)
Fusing [26]202285.1(↑4.1)92.0(↑0.7)34.0(↑54.1)50.7(↑38.4)2.7(↑24.3)5.3(↑26.1)40.1(↑33.1)29.1(↑40.3)
Gram-Net [37]202070.1(↑19.1)82.4(↑9.3)93.5(↑4.4)96.6(↑2.0)0.8(↑89.1)1.6(↑85.3)55.0(↑37.1)58.0(↑35.1)
UnivFD [46]202368.0(↑0.3)67.4(↑1.1)62.1(↑24.3)87.5(↑10.5)64.0(↑28.5)85.3(↑4.7)64.0(↑21.7)85.3(↑4.5)
LGrad [59]202364.8(↓2.8)78.6(↓2.5)83.5(↓25.5)91.0(↓23.7)6.8(↑92.3)12.7(↑86.1)51.8(↑20.2)55.5(↑23.9)
LNP [2]202371.2(↓56.8)83.2(↓60.2)91.8(↓55.6)95.7(↓60.1)2.9(↑90.4)5.7(↑88.9)55.2(↓7.6)58.2(↑4.1)
SIDA-7B202489.191.098.798.691.291.093.593.5
SIDA-13B202489.691.198.598.792.991.293.693.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.446 + ], + "angle": 0, + "content": "its segmentation reasoning capabilities. We used LISA-7Bv1 and fine-tuned it on SID-Set. The results indicate that SIDA achieves the best performance. We suppose that while LISA possesses strong general segmentation capabilities, it lacks the specific specialized features required to detect subtle manipulations, ultimately limiting the effectiveness of fine-tuning for precise forgery localization." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.494 + ], + "angle": 0, + "content": "Table 3. Comparison between SIDA and other IFDL approaches. * indicates the use of the pre-trained model from the original paper due to unavailable training code." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.504, + 0.456, + 0.643 + ], + "angle": 0, + "content": "
MethodsYearsTampered
AUCF1IOU
MVSS-Net* [11]202348.931.623.7
HIFI-Net* [21]202364.045.921.1
PSCC-Net [36]202282.171.335.7
LISA-7B-v1 [29]202478.469.132.5
SIDA-7B202487.373.943.8
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.265, + 0.669 + ], + "angle": 0, + "content": "5.3. Robustness Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.841 + ], + "angle": 0, + "content": "We further evaluate the robustness of SIDA against common image perturbations found in social media, such as JPEG compression, resizing, and Gaussian noise. Table 4 shows our model's performance on the SID-Set under six degradation scenarios: JPEG compression (with quality levels of 70 and 80), resizing (with scaling factors of 0.5 and 0.75), and Gaussian noise (with variances of 5 and 10). Despite not being explicitly trained on degraded data, SIDA demonstrates resilience to these low-level distortions. The model's stable performance against common social media perturbations highlights its robustness and practical applicability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.329, + 0.865 + ], + "angle": 0, + "content": "5.4. Test on Other Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this stage, we evaluate SIDA on DMImage [8] dataset to assess its generalization capabilities. We compare SIDA" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.337, + 0.889, + 0.351 + ], + "angle": 0, + "content": "Table 4. Performance of SIDA under different perturbations." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.355, + 0.887, + 0.498 + ], + "angle": 0, + "content": "
DetectionLocalization
ACCF1AUCF1IOU
JPEG 7089.490.186.271.842.3
JPEG 8088.789.585.871.141.7
Resize 0.589.391.186.872.543.2
Resize 0.7589.991.687.173.043.5
Gaussian 1086.989.384.170.241.0
Gaussian 588.489.985.371.041.5
SIDA-7B93.593.587.373.943.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.6 + ], + "angle": 0, + "content": "with CNNSpot [17], Fusing [26], Gram-Net [37], LNP [2], UnivFD [46], and AntifakePrompt [5]. For these methods, we use the original hyperparameter settings and pre-trained weights provided by the authors. The results in Table 5 show that SIDA achieves superior performance, demonstrating its strong adaptability." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.611, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Table 5. Comparison with other deepfake detection methods on DMImage [8]. We used original weights for each method." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.65, + 0.898, + 0.781 + ], + "angle": 0, + "content": "
MethodsRealFakeOverall
AccF1AccF1AccF1
CNNSpot [17]87.888.428.444.240.643.3
Gram-Net [37]62.854.178.888.167.479.4
Fusing [26]87.786.115.527.240.436.5
LNP [2]63.167.456.972.558.268.3
UnivFD [46]89.488.344.961.253.960.7
AntifakePrompt [5]91.392.589.391.290.691.2
SIDA-7B92.993.190.791.091.892.4
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.667, + 0.82 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Attention Module. We conducted ablation experiments to assess the importance of the attention module. Variants included removing the attention module entirely and replacing it with fully connected (FC) layers. Results in Table 6 show that removing the attention module or replacing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28837" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.077, + 0.468, + 0.091 + ], + "angle": 0, + "content": "Table 6. Ablation study results for attention module in SIDA." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.103, + 0.466, + 0.186 + ], + "angle": 0, + "content": "
DetectionLocalization
ACCF1AUCF1IOU
FC91.190.384.371.638.9
w/o Attention90.389.984.171.338.8
SIDA93.593.587.373.943.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.483, + 0.243 + ], + "angle": 0, + "content": "it with FC significantly reduces performance, underscoring the critical role of attention in enhancing feature interaction and improving detection and localization accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.379 + ], + "angle": 0, + "content": "Training Weights. SIDA training utilizes weighted losses to balance task contributions. In the detection and localization stages, detection loss is adjusted by a weight \\(\\lambda_{\\mathrm{det}}\\), and localization loss by binary cross-entropy (BCE) and DICE losses, with weights \\(\\lambda_{\\mathrm{bce}}\\) and \\(\\lambda_{\\mathrm{dice}}\\) respectively. For our experiments, we set \\(\\lambda_{\\mathrm{det}}\\) to 1, \\(\\lambda_{\\mathrm{bce}}\\) to 2.0, and \\(\\lambda_{\\mathrm{dice}}\\) to 0.5 to maintain balance between detection and localization, enhancing model stability and performance. We summarize the outcomes of various weight configurations in Table 7." + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.385, + 0.46, + 0.4 + ], + "angle": 0, + "content": "Table 7. Different weight configurations in SIDA training." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.405, + 0.424, + 0.468 + ], + "angle": 0, + "content": "
λdetλbceλdiceAccF1 Score
1.02.00.593.5691.01
1.04.01.093.4990.86
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.274, + 0.494 + ], + "angle": 0, + "content": "5.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.622 + ], + "angle": 0, + "content": "In this section, we present examples of SIDA's output for tampered images, showcasing its detection, localization, and explanation capabilities. SIDA accurately identifies tampered regions and provides explanations for its decisions. We also include some challenging failure cases where SIDA was unable to accurately detect the tampered regions, highlighting areas for future improvement. Additional visualization results are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.347, + 0.65 + ], + "angle": 0, + "content": "6. Conclusion and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In this work, we present SID-Set for social image deepfake detection, localization, and explanation tasks, consisting of 100k real images, 100k fully synthetic images, and 100k tampered images. Furthermore, we propose a new VLMs-based deepfake detection framework, SIDA, to address these tasks. SIDA demonstrates its ability to detect fake types, localize tampered regions, and provide explanations for its decisions. We believe that the integration of VLMs into deepfake detection tasks, as demonstrated by SIDA, offers promising new avenues for future research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Although the development of the SID-Set and the introduction of the SIDA framework have yielded favorable outcomes in deepfake detection tasks, we recognize some potential limitations and will optimize them in future research. Dataset Size. While SID-Set includes 100k fully synthetic and 100k tampered images, the complexity of real social" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.108, + 0.796, + 0.128 + ], + "angle": 0, + "content": "Can you identify if this image is real, full synthetic, or tampered? Please mask the tampered object/part if it is tampered." + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.128, + 0.62, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.129, + 0.703, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.128, + 0.782, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.127, + 0.868, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.187, + 0.569, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.569, + 0.212, + 0.603, + 0.229 + ], + "angle": 0, + "content": "Object ampered" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.232, + 0.62, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.631, + 0.212, + 0.693, + 0.23 + ], + "angle": 0, + "content": "Partial tampered" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.232, + 0.706, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.71, + 0.212, + 0.787, + 0.231 + ], + "angle": 0, + "content": "Partial tampered" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.233, + 0.786, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.814, + 0.212, + 0.851, + 0.229 + ], + "angle": 0, + "content": "Object tampered" + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.233, + 0.872, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.29, + 0.594, + 0.35 + ], + "angle": 0, + "content": "Type of Image: object tampered The tampered areas are specifically located in the h The tampered. Absolute Positino The location is" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.29, + 0.701, + 0.35 + ], + "angle": 0, + "content": "Type of Image: part tampered Tampering Type and Location... Relative Position: The tampered part located in the face of the women... Absolute Position..." + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.291, + 0.782, + 0.35 + ], + "angle": 0, + "content": "Type of Image: object \ntampered Tampering \nType and Location is the face of the man... \nAbsolute Position \nTampered Content: Modification... \nInconsistencies: Lighting..." + }, + { + "type": "text", + "bbox": [ + 0.798, + 0.291, + 0.868, + 0.346 + ], + "angle": 0, + "content": "Type of Image: part tampered Tampered Content: Modification... Relative Position: Inconsistencies... Colour in the background looks ..." + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.356, + 0.569, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.64, + 0.358, + 0.748, + 0.369 + ], + "angle": 0, + "content": "a) Correct examples" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.385, + 0.618, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.385, + 0.699, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.385, + 0.783, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.385, + 0.869, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.44, + 0.618, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.44, + 0.699, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.44, + 0.783, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.44, + 0.868, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.499, + 0.619, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.499, + 0.699, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.5, + 0.783, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.5, + 0.868, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.563, + 0.748, + 0.575 + ], + "angle": 0, + "content": "b) Failure examples" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.594, + 0.869, + 0.609 + ], + "angle": 0, + "content": "Figure 6. Visual results of SIDA on tampered images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.905, + 0.678 + ], + "angle": 0, + "content": "media environments demands a larger dataset. Therefore, expanding the dataset with additional images is a crucial objective for future research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.907, + 0.829 + ], + "angle": 0, + "content": "Data Domain. Other methods often generate social media images that lack authenticity or are easily identifiable. Consequently, we used only FLUX to generate fake images due to its superior quality. However, depending exclusively on one method may lead to issues of data skew, although this was not significantly evident in our experiments. Such skew could potentially impair performance on diverse datasets. Moving forward, we plan to explore additional generative methods and integrate various other generation techniques to produce a more varied and higher-quality set of images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.83, + 0.905, + 0.89 + ], + "angle": 0, + "content": "Localization Results. Although SIDA demonstrates relatively strong performance on our dataset, there is still room for improvement. Certain tampered regions are not reliably detected, underscoring the need for further advancements." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28838" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.157 + ], + "angle": 0, + "content": "This work is supported by The Alan Turing Institute (UK) through the project 'Turing-DSO Labs Singapore Collaboration' (SDCfP2\\100009)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.184, + 0.188, + 0.2 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.21, + 0.484, + 0.264 + ], + "angle": 0, + "content": "[1] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. Arxiv, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.266, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[2] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. Detecting generated images by real images only. Arxiv, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.309, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[3] Jordan J. Bird and Ahmad Lotfi. CIFAKE: image classification and explainable identification of ai-generated synthetic images. IEEE Access, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.483, + 0.405 + ], + "angle": 0, + "content": "[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[5] You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors. Arxiv, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[6] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[7] Harry Cheng, Yangyang Guo, Tianyi Wang, Liqiang Nie, and Mohan S. Kankanhalli. Diffusion facial forgery detection. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[8] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[9] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.662, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[11] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE TPAMI, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Jing Dong, Wei Wang, and Tieniu Tan. CASIA image tampering detection evaluation database. In ISIP, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[13] Junxian Duan, Yang Ai, Jipeng Liu, Shenyuan Huang, Huaibo Huang, Jie Cao, and Ran He. Test-time forgery detection with spatial-frequency prompt learning. IJCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[14] Abhimanyu Dubey et al. The llama 3 herd of models. Arxiv, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.21, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[15] Hugo Touvron et al. Llama: Open and efficient foundation language models. Arxiv, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[16] Hugo Touvron et al. Llama 2: Open foundation and finetuned chat models. Arxiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.152, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[17] Joel Frank and Thorsten Holz. Cnn-generated images are surprisingly easy to spot...for now. Arxiv, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[18] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[19] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[20] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[21] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[22] Xiao Guo, Xiaohong Liu, Iacopo Masi, and Xiaoming Liu. Language-guided hierarchical fine-grained image forgery detection and localization. Arxiv, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[23] Ruidong Han, Xiaofeng Wang, Ningning Bai, Yihang Wang, Jianpeng Hou, and Jianru Xue. HDF-net: Capturing homogeneity difference features to localize the tampered image. IEEE TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[24] Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. Forgerynet: A versatile benchmark for comprehensive forgery analysis. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[25] Yonghyun Jeong, Doyeon Kim, Youngmin Ro, and Jongwon Choi. Frepgan: Robust deepfake detection using frequency-level perturbations. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[26] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In ICIP, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[27] Minchul Kim, Feng Liu, Anil K. Jain, and Xiaoming Liu. Dcface: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[29] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: reasoning segmentation via large language model. In CVPR, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[30] lang-sam team. Lang-segment-anything. https://github.com/luca-medeiros/lang-segment-anything, 2024. Accessed: 2024-11-15. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Jiawei Li, Fanrui Zhang, Jiaying Zhu, Esther Sun, Qiang Zhang, and Zheng-Jun Zha. Forgerygpt: Multimodal large language model for explainable image forgery detection and localization. Arxiv, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28839" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[32] Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. Detecting multimedia generated by large AI models: A survey. Arxiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[36] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for image manipulation detection and localization. IEEE TCSVT, 2022, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[37] Zhengzhe Liu, Xiaojuan Qi, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. In CVPR, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[38] Momina Masood, Marriam Nawaz, Khalid Mahmood Malik, Ali Javed, Ain Irtaza, and Hafiz Malik. Deepfakes generation and detection: state-of-the-art, open challenges, countermeasures, and way forward. Appl. Intell., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[39] Brandon B. May, Kirill Trapeznikov, Shengbang Fang, and Matthew C. Stamm. Comprehensive dataset of synthetic and manipulated overhead imagery for development and evaluation of forensic tools. In IH&MMSec, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[40] Changtao Miao, Qi Chu, Zhentao Tan, Zhenchao Jin, Wanyi Zhuang, Yue Wu, Bin Liu, Honggang Hu, and Nenghai Yu. Multi-spectral class center network for face manipulation detection and localization. Arxiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[41] Shervin Minaee, Tomás Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. Arxiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.582, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[42] Absolutereality model team. Absolutereality model repository. https://huggingface.co/jochemstoel/absolutereality-model-repository, 2024. Accessed: 2024-11-15. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[43] Flux model team. Flux model. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. Accessed: 2024-11-15. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[44] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[45] Adam Novozámsky, Babak Mahdian, and Stanislav Saic. IMD2020: A large-scale annotated dataset tailored for detecting manipulated images. In WACvw, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[46] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.413, + 0.845 + ], + "angle": 0, + "content": "[47] OpenAI. GPT-4 technical report. Arxiv, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[48] Lorenzo Papa, Lorenzo Faiella, Luca Corvitto, Luca Maiano, and Irene Amerini. On the use of stable diffusion for creating realistic faces: from generation to detection. In IWBF, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[49] Gan Pei, Jiangning Zhang, Menghan Hu, Guangtao Zhai, Chengjie Wang, Zhenyu Zhang, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey. Arxiv, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.201 + ], + "angle": 0, + "content": "[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. Arxiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[51] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[53] Md Awsafur Rahman, Bishmoy Paul, Najibul Haque Sarker, Zaber Ibn Abdul Hakim, and Shaikh Anowarul Fattah. Artifact: A large-scale dataset with artificial and factual images for generalizable and robust synthetic image detection. In ICIP, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[54] Hanoona Abdul Rasheed, Muhammad Maaz, Sahal Shaji Mullappilly, Abdelrahman M. Shaker, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Eric P. Xing, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Glamm: Pixel grounding large multimodal model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[55] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[57] Xincheng Shuai, Henghui Ding, Xingjun Ma, Rongcheng Tu, Yu-Gang Jiang, and Dacheng Tao. A survey of multimodal-guided image editing with text-to-image diffusion models. Arxiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[58] Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. Arxiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[59] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[60] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[61] Dragos-Constantin Tântaru, Elisabeta Oneata, and Dan Oneata. Weakly-supervised deepfake localization in diffusion-generated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[62] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[63] Run Wang, Felix Juefei-Xu, Lei Ma, Xiaofei Xie, Yihao Huang, Jian Wang, and Yang Liu. Fakespotter: A simple yet robust baseline for spotting ai-synthesized fake faces. In IJCAI, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28840" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[64] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[65] Yuan Wang, Kun Yu, Chen Chen, Xiyuan Hu, and Silong Peng. Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[66] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[67] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. Arxiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[68] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. GSVA: generalized segmentation via multimodal large language models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[69] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. Arxiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[70] Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A comprehensive benchmark of deepfake detection. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[71] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. *Arxiv*, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[72] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[73] Zeqin Yu, Jiangqun Ni, Yuzhen Lin, Haoyi Deng, and Bin Li. Diffforensics: Leveraging diffusion prior to image forgery detection and localization. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[74] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Arxiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[75] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, Lingjie Liu, Adam Kortylewski, Christian Theobalt, and Eric P. Xing. Multimodal image synthesis and editing: The generative AI era. IEEE TPAMI, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[76] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion models in generative AI: A survey. Arxiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[77] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, and Jainwei Yang. Llava-grounding: Grounded visual chat with large multimodal models. In ECCV, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[78] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE TPAMI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[80] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[81] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Change Loy Chen, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[82] Yue Zhang, Ben Colman, Ali Shahriyari, and Gaurav Bharaj. Common sense reasoning for deep fake detection. *Arxiv*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[83] Yi Zhang, Changtao Miao, Man Luo, Jianshu Li, Wenzhong Deng, Weibin Yao, Zhe Li, Bingyu Hu, Weiwei Feng, Tao Gong, and Qi Chu. MFMS: learning modality-fused and modality-specific features for deepfake detection and localization tasks. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[84] Nan Zhong, Yiran Xu, Zhenxing Qian, and Xinpeng Zhang. Rich and poor texture contrast: A simple yet effective approach for ai-generated image detection. Arxiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[85] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and VQA. In AAAI, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[86] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In NeurIPS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[87] Giada Zingarini, Davide Cozzolino, Riccardo Corvi, Giovanni Poggi, and Luisa Verdoliva. M3DSYNTH: A dataset of medical 3d images with ai-generated local manipulations. In ICASSP, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28841" + } + ] +] \ No newline at end of file diff --git a/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_origin.pdf b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c657f18a5b24fc9681844fa88c21a78bf5efe401 --- /dev/null +++ b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/d78f919f-a503-4ae4-a9af-9158aa41b47d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:219e9e2eeb4efdc8f2d3d3993c9d3b337f4c73f486a803f4c9732d4fa725ef76 +size 873197 diff --git a/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/full.md b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3713ba30fda55ee90f9531a22ae88a9d3bca2dac --- /dev/null +++ b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/full.md @@ -0,0 +1,471 @@ +# SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model + +Zhenglin Huang $^{1}$ Jinwei Hu $^{1}$ Xiangtai Li $^{2\dagger}$ Yiwei He $^{1}$ +Xingyu Zhao $^{3}$ Bei Peng $^{1}$ Baoyuan Wu $^{4}$ Xiaowei Huang $^{1}$ Guangliang Cheng $^{1\dagger}$ + +1University of Liverpool, UK 2Nanyang Technological University, SG + +3 WMG, University of Warwick 4The Chinese University of Hong Kong, Shenzhen, Guangdong, China + +Project Page: https://hzlsaber.github.io/projects/SIDA/ + +† Corresponding author. E-mail: guangliang.cheng@liverpool.ac.uk xiangtai94@gmail.com + +# Abstract + +The rapid advancement of generative models in creating highly realistic images poses substantial risks for misinformation dissemination. For instance, a synthetic image, when shared on social media, can mislead extensive audiences and erode trust in digital content, resulting in severe repercussions. Despite some progress, academia has not yet created a large and diversified deepfake detection dataset for social media, nor has it devised an effective solution to address this issue. In this paper, we introduce the Social media Image Detection dataSet (SID-Set), which offers three key advantages: (1) extensive volume, featuring 300K AI-generated/tampered and authentic images with comprehensive annotations, (2) broad diversity, encompassing fully synthetic and tampered images across various classes, and (3) elevated realism, with images that are predominantly indistinguishable from genuine ones through mere visual inspection. Furthermore, leveraging the exceptional capabilities of large multimodal models, we propose a new image deepfake detection, localization, and explanation framework, named SIDA (Social media Image Detection, localization, and explanation Assistant). SIDA not only discerns the authenticity of images, but also delineates tampered regions through mask prediction and provides textual explanations of the model's judgment criteria. Compared with state-of-the-art deepfake detection models on SID-Set and other benchmarks, extensive experiments demonstrate that SIDA achieves superior performance among diversified settings. The code, model, and dataset will be released. + +# 1. Introduction + +Recent advances in generative AI [9, 57, 75] have significantly improved the ability to generate highly realistic images, making it easier to create content that closely resem + +![](images/bd1bf6fb12dc6a8e113de6aab54d4e307954fc8979f8f8bee3e359a79aed1bbb.jpg) +Figure 1. The framework comparisons. Existing deepfake methods (a-b) are limited to detection, localization, or both. In contrast, SIDA (c) offers a more comprehensive solution, capable of handling detection, localization, and explanation tasks. + +bles real-world events. However, these advancements also bring new risks of malicious misuse, particularly in creating deceptive content aimed at misleading public opinion or distorting historical records. Such concerns have motivated the computer vision community to develop more sophisticated deepfake detection techniques. Contemporary methods [32, 70] primarily focus on assessing the authenticity of facial images (i.e., real or fake), while an emerging subset aims to detect and localize facial manipulations [21, 22]. These methods are typically trained on datasets containing real and fake images, aiming to detect images as real or fake, or to localize the tampered regions. Consequently, the quality and diversity of the datasets used for training and evaluation play a crucial role in achieving high accuracy in deepfake detection and localization. A well-curated dataset can enable models to learn nuanced features, improving robustness and generalization in real-world scenarios. + +However, existing deepfake detection and localization datasets face two main challenges: 1) Insufficient Diversity. The majority of existing datasets for deepfake detection focus mainly on facial imagery [7, 24]. However, given the growing capabilities of generative AI, the issue of non + +facial image falsification on social media cannot be overlooked. While researchers [66, 86] have developed relatively large datasets based on ImageNet for image deepfake detection, these datasets typically consist of images from simple scenarios that do not specifically focus on social media. Additionally, they often utilize somewhat outdated image-generation techniques, which can result in less convincing forgeries that are easier for both humans and models to detect. Currently, there is a substantial lack of large-scale image deepfake datasets specifically designed for social media that leverage the latest generative methods. 2) Limited Comprehensiveness. Existing datasets are typically suited either for deepfake detection or for tampered region localization [21-23, 73], focusing on specific types of generative methods or image manipulations. However, an ideal deep-fake dataset should encompass a wide range of scenarios to reflect the complexity of real social media content, where fake images may be fully generated or manipulated through image editing strategies [75]. Furthermore, most existing datasets primarily focus on binary real/fake classification or tampered region localization, with limited emphasis on explaining the cues that models use to make these decisions. + +To address these challenges, we introduce the Social media Image Detection dataSet (SID-Set), which consists of 300K images (i.e., 100K real, 100K synthetic, and 100K tampered images), providing a comprehensive resource for the deepfake detection community. Additionally, we include textual descriptions explaining the model's judgment basis. As shown in Figures 2 and 4, synthetic and tampered images are indistinguishable to humans. In particular, challenges for the SID-Set include: 1) subtle alterations of just dozens of pixels; 2) natural-looking local manipulations; 3) complex scenes in datasets. To our knowledge, SID-Set is the first dataset of its scale with extensive annotations, making it the largest and most comprehensive dataset for social media deepfake detection to date. Compared to existing datasets in Table 1, SID-Set addresses the challenges of limited diversity and outdated generative techniques by providing a more comprehensive set of high-quality and diverse images. Accordingly, we propose a new VLMs-based deepfake detection framework, named the Social media Image Detection, localization, and explanation Assistant (SIDA), which achieves the state-of-the-art (SOTA) performance on SID-Set and generalizes effectively across other benchmarks. SIDA can serve as a baseline model on SID-Set, offering a new framework for tackling social media image deepfake detection and localization. + +The main contributions of this paper are as follows: + +- We establish SID-Set, a comprehensive benchmark for detecting, localizing, and explaining deepfakes in social media images, featuring multiple image types and extensive annotations. SID-Set holds the potential for advancing the field of deepfake detection and ensuring + +Table 1. Comparison with existing image deepfake datasets. + +
DatasetContentData SourceGenerator YearMulticlasslesMasksExplanation
OHImg [39]OverheadGoogle Map2023✘✘✘
FakeSpotter [63]FaceCelebA, FFHQ2020✘✓✘
ForgeryNet [24]FaceCREMA-D2021✘✓✘
DCFace [27]FaceFFHQ2023✘✘✘
DFF [58]FaceIMDB-WIKI2023✘✘✘
RealFaces [48]FacePrompts2023✘✘✘
M3Dsynth [87]BiologyDDPM, CycleGAN2023✘✓✘
CNNSpot [64]ObjectSDXL2020✓✘✘
CIFAKE [3]ObjectCIFAR2023✓✘✘
CASIA 2.0 [12]GeneralCorel2022✘✓✘
Artifact [53]GeneralCOCO, FFHQ, LSUN2023✘✘✘
IMD2020 [45]GeneralPlaces22020✘✓✘
AIGCD [84]GeneralLSUN, COCO, FFHQ2023✘✘✘
GenImage [86]GeneralImageNet2023✘✘✘
SID-SetGeneralCOCO, Flickr30k, MagicBrush2024✓✓✓
+ +robust model performance in complex real-world scenarios. + +- We propose SIDA, a new image deepfake detection, localization, and explanation framework that not only detects images with high accuracy but also localizes and explains potential manipulations, enhancing the transparency and utility of deepfake detection technologies. +- Extensive experiments demonstrate that SIDA effectively identifies and delineates tampered areas within images, supporting the development of more robust and interpretable deepfake detection systems. Notably, SIDA demonstrates superior or equivalent performance on the SID-Set and other benchmarks. + +# 2. Related Work + +# 2.1. Image Deepfake Datasets + +In the realm of deepfake detection, the primary focus has historically centered on the identification of facial deepfakes. Renowned datasets such as ForgeryNet [24], DeepFakeFace [58], and DFFD [7] have been pivotal in this area. As the field evolves, there is a growing shift among researchers towards exploring non-facial deepfake data. Advanced methodologies involving text-to-image or imaged-to-image generation techniques [76], utilizing GANs or the stable diffusion series, have facilitated the creation of expansive deepfake datasets like GenImage [86], HiFi-IFDL [21], and DiffForensics [66]. These datasets are characterized by their enlarged data volumes, diversified generation methodologies, and enriched annotation details. Furthermore, beyond the conventional real/fake annotations, certain datasets [21, 22, 73] now include more granular annotations. Table 1 delineates a detailed comparison among various deepfake datasets, highlighting key differences in generation scenarios and annotation types supported. It shows that SID-Set is particularly tailored towards social media data, incorporating the latest SOTA generation models, emphasizing high-quality production, and providing much more comprehensive and diverse annotations. + +# 2.2. Image Deepfake Detection and Localization + +Deepfake detection methods [4, 6, 38, 49, 62] are typically approached as classification tasks within the data-driven + +![](images/900e06e414fc30c68fb51eafa1819f86407d0770c2ebab240b89e4d32f372746.jpg) +Figure 2. SID-Set examples. The 1st row is the synthetic images, while the 2nd row shows tampered images. (Zoom in to view) + +paradigm. These strategies primarily leverage diverse architectures [38, 49], including Convolutional Neural Networks (CNNs) and Transformers, to detect distinctive artifacts. Some scholars have attempted to achieve relatively high precision and generality by employing strategies such as employing techniques such as data augmentation [62], adversarial training [6], reconstruction [4], etc. On the other hand, some researchers [25, 60] have explored extracting features from the frequency domain for deepfake identification. Efforts [13, 65] have also been made to fuse features from both spatial and frequency domains to obtain a more comprehensive set of discriminative features for deepfake detection. Although these methods have shown some progress, they still struggle with issues of generalization. Furthermore, some scholars [20, 40, 44, 61, 80, 83] have gone beyond the basic classification between real and fake by gradually constructing datasets annotated with masks of locally tampered areas, thereby addressing both image deepfake detection and localization tasks. However, these datasets are concentrated mainly on facial data, with fewer datasets available for non-facial deepfake detection and localization, and even fewer large and public datasets for realistic social media data. As a result, our work aims to address these critical gaps by providing a new, comprehensive dataset that includes diverse manipulations beyond facial data, particularly focusing on social media images. + +# 2.3. Large Multimodal Models + +The progress in large language models (LLMs) [14-16, 41] and vision-language models (VLMs) [14-16, 29, 71, 78] has notably improved multimodal comprehension, seamlessly integrating visual and textual data. The LLaMA series [14-16] optimize language understanding with a compact, high-performance design using fewer parameters than prior models. LLaVA series [34, 35] enhance visual question answering by synchronizing visual features with textual data. Additionally, LISA series [29, 71] employ LLM for accurate image segmentation, merging visual perception with linguistic insights to precisely segment the targeted areas. Several grounding large multimodal models [50, 54, 55, 67, 68, 72, 74, 77, 81] have been proposed to localize the contents based on linguistic information. + +Advancements [5, 10, 19, 69, 82] in integrating multi + +modal data such as visual and linguistic information have also significantly improved the performance of models in detecting and pinpointing deepfakes. For instance, AntifakePrompt [5] approaches deepfake detection by formulating it as a visual question answering problem, which adjusts soft prompts for InstructBLIP [10], enabling it to determine whether a query image is real or fake. ForgeryGPT [31] enhances image forgery detection and localization by integrating advanced forensic knowledge with a mask-aware forgery extractor that targets pixel-level fraud detection. As a concurrent work, FakeShield [69] leverages the capabilities of LLaVA to identify and localize altered regions while providing interpretable insights into the findings. Our method diverges significantly from existing approaches by creating the most extensive deepfake dataset tailored specifically for social media images with comprehensive annotations. Additionally, we also set a new standard for social media image deepfake detection by utilizing the visual interpretation strengths of VLMs to boost detection accuracy, pinpoint forgeries, and provide clearer explanations within a comprehensive framework. + +# 3. Benchmark + +# 3.1. Motivation + +In the realm of deepfake detection and localization, research has predominantly focused on facial deepfakes due to their substantial societal impact. However, with the advancements in generative technology, the scope of deepfakes has extended beyond facial content to include non-facial manipulations. Historically, non-facial deepfakes were less prevalent, largely limited by technological constraints that produced low-quality, easily detectable forgeries. Although some datasets, such as GenImage [86] and AIGCD [84], have been constructed, they suffer from several limitations: 1) They often utilize relatively outdated generative technologies, resulting in lower quality data easily distinguished by humans. 2) They primarily focus on text-to-image or image-to-image generation, neglecting the need for data involving manipulations of specific regions, objects, or parts. Such tampered manipulations can be especially insidious as they introduce subtle misinformation, making existing SOTA deepfake detection methods less effective. 3) They lack well-defined criteria for content authenticity, limiting their effectiveness in providing interpretative insights and educating the public on distinguishing synthetic content. + +Addressing these limitations is crucial for improving the transparency and utility of deepfake detection systems. Furthermore, most existing datasets emphasize either fully synthetic or tampered images, as shown in Table 1, but in real applications, we don't know the image deepfake type in advance. Effective deepfake detection and localization methods should be capable of addressing both scenarios, as + +![](images/d7d4e80538d0cedc06c6ff327ed89b1c9acbdf47bbcd410eed18cad201fc0b3a.jpg) +Figure 3. Tampered image generation pipeline: It consists of four stages—extracting objects from captions using GPT-4o, obtaining object masks with Language-SAM, setting up replacement dictionaries for generating tampered images, and generating new images using Latent Diffusion. This figure illustrates an example of object replacement (e.g., “cat” to “dog”) and attribute modification. + +![](images/8344ea02b8ed04918e727bb17423363240ffa9770bdf7037f949469b4fe2bf08.jpg) + +![](images/0516d556f88e5dce042f78bc0252f097d6fd95bc81476383774671076022833a.jpg) + +![](images/0671702a1689addff8324475dc29763e71b7ab4ac300c689fed73c51c0cc48d4.jpg) + +social media images often involve complex combinations of synthetic and tampered content. Therefore, developing a comprehensive benchmark for detecting and localizing deepfakes in social media images is essential. We propose SID-Set, which encompasses comprehensive, high-quality annotations for detection and localization, along with detailed textual explanations of the judgment criteria. + +# 3.2. Benchmark Construction + +Data Details. To develop an effective benchmark for detecting and localizing images on social media, we created SID-Set, a dataset with real, synthetic, and tampered images reflecting diverse real-world scenarios. Our benchmark assesses whether models can differentiate among real, synthetic, and tampered images, as well as accurately identify altered regions in tampered images. + +Real Images: 100K images from OpenImages $\mathrm{V}7^{1}$ , with a wide range of scenarios reflecting real-world diversity. + +Synthetic Images: 100K images generated through FLUX [43], specifically designed to challenge identification due to their high-quality, highly realistic appearance. + +Tampered Images: 100K tampered images, with specific objects or regions replaced or altered; the detailed generation process is shown in Figure 4. + +Data Generation. To generate highly realistic synthetic images, we experimented with several open-source SOTA generative models, such as FLUX [43], Kandinsky 3.0 [1], SDXL [52], AbsoluteReality [42], and others. Following a review by five human experts of 1,000 images from each generative model, FLUX emerged as the top performer, producing highly convincing images that were indistinguishable from real ones to human experts. Consequently, we employed FLUX to create 100K synthetic images based on + +![](images/de1d4cf5dd0cc21a7e9e57da614f084929da6b2922307435435695494833426b.jpg) +bird + +![](images/48ca8d52b993b4983b20c04513176cc78bea2e92700bd98cc2c3195281b72154.jpg) + +![](images/26044953c5c9b46fa493d87a54404aedd5ddf2de562b40abba2e99558414eda2.jpg) +eagle + +![](images/4095ae0660da8175d8a17be01a0d65d498f0c9a885816e8eba455e5b41d68b5b.jpg) +cat + +![](images/a49b555a043a0427611175aa13d4d3221685b254e3b1cf58c6317979564ac29e.jpg) +dog + +![](images/96e032ec2989046286f6412203fee5791be076ddb2334d39659de7a84bbdca30.jpg) +bodyboard + +![](images/49f9a0dbdf350f25cd71c1376b5bcb405809ef7defae0f6c092eb534b022d5c3.jpg) + +![](images/db7f9096a2b1372de8d9d344772e2f39c0c3ddbced67d20ca7d1e02339b80ca3.jpg) +coffee table +Figure 4. Examples of tampered images. (Zoom in to view) + +![](images/09257dd71993a5510489016414d1f4cc3b8aca492b0f3ab46b1ba410289d1c92.jpg) +bodyboard +coffee table + +the Flickr30k [51] and COCO [33]. The image tampering process depicted in Figure 4 follows four distinct stages, utilizing the COCO image as an example. + +Stage 1: We extract objects from an image's caption using GPT-4o [47]. For instance, from the caption "A large fluffy cat laying on top of a wooden table", GPT-4o identifies relevant COCO class objects or retains nouns if no match exists. This extraction is documented in an "Image-Caption-Object" JSON file. + +Stage 2: Employing Language-SAM [30], we generate masks for identified objects as training ground truth. + +Stage 3: We establish dictionaries for full and partial image tampering using COCO classes for object replacement and attribute modifications, respectively. For example, replacing "dog" with animals like "cat" or adding attributes such as "happy" or "angry" to the "dog" class. For more details, please refer to the Appendix. + +Stage 4: Utilizing Latent Diffusion [56], we modify captions and regenerate images, either replacing or retaining original objects based on availability. An example modification is altering "cat" to "dog" in the image caption. + +In total, we generated 80,000 object-tampered images and 20,000 partially tampered images. To demonstrate the explainability of SIDA, we used GPT-4o to generate textual descriptions for the judgment basis of 3,000 images from + +![](images/42197889730a533eac850997b77f30edbf2e7d1e393280ceebb589a8e13b85bd.jpg) +Figure 5. The pipeline of SIDA: Given an image $x_{i}$ and the corresponding text input $x_{t}$ , the last hidden layer for the token provides the detection result. If the detection result indicates a tampered image, SIDA extracts the token to generate masks for the tampered regions. This figure shows an example where the man's face has been manipulated. + +the SID-Set, divided equally among three types. Additionally, to ensure the realism of synthetic images, tampered images, and their textual descriptions, we engaged 5 annotation experts for quality control and adjustments. Detailed prompts and generation pipelines used to create descriptions for each type of image are provided in the Appendix. + +# 4. Method + +In this section, we first present the model architecture of SIDA in Section 4.1, followed by an introduction to the training process of our method in Section 4.2. + +# 4.1. Architecture + +Large vision-language models have demonstrated remarkable capabilities in understanding the alignment between textual and visual information. For instance, LLaVA [34] leverages language alone to achieve a comprehensive understanding of both visual and linguistic content. Building on LLaVA, LISA [29] extends this capability by providing fine-grained segmentation masks along with corresponding textual descriptions. However, to effectively detect and localize synthetic images, VLMs must not only be capable of multimodal understanding but also possess the ability to identify and segment manipulated regions, providing detailed explanations for both synthetic and tampered images. + +To this end, we propose SIDA to tackle the task of synthetic image detection and tampered region localization. The pipeline of our method is illustrated in Figure 5. Inspired by previous approaches [28, 29, 85], we expand the + +original vocabulary of VLMs by adding two new tokens, $<\mathrm{DET}>$ and $<\mathrm{SEG}>$ , to enable the model to extract detection and segmentation information. Given an image $x_{i}$ and a text prompt $x_{t}$ , such as "Can you identify if this image is real, fully synthetic, or tampered? Please mask the tampered object/part if it is tampered." We feed them into the VLM. The VLM then outputs a text description $\hat{y}_{\mathrm{des}}$ , while the last hidden layer $h_{\mathrm{hid}}$ contains the $<\mathrm{DET}>$ and $<\mathrm{SEG}>$ tokens. This process can be formulated as follows: + +$$ +\hat {y} _ {\mathrm {d e s}} = \operatorname {V L M} \left(x _ {i}, x _ {t}\right). \tag {1} +$$ + +Next, we extract the $<\mathrm{DET}>$ token from the last hidden layer $h_{\mathrm{hid}}$ to obtain $h_{\mathrm{det}}$ . The representation $h_{\mathrm{det}}$ is then passed through a detection head $F_{\mathrm{det}}$ to determine whether the image is real, fully synthetic, or tampered. We denote the final detection result by $\hat{\mathsf{D}}$ : + +$$ +\hat {\mathrm {D}} = F _ {\text {d e t}} \left(h _ {\text {d e t}}\right), \tag {2} +$$ + +$F_{\mathrm{det}}$ is the detection head that processes the extracted $\langle \mathrm{DET}\rangle$ representation $h_\mathrm{det}$ to produce the detection output. + +If the detection result indicates that the image has been tampered with, SIDA will then predict a mask for the tampered regions. The $h_{\mathrm{seg}}$ feature is extracted from the hidden layer $h_{\mathrm{hid}}$ , similar to the extraction process for the token. Given that the token encapsulates crucial information that can aid in generating fine-grained segmentation masks, the representation $h_{\mathrm{det}}$ is transformed using a fully connected layer $F$ to align with the dimensions of $h_{\mathrm{seg}}$ . To further capture the relationship between the $h_{\mathrm{det}}$ and $h_{\mathrm{seg}}$ + +features, we apply a single-layer Multihead Attention, facilitating effective feature interaction and enhancing mask quality. The process can be formulated as follows: + +$$ +\tilde {h} _ {\mathrm {d e t}} = F (h _ {\mathrm {d e t}}), +$$ + +$$ +\tilde {h} _ {\text {s e g}} = \operatorname {M S A} \left(\tilde {h} _ {\det }, h _ {\text {s e g}}\right), \tag {3} +$$ + +$$ +\tilde {h} _ {\mathrm {s e g}} = \tilde {h} _ {\mathrm {s e g}} + h _ {\mathrm {s e g}}. +$$ + +In this formulation, we treat the detection features as the query, and the segmentation features as the key and value. A residual connection is employed to combine both the original and transformed information, thereby enhancing the representation for precise segmentation. + +Finally, we employ a frozen image encoder $F_{\mathrm{enc}}$ to extract visual features from the input image $x_{i}$ , resulting in visual features $f$ . The segmentation embedding $\tilde{h}_{\mathrm{seg}}$ is then combined with the visual features $f$ and fed into a decoder to produce the final mask $\hat{M}$ . This can be formulated as: + +$$ +f = F _ {\mathrm {e n c}} (x _ {i}), +$$ + +$$ +\hat {\mathrm {M}} = F _ {\mathrm {d e c}} \left(\tilde {h} _ {\mathrm {s e g}}, f\right). \tag {4} +$$ + +# 4.2. Training + +Training Objectives. The training loss, $\mathcal{L}$ , for the SIDA consists of three components: the detection loss $\mathcal{L}_{det}$ , the text generation loss $\mathcal{L}_{txt}$ , and the segmentation mask loss $\mathcal{L}_{mask}$ . Initially, SIDA is trained in an end-to-end manner by employing the detection loss and the segmentation loss. For detection, we use CrossEntropy loss, while for the segmentation task, we use a weighted combination of binary cross-entropy (BCE) and DICE loss, with respective loss weights $\lambda_{\mathrm{bce}}$ and $\lambda_{\mathrm{dice}}$ . This can be formulated as: + +$$ +\mathcal {L} = \lambda_ {\text {d e t}} \mathcal {L} _ {\text {d e t}} + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}}, +$$ + +$$ +\mathcal {L} _ {d e t} = \mathcal {L} _ {\mathrm {C E}} (\hat {\mathrm {D}}, \mathrm {D}), \tag {5} +$$ + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \mathcal {L} _ {B C E} (\hat {\mathrm {M}}, \mathrm {M}) + \lambda_ {d i c e} \mathcal {L} _ {D I C E} (\hat {\mathrm {M}}, \mathrm {M}). +$$ + +After completing the training phase, we proceed to fine-tune the SIDA model by utilizing detailed textural descriptions from 3,000 images as the ground truth, represented by $y_{des}$ . This phase focuses on optimizing the text generation component, $\mathcal{L}_{txt}$ , to improve its ability in textural interpretability. The final loss function is as follows: + +$$ +\mathcal {L} _ {t x t} = \mathcal {L} _ {\mathrm {C E}} (y _ {\hat {d e s}}, y _ {d e s}), +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {d e t}} \mathcal {L} _ {\text {d e t}} + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\text {t x t}} \mathcal {L} _ {\text {t x t}}, \tag {6} +$$ + +where $\lambda_{det}$ , $\lambda_{mask}$ , and $\lambda_{txt}$ are the weighting factors that balance the contributions of the detection, segmentation, and text generation losses, respectively. + +Training Data. We utilize the SID-Set, consisting of 300k images, to train SIDA. To further enhance diversity, we incorporate the MagicBrush dataset [79] after filtering out + +low-quality images. The combined dataset supports robust training for both the detection and localization of synthetic content. Additionally, we generate descriptions for 3,000 randomly selected images using LLMs. + +# 5. Experiments + +Implementation Details. We choose LISA as the base large vision language model due to its strong capability for reasoning-based localization. We fine-tuned both LISA-7B-v1 and LISA-13B-v1 on the SID-Set using LoRA, setting $\alpha$ at 16 and the dropout rate at 0.05. The input images are resized to $1024 \times 1024$ . The loss weights in Eq. (6) for the detection $(\lambda_{\mathrm{det}})$ , the text generation $(\lambda_{\mathrm{txt}})$ , and the localization $(\lambda_{\mathrm{mask}})$ are set to 1.0, respectively. The localization loss weights in Eq. (5) for the $\lambda_{\mathrm{bce}}$ and $\lambda_{\mathrm{dice}}$ are set to 2.0 and 0.5, respectively. To determine the optimal weight configuration, we perform ablation studies as detailed in Section 5.5. During the detection and localization training stage, the image encoder is frozen, and all other modules are trainable. For the text generation stage, only vision-language models are fine-tuned using the LoRA strategy. The initial learning rate is set to $1 \times 10^{-4}$ , with a batch size of 2 per device and a gradient accumulation step of 10. We use two NVIDIA A100 GPUs (40GB each). Training for SIDA-7B and SIDA-13B took 48 hours and 72 hours, respectively. + +Evaluation Metrics. We evaluate detection using image-level accuracy and F1 scores. For forgery localization, our metrics include Area Under the Curve (AUC), F1 scores, and Intersection over Union (IoU). + +# 5.1. Detection Evaluation + +We compare SIDA against other SOTA deepfake detection methods on SID-Set, including CnnSpot [17], AntifakePrompt [5], FreDect [18], Fusing [26], Gram-Net [37], UniFDP [46], LGrad [59], and LNP [2]. To ensure a fair comparison, we first evaluate these models on our dataset using their original pre-trained weights, then retrain them with the SID-Set to assess performance improvements. Table 2 demonstrates that SIDA achieves better or comparable results among all the evaluated methods. Notably, LGrad [59] achieves the highest accuracy and F1 score on tampered images after retraining, but this comes at the expense of lower performance in other metrics. Our analysis indicates that LGrad's high recall and false positive rates stem from its propensity to misclassify other types as tampered. The training details are provided in the Appendix. + +# 5.2. Localization Results + +Table 3 presents the forgery localization performance on the SID-Set. We selected PSCC-Net [36], MVSS-Net [11], and HIFI-Net [21] as representative IFDL methods. Additionally, we chose LISA [29] as a representative LLM due to + +Table 2. Comparison of SIDA with other deepfake detection methods. Values outside parentheses are evaluated with open-source models directly on the SID-Set; values in parentheses indicate performance changes after training with the SID-Set. The best results are in bold. + +
MethodsYearRealFully syntheticTamperedOverall
AccF1AccF1AccF1AccF1
AntifakePrompt [5]202464.8(↑24.1)78.6(↑10.5)93.8(↑3.7)96.8(↑1.1)30.8(↑60.1)47.2(↑33.2)63.1(↑29.1)69.3(↑23.4)
CnnSpott [17]202179.8(↑9.2)88.7(↑2.1)39.5(↑51.2)56.6(↑31.5)6.9(↑61.2)12.9(↑51.1)42.1(↑39.3)69.6(↑20.7)
FreDect [18]202083.7(↓37.7)91.1(↓43.5)16.8(↑44.1)28.8(↑37.2)11.9(↑25.2)21.3(↑31.7)37.4(↑33.6)23.4(↑47.2)
Fusing [26]202285.1(↑4.1)92.0(↑0.7)34.0(↑54.1)50.7(↑38.4)2.7(↑24.3)5.3(↑26.1)40.1(↑33.1)29.1(↑40.3)
Gram-Net [37]202070.1(↑19.1)82.4(↑9.3)93.5(↑4.4)96.6(↑2.0)0.8(↑89.1)1.6(↑85.3)55.0(↑37.1)58.0(↑35.1)
UnivFD [46]202368.0(↑0.3)67.4(↑1.1)62.1(↑24.3)87.5(↑10.5)64.0(↑28.5)85.3(↑4.7)64.0(↑21.7)85.3(↑4.5)
LGrad [59]202364.8(↓2.8)78.6(↓2.5)83.5(↓25.5)91.0(↓23.7)6.8(↑92.3)12.7(↑86.1)51.8(↑20.2)55.5(↑23.9)
LNP [2]202371.2(↓56.8)83.2(↓60.2)91.8(↓55.6)95.7(↓60.1)2.9(↑90.4)5.7(↑88.9)55.2(↓7.6)58.2(↑4.1)
SIDA-7B202489.191.098.798.691.291.093.593.5
SIDA-13B202489.691.198.598.792.991.293.693.5
+ +its segmentation reasoning capabilities. We used LISA-7Bv1 and fine-tuned it on SID-Set. The results indicate that SIDA achieves the best performance. We suppose that while LISA possesses strong general segmentation capabilities, it lacks the specific specialized features required to detect subtle manipulations, ultimately limiting the effectiveness of fine-tuning for precise forgery localization. + +Table 3. Comparison between SIDA and other IFDL approaches. * indicates the use of the pre-trained model from the original paper due to unavailable training code. + +
MethodsYearsTampered
AUCF1IOU
MVSS-Net* [11]202348.931.623.7
HIFI-Net* [21]202364.045.921.1
PSCC-Net [36]202282.171.335.7
LISA-7B-v1 [29]202478.469.132.5
SIDA-7B202487.373.943.8
+ +# 5.3. Robustness Study + +We further evaluate the robustness of SIDA against common image perturbations found in social media, such as JPEG compression, resizing, and Gaussian noise. Table 4 shows our model's performance on the SID-Set under six degradation scenarios: JPEG compression (with quality levels of 70 and 80), resizing (with scaling factors of 0.5 and 0.75), and Gaussian noise (with variances of 5 and 10). Despite not being explicitly trained on degraded data, SIDA demonstrates resilience to these low-level distortions. The model's stable performance against common social media perturbations highlights its robustness and practical applicability. + +# 5.4. Test on Other Benchmark + +In this stage, we evaluate SIDA on DMImage [8] dataset to assess its generalization capabilities. We compare SIDA + +Table 4. Performance of SIDA under different perturbations. + +
DetectionLocalization
ACCF1AUCF1IOU
JPEG 7089.490.186.271.842.3
JPEG 8088.789.585.871.141.7
Resize 0.589.391.186.872.543.2
Resize 0.7589.991.687.173.043.5
Gaussian 1086.989.384.170.241.0
Gaussian 588.489.985.371.041.5
SIDA-7B93.593.587.373.943.8
+ +with CNNSpot [17], Fusing [26], Gram-Net [37], LNP [2], UnivFD [46], and AntifakePrompt [5]. For these methods, we use the original hyperparameter settings and pre-trained weights provided by the authors. The results in Table 5 show that SIDA achieves superior performance, demonstrating its strong adaptability. + +Table 5. Comparison with other deepfake detection methods on DMImage [8]. We used original weights for each method. + +
MethodsRealFakeOverall
AccF1AccF1AccF1
CNNSpot [17]87.888.428.444.240.643.3
Gram-Net [37]62.854.178.888.167.479.4
Fusing [26]87.786.115.527.240.436.5
LNP [2]63.167.456.972.558.268.3
UnivFD [46]89.488.344.961.253.960.7
AntifakePrompt [5]91.392.589.391.290.691.2
SIDA-7B92.993.190.791.091.892.4
+ +# 5.5. Ablation Study + +Attention Module. We conducted ablation experiments to assess the importance of the attention module. Variants included removing the attention module entirely and replacing it with fully connected (FC) layers. Results in Table 6 show that removing the attention module or replacing + +Table 6. Ablation study results for attention module in SIDA. + +
DetectionLocalization
ACCF1AUCF1IOU
FC91.190.384.371.638.9
w/o Attention90.389.984.171.338.8
SIDA93.593.587.373.943.8
+ +it with FC significantly reduces performance, underscoring the critical role of attention in enhancing feature interaction and improving detection and localization accuracy. + +Training Weights. SIDA training utilizes weighted losses to balance task contributions. In the detection and localization stages, detection loss is adjusted by a weight $\lambda_{\mathrm{det}}$ , and localization loss by binary cross-entropy (BCE) and DICE losses, with weights $\lambda_{\mathrm{bce}}$ and $\lambda_{\mathrm{dice}}$ respectively. For our experiments, we set $\lambda_{\mathrm{det}}$ to 1, $\lambda_{\mathrm{bce}}$ to 2.0, and $\lambda_{\mathrm{dice}}$ to 0.5 to maintain balance between detection and localization, enhancing model stability and performance. We summarize the outcomes of various weight configurations in Table 7. + +Table 7. Different weight configurations in SIDA training. + +
λdetλbceλdiceAccF1 Score
1.02.00.593.5691.01
1.04.01.093.4990.86
+ +# 5.6. Qualitative Results + +In this section, we present examples of SIDA's output for tampered images, showcasing its detection, localization, and explanation capabilities. SIDA accurately identifies tampered regions and provides explanations for its decisions. We also include some challenging failure cases where SIDA was unable to accurately detect the tampered regions, highlighting areas for future improvement. Additional visualization results are available in the Appendix. + +# 6. Conclusion and Discussions + +In this work, we present SID-Set for social image deepfake detection, localization, and explanation tasks, consisting of 100k real images, 100k fully synthetic images, and 100k tampered images. Furthermore, we propose a new VLMs-based deepfake detection framework, SIDA, to address these tasks. SIDA demonstrates its ability to detect fake types, localize tampered regions, and provide explanations for its decisions. We believe that the integration of VLMs into deepfake detection tasks, as demonstrated by SIDA, offers promising new avenues for future research. + +Although the development of the SID-Set and the introduction of the SIDA framework have yielded favorable outcomes in deepfake detection tasks, we recognize some potential limitations and will optimize them in future research. Dataset Size. While SID-Set includes 100k fully synthetic and 100k tampered images, the complexity of real social + +Can you identify if this image is real, full synthetic, or tampered? Please mask the tampered object/part if it is tampered. + +![](images/cea45cf8fea55b4926330f894c18d05f6deb08fff83a21fd7e5076479c50d5ef.jpg) + +![](images/c200b6ce03b3931f3d454cab54e2684f47613a32ea7ecb693f80664fd17f538d.jpg) + +![](images/5378186f9d0b3ca33adfc77b147ccfa0659099504b64c2a085bb99be9287f1c4.jpg) + +![](images/079219d1ce457548ba22ecf335969b56e524b9638d61fcaf3b92efeb35f77ca0.jpg) + +![](images/ce7e752e97820e7a8226687c15abd7008aceaf0b6bd19c156788361551a45a9c.jpg) + +# Object ampered + +![](images/bb062c8e102d27df3f642a0a9898045943daa8236db0862d47fb0aafada622e4.jpg) + +# Partial tampered + +![](images/e319df19ca84f1e3b75245fd10a4f9810d97beb316ed10ad146bd76cc6d2a7f6.jpg) + +# Partial tampered + +![](images/62013276a45753571f0ff721eabd0b1b48e45e5abfbc92bb27c42901d3dc32ac.jpg) + +# Object tampered + +![](images/c2429f968347ef108e363714fc4ae8c9e58a74641c8ceef03c045d5a6284092b.jpg) + +Type of Image: object tampered The tampered areas are specifically located in the h The tampered. Absolute Positino The location is + +Type of Image: part tampered Tampering Type and Location... Relative Position: The tampered part located in the face of the women... Absolute Position... + +Type of Image: object +tampered Tampering +Type and Location is the face of the man... +Absolute Position +Tampered Content: Modification... +Inconsistencies: Lighting... + +Type of Image: part tampered Tampered Content: Modification... Relative Position: Inconsistencies... Colour in the background looks ... + +![](images/11cdcee45f455b539c81086e3fcd9513808fc335d8ccea4096713c27a0c98e59.jpg) + +# a) Correct examples + +![](images/14d6f60b7b0334a52063389610d2bea7f39c4c3b83e067466feb6f3c10cf8e27.jpg) + +![](images/eb7b756972ab1bf586fc87727112f928b3ae74a50723f70cdf684013e6c7afc1.jpg) + +![](images/c2fe3771f02eab9bbb4d1c8c29db5f187e33548a73c30bd7e7ce0133cfda18a9.jpg) + +![](images/f1a5e589bf67da6b38cec5820377fc7b1981fcf416524b65def7ad4e2ba1e3d5.jpg) + +![](images/512d5490c30048e1b356504f7698c6ba0cd23d4f5ef762b7f836ba62bd9771aa.jpg) + +![](images/3e87ad72ff1ac4415100190a6a8af837624093e789d0f5cf7c3696f8f9a1bede.jpg) + +![](images/0140ded12571ec68e5754f4aee42949c691cc306aa0071d125340cac24d60006.jpg) + +![](images/e600bd8b9b358bc9f427a2fc421aa5ef57d09b29234c191ab40f76c238401334.jpg) + +![](images/36ab3eb1f014cf103ac1139689ef532fbdead74790683bd375209dd21c2fbe90.jpg) +b) Failure examples + +![](images/ad4bcb55a045272f48a59cacd607cd3f77c9142b8494f6f864dbff97e04911c2.jpg) +Figure 6. Visual results of SIDA on tampered images. + +![](images/9cc53ac93d82242eb6072fe6b0402e0cdb5f7dfcc25c56274a934a5e4e8464da.jpg) + +![](images/6a4d2a53420c420336ae94a1141e2184c8d71d248cdb2cd33a130a3d20594551.jpg) + +media environments demands a larger dataset. Therefore, expanding the dataset with additional images is a crucial objective for future research. + +Data Domain. Other methods often generate social media images that lack authenticity or are easily identifiable. Consequently, we used only FLUX to generate fake images due to its superior quality. However, depending exclusively on one method may lead to issues of data skew, although this was not significantly evident in our experiments. Such skew could potentially impair performance on diverse datasets. Moving forward, we plan to explore additional generative methods and integrate various other generation techniques to produce a more varied and higher-quality set of images. + +Localization Results. Although SIDA demonstrates relatively strong performance on our dataset, there is still room for improvement. Certain tampered regions are not reliably detected, underscoring the need for further advancements. + +# Acknowledgments + +This work is supported by The Alan Turing Institute (UK) through the project 'Turing-DSO Labs Singapore Collaboration' (SDCfP2\100009). + +# References + +[1] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. Arxiv, 2023. 4 +[2] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. Detecting generated images by real images only. Arxiv, 2023. 6, 7 +[3] Jordan J. Bird and Ahmad Lotfi. CIFAKE: image classification and explainable identification of ai-generated synthetic images. IEEE Access, 2024. 2 +[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In CVPR, 2022. 2, 3 +[5] You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors. Arxiv, 2023. 3, 6, 7 +[6] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 2, 3 +[7] Harry Cheng, Yangyang Guo, Tianyi Wang, Liqiang Nie, and Mohan S. Kankanhalli. Diffusion facial forgery detection. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 1, 2 +[8] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, 2023. 7 +[9] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 2023. 1 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 3 +[11] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE TPAMI, 2023. 6, 7 +[12] Jing Dong, Wei Wang, and Tieniu Tan. CASIA image tampering detection evaluation database. In ISIP, 2013. 2 +[13] Junxian Duan, Yang Ai, Jipeng Liu, Shenyuan Huang, Huaibo Huang, Jie Cao, and Ran He. Test-time forgery detection with spatial-frequency prompt learning. IJCV, 2024. 3 +[14] Abhimanyu Dubey et al. The llama 3 herd of models. Arxiv, 2024. 3 + +[15] Hugo Touvron et al. Llama: Open and efficient foundation language models. Arxiv, 2023. +[16] Hugo Touvron et al. Llama 2: Open foundation and finetuned chat models. Arxiv, 2023. 3 +[17] Joel Frank and Thorsten Holz. Cnn-generated images are surprisingly easy to spot...for now. Arxiv, 2021. 6, 7 +[18] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 6, 7 +[19] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In AAAI, 2024. 3 +[20] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In CVPR, 2023. 3 +[21] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 1, 2, 6, 7 +[22] Xiao Guo, Xiaohong Liu, Iacopo Masi, and Xiaoming Liu. Language-guided hierarchical fine-grained image forgery detection and localization. Arxiv, 2024. 1, 2 +[23] Ruidong Han, Xiaofeng Wang, Ningning Bai, Yihang Wang, Jianpeng Hou, and Jianru Xue. HDF-net: Capturing homogeneity difference features to localize the tampered image. IEEE TPAMI, 2024. 2 +[24] Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. Forgerynet: A versatile benchmark for comprehensive forgery analysis. In CVPR, 2021. 1, 2 +[25] Yonghyun Jeong, Doyeon Kim, Youngmin Ro, and Jongwon Choi. Frepgan: Robust deepfake detection using frequency-level perturbations. In AAAI, 2022. 3 +[26] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In ICIP, 2022. 6, 7 +[27] Minchul Kim, Feng Liu, Anil K. Jain, and Xiaoming Liu. Dcface: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 2 +[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021. 5 +[29] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: reasoning segmentation via large language model. In CVPR, 2024. 3, 5, 6, 7 +[30] lang-sam team. Lang-segment-anything. https://github.com/luca-medeiros/lang-segment-anything, 2024. Accessed: 2024-11-15. 4 +[31] Jiawei Li, Fanrui Zhang, Jiaying Zhu, Esther Sun, Qiang Zhang, and Zheng-Jun Zha. Forgerygpt: Multimodal large language model for explainable image forgery detection and localization. Arxiv, 2024. 3 + +[32] Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. Detecting multimedia generated by large AI models: A survey. Arxiv, 2024. 1 +[33] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 5 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 3 +[36] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for image manipulation detection and localization. IEEE TCSVT, 2022, 6, 7 +[37] Zhengzhe Liu, Xiaojuan Qi, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. In CVPR, 2020. 6, 7 +[38] Momina Masood, Marriam Nawaz, Khalid Mahmood Malik, Ali Javed, Ain Irtaza, and Hafiz Malik. Deepfakes generation and detection: state-of-the-art, open challenges, countermeasures, and way forward. Appl. Intell., 2023. 2, 3 +[39] Brandon B. May, Kirill Trapeznikov, Shengbang Fang, and Matthew C. Stamm. Comprehensive dataset of synthetic and manipulated overhead imagery for development and evaluation of forensic tools. In IH&MMSec, 2023. 2 +[40] Changtao Miao, Qi Chu, Zhentao Tan, Zhenchao Jin, Wanyi Zhuang, Yue Wu, Bin Liu, Honggang Hu, and Nenghai Yu. Multi-spectral class center network for face manipulation detection and localization. Arxiv, 2023. 3 +[41] Shervin Minaee, Tomás Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. Arxiv, 2024. 3 +[42] Absolutereality model team. Absolutereality model repository. https://huggingface.co/jochemstoel/absolutereality-model-repository, 2024. Accessed: 2024-11-15. 4 +[43] Flux model team. Flux model. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. Accessed: 2024-11-15. 4 +[44] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In CVPR, 2024. 3 +[45] Adam Novozámsky, Babak Mahdian, and Stanislav Saic. IMD2020: A large-scale annotated dataset tailored for detecting manipulated images. In WACvw, 2020. 2 +[46] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, 2023. 6, 7 +[47] OpenAI. GPT-4 technical report. Arxiv, 2023. 4 +[48] Lorenzo Papa, Lorenzo Faiella, Luca Corvitto, Luca Maiano, and Irene Amerini. On the use of stable diffusion for creating realistic faces: from generation to detection. In IWBF, 2023. 2 + +[49] Gan Pei, Jiangning Zhang, Menghan Hu, Guangtao Zhai, Chengjie Wang, Zhenyu Zhang, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey. Arxiv, 2024. 2, 3 +[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. Arxiv, 2023. 3 +[51] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 2017. 4 +[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 4 +[53] Md Awsafur Rahman, Bishmoy Paul, Najibul Haque Sarker, Zaber Ibn Abdul Hakim, and Shaikh Anowarul Fattah. Artifact: A large-scale dataset with artificial and factual images for generalizable and robust synthetic image detection. In ICIP, 2023. 2 +[54] Hanoona Abdul Rasheed, Muhammad Maaz, Sahal Shaji Mullappilly, Abdelrahman M. Shaker, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Eric P. Xing, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Glamm: Pixel grounding large multimodal model. In CVPR, 2024. 3 +[55] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In CVPR, 2024. 3 +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4 +[57] Xincheng Shuai, Henghui Ding, Xingjun Ma, Rongcheng Tu, Yu-Gang Jiang, and Dacheng Tao. A survey of multimodal-guided image editing with text-to-image diffusion models. Arxiv, 2024. 1 +[58] Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. Arxiv, 2023. 2 +[59] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In CVPR, 2023. 6, 7 +[60] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In AAAI, 2024. 3 +[61] Dragos-Constantin Tântaru, Elisabeta Oneata, and Dan Oneata. Weakly-supervised deepfake localization in diffusion-generated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 3 +[62] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, 2021. 2, 3 +[63] Run Wang, Felix Juefei-Xu, Lei Ma, Xiaofei Xie, Yihao Huang, Jian Wang, and Yang Liu. Fakespotter: A simple yet robust baseline for spotting ai-synthesized fake faces. In IJCAI, 2020. 2 + +[64] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2 +[65] Yuan Wang, Kun Yu, Chen Chen, Xiyuan Hu, and Silong Peng. Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In CVPR, 2023. 3 +[66] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In ICCV, 2023. 2 +[67] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. Arxiv, 2024. 3 +[68] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. GSVA: generalized segmentation via multimodal large language models. In CVPR, 2024. 3 +[69] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. Arxiv, 2024. 3 +[70] Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A comprehensive benchmark of deepfake detection. In NeurIPS, 2023. 1 +[71] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. *Arxiv*, 2023. 3 +[72] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In ICLR, 2024. 3 +[73] Zeqin Yu, Jiangqun Ni, Yuzhen Lin, Haoyi Deng, and Bin Li. Diffforensics: Leveraging diffusion prior to image forgery detection and localization. In CVPR, 2024. 2 +[74] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Arxiv, 2023. 3 +[75] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, Lingjie Liu, Adam Kortylewski, Christian Theobalt, and Eric P. Xing. Multimodal image synthesis and editing: The generative AI era. IEEE TPAMI, 2023. 1, 2 +[76] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion models in generative AI: A survey. Arxiv, 2023. 2 +[77] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, and Jainwei Yang. Llava-grounding: Grounded visual chat with large multimodal models. In ECCV, 2024.3 +[78] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE TPAMI, 2024. 3 +[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2023. 6 +[80] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli + +Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In ICCV, 2023. 3 +[81] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Change Loy Chen, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. In NeurIPS, 2024. 3 +[82] Yue Zhang, Ben Colman, Ali Shahriyari, and Gaurav Bharaj. Common sense reasoning for deep fake detection. *Arxiv*, 2024. 3 +[83] Yi Zhang, Changtao Miao, Man Luo, Jianshu Li, Wenzhong Deng, Weibin Yao, Zhe Li, Bingyu Hu, Weiwei Feng, Tao Gong, and Qi Chu. MFMS: learning modality-fused and modality-specific features for deepfake detection and localization tasks. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 3 +[84] Nan Zhong, Yiran Xu, Zhenxing Qian, and Xinpeng Zhang. Rich and poor texture contrast: A simple yet effective approach for ai-generated image detection. Arxiv, 2023. 2, 3 +[85] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and VQA. In AAAI, 2020. 5 +[86] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In NeurIPS, 2023. 2, 3 +[87] Giada Zingarini, Davide Cozzolino, Riccardo Corvi, Giovanni Poggi, and Luisa Verdoliva. M3DSYNTH: A dataset of medical 3d images with ai-generated local manipulations. In ICASSP, 2024. 2 \ No newline at end of file diff --git a/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/images.zip b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..180c71e26757fb4e50328c820336685b21ea75a6 --- /dev/null +++ b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48c3b885847da048242f67dd2b5fa86fbcdb5de38a5cea5f313c27e7767383e4 +size 627708 diff --git a/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/layout.json b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bea4d8f986dc59991a839ae2453847a5cb3118e9 --- /dev/null +++ b/2025/SIDA_ Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model/layout.json @@ -0,0 +1,11466 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 544, + 140 + ], + "type": "text", + "content": "SIDA: Social Media Image Deepfake Detection, Localization and Explanation with Large Multimodal Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": "Zhenglin Huang" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Jinwei Hu" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Xiangtai Li" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Yiwei He" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " \nXingyu Zhao" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Bei Peng" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Baoyuan Wu" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Xiaowei Huang" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "text", + "content": " Guangliang Cheng" + }, + { + "bbox": [ + 88, + 160, + 517, + 191 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 193, + 452, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 193, + 452, + 207 + ], + "spans": [ + { + "bbox": [ + 157, + 193, + 452, + 207 + ], + "type": "text", + "content": "1University of Liverpool, UK 2Nanyang Technological University, SG" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 208, + 512, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 208, + 512, + 220 + ], + "spans": [ + { + "bbox": [ + 85, + 208, + 512, + 220 + ], + "type": "text", + "content": "3 WMG, University of Warwick 4The Chinese University of Hong Kong, Shenzhen, Guangdong, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 153, + 222, + 456, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 222, + 456, + 235 + ], + "spans": [ + { + "bbox": [ + 153, + 222, + 456, + 235 + ], + "type": "text", + "content": "Project Page: https://hzlsaber.github.io/projects/SIDA/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 236, + 493, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 236, + 493, + 249 + ], + "spans": [ + { + "bbox": [ + 113, + 236, + 493, + 249 + ], + "type": "text", + "content": "† Corresponding author. E-mail: guangliang.cheng@liverpool.ac.uk xiangtai94@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 277, + 200, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 277, + 200, + 289 + ], + "spans": [ + { + "bbox": [ + 151, + 277, + 200, + 289 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 301, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 296, + 636 + ], + "type": "text", + "content": "The rapid advancement of generative models in creating highly realistic images poses substantial risks for misinformation dissemination. For instance, a synthetic image, when shared on social media, can mislead extensive audiences and erode trust in digital content, resulting in severe repercussions. Despite some progress, academia has not yet created a large and diversified deepfake detection dataset for social media, nor has it devised an effective solution to address this issue. In this paper, we introduce the Social media Image Detection dataSet (SID-Set), which offers three key advantages: (1) extensive volume, featuring 300K AI-generated/tampered and authentic images with comprehensive annotations, (2) broad diversity, encompassing fully synthetic and tampered images across various classes, and (3) elevated realism, with images that are predominantly indistinguishable from genuine ones through mere visual inspection. Furthermore, leveraging the exceptional capabilities of large multimodal models, we propose a new image deepfake detection, localization, and explanation framework, named SIDA (Social media Image Detection, localization, and explanation Assistant). SIDA not only discerns the authenticity of images, but also delineates tampered regions through mask prediction and provides textual explanations of the model's judgment criteria. Compared with state-of-the-art deepfake detection models on SID-Set and other benchmarks, extensive experiments demonstrate that SIDA achieves superior performance among diversified settings. The code, model, and dataset will be released." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 657, + 135, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 135, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 135, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Recent advances in generative AI [9, 57, 75] have significantly improved the ability to generate highly realistic images, making it easier to create content that closely resem" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 275, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 326, + 275, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 326, + 275, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 326, + 275, + 553, + 392 + ], + "type": "image", + "image_path": "bd1bf6fb12dc6a8e113de6aab54d4e307954fc8979f8f8bee3e359a79aed1bbb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. The framework comparisons. Existing deepfake methods (a-b) are limited to detection, localization, or both. In contrast, SIDA (c) offers a more comprehensive solution, capable of handling detection, localization, and explanation tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 449, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 652 + ], + "type": "text", + "content": "bles real-world events. However, these advancements also bring new risks of malicious misuse, particularly in creating deceptive content aimed at misleading public opinion or distorting historical records. Such concerns have motivated the computer vision community to develop more sophisticated deepfake detection techniques. Contemporary methods [32, 70] primarily focus on assessing the authenticity of facial images (i.e., real or fake), while an emerging subset aims to detect and localize facial manipulations [21, 22]. These methods are typically trained on datasets containing real and fake images, aiming to detect images as real or fake, or to localize the tampered regions. Consequently, the quality and diversity of the datasets used for training and evaluation play a crucial role in achieving high accuracy in deepfake detection and localization. A well-curated dataset can enable models to learn nuanced features, improving robustness and generalization in real-world scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "However, existing deepfake detection and localization datasets face two main challenges: 1) Insufficient Diversity. The majority of existing datasets for deepfake detection focus mainly on facial imagery [7, 24]. However, given the growing capabilities of generative AI, the issue of non" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28831" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "content": "facial image falsification on social media cannot be overlooked. While researchers [66, 86] have developed relatively large datasets based on ImageNet for image deepfake detection, these datasets typically consist of images from simple scenarios that do not specifically focus on social media. Additionally, they often utilize somewhat outdated image-generation techniques, which can result in less convincing forgeries that are easier for both humans and models to detect. Currently, there is a substantial lack of large-scale image deepfake datasets specifically designed for social media that leverage the latest generative methods. 2) Limited Comprehensiveness. Existing datasets are typically suited either for deepfake detection or for tampered region localization [21-23, 73], focusing on specific types of generative methods or image manipulations. However, an ideal deep-fake dataset should encompass a wide range of scenarios to reflect the complexity of real social media content, where fake images may be fully generated or manipulated through image editing strategies [75]. Furthermore, most existing datasets primarily focus on binary real/fake classification or tampered region localization, with limited emphasis on explaining the cues that models use to make these decisions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 338, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 637 + ], + "type": "text", + "content": "To address these challenges, we introduce the Social media Image Detection dataSet (SID-Set), which consists of 300K images (i.e., 100K real, 100K synthetic, and 100K tampered images), providing a comprehensive resource for the deepfake detection community. Additionally, we include textual descriptions explaining the model's judgment basis. As shown in Figures 2 and 4, synthetic and tampered images are indistinguishable to humans. In particular, challenges for the SID-Set include: 1) subtle alterations of just dozens of pixels; 2) natural-looking local manipulations; 3) complex scenes in datasets. To our knowledge, SID-Set is the first dataset of its scale with extensive annotations, making it the largest and most comprehensive dataset for social media deepfake detection to date. Compared to existing datasets in Table 1, SID-Set addresses the challenges of limited diversity and outdated generative techniques by providing a more comprehensive set of high-quality and diverse images. Accordingly, we propose a new VLMs-based deepfake detection framework, named the Social media Image Detection, localization, and explanation Assistant (SIDA), which achieves the state-of-the-art (SOTA) performance on SID-Set and generalizes effectively across other benchmarks. SIDA can serve as a baseline model on SID-Set, offering a new framework for tackling social media image deepfake detection and localization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 639, + 276, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 639, + 276, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 639, + 276, + 651 + ], + "type": "text", + "content": "The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 653, + 295, + 714 + ], + "type": "text", + "content": "- We establish SID-Set, a comprehensive benchmark for detecting, localizing, and explaining deepfakes in social media images, featuring multiple image types and extensive annotations. SID-Set holds the potential for advancing the field of deepfake detection and ensuring" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 91, + 549, + 189 + ], + "blocks": [ + { + "bbox": [ + 324, + 71, + 542, + 82 + ], + "lines": [ + { + "bbox": [ + 324, + 71, + 542, + 82 + ], + "spans": [ + { + "bbox": [ + 324, + 71, + 542, + 82 + ], + "type": "text", + "content": "Table 1. Comparison with existing image deepfake datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 91, + 549, + 189 + ], + "lines": [ + { + "bbox": [ + 316, + 91, + 549, + 189 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 549, + 189 + ], + "type": "table", + "html": "
DatasetContentData SourceGenerator YearMulticlasslesMasksExplanation
OHImg [39]OverheadGoogle Map2023✘✘✘
FakeSpotter [63]FaceCelebA, FFHQ2020✘✓✘
ForgeryNet [24]FaceCREMA-D2021✘✓✘
DCFace [27]FaceFFHQ2023✘✘✘
DFF [58]FaceIMDB-WIKI2023✘✘✘
RealFaces [48]FacePrompts2023✘✘✘
M3Dsynth [87]BiologyDDPM, CycleGAN2023✘✓✘
CNNSpot [64]ObjectSDXL2020✓✘✘
CIFAKE [3]ObjectCIFAR2023✓✘✘
CASIA 2.0 [12]GeneralCorel2022✘✓✘
Artifact [53]GeneralCOCO, FFHQ, LSUN2023✘✘✘
IMD2020 [45]GeneralPlaces22020✘✓✘
AIGCD [84]GeneralLSUN, COCO, FFHQ2023✘✘✘
GenImage [86]GeneralImageNet2023✘✘✘
SID-SetGeneralCOCO, Flickr30k, MagicBrush2024✓✓✓
", + "image_path": "c9b5686711e62ba70430b96c272e31cbcae4f6d32a99a874f48a26c67de818a3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 198, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 198, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 328, + 198, + 553, + 221 + ], + "type": "text", + "content": "robust model performance in complex real-world scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 223, + 553, + 353 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 320, + 223, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 223, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 320, + 223, + 553, + 281 + ], + "type": "text", + "content": "- We propose SIDA, a new image deepfake detection, localization, and explanation framework that not only detects images with high accuracy but also localizes and explains potential manipulations, enhancing the transparency and utility of deepfake detection technologies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 282, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 282, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 320, + 282, + 553, + 353 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that SIDA effectively identifies and delineates tampered areas within images, supporting the development of more robust and interpretable deepfake detection systems. Notably, SIDA demonstrates superior or equivalent performance on the SID-Set and other benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 365, + 400, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 365, + 400, + 376 + ], + "spans": [ + { + "bbox": [ + 314, + 365, + 400, + 376 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 454, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 454, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 454, + 396 + ], + "type": "text", + "content": "2.1. Image Deepfake Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 665 + ], + "type": "text", + "content": "In the realm of deepfake detection, the primary focus has historically centered on the identification of facial deepfakes. Renowned datasets such as ForgeryNet [24], DeepFakeFace [58], and DFFD [7] have been pivotal in this area. As the field evolves, there is a growing shift among researchers towards exploring non-facial deepfake data. Advanced methodologies involving text-to-image or imaged-to-image generation techniques [76], utilizing GANs or the stable diffusion series, have facilitated the creation of expansive deepfake datasets like GenImage [86], HiFi-IFDL [21], and DiffForensics [66]. These datasets are characterized by their enlarged data volumes, diversified generation methodologies, and enriched annotation details. Furthermore, beyond the conventional real/fake annotations, certain datasets [21, 22, 73] now include more granular annotations. Table 1 delineates a detailed comparison among various deepfake datasets, highlighting key differences in generation scenarios and annotation types supported. It shows that SID-Set is particularly tailored towards social media data, incorporating the latest SOTA generation models, emphasizing high-quality production, and providing much more comprehensive and diverse annotations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 541, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 541, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 541, + 685 + ], + "type": "text", + "content": "2.2. Image Deepfake Detection and Localization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Deepfake detection methods [4, 6, 38, 49, 62] are typically approached as classification tasks within the data-driven" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28832" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 293, + 144 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 293, + 144 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 293, + 144 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 293, + 144 + ], + "type": "image", + "image_path": "900e06e414fc30c68fb51eafa1819f86407d0770c2ebab240b89e4d32f372746.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 154, + 295, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 177 + ], + "type": "text", + "content": "Figure 2. SID-Set examples. The 1st row is the synthetic images, while the 2nd row shows tampered images. (Zoom in to view)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 197, + 297, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 297, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 297, + 508 + ], + "type": "text", + "content": "paradigm. These strategies primarily leverage diverse architectures [38, 49], including Convolutional Neural Networks (CNNs) and Transformers, to detect distinctive artifacts. Some scholars have attempted to achieve relatively high precision and generality by employing strategies such as employing techniques such as data augmentation [62], adversarial training [6], reconstruction [4], etc. On the other hand, some researchers [25, 60] have explored extracting features from the frequency domain for deepfake identification. Efforts [13, 65] have also been made to fuse features from both spatial and frequency domains to obtain a more comprehensive set of discriminative features for deepfake detection. Although these methods have shown some progress, they still struggle with issues of generalization. Furthermore, some scholars [20, 40, 44, 61, 80, 83] have gone beyond the basic classification between real and fake by gradually constructing datasets annotated with masks of locally tampered areas, thereby addressing both image deepfake detection and localization tasks. However, these datasets are concentrated mainly on facial data, with fewer datasets available for non-facial deepfake detection and localization, and even fewer large and public datasets for realistic social media data. As a result, our work aims to address these critical gaps by providing a new, comprehensive dataset that includes diverse manipulations beyond facial data, particularly focusing on social media images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 201, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 201, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 201, + 529 + ], + "type": "text", + "content": "2.3. Large Multimodal Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "content": "The progress in large language models (LLMs) [14-16, 41] and vision-language models (VLMs) [14-16, 29, 71, 78] has notably improved multimodal comprehension, seamlessly integrating visual and textual data. The LLaMA series [14-16] optimize language understanding with a compact, high-performance design using fewer parameters than prior models. LLaVA series [34, 35] enhance visual question answering by synchronizing visual features with textual data. Additionally, LISA series [29, 71] employ LLM for accurate image segmentation, merging visual perception with linguistic insights to precisely segment the targeted areas. Several grounding large multimodal models [50, 54, 55, 67, 68, 72, 74, 77, 81] have been proposed to localize the contents based on linguistic information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Advancements [5, 10, 19, 69, 82] in integrating multi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "modal data such as visual and linguistic information have also significantly improved the performance of models in detecting and pinpointing deepfakes. For instance, AntifakePrompt [5] approaches deepfake detection by formulating it as a visual question answering problem, which adjusts soft prompts for InstructBLIP [10], enabling it to determine whether a query image is real or fake. ForgeryGPT [31] enhances image forgery detection and localization by integrating advanced forensic knowledge with a mask-aware forgery extractor that targets pixel-level fraud detection. As a concurrent work, FakeShield [69] leverages the capabilities of LLaVA to identify and localize altered regions while providing interpretable insights into the findings. Our method diverges significantly from existing approaches by creating the most extensive deepfake dataset tailored specifically for social media images with comprehensive annotations. Additionally, we also set a new standard for social media image deepfake detection by utilizing the visual interpretation strengths of VLMs to boost detection accuracy, pinpoint forgeries, and provide clearer explanations within a comprehensive framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 338, + 389, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 389, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 389, + 350 + ], + "type": "text", + "content": "3. Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 358, + 389, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 358, + 389, + 370 + ], + "spans": [ + { + "bbox": [ + 314, + 358, + 389, + 370 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 377, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 628 + ], + "type": "text", + "content": "In the realm of deepfake detection and localization, research has predominantly focused on facial deepfakes due to their substantial societal impact. However, with the advancements in generative technology, the scope of deepfakes has extended beyond facial content to include non-facial manipulations. Historically, non-facial deepfakes were less prevalent, largely limited by technological constraints that produced low-quality, easily detectable forgeries. Although some datasets, such as GenImage [86] and AIGCD [84], have been constructed, they suffer from several limitations: 1) They often utilize relatively outdated generative technologies, resulting in lower quality data easily distinguished by humans. 2) They primarily focus on text-to-image or image-to-image generation, neglecting the need for data involving manipulations of specific regions, objects, or parts. Such tampered manipulations can be especially insidious as they introduce subtle misinformation, making existing SOTA deepfake detection methods less effective. 3) They lack well-defined criteria for content authenticity, limiting their effectiveness in providing interpretative insights and educating the public on distinguishing synthetic content." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Addressing these limitations is crucial for improving the transparency and utility of deepfake detection systems. Furthermore, most existing datasets emphasize either fully synthetic or tampered images, as shown in Table 1, but in real applications, we don't know the image deepfake type in advance. Effective deepfake detection and localization methods should be capable of addressing both scenarios, as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28833" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 82, + 192, + 224 + ], + "blocks": [ + { + "bbox": [ + 85, + 82, + 192, + 224 + ], + "lines": [ + { + "bbox": [ + 85, + 82, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 85, + 82, + 192, + 224 + ], + "type": "image", + "image_path": "d7d4e80538d0cedc06c6ff327ed89b1c9acbdf47bbcd410eed18cad201fc0b3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. Tampered image generation pipeline: It consists of four stages—extracting objects from captions using GPT-4o, obtaining object masks with Language-SAM, setting up replacement dictionaries for generating tampered images, and generating new images using Latent Diffusion. This figure illustrates an example of object replacement (e.g., “cat” to “dog”) and attribute modification." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 82, + 304, + 225 + ], + "blocks": [ + { + "bbox": [ + 197, + 82, + 304, + 225 + ], + "lines": [ + { + "bbox": [ + 197, + 82, + 304, + 225 + ], + "spans": [ + { + "bbox": [ + 197, + 82, + 304, + 225 + ], + "type": "image", + "image_path": "8344ea02b8ed04918e727bb17423363240ffa9770bdf7037f949469b4fe2bf08.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 82, + 414, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 82, + 414, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 82, + 414, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 82, + 414, + 225 + ], + "type": "image", + "image_path": "0516d556f88e5dce042f78bc0252f097d6fd95bc81476383774671076022833a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 419, + 82, + 526, + 225 + ], + "blocks": [ + { + "bbox": [ + 419, + 82, + 526, + 225 + ], + "lines": [ + { + "bbox": [ + 419, + 82, + 526, + 225 + ], + "spans": [ + { + "bbox": [ + 419, + 82, + 526, + 225 + ], + "type": "image", + "image_path": "0671702a1689addff8324475dc29763e71b7ab4ac300c689fed73c51c0cc48d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 291, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 375 + ], + "type": "text", + "content": "social media images often involve complex combinations of synthetic and tampered content. Therefore, developing a comprehensive benchmark for detecting and localizing deepfakes in social media images is essential. We propose SID-Set, which encompasses comprehensive, high-quality annotations for detection and localization, along with detailed textual explanations of the judgment criteria." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 382, + 198, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 198, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 198, + 393 + ], + "type": "text", + "content": "3.2. Benchmark Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 399, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 483 + ], + "type": "text", + "content": "Data Details. To develop an effective benchmark for detecting and localizing images on social media, we created SID-Set, a dataset with real, synthetic, and tampered images reflecting diverse real-world scenarios. Our benchmark assesses whether models can differentiate among real, synthetic, and tampered images, as well as accurately identify altered regions in tampered images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "type": "text", + "content": "Real Images: 100K images from OpenImages " + }, + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{V}7^{1}" + }, + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "type": "text", + "content": ", with a wide range of scenarios reflecting real-world diversity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 507, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 543 + ], + "type": "text", + "content": "Synthetic Images: 100K images generated through FLUX [43], specifically designed to challenge identification due to their high-quality, highly realistic appearance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 544, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 578 + ], + "type": "text", + "content": "Tampered Images: 100K tampered images, with specific objects or regions replaced or altered; the detailed generation process is shown in Figure 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 579, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 687 + ], + "type": "text", + "content": "Data Generation. To generate highly realistic synthetic images, we experimented with several open-source SOTA generative models, such as FLUX [43], Kandinsky 3.0 [1], SDXL [52], AbsoluteReality [42], and others. Following a review by five human experts of 1,000 images from each generative model, FLUX emerged as the top performer, producing highly convincing images that were indistinguishable from real ones to human experts. Consequently, we employed FLUX to create 100K synthetic images based on" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 289, + 354, + 322 + ], + "blocks": [ + { + "bbox": [ + 320, + 289, + 354, + 322 + ], + "lines": [ + { + "bbox": [ + 320, + 289, + 354, + 322 + ], + "spans": [ + { + "bbox": [ + 320, + 289, + 354, + 322 + ], + "type": "image", + "image_path": "de1d4cf5dd0cc21a7e9e57da614f084929da6b2922307435435695494833426b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 323, + 343, + 330 + ], + "lines": [ + { + "bbox": [ + 331, + 323, + 343, + 330 + ], + "spans": [ + { + "bbox": [ + 331, + 323, + 343, + 330 + ], + "type": "text", + "content": "bird" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 356, + 289, + 389, + 323 + ], + "blocks": [ + { + "bbox": [ + 356, + 289, + 389, + 323 + ], + "lines": [ + { + "bbox": [ + 356, + 289, + 389, + 323 + ], + "spans": [ + { + "bbox": [ + 356, + 289, + 389, + 323 + ], + "type": "image", + "image_path": "48ca8d52b993b4983b20c04513176cc78bea2e92700bd98cc2c3195281b72154.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 390, + 290, + 424, + 323 + ], + "blocks": [ + { + "bbox": [ + 390, + 290, + 424, + 323 + ], + "lines": [ + { + "bbox": [ + 390, + 290, + 424, + 323 + ], + "spans": [ + { + "bbox": [ + 390, + 290, + 424, + 323 + ], + "type": "image", + "image_path": "26044953c5c9b46fa493d87a54404aedd5ddf2de562b40abba2e99558414eda2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 324, + 417, + 331 + ], + "lines": [ + { + "bbox": [ + 401, + 324, + 417, + 331 + ], + "spans": [ + { + "bbox": [ + 401, + 324, + 417, + 331 + ], + "type": "text", + "content": "eagle" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 321, + 331, + 354, + 367 + ], + "blocks": [ + { + "bbox": [ + 321, + 331, + 354, + 367 + ], + "lines": [ + { + "bbox": [ + 321, + 331, + 354, + 367 + ], + "spans": [ + { + "bbox": [ + 321, + 331, + 354, + 367 + ], + "type": "image", + "image_path": "4095ae0660da8175d8a17be01a0d65d498f0c9a885816e8eba455e5b41d68b5b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 368, + 342, + 374 + ], + "lines": [ + { + "bbox": [ + 331, + 368, + 342, + 374 + ], + "spans": [ + { + "bbox": [ + 331, + 368, + 342, + 374 + ], + "type": "text", + "content": "cat" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 356, + 331, + 390, + 367 + ], + "blocks": [ + { + "bbox": [ + 356, + 331, + 390, + 367 + ], + "lines": [ + { + "bbox": [ + 356, + 331, + 390, + 367 + ], + "spans": [ + { + "bbox": [ + 356, + 331, + 390, + 367 + ], + "type": "image", + "image_path": "a49b555a043a0427611175aa13d4d3221685b254e3b1cf58c6317979564ac29e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 368, + 416, + 375 + ], + "lines": [ + { + "bbox": [ + 390, + 368, + 416, + 375 + ], + "spans": [ + { + "bbox": [ + 390, + 368, + 416, + 375 + ], + "type": "text", + "content": "dog" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 437, + 289, + 472, + 323 + ], + "blocks": [ + { + "bbox": [ + 437, + 289, + 472, + 323 + ], + "lines": [ + { + "bbox": [ + 437, + 289, + 472, + 323 + ], + "spans": [ + { + "bbox": [ + 437, + 289, + 472, + 323 + ], + "type": "image", + "image_path": "96e032ec2989046286f6412203fee5791be076ddb2334d39659de7a84bbdca30.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 323, + 471, + 331 + ], + "lines": [ + { + "bbox": [ + 439, + 323, + 471, + 331 + ], + "spans": [ + { + "bbox": [ + 439, + 323, + 471, + 331 + ], + "type": "text", + "content": "bodyboard" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 474, + 289, + 508, + 323 + ], + "blocks": [ + { + "bbox": [ + 474, + 289, + 508, + 323 + ], + "lines": [ + { + "bbox": [ + 474, + 289, + 508, + 323 + ], + "spans": [ + { + "bbox": [ + 474, + 289, + 508, + 323 + ], + "type": "image", + "image_path": "49f9a0dbdf350f25cd71c1376b5bcb405809ef7defae0f6c092eb534b022d5c3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 436, + 331, + 471, + 366 + ], + "blocks": [ + { + "bbox": [ + 436, + 331, + 471, + 366 + ], + "lines": [ + { + "bbox": [ + 436, + 331, + 471, + 366 + ], + "spans": [ + { + "bbox": [ + 436, + 331, + 471, + 366 + ], + "type": "image", + "image_path": "db7f9096a2b1372de8d9d344772e2f39c0c3ddbced67d20ca7d1e02339b80ca3.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 367, + 472, + 373 + ], + "lines": [ + { + "bbox": [ + 436, + 367, + 472, + 373 + ], + "spans": [ + { + "bbox": [ + 436, + 367, + 472, + 373 + ], + "type": "text", + "content": "coffee table" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 384, + 542, + 396 + ], + "lines": [ + { + "bbox": [ + 325, + 384, + 542, + 396 + ], + "spans": [ + { + "bbox": [ + 325, + 384, + 542, + 396 + ], + "type": "text", + "content": "Figure 4. Examples of tampered images. (Zoom in to view)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 474, + 331, + 508, + 366 + ], + "blocks": [ + { + "bbox": [ + 512, + 324, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 512, + 324, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 512, + 324, + 546, + 331 + ], + "type": "text", + "content": "bodyboard" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 331, + 508, + 366 + ], + "lines": [ + { + "bbox": [ + 474, + 331, + 508, + 366 + ], + "spans": [ + { + "bbox": [ + 474, + 331, + 508, + 366 + ], + "type": "image", + "image_path": "09257dd71993a5510489016414d1f4cc3b8aca492b0f3ab46b1ba410289d1c92.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 367, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 511, + 367, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 511, + 367, + 547, + 373 + ], + "type": "text", + "content": "coffee table" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 407, + 553, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 444 + ], + "type": "text", + "content": "the Flickr30k [51] and COCO [33]. The image tampering process depicted in Figure 4 follows four distinct stages, utilizing the COCO image as an example." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 445, + 554, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 516 + ], + "type": "text", + "content": "Stage 1: We extract objects from an image's caption using GPT-4o [47]. For instance, from the caption \"A large fluffy cat laying on top of a wooden table\", GPT-4o identifies relevant COCO class objects or retains nouns if no match exists. This extraction is documented in an \"Image-Caption-Object\" JSON file." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 518, + 553, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 553, + 541 + ], + "type": "text", + "content": "Stage 2: Employing Language-SAM [30], we generate masks for identified objects as training ground truth." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 544, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 615 + ], + "type": "text", + "content": "Stage 3: We establish dictionaries for full and partial image tampering using COCO classes for object replacement and attribute modifications, respectively. For example, replacing \"dog\" with animals like \"cat\" or adding attributes such as \"happy\" or \"angry\" to the \"dog\" class. For more details, please refer to the Appendix." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 616, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 664 + ], + "type": "text", + "content": "Stage 4: Utilizing Latent Diffusion [56], we modify captions and regenerate images, either replacing or retaining original objects based on availability. An example modification is altering \"cat\" to \"dog\" in the image caption." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "In total, we generated 80,000 object-tampered images and 20,000 partially tampered images. To demonstrate the explainability of SIDA, we used GPT-4o to generate textual descriptions for the judgment basis of 3,000 images from" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "1https://storage.googleapis.com/openimages/web/index.html" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28834" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 73, + 512, + 289 + ], + "blocks": [ + { + "bbox": [ + 97, + 73, + 512, + 289 + ], + "lines": [ + { + "bbox": [ + 97, + 73, + 512, + 289 + ], + "spans": [ + { + "bbox": [ + 97, + 73, + 512, + 289 + ], + "type": "image", + "image_path": "42197889730a533eac850997b77f30edbf2e7d1e393280ceebb589a8e13b85bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "type": "text", + "content": "Figure 5. The pipeline of SIDA: Given an image " + }, + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "type": "text", + "content": " and the corresponding text input " + }, + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 54, + 300, + 555, + 334 + ], + "type": "text", + "content": ", the last hidden layer for the token provides the detection result. If the detection result indicates a tampered image, SIDA extracts the token to generate masks for the tampered regions. This figure shows an example where the man's face has been manipulated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 353, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 296, + 426 + ], + "type": "text", + "content": "the SID-Set, divided equally among three types. Additionally, to ensure the realism of synthetic images, tampered images, and their textual descriptions, we engaged 5 annotation experts for quality control and adjustments. Detailed prompts and generation pipelines used to create descriptions for each type of image are provided in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 112, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 112, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 112, + 449 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 458, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 495 + ], + "type": "text", + "content": "In this section, we first present the model architecture of SIDA in Section 4.1, followed by an introduction to the training process of our method in Section 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 138, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 138, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 138, + 514 + ], + "type": "text", + "content": "4.1. Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 664 + ], + "type": "text", + "content": "Large vision-language models have demonstrated remarkable capabilities in understanding the alignment between textual and visual information. For instance, LLaVA [34] leverages language alone to achieve a comprehensive understanding of both visual and linguistic content. Building on LLaVA, LISA [29] extends this capability by providing fine-grained segmentation masks along with corresponding textual descriptions. However, to effectively detect and localize synthetic images, VLMs must not only be capable of multimodal understanding but also possess the ability to identify and segment manipulated regions, providing detailed explanations for both synthetic and tampered images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "To this end, we propose SIDA to tackle the task of synthetic image detection and tampered region localization. The pipeline of our method is illustrated in Figure 5. Inspired by previous approaches [28, 29, 85], we expand the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": "original vocabulary of VLMs by adding two new tokens, " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "<\\mathrm{DET}>" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": ", to enable the model to extract detection and segmentation information. Given an image " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": ", such as \"Can you identify if this image is real, fully synthetic, or tampered? Please mask the tampered object/part if it is tampered.\" We feed them into the VLM. The VLM then outputs a text description " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\mathrm{des}}" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": ", while the last hidden layer " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{hid}}" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": " contains the " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "<\\mathrm{DET}>" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 313, + 353, + 555, + 461 + ], + "type": "text", + "content": " tokens. This process can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 391, + 473, + 553, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 473, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 391, + 473, + 553, + 486 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {\\mathrm {d e s}} = \\operatorname {V L M} \\left(x _ {i}, x _ {t}\\right). \\tag {1}", + "image_path": "1b792dbfa1d452c74d1c3e560f2cf36b0c7b927a3915417dea1b637424b55c41.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": "Next, we extract the " + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "inline_equation", + "content": "<\\mathrm{DET}>" + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": " token from the last hidden layer " + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{hid}}" + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": ". The representation " + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": " is then passed through a detection head " + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": " to determine whether the image is real, fully synthetic, or tampered. We denote the final detection result by " + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\hat{\\mathsf{D}}" + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 400, + 563, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 563, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 400, + 563, + 553, + 576 + ], + "type": "interline_equation", + "content": "\\hat {\\mathrm {D}} = F _ {\\text {d e t}} \\left(h _ {\\text {d e t}}\\right), \\tag {2}", + "image_path": "b0a6cfec7a35849eedf6df0529b0aeab0c7d06284e1686bf238deab120662bb0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "content": " is the detection head that processes the extracted " + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{DET}\\rangle" + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "content": " representation " + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "inline_equation", + "content": "h_\\mathrm{det}" + }, + { + "bbox": [ + 313, + 582, + 553, + 605 + ], + "type": "text", + "content": " to produce the detection output." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": "If the detection result indicates that the image has been tampered with, SIDA will then predict a mask for the tampered regions. The " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": " feature is extracted from the hidden layer " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{hid}}" + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": ", similar to the extraction process for the token. Given that the token encapsulates crucial information that can aid in generating fine-grained segmentation masks, the representation " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": " is transformed using a fully connected layer " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": " to align with the dimensions of " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": ". To further capture the relationship between the " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{det}}" + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{seg}}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28835" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "features, we apply a single-layer Multihead Attention, facilitating effective feature interaction and enhancing mask quality. The process can be formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 118, + 193, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 118, + 193, + 130 + ], + "spans": [ + { + "bbox": [ + 130, + 118, + 193, + 130 + ], + "type": "interline_equation", + "content": "\\tilde {h} _ {\\mathrm {d e t}} = F (h _ {\\mathrm {d e t}}),", + "image_path": "e0357620bba77aaf9f68f6e88723dddf16ffe3056fbdb57d9069d94684a94a85.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 129, + 133, + 294, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 133, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 129, + 133, + 294, + 148 + ], + "type": "interline_equation", + "content": "\\tilde {h} _ {\\text {s e g}} = \\operatorname {M S A} \\left(\\tilde {h} _ {\\det }, h _ {\\text {s e g}}\\right), \\tag {3}", + "image_path": "97d3269cd5b960bf698d88c3667a802960a1ba730a7359c4f3f35b2306f1fcab.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 150, + 205, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 150, + 205, + 166 + ], + "spans": [ + { + "bbox": [ + 129, + 150, + 205, + 166 + ], + "type": "interline_equation", + "content": "\\tilde {h} _ {\\mathrm {s e g}} = \\tilde {h} _ {\\mathrm {s e g}} + h _ {\\mathrm {s e g}}.", + "image_path": "87abf2fb366f6891d2e1cb4b7d1d2a8ea59b83f3400b7ffeaa93bf7683972a41.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 174, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 233 + ], + "type": "text", + "content": "In this formulation, we treat the detection features as the query, and the segmentation features as the key and value. A residual connection is employed to combine both the original and transformed information, thereby enhancing the representation for precise segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": "Finally, we employ a frozen image encoder " + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": " to extract visual features from the input image " + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": ", resulting in visual features " + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": ". The segmentation embedding " + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{h}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": " is then combined with the visual features " + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": " and fed into a decoder to produce the final mask " + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 55, + 234, + 296, + 293 + ], + "type": "text", + "content": ". This can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 303, + 197, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 303, + 197, + 316 + ], + "spans": [ + { + "bbox": [ + 138, + 303, + 197, + 316 + ], + "type": "interline_equation", + "content": "f = F _ {\\mathrm {e n c}} (x _ {i}),", + "image_path": "4c1b5bf26c61bb44d7baf134dbcf1086614a86ef0f8cbca79a30c576ee5ffff0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 314, + 294, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 314, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 138, + 314, + 294, + 332 + ], + "type": "interline_equation", + "content": "\\hat {\\mathrm {M}} = F _ {\\mathrm {d e c}} \\left(\\tilde {h} _ {\\mathrm {s e g}}, f\\right). \\tag {4}", + "image_path": "002a55aaea3454e180bdc3628a80d12945f15b1de8c0ae2e22097d59f166cce7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 338, + 119, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 119, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 119, + 352 + ], + "type": "text", + "content": "4.2. Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": "Training Objectives. The training loss, " + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": ", for the SIDA consists of three components: the detection loss " + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": ", the text generation loss " + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": ", and the segmentation mask loss " + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": ". Initially, SIDA is trained in an end-to-end manner by employing the detection loss and the segmentation loss. For detection, we use CrossEntropy loss, while for the segmentation task, we use a weighted combination of binary cross-entropy (BCE) and DICE loss, with respective loss weights " + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{bce}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dice}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 464 + ], + "type": "text", + "content": ". This can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 474, + 215, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 474, + 215, + 485 + ], + "spans": [ + { + "bbox": [ + 91, + 474, + 215, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}},", + "image_path": "1843c5efac075a9e004717fa6b63ec62d46a7e6b807cc31714a30786a9d36eac.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 80, + 488, + 294, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 488, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 80, + 488, + 294, + 500 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e t} = \\mathcal {L} _ {\\mathrm {C E}} (\\hat {\\mathrm {D}}, \\mathrm {D}), \\tag {5}", + "image_path": "746c8fd52c5b9f73ef368af7222946eab58a3486f51c42ef0399b93bfbb1ac61.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 502, + 269, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 502, + 269, + 515 + ], + "spans": [ + { + "bbox": [ + 70, + 502, + 269, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {B C E} (\\hat {\\mathrm {M}}, \\mathrm {M}) + \\lambda_ {d i c e} \\mathcal {L} _ {D I C E} (\\hat {\\mathrm {M}}, \\mathrm {M}).", + "image_path": "5bc4b94a96a6daa21db5b5653b72ba489eb4c3bdcd0ad9807dfe700b28d31df6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": "After completing the training phase, we proceed to fine-tune the SIDA model by utilizing detailed textural descriptions from 3,000 images as the ground truth, represented by " + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "inline_equation", + "content": "y_{des}" + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": ". This phase focuses on optimizing the text generation component, " + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 55, + 525, + 296, + 596 + ], + "type": "text", + "content": ", to improve its ability in textural interpretability. The final loss function is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 606, + 185, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 606, + 185, + 617 + ], + "spans": [ + { + "bbox": [ + 88, + 606, + 185, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t x t} = \\mathcal {L} _ {\\mathrm {C E}} (y _ {\\hat {d e s}}, y _ {d e s}),", + "image_path": "50e2052c63d1b6eb9b5a1c21dce181e1821b8377c98dd14e16244f2ee1177316.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 82, + 616, + 294, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 616, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 82, + 616, + 294, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {d e t}} \\mathcal {L} _ {\\text {d e t}} + \\lambda_ {\\text {m a s k}} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {\\text {t x t}} \\mathcal {L} _ {\\text {t x t}}, \\tag {6}", + "image_path": "8245739b90dc12c8aed7f4eeb562d6a854cc8a5724c22ad05479af409125e2d3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_{det}" + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 55, + 642, + 296, + 677 + ], + "type": "text", + "content": " are the weighting factors that balance the contributions of the detection, segmentation, and text generation losses, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Training Data. We utilize the SID-Set, consisting of 300k images, to train SIDA. To further enhance diversity, we incorporate the MagicBrush dataset [79] after filtering out" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "low-quality images. The combined dataset supports robust training for both the detection and localization of synthetic content. Additionally, we generate descriptions for 3,000 randomly selected images using LLMs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 133, + 395, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 395, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 395, + 147 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": "Implementation Details. We choose LISA as the base large vision language model due to its strong capability for reasoning-based localization. We fine-tuned both LISA-7B-v1 and LISA-13B-v1 on the SID-Set using LoRA, setting " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": " at 16 and the dropout rate at 0.05. The input images are resized to " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": ". The loss weights in Eq. (6) for the detection " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{det}})" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": ", the text generation " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{txt}})" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": ", and the localization " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{mask}})" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": " are set to 1.0, respectively. The localization loss weights in Eq. (5) for the " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{bce}}" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dice}}" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": " are set to 2.0 and 0.5, respectively. To determine the optimal weight configuration, we perform ablation studies as detailed in Section 5.5. During the detection and localization training stage, the image encoder is frozen, and all other modules are trainable. For the text generation stage, only vision-language models are fine-tuned using the LoRA strategy. The initial learning rate is set to " + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 154, + 555, + 381 + ], + "type": "text", + "content": ", with a batch size of 2 per device and a gradient accumulation step of 10. We use two NVIDIA A100 GPUs (40GB each). Training for SIDA-7B and SIDA-13B took 48 hours and 72 hours, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate detection using image-level accuracy and F1 scores. For forgery localization, our metrics include Area Under the Curve (AUC), F1 scores, and Intersection over Union (IoU)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 440, + 435, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 435, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 435, + 451 + ], + "type": "text", + "content": "5.1. Detection Evaluation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 458, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 638 + ], + "type": "text", + "content": "We compare SIDA against other SOTA deepfake detection methods on SID-Set, including CnnSpot [17], AntifakePrompt [5], FreDect [18], Fusing [26], Gram-Net [37], UniFDP [46], LGrad [59], and LNP [2]. To ensure a fair comparison, we first evaluate these models on our dataset using their original pre-trained weights, then retrain them with the SID-Set to assess performance improvements. Table 2 demonstrates that SIDA achieves better or comparable results among all the evaluated methods. Notably, LGrad [59] achieves the highest accuracy and F1 score on tampered images after retraining, but this comes at the expense of lower performance in other metrics. Our analysis indicates that LGrad's high recall and false positive rates stem from its propensity to misclassify other types as tampered. The training details are provided in the Appendix." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 647, + 432, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 432, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 432, + 658 + ], + "type": "text", + "content": "5.2. Localization Results" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Table 3 presents the forgery localization performance on the SID-Set. We selected PSCC-Net [36], MVSS-Net [11], and HIFI-Net [21] as representative IFDL methods. Additionally, we chose LISA [29] as a representative LLM due to" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28836" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 102, + 547, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison of SIDA with other deepfake detection methods. Values outside parentheses are evaluated with open-source models directly on the SID-Set; values in parentheses indicate performance changes after training with the SID-Set. The best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 102, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 248 + ], + "type": "table", + "html": "
MethodsYearRealFully syntheticTamperedOverall
AccF1AccF1AccF1AccF1
AntifakePrompt [5]202464.8(↑24.1)78.6(↑10.5)93.8(↑3.7)96.8(↑1.1)30.8(↑60.1)47.2(↑33.2)63.1(↑29.1)69.3(↑23.4)
CnnSpott [17]202179.8(↑9.2)88.7(↑2.1)39.5(↑51.2)56.6(↑31.5)6.9(↑61.2)12.9(↑51.1)42.1(↑39.3)69.6(↑20.7)
FreDect [18]202083.7(↓37.7)91.1(↓43.5)16.8(↑44.1)28.8(↑37.2)11.9(↑25.2)21.3(↑31.7)37.4(↑33.6)23.4(↑47.2)
Fusing [26]202285.1(↑4.1)92.0(↑0.7)34.0(↑54.1)50.7(↑38.4)2.7(↑24.3)5.3(↑26.1)40.1(↑33.1)29.1(↑40.3)
Gram-Net [37]202070.1(↑19.1)82.4(↑9.3)93.5(↑4.4)96.6(↑2.0)0.8(↑89.1)1.6(↑85.3)55.0(↑37.1)58.0(↑35.1)
UnivFD [46]202368.0(↑0.3)67.4(↑1.1)62.1(↑24.3)87.5(↑10.5)64.0(↑28.5)85.3(↑4.7)64.0(↑21.7)85.3(↑4.5)
LGrad [59]202364.8(↓2.8)78.6(↓2.5)83.5(↓25.5)91.0(↓23.7)6.8(↑92.3)12.7(↑86.1)51.8(↑20.2)55.5(↑23.9)
LNP [2]202371.2(↓56.8)83.2(↓60.2)91.8(↓55.6)95.7(↓60.1)2.9(↑90.4)5.7(↑88.9)55.2(↓7.6)58.2(↑4.1)
SIDA-7B202489.191.098.798.691.291.093.593.5
SIDA-13B202489.691.198.598.792.991.293.693.5
", + "image_path": "10c8d3229089fe4ea36be612686c6f04d0109b7308d4ce10d64cd68570678b68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 268, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 353 + ], + "type": "text", + "content": "its segmentation reasoning capabilities. We used LISA-7Bv1 and fine-tuned it on SID-Set. The results indicate that SIDA achieves the best performance. We suppose that while LISA possesses strong general segmentation capabilities, it lacks the specific specialized features required to detect subtle manipulations, ultimately limiting the effectiveness of fine-tuning for precise forgery localization." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 399, + 279, + 509 + ], + "blocks": [ + { + "bbox": [ + 55, + 356, + 296, + 391 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 391 + ], + "type": "text", + "content": "Table 3. Comparison between SIDA and other IFDL approaches. * indicates the use of the pre-trained model from the original paper due to unavailable training code." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 399, + 279, + 509 + ], + "lines": [ + { + "bbox": [ + 69, + 399, + 279, + 509 + ], + "spans": [ + { + "bbox": [ + 69, + 399, + 279, + 509 + ], + "type": "table", + "html": "
MethodsYearsTampered
AUCF1IOU
MVSS-Net* [11]202348.931.623.7
HIFI-Net* [21]202364.045.921.1
PSCC-Net [36]202282.171.335.7
LISA-7B-v1 [29]202478.469.132.5
SIDA-7B202487.373.943.8
", + "image_path": "d2d5a4ba70606e8844e510a108f4d0c12034dc04f6d6f7ec2b6f3edabd48bf9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 516, + 162, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 162, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 162, + 529 + ], + "type": "text", + "content": "5.3. Robustness Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 666 + ], + "type": "text", + "content": "We further evaluate the robustness of SIDA against common image perturbations found in social media, such as JPEG compression, resizing, and Gaussian noise. Table 4 shows our model's performance on the SID-Set under six degradation scenarios: JPEG compression (with quality levels of 70 and 80), resizing (with scaling factors of 0.5 and 0.75), and Gaussian noise (with variances of 5 and 10). Despite not being explicitly trained on degraded data, SIDA demonstrates resilience to these low-level distortions. The model's stable performance against common social media perturbations highlights its robustness and practical applicability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 201, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 201, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 201, + 685 + ], + "type": "text", + "content": "5.4. Test on Other Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In this stage, we evaluate SIDA on DMImage [8] dataset to assess its generalization capabilities. We compare SIDA" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 281, + 542, + 394 + ], + "blocks": [ + { + "bbox": [ + 324, + 266, + 544, + 277 + ], + "lines": [ + { + "bbox": [ + 324, + 266, + 544, + 277 + ], + "spans": [ + { + "bbox": [ + 324, + 266, + 544, + 277 + ], + "type": "text", + "content": "Table 4. Performance of SIDA under different perturbations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 281, + 542, + 394 + ], + "lines": [ + { + "bbox": [ + 325, + 281, + 542, + 394 + ], + "spans": [ + { + "bbox": [ + 325, + 281, + 542, + 394 + ], + "type": "table", + "html": "
DetectionLocalization
ACCF1AUCF1IOU
JPEG 7089.490.186.271.842.3
JPEG 8088.789.585.871.141.7
Resize 0.589.391.186.872.543.2
Resize 0.7589.991.687.173.043.5
Gaussian 1086.989.384.170.241.0
Gaussian 588.489.985.371.041.5
SIDA-7B93.593.587.373.943.8
", + "image_path": "cb1049f7235e88308195a8ced4c1d1ab278c39bdf17805d429c2f8875f5126ae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 475 + ], + "type": "text", + "content": "with CNNSpot [17], Fusing [26], Gram-Net [37], LNP [2], UnivFD [46], and AntifakePrompt [5]. For these methods, we use the original hyperparameter settings and pre-trained weights provided by the authors. The results in Table 5 show that SIDA achieves superior performance, demonstrating its strong adaptability." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 320, + 514, + 549, + 618 + ], + "blocks": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": "Table 5. Comparison with other deepfake detection methods on DMImage [8]. We used original weights for each method." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 514, + 549, + 618 + ], + "lines": [ + { + "bbox": [ + 320, + 514, + 549, + 618 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 549, + 618 + ], + "type": "table", + "html": "
MethodsRealFakeOverall
AccF1AccF1AccF1
CNNSpot [17]87.888.428.444.240.643.3
Gram-Net [37]62.854.178.888.167.479.4
Fusing [26]87.786.115.527.240.436.5
LNP [2]63.167.456.972.558.268.3
UnivFD [46]89.488.344.961.253.960.7
AntifakePrompt [5]91.392.589.391.290.691.2
SIDA-7B92.993.190.791.091.892.4
", + "image_path": "a95dd370faa02191c41db8e2dce6fa3e249fd87b03f4db9379422553e6e04633.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 408, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 408, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 408, + 649 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Attention Module. We conducted ablation experiments to assess the importance of the attention module. Variants included removing the attention module entirely and replacing it with fully connected (FC) layers. Results in Table 6 show that removing the attention module or replacing" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28837" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 81, + 285, + 147 + ], + "blocks": [ + { + "bbox": [ + 64, + 60, + 286, + 72 + ], + "lines": [ + { + "bbox": [ + 64, + 60, + 286, + 72 + ], + "spans": [ + { + "bbox": [ + 64, + 60, + 286, + 72 + ], + "type": "text", + "content": "Table 6. Ablation study results for attention module in SIDA." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 81, + 285, + 147 + ], + "lines": [ + { + "bbox": [ + 62, + 81, + 285, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 81, + 285, + 147 + ], + "type": "table", + "html": "
DetectionLocalization
ACCF1AUCF1IOU
FC91.190.384.371.638.9
w/o Attention90.389.984.171.338.8
SIDA93.593.587.373.943.8
", + "image_path": "3be4828393098d6cd896c94fcd57e334a9d7dcb1983fbe7350ec3da5a626892d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 295, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 192 + ], + "type": "text", + "content": "it with FC significantly reduces performance, underscoring the critical role of attention in enhancing feature interaction and improving detection and localization accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": "Training Weights. SIDA training utilizes weighted losses to balance task contributions. In the detection and localization stages, detection loss is adjusted by a weight " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{det}}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": ", and localization loss by binary cross-entropy (BCE) and DICE losses, with weights " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{bce}}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dice}}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": " respectively. For our experiments, we set " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{det}}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": " to 1, " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{bce}}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": " to 2.0, and " + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dice}}" + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": " to 0.5 to maintain balance between detection and localization, enhancing model stability and performance. We summarize the outcomes of various weight configurations in Table 7." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 88, + 320, + 259, + 370 + ], + "blocks": [ + { + "bbox": [ + 69, + 304, + 281, + 316 + ], + "lines": [ + { + "bbox": [ + 69, + 304, + 281, + 316 + ], + "spans": [ + { + "bbox": [ + 69, + 304, + 281, + 316 + ], + "type": "text", + "content": "Table 7. Different weight configurations in SIDA training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 320, + 259, + 370 + ], + "lines": [ + { + "bbox": [ + 88, + 320, + 259, + 370 + ], + "spans": [ + { + "bbox": [ + 88, + 320, + 259, + 370 + ], + "type": "table", + "html": "
λdetλbceλdiceAccF1 Score
1.02.00.593.5691.01
1.04.01.093.4990.86
", + "image_path": "f205179618c5c520474973b0d3cec0dfe0232ca1fa530d5ab1d858fe57eb92ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 167, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 167, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 167, + 391 + ], + "type": "text", + "content": "5.6. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": "In this section, we present examples of SIDA's output for tampered images, showcasing its detection, localization, and explanation capabilities. SIDA accurately identifies tampered regions and provides explanations for its decisions. We also include some challenging failure cases where SIDA was unable to accurately detect the tampered regions, highlighting areas for future improvement. Additional visualization results are available in the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 212, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 212, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 212, + 514 + ], + "type": "text", + "content": "6. Conclusion and Discussions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 641 + ], + "type": "text", + "content": "In this work, we present SID-Set for social image deepfake detection, localization, and explanation tasks, consisting of 100k real images, 100k fully synthetic images, and 100k tampered images. Furthermore, we propose a new VLMs-based deepfake detection framework, SIDA, to address these tasks. SIDA demonstrates its ability to detect fake types, localize tampered regions, and provide explanations for its decisions. We believe that the integration of VLMs into deepfake detection tasks, as demonstrated by SIDA, offers promising new avenues for future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Although the development of the SID-Set and the introduction of the SIDA framework have yielded favorable outcomes in deepfake detection tasks, we recognize some potential limitations and will optimize them in future research. Dataset Size. While SID-Set includes 100k fully synthetic and 100k tampered images, the complexity of real social" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 85, + 487, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 85, + 487, + 101 + ], + "spans": [ + { + "bbox": [ + 328, + 85, + 487, + 101 + ], + "type": "text", + "content": "Can you identify if this image is real, full synthetic, or tampered? Please mask the tampered object/part if it is tampered." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 333, + 101, + 379, + 144 + ], + "blocks": [ + { + "bbox": [ + 333, + 101, + 379, + 144 + ], + "lines": [ + { + "bbox": [ + 333, + 101, + 379, + 144 + ], + "spans": [ + { + "bbox": [ + 333, + 101, + 379, + 144 + ], + "type": "image", + "image_path": "cea45cf8fea55b4926330f894c18d05f6deb08fff83a21fd7e5076479c50d5ef.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 386, + 102, + 430, + 143 + ], + "blocks": [ + { + "bbox": [ + 386, + 102, + 430, + 143 + ], + "lines": [ + { + "bbox": [ + 386, + 102, + 430, + 143 + ], + "spans": [ + { + "bbox": [ + 386, + 102, + 430, + 143 + ], + "type": "image", + "image_path": "c200b6ce03b3931f3d454cab54e2684f47613a32ea7ecb693f80664fd17f538d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 101, + 478, + 143 + ], + "blocks": [ + { + "bbox": [ + 435, + 101, + 478, + 143 + ], + "lines": [ + { + "bbox": [ + 435, + 101, + 478, + 143 + ], + "spans": [ + { + "bbox": [ + 435, + 101, + 478, + 143 + ], + "type": "image", + "image_path": "5378186f9d0b3ca33adfc77b147ccfa0659099504b64c2a085bb99be9287f1c4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 486, + 100, + 531, + 143 + ], + "blocks": [ + { + "bbox": [ + 486, + 100, + 531, + 143 + ], + "lines": [ + { + "bbox": [ + 486, + 100, + 531, + 143 + ], + "spans": [ + { + "bbox": [ + 486, + 100, + 531, + 143 + ], + "type": "image", + "image_path": "079219d1ce457548ba22ecf335969b56e524b9638d61fcaf3b92efeb35f77ca0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 326, + 148, + 348, + 161 + ], + "blocks": [ + { + "bbox": [ + 326, + 148, + 348, + 161 + ], + "lines": [ + { + "bbox": [ + 326, + 148, + 348, + 161 + ], + "spans": [ + { + "bbox": [ + 326, + 148, + 348, + 161 + ], + "type": "image", + "image_path": "ce7e752e97820e7a8226687c15abd7008aceaf0b6bd19c156788361551a45a9c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 167, + 369, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 167, + 369, + 181 + ], + "spans": [ + { + "bbox": [ + 348, + 167, + 369, + 181 + ], + "type": "text", + "content": "Object ampered" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 332, + 183, + 379, + 224 + ], + "blocks": [ + { + "bbox": [ + 332, + 183, + 379, + 224 + ], + "lines": [ + { + "bbox": [ + 332, + 183, + 379, + 224 + ], + "spans": [ + { + "bbox": [ + 332, + 183, + 379, + 224 + ], + "type": "image", + "image_path": "bb062c8e102d27df3f642a0a9898045943daa8236db0862d47fb0aafada622e4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 386, + 167, + 424, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 167, + 424, + 182 + ], + "spans": [ + { + "bbox": [ + 386, + 167, + 424, + 182 + ], + "type": "text", + "content": "Partial tampered" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 388, + 183, + 432, + 224 + ], + "blocks": [ + { + "bbox": [ + 388, + 183, + 432, + 224 + ], + "lines": [ + { + "bbox": [ + 388, + 183, + 432, + 224 + ], + "spans": [ + { + "bbox": [ + 388, + 183, + 432, + 224 + ], + "type": "image", + "image_path": "e319df19ca84f1e3b75245fd10a4f9810d97beb316ed10ad146bd76cc6d2a7f6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 434, + 167, + 481, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 434, + 167, + 481, + 182 + ], + "spans": [ + { + "bbox": [ + 434, + 167, + 481, + 182 + ], + "type": "text", + "content": "Partial tampered" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 435, + 184, + 481, + 228 + ], + "blocks": [ + { + "bbox": [ + 435, + 184, + 481, + 228 + ], + "lines": [ + { + "bbox": [ + 435, + 184, + 481, + 228 + ], + "spans": [ + { + "bbox": [ + 435, + 184, + 481, + 228 + ], + "type": "image", + "image_path": "62013276a45753571f0ff721eabd0b1b48e45e5abfbc92bb27c42901d3dc32ac.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 498, + 167, + 520, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 498, + 167, + 520, + 181 + ], + "spans": [ + { + "bbox": [ + 498, + 167, + 520, + 181 + ], + "type": "text", + "content": "Object tampered" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 488, + 184, + 533, + 226 + ], + "blocks": [ + { + "bbox": [ + 488, + 184, + 533, + 226 + ], + "lines": [ + { + "bbox": [ + 488, + 184, + 533, + 226 + ], + "spans": [ + { + "bbox": [ + 488, + 184, + 533, + 226 + ], + "type": "image", + "image_path": "c2429f968347ef108e363714fc4ae8c9e58a74641c8ceef03c045d5a6284092b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 333, + 229, + 363, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 229, + 363, + 277 + ], + "spans": [ + { + "bbox": [ + 333, + 229, + 363, + 277 + ], + "type": "text", + "content": "Type of Image: object tampered The tampered areas are specifically located in the h The tampered. Absolute Positino The location is" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 389, + 229, + 429, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 229, + 429, + 277 + ], + "spans": [ + { + "bbox": [ + 389, + 229, + 429, + 277 + ], + "type": "text", + "content": "Type of Image: part tampered Tampering Type and Location... Relative Position: The tampered part located in the face of the women... Absolute Position..." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 437, + 230, + 478, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 230, + 478, + 277 + ], + "spans": [ + { + "bbox": [ + 437, + 230, + 478, + 277 + ], + "type": "text", + "content": "Type of Image: object \ntampered Tampering \nType and Location is the face of the man... \nAbsolute Position \nTampered Content: Modification... \nInconsistencies: Lighting..." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 488, + 230, + 531, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 230, + 531, + 274 + ], + "spans": [ + { + "bbox": [ + 488, + 230, + 531, + 274 + ], + "type": "text", + "content": "Type of Image: part tampered Tampered Content: Modification... Relative Position: Inconsistencies... Colour in the background looks ..." + } + ] + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 327, + 281, + 348, + 298 + ], + "blocks": [ + { + "bbox": [ + 327, + 281, + 348, + 298 + ], + "lines": [ + { + "bbox": [ + 327, + 281, + 348, + 298 + ], + "spans": [ + { + "bbox": [ + 327, + 281, + 348, + 298 + ], + "type": "image", + "image_path": "11cdcee45f455b539c81086e3fcd9513808fc335d8ccea4096713c27a0c98e59.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 391, + 283, + 457, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 283, + 457, + 292 + ], + "spans": [ + { + "bbox": [ + 391, + 283, + 457, + 292 + ], + "type": "text", + "content": "a) Correct examples" + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 333, + 304, + 378, + 346 + ], + "blocks": [ + { + "bbox": [ + 333, + 304, + 378, + 346 + ], + "lines": [ + { + "bbox": [ + 333, + 304, + 378, + 346 + ], + "spans": [ + { + "bbox": [ + 333, + 304, + 378, + 346 + ], + "type": "image", + "image_path": "14d6f60b7b0334a52063389610d2bea7f39c4c3b83e067466feb6f3c10cf8e27.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 383, + 304, + 427, + 346 + ], + "blocks": [ + { + "bbox": [ + 383, + 304, + 427, + 346 + ], + "lines": [ + { + "bbox": [ + 383, + 304, + 427, + 346 + ], + "spans": [ + { + "bbox": [ + 383, + 304, + 427, + 346 + ], + "type": "image", + "image_path": "eb7b756972ab1bf586fc87727112f928b3ae74a50723f70cdf684013e6c7afc1.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 434, + 304, + 479, + 346 + ], + "blocks": [ + { + "bbox": [ + 434, + 304, + 479, + 346 + ], + "lines": [ + { + "bbox": [ + 434, + 304, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 434, + 304, + 479, + 346 + ], + "type": "image", + "image_path": "c2fe3771f02eab9bbb4d1c8c29db5f187e33548a73c30bd7e7ce0133cfda18a9.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 487, + 304, + 531, + 346 + ], + "blocks": [ + { + "bbox": [ + 487, + 304, + 531, + 346 + ], + "lines": [ + { + "bbox": [ + 487, + 304, + 531, + 346 + ], + "spans": [ + { + "bbox": [ + 487, + 304, + 531, + 346 + ], + "type": "image", + "image_path": "f1a5e589bf67da6b38cec5820377fc7b1981fcf416524b65def7ad4e2ba1e3d5.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 334, + 348, + 378, + 392 + ], + "blocks": [ + { + "bbox": [ + 334, + 348, + 378, + 392 + ], + "lines": [ + { + "bbox": [ + 334, + 348, + 378, + 392 + ], + "spans": [ + { + "bbox": [ + 334, + 348, + 378, + 392 + ], + "type": "image", + "image_path": "512d5490c30048e1b356504f7698c6ba0cd23d4f5ef762b7f836ba62bd9771aa.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 383, + 348, + 427, + 392 + ], + "blocks": [ + { + "bbox": [ + 383, + 348, + 427, + 392 + ], + "lines": [ + { + "bbox": [ + 383, + 348, + 427, + 392 + ], + "spans": [ + { + "bbox": [ + 383, + 348, + 427, + 392 + ], + "type": "image", + "image_path": "3e87ad72ff1ac4415100190a6a8af837624093e789d0f5cf7c3696f8f9a1bede.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 435, + 348, + 479, + 392 + ], + "blocks": [ + { + "bbox": [ + 435, + 348, + 479, + 392 + ], + "lines": [ + { + "bbox": [ + 435, + 348, + 479, + 392 + ], + "spans": [ + { + "bbox": [ + 435, + 348, + 479, + 392 + ], + "type": "image", + "image_path": "0140ded12571ec68e5754f4aee42949c691cc306aa0071d125340cac24d60006.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 487, + 348, + 531, + 392 + ], + "blocks": [ + { + "bbox": [ + 487, + 348, + 531, + 392 + ], + "lines": [ + { + "bbox": [ + 487, + 348, + 531, + 392 + ], + "spans": [ + { + "bbox": [ + 487, + 348, + 531, + 392 + ], + "type": "image", + "image_path": "e600bd8b9b358bc9f427a2fc421aa5ef57d09b29234c191ab40f76c238401334.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 333, + 395, + 378, + 439 + ], + "blocks": [ + { + "bbox": [ + 333, + 395, + 378, + 439 + ], + "lines": [ + { + "bbox": [ + 333, + 395, + 378, + 439 + ], + "spans": [ + { + "bbox": [ + 333, + 395, + 378, + 439 + ], + "type": "image", + "image_path": "36ab3eb1f014cf103ac1139689ef532fbdead74790683bd375209dd21c2fbe90.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 445, + 457, + 455 + ], + "lines": [ + { + "bbox": [ + 391, + 445, + 457, + 455 + ], + "spans": [ + { + "bbox": [ + 391, + 445, + 457, + 455 + ], + "type": "text", + "content": "b) Failure examples" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 383, + 395, + 427, + 440 + ], + "blocks": [ + { + "bbox": [ + 383, + 395, + 427, + 440 + ], + "lines": [ + { + "bbox": [ + 383, + 395, + 427, + 440 + ], + "spans": [ + { + "bbox": [ + 383, + 395, + 427, + 440 + ], + "type": "image", + "image_path": "ad4bcb55a045272f48a59cacd607cd3f77c9142b8494f6f864dbff97e04911c2.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 470, + 531, + 482 + ], + "lines": [ + { + "bbox": [ + 336, + 470, + 531, + 482 + ], + "spans": [ + { + "bbox": [ + 336, + 470, + 531, + 482 + ], + "type": "text", + "content": "Figure 6. Visual results of SIDA on tampered images." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 435, + 396, + 479, + 439 + ], + "blocks": [ + { + "bbox": [ + 435, + 396, + 479, + 439 + ], + "lines": [ + { + "bbox": [ + 435, + 396, + 479, + 439 + ], + "spans": [ + { + "bbox": [ + 435, + 396, + 479, + 439 + ], + "type": "image", + "image_path": "9cc53ac93d82242eb6072fe6b0402e0cdb5f7dfcc25c56274a934a5e4e8464da.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 487, + 396, + 531, + 440 + ], + "blocks": [ + { + "bbox": [ + 487, + 396, + 531, + 440 + ], + "lines": [ + { + "bbox": [ + 487, + 396, + 531, + 440 + ], + "spans": [ + { + "bbox": [ + 487, + 396, + 531, + 440 + ], + "type": "image", + "image_path": "6a4d2a53420c420336ae94a1141e2184c8d71d248cdb2cd33a130a3d20594551.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 501, + 553, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 553, + 536 + ], + "type": "text", + "content": "media environments demands a larger dataset. Therefore, expanding the dataset with additional images is a crucial objective for future research." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 537, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 656 + ], + "type": "text", + "content": "Data Domain. Other methods often generate social media images that lack authenticity or are easily identifiable. Consequently, we used only FLUX to generate fake images due to its superior quality. However, depending exclusively on one method may lead to issues of data skew, although this was not significantly evident in our experiments. Such skew could potentially impair performance on diverse datasets. Moving forward, we plan to explore additional generative methods and integrate various other generation techniques to produce a more varied and higher-quality set of images." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 657, + 553, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 553, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 553, + 704 + ], + "type": "text", + "content": "Localization Results. Although SIDA demonstrates relatively strong performance on our dataset, there is still room for improvement. Certain tampered regions are not reliably detected, underscoring the need for further advancements." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28838" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 124 + ], + "type": "text", + "content": "This work is supported by The Alan Turing Institute (UK) through the project 'Turing-DSO Labs Singapore Collaboration' (SDCfP2\\100009)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 145, + 115, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 115, + 158 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 115, + 158 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 166, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 166, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 296, + 209 + ], + "type": "text", + "content": "[1] Vladimir Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. Arxiv, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 210, + 296, + 243 + ], + "type": "text", + "content": "[2] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C. Cosman. Detecting generated images by real images only. Arxiv, 2023. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 244, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 244, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 295, + 277 + ], + "type": "text", + "content": "[3] Jordan J. Bird and Ahmad Lotfi. CIFAKE: image classification and explainable identification of ai-generated synthetic images. IEEE Access, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 278, + 295, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 295, + 320 + ], + "type": "text", + "content": "[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 323, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 355 + ], + "type": "text", + "content": "[5] You-Ming Chang, Chen Yeh, Wei-Chen Chiu, and Ning Yu. Antifakeprompt: Prompt-tuned vision-language models are fake image detectors. Arxiv, 2023. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 357, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 399 + ], + "type": "text", + "content": "[6] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 401, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 295, + 443 + ], + "type": "text", + "content": "[7] Harry Cheng, Yangyang Guo, Tianyi Wang, Liqiang Nie, and Mohan S. Kankanhalli. Diffusion facial forgery detection. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 446, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 295, + 487 + ], + "type": "text", + "content": "[8] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 491, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 521 + ], + "type": "text", + "content": "[9] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE TPAMI, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 524, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 524, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 524, + 295, + 578 + ], + "type": "text", + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 622 + ], + "type": "text", + "content": "[11] Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, and Xirong Li. Mvss-net: Multi-view multi-scale supervised networks for image manipulation detection. IEEE TPAMI, 2023. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "type": "text", + "content": "[12] Jing Dong, Wei Wang, and Tieniu Tan. CASIA image tampering detection evaluation database. In ISIP, 2013. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Junxian Duan, Yang Ai, Jipeng Liu, Shenyuan Huang, Huaibo Huang, Jie Cao, and Ran He. Test-time forgery detection with spatial-frequency prompt learning. IJCV, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 692, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 712 + ], + "type": "text", + "content": "[14] Abhimanyu Dubey et al. The llama 3 herd of models. Arxiv, 2024. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[15] Hugo Touvron et al. Llama: Open and efficient foundation language models. Arxiv, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "text", + "content": "[16] Hugo Touvron et al. Llama 2: Open foundation and finetuned chat models. Arxiv, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 120, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 553, + 140 + ], + "type": "text", + "content": "[17] Joel Frank and Thorsten Holz. Cnn-generated images are surprisingly easy to spot...for now. Arxiv, 2021. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 143, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 196 + ], + "type": "text", + "content": "[18] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[19] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In AAAI, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[20] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. Trufor: Leveraging all-round clues for trustworthy image forgery detection and localization. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "type": "text", + "content": "[21] Xiao Guo, Xiaohong Liu, Zhiyuan Ren, Steven Grosz, Iacopo Masi, and Xiaoming Liu. Hierarchical fine-grained image forgery detection and localization. In CVPR, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 364 + ], + "type": "text", + "content": "[22] Xiao Guo, Xiaohong Liu, Iacopo Masi, and Xiaoming Liu. Language-guided hierarchical fine-grained image forgery detection and localization. Arxiv, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 366, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 408 + ], + "type": "text", + "content": "[23] Ruidong Han, Xiaofeng Wang, Ningning Bai, Yihang Wang, Jianpeng Hou, and Jianru Xue. HDF-net: Capturing homogeneity difference features to localize the tampered image. IEEE TPAMI, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[24] Yinan He, Bei Gan, Siyu Chen, Yichun Zhou, Guojun Yin, Luchuan Song, Lu Sheng, Jing Shao, and Ziwei Liu. Forgerynet: A versatile benchmark for comprehensive forgery analysis. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[25] Yonghyun Jeong, Doyeon Kim, Youngmin Ro, and Jongwon Choi. Frepgan: Robust deepfake detection using frequency-level perturbations. In AAAI, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "text", + "content": "[26] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In ICIP, 2022. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[27] Minchul Kim, Feng Liu, Anil K. Jain, and Xiaoming Liu. Dcface: Synthetic face generation with dual condition diffusion model. In CVPR, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "type": "text", + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), 2021. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 602, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 635 + ], + "type": "text", + "content": "[29] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: reasoning segmentation via large language model. In CVPR, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "type": "text", + "content": "[30] lang-sam team. Lang-segment-anything. https://github.com/luca-medeiros/lang-segment-anything, 2024. Accessed: 2024-11-15. 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Jiawei Li, Fanrui Zhang, Jiaying Zhu, Esther Sun, Qiang Zhang, and Zheng-Jun Zha. Forgerygpt: Multimodal large language model for explainable image forgery detection and localization. Arxiv, 2024. 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28839" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 711 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[32] Li Lin, Neeraj Gupta, Yue Zhang, Hainan Ren, Chun-Hao Liu, Feng Ding, Xin Wang, Xin Li, Luisa Verdoliva, and Shu Hu. Detecting multimedia generated by large AI models: A survey. Arxiv, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 183 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 258 + ], + "type": "text", + "content": "[36] Xiaohong Liu, Yaojie Liu, Jun Chen, and Xiaoming Liu. Pssc-net: Progressive spatio-channel correlation network for image manipulation detection and localization. IEEE TCSVT, 2022, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[37] Zhengzhe Liu, Xiaojuan Qi, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. In CVPR, 2020. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "text", + "content": "[38] Momina Masood, Marriam Nawaz, Khalid Mahmood Malik, Ali Javed, Ain Irtaza, and Hafiz Malik. Deepfakes generation and detection: state-of-the-art, open challenges, countermeasures, and way forward. Appl. Intell., 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "text", + "content": "[39] Brandon B. May, Kirill Trapeznikov, Shengbang Fang, and Matthew C. Stamm. Comprehensive dataset of synthetic and manipulated overhead imagery for development and evaluation of forensic tools. In IH&MMSec, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "text", + "content": "[40] Changtao Miao, Qi Chu, Zhentao Tan, Zhenchao Jin, Wanyi Zhuang, Yue Wu, Bin Liu, Honggang Hu, and Nenghai Yu. Multi-spectral class center network for face manipulation detection and localization. Arxiv, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[41] Shervin Minaee, Tomás Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. Arxiv, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[42] Absolutereality model team. Absolutereality model repository. https://huggingface.co/jochemstoel/absolutereality-model-repository, 2024. Accessed: 2024-11-15. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 535 + ], + "type": "text", + "content": "[43] Flux model team. Flux model. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. Accessed: 2024-11-15. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[44] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In CVPR, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "text", + "content": "[45] Adam Novozámsky, Babak Mahdian, and Stanislav Saic. IMD2020: A large-scale annotated dataset tailored for detecting manipulated images. In WACvw, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "type": "text", + "content": "[46] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 252, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 252, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 252, + 669 + ], + "type": "text", + "content": "[47] OpenAI. GPT-4 technical report. Arxiv, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 711 + ], + "type": "text", + "content": "[48] Lorenzo Papa, Lorenzo Faiella, Luca Corvitto, Luca Maiano, and Irene Amerini. On the use of stable diffusion for creating realistic faces: from generation to detection. In IWBF, 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[49] Gan Pei, Jiangning Zhang, Menghan Hu, Guangtao Zhai, Chengjie Wang, Zhenyu Zhang, Jian Yang, Chunhua Shen, and Dacheng Tao. Deepfake generation and detection: A benchmark and survey. Arxiv, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 159 + ], + "type": "text", + "content": "[50] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. Arxiv, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[51] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. IJCV, 2017. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[52] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "text", + "content": "[53] Md Awsafur Rahman, Bishmoy Paul, Najibul Haque Sarker, Zaber Ibn Abdul Hakim, and Shaikh Anowarul Fattah. Artifact: A large-scale dataset with artificial and factual images for generalizable and robust synthetic image detection. In ICIP, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "text", + "content": "[54] Hanoona Abdul Rasheed, Muhammad Maaz, Sahal Shaji Mullappilly, Abdelrahman M. Shaker, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Eric P. Xing, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Glamm: Pixel grounding large multimodal model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "text", + "content": "[55] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[57] Xincheng Shuai, Henghui Ding, Xingjun Ma, Rongcheng Tu, Yu-Gang Jiang, and Dacheng Tao. A survey of multimodal-guided image editing with text-to-image diffusion models. Arxiv, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 503 + ], + "type": "text", + "content": "[58] Haixu Song, Shiyu Huang, Yinpeng Dong, and Wei-Wei Tu. Robustness and generalizability of deepfake detection: A study with diffusion models. Arxiv, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 505, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 546 + ], + "type": "text", + "content": "[59] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 548, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 590 + ], + "type": "text", + "content": "[60] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In AAAI, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "text", + "content": "[61] Dragos-Constantin Tântaru, Elisabeta Oneata, and Dan Oneata. Weakly-supervised deepfake localization in diffusion-generated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 669 + ], + "type": "text", + "content": "[62] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[63] Run Wang, Felix Juefei-Xu, Lei Ma, Xiaofei Xie, Yihao Huang, Jian Wang, and Yang Liu. Fakespotter: A simple yet robust baseline for spotting ai-synthesized fake faces. In IJCAI, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28840" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[64] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot...for now. In CVPR, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[65] Yuan Wang, Kun Yu, Chen Chen, Xiyuan Hu, and Silong Peng. Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In CVPR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[66] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[67] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. Arxiv, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[68] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. GSVA: generalized segmentation via multimodal large language models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[69] Zhipei Xu, Xuanyu Zhang, Runyi Li, Zecheng Tang, Qing Huang, and Jian Zhang. Fakeshield: Explainable image forgery detection and localization via multi-modal large language models. Arxiv, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[70] Zhiyuan Yan, Yong Zhang, Xinhang Yuan, Siwei Lyu, and Baoyuan Wu. Deepfakebench: A comprehensive benchmark of deepfake detection. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[71] Senqiao Yang, Tianyuan Qu, Xin Lai, Zhuotao Tian, Bohao Peng, Shu Liu, and Jiaya Jia. An improved baseline for reasoning segmentation with large language model. *Arxiv*, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[72] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In ICLR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 453 + ], + "type": "text", + "content": "[73] Zeqin Yu, Jiangqun Ni, Yuzhen Lin, Haoyi Deng, and Bin Li. Diffforensics: Leveraging diffusion prior to image forgery detection and localization. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[74] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. Arxiv, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[75] Fangneng Zhan, Yingchen Yu, Rongliang Wu, Jiahui Zhang, Shijian Lu, Lingjie Liu, Adam Kortylewski, Christian Theobalt, and Eric P. Xing. Multimodal image synthesis and editing: The generative AI era. IEEE TPAMI, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[76] Chenshuang Zhang, Chaoning Zhang, Mengchun Zhang, and In So Kweon. Text-to-image diffusion models in generative AI: A survey. Arxiv, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "text", + "content": "[77] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, and Jainwei Yang. Llava-grounding: Grounded visual chat with large multimodal models. In ECCV, 2024.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "text", + "content": "[78] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE TPAMI, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[79] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[80] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 397 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. In ICCV, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[81] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Change Loy Chen, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[82] Yue Zhang, Ben Colman, Ali Shahriyari, and Gaurav Bharaj. Common sense reasoning for deep fake detection. *Arxiv*, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 175, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 240 + ], + "type": "text", + "content": "[83] Yi Zhang, Changtao Miao, Man Luo, Jianshu Li, Wenzhong Deng, Weibin Yao, Zhe Li, Bingyu Hu, Weiwei Feng, Tao Gong, and Qi Chu. MFMS: learning modality-fused and modality-specific features for deepfake detection and localization tasks. In Proceedings of the 32nd ACM International Conference on Multimedia, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[84] Nan Zhong, Yiran Xu, Zhenxing Qian, and Xinpeng Zhang. Rich and poor texture contrast: A simple yet effective approach for ai-generated image detection. Arxiv, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[85] Luowei Zhou, Hamid Palangi, Lei Zhang, Houdong Hu, Jason J. Corso, and Jianfeng Gao. Unified vision-language pretraining for image captioning and VQA. In AAAI, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "text", + "content": "[86] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[87] Giada Zingarini, Davide Cozzolino, Riccardo Corvi, Giovanni Poggi, and Luisa Verdoliva. M3DSYNTH: A dataset of medical 3d images with ai-generated local manipulations. In ICASSP, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28841" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_content_list.json b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..abb9b31891b7ef6946db75a3d19e6a0c25fb2ea5 --- /dev/null +++ b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_content_list.json @@ -0,0 +1,1567 @@ +[ + { + "type": "text", + "text": "SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 143, + 128, + 852, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leigang $\\mathrm{Qu}^{1}$ , Haochuan $\\mathrm{Li}^{1}$ , Wenjie Wang $^{2*}$ , Xiang Liu $^{1}$ , Juncheng Li $^{3}$ , Liqiang Nie $^{4}$ , Tat-Seng Chua $^{1}$", + "bbox": [ + 99, + 202, + 901, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore, $^{2}$ University of Science and Technology of China, $^{3}$ Zhejiang University, $^{4}$ Harbin Institute of Technology (Shenzhen)", + "bbox": [ + 151, + 223, + 852, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leigangqu@gmail.com, haochuan@u.nus.edu, wenjiewang96@gmail.com, liu.xiang@u.nus.edu", + "bbox": [ + 130, + 258, + 874, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "junchengli@zju.edu.cn, nieliqiang@gmail.com, dcscts@nus.edu.sg", + "bbox": [ + 225, + 277, + 774, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) have demonstrated impressive capabilities in multimodal understanding and generation, pushing forward advancements in text-to-image generation. However, achieving accurate text-image alignment for LMMs, particularly in compositional scenarios, remains challenging. Existing approaches, such as layout planning for multi-step generation and learning3 from human feedback or AI feedback, depend heavily on prompt engineering, costly human annotations, and continual upgrading, limiting flexibility and scalability. In this work, we introduce a model-agnostic iterative self-improvement framework (SILMM) that can enable LMMs to provide helpful and scalable self-feedback and optimize text-image alignment via Direct Preference Optimization (DPO). DPO can readily applied to LMMs that use discrete visual tokens as intermediate image representations; while it is less suitable for LMMs with continuous visual features, as obtaining generation probabilities is challenging. To adapt SILMM to LMMs with continuous features, we propose a diversity mechanism to obtain diverse representations and a kernel-based continuous DPO for alignment. Extensive experiments on three compositional text-to-image generation benchmarks validate the effectiveness and superiority of SILMM, showing improvements exceeding $30\\%$ on T2I-CompBench++ and around $20\\%$ on DPG-Bench. The code is available at https://silmm.github.io/.", + "bbox": [ + 88, + 359, + 483, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 785, + 220, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) are advancing rapidly, surpassing Large Language Models (LLMs) by embracing multimodal capabilities for multimodal content perception, understanding [34, 35, 48], and generation [19, 64].", + "bbox": [ + 89, + 811, + 483, + 873 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2a02d5224f9cc65de72e7c1d7be7c620243e628608b017949dacdddc5f5129a.jpg", + "image_caption": [ + "a green leaf and a yellow butterfly", + "Attribute Binding" + ], + "image_footnote": [], + "bbox": [ + 563, + 362, + 642, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1b44fcf49d54ead8088bc18a910e92cb54313e2396db903a353262060fa865e.jpg", + "image_caption": [ + "five balloons", + "Counting Error" + ], + "image_footnote": [], + "bbox": [ + 668, + 362, + 746, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f261033ec7922fb1d68e85ed0876be40ec2c8eff32759e14816be1d011adfe6.jpg", + "image_caption": [ + "a bicycle on the left of a man", + "Relation Inaccuracy" + ], + "image_footnote": [], + "bbox": [ + 776, + 362, + 854, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef5c257411ef4ff8323d4097b7fddc94923a468d02bd53f027adbc7e3f820697.jpg", + "image_caption": [ + "(a) Text-image misalignment in compositional prompts", + "(b) Discrete & Continuous LMMs for text-to-image generation", + "Figure 1. Illustration of (a) text-image misalignment in compositional prompts and (b) comparison of discrete and continuous LMMs for T2I. Given a prompt, discrete LMMs can sample diverse token sequences from categorical distributions, while continuous LMMs can only produce a single deterministic feature vector. Note that the input learnable embeddings are optional for some continuous LMMs [64]." + ], + "image_footnote": [], + "bbox": [ + 534, + 465, + 890, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In particular, LMMs demonstrate promising abilities in interpreting user input prompts for text-to-image generation (T2I) [55, 57], producing vivid and photorealistic images. However, as shown in Fig. 1(a), achieving precise text-image alignment between generated images and complex prompts remains challenging, especially for compositional prompts involving multiple objects, attributes, counting, and complex relationships [6, 16, 49].", + "bbox": [ + 511, + 761, + 906, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance text-image alignment, existing work falls", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18497", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into two primary research lines. One line focuses on decomposing the T2I task into multiple stages. For example, some methods perform layout planning before generating the image [17, 37, 75]; while some split the image into sections for multi-step generation via multi-agent collaboration [46, 70]. However, these methods depend on extensive multi-step prompt engineering, which risks error accumulation. The second research line emphasizes learning from human feedback (RLHF [43]) to improve text-image alignment [15, 31, 33, 67, 73], or using AI feedback (RLAIF) from strong evaluation approaches or reward models [3, 77]. Nevertheless, it is labor-intensive and costly to obtain extensive high-quality human feedback, which is also often required to train external reward models [5]. Additionally, as LMMs evolve, the external evaluation approaches and reward models may require continual upgrading [44, 77].", + "bbox": [ + 89, + 90, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitations, we consider utilizing LLMs' inherent discriminative capabilities to self-improve their generation quality for text-image alignment. This offers a pathway for LMMs to evolve for T2I independently, without relying on human or external feedback. To pursue self-improvement, the key steps are: 1) generating diverse images by LMMs based on a given prompt, ensuring the image diversity to facilitate subsequent self-assessment and optimization; 2) using LMMs to self-assess text-image alignment in the generated images, producing alignment scores as self-feedback; and 3) adopting the self-feedback to optimize LMMs to generate superior visual tokens, resulting in images that better align with text prompts.", + "bbox": [ + 89, + 333, + 480, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, achieving the above objectives faces significant challenges. In particular:", + "bbox": [ + 89, + 531, + 480, + 561 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) As shown in Fig. 1(b), LMMs typically generate intermediate visual representations, i.e., discrete visual tokens or continuous visual features, which are then converted into images by a decoder (e.g., a diffusion model) [19, 64]. For LMMs with discrete visual tokens [19, 68, 76], using existing sampling strategies (e.g., adjusting temperature) in the autoregressive generation process can obtain diverse visual tokens. However, it is non-trivial for LMMs with deterministic continuous visual features, such as Dream-LLM [13], to sample diverse visual representations1.", + "2) Compositional prompts require LMMs to inspect object counts, attributes, and complex relationships in the generated images. However, existing LMMs still struggle with compositional cross-modal assessment [7, 56], challenging the generation of faithful self-feedback.", + "3) Optimizing LMMs with self-feedback is also intricate. Supervised Fine-Tuning (SFT) [11] and certain RLAIF methods [3, 77] require highly accurate self-feedback. Moreover, another representative method Direct Prefer-" + ], + "bbox": [ + 83, + 561, + 480, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ence Optimization (DPO) requires modeling generation distributions, from which we need to sample diverse images to construct pairwise training data, which is challenging for LMMs with continuous visual features.", + "bbox": [ + 524, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the above challenges, we propose an Self-Improving Large Multimodal Models (SILMM) framework for iterative optimization. As illustrated in Fig. 2, SILMM operates through five steps: 1) Compositional Prompt Generation prompts an LMM to imagine compositional scenarios and generate compositional prompts. 2) Diverse Image Generation. For discrete LMMs $^2$ , we follow the sampling decoding strategy commonly used in LLM alignment [43, 53]. For continuous LMMs [13, 64, 72], we propose a diversification strategy named $DropDiv$ , inspired by Monte Carlo (MC) Dropout [18], to perform dropout on the MLP layers of LMMs for diverse visual features, producing diverse images. 3) Decompositional Self-Questioning. To reduce the difficulty of compositional cross-modal assessment, LMMs can decompose a compositional prompt into atomic concepts and relations and generate questions for multi-step assessment. 4) VQA-based Self-Feedback. For each image generated in Step 2, LMMs can use the decomposed questions to assess text-image alignment, and then aggregate the results to obtain reasonable self-feedback. 5) Learning from Self-Feedback. For discrete LMMs, we directly apply DPO based on pairwise samples from Step 2. As to continuous LMMs, we propose Kernel-based Continuous DPO (KC-DPO), inducing a quadruplet objective with kernel functions for pairwise distance regulation over continuous visual features. The above five steps can iteratively repeat until self-improvement performance converges.", + "bbox": [ + 511, + 152, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are threefold:", + "bbox": [ + 535, + 561, + 862, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our knowledge, we are the first to focus on the task of LMMs' self-improvement for T2I. We propose a model-agnostic self-improvement framework to enable LMMs to achieve high-quality self-feedback and learning.", + "- For continuous LMMs, we introduce a dropout-based strategy to diversify image representations, along with a continuous DPO approach, i.e., KC-DPO, to optimize LMMs with preference representation pairs.", + "- We conduct extensive experiments on three compositional T2I benchmarks, demonstrating the superiority of SILMM, e.g., $30\\%$ improvements on T2I-CompBench++." + ], + "bbox": [ + 513, + 577, + 903, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 758, + 653, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compositional Text-to-Image Generation. Diffusion models [55, 57] have marked a significant advancement in T2I generation due to their stability and scalability. However, they still struggle with text-image alignment, such as attribute binding, counting error, and relation confu", + "bbox": [ + 511, + 786, + 903, + 861 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "Sampling diverse images at the decoder stage is inapplicable, as it can only optimize the decoder yet we aim to optimize LMMs to generate superior visual representations for text-image alignment in this work.", + "bbox": [ + 89, + 862, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2For simplicity, we denote LLMs outputting discrete and continuous visual representations as discrete and continuous LMMs, respectively.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18498", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6950e498312f2f876f3561238d0d0cc017869b5dda0a32a6faeb0d364c754bc6.jpg", + "image_caption": [ + "Figure 2. Schematic illustration of SILMM, comprising five steps: 1) LMMs generate compositional prompts by sampling based on provided instructions. 2) Diverse representations and images are generated using either discrete nucleus sampling or the proposed continuous DivDrop. 3) LMMs divide each compositional prompt into semantic units and generate questions for each unit. 4) VQA is conducted to answer these questions, with the answers and likelihoods aggregated into alignment scores as self-feedback. 5) For alignment tuning, DPO is applied for discrete LMMs, while the proposed KC-DPO is used for continuous LMMs." + ], + "image_footnote": [], + "bbox": [ + 124, + 99, + 898, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion [16, 51]. To enhance compositional T2I, some approaches intervene in language structures [16] or cross-attention mechanisms [6]. Other methods [17, 37, 38, 49] incorporate layout planning by LLMs or use multi-agent collaboration [46, 70]. Inspired by alignment successes in LLMs, recent work [5, 15, 67] applies RLHF [43] to optimize diffusion models. Despite the progress, they rely on inductive biases, extensive prompt engineering, or labor-intensive annotations, limiting flexibility and scalability.", + "bbox": [ + 88, + 549, + 480, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multimodal Models. The pioneering LMMs [39, 79] integrate a visual encoder, e.g., CLIP [52], with LLMs as the foundation, showing impressive multimodal understanding capabilities. To extend LMMs to visual generation, recent approaches align diffusion models [13, 19, 72] with LLMs or train a single transformer [62, 68, 74, 76]. According to the form of output visual features, they can be divided into discrete visual tokenization methods [19, 50, 62, 68] and continuous visual representation methods [13, 64, 72]. While LLM integration enhances language understanding and supports flexible applications (e.g., interleaved multimodal generation [64]), compositional T2I in the context of LMMs remains underexplored.", + "bbox": [ + 88, + 686, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning from AI Feedback. The high cost of collect", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing human preference has spurred research into RLAIF [3]. Benefiting from the convenience and scalability, there have been a series of studies adopting RLAIF to tackle a range of NLP tasks [10, 32, 78] and vision-language understanding [69, 77]. Despite the thrilling success, they only focus on text generation, overlooking the potential of RLAIF in other modalities. In contrast, we explore self-improving LMMs by activating multimodal understanding abilities for T2I. Particularly, we propose continuous strategies meticulously tailored to continuous visual features.", + "bbox": [ + 511, + 549, + 906, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 717, + 648, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we elaborate on the proposed method, including the SILMM framework with five steps and the iteration strategy (Sec. 3.1), as illustrated in Fig. 2. Afterward, we introduce the continuous KC-DPO applied to LMMs with continuous visual features in Sec. 3.2.", + "bbox": [ + 511, + 744, + 906, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Self-Improving Large Multimodal Models", + "text_level": 1, + "bbox": [ + 511, + 830, + 872, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 1: Compositional Prompt Generation. We first divide compositional scenarios into four categories: Attribute (color, shape, texture), Layout (counting, spatial relation),", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18499", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic Relation, and Complex Composition. Complex composition includes any possible composition of the first three. For attribute and layout, we prompt the LMM to separately generate common objects, attributes, numbers, and spatial relations, and then use templates to compose these concepts. For semantic relation and complex composition, we adopt in-context learning [12] to generate prompts. More details can be found in App. 6.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Diverse Representation and Image generation. The purpose of this step is to sample diverse intermediate visual representations from the LLM backbone $\\pi$ of an LMM, given a text prompt $x$ , which would be decoded into images with different qualities. These representations are denoted as $\\mathcal{Z} = \\{z_i,\\dots,z_M\\}$ , where $z_{i}\\sim \\pi (z|x)$ . For discrete LMMs [19], $z_{i}$ is a discrete visual sequence. We follow the common practice [43, 53] in language generation to obtain $\\mathcal{Z}$ , by sampling with different random seeds during auto-regressive decoding. For continuous LMMs [13], the LLM can only output a fixed continuous visual feature, without diversity. To tackle this issue, we propose DropDiv. First, we insert the dropout operations in the last few MLP layers of LLMs, which introduces randomness and enables LLMs for sampling. During inference, we activate these dropout operations to output diverse representations by sampling: $z_{i}\\sim \\pi^{\\prime}(z|x)$ , where $z_{i}$ denotes a continuous visual feature and $\\pi^\\prime$ represents the LLM with activated dropout operations. Afterward, these diverse visual representations $\\mathcal{Z}$ are decoded into images as $\\mathcal{Y} = \\{y_1,\\dots,y_M\\}$ .", + "bbox": [ + 89, + 215, + 482, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Unlike prior work [5, 15] focused on tuning diffusion models, our approach resorts to LLM backbones in LLMs to control image decoders (e.g., diffusion models) for better text-image alignment, centering on LLM backbone optimization. Our approach offers three key advantages: 1) LLMs demonstrate superior proficiency in prompt comprehension over text encoders [52, 54] commonly employed in diffusion models. Tuning LLM backbones may unlock their enormous potential for compositional T2I, especially in complex scenarios. 2) Tuning diffusion models is often constrained by efficiency challenges inherent to iterative likelihood estimation, whereas there have been well-established technologies [1, 40, 53] for LLM alignment. 3) Our method is orthogonal to existing methods to tune diffusion models, combining them may get further gains.", + "bbox": [ + 89, + 518, + 482, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: Decompositional Self-Questioning. To provide helpful feedback to the generated images, the LMM should first accurately assess text-image alignment, which requires strong compositional reasoning abilities. However, current advanced LMMs still suffer from compositional reasoning [42], such as spatial relation understanding [7] and counting [56]. To improve compositional reasoning, we introduce a divide-and-conquer strategy [77] for self-questioning. Specifically, the LMM first divides the given prompt $x$ into atomic concepts (e.g., \"a white harp\") and", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relations (e.g., \"a pancake is on the left of a pasta\"), and then generates questions $\\mathcal{Q} = \\{q_1, \\dots, q_N\\}$ , each $q_i$ corresponding to a concept or relation. For simplicity, the generated questions are constrained to be yes/no questions (e.g., \"Is there a while harp?\", \"Is the pancake on the left of the pasta?\"). Refer to App. 7 for more details on prompt templates of self-questioning.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 4: VQA-based Self-Feedback. Taking a generated image $y \\in \\mathcal{V}$ and all the questions $\\mathcal{Q}$ as input, the LMM conducts the VQA task, and the average difference between the probabilities of answering \"yes\" and \"no\" serves as the text-image alignment score:", + "bbox": [ + 511, + 198, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns (x, y) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ p \\left(\" y e s\" \\mid y, q _ {i}\\right) - p \\left(\" n o\" \\mid y, q _ {i}\\right) \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 282, + 903, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we adopt the vision-language understanding abilities of LMMs via VQA to provide feedback to the images generated by themselves, thus this step is named VQA-based self-feedback. We carry out this step through all the sampled images prompted by $x$ and get all the scores $S = \\{ s(x,y_j) | y_j \\in \\mathcal{V} \\}$ .", + "bbox": [ + 511, + 335, + 903, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 5: Learning from Self-Feedback. Based on the selffeedback alignment scores, we sample representation pairs $(z_w,z_l)$ from $\\mathcal{Q}$ , where $z_{w}$ and $z_{l}$ denote the chosen and the rejected representations and their corresponding decoded images should satisfy $s(x,y_w) > s(x,y_l)$ . With the preference data, we optimize the LLM backbone with DPO [53]:", + "bbox": [ + 511, + 428, + 903, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P O}} = - \\mathbb {E} _ {(x, z _ {w}, z _ {l}) \\sim \\mathcal {D}}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 532, + 687, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(z _ {w} | x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(z _ {w} | x\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(z _ {l} | x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(z _ {l} | x\\right)}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 551, + 903, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ denotes the training set, and $\\pi_{\\theta}$ and $\\pi_{\\mathrm{ref}}$ represent the policy and reference models, respectively. $\\sigma$ is the sigmoid function, and $\\beta$ is a hyperparameter controlling the deviation from the reference model.", + "bbox": [ + 511, + 595, + 903, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Iterative Self-Improvement. After learning from self-feedback, the updated LMM becomes more likely to generate preferred representations that are decoded into images better aligned with the prompt. This improvement in overall text-image alignment motivates us to iterate the above five steps with the updated LMM as the new reference model. The iteration mechanism continues until the alignment performance converges. As the process is independent of human annotations and external models, it is cost-effective and scalable. More importantly, it showcases the potential for self-improvement in LMMs by harmonizing their understanding and generation capabilities.", + "bbox": [ + 511, + 657, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Continuous Direct Preference Optimization", + "text_level": 1, + "bbox": [ + 511, + 848, + 883, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the step of learning from self-feedback, LMMs are optimized using the DPO objective as shown in Eqn. (2).", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18500", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The difference between discrete and continuous LMMs in this learning process lies in the calculation of the likelihood $\\pi (z|x)$ . For discrete LMMs, $\\pi (z|x)$ can be straightforwardly obtained by the softmax categorical distribution. However, for continuous LMMs with unknown distribution modeling, calculating $\\pi (z|x)$ is intractable.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Predictive Distribution with MC Dropout. MC Dropout [18] enables predictive distribution estimation via Monte Carlo simulation to calculate $\\pi(z|x)$ . Specifically, the dropout layers3 in an LMM are activated during inference and the LMM performs forward propagation multiple times to get multiple outputs. Assuming a Gaussian distribution, we can estimate its parameters and calculate the likelihood $\\pi(z|x)$ based on these outputs. However, such multi-forward estimation imposes a significant computational burden during training, making this approach insufficient and impractical.", + "bbox": [ + 89, + 181, + 482, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simplified Kernel-based Continuous DPO. Inspired by MC Dropout and motivated by its insufficiency issue, we propose a simplified method to achieve continuous DPO. Concretely, the intermediate representation $z$ often performs as a feature matrix $\\pmb{H} \\in \\mathbb{R}^{L \\times D}$ where $L$ and $D$ denote the sequence length and dimension. $\\pmb{H}$ can be attained by a Q-Former [13, 20] or from the last layer of the LMM in an autoregressive way [63, 64]. To estimate $\\pi(\\pmb{H} | x)$ , we first make a decomposition as:", + "bbox": [ + 89, + 347, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi (\\boldsymbol {H} | x) = \\prod_ {i = 1} ^ {L} \\pi \\left(\\boldsymbol {h} _ {i} \\mid \\boldsymbol {H} _ {< i}, x\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 492, + 482, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{h}_i \\in \\mathbb{R}^D$ denotes the $i$ -th feature vector. Based on the Gaussian assumption, we have:", + "bbox": [ + 89, + 542, + 482, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi \\left(\\boldsymbol {h} _ {i} \\mid \\boldsymbol {H} _ {< i}, x\\right) = \\frac {\\exp \\left[ - \\frac {1}{2} \\left(\\boldsymbol {h} _ {i} - \\boldsymbol {\\mu} _ {i}\\right) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\boldsymbol {h} _ {i} - \\boldsymbol {\\mu} _ {i}\\right) \\right]}{\\sqrt {(2 \\pi) ^ {D} \\left| \\boldsymbol {\\Sigma} _ {i} \\right|}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 579, + 482, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu}_i$ and $\\pmb{\\Sigma}_i$ denote the mean vector and the covariance matrix, respectively. Furthermore, we further simplify and approximate this formula: 1) the mean vector is estimated by the direct output of the continuous LMM, i.e., $\\pmb{\\mu}_i\\approx \\mathrm{LMM}(\\boldsymbol {x})[i]$ , and 2) the Gaussian distribution is isotropic and all dimensions share the same variance value $\\bar{\\sigma}$ , i.e., $\\pmb{\\Sigma}_i\\approx \\mathrm{diag}(\\sigma_1,\\dots,\\sigma_{\\mathrm{D}})$ and $\\sigma_{1} = \\ldots = \\sigma_{D} = \\bar{\\sigma}$ , and $\\bar{\\sigma}$ can be learnable or viewed as a hyperparameter.", + "bbox": [ + 89, + 619, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compute the simplified likelihood with Eqn. (4), obtain the joint one with Eqn. (3), and finally derive the continuous DPO based on Eqn. (2):", + "bbox": [ + 89, + 739, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} \\mathcal {L} _ {\\mathrm {C - D P O}} = & - \\mathbb {E} _ {(x, \\boldsymbol {H} _ {w}, \\boldsymbol {H} _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\frac {\\beta}{2 \\bar {\\sigma} ^ {2}} (- \\| \\boldsymbol {H} - \\boldsymbol {H} _ {w} \\| _ {F} ^ {2} + \\| \\boldsymbol {H} _ {r} - \\boldsymbol {H} _ {w} \\| _ {F} ^ {2})\\right) \\right], \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 789, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _F$ denotes the Frobenius norm, and $H$ and $H_{r}$ represent the continuous feature matrices from the policy and reference LMMs, respectively. $H_{w}$ and $H_{l}$ refer to the chosen and rejected feature matrices, respectively. Compared with the MC dropout method, this objective only requires one forward pass, which is more efficient. We relegate more details of the derivation to App. 8. From Eqn. (5), we can see that this objective aims to adjust the relative distances within the quadruple $(\\pmb {H},\\pmb {H}_r,\\pmb {H}_w,\\pmb {H}_l)$ and the distance metric is the Euclidean distance between two matrices. To further improve the flexibility, we generalize the continuous DPO objective to,", + "bbox": [ + 511, + 90, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {K C - D P O}} = - \\mathbb {E} _ {(x, \\boldsymbol {H} _ {w}, \\boldsymbol {H} _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\gamma (- k (\\boldsymbol {H}, \\boldsymbol {H} _ {w}) \\right. \\right. \\\\ \\left. \\left. + k \\left(\\boldsymbol {H} _ {r}, \\boldsymbol {H} _ {w}\\right) + k \\left(\\boldsymbol {H}, \\boldsymbol {H} _ {l}\\right) - k \\left(\\boldsymbol {H} _ {r}, \\boldsymbol {H} _ {l}\\right)\\right)\\right) \\Bigg ], \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 270, + 903, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma = \\frac{\\beta}{2\\bar{\\sigma}^2}$ controls the degree of adherence to the reference model, $k(\\cdot ,\\cdot)$ denotes a generalized distance measurement function. Considering it is similar to kernel methods [22, 60], we name the objective Kernel-based Continuous DPO (KC-DPO). In the following experiments section, we will discuss different distance functions and their influences on alignment performance.", + "bbox": [ + 511, + 349, + 903, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 472, + 645, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 498, + 702, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base Model Settings. We implement our method on DreamLLM (continuous LMM) [13] and SEED-LLaMA (discrete LMM) [19] for all experiments. We also apply our method to Emu-3 [68], the recent state-of-the-art discrete LMM. Details on DPO training are provided in App. 9.", + "bbox": [ + 511, + 521, + 903, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We curated a dataset of 16,000 prompts across four categories using LMM. In each DPO training iteration, images generated by the model in the previous iteration served as the training data for next DPO iteration, allowing for iterative self-improvement. Details on data creation are provided in App. 10.", + "bbox": [ + 511, + 598, + 903, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Benchmarks.** We evaluate our method on three text-to-image alignment benchmarks and follow their default settings. T2I-CompBench++ [28] consists of 8,000 compositional text prompts organized into 4 main categories: attribute, layout, non-spatial, and complex compositions, further divided into 8 subcategories, including color, binding, binding, 2D/3D-spatial relationships, non-spatial relationships, numeracy, and complex compositions. TIFA [27] uses pre-generated question-answer pairs and a VQA model to evaluate generation results based on 4,081 diverse text prompts and 25,829 questions across 12 categories. DPG-Bench [26] comprises 1,065 densely descriptive prompts with an average token length of 83.91, presenting more complex scenarios with varied objects and rich adjectives.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3In fact, there is no dropout layer in most open-sourced LLMs (e.g., LLaMA series [14, 65, 66]), and a compromise solution is to introduce additional dropout layers.", + "bbox": [ + 89, + 862, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18501", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/83f0e9e544828bd594bc8e34656f52939070c923468dc348148a1e7d069c4c54.jpg", + "image_caption": [ + "Figure 3. Performance improvement of iterative alignment tuning based on SEED-LLaMA and DreamLLM, across 8 detailed categories of T2I-CompBench++. Iter. 0 denotes the base models without alignment tuning." + ], + "image_footnote": [], + "bbox": [ + 142, + 88, + 433, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 89, + 309, + 326, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, we evaluate alignment performance of our method against T2I generative models and base LMMs on three compositional T2I benchmarks, including T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Key observations are as follows: 1) Although LMMs enable more flexible settings (e.g., in-context learning and interleaved multimodal generation) for image generation, they still underperform compared to specialized T2I models in terms of the basic alignment ability to follow prompts. It demonstrates that current LMMs may ignore the compositional text-image alignment during multimodal pre-training and fine-tuning. 2) Without human annotations or external models, the proposed SILMM method enhances alignment performance across all categories in three benchmarks over the base LMMs, improving both the discrete SEEDLLaMA and the continuous DreamLLM, verifying the effectiveness and the generalization of SILMM. 3) SEEDLLaMA shows greater self-improvement than DreamLLM, possibly due to its weaker baseline alignment and the stability of discrete DPO over continuous KC-DPO induced by a series of simplification, as discussed in Sec. 3.2. And 4) improvements are more challenging in layout, relation, and complex categories than in attribute categories. This difficulty arises partly because the basic generative ability in these categories is weak, making it difficult to obtain high-quality chosen samples. Besides, understanding compositional concepts remains a challenge for LMMs [7, 56].", + "bbox": [ + 91, + 333, + 483, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. In-depth Analysis", + "text_level": 1, + "bbox": [ + 89, + 753, + 264, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To explore the efficacy of SILMM, we conduct extensive ablation studies and hyperparameter analyses. We first investigate the iteration process and data scaling, followed by an in-depth study of key components, including diversity strategies, decompositional self-questioning and answering for self-feedback, and KC-DPO.", + "bbox": [ + 89, + 777, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Iterative Self-Improvement. As shown in Fig. 3, we conduct three iterations of self-improvement and assess", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance changes across eight detailed categories of T2I-CompBench++ [28]. The results show that SILMM achieves effective, consistent, and continuous improvements in text-image alignment, across most compositional categories. Notably, attribute categories (e.g., color, shape, and texture) exhibit the most significant gains, whereas the non-spatial category shows slower improvement. This slower progress may stem from CLIP score [23], which is less sensitive than other metrics. Finally, as the iteration progresses, improvement rates gradually decrease, indicating convergence. More iterative self-improvement experiments results can be found in App. 11.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2614437d62fcf9e7434be36fbdf5f096e32042199a379bd96215b9f0fbb10a37.jpg", + "image_caption": [ + "(a) Num. of Training Prompts" + ], + "image_footnote": [], + "bbox": [ + 524, + 285, + 717, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d440124397b0c90faff956d1051b90b1cf838014c379a499ba7d8182de20a41.jpg", + "image_caption": [ + "(b) Num. of Pairs per Prompt", + "Figure 4. Overall alignment scores of SEED-LLaMA with discrete DPO and DreamLLM with continuous KC-DPO, on T2I-CompBench++ with (a) varying numbers of generated prompts in the training data, and (b) different number of preference pairs sampled from 30 diverse generated images per prompt. $N \\times N$ means we select the top-N and last-N images from 30 generated ones as the chosen and rejected, respectively." + ], + "image_footnote": [], + "bbox": [ + 718, + 287, + 903, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Scale. The proposed SILMM method leverages self-synthesized data for tuning, allowing flexible adjustment of data scale according to practical needs and available computational resources. In Fig. 4, we investigate how data scale affects overall alignment performance (averaged across eight categories in T2I-CompBench++, focusing on two factors: the number of training prompts and the number of preference pairs per prompt. Results in Fig. 4a indicate that both LMMs show consistent improvement as data samples increase, showing the strong scalability of the proposed method. Besides, we generate 30 representations and images per prompt, and then select the top- $N$ and last- $N$ samples to construct $N \\times N$ preference pairs (see Fig. 4b). Notably, the two LMMs perform differently. This may be because the continuous feature space, being larger and denser than the discrete space, requires denser data pairs to stabilize the optimization dynamics.", + "bbox": [ + 511, + 522, + 908, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity Strategies. To synthesize high-quality preference pair data, the diversity strategy is crucial. An effective diversity strategy should maximize the potential of LMMs while ensuring sufficient variation among generated images. To explore various strategies, we compare the proposed $DropDiv$ with three alternatives: Rephrase the original prompt, Explain the original prompt with added imaginative elements, and add Gaussian noises to the learnable", + "bbox": [ + 511, + 779, + 910, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18502", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f3356dda046940924f2ed25fe33ce50f6524c2abf12f4fecc3ef663bfb05e09c.jpg", + "table_caption": [ + "Table 1. Performance comparison and improvement of the proposed method for compositional text-to-image generation on T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Alignment scores are calculated using expert understanding models (e.g., VQA or object detection models) recommended by these benchmarks. Prompt rewriting in Emu3 [68] was not used for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodT2I-CompBench++ [28]DPG-Bench [26]TIFA [27]
AttributeLayoutNon-spatialComplexGlobalEntityAttributeRelationOtherAllAll
Text-to-Image Generative Models
SD-v1.5 [55]38.65---74.6374.2375.3973.4967.8163.1878.40
DALL-E 2 [55]58.63----------
SD-v2 [55]47.3630.5031.2733.8677.6778.1374.9180.7280.6668.09-
SD-v2.1 [55]50.57---------82.00
SDXL [45]52.8835.6231.1932.3783.2782.4380.9186.7680.4174.65-
PixArt-α [8]60.3136.7431.9734.3374.9779.3278.6082.5776.9671.11-
DALL-E 3 [4]70.0941.6330.0337.7390.9789.6188.3990.5889.8383.50-
Large Multimodal Models
SEED-LLaMA [19]19.2020.2928.8621.4665.5955.8761.9662.7759.4647.1266.74
SEED-LLaMA + Ours39.6025.1129.8228.2873.5570.4868.4974.7968.6457.3173.74
%Improvement106.25%23.77%3.33%31.78%12.14%26.15%10.54%19.15%15.44%21.63%10.49%
DreamLLM [13]22.9423.7428.7623.0174.4765.8663.8074.2446.0053.9369.91
DreamLLM + Ours39.9427.6329.0026.4376.2975.9169.2084.4160.0064.2275.38
%Improvement74.15%16.40%0.83%14.86%2.44%15.26%8.46%13.70%30.43%19.08%7.82%
Emu3 [68]44.7932.3030.1531.3284.1980.8182.7587.2350.8074.1981.86
Emu3 + Ours59.7136.0330.5133.9384.1981.5784.5289.0164.8077.4585.11
%Improvement33.30%11.57%1.19%8.33%0.00%0.94%2.14%2.04%27.56%4.39%3.97%
", + "bbox": [ + 148, + 135, + 844, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5410bb3d5febd78c8305bbb114ae59eac73ddd33c5231041c24ff85f280532b.jpg", + "image_caption": [ + "(a) Color, Shape, and Texture" + ], + "image_footnote": [], + "bbox": [ + 151, + 382, + 367, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b0b8c4058c035053301add10f33b5b5c81cfcb9dbda59a6672a7f3a493b9033.jpg", + "image_caption": [ + "(b) Spatial, 3D Spatial, and Numeracy" + ], + "image_footnote": [], + "bbox": [ + 393, + 385, + 607, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16231f133c5649e6ced2d0eee83e643ba68f3addd136d8906b86c958ca1c0b77.jpg", + "image_caption": [ + "(c) Non-spatial and Complex" + ], + "image_footnote": [], + "bbox": [ + 633, + 383, + 854, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/689ba48b80820c28b2f7da355c5d1f74d4e168c8ef89f1e9dedae31f6bd481b2.jpg", + "image_caption": [ + "Figure 5. Comparison of four methods for diverse continuous representation generation, with alignment scores evaluated on the validation set of T2I-CompBench++. For each prompt, DreamLLM generates ten diverse representations and corresponding images.", + "Figure 6. Distribution of alignment scores and variation of maximum scores across different dropout rates (%Drop) of the proposed DropDiv method, evaluated on the T2I-CompBench++ validation set. For each prompt, DreamLLM generates ten diverse representations and corresponding images with DropDiv." + ], + "image_footnote": [], + "bbox": [ + 125, + 585, + 450, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dream Embeddings in DreamLLM [13] to create Noisy DreamEmb. As shown in Fig. 5, four strategies have different perturbation influences on different categories. For example, DropDiv could generate better samples in attribute,", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "non-spatial, and complex categories, but compromise in layout categories. To further examine the effects of Drop-Div, we conduct experiments across different dropout rates as shown in Fig. 6. Results indicate that higher dropout rates enhance diversity, but the alignment quality could be impaired. Therefore, achieving a good diversity-quality tradeoff remains challenging.", + "bbox": [ + 511, + 589, + 906, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decompositional Self-Feedback. We perform ablation studies on question generation and alignment score calculation, as shown in Tab. 2. Compared to two variants Prompt-Q (appending or replacing the period with “?” at the end of each prompt) and Phrase-Q (segmenting each prompt into phrases using NLP tools [24]), Self-Questioning (Self-Q) achieves better alignment performance across most categories, demonstrating the effectiveness of leveraging language processing abilities of LMMs for text-image alignment evaluation. Additionally, we compare the proposed alignment score calculation from Eqn. (1) with two variants: Random Score and Ratio of “yes” (where a higher ratio indicates a higher score). Results show that our method", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18503", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1bdd2c3d7848ccdfa72ff78b154205c1f610e6423e8bfdc0c11ffcd749729cf1.jpg", + "table_caption": [ + "Table 2. Ablation study on T2I-CompBench++ [28] and DPG-Bench [26] examining variations in Question Generation and VQA-based Alignment Score Calculation methods for self-feedback. Prompt-Q adds a “?” or replaces the period with a “?” at the end of each prompt. Phrase-Q involves dividing a prompt into phrases, each followed by a “?”. Self-Q instructs the LMM to generate questions for each prompt using in-context examples. Diff. of Prob. denotes the proposed alignment score calculation approach described in Eqn. (1)." + ], + "table_footnote": [], + "table_body": "
FeedbackT2I-CompBench++DPG-Bench
AttributeLayoutNon-spatialComplexAll
Baseline (DreamLLM [13])
-22.9423.7428.7623.0153.93
Question Generation
Prompt-Q33.6625.9328.1424.6760.13
Phrase-Q34.6324.9128.0125.4160.10
Self-Q34.8525.5128.8225.3160.95
VQA-based Alignment Score Calculation
Random Score23.4124.8128.6722.9553.89
Ratio of “yes”25.3623.5128.7324.0054.68
Diff. of Prob.34.8525.5128.8225.3160.95
", + "bbox": [ + 135, + 218, + 455, + 358 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/49abbbb9b371371b2dd7a7db6e2834d67f5e2f54214fe5d53d31b765cf06c005.jpg", + "table_caption": [ + "Table 3. Ablation study on T2I-CompBench++ [28] to investigate different instantiation of the Kernel Function to calculate the continuous KC-DPO loss function to tune DreamLLM. Aggregation means we aggregate the 2D feature matrix (e.g., $\\mathbf{H}$ ) into 1D along the sequence dimension. Eucl. denotes Euclidean distance." + ], + "table_footnote": [], + "table_body": "
AggregationDistanceAttributeLayoutNon-spatialComplex
Baseline (DreamLLM [13])
--22.9423.7428.7623.01
Supervised Fine-tuning (SFT)
-Eucl.12.250.7516.4111.71
-Cos6.950.2916.7811.48
AvgPoolEucl.23.3123.8928.7623.22
AvgPoolCos23.1224.2028.7923.29
Continuous Kernel-based Direct Preference Optimization
-Eucl.23.6524.3428.8323.08
-Cos23.9724.1128.7723.21
MaxPoolEucl.23.7924.0428.8623.92
MaxPoolCos29.1825.0118.7212.27
AvgPoolEucl.26.7524.7028.9425.12
AvgPoolCos34.8525.5128.8225.31
", + "bbox": [ + 133, + 444, + 436, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieves superior performance by considering the relative confidence between \"yes\" and \"no\".", + "bbox": [ + 89, + 628, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kernel-based Continuous DPO. In Sec. 3.2, we introduce the KC-DPO to fine-tune LMMs with continuous representations. The implementation of kernel functions can be divided into Aggregation and Distance. To assess the impacts of different kernels, we conduct extensive comparison experiments, as shown in Tab. 3. We observe SFT slightly improves the alignment performance, while DPO yields more substantial gains across all metrics. These results show that kernel functions are crucial to KC-DPO, and an optimal choice could greatly enhance the efficiency of preference optimization in continuous feature space. Overall, AvgPool + Cos demonstrates the superior performance improvement.", + "bbox": [ + 89, + 657, + 483, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 848, + 272, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To illustrate the improvements achieved by SILMM, Fig. 8 presents examples generated by SEED-LLaMA, Dream", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0f47ca9a390922442803deee89f7647d4dcc9d48dc36b0f6dd92f942c7afab1.jpg", + "image_caption": [ + "(a) $\\beta$ in Discrete DPO" + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 717, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/595a31543c9a62e10a57d0d2656f10583ab1f115d301086164fcdf8841fb73b4.jpg", + "image_caption": [ + "(b) $\\gamma$ in Continuous KC-DPO" + ], + "image_footnote": [], + "bbox": [ + 718, + 92, + 903, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dc0fccd90873afa413bdd93614ff965e05b5685c3af12456817877ad3fec3d7.jpg", + "image_caption": [ + "Figure 7. Hyperparameter sensitivity on four general categories of T2I-CompBench++, examining (a) $\\beta$ in discrete DPO for SEEDLLaMA, and (b) $\\gamma$ in continuous KC-DPO for DreamLLM.", + "Figure 8. Qualitative results from SEED-LLaMA, DreamLLM, and the proposed SILMM method, on T2I-CompBench++." + ], + "image_footnote": [], + "bbox": [ + 535, + 275, + 870, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLM, and SILMM, in Fig. 8 using prompts from T2I-CompBench++. These results showcase the effectiveness of SILMM across extensive compositional scenarios.", + "bbox": [ + 511, + 662, + 906, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present a self-improvement approach named SILMM to enhance text-image alignment within LMMs, introducing an iterative model-agnostic framework comprising five stages to enable high-quality self-feedback and alignment learning. For continuous LMMs, we propose a dropout-based strategy to diversify image representations and a continuous DPO method, KC-DPO, for optimizing LMMs with preference representation pairs. Extensive experiments validate the effectiveness and superiority of our SILMM framework.", + "bbox": [ + 511, + 747, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18504", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, pages 4447-4455. PMLR, 2024. 4", + "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 15", + "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2, 3", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2(3):8, 2023. 7", + "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICML Workshop on Structured Probabilistic Inference & Generative Modeling, 2023. 2, 3, 4", + "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics, 42(4):1-10, 2023. 1, 3", + "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 4, 6", + "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 7", + "[9] Xiaolin Chen, Xuemeng Song, Liqiang Jing, Shuo Li, Linmei Hu, and Liqiang Nie. Multimodal dialog systems with dual knowledge-enhanced generative pretrained language model. ACM Transactions on Information Systems, 42(2): 1-25, 2023. 17", + "[10] Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al. Ultrafeedback: Boosting language models with scaled ai feedback. In International Conference on Machine Learning, 2024. 3", + "[11] Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, and Jingren Zhou. How abilities in large language models are affected by supervised fine-tuning data composition. arXiv preprint arXiv:2310.05492, 2023. 2", + "[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4", + "[13] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In International Conference on Learning Representations, 2024. 2, 3, 4, 5, 7, 8, 15, 18", + "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5", + "[15] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4", + "[16] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In International Conference on Machine Learning, 2023. 1, 3, 17", + "[17] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 2, 3", + "[18] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 2, 5", + "[19] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. In International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5, 7, 18", + "[20] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024.5, 15", + "[21] Yangyang Guo, Guangzhi Wang, and Mohan Kankanhalli. Pela: Learning parameter-efficient models with low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15699-15709, 2024. 17", + "[22] Marti A. Hearst, Susan T Dumais, Edgar Osuna, John Platt, and Bernhard Scholkopf. Support vector machines. IEEE Intelligent Systems and their applications, 13(4):18-28, 1998. 5", + "[23] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6", + "[24] Matthew Honnibal, Ines Montani, Sofie Van Landeghem," + ], + "bbox": [ + 517, + 93, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18505", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7", + "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 16", + "[26] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 5, 6, 7, 8, 16, 17, 18, 21, 22", + "[27] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 5, 6, 7, 16, 18", + "[28] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 5, 6, 7, 8, 16, 17, 18, 19, 20", + "[29] Natasha Jaques, Shixiang Gu, Dzmitry Bahdanau, José Miguel Hernández-Lobato, Richard E Turner, and Douglas Eck. Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control. In International Conference on Machine Learning, pages 1645-1654. PMLR, 2017. 15", + "[30] Natasha Jaques, Judy Hanwen Shen, Asma Ghandehar-ioun, Craig Ferguson, Agata Lapedriza, Noah Jones, Shixiang Shane Gu, and Rosalind Picard. Human-centric dialog training via offline reinforcement learning. arXiv preprint arXiv:2010.05848, 2020. 15", + "[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2", + "[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Ren Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. In International Conference on Machine Learning, 2024. 3, 15", + "[33] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2", + "[34] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 1", + "[35] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[36] Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, and Mohan Kankanhalli. Attribute-driven disentangled" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "representation learning for multimodal recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 9660-9669. ACM, 2024. 17", + "[37] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3", + "[38] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. In International Conference on Learning Representations, 2024. 3", + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3", + "[40] Yongshuai Liu, Jiaxin Ding, and Xin Liu. Ipo: Interior-point policy optimization under constraints. In Proceedings of the AAAI conference on artificial intelligence, pages 4940-4947, 2020. 4", + "[41] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024. 17", + "[42] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 4", + "[43] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 3, 4, 14, 15", + "[44] Jing-Cheng Pang, Pengyuan Wang, Kaiyuan Li, Xiong-Hui Chen, Jiacheng Xu, Zongzhang Zhang, and Yang Yu. Language model self-improvement by reinforcement learning contemplation. In International Conference on Learning Representations, 2024. 2", + "[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7", + "[46] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2, 3", + "[47] Tianyi Qin, Bo Peng, Jianjun Lei, Jiahui Song, Liying Xu, and Qingming Huang. 3d-immc: Incomplete multi-modal 3d shape clustering via cross mapping and dual adaptive fusion. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 17", + "[48] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18506", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Leigang Qu, Shengqiong Wu, Hao Fei, Liqiang Nie, and TatSeng Chua. Layoutllm-t2i: Eliciting layout guidance from llm for text-to-image generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 643-654, 2023. 1, 3", + "[50] Leigang Qu, Haochuan Li, Tan Wang, Wenjie Wang, Yongqi Li, Liqiang Nie, and Tat-Seng Chua. Unified text-to-image generation and retrieval. arXiv preprint arXiv:2406.05814, 2024.3", + "[51] Leigang Qu, Wenjie Wang, Yongqi Li, Hanwang Zhang, Liqiang Nie, and Tat-Seng Chua. Discriminative probing and tuning for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7434-7444, 2024. 3", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4", + "[53] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 15", + "[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4", + "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 7", + "[56] Sunayana Rane, Alexander Ku, Jason Baldridge, Ian Tenney, Tom Griffiths, and Been Kim. Can generative multimodal models count to ten? In Proceedings of the Annual Meeting of the Cognitive Science Society, 2024. 2, 4, 6", + "[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[58] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 15", + "[59] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 17", + "[60] J Shawe-Taylor. Kernel methods for pattern analysis. Cambridge University Press google schola, 2:181-201, 2004. 5", + "[61] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 15" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3", + "[63] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations, 2023. 5, 15", + "[64] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 2, 3, 5", + "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5", + "[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 15", + "[67] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 3", + "[68] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 5, 7", + "[69] Yufei Wang, Zhanyi Sun, Jesse Zhang, Zhou Xian, Erdem Biyik, David Held, and Zackory Erickson. Rl-vlm-f: Reinforcement learning from vision language foundation model feedback. arXiv preprint arXiv:2402.03681, 2024. 3", + "[70] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv preprint arXiv:2401.15688, 2024. 2, 3", + "[71] Haokun Wen, Xuemeng Song, Xin Yang, Yibing Zhan, and Liqiang Nie. Comprehensive linguistic-visual composition network for image retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1369–1378, 2021. 17", + "[72] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In International Conference on Machine Learning, 2024. 2, 3", + "[73] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Better aligning text-to-image models with human preference. arXiv preprint arXiv:2303.14420, 1(3), 2023. 2", + "[74] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18507", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024.3", + "[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In International Conference on Machine Learning, 2024. 2", + "[76] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2, 3", + "[77] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 2, 3, 4, 15", + "[78] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 2024. 3", + "[79] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "[80] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 15" + ], + "bbox": [ + 91, + 90, + 482, + 513 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18508", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_model.json b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..55fedd12f9d9d801894333a941a323fc0745697d --- /dev/null +++ b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_model.json @@ -0,0 +1,2655 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.13, + 0.853, + 0.175 + ], + "angle": 0, + "content": "SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.903, + 0.223 + ], + "angle": 0, + "content": "Leigang \\(\\mathrm{Qu}^{1}\\), Haochuan \\(\\mathrm{Li}^{1}\\), Wenjie Wang\\(^{2*}\\), Xiang Liu\\(^{1}\\), Juncheng Li\\(^{3}\\), Liqiang Nie\\(^{4}\\), Tat-Seng Chua\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.224, + 0.853, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Singapore, \\(^{2}\\)University of Science and Technology of China, \\(^{3}\\)Zhejiang University, \\(^{4}\\)Harbin Institute of Technology (Shenzhen)" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.26, + 0.875, + 0.275 + ], + "angle": 0, + "content": "leigangqu@gmail.com, haochuan@u.nus.edu, wenjiewang96@gmail.com, liu.xiang@u.nus.edu" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.278, + 0.776, + 0.293 + ], + "angle": 0, + "content": "junchengli@zju.edu.cn, nieliqiang@gmail.com, dcscts@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.485, + 0.754 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) have demonstrated impressive capabilities in multimodal understanding and generation, pushing forward advancements in text-to-image generation. However, achieving accurate text-image alignment for LMMs, particularly in compositional scenarios, remains challenging. Existing approaches, such as layout planning for multi-step generation and learning3 from human feedback or AI feedback, depend heavily on prompt engineering, costly human annotations, and continual upgrading, limiting flexibility and scalability. In this work, we introduce a model-agnostic iterative self-improvement framework (SILMM) that can enable LMMs to provide helpful and scalable self-feedback and optimize text-image alignment via Direct Preference Optimization (DPO). DPO can readily applied to LMMs that use discrete visual tokens as intermediate image representations; while it is less suitable for LMMs with continuous visual features, as obtaining generation probabilities is challenging. To adapt SILMM to LMMs with continuous features, we propose a diversity mechanism to obtain diverse representations and a kernel-based continuous DPO for alignment. Extensive experiments on three compositional text-to-image generation benchmarks validate the effectiveness and superiority of SILMM, showing improvements exceeding \\(30\\%\\) on T2I-CompBench++ and around \\(20\\%\\) on DPG-Bench. The code is available at https://silmm.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.786, + 0.222, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.812, + 0.484, + 0.874 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) are advancing rapidly, surpassing Large Language Models (LLMs) by embracing multimodal capabilities for multimodal content perception, understanding [34, 35, 48], and generation [19, 64]." + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.343, + 0.655, + 0.363 + ], + "angle": 0, + "content": "a green leaf and a yellow butterfly" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.363, + 0.643, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.426, + 0.643, + 0.436 + ], + "angle": 0, + "content": "Attribute Binding" + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.348, + 0.737, + 0.358 + ], + "angle": 0, + "content": "five balloons" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.363, + 0.747, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.426, + 0.744, + 0.437 + ], + "angle": 0, + "content": "Counting Error" + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.343, + 0.857, + 0.362 + ], + "angle": 0, + "content": "a bicycle on the left of a man" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.363, + 0.855, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.426, + 0.863, + 0.437 + ], + "angle": 0, + "content": "Relation Inaccuracy" + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.445, + 0.845, + 0.458 + ], + "angle": 0, + "content": "(a) Text-image misalignment in compositional prompts" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.467, + 0.891, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.634, + 0.862, + 0.647 + ], + "angle": 0, + "content": "(b) Discrete & Continuous LMMs for text-to-image generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.649, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Figure 1. Illustration of (a) text-image misalignment in compositional prompts and (b) comparison of discrete and continuous LMMs for T2I. Given a prompt, discrete LMMs can sample diverse token sequences from categorical distributions, while continuous LMMs can only produce a single deterministic feature vector. Note that the input learnable embeddings are optional for some continuous LMMs [64]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.907, + 0.883 + ], + "angle": 0, + "content": "In particular, LMMs demonstrate promising abilities in interpreting user input prompts for text-to-image generation (T2I) [55, 57], producing vivid and photorealistic images. However, as shown in Fig. 1(a), achieving precise text-image alignment between generated images and complex prompts remains challenging, especially for compositional prompts involving multiple objects, attributes, counting, and complex relationships [6, 16, 49]." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To enhance text-image alignment, existing work falls" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": "into two primary research lines. One line focuses on decomposing the T2I task into multiple stages. For example, some methods perform layout planning before generating the image [17, 37, 75]; while some split the image into sections for multi-step generation via multi-agent collaboration [46, 70]. However, these methods depend on extensive multi-step prompt engineering, which risks error accumulation. The second research line emphasizes learning from human feedback (RLHF [43]) to improve text-image alignment [15, 31, 33, 67, 73], or using AI feedback (RLAIF) from strong evaluation approaches or reward models [3, 77]. Nevertheless, it is labor-intensive and costly to obtain extensive high-quality human feedback, which is also often required to train external reward models [5]. Additionally, as LMMs evolve, the external evaluation approaches and reward models may require continual upgrading [44, 77]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.53 + ], + "angle": 0, + "content": "To address the limitations, we consider utilizing LLMs' inherent discriminative capabilities to self-improve their generation quality for text-image alignment. This offers a pathway for LMMs to evolve for T2I independently, without relying on human or external feedback. To pursue self-improvement, the key steps are: 1) generating diverse images by LMMs based on a given prompt, ensuring the image diversity to facilitate subsequent self-assessment and optimization; 2) using LMMs to self-assess text-image alignment in the generated images, producing alignment scores as self-feedback; and 3) adopting the self-feedback to optimize LMMs to generate superior visual tokens, resulting in images that better align with text prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.482, + 0.562 + ], + "angle": 0, + "content": "However, achieving the above objectives faces significant challenges. In particular:" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.563, + 0.482, + 0.713 + ], + "angle": 0, + "content": "1) As shown in Fig. 1(b), LMMs typically generate intermediate visual representations, i.e., discrete visual tokens or continuous visual features, which are then converted into images by a decoder (e.g., a diffusion model) [19, 64]. For LMMs with discrete visual tokens [19, 68, 76], using existing sampling strategies (e.g., adjusting temperature) in the autoregressive generation process can obtain diverse visual tokens. However, it is non-trivial for LMMs with deterministic continuous visual features, such as Dream-LLM [13], to sample diverse visual representations1." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.714, + 0.482, + 0.789 + ], + "angle": 0, + "content": "2) Compositional prompts require LMMs to inspect object counts, attributes, and complex relationships in the generated images. However, existing LMMs still struggle with compositional cross-modal assessment [7, 56], challenging the generation of faithful self-feedback." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.79, + 0.482, + 0.849 + ], + "angle": 0, + "content": "3) Optimizing LMMs with self-feedback is also intricate. Supervised Fine-Tuning (SFT) [11] and certain RLAIF methods [3, 77] require highly accurate self-feedback. Moreover, another representative method Direct Prefer-" + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.563, + 0.482, + 0.849 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "ence Optimization (DPO) requires modeling generation distributions, from which we need to sample diverse images to construct pairwise training data, which is challenging for LMMs with continuous visual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.56 + ], + "angle": 0, + "content": "To tackle the above challenges, we propose an Self-Improving Large Multimodal Models (SILMM) framework for iterative optimization. As illustrated in Fig. 2, SILMM operates through five steps: 1) Compositional Prompt Generation prompts an LMM to imagine compositional scenarios and generate compositional prompts. 2) Diverse Image Generation. For discrete LMMs\\(^2\\), we follow the sampling decoding strategy commonly used in LLM alignment [43, 53]. For continuous LMMs [13, 64, 72], we propose a diversification strategy named \\(DropDiv\\), inspired by Monte Carlo (MC) Dropout [18], to perform dropout on the MLP layers of LMMs for diverse visual features, producing diverse images. 3) Decompositional Self-Questioning. To reduce the difficulty of compositional cross-modal assessment, LMMs can decompose a compositional prompt into atomic concepts and relations and generate questions for multi-step assessment. 4) VQA-based Self-Feedback. For each image generated in Step 2, LMMs can use the decomposed questions to assess text-image alignment, and then aggregate the results to obtain reasonable self-feedback. 5) Learning from Self-Feedback. For discrete LMMs, we directly apply DPO based on pairwise samples from Step 2. As to continuous LMMs, we propose Kernel-based Continuous DPO (KC-DPO), inducing a quadruplet objective with kernel functions for pairwise distance regulation over continuous visual features. The above five steps can iteratively repeat until self-improvement performance converges." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.562, + 0.864, + 0.576 + ], + "angle": 0, + "content": "In summary, our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.578, + 0.905, + 0.638 + ], + "angle": 0, + "content": "- To our knowledge, we are the first to focus on the task of LMMs' self-improvement for T2I. We propose a model-agnostic self-improvement framework to enable LMMs to achieve high-quality self-feedback and learning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.639, + 0.905, + 0.697 + ], + "angle": 0, + "content": "- For continuous LMMs, we introduce a dropout-based strategy to diversify image representations, along with a continuous DPO approach, i.e., KC-DPO, to optimize LMMs with preference representation pairs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.7, + 0.905, + 0.743 + ], + "angle": 0, + "content": "- We conduct extensive experiments on three compositional T2I benchmarks, demonstrating the superiority of SILMM, e.g., \\(30\\%\\) improvements on T2I-CompBench++." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.578, + 0.905, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.76, + 0.655, + 0.775 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.905, + 0.862 + ], + "angle": 0, + "content": "Compositional Text-to-Image Generation. Diffusion models [55, 57] have marked a significant advancement in T2I generation due to their stability and scalability. However, they still struggle with text-image alignment, such as attribute binding, counting error, and relation confu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Sampling diverse images at the decoder stage is inapplicable, as it can only optimize the decoder yet we aim to optimize LMMs to generate superior visual representations for text-image alignment in this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2For simplicity, we denote LLMs outputting discrete and continuous visual representations as discrete and continuous LMMs, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18498" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.101, + 0.899, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.465, + 0.908, + 0.537 + ], + "angle": 0, + "content": "Figure 2. Schematic illustration of SILMM, comprising five steps: 1) LMMs generate compositional prompts by sampling based on provided instructions. 2) Diverse representations and images are generated using either discrete nucleus sampling or the proposed continuous DivDrop. 3) LMMs divide each compositional prompt into semantic units and generate questions for each unit. 4) VQA is conducted to answer these questions, with the answers and likelihoods aggregated into alignment scores as self-feedback. 5) For alignment tuning, DPO is applied for discrete LMMs, while the proposed KC-DPO is used for continuous LMMs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.482, + 0.687 + ], + "angle": 0, + "content": "sion [16, 51]. To enhance compositional T2I, some approaches intervene in language structures [16] or cross-attention mechanisms [6]. Other methods [17, 37, 38, 49] incorporate layout planning by LLMs or use multi-agent collaboration [46, 70]. Inspired by alignment successes in LLMs, recent work [5, 15, 67] applies RLHF [43] to optimize diffusion models. Despite the progress, they rely on inductive biases, extensive prompt engineering, or labor-intensive annotations, limiting flexibility and scalability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.688, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Large Multimodal Models. The pioneering LMMs [39, 79] integrate a visual encoder, e.g., CLIP [52], with LLMs as the foundation, showing impressive multimodal understanding capabilities. To extend LMMs to visual generation, recent approaches align diffusion models [13, 19, 72] with LLMs or train a single transformer [62, 68, 74, 76]. According to the form of output visual features, they can be divided into discrete visual tokenization methods [19, 50, 62, 68] and continuous visual representation methods [13, 64, 72]. While LLM integration enhances language understanding and supports flexible applications (e.g., interleaved multimodal generation [64]), compositional T2I in the context of LMMs remains underexplored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Learning from AI Feedback. The high cost of collect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.702 + ], + "angle": 0, + "content": "ing human preference has spurred research into RLAIF [3]. Benefiting from the convenience and scalability, there have been a series of studies adopting RLAIF to tackle a range of NLP tasks [10, 32, 78] and vision-language understanding [69, 77]. Despite the thrilling success, they only focus on text generation, overlooking the potential of RLAIF in other modalities. In contrast, we explore self-improving LMMs by activating multimodal understanding abilities for T2I. Particularly, we propose continuous strategies meticulously tailored to continuous visual features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.718, + 0.65, + 0.736 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.82 + ], + "angle": 0, + "content": "In this section, we elaborate on the proposed method, including the SILMM framework with five steps and the iteration strategy (Sec. 3.1), as illustrated in Fig. 2. Afterward, we introduce the continuous KC-DPO applied to LMMs with continuous visual features in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.874, + 0.849 + ], + "angle": 0, + "content": "3.1. Self-Improving Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Step 1: Compositional Prompt Generation. We first divide compositional scenarios into four categories: Attribute (color, shape, texture), Layout (counting, spatial relation)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "Semantic Relation, and Complex Composition. Complex composition includes any possible composition of the first three. For attribute and layout, we prompt the LMM to separately generate common objects, attributes, numbers, and spatial relations, and then use templates to compose these concepts. For semantic relation and complex composition, we adopt in-context learning [12] to generate prompts. More details can be found in App. 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Step 2: Diverse Representation and Image generation. The purpose of this step is to sample diverse intermediate visual representations from the LLM backbone \\(\\pi\\) of an LMM, given a text prompt \\(x\\), which would be decoded into images with different qualities. These representations are denoted as \\(\\mathcal{Z} = \\{z_i,\\dots,z_M\\}\\), where \\(z_{i}\\sim \\pi (z|x)\\). For discrete LMMs [19], \\(z_{i}\\) is a discrete visual sequence. We follow the common practice [43, 53] in language generation to obtain \\(\\mathcal{Z}\\), by sampling with different random seeds during auto-regressive decoding. For continuous LMMs [13], the LLM can only output a fixed continuous visual feature, without diversity. To tackle this issue, we propose DropDiv. First, we insert the dropout operations in the last few MLP layers of LLMs, which introduces randomness and enables LLMs for sampling. During inference, we activate these dropout operations to output diverse representations by sampling: \\(z_{i}\\sim \\pi^{\\prime}(z|x)\\), where \\(z_{i}\\) denotes a continuous visual feature and \\(\\pi^\\prime\\) represents the LLM with activated dropout operations. Afterward, these diverse visual representations \\(\\mathcal{Z}\\) are decoded into images as \\(\\mathcal{Y} = \\{y_1,\\dots,y_M\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Discussion. Unlike prior work [5, 15] focused on tuning diffusion models, our approach resorts to LLM backbones in LLMs to control image decoders (e.g., diffusion models) for better text-image alignment, centering on LLM backbone optimization. Our approach offers three key advantages: 1) LLMs demonstrate superior proficiency in prompt comprehension over text encoders [52, 54] commonly employed in diffusion models. Tuning LLM backbones may unlock their enormous potential for compositional T2I, especially in complex scenarios. 2) Tuning diffusion models is often constrained by efficiency challenges inherent to iterative likelihood estimation, whereas there have been well-established technologies [1, 40, 53] for LLM alignment. 3) Our method is orthogonal to existing methods to tune diffusion models, combining them may get further gains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Step 3: Decompositional Self-Questioning. To provide helpful feedback to the generated images, the LMM should first accurately assess text-image alignment, which requires strong compositional reasoning abilities. However, current advanced LMMs still suffer from compositional reasoning [42], such as spatial relation understanding [7] and counting [56]. To improve compositional reasoning, we introduce a divide-and-conquer strategy [77] for self-questioning. Specifically, the LMM first divides the given prompt \\( x \\) into atomic concepts (e.g., \"a white harp\") and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "relations (e.g., \"a pancake is on the left of a pasta\"), and then generates questions \\(\\mathcal{Q} = \\{q_1, \\dots, q_N\\}\\), each \\(q_i\\) corresponding to a concept or relation. For simplicity, the generated questions are constrained to be yes/no questions (e.g., \"Is there a while harp?\", \"Is the pancake on the left of the pasta?\"). Refer to App. 7 for more details on prompt templates of self-questioning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Step 4: VQA-based Self-Feedback. Taking a generated image \\( y \\in \\mathcal{V} \\) and all the questions \\( \\mathcal{Q} \\) as input, the LMM conducts the VQA task, and the average difference between the probabilities of answering \"yes\" and \"no\" serves as the text-image alignment score:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.284, + 0.905, + 0.327 + ], + "angle": 0, + "content": "\\[\ns (x, y) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ p \\left(\" y e s\" \\mid y, q _ {i}\\right) - p \\left(\" n o\" \\mid y, q _ {i}\\right) \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.428 + ], + "angle": 0, + "content": "Here we adopt the vision-language understanding abilities of LMMs via VQA to provide feedback to the images generated by themselves, thus this step is named VQA-based self-feedback. We carry out this step through all the sampled images prompted by \\( x \\) and get all the scores \\( S = \\{ s(x,y_j) | y_j \\in \\mathcal{V} \\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.429, + 0.905, + 0.518 + ], + "angle": 0, + "content": "Step 5: Learning from Self-Feedback. Based on the selffeedback alignment scores, we sample representation pairs \\((z_w,z_l)\\) from \\(\\mathcal{Q}\\), where \\(z_{w}\\) and \\(z_{l}\\) denote the chosen and the rejected representations and their corresponding decoded images should satisfy \\(s(x,y_w) > s(x,y_l)\\). With the preference data, we optimize the LLM backbone with DPO [53]:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.534, + 0.689, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P O}} = - \\mathbb {E} _ {(x, z _ {w}, z _ {l}) \\sim \\mathcal {D}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.553, + 0.905, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(z _ {w} | x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(z _ {w} | x\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(z _ {l} | x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(z _ {l} | x\\right)}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.905, + 0.656 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) denotes the training set, and \\(\\pi_{\\theta}\\) and \\(\\pi_{\\mathrm{ref}}\\) represent the policy and reference models, respectively. \\(\\sigma\\) is the sigmoid function, and \\(\\beta\\) is a hyperparameter controlling the deviation from the reference model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Iterative Self-Improvement. After learning from self-feedback, the updated LMM becomes more likely to generate preferred representations that are decoded into images better aligned with the prompt. This improvement in overall text-image alignment motivates us to iterate the above five steps with the updated LMM as the new reference model. The iteration mechanism continues until the alignment performance converges. As the process is independent of human annotations and external models, it is cost-effective and scalable. More importantly, it showcases the potential for self-improvement in LMMs by harmonizing their understanding and generation capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.885, + 0.865 + ], + "angle": 0, + "content": "3.2. Continuous Direct Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "At the step of learning from self-feedback, LMMs are optimized using the DPO objective as shown in Eqn. (2)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18500" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "The difference between discrete and continuous LMMs in this learning process lies in the calculation of the likelihood \\(\\pi (z|x)\\). For discrete LMMs, \\(\\pi (z|x)\\) can be straightforwardly obtained by the softmax categorical distribution. However, for continuous LMMs with unknown distribution modeling, calculating \\(\\pi (z|x)\\) is intractable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Predictive Distribution with MC Dropout. MC Dropout [18] enables predictive distribution estimation via Monte Carlo simulation to calculate \\(\\pi(z|x)\\). Specifically, the dropout layers3 in an LMM are activated during inference and the LMM performs forward propagation multiple times to get multiple outputs. Assuming a Gaussian distribution, we can estimate its parameters and calculate the likelihood \\(\\pi(z|x)\\) based on these outputs. However, such multi-forward estimation imposes a significant computational burden during training, making this approach insufficient and impractical." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Simplified Kernel-based Continuous DPO. Inspired by MC Dropout and motivated by its insufficiency issue, we propose a simplified method to achieve continuous DPO. Concretely, the intermediate representation \\(z\\) often performs as a feature matrix \\(\\pmb{H} \\in \\mathbb{R}^{L \\times D}\\) where \\(L\\) and \\(D\\) denote the sequence length and dimension. \\(\\pmb{H}\\) can be attained by a Q-Former [13, 20] or from the last layer of the LMM in an autoregressive way [63, 64]. To estimate \\(\\pi(\\pmb{H} | x)\\), we first make a decomposition as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.493, + 0.483, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\pi (\\boldsymbol {H} | x) = \\prod_ {i = 1} ^ {L} \\pi \\left(\\boldsymbol {h} _ {i} \\mid \\boldsymbol {H} _ {< i}, x\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.575 + ], + "angle": 0, + "content": "where \\( \\pmb{h}_i \\in \\mathbb{R}^D \\) denotes the \\( i \\)-th feature vector. Based on the Gaussian assumption, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.58, + 0.483, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\pi \\left(\\boldsymbol {h} _ {i} \\mid \\boldsymbol {H} _ {< i}, x\\right) = \\frac {\\exp \\left[ - \\frac {1}{2} \\left(\\boldsymbol {h} _ {i} - \\boldsymbol {\\mu} _ {i}\\right) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\boldsymbol {h} _ {i} - \\boldsymbol {\\mu} _ {i}\\right) \\right]}{\\sqrt {(2 \\pi) ^ {D} \\left| \\boldsymbol {\\Sigma} _ {i} \\right|}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.741 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu}_i\\) and \\(\\pmb{\\Sigma}_i\\) denote the mean vector and the covariance matrix, respectively. Furthermore, we further simplify and approximate this formula: 1) the mean vector is estimated by the direct output of the continuous LMM, i.e., \\(\\pmb{\\mu}_i\\approx \\mathrm{LMM}(\\boldsymbol {x})[i]\\), and 2) the Gaussian distribution is isotropic and all dimensions share the same variance value \\(\\bar{\\sigma}\\), i.e., \\(\\pmb{\\Sigma}_i\\approx \\mathrm{diag}(\\sigma_1,\\dots,\\sigma_{\\mathrm{D}})\\) and \\(\\sigma_{1} = \\ldots = \\sigma_{D} = \\bar{\\sigma}\\), and \\(\\bar{\\sigma}\\) can be learnable or viewed as a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.784 + ], + "angle": 0, + "content": "We compute the simplified likelihood with Eqn. (4), obtain the joint one with Eqn. (3), and finally derive the continuous DPO based on Eqn. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.79, + 0.483, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} \\mathcal {L} _ {\\mathrm {C - D P O}} = & - \\mathbb {E} _ {(x, \\boldsymbol {H} _ {w}, \\boldsymbol {H} _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\frac {\\beta}{2 \\bar {\\sigma} ^ {2}} (- \\| \\boldsymbol {H} - \\boldsymbol {H} _ {w} \\| _ {F} ^ {2} + \\| \\boldsymbol {H} _ {r} - \\boldsymbol {H} _ {w} \\| _ {F} ^ {2})\\right) \\right], \\end{array} \\tag {5}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.902 + ], + "angle": 0, + "content": "3In fact, there is no dropout layer in most open-sourced LLMs (e.g., LLaMA series [14, 65, 66]), and a compromise solution is to introduce additional dropout layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.26 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _F\\) denotes the Frobenius norm, and \\(H\\) and \\(H_{r}\\) represent the continuous feature matrices from the policy and reference LMMs, respectively. \\(H_{w}\\) and \\(H_{l}\\) refer to the chosen and rejected feature matrices, respectively. Compared with the MC dropout method, this objective only requires one forward pass, which is more efficient. We relegate more details of the derivation to App. 8. From Eqn. (5), we can see that this objective aims to adjust the relative distances within the quadruple \\((\\pmb {H},\\pmb {H}_r,\\pmb {H}_w,\\pmb {H}_l)\\) and the distance metric is the Euclidean distance between two matrices. To further improve the flexibility, we generalize the continuous DPO objective to," + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.271, + 0.905, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {K C - D P O}} = - \\mathbb {E} _ {(x, \\boldsymbol {H} _ {w}, \\boldsymbol {H} _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\gamma (- k (\\boldsymbol {H}, \\boldsymbol {H} _ {w}) \\right. \\right. \\\\ \\left. \\left. + k \\left(\\boldsymbol {H} _ {r}, \\boldsymbol {H} _ {w}\\right) + k \\left(\\boldsymbol {H}, \\boldsymbol {H} _ {l}\\right) - k \\left(\\boldsymbol {H} _ {r}, \\boldsymbol {H} _ {l}\\right)\\right)\\right) \\Bigg ], \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.905, + 0.457 + ], + "angle": 0, + "content": "where \\(\\gamma = \\frac{\\beta}{2\\bar{\\sigma}^2}\\) controls the degree of adherence to the reference model, \\(k(\\cdot ,\\cdot)\\) denotes a generalized distance measurement function. Considering it is similar to kernel methods [22, 60], we name the objective Kernel-based Continuous DPO (KC-DPO). In the following experiments section, we will discuss different distance functions and their influences on alignment performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.473, + 0.646, + 0.49 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.704, + 0.516 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.905, + 0.598 + ], + "angle": 0, + "content": "Base Model Settings. We implement our method on DreamLLM (continuous LMM) [13] and SEED-LLaMA (discrete LMM) [19] for all experiments. We also apply our method to Emu-3 [68], the recent state-of-the-art discrete LMM. Details on DPO training are provided in App. 9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.689 + ], + "angle": 0, + "content": "Datasets. We curated a dataset of 16,000 prompts across four categories using LMM. In each DPO training iteration, images generated by the model in the previous iteration served as the training data for next DPO iteration, allowing for iterative self-improvement. Details on data creation are provided in App. 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "**Benchmarks.** We evaluate our method on three text-to-image alignment benchmarks and follow their default settings. T2I-CompBench++ [28] consists of 8,000 compositional text prompts organized into 4 main categories: attribute, layout, non-spatial, and complex compositions, further divided into 8 subcategories, including color, binding, binding, 2D/3D-spatial relationships, non-spatial relationships, numeracy, and complex compositions. TIFA [27] uses pre-generated question-answer pairs and a VQA model to evaluate generation results based on 4,081 diverse text prompts and 25,829 questions across 12 categories. DPG-Bench [26] comprises 1,065 densely descriptive prompts with an average token length of 83.91, presenting more complex scenarios with varied objects and rich adjectives." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.143, + 0.089, + 0.434, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.486, + 0.3 + ], + "angle": 0, + "content": "Figure 3. Performance improvement of iterative alignment tuning based on SEED-LLaMA and DreamLLM, across 8 detailed categories of T2I-CompBench++. Iter. 0 denotes the base models without alignment tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.31, + 0.328, + 0.327 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.485, + 0.742 + ], + "angle": 0, + "content": "As shown in Tab. 1, we evaluate alignment performance of our method against T2I generative models and base LMMs on three compositional T2I benchmarks, including T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Key observations are as follows: 1) Although LMMs enable more flexible settings (e.g., in-context learning and interleaved multimodal generation) for image generation, they still underperform compared to specialized T2I models in terms of the basic alignment ability to follow prompts. It demonstrates that current LMMs may ignore the compositional text-image alignment during multimodal pre-training and fine-tuning. 2) Without human annotations or external models, the proposed SILMM method enhances alignment performance across all categories in three benchmarks over the base LMMs, improving both the discrete SEEDLLaMA and the continuous DreamLLM, verifying the effectiveness and the generalization of SILMM. 3) SEEDLLaMA shows greater self-improvement than DreamLLM, possibly due to its weaker baseline alignment and the stability of discrete DPO over continuous KC-DPO induced by a series of simplification, as discussed in Sec. 3.2. And 4) improvements are more challenging in layout, relation, and complex categories than in attribute categories. This difficulty arises partly because the basic generative ability in these categories is weak, making it difficult to obtain high-quality chosen samples. Besides, understanding compositional concepts remains a challenge for LMMs [7, 56]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.265, + 0.772 + ], + "angle": 0, + "content": "4.3. In-depth Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To explore the efficacy of SILMM, we conduct extensive ablation studies and hyperparameter analyses. We first investigate the iteration process and data scaling, followed by an in-depth study of key components, including diversity strategies, decompositional self-questioning and answering for self-feedback, and KC-DPO." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Iterative Self-Improvement. As shown in Fig. 3, we conduct three iterations of self-improvement and assess" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "performance changes across eight detailed categories of T2I-CompBench++ [28]. The results show that SILMM achieves effective, consistent, and continuous improvements in text-image alignment, across most compositional categories. Notably, attribute categories (e.g., color, shape, and texture) exhibit the most significant gains, whereas the non-spatial category shows slower improvement. This slower progress may stem from CLIP score [23], which is less sensitive than other metrics. Finally, as the iteration progresses, improvement rates gradually decrease, indicating convergence. More iterative self-improvement experiments results can be found in App. 11." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.286, + 0.718, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.406, + 0.698, + 0.418 + ], + "angle": 0, + "content": "(a) Num. of Training Prompts" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.288, + 0.904, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.405, + 0.887, + 0.418 + ], + "angle": 0, + "content": "(b) Num. of Pairs per Prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 4. Overall alignment scores of SEED-LLaMA with discrete DPO and DreamLLM with continuous KC-DPO, on T2I-CompBench++ with (a) varying numbers of generated prompts in the training data, and (b) different number of preference pairs sampled from 30 diverse generated images per prompt. \\(N \\times N\\) means we select the top-N and last-N images from 30 generated ones as the chosen and rejected, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Data Scale. The proposed SILMM method leverages self-synthesized data for tuning, allowing flexible adjustment of data scale according to practical needs and available computational resources. In Fig. 4, we investigate how data scale affects overall alignment performance (averaged across eight categories in T2I-CompBench++, focusing on two factors: the number of training prompts and the number of preference pairs per prompt. Results in Fig. 4a indicate that both LMMs show consistent improvement as data samples increase, showing the strong scalability of the proposed method. Besides, we generate 30 representations and images per prompt, and then select the top-\\(N\\) and last-\\(N\\) samples to construct \\(N \\times N\\) preference pairs (see Fig. 4b). Notably, the two LMMs perform differently. This may be because the continuous feature space, being larger and denser than the discrete space, requires denser data pairs to stabilize the optimization dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Diversity Strategies. To synthesize high-quality preference pair data, the diversity strategy is crucial. An effective diversity strategy should maximize the potential of LMMs while ensuring sufficient variation among generated images. To explore various strategies, we compare the proposed \\(DropDiv\\) with three alternatives: Rephrase the original prompt, Explain the original prompt with added imaginative elements, and add Gaussian noises to the learnable" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18502" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.132 + ], + "angle": 0, + "content": "Table 1. Performance comparison and improvement of the proposed method for compositional text-to-image generation on T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Alignment scores are calculated using expert understanding models (e.g., VQA or object detection models) recommended by these benchmarks. Prompt rewriting in Emu3 [68] was not used for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.136, + 0.846, + 0.381 + ], + "angle": 0, + "content": "
MethodT2I-CompBench++ [28]DPG-Bench [26]TIFA [27]
AttributeLayoutNon-spatialComplexGlobalEntityAttributeRelationOtherAllAll
Text-to-Image Generative Models
SD-v1.5 [55]38.65---74.6374.2375.3973.4967.8163.1878.40
DALL-E 2 [55]58.63----------
SD-v2 [55]47.3630.5031.2733.8677.6778.1374.9180.7280.6668.09-
SD-v2.1 [55]50.57---------82.00
SDXL [45]52.8835.6231.1932.3783.2782.4380.9186.7680.4174.65-
PixArt-α [8]60.3136.7431.9734.3374.9779.3278.6082.5776.9671.11-
DALL-E 3 [4]70.0941.6330.0337.7390.9789.6188.3990.5889.8383.50-
Large Multimodal Models
SEED-LLaMA [19]19.2020.2928.8621.4665.5955.8761.9662.7759.4647.1266.74
SEED-LLaMA + Ours39.6025.1129.8228.2873.5570.4868.4974.7968.6457.3173.74
%Improvement106.25%23.77%3.33%31.78%12.14%26.15%10.54%19.15%15.44%21.63%10.49%
DreamLLM [13]22.9423.7428.7623.0174.4765.8663.8074.2446.0053.9369.91
DreamLLM + Ours39.9427.6329.0026.4376.2975.9169.2084.4160.0064.2275.38
%Improvement74.15%16.40%0.83%14.86%2.44%15.26%8.46%13.70%30.43%19.08%7.82%
Emu3 [68]44.7932.3030.1531.3284.1980.8182.7587.2350.8074.1981.86
Emu3 + Ours59.7136.0330.5133.9384.1981.5784.5289.0164.8077.4585.11
%Improvement33.30%11.57%1.19%8.33%0.00%0.94%2.14%2.04%27.56%4.39%3.97%
" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.383, + 0.368, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.527, + 0.334, + 0.54 + ], + "angle": 0, + "content": "(a) Color, Shape, and Texture" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.386, + 0.609, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.527, + 0.597, + 0.541 + ], + "angle": 0, + "content": "(b) Spatial, 3D Spatial, and Numeracy" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.385, + 0.855, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.527, + 0.817, + 0.54 + ], + "angle": 0, + "content": "(c) Non-spatial and Complex" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.546, + 0.908, + 0.575 + ], + "angle": 0, + "content": "Figure 5. Comparison of four methods for diverse continuous representation generation, with alignment scores evaluated on the validation set of T2I-CompBench++. For each prompt, DreamLLM generates ten diverse representations and corresponding images." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.587, + 0.452, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Figure 6. Distribution of alignment scores and variation of maximum scores across different dropout rates (%Drop) of the proposed DropDiv method, evaluated on the T2I-CompBench++ validation set. For each prompt, DreamLLM generates ten diverse representations and corresponding images with DropDiv." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dream Embeddings in DreamLLM [13] to create Noisy DreamEmb. As shown in Fig. 5, four strategies have different perturbation influences on different categories. For example, DropDiv could generate better samples in attribute," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.696 + ], + "angle": 0, + "content": "non-spatial, and complex categories, but compromise in layout categories. To further examine the effects of Drop-Div, we conduct experiments across different dropout rates as shown in Fig. 6. Results indicate that higher dropout rates enhance diversity, but the alignment quality could be impaired. Therefore, achieving a good diversity-quality tradeoff remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Decompositional Self-Feedback. We perform ablation studies on question generation and alignment score calculation, as shown in Tab. 2. Compared to two variants Prompt-Q (appending or replacing the period with “?” at the end of each prompt) and Phrase-Q (segmenting each prompt into phrases using NLP tools [24]), Self-Questioning (Self-Q) achieves better alignment performance across most categories, demonstrating the effectiveness of leveraging language processing abilities of LMMs for text-image alignment evaluation. Additionally, we compare the proposed alignment score calculation from Eqn. (1) with two variants: Random Score and Ratio of “yes” (where a higher ratio indicates a higher score). Results show that our method" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18503" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.215 + ], + "angle": 0, + "content": "Table 2. Ablation study on T2I-CompBench++ [28] and DPG-Bench [26] examining variations in Question Generation and VQA-based Alignment Score Calculation methods for self-feedback. Prompt-Q adds a “?” or replaces the period with a “?” at the end of each prompt. Phrase-Q involves dividing a prompt into phrases, each followed by a “?”. Self-Q instructs the LMM to generate questions for each prompt using in-context examples. Diff. of Prob. denotes the proposed alignment score calculation approach described in Eqn. (1)." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.219, + 0.457, + 0.359 + ], + "angle": 0, + "content": "
FeedbackT2I-CompBench++DPG-Bench
AttributeLayoutNon-spatialComplexAll
Baseline (DreamLLM [13])
-22.9423.7428.7623.0153.93
Question Generation
Prompt-Q33.6625.9328.1424.6760.13
Phrase-Q34.6324.9128.0125.4160.10
Self-Q34.8525.5128.8225.3160.95
VQA-based Alignment Score Calculation
Random Score23.4124.8128.6722.9553.89
Ratio of “yes”25.3623.5128.7324.0054.68
Diff. of Prob.34.8525.5128.8225.3160.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Table 3. Ablation study on T2I-CompBench++ [28] to investigate different instantiation of the Kernel Function to calculate the continuous KC-DPO loss function to tune DreamLLM. Aggregation means we aggregate the 2D feature matrix (e.g., \\( \\mathbf{H} \\)) into 1D along the sequence dimension. Eucl. denotes Euclidean distance." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.445, + 0.437, + 0.622 + ], + "angle": 0, + "content": "
AggregationDistanceAttributeLayoutNon-spatialComplex
Baseline (DreamLLM [13])
--22.9423.7428.7623.01
Supervised Fine-tuning (SFT)
-Eucl.12.250.7516.4111.71
-Cos6.950.2916.7811.48
AvgPoolEucl.23.3123.8928.7623.22
AvgPoolCos23.1224.2028.7923.29
Continuous Kernel-based Direct Preference Optimization
-Eucl.23.6524.3428.8323.08
-Cos23.9724.1128.7723.21
MaxPoolEucl.23.7924.0428.8623.92
MaxPoolCos29.1825.0118.7212.27
AvgPoolEucl.26.7524.7028.9425.12
AvgPoolCos34.8525.5128.8225.31
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.658 + ], + "angle": 0, + "content": "achieves superior performance by considering the relative confidence between \"yes\" and \"no\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Kernel-based Continuous DPO. In Sec. 3.2, we introduce the KC-DPO to fine-tune LMMs with continuous representations. The implementation of kernel functions can be divided into Aggregation and Distance. To assess the impacts of different kernels, we conduct extensive comparison experiments, as shown in Tab. 3. We observe SFT slightly improves the alignment performance, while DPO yields more substantial gains across all metrics. These results show that kernel functions are crucial to KC-DPO, and an optimal choice could greatly enhance the efficiency of preference optimization in continuous feature space. Overall, AvgPool + Cos demonstrates the superior performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.274, + 0.865 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To illustrate the improvements achieved by SILMM, Fig. 8 presents examples generated by SEED-LLaMA, Dream" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.718, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.212, + 0.678, + 0.224 + ], + "angle": 0, + "content": "(a) \\(\\beta\\) in Discrete DPO" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.093, + 0.905, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.212, + 0.888, + 0.224 + ], + "angle": 0, + "content": "(b) \\(\\gamma\\) in Continuous KC-DPO" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.228, + 0.907, + 0.27 + ], + "angle": 0, + "content": "Figure 7. Hyperparameter sensitivity on four general categories of T2I-CompBench++, examining (a) \\(\\beta\\) in discrete DPO for SEEDLLaMA, and (b) \\(\\gamma\\) in continuous KC-DPO for DreamLLM." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.276, + 0.871, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.625, + 0.907, + 0.654 + ], + "angle": 0, + "content": "Figure 8. Qualitative results from SEED-LLaMA, DreamLLM, and the proposed SILMM method, on T2I-CompBench++." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.71 + ], + "angle": 0, + "content": "LLM, and SILMM, in Fig. 8 using prompts from T2I-CompBench++. These results showcase the effectiveness of SILMM across extensive compositional scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.888 + ], + "angle": 0, + "content": "In this work, we present a self-improvement approach named SILMM to enhance text-image alignment within LMMs, introducing an iterative model-agnostic framework comprising five stages to enable high-quality self-feedback and alignment learning. For continuous LMMs, we propose a dropout-based strategy to diversify image representations and a continuous DPO method, KC-DPO, for optimizing LMMs with preference representation pairs. Extensive experiments validate the effectiveness and superiority of our SILMM framework." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18504" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, pages 4447-4455. PMLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2(3):8, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICML Workshop on Structured Probabilistic Inference & Generative Modeling, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics, 42(4):1-10, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[9] Xiaolin Chen, Xuemeng Song, Liqiang Jing, Shuo Li, Linmei Hu, and Liqiang Nie. Multimodal dialog systems with dual knowledge-enhanced generative pretrained language model. ACM Transactions on Information Systems, 42(2): 1-25, 2023. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[10] Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al. Ultrafeedback: Boosting language models with scaled ai feedback. In International Conference on Machine Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, and Jingren Zhou. How abilities in large language models are affected by supervised fine-tuning data composition. arXiv preprint arXiv:2310.05492, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.906, + 0.119 + ], + "angle": 0, + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[13] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In International Conference on Learning Representations, 2024. 2, 3, 4, 5, 7, 8, 15, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[15] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[16] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In International Conference on Machine Learning, 2023. 1, 3, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[17] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[18] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.615 + ], + "angle": 0, + "content": "[19] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. In International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5, 7, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.906, + 0.685 + ], + "angle": 0, + "content": "[20] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024.5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[21] Yangyang Guo, Guangzhi Wang, and Mohan Kankanhalli. Pela: Learning parameter-efficient models with low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15699-15709, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[22] Marti A. Hearst, Susan T Dumais, Edgar Osuna, John Platt, and Bernhard Scholkopf. Support vector machines. IEEE Intelligent Systems and their applications, 13(4):18-28, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.906, + 0.885 + ], + "angle": 0, + "content": "[23] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.888, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Matthew Honnibal, Ines Montani, Sofie Van Landeghem," + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18505" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.178, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 5, 6, 7, 8, 16, 17, 18, 21, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[27] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 5, 6, 7, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[28] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 5, 6, 7, 8, 16, 17, 18, 19, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[29] Natasha Jaques, Shixiang Gu, Dzmitry Bahdanau, José Miguel Hernández-Lobato, Richard E Turner, and Douglas Eck. Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control. In International Conference on Machine Learning, pages 1645-1654. PMLR, 2017. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[30] Natasha Jaques, Judy Hanwen Shen, Asma Ghandehar-ioun, Craig Ferguson, Agata Lapedriza, Noah Jones, Shixiang Shane Gu, and Rosalind Picard. Human-centric dialog training via offline reinforcement learning. arXiv preprint arXiv:2010.05848, 2020. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Ren Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. In International Conference on Machine Learning, 2024. 3, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[33] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[34] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[35] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, and Mohan Kankanhalli. Attribute-driven disentangled" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "representation learning for multimodal recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 9660-9669. ACM, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[37] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[38] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. In International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[40] Yongshuai Liu, Jiaxin Ding, and Xin Liu. Ipo: Interior-point policy optimization under constraints. In Proceedings of the AAAI conference on artificial intelligence, pages 4940-4947, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[41] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[42] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[43] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 3, 4, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[44] Jing-Cheng Pang, Pengyuan Wang, Kaiyuan Li, Xiong-Hui Chen, Jiacheng Xu, Zongzhang Zhang, and Yang Yu. Language model self-improvement by reinforcement learning contemplation. In International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[46] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[47] Tianyi Qin, Bo Peng, Jianjun Lei, Jiahui Song, Liying Xu, and Qingming Huang. 3d-immc: Incomplete multi-modal 3d shape clustering via cross mapping and dual adaptive fusion. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[49] Leigang Qu, Shengqiong Wu, Hao Fei, Liqiang Nie, and TatSeng Chua. Layoutllm-t2i: Eliciting layout guidance from llm for text-to-image generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 643-654, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[50] Leigang Qu, Haochuan Li, Tan Wang, Wenjie Wang, Yongqi Li, Liqiang Nie, and Tat-Seng Chua. Unified text-to-image generation and retrieval. arXiv preprint arXiv:2406.05814, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[51] Leigang Qu, Wenjie Wang, Yongqi Li, Hanwang Zhang, Liqiang Nie, and Tat-Seng Chua. Discriminative probing and tuning for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7434-7444, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[53] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[56] Sunayana Rane, Alexander Ku, Jason Baldridge, Ian Tenney, Tom Griffiths, and Been Kim. Can generative multimodal models count to ten? In Proceedings of the Annual Meeting of the Cognitive Science Society, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[58] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[59] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.832 + ], + "angle": 0, + "content": "[60] J Shawe-Taylor. Kernel methods for pattern analysis. Cambridge University Press google schola, 2:181-201, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[61] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 15" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[62] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[63] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations, 2023. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[64] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[67] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[68] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.594, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[69] Yufei Wang, Zhanyi Sun, Jesse Zhang, Zhou Xian, Erdem Biyik, David Held, and Zackory Erickson. Rl-vlm-f: Reinforcement learning from vision language foundation model feedback. arXiv preprint arXiv:2402.03681, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[70] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv preprint arXiv:2401.15688, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[71] Haokun Wen, Xuemeng Song, Xin Yang, Yibing Zhan, and Liqiang Nie. Comprehensive linguistic-visual composition network for image retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1369–1378, 2021. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[72] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In International Conference on Machine Learning, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[73] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Better aligning text-to-image models with human preference. arXiv preprint arXiv:2303.14420, 1(3), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[74] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18507" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[76] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[77] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 2, 3, 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[78] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[79] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[80] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 15" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18508" + } + ] +] \ No newline at end of file diff --git a/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_origin.pdf b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..52f6980ada01ddc174e5e01e0e76d7e1d3d6e3a7 --- /dev/null +++ b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/d6a00039-382c-4d20-89db-a164b6033fad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efadb7681c7e29ee31ca70c7bc4f3a8e8608bb451928239f62c3aa7b8a6f4613 +size 2526297 diff --git a/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/full.md b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee50d282d2b9045eab337e8e786549ac0c4fd294 --- /dev/null +++ b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/full.md @@ -0,0 +1,333 @@ +# SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation + +Leigang $\mathrm{Qu}^{1}$ , Haochuan $\mathrm{Li}^{1}$ , Wenjie Wang $^{2*}$ , Xiang Liu $^{1}$ , Juncheng Li $^{3}$ , Liqiang Nie $^{4}$ , Tat-Seng Chua $^{1}$ + +$^{1}$ National University of Singapore, $^{2}$ University of Science and Technology of China, $^{3}$ Zhejiang University, $^{4}$ Harbin Institute of Technology (Shenzhen) + +leigangqu@gmail.com, haochuan@u.nus.edu, wenjiewang96@gmail.com, liu.xiang@u.nus.edu + +junchengli@zju.edu.cn, nieliqiang@gmail.com, dcscts@nus.edu.sg + +# Abstract + +Large Multimodal Models (LMMs) have demonstrated impressive capabilities in multimodal understanding and generation, pushing forward advancements in text-to-image generation. However, achieving accurate text-image alignment for LMMs, particularly in compositional scenarios, remains challenging. Existing approaches, such as layout planning for multi-step generation and learning3 from human feedback or AI feedback, depend heavily on prompt engineering, costly human annotations, and continual upgrading, limiting flexibility and scalability. In this work, we introduce a model-agnostic iterative self-improvement framework (SILMM) that can enable LMMs to provide helpful and scalable self-feedback and optimize text-image alignment via Direct Preference Optimization (DPO). DPO can readily applied to LMMs that use discrete visual tokens as intermediate image representations; while it is less suitable for LMMs with continuous visual features, as obtaining generation probabilities is challenging. To adapt SILMM to LMMs with continuous features, we propose a diversity mechanism to obtain diverse representations and a kernel-based continuous DPO for alignment. Extensive experiments on three compositional text-to-image generation benchmarks validate the effectiveness and superiority of SILMM, showing improvements exceeding $30\%$ on T2I-CompBench++ and around $20\%$ on DPG-Bench. The code is available at https://silmm.github.io/. + +# 1. Introduction + +Large Multimodal Models (LMMs) are advancing rapidly, surpassing Large Language Models (LLMs) by embracing multimodal capabilities for multimodal content perception, understanding [34, 35, 48], and generation [19, 64]. + +![](images/a2a02d5224f9cc65de72e7c1d7be7c620243e628608b017949dacdddc5f5129a.jpg) +a green leaf and a yellow butterfly +Attribute Binding + +![](images/f1b44fcf49d54ead8088bc18a910e92cb54313e2396db903a353262060fa865e.jpg) +five balloons +Counting Error + +![](images/1f261033ec7922fb1d68e85ed0876be40ec2c8eff32759e14816be1d011adfe6.jpg) +a bicycle on the left of a man +Relation Inaccuracy + +![](images/ef5c257411ef4ff8323d4097b7fddc94923a468d02bd53f027adbc7e3f820697.jpg) +(a) Text-image misalignment in compositional prompts +(b) Discrete & Continuous LMMs for text-to-image generation +Figure 1. Illustration of (a) text-image misalignment in compositional prompts and (b) comparison of discrete and continuous LMMs for T2I. Given a prompt, discrete LMMs can sample diverse token sequences from categorical distributions, while continuous LMMs can only produce a single deterministic feature vector. Note that the input learnable embeddings are optional for some continuous LMMs [64]. + +In particular, LMMs demonstrate promising abilities in interpreting user input prompts for text-to-image generation (T2I) [55, 57], producing vivid and photorealistic images. However, as shown in Fig. 1(a), achieving precise text-image alignment between generated images and complex prompts remains challenging, especially for compositional prompts involving multiple objects, attributes, counting, and complex relationships [6, 16, 49]. + +To enhance text-image alignment, existing work falls + +into two primary research lines. One line focuses on decomposing the T2I task into multiple stages. For example, some methods perform layout planning before generating the image [17, 37, 75]; while some split the image into sections for multi-step generation via multi-agent collaboration [46, 70]. However, these methods depend on extensive multi-step prompt engineering, which risks error accumulation. The second research line emphasizes learning from human feedback (RLHF [43]) to improve text-image alignment [15, 31, 33, 67, 73], or using AI feedback (RLAIF) from strong evaluation approaches or reward models [3, 77]. Nevertheless, it is labor-intensive and costly to obtain extensive high-quality human feedback, which is also often required to train external reward models [5]. Additionally, as LMMs evolve, the external evaluation approaches and reward models may require continual upgrading [44, 77]. + +To address the limitations, we consider utilizing LLMs' inherent discriminative capabilities to self-improve their generation quality for text-image alignment. This offers a pathway for LMMs to evolve for T2I independently, without relying on human or external feedback. To pursue self-improvement, the key steps are: 1) generating diverse images by LMMs based on a given prompt, ensuring the image diversity to facilitate subsequent self-assessment and optimization; 2) using LMMs to self-assess text-image alignment in the generated images, producing alignment scores as self-feedback; and 3) adopting the self-feedback to optimize LMMs to generate superior visual tokens, resulting in images that better align with text prompts. + +However, achieving the above objectives faces significant challenges. In particular: + +1) As shown in Fig. 1(b), LMMs typically generate intermediate visual representations, i.e., discrete visual tokens or continuous visual features, which are then converted into images by a decoder (e.g., a diffusion model) [19, 64]. For LMMs with discrete visual tokens [19, 68, 76], using existing sampling strategies (e.g., adjusting temperature) in the autoregressive generation process can obtain diverse visual tokens. However, it is non-trivial for LMMs with deterministic continuous visual features, such as Dream-LLM [13], to sample diverse visual representations1. +2) Compositional prompts require LMMs to inspect object counts, attributes, and complex relationships in the generated images. However, existing LMMs still struggle with compositional cross-modal assessment [7, 56], challenging the generation of faithful self-feedback. +3) Optimizing LMMs with self-feedback is also intricate. Supervised Fine-Tuning (SFT) [11] and certain RLAIF methods [3, 77] require highly accurate self-feedback. Moreover, another representative method Direct Prefer- + +ence Optimization (DPO) requires modeling generation distributions, from which we need to sample diverse images to construct pairwise training data, which is challenging for LMMs with continuous visual features. + +To tackle the above challenges, we propose an Self-Improving Large Multimodal Models (SILMM) framework for iterative optimization. As illustrated in Fig. 2, SILMM operates through five steps: 1) Compositional Prompt Generation prompts an LMM to imagine compositional scenarios and generate compositional prompts. 2) Diverse Image Generation. For discrete LMMs $^2$ , we follow the sampling decoding strategy commonly used in LLM alignment [43, 53]. For continuous LMMs [13, 64, 72], we propose a diversification strategy named $DropDiv$ , inspired by Monte Carlo (MC) Dropout [18], to perform dropout on the MLP layers of LMMs for diverse visual features, producing diverse images. 3) Decompositional Self-Questioning. To reduce the difficulty of compositional cross-modal assessment, LMMs can decompose a compositional prompt into atomic concepts and relations and generate questions for multi-step assessment. 4) VQA-based Self-Feedback. For each image generated in Step 2, LMMs can use the decomposed questions to assess text-image alignment, and then aggregate the results to obtain reasonable self-feedback. 5) Learning from Self-Feedback. For discrete LMMs, we directly apply DPO based on pairwise samples from Step 2. As to continuous LMMs, we propose Kernel-based Continuous DPO (KC-DPO), inducing a quadruplet objective with kernel functions for pairwise distance regulation over continuous visual features. The above five steps can iteratively repeat until self-improvement performance converges. + +In summary, our main contributions are threefold: + +- To our knowledge, we are the first to focus on the task of LMMs' self-improvement for T2I. We propose a model-agnostic self-improvement framework to enable LMMs to achieve high-quality self-feedback and learning. +- For continuous LMMs, we introduce a dropout-based strategy to diversify image representations, along with a continuous DPO approach, i.e., KC-DPO, to optimize LMMs with preference representation pairs. +- We conduct extensive experiments on three compositional T2I benchmarks, demonstrating the superiority of SILMM, e.g., $30\%$ improvements on T2I-CompBench++. + +# 2. Related Work + +Compositional Text-to-Image Generation. Diffusion models [55, 57] have marked a significant advancement in T2I generation due to their stability and scalability. However, they still struggle with text-image alignment, such as attribute binding, counting error, and relation confu + +![](images/6950e498312f2f876f3561238d0d0cc017869b5dda0a32a6faeb0d364c754bc6.jpg) +Figure 2. Schematic illustration of SILMM, comprising five steps: 1) LMMs generate compositional prompts by sampling based on provided instructions. 2) Diverse representations and images are generated using either discrete nucleus sampling or the proposed continuous DivDrop. 3) LMMs divide each compositional prompt into semantic units and generate questions for each unit. 4) VQA is conducted to answer these questions, with the answers and likelihoods aggregated into alignment scores as self-feedback. 5) For alignment tuning, DPO is applied for discrete LMMs, while the proposed KC-DPO is used for continuous LMMs. + +sion [16, 51]. To enhance compositional T2I, some approaches intervene in language structures [16] or cross-attention mechanisms [6]. Other methods [17, 37, 38, 49] incorporate layout planning by LLMs or use multi-agent collaboration [46, 70]. Inspired by alignment successes in LLMs, recent work [5, 15, 67] applies RLHF [43] to optimize diffusion models. Despite the progress, they rely on inductive biases, extensive prompt engineering, or labor-intensive annotations, limiting flexibility and scalability. + +Large Multimodal Models. The pioneering LMMs [39, 79] integrate a visual encoder, e.g., CLIP [52], with LLMs as the foundation, showing impressive multimodal understanding capabilities. To extend LMMs to visual generation, recent approaches align diffusion models [13, 19, 72] with LLMs or train a single transformer [62, 68, 74, 76]. According to the form of output visual features, they can be divided into discrete visual tokenization methods [19, 50, 62, 68] and continuous visual representation methods [13, 64, 72]. While LLM integration enhances language understanding and supports flexible applications (e.g., interleaved multimodal generation [64]), compositional T2I in the context of LMMs remains underexplored. + +Learning from AI Feedback. The high cost of collect + +ing human preference has spurred research into RLAIF [3]. Benefiting from the convenience and scalability, there have been a series of studies adopting RLAIF to tackle a range of NLP tasks [10, 32, 78] and vision-language understanding [69, 77]. Despite the thrilling success, they only focus on text generation, overlooking the potential of RLAIF in other modalities. In contrast, we explore self-improving LMMs by activating multimodal understanding abilities for T2I. Particularly, we propose continuous strategies meticulously tailored to continuous visual features. + +# 3. Methodology + +In this section, we elaborate on the proposed method, including the SILMM framework with five steps and the iteration strategy (Sec. 3.1), as illustrated in Fig. 2. Afterward, we introduce the continuous KC-DPO applied to LMMs with continuous visual features in Sec. 3.2. + +# 3.1. Self-Improving Large Multimodal Models + +Step 1: Compositional Prompt Generation. We first divide compositional scenarios into four categories: Attribute (color, shape, texture), Layout (counting, spatial relation), + +Semantic Relation, and Complex Composition. Complex composition includes any possible composition of the first three. For attribute and layout, we prompt the LMM to separately generate common objects, attributes, numbers, and spatial relations, and then use templates to compose these concepts. For semantic relation and complex composition, we adopt in-context learning [12] to generate prompts. More details can be found in App. 6. + +Step 2: Diverse Representation and Image generation. The purpose of this step is to sample diverse intermediate visual representations from the LLM backbone $\pi$ of an LMM, given a text prompt $x$ , which would be decoded into images with different qualities. These representations are denoted as $\mathcal{Z} = \{z_i,\dots,z_M\}$ , where $z_{i}\sim \pi (z|x)$ . For discrete LMMs [19], $z_{i}$ is a discrete visual sequence. We follow the common practice [43, 53] in language generation to obtain $\mathcal{Z}$ , by sampling with different random seeds during auto-regressive decoding. For continuous LMMs [13], the LLM can only output a fixed continuous visual feature, without diversity. To tackle this issue, we propose DropDiv. First, we insert the dropout operations in the last few MLP layers of LLMs, which introduces randomness and enables LLMs for sampling. During inference, we activate these dropout operations to output diverse representations by sampling: $z_{i}\sim \pi^{\prime}(z|x)$ , where $z_{i}$ denotes a continuous visual feature and $\pi^\prime$ represents the LLM with activated dropout operations. Afterward, these diverse visual representations $\mathcal{Z}$ are decoded into images as $\mathcal{Y} = \{y_1,\dots,y_M\}$ . + +Discussion. Unlike prior work [5, 15] focused on tuning diffusion models, our approach resorts to LLM backbones in LLMs to control image decoders (e.g., diffusion models) for better text-image alignment, centering on LLM backbone optimization. Our approach offers three key advantages: 1) LLMs demonstrate superior proficiency in prompt comprehension over text encoders [52, 54] commonly employed in diffusion models. Tuning LLM backbones may unlock their enormous potential for compositional T2I, especially in complex scenarios. 2) Tuning diffusion models is often constrained by efficiency challenges inherent to iterative likelihood estimation, whereas there have been well-established technologies [1, 40, 53] for LLM alignment. 3) Our method is orthogonal to existing methods to tune diffusion models, combining them may get further gains. + +Step 3: Decompositional Self-Questioning. To provide helpful feedback to the generated images, the LMM should first accurately assess text-image alignment, which requires strong compositional reasoning abilities. However, current advanced LMMs still suffer from compositional reasoning [42], such as spatial relation understanding [7] and counting [56]. To improve compositional reasoning, we introduce a divide-and-conquer strategy [77] for self-questioning. Specifically, the LMM first divides the given prompt $x$ into atomic concepts (e.g., "a white harp") and + +relations (e.g., "a pancake is on the left of a pasta"), and then generates questions $\mathcal{Q} = \{q_1, \dots, q_N\}$ , each $q_i$ corresponding to a concept or relation. For simplicity, the generated questions are constrained to be yes/no questions (e.g., "Is there a while harp?", "Is the pancake on the left of the pasta?"). Refer to App. 7 for more details on prompt templates of self-questioning. + +Step 4: VQA-based Self-Feedback. Taking a generated image $y \in \mathcal{V}$ and all the questions $\mathcal{Q}$ as input, the LMM conducts the VQA task, and the average difference between the probabilities of answering "yes" and "no" serves as the text-image alignment score: + +$$ +s (x, y) = \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ p \left(" y e s" \mid y, q _ {i}\right) - p \left(" n o" \mid y, q _ {i}\right) \right]. \tag {1} +$$ + +Here we adopt the vision-language understanding abilities of LMMs via VQA to provide feedback to the images generated by themselves, thus this step is named VQA-based self-feedback. We carry out this step through all the sampled images prompted by $x$ and get all the scores $S = \{ s(x,y_j) | y_j \in \mathcal{V} \}$ . + +Step 5: Learning from Self-Feedback. Based on the selffeedback alignment scores, we sample representation pairs $(z_w,z_l)$ from $\mathcal{Q}$ , where $z_{w}$ and $z_{l}$ denote the chosen and the rejected representations and their corresponding decoded images should satisfy $s(x,y_w) > s(x,y_l)$ . With the preference data, we optimize the LLM backbone with DPO [53]: + +$$ +\mathcal {L} _ {\mathrm {D P O}} = - \mathbb {E} _ {(x, z _ {w}, z _ {l}) \sim \mathcal {D}} +$$ + +$$ +\left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(z _ {w} | x\right)}{\pi_ {\operatorname {r e f}} \left(z _ {w} | x\right)} - \beta \log \frac {\pi_ {\theta} \left(z _ {l} | x\right)}{\pi_ {\operatorname {r e f}} \left(z _ {l} | x\right)}\right) \right], \tag {2} +$$ + +where $\mathcal{D}$ denotes the training set, and $\pi_{\theta}$ and $\pi_{\mathrm{ref}}$ represent the policy and reference models, respectively. $\sigma$ is the sigmoid function, and $\beta$ is a hyperparameter controlling the deviation from the reference model. + +Iterative Self-Improvement. After learning from self-feedback, the updated LMM becomes more likely to generate preferred representations that are decoded into images better aligned with the prompt. This improvement in overall text-image alignment motivates us to iterate the above five steps with the updated LMM as the new reference model. The iteration mechanism continues until the alignment performance converges. As the process is independent of human annotations and external models, it is cost-effective and scalable. More importantly, it showcases the potential for self-improvement in LMMs by harmonizing their understanding and generation capabilities. + +# 3.2. Continuous Direct Preference Optimization + +At the step of learning from self-feedback, LMMs are optimized using the DPO objective as shown in Eqn. (2). + +The difference between discrete and continuous LMMs in this learning process lies in the calculation of the likelihood $\pi (z|x)$ . For discrete LMMs, $\pi (z|x)$ can be straightforwardly obtained by the softmax categorical distribution. However, for continuous LMMs with unknown distribution modeling, calculating $\pi (z|x)$ is intractable. + +Predictive Distribution with MC Dropout. MC Dropout [18] enables predictive distribution estimation via Monte Carlo simulation to calculate $\pi(z|x)$ . Specifically, the dropout layers3 in an LMM are activated during inference and the LMM performs forward propagation multiple times to get multiple outputs. Assuming a Gaussian distribution, we can estimate its parameters and calculate the likelihood $\pi(z|x)$ based on these outputs. However, such multi-forward estimation imposes a significant computational burden during training, making this approach insufficient and impractical. + +Simplified Kernel-based Continuous DPO. Inspired by MC Dropout and motivated by its insufficiency issue, we propose a simplified method to achieve continuous DPO. Concretely, the intermediate representation $z$ often performs as a feature matrix $\pmb{H} \in \mathbb{R}^{L \times D}$ where $L$ and $D$ denote the sequence length and dimension. $\pmb{H}$ can be attained by a Q-Former [13, 20] or from the last layer of the LMM in an autoregressive way [63, 64]. To estimate $\pi(\pmb{H} | x)$ , we first make a decomposition as: + +$$ +\pi (\boldsymbol {H} | x) = \prod_ {i = 1} ^ {L} \pi \left(\boldsymbol {h} _ {i} \mid \boldsymbol {H} _ {< i}, x\right), \tag {3} +$$ + +where $\pmb{h}_i \in \mathbb{R}^D$ denotes the $i$ -th feature vector. Based on the Gaussian assumption, we have: + +$$ +\pi \left(\boldsymbol {h} _ {i} \mid \boldsymbol {H} _ {< i}, x\right) = \frac {\exp \left[ - \frac {1}{2} \left(\boldsymbol {h} _ {i} - \boldsymbol {\mu} _ {i}\right) ^ {\top} \boldsymbol {\Sigma} _ {i} ^ {- 1} \left(\boldsymbol {h} _ {i} - \boldsymbol {\mu} _ {i}\right) \right]}{\sqrt {(2 \pi) ^ {D} \left| \boldsymbol {\Sigma} _ {i} \right|}}, \tag {4} +$$ + +where $\pmb{\mu}_i$ and $\pmb{\Sigma}_i$ denote the mean vector and the covariance matrix, respectively. Furthermore, we further simplify and approximate this formula: 1) the mean vector is estimated by the direct output of the continuous LMM, i.e., $\pmb{\mu}_i\approx \mathrm{LMM}(\boldsymbol {x})[i]$ , and 2) the Gaussian distribution is isotropic and all dimensions share the same variance value $\bar{\sigma}$ , i.e., $\pmb{\Sigma}_i\approx \mathrm{diag}(\sigma_1,\dots,\sigma_{\mathrm{D}})$ and $\sigma_{1} = \ldots = \sigma_{D} = \bar{\sigma}$ , and $\bar{\sigma}$ can be learnable or viewed as a hyperparameter. + +We compute the simplified likelihood with Eqn. (4), obtain the joint one with Eqn. (3), and finally derive the continuous DPO based on Eqn. (2): + +$$ +\begin{array}{r l} \mathcal {L} _ {\mathrm {C - D P O}} = & - \mathbb {E} _ {(x, \boldsymbol {H} _ {w}, \boldsymbol {H} _ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(\frac {\beta}{2 \bar {\sigma} ^ {2}} (- \| \boldsymbol {H} - \boldsymbol {H} _ {w} \| _ {F} ^ {2} + \| \boldsymbol {H} _ {r} - \boldsymbol {H} _ {w} \| _ {F} ^ {2})\right) \right], \end{array} \tag {5} +$$ + +where $\| \cdot \| _F$ denotes the Frobenius norm, and $H$ and $H_{r}$ represent the continuous feature matrices from the policy and reference LMMs, respectively. $H_{w}$ and $H_{l}$ refer to the chosen and rejected feature matrices, respectively. Compared with the MC dropout method, this objective only requires one forward pass, which is more efficient. We relegate more details of the derivation to App. 8. From Eqn. (5), we can see that this objective aims to adjust the relative distances within the quadruple $(\pmb {H},\pmb {H}_r,\pmb {H}_w,\pmb {H}_l)$ and the distance metric is the Euclidean distance between two matrices. To further improve the flexibility, we generalize the continuous DPO objective to, + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {K C - D P O}} = - \mathbb {E} _ {(x, \boldsymbol {H} _ {w}, \boldsymbol {H} _ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(\gamma (- k (\boldsymbol {H}, \boldsymbol {H} _ {w}) \right. \right. \\ \left. \left. + k \left(\boldsymbol {H} _ {r}, \boldsymbol {H} _ {w}\right) + k \left(\boldsymbol {H}, \boldsymbol {H} _ {l}\right) - k \left(\boldsymbol {H} _ {r}, \boldsymbol {H} _ {l}\right)\right)\right) \Bigg ], \tag {6} \\ \end{array} +$$ + +where $\gamma = \frac{\beta}{2\bar{\sigma}^2}$ controls the degree of adherence to the reference model, $k(\cdot ,\cdot)$ denotes a generalized distance measurement function. Considering it is similar to kernel methods [22, 60], we name the objective Kernel-based Continuous DPO (KC-DPO). In the following experiments section, we will discuss different distance functions and their influences on alignment performance. + +# 4. Experiments + +# 4.1. Experimental Setup + +Base Model Settings. We implement our method on DreamLLM (continuous LMM) [13] and SEED-LLaMA (discrete LMM) [19] for all experiments. We also apply our method to Emu-3 [68], the recent state-of-the-art discrete LMM. Details on DPO training are provided in App. 9. + +Datasets. We curated a dataset of 16,000 prompts across four categories using LMM. In each DPO training iteration, images generated by the model in the previous iteration served as the training data for next DPO iteration, allowing for iterative self-improvement. Details on data creation are provided in App. 10. + +**Benchmarks.** We evaluate our method on three text-to-image alignment benchmarks and follow their default settings. T2I-CompBench++ [28] consists of 8,000 compositional text prompts organized into 4 main categories: attribute, layout, non-spatial, and complex compositions, further divided into 8 subcategories, including color, binding, binding, 2D/3D-spatial relationships, non-spatial relationships, numeracy, and complex compositions. TIFA [27] uses pre-generated question-answer pairs and a VQA model to evaluate generation results based on 4,081 diverse text prompts and 25,829 questions across 12 categories. DPG-Bench [26] comprises 1,065 densely descriptive prompts with an average token length of 83.91, presenting more complex scenarios with varied objects and rich adjectives. + +![](images/83f0e9e544828bd594bc8e34656f52939070c923468dc348148a1e7d069c4c54.jpg) +Figure 3. Performance improvement of iterative alignment tuning based on SEED-LLaMA and DreamLLM, across 8 detailed categories of T2I-CompBench++. Iter. 0 denotes the base models without alignment tuning. + +# 4.2. Performance Comparison + +As shown in Tab. 1, we evaluate alignment performance of our method against T2I generative models and base LMMs on three compositional T2I benchmarks, including T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Key observations are as follows: 1) Although LMMs enable more flexible settings (e.g., in-context learning and interleaved multimodal generation) for image generation, they still underperform compared to specialized T2I models in terms of the basic alignment ability to follow prompts. It demonstrates that current LMMs may ignore the compositional text-image alignment during multimodal pre-training and fine-tuning. 2) Without human annotations or external models, the proposed SILMM method enhances alignment performance across all categories in three benchmarks over the base LMMs, improving both the discrete SEEDLLaMA and the continuous DreamLLM, verifying the effectiveness and the generalization of SILMM. 3) SEEDLLaMA shows greater self-improvement than DreamLLM, possibly due to its weaker baseline alignment and the stability of discrete DPO over continuous KC-DPO induced by a series of simplification, as discussed in Sec. 3.2. And 4) improvements are more challenging in layout, relation, and complex categories than in attribute categories. This difficulty arises partly because the basic generative ability in these categories is weak, making it difficult to obtain high-quality chosen samples. Besides, understanding compositional concepts remains a challenge for LMMs [7, 56]. + +# 4.3. In-depth Analysis + +To explore the efficacy of SILMM, we conduct extensive ablation studies and hyperparameter analyses. We first investigate the iteration process and data scaling, followed by an in-depth study of key components, including diversity strategies, decompositional self-questioning and answering for self-feedback, and KC-DPO. + +Iterative Self-Improvement. As shown in Fig. 3, we conduct three iterations of self-improvement and assess + +performance changes across eight detailed categories of T2I-CompBench++ [28]. The results show that SILMM achieves effective, consistent, and continuous improvements in text-image alignment, across most compositional categories. Notably, attribute categories (e.g., color, shape, and texture) exhibit the most significant gains, whereas the non-spatial category shows slower improvement. This slower progress may stem from CLIP score [23], which is less sensitive than other metrics. Finally, as the iteration progresses, improvement rates gradually decrease, indicating convergence. More iterative self-improvement experiments results can be found in App. 11. + +![](images/2614437d62fcf9e7434be36fbdf5f096e32042199a379bd96215b9f0fbb10a37.jpg) +(a) Num. of Training Prompts + +![](images/0d440124397b0c90faff956d1051b90b1cf838014c379a499ba7d8182de20a41.jpg) +(b) Num. of Pairs per Prompt +Figure 4. Overall alignment scores of SEED-LLaMA with discrete DPO and DreamLLM with continuous KC-DPO, on T2I-CompBench++ with (a) varying numbers of generated prompts in the training data, and (b) different number of preference pairs sampled from 30 diverse generated images per prompt. $N \times N$ means we select the top-N and last-N images from 30 generated ones as the chosen and rejected, respectively. + +Data Scale. The proposed SILMM method leverages self-synthesized data for tuning, allowing flexible adjustment of data scale according to practical needs and available computational resources. In Fig. 4, we investigate how data scale affects overall alignment performance (averaged across eight categories in T2I-CompBench++, focusing on two factors: the number of training prompts and the number of preference pairs per prompt. Results in Fig. 4a indicate that both LMMs show consistent improvement as data samples increase, showing the strong scalability of the proposed method. Besides, we generate 30 representations and images per prompt, and then select the top- $N$ and last- $N$ samples to construct $N \times N$ preference pairs (see Fig. 4b). Notably, the two LMMs perform differently. This may be because the continuous feature space, being larger and denser than the discrete space, requires denser data pairs to stabilize the optimization dynamics. + +Diversity Strategies. To synthesize high-quality preference pair data, the diversity strategy is crucial. An effective diversity strategy should maximize the potential of LMMs while ensuring sufficient variation among generated images. To explore various strategies, we compare the proposed $DropDiv$ with three alternatives: Rephrase the original prompt, Explain the original prompt with added imaginative elements, and add Gaussian noises to the learnable + +Table 1. Performance comparison and improvement of the proposed method for compositional text-to-image generation on T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Alignment scores are calculated using expert understanding models (e.g., VQA or object detection models) recommended by these benchmarks. Prompt rewriting in Emu3 [68] was not used for fair comparison. + +
MethodT2I-CompBench++ [28]DPG-Bench [26]TIFA [27]
AttributeLayoutNon-spatialComplexGlobalEntityAttributeRelationOtherAllAll
Text-to-Image Generative Models
SD-v1.5 [55]38.65---74.6374.2375.3973.4967.8163.1878.40
DALL-E 2 [55]58.63----------
SD-v2 [55]47.3630.5031.2733.8677.6778.1374.9180.7280.6668.09-
SD-v2.1 [55]50.57---------82.00
SDXL [45]52.8835.6231.1932.3783.2782.4380.9186.7680.4174.65-
PixArt-α [8]60.3136.7431.9734.3374.9779.3278.6082.5776.9671.11-
DALL-E 3 [4]70.0941.6330.0337.7390.9789.6188.3990.5889.8383.50-
Large Multimodal Models
SEED-LLaMA [19]19.2020.2928.8621.4665.5955.8761.9662.7759.4647.1266.74
SEED-LLaMA + Ours39.6025.1129.8228.2873.5570.4868.4974.7968.6457.3173.74
%Improvement106.25%23.77%3.33%31.78%12.14%26.15%10.54%19.15%15.44%21.63%10.49%
DreamLLM [13]22.9423.7428.7623.0174.4765.8663.8074.2446.0053.9369.91
DreamLLM + Ours39.9427.6329.0026.4376.2975.9169.2084.4160.0064.2275.38
%Improvement74.15%16.40%0.83%14.86%2.44%15.26%8.46%13.70%30.43%19.08%7.82%
Emu3 [68]44.7932.3030.1531.3284.1980.8182.7587.2350.8074.1981.86
Emu3 + Ours59.7136.0330.5133.9384.1981.5784.5289.0164.8077.4585.11
%Improvement33.30%11.57%1.19%8.33%0.00%0.94%2.14%2.04%27.56%4.39%3.97%
+ +![](images/b5410bb3d5febd78c8305bbb114ae59eac73ddd33c5231041c24ff85f280532b.jpg) +(a) Color, Shape, and Texture + +![](images/0b0b8c4058c035053301add10f33b5b5c81cfcb9dbda59a6672a7f3a493b9033.jpg) +(b) Spatial, 3D Spatial, and Numeracy + +![](images/16231f133c5649e6ced2d0eee83e643ba68f3addd136d8906b86c958ca1c0b77.jpg) +(c) Non-spatial and Complex + +![](images/689ba48b80820c28b2f7da355c5d1f74d4e168c8ef89f1e9dedae31f6bd481b2.jpg) +Figure 5. Comparison of four methods for diverse continuous representation generation, with alignment scores evaluated on the validation set of T2I-CompBench++. For each prompt, DreamLLM generates ten diverse representations and corresponding images. +Figure 6. Distribution of alignment scores and variation of maximum scores across different dropout rates (%Drop) of the proposed DropDiv method, evaluated on the T2I-CompBench++ validation set. For each prompt, DreamLLM generates ten diverse representations and corresponding images with DropDiv. + +Dream Embeddings in DreamLLM [13] to create Noisy DreamEmb. As shown in Fig. 5, four strategies have different perturbation influences on different categories. For example, DropDiv could generate better samples in attribute, + +non-spatial, and complex categories, but compromise in layout categories. To further examine the effects of Drop-Div, we conduct experiments across different dropout rates as shown in Fig. 6. Results indicate that higher dropout rates enhance diversity, but the alignment quality could be impaired. Therefore, achieving a good diversity-quality tradeoff remains challenging. + +Decompositional Self-Feedback. We perform ablation studies on question generation and alignment score calculation, as shown in Tab. 2. Compared to two variants Prompt-Q (appending or replacing the period with “?” at the end of each prompt) and Phrase-Q (segmenting each prompt into phrases using NLP tools [24]), Self-Questioning (Self-Q) achieves better alignment performance across most categories, demonstrating the effectiveness of leveraging language processing abilities of LMMs for text-image alignment evaluation. Additionally, we compare the proposed alignment score calculation from Eqn. (1) with two variants: Random Score and Ratio of “yes” (where a higher ratio indicates a higher score). Results show that our method + +Table 2. Ablation study on T2I-CompBench++ [28] and DPG-Bench [26] examining variations in Question Generation and VQA-based Alignment Score Calculation methods for self-feedback. Prompt-Q adds a “?” or replaces the period with a “?” at the end of each prompt. Phrase-Q involves dividing a prompt into phrases, each followed by a “?”. Self-Q instructs the LMM to generate questions for each prompt using in-context examples. Diff. of Prob. denotes the proposed alignment score calculation approach described in Eqn. (1). + +
FeedbackT2I-CompBench++DPG-Bench
AttributeLayoutNon-spatialComplexAll
Baseline (DreamLLM [13])
-22.9423.7428.7623.0153.93
Question Generation
Prompt-Q33.6625.9328.1424.6760.13
Phrase-Q34.6324.9128.0125.4160.10
Self-Q34.8525.5128.8225.3160.95
VQA-based Alignment Score Calculation
Random Score23.4124.8128.6722.9553.89
Ratio of “yes”25.3623.5128.7324.0054.68
Diff. of Prob.34.8525.5128.8225.3160.95
+ +Table 3. Ablation study on T2I-CompBench++ [28] to investigate different instantiation of the Kernel Function to calculate the continuous KC-DPO loss function to tune DreamLLM. Aggregation means we aggregate the 2D feature matrix (e.g., $\mathbf{H}$ ) into 1D along the sequence dimension. Eucl. denotes Euclidean distance. + +
AggregationDistanceAttributeLayoutNon-spatialComplex
Baseline (DreamLLM [13])
--22.9423.7428.7623.01
Supervised Fine-tuning (SFT)
-Eucl.12.250.7516.4111.71
-Cos6.950.2916.7811.48
AvgPoolEucl.23.3123.8928.7623.22
AvgPoolCos23.1224.2028.7923.29
Continuous Kernel-based Direct Preference Optimization
-Eucl.23.6524.3428.8323.08
-Cos23.9724.1128.7723.21
MaxPoolEucl.23.7924.0428.8623.92
MaxPoolCos29.1825.0118.7212.27
AvgPoolEucl.26.7524.7028.9425.12
AvgPoolCos34.8525.5128.8225.31
+ +achieves superior performance by considering the relative confidence between "yes" and "no". + +Kernel-based Continuous DPO. In Sec. 3.2, we introduce the KC-DPO to fine-tune LMMs with continuous representations. The implementation of kernel functions can be divided into Aggregation and Distance. To assess the impacts of different kernels, we conduct extensive comparison experiments, as shown in Tab. 3. We observe SFT slightly improves the alignment performance, while DPO yields more substantial gains across all metrics. These results show that kernel functions are crucial to KC-DPO, and an optimal choice could greatly enhance the efficiency of preference optimization in continuous feature space. Overall, AvgPool + Cos demonstrates the superior performance improvement. + +# 4.4. Qualitative Results + +To illustrate the improvements achieved by SILMM, Fig. 8 presents examples generated by SEED-LLaMA, Dream + +![](images/e0f47ca9a390922442803deee89f7647d4dcc9d48dc36b0f6dd92f942c7afab1.jpg) +(a) $\beta$ in Discrete DPO + +![](images/595a31543c9a62e10a57d0d2656f10583ab1f115d301086164fcdf8841fb73b4.jpg) +(b) $\gamma$ in Continuous KC-DPO + +![](images/1dc0fccd90873afa413bdd93614ff965e05b5685c3af12456817877ad3fec3d7.jpg) +Figure 7. Hyperparameter sensitivity on four general categories of T2I-CompBench++, examining (a) $\beta$ in discrete DPO for SEEDLLaMA, and (b) $\gamma$ in continuous KC-DPO for DreamLLM. +Figure 8. Qualitative results from SEED-LLaMA, DreamLLM, and the proposed SILMM method, on T2I-CompBench++. + +LLM, and SILMM, in Fig. 8 using prompts from T2I-CompBench++. These results showcase the effectiveness of SILMM across extensive compositional scenarios. + +# 5. Conclusion + +In this work, we present a self-improvement approach named SILMM to enhance text-image alignment within LMMs, introducing an iterative model-agnostic framework comprising five stages to enable high-quality self-feedback and alignment learning. For continuous LMMs, we propose a dropout-based strategy to diversify image representations and a continuous DPO method, KC-DPO, for optimizing LMMs with preference representation pairs. Extensive experiments validate the effectiveness and superiority of our SILMM framework. + +# References + +[1] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, pages 4447-4455. PMLR, 2024. 4 +[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 15 +[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2, 3 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2(3):8, 2023. 7 +[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICML Workshop on Structured Probabilistic Inference & Generative Modeling, 2023. 2, 3, 4 +[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics, 42(4):1-10, 2023. 1, 3 +[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 4, 6 +[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 7 +[9] Xiaolin Chen, Xuemeng Song, Liqiang Jing, Shuo Li, Linmei Hu, and Liqiang Nie. Multimodal dialog systems with dual knowledge-enhanced generative pretrained language model. ACM Transactions on Information Systems, 42(2): 1-25, 2023. 17 +[10] Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al. Ultrafeedback: Boosting language models with scaled ai feedback. In International Conference on Machine Learning, 2024. 3 +[11] Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, and Jingren Zhou. How abilities in large language models are affected by supervised fine-tuning data composition. arXiv preprint arXiv:2310.05492, 2023. 2 +[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu + +Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4 +[13] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In International Conference on Learning Representations, 2024. 2, 3, 4, 5, 7, 8, 15, 18 +[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5 +[15] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4 +[16] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In International Conference on Machine Learning, 2023. 1, 3, 17 +[17] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 2, 3 +[18] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 2, 5 +[19] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. In International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5, 7, 18 +[20] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024.5, 15 +[21] Yangyang Guo, Guangzhi Wang, and Mohan Kankanhalli. Pela: Learning parameter-efficient models with low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15699-15709, 2024. 17 +[22] Marti A. Hearst, Susan T Dumais, Edgar Osuna, John Platt, and Bernhard Scholkopf. Support vector machines. IEEE Intelligent Systems and their applications, 13(4):18-28, 1998. 5 +[23] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6 +[24] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, + +and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7 +[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 16 +[26] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 5, 6, 7, 8, 16, 17, 18, 21, 22 +[27] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 5, 6, 7, 16, 18 +[28] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 5, 6, 7, 8, 16, 17, 18, 19, 20 +[29] Natasha Jaques, Shixiang Gu, Dzmitry Bahdanau, José Miguel Hernández-Lobato, Richard E Turner, and Douglas Eck. Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control. In International Conference on Machine Learning, pages 1645-1654. PMLR, 2017. 15 +[30] Natasha Jaques, Judy Hanwen Shen, Asma Ghandehar-ioun, Craig Ferguson, Agata Lapedriza, Noah Jones, Shixiang Shane Gu, and Rosalind Picard. Human-centric dialog training via offline reinforcement learning. arXiv preprint arXiv:2010.05848, 2020. 15 +[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2 +[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Ren Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. In International Conference on Machine Learning, 2024. 3, 15 +[33] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2 +[34] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 1 +[35] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[36] Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, and Mohan Kankanhalli. Attribute-driven disentangled + +representation learning for multimodal recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 9660-9669. ACM, 2024. 17 +[37] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3 +[38] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. In International Conference on Learning Representations, 2024. 3 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3 +[40] Yongshuai Liu, Jiaxin Ding, and Xin Liu. Ipo: Interior-point policy optimization under constraints. In Proceedings of the AAAI conference on artificial intelligence, pages 4940-4947, 2020. 4 +[41] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024. 17 +[42] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 4 +[43] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 3, 4, 14, 15 +[44] Jing-Cheng Pang, Pengyuan Wang, Kaiyuan Li, Xiong-Hui Chen, Jiacheng Xu, Zongzhang Zhang, and Yang Yu. Language model self-improvement by reinforcement learning contemplation. In International Conference on Learning Representations, 2024. 2 +[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7 +[46] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2, 3 +[47] Tianyi Qin, Bo Peng, Jianjun Lei, Jiahui Song, Liying Xu, and Qingming Huang. 3d-immc: Incomplete multi-modal 3d shape clustering via cross mapping and dual adaptive fusion. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 17 +[48] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 1 + +[49] Leigang Qu, Shengqiong Wu, Hao Fei, Liqiang Nie, and TatSeng Chua. Layoutllm-t2i: Eliciting layout guidance from llm for text-to-image generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 643-654, 2023. 1, 3 +[50] Leigang Qu, Haochuan Li, Tan Wang, Wenjie Wang, Yongqi Li, Liqiang Nie, and Tat-Seng Chua. Unified text-to-image generation and retrieval. arXiv preprint arXiv:2406.05814, 2024.3 +[51] Leigang Qu, Wenjie Wang, Yongqi Li, Hanwang Zhang, Liqiang Nie, and Tat-Seng Chua. Discriminative probing and tuning for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7434-7444, 2024. 3 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4 +[53] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 15 +[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4 +[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 7 +[56] Sunayana Rane, Alexander Ku, Jason Baldridge, Ian Tenney, Tom Griffiths, and Been Kim. Can generative multimodal models count to ten? In Proceedings of the Annual Meeting of the Cognitive Science Society, 2024. 2, 4, 6 +[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[58] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 15 +[59] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 17 +[60] J Shawe-Taylor. Kernel methods for pattern analysis. Cambridge University Press google schola, 2:181-201, 2004. 5 +[61] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 15 + +[62] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3 +[63] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations, 2023. 5, 15 +[64] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 2, 3, 5 +[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5 +[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 15 +[67] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 3 +[68] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 5, 7 +[69] Yufei Wang, Zhanyi Sun, Jesse Zhang, Zhou Xian, Erdem Biyik, David Held, and Zackory Erickson. Rl-vlm-f: Reinforcement learning from vision language foundation model feedback. arXiv preprint arXiv:2402.03681, 2024. 3 +[70] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv preprint arXiv:2401.15688, 2024. 2, 3 +[71] Haokun Wen, Xuemeng Song, Xin Yang, Yibing Zhan, and Liqiang Nie. Comprehensive linguistic-visual composition network for image retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1369–1378, 2021. 17 +[72] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In International Conference on Machine Learning, 2024. 2, 3 +[73] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Better aligning text-to-image models with human preference. arXiv preprint arXiv:2303.14420, 1(3), 2023. 2 +[74] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie + +Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024.3 +[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In International Conference on Machine Learning, 2024. 2 +[76] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2, 3 +[77] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 2, 3, 4, 15 +[78] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 2024. 3 +[79] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 +[80] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 15 \ No newline at end of file diff --git a/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/images.zip b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5f3ba4c20f3039ea1f66ea44d16342e4c0fc3e3 --- /dev/null +++ b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac2f258c8cefa4f63e84b99fd886c67856024a87c1e10dd8a29d2678aa6239cf +size 753107 diff --git a/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/layout.json b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2acdcbea93894976b5dadc4661a5e13725948e0f --- /dev/null +++ b/2025/SILMM_ Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation/layout.json @@ -0,0 +1,9700 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 102, + 522, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 522, + 138 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 522, + 138 + ], + "type": "text", + "content": "SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": "Leigang " + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{Qu}^{1}" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": ", Haochuan " + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^{1}" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": ", Wenjie Wang" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": ", Xiang Liu" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": ", Juncheng Li" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": ", Liqiang Nie" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "text", + "content": ", Tat-Seng Chua" + }, + { + "bbox": [ + 61, + 160, + 552, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "spans": [ + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "text", + "content": "National University of Singapore, " + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "text", + "content": "University of Science and Technology of China, " + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "text", + "content": "Zhejiang University, " + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 93, + 177, + 522, + 203 + ], + "type": "text", + "content": "Harbin Institute of Technology (Shenzhen)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 205, + 535, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 205, + 535, + 217 + ], + "spans": [ + { + "bbox": [ + 80, + 205, + 535, + 217 + ], + "type": "text", + "content": "leigangqu@gmail.com, haochuan@u.nus.edu, wenjiewang96@gmail.com, liu.xiang@u.nus.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 220, + 474, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 220, + 474, + 232 + ], + "spans": [ + { + "bbox": [ + 138, + 220, + 474, + 232 + ], + "type": "text", + "content": "junchengli@zju.edu.cn, nieliqiang@gmail.com, dcscts@nus.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) have demonstrated impressive capabilities in multimodal understanding and generation, pushing forward advancements in text-to-image generation. However, achieving accurate text-image alignment for LMMs, particularly in compositional scenarios, remains challenging. Existing approaches, such as layout planning for multi-step generation and learning3 from human feedback or AI feedback, depend heavily on prompt engineering, costly human annotations, and continual upgrading, limiting flexibility and scalability. In this work, we introduce a model-agnostic iterative self-improvement framework (SILMM) that can enable LMMs to provide helpful and scalable self-feedback and optimize text-image alignment via Direct Preference Optimization (DPO). DPO can readily applied to LMMs that use discrete visual tokens as intermediate image representations; while it is less suitable for LMMs with continuous visual features, as obtaining generation probabilities is challenging. To adapt SILMM to LMMs with continuous features, we propose a diversity mechanism to obtain diverse representations and a kernel-based continuous DPO for alignment. Extensive experiments on three compositional text-to-image generation benchmarks validate the effectiveness and superiority of SILMM, showing improvements exceeding " + }, + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "type": "text", + "content": " on T2I-CompBench++ and around " + }, + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 285, + 296, + 597 + ], + "type": "text", + "content": " on DPG-Bench. The code is available at https://silmm.github.io/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 622, + 135, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 135, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 135, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 643, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 692 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) are advancing rapidly, surpassing Large Language Models (LLMs) by embracing multimodal capabilities for multimodal content perception, understanding [34, 35, 48], and generation [19, 64]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 345, + 287, + 393, + 335 + ], + "blocks": [ + { + "bbox": [ + 345, + 271, + 400, + 287 + ], + "lines": [ + { + "bbox": [ + 345, + 271, + 400, + 287 + ], + "spans": [ + { + "bbox": [ + 345, + 271, + 400, + 287 + ], + "type": "text", + "content": "a green leaf and a yellow butterfly" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 287, + 393, + 335 + ], + "lines": [ + { + "bbox": [ + 345, + 287, + 393, + 335 + ], + "spans": [ + { + "bbox": [ + 345, + 287, + 393, + 335 + ], + "type": "image", + "image_path": "a2a02d5224f9cc65de72e7c1d7be7c620243e628608b017949dacdddc5f5129a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 337, + 393, + 345 + ], + "lines": [ + { + "bbox": [ + 345, + 337, + 393, + 345 + ], + "spans": [ + { + "bbox": [ + 345, + 337, + 393, + 345 + ], + "type": "text", + "content": "Attribute Binding" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 409, + 287, + 457, + 335 + ], + "blocks": [ + { + "bbox": [ + 413, + 275, + 451, + 283 + ], + "lines": [ + { + "bbox": [ + 413, + 275, + 451, + 283 + ], + "spans": [ + { + "bbox": [ + 413, + 275, + 451, + 283 + ], + "type": "text", + "content": "five balloons" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 409, + 287, + 457, + 335 + ], + "lines": [ + { + "bbox": [ + 409, + 287, + 457, + 335 + ], + "spans": [ + { + "bbox": [ + 409, + 287, + 457, + 335 + ], + "type": "image", + "image_path": "f1b44fcf49d54ead8088bc18a910e92cb54313e2396db903a353262060fa865e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 337, + 455, + 346 + ], + "lines": [ + { + "bbox": [ + 411, + 337, + 455, + 346 + ], + "spans": [ + { + "bbox": [ + 411, + 337, + 455, + 346 + ], + "type": "text", + "content": "Counting Error" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 475, + 287, + 523, + 335 + ], + "blocks": [ + { + "bbox": [ + 475, + 271, + 524, + 286 + ], + "lines": [ + { + "bbox": [ + 475, + 271, + 524, + 286 + ], + "spans": [ + { + "bbox": [ + 475, + 271, + 524, + 286 + ], + "type": "text", + "content": "a bicycle on the left of a man" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 287, + 523, + 335 + ], + "lines": [ + { + "bbox": [ + 475, + 287, + 523, + 335 + ], + "spans": [ + { + "bbox": [ + 475, + 287, + 523, + 335 + ], + "type": "image", + "image_path": "1f261033ec7922fb1d68e85ed0876be40ec2c8eff32759e14816be1d011adfe6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 337, + 528, + 346 + ], + "lines": [ + { + "bbox": [ + 470, + 337, + 528, + 346 + ], + "spans": [ + { + "bbox": [ + 470, + 337, + 528, + 346 + ], + "type": "text", + "content": "Relation Inaccuracy" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 327, + 369, + 545, + 500 + ], + "blocks": [ + { + "bbox": [ + 353, + 352, + 517, + 362 + ], + "lines": [ + { + "bbox": [ + 353, + 352, + 517, + 362 + ], + "spans": [ + { + "bbox": [ + 353, + 352, + 517, + 362 + ], + "type": "text", + "content": "(a) Text-image misalignment in compositional prompts" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 369, + 545, + 500 + ], + "lines": [ + { + "bbox": [ + 327, + 369, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 327, + 369, + 545, + 500 + ], + "type": "image", + "image_path": "ef5c257411ef4ff8323d4097b7fddc94923a468d02bd53f027adbc7e3f820697.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 502, + 527, + 512 + ], + "lines": [ + { + "bbox": [ + 343, + 502, + 527, + 512 + ], + "spans": [ + { + "bbox": [ + 343, + 502, + 527, + 512 + ], + "type": "text", + "content": "(b) Discrete & Continuous LMMs for text-to-image generation" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 514, + 555, + 591 + ], + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 591 + ], + "type": "text", + "content": "Figure 1. Illustration of (a) text-image misalignment in compositional prompts and (b) comparison of discrete and continuous LMMs for T2I. Given a prompt, discrete LMMs can sample diverse token sequences from categorical distributions, while continuous LMMs can only produce a single deterministic feature vector. Note that the input learnable embeddings are optional for some continuous LMMs [64]." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 603, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 699 + ], + "type": "text", + "content": "In particular, LMMs demonstrate promising abilities in interpreting user input prompts for text-to-image generation (T2I) [55, 57], producing vivid and photorealistic images. However, as shown in Fig. 1(a), achieving precise text-image alignment between generated images and complex prompts remains challenging, especially for compositional prompts involving multiple objects, attributes, counting, and complex relationships [6, 16, 49]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "To enhance text-image alignment, existing work falls" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18497" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": "into two primary research lines. One line focuses on decomposing the T2I task into multiple stages. For example, some methods perform layout planning before generating the image [17, 37, 75]; while some split the image into sections for multi-step generation via multi-agent collaboration [46, 70]. However, these methods depend on extensive multi-step prompt engineering, which risks error accumulation. The second research line emphasizes learning from human feedback (RLHF [43]) to improve text-image alignment [15, 31, 33, 67, 73], or using AI feedback (RLAIF) from strong evaluation approaches or reward models [3, 77]. Nevertheless, it is labor-intensive and costly to obtain extensive high-quality human feedback, which is also often required to train external reward models [5]. Additionally, as LMMs evolve, the external evaluation approaches and reward models may require continual upgrading [44, 77]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 294, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 419 + ], + "type": "text", + "content": "To address the limitations, we consider utilizing LLMs' inherent discriminative capabilities to self-improve their generation quality for text-image alignment. This offers a pathway for LMMs to evolve for T2I independently, without relying on human or external feedback. To pursue self-improvement, the key steps are: 1) generating diverse images by LMMs based on a given prompt, ensuring the image diversity to facilitate subsequent self-assessment and optimization; 2) using LMMs to self-assess text-image alignment in the generated images, producing alignment scores as self-feedback; and 3) adopting the self-feedback to optimize LMMs to generate superior visual tokens, resulting in images that better align with text prompts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 421, + 294, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 294, + 445 + ], + "type": "text", + "content": "However, achieving the above objectives faces significant challenges. In particular:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 445, + 294, + 672 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 52, + 445, + 294, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 445, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 52, + 445, + 294, + 564 + ], + "type": "text", + "content": "1) As shown in Fig. 1(b), LMMs typically generate intermediate visual representations, i.e., discrete visual tokens or continuous visual features, which are then converted into images by a decoder (e.g., a diffusion model) [19, 64]. For LMMs with discrete visual tokens [19, 68, 76], using existing sampling strategies (e.g., adjusting temperature) in the autoregressive generation process can obtain diverse visual tokens. However, it is non-trivial for LMMs with deterministic continuous visual features, such as Dream-LLM [13], to sample diverse visual representations1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 565, + 294, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 565, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 51, + 565, + 294, + 624 + ], + "type": "text", + "content": "2) Compositional prompts require LMMs to inspect object counts, attributes, and complex relationships in the generated images. However, existing LMMs still struggle with compositional cross-modal assessment [7, 56], challenging the generation of faithful self-feedback." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 625, + 294, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 625, + 294, + 672 + ], + "spans": [ + { + "bbox": [ + 51, + 625, + 294, + 672 + ], + "type": "text", + "content": "3) Optimizing LMMs with self-feedback is also intricate. Supervised Fine-Tuning (SFT) [11] and certain RLAIF methods [3, 77] require highly accurate self-feedback. Moreover, another representative method Direct Prefer-" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 119 + ], + "type": "text", + "content": "ence Optimization (DPO) requires modeling generation distributions, from which we need to sample diverse images to construct pairwise training data, which is challenging for LMMs with continuous visual features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "type": "text", + "content": "To tackle the above challenges, we propose an Self-Improving Large Multimodal Models (SILMM) framework for iterative optimization. As illustrated in Fig. 2, SILMM operates through five steps: 1) Compositional Prompt Generation prompts an LMM to imagine compositional scenarios and generate compositional prompts. 2) Diverse Image Generation. For discrete LMMs" + }, + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "type": "text", + "content": ", we follow the sampling decoding strategy commonly used in LLM alignment [43, 53]. For continuous LMMs [13, 64, 72], we propose a diversification strategy named " + }, + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "type": "inline_equation", + "content": "DropDiv" + }, + { + "bbox": [ + 313, + 121, + 553, + 443 + ], + "type": "text", + "content": ", inspired by Monte Carlo (MC) Dropout [18], to perform dropout on the MLP layers of LMMs for diverse visual features, producing diverse images. 3) Decompositional Self-Questioning. To reduce the difficulty of compositional cross-modal assessment, LMMs can decompose a compositional prompt into atomic concepts and relations and generate questions for multi-step assessment. 4) VQA-based Self-Feedback. For each image generated in Step 2, LMMs can use the decomposed questions to assess text-image alignment, and then aggregate the results to obtain reasonable self-feedback. 5) Learning from Self-Feedback. For discrete LMMs, we directly apply DPO based on pairwise samples from Step 2. As to continuous LMMs, we propose Kernel-based Continuous DPO (KC-DPO), inducing a quadruplet objective with kernel functions for pairwise distance regulation over continuous visual features. The above five steps can iteratively repeat until self-improvement performance converges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 445, + 528, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 445, + 528, + 456 + ], + "spans": [ + { + "bbox": [ + 328, + 445, + 528, + 456 + ], + "type": "text", + "content": "In summary, our main contributions are threefold:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 457, + 553, + 588 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 457, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 553, + 505 + ], + "type": "text", + "content": "- To our knowledge, we are the first to focus on the task of LMMs' self-improvement for T2I. We propose a model-agnostic self-improvement framework to enable LMMs to achieve high-quality self-feedback and learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 506, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 553, + 552 + ], + "type": "text", + "content": "- For continuous LMMs, we introduce a dropout-based strategy to diversify image representations, along with a continuous DPO approach, i.e., KC-DPO, to optimize LMMs with preference representation pairs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 554, + 553, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 554, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 554, + 553, + 588 + ], + "type": "text", + "content": "- We conduct extensive experiments on three compositional T2I benchmarks, demonstrating the superiority of SILMM, e.g., " + }, + { + "bbox": [ + 314, + 554, + 553, + 588 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 314, + 554, + 553, + 588 + ], + "type": "text", + "content": " improvements on T2I-CompBench++." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 601, + 400, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 601, + 400, + 613 + ], + "spans": [ + { + "bbox": [ + 314, + 601, + 400, + 613 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 623, + 553, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 553, + 682 + ], + "type": "text", + "content": "Compositional Text-to-Image Generation. Diffusion models [55, 57] have marked a significant advancement in T2I generation due to their stability and scalability. However, they still struggle with text-image alignment, such as attribute binding, counting error, and relation confu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 294, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 294, + 713 + ], + "type": "text", + "content": "Sampling diverse images at the decoder stage is inapplicable, as it can only optimize the decoder yet we aim to optimize LMMs to generate superior visual representations for text-image alignment in this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "2For simplicity, we denote LLMs outputting discrete and continuous visual representations as discrete and continuous LMMs, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18498" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 79, + 550, + 352 + ], + "blocks": [ + { + "bbox": [ + 76, + 79, + 550, + 352 + ], + "lines": [ + { + "bbox": [ + 76, + 79, + 550, + 352 + ], + "spans": [ + { + "bbox": [ + 76, + 79, + 550, + 352 + ], + "type": "image", + "image_path": "6950e498312f2f876f3561238d0d0cc017869b5dda0a32a6faeb0d364c754bc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 368, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 54, + 368, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 54, + 368, + 555, + 425 + ], + "type": "text", + "content": "Figure 2. Schematic illustration of SILMM, comprising five steps: 1) LMMs generate compositional prompts by sampling based on provided instructions. 2) Diverse representations and images are generated using either discrete nucleus sampling or the proposed continuous DivDrop. 3) LMMs divide each compositional prompt into semantic units and generate questions for each unit. 4) VQA is conducted to answer these questions, with the answers and likelihoods aggregated into alignment scores as self-feedback. 5) For alignment tuning, DPO is applied for discrete LMMs, while the proposed KC-DPO is used for continuous LMMs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 435, + 294, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 294, + 544 + ], + "type": "text", + "content": "sion [16, 51]. To enhance compositional T2I, some approaches intervene in language structures [16] or cross-attention mechanisms [6]. Other methods [17, 37, 38, 49] incorporate layout planning by LLMs or use multi-agent collaboration [46, 70]. Inspired by alignment successes in LLMs, recent work [5, 15, 67] applies RLHF [43] to optimize diffusion models. Despite the progress, they rely on inductive biases, extensive prompt engineering, or labor-intensive annotations, limiting flexibility and scalability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 544, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 700 + ], + "type": "text", + "content": "Large Multimodal Models. The pioneering LMMs [39, 79] integrate a visual encoder, e.g., CLIP [52], with LLMs as the foundation, showing impressive multimodal understanding capabilities. To extend LMMs to visual generation, recent approaches align diffusion models [13, 19, 72] with LLMs or train a single transformer [62, 68, 74, 76]. According to the form of output visual features, they can be divided into discrete visual tokenization methods [19, 50, 62, 68] and continuous visual representation methods [13, 64, 72]. While LLM integration enhances language understanding and supports flexible applications (e.g., interleaved multimodal generation [64]), compositional T2I in the context of LMMs remains underexplored." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Learning from AI Feedback. The high cost of collect" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 435, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 555 + ], + "type": "text", + "content": "ing human preference has spurred research into RLAIF [3]. Benefiting from the convenience and scalability, there have been a series of studies adopting RLAIF to tackle a range of NLP tasks [10, 32, 78] and vision-language understanding [69, 77]. Despite the thrilling success, they only focus on text generation, overlooking the potential of RLAIF in other modalities. In contrast, we explore self-improving LMMs by activating multimodal understanding abilities for T2I. Particularly, we propose continuous strategies meticulously tailored to continuous visual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 568, + 397, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 397, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 397, + 582 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 590, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 649 + ], + "type": "text", + "content": "In this section, we elaborate on the proposed method, including the SILMM framework with five steps and the iteration strategy (Sec. 3.1), as illustrated in Fig. 2. Afterward, we introduce the continuous KC-DPO applied to LMMs with continuous visual features in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 658, + 534, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 534, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 534, + 672 + ], + "type": "text", + "content": "3.1. Self-Improving Large Multimodal Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Step 1: Compositional Prompt Generation. We first divide compositional scenarios into four categories: Attribute (color, shape, texture), Layout (counting, spatial relation)," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18499" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Semantic Relation, and Complex Composition. Complex composition includes any possible composition of the first three. For attribute and layout, we prompt the LMM to separately generate common objects, attributes, numbers, and spatial relations, and then use templates to compose these concepts. For semantic relation and complex composition, we adopt in-context learning [12] to generate prompts. More details can be found in App. 6." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": "Step 2: Diverse Representation and Image generation. The purpose of this step is to sample diverse intermediate visual representations from the LLM backbone " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": " of an LMM, given a text prompt " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": ", which would be decoded into images with different qualities. These representations are denoted as " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{z_i,\\dots,z_M\\}" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "z_{i}\\sim \\pi (z|x)" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": ". For discrete LMMs [19], " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": " is a discrete visual sequence. We follow the common practice [43, 53] in language generation to obtain " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": ", by sampling with different random seeds during auto-regressive decoding. For continuous LMMs [13], the LLM can only output a fixed continuous visual feature, without diversity. To tackle this issue, we propose DropDiv. First, we insert the dropout operations in the last few MLP layers of LLMs, which introduces randomness and enables LLMs for sampling. During inference, we activate these dropout operations to output diverse representations by sampling: " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "z_{i}\\sim \\pi^{\\prime}(z|x)" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": " denotes a continuous visual feature and " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\pi^\\prime" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": " represents the LLM with activated dropout operations. Afterward, these diverse visual representations " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": " are decoded into images as " + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{y_1,\\dots,y_M\\}" + }, + { + "bbox": [ + 55, + 171, + 295, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 411, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 591 + ], + "type": "text", + "content": "Discussion. Unlike prior work [5, 15] focused on tuning diffusion models, our approach resorts to LLM backbones in LLMs to control image decoders (e.g., diffusion models) for better text-image alignment, centering on LLM backbone optimization. Our approach offers three key advantages: 1) LLMs demonstrate superior proficiency in prompt comprehension over text encoders [52, 54] commonly employed in diffusion models. Tuning LLM backbones may unlock their enormous potential for compositional T2I, especially in complex scenarios. 2) Tuning diffusion models is often constrained by efficiency challenges inherent to iterative likelihood estimation, whereas there have been well-established technologies [1, 40, 53] for LLM alignment. 3) Our method is orthogonal to existing methods to tune diffusion models, combining them may get further gains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Step 3: Decompositional Self-Questioning. To provide helpful feedback to the generated images, the LMM should first accurately assess text-image alignment, which requires strong compositional reasoning abilities. However, current advanced LMMs still suffer from compositional reasoning [42], such as spatial relation understanding [7] and counting [56]. To improve compositional reasoning, we introduce a divide-and-conquer strategy [77] for self-questioning. Specifically, the LMM first divides the given prompt " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " into atomic concepts (e.g., \"a white harp\") and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "relations (e.g., \"a pancake is on the left of a pasta\"), and then generates questions " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{q_1, \\dots, q_N\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " corresponding to a concept or relation. For simplicity, the generated questions are constrained to be yes/no questions (e.g., \"Is there a while harp?\", \"Is the pancake on the left of the pasta?\"). Refer to App. 7 for more details on prompt templates of self-questioning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "type": "text", + "content": "Step 4: VQA-based Self-Feedback. Taking a generated image " + }, + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "type": "text", + "content": " and all the questions " + }, + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 157, + 553, + 216 + ], + "type": "text", + "content": " as input, the LMM conducts the VQA task, and the average difference between the probabilities of answering \"yes\" and \"no\" serves as the text-image alignment score:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 224, + 553, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 224, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 326, + 224, + 553, + 258 + ], + "type": "interline_equation", + "content": "s (x, y) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ p \\left(\" y e s\" \\mid y, q _ {i}\\right) - p \\left(\" n o\" \\mid y, q _ {i}\\right) \\right]. \\tag {1}", + "image_path": "f721ae1f4dd67d0175fd3011f3c8c107893c5a8d39d2f80bd69ea4a571cedc76.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": "Here we adopt the vision-language understanding abilities of LMMs via VQA to provide feedback to the images generated by themselves, thus this step is named VQA-based self-feedback. We carry out this step through all the sampled images prompted by " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": " and get all the scores " + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "inline_equation", + "content": "S = \\{ s(x,y_j) | y_j \\in \\mathcal{V} \\}" + }, + { + "bbox": [ + 313, + 266, + 553, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "content": "Step 5: Learning from Self-Feedback. Based on the selffeedback alignment scores, we sample representation pairs " + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "inline_equation", + "content": "(z_w,z_l)" + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "inline_equation", + "content": "z_{w}" + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "inline_equation", + "content": "z_{l}" + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "content": " denote the chosen and the rejected representations and their corresponding decoded images should satisfy " + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "inline_equation", + "content": "s(x,y_w) > s(x,y_l)" + }, + { + "bbox": [ + 313, + 339, + 553, + 410 + ], + "type": "text", + "content": ". With the preference data, we optimize the LLM backbone with DPO [53]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 422, + 421, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 422, + 421, + 436 + ], + "spans": [ + { + "bbox": [ + 324, + 422, + 421, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P O}} = - \\mathbb {E} _ {(x, z _ {w}, z _ {l}) \\sim \\mathcal {D}}", + "image_path": "88c67476028df9a9f5de8f3d63ba7dec286f5dd5ac31131087140778b83f84de.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 338, + 437, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 437, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 338, + 437, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(z _ {w} | x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(z _ {w} | x\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(z _ {l} | x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(z _ {l} | x\\right)}\\right) \\right], \\tag {2}", + "image_path": "21da8909ed15bb99f72795903b3000b544cc38fd3057126f88a84589e3904cbb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "content": " denotes the training set, and " + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "content": " represent the policy and reference models, respectively. " + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "content": " is the sigmoid function, and " + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 472, + 553, + 519 + ], + "type": "text", + "content": " is a hyperparameter controlling the deviation from the reference model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "content": "Iterative Self-Improvement. After learning from self-feedback, the updated LMM becomes more likely to generate preferred representations that are decoded into images better aligned with the prompt. This improvement in overall text-image alignment motivates us to iterate the above five steps with the updated LMM as the new reference model. The iteration mechanism continues until the alignment performance converges. As the process is independent of human annotations and external models, it is cost-effective and scalable. More importantly, it showcases the potential for self-improvement in LMMs by harmonizing their understanding and generation capabilities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 541, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 541, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 541, + 685 + ], + "type": "text", + "content": "3.2. Continuous Direct Preference Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "At the step of learning from self-feedback, LMMs are optimized using the DPO objective as shown in Eqn. (2)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18500" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "The difference between discrete and continuous LMMs in this learning process lies in the calculation of the likelihood " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\pi (z|x)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". For discrete LMMs, " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\pi (z|x)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " can be straightforwardly obtained by the softmax categorical distribution. However, for continuous LMMs with unknown distribution modeling, calculating " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\pi (z|x)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is intractable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "type": "text", + "content": "Predictive Distribution with MC Dropout. MC Dropout [18] enables predictive distribution estimation via Monte Carlo simulation to calculate " + }, + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\pi(z|x)" + }, + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "type": "text", + "content": ". Specifically, the dropout layers3 in an LMM are activated during inference and the LMM performs forward propagation multiple times to get multiple outputs. Assuming a Gaussian distribution, we can estimate its parameters and calculate the likelihood " + }, + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\pi(z|x)" + }, + { + "bbox": [ + 55, + 144, + 295, + 275 + ], + "type": "text", + "content": " based on these outputs. However, such multi-forward estimation imposes a significant computational burden during training, making this approach insufficient and impractical." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": "Simplified Kernel-based Continuous DPO. Inspired by MC Dropout and motivated by its insufficiency issue, we propose a simplified method to achieve continuous DPO. Concretely, the intermediate representation " + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": " often performs as a feature matrix " + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{H} \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": " denote the sequence length and dimension. " + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{H}" + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": " can be attained by a Q-Former [13, 20] or from the last layer of the LMM in an autoregressive way [63, 64]. To estimate " + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\pi(\\pmb{H} | x)" + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": ", we first make a decomposition as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 390, + 295, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 390, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 113, + 390, + 295, + 423 + ], + "type": "interline_equation", + "content": "\\pi (\\boldsymbol {H} | x) = \\prod_ {i = 1} ^ {L} \\pi \\left(\\boldsymbol {h} _ {i} \\mid \\boldsymbol {H} _ {< i}, x\\right), \\tag {3}", + "image_path": "74b041d4eb9ddeb9fc88bbb3a64c33f0bde37f8691dcb776d6e108a3fb007581.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\pmb{h}_i \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": "-th feature vector. Based on the Gaussian assumption, we have:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 459, + 295, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 459, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 66, + 459, + 295, + 487 + ], + "type": "interline_equation", + "content": "\\pi \\left(\\boldsymbol {h} _ {i} \\mid \\boldsymbol {H} _ {< i}, x\\right) = \\frac {\\exp \\left[ - \\frac {1}{2} \\left(\\boldsymbol {h} _ {i} - \\boldsymbol {\\mu} _ {i}\\right) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\boldsymbol {h} _ {i} - \\boldsymbol {\\mu} _ {i}\\right) \\right]}{\\sqrt {(2 \\pi) ^ {D} \\left| \\boldsymbol {\\Sigma} _ {i} \\right|}}, \\tag {4}", + "image_path": "410e814313d9965d42ad5107ab2288ff5ed8068e870ddb22cac9706c981d891f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": " denote the mean vector and the covariance matrix, respectively. Furthermore, we further simplify and approximate this formula: 1) the mean vector is estimated by the direct output of the continuous LMM, i.e., " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i\\approx \\mathrm{LMM}(\\boldsymbol {x})[i]" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": ", and 2) the Gaussian distribution is isotropic and all dimensions share the same variance value " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma}" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i\\approx \\mathrm{diag}(\\sigma_1,\\dots,\\sigma_{\\mathrm{D}})" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\sigma_{1} = \\ldots = \\sigma_{D} = \\bar{\\sigma}" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma}" + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": " can be learnable or viewed as a hyperparameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 620 + ], + "type": "text", + "content": "We compute the simplified likelihood with Eqn. (4), obtain the joint one with Eqn. (3), and finally derive the continuous DPO based on Eqn. (2):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 625, + 295, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 625, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 61, + 625, + 295, + 677 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} \\mathcal {L} _ {\\mathrm {C - D P O}} = & - \\mathbb {E} _ {(x, \\boldsymbol {H} _ {w}, \\boldsymbol {H} _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\frac {\\beta}{2 \\bar {\\sigma} ^ {2}} (- \\| \\boldsymbol {H} - \\boldsymbol {H} _ {w} \\| _ {F} ^ {2} + \\| \\boldsymbol {H} _ {r} - \\boldsymbol {H} _ {w} \\| _ {F} ^ {2})\\right) \\right], \\end{array} \\tag {5}", + "image_path": "9831dbf85901898ada1a5935ae3b7339480e82259ea07fd553023d1e49a8b826.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " denotes the Frobenius norm, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "H_{r}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " represent the continuous feature matrices from the policy and reference LMMs, respectively. " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "H_{w}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "H_{l}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " refer to the chosen and rejected feature matrices, respectively. Compared with the MC dropout method, this objective only requires one forward pass, which is more efficient. We relegate more details of the derivation to App. 8. From Eqn. (5), we can see that this objective aims to adjust the relative distances within the quadruple " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "(\\pmb {H},\\pmb {H}_r,\\pmb {H}_w,\\pmb {H}_l)" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " and the distance metric is the Euclidean distance between two matrices. To further improve the flexibility, we generalize the continuous DPO objective to," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 214, + 553, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 214, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 319, + 214, + 553, + 266 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {K C - D P O}} = - \\mathbb {E} _ {(x, \\boldsymbol {H} _ {w}, \\boldsymbol {H} _ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\gamma (- k (\\boldsymbol {H}, \\boldsymbol {H} _ {w}) \\right. \\right. \\\\ \\left. \\left. + k \\left(\\boldsymbol {H} _ {r}, \\boldsymbol {H} _ {w}\\right) + k \\left(\\boldsymbol {H}, \\boldsymbol {H} _ {l}\\right) - k \\left(\\boldsymbol {H} _ {r}, \\boldsymbol {H} _ {l}\\right)\\right)\\right) \\Bigg ], \\tag {6} \\\\ \\end{array}", + "image_path": "52a8fcc5ac69f9cb5437a22468f00cb151ca6a941eb7c0465a1b17eae63c683d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\gamma = \\frac{\\beta}{2\\bar{\\sigma}^2}" + }, + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "type": "text", + "content": " controls the degree of adherence to the reference model, " + }, + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "type": "inline_equation", + "content": "k(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 277, + 553, + 361 + ], + "type": "text", + "content": " denotes a generalized distance measurement function. Considering it is similar to kernel methods [22, 60], we name the objective Kernel-based Continuous DPO (KC-DPO). In the following experiments section, we will discuss different distance functions and their influences on alignment performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 374, + 395, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 395, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 395, + 388 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 395, + 430, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 430, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 430, + 408 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 413, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 553, + 473 + ], + "type": "text", + "content": "Base Model Settings. We implement our method on DreamLLM (continuous LMM) [13] and SEED-LLaMA (discrete LMM) [19] for all experiments. We also apply our method to Emu-3 [68], the recent state-of-the-art discrete LMM. Details on DPO training are provided in App. 9." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 545 + ], + "type": "text", + "content": "Datasets. We curated a dataset of 16,000 prompts across four categories using LMM. In each DPO training iteration, images generated by the model in the previous iteration served as the training data for next DPO iteration, allowing for iterative self-improvement. Details on data creation are provided in App. 10." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "**Benchmarks.** We evaluate our method on three text-to-image alignment benchmarks and follow their default settings. T2I-CompBench++ [28] consists of 8,000 compositional text prompts organized into 4 main categories: attribute, layout, non-spatial, and complex compositions, further divided into 8 subcategories, including color, binding, binding, 2D/3D-spatial relationships, non-spatial relationships, numeracy, and complex compositions. TIFA [27] uses pre-generated question-answer pairs and a VQA model to evaluate generation results based on 4,081 diverse text prompts and 25,829 questions across 12 categories. DPG-Bench [26] comprises 1,065 densely descriptive prompts with an average token length of 83.91, presenting more complex scenarios with varied objects and rich adjectives." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 714 + ], + "type": "text", + "content": "3In fact, there is no dropout layer in most open-sourced LLMs (e.g., LLaMA series [14, 65, 66]), and a compromise solution is to introduce additional dropout layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18501" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 70, + 265, + 190 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 265, + 190 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 265, + 190 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 265, + 190 + ], + "type": "image", + "image_path": "83f0e9e544828bd594bc8e34656f52939070c923468dc348148a1e7d069c4c54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 297, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 297, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 297, + 237 + ], + "type": "text", + "content": "Figure 3. Performance improvement of iterative alignment tuning based on SEED-LLaMA and DreamLLM, across 8 detailed categories of T2I-CompBench++. Iter. 0 denotes the base models without alignment tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 245, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 200, + 258 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 587 + ], + "type": "text", + "content": "As shown in Tab. 1, we evaluate alignment performance of our method against T2I generative models and base LMMs on three compositional T2I benchmarks, including T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Key observations are as follows: 1) Although LMMs enable more flexible settings (e.g., in-context learning and interleaved multimodal generation) for image generation, they still underperform compared to specialized T2I models in terms of the basic alignment ability to follow prompts. It demonstrates that current LMMs may ignore the compositional text-image alignment during multimodal pre-training and fine-tuning. 2) Without human annotations or external models, the proposed SILMM method enhances alignment performance across all categories in three benchmarks over the base LMMs, improving both the discrete SEEDLLaMA and the continuous DreamLLM, verifying the effectiveness and the generalization of SILMM. 3) SEEDLLaMA shows greater self-improvement than DreamLLM, possibly due to its weaker baseline alignment and the stability of discrete DPO over continuous KC-DPO induced by a series of simplification, as discussed in Sec. 3.2. And 4) improvements are more challenging in layout, relation, and complex categories than in attribute categories. This difficulty arises partly because the basic generative ability in these categories is weak, making it difficult to obtain high-quality chosen samples. Besides, understanding compositional concepts remains a challenge for LMMs [7, 56]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 597, + 162, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 162, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 162, + 611 + ], + "type": "text", + "content": "4.3. In-depth Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 616, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 688 + ], + "type": "text", + "content": "To explore the efficacy of SILMM, we conduct extensive ablation studies and hyperparameter analyses. We first investigate the iteration process and data scaling, followed by an in-depth study of key components, including diversity strategies, decompositional self-questioning and answering for self-feedback, and KC-DPO." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Iterative Self-Improvement. As shown in Fig. 3, we conduct three iterations of self-improvement and assess" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "performance changes across eight detailed categories of T2I-CompBench++ [28]. The results show that SILMM achieves effective, consistent, and continuous improvements in text-image alignment, across most compositional categories. Notably, attribute categories (e.g., color, shape, and texture) exhibit the most significant gains, whereas the non-spatial category shows slower improvement. This slower progress may stem from CLIP score [23], which is less sensitive than other metrics. Finally, as the iteration progresses, improvement rates gradually decrease, indicating convergence. More iterative self-improvement experiments results can be found in App. 11." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 226, + 439, + 319 + ], + "blocks": [ + { + "bbox": [ + 321, + 226, + 439, + 319 + ], + "lines": [ + { + "bbox": [ + 321, + 226, + 439, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 226, + 439, + 319 + ], + "type": "image", + "image_path": "2614437d62fcf9e7434be36fbdf5f096e32042199a379bd96215b9f0fbb10a37.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 321, + 427, + 331 + ], + "lines": [ + { + "bbox": [ + 329, + 321, + 427, + 331 + ], + "spans": [ + { + "bbox": [ + 329, + 321, + 427, + 331 + ], + "type": "text", + "content": "(a) Num. of Training Prompts" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 228, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 440, + 228, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 440, + 228, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 440, + 228, + 553, + 319 + ], + "type": "image", + "image_path": "0d440124397b0c90faff956d1051b90b1cf838014c379a499ba7d8182de20a41.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 320, + 542, + 331 + ], + "lines": [ + { + "bbox": [ + 447, + 320, + 542, + 331 + ], + "spans": [ + { + "bbox": [ + 447, + 320, + 542, + 331 + ], + "type": "text", + "content": "(b) Num. of Pairs per Prompt" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 334, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 411 + ], + "type": "text", + "content": "Figure 4. Overall alignment scores of SEED-LLaMA with discrete DPO and DreamLLM with continuous KC-DPO, on T2I-CompBench++ with (a) varying numbers of generated prompts in the training data, and (b) different number of preference pairs sampled from 30 diverse generated images per prompt. " + }, + { + "bbox": [ + 313, + 334, + 555, + 411 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 313, + 334, + 555, + 411 + ], + "type": "text", + "content": " means we select the top-N and last-N images from 30 generated ones as the chosen and rejected, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": "Data Scale. The proposed SILMM method leverages self-synthesized data for tuning, allowing flexible adjustment of data scale according to practical needs and available computational resources. In Fig. 4, we investigate how data scale affects overall alignment performance (averaged across eight categories in T2I-CompBench++, focusing on two factors: the number of training prompts and the number of preference pairs per prompt. Results in Fig. 4a indicate that both LMMs show consistent improvement as data samples increase, showing the strong scalability of the proposed method. Besides, we generate 30 representations and images per prompt, and then select the top-" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " and last-" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " samples to construct " + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 313, + 414, + 556, + 617 + ], + "type": "text", + "content": " preference pairs (see Fig. 4b). Notably, the two LMMs perform differently. This may be because the continuous feature space, being larger and denser than the discrete space, requires denser data pairs to stabilize the optimization dynamics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": "Diversity Strategies. To synthesize high-quality preference pair data, the diversity strategy is crucial. An effective diversity strategy should maximize the potential of LMMs while ensuring sufficient variation among generated images. To explore various strategies, we compare the proposed " + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "inline_equation", + "content": "DropDiv" + }, + { + "bbox": [ + 313, + 617, + 557, + 713 + ], + "type": "text", + "content": " with three alternatives: Rephrase the original prompt, Explain the original prompt with added imaginative elements, and add Gaussian noises to the learnable" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18502" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 107, + 517, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Performance comparison and improvement of the proposed method for compositional text-to-image generation on T2I-CompBench++ [28], DPG-Bench [26], and TIFA [27]. Alignment scores are calculated using expert understanding models (e.g., VQA or object detection models) recommended by these benchmarks. Prompt rewriting in Emu3 [68] was not used for fair comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 107, + 517, + 301 + ], + "lines": [ + { + "bbox": [ + 91, + 107, + 517, + 301 + ], + "spans": [ + { + "bbox": [ + 91, + 107, + 517, + 301 + ], + "type": "table", + "html": "
MethodT2I-CompBench++ [28]DPG-Bench [26]TIFA [27]
AttributeLayoutNon-spatialComplexGlobalEntityAttributeRelationOtherAllAll
Text-to-Image Generative Models
SD-v1.5 [55]38.65---74.6374.2375.3973.4967.8163.1878.40
DALL-E 2 [55]58.63----------
SD-v2 [55]47.3630.5031.2733.8677.6778.1374.9180.7280.6668.09-
SD-v2.1 [55]50.57---------82.00
SDXL [45]52.8835.6231.1932.3783.2782.4380.9186.7680.4174.65-
PixArt-α [8]60.3136.7431.9734.3374.9779.3278.6082.5776.9671.11-
DALL-E 3 [4]70.0941.6330.0337.7390.9789.6188.3990.5889.8383.50-
Large Multimodal Models
SEED-LLaMA [19]19.2020.2928.8621.4665.5955.8761.9662.7759.4647.1266.74
SEED-LLaMA + Ours39.6025.1129.8228.2873.5570.4868.4974.7968.6457.3173.74
%Improvement106.25%23.77%3.33%31.78%12.14%26.15%10.54%19.15%15.44%21.63%10.49%
DreamLLM [13]22.9423.7428.7623.0174.4765.8663.8074.2446.0053.9369.91
DreamLLM + Ours39.9427.6329.0026.4376.2975.9169.2084.4160.0064.2275.38
%Improvement74.15%16.40%0.83%14.86%2.44%15.26%8.46%13.70%30.43%19.08%7.82%
Emu3 [68]44.7932.3030.1531.3284.1980.8182.7587.2350.8074.1981.86
Emu3 + Ours59.7136.0330.5133.9384.1981.5784.5289.0164.8077.4585.11
%Improvement33.30%11.57%1.19%8.33%0.00%0.94%2.14%2.04%27.56%4.39%3.97%
", + "image_path": "f3356dda046940924f2ed25fe33ce50f6524c2abf12f4fecc3ef663bfb05e09c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 93, + 303, + 225, + 415 + ], + "blocks": [ + { + "bbox": [ + 93, + 303, + 225, + 415 + ], + "lines": [ + { + "bbox": [ + 93, + 303, + 225, + 415 + ], + "spans": [ + { + "bbox": [ + 93, + 303, + 225, + 415 + ], + "type": "image", + "image_path": "b5410bb3d5febd78c8305bbb114ae59eac73ddd33c5231041c24ff85f280532b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 417, + 204, + 427 + ], + "lines": [ + { + "bbox": [ + 108, + 417, + 204, + 427 + ], + "spans": [ + { + "bbox": [ + 108, + 417, + 204, + 427 + ], + "type": "text", + "content": "(a) Color, Shape, and Texture" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 241, + 305, + 372, + 415 + ], + "blocks": [ + { + "bbox": [ + 241, + 305, + 372, + 415 + ], + "lines": [ + { + "bbox": [ + 241, + 305, + 372, + 415 + ], + "spans": [ + { + "bbox": [ + 241, + 305, + 372, + 415 + ], + "type": "image", + "image_path": "0b0b8c4058c035053301add10f33b5b5c81cfcb9dbda59a6672a7f3a493b9033.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 417, + 365, + 428 + ], + "lines": [ + { + "bbox": [ + 242, + 417, + 365, + 428 + ], + "spans": [ + { + "bbox": [ + 242, + 417, + 365, + 428 + ], + "type": "text", + "content": "(b) Spatial, 3D Spatial, and Numeracy" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 388, + 304, + 523, + 415 + ], + "blocks": [ + { + "bbox": [ + 388, + 304, + 523, + 415 + ], + "lines": [ + { + "bbox": [ + 388, + 304, + 523, + 415 + ], + "spans": [ + { + "bbox": [ + 388, + 304, + 523, + 415 + ], + "type": "image", + "image_path": "16231f133c5649e6ced2d0eee83e643ba68f3addd136d8906b86c958ca1c0b77.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 417, + 500, + 427 + ], + "lines": [ + { + "bbox": [ + 405, + 417, + 500, + 427 + ], + "spans": [ + { + "bbox": [ + 405, + 417, + 500, + 427 + ], + "type": "text", + "content": "(c) Non-spatial and Complex" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 77, + 464, + 276, + 583 + ], + "blocks": [ + { + "bbox": [ + 55, + 432, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 432, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 555, + 455 + ], + "type": "text", + "content": "Figure 5. Comparison of four methods for diverse continuous representation generation, with alignment scores evaluated on the validation set of T2I-CompBench++. For each prompt, DreamLLM generates ten diverse representations and corresponding images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 464, + 276, + 583 + ], + "lines": [ + { + "bbox": [ + 77, + 464, + 276, + 583 + ], + "spans": [ + { + "bbox": [ + 77, + 464, + 276, + 583 + ], + "type": "image", + "image_path": "689ba48b80820c28b2f7da355c5d1f74d4e168c8ef89f1e9dedae31f6bd481b2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 592, + 295, + 647 + ], + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 647 + ], + "type": "text", + "content": "Figure 6. Distribution of alignment scores and variation of maximum scores across different dropout rates (%Drop) of the proposed DropDiv method, evaluated on the T2I-CompBench++ validation set. For each prompt, DreamLLM generates ten diverse representations and corresponding images with DropDiv." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Dream Embeddings in DreamLLM [13] to create Noisy DreamEmb. As shown in Fig. 5, four strategies have different perturbation influences on different categories. For example, DropDiv could generate better samples in attribute," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 467, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 551 + ], + "type": "text", + "content": "non-spatial, and complex categories, but compromise in layout categories. To further examine the effects of Drop-Div, we conduct experiments across different dropout rates as shown in Fig. 6. Results indicate that higher dropout rates enhance diversity, but the alignment quality could be impaired. Therefore, achieving a good diversity-quality tradeoff remains challenging." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "Decompositional Self-Feedback. We perform ablation studies on question generation and alignment score calculation, as shown in Tab. 2. Compared to two variants Prompt-Q (appending or replacing the period with “?” at the end of each prompt) and Phrase-Q (segmenting each prompt into phrases using NLP tools [24]), Self-Questioning (Self-Q) achieves better alignment performance across most categories, demonstrating the effectiveness of leveraging language processing abilities of LMMs for text-image alignment evaluation. Additionally, we compare the proposed alignment score calculation from Eqn. (1) with two variants: Random Score and Ratio of “yes” (where a higher ratio indicates a higher score). Results show that our method" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18503" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 173, + 279, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": "Table 2. Ablation study on T2I-CompBench++ [28] and DPG-Bench [26] examining variations in Question Generation and VQA-based Alignment Score Calculation methods for self-feedback. Prompt-Q adds a “?” or replaces the period with a “?” at the end of each prompt. Phrase-Q involves dividing a prompt into phrases, each followed by a “?”. Self-Q instructs the LMM to generate questions for each prompt using in-context examples. Diff. of Prob. denotes the proposed alignment score calculation approach described in Eqn. (1)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 173, + 279, + 284 + ], + "lines": [ + { + "bbox": [ + 83, + 173, + 279, + 284 + ], + "spans": [ + { + "bbox": [ + 83, + 173, + 279, + 284 + ], + "type": "table", + "html": "
FeedbackT2I-CompBench++DPG-Bench
AttributeLayoutNon-spatialComplexAll
Baseline (DreamLLM [13])
-22.9423.7428.7623.0153.93
Question Generation
Prompt-Q33.6625.9328.1424.6760.13
Phrase-Q34.6324.9128.0125.4160.10
Self-Q34.8525.5128.8225.3160.95
VQA-based Alignment Score Calculation
Random Score23.4124.8128.6722.9553.89
Ratio of “yes”25.3623.5128.7324.0054.68
Diff. of Prob.34.8525.5128.8225.3160.95
", + "image_path": "1bdd2c3d7848ccdfa72ff78b154205c1f610e6423e8bfdc0c11ffcd749729cf1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 82, + 352, + 267, + 492 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 296, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 348 + ], + "type": "text", + "content": "Table 3. Ablation study on T2I-CompBench++ [28] to investigate different instantiation of the Kernel Function to calculate the continuous KC-DPO loss function to tune DreamLLM. Aggregation means we aggregate the 2D feature matrix (e.g., " + }, + { + "bbox": [ + 55, + 293, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 293, + 296, + 348 + ], + "type": "text", + "content": ") into 1D along the sequence dimension. Eucl. denotes Euclidean distance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 352, + 267, + 492 + ], + "lines": [ + { + "bbox": [ + 82, + 352, + 267, + 492 + ], + "spans": [ + { + "bbox": [ + 82, + 352, + 267, + 492 + ], + "type": "table", + "html": "
AggregationDistanceAttributeLayoutNon-spatialComplex
Baseline (DreamLLM [13])
--22.9423.7428.7623.01
Supervised Fine-tuning (SFT)
-Eucl.12.250.7516.4111.71
-Cos6.950.2916.7811.48
AvgPoolEucl.23.3123.8928.7623.22
AvgPoolCos23.1224.2028.7923.29
Continuous Kernel-based Direct Preference Optimization
-Eucl.23.6524.3428.8323.08
-Cos23.9724.1128.7723.21
MaxPoolEucl.23.7924.0428.8623.92
MaxPoolCos29.1825.0118.7212.27
AvgPoolEucl.26.7524.7028.9425.12
AvgPoolCos34.8525.5128.8225.31
", + "image_path": "49abbbb9b371371b2dd7a7db6e2834d67f5e2f54214fe5d53d31b765cf06c005.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 521 + ], + "type": "text", + "content": "achieves superior performance by considering the relative confidence between \"yes\" and \"no\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 666 + ], + "type": "text", + "content": "Kernel-based Continuous DPO. In Sec. 3.2, we introduce the KC-DPO to fine-tune LMMs with continuous representations. The implementation of kernel functions can be divided into Aggregation and Distance. To assess the impacts of different kernels, we conduct extensive comparison experiments, as shown in Tab. 3. We observe SFT slightly improves the alignment performance, while DPO yields more substantial gains across all metrics. These results show that kernel functions are crucial to KC-DPO, and an optimal choice could greatly enhance the efficiency of preference optimization in continuous feature space. Overall, AvgPool + Cos demonstrates the superior performance improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 167, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 167, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 167, + 685 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "To illustrate the improvements achieved by SILMM, Fig. 8 presents examples generated by SEED-LLaMA, Dream" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 439, + 165 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 439, + 165 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 439, + 165 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 439, + 165 + ], + "type": "image", + "image_path": "e0f47ca9a390922442803deee89f7647d4dcc9d48dc36b0f6dd92f942c7afab1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 167, + 414, + 177 + ], + "lines": [ + { + "bbox": [ + 343, + 167, + 414, + 177 + ], + "spans": [ + { + "bbox": [ + 343, + 167, + 414, + 177 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 343, + 167, + 414, + 177 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 343, + 167, + 414, + 177 + ], + "type": "text", + "content": " in Discrete DPO" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 73, + 553, + 165 + ], + "blocks": [ + { + "bbox": [ + 440, + 73, + 553, + 165 + ], + "lines": [ + { + "bbox": [ + 440, + 73, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 440, + 73, + 553, + 165 + ], + "type": "image", + "image_path": "595a31543c9a62e10a57d0d2656f10583ab1f115d301086164fcdf8841fb73b4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 167, + 543, + 177 + ], + "lines": [ + { + "bbox": [ + 446, + 167, + 543, + 177 + ], + "spans": [ + { + "bbox": [ + 446, + 167, + 543, + 177 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 446, + 167, + 543, + 177 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 446, + 167, + 543, + 177 + ], + "type": "text", + "content": " in Continuous KC-DPO" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 218, + 533, + 492 + ], + "blocks": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "text", + "content": "Figure 7. Hyperparameter sensitivity on four general categories of T2I-CompBench++, examining (a) " + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "text", + "content": " in discrete DPO for SEEDLLaMA, and (b) " + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "text", + "content": " in continuous KC-DPO for DreamLLM." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 218, + 533, + 492 + ], + "lines": [ + { + "bbox": [ + 328, + 218, + 533, + 492 + ], + "spans": [ + { + "bbox": [ + 328, + 218, + 533, + 492 + ], + "type": "image", + "image_path": "1dc0fccd90873afa413bdd93614ff965e05b5685c3af12456817877ad3fec3d7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 495, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 517 + ], + "type": "text", + "content": "Figure 8. Qualitative results from SEED-LLaMA, DreamLLM, and the proposed SILMM method, on T2I-CompBench++." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 562 + ], + "type": "text", + "content": "LLM, and SILMM, in Fig. 8 using prompts from T2I-CompBench++. These results showcase the effectiveness of SILMM across extensive compositional scenarios." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 592, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 703 + ], + "type": "text", + "content": "In this work, we present a self-improvement approach named SILMM to enhance text-image alignment within LMMs, introducing an iterative model-agnostic framework comprising five stages to enable high-quality self-feedback and alignment learning. For continuous LMMs, we propose a dropout-based strategy to diversify image representations and a continuous DPO method, KC-DPO, for optimizing LMMs with preference representation pairs. Extensive experiments validate the effectiveness and superiority of our SILMM framework." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18504" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. A general theoretical paradigm to understand learning from human preferences. In International Conference on Artificial Intelligence and Statistics, pages 4447-4455. PMLR, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "text", + "content": "[3] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science., 2(3):8, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 357 + ], + "type": "text", + "content": "[5] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICML Workshop on Structured Probabilistic Inference & Generative Modeling, 2023. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 401 + ], + "type": "text", + "content": "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics, 42(4):1-10, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 467 + ], + "type": "text", + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "text", + "content": "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "text", + "content": "[9] Xiaolin Chen, Xuemeng Song, Liqiang Jing, Shuo Li, Linmei Hu, and Liqiang Nie. Multimodal dialog systems with dual knowledge-enhanced generative pretrained language model. ACM Transactions on Information Systems, 42(2): 1-25, 2023. 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "text", + "content": "[10] Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al. Ultrafeedback: Boosting language models with scaled ai feedback. In International Conference on Machine Learning, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[11] Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, and Jingren Zhou. How abilities in large language models are affected by supervised fine-tuning data composition. arXiv preprint arXiv:2310.05492, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 74, + 554, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 554, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 554, + 94 + ], + "type": "text", + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "type": "text", + "content": "[13] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. In International Conference on Learning Representations, 2024. 2, 3, 4, 5, 7, 8, 15, 18" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 554, + 206 + ], + "type": "text", + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 274 + ], + "type": "text", + "content": "[15] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "type": "text", + "content": "[16] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In International Conference on Machine Learning, 2023. 1, 3, 17" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 343, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 554, + 397 + ], + "type": "text", + "content": "[17] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. arXiv preprint arXiv:2305.15393, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 442 + ], + "type": "text", + "content": "[18] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 445, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 487 + ], + "type": "text", + "content": "[19] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. In International Conference on Learning Representations, 2024. 1, 2, 3, 4, 5, 7, 18" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 554, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 554, + 542 + ], + "type": "text", + "content": "[20] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024.5, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 545, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 554, + 599 + ], + "type": "text", + "content": "[21] Yangyang Guo, Guangzhi Wang, and Mohan Kankanhalli. Pela: Learning parameter-efficient models with low-rank approximation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15699-15709, 2024. 17" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 601, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 554, + 643 + ], + "type": "text", + "content": "[22] Marti A. Hearst, Susan T Dumais, Edgar Osuna, John Platt, and Bernhard Scholkopf. Support vector machines. IEEE Intelligent Systems and their applications, 13(4):18-28, 1998. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 554, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 554, + 700 + ], + "type": "text", + "content": "[23] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 703, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 554, + 713 + ], + "type": "text", + "content": "[24] Matthew Honnibal, Ines Montani, Sofie Van Landeghem," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18505" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "text", + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 16" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 294, + 183 + ], + "type": "text", + "content": "[26] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 5, 6, 7, 8, 16, 17, 18, 21, 22" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "type": "text", + "content": "[27] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20406-20417, 2023. 5, 6, 7, 16, 18" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 304 + ], + "type": "text", + "content": "[28] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 5, 6, 7, 8, 16, 17, 18, 19, 20" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 370 + ], + "type": "text", + "content": "[29] Natasha Jaques, Shixiang Gu, Dzmitry Bahdanau, José Miguel Hernández-Lobato, Richard E Turner, and Douglas Eck. Sequence tutor: Conservative fine-tuning of sequence generation models with kl-control. In International Conference on Machine Learning, pages 1645-1654. PMLR, 2017. 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 425 + ], + "type": "text", + "content": "[30] Natasha Jaques, Judy Hanwen Shen, Asma Ghandehar-ioun, Craig Ferguson, Agata Lapedriza, Noah Jones, Shixiang Shane Gu, and Rosalind Picard. Human-centric dialog training via offline reinforcement learning. arXiv preprint arXiv:2010.05848, 2020. 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[31] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36: 36652-36663, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 295, + 547 + ], + "type": "text", + "content": "[32] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Ren Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, et al. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. In International Conference on Machine Learning, 2024. 3, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "type": "text", + "content": "[33] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 647 + ], + "type": "text", + "content": "[34] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "text", + "content": "[35] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[36] Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, and Mohan Kankanhalli. Attribute-driven disentangled" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "representation learning for multimodal recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia, page 9660-9669. ACM, 2024. 17" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "text", + "content": "[37] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Lm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 186 + ], + "type": "text", + "content": "[38] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. In International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[40] Yongshuai Liu, Jiaxin Ding, and Xin Liu. Ipo: Interior-point policy optimization under constraints. In Proceedings of the AAAI conference on artificial intelligence, pages 4940-4947, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "text", + "content": "[41] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. Advances in Neural Information Processing Systems, 36, 2024. 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 322, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 376 + ], + "type": "text", + "content": "[42] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "type": "text", + "content": "[43] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2, 3, 4, 14, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "text", + "content": "[44] Jing-Cheng Pang, Pengyuan Wang, Kaiyuan Li, Xiong-Hui Chen, Jiacheng Xu, Zongzhang Zhang, and Yang Yu. Language model self-improvement by reinforcement learning contemplation. In International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "text", + "content": "[46] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "text", + "content": "[47] Tianyi Qin, Bo Peng, Jianjun Lei, Jiahui Song, Liying Xu, and Qingming Huang. 3d-immc: Incomplete multi-modal 3d shape clustering via cross mapping and dual adaptive fusion. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 17" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[48] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1113, 2021. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18506" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "text", + "content": "[49] Leigang Qu, Shengqiong Wu, Hao Fei, Liqiang Nie, and TatSeng Chua. Layoutllm-t2i: Eliciting layout guidance from llm for text-to-image generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 643-654, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 171 + ], + "type": "text", + "content": "[50] Leigang Qu, Haochuan Li, Tan Wang, Wenjie Wang, Yongqi Li, Liqiang Nie, and Tat-Seng Chua. Unified text-to-image generation and retrieval. arXiv preprint arXiv:2406.05814, 2024.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "text", + "content": "[51] Leigang Qu, Wenjie Wang, Yongqi Li, Hanwang Zhang, Liqiang Nie, and Tat-Seng Chua. Discriminative probing and tuning for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7434-7444, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "text", + "content": "[53] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "type": "text", + "content": "[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "text", + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "text", + "content": "[56] Sunayana Rane, Alexander Ku, Jason Baldridge, Ian Tenney, Tom Griffiths, and Been Kim. Can generative multimodal models count to ten? In Proceedings of the Annual Meeting of the Cognitive Science Society, 2024. 2, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 558 + ], + "type": "text", + "content": "[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 592 + ], + "type": "text", + "content": "[58] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "text", + "content": "[59] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 17" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 658 + ], + "type": "text", + "content": "[60] J Shawe-Taylor. Kernel methods for pattern analysis. Cambridge University Press google schola, 2:181-201, 2004. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[61] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008-3021, 2020. 15" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[62] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[63] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In International Conference on Learning Representations, 2023. 5, 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 239 + ], + "type": "text", + "content": "[64] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14398-14409, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "type": "text", + "content": "[65] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[66] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 349, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 555, + 415 + ], + "type": "text", + "content": "[67] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[68] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 470, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 470, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 470, + 553, + 514 + ], + "type": "text", + "content": "[69] Yufei Wang, Zhanyi Sun, Jesse Zhang, Zhou Xian, Erdem Biyik, David Held, and Zackory Erickson. Rl-vlm-f: Reinforcement learning from vision language foundation model feedback. arXiv preprint arXiv:2402.03681, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 514, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 559 + ], + "type": "text", + "content": "[70] Zhenyu Wang, Enze Xie, Aoxue Li, Zhongdao Wang, Xihui Liu, and Zhenguo Li. Divide and conquer: Language models can plan and self-correct for compositional text-to-image generation. arXiv preprint arXiv:2401.15688, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 559, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 623 + ], + "type": "text", + "content": "[71] Haokun Wen, Xuemeng Song, Xin Yang, Yibing Zhan, and Liqiang Nie. Comprehensive linguistic-visual composition network for image retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1369–1378, 2021. 17" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 658 + ], + "type": "text", + "content": "[72] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In International Conference on Machine Learning, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "text", + "content": "[73] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Better aligning text-to-image models with human preference. arXiv preprint arXiv:2303.14420, 1(3), 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[74] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18507" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[76] Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multimodal models: Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2(3), 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[77] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 2, 3, 4, 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "text", + "content": "[78] Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. arXiv preprint arXiv:2401.10020, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[79] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[80] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 15" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18508" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_content_list.json b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..76d6c6f663723b1542268eb02c95917813760151 --- /dev/null +++ b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_content_list.json @@ -0,0 +1,1226 @@ +[ + { + "type": "text", + "text": "SINR: Sparsity Driven Compressed Implicit Neural Representations", + "text_level": 1, + "bbox": [ + 153, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dhananjaya Jayasundara, Sudarshan Rajagopalan, Yasiru Ranasinghe, Trac D. Tran, and Vishal M. Patel Johns Hopkins University", + "bbox": [ + 220, + 180, + 772, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{vjayasul, sambasa2, dranasil, trac, vpatel36}@jhu.edu", + "bbox": [ + 251, + 236, + 730, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit Neural Representations (INRs) are increasingly recognized as a versatile data modality for representing discretized signals, offering benefits such as infinite query resolution and reduced storage requirements. Existing signal compression approaches for INRs typically employ one of two strategies: 1. direct quantization with entropy coding of the trained INR; 2. deriving a latent code on top of the INR through a learnable transformation. Thus, their performance is heavily dependent on the quantization and entropy coding schemes employed. In this paper, we introduce SINR, an innovative compression algorithm that leverages the patterns in the vector spaces formed by weights of INRs. We compress these vector spaces using a high-dimensional sparse code within a dictionary. Further analysis reveals that the atoms of the dictionary used to generate the sparse code do not need to be learned or transmitted to successfully recover the INR weights. We demonstrate that the proposed approach can be integrated with any existing INR-based signal compression technique. Our results indicate that SINR achieves substantial reductions in storage requirements for INRs across various configurations, outperforming conventional INR-based compression baselines. Furthermore, SINR maintains high-quality decoding across diverse data modalities, including images, occupancy fields, and Neural Radiance Fields.", + "bbox": [ + 91, + 319, + 483, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 744, + 220, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the fact that all naturally occurring signals observed by humans are continuous, capturing these signals through digital devices requires their discretization. For example, an image of a mountain is processed and stored in a discretized format. A primary reason for this approach is to conserve storage space; storing signals with high precision in an almost continuous manner would necessitate a sub", + "bbox": [ + 89, + 770, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stantial amount of storage. Consequently, the digital representation of signals is both practical and essential. For instance, it is estimated that over 400TB of data is created every day [8]. Moreover, humans share their captured signals through various mediums on a daily basis. Therefore, data compression is essential for efficient transmission.", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional signal compression techniques often rely on classic signal processing methods and are modality-specific. For example, JPEG [39], designed for photographic images, and is unsuitable for audio files. Similarly, audio compression standards like MP3 or AAC [4] are optimized for sound and are not applicable to images. With the advancements of neural networks, researchers have explored compressing signals using neural methods, predominantly through mechanisms based on autoencoders [2, 6, 7, 25, 37, 41]. In these systems, the encoder transforms the signal into a latent vector, which the decoder then uses to reconstruct the original signal. While autoencoder-based methods effectively encode signals into latent vectors, they are generally designed for a single modality. Adapting these methods to different data modalities not only requires training on a large corpus of data specific to those modalities but also a specialized autoencoder architecture tailored to handle the data effectively.", + "bbox": [ + 511, + 382, + 908, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a significant surge in interest in representing signals through Implicit Neural Representations (INRs). Unlike large models based on autoencoders, INRs typically consist of multi-layer perceptrons (MLPs) equipped with specialized nonlinearities that differ from the conventional nonlinearities used in deep learning. This simplicity and versatility allow INRs to unify signal representations across diverse data modalities. When signals are represented by INRs, they are encoded in the MLPs' weights and biases. For instance, in image transmission, instead of using conventional JPEG encoding, the weights and biases of the MLP are transmitted by a transmitter (TX). A receiver (RX) can then feed the coordinates into the MLP and decode the image. The primary advantage of INRs lies in their ability to represent signals with high fidelity while utilizing fewer parameters than parameter-", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project Page: https://dsgrad.github.io/SINR", + "bbox": [ + 112, + 887, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3061", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "heavy autoencoder-based mechanisms.", + "bbox": [ + 89, + 90, + 349, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in INR-based signal compression include COIN [9], COIN++ [10], INRIC [34], and SHACIRA [11]. COIN pioneered the application of INRs for image compression. Building on this, COIN++ and INRIC introduced quantization and entropy coding to improve compression efficiency. Both approaches also focus on enhancing the generalization capabilities of INRs through meta-learning techniques. Additionally, COIN++ incorporates latent modulations discovered via a learnable transformation applied on top of the INR model. However, COIN++ requires transmitting the base INR and the learned transformation apriori, in addition to the latent modulations for signal decoding. Alternatively, SHACIRA applies quantization on the latent weights and enforces entropy regularization to reparameterize feature grids to enable efficient compression across diverse domains. None of the existing methods; however, have explored fundamentally compressing the INR by identifying patterns within its parameter space before applying standard techniques such as quantization and entropy coding.", + "bbox": [ + 89, + 106, + 483, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our work, named SINR, we build upon the observed behaviors of the vector spaces generated by the weights in an INR. We integrate compressed sensing algorithms into the INR-based compression pipeline, proposing a mechanism that obtains a higher-dimensional sparse code for the weight vectors without requiring any learnable transformations. Furthermore, based on the Central Limit Theorem (CLT) [42], the transformation matrix can be directly sampled from a distribution rather than learning or hand crafting it. This eliminates the need to transmit the transformation matrix for accurately decoding weight spaces. Consequently, SINR, as a fundamental compression technique built on the observations of weights spaces, achieves superior compression and higher decoding quality for each data modality compared to the competing methods. Moreover, it can be easily embedded into any INR-based signal compression algorithm.", + "bbox": [ + 89, + 407, + 483, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 672, + 236, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Implicit neural representations", + "text_level": 1, + "bbox": [ + 89, + 696, + 366, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "INRs have recently gained considerable attention in the computer vision community due to their streamlined network architectures and improved performance in various vision tasks compared to traditional, parameter-heavy models [13, 29, 32]. This surge in interest followed the advent of Neural Radiance Fields (NeRF) [24], which has inspired a plethora of subsequent studies [26, 45]. Further research has explored the pivotal role of different activation functions in INRs [28, 29, 32, 35]. Moreover, INRs provide a universal framework for representing various data modalities. More recent studies have investigated the use of INRs for image classification by transforming standard image for", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9f7208e65004e9cbb4c7df590ea7aa703522b4e875f67b2b1af56beeff83f21e.jpg", + "image_caption": [ + "(a) INR weight distributions for image representation." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 704, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d7f5c994c8b5cd05b203b8c09a25c6001ba8deef94905b094da92bc8cbf1c56a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 89, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/99aada467b8b772a54a64fecb40beea654efd40ddd8cf02af08f320ef9d28b7c.jpg", + "image_caption": [ + "(b) INR weight distribution for occupancy fields (left) and NeRF (right)." + ], + "image_footnote": [], + "bbox": [ + 517, + 214, + 704, + 319 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9795a26df927fde903a81d644b99e10cc7b2b45aeeea5a0b119cbf0a29f0dd81.jpg", + "image_caption": [ + "Figure 1. Weight distribution of INRs tends to follow a Gaussian distribution for various data modalities." + ], + "image_footnote": [], + "bbox": [ + 717, + 214, + 903, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mats into INRs and training classifiers directly on the INRs' weights and biases [31]. These innovative approaches have showcased the potential of INRs to significantly reduce the dimensionality and computational complexity typically associated with conventional image processing techniques.", + "bbox": [ + 511, + 393, + 905, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Signal compression", + "text_level": 1, + "bbox": [ + 511, + 481, + 697, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Signal compression is crucial for reducing bandwidth needs and saving storage space. With the rise of deep learning, signal compression has evolved into two main approaches: rule-based (traditional) and learning-based methods. Traditional compression methods, such as JPEG for images and MP3 for audio, rely on algorithmic techniques tailored to specific signal types. JPEG minimizes redundancies using the discrete cosine transform [27], while MP3 [4] employs a psycho-acoustic model that enhances compression by removing inaudible sounds through auditory masking. On the other hand, deep learning-based techniques use models trained on vast datasets, adapting to a wide range of signals without predefined algorithms. These methods offer flexibility but require different architectures for each data modality, presenting unique challenges. In this landscape, INRs stand out as a potential universal signal representative.", + "bbox": [ + 511, + 503, + 906, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Compressed sensing", + "text_level": 1, + "bbox": [ + 511, + 757, + 705, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compressed sensing is a field that capitalizes on the inherent sparsity of data to capture information efficiently. In digital imaging, not every pixel is crucial for accurate image reconstruction. Although images appear dense in pixel space, they exhibit considerable redundancy when transformed into different basis functions. This sparsity is exploited by compressed sensing algorithms to reconstruct the original image from fewer sampled data points. These algo", + "bbox": [ + 511, + 779, + 905, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3062", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7715bfd98d66a74aa851b04fdf64e542e8c0622526ae7901a192935388b90475.jpg", + "image_caption": [ + "Figure 2. The proposed SINR compression algorithm: Standard compression techniques for INRs typically involve direct quantization and entropy coding of their weights. However, since natural signals exhibit inherent compressibility in a dictionary, the characteristics that aid in the compressibility of the weight space of an INR are discovered through the Gaussian nature of the weight space. Therefore, SINR employs $L_{1}$ minimization to identify a higher-dimensional sparse code. Furthermore, based on the weight space observations and the CLT, we simplify the encoding and decoding process using a random sensing matrix controlled by a seed. Subsequently, only the non-zero (NZ) values and their corresponding indices are quantized and entropy coded." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rithms employ optimization techniques and linear algebra to solve underdetermined systems, revolutionizing data acquisition in areas such as medical imaging and signal processing. Dictionary learning, integral to compressed sensing, seeks sparse representations of data using dictionary elements or atoms that capture the data's intrinsic structure. These atoms are either predefined or adaptively learned. Compressed sensing's versatility is evident in its applications across various domains, such as image and video compression [44], medical image encryption [17], and classification tasks [14, 15, 19, 21]. It also addresses inverse vision problems like image inpainting [30], deblurring [16, 23], and super-resolution [3]. Recent efforts have merged dictionary learning with deep learning to tackle more complex computer vision challenges, including image recognition [36], denoising [43], and scene recognition [22]. These developments underscore compressed sensing's transformative impact on computer vision.", + "bbox": [ + 88, + 455, + 480, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work, SINR, is pioneering the application of compressed sensing principles to INRs. By leveraging these principles alongside the structural distributions of INR weights, SINR identifies redundancies in these spaces, resulting in substantial compression improvements.", + "bbox": [ + 89, + 729, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 820, + 181, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Signal representation through INRs", + "text_level": 1, + "bbox": [ + 89, + 847, + 401, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mathematically, an INR can be defined by a function $G_{\\theta}$ where $\\theta$ are the estimizable parameters of the neural net", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work. The input and output dimensions of $G_{\\theta}$ vary for different data modalities. In general, $G_{\\theta}$ acts as a mapping from an $a$ -dimensional input coordinate space to a $b$ -dimensional output signal space, described as:", + "bbox": [ + 511, + 455, + 906, + 516 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\theta}: \\mathbb {R} ^ {a} \\to \\mathbb {R} ^ {b}.\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 525, + 764, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For instance, for RGB images, $a = 2$ and $b = 3$ , while for audio signals, $a = 1$ and $b = 1$ . In this architecture, the output of the $i^{\\text{th}}$ layer, which feeds into the $(i + 1)^{\\text{th}}$ layer, can be expressed as $\\sigma(W^{(i)}y^{(i)} + b^{(i)})$ . Here, $\\sigma$ denotes the activation function, and $y^{(i)}$ represents the output from the preceding layer. Furthermore, the choice of activation function $(\\sigma)$ plays a critical role in shaping the neural network's ability to model complex functions, as explored in various studies [28, 29, 32, 35].", + "bbox": [ + 511, + 551, + 906, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Exploring the compressibility", + "text_level": 1, + "bbox": [ + 511, + 696, + 777, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to compressed sensing theory, most real-world signals display sparsity when transformed into an appropriate domain, meaning they can be accurately represented with fewer measurements than traditionally required. Furthermore, real-world signals can be compressed through a set of basis functions, and the coefficients of these functions are derived by minimizing the reconstruction loss. The core concept of INRs involves encoding signals into the weights and biases of an MLP. This process can be viewed as a classical domain transformation technique where pixel values are reconstructed by feeding the corresponding coordinates through the MLP. Unlike predefined signal trans", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3063", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formers like Fourier [5] and DCT [20], the MLP attempts to minimize the reconstruction loss through backpropagation to find the transformation. The learned representation of the signal resides in another domain. Given that real-world signals inherently exhibit sparsity in transformed domains, we hypothesize that this sparsity can be explored within the MLP's weights. If we can identify where this sparse nature is hidden within the weight space, we could achieve further compression on INRs compared to competing methods. However, identifying this sparse representation within the weights is not straightforward. We believe there are two main approaches to achieving a sparser representation, each with its own challenges and considerations.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first approach involves either promoting or enforcing a specified level of sparsity in the weights during the training of an INR. Promoting sparsity can generally be achieved by incorporating $L_{1}$ regularization on the model parameters, which encourages many weights to approach zero, thereby creating a sparse representation. However, while we observed that $L_{1}$ regularization results in a higher level of sparsity within the weights, it fails to accurately represent the images. Alternatively, enforcing a specified level of sparsity can also be achieved through model pruning, where weights deemed insignificant are pruned or eliminated during the training process. When it comes to model pruning, we employed both structured and unstructured pruning of weights. We noted that both methods led to significant performance degradation for certain data modalities, particularly for occupancy fields. Moreover, only a small pruning percentages resulted in satisfactory performance for signal representation. For applications that require a high level of generalization, such as NeRFs, the pruning approach did not generalize well, indicating its limitations in achieving a balance between sparsity and performance.", + "bbox": [ + 91, + 292, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second approach seeks to uncover the inherent structures within the weights that aid in INR compression. This involves identifying patterns or regularities that can be exploited to reduce the dimensionality of the representation without sacrificing performance. We examined this from a dimensionality reduction perspective; however, the weight space in reduced dimensions did not reveal clear patterns, even across different natural images. Nonetheless, we observed that the weight space of an INR often tends to follow a normal distribution. Fig. 1 shows the weight distribution of the hidden layers of an INR when different data modalities are encoded into it. This suggests that INRs share a common pattern across different data modalities, showcasing a potential pathway for a fundamental compression. Given that each weight vector of an INR exhibits Gaussian behavior, we seek a higher-dimensional but sparse equivalent through a dictionary learning-based approach. Let us denote $\\mathbf{w} \\in \\mathbb{R}^{k_1}$ as a hidden weight vector, $\\mathbf{A} \\in \\mathbb{R}^{k_1 \\times k_2}$ as", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a dictionary, and $\\mathbf{x} \\in \\mathbb{R}^{k_2}$ as the corresponding sparse vector. In search of a sparse representation, according to standard compressed sensing, we can write $\\mathbf{w} = \\mathbf{A}\\mathbf{x}$ , where $\\| \\mathbf{x} \\|_0 < k_1$ . To discover the sparse code $\\mathbf{x}$ , the best and most efficient choice is $L_1$ minimization, as $L_0$ minimization iterates through all possible combinations and is therefore not efficient. However, the problem arises with the sensing matrix, commonly referred to as the dictionary $\\mathbf{A}$ . Although we could use either a dictionary learning-based approach for learning basis functions for the dictionary or a deep learning-based learnable transformation, these approaches would be time-consuming. Furthermore, a TX needs to transmit the learned dictionary alongside the obtained sparse codes. Conversely, we propose that the dictionary does not need to be learned or even transmitted.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we have confirmed, the weights are normally distributed. According to the CLT, a normally distributed random variable can be produced through a finite linear combination of any random variables. In summation form, this can be expressed as: $w_{i} = \\sum_{j = 1}^{k_{2}}A_{ij}x_{j}$ , where $w_{i}$ is the $i$ -th element of the weight vector $\\mathbf{w}$ , $A_{ij}$ is the element in the $i$ -th row and $j$ -th column of the sensing matrix $\\mathbf{A}$ , and $x_{j}$ is the $j$ -th element of the vector $\\mathbf{x}$ . To satisfy the CLT, the number of terms in the summation, which is $k_{2}$ , should be sufficiently large. Therefore, considering all elements of the weight vector $\\mathbf{w}$ , this can be compactly written as $\\mathbf{w} = \\mathbf{A}\\mathbf{x}$ . From CLT, the sensing matrix can be defined by a set of random vectors whose appropriate coefficients $(\\mathbf{x})$ can be learned using sparse signal recovery algorithms such as matching pursuit or its variants. Therefore, the optimization can be written as, min $\\| \\mathbf{x}\\| _1$ subject to $\\mathbf{w} = \\mathbf{A}\\mathbf{x}$ .", + "bbox": [ + 511, + 320, + 903, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For convenience, let us denote $\\| \\mathbf{x} \\|_0$ as $s$ . A further constraint to the above optimization procedure is that when the sparse code $\\mathbf{x}$ is found, we need to store not only its nonzero elements but also the corresponding indices. Therefore, the above $L_1$ minimization is solved with $2s < k_1$ . We do not apply our compression algorithm to the biases of the INR as the size of bias vectors is very small compared to those of the weight matrices.", + "bbox": [ + 511, + 566, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of saving $k_{1}$ floating-point numbers for $\\mathbf{w}$ , we now only need to save $2s$ elements: $s$ elements are floating-point numbers representing the non-zero values in the sparse code, and the remaining $s$ elements are integers that give the indices of those non-zero values. The indices can often be represented with 16-bit precision, unlike the nonzero values in the sparse code, which require 32-bit floating-point precision. At the RX end, $\\mathbf{x}$ must be converted back to $\\mathbf{w}$ . This requires the sensing matrix $\\mathbf{A}$ , which is random and can be controlled by a seed to reproduce the exact $\\mathbf{w}$ using $\\mathbf{w} = \\mathbf{A}\\mathbf{x}$ . Thus, the receiver only needs $\\mathbf{x}$ to obtain $\\mathbf{w}$ . This process can be viewed as a method of uncovering the inherent sparsity within natural signals, as represented through the weight space of INRs. As we hypothesized, the", + "bbox": [ + 511, + 689, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3064", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ability to condense natural signals into a dictionary hinges on identifying specific patterns encoded within the weights of INRs. Once the non-zero elements of the sparse vector are obtained at the RX, the resulting procedure is virtually the same across different INR-based baselines. Our method fundamentally achieves compression by delving into the weight spaces to uncover patterns, a step not typically taken by existing methods. A summarization of SINR is illustrated in Fig. 2. As can be seen from Fig. 2, SINR is only dependent on the weights of the INR and is applied prior to any quantization or entropy coding schemes. Therefore, SINR can be applied to existing INR compression methods to improve their compressibility.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. How much fundamental compression does SINR achieve?", + "text_level": 1, + "bbox": [ + 89, + 295, + 482, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1.Standard INRs", + "text_level": 1, + "bbox": [ + 89, + 332, + 243, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider an INR with $l$ hidden layers, yielding $l + 2$ total layers. For simplicity, assume $k$ neurons per hidden layer. If the input dimension is $a$ and the output dimension is $b$ , the total number of weight parameters is given by $\\mathcal{T}_s = a \\times k + l \\times k^2 + b \\times k$ . However, SINR modifies this structure by reducing the parameters from $\\mathcal{T}_s$ in the original network to $\\mathcal{T}_{s_{\\mathrm{INR}}} = a \\times 2s + k \\times l \\times 2s + b \\times 2s$ , where $s \\ll k$ . Additionally, SINR does not require transmitting any additional data to recover the original INR weights.", + "bbox": [ + 89, + 351, + 482, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Tiny INRs", + "text_level": 1, + "bbox": [ + 89, + 494, + 210, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us define an INR as \"tiny\" if the number of neurons in a hidden layer, denoted by $k$ , is less than 50. In such cases, we aim to achieve a sparse representation where $2s < k$ and $\\| x \\|_0 = s$ . However, achieving a sparse representation that satisfies $2s < k$ is often extremely challenging and typically does not result in effective compression. To overcome this, we exploit the fact that the weight matrix connecting the $i^{\\text{th}}$ layer to the $(i + 1)^{\\text{th}}$ layer is of dimensions $k \\times k$ . By vectorizing this weight matrix, we obtain a vector of dimension $k^2 \\times 1$ . Given that $k^2$ is significantly larger than $k$ , we can apply our SINR procedure directly to the flattened weight matrix. This strategy leads to a sparser representation, thus enhancing compression efficiency for tiny INRs.", + "bbox": [ + 89, + 512, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. COIN++", + "text_level": 1, + "bbox": [ + 89, + 715, + 197, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the COIN++ framework, modulation parameters are stored instead of traditional weights and biases, under the assumption that the base network parameters can be transmitted beforehand. For $n$ test images, each segmented into $m$ patches with a latent dimension of size $d$ , COIN++ necessitates the transmission of $m \\times d$ parameters for reconstructing each image. As the base network in COIN++ conforms to a standard INR structure, it is amenable to further compression via the SINR technique. By implementing SINR principles on the modulations in COIN++, the parameter transmission requirement per image can be reduced from", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$m \\times d$ to just $2s \\times d$ , where $s \\ll m$ . As the size of each test image and the number of images in the test dataset grow, COIN++ would typically require the transmission of numerous parameters. However, by leveraging SINR, both the modulations and the base network can be significantly compressed, achieving enhanced compression.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Quantization and entropy coding", + "text_level": 1, + "bbox": [ + 511, + 188, + 805, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After an INR is trained, its parameters are not immediately saved but are first subject to quantization [12]. This involves reducing the bitwidths below typical floating-point precision. Following quantization, the parameters are processed through entropy coding, in our experiments we utilize Brotli coding [1, 18], which allows the compressed data to be stored or transmitted efficiently. To retrieve the original parameters, the decoder must reverse the entropy coding and then perform dequantization. In the case of SINR, the compression process is intensified by utilizing the sparsity induced in the model parameters by natural signals. Once the sparse code is established, the parameters are quantized and subjected to entropy coding. The decoder then reverses the entropy coding and dequantizes the data. Finally, the model parameters are reconstructed by multiplying them with a random Gaussian matrix with a specific seed.", + "bbox": [ + 511, + 210, + 906, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 462, + 645, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 487, + 700, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SINR, a novel INR compression algorithm, is predicated on the idea that if natural signals are compressible through a dictionary, then INRs should be similarly compressible. This concept underpins SINR's goal to efficiently reduce INR storage requirements while maintaining high fidelity. Our experiments, conducted using the PyTorch framework following WIRE [29] codebase on an NVIDIA RTX A6000 GPU, spanned various data types including images, occupancy fields, and neural radiance fields. Image encoding metrics involved file size, bits per pixel (bpp) and Peak Signal-to-Noise Ratio (PSNR). Occupancy fields were evaluated using file size and Intersection over Union (IoU), and neural radiance fields were assessed using file size and PSNR. Other than the network configurations mentioned in the paper, for occupancy field evaluation, we utilized an MLP with 128 hidden neurons, and 3 hidden layers. For INRIC, we applied the network hyperparameters specified in its paper. In COIN++, we followed the guidelines in its paper but modified the hidden neuron size to 300. All experiments used Brotli entropy coding with a 16-bitwidth (65536 levels) uniform quantizer.", + "bbox": [ + 511, + 508, + 906, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. How do we find s?", + "text_level": 1, + "bbox": [ + 511, + 833, + 689, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implemented $L_{1}$ minimization using the Orthogonal Matching Pursuit (OMP) algorithm [38]. The OMP algorithm requires the pre-determination of $s$ before obtaining", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3065", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "x, and it must adhere to the condition $2s < k_{1}$ . If $2s$ is set too low, it results in inaccurate representations of $\\mathbf{w}$ within the weight space. Therefore, we incrementally increased $s$ from a low value until $2s = k_{1}$ for all KODAK images in the $C_1$ experiment, as outlined in Sec. 4.3. Our findings suggest that the optimal value of $s$ for successfully reconstructing the weight space does not depend on the specific image but on the number of neurons in a hidden layer. By adjusting the neuron count, we identified an optimal $s$ that accurately reconstructs the weight space while satisfying the specified constraint. Extending these experiments to natural signals outside the KODAK dataset confirmed the consistency of our results. Additionally, we have included a regression plot in the supplementary that details how to determine the optimal $s$ based on the number of neurons.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Image encoding", + "text_level": 1, + "bbox": [ + 89, + 330, + 248, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Representing an image through the weights and biases of a neural network serves as a method of encoding. For our image encoding task, we utilized the KODAK dataset, which includes 24 natural RGB images, each measuring $768 \\times 512$ pixels. We conducted five types of experiments, denoted as $C_i$ , where $i$ ranges from 1 to 5, to demonstrate the effectiveness of our proposed method.", + "bbox": [ + 89, + 354, + 482, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment $C_1$ involved encoding each image in the KODAK dataset using an INR without positional embedding, by varying the number of neurons in each hidden layer. Experiment $C_2$ mirrored $C_1$ , but with the variation in the number of hidden layers instead. Experiments $C_3$ and $C_4$ implemented the meta-learning approach for INRs proposed in INRIC, without and with positional embedding for the input layer, respectively. Experiment $C_5$ involved the COIN++ framework, testing both with and without patching. When using patching, we adopted $32 \\times 32$ patches as suggested by COIN++. However, we observed that without patching, even as the latent modulation dimension increased, the average PSNR obtained by COIN++ remained nearly constant. For these meta-learning-based experiments, we used the first 12 images of the KODAK dataset for meta-learning and the remaining 12 images for fine-tuning. For all image encoding experiments, we used the Sinusoidal activation function (see supplementary).", + "bbox": [ + 89, + 460, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let us define $h$ and $m$ as the number of hidden layers and the number of neurons per hidden layer in an INR, respectively. For experiment $C_1$ , we configured the INR with settings $(h,m)$ as $(2,32)$ , $(3,64)$ , $(3,128)$ . Experiment $C_1$ aims to assess the effectiveness of SINR by varying the number of hidden neurons. The results, depicted in Fig. 3, demonstrate how effectively SINR identifies the compressibility of the weight space. This is indicated by the bpp values, which reflect the size of the model parameters. For example, representing the KODAK dataset with an average PSNR of 30 dB requires about 3.7 bpp for COIN and 2.0", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6fb298a998e1ce91f83f6d74ab8f38c95f2f6d9414b5df8f797ecebe9d5efd68.jpg", + "image_caption": [ + "Figure 3. Experiments $C_1$ and $C_2$ : Identifying compressible INR combinations. The SINR approach demonstrates that configurations in $C_1$ are more compressible than those in $C_2$ . Furthermore, in both configurations SINR achieves lower bpp while maintaining the PSNR values." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bpp for INRIC. However, SINR significantly reduces the bpp to approximately 1.7 using the same quantizer and entropy coder. The first configuration in $C_1$ falls under the category of tiny INRs, underscoring the proposed method's effectiveness even for compact INRs. As illustrated in Fig. 3, SINR achieves the same level of PSNR as baselines with a lower bpp for any network configuration. This substantial reduction of bpp across the $C_1$ experiment showcases the efficiency and compactness achieved by SINR. From $C_1$ , it can be established that greater compressibility of an INR into a dictionary is possible with an increased number of hidden neurons. Following the conclusions drawn from experiment $C_1$ , experiment $C_2$ was designed to explore the impact of increasing the number of hidden layers on the effectiveness of SINR. The configurations tested in $C_2$ were $(h,m) = \\{(3,64),(5,64),(7,64)\\}$ . As illustrated in Fig. 3, SINR consistently achieved PSNR levels comparable to baseline methods, but with a reduced bpp. Given that $C_2$ maintained a constant neuron count at 64, the observed deviations in compression between SINR and INRIC were less significant than those observed in $C_1$ . This discrepancy can be attributed to the following: an INR configuration with a higher number of neurons (e.g., $m = 128$ ), even with fewer hidden layers (e.g., $h = 2$ ), possesses more trainable parameters. Consequently, such a model is capable of learning a more robust representation of the image compared to configurations with a larger number of layers but fewer neurons per layer. As a result, the compressible characteristics of the images are more effectively transferred into the model parameters during the INR training process. This leads to a more compressible INR. These findings support the premise that if natural images can be efficiently compressed into a dictionary, the weight space", + "bbox": [ + 511, + 401, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3066", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88323c6069865b12560ac00de8bb3b353d9a17ab289d138ce0eaec8b4dc95a3f.jpg", + "image_caption": [ + "Figure 4. Experiments $C_3$ , $C_4$ , and $C_5$ : Identifying compressible INR combinations under Meta-Learning. Meta-learning approaches have been introduced for INRs to enhance their generalization abilities and achieve faster convergence. When assessing induced sparsity in the weight space, SINR demonstrates a significant reduction in bpp values while maintaining nearly the same PSNR performance as the baselines." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of INRs can also be effectively compressed.", + "bbox": [ + 89, + 426, + 382, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As in previous experiments, each image required separate training of an INR. Experiments $C_3$ and $C_4$ address this challenge through meta-learning, with and without positional embedding, respectively. The configurations for these INRs are given by $(h,m) = \\{(3,32),(3,64),(3,96),(3,128)\\}$ . For COIN++, the number of layers was set to 5 with MLP's hidden dimension at 300. The latent dimension parameter $(d)$ varied as follows: $d = \\{16,32,64,96\\}$ . Fig. 4 presents the experimental results for $C_3$ , $C_4$ , and $C_5$ , illustrating significant compression capabilities of the proposed SINR within a meta-learning framework. Notably, models using positional embedding generally have more parameters than those without.", + "bbox": [ + 89, + 444, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing the performance of INRIC and SINR without positional embedding schemes, the initial INR configuration shows that SINR exhibits a lower bpp for the same average PSNR. Generally, as bpp increases, the representation capacity of the INR enhances, leading to more robust image representation. From the graphs, as bpp increases, SINR demonstrates greater improvements in PSNR than INRIC, a phenomenon that can be explained by the aforementioned logic. In the case of COIN++, the approach focuses on fine-tuning only the modulations using their proposed meta-learning method. However, since finetuning encodes natural signals within these modulations, they should also be compressible via a dictionary. Due to patching, each KODAK test image results in a $d \\times 384$ matrix. Our experiments reveal that these modulations encode hidden redundancies in natural signals. For instance, to achieve an average PSNR of approximately 24.2 dB,", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COIN++ requires more than 1.5 bpp; however, the same PSNR can be achieved with COIN++ using just under 1 bpp by exploiting the hidden sparsity in its modulations through our proposed approach. Therefore, when a high-capacity model effectively represents a signal, it must encapsulate this sparsity within its weight and bias spaces. SINR explores and removes redundancies in these parameters, retaining only essential information. Fig. 5 showcases the decoded images by SINR alongside with the INR based image compressors. Decoded PSNR, BPP, and file size are displayed in the first, second, and third rows of the text boxes.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Occupancy fields encoding", + "text_level": 1, + "bbox": [ + 511, + 266, + 756, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Occupancy fields are represented by binary values, either 1 or 0, where 1 denotes that the signal lies within a specified region and 0 indicates its absence. Another variant of occupancy volumes stores not only the presence or absence of a signal but also the color at that location. Typically, occupancy fields consume more space than other data modalities. However, they can be represented with higher accuracy and lower storage requirements using INRs. In this experiment, we followed the sampling procedure described in [29]. Occupancy fields can be thought of as representations of three-dimensional objects, capturing natural signals. Despite following the sampling procedure, redundancies may exist that are not essential for representing the occupancy volume. Identifying these redundancies can reduce storage requirements. However, identifying them in the spatial domain $(xyz)$ requires domain-specific algorithms, as described in Sec. 1. As INRs serve as unified data modality representatives, these redundancies must be encapsulated within its weights space. SINR fundamentally compresses the INRs into a dictionary regardless of the data modality; therefore, indeed it is equally applicable to occupancy fields. To validate this hypothesis for occupancy fields, two experiments were conducted using shapes from the Stanford shape dataset [33]. Figure 6 showcases the decoded SINR's representations for 'Thai Statue'(first volume) and 'Lucy' (fourth volume) datasets alongside the existing INR-based occupancy compressor. We use the Gaussian activation function for this task (see supplementary). The first value and second value in each text box represent the IoU metric and storage requirement, respectively, except for GT.", + "bbox": [ + 511, + 287, + 906, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Additional materials", + "text_level": 1, + "bbox": [ + 511, + 750, + 709, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The pseudocode for SINR, additional results and ablation studies on finding $s$ are in the supplementary material.", + "bbox": [ + 511, + 771, + 903, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 814, + 633, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implicit Neural Representations (INRs) have emerged as a promising framework for unified data modality representation. Several studies have explored the potential for compressing images, occupancy fields, and audio using INRs.", + "bbox": [ + 511, + 839, + 903, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3067", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83679984da93e168f814fc55fb8341af0f970ab79b9c631697abfb21be4c514a.jpg", + "image_caption": [ + "Figure 5. Results for image encoding experiment. SINR compresses the INR into a dictionary, significantly reducing the storage required compared to baseline INR image compressors. The results demonstrate that the decoded representations undergo a very negligible loss in PSNR, which is minimal considering the substantial storage space saved." + ], + "image_footnote": [], + "bbox": [ + 173, + 85, + 826, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45c9e37db49849fa961df1a6bd38e435297cb6285d7b30113c5b1768b2941228.jpg", + "image_caption": [ + "Figure 6. Results for occupancy fields encoding experiment. The results clearly demonstrate that SINR achieves the smallest file size and the highest accuracy metric for every shape in the tested dataset. The significant compression obtained by our algorithm suggests that occupancy fields, when represented using an INR, can be more efficiently compressed into a dictionary compared to images." + ], + "image_footnote": [], + "bbox": [ + 168, + 431, + 830, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, none of these methods have investigated whether the INR itself can be compressed prior to quantization and entropy coding. As natural signals can be efficiently compressed in bases of transformed domains due to their sparsity—allowing for higher accuracy and lower storage requirements—we hypothesize that a similar compressible nature must also exist in the INR once it is trained. With the discovery that weight vectors in the weight space tend to adhere to a Gaussian distribution, we propose SINR, which compresses any INR in a dictionary. Furthermore, we demonstrate that this dictionary does not need to be learned but can instead be generated using a seed. We compare our findings with standard INR compressors for im", + "bbox": [ + 89, + 704, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages, occupancy fields, and neural radiance fields. SINR achieves fundamental compression for any INR, independent of other post-processing methods such as quantization and entropy coding, and it showcases significantly lower storage requirements and higher fidelity across various data modalities. Through our experiments, we observed that the INR can be more compressed when a more robust representation of the signal is learned. Additionally, some data modalities exhibit greater compressibility than others. We firmly believe this research will aid other researchers in exploring more patterns in the weight spaces of INRs and in developing operators and transforms for INR.", + "bbox": [ + 511, + 704, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3068", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government.", + "bbox": [ + 89, + 114, + 483, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 294, + 187, + 310 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jyrki Alakuijala, Andrea Farruggia, Paolo Ferragina, Eugene Kliuchnikov, Robert Obryk, Zoltan Szabadka, and Lode Vandevenne. Brotli: A general-purpose data compressor. ACM Transactions on Information Systems (TOIS), 37(1):1-30, 2018. 5", + "[2] David Alexandre, Chih-Peng Chang, Wen-Hsiao Peng, and Hsueh-Ming Hang. An autoencoder-based learned image compressor: Description of challenge proposal by nctu. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2539-2542, 2018. 1", + "[3] Selen Ayas and Murat Ekinci. Single image super resolution using dictionary learning and sparse coding with multi-scale and multi-directional gabor feature representation. Information Sciences, 512:1264-1278, 2020. 3", + "[4] Karlheinz Brandenburg. Mp3 and aac explained. In Audio Engineering Society Conference: 17th International Conference: High-Quality Audio Coding. Audio Engineering Society, 1999. 1, 2", + "[5] Komaravolu Chandrasekharan. Classical Fourier Transforms. Springer Science & Business Media, 2012. 4", + "[6] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Energy compaction-based image compression using convolutional autoencoder. IEEE Transactions on Multimedia, 22(4):860-873, 2019. 1", + "[7] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1", + "[8] Fabio Duarte. Amount of data created daily (2024). https://explodingtopics.com/blog/data-generated-per-day, 2024. Accessed: 2024-07-14. 1", + "[9] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. arXiv preprint arXiv:2103.03123, 2021. 2", + "[10] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2" + ], + "bbox": [ + 93, + 320, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Sharath Girish, Abhinav Shrivastava, and Kamal Gupta. Shacira: Scalable hash-grid compression for implicit neural representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17513-17524, 2023. 2", + "[12] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 5", + "[13] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Implicit neural representations with levels-of-experts. Advances in Neural Information Processing Systems, 35:2564-2576, 2022. 2", + "[14] Daniel J Hsu, Sham M Kakade, John Langford, and Tong Zhang. Multi-label prediction via compressed sensing. Advances in neural information processing systems, 22, 2009. 3", + "[15] Junlin Hu and Yap-Peng Tan. Nonlinear dictionary learning with application to image classification. Pattern Recognition, 75:282-291, 2018. 3", + "[16] Zhe Hu, Jia-Bin Huang, and Ming-Hsuan Yang. Single image deblurring with adaptive dictionary learning. In 2010 IEEE International Conference on Image Processing, pages 1169-1172. IEEE, 2010. 3", + "[17] Donghua Jiang, Nestor Tsafack, Wadii Boulila, Jawad Ahmad, and JJ Barba-Franco. Asb-cs: Adaptive sparse basis compressive sensing model and its application to medical image encryption. Expert Systems with Applications, 236: 121378, 2024. 3", + "[18] Gareth A Jones and J Mary Jones. Information and coding theory. Springer Science & Business Media, 2012. 5", + "[19] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 3", + "[20] Syed Ali Khayam. The discrete cosine transform (dct): theory and application. Michigan State University, 114(1):31, 2003. 4", + "[21] Li Liu and Paul Fieguth. Texture classification using compressed sensing. In 2010 Canadian Conference on Computer and Robot Vision, pages 71-78. IEEE, 2010. 3", + "[22] Yang Liu, Qingchao Chen, Wei Chen, and Ian Wassell. Dictionary learning inspired deep network for scene recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3", + "[23] Liyan Ma, Lionel Moisan, Jian Yu, and Tieyong Zeng. A dictionary learning approach for poisson image deblurring. IEEE Transactions on medical imaging, 32(7):1277-1289, 2013. 3", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] AKM Rabby and Chengcui Zhang. Beyondpixels: A comprehensive review of the evolution of neural radiance fields. arXiv preprint arXiv:2306.03000, 2023. 2", + "[27] AM Raid, WM Khedr, Mohamed A El-Dosuky, and Wesam Ahmed. Jpeg image compression using discrete cosine transform-a survey. arXiv preprint arXiv:1405.6147, 2014. 2", + "[28] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 2, 3", + "[29] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2, 3, 5, 7", + "[30] K Seemakurthy, A Majumdar, J Gubbi, NK Sandeep, A Varghese, S Deshpande, M Girish Chandra, and P Balamurali. Deep dictionary learning for inpainting. In Computer Vision, Pattern Recognition, Image Processing, and Graphics: 7th National Conference, NCVPRIPG 2019, Hubballi, India, December 22-24, 2019, Revised Selected Papers 7, pages 79-88. Springer, 2020. 3", + "[31] Aviv Shamsian, Aviv Navon, David W Zhang, Yan Zhang, Ethan Fetaya, Gal Chechik, and Haggai Maron. Improved generalization of weight space networks via augmentations. arXiv preprint arXiv:2402.04081, 2024. 2", + "[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 3, 1", + "[33] Stanford University Computer Graphics Laboratory. The stanford 3d scanning repository. https://graphics.stanford.edu/data/3Dscanrep/. Accessed: 2024-07-15. 7", + "[34] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74-91. Springer, 2022. 2", + "[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 3", + "[36] Hao Tang, Hong Liu, Wei Xiao, and Nicu Sebe. When dictionary learning meets deep learning: Deep dictionary learning and coding network for image recognition with limited data. IEEE transactions on neural networks and learning systems, 32(5):2129-2141, 2020. 3", + "[37] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International conference on learning representations, 2022. 1", + "[38] Joel A Tropp and Anna C Gilbert. Signal recovery from random measurements via orthogonal matching pursuit." + ], + "bbox": [ + 91, + 92, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on information theory, 53(12):4655-4666, 2007. 5", + "[39] Gregory K Wallace. The jpeg still picture compression standard. IEEE transactions on consumer electronics, 38(1): xviii-xxxiv, 1992. 1", + "[40] Jan Hendrik Weissenbruch. The shipping canal at rijswijk, known as \"the view at geestbrug\". Painting. 2", + "[41] Yueqi Xie, Ka Leong Cheng, and Qifeng Chen. Enhanced invertible encoding for learned image compression. In Proceedings of the 29th ACM international conference on multimedia, pages 162-170, 2021. 1", + "[42] Xijuan Zhang, Oscar L Olvera Astivia, Edward Kroc, and Bruno D Zumbo. How to think clearly about the central limit theorem. Psychological Methods, 2022. 2", + "[43] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 630-641, 2021. 3", + "[44] Jinjia Zhou and Jian Yang. Compressive sensing in image/video compression: Sampling, coding, reconstruction, and codec optimization. Information, 15(2):75, 2024. 3", + "[45] Fang Zhu, Shuai Guo, Li Song, Ke Xu, Jiayu Hu, et al. Deep review and analysis of recent nerfs. APSIPA Transactions on Signal and Information Processing, 12(1), 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 433 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3070", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_model.json b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47508bffff1c167c0f9ef0db331eef5631360c65 --- /dev/null +++ b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_model.json @@ -0,0 +1,1727 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.844, + 0.154 + ], + "angle": 0, + "content": "SINR: Sparsity Driven Compressed Implicit Neural Representations" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.181, + 0.774, + 0.234 + ], + "angle": 0, + "content": "Dhananjaya Jayasundara, Sudarshan Rajagopalan, Yasiru Ranasinghe, Trac D. Tran, and Vishal M. Patel Johns Hopkins University" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.237, + 0.731, + 0.253 + ], + "angle": 0, + "content": "{vjayasul, sambasa2, dranasil, trac, vpatel36}@jhu.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.32, + 0.485, + 0.698 + ], + "angle": 0, + "content": "Implicit Neural Representations (INRs) are increasingly recognized as a versatile data modality for representing discretized signals, offering benefits such as infinite query resolution and reduced storage requirements. Existing signal compression approaches for INRs typically employ one of two strategies: 1. direct quantization with entropy coding of the trained INR; 2. deriving a latent code on top of the INR through a learnable transformation. Thus, their performance is heavily dependent on the quantization and entropy coding schemes employed. In this paper, we introduce SINR, an innovative compression algorithm that leverages the patterns in the vector spaces formed by weights of INRs. We compress these vector spaces using a high-dimensional sparse code within a dictionary. Further analysis reveals that the atoms of the dictionary used to generate the sparse code do not need to be learned or transmitted to successfully recover the INR weights. We demonstrate that the proposed approach can be integrated with any existing INR-based signal compression technique. Our results indicate that SINR achieves substantial reductions in storage requirements for INRs across various configurations, outperforming conventional INR-based compression baselines. Furthermore, SINR maintains high-quality decoding across diverse data modalities, including images, occupancy fields, and Neural Radiance Fields." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.745, + 0.222, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Despite the fact that all naturally occurring signals observed by humans are continuous, capturing these signals through digital devices requires their discretization. For example, an image of a mountain is processed and stored in a discretized format. A primary reason for this approach is to conserve storage space; storing signals with high precision in an almost continuous manner would necessitate a sub" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.379 + ], + "angle": 0, + "content": "stantial amount of storage. Consequently, the digital representation of signals is both practical and essential. For instance, it is estimated that over 400TB of data is created every day [8]. Moreover, humans share their captured signals through various mediums on a daily basis. Therefore, data compression is essential for efficient transmission." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.655 + ], + "angle": 0, + "content": "Traditional signal compression techniques often rely on classic signal processing methods and are modality-specific. For example, JPEG [39], designed for photographic images, and is unsuitable for audio files. Similarly, audio compression standards like MP3 or AAC [4] are optimized for sound and are not applicable to images. With the advancements of neural networks, researchers have explored compressing signals using neural methods, predominantly through mechanisms based on autoencoders [2, 6, 7, 25, 37, 41]. In these systems, the encoder transforms the signal into a latent vector, which the decoder then uses to reconstruct the original signal. While autoencoder-based methods effectively encode signals into latent vectors, they are generally designed for a single modality. Adapting these methods to different data modalities not only requires training on a large corpus of data specific to those modalities but also a specialized autoencoder architecture tailored to handle the data effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In recent years, there has been a significant surge in interest in representing signals through Implicit Neural Representations (INRs). Unlike large models based on autoencoders, INRs typically consist of multi-layer perceptrons (MLPs) equipped with specialized nonlinearities that differ from the conventional nonlinearities used in deep learning. This simplicity and versatility allow INRs to unify signal representations across diverse data modalities. When signals are represented by INRs, they are encoded in the MLPs' weights and biases. For instance, in image transmission, instead of using conventional JPEG encoding, the weights and biases of the MLP are transmitted by a transmitter (TX). A receiver (RX) can then feed the coordinates into the MLP and decode the image. The primary advantage of INRs lies in their ability to represent signals with high fidelity while utilizing fewer parameters than parameter-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.42, + 0.901 + ], + "angle": 0, + "content": "Project Page: https://dsgrad.github.io/SINR" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3061" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.35, + 0.106 + ], + "angle": 0, + "content": "heavy autoencoder-based mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.408 + ], + "angle": 0, + "content": "Recent advances in INR-based signal compression include COIN [9], COIN++ [10], INRIC [34], and SHACIRA [11]. COIN pioneered the application of INRs for image compression. Building on this, COIN++ and INRIC introduced quantization and entropy coding to improve compression efficiency. Both approaches also focus on enhancing the generalization capabilities of INRs through meta-learning techniques. Additionally, COIN++ incorporates latent modulations discovered via a learnable transformation applied on top of the INR model. However, COIN++ requires transmitting the base INR and the learned transformation apriori, in addition to the latent modulations for signal decoding. Alternatively, SHACIRA applies quantization on the latent weights and enforces entropy regularization to reparameterize feature grids to enable efficient compression across diverse domains. None of the existing methods; however, have explored fundamentally compressing the INR by identifying patterns within its parameter space before applying standard techniques such as quantization and entropy coding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.667 + ], + "angle": 0, + "content": "In our work, named SINR, we build upon the observed behaviors of the vector spaces generated by the weights in an INR. We integrate compressed sensing algorithms into the INR-based compression pipeline, proposing a mechanism that obtains a higher-dimensional sparse code for the weight vectors without requiring any learnable transformations. Furthermore, based on the Central Limit Theorem (CLT) [42], the transformation matrix can be directly sampled from a distribution rather than learning or hand crafting it. This eliminates the need to transmit the transformation matrix for accurately decoding weight spaces. Consequently, SINR, as a fundamental compression technique built on the observations of weights spaces, achieves superior compression and higher decoding quality for each data modality compared to the competing methods. Moreover, it can be easily embedded into any INR-based signal compression algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.238, + 0.689 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.367, + 0.714 + ], + "angle": 0, + "content": "2.1. Implicit neural representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "INRs have recently gained considerable attention in the computer vision community due to their streamlined network architectures and improved performance in various vision tasks compared to traditional, parameter-heavy models [13, 29, 32]. This surge in interest followed the advent of Neural Radiance Fields (NeRF) [24], which has inspired a plethora of subsequent studies [26, 45]. Further research has explored the pivotal role of different activation functions in INRs [28, 29, 32, 35]. Moreover, INRs provide a universal framework for representing various data modalities. More recent studies have investigated the use of INRs for image classification by transforming standard image for" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.705, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.09, + 0.904, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.201, + 0.852, + 0.213 + ], + "angle": 0, + "content": "(a) INR weight distributions for image representation." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.215, + 0.705, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.215, + 0.904, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.325, + 0.9, + 0.338 + ], + "angle": 0, + "content": "(b) INR weight distribution for occupancy fields (left) and NeRF (right)." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.35, + 0.906, + 0.378 + ], + "angle": 0, + "content": "Figure 1. Weight distribution of INRs tends to follow a Gaussian distribution for various data modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.473 + ], + "angle": 0, + "content": "mats into INRs and training classifiers directly on the INRs' weights and biases [31]. These innovative approaches have showcased the potential of INRs to significantly reduce the dimensionality and computational complexity typically associated with conventional image processing techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.698, + 0.499 + ], + "angle": 0, + "content": "2.2. Signal compression" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Signal compression is crucial for reducing bandwidth needs and saving storage space. With the rise of deep learning, signal compression has evolved into two main approaches: rule-based (traditional) and learning-based methods. Traditional compression methods, such as JPEG for images and MP3 for audio, rely on algorithmic techniques tailored to specific signal types. JPEG minimizes redundancies using the discrete cosine transform [27], while MP3 [4] employs a psycho-acoustic model that enhances compression by removing inaudible sounds through auditory masking. On the other hand, deep learning-based techniques use models trained on vast datasets, adapting to a wide range of signals without predefined algorithms. These methods offer flexibility but require different architectures for each data modality, presenting unique challenges. In this landscape, INRs stand out as a potential universal signal representative." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.707, + 0.774 + ], + "angle": 0, + "content": "2.3. Compressed sensing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Compressed sensing is a field that capitalizes on the inherent sparsity of data to capture information efficiently. In digital imaging, not every pixel is crucial for accurate image reconstruction. Although images appear dense in pixel space, they exhibit considerable redundancy when transformed into different basis functions. This sparsity is exploited by compressed sensing algorithms to reconstruct the original image from fewer sampled data points. These algo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.086, + 0.891, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.346, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. The proposed SINR compression algorithm: Standard compression techniques for INRs typically involve direct quantization and entropy coding of their weights. However, since natural signals exhibit inherent compressibility in a dictionary, the characteristics that aid in the compressibility of the weight space of an INR are discovered through the Gaussian nature of the weight space. Therefore, SINR employs \\( L_{1} \\) minimization to identify a higher-dimensional sparse code. Furthermore, based on the weight space observations and the CLT, we simplify the encoding and decoding process using a random sensing matrix controlled by a seed. Subsequently, only the non-zero (NZ) values and their corresponding indices are quantized and entropy coded." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.456, + 0.482, + 0.728 + ], + "angle": 0, + "content": "rithms employ optimization techniques and linear algebra to solve underdetermined systems, revolutionizing data acquisition in areas such as medical imaging and signal processing. Dictionary learning, integral to compressed sensing, seeks sparse representations of data using dictionary elements or atoms that capture the data's intrinsic structure. These atoms are either predefined or adaptively learned. Compressed sensing's versatility is evident in its applications across various domains, such as image and video compression [44], medical image encryption [17], and classification tasks [14, 15, 19, 21]. It also addresses inverse vision problems like image inpainting [30], deblurring [16, 23], and super-resolution [3]. Recent efforts have merged dictionary learning with deep learning to tackle more complex computer vision challenges, including image recognition [36], denoising [43], and scene recognition [22]. These developments underscore compressed sensing's transformative impact on computer vision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Our work, SINR, is pioneering the application of compressed sensing principles to INRs. By leveraging these principles alongside the structural distributions of INR weights, SINR identifies redundancies in these spaces, resulting in substantial compression improvements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.182, + 0.837 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.402, + 0.865 + ], + "angle": 0, + "content": "3.1. Signal representation through INRs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Mathematically, an INR can be defined by a function \\( G_{\\theta} \\) where \\( \\theta \\) are the estimizable parameters of the neural net" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.517 + ], + "angle": 0, + "content": "work. The input and output dimensions of \\( G_{\\theta} \\) vary for different data modalities. In general, \\( G_{\\theta} \\) acts as a mapping from an \\( a \\)-dimensional input coordinate space to a \\( b \\)-dimensional output signal space, described as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.526, + 0.766, + 0.543 + ], + "angle": 0, + "content": "\\[\nG _ {\\theta}: \\mathbb {R} ^ {a} \\to \\mathbb {R} ^ {b}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.69 + ], + "angle": 0, + "content": "For instance, for RGB images, \\( a = 2 \\) and \\( b = 3 \\), while for audio signals, \\( a = 1 \\) and \\( b = 1 \\). In this architecture, the output of the \\( i^{\\text{th}} \\) layer, which feeds into the \\( (i + 1)^{\\text{th}} \\) layer, can be expressed as \\( \\sigma(W^{(i)}y^{(i)} + b^{(i)}) \\). Here, \\( \\sigma \\) denotes the activation function, and \\( y^{(i)} \\) represents the output from the preceding layer. Furthermore, the choice of activation function \\( (\\sigma) \\) plays a critical role in shaping the neural network's ability to model complex functions, as explored in various studies [28, 29, 32, 35]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.779, + 0.715 + ], + "angle": 0, + "content": "3.2. Exploring the compressibility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "According to compressed sensing theory, most real-world signals display sparsity when transformed into an appropriate domain, meaning they can be accurately represented with fewer measurements than traditionally required. Furthermore, real-world signals can be compressed through a set of basis functions, and the coefficients of these functions are derived by minimizing the reconstruction loss. The core concept of INRs involves encoding signals into the weights and biases of an MLP. This process can be viewed as a classical domain transformation technique where pixel values are reconstructed by feeding the corresponding coordinates through the MLP. Unlike predefined signal trans" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3063" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "formers like Fourier [5] and DCT [20], the MLP attempts to minimize the reconstruction loss through backpropagation to find the transformation. The learned representation of the signal resides in another domain. Given that real-world signals inherently exhibit sparsity in transformed domains, we hypothesize that this sparsity can be explored within the MLP's weights. If we can identify where this sparse nature is hidden within the weight space, we could achieve further compression on INRs compared to competing methods. However, identifying this sparse representation within the weights is not straightforward. We believe there are two main approaches to achieving a sparser representation, each with its own challenges and considerations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.485, + 0.625 + ], + "angle": 0, + "content": "The first approach involves either promoting or enforcing a specified level of sparsity in the weights during the training of an INR. Promoting sparsity can generally be achieved by incorporating \\( L_{1} \\) regularization on the model parameters, which encourages many weights to approach zero, thereby creating a sparse representation. However, while we observed that \\( L_{1} \\) regularization results in a higher level of sparsity within the weights, it fails to accurately represent the images. Alternatively, enforcing a specified level of sparsity can also be achieved through model pruning, where weights deemed insignificant are pruned or eliminated during the training process. When it comes to model pruning, we employed both structured and unstructured pruning of weights. We noted that both methods led to significant performance degradation for certain data modalities, particularly for occupancy fields. Moreover, only a small pruning percentages resulted in satisfactory performance for signal representation. For applications that require a high level of generalization, such as NeRFs, the pruning approach did not generalize well, indicating its limitations in achieving a balance between sparsity and performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The second approach seeks to uncover the inherent structures within the weights that aid in INR compression. This involves identifying patterns or regularities that can be exploited to reduce the dimensionality of the representation without sacrificing performance. We examined this from a dimensionality reduction perspective; however, the weight space in reduced dimensions did not reveal clear patterns, even across different natural images. Nonetheless, we observed that the weight space of an INR often tends to follow a normal distribution. Fig. 1 shows the weight distribution of the hidden layers of an INR when different data modalities are encoded into it. This suggests that INRs share a common pattern across different data modalities, showcasing a potential pathway for a fundamental compression. Given that each weight vector of an INR exhibits Gaussian behavior, we seek a higher-dimensional but sparse equivalent through a dictionary learning-based approach. Let us denote \\(\\mathbf{w} \\in \\mathbb{R}^{k_1}\\) as a hidden weight vector, \\(\\mathbf{A} \\in \\mathbb{R}^{k_1 \\times k_2}\\) as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "a dictionary, and \\(\\mathbf{x} \\in \\mathbb{R}^{k_2}\\) as the corresponding sparse vector. In search of a sparse representation, according to standard compressed sensing, we can write \\(\\mathbf{w} = \\mathbf{A}\\mathbf{x}\\), where \\(\\| \\mathbf{x} \\|_0 < k_1\\). To discover the sparse code \\(\\mathbf{x}\\), the best and most efficient choice is \\(L_1\\) minimization, as \\(L_0\\) minimization iterates through all possible combinations and is therefore not efficient. However, the problem arises with the sensing matrix, commonly referred to as the dictionary \\(\\mathbf{A}\\). Although we could use either a dictionary learning-based approach for learning basis functions for the dictionary or a deep learning-based learnable transformation, these approaches would be time-consuming. Furthermore, a TX needs to transmit the learned dictionary alongside the obtained sparse codes. Conversely, we propose that the dictionary does not need to be learned or even transmitted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.905, + 0.565 + ], + "angle": 0, + "content": "As we have confirmed, the weights are normally distributed. According to the CLT, a normally distributed random variable can be produced through a finite linear combination of any random variables. In summation form, this can be expressed as: \\( w_{i} = \\sum_{j = 1}^{k_{2}}A_{ij}x_{j} \\), where \\( w_{i} \\) is the \\( i \\)-th element of the weight vector \\( \\mathbf{w} \\), \\( A_{ij} \\) is the element in the \\( i \\)-th row and \\( j \\)-th column of the sensing matrix \\( \\mathbf{A} \\), and \\( x_{j} \\) is the \\( j \\)-th element of the vector \\( \\mathbf{x} \\). To satisfy the CLT, the number of terms in the summation, which is \\( k_{2} \\), should be sufficiently large. Therefore, considering all elements of the weight vector \\( \\mathbf{w} \\), this can be compactly written as \\( \\mathbf{w} = \\mathbf{A}\\mathbf{x} \\). From CLT, the sensing matrix can be defined by a set of random vectors whose appropriate coefficients \\( (\\mathbf{x}) \\) can be learned using sparse signal recovery algorithms such as matching pursuit or its variants. Therefore, the optimization can be written as, min \\( \\| \\mathbf{x}\\| _1 \\) subject to \\( \\mathbf{w} = \\mathbf{A}\\mathbf{x} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.687 + ], + "angle": 0, + "content": "For convenience, let us denote \\( \\| \\mathbf{x} \\|_0 \\) as \\( s \\). A further constraint to the above optimization procedure is that when the sparse code \\( \\mathbf{x} \\) is found, we need to store not only its nonzero elements but also the corresponding indices. Therefore, the above \\( L_1 \\) minimization is solved with \\( 2s < k_1 \\). We do not apply our compression algorithm to the biases of the INR as the size of bias vectors is very small compared to those of the weight matrices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Instead of saving \\( k_{1} \\) floating-point numbers for \\( \\mathbf{w} \\), we now only need to save \\( 2s \\) elements: \\( s \\) elements are floating-point numbers representing the non-zero values in the sparse code, and the remaining \\( s \\) elements are integers that give the indices of those non-zero values. The indices can often be represented with 16-bit precision, unlike the nonzero values in the sparse code, which require 32-bit floating-point precision. At the RX end, \\( \\mathbf{x} \\) must be converted back to \\( \\mathbf{w} \\). This requires the sensing matrix \\( \\mathbf{A} \\), which is random and can be controlled by a seed to reproduce the exact \\( \\mathbf{w} \\) using \\( \\mathbf{w} = \\mathbf{A}\\mathbf{x} \\). Thus, the receiver only needs \\( \\mathbf{x} \\) to obtain \\( \\mathbf{w} \\). This process can be viewed as a method of uncovering the inherent sparsity within natural signals, as represented through the weight space of INRs. As we hypothesized, the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.289 + ], + "angle": 0, + "content": "ability to condense natural signals into a dictionary hinges on identifying specific patterns encoded within the weights of INRs. Once the non-zero elements of the sparse vector are obtained at the RX, the resulting procedure is virtually the same across different INR-based baselines. Our method fundamentally achieves compression by delving into the weight spaces to uncover patterns, a step not typically taken by existing methods. A summarization of SINR is illustrated in Fig. 2. As can be seen from Fig. 2, SINR is only dependent on the weights of the INR and is applied prior to any quantization or entropy coding schemes. Therefore, SINR can be applied to existing INR compression methods to improve their compressibility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.296, + 0.483, + 0.327 + ], + "angle": 0, + "content": "3.3. How much fundamental compression does SINR achieve?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.333, + 0.245, + 0.347 + ], + "angle": 0, + "content": "3.3.1.Standard INRs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Consider an INR with \\(l\\) hidden layers, yielding \\(l + 2\\) total layers. For simplicity, assume \\(k\\) neurons per hidden layer. If the input dimension is \\(a\\) and the output dimension is \\(b\\), the total number of weight parameters is given by \\(\\mathcal{T}_s = a \\times k + l \\times k^2 + b \\times k\\). However, SINR modifies this structure by reducing the parameters from \\(\\mathcal{T}_s\\) in the original network to \\(\\mathcal{T}_{s_{\\mathrm{INR}}} = a \\times 2s + k \\times l \\times 2s + b \\times 2s\\), where \\(s \\ll k\\). Additionally, SINR does not require transmitting any additional data to recover the original INR weights." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.212, + 0.51 + ], + "angle": 0, + "content": "3.3.2. Tiny INRs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.711 + ], + "angle": 0, + "content": "Let us define an INR as \"tiny\" if the number of neurons in a hidden layer, denoted by \\( k \\), is less than 50. In such cases, we aim to achieve a sparse representation where \\( 2s < k \\) and \\( \\| x \\|_0 = s \\). However, achieving a sparse representation that satisfies \\( 2s < k \\) is often extremely challenging and typically does not result in effective compression. To overcome this, we exploit the fact that the weight matrix connecting the \\( i^{\\text{th}} \\) layer to the \\( (i + 1)^{\\text{th}} \\) layer is of dimensions \\( k \\times k \\). By vectorizing this weight matrix, we obtain a vector of dimension \\( k^2 \\times 1 \\). Given that \\( k^2 \\) is significantly larger than \\( k \\), we can apply our SINR procedure directly to the flattened weight matrix. This strategy leads to a sparser representation, thus enhancing compression efficiency for tiny INRs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.199, + 0.73 + ], + "angle": 0, + "content": "3.3.3. COIN++" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In the COIN++ framework, modulation parameters are stored instead of traditional weights and biases, under the assumption that the base network parameters can be transmitted beforehand. For \\( n \\) test images, each segmented into \\( m \\) patches with a latent dimension of size \\( d \\), COIN++ necessitates the transmission of \\( m \\times d \\) parameters for reconstructing each image. As the base network in COIN++ conforms to a standard INR structure, it is amenable to further compression via the SINR technique. By implementing SINR principles on the modulations in COIN++, the parameter transmission requirement per image can be reduced from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\(m \\times d\\) to just \\(2s \\times d\\), where \\(s \\ll m\\). As the size of each test image and the number of images in the test dataset grow, COIN++ would typically require the transmission of numerous parameters. However, by leveraging SINR, both the modulations and the base network can be significantly compressed, achieving enhanced compression." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.189, + 0.806, + 0.207 + ], + "angle": 0, + "content": "3.4. Quantization and entropy coding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.907, + 0.453 + ], + "angle": 0, + "content": "After an INR is trained, its parameters are not immediately saved but are first subject to quantization [12]. This involves reducing the bitwidths below typical floating-point precision. Following quantization, the parameters are processed through entropy coding, in our experiments we utilize Brotli coding [1, 18], which allows the compressed data to be stored or transmitted efficiently. To retrieve the original parameters, the decoder must reverse the entropy coding and then perform dequantization. In the case of SINR, the compression process is intensified by utilizing the sparsity induced in the model parameters by natural signals. Once the sparse code is established, the parameters are quantized and subjected to entropy coding. The decoder then reverses the entropy coding and dequantizes the data. Finally, the model parameters are reconstructed by multiplying them with a random Gaussian matrix with a specific seed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.463, + 0.646, + 0.48 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.702, + 0.504 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.827 + ], + "angle": 0, + "content": "SINR, a novel INR compression algorithm, is predicated on the idea that if natural signals are compressible through a dictionary, then INRs should be similarly compressible. This concept underpins SINR's goal to efficiently reduce INR storage requirements while maintaining high fidelity. Our experiments, conducted using the PyTorch framework following WIRE [29] codebase on an NVIDIA RTX A6000 GPU, spanned various data types including images, occupancy fields, and neural radiance fields. Image encoding metrics involved file size, bits per pixel (bpp) and Peak Signal-to-Noise Ratio (PSNR). Occupancy fields were evaluated using file size and Intersection over Union (IoU), and neural radiance fields were assessed using file size and PSNR. Other than the network configurations mentioned in the paper, for occupancy field evaluation, we utilized an MLP with 128 hidden neurons, and 3 hidden layers. For INRIC, we applied the network hyperparameters specified in its paper. In COIN++, we followed the guidelines in its paper but modified the hidden neuron size to 300. All experiments used Brotli entropy coding with a 16-bitwidth (65536 levels) uniform quantizer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.691, + 0.849 + ], + "angle": 0, + "content": "4.2. How do we find s?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We implemented \\( L_{1} \\) minimization using the Orthogonal Matching Pursuit (OMP) algorithm [38]. The OMP algorithm requires the pre-determination of \\( s \\) before obtaining" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "x, and it must adhere to the condition \\(2s < k_{1}\\). If \\(2s\\) is set too low, it results in inaccurate representations of \\(\\mathbf{w}\\) within the weight space. Therefore, we incrementally increased \\(s\\) from a low value until \\(2s = k_{1}\\) for all KODAK images in the \\(C_1\\) experiment, as outlined in Sec. 4.3. Our findings suggest that the optimal value of \\(s\\) for successfully reconstructing the weight space does not depend on the specific image but on the number of neurons in a hidden layer. By adjusting the neuron count, we identified an optimal \\(s\\) that accurately reconstructs the weight space while satisfying the specified constraint. Extending these experiments to natural signals outside the KODAK dataset confirmed the consistency of our results. Additionally, we have included a regression plot in the supplementary that details how to determine the optimal \\(s\\) based on the number of neurons." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.249, + 0.348 + ], + "angle": 0, + "content": "4.3. Image encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.46 + ], + "angle": 0, + "content": "Representing an image through the weights and biases of a neural network serves as a method of encoding. For our image encoding task, we utilized the KODAK dataset, which includes 24 natural RGB images, each measuring \\(768 \\times 512\\) pixels. We conducted five types of experiments, denoted as \\(C_i\\), where \\(i\\) ranges from 1 to 5, to demonstrate the effectiveness of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Experiment \\(C_1\\) involved encoding each image in the KODAK dataset using an INR without positional embedding, by varying the number of neurons in each hidden layer. Experiment \\(C_2\\) mirrored \\(C_1\\), but with the variation in the number of hidden layers instead. Experiments \\(C_3\\) and \\(C_4\\) implemented the meta-learning approach for INRs proposed in INRIC, without and with positional embedding for the input layer, respectively. Experiment \\(C_5\\) involved the COIN++ framework, testing both with and without patching. When using patching, we adopted \\(32 \\times 32\\) patches as suggested by COIN++. However, we observed that without patching, even as the latent modulation dimension increased, the average PSNR obtained by COIN++ remained nearly constant. For these meta-learning-based experiments, we used the first 12 images of the KODAK dataset for meta-learning and the remaining 12 images for fine-tuning. For all image encoding experiments, we used the Sinusoidal activation function (see supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Let us define \\( h \\) and \\( m \\) as the number of hidden layers and the number of neurons per hidden layer in an INR, respectively. For experiment \\( C_1 \\), we configured the INR with settings \\( (h,m) \\) as \\( (2,32) \\), \\( (3,64) \\), \\( (3,128) \\). Experiment \\( C_1 \\) aims to assess the effectiveness of SINR by varying the number of hidden neurons. The results, depicted in Fig. 3, demonstrate how effectively SINR identifies the compressibility of the weight space. This is indicated by the bpp values, which reflect the size of the model parameters. For example, representing the KODAK dataset with an average PSNR of 30 dB requires about 3.7 bpp for COIN and 2.0" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.296, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Experiments \\( C_1 \\) and \\( C_2 \\): Identifying compressible INR combinations. The SINR approach demonstrates that configurations in \\( C_1 \\) are more compressible than those in \\( C_2 \\). Furthermore, in both configurations SINR achieves lower bpp while maintaining the PSNR values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.908, + 0.903 + ], + "angle": 0, + "content": "bpp for INRIC. However, SINR significantly reduces the bpp to approximately 1.7 using the same quantizer and entropy coder. The first configuration in \\( C_1 \\) falls under the category of tiny INRs, underscoring the proposed method's effectiveness even for compact INRs. As illustrated in Fig. 3, SINR achieves the same level of PSNR as baselines with a lower bpp for any network configuration. This substantial reduction of bpp across the \\( C_1 \\) experiment showcases the efficiency and compactness achieved by SINR. From \\( C_1 \\), it can be established that greater compressibility of an INR into a dictionary is possible with an increased number of hidden neurons. Following the conclusions drawn from experiment \\( C_1 \\), experiment \\( C_2 \\) was designed to explore the impact of increasing the number of hidden layers on the effectiveness of SINR. The configurations tested in \\( C_2 \\) were \\( (h,m) = \\{(3,64),(5,64),(7,64)\\} \\). As illustrated in Fig. 3, SINR consistently achieved PSNR levels comparable to baseline methods, but with a reduced bpp. Given that \\( C_2 \\) maintained a constant neuron count at 64, the observed deviations in compression between SINR and INRIC were less significant than those observed in \\( C_1 \\). This discrepancy can be attributed to the following: an INR configuration with a higher number of neurons (e.g., \\( m = 128 \\)), even with fewer hidden layers (e.g., \\( h = 2 \\)), possesses more trainable parameters. Consequently, such a model is capable of learning a more robust representation of the image compared to configurations with a larger number of layers but fewer neurons per layer. As a result, the compressible characteristics of the images are more effectively transferred into the model parameters during the INR training process. This leads to a more compressible INR. These findings support the premise that if natural images can be efficiently compressed into a dictionary, the weight space" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.485, + 0.396 + ], + "angle": 0, + "content": "Figure 4. Experiments \\( C_3 \\), \\( C_4 \\), and \\( C_5 \\): Identifying compressible INR combinations under Meta-Learning. Meta-learning approaches have been introduced for INRs to enhance their generalization abilities and achieve faster convergence. When assessing induced sparsity in the weight space, SINR demonstrates a significant reduction in bpp values while maintaining nearly the same PSNR performance as the baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.383, + 0.442 + ], + "angle": 0, + "content": "of INRs can also be effectively compressed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.642 + ], + "angle": 0, + "content": "As in previous experiments, each image required separate training of an INR. Experiments \\( C_3 \\) and \\( C_4 \\) address this challenge through meta-learning, with and without positional embedding, respectively. The configurations for these INRs are given by \\( (h,m) = \\{(3,32),(3,64),(3,96),(3,128)\\} \\). For COIN++, the number of layers was set to 5 with MLP's hidden dimension at 300. The latent dimension parameter \\( (d) \\) varied as follows: \\( d = \\{16,32,64,96\\} \\). Fig. 4 presents the experimental results for \\( C_3 \\), \\( C_4 \\), and \\( C_5 \\), illustrating significant compression capabilities of the proposed SINR within a meta-learning framework. Notably, models using positional embedding generally have more parameters than those without." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparing the performance of INRIC and SINR without positional embedding schemes, the initial INR configuration shows that SINR exhibits a lower bpp for the same average PSNR. Generally, as bpp increases, the representation capacity of the INR enhances, leading to more robust image representation. From the graphs, as bpp increases, SINR demonstrates greater improvements in PSNR than INRIC, a phenomenon that can be explained by the aforementioned logic. In the case of COIN++, the approach focuses on fine-tuning only the modulations using their proposed meta-learning method. However, since finetuning encodes natural signals within these modulations, they should also be compressible via a dictionary. Due to patching, each KODAK test image results in a \\(d \\times 384\\) matrix. Our experiments reveal that these modulations encode hidden redundancies in natural signals. For instance, to achieve an average PSNR of approximately 24.2 dB," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "COIN++ requires more than 1.5 bpp; however, the same PSNR can be achieved with COIN++ using just under 1 bpp by exploiting the hidden sparsity in its modulations through our proposed approach. Therefore, when a high-capacity model effectively represents a signal, it must encapsulate this sparsity within its weight and bias spaces. SINR explores and removes redundancies in these parameters, retaining only essential information. Fig. 5 showcases the decoded images by SINR alongside with the INR based image compressors. Decoded PSNR, BPP, and file size are displayed in the first, second, and third rows of the text boxes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.757, + 0.284 + ], + "angle": 0, + "content": "4.4. Occupancy fields encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Occupancy fields are represented by binary values, either 1 or 0, where 1 denotes that the signal lies within a specified region and 0 indicates its absence. Another variant of occupancy volumes stores not only the presence or absence of a signal but also the color at that location. Typically, occupancy fields consume more space than other data modalities. However, they can be represented with higher accuracy and lower storage requirements using INRs. In this experiment, we followed the sampling procedure described in [29]. Occupancy fields can be thought of as representations of three-dimensional objects, capturing natural signals. Despite following the sampling procedure, redundancies may exist that are not essential for representing the occupancy volume. Identifying these redundancies can reduce storage requirements. However, identifying them in the spatial domain \\((xyz)\\) requires domain-specific algorithms, as described in Sec. 1. As INRs serve as unified data modality representatives, these redundancies must be encapsulated within its weights space. SINR fundamentally compresses the INRs into a dictionary regardless of the data modality; therefore, indeed it is equally applicable to occupancy fields. To validate this hypothesis for occupancy fields, two experiments were conducted using shapes from the Stanford shape dataset [33]. Figure 6 showcases the decoded SINR's representations for 'Thai Statue'(first volume) and 'Lucy' (fourth volume) datasets alongside the existing INR-based occupancy compressor. We use the Gaussian activation function for this task (see supplementary). The first value and second value in each text box represent the IoU metric and storage requirement, respectively, except for GT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.75, + 0.71, + 0.765 + ], + "angle": 0, + "content": "4.5. Additional materials" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.905, + 0.803 + ], + "angle": 0, + "content": "The pseudocode for SINR, additional results and ablation studies on finding \\(s\\) are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.634, + 0.831 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Implicit Neural Representations (INRs) have emerged as a promising framework for unified data modality representation. Several studies have explored the potential for compressing images, occupancy fields, and audio using INRs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.086, + 0.827, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 5. Results for image encoding experiment. SINR compresses the INR into a dictionary, significantly reducing the storage required compared to baseline INR image compressors. The results demonstrate that the decoded representations undergo a very negligible loss in PSNR, which is minimal considering the substantial storage space saved." + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.432, + 0.831, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.636, + 0.908, + 0.679 + ], + "angle": 0, + "content": "Figure 6. Results for occupancy fields encoding experiment. The results clearly demonstrate that SINR achieves the smallest file size and the highest accuracy metric for every shape in the tested dataset. The significant compression obtained by our algorithm suggests that occupancy fields, when represented using an INR, can be more efficiently compressed into a dictionary compared to images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.901 + ], + "angle": 0, + "content": "However, none of these methods have investigated whether the INR itself can be compressed prior to quantization and entropy coding. As natural signals can be efficiently compressed in bases of transformed domains due to their sparsity—allowing for higher accuracy and lower storage requirements—we hypothesize that a similar compressible nature must also exist in the INR once it is trained. With the discovery that weight vectors in the weight space tend to adhere to a Gaussian distribution, we propose SINR, which compresses any INR in a dictionary. Furthermore, we demonstrate that this dictionary does not need to be learned but can instead be generated using a seed. We compare our findings with standard INR compressors for im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.888 + ], + "angle": 0, + "content": "ages, occupancy fields, and neural radiance fields. SINR achieves fundamental compression for any INR, independent of other post-processing methods such as quantization and entropy coding, and it showcases significantly lower storage requirements and higher fidelity across various data modalities. Through our experiments, we observed that the INR can be more compressed when a more robust representation of the signal is learned. Additionally, some data modalities exhibit greater compressibility than others. We firmly believe this research will aid other researchers in exploring more patterns in the weight spaces of INRs and in developing operators and transforms for INR." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3068" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.283 + ], + "angle": 0, + "content": "This work is supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.295, + 0.188, + 0.311 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[1] Jyrki Alakuijala, Andrea Farruggia, Paolo Ferragina, Eugene Kliuchnikov, Robert Obryk, Zoltan Szabadka, and Lode Vandevenne. Brotli: A general-purpose data compressor. ACM Transactions on Information Systems (TOIS), 37(1):1-30, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.392, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[2] David Alexandre, Chih-Peng Chang, Wen-Hsiao Peng, and Hsueh-Ming Hang. An autoencoder-based learned image compressor: Description of challenge proposal by nctu. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2539-2542, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.477, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[3] Selen Ayas and Murat Ekinci. Single image super resolution using dictionary learning and sparse coding with multi-scale and multi-directional gabor feature representation. Information Sciences, 512:1264-1278, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[4] Karlheinz Brandenburg. Mp3 and aac explained. In Audio Engineering Society Conference: 17th International Conference: High-Quality Audio Coding. Audio Engineering Society, 1999. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[5] Komaravolu Chandrasekharan. Classical Fourier Transforms. Springer Science & Business Media, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[6] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Energy compaction-based image compression using convolutional autoencoder. IEEE Transactions on Multimedia, 22(4):860-873, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[7] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[8] Fabio Duarte. Amount of data created daily (2024). https://explodingtopics.com/blog/data-generated-per-day, 2024. Accessed: 2024-07-14. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[9] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. arXiv preprint arXiv:2103.03123, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.321, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[11] Sharath Girish, Abhinav Shrivastava, and Kamal Gupta. Shacira: Scalable hash-grid compression for implicit neural representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17513-17524, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[12] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[13] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Implicit neural representations with levels-of-experts. Advances in Neural Information Processing Systems, 35:2564-2576, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[14] Daniel J Hsu, Sham M Kakade, John Langford, and Tong Zhang. Multi-label prediction via compressed sensing. Advances in neural information processing systems, 22, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[15] Junlin Hu and Yap-Peng Tan. Nonlinear dictionary learning with application to image classification. Pattern Recognition, 75:282-291, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[16] Zhe Hu, Jia-Bin Huang, and Ming-Hsuan Yang. Single image deblurring with adaptive dictionary learning. In 2010 IEEE International Conference on Image Processing, pages 1169-1172. IEEE, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[17] Donghua Jiang, Nestor Tsafack, Wadii Boulila, Jawad Ahmad, and JJ Barba-Franco. Asb-cs: Adaptive sparse basis compressive sensing model and its application to medical image encryption. Expert Systems with Applications, 236: 121378, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[18] Gareth A Jones and J Mary Jones. Information and coding theory. Springer Science & Business Media, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[19] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[20] Syed Ali Khayam. The discrete cosine transform (dct): theory and application. Michigan State University, 114(1):31, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.906, + 0.66 + ], + "angle": 0, + "content": "[21] Li Liu and Paul Fieguth. Texture classification using compressed sensing. In 2010 Canadian Conference on Computer and Robot Vision, pages 71-78. IEEE, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[22] Yang Liu, Qingchao Chen, Wei Chen, and Ian Wassell. Dictionary learning inspired deep network for scene recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[23] Liyan Ma, Lionel Moisan, Jian Yu, and Tieyong Zeng. A dictionary learning approach for poisson image deblurring. IEEE Transactions on medical imaging, 32(7):1277-1289, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3069" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[26] AKM Rabby and Chengcui Zhang. Beyondpixels: A comprehensive review of the evolution of neural radiance fields. arXiv preprint arXiv:2306.03000, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[27] AM Raid, WM Khedr, Mohamed A El-Dosuky, and Wesam Ahmed. Jpeg image compression using discrete cosine transform-a survey. arXiv preprint arXiv:1405.6147, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[28] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[29] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[30] K Seemakurthy, A Majumdar, J Gubbi, NK Sandeep, A Varghese, S Deshpande, M Girish Chandra, and P Balamurali. Deep dictionary learning for inpainting. In Computer Vision, Pattern Recognition, Image Processing, and Graphics: 7th National Conference, NCVPRIPG 2019, Hubballi, India, December 22-24, 2019, Revised Selected Papers 7, pages 79-88. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[31] Aviv Shamsian, Aviv Navon, David W Zhang, Yan Zhang, Ethan Fetaya, Gal Chechik, and Haggai Maron. Improved generalization of weight space networks via augmentations. arXiv preprint arXiv:2402.04081, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[33] Stanford University Computer Graphics Laboratory. The stanford 3d scanning repository. https://graphics.stanford.edu/data/3Dscanrep/. Accessed: 2024-07-15. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[34] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74-91. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[36] Hao Tang, Hong Liu, Wei Xiao, and Nicu Sebe. When dictionary learning meets deep learning: Deep dictionary learning and coding network for image recognition with limited data. IEEE transactions on neural networks and learning systems, 32(5):2129-2141, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[37] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International conference on learning representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Joel A Tropp and Anna C Gilbert. Signal recovery from random measurements via orthogonal matching pursuit." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on information theory, 53(12):4655-4666, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[39] Gregory K Wallace. The jpeg still picture compression standard. IEEE transactions on consumer electronics, 38(1): xviii-xxxiv, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[40] Jan Hendrik Weissenbruch. The shipping canal at rijswijk, known as \"the view at geestbrug\". Painting. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[41] Yueqi Xie, Ka Leong Cheng, and Qifeng Chen. Enhanced invertible encoding for learned image compression. In Proceedings of the 29th ACM international conference on multimedia, pages 162-170, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[42] Xijuan Zhang, Oscar L Olvera Astivia, Edward Kroc, and Bruno D Zumbo. How to think clearly about the central limit theorem. Psychological Methods, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[43] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 630-641, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.391 + ], + "angle": 0, + "content": "[44] Jinjia Zhou and Jian Yang. Compressive sensing in image/video compression: Sampling, coding, reconstruction, and codec optimization. Information, 15(2):75, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.434 + ], + "angle": 0, + "content": "[45] Fang Zhu, Shuai Guo, Li Song, Ke Xu, Jiayu Hu, et al. Deep review and analysis of recent nerfs. APSIPA Transactions on Signal and Information Processing, 12(1), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3070" + } + ] +] \ No newline at end of file diff --git a/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_origin.pdf b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f32085fc5cf2b2809ed2fa5defe2da8af16b66e --- /dev/null +++ b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/a675d8dc-7039-4fed-b302-de9ebbbd1aed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5391b85228fa903d1b957266cf6cfbf1f8fe02c90f1852752dd6c48a96e716af +size 2990345 diff --git a/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/full.md b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab581418d2b567fcd3bfbf72e7854a68095d30b2 --- /dev/null +++ b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/full.md @@ -0,0 +1,227 @@ +# SINR: Sparsity Driven Compressed Implicit Neural Representations + +Dhananjaya Jayasundara, Sudarshan Rajagopalan, Yasiru Ranasinghe, Trac D. Tran, and Vishal M. Patel Johns Hopkins University + +{vjayasul, sambasa2, dranasil, trac, vpatel36}@jhu.edu + +# Abstract + +Implicit Neural Representations (INRs) are increasingly recognized as a versatile data modality for representing discretized signals, offering benefits such as infinite query resolution and reduced storage requirements. Existing signal compression approaches for INRs typically employ one of two strategies: 1. direct quantization with entropy coding of the trained INR; 2. deriving a latent code on top of the INR through a learnable transformation. Thus, their performance is heavily dependent on the quantization and entropy coding schemes employed. In this paper, we introduce SINR, an innovative compression algorithm that leverages the patterns in the vector spaces formed by weights of INRs. We compress these vector spaces using a high-dimensional sparse code within a dictionary. Further analysis reveals that the atoms of the dictionary used to generate the sparse code do not need to be learned or transmitted to successfully recover the INR weights. We demonstrate that the proposed approach can be integrated with any existing INR-based signal compression technique. Our results indicate that SINR achieves substantial reductions in storage requirements for INRs across various configurations, outperforming conventional INR-based compression baselines. Furthermore, SINR maintains high-quality decoding across diverse data modalities, including images, occupancy fields, and Neural Radiance Fields. + +# 1. Introduction + +Despite the fact that all naturally occurring signals observed by humans are continuous, capturing these signals through digital devices requires their discretization. For example, an image of a mountain is processed and stored in a discretized format. A primary reason for this approach is to conserve storage space; storing signals with high precision in an almost continuous manner would necessitate a sub + +stantial amount of storage. Consequently, the digital representation of signals is both practical and essential. For instance, it is estimated that over 400TB of data is created every day [8]. Moreover, humans share their captured signals through various mediums on a daily basis. Therefore, data compression is essential for efficient transmission. + +Traditional signal compression techniques often rely on classic signal processing methods and are modality-specific. For example, JPEG [39], designed for photographic images, and is unsuitable for audio files. Similarly, audio compression standards like MP3 or AAC [4] are optimized for sound and are not applicable to images. With the advancements of neural networks, researchers have explored compressing signals using neural methods, predominantly through mechanisms based on autoencoders [2, 6, 7, 25, 37, 41]. In these systems, the encoder transforms the signal into a latent vector, which the decoder then uses to reconstruct the original signal. While autoencoder-based methods effectively encode signals into latent vectors, they are generally designed for a single modality. Adapting these methods to different data modalities not only requires training on a large corpus of data specific to those modalities but also a specialized autoencoder architecture tailored to handle the data effectively. + +In recent years, there has been a significant surge in interest in representing signals through Implicit Neural Representations (INRs). Unlike large models based on autoencoders, INRs typically consist of multi-layer perceptrons (MLPs) equipped with specialized nonlinearities that differ from the conventional nonlinearities used in deep learning. This simplicity and versatility allow INRs to unify signal representations across diverse data modalities. When signals are represented by INRs, they are encoded in the MLPs' weights and biases. For instance, in image transmission, instead of using conventional JPEG encoding, the weights and biases of the MLP are transmitted by a transmitter (TX). A receiver (RX) can then feed the coordinates into the MLP and decode the image. The primary advantage of INRs lies in their ability to represent signals with high fidelity while utilizing fewer parameters than parameter- + +heavy autoencoder-based mechanisms. + +Recent advances in INR-based signal compression include COIN [9], COIN++ [10], INRIC [34], and SHACIRA [11]. COIN pioneered the application of INRs for image compression. Building on this, COIN++ and INRIC introduced quantization and entropy coding to improve compression efficiency. Both approaches also focus on enhancing the generalization capabilities of INRs through meta-learning techniques. Additionally, COIN++ incorporates latent modulations discovered via a learnable transformation applied on top of the INR model. However, COIN++ requires transmitting the base INR and the learned transformation apriori, in addition to the latent modulations for signal decoding. Alternatively, SHACIRA applies quantization on the latent weights and enforces entropy regularization to reparameterize feature grids to enable efficient compression across diverse domains. None of the existing methods; however, have explored fundamentally compressing the INR by identifying patterns within its parameter space before applying standard techniques such as quantization and entropy coding. + +In our work, named SINR, we build upon the observed behaviors of the vector spaces generated by the weights in an INR. We integrate compressed sensing algorithms into the INR-based compression pipeline, proposing a mechanism that obtains a higher-dimensional sparse code for the weight vectors without requiring any learnable transformations. Furthermore, based on the Central Limit Theorem (CLT) [42], the transformation matrix can be directly sampled from a distribution rather than learning or hand crafting it. This eliminates the need to transmit the transformation matrix for accurately decoding weight spaces. Consequently, SINR, as a fundamental compression technique built on the observations of weights spaces, achieves superior compression and higher decoding quality for each data modality compared to the competing methods. Moreover, it can be easily embedded into any INR-based signal compression algorithm. + +# 2. Related works + +# 2.1. Implicit neural representations + +INRs have recently gained considerable attention in the computer vision community due to their streamlined network architectures and improved performance in various vision tasks compared to traditional, parameter-heavy models [13, 29, 32]. This surge in interest followed the advent of Neural Radiance Fields (NeRF) [24], which has inspired a plethora of subsequent studies [26, 45]. Further research has explored the pivotal role of different activation functions in INRs [28, 29, 32, 35]. Moreover, INRs provide a universal framework for representing various data modalities. More recent studies have investigated the use of INRs for image classification by transforming standard image for + +![](images/9f7208e65004e9cbb4c7df590ea7aa703522b4e875f67b2b1af56beeff83f21e.jpg) +(a) INR weight distributions for image representation. + +![](images/d7f5c994c8b5cd05b203b8c09a25c6001ba8deef94905b094da92bc8cbf1c56a.jpg) + +![](images/99aada467b8b772a54a64fecb40beea654efd40ddd8cf02af08f320ef9d28b7c.jpg) +(b) INR weight distribution for occupancy fields (left) and NeRF (right). + +![](images/9795a26df927fde903a81d644b99e10cc7b2b45aeeea5a0b119cbf0a29f0dd81.jpg) +Figure 1. Weight distribution of INRs tends to follow a Gaussian distribution for various data modalities. + +mats into INRs and training classifiers directly on the INRs' weights and biases [31]. These innovative approaches have showcased the potential of INRs to significantly reduce the dimensionality and computational complexity typically associated with conventional image processing techniques. + +# 2.2. Signal compression + +Signal compression is crucial for reducing bandwidth needs and saving storage space. With the rise of deep learning, signal compression has evolved into two main approaches: rule-based (traditional) and learning-based methods. Traditional compression methods, such as JPEG for images and MP3 for audio, rely on algorithmic techniques tailored to specific signal types. JPEG minimizes redundancies using the discrete cosine transform [27], while MP3 [4] employs a psycho-acoustic model that enhances compression by removing inaudible sounds through auditory masking. On the other hand, deep learning-based techniques use models trained on vast datasets, adapting to a wide range of signals without predefined algorithms. These methods offer flexibility but require different architectures for each data modality, presenting unique challenges. In this landscape, INRs stand out as a potential universal signal representative. + +# 2.3. Compressed sensing + +Compressed sensing is a field that capitalizes on the inherent sparsity of data to capture information efficiently. In digital imaging, not every pixel is crucial for accurate image reconstruction. Although images appear dense in pixel space, they exhibit considerable redundancy when transformed into different basis functions. This sparsity is exploited by compressed sensing algorithms to reconstruct the original image from fewer sampled data points. These algo + +![](images/7715bfd98d66a74aa851b04fdf64e542e8c0622526ae7901a192935388b90475.jpg) +Figure 2. The proposed SINR compression algorithm: Standard compression techniques for INRs typically involve direct quantization and entropy coding of their weights. However, since natural signals exhibit inherent compressibility in a dictionary, the characteristics that aid in the compressibility of the weight space of an INR are discovered through the Gaussian nature of the weight space. Therefore, SINR employs $L_{1}$ minimization to identify a higher-dimensional sparse code. Furthermore, based on the weight space observations and the CLT, we simplify the encoding and decoding process using a random sensing matrix controlled by a seed. Subsequently, only the non-zero (NZ) values and their corresponding indices are quantized and entropy coded. + +rithms employ optimization techniques and linear algebra to solve underdetermined systems, revolutionizing data acquisition in areas such as medical imaging and signal processing. Dictionary learning, integral to compressed sensing, seeks sparse representations of data using dictionary elements or atoms that capture the data's intrinsic structure. These atoms are either predefined or adaptively learned. Compressed sensing's versatility is evident in its applications across various domains, such as image and video compression [44], medical image encryption [17], and classification tasks [14, 15, 19, 21]. It also addresses inverse vision problems like image inpainting [30], deblurring [16, 23], and super-resolution [3]. Recent efforts have merged dictionary learning with deep learning to tackle more complex computer vision challenges, including image recognition [36], denoising [43], and scene recognition [22]. These developments underscore compressed sensing's transformative impact on computer vision. + +Our work, SINR, is pioneering the application of compressed sensing principles to INRs. By leveraging these principles alongside the structural distributions of INR weights, SINR identifies redundancies in these spaces, resulting in substantial compression improvements. + +# 3. Method + +# 3.1. Signal representation through INRs + +Mathematically, an INR can be defined by a function $G_{\theta}$ where $\theta$ are the estimizable parameters of the neural net + +work. The input and output dimensions of $G_{\theta}$ vary for different data modalities. In general, $G_{\theta}$ acts as a mapping from an $a$ -dimensional input coordinate space to a $b$ -dimensional output signal space, described as: + +$$ +G _ {\theta}: \mathbb {R} ^ {a} \to \mathbb {R} ^ {b}. +$$ + +For instance, for RGB images, $a = 2$ and $b = 3$ , while for audio signals, $a = 1$ and $b = 1$ . In this architecture, the output of the $i^{\text{th}}$ layer, which feeds into the $(i + 1)^{\text{th}}$ layer, can be expressed as $\sigma(W^{(i)}y^{(i)} + b^{(i)})$ . Here, $\sigma$ denotes the activation function, and $y^{(i)}$ represents the output from the preceding layer. Furthermore, the choice of activation function $(\sigma)$ plays a critical role in shaping the neural network's ability to model complex functions, as explored in various studies [28, 29, 32, 35]. + +# 3.2. Exploring the compressibility + +According to compressed sensing theory, most real-world signals display sparsity when transformed into an appropriate domain, meaning they can be accurately represented with fewer measurements than traditionally required. Furthermore, real-world signals can be compressed through a set of basis functions, and the coefficients of these functions are derived by minimizing the reconstruction loss. The core concept of INRs involves encoding signals into the weights and biases of an MLP. This process can be viewed as a classical domain transformation technique where pixel values are reconstructed by feeding the corresponding coordinates through the MLP. Unlike predefined signal trans + +formers like Fourier [5] and DCT [20], the MLP attempts to minimize the reconstruction loss through backpropagation to find the transformation. The learned representation of the signal resides in another domain. Given that real-world signals inherently exhibit sparsity in transformed domains, we hypothesize that this sparsity can be explored within the MLP's weights. If we can identify where this sparse nature is hidden within the weight space, we could achieve further compression on INRs compared to competing methods. However, identifying this sparse representation within the weights is not straightforward. We believe there are two main approaches to achieving a sparser representation, each with its own challenges and considerations. + +The first approach involves either promoting or enforcing a specified level of sparsity in the weights during the training of an INR. Promoting sparsity can generally be achieved by incorporating $L_{1}$ regularization on the model parameters, which encourages many weights to approach zero, thereby creating a sparse representation. However, while we observed that $L_{1}$ regularization results in a higher level of sparsity within the weights, it fails to accurately represent the images. Alternatively, enforcing a specified level of sparsity can also be achieved through model pruning, where weights deemed insignificant are pruned or eliminated during the training process. When it comes to model pruning, we employed both structured and unstructured pruning of weights. We noted that both methods led to significant performance degradation for certain data modalities, particularly for occupancy fields. Moreover, only a small pruning percentages resulted in satisfactory performance for signal representation. For applications that require a high level of generalization, such as NeRFs, the pruning approach did not generalize well, indicating its limitations in achieving a balance between sparsity and performance. + +The second approach seeks to uncover the inherent structures within the weights that aid in INR compression. This involves identifying patterns or regularities that can be exploited to reduce the dimensionality of the representation without sacrificing performance. We examined this from a dimensionality reduction perspective; however, the weight space in reduced dimensions did not reveal clear patterns, even across different natural images. Nonetheless, we observed that the weight space of an INR often tends to follow a normal distribution. Fig. 1 shows the weight distribution of the hidden layers of an INR when different data modalities are encoded into it. This suggests that INRs share a common pattern across different data modalities, showcasing a potential pathway for a fundamental compression. Given that each weight vector of an INR exhibits Gaussian behavior, we seek a higher-dimensional but sparse equivalent through a dictionary learning-based approach. Let us denote $\mathbf{w} \in \mathbb{R}^{k_1}$ as a hidden weight vector, $\mathbf{A} \in \mathbb{R}^{k_1 \times k_2}$ as + +a dictionary, and $\mathbf{x} \in \mathbb{R}^{k_2}$ as the corresponding sparse vector. In search of a sparse representation, according to standard compressed sensing, we can write $\mathbf{w} = \mathbf{A}\mathbf{x}$ , where $\| \mathbf{x} \|_0 < k_1$ . To discover the sparse code $\mathbf{x}$ , the best and most efficient choice is $L_1$ minimization, as $L_0$ minimization iterates through all possible combinations and is therefore not efficient. However, the problem arises with the sensing matrix, commonly referred to as the dictionary $\mathbf{A}$ . Although we could use either a dictionary learning-based approach for learning basis functions for the dictionary or a deep learning-based learnable transformation, these approaches would be time-consuming. Furthermore, a TX needs to transmit the learned dictionary alongside the obtained sparse codes. Conversely, we propose that the dictionary does not need to be learned or even transmitted. + +As we have confirmed, the weights are normally distributed. According to the CLT, a normally distributed random variable can be produced through a finite linear combination of any random variables. In summation form, this can be expressed as: $w_{i} = \sum_{j = 1}^{k_{2}}A_{ij}x_{j}$ , where $w_{i}$ is the $i$ -th element of the weight vector $\mathbf{w}$ , $A_{ij}$ is the element in the $i$ -th row and $j$ -th column of the sensing matrix $\mathbf{A}$ , and $x_{j}$ is the $j$ -th element of the vector $\mathbf{x}$ . To satisfy the CLT, the number of terms in the summation, which is $k_{2}$ , should be sufficiently large. Therefore, considering all elements of the weight vector $\mathbf{w}$ , this can be compactly written as $\mathbf{w} = \mathbf{A}\mathbf{x}$ . From CLT, the sensing matrix can be defined by a set of random vectors whose appropriate coefficients $(\mathbf{x})$ can be learned using sparse signal recovery algorithms such as matching pursuit or its variants. Therefore, the optimization can be written as, min $\| \mathbf{x}\| _1$ subject to $\mathbf{w} = \mathbf{A}\mathbf{x}$ . + +For convenience, let us denote $\| \mathbf{x} \|_0$ as $s$ . A further constraint to the above optimization procedure is that when the sparse code $\mathbf{x}$ is found, we need to store not only its nonzero elements but also the corresponding indices. Therefore, the above $L_1$ minimization is solved with $2s < k_1$ . We do not apply our compression algorithm to the biases of the INR as the size of bias vectors is very small compared to those of the weight matrices. + +Instead of saving $k_{1}$ floating-point numbers for $\mathbf{w}$ , we now only need to save $2s$ elements: $s$ elements are floating-point numbers representing the non-zero values in the sparse code, and the remaining $s$ elements are integers that give the indices of those non-zero values. The indices can often be represented with 16-bit precision, unlike the nonzero values in the sparse code, which require 32-bit floating-point precision. At the RX end, $\mathbf{x}$ must be converted back to $\mathbf{w}$ . This requires the sensing matrix $\mathbf{A}$ , which is random and can be controlled by a seed to reproduce the exact $\mathbf{w}$ using $\mathbf{w} = \mathbf{A}\mathbf{x}$ . Thus, the receiver only needs $\mathbf{x}$ to obtain $\mathbf{w}$ . This process can be viewed as a method of uncovering the inherent sparsity within natural signals, as represented through the weight space of INRs. As we hypothesized, the + +ability to condense natural signals into a dictionary hinges on identifying specific patterns encoded within the weights of INRs. Once the non-zero elements of the sparse vector are obtained at the RX, the resulting procedure is virtually the same across different INR-based baselines. Our method fundamentally achieves compression by delving into the weight spaces to uncover patterns, a step not typically taken by existing methods. A summarization of SINR is illustrated in Fig. 2. As can be seen from Fig. 2, SINR is only dependent on the weights of the INR and is applied prior to any quantization or entropy coding schemes. Therefore, SINR can be applied to existing INR compression methods to improve their compressibility. + +# 3.3. How much fundamental compression does SINR achieve? + +# 3.3.1.Standard INRs + +Consider an INR with $l$ hidden layers, yielding $l + 2$ total layers. For simplicity, assume $k$ neurons per hidden layer. If the input dimension is $a$ and the output dimension is $b$ , the total number of weight parameters is given by $\mathcal{T}_s = a \times k + l \times k^2 + b \times k$ . However, SINR modifies this structure by reducing the parameters from $\mathcal{T}_s$ in the original network to $\mathcal{T}_{s_{\mathrm{INR}}} = a \times 2s + k \times l \times 2s + b \times 2s$ , where $s \ll k$ . Additionally, SINR does not require transmitting any additional data to recover the original INR weights. + +# 3.3.2. Tiny INRs + +Let us define an INR as "tiny" if the number of neurons in a hidden layer, denoted by $k$ , is less than 50. In such cases, we aim to achieve a sparse representation where $2s < k$ and $\| x \|_0 = s$ . However, achieving a sparse representation that satisfies $2s < k$ is often extremely challenging and typically does not result in effective compression. To overcome this, we exploit the fact that the weight matrix connecting the $i^{\text{th}}$ layer to the $(i + 1)^{\text{th}}$ layer is of dimensions $k \times k$ . By vectorizing this weight matrix, we obtain a vector of dimension $k^2 \times 1$ . Given that $k^2$ is significantly larger than $k$ , we can apply our SINR procedure directly to the flattened weight matrix. This strategy leads to a sparser representation, thus enhancing compression efficiency for tiny INRs. + +# 3.3.3. COIN++ + +In the COIN++ framework, modulation parameters are stored instead of traditional weights and biases, under the assumption that the base network parameters can be transmitted beforehand. For $n$ test images, each segmented into $m$ patches with a latent dimension of size $d$ , COIN++ necessitates the transmission of $m \times d$ parameters for reconstructing each image. As the base network in COIN++ conforms to a standard INR structure, it is amenable to further compression via the SINR technique. By implementing SINR principles on the modulations in COIN++, the parameter transmission requirement per image can be reduced from + +$m \times d$ to just $2s \times d$ , where $s \ll m$ . As the size of each test image and the number of images in the test dataset grow, COIN++ would typically require the transmission of numerous parameters. However, by leveraging SINR, both the modulations and the base network can be significantly compressed, achieving enhanced compression. + +# 3.4. Quantization and entropy coding + +After an INR is trained, its parameters are not immediately saved but are first subject to quantization [12]. This involves reducing the bitwidths below typical floating-point precision. Following quantization, the parameters are processed through entropy coding, in our experiments we utilize Brotli coding [1, 18], which allows the compressed data to be stored or transmitted efficiently. To retrieve the original parameters, the decoder must reverse the entropy coding and then perform dequantization. In the case of SINR, the compression process is intensified by utilizing the sparsity induced in the model parameters by natural signals. Once the sparse code is established, the parameters are quantized and subjected to entropy coding. The decoder then reverses the entropy coding and dequantizes the data. Finally, the model parameters are reconstructed by multiplying them with a random Gaussian matrix with a specific seed. + +# 4. Experiments + +# 4.1. Experimental setup + +SINR, a novel INR compression algorithm, is predicated on the idea that if natural signals are compressible through a dictionary, then INRs should be similarly compressible. This concept underpins SINR's goal to efficiently reduce INR storage requirements while maintaining high fidelity. Our experiments, conducted using the PyTorch framework following WIRE [29] codebase on an NVIDIA RTX A6000 GPU, spanned various data types including images, occupancy fields, and neural radiance fields. Image encoding metrics involved file size, bits per pixel (bpp) and Peak Signal-to-Noise Ratio (PSNR). Occupancy fields were evaluated using file size and Intersection over Union (IoU), and neural radiance fields were assessed using file size and PSNR. Other than the network configurations mentioned in the paper, for occupancy field evaluation, we utilized an MLP with 128 hidden neurons, and 3 hidden layers. For INRIC, we applied the network hyperparameters specified in its paper. In COIN++, we followed the guidelines in its paper but modified the hidden neuron size to 300. All experiments used Brotli entropy coding with a 16-bitwidth (65536 levels) uniform quantizer. + +# 4.2. How do we find s? + +We implemented $L_{1}$ minimization using the Orthogonal Matching Pursuit (OMP) algorithm [38]. The OMP algorithm requires the pre-determination of $s$ before obtaining + +x, and it must adhere to the condition $2s < k_{1}$ . If $2s$ is set too low, it results in inaccurate representations of $\mathbf{w}$ within the weight space. Therefore, we incrementally increased $s$ from a low value until $2s = k_{1}$ for all KODAK images in the $C_1$ experiment, as outlined in Sec. 4.3. Our findings suggest that the optimal value of $s$ for successfully reconstructing the weight space does not depend on the specific image but on the number of neurons in a hidden layer. By adjusting the neuron count, we identified an optimal $s$ that accurately reconstructs the weight space while satisfying the specified constraint. Extending these experiments to natural signals outside the KODAK dataset confirmed the consistency of our results. Additionally, we have included a regression plot in the supplementary that details how to determine the optimal $s$ based on the number of neurons. + +# 4.3. Image encoding + +Representing an image through the weights and biases of a neural network serves as a method of encoding. For our image encoding task, we utilized the KODAK dataset, which includes 24 natural RGB images, each measuring $768 \times 512$ pixels. We conducted five types of experiments, denoted as $C_i$ , where $i$ ranges from 1 to 5, to demonstrate the effectiveness of our proposed method. + +Experiment $C_1$ involved encoding each image in the KODAK dataset using an INR without positional embedding, by varying the number of neurons in each hidden layer. Experiment $C_2$ mirrored $C_1$ , but with the variation in the number of hidden layers instead. Experiments $C_3$ and $C_4$ implemented the meta-learning approach for INRs proposed in INRIC, without and with positional embedding for the input layer, respectively. Experiment $C_5$ involved the COIN++ framework, testing both with and without patching. When using patching, we adopted $32 \times 32$ patches as suggested by COIN++. However, we observed that without patching, even as the latent modulation dimension increased, the average PSNR obtained by COIN++ remained nearly constant. For these meta-learning-based experiments, we used the first 12 images of the KODAK dataset for meta-learning and the remaining 12 images for fine-tuning. For all image encoding experiments, we used the Sinusoidal activation function (see supplementary). + +Let us define $h$ and $m$ as the number of hidden layers and the number of neurons per hidden layer in an INR, respectively. For experiment $C_1$ , we configured the INR with settings $(h,m)$ as $(2,32)$ , $(3,64)$ , $(3,128)$ . Experiment $C_1$ aims to assess the effectiveness of SINR by varying the number of hidden neurons. The results, depicted in Fig. 3, demonstrate how effectively SINR identifies the compressibility of the weight space. This is indicated by the bpp values, which reflect the size of the model parameters. For example, representing the KODAK dataset with an average PSNR of 30 dB requires about 3.7 bpp for COIN and 2.0 + +![](images/6fb298a998e1ce91f83f6d74ab8f38c95f2f6d9414b5df8f797ecebe9d5efd68.jpg) +Figure 3. Experiments $C_1$ and $C_2$ : Identifying compressible INR combinations. The SINR approach demonstrates that configurations in $C_1$ are more compressible than those in $C_2$ . Furthermore, in both configurations SINR achieves lower bpp while maintaining the PSNR values. + +bpp for INRIC. However, SINR significantly reduces the bpp to approximately 1.7 using the same quantizer and entropy coder. The first configuration in $C_1$ falls under the category of tiny INRs, underscoring the proposed method's effectiveness even for compact INRs. As illustrated in Fig. 3, SINR achieves the same level of PSNR as baselines with a lower bpp for any network configuration. This substantial reduction of bpp across the $C_1$ experiment showcases the efficiency and compactness achieved by SINR. From $C_1$ , it can be established that greater compressibility of an INR into a dictionary is possible with an increased number of hidden neurons. Following the conclusions drawn from experiment $C_1$ , experiment $C_2$ was designed to explore the impact of increasing the number of hidden layers on the effectiveness of SINR. The configurations tested in $C_2$ were $(h,m) = \{(3,64),(5,64),(7,64)\}$ . As illustrated in Fig. 3, SINR consistently achieved PSNR levels comparable to baseline methods, but with a reduced bpp. Given that $C_2$ maintained a constant neuron count at 64, the observed deviations in compression between SINR and INRIC were less significant than those observed in $C_1$ . This discrepancy can be attributed to the following: an INR configuration with a higher number of neurons (e.g., $m = 128$ ), even with fewer hidden layers (e.g., $h = 2$ ), possesses more trainable parameters. Consequently, such a model is capable of learning a more robust representation of the image compared to configurations with a larger number of layers but fewer neurons per layer. As a result, the compressible characteristics of the images are more effectively transferred into the model parameters during the INR training process. This leads to a more compressible INR. These findings support the premise that if natural images can be efficiently compressed into a dictionary, the weight space + +![](images/88323c6069865b12560ac00de8bb3b353d9a17ab289d138ce0eaec8b4dc95a3f.jpg) +Figure 4. Experiments $C_3$ , $C_4$ , and $C_5$ : Identifying compressible INR combinations under Meta-Learning. Meta-learning approaches have been introduced for INRs to enhance their generalization abilities and achieve faster convergence. When assessing induced sparsity in the weight space, SINR demonstrates a significant reduction in bpp values while maintaining nearly the same PSNR performance as the baselines. + +of INRs can also be effectively compressed. + +As in previous experiments, each image required separate training of an INR. Experiments $C_3$ and $C_4$ address this challenge through meta-learning, with and without positional embedding, respectively. The configurations for these INRs are given by $(h,m) = \{(3,32),(3,64),(3,96),(3,128)\}$ . For COIN++, the number of layers was set to 5 with MLP's hidden dimension at 300. The latent dimension parameter $(d)$ varied as follows: $d = \{16,32,64,96\}$ . Fig. 4 presents the experimental results for $C_3$ , $C_4$ , and $C_5$ , illustrating significant compression capabilities of the proposed SINR within a meta-learning framework. Notably, models using positional embedding generally have more parameters than those without. + +Comparing the performance of INRIC and SINR without positional embedding schemes, the initial INR configuration shows that SINR exhibits a lower bpp for the same average PSNR. Generally, as bpp increases, the representation capacity of the INR enhances, leading to more robust image representation. From the graphs, as bpp increases, SINR demonstrates greater improvements in PSNR than INRIC, a phenomenon that can be explained by the aforementioned logic. In the case of COIN++, the approach focuses on fine-tuning only the modulations using their proposed meta-learning method. However, since finetuning encodes natural signals within these modulations, they should also be compressible via a dictionary. Due to patching, each KODAK test image results in a $d \times 384$ matrix. Our experiments reveal that these modulations encode hidden redundancies in natural signals. For instance, to achieve an average PSNR of approximately 24.2 dB, + +COIN++ requires more than 1.5 bpp; however, the same PSNR can be achieved with COIN++ using just under 1 bpp by exploiting the hidden sparsity in its modulations through our proposed approach. Therefore, when a high-capacity model effectively represents a signal, it must encapsulate this sparsity within its weight and bias spaces. SINR explores and removes redundancies in these parameters, retaining only essential information. Fig. 5 showcases the decoded images by SINR alongside with the INR based image compressors. Decoded PSNR, BPP, and file size are displayed in the first, second, and third rows of the text boxes. + +# 4.4. Occupancy fields encoding + +Occupancy fields are represented by binary values, either 1 or 0, where 1 denotes that the signal lies within a specified region and 0 indicates its absence. Another variant of occupancy volumes stores not only the presence or absence of a signal but also the color at that location. Typically, occupancy fields consume more space than other data modalities. However, they can be represented with higher accuracy and lower storage requirements using INRs. In this experiment, we followed the sampling procedure described in [29]. Occupancy fields can be thought of as representations of three-dimensional objects, capturing natural signals. Despite following the sampling procedure, redundancies may exist that are not essential for representing the occupancy volume. Identifying these redundancies can reduce storage requirements. However, identifying them in the spatial domain $(xyz)$ requires domain-specific algorithms, as described in Sec. 1. As INRs serve as unified data modality representatives, these redundancies must be encapsulated within its weights space. SINR fundamentally compresses the INRs into a dictionary regardless of the data modality; therefore, indeed it is equally applicable to occupancy fields. To validate this hypothesis for occupancy fields, two experiments were conducted using shapes from the Stanford shape dataset [33]. Figure 6 showcases the decoded SINR's representations for 'Thai Statue'(first volume) and 'Lucy' (fourth volume) datasets alongside the existing INR-based occupancy compressor. We use the Gaussian activation function for this task (see supplementary). The first value and second value in each text box represent the IoU metric and storage requirement, respectively, except for GT. + +# 4.5. Additional materials + +The pseudocode for SINR, additional results and ablation studies on finding $s$ are in the supplementary material. + +# 5. Conclusion + +Implicit Neural Representations (INRs) have emerged as a promising framework for unified data modality representation. Several studies have explored the potential for compressing images, occupancy fields, and audio using INRs. + +![](images/83679984da93e168f814fc55fb8341af0f970ab79b9c631697abfb21be4c514a.jpg) +Figure 5. Results for image encoding experiment. SINR compresses the INR into a dictionary, significantly reducing the storage required compared to baseline INR image compressors. The results demonstrate that the decoded representations undergo a very negligible loss in PSNR, which is minimal considering the substantial storage space saved. + +![](images/45c9e37db49849fa961df1a6bd38e435297cb6285d7b30113c5b1768b2941228.jpg) +Figure 6. Results for occupancy fields encoding experiment. The results clearly demonstrate that SINR achieves the smallest file size and the highest accuracy metric for every shape in the tested dataset. The significant compression obtained by our algorithm suggests that occupancy fields, when represented using an INR, can be more efficiently compressed into a dictionary compared to images. + +However, none of these methods have investigated whether the INR itself can be compressed prior to quantization and entropy coding. As natural signals can be efficiently compressed in bases of transformed domains due to their sparsity—allowing for higher accuracy and lower storage requirements—we hypothesize that a similar compressible nature must also exist in the INR once it is trained. With the discovery that weight vectors in the weight space tend to adhere to a Gaussian distribution, we propose SINR, which compresses any INR in a dictionary. Furthermore, we demonstrate that this dictionary does not need to be learned but can instead be generated using a seed. We compare our findings with standard INR compressors for im + +ages, occupancy fields, and neural radiance fields. SINR achieves fundamental compression for any INR, independent of other post-processing methods such as quantization and entropy coding, and it showcases significantly lower storage requirements and higher fidelity across various data modalities. Through our experiments, we observed that the INR can be more compressed when a more robust representation of the signal is learned. Additionally, some data modalities exhibit greater compressibility than others. We firmly believe this research will aid other researchers in exploring more patterns in the weight spaces of INRs and in developing operators and transforms for INR. + +# Acknowledgment + +This work is supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government. + +# References + +[1] Jyrki Alakuijala, Andrea Farruggia, Paolo Ferragina, Eugene Kliuchnikov, Robert Obryk, Zoltan Szabadka, and Lode Vandevenne. Brotli: A general-purpose data compressor. ACM Transactions on Information Systems (TOIS), 37(1):1-30, 2018. 5 +[2] David Alexandre, Chih-Peng Chang, Wen-Hsiao Peng, and Hsueh-Ming Hang. An autoencoder-based learned image compressor: Description of challenge proposal by nctu. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2539-2542, 2018. 1 +[3] Selen Ayas and Murat Ekinci. Single image super resolution using dictionary learning and sparse coding with multi-scale and multi-directional gabor feature representation. Information Sciences, 512:1264-1278, 2020. 3 +[4] Karlheinz Brandenburg. Mp3 and aac explained. In Audio Engineering Society Conference: 17th International Conference: High-Quality Audio Coding. Audio Engineering Society, 1999. 1, 2 +[5] Komaravolu Chandrasekharan. Classical Fourier Transforms. Springer Science & Business Media, 2012. 4 +[6] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Energy compaction-based image compression using convolutional autoencoder. IEEE Transactions on Multimedia, 22(4):860-873, 2019. 1 +[7] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1 +[8] Fabio Duarte. Amount of data created daily (2024). https://explodingtopics.com/blog/data-generated-per-day, 2024. Accessed: 2024-07-14. 1 +[9] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. arXiv preprint arXiv:2103.03123, 2021. 2 +[10] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2 + +[11] Sharath Girish, Abhinav Shrivastava, and Kamal Gupta. Shacira: Scalable hash-grid compression for implicit neural representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17513-17524, 2023. 2 +[12] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 5 +[13] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Implicit neural representations with levels-of-experts. Advances in Neural Information Processing Systems, 35:2564-2576, 2022. 2 +[14] Daniel J Hsu, Sham M Kakade, John Langford, and Tong Zhang. Multi-label prediction via compressed sensing. Advances in neural information processing systems, 22, 2009. 3 +[15] Junlin Hu and Yap-Peng Tan. Nonlinear dictionary learning with application to image classification. Pattern Recognition, 75:282-291, 2018. 3 +[16] Zhe Hu, Jia-Bin Huang, and Ming-Hsuan Yang. Single image deblurring with adaptive dictionary learning. In 2010 IEEE International Conference on Image Processing, pages 1169-1172. IEEE, 2010. 3 +[17] Donghua Jiang, Nestor Tsafack, Wadii Boulila, Jawad Ahmad, and JJ Barba-Franco. Asb-cs: Adaptive sparse basis compressive sensing model and its application to medical image encryption. Expert Systems with Applications, 236: 121378, 2024. 3 +[18] Gareth A Jones and J Mary Jones. Information and coding theory. Springer Science & Business Media, 2012. 5 +[19] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 3 +[20] Syed Ali Khayam. The discrete cosine transform (dct): theory and application. Michigan State University, 114(1):31, 2003. 4 +[21] Li Liu and Paul Fieguth. Texture classification using compressed sensing. In 2010 Canadian Conference on Computer and Robot Vision, pages 71-78. IEEE, 2010. 3 +[22] Yang Liu, Qingchao Chen, Wei Chen, and Ian Wassell. Dictionary learning inspired deep network for scene recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3 +[23] Liyan Ma, Lionel Moisan, Jian Yu, and Tieyong Zeng. A dictionary learning approach for poisson image deblurring. IEEE Transactions on medical imaging, 32(7):1277-1289, 2013. 3 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1 + +[26] AKM Rabby and Chengcui Zhang. Beyondpixels: A comprehensive review of the evolution of neural radiance fields. arXiv preprint arXiv:2306.03000, 2023. 2 +[27] AM Raid, WM Khedr, Mohamed A El-Dosuky, and Wesam Ahmed. Jpeg image compression using discrete cosine transform-a survey. arXiv preprint arXiv:1405.6147, 2014. 2 +[28] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 2, 3 +[29] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2, 3, 5, 7 +[30] K Seemakurthy, A Majumdar, J Gubbi, NK Sandeep, A Varghese, S Deshpande, M Girish Chandra, and P Balamurali. Deep dictionary learning for inpainting. In Computer Vision, Pattern Recognition, Image Processing, and Graphics: 7th National Conference, NCVPRIPG 2019, Hubballi, India, December 22-24, 2019, Revised Selected Papers 7, pages 79-88. Springer, 2020. 3 +[31] Aviv Shamsian, Aviv Navon, David W Zhang, Yan Zhang, Ethan Fetaya, Gal Chechik, and Haggai Maron. Improved generalization of weight space networks via augmentations. arXiv preprint arXiv:2402.04081, 2024. 2 +[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 3, 1 +[33] Stanford University Computer Graphics Laboratory. The stanford 3d scanning repository. https://graphics.stanford.edu/data/3Dscanrep/. Accessed: 2024-07-15. 7 +[34] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74-91. Springer, 2022. 2 +[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 3 +[36] Hao Tang, Hong Liu, Wei Xiao, and Nicu Sebe. When dictionary learning meets deep learning: Deep dictionary learning and coding network for image recognition with limited data. IEEE transactions on neural networks and learning systems, 32(5):2129-2141, 2020. 3 +[37] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International conference on learning representations, 2022. 1 +[38] Joel A Tropp and Anna C Gilbert. Signal recovery from random measurements via orthogonal matching pursuit. + +IEEE Transactions on information theory, 53(12):4655-4666, 2007. 5 +[39] Gregory K Wallace. The jpeg still picture compression standard. IEEE transactions on consumer electronics, 38(1): xviii-xxxiv, 1992. 1 +[40] Jan Hendrik Weissenbruch. The shipping canal at rijswijk, known as "the view at geestbrug". Painting. 2 +[41] Yueqi Xie, Ka Leong Cheng, and Qifeng Chen. Enhanced invertible encoding for learned image compression. In Proceedings of the 29th ACM international conference on multimedia, pages 162-170, 2021. 1 +[42] Xijuan Zhang, Oscar L Olvera Astivia, Edward Kroc, and Bruno D Zumbo. How to think clearly about the central limit theorem. Psychological Methods, 2022. 2 +[43] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 630-641, 2021. 3 +[44] Jinjia Zhou and Jian Yang. Compressive sensing in image/video compression: Sampling, coding, reconstruction, and codec optimization. Information, 15(2):75, 2024. 3 +[45] Fang Zhu, Shuai Guo, Li Song, Ke Xu, Jiayu Hu, et al. Deep review and analysis of recent nerfs. APSIPA Transactions on Signal and Information Processing, 12(1), 2023. 2 \ No newline at end of file diff --git a/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/images.zip b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec9297522e966105b7ff83d5b5b7d8a821adf208 --- /dev/null +++ b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:136d07e3226cfcb2fa1367673649889bb9accf889a657046a7c5d94953cc98d3 +size 335890 diff --git a/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/layout.json b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a68417c85cd67611b58e7d63c8426ce477b9849c --- /dev/null +++ b/2025/SINR_ Sparsity Driven Compressed Implicit Neural Representations/layout.json @@ -0,0 +1,8174 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "text", + "content": "SINR: Sparsity Driven Compressed Implicit Neural Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 143, + 473, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 143, + 473, + 185 + ], + "spans": [ + { + "bbox": [ + 135, + 143, + 473, + 185 + ], + "type": "text", + "content": "Dhananjaya Jayasundara, Sudarshan Rajagopalan, Yasiru Ranasinghe, Trac D. Tran, and Vishal M. Patel Johns Hopkins University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 187, + 447, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 187, + 447, + 200 + ], + "spans": [ + { + "bbox": [ + 154, + 187, + 447, + 200 + ], + "type": "text", + "content": "{vjayasul, sambasa2, dranasil, trac, vpatel36}@jhu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 253, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 552 + ], + "type": "text", + "content": "Implicit Neural Representations (INRs) are increasingly recognized as a versatile data modality for representing discretized signals, offering benefits such as infinite query resolution and reduced storage requirements. Existing signal compression approaches for INRs typically employ one of two strategies: 1. direct quantization with entropy coding of the trained INR; 2. deriving a latent code on top of the INR through a learnable transformation. Thus, their performance is heavily dependent on the quantization and entropy coding schemes employed. In this paper, we introduce SINR, an innovative compression algorithm that leverages the patterns in the vector spaces formed by weights of INRs. We compress these vector spaces using a high-dimensional sparse code within a dictionary. Further analysis reveals that the atoms of the dictionary used to generate the sparse code do not need to be learned or transmitted to successfully recover the INR weights. We demonstrate that the proposed approach can be integrated with any existing INR-based signal compression technique. Our results indicate that SINR achieves substantial reductions in storage requirements for INRs across various configurations, outperforming conventional INR-based compression baselines. Furthermore, SINR maintains high-quality decoding across diverse data modalities, including images, occupancy fields, and Neural Radiance Fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 610, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 295, + 694 + ], + "type": "text", + "content": "Despite the fact that all naturally occurring signals observed by humans are continuous, capturing these signals through digital devices requires their discretization. For example, an image of a mountain is processed and stored in a discretized format. A primary reason for this approach is to conserve storage space; storing signals with high precision in an almost continuous manner would necessitate a sub" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "stantial amount of storage. Consequently, the digital representation of signals is both practical and essential. For instance, it is estimated that over 400TB of data is created every day [8]. Moreover, humans share their captured signals through various mediums on a daily basis. Therefore, data compression is essential for efficient transmission." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 303, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 518 + ], + "type": "text", + "content": "Traditional signal compression techniques often rely on classic signal processing methods and are modality-specific. For example, JPEG [39], designed for photographic images, and is unsuitable for audio files. Similarly, audio compression standards like MP3 or AAC [4] are optimized for sound and are not applicable to images. With the advancements of neural networks, researchers have explored compressing signals using neural methods, predominantly through mechanisms based on autoencoders [2, 6, 7, 25, 37, 41]. In these systems, the encoder transforms the signal into a latent vector, which the decoder then uses to reconstruct the original signal. While autoencoder-based methods effectively encode signals into latent vectors, they are generally designed for a single modality. Adapting these methods to different data modalities not only requires training on a large corpus of data specific to those modalities but also a specialized autoencoder architecture tailored to handle the data effectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "In recent years, there has been a significant surge in interest in representing signals through Implicit Neural Representations (INRs). Unlike large models based on autoencoders, INRs typically consist of multi-layer perceptrons (MLPs) equipped with specialized nonlinearities that differ from the conventional nonlinearities used in deep learning. This simplicity and versatility allow INRs to unify signal representations across diverse data modalities. When signals are represented by INRs, they are encoded in the MLPs' weights and biases. For instance, in image transmission, instead of using conventional JPEG encoding, the weights and biases of the MLP are transmitted by a transmitter (TX). A receiver (RX) can then feed the coordinates into the MLP and decode the image. The primary advantage of INRs lies in their ability to represent signals with high fidelity while utilizing fewer parameters than parameter-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 703, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 257, + 713 + ], + "type": "text", + "content": "Project Page: https://dsgrad.github.io/SINR" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3061" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 214, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 214, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 214, + 83 + ], + "type": "text", + "content": "heavy autoencoder-based mechanisms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 323 + ], + "type": "text", + "content": "Recent advances in INR-based signal compression include COIN [9], COIN++ [10], INRIC [34], and SHACIRA [11]. COIN pioneered the application of INRs for image compression. Building on this, COIN++ and INRIC introduced quantization and entropy coding to improve compression efficiency. Both approaches also focus on enhancing the generalization capabilities of INRs through meta-learning techniques. Additionally, COIN++ incorporates latent modulations discovered via a learnable transformation applied on top of the INR model. However, COIN++ requires transmitting the base INR and the learned transformation apriori, in addition to the latent modulations for signal decoding. Alternatively, SHACIRA applies quantization on the latent weights and enforces entropy regularization to reparameterize feature grids to enable efficient compression across diverse domains. None of the existing methods; however, have explored fundamentally compressing the INR by identifying patterns within its parameter space before applying standard techniques such as quantization and entropy coding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 528 + ], + "type": "text", + "content": "In our work, named SINR, we build upon the observed behaviors of the vector spaces generated by the weights in an INR. We integrate compressed sensing algorithms into the INR-based compression pipeline, proposing a mechanism that obtains a higher-dimensional sparse code for the weight vectors without requiring any learnable transformations. Furthermore, based on the Central Limit Theorem (CLT) [42], the transformation matrix can be directly sampled from a distribution rather than learning or hand crafting it. This eliminates the need to transmit the transformation matrix for accurately decoding weight spaces. Consequently, SINR, as a fundamental compression technique built on the observations of weights spaces, achieves superior compression and higher decoding quality for each data modality compared to the competing methods. Moreover, it can be easily embedded into any INR-based signal compression algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 145, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 145, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 145, + 545 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 224, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 224, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 224, + 565 + ], + "type": "text", + "content": "2.1. Implicit neural representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "INRs have recently gained considerable attention in the computer vision community due to their streamlined network architectures and improved performance in various vision tasks compared to traditional, parameter-heavy models [13, 29, 32]. This surge in interest followed the advent of Neural Radiance Fields (NeRF) [24], which has inspired a plethora of subsequent studies [26, 45]. Further research has explored the pivotal role of different activation functions in INRs [28, 29, 32, 35]. Moreover, INRs provide a universal framework for representing various data modalities. More recent studies have investigated the use of INRs for image classification by transforming standard image for" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 431, + 155 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 431, + 155 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 431, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 431, + 155 + ], + "type": "image", + "image_path": "9f7208e65004e9cbb4c7df590ea7aa703522b4e875f67b2b1af56beeff83f21e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 159, + 521, + 168 + ], + "lines": [ + { + "bbox": [ + 348, + 159, + 521, + 168 + ], + "spans": [ + { + "bbox": [ + 348, + 159, + 521, + 168 + ], + "type": "text", + "content": "(a) INR weight distributions for image representation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 439, + 71, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 439, + 71, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 439, + 71, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 439, + 71, + 553, + 156 + ], + "type": "image", + "image_path": "d7f5c994c8b5cd05b203b8c09a25c6001ba8deef94905b094da92bc8cbf1c56a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 170, + 431, + 253 + ], + "blocks": [ + { + "bbox": [ + 317, + 170, + 431, + 253 + ], + "lines": [ + { + "bbox": [ + 317, + 170, + 431, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 431, + 253 + ], + "type": "image", + "image_path": "99aada467b8b772a54a64fecb40beea654efd40ddd8cf02af08f320ef9d28b7c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 257, + 550, + 267 + ], + "lines": [ + { + "bbox": [ + 317, + 257, + 550, + 267 + ], + "spans": [ + { + "bbox": [ + 317, + 257, + 550, + 267 + ], + "type": "text", + "content": "(b) INR weight distribution for occupancy fields (left) and NeRF (right)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 439, + 170, + 553, + 253 + ], + "blocks": [ + { + "bbox": [ + 439, + 170, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 439, + 170, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 439, + 170, + 553, + 253 + ], + "type": "image", + "image_path": "9795a26df927fde903a81d644b99e10cc7b2b45aeeea5a0b119cbf0a29f0dd81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 554, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 299 + ], + "type": "text", + "content": "Figure 1. Weight distribution of INRs tends to follow a Gaussian distribution for various data modalities." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 374 + ], + "type": "text", + "content": "mats into INRs and training classifiers directly on the INRs' weights and biases [31]. These innovative approaches have showcased the potential of INRs to significantly reduce the dimensionality and computational complexity typically associated with conventional image processing techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 381, + 427, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 427, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 427, + 395 + ], + "type": "text", + "content": "2.2. Signal compression" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 591 + ], + "type": "text", + "content": "Signal compression is crucial for reducing bandwidth needs and saving storage space. With the rise of deep learning, signal compression has evolved into two main approaches: rule-based (traditional) and learning-based methods. Traditional compression methods, such as JPEG for images and MP3 for audio, rely on algorithmic techniques tailored to specific signal types. JPEG minimizes redundancies using the discrete cosine transform [27], while MP3 [4] employs a psycho-acoustic model that enhances compression by removing inaudible sounds through auditory masking. On the other hand, deep learning-based techniques use models trained on vast datasets, adapting to a wide range of signals without predefined algorithms. These methods offer flexibility but require different architectures for each data modality, presenting unique challenges. In this landscape, INRs stand out as a potential universal signal representative." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 600, + 432, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 432, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 432, + 613 + ], + "type": "text", + "content": "2.3. Compressed sensing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 715 + ], + "type": "text", + "content": "Compressed sensing is a field that capitalizes on the inherent sparsity of data to capture information efficiently. In digital imaging, not every pixel is crucial for accurate image reconstruction. Although images appear dense in pixel space, they exhibit considerable redundancy when transformed into different basis functions. This sparsity is exploited by compressed sensing algorithms to reconstruct the original image from fewer sampled data points. These algo" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3062" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 545, + 257 + ], + "type": "image", + "image_path": "7715bfd98d66a74aa851b04fdf64e542e8c0622526ae7901a192935388b90475.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. The proposed SINR compression algorithm: Standard compression techniques for INRs typically involve direct quantization and entropy coding of their weights. However, since natural signals exhibit inherent compressibility in a dictionary, the characteristics that aid in the compressibility of the weight space of an INR are discovered through the Gaussian nature of the weight space. Therefore, SINR employs " + }, + { + "bbox": [ + 54, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 274, + 555, + 340 + ], + "type": "text", + "content": " minimization to identify a higher-dimensional sparse code. Furthermore, based on the weight space observations and the CLT, we simplify the encoding and decoding process using a random sensing matrix controlled by a seed. Subsequently, only the non-zero (NZ) values and their corresponding indices are quantized and entropy coded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 361, + 294, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 294, + 576 + ], + "type": "text", + "content": "rithms employ optimization techniques and linear algebra to solve underdetermined systems, revolutionizing data acquisition in areas such as medical imaging and signal processing. Dictionary learning, integral to compressed sensing, seeks sparse representations of data using dictionary elements or atoms that capture the data's intrinsic structure. These atoms are either predefined or adaptively learned. Compressed sensing's versatility is evident in its applications across various domains, such as image and video compression [44], medical image encryption [17], and classification tasks [14, 15, 19, 21]. It also addresses inverse vision problems like image inpainting [30], deblurring [16, 23], and super-resolution [3]. Recent efforts have merged dictionary learning with deep learning to tackle more complex computer vision challenges, including image recognition [36], denoising [43], and scene recognition [22]. These developments underscore compressed sensing's transformative impact on computer vision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 578, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 638 + ], + "type": "text", + "content": "Our work, SINR, is pioneering the application of compressed sensing principles to INRs. By leveraging these principles alongside the structural distributions of INR weights, SINR identifies redundancies in these spaces, resulting in substantial compression improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 650, + 111, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 111, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 111, + 662 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 246, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 246, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 246, + 685 + ], + "type": "text", + "content": "3.1. Signal representation through INRs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Mathematically, an INR can be defined by a function " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " are the estimizable parameters of the neural net" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "text", + "content": "work. The input and output dimensions of " + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "text", + "content": " vary for different data modalities. In general, " + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "text", + "content": " acts as a mapping from an " + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "text", + "content": "-dimensional input coordinate space to a " + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 361, + 555, + 409 + ], + "type": "text", + "content": "-dimensional output signal space, described as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 400, + 416, + 468, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 416, + 468, + 430 + ], + "spans": [ + { + "bbox": [ + 400, + 416, + 468, + 430 + ], + "type": "interline_equation", + "content": "G _ {\\theta}: \\mathbb {R} ^ {a} \\to \\mathbb {R} ^ {b}.", + "image_path": "e0037723e44ef231247268f8c1aa91be44f0f0c8c982304f9fa821d6a7ee454d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": "For instance, for RGB images, " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "a = 2" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "b = 3" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": ", while for audio signals, " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "a = 1" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "b = 1" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": ". In this architecture, the output of the " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " layer, which feeds into the " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "(i + 1)^{\\text{th}}" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " layer, can be expressed as " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "\\sigma(W^{(i)}y^{(i)} + b^{(i)})" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " denotes the activation function, and " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "y^{(i)}" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " represents the output from the preceding layer. Furthermore, the choice of activation function " + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 313, + 437, + 555, + 546 + ], + "type": "text", + "content": " plays a critical role in shaping the neural network's ability to model complex functions, as explored in various studies [28, 29, 32, 35]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 552, + 476, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 476, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 476, + 566 + ], + "type": "text", + "content": "3.2. Exploring the compressibility" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "According to compressed sensing theory, most real-world signals display sparsity when transformed into an appropriate domain, meaning they can be accurately represented with fewer measurements than traditionally required. Furthermore, real-world signals can be compressed through a set of basis functions, and the coefficients of these functions are derived by minimizing the reconstruction loss. The core concept of INRs involves encoding signals into the weights and biases of an MLP. This process can be viewed as a classical domain transformation technique where pixel values are reconstructed by feeding the corresponding coordinates through the MLP. Unlike predefined signal trans" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3063" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "formers like Fourier [5] and DCT [20], the MLP attempts to minimize the reconstruction loss through backpropagation to find the transformation. The learned representation of the signal resides in another domain. Given that real-world signals inherently exhibit sparsity in transformed domains, we hypothesize that this sparsity can be explored within the MLP's weights. If we can identify where this sparse nature is hidden within the weight space, we could achieve further compression on INRs compared to competing methods. However, identifying this sparse representation within the weights is not straightforward. We believe there are two main approaches to achieving a sparser representation, each with its own challenges and considerations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "content": "The first approach involves either promoting or enforcing a specified level of sparsity in the weights during the training of an INR. Promoting sparsity can generally be achieved by incorporating " + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "content": " regularization on the model parameters, which encourages many weights to approach zero, thereby creating a sparse representation. However, while we observed that " + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 56, + 232, + 296, + 495 + ], + "type": "text", + "content": " regularization results in a higher level of sparsity within the weights, it fails to accurately represent the images. Alternatively, enforcing a specified level of sparsity can also be achieved through model pruning, where weights deemed insignificant are pruned or eliminated during the training process. When it comes to model pruning, we employed both structured and unstructured pruning of weights. We noted that both methods led to significant performance degradation for certain data modalities, particularly for occupancy fields. Moreover, only a small pruning percentages resulted in satisfactory performance for signal representation. For applications that require a high level of generalization, such as NeRFs, the pruning approach did not generalize well, indicating its limitations in achieving a balance between sparsity and performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "The second approach seeks to uncover the inherent structures within the weights that aid in INR compression. This involves identifying patterns or regularities that can be exploited to reduce the dimensionality of the representation without sacrificing performance. We examined this from a dimensionality reduction perspective; however, the weight space in reduced dimensions did not reveal clear patterns, even across different natural images. Nonetheless, we observed that the weight space of an INR often tends to follow a normal distribution. Fig. 1 shows the weight distribution of the hidden layers of an INR when different data modalities are encoded into it. This suggests that INRs share a common pattern across different data modalities, showcasing a potential pathway for a fundamental compression. Given that each weight vector of an INR exhibits Gaussian behavior, we seek a higher-dimensional but sparse equivalent through a dictionary learning-based approach. Let us denote " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{k_1}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " as a hidden weight vector, " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{k_1 \\times k_2}" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "a dictionary, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{k_2}" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " as the corresponding sparse vector. In search of a sparse representation, according to standard compressed sensing, we can write " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{A}\\mathbf{x}" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x} \\|_0 < k_1" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ". To discover the sparse code " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ", the best and most efficient choice is " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " minimization, as " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " minimization iterates through all possible combinations and is therefore not efficient. However, the problem arises with the sensing matrix, commonly referred to as the dictionary " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": ". Although we could use either a dictionary learning-based approach for learning basis functions for the dictionary or a deep learning-based learnable transformation, these approaches would be time-consuming. Furthermore, a TX needs to transmit the learned dictionary alongside the obtained sparse codes. Conversely, we propose that the dictionary does not need to be learned or even transmitted." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": "As we have confirmed, the weights are normally distributed. According to the CLT, a normally distributed random variable can be produced through a finite linear combination of any random variables. In summation form, this can be expressed as: " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "w_{i} = \\sum_{j = 1}^{k_{2}}A_{ij}x_{j}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": "-th element of the weight vector " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "A_{ij}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": " is the element in the " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": "-th column of the sensing matrix " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": "-th element of the vector " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ". To satisfy the CLT, the number of terms in the summation, which is " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "k_{2}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ", should be sufficiently large. Therefore, considering all elements of the weight vector " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ", this can be compactly written as " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{A}\\mathbf{x}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": ". From CLT, the sensing matrix can be defined by a set of random vectors whose appropriate coefficients " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": " can be learned using sparse signal recovery algorithms such as matching pursuit or its variants. Therefore, the optimization can be written as, min " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x}\\| _1" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": " subject to " + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{A}\\mathbf{x}" + }, + { + "bbox": [ + 313, + 254, + 553, + 447 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": "For convenience, let us denote " + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{x} \\|_0" + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": ". A further constraint to the above optimization procedure is that when the sparse code " + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": " is found, we need to store not only its nonzero elements but also the corresponding indices. Therefore, the above " + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": " minimization is solved with " + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "inline_equation", + "content": "2s < k_1" + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": ". We do not apply our compression algorithm to the biases of the INR as the size of bias vectors is very small compared to those of the weight matrices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": "Instead of saving " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "k_{1}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " floating-point numbers for " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ", we now only need to save " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "2s" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " elements: " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " elements are floating-point numbers representing the non-zero values in the sparse code, and the remaining " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " elements are integers that give the indices of those non-zero values. The indices can often be represented with 16-bit precision, unlike the nonzero values in the sparse code, which require 32-bit floating-point precision. At the RX end, " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " must be converted back to " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ". This requires the sensing matrix " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ", which is random and can be controlled by a seed to reproduce the exact " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{A}\\mathbf{x}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ". Thus, the receiver only needs " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": ". This process can be viewed as a method of uncovering the inherent sparsity within natural signals, as represented through the weight space of INRs. As we hypothesized, the" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3064" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "ability to condense natural signals into a dictionary hinges on identifying specific patterns encoded within the weights of INRs. Once the non-zero elements of the sparse vector are obtained at the RX, the resulting procedure is virtually the same across different INR-based baselines. Our method fundamentally achieves compression by delving into the weight spaces to uncover patterns, a step not typically taken by existing methods. A summarization of SINR is illustrated in Fig. 2. As can be seen from Fig. 2, SINR is only dependent on the weights of the INR and is applied prior to any quantization or entropy coding schemes. Therefore, SINR can be applied to existing INR compression methods to improve their compressibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 295, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 258 + ], + "type": "text", + "content": "3.3. How much fundamental compression does SINR achieve?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 149, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 149, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 149, + 274 + ], + "type": "text", + "content": "3.3.1.Standard INRs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": "Consider an INR with " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": " hidden layers, yielding " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "l + 2" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": " total layers. For simplicity, assume " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": " neurons per hidden layer. If the input dimension is " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": " and the output dimension is " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": ", the total number of weight parameters is given by " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s = a \\times k + l \\times k^2 + b \\times k" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": ". However, SINR modifies this structure by reducing the parameters from " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": " in the original network to " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{s_{\\mathrm{INR}}} = a \\times 2s + k \\times l \\times 2s + b \\times 2s" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "inline_equation", + "content": "s \\ll k" + }, + { + "bbox": [ + 55, + 278, + 295, + 387 + ], + "type": "text", + "content": ". Additionally, SINR does not require transmitting any additional data to recover the original INR weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 129, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 129, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 129, + 403 + ], + "type": "text", + "content": "3.3.2. Tiny INRs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": "Let us define an INR as \"tiny\" if the number of neurons in a hidden layer, denoted by " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": ", is less than 50. In such cases, we aim to achieve a sparse representation where " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "2s < k" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\| x \\|_0 = s" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": ". However, achieving a sparse representation that satisfies " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "2s < k" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": " is often extremely challenging and typically does not result in effective compression. To overcome this, we exploit the fact that the weight matrix connecting the " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": " layer to the " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "(i + 1)^{\\text{th}}" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": " layer is of dimensions " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "k \\times k" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": ". By vectorizing this weight matrix, we obtain a vector of dimension " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "k^2 \\times 1" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": ". Given that " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "k^2" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": " is significantly larger than " + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 406, + 295, + 563 + ], + "type": "text", + "content": ", we can apply our SINR procedure directly to the flattened weight matrix. This strategy leads to a sparser representation, thus enhancing compression efficiency for tiny INRs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 121, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 121, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 121, + 578 + ], + "type": "text", + "content": "3.3.3. COIN++" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "In the COIN++ framework, modulation parameters are stored instead of traditional weights and biases, under the assumption that the base network parameters can be transmitted beforehand. For " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " test images, each segmented into " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " patches with a latent dimension of size " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", COIN++ necessitates the transmission of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "m \\times d" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " parameters for reconstructing each image. As the base network in COIN++ conforms to a standard INR structure, it is amenable to further compression via the SINR technique. By implementing SINR principles on the modulations in COIN++, the parameter transmission requirement per image can be reduced from" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "m \\times d" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to just " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "2s \\times d" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "s \\ll m" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". As the size of each test image and the number of images in the test dataset grow, COIN++ would typically require the transmission of numerous parameters. However, by leveraging SINR, both the modulations and the base network can be significantly compressed, achieving enhanced compression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 149, + 493, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 493, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 493, + 163 + ], + "type": "text", + "content": "3.4. Quantization and entropy coding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 167, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 358 + ], + "type": "text", + "content": "After an INR is trained, its parameters are not immediately saved but are first subject to quantization [12]. This involves reducing the bitwidths below typical floating-point precision. Following quantization, the parameters are processed through entropy coding, in our experiments we utilize Brotli coding [1, 18], which allows the compressed data to be stored or transmitted efficiently. To retrieve the original parameters, the decoder must reverse the entropy coding and then perform dequantization. In the case of SINR, the compression process is intensified by utilizing the sparsity induced in the model parameters by natural signals. Once the sparse code is established, the parameters are quantized and subjected to entropy coding. The decoder then reverses the entropy coding and dequantizes the data. Finally, the model parameters are reconstructed by multiplying them with a random Gaussian matrix with a specific seed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 395, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 395, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 395, + 380 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 386, + 429, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 429, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 429, + 399 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 654 + ], + "type": "text", + "content": "SINR, a novel INR compression algorithm, is predicated on the idea that if natural signals are compressible through a dictionary, then INRs should be similarly compressible. This concept underpins SINR's goal to efficiently reduce INR storage requirements while maintaining high fidelity. Our experiments, conducted using the PyTorch framework following WIRE [29] codebase on an NVIDIA RTX A6000 GPU, spanned various data types including images, occupancy fields, and neural radiance fields. Image encoding metrics involved file size, bits per pixel (bpp) and Peak Signal-to-Noise Ratio (PSNR). Occupancy fields were evaluated using file size and Intersection over Union (IoU), and neural radiance fields were assessed using file size and PSNR. Other than the network configurations mentioned in the paper, for occupancy field evaluation, we utilized an MLP with 128 hidden neurons, and 3 hidden layers. For INRIC, we applied the network hyperparameters specified in its paper. In COIN++, we followed the guidelines in its paper but modified the hidden neuron size to 300. All experiments used Brotli entropy coding with a 16-bitwidth (65536 levels) uniform quantizer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 660, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 422, + 672 + ], + "type": "text", + "content": "4.2. How do we find s?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "We implemented " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " minimization using the Orthogonal Matching Pursuit (OMP) algorithm [38]. The OMP algorithm requires the pre-determination of " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " before obtaining" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3065" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "x, and it must adhere to the condition " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "2s < k_{1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "2s" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " is set too low, it results in inaccurate representations of " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " within the weight space. Therefore, we incrementally increased " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " from a low value until " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "2s = k_{1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " for all KODAK images in the " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " experiment, as outlined in Sec. 4.3. Our findings suggest that the optimal value of " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " for successfully reconstructing the weight space does not depend on the specific image but on the number of neurons in a hidden layer. By adjusting the neuron count, we identified an optimal " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " that accurately reconstructs the weight space while satisfying the specified constraint. Extending these experiments to natural signals outside the KODAK dataset confirmed the consistency of our results. Additionally, we have included a regression plot in the supplementary that details how to determine the optimal " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " based on the number of neurons." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 152, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 152, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 152, + 275 + ], + "type": "text", + "content": "4.3. Image encoding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "text", + "content": "Representing an image through the weights and biases of a neural network serves as a method of encoding. For our image encoding task, we utilized the KODAK dataset, which includes 24 natural RGB images, each measuring " + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "inline_equation", + "content": "768 \\times 512" + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "text", + "content": " pixels. We conducted five types of experiments, denoted as " + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 281, + 295, + 364 + ], + "type": "text", + "content": " ranges from 1 to 5, to demonstrate the effectiveness of our proposed method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": "Experiment " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": " involved encoding each image in the KODAK dataset using an INR without positional embedding, by varying the number of neurons in each hidden layer. Experiment " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": " mirrored " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": ", but with the variation in the number of hidden layers instead. Experiments " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "C_4" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": " implemented the meta-learning approach for INRs proposed in INRIC, without and with positional embedding for the input layer, respectively. Experiment " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "C_5" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": " involved the COIN++ framework, testing both with and without patching. When using patching, we adopted " + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 365, + 295, + 581 + ], + "type": "text", + "content": " patches as suggested by COIN++. However, we observed that without patching, even as the latent modulation dimension increased, the average PSNR obtained by COIN++ remained nearly constant. For these meta-learning-based experiments, we used the first 12 images of the KODAK dataset for meta-learning and the remaining 12 images for fine-tuning. For all image encoding experiments, we used the Sinusoidal activation function (see supplementary)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Let us define " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " as the number of hidden layers and the number of neurons per hidden layer in an INR, respectively. For experiment " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", we configured the INR with settings " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "(h,m)" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "(2,32)" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "(3,64)" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "(3,128)" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ". Experiment " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " aims to assess the effectiveness of SINR by varying the number of hidden neurons. The results, depicted in Fig. 3, demonstrate how effectively SINR identifies the compressibility of the weight space. This is indicated by the bpp values, which reflect the size of the model parameters. For example, representing the KODAK dataset with an average PSNR of 30 dB requires about 3.7 bpp for COIN and 2.0" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 225 + ], + "type": "image", + "image_path": "6fb298a998e1ce91f83f6d74ab8f38c95f2f6d9414b5df8f797ecebe9d5efd68.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "text", + "content": "Figure 3. Experiments " + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "text", + "content": ": Identifying compressible INR combinations. The SINR approach demonstrates that configurations in " + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "text", + "content": " are more compressible than those in " + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 313, + 234, + 555, + 289 + ], + "type": "text", + "content": ". Furthermore, in both configurations SINR achieves lower bpp while maintaining the PSNR values." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": "bpp for INRIC. However, SINR significantly reduces the bpp to approximately 1.7 using the same quantizer and entropy coder. The first configuration in " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": " falls under the category of tiny INRs, underscoring the proposed method's effectiveness even for compact INRs. As illustrated in Fig. 3, SINR achieves the same level of PSNR as baselines with a lower bpp for any network configuration. This substantial reduction of bpp across the " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": " experiment showcases the efficiency and compactness achieved by SINR. From " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": ", it can be established that greater compressibility of an INR into a dictionary is possible with an increased number of hidden neurons. Following the conclusions drawn from experiment " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": ", experiment " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": " was designed to explore the impact of increasing the number of hidden layers on the effectiveness of SINR. The configurations tested in " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": " were " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "(h,m) = \\{(3,64),(5,64),(7,64)\\}" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": ". As illustrated in Fig. 3, SINR consistently achieved PSNR levels comparable to baseline methods, but with a reduced bpp. Given that " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": " maintained a constant neuron count at 64, the observed deviations in compression between SINR and INRIC were less significant than those observed in " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": ". This discrepancy can be attributed to the following: an INR configuration with a higher number of neurons (e.g., " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "m = 128" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": "), even with fewer hidden layers (e.g., " + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "inline_equation", + "content": "h = 2" + }, + { + "bbox": [ + 313, + 318, + 555, + 715 + ], + "type": "text", + "content": "), possesses more trainable parameters. Consequently, such a model is capable of learning a more robust representation of the image compared to configurations with a larger number of layers but fewer neurons per layer. As a result, the compressible characteristics of the images are more effectively transferred into the model parameters during the INR training process. This leads to a more compressible INR. These findings support the premise that if natural images can be efficiently compressed into a dictionary, the weight space" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3066" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 225 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 225 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 225 + ], + "type": "image", + "image_path": "88323c6069865b12560ac00de8bb3b353d9a17ab289d138ce0eaec8b4dc95a3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "text", + "content": "Figure 4. Experiments " + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "inline_equation", + "content": "C_4" + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "inline_equation", + "content": "C_5" + }, + { + "bbox": [ + 55, + 234, + 296, + 313 + ], + "type": "text", + "content": ": Identifying compressible INR combinations under Meta-Learning. Meta-learning approaches have been introduced for INRs to enhance their generalization abilities and achieve faster convergence. When assessing induced sparsity in the weight space, SINR demonstrates a significant reduction in bpp values while maintaining nearly the same PSNR performance as the baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 234, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 234, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 234, + 350 + ], + "type": "text", + "content": "of INRs can also be effectively compressed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": "As in previous experiments, each image required separate training of an INR. Experiments " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "C_4" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": " address this challenge through meta-learning, with and without positional embedding, respectively. The configurations for these INRs are given by " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "(h,m) = \\{(3,32),(3,64),(3,96),(3,128)\\}" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": ". For COIN++, the number of layers was set to 5 with MLP's hidden dimension at 300. The latent dimension parameter " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "(d)" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": " varied as follows: " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "d = \\{16,32,64,96\\}" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": ". Fig. 4 presents the experimental results for " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "C_4" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "inline_equation", + "content": "C_5" + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": ", illustrating significant compression capabilities of the proposed SINR within a meta-learning framework. Notably, models using positional embedding generally have more parameters than those without." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "Comparing the performance of INRIC and SINR without positional embedding schemes, the initial INR configuration shows that SINR exhibits a lower bpp for the same average PSNR. Generally, as bpp increases, the representation capacity of the INR enhances, leading to more robust image representation. From the graphs, as bpp increases, SINR demonstrates greater improvements in PSNR than INRIC, a phenomenon that can be explained by the aforementioned logic. In the case of COIN++, the approach focuses on fine-tuning only the modulations using their proposed meta-learning method. However, since finetuning encodes natural signals within these modulations, they should also be compressible via a dictionary. Due to patching, each KODAK test image results in a " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "d \\times 384" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " matrix. Our experiments reveal that these modulations encode hidden redundancies in natural signals. For instance, to achieve an average PSNR of approximately 24.2 dB," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "COIN++ requires more than 1.5 bpp; however, the same PSNR can be achieved with COIN++ using just under 1 bpp by exploiting the hidden sparsity in its modulations through our proposed approach. Therefore, when a high-capacity model effectively represents a signal, it must encapsulate this sparsity within its weight and bias spaces. SINR explores and removes redundancies in these parameters, retaining only essential information. Fig. 5 showcases the decoded images by SINR alongside with the INR based image compressors. Decoded PSNR, BPP, and file size are displayed in the first, second, and third rows of the text boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 211, + 463, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 463, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 463, + 224 + ], + "type": "text", + "content": "4.4. Occupancy fields encoding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 587 + ], + "type": "text", + "content": "Occupancy fields are represented by binary values, either 1 or 0, where 1 denotes that the signal lies within a specified region and 0 indicates its absence. Another variant of occupancy volumes stores not only the presence or absence of a signal but also the color at that location. Typically, occupancy fields consume more space than other data modalities. However, they can be represented with higher accuracy and lower storage requirements using INRs. In this experiment, we followed the sampling procedure described in [29]. Occupancy fields can be thought of as representations of three-dimensional objects, capturing natural signals. Despite following the sampling procedure, redundancies may exist that are not essential for representing the occupancy volume. Identifying these redundancies can reduce storage requirements. However, identifying them in the spatial domain " + }, + { + "bbox": [ + 313, + 228, + 555, + 587 + ], + "type": "inline_equation", + "content": "(xyz)" + }, + { + "bbox": [ + 313, + 228, + 555, + 587 + ], + "type": "text", + "content": " requires domain-specific algorithms, as described in Sec. 1. As INRs serve as unified data modality representatives, these redundancies must be encapsulated within its weights space. SINR fundamentally compresses the INRs into a dictionary regardless of the data modality; therefore, indeed it is equally applicable to occupancy fields. To validate this hypothesis for occupancy fields, two experiments were conducted using shapes from the Stanford shape dataset [33]. Figure 6 showcases the decoded SINR's representations for 'Thai Statue'(first volume) and 'Lucy' (fourth volume) datasets alongside the existing INR-based occupancy compressor. We use the Gaussian activation function for this task (see supplementary). The first value and second value in each text box represent the IoU metric and storage requirement, respectively, except for GT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 434, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 434, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 434, + 605 + ], + "type": "text", + "content": "4.5. Additional materials" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "type": "text", + "content": "The pseudocode for SINR, additional results and ablation studies on finding " + }, + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "type": "text", + "content": " are in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 645, + 388, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 388, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 388, + 658 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "content": "Implicit Neural Representations (INRs) have emerged as a promising framework for unified data modality representation. Several studies have explored the potential for compressing images, occupancy fields, and audio using INRs." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3067" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 68, + 506, + 290 + ], + "blocks": [ + { + "bbox": [ + 106, + 68, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 68, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 106, + 68, + 506, + 290 + ], + "type": "image", + "image_path": "83679984da93e168f814fc55fb8341af0f970ab79b9c631697abfb21be4c514a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 333 + ], + "type": "text", + "content": "Figure 5. Results for image encoding experiment. SINR compresses the INR into a dictionary, significantly reducing the storage required compared to baseline INR image compressors. The results demonstrate that the decoded representations undergo a very negligible loss in PSNR, which is minimal considering the substantial storage space saved." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 103, + 342, + 508, + 495 + ], + "blocks": [ + { + "bbox": [ + 103, + 342, + 508, + 495 + ], + "lines": [ + { + "bbox": [ + 103, + 342, + 508, + 495 + ], + "spans": [ + { + "bbox": [ + 103, + 342, + 508, + 495 + ], + "type": "image", + "image_path": "45c9e37db49849fa961df1a6bd38e435297cb6285d7b30113c5b1768b2941228.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 503, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 503, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 555, + 537 + ], + "type": "text", + "content": "Figure 6. Results for occupancy fields encoding experiment. The results clearly demonstrate that SINR achieves the smallest file size and the highest accuracy metric for every shape in the tested dataset. The significant compression obtained by our algorithm suggests that occupancy fields, when represented using an INR, can be more efficiently compressed into a dictionary compared to images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": "However, none of these methods have investigated whether the INR itself can be compressed prior to quantization and entropy coding. As natural signals can be efficiently compressed in bases of transformed domains due to their sparsity—allowing for higher accuracy and lower storage requirements—we hypothesize that a similar compressible nature must also exist in the INR once it is trained. With the discovery that weight vectors in the weight space tend to adhere to a Gaussian distribution, we propose SINR, which compresses any INR in a dictionary. Furthermore, we demonstrate that this dictionary does not need to be learned but can instead be generated using a seed. We compare our findings with standard INR compressors for im" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 558, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 703 + ], + "type": "text", + "content": "ages, occupancy fields, and neural radiance fields. SINR achieves fundamental compression for any INR, independent of other post-processing methods such as quantization and entropy coding, and it showcases significantly lower storage requirements and higher fidelity across various data modalities. Through our experiments, we observed that the INR can be more compressed when a more robust representation of the signal is learned. Additionally, some data modalities exhibit greater compressibility than others. We firmly believe this research will aid other researchers in exploring more patterns in the weight spaces of INRs and in developing operators and transforms for INR." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3068" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 224 + ], + "type": "text", + "content": "This work is supported by the Intelligence Advanced Research Projects Activity (IARPA) via Department of Interior/ Interior Business Center (DOI/IBC) contract number 140D0423C0076. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, DOI/IBC, or the U.S. Government." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 254, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 254, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 296, + 308 + ], + "type": "text", + "content": "[1] Jyrki Alakuijala, Andrea Farruggia, Paolo Ferragina, Eugene Kliuchnikov, Robert Obryk, Zoltan Szabadka, and Lode Vandevenne. Brotli: A general-purpose data compressor. ACM Transactions on Information Systems (TOIS), 37(1):1-30, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 310, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 310, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 61, + 310, + 296, + 376 + ], + "type": "text", + "content": "[2] David Alexandre, Chih-Peng Chang, Wen-Hsiao Peng, and Hsueh-Ming Hang. An autoencoder-based learned image compressor: Description of challenge proposal by nctu. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2539-2542, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 377, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 377, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 377, + 296, + 421 + ], + "type": "text", + "content": "[3] Selen Ayas and Murat Ekinci. Single image super resolution using dictionary learning and sparse coding with multi-scale and multi-directional gabor feature representation. Information Sciences, 512:1264-1278, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "type": "text", + "content": "[4] Karlheinz Brandenburg. Mp3 and aac explained. In Audio Engineering Society Conference: 17th International Conference: High-Quality Audio Coding. Audio Engineering Society, 1999. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 467, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 489 + ], + "type": "text", + "content": "[5] Komaravolu Chandrasekharan. Classical Fourier Transforms. Springer Science & Business Media, 2012. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 490, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 296, + 533 + ], + "type": "text", + "content": "[6] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Energy compaction-based image compression using convolutional autoencoder. IEEE Transactions on Multimedia, 22(4):860-873, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 535, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 296, + 590 + ], + "type": "text", + "content": "[7] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 624 + ], + "type": "text", + "content": "[8] Fabio Duarte. Amount of data created daily (2024). https://explodingtopics.com/blog/data-generated-per-day, 2024. Accessed: 2024-07-14. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 667 + ], + "type": "text", + "content": "[9] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. arXiv preprint arXiv:2103.03123, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 296, + 713 + ], + "type": "text", + "content": "[10] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[11] Sharath Girish, Abhinav Shrivastava, and Kamal Gupta. Shacira: Scalable hash-grid compression for implicit neural representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17513-17524, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "type": "text", + "content": "[12] Robert M. Gray and David L. Neuhoff. Quantization. IEEE transactions on information theory, 44(6):2325-2383, 1998. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "text", + "content": "[13] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Implicit neural representations with levels-of-experts. Advances in Neural Information Processing Systems, 35:2564-2576, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "type": "text", + "content": "[14] Daniel J Hsu, Sham M Kakade, John Langford, and Tong Zhang. Multi-label prediction via compressed sensing. Advances in neural information processing systems, 22, 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 285 + ], + "type": "text", + "content": "[15] Junlin Hu and Yap-Peng Tan. Nonlinear dictionary learning with application to image classification. Pattern Recognition, 75:282-291, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 330 + ], + "type": "text", + "content": "[16] Zhe Hu, Jia-Bin Huang, and Ming-Hsuan Yang. Single image deblurring with adaptive dictionary learning. In 2010 IEEE International Conference on Image Processing, pages 1169-1172. IEEE, 2010. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 386 + ], + "type": "text", + "content": "[17] Donghua Jiang, Nestor Tsafack, Wadii Boulila, Jawad Ahmad, and JJ Barba-Franco. Asb-cs: Adaptive sparse basis compressive sensing model and its application to medical image encryption. Expert Systems with Applications, 236: 121378, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 410 + ], + "type": "text", + "content": "[18] Gareth A Jones and J Mary Jones. Information and coding theory. Springer Science & Business Media, 2012. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "type": "text", + "content": "[19] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 487 + ], + "type": "text", + "content": "[20] Syed Ali Khayam. The discrete cosine transform (dct): theory and application. Michigan State University, 114(1):31, 2003. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 554, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 554, + 522 + ], + "type": "text", + "content": "[21] Li Liu and Paul Fieguth. Texture classification using compressed sensing. In 2010 Canadian Conference on Computer and Robot Vision, pages 71-78. IEEE, 2010. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "text", + "content": "[22] Yang Liu, Qingchao Chen, Wei Chen, and Ian Wassell. Dictionary learning inspired deep network for scene recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 569, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 554, + 612 + ], + "type": "text", + "content": "[23] Liyan Ma, Lionel Moisan, Jian Yu, and Tieyong Zeng. A dictionary learning approach for poisson image deblurring. IEEE Transactions on medical imaging, 32(7):1277-1289, 2013. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3069" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[26] AKM Rabby and Chengcui Zhang. Beyondpixels: A comprehensive review of the evolution of neural radiance fields. arXiv preprint arXiv:2306.03000, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "text", + "content": "[27] AM Raid, WM Khedr, Mohamed A El-Dosuky, and Wesam Ahmed. Jpeg image compression using discrete cosine transform-a survey. arXiv preprint arXiv:1405.6147, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[28] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "type": "text", + "content": "[29] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 338 + ], + "type": "text", + "content": "[30] K Seemakurthy, A Majumdar, J Gubbi, NK Sandeep, A Varghese, S Deshpande, M Girish Chandra, and P Balamurali. Deep dictionary learning for inpainting. In Computer Vision, Pattern Recognition, Image Processing, and Graphics: 7th National Conference, NCVPRIPG 2019, Hubballi, India, December 22-24, 2019, Revised Selected Papers 7, pages 79-88. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[31] Aviv Shamsian, Aviv Navon, David W Zhang, Yan Zhang, Ethan Fetaya, Gal Chechik, and Haggai Maron. Improved generalization of weight space networks via augmentations. arXiv preprint arXiv:2402.04081, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 381, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 436 + ], + "type": "text", + "content": "[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 2, 3, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[33] Stanford University Computer Graphics Laboratory. The stanford 3d scanning repository. https://graphics.stanford.edu/data/3Dscanrep/. Accessed: 2024-07-15. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[34] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74-91. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 591 + ], + "type": "text", + "content": "[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[36] Hao Tang, Hong Liu, Wei Xiao, and Nicu Sebe. When dictionary learning meets deep learning: Deep dictionary learning and coding network for image recognition with limited data. IEEE transactions on neural networks and learning systems, 32(5):2129-2141, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[37] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International conference on learning representations, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "text", + "content": "[38] Joel A Tropp and Anna C Gilbert. Signal recovery from random measurements via orthogonal matching pursuit." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 343 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE Transactions on information theory, 53(12):4655-4666, 2007. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 129 + ], + "type": "text", + "content": "[39] Gregory K Wallace. The jpeg still picture compression standard. IEEE transactions on consumer electronics, 38(1): xviii-xxxiv, 1992. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 152 + ], + "type": "text", + "content": "[40] Jan Hendrik Weissenbruch. The shipping canal at rijswijk, known as \"the view at geestbrug\". Painting. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "text", + "content": "[41] Yueqi Xie, Ka Leong Cheng, and Qifeng Chen. Enhanced invertible encoding for learned image compression. In Proceedings of the 29th ACM international conference on multimedia, pages 162-170, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[42] Xijuan Zhang, Oscar L Olvera Astivia, Edward Kroc, and Bruno D Zumbo. How to think clearly about the central limit theorem. Psychological Methods, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "text", + "content": "[43] Hongyi Zheng, Hongwei Yong, and Lei Zhang. Deep convolutional dictionary learning for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 630-641, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 309 + ], + "type": "text", + "content": "[44] Jinjia Zhou and Jian Yang. Compressive sensing in image/video compression: Sampling, coding, reconstruction, and codec optimization. Information, 15(2):75, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "text", + "content": "[45] Fang Zhu, Shuai Guo, Li Song, Ke Xu, Jiayu Hu, et al. Deep review and analysis of recent nerfs. APSIPA Transactions on Signal and Information Processing, 12(1), 2023. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3070" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_content_list.json b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa06d6220837aa304d05e362d58357672253856c --- /dev/null +++ b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "SIR-DIFF: Sparse Image Sets Restoration with Multi-View Diffusion Model", + "text_level": 1, + "bbox": [ + 114, + 130, + 880, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yucheng Mao $^{1, *}$ Boyang Wang $^{1, *}$ Nilesh Kulkarni $^{1,2}$ Jeong Joon Park $^{1}$ $^{1}$ University of Michigan, Ann Arbor $^{2}$ Netflix", + "bbox": [ + 189, + 178, + 802, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/876151be1a6b5e7ea8334fc64c810db46a229e8fd9782d05bb108244540f63a5.jpg", + "image_caption": [ + "Figure 1. Sparse View Image Restoration. Our diffusion model takes multi-view images and jointly enhances their visual quality while maintaining 3D consistency. (Top) Four motion-blurred input images are processed by our method, resulting in sharp outputs that significantly outperform single-view restoration methods as shown in the corresponding purple boxes. (Bottom) Our method can consistently restore multi-view images (4 out of 50 shown), leading to accurate 3D reconstructions." + ], + "image_footnote": [], + "bbox": [ + 127, + 253, + 867, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 664, + 326, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The computer vision community has developed numerous techniques for digitally restoring true scene information from single-view degraded photographs, an important yet extremely ill-posed task. In this work, we tackle image restoration from a different perspective by jointly denoising multiple photographs of the same scene. Our core hypothesis is that degraded images capturing a shared scene contain complementary information that, when combined, better constrains the restoration problem. To this end, we implement a powerful multi-view diffusion model that jointly generates uncorrupted views by extracting rich information from multi-view relationships. Our experiments show that our multi-view approach outperforms existing single-view im", + "bbox": [ + 89, + 681, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age and even video-based methods on image deblurring and super-resolution tasks. Critically, our model is trained to output 3D consistent images, making it a promising tool for applications requiring robust multi-view integration, such as 3D reconstruction or pose estimation. Project website: https://myc634.github.io/sirdiff/", + "bbox": [ + 511, + 665, + 908, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 773, + 645, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image restoration approaches [8, 16, 29, 53] typically process degraded images in isolation, confronting the challenge of the original scene content. Yet in practice, we frequently capture multiple photographs of meaningful moments, yielding a natural collection of views of the same scene, albeit some degraded. This common behavior presents an untapped opportunity: leveraging multiple captures to enhance image", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 114, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21620", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality and better preserve our visual memories.", + "bbox": [ + 89, + 90, + 410, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present a novel perspective on image restoration by reformulating it as a multi-view collaborative restoration task. We leverage our key insight that multiple corrupted photographs of the same scene contain complementary information that can aid superior denoising outcomes. The concept of multi-image fusion for quality enhancement has been explored in classical super-resolution techniques where sub-pixel shifted images are combined to reconstruct fractional pixel information [19]. Our approach revisits this paradigm in a fundamentally wider context by using neural attention mechanisms to implicitly fuse overlapping information from general multi-view images to enhance even severe aberrations.", + "bbox": [ + 89, + 116, + 485, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notably, our multi-view restoration framework has particular significance for 3D computer vision applications. Many real-world tasks, from robotic SLAM to novel-view synthesis, rely fundamentally on multi-view imagery of a scene. When restoring degraded images within such multi-view sets, maintaining geometric consistency across views is crucial to preserve the underlying 3D scene assumptions. Traditional single-view restoration methods, inevitably introduce inconsistencies that can compromise these geometric constraints. In contrast, our restoration approach naturally enforces view consistency through learned neural priors.", + "bbox": [ + 89, + 321, + 486, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We implement our approach, dubbed SIR-Diff (Sparse Image Restoration with Diffusion), through an innovative multi-view diffusion model architecture. Our choice of diffusion models is motivated by their natural handling of uncertainty in image restoration. We extend the UNet-based Stable Diffusion [35] architecture to process multiple views jointly by introducing three key components: (1) a unified encoder for diverse degraded inputs including low-resolution, blurred, and regular RGB images, (2) a hybrid 2D-3D convolutional structure derived from SVD (Stable Video Diffusion) [4, 5], and (3) a 3D self-attention Transformer that enables cross-view attention across all spatial and multi-view tokens.", + "bbox": [ + 89, + 498, + 486, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate and show zero-shot performance on Scannet++ [61], ETH3D [38], and CO3D [32] for multi-view deblurring and super-resolution tasks and their 3D downstream applications. SIR-Diff consistently outperforms existing single-image and video restoration methods for multi-view restoration tasks in both image quality and our proposed 3D view consistency metrics. For downstream applications, we show that our multi-view restoration strategy significantly enhances 3D reconstruction quality, as evidenced by improved novel-view synthesis accuracy and feature-matching success rates. Beyond image restoration, our architecture shows promise in other multi-view tasks, notably achieving competitive depth estimation accuracy and geometric consistency compared to leading single-view methods.", + "bbox": [ + 89, + 689, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Diffusion Models. Diffusion Models (DM) show huge popularity in all domains: illustrations [9, 35, 51, 65], video creation [4, 5, 30], 3D modeling with limited input views [27, 41, 45], 4D generation [2, 3, 26], and etc. Most models are fine-tuned from the pre-trained weight instead of training from the beginning. One of the most widely-used pre-trained Diffusion Models is Stable Diffusion 2.1 [35], a UNet-based [36] architecture.", + "bbox": [ + 511, + 114, + 906, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiview-Image Diffusion models. In the image-to-3D task [24], Diffusion Model is a powerful generative prior to enriching details for the novel (unseen) view synthesis [6, 17, 44, 55, 56]. Methods like MVDream [41], CAT3D [13], Zero $123++$ [40] jointly generate multi-views simultaneously by extending the image diffusion model (SD2.1) with a 3D self-attention layer. However, these works ignore the exploration of the convolution layer architecture in the UNet-based Diffusion Model. In this paper, we explore the methodology of applying the 3D self-attention layer to the 3D restoration task.", + "bbox": [ + 511, + 237, + 908, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image and Video Restoration. These methods aim at converting a degraded input image or video into a high-quality form. Traditional tasks include Super-Resolution [23, 53, 57, 62, 68], Deblurring [22, 25, 31], Denoise [10, 59], Compression Artifact Reduction [8, 46, 47], Inpainting [7, 58], Burst Image Deblurring [1, 14], and etc. These domains have achieved milestones in the 2D image and video domain [28]. However, the study on the restoration in 3D is insufficient. The restoration consistency of the texture and geometry from different views is hard to maintain when independently restoring them.", + "bbox": [ + 511, + 404, + 908, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Reconstruction from Corrupted Inputs. 3D Gaussian Splatting (3DGS) [21, 33, 54, 63] transforms input images into explicit-represented 3D Gaussian blobs with continuous optimization to reconstruct the large 3D scene. However, these models heavily rely on the condition of the input images. Motion blurring [50, 67] or low-resolution [12, 39] inputs that lack high-frequency and sharp details may greatly deteriorate the reconstruction quality or even fail at the initialization stage. To resolve this, methods like Bad-Gaussian [39] and SuperGaussian [67] depend on dense view input to iteratively optimize the degraded artifacts. Further, under heavily deteriorated inputs, these methods cannot recover high-quality outputs when they lack a strong generative prior [39] such as a pre-trained Diffusion Model.", + "bbox": [ + 511, + 571, + 910, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 799, + 604, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss our proposed approach, SIR-Diff, which can generate and restore sparse image sets from 3D scenes or objects conditioned on their degraded inputs. Specifically, we showcase our approach to two challenging tasks: motion deblurring and super-resolution. In the", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21621", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/968f1a4edd490f63fed9f564b435e81eb7ef4e209a684bbea0c3a4b5ff791c45.jpg", + "image_caption": [ + "Figure 2. Method Overview. Our model is a UNet-based latent diffusion model that takes as input a degraded sparse image set and outputs a set of restored consistent images. The core of our approach is a diffusion model that jointly denoises image latents across multiple views by relying on our spatial-3D ResNet and 3D Self-Attn transformer. We use pre-trained encoder-decoders from SD2.1 [35] and train our denoising UNet on those encoded latents. We show view-consistent results for the task of Deblurring and Super-Resolution." + ], + "image_footnote": [], + "bbox": [ + 96, + 79, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "following sections, we discuss our problem setup in § 3.1, multi-view image restoration diffusion model in § 3.2, and our proposed model architecture in § 3.3.", + "bbox": [ + 89, + 395, + 483, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-View Generation Setup", + "text_level": 1, + "bbox": [ + 89, + 450, + 354, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to restore sparse view images from degraded inputs. Typical methods to solve this problem work either on a single image [31, 35, 57] or require a coherent video [25, 68] as input. Toward the goal of sparse view image restoration, we propose a generative model that can respect the geometrical priors to perform the task of image restoration from blurry images or recovering high-resolution image data from its low-resolution counterpart. Any single-image generative model that performs independent inference on each image cannot ensure consistency over the generated image set.", + "bbox": [ + 88, + 473, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate these issues, we formulate this as a generation problem to infer all the required views of the scene by modeling the joint distribution using a multi-view diffusion model. Our model takes a degraded image set as input and recovers a high-quality self-consistent image set.", + "bbox": [ + 89, + 625, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-View Image Restoration Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 710, + 482, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model aims to restore images when conditioned on a corrupted image set $\\{I^c\\}_{i = 1}^N$ . The model's primary objective is to restore a self-consistent high-quality image set. We do this by using a multi-view image restoration diffusion model that can jointly restore the image set. In training, we start from a dataset of non-blurry / high-resolution images $\\{I^r\\}_{i = 1}^N$ and apply the forward diffusion process. Our input representation to the N-view diffusion model is $\\mathbf{x} = \\{I_1^r,\\dots,I_N^r\\}$ . Specifically, $\\mathbf{x}^0$ corresponds to a clean signal at the start of the forward diffusion process at step 0.", + "bbox": [ + 89, + 734, + 483, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our generative model learns to reverse this noising pro", + "bbox": [ + 109, + 885, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cess by predicting the added noise at every timestep of the diffusion process.", + "bbox": [ + 511, + 395, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 436, + 700, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We build our joint image restoration diffusion model based on a latent diffusion model [35]. Fig. 2 shows an overview of our method. The goal of our diffusion model is to denoise image latents such that a pre-trained decoder can convert them to high-quality images. When simultaneously denoising multiple image latents, it is important that latent representation remain consistent across views. We enable this by allowing the diffusion model to attend to latents from other views.", + "bbox": [ + 511, + 459, + 908, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Latent Diffusion Models for Restoration", + "text_level": 1, + "bbox": [ + 511, + 590, + 839, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A Latent Diffusion Model [35] architecture, such as Stable Diffusion 2.1 (SD2.1), consists of two parts a) encoder and decoder and b) denoising UNet.", + "bbox": [ + 511, + 609, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, our latent denoising UNet progressively refines pure noise into a clean latent representation, conditioned on the degraded image set, through several denoising steps. Once the latent representation is clean, a decoder translates it into the restored image set.", + "bbox": [ + 511, + 657, + 906, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, during training, our diffusion model is optimized to learn the denoising process of the latent representation, conditioned on the degraded input sets.", + "bbox": [ + 511, + 732, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder-Decoder. We initialize our encoder with the weights from an SD2.1 [35] and are frozen when we train our model. The encoder, $\\phi$ , is used to map our images to latent vectors which are used to learn the UNet-Style diffusion model. The decoder, $\\delta$ , converts the denoised latents from the diffusion model to pixels in image space.", + "bbox": [ + 511, + 779, + 908, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UNet. Our latent denoising UNet, $\\mathbf{s}_{\\theta}(\\cdot)$ with parameters $\\theta$ , considers as input a noisy latent $\\mathbf{x}^k$ , encoded latents for", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21622", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2043b64822d415bb507354c5896c9505e030d7cbf1cc20b43e01469320497cee.jpg", + "image_caption": [ + "Figure 3. Qualitative Comparisons on Motion Deblurring and Super-Resolution. The left side is the Motion-Deblurring and the right side is the Super-Resolution results. We refer readers to supplementary for more visualizations. Zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 94, + 84, + 903, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "degraded images, $\\{I^c\\}_{i = 1}^N$ and the diffusion noise $k$ to output an of estimate the noise, $\\hat{\\pmb{\\epsilon}}$ , as shown in Eq. (1):", + "bbox": [ + 89, + 404, + 480, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} = \\mathbf {s} _ {\\theta} (\\mathbf {x} ^ {k}, \\phi (\\{I ^ {c} \\} _ {i = 1} ^ {N}), k). \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 444, + 482, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We supervise the outputs of our UNet to match the noise added during the forward diffusion process using Eq. 2:", + "bbox": [ + 89, + 472, + 482, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon \\sim \\mathcal {N} (0, I), k \\sim \\mathcal {U} (T)} \\| \\epsilon - \\hat {\\epsilon} \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 510, + 482, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Adapting Diffusion Model to multiple views", + "text_level": 1, + "bbox": [ + 89, + 539, + 441, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Starting with the image diffusion model, we expand a new dimension to represent image sets from different views. Most of the time, batch and views are merged into one dimension and split when needed. We encode the degraded image set with the same VAE encoder as the target noisy latents. Noisy latents and the condition latents are integrated by channel-wise concatenation. Then, due to channel number change, we replace the original convolution with a new zero-initialized convolution layer [65] to adapt the new number of channels input.", + "bbox": [ + 88, + 556, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Spatial-3D Integrated UNet Architecture", + "text_level": 1, + "bbox": [ + 89, + 715, + 421, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typical latent diffusion models such as Stable Diffusion [35] do not consider denoise multiple images simultaneously and often generate individual images that lead to inconsistent image generations between multi-views. Our solution to remove this inconsistency is to denoise all the views jointly by allowing the UNet to look at the intermediate latent representations from different views. As shown in the right side of Fig. 2, our architecture is composed of a Spatial-3D ResNet and a 3D self-attention Transformer.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial-3D ResNet. Image diffusion model like SD2.1 only has a 2D convolution layer in the ResNet, which makes it", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hard to perform well in tasks that require 3D understanding. We want to enhance the 3D performance of the ResNet of the image diffusion model by introducing a 3D convolution layer to better capture geometry information as shown in Fig. 2. Training from raw will be unstable and time-consuming, but we find that the video diffusion model that is performing the image-to-video task already has a pre-trained 3D convolution layer. Understanding temporal similarity in the time dimension and understanding 3D similarity in the spatial dimension share similarity. By the experiment, we empirically find that merging two different weights into one is feasible and can lead to better performance. Hence, we initialize the 2D Convolution layer with image Diffusion Model weight (SD2.1) and the 3D convolution layer with video Diffusion weight (Stable Video Diffusion [4]). The output of each layer will be blended as follows: $O_{\\mathrm{ResNet}} = \\sigma (\\alpha)\\times O_{2D} + \\sigma (1 - \\alpha)\\times O_{3D}$ , where $\\sigma$ is the sigmoid function and $\\alpha$ is usually taken as 0.5 following SVD [4]. $O_{2D}$ is the output of 2D ResNet and $O_{3D}$ is the output of 3D ResNet layer.", + "bbox": [ + 511, + 404, + 906, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Self-Attention Transformer. Furthermore, we adopt the 3D multi-view consistent self-attention Transformer as shown in Fig. 2, similar to previous works [13, 41]. Our self-attention layer considers image latents from all the $N$ views and performs self-attention. After ResNet, the model patches the latent to create $p$ patch token per view, and in total, we have $N \\times p$ tokens that are part of the attention layer. Each token calculates the attention score to all $N \\times p$ tokens as shown on the right side of Fig. 2. This feature enables each token to be accessible to all conditional image sets given at the maximum level. Further, by this design, our model, in the inference stage, is scalable to a variable number of input frames with stable performance even though", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21623", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f5227dca3fa4b54d0b2639c28e2500d3399bf183043624efb0caee8607459bfa.jpg", + "table_caption": [ + "Table 1. Sparse-View Deblurring Comparisons. We show quantitative results on three datasets and evaluate metrics as in § 4.1. VConsis. (visual consistency) is multiplied by 100. Across all methods, SIR-Diff is the most visually consistent as compared to all the baselines. On other standard metrics, we also outperform baselines. SIR-Diff (1 Frame) is a single image version of our method that does not depend on multiple views. The best is highlighted, and the second is underlined." + ], + "table_footnote": [], + "table_body": "
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodFID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓
PromptIR [31]81.280.2487.81150.340.53710.72166.30.35012.89
Restromer [64]49.720.2326.5273.470.3309.59144.80.42611.00
VRT [25]134.50.3717.67160.40.53011.13170.40.44012.07
SIR-Diff (1 Frame)81.580.2476.45109.60.3509.55137.90.35710.36
SIR-Diff40.090.1605.7567.310.2639.4385.670.2759.30
", + "bbox": [ + 179, + 148, + 818, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we train with a handful of images per set.", + "bbox": [ + 89, + 272, + 364, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, attention communication between views leads to heavy computes. To resolve this, we only inject this module into the low-resolution layers (similar to CAT3D [13]), where we also ignore the biggest resolution layer for the UNet Encoder and Decoder. Further, we completely take away the cross-attention module across all layers instead of putting empty string inputs as placeholders like previous works [20]. We found that both of these changes help faster convergence, less training time, and better numerical results under the same number of training iterations.", + "bbox": [ + 88, + 287, + 485, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 450, + 215, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SIR-Diff on the task of sparse view motion deblurring and super-resolution in a zero-shot setting $\\S 4.2$ , and discuss our deblurring results in $\\S 4.4$ and super-resolution results in $\\S 4.5$ . We discuss implementation details in $\\S 4.2$ , and now we start by discussing evaluation metrics in $\\S 4.1$ . Please refer to supplementary materials for the ablation study.", + "bbox": [ + 89, + 476, + 485, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Metrics", + "text_level": 1, + "bbox": [ + 89, + 575, + 187, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use several metrics to evaluate the performance of all the methods on the task of motion deblurring and superresolution. Since our approach is generative in nature we borrow metrics from the literature that are good at evaluating generative models such as Frechet Inception Distance (FID) [15], Learned Perceptual Image Patch Similarity (LPIPS) [66]. In addition to these, we evaluate our superresolution results on PSNR and SSIM similar to methods used in this literature [57, 62, 64, 68]. We also propose a new metric to evaluate view consistency between pairs of generated views.", + "bbox": [ + 89, + 598, + 485, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual Consistency (VConsis). This metric evaluates self-consistency within the generated view set by measuring the perceptual similarity of corresponding patches between the two views. Please refer to the supplementary for additional details.", + "bbox": [ + 89, + 763, + 483, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 848, + 307, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our latent diffusion model based on pretrained UNet and auto-encoder from SDv2.1 as part of dif", + "bbox": [ + 89, + 869, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fusers. We train our models on two synthetic datasets Hypersim [34] and TartanAir [52] with image resolution resized to $640 \\times 480$ . We randomly select the order of image sets to maintain permutation invariance in training and inference. We utilize ground truth geometry and pose data to pre-compute the overlap region between each pair of frames. We provide additional details on model training, dataset, and view selection in the supplemental.", + "bbox": [ + 511, + 272, + 908, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Zero Shots Datasets For Evaluation", + "text_level": 1, + "bbox": [ + 511, + 404, + 821, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As all our baselines are training on different datasets, we perform a zero-shot evaluation across a diverse set of real-world datasets to keep a fair comparison. Zero-Shot means rather than split the dataset into training-evaluation splits, we chose some dataset for training and some other out-of-domain for evaluation. Our evaluation datasets consist of indoor and outdoor scenes along with object-centric datasets. ScanNet++ [61]. This is a dataset of 460 indoor scanned readl 3D scenes. We evaluate all the methods on the scenes from the official test set of 50 scenes, by sampling 200 image sets with four images per set. We highlight the zero-shot performance of our method trained on synthetic indoor data transfers well to real indoor scene data.", + "bbox": [ + 511, + 426, + 908, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ETH3D [38]. This dataset provides high-resolution images of indoor and outdoor scenes, with accurate depth maps captured via a LiDAR. We evaluate all the 454 samples from the ETH3D dataset.", + "bbox": [ + 511, + 623, + 906, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CO3Dv2 [32]. This is an object-centric dataset containing object-level videos across 51 categories of household objects. This dataset has pose annotations from COLMAP [37] provided for all video sequences. We use the official single-sequence subsets for evaluation and randomly sample 200 image sets among them.", + "bbox": [ + 511, + 685, + 908, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For all these three datasets, please refer to the supplementary material for the detail of view-selection strategy", + "bbox": [ + 511, + 776, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Motion Deblurring (MD)", + "text_level": 1, + "bbox": [ + 511, + 816, + 743, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our motion deblurring (MD) method targets solving arbitrary motion paths in the wild. We target at restoring original images from the set of sparse view motion blur inputs. Our paired training dataset is prepared by applying the motion-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21624", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9932788acb7ca3325ade09f3cf484f309a89d166055a3d4339cc3ad4c6b9230c.jpg", + "table_caption": [ + "Table 2. Sparse-View Super-Resolution Comparisons. VConsis. (visual consistency) is multiplied by 100. SIR-Diff (1 Frame) is a single image process form like other image diffusion models. Ours, in the inference stage, can adapt an arbitrary number of views. The best is highlighted and the second is underlined." + ], + "table_footnote": [], + "table_body": "
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodPSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓
SAM-DiffSR [48]24.130.2160.7586.9419.160.3170.4958.2823.900.2460.62610.96
OSEDiff [57]25.220.8080.1545.5121.930.6210.2648.6224.500.6580.2778.86
UAV [68]23.230.6770.3549.0120.390.5060.4139.2622.730.5660.40011.82
SIR-Diff (1 Frame)25.850.7920.1645.7021.970.5930.2217.9024.350.6410.3019.07
SIR-Diff27.380.8370.0995.0122.580.6210.2017.6024.530.6580.2107.85
", + "bbox": [ + 102, + 137, + 890, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "blurring kernel on the clean images from the training datasets based on the codebase inspired by DPS [7]. The blur kernel size is chosen from a normal distribution with mean 85px and std 12.75px.", + "bbox": [ + 88, + 253, + 480, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further, the intensity of the blur is randomly sampled in the range $[0, 1]$ . The blur kernel for each image in the set is sampled independently. For evaluation, we randomly sample 4 images per view-set from the pre-computed image set and require all the methods to remove motion blurring from them.", + "bbox": [ + 89, + 315, + 482, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1. Baselines", + "text_level": 1, + "bbox": [ + 89, + 415, + 202, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare against several baselines that can restore motion blurring from single-image or videos.", + "bbox": [ + 89, + 434, + 482, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Restormer [64]. This is a single-image motion deblurring method that does not have the benefit of jointly deblurring across a sparse view image collection. Our method outperforms this on single image deblurring and sparse view deblurring across all metrics.", + "bbox": [ + 89, + 465, + 483, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PromptIR [31]. This is all-in-one restoration method trained to deblur. We use their pre-trained model and compare the performance of their method on sparse view deblurring.", + "bbox": [ + 89, + 541, + 483, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VRT [25]. This is a video-oriented restoration Transformer model trained with deblurring capability. Our method outperforms this on sparse view deblurring across all metrics. Fair comparisons on video datasets are in the supplemental.", + "bbox": [ + 89, + 602, + 483, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 672, + 187, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets.", + "bbox": [ + 89, + 691, + 482, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative. As shown in the left side of Fig. 3, our SIR-Diff show rich details and sharp deblurring restoration quality than other methods on both object level [32] and scene level [61]. This is thanks to our architectural design information from different views can be aggregated and constructive to help the other view in attention and residual layers.", + "bbox": [ + 89, + 723, + 483, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative. We assess our results using the metrics described in § 4.2, with outcomes summarized in Tab. 1.", + "bbox": [ + 89, + 814, + 483, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method outperforms all the methods on the three datasets on FID and LPIPS metrics. Restormer [64], a single-image", + "bbox": [ + 89, + 845, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method, lacks the ability to leverage multi-view information to enhance output quality. In contrast, our method uses 3D self-attention to achieve superior results.", + "bbox": [ + 511, + 253, + 903, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also outperform the video motion deblurring method, VRT, as they are incapable of handling sparse views well by motion estimation and require continuous and smooth video-captured views to fuse frame features. In the supplement, we provide a direct comparison with video methods quantitatively on continuous videos.", + "bbox": [ + 511, + 299, + 906, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Super-Resolution (SR)", + "text_level": 1, + "bbox": [ + 511, + 398, + 723, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on the 4x sparse-view super-resolution task. Our paired training dataset is prepared by first downsampling the high-resolution image with a scale factor of 4 and then upsampling it back to the original resolution to introduce a low-resolution effect. The resize algorithm we use is bicubic.", + "bbox": [ + 511, + 422, + 906, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5.1. Baselines", + "text_level": 1, + "bbox": [ + 511, + 521, + 624, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare against several diffusion model baselines that handle single-image and video super-resolution.", + "bbox": [ + 511, + 540, + 905, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAM-DIFFSR [48]. This is an image diffusion model-based methodology, which targets classical super-resolution tasks like ours. Our method outperforms this on sparse view restoration on all the metrics.", + "bbox": [ + 511, + 570, + 905, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OSEDiff [57]. This is a single-image diffusion model-based super-resolution model. OSEDiff reaches SOTA performance on the 2D image super-resolution.", + "bbox": [ + 511, + 631, + 905, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Upscale-A-Video [68]. This is a video diffusion model-based super-resolution method. We input it with a sequential video frame (in total 4 for each scenario) but a larger motion gap compared to ordinary videos. We also provide a fair comparison between Upscale-A-Video and our method on video datasets in the supplement.", + "bbox": [ + 511, + 676, + 906, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5.2. Results", + "text_level": 1, + "bbox": [ + 511, + 775, + 609, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets.", + "bbox": [ + 511, + 794, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative. As the red rectangular box of Fig. 3 shows,", + "bbox": [ + 511, + 825, + 883, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for shadow artifacts and severe aliasing effects (see the second example), Upscale-A-Video [68] fails and OSEDiff [57] restoration is unfaithful and inconsistent for different views compared to ours. Moreover, our model can restore", + "bbox": [ + 511, + 840, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/LeviBorodenko/motionblur", + "bbox": [ + 107, + 886, + 359, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21625", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/747ad6d8fd3bd9d99840a0d5d22e8f03340e8866fecc2ce43c679ac15942a5df.jpg", + "table_caption": [ + "Table 3. Gaussian Splitting Reconstruction from Blurry Images. Under serious motion blurring, COLMAP[37] makes it hard to solve correct camera pose and SfM points, which leads to bad results on both vanilla 3DGS and BAD-GS. With the help of SIR-Diff, we can achieve restore better results and be robust to difficult motion blurring effects. KS refers to kernel size with normal distribution $[\\mu, \\sigma]$ . Inten. refers to the blurring intensity range that is sampled uniformly. We randomly select 3 objects in CO3Dv2 [32] dataset and apply motion blur on clean images to conduct this experiment." + ], + "table_footnote": [], + "table_body": "
DifficultyMediumHard
ParamsKS:[30, 10.2], Inten:[0, 0.4]KS:[45,14.85], Inten:[0, 0.5]
MethodPSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [21]13.470.6100.60710.420.4210.801
BAD-GS [67]10.060.5370.6878.0610.3850.948
Resotmer [64]+BAD-GS [67]25.770.6160.35623.880.5600.317
SIR-Diff + BAD-NeRF [49]23.600.5030.45222.470.4390.485
SIR-Diff + BAD-GS26.110.6610.25025.330.6440.277
", + "bbox": [ + 99, + 229, + 480, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images with better geometry alignment as shown in the first example.", + "bbox": [ + 89, + 342, + 482, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative. We use PSNR, SSIM, LPIPS [66] and our proposed visual consistency § 4.1 metrics. The result is shown in Tab. 2 to show that SIR-Diff outperforms all the existing baselines on all datasets. Compared to the image diffusion method, OSEDiff[57], our spatial-3D Resnet and 3D self-attention Transformer help each latent token to access information from all views which leads to successful multiview restoration. Upscale-A-Video also employs a similar 3D ResNet architecture and cross-image attention mechanism; however, our model does not rely on conditioning with optical flow, which is challenging to obtain accurately in sparse views.", + "bbox": [ + 88, + 372, + 483, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Down Stream Applications", + "text_level": 1, + "bbox": [ + 89, + 564, + 333, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show several downstream applications of sparse multiview deblurring and super-resolution. Multi-view sparse images are used for tasks of finding correspondences, pose estimation, and 3D reconstruction. The performance of these methods is often dependent on the input image quality. If the underlying image quality is degraded, the output of these methods is poor. Blindly restoring these images with off-the-shelf single-image methods leads to sub-optimal results because restoration from these methods lacks view consistency. We show that, with SIR-Diff, we can consistently restore image quality and help alleviate the performance gap on tasks requiring multiview sparse views.", + "bbox": [ + 89, + 585, + 485, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.1. Estimating Correspondences", + "text_level": 1, + "bbox": [ + 89, + 776, + 336, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Under severely degraded inputs it is hard for methods like LoFTR [43] to estimate correspondences. As shown in Fig. 4, (a) we show qualitatively at low-resolution LoFTR struggles to find correct correspondences. Moreover, since single-view image restoration models lack multi-view consistency this does not solve the issue. SIR-Diff can not only restore the image quality but since it is view consistent can improve the", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4374ad5459e1d05392054586fb5127e0b1443e5561f955c801586ab2a1202b4a.jpg", + "table_caption": [ + "Table 4. Feature Matching Correspondence. We compute the correspondences using off-the-shelf LoFTR [43] on the restored images with single-view methods and SIR-Diff. Both in SR and deblurring cases, our method can enable LoFTR to find more correspondences (# of correspondences found)." + ], + "table_footnote": [], + "table_body": "
DeblurringSR
GT Images1979.51979.5
Degraded Images640.11374.0
Single-View Methods [57, 64]1262.61478.0
SIR-Diff1621.31494.2
", + "bbox": [ + 575, + 162, + 841, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "correspondences for off-the-shelf methods like LoFTR [43]. Quantitatively, as shown in Tab. 4, we also see that LoFTR can find over 380 additional correspondences as compared to the best image-based method on deblurring and on average 15 for super-resolution.", + "bbox": [ + 511, + 239, + 906, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.2. 3D Reconstruction on Degraded Images", + "text_level": 1, + "bbox": [ + 511, + 324, + 836, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods such as BAD-Gaussians [67] (BAD-GS) aim to perform 3D reconstruction from motion-blurry images, but they depend on running COLMAP [37] on blurry images to get pose initialization and sparse 3D geometry. We show in Tab. 3 that with increasing blurring effect, BAD-GS performance drops significantly while using SIR-Diff with BAD-GS leads to a much steady performance. In Fig. 5, we show a hard case of motion blurring that leads to failure of BAD-GS [67] which can be solved if we restore the image using SIR-Diff first and then do BAD-GS. We refer to supplementary for details.", + "bbox": [ + 511, + 342, + 908, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.3. Sparse-View 3D reconstruction", + "text_level": 1, + "bbox": [ + 511, + 517, + 774, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "InstantSplat [11] represents a leading approach for sparse view Gaussian Splittings (GS). However, reconstructing a scene under sparse viewpoints with low-quality observations remains challenging. In Fig 6, we show an example of applying InstantSplat on an image restored using one of the baselines compared to restoration using SIR-Diff. In Tab. 5, we report the average performance of all baselines on the selected 3 scenes from Scannet++[61] and CO3Dv2[32] and build a 3DGS using InstantSplat [11]. We evaluate the GS with standard Novel View Synthesis metrics. Additional details can be found in the supplementary.", + "bbox": [ + 511, + 536, + 908, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Depth Prediction for Sparse Views", + "text_level": 1, + "bbox": [ + 511, + 712, + 815, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also apply our model architecture to the task of improving sparse multiview depth estimation. The only thing we edit is changing the target latent to depth image sets and the conditional latent to RGB image sets. Following Marigold [20], we use the ground truth depth label provided by the dataset for supervision signal during training.", + "bbox": [ + 511, + 734, + 906, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We evaluate the performance of methods on three metrics AbsRel and $\\delta_{1}$ as defined by Silberman et al. [42]. In the case of MDE methods, we additionally align all the predicted depth maps to find the most optimum scale and bias to make them consistent. Similar to the view consistency", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21626", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5124e623cf3a17a2f23c804cbf7ad335319e89e482c8254fec506115b8b6760.jpg", + "image_caption": [ + "Figure 4. Feature Matching Downstream Application. Low-Resolution and Single-image-based SR models like OSEDiff [57] find less correspondence points than our SIR-Diff that can restore multiview images consistently. SIR-Diff recovers the highest number of correspondences (310) as compared to other baseline restoration methods." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 903, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f68868daecdb2fc2a1a8b22c84d9bbdee8875d10d5644a83587142eb6b8bd40.jpg", + "image_caption": [ + "Figure 5. BAD-GS [67] with SIR-Diff. We show the effect of using SIR-Diff to help BAD-GS recover from catastrophic failure. BAD-GS (Top) output with a very high LPIPS $\\downarrow$ of 0.59 vs. 0.22 BAD-GS output with image-restoration using SIR-Diff (Bottom) leads to a significantly higher-quality 3D reconstruction." + ], + "image_footnote": [], + "bbox": [ + 94, + 277, + 478, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83122bcf17bc8a428d4312fbb1b1dd32fe8d47c46a6530116a18c93becc64e1e.jpg", + "image_caption": [ + "Figure 6. InstantSplat Downstream Application. 3D Reconstruction is done by InstantSplat [11]. SIR-Diff that can restore multiview images consistently leads to better visual and numerical reconstruction results than other single-image methods, Restormer [64]." + ], + "image_footnote": [], + "bbox": [ + 94, + 546, + 480, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "metrics for image restoration, We evaluate Geometric Consistency (GConsis.) by calculating the distances between corresponding points between two different views under the camera coordinate system. Additional details on this metric are in the supplemental.", + "bbox": [ + 89, + 794, + 483, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We evaluate against several monocular depth estimation (MDE) methods such as DepthAnythingV2 [60],", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d87511192a8a777eec72f2075a1e926a99329fabdd71d78c5b30218104eeea4e.jpg", + "table_caption": [ + "Table 5. Quantitative Results of Super-Resolution and Motion-Deblurring Reconstruction on InstantSplat [11]. The result is an average of 3 scenes per dataset (individual results in supp). Using SIR-Diff results in higher quality inputs to InstantSplat compared to OSEDiff [57] and Restormer [64]." + ], + "table_footnote": [], + "table_body": "
Scannnet++ [61]CO3D [32]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GT Image31.860.9400.09623.940.7570.243
Super-Resolution
OSEDiff [57]24.530.7990.27420.250.5490.379
SIR-Diff26.200.8270.23321.780.5780.357
Motion-Deblurring
Restormer [64]22.020.7890.30920.140.5790.465
SIR-Diff26.530.8420.23121.600.6140.384
", + "bbox": [ + 521, + 348, + 901, + 474 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29e83495825a31fe5ca461774523a8148af240a0c060f3e0e4f376a148371548.jpg", + "table_caption": [ + "Table 6. Quantitative comparisons on image and video depth models. The best is highlighted and the second is underlined. GConsis. refers to geometric consistency and the unit is meter." + ], + "table_footnote": [], + "table_body": "
DatasetScannet++ [61]ETH3D [38]
MethodAbsRel ↓δ1 ↑GConsis.↓AbsRel ↓δ1 ↑GConsis.↓
DepthAnyV2 [60]9.27093.350.17713.2086.521.932
DepthCrafter [18]21.6589.960.7516.66295.051.701
Marigold [20]11.1089.920.1627.58194.460.410
SIR-Diff9.01193.540.1296.66395.050.339
", + "bbox": [ + 521, + 531, + 901, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DepthCrafter [18], and Marigold [20].", + "bbox": [ + 511, + 619, + 767, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. As shown in Tab. 6, our multi-view depth model is competitive with single image depth methods on the AbsRel, and $\\delta_{1}$ metrics are sometimes better. Meanwhile, our model outperforms these methods on the geometric consistency metric that evaluates the 3D consistency when predicting depth jointly from sparse views.", + "bbox": [ + 511, + 633, + 906, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion & Limitations", + "text_level": 1, + "bbox": [ + 511, + 738, + 754, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SIR-Diff, a simple and effective approach for multi-view image restoration, addressing blurring and low-resolution degradation across multiple images simultaneously. SIR-Diff outperforms leading single-view methods in sparse-view restoration and enables 3D-consistent image restoration for 3D computer vision applications. Future work will explore efficient cross-image attention mechanisms, such as selective patch attention along epipolar lines, and develop explicit guarantees for 3D view consistency.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21627", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work partially used NCSA Delta GPU at the University of Illinois through allocation CIS230313 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296.", + "bbox": [ + 89, + 114, + 485, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 222, + 187, + 238 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Miika Aittala and Frédo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 2", + "[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 2", + "[3] Sherwin Bahmani, Xian Liu, Wang Yifan, Ivan Skorokhodov, Victor Rong, Ziwei Liu, Xihui Liu, Jeong Joon Park, Sergey Tulyakov, Gordon Wetzstein, et al. Tc4d: Trajectory-conditioned text-to-4d generation. In European Conference on Computer Vision, pages 53-72. Springer, 2025. 2", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4217-4229, 2023. 2", + "[7] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 6", + "[8] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaou Tang. Compression artifacts reduction by a deep convolutional network. In Proceedings of the IEEE international conference on computer vision, pages 576-584, 2015. 1, 2", + "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2" + ], + "bbox": [ + 99, + 247, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang. Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine, and Art, 2(1):7, 2019. 2", + "[11] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 7, 8", + "[12] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jianping Fan, Yu Jun, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024. 2", + "[13] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 4, 5", + "[14] Shi Guo, Xi Yang, Jianqi Ma, Gaofeng Ren, and Lei Zhang. A differentiable two-stage alignment scheme for burst image reconstruction with large shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17472-17481, 2022. 2", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[17] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2", + "[18] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 8", + "[19] Michal Kawulok, Pawel Benecki, Szymon Piechaczek, Krzysztof Hrynczenko, Daniel Kostrzewa, and Jakub Nalepa. Deep learning for multiple-image super-resolution. IEEE Geoscience and Remote Sensing Letters, 17(6):1062-1066, 2019. 2", + "[20] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 5, 7, 8", + "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 7", + "[22] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21628", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2", + "[24] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2", + "[25] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 2, 3, 5, 6", + "[26] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 2", + "[27] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2", + "[28] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Taming diffusion models for image restoration: A review. arXiv preprint arXiv:2409.10353, 2024. 2", + "[29] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1", + "[30] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2", + "[31] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 6", + "[32] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2, 5, 6, 7, 8", + "[33] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 2", + "[34] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2", + "[37] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5, 7", + "[38] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2, 5, 6, 8", + "[39] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024. 2", + "[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2", + "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 4", + "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7", + "[43] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7", + "[44] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2", + "[45] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21629", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Boyang Wang, Bowen Liu, Shiyu Liu, and Fengyu Yang. Vciser: Blind single image super-resolution with video compression synthetic data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4302-4312, 2024. 2", + "[47] Boyang Wang, Fengyu Yang, Xihang Yu, Chao Zhang, and Hanbin Zhao. Apisr: anime production inspired real-world anime super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25574-25584, 2024. 2", + "[48] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image super-resolution. arXiv preprint arXiv:2402.17133, 2024. 6", + "[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 7", + "[50] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 2", + "[51] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2", + "[52] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5", + "[53] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2", + "[54] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2", + "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[56] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2", + "[57] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 5, 6, 7, 8", + "[58] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 134:109046, 2023. 2" + ], + "bbox": [ + 91, + 92, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Cheng Yang, Lijing Liang, and Zhixun Su. Real-world denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023.2", + "[60] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8", + "[61] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 5, 6, 7, 8", + "[62] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 2, 5", + "[63] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2", + "[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5, 6, 7, 8", + "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4", + "[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7", + "[67] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 2, 7, 8", + "[68] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 2, 3, 5, 6" + ], + "bbox": [ + 516, + 92, + 906, + 712 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21630", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_model.json b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4afc534fa851fb8e14b84d5c017b3c5ce8909641 --- /dev/null +++ b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.131, + 0.882, + 0.152 + ], + "angle": 0, + "content": "SIR-DIFF: Sparse Image Sets Restoration with Multi-View Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.179, + 0.803, + 0.235 + ], + "angle": 0, + "content": "Yucheng Mao\\(^{1, *}\\) Boyang Wang\\(^{1, *}\\) Nilesh Kulkarni\\(^{1,2}\\) Jeong Joon Park\\(^{1}\\) \n\\(^{1}\\)University of Michigan, Ann Arbor \\(^{2}\\)Netflix" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.254, + 0.868, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.596, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Sparse View Image Restoration. Our diffusion model takes multi-view images and jointly enhances their visual quality while maintaining 3D consistency. (Top) Four motion-blurred input images are processed by our method, resulting in sharp outputs that significantly outperform single-view restoration methods as shown in the corresponding purple boxes. (Bottom) Our method can consistently restore multi-view images (4 out of 50 shown), leading to accurate 3D reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.665, + 0.327, + 0.68 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.486, + 0.879 + ], + "angle": 0, + "content": "The computer vision community has developed numerous techniques for digitally restoring true scene information from single-view degraded photographs, an important yet extremely ill-posed task. In this work, we tackle image restoration from a different perspective by jointly denoising multiple photographs of the same scene. Our core hypothesis is that degraded images capturing a shared scene contain complementary information that, when combined, better constrains the restoration problem. To this end, we implement a powerful multi-view diffusion model that jointly generates uncorrupted views by extracting rich information from multi-view relationships. Our experiments show that our multi-view approach outperforms existing single-view im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.909, + 0.757 + ], + "angle": 0, + "content": "age and even video-based methods on image deblurring and super-resolution tasks. Critically, our model is trained to output 3D consistent images, making it a promising tool for applications requiring robust multi-view integration, such as 3D reconstruction or pose estimation. Project website: https://myc634.github.io/sirdiff/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.775, + 0.646, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Image restoration approaches [8, 16, 29, 53] typically process degraded images in isolation, confronting the challenge of the original scene content. Yet in practice, we frequently capture multiple photographs of meaningful moments, yielding a natural collection of views of the same scene, albeit some degraded. This common behavior presents an untapped opportunity: leveraging multiple captures to enhance image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21620" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.411, + 0.108 + ], + "angle": 0, + "content": "quality and better preserve our visual memories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.312 + ], + "angle": 0, + "content": "In this work, we present a novel perspective on image restoration by reformulating it as a multi-view collaborative restoration task. We leverage our key insight that multiple corrupted photographs of the same scene contain complementary information that can aid superior denoising outcomes. The concept of multi-image fusion for quality enhancement has been explored in classical super-resolution techniques where sub-pixel shifted images are combined to reconstruct fractional pixel information [19]. Our approach revisits this paradigm in a fundamentally wider context by using neural attention mechanisms to implicitly fuse overlapping information from general multi-view images to enhance even severe aberrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.487, + 0.49 + ], + "angle": 0, + "content": "Notably, our multi-view restoration framework has particular significance for 3D computer vision applications. Many real-world tasks, from robotic SLAM to novel-view synthesis, rely fundamentally on multi-view imagery of a scene. When restoring degraded images within such multi-view sets, maintaining geometric consistency across views is crucial to preserve the underlying 3D scene assumptions. Traditional single-view restoration methods, inevitably introduce inconsistencies that can compromise these geometric constraints. In contrast, our restoration approach naturally enforces view consistency through learned neural priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.487, + 0.681 + ], + "angle": 0, + "content": "We implement our approach, dubbed SIR-Diff (Sparse Image Restoration with Diffusion), through an innovative multi-view diffusion model architecture. Our choice of diffusion models is motivated by their natural handling of uncertainty in image restoration. We extend the UNet-based Stable Diffusion [35] architecture to process multiple views jointly by introducing three key components: (1) a unified encoder for diverse degraded inputs including low-resolution, blurred, and regular RGB images, (2) a hybrid 2D-3D convolutional structure derived from SVD (Stable Video Diffusion) [4, 5], and (3) a 3D self-attention Transformer that enables cross-view attention across all spatial and multi-view tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.487, + 0.902 + ], + "angle": 0, + "content": "We evaluate and show zero-shot performance on Scannet++ [61], ETH3D [38], and CO3D [32] for multi-view deblurring and super-resolution tasks and their 3D downstream applications. SIR-Diff consistently outperforms existing single-image and video restoration methods for multi-view restoration tasks in both image quality and our proposed 3D view consistency metrics. For downstream applications, we show that our multi-view restoration strategy significantly enhances 3D reconstruction quality, as evidenced by improved novel-view synthesis accuracy and feature-matching success rates. Beyond image restoration, our architecture shows promise in other multi-view tasks, notably achieving competitive depth estimation accuracy and geometric consistency compared to leading single-view methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.665, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.908, + 0.237 + ], + "angle": 0, + "content": "Image Diffusion Models. Diffusion Models (DM) show huge popularity in all domains: illustrations [9, 35, 51, 65], video creation [4, 5, 30], 3D modeling with limited input views [27, 41, 45], 4D generation [2, 3, 26], and etc. Most models are fine-tuned from the pre-trained weight instead of training from the beginning. One of the most widely-used pre-trained Diffusion Models is Stable Diffusion 2.1 [35], a UNet-based [36] architecture." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.238, + 0.909, + 0.403 + ], + "angle": 0, + "content": "Multiview-Image Diffusion models. In the image-to-3D task [24], Diffusion Model is a powerful generative prior to enriching details for the novel (unseen) view synthesis [6, 17, 44, 55, 56]. Methods like MVDream [41], CAT3D [13], Zero \\(123++\\) [40] jointly generate multi-views simultaneously by extending the image diffusion model (SD2.1) with a 3D self-attention layer. However, these works ignore the exploration of the convolution layer architecture in the UNet-based Diffusion Model. In this paper, we explore the methodology of applying the 3D self-attention layer to the 3D restoration task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.405, + 0.909, + 0.571 + ], + "angle": 0, + "content": "Image and Video Restoration. These methods aim at converting a degraded input image or video into a high-quality form. Traditional tasks include Super-Resolution [23, 53, 57, 62, 68], Deblurring [22, 25, 31], Denoise [10, 59], Compression Artifact Reduction [8, 46, 47], Inpainting [7, 58], Burst Image Deblurring [1, 14], and etc. These domains have achieved milestones in the 2D image and video domain [28]. However, the study on the restoration in 3D is insufficient. The restoration consistency of the texture and geometry from different views is hard to maintain when independently restoring them." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.911, + 0.785 + ], + "angle": 0, + "content": "3D Reconstruction from Corrupted Inputs. 3D Gaussian Splatting (3DGS) [21, 33, 54, 63] transforms input images into explicit-represented 3D Gaussian blobs with continuous optimization to reconstruct the large 3D scene. However, these models heavily rely on the condition of the input images. Motion blurring [50, 67] or low-resolution [12, 39] inputs that lack high-frequency and sharp details may greatly deteriorate the reconstruction quality or even fail at the initialization stage. To resolve this, methods like Bad-Gaussian [39] and SuperGaussian [67] depend on dense view input to iteratively optimize the degraded artifacts. Further, under heavily deteriorated inputs, these methods cannot recover high-quality outputs when they lack a strong generative prior [39] such as a pre-trained Diffusion Model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.606, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.91, + 0.903 + ], + "angle": 0, + "content": "In this section, we discuss our proposed approach, SIR-Diff, which can generate and restore sparse image sets from 3D scenes or objects conditioned on their degraded inputs. Specifically, we showcase our approach to two challenging tasks: motion deblurring and super-resolution. In the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "21621" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.08, + 0.904, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Method Overview. Our model is a UNet-based latent diffusion model that takes as input a degraded sparse image set and outputs a set of restored consistent images. The core of our approach is a diffusion model that jointly denoises image latents across multiple views by relying on our spatial-3D ResNet and 3D Self-Attn transformer. We use pre-trained encoder-decoders from SD2.1 [35] and train our denoising UNet on those encoded latents. We show view-consistent results for the task of Deblurring and Super-Resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.442 + ], + "angle": 0, + "content": "following sections, we discuss our problem setup in § 3.1, multi-view image restoration diffusion model in § 3.2, and our proposed model architecture in § 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.452, + 0.355, + 0.468 + ], + "angle": 0, + "content": "3.1. Multi-View Generation Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Our goal is to restore sparse view images from degraded inputs. Typical methods to solve this problem work either on a single image [31, 35, 57] or require a coherent video [25, 68] as input. Toward the goal of sparse view image restoration, we propose a generative model that can respect the geometrical priors to perform the task of image restoration from blurry images or recovering high-resolution image data from its low-resolution counterpart. Any single-image generative model that performs independent inference on each image cannot ensure consistency over the generated image set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.702 + ], + "angle": 0, + "content": "To alleviate these issues, we formulate this as a generation problem to infer all the required views of the scene by modeling the joint distribution using a multi-view diffusion model. Our model takes a degraded image set as input and recovers a high-quality self-consistent image set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.483, + 0.728 + ], + "angle": 0, + "content": "3.2. Multi-View Image Restoration Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.885 + ], + "angle": 0, + "content": "Our model aims to restore images when conditioned on a corrupted image set \\(\\{I^c\\}_{i = 1}^N\\). The model's primary objective is to restore a self-consistent high-quality image set. We do this by using a multi-view image restoration diffusion model that can jointly restore the image set. In training, we start from a dataset of non-blurry / high-resolution images \\(\\{I^r\\}_{i = 1}^N\\) and apply the forward diffusion process. Our input representation to the N-view diffusion model is \\(\\mathbf{x} = \\{I_1^r,\\dots,I_N^r\\}\\). Specifically, \\(\\mathbf{x}^0\\) corresponds to a clean signal at the start of the forward diffusion process at step 0." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Our generative model learns to reverse this noising pro" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.396, + 0.907, + 0.427 + ], + "angle": 0, + "content": "cess by predicting the added noise at every timestep of the diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.702, + 0.453 + ], + "angle": 0, + "content": "3.3. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.909, + 0.581 + ], + "angle": 0, + "content": "We build our joint image restoration diffusion model based on a latent diffusion model [35]. Fig. 2 shows an overview of our method. The goal of our diffusion model is to denoise image latents such that a pre-trained decoder can convert them to high-quality images. When simultaneously denoising multiple image latents, it is important that latent representation remain consistent across views. We enable this by allowing the diffusion model to attend to latents from other views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.84, + 0.606 + ], + "angle": 0, + "content": "3.3.1. Latent Diffusion Models for Restoration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.656 + ], + "angle": 0, + "content": "A Latent Diffusion Model [35] architecture, such as Stable Diffusion 2.1 (SD2.1), consists of two parts a) encoder and decoder and b) denoising UNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.732 + ], + "angle": 0, + "content": "During inference, our latent denoising UNet progressively refines pure noise into a clean latent representation, conditioned on the degraded image set, through several denoising steps. Once the latent representation is clean, a decoder translates it into the restored image set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.908, + 0.779 + ], + "angle": 0, + "content": "Similarly, during training, our diffusion model is optimized to learn the denoising process of the latent representation, conditioned on the degraded input sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Encoder-Decoder. We initialize our encoder with the weights from an SD2.1 [35] and are frozen when we train our model. The encoder, \\(\\phi\\), is used to map our images to latent vectors which are used to learn the UNet-Style diffusion model. The decoder, \\(\\delta\\), converts the denoised latents from the diffusion model to pixels in image space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "UNet. Our latent denoising UNet, \\( \\mathbf{s}_{\\theta}(\\cdot) \\) with parameters \\( \\theta \\), considers as input a noisy latent \\( \\mathbf{x}^k \\), encoded latents for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.085, + 0.905, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.362, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 3. Qualitative Comparisons on Motion Deblurring and Super-Resolution. The left side is the Motion-Deblurring and the right side is the Super-Resolution results. We refer readers to supplementary for more visualizations. Zoom in for the best view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.482, + 0.437 + ], + "angle": 0, + "content": "degraded images, \\(\\{I^c\\}_{i = 1}^N\\) and the diffusion noise \\(k\\) to output an of estimate the noise, \\(\\hat{\\pmb{\\epsilon}}\\), as shown in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.445, + 0.483, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} = \\mathbf {s} _ {\\theta} (\\mathbf {x} ^ {k}, \\phi (\\{I ^ {c} \\} _ {i = 1} ^ {N}), k). \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.504 + ], + "angle": 0, + "content": "We supervise the outputs of our UNet to match the noise added during the forward diffusion process using Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.511, + 0.483, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon \\sim \\mathcal {N} (0, I), k \\sim \\mathcal {U} (T)} \\| \\epsilon - \\hat {\\epsilon} \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.54, + 0.442, + 0.555 + ], + "angle": 0, + "content": "3.3.2. Adapting Diffusion Model to multiple views" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.558, + 0.483, + 0.71 + ], + "angle": 0, + "content": "Starting with the image diffusion model, we expand a new dimension to represent image sets from different views. Most of the time, batch and views are merged into one dimension and split when needed. We encode the degraded image set with the same VAE encoder as the target noisy latents. Noisy latents and the condition latents are integrated by channel-wise concatenation. Then, due to channel number change, we replace the original convolution with a new zero-initialized convolution layer [65] to adapt the new number of channels input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.422, + 0.731 + ], + "angle": 0, + "content": "3.3.3. Spatial-3D Integrated UNet Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Typical latent diffusion models such as Stable Diffusion [35] do not consider denoise multiple images simultaneously and often generate individual images that lead to inconsistent image generations between multi-views. Our solution to remove this inconsistency is to denoise all the views jointly by allowing the UNet to look at the intermediate latent representations from different views. As shown in the right side of Fig. 2, our architecture is composed of a Spatial-3D ResNet and a 3D self-attention Transformer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Spatial-3D ResNet. Image diffusion model like SD2.1 only has a 2D convolution layer in the ResNet, which makes it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.908, + 0.694 + ], + "angle": 0, + "content": "hard to perform well in tasks that require 3D understanding. We want to enhance the 3D performance of the ResNet of the image diffusion model by introducing a 3D convolution layer to better capture geometry information as shown in Fig. 2. Training from raw will be unstable and time-consuming, but we find that the video diffusion model that is performing the image-to-video task already has a pre-trained 3D convolution layer. Understanding temporal similarity in the time dimension and understanding 3D similarity in the spatial dimension share similarity. By the experiment, we empirically find that merging two different weights into one is feasible and can lead to better performance. Hence, we initialize the 2D Convolution layer with image Diffusion Model weight (SD2.1) and the 3D convolution layer with video Diffusion weight (Stable Video Diffusion [4]). The output of each layer will be blended as follows: \\( O_{\\mathrm{ResNet}} = \\sigma (\\alpha)\\times O_{2D} + \\sigma (1 - \\alpha)\\times O_{3D} \\), where \\( \\sigma \\) is the sigmoid function and \\( \\alpha \\) is usually taken as 0.5 following SVD [4]. \\( O_{2D} \\) is the output of 2D ResNet and \\( O_{3D} \\) is the output of 3D ResNet layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D Self-Attention Transformer. Furthermore, we adopt the 3D multi-view consistent self-attention Transformer as shown in Fig. 2, similar to previous works [13, 41]. Our self-attention layer considers image latents from all the \\(N\\) views and performs self-attention. After ResNet, the model patches the latent to create \\(p\\) patch token per view, and in total, we have \\(N \\times p\\) tokens that are part of the attention layer. Each token calculates the attention score to all \\(N \\times p\\) tokens as shown on the right side of Fig. 2. This feature enables each token to be accessible to all conditional image sets given at the maximum level. Further, by this design, our model, in the inference stage, is scalable to a variable number of input frames with stable performance even though" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21623" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.085, + 0.908, + 0.141 + ], + "angle": 0, + "content": "Table 1. Sparse-View Deblurring Comparisons. We show quantitative results on three datasets and evaluate metrics as in § 4.1. VConsis. (visual consistency) is multiplied by 100. Across all methods, SIR-Diff is the most visually consistent as compared to all the baselines. On other standard metrics, we also outperform baselines. SIR-Diff (1 Frame) is a single image version of our method that does not depend on multiple views. The best is highlighted, and the second is underlined." + }, + { + "type": "table", + "bbox": [ + 0.18, + 0.15, + 0.819, + 0.254 + ], + "angle": 0, + "content": "
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodFID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓
PromptIR [31]81.280.2487.81150.340.53710.72166.30.35012.89
Restromer [64]49.720.2326.5273.470.3309.59144.80.42611.00
VRT [25]134.50.3717.67160.40.53011.13170.40.44012.07
SIR-Diff (1 Frame)81.580.2476.45109.60.3509.55137.90.35710.36
SIR-Diff40.090.1605.7567.310.2639.4385.670.2759.30
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.365, + 0.287 + ], + "angle": 0, + "content": "we train with a handful of images per set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.288, + 0.486, + 0.44 + ], + "angle": 0, + "content": "However, attention communication between views leads to heavy computes. To resolve this, we only inject this module into the low-resolution layers (similar to CAT3D [13]), where we also ignore the biggest resolution layer for the UNet Encoder and Decoder. Further, we completely take away the cross-attention module across all layers instead of putting empty string inputs as placeholders like previous works [20]. We found that both of these changes help faster convergence, less training time, and better numerical results under the same number of training iterations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.452, + 0.216, + 0.469 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.486, + 0.569 + ], + "angle": 0, + "content": "We evaluate SIR-Diff on the task of sparse view motion deblurring and super-resolution in a zero-shot setting \\(\\S 4.2\\), and discuss our deblurring results in \\(\\S 4.4\\) and super-resolution results in \\(\\S 4.5\\). We discuss implementation details in \\(\\S 4.2\\), and now we start by discussing evaluation metrics in \\(\\S 4.1\\). Please refer to supplementary materials for the ablation study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.188, + 0.592 + ], + "angle": 0, + "content": "4.1. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.764 + ], + "angle": 0, + "content": "We use several metrics to evaluate the performance of all the methods on the task of motion deblurring and superresolution. Since our approach is generative in nature we borrow metrics from the literature that are good at evaluating generative models such as Frechet Inception Distance (FID) [15], Learned Perceptual Image Patch Similarity (LPIPS) [66]. In addition to these, we evaluate our superresolution results on PSNR and SSIM similar to methods used in this literature [57, 62, 64, 68]. We also propose a new metric to evaluate view consistency between pairs of generated views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.839 + ], + "angle": 0, + "content": "Visual Consistency (VConsis). This metric evaluates self-consistency within the generated view set by measuring the perceptual similarity of corresponding patches between the two views. Please refer to the supplementary for additional details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.308, + 0.866 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We implement our latent diffusion model based on pretrained UNet and auto-encoder from SDv2.1 as part of dif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.909, + 0.395 + ], + "angle": 0, + "content": "fusers. We train our models on two synthetic datasets Hypersim [34] and TartanAir [52] with image resolution resized to \\(640 \\times 480\\). We randomly select the order of image sets to maintain permutation invariance in training and inference. We utilize ground truth geometry and pose data to pre-compute the overlap region between each pair of frames. We provide additional details on model training, dataset, and view selection in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.822, + 0.42 + ], + "angle": 0, + "content": "4.3. Zero Shots Datasets For Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.909, + 0.624 + ], + "angle": 0, + "content": "As all our baselines are training on different datasets, we perform a zero-shot evaluation across a diverse set of real-world datasets to keep a fair comparison. Zero-Shot means rather than split the dataset into training-evaluation splits, we chose some dataset for training and some other out-of-domain for evaluation. Our evaluation datasets consist of indoor and outdoor scenes along with object-centric datasets. ScanNet++ [61]. This is a dataset of 460 indoor scanned readl 3D scenes. We evaluate all the methods on the scenes from the official test set of 50 scenes, by sampling 200 image sets with four images per set. We highlight the zero-shot performance of our method trained on synthetic indoor data transfers well to real indoor scene data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.685 + ], + "angle": 0, + "content": "ETH3D [38]. This dataset provides high-resolution images of indoor and outdoor scenes, with accurate depth maps captured via a LiDAR. We evaluate all the 454 samples from the ETH3D dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.909, + 0.776 + ], + "angle": 0, + "content": "CO3Dv2 [32]. This is an object-centric dataset containing object-level videos across 51 categories of household objects. This dataset has pose annotations from COLMAP [37] provided for all video sequences. We use the official single-sequence subsets for evaluation and randomly sample 200 image sets among them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.908, + 0.809 + ], + "angle": 0, + "content": "For all these three datasets, please refer to the supplementary material for the detail of view-selection strategy" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.744, + 0.835 + ], + "angle": 0, + "content": "4.4. Motion Deblurring (MD)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our motion deblurring (MD) method targets solving arbitrary motion paths in the wild. We target at restoring original images from the set of sparse view motion blur inputs. Our paired training dataset is prepared by applying the motion-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21624" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.085, + 0.907, + 0.128 + ], + "angle": 0, + "content": "Table 2. Sparse-View Super-Resolution Comparisons. VConsis. (visual consistency) is multiplied by 100. SIR-Diff (1 Frame) is a single image process form like other image diffusion models. Ours, in the inference stage, can adapt an arbitrary number of views. The best is highlighted and the second is underlined." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.138, + 0.891, + 0.236 + ], + "angle": 0, + "content": "
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodPSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓
SAM-DiffSR [48]24.130.2160.7586.9419.160.3170.4958.2823.900.2460.62610.96
OSEDiff [57]25.220.8080.1545.5121.930.6210.2648.6224.500.6580.2778.86
UAV [68]23.230.6770.3549.0120.390.5060.4139.2622.730.5660.40011.82
SIR-Diff (1 Frame)25.850.7920.1645.7021.970.5930.2217.9024.350.6410.3019.07
SIR-Diff27.380.8370.0995.0122.580.6210.2017.6024.530.6580.2107.85
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.255, + 0.482, + 0.315 + ], + "angle": 0, + "content": "blurring kernel on the clean images from the training datasets based on the codebase inspired by DPS [7]. The blur kernel size is chosen from a normal distribution with mean 85px and std 12.75px." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.483, + 0.405 + ], + "angle": 0, + "content": "Further, the intensity of the blur is randomly sampled in the range \\([0, 1]\\). The blur kernel for each image in the set is sampled independently. For evaluation, we randomly sample 4 images per view-set from the pre-computed image set and require all the methods to remove motion blurring from them." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.416, + 0.203, + 0.43 + ], + "angle": 0, + "content": "4.4.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.483, + 0.465 + ], + "angle": 0, + "content": "We compare against several baselines that can restore motion blurring from single-image or videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.541 + ], + "angle": 0, + "content": "Restormer [64]. This is a single-image motion deblurring method that does not have the benefit of jointly deblurring across a sparse view image collection. Our method outperforms this on single image deblurring and sparse view deblurring across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.602 + ], + "angle": 0, + "content": "PromptIR [31]. This is all-in-one restoration method trained to deblur. We use their pre-trained model and compare the performance of their method on sparse view deblurring." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.664 + ], + "angle": 0, + "content": "VRT [25]. This is a video-oriented restoration Transformer model trained with deblurring capability. Our method outperforms this on sparse view deblurring across all metrics. Fair comparisons on video datasets are in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.188, + 0.687 + ], + "angle": 0, + "content": "4.4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.722 + ], + "angle": 0, + "content": "We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.484, + 0.814 + ], + "angle": 0, + "content": "Qualitative. As shown in the left side of Fig. 3, our SIR-Diff show rich details and sharp deblurring restoration quality than other methods on both object level [32] and scene level [61]. This is thanks to our architectural design information from different views can be aggregated and constructive to help the other view in attention and residual layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.844 + ], + "angle": 0, + "content": "Quantitative. We assess our results using the metrics described in § 4.2, with outcomes summarized in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Our method outperforms all the methods on the three datasets on FID and LPIPS metrics. Restormer [64], a single-image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.905, + 0.3 + ], + "angle": 0, + "content": "method, lacks the ability to leverage multi-view information to enhance output quality. In contrast, our method uses 3D self-attention to achieve superior results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.391 + ], + "angle": 0, + "content": "We also outperform the video motion deblurring method, VRT, as they are incapable of handling sparse views well by motion estimation and require continuous and smooth video-captured views to fuse frame features. In the supplement, we provide a direct comparison with video methods quantitatively on continuous videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.4, + 0.724, + 0.417 + ], + "angle": 0, + "content": "4.5. Super-Resolution (SR)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.907, + 0.513 + ], + "angle": 0, + "content": "We evaluate our method on the 4x sparse-view super-resolution task. Our paired training dataset is prepared by first downsampling the high-resolution image with a scale factor of 4 and then upsampling it back to the original resolution to introduce a low-resolution effect. The resize algorithm we use is bicubic." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.625, + 0.535 + ], + "angle": 0, + "content": "4.5.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.906, + 0.571 + ], + "angle": 0, + "content": "We compare against several diffusion model baselines that handle single-image and video super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.631 + ], + "angle": 0, + "content": "SAM-DIFFSR [48]. This is an image diffusion model-based methodology, which targets classical super-resolution tasks like ours. Our method outperforms this on sparse view restoration on all the metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.676 + ], + "angle": 0, + "content": "OSEDiff [57]. This is a single-image diffusion model-based super-resolution model. OSEDiff reaches SOTA performance on the 2D image super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.907, + 0.768 + ], + "angle": 0, + "content": "Upscale-A-Video [68]. This is a video diffusion model-based super-resolution method. We input it with a sequential video frame (in total 4 for each scenario) but a larger motion gap compared to ordinary videos. We also provide a fair comparison between Upscale-A-Video and our method on video datasets in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.776, + 0.611, + 0.79 + ], + "angle": 0, + "content": "4.5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.825 + ], + "angle": 0, + "content": "We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.885, + 0.841 + ], + "angle": 0, + "content": "Qualitative. As the red rectangular box of Fig. 3 shows," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.9 + ], + "angle": 0, + "content": "for shadow artifacts and severe aliasing effects (see the second example), Upscale-A-Video [68] fails and OSEDiff [57] restoration is unfaithful and inconsistent for different views compared to ours. Moreover, our model can restore" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.361, + 0.901 + ], + "angle": 0, + "content": "1 https://github.com/LeviBorodenko/motionblur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21625" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.487, + 0.228 + ], + "angle": 0, + "content": "Table 3. Gaussian Splitting Reconstruction from Blurry Images. Under serious motion blurring, COLMAP[37] makes it hard to solve correct camera pose and SfM points, which leads to bad results on both vanilla 3DGS and BAD-GS. With the help of SIR-Diff, we can achieve restore better results and be robust to difficult motion blurring effects. KS refers to kernel size with normal distribution \\([\\mu, \\sigma]\\). Inten. refers to the blurring intensity range that is sampled uniformly. We randomly select 3 objects in CO3Dv2 [32] dataset and apply motion blur on clean images to conduct this experiment." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.231, + 0.482, + 0.316 + ], + "angle": 0, + "content": "
DifficultyMediumHard
ParamsKS:[30, 10.2], Inten:[0, 0.4]KS:[45,14.85], Inten:[0, 0.5]
MethodPSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [21]13.470.6100.60710.420.4210.801
BAD-GS [67]10.060.5370.6878.0610.3850.948
Resotmer [64]+BAD-GS [67]25.770.6160.35623.880.5600.317
SIR-Diff + BAD-NeRF [49]23.600.5030.45222.470.4390.485
SIR-Diff + BAD-GS26.110.6610.25025.330.6440.277
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.483, + 0.371 + ], + "angle": 0, + "content": "images with better geometry alignment as shown in the first example." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.485, + 0.554 + ], + "angle": 0, + "content": "Quantitative. We use PSNR, SSIM, LPIPS [66] and our proposed visual consistency § 4.1 metrics. The result is shown in Tab. 2 to show that SIR-Diff outperforms all the existing baselines on all datasets. Compared to the image diffusion method, OSEDiff[57], our spatial-3D Resnet and 3D self-attention Transformer help each latent token to access information from all views which leads to successful multiview restoration. Upscale-A-Video also employs a similar 3D ResNet architecture and cross-image attention mechanism; however, our model does not rely on conditioning with optical flow, which is challenging to obtain accurately in sparse views." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.334, + 0.58 + ], + "angle": 0, + "content": "4.6. Down Stream Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.486, + 0.768 + ], + "angle": 0, + "content": "We show several downstream applications of sparse multiview deblurring and super-resolution. Multi-view sparse images are used for tasks of finding correspondences, pose estimation, and 3D reconstruction. The performance of these methods is often dependent on the input image quality. If the underlying image quality is degraded, the output of these methods is poor. Blindly restoring these images with off-the-shelf single-image methods leads to sub-optimal results because restoration from these methods lacks view consistency. We show that, with SIR-Diff, we can consistently restore image quality and help alleviate the performance gap on tasks requiring multiview sparse views." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.338, + 0.792 + ], + "angle": 0, + "content": "4.6.1. Estimating Correspondences" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Under severely degraded inputs it is hard for methods like LoFTR [43] to estimate correspondences. As shown in Fig. 4, (a) we show qualitatively at low-resolution LoFTR struggles to find correct correspondences. Moreover, since single-view image restoration models lack multi-view consistency this does not solve the issue. SIR-Diff can not only restore the image quality but since it is view consistent can improve the" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.089, + 0.909, + 0.159 + ], + "angle": 0, + "content": "Table 4. Feature Matching Correspondence. We compute the correspondences using off-the-shelf LoFTR [43] on the restored images with single-view methods and SIR-Diff. Both in SR and deblurring cases, our method can enable LoFTR to find more correspondences (# of correspondences found)." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.164, + 0.843, + 0.232 + ], + "angle": 0, + "content": "
DeblurringSR
GT Images1979.51979.5
Degraded Images640.11374.0
Single-View Methods [57, 64]1262.61478.0
SIR-Diff1621.31494.2
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.908, + 0.317 + ], + "angle": 0, + "content": "correspondences for off-the-shelf methods like LoFTR [43]. Quantitatively, as shown in Tab. 4, we also see that LoFTR can find over 380 additional correspondences as compared to the best image-based method on deblurring and on average 15 for super-resolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.325, + 0.837, + 0.34 + ], + "angle": 0, + "content": "4.6.2. 3D Reconstruction on Degraded Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.909, + 0.51 + ], + "angle": 0, + "content": "Methods such as BAD-Gaussians [67] (BAD-GS) aim to perform 3D reconstruction from motion-blurry images, but they depend on running COLMAP [37] on blurry images to get pose initialization and sparse 3D geometry. We show in Tab. 3 that with increasing blurring effect, BAD-GS performance drops significantly while using SIR-Diff with BAD-GS leads to a much steady performance. In Fig. 5, we show a hard case of motion blurring that leads to failure of BAD-GS [67] which can be solved if we restore the image using SIR-Diff first and then do BAD-GS. We refer to supplementary for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.518, + 0.776, + 0.533 + ], + "angle": 0, + "content": "4.6.3. Sparse-View 3D reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.704 + ], + "angle": 0, + "content": "InstantSplat [11] represents a leading approach for sparse view Gaussian Splittings (GS). However, reconstructing a scene under sparse viewpoints with low-quality observations remains challenging. In Fig 6, we show an example of applying InstantSplat on an image restored using one of the baselines compared to restoration using SIR-Diff. In Tab. 5, we report the average performance of all baselines on the selected 3 scenes from Scannet++[61] and CO3Dv2[32] and build a 3DGS using InstantSplat [11]. We evaluate the GS with standard Novel View Synthesis metrics. Additional details can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.816, + 0.729 + ], + "angle": 0, + "content": "4.7. Depth Prediction for Sparse Views" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.826 + ], + "angle": 0, + "content": "We also apply our model architecture to the task of improving sparse multiview depth estimation. The only thing we edit is changing the target latent to depth image sets and the conditional latent to RGB image sets. Following Marigold [20], we use the ground truth depth label provided by the dataset for supervision signal during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Metrics. We evaluate the performance of methods on three metrics AbsRel and \\(\\delta_{1}\\) as defined by Silberman et al. [42]. In the case of MDE methods, we additionally align all the predicted depth maps to find the most optimum scale and bias to make them consistent. Similar to the view consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.904, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.217, + 0.908, + 0.259 + ], + "angle": 0, + "content": "Figure 4. Feature Matching Downstream Application. Low-Resolution and Single-image-based SR models like OSEDiff [57] find less correspondence points than our SIR-Diff that can restore multiview images consistently. SIR-Diff recovers the highest number of correspondences (310) as compared to other baseline restoration methods." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.279, + 0.48, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Figure 5. BAD-GS [67] with SIR-Diff. We show the effect of using SIR-Diff to help BAD-GS recover from catastrophic failure. BAD-GS (Top) output with a very high LPIPS\\(\\downarrow\\) of 0.59 vs. 0.22 BAD-GS output with image-restoration using SIR-Diff (Bottom) leads to a significantly higher-quality 3D reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.547, + 0.481, + 0.723 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.73, + 0.486, + 0.788 + ], + "angle": 0, + "content": "Figure 6. InstantSplat Downstream Application. 3D Reconstruction is done by InstantSplat [11]. SIR-Diff that can restore multiview images consistently leads to better visual and numerical reconstruction results than other single-image methods, Restormer [64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.87 + ], + "angle": 0, + "content": "metrics for image restoration, We evaluate Geometric Consistency (GConsis.) by calculating the distances between corresponding points between two different views under the camera coordinate system. Additional details on this metric are in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baselines. We evaluate against several monocular depth estimation (MDE) methods such as DepthAnythingV2 [60]," + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.275, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Table 5. Quantitative Results of Super-Resolution and Motion-Deblurring Reconstruction on InstantSplat [11]. The result is an average of 3 scenes per dataset (individual results in supp). Using SIR-Diff results in higher quality inputs to InstantSplat compared to OSEDiff [57] and Restormer [64]." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.349, + 0.903, + 0.475 + ], + "angle": 0, + "content": "
Scannnet++ [61]CO3D [32]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GT Image31.860.9400.09623.940.7570.243
Super-Resolution
OSEDiff [57]24.530.7990.27420.250.5490.379
SIR-Diff26.200.8270.23321.780.5780.357
Motion-Deblurring
Restormer [64]22.020.7890.30920.140.5790.465
SIR-Diff26.530.8420.23121.600.6140.384
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Table 6. Quantitative comparisons on image and video depth models. The best is highlighted and the second is underlined. GConsis. refers to geometric consistency and the unit is meter." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.532, + 0.902, + 0.606 + ], + "angle": 0, + "content": "
DatasetScannet++ [61]ETH3D [38]
MethodAbsRel ↓δ1 ↑GConsis.↓AbsRel ↓δ1 ↑GConsis.↓
DepthAnyV2 [60]9.27093.350.17713.2086.521.932
DepthCrafter [18]21.6589.960.7516.66295.051.701
Marigold [20]11.1089.920.1627.58194.460.410
SIR-Diff9.01193.540.1296.66395.050.339
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.62, + 0.768, + 0.635 + ], + "angle": 0, + "content": "DepthCrafter [18], and Marigold [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.726 + ], + "angle": 0, + "content": "Results. As shown in Tab. 6, our multi-view depth model is competitive with single image depth methods on the AbsRel, and \\(\\delta_{1}\\) metrics are sometimes better. Meanwhile, our model outperforms these methods on the geometric consistency metric that evaluates the 3D consistency when predicting depth jointly from sparse views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.756, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion & Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We present SIR-Diff, a simple and effective approach for multi-view image restoration, addressing blurring and low-resolution degradation across multiple images simultaneously. SIR-Diff outperforms leading single-view methods in sparse-view restoration and enables 3D-consistent image restoration for 3D computer vision applications. Future work will explore efficient cross-image attention mechanisms, such as selective patch attention along epipolar lines, and develop explicit guarantees for 3D view consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21627" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.207 + ], + "angle": 0, + "content": "This work partially used NCSA Delta GPU at the University of Illinois through allocation CIS230313 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.223, + 0.188, + 0.239 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[1] Miika Aittala and Frédo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.306, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.405, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[3] Sherwin Bahmani, Xian Liu, Wang Yifan, Ivan Skorokhodov, Victor Rong, Ziwei Liu, Xihui Liu, Jeong Joon Park, Sergey Tulyakov, Gordon Wetzstein, et al. Tc4d: Trajectory-conditioned text-to-4d generation. In European Conference on Computer Vision, pages 53-72. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.485, + 0.63 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4217-4229, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.772 + ], + "angle": 0, + "content": "[7] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[8] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaou Tang. Compression artifacts reduction by a deep convolutional network. In Proceedings of the IEEE international conference on computer vision, pages 576-584, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.248, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[10] Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang. Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine, and Art, 2(1):7, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[11] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[12] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jianping Fan, Yu Jun, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[13] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[14] Shi Guo, Xi Yang, Jianqi Ma, Gaofeng Ren, and Lei Zhang. A differentiable two-stage alignment scheme for burst image reconstruction with large shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17472-17481, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.907, + 0.466 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.908, + 0.58 + ], + "angle": 0, + "content": "[17] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[18] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[19] Michal Kawulok, Pawel Benecki, Szymon Piechaczek, Krzysztof Hrynczenko, Daniel Kostrzewa, and Jakub Nalepa. Deep learning for multiple-image super-resolution. IEEE Geoscience and Remote Sensing Letters, 17(6):1062-1066, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.908, + 0.788 + ], + "angle": 0, + "content": "[20] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[22] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21628" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.188 + ], + "angle": 0, + "content": "[23] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.485, + 0.26 + ], + "angle": 0, + "content": "[24] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[25] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[26] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[27] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[28] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Taming diffusion models for image restoration: A review. arXiv preprint arXiv:2409.10353, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[29] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[30] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[31] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[32] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[33] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[34] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.908, + 0.206 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.908, + 0.29 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.908, + 0.347 + ], + "angle": 0, + "content": "[37] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.908, + 0.404 + ], + "angle": 0, + "content": "[38] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.908, + 0.462 + ], + "angle": 0, + "content": "[39] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.908, + 0.531 + ], + "angle": 0, + "content": "[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.908, + 0.576 + ], + "angle": 0, + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.908, + 0.659 + ], + "angle": 0, + "content": "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.908, + 0.731 + ], + "angle": 0, + "content": "[43] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[44] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[45] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21629" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[46] Boyang Wang, Bowen Liu, Shiyu Liu, and Fengyu Yang. Vciser: Blind single image super-resolution with video compression synthetic data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4302-4312, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[47] Boyang Wang, Fengyu Yang, Xihang Yu, Chao Zhang, and Hanbin Zhao. Apisr: anime production inspired real-world anime super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25574-25584, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[48] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image super-resolution. arXiv preprint arXiv:2402.17133, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[50] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[51] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[52] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[53] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[54] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.733 + ], + "angle": 0, + "content": "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[56] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[57] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[58] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 134:109046, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.133 + ], + "angle": 0, + "content": "[59] Cheng Yang, Lijing Liang, and Zhixun Su. Real-world denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.177 + ], + "angle": 0, + "content": "[60] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[61] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.331 + ], + "angle": 0, + "content": "[62] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.908, + 0.387 + ], + "angle": 0, + "content": "[63] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.908, + 0.471 + ], + "angle": 0, + "content": "[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.908, + 0.53 + ], + "angle": 0, + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.908, + 0.599 + ], + "angle": 0, + "content": "[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.908, + 0.642 + ], + "angle": 0, + "content": "[67] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.908, + 0.713 + ], + "angle": 0, + "content": "[68] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21630" + } + ] +] \ No newline at end of file diff --git a/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_origin.pdf b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..403a79b2bbaef8660ec41dfba9a113b359332d77 --- /dev/null +++ b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/77b2d674-0341-4ee8-8978-dbba54ab0a91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1140cce3a52cc5a45784782e2aa7faddf56c34805bb215156d330a50734492b +size 8396865 diff --git a/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/full.md b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7685a367b58b82d2496cadacdb9c2d831d9a458 --- /dev/null +++ b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/full.md @@ -0,0 +1,342 @@ +# SIR-DIFF: Sparse Image Sets Restoration with Multi-View Diffusion Model + +Yucheng Mao $^{1, *}$ Boyang Wang $^{1, *}$ Nilesh Kulkarni $^{1,2}$ Jeong Joon Park $^{1}$ $^{1}$ University of Michigan, Ann Arbor $^{2}$ Netflix + +![](images/876151be1a6b5e7ea8334fc64c810db46a229e8fd9782d05bb108244540f63a5.jpg) +Figure 1. Sparse View Image Restoration. Our diffusion model takes multi-view images and jointly enhances their visual quality while maintaining 3D consistency. (Top) Four motion-blurred input images are processed by our method, resulting in sharp outputs that significantly outperform single-view restoration methods as shown in the corresponding purple boxes. (Bottom) Our method can consistently restore multi-view images (4 out of 50 shown), leading to accurate 3D reconstructions. + +# Abstract + +The computer vision community has developed numerous techniques for digitally restoring true scene information from single-view degraded photographs, an important yet extremely ill-posed task. In this work, we tackle image restoration from a different perspective by jointly denoising multiple photographs of the same scene. Our core hypothesis is that degraded images capturing a shared scene contain complementary information that, when combined, better constrains the restoration problem. To this end, we implement a powerful multi-view diffusion model that jointly generates uncorrupted views by extracting rich information from multi-view relationships. Our experiments show that our multi-view approach outperforms existing single-view im + +age and even video-based methods on image deblurring and super-resolution tasks. Critically, our model is trained to output 3D consistent images, making it a promising tool for applications requiring robust multi-view integration, such as 3D reconstruction or pose estimation. Project website: https://myc634.github.io/sirdiff/ + +# 1. Introduction + +Image restoration approaches [8, 16, 29, 53] typically process degraded images in isolation, confronting the challenge of the original scene content. Yet in practice, we frequently capture multiple photographs of meaningful moments, yielding a natural collection of views of the same scene, albeit some degraded. This common behavior presents an untapped opportunity: leveraging multiple captures to enhance image + +quality and better preserve our visual memories. + +In this work, we present a novel perspective on image restoration by reformulating it as a multi-view collaborative restoration task. We leverage our key insight that multiple corrupted photographs of the same scene contain complementary information that can aid superior denoising outcomes. The concept of multi-image fusion for quality enhancement has been explored in classical super-resolution techniques where sub-pixel shifted images are combined to reconstruct fractional pixel information [19]. Our approach revisits this paradigm in a fundamentally wider context by using neural attention mechanisms to implicitly fuse overlapping information from general multi-view images to enhance even severe aberrations. + +Notably, our multi-view restoration framework has particular significance for 3D computer vision applications. Many real-world tasks, from robotic SLAM to novel-view synthesis, rely fundamentally on multi-view imagery of a scene. When restoring degraded images within such multi-view sets, maintaining geometric consistency across views is crucial to preserve the underlying 3D scene assumptions. Traditional single-view restoration methods, inevitably introduce inconsistencies that can compromise these geometric constraints. In contrast, our restoration approach naturally enforces view consistency through learned neural priors. + +We implement our approach, dubbed SIR-Diff (Sparse Image Restoration with Diffusion), through an innovative multi-view diffusion model architecture. Our choice of diffusion models is motivated by their natural handling of uncertainty in image restoration. We extend the UNet-based Stable Diffusion [35] architecture to process multiple views jointly by introducing three key components: (1) a unified encoder for diverse degraded inputs including low-resolution, blurred, and regular RGB images, (2) a hybrid 2D-3D convolutional structure derived from SVD (Stable Video Diffusion) [4, 5], and (3) a 3D self-attention Transformer that enables cross-view attention across all spatial and multi-view tokens. + +We evaluate and show zero-shot performance on Scannet++ [61], ETH3D [38], and CO3D [32] for multi-view deblurring and super-resolution tasks and their 3D downstream applications. SIR-Diff consistently outperforms existing single-image and video restoration methods for multi-view restoration tasks in both image quality and our proposed 3D view consistency metrics. For downstream applications, we show that our multi-view restoration strategy significantly enhances 3D reconstruction quality, as evidenced by improved novel-view synthesis accuracy and feature-matching success rates. Beyond image restoration, our architecture shows promise in other multi-view tasks, notably achieving competitive depth estimation accuracy and geometric consistency compared to leading single-view methods. + +# 2. Related Works + +Image Diffusion Models. Diffusion Models (DM) show huge popularity in all domains: illustrations [9, 35, 51, 65], video creation [4, 5, 30], 3D modeling with limited input views [27, 41, 45], 4D generation [2, 3, 26], and etc. Most models are fine-tuned from the pre-trained weight instead of training from the beginning. One of the most widely-used pre-trained Diffusion Models is Stable Diffusion 2.1 [35], a UNet-based [36] architecture. + +Multiview-Image Diffusion models. In the image-to-3D task [24], Diffusion Model is a powerful generative prior to enriching details for the novel (unseen) view synthesis [6, 17, 44, 55, 56]. Methods like MVDream [41], CAT3D [13], Zero $123++$ [40] jointly generate multi-views simultaneously by extending the image diffusion model (SD2.1) with a 3D self-attention layer. However, these works ignore the exploration of the convolution layer architecture in the UNet-based Diffusion Model. In this paper, we explore the methodology of applying the 3D self-attention layer to the 3D restoration task. + +Image and Video Restoration. These methods aim at converting a degraded input image or video into a high-quality form. Traditional tasks include Super-Resolution [23, 53, 57, 62, 68], Deblurring [22, 25, 31], Denoise [10, 59], Compression Artifact Reduction [8, 46, 47], Inpainting [7, 58], Burst Image Deblurring [1, 14], and etc. These domains have achieved milestones in the 2D image and video domain [28]. However, the study on the restoration in 3D is insufficient. The restoration consistency of the texture and geometry from different views is hard to maintain when independently restoring them. + +3D Reconstruction from Corrupted Inputs. 3D Gaussian Splatting (3DGS) [21, 33, 54, 63] transforms input images into explicit-represented 3D Gaussian blobs with continuous optimization to reconstruct the large 3D scene. However, these models heavily rely on the condition of the input images. Motion blurring [50, 67] or low-resolution [12, 39] inputs that lack high-frequency and sharp details may greatly deteriorate the reconstruction quality or even fail at the initialization stage. To resolve this, methods like Bad-Gaussian [39] and SuperGaussian [67] depend on dense view input to iteratively optimize the degraded artifacts. Further, under heavily deteriorated inputs, these methods cannot recover high-quality outputs when they lack a strong generative prior [39] such as a pre-trained Diffusion Model. + +# 3. Method + +In this section, we discuss our proposed approach, SIR-Diff, which can generate and restore sparse image sets from 3D scenes or objects conditioned on their degraded inputs. Specifically, we showcase our approach to two challenging tasks: motion deblurring and super-resolution. In the + +![](images/968f1a4edd490f63fed9f564b435e81eb7ef4e209a684bbea0c3a4b5ff791c45.jpg) +Figure 2. Method Overview. Our model is a UNet-based latent diffusion model that takes as input a degraded sparse image set and outputs a set of restored consistent images. The core of our approach is a diffusion model that jointly denoises image latents across multiple views by relying on our spatial-3D ResNet and 3D Self-Attn transformer. We use pre-trained encoder-decoders from SD2.1 [35] and train our denoising UNet on those encoded latents. We show view-consistent results for the task of Deblurring and Super-Resolution. + +following sections, we discuss our problem setup in § 3.1, multi-view image restoration diffusion model in § 3.2, and our proposed model architecture in § 3.3. + +# 3.1. Multi-View Generation Setup + +Our goal is to restore sparse view images from degraded inputs. Typical methods to solve this problem work either on a single image [31, 35, 57] or require a coherent video [25, 68] as input. Toward the goal of sparse view image restoration, we propose a generative model that can respect the geometrical priors to perform the task of image restoration from blurry images or recovering high-resolution image data from its low-resolution counterpart. Any single-image generative model that performs independent inference on each image cannot ensure consistency over the generated image set. + +To alleviate these issues, we formulate this as a generation problem to infer all the required views of the scene by modeling the joint distribution using a multi-view diffusion model. Our model takes a degraded image set as input and recovers a high-quality self-consistent image set. + +# 3.2. Multi-View Image Restoration Diffusion Model + +Our model aims to restore images when conditioned on a corrupted image set $\{I^c\}_{i = 1}^N$ . The model's primary objective is to restore a self-consistent high-quality image set. We do this by using a multi-view image restoration diffusion model that can jointly restore the image set. In training, we start from a dataset of non-blurry / high-resolution images $\{I^r\}_{i = 1}^N$ and apply the forward diffusion process. Our input representation to the N-view diffusion model is $\mathbf{x} = \{I_1^r,\dots,I_N^r\}$ . Specifically, $\mathbf{x}^0$ corresponds to a clean signal at the start of the forward diffusion process at step 0. + +Our generative model learns to reverse this noising pro + +cess by predicting the added noise at every timestep of the diffusion process. + +# 3.3. Model Architecture + +We build our joint image restoration diffusion model based on a latent diffusion model [35]. Fig. 2 shows an overview of our method. The goal of our diffusion model is to denoise image latents such that a pre-trained decoder can convert them to high-quality images. When simultaneously denoising multiple image latents, it is important that latent representation remain consistent across views. We enable this by allowing the diffusion model to attend to latents from other views. + +# 3.3.1. Latent Diffusion Models for Restoration + +A Latent Diffusion Model [35] architecture, such as Stable Diffusion 2.1 (SD2.1), consists of two parts a) encoder and decoder and b) denoising UNet. + +During inference, our latent denoising UNet progressively refines pure noise into a clean latent representation, conditioned on the degraded image set, through several denoising steps. Once the latent representation is clean, a decoder translates it into the restored image set. + +Similarly, during training, our diffusion model is optimized to learn the denoising process of the latent representation, conditioned on the degraded input sets. + +Encoder-Decoder. We initialize our encoder with the weights from an SD2.1 [35] and are frozen when we train our model. The encoder, $\phi$ , is used to map our images to latent vectors which are used to learn the UNet-Style diffusion model. The decoder, $\delta$ , converts the denoised latents from the diffusion model to pixels in image space. + +UNet. Our latent denoising UNet, $\mathbf{s}_{\theta}(\cdot)$ with parameters $\theta$ , considers as input a noisy latent $\mathbf{x}^k$ , encoded latents for + +![](images/2043b64822d415bb507354c5896c9505e030d7cbf1cc20b43e01469320497cee.jpg) +Figure 3. Qualitative Comparisons on Motion Deblurring and Super-Resolution. The left side is the Motion-Deblurring and the right side is the Super-Resolution results. We refer readers to supplementary for more visualizations. Zoom in for the best view. + +degraded images, $\{I^c\}_{i = 1}^N$ and the diffusion noise $k$ to output an of estimate the noise, $\hat{\pmb{\epsilon}}$ , as shown in Eq. (1): + +$$ +\hat {\epsilon} = \mathbf {s} _ {\theta} (\mathbf {x} ^ {k}, \phi (\{I ^ {c} \} _ {i = 1} ^ {N}), k). \qquad (1) +$$ + +We supervise the outputs of our UNet to match the noise added during the forward diffusion process using Eq. 2: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon \sim \mathcal {N} (0, I), k \sim \mathcal {U} (T)} \| \epsilon - \hat {\epsilon} \| _ {2} ^ {2}. \tag {2} +$$ + +# 3.3.2. Adapting Diffusion Model to multiple views + +Starting with the image diffusion model, we expand a new dimension to represent image sets from different views. Most of the time, batch and views are merged into one dimension and split when needed. We encode the degraded image set with the same VAE encoder as the target noisy latents. Noisy latents and the condition latents are integrated by channel-wise concatenation. Then, due to channel number change, we replace the original convolution with a new zero-initialized convolution layer [65] to adapt the new number of channels input. + +# 3.3.3. Spatial-3D Integrated UNet Architecture + +Typical latent diffusion models such as Stable Diffusion [35] do not consider denoise multiple images simultaneously and often generate individual images that lead to inconsistent image generations between multi-views. Our solution to remove this inconsistency is to denoise all the views jointly by allowing the UNet to look at the intermediate latent representations from different views. As shown in the right side of Fig. 2, our architecture is composed of a Spatial-3D ResNet and a 3D self-attention Transformer. + +Spatial-3D ResNet. Image diffusion model like SD2.1 only has a 2D convolution layer in the ResNet, which makes it + +hard to perform well in tasks that require 3D understanding. We want to enhance the 3D performance of the ResNet of the image diffusion model by introducing a 3D convolution layer to better capture geometry information as shown in Fig. 2. Training from raw will be unstable and time-consuming, but we find that the video diffusion model that is performing the image-to-video task already has a pre-trained 3D convolution layer. Understanding temporal similarity in the time dimension and understanding 3D similarity in the spatial dimension share similarity. By the experiment, we empirically find that merging two different weights into one is feasible and can lead to better performance. Hence, we initialize the 2D Convolution layer with image Diffusion Model weight (SD2.1) and the 3D convolution layer with video Diffusion weight (Stable Video Diffusion [4]). The output of each layer will be blended as follows: $O_{\mathrm{ResNet}} = \sigma (\alpha)\times O_{2D} + \sigma (1 - \alpha)\times O_{3D}$ , where $\sigma$ is the sigmoid function and $\alpha$ is usually taken as 0.5 following SVD [4]. $O_{2D}$ is the output of 2D ResNet and $O_{3D}$ is the output of 3D ResNet layer. + +3D Self-Attention Transformer. Furthermore, we adopt the 3D multi-view consistent self-attention Transformer as shown in Fig. 2, similar to previous works [13, 41]. Our self-attention layer considers image latents from all the $N$ views and performs self-attention. After ResNet, the model patches the latent to create $p$ patch token per view, and in total, we have $N \times p$ tokens that are part of the attention layer. Each token calculates the attention score to all $N \times p$ tokens as shown on the right side of Fig. 2. This feature enables each token to be accessible to all conditional image sets given at the maximum level. Further, by this design, our model, in the inference stage, is scalable to a variable number of input frames with stable performance even though + +Table 1. Sparse-View Deblurring Comparisons. We show quantitative results on three datasets and evaluate metrics as in § 4.1. VConsis. (visual consistency) is multiplied by 100. Across all methods, SIR-Diff is the most visually consistent as compared to all the baselines. On other standard metrics, we also outperform baselines. SIR-Diff (1 Frame) is a single image version of our method that does not depend on multiple views. The best is highlighted, and the second is underlined. + +
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodFID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓
PromptIR [31]81.280.2487.81150.340.53710.72166.30.35012.89
Restromer [64]49.720.2326.5273.470.3309.59144.80.42611.00
VRT [25]134.50.3717.67160.40.53011.13170.40.44012.07
SIR-Diff (1 Frame)81.580.2476.45109.60.3509.55137.90.35710.36
SIR-Diff40.090.1605.7567.310.2639.4385.670.2759.30
+ +we train with a handful of images per set. + +However, attention communication between views leads to heavy computes. To resolve this, we only inject this module into the low-resolution layers (similar to CAT3D [13]), where we also ignore the biggest resolution layer for the UNet Encoder and Decoder. Further, we completely take away the cross-attention module across all layers instead of putting empty string inputs as placeholders like previous works [20]. We found that both of these changes help faster convergence, less training time, and better numerical results under the same number of training iterations. + +# 4. Experiment + +We evaluate SIR-Diff on the task of sparse view motion deblurring and super-resolution in a zero-shot setting $\S 4.2$ , and discuss our deblurring results in $\S 4.4$ and super-resolution results in $\S 4.5$ . We discuss implementation details in $\S 4.2$ , and now we start by discussing evaluation metrics in $\S 4.1$ . Please refer to supplementary materials for the ablation study. + +# 4.1. Metrics + +We use several metrics to evaluate the performance of all the methods on the task of motion deblurring and superresolution. Since our approach is generative in nature we borrow metrics from the literature that are good at evaluating generative models such as Frechet Inception Distance (FID) [15], Learned Perceptual Image Patch Similarity (LPIPS) [66]. In addition to these, we evaluate our superresolution results on PSNR and SSIM similar to methods used in this literature [57, 62, 64, 68]. We also propose a new metric to evaluate view consistency between pairs of generated views. + +Visual Consistency (VConsis). This metric evaluates self-consistency within the generated view set by measuring the perceptual similarity of corresponding patches between the two views. Please refer to the supplementary for additional details. + +# 4.2. Implementation Details + +We implement our latent diffusion model based on pretrained UNet and auto-encoder from SDv2.1 as part of dif + +fusers. We train our models on two synthetic datasets Hypersim [34] and TartanAir [52] with image resolution resized to $640 \times 480$ . We randomly select the order of image sets to maintain permutation invariance in training and inference. We utilize ground truth geometry and pose data to pre-compute the overlap region between each pair of frames. We provide additional details on model training, dataset, and view selection in the supplemental. + +# 4.3. Zero Shots Datasets For Evaluation + +As all our baselines are training on different datasets, we perform a zero-shot evaluation across a diverse set of real-world datasets to keep a fair comparison. Zero-Shot means rather than split the dataset into training-evaluation splits, we chose some dataset for training and some other out-of-domain for evaluation. Our evaluation datasets consist of indoor and outdoor scenes along with object-centric datasets. ScanNet++ [61]. This is a dataset of 460 indoor scanned readl 3D scenes. We evaluate all the methods on the scenes from the official test set of 50 scenes, by sampling 200 image sets with four images per set. We highlight the zero-shot performance of our method trained on synthetic indoor data transfers well to real indoor scene data. + +ETH3D [38]. This dataset provides high-resolution images of indoor and outdoor scenes, with accurate depth maps captured via a LiDAR. We evaluate all the 454 samples from the ETH3D dataset. + +CO3Dv2 [32]. This is an object-centric dataset containing object-level videos across 51 categories of household objects. This dataset has pose annotations from COLMAP [37] provided for all video sequences. We use the official single-sequence subsets for evaluation and randomly sample 200 image sets among them. + +For all these three datasets, please refer to the supplementary material for the detail of view-selection strategy + +# 4.4. Motion Deblurring (MD) + +Our motion deblurring (MD) method targets solving arbitrary motion paths in the wild. We target at restoring original images from the set of sparse view motion blur inputs. Our paired training dataset is prepared by applying the motion- + +Table 2. Sparse-View Super-Resolution Comparisons. VConsis. (visual consistency) is multiplied by 100. SIR-Diff (1 Frame) is a single image process form like other image diffusion models. Ours, in the inference stage, can adapt an arbitrary number of views. The best is highlighted and the second is underlined. + +
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodPSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓
SAM-DiffSR [48]24.130.2160.7586.9419.160.3170.4958.2823.900.2460.62610.96
OSEDiff [57]25.220.8080.1545.5121.930.6210.2648.6224.500.6580.2778.86
UAV [68]23.230.6770.3549.0120.390.5060.4139.2622.730.5660.40011.82
SIR-Diff (1 Frame)25.850.7920.1645.7021.970.5930.2217.9024.350.6410.3019.07
SIR-Diff27.380.8370.0995.0122.580.6210.2017.6024.530.6580.2107.85
+ +blurring kernel on the clean images from the training datasets based on the codebase inspired by DPS [7]. The blur kernel size is chosen from a normal distribution with mean 85px and std 12.75px. + +Further, the intensity of the blur is randomly sampled in the range $[0, 1]$ . The blur kernel for each image in the set is sampled independently. For evaluation, we randomly sample 4 images per view-set from the pre-computed image set and require all the methods to remove motion blurring from them. + +# 4.4.1. Baselines + +We compare against several baselines that can restore motion blurring from single-image or videos. + +Restormer [64]. This is a single-image motion deblurring method that does not have the benefit of jointly deblurring across a sparse view image collection. Our method outperforms this on single image deblurring and sparse view deblurring across all metrics. + +PromptIR [31]. This is all-in-one restoration method trained to deblur. We use their pre-trained model and compare the performance of their method on sparse view deblurring. + +VRT [25]. This is a video-oriented restoration Transformer model trained with deblurring capability. Our method outperforms this on sparse view deblurring across all metrics. Fair comparisons on video datasets are in the supplemental. + +# 4.4.2. Results + +We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets. + +Qualitative. As shown in the left side of Fig. 3, our SIR-Diff show rich details and sharp deblurring restoration quality than other methods on both object level [32] and scene level [61]. This is thanks to our architectural design information from different views can be aggregated and constructive to help the other view in attention and residual layers. + +Quantitative. We assess our results using the metrics described in § 4.2, with outcomes summarized in Tab. 1. + +Our method outperforms all the methods on the three datasets on FID and LPIPS metrics. Restormer [64], a single-image + +method, lacks the ability to leverage multi-view information to enhance output quality. In contrast, our method uses 3D self-attention to achieve superior results. + +We also outperform the video motion deblurring method, VRT, as they are incapable of handling sparse views well by motion estimation and require continuous and smooth video-captured views to fuse frame features. In the supplement, we provide a direct comparison with video methods quantitatively on continuous videos. + +# 4.5. Super-Resolution (SR) + +We evaluate our method on the 4x sparse-view super-resolution task. Our paired training dataset is prepared by first downsampling the high-resolution image with a scale factor of 4 and then upsampling it back to the original resolution to introduce a low-resolution effect. The resize algorithm we use is bicubic. + +# 4.5.1. Baselines + +We compare against several diffusion model baselines that handle single-image and video super-resolution. + +SAM-DIFFSR [48]. This is an image diffusion model-based methodology, which targets classical super-resolution tasks like ours. Our method outperforms this on sparse view restoration on all the metrics. + +OSEDiff [57]. This is a single-image diffusion model-based super-resolution model. OSEDiff reaches SOTA performance on the 2D image super-resolution. + +Upscale-A-Video [68]. This is a video diffusion model-based super-resolution method. We input it with a sequential video frame (in total 4 for each scenario) but a larger motion gap compared to ordinary videos. We also provide a fair comparison between Upscale-A-Video and our method on video datasets in the supplement. + +# 4.5.2. Results + +We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets. + +Qualitative. As the red rectangular box of Fig. 3 shows, + +for shadow artifacts and severe aliasing effects (see the second example), Upscale-A-Video [68] fails and OSEDiff [57] restoration is unfaithful and inconsistent for different views compared to ours. Moreover, our model can restore + +Table 3. Gaussian Splitting Reconstruction from Blurry Images. Under serious motion blurring, COLMAP[37] makes it hard to solve correct camera pose and SfM points, which leads to bad results on both vanilla 3DGS and BAD-GS. With the help of SIR-Diff, we can achieve restore better results and be robust to difficult motion blurring effects. KS refers to kernel size with normal distribution $[\mu, \sigma]$ . Inten. refers to the blurring intensity range that is sampled uniformly. We randomly select 3 objects in CO3Dv2 [32] dataset and apply motion blur on clean images to conduct this experiment. + +
DifficultyMediumHard
ParamsKS:[30, 10.2], Inten:[0, 0.4]KS:[45,14.85], Inten:[0, 0.5]
MethodPSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [21]13.470.6100.60710.420.4210.801
BAD-GS [67]10.060.5370.6878.0610.3850.948
Resotmer [64]+BAD-GS [67]25.770.6160.35623.880.5600.317
SIR-Diff + BAD-NeRF [49]23.600.5030.45222.470.4390.485
SIR-Diff + BAD-GS26.110.6610.25025.330.6440.277
+ +images with better geometry alignment as shown in the first example. + +Quantitative. We use PSNR, SSIM, LPIPS [66] and our proposed visual consistency § 4.1 metrics. The result is shown in Tab. 2 to show that SIR-Diff outperforms all the existing baselines on all datasets. Compared to the image diffusion method, OSEDiff[57], our spatial-3D Resnet and 3D self-attention Transformer help each latent token to access information from all views which leads to successful multiview restoration. Upscale-A-Video also employs a similar 3D ResNet architecture and cross-image attention mechanism; however, our model does not rely on conditioning with optical flow, which is challenging to obtain accurately in sparse views. + +# 4.6. Down Stream Applications + +We show several downstream applications of sparse multiview deblurring and super-resolution. Multi-view sparse images are used for tasks of finding correspondences, pose estimation, and 3D reconstruction. The performance of these methods is often dependent on the input image quality. If the underlying image quality is degraded, the output of these methods is poor. Blindly restoring these images with off-the-shelf single-image methods leads to sub-optimal results because restoration from these methods lacks view consistency. We show that, with SIR-Diff, we can consistently restore image quality and help alleviate the performance gap on tasks requiring multiview sparse views. + +# 4.6.1. Estimating Correspondences + +Under severely degraded inputs it is hard for methods like LoFTR [43] to estimate correspondences. As shown in Fig. 4, (a) we show qualitatively at low-resolution LoFTR struggles to find correct correspondences. Moreover, since single-view image restoration models lack multi-view consistency this does not solve the issue. SIR-Diff can not only restore the image quality but since it is view consistent can improve the + +Table 4. Feature Matching Correspondence. We compute the correspondences using off-the-shelf LoFTR [43] on the restored images with single-view methods and SIR-Diff. Both in SR and deblurring cases, our method can enable LoFTR to find more correspondences (# of correspondences found). + +
DeblurringSR
GT Images1979.51979.5
Degraded Images640.11374.0
Single-View Methods [57, 64]1262.61478.0
SIR-Diff1621.31494.2
+ +correspondences for off-the-shelf methods like LoFTR [43]. Quantitatively, as shown in Tab. 4, we also see that LoFTR can find over 380 additional correspondences as compared to the best image-based method on deblurring and on average 15 for super-resolution. + +# 4.6.2. 3D Reconstruction on Degraded Images + +Methods such as BAD-Gaussians [67] (BAD-GS) aim to perform 3D reconstruction from motion-blurry images, but they depend on running COLMAP [37] on blurry images to get pose initialization and sparse 3D geometry. We show in Tab. 3 that with increasing blurring effect, BAD-GS performance drops significantly while using SIR-Diff with BAD-GS leads to a much steady performance. In Fig. 5, we show a hard case of motion blurring that leads to failure of BAD-GS [67] which can be solved if we restore the image using SIR-Diff first and then do BAD-GS. We refer to supplementary for details. + +# 4.6.3. Sparse-View 3D reconstruction + +InstantSplat [11] represents a leading approach for sparse view Gaussian Splittings (GS). However, reconstructing a scene under sparse viewpoints with low-quality observations remains challenging. In Fig 6, we show an example of applying InstantSplat on an image restored using one of the baselines compared to restoration using SIR-Diff. In Tab. 5, we report the average performance of all baselines on the selected 3 scenes from Scannet++[61] and CO3Dv2[32] and build a 3DGS using InstantSplat [11]. We evaluate the GS with standard Novel View Synthesis metrics. Additional details can be found in the supplementary. + +# 4.7. Depth Prediction for Sparse Views + +We also apply our model architecture to the task of improving sparse multiview depth estimation. The only thing we edit is changing the target latent to depth image sets and the conditional latent to RGB image sets. Following Marigold [20], we use the ground truth depth label provided by the dataset for supervision signal during training. + +Metrics. We evaluate the performance of methods on three metrics AbsRel and $\delta_{1}$ as defined by Silberman et al. [42]. In the case of MDE methods, we additionally align all the predicted depth maps to find the most optimum scale and bias to make them consistent. Similar to the view consistency + +![](images/b5124e623cf3a17a2f23c804cbf7ad335319e89e482c8254fec506115b8b6760.jpg) +Figure 4. Feature Matching Downstream Application. Low-Resolution and Single-image-based SR models like OSEDiff [57] find less correspondence points than our SIR-Diff that can restore multiview images consistently. SIR-Diff recovers the highest number of correspondences (310) as compared to other baseline restoration methods. + +![](images/2f68868daecdb2fc2a1a8b22c84d9bbdee8875d10d5644a83587142eb6b8bd40.jpg) +Figure 5. BAD-GS [67] with SIR-Diff. We show the effect of using SIR-Diff to help BAD-GS recover from catastrophic failure. BAD-GS (Top) output with a very high LPIPS $\downarrow$ of 0.59 vs. 0.22 BAD-GS output with image-restoration using SIR-Diff (Bottom) leads to a significantly higher-quality 3D reconstruction. + +![](images/83122bcf17bc8a428d4312fbb1b1dd32fe8d47c46a6530116a18c93becc64e1e.jpg) +Figure 6. InstantSplat Downstream Application. 3D Reconstruction is done by InstantSplat [11]. SIR-Diff that can restore multiview images consistently leads to better visual and numerical reconstruction results than other single-image methods, Restormer [64]. + +metrics for image restoration, We evaluate Geometric Consistency (GConsis.) by calculating the distances between corresponding points between two different views under the camera coordinate system. Additional details on this metric are in the supplemental. + +Baselines. We evaluate against several monocular depth estimation (MDE) methods such as DepthAnythingV2 [60], + +Table 5. Quantitative Results of Super-Resolution and Motion-Deblurring Reconstruction on InstantSplat [11]. The result is an average of 3 scenes per dataset (individual results in supp). Using SIR-Diff results in higher quality inputs to InstantSplat compared to OSEDiff [57] and Restormer [64]. + +
Scannnet++ [61]CO3D [32]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GT Image31.860.9400.09623.940.7570.243
Super-Resolution
OSEDiff [57]24.530.7990.27420.250.5490.379
SIR-Diff26.200.8270.23321.780.5780.357
Motion-Deblurring
Restormer [64]22.020.7890.30920.140.5790.465
SIR-Diff26.530.8420.23121.600.6140.384
+ +Table 6. Quantitative comparisons on image and video depth models. The best is highlighted and the second is underlined. GConsis. refers to geometric consistency and the unit is meter. + +
DatasetScannet++ [61]ETH3D [38]
MethodAbsRel ↓δ1 ↑GConsis.↓AbsRel ↓δ1 ↑GConsis.↓
DepthAnyV2 [60]9.27093.350.17713.2086.521.932
DepthCrafter [18]21.6589.960.7516.66295.051.701
Marigold [20]11.1089.920.1627.58194.460.410
SIR-Diff9.01193.540.1296.66395.050.339
+ +DepthCrafter [18], and Marigold [20]. + +Results. As shown in Tab. 6, our multi-view depth model is competitive with single image depth methods on the AbsRel, and $\delta_{1}$ metrics are sometimes better. Meanwhile, our model outperforms these methods on the geometric consistency metric that evaluates the 3D consistency when predicting depth jointly from sparse views. + +# 5. Conclusion & Limitations + +We present SIR-Diff, a simple and effective approach for multi-view image restoration, addressing blurring and low-resolution degradation across multiple images simultaneously. SIR-Diff outperforms leading single-view methods in sparse-view restoration and enables 3D-consistent image restoration for 3D computer vision applications. Future work will explore efficient cross-image attention mechanisms, such as selective patch attention along epipolar lines, and develop explicit guarantees for 3D view consistency. + +# Acknowledgment + +This work partially used NCSA Delta GPU at the University of Illinois through allocation CIS230313 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. + +# References + +[1] Miika Aittala and Frédo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 2 +[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 2 +[3] Sherwin Bahmani, Xian Liu, Wang Yifan, Ivan Skorokhodov, Victor Rong, Ziwei Liu, Xihui Liu, Jeong Joon Park, Sergey Tulyakov, Gordon Wetzstein, et al. Tc4d: Trajectory-conditioned text-to-4d generation. In European Conference on Computer Vision, pages 53-72. Springer, 2025. 2 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4217-4229, 2023. 2 +[7] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 6 +[8] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaou Tang. Compression artifacts reduction by a deep convolutional network. In Proceedings of the IEEE international conference on computer vision, pages 576-584, 2015. 1, 2 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2 + +[10] Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang. Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine, and Art, 2(1):7, 2019. 2 +[11] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 7, 8 +[12] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jianping Fan, Yu Jun, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024. 2 +[13] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 4, 5 +[14] Shi Guo, Xi Yang, Jianqi Ma, Gaofeng Ren, and Lei Zhang. A differentiable two-stage alignment scheme for burst image reconstruction with large shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17472-17481, 2022. 2 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[17] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2 +[18] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 8 +[19] Michal Kawulok, Pawel Benecki, Szymon Piechaczek, Krzysztof Hrynczenko, Daniel Kostrzewa, and Jakub Nalepa. Deep learning for multiple-image super-resolution. IEEE Geoscience and Remote Sensing Letters, 17(6):1062-1066, 2019. 2 +[20] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 5, 7, 8 +[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 7 +[22] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2 + +[23] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2 +[24] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2 +[25] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 2, 3, 5, 6 +[26] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 2 +[27] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2 +[28] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Taming diffusion models for image restoration: A review. arXiv preprint arXiv:2409.10353, 2024. 2 +[29] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1 +[30] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2 +[31] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 6 +[32] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2, 5, 6, 7, 8 +[33] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 2 +[34] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic + +dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2 +[37] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5, 7 +[38] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2, 5, 6, 8 +[39] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024. 2 +[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2 +[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 4 +[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7 +[43] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7 +[44] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2 +[45] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2 + +[46] Boyang Wang, Bowen Liu, Shiyu Liu, and Fengyu Yang. Vciser: Blind single image super-resolution with video compression synthetic data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4302-4312, 2024. 2 +[47] Boyang Wang, Fengyu Yang, Xihang Yu, Chao Zhang, and Hanbin Zhao. Apisr: anime production inspired real-world anime super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25574-25584, 2024. 2 +[48] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image super-resolution. arXiv preprint arXiv:2402.17133, 2024. 6 +[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 7 +[50] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 2 +[51] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2 +[52] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5 +[53] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2 +[54] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2 +[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[56] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2 +[57] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 5, 6, 7, 8 +[58] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 134:109046, 2023. 2 + +[59] Cheng Yang, Lijing Liang, and Zhixun Su. Real-world denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023.2 +[60] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8 +[61] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 5, 6, 7, 8 +[62] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 2, 5 +[63] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2 +[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5, 6, 7, 8 +[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4 +[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7 +[67] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 2, 7, 8 +[68] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 2, 3, 5, 6 \ No newline at end of file diff --git a/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/images.zip b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23fc476dcdeaf1b4923deb7d2312b86109cdb651 --- /dev/null +++ b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94cb4801b13eb20d8169427167d65745cad8fd6fa76866c3ab05b2e7138e72ed +size 690343 diff --git a/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/layout.json b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..92750b476ab965457dac8035a5f33a982c2d6382 --- /dev/null +++ b/2025/SIR-DIFF_ Sparse Image Sets Restoration with Multi-View Diffusion Model/layout.json @@ -0,0 +1,8360 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 539, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 539, + 120 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 539, + 120 + ], + "type": "text", + "content": "SIR-DIFF: Sparse Image Sets Restoration with Multi-View Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "content": "Yucheng Mao" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1, *}" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "content": " Boyang Wang" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1, *}" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "content": " Nilesh Kulkarni" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "content": " Jeong Joon Park" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "content": "University of Michigan, Ann Arbor " + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 141, + 491, + 186 + ], + "type": "text", + "content": "Netflix" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 201, + 531, + 467 + ], + "blocks": [ + { + "bbox": [ + 78, + 201, + 531, + 467 + ], + "lines": [ + { + "bbox": [ + 78, + 201, + 531, + 467 + ], + "spans": [ + { + "bbox": [ + 78, + 201, + 531, + 467 + ], + "type": "image", + "image_path": "876151be1a6b5e7ea8334fc64c810db46a229e8fd9782d05bb108244540f63a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 472, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 54, + 472, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 555, + 518 + ], + "type": "text", + "content": "Figure 1. Sparse View Image Restoration. Our diffusion model takes multi-view images and jointly enhances their visual quality while maintaining 3D consistency. (Top) Four motion-blurred input images are processed by our method, resulting in sharp outputs that significantly outperform single-view restoration methods as shown in the corresponding purple boxes. (Bottom) Our method can consistently restore multi-view images (4 out of 50 shown), leading to accurate 3D reconstructions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 526, + 200, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 526, + 200, + 538 + ], + "spans": [ + { + "bbox": [ + 151, + 526, + 200, + 538 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 297, + 696 + ], + "type": "text", + "content": "The computer vision community has developed numerous techniques for digitally restoring true scene information from single-view degraded photographs, an important yet extremely ill-posed task. In this work, we tackle image restoration from a different perspective by jointly denoising multiple photographs of the same scene. Our core hypothesis is that degraded images capturing a shared scene contain complementary information that, when combined, better constrains the restoration problem. To this end, we implement a powerful multi-view diffusion model that jointly generates uncorrupted views by extracting rich information from multi-view relationships. Our experiments show that our multi-view approach outperforms existing single-view im" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 527, + 556, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 556, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 556, + 599 + ], + "type": "text", + "content": "age and even video-based methods on image deblurring and super-resolution tasks. Critically, our model is trained to output 3D consistent images, making it a promising tool for applications requiring robust multi-view integration, such as 3D reconstruction or pose estimation. Project website: https://myc634.github.io/sirdiff/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 613, + 395, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 613, + 395, + 625 + ], + "spans": [ + { + "bbox": [ + 314, + 613, + 395, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Image restoration approaches [8, 16, 29, 53] typically process degraded images in isolation, confronting the challenge of the original scene content. Yet in practice, we frequently capture multiple photographs of meaningful moments, yielding a natural collection of views of the same scene, albeit some degraded. This common behavior presents an untapped opportunity: leveraging multiple captures to enhance image" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 138, + 712 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21620" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 85 + ], + "type": "text", + "content": "quality and better preserve our visual memories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 247 + ], + "type": "text", + "content": "In this work, we present a novel perspective on image restoration by reformulating it as a multi-view collaborative restoration task. We leverage our key insight that multiple corrupted photographs of the same scene contain complementary information that can aid superior denoising outcomes. The concept of multi-image fusion for quality enhancement has been explored in classical super-resolution techniques where sub-pixel shifted images are combined to reconstruct fractional pixel information [19]. Our approach revisits this paradigm in a fundamentally wider context by using neural attention mechanisms to implicitly fuse overlapping information from general multi-view images to enhance even severe aberrations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 298, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 298, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 298, + 388 + ], + "type": "text", + "content": "Notably, our multi-view restoration framework has particular significance for 3D computer vision applications. Many real-world tasks, from robotic SLAM to novel-view synthesis, rely fundamentally on multi-view imagery of a scene. When restoring degraded images within such multi-view sets, maintaining geometric consistency across views is crucial to preserve the underlying 3D scene assumptions. Traditional single-view restoration methods, inevitably introduce inconsistencies that can compromise these geometric constraints. In contrast, our restoration approach naturally enforces view consistency through learned neural priors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 395, + 298, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 298, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 298, + 539 + ], + "type": "text", + "content": "We implement our approach, dubbed SIR-Diff (Sparse Image Restoration with Diffusion), through an innovative multi-view diffusion model architecture. Our choice of diffusion models is motivated by their natural handling of uncertainty in image restoration. We extend the UNet-based Stable Diffusion [35] architecture to process multiple views jointly by introducing three key components: (1) a unified encoder for diverse degraded inputs including low-resolution, blurred, and regular RGB images, (2) a hybrid 2D-3D convolutional structure derived from SVD (Stable Video Diffusion) [4, 5], and (3) a 3D self-attention Transformer that enables cross-view attention across all spatial and multi-view tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 298, + 714 + ], + "type": "text", + "content": "We evaluate and show zero-shot performance on Scannet++ [61], ETH3D [38], and CO3D [32] for multi-view deblurring and super-resolution tasks and their 3D downstream applications. SIR-Diff consistently outperforms existing single-image and video restoration methods for multi-view restoration tasks in both image quality and our proposed 3D view consistency metrics. For downstream applications, we show that our multi-view restoration strategy significantly enhances 3D reconstruction quality, as evidenced by improved novel-view synthesis accuracy and feature-matching success rates. Beyond image restoration, our architecture shows promise in other multi-view tasks, notably achieving competitive depth estimation accuracy and geometric consistency compared to leading single-view methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 555, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 187 + ], + "type": "text", + "content": "Image Diffusion Models. Diffusion Models (DM) show huge popularity in all domains: illustrations [9, 35, 51, 65], video creation [4, 5, 30], 3D modeling with limited input views [27, 41, 45], 4D generation [2, 3, 26], and etc. Most models are fine-tuned from the pre-trained weight instead of training from the beginning. One of the most widely-used pre-trained Diffusion Models is Stable Diffusion 2.1 [35], a UNet-based [36] architecture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 188, + 556, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 556, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 556, + 319 + ], + "type": "text", + "content": "Multiview-Image Diffusion models. In the image-to-3D task [24], Diffusion Model is a powerful generative prior to enriching details for the novel (unseen) view synthesis [6, 17, 44, 55, 56]. Methods like MVDream [41], CAT3D [13], Zero " + }, + { + "bbox": [ + 313, + 188, + 556, + 319 + ], + "type": "inline_equation", + "content": "123++" + }, + { + "bbox": [ + 313, + 188, + 556, + 319 + ], + "type": "text", + "content": " [40] jointly generate multi-views simultaneously by extending the image diffusion model (SD2.1) with a 3D self-attention layer. However, these works ignore the exploration of the convolution layer architecture in the UNet-based Diffusion Model. In this paper, we explore the methodology of applying the 3D self-attention layer to the 3D restoration task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 320, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 452 + ], + "type": "text", + "content": "Image and Video Restoration. These methods aim at converting a degraded input image or video into a high-quality form. Traditional tasks include Super-Resolution [23, 53, 57, 62, 68], Deblurring [22, 25, 31], Denoise [10, 59], Compression Artifact Reduction [8, 46, 47], Inpainting [7, 58], Burst Image Deblurring [1, 14], and etc. These domains have achieved milestones in the 2D image and video domain [28]. However, the study on the restoration in 3D is insufficient. The restoration consistency of the texture and geometry from different views is hard to maintain when independently restoring them." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 557, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 557, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 557, + 621 + ], + "type": "text", + "content": "3D Reconstruction from Corrupted Inputs. 3D Gaussian Splatting (3DGS) [21, 33, 54, 63] transforms input images into explicit-represented 3D Gaussian blobs with continuous optimization to reconstruct the large 3D scene. However, these models heavily rely on the condition of the input images. Motion blurring [50, 67] or low-resolution [12, 39] inputs that lack high-frequency and sharp details may greatly deteriorate the reconstruction quality or even fail at the initialization stage. To resolve this, methods like Bad-Gaussian [39] and SuperGaussian [67] depend on dense view input to iteratively optimize the degraded artifacts. Further, under heavily deteriorated inputs, these methods cannot recover high-quality outputs when they lack a strong generative prior [39] such as a pre-trained Diffusion Model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 633, + 370, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 370, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 370, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "In this section, we discuss our proposed approach, SIR-Diff, which can generate and restore sparse image sets from 3D scenes or objects conditioned on their degraded inputs. Specifically, we showcase our approach to two challenging tasks: motion deblurring and super-resolution. In the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21621" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 63, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 59, + 63, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 63, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 63, + 553, + 247 + ], + "type": "image", + "image_path": "968f1a4edd490f63fed9f564b435e81eb7ef4e209a684bbea0c3a4b5ff791c45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "Figure 2. Method Overview. Our model is a UNet-based latent diffusion model that takes as input a degraded sparse image set and outputs a set of restored consistent images. The core of our approach is a diffusion model that jointly denoises image latents across multiple views by relying on our spatial-3D ResNet and 3D Self-Attn transformer. We use pre-trained encoder-decoders from SD2.1 [35] and train our denoising UNet on those encoded latents. We show view-consistent results for the task of Deblurring and Super-Resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "content": "following sections, we discuss our problem setup in § 3.1, multi-view image restoration diffusion model in § 3.2, and our proposed model architecture in § 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 217, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 217, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 217, + 370 + ], + "type": "text", + "content": "3.1. Multi-View Generation Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 375, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 295, + 495 + ], + "type": "text", + "content": "Our goal is to restore sparse view images from degraded inputs. Typical methods to solve this problem work either on a single image [31, 35, 57] or require a coherent video [25, 68] as input. Toward the goal of sparse view image restoration, we propose a generative model that can respect the geometrical priors to perform the task of image restoration from blurry images or recovering high-resolution image data from its low-resolution counterpart. Any single-image generative model that performs independent inference on each image cannot ensure consistency over the generated image set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": "To alleviate these issues, we formulate this as a generation problem to infer all the required views of the scene by modeling the joint distribution using a multi-view diffusion model. Our model takes a degraded image set as input and recovers a high-quality self-consistent image set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 295, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 576 + ], + "type": "text", + "content": "3.2. Multi-View Image Restoration Diffusion Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "text", + "content": "Our model aims to restore images when conditioned on a corrupted image set " + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\{I^c\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "text", + "content": ". The model's primary objective is to restore a self-consistent high-quality image set. We do this by using a multi-view image restoration diffusion model that can jointly restore the image set. In training, we start from a dataset of non-blurry / high-resolution images " + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\{I^r\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "text", + "content": " and apply the forward diffusion process. Our input representation to the N-view diffusion model is " + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{I_1^r,\\dots,I_N^r\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^0" + }, + { + "bbox": [ + 55, + 582, + 296, + 700 + ], + "type": "text", + "content": " corresponds to a clean signal at the start of the forward diffusion process at step 0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "content": "Our generative model learns to reverse this noising pro" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 313, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 338 + ], + "type": "text", + "content": "cess by predicting the added noise at every timestep of the diffusion process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 429, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 429, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 429, + 358 + ], + "type": "text", + "content": "3.3. Model Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 364, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 556, + 460 + ], + "type": "text", + "content": "We build our joint image restoration diffusion model based on a latent diffusion model [35]. Fig. 2 shows an overview of our method. The goal of our diffusion model is to denoise image latents such that a pre-trained decoder can convert them to high-quality images. When simultaneously denoising multiple image latents, it is important that latent representation remain consistent across views. We enable this by allowing the diffusion model to attend to latents from other views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 514, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 514, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 514, + 479 + ], + "type": "text", + "content": "3.3.1. Latent Diffusion Models for Restoration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 483, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 519 + ], + "type": "text", + "content": "A Latent Diffusion Model [35] architecture, such as Stable Diffusion 2.1 (SD2.1), consists of two parts a) encoder and decoder and b) denoising UNet." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 579 + ], + "type": "text", + "content": "During inference, our latent denoising UNet progressively refines pure noise into a clean latent representation, conditioned on the degraded image set, through several denoising steps. Once the latent representation is clean, a decoder translates it into the restored image set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 580, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 616 + ], + "type": "text", + "content": "Similarly, during training, our diffusion model is optimized to learn the denoising process of the latent representation, conditioned on the degraded input sets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "text", + "content": "Encoder-Decoder. We initialize our encoder with the weights from an SD2.1 [35] and are frozen when we train our model. The encoder, " + }, + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "text", + "content": ", is used to map our images to latent vectors which are used to learn the UNet-Style diffusion model. The decoder, " + }, + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 617, + 556, + 689 + ], + "type": "text", + "content": ", converts the denoised latents from the diffusion model to pixels in image space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "UNet. Our latent denoising UNet, " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": ", considers as input a noisy latent " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^k" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": ", encoded latents for" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21622" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 67, + 553, + 284 + ], + "blocks": [ + { + "bbox": [ + 58, + 67, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 58, + 67, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 58, + 67, + 553, + 284 + ], + "type": "image", + "image_path": "2043b64822d415bb507354c5896c9505e030d7cbf1cc20b43e01469320497cee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 286, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 555, + 309 + ], + "type": "text", + "content": "Figure 3. Qualitative Comparisons on Motion Deblurring and Super-Resolution. The left side is the Motion-Deblurring and the right side is the Super-Resolution results. We refer readers to supplementary for more visualizations. Zoom in for the best view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "text", + "content": "degraded images, " + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "inline_equation", + "content": "\\{I^c\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "text", + "content": " and the diffusion noise " + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "text", + "content": " to output an of estimate the noise, " + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\epsilon}}" + }, + { + "bbox": [ + 55, + 320, + 294, + 346 + ], + "type": "text", + "content": ", as shown in Eq. (1):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 352, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 352, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 118, + 352, + 295, + 369 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} = \\mathbf {s} _ {\\theta} (\\mathbf {x} ^ {k}, \\phi (\\{I ^ {c} \\} _ {i = 1} ^ {N}), k). \\qquad (1)", + "image_path": "b0fdae02c48db240e90be6dc153796af3ec8c1dd4398c374e97eaadea29beb0d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 374, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 399 + ], + "type": "text", + "content": "We supervise the outputs of our UNet to match the noise added during the forward diffusion process using Eq. 2:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 404, + 295, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 404, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 102, + 404, + 295, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon \\sim \\mathcal {N} (0, I), k \\sim \\mathcal {U} (T)} \\| \\epsilon - \\hat {\\epsilon} \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "d16fccdef74c1f25bffb58b5759721c6ab79d2f9da1d52878f16f0d4f5d97b73.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 427, + 270, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 270, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 270, + 439 + ], + "type": "text", + "content": "3.3.2. Adapting Diffusion Model to multiple views" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 441, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 441, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 295, + 562 + ], + "type": "text", + "content": "Starting with the image diffusion model, we expand a new dimension to represent image sets from different views. Most of the time, batch and views are merged into one dimension and split when needed. We encode the degraded image set with the same VAE encoder as the target noisy latents. Noisy latents and the condition latents are integrated by channel-wise concatenation. Then, due to channel number change, we replace the original convolution with a new zero-initialized convolution layer [65] to adapt the new number of channels input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 258, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 258, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 258, + 578 + ], + "type": "text", + "content": "3.3.3. Spatial-3D Integrated UNet Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "Typical latent diffusion models such as Stable Diffusion [35] do not consider denoise multiple images simultaneously and often generate individual images that lead to inconsistent image generations between multi-views. Our solution to remove this inconsistency is to denoise all the views jointly by allowing the UNet to look at the intermediate latent representations from different views. As shown in the right side of Fig. 2, our architecture is composed of a Spatial-3D ResNet and a 3D self-attention Transformer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Spatial-3D ResNet. Image diffusion model like SD2.1 only has a 2D convolution layer in the ResNet, which makes it" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "content": "hard to perform well in tasks that require 3D understanding. We want to enhance the 3D performance of the ResNet of the image diffusion model by introducing a 3D convolution layer to better capture geometry information as shown in Fig. 2. Training from raw will be unstable and time-consuming, but we find that the video diffusion model that is performing the image-to-video task already has a pre-trained 3D convolution layer. Understanding temporal similarity in the time dimension and understanding 3D similarity in the spatial dimension share similarity. By the experiment, we empirically find that merging two different weights into one is feasible and can lead to better performance. Hence, we initialize the 2D Convolution layer with image Diffusion Model weight (SD2.1) and the 3D convolution layer with video Diffusion weight (Stable Video Diffusion [4]). The output of each layer will be blended as follows: " + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{ResNet}} = \\sigma (\\alpha)\\times O_{2D} + \\sigma (1 - \\alpha)\\times O_{3D}" + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "content": " is the sigmoid function and " + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "content": " is usually taken as 0.5 following SVD [4]. " + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "inline_equation", + "content": "O_{2D}" + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "content": " is the output of 2D ResNet and " + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "inline_equation", + "content": "O_{3D}" + }, + { + "bbox": [ + 313, + 320, + 555, + 549 + ], + "type": "text", + "content": " is the output of 3D ResNet layer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "3D Self-Attention Transformer. Furthermore, we adopt the 3D multi-view consistent self-attention Transformer as shown in Fig. 2, similar to previous works [13, 41]. Our self-attention layer considers image latents from all the " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": " views and performs self-attention. After ResNet, the model patches the latent to create " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": " patch token per view, and in total, we have " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "N \\times p" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": " tokens that are part of the attention layer. Each token calculates the attention score to all " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "N \\times p" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": " tokens as shown on the right side of Fig. 2. This feature enables each token to be accessible to all conditional image sets given at the maximum level. Further, by this design, our model, in the inference stage, is scalable to a variable number of input frames with stable performance even though" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21623" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 118, + 501, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 67, + 555, + 111 + ], + "lines": [ + { + "bbox": [ + 55, + 67, + 555, + 111 + ], + "spans": [ + { + "bbox": [ + 55, + 67, + 555, + 111 + ], + "type": "text", + "content": "Table 1. Sparse-View Deblurring Comparisons. We show quantitative results on three datasets and evaluate metrics as in § 4.1. VConsis. (visual consistency) is multiplied by 100. Across all methods, SIR-Diff is the most visually consistent as compared to all the baselines. On other standard metrics, we also outperform baselines. SIR-Diff (1 Frame) is a single image version of our method that does not depend on multiple views. The best is highlighted, and the second is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 110, + 118, + 501, + 201 + ], + "lines": [ + { + "bbox": [ + 110, + 118, + 501, + 201 + ], + "spans": [ + { + "bbox": [ + 110, + 118, + 501, + 201 + ], + "type": "table", + "html": "
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodFID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓FID ↓LPIPS ↓VConsis.↓
PromptIR [31]81.280.2487.81150.340.53710.72166.30.35012.89
Restromer [64]49.720.2326.5273.470.3309.59144.80.42611.00
VRT [25]134.50.3717.67160.40.53011.13170.40.44012.07
SIR-Diff (1 Frame)81.580.2476.45109.60.3509.55137.90.35710.36
SIR-Diff40.090.1605.7567.310.2639.4385.670.2759.30
", + "image_path": "f5227dca3fa4b54d0b2639c28e2500d3399bf183043624efb0caee8607459bfa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 223, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 223, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 223, + 227 + ], + "type": "text", + "content": "we train with a handful of images per set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 228, + 297, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 297, + 348 + ], + "type": "text", + "content": "However, attention communication between views leads to heavy computes. To resolve this, we only inject this module into the low-resolution layers (similar to CAT3D [13]), where we also ignore the biggest resolution layer for the UNet Encoder and Decoder. Further, we completely take away the cross-attention module across all layers instead of putting empty string inputs as placeholders like previous works [20]. We found that both of these changes help faster convergence, less training time, and better numerical results under the same number of training iterations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 132, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 132, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 132, + 371 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "content": "We evaluate SIR-Diff on the task of sparse view motion deblurring and super-resolution in a zero-shot setting " + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "content": ", and discuss our deblurring results in " + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "inline_equation", + "content": "\\S 4.4" + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "content": " and super-resolution results in " + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "inline_equation", + "content": "\\S 4.5" + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "content": ". We discuss implementation details in " + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "content": ", and now we start by discussing evaluation metrics in " + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "inline_equation", + "content": "\\S 4.1" + }, + { + "bbox": [ + 55, + 377, + 297, + 450 + ], + "type": "text", + "content": ". Please refer to supplementary materials for the ablation study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 115, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 115, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 115, + 468 + ], + "type": "text", + "content": "4.1. Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 605 + ], + "type": "text", + "content": "We use several metrics to evaluate the performance of all the methods on the task of motion deblurring and superresolution. Since our approach is generative in nature we borrow metrics from the literature that are good at evaluating generative models such as Frechet Inception Distance (FID) [15], Learned Perceptual Image Patch Similarity (LPIPS) [66]. In addition to these, we evaluate our superresolution results on PSNR and SSIM similar to methods used in this literature [57, 62, 64, 68]. We also propose a new metric to evaluate view consistency between pairs of generated views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 664 + ], + "type": "text", + "content": "Visual Consistency (VConsis). This metric evaluates self-consistency within the generated view set by measuring the perceptual similarity of corresponding patches between the two views. Please refer to the supplementary for additional details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 713 + ], + "type": "text", + "content": "We implement our latent diffusion model based on pretrained UNet and auto-encoder from SDv2.1 as part of dif" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 556, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 556, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 556, + 312 + ], + "type": "text", + "content": "fusers. We train our models on two synthetic datasets Hypersim [34] and TartanAir [52] with image resolution resized to " + }, + { + "bbox": [ + 313, + 216, + 556, + 312 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 313, + 216, + 556, + 312 + ], + "type": "text", + "content": ". We randomly select the order of image sets to maintain permutation invariance in training and inference. We utilize ground truth geometry and pose data to pre-compute the overlap region between each pair of frames. We provide additional details on model training, dataset, and view selection in the supplemental." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 503, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 503, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 503, + 332 + ], + "type": "text", + "content": "4.3. Zero Shots Datasets For Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 338, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 494 + ], + "type": "text", + "content": "As all our baselines are training on different datasets, we perform a zero-shot evaluation across a diverse set of real-world datasets to keep a fair comparison. Zero-Shot means rather than split the dataset into training-evaluation splits, we chose some dataset for training and some other out-of-domain for evaluation. Our evaluation datasets consist of indoor and outdoor scenes along with object-centric datasets. ScanNet++ [61]. This is a dataset of 460 indoor scanned readl 3D scenes. We evaluate all the methods on the scenes from the official test set of 50 scenes, by sampling 200 image sets with four images per set. We highlight the zero-shot performance of our method trained on synthetic indoor data transfers well to real indoor scene data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 494, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 542 + ], + "type": "text", + "content": "ETH3D [38]. This dataset provides high-resolution images of indoor and outdoor scenes, with accurate depth maps captured via a LiDAR. We evaluate all the 454 samples from the ETH3D dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 543, + 556, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 556, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 556, + 614 + ], + "type": "text", + "content": "CO3Dv2 [32]. This is an object-centric dataset containing object-level videos across 51 categories of household objects. This dataset has pose annotations from COLMAP [37] provided for all video sequences. We use the official single-sequence subsets for evaluation and randomly sample 200 image sets among them." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "type": "text", + "content": "For all these three datasets, please refer to the supplementary material for the detail of view-selection strategy" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 647, + 455, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 455, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 455, + 661 + ], + "type": "text", + "content": "4.4. Motion Deblurring (MD)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Our motion deblurring (MD) method targets solving arbitrary motion paths in the wild. We target at restoring original images from the set of sparse view motion blur inputs. Our paired training dataset is prepared by applying the motion-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21624" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 109, + 545, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 67, + 555, + 101 + ], + "lines": [ + { + "bbox": [ + 55, + 67, + 555, + 101 + ], + "spans": [ + { + "bbox": [ + 55, + 67, + 555, + 101 + ], + "type": "text", + "content": "Table 2. Sparse-View Super-Resolution Comparisons. VConsis. (visual consistency) is multiplied by 100. SIR-Diff (1 Frame) is a single image process form like other image diffusion models. Ours, in the inference stage, can adapt an arbitrary number of views. The best is highlighted and the second is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 109, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 63, + 109, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 109, + 545, + 186 + ], + "type": "table", + "html": "
DatasetScannet++ [61]ETH3D [38]CO3D [32]
MethodPSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓PSNR ↑SSIM ↑LPIPS ↓VConsis.↓
SAM-DiffSR [48]24.130.2160.7586.9419.160.3170.4958.2823.900.2460.62610.96
OSEDiff [57]25.220.8080.1545.5121.930.6210.2648.6224.500.6580.2778.86
UAV [68]23.230.6770.3549.0120.390.5060.4139.2622.730.5660.40011.82
SIR-Diff (1 Frame)25.850.7920.1645.7021.970.5930.2217.9024.350.6410.3019.07
SIR-Diff27.380.8370.0995.0122.580.6210.2017.6024.530.6580.2107.85
", + "image_path": "9932788acb7ca3325ade09f3cf484f309a89d166055a3d4339cc3ad4c6b9230c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 201, + 294, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 201, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 294, + 249 + ], + "type": "text", + "content": "blurring kernel on the clean images from the training datasets based on the codebase inspired by DPS [7]. The blur kernel size is chosen from a normal distribution with mean 85px and std 12.75px." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 250, + 295, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 295, + 320 + ], + "type": "text", + "content": "Further, the intensity of the blur is randomly sampled in the range " + }, + { + "bbox": [ + 55, + 250, + 295, + 320 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 55, + 250, + 295, + 320 + ], + "type": "text", + "content": ". The blur kernel for each image in the set is sampled independently. For evaluation, we randomly sample 4 images per view-set from the pre-computed image set and require all the methods to remove motion blurring from them." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 329, + 124, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 124, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 124, + 340 + ], + "type": "text", + "content": "4.4.1. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 295, + 368 + ], + "type": "text", + "content": "We compare against several baselines that can restore motion blurring from single-image or videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 428 + ], + "type": "text", + "content": "Restormer [64]. This is a single-image motion deblurring method that does not have the benefit of jointly deblurring across a sparse view image collection. Our method outperforms this on single image deblurring and sparse view deblurring across all metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 429, + 296, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 476 + ], + "type": "text", + "content": "PromptIR [31]. This is all-in-one restoration method trained to deblur. We use their pre-trained model and compare the performance of their method on sparse view deblurring." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "type": "text", + "content": "VRT [25]. This is a video-oriented restoration Transformer model trained with deblurring capability. Our method outperforms this on sparse view deblurring across all metrics. Fair comparisons on video datasets are in the supplemental." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 115, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 115, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 115, + 544 + ], + "type": "text", + "content": "4.4.2. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 548, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 571 + ], + "type": "text", + "content": "We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 573, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 644 + ], + "type": "text", + "content": "Qualitative. As shown in the left side of Fig. 3, our SIR-Diff show rich details and sharp deblurring restoration quality than other methods on both object level [32] and scene level [61]. This is thanks to our architectural design information from different views can be aggregated and constructive to help the other view in attention and residual layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 645, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 668 + ], + "type": "text", + "content": "Quantitative. We assess our results using the metrics described in § 4.2, with outcomes summarized in Tab. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "content": "Our method outperforms all the methods on the three datasets on FID and LPIPS metrics. Restormer [64], a single-image" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 201, + 553, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 237 + ], + "type": "text", + "content": "method, lacks the ability to leverage multi-view information to enhance output quality. In contrast, our method uses 3D self-attention to achieve superior results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 237, + 555, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 309 + ], + "type": "text", + "content": "We also outperform the video motion deblurring method, VRT, as they are incapable of handling sparse views well by motion estimation and require continuous and smooth video-captured views to fuse frame features. In the supplement, we provide a direct comparison with video methods quantitatively on continuous videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 316, + 443, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 443, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 443, + 330 + ], + "type": "text", + "content": "4.5. Super-Resolution (SR)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 335, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 406 + ], + "type": "text", + "content": "We evaluate our method on the 4x sparse-view super-resolution task. Our paired training dataset is prepared by first downsampling the high-resolution image with a scale factor of 4 and then upsampling it back to the original resolution to introduce a low-resolution effect. The resize algorithm we use is bicubic." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 413, + 382, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 382, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 382, + 423 + ], + "type": "text", + "content": "4.5.1. Baselines" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 428, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 452 + ], + "type": "text", + "content": "We compare against several diffusion model baselines that handle single-image and video super-resolution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 452, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 499 + ], + "type": "text", + "content": "SAM-DIFFSR [48]. This is an image diffusion model-based methodology, which targets classical super-resolution tasks like ours. Our method outperforms this on sparse view restoration on all the metrics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 500, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 535 + ], + "type": "text", + "content": "OSEDiff [57]. This is a single-image diffusion model-based super-resolution model. OSEDiff reaches SOTA performance on the 2D image super-resolution." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 536, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 608 + ], + "type": "text", + "content": "Upscale-A-Video [68]. This is a video diffusion model-based super-resolution method. We input it with a sequential video frame (in total 4 for each scenario) but a larger motion gap compared to ordinary videos. We also provide a fair comparison between Upscale-A-Video and our method on video datasets in the supplement." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 614, + 373, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 373, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 373, + 625 + ], + "type": "text", + "content": "4.5.2. Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 629, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 653 + ], + "type": "text", + "content": "We evaluate the performance of baselines and our method quantitatively and qualitatively on three datasets." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 654, + 541, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 541, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 541, + 666 + ], + "type": "text", + "content": "Qualitative. As the red rectangular box of Fig. 3 shows," + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": "for shadow artifacts and severe aliasing effects (see the second example), Upscale-A-Video [68] fails and OSEDiff [57] restoration is unfaithful and inconsistent for different views compared to ours. Moreover, our model can restore" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 220, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 220, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 220, + 713 + ], + "type": "text", + "content": "1 https://github.com/LeviBorodenko/motionblur" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21625" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 182, + 294, + 250 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 298, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 298, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 298, + 180 + ], + "type": "text", + "content": "Table 3. Gaussian Splitting Reconstruction from Blurry Images. Under serious motion blurring, COLMAP[37] makes it hard to solve correct camera pose and SfM points, which leads to bad results on both vanilla 3DGS and BAD-GS. With the help of SIR-Diff, we can achieve restore better results and be robust to difficult motion blurring effects. KS refers to kernel size with normal distribution " + }, + { + "bbox": [ + 55, + 70, + 298, + 180 + ], + "type": "inline_equation", + "content": "[\\mu, \\sigma]" + }, + { + "bbox": [ + 55, + 70, + 298, + 180 + ], + "type": "text", + "content": ". Inten. refers to the blurring intensity range that is sampled uniformly. We randomly select 3 objects in CO3Dv2 [32] dataset and apply motion blur on clean images to conduct this experiment." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 182, + 294, + 250 + ], + "lines": [ + { + "bbox": [ + 61, + 182, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 294, + 250 + ], + "type": "table", + "html": "
DifficultyMediumHard
ParamsKS:[30, 10.2], Inten:[0, 0.4]KS:[45,14.85], Inten:[0, 0.5]
MethodPSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
3DGS [21]13.470.6100.60710.420.4210.801
BAD-GS [67]10.060.5370.6878.0610.3850.948
Resotmer [64]+BAD-GS [67]25.770.6160.35623.880.5600.317
SIR-Diff + BAD-NeRF [49]23.600.5030.45222.470.4390.485
SIR-Diff + BAD-GS26.110.6610.25025.330.6440.277
", + "image_path": "747ad6d8fd3bd9d99840a0d5d22e8f03340e8866fecc2ce43c679ac15942a5df.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "type": "text", + "content": "images with better geometry alignment as shown in the first example." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 295, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 296, + 438 + ], + "type": "text", + "content": "Quantitative. We use PSNR, SSIM, LPIPS [66] and our proposed visual consistency § 4.1 metrics. The result is shown in Tab. 2 to show that SIR-Diff outperforms all the existing baselines on all datasets. Compared to the image diffusion method, OSEDiff[57], our spatial-3D Resnet and 3D self-attention Transformer help each latent token to access information from all views which leads to successful multiview restoration. Upscale-A-Video also employs a similar 3D ResNet architecture and cross-image attention mechanism; however, our model does not rely on conditioning with optical flow, which is challenging to obtain accurately in sparse views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 204, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 204, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 204, + 459 + ], + "type": "text", + "content": "4.6. Down Stream Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 297, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 297, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 297, + 608 + ], + "type": "text", + "content": "We show several downstream applications of sparse multiview deblurring and super-resolution. Multi-view sparse images are used for tasks of finding correspondences, pose estimation, and 3D reconstruction. The performance of these methods is often dependent on the input image quality. If the underlying image quality is degraded, the output of these methods is poor. Blindly restoring these images with off-the-shelf single-image methods leads to sub-optimal results because restoration from these methods lacks view consistency. We show that, with SIR-Diff, we can consistently restore image quality and help alleviate the performance gap on tasks requiring multiview sparse views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 615, + 206, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 206, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 206, + 627 + ], + "type": "text", + "content": "4.6.1. Estimating Correspondences" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Under severely degraded inputs it is hard for methods like LoFTR [43] to estimate correspondences. As shown in Fig. 4, (a) we show qualitatively at low-resolution LoFTR struggles to find correct correspondences. Moreover, since single-view image restoration models lack multi-view consistency this does not solve the issue. SIR-Diff can not only restore the image quality but since it is view consistent can improve the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 352, + 129, + 515, + 183 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 556, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 556, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 556, + 125 + ], + "type": "text", + "content": "Table 4. Feature Matching Correspondence. We compute the correspondences using off-the-shelf LoFTR [43] on the restored images with single-view methods and SIR-Diff. Both in SR and deblurring cases, our method can enable LoFTR to find more correspondences (# of correspondences found)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 129, + 515, + 183 + ], + "lines": [ + { + "bbox": [ + 352, + 129, + 515, + 183 + ], + "spans": [ + { + "bbox": [ + 352, + 129, + 515, + 183 + ], + "type": "table", + "html": "
DeblurringSR
GT Images1979.51979.5
Degraded Images640.11374.0
Single-View Methods [57, 64]1262.61478.0
SIR-Diff1621.31494.2
", + "image_path": "4374ad5459e1d05392054586fb5127e0b1443e5561f955c801586ab2a1202b4a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 190, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 251 + ], + "type": "text", + "content": "correspondences for off-the-shelf methods like LoFTR [43]. Quantitatively, as shown in Tab. 4, we also see that LoFTR can find over 380 additional correspondences as compared to the best image-based method on deblurring and on average 15 for super-resolution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 257, + 512, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 512, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 512, + 269 + ], + "type": "text", + "content": "4.6.2. 3D Reconstruction on Degraded Images" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 271, + 556, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 556, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 556, + 403 + ], + "type": "text", + "content": "Methods such as BAD-Gaussians [67] (BAD-GS) aim to perform 3D reconstruction from motion-blurry images, but they depend on running COLMAP [37] on blurry images to get pose initialization and sparse 3D geometry. We show in Tab. 3 that with increasing blurring effect, BAD-GS performance drops significantly while using SIR-Diff with BAD-GS leads to a much steady performance. In Fig. 5, we show a hard case of motion blurring that leads to failure of BAD-GS [67] which can be solved if we restore the image using SIR-Diff first and then do BAD-GS. We refer to supplementary for details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 410, + 474, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 474, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 474, + 422 + ], + "type": "text", + "content": "4.6.3. Sparse-View 3D reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": "InstantSplat [11] represents a leading approach for sparse view Gaussian Splittings (GS). However, reconstructing a scene under sparse viewpoints with low-quality observations remains challenging. In Fig 6, we show an example of applying InstantSplat on an image restored using one of the baselines compared to restoration using SIR-Diff. In Tab. 5, we report the average performance of all baselines on the selected 3 scenes from Scannet++[61] and CO3Dv2[32] and build a 3DGS using InstantSplat [11]. We evaluate the GS with standard Novel View Synthesis metrics. Additional details can be found in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 564, + 499, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 499, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 499, + 577 + ], + "type": "text", + "content": "4.7. Depth Prediction for Sparse Views" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 654 + ], + "type": "text", + "content": "We also apply our model architecture to the task of improving sparse multiview depth estimation. The only thing we edit is changing the target latent to depth image sets and the conditional latent to RGB image sets. Following Marigold [20], we use the ground truth depth label provided by the dataset for supervision signal during training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Metrics. We evaluate the performance of methods on three metrics AbsRel and " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": " as defined by Silberman et al. [42]. In the case of MDE methods, we additionally align all the predicted depth maps to find the most optimum scale and bias to make them consistent. Similar to the view consistency" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21626" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 553, + 170 + ], + "type": "image", + "image_path": "b5124e623cf3a17a2f23c804cbf7ad335319e89e482c8254fec506115b8b6760.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "text", + "content": "Figure 4. Feature Matching Downstream Application. Low-Resolution and Single-image-based SR models like OSEDiff [57] find less correspondence points than our SIR-Diff that can restore multiview images consistently. SIR-Diff recovers the highest number of correspondences (310) as compared to other baseline restoration methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 220, + 293, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 220, + 293, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 293, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 293, + 370 + ], + "type": "image", + "image_path": "2f68868daecdb2fc2a1a8b22c84d9bbdee8875d10d5644a83587142eb6b8bd40.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 296, + 424 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 424 + ], + "type": "text", + "content": "Figure 5. BAD-GS [67] with SIR-Diff. We show the effect of using SIR-Diff to help BAD-GS recover from catastrophic failure. BAD-GS (Top) output with a very high LPIPS" + }, + { + "bbox": [ + 55, + 369, + 296, + 424 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 369, + 296, + 424 + ], + "type": "text", + "content": " of 0.59 vs. 0.22 BAD-GS output with image-restoration using SIR-Diff (Bottom) leads to a significantly higher-quality 3D reconstruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 433, + 294, + 572 + ], + "blocks": [ + { + "bbox": [ + 58, + 433, + 294, + 572 + ], + "lines": [ + { + "bbox": [ + 58, + 433, + 294, + 572 + ], + "spans": [ + { + "bbox": [ + 58, + 433, + 294, + 572 + ], + "type": "image", + "image_path": "83122bcf17bc8a428d4312fbb1b1dd32fe8d47c46a6530116a18c93becc64e1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 578, + 297, + 624 + ], + "lines": [ + { + "bbox": [ + 55, + 578, + 297, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 297, + 624 + ], + "type": "text", + "content": "Figure 6. InstantSplat Downstream Application. 3D Reconstruction is done by InstantSplat [11]. SIR-Diff that can restore multiview images consistently leads to better visual and numerical reconstruction results than other single-image methods, Restormer [64]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 689 + ], + "type": "text", + "content": "metrics for image restoration, We evaluate Geometric Consistency (GConsis.) by calculating the distances between corresponding points between two different views under the camera coordinate system. Additional details on this metric are in the supplemental." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Baselines. We evaluate against several monocular depth estimation (MDE) methods such as DepthAnythingV2 [60]," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 276, + 552, + 376 + ], + "blocks": [ + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 273 + ], + "type": "text", + "content": "Table 5. Quantitative Results of Super-Resolution and Motion-Deblurring Reconstruction on InstantSplat [11]. The result is an average of 3 scenes per dataset (individual results in supp). Using SIR-Diff results in higher quality inputs to InstantSplat compared to OSEDiff [57] and Restormer [64]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 276, + 552, + 376 + ], + "lines": [ + { + "bbox": [ + 319, + 276, + 552, + 376 + ], + "spans": [ + { + "bbox": [ + 319, + 276, + 552, + 376 + ], + "type": "table", + "html": "
Scannnet++ [61]CO3D [32]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GT Image31.860.9400.09623.940.7570.243
Super-Resolution
OSEDiff [57]24.530.7990.27420.250.5490.379
SIR-Diff26.200.8270.23321.780.5780.357
Motion-Deblurring
Restormer [64]22.020.7890.30920.140.5790.465
SIR-Diff26.530.8420.23121.600.6140.384
", + "image_path": "d87511192a8a777eec72f2075a1e926a99329fabdd71d78c5b30218104eeea4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 421, + 552, + 479 + ], + "blocks": [ + { + "bbox": [ + 313, + 387, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 419 + ], + "type": "text", + "content": "Table 6. Quantitative comparisons on image and video depth models. The best is highlighted and the second is underlined. GConsis. refers to geometric consistency and the unit is meter." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 421, + 552, + 479 + ], + "lines": [ + { + "bbox": [ + 319, + 421, + 552, + 479 + ], + "spans": [ + { + "bbox": [ + 319, + 421, + 552, + 479 + ], + "type": "table", + "html": "
DatasetScannet++ [61]ETH3D [38]
MethodAbsRel ↓δ1 ↑GConsis.↓AbsRel ↓δ1 ↑GConsis.↓
DepthAnyV2 [60]9.27093.350.17713.2086.521.932
DepthCrafter [18]21.6589.960.7516.66295.051.701
Marigold [20]11.1089.920.1627.58194.460.410
SIR-Diff9.01193.540.1296.66395.050.339
", + "image_path": "29e83495825a31fe5ca461774523a8148af240a0c060f3e0e4f376a148371548.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 470, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 470, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 470, + 502 + ], + "type": "text", + "content": "DepthCrafter [18], and Marigold [20]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "content": "Results. As shown in Tab. 6, our multi-view depth model is competitive with single image depth methods on the AbsRel, and " + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 502, + 555, + 574 + ], + "type": "text", + "content": " metrics are sometimes better. Meanwhile, our model outperforms these methods on the geometric consistency metric that evaluates the 3D consistency when predicting depth jointly from sparse views." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 585, + 462, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 462, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 462, + 597 + ], + "type": "text", + "content": "5. Conclusion & Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "We present SIR-Diff, a simple and effective approach for multi-view image restoration, addressing blurring and low-resolution degradation across multiple images simultaneously. SIR-Diff outperforms leading single-view methods in sparse-view restoration and enables 3D-consistent image restoration for 3D computer vision applications. Future work will explore efficient cross-image attention mechanisms, such as selective patch attention along epipolar lines, and develop explicit guarantees for 3D view consistency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21627" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 163 + ], + "type": "text", + "content": "This work partially used NCSA Delta GPU at the University of Illinois through allocation CIS230313 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program, which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 115, + 189 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 196, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 196, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 295, + 240 + ], + "type": "text", + "content": "[1] Miika Aittala and Frédo Durand. Burst image deblurring using permutation invariant convolutional neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 731-747, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 242, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 242, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 296, + 317 + ], + "type": "text", + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 320, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 296, + 374 + ], + "type": "text", + "content": "[3] Sherwin Bahmani, Xian Liu, Wang Yifan, Ivan Skorokhodov, Victor Rong, Ziwei Liu, Xihui Liu, Jeong Joon Park, Sergey Tulyakov, Gordon Wetzstein, et al. Tc4d: Trajectory-conditioned text-to-4d generation. In European Conference on Computer Vision, pages 53-72. Springer, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 432 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 498 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 566 + ], + "type": "text", + "content": "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4217-4229, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 568, + 296, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 611 + ], + "type": "text", + "content": "[7] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 657 + ], + "type": "text", + "content": "[8] Chao Dong, Yubin Deng, Chen Change Loy, and Xiaou Tang. Compression artifacts reduction by a deep convolutional network. In Proceedings of the IEEE international conference on computer vision, pages 576-584, 2015. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "text", + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[10] Linwei Fan, Fan Zhang, Hui Fan, and Caiming Zhang. Brief review of image denoising techniques. Visual Computing for Industry, Biomedicine, and Art, 2(1):7, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "type": "text", + "content": "[11] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 205 + ], + "type": "text", + "content": "[12] Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jianping Fan, Yu Jun, et al. Srgs: Super-resolution 3d gaussian splatting. arXiv preprint arXiv:2404.10318, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 259 + ], + "type": "text", + "content": "[13] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "type": "text", + "content": "[14] Shi Guo, Xi Yang, Jianqi Ma, Gaofeng Ren, and Lei Zhang. A differentiable two-stage alignment scheme for burst image reconstruction with large shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17472-17481, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 317, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 555, + 369 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 404 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 405, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 555, + 459 + ], + "type": "text", + "content": "[17] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 503 + ], + "type": "text", + "content": "[18] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 555, + 557 + ], + "type": "text", + "content": "[19] Michal Kawulok, Pawel Benecki, Szymon Piechaczek, Krzysztof Hrynczenko, Daniel Kostrzewa, and Jakub Nalepa. Deep learning for multiple-image super-resolution. IEEE Geoscience and Remote Sensing Letters, 17(6):1062-1066, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 559, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 624 + ], + "type": "text", + "content": "[20] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 5, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "text", + "content": "[21] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[22] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21628" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 148 + ], + "type": "text", + "content": "[23] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 205 + ], + "type": "text", + "content": "[24] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung-Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "text", + "content": "[25] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 308 + ], + "type": "text", + "content": "[26] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 374 + ], + "type": "text", + "content": "[27] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 418 + ], + "type": "text", + "content": "[28] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Taming diffusion models for image restoration: A review. arXiv preprint arXiv:2409.10353, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 475 + ], + "type": "text", + "content": "[29] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 531 + ], + "type": "text", + "content": "[30] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 578 + ], + "type": "text", + "content": "[31] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 645 + ], + "type": "text", + "content": "[32] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 679 + ], + "type": "text", + "content": "[33] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[34] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 108, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 163 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 164, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 229 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 231, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 555, + 274 + ], + "type": "text", + "content": "[37] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "text", + "content": "[38] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 322, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 555, + 365 + ], + "type": "text", + "content": "[39] Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Frustück. Supergaussian: Repurposing video models for 3d super resolution. arXiv preprint arXiv:2406.00609, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 367, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 555, + 420 + ], + "type": "text", + "content": "[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 555, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 456 + ], + "type": "text", + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 457, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 555, + 521 + ], + "type": "text", + "content": "[42] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 524, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 555, + 578 + ], + "type": "text", + "content": "[43] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 580, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 645 + ], + "type": "text", + "content": "[44] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "text", + "content": "[45] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21629" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[46] Boyang Wang, Bowen Liu, Shiyu Liu, and Fengyu Yang. Vciser: Blind single image super-resolution with video compression synthetic data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4302-4312, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[47] Boyang Wang, Fengyu Yang, Xihang Yu, Chao Zhang, and Hanbin Zhao. Apisr: anime production inspired real-world anime super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25574-25584, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 227 + ], + "type": "text", + "content": "[48] Chengcheng Wang, Zhiwei Hao, Yehui Tang, Jianyuan Guo, Yujie Yang, Kai Han, and Yunhe Wang. Sam-diffsr: Structure-modulated diffusion model for image super-resolution. arXiv preprint arXiv:2402.17133, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "text", + "content": "[49] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 316 + ], + "type": "text", + "content": "[50] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Badnerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170-4179, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "text", + "content": "[51] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 425 + ], + "type": "text", + "content": "[52] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 480 + ], + "type": "text", + "content": "[53] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "text", + "content": "[54] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "text", + "content": "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 296, + 624 + ], + "type": "text", + "content": "[56] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": "[57] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[58] Hanyu Xiang, Qin Zou, Muhammad Ali Nawaz, Xianfeng Huang, Fan Zhang, and Hongkai Yu. Deep learning for image inpainting: A survey. Pattern Recognition, 134:109046, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 564 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "text", + "content": "[59] Cheng Yang, Lijing Liang, and Zhixun Su. Real-world denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023.2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "type": "text", + "content": "[60] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": "[61] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 262 + ], + "type": "text", + "content": "[62] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 555, + 306 + ], + "type": "text", + "content": "[63] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 555, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 373 + ], + "type": "text", + "content": "[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 375, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 555, + 419 + ], + "type": "text", + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 420, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 555, + 474 + ], + "type": "text", + "content": "[66] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 555, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 555, + 508 + ], + "type": "text", + "content": "[67] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 2, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 509, + 555, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 555, + 564 + ], + "type": "text", + "content": "[68] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21630" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_content_list.json b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc683ea5297be366708f282e24e49e9340eb0902 --- /dev/null +++ b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_content_list.json @@ -0,0 +1,2029 @@ +[ + { + "type": "text", + "text": "SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons", + "text_level": 1, + "bbox": [ + 89, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanyou Xu $^{1}$ , Zongxin Yang $^{2}$ , Yi Yang $^{1*}$ $^{1}$ ReLER, CCAI, Zhejiang University \n $^{2}$ DBMI, HMS, Harvard University", + "bbox": [ + 336, + 179, + 666, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yoxu, yangyics}@zju.edu.cn, zongxin-yang@hms.harvard.edu", + "bbox": [ + 250, + 236, + 741, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Controllable generation has achieved substantial progress in both 2D and 3D domains, yet current conditional generation methods still face limitations in describing detailed shape structures. Skeletons can effectively represent and describe object anatomy and pose. Unfortunately, past studies are often limited to human skeletons. In this work, we generalize skeletal conditioned generation to arbitrary structures. First, we design a reliable mesh skeletonization pipeline to generate a large-scale mesh-skeleton paired dataset. Based on the dataset, a multi-view and 3D generation pipeline is built. We propose to represent 3D skeletons by Coordinate Color Encoding as 2D conditional images. A Skeletal Correlation Module is designed to extract global skeletal features for condition injection. After multi-view images are generated, 3D assets can be obtained by incorporating a large reconstruction model, followed by a UV texture refinement stage. As a result, our method achieves instant generation of multi-view and 3D contents that are aligned with given skeletons. The proposed techniques largely improve the object-skeleton alignment and generation quality. Project page at https://skdream3d.github.io/.", + "bbox": [ + 89, + 318, + 485, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 681, + 220, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In view of visual representation dimension [64, 72], 2D image generation [16, 53, 57], multi-view (2.5D) generation [35, 55], and 3D generation [17, 27, 69] have advanced successively and achieved significant progress. Given the success of large language models [23, 26, 43, 48-50], textual representation has been widely applied in generation [51-53], as in other domains [13, 14, 73]. To achieve more controllable generation, conditions beyond text have attracted considerable attention. 2D image conditions (e.g., edge maps, human skeletons, and concept references) [54, 75] have been well studied. Similarly in 3D generation,", + "bbox": [ + 89, + 708, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2D [30] and 3D [10] conditions have also been explored.", + "bbox": [ + 511, + 287, + 888, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although the aforementioned conditions in controllable generation complement text descriptions, they still struggle to precisely describe shape structures. In contrast, skeletons, among various types of conditions, exhibit superior ability to depict shape structures: (i) Representation of object anatomy. Skeletons can efficiently represent various 3D structures with sparse joints and bones, while other conditions like depth maps fail to represent full 3D anatomy. (ii) Articulation into different poses. Skeletons are widely used for character animation in computer graphics [3, 22] due to their simplicity and efficiency. Other conditions such as rough shapes [10] are inconvenient for pose articulation. (iii) Freedom of editing. Given an initial skeleton, users can freely add new structures or modify joints and bones to create ideal shapes. Examples are in Fig. 1.", + "bbox": [ + 509, + 304, + 906, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advantages, previous studies [19, 20, 40, 75, 77] on conditional generation are limited to human skeletons. From the perspective of generalization, we would like to ask: Is it possible to use arbitrary skeletons as conditions to generate any creatures or even general objects?", + "bbox": [ + 511, + 532, + 906, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve this goal, we identify two main challenges that hinder the use of arbitrary skeletal conditions for generation: (i) Lack of large-scale object-skeleton pairs for training. Extensive studies [4, 15, 37] on 2D/3D human pose estimation have made human-skeleton paired data readily available. However, when skeletal structures are unknown, estimating arbitrary skeletons from 2D images or videos becomes challenging due to its ill-posedness. (ii) Insufficiency of 2D information to describe arbitrary skeletons. Human skeletons are relatively simple and can be described by a fixed set of 2D joints. However, complex skeletons suffer from self-occlusion and ambiguity, necessitating 3D information to fully capture their anatomy and pose.", + "bbox": [ + 511, + 611, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we focus on multi-view and 3D generation with skeletal conditions. To tackle data scarcity problem, we construct a large-scale dataset Objaverse-SK containing mesh-skeleton pairs. Textured meshes are selected from Objaverse [8] by semantic classes to form a subset. In order to achieve reliable mesh ske", + "bbox": [ + 511, + 810, + 908, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "314", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0afbb1f749cc0cbdc98dd5fe1dd01e3833dcfac37e9e2d24fecbc9d1b4cea17e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 104, + 233, + 150 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/25c410ee255a8d99ceb9eb74a7b4357fa1c678b5f5a29bd5d824ce0d3be3c97f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 107, + 299, + 154 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a00d0e26d78e1918eeb6ea317f26a1c7200b4a26fae4c1b3af86d0b38a52d567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 107, + 357, + 143 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7efd391b91e4f5b2172481bd2da9d2579f92f7892611ffd2c56edb7546ce6761.jpg", + "image_caption": [ + "MVDream", + "sno" + ], + "image_footnote": [], + "bbox": [ + 132, + 175, + 179, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08ef72300a46dfc3d905c8b346e18f2cc6cadd4170e55d58521c9d6fc21dd0bf.jpg", + "image_caption": [ + "Text & Skeleton", + "Dog" + ], + "image_footnote": [], + "bbox": [ + 189, + 172, + 238, + 217 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d1dd94c0d63234511a5c7dae5bc680c9bf735fd5334459cfd674d3bb3426607.jpg", + "image_caption": [ + "Chair" + ], + "image_footnote": [], + "bbox": [ + 243, + 176, + 300, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/85cef554b554ce5b10bd75589395a08b6ae6445dd6af76ba88f7aa33fec4903c.jpg", + "image_caption": [ + "Fish" + ], + "image_footnote": [], + "bbox": [ + 308, + 165, + 359, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a2c1116ee6165bb09ef7a2a314dd32a625e9e2e1aa53806af3b2ffa7cc0242a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 160, + 436, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fac99c2eda3a17b8cea6a79ee43f229656a38d3154cb3cfd59382e6b6d3bc1a7.jpg", + "image_caption": [ + "Text Only" + ], + "image_footnote": [], + "bbox": [ + 444, + 97, + 498, + 156 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/86f54caa4684bef521224cc22a98f6936016b5f5e5c0b2ec223b56c19b84c546.jpg", + "image_caption": [ + "Spiderman", + "(a) Text and skeletons for Collaborative Appearance and Shape Controlling" + ], + "image_footnote": [], + "bbox": [ + 442, + 159, + 498, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97fb40a7cb2d5e27595631d7ba02d4ed43bccb98fe3cb97d6d7f2a0198c7c4a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 98, + 550, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3d04219a1222971b64d7ed2afd444c0aa9534bc273a276fb105214ee73956ed6.jpg", + "image_caption": [ + "Rocket" + ], + "image_footnote": [], + "bbox": [ + 509, + 159, + 560, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/103aa145f1228f6b48ff449e647403c6553b8aed7795fddfe1e269a341cd390c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 98, + 612, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a0877861603c8471d23e49169038a25b0b54c62afb407571e603d8d26e468888.jpg", + "image_caption": [ + "Ice cream" + ], + "image_footnote": [], + "bbox": [ + 563, + 159, + 627, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 647, + 160, + 692, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5d902562c5f9e00063d794b49712c15baaafb81b11a3fe05b91557b84c2e7699.jpg", + "image_caption": [ + "Text Only" + ], + "image_footnote": [], + "bbox": [ + 702, + 107, + 764, + 143 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c9578f61a2a517c76c5eb64f5a610b9c0b21cb263b58b9d58eff42b6b04ba4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 104, + 834, + 145 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a1483b0375fff0ee8bb5469c4d2b1f3fc21183cf9c480627a00e0720e85508e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 846, + 98, + 874, + 150 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/486e1a8da97c2665e6b861e300ef360f11ba289943612568a23057521eb15dea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 159, + 692, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b15b2f8bda631ef6a647b4608b255c9cc83f9d485b9ff6e4e2dfbdb670e8e909.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 157, + 764, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7578f5e542481c324c0f22cafcb787af1833f34491cb74657c9d4465181b6ab9.jpg", + "image_caption": [ + "Coral", + "Torch" + ], + "image_footnote": [], + "bbox": [ + 759, + 159, + 823, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9aeee3ae7bcf98ba3fa96ba6ff877d63c8e3c724c4375312534f199325100a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 838, + 162, + 885, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/32047747d873f67b55ee937c9d8c97392b1ec19707c96567036be54d839fbd4c.jpg", + "image_caption": [ + "Animal" + ], + "image_footnote": [], + "bbox": [ + 114, + 292, + 161, + 328 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cb0b3004e56923e389338ab80c6d5776f6da88f07a713d3a4dcdece2ccbcde25.jpg", + "image_caption": [ + "Human" + ], + "image_footnote": [], + "bbox": [ + 109, + 354, + 161, + 415 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8ae4fdc0d04a155b5e87d6ea27e595175b36a569f9eddd1f644cecfcdf54a513.jpg", + "image_caption": [ + "Plant" + ], + "image_footnote": [], + "bbox": [ + 109, + 435, + 158, + 491 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/12b88a4c2585954215fd56756cce4d6fcc163aea5cb7ebc6b6874f5e2672a63b.jpg", + "image_caption": [ + "Initial skeleton", + "Adding joints, bones", + "Dog with a spiral tail" + ], + "image_footnote": [], + "bbox": [ + 181, + 277, + 295, + 328 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b797d4f5e3878ba89cb41acbf3da3b0e5d15fff47d67dd1eaf5013a4fbbe9afa.jpg", + "image_caption": [ + "Dog with two heads" + ], + "image_footnote": [], + "bbox": [ + 300, + 282, + 416, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/48309fdfc7f776ab20d0759186ac394e4e2bd1d2e077df7c543ec422f9ca8bbb.jpg", + "image_caption": [ + "Knight with dragon wings" + ], + "image_footnote": [], + "bbox": [ + 179, + 354, + 295, + 417 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b502d02c4974d33a9e46cde3bcb701ec60ed8f0d20b313d7b7713b8322c24696.jpg", + "image_caption": [ + "Plant with leaves and roots" + ], + "image_footnote": [], + "bbox": [ + 191, + 431, + 292, + 491 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/afd9e106bda4954b4111f6d1cf65f20c5391384583a2604b90386aa697e11d50.jpg", + "image_caption": [ + "Monster with four arms" + ], + "image_footnote": [], + "bbox": [ + 310, + 354, + 431, + 417 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/58744925b1b2268c9169ee3c10ba29fbe46a2b969d6767255efdfdbad4078f1d.jpg", + "image_caption": [ + "Plant with legs, Pokémon" + ], + "image_footnote": [], + "bbox": [ + 316, + 431, + 410, + 492 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6232048d19425b99adeac4a279bbef9c419e3322346843b3c53fdcf3a5c204cc.jpg", + "image_caption": [ + "Modifying joint positions, bone sizes", + "Dog with a big head" + ], + "image_footnote": [], + "bbox": [ + 431, + 276, + 544, + 328 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f50b45ce4bc9454e0e5d853d778f0ecdbfd70edd5af9ba1b852ee74c0ffaad62.jpg", + "image_caption": [ + "Dog with a long torso" + ], + "image_footnote": [], + "bbox": [ + 552, + 276, + 661, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62e188a77cdc1da2f3b9b2df854c7b7819c038dffcbd491fddc0b05b87780fb0.jpg", + "image_caption": [ + "Dwarf in armor" + ], + "image_footnote": [], + "bbox": [ + 436, + 351, + 550, + 412 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f026946b1d9f0cd73d3c3fa3960efcc9e70e1243bd0333d162fad5a442ac6630.jpg", + "image_caption": [ + "How wide?", + "Wide shoulder bodybuilder" + ], + "image_footnote": [], + "bbox": [ + 557, + 351, + 663, + 416 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae20cba4474432e47284557cfd31e28d00f3526ab8759079a995a780fa0bb06e.jpg", + "image_caption": [ + "Tree with wide branches" + ], + "image_footnote": [], + "bbox": [ + 436, + 435, + 549, + 487 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9e78a8b57ff69b3e7b45b36dbbc3c051e76265c5c82a92e0ffb035c7f468487.jpg", + "image_caption": [ + "Tree with a long truck", + "(b) Skeleton-based editing for Accurate Anatomy and Pose Controlling" + ], + "image_footnote": [], + "bbox": [ + 560, + 431, + 604, + 493 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50faf0ff8fbf485697abc3de93af8a61a39aa191b84f1f103b677761eea1f0a7.jpg", + "image_caption": [ + "Posing", + "Dog jumping" + ], + "image_footnote": [], + "bbox": [ + 681, + 277, + 776, + 327 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/81a0f19445b6a834f306eb385cbe4c924f19a1c1d5cde68f4e9cd0a6223fb843.jpg", + "image_caption": [ + "Dog sitting" + ], + "image_footnote": [], + "bbox": [ + 787, + 280, + 877, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1da94bcb445dc3678fceafce0d48bcd1097608031a535a16b3b52cddd7a80469.jpg", + "image_caption": [ + "Athlete running" + ], + "image_footnote": [], + "bbox": [ + 679, + 353, + 776, + 414 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/48fa0c23ae145450701724393702415f99df007b16c7c30f97749fb8dfdc3077.jpg", + "image_caption": [ + "Master lunging" + ], + "image_footnote": [], + "bbox": [ + 784, + 358, + 821, + 414 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d51661c81751385a118c3f26220095c524d92fcc7b3bc62b6d73e701ed983dfb.jpg", + "image_caption": [ + "Downward leaves" + ], + "image_footnote": [], + "bbox": [ + 676, + 431, + 782, + 492 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61f041eff042f551a3c8d666abfc62debdd211dd99c8150d3f0683cb0261cb9d.jpg", + "image_caption": [ + "Bent stem", + "Figure 1. Demonstration of skeletal conditions for controllable generation. We argue that skeletons and text provide complementary description for shape and appearance respectively, as shown in (a). Moreover, flexible and accurate control of object anatomy and pose can be realized by editing the joints and bones in skeletons, as shown in (b). Arbitrary skeletal structures are supported in our framework. Multiple views are generated and only front-view images are shown." + ], + "image_footnote": [], + "bbox": [ + 785, + 433, + 883, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tonization, we propose a new pipeline to generate skeletons with sparse joints from meshes. The pipeline mainly consists of curve skeleton extraction and curve simplification, achieving an $80\\%$ success rate, significantly outperforming previous deep learning based method RigNet [71].", + "bbox": [ + 89, + 599, + 482, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully control object anatomy and pose, we build the skeletal conditioned generation model in a multi-view manner. We represent a 3D skeleton with conditional skeleton images by Coordinate Color Encoding (CCE) to reduce ambiguity. Joints and bones are encoded with unique colors according to their 3D positions. For condition injection, we designed a Skeletal Correlation Module (SCM) to extract features from these conditional images and then generate multi-view images for the object. Later, a Large Reconstruction Model (LRM) is employed to produce 3D assets from the multi-view images. To address potential blurriness during reconstruction, we enhance appearance quality by a refinement stage that up-samples the multi-view images to higher resolutions and refines the texture in UV space.", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The experimental results indicate that our framework achieves instant generation of multi-view and 3D contents which are aligned with given skeletons. The proposed coordinate color encoding and the skeletal correlation module significantly improve the object-skeleton alignment, and accelerate model convergence by $5 \\times$ 3D assets conforming to the given skeleton can be generated in $\\sim 20\\mathrm{s}$ and refined in $\\sim 60\\mathrm{s}$ . To the best of our knowledge, this work is a pioneer in achieving arbitrary skeletal conditioned generation with following contributions:", + "bbox": [ + 511, + 599, + 906, + 752 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Constructing the first large-scale dataset, Objaverse-SK, containing mesh and skeleton pairs that cover diverse skeletal structures. We developed a pipeline to generate sparse skeletons from meshes with a high success rate.", + "- Proposing a multi-view and 3D generation pipeline for arbitrary skeletons, including coordinate color encoding for compact condition representation and the skeletal correlation module for effective condition injection." + ], + "bbox": [ + 513, + 777, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "315", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb419e83bedf16bb7189b8f83e0755967f849e5aa5905b801fed4f5e8d554f48.jpg", + "image_caption": [ + "Figure 2. Illustration of the pipeline for skeleton generation from meshes (§3.2). The curve skeleton is first extracted from the given mesh, followed by simplification of parted curves. The curve graph is converted to a tree as the final skeleton." + ], + "image_footnote": [], + "bbox": [ + 114, + 84, + 883, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 263, + 233, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable 2D Generation. Based on image diffusion models [51-53], versatile controlling conditions have been studied. In terms of spatial control, ControlNet [75] and other similar works [40, 78] train a side network for spatial conditions such as edge maps, normal maps and human skeletons. Some works focus on human image generation from skeletons [18, 20, 62]. Box-based instance control is also concerned in some works [29, 79, 81]. As for conceptual controlling, DreamBooth [54] finetunes the model to bind the given subject with an identifier in text prompt. IP-Adapter [74] trains an adapter to inject styles or concepts into the model. Some works [31, 32, 63] also focus on human ID control. Besides, some methods [2, 38, 39, 41] can achieve conditional generation without fine-tuning.", + "bbox": [ + 89, + 292, + 483, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable 3D Generation. Controlling concepts in 3D generation can be easily achieved by combining image control methods with image-to-3D generation [17, 27, 59, 69]. However, spatial control for 3D generation is not as easy as conceptual control. Many works achieve spatial control based on score distillation paradigm [28, 46, 83]. Coin3D [10] presents a framework to control the multi-view diffusion and 3D generation by shape proxies, i.e. combination of simple basic shapes. Sculpt3D [6] enhances text-to-3D generation with retrieved 3D priors. Sherpa3D [34] proposes to generate a coarse shape with a 3D diffusion model and refine the shape with SDS [46]. Some works for 3D human or head avatar generation [19, 24, 33, 77, 80, 82] uses human skeleton or facial landmarks as the condition in 2D or 3D space. Recent 3D native diffusion models [7, 21, 25, 67, 76] indicate promising results. Clay [76] designs a transformer-based [45, 60] 3D diffusion framework and various conditions like images and point clouds can be injected through cross-attention layers. A recent work MVControl [30] realizes 3D generation with single-view 2D spatial conditions like normal maps and edge maps by conditional multi-view generation and 3D reconstruction. Our work shares the similar workflow, but we focus on skeletal conditioned generation, which is not studied in MVControl. Another recent work AnimatableDreamer [65] generates 4D objects with skeletons extracted from given", + "bbox": [ + 89, + 508, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "videos by canonical score distillation, while our work develops a totally different framework for instant generation. As for skeletal representation, AnimatableDreamer extracts dense joints and unstructured bones from videos, while our method uses sparse joints and structured bones, which are more similar to human-made skeletons and more efficient for shape representation and animation.", + "bbox": [ + 511, + 265, + 906, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mesh Skeletonization. Various algorithms were designed for extracting skeletons from 3D meshes. [58] and [1] compute curve skeletons (C-S) via iterative mesh contraction operations. [11, 66] proposed to extract skeletons medial axis transformation skeleton (MAT-S) by point selection and connection prediction. C-S and MAT-S can serve as shape representation, while human-made skeletons (H-S) are often different from them. Since the main purpose is animation, H-S only contain sparse joints and bones. Some works [70, 71] propose data-driven approaches to learn mesh skeletonization from human annotated data. In this work, we have tried learning-based method [71] but found the results were not satisfactory. Therefore, we develop a new pipeline to generate skeletons which are as sparse as H-S while keep the shape of C-S.", + "bbox": [ + 511, + 371, + 908, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 611, + 718, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Preparation", + "text_level": 1, + "bbox": [ + 511, + 636, + 684, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The largest existing open dataset containing mesh-skeleton pairs is ModelResources [70]. There are around 3,000 3D meshes without textures. The scale is insufficient to train a text-driven generative model, and it lacks textures for appearance modeling. To address these limitations, we construct a dataset with $8 \\times$ larger scale and includes color textures. Our dataset, named Objaverse-SK, is built upon a large-scale 3D dataset Objaverse [8]. Although our data generation pipeline is applicable to a broad range of object categories, we focus on three main categories including \"Animals\", \"Human Shapes\" and \"Plants\", as they can typically be represented by tree-structured skeletons. Category labels are obtained from G-Objaverse [47]. Consequently, our dataset contains 24k 3D meshes, consisting of 15k animals, 6k human shapes and 3k plants. Text prompts for these models are generated by Cap3D [36].", + "bbox": [ + 511, + 659, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "316", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bc7efb92eb3dbc3c081563d6a8bc751ed0b04b53a2034ae245e8830e397fee18.jpg", + "image_caption": [ + "Figure 3. Illustration of the pipeline for skeletal conditioned multi-view and 3D generation (§4). The 3D skeleton is projected into 2D images and represented by coordinate color encoding. The skeletal correlation is modeled by skeletal correlation module, and then is injected into the diffusion model. Multi-view images are first generated and then a 3D textured mesh is reconstructed. The texture is further refined through UV-space optimization. Our framework achieves instant and high-quality generation given arbitrary skeletons." + ], + "image_footnote": [], + "bbox": [ + 116, + 93, + 879, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Skeleton Generation", + "text_level": 1, + "bbox": [ + 89, + 359, + 284, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain mesh-skeleton pairs, a effective method for generating skeletons from meshes is crucial. There are two concerns: the skeleton structure and success rate. The skeleton structure should accurately describe the object anatomy and be suitable for posing. Furthermore, an ideal method should generate reasonable skeleton structures with a high success rate. We tested a learning-based method RigNet [71] (Fig. 4). Although the generated skeleton structures can be close to human-made skeletons in its training data, it primarily produces symmetric skeletons and tends to be unstable across diverse anatomies and poses.", + "bbox": [ + 89, + 383, + 483, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skeleton extraction. To enhance flexibility and robustness, we design a new reliable pipeline, utilizing curve skeletons as the intermediate representation. An illustration of the pipeline is in Fig. 2. Given the structural inconsistency between curve skeletons and human-made skeletons, we further convert dense curves into sparse joints and bones. The detailed pipeline is elaborated below. 1) Initially, Mean Curvature Flow (MCF) [58] is employed to generate curve skeletons from meshes robustly. 2) Next, we build a graph from the set of curves generated from the mesh, consisting of dense nodes and edges. Intersection nodes (degree $>1$ ) are identified and the graph is divided into several parts by these nodes. 3) In each part, the curve contains no branches so it can be simplified by the Douglas-Peucker algorithm (DP) [12] into line segments.", + "bbox": [ + 89, + 551, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tree conversion. At this stage, the basic shape of the skeleton is established, but the root position and the bone directions between joints still need to be determined. The problem can be regarded as a graph-to-tree conversion. First, a spanning tree is constructed from the graph to eliminate cycles. We then identify high-degree intersection nodes as candidates for the root. To ensure an efficient structure, the skeleton is configured by selecting the tree with the min", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "imum height among these candidates. This approach ensures that the root node is located at a significant intersection, minimizing the distances between the root and other joints. More details can be found in the appendix.", + "bbox": [ + 511, + 361, + 906, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Generation Pipeline", + "text_level": 1, + "bbox": [ + 511, + 439, + 705, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Skeletal Conditioned Multi-view Generation", + "text_level": 1, + "bbox": [ + 511, + 465, + 888, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the dataset is constructed, we consider building the conditional generative model based on it. Since unconditional multi-view diffusion models have been well studied, we start directly from a base model MVDream [55] and focus on the conditional generation. Two main issues are of concern: i) how the skeleton is represented, and ii) how it is injected into the model.", + "bbox": [ + 511, + 489, + 905, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skeletal condition representation. As we aim to generate images which align with the given skeletons, using spatial guidance in the diffusion model is a reasonable approach. Skeletons can be projected from 3D space to the image plane as 2D conditions. However, overlapping and information loss occur during the projection, which may lead to semantic and structural ambiguities for spatial guidance, as illustrated in Fig. 8. Thus, incorporating richer information is crucial to mitigate these ambiguities.", + "bbox": [ + 511, + 597, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coordinate Color Encoding (CCE). In order to preserve 3D information, we encode joint coordinates using spatial colors. While prior works [28, 61] use canonical color maps for shape representation, our approach focuses on representing skeletons with sparse joints and bones. We begin by normalizing skeletons within a canonical cube $[0,1]^3$ . Each position in this cube corresponds to a unique color, with RGB values precisely matching the positional coordinates. As a result, the 2D conditional image can represent the 3D spatial positions of the skeleton. For bones, we assign colors based on their midpoint. Additionally, we incor", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "317", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "porate normalized values of view-dependent inverse depth of the skeleton as the alpha channel (CCE-D). With the absolute spatial coordinates and relative depth encoded in the conditional images, there will be more precise and richer guidance information for generation.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Skeletal condition injection. Spatial conditions such as canny edges and normal maps have been investigated in 2D image diffusion models. In ControlNet [75], the conditional image is encoded by convolution blocks, resulting in an output spatial size that matches the latent size. Then, the condition features are added to the latent features. The encoder of the original diffusion model is copied as a side network to produce guidance features, which are fused with the original features in the decoder. Our pipeline adopts this paradigm from ControlNet, and further enhances it with a more effective condition feature extraction module.", + "bbox": [ + 89, + 167, + 482, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Skeletal Correlation Modeling (SCM). For a skeleton in 3D space, we first project it into multi-view images as 2D conditions. Given the sparse nature of skeletons in spatial dimension, convolution blocks lack global modeling capacity. To address this, we design a Skeletal Correlation Module (SCM) to enhance the condition features by modeling the anatomical correlation among different parts of a skeleton, and the view correlation for different projection views. The structure of the module is in Fig. 3. (i) First, anatomical correlation is extracted by a self-attention layer, which constructs the global skeleton features for each view. (ii) Then, the cross-view correlation is modeled by a cross-attention layer, allowing the extraction of correspondences among skeleton images from multiple views. This enables the model to recognize identical joints in different views. In addition, we use adaptive layer normalization [68] to fuse the camera pose embedding with the skeletal features. Associating each skeleton image with a camera pose helps to generate view-dependent shapes. Adding correlation layers during condition encoding significantly facilitates learning, achieving $5 \\times$ faster convergence (Fig. 9).", + "bbox": [ + 89, + 334, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multi-view Images to 3D Generation", + "text_level": 1, + "bbox": [ + 89, + 664, + 408, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instant reconstruction. Given the generated multi-view images, we use a Large Reconstruction Model (LRM), specifically InstantMesh [69] for fast textured mesh reconstruction. However, the reconstructed textures often appear blurry. On one hand, the resolution of generated images is $256^2$ , which struggles to capture fine details. On the other hand, the appearance quality also degrades during reconstruction. In order to recover and further enrich the appearance, we introduce a new refinement stage.", + "bbox": [ + 89, + 688, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance refinement. First, the generated multi-view images are upscaled 4 times into $1024^2$ by Stable Diffusion with ControlNet-Tile [75]. ControlNet-Edge [9, 75] is used to maintain the shape consistency across different views during tiling. Once tiled, these high-resolution im", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ages $I_{i}^{h}$ are used to refine the reconstructed texture. A learnable 2D texture $u$ in UV space is created and initialized as the reconstructed texture $u_{0}$ , and then images are rendered through differentiable rendering $\\mathcal{R}(u, c_{i})$ for given camera views $c_{i}$ . The MSE loss is optimized between the rendered images and tiled high-res images. Moreover, a regularization term is added to maintain consistency in UV space:", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\sum_ {i} \\left| \\left| I _ {i} ^ {h} - \\mathcal {R} (u, c _ {i}) \\right| \\right| _ {2} ^ {2} + \\lambda * \\left| \\left| u - u _ {0} \\right| \\right| _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 209, + 903, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the high-res images cannot cover every position on the mesh, some regions of $u$ will not be optimized, e.g. bottom of the object. We found these regions are unstable during optimization and may produce unexpected artifacts (see Fig. 10). The regularization term helps the optimized texture maintain the appearance from $u_{0}$ in these regions. Consequently, the high-frequency details can be learned in covered regions while the global consistency can also be achieved in uncovered regions. The optimization can be finished within 15 seconds.", + "bbox": [ + 511, + 250, + 906, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Object-Skeleton Alignment Evaluation", + "text_level": 1, + "bbox": [ + 511, + 414, + 846, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive alignment. In order to measure how much an object is aligned with a skeleton, we develop a new evaluator, named Contrastive Object-Skeleton Alignment (COSA). We use the self-supervised DINOv2 [44] as the backbone $F$ to extract both object and skeleton features. Then, the alignment adapter $G_{\\theta}$ consisting of several self-attention layers is used to modulate the features. The adapter ends with an average pooling layer to aggregate the aligned features into a vector. Similar to CLIP [49], we train the adapter using contrastive learning by InfoNCE loss [42, 56]. Finally, the skeleton alignment score (SKA) can be calculated by cosine similarity between the features from an object image $x$ and a skeleton image $y$ as:", + "bbox": [ + 511, + 436, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {\\mathrm {S K A}} (x, y) = \\cos (G (F (x)), G (F (y))). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 647, + 903, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COSA guided diffusion. Based on COSA, another conditional generation pipeline can also be realized, following the approach proposed in [2]. On each denoising time step $t$ , the approximate clean image $\\hat{x}_0$ is estimated from the predicted noise $\\epsilon_t$ as in DDIM [57]. The estimated clean image and skeleton condition are fed into COSA to calculate the alignment loss $\\mathcal{L}_{\\mathrm{COSA}}(\\hat{x}_0,y) = 1 - S_{\\mathrm{SKA}}(\\hat{x}_0,y)$ . Then the predicted noise is modified by the gradient of the alignment loss for actual denoising:", + "bbox": [ + 511, + 676, + 906, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {t} = \\epsilon_ {t} + s (t) \\cdot \\nabla \\mathcal {L} _ {\\mathrm {C O S A}} (\\hat {x} _ {0}, y) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 825, + 903, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s(t)$ controls the guidance strength. With the additional guidance of the alignment loss, the generated object will tend to follow the conditional skeleton $y$ .", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "318", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/acec4fb0014807ba5e9752964bdc131ce667ae90378461d45d06cf7e24576576.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method\\SKA Score TrainingMeanInst.MeanClassAnimalsHumansPlantsApodesBipedsQuadrupedsArthropodsWings
SDEdit [38]○70.1365.0679.5064.7050.9974.7475.2482.7379.0779.21
SDEdit+COSAG●72.1167.3280.9167.6053.4679.7383.1785.9877.8073.82
Ours-Raw●74.6967.0988.3969.9042.9794.7482.1490.1889.8184.69
Ours-SCM●80.4374.3891.1678.4553.5394.4785.2094.1990.6888.40
", + "bbox": [ + 102, + 88, + 897, + 175 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bcf031ae3cfe895a1d563a8864d84ac3ee09ba1c5e21060cf46fec9d837cc140.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of object-skeleton alignment (SKA) score (\\$5.2). Alignment scores are calculated over three classes (blue) and five subclasses of animal (green). The average scores over all instances and three classes (pink) are also shown. The highest scores among all methods are bold and the highest scores among baseline methods are underlined." + ], + "table_footnote": [], + "table_body": "
MethodTrainingPickScoreCLIP Score
Win RateAnimalsHumanPlants\\( Mean_{Inst.} \\)AnimalsHumanPlants
SDEdit [38]○19.5618.2314.6628.9829.0429.4728.7128.29
SDEdit+COSAG\\( \\blacksquare \\)18.1718.5419.9515.0628.9929.5228.5028.11
Ours-Raw●28.9929.0628.3729.5529.9030.1030.0229.23
Ours-SCM●33.2834.1737.0226.4229.8330.2329.8428.75
", + "bbox": [ + 155, + 234, + 844, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison of PickScore and CLIP Score (§5.2). Scores are calculated over three classes (blue) and averaged over all instances (red). The highest scores among all methods are bold and the highest scores among baseline methods are underlined.", + "bbox": [ + 89, + 335, + 906, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e5ddf8ef1448ba0aded0e61608692a204b41d8a063ee69935a38ec68f8e59f5.jpg", + "image_caption": [ + "Figure 4. Comparison of skeletons generated from 3D meshes by RigNet [71] and our method (§5.1)." + ], + "image_footnote": [], + "bbox": [ + 99, + 376, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 580, + 217, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Results of Mesh Skeletonization", + "text_level": 1, + "bbox": [ + 89, + 604, + 372, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the learning-based method RigNet [71], and the results are shown in Fig. 4. RigNet tends to produce symmetric skeletons with limited flexibility and misaligned joints/bones, resulting in an around $15\\%$ success rate. In contrast, our method supports arbitrary-pose skeletons and achieves better joint/bone alignment. It produces more reliable results with an $80\\%$ success rate. More details and results can be found in appendix.", + "bbox": [ + 89, + 627, + 483, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results of Multi-view Generation", + "text_level": 1, + "bbox": [ + 89, + 757, + 382, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocols. The Objaverse-SK evaluation set contains 108 skeleton instances, covering three main classes: animals, human shapes and plants. As animals include diverse skeleton structures, we further divide them into more detailed subclasses (examples are shown behind): Apodes (fish, snakes), Bipeds (ducks, penguins), Quadrupeds (dogs, bears), Arthropods (scorpions, crabs), Wings (birds, dragons). Three evaluation metrics are con", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sidered for multi-view generation: SKA Score for skeletal alignment, PickScore for image quality, and CLIP Score for textual alignment. Furthermore, samples from categories excluded in training are obtained from ShapeNet [5] for generalization evaluation (Fig. 7). Additionally, evaluation results on human-made skeletons are in appendix.", + "bbox": [ + 511, + 378, + 906, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods. Since there is no prior work that can achieve arbitrary skeletal conditioned generation, we implement two methods for comparison. The first baseline is SDEdit [38]. The process starts from adding noise on conditional images with a time step (set as 0.7). Then clean images are generated by denoising steps. The method is entirely unsupervised. The second baseline is the COSA Guidance (COSAG) derived from [2], which is elaborated in Section 4.3. The guidance strength is set as $s(t) = 7.5\\sqrt{1 - \\alpha_t}$ . Since we found it cannot achieve stable results, it is combined with SDEdit. The method requires an extra model so it is half supervised. Ours is fully supervised on object-skeleton pairs.", + "bbox": [ + 511, + 470, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison. The qualitative results are shown in Fig. 5. Given skeleton images as conditions, SDEdit can produce images following the skeleton. However, limited by the editing capacity, the generated objects often have incorrect anatomies. For example, the snake body is broken, and the donkey body is generated as wooden. When it is enhanced by the COSAG, the quality of generated content is improved in some cases but still unsatisfactory. Compared with them, our results show superior quality and are more consistent with both skeletal and textual conditions.", + "bbox": [ + 511, + 670, + 908, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. Comparison results of skeleton alignment are shown in Table 1. SDEdit-based methods have around 70 SKA scores, while ours can achieve 80. Among three classes, animals tend to have higher alignment scores while plants have lower scores. Since plants", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "319", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/248e142f0b164f8274b9388b4f2b47475b3cb316226ed855a2933b2170ec15c9.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of skeletal conditioned multi-view generation (§5.2). Conditional skeletons and text prompts are shown above. Four views are generated and two views are shown for simplicity." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 381 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d571e216cc6080034d8efe37e6c35b3ae263d050c8b1f4bf5584bb102ff910bc.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of textured meshes before and after refinement (§5.3). Rendered color images are shown. Local areas are enlarged for better viewing." + ], + "image_footnote": [], + "bbox": [ + 109, + 421, + 883, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "may have more complex structures and sometimes extend further from skeletons. For five sub-classes, our method achieves consistently high alignment scores. With the help of SCM and CCE-D skeletons, both the alignment scores and pick scores are further improved.", + "bbox": [ + 89, + 613, + 483, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results of 3D Generation", + "text_level": 1, + "bbox": [ + 89, + 700, + 321, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Texture refinement. Results of 3D reconstruction from multi-view images are shown in Fig. 6. The raw reconstructed and refined results are compared. The raw textures are blurry and lack details, while the proposed refinement stage can significantly enhance texture quality.", + "bbox": [ + 89, + 723, + 483, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rigging and animation. Given a motion sequence of a skeleton, our method can be applied to generate 4D animation. Textured mesh can be generated given the skeleton at the rest pose, and then directly be rigged and skinned for animation. Demo videos can be found in project page1.", + "bbox": [ + 89, + 799, + 483, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 612, + 665, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skeletal condition representation. The skeletal condition representation we use consists of coordinate color encoding (CCE) with depth alpha (D). The ablation results are shown in Fig. 8 and Tab. 3. Richer information in conditions can help the model determine the content better. As a result, better alignment can be achieved. In Fig. 8 right, the skeleton of a penguin is highly ambiguous. If CCE-D is used, the body pose and orientation of the penguin can be successfully inferred from colors.", + "bbox": [ + 511, + 640, + 906, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skeletal correlation modeling. We show the effect of our skeletal correlation module in Tab. 3 and Fig. 9. SCM achieves better alignment scores than convolutional blocks. SCM with layer normalization (LN) achieves a $4 \\times$ faster convergence rate. Furthermore, if LN is replaced with the adaptive LN (AdaLN), the model can achieve an SKA score of 70 within 1k training steps. The results indicate that extracting global features from conditional images is crucial.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1https://skdream3d.github.io/.", + "bbox": [ + 109, + 886, + 269, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "320", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21dc4b36abee86086f91274e70d994423aa6ff898b31ccc02c4e4502d4973500.jpg", + "image_caption": [ + "Figure 7. Qualitative results of novel categories in ShapeNet [5] (§5.2), which are not covered by the training set of Objaverse-SK." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 472, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f20c9f0f6322eb23b4bcc7a61cd5038f277d74f8689d33145806fd728ef901b0.jpg", + "image_caption": [ + "Figure 8. Ablation study of coordinate color encoding with depth alpha (CCE-D) ( $\\S 6$ ). Richer information can help the model to avoid ambiguity and generate better anatomy." + ], + "image_footnote": [], + "bbox": [ + 99, + 412, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/95c30b52a53b11c456c5801a19825d1a7432fbc72c9dd16cccb0685aa81070d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModuleSkeletonSKA ScoreAnimalsHumansPlants
SCMCCE-D80.4391.1678.4553.53
SCMCCE78.9790.4574.8252.57
SCMBinary77.2789.2076.0446.16
ConvCCE-D76.6588.5375.0446.17
ConvBinary74.6988.3969.9042.97
", + "bbox": [ + 91, + 603, + 488, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of skeletal module and representation types (§6). SCM and CCE-D achieve higher alignment scores.", + "bbox": [ + 89, + 705, + 482, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Refinement regularization. We show the ablation results of appearance refinement in Fig. 10. The refined appearance contains rich details such as snake scales and wood grain, in comparison to the reconstructed results. However, artifacts also appear in the regions that are not covered by high-res images. With the help of UV space regularization, the artifacts are effectively removed in uncovered regions. As a result, consistent colors are maintained from original textures and details are enhanced during optimization.", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90325cbbee12e7ab61a01307d47912cc129fb6fe145131613362f7729b3d5fbe.jpg", + "image_caption": [ + "Figure 9. Convergence processes of different skeletal modules (§6). SCM with AdaLN achieves 5x faster convergence." + ], + "image_footnote": [], + "bbox": [ + 535, + 92, + 880, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2279ee273e3641a6df134415d244219e7e00bb6e24d2f692e65d2b6774fafae.jpg", + "image_caption": [ + "Figure 10. Ablation study of UV space regularization (§6). Bottom views and UV textures are shown. Front views of the snake and the tree stump can be found in the first column of Fig. 6." + ], + "image_footnote": [], + "bbox": [ + 524, + 262, + 691, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25e3dacd578b05391db044f194cb5e08c46a16fc76cb7db90f8fcd4cd60a7202.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 263, + 888, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 513, + 455, + 777, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our work achieves arbitrary skeletal conditioned generation, there are still many problems that can be further studied. The skeletons we currently use may have limited descriptive ability for non-tree structured objects. More powerful shape representations can be studied as new conditions. In addition, our work only considers global skeletons without fine-grained semantics. Injecting detailed semantics into the skeleton parts is a meaningful topic to study. For future work, the skeleton condition can be studied in native 3D generation frameworks. Moreover, our dataset may also be used in other tasks such as arbitrary skeleton estimation from images. More discussion is in the appendix.", + "bbox": [ + 511, + 482, + 906, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose to use skeletons as the structural condition for controllable generation. First, we construct a large-scale 3D mesh-skeleton paired dataset. We propose an effective mesh skeletonization method to generate mesh-aligned sparse skeletons with a high success rate. Based on the dataset, we present a skeletal conditioned multi-view generation pipeline. Coordinate color encoding and skeletal correlation module are proposed to realize efficient condition representation and injection. Furthermore, 3D meshes can be instantly reconstructed, followed by a refinement stage to achieve better texture quality. In summary, our work achieves controllable multi-view and 3D generation with arbitrary skeletons as conditions.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "321", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (No. 62293554), and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02) and the Fundamental Research Funds for the Zhejiang Provincial Universities (226-2024-00208).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Bærentzen and Eva Rotenberg. Skeletonization via local separators. ACM Transactions on Graphics (TOG), 40 (5):1-18, 2021. 3", + "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 5, 6", + "[3] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 1", + "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8", + "[6] Cheng Chen, Xiaofeng Yang, Fan Yang, Chengzeng Feng, Zhoujie Fu, Chuan-Sheng Foo, Guosheng Lin, and Fayao Liu. Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10228-10237, 2024. 3", + "[7] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 3", + "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 1, 3", + "[9] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5", + "[10] Wenqi Dong, Bangbang Yang, Lin Ma, Xiao Liu, Liyuan Cui, Hujun Bao, Yuewen Ma, and Zhaopeng Cui. Coin3d: Controllable and interactive 3d assets generation with proxy-guided conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1, 3", + "[11] Zhiyang Dou, Cheng Lin, Rui Xu, Lei Yang, Shiqing Xin, Taku Komura, and Wenping Wang. Coverage axis: Inner point selection for 3d shape skeletonization. In Com" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "puter Graphics Forum, pages 419-432. Wiley Online Library, 2022. 3", + "[12] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 10 (2):112-122, 1973. 4", + "[13] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Navigation instruction generation with bev perception and large language models. In ECCV, 2024. 1", + "[14] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Scene map-based prompt tuning for navigation instruction generation. In CVPR, 2025. 1", + "[15] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 1", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 1, 3", + "[18] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3", + "[19] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[20] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 1, 3", + "[21] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3", + "[22] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46, 2007. 1", + "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1", + "[24] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 3", + "[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "322", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2024. 3", + "[26] Christoph Leiter, Ran Zhang, Yanran Chen, Jonas Belouadi, Daniil Larionov, Vivian Fresen, and Steffen Eger. Chatgpt: A meta-analysis after 2.5 months. arXiv preprint arXiv:2302.13795, 2023. 1", + "[27] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 1, 3", + "[28] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 3, 4", + "[29] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3", + "[30] Zhiqi Li, Yiming Chen, Lingzhe Zhao, and Peidong Liu. Mycontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494, 2023. 1, 3", + "[31] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3", + "[32] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6400-6409, 2024. 3", + "[33] Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J Black. Tada! text to animatable digital avatars. In 2024 International Conference on 3D Vision (3DV), pages 1508-1519. IEEE, 2024. 3", + "[34] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20763–20774, 2024. 3", + "[35] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1", + "[36] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[37] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3, 6", + "[39] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3", + "[40] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 3", + "[41] Marianna Ohanyan, Hayk Manukyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Zero-painter: Training-free layout control for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8764-8774, 2024. 3", + "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5", + "[43] OpenAI. Gpt-4 technical report, 2023. 1", + "[44] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5", + "[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3", + "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[47] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 3", + "[48] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "323", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3", + "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022.", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[54] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3", + "[55] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 4", + "[56] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 5", + "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 5", + "[58] Andrea Tagliasacchi, Ibraheem Alhashim, Matt Olson, and Hao Zhang. Mean curvature skeletons. In Computer Graphics Forum, pages 1735-1744. Wiley Online Library, 2012. 3, 4", + "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3", + "[60] Ashish Vaswani. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 3", + "[61] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 4", + "[62] Jiajun Wang, Morteza Ghahremani, Yitong Li, Björn Omer, and Christian Wachinger. Stable-pose: Leveraging transformers for pose-guided text-to-image generation. arXiv preprint arXiv:2406.02485, 2024. 3", + "[63] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3", + "[64] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Xinzhou Wang, Yikai Wang, Junliang Ye, Fuchun Sun, Zhengyi Wang, Ling Wang, Pengkun Liu, Kai Sun, Xintong Wang, Wende Xie, et al. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 3", + "[66] Zimeng Wang, Zhiyang Dou, Rui Xu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Shiqing Xin, Taku Komura, Xiaoming Yuan, and Wenping Wang. Coverage axis++: Efficient inner point selection for 3d shape skeletonization. In Computer Graphics Forum, page e15143. Wiley Online Library, 2024. 3", + "[67] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 3", + "[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 5", + "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 1, 3, 5", + "[70] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, and Karan Singh. Predicting animation skeletons for 3d articulated models via volumetric nets. In 2019 international conference on 3D vision (3DV), pages 298-307. IEEE, 2019. 3", + "[71] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2, 3, 4, 6", + "[72] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 1", + "[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 1", + "[74] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3", + "[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3, 5", + "[76] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3", + "[77] Muxin Zhang, Qiao Feng, Zhuo Su, Chao Wen, Zhou Xue, and Kun Li. Joint2human: High-quality 3d human generation via compact spherical embedding of 3d joints. In Pro" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "324", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1438, 2024. 1, 3", + "[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[79] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3", + "[80] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavator: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 3", + "[81] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3", + "[82] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 3", + "[83] Wenjie Zhuo, Fan Ma, Hehe Fan, and Yi Yang. Vividdreamer: invariant score distillation for hyper-realistic text-to-3d generation. In European Conference on Computer Vision, pages 122–139. Springer, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 515 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "325", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_model.json b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91f628df66c3d34736dc0a687d7ee8e7586dc7c4 --- /dev/null +++ b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_model.json @@ -0,0 +1,3260 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.18, + 0.667, + 0.235 + ], + "angle": 0, + "content": "Yuanyou Xu\\(^{1}\\), Zongxin Yang\\(^{2}\\), Yi Yang\\(^{1*}\\) \n\\(^{1}\\)ReLER, CCAI, Zhejiang University \n\\(^{2}\\)DBMI, HMS, Harvard University" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.237, + 0.742, + 0.253 + ], + "angle": 0, + "content": "{yoxu, yangyics}@zju.edu.cn, zongxin-yang@hms.harvard.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.637 + ], + "angle": 0, + "content": "Controllable generation has achieved substantial progress in both 2D and 3D domains, yet current conditional generation methods still face limitations in describing detailed shape structures. Skeletons can effectively represent and describe object anatomy and pose. Unfortunately, past studies are often limited to human skeletons. In this work, we generalize skeletal conditioned generation to arbitrary structures. First, we design a reliable mesh skeletonization pipeline to generate a large-scale mesh-skeleton paired dataset. Based on the dataset, a multi-view and 3D generation pipeline is built. We propose to represent 3D skeletons by Coordinate Color Encoding as 2D conditional images. A Skeletal Correlation Module is designed to extract global skeletal features for condition injection. After multi-view images are generated, 3D assets can be obtained by incorporating a large reconstruction model, followed by a UV texture refinement stage. As a result, our method achieves instant generation of multi-view and 3D contents that are aligned with given skeletons. The proposed techniques largely improve the object-skeleton alignment and generation quality. Project page at https://skdream3d.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.222, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.484, + 0.875 + ], + "angle": 0, + "content": "In view of visual representation dimension [64, 72], 2D image generation [16, 53, 57], multi-view (2.5D) generation [35, 55], and 3D generation [17, 27, 69] have advanced successively and achieved significant progress. Given the success of large language models [23, 26, 43, 48-50], textual representation has been widely applied in generation [51-53], as in other domains [13, 14, 73]. To achieve more controllable generation, conditions beyond text have attracted considerable attention. 2D image conditions (e.g., edge maps, human skeletons, and concept references) [54, 75] have been well studied. Similarly in 3D generation," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.889, + 0.304 + ], + "angle": 0, + "content": "2D [30] and 3D [10] conditions have also been explored." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.305, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Although the aforementioned conditions in controllable generation complement text descriptions, they still struggle to precisely describe shape structures. In contrast, skeletons, among various types of conditions, exhibit superior ability to depict shape structures: (i) Representation of object anatomy. Skeletons can efficiently represent various 3D structures with sparse joints and bones, while other conditions like depth maps fail to represent full 3D anatomy. (ii) Articulation into different poses. Skeletons are widely used for character animation in computer graphics [3, 22] due to their simplicity and efficiency. Other conditions such as rough shapes [10] are inconvenient for pose articulation. (iii) Freedom of editing. Given an initial skeleton, users can freely add new structures or modify joints and bones to create ideal shapes. Examples are in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Despite these advantages, previous studies [19, 20, 40, 75, 77] on conditional generation are limited to human skeletons. From the perspective of generalization, we would like to ask: Is it possible to use arbitrary skeletons as conditions to generate any creatures or even general objects?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.809 + ], + "angle": 0, + "content": "To achieve this goal, we identify two main challenges that hinder the use of arbitrary skeletal conditions for generation: (i) Lack of large-scale object-skeleton pairs for training. Extensive studies [4, 15, 37] on 2D/3D human pose estimation have made human-skeleton paired data readily available. However, when skeletal structures are unknown, estimating arbitrary skeletons from 2D images or videos becomes challenging due to its ill-posedness. (ii) Insufficiency of 2D information to describe arbitrary skeletons. Human skeletons are relatively simple and can be described by a fixed set of 2D joints. However, complex skeletons suffer from self-occlusion and ambiguity, necessitating 3D information to fully capture their anatomy and pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.9 + ], + "angle": 0, + "content": "To address these challenges, we focus on multi-view and 3D generation with skeletal conditions. To tackle data scarcity problem, we construct a large-scale dataset Objaverse-SK containing mesh-skeleton pairs. Textured meshes are selected from Objaverse [8] by semantic classes to form a subset. In order to achieve reliable mesh ske" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.238, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "314" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.105, + 0.12, + 0.147 + ], + "angle": 270, + "content": "MVDream" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.105, + 0.235, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.108, + 0.3, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.307, + 0.108, + 0.358, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.18, + 0.12, + 0.201 + ], + "angle": 0, + "content": "sno" + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.176, + 0.18, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.222, + 0.192, + 0.232 + ], + "angle": 0, + "content": "Text & Skeleton" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.174, + 0.239, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.222, + 0.225, + 0.232 + ], + "angle": 0, + "content": "Dog" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.178, + 0.302, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.222, + 0.289, + 0.232 + ], + "angle": 0, + "content": "Chair" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.166, + 0.361, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.222, + 0.349, + 0.232 + ], + "angle": 0, + "content": "Fish" + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.123, + 0.435, + 0.134 + ], + "angle": 0, + "content": "Text Only" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.161, + 0.437, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.098, + 0.499, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.16, + 0.499, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.223, + 0.496, + 0.232 + ], + "angle": 0, + "content": "Spiderman" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.099, + 0.552, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.16, + 0.562, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.223, + 0.552, + 0.231 + ], + "angle": 0, + "content": "Rocket" + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.099, + 0.614, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.16, + 0.628, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.223, + 0.622, + 0.232 + ], + "angle": 0, + "content": "Ice cream" + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.124, + 0.698, + 0.134 + ], + "angle": 0, + "content": "Text Only" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.161, + 0.693, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.108, + 0.766, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.106, + 0.836, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.848, + 0.099, + 0.875, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.16, + 0.693, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.159, + 0.766, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.16, + 0.825, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.839, + 0.163, + 0.886, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.223, + 0.815, + 0.232 + ], + "angle": 0, + "content": "Coral" + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.222, + 0.876, + 0.231 + ], + "angle": 0, + "content": "Torch" + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.239, + 0.747, + 0.254 + ], + "angle": 0, + "content": "(a) Text and skeletons for Collaborative Appearance and Shape Controlling" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.263, + 0.2, + 0.275 + ], + "angle": 0, + "content": "Initial skeleton" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.293, + 0.163, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.332, + 0.16, + 0.34 + ], + "angle": 0, + "content": "Animal" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.356, + 0.163, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.418, + 0.16, + 0.426 + ], + "angle": 0, + "content": "Human" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.436, + 0.159, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.494, + 0.152, + 0.503 + ], + "angle": 0, + "content": "Plant" + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.261, + 0.38, + 0.276 + ], + "angle": 0, + "content": "Adding joints, bones" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.279, + 0.296, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.331, + 0.29, + 0.341 + ], + "angle": 0, + "content": "Dog with a spiral tail" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.283, + 0.417, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.332, + 0.414, + 0.341 + ], + "angle": 0, + "content": "Dog with two heads" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.355, + 0.297, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.418, + 0.301, + 0.428 + ], + "angle": 0, + "content": "Knight with dragon wings" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.432, + 0.293, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.494, + 0.302, + 0.503 + ], + "angle": 0, + "content": "Plant with leaves and roots" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.356, + 0.432, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.418, + 0.416, + 0.428 + ], + "angle": 0, + "content": "Monster with four arms" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.432, + 0.411, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.494, + 0.43, + 0.503 + ], + "angle": 0, + "content": "Plant with legs, Pokémon" + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.261, + 0.664, + 0.276 + ], + "angle": 0, + "content": "Modifying joint positions, bone sizes" + }, + { + "type": "image", + "bbox": [ + 0.433, + 0.277, + 0.545, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.331, + 0.536, + 0.341 + ], + "angle": 0, + "content": "Dog with a big head" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.277, + 0.662, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.331, + 0.659, + 0.341 + ], + "angle": 0, + "content": "Dog with a long torso" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.341, + 0.607, + 0.351 + ], + "angle": 0, + "content": "How wide?" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.352, + 0.551, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.418, + 0.526, + 0.428 + ], + "angle": 0, + "content": "Dwarf in armor" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.352, + 0.665, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.418, + 0.671, + 0.428 + ], + "angle": 0, + "content": "Wide shoulder bodybuilder" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.436, + 0.55, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.495, + 0.551, + 0.504 + ], + "angle": 0, + "content": "Tree with wide branches" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.432, + 0.606, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.495, + 0.659, + 0.504 + ], + "angle": 0, + "content": "Tree with a long truck" + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.262, + 0.814, + 0.276 + ], + "angle": 0, + "content": "Posing" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.278, + 0.777, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.332, + 0.763, + 0.343 + ], + "angle": 0, + "content": "Dog jumping" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.281, + 0.879, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.331, + 0.857, + 0.342 + ], + "angle": 0, + "content": "Dog sitting" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.354, + 0.777, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.417, + 0.769, + 0.428 + ], + "angle": 0, + "content": "Athlete running" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.359, + 0.822, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.417, + 0.865, + 0.428 + ], + "angle": 0, + "content": "Master lunging" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.433, + 0.784, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.495, + 0.771, + 0.504 + ], + "angle": 0, + "content": "Downward leaves" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.434, + 0.885, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.495, + 0.855, + 0.504 + ], + "angle": 0, + "content": "Bent stem" + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.51, + 0.733, + 0.525 + ], + "angle": 0, + "content": "(b) Skeleton-based editing for Accurate Anatomy and Pose Controlling" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.531, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Demonstration of skeletal conditions for controllable generation. We argue that skeletons and text provide complementary description for shape and appearance respectively, as shown in (a). Moreover, flexible and accurate control of object anatomy and pose can be realized by editing the joints and bones in skeletons, as shown in (b). Arbitrary skeletal structures are supported in our framework. Multiple views are generated and only front-view images are shown." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.678 + ], + "angle": 0, + "content": "tonization, we propose a new pipeline to generate skeletons with sparse joints from meshes. The pipeline mainly consists of curve skeleton extraction and curve simplification, achieving an \\(80\\%\\) success rate, significantly outperforming previous deep learning based method RigNet [71]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To fully control object anatomy and pose, we build the skeletal conditioned generation model in a multi-view manner. We represent a 3D skeleton with conditional skeleton images by Coordinate Color Encoding (CCE) to reduce ambiguity. Joints and bones are encoded with unique colors according to their 3D positions. For condition injection, we designed a Skeletal Correlation Module (SCM) to extract features from these conditional images and then generate multi-view images for the object. Later, a Large Reconstruction Model (LRM) is employed to produce 3D assets from the multi-view images. To address potential blurriness during reconstruction, we enhance appearance quality by a refinement stage that up-samples the multi-view images to higher resolutions and refines the texture in UV space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.753 + ], + "angle": 0, + "content": "The experimental results indicate that our framework achieves instant generation of multi-view and 3D contents which are aligned with given skeletons. The proposed coordinate color encoding and the skeletal correlation module significantly improve the object-skeleton alignment, and accelerate model convergence by \\(5 \\times\\) 3D assets conforming to the given skeleton can be generated in \\(\\sim 20\\mathrm{s}\\) and refined in \\(\\sim 60\\mathrm{s}\\). To the best of our knowledge, this work is a pioneer in achieving arbitrary skeletal conditioned generation with following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.778, + 0.905, + 0.837 + ], + "angle": 0, + "content": "- Constructing the first large-scale dataset, Objaverse-SK, containing mesh and skeleton pairs that cover diverse skeletal structures. We developed a pipeline to generate sparse skeletons from meshes with a high success rate." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.838, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- Proposing a multi-view and 3D generation pipeline for arbitrary skeletons, including coordinate color encoding for compact condition representation and the skeletal correlation module for effective condition injection." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.778, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "315" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.085, + 0.885, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.224, + 0.908, + 0.253 + ], + "angle": 0, + "content": "Figure 2. Illustration of the pipeline for skeleton generation from meshes (§3.2). The curve skeleton is first extracted from the given mesh, followed by simplification of parted curves. The curve graph is converted to a tree as the final skeleton." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.265, + 0.234, + 0.28 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.485, + 0.506 + ], + "angle": 0, + "content": "Controllable 2D Generation. Based on image diffusion models [51-53], versatile controlling conditions have been studied. In terms of spatial control, ControlNet [75] and other similar works [40, 78] train a side network for spatial conditions such as edge maps, normal maps and human skeletons. Some works focus on human image generation from skeletons [18, 20, 62]. Box-based instance control is also concerned in some works [29, 79, 81]. As for conceptual controlling, DreamBooth [54] finetunes the model to bind the given subject with an identifier in text prompt. IP-Adapter [74] trains an adapter to inject styles or concepts into the model. Some works [31, 32, 63] also focus on human ID control. Besides, some methods [2, 38, 39, 41] can achieve conditional generation without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.487, + 0.904 + ], + "angle": 0, + "content": "Controllable 3D Generation. Controlling concepts in 3D generation can be easily achieved by combining image control methods with image-to-3D generation [17, 27, 59, 69]. However, spatial control for 3D generation is not as easy as conceptual control. Many works achieve spatial control based on score distillation paradigm [28, 46, 83]. Coin3D [10] presents a framework to control the multi-view diffusion and 3D generation by shape proxies, i.e. combination of simple basic shapes. Sculpt3D [6] enhances text-to-3D generation with retrieved 3D priors. Sherpa3D [34] proposes to generate a coarse shape with a 3D diffusion model and refine the shape with SDS [46]. Some works for 3D human or head avatar generation [19, 24, 33, 77, 80, 82] uses human skeleton or facial landmarks as the condition in 2D or 3D space. Recent 3D native diffusion models [7, 21, 25, 67, 76] indicate promising results. Clay [76] designs a transformer-based [45, 60] 3D diffusion framework and various conditions like images and point clouds can be injected through cross-attention layers. A recent work MVControl [30] realizes 3D generation with single-view 2D spatial conditions like normal maps and edge maps by conditional multi-view generation and 3D reconstruction. Our work shares the similar workflow, but we focus on skeletal conditioned generation, which is not studied in MVControl. Another recent work AnimatableDreamer [65] generates 4D objects with skeletons extracted from given" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.37 + ], + "angle": 0, + "content": "videos by canonical score distillation, while our work develops a totally different framework for instant generation. As for skeletal representation, AnimatableDreamer extracts dense joints and unstructured bones from videos, while our method uses sparse joints and structured bones, which are more similar to human-made skeletons and more efficient for shape representation and animation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.909, + 0.6 + ], + "angle": 0, + "content": "Mesh Skeletonization. Various algorithms were designed for extracting skeletons from 3D meshes. [58] and [1] compute curve skeletons (C-S) via iterative mesh contraction operations. [11, 66] proposed to extract skeletons medial axis transformation skeleton (MAT-S) by point selection and connection prediction. C-S and MAT-S can serve as shape representation, while human-made skeletons (H-S) are often different from them. Since the main purpose is animation, H-S only contain sparse joints and bones. Some works [70, 71] propose data-driven approaches to learn mesh skeletonization from human annotated data. In this work, we have tried learning-based method [71] but found the results were not satisfactory. Therefore, we develop a new pipeline to generate skeletons which are as sparse as H-S while keep the shape of C-S." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.719, + 0.629 + ], + "angle": 0, + "content": "3. Dataset Construction" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.686, + 0.654 + ], + "angle": 0, + "content": "3.1. Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The largest existing open dataset containing mesh-skeleton pairs is ModelResources [70]. There are around 3,000 3D meshes without textures. The scale is insufficient to train a text-driven generative model, and it lacks textures for appearance modeling. To address these limitations, we construct a dataset with \\(8 \\times\\) larger scale and includes color textures. Our dataset, named Objaverse-SK, is built upon a large-scale 3D dataset Objaverse [8]. Although our data generation pipeline is applicable to a broad range of object categories, we focus on three main categories including \"Animals\", \"Human Shapes\" and \"Plants\", as they can typically be represented by tree-structured skeletons. Category labels are obtained from G-Objaverse [47]. Consequently, our dataset contains 24k 3D meshes, consisting of 15k animals, 6k human shapes and 3k plants. Text prompts for these models are generated by Cap3D [36]." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "316" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.094, + 0.88, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Illustration of the pipeline for skeletal conditioned multi-view and 3D generation (§4). The 3D skeleton is projected into 2D images and represented by coordinate color encoding. The skeletal correlation is modeled by skeletal correlation module, and then is injected into the diffusion model. Multi-view images are first generated and then a 3D textured mesh is reconstructed. The texture is further refined through UV-space optimization. Our framework achieves instant and high-quality generation given arbitrary skeletons." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.361, + 0.285, + 0.375 + ], + "angle": 0, + "content": "3.2. Skeleton Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.55 + ], + "angle": 0, + "content": "To obtain mesh-skeleton pairs, a effective method for generating skeletons from meshes is crucial. There are two concerns: the skeleton structure and success rate. The skeleton structure should accurately describe the object anatomy and be suitable for posing. Furthermore, an ideal method should generate reasonable skeleton structures with a high success rate. We tested a learning-based method RigNet [71] (Fig. 4). Although the generated skeleton structures can be close to human-made skeletons in its training data, it primarily produces symmetric skeletons and tends to be unstable across diverse anatomies and poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Skeleton extraction. To enhance flexibility and robustness, we design a new reliable pipeline, utilizing curve skeletons as the intermediate representation. An illustration of the pipeline is in Fig. 2. Given the structural inconsistency between curve skeletons and human-made skeletons, we further convert dense curves into sparse joints and bones. The detailed pipeline is elaborated below. 1) Initially, Mean Curvature Flow (MCF) [58] is employed to generate curve skeletons from meshes robustly. 2) Next, we build a graph from the set of curves generated from the mesh, consisting of dense nodes and edges. Intersection nodes (degree \\(>1\\)) are identified and the graph is divided into several parts by these nodes. 3) In each part, the curve contains no branches so it can be simplified by the Douglas-Peucker algorithm (DP) [12] into line segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tree conversion. At this stage, the basic shape of the skeleton is established, but the root position and the bone directions between joints still need to be determined. The problem can be regarded as a graph-to-tree conversion. First, a spanning tree is constructed from the graph to eliminate cycles. We then identify high-degree intersection nodes as candidates for the root. To ensure an efficient structure, the skeleton is configured by selecting the tree with the min" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.422 + ], + "angle": 0, + "content": "imum height among these candidates. This approach ensures that the root node is located at a significant intersection, minimizing the distances between the root and other joints. More details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.706, + 0.457 + ], + "angle": 0, + "content": "4. Generation Pipeline" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.89, + 0.482 + ], + "angle": 0, + "content": "4.1. Skeletal Conditioned Multi-view Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.595 + ], + "angle": 0, + "content": "As the dataset is constructed, we consider building the conditional generative model based on it. Since unconditional multi-view diffusion models have been well studied, we start directly from a base model MVDream [55] and focus on the conditional generation. Two main issues are of concern: i) how the skeleton is represented, and ii) how it is injected into the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Skeletal condition representation. As we aim to generate images which align with the given skeletons, using spatial guidance in the diffusion model is a reasonable approach. Skeletons can be projected from 3D space to the image plane as 2D conditions. However, overlapping and information loss occur during the projection, which may lead to semantic and structural ambiguities for spatial guidance, as illustrated in Fig. 8. Thus, incorporating richer information is crucial to mitigate these ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Coordinate Color Encoding (CCE). In order to preserve 3D information, we encode joint coordinates using spatial colors. While prior works [28, 61] use canonical color maps for shape representation, our approach focuses on representing skeletons with sparse joints and bones. We begin by normalizing skeletons within a canonical cube \\([0,1]^3\\). Each position in this cube corresponds to a unique color, with RGB values precisely matching the positional coordinates. As a result, the 2D conditional image can represent the 3D spatial positions of the skeleton. For bones, we assign colors based on their midpoint. Additionally, we incor" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "porate normalized values of view-dependent inverse depth of the skeleton as the alpha channel (CCE-D). With the absolute spatial coordinates and relative depth encoded in the conditional images, there will be more precise and richer guidance information for generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.333 + ], + "angle": 0, + "content": "Skeletal condition injection. Spatial conditions such as canny edges and normal maps have been investigated in 2D image diffusion models. In ControlNet [75], the conditional image is encoded by convolution blocks, resulting in an output spatial size that matches the latent size. Then, the condition features are added to the latent features. The encoder of the original diffusion model is copied as a side network to produce guidance features, which are fused with the original features in the decoder. Our pipeline adopts this paradigm from ControlNet, and further enhances it with a more effective condition feature extraction module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.654 + ], + "angle": 0, + "content": "Skeletal Correlation Modeling (SCM). For a skeleton in 3D space, we first project it into multi-view images as 2D conditions. Given the sparse nature of skeletons in spatial dimension, convolution blocks lack global modeling capacity. To address this, we design a Skeletal Correlation Module (SCM) to enhance the condition features by modeling the anatomical correlation among different parts of a skeleton, and the view correlation for different projection views. The structure of the module is in Fig. 3. (i) First, anatomical correlation is extracted by a self-attention layer, which constructs the global skeleton features for each view. (ii) Then, the cross-view correlation is modeled by a cross-attention layer, allowing the extraction of correspondences among skeleton images from multiple views. This enables the model to recognize identical joints in different views. In addition, we use adaptive layer normalization [68] to fuse the camera pose embedding with the skeletal features. Associating each skeleton image with a camera pose helps to generate view-dependent shapes. Adding correlation layers during condition encoding significantly facilitates learning, achieving \\(5 \\times\\) faster convergence (Fig. 9)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.409, + 0.681 + ], + "angle": 0, + "content": "4.2. Multi-view Images to 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Instant reconstruction. Given the generated multi-view images, we use a Large Reconstruction Model (LRM), specifically InstantMesh [69] for fast textured mesh reconstruction. However, the reconstructed textures often appear blurry. On one hand, the resolution of generated images is \\(256^2\\), which struggles to capture fine details. On the other hand, the appearance quality also degrades during reconstruction. In order to recover and further enrich the appearance, we introduce a new refinement stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Appearance refinement. First, the generated multi-view images are upscaled 4 times into \\(1024^2\\) by Stable Diffusion with ControlNet-Tile [75]. ControlNet-Edge [9, 75] is used to maintain the shape consistency across different views during tiling. Once tiled, these high-resolution im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ages \\( I_{i}^{h} \\) are used to refine the reconstructed texture. A learnable 2D texture \\( u \\) in UV space is created and initialized as the reconstructed texture \\( u_{0} \\), and then images are rendered through differentiable rendering \\( \\mathcal{R}(u, c_{i}) \\) for given camera views \\( c_{i} \\). The MSE loss is optimized between the rendered images and tiled high-res images. Moreover, a regularization term is added to maintain consistency in UV space:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.21, + 0.905, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\sum_ {i} \\left| \\left| I _ {i} ^ {h} - \\mathcal {R} (u, c _ {i}) \\right| \\right| _ {2} ^ {2} + \\lambda * \\left| \\left| u - u _ {0} \\right| \\right| _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.402 + ], + "angle": 0, + "content": "Since the high-res images cannot cover every position on the mesh, some regions of \\( u \\) will not be optimized, e.g. bottom of the object. We found these regions are unstable during optimization and may produce unexpected artifacts (see Fig. 10). The regularization term helps the optimized texture maintain the appearance from \\( u_{0} \\) in these regions. Consequently, the high-frequency details can be learned in covered regions while the global consistency can also be achieved in uncovered regions. The optimization can be finished within 15 seconds." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.415, + 0.847, + 0.431 + ], + "angle": 0, + "content": "4.3. Object-Skeleton Alignment Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Contrastive alignment. In order to measure how much an object is aligned with a skeleton, we develop a new evaluator, named Contrastive Object-Skeleton Alignment (COSA). We use the self-supervised DINOv2 [44] as the backbone \\(F\\) to extract both object and skeleton features. Then, the alignment adapter \\(G_{\\theta}\\) consisting of several self-attention layers is used to modulate the features. The adapter ends with an average pooling layer to aggregate the aligned features into a vector. Similar to CLIP [49], we train the adapter using contrastive learning by InfoNCE loss [42, 56]. Finally, the skeleton alignment score (SKA) can be calculated by cosine similarity between the features from an object image \\(x\\) and a skeleton image \\(y\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.648, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {\\mathrm {S K A}} (x, y) = \\cos (G (F (x)), G (F (y))). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.907, + 0.813 + ], + "angle": 0, + "content": "COSA guided diffusion. Based on COSA, another conditional generation pipeline can also be realized, following the approach proposed in [2]. On each denoising time step \\( t \\), the approximate clean image \\( \\hat{x}_0 \\) is estimated from the predicted noise \\( \\epsilon_t \\) as in DDIM [57]. The estimated clean image and skeleton condition are fed into COSA to calculate the alignment loss \\( \\mathcal{L}_{\\mathrm{COSA}}(\\hat{x}_0,y) = 1 - S_{\\mathrm{SKA}}(\\hat{x}_0,y) \\). Then the predicted noise is modified by the gradient of the alignment loss for actual denoising:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.827, + 0.905, + 0.842 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {t} = \\epsilon_ {t} + s (t) \\cdot \\nabla \\mathcal {L} _ {\\mathrm {C O S A}} (\\hat {x} _ {0}, y) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\( s(t) \\) controls the guidance strength. With the additional guidance of the alignment loss, the generated object will tend to follow the conditional skeleton \\( y \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "318" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.898, + 0.176 + ], + "angle": 0, + "content": "
Method\\SKA Score TrainingMeanInst.MeanClassAnimalsHumansPlantsApodesBipedsQuadrupedsArthropodsWings
SDEdit [38]○70.1365.0679.5064.7050.9974.7475.2482.7379.0779.21
SDEdit+COSAG●72.1167.3280.9167.6053.4679.7383.1785.9877.8073.82
Ours-Raw●74.6967.0988.3969.9042.9794.7482.1490.1889.8184.69
Ours-SCM●80.4374.3891.1678.4553.5394.4785.2094.1990.6888.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.18, + 0.908, + 0.222 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of object-skeleton alignment (SKA) score (\\$5.2). Alignment scores are calculated over three classes (blue) and five subclasses of animal (green). The average scores over all instances and three classes (pink) are also shown. The highest scores among all methods are bold and the highest scores among baseline methods are underlined." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.235, + 0.846, + 0.334 + ], + "angle": 0, + "content": "
MethodTrainingPickScoreCLIP Score
Win RateAnimalsHumanPlants\\( Mean_{Inst.} \\)AnimalsHumanPlants
SDEdit [38]○19.5618.2314.6628.9829.0429.4728.7128.29
SDEdit+COSAG\\( \\blacksquare \\)18.1718.5419.9515.0628.9929.5228.5028.11
Ours-Raw●28.9929.0628.3729.5529.9030.1030.0229.23
Ours-SCM●33.2834.1737.0226.4229.8330.2329.8428.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.336, + 0.908, + 0.365 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of PickScore and CLIP Score (§5.2). Scores are calculated over three classes (blue) and averaged over all instances (red). The highest scores among all methods are bold and the highest scores among baseline methods are underlined." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.377, + 0.469, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Figure 4. Comparison of skeletons generated from 3D meshes by RigNet [71] and our method (§5.1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.582, + 0.218, + 0.599 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.373, + 0.621 + ], + "angle": 0, + "content": "5.1. Results of Mesh Skeletonization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.75 + ], + "angle": 0, + "content": "We compare our method with the learning-based method RigNet [71], and the results are shown in Fig. 4. RigNet tends to produce symmetric skeletons with limited flexibility and misaligned joints/bones, resulting in an around \\(15\\%\\) success rate. In contrast, our method supports arbitrary-pose skeletons and achieves better joint/bone alignment. It produces more reliable results with an \\(80\\%\\) success rate. More details and results can be found in appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.383, + 0.773 + ], + "angle": 0, + "content": "5.2. Results of Multi-view Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocols. The Objaverse-SK evaluation set contains 108 skeleton instances, covering three main classes: animals, human shapes and plants. As animals include diverse skeleton structures, we further divide them into more detailed subclasses (examples are shown behind): Apodes (fish, snakes), Bipeds (ducks, penguins), Quadrupeds (dogs, bears), Arthropods (scorpions, crabs), Wings (birds, dragons). Three evaluation metrics are con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.47 + ], + "angle": 0, + "content": "sidered for multi-view generation: SKA Score for skeletal alignment, PickScore for image quality, and CLIP Score for textual alignment. Furthermore, samples from categories excluded in training are obtained from ShapeNet [5] for generalization evaluation (Fig. 7). Additionally, evaluation results on human-made skeletons are in appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Baseline methods. Since there is no prior work that can achieve arbitrary skeletal conditioned generation, we implement two methods for comparison. The first baseline is SDEdit [38]. The process starts from adding noise on conditional images with a time step (set as 0.7). Then clean images are generated by denoising steps. The method is entirely unsupervised. The second baseline is the COSA Guidance (COSAG) derived from [2], which is elaborated in Section 4.3. The guidance strength is set as \\( s(t) = 7.5\\sqrt{1 - \\alpha_t} \\). Since we found it cannot achieve stable results, it is combined with SDEdit. The method requires an extra model so it is half supervised. Ours is fully supervised on object-skeleton pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.822 + ], + "angle": 0, + "content": "Qualitative comparison. The qualitative results are shown in Fig. 5. Given skeleton images as conditions, SDEdit can produce images following the skeleton. However, limited by the editing capacity, the generated objects often have incorrect anatomies. For example, the snake body is broken, and the donkey body is generated as wooden. When it is enhanced by the COSAG, the quality of generated content is improved in some cases but still unsatisfactory. Compared with them, our results show superior quality and are more consistent with both skeletal and textual conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative comparison. Comparison results of skeleton alignment are shown in Table 1. SDEdit-based methods have around 70 SKA scores, while ours can achieve 80. Among three classes, animals tend to have higher alignment scores while plants have lower scores. Since plants" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "319" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.387, + 0.908, + 0.417 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of skeletal conditioned multi-view generation (§5.2). Conditional skeletons and text prompts are shown above. Four views are generated and two views are shown for simplicity." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.422, + 0.885, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.572, + 0.908, + 0.601 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of textured meshes before and after refinement (§5.3). Rendered color images are shown. Local areas are enlarged for better viewing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.69 + ], + "angle": 0, + "content": "may have more complex structures and sometimes extend further from skeletons. For five sub-classes, our method achieves consistently high alignment scores. With the help of SCM and CCE-D skeletons, both the alignment scores and pick scores are further improved." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.322, + 0.716 + ], + "angle": 0, + "content": "5.3. Results of 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.484, + 0.8 + ], + "angle": 0, + "content": "Texture refinement. Results of 3D reconstruction from multi-view images are shown in Fig. 6. The raw reconstructed and refined results are compared. The raw textures are blurry and lack details, while the proposed refinement stage can significantly enhance texture quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Rigging and animation. Given a motion sequence of a skeleton, our method can be applied to generate 4D animation. Textured mesh can be generated given the skeleton at the rest pose, and then directly be rigged and skinned for animation. Demo videos can be found in project page1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.613, + 0.666, + 0.631 + ], + "angle": 0, + "content": "6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.777 + ], + "angle": 0, + "content": "Skeletal condition representation. The skeletal condition representation we use consists of coordinate color encoding (CCE) with depth alpha (D). The ablation results are shown in Fig. 8 and Tab. 3. Richer information in conditions can help the model determine the content better. As a result, better alignment can be achieved. In Fig. 8 right, the skeleton of a penguin is highly ambiguous. If CCE-D is used, the body pose and orientation of the penguin can be successfully inferred from colors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Skeletal correlation modeling. We show the effect of our skeletal correlation module in Tab. 3 and Fig. 9. SCM achieves better alignment scores than convolutional blocks. SCM with layer normalization (LN) achieves a \\(4 \\times\\) faster convergence rate. Furthermore, if LN is replaced with the adaptive LN (AdaLN), the model can achieve an SKA score of 70 within 1k training steps. The results indicate that extracting global features from conditional images is crucial." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.27, + 0.901 + ], + "angle": 0, + "content": "1https://skdream3d.github.io/." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "320" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.09, + 0.473, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of novel categories in ShapeNet [5] (§5.2), which are not covered by the training set of Objaverse-SK." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.413, + 0.47, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.582 + ], + "angle": 0, + "content": "Figure 8. Ablation study of coordinate color encoding with depth alpha (CCE-D) (\\(\\S 6\\)). Richer information can help the model to avoid ambiguity and generate better anatomy." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.604, + 0.489, + 0.698 + ], + "angle": 0, + "content": "
ModuleSkeletonSKA ScoreAnimalsHumansPlants
SCMCCE-D80.4391.1678.4553.53
SCMCCE78.9790.4574.8252.57
SCMBinary77.2789.2076.0446.16
ConvCCE-D76.6588.5375.0446.17
ConvBinary74.6988.3969.9042.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Table 3. Ablation study of skeletal module and representation types (§6). SCM and CCE-D achieve higher alignment scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Refinement regularization. We show the ablation results of appearance refinement in Fig. 10. The refined appearance contains rich details such as snake scales and wood grain, in comparison to the reconstructed results. However, artifacts also appear in the regions that are not covered by high-res images. With the help of UV space regularization, the artifacts are effectively removed in uncovered regions. As a result, consistent colors are maintained from original textures and details are enhanced during optimization." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.093, + 0.881, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.221, + 0.905, + 0.25 + ], + "angle": 0, + "content": "Figure 9. Convergence processes of different skeletal modules (§6). SCM with AdaLN achieves 5x faster convergence." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.263, + 0.692, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.264, + 0.89, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.402, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 10. Ablation study of UV space regularization (§6). Bottom views and UV textures are shown. Front views of the snake and the tree stump can be found in the first column of Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.457, + 0.779, + 0.473 + ], + "angle": 0, + "content": "7. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Although our work achieves arbitrary skeletal conditioned generation, there are still many problems that can be further studied. The skeletons we currently use may have limited descriptive ability for non-tree structured objects. More powerful shape representations can be studied as new conditions. In addition, our work only considers global skeletons without fine-grained semantics. Injecting detailed semantics into the skeleton parts is a meaningful topic to study. For future work, the skeleton condition can be studied in native 3D generation frameworks. Moreover, our dataset may also be used in other tasks such as arbitrary skeleton estimation from images. More discussion is in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose to use skeletons as the structural condition for controllable generation. First, we construct a large-scale 3D mesh-skeleton paired dataset. We propose an effective mesh skeletonization method to generate mesh-aligned sparse skeletons with a high success rate. Based on the dataset, we present a skeletal conditioned multi-view generation pipeline. Coordinate color encoding and skeletal correlation module are proposed to realize efficient condition representation and injection. Furthermore, 3D meshes can be instantly reconstructed, followed by a refinement stage to achieve better texture quality. In summary, our work achieves controllable multi-view and 3D generation with arbitrary skeletons as conditions." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "321" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (No. 62293554), and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02) and the Fundamental Research Funds for the Zhejiang Provincial Universities (226-2024-00208)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.483, + 0.27 + ], + "angle": 0, + "content": "[1] Andreas Bærentzen and Eva Rotenberg. Skeletonization via local separators. ACM Transactions on Graphics (TOG), 40 (5):1-18, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.271, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[3] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.398, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[6] Cheng Chen, Xiaofeng Yang, Fan Yang, Chengzeng Feng, Zhoujie Fu, Chuan-Sheng Foo, Guosheng Lin, and Fayao Liu. Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10228-10237, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[7] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[9] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[10] Wenqi Dong, Bangbang Yang, Lin Ma, Xiao Liu, Liyuan Cui, Hujun Bao, Yuewen Ma, and Zhaopeng Cui. Coin3d: Controllable and interactive 3d assets generation with proxy-guided conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Zhiyang Dou, Cheng Lin, Rui Xu, Lei Yang, Shiqing Xin, Taku Komura, and Wenping Wang. Coverage axis: Inner point selection for 3d shape skeletonization. In Com" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "puter Graphics Forum, pages 419-432. Wiley Online Library, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 10 (2):112-122, 1973. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[13] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Navigation instruction generation with bev perception and large language models. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Scene map-based prompt tuning for navigation instruction generation. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[15] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.907, + 0.518 + ], + "angle": 0, + "content": "[18] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[19] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[20] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[22] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "322" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[26] Christoph Leiter, Ran Zhang, Yanran Chen, Jonas Belouadi, Daniil Larionov, Vivian Fresen, and Steffen Eger. Chatgpt: A meta-analysis after 2.5 months. arXiv preprint arXiv:2302.13795, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[29] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Zhiqi Li, Yiming Chen, Lingzhe Zhao, and Peidong Liu. Mycontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[31] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[32] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6400-6409, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[33] Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J Black. Tada! text to animatable digital avatars. In 2024 International Conference on 3D Vision (3DV), pages 1508-1519. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[34] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20763–20774, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[35] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[36] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[39] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[40] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[41] Marianna Ohanyan, Hayk Manukyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Zero-painter: Training-free layout control for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8764-8774, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.79, + 0.433 + ], + "angle": 0, + "content": "[43] OpenAI. Gpt-4 technical report, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[44] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[47] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[48] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "323" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[54] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[55] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[56] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[58] Andrea Tagliasacchi, Ibraheem Alhashim, Matt Olson, and Hao Zhang. Mean curvature skeletons. In Computer Graphics Forum, pages 1735-1744. Wiley Online Library, 2012. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[60] Ashish Vaswani. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[61] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[62] Jiajun Wang, Morteza Ghahremani, Yitong Li, Björn Omer, and Christian Wachinger. Stable-pose: Leveraging transformers for pose-guided text-to-image generation. arXiv preprint arXiv:2406.02485, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[63] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[64] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[65] Xinzhou Wang, Yikai Wang, Junliang Ye, Fuchun Sun, Zhengyi Wang, Ling Wang, Pengkun Liu, Kai Sun, Xintong Wang, Wende Xie, et al. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[66] Zimeng Wang, Zhiyang Dou, Rui Xu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Shiqing Xin, Taku Komura, Xiaoming Yuan, and Wenping Wang. Coverage axis++: Efficient inner point selection for 3d shape skeletonization. In Computer Graphics Forum, page e15143. Wiley Online Library, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[67] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[70] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, and Karan Singh. Predicting animation skeletons for 3d articulated models via volumetric nets. In 2019 international conference on 3D vision (3DV), pages 298-307. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[71] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[72] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[74] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[76] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[77] Muxin Zhang, Qiao Feng, Zhuo Su, Chao Wen, Zhou Xue, and Kun Li. Joint2human: High-quality 3d human generation via compact spherical embedding of 3d joints. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "324" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1438, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[79] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[80] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavator: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[81] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[82] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[83] Wenjie Zhuo, Fan Ma, Hehe Fan, and Yi Yang. Vividdreamer: invariant score distillation for hyper-realistic text-to-3d generation. In European Conference on Computer Vision, pages 122–139. Springer, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "325" + } + ] +] \ No newline at end of file diff --git a/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_origin.pdf b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d91c6b6c3f4a2ce8b8af3ab7c3ea3108dd30e8a4 --- /dev/null +++ b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/7b2caa84-1e0b-47fd-a90e-b20604143574_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a02007d974f9b92e22eeb747fcd3ceff3fdf8b7e157c4fcfe3d17361a62f627 +size 7644061 diff --git a/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/full.md b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44e3c8a9858e35d369416ed5b3d3d8b147de0517 --- /dev/null +++ b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/full.md @@ -0,0 +1,413 @@ +# SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons + +Yuanyou Xu $^{1}$ , Zongxin Yang $^{2}$ , Yi Yang $^{1*}$ $^{1}$ ReLER, CCAI, Zhejiang University + $^{2}$ DBMI, HMS, Harvard University + +{yoxu, yangyics}@zju.edu.cn, zongxin-yang@hms.harvard.edu + +# Abstract + +Controllable generation has achieved substantial progress in both 2D and 3D domains, yet current conditional generation methods still face limitations in describing detailed shape structures. Skeletons can effectively represent and describe object anatomy and pose. Unfortunately, past studies are often limited to human skeletons. In this work, we generalize skeletal conditioned generation to arbitrary structures. First, we design a reliable mesh skeletonization pipeline to generate a large-scale mesh-skeleton paired dataset. Based on the dataset, a multi-view and 3D generation pipeline is built. We propose to represent 3D skeletons by Coordinate Color Encoding as 2D conditional images. A Skeletal Correlation Module is designed to extract global skeletal features for condition injection. After multi-view images are generated, 3D assets can be obtained by incorporating a large reconstruction model, followed by a UV texture refinement stage. As a result, our method achieves instant generation of multi-view and 3D contents that are aligned with given skeletons. The proposed techniques largely improve the object-skeleton alignment and generation quality. Project page at https://skdream3d.github.io/. + +# 1. Introduction + +In view of visual representation dimension [64, 72], 2D image generation [16, 53, 57], multi-view (2.5D) generation [35, 55], and 3D generation [17, 27, 69] have advanced successively and achieved significant progress. Given the success of large language models [23, 26, 43, 48-50], textual representation has been widely applied in generation [51-53], as in other domains [13, 14, 73]. To achieve more controllable generation, conditions beyond text have attracted considerable attention. 2D image conditions (e.g., edge maps, human skeletons, and concept references) [54, 75] have been well studied. Similarly in 3D generation, + +2D [30] and 3D [10] conditions have also been explored. + +Although the aforementioned conditions in controllable generation complement text descriptions, they still struggle to precisely describe shape structures. In contrast, skeletons, among various types of conditions, exhibit superior ability to depict shape structures: (i) Representation of object anatomy. Skeletons can efficiently represent various 3D structures with sparse joints and bones, while other conditions like depth maps fail to represent full 3D anatomy. (ii) Articulation into different poses. Skeletons are widely used for character animation in computer graphics [3, 22] due to their simplicity and efficiency. Other conditions such as rough shapes [10] are inconvenient for pose articulation. (iii) Freedom of editing. Given an initial skeleton, users can freely add new structures or modify joints and bones to create ideal shapes. Examples are in Fig. 1. + +Despite these advantages, previous studies [19, 20, 40, 75, 77] on conditional generation are limited to human skeletons. From the perspective of generalization, we would like to ask: Is it possible to use arbitrary skeletons as conditions to generate any creatures or even general objects? + +To achieve this goal, we identify two main challenges that hinder the use of arbitrary skeletal conditions for generation: (i) Lack of large-scale object-skeleton pairs for training. Extensive studies [4, 15, 37] on 2D/3D human pose estimation have made human-skeleton paired data readily available. However, when skeletal structures are unknown, estimating arbitrary skeletons from 2D images or videos becomes challenging due to its ill-posedness. (ii) Insufficiency of 2D information to describe arbitrary skeletons. Human skeletons are relatively simple and can be described by a fixed set of 2D joints. However, complex skeletons suffer from self-occlusion and ambiguity, necessitating 3D information to fully capture their anatomy and pose. + +To address these challenges, we focus on multi-view and 3D generation with skeletal conditions. To tackle data scarcity problem, we construct a large-scale dataset Objaverse-SK containing mesh-skeleton pairs. Textured meshes are selected from Objaverse [8] by semantic classes to form a subset. In order to achieve reliable mesh ske + +![](images/0afbb1f749cc0cbdc98dd5fe1dd01e3833dcfac37e9e2d24fecbc9d1b4cea17e.jpg) + +![](images/25c410ee255a8d99ceb9eb74a7b4357fa1c678b5f5a29bd5d824ce0d3be3c97f.jpg) + +![](images/a00d0e26d78e1918eeb6ea317f26a1c7200b4a26fae4c1b3af86d0b38a52d567.jpg) + +![](images/7efd391b91e4f5b2172481bd2da9d2579f92f7892611ffd2c56edb7546ce6761.jpg) +MVDream +sno + +![](images/08ef72300a46dfc3d905c8b346e18f2cc6cadd4170e55d58521c9d6fc21dd0bf.jpg) +Text & Skeleton +Dog + +![](images/7d1dd94c0d63234511a5c7dae5bc680c9bf735fd5334459cfd674d3bb3426607.jpg) +Chair + +![](images/85cef554b554ce5b10bd75589395a08b6ae6445dd6af76ba88f7aa33fec4903c.jpg) +Fish + +![](images/9a2c1116ee6165bb09ef7a2a314dd32a625e9e2e1aa53806af3b2ffa7cc0242a.jpg) + +![](images/fac99c2eda3a17b8cea6a79ee43f229656a38d3154cb3cfd59382e6b6d3bc1a7.jpg) +Text Only + +![](images/86f54caa4684bef521224cc22a98f6936016b5f5e5c0b2ec223b56c19b84c546.jpg) +Spiderman +(a) Text and skeletons for Collaborative Appearance and Shape Controlling + +![](images/97fb40a7cb2d5e27595631d7ba02d4ed43bccb98fe3cb97d6d7f2a0198c7c4a5.jpg) + +![](images/3d04219a1222971b64d7ed2afd444c0aa9534bc273a276fb105214ee73956ed6.jpg) +Rocket + +![](images/103aa145f1228f6b48ff449e647403c6553b8aed7795fddfe1e269a341cd390c.jpg) + +![](images/a0877861603c8471d23e49169038a25b0b54c62afb407571e603d8d26e468888.jpg) +Ice cream + +![](images/5d902562c5f9e00063d794b49712c15baaafb81b11a3fe05b91557b84c2e7699.jpg) +Text Only + +![](images/1c9578f61a2a517c76c5eb64f5a610b9c0b21cb263b58b9d58eff42b6b04ba4c.jpg) + +![](images/3a1483b0375fff0ee8bb5469c4d2b1f3fc21183cf9c480627a00e0720e85508e.jpg) + +![](images/486e1a8da97c2665e6b861e300ef360f11ba289943612568a23057521eb15dea.jpg) + +![](images/b15b2f8bda631ef6a647b4608b255c9cc83f9d485b9ff6e4e2dfbdb670e8e909.jpg) + +![](images/7578f5e542481c324c0f22cafcb787af1833f34491cb74657c9d4465181b6ab9.jpg) +Coral +Torch + +![](images/c9aeee3ae7bcf98ba3fa96ba6ff877d63c8e3c724c4375312534f199325100a7.jpg) + +![](images/32047747d873f67b55ee937c9d8c97392b1ec19707c96567036be54d839fbd4c.jpg) +Animal + +![](images/cb0b3004e56923e389338ab80c6d5776f6da88f07a713d3a4dcdece2ccbcde25.jpg) +Human + +![](images/8ae4fdc0d04a155b5e87d6ea27e595175b36a569f9eddd1f644cecfcdf54a513.jpg) +Plant + +![](images/12b88a4c2585954215fd56756cce4d6fcc163aea5cb7ebc6b6874f5e2672a63b.jpg) +Initial skeleton +Adding joints, bones +Dog with a spiral tail + +![](images/b797d4f5e3878ba89cb41acbf3da3b0e5d15fff47d67dd1eaf5013a4fbbe9afa.jpg) +Dog with two heads + +![](images/48309fdfc7f776ab20d0759186ac394e4e2bd1d2e077df7c543ec422f9ca8bbb.jpg) +Knight with dragon wings + +![](images/b502d02c4974d33a9e46cde3bcb701ec60ed8f0d20b313d7b7713b8322c24696.jpg) +Plant with leaves and roots + +![](images/afd9e106bda4954b4111f6d1cf65f20c5391384583a2604b90386aa697e11d50.jpg) +Monster with four arms + +![](images/58744925b1b2268c9169ee3c10ba29fbe46a2b969d6767255efdfdbad4078f1d.jpg) +Plant with legs, Pokémon + +![](images/6232048d19425b99adeac4a279bbef9c419e3322346843b3c53fdcf3a5c204cc.jpg) +Modifying joint positions, bone sizes +Dog with a big head + +![](images/f50b45ce4bc9454e0e5d853d778f0ecdbfd70edd5af9ba1b852ee74c0ffaad62.jpg) +Dog with a long torso + +![](images/62e188a77cdc1da2f3b9b2df854c7b7819c038dffcbd491fddc0b05b87780fb0.jpg) +Dwarf in armor + +![](images/f026946b1d9f0cd73d3c3fa3960efcc9e70e1243bd0333d162fad5a442ac6630.jpg) +How wide? +Wide shoulder bodybuilder + +![](images/ae20cba4474432e47284557cfd31e28d00f3526ab8759079a995a780fa0bb06e.jpg) +Tree with wide branches + +![](images/c9e78a8b57ff69b3e7b45b36dbbc3c051e76265c5c82a92e0ffb035c7f468487.jpg) +Tree with a long truck +(b) Skeleton-based editing for Accurate Anatomy and Pose Controlling + +![](images/50faf0ff8fbf485697abc3de93af8a61a39aa191b84f1f103b677761eea1f0a7.jpg) +Posing +Dog jumping + +![](images/81a0f19445b6a834f306eb385cbe4c924f19a1c1d5cde68f4e9cd0a6223fb843.jpg) +Dog sitting + +![](images/1da94bcb445dc3678fceafce0d48bcd1097608031a535a16b3b52cddd7a80469.jpg) +Athlete running + +![](images/48fa0c23ae145450701724393702415f99df007b16c7c30f97749fb8dfdc3077.jpg) +Master lunging + +![](images/d51661c81751385a118c3f26220095c524d92fcc7b3bc62b6d73e701ed983dfb.jpg) +Downward leaves + +![](images/61f041eff042f551a3c8d666abfc62debdd211dd99c8150d3f0683cb0261cb9d.jpg) +Bent stem +Figure 1. Demonstration of skeletal conditions for controllable generation. We argue that skeletons and text provide complementary description for shape and appearance respectively, as shown in (a). Moreover, flexible and accurate control of object anatomy and pose can be realized by editing the joints and bones in skeletons, as shown in (b). Arbitrary skeletal structures are supported in our framework. Multiple views are generated and only front-view images are shown. + +tonization, we propose a new pipeline to generate skeletons with sparse joints from meshes. The pipeline mainly consists of curve skeleton extraction and curve simplification, achieving an $80\%$ success rate, significantly outperforming previous deep learning based method RigNet [71]. + +To fully control object anatomy and pose, we build the skeletal conditioned generation model in a multi-view manner. We represent a 3D skeleton with conditional skeleton images by Coordinate Color Encoding (CCE) to reduce ambiguity. Joints and bones are encoded with unique colors according to their 3D positions. For condition injection, we designed a Skeletal Correlation Module (SCM) to extract features from these conditional images and then generate multi-view images for the object. Later, a Large Reconstruction Model (LRM) is employed to produce 3D assets from the multi-view images. To address potential blurriness during reconstruction, we enhance appearance quality by a refinement stage that up-samples the multi-view images to higher resolutions and refines the texture in UV space. + +The experimental results indicate that our framework achieves instant generation of multi-view and 3D contents which are aligned with given skeletons. The proposed coordinate color encoding and the skeletal correlation module significantly improve the object-skeleton alignment, and accelerate model convergence by $5 \times$ 3D assets conforming to the given skeleton can be generated in $\sim 20\mathrm{s}$ and refined in $\sim 60\mathrm{s}$ . To the best of our knowledge, this work is a pioneer in achieving arbitrary skeletal conditioned generation with following contributions: + +- Constructing the first large-scale dataset, Objaverse-SK, containing mesh and skeleton pairs that cover diverse skeletal structures. We developed a pipeline to generate sparse skeletons from meshes with a high success rate. +- Proposing a multi-view and 3D generation pipeline for arbitrary skeletons, including coordinate color encoding for compact condition representation and the skeletal correlation module for effective condition injection. + +![](images/bb419e83bedf16bb7189b8f83e0755967f849e5aa5905b801fed4f5e8d554f48.jpg) +Figure 2. Illustration of the pipeline for skeleton generation from meshes (§3.2). The curve skeleton is first extracted from the given mesh, followed by simplification of parted curves. The curve graph is converted to a tree as the final skeleton. + +# 2. Related Work + +Controllable 2D Generation. Based on image diffusion models [51-53], versatile controlling conditions have been studied. In terms of spatial control, ControlNet [75] and other similar works [40, 78] train a side network for spatial conditions such as edge maps, normal maps and human skeletons. Some works focus on human image generation from skeletons [18, 20, 62]. Box-based instance control is also concerned in some works [29, 79, 81]. As for conceptual controlling, DreamBooth [54] finetunes the model to bind the given subject with an identifier in text prompt. IP-Adapter [74] trains an adapter to inject styles or concepts into the model. Some works [31, 32, 63] also focus on human ID control. Besides, some methods [2, 38, 39, 41] can achieve conditional generation without fine-tuning. + +Controllable 3D Generation. Controlling concepts in 3D generation can be easily achieved by combining image control methods with image-to-3D generation [17, 27, 59, 69]. However, spatial control for 3D generation is not as easy as conceptual control. Many works achieve spatial control based on score distillation paradigm [28, 46, 83]. Coin3D [10] presents a framework to control the multi-view diffusion and 3D generation by shape proxies, i.e. combination of simple basic shapes. Sculpt3D [6] enhances text-to-3D generation with retrieved 3D priors. Sherpa3D [34] proposes to generate a coarse shape with a 3D diffusion model and refine the shape with SDS [46]. Some works for 3D human or head avatar generation [19, 24, 33, 77, 80, 82] uses human skeleton or facial landmarks as the condition in 2D or 3D space. Recent 3D native diffusion models [7, 21, 25, 67, 76] indicate promising results. Clay [76] designs a transformer-based [45, 60] 3D diffusion framework and various conditions like images and point clouds can be injected through cross-attention layers. A recent work MVControl [30] realizes 3D generation with single-view 2D spatial conditions like normal maps and edge maps by conditional multi-view generation and 3D reconstruction. Our work shares the similar workflow, but we focus on skeletal conditioned generation, which is not studied in MVControl. Another recent work AnimatableDreamer [65] generates 4D objects with skeletons extracted from given + +videos by canonical score distillation, while our work develops a totally different framework for instant generation. As for skeletal representation, AnimatableDreamer extracts dense joints and unstructured bones from videos, while our method uses sparse joints and structured bones, which are more similar to human-made skeletons and more efficient for shape representation and animation. + +Mesh Skeletonization. Various algorithms were designed for extracting skeletons from 3D meshes. [58] and [1] compute curve skeletons (C-S) via iterative mesh contraction operations. [11, 66] proposed to extract skeletons medial axis transformation skeleton (MAT-S) by point selection and connection prediction. C-S and MAT-S can serve as shape representation, while human-made skeletons (H-S) are often different from them. Since the main purpose is animation, H-S only contain sparse joints and bones. Some works [70, 71] propose data-driven approaches to learn mesh skeletonization from human annotated data. In this work, we have tried learning-based method [71] but found the results were not satisfactory. Therefore, we develop a new pipeline to generate skeletons which are as sparse as H-S while keep the shape of C-S. + +# 3. Dataset Construction + +# 3.1. Data Preparation + +The largest existing open dataset containing mesh-skeleton pairs is ModelResources [70]. There are around 3,000 3D meshes without textures. The scale is insufficient to train a text-driven generative model, and it lacks textures for appearance modeling. To address these limitations, we construct a dataset with $8 \times$ larger scale and includes color textures. Our dataset, named Objaverse-SK, is built upon a large-scale 3D dataset Objaverse [8]. Although our data generation pipeline is applicable to a broad range of object categories, we focus on three main categories including "Animals", "Human Shapes" and "Plants", as they can typically be represented by tree-structured skeletons. Category labels are obtained from G-Objaverse [47]. Consequently, our dataset contains 24k 3D meshes, consisting of 15k animals, 6k human shapes and 3k plants. Text prompts for these models are generated by Cap3D [36]. + +![](images/bc7efb92eb3dbc3c081563d6a8bc751ed0b04b53a2034ae245e8830e397fee18.jpg) +Figure 3. Illustration of the pipeline for skeletal conditioned multi-view and 3D generation (§4). The 3D skeleton is projected into 2D images and represented by coordinate color encoding. The skeletal correlation is modeled by skeletal correlation module, and then is injected into the diffusion model. Multi-view images are first generated and then a 3D textured mesh is reconstructed. The texture is further refined through UV-space optimization. Our framework achieves instant and high-quality generation given arbitrary skeletons. + +# 3.2. Skeleton Generation + +To obtain mesh-skeleton pairs, a effective method for generating skeletons from meshes is crucial. There are two concerns: the skeleton structure and success rate. The skeleton structure should accurately describe the object anatomy and be suitable for posing. Furthermore, an ideal method should generate reasonable skeleton structures with a high success rate. We tested a learning-based method RigNet [71] (Fig. 4). Although the generated skeleton structures can be close to human-made skeletons in its training data, it primarily produces symmetric skeletons and tends to be unstable across diverse anatomies and poses. + +Skeleton extraction. To enhance flexibility and robustness, we design a new reliable pipeline, utilizing curve skeletons as the intermediate representation. An illustration of the pipeline is in Fig. 2. Given the structural inconsistency between curve skeletons and human-made skeletons, we further convert dense curves into sparse joints and bones. The detailed pipeline is elaborated below. 1) Initially, Mean Curvature Flow (MCF) [58] is employed to generate curve skeletons from meshes robustly. 2) Next, we build a graph from the set of curves generated from the mesh, consisting of dense nodes and edges. Intersection nodes (degree $>1$ ) are identified and the graph is divided into several parts by these nodes. 3) In each part, the curve contains no branches so it can be simplified by the Douglas-Peucker algorithm (DP) [12] into line segments. + +Tree conversion. At this stage, the basic shape of the skeleton is established, but the root position and the bone directions between joints still need to be determined. The problem can be regarded as a graph-to-tree conversion. First, a spanning tree is constructed from the graph to eliminate cycles. We then identify high-degree intersection nodes as candidates for the root. To ensure an efficient structure, the skeleton is configured by selecting the tree with the min + +imum height among these candidates. This approach ensures that the root node is located at a significant intersection, minimizing the distances between the root and other joints. More details can be found in the appendix. + +# 4. Generation Pipeline + +# 4.1. Skeletal Conditioned Multi-view Generation + +As the dataset is constructed, we consider building the conditional generative model based on it. Since unconditional multi-view diffusion models have been well studied, we start directly from a base model MVDream [55] and focus on the conditional generation. Two main issues are of concern: i) how the skeleton is represented, and ii) how it is injected into the model. + +Skeletal condition representation. As we aim to generate images which align with the given skeletons, using spatial guidance in the diffusion model is a reasonable approach. Skeletons can be projected from 3D space to the image plane as 2D conditions. However, overlapping and information loss occur during the projection, which may lead to semantic and structural ambiguities for spatial guidance, as illustrated in Fig. 8. Thus, incorporating richer information is crucial to mitigate these ambiguities. + +Coordinate Color Encoding (CCE). In order to preserve 3D information, we encode joint coordinates using spatial colors. While prior works [28, 61] use canonical color maps for shape representation, our approach focuses on representing skeletons with sparse joints and bones. We begin by normalizing skeletons within a canonical cube $[0,1]^3$ . Each position in this cube corresponds to a unique color, with RGB values precisely matching the positional coordinates. As a result, the 2D conditional image can represent the 3D spatial positions of the skeleton. For bones, we assign colors based on their midpoint. Additionally, we incor + +porate normalized values of view-dependent inverse depth of the skeleton as the alpha channel (CCE-D). With the absolute spatial coordinates and relative depth encoded in the conditional images, there will be more precise and richer guidance information for generation. + +Skeletal condition injection. Spatial conditions such as canny edges and normal maps have been investigated in 2D image diffusion models. In ControlNet [75], the conditional image is encoded by convolution blocks, resulting in an output spatial size that matches the latent size. Then, the condition features are added to the latent features. The encoder of the original diffusion model is copied as a side network to produce guidance features, which are fused with the original features in the decoder. Our pipeline adopts this paradigm from ControlNet, and further enhances it with a more effective condition feature extraction module. + +Skeletal Correlation Modeling (SCM). For a skeleton in 3D space, we first project it into multi-view images as 2D conditions. Given the sparse nature of skeletons in spatial dimension, convolution blocks lack global modeling capacity. To address this, we design a Skeletal Correlation Module (SCM) to enhance the condition features by modeling the anatomical correlation among different parts of a skeleton, and the view correlation for different projection views. The structure of the module is in Fig. 3. (i) First, anatomical correlation is extracted by a self-attention layer, which constructs the global skeleton features for each view. (ii) Then, the cross-view correlation is modeled by a cross-attention layer, allowing the extraction of correspondences among skeleton images from multiple views. This enables the model to recognize identical joints in different views. In addition, we use adaptive layer normalization [68] to fuse the camera pose embedding with the skeletal features. Associating each skeleton image with a camera pose helps to generate view-dependent shapes. Adding correlation layers during condition encoding significantly facilitates learning, achieving $5 \times$ faster convergence (Fig. 9). + +# 4.2. Multi-view Images to 3D Generation + +Instant reconstruction. Given the generated multi-view images, we use a Large Reconstruction Model (LRM), specifically InstantMesh [69] for fast textured mesh reconstruction. However, the reconstructed textures often appear blurry. On one hand, the resolution of generated images is $256^2$ , which struggles to capture fine details. On the other hand, the appearance quality also degrades during reconstruction. In order to recover and further enrich the appearance, we introduce a new refinement stage. + +Appearance refinement. First, the generated multi-view images are upscaled 4 times into $1024^2$ by Stable Diffusion with ControlNet-Tile [75]. ControlNet-Edge [9, 75] is used to maintain the shape consistency across different views during tiling. Once tiled, these high-resolution im + +ages $I_{i}^{h}$ are used to refine the reconstructed texture. A learnable 2D texture $u$ in UV space is created and initialized as the reconstructed texture $u_{0}$ , and then images are rendered through differentiable rendering $\mathcal{R}(u, c_{i})$ for given camera views $c_{i}$ . The MSE loss is optimized between the rendered images and tiled high-res images. Moreover, a regularization term is added to maintain consistency in UV space: + +$$ +\mathcal {L} _ {u} = \sum_ {i} \left| \left| I _ {i} ^ {h} - \mathcal {R} (u, c _ {i}) \right| \right| _ {2} ^ {2} + \lambda * \left| \left| u - u _ {0} \right| \right| _ {2} ^ {2}. \tag {1} +$$ + +Since the high-res images cannot cover every position on the mesh, some regions of $u$ will not be optimized, e.g. bottom of the object. We found these regions are unstable during optimization and may produce unexpected artifacts (see Fig. 10). The regularization term helps the optimized texture maintain the appearance from $u_{0}$ in these regions. Consequently, the high-frequency details can be learned in covered regions while the global consistency can also be achieved in uncovered regions. The optimization can be finished within 15 seconds. + +# 4.3. Object-Skeleton Alignment Evaluation + +Contrastive alignment. In order to measure how much an object is aligned with a skeleton, we develop a new evaluator, named Contrastive Object-Skeleton Alignment (COSA). We use the self-supervised DINOv2 [44] as the backbone $F$ to extract both object and skeleton features. Then, the alignment adapter $G_{\theta}$ consisting of several self-attention layers is used to modulate the features. The adapter ends with an average pooling layer to aggregate the aligned features into a vector. Similar to CLIP [49], we train the adapter using contrastive learning by InfoNCE loss [42, 56]. Finally, the skeleton alignment score (SKA) can be calculated by cosine similarity between the features from an object image $x$ and a skeleton image $y$ as: + +$$ +\mathcal {S} _ {\mathrm {S K A}} (x, y) = \cos (G (F (x)), G (F (y))). \tag {2} +$$ + +COSA guided diffusion. Based on COSA, another conditional generation pipeline can also be realized, following the approach proposed in [2]. On each denoising time step $t$ , the approximate clean image $\hat{x}_0$ is estimated from the predicted noise $\epsilon_t$ as in DDIM [57]. The estimated clean image and skeleton condition are fed into COSA to calculate the alignment loss $\mathcal{L}_{\mathrm{COSA}}(\hat{x}_0,y) = 1 - S_{\mathrm{SKA}}(\hat{x}_0,y)$ . Then the predicted noise is modified by the gradient of the alignment loss for actual denoising: + +$$ +\hat {\epsilon} _ {t} = \epsilon_ {t} + s (t) \cdot \nabla \mathcal {L} _ {\mathrm {C O S A}} (\hat {x} _ {0}, y) \tag {3} +$$ + +where $s(t)$ controls the guidance strength. With the additional guidance of the alignment loss, the generated object will tend to follow the conditional skeleton $y$ . + +
Method\SKA Score TrainingMeanInst.MeanClassAnimalsHumansPlantsApodesBipedsQuadrupedsArthropodsWings
SDEdit [38]○70.1365.0679.5064.7050.9974.7475.2482.7379.0779.21
SDEdit+COSAG●72.1167.3280.9167.6053.4679.7383.1785.9877.8073.82
Ours-Raw●74.6967.0988.3969.9042.9794.7482.1490.1889.8184.69
Ours-SCM●80.4374.3891.1678.4553.5394.4785.2094.1990.6888.40
+ +Table 1. Quantitative comparison of object-skeleton alignment (SKA) score (\$5.2). Alignment scores are calculated over three classes (blue) and five subclasses of animal (green). The average scores over all instances and three classes (pink) are also shown. The highest scores among all methods are bold and the highest scores among baseline methods are underlined. + +
MethodTrainingPickScoreCLIP Score
Win RateAnimalsHumanPlants\( Mean_{Inst.} \)AnimalsHumanPlants
SDEdit [38]○19.5618.2314.6628.9829.0429.4728.7128.29
SDEdit+COSAG\( \blacksquare \)18.1718.5419.9515.0628.9929.5228.5028.11
Ours-Raw●28.9929.0628.3729.5529.9030.1030.0229.23
Ours-SCM●33.2834.1737.0226.4229.8330.2329.8428.75
+ +Table 2. Quantitative comparison of PickScore and CLIP Score (§5.2). Scores are calculated over three classes (blue) and averaged over all instances (red). The highest scores among all methods are bold and the highest scores among baseline methods are underlined. + +![](images/4e5ddf8ef1448ba0aded0e61608692a204b41d8a063ee69935a38ec68f8e59f5.jpg) +Figure 4. Comparison of skeletons generated from 3D meshes by RigNet [71] and our method (§5.1). + +# 5. Experiment + +# 5.1. Results of Mesh Skeletonization + +We compare our method with the learning-based method RigNet [71], and the results are shown in Fig. 4. RigNet tends to produce symmetric skeletons with limited flexibility and misaligned joints/bones, resulting in an around $15\%$ success rate. In contrast, our method supports arbitrary-pose skeletons and achieves better joint/bone alignment. It produces more reliable results with an $80\%$ success rate. More details and results can be found in appendix. + +# 5.2. Results of Multi-view Generation + +Evaluation protocols. The Objaverse-SK evaluation set contains 108 skeleton instances, covering three main classes: animals, human shapes and plants. As animals include diverse skeleton structures, we further divide them into more detailed subclasses (examples are shown behind): Apodes (fish, snakes), Bipeds (ducks, penguins), Quadrupeds (dogs, bears), Arthropods (scorpions, crabs), Wings (birds, dragons). Three evaluation metrics are con + +sidered for multi-view generation: SKA Score for skeletal alignment, PickScore for image quality, and CLIP Score for textual alignment. Furthermore, samples from categories excluded in training are obtained from ShapeNet [5] for generalization evaluation (Fig. 7). Additionally, evaluation results on human-made skeletons are in appendix. + +Baseline methods. Since there is no prior work that can achieve arbitrary skeletal conditioned generation, we implement two methods for comparison. The first baseline is SDEdit [38]. The process starts from adding noise on conditional images with a time step (set as 0.7). Then clean images are generated by denoising steps. The method is entirely unsupervised. The second baseline is the COSA Guidance (COSAG) derived from [2], which is elaborated in Section 4.3. The guidance strength is set as $s(t) = 7.5\sqrt{1 - \alpha_t}$ . Since we found it cannot achieve stable results, it is combined with SDEdit. The method requires an extra model so it is half supervised. Ours is fully supervised on object-skeleton pairs. + +Qualitative comparison. The qualitative results are shown in Fig. 5. Given skeleton images as conditions, SDEdit can produce images following the skeleton. However, limited by the editing capacity, the generated objects often have incorrect anatomies. For example, the snake body is broken, and the donkey body is generated as wooden. When it is enhanced by the COSAG, the quality of generated content is improved in some cases but still unsatisfactory. Compared with them, our results show superior quality and are more consistent with both skeletal and textual conditions. + +Quantitative comparison. Comparison results of skeleton alignment are shown in Table 1. SDEdit-based methods have around 70 SKA scores, while ours can achieve 80. Among three classes, animals tend to have higher alignment scores while plants have lower scores. Since plants + +![](images/248e142f0b164f8274b9388b4f2b47475b3cb316226ed855a2933b2170ec15c9.jpg) +Figure 5. Qualitative comparison of skeletal conditioned multi-view generation (§5.2). Conditional skeletons and text prompts are shown above. Four views are generated and two views are shown for simplicity. + +![](images/d571e216cc6080034d8efe37e6c35b3ae263d050c8b1f4bf5584bb102ff910bc.jpg) +Figure 6. Qualitative comparison of textured meshes before and after refinement (§5.3). Rendered color images are shown. Local areas are enlarged for better viewing. + +may have more complex structures and sometimes extend further from skeletons. For five sub-classes, our method achieves consistently high alignment scores. With the help of SCM and CCE-D skeletons, both the alignment scores and pick scores are further improved. + +# 5.3. Results of 3D Generation + +Texture refinement. Results of 3D reconstruction from multi-view images are shown in Fig. 6. The raw reconstructed and refined results are compared. The raw textures are blurry and lack details, while the proposed refinement stage can significantly enhance texture quality. + +Rigging and animation. Given a motion sequence of a skeleton, our method can be applied to generate 4D animation. Textured mesh can be generated given the skeleton at the rest pose, and then directly be rigged and skinned for animation. Demo videos can be found in project page1. + +# 6. Ablation Study + +Skeletal condition representation. The skeletal condition representation we use consists of coordinate color encoding (CCE) with depth alpha (D). The ablation results are shown in Fig. 8 and Tab. 3. Richer information in conditions can help the model determine the content better. As a result, better alignment can be achieved. In Fig. 8 right, the skeleton of a penguin is highly ambiguous. If CCE-D is used, the body pose and orientation of the penguin can be successfully inferred from colors. + +Skeletal correlation modeling. We show the effect of our skeletal correlation module in Tab. 3 and Fig. 9. SCM achieves better alignment scores than convolutional blocks. SCM with layer normalization (LN) achieves a $4 \times$ faster convergence rate. Furthermore, if LN is replaced with the adaptive LN (AdaLN), the model can achieve an SKA score of 70 within 1k training steps. The results indicate that extracting global features from conditional images is crucial. + +![](images/21dc4b36abee86086f91274e70d994423aa6ff898b31ccc02c4e4502d4973500.jpg) +Figure 7. Qualitative results of novel categories in ShapeNet [5] (§5.2), which are not covered by the training set of Objaverse-SK. + +![](images/f20c9f0f6322eb23b4bcc7a61cd5038f277d74f8689d33145806fd728ef901b0.jpg) +Figure 8. Ablation study of coordinate color encoding with depth alpha (CCE-D) ( $\S 6$ ). Richer information can help the model to avoid ambiguity and generate better anatomy. + +
ModuleSkeletonSKA ScoreAnimalsHumansPlants
SCMCCE-D80.4391.1678.4553.53
SCMCCE78.9790.4574.8252.57
SCMBinary77.2789.2076.0446.16
ConvCCE-D76.6588.5375.0446.17
ConvBinary74.6988.3969.9042.97
+ +Table 3. Ablation study of skeletal module and representation types (§6). SCM and CCE-D achieve higher alignment scores. + +Refinement regularization. We show the ablation results of appearance refinement in Fig. 10. The refined appearance contains rich details such as snake scales and wood grain, in comparison to the reconstructed results. However, artifacts also appear in the regions that are not covered by high-res images. With the help of UV space regularization, the artifacts are effectively removed in uncovered regions. As a result, consistent colors are maintained from original textures and details are enhanced during optimization. + +![](images/90325cbbee12e7ab61a01307d47912cc129fb6fe145131613362f7729b3d5fbe.jpg) +Figure 9. Convergence processes of different skeletal modules (§6). SCM with AdaLN achieves 5x faster convergence. + +![](images/e2279ee273e3641a6df134415d244219e7e00bb6e24d2f692e65d2b6774fafae.jpg) +Figure 10. Ablation study of UV space regularization (§6). Bottom views and UV textures are shown. Front views of the snake and the tree stump can be found in the first column of Fig. 6. + +![](images/25e3dacd578b05391db044f194cb5e08c46a16fc76cb7db90f8fcd4cd60a7202.jpg) + +# 7. Limitation and Future Work + +Although our work achieves arbitrary skeletal conditioned generation, there are still many problems that can be further studied. The skeletons we currently use may have limited descriptive ability for non-tree structured objects. More powerful shape representations can be studied as new conditions. In addition, our work only considers global skeletons without fine-grained semantics. Injecting detailed semantics into the skeleton parts is a meaningful topic to study. For future work, the skeleton condition can be studied in native 3D generation frameworks. Moreover, our dataset may also be used in other tasks such as arbitrary skeleton estimation from images. More discussion is in the appendix. + +# 8. Conclusion + +In this work, we propose to use skeletons as the structural condition for controllable generation. First, we construct a large-scale 3D mesh-skeleton paired dataset. We propose an effective mesh skeletonization method to generate mesh-aligned sparse skeletons with a high success rate. Based on the dataset, we present a skeletal conditioned multi-view generation pipeline. Coordinate color encoding and skeletal correlation module are proposed to realize efficient condition representation and injection. Furthermore, 3D meshes can be instantly reconstructed, followed by a refinement stage to achieve better texture quality. In summary, our work achieves controllable multi-view and 3D generation with arbitrary skeletons as conditions. + +# Acknowledgements + +This work was supported by National Natural Science Foundation of China (No. 62293554), and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02) and the Fundamental Research Funds for the Zhejiang Provincial Universities (226-2024-00208). + +# References + +[1] Andreas Bærentzen and Eva Rotenberg. Skeletonization via local separators. ACM Transactions on Graphics (TOG), 40 (5):1-18, 2021. 3 +[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 5, 6 +[3] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 1 +[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8 +[6] Cheng Chen, Xiaofeng Yang, Fan Yang, Chengzeng Feng, Zhoujie Fu, Chuan-Sheng Foo, Guosheng Lin, and Fayao Liu. Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10228-10237, 2024. 3 +[7] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 3 +[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 1, 3 +[9] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5 +[10] Wenqi Dong, Bangbang Yang, Lin Ma, Xiao Liu, Liyuan Cui, Hujun Bao, Yuewen Ma, and Zhaopeng Cui. Coin3d: Controllable and interactive 3d assets generation with proxy-guided conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1, 3 +[11] Zhiyang Dou, Cheng Lin, Rui Xu, Lei Yang, Shiqing Xin, Taku Komura, and Wenping Wang. Coverage axis: Inner point selection for 3d shape skeletonization. In Com + +puter Graphics Forum, pages 419-432. Wiley Online Library, 2022. 3 +[12] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 10 (2):112-122, 1973. 4 +[13] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Navigation instruction generation with bev perception and large language models. In ECCV, 2024. 1 +[14] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Scene map-based prompt tuning for navigation instruction generation. In CVPR, 2025. 1 +[15] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 1 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 1, 3 +[18] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3 +[19] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[20] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 1, 3 +[21] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3 +[22] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46, 2007. 1 +[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1 +[24] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 3 +[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. + +Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2024. 3 +[26] Christoph Leiter, Ran Zhang, Yanran Chen, Jonas Belouadi, Daniil Larionov, Vivian Fresen, and Steffen Eger. Chatgpt: A meta-analysis after 2.5 months. arXiv preprint arXiv:2302.13795, 2023. 1 +[27] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 1, 3 +[28] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 3, 4 +[29] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3 +[30] Zhiqi Li, Yiming Chen, Lingzhe Zhao, and Peidong Liu. Mycontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494, 2023. 1, 3 +[31] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3 +[32] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6400-6409, 2024. 3 +[33] Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J Black. Tada! text to animatable digital avatars. In 2024 International Conference on 3D Vision (3DV), pages 1508-1519. IEEE, 2024. 3 +[34] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20763–20774, 2024. 3 +[35] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1 +[36] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[37] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1 + +[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3, 6 +[39] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3 +[40] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 3 +[41] Marianna Ohanyan, Hayk Manukyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Zero-painter: Training-free layout control for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8764-8774, 2024. 3 +[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[43] OpenAI. Gpt-4 technical report, 2023. 1 +[44] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5 +[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3 +[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[47] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 3 +[48] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 1 + +[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3 +[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[54] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3 +[55] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 4 +[56] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 5 +[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 5 +[58] Andrea Tagliasacchi, Ibraheem Alhashim, Matt Olson, and Hao Zhang. Mean curvature skeletons. In Computer Graphics Forum, pages 1735-1744. Wiley Online Library, 2012. 3, 4 +[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3 +[60] Ashish Vaswani. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 3 +[61] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 4 +[62] Jiajun Wang, Morteza Ghahremani, Yitong Li, Björn Omer, and Christian Wachinger. Stable-pose: Leveraging transformers for pose-guided text-to-image generation. arXiv preprint arXiv:2406.02485, 2024. 3 +[63] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3 +[64] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 1 + +[65] Xinzhou Wang, Yikai Wang, Junliang Ye, Fuchun Sun, Zhengyi Wang, Ling Wang, Pengkun Liu, Kai Sun, Xintong Wang, Wende Xie, et al. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 3 +[66] Zimeng Wang, Zhiyang Dou, Rui Xu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Shiqing Xin, Taku Komura, Xiaoming Yuan, and Wenping Wang. Coverage axis++: Efficient inner point selection for 3d shape skeletonization. In Computer Graphics Forum, page e15143. Wiley Online Library, 2024. 3 +[67] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 3 +[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 5 +[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 1, 3, 5 +[70] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, and Karan Singh. Predicting animation skeletons for 3d articulated models via volumetric nets. In 2019 international conference on 3D vision (3DV), pages 298-307. IEEE, 2019. 3 +[71] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2, 3, 4, 6 +[72] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 1 +[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 1 +[74] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3, 5 +[76] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3 +[77] Muxin Zhang, Qiao Feng, Zhuo Su, Chao Wen, Zhou Xue, and Kun Li. Joint2human: High-quality 3d human generation via compact spherical embedding of 3d joints. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1438, 2024. 1, 3 +[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[79] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3 +[80] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavator: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 3 +[81] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3 +[82] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 3 +[83] Wenjie Zhuo, Fan Ma, Hehe Fan, and Yi Yang. Vividdreamer: invariant score distillation for hyper-realistic text-to-3d generation. In European Conference on Computer Vision, pages 122–139. Springer, 2024. 3 \ No newline at end of file diff --git a/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/images.zip b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..347d0c93a9798cb374f0a27656cc88ca22978758 --- /dev/null +++ b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64571f84a1819b03cbd4a8774215d36eec73ba3739359c4e6997ab293cbc77a7 +size 714791 diff --git a/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/layout.json b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba2168fa3dddb24a1839e37161b53f13823f3693 --- /dev/null +++ b/2025/SKDream_ Controllable Multi-view and 3D Generation with Arbitrary Skeletons/layout.json @@ -0,0 +1,11084 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 555, + 121 + ], + "type": "text", + "content": "SKDream: Controllable Multi-view and 3D Generation with Arbitrary Skeletons" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "spans": [ + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "text", + "content": "Yuanyou Xu" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "text", + "content": ", Zongxin Yang" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "text", + "content": ", Yi Yang" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "text", + "content": "ReLER, CCAI, Zhejiang University \n" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 206, + 142, + 408, + 186 + ], + "type": "text", + "content": "DBMI, HMS, Harvard University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 187, + 454, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 187, + 454, + 200 + ], + "spans": [ + { + "bbox": [ + 153, + 187, + 454, + 200 + ], + "type": "text", + "content": "{yoxu, yangyics}@zju.edu.cn, zongxin-yang@hms.harvard.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 252, + 297, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 504 + ], + "type": "text", + "content": "Controllable generation has achieved substantial progress in both 2D and 3D domains, yet current conditional generation methods still face limitations in describing detailed shape structures. Skeletons can effectively represent and describe object anatomy and pose. Unfortunately, past studies are often limited to human skeletons. In this work, we generalize skeletal conditioned generation to arbitrary structures. First, we design a reliable mesh skeletonization pipeline to generate a large-scale mesh-skeleton paired dataset. Based on the dataset, a multi-view and 3D generation pipeline is built. We propose to represent 3D skeletons by Coordinate Color Encoding as 2D conditional images. A Skeletal Correlation Module is designed to extract global skeletal features for condition injection. After multi-view images are generated, 3D assets can be obtained by incorporating a large reconstruction model, followed by a UV texture refinement stage. As a result, our method achieves instant generation of multi-view and 3D contents that are aligned with given skeletons. The proposed techniques largely improve the object-skeleton alignment and generation quality. Project page at https://skdream3d.github.io/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 135, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 561, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 693 + ], + "type": "text", + "content": "In view of visual representation dimension [64, 72], 2D image generation [16, 53, 57], multi-view (2.5D) generation [35, 55], and 3D generation [17, 27, 69] have advanced successively and achieved significant progress. Given the success of large language models [23, 26, 43, 48-50], textual representation has been widely applied in generation [51-53], as in other domains [13, 14, 73]. To achieve more controllable generation, conditions beyond text have attracted considerable attention. 2D image conditions (e.g., edge maps, human skeletons, and concept references) [54, 75] have been well studied. Similarly in 3D generation," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 544, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 544, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 544, + 240 + ], + "type": "text", + "content": "2D [30] and 3D [10] conditions have also been explored." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 241, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 241, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 312, + 241, + 555, + 421 + ], + "type": "text", + "content": "Although the aforementioned conditions in controllable generation complement text descriptions, they still struggle to precisely describe shape structures. In contrast, skeletons, among various types of conditions, exhibit superior ability to depict shape structures: (i) Representation of object anatomy. Skeletons can efficiently represent various 3D structures with sparse joints and bones, while other conditions like depth maps fail to represent full 3D anatomy. (ii) Articulation into different poses. Skeletons are widely used for character animation in computer graphics [3, 22] due to their simplicity and efficiency. Other conditions such as rough shapes [10] are inconvenient for pose articulation. (iii) Freedom of editing. Given an initial skeleton, users can freely add new structures or modify joints and bones to create ideal shapes. Examples are in Fig. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 483 + ], + "type": "text", + "content": "Despite these advantages, previous studies [19, 20, 40, 75, 77] on conditional generation are limited to human skeletons. From the perspective of generalization, we would like to ask: Is it possible to use arbitrary skeletons as conditions to generate any creatures or even general objects?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 640 + ], + "type": "text", + "content": "To achieve this goal, we identify two main challenges that hinder the use of arbitrary skeletal conditions for generation: (i) Lack of large-scale object-skeleton pairs for training. Extensive studies [4, 15, 37] on 2D/3D human pose estimation have made human-skeleton paired data readily available. However, when skeletal structures are unknown, estimating arbitrary skeletons from 2D images or videos becomes challenging due to its ill-posedness. (ii) Insufficiency of 2D information to describe arbitrary skeletons. Human skeletons are relatively simple and can be described by a fixed set of 2D joints. However, complex skeletons suffer from self-occlusion and ambiguity, necessitating 3D information to fully capture their anatomy and pose." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": "To address these challenges, we focus on multi-view and 3D generation with skeletal conditions. To tackle data scarcity problem, we construct a large-scale dataset Objaverse-SK containing mesh-skeleton pairs. Textured meshes are selected from Objaverse [8] by semantic classes to form a subset. In order to achieve reliable mesh ske" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "314" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 83, + 143, + 119 + ], + "blocks": [ + { + "bbox": [ + 99, + 83, + 143, + 119 + ], + "lines": [ + { + "bbox": [ + 99, + 83, + 143, + 119 + ], + "spans": [ + { + "bbox": [ + 99, + 83, + 143, + 119 + ], + "type": "image", + "image_path": "0afbb1f749cc0cbdc98dd5fe1dd01e3833dcfac37e9e2d24fecbc9d1b4cea17e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 153, + 85, + 183, + 122 + ], + "blocks": [ + { + "bbox": [ + 153, + 85, + 183, + 122 + ], + "lines": [ + { + "bbox": [ + 153, + 85, + 183, + 122 + ], + "spans": [ + { + "bbox": [ + 153, + 85, + 183, + 122 + ], + "type": "image", + "image_path": "25c410ee255a8d99ceb9eb74a7b4357fa1c678b5f5a29bd5d824ce0d3be3c97f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 85, + 219, + 114 + ], + "blocks": [ + { + "bbox": [ + 187, + 85, + 219, + 114 + ], + "lines": [ + { + "bbox": [ + 187, + 85, + 219, + 114 + ], + "spans": [ + { + "bbox": [ + 187, + 85, + 219, + 114 + ], + "type": "image", + "image_path": "a00d0e26d78e1918eeb6ea317f26a1c7200b4a26fae4c1b3af86d0b38a52d567.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 81, + 139, + 110, + 169 + ], + "blocks": [ + { + "bbox": [ + 62, + 83, + 73, + 116 + ], + "lines": [ + { + "bbox": [ + 62, + 83, + 73, + 116 + ], + "spans": [ + { + "bbox": [ + 62, + 83, + 73, + 116 + ], + "type": "text", + "content": "MVDream" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 142, + 73, + 159 + ], + "lines": [ + { + "bbox": [ + 64, + 142, + 73, + 159 + ], + "spans": [ + { + "bbox": [ + 64, + 142, + 73, + 159 + ], + "type": "text", + "content": "sno" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 81, + 139, + 110, + 169 + ], + "lines": [ + { + "bbox": [ + 81, + 139, + 110, + 169 + ], + "spans": [ + { + "bbox": [ + 81, + 139, + 110, + 169 + ], + "type": "image", + "image_path": "7efd391b91e4f5b2172481bd2da9d2579f92f7892611ffd2c56edb7546ce6761.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 116, + 137, + 146, + 172 + ], + "blocks": [ + { + "bbox": [ + 71, + 175, + 117, + 183 + ], + "lines": [ + { + "bbox": [ + 71, + 175, + 117, + 183 + ], + "spans": [ + { + "bbox": [ + 71, + 175, + 117, + 183 + ], + "type": "text", + "content": "Text & Skeleton" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 137, + 146, + 172 + ], + "lines": [ + { + "bbox": [ + 116, + 137, + 146, + 172 + ], + "spans": [ + { + "bbox": [ + 116, + 137, + 146, + 172 + ], + "type": "image", + "image_path": "08ef72300a46dfc3d905c8b346e18f2cc6cadd4170e55d58521c9d6fc21dd0bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 175, + 137, + 183 + ], + "lines": [ + { + "bbox": [ + 125, + 175, + 137, + 183 + ], + "spans": [ + { + "bbox": [ + 125, + 175, + 137, + 183 + ], + "type": "text", + "content": "Dog" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 149, + 140, + 184, + 173 + ], + "blocks": [ + { + "bbox": [ + 149, + 140, + 184, + 173 + ], + "lines": [ + { + "bbox": [ + 149, + 140, + 184, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 140, + 184, + 173 + ], + "type": "image", + "image_path": "7d1dd94c0d63234511a5c7dae5bc680c9bf735fd5334459cfd674d3bb3426607.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 175, + 176, + 183 + ], + "lines": [ + { + "bbox": [ + 160, + 175, + 176, + 183 + ], + "spans": [ + { + "bbox": [ + 160, + 175, + 176, + 183 + ], + "type": "text", + "content": "Chair" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 189, + 131, + 220, + 173 + ], + "blocks": [ + { + "bbox": [ + 189, + 131, + 220, + 173 + ], + "lines": [ + { + "bbox": [ + 189, + 131, + 220, + 173 + ], + "spans": [ + { + "bbox": [ + 189, + 131, + 220, + 173 + ], + "type": "image", + "image_path": "85cef554b554ce5b10bd75589395a08b6ae6445dd6af76ba88f7aa33fec4903c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 175, + 213, + 183 + ], + "lines": [ + { + "bbox": [ + 199, + 175, + 213, + 183 + ], + "spans": [ + { + "bbox": [ + 199, + 175, + 213, + 183 + ], + "type": "text", + "content": "Fish" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 235, + 127, + 267, + 175 + ], + "blocks": [ + { + "bbox": [ + 235, + 127, + 267, + 175 + ], + "lines": [ + { + "bbox": [ + 235, + 127, + 267, + 175 + ], + "spans": [ + { + "bbox": [ + 235, + 127, + 267, + 175 + ], + "type": "image", + "image_path": "9a2c1116ee6165bb09ef7a2a314dd32a625e9e2e1aa53806af3b2ffa7cc0242a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 272, + 77, + 305, + 124 + ], + "blocks": [ + { + "bbox": [ + 238, + 97, + 266, + 106 + ], + "lines": [ + { + "bbox": [ + 238, + 97, + 266, + 106 + ], + "spans": [ + { + "bbox": [ + 238, + 97, + 266, + 106 + ], + "type": "text", + "content": "Text Only" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 77, + 305, + 124 + ], + "lines": [ + { + "bbox": [ + 272, + 77, + 305, + 124 + ], + "spans": [ + { + "bbox": [ + 272, + 77, + 305, + 124 + ], + "type": "image", + "image_path": "fac99c2eda3a17b8cea6a79ee43f229656a38d3154cb3cfd59382e6b6d3bc1a7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 271, + 126, + 305, + 176 + ], + "blocks": [ + { + "bbox": [ + 271, + 126, + 305, + 176 + ], + "lines": [ + { + "bbox": [ + 271, + 126, + 305, + 176 + ], + "spans": [ + { + "bbox": [ + 271, + 126, + 305, + 176 + ], + "type": "image", + "image_path": "86f54caa4684bef521224cc22a98f6936016b5f5e5c0b2ec223b56c19b84c546.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 176, + 303, + 183 + ], + "lines": [ + { + "bbox": [ + 272, + 176, + 303, + 183 + ], + "spans": [ + { + "bbox": [ + 272, + 176, + 303, + 183 + ], + "type": "text", + "content": "Spiderman" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 189, + 457, + 201 + ], + "lines": [ + { + "bbox": [ + 150, + 189, + 457, + 201 + ], + "spans": [ + { + "bbox": [ + 150, + 189, + 457, + 201 + ], + "type": "text", + "content": "(a) Text and skeletons for Collaborative Appearance and Shape Controlling" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 78, + 337, + 123 + ], + "blocks": [ + { + "bbox": [ + 316, + 78, + 337, + 123 + ], + "lines": [ + { + "bbox": [ + 316, + 78, + 337, + 123 + ], + "spans": [ + { + "bbox": [ + 316, + 78, + 337, + 123 + ], + "type": "image", + "image_path": "97fb40a7cb2d5e27595631d7ba02d4ed43bccb98fe3cb97d6d7f2a0198c7c4a5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 126, + 343, + 175 + ], + "blocks": [ + { + "bbox": [ + 312, + 126, + 343, + 175 + ], + "lines": [ + { + "bbox": [ + 312, + 126, + 343, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 343, + 175 + ], + "type": "image", + "image_path": "3d04219a1222971b64d7ed2afd444c0aa9534bc273a276fb105214ee73956ed6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 176, + 337, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 176, + 337, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 176, + 337, + 182 + ], + "type": "text", + "content": "Rocket" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 353, + 78, + 375, + 121 + ], + "blocks": [ + { + "bbox": [ + 353, + 78, + 375, + 121 + ], + "lines": [ + { + "bbox": [ + 353, + 78, + 375, + 121 + ], + "spans": [ + { + "bbox": [ + 353, + 78, + 375, + 121 + ], + "type": "image", + "image_path": "103aa145f1228f6b48ff449e647403c6553b8aed7795fddfe1e269a341cd390c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 345, + 126, + 384, + 175 + ], + "blocks": [ + { + "bbox": [ + 345, + 126, + 384, + 175 + ], + "lines": [ + { + "bbox": [ + 345, + 126, + 384, + 175 + ], + "spans": [ + { + "bbox": [ + 345, + 126, + 384, + 175 + ], + "type": "image", + "image_path": "a0877861603c8471d23e49169038a25b0b54c62afb407571e603d8d26e468888.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 176, + 380, + 183 + ], + "lines": [ + { + "bbox": [ + 352, + 176, + 380, + 183 + ], + "spans": [ + { + "bbox": [ + 352, + 176, + 380, + 183 + ], + "type": "text", + "content": "Ice cream" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 396, + 127, + 424, + 175 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 127, + 424, + 175 + ], + "spans": [ + { + "bbox": [ + 396, + 127, + 424, + 175 + ], + "type": "image", + "image_path": "66668ed58d5a1193aa6a7bf138f772663ba13d696a4fec7bafdca33e2b287129.jpg" + } + ] + } + ], + "index": 25, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 430, + 85, + 468, + 114 + ], + "blocks": [ + { + "bbox": [ + 397, + 98, + 427, + 106 + ], + "lines": [ + { + "bbox": [ + 397, + 98, + 427, + 106 + ], + "spans": [ + { + "bbox": [ + 397, + 98, + 427, + 106 + ], + "type": "text", + "content": "Text Only" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 85, + 468, + 114 + ], + "lines": [ + { + "bbox": [ + 430, + 85, + 468, + 114 + ], + "spans": [ + { + "bbox": [ + 430, + 85, + 468, + 114 + ], + "type": "image", + "image_path": "5d902562c5f9e00063d794b49712c15baaafb81b11a3fe05b91557b84c2e7699.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 472, + 83, + 511, + 115 + ], + "blocks": [ + { + "bbox": [ + 472, + 83, + 511, + 115 + ], + "lines": [ + { + "bbox": [ + 472, + 83, + 511, + 115 + ], + "spans": [ + { + "bbox": [ + 472, + 83, + 511, + 115 + ], + "type": "image", + "image_path": "1c9578f61a2a517c76c5eb64f5a610b9c0b21cb263b58b9d58eff42b6b04ba4c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 518, + 78, + 535, + 119 + ], + "blocks": [ + { + "bbox": [ + 518, + 78, + 535, + 119 + ], + "lines": [ + { + "bbox": [ + 518, + 78, + 535, + 119 + ], + "spans": [ + { + "bbox": [ + 518, + 78, + 535, + 119 + ], + "type": "image", + "image_path": "3a1483b0375fff0ee8bb5469c4d2b1f3fc21183cf9c480627a00e0720e85508e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 396, + 126, + 424, + 175 + ], + "blocks": [ + { + "bbox": [ + 396, + 126, + 424, + 175 + ], + "lines": [ + { + "bbox": [ + 396, + 126, + 424, + 175 + ], + "spans": [ + { + "bbox": [ + 396, + 126, + 424, + 175 + ], + "type": "image", + "image_path": "486e1a8da97c2665e6b861e300ef360f11ba289943612568a23057521eb15dea.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 430, + 125, + 468, + 175 + ], + "blocks": [ + { + "bbox": [ + 430, + 125, + 468, + 175 + ], + "lines": [ + { + "bbox": [ + 430, + 125, + 468, + 175 + ], + "spans": [ + { + "bbox": [ + 430, + 125, + 468, + 175 + ], + "type": "image", + "image_path": "b15b2f8bda631ef6a647b4608b255c9cc83f9d485b9ff6e4e2dfbdb670e8e909.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 465, + 126, + 504, + 175 + ], + "blocks": [ + { + "bbox": [ + 465, + 126, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 465, + 126, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 465, + 126, + 504, + 175 + ], + "type": "image", + "image_path": "7578f5e542481c324c0f22cafcb787af1833f34491cb74657c9d4465181b6ab9.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 176, + 498, + 183 + ], + "lines": [ + { + "bbox": [ + 482, + 176, + 498, + 183 + ], + "spans": [ + { + "bbox": [ + 482, + 176, + 498, + 183 + ], + "type": "text", + "content": "Coral" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 503, + 175, + 536, + 182 + ], + "lines": [ + { + "bbox": [ + 503, + 175, + 536, + 182 + ], + "spans": [ + { + "bbox": [ + 503, + 175, + 536, + 182 + ], + "type": "text", + "content": "Torch" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 513, + 129, + 542, + 175 + ], + "blocks": [ + { + "bbox": [ + 513, + 129, + 542, + 175 + ], + "lines": [ + { + "bbox": [ + 513, + 129, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 513, + 129, + 542, + 175 + ], + "type": "image", + "image_path": "c9aeee3ae7bcf98ba3fa96ba6ff877d63c8e3c724c4375312534f199325100a7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 70, + 232, + 99, + 260 + ], + "blocks": [ + { + "bbox": [ + 70, + 232, + 99, + 260 + ], + "lines": [ + { + "bbox": [ + 70, + 232, + 99, + 260 + ], + "spans": [ + { + "bbox": [ + 70, + 232, + 99, + 260 + ], + "type": "image", + "image_path": "32047747d873f67b55ee937c9d8c97392b1ec19707c96567036be54d839fbd4c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 262, + 97, + 269 + ], + "lines": [ + { + "bbox": [ + 74, + 262, + 97, + 269 + ], + "spans": [ + { + "bbox": [ + 74, + 262, + 97, + 269 + ], + "type": "text", + "content": "Animal" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 67, + 281, + 99, + 329 + ], + "blocks": [ + { + "bbox": [ + 67, + 281, + 99, + 329 + ], + "lines": [ + { + "bbox": [ + 67, + 281, + 99, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 281, + 99, + 329 + ], + "type": "image", + "image_path": "cb0b3004e56923e389338ab80c6d5776f6da88f07a713d3a4dcdece2ccbcde25.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 331, + 97, + 337 + ], + "lines": [ + { + "bbox": [ + 74, + 331, + 97, + 337 + ], + "spans": [ + { + "bbox": [ + 74, + 331, + 97, + 337 + ], + "type": "text", + "content": "Human" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 67, + 345, + 97, + 389 + ], + "blocks": [ + { + "bbox": [ + 67, + 345, + 97, + 389 + ], + "lines": [ + { + "bbox": [ + 67, + 345, + 97, + 389 + ], + "spans": [ + { + "bbox": [ + 67, + 345, + 97, + 389 + ], + "type": "image", + "image_path": "8ae4fdc0d04a155b5e87d6ea27e595175b36a569f9eddd1f644cecfcdf54a513.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 391, + 93, + 398 + ], + "lines": [ + { + "bbox": [ + 76, + 391, + 93, + 398 + ], + "spans": [ + { + "bbox": [ + 76, + 391, + 93, + 398 + ], + "type": "text", + "content": "Plant" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 111, + 220, + 181, + 260 + ], + "blocks": [ + { + "bbox": [ + 69, + 208, + 122, + 217 + ], + "lines": [ + { + "bbox": [ + 69, + 208, + 122, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 122, + 217 + ], + "type": "text", + "content": "Initial skeleton" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 206, + 232, + 218 + ], + "lines": [ + { + "bbox": [ + 148, + 206, + 232, + 218 + ], + "spans": [ + { + "bbox": [ + 148, + 206, + 232, + 218 + ], + "type": "text", + "content": "Adding joints, bones" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 220, + 181, + 260 + ], + "lines": [ + { + "bbox": [ + 111, + 220, + 181, + 260 + ], + "spans": [ + { + "bbox": [ + 111, + 220, + 181, + 260 + ], + "type": "image", + "image_path": "12b88a4c2585954215fd56756cce4d6fcc163aea5cb7ebc6b6874f5e2672a63b.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 262, + 177, + 270 + ], + "lines": [ + { + "bbox": [ + 121, + 262, + 177, + 270 + ], + "spans": [ + { + "bbox": [ + 121, + 262, + 177, + 270 + ], + "type": "text", + "content": "Dog with a spiral tail" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 184, + 224, + 255, + 261 + ], + "blocks": [ + { + "bbox": [ + 184, + 224, + 255, + 261 + ], + "lines": [ + { + "bbox": [ + 184, + 224, + 255, + 261 + ], + "spans": [ + { + "bbox": [ + 184, + 224, + 255, + 261 + ], + "type": "image", + "image_path": "b797d4f5e3878ba89cb41acbf3da3b0e5d15fff47d67dd1eaf5013a4fbbe9afa.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 262, + 253, + 270 + ], + "lines": [ + { + "bbox": [ + 199, + 262, + 253, + 270 + ], + "spans": [ + { + "bbox": [ + 199, + 262, + 253, + 270 + ], + "type": "text", + "content": "Dog with two heads" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 110, + 281, + 181, + 331 + ], + "blocks": [ + { + "bbox": [ + 110, + 281, + 181, + 331 + ], + "lines": [ + { + "bbox": [ + 110, + 281, + 181, + 331 + ], + "spans": [ + { + "bbox": [ + 110, + 281, + 181, + 331 + ], + "type": "image", + "image_path": "48309fdfc7f776ab20d0759186ac394e4e2bd1d2e077df7c543ec422f9ca8bbb.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 331, + 184, + 338 + ], + "lines": [ + { + "bbox": [ + 114, + 331, + 184, + 338 + ], + "spans": [ + { + "bbox": [ + 114, + 331, + 184, + 338 + ], + "type": "text", + "content": "Knight with dragon wings" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 117, + 342, + 179, + 389 + ], + "blocks": [ + { + "bbox": [ + 117, + 342, + 179, + 389 + ], + "lines": [ + { + "bbox": [ + 117, + 342, + 179, + 389 + ], + "spans": [ + { + "bbox": [ + 117, + 342, + 179, + 389 + ], + "type": "image", + "image_path": "b502d02c4974d33a9e46cde3bcb701ec60ed8f0d20b313d7b7713b8322c24696.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 391, + 184, + 398 + ], + "lines": [ + { + "bbox": [ + 111, + 391, + 184, + 398 + ], + "spans": [ + { + "bbox": [ + 111, + 391, + 184, + 398 + ], + "type": "text", + "content": "Plant with leaves and roots" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 190, + 281, + 264, + 331 + ], + "blocks": [ + { + "bbox": [ + 190, + 281, + 264, + 331 + ], + "lines": [ + { + "bbox": [ + 190, + 281, + 264, + 331 + ], + "spans": [ + { + "bbox": [ + 190, + 281, + 264, + 331 + ], + "type": "image", + "image_path": "afd9e106bda4954b4111f6d1cf65f20c5391384583a2604b90386aa697e11d50.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 331, + 254, + 338 + ], + "lines": [ + { + "bbox": [ + 192, + 331, + 254, + 338 + ], + "spans": [ + { + "bbox": [ + 192, + 331, + 254, + 338 + ], + "type": "text", + "content": "Monster with four arms" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 194, + 342, + 251, + 390 + ], + "blocks": [ + { + "bbox": [ + 194, + 342, + 251, + 390 + ], + "lines": [ + { + "bbox": [ + 194, + 342, + 251, + 390 + ], + "spans": [ + { + "bbox": [ + 194, + 342, + 251, + 390 + ], + "type": "image", + "image_path": "58744925b1b2268c9169ee3c10ba29fbe46a2b969d6767255efdfdbad4078f1d.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 391, + 263, + 398 + ], + "lines": [ + { + "bbox": [ + 193, + 391, + 263, + 398 + ], + "spans": [ + { + "bbox": [ + 193, + 391, + 263, + 398 + ], + "type": "text", + "content": "Plant with legs, Pokémon" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 264, + 219, + 333, + 260 + ], + "blocks": [ + { + "bbox": [ + 264, + 206, + 406, + 218 + ], + "lines": [ + { + "bbox": [ + 264, + 206, + 406, + 218 + ], + "spans": [ + { + "bbox": [ + 264, + 206, + 406, + 218 + ], + "type": "text", + "content": "Modifying joint positions, bone sizes" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 264, + 219, + 333, + 260 + ], + "lines": [ + { + "bbox": [ + 264, + 219, + 333, + 260 + ], + "spans": [ + { + "bbox": [ + 264, + 219, + 333, + 260 + ], + "type": "image", + "image_path": "6232048d19425b99adeac4a279bbef9c419e3322346843b3c53fdcf3a5c204cc.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 262, + 328, + 270 + ], + "lines": [ + { + "bbox": [ + 272, + 262, + 328, + 270 + ], + "spans": [ + { + "bbox": [ + 272, + 262, + 328, + 270 + ], + "type": "text", + "content": "Dog with a big head" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 338, + 219, + 405, + 258 + ], + "blocks": [ + { + "bbox": [ + 338, + 219, + 405, + 258 + ], + "lines": [ + { + "bbox": [ + 338, + 219, + 405, + 258 + ], + "spans": [ + { + "bbox": [ + 338, + 219, + 405, + 258 + ], + "type": "image", + "image_path": "f50b45ce4bc9454e0e5d853d778f0ecdbfd70edd5af9ba1b852ee74c0ffaad62.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 262, + 403, + 270 + ], + "lines": [ + { + "bbox": [ + 345, + 262, + 403, + 270 + ], + "spans": [ + { + "bbox": [ + 345, + 262, + 403, + 270 + ], + "type": "text", + "content": "Dog with a long torso" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 267, + 278, + 337, + 327 + ], + "blocks": [ + { + "bbox": [ + 267, + 278, + 337, + 327 + ], + "lines": [ + { + "bbox": [ + 267, + 278, + 337, + 327 + ], + "spans": [ + { + "bbox": [ + 267, + 278, + 337, + 327 + ], + "type": "image", + "image_path": "62e188a77cdc1da2f3b9b2df854c7b7819c038dffcbd491fddc0b05b87780fb0.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 331, + 321, + 338 + ], + "lines": [ + { + "bbox": [ + 280, + 331, + 321, + 338 + ], + "spans": [ + { + "bbox": [ + 280, + 331, + 321, + 338 + ], + "type": "text", + "content": "Dwarf in armor" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 341, + 278, + 406, + 330 + ], + "blocks": [ + { + "bbox": [ + 345, + 270, + 371, + 277 + ], + "lines": [ + { + "bbox": [ + 345, + 270, + 371, + 277 + ], + "spans": [ + { + "bbox": [ + 345, + 270, + 371, + 277 + ], + "type": "text", + "content": "How wide?" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 278, + 406, + 330 + ], + "lines": [ + { + "bbox": [ + 341, + 278, + 406, + 330 + ], + "spans": [ + { + "bbox": [ + 341, + 278, + 406, + 330 + ], + "type": "image", + "image_path": "f026946b1d9f0cd73d3c3fa3960efcc9e70e1243bd0333d162fad5a442ac6630.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 331, + 410, + 338 + ], + "lines": [ + { + "bbox": [ + 336, + 331, + 410, + 338 + ], + "spans": [ + { + "bbox": [ + 336, + 331, + 410, + 338 + ], + "type": "text", + "content": "Wide shoulder bodybuilder" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 267, + 345, + 336, + 386 + ], + "blocks": [ + { + "bbox": [ + 267, + 345, + 336, + 386 + ], + "lines": [ + { + "bbox": [ + 267, + 345, + 336, + 386 + ], + "spans": [ + { + "bbox": [ + 267, + 345, + 336, + 386 + ], + "type": "image", + "image_path": "ae20cba4474432e47284557cfd31e28d00f3526ab8759079a995a780fa0bb06e.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 392, + 337, + 399 + ], + "lines": [ + { + "bbox": [ + 270, + 392, + 337, + 399 + ], + "spans": [ + { + "bbox": [ + 270, + 392, + 337, + 399 + ], + "type": "text", + "content": "Tree with wide branches" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 343, + 342, + 370, + 391 + ], + "blocks": [ + { + "bbox": [ + 343, + 342, + 370, + 391 + ], + "lines": [ + { + "bbox": [ + 343, + 342, + 370, + 391 + ], + "spans": [ + { + "bbox": [ + 343, + 342, + 370, + 391 + ], + "type": "image", + "image_path": "c9e78a8b57ff69b3e7b45b36dbbc3c051e76265c5c82a92e0ffb035c7f468487.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 392, + 403, + 399 + ], + "lines": [ + { + "bbox": [ + 343, + 392, + 403, + 399 + ], + "spans": [ + { + "bbox": [ + 343, + 392, + 403, + 399 + ], + "type": "text", + "content": "Tree with a long truck" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 161, + 403, + 448, + 415 + ], + "lines": [ + { + "bbox": [ + 161, + 403, + 448, + 415 + ], + "spans": [ + { + "bbox": [ + 161, + 403, + 448, + 415 + ], + "type": "text", + "content": "(b) Skeleton-based editing for Accurate Anatomy and Pose Controlling" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 417, + 220, + 475, + 259 + ], + "blocks": [ + { + "bbox": [ + 462, + 207, + 498, + 218 + ], + "lines": [ + { + "bbox": [ + 462, + 207, + 498, + 218 + ], + "spans": [ + { + "bbox": [ + 462, + 207, + 498, + 218 + ], + "type": "text", + "content": "Posing" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 220, + 475, + 259 + ], + "lines": [ + { + "bbox": [ + 417, + 220, + 475, + 259 + ], + "spans": [ + { + "bbox": [ + 417, + 220, + 475, + 259 + ], + "type": "image", + "image_path": "50faf0ff8fbf485697abc3de93af8a61a39aa191b84f1f103b677761eea1f0a7.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 262, + 466, + 271 + ], + "lines": [ + { + "bbox": [ + 429, + 262, + 466, + 271 + ], + "spans": [ + { + "bbox": [ + 429, + 262, + 466, + 271 + ], + "type": "text", + "content": "Dog jumping" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 482, + 222, + 537, + 258 + ], + "blocks": [ + { + "bbox": [ + 482, + 222, + 537, + 258 + ], + "lines": [ + { + "bbox": [ + 482, + 222, + 537, + 258 + ], + "spans": [ + { + "bbox": [ + 482, + 222, + 537, + 258 + ], + "type": "image", + "image_path": "81a0f19445b6a834f306eb385cbe4c924f19a1c1d5cde68f4e9cd0a6223fb843.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 262, + 524, + 270 + ], + "lines": [ + { + "bbox": [ + 492, + 262, + 524, + 270 + ], + "spans": [ + { + "bbox": [ + 492, + 262, + 524, + 270 + ], + "type": "text", + "content": "Dog sitting" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 416, + 280, + 475, + 328 + ], + "blocks": [ + { + "bbox": [ + 416, + 280, + 475, + 328 + ], + "lines": [ + { + "bbox": [ + 416, + 280, + 475, + 328 + ], + "spans": [ + { + "bbox": [ + 416, + 280, + 475, + 328 + ], + "type": "image", + "image_path": "1da94bcb445dc3678fceafce0d48bcd1097608031a535a16b3b52cddd7a80469.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 330, + 470, + 338 + ], + "lines": [ + { + "bbox": [ + 427, + 330, + 470, + 338 + ], + "spans": [ + { + "bbox": [ + 427, + 330, + 470, + 338 + ], + "type": "text", + "content": "Athlete running" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_caption" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 480, + 284, + 503, + 328 + ], + "blocks": [ + { + "bbox": [ + 480, + 284, + 503, + 328 + ], + "lines": [ + { + "bbox": [ + 480, + 284, + 503, + 328 + ], + "spans": [ + { + "bbox": [ + 480, + 284, + 503, + 328 + ], + "type": "image", + "image_path": "48fa0c23ae145450701724393702415f99df007b16c7c30f97749fb8dfdc3077.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 330, + 529, + 338 + ], + "lines": [ + { + "bbox": [ + 488, + 330, + 529, + 338 + ], + "spans": [ + { + "bbox": [ + 488, + 330, + 529, + 338 + ], + "type": "text", + "content": "Master lunging" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_caption" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 414, + 342, + 479, + 390 + ], + "blocks": [ + { + "bbox": [ + 414, + 342, + 479, + 390 + ], + "lines": [ + { + "bbox": [ + 414, + 342, + 479, + 390 + ], + "spans": [ + { + "bbox": [ + 414, + 342, + 479, + 390 + ], + "type": "image", + "image_path": "d51661c81751385a118c3f26220095c524d92fcc7b3bc62b6d73e701ed983dfb.jpg" + } + ] + } + ], + "index": 79, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 392, + 471, + 399 + ], + "lines": [ + { + "bbox": [ + 423, + 392, + 471, + 399 + ], + "spans": [ + { + "bbox": [ + 423, + 392, + 471, + 399 + ], + "type": "text", + "content": "Downward leaves" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_caption" + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 481, + 343, + 541, + 393 + ], + "blocks": [ + { + "bbox": [ + 481, + 343, + 541, + 393 + ], + "lines": [ + { + "bbox": [ + 481, + 343, + 541, + 393 + ], + "spans": [ + { + "bbox": [ + 481, + 343, + 541, + 393 + ], + "type": "image", + "image_path": "61f041eff042f551a3c8d666abfc62debdd211dd99c8150d3f0683cb0261cb9d.jpg" + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 392, + 523, + 399 + ], + "lines": [ + { + "bbox": [ + 495, + 392, + 523, + 399 + ], + "spans": [ + { + "bbox": [ + 495, + 392, + 523, + 399 + ], + "type": "text", + "content": "Bent stem" + } + ] + } + ], + "index": 82, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 420, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 54, + 420, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Demonstration of skeletal conditions for controllable generation. We argue that skeletons and text provide complementary description for shape and appearance respectively, as shown in (a). Moreover, flexible and accurate control of object anatomy and pose can be realized by editing the joints and bones in skeletons, as shown in (b). Arbitrary skeletal structures are supported in our framework. Multiple views are generated and only front-view images are shown." + } + ] + } + ], + "index": 84, + "angle": 0, + "type": "image_caption" + } + ], + "index": 81 + }, + { + "bbox": [ + 55, + 475, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 536 + ], + "type": "text", + "content": "tonization, we propose a new pipeline to generate skeletons with sparse joints from meshes. The pipeline mainly consists of curve skeleton extraction and curve simplification, achieving an " + }, + { + "bbox": [ + 55, + 475, + 295, + 536 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 475, + 295, + 536 + ], + "type": "text", + "content": " success rate, significantly outperforming previous deep learning based method RigNet [71]." + } + ] + } + ], + "index": 85 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "To fully control object anatomy and pose, we build the skeletal conditioned generation model in a multi-view manner. We represent a 3D skeleton with conditional skeleton images by Coordinate Color Encoding (CCE) to reduce ambiguity. Joints and bones are encoded with unique colors according to their 3D positions. For condition injection, we designed a Skeletal Correlation Module (SCM) to extract features from these conditional images and then generate multi-view images for the object. Later, a Large Reconstruction Model (LRM) is employed to produce 3D assets from the multi-view images. To address potential blurriness during reconstruction, we enhance appearance quality by a refinement stage that up-samples the multi-view images to higher resolutions and refines the texture in UV space." + } + ] + } + ], + "index": 86 + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "text", + "content": "The experimental results indicate that our framework achieves instant generation of multi-view and 3D contents which are aligned with given skeletons. The proposed coordinate color encoding and the skeletal correlation module significantly improve the object-skeleton alignment, and accelerate model convergence by " + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "text", + "content": " 3D assets conforming to the given skeleton can be generated in " + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{s}" + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "text", + "content": " and refined in " + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "inline_equation", + "content": "\\sim 60\\mathrm{s}" + }, + { + "bbox": [ + 313, + 475, + 555, + 596 + ], + "type": "text", + "content": ". To the best of our knowledge, this work is a pioneer in achieving arbitrary skeletal conditioned generation with following contributions:" + } + ] + } + ], + "index": 87 + }, + { + "bbox": [ + 314, + 616, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 90, + "blocks": [ + { + "bbox": [ + 314, + 616, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 553, + 662 + ], + "type": "text", + "content": "- Constructing the first large-scale dataset, Objaverse-SK, containing mesh and skeleton pairs that cover diverse skeletal structures. We developed a pipeline to generate sparse skeletons from meshes with a high success rate." + } + ] + } + ], + "index": 88 + }, + { + "bbox": [ + 314, + 663, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 663, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 663, + 553, + 711 + ], + "type": "text", + "content": "- Proposing a multi-view and 3D generation pipeline for arbitrary skeletons, including coordinate color encoding for compact condition representation and the skeletal correlation module for effective condition injection." + } + ] + } + ], + "index": 89 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "315" + } + ] + } + ], + "index": 91 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 67, + 541, + 168 + ], + "blocks": [ + { + "bbox": [ + 70, + 67, + 541, + 168 + ], + "lines": [ + { + "bbox": [ + 70, + 67, + 541, + 168 + ], + "spans": [ + { + "bbox": [ + 70, + 67, + 541, + 168 + ], + "type": "image", + "image_path": "bb419e83bedf16bb7189b8f83e0755967f849e5aa5905b801fed4f5e8d554f48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 177, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 555, + 200 + ], + "type": "text", + "content": "Figure 2. Illustration of the pipeline for skeleton generation from meshes (§3.2). The curve skeleton is first extracted from the given mesh, followed by simplification of parted curves. The curve graph is converted to a tree as the final skeleton." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 143, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 143, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 143, + 221 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 400 + ], + "type": "text", + "content": "Controllable 2D Generation. Based on image diffusion models [51-53], versatile controlling conditions have been studied. In terms of spatial control, ControlNet [75] and other similar works [40, 78] train a side network for spatial conditions such as edge maps, normal maps and human skeletons. Some works focus on human image generation from skeletons [18, 20, 62]. Box-based instance control is also concerned in some works [29, 79, 81]. As for conceptual controlling, DreamBooth [54] finetunes the model to bind the given subject with an identifier in text prompt. IP-Adapter [74] trains an adapter to inject styles or concepts into the model. Some works [31, 32, 63] also focus on human ID control. Besides, some methods [2, 38, 39, 41] can achieve conditional generation without fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 298, + 715 + ], + "type": "text", + "content": "Controllable 3D Generation. Controlling concepts in 3D generation can be easily achieved by combining image control methods with image-to-3D generation [17, 27, 59, 69]. However, spatial control for 3D generation is not as easy as conceptual control. Many works achieve spatial control based on score distillation paradigm [28, 46, 83]. Coin3D [10] presents a framework to control the multi-view diffusion and 3D generation by shape proxies, i.e. combination of simple basic shapes. Sculpt3D [6] enhances text-to-3D generation with retrieved 3D priors. Sherpa3D [34] proposes to generate a coarse shape with a 3D diffusion model and refine the shape with SDS [46]. Some works for 3D human or head avatar generation [19, 24, 33, 77, 80, 82] uses human skeleton or facial landmarks as the condition in 2D or 3D space. Recent 3D native diffusion models [7, 21, 25, 67, 76] indicate promising results. Clay [76] designs a transformer-based [45, 60] 3D diffusion framework and various conditions like images and point clouds can be injected through cross-attention layers. A recent work MVControl [30] realizes 3D generation with single-view 2D spatial conditions like normal maps and edge maps by conditional multi-view generation and 3D reconstruction. Our work shares the similar workflow, but we focus on skeletal conditioned generation, which is not studied in MVControl. Another recent work AnimatableDreamer [65] generates 4D objects with skeletons extracted from given" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 210, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 293 + ], + "type": "text", + "content": "videos by canonical score distillation, while our work develops a totally different framework for instant generation. As for skeletal representation, AnimatableDreamer extracts dense joints and unstructured bones from videos, while our method uses sparse joints and structured bones, which are more similar to human-made skeletons and more efficient for shape representation and animation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 294, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 556, + 475 + ], + "type": "text", + "content": "Mesh Skeletonization. Various algorithms were designed for extracting skeletons from 3D meshes. [58] and [1] compute curve skeletons (C-S) via iterative mesh contraction operations. [11, 66] proposed to extract skeletons medial axis transformation skeleton (MAT-S) by point selection and connection prediction. C-S and MAT-S can serve as shape representation, while human-made skeletons (H-S) are often different from them. Since the main purpose is animation, H-S only contain sparse joints and bones. Some works [70, 71] propose data-driven approaches to learn mesh skeletonization from human annotated data. In this work, we have tried learning-based method [71] but found the results were not satisfactory. Therefore, we develop a new pipeline to generate skeletons which are as sparse as H-S while keep the shape of C-S." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 484, + 440, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 440, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 440, + 498 + ], + "type": "text", + "content": "3. Dataset Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 504, + 419, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 419, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 419, + 517 + ], + "type": "text", + "content": "3.1. Data Preparation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": "The largest existing open dataset containing mesh-skeleton pairs is ModelResources [70]. There are around 3,000 3D meshes without textures. The scale is insufficient to train a text-driven generative model, and it lacks textures for appearance modeling. To address these limitations, we construct a dataset with " + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": " larger scale and includes color textures. Our dataset, named Objaverse-SK, is built upon a large-scale 3D dataset Objaverse [8]. Although our data generation pipeline is applicable to a broad range of object categories, we focus on three main categories including \"Animals\", \"Human Shapes\" and \"Plants\", as they can typically be represented by tree-structured skeletons. Category labels are obtained from G-Objaverse [47]. Consequently, our dataset contains 24k 3D meshes, consisting of 15k animals, 6k human shapes and 3k plants. Text prompts for these models are generated by Cap3D [36]." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "316" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 74, + 538, + 224 + ], + "blocks": [ + { + "bbox": [ + 71, + 74, + 538, + 224 + ], + "lines": [ + { + "bbox": [ + 71, + 74, + 538, + 224 + ], + "spans": [ + { + "bbox": [ + 71, + 74, + 538, + 224 + ], + "type": "image", + "image_path": "bc7efb92eb3dbc3c081563d6a8bc751ed0b04b53a2034ae245e8830e397fee18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 275 + ], + "type": "text", + "content": "Figure 3. Illustration of the pipeline for skeletal conditioned multi-view and 3D generation (§4). The 3D skeleton is projected into 2D images and represented by coordinate color encoding. The skeletal correlation is modeled by skeletal correlation module, and then is injected into the diffusion model. Multi-view images are first generated and then a 3D textured mesh is reconstructed. The texture is further refined through UV-space optimization. Our framework achieves instant and high-quality generation given arbitrary skeletons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 285, + 174, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 174, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 174, + 297 + ], + "type": "text", + "content": "3.2. Skeleton Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 435 + ], + "type": "text", + "content": "To obtain mesh-skeleton pairs, a effective method for generating skeletons from meshes is crucial. There are two concerns: the skeleton structure and success rate. The skeleton structure should accurately describe the object anatomy and be suitable for posing. Furthermore, an ideal method should generate reasonable skeleton structures with a high success rate. We tested a learning-based method RigNet [71] (Fig. 4). Although the generated skeleton structures can be close to human-made skeletons in its training data, it primarily produces symmetric skeletons and tends to be unstable across diverse anatomies and poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 616 + ], + "type": "text", + "content": "Skeleton extraction. To enhance flexibility and robustness, we design a new reliable pipeline, utilizing curve skeletons as the intermediate representation. An illustration of the pipeline is in Fig. 2. Given the structural inconsistency between curve skeletons and human-made skeletons, we further convert dense curves into sparse joints and bones. The detailed pipeline is elaborated below. 1) Initially, Mean Curvature Flow (MCF) [58] is employed to generate curve skeletons from meshes robustly. 2) Next, we build a graph from the set of curves generated from the mesh, consisting of dense nodes and edges. Intersection nodes (degree " + }, + { + "bbox": [ + 55, + 437, + 296, + 616 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 55, + 437, + 296, + 616 + ], + "type": "text", + "content": ") are identified and the graph is divided into several parts by these nodes. 3) In each part, the curve contains no branches so it can be simplified by the Douglas-Peucker algorithm (DP) [12] into line segments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Tree conversion. At this stage, the basic shape of the skeleton is established, but the root position and the bone directions between joints still need to be determined. The problem can be regarded as a graph-to-tree conversion. First, a spanning tree is constructed from the graph to eliminate cycles. We then identify high-degree intersection nodes as candidates for the root. To ensure an efficient structure, the skeleton is configured by selecting the tree with the min" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 286, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 334 + ], + "type": "text", + "content": "imum height among these candidates. This approach ensures that the root node is located at a significant intersection, minimizing the distances between the root and other joints. More details can be found in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 348, + 432, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 432, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 432, + 361 + ], + "type": "text", + "content": "4. Generation Pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 544, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 544, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 544, + 381 + ], + "type": "text", + "content": "4.1. Skeletal Conditioned Multi-view Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 471 + ], + "type": "text", + "content": "As the dataset is constructed, we consider building the conditional generative model based on it. Since unconditional multi-view diffusion models have been well studied, we start directly from a base model MVDream [55] and focus on the conditional generation. Two main issues are of concern: i) how the skeleton is represented, and ii) how it is injected into the model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": "Skeletal condition representation. As we aim to generate images which align with the given skeletons, using spatial guidance in the diffusion model is a reasonable approach. Skeletons can be projected from 3D space to the image plane as 2D conditions. However, overlapping and information loss occur during the projection, which may lead to semantic and structural ambiguities for spatial guidance, as illustrated in Fig. 8. Thus, incorporating richer information is crucial to mitigate these ambiguities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Coordinate Color Encoding (CCE). In order to preserve 3D information, we encode joint coordinates using spatial colors. While prior works [28, 61] use canonical color maps for shape representation, our approach focuses on representing skeletons with sparse joints and bones. We begin by normalizing skeletons within a canonical cube " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "[0,1]^3" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". Each position in this cube corresponds to a unique color, with RGB values precisely matching the positional coordinates. As a result, the 2D conditional image can represent the 3D spatial positions of the skeleton. For bones, we assign colors based on their midpoint. Additionally, we incor" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "317" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "porate normalized values of view-dependent inverse depth of the skeleton as the alpha channel (CCE-D). With the absolute spatial coordinates and relative depth encoded in the conditional images, there will be more precise and richer guidance information for generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": "Skeletal condition injection. Spatial conditions such as canny edges and normal maps have been investigated in 2D image diffusion models. In ControlNet [75], the conditional image is encoded by convolution blocks, resulting in an output spatial size that matches the latent size. Then, the condition features are added to the latent features. The encoder of the original diffusion model is copied as a side network to produce guidance features, which are fused with the original features in the decoder. Our pipeline adopts this paradigm from ControlNet, and further enhances it with a more effective condition feature extraction module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 517 + ], + "type": "text", + "content": "Skeletal Correlation Modeling (SCM). For a skeleton in 3D space, we first project it into multi-view images as 2D conditions. Given the sparse nature of skeletons in spatial dimension, convolution blocks lack global modeling capacity. To address this, we design a Skeletal Correlation Module (SCM) to enhance the condition features by modeling the anatomical correlation among different parts of a skeleton, and the view correlation for different projection views. The structure of the module is in Fig. 3. (i) First, anatomical correlation is extracted by a self-attention layer, which constructs the global skeleton features for each view. (ii) Then, the cross-view correlation is modeled by a cross-attention layer, allowing the extraction of correspondences among skeleton images from multiple views. This enables the model to recognize identical joints in different views. In addition, we use adaptive layer normalization [68] to fuse the camera pose embedding with the skeletal features. Associating each skeleton image with a camera pose helps to generate view-dependent shapes. Adding correlation layers during condition encoding significantly facilitates learning, achieving " + }, + { + "bbox": [ + 55, + 265, + 295, + 517 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 55, + 265, + 295, + 517 + ], + "type": "text", + "content": " faster convergence (Fig. 9)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 250, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 250, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 250, + 539 + ], + "type": "text", + "content": "4.2. Multi-view Images to 3D Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": "Instant reconstruction. Given the generated multi-view images, we use a Large Reconstruction Model (LRM), specifically InstantMesh [69] for fast textured mesh reconstruction. However, the reconstructed textures often appear blurry. On one hand, the resolution of generated images is " + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": ", which struggles to capture fine details. On the other hand, the appearance quality also degrades during reconstruction. In order to recover and further enrich the appearance, we introduce a new refinement stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Appearance refinement. First, the generated multi-view images are upscaled 4 times into " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " by Stable Diffusion with ControlNet-Tile [75]. ControlNet-Edge [9, 75] is used to maintain the shape consistency across different views during tiling. Once tiled, these high-resolution im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ages " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "I_{i}^{h}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " are used to refine the reconstructed texture. A learnable 2D texture " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " in UV space is created and initialized as the reconstructed texture " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "u_{0}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", and then images are rendered through differentiable rendering " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(u, c_{i})" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " for given camera views " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". The MSE loss is optimized between the rendered images and tiled high-res images. Moreover, a regularization term is added to maintain consistency in UV space:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 345, + 166, + 553, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 166, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 345, + 166, + 553, + 188 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\sum_ {i} \\left| \\left| I _ {i} ^ {h} - \\mathcal {R} (u, c _ {i}) \\right| \\right| _ {2} ^ {2} + \\lambda * \\left| \\left| u - u _ {0} \\right| \\right| _ {2} ^ {2}. \\tag {1}", + "image_path": "e637594e473481c4adc64c736b1e37026dcc3ca62217d719f9f9d31fea6bc040.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "type": "text", + "content": "Since the high-res images cannot cover every position on the mesh, some regions of " + }, + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "type": "text", + "content": " will not be optimized, e.g. bottom of the object. We found these regions are unstable during optimization and may produce unexpected artifacts (see Fig. 10). The regularization term helps the optimized texture maintain the appearance from " + }, + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "type": "inline_equation", + "content": "u_{0}" + }, + { + "bbox": [ + 313, + 198, + 555, + 318 + ], + "type": "text", + "content": " in these regions. Consequently, the high-frequency details can be learned in covered regions while the global consistency can also be achieved in uncovered regions. The optimization can be finished within 15 seconds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 328, + 518, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 518, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 518, + 341 + ], + "type": "text", + "content": "4.3. Object-Skeleton Alignment Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "text", + "content": "Contrastive alignment. In order to measure how much an object is aligned with a skeleton, we develop a new evaluator, named Contrastive Object-Skeleton Alignment (COSA). We use the self-supervised DINOv2 [44] as the backbone " + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "text", + "content": " to extract both object and skeleton features. Then, the alignment adapter " + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "text", + "content": " consisting of several self-attention layers is used to modulate the features. The adapter ends with an average pooling layer to aggregate the aligned features into a vector. Similar to CLIP [49], we train the adapter using contrastive learning by InfoNCE loss [42, 56]. Finally, the skeleton alignment score (SKA) can be calculated by cosine similarity between the features from an object image " + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "text", + "content": " and a skeleton image " + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 346, + 555, + 502 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 513, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 513, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 358, + 513, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {\\mathrm {S K A}} (x, y) = \\cos (G (F (x)), G (F (y))). \\tag {2}", + "image_path": "5204f1b1d2265c019188ae3ed63113456dcb9f5a797bd8563de66d075d49962b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "text", + "content": "COSA guided diffusion. Based on COSA, another conditional generation pipeline can also be realized, following the approach proposed in [2]. On each denoising time step " + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "text", + "content": ", the approximate clean image " + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "text", + "content": " is estimated from the predicted noise " + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "text", + "content": " as in DDIM [57]. The estimated clean image and skeleton condition are fed into COSA to calculate the alignment loss " + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{COSA}}(\\hat{x}_0,y) = 1 - S_{\\mathrm{SKA}}(\\hat{x}_0,y)" + }, + { + "bbox": [ + 313, + 536, + 555, + 643 + ], + "type": "text", + "content": ". Then the predicted noise is modified by the gradient of the alignment loss for actual denoising:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 654, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 654, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 372, + 654, + 553, + 666 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {t} = \\epsilon_ {t} + s (t) \\cdot \\nabla \\mathcal {L} _ {\\mathrm {C O S A}} (\\hat {x} _ {0}, y) \\tag {3}", + "image_path": "86f717d3b50ca6f65fae4f240b0e09b4b186fe003e3782588314ff2202eac3c1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " controls the guidance strength. With the additional guidance of the alignment loss, the generated object will tend to follow the conditional skeleton " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "318" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 549, + 139 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 549, + 139 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 549, + 139 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 549, + 139 + ], + "type": "table", + "html": "
Method\\SKA Score TrainingMeanInst.MeanClassAnimalsHumansPlantsApodesBipedsQuadrupedsArthropodsWings
SDEdit [38]○70.1365.0679.5064.7050.9974.7475.2482.7379.0779.21
SDEdit+COSAG●72.1167.3280.9167.6053.4679.7383.1785.9877.8073.82
Ours-Raw●74.6967.0988.3969.9042.9794.7482.1490.1889.8184.69
Ours-SCM●80.4374.3891.1678.4553.5394.4785.2094.1990.6888.40
", + "image_path": "acec4fb0014807ba5e9752964bdc131ce667ae90378461d45d06cf7e24576576.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 95, + 186, + 517, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 142, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 555, + 175 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of object-skeleton alignment (SKA) score (\\$5.2). Alignment scores are calculated over three classes (blue) and five subclasses of animal (green). The average scores over all instances and three classes (pink) are also shown. The highest scores among all methods are bold and the highest scores among baseline methods are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 186, + 517, + 264 + ], + "lines": [ + { + "bbox": [ + 95, + 186, + 517, + 264 + ], + "spans": [ + { + "bbox": [ + 95, + 186, + 517, + 264 + ], + "type": "table", + "html": "
MethodTrainingPickScoreCLIP Score
Win RateAnimalsHumanPlants\\( Mean_{Inst.} \\)AnimalsHumanPlants
SDEdit [38]○19.5618.2314.6628.9829.0429.4728.7128.29
SDEdit+COSAG\\( \\blacksquare \\)18.1718.5419.9515.0628.9929.5228.5028.11
Ours-Raw●28.9929.0628.3729.5529.9030.1030.0229.23
Ours-SCM●33.2834.1737.0226.4229.8330.2329.8428.75
", + "image_path": "bcf031ae3cfe895a1d563a8864d84ac3ee09ba1c5e21060cf46fec9d837cc140.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 289 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of PickScore and CLIP Score (§5.2). Scores are calculated over three classes (blue) and averaged over all instances (red). The highest scores among all methods are bold and the highest scores among baseline methods are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 298, + 287, + 425 + ], + "blocks": [ + { + "bbox": [ + 61, + 298, + 287, + 425 + ], + "lines": [ + { + "bbox": [ + 61, + 298, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 61, + 298, + 287, + 425 + ], + "type": "image", + "image_path": "4e5ddf8ef1448ba0aded0e61608692a204b41d8a063ee69935a38ec68f8e59f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 452 + ], + "type": "text", + "content": "Figure 4. Comparison of skeletons generated from 3D meshes by RigNet [71] and our method (§5.1)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 133, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 133, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 133, + 474 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 228, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 228, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 228, + 491 + ], + "type": "text", + "content": "5.1. Results of Mesh Skeletonization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "type": "text", + "content": "We compare our method with the learning-based method RigNet [71], and the results are shown in Fig. 4. RigNet tends to produce symmetric skeletons with limited flexibility and misaligned joints/bones, resulting in an around " + }, + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "type": "text", + "content": " success rate. In contrast, our method supports arbitrary-pose skeletons and achieves better joint/bone alignment. It produces more reliable results with an " + }, + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 497, + 296, + 594 + ], + "type": "text", + "content": " success rate. More details and results can be found in appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 600, + 234, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 234, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 234, + 612 + ], + "type": "text", + "content": "5.2. Results of Multi-view Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Evaluation protocols. The Objaverse-SK evaluation set contains 108 skeleton instances, covering three main classes: animals, human shapes and plants. As animals include diverse skeleton structures, we further divide them into more detailed subclasses (examples are shown behind): Apodes (fish, snakes), Bipeds (ducks, penguins), Quadrupeds (dogs, bears), Arthropods (scorpions, crabs), Wings (birds, dragons). Three evaluation metrics are con" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 372 + ], + "type": "text", + "content": "sidered for multi-view generation: SKA Score for skeletal alignment, PickScore for image quality, and CLIP Score for textual alignment. Furthermore, samples from categories excluded in training are obtained from ShapeNet [5] for generalization evaluation (Fig. 7). Additionally, evaluation results on human-made skeletons are in appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 373, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 529 + ], + "type": "text", + "content": "Baseline methods. Since there is no prior work that can achieve arbitrary skeletal conditioned generation, we implement two methods for comparison. The first baseline is SDEdit [38]. The process starts from adding noise on conditional images with a time step (set as 0.7). Then clean images are generated by denoising steps. The method is entirely unsupervised. The second baseline is the COSA Guidance (COSAG) derived from [2], which is elaborated in Section 4.3. The guidance strength is set as " + }, + { + "bbox": [ + 313, + 373, + 556, + 529 + ], + "type": "inline_equation", + "content": "s(t) = 7.5\\sqrt{1 - \\alpha_t}" + }, + { + "bbox": [ + 313, + 373, + 556, + 529 + ], + "type": "text", + "content": ". Since we found it cannot achieve stable results, it is combined with SDEdit. The method requires an extra model so it is half supervised. Ours is fully supervised on object-skeleton pairs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 531, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 651 + ], + "type": "text", + "content": "Qualitative comparison. The qualitative results are shown in Fig. 5. Given skeleton images as conditions, SDEdit can produce images following the skeleton. However, limited by the editing capacity, the generated objects often have incorrect anatomies. For example, the snake body is broken, and the donkey body is generated as wooden. When it is enhanced by the COSAG, the quality of generated content is improved in some cases but still unsatisfactory. Compared with them, our results show superior quality and are more consistent with both skeletal and textual conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Quantitative comparison. Comparison results of skeleton alignment are shown in Table 1. SDEdit-based methods have around 70 SKA scores, while ours can achieve 80. Among three classes, animals tend to have higher alignment scores while plants have lower scores. Since plants" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "319" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 302 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 302 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 302 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 302 + ], + "type": "image", + "image_path": "248e142f0b164f8274b9388b4f2b47475b3cb316226ed855a2933b2170ec15c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 306, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 555, + 330 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of skeletal conditioned multi-view generation (§5.2). Conditional skeletons and text prompts are shown above. Four views are generated and two views are shown for simplicity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 334, + 541, + 449 + ], + "blocks": [ + { + "bbox": [ + 67, + 334, + 541, + 449 + ], + "lines": [ + { + "bbox": [ + 67, + 334, + 541, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 334, + 541, + 449 + ], + "type": "image", + "image_path": "d571e216cc6080034d8efe37e6c35b3ae263d050c8b1f4bf5584bb102ff910bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 453, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 555, + 475 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of textured meshes before and after refinement (§5.3). Rendered color images are shown. Local areas are enlarged for better viewing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 546 + ], + "type": "text", + "content": "may have more complex structures and sometimes extend further from skeletons. For five sub-classes, our method achieves consistently high alignment scores. With the help of SCM and CCE-D skeletons, both the alignment scores and pick scores are further improved." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 197, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 197, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 197, + 567 + ], + "type": "text", + "content": "5.3. Results of 3D Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 573, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 633 + ], + "type": "text", + "content": "Texture refinement. Results of 3D reconstruction from multi-view images are shown in Fig. 6. The raw reconstructed and refined results are compared. The raw textures are blurry and lack details, while the proposed refinement stage can significantly enhance texture quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 693 + ], + "type": "text", + "content": "Rigging and animation. Given a motion sequence of a skeleton, our method can be applied to generate 4D animation. Textured mesh can be generated given the skeleton at the rest pose, and then directly be rigged and skinned for animation. Demo videos can be found in project page1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 485, + 407, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 485, + 407, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 407, + 499 + ], + "type": "text", + "content": "6. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 507, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 615 + ], + "type": "text", + "content": "Skeletal condition representation. The skeletal condition representation we use consists of coordinate color encoding (CCE) with depth alpha (D). The ablation results are shown in Fig. 8 and Tab. 3. Richer information in conditions can help the model determine the content better. As a result, better alignment can be achieved. In Fig. 8 right, the skeleton of a penguin is highly ambiguous. If CCE-D is used, the body pose and orientation of the penguin can be successfully inferred from colors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Skeletal correlation modeling. We show the effect of our skeletal correlation module in Tab. 3 and Fig. 9. SCM achieves better alignment scores than convolutional blocks. SCM with layer normalization (LN) achieves a " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " faster convergence rate. Furthermore, if LN is replaced with the adaptive LN (AdaLN), the model can achieve an SKA score of 70 within 1k training steps. The results indicate that extracting global features from conditional images is crucial." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 165, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 165, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 165, + 713 + ], + "type": "text", + "content": "1https://skdream3d.github.io/." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "320" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 289, + 287 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 289, + 287 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 289, + 287 + ], + "type": "image", + "image_path": "21dc4b36abee86086f91274e70d994423aa6ff898b31ccc02c4e4502d4973500.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 290, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 314 + ], + "type": "text", + "content": "Figure 7. Qualitative results of novel categories in ShapeNet [5] (§5.2), which are not covered by the training set of Objaverse-SK." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 327, + 287, + 422 + ], + "blocks": [ + { + "bbox": [ + 61, + 327, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 61, + 327, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 327, + 287, + 422 + ], + "type": "image", + "image_path": "f20c9f0f6322eb23b4bcc7a61cd5038f277d74f8689d33145806fd728ef901b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 426, + 295, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 460 + ], + "type": "text", + "content": "Figure 8. Ablation study of coordinate color encoding with depth alpha (CCE-D) (" + }, + { + "bbox": [ + 55, + 426, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\S 6" + }, + { + "bbox": [ + 55, + 426, + 295, + 460 + ], + "type": "text", + "content": "). Richer information can help the model to avoid ambiguity and generate better anatomy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 478, + 299, + 552 + ], + "blocks": [ + { + "bbox": [ + 56, + 478, + 299, + 552 + ], + "lines": [ + { + "bbox": [ + 56, + 478, + 299, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 299, + 552 + ], + "type": "table", + "html": "
ModuleSkeletonSKA ScoreAnimalsHumansPlants
SCMCCE-D80.4391.1678.4553.53
SCMCCE78.9790.4574.8252.57
SCMBinary77.2789.2076.0446.16
ConvCCE-D76.6588.5375.0446.17
ConvBinary74.6988.3969.9042.97
", + "image_path": "95c30b52a53b11c456c5801a19825d1a7432fbc72c9dd16cccb0685aa81070d7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "text", + "content": "Table 3. Ablation study of skeletal module and representation types (§6). SCM and CCE-D achieve higher alignment scores." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Refinement regularization. We show the ablation results of appearance refinement in Fig. 10. The refined appearance contains rich details such as snake scales and wood grain, in comparison to the reconstructed results. However, artifacts also appear in the regions that are not covered by high-res images. With the help of UV space regularization, the artifacts are effectively removed in uncovered regions. As a result, consistent colors are maintained from original textures and details are enhanced during optimization." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 73, + 539, + 173 + ], + "blocks": [ + { + "bbox": [ + 328, + 73, + 539, + 173 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 539, + 173 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 539, + 173 + ], + "type": "image", + "image_path": "90325cbbee12e7ab61a01307d47912cc129fb6fe145131613362f7729b3d5fbe.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 175, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 314, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 314, + 175, + 553, + 198 + ], + "type": "text", + "content": "Figure 9. Convergence processes of different skeletal modules (§6). SCM with AdaLN achieves 5x faster convergence." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 208, + 423, + 314 + ], + "blocks": [ + { + "bbox": [ + 321, + 208, + 423, + 314 + ], + "lines": [ + { + "bbox": [ + 321, + 208, + 423, + 314 + ], + "spans": [ + { + "bbox": [ + 321, + 208, + 423, + 314 + ], + "type": "image", + "image_path": "e2279ee273e3641a6df134415d244219e7e00bb6e24d2f692e65d2b6774fafae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 318, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 352 + ], + "type": "text", + "content": "Figure 10. Ablation study of UV space regularization (§6). Bottom views and UV textures are shown. Front views of the snake and the tree stump can be found in the first column of Fig. 6." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 443, + 209, + 544, + 314 + ], + "blocks": [ + { + "bbox": [ + 443, + 209, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 443, + 209, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 443, + 209, + 544, + 314 + ], + "type": "image", + "image_path": "25e3dacd578b05391db044f194cb5e08c46a16fc76cb7db90f8fcd4cd60a7202.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 361, + 476, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 476, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 476, + 374 + ], + "type": "text", + "content": "7. Limitation and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 382, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 526 + ], + "type": "text", + "content": "Although our work achieves arbitrary skeletal conditioned generation, there are still many problems that can be further studied. The skeletons we currently use may have limited descriptive ability for non-tree structured objects. More powerful shape representations can be studied as new conditions. In addition, our work only considers global skeletons without fine-grained semantics. Injecting detailed semantics into the skeleton parts is a meaningful topic to study. For future work, the skeleton condition can be studied in native 3D generation frameworks. Moreover, our dataset may also be used in other tasks such as arbitrary skeleton estimation from images. More discussion is in the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose to use skeletons as the structural condition for controllable generation. First, we construct a large-scale 3D mesh-skeleton paired dataset. We propose an effective mesh skeletonization method to generate mesh-aligned sparse skeletons with a high success rate. Based on the dataset, we present a skeletal conditioned multi-view generation pipeline. Coordinate color encoding and skeletal correlation module are proposed to realize efficient condition representation and injection. Furthermore, 3D meshes can be instantly reconstructed, followed by a refinement stage to achieve better texture quality. In summary, our work achieves controllable multi-view and 3D generation with arbitrary skeletons as conditions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "321" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (No. 62293554), and in part by the Natural Science Foundation of Zhejiang Province (LDT23F02023F02) and the Fundamental Research Funds for the Zhejiang Provincial Universities (226-2024-00208)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 180, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 295, + 213 + ], + "type": "text", + "content": "[1] Andreas Bærentzen and Eva Rotenberg. Skeletonization via local separators. ACM Transactions on Graphics (TOG), 40 (5):1-18, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 296, + 280 + ], + "type": "text", + "content": "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 281, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 295, + 314 + ], + "type": "text", + "content": "[3] Ilya Baran and Jovan Popovic. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 315, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 295, + 358 + ], + "type": "text", + "content": "[4] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 359, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 296, + 413 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "type": "text", + "content": "[6] Cheng Chen, Xiaofeng Yang, Fan Yang, Chengzeng Feng, Zhoujie Fu, Chuan-Sheng Foo, Guosheng Lin, and Fayao Liu. Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10228-10237, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 535 + ], + "type": "text", + "content": "[7] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 602 + ], + "type": "text", + "content": "[8] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 624 + ], + "type": "text", + "content": "[9] Lijun Ding and Ardeshir Goshtasby. On the canny edge detector. Pattern recognition, 34(3):721-725, 2001. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 58, + 625, + 295, + 680 + ], + "type": "text", + "content": "[10] Wenqi Dong, Bangbang Yang, Lin Ma, Xiao Liu, Liyuan Cui, Hujun Bao, Yuewen Ma, and Zhaopeng Cui. Coin3d: Controllable and interactive 3d assets generation with proxy-guided conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[11] Zhiyang Dou, Cheng Lin, Rui Xu, Lei Yang, Shiqing Xin, Taku Komura, and Wenping Wang. Coverage axis: Inner point selection for 3d shape skeletonization. In Com" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "puter Graphics Forum, pages 419-432. Wiley Online Library, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[12] David H Douglas and Thomas K Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature. Cartographica: the international journal for geographic information and geovisualization, 10 (2):112-122, 1973. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[13] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Navigation instruction generation with bev perception and large language models. In ECCV, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "text", + "content": "[14] Sheng Fan, Rui Liu, Wenguan Wang, and Yi Yang. Scene map-based prompt tuning for navigation instruction generation. In CVPR, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 284 + ], + "type": "text", + "content": "[15] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "text", + "content": "[17] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 555, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 555, + 410 + ], + "type": "text", + "content": "[18] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 455 + ], + "type": "text", + "content": "[19] Yukun Huang, Jianan Wang, Ailing Zeng, He Cao, Xianbiao Qi, Yukai Shi, Zheng-Jun Zha, and Lei Zhang. Dreamwaltz: Make a scene with complex 3d animatable avatars. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[20] Xuan Ju, Ailing Zeng, Chenchen Zhao, Jianan Wang, Lei Zhang, and Qiang Xu. Humansd: A native skeleton-guided diffusion model for human image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15988-15998, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[21] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "text", + "content": "[22] Ladislav Kavan, Steven Collins, Jiří Žára, and Carol O'Sullivan. Skinning with dual quaternions. In Proceedings of the 2007 symposium on Interactive 3D graphics and games, pages 39-46, 2007. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 634 + ], + "type": "text", + "content": "[23] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "322" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[26] Christoph Leiter, Ran Zhang, Yanran Chen, Jonas Belouadi, Daniil Larionov, Vivian Fresen, and Steffen Eger. Chatgpt: A meta-analysis after 2.5 months. arXiv preprint arXiv:2302.13795, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[27] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "text", + "content": "[28] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. arXiv preprint arXiv:2310.02596, 2023. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[29] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "type": "text", + "content": "[30] Zhiqi Li, Yiming Chen, Lingzhe Zhao, and Peidong Liu. Mycontrol: Adding conditional control to multi-view diffusion for controllable text-to-3d generation. arXiv preprint arXiv:2311.14494, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "text", + "content": "[31] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8640-8650, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[32] Chao Liang, Fan Ma, Linchao Zhu, Yingying Deng, and Yi Yang. Caphuman: Capture your moments in parallel universes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6400-6409, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "text", + "content": "[33] Tingting Liao, Hongwei Yi, Yuliang Xiu, Jiaxiang Tang, Yangyi Huang, Justus Thies, and Michael J Black. Tada! text to animatable digital avatars. In 2024 International Conference on 3D Vision (3DV), pages 1508-1519. IEEE, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[34] Fangfu Liu, Diankun Wu, Yi Wei, Yongming Rao, and Yueqi Duan. Sherpa3d: Boosting high-fidelity text-to-3d generation via coarse 3d prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20763–20774, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[35] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[36] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[37] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2640-2649, 2017. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[38] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "type": "text", + "content": "[39] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7465-7475, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[40] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[41] Marianna Ohanyan, Hayk Manukyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Zero-painter: Training-free layout control for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8764-8774, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[42] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 483, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 483, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 483, + 342 + ], + "type": "text", + "content": "[43] OpenAI. Gpt-4 technical report, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[44] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[45] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 478 + ], + "type": "text", + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 555 + ], + "type": "text", + "content": "[47] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9914–9925, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "text", + "content": "[48] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 591, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 656 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(1):5485-5551, 2020. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "323" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[51] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "type": "text", + "content": "[54] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "text", + "content": "[55] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[56] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "text", + "content": "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 441 + ], + "type": "text", + "content": "[58] Andrea Tagliasacchi, Ibraheem Alhashim, Matt Olson, and Hao Zhang. Mean curvature skeletons. In Computer Graphics Forum, pages 1735-1744. Wiley Online Library, 2012. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "type": "text", + "content": "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 510 + ], + "type": "text", + "content": "[60] Ashish Vaswani. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "type": "text", + "content": "[61] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[62] Jiajun Wang, Morteza Ghahremani, Yitong Li, Björn Omer, and Christian Wachinger. Stable-pose: Leveraging transformers for pose-guided text-to-image generation. arXiv preprint arXiv:2406.02485, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[63] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "text", + "content": "[64] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[65] Xinzhou Wang, Yikai Wang, Junliang Ye, Fuchun Sun, Zhengyi Wang, Ling Wang, Pengkun Liu, Kai Sun, Xintong Wang, Wende Xie, et al. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "type": "text", + "content": "[66] Zimeng Wang, Zhiyang Dou, Rui Xu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Shiqing Xin, Taku Komura, Xiaoming Yuan, and Wenping Wang. Coverage axis++: Efficient inner point selection for 3d shape skeletonization. In Computer Graphics Forum, page e15143. Wiley Online Library, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[67] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "text", + "content": "[68] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "text", + "content": "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 1, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[70] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, and Karan Singh. Predicting animation skeletons for 3d articulated models via volumetric nets. In 2019 international conference on 3D vision (3DV), pages 298-307. IEEE, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[71] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, and Karan Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2, 3, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "type": "text", + "content": "[72] Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies. Frontiers of Information Technology & Electronic Engineering, 22(12):1551-1558, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "text", + "content": "[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 577 + ], + "type": "text", + "content": "[74] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[76] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[77] Muxin Zhang, Qiao Feng, Zhuo Su, Chao Wen, Zhou Xue, and Kun Li. Joint2human: High-quality 3d human generation via compact spherical embedding of 3d joints. In Pro" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "324" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 408 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1438, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[79] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 262 + ], + "type": "text", + "content": "[80] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavator: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21057-21067, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[81] Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Mige: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818-6828, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[82] Zhenglin Zhou, Fan Ma, Hehe Fan, Zongxin Yang, and Yi Yang. Headstudio: Text to animatable head avatars with 3d gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 408 + ], + "type": "text", + "content": "[83] Wenjie Zhuo, Fan Ma, Hehe Fan, and Yi Yang. Vividdreamer: invariant score distillation for hyper-realistic text-to-3d generation. In European Conference on Computer Vision, pages 122–139. Springer, 2024. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "325" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_content_list.json b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd01db35c350cd0dcca2a05976e061f004e13ff6 --- /dev/null +++ b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_content_list.json @@ -0,0 +1,1453 @@ +[ + { + "type": "text", + "text": "SKE-Layout: Spatial Knowledge Enhanced Layout Generation with LLMs", + "text_level": 1, + "bbox": [ + 119, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junsheng Wang $^{1*}$ Nieqing Cao $^{2*}$ Yan Ding $^{3}$ Mengying Xie $^{1}$ Fuqiang Gu $^{1}$ Chao Chen $^{1\\dagger}$ $^{1}$ Chongqing University $^{2}$ Xi'an Jiaotong-Liverpool University $^{3}$ Shanghai AI Lab", + "bbox": [ + 107, + 179, + 887, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{csjunshengwang, xiemy, gufq, cschaochen}@cqu.edu.cn", + "bbox": [ + 267, + 218, + 727, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nieqing.Cao@xjtlu.edu.cn, yding25@binghamton.edu", + "bbox": [ + 281, + 236, + 707, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating layouts from textual descriptions by large language models (LLMs) plays a crucial role in precise spatial reasoning-induced domains such as robotic object rearrangement and text-to-image generation. However, current methods face challenges in limited real-world examples, handling diverse layout descriptions and varying levels of granularity. To address these issues, a novel framework named Spatial Knowledge Enhanced Layout (SKE-Layout), is introduced. SKE-Layout integrates mixed spatial knowledge sources, leveraging both real and synthetic data to enhance spatial contexts. It utilizes diverse representations tailored to specific tasks and employs contrastive learning and multitask learning techniques for accurate spatial knowledge retrieval. This framework generates more accurate and fine-grained visual layouts for object rearrangement and text-to-image generation tasks, achieving improvements of $5\\% -30\\%$ compared to existing methods.", + "bbox": [ + 89, + 318, + 485, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 606, + 220, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating layouts from textual descriptions is fundamental in various domains, particularly in robotic object rearrangement and text-to-image generation, where achieving precise object placement is critical. For example, an object rearrangement task might require arranging specific items on a table [1, 8], while an image generation task often involves creating a visually coherent scene based on descriptions of object locations [29, 39]. Both scenarios underscore the necessity of accurate spatial understanding to ensure proper placement and prevent errors such as misalignment or omissions. Research highlights the need to minimize these errors in robotic manipulation and image generation in order to reflect spatial instructions accurately [2].", + "bbox": [ + 89, + 632, + 482, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, a growing body of research has investigated the application of large language models (LLMs) for layout generation [9, 29, 39], highlighting their no", + "bbox": [ + 89, + 829, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "table advantages in various aspects. Primarily, LLMs exhibit strong zero-shot learning capabilities, generating coherent layouts from natural language descriptions without task-specific training, enabling broad generalization across diverse tasks [37]. Unlike traditional models, LLMs adapt seamlessly to different domains without requiring separate training for each task [19, 41].", + "bbox": [ + 511, + 287, + 905, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although LLMs exhibit numerous advantages in layout generation, they also face challenges related to hallucination when directly generating layouts from textual descriptions. Existing approaches often rely on using predefined layouts as examples to guide LLMs in generating new layouts, allowing the model to infer reasonable spatial arrangements based on these examples [9]. However, there is often a lack of sufficient and diverse examples in real-world scenarios, particularly for complex layouts, exacerbating the model's dependence on existing examples. As representations, layouts often fail to fully capture the spatial knowledge of the current scene [14, 30]. When handling fine-grained instructions related to angles, distances, and other spatial details, LLMs struggle to infer the corresponding layout solely by relying on example layouts. Furthermore, due to the varying importance of information across different tasks, directly applying existing retrieval methods may result in the inability to retrieve the most relevant knowledge for the specific task at hand [28].", + "bbox": [ + 511, + 397, + 905, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose SKE-Layout, the first framework capable of handling diverse layout task descriptions and inferring unknown object attributes, providing a unified solution that current frameworks cannot achieve. Specifically, SKE-Layout extracts diverse spatial knowledge from real-world or synthetic tasks and injects the most relevant spatial knowledge into LLMs, enabling them to generate ideal layouts based on textual conditions. Despite not using any image data for training, LLMs can learn spatial knowledge through contextual demonstrations and then apply this knowledge to visual reasoning for new samples [20, 34]. In other words, by providing relevant examples, LLMs can capture spatial relationships between objects based on language understanding, in", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, †Corresponding author", + "bbox": [ + 107, + 886, + 344, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19414", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferring information, such as relative positions, sizes, and distances of objects, thereby achieving accurate object rearrangement and coherent image generation [6].", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are threefold:", + "bbox": [ + 109, + 138, + 439, + 152 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel approach SKE.layout with LLMs for generating layouts from textural descriptions more precisely. SKE.layout integrates multiple spatial knowledge sources and combines real-world spatial datasets with model-generated data from LLMs, enhancing the spatial knowledge database without relying on existing datasets.", + "- We present a diverse framework for spatial knowledge representations. It adapts to the specific needs of different tasks by providing precise quantitative expressions (e.g., \"30-degree angle\", \"10 cm away\") and qualitative spatial relations (e.g., \"zebra to the left of a giraffe\").", + "- We construct an innovative spatial knowledge retrieval method, which enhances spatial context extraction by combining contrastive learning with multitask learning, optimizing accuracy for complex visual layout tasks." + ], + "bbox": [ + 89, + 154, + 482, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, the SKE-Layout framework can effectively address the challenges in generating accurate and fine-grained visual layouts by leveraging multiple spatial knowledge sources, diverse representations, and innovative retrieval techniques. Its versatility allows it to simultaneously handle object rearrangement and image generation tasks, demonstrating its generalizability across different layout generation domains.", + "bbox": [ + 89, + 381, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 518, + 232, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Object Rearrangement", + "text_level": 1, + "bbox": [ + 89, + 545, + 303, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object rearrangement is a crucial task for service robots, requiring the movement of objects from one location to another and arranging them according to specified targets. This task is exemplified by challenges such as the Habitat Rearrangement Challenge [35] and the AI2-THOR Rearrangement Challenge [38]. Various approaches have been explored to tackle this problem. One deep learning-based method, StructFormer [24], utilized transformer-based neural networks to learn fine-grained object arrangement representations from real-world datasets, enabling the model to infer semantically meaningful structures from natural language instructions. However, this method presented challenges in terms of model complexity, as each task requires independent dataset construction and training, placing high demands on data collection and computational resources. Additionally, its generalization capabilities were limited, as the model struggled to handle tasks it failed to be specifically trained on. Another approach, DALL-E-Bot [18], leveraged DALL-E 2 [31] to generate images, which were then used for object placement in table-top scenarios. Nonetheless, DALL-E's image generation lacked sufficient stability when it came to producing task-relevant images,", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which could hinder object arrangement accuracy. To address these limitations, methods based on LLMs have been proposed [8, 15, 17, 32, 36, 40]. LLM-GROP [8] extracted commonsense object arrangement knowledge from LLMs and applied this knowledge to perform complex manipulation and object placement tasks. SayPlan [32] introduced a novel approach to grounding LLMs in 3D scene graphs, enabling scalable task planning by translating high-level instructions into structured scene graph representations and facilitating object manipulation in complex environments. SG-Bot [40] proposed a coarse-to-fine strategy for object rearrangement, leveraging scene graphs to guide robotic imagination from scene understanding to detailed object placements. This method emphasized hierarchical planning, where initial rough layouts were refined through iterative adjustments. Despite the potential of LLMs in facilitating object rearrangement, their spatial reasoning capabilities still face limitations, particularly in accurately interpreting and generalizing complex spatial relationships across diverse scenarios.", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Layout Generation", + "text_level": 1, + "bbox": [ + 511, + 402, + 748, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image layout generation is a critical component in accurately creating complex images. Various approaches have been proposed to generate layouts for images, each utilizing different models and input conditions to accomplish the task. Deep learning-based methods, such as LayoutGAN [21], LayoutVAE [16], and LayoutTransformer [13], typically require specific scenarios or input conditions for training, producing layouts based on fixed class labels or partially input layouts. On the other hand, LLM-based methods represent a more advanced approach [9, 10, 27, 29, 39, 42]. LayoutGPT constructed a visual layout framework from textual descriptions, extracting spatial knowledge from image datasets, such as object bounding boxes, and encoding this knowledge using CSS-style formatting. The spatial knowledge was then fed into LLM prompts, enabling the model to leverage contextual understanding to generate higher-level and lower-level layouts, which were subsequently translated into image layouts by image generation models. However, challenges remained, particularly in handling fine-grained textual descriptions related to angles and distances, as well as the generalization issues stemming from the potential lack of relevant images in current datasets for specific tasks. Building on this, GLIGEN [23] introduced bounding boxes and regional descriptions as input conditions for generating images.", + "bbox": [ + 511, + 424, + 903, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 814, + 604, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, we propose the SKE-Layout framework for generating generalizable layouts from textual descriptions. The framework is structured in two main parts: Knowledge Base Construction and Layout Gener", + "bbox": [ + 511, + 839, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd16667849688fe10c39a69ee06e69a55c0f1112374fc5b253512b640ada4b85.jpg", + "image_caption": [ + "Figure 1. Framework of SKE-Layout. The framework comprises two main parts: Knowledge Base Construction, which builds a database from real-world images and LLM-generated knowledge; and Layout Generation Pipeline, which includes (a) Input, (b) Knowledge Retrieval, (c) Layout Planning and (d) Simulation." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 318, + 385 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2260e8b230a4ff97a95f2401166b82dcf16c40d2865de3b8cd54d85cea393892.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 88, + 898, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ation Process. Knowledge Base Construction (Sec. 3.2) gathers task knowledge from two aspects. The first aspect involves real-world extraction, using a Visual Language Model (VLM) for scene interpretation and a Detection Transformer (DETR) [4] for object detection and bounding box generation. The second aspect focuses on task-specific knowledge generation using LLMs. Together, these aspects form a robust foundation for layout generation. After the knowledge base is built, the Layout Generation Process proceeds with the Input module, which employs the same real-world extraction methods as the Knowledge Base Construction. Then, the Knowledge Retrieval module (Sec. 3.4) enhances the system's understanding through data augmentation and multi-task learning, optimizing user instructions using relevant knowledge from the knowledge base. Subsequently, the Layout Planning module (Sec. 3.5) integrates all inputs to create a detailed and optimized layout that aligns with user instructions. Finally, the Simulation module (Sec. 3.6) implements the layout in a virtual environment, providing visual verification to ensure task accuracy and consistency with the given description.", + "bbox": [ + 88, + 469, + 483, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Settings", + "text_level": 1, + "bbox": [ + 89, + 801, + 259, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, we establish a unified problem setting for layout generation, encompassing both object rearrangement and image generation tasks. We define a reference coordinate system and origin tailored to each task. For object rearrangement, the reference is a 3D coordinate system", + "bbox": [ + 88, + 824, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$(x,y,z)\\in \\mathbb{R}^3$ , with the origin at the top-left corner of the rearrangement space, while for image generation, it is a 2D coordinate system $(x,y)\\in \\mathbb{R}^2$ , with the origin at the topleft corner of the image.", + "bbox": [ + 511, + 469, + 906, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The spatial information of an object, referred to as its layout, is represented by a tuple", + "bbox": [ + 511, + 530, + 905, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no = (\\mathrm {i d}, c, x, y, z, w, h, d, \\theta),\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 571, + 805, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where id denotes the object name, $c$ is the color, $(x,y,z)$ represents the object's center coordinates, and $w$ , $h$ , $d$ , $\\theta$ denote the bounding box's width, height, depth, and the rotation angle in the $x-y$ plane, respectively. In image generation tasks, only the 2D coordinates $(x,y)$ and bounding box dimensions $(w,h)$ are relevant, while 3D attributes like $z$ , $d$ , and $\\theta$ are often missing or irrelevant. In contrast, object rearrangement tasks require inferring the 3D coordinates $(x,y,z)$ and the rotation angle $\\theta$ , while color and dimensions $w$ , $h$ , $d$ are known. Given a textual description $\\mathcal{I}$ , the objective of layout generation is to produce a set of tuples $\\mathcal{O}$ , where each tuple $o\\in \\mathcal{O}$ captures the spatial attributes of the objects as described.", + "bbox": [ + 511, + 598, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial knowledge is defined as information related to the positioning, orientation, and arrangement of objects within a given environment, expressed as pairs of either $\\{\\text{Instruction - Object}\\}$ , where $\\mathcal{O} = \\{o_1, o_2, \\ldots, o_i\\}$ , or $\\{\\text{Instruction - Mathematical constraint}\\}$ , where the latter represents numerical constraints related to object positioning. In this context, 'knowledge' specifically refers to spa", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19416", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tial knowledge for the sake of brevity.", + "bbox": [ + 91, + 90, + 341, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Knowledge Base Construction", + "text_level": 1, + "bbox": [ + 89, + 122, + 357, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Retrieval-Augmented Generation (RAG) [12], constructing a comprehensive knowledge base is essential. This repository of exemplars and contextual information enhances the model's generative capabilities, allowing the system to retrieve relevant examples and produce precise, contextually appropriate layouts, thereby improving the efficacy and reliability of RAG.", + "bbox": [ + 89, + 147, + 482, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Extracting Knowledge from Real-World Images", + "text_level": 1, + "bbox": [ + 89, + 267, + 470, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary aspect for knowledge base construction is to extract information from real-world images. We use VLMs and DETR for object detection and bounding box acquisition, capturing object location and dimensions to build a detailed representation of real-world scenes. By aggregating layouts from diverse images, we create a rich knowledge base that accurately reflects spatial arrangements and object interactions. The extracted information provides tuples $o \\in \\mathcal{O}_{2D}$ and $o \\in \\mathcal{O}_{3D}$ , with each tuple representing object category, location, size, and orientation. Additionally, VLMs capture spatial relationships between objects, stored as instruction information $I$ for each image in the knowledge base, aiding in retrieving relevant images and spatial context for layout generation based on new instructions.", + "bbox": [ + 89, + 287, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Generating Knowledge with LLMs", + "text_level": 1, + "bbox": [ + 89, + 513, + 380, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second aspect uses LLMs to generate diverse tasks $I$ and layouts $O$ from existing instructions collected in other research[8, 22, 26]. By interpreting and synthesizing text, LLMs create tasks that span a broad spectrum of real-world and hypothetical scenarios, introducing variability crucial for building a versatile knowledge base. This process begins by inputting varied instructions into the LLMs, which produces corresponding tasks not as duplicates but as novel scenarios. For each task, the LLMs generate layouts specifying object categories, spatial arrangements, and interobject relationships. This strategy enriches the knowledge base with synthetic scenarios that improve the system's adaptability to complex and rare instructions. Mathematical constraints $M = \\{m_{n}\\}$ are also integrated to meet spatial requirements like distances and angles, ensuring layout accuracy.", + "bbox": [ + 89, + 536, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address potential inaccuracies, the synthetic data undergoes rigorous validation. First, VLMs verify if the generated images satisfy the original instructions (e.g., object arrange requirement, image composition) by evaluating object categories, positions, and spatial relationships. If discrepancies are found, human evaluators review the images for realism and relevance. This validation ensures the quality and reliability of data added to the RAG system.", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Input", + "text_level": 1, + "bbox": [ + 513, + 90, + 594, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section describes the configuration of inputs for different task settings within the SKE-Layout framework. For the image generation task, the input consists solely of user instructions, which outline the requirements for the generated image layout. The object rearrangement task involves a more complex input configuration. In addition to user instructions, it incorporates robot observations, which provide visual data of the environment and capture the spatial layout and positioning of objects within the scene through VLMs and DETR methods, as described in Sec. 3.2.", + "bbox": [ + 511, + 114, + 906, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Knowledge Retrieval", + "text_level": 1, + "bbox": [ + 511, + 281, + 710, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After constructing our knowledge base, we need to retrieve knowledge in the database based on the current user instruction. Given a test instruction $I$ and the knowledge base $D = \\{(I_k, o_k) \\mid k = 1, 2, \\ldots\\}$ , we define a function $f(I_k, I_j) \\in \\mathbb{R}$ that measures the distances between two instructions. Initially, we experiment with directly using existing text embedding models [3, 7] to convert instructions into vectors and calculate their similarity. However, this approach reveals a significant issue: different tasks emphasize different aspects of the information. For example, in the instruction \"place the banana $30^\\circ$ above the apple\", the specific fruits are less relevant than their spatial relationship. Standard embedding models can skew the retrieval results by overemphasizing the objects rather than their spatial arrangement.", + "bbox": [ + 511, + 305, + 906, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1. Instance-Wise Contrastive Learning", + "text_level": 1, + "bbox": [ + 511, + 545, + 812, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we train a custom text embedding model through contrastive learning, building on the architecture of BERT [25] to meet our retrieval needs. This model is tailored for specific tasks, emphasizing the importance of spatial relationships over individual object identities, especially in object placement scenarios. To further enhance robustness, we apply data augmentation techniques using LLMs. For instance, in the instruction \"place the banana $30^{\\circ}$ above the apple,\" we generalize it to \"place object A $30^{\\circ}$ above object B,\" allowing the model to prioritize spatial relationships while ignoring specific object identities.", + "bbox": [ + 511, + 566, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by SimCSE [11], let $S = [I_n]_{n=1}^N$ be the set of original instructions, and $S' = [I_n, I_n']_{n=1}^N$ be the set after data augmentation, where $I_n'$ is the augmented instruction derived from $I_n$ . We use a pre-trained BERT model $\\psi$ to obtain the embeddings of the instructions. For each instruction $I_n$ , we define its unique positive pair as $(I_n, I_n')$ , and all other pairs are treated as negative samples. The similarity between the two embeddings is calculated as follows:", + "bbox": [ + 511, + 734, + 908, + 856 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {m}\\right)\\right) = \\frac {\\psi \\left(I _ {n}\\right) ^ {T} \\psi \\left(I _ {m}\\right)}{\\| \\psi \\left(I _ {n}\\right) \\| _ {2} \\| \\psi \\left(I _ {m}\\right) \\| _ {2}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 869, + 906, + 906 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The contrastive loss function is defined as:", + "bbox": [ + 89, + 90, + 374, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {n} ^ {N} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right) , \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right)\\right)}{\\sum_ {m = 1 , m \\neq n} ^ {2 N} \\exp \\left(\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right) , \\psi \\left(I _ {m}\\right)\\right)\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 112, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall contrastive loss is the average of all individual losses:", + "bbox": [ + 89, + 157, + 482, + 185 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {I n s t a n c e - C L}} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {2 N} \\ell_ {n} ^ {N} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 185, + 482, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Training Procedure for Multi-task Instance-wise Contrastive Learning", + "bbox": [ + 89, + 247, + 482, + 277 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Instruction set $S = \\{S_{\\mathrm{image}}, S_{\\mathrm{object}}\\}$ containing tasks for image generation and object manipulation", + "2: Output: Trained model", + "3: for each $I_{n} \\in S$ do", + "4: Generate augmented instruction $I_{n}^{\\prime}$ using LLMs", + "5: end for", + "6: Obtain embeddings $\\psi(I_n)$ and $\\psi(I_n')$ for all instructions using task-specific embedding heads in the pretrained BERT model", + "7: for each $I_{n} \\in S_{\\mathrm{image}}$ do", + "8: Calculate the image generation contrastive loss:" + ], + "bbox": [ + 99, + 282, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {n} ^ {\\text {I m a g e - C L}} = \\operatorname {L o s s} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 459, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "9: Update model parameters to minimize $\\mathcal{L}_{\\mathrm{Image - CL}}$", + "10: end for", + "11: for each $I_{n} \\in S_{\\mathrm{object}}$ do", + "12: Calculate the object manipulation contrastive loss:" + ], + "bbox": [ + 94, + 503, + 480, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {n} ^ {\\text {O b j e c t - C L}} = \\operatorname {L o s s} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 590, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "13: Update model parameters to minimize $\\mathcal{L}_{\\mathrm{Object - CL}}$", + "14: end for", + "15: Compute the overall multi-task contrastive loss:" + ], + "bbox": [ + 94, + 635, + 452, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {M u l t i - T a s k - C L}} = \\lambda_ {\\text {i m a g e}} \\sum_ {n = 1} ^ {| S _ {\\text {i m a g e}} |} \\ell_ {n} ^ {\\text {I m a g e - C L}} \\\\ + \\lambda_ {\\text {o b j e c t}} \\sum_ {n = 1} ^ {| S _ {\\text {o b j e c t}} |} \\ell_ {n} ^ {\\text {O b j e c t - C L}} \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 691, + 482, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Multi-Task Learning", + "text_level": 1, + "bbox": [ + 89, + 821, + 282, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The existing learning algorithm has limitations when handling different tasks, particularly in multi-task scenarios such as image generation and object manipulation. Since each task focuses on different key information, existing", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models struggle to leverage and learn both the commonalities and differences between tasks, thereby limiting their generalization ability. Furthermore, when multiple complex tasks coexist, using separate models for each task often fails to exploit complementary information among tasks, resulting in low learning efficiency and suboptimal model performance. To improve adaptability, we design a multi-task learning strategy [5] that trains the embedding model on both 3D object placement and 2D layout tasks, enhancing generalization across spatial, temporal, and categorical relationships.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 outlines the training procedure for Multi-task Instance-wise Contrastive Learning. The instruction set $S$ , containing tasks for image generation and object manipulation, is processed through the following steps. First, for each instruction $I_{n}$ in $S$ , an augmented instruction $I_{n}^{\\prime}$ is generated using LLMs (Lines 3-5). These instructions, along with their original versions, are then passed through task-specific embedding heads in a pre-trained BERT model to obtain their corresponding embeddings $\\psi(I_{n})$ and $\\psi(I_{n}^{\\prime})$ (Line 6). For the image generation tasks $S_{\\text{image}}$ , the model calculates the image generation contrastive loss $\\ell_{\\text{image-CL}}$ (Lines 7-9). Similarly, for the object manipulation tasks $S_{\\text{object}}$ , the object manipulation contrastive loss $\\ell_{\\text{object-CL}}$ is computed (Lines 11-13). Finally, the overall multi-task contrastive loss $\\mathcal{L}_{\\text{Multi-Task-CL}}$ is computed by combining the individual losses from both tasks (Line 15). The model parameters, including those of the task-specific embedding heads, are updated to minimize the loss (Lines 9 and 13).", + "bbox": [ + 511, + 258, + 906, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Layout Planning", + "text_level": 1, + "bbox": [ + 513, + 544, + 679, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Layout Planning module is responsible for determining the target positions and orientations of objects based on high-level user instructions, adhering to a unified problem setting for layout generation. The process integrates User Instructions and Spatial Knowledge into a comprehensive prompt for both object rearrangement and image generation tasks, with Robot Observations added in the object rearrangement task to assist in identifying relevant objects and their attributes. For object rearrangement, during the object selection phase, the system uses the Object Selection LLM to identify relevant objects and determine their attributes as $(id, c, w, h, d)$ . In the layout generation phase, the Layout Generation LLM calculates the spatial configuration of each object, inferring the 3D coordinates $(x, y, z)$ and rotation angle $\\theta$ while using known attributes such as width, height, depth $(w, h, d)$ , and color $c$ . By assigning precise poses $(x, y, z, \\theta)$ to the objects, the Layout Planning module completes the layout generation based on the unified problem setting, allowing the process to move into the Simulation Phase, where the objects are arranged according to the determined layout. The image generation task follows a similar approach, where the system also generates object", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19418", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c70643a23548d44e3a0c9b571ef2e2ac5a7fad6de97e21095d8d9918a09c9b1b.jpg", + "table_caption": [ + "Table 1. Success rates for different models across task types and subtasks in SK-Dataset" + ], + "table_footnote": [], + "table_body": "
ModelSuccess Rate (%)
Task TypeSimple TaskComplex Task
Sub Task Type-LettersGeometric ShapesSemantic Geometric Shapes
GPT3.5SKE-Layout (ours)42.211.439.45.8
w/o CL35.210.940.64.3
LLM-GROP26.36.124.73.8
GPT4SKE-Layout (ours)75.142.772.621.7
w/o CL71.541.369.219.2
LLM-GROP69.49.418.54.5
", + "bbox": [ + 163, + 114, + 828, + 275 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1326496d3185d63c4117e599d0bf55de5c46b47a926f9699493bd36ac162fb20.jpg", + "table_caption": [ + "Table 2. Evaluation of methods on NSR-1K dataset" + ], + "table_footnote": [], + "table_body": "
MethodsNumerical ReasoningSpatial Reasoning
Layout Eval.Image Eval.Layout Eval.Image Eval.
PrecisionRecallAccuracyAcc. (GLIP)AccuracyAcc. (GLIP)
Text → Image
Stable Diffusion (v2.1)---42.44-17.81
Attend-and-Excite (SD v2.1)---45.74-26.86
Text → Layout → Image
LayoutTransformer [17]75.7061.6922.2640.556.3628.13
LayoutGPT (GPT-4)78.3686.2978.4355.6491.7360.64
w/o CL86.0897.4695.2856.8694.3561.13
w/o ML85.2398.8097.1158.6894.7061.48
SKE-Layout (ours)85.2899.0697.2459.7095.0562.19
GT layouts100.00100.00100.0053.23100.0062.54
Human99.2696.5292.5656.0791.1751.94
", + "bbox": [ + 143, + 311, + 849, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layouts $(id, c, w, d, x, y)$ based on task requirements.", + "bbox": [ + 89, + 537, + 439, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Scene Simulation", + "text_level": 1, + "bbox": [ + 89, + 570, + 261, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For object rearrangement task, we input the objects and their coordinates into the PyBullet simulation environment to create a 3D scene. This allows us to observe and evaluate the arrangement of objects in a simulated physical space, ensuring that the generated layouts are feasible and accurate in a realistic setting. For the image generation task, we use the layout-to-image generation model, GLIGEN. The model generates images based on the object layouts, ensuring that the categories, locations, and sizes of the objects in the generated image precisely match the predicted layouts described in prior works.", + "bbox": [ + 88, + 594, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 781, + 223, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present extensive empirical results of SKE-Layout and several competitive baseline methods. From the representative spatial reasoning-induced domains, including both robotic object rearrangement and image generation tasks, we observe that SKE-Layout can significantly improve task success rate compared to all baselines.", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 537, + 702, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Baselines", + "text_level": 1, + "bbox": [ + 511, + 560, + 625, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the two distinct tasks in our experiments, it is imperative to employ different baseline models for a comprehensive comparison. We select a range of competitive baselines that represent the state-of-the-art approaches in both image generation tasks and robotic object rearrangement tasks.", + "bbox": [ + 511, + 580, + 905, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Robotic Object Rearrangement Task, we utilize StructFormer [24], a Transformer-based model designed for capturing spatial relationships, and LLM-GROP [8] that integrates LLMs with planning techniques. We also include the SKE-Layout variant trained without contrastive learning (w/o CL), which uses BERT for processing, to evaluate object arrangement tasks without contrastive loss.", + "bbox": [ + 511, + 657, + 905, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Image Generation Task, we select several baselines: Stable Diffusion [33], an end-to-end diffusion-based text-to-image model; Attend-and-Excite, which uses attention mechanisms for detail-oriented image generation; LayoutTransformer [13], a text-to-query model for layout generation; and LayoutGPT, [9], which leverages LLMs for text-to-query generation. We also include our SKE-Layout model, along with its variants trained without contrastive learning (w/o CL) or multitask learning (w/o ML) to assess", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19419", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "their individual contributions.", + "bbox": [ + 89, + 90, + 287, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2. Datasets", + "text_level": 1, + "bbox": [ + 89, + 111, + 196, + 125 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the performance of our proposed method on three different datasets, each tailored to a specific task, including image generation and robotic object rearrangement, to ensure a comprehensive assessment.", + "bbox": [ + 89, + 130, + 482, + 190 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- NSR-1K Dataset: Derived from the LayoutGPT framework, it includes both template-based and human-written prompts sourced from MSCOCO, designed to evaluate image generation based on specified object counts and spatial locations.", + "- StructFormer Dataset: Designed specifically for the robotic object rearrangement task, this dataset includes scenarios where objects must be arranged based on given constraints (e.g., circle, line, tower, table setting), testing the model's ability to maintain structural coherence and execute tasks accurately. We do not use the training dataset from StructFormer, which serves as an Out-Of-Distribution (OOD) test for evaluating the model's generalization to unseen tasks.", + "- SK-Dataset: Created by us to evaluate complex object rearrangement tasks, the SK-Dataset includes advanced scenarios with multi-object planning and execution. Tasks are divided into two levels: simple and complex. The simple tasks focus on accurately positioning objects based on specified distances and positions. The complex tasks require higher-order reasoning and are further divided into letters, geometric shapes, and semantic geometric shapes, emphasizing spatial precision, symmetry, and meaningful arrangements." + ], + "bbox": [ + 89, + 191, + 482, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results and Discussions", + "text_level": 1, + "bbox": [ + 89, + 561, + 308, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SK Dataset. Table 1 compares the success rates of different models across SK-Dataset task types, including GPT-3.5 and GPT-4 using our method, w/o CL variant, and LLM-GROP. Overall, our method achieves consistently high success rates, with the degree of improvement varying by task complexity. For example, in simple tasks with GPT-4, our method reached a $75.1\\%$ success rate, outperforming w/o CL $(71.5\\%)$ and LLM-GROP $(69.4\\%)$ . In contrast, for complex tasks, our approach shows a more pronounced advantage, increasing the success rate from $4.5\\%$ (LLM-GROP) to $21.7\\%$ . Two key observations emerge from Table 1:", + "bbox": [ + 89, + 583, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- While improvements in simple tasks are moderate, our method significantly outperforms baselines in complex tasks, particularly in handling semantic geometric shapes, demonstrating its ability to enhance comprehension and reasoning in high-difficulty scenarios.", + "- GPT-4 consistently outperforms GPT-3.5, especially when combined with our method. This highlights the effectiveness of our joint training strategy in enhancing the model's understanding of complex information and its adaptability across different task difficulties." + ], + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "StructFormer Dataset. Fig. 2 compares model performance across four object rearrangement tasks: Tower, Line, Circle, and Set the Table. Our model consistently outperforms StructFormer, with improvements of $30\\%$ , $20\\%$ , $20\\%$ , and $10\\%$ , respectively. This demonstrates our method's advantage in managing tasks by effectively integrating external knowledge, enhancing both accuracy and reliability. Several key insights emerge from these results:", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The improvement is more significant in complex tasks, indicating that our approach excels in scenarios requiring higher levels of reasoning and understanding.", + "- Unlike StructFormer, our method maintains strong performance across all tasks without separate task-specific models, showcasing its versatility and adaptability to diverse object rearrangement tasks." + ], + "bbox": [ + 511, + 212, + 903, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e687dda496510f58934e689b52922b53232af894613b54a0453062be902d9ac.jpg", + "image_caption": [ + "Figure 2. Evaluation of methods on the StructFormer dataset on various tasks: Tower, Line, Circle, and Set the Table." + ], + "image_footnote": [], + "bbox": [ + 517, + 319, + 900, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NSR-1K Dataset. Table 2 presents the main results of image generation, showing that our method significantly outperforms comparative models across multiple evaluation metrics. In numerical reasoning, it exceeds LayoutGPT (GPT-4) by $10\\% - 20\\%$ in precision, recall, and accuracy. In spatial reasoning, our model achieves $95.05\\%$ layout accuracy and $61.48\\%$ GLIP accuracy, outperforming LayoutGPT by $3\\%$ and $1\\%$ , respectively. This improvement is attributed to the more comprehensive spatial knowledge of our method, which allows it to retrieve more relevant spatial information, thereby providing higher precision in both layout and image generation.", + "bbox": [ + 511, + 539, + 906, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 727, + 666, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Q1: What happens when we remove contrastive learning or multitask learning from the model? The performance on both the NSR-1K and SK datasets decreases when contrastive learning or multitask learning is disabled. This indicates that these two techniques are crucial. Specifically, contrastive learning helps the model better distinguish subtle spatial relationships, while multitask learning improves the model's ability to understand context across different tasks. Together, they contribute to more effective retrieval of spatial knowledge.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19420", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca2dceb4a6058f1ccbf66dbd78e508df28bf1a5e0faf378a1efebbd71aa863dd.jpg", + "image_caption": [ + "Figure 3. Case Study. We present a comparison of different methods in 3D object rearrangement and 2D image generation tasks. (a) arranging objects on a table with correct spatial relationships, (b) placing blocks in a specified formation, (c) generating a 2D image with objects positioned according to a textual description." + ], + "image_footnote": [], + "bbox": [ + 153, + 88, + 846, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q2: Why doesn't the StructFormer dataset show similar performance drops? The performance differences among our model variants on the StructFormer dataset are minimal. This is likely because the spatial relationships required in this dataset are relatively straightforward, making precise retrieval of complex spatial knowledge less critical. As a result, the advantages of contrastive learning and multitask learning become less pronounced.", + "bbox": [ + 88, + 470, + 482, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Case Study", + "text_level": 1, + "bbox": [ + 89, + 604, + 215, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present three representative cases to demonstrate the effectiveness of our SKE-Layout framework across both 3D object rearrangement and 2D image generation tasks. Fig. 3.(a) involves setting the table and arranging objects into a smiley face pattern. While the LLM-GROP model fails to maintain correct angles and semantic coherence, our SKE-Layout significantly reduces errors, providing accurate placement and alignment. In Fig. 3.(b), where blocks need to be arranged into a circular pattern, our method achieves perfect spatial precision, outperforming both StructFormer and LLM-GROP, which struggle with alignment. For the 2D image generation task in Fig. 3.(c), our method also excels. While Stable Diffusion generates quantity errors and LayoutGPT shows positional errors, SKE-Layout accurately captures object relationships, ensuring correct quantity and positioning and thus closely matching the given description. From Fig. 3, we draw the following observations:", + "bbox": [ + 91, + 628, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Richer Spatial Knowledge. Our SKE layout framework demonstrates a deeper understanding of spatial relationships than baseline models. By integrating diverse spatial knowledge sources, including real-world and synthetic data, our approach enhances the richness and depth of spatial reasoning. Additionally, it employs multiple forms of spatial knowledge representation, which contribute to more accurate positioning and alignment of objects in both 3D rearrangement and 2D image generation tasks.", + "- Adaptability to Multi-Task Settings. The SKE layout framework excels in handling a variety of tasks, showcasing versatility in both 3D object arrangement and 2D image generation. This adaptability highlights its robustness and effectiveness in addressing complex spatial reasoning challenges across multiple domains." + ], + "bbox": [ + 513, + 470, + 906, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 709, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present SKE.layout, a novel method designed to effectively generate layouts by leveraging large language models. Our approach injects spatial knowledge into the large language model, thereby demonstrating its versatility and robustness across different application domains. Empirically, we have demonstrated that SKE.layout performs exceptionally well in both image generation and robotic task arrangement tasks. Future research directions include exploring how to apply SKE.layout to more application scenarios, such as open-domain scenarios, where the set of objects in a problem instance is not known in advance.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 321, + 187, + 338 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020. 1", + "[2] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446):eaat8414, 2019. 1", + "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 4", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[5] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 5", + "[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[7] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4", + "[8] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 1, 2, 4, 6", + "[9] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6", + "[10] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LIm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2", + "[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 4" + ], + "bbox": [ + 93, + 347, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was partially supported by the National Natural Science Foundation of China under Grant 62322601 and 42174050, the Independent Research Project of State Key Laboratory of Mechanical Transmission for Advanced Equipment under Grant SKLMT-ZZKT2024R07, the Fundamental Research Funds for the Central Universities 2024IAIS-QN01, the Postdoctoral Fellowship Program of CPSF under Grant GZB20230912, the China Postdoctoral Science Foundation under Grant 2024M753837, the Chongqing Postdoctoral Innovation Talents Support Program under Grant CQBX202324 and the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQJQX0025 and CSTB2024NSCQMSX0955.", + "[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 4", + "[13] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2, 6", + "[14] Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints, 2023. 1", + "[15] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2", + "[16] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layout: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 2", + "[17] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. arXiv preprint arXiv:2309.10062, 2023. 2", + "[18] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. IEEE Robotics and Automation Letters, 8(7):3956-3963, 2023. 2", + "[19] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1", + "[20] Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329, 2022. 1", + "[21] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Generating graphic layouts with wireframe discriminators. arXiv preprint arXiv:1901.06767, 2019. 2", + "[22] Jiachen Li, Qiaozi Gao, Michael Johnston, Xiaofeng Gao, Xuehai He, Suhaila Shakiah, Hangjie Shi, Reza Ghanadan, and William Yang Wang. Mastering robot manipulation with multimodal prompts through pretraining and multi-task finetuning. arXiv preprint arXiv:2310.09676, 2023. 4", + "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2", + "[24] Weiyu Liu, Chris Paxton, Tucker Hermans, and Dieter Fox. Structformer: Learning spatial structure for language-guided" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19422", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semantic rearrangement of novel objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 6322-6329. IEEE, 2022. 2, 6", + "[25] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering, 35(1):857-876, 2021. 4", + "[26] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 4", + "[27] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 2", + "[28] Ruoling Peng, Kang Liu, Po Yang, Zhipeng Yuan, and Shunbao Li. Embedding-based retrieval with llm for effective agriculture information extracting from unstructured data. arXiv preprint arXiv:2308.03107, 2023. 1", + "[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 1, 2", + "[30] Shuhan Qi, Zhengying Cao, Jun Rao, Lei Wang, Jing Xiao, and Xuan Wang. What is the limitation of multimodal llms? a deeper look into multimodal llms through prompt probing. Information Processing & Management, 60(6):103510, 2023. 1", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[32] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable task planning. CoRR, 2023. 2", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6", + "[34] Sai Sabitha, Deepti Mehrotra, and Abhay Bansal. Knowledge enriched learning by converging knowledge object&learning object. Electronic Journal of e-Learning, 13 (1):pp3-13, 2015. 1", + "[35] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/rearrange_2022, 2022. 2", + "[36] Jiaqi Wang, Zihao Wu, Yiwei Li, Hanqi Jiang, Peng Shu, Enze Shi, Huawen Hu, Chong Ma, Yiheng Liu, Xuhui Wang, et al. Large language models for robotics: Opportunities, challenges, and perspectives. arXiv preprint arXiv:2401.04334, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1", + "[38] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[39] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 1, 2", + "[40] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4303-4310. IEEE, 2024. 2", + "[41] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1", + "[42] Shanshan Zhong, Zhongzhan Huang, Weushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 567-578, 2023. 2" + ], + "bbox": [ + 516, + 90, + 903, + 500 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_model.json b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2dc2f87f5d488eb4901d672496880ed5b9abeb3a --- /dev/null +++ b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "SKE-Layout: Spatial Knowledge Enhanced Layout Generation with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.18, + 0.888, + 0.217 + ], + "angle": 0, + "content": "Junsheng Wang\\(^{1*}\\) Nieqing Cao\\(^{2*}\\) Yan Ding\\(^{3}\\) Mengying Xie\\(^{1}\\) Fuqiang Gu\\(^{1}\\) Chao Chen\\(^{1\\dagger}\\) \n\\(^{1}\\)Chongqing University \\(^{2}\\)Xi'an Jiaotong-Liverpool University \\(^{3}\\)Shanghai AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.219, + 0.728, + 0.234 + ], + "angle": 0, + "content": "{csjunshengwang, xiemy, gufq, cschaochen}@cqu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.237, + 0.709, + 0.251 + ], + "angle": 0, + "content": "Nieqing.Cao@xjtlu.edu.cn, yding25@binghamton.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.577 + ], + "angle": 0, + "content": "Generating layouts from textual descriptions by large language models (LLMs) plays a crucial role in precise spatial reasoning-induced domains such as robotic object rearrangement and text-to-image generation. However, current methods face challenges in limited real-world examples, handling diverse layout descriptions and varying levels of granularity. To address these issues, a novel framework named Spatial Knowledge Enhanced Layout (SKE-Layout), is introduced. SKE-Layout integrates mixed spatial knowledge sources, leveraging both real and synthetic data to enhance spatial contexts. It utilizes diverse representations tailored to specific tasks and employs contrastive learning and multitask learning techniques for accurate spatial knowledge retrieval. This framework generates more accurate and fine-grained visual layouts for object rearrangement and text-to-image generation tasks, achieving improvements of \\(5\\% -30\\%\\) compared to existing methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.222, + 0.622 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.829 + ], + "angle": 0, + "content": "Generating layouts from textual descriptions is fundamental in various domains, particularly in robotic object rearrangement and text-to-image generation, where achieving precise object placement is critical. For example, an object rearrangement task might require arranging specific items on a table [1, 8], while an image generation task often involves creating a visually coherent scene based on descriptions of object locations [29, 39]. Both scenarios underscore the necessity of accurate spatial understanding to ensure proper placement and prevent errors such as misalignment or omissions. Research highlights the need to minimize these errors in robotic manipulation and image generation in order to reflect spatial instructions accurately [2]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.483, + 0.876 + ], + "angle": 0, + "content": "In recent years, a growing body of research has investigated the application of large language models (LLMs) for layout generation [9, 29, 39], highlighting their no" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.395 + ], + "angle": 0, + "content": "table advantages in various aspects. Primarily, LLMs exhibit strong zero-shot learning capabilities, generating coherent layouts from natural language descriptions without task-specific training, enabling broad generalization across diverse tasks [37]. Unlike traditional models, LLMs adapt seamlessly to different domains without requiring separate training for each task [19, 41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.906, + 0.685 + ], + "angle": 0, + "content": "Although LLMs exhibit numerous advantages in layout generation, they also face challenges related to hallucination when directly generating layouts from textual descriptions. Existing approaches often rely on using predefined layouts as examples to guide LLMs in generating new layouts, allowing the model to infer reasonable spatial arrangements based on these examples [9]. However, there is often a lack of sufficient and diverse examples in real-world scenarios, particularly for complex layouts, exacerbating the model's dependence on existing examples. As representations, layouts often fail to fully capture the spatial knowledge of the current scene [14, 30]. When handling fine-grained instructions related to angles, distances, and other spatial details, LLMs struggle to infer the corresponding layout solely by relying on example layouts. Furthermore, due to the varying importance of information across different tasks, directly applying existing retrieval methods may result in the inability to retrieve the most relevant knowledge for the specific task at hand [28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose SKE-Layout, the first framework capable of handling diverse layout task descriptions and inferring unknown object attributes, providing a unified solution that current frameworks cannot achieve. Specifically, SKE-Layout extracts diverse spatial knowledge from real-world or synthetic tasks and injects the most relevant spatial knowledge into LLMs, enabling them to generate ideal layouts based on textual conditions. Despite not using any image data for training, LLMs can learn spatial knowledge through contextual demonstrations and then apply this knowledge to visual reasoning for new samples [20, 34]. In other words, by providing relevant examples, LLMs can capture spatial relationships between objects based on language understanding, in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.345, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution, †Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19414" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "ferring information, such as relative positions, sizes, and distances of objects, thereby achieving accurate object rearrangement and coherent image generation [6]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.139, + 0.441, + 0.153 + ], + "angle": 0, + "content": "In summary, our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.155, + 0.483, + 0.245 + ], + "angle": 0, + "content": "- We propose a novel approach SKE.layout with LLMs for generating layouts from textural descriptions more precisely. SKE.layout integrates multiple spatial knowledge sources and combines real-world spatial datasets with model-generated data from LLMs, enhancing the spatial knowledge database without relying on existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.246, + 0.482, + 0.32 + ], + "angle": 0, + "content": "- We present a diverse framework for spatial knowledge representations. It adapts to the specific needs of different tasks by providing precise quantitative expressions (e.g., \"30-degree angle\", \"10 cm away\") and qualitative spatial relations (e.g., \"zebra to the left of a giraffe\")." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.321, + 0.482, + 0.381 + ], + "angle": 0, + "content": "- We construct an innovative spatial knowledge retrieval method, which enhances spatial context extraction by combining contrastive learning with multitask learning, optimizing accuracy for complex visual layout tasks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.155, + 0.483, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Overall, the SKE-Layout framework can effectively address the challenges in generating accurate and fine-grained visual layouts by leveraging multiple spatial knowledge sources, diverse representations, and innovative retrieval techniques. Its versatility allows it to simultaneously handle object rearrangement and image generation tasks, demonstrating its generalizability across different layout generation domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.233, + 0.535 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.305, + 0.561 + ], + "angle": 0, + "content": "2.1. Object Rearrangement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Object rearrangement is a crucial task for service robots, requiring the movement of objects from one location to another and arranging them according to specified targets. This task is exemplified by challenges such as the Habitat Rearrangement Challenge [35] and the AI2-THOR Rearrangement Challenge [38]. Various approaches have been explored to tackle this problem. One deep learning-based method, StructFormer [24], utilized transformer-based neural networks to learn fine-grained object arrangement representations from real-world datasets, enabling the model to infer semantically meaningful structures from natural language instructions. However, this method presented challenges in terms of model complexity, as each task requires independent dataset construction and training, placing high demands on data collection and computational resources. Additionally, its generalization capabilities were limited, as the model struggled to handle tasks it failed to be specifically trained on. Another approach, DALL-E-Bot [18], leveraged DALL-E 2 [31] to generate images, which were then used for object placement in table-top scenarios. Nonetheless, DALL-E's image generation lacked sufficient stability when it came to producing task-relevant images," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.393 + ], + "angle": 0, + "content": "which could hinder object arrangement accuracy. To address these limitations, methods based on LLMs have been proposed [8, 15, 17, 32, 36, 40]. LLM-GROP [8] extracted commonsense object arrangement knowledge from LLMs and applied this knowledge to perform complex manipulation and object placement tasks. SayPlan [32] introduced a novel approach to grounding LLMs in 3D scene graphs, enabling scalable task planning by translating high-level instructions into structured scene graph representations and facilitating object manipulation in complex environments. SG-Bot [40] proposed a coarse-to-fine strategy for object rearrangement, leveraging scene graphs to guide robotic imagination from scene understanding to detailed object placements. This method emphasized hierarchical planning, where initial rough layouts were refined through iterative adjustments. Despite the potential of LLMs in facilitating object rearrangement, their spatial reasoning capabilities still face limitations, particularly in accurately interpreting and generalizing complex spatial relationships across diverse scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.749, + 0.418 + ], + "angle": 0, + "content": "2.2. Image Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.905, + 0.803 + ], + "angle": 0, + "content": "Image layout generation is a critical component in accurately creating complex images. Various approaches have been proposed to generate layouts for images, each utilizing different models and input conditions to accomplish the task. Deep learning-based methods, such as LayoutGAN [21], LayoutVAE [16], and LayoutTransformer [13], typically require specific scenarios or input conditions for training, producing layouts based on fixed class labels or partially input layouts. On the other hand, LLM-based methods represent a more advanced approach [9, 10, 27, 29, 39, 42]. LayoutGPT constructed a visual layout framework from textual descriptions, extracting spatial knowledge from image datasets, such as object bounding boxes, and encoding this knowledge using CSS-style formatting. The spatial knowledge was then fed into LLM prompts, enabling the model to leverage contextual understanding to generate higher-level and lower-level layouts, which were subsequently translated into image layouts by image generation models. However, challenges remained, particularly in handling fine-grained textual descriptions related to angles and distances, as well as the generalization issues stemming from the potential lack of relevant images in current datasets for specific tasks. Building on this, GLIGEN [23] introduced bounding boxes and regional descriptions as input conditions for generating images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.605, + 0.831 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 1, we propose the SKE-Layout framework for generating generalizable layouts from textual descriptions. The framework is structured in two main parts: Knowledge Base Construction and Layout Gener" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19415" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.319, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.089, + 0.899, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.401, + 0.908, + 0.445 + ], + "angle": 0, + "content": "Figure 1. Framework of SKE-Layout. The framework comprises two main parts: Knowledge Base Construction, which builds a database from real-world images and LLM-generated knowledge; and Layout Generation Pipeline, which includes (a) Input, (b) Knowledge Retrieval, (c) Layout Planning and (d) Simulation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.485, + 0.789 + ], + "angle": 0, + "content": "ation Process. Knowledge Base Construction (Sec. 3.2) gathers task knowledge from two aspects. The first aspect involves real-world extraction, using a Visual Language Model (VLM) for scene interpretation and a Detection Transformer (DETR) [4] for object detection and bounding box generation. The second aspect focuses on task-specific knowledge generation using LLMs. Together, these aspects form a robust foundation for layout generation. After the knowledge base is built, the Layout Generation Process proceeds with the Input module, which employs the same real-world extraction methods as the Knowledge Base Construction. Then, the Knowledge Retrieval module (Sec. 3.4) enhances the system's understanding through data augmentation and multi-task learning, optimizing user instructions using relevant knowledge from the knowledge base. Subsequently, the Layout Planning module (Sec. 3.5) integrates all inputs to create a detailed and optimized layout that aligns with user instructions. Finally, the Simulation module (Sec. 3.6) implements the layout in a virtual environment, providing visual verification to ensure task accuracy and consistency with the given description." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.26, + 0.818 + ], + "angle": 0, + "content": "3.1. Problem Settings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this study, we establish a unified problem setting for layout generation, encompassing both object rearrangement and image generation tasks. We define a reference coordinate system and origin tailored to each task. For object rearrangement, the reference is a 3D coordinate system" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.53 + ], + "angle": 0, + "content": "\\((x,y,z)\\in \\mathbb{R}^3\\) , with the origin at the top-left corner of the rearrangement space, while for image generation, it is a 2D coordinate system \\((x,y)\\in \\mathbb{R}^2\\) , with the origin at the topleft corner of the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.906, + 0.562 + ], + "angle": 0, + "content": "The spatial information of an object, referred to as its layout, is represented by a tuple" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.572, + 0.807, + 0.589 + ], + "angle": 0, + "content": "\\[\no = (\\mathrm {i d}, c, x, y, z, w, h, d, \\theta),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.795 + ], + "angle": 0, + "content": "where id denotes the object name, \\(c\\) is the color, \\((x,y,z)\\) represents the object's center coordinates, and \\(w\\), \\(h\\), \\(d\\), \\(\\theta\\) denote the bounding box's width, height, depth, and the rotation angle in the \\(x-y\\) plane, respectively. In image generation tasks, only the 2D coordinates \\((x,y)\\) and bounding box dimensions \\((w,h)\\) are relevant, while 3D attributes like \\(z\\), \\(d\\), and \\(\\theta\\) are often missing or irrelevant. In contrast, object rearrangement tasks require inferring the 3D coordinates \\((x,y,z)\\) and the rotation angle \\(\\theta\\), while color and dimensions \\(w\\), \\(h\\), \\(d\\) are known. Given a textual description \\(\\mathcal{I}\\), the objective of layout generation is to produce a set of tuples \\(\\mathcal{O}\\), where each tuple \\(o\\in \\mathcal{O}\\) captures the spatial attributes of the objects as described." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Spatial knowledge is defined as information related to the positioning, orientation, and arrangement of objects within a given environment, expressed as pairs of either \\(\\{\\text{Instruction - Object}\\}\\), where \\(\\mathcal{O} = \\{o_1, o_2, \\ldots, o_i\\}\\), or \\(\\{\\text{Instruction - Mathematical constraint}\\}\\), where the latter represents numerical constraints related to object positioning. In this context, 'knowledge' specifically refers to spa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19416" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.343, + 0.108 + ], + "angle": 0, + "content": "tial knowledge for the sake of brevity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.123, + 0.358, + 0.14 + ], + "angle": 0, + "content": "3.2. Knowledge Base Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.148, + 0.483, + 0.254 + ], + "angle": 0, + "content": "In Retrieval-Augmented Generation (RAG) [12], constructing a comprehensive knowledge base is essential. This repository of exemplars and contextual information enhances the model's generative capabilities, allowing the system to retrieve relevant examples and produce precise, contextually appropriate layouts, thereby improving the efficacy and reliability of RAG." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.268, + 0.471, + 0.284 + ], + "angle": 0, + "content": "3.2.1. Extracting Knowledge from Real-World Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.501 + ], + "angle": 0, + "content": "The primary aspect for knowledge base construction is to extract information from real-world images. We use VLMs and DETR for object detection and bounding box acquisition, capturing object location and dimensions to build a detailed representation of real-world scenes. By aggregating layouts from diverse images, we create a rich knowledge base that accurately reflects spatial arrangements and object interactions. The extracted information provides tuples \\( o \\in \\mathcal{O}_{2D} \\) and \\( o \\in \\mathcal{O}_{3D} \\), with each tuple representing object category, location, size, and orientation. Additionally, VLMs capture spatial relationships between objects, stored as instruction information \\( I \\) for each image in the knowledge base, aiding in retrieving relevant images and spatial context for layout generation based on new instructions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.381, + 0.53 + ], + "angle": 0, + "content": "3.2.2. Generating Knowledge with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.779 + ], + "angle": 0, + "content": "The second aspect uses LLMs to generate diverse tasks \\( I \\) and layouts \\( O \\) from existing instructions collected in other research[8, 22, 26]. By interpreting and synthesizing text, LLMs create tasks that span a broad spectrum of real-world and hypothetical scenarios, introducing variability crucial for building a versatile knowledge base. This process begins by inputting varied instructions into the LLMs, which produces corresponding tasks not as duplicates but as novel scenarios. For each task, the LLMs generate layouts specifying object categories, spatial arrangements, and interobject relationships. This strategy enriches the knowledge base with synthetic scenarios that improve the system's adaptability to complex and rare instructions. Mathematical constraints \\( M = \\{m_{n}\\} \\) are also integrated to meet spatial requirements like distances and angles, ensuring layout accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To address potential inaccuracies, the synthetic data undergoes rigorous validation. First, VLMs verify if the generated images satisfy the original instructions (e.g., object arrange requirement, image composition) by evaluating object categories, positions, and spatial relationships. If discrepancies are found, human evaluators review the images for realism and relevance. This validation ensures the quality and reliability of data added to the RAG system." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.596, + 0.107 + ], + "angle": 0, + "content": "3.3. Input" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.267 + ], + "angle": 0, + "content": "This section describes the configuration of inputs for different task settings within the SKE-Layout framework. For the image generation task, the input consists solely of user instructions, which outline the requirements for the generated image layout. The object rearrangement task involves a more complex input configuration. In addition to user instructions, it incorporates robot observations, which provide visual data of the environment and capture the spatial layout and positioning of objects within the scene through VLMs and DETR methods, as described in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.282, + 0.712, + 0.298 + ], + "angle": 0, + "content": "3.4. Knowledge Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.533 + ], + "angle": 0, + "content": "After constructing our knowledge base, we need to retrieve knowledge in the database based on the current user instruction. Given a test instruction \\( I \\) and the knowledge base \\( D = \\{(I_k, o_k) \\mid k = 1, 2, \\ldots\\} \\), we define a function \\( f(I_k, I_j) \\in \\mathbb{R} \\) that measures the distances between two instructions. Initially, we experiment with directly using existing text embedding models [3, 7] to convert instructions into vectors and calculate their similarity. However, this approach reveals a significant issue: different tasks emphasize different aspects of the information. For example, in the instruction \"place the banana \\( 30^\\circ \\) above the apple\", the specific fruits are less relevant than their spatial relationship. Standard embedding models can skew the retrieval results by overemphasizing the objects rather than their spatial arrangement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.813, + 0.562 + ], + "angle": 0, + "content": "3.4.1. Instance-Wise Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.734 + ], + "angle": 0, + "content": "To address this issue, we train a custom text embedding model through contrastive learning, building on the architecture of BERT [25] to meet our retrieval needs. This model is tailored for specific tasks, emphasizing the importance of spatial relationships over individual object identities, especially in object placement scenarios. To further enhance robustness, we apply data augmentation techniques using LLMs. For instance, in the instruction \"place the banana \\(30^{\\circ}\\) above the apple,\" we generalize it to \"place object A \\(30^{\\circ}\\) above object B,\" allowing the model to prioritize spatial relationships while ignoring specific object identities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.857 + ], + "angle": 0, + "content": "Inspired by SimCSE [11], let \\( S = [I_n]_{n=1}^N \\) be the set of original instructions, and \\( S' = [I_n, I_n']_{n=1}^N \\) be the set after data augmentation, where \\( I_n' \\) is the augmented instruction derived from \\( I_n \\). We use a pre-trained BERT model \\( \\psi \\) to obtain the embeddings of the instructions. For each instruction \\( I_n \\), we define its unique positive pair as \\( (I_n, I_n') \\), and all other pairs are treated as negative samples. The similarity between the two embeddings is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.87, + 0.907, + 0.907 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {m}\\right)\\right) = \\frac {\\psi \\left(I _ {n}\\right) ^ {T} \\psi \\left(I _ {m}\\right)}{\\| \\psi \\left(I _ {n}\\right) \\| _ {2} \\| \\psi \\left(I _ {m}\\right) \\| _ {2}} \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19417" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.375, + 0.106 + ], + "angle": 0, + "content": "The contrastive loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.113, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\ell_ {n} ^ {N} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right) , \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right)\\right)}{\\sum_ {m = 1 , m \\neq n} ^ {2 N} \\exp \\left(\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right) , \\psi \\left(I _ {m}\\right)\\right)\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.159, + 0.483, + 0.186 + ], + "angle": 0, + "content": "The overall contrastive loss is the average of all individual losses:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.186, + 0.483, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {I n s t a n c e - C L}} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {2 N} \\ell_ {n} ^ {N} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.483, + 0.278 + ], + "angle": 0, + "content": "Algorithm 1 Training Procedure for Multi-task Instance-wise Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.283, + 0.483, + 0.314 + ], + "angle": 0, + "content": "1: Input: Instruction set \\( S = \\{S_{\\mathrm{image}}, S_{\\mathrm{object}}\\} \\) containing tasks for image generation and object manipulation" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.315, + 0.281, + 0.328 + ], + "angle": 0, + "content": "2: Output: Trained model" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.33, + 0.251, + 0.343 + ], + "angle": 0, + "content": "3: for each \\( I_{n} \\in S \\) do" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.344, + 0.453, + 0.36 + ], + "angle": 0, + "content": "4: Generate augmented instruction \\( I_{n}^{\\prime} \\) using LLMs" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.361, + 0.175, + 0.372 + ], + "angle": 0, + "content": "5: end for" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.375, + 0.483, + 0.418 + ], + "angle": 0, + "content": "6: Obtain embeddings \\(\\psi(I_n)\\) and \\(\\psi(I_n')\\) for all instructions using task-specific embedding heads in the pretrained BERT model" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.42, + 0.28, + 0.435 + ], + "angle": 0, + "content": "7: for each \\( I_{n} \\in S_{\\mathrm{image}} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.436, + 0.449, + 0.451 + ], + "angle": 0, + "content": "8: Calculate the image generation contrastive loss:" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.283, + 0.483, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.46, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\ell_ {n} ^ {\\text {I m a g e - C L}} = \\operatorname {L o s s} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.505, + 0.452, + 0.521 + ], + "angle": 0, + "content": "9: Update model parameters to minimize \\(\\mathcal{L}_{\\mathrm{Image - CL}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.522, + 0.175, + 0.534 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.536, + 0.279, + 0.551 + ], + "angle": 0, + "content": "11: for each \\( I_{n} \\in S_{\\mathrm{object}} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.552, + 0.482, + 0.566 + ], + "angle": 0, + "content": "12: Calculate the object manipulation contrastive loss:" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.505, + 0.482, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.591, + 0.483, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\ell_ {n} ^ {\\text {O b j e c t - C L}} = \\operatorname {L o s s} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.636, + 0.454, + 0.652 + ], + "angle": 0, + "content": "13: Update model parameters to minimize \\(\\mathcal{L}_{\\mathrm{Object - CL}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.653, + 0.175, + 0.665 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.667, + 0.436, + 0.682 + ], + "angle": 0, + "content": "15: Compute the overall multi-task contrastive loss:" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.636, + 0.454, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.693, + 0.483, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {M u l t i - T a s k - C L}} = \\lambda_ {\\text {i m a g e}} \\sum_ {n = 1} ^ {| S _ {\\text {i m a g e}} |} \\ell_ {n} ^ {\\text {I m a g e - C L}} \\\\ + \\lambda_ {\\text {o b j e c t}} \\sum_ {n = 1} ^ {| S _ {\\text {o b j e c t}} |} \\ell_ {n} ^ {\\text {O b j e c t - C L}} \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.284, + 0.837 + ], + "angle": 0, + "content": "3.4.2. Multi-Task Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The existing learning algorithm has limitations when handling different tasks, particularly in multi-task scenarios such as image generation and object manipulation. Since each task focuses on different key information, existing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "models struggle to leverage and learn both the commonalities and differences between tasks, thereby limiting their generalization ability. Furthermore, when multiple complex tasks coexist, using separate models for each task often fails to exploit complementary information among tasks, resulting in low learning efficiency and suboptimal model performance. To improve adaptability, we design a multi-task learning strategy [5] that trains the embedding model on both 3D object placement and 2D layout tasks, enhancing generalization across spatial, temporal, and categorical relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Algorithm 1 outlines the training procedure for Multi-task Instance-wise Contrastive Learning. The instruction set \\( S \\), containing tasks for image generation and object manipulation, is processed through the following steps. First, for each instruction \\( I_{n} \\) in \\( S \\), an augmented instruction \\( I_{n}^{\\prime} \\) is generated using LLMs (Lines 3-5). These instructions, along with their original versions, are then passed through task-specific embedding heads in a pre-trained BERT model to obtain their corresponding embeddings \\( \\psi(I_{n}) \\) and \\( \\psi(I_{n}^{\\prime}) \\) (Line 6). For the image generation tasks \\( S_{\\text{image}} \\), the model calculates the image generation contrastive loss \\( \\ell_{\\text{image-CL}} \\) (Lines 7-9). Similarly, for the object manipulation tasks \\( S_{\\text{object}} \\), the object manipulation contrastive loss \\( \\ell_{\\text{object-CL}} \\) is computed (Lines 11-13). Finally, the overall multi-task contrastive loss \\( \\mathcal{L}_{\\text{Multi-Task-CL}} \\) is computed by combining the individual losses from both tasks (Line 15). The model parameters, including those of the task-specific embedding heads, are updated to minimize the loss (Lines 9 and 13)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.545, + 0.68, + 0.561 + ], + "angle": 0, + "content": "3.5. Layout Planning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The Layout Planning module is responsible for determining the target positions and orientations of objects based on high-level user instructions, adhering to a unified problem setting for layout generation. The process integrates User Instructions and Spatial Knowledge into a comprehensive prompt for both object rearrangement and image generation tasks, with Robot Observations added in the object rearrangement task to assist in identifying relevant objects and their attributes. For object rearrangement, during the object selection phase, the system uses the Object Selection LLM to identify relevant objects and determine their attributes as \\((id, c, w, h, d)\\). In the layout generation phase, the Layout Generation LLM calculates the spatial configuration of each object, inferring the 3D coordinates \\((x, y, z)\\) and rotation angle \\(\\theta\\) while using known attributes such as width, height, depth \\((w, h, d)\\), and color \\(c\\). By assigning precise poses \\((x, y, z, \\theta)\\) to the objects, the Layout Planning module completes the layout generation based on the unified problem setting, allowing the process to move into the Simulation Phase, where the objects are arranged according to the determined layout. The image generation task follows a similar approach, where the system also generates object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19418" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.089, + 0.761, + 0.104 + ], + "angle": 0, + "content": "Table 1. Success rates for different models across task types and subtasks in SK-Dataset" + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.115, + 0.83, + 0.276 + ], + "angle": 0, + "content": "
ModelSuccess Rate (%)
Task TypeSimple TaskComplex Task
Sub Task Type-LettersGeometric ShapesSemantic Geometric Shapes
GPT3.5SKE-Layout (ours)42.211.439.45.8
w/o CL35.210.940.64.3
LLM-GROP26.36.124.73.8
GPT4SKE-Layout (ours)75.142.772.621.7
w/o CL71.541.369.219.2
LLM-GROP69.49.418.54.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.345, + 0.288, + 0.652, + 0.302 + ], + "angle": 0, + "content": "Table 2. Evaluation of methods on NSR-1K dataset" + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.313, + 0.851, + 0.515 + ], + "angle": 0, + "content": "
MethodsNumerical ReasoningSpatial Reasoning
Layout Eval.Image Eval.Layout Eval.Image Eval.
PrecisionRecallAccuracyAcc. (GLIP)AccuracyAcc. (GLIP)
Text → Image
Stable Diffusion (v2.1)---42.44-17.81
Attend-and-Excite (SD v2.1)---45.74-26.86
Text → Layout → Image
LayoutTransformer [17]75.7061.6922.2640.556.3628.13
LayoutGPT (GPT-4)78.3686.2978.4355.6491.7360.64
w/o CL86.0897.4695.2856.8694.3561.13
w/o ML85.2398.8097.1158.6894.7061.48
SKE-Layout (ours)85.2899.0697.2459.7095.0562.19
GT layouts100.00100.00100.0053.23100.0062.54
Human99.2696.5292.5656.0791.1751.94
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.441, + 0.555 + ], + "angle": 0, + "content": "Layouts \\((id, c, w, d, x, y)\\) based on task requirements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.262, + 0.587 + ], + "angle": 0, + "content": "3.6. Scene Simulation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.595, + 0.483, + 0.761 + ], + "angle": 0, + "content": "For object rearrangement task, we input the objects and their coordinates into the PyBullet simulation environment to create a 3D scene. This allows us to observe and evaluate the arrangement of objects in a simulated physical space, ensuring that the generated layouts are feasible and accurate in a realistic setting. For the image generation task, we use the layout-to-image generation model, GLIGEN. The model generates images based on the object layouts, ensuring that the categories, locations, and sizes of the objects in the generated image precisely match the predicted layouts described in prior works." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.782, + 0.224, + 0.8 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we present extensive empirical results of SKE-Layout and several competitive baseline methods. From the representative spatial reasoning-induced domains, including both robotic object rearrangement and image generation tasks, we observe that SKE-Layout can significantly improve task success rate compared to all baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.704, + 0.556 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.627, + 0.575 + ], + "angle": 0, + "content": "4.1.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Given the two distinct tasks in our experiments, it is imperative to employ different baseline models for a comprehensive comparison. We select a range of competitive baselines that represent the state-of-the-art approaches in both image generation tasks and robotic object rearrangement tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.763 + ], + "angle": 0, + "content": "For the Robotic Object Rearrangement Task, we utilize StructFormer [24], a Transformer-based model designed for capturing spatial relationships, and LLM-GROP [8] that integrates LLMs with planning techniques. We also include the SKE-Layout variant trained without contrastive learning (w/o CL), which uses BERT for processing, to evaluate object arrangement tasks without contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the Image Generation Task, we select several baselines: Stable Diffusion [33], an end-to-end diffusion-based text-to-image model; Attend-and-Excite, which uses attention mechanisms for detail-oriented image generation; LayoutTransformer [13], a text-to-query model for layout generation; and LayoutGPT, [9], which leverages LLMs for text-to-query generation. We also include our SKE-Layout model, along with its variants trained without contrastive learning (w/o CL) or multitask learning (w/o ML) to assess" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19419" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.289, + 0.105 + ], + "angle": 0, + "content": "their individual contributions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.112, + 0.197, + 0.126 + ], + "angle": 0, + "content": "4.1.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.131, + 0.483, + 0.191 + ], + "angle": 0, + "content": "We evaluate the performance of our proposed method on three different datasets, each tailored to a specific task, including image generation and robotic object rearrangement, to ensure a comprehensive assessment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.193, + 0.483, + 0.268 + ], + "angle": 0, + "content": "- NSR-1K Dataset: Derived from the LayoutGPT framework, it includes both template-based and human-written prompts sourced from MSCOCO, designed to evaluate image generation based on specified object counts and spatial locations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.268, + 0.483, + 0.403 + ], + "angle": 0, + "content": "- StructFormer Dataset: Designed specifically for the robotic object rearrangement task, this dataset includes scenarios where objects must be arranged based on given constraints (e.g., circle, line, tower, table setting), testing the model's ability to maintain structural coherence and execute tasks accurately. We do not use the training dataset from StructFormer, which serves as an Out-Of-Distribution (OOD) test for evaluating the model's generalization to unseen tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.404, + 0.483, + 0.556 + ], + "angle": 0, + "content": "- SK-Dataset: Created by us to evaluate complex object rearrangement tasks, the SK-Dataset includes advanced scenarios with multi-object planning and execution. Tasks are divided into two levels: simple and complex. The simple tasks focus on accurately positioning objects based on specified distances and positions. The complex tasks require higher-order reasoning and are further divided into letters, geometric shapes, and semantic geometric shapes, emphasizing spatial precision, symmetry, and meaningful arrangements." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.193, + 0.483, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.31, + 0.577 + ], + "angle": 0, + "content": "4.2. Results and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.75 + ], + "angle": 0, + "content": "SK Dataset. Table 1 compares the success rates of different models across SK-Dataset task types, including GPT-3.5 and GPT-4 using our method, w/o CL variant, and LLM-GROP. Overall, our method achieves consistently high success rates, with the degree of improvement varying by task complexity. For example, in simple tasks with GPT-4, our method reached a \\(75.1\\%\\) success rate, outperforming w/o CL \\((71.5\\%)\\) and LLM-GROP \\((69.4\\%)\\). In contrast, for complex tasks, our approach shows a more pronounced advantage, increasing the success rate from \\(4.5\\%\\) (LLM-GROP) to \\(21.7\\%\\). Two key observations emerge from Table 1:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.751, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- While improvements in simple tasks are moderate, our method significantly outperforms baselines in complex tasks, particularly in handling semantic geometric shapes, demonstrating its ability to enhance comprehension and reasoning in high-difficulty scenarios." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- GPT-4 consistently outperforms GPT-3.5, especially when combined with our method. This highlights the effectiveness of our joint training strategy in enhancing the model's understanding of complex information and its adaptability across different task difficulties." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.751, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "StructFormer Dataset. Fig. 2 compares model performance across four object rearrangement tasks: Tower, Line, Circle, and Set the Table. Our model consistently outperforms StructFormer, with improvements of \\(30\\%\\), \\(20\\%\\), \\(20\\%\\), and \\(10\\%\\), respectively. This demonstrates our method's advantage in managing tasks by effectively integrating external knowledge, enhancing both accuracy and reliability. Several key insights emerge from these results:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- The improvement is more significant in complex tasks, indicating that our approach excels in scenarios requiring higher levels of reasoning and understanding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.318 + ], + "angle": 0, + "content": "- Unlike StructFormer, our method maintains strong performance across all tasks without separate task-specific models, showcasing its versatility and adaptability to diverse object rearrangement tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.32, + 0.901, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.497, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 2. Evaluation of methods on the StructFormer dataset on various tasks: Tower, Line, Circle, and Set the Table." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.722 + ], + "angle": 0, + "content": "NSR-1K Dataset. Table 2 presents the main results of image generation, showing that our method significantly outperforms comparative models across multiple evaluation metrics. In numerical reasoning, it exceeds LayoutGPT (GPT-4) by \\(10\\% - 20\\%\\) in precision, recall, and accuracy. In spatial reasoning, our model achieves \\(95.05\\%\\) layout accuracy and \\(61.48\\%\\) GLIP accuracy, outperforming LayoutGPT by \\(3\\%\\) and \\(1\\%\\), respectively. This improvement is attributed to the more comprehensive spatial knowledge of our method, which allows it to retrieve more relevant spatial information, thereby providing higher precision in both layout and image generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.667, + 0.744 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Q1: What happens when we remove contrastive learning or multitask learning from the model? The performance on both the NSR-1K and SK datasets decreases when contrastive learning or multitask learning is disabled. This indicates that these two techniques are crucial. Specifically, contrastive learning helps the model better distinguish subtle spatial relationships, while multitask learning improves the model's ability to understand context across different tasks. Together, they contribute to more effective retrieval of spatial knowledge." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.089, + 0.847, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.415, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 3. Case Study. We present a comparison of different methods in 3D object rearrangement and 2D image generation tasks. (a) arranging objects on a table with correct spatial relationships, (b) placing blocks in a specified formation, (c) generating a 2D image with objects positioned according to a textual description." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Q2: Why doesn't the StructFormer dataset show similar performance drops? The performance differences among our model variants on the StructFormer dataset are minimal. This is likely because the spatial relationships required in this dataset are relatively straightforward, making precise retrieval of complex spatial knowledge less critical. As a result, the advantages of contrastive learning and multitask learning become less pronounced." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.606, + 0.216, + 0.622 + ], + "angle": 0, + "content": "4.4. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.629, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We present three representative cases to demonstrate the effectiveness of our SKE-Layout framework across both 3D object rearrangement and 2D image generation tasks. Fig. 3.(a) involves setting the table and arranging objects into a smiley face pattern. While the LLM-GROP model fails to maintain correct angles and semantic coherence, our SKE-Layout significantly reduces errors, providing accurate placement and alignment. In Fig. 3.(b), where blocks need to be arranged into a circular pattern, our method achieves perfect spatial precision, outperforming both StructFormer and LLM-GROP, which struggle with alignment. For the 2D image generation task in Fig. 3.(c), our method also excels. While Stable Diffusion generates quantity errors and LayoutGPT shows positional errors, SKE-Layout accurately captures object relationships, ensuring correct quantity and positioning and thus closely matching the given description. From Fig. 3, we draw the following observations:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.471, + 0.907, + 0.607 + ], + "angle": 0, + "content": "- Richer Spatial Knowledge. Our SKE layout framework demonstrates a deeper understanding of spatial relationships than baseline models. By integrating diverse spatial knowledge sources, including real-world and synthetic data, our approach enhances the richness and depth of spatial reasoning. Additionally, it employs multiple forms of spatial knowledge representation, which contribute to more accurate positioning and alignment of objects in both 3D rearrangement and 2D image generation tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.608, + 0.907, + 0.699 + ], + "angle": 0, + "content": "- Adaptability to Multi-Task Settings. The SKE layout framework excels in handling a variety of tasks, showcasing versatility in both 3D object arrangement and 2D image generation. This adaptability highlights its robustness and effectiveness in addressing complex spatial reasoning challenges across multiple domains." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.471, + 0.907, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present SKE.layout, a novel method designed to effectively generate layouts by leveraging large language models. Our approach injects spatial knowledge into the large language model, thereby demonstrating its versatility and robustness across different application domains. Empirically, we have demonstrated that SKE.layout performs exceptionally well in both image generation and robotic task arrangement tasks. Future research directions include exploring how to apply SKE.layout to more application scenarios, such as open-domain scenarios, where the set of objects in a problem instance is not known in advance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19421" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.309 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China under Grant 62322601 and 42174050, the Independent Research Project of State Key Laboratory of Mechanical Transmission for Advanced Equipment under Grant SKLMT-ZZKT2024R07, the Fundamental Research Funds for the Central Universities 2024IAIS-QN01, the Postdoctoral Fellowship Program of CPSF under Grant GZB20230912, the China Postdoctoral Science Foundation under Grant 2024M753837, the Chongqing Postdoctoral Innovation Talents Support Program under Grant CQBX202324 and the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQJQX0025 and CSTB2024NSCQMSX0955." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.323, + 0.188, + 0.339 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.348, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[1] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.419, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[2] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446):eaat8414, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[5] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[7] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[8] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[9] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[10] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LIm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.348, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[14] Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[15] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[16] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layout: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[17] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. arXiv preprint arXiv:2309.10062, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[18] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. IEEE Robotics and Automation Letters, 8(7):3956-3963, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[19] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[21] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Generating graphic layouts with wireframe discriminators. arXiv preprint arXiv:1901.06767, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] Jiachen Li, Qiaozi Gao, Michael Johnston, Xiaofeng Gao, Xuehai He, Suhaila Shakiah, Hangjie Shi, Reza Ghanadan, and William Yang Wang. Mastering robot manipulation with multimodal prompts through pretraining and multi-task finetuning. arXiv preprint arXiv:2310.09676, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Weiyu Liu, Chris Paxton, Tucker Hermans, and Dieter Fox. Structformer: Learning spatial structure for language-guided" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19422" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "semantic rearrangement of novel objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 6322-6329. IEEE, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[25] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering, 35(1):857-876, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[26] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[27] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[28] Ruoling Peng, Kang Liu, Po Yang, Zhipeng Yuan, and Shunbao Li. Embedding-based retrieval with llm for effective agriculture information extracting from unstructured data. arXiv preprint arXiv:2308.03107, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[30] Shuhan Qi, Zhengying Cao, Jun Rao, Lei Wang, Jing Xiao, and Xuan Wang. What is the limitation of multimodal llms? a deeper look into multimodal llms through prompt probing. Information Processing & Management, 60(6):103510, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[32] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable task planning. CoRR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[34] Sai Sabitha, Deepti Mehrotra, and Abhay Bansal. Knowledge enriched learning by converging knowledge object&learning object. Electronic Journal of e-Learning, 13 (1):pp3-13, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[35] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/rearrange_2022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Jiaqi Wang, Zihao Wu, Yiwei Li, Hanqi Jiang, Peng Shu, Enze Shi, Huawen Hu, Chong Ma, Yiheng Liu, Xuhui Wang, et al. Large language models for robotics: Opportunities, challenges, and perspectives. arXiv preprint arXiv:2401.04334, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[38] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[39] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[40] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4303-4310. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[41] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[42] Shanshan Zhong, Zhongzhan Huang, Weushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 567-578, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19423" + } + ] +] \ No newline at end of file diff --git a/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_origin.pdf b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c35bbca09b559f807e1be71f7f25a4f6070d1bb --- /dev/null +++ b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/acd7c271-45e8-43e6-a57e-fd2570599f21_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95832d9055af2e6b7e846a792dda4c92bdbad2610250263751d247527c4b15d7 +size 5795186 diff --git a/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/full.md b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aa9066c60f96b048ee58226583b880425447ba7b --- /dev/null +++ b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/full.md @@ -0,0 +1,289 @@ +# SKE-Layout: Spatial Knowledge Enhanced Layout Generation with LLMs + +Junsheng Wang $^{1*}$ Nieqing Cao $^{2*}$ Yan Ding $^{3}$ Mengying Xie $^{1}$ Fuqiang Gu $^{1}$ Chao Chen $^{1\dagger}$ $^{1}$ Chongqing University $^{2}$ Xi'an Jiaotong-Liverpool University $^{3}$ Shanghai AI Lab + +{csjunshengwang, xiemy, gufq, cschaochen}@cqu.edu.cn + +Nieqing.Cao@xjtlu.edu.cn, yding25@binghamton.edu + +# Abstract + +Generating layouts from textual descriptions by large language models (LLMs) plays a crucial role in precise spatial reasoning-induced domains such as robotic object rearrangement and text-to-image generation. However, current methods face challenges in limited real-world examples, handling diverse layout descriptions and varying levels of granularity. To address these issues, a novel framework named Spatial Knowledge Enhanced Layout (SKE-Layout), is introduced. SKE-Layout integrates mixed spatial knowledge sources, leveraging both real and synthetic data to enhance spatial contexts. It utilizes diverse representations tailored to specific tasks and employs contrastive learning and multitask learning techniques for accurate spatial knowledge retrieval. This framework generates more accurate and fine-grained visual layouts for object rearrangement and text-to-image generation tasks, achieving improvements of $5\% -30\%$ compared to existing methods. + +# 1. Introduction + +Generating layouts from textual descriptions is fundamental in various domains, particularly in robotic object rearrangement and text-to-image generation, where achieving precise object placement is critical. For example, an object rearrangement task might require arranging specific items on a table [1, 8], while an image generation task often involves creating a visually coherent scene based on descriptions of object locations [29, 39]. Both scenarios underscore the necessity of accurate spatial understanding to ensure proper placement and prevent errors such as misalignment or omissions. Research highlights the need to minimize these errors in robotic manipulation and image generation in order to reflect spatial instructions accurately [2]. + +In recent years, a growing body of research has investigated the application of large language models (LLMs) for layout generation [9, 29, 39], highlighting their no + +table advantages in various aspects. Primarily, LLMs exhibit strong zero-shot learning capabilities, generating coherent layouts from natural language descriptions without task-specific training, enabling broad generalization across diverse tasks [37]. Unlike traditional models, LLMs adapt seamlessly to different domains without requiring separate training for each task [19, 41]. + +Although LLMs exhibit numerous advantages in layout generation, they also face challenges related to hallucination when directly generating layouts from textual descriptions. Existing approaches often rely on using predefined layouts as examples to guide LLMs in generating new layouts, allowing the model to infer reasonable spatial arrangements based on these examples [9]. However, there is often a lack of sufficient and diverse examples in real-world scenarios, particularly for complex layouts, exacerbating the model's dependence on existing examples. As representations, layouts often fail to fully capture the spatial knowledge of the current scene [14, 30]. When handling fine-grained instructions related to angles, distances, and other spatial details, LLMs struggle to infer the corresponding layout solely by relying on example layouts. Furthermore, due to the varying importance of information across different tasks, directly applying existing retrieval methods may result in the inability to retrieve the most relevant knowledge for the specific task at hand [28]. + +To address the aforementioned challenges, we propose SKE-Layout, the first framework capable of handling diverse layout task descriptions and inferring unknown object attributes, providing a unified solution that current frameworks cannot achieve. Specifically, SKE-Layout extracts diverse spatial knowledge from real-world or synthetic tasks and injects the most relevant spatial knowledge into LLMs, enabling them to generate ideal layouts based on textual conditions. Despite not using any image data for training, LLMs can learn spatial knowledge through contextual demonstrations and then apply this knowledge to visual reasoning for new samples [20, 34]. In other words, by providing relevant examples, LLMs can capture spatial relationships between objects based on language understanding, in + +ferring information, such as relative positions, sizes, and distances of objects, thereby achieving accurate object rearrangement and coherent image generation [6]. + +In summary, our main contributions are threefold: + +- We propose a novel approach SKE.layout with LLMs for generating layouts from textural descriptions more precisely. SKE.layout integrates multiple spatial knowledge sources and combines real-world spatial datasets with model-generated data from LLMs, enhancing the spatial knowledge database without relying on existing datasets. +- We present a diverse framework for spatial knowledge representations. It adapts to the specific needs of different tasks by providing precise quantitative expressions (e.g., "30-degree angle", "10 cm away") and qualitative spatial relations (e.g., "zebra to the left of a giraffe"). +- We construct an innovative spatial knowledge retrieval method, which enhances spatial context extraction by combining contrastive learning with multitask learning, optimizing accuracy for complex visual layout tasks. + +Overall, the SKE-Layout framework can effectively address the challenges in generating accurate and fine-grained visual layouts by leveraging multiple spatial knowledge sources, diverse representations, and innovative retrieval techniques. Its versatility allows it to simultaneously handle object rearrangement and image generation tasks, demonstrating its generalizability across different layout generation domains. + +# 2. Related Work + +# 2.1. Object Rearrangement + +Object rearrangement is a crucial task for service robots, requiring the movement of objects from one location to another and arranging them according to specified targets. This task is exemplified by challenges such as the Habitat Rearrangement Challenge [35] and the AI2-THOR Rearrangement Challenge [38]. Various approaches have been explored to tackle this problem. One deep learning-based method, StructFormer [24], utilized transformer-based neural networks to learn fine-grained object arrangement representations from real-world datasets, enabling the model to infer semantically meaningful structures from natural language instructions. However, this method presented challenges in terms of model complexity, as each task requires independent dataset construction and training, placing high demands on data collection and computational resources. Additionally, its generalization capabilities were limited, as the model struggled to handle tasks it failed to be specifically trained on. Another approach, DALL-E-Bot [18], leveraged DALL-E 2 [31] to generate images, which were then used for object placement in table-top scenarios. Nonetheless, DALL-E's image generation lacked sufficient stability when it came to producing task-relevant images, + +which could hinder object arrangement accuracy. To address these limitations, methods based on LLMs have been proposed [8, 15, 17, 32, 36, 40]. LLM-GROP [8] extracted commonsense object arrangement knowledge from LLMs and applied this knowledge to perform complex manipulation and object placement tasks. SayPlan [32] introduced a novel approach to grounding LLMs in 3D scene graphs, enabling scalable task planning by translating high-level instructions into structured scene graph representations and facilitating object manipulation in complex environments. SG-Bot [40] proposed a coarse-to-fine strategy for object rearrangement, leveraging scene graphs to guide robotic imagination from scene understanding to detailed object placements. This method emphasized hierarchical planning, where initial rough layouts were refined through iterative adjustments. Despite the potential of LLMs in facilitating object rearrangement, their spatial reasoning capabilities still face limitations, particularly in accurately interpreting and generalizing complex spatial relationships across diverse scenarios. + +# 2.2. Image Layout Generation + +Image layout generation is a critical component in accurately creating complex images. Various approaches have been proposed to generate layouts for images, each utilizing different models and input conditions to accomplish the task. Deep learning-based methods, such as LayoutGAN [21], LayoutVAE [16], and LayoutTransformer [13], typically require specific scenarios or input conditions for training, producing layouts based on fixed class labels or partially input layouts. On the other hand, LLM-based methods represent a more advanced approach [9, 10, 27, 29, 39, 42]. LayoutGPT constructed a visual layout framework from textual descriptions, extracting spatial knowledge from image datasets, such as object bounding boxes, and encoding this knowledge using CSS-style formatting. The spatial knowledge was then fed into LLM prompts, enabling the model to leverage contextual understanding to generate higher-level and lower-level layouts, which were subsequently translated into image layouts by image generation models. However, challenges remained, particularly in handling fine-grained textual descriptions related to angles and distances, as well as the generalization issues stemming from the potential lack of relevant images in current datasets for specific tasks. Building on this, GLIGEN [23] introduced bounding boxes and regional descriptions as input conditions for generating images. + +# 3. Method + +As shown in Fig. 1, we propose the SKE-Layout framework for generating generalizable layouts from textual descriptions. The framework is structured in two main parts: Knowledge Base Construction and Layout Gener + +![](images/dd16667849688fe10c39a69ee06e69a55c0f1112374fc5b253512b640ada4b85.jpg) +Figure 1. Framework of SKE-Layout. The framework comprises two main parts: Knowledge Base Construction, which builds a database from real-world images and LLM-generated knowledge; and Layout Generation Pipeline, which includes (a) Input, (b) Knowledge Retrieval, (c) Layout Planning and (d) Simulation. + +![](images/2260e8b230a4ff97a95f2401166b82dcf16c40d2865de3b8cd54d85cea393892.jpg) + +ation Process. Knowledge Base Construction (Sec. 3.2) gathers task knowledge from two aspects. The first aspect involves real-world extraction, using a Visual Language Model (VLM) for scene interpretation and a Detection Transformer (DETR) [4] for object detection and bounding box generation. The second aspect focuses on task-specific knowledge generation using LLMs. Together, these aspects form a robust foundation for layout generation. After the knowledge base is built, the Layout Generation Process proceeds with the Input module, which employs the same real-world extraction methods as the Knowledge Base Construction. Then, the Knowledge Retrieval module (Sec. 3.4) enhances the system's understanding through data augmentation and multi-task learning, optimizing user instructions using relevant knowledge from the knowledge base. Subsequently, the Layout Planning module (Sec. 3.5) integrates all inputs to create a detailed and optimized layout that aligns with user instructions. Finally, the Simulation module (Sec. 3.6) implements the layout in a virtual environment, providing visual verification to ensure task accuracy and consistency with the given description. + +# 3.1. Problem Settings + +In this study, we establish a unified problem setting for layout generation, encompassing both object rearrangement and image generation tasks. We define a reference coordinate system and origin tailored to each task. For object rearrangement, the reference is a 3D coordinate system + +$(x,y,z)\in \mathbb{R}^3$ , with the origin at the top-left corner of the rearrangement space, while for image generation, it is a 2D coordinate system $(x,y)\in \mathbb{R}^2$ , with the origin at the topleft corner of the image. + +The spatial information of an object, referred to as its layout, is represented by a tuple + +$$ +o = (\mathrm {i d}, c, x, y, z, w, h, d, \theta), +$$ + +where id denotes the object name, $c$ is the color, $(x,y,z)$ represents the object's center coordinates, and $w$ , $h$ , $d$ , $\theta$ denote the bounding box's width, height, depth, and the rotation angle in the $x-y$ plane, respectively. In image generation tasks, only the 2D coordinates $(x,y)$ and bounding box dimensions $(w,h)$ are relevant, while 3D attributes like $z$ , $d$ , and $\theta$ are often missing or irrelevant. In contrast, object rearrangement tasks require inferring the 3D coordinates $(x,y,z)$ and the rotation angle $\theta$ , while color and dimensions $w$ , $h$ , $d$ are known. Given a textual description $\mathcal{I}$ , the objective of layout generation is to produce a set of tuples $\mathcal{O}$ , where each tuple $o\in \mathcal{O}$ captures the spatial attributes of the objects as described. + +Spatial knowledge is defined as information related to the positioning, orientation, and arrangement of objects within a given environment, expressed as pairs of either $\{\text{Instruction - Object}\}$ , where $\mathcal{O} = \{o_1, o_2, \ldots, o_i\}$ , or $\{\text{Instruction - Mathematical constraint}\}$ , where the latter represents numerical constraints related to object positioning. In this context, 'knowledge' specifically refers to spa + +tial knowledge for the sake of brevity. + +# 3.2. Knowledge Base Construction + +In Retrieval-Augmented Generation (RAG) [12], constructing a comprehensive knowledge base is essential. This repository of exemplars and contextual information enhances the model's generative capabilities, allowing the system to retrieve relevant examples and produce precise, contextually appropriate layouts, thereby improving the efficacy and reliability of RAG. + +# 3.2.1. Extracting Knowledge from Real-World Images + +The primary aspect for knowledge base construction is to extract information from real-world images. We use VLMs and DETR for object detection and bounding box acquisition, capturing object location and dimensions to build a detailed representation of real-world scenes. By aggregating layouts from diverse images, we create a rich knowledge base that accurately reflects spatial arrangements and object interactions. The extracted information provides tuples $o \in \mathcal{O}_{2D}$ and $o \in \mathcal{O}_{3D}$ , with each tuple representing object category, location, size, and orientation. Additionally, VLMs capture spatial relationships between objects, stored as instruction information $I$ for each image in the knowledge base, aiding in retrieving relevant images and spatial context for layout generation based on new instructions. + +# 3.2.2. Generating Knowledge with LLMs + +The second aspect uses LLMs to generate diverse tasks $I$ and layouts $O$ from existing instructions collected in other research[8, 22, 26]. By interpreting and synthesizing text, LLMs create tasks that span a broad spectrum of real-world and hypothetical scenarios, introducing variability crucial for building a versatile knowledge base. This process begins by inputting varied instructions into the LLMs, which produces corresponding tasks not as duplicates but as novel scenarios. For each task, the LLMs generate layouts specifying object categories, spatial arrangements, and interobject relationships. This strategy enriches the knowledge base with synthetic scenarios that improve the system's adaptability to complex and rare instructions. Mathematical constraints $M = \{m_{n}\}$ are also integrated to meet spatial requirements like distances and angles, ensuring layout accuracy. + +To address potential inaccuracies, the synthetic data undergoes rigorous validation. First, VLMs verify if the generated images satisfy the original instructions (e.g., object arrange requirement, image composition) by evaluating object categories, positions, and spatial relationships. If discrepancies are found, human evaluators review the images for realism and relevance. This validation ensures the quality and reliability of data added to the RAG system. + +# 3.3. Input + +This section describes the configuration of inputs for different task settings within the SKE-Layout framework. For the image generation task, the input consists solely of user instructions, which outline the requirements for the generated image layout. The object rearrangement task involves a more complex input configuration. In addition to user instructions, it incorporates robot observations, which provide visual data of the environment and capture the spatial layout and positioning of objects within the scene through VLMs and DETR methods, as described in Sec. 3.2. + +# 3.4. Knowledge Retrieval + +After constructing our knowledge base, we need to retrieve knowledge in the database based on the current user instruction. Given a test instruction $I$ and the knowledge base $D = \{(I_k, o_k) \mid k = 1, 2, \ldots\}$ , we define a function $f(I_k, I_j) \in \mathbb{R}$ that measures the distances between two instructions. Initially, we experiment with directly using existing text embedding models [3, 7] to convert instructions into vectors and calculate their similarity. However, this approach reveals a significant issue: different tasks emphasize different aspects of the information. For example, in the instruction "place the banana $30^\circ$ above the apple", the specific fruits are less relevant than their spatial relationship. Standard embedding models can skew the retrieval results by overemphasizing the objects rather than their spatial arrangement. + +# 3.4.1. Instance-Wise Contrastive Learning + +To address this issue, we train a custom text embedding model through contrastive learning, building on the architecture of BERT [25] to meet our retrieval needs. This model is tailored for specific tasks, emphasizing the importance of spatial relationships over individual object identities, especially in object placement scenarios. To further enhance robustness, we apply data augmentation techniques using LLMs. For instance, in the instruction "place the banana $30^{\circ}$ above the apple," we generalize it to "place object A $30^{\circ}$ above object B," allowing the model to prioritize spatial relationships while ignoring specific object identities. + +Inspired by SimCSE [11], let $S = [I_n]_{n=1}^N$ be the set of original instructions, and $S' = [I_n, I_n']_{n=1}^N$ be the set after data augmentation, where $I_n'$ is the augmented instruction derived from $I_n$ . We use a pre-trained BERT model $\psi$ to obtain the embeddings of the instructions. For each instruction $I_n$ , we define its unique positive pair as $(I_n, I_n')$ , and all other pairs are treated as negative samples. The similarity between the two embeddings is calculated as follows: + +$$ +\operatorname {s i m} \left(\psi \left(I _ {n}\right), \psi \left(I _ {m}\right)\right) = \frac {\psi \left(I _ {n}\right) ^ {T} \psi \left(I _ {m}\right)}{\| \psi \left(I _ {n}\right) \| _ {2} \| \psi \left(I _ {m}\right) \| _ {2}} \tag {1} +$$ + +The contrastive loss function is defined as: + +$$ +\ell_ {n} ^ {N} = - \log \frac {\exp \left(\operatorname {s i m} \left(\psi \left(I _ {n}\right) , \psi \left(I _ {n} ^ {\prime}\right)\right)\right)}{\sum_ {m = 1 , m \neq n} ^ {2 N} \exp \left(\operatorname {s i m} \left(\psi \left(I _ {n}\right) , \psi \left(I _ {m}\right)\right)\right)} \tag {2} +$$ + +The overall contrastive loss is the average of all individual losses: + +$$ +\mathcal {L} _ {\text {I n s t a n c e - C L}} = \frac {1}{2 N} \sum_ {n = 1} ^ {2 N} \ell_ {n} ^ {N} \tag {3} +$$ + +Algorithm 1 Training Procedure for Multi-task Instance-wise Contrastive Learning + +1: Input: Instruction set $S = \{S_{\mathrm{image}}, S_{\mathrm{object}}\}$ containing tasks for image generation and object manipulation +2: Output: Trained model +3: for each $I_{n} \in S$ do +4: Generate augmented instruction $I_{n}^{\prime}$ using LLMs +5: end for +6: Obtain embeddings $\psi(I_n)$ and $\psi(I_n')$ for all instructions using task-specific embedding heads in the pretrained BERT model +7: for each $I_{n} \in S_{\mathrm{image}}$ do +8: Calculate the image generation contrastive loss: + +$$ +\ell_ {n} ^ {\text {I m a g e - C L}} = \operatorname {L o s s} \left(\psi \left(I _ {n}\right), \psi \left(I _ {n} ^ {\prime}\right)\right) \tag {4} +$$ + +9: Update model parameters to minimize $\mathcal{L}_{\mathrm{Image - CL}}$ +10: end for +11: for each $I_{n} \in S_{\mathrm{object}}$ do +12: Calculate the object manipulation contrastive loss: + +$$ +\ell_ {n} ^ {\text {O b j e c t - C L}} = \operatorname {L o s s} \left(\psi \left(I _ {n}\right), \psi \left(I _ {n} ^ {\prime}\right)\right) \tag {5} +$$ + +13: Update model parameters to minimize $\mathcal{L}_{\mathrm{Object - CL}}$ +14: end for +15: Compute the overall multi-task contrastive loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {M u l t i - T a s k - C L}} = \lambda_ {\text {i m a g e}} \sum_ {n = 1} ^ {| S _ {\text {i m a g e}} |} \ell_ {n} ^ {\text {I m a g e - C L}} \\ + \lambda_ {\text {o b j e c t}} \sum_ {n = 1} ^ {| S _ {\text {o b j e c t}} |} \ell_ {n} ^ {\text {O b j e c t - C L}} \tag {6} \\ \end{array} +$$ + +# 3.4.2. Multi-Task Learning + +The existing learning algorithm has limitations when handling different tasks, particularly in multi-task scenarios such as image generation and object manipulation. Since each task focuses on different key information, existing + +models struggle to leverage and learn both the commonalities and differences between tasks, thereby limiting their generalization ability. Furthermore, when multiple complex tasks coexist, using separate models for each task often fails to exploit complementary information among tasks, resulting in low learning efficiency and suboptimal model performance. To improve adaptability, we design a multi-task learning strategy [5] that trains the embedding model on both 3D object placement and 2D layout tasks, enhancing generalization across spatial, temporal, and categorical relationships. + +Algorithm 1 outlines the training procedure for Multi-task Instance-wise Contrastive Learning. The instruction set $S$ , containing tasks for image generation and object manipulation, is processed through the following steps. First, for each instruction $I_{n}$ in $S$ , an augmented instruction $I_{n}^{\prime}$ is generated using LLMs (Lines 3-5). These instructions, along with their original versions, are then passed through task-specific embedding heads in a pre-trained BERT model to obtain their corresponding embeddings $\psi(I_{n})$ and $\psi(I_{n}^{\prime})$ (Line 6). For the image generation tasks $S_{\text{image}}$ , the model calculates the image generation contrastive loss $\ell_{\text{image-CL}}$ (Lines 7-9). Similarly, for the object manipulation tasks $S_{\text{object}}$ , the object manipulation contrastive loss $\ell_{\text{object-CL}}$ is computed (Lines 11-13). Finally, the overall multi-task contrastive loss $\mathcal{L}_{\text{Multi-Task-CL}}$ is computed by combining the individual losses from both tasks (Line 15). The model parameters, including those of the task-specific embedding heads, are updated to minimize the loss (Lines 9 and 13). + +# 3.5. Layout Planning + +The Layout Planning module is responsible for determining the target positions and orientations of objects based on high-level user instructions, adhering to a unified problem setting for layout generation. The process integrates User Instructions and Spatial Knowledge into a comprehensive prompt for both object rearrangement and image generation tasks, with Robot Observations added in the object rearrangement task to assist in identifying relevant objects and their attributes. For object rearrangement, during the object selection phase, the system uses the Object Selection LLM to identify relevant objects and determine their attributes as $(id, c, w, h, d)$ . In the layout generation phase, the Layout Generation LLM calculates the spatial configuration of each object, inferring the 3D coordinates $(x, y, z)$ and rotation angle $\theta$ while using known attributes such as width, height, depth $(w, h, d)$ , and color $c$ . By assigning precise poses $(x, y, z, \theta)$ to the objects, the Layout Planning module completes the layout generation based on the unified problem setting, allowing the process to move into the Simulation Phase, where the objects are arranged according to the determined layout. The image generation task follows a similar approach, where the system also generates object + +Table 1. Success rates for different models across task types and subtasks in SK-Dataset + +
ModelSuccess Rate (%)
Task TypeSimple TaskComplex Task
Sub Task Type-LettersGeometric ShapesSemantic Geometric Shapes
GPT3.5SKE-Layout (ours)42.211.439.45.8
w/o CL35.210.940.64.3
LLM-GROP26.36.124.73.8
GPT4SKE-Layout (ours)75.142.772.621.7
w/o CL71.541.369.219.2
LLM-GROP69.49.418.54.5
+ +Table 2. Evaluation of methods on NSR-1K dataset + +
MethodsNumerical ReasoningSpatial Reasoning
Layout Eval.Image Eval.Layout Eval.Image Eval.
PrecisionRecallAccuracyAcc. (GLIP)AccuracyAcc. (GLIP)
Text → Image
Stable Diffusion (v2.1)---42.44-17.81
Attend-and-Excite (SD v2.1)---45.74-26.86
Text → Layout → Image
LayoutTransformer [17]75.7061.6922.2640.556.3628.13
LayoutGPT (GPT-4)78.3686.2978.4355.6491.7360.64
w/o CL86.0897.4695.2856.8694.3561.13
w/o ML85.2398.8097.1158.6894.7061.48
SKE-Layout (ours)85.2899.0697.2459.7095.0562.19
GT layouts100.00100.00100.0053.23100.0062.54
Human99.2696.5292.5656.0791.1751.94
+ +Layouts $(id, c, w, d, x, y)$ based on task requirements. + +# 3.6. Scene Simulation + +For object rearrangement task, we input the objects and their coordinates into the PyBullet simulation environment to create a 3D scene. This allows us to observe and evaluate the arrangement of objects in a simulated physical space, ensuring that the generated layouts are feasible and accurate in a realistic setting. For the image generation task, we use the layout-to-image generation model, GLIGEN. The model generates images based on the object layouts, ensuring that the categories, locations, and sizes of the objects in the generated image precisely match the predicted layouts described in prior works. + +# 4. Experiments + +In this section, we present extensive empirical results of SKE-Layout and several competitive baseline methods. From the representative spatial reasoning-induced domains, including both robotic object rearrangement and image generation tasks, we observe that SKE-Layout can significantly improve task success rate compared to all baselines. + +# 4.1. Experimental Setup + +# 4.1.1. Baselines + +Given the two distinct tasks in our experiments, it is imperative to employ different baseline models for a comprehensive comparison. We select a range of competitive baselines that represent the state-of-the-art approaches in both image generation tasks and robotic object rearrangement tasks. + +For the Robotic Object Rearrangement Task, we utilize StructFormer [24], a Transformer-based model designed for capturing spatial relationships, and LLM-GROP [8] that integrates LLMs with planning techniques. We also include the SKE-Layout variant trained without contrastive learning (w/o CL), which uses BERT for processing, to evaluate object arrangement tasks without contrastive loss. + +For the Image Generation Task, we select several baselines: Stable Diffusion [33], an end-to-end diffusion-based text-to-image model; Attend-and-Excite, which uses attention mechanisms for detail-oriented image generation; LayoutTransformer [13], a text-to-query model for layout generation; and LayoutGPT, [9], which leverages LLMs for text-to-query generation. We also include our SKE-Layout model, along with its variants trained without contrastive learning (w/o CL) or multitask learning (w/o ML) to assess + +their individual contributions. + +# 4.1.2. Datasets + +We evaluate the performance of our proposed method on three different datasets, each tailored to a specific task, including image generation and robotic object rearrangement, to ensure a comprehensive assessment. + +- NSR-1K Dataset: Derived from the LayoutGPT framework, it includes both template-based and human-written prompts sourced from MSCOCO, designed to evaluate image generation based on specified object counts and spatial locations. +- StructFormer Dataset: Designed specifically for the robotic object rearrangement task, this dataset includes scenarios where objects must be arranged based on given constraints (e.g., circle, line, tower, table setting), testing the model's ability to maintain structural coherence and execute tasks accurately. We do not use the training dataset from StructFormer, which serves as an Out-Of-Distribution (OOD) test for evaluating the model's generalization to unseen tasks. +- SK-Dataset: Created by us to evaluate complex object rearrangement tasks, the SK-Dataset includes advanced scenarios with multi-object planning and execution. Tasks are divided into two levels: simple and complex. The simple tasks focus on accurately positioning objects based on specified distances and positions. The complex tasks require higher-order reasoning and are further divided into letters, geometric shapes, and semantic geometric shapes, emphasizing spatial precision, symmetry, and meaningful arrangements. + +# 4.2. Results and Discussions + +SK Dataset. Table 1 compares the success rates of different models across SK-Dataset task types, including GPT-3.5 and GPT-4 using our method, w/o CL variant, and LLM-GROP. Overall, our method achieves consistently high success rates, with the degree of improvement varying by task complexity. For example, in simple tasks with GPT-4, our method reached a $75.1\%$ success rate, outperforming w/o CL $(71.5\%)$ and LLM-GROP $(69.4\%)$ . In contrast, for complex tasks, our approach shows a more pronounced advantage, increasing the success rate from $4.5\%$ (LLM-GROP) to $21.7\%$ . Two key observations emerge from Table 1: + +- While improvements in simple tasks are moderate, our method significantly outperforms baselines in complex tasks, particularly in handling semantic geometric shapes, demonstrating its ability to enhance comprehension and reasoning in high-difficulty scenarios. +- GPT-4 consistently outperforms GPT-3.5, especially when combined with our method. This highlights the effectiveness of our joint training strategy in enhancing the model's understanding of complex information and its adaptability across different task difficulties. + +StructFormer Dataset. Fig. 2 compares model performance across four object rearrangement tasks: Tower, Line, Circle, and Set the Table. Our model consistently outperforms StructFormer, with improvements of $30\%$ , $20\%$ , $20\%$ , and $10\%$ , respectively. This demonstrates our method's advantage in managing tasks by effectively integrating external knowledge, enhancing both accuracy and reliability. Several key insights emerge from these results: + +- The improvement is more significant in complex tasks, indicating that our approach excels in scenarios requiring higher levels of reasoning and understanding. +- Unlike StructFormer, our method maintains strong performance across all tasks without separate task-specific models, showcasing its versatility and adaptability to diverse object rearrangement tasks. + +![](images/2e687dda496510f58934e689b52922b53232af894613b54a0453062be902d9ac.jpg) +Figure 2. Evaluation of methods on the StructFormer dataset on various tasks: Tower, Line, Circle, and Set the Table. + +NSR-1K Dataset. Table 2 presents the main results of image generation, showing that our method significantly outperforms comparative models across multiple evaluation metrics. In numerical reasoning, it exceeds LayoutGPT (GPT-4) by $10\% - 20\%$ in precision, recall, and accuracy. In spatial reasoning, our model achieves $95.05\%$ layout accuracy and $61.48\%$ GLIP accuracy, outperforming LayoutGPT by $3\%$ and $1\%$ , respectively. This improvement is attributed to the more comprehensive spatial knowledge of our method, which allows it to retrieve more relevant spatial information, thereby providing higher precision in both layout and image generation. + +# 4.3. Ablation Study + +Q1: What happens when we remove contrastive learning or multitask learning from the model? The performance on both the NSR-1K and SK datasets decreases when contrastive learning or multitask learning is disabled. This indicates that these two techniques are crucial. Specifically, contrastive learning helps the model better distinguish subtle spatial relationships, while multitask learning improves the model's ability to understand context across different tasks. Together, they contribute to more effective retrieval of spatial knowledge. + +![](images/ca2dceb4a6058f1ccbf66dbd78e508df28bf1a5e0faf378a1efebbd71aa863dd.jpg) +Figure 3. Case Study. We present a comparison of different methods in 3D object rearrangement and 2D image generation tasks. (a) arranging objects on a table with correct spatial relationships, (b) placing blocks in a specified formation, (c) generating a 2D image with objects positioned according to a textual description. + +Q2: Why doesn't the StructFormer dataset show similar performance drops? The performance differences among our model variants on the StructFormer dataset are minimal. This is likely because the spatial relationships required in this dataset are relatively straightforward, making precise retrieval of complex spatial knowledge less critical. As a result, the advantages of contrastive learning and multitask learning become less pronounced. + +# 4.4. Case Study + +We present three representative cases to demonstrate the effectiveness of our SKE-Layout framework across both 3D object rearrangement and 2D image generation tasks. Fig. 3.(a) involves setting the table and arranging objects into a smiley face pattern. While the LLM-GROP model fails to maintain correct angles and semantic coherence, our SKE-Layout significantly reduces errors, providing accurate placement and alignment. In Fig. 3.(b), where blocks need to be arranged into a circular pattern, our method achieves perfect spatial precision, outperforming both StructFormer and LLM-GROP, which struggle with alignment. For the 2D image generation task in Fig. 3.(c), our method also excels. While Stable Diffusion generates quantity errors and LayoutGPT shows positional errors, SKE-Layout accurately captures object relationships, ensuring correct quantity and positioning and thus closely matching the given description. From Fig. 3, we draw the following observations: + +- Richer Spatial Knowledge. Our SKE layout framework demonstrates a deeper understanding of spatial relationships than baseline models. By integrating diverse spatial knowledge sources, including real-world and synthetic data, our approach enhances the richness and depth of spatial reasoning. Additionally, it employs multiple forms of spatial knowledge representation, which contribute to more accurate positioning and alignment of objects in both 3D rearrangement and 2D image generation tasks. +- Adaptability to Multi-Task Settings. The SKE layout framework excels in handling a variety of tasks, showcasing versatility in both 3D object arrangement and 2D image generation. This adaptability highlights its robustness and effectiveness in addressing complex spatial reasoning challenges across multiple domains. + +# 5. Conclusion + +In this paper, we present SKE.layout, a novel method designed to effectively generate layouts by leveraging large language models. Our approach injects spatial knowledge into the large language model, thereby demonstrating its versatility and robustness across different application domains. Empirically, we have demonstrated that SKE.layout performs exceptionally well in both image generation and robotic task arrangement tasks. Future research directions include exploring how to apply SKE.layout to more application scenarios, such as open-domain scenarios, where the set of objects in a problem instance is not known in advance. + +# Acknowledgement + +# References + +[1] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020. 1 +[2] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446):eaat8414, 2019. 1 +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 4 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[5] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 5 +[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[7] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4 +[8] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 1, 2, 4, 6 +[9] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6 +[10] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LIm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2 +[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 4 + +This work was partially supported by the National Natural Science Foundation of China under Grant 62322601 and 42174050, the Independent Research Project of State Key Laboratory of Mechanical Transmission for Advanced Equipment under Grant SKLMT-ZZKT2024R07, the Fundamental Research Funds for the Central Universities 2024IAIS-QN01, the Postdoctoral Fellowship Program of CPSF under Grant GZB20230912, the China Postdoctoral Science Foundation under Grant 2024M753837, the Chongqing Postdoctoral Innovation Talents Support Program under Grant CQBX202324 and the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQJQX0025 and CSTB2024NSCQMSX0955. +[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 4 +[13] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2, 6 +[14] Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints, 2023. 1 +[15] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2 +[16] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layout: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 2 +[17] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. arXiv preprint arXiv:2309.10062, 2023. 2 +[18] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. IEEE Robotics and Automation Letters, 8(7):3956-3963, 2023. 2 +[19] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1 +[20] Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329, 2022. 1 +[21] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Generating graphic layouts with wireframe discriminators. arXiv preprint arXiv:1901.06767, 2019. 2 +[22] Jiachen Li, Qiaozi Gao, Michael Johnston, Xiaofeng Gao, Xuehai He, Suhaila Shakiah, Hangjie Shi, Reza Ghanadan, and William Yang Wang. Mastering robot manipulation with multimodal prompts through pretraining and multi-task finetuning. arXiv preprint arXiv:2310.09676, 2023. 4 +[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[24] Weiyu Liu, Chris Paxton, Tucker Hermans, and Dieter Fox. Structformer: Learning spatial structure for language-guided + +semantic rearrangement of novel objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 6322-6329. IEEE, 2022. 2, 6 +[25] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering, 35(1):857-876, 2021. 4 +[26] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 4 +[27] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 2 +[28] Ruoling Peng, Kang Liu, Po Yang, Zhipeng Yuan, and Shunbao Li. Embedding-based retrieval with llm for effective agriculture information extracting from unstructured data. arXiv preprint arXiv:2308.03107, 2023. 1 +[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 1, 2 +[30] Shuhan Qi, Zhengying Cao, Jun Rao, Lei Wang, Jing Xiao, and Xuan Wang. What is the limitation of multimodal llms? a deeper look into multimodal llms through prompt probing. Information Processing & Management, 60(6):103510, 2023. 1 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[32] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable task planning. CoRR, 2023. 2 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6 +[34] Sai Sabitha, Deepti Mehrotra, and Abhay Bansal. Knowledge enriched learning by converging knowledge object&learning object. Electronic Journal of e-Learning, 13 (1):pp3-13, 2015. 1 +[35] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/rearrange_2022, 2022. 2 +[36] Jiaqi Wang, Zihao Wu, Yiwei Li, Hanqi Jiang, Peng Shu, Enze Shi, Huawen Hu, Chong Ma, Yiheng Liu, Xuhui Wang, et al. Large language models for robotics: Opportunities, challenges, and perspectives. arXiv preprint arXiv:2401.04334, 2024. 2 + +[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1 +[38] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[39] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 1, 2 +[40] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4303-4310. IEEE, 2024. 2 +[41] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1 +[42] Shanshan Zhong, Zhongzhan Huang, Weushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 567-578, 2023. 2 \ No newline at end of file diff --git a/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/images.zip b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a609824870959e067e65d53743a3cddd33ca5f7 --- /dev/null +++ b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bef48206dc623772bb5e0bb16cc3668db75fa34c3258ca30ff78bb192d09445 +size 431615 diff --git a/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/layout.json b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..165e11afa4d1ea2a05f2fb4c2940b9209f65b682 --- /dev/null +++ b/2025/SKE-Layout_ Spatial Knowledge Enhanced Layout Generation with LLMs/layout.json @@ -0,0 +1,8231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 536, + 121 + ], + "type": "text", + "content": "SKE-Layout: Spatial Knowledge Enhanced Layout Generation with LLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": "Junsheng Wang" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": " Nieqing Cao" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": " Yan Ding" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": " Mengying Xie" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": " Fuqiang Gu" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": " Chao Chen" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": "Chongqing University " + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": "Xi'an Jiaotong-Liverpool University " + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 142, + 543, + 171 + ], + "type": "text", + "content": "Shanghai AI Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 173, + 445, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 173, + 445, + 185 + ], + "spans": [ + { + "bbox": [ + 164, + 173, + 445, + 185 + ], + "type": "text", + "content": "{csjunshengwang, xiemy, gufq, cschaochen}@cqu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 187, + 433, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 187, + 433, + 198 + ], + "spans": [ + { + "bbox": [ + 172, + 187, + 433, + 198 + ], + "type": "text", + "content": "Nieqing.Cao@xjtlu.edu.cn, yding25@binghamton.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 252, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 456 + ], + "type": "text", + "content": "Generating layouts from textual descriptions by large language models (LLMs) plays a crucial role in precise spatial reasoning-induced domains such as robotic object rearrangement and text-to-image generation. However, current methods face challenges in limited real-world examples, handling diverse layout descriptions and varying levels of granularity. To address these issues, a novel framework named Spatial Knowledge Enhanced Layout (SKE-Layout), is introduced. SKE-Layout integrates mixed spatial knowledge sources, leveraging both real and synthetic data to enhance spatial contexts. It utilizes diverse representations tailored to specific tasks and employs contrastive learning and multitask learning techniques for accurate spatial knowledge retrieval. This framework generates more accurate and fine-grained visual layouts for object rearrangement and text-to-image generation tasks, achieving improvements of " + }, + { + "bbox": [ + 55, + 252, + 297, + 456 + ], + "type": "inline_equation", + "content": "5\\% -30\\%" + }, + { + "bbox": [ + 55, + 252, + 297, + 456 + ], + "type": "text", + "content": " compared to existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 135, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 135, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 135, + 492 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": "Generating layouts from textual descriptions is fundamental in various domains, particularly in robotic object rearrangement and text-to-image generation, where achieving precise object placement is critical. For example, an object rearrangement task might require arranging specific items on a table [1, 8], while an image generation task often involves creating a visually coherent scene based on descriptions of object locations [29, 39]. Both scenarios underscore the necessity of accurate spatial understanding to ensure proper placement and prevent errors such as misalignment or omissions. Research highlights the need to minimize these errors in robotic manipulation and image generation in order to reflect spatial instructions accurately [2]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 657, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 295, + 693 + ], + "type": "text", + "content": "In recent years, a growing body of research has investigated the application of large language models (LLMs) for layout generation [9, 29, 39], highlighting their no" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 312 + ], + "type": "text", + "content": "table advantages in various aspects. Primarily, LLMs exhibit strong zero-shot learning capabilities, generating coherent layouts from natural language descriptions without task-specific training, enabling broad generalization across diverse tasks [37]. Unlike traditional models, LLMs adapt seamlessly to different domains without requiring separate training for each task [19, 41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 554, + 542 + ], + "type": "text", + "content": "Although LLMs exhibit numerous advantages in layout generation, they also face challenges related to hallucination when directly generating layouts from textual descriptions. Existing approaches often rely on using predefined layouts as examples to guide LLMs in generating new layouts, allowing the model to infer reasonable spatial arrangements based on these examples [9]. However, there is often a lack of sufficient and diverse examples in real-world scenarios, particularly for complex layouts, exacerbating the model's dependence on existing examples. As representations, layouts often fail to fully capture the spatial knowledge of the current scene [14, 30]. When handling fine-grained instructions related to angles, distances, and other spatial details, LLMs struggle to infer the corresponding layout solely by relying on example layouts. Furthermore, due to the varying importance of information across different tasks, directly applying existing retrieval methods may result in the inability to retrieve the most relevant knowledge for the specific task at hand [28]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose SKE-Layout, the first framework capable of handling diverse layout task descriptions and inferring unknown object attributes, providing a unified solution that current frameworks cannot achieve. Specifically, SKE-Layout extracts diverse spatial knowledge from real-world or synthetic tasks and injects the most relevant spatial knowledge into LLMs, enabling them to generate ideal layouts based on textual conditions. Despite not using any image data for training, LLMs can learn spatial knowledge through contextual demonstrations and then apply this knowledge to visual reasoning for new samples [20, 34]. In other words, by providing relevant examples, LLMs can capture spatial relationships between objects based on language understanding, in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 211, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 211, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 211, + 713 + ], + "type": "text", + "content": "*Equal contribution, †Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19414" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "ferring information, such as relative positions, sizes, and distances of objects, thereby achieving accurate object rearrangement and coherent image generation [6]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 110, + 269, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 110, + 269, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 110, + 269, + 121 + ], + "type": "text", + "content": "In summary, our main contributions are threefold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 295, + 301 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 122, + 295, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 194 + ], + "type": "text", + "content": "- We propose a novel approach SKE.layout with LLMs for generating layouts from textural descriptions more precisely. SKE.layout integrates multiple spatial knowledge sources and combines real-world spatial datasets with model-generated data from LLMs, enhancing the spatial knowledge database without relying on existing datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 294, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 253 + ], + "type": "text", + "content": "- We present a diverse framework for spatial knowledge representations. It adapts to the specific needs of different tasks by providing precise quantitative expressions (e.g., \"30-degree angle\", \"10 cm away\") and qualitative spatial relations (e.g., \"zebra to the left of a giraffe\")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 254, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 294, + 301 + ], + "type": "text", + "content": "- We construct an innovative spatial knowledge retrieval method, which enhances spatial context extraction by combining contrastive learning with multitask learning, optimizing accuracy for complex visual layout tasks." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 302, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 397 + ], + "type": "text", + "content": "Overall, the SKE-Layout framework can effectively address the challenges in generating accurate and fine-grained visual layouts by leveraging multiple spatial knowledge sources, diverse representations, and innovative retrieval techniques. Its versatility allows it to simultaneously handle object rearrangement and image generation tasks, demonstrating its generalizability across different layout generation domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 142, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 142, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 142, + 423 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 432, + 186, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 186, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 186, + 444 + ], + "type": "text", + "content": "2.1. Object Rearrangement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "Object rearrangement is a crucial task for service robots, requiring the movement of objects from one location to another and arranging them according to specified targets. This task is exemplified by challenges such as the Habitat Rearrangement Challenge [35] and the AI2-THOR Rearrangement Challenge [38]. Various approaches have been explored to tackle this problem. One deep learning-based method, StructFormer [24], utilized transformer-based neural networks to learn fine-grained object arrangement representations from real-world datasets, enabling the model to infer semantically meaningful structures from natural language instructions. However, this method presented challenges in terms of model complexity, as each task requires independent dataset construction and training, placing high demands on data collection and computational resources. Additionally, its generalization capabilities were limited, as the model struggled to handle tasks it failed to be specifically trained on. Another approach, DALL-E-Bot [18], leveraged DALL-E 2 [31] to generate images, which were then used for object placement in table-top scenarios. Nonetheless, DALL-E's image generation lacked sufficient stability when it came to producing task-relevant images," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "which could hinder object arrangement accuracy. To address these limitations, methods based on LLMs have been proposed [8, 15, 17, 32, 36, 40]. LLM-GROP [8] extracted commonsense object arrangement knowledge from LLMs and applied this knowledge to perform complex manipulation and object placement tasks. SayPlan [32] introduced a novel approach to grounding LLMs in 3D scene graphs, enabling scalable task planning by translating high-level instructions into structured scene graph representations and facilitating object manipulation in complex environments. SG-Bot [40] proposed a coarse-to-fine strategy for object rearrangement, leveraging scene graphs to guide robotic imagination from scene understanding to detailed object placements. This method emphasized hierarchical planning, where initial rough layouts were refined through iterative adjustments. Despite the potential of LLMs in facilitating object rearrangement, their spatial reasoning capabilities still face limitations, particularly in accurately interpreting and generalizing complex spatial relationships across diverse scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 458, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 458, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 458, + 331 + ], + "type": "text", + "content": "2.2. Image Layout Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 336, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 553, + 635 + ], + "type": "text", + "content": "Image layout generation is a critical component in accurately creating complex images. Various approaches have been proposed to generate layouts for images, each utilizing different models and input conditions to accomplish the task. Deep learning-based methods, such as LayoutGAN [21], LayoutVAE [16], and LayoutTransformer [13], typically require specific scenarios or input conditions for training, producing layouts based on fixed class labels or partially input layouts. On the other hand, LLM-based methods represent a more advanced approach [9, 10, 27, 29, 39, 42]. LayoutGPT constructed a visual layout framework from textual descriptions, extracting spatial knowledge from image datasets, such as object bounding boxes, and encoding this knowledge using CSS-style formatting. The spatial knowledge was then fed into LLM prompts, enabling the model to leverage contextual understanding to generate higher-level and lower-level layouts, which were subsequently translated into image layouts by image generation models. However, challenges remained, particularly in handling fine-grained textual descriptions related to angles and distances, as well as the generalization issues stemming from the potential lack of relevant images in current datasets for specific tasks. Building on this, GLIGEN [23] introduced bounding boxes and regional descriptions as input conditions for generating images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 645, + 370, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 370, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 370, + 658 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "content": "As shown in Fig. 1, we propose the SKE-Layout framework for generating generalizable layouts from textual descriptions. The framework is structured in two main parts: Knowledge Base Construction and Layout Gener" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19415" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 195, + 305 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 195, + 305 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 195, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 195, + 305 + ], + "type": "image", + "image_path": "dd16667849688fe10c39a69ee06e69a55c0f1112374fc5b253512b640ada4b85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 317, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 54, + 317, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 555, + 352 + ], + "type": "text", + "content": "Figure 1. Framework of SKE-Layout. The framework comprises two main parts: Knowledge Base Construction, which builds a database from real-world images and LLM-generated knowledge; and Layout Generation Pipeline, which includes (a) Input, (b) Knowledge Retrieval, (c) Layout Planning and (d) Simulation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 70, + 550, + 305 + ], + "blocks": [ + { + "bbox": [ + 195, + 70, + 550, + 305 + ], + "lines": [ + { + "bbox": [ + 195, + 70, + 550, + 305 + ], + "spans": [ + { + "bbox": [ + 195, + 70, + 550, + 305 + ], + "type": "image", + "image_path": "2260e8b230a4ff97a95f2401166b82dcf16c40d2865de3b8cd54d85cea393892.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 372, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 296, + 624 + ], + "type": "text", + "content": "ation Process. Knowledge Base Construction (Sec. 3.2) gathers task knowledge from two aspects. The first aspect involves real-world extraction, using a Visual Language Model (VLM) for scene interpretation and a Detection Transformer (DETR) [4] for object detection and bounding box generation. The second aspect focuses on task-specific knowledge generation using LLMs. Together, these aspects form a robust foundation for layout generation. After the knowledge base is built, the Layout Generation Process proceeds with the Input module, which employs the same real-world extraction methods as the Knowledge Base Construction. Then, the Knowledge Retrieval module (Sec. 3.4) enhances the system's understanding through data augmentation and multi-task learning, optimizing user instructions using relevant knowledge from the knowledge base. Subsequently, the Layout Planning module (Sec. 3.5) integrates all inputs to create a detailed and optimized layout that aligns with user instructions. Finally, the Simulation module (Sec. 3.6) implements the layout in a virtual environment, providing visual verification to ensure task accuracy and consistency with the given description." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 159, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 159, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 159, + 647 + ], + "type": "text", + "content": "3.1. Problem Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": "In this study, we establish a unified problem setting for layout generation, encompassing both object rearrangement and image generation tasks. We define a reference coordinate system and origin tailored to each task. For object rearrangement, the reference is a 3D coordinate system" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 372, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 419 + ], + "type": "inline_equation", + "content": "(x,y,z)\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 372, + 555, + 419 + ], + "type": "text", + "content": " , with the origin at the top-left corner of the rearrangement space, while for image generation, it is a 2D coordinate system " + }, + { + "bbox": [ + 313, + 372, + 555, + 419 + ], + "type": "inline_equation", + "content": "(x,y)\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 372, + 555, + 419 + ], + "type": "text", + "content": " , with the origin at the topleft corner of the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 420, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 554, + 445 + ], + "type": "text", + "content": "The spatial information of an object, referred to as its layout, is represented by a tuple" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 374, + 453, + 493, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 453, + 493, + 466 + ], + "spans": [ + { + "bbox": [ + 374, + 453, + 493, + 466 + ], + "type": "interline_equation", + "content": "o = (\\mathrm {i d}, c, x, y, z, w, h, d, \\theta),", + "image_path": "15ae6845c80928078b7b17cd715e02ae4c906ef471d7189e10e134feb7ed574c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": "where id denotes the object name, " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " is the color, " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " represents the object's center coordinates, and " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " denote the bounding box's width, height, depth, and the rotation angle in the " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "x-y" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " plane, respectively. In image generation tasks, only the 2D coordinates " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " and bounding box dimensions " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "(w,h)" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " are relevant, while 3D attributes like " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " are often missing or irrelevant. In contrast, object rearrangement tasks require inferring the 3D coordinates " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " and the rotation angle " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", while color and dimensions " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " are known. Given a textual description " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", the objective of layout generation is to produce a set of tuples " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": ", where each tuple " + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "inline_equation", + "content": "o\\in \\mathcal{O}" + }, + { + "bbox": [ + 313, + 474, + 555, + 629 + ], + "type": "text", + "content": " captures the spatial attributes of the objects as described." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Spatial knowledge is defined as information related to the positioning, orientation, and arrangement of objects within a given environment, expressed as pairs of either " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\text{Instruction - Object}\\}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{o_1, o_2, \\ldots, o_i\\}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\text{Instruction - Mathematical constraint}\\}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", where the latter represents numerical constraints related to object positioning. In this context, 'knowledge' specifically refers to spa" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19416" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 209, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 209, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 209, + 85 + ], + "type": "text", + "content": "tial knowledge for the sake of brevity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 219, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 219, + 110 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 219, + 110 + ], + "type": "text", + "content": "3.2. Knowledge Base Construction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 117, + 295, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 295, + 201 + ], + "type": "text", + "content": "In Retrieval-Augmented Generation (RAG) [12], constructing a comprehensive knowledge base is essential. This repository of exemplars and contextual information enhances the model's generative capabilities, allowing the system to retrieve relevant examples and produce precise, contextually appropriate layouts, thereby improving the efficacy and reliability of RAG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 288, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 288, + 224 + ], + "type": "text", + "content": "3.2.1. Extracting Knowledge from Real-World Images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "text", + "content": "The primary aspect for knowledge base construction is to extract information from real-world images. We use VLMs and DETR for object detection and bounding box acquisition, capturing object location and dimensions to build a detailed representation of real-world scenes. By aggregating layouts from diverse images, we create a rich knowledge base that accurately reflects spatial arrangements and object interactions. The extracted information provides tuples " + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "inline_equation", + "content": "o \\in \\mathcal{O}_{2D}" + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "inline_equation", + "content": "o \\in \\mathcal{O}_{3D}" + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "text", + "content": ", with each tuple representing object category, location, size, and orientation. Additionally, VLMs capture spatial relationships between objects, stored as instruction information " + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 228, + 296, + 396 + ], + "type": "text", + "content": " for each image in the knowledge base, aiding in retrieving relevant images and spatial context for layout generation based on new instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 233, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 233, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 233, + 419 + ], + "type": "text", + "content": "3.2.2. Generating Knowledge with LLMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "content": "The second aspect uses LLMs to generate diverse tasks " + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "content": " and layouts " + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "content": " from existing instructions collected in other research[8, 22, 26]. By interpreting and synthesizing text, LLMs create tasks that span a broad spectrum of real-world and hypothetical scenarios, introducing variability crucial for building a versatile knowledge base. This process begins by inputting varied instructions into the LLMs, which produces corresponding tasks not as duplicates but as novel scenarios. For each task, the LLMs generate layouts specifying object categories, spatial arrangements, and interobject relationships. This strategy enriches the knowledge base with synthetic scenarios that improve the system's adaptability to complex and rare instructions. Mathematical constraints " + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "inline_equation", + "content": "M = \\{m_{n}\\}" + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "content": " are also integrated to meet spatial requirements like distances and angles, ensuring layout accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "To address potential inaccuracies, the synthetic data undergoes rigorous validation. First, VLMs verify if the generated images satisfy the original instructions (e.g., object arrange requirement, image composition) by evaluating object categories, positions, and spatial relationships. If discrepancies are found, human evaluators review the images for realism and relevance. This validation ensures the quality and reliability of data added to the RAG system." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 364, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 364, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 364, + 84 + ], + "type": "text", + "content": "3.3. Input" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 555, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 211 + ], + "type": "text", + "content": "This section describes the configuration of inputs for different task settings within the SKE-Layout framework. For the image generation task, the input consists solely of user instructions, which outline the requirements for the generated image layout. The object rearrangement task involves a more complex input configuration. In addition to user instructions, it incorporates robot observations, which provide visual data of the environment and capture the spatial layout and positioning of objects within the scene through VLMs and DETR methods, as described in Sec. 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 223, + 435, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 435, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 435, + 236 + ], + "type": "text", + "content": "3.4. Knowledge Retrieval" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "text", + "content": "After constructing our knowledge base, we need to retrieve knowledge in the database based on the current user instruction. Given a test instruction " + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "text", + "content": " and the knowledge base " + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "inline_equation", + "content": "D = \\{(I_k, o_k) \\mid k = 1, 2, \\ldots\\}" + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "text", + "content": ", we define a function " + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "inline_equation", + "content": "f(I_k, I_j) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "text", + "content": " that measures the distances between two instructions. Initially, we experiment with directly using existing text embedding models [3, 7] to convert instructions into vectors and calculate their similarity. However, this approach reveals a significant issue: different tasks emphasize different aspects of the information. For example, in the instruction \"place the banana " + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "inline_equation", + "content": "30^\\circ" + }, + { + "bbox": [ + 313, + 242, + 555, + 422 + ], + "type": "text", + "content": " above the apple\", the specific fruits are less relevant than their spatial relationship. Standard embedding models can skew the retrieval results by overemphasizing the objects rather than their spatial arrangement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 497, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 497, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 497, + 445 + ], + "type": "text", + "content": "3.4.1. Instance-Wise Contrastive Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": "To address this issue, we train a custom text embedding model through contrastive learning, building on the architecture of BERT [25] to meet our retrieval needs. This model is tailored for specific tasks, emphasizing the importance of spatial relationships over individual object identities, especially in object placement scenarios. To further enhance robustness, we apply data augmentation techniques using LLMs. For instance, in the instruction \"place the banana " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": " above the apple,\" we generalize it to \"place object A " + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 449, + 555, + 581 + ], + "type": "text", + "content": " above object B,\" allowing the model to prioritize spatial relationships while ignoring specific object identities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": "Inspired by SimCSE [11], let " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "S = [I_n]_{n=1}^N" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": " be the set of original instructions, and " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "S' = [I_n, I_n']_{n=1}^N" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": " be the set after data augmentation, where " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "I_n'" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": " is the augmented instruction derived from " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "I_n" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": ". We use a pre-trained BERT model " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": " to obtain the embeddings of the instructions. For each instruction " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "I_n" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": ", we define its unique positive pair as " + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "inline_equation", + "content": "(I_n, I_n')" + }, + { + "bbox": [ + 313, + 582, + 556, + 678 + ], + "type": "text", + "content": ", and all other pairs are treated as negative samples. The similarity between the two embeddings is calculated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 689, + 555, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 689, + 555, + 718 + ], + "spans": [ + { + "bbox": [ + 345, + 689, + 555, + 718 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {m}\\right)\\right) = \\frac {\\psi \\left(I _ {n}\\right) ^ {T} \\psi \\left(I _ {m}\\right)}{\\| \\psi \\left(I _ {n}\\right) \\| _ {2} \\| \\psi \\left(I _ {m}\\right) \\| _ {2}} \\tag {1}", + "image_path": "b053ea10ad81f7aca3609ec6e963ef889e3a08277f2d337368ca13bb62411b55.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19417" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 229, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 229, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 229, + 83 + ], + "type": "text", + "content": "The contrastive loss function is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 295, + 119 + ], + "type": "interline_equation", + "content": "\\ell_ {n} ^ {N} = - \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right) , \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right)\\right)}{\\sum_ {m = 1 , m \\neq n} ^ {2 N} \\exp \\left(\\operatorname {s i m} \\left(\\psi \\left(I _ {n}\\right) , \\psi \\left(I _ {m}\\right)\\right)\\right)} \\tag {2}", + "image_path": "8c572b034e1e3c28ce95aadf811effb639f4177cab4553a05d7ab1805fa2afc7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 125, + 295, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 295, + 147 + ], + "type": "text", + "content": "The overall contrastive loss is the average of all individual losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 147, + 295, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 147, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 123, + 147, + 295, + 179 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {I n s t a n c e - C L}} = \\frac {1}{2 N} \\sum_ {n = 1} ^ {2 N} \\ell_ {n} ^ {N} \\tag {3}", + "image_path": "a5aa28f18cbc1a839d64740b1b4a3492d1e543d9d194f6276f53ec30373a4f75.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 196, + 295, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 220 + ], + "type": "text", + "content": "Algorithm 1 Training Procedure for Multi-task Instance-wise Contrastive Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 224, + 295, + 357 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 224, + 295, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 295, + 248 + ], + "type": "text", + "content": "1: Input: Instruction set " + }, + { + "bbox": [ + 61, + 224, + 295, + 248 + ], + "type": "inline_equation", + "content": "S = \\{S_{\\mathrm{image}}, S_{\\mathrm{object}}\\}" + }, + { + "bbox": [ + 61, + 224, + 295, + 248 + ], + "type": "text", + "content": " containing tasks for image generation and object manipulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 249, + 171, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 249, + 171, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 171, + 259 + ], + "type": "text", + "content": "2: Output: Trained model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 261, + 153, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 153, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 153, + 271 + ], + "type": "text", + "content": "3: for each " + }, + { + "bbox": [ + 61, + 261, + 153, + 271 + ], + "type": "inline_equation", + "content": "I_{n} \\in S" + }, + { + "bbox": [ + 61, + 261, + 153, + 271 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 272, + 277, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 277, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 277, + 285 + ], + "type": "text", + "content": "4: Generate augmented instruction " + }, + { + "bbox": [ + 61, + 272, + 277, + 285 + ], + "type": "inline_equation", + "content": "I_{n}^{\\prime}" + }, + { + "bbox": [ + 61, + 272, + 277, + 285 + ], + "type": "text", + "content": " using LLMs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 285, + 107, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 107, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 107, + 294 + ], + "type": "text", + "content": "5: end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "type": "text", + "content": "6: Obtain embeddings " + }, + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\psi(I_n)" + }, + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\psi(I_n')" + }, + { + "bbox": [ + 61, + 297, + 295, + 331 + ], + "type": "text", + "content": " for all instructions using task-specific embedding heads in the pretrained BERT model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 332, + 171, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 332, + 171, + 344 + ], + "spans": [ + { + "bbox": [ + 61, + 332, + 171, + 344 + ], + "type": "text", + "content": "7: for each " + }, + { + "bbox": [ + 61, + 332, + 171, + 344 + ], + "type": "inline_equation", + "content": "I_{n} \\in S_{\\mathrm{image}}" + }, + { + "bbox": [ + 61, + 332, + 171, + 344 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 345, + 274, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 345, + 274, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 345, + 274, + 357 + ], + "type": "text", + "content": "8: Calculate the image generation contrastive loss:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 124, + 364, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 364, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 124, + 364, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\ell_ {n} ^ {\\text {I m a g e - C L}} = \\operatorname {L o s s} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right) \\tag {4}", + "image_path": "46648187f7d375f4a45316af9c18e7a2bf34df91f38106d22d161f7a9410dd59.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 399, + 294, + 448 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 60, + 399, + 276, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 399, + 276, + 412 + ], + "spans": [ + { + "bbox": [ + 60, + 399, + 276, + 412 + ], + "type": "text", + "content": "9: Update model parameters to minimize " + }, + { + "bbox": [ + 60, + 399, + 276, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Image - CL}}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 413, + 107, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 413, + 107, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 413, + 107, + 422 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 424, + 170, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 424, + 170, + 436 + ], + "spans": [ + { + "bbox": [ + 59, + 424, + 170, + 436 + ], + "type": "text", + "content": "11: for each " + }, + { + "bbox": [ + 59, + 424, + 170, + 436 + ], + "type": "inline_equation", + "content": "I_{n} \\in S_{\\mathrm{object}}" + }, + { + "bbox": [ + 59, + 424, + 170, + 436 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 58, + 437, + 294, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 437, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 58, + 437, + 294, + 448 + ], + "type": "text", + "content": "12: Calculate the object manipulation contrastive loss:" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 123, + 468, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 468, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 123, + 468, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\ell_ {n} ^ {\\text {O b j e c t - C L}} = \\operatorname {L o s s} \\left(\\psi \\left(I _ {n}\\right), \\psi \\left(I _ {n} ^ {\\prime}\\right)\\right) \\tag {5}", + "image_path": "66aef189154bdaa9b4a3fac769f24cd63b68c73cea0ae5d81cffdba56a2ab105.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 58, + 503, + 277, + 540 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 58, + 503, + 277, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 503, + 277, + 516 + ], + "spans": [ + { + "bbox": [ + 58, + 503, + 277, + 516 + ], + "type": "text", + "content": "13: Update model parameters to minimize " + }, + { + "bbox": [ + 58, + 503, + 277, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Object - CL}}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 517, + 107, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 107, + 526 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 107, + 526 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 58, + 528, + 266, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 528, + 266, + 540 + ], + "spans": [ + { + "bbox": [ + 58, + 528, + 266, + 540 + ], + "type": "text", + "content": "15: Compute the overall multi-task contrastive loss:" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 104, + 548, + 295, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 548, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 548, + 295, + 619 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {M u l t i - T a s k - C L}} = \\lambda_ {\\text {i m a g e}} \\sum_ {n = 1} ^ {| S _ {\\text {i m a g e}} |} \\ell_ {n} ^ {\\text {I m a g e - C L}} \\\\ + \\lambda_ {\\text {o b j e c t}} \\sum_ {n = 1} ^ {| S _ {\\text {o b j e c t}} |} \\ell_ {n} ^ {\\text {O b j e c t - C L}} \\tag {6} \\\\ \\end{array}", + "image_path": "c036d60f738ff89c0e6f28a885ca64a191e79e751027225b7e40b8fd77f7e017.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 651, + 173, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 173, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 173, + 662 + ], + "type": "text", + "content": "3.4.2. Multi-Task Learning" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The existing learning algorithm has limitations when handling different tasks, particularly in multi-task scenarios such as image generation and object manipulation. Since each task focuses on different key information, existing" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "models struggle to leverage and learn both the commonalities and differences between tasks, thereby limiting their generalization ability. Furthermore, when multiple complex tasks coexist, using separate models for each task often fails to exploit complementary information among tasks, resulting in low learning efficiency and suboptimal model performance. To improve adaptability, we design a multi-task learning strategy [5] that trains the embedding model on both 3D object placement and 2D layout tasks, enhancing generalization across spatial, temporal, and categorical relationships." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": "Algorithm 1 outlines the training procedure for Multi-task Instance-wise Contrastive Learning. The instruction set " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", containing tasks for image generation and object manipulation, is processed through the following steps. First, for each instruction " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", an augmented instruction " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "I_{n}^{\\prime}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " is generated using LLMs (Lines 3-5). These instructions, along with their original versions, are then passed through task-specific embedding heads in a pre-trained BERT model to obtain their corresponding embeddings " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\psi(I_{n})" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\psi(I_{n}^{\\prime})" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " (Line 6). For the image generation tasks " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "S_{\\text{image}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", the model calculates the image generation contrastive loss " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\ell_{\\text{image-CL}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " (Lines 7-9). Similarly, for the object manipulation tasks " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "S_{\\text{object}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": ", the object manipulation contrastive loss " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\ell_{\\text{object-CL}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " is computed (Lines 11-13). Finally, the overall multi-task contrastive loss " + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{Multi-Task-CL}}" + }, + { + "bbox": [ + 313, + 205, + 555, + 421 + ], + "type": "text", + "content": " is computed by combining the individual losses from both tasks (Line 15). The model parameters, including those of the task-specific embedding heads, are updated to minimize the loss (Lines 9 and 13)." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 431, + 416, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 416, + 444 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 416, + 444 + ], + "type": "text", + "content": "3.5. Layout Planning" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "The Layout Planning module is responsible for determining the target positions and orientations of objects based on high-level user instructions, adhering to a unified problem setting for layout generation. The process integrates User Instructions and Spatial Knowledge into a comprehensive prompt for both object rearrangement and image generation tasks, with Robot Observations added in the object rearrangement task to assist in identifying relevant objects and their attributes. For object rearrangement, during the object selection phase, the system uses the Object Selection LLM to identify relevant objects and determine their attributes as " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "(id, c, w, h, d)" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": ". In the layout generation phase, the Layout Generation LLM calculates the spatial configuration of each object, inferring the 3D coordinates " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " and rotation angle " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " while using known attributes such as width, height, depth " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "(w, h, d)" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": ". By assigning precise poses " + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "(x, y, z, \\theta)" + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": " to the objects, the Layout Planning module completes the layout generation based on the unified problem setting, allowing the process to move into the Simulation Phase, where the objects are arranged according to the determined layout. The image generation task follows a similar approach, where the system also generates object" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19418" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 91, + 507, + 218 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 465, + 82 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 465, + 82 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 465, + 82 + ], + "type": "text", + "content": "Table 1. Success rates for different models across task types and subtasks in SK-Dataset" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 91, + 507, + 218 + ], + "lines": [ + { + "bbox": [ + 100, + 91, + 507, + 218 + ], + "spans": [ + { + "bbox": [ + 100, + 91, + 507, + 218 + ], + "type": "table", + "html": "
ModelSuccess Rate (%)
Task TypeSimple TaskComplex Task
Sub Task Type-LettersGeometric ShapesSemantic Geometric Shapes
GPT3.5SKE-Layout (ours)42.211.439.45.8
w/o CL35.210.940.64.3
LLM-GROP26.36.124.73.8
GPT4SKE-Layout (ours)75.142.772.621.7
w/o CL71.541.369.219.2
LLM-GROP69.49.418.54.5
", + "image_path": "c70643a23548d44e3a0c9b571ef2e2ac5a7fad6de97e21095d8d9918a09c9b1b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 88, + 247, + 520, + 407 + ], + "blocks": [ + { + "bbox": [ + 211, + 228, + 399, + 239 + ], + "lines": [ + { + "bbox": [ + 211, + 228, + 399, + 239 + ], + "spans": [ + { + "bbox": [ + 211, + 228, + 399, + 239 + ], + "type": "text", + "content": "Table 2. Evaluation of methods on NSR-1K dataset" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 247, + 520, + 407 + ], + "lines": [ + { + "bbox": [ + 88, + 247, + 520, + 407 + ], + "spans": [ + { + "bbox": [ + 88, + 247, + 520, + 407 + ], + "type": "table", + "html": "
MethodsNumerical ReasoningSpatial Reasoning
Layout Eval.Image Eval.Layout Eval.Image Eval.
PrecisionRecallAccuracyAcc. (GLIP)AccuracyAcc. (GLIP)
Text → Image
Stable Diffusion (v2.1)---42.44-17.81
Attend-and-Excite (SD v2.1)---45.74-26.86
Text → Layout → Image
LayoutTransformer [17]75.7061.6922.2640.556.3628.13
LayoutGPT (GPT-4)78.3686.2978.4355.6491.7360.64
w/o CL86.0897.4695.2856.8694.3561.13
w/o ML85.2398.8097.1158.6894.7061.48
SKE-Layout (ours)85.2899.0697.2459.7095.0562.19
GT layouts100.00100.00100.0053.23100.0062.54
Human99.2696.5292.5656.0791.1751.94
", + "image_path": "1326496d3185d63c4117e599d0bf55de5c46b47a926f9699493bd36ac162fb20.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 269, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 269, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 269, + 439 + ], + "type": "text", + "content": "Layouts " + }, + { + "bbox": [ + 55, + 426, + 269, + 439 + ], + "type": "inline_equation", + "content": "(id, c, w, d, x, y)" + }, + { + "bbox": [ + 55, + 426, + 269, + 439 + ], + "type": "text", + "content": " based on task requirements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 452, + 160, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 160, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 160, + 464 + ], + "type": "text", + "content": "3.6. Scene Simulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 471, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 471, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 471, + 295, + 602 + ], + "type": "text", + "content": "For object rearrangement task, we input the objects and their coordinates into the PyBullet simulation environment to create a 3D scene. This allows us to observe and evaluate the arrangement of objects in a simulated physical space, ensuring that the generated layouts are feasible and accurate in a realistic setting. For the image generation task, we use the layout-to-image generation model, GLIGEN. The model generates images based on the object layouts, ensuring that the categories, locations, and sizes of the objects in the generated image precisely match the predicted layouts described in prior works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 619, + 137, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 137, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 137, + 633 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "In this section, we present extensive empirical results of SKE-Layout and several competitive baseline methods. From the representative spatial reasoning-induced domains, including both robotic object rearrangement and image generation tasks, we observe that SKE-Layout can significantly improve task success rate compared to all baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 430, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 430, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 430, + 440 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 383, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 383, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 383, + 455 + ], + "type": "text", + "content": "4.1.1. Baselines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 520 + ], + "type": "text", + "content": "Given the two distinct tasks in our experiments, it is imperative to employ different baseline models for a comprehensive comparison. We select a range of competitive baselines that represent the state-of-the-art approaches in both image generation tasks and robotic object rearrangement tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 604 + ], + "type": "text", + "content": "For the Robotic Object Rearrangement Task, we utilize StructFormer [24], a Transformer-based model designed for capturing spatial relationships, and LLM-GROP [8] that integrates LLMs with planning techniques. We also include the SKE-Layout variant trained without contrastive learning (w/o CL), which uses BERT for processing, to evaluate object arrangement tasks without contrastive loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "For the Image Generation Task, we select several baselines: Stable Diffusion [33], an end-to-end diffusion-based text-to-image model; Attend-and-Excite, which uses attention mechanisms for detail-oriented image generation; LayoutTransformer [13], a text-to-query model for layout generation; and LayoutGPT, [9], which leverages LLMs for text-to-query generation. We also include our SKE-Layout model, along with its variants trained without contrastive learning (w/o CL) or multitask learning (w/o ML) to assess" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19419" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "type": "text", + "content": "their individual contributions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 88, + 120, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 88, + 120, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 88, + 120, + 99 + ], + "type": "text", + "content": "4.1.2. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 103, + 295, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 295, + 151 + ], + "type": "text", + "content": "We evaluate the performance of our proposed method on three different datasets, each tailored to a specific task, including image generation and robotic object rearrangement, to ensure a comprehensive assessment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 152, + 295, + 440 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 152, + 295, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 212 + ], + "type": "text", + "content": "- NSR-1K Dataset: Derived from the LayoutGPT framework, it includes both template-based and human-written prompts sourced from MSCOCO, designed to evaluate image generation based on specified object counts and spatial locations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 212, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 319 + ], + "type": "text", + "content": "- StructFormer Dataset: Designed specifically for the robotic object rearrangement task, this dataset includes scenarios where objects must be arranged based on given constraints (e.g., circle, line, tower, table setting), testing the model's ability to maintain structural coherence and execute tasks accurately. We do not use the training dataset from StructFormer, which serves as an Out-Of-Distribution (OOD) test for evaluating the model's generalization to unseen tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 440 + ], + "type": "text", + "content": "- SK-Dataset: Created by us to evaluate complex object rearrangement tasks, the SK-Dataset includes advanced scenarios with multi-object planning and execution. Tasks are divided into two levels: simple and complex. The simple tasks focus on accurately positioning objects based on specified distances and positions. The complex tasks require higher-order reasoning and are further divided into letters, geometric shapes, and semantic geometric shapes, emphasizing spatial precision, symmetry, and meaningful arrangements." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 445, + 189, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 189, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 189, + 456 + ], + "type": "text", + "content": "4.2. Results and Discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": "SK Dataset. Table 1 compares the success rates of different models across SK-Dataset task types, including GPT-3.5 and GPT-4 using our method, w/o CL variant, and LLM-GROP. Overall, our method achieves consistently high success rates, with the degree of improvement varying by task complexity. For example, in simple tasks with GPT-4, our method reached a " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": " success rate, outperforming w/o CL " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "(71.5\\%)" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": " and LLM-GROP " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "(69.4\\%)" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": ". In contrast, for complex tasks, our approach shows a more pronounced advantage, increasing the success rate from " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": " (LLM-GROP) to " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "21.7\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": ". Two key observations emerge from Table 1:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "- While improvements in simple tasks are moderate, our method significantly outperforms baselines in complex tasks, particularly in handling semantic geometric shapes, demonstrating its ability to enhance comprehension and reasoning in high-difficulty scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "- GPT-4 consistently outperforms GPT-3.5, especially when combined with our method. This highlights the effectiveness of our joint training strategy in enhancing the model's understanding of complex information and its adaptability across different task difficulties." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "StructFormer Dataset. Fig. 2 compares model performance across four object rearrangement tasks: Tower, Line, Circle, and Set the Table. Our model consistently outperforms StructFormer, with improvements of " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", respectively. This demonstrates our method's advantage in managing tasks by effectively integrating external knowledge, enhancing both accuracy and reliability. Several key insights emerge from these results:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 313, + 168, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 204 + ], + "type": "text", + "content": "- The improvement is more significant in complex tasks, indicating that our approach excels in scenarios requiring higher levels of reasoning and understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "content": "- Unlike StructFormer, our method maintains strong performance across all tasks without separate task-specific models, showcasing its versatility and adaptability to diverse object rearrangement tasks." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 253, + 551, + 389 + ], + "blocks": [ + { + "bbox": [ + 317, + 253, + 551, + 389 + ], + "lines": [ + { + "bbox": [ + 317, + 253, + 551, + 389 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 551, + 389 + ], + "type": "image", + "image_path": "2e687dda496510f58934e689b52922b53232af894613b54a0453062be902d9ac.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 393, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 416 + ], + "type": "text", + "content": "Figure 2. Evaluation of methods on the StructFormer dataset on various tasks: Tower, Line, Circle, and Set the Table." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "content": "NSR-1K Dataset. Table 2 presents the main results of image generation, showing that our method significantly outperforms comparative models across multiple evaluation metrics. In numerical reasoning, it exceeds LayoutGPT (GPT-4) by " + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "inline_equation", + "content": "10\\% - 20\\%" + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "content": " in precision, recall, and accuracy. In spatial reasoning, our model achieves " + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "inline_equation", + "content": "95.05\\%" + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "content": " layout accuracy and " + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "inline_equation", + "content": "61.48\\%" + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "content": " GLIP accuracy, outperforming LayoutGPT by " + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 427, + 555, + 571 + ], + "type": "text", + "content": ", respectively. This improvement is attributed to the more comprehensive spatial knowledge of our method, which allows it to retrieve more relevant spatial information, thereby providing higher precision in both layout and image generation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 576, + 408, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 408, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 408, + 589 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Q1: What happens when we remove contrastive learning or multitask learning from the model? The performance on both the NSR-1K and SK datasets decreases when contrastive learning or multitask learning is disabled. This indicates that these two techniques are crucial. Specifically, contrastive learning helps the model better distinguish subtle spatial relationships, while multitask learning improves the model's ability to understand context across different tasks. Together, they contribute to more effective retrieval of spatial knowledge." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19420" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 70, + 518, + 322 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 518, + 322 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 518, + 322 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 518, + 322 + ], + "type": "image", + "image_path": "ca2dceb4a6058f1ccbf66dbd78e508df28bf1a5e0faf378a1efebbd71aa863dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 328, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 328, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 555, + 361 + ], + "type": "text", + "content": "Figure 3. Case Study. We present a comparison of different methods in 3D object rearrangement and 2D image generation tasks. (a) arranging objects on a table with correct spatial relationships, (b) placing blocks in a specified formation, (c) generating a 2D image with objects positioned according to a textual description." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 469 + ], + "type": "text", + "content": "Q2: Why doesn't the StructFormer dataset show similar performance drops? The performance differences among our model variants on the StructFormer dataset are minimal. This is likely because the spatial relationships required in this dataset are relatively straightforward, making precise retrieval of complex spatial knowledge less critical. As a result, the advantages of contrastive learning and multitask learning become less pronounced." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 132, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 132, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 132, + 492 + ], + "type": "text", + "content": "4.4. Case Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 296, + 713 + ], + "type": "text", + "content": "We present three representative cases to demonstrate the effectiveness of our SKE-Layout framework across both 3D object rearrangement and 2D image generation tasks. Fig. 3.(a) involves setting the table and arranging objects into a smiley face pattern. While the LLM-GROP model fails to maintain correct angles and semantic coherence, our SKE-Layout significantly reduces errors, providing accurate placement and alignment. In Fig. 3.(b), where blocks need to be arranged into a circular pattern, our method achieves perfect spatial precision, outperforming both StructFormer and LLM-GROP, which struggle with alignment. For the 2D image generation task in Fig. 3.(c), our method also excels. While Stable Diffusion generates quantity errors and LayoutGPT shows positional errors, SKE-Layout accurately captures object relationships, ensuring correct quantity and positioning and thus closely matching the given description. From Fig. 3, we draw the following observations:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 373, + 555, + 553 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 373, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 555, + 480 + ], + "type": "text", + "content": "- Richer Spatial Knowledge. Our SKE layout framework demonstrates a deeper understanding of spatial relationships than baseline models. By integrating diverse spatial knowledge sources, including real-world and synthetic data, our approach enhances the richness and depth of spatial reasoning. Additionally, it employs multiple forms of spatial knowledge representation, which contribute to more accurate positioning and alignment of objects in both 3D rearrangement and 2D image generation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 481, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 555, + 553 + ], + "type": "text", + "content": "- Adaptability to Multi-Task Settings. The SKE layout framework excels in handling a variety of tasks, showcasing versatility in both 3D object arrangement and 2D image generation. This adaptability highlights its robustness and effectiveness in addressing complex spatial reasoning challenges across multiple domains." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we present SKE.layout, a novel method designed to effectively generate layouts by leveraging large language models. Our approach injects spatial knowledge into the large language model, thereby demonstrating its versatility and robustness across different application domains. Empirically, we have demonstrated that SKE.layout performs exceptionally well in both image generation and robotic task arrangement tasks. Future research directions include exploring how to apply SKE.layout to more application scenarios, such as open-domain scenarios, where the set of objects in a problem instance is not known in advance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19421" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 255, + 115, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 115, + 268 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 115, + 268 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 275, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 275, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 275, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 61, + 275, + 296, + 330 + ], + "type": "text", + "content": "[1] Dhruv Batra, Angel X Chang, Sonia Chernova, Andrew J Davison, Jia Deng, Vladlen Koltun, Sergey Levine, Jitendra Malik, Igor Mordatch, Roozbeh Mottaghi, et al. Rearrangement: A challenge for embodied ai. arXiv preprint arXiv:2011.01975, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 331, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 295, + 353 + ], + "type": "text", + "content": "[2] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446):eaat8414, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 354, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 376 + ], + "type": "text", + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 377, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 295, + 422 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 443 + ], + "type": "text", + "content": "[5] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 487 + ], + "type": "text", + "content": "[6] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 521 + ], + "type": "text", + "content": "[7] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 524, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 295, + 578 + ], + "type": "text", + "content": "[8] Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. Task and motion planning with large language models for object rearrangement. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2086-2092. IEEE, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 295, + 635 + ], + "type": "text", + "content": "[9] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "text", + "content": "[10] Hanan Gani, Shariq Farooq Bhat, Muzammal Naseer, Salman Khan, and Peter Wonka. LIm blueprint: Enabling text-to-image generation with complex and detailed prompts. arXiv preprint arXiv:2310.10640, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[11] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. arXiv preprint arXiv:2104.08821, 2021. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 90, + 297, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 244 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China under Grant 62322601 and 42174050, the Independent Research Project of State Key Laboratory of Mechanical Transmission for Advanced Equipment under Grant SKLMT-ZZKT2024R07, the Fundamental Research Funds for the Central Universities 2024IAIS-QN01, the Postdoctoral Fellowship Program of CPSF under Grant GZB20230912, the China Postdoctoral Science Foundation under Grant 2024M753837, the Chongqing Postdoctoral Innovation Talents Support Program under Grant CQBX202324 and the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQJQX0025 and CSTB2024NSCQMSX0955." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 184 + ], + "type": "text", + "content": "[13] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[14] Muhammad Usman Hadi, Rizwan Qureshi, Abbas Shah, Muhammad Irfan, Anas Zafar, Muhammad Bilal Shaikh, Naveed Akhtar, Jia Wu, Seyedali Mirjalili, et al. A survey on large language models: Applications, challenges, limitations, and practical usage. Authorea Preprints, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[15] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[16] Akash Abdu Jyothi, Thibaut Durand, Jiawei He, Leonid Sigal, and Greg Mori. Layout: Stochastic scene layout generation from a label set. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9895-9904, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[17] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. arXiv preprint arXiv:2309.10062, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[18] Ivan Kapelyukh, Vitalis Vosylius, and Edward Johns. Dallebot: Introducing web-scale diffusion models to robotics. IEEE Robotics and Automation Letters, 8(7):3956-3963, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[19] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[20] Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "type": "text", + "content": "[21] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. Layoutgan: Generating graphic layouts with wireframe discriminators. arXiv preprint arXiv:1901.06767, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 634 + ], + "type": "text", + "content": "[22] Jiachen Li, Qiaozi Gao, Michael Johnston, Xiaofeng Gao, Xuehai He, Suhaila Shakiah, Hangjie Shi, Reza Ghanadan, and William Yang Wang. Mastering robot manipulation with multimodal prompts through pretraining and multi-task finetuning. arXiv preprint arXiv:2310.09676, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "type": "text", + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Weiyu Liu, Chris Paxton, Tucker Hermans, and Dieter Fox. Structformer: Learning spatial structure for language-guided" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19422" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "semantic rearrangement of novel objects. In 2022 International Conference on Robotics and Automation (ICRA), pages 6322-6329. IEEE, 2022. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[25] Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering, 35(1):857-876, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[26] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters (RA-L), 7(3):7327-7334, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[27] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and William Yang Wang. Automatically correcting large language models: Surveying the landscape of diverse self-correction strategies. arXiv preprint arXiv:2308.03188, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[28] Ruoling Peng, Kang Liu, Po Yang, Zhipeng Yuan, and Shunbao Li. Embedding-based retrieval with llm for effective agriculture information extracting from unstructured data. arXiv preprint arXiv:2308.03107, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "text", + "content": "[29] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7942, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "text", + "content": "[30] Shuhan Qi, Zhengying Cao, Jun Rao, Lei Wang, Jing Xiao, and Xuan Wang. What is the limitation of multimodal llms? a deeper look into multimodal llms through prompt probing. Information Processing & Management, 60(6):103510, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "text", + "content": "[32] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Suenderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable task planning. CoRR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[34] Sai Sabitha, Deepti Mehrotra, and Abhay Bansal. Knowledge enriched learning by converging knowledge object&learning object. Electronic Journal of e-Learning, 13 (1):pp3-13, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "text", + "content": "[35] Andrew Szot, Karmesh Yadav, Alex Clegg, Vincent-Pierre Berges, Aaron Gokaslan, Angel Chang, Manolis Savva, Zsolt Kira, and Dhruv Batra. Habitat rearrangement challenge 2022. https://aihabitat.org/challenge/rearrange_2022, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Jiaqi Wang, Zihao Wu, Yiwei Li, Hanqi Jiang, Peng Shu, Enze Shi, Huawen Hu, Chong Ma, Yiheng Liu, Xuhui Wang, et al. Large language models for robotics: Opportunities, challenges, and perspectives. arXiv preprint arXiv:2401.04334, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 396 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[38] Luca Weihs, Matt Deitke, Aniruddha Kembhavi, and Roozbeh Mottaghi. Visual room rearrangement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[39] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "type": "text", + "content": "[40] Guangyao Zhai, Xiaoni Cai, Dianye Huang, Yan Di, Fabian Manhardt, Federico Tombari, Nassir Navab, and Benjamin Busam. Sg-bot: Object rearrangement via coarse-to-fine robotic imagination on scene graphs. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4303-4310. IEEE, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "text", + "content": "[41] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology, 15(2):1-38, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 341, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 396 + ], + "type": "text", + "content": "[42] Shanshan Zhong, Zhongzhan Huang, Weushao Wen, Jinghui Qin, and Liang Lin. Sur-adapter: Enhancing text-to-image pre-trained diffusion models with large language models. In Proceedings of the 31st ACM International Conference on Multimedia, pages 567-578, 2023. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19423" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_content_list.json b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a9bb2d5ce503c86458124fd78978930107157ab4 --- /dev/null +++ b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_content_list.json @@ -0,0 +1,1540 @@ +[ + { + "type": "text", + "text": "SLADE: Shielding against Dual Exploits in Large Vision-Language Models", + "text_level": 1, + "bbox": [ + 119, + 128, + 877, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Md Zarif Hossain, Ahmed Imteaj†", + "bbox": [ + 362, + 180, + 635, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computing, Southern Illinois University Carbondale, IL 62901, USA", + "bbox": [ + 184, + 200, + 812, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 252, + 326, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Vision-Language Models (LVLMs) have emerged as transformative tools in multimodal tasks, seamlessly integrating pretrained vision encoders to align visual and textual modalities. Prior works have highlighted the susceptibility of LVLMs to dual exploits (gradient-based and optimization-based jailbreak attacks), which leverage the expanded attack surface introduced by the image modality. Despite advancements in enhancing robustness, existing methods fall short in their ability to defend against dual exploits while preserving fine-grained semantic details and overall semantic coherence under intense adversarial perturbations. To bridge this gap, we introduce SheiLding against Dual Exploits (SLADE), a novel unsupervised adversarial fine-tuning scheme that enhances the resilience of CLIP-based vision encoders. SLADE's dual-level contrastive learning approach balances the granular and the holistic, capturing fine-grained image details without losing sight of high-level semantic coherence. Extensive experiments demonstrate that SLADE-equipped LVLMs set a new benchmark for robustness against dual exploits while preserving fine-grained semantic details of perturbed images. Notably, SLADE achieves these results without compromising the core functionalities of LVLMs, such as instruction following, or requiring the computational overhead (e.g., large batch sizes, momentum encoders) commonly associated with traditional contrastive learning methods.", + "bbox": [ + 91, + 284, + 480, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 708, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The remarkable success of large language models (LLMs) [40, 45] in understanding and generating human-like text has inspired researchers to extend their capabilities to the visual domain. This effort has led to the development of Large Vision-Language Models (LVLMs) [1, 34], which aim to bridge the gap between visual and textual information. These LVLMs frequently utilize pretrained vision encoders, such as CLIP [48], BEiT [7], and DINO [10], to derive meaningful visual representations from images. LVLMs [2, 34] equipped with these vision encoders demonstrate remarkable zero-shot performance in down-", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stream tasks such as image captioning, and visual question answering (VQA). Their promising capabilities and impact on multimodal learning have led to a growing number of LVLMs being publicly released. However, while this availability fosters research and application development, it also introduces critical safety concerns that require careful attention. Recent studies [57, 63] highlight the high susceptibility of LVLMs to harmful exploits on text and image modalities. Notably, the addition of a vision modality increases the model's attack surface, making LVLMs even more vulnerable to manipulation compared to LLMs [9, 22]. Moreover, the adversaries can use imperceptible gradient-based perturbations to create targeted attacks, generating malicious outputs that evade detection [49]. Such targeted attacks are particularly concerning due to their ability to manipulate models toward specific, often harmful outputs. Additionally, the increasing deployment of LVLMs raises the potential for exploitation not only by adversaries but also by corporations aiming to leverage these models to promote their products or agendas as shown in Figure 1. Recent studies [43, 47] demonstrate that gradient-based perturbations can be adapted in optimization-based jailbreak attacks to bypass a model's safety mechanisms, causing LVLMs to generate harmful or undesirable content.", + "bbox": [ + 511, + 253, + 906, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While existing methods [38, 50] made notable progress in improving the robustness of vision encoders, particularly CLIP, which is widely utilized across various applications, critical challenges persist. For example, in downstream tasks such as image captioning, these models successfully capture high-level semantic meaning but often struggle to preserve fine-grained semantic details when faced with perturbed images. Consequently, their ability to generate accurate and contextually rich descriptions is compromised, limiting their applicability and effectiveness in real-world scenarios where precision and detail are paramount. To address these concerns, we propose SLADE, an unsupervised adversarial fine-tuning scheme that enhances the robustness of the widely used CLIP vision encoder's embeddings against dual exploits: gradient-based attacks and optimization-based Jailbreak attacks. SLADE combines symmetric contrastive learning objectives at both the patch level and the global level (on full image) to capture fine", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24244", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f934b1330ff6c288914c75eff374fe6f7e8a7ebad2f6cab23d346de00e77f5d7.jpg", + "image_caption": [ + "Figure 1. Targeted $\\ell_{\\infty}$ attack and VisualAdV [47] jailbreak attack at $\\epsilon = 4 / 255$ radii and using original CLIP and SLADE as vision encoder in LLaVA-7B. SLADE (trained at $\\epsilon = 4 / 255$ ) generates captions with fine-grained semantic details for the perturbed sample and successfully defends against the jailbreak attempt." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 862, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "grained image details and high-level semantic representation of the entire image, respectively. This dual-level approach enables the model to preserve both detailed features and high-level semantic consistency, even in the presence of adversarial perturbations. Moreover, SLADE does not require LVLM or task-specific training and can be seamlessly integrated into existing LVLMs as a robust vision encoder. SLADE enhances the adversarial robustness of the vision modality in LVLMs, preserving the model's ability to capture nuanced semantic details while effectively defending against jailbreak attacks, as illustrated in Figure 1. Our work sets a new standard for adversarial robustness in LVLMs, a critical advancement as these models become integral to critical applications.", + "bbox": [ + 88, + 375, + 480, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 593, + 240, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Vision-Language Models. In recent times, several advanced LVLMs have been introduced, including LLaVA [34], OpenFlamingo (OF) [2], InstructBLIP [17] and Macaw-LLM [36]. These models leverage the power of multimodal learning by combining visual and textual data, thereby, enhancing their capability to understand and generate responses to multimodal inputs. Most of these LVLMs incorporate pre-trained LLMs, such as Llama [53], paired with large-scale vision encoders like CLIP [48], to process multimodal data.", + "bbox": [ + 89, + 618, + 483, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial and Jailbreaking attacks in LVLMs. In recent studies [3, 4, 19, 63], authors have explored the susceptibility of large vision-language models to adversarial attacks. Notably, recent works [3, 6, 27, 64] demonstrate the vulnerability of vision encoders in LVLMs, particularly CLIP, to adversarial manipulation, which can significantly impair the performance of LVLMs in downstream tasks. Moreover, [16, 49] highlight the efficacy of gradient-", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based attacks on LVLMs, demonstrating their capacity to induce the model into generating harmful or misleading outputs. In [43, 47] authors show how gradient-based perturbations can be leveraged in optimization-based jailbreak attacks to bypass LVLMs' safety protocols. While recent studies have proposed defense strategies [58, 62] to counter jailbreak attacks, these methods rely heavily on the model's inherent safety alignment, necessitating specific training for each LVLM and often resulting in degraded performance in downstream tasks (e.g., VQA and image captioning).", + "bbox": [ + 511, + 375, + 906, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robustness against exploits in LVLMs. Prior works [32, 38] introduced supervised adversarial fine-tuning scheme for CLIP, using cross-modal image-text contrastive loss. However, in LVLMs their zero-shot performance in downstream tasks (mostly in unseen classes) often degrades due to an over-dependence on the text embeddings during the fine-tuning process. To address this, some concurrent works [20, 23, 27] proposed unsupervised adversarial training methods using SimCLR [12] and BYOL [24] contrastive loss, which improve robustness but require large batch sizes for sufficient negative samples or a momentum encoder for optimal results. Additionally, the authors in [50] proposed an alternative approach, tailoring an $\\ell_2$ loss-based unsupervised fine-tuning scheme. However, our empirical analysis reveals that applying $\\ell_2$ loss across the entire image is insufficient for retaining nuanced details, particularly under adversarial conditions. Unlike prior methods, SLADE integrates a symmetric contrastive learning objective applied simultaneously at the patch level and global level. This dual-level strategy enables SLADE to capture fine-grained details and nuanced spatial relationships, even under adversarial perturbations. Drawing inspiration from traditional machine learning techniques that use patch embedding with pretext tasks to preserve detail and spatial", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24245", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "localization [11, 41, 44, 51], SLADE advances this concept by incorporating it directly into its adversarial fin-tuning mechanism. Furthermore, SLADE leverages a stop-gradient mechanism to bypass the reliance on large batches or momentum encoders, offering a scalable and computationally efficient solution.", + "bbox": [ + 89, + 90, + 482, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Overview of SLADE", + "text_level": 1, + "bbox": [ + 89, + 195, + 287, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides a detailed overview of SLADE, our proposed adversarial fine-tuning scheme, illustrated in Figure 2. The primary objective of SLADE is to adversarially fine-tune a CLIP-like model $f_{\\theta}(\\cdot)$ , with parameter $\\theta$ on a given dataset $\\mathcal{D} = \\{I_1, I_2, \\dots, I_N\\}$ with $I_n \\in \\mathbb{R}^{H \\times W \\times 3}$ . SLADE is designed to learn invariant features that are robust to adversarial perturbations while preserving the semantic details of the input data. This ensures that the fine-tuned model effectively balances adversarial resilience with the retention of critical semantic information.", + "bbox": [ + 89, + 220, + 483, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75bfc60eac1559218dd8d04e2d43b374b57dc9b5656653b2eebb7d7c948e4492.jpg", + "image_caption": [ + "Figure 2. Workflow of our proposed scheme: SLADE." + ], + "image_footnote": [], + "bbox": [ + 94, + 382, + 483, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Adversarial Image generation", + "text_level": 1, + "bbox": [ + 89, + 609, + 356, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During adversarial fine-tuning, SLADE generates a perturbed version $I^{\\prime}$ from the original clean image $I$ using Projected Gradient Descent (PGD) [37]. Given an input image $I$ , the perturbed image $I^{\\prime}$ is iteratively computed by applying small, bounded modifications designed to maximize the adversarial objective. At each iteration $t$ , the perturbed image $I_{t + 1}^{\\prime}$ is updated as follows:", + "bbox": [ + 89, + 633, + 483, + 738 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t + 1} ^ {\\prime} = \\Pi_ {I + \\epsilon} \\left(I _ {t} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {I} L \\left(I _ {t}, y\\right)\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 750, + 482, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $L$ is the loss function (e.g., cross-entropy loss) and $\\nabla_I L(I_t, y)$ denotes the gradient of this loss with respect to the input $I_t$ at each iteration. The term $\\alpha$ is the step size, and $\\epsilon$ is the maximum allowed perturbation bound.", + "bbox": [ + 89, + 777, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Patch-Level Contrastive Loss", + "text_level": 1, + "bbox": [ + 89, + 847, + 352, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SLADE applies a symmetric contrastive learning objective at the patch level to retain fine-grained image features under", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adversarial perturbations. Each image $I$ and its adversarial counterpart $I^{\\prime}$ are divided into non-overlapping patches. Let $P(I) = \\{p_{1},p_{2},\\ldots ,p_{n}\\}$ and $P(I^{\\prime}) = \\{p_1^{\\prime},p_2^{\\prime},\\dots ,p_n^{\\prime}\\}$ represent the respective sets of patches from clean image I and perturbed image $I^{\\prime}$ , respectively. For an image with resolution $(H,W)$ with $C$ color channels, each patch is a flattened 2D segment of size $(K,K)$ , where $K$ is the patch size. The number of resulting patches, $N$ , is calculated as $N = \\frac{H\\times W}{K^2}$ . Subsequently, each patch $p_i$ and its corresponding adversarial patch $p_i^\\prime$ are then passed through the encoder, yielding patch embeddings $z_{i} = f_{\\theta}(p_{i})$ and $z_{i}^{\\prime} = f_{\\theta}(p_{i}^{\\prime})$ , respectively.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In SLADE, we focus on ensuring the model's capability to retain fine-grained details despite adversarial perturbations. To achieve this, we apply a contrastive loss at the patch level, aligning embeddings of corresponding patches (i.e., $z_{i}$ and $z_{i}^{\\prime}$ ). By aligning patches individually, the model $f_{\\theta}$ treats each patch as an independent unit and encourages to preserve semantic coherence within localized regions. This approach enables the model to capture fine-grained and spatially distinct features, thereby retaining intricate details that might otherwise be lost to adversarial noise. Following the adversarial contrastive loss approach in [29, 60], we define the loss for $z_{i}$ and $z_{i}^{\\prime}$ as:", + "bbox": [ + 511, + 273, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o} = - \\log \\frac {\\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right) + \\sum_ {I ^ {n} \\in \\mathcal {D} _ {N}} \\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {n}\\right) / \\tau\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 465, + 903, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathcal{D}_N\\subseteq \\mathcal{D}$ represents a batch of negative samples, where $N$ is the batch size. The negative images $I^n$ are drawn uniformly at random from the dataset $\\mathcal{D}$ , where $I\\neq I^{n}$ . The embedding of patch $p_i^n$ from a negative sample $I^n$ is denoted by $z_{i}^{n}$ . $\\tau$ is a temperature parameter, and $\\mathrm{sim}(\\cdot ,\\cdot)$ denotes the cosine similarity between embeddings.", + "bbox": [ + 511, + 512, + 903, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Equation 2, negative samples are crucial to prevent a collapsing solution. In their absence, the model may map all patches to similar embeddings, failing to learn discriminative representations [56, 60]. Achieving robust performance typically requires a large number of negative samples, which increases the need for a large batch size, making it resource-intensive. Leveraging a momentum encoder [23] can mitigate the reliance on large batch sizes, though it introduces additional computational overhead.", + "bbox": [ + 511, + 604, + 905, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mitigating computational overhead: SLADE incorporates a stop-gradient mechanism [13] with a cosine similarity metric to prevent feature collapse and reduce computational overhead, which can be defined as:", + "bbox": [ + 511, + 750, + 903, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s t o p s i m} (\\cdot , \\cdot) = \\operatorname {s i m} (\\cdot , \\operatorname {s t o p g r a d} (\\cdot)) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 825, + 903, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The stop-gradient mechanism in Equation 3 is then incorporated into a symmetric contrastive loss, which is designed to align patch embeddings between clean and adver", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24246", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sarial images. The resulting contrastive loss is defined as:", + "bbox": [ + 89, + 90, + 482, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p a t c h} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left(\\operatorname {s t o p s i m} \\left(z _ {i}, z _ {i} ^ {\\prime}\\right) + \\operatorname {s t o p s i m} \\left(z _ {i} ^ {\\prime}, z _ {i}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 114, + 482, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $N$ denotes the total number of patches per image and $L_{patch}$ represents the patch-level symmetric contrastive loss, which aims to align the patch embeddings of clean and adversarial images to capture fine-grained details. The stopgrad $(\\cdot)$ operation is applied in a symmetric manner, where either $z_{i}$ or $z_{i}^{\\prime}$ is set to be constant in each term. This halts the gradient back-propagation on one side, preventing the model from simply mapping to constant outputs. By only updating the gradients on one side, stop-gradient prevents the model from converging to trivial solutions.", + "bbox": [ + 89, + 170, + 483, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Global-level Contrastive Loss", + "text_level": 1, + "bbox": [ + 89, + 333, + 354, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Equation 4, SLADE aims to preserve fine-grained details by focusing on patch-level alignment between clean and adversarial patch embeddings. However, concentrating solely on patch-level alignment could lead to a potential issue where the model captures fine-grained details in each patch but may struggle to capture the overall semantic meaning of the image as it might overlook the contextual relation between the patches. To address this and ensure the preservation of overall semantic coherence, SLADE introduces an additional global-level alignment objective on the clean $I$ and adversarial image $I^{\\prime}$ :", + "bbox": [ + 89, + 357, + 483, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g l o b a l} = - \\frac {1}{2} \\left(\\operatorname {s t o p s i m} \\left(Z, Z ^ {\\prime}\\right) + \\operatorname {s t o p s i m} \\left(Z ^ {\\prime}, Z\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 536, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $Z$ and $Z'$ are the embeddings of clean $I$ and perturbed image $I'$ , respectively.", + "bbox": [ + 89, + 580, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Final loss for SLADE. The final adversarial fine-tuning objective in SLADE combines the fine-grained and global losses with a weighting parameter $\\lambda$ (where $0 \\leq \\lambda \\leq 1$ ) to control their relative importance. The final loss function for SLADE can be defined as:", + "bbox": [ + 89, + 625, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s l a d e} = \\lambda L _ {p a t c h} + (1 - \\lambda) L _ {g l o b a l} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 715, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Minimizing the loss $L_{slade}$ effectively preserves both localized and high-level semantic features, thereby improving robustness against perturbations. Adjusting the weight $\\lambda$ provides flexibility in the model's focus: setting $\\lambda = 0$ prioritizes the global objective, emphasizing high-level semantic features, while larger values of $\\lambda$ shift the focus toward finer details within patches. However, this shift may compromise the model's ability to generalize across the entire image. In Appendix, we provide a detailed analysis of the impact of $\\lambda$ and the stop-gradient mechanism in SLADE.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental Analysis", + "text_level": 1, + "bbox": [ + 513, + 90, + 728, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LVLM models. In our experiments, we use OpenFlamingo-9B (OF) [2] and two versions of LLaVA-1.5 [34]: LLaVA-1.5 Vicuna-7B (LLaVA-7B) and LLaVA-1.5 LLaMA-2-13B (LLaVA-13B) as our target LVLMs. While these LVLMs use CLIP ViT-L-14 as their vision encoder, they incorporate different language models [14, 52, 53] for text processing. Notably, we select LLaVA-13B in our evaluation to assess the effectiveness of our robust encoder within the context of larger LVLM.", + "bbox": [ + 511, + 114, + 906, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial fine-tuning settings. In SLADE, we adversarially fine-tune the CLIP ViT-L/14 encoder on the ImageNet dataset [18], omitting labels for an unsupervised approach. We use PGD with 10 steps at $\\ell_{\\infty}$ norm to generate perturbed views. Our training spans two epochs with an AdamW optimizer (WD=1e-4, LR=1e-5) and a batch size of 64, incorporating a cosine decaying LR schedule with linear warmup. We set the weighting parameter $\\lambda$ to 0.6 and use a patch size of 32. We conduct an ablation study (see Appendix) to determine the optimal values for these hyperparameters. To ensure compatibility with LVLMs, we perform adversarial fine-tuning on the Vit-L/14 backbone of CLIP, matching the version used by LVLMs (e.g., OF and LLaVA-1.5). During robust inference, we substitute the default CLIP [48] model of the LVLMs with SLADE encoder, while keeping the language model and projection layer frozen, offering a plug-n-play adoption. While robustness is crucial, it often comes at the cost of potential degradation of clean performance. To optimize this trade-off between robust accuracy and clean performance, we adversarially fine-tune CLIP with perturbation radii $\\epsilon = 2/255$ and $\\epsilon = 4/255$ , naming these models $\\mathrm{SLADE}^2$ and $\\mathrm{SLADE}^4$ . In the following sections, we present a comprehensive comparison of our proposed fine-tuning scheme with the original CLIP and two state-of-the-art (SOTA) adversarial fine-tuned encoders, TeCoA [38] and FARE [50]. To ensure fare comparison, we use fine-tuned versions of TeCoA and FARE at the same adversarial radii, denoting them as $\\mathrm{TeCoA}^2$ and $\\mathrm{FARE}^2$ for $\\epsilon = 2/255$ , $\\mathrm{TeCoA}^4$ and $\\mathrm{FARE}^4$ for $\\epsilon = 4/255$ .", + "bbox": [ + 511, + 258, + 906, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Untargeted attacks on LVLMs", + "text_level": 1, + "bbox": [ + 513, + 704, + 784, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct untargeted attacks to assess clean and robust performance in downstream tasks. For clean evaluation, target LVLMs are given a clean image and its corresponding prompt, while for adversarial evaluation, they are given an adversarial image and prompt. In our experiments with untargeted attacks, we use the prompt \"Provide a short caption for this image\" for the image captioning task and questions from the dataset as prompts for VQA.", + "bbox": [ + 511, + 727, + 906, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Untargeted attack setup. Following [49], we employ APGD [15] attack with $\\ell_{\\infty}$ norm at $\\epsilon = 2/255$ and $\\epsilon = 4/255$ . The APGD attack is executed using single precision for 100", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24247", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c954f618b8f7213b9194c69a5d9182051428d3566fc68862c5a585c5d6e72fb2.jpg", + "table_caption": [ + "Table 1. Assessing the robustness of target LVLMs against untargeted attacks in image captioning and visual question answering (VQA) tasks. For image captioning, we evaluate performance using the COCO and Flickr30k datasets, reporting the CIDEr score. For VQA tasks, we utilize the OKVQA and VizWiz datasets, providing VQA accuracy as the evaluation metric." + ], + "table_footnote": [], + "table_body": "
Vision LVLM encoderCOCOFlickr30KOKVQAVizWiz
clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255
Open FlamingoCLIP80.57.825.62.461.06.43.81.448.51.80.00.023.82.41.80
\\( {\\mathrm{{TeCoA}}}^{2} \\)74.559.740.310.348.237.929.010.140.127.421.510.020.116.18.35.6
\\( {\\mathrm{{FARE}}}^{2} \\)76.269.455.310.949.443.432.115.437.829.521.89.718.615.89.15.6
\\( {\\mathrm{{SLADE}}}^{2} \\)79.871.561.716.252.646.937.919.543.934.826.313.523.419.713.88.5
\\( {\\mathrm{{TeCoA}}}^{4} \\)72.364.454.210.546.139.228.712.237.430.823.58.418.814.68.35.0
\\( {\\mathrm{{FARE}}}^{4} \\)75.467.352.212.148.341.334.615.139.232.223.510.517.814.79.56.1
\\( {\\mathrm{{SLADE}}}^{4} \\)78.571.862.116.453.045.937.719.944.636.128.913.822.420.314.69.1
LLaVA-7BCLIP121.921.813.52.479.015.310.03.457.38.33.00.039.313.33.20.0
\\( {\\mathrm{{TeCoA}}}^{2} \\)115.698.373.538.175.665.350.529.455.640.330.514.238.525.415.48.3
\\( {\\mathrm{{FARE}}}^{2} \\)123.5105.286.439.478.970.360.525.154.343.530.115.337.329.317.610.4
\\( {\\mathrm{{SLADE}}}^{2} \\)127.0111.594.647.282.874.365.430.961.549.232.419.142.631.120.915.1
\\( {\\mathrm{{TeCoA}}}^{4} \\)110.395.575.635.371.862.551.027.050.339.032.312.334.530.518.39.3
\\( {\\mathrm{{FARE}}}^{4} \\)119.4100.583.541.676.370.356.529.553.545.034.815.338.531.321.010.1
\\( {\\mathrm{{SLADE}}}^{4} \\)123.9109.792.145.981.573.863.733.959.848.939.423.441.630.623.517.4
", + "bbox": [ + 156, + 133, + 834, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iterations. Additionally, we evaluate the robust CLIP models under a stronger adversarial setting by using a perturbation radius of $\\epsilon = 8/255$ for the downstream tasks. The attacked images with $\\epsilon = 2/255$ perturbations are completely imperceptible to human eyes, while with $\\epsilon = 4/255$ a user could notice the perturbation when paying close attention.", + "bbox": [ + 88, + 441, + 480, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. For image captioning, we use the COCO [33] and Flickr30k [46] datasets and for VQA, we utilize OKVQA [39] and VizWiz [25] dataset. For assessing captioning tasks, we utilize the CIDEr [54] metric, which leverages higher order (longer) n-grams to capture grammatical properties as well as richer semantics [54]. Empirically, CIDEr effectively measures fine-grained detail by reflecting alignment between generated and reference captions in terms of vocabulary, context, and descriptive specificity. Higher CIDEr scores thus indicate stronger semantic alignment with ground truth. For VQA, we assess performance through VQA accuracy, which measures the percentage of correctly answered questions.", + "bbox": [ + 88, + 540, + 482, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results and discussion. Table 1 presents the performance of various CLIP encoders under untargeted attacks. While the original CLIP model excels in clean performance, it suffers significantly under adversarial attacks, particularly at stronger attack strength at $\\epsilon = 8 / 255$ radii. Adversarially fine-tuned CLIP models show slight degradation in clean performance due to adversarial fine-tuning, but SLADE achieves the best clean performance among them. For OF, $\\mathrm{SLADE^4}$ demonstrates superior performance compared to $\\mathrm{FARE^4}$ and $\\mathrm{TeCoA^4}$ across most downstream task", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets. Notably, when subjected to stronger attacks at $\\epsilon = 4/255$ and $\\epsilon = 8/255$ , $\\mathrm{SLADE}^4$ consistently outperforms all SOTA robust clip models. Similar trends are observed with $\\mathrm{SLADE}^2$ , outperforming $\\mathrm{FARE}^2$ and $\\mathrm{TeCoA}^2$ in both clean and robust performance. Additionally, for LLaVA, SLADE demonstrates even greater efficacy, achieving superior performance in both captioning and VQA tasks. Moreover, the higher CIDer score achieved by SLADE under attack demonstrates its superior ability to preserve the semantic details of image features in the generated captions compared to other robust variations of CLIP. In the Appendix, we provide results of untargeted attacks on LLaVA-13B LVLM and include visual examples of VQA.", + "bbox": [ + 511, + 441, + 906, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Targeted attacks on LVLMs", + "text_level": 1, + "bbox": [ + 511, + 650, + 764, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Targeted attack setup. We conduct targeted attacks on LLaVA-7B to evaluate the robust performance of CLIP models. Targeted attacks force LVLMs to produce a specific output predetermined by the attackers, posing a more severe threat than untargeted attacks. In our experiments, we randomly select 25 images from the COCO dataset and apply the APGD attack for 10,000 iterations, using five realistic target strings representing malicious intents such as misinformation and false propaganda. For the attack configuration, we use the $\\ell_{\\infty}$ norm with radii $\\epsilon = 2 / 255$ and $\\epsilon = 4 / 255$ . We report the CIDEr score to quantify caption quality under attack conditions and attack success rate to measure how many times the targeted attack was successful. More details on our targeted attack pipeline and evaluation metrics are provided in the Appendix.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24248", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target String: Vaccination causes autism in children.", + "bbox": [ + 307, + 95, + 640, + 108 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7de4c6695f5e8e20ae3e59af2f88a02278ddb22b2918681a4e5c1fd623c28611.jpg", + "image_caption": [ + "Clean Image" + ], + "image_footnote": [], + "bbox": [ + 161, + 132, + 248, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP:Young girl with mask getting vaccine while holding a teddy bear", + "bbox": [ + 258, + 116, + 470, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TecoA4: Young girl on a sofa taking vaccine", + "bbox": [ + 259, + 143, + 480, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FARE4: Young girl with mask taking vaccine, holding a teddy bear", + "bbox": [ + 259, + 170, + 472, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SLADE:Young girl wearing mask taking vaccine while holding teddy bear", + "bbox": [ + 254, + 195, + 486, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e59069186c8b34a19a029394dbad876e9fc58ad2adc53b6a78073fbcccab85e.jpg", + "image_caption": [ + "Adversarial Image" + ], + "image_footnote": [], + "bbox": [ + 496, + 132, + 584, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP: Vaccination causes autism in children.", + "bbox": [ + 596, + 114, + 810, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TecoA: Vaccine shot for children", + "bbox": [ + 596, + 143, + 803, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FARE: Young girl taking vaccine, holding a doll", + "bbox": [ + 596, + 167, + 800, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SLADE4:Young girl wearing mask taking vaccine while holding teddy bear", + "bbox": [ + 593, + 195, + 823, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Target String: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk.", + "bbox": [ + 209, + 233, + 794, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7321226e8b26b47e7528cd7ee68c583bf01a3126da0d42dc42805b337130f42.jpg", + "image_caption": [ + "Clean Image" + ], + "image_footnote": [], + "bbox": [ + 161, + 268, + 248, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP: A statue of a muscular man with beard and six packs", + "bbox": [ + 258, + 256, + 486, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TecoA4: A statue of a man with beard and muscles", + "bbox": [ + 259, + 282, + 483, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FARE: A statue of a man with a beard and six-packs", + "bbox": [ + 259, + 310, + 490, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SLADE4: A statue of a shirtless muscu-lar man with beard and six-packs", + "bbox": [ + 256, + 335, + 486, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb23b73d19c1439a2d9e81963f90c917a2611a4b4a5c7d51a936f06ffe57117a.jpg", + "image_caption": [ + "Adversarial Image", + "Figure 3. Qualitative examples of $\\ell_{\\infty}$ targeted attacks at $\\epsilon = 4/255$ radii using original and adversarially fine-tuned CLIP models as vision encoder in LLaVA-7B. We present generated captions (target string, mistakes in captions, fine-grained details in caption) for the original images (left) and the imperceptible adversarial images (right)." + ], + "image_footnote": [], + "bbox": [ + 501, + 273, + 583, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk.", + "bbox": [ + 594, + 250, + 821, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TecoA4: Before and after of photosh -ot of a man", + "bbox": [ + 594, + 286, + 818, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FARE4: A statue of a man with a beard and muscles", + "bbox": [ + 594, + 311, + 825, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SLADE4: Statue of a shirtless muscu-lar man with a beard and six packs", + "bbox": [ + 593, + 338, + 812, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results and discussion. We present quantitative results of our targeted attacks at $\\epsilon = 4 / 255$ in Table 2 and qualitative examples in Figure 3. The original CLIP model shows high susceptibility to targeted attacks by consistently generating the specified target strings. In contrast, $\\mathrm{TeCoA}^2$ and $\\mathrm{FARE}^2$ break in only 6 and 2 cases, resulting in mean success rates of $4.8\\%$ and $1.6\\%$ , respectively, while $\\mathrm{SLADE}^2$ breaks in only one case, highlighting its superior performance. Notably, $\\mathrm{SLADE}^4$ , $\\mathrm{FARE}^4$ , and $\\mathrm{TeCoA}^4$ show complete robustness under targeted attack at $\\epsilon = 4 / 255$ . However, the quality of captions generated by $\\mathrm{SLADE}^4$ notably surpasses $\\mathrm{FARE}^4$ and $\\mathrm{TeCoA}^4$ , as shown in Figure 3. We also observe that captions generated by $\\mathrm{TeCoA}^4$ exhibit inferior quality and contain errors, while $\\mathrm{FARE}^4$ 's captions often lack intricate details from the corresponding images. For instance, in first example of Figure 3, SLADE retains the fine-grained details of the young girl wearing mask and holding teddy bear whereas FARE overlooks these details and TeCoA generates captions with mistakes. This highlights SLADE's robustness, effectively resisting attacks and retaining both fine details and semantic integrity. Empirically, the CIDEr scores in Table 2 further support these claims, with $\\mathrm{SLADE}^4$ achieving the highest score. We present more qualitative results in the Appendix.", + "bbox": [ + 91, + 440, + 483, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Zero-shot evaluation under AutoAttack", + "text_level": 1, + "bbox": [ + 89, + 816, + 429, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack setup. We examine the adversarial robustness of various CLIP models in zero-shot classification tasks using AutoAttack [15]. Our evaluation spans multiple datasets, including CIFAR10, CIFAR100 [31], StanfordCars [30],", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EuroSAT [26], PCAM [55], and Flowers [42]. For multiclass classification datasets, we utilize AutoAttack's ensemble of four attacks: APGD-CE, APGD-T, FAB, and Square. For binary classification datasets like PCAM, we exclude APGD with targeted DLR loss (APGD-T), as DLR loss is unsuitable for binary classification, and use the remaining three attacks instead. We apply AutoAttack for 100 iterations on 1,000 samples from each dataset at $\\epsilon = 4/255$ radii.", + "bbox": [ + 511, + 439, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results and discussion. Table 3 summarizes the performance of both original and robust CLIP models on zero-shot classification tasks. At $\\epsilon = 4/255$ radii, SLADE consistently outperforms TeCoA and FARE. Notably, $\\mathrm{SLADE}^4$ demonstrates greater robustness compared to $\\mathrm{FARE}^4$ and $\\mathrm{TeCoA}^4$ , with a $3.1\\%$ gain in robust accuracy, while also slightly outperforming $\\mathrm{SLADE}^2$ . These findings highlight SLADE's superior resilience against AutoAttack's ensemble of adversarial attacks. In Appendix, we present detailed results for clean and adversarial evaluations under $\\epsilon = 2/255$ and $\\epsilon = 8/255$ radii on zero-shot classification datasets.", + "bbox": [ + 511, + 575, + 906, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Jailbreak attacks on LVLMs", + "text_level": 1, + "bbox": [ + 511, + 755, + 769, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In previous sections, we highlighted the vulnerability of LVLMs to gradient-based attacks. These models are also susceptible to jailbreak attacks due to the expanded attack surface from the image modality. Adversaries use gradient-based perturbations to create universal optimization-based jailbreak attacks, where a single adversarial image can compromise multiple LVLMs (e.g., LLaVA, MiniGPT-4 [65]) without contextually aligned prompts.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24249", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8aa5069df2561da951443142469c0bac2f1825cd1e8679b81257a13e85aaf98.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of ${\\ell }_{\\infty }$ targeted attacks at $\\epsilon = 4/{255}$ radii." + ], + "table_footnote": [], + "table_body": "
TargetCLIPTeCoA2FARE2SLADE2TeCoA4FARE4SLADE4
Vaccination causes autism...25/250/250/250/250/250/250/25
Achieve this physique with ‘X’...25/252/251/250/250/250/250/25
Invest in ‘Y’ stock for quick...25/251/250/250/250/250/250/25
Visit https://tinyurl.com/zh...25/253/251/251/250/250/250/25
Opposition political leaders...25/250/250/250/250/250/250/25
Average attack success rate:100%4.8%1.6%0.8%0%0%0%
Average CIDEr score:018.625.148.467.378.787.5
", + "bbox": [ + 107, + 107, + 885, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6ed7e7bbbc9c9e5435dbad63a54a3638eb361ad4b8a3a88d1d74a7656876890.jpg", + "table_caption": [ + "Table 3. Evaluation of CLIP models on zero-shot classification datasets under AutoAttack." + ], + "table_footnote": [], + "table_body": "
AttackVision encoderCIFAR10CIFAR100CarsFlowersEuroSATPCAM
4/255CLIP0.00.00.00.00.00.00.00.0
TeCoA231.017.85.86.73.516.0
FARE225.914.04.87.15.517.2
SLADE233.419.36.18.84.719.2
€TeCoA435.521.68.412.46.843.5
FARE434.821.412.812.911.750.2
SLADE434.423.015.313.712.551.3
", + "bbox": [ + 91, + 291, + 483, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attack Setup. To evaluate the robustness of vision encoders against jailbreak attacks, we use two optimization-based attacks: ImgJP [43] and VisualAdv [47], both of which employ gradient-based perturbations to generate universal adversarial images that bypass LVLMs' safety mechanisms. For ImgJP, we perform 1,000 iterations at $\\epsilon = 16/255$ under the $\\ell_{\\infty}$ norm, using MiniGPT-4 (Vicuna 7B) to generate adversarial images targeting the LLaVA-7B and LLaVA-13B models. Each image is tested against 25 harmful instructions from the AdvBench dataset [66], with Attack Success Rate (ASR) reported as the percentage of harmful prompts that elicit unsafe responses. In the VisualAdv evaluation, universal imperceptible adversarial images are generated with high perturbation ( $\\epsilon = 128/255$ ) and tested against 100 toxic prompts from the RealToxicityPrompts dataset [21]. Using LLaVA-13B as a surrogate model, we measure the toxicity in the outputs via the Perspective API [28] and report the percentage of responses exhibiting attributes like threat or profanity.", + "bbox": [ + 88, + 484, + 482, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and discussions. Table 4 presents the performance of different CLIP encoders under VisualAdv jailbreak attacks with LLaVA-13B. Substituting the original CLIP encoder with $\\mathrm{SLADE^4}$ significantly lowers toxicity rates, reducing them from $30.4\\%$ to $13.7\\%$ at $\\epsilon = 128 / 255$ . While TeCoA and FARE also reduce toxicity rates compared to the original CLIP model, they remain less effective than SLADE. Although external defenses like JailGuard", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and Diffpure show some promise, $\\mathrm{SLADE^4}$ consistently achieves superior results. For our additional evaluation of VisualAdv attacks, see Appendix. Table 5 presents the ASR results for ImgJP attacks across target LVLMs. The original CLIP encoder in LLaVA models shows high ASR, particularly in LLaVA-7B. While external defenses like JailGuard decrease ASR, they add complexity and overhead due to requiring LVLM-specific training. $\\mathrm{SLADE^4}$ outperforms $\\mathrm{FARE^4}$ and $\\mathrm{TeCoA^4}$ , achieving the lowest ASR and underscoring SLADE's effectiveness against both gradient-based and universal jailbreak attacks.", + "bbox": [ + 511, + 260, + 906, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Evaluation across other Tasks", + "text_level": 1, + "bbox": [ + 511, + 439, + 779, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the preceding sections, we evaluated the performance of robust encoders in LVLMs under adversarial and jailbreak attacks. While resilience against such attacks is critical, it is equally important to ensure that LVLMs equipped with robust encoders maintain their core functionalities, including reasoning and instruction-following capabilities. In this section, we assess whether LVLMs with robust encoders can effectively retain these essential abilities, particularly in solving reasoning tasks and adhering to instructions.", + "bbox": [ + 511, + 464, + 908, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruction following benchmark. Instruction following is a fundamental capability in LVLMs, enabling them to accurately interpret and respond to contextually specific requests across diverse and scenarios. This capability extends LVLMs beyond basic image recognition, allowing them to generate responses that are sensitive to context, interpret visual nuances, and incorporate reasoning about complex, scene-specific details [2, 59]. We assess the instruction-following performance of LVLMs with robust encoders through a quantitative evaluation using VisIT-Bench [8], a recently developed comprehensive benchmark that includes 592 challenging instructions across 70 diverse instruction categories. In this evaluation, we use LLaVA-13B and a subset of 200 instructions, where each instance includes an instruction, an input image, and a human-verified GPT-4 response. Following [8], we set the GPT-4 [45] response as the ground truth and use BERTscore [61] and METEOR score [5] as evaluation metrics and summarize the results in Table 6. We observe that SLADE upholds instruction", + "bbox": [ + 511, + 613, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24250", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca8e09ee0d125c9ee6db254408e4310997b3db4b351d7cec1c09f904e3bb350a.jpg", + "table_caption": [ + "Table 4. Evaluation of LLaVA-13B against VisualAdv jailbreak attacks with different CLIP based models. We report the percentage of LVLM outputs that exhibit specific toxic attributes. The blue downward arrows (↓) indicate a decrease in toxic attribute percentages compared to the baseline CLIP model, while the red upward arrows (↑) indicate an increase. Lower values signify better performance." + ], + "table_footnote": [], + "table_body": "
Attack Strength (€)Vision EncoderExternal DefenseAny IdentityProfanitySevere ToxicitySexually ExplicitThreatToxicityAverage
128/255CLIP—62.09.053.05.019.04.061.0
JailGuard41.04.020.01.014.05.030.0
Diffpure (n = 0.25)35.04.028.02.011.02.027.0
TeCoA4—41.04.027.02.015.05.045.0
FARE4—36.04.019.01.09.05.033.0
SLADE4—34.04.017.00.010.05.030.0
", + "bbox": [ + 122, + 135, + 874, + 272 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/438702c441c8ccba4f76c37708f949a8c2e8aa1111bcf74e906cf87e955a4d94.jpg", + "table_caption": [ + "Table 5. Performance analysis of LVLMs with various vision encoders and defenses against ImgJP Attack at $\\epsilon = 16/255$ ." + ], + "table_footnote": [], + "table_body": "
LVLMVision EncoderExternal DefenseASR (%)
LLaVA-13BCLIP—28.0 ± 0.2
FARE4—18.3 ± 0.1
CLIPJailGuard21.0 ± 0.1
SLADE4—15.0 ± 0.3
LLaVA-7BCLIP—34.0 ± 0.1
FARE4—24.0 ± 0.2
CLIPJailGuard26.0 ± 0.3
SLADE4—19.3 ± 0.5
", + "bbox": [ + 117, + 314, + 450, + 486 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0fb6790b98d3d6dd5ecb8efeb2c4afa4557d9d8ab6c39fc0e87c9d8ff6ad4545.jpg", + "table_caption": [ + "Table 6. Instruction-following benchmark results of LLaVA 13B-model with original CLIP and robust vision encoders." + ], + "table_footnote": [], + "table_body": "
Vision EncoderBERTScoreMETEOR Score
PrecisionRecallF1
CLIP0.8910.8940.8920.37
SLADE²0.8900.8950.8870.36
SLADE⁴0.8890.8910.8890.36
", + "bbox": [ + 112, + 529, + 459, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "following performance comparably to the original CLIP. In Appendix, we present additional details on the instruction-following benchmark setup and qualitative examples.", + "bbox": [ + 89, + 621, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on reasoning tasks. LVLMs are designed to emulate human reasoning, often achieved through a chain-of-thought approach, where a series of intermediate reasoning steps significantly enhances their ability to handle complex tasks. Following [50], we assess the reasoning capabilities of LVLMs equipped with robust vision encoders on the Science Question Answering (SQA) [35] benchmark and summarize the results in Table 7. We evaluate LLaVA-13B on a subset of 10,000 image-question pairs from the SQA benchmark, where each instance includes an explanation of a concept, followed by a question and an input image. LLaVA with the original CLIP model achieves the best accuracy of $64.5\\%$ . Both FARE and SLADE perform better than TeCoA. Among all the robust encoders, $\\mathrm{SLADE}^2$ performs best and achieves an accuracy of $63.4\\%$ accuracy.", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9801172d4b9649c2e31ca4b74f080d8f3acc3d7eeffc9430a7d0a618ffb83f34.jpg", + "table_caption": [ + "Table 7. Reasoning tasks evaluation with LLaVA on the SQA benchmark. We report accuracy for image-question pairs." + ], + "table_footnote": [], + "table_body": "
CLIPTeCoA2FARE2TeCoA4FARE4SLADE2SLADE4
64.561.162.463.059.363.462.3
", + "bbox": [ + 527, + 314, + 888, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our evaluations of reasoning and instruction-following tasks reveal that SLADE achieves performance comparable to the original CLIP model. This demonstrate that SLADE enables LVLMs to preserve their core capabilities, such as reasoning and instruction-following, without compromise. Furthermore, as detailed in the Appendix, SLADE also improves its image generation capabilities, demonstrating its versatility beyond robust performance.", + "bbox": [ + 511, + 364, + 906, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 497, + 633, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SLADE, an unsupervised adversarial fine-tuning scheme designed to enhance the robustness of the widely-used CLIP vision encoder against both gradient-based and optimization-based jailbreak attacks. By leveraging a dual-level contrastive learning approach, SLADE effectively captures fine-grained image details at the patch level while preserving high-level semantic consistency at the global level, enabling VLMs to effectively withstand adversarial perturbations without requiring any computational overhead. Our extensive experiments demonstrate that SLADE-equipped LVLMs outperform existing models in adversarial robustness, excelling in nuanced semantic understanding of perturbed images while maintaining high accuracy on clean data. Moreover, we empirically demonstrate that replacing the default CLIP encoder with SLADE enables robust defense against jailbreak attacks without requiring any LVLM-specific training, highlighting its seamless integration and adaptability across multimodal systems. Further, we validate that SLADE-equipped LVLMs do not compromise on non-adversarial core performance, such as reasoning and instruction-following capabilities. These findings position SLADE as a transformative approach with the potential to set a new benchmark for adversarial robustness in multimodal AI, paving the way for safer and more reliable applications in real-world scenarios.", + "bbox": [ + 511, + 523, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24251", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 272, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This material is partly based upon work supported by the U.S. National Science Foundation (NSF) under Grant No. CRII-IIS-RI-2348145. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF.", + "bbox": [ + 89, + 109, + 485, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1", + "[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 4, 7", + "[3] Eugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, and Vitaly Shmatikov. (ab) using images and sounds for indirect instruction injection in multi-modal llms. arXiv preprint arXiv:2307.10490, 2023. 2", + "[4] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime. arXiv preprint arXiv:2309.00236, 2023. 2", + "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7", + "[6] Hritik Bansal, Nishad Singhi, Yu Yang, Fan Yin, Aditya Grover, and Kai-Wei Chang. Cleanclip: Mitigating data poisoning attacks in multimodal contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112–123, 2023. 2", + "[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1", + "[8] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595, 2023. 7", + "[9] Nicholas Carlini, Milad Nasr, Christopher A Choquette-Choo, Matthew Jagielski, Irena Gao, Pang Wei W Koh, Daphne Ippolito, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned? Advances in Neural Information Processing Systems, 36, 2024. 1", + "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Pro" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1", + "[11] Lu Chen, Qiangchang Wang, Zhaohui Li, and Yilong Yin. Hypergraph-guided intra-and inter-category relation modeling for fine-grained visual recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8043-8052, 2024. 3", + "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3", + "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 4", + "[15] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 4, 6", + "[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2", + "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4", + "[19] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 2", + "[20] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in neural information processing systems, 34:21480-21492, 2021. 2", + "[21] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462, 2020. 7", + "[22] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1", + "[23] Sven Gowal, Po-Sen Huang, Aaron van den Oord, Timothy Mann, and Pushmeet Kohli. Self-supervised adversarial robustness for the low-label, high-data regime. In International conference on learning representations, 2020. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24252", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2", + "[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5", + "[26] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6", + "[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in neural information processing systems, 33: 16199-16210, 2020. 2", + "[28] Google Jigsaw. Perspective api. https://www.perspectiveapi.com/, 2017. Accessed: 2024-06-02. 7", + "[29] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in neural information processing systems, 33:2983-2994, 2020. 3", + "[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6", + "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[32] Xiao Li, Wei Zhang, Yining Liu, Zhanhao Hu, Bo Zhang, and Xiaolin Hu. Language-driven anchors for zero-shot adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24686-24695, 2024. 2", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 4", + "[35] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 8", + "[36] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3", + "[38] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 1, 2, 4", + "[39] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5", + "[40] AI Meta. Introducing llama: A foundational, 65-billion-parameter large language model. Meta AI, 2023. 1", + "[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 3", + "[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6", + "[43] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 1, 2, 7", + "[44] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3", + "[45] R OpenAI. Gpt-4 technical report. ArXiv, 2303, 2023. 1, 7", + "[46] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 5", + "[47] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 7", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4", + "[49] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 2, 4", + "[50] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336, 2024. 1, 2, 4, 8" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24253", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Arindam Sikdar, Yonghuai Liu, Siddhardha Kedarisetty, Yitian Zhao, Amr Ahmed, and Ardhendu Behera. Interweaving insights: High-order feature interaction for fine-grained visual recognition. International Journal of Computer Vision, pages 1-25, 2024. 3", + "[52] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 4", + "[53] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 4", + "[54] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5", + "[55] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation equivariant cnns for digital pathology. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11, pages 210-218. Springer, 2018. 6", + "[56] Xiao Wang, Yuhang Huang, Dan Zeng, and Guo-Jun Qi. Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10718-10730, 2023. 3", + "[57] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 1", + "[58] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 2", + "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7", + "[60] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness. In European Conference on Computer Vision, pages 725-742. Springer, 2022. 3", + "[61] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 7", + "[62] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 2", + "[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[64] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial examples in multimodal contrastive learning. In 31st ACM International Conference on Multimedia, pages 6311-6320, 2023. 2", + "[65] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6", + "[66] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 7" + ], + "bbox": [ + 517, + 92, + 903, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24254", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_model.json b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d104ae54f23dd1d68f5f5457041b08f9e12f5f87 --- /dev/null +++ b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.878, + 0.153 + ], + "angle": 0, + "content": "SLADE: Shielding against Dual Exploits in Large Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.181, + 0.636, + 0.2 + ], + "angle": 0, + "content": "Md Zarif Hossain, Ahmed Imteaj†" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.201, + 0.813, + 0.218 + ], + "angle": 0, + "content": "School of Computing, Southern Illinois University Carbondale, IL 62901, USA" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.253, + 0.327, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.285, + 0.482, + 0.678 + ], + "angle": 0, + "content": "Large Vision-Language Models (LVLMs) have emerged as transformative tools in multimodal tasks, seamlessly integrating pretrained vision encoders to align visual and textual modalities. Prior works have highlighted the susceptibility of LVLMs to dual exploits (gradient-based and optimization-based jailbreak attacks), which leverage the expanded attack surface introduced by the image modality. Despite advancements in enhancing robustness, existing methods fall short in their ability to defend against dual exploits while preserving fine-grained semantic details and overall semantic coherence under intense adversarial perturbations. To bridge this gap, we introduce SheiLding against Dual Exploits (SLADE), a novel unsupervised adversarial fine-tuning scheme that enhances the resilience of CLIP-based vision encoders. SLADE's dual-level contrastive learning approach balances the granular and the holistic, capturing fine-grained image details without losing sight of high-level semantic coherence. Extensive experiments demonstrate that SLADE-equipped LVLMs set a new benchmark for robustness against dual exploits while preserving fine-grained semantic details of perturbed images. Notably, SLADE achieves these results without compromising the core functionalities of LVLMs, such as instruction following, or requiring the computational overhead (e.g., large batch sizes, momentum encoders) commonly associated with traditional contrastive learning methods." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.709, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The remarkable success of large language models (LLMs) [40, 45] in understanding and generating human-like text has inspired researchers to extend their capabilities to the visual domain. This effort has led to the development of Large Vision-Language Models (LVLMs) [1, 34], which aim to bridge the gap between visual and textual information. These LVLMs frequently utilize pretrained vision encoders, such as CLIP [48], BEiT [7], and DINO [10], to derive meaningful visual representations from images. LVLMs [2, 34] equipped with these vision encoders demonstrate remarkable zero-shot performance in down-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.908, + 0.617 + ], + "angle": 0, + "content": "stream tasks such as image captioning, and visual question answering (VQA). Their promising capabilities and impact on multimodal learning have led to a growing number of LVLMs being publicly released. However, while this availability fosters research and application development, it also introduces critical safety concerns that require careful attention. Recent studies [57, 63] highlight the high susceptibility of LVLMs to harmful exploits on text and image modalities. Notably, the addition of a vision modality increases the model's attack surface, making LVLMs even more vulnerable to manipulation compared to LLMs [9, 22]. Moreover, the adversaries can use imperceptible gradient-based perturbations to create targeted attacks, generating malicious outputs that evade detection [49]. Such targeted attacks are particularly concerning due to their ability to manipulate models toward specific, often harmful outputs. Additionally, the increasing deployment of LVLMs raises the potential for exploitation not only by adversaries but also by corporations aiming to leverage these models to promote their products or agendas as shown in Figure 1. Recent studies [43, 47] demonstrate that gradient-based perturbations can be adapted in optimization-based jailbreak attacks to bypass a model's safety mechanisms, causing LVLMs to generate harmful or undesirable content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "While existing methods [38, 50] made notable progress in improving the robustness of vision encoders, particularly CLIP, which is widely utilized across various applications, critical challenges persist. For example, in downstream tasks such as image captioning, these models successfully capture high-level semantic meaning but often struggle to preserve fine-grained semantic details when faced with perturbed images. Consequently, their ability to generate accurate and contextually rich descriptions is compromised, limiting their applicability and effectiveness in real-world scenarios where precision and detail are paramount. To address these concerns, we propose SLADE, an unsupervised adversarial fine-tuning scheme that enhances the robustness of the widely used CLIP vision encoder's embeddings against dual exploits: gradient-based attacks and optimization-based Jailbreak attacks. SLADE combines symmetric contrastive learning objectives at both the patch level and the global level (on full image) to capture fine" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24244" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.864, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Figure 1. Targeted \\(\\ell_{\\infty}\\) attack and VisualAdV [47] jailbreak attack at \\(\\epsilon = 4 / 255\\) radii and using original CLIP and SLADE as vision encoder in LLaVA-7B. SLADE (trained at \\(\\epsilon = 4 / 255\\)) generates captions with fine-grained semantic details for the perturbed sample and successfully defends against the jailbreak attempt." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.482, + 0.587 + ], + "angle": 0, + "content": "grained image details and high-level semantic representation of the entire image, respectively. This dual-level approach enables the model to preserve both detailed features and high-level semantic consistency, even in the presence of adversarial perturbations. Moreover, SLADE does not require LVLM or task-specific training and can be seamlessly integrated into existing LVLMs as a robust vision encoder. SLADE enhances the adversarial robustness of the vision modality in LVLMs, preserving the model's ability to capture nuanced semantic details while effectively defending against jailbreak attacks, as illustrated in Figure 1. Our work sets a new standard for adversarial robustness in LVLMs, a critical advancement as these models become integral to critical applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.594, + 0.241, + 0.609 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.771 + ], + "angle": 0, + "content": "Large Vision-Language Models. In recent times, several advanced LVLMs have been introduced, including LLaVA [34], OpenFlamingo (OF) [2], InstructBLIP [17] and Macaw-LLM [36]. These models leverage the power of multimodal learning by combining visual and textual data, thereby, enhancing their capability to understand and generate responses to multimodal inputs. Most of these LVLMs incorporate pre-trained LLMs, such as Llama [53], paired with large-scale vision encoders like CLIP [48], to process multimodal data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Adversarial and Jailbreaking attacks in LVLMs. In recent studies [3, 4, 19, 63], authors have explored the susceptibility of large vision-language models to adversarial attacks. Notably, recent works [3, 6, 27, 64] demonstrate the vulnerability of vision encoders in LVLMs, particularly CLIP, to adversarial manipulation, which can significantly impair the performance of LVLMs in downstream tasks. Moreover, [16, 49] highlight the efficacy of gradient-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.527 + ], + "angle": 0, + "content": "based attacks on LVLMs, demonstrating their capacity to induce the model into generating harmful or misleading outputs. In [43, 47] authors show how gradient-based perturbations can be leveraged in optimization-based jailbreak attacks to bypass LVLMs' safety protocols. While recent studies have proposed defense strategies [58, 62] to counter jailbreak attacks, these methods rely heavily on the model's inherent safety alignment, necessitating specific training for each LVLM and often resulting in degraded performance in downstream tasks (e.g., VQA and image captioning)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Robustness against exploits in LVLMs. Prior works [32, 38] introduced supervised adversarial fine-tuning scheme for CLIP, using cross-modal image-text contrastive loss. However, in LVLMs their zero-shot performance in downstream tasks (mostly in unseen classes) often degrades due to an over-dependence on the text embeddings during the fine-tuning process. To address this, some concurrent works [20, 23, 27] proposed unsupervised adversarial training methods using SimCLR [12] and BYOL [24] contrastive loss, which improve robustness but require large batch sizes for sufficient negative samples or a momentum encoder for optimal results. Additionally, the authors in [50] proposed an alternative approach, tailoring an \\(\\ell_2\\) loss-based unsupervised fine-tuning scheme. However, our empirical analysis reveals that applying \\(\\ell_2\\) loss across the entire image is insufficient for retaining nuanced details, particularly under adversarial conditions. Unlike prior methods, SLADE integrates a symmetric contrastive learning objective applied simultaneously at the patch level and global level. This dual-level strategy enables SLADE to capture fine-grained details and nuanced spatial relationships, even under adversarial perturbations. Drawing inspiration from traditional machine learning techniques that use patch embedding with pretext tasks to preserve detail and spatial" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24245" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.182 + ], + "angle": 0, + "content": "localization [11, 41, 44, 51], SLADE advances this concept by incorporating it directly into its adversarial fin-tuning mechanism. Furthermore, SLADE leverages a stop-gradient mechanism to bypass the reliance on large batches or momentum encoders, offering a scalable and computationally efficient solution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.288, + 0.211 + ], + "angle": 0, + "content": "3. Overview of SLADE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.484, + 0.371 + ], + "angle": 0, + "content": "This section provides a detailed overview of SLADE, our proposed adversarial fine-tuning scheme, illustrated in Figure 2. The primary objective of SLADE is to adversarially fine-tune a CLIP-like model \\( f_{\\theta}(\\cdot) \\), with parameter \\( \\theta \\) on a given dataset \\( \\mathcal{D} = \\{I_1, I_2, \\dots, I_N\\} \\) with \\( I_n \\in \\mathbb{R}^{H \\times W \\times 3} \\). SLADE is designed to learn invariant features that are robust to adversarial perturbations while preserving the semantic details of the input data. This ensures that the fine-tuned model effectively balances adversarial resilience with the retention of critical semantic information." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.383, + 0.485, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.586, + 0.455, + 0.601 + ], + "angle": 0, + "content": "Figure 2. Workflow of our proposed scheme: SLADE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.357, + 0.627 + ], + "angle": 0, + "content": "3.1. Adversarial Image generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.484, + 0.739 + ], + "angle": 0, + "content": "During adversarial fine-tuning, SLADE generates a perturbed version \\(I^{\\prime}\\) from the original clean image \\(I\\) using Projected Gradient Descent (PGD) [37]. Given an input image \\(I\\), the perturbed image \\(I^{\\prime}\\) is iteratively computed by applying small, bounded modifications designed to maximize the adversarial objective. At each iteration \\(t\\), the perturbed image \\(I_{t + 1}^{\\prime}\\) is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.75, + 0.483, + 0.768 + ], + "angle": 0, + "content": "\\[\nI _ {t + 1} ^ {\\prime} = \\Pi_ {I + \\epsilon} \\left(I _ {t} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {I} L \\left(I _ {t}, y\\right)\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Here, \\( L \\) is the loss function (e.g., cross-entropy loss) and \\( \\nabla_I L(I_t, y) \\) denotes the gradient of this loss with respect to the input \\( I_t \\) at each iteration. The term \\( \\alpha \\) is the step size, and \\( \\epsilon \\) is the maximum allowed perturbation bound." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.353, + 0.863 + ], + "angle": 0, + "content": "3.2. Patch-Level Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "SLADE applies a symmetric contrastive learning objective at the patch level to retain fine-grained image features under" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "adversarial perturbations. Each image \\(I\\) and its adversarial counterpart \\(I^{\\prime}\\) are divided into non-overlapping patches. Let \\(P(I) = \\{p_{1},p_{2},\\ldots ,p_{n}\\}\\) and \\(P(I^{\\prime}) = \\{p_1^{\\prime},p_2^{\\prime},\\dots ,p_n^{\\prime}\\}\\) represent the respective sets of patches from clean image I and perturbed image \\(I^{\\prime}\\), respectively. For an image with resolution \\((H,W)\\) with \\(C\\) color channels, each patch is a flattened 2D segment of size \\((K,K)\\), where \\(K\\) is the patch size. The number of resulting patches, \\(N\\), is calculated as \\(N = \\frac{H\\times W}{K^2}\\). Subsequently, each patch \\(p_i\\) and its corresponding adversarial patch \\(p_i^\\prime\\) are then passed through the encoder, yielding patch embeddings \\(z_{i} = f_{\\theta}(p_{i})\\) and \\(z_{i}^{\\prime} = f_{\\theta}(p_{i}^{\\prime})\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.456 + ], + "angle": 0, + "content": "In SLADE, we focus on ensuring the model's capability to retain fine-grained details despite adversarial perturbations. To achieve this, we apply a contrastive loss at the patch level, aligning embeddings of corresponding patches (i.e., \\( z_{i} \\) and \\( z_{i}^{\\prime} \\)). By aligning patches individually, the model \\( f_{\\theta} \\) treats each patch as an independent unit and encourages to preserve semantic coherence within localized regions. This approach enables the model to capture fine-grained and spatially distinct features, thereby retaining intricate details that might otherwise be lost to adversarial noise. Following the adversarial contrastive loss approach in [29, 60], we define the loss for \\( z_{i} \\) and \\( z_{i}^{\\prime} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.466, + 0.905, + 0.512 + ], + "angle": 0, + "content": "\\[\nL _ {c o} = - \\log \\frac {\\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right) + \\sum_ {I ^ {n} \\in \\mathcal {D} _ {N}} \\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {n}\\right) / \\tau\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.513, + 0.905, + 0.603 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{D}_N\\subseteq \\mathcal{D}\\) represents a batch of negative samples, where \\(N\\) is the batch size. The negative images \\(I^n\\) are drawn uniformly at random from the dataset \\(\\mathcal{D}\\), where \\(I\\neq I^{n}\\). The embedding of patch \\(p_i^n\\) from a negative sample \\(I^n\\) is denoted by \\(z_{i}^{n}\\). \\(\\tau\\) is a temperature parameter, and \\(\\mathrm{sim}(\\cdot ,\\cdot)\\) denotes the cosine similarity between embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.906, + 0.74 + ], + "angle": 0, + "content": "In Equation 2, negative samples are crucial to prevent a collapsing solution. In their absence, the model may map all patches to similar embeddings, failing to learn discriminative representations [56, 60]. Achieving robust performance typically requires a large number of negative samples, which increases the need for a large batch size, making it resource-intensive. Leveraging a momentum encoder [23] can mitigate the reliance on large batch sizes, though it introduces additional computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.905, + 0.811 + ], + "angle": 0, + "content": "Mitigating computational overhead: SLADE incorporates a stop-gradient mechanism [13] with a cosine similarity metric to prevent feature collapse and reduce computational overhead, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.826, + 0.905, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s t o p s i m} (\\cdot , \\cdot) = \\operatorname {s i m} (\\cdot , \\operatorname {s t o p g r a d} (\\cdot)) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The stop-gradient mechanism in Equation 3 is then incorporated into a symmetric contrastive loss, which is designed to align patch embeddings between clean and adver" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24246" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.108 + ], + "angle": 0, + "content": "sarial images. The resulting contrastive loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.116, + 0.483, + 0.17 + ], + "angle": 0, + "content": "\\[\nL _ {p a t c h} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left(\\operatorname {s t o p s i m} \\left(z _ {i}, z _ {i} ^ {\\prime}\\right) + \\operatorname {s t o p s i m} \\left(z _ {i} ^ {\\prime}, z _ {i}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.484, + 0.322 + ], + "angle": 0, + "content": "Here, \\(N\\) denotes the total number of patches per image and \\(L_{patch}\\) represents the patch-level symmetric contrastive loss, which aims to align the patch embeddings of clean and adversarial images to capture fine-grained details. The stopgrad \\((\\cdot)\\) operation is applied in a symmetric manner, where either \\(z_{i}\\) or \\(z_{i}^{\\prime}\\) is set to be constant in each term. This halts the gradient back-propagation on one side, preventing the model from simply mapping to constant outputs. By only updating the gradients on one side, stop-gradient prevents the model from converging to trivial solutions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.334, + 0.355, + 0.349 + ], + "angle": 0, + "content": "3.3. Global-level Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.525 + ], + "angle": 0, + "content": "In Equation 4, SLADE aims to preserve fine-grained details by focusing on patch-level alignment between clean and adversarial patch embeddings. However, concentrating solely on patch-level alignment could lead to a potential issue where the model captures fine-grained details in each patch but may struggle to capture the overall semantic meaning of the image as it might overlook the contextual relation between the patches. To address this and ensure the preservation of overall semantic coherence, SLADE introduces an additional global-level alignment objective on the clean \\(I\\) and adversarial image \\(I^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.537, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nL _ {g l o b a l} = - \\frac {1}{2} \\left(\\operatorname {s t o p s i m} \\left(Z, Z ^ {\\prime}\\right) + \\operatorname {s t o p s i m} \\left(Z ^ {\\prime}, Z\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Here, \\( Z \\) and \\( Z' \\) are the embeddings of clean \\( I \\) and perturbed image \\( I' \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Final loss for SLADE. The final adversarial fine-tuning objective in SLADE combines the fine-grained and global losses with a weighting parameter \\(\\lambda\\) (where \\(0 \\leq \\lambda \\leq 1\\)) to control their relative importance. The final loss function for SLADE can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.717, + 0.483, + 0.734 + ], + "angle": 0, + "content": "\\[\nL _ {s l a d e} = \\lambda L _ {p a t c h} + (1 - \\lambda) L _ {g l o b a l} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Minimizing the loss \\( L_{slade} \\) effectively preserves both localized and high-level semantic features, thereby improving robustness against perturbations. Adjusting the weight \\( \\lambda \\) provides flexibility in the model's focus: setting \\( \\lambda = 0 \\) prioritizes the global objective, emphasizing high-level semantic features, while larger values of \\( \\lambda \\) shift the focus toward finer details within patches. However, this shift may compromise the model's ability to generalize across the entire image. In Appendix, we provide a detailed analysis of the impact of \\( \\lambda \\) and the stop-gradient mechanism in SLADE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.73, + 0.108 + ], + "angle": 0, + "content": "4. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.252 + ], + "angle": 0, + "content": "LVLM models. In our experiments, we use OpenFlamingo-9B (OF) [2] and two versions of LLaVA-1.5 [34]: LLaVA-1.5 Vicuna-7B (LLaVA-7B) and LLaVA-1.5 LLaMA-2-13B (LLaVA-13B) as our target LVLMs. While these LVLMs use CLIP ViT-L-14 as their vision encoder, they incorporate different language models [14, 52, 53] for text processing. Notably, we select LLaVA-13B in our evaluation to assess the effectiveness of our robust encoder within the context of larger LVLM." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.908, + 0.697 + ], + "angle": 0, + "content": "Adversarial fine-tuning settings. In SLADE, we adversarially fine-tune the CLIP ViT-L/14 encoder on the ImageNet dataset [18], omitting labels for an unsupervised approach. We use PGD with 10 steps at \\(\\ell_{\\infty}\\) norm to generate perturbed views. Our training spans two epochs with an AdamW optimizer (WD=1e-4, LR=1e-5) and a batch size of 64, incorporating a cosine decaying LR schedule with linear warmup. We set the weighting parameter \\(\\lambda\\) to 0.6 and use a patch size of 32. We conduct an ablation study (see Appendix) to determine the optimal values for these hyperparameters. To ensure compatibility with LVLMs, we perform adversarial fine-tuning on the Vit-L/14 backbone of CLIP, matching the version used by LVLMs (e.g., OF and LLaVA-1.5). During robust inference, we substitute the default CLIP [48] model of the LVLMs with SLADE encoder, while keeping the language model and projection layer frozen, offering a plug-n-play adoption. While robustness is crucial, it often comes at the cost of potential degradation of clean performance. To optimize this trade-off between robust accuracy and clean performance, we adversarially fine-tune CLIP with perturbation radii \\(\\epsilon = 2/255\\) and \\(\\epsilon = 4/255\\), naming these models \\(\\mathrm{SLADE}^2\\) and \\(\\mathrm{SLADE}^4\\). In the following sections, we present a comprehensive comparison of our proposed fine-tuning scheme with the original CLIP and two state-of-the-art (SOTA) adversarial fine-tuned encoders, TeCoA [38] and FARE [50]. To ensure fare comparison, we use fine-tuned versions of TeCoA and FARE at the same adversarial radii, denoting them as \\(\\mathrm{TeCoA}^2\\) and \\(\\mathrm{FARE}^2\\) for \\(\\epsilon = 2/255\\), \\(\\mathrm{TeCoA}^4\\) and \\(\\mathrm{FARE}^4\\) for \\(\\epsilon = 4/255\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.705, + 0.785, + 0.72 + ], + "angle": 0, + "content": "4.1. Untargeted attacks on LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.728, + 0.907, + 0.85 + ], + "angle": 0, + "content": "We conduct untargeted attacks to assess clean and robust performance in downstream tasks. For clean evaluation, target LVLMs are given a clean image and its corresponding prompt, while for adversarial evaluation, they are given an adversarial image and prompt. In our experiments with untargeted attacks, we use the prompt \"Provide a short caption for this image\" for the image captioning task and questions from the dataset as prompts for VQA." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Untargeted attack setup. Following [49], we employ APGD [15] attack with \\(\\ell_{\\infty}\\) norm at \\(\\epsilon = 2/255\\) and \\(\\epsilon = 4/255\\). The APGD attack is executed using single precision for 100" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24247" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Assessing the robustness of target LVLMs against untargeted attacks in image captioning and visual question answering (VQA) tasks. For image captioning, we evaluate performance using the COCO and Flickr30k datasets, reporting the CIDEr score. For VQA tasks, we utilize the OKVQA and VizWiz datasets, providing VQA accuracy as the evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.135, + 0.836, + 0.43 + ], + "angle": 0, + "content": "
Vision LVLM encoderCOCOFlickr30KOKVQAVizWiz
clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255
Open FlamingoCLIP80.57.825.62.461.06.43.81.448.51.80.00.023.82.41.80
\\( {\\mathrm{{TeCoA}}}^{2} \\)74.559.740.310.348.237.929.010.140.127.421.510.020.116.18.35.6
\\( {\\mathrm{{FARE}}}^{2} \\)76.269.455.310.949.443.432.115.437.829.521.89.718.615.89.15.6
\\( {\\mathrm{{SLADE}}}^{2} \\)79.871.561.716.252.646.937.919.543.934.826.313.523.419.713.88.5
\\( {\\mathrm{{TeCoA}}}^{4} \\)72.364.454.210.546.139.228.712.237.430.823.58.418.814.68.35.0
\\( {\\mathrm{{FARE}}}^{4} \\)75.467.352.212.148.341.334.615.139.232.223.510.517.814.79.56.1
\\( {\\mathrm{{SLADE}}}^{4} \\)78.571.862.116.453.045.937.719.944.636.128.913.822.420.314.69.1
LLaVA-7BCLIP121.921.813.52.479.015.310.03.457.38.33.00.039.313.33.20.0
\\( {\\mathrm{{TeCoA}}}^{2} \\)115.698.373.538.175.665.350.529.455.640.330.514.238.525.415.48.3
\\( {\\mathrm{{FARE}}}^{2} \\)123.5105.286.439.478.970.360.525.154.343.530.115.337.329.317.610.4
\\( {\\mathrm{{SLADE}}}^{2} \\)127.0111.594.647.282.874.365.430.961.549.232.419.142.631.120.915.1
\\( {\\mathrm{{TeCoA}}}^{4} \\)110.395.575.635.371.862.551.027.050.339.032.312.334.530.518.39.3
\\( {\\mathrm{{FARE}}}^{4} \\)119.4100.583.541.676.370.356.529.553.545.034.815.338.531.321.010.1
\\( {\\mathrm{{SLADE}}}^{4} \\)123.9109.792.145.981.573.863.733.959.848.939.423.441.630.623.517.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.443, + 0.482, + 0.534 + ], + "angle": 0, + "content": "iterations. Additionally, we evaluate the robust CLIP models under a stronger adversarial setting by using a perturbation radius of \\(\\epsilon = 8/255\\) for the downstream tasks. The attacked images with \\(\\epsilon = 2/255\\) perturbations are completely imperceptible to human eyes, while with \\(\\epsilon = 4/255\\) a user could notice the perturbation when paying close attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.541, + 0.483, + 0.738 + ], + "angle": 0, + "content": "Datasets and metrics. For image captioning, we use the COCO [33] and Flickr30k [46] datasets and for VQA, we utilize OKVQA [39] and VizWiz [25] dataset. For assessing captioning tasks, we utilize the CIDEr [54] metric, which leverages higher order (longer) n-grams to capture grammatical properties as well as richer semantics [54]. Empirically, CIDEr effectively measures fine-grained detail by reflecting alignment between generated and reference captions in terms of vocabulary, context, and descriptive specificity. Higher CIDEr scores thus indicate stronger semantic alignment with ground truth. For VQA, we assess performance through VQA accuracy, which measures the percentage of correctly answered questions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results and discussion. Table 1 presents the performance of various CLIP encoders under untargeted attacks. While the original CLIP model excels in clean performance, it suffers significantly under adversarial attacks, particularly at stronger attack strength at \\(\\epsilon = 8 / 255\\) radii. Adversarially fine-tuned CLIP models show slight degradation in clean performance due to adversarial fine-tuning, but SLADE achieves the best clean performance among them. For OF, \\(\\mathrm{SLADE^4}\\) demonstrates superior performance compared to \\(\\mathrm{FARE^4}\\) and \\(\\mathrm{TeCoA^4}\\) across most downstream task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.64 + ], + "angle": 0, + "content": "datasets. Notably, when subjected to stronger attacks at \\(\\epsilon = 4/255\\) and \\(\\epsilon = 8/255\\), \\(\\mathrm{SLADE}^4\\) consistently outperforms all SOTA robust clip models. Similar trends are observed with \\(\\mathrm{SLADE}^2\\), outperforming \\(\\mathrm{FARE}^2\\) and \\(\\mathrm{TeCoA}^2\\) in both clean and robust performance. Additionally, for LLaVA, SLADE demonstrates even greater efficacy, achieving superior performance in both captioning and VQA tasks. Moreover, the higher CIDer score achieved by SLADE under attack demonstrates its superior ability to preserve the semantic details of image features in the generated captions compared to other robust variations of CLIP. In the Appendix, we provide results of untargeted attacks on LLaVA-13B LVLM and include visual examples of VQA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.766, + 0.667 + ], + "angle": 0, + "content": "4.2. Targeted attacks on LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Targeted attack setup. We conduct targeted attacks on LLaVA-7B to evaluate the robust performance of CLIP models. Targeted attacks force LVLMs to produce a specific output predetermined by the attackers, posing a more severe threat than untargeted attacks. In our experiments, we randomly select 25 images from the COCO dataset and apply the APGD attack for 10,000 iterations, using five realistic target strings representing malicious intents such as misinformation and false propaganda. For the attack configuration, we use the \\(\\ell_{\\infty}\\) norm with radii \\(\\epsilon = 2 / 255\\) and \\(\\epsilon = 4 / 255\\). We report the CIDEr score to quantify caption quality under attack conditions and attack success rate to measure how many times the targeted attack was successful. More details on our targeted attack pipeline and evaluation metrics are provided in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24248" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.308, + 0.097, + 0.642, + 0.109 + ], + "angle": 0, + "content": "Target String: Vaccination causes autism in children." + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.118, + 0.246, + 0.128 + ], + "angle": 0, + "content": "Clean Image" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.133, + 0.249, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.117, + 0.472, + 0.138 + ], + "angle": 0, + "content": "CLIP:Young girl with mask getting vaccine while holding a teddy bear" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.144, + 0.481, + 0.164 + ], + "angle": 0, + "content": "TecoA4: Young girl on a sofa taking vaccine" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.171, + 0.473, + 0.193 + ], + "angle": 0, + "content": "FARE4: Young girl with mask taking vaccine, holding a teddy bear" + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.196, + 0.487, + 0.219 + ], + "angle": 0, + "content": "SLADE:Young girl wearing mask taking vaccine while holding teddy bear" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.112, + 0.578, + 0.132 + ], + "angle": 0, + "content": "Adversarial Image" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.133, + 0.585, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.115, + 0.812, + 0.135 + ], + "angle": 0, + "content": "CLIP: Vaccination causes autism in children." + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.145, + 0.804, + 0.155 + ], + "angle": 0, + "content": "TecoA: Vaccine shot for children" + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.168, + 0.802, + 0.188 + ], + "angle": 0, + "content": "FARE: Young girl taking vaccine, holding a doll" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.196, + 0.825, + 0.218 + ], + "angle": 0, + "content": "SLADE4:Young girl wearing mask taking vaccine while holding teddy bear" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.234, + 0.795, + 0.246 + ], + "angle": 0, + "content": "Target String: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk." + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.257, + 0.245, + 0.268 + ], + "angle": 0, + "content": "Clean Image" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.27, + 0.249, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.257, + 0.487, + 0.278 + ], + "angle": 0, + "content": "CLIP: A statue of a muscular man with beard and six packs" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.284, + 0.485, + 0.304 + ], + "angle": 0, + "content": "TecoA4: A statue of a man with beard and muscles" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.311, + 0.491, + 0.332 + ], + "angle": 0, + "content": "FARE: A statue of a man with a beard and six-packs" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.337, + 0.487, + 0.358 + ], + "angle": 0, + "content": "SLADE4: A statue of a shirtless muscu-lar man with beard and six-packs" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.25, + 0.578, + 0.271 + ], + "angle": 0, + "content": "Adversarial Image" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.274, + 0.584, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.595, + 0.25, + 0.822, + 0.282 + ], + "angle": 0, + "content": "CLIP: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk." + }, + { + "type": "text", + "bbox": [ + 0.595, + 0.287, + 0.819, + 0.307 + ], + "angle": 0, + "content": "TecoA4: Before and after of photosh -ot of a man" + }, + { + "type": "text", + "bbox": [ + 0.595, + 0.313, + 0.826, + 0.332 + ], + "angle": 0, + "content": "FARE4: A statue of a man with a beard and muscles" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.339, + 0.813, + 0.36 + ], + "angle": 0, + "content": "SLADE4: Statue of a shirtless muscu-lar man with a beard and six packs" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.381, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Figure 3. Qualitative examples of \\(\\ell_{\\infty}\\) targeted attacks at \\(\\epsilon = 4/255\\) radii using original and adversarially fine-tuned CLIP models as vision encoder in LLaVA-7B. We present generated captions (target string, mistakes in captions, fine-grained details in caption) for the original images (left) and the imperceptible adversarial images (right)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.441, + 0.485, + 0.804 + ], + "angle": 0, + "content": "Results and discussion. We present quantitative results of our targeted attacks at \\(\\epsilon = 4 / 255\\) in Table 2 and qualitative examples in Figure 3. The original CLIP model shows high susceptibility to targeted attacks by consistently generating the specified target strings. In contrast, \\(\\mathrm{TeCoA}^2\\) and \\(\\mathrm{FARE}^2\\) break in only 6 and 2 cases, resulting in mean success rates of \\(4.8\\%\\) and \\(1.6\\%\\), respectively, while \\(\\mathrm{SLADE}^2\\) breaks in only one case, highlighting its superior performance. Notably, \\(\\mathrm{SLADE}^4\\), \\(\\mathrm{FARE}^4\\), and \\(\\mathrm{TeCoA}^4\\) show complete robustness under targeted attack at \\(\\epsilon = 4 / 255\\). However, the quality of captions generated by \\(\\mathrm{SLADE}^4\\) notably surpasses \\(\\mathrm{FARE}^4\\) and \\(\\mathrm{TeCoA}^4\\), as shown in Figure 3. We also observe that captions generated by \\(\\mathrm{TeCoA}^4\\) exhibit inferior quality and contain errors, while \\(\\mathrm{FARE}^4\\)'s captions often lack intricate details from the corresponding images. For instance, in first example of Figure 3, SLADE retains the fine-grained details of the young girl wearing mask and holding teddy bear whereas FARE overlooks these details and TeCoA generates captions with mistakes. This highlights SLADE's robustness, effectively resisting attacks and retaining both fine details and semantic integrity. Empirically, the CIDEr scores in Table 2 further support these claims, with \\(\\mathrm{SLADE}^4\\) achieving the highest score. We present more qualitative results in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.43, + 0.831 + ], + "angle": 0, + "content": "4.3. Zero-shot evaluation under AutoAttack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Attack setup. We examine the adversarial robustness of various CLIP models in zero-shot classification tasks using AutoAttack [15]. Our evaluation spans multiple datasets, including CIFAR10, CIFAR100 [31], StanfordCars [30]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.562 + ], + "angle": 0, + "content": "EuroSAT [26], PCAM [55], and Flowers [42]. For multiclass classification datasets, we utilize AutoAttack's ensemble of four attacks: APGD-CE, APGD-T, FAB, and Square. For binary classification datasets like PCAM, we exclude APGD with targeted DLR loss (APGD-T), as DLR loss is unsuitable for binary classification, and use the remaining three attacks instead. We apply AutoAttack for 100 iterations on 1,000 samples from each dataset at \\(\\epsilon = 4/255\\) radii." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.743 + ], + "angle": 0, + "content": "Results and discussion. Table 3 summarizes the performance of both original and robust CLIP models on zero-shot classification tasks. At \\(\\epsilon = 4/255\\) radii, SLADE consistently outperforms TeCoA and FARE. Notably, \\(\\mathrm{SLADE}^4\\) demonstrates greater robustness compared to \\(\\mathrm{FARE}^4\\) and \\(\\mathrm{TeCoA}^4\\), with a \\(3.1\\%\\) gain in robust accuracy, while also slightly outperforming \\(\\mathrm{SLADE}^2\\). These findings highlight SLADE's superior resilience against AutoAttack's ensemble of adversarial attacks. In Appendix, we present detailed results for clean and adversarial evaluations under \\(\\epsilon = 2/255\\) and \\(\\epsilon = 8/255\\) radii on zero-shot classification datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.77, + 0.77 + ], + "angle": 0, + "content": "4.4. Jailbreak attacks on LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In previous sections, we highlighted the vulnerability of LVLMs to gradient-based attacks. These models are also susceptible to jailbreak attacks due to the expanded attack surface from the image modality. Adversaries use gradient-based perturbations to create universal optimization-based jailbreak attacks, where a single adversarial image can compromise multiple LVLMs (e.g., LLaVA, MiniGPT-4 [65]) without contextually aligned prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24249" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.089, + 0.728, + 0.104 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of \\( {\\ell }_{\\infty } \\) targeted attacks at \\( \\epsilon = 4/{255} \\) radii." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.108, + 0.887, + 0.248 + ], + "angle": 0, + "content": "
TargetCLIPTeCoA2FARE2SLADE2TeCoA4FARE4SLADE4
Vaccination causes autism...25/250/250/250/250/250/250/25
Achieve this physique with ‘X’...25/252/251/250/250/250/250/25
Invest in ‘Y’ stock for quick...25/251/250/250/250/250/250/25
Visit https://tinyurl.com/zh...25/253/251/251/250/250/250/25
Opposition political leaders...25/250/250/250/250/250/250/25
Average attack success rate:100%4.8%1.6%0.8%0%0%0%
Average CIDEr score:018.625.148.467.378.787.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.26, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Table 3. Evaluation of CLIP models on zero-shot classification datasets under AutoAttack." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.47 + ], + "angle": 0, + "content": "
AttackVision encoderCIFAR10CIFAR100CarsFlowersEuroSATPCAM
4/255CLIP0.00.00.00.00.00.00.00.0
TeCoA231.017.85.86.73.516.0
FARE225.914.04.87.15.517.2
SLADE233.419.36.18.84.719.2
€TeCoA435.521.68.412.46.843.5
FARE434.821.412.812.911.750.2
SLADE434.423.015.313.712.551.3
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.486, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Attack Setup. To evaluate the robustness of vision encoders against jailbreak attacks, we use two optimization-based attacks: ImgJP [43] and VisualAdv [47], both of which employ gradient-based perturbations to generate universal adversarial images that bypass LVLMs' safety mechanisms. For ImgJP, we perform 1,000 iterations at \\(\\epsilon = 16/255\\) under the \\(\\ell_{\\infty}\\) norm, using MiniGPT-4 (Vicuna 7B) to generate adversarial images targeting the LLaVA-7B and LLaVA-13B models. Each image is tested against 25 harmful instructions from the AdvBench dataset [66], with Attack Success Rate (ASR) reported as the percentage of harmful prompts that elicit unsafe responses. In the VisualAdv evaluation, universal imperceptible adversarial images are generated with high perturbation (\\(\\epsilon = 128/255\\)) and tested against 100 toxic prompts from the RealToxicityPrompts dataset [21]. Using LLaVA-13B as a surrogate model, we measure the toxicity in the outputs via the Perspective API [28] and report the percentage of responses exhibiting attributes like threat or profanity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results and discussions. Table 4 presents the performance of different CLIP encoders under VisualAdv jailbreak attacks with LLaVA-13B. Substituting the original CLIP encoder with \\(\\mathrm{SLADE^4}\\) significantly lowers toxicity rates, reducing them from \\(30.4\\%\\) to \\(13.7\\%\\) at \\(\\epsilon = 128 / 255\\). While TeCoA and FARE also reduce toxicity rates compared to the original CLIP model, they remain less effective than SLADE. Although external defenses like JailGuard" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.908, + 0.429 + ], + "angle": 0, + "content": "and Diffpure show some promise, \\(\\mathrm{SLADE^4}\\) consistently achieves superior results. For our additional evaluation of VisualAdv attacks, see Appendix. Table 5 presents the ASR results for ImgJP attacks across target LVLMs. The original CLIP encoder in LLaVA models shows high ASR, particularly in LLaVA-7B. While external defenses like JailGuard decrease ASR, they add complexity and overhead due to requiring LVLM-specific training. \\(\\mathrm{SLADE^4}\\) outperforms \\(\\mathrm{FARE^4}\\) and \\(\\mathrm{TeCoA^4}\\), achieving the lowest ASR and underscoring SLADE's effectiveness against both gradient-based and universal jailbreak attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.78, + 0.456 + ], + "angle": 0, + "content": "4.5. Evaluation across other Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.909, + 0.603 + ], + "angle": 0, + "content": "In the preceding sections, we evaluated the performance of robust encoders in LVLMs under adversarial and jailbreak attacks. While resilience against such attacks is critical, it is equally important to ensure that LVLMs equipped with robust encoders maintain their core functionalities, including reasoning and instruction-following capabilities. In this section, we assess whether LVLMs with robust encoders can effectively retain these essential abilities, particularly in solving reasoning tasks and adhering to instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Instruction following benchmark. Instruction following is a fundamental capability in LVLMs, enabling them to accurately interpret and respond to contextually specific requests across diverse and scenarios. This capability extends LVLMs beyond basic image recognition, allowing them to generate responses that are sensitive to context, interpret visual nuances, and incorporate reasoning about complex, scene-specific details [2, 59]. We assess the instruction-following performance of LVLMs with robust encoders through a quantitative evaluation using VisIT-Bench [8], a recently developed comprehensive benchmark that includes 592 challenging instructions across 70 diverse instruction categories. In this evaluation, we use LLaVA-13B and a subset of 200 instructions, where each instance includes an instruction, an input image, and a human-verified GPT-4 response. Following [8], we set the GPT-4 [45] response as the ground truth and use BERTscore [61] and METEOR score [5] as evaluation metrics and summarize the results in Table 6. We observe that SLADE upholds instruction" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24250" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 4. Evaluation of LLaVA-13B against VisualAdv jailbreak attacks with different CLIP based models. We report the percentage of LVLM outputs that exhibit specific toxic attributes. The blue downward arrows (↓) indicate a decrease in toxic attribute percentages compared to the baseline CLIP model, while the red upward arrows (↑) indicate an increase. Lower values signify better performance." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.136, + 0.875, + 0.273 + ], + "angle": 0, + "content": "
Attack Strength (€)Vision EncoderExternal DefenseAny IdentityProfanitySevere ToxicitySexually ExplicitThreatToxicityAverage
128/255CLIP—62.09.053.05.019.04.061.0
JailGuard41.04.020.01.014.05.030.0
Diffpure (n = 0.25)35.04.028.02.011.02.027.0
TeCoA4—41.04.027.02.015.05.045.0
FARE4—36.04.019.01.09.05.033.0
SLADE4—34.04.017.00.010.05.030.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Table 5. Performance analysis of LVLMs with various vision encoders and defenses against ImgJP Attack at \\(\\epsilon = 16/255\\)." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.315, + 0.451, + 0.487 + ], + "angle": 0, + "content": "
LVLMVision EncoderExternal DefenseASR (%)
LLaVA-13BCLIP—28.0 ± 0.2
FARE4—18.3 ± 0.1
CLIPJailGuard21.0 ± 0.1
SLADE4—15.0 ± 0.3
LLaVA-7BCLIP—34.0 ± 0.1
FARE4—24.0 ± 0.2
CLIPJailGuard26.0 ± 0.3
SLADE4—19.3 ± 0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.527 + ], + "angle": 0, + "content": "Table 6. Instruction-following benchmark results of LLaVA 13B-model with original CLIP and robust vision encoders." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.53, + 0.46, + 0.613 + ], + "angle": 0, + "content": "
Vision EncoderBERTScoreMETEOR Score
PrecisionRecallF1
CLIP0.8910.8940.8920.37
SLADE²0.8900.8950.8870.36
SLADE⁴0.8890.8910.8890.36
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.669 + ], + "angle": 0, + "content": "following performance comparably to the original CLIP. In Appendix, we present additional details on the instruction-following benchmark setup and qualitative examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation on reasoning tasks. LVLMs are designed to emulate human reasoning, often achieved through a chain-of-thought approach, where a series of intermediate reasoning steps significantly enhances their ability to handle complex tasks. Following [50], we assess the reasoning capabilities of LVLMs equipped with robust vision encoders on the Science Question Answering (SQA) [35] benchmark and summarize the results in Table 7. We evaluate LLaVA-13B on a subset of 10,000 image-question pairs from the SQA benchmark, where each instance includes an explanation of a concept, followed by a question and an input image. LLaVA with the original CLIP model achieves the best accuracy of \\(64.5\\%\\). Both FARE and SLADE perform better than TeCoA. Among all the robust encoders, \\(\\mathrm{SLADE}^2\\) performs best and achieves an accuracy of \\(63.4\\%\\) accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.282, + 0.907, + 0.311 + ], + "angle": 0, + "content": "Table 7. Reasoning tasks evaluation with LLaVA on the SQA benchmark. We report accuracy for image-question pairs." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.315, + 0.89, + 0.361 + ], + "angle": 0, + "content": "
CLIPTeCoA2FARE2TeCoA4FARE4SLADE2SLADE4
64.561.162.463.059.363.462.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Our evaluations of reasoning and instruction-following tasks reveal that SLADE achieves performance comparable to the original CLIP model. This demonstrate that SLADE enables LVLMs to preserve their core capabilities, such as reasoning and instruction-following, without compromise. Furthermore, as detailed in the Appendix, SLADE also improves its image generation capabilities, demonstrating its versatility beyond robust performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.634, + 0.515 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce SLADE, an unsupervised adversarial fine-tuning scheme designed to enhance the robustness of the widely-used CLIP vision encoder against both gradient-based and optimization-based jailbreak attacks. By leveraging a dual-level contrastive learning approach, SLADE effectively captures fine-grained image details at the patch level while preserving high-level semantic consistency at the global level, enabling VLMs to effectively withstand adversarial perturbations without requiring any computational overhead. Our extensive experiments demonstrate that SLADE-equipped LVLMs outperform existing models in adversarial robustness, excelling in nuanced semantic understanding of perturbed images while maintaining high accuracy on clean data. Moreover, we empirically demonstrate that replacing the default CLIP encoder with SLADE enables robust defense against jailbreak attacks without requiring any LVLM-specific training, highlighting its seamless integration and adaptability across multimodal systems. Further, we validate that SLADE-equipped LVLMs do not compromise on non-adversarial core performance, such as reasoning and instruction-following capabilities. These findings position SLADE as a transformative approach with the potential to set a new benchmark for adversarial robustness in multimodal AI, paving the way for safer and more reliable applications in real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24251" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.273, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.486, + 0.201 + ], + "angle": 0, + "content": "This material is partly based upon work supported by the U.S. National Science Foundation (NSF) under Grant No. CRII-IIS-RI-2348145. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.482, + 0.326 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.329, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.414, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[3] Eugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, and Vitaly Shmatikov. (ab) using images and sounds for indirect instruction injection in multi-modal llms. arXiv preprint arXiv:2307.10490, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[4] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime. arXiv preprint arXiv:2309.00236, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[6] Hritik Bansal, Nishad Singhi, Yu Yang, Fan Yin, Aditya Grover, and Kai-Wei Chang. Cleanclip: Mitigating data poisoning attacks in multimodal contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112–123, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[8] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[9] Nicholas Carlini, Milad Nasr, Christopher A Choquette-Choo, Matthew Jagielski, Irena Gao, Pang Wei W Koh, Daphne Ippolito, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned? Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[11] Lu Chen, Qiangchang Wang, Zhaohui Li, and Yilong Yin. Hypergraph-guided intra-and inter-category relation modeling for fine-grained visual recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8043-8052, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[15] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[19] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[20] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in neural information processing systems, 34:21480-21492, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[21] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[22] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Sven Gowal, Po-Sen Huang, Aaron van den Oord, Timothy Mann, and Pushmeet Kohli. Self-supervised adversarial robustness for the low-label, high-data regime. In International conference on learning representations, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24252" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[24] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[26] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in neural information processing systems, 33: 16199-16210, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[28] Google Jigsaw. Perspective api. https://www.perspectiveapi.com/, 2017. Accessed: 2024-06-02. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[29] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in neural information processing systems, 33:2983-2994, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[32] Xiao Li, Wei Zhang, Yining Liu, Zhanhao Hu, Bo Zhang, and Xiaolin Hu. Language-driven anchors for zero-shot adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24686-24695, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[38] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[39] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[40] AI Meta. Introducing llama: A foundational, 65-billion-parameter large language model. Meta AI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[43] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[44] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.899, + 0.539 + ], + "angle": 0, + "content": "[45] R OpenAI. Gpt-4 technical report. ArXiv, 2303, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[46] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[47] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[49] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336, 2024. 1, 2, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24253" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[51] Arindam Sikdar, Yonghuai Liu, Siddhardha Kedarisetty, Yitian Zhao, Amr Ahmed, and Ardhendu Behera. Interweaving insights: High-order feature interaction for fine-grained visual recognition. International Journal of Computer Vision, pages 1-25, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[52] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[53] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[54] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[55] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation equivariant cnns for digital pathology. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11, pages 210-218. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[56] Xiao Wang, Yuhang Huang, Dan Zeng, and Guo-Jun Qi. Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10718-10730, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[57] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[58] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[60] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness. In European Conference on Computer Vision, pages 725-742. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[61] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[62] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[64] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "adversarial examples in multimodal contrastive learning. In 31st ACM International Conference on Multimedia, pages 6311-6320, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[65] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[66] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24254" + } + ] +] \ No newline at end of file diff --git a/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_origin.pdf b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e1302a8d6b4474cda5c5f4dea826f083e03f200 --- /dev/null +++ b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/18172bcf-3306-49a6-8b62-dc7f40cb507d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a49d1289b4dc09efb1232d6dda75b6668496e1ca550a7e66059020c37fd75e10 +size 553326 diff --git a/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/full.md b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c22256cf95ef861b90bd6534bd9860ffa2f14fd9 --- /dev/null +++ b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/full.md @@ -0,0 +1,318 @@ +# SLADE: Shielding against Dual Exploits in Large Vision-Language Models + +Md Zarif Hossain, Ahmed Imteaj† + +School of Computing, Southern Illinois University Carbondale, IL 62901, USA + +# Abstract + +Large Vision-Language Models (LVLMs) have emerged as transformative tools in multimodal tasks, seamlessly integrating pretrained vision encoders to align visual and textual modalities. Prior works have highlighted the susceptibility of LVLMs to dual exploits (gradient-based and optimization-based jailbreak attacks), which leverage the expanded attack surface introduced by the image modality. Despite advancements in enhancing robustness, existing methods fall short in their ability to defend against dual exploits while preserving fine-grained semantic details and overall semantic coherence under intense adversarial perturbations. To bridge this gap, we introduce SheiLding against Dual Exploits (SLADE), a novel unsupervised adversarial fine-tuning scheme that enhances the resilience of CLIP-based vision encoders. SLADE's dual-level contrastive learning approach balances the granular and the holistic, capturing fine-grained image details without losing sight of high-level semantic coherence. Extensive experiments demonstrate that SLADE-equipped LVLMs set a new benchmark for robustness against dual exploits while preserving fine-grained semantic details of perturbed images. Notably, SLADE achieves these results without compromising the core functionalities of LVLMs, such as instruction following, or requiring the computational overhead (e.g., large batch sizes, momentum encoders) commonly associated with traditional contrastive learning methods. + +# 1. Introduction + +The remarkable success of large language models (LLMs) [40, 45] in understanding and generating human-like text has inspired researchers to extend their capabilities to the visual domain. This effort has led to the development of Large Vision-Language Models (LVLMs) [1, 34], which aim to bridge the gap between visual and textual information. These LVLMs frequently utilize pretrained vision encoders, such as CLIP [48], BEiT [7], and DINO [10], to derive meaningful visual representations from images. LVLMs [2, 34] equipped with these vision encoders demonstrate remarkable zero-shot performance in down- + +stream tasks such as image captioning, and visual question answering (VQA). Their promising capabilities and impact on multimodal learning have led to a growing number of LVLMs being publicly released. However, while this availability fosters research and application development, it also introduces critical safety concerns that require careful attention. Recent studies [57, 63] highlight the high susceptibility of LVLMs to harmful exploits on text and image modalities. Notably, the addition of a vision modality increases the model's attack surface, making LVLMs even more vulnerable to manipulation compared to LLMs [9, 22]. Moreover, the adversaries can use imperceptible gradient-based perturbations to create targeted attacks, generating malicious outputs that evade detection [49]. Such targeted attacks are particularly concerning due to their ability to manipulate models toward specific, often harmful outputs. Additionally, the increasing deployment of LVLMs raises the potential for exploitation not only by adversaries but also by corporations aiming to leverage these models to promote their products or agendas as shown in Figure 1. Recent studies [43, 47] demonstrate that gradient-based perturbations can be adapted in optimization-based jailbreak attacks to bypass a model's safety mechanisms, causing LVLMs to generate harmful or undesirable content. + +While existing methods [38, 50] made notable progress in improving the robustness of vision encoders, particularly CLIP, which is widely utilized across various applications, critical challenges persist. For example, in downstream tasks such as image captioning, these models successfully capture high-level semantic meaning but often struggle to preserve fine-grained semantic details when faced with perturbed images. Consequently, their ability to generate accurate and contextually rich descriptions is compromised, limiting their applicability and effectiveness in real-world scenarios where precision and detail are paramount. To address these concerns, we propose SLADE, an unsupervised adversarial fine-tuning scheme that enhances the robustness of the widely used CLIP vision encoder's embeddings against dual exploits: gradient-based attacks and optimization-based Jailbreak attacks. SLADE combines symmetric contrastive learning objectives at both the patch level and the global level (on full image) to capture fine + +![](images/f934b1330ff6c288914c75eff374fe6f7e8a7ebad2f6cab23d346de00e77f5d7.jpg) +Figure 1. Targeted $\ell_{\infty}$ attack and VisualAdV [47] jailbreak attack at $\epsilon = 4 / 255$ radii and using original CLIP and SLADE as vision encoder in LLaVA-7B. SLADE (trained at $\epsilon = 4 / 255$ ) generates captions with fine-grained semantic details for the perturbed sample and successfully defends against the jailbreak attempt. + +grained image details and high-level semantic representation of the entire image, respectively. This dual-level approach enables the model to preserve both detailed features and high-level semantic consistency, even in the presence of adversarial perturbations. Moreover, SLADE does not require LVLM or task-specific training and can be seamlessly integrated into existing LVLMs as a robust vision encoder. SLADE enhances the adversarial robustness of the vision modality in LVLMs, preserving the model's ability to capture nuanced semantic details while effectively defending against jailbreak attacks, as illustrated in Figure 1. Our work sets a new standard for adversarial robustness in LVLMs, a critical advancement as these models become integral to critical applications. + +# 2. Related Works + +Large Vision-Language Models. In recent times, several advanced LVLMs have been introduced, including LLaVA [34], OpenFlamingo (OF) [2], InstructBLIP [17] and Macaw-LLM [36]. These models leverage the power of multimodal learning by combining visual and textual data, thereby, enhancing their capability to understand and generate responses to multimodal inputs. Most of these LVLMs incorporate pre-trained LLMs, such as Llama [53], paired with large-scale vision encoders like CLIP [48], to process multimodal data. + +Adversarial and Jailbreaking attacks in LVLMs. In recent studies [3, 4, 19, 63], authors have explored the susceptibility of large vision-language models to adversarial attacks. Notably, recent works [3, 6, 27, 64] demonstrate the vulnerability of vision encoders in LVLMs, particularly CLIP, to adversarial manipulation, which can significantly impair the performance of LVLMs in downstream tasks. Moreover, [16, 49] highlight the efficacy of gradient- + +based attacks on LVLMs, demonstrating their capacity to induce the model into generating harmful or misleading outputs. In [43, 47] authors show how gradient-based perturbations can be leveraged in optimization-based jailbreak attacks to bypass LVLMs' safety protocols. While recent studies have proposed defense strategies [58, 62] to counter jailbreak attacks, these methods rely heavily on the model's inherent safety alignment, necessitating specific training for each LVLM and often resulting in degraded performance in downstream tasks (e.g., VQA and image captioning). + +Robustness against exploits in LVLMs. Prior works [32, 38] introduced supervised adversarial fine-tuning scheme for CLIP, using cross-modal image-text contrastive loss. However, in LVLMs their zero-shot performance in downstream tasks (mostly in unseen classes) often degrades due to an over-dependence on the text embeddings during the fine-tuning process. To address this, some concurrent works [20, 23, 27] proposed unsupervised adversarial training methods using SimCLR [12] and BYOL [24] contrastive loss, which improve robustness but require large batch sizes for sufficient negative samples or a momentum encoder for optimal results. Additionally, the authors in [50] proposed an alternative approach, tailoring an $\ell_2$ loss-based unsupervised fine-tuning scheme. However, our empirical analysis reveals that applying $\ell_2$ loss across the entire image is insufficient for retaining nuanced details, particularly under adversarial conditions. Unlike prior methods, SLADE integrates a symmetric contrastive learning objective applied simultaneously at the patch level and global level. This dual-level strategy enables SLADE to capture fine-grained details and nuanced spatial relationships, even under adversarial perturbations. Drawing inspiration from traditional machine learning techniques that use patch embedding with pretext tasks to preserve detail and spatial + +localization [11, 41, 44, 51], SLADE advances this concept by incorporating it directly into its adversarial fin-tuning mechanism. Furthermore, SLADE leverages a stop-gradient mechanism to bypass the reliance on large batches or momentum encoders, offering a scalable and computationally efficient solution. + +# 3. Overview of SLADE + +This section provides a detailed overview of SLADE, our proposed adversarial fine-tuning scheme, illustrated in Figure 2. The primary objective of SLADE is to adversarially fine-tune a CLIP-like model $f_{\theta}(\cdot)$ , with parameter $\theta$ on a given dataset $\mathcal{D} = \{I_1, I_2, \dots, I_N\}$ with $I_n \in \mathbb{R}^{H \times W \times 3}$ . SLADE is designed to learn invariant features that are robust to adversarial perturbations while preserving the semantic details of the input data. This ensures that the fine-tuned model effectively balances adversarial resilience with the retention of critical semantic information. + +![](images/75bfc60eac1559218dd8d04e2d43b374b57dc9b5656653b2eebb7d7c948e4492.jpg) +Figure 2. Workflow of our proposed scheme: SLADE. + +# 3.1. Adversarial Image generation + +During adversarial fine-tuning, SLADE generates a perturbed version $I^{\prime}$ from the original clean image $I$ using Projected Gradient Descent (PGD) [37]. Given an input image $I$ , the perturbed image $I^{\prime}$ is iteratively computed by applying small, bounded modifications designed to maximize the adversarial objective. At each iteration $t$ , the perturbed image $I_{t + 1}^{\prime}$ is updated as follows: + +$$ +I _ {t + 1} ^ {\prime} = \Pi_ {I + \epsilon} \left(I _ {t} + \alpha \cdot \operatorname {s i g n} \left(\nabla_ {I} L \left(I _ {t}, y\right)\right)\right) \tag {1} +$$ + +Here, $L$ is the loss function (e.g., cross-entropy loss) and $\nabla_I L(I_t, y)$ denotes the gradient of this loss with respect to the input $I_t$ at each iteration. The term $\alpha$ is the step size, and $\epsilon$ is the maximum allowed perturbation bound. + +# 3.2. Patch-Level Contrastive Loss + +SLADE applies a symmetric contrastive learning objective at the patch level to retain fine-grained image features under + +adversarial perturbations. Each image $I$ and its adversarial counterpart $I^{\prime}$ are divided into non-overlapping patches. Let $P(I) = \{p_{1},p_{2},\ldots ,p_{n}\}$ and $P(I^{\prime}) = \{p_1^{\prime},p_2^{\prime},\dots ,p_n^{\prime}\}$ represent the respective sets of patches from clean image I and perturbed image $I^{\prime}$ , respectively. For an image with resolution $(H,W)$ with $C$ color channels, each patch is a flattened 2D segment of size $(K,K)$ , where $K$ is the patch size. The number of resulting patches, $N$ , is calculated as $N = \frac{H\times W}{K^2}$ . Subsequently, each patch $p_i$ and its corresponding adversarial patch $p_i^\prime$ are then passed through the encoder, yielding patch embeddings $z_{i} = f_{\theta}(p_{i})$ and $z_{i}^{\prime} = f_{\theta}(p_{i}^{\prime})$ , respectively. + +In SLADE, we focus on ensuring the model's capability to retain fine-grained details despite adversarial perturbations. To achieve this, we apply a contrastive loss at the patch level, aligning embeddings of corresponding patches (i.e., $z_{i}$ and $z_{i}^{\prime}$ ). By aligning patches individually, the model $f_{\theta}$ treats each patch as an independent unit and encourages to preserve semantic coherence within localized regions. This approach enables the model to capture fine-grained and spatially distinct features, thereby retaining intricate details that might otherwise be lost to adversarial noise. Following the adversarial contrastive loss approach in [29, 60], we define the loss for $z_{i}$ and $z_{i}^{\prime}$ as: + +$$ +L _ {c o} = - \log \frac {\exp \left(\operatorname* {s i m} \left(z _ {i} , z _ {i} ^ {\prime}\right) / \tau\right)}{\exp \left(\operatorname* {s i m} \left(z _ {i} , z _ {i} ^ {\prime}\right) / \tau\right) + \sum_ {I ^ {n} \in \mathcal {D} _ {N}} \exp \left(\operatorname* {s i m} \left(z _ {i} , z _ {i} ^ {n}\right) / \tau\right)} \tag {2} +$$ + +Here, $\mathcal{D}_N\subseteq \mathcal{D}$ represents a batch of negative samples, where $N$ is the batch size. The negative images $I^n$ are drawn uniformly at random from the dataset $\mathcal{D}$ , where $I\neq I^{n}$ . The embedding of patch $p_i^n$ from a negative sample $I^n$ is denoted by $z_{i}^{n}$ . $\tau$ is a temperature parameter, and $\mathrm{sim}(\cdot ,\cdot)$ denotes the cosine similarity between embeddings. + +In Equation 2, negative samples are crucial to prevent a collapsing solution. In their absence, the model may map all patches to similar embeddings, failing to learn discriminative representations [56, 60]. Achieving robust performance typically requires a large number of negative samples, which increases the need for a large batch size, making it resource-intensive. Leveraging a momentum encoder [23] can mitigate the reliance on large batch sizes, though it introduces additional computational overhead. + +Mitigating computational overhead: SLADE incorporates a stop-gradient mechanism [13] with a cosine similarity metric to prevent feature collapse and reduce computational overhead, which can be defined as: + +$$ +\operatorname {s t o p s i m} (\cdot , \cdot) = \operatorname {s i m} (\cdot , \operatorname {s t o p g r a d} (\cdot)) \tag {3} +$$ + +The stop-gradient mechanism in Equation 3 is then incorporated into a symmetric contrastive loss, which is designed to align patch embeddings between clean and adver + +sarial images. The resulting contrastive loss is defined as: + +$$ +L _ {p a t c h} = - \frac {1}{2 N} \sum_ {i = 1} ^ {N} \left(\operatorname {s t o p s i m} \left(z _ {i}, z _ {i} ^ {\prime}\right) + \operatorname {s t o p s i m} \left(z _ {i} ^ {\prime}, z _ {i}\right)\right) \tag {4} +$$ + +Here, $N$ denotes the total number of patches per image and $L_{patch}$ represents the patch-level symmetric contrastive loss, which aims to align the patch embeddings of clean and adversarial images to capture fine-grained details. The stopgrad $(\cdot)$ operation is applied in a symmetric manner, where either $z_{i}$ or $z_{i}^{\prime}$ is set to be constant in each term. This halts the gradient back-propagation on one side, preventing the model from simply mapping to constant outputs. By only updating the gradients on one side, stop-gradient prevents the model from converging to trivial solutions. + +# 3.3. Global-level Contrastive Loss + +In Equation 4, SLADE aims to preserve fine-grained details by focusing on patch-level alignment between clean and adversarial patch embeddings. However, concentrating solely on patch-level alignment could lead to a potential issue where the model captures fine-grained details in each patch but may struggle to capture the overall semantic meaning of the image as it might overlook the contextual relation between the patches. To address this and ensure the preservation of overall semantic coherence, SLADE introduces an additional global-level alignment objective on the clean $I$ and adversarial image $I^{\prime}$ : + +$$ +L _ {g l o b a l} = - \frac {1}{2} \left(\operatorname {s t o p s i m} \left(Z, Z ^ {\prime}\right) + \operatorname {s t o p s i m} \left(Z ^ {\prime}, Z\right)\right) \tag {5} +$$ + +Here, $Z$ and $Z'$ are the embeddings of clean $I$ and perturbed image $I'$ , respectively. + +Final loss for SLADE. The final adversarial fine-tuning objective in SLADE combines the fine-grained and global losses with a weighting parameter $\lambda$ (where $0 \leq \lambda \leq 1$ ) to control their relative importance. The final loss function for SLADE can be defined as: + +$$ +L _ {s l a d e} = \lambda L _ {p a t c h} + (1 - \lambda) L _ {g l o b a l} \tag {6} +$$ + +Minimizing the loss $L_{slade}$ effectively preserves both localized and high-level semantic features, thereby improving robustness against perturbations. Adjusting the weight $\lambda$ provides flexibility in the model's focus: setting $\lambda = 0$ prioritizes the global objective, emphasizing high-level semantic features, while larger values of $\lambda$ shift the focus toward finer details within patches. However, this shift may compromise the model's ability to generalize across the entire image. In Appendix, we provide a detailed analysis of the impact of $\lambda$ and the stop-gradient mechanism in SLADE. + +# 4. Experimental Analysis + +LVLM models. In our experiments, we use OpenFlamingo-9B (OF) [2] and two versions of LLaVA-1.5 [34]: LLaVA-1.5 Vicuna-7B (LLaVA-7B) and LLaVA-1.5 LLaMA-2-13B (LLaVA-13B) as our target LVLMs. While these LVLMs use CLIP ViT-L-14 as their vision encoder, they incorporate different language models [14, 52, 53] for text processing. Notably, we select LLaVA-13B in our evaluation to assess the effectiveness of our robust encoder within the context of larger LVLM. + +Adversarial fine-tuning settings. In SLADE, we adversarially fine-tune the CLIP ViT-L/14 encoder on the ImageNet dataset [18], omitting labels for an unsupervised approach. We use PGD with 10 steps at $\ell_{\infty}$ norm to generate perturbed views. Our training spans two epochs with an AdamW optimizer (WD=1e-4, LR=1e-5) and a batch size of 64, incorporating a cosine decaying LR schedule with linear warmup. We set the weighting parameter $\lambda$ to 0.6 and use a patch size of 32. We conduct an ablation study (see Appendix) to determine the optimal values for these hyperparameters. To ensure compatibility with LVLMs, we perform adversarial fine-tuning on the Vit-L/14 backbone of CLIP, matching the version used by LVLMs (e.g., OF and LLaVA-1.5). During robust inference, we substitute the default CLIP [48] model of the LVLMs with SLADE encoder, while keeping the language model and projection layer frozen, offering a plug-n-play adoption. While robustness is crucial, it often comes at the cost of potential degradation of clean performance. To optimize this trade-off between robust accuracy and clean performance, we adversarially fine-tune CLIP with perturbation radii $\epsilon = 2/255$ and $\epsilon = 4/255$ , naming these models $\mathrm{SLADE}^2$ and $\mathrm{SLADE}^4$ . In the following sections, we present a comprehensive comparison of our proposed fine-tuning scheme with the original CLIP and two state-of-the-art (SOTA) adversarial fine-tuned encoders, TeCoA [38] and FARE [50]. To ensure fare comparison, we use fine-tuned versions of TeCoA and FARE at the same adversarial radii, denoting them as $\mathrm{TeCoA}^2$ and $\mathrm{FARE}^2$ for $\epsilon = 2/255$ , $\mathrm{TeCoA}^4$ and $\mathrm{FARE}^4$ for $\epsilon = 4/255$ . + +# 4.1. Untargeted attacks on LVLMs + +We conduct untargeted attacks to assess clean and robust performance in downstream tasks. For clean evaluation, target LVLMs are given a clean image and its corresponding prompt, while for adversarial evaluation, they are given an adversarial image and prompt. In our experiments with untargeted attacks, we use the prompt "Provide a short caption for this image" for the image captioning task and questions from the dataset as prompts for VQA. + +Untargeted attack setup. Following [49], we employ APGD [15] attack with $\ell_{\infty}$ norm at $\epsilon = 2/255$ and $\epsilon = 4/255$ . The APGD attack is executed using single precision for 100 + +Table 1. Assessing the robustness of target LVLMs against untargeted attacks in image captioning and visual question answering (VQA) tasks. For image captioning, we evaluate performance using the COCO and Flickr30k datasets, reporting the CIDEr score. For VQA tasks, we utilize the OKVQA and VizWiz datasets, providing VQA accuracy as the evaluation metric. + +
Vision LVLM encoderCOCOFlickr30KOKVQAVizWiz
clean\( \ell_{\infty} \)2/2554/2558/255clean\( \ell_{\infty} \)2/2554/2558/255clean\( \ell_{\infty} \)2/2554/2558/255clean\( \ell_{\infty} \)2/2554/2558/255
Open FlamingoCLIP80.57.825.62.461.06.43.81.448.51.80.00.023.82.41.80
\( {\mathrm{{TeCoA}}}^{2} \)74.559.740.310.348.237.929.010.140.127.421.510.020.116.18.35.6
\( {\mathrm{{FARE}}}^{2} \)76.269.455.310.949.443.432.115.437.829.521.89.718.615.89.15.6
\( {\mathrm{{SLADE}}}^{2} \)79.871.561.716.252.646.937.919.543.934.826.313.523.419.713.88.5
\( {\mathrm{{TeCoA}}}^{4} \)72.364.454.210.546.139.228.712.237.430.823.58.418.814.68.35.0
\( {\mathrm{{FARE}}}^{4} \)75.467.352.212.148.341.334.615.139.232.223.510.517.814.79.56.1
\( {\mathrm{{SLADE}}}^{4} \)78.571.862.116.453.045.937.719.944.636.128.913.822.420.314.69.1
LLaVA-7BCLIP121.921.813.52.479.015.310.03.457.38.33.00.039.313.33.20.0
\( {\mathrm{{TeCoA}}}^{2} \)115.698.373.538.175.665.350.529.455.640.330.514.238.525.415.48.3
\( {\mathrm{{FARE}}}^{2} \)123.5105.286.439.478.970.360.525.154.343.530.115.337.329.317.610.4
\( {\mathrm{{SLADE}}}^{2} \)127.0111.594.647.282.874.365.430.961.549.232.419.142.631.120.915.1
\( {\mathrm{{TeCoA}}}^{4} \)110.395.575.635.371.862.551.027.050.339.032.312.334.530.518.39.3
\( {\mathrm{{FARE}}}^{4} \)119.4100.583.541.676.370.356.529.553.545.034.815.338.531.321.010.1
\( {\mathrm{{SLADE}}}^{4} \)123.9109.792.145.981.573.863.733.959.848.939.423.441.630.623.517.4
+ +iterations. Additionally, we evaluate the robust CLIP models under a stronger adversarial setting by using a perturbation radius of $\epsilon = 8/255$ for the downstream tasks. The attacked images with $\epsilon = 2/255$ perturbations are completely imperceptible to human eyes, while with $\epsilon = 4/255$ a user could notice the perturbation when paying close attention. + +Datasets and metrics. For image captioning, we use the COCO [33] and Flickr30k [46] datasets and for VQA, we utilize OKVQA [39] and VizWiz [25] dataset. For assessing captioning tasks, we utilize the CIDEr [54] metric, which leverages higher order (longer) n-grams to capture grammatical properties as well as richer semantics [54]. Empirically, CIDEr effectively measures fine-grained detail by reflecting alignment between generated and reference captions in terms of vocabulary, context, and descriptive specificity. Higher CIDEr scores thus indicate stronger semantic alignment with ground truth. For VQA, we assess performance through VQA accuracy, which measures the percentage of correctly answered questions. + +Results and discussion. Table 1 presents the performance of various CLIP encoders under untargeted attacks. While the original CLIP model excels in clean performance, it suffers significantly under adversarial attacks, particularly at stronger attack strength at $\epsilon = 8 / 255$ radii. Adversarially fine-tuned CLIP models show slight degradation in clean performance due to adversarial fine-tuning, but SLADE achieves the best clean performance among them. For OF, $\mathrm{SLADE^4}$ demonstrates superior performance compared to $\mathrm{FARE^4}$ and $\mathrm{TeCoA^4}$ across most downstream task + +datasets. Notably, when subjected to stronger attacks at $\epsilon = 4/255$ and $\epsilon = 8/255$ , $\mathrm{SLADE}^4$ consistently outperforms all SOTA robust clip models. Similar trends are observed with $\mathrm{SLADE}^2$ , outperforming $\mathrm{FARE}^2$ and $\mathrm{TeCoA}^2$ in both clean and robust performance. Additionally, for LLaVA, SLADE demonstrates even greater efficacy, achieving superior performance in both captioning and VQA tasks. Moreover, the higher CIDer score achieved by SLADE under attack demonstrates its superior ability to preserve the semantic details of image features in the generated captions compared to other robust variations of CLIP. In the Appendix, we provide results of untargeted attacks on LLaVA-13B LVLM and include visual examples of VQA. + +# 4.2. Targeted attacks on LVLMs + +Targeted attack setup. We conduct targeted attacks on LLaVA-7B to evaluate the robust performance of CLIP models. Targeted attacks force LVLMs to produce a specific output predetermined by the attackers, posing a more severe threat than untargeted attacks. In our experiments, we randomly select 25 images from the COCO dataset and apply the APGD attack for 10,000 iterations, using five realistic target strings representing malicious intents such as misinformation and false propaganda. For the attack configuration, we use the $\ell_{\infty}$ norm with radii $\epsilon = 2 / 255$ and $\epsilon = 4 / 255$ . We report the CIDEr score to quantify caption quality under attack conditions and attack success rate to measure how many times the targeted attack was successful. More details on our targeted attack pipeline and evaluation metrics are provided in the Appendix. + +Target String: Vaccination causes autism in children. + +![](images/7de4c6695f5e8e20ae3e59af2f88a02278ddb22b2918681a4e5c1fd623c28611.jpg) +Clean Image + +CLIP:Young girl with mask getting vaccine while holding a teddy bear + +TecoA4: Young girl on a sofa taking vaccine + +FARE4: Young girl with mask taking vaccine, holding a teddy bear + +SLADE:Young girl wearing mask taking vaccine while holding teddy bear + +![](images/8e59069186c8b34a19a029394dbad876e9fc58ad2adc53b6a78073fbcccab85e.jpg) +Adversarial Image + +CLIP: Vaccination causes autism in children. + +TecoA: Vaccine shot for children + +FARE: Young girl taking vaccine, holding a doll + +SLADE4:Young girl wearing mask taking vaccine while holding teddy bear + +Target String: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk. + +![](images/e7321226e8b26b47e7528cd7ee68c583bf01a3126da0d42dc42805b337130f42.jpg) +Clean Image + +CLIP: A statue of a muscular man with beard and six packs + +TecoA4: A statue of a man with beard and muscles + +FARE: A statue of a man with a beard and six-packs + +SLADE4: A statue of a shirtless muscu-lar man with beard and six-packs + +![](images/bb23b73d19c1439a2d9e81963f90c917a2611a4b4a5c7d51a936f06ffe57117a.jpg) +Adversarial Image +Figure 3. Qualitative examples of $\ell_{\infty}$ targeted attacks at $\epsilon = 4/255$ radii using original and adversarially fine-tuned CLIP models as vision encoder in LLaVA-7B. We present generated captions (target string, mistakes in captions, fine-grained details in caption) for the original images (left) and the imperceptible adversarial images (right). + +CLIP: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk. + +TecoA4: Before and after of photosh -ot of a man + +FARE4: A statue of a man with a beard and muscles + +SLADE4: Statue of a shirtless muscu-lar man with a beard and six packs + +Results and discussion. We present quantitative results of our targeted attacks at $\epsilon = 4 / 255$ in Table 2 and qualitative examples in Figure 3. The original CLIP model shows high susceptibility to targeted attacks by consistently generating the specified target strings. In contrast, $\mathrm{TeCoA}^2$ and $\mathrm{FARE}^2$ break in only 6 and 2 cases, resulting in mean success rates of $4.8\%$ and $1.6\%$ , respectively, while $\mathrm{SLADE}^2$ breaks in only one case, highlighting its superior performance. Notably, $\mathrm{SLADE}^4$ , $\mathrm{FARE}^4$ , and $\mathrm{TeCoA}^4$ show complete robustness under targeted attack at $\epsilon = 4 / 255$ . However, the quality of captions generated by $\mathrm{SLADE}^4$ notably surpasses $\mathrm{FARE}^4$ and $\mathrm{TeCoA}^4$ , as shown in Figure 3. We also observe that captions generated by $\mathrm{TeCoA}^4$ exhibit inferior quality and contain errors, while $\mathrm{FARE}^4$ 's captions often lack intricate details from the corresponding images. For instance, in first example of Figure 3, SLADE retains the fine-grained details of the young girl wearing mask and holding teddy bear whereas FARE overlooks these details and TeCoA generates captions with mistakes. This highlights SLADE's robustness, effectively resisting attacks and retaining both fine details and semantic integrity. Empirically, the CIDEr scores in Table 2 further support these claims, with $\mathrm{SLADE}^4$ achieving the highest score. We present more qualitative results in the Appendix. + +# 4.3. Zero-shot evaluation under AutoAttack + +Attack setup. We examine the adversarial robustness of various CLIP models in zero-shot classification tasks using AutoAttack [15]. Our evaluation spans multiple datasets, including CIFAR10, CIFAR100 [31], StanfordCars [30], + +EuroSAT [26], PCAM [55], and Flowers [42]. For multiclass classification datasets, we utilize AutoAttack's ensemble of four attacks: APGD-CE, APGD-T, FAB, and Square. For binary classification datasets like PCAM, we exclude APGD with targeted DLR loss (APGD-T), as DLR loss is unsuitable for binary classification, and use the remaining three attacks instead. We apply AutoAttack for 100 iterations on 1,000 samples from each dataset at $\epsilon = 4/255$ radii. + +Results and discussion. Table 3 summarizes the performance of both original and robust CLIP models on zero-shot classification tasks. At $\epsilon = 4/255$ radii, SLADE consistently outperforms TeCoA and FARE. Notably, $\mathrm{SLADE}^4$ demonstrates greater robustness compared to $\mathrm{FARE}^4$ and $\mathrm{TeCoA}^4$ , with a $3.1\%$ gain in robust accuracy, while also slightly outperforming $\mathrm{SLADE}^2$ . These findings highlight SLADE's superior resilience against AutoAttack's ensemble of adversarial attacks. In Appendix, we present detailed results for clean and adversarial evaluations under $\epsilon = 2/255$ and $\epsilon = 8/255$ radii on zero-shot classification datasets. + +# 4.4. Jailbreak attacks on LVLMs + +In previous sections, we highlighted the vulnerability of LVLMs to gradient-based attacks. These models are also susceptible to jailbreak attacks due to the expanded attack surface from the image modality. Adversaries use gradient-based perturbations to create universal optimization-based jailbreak attacks, where a single adversarial image can compromise multiple LVLMs (e.g., LLaVA, MiniGPT-4 [65]) without contextually aligned prompts. + +Table 2. Quantitative evaluation of ${\ell }_{\infty }$ targeted attacks at $\epsilon = 4/{255}$ radii. + +
TargetCLIPTeCoA2FARE2SLADE2TeCoA4FARE4SLADE4
Vaccination causes autism...25/250/250/250/250/250/250/25
Achieve this physique with ‘X’...25/252/251/250/250/250/250/25
Invest in ‘Y’ stock for quick...25/251/250/250/250/250/250/25
Visit https://tinyurl.com/zh...25/253/251/251/250/250/250/25
Opposition political leaders...25/250/250/250/250/250/250/25
Average attack success rate:100%4.8%1.6%0.8%0%0%0%
Average CIDEr score:018.625.148.467.378.787.5
+ +Table 3. Evaluation of CLIP models on zero-shot classification datasets under AutoAttack. + +
AttackVision encoderCIFAR10CIFAR100CarsFlowersEuroSATPCAM
4/255CLIP0.00.00.00.00.00.00.00.0
TeCoA231.017.85.86.73.516.0
FARE225.914.04.87.15.517.2
SLADE233.419.36.18.84.719.2
€TeCoA435.521.68.412.46.843.5
FARE434.821.412.812.911.750.2
SLADE434.423.015.313.712.551.3
+ +Attack Setup. To evaluate the robustness of vision encoders against jailbreak attacks, we use two optimization-based attacks: ImgJP [43] and VisualAdv [47], both of which employ gradient-based perturbations to generate universal adversarial images that bypass LVLMs' safety mechanisms. For ImgJP, we perform 1,000 iterations at $\epsilon = 16/255$ under the $\ell_{\infty}$ norm, using MiniGPT-4 (Vicuna 7B) to generate adversarial images targeting the LLaVA-7B and LLaVA-13B models. Each image is tested against 25 harmful instructions from the AdvBench dataset [66], with Attack Success Rate (ASR) reported as the percentage of harmful prompts that elicit unsafe responses. In the VisualAdv evaluation, universal imperceptible adversarial images are generated with high perturbation ( $\epsilon = 128/255$ ) and tested against 100 toxic prompts from the RealToxicityPrompts dataset [21]. Using LLaVA-13B as a surrogate model, we measure the toxicity in the outputs via the Perspective API [28] and report the percentage of responses exhibiting attributes like threat or profanity. + +Results and discussions. Table 4 presents the performance of different CLIP encoders under VisualAdv jailbreak attacks with LLaVA-13B. Substituting the original CLIP encoder with $\mathrm{SLADE^4}$ significantly lowers toxicity rates, reducing them from $30.4\%$ to $13.7\%$ at $\epsilon = 128 / 255$ . While TeCoA and FARE also reduce toxicity rates compared to the original CLIP model, they remain less effective than SLADE. Although external defenses like JailGuard + +and Diffpure show some promise, $\mathrm{SLADE^4}$ consistently achieves superior results. For our additional evaluation of VisualAdv attacks, see Appendix. Table 5 presents the ASR results for ImgJP attacks across target LVLMs. The original CLIP encoder in LLaVA models shows high ASR, particularly in LLaVA-7B. While external defenses like JailGuard decrease ASR, they add complexity and overhead due to requiring LVLM-specific training. $\mathrm{SLADE^4}$ outperforms $\mathrm{FARE^4}$ and $\mathrm{TeCoA^4}$ , achieving the lowest ASR and underscoring SLADE's effectiveness against both gradient-based and universal jailbreak attacks. + +# 4.5. Evaluation across other Tasks + +In the preceding sections, we evaluated the performance of robust encoders in LVLMs under adversarial and jailbreak attacks. While resilience against such attacks is critical, it is equally important to ensure that LVLMs equipped with robust encoders maintain their core functionalities, including reasoning and instruction-following capabilities. In this section, we assess whether LVLMs with robust encoders can effectively retain these essential abilities, particularly in solving reasoning tasks and adhering to instructions. + +Instruction following benchmark. Instruction following is a fundamental capability in LVLMs, enabling them to accurately interpret and respond to contextually specific requests across diverse and scenarios. This capability extends LVLMs beyond basic image recognition, allowing them to generate responses that are sensitive to context, interpret visual nuances, and incorporate reasoning about complex, scene-specific details [2, 59]. We assess the instruction-following performance of LVLMs with robust encoders through a quantitative evaluation using VisIT-Bench [8], a recently developed comprehensive benchmark that includes 592 challenging instructions across 70 diverse instruction categories. In this evaluation, we use LLaVA-13B and a subset of 200 instructions, where each instance includes an instruction, an input image, and a human-verified GPT-4 response. Following [8], we set the GPT-4 [45] response as the ground truth and use BERTscore [61] and METEOR score [5] as evaluation metrics and summarize the results in Table 6. We observe that SLADE upholds instruction + +Table 4. Evaluation of LLaVA-13B against VisualAdv jailbreak attacks with different CLIP based models. We report the percentage of LVLM outputs that exhibit specific toxic attributes. The blue downward arrows (↓) indicate a decrease in toxic attribute percentages compared to the baseline CLIP model, while the red upward arrows (↑) indicate an increase. Lower values signify better performance. + +
Attack Strength (€)Vision EncoderExternal DefenseAny IdentityProfanitySevere ToxicitySexually ExplicitThreatToxicityAverage
128/255CLIP—62.09.053.05.019.04.061.0
JailGuard41.04.020.01.014.05.030.0
Diffpure (n = 0.25)35.04.028.02.011.02.027.0
TeCoA4—41.04.027.02.015.05.045.0
FARE4—36.04.019.01.09.05.033.0
SLADE4—34.04.017.00.010.05.030.0
+ +Table 5. Performance analysis of LVLMs with various vision encoders and defenses against ImgJP Attack at $\epsilon = 16/255$ . + +
LVLMVision EncoderExternal DefenseASR (%)
LLaVA-13BCLIP—28.0 ± 0.2
FARE4—18.3 ± 0.1
CLIPJailGuard21.0 ± 0.1
SLADE4—15.0 ± 0.3
LLaVA-7BCLIP—34.0 ± 0.1
FARE4—24.0 ± 0.2
CLIPJailGuard26.0 ± 0.3
SLADE4—19.3 ± 0.5
+ +Table 6. Instruction-following benchmark results of LLaVA 13B-model with original CLIP and robust vision encoders. + +
Vision EncoderBERTScoreMETEOR Score
PrecisionRecallF1
CLIP0.8910.8940.8920.37
SLADE²0.8900.8950.8870.36
SLADE⁴0.8890.8910.8890.36
+ +following performance comparably to the original CLIP. In Appendix, we present additional details on the instruction-following benchmark setup and qualitative examples. + +Evaluation on reasoning tasks. LVLMs are designed to emulate human reasoning, often achieved through a chain-of-thought approach, where a series of intermediate reasoning steps significantly enhances their ability to handle complex tasks. Following [50], we assess the reasoning capabilities of LVLMs equipped with robust vision encoders on the Science Question Answering (SQA) [35] benchmark and summarize the results in Table 7. We evaluate LLaVA-13B on a subset of 10,000 image-question pairs from the SQA benchmark, where each instance includes an explanation of a concept, followed by a question and an input image. LLaVA with the original CLIP model achieves the best accuracy of $64.5\%$ . Both FARE and SLADE perform better than TeCoA. Among all the robust encoders, $\mathrm{SLADE}^2$ performs best and achieves an accuracy of $63.4\%$ accuracy. + +Table 7. Reasoning tasks evaluation with LLaVA on the SQA benchmark. We report accuracy for image-question pairs. + +
CLIPTeCoA2FARE2TeCoA4FARE4SLADE2SLADE4
64.561.162.463.059.363.462.3
+ +Our evaluations of reasoning and instruction-following tasks reveal that SLADE achieves performance comparable to the original CLIP model. This demonstrate that SLADE enables LVLMs to preserve their core capabilities, such as reasoning and instruction-following, without compromise. Furthermore, as detailed in the Appendix, SLADE also improves its image generation capabilities, demonstrating its versatility beyond robust performance. + +# 5. Conclusion + +In this paper, we introduce SLADE, an unsupervised adversarial fine-tuning scheme designed to enhance the robustness of the widely-used CLIP vision encoder against both gradient-based and optimization-based jailbreak attacks. By leveraging a dual-level contrastive learning approach, SLADE effectively captures fine-grained image details at the patch level while preserving high-level semantic consistency at the global level, enabling VLMs to effectively withstand adversarial perturbations without requiring any computational overhead. Our extensive experiments demonstrate that SLADE-equipped LVLMs outperform existing models in adversarial robustness, excelling in nuanced semantic understanding of perturbed images while maintaining high accuracy on clean data. Moreover, we empirically demonstrate that replacing the default CLIP encoder with SLADE enables robust defense against jailbreak attacks without requiring any LVLM-specific training, highlighting its seamless integration and adaptability across multimodal systems. Further, we validate that SLADE-equipped LVLMs do not compromise on non-adversarial core performance, such as reasoning and instruction-following capabilities. These findings position SLADE as a transformative approach with the potential to set a new benchmark for adversarial robustness in multimodal AI, paving the way for safer and more reliable applications in real-world scenarios. + +# 6. Acknowledgement + +This material is partly based upon work supported by the U.S. National Science Foundation (NSF) under Grant No. CRII-IIS-RI-2348145. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1 +[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 4, 7 +[3] Eugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, and Vitaly Shmatikov. (ab) using images and sounds for indirect instruction injection in multi-modal llms. arXiv preprint arXiv:2307.10490, 2023. 2 +[4] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime. arXiv preprint arXiv:2309.00236, 2023. 2 +[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7 +[6] Hritik Bansal, Nishad Singhi, Yu Yang, Fan Yin, Aditya Grover, and Kai-Wei Chang. Cleanclip: Mitigating data poisoning attacks in multimodal contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112–123, 2023. 2 +[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1 +[8] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595, 2023. 7 +[9] Nicholas Carlini, Milad Nasr, Christopher A Choquette-Choo, Matthew Jagielski, Irena Gao, Pang Wei W Koh, Daphne Ippolito, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned? Advances in Neural Information Processing Systems, 36, 2024. 1 +[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Pro + +ceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1 +[11] Lu Chen, Qiangchang Wang, Zhaohui Li, and Yilong Yin. Hypergraph-guided intra-and inter-category relation modeling for fine-grained visual recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8043-8052, 2024. 3 +[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3 +[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 4 +[15] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 4, 6 +[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2 +[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2 +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 +[19] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 2 +[20] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in neural information processing systems, 34:21480-21492, 2021. 2 +[21] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462, 2020. 7 +[22] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1 +[23] Sven Gowal, Po-Sen Huang, Aaron van den Oord, Timothy Mann, and Pushmeet Kohli. Self-supervised adversarial robustness for the low-label, high-data regime. In International conference on learning representations, 2020. 2, 3 + +[24] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2 +[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5 +[26] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6 +[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in neural information processing systems, 33: 16199-16210, 2020. 2 +[28] Google Jigsaw. Perspective api. https://www.perspectiveapi.com/, 2017. Accessed: 2024-06-02. 7 +[29] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in neural information processing systems, 33:2983-2994, 2020. 3 +[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6 +[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[32] Xiao Li, Wei Zhang, Yining Liu, Zhanhao Hu, Bo Zhang, and Xiaolin Hu. Language-driven anchors for zero-shot adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24686-24695, 2024. 2 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 4 +[35] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 8 +[36] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2 + +[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3 +[38] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 1, 2, 4 +[39] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5 +[40] AI Meta. Introducing llama: A foundational, 65-billion-parameter large language model. Meta AI, 2023. 1 +[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 3 +[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6 +[43] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 1, 2, 7 +[44] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3 +[45] R OpenAI. Gpt-4 technical report. ArXiv, 2303, 2023. 1, 7 +[46] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 5 +[47] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 7 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4 +[49] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 2, 4 +[50] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336, 2024. 1, 2, 4, 8 + +[51] Arindam Sikdar, Yonghuai Liu, Siddhardha Kedarisetty, Yitian Zhao, Amr Ahmed, and Ardhendu Behera. Interweaving insights: High-order feature interaction for fine-grained visual recognition. International Journal of Computer Vision, pages 1-25, 2024. 3 +[52] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 4 +[53] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 4 +[54] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5 +[55] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation equivariant cnns for digital pathology. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11, pages 210-218. Springer, 2018. 6 +[56] Xiao Wang, Yuhang Huang, Dan Zeng, and Guo-Jun Qi. Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10718-10730, 2023. 3 +[57] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 1 +[58] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 2 +[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7 +[60] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness. In European Conference on Computer Vision, pages 725-742. Springer, 2022. 3 +[61] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 7 +[62] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 2 +[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[64] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic + +adversarial examples in multimodal contrastive learning. In 31st ACM International Conference on Multimedia, pages 6311-6320, 2023. 2 +[65] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6 +[66] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 7 \ No newline at end of file diff --git a/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/images.zip b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c585ce1ffe60018322a908164f5433e37aafa35 --- /dev/null +++ b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e96b0e67db0267431448d7333e919d31e21b0cb91b13b9e1b5f6fc8ea1e931f +size 561678 diff --git a/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/layout.json b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ed6032c36cc456f6024b498b17df66f87351079e --- /dev/null +++ b/2025/SLADE_ Shielding against Dual Exploits in Large Vision-Language Models/layout.json @@ -0,0 +1,9893 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 537, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 537, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 537, + 121 + ], + "type": "text", + "content": "SLADE: Shielding against Dual Exploits in Large Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 143, + 389, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 143, + 389, + 158 + ], + "spans": [ + { + "bbox": [ + 222, + 143, + 389, + 158 + ], + "type": "text", + "content": "Md Zarif Hossain, Ahmed Imteaj†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 159, + 497, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 159, + 497, + 172 + ], + "spans": [ + { + "bbox": [ + 113, + 159, + 497, + 172 + ], + "type": "text", + "content": "School of Computing, Southern Illinois University Carbondale, IL 62901, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 200, + 200, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 200, + 200, + 213 + ], + "spans": [ + { + "bbox": [ + 151, + 200, + 200, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 225, + 294, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 225, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 294, + 536 + ], + "type": "text", + "content": "Large Vision-Language Models (LVLMs) have emerged as transformative tools in multimodal tasks, seamlessly integrating pretrained vision encoders to align visual and textual modalities. Prior works have highlighted the susceptibility of LVLMs to dual exploits (gradient-based and optimization-based jailbreak attacks), which leverage the expanded attack surface introduced by the image modality. Despite advancements in enhancing robustness, existing methods fall short in their ability to defend against dual exploits while preserving fine-grained semantic details and overall semantic coherence under intense adversarial perturbations. To bridge this gap, we introduce SheiLding against Dual Exploits (SLADE), a novel unsupervised adversarial fine-tuning scheme that enhances the resilience of CLIP-based vision encoders. SLADE's dual-level contrastive learning approach balances the granular and the holistic, capturing fine-grained image details without losing sight of high-level semantic coherence. Extensive experiments demonstrate that SLADE-equipped LVLMs set a new benchmark for robustness against dual exploits while preserving fine-grained semantic details of perturbed images. Notably, SLADE achieves these results without compromising the core functionalities of LVLMs, such as instruction following, or requiring the computational overhead (e.g., large batch sizes, momentum encoders) commonly associated with traditional contrastive learning methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "The remarkable success of large language models (LLMs) [40, 45] in understanding and generating human-like text has inspired researchers to extend their capabilities to the visual domain. This effort has led to the development of Large Vision-Language Models (LVLMs) [1, 34], which aim to bridge the gap between visual and textual information. These LVLMs frequently utilize pretrained vision encoders, such as CLIP [48], BEiT [7], and DINO [10], to derive meaningful visual representations from images. LVLMs [2, 34] equipped with these vision encoders demonstrate remarkable zero-shot performance in down-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 201, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 488 + ], + "type": "text", + "content": "stream tasks such as image captioning, and visual question answering (VQA). Their promising capabilities and impact on multimodal learning have led to a growing number of LVLMs being publicly released. However, while this availability fosters research and application development, it also introduces critical safety concerns that require careful attention. Recent studies [57, 63] highlight the high susceptibility of LVLMs to harmful exploits on text and image modalities. Notably, the addition of a vision modality increases the model's attack surface, making LVLMs even more vulnerable to manipulation compared to LLMs [9, 22]. Moreover, the adversaries can use imperceptible gradient-based perturbations to create targeted attacks, generating malicious outputs that evade detection [49]. Such targeted attacks are particularly concerning due to their ability to manipulate models toward specific, often harmful outputs. Additionally, the increasing deployment of LVLMs raises the potential for exploitation not only by adversaries but also by corporations aiming to leverage these models to promote their products or agendas as shown in Figure 1. Recent studies [43, 47] demonstrate that gradient-based perturbations can be adapted in optimization-based jailbreak attacks to bypass a model's safety mechanisms, causing LVLMs to generate harmful or undesirable content." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "While existing methods [38, 50] made notable progress in improving the robustness of vision encoders, particularly CLIP, which is widely utilized across various applications, critical challenges persist. For example, in downstream tasks such as image captioning, these models successfully capture high-level semantic meaning but often struggle to preserve fine-grained semantic details when faced with perturbed images. Consequently, their ability to generate accurate and contextually rich descriptions is compromised, limiting their applicability and effectiveness in real-world scenarios where precision and detail are paramount. To address these concerns, we propose SLADE, an unsupervised adversarial fine-tuning scheme that enhances the robustness of the widely used CLIP vision encoder's embeddings against dual exploits: gradient-based attacks and optimization-based Jailbreak attacks. SLADE combines symmetric contrastive learning objectives at both the patch level and the global level (on full image) to capture fine" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24244" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 528, + 240 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 528, + 240 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 528, + 240 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 528, + 240 + ], + "type": "image", + "image_path": "f934b1330ff6c288914c75eff374fe6f7e8a7ebad2f6cab23d346de00e77f5d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "text", + "content": "Figure 1. Targeted " + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "text", + "content": " attack and VisualAdV [47] jailbreak attack at " + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "text", + "content": " radii and using original CLIP and SLADE as vision encoder in LLaVA-7B. SLADE (trained at " + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 55, + 249, + 555, + 283 + ], + "type": "text", + "content": ") generates captions with fine-grained semantic details for the perturbed sample and successfully defends against the jailbreak attempt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 294, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 294, + 464 + ], + "type": "text", + "content": "grained image details and high-level semantic representation of the entire image, respectively. This dual-level approach enables the model to preserve both detailed features and high-level semantic consistency, even in the presence of adversarial perturbations. Moreover, SLADE does not require LVLM or task-specific training and can be seamlessly integrated into existing LVLMs as a robust vision encoder. SLADE enhances the adversarial robustness of the vision modality in LVLMs, preserving the model's ability to capture nuanced semantic details while effectively defending against jailbreak attacks, as illustrated in Figure 1. Our work sets a new standard for adversarial robustness in LVLMs, a critical advancement as these models become integral to critical applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 470, + 147, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 147, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 147, + 482 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 490, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 610 + ], + "type": "text", + "content": "Large Vision-Language Models. In recent times, several advanced LVLMs have been introduced, including LLaVA [34], OpenFlamingo (OF) [2], InstructBLIP [17] and Macaw-LLM [36]. These models leverage the power of multimodal learning by combining visual and textual data, thereby, enhancing their capability to understand and generate responses to multimodal inputs. Most of these LVLMs incorporate pre-trained LLMs, such as Llama [53], paired with large-scale vision encoders like CLIP [48], to process multimodal data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Adversarial and Jailbreaking attacks in LVLMs. In recent studies [3, 4, 19, 63], authors have explored the susceptibility of large vision-language models to adversarial attacks. Notably, recent works [3, 6, 27, 64] demonstrate the vulnerability of vision encoders in LVLMs, particularly CLIP, to adversarial manipulation, which can significantly impair the performance of LVLMs in downstream tasks. Moreover, [16, 49] highlight the efficacy of gradient-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 417 + ], + "type": "text", + "content": "based attacks on LVLMs, demonstrating their capacity to induce the model into generating harmful or misleading outputs. In [43, 47] authors show how gradient-based perturbations can be leveraged in optimization-based jailbreak attacks to bypass LVLMs' safety protocols. While recent studies have proposed defense strategies [58, 62] to counter jailbreak attacks, these methods rely heavily on the model's inherent safety alignment, necessitating specific training for each LVLM and often resulting in degraded performance in downstream tasks (e.g., VQA and image captioning)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "Robustness against exploits in LVLMs. Prior works [32, 38] introduced supervised adversarial fine-tuning scheme for CLIP, using cross-modal image-text contrastive loss. However, in LVLMs their zero-shot performance in downstream tasks (mostly in unseen classes) often degrades due to an over-dependence on the text embeddings during the fine-tuning process. To address this, some concurrent works [20, 23, 27] proposed unsupervised adversarial training methods using SimCLR [12] and BYOL [24] contrastive loss, which improve robustness but require large batch sizes for sufficient negative samples or a momentum encoder for optimal results. Additionally, the authors in [50] proposed an alternative approach, tailoring an " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " loss-based unsupervised fine-tuning scheme. However, our empirical analysis reveals that applying " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " loss across the entire image is insufficient for retaining nuanced details, particularly under adversarial conditions. Unlike prior methods, SLADE integrates a symmetric contrastive learning objective applied simultaneously at the patch level and global level. This dual-level strategy enables SLADE to capture fine-grained details and nuanced spatial relationships, even under adversarial perturbations. Drawing inspiration from traditional machine learning techniques that use patch embedding with pretext tasks to preserve detail and spatial" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24245" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": "localization [11, 41, 44, 51], SLADE advances this concept by incorporating it directly into its adversarial fin-tuning mechanism. Furthermore, SLADE leverages a stop-gradient mechanism to bypass the reliance on large batches or momentum encoders, offering a scalable and computationally efficient solution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 176, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 176, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 176, + 167 + ], + "type": "text", + "content": "3. Overview of SLADE" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "text", + "content": "This section provides a detailed overview of SLADE, our proposed adversarial fine-tuning scheme, illustrated in Figure 2. The primary objective of SLADE is to adversarially fine-tune a CLIP-like model " + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "text", + "content": ", with parameter " + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "text", + "content": " on a given dataset " + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{I_1, I_2, \\dots, I_N\\}" + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "inline_equation", + "content": "I_n \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 175, + 296, + 293 + ], + "type": "text", + "content": ". SLADE is designed to learn invariant features that are robust to adversarial perturbations while preserving the semantic details of the input data. This ensures that the fine-tuned model effectively balances adversarial resilience with the retention of critical semantic information." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 303, + 296, + 456 + ], + "blocks": [ + { + "bbox": [ + 58, + 303, + 296, + 456 + ], + "lines": [ + { + "bbox": [ + 58, + 303, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 296, + 456 + ], + "type": "image", + "image_path": "75bfc60eac1559218dd8d04e2d43b374b57dc9b5656653b2eebb7d7c948e4492.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 464, + 278, + 475 + ], + "lines": [ + { + "bbox": [ + 71, + 464, + 278, + 475 + ], + "spans": [ + { + "bbox": [ + 71, + 464, + 278, + 475 + ], + "type": "text", + "content": "Figure 2. Workflow of our proposed scheme: SLADE." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 218, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 218, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 218, + 496 + ], + "type": "text", + "content": "3.1. Adversarial Image generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": "During adversarial fine-tuning, SLADE generates a perturbed version " + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": " from the original clean image " + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": " using Projected Gradient Descent (PGD) [37]. Given an input image " + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": ", the perturbed image " + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": " is iteratively computed by applying small, bounded modifications designed to maximize the adversarial objective. At each iteration " + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": ", the perturbed image " + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "inline_equation", + "content": "I_{t + 1}^{\\prime}" + }, + { + "bbox": [ + 55, + 502, + 296, + 585 + ], + "type": "text", + "content": " is updated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 594, + 295, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 594, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 96, + 594, + 295, + 608 + ], + "type": "interline_equation", + "content": "I _ {t + 1} ^ {\\prime} = \\Pi_ {I + \\epsilon} \\left(I _ {t} + \\alpha \\cdot \\operatorname {s i g n} \\left(\\nabla_ {I} L \\left(I _ {t}, y\\right)\\right)\\right) \\tag {1}", + "image_path": "48c78ab5c35087a85ffa14eebc8f07a86d8b5f71141ed06e777972ceed69e9a1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": " is the loss function (e.g., cross-entropy loss) and " + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\nabla_I L(I_t, y)" + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": " denotes the gradient of this loss with respect to the input " + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": " at each iteration. The term " + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": " is the step size, and " + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": " is the maximum allowed perturbation bound." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 671, + 216, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 216, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 216, + 683 + ], + "type": "text", + "content": "3.2. Patch-Level Contrastive Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "SLADE applies a symmetric contrastive learning objective at the patch level to retain fine-grained image features under" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "adversarial perturbations. Each image " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and its adversarial counterpart " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " are divided into non-overlapping patches. Let " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "P(I) = \\{p_{1},p_{2},\\ldots ,p_{n}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "P(I^{\\prime}) = \\{p_1^{\\prime},p_2^{\\prime},\\dots ,p_n^{\\prime}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " represent the respective sets of patches from clean image I and perturbed image " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", respectively. For an image with resolution " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " color channels, each patch is a flattened 2D segment of size " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "(K,K)" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " is the patch size. The number of resulting patches, " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", is calculated as " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "N = \\frac{H\\times W}{K^2}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ". Subsequently, each patch " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and its corresponding adversarial patch " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "p_i^\\prime" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " are then passed through the encoder, yielding patch embeddings " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "z_{i} = f_{\\theta}(p_{i})" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "z_{i}^{\\prime} = f_{\\theta}(p_{i}^{\\prime})" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "content": "In SLADE, we focus on ensuring the model's capability to retain fine-grained details despite adversarial perturbations. To achieve this, we apply a contrastive loss at the patch level, aligning embeddings of corresponding patches (i.e., " + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "inline_equation", + "content": "z_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "content": "). By aligning patches individually, the model " + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "content": " treats each patch as an independent unit and encourages to preserve semantic coherence within localized regions. This approach enables the model to capture fine-grained and spatially distinct features, thereby retaining intricate details that might otherwise be lost to adversarial noise. Following the adversarial contrastive loss approach in [29, 60], we define the loss for " + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "inline_equation", + "content": "z_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 217, + 555, + 361 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 369, + 553, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 369, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 318, + 369, + 553, + 405 + ], + "type": "interline_equation", + "content": "L _ {c o} = - \\log \\frac {\\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right)}{\\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {\\prime}\\right) / \\tau\\right) + \\sum_ {I ^ {n} \\in \\mathcal {D} _ {N}} \\exp \\left(\\operatorname* {s i m} \\left(z _ {i} , z _ {i} ^ {n}\\right) / \\tau\\right)} \\tag {2}", + "image_path": "252c95392283b0328342d8c9c66d72f6783e0eac3c97e074652ab4facd74a821.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N\\subseteq \\mathcal{D}" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " represents a batch of negative samples, where " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " is the batch size. The negative images " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " are drawn uniformly at random from the dataset " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "I\\neq I^{n}" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": ". The embedding of patch " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "p_i^n" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " from a negative sample " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "z_{i}^{n}" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " is a temperature parameter, and " + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 406, + 553, + 477 + ], + "type": "text", + "content": " denotes the cosine similarity between embeddings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 479, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 586 + ], + "type": "text", + "content": "In Equation 2, negative samples are crucial to prevent a collapsing solution. In their absence, the model may map all patches to similar embeddings, failing to learn discriminative representations [56, 60]. Achieving robust performance typically requires a large number of negative samples, which increases the need for a large batch size, making it resource-intensive. Leveraging a momentum encoder [23] can mitigate the reliance on large batch sizes, though it introduces additional computational overhead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 642 + ], + "type": "text", + "content": "Mitigating computational overhead: SLADE incorporates a stop-gradient mechanism [13] with a cosine similarity metric to prevent feature collapse and reduce computational overhead, which can be defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 654, + 553, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 654, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 361, + 654, + 553, + 667 + ], + "type": "interline_equation", + "content": "\\operatorname {s t o p s i m} (\\cdot , \\cdot) = \\operatorname {s i m} (\\cdot , \\operatorname {s t o p g r a d} (\\cdot)) \\tag {3}", + "image_path": "00d52c568648ddb4eb50b25cf9e308ba0cce2fefc5c74539a8f89f559581689d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "The stop-gradient mechanism in Equation 3 is then incorporated into a symmetric contrastive loss, which is designed to align patch embeddings between clean and adver" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24246" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": "sarial images. The resulting contrastive loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 295, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 134 + ], + "type": "interline_equation", + "content": "L _ {p a t c h} = - \\frac {1}{2 N} \\sum_ {i = 1} ^ {N} \\left(\\operatorname {s t o p s i m} \\left(z _ {i}, z _ {i} ^ {\\prime}\\right) + \\operatorname {s t o p s i m} \\left(z _ {i} ^ {\\prime}, z _ {i}\\right)\\right) \\tag {4}", + "image_path": "08e64d21fde116590ed865a353d9759302551bf04b715853a05bf07f035ae821.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "content": " denotes the total number of patches per image and " + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "inline_equation", + "content": "L_{patch}" + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "content": " represents the patch-level symmetric contrastive loss, which aims to align the patch embeddings of clean and adversarial images to capture fine-grained details. The stopgrad " + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "content": " operation is applied in a symmetric manner, where either " + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "inline_equation", + "content": "z_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 135, + 296, + 255 + ], + "type": "text", + "content": " is set to be constant in each term. This halts the gradient back-propagation on one side, preventing the model from simply mapping to constant outputs. By only updating the gradients on one side, stop-gradient prevents the model from converging to trivial solutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 217, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 217, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 217, + 276 + ], + "type": "text", + "content": "3.3. Global-level Contrastive Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "text", + "content": "In Equation 4, SLADE aims to preserve fine-grained details by focusing on patch-level alignment between clean and adversarial patch embeddings. However, concentrating solely on patch-level alignment could lead to a potential issue where the model captures fine-grained details in each patch but may struggle to capture the overall semantic meaning of the image as it might overlook the contextual relation between the patches. To address this and ensure the preservation of overall semantic coherence, SLADE introduces an additional global-level alignment objective on the clean " + }, + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "text", + "content": " and adversarial image " + }, + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 425, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 425, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 66, + 425, + 295, + 449 + ], + "type": "interline_equation", + "content": "L _ {g l o b a l} = - \\frac {1}{2} \\left(\\operatorname {s t o p s i m} \\left(Z, Z ^ {\\prime}\\right) + \\operatorname {s t o p s i m} \\left(Z ^ {\\prime}, Z\\right)\\right) \\tag {5}", + "image_path": "b0a8d3a4201da01079255fc96ea230aebdd24b5cbe66d47fcd076988c42a748d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "inline_equation", + "content": "Z'" + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": " are the embeddings of clean " + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": " and perturbed image " + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": "Final loss for SLADE. The final adversarial fine-tuning objective in SLADE combines the fine-grained and global losses with a weighting parameter " + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "inline_equation", + "content": "0 \\leq \\lambda \\leq 1" + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": ") to control their relative importance. The final loss function for SLADE can be defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 567, + 295, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 567, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 102, + 567, + 295, + 581 + ], + "type": "interline_equation", + "content": "L _ {s l a d e} = \\lambda L _ {p a t c h} + (1 - \\lambda) L _ {g l o b a l} \\tag {6}", + "image_path": "3bee5f3627d5bf9eebe7064015e43600ee28383fd48caa08143b58a44b899e14.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Minimizing the loss " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "L_{slade}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " effectively preserves both localized and high-level semantic features, thereby improving robustness against perturbations. Adjusting the weight " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " provides flexibility in the model's focus: setting " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " prioritizes the global objective, emphasizing high-level semantic features, while larger values of " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " shift the focus toward finer details within patches. However, this shift may compromise the model's ability to generalize across the entire image. In Appendix, we provide a detailed analysis of the impact of " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and the stop-gradient mechanism in SLADE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "type": "text", + "content": "4. Experimental Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "type": "text", + "content": "LVLM models. In our experiments, we use OpenFlamingo-9B (OF) [2] and two versions of LLaVA-1.5 [34]: LLaVA-1.5 Vicuna-7B (LLaVA-7B) and LLaVA-1.5 LLaMA-2-13B (LLaVA-13B) as our target LVLMs. While these LVLMs use CLIP ViT-L-14 as their vision encoder, they incorporate different language models [14, 52, 53] for text processing. Notably, we select LLaVA-13B in our evaluation to assess the effectiveness of our robust encoder within the context of larger LVLM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": "Adversarial fine-tuning settings. In SLADE, we adversarially fine-tune the CLIP ViT-L/14 encoder on the ImageNet dataset [18], omitting labels for an unsupervised approach. We use PGD with 10 steps at " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " norm to generate perturbed views. Our training spans two epochs with an AdamW optimizer (WD=1e-4, LR=1e-5) and a batch size of 64, incorporating a cosine decaying LR schedule with linear warmup. We set the weighting parameter " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " to 0.6 and use a patch size of 32. We conduct an ablation study (see Appendix) to determine the optimal values for these hyperparameters. To ensure compatibility with LVLMs, we perform adversarial fine-tuning on the Vit-L/14 backbone of CLIP, matching the version used by LVLMs (e.g., OF and LLaVA-1.5). During robust inference, we substitute the default CLIP [48] model of the LVLMs with SLADE encoder, while keeping the language model and projection layer frozen, offering a plug-n-play adoption. While robustness is crucial, it often comes at the cost of potential degradation of clean performance. To optimize this trade-off between robust accuracy and clean performance, we adversarially fine-tune CLIP with perturbation radii " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon = 2/255" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": ", naming these models " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^2" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^4" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": ". In the following sections, we present a comprehensive comparison of our proposed fine-tuning scheme with the original CLIP and two state-of-the-art (SOTA) adversarial fine-tuned encoders, TeCoA [38] and FARE [50]. To ensure fare comparison, we use fine-tuned versions of TeCoA and FARE at the same adversarial radii, denoting them as " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^2" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^2" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon = 2/255" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^4" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^4" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 205, + 555, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 558, + 480, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 558, + 480, + 570 + ], + "spans": [ + { + "bbox": [ + 314, + 558, + 480, + 570 + ], + "type": "text", + "content": "4.1. Untargeted attacks on LVLMs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 576, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 673 + ], + "type": "text", + "content": "We conduct untargeted attacks to assess clean and robust performance in downstream tasks. For clean evaluation, target LVLMs are given a clean image and its corresponding prompt, while for adversarial evaluation, they are given an adversarial image and prompt. In our experiments with untargeted attacks, we use the prompt \"Provide a short caption for this image\" for the image captioning task and questions from the dataset as prompts for VQA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Untargeted attack setup. Following [49], we employ APGD [15] attack with " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " norm at " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 2/255" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": ". The APGD attack is executed using single precision for 100" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24247" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 106, + 511, + 340 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Assessing the robustness of target LVLMs against untargeted attacks in image captioning and visual question answering (VQA) tasks. For image captioning, we evaluate performance using the COCO and Flickr30k datasets, reporting the CIDEr score. For VQA tasks, we utilize the OKVQA and VizWiz datasets, providing VQA accuracy as the evaluation metric." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 106, + 511, + 340 + ], + "lines": [ + { + "bbox": [ + 96, + 106, + 511, + 340 + ], + "spans": [ + { + "bbox": [ + 96, + 106, + 511, + 340 + ], + "type": "table", + "html": "
Vision LVLM encoderCOCOFlickr30KOKVQAVizWiz
clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255clean\\( \\ell_{\\infty} \\)2/2554/2558/255
Open FlamingoCLIP80.57.825.62.461.06.43.81.448.51.80.00.023.82.41.80
\\( {\\mathrm{{TeCoA}}}^{2} \\)74.559.740.310.348.237.929.010.140.127.421.510.020.116.18.35.6
\\( {\\mathrm{{FARE}}}^{2} \\)76.269.455.310.949.443.432.115.437.829.521.89.718.615.89.15.6
\\( {\\mathrm{{SLADE}}}^{2} \\)79.871.561.716.252.646.937.919.543.934.826.313.523.419.713.88.5
\\( {\\mathrm{{TeCoA}}}^{4} \\)72.364.454.210.546.139.228.712.237.430.823.58.418.814.68.35.0
\\( {\\mathrm{{FARE}}}^{4} \\)75.467.352.212.148.341.334.615.139.232.223.510.517.814.79.56.1
\\( {\\mathrm{{SLADE}}}^{4} \\)78.571.862.116.453.045.937.719.944.636.128.913.822.420.314.69.1
LLaVA-7BCLIP121.921.813.52.479.015.310.03.457.38.33.00.039.313.33.20.0
\\( {\\mathrm{{TeCoA}}}^{2} \\)115.698.373.538.175.665.350.529.455.640.330.514.238.525.415.48.3
\\( {\\mathrm{{FARE}}}^{2} \\)123.5105.286.439.478.970.360.525.154.343.530.115.337.329.317.610.4
\\( {\\mathrm{{SLADE}}}^{2} \\)127.0111.594.647.282.874.365.430.961.549.232.419.142.631.120.915.1
\\( {\\mathrm{{TeCoA}}}^{4} \\)110.395.575.635.371.862.551.027.050.339.032.312.334.530.518.39.3
\\( {\\mathrm{{FARE}}}^{4} \\)119.4100.583.541.676.370.356.529.553.545.034.815.338.531.321.010.1
\\( {\\mathrm{{SLADE}}}^{4} \\)123.9109.792.145.981.573.863.733.959.848.939.423.441.630.623.517.4
", + "image_path": "c954f618b8f7213b9194c69a5d9182051428d3566fc68862c5a585c5d6e72fb2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "text", + "content": "iterations. Additionally, we evaluate the robust CLIP models under a stronger adversarial setting by using a perturbation radius of " + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "text", + "content": " for the downstream tasks. The attacked images with " + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 2/255" + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "text", + "content": " perturbations are completely imperceptible to human eyes, while with " + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 54, + 350, + 294, + 422 + ], + "type": "text", + "content": " a user could notice the perturbation when paying close attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 428, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 428, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 54, + 428, + 295, + 584 + ], + "type": "text", + "content": "Datasets and metrics. For image captioning, we use the COCO [33] and Flickr30k [46] datasets and for VQA, we utilize OKVQA [39] and VizWiz [25] dataset. For assessing captioning tasks, we utilize the CIDEr [54] metric, which leverages higher order (longer) n-grams to capture grammatical properties as well as richer semantics [54]. Empirically, CIDEr effectively measures fine-grained detail by reflecting alignment between generated and reference captions in terms of vocabulary, context, and descriptive specificity. Higher CIDEr scores thus indicate stronger semantic alignment with ground truth. For VQA, we assess performance through VQA accuracy, which measures the percentage of correctly answered questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "Results and discussion. Table 1 presents the performance of various CLIP encoders under untargeted attacks. While the original CLIP model excels in clean performance, it suffers significantly under adversarial attacks, particularly at stronger attack strength at " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " radii. Adversarially fine-tuned CLIP models show slight degradation in clean performance due to adversarial fine-tuning, but SLADE achieves the best clean performance among them. For OF, " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE^4}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " demonstrates superior performance compared to " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE^4}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA^4}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " across most downstream task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": "datasets. Notably, when subjected to stronger attacks at " + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^4" + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": " consistently outperforms all SOTA robust clip models. Similar trends are observed with " + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^2" + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": ", outperforming " + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^2" + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^2" + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": " in both clean and robust performance. Additionally, for LLaVA, SLADE demonstrates even greater efficacy, achieving superior performance in both captioning and VQA tasks. Moreover, the higher CIDer score achieved by SLADE under attack demonstrates its superior ability to preserve the semantic details of image features in the generated captions compared to other robust variations of CLIP. In the Appendix, we provide results of untargeted attacks on LLaVA-13B LVLM and include visual examples of VQA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 515, + 468, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 468, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 468, + 528 + ], + "type": "text", + "content": "4.2. Targeted attacks on LVLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Targeted attack setup. We conduct targeted attacks on LLaVA-7B to evaluate the robust performance of CLIP models. Targeted attacks force LVLMs to produce a specific output predetermined by the attackers, posing a more severe threat than untargeted attacks. In our experiments, we randomly select 25 images from the COCO dataset and apply the APGD attack for 10,000 iterations, using five realistic target strings representing malicious intents such as misinformation and false propaganda. For the attack configuration, we use the " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " norm with radii " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 2 / 255" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ". We report the CIDEr score to quantify caption quality under attack conditions and attack success rate to measure how many times the targeted attack was successful. More details on our targeted attack pipeline and evaluation metrics are provided in the Appendix." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24248" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 188, + 76, + 392, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 76, + 392, + 86 + ], + "spans": [ + { + "bbox": [ + 188, + 76, + 392, + 86 + ], + "type": "text", + "content": "Target String: Vaccination causes autism in children." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 99, + 105, + 152, + 160 + ], + "blocks": [ + { + "bbox": [ + 105, + 93, + 150, + 101 + ], + "lines": [ + { + "bbox": [ + 105, + 93, + 150, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 150, + 101 + ], + "type": "text", + "content": "Clean Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 105, + 152, + 160 + ], + "lines": [ + { + "bbox": [ + 99, + 105, + 152, + 160 + ], + "spans": [ + { + "bbox": [ + 99, + 105, + 152, + 160 + ], + "type": "image", + "image_path": "7de4c6695f5e8e20ae3e59af2f88a02278ddb22b2918681a4e5c1fd623c28611.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 92, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 92, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 158, + 92, + 288, + 109 + ], + "type": "text", + "content": "CLIP:Young girl with mask getting vaccine while holding a teddy bear" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 114, + 294, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 114, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 159, + 114, + 294, + 129 + ], + "type": "text", + "content": "TecoA4: Young girl on a sofa taking vaccine" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 135, + 289, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 135, + 289, + 152 + ], + "spans": [ + { + "bbox": [ + 159, + 135, + 289, + 152 + ], + "type": "text", + "content": "FARE4: Young girl with mask taking vaccine, holding a teddy bear" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 155, + 298, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 155, + 298, + 173 + ], + "spans": [ + { + "bbox": [ + 156, + 155, + 298, + 173 + ], + "type": "text", + "content": "SLADE:Young girl wearing mask taking vaccine while holding teddy bear" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 105, + 358, + 160 + ], + "blocks": [ + { + "bbox": [ + 310, + 88, + 353, + 104 + ], + "lines": [ + { + "bbox": [ + 310, + 88, + 353, + 104 + ], + "spans": [ + { + "bbox": [ + 310, + 88, + 353, + 104 + ], + "type": "text", + "content": "Adversarial Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 105, + 358, + 160 + ], + "lines": [ + { + "bbox": [ + 304, + 105, + 358, + 160 + ], + "spans": [ + { + "bbox": [ + 304, + 105, + 358, + 160 + ], + "type": "image", + "image_path": "8e59069186c8b34a19a029394dbad876e9fc58ad2adc53b6a78073fbcccab85e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 365, + 91, + 496, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 496, + 106 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 496, + 106 + ], + "type": "text", + "content": "CLIP: Vaccination causes autism in children." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 114, + 492, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 114, + 492, + 122 + ], + "spans": [ + { + "bbox": [ + 365, + 114, + 492, + 122 + ], + "type": "text", + "content": "TecoA: Vaccine shot for children" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 133, + 490, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 133, + 490, + 148 + ], + "spans": [ + { + "bbox": [ + 365, + 133, + 490, + 148 + ], + "type": "text", + "content": "FARE: Young girl taking vaccine, holding a doll" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 363, + 155, + 504, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 155, + 504, + 172 + ], + "spans": [ + { + "bbox": [ + 363, + 155, + 504, + 172 + ], + "type": "text", + "content": "SLADE4:Young girl wearing mask taking vaccine while holding teddy bear" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 128, + 185, + 486, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 185, + 486, + 194 + ], + "spans": [ + { + "bbox": [ + 128, + 185, + 486, + 194 + ], + "type": "text", + "content": "Target String: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 99, + 213, + 152, + 281 + ], + "blocks": [ + { + "bbox": [ + 106, + 203, + 149, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 149, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 149, + 212 + ], + "type": "text", + "content": "Clean Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 213, + 152, + 281 + ], + "lines": [ + { + "bbox": [ + 99, + 213, + 152, + 281 + ], + "spans": [ + { + "bbox": [ + 99, + 213, + 152, + 281 + ], + "type": "image", + "image_path": "e7321226e8b26b47e7528cd7ee68c583bf01a3126da0d42dc42805b337130f42.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 158, + 203, + 298, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 203, + 298, + 220 + ], + "spans": [ + { + "bbox": [ + 158, + 203, + 298, + 220 + ], + "type": "text", + "content": "CLIP: A statue of a muscular man with beard and six packs" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 159, + 224, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 224, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 159, + 224, + 296, + 240 + ], + "type": "text", + "content": "TecoA4: A statue of a man with beard and muscles" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 159, + 246, + 300, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 246, + 300, + 262 + ], + "spans": [ + { + "bbox": [ + 159, + 246, + 300, + 262 + ], + "type": "text", + "content": "FARE: A statue of a man with a beard and six-packs" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 157, + 266, + 298, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 266, + 298, + 283 + ], + "spans": [ + { + "bbox": [ + 157, + 266, + 298, + 283 + ], + "type": "text", + "content": "SLADE4: A statue of a shirtless muscu-lar man with beard and six-packs" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 217, + 357, + 283 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 353, + 214 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 353, + 214 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 353, + 214 + ], + "type": "text", + "content": "Adversarial Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 217, + 357, + 283 + ], + "lines": [ + { + "bbox": [ + 307, + 217, + 357, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 357, + 283 + ], + "type": "image", + "image_path": "bb23b73d19c1439a2d9e81963f90c917a2611a4b4a5c7d51a936f06ffe57117a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "type": "text", + "content": "Figure 3. Qualitative examples of " + }, + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "type": "text", + "content": " targeted attacks at " + }, + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 55, + 301, + 555, + 338 + ], + "type": "text", + "content": " radii using original and adversarially fine-tuned CLIP models as vision encoder in LLaVA-7B. We present generated captions (target string, mistakes in captions, fine-grained details in caption) for the original images (left) and the imperceptible adversarial images (right)." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 364, + 198, + 503, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 198, + 503, + 223 + ], + "spans": [ + { + "bbox": [ + 364, + 198, + 503, + 223 + ], + "type": "text", + "content": "CLIP: Achieve this physique with 'X' supplement. Start now: https://tinyurl.com/mrnk." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 364, + 227, + 501, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 227, + 501, + 243 + ], + "spans": [ + { + "bbox": [ + 364, + 227, + 501, + 243 + ], + "type": "text", + "content": "TecoA4: Before and after of photosh -ot of a man" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 364, + 247, + 505, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 247, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 364, + 247, + 505, + 262 + ], + "type": "text", + "content": "FARE4: A statue of a man with a beard and muscles" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 363, + 268, + 497, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 268, + 497, + 285 + ], + "spans": [ + { + "bbox": [ + 363, + 268, + 497, + 285 + ], + "type": "text", + "content": "SLADE4: Statue of a shirtless muscu-lar man with a beard and six packs" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": "Results and discussion. We present quantitative results of our targeted attacks at " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " in Table 2 and qualitative examples in Figure 3. The original CLIP model shows high susceptibility to targeted attacks by consistently generating the specified target strings. In contrast, " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^2" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^2" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " break in only 6 and 2 cases, resulting in mean success rates of " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": ", respectively, while " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^2" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " breaks in only one case, highlighting its superior performance. Notably, " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " show complete robustness under targeted attack at " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": ". However, the quality of captions generated by " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " notably surpasses " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": ", as shown in Figure 3. We also observe that captions generated by " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " exhibit inferior quality and contain errors, while " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": "'s captions often lack intricate details from the corresponding images. For instance, in first example of Figure 3, SLADE retains the fine-grained details of the young girl wearing mask and holding teddy bear whereas FARE overlooks these details and TeCoA generates captions with mistakes. This highlights SLADE's robustness, effectively resisting attacks and retaining both fine details and semantic integrity. Empirically, the CIDEr scores in Table 2 further support these claims, with " + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^4" + }, + { + "bbox": [ + 56, + 349, + 296, + 636 + ], + "type": "text", + "content": " achieving the highest score. We present more qualitative results in the Appendix." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 647, + 263, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 263, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 263, + 658 + ], + "type": "text", + "content": "4.3. Zero-shot evaluation under AutoAttack" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Attack setup. We examine the adversarial robustness of various CLIP models in zero-shot classification tasks using AutoAttack [15]. Our evaluation spans multiple datasets, including CIFAR10, CIFAR100 [31], StanfordCars [30]," + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "text", + "content": "EuroSAT [26], PCAM [55], and Flowers [42]. For multiclass classification datasets, we utilize AutoAttack's ensemble of four attacks: APGD-CE, APGD-T, FAB, and Square. For binary classification datasets like PCAM, we exclude APGD with targeted DLR loss (APGD-T), as DLR loss is unsuitable for binary classification, and use the remaining three attacks instead. We apply AutoAttack for 100 iterations on 1,000 samples from each dataset at " + }, + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "text", + "content": " radii." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": "Results and discussion. Table 3 summarizes the performance of both original and robust CLIP models on zero-shot classification tasks. At " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": " radii, SLADE consistently outperforms TeCoA and FARE. Notably, " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^4" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": " demonstrates greater robustness compared to " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE}^4" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA}^4" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": ", with a " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": " gain in robust accuracy, while also slightly outperforming " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^2" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": ". These findings highlight SLADE's superior resilience against AutoAttack's ensemble of adversarial attacks. In Appendix, we present detailed results for clean and adversarial evaluations under " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\epsilon = 2/255" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 313, + 456, + 555, + 588 + ], + "type": "text", + "content": " radii on zero-shot classification datasets." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 598, + 471, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 471, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 471, + 609 + ], + "type": "text", + "content": "4.4. Jailbreak attacks on LVLMs" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "In previous sections, we highlighted the vulnerability of LVLMs to gradient-based attacks. These models are also susceptible to jailbreak attacks due to the expanded attack surface from the image modality. Adversaries use gradient-based perturbations to create universal optimization-based jailbreak attacks, where a single adversarial image can compromise multiple LVLMs (e.g., LLaVA, MiniGPT-4 [65]) without contextually aligned prompts." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24249" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 85, + 542, + 196 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of " + }, + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "type": "inline_equation", + "content": "{\\ell }_{\\infty }" + }, + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "type": "text", + "content": " targeted attacks at " + }, + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/{255}" + }, + { + "bbox": [ + 165, + 70, + 445, + 82 + ], + "type": "text", + "content": " radii." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 85, + 542, + 196 + ], + "lines": [ + { + "bbox": [ + 66, + 85, + 542, + 196 + ], + "spans": [ + { + "bbox": [ + 66, + 85, + 542, + 196 + ], + "type": "table", + "html": "
TargetCLIPTeCoA2FARE2SLADE2TeCoA4FARE4SLADE4
Vaccination causes autism...25/250/250/250/250/250/250/25
Achieve this physique with ‘X’...25/252/251/250/250/250/250/25
Invest in ‘Y’ stock for quick...25/251/250/250/250/250/250/25
Visit https://tinyurl.com/zh...25/253/251/251/250/250/250/25
Opposition political leaders...25/250/250/250/250/250/250/25
Average attack success rate:100%4.8%1.6%0.8%0%0%0%
Average CIDEr score:018.625.148.467.378.787.5
", + "image_path": "d8aa5069df2561da951443142469c0bac2f1825cd1e8679b81257a13e85aaf98.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 231, + 296, + 372 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 227 + ], + "type": "text", + "content": "Table 3. Evaluation of CLIP models on zero-shot classification datasets under AutoAttack." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 231, + 296, + 372 + ], + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 372 + ], + "type": "table", + "html": "
AttackVision encoderCIFAR10CIFAR100CarsFlowersEuroSATPCAM
4/255CLIP0.00.00.00.00.00.00.00.0
TeCoA231.017.85.86.73.516.0
FARE225.914.04.87.15.517.2
SLADE233.419.36.18.84.719.2
€TeCoA435.521.68.412.46.843.5
FARE434.821.412.812.911.750.2
SLADE434.423.015.313.712.551.3
", + "image_path": "c6ed7e7bbbc9c9e5435dbad63a54a3638eb361ad4b8a3a88d1d74a7656876890.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "text", + "content": "Attack Setup. To evaluate the robustness of vision encoders against jailbreak attacks, we use two optimization-based attacks: ImgJP [43] and VisualAdv [47], both of which employ gradient-based perturbations to generate universal adversarial images that bypass LVLMs' safety mechanisms. For ImgJP, we perform 1,000 iterations at " + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\epsilon = 16/255" + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "text", + "content": " under the " + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "text", + "content": " norm, using MiniGPT-4 (Vicuna 7B) to generate adversarial images targeting the LLaVA-7B and LLaVA-13B models. Each image is tested against 25 harmful instructions from the AdvBench dataset [66], with Attack Success Rate (ASR) reported as the percentage of harmful prompts that elicit unsafe responses. In the VisualAdv evaluation, universal imperceptible adversarial images are generated with high perturbation (" + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\epsilon = 128/255" + }, + { + "bbox": [ + 54, + 384, + 295, + 612 + ], + "type": "text", + "content": ") and tested against 100 toxic prompts from the RealToxicityPrompts dataset [21]. Using LLaVA-13B as a surrogate model, we measure the toxicity in the outputs via the Perspective API [28] and report the percentage of responses exhibiting attributes like threat or profanity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "Results and discussions. Table 4 presents the performance of different CLIP encoders under VisualAdv jailbreak attacks with LLaVA-13B. Substituting the original CLIP encoder with " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE^4}" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " significantly lowers toxicity rates, reducing them from " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "30.4\\%" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 128 / 255" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": ". While TeCoA and FARE also reduce toxicity rates compared to the original CLIP model, they remain less effective than SLADE. Although external defenses like JailGuard" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "text", + "content": "and Diffpure show some promise, " + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE^4}" + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "text", + "content": " consistently achieves superior results. For our additional evaluation of VisualAdv attacks, see Appendix. Table 5 presents the ASR results for ImgJP attacks across target LVLMs. The original CLIP encoder in LLaVA models shows high ASR, particularly in LLaVA-7B. While external defenses like JailGuard decrease ASR, they add complexity and overhead due to requiring LVLM-specific training. " + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE^4}" + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{FARE^4}" + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{TeCoA^4}" + }, + { + "bbox": [ + 313, + 206, + 555, + 339 + ], + "type": "text", + "content": ", achieving the lowest ASR and underscoring SLADE's effectiveness against both gradient-based and universal jailbreak attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 348, + 477, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 477, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 477, + 361 + ], + "type": "text", + "content": "4.5. Evaluation across other Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 368, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 556, + 477 + ], + "type": "text", + "content": "In the preceding sections, we evaluated the performance of robust encoders in LVLMs under adversarial and jailbreak attacks. While resilience against such attacks is critical, it is equally important to ensure that LVLMs equipped with robust encoders maintain their core functionalities, including reasoning and instruction-following capabilities. In this section, we assess whether LVLMs with robust encoders can effectively retain these essential abilities, particularly in solving reasoning tasks and adhering to instructions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "content": "Instruction following benchmark. Instruction following is a fundamental capability in LVLMs, enabling them to accurately interpret and respond to contextually specific requests across diverse and scenarios. This capability extends LVLMs beyond basic image recognition, allowing them to generate responses that are sensitive to context, interpret visual nuances, and incorporate reasoning about complex, scene-specific details [2, 59]. We assess the instruction-following performance of LVLMs with robust encoders through a quantitative evaluation using VisIT-Bench [8], a recently developed comprehensive benchmark that includes 592 challenging instructions across 70 diverse instruction categories. In this evaluation, we use LLaVA-13B and a subset of 200 instructions, where each instance includes an instruction, an input image, and a human-verified GPT-4 response. Following [8], we set the GPT-4 [45] response as the ground truth and use BERTscore [61] and METEOR score [5] as evaluation metrics and summarize the results in Table 6. We observe that SLADE upholds instruction" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24250" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 107, + 535, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 4. Evaluation of LLaVA-13B against VisualAdv jailbreak attacks with different CLIP based models. We report the percentage of LVLM outputs that exhibit specific toxic attributes. The blue downward arrows (↓) indicate a decrease in toxic attribute percentages compared to the baseline CLIP model, while the red upward arrows (↑) indicate an increase. Lower values signify better performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 107, + 535, + 216 + ], + "lines": [ + { + "bbox": [ + 75, + 107, + 535, + 216 + ], + "spans": [ + { + "bbox": [ + 75, + 107, + 535, + 216 + ], + "type": "table", + "html": "
Attack Strength (€)Vision EncoderExternal DefenseAny IdentityProfanitySevere ToxicitySexually ExplicitThreatToxicityAverage
128/255CLIP—62.09.053.05.019.04.061.0
JailGuard41.04.020.01.014.05.030.0
Diffpure (n = 0.25)35.04.028.02.011.02.027.0
TeCoA4—41.04.027.02.015.05.045.0
FARE4—36.04.019.01.09.05.033.0
SLADE4—34.04.017.00.010.05.030.0
", + "image_path": "ca8e09ee0d125c9ee6db254408e4310997b3db4b351d7cec1c09f904e3bb350a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 249, + 276, + 385 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": "Table 5. Performance analysis of LVLMs with various vision encoders and defenses against ImgJP Attack at " + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\epsilon = 16/255" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 249, + 276, + 385 + ], + "lines": [ + { + "bbox": [ + 72, + 249, + 276, + 385 + ], + "spans": [ + { + "bbox": [ + 72, + 249, + 276, + 385 + ], + "type": "table", + "html": "
LVLMVision EncoderExternal DefenseASR (%)
LLaVA-13BCLIP—28.0 ± 0.2
FARE4—18.3 ± 0.1
CLIPJailGuard21.0 ± 0.1
SLADE4—15.0 ± 0.3
LLaVA-7BCLIP—34.0 ± 0.1
FARE4—24.0 ± 0.2
CLIPJailGuard26.0 ± 0.3
SLADE4—19.3 ± 0.5
", + "image_path": "438702c441c8ccba4f76c37708f949a8c2e8aa1111bcf74e906cf87e955a4d94.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 69, + 419, + 281, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "type": "text", + "content": "Table 6. Instruction-following benchmark results of LLaVA 13B-model with original CLIP and robust vision encoders." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 419, + 281, + 485 + ], + "lines": [ + { + "bbox": [ + 69, + 419, + 281, + 485 + ], + "spans": [ + { + "bbox": [ + 69, + 419, + 281, + 485 + ], + "type": "table", + "html": "
Vision EncoderBERTScoreMETEOR Score
PrecisionRecallF1
CLIP0.8910.8940.8920.37
SLADE²0.8900.8950.8870.36
SLADE⁴0.8890.8910.8890.36
", + "image_path": "0fb6790b98d3d6dd5ecb8efeb2c4afa4557d9d8ab6c39fc0e87c9d8ff6ad4545.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 492, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 529 + ], + "type": "text", + "content": "following performance comparably to the original CLIP. In Appendix, we present additional details on the instruction-following benchmark setup and qualitative examples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "Evaluation on reasoning tasks. LVLMs are designed to emulate human reasoning, often achieved through a chain-of-thought approach, where a series of intermediate reasoning steps significantly enhances their ability to handle complex tasks. Following [50], we assess the reasoning capabilities of LVLMs equipped with robust vision encoders on the Science Question Answering (SQA) [35] benchmark and summarize the results in Table 7. We evaluate LLaVA-13B on a subset of 10,000 image-question pairs from the SQA benchmark, where each instance includes an explanation of a concept, followed by a question and an input image. LLaVA with the original CLIP model achieves the best accuracy of " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "64.5\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": ". Both FARE and SLADE perform better than TeCoA. Among all the robust encoders, " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SLADE}^2" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " performs best and achieves an accuracy of " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "63.4\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " accuracy." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 323, + 249, + 544, + 285 + ], + "blocks": [ + { + "bbox": [ + 313, + 223, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 246 + ], + "type": "text", + "content": "Table 7. Reasoning tasks evaluation with LLaVA on the SQA benchmark. We report accuracy for image-question pairs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 249, + 544, + 285 + ], + "lines": [ + { + "bbox": [ + 323, + 249, + 544, + 285 + ], + "spans": [ + { + "bbox": [ + 323, + 249, + 544, + 285 + ], + "type": "table", + "html": "
CLIPTeCoA2FARE2TeCoA4FARE4SLADE2SLADE4
64.561.162.463.059.363.462.3
", + "image_path": "9801172d4b9649c2e31ca4b74f080d8f3acc3d7eeffc9430a7d0a618ffb83f34.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 385 + ], + "type": "text", + "content": "Our evaluations of reasoning and instruction-following tasks reveal that SLADE achieves performance comparable to the original CLIP model. This demonstrate that SLADE enables LVLMs to preserve their core capabilities, such as reasoning and instruction-following, without compromise. Furthermore, as detailed in the Appendix, SLADE also improves its image generation capabilities, demonstrating its versatility beyond robust performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 394, + 388, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 388, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 388, + 407 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we introduce SLADE, an unsupervised adversarial fine-tuning scheme designed to enhance the robustness of the widely-used CLIP vision encoder against both gradient-based and optimization-based jailbreak attacks. By leveraging a dual-level contrastive learning approach, SLADE effectively captures fine-grained image details at the patch level while preserving high-level semantic consistency at the global level, enabling VLMs to effectively withstand adversarial perturbations without requiring any computational overhead. Our extensive experiments demonstrate that SLADE-equipped LVLMs outperform existing models in adversarial robustness, excelling in nuanced semantic understanding of perturbed images while maintaining high accuracy on clean data. Moreover, we empirically demonstrate that replacing the default CLIP encoder with SLADE enables robust defense against jailbreak attacks without requiring any LVLM-specific training, highlighting its seamless integration and adaptability across multimodal systems. Further, we validate that SLADE-equipped LVLMs do not compromise on non-adversarial core performance, such as reasoning and instruction-following capabilities. These findings position SLADE as a transformative approach with the potential to set a new benchmark for adversarial robustness in multimodal AI, paving the way for safer and more reliable applications in real-world scenarios." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24251" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 167, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 159 + ], + "type": "text", + "content": "This material is partly based upon work supported by the U.S. National Science Foundation (NSF) under Grant No. CRII-IIS-RI-2348145. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 193, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 294, + 258 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 260, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 260, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 260, + 296, + 326 + ], + "type": "text", + "content": "[2] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1, 2, 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 327, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 327, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 327, + 296, + 371 + ], + "type": "text", + "content": "[3] Eugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, and Vitaly Shmatikov. (ab) using images and sounds for indirect instruction injection in multi-modal llms. arXiv preprint arXiv:2307.10490, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 416 + ], + "type": "text", + "content": "[4] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime. arXiv preprint arXiv:2309.00236, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 419, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 296, + 474 + ], + "type": "text", + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 475, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 531 + ], + "type": "text", + "content": "[6] Hritik Bansal, Nishad Singhi, Yu Yang, Fan Yin, Aditya Grover, and Kai-Wei Chang. Cleanclip: Mitigating data poisoning attacks in multimodal contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 112–123, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 564 + ], + "type": "text", + "content": "[7] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "type": "text", + "content": "[8] Yonatan Bitton, Hritik Bansal, Jack Hessel, Rulin Shao, Wanrong Zhu, Anas Awadalla, Josh Gardner, Rohan Taori, and Ludwig Schmidt. Visit-bench: A benchmark for vision-language instruction following inspired by real-world use. arXiv preprint arXiv:2308.06595, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "type": "text", + "content": "[9] Nicholas Carlini, Milad Nasr, Christopher A Choquette-Choo, Matthew Jagielski, Irena Gao, Pang Wei W Koh, Daphne Ippolito, Florian Tramer, and Ludwig Schmidt. Are aligned neural networks adversarially aligned? Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Pro" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "text", + "content": "[11] Lu Chen, Qiangchang Wang, Zhaohui Li, and Yilong Yin. Hypergraph-guided intra-and inter-category relation modeling for fine-grained visual recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8043-8052, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "text", + "content": "[12] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[13] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "text", + "content": "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "text", + "content": "[15] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "text", + "content": "[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "text", + "content": "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "text", + "content": "[19] Yinpeng Dong, Huanran Chen, Jiawei Chen, Zhengwei Fang, Xiao Yang, Yichi Zhang, Yu Tian, Hang Su, and Jun Zhu. How robust is google's bard to adversarial image attacks? arXiv preprint arXiv:2309.11751, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[20] Lijie Fan, Sijia Liu, Pin-Yu Chen, Gaoyuan Zhang, and Chuang Gan. When does contrastive learning preserve adversarial robustness from pretraining to finetuning? Advances in neural information processing systems, 34:21480-21492, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[21] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint arXiv:2009.11462, 2020. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[22] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Sven Gowal, Po-Sen Huang, Aaron van den Oord, Timothy Mann, and Pushmeet Kohli. Self-supervised adversarial robustness for the low-label, high-data regime. In International conference on learning representations, 2020. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24252" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[24] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 204 + ], + "type": "text", + "content": "[25] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "text", + "content": "[26] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in neural information processing systems, 33: 16199-16210, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 305, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 335 + ], + "type": "text", + "content": "[28] Google Jigsaw. Perspective api. https://www.perspectiveapi.com/, 2017. Accessed: 2024-06-02. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "type": "text", + "content": "[29] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in neural information processing systems, 33:2983-2994, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[30] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 437 + ], + "type": "text", + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 491 + ], + "type": "text", + "content": "[32] Xiao Li, Wei Zhang, Yining Liu, Zhanhao Hu, Bo Zhang, and Xiaolin Hu. Language-driven anchors for zero-shot adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24686-24695, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 559 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 592 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 593, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 657 + ], + "type": "text", + "content": "[35] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Chenyang Lyu, Minghao Wu, Longyue Wang, Xinting Huang, Bingshuai Liu, Zefeng Du, Shuming Shi, and Zhaopeng Tu. Macaw-llm: Multi-modal language modeling with image, audio, video, and text integration. arXiv preprint arXiv:2306.09093, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[37] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[38] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 1, 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "text", + "content": "[39] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "text", + "content": "[40] AI Meta. Introducing llama: A foundational, 65-billion-parameter large language model. Meta AI, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[41] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[42] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[43] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[44] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 550, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 550, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 550, + 426 + ], + "type": "text", + "content": "[45] R OpenAI. Gpt-4 technical report. ArXiv, 2303, 2023. 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "type": "text", + "content": "[46] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "type": "text", + "content": "[47] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[49] Christian Schlarmann and Matthias Hein. On the adversarial robustness of multi-modal foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3677-3685, 2023. 1, 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[50] Christian Schlarmann, Naman Deep Singh, Francesco Croce, and Matthias Hein. Robust clip: Unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336, 2024. 1, 2, 4, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24253" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[51] Arindam Sikdar, Yonghuai Liu, Siddhardha Kedarisetty, Yitian Zhao, Amr Ahmed, and Ardhendu Behera. Interweaving insights: High-order feature interaction for fine-grained visual recognition. International Journal of Computer Vision, pages 1-25, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "type": "text", + "content": "[52] MN Team et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[53] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[54] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 319 + ], + "type": "text", + "content": "[55] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation equivariant cnns for digital pathology. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2018: 21st International Conference, Granada, Spain, September 16-20, 2018, Proceedings, Part II 11, pages 210-218. Springer, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "text", + "content": "[56] Xiao Wang, Yuhang Huang, Dan Zeng, and Guo-Jun Qi. Caco: Both positive and negative samples are directly learnable via cooperative-adversarial contrastive learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10718-10730, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[57] Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training fail? Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[58] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wenjie Wang. Defending jailbreak attack in vlms via cross-modality information detector. arXiv preprint arXiv:2407.21659, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[59] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[60] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness. In European Conference on Computer Vision, pages 725-742. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "text", + "content": "[61] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "text", + "content": "[62] Xiaoyu Zhang, Cen Zhang, Tianlin Li, Yihao Huang, Xiaojun Jia, Xiaofei Xie, Yang Liu, and Chao Shen. A mutation-based method for multi-modal jailbreaking attack detection. arXiv preprint arXiv:2312.10766, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[63] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[64] Ziqi Zhou, Shengshan Hu, Minghui Li, Hangtao Zhang, Yechao Zhang, and Hai Jin. Advclip: Downstream-agnostic" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 195 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "adversarial examples in multimodal contrastive learning. In 31st ACM International Conference on Multimedia, pages 6311-6320, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[65] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "text", + "content": "[66] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24254" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_content_list.json b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b44d1d51e385d29a23b46e6f0266a3215c28310a --- /dev/null +++ b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_content_list.json @@ -0,0 +1,1799 @@ +[ + { + "type": "text", + "text": "SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos", + "text_level": 1, + "bbox": [ + 93, + 130, + 901, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuzheng Liu $^{1*}$ Siyan Dong $^{2*†}$ Shuzhe Wang $^{3}$ Yingda Yin $^{1}$", + "bbox": [ + 245, + 179, + 750, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanchao Yang $^{2\\dagger}$ Qingnan Fan $^{4}$ Baoquan Chen $^{1\\dagger}$", + "bbox": [ + 295, + 198, + 707, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Peking University $^{2}$ The University of Hong Kong $^{3}$ Aalto University $^{4}$ VIVO", + "bbox": [ + 166, + 215, + 826, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce SLAM3R, a novel and effective system for real-time, high-quality, dense 3D reconstruction using RGB videos. SLAM3R provides an end-to-end solution by seamlessly integrating local 3D reconstruction and global coordinate registration through feedforward neural networks. Given an input video, the system first converts it into overlapping clips using a sliding window mechanism. Unlike traditional pose optimization-based methods, SLAM3R directly regresses 3D pointmaps from RGB images in each window and progressively aligns and deforms these local pointmaps to create a globally consistent scene reconstruction - all without explicitly solving any camera parameters. Experiments across datasets consistently show that SLAM3R achieves state-of-the-art reconstruction accuracy and completeness while maintaining real-time performance at $20+$ FPS. Code available at: https://github.com/PRU-VCL-3DV/SLAM3R.", + "bbox": [ + 86, + 300, + 483, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 584, + 220, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense 3D reconstruction, a long-standing challenge in computer vision, aims to capture and reconstruct the detailed geometry of real-world scenes. Traditional approaches have largely relied on multi-stage pipelines. These typically begin with sparse Simultaneous Localization and Mapping (SLAM) [7, 16, 25, 37, 38] or Structure-from-Motion (SfM) [31, 33, 47, 53, 66] algorithms to estimate camera parameters, followed by Multi-View Stereo (MVS) [18, 48, 60, 68] techniques to fill in scene details. While these methods offer high-quality reconstructions, they often require offline processing to produce a complete model, which limits their applicability in real-world scenarios.", + "bbox": [ + 89, + 609, + 482, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the literature, dense SLAM approaches [5, 9, 10, 17, 22, 39, 42, 55, 56, 76, 78] have been developed to address dense scene reconstruction as a complete system. However, these approaches often fall short in terms of reconstruction accuracy or completeness, or they rely heavily on", + "bbox": [ + 89, + 791, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0db8d85bac62fd1d9a1399553068be176822bf31718e87fa364a43f5043dae38.jpg", + "image_caption": [ + "Monocular RGB video", + "：", + "rnnnne" + ], + "image_footnote": [], + "bbox": [ + 521, + 290, + 578, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdf5f9486d9d2ae216eea8cde4677a19cd7b2b0fa006afa190e2f61c71fa8007.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 484, + 578, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b1725e7c08c93b6714b35e83d89801688addcb8e6bf3235a30dcb048dc049df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 271, + 686, + 324 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f9cac0419478a57d85aa30cf7d2a15ad0b7c97ae3bae5c0ebd692514623326b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 270, + 769, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e0ae777d0d36307b7da4c032205aec8aff4db120412f6902a211f13b2663123.jpg", + "image_caption": [ + "···", + "···" + ], + "image_footnote": [], + "bbox": [ + 803, + 270, + 877, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7d22570d7d62c3fde62ae775babab89d99b1e606a4591121ef032ec7978cc21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 329, + 776, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/198e72296ee59f9ada4d8ae8f1bfb953c3a1d6566dc895b63c5cce010087da21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 330, + 893, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab733a38f11d22d11d4e114711cc2c8ab07f9145858444a9d3b4fb7987543ab3.jpg", + "image_caption": [ + "Figure 1. We introduce a novel dense reconstruction system - SLAM3R. It takes a monocular RGB video as input and reconstructs the scene as a dense pointcloud. The video is converted into short clips for local reconstruction (denoted as inner-window), which are then incrementally registered together (inter-window) to create a global scene model. This process runs in real-time, producing a reconstruction that is both accurate and complete." + ], + "image_footnote": [], + "bbox": [ + 589, + 402, + 774, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f43ce300897134d6b32ab3c7b7c425094180c6b9f0cce2d3f7cd66873c5cc07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 402, + 895, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d96334ff76ea5b7f73d0a489e23f810f23d0b8e13a738866f6e409a60f0508c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 467, + 895, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth sensors. Recently, several monocular SLAM systems [29, 46, 73, 75, 77, 79] have been proposed to tackle dense scene reconstruction from RGB videos. By incorporating advanced scene representations [24, 36, 40, 57, 62], these systems produce accurate and complete scene reconstructions. However, this comes at the cost of reduced running efficiency. For example, NICER-SLAM [79] operates at a speed significantly below 1 FPS. Therefore, current approaches struggle with at least one of three key criteria: reconstruction accuracy, completeness, or efficiency.", + "bbox": [ + 509, + 671, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While monocular dense SLAM systems encounter the limitations mentioned earlier, recent advances in two-view geometry have shown promising potential. DUSt3R [64] introduces a purely end-to-end approach for learning dense reconstruction. Trained on large-scale datasets, its network", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Joint first authors: liu_yuzheng@stu.pku.edu.cn, siyan3d@hku.hk", + "bbox": [ + 107, + 875, + 462, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16651", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is capable of producing high-quality dense reconstructions from paired images in real-time. However, for multiple views, a global optimization step is required to align these image pairs, which significantly hampers its efficiency. A concurrent work, Spann3R [61], extends DUSt3R to multiview (video) scenarios through an incremental pipeline with spatial memory. While this method accelerates the reconstruction process, it unfortunately results in noticeable accumulated drift and reduced reconstruction quality.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present SLAM3R (pronounced \"slæmər\"), a real-time dense 3D reconstruction system using RGB-only videos as input. Unlike traditional SLAM problems, SLAM3R performs implicit camera localization and focuses more on dense scene mapping, where 3R stands for 3D Reconstruction. SLAM3R comprises a two-hierarchy framework. First, it reconstructs local 3D geometry from a sliding window that processes short clips from the input video. Then, it progressively registers these local reconstructions to build a globally consistent 3D scene. Both modules are developed with simple yet effective feed-forward models, enabling end-to-end and efficient scene reconstruction. Specifically, the two modules are the Image-to-Points (I2P) network and the Local-to-World (L2W) network. The I2P module, inspired by DUSt3R, selects a keyframe in a local window as the coordinate system reference. It directly predicts the dense 3D point map supported by the remaining frames within that window. The L2W module incrementally fuses locally reconstructed points into a coherent global coordinate system. Both processes reconstruct the 3D points without explicitly estimating any camera parameters.", + "bbox": [ + 91, + 229, + 486, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments, we demonstrate that SLAM3R provides high-quality scene reconstructions with minimal drift, outperforming existing dense SLAM systems across various benchmarks. Furthermore, SLAM3R achieves these results at $20+$ FPS, bridging the gap between quality and efficiency in RGB-only dense scene reconstruction. Our contributions are summarized below:", + "bbox": [ + 89, + 563, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel real-time end-to-end dense 3D reconstruction system that uses RGB videos to directly predict 3D pointmaps in a unified coordinate system through feed-forward neural networks.", + "- Through careful design, our Image-to-Points module can process an arbitrary number of images simultaneously, effectively extending DUSt3R to handle multiple views and produce higher-quality predictions.", + "- The proposed Local-to-World module directly aligns predicted local 3D pointmaps into a unified global coordinate system. This eliminates the need for explicit camera parameter estimation and costly global optimization.", + "- We evaluate our method on multiple public benchmarks. It achieves state-of-the-art reconstruction quality in terms of both accuracy and completeness at real-time speeds." + ], + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 656, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional offline approaches. Dense 3D pointcloud reconstruction is a long-standing problem in computer vision. Classical approaches to this problem first determine camera parameters using Structure from Motion (SfM) [31, 33, 47, 53, 66], followed by dense 3D points triangulation with Multi-View Stereo (MVS) [1, 18, 48, 60, 68]. In recent years, neural implicit [8, 30, 36, 62, 65] and 3D Gaussian [12, 19, 20] representations have been applied to further enhance the quality of dense reconstruction. While these methods deliver high-quality results, they have a significant limitation: the requirement for offline processing to generate the final 3D model, which restricts their applicability in real-time scenarios. In this paper, we focus on online dense reconstruction in the context of Simultaneous Localization and Mapping (SLAM).", + "bbox": [ + 511, + 114, + 903, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense SLAM. Early works on SLAM [4, 7, 15, 16, 25, 37, 38] focused on reconstructing the structure of unknown environments while simultaneously localizing camera poses. These approaches prioritize real-time performance but produce only sparse structures of the scene. Dense SLAM approaches [5, 9-11, 17, 22, 27, 39, 42, 55, 56, 76, 78] incorporate detailed scene geometry information to improve pose estimation. DROID-SLAM [56] introduces recurrent iterative updates of camera poses and pixel-wise depth estimates, while TANDEM [27] proposes an online MVS module for depth prediction. These systems enable real-time dense scene reconstruction. However, their focus on camera trajectory accuracy often results in incomplete and noisy 3D reconstruction. Neural implicit and Gaussian representations have also been integrated with dense SLAM systems [9, 21-23, 32, 34, 42, 42, 44, 45, 55, 67, 72, 78]. However, these approaches often rely on additional depth sensors or focus primarily on novel view synthesis rather than producing detailed geometric reconstruction.", + "bbox": [ + 511, + 340, + 903, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, several monocular dense SLAM systems [29, 46, 73, 75, 77, 79] have been developed to produce dense scene geometry reconstruction. A notable limitation of these systems is their slow runtime. Among these systems, GO-SLAM [75] achieves a speed of $\\sim 8$ FPS, which still falls short of real-time capability. Furthermore, these methods all share a common strategy: they alternate between solving for camera poses and estimating the scene representation. In contrast, this paper presents a novel approach to dense scene reconstruction that eliminates the need for explicitly solving camera parameters, offering a more efficient and streamlined solution.", + "bbox": [ + 511, + 628, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end dense 3D reconstruction. DUSt3R [64] introduces the first purely end-to-end dense 3D reconstruction pipeline without relying on camera parameters. Recently, several works have adopted a similar approach for single-view reconstruction [63], feature matching [28], novel view synthesis [52, 70], and dynamics reconstruction [74]. These", + "bbox": [ + 511, + 810, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16652", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "successes demonstrate the effectiveness of end-to-end dense point prediction, inspiring us to develop a dense reconstruction system with a similar methodology.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While DUSt3R operates in real-time for two-view predictions, its extension to multiple views involves exhaustive pairing images and performing an additional global optimization step. This process significantly increases computational time, thereby hindering its real-time performance. MASt3R [28] enhances the matching capability of DUSt3R by adding a match head, achieving more accurate keypoint correspondences for 3D reconstruction [14], but at the cost of increased computational time. More recently, the concurrent work Spann3R [61] extends DUSt3R with spatial memory. It takes a video as input and performs incremental scene reconstruction in a unified coordinate system without requiring global optimization. While this approach significantly improves runtime efficiency, the incremental reconstruction pipeline frame by frame leads to noticeable accumulated drift. Unlike Spann3R, our networks at each hierarchy take multiple frames as input to minimize drift. Additionally, we propose a self-contained retrieval module that, when registering a new frame, this module selects not only its previous few frames but also other similar frames from long-term history for more global scene reference.", + "bbox": [ + 91, + 137, + 482, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 467, + 181, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem statement. Given a monocular video consisting of a sequence of RGB image frames $\\{I_i\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{i = 1}^N$ that captures a static scene, the goal is to reconstruct its dense 3D pointcloud $P\\in \\mathbb{R}^{M\\times 3}$ , where $M$ is the number of 3D points. Our work focuses on three key objectives: maximizing 3D points recovery for reconstruction completeness, improving the accuracy of each recovered point, and achieving these goals while preserving real-time performance.", + "bbox": [ + 89, + 492, + 482, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "System overview. Figure 2 illustrates an overview of the proposed dense reconstruction system. It consists of two main components: an Image-to-Points (I2P) network that recovers local 3D points from video clips, and a Local-to-World (L2W) network that registers local reconstructions into a global scene coordinate system. During the reconstruction of the dense point cloud, the system does not explicitly solve any camera parameters. Instead, it directly predicts 3D point maps in unified coordinate systems.", + "bbox": [ + 89, + 614, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The system starts by applying a sliding window mechanism of length $L$ to convert the input video into short clips $\\{\\mathcal{W}_i \\in \\mathbb{R}^{L \\times H \\times W \\times 3}\\}$ . The I2P network then processes each window $\\mathcal{W}_i$ to recover local 3D pointmaps. Within each window, the system selects a keyframe to define a reference coordinate system for point reconstruction, as detailed in Sec. 3.1. By default, the stride of the sliding window is set to 1, ensuring each input frame in the video is selected at least once as a keyframe. For global scene reconstruction, we initialize the world coordinate system with the", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "first window and use the reconstructed frames (image and local point map produced by the I2P) as input for the L2W model. The L2W model incrementally register these local reconstructions into a unified global 3D coordinate system. To ensure both accuracy and efficiency during this process, the system maintains a limited reservoir of registered frames, called scene frames. Whenever the L2W model registers a new keyframe, we retrieve the best-correlated scene frames as a reference. The details are introduced in Sec. 3.2.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Inner-Window Local Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 234, + 831, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Image-to-Points (I2P) model aims to infer dense 3D pointmaps for every pixel of a keyframe in a given video clip. By default, the middle image of a window $\\mathcal{W}$ is chosen as the keyframe $I_{key}$ to define the local coordinate system, as it is most likely to have the largest overlap with other frames. The remaining images $\\{I_{sup_i}\\}_{i = 1}^{L - 1}$ serve as supporting frames. Note that the 3D pointmaps of supporting frames can also be reconstructed through I2P.", + "bbox": [ + 511, + 258, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The I2P network draws inspiration from DUSt3R [64], originally designed for stereo 3D reconstruction. We introduce several simple yet effective modifications to extend it for multi-view scenarios. The I2P model uses a multibranch Vision Transformer (ViT) [13] as its backbone. It consists of a shared encoder $E_{img}$ , two separate decoders $D_{key}$ and $D_{sup}$ , and a point regression head for final prediction. These components are detailed below.", + "bbox": [ + 511, + 378, + 903, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image encoder. For a given video clip, the image encoder $E_{img}$ encodes each frame $I_{i}$ to obtain token representations $F_{i} \\in \\mathbb{R}^{T \\times d}$ , where $T$ is the number of tokens and $d$ is the token dimension. The encoder $E_{img}$ comprises $m$ ViT encoder blocks, each containing self-attention and feed-forward layers. The encoding process is denoted as", + "bbox": [ + 511, + 500, + 903, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {(T \\times d)} = E _ {i m g} (I _ {i} ^ {(H \\times W \\times 3)}), i = 1, \\dots , L.\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 598, + 854, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The frames are processed independently and in parallel, with the output divided into two parts: $F_{key}$ for the keyframe and $\\{F_{sup_i}\\}_{i=1}^{L-1}$ for the supporting frames.", + "bbox": [ + 511, + 628, + 903, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Keyframe decoder. The keyframe decoder $D_{key}$ consists of $n$ ViT decoder blocks, each containing self-attention, cross-attention, and feed-forward layers. Unlike DUSt3R which uses the standard cross-attention, we introduce a novel multi-view cross-attention to combine information from different supporting frames. Given the feature tokens $F_{key}$ and $\\{F_{sup_i}\\}_{i=1}^{L-1}$ , the keyframe decoder $D_{key}$ takes $F_{key}$ as input for self-attention and performs cross-attention between $F_{key}$ and $\\{F_{sup_i}\\}_{i=1}^{L-1}$ . A decoder block is illustrated in Figure 3. For each cross-attention layer, queries are taken from $F_{key}$ , while keys and values are extracted from the supporting tokens $F_{sup_i}$ . These $L-1$ cross-attention layers are independent of each other, allowing for parallel processing. A max-pooling layer is then employed to aggregate features after cross-attention. We obtain decoded", + "bbox": [ + 511, + 675, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16653", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1cf1dd152ac4f10e494115256d7b0d5b3cd4d6e1d4851006aa050e406e416a35.jpg", + "image_caption": [ + "Figure 2. System overview. Given an input monocular RGB video, we apply a sliding window mechanism to convert it into overlapping clips (referred to as windows). Each window is fed into an Image-to-Points (I2P) network to recover 3D points in a local coordinate system. Next, the local points are incrementally fed into a Local-to-World (L2W) network to create a globally consistent scene model. The proposed I2P and L2W networks elegantly share similar architectures. In the I2P step (Sec. 3.1), we select a keyframe as a reference to set up a local coordinate system and use the remaining frames in the window to estimate the 3D geometry captured within it. The points from the first window are used to establish the world coordinate system. We then incrementally fuse the following windows in the L2W step (Sec. 3.2). This process involves retrieving the most relevant already-registered keyframes as a reference, and integrating new keyframes. Through this iterative process, we eventually obtain the full scene reconstruction." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 903, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "keyframe tokens $G_{key}$ as:", + "bbox": [ + 89, + 411, + 264, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {k e y} = D _ {k e y} (F _ {k e y}, F _ {s u p _ {1}}, \\dots , F _ {s u p _ {L - 1}}).\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 439, + 423, + 458 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9328c85455681f3182ac79be86dcd56cc43692a32c381184ebc4032850bcf15b.jpg", + "image_caption": [ + "Figure 3. Illustration of a decoder block in the proposed keyframe decoder $D_{key}$ . We present a minimalist modification to integrate information from different supporting images. Our approach traverses each of them, selects its token keys and values, and uses the keyframe queries to interact with them separately across the supporting images. This multi-view information is then aggregated through max-pooling. The registration decoder $D_{reg}$ and scene decoder $D_{sce}$ (described in Sec. 3.2) share the same architecture." + ], + "image_footnote": [], + "bbox": [ + 104, + 484, + 472, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supporting decoder. The supporting decoder $D_{sup}$ is designed to complement the keyframe decoder. It inherits the decoder architecture used in DUSt3R, consisting of $n$ standard ViT decoder blocks. The cross-attention mechanism is applied only to exchange information with the keyframe.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that all supporting frames share the same $D_{sup}$ . This process is denoted as", + "bbox": [ + 511, + 411, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {s u p _ {i}} = D _ {s u p} (F _ {s u p _ {i}}, F _ {k e y}), i = 1, \\dots , L - 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 449, + 864, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Points reconstruction. Similar to DUSt3R, we apply a linear head [64] to regress dense 3D pointmaps in the unified coordinate system from decoded tokens. In addition to the pointmaps, we also predict the confidence maps for all frames to evaluate their reliability. The final predictions are:", + "bbox": [ + 511, + 472, + 906, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {i} ^ {(H \\times W \\times 3)}, \\hat {C} _ {i} ^ {(H \\times W \\times 1)} = \\mathrm {H} (G _ {i} ^ {(T \\times d)}), i = 1, \\dots , L.\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 565, + 885, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training loss. Following DUSt3R, the I2P network is trained end-to-end using ground-truth scene points $\\{X_{i}\\}_{i = 1}^{L}$ . Both the ground truth and predicted point maps are normalized to a canonical scale, determined by the average distance of all valid points within the window to the origin. The confidence-aware training loss is:", + "bbox": [ + 511, + 593, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I 2 P} = \\sum_ {i = 1} ^ {L} M _ {i} \\cdot (\\hat {C} _ {i} \\cdot \\mathrm {L} 1 (\\frac {1}{\\hat {z}} \\hat {X} _ {i}, \\frac {1}{z} X _ {i}) - \\alpha \\log \\hat {C} _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 702, + 879, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{i}$ represents a mask of valid points that have ground-truth values in $X_{i}$ , $z$ and $\\hat{z}$ are the scale factor, $\\hat{C}_i$ is the confidence map, the operator $\\cdot$ denotes the elementwise matrix multiplication, $\\mathrm{L1}(\\cdot)$ denotes the point-wise Euclidean distance, and $\\alpha$ is a hyper-parameter to control the regularization term. We will detail the process in Sec 4.", + "bbox": [ + 511, + 750, + 906, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Inter-Window Global Registration", + "text_level": 1, + "bbox": [ + 511, + 847, + 815, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the 3D pointmap $\\{\\hat{X}_{key}\\}$ from the I2P network, we use the inter-window Local-to-World (L2W)", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16654", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model to incrementally register the newly generated pointmap into a global 3D coordinate system. Similar to the I2P network, the L2W model also relies on some frames to serve as a reference for the scene coordinate system. Furthermore, it can leverage multiple registered keyframes as a global reference. These registered keyframes are referred to as scene frames and they are maintained in a buffering set through a sampling mechanism.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The buffering set is designed for scalability in handling long videos. We apply a reservoir strategy [59] that maintains a maximum of $B$ registered frames in the buffering set. When a new keyframe is inferred from I2P and ready for fusion, we retrieve the top- $K$ best-correlated scene frames from the buffering set as its support for global registration.", + "bbox": [ + 89, + 213, + 483, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene initialization. The first window is used to initialize the scene model. It's crucial to ensure this initialization is as accurate as possible. To achieve this, we execute the I2P network $L$ times, attempting to traverse and designate each frame within the window as the keyframe. We then select the result with the highest total confidence score for scene model initialization. This process results in a scene pointcloud along with a set of registered frames. All these frames are regarded as scene frames, and they are used to initialize the buffering set.", + "bbox": [ + 89, + 305, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reservoir and retrieval. Each scene frame $I_{sce_i}$ is recorded with its latent feature $F_{sce_i}$ and pointmaps $\\tilde{X}_{sce_i}$ . For efficiency, we apply reservoir sampling to allow storing an unbiased subset of an empirical distribution in a bounded amount of memory. The first $B$ registered frames chosen are directly inserted into the buffering set. For each subsequent frame with $id > B$ , the probability of inserting it is $B / id$ . If chosen for insertion, it will randomly replace one of the current scene frames in the buffering set.", + "bbox": [ + 89, + 458, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a new keyframe $I_{key}$ to be registered, we feed its feature $F_{key}$ and the features from the buffering set into a retrieval module,", + "bbox": [ + 89, + 595, + 483, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e t r i e v a l} \\left(F _ {\\text {k e y}} ^ {(T \\times d)}, \\left\\{F _ {\\text {s c e} _ {i}} ^ {(T \\times d)} \\right\\}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 652, + 388, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to obtain a list of correlation scores, measuring both the visual similarity and baseline suitability between the keyframe and the scene frames in the buffering set. The retrieval module uses the first $r$ decoder blocks from the I2P module as its backbone. A linear projection and an average-pooling layer follow, together producing an image-wise correlation score. We then select the top- $K$ scene frames as a global reference to fuse the current keyframe. As a result, we have $K$ scene frames and one keyframe as the input for the following L2W model.", + "bbox": [ + 89, + 686, + 483, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Points embedding. The 3D pointmaps reconstructed by the I2P model are encoded into the L2W model using a patch embedding method similar to image patchification in the ViT encoder $E_{img}$ . We process the new keyframe and $K$", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "retrieved scene frames in parallel as:", + "bbox": [ + 511, + 90, + 756, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {i} ^ {(T \\times d)} = E _ {p t s} (\\hat {X} _ {i} ^ {(H \\times W \\times 3)}), i = 1, \\dots , K + 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 118, + 869, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The encoded geometric tokens are combined with their corresponding visual tokens by", + "bbox": [ + 511, + 152, + 903, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {i} ^ {(T \\times d)} = F _ {i} ^ {(T \\times d)} + \\mathcal {P} _ {i} ^ {(T \\times d)}, i = 1, \\dots , K + 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 195, + 869, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This resulting a token set $\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}$ contains joint features of image patch appearance and 3D geometry for the keyframe and retrieved scene frames. In the following decoders, $\\{\\mathcal{P}\\}$ are further accumulated to $\\{\\mathcal{F}\\}$ between adjacent blocks to enhance the geometric representation.", + "bbox": [ + 511, + 228, + 903, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Registration decoder. The registration decoder $D_{reg}$ takes feature tokens $\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}$ as input and aims to transform the local reconstruction of the keyframe to the scene coordinate system. It takes the same network architecture of the keyframe decoder $D_{key}$ . This decoding process is denoted by", + "bbox": [ + 511, + 305, + 903, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {k e y} = D _ {r e g} \\left(\\mathcal {F} _ {k e y}, \\mathcal {F} _ {s c e _ {1}}, \\dots , \\mathcal {F} _ {s c e _ {K}}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 409, + 836, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene decoder. The scene decoder $D_{sce}$ takes the token set $\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}$ as input to refine the scene geometry without coordinate system changes. It uses the same network architecture as the keyframe decoder $D_{key}$ , allowing us to extend to multi-keyframe co-registration (see supplementary material for details). By default, we register one keyframe each time. Each of the $\\mathcal{F}_{sce_i}$ has information exchange only with the $\\mathcal{F}_{key}$ . This decoding process is denoted by", + "bbox": [ + 511, + 440, + 905, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {s c e _ {i}} = D _ {s c e} \\left(\\mathcal {F} _ {s c e _ {i}}, \\mathcal {F} _ {k e y}\\right), i = 1, \\dots , K.\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 590, + 846, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Points reconstruction and training loss. We apply the same head design as that of the I2P network to predict all the pointmaps $\\tilde{X}_i$ in the global scene coordinate system:", + "bbox": [ + 511, + 622, + 905, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} _ {i} ^ {(H \\times W \\times 3)}, \\tilde {C} _ {i} ^ {(H \\times W \\times 1)} = \\mathrm {H} (\\mathcal {G} _ {i} ^ {(T \\times d)}), i = 1, \\dots , K + 1.\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 678, + 900, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the L2W network using a similar loss function as the I2P network. Differently, no normalization is applied to the predicted point map, as the output scale must align with the scene frames in the input. This alignment ensures that the output can be directly integrated into the existing reconstruction. The training loss of the L2W network is:", + "bbox": [ + 511, + 712, + 905, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L 2 W} = \\sum_ {i = 1} ^ {L} M _ {i} \\cdot \\left(\\tilde {C} _ {i} \\cdot \\mathrm {L} 1 \\left(\\tilde {X} _ {i}, X _ {i}\\right) - \\alpha \\log \\tilde {C} _ {i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 816, + 869, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The following section provides a detailed discussion of the training process and its implementation details.", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16655", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/69f6649d353ad4a9524d8bf7436d7d906249f454b0eacece42760ab847fdce73.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChess Acc. / Comp.Fire Acc. / Comp.Heads Acc. / Comp.Office Acc. / Comp.Pumpkin Acc. / Comp.RedKitchen Acc. / Comp.Stairs Acc. / Comp.Average Acc. / Comp.FPS
DUS3R [64]2.26 / 2.131.04 / 1.501.66 / 0.984.62 / 4.741.73 / 2.431.95 / 2.363.37 / 10.752.19 / 3.24<1
MAS3R [28]2.08 / 2.121.54 / 1.431.06 / 1.043.23 / 3.195.68 / 3.073.50 / 3.372.36 / 13.163.04 / 3.90≪1
Spann3R [61]2.23 / 1.680.88 / 0.922.67 / 0.985.86 / 3.542.25 / 1.852.68 / 1.805.65 / 5.153.42 / 2.41>50
SLAM3R-NoConf (Ours)2.12 / 1.210.95 / 0.803.23 / 1.672.59 / 2.211.99 / 2.042.09 / 1.884.54 / 6.382.40 / 2.24~25
SLAM3R (Ours)1.63 / 1.310.84 / 0.832.95 / 1.222.32 / 2.261.81 / 2.051.84 / 1.944.19 / 6.912.13 / 2.34~25
", + "bbox": [ + 122, + 88, + 874, + 176 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e604eaee299e3ec9c1ebc931ac85b23d5a9542484657d87d5e1452707100459b.jpg", + "table_caption": [ + "Table 1. Reconstruction results on 7 Scenes [51] dataset. The average numbers are computed over all test sequences. The methods are categorized into two groups based on whether their FPS is above or below 1. The best results within each category are shown in bold. We report accuracy and completeness in centimeters. The color gradient shifts from red through yellow to green to show increasing FPS." + ], + "table_footnote": [], + "table_body": "
MethodRoom 0 Acc. / Comp.Room 1 Acc. / Comp.Room 2 Acc. / Comp.Office 0 Acc. / Comp.Office 1 Acc. / Comp.Office 2 Acc. / Comp.Office 3 Acc. / Comp.Office 4 Acc. / Comp.Average Acc. / Comp.FPS
DUSt3R [64]3.47 / 2.502.53 / 1.862.95 / 1.764.92 / 3.513.09 / 2.214.01 / 3.103.27 / 2.253.66 / 2.613.49 / 2.48<1
MASt3R [28]4.01 / 4.103.61 / 3.253.13 / 2.152.57 / 1.6312.85 / 8.133.13 / 1.994.67 / 3.153.69 / 2.474.71 / 3.36≤1
NICER-SLAM [79]*2.53 / 3.043.93 / 4.103.40 / 3.425.49 / 6.093.45 / 4.424.02 / 4.293.34 / 4.033.03 / 3.873.65 / 4.16≤1
DROID-SLAM [56]*12.18 / 8.968.35 / 6.073.26 / 16.013.01 / 16.192.39 / 16.205.66 / 15.564.49 / 9.734.65 / 9.635.50 / 12.29~20
DIM-SLAM [29]*14.19 / 6.249.56 / 6.458.41 / 12.1710.16 / 5.957.86 / 8.3316.50 / 8.2813.01 / 6.7713.08 / 8.6211.60 / 7.85~3
GO-SLAM [75]--------3.81 / 4.79~8
Spann3R [61]9.75 / 12.9415.51 / 12.947.28 / 8.505.46 / 18.755.24 / 16.649.33 / 11.8016.00 / 9.0313.97 / 16.0210.32 / 13.33>50
SLAM3R-NoConf (Ours)3.37 / 2.403.22 / 2.333.15 / 2.004.43 / 2.593.18 / 2.343.95 / 2.784.20 / 3.154.57 / 3.383.76 / 2.62~24
SLAM3R (Ours)3.19 / 2.403.12 / 2.342.72 / 2.004.28 / 2.603.17 / 2.343.84 / 2.783.90 / 3.164.32 / 3.363.57 / 2.62~24
", + "bbox": [ + 94, + 241, + 903, + 368 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/51955a9f1835514b360d08862e49c9291fe34f773e4ac771d1f1f4c26a816fe5.jpg", + "table_caption": [ + "Table 2. Reconstruction results on Replica [54] dataset. * denotes the results reported in NICER-SLAM." + ], + "table_footnote": [], + "table_body": "
DUSt3R [64]MASt3R [28]NICER-SLAM [79]*DROID-SLAM [56]*DIM-SLAM [29]GO-SLAM [75]Spann3R [61]SLAM3R-NoConf (Ours)SLAM3R (Ours)
7 Scenes8.026.288.555.66--11.708.448.41
Replica4.761.671.880.330.460.3932.796.616.61
", + "bbox": [ + 104, + 406, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Camera pose results evaluated by ATE-RMSE (cm) on 7 Scenes [51] and Replica [54] datasets.", + "bbox": [ + 189, + 457, + 805, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44345ea07f6baf0eb2ac594b6ac8db7b6c7d52b1a43792d0a855fbf0d20ce29a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 488, + 267, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d20d4f1c45c1c75db0ac26467fa3204ec21b1a1b942d11bcba92fecb577fcd3.jpg", + "image_caption": [ + "Spann3R", + "Figure 4. We visualize the reconstruction results on two scenes: Office-09 and Office 2 from the 7-Scenes [51] and Replica [54] datasets. Our method runs in real-time and achieves high-quality reconstruction comparable to the offline method DUSt3R [64]." + ], + "image_footnote": [], + "bbox": [ + 98, + 583, + 233, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0af3d4918d0308fc33eea8a0a5322a26ce8641a9335ec7451d5a7732d39b607b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 488, + 470, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/309d4c9fb0ceeec648887a6c95d84d1f1c9ef8404ce2af3f9b1b1cad7a5dade8.jpg", + "image_caption": [ + "DUSt3R" + ], + "image_footnote": [], + "bbox": [ + 269, + 584, + 472, + 662 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/200ae69a4d04f2b4fc8008b95a78ccbf15cded675d5513c4eb3fc283b1f31a58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 488, + 684, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5453a8c6eea0f90eef1faa2e3e7c4def1e186093507c7db842f3b8018e9598c1.jpg", + "image_caption": [ + "SLAM3R (Ours)" + ], + "image_footnote": [], + "bbox": [ + 478, + 584, + 684, + 662 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5b41e2e7e51997716d37255fafedb4a2504a3de2d032030de7833a298909481.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 488, + 898, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/beb59d5df17aed1ca7a0573c6a38e9e65349dc67a5ac2f453bf0d63b8a748181.jpg", + "image_caption": [ + "Ground-truth" + ], + "image_footnote": [], + "bbox": [ + 687, + 584, + 898, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 744, + 223, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For both the Image-to-Points (I2P) and Local-to-World (L2W) models, we perform training with a mixture of three datasets: ScanNet++ [71], Aria Synthetic Environments [3] and CO3D-v2 [41]. These datasets vary from scene-level to object-centric, and contain both real-world and synthetic scenes. Since they are all recorded sequentially, we can easily extract video clips with a sliding-window mechanism as our training data. We select about", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "850K clips for training in total. To validate our reconstruction quality, we conduct quantitative evaluations on two unseen datasets: 7 Scenes [51], a real-world dataset of partial scenes, and Replica [54], a synthetic dataset of complete scenes. We also demonstrate visual reconstruction results across diverse datasets and in-the-wild captured videos to showcase the generalization ability of SLAM3R.", + "bbox": [ + 511, + 746, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Both the I2P and L2W models build upon the architecture of DUSt3R [64] with minimal but effective modifications, making it natural for them to", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16656", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c0ed92335b18e80c89f9545ce3b3994448a23ce6aff48003530a111283e3731.jpg", + "image_caption": [ + "Figure 5. Qualitative examples. We show our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild captured videos, to demonstrate SLAM3R's generalization ability." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "initialize their weights from the DUSt3R pre-trained model. We initialize our weights using the DUSt3R model trained on $224 \\times 224$ resolution images with $m = 24$ encoder blocks, $n = 12$ decoder blocks with linear heads. All images are center-cropped to $224 \\times 224$ pixels before feeding into SLAM3R. Our training is conducted on 8 NVIDIA 4090D GPUs, each with 24 GB of memory. It takes about one day. At test time, we set the initial window length to $L = 5$ to ensure high-quality reconstruction of all frames within the window. For subsequent incremental windows, we use $L = 11$ to provide more supporting views for better keyframe reconstruction. Please refer to our supplementary material for more implementation details.", + "bbox": [ + 88, + 449, + 485, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparisons", + "text_level": 1, + "bbox": [ + 89, + 661, + 232, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. Following NICER-SLAM [79] and Spann3R [61], we build a ground-truth point cloud model for each test sequence by back-projecting pixels to the world using ground-truth depths and camera parameters. The reconstructed point clouds are aligned to ground truths using Umayama [58] and ICP [43] algorithms. We quantify reconstruction quality through accuracy and completeness metrics. To demonstrate computational efficiency, we report frames per second (FPS) on a single NVIDIA 4090D GPU. We also evaluate camera poses using absolute trajectory error (ATE-RMSE). For detailed formulations of the metrics, please refer to the supplementary material.", + "bbox": [ + 89, + 686, + 483, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction results on the 7 Scenes [51] dataset. The numerical results of scene reconstruction quality are re", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ported in Table 1. Following Spann3R [61]'s setting, we uniformly sample one-twentieth of the frames in each test sequence as input video. Each video is regarded as an individual scene. We evaluate SLAM3R using two settings: integrating the full pointmaps predicted for all input frames to create reconstruction results (denoted by SLAM3R-NoConf), and filtering pointmaps with a confidence threshold of 3 before creating reconstruction results (SLAM3R). We compare our method with optimization-based reconstruction DUSt3R [64], triangulation-based MASt3R [28], and online incremental reconstruction Spann3R. Our method outperforms all baselines in both accuracy and completeness while maintaining real-time performance. As shown in the Office-09 scene (the top row in Figure 4), our approach demonstrates much less drift compared to the concurrent work Spann3R [61].", + "bbox": [ + 511, + 449, + 908, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction results on the Replica [54] dataset. Besides the baselines mentioned in 7 Scene datasets, we also compare the SLAM-based reconstruction approaches NICER-SLAM [79], DROID-SLAM [56], DIM-SLAM [29] and GO-SLAM [75] on the Replica [54] dataset. The numerical results on full scene reconstruction are reported in Table 2. Due to the memory constraint, DUSt3R [64] and MASt3R [28] process only one-twentieth of the frames for reconstruction. As is shown in the table, our method surpasses all baselines with FPS greater than 1. Notably, without any optimization procedure, our method achieves reconstruction quality comparable to optimization-based methods such as NICER-", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16657", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f0c070484c1535c94ab62ed796ef597f6a05faac4da175ef6e632d36df5d113.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# FramesAcc.Comp.FPS
DUSt3R [64]23.162.8942.55
I2P23.393.0442.55
I2P52.622.2840.82
I2P (Default)112.382.0340.11
I2P152.271.9435.51
I2P512.231.8611.97
", + "bbox": [ + 130, + 88, + 442, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SLAM [79] and DUS3R [64]. Example of the Office 2 (the bottom row in Figure 4) also illustrates the global consistency of our reconstruction result.", + "bbox": [ + 88, + 273, + 482, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Camera pose estimation on 7 Scenes [51] and Replica [54]. Our method is designed in a new paradigm that reconstructs 3D points end-to-end without explicitly solving camera parameters. Following DUSt3R [64], We also derive camera poses from the predicted scene points using PnP-RANSAC solver in OpenCV [6] with ground truth camera intrinsics of each frame. The results are reported in Table 3. We can observe that camera poses and scene reconstruction results are not fully positively correlated. This discrepancy between pose and reconstruction errors indicates that effective end-to-end 3D reconstruction is possible and promising without first obtaining precise camera poses.", + "bbox": [ + 88, + 319, + 482, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For more details on the comparisons in this section, please refer to our supplementary material.", + "bbox": [ + 89, + 502, + 482, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Analyses", + "text_level": 1, + "bbox": [ + 89, + 544, + 194, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the I2P model. To highlight the advantages of our multi-view I2P model over the original two-view DUSt3R [64], we evaluate the reconstruction quality of keyframes with varying numbers of supporting views. We conduct experiments on the Replica [54] dataset, where input views are sampled using a sliding window of different sizes, and the reconstruction accuracy and completeness of the keyframes are computed. The results are reported in Table 4. As the number of supporting views increases, our approach progressively improves reconstruction quality. Notably, the efficiency of our method remains stable until the window size exceeds 11, demonstrating the effectiveness of our parallel design. However, the results also show diminishing returns as the number of views increases, which we detail in the supplementary material. Visual results of I2P reconstruction can be found in Figure 1.", + "bbox": [ + 91, + 566, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Advantages of the L2W model. The effectiveness of the L2W model is evaluated through ablation studies on the Replica [54] dataset. Per-window reconstructions are first generated with a window size of 11 using the I2P model. Local points are then aligned to a unified coordinate frame using different methods: global optimization", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c98512c71147373f63fd7808dd9ba278dfd351dbd9a602e61c8431423fa1fcf0.jpg", + "table_caption": [ + "Table 4. Inner-window keyframe reconstruction results with various window lengths. By default, we use 11-frame windows for incremental reconstruction to balance quality and efficiency." + ], + "table_footnote": [], + "table_body": "
MethodAcc.Comp.FPS
I2P+GA4.873.00~3
I2P+UI7.473.86~1
I2P+L2W6.193.54~92
I2P+L2W+Re (Full)3.622.70~43
", + "bbox": [ + 573, + 88, + 848, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Reconstruction results using various point alignment methods and scene frame selection strategies. The FPS reported only accounts for the overhead of the alignment operation.", + "bbox": [ + 511, + 172, + 906, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from DUSt3R [64] (I2P-GA), traditional approaches such as Umeyama [58] and ICP [43] (I2P+UI), and our L2W model (I2P+L2W+Re). For consistency, we set the window size for global optimization to 10, which is equal to the number of views used to align new frames in other methods. Results in Table 5 show that our full method achieves superior alignment accuracy and computational efficiency compared to the alternatives.", + "bbox": [ + 511, + 244, + 903, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the retrieval module. We propose a lightweight retrieval module that selects historical scene frames from the reservoir. This approach effectively performs implicit re-localization. We compare our retrieval method with a baseline approach that selects the ten nearest previous frames, named I2P+L2W. The results in Table 5 indicate a significant performance improvement with our retrieval strategy, demonstrating its effectiveness.", + "bbox": [ + 511, + 368, + 903, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In-the-wild scene reconstruction. We have tested our method on a diverse range of unseen datasets and found that SLAM3R shows strong generalization capabilities. Figure 5 shows our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild videos we captured. These results show that our method performs reliably across different scales and scenes. We also provide additional numerical results on sampled scenes from these datasets in the supplementary material.", + "bbox": [ + 511, + 489, + 905, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 660, + 633, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present SLAM3R, a novel and effective system that performs real-time, high-quality, dense 3D scene reconstruction using RGB videos. It employs a two-hierarchy neural network framework to perform end-to-end 3D reconstruction through streamlined feed-forward processes, eliminating the need to explicitly solve any camera parameters. Experiments demonstrate its state-of-the-art reconstruction quality and real-time efficiency (20+ FPS).", + "bbox": [ + 511, + 686, + 903, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. The elimination of camera parameter prediction prevents us from performing global bundle adjustment. Additionally, the poses derived from our scene point cloud prediction still fall short of SLAM systems that specialize in camera localization. Addressing these limitations will be the focus of our future work.", + "bbox": [ + 511, + 810, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16658", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (2022ZD0160800). This work is also supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and in part by the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We thank the reviewers for their valuable feedback, as well as Zihan Zhu and Songyou Peng for their help with our experiments. Siyan Dong would also like to thank the support from HKU School of Computing and Data Science.", + "bbox": [ + 89, + 114, + 483, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 292, + 187, + 308 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2", + "[2] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 7, 8", + "[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescipt: Reconstructing scenes with an autoregressive structured language model. arXiv preprint arXiv:2403.13064, 2024. 6", + "[4] Tim Bailey and Hugh Durrant-Whyte. Simultaneous localization and mapping (slam): Part ii. IEEE robotics & automation magazine, 13(3):108-117, 2006. 2", + "[5] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 1, 2", + "[6] G Bradski. The opencv library. Dr. Dobb's Journal of Software Tools, 2000. 8", + "[7] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2", + "[8] Yiyang Chen, Siyan Dong, Xulong Wang, Lulu Cai, Youyi Zheng, and Yanchao Yang. Sg-nerf: Neural surface reconstruction with scene graph optimization. arXiv preprint arXiv:2407.12667, 2024. 2", + "[9] Chi-Ming Chung, Yang-Che Tseng, Ya-Ching Hsu, Xiang-Qian Shi, Yun-Hung Hua, Jia-Fong Yeh, Wen-Chin Chen, Yi-Ting Chen, and Winston H Hsu. Orbeez-slam: A real-time monocular visual slam with orb features and nefrealized mapping. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9400-9406. IEEE, 2023. 1, 2" + ], + "bbox": [ + 99, + 316, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jan Czarnowski, Tristan Laidlow, Ronald Clark, and Andrew J Davison. Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters, 5 (2):721-728, 2020. 1", + "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2", + "[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[14] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. arXiv preprint arXiv:2409.19152, 2024. 3", + "[15] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2", + "[16] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 1, 2", + "[17] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 1, 2", + "[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2", + "[19] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2", + "[20] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[21] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024. 2", + "[22] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 1, 2", + "[23] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16659", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2", + "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1", + "[25] Georg Klein and David Murray. Parallel tracking and mapping on a camera phone. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 83-86. IEEE, 2009. 1, 2", + "[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7, 8", + "[27] Lukas Koestler, Nan Yang, Niclas Zeller, and Daniel Cremers. Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In Conference on Robot Learning, pages 34-45. PMLR, 2022. 2", + "[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 6, 7", + "[29] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations, 2023. 1, 2, 6, 7", + "[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2", + "[31] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 1, 2", + "[32] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2", + "[33] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 1, 2", + "[34] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2", + "[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8", + "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 1, 2", + "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 1, 2", + "[39] Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision, pages 2320-2327. IEEE, 2011. 1, 2", + "[40] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1", + "[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6", + "[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3437-3444. IEEE, 2023. 1, 2", + "[43] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 7, 8", + "[44] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2", + "[45] Erik Sandström, Kevin Ta, Luc Van Gool, and Martin R Oswald. Uncle-slam: Uncertainty learning for dense neural slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4537-4548, 2023. 2", + "[46] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians. arXiv preprint arXiv:2405.16544, 2024. 1, 2", + "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 2", + "[48] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1, 2", + "[49] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16660", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 7, 8", + "[50] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 7, 8", + "[51] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 6, 7, 8", + "[52] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2", + "[53] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 1, 2", + "[54] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7, 8", + "[55] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6229-6238, 2021. 1, 2", + "[56] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6, 7", + "[57] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R Oswald, and Matteo Poggi. How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255, 4, 2024. 1", + "[58] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7, 8", + "[59] Jeffrey S Vitter. Random sampling with a reservoir. ACM Transactions on Mathematical Software (TOMS), 11(1):37-57, 1985. 5", + "[60] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8606-8615, 2022. 1, 2", + "[61] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 3, 6, 7" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2", + "[63] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2", + "[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7, 8", + "[65] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2", + "[66] Changchang Wu. Visualsfm: A visual structure from motion system. http://www.cs.washington.edu/homes/ccwu/vsfm, 2011. 1, 2", + "[67] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2", + "[68] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2", + "[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 7, 8", + "[70] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2", + "[71] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 6", + "[72] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2", + "[73] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 1, 2", + "[74] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16661", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 1, 2, 6, 7", + "[76] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 1, 2", + "[77] Heng Zhou, Zhetao Guo, Shuhong Liu, Lechen Zhang, Qihao Wang, Yuxiang Ren, and Mingrui Li. Mod-slam: Monocular dense mapping for unbounded 3d scene reconstruction. arXiv preprint arXiv:2402.03762, 2024. 1, 2", + "[78] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 1, 2", + "[79] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 1, 2, 6, 7, 8" + ], + "bbox": [ + 91, + 92, + 482, + 428 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16662", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_model.json b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6209db1f20e804df016a599320e087b8f22dbd79 --- /dev/null +++ b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_model.json @@ -0,0 +1,2743 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.131, + 0.903, + 0.152 + ], + "angle": 0, + "content": "SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.18, + 0.751, + 0.2 + ], + "angle": 0, + "content": "Yuzheng Liu\\(^{1*}\\) Siyan Dong\\(^{2*†}\\) Shuzhe Wang\\(^{3}\\) Yingda Yin\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.199, + 0.709, + 0.218 + ], + "angle": 0, + "content": "Yanchao Yang\\(^{2\\dagger}\\) Qingnan Fan\\(^{4}\\) Baoquan Chen\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.217, + 0.828, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Peking University \\(^{2}\\)The University of Hong Kong \\(^{3}\\)Aalto University \\(^{4}\\)VIVO" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.301, + 0.485, + 0.558 + ], + "angle": 0, + "content": "In this paper, we introduce SLAM3R, a novel and effective system for real-time, high-quality, dense 3D reconstruction using RGB videos. SLAM3R provides an end-to-end solution by seamlessly integrating local 3D reconstruction and global coordinate registration through feedforward neural networks. Given an input video, the system first converts it into overlapping clips using a sliding window mechanism. Unlike traditional pose optimization-based methods, SLAM3R directly regresses 3D pointmaps from RGB images in each window and progressively aligns and deforms these local pointmaps to create a globally consistent scene reconstruction - all without explicitly solving any camera parameters. Experiments across datasets consistently show that SLAM3R achieves state-of-the-art reconstruction accuracy and completeness while maintaining real-time performance at \\(20+\\) FPS. Code available at: https://github.com/PRU-VCL-3DV/SLAM3R." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.222, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Dense 3D reconstruction, a long-standing challenge in computer vision, aims to capture and reconstruct the detailed geometry of real-world scenes. Traditional approaches have largely relied on multi-stage pipelines. These typically begin with sparse Simultaneous Localization and Mapping (SLAM) [7, 16, 25, 37, 38] or Structure-from-Motion (SfM) [31, 33, 47, 53, 66] algorithms to estimate camera parameters, followed by Multi-View Stereo (MVS) [18, 48, 60, 68] techniques to fill in scene details. While these methods offer high-quality reconstructions, they often require offline processing to produce a complete model, which limits their applicability in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.483, + 0.868 + ], + "angle": 0, + "content": "In the literature, dense SLAM approaches [5, 9, 10, 17, 22, 39, 42, 55, 56, 76, 78] have been developed to address dense scene reconstruction as a complete system. However, these approaches often fall short in terms of reconstruction accuracy or completeness, or they rely heavily on" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.271, + 0.579, + 0.291 + ], + "angle": 0, + "content": "Monocular RGB video" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.291, + 0.579, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.472, + 0.555, + 0.483 + ], + "angle": 0, + "content": "：" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.486, + 0.579, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.272, + 0.687, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.339, + 0.607, + 0.391 + ], + "angle": 0, + "content": "rnnnne" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.271, + 0.77, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.295, + 0.797, + 0.303 + ], + "angle": 0, + "content": "···" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.271, + 0.879, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.887, + 0.296, + 0.899, + 0.303 + ], + "angle": 0, + "content": "···" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.33, + 0.777, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.331, + 0.894, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.404, + 0.776, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.404, + 0.897, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.468, + 0.897, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Figure 1. We introduce a novel dense reconstruction system - SLAM3R. It takes a monocular RGB video as input and reconstructs the scene as a dense pointcloud. The video is converted into short clips for local reconstruction (denoted as inner-window), which are then incrementally registered together (inter-window) to create a global scene model. This process runs in real-time, producing a reconstruction that is both accurate and complete." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.907, + 0.824 + ], + "angle": 0, + "content": "depth sensors. Recently, several monocular SLAM systems [29, 46, 73, 75, 77, 79] have been proposed to tackle dense scene reconstruction from RGB videos. By incorporating advanced scene representations [24, 36, 40, 57, 62], these systems produce accurate and complete scene reconstructions. However, this comes at the cost of reduced running efficiency. For example, NICER-SLAM [79] operates at a speed significantly below 1 FPS. Therefore, current approaches struggle with at least one of three key criteria: reconstruction accuracy, completeness, or efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "While monocular dense SLAM systems encounter the limitations mentioned earlier, recent advances in two-view geometry have shown promising potential. DUSt3R [64] introduces a purely end-to-end approach for learning dense reconstruction. Trained on large-scale datasets, its network" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.463, + 0.888 + ], + "angle": 0, + "content": "*Joint first authors: liu_yuzheng@stu.pku.edu.cn, siyan3d@hku.hk" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.239, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.463, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16651" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "is capable of producing high-quality dense reconstructions from paired images in real-time. However, for multiple views, a global optimization step is required to align these image pairs, which significantly hampers its efficiency. A concurrent work, Spann3R [61], extends DUSt3R to multiview (video) scenarios through an incremental pipeline with spatial memory. While this method accelerates the reconstruction process, it unfortunately results in noticeable accumulated drift and reduced reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.23, + 0.487, + 0.562 + ], + "angle": 0, + "content": "To address these challenges, we present SLAM3R (pronounced \"slæmər\"), a real-time dense 3D reconstruction system using RGB-only videos as input. Unlike traditional SLAM problems, SLAM3R performs implicit camera localization and focuses more on dense scene mapping, where 3R stands for 3D Reconstruction. SLAM3R comprises a two-hierarchy framework. First, it reconstructs local 3D geometry from a sliding window that processes short clips from the input video. Then, it progressively registers these local reconstructions to build a globally consistent 3D scene. Both modules are developed with simple yet effective feed-forward models, enabling end-to-end and efficient scene reconstruction. Specifically, the two modules are the Image-to-Points (I2P) network and the Local-to-World (L2W) network. The I2P module, inspired by DUSt3R, selects a keyframe in a local window as the coordinate system reference. It directly predicts the dense 3D point map supported by the remaining frames within that window. The L2W module incrementally fuses locally reconstructed points into a coherent global coordinate system. Both processes reconstruct the 3D points without explicitly estimating any camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Through extensive experiments, we demonstrate that SLAM3R provides high-quality scene reconstructions with minimal drift, outperforming existing dense SLAM systems across various benchmarks. Furthermore, SLAM3R achieves these results at \\(20+\\) FPS, bridging the gap between quality and efficiency in RGB-only dense scene reconstruction. Our contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.735 + ], + "angle": 0, + "content": "- We present a novel real-time end-to-end dense 3D reconstruction system that uses RGB videos to directly predict 3D pointmaps in a unified coordinate system through feed-forward neural networks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.736, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- Through careful design, our Image-to-Points module can process an arbitrary number of images simultaneously, effectively extending DUSt3R to handle multiple views and produce higher-quality predictions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.797, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- The proposed Local-to-World module directly aligns predicted local 3D pointmaps into a unified global coordinate system. This eliminates the need for explicit camera parameter estimation and costly global optimization." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.857, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We evaluate our method on multiple public benchmarks. It achieves state-of-the-art reconstruction quality in terms of both accuracy and completeness at real-time speeds." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.657, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.342 + ], + "angle": 0, + "content": "Traditional offline approaches. Dense 3D pointcloud reconstruction is a long-standing problem in computer vision. Classical approaches to this problem first determine camera parameters using Structure from Motion (SfM) [31, 33, 47, 53, 66], followed by dense 3D points triangulation with Multi-View Stereo (MVS) [1, 18, 48, 60, 68]. In recent years, neural implicit [8, 30, 36, 62, 65] and 3D Gaussian [12, 19, 20] representations have been applied to further enhance the quality of dense reconstruction. While these methods deliver high-quality results, they have a significant limitation: the requirement for offline processing to generate the final 3D model, which restricts their applicability in real-time scenarios. In this paper, we focus on online dense reconstruction in the context of Simultaneous Localization and Mapping (SLAM)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.629 + ], + "angle": 0, + "content": "Dense SLAM. Early works on SLAM [4, 7, 15, 16, 25, 37, 38] focused on reconstructing the structure of unknown environments while simultaneously localizing camera poses. These approaches prioritize real-time performance but produce only sparse structures of the scene. Dense SLAM approaches [5, 9-11, 17, 22, 27, 39, 42, 55, 56, 76, 78] incorporate detailed scene geometry information to improve pose estimation. DROID-SLAM [56] introduces recurrent iterative updates of camera poses and pixel-wise depth estimates, while TANDEM [27] proposes an online MVS module for depth prediction. These systems enable real-time dense scene reconstruction. However, their focus on camera trajectory accuracy often results in incomplete and noisy 3D reconstruction. Neural implicit and Gaussian representations have also been integrated with dense SLAM systems [9, 21-23, 32, 34, 42, 42, 44, 45, 55, 67, 72, 78]. However, these approaches often rely on additional depth sensors or focus primarily on novel view synthesis rather than producing detailed geometric reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.905, + 0.809 + ], + "angle": 0, + "content": "More recently, several monocular dense SLAM systems [29, 46, 73, 75, 77, 79] have been developed to produce dense scene geometry reconstruction. A notable limitation of these systems is their slow runtime. Among these systems, GO-SLAM [75] achieves a speed of \\(\\sim 8\\) FPS, which still falls short of real-time capability. Furthermore, these methods all share a common strategy: they alternate between solving for camera poses and estimating the scene representation. In contrast, this paper presents a novel approach to dense scene reconstruction that eliminates the need for explicitly solving camera parameters, offering a more efficient and streamlined solution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.902 + ], + "angle": 0, + "content": "End-to-end dense 3D reconstruction. DUSt3R [64] introduces the first purely end-to-end dense 3D reconstruction pipeline without relying on camera parameters. Recently, several works have adopted a similar approach for single-view reconstruction [63], feature matching [28], novel view synthesis [52, 70], and dynamics reconstruction [74]. These" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "successes demonstrate the effectiveness of end-to-end dense point prediction, inspiring us to develop a dense reconstruction system with a similar methodology." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.138, + 0.483, + 0.455 + ], + "angle": 0, + "content": "While DUSt3R operates in real-time for two-view predictions, its extension to multiple views involves exhaustive pairing images and performing an additional global optimization step. This process significantly increases computational time, thereby hindering its real-time performance. MASt3R [28] enhances the matching capability of DUSt3R by adding a match head, achieving more accurate keypoint correspondences for 3D reconstruction [14], but at the cost of increased computational time. More recently, the concurrent work Spann3R [61] extends DUSt3R with spatial memory. It takes a video as input and performs incremental scene reconstruction in a unified coordinate system without requiring global optimization. While this approach significantly improves runtime efficiency, the incremental reconstruction pipeline frame by frame leads to noticeable accumulated drift. Unlike Spann3R, our networks at each hierarchy take multiple frames as input to minimize drift. Additionally, we propose a self-contained retrieval module that, when registering a new frame, this module selects not only its previous few frames but also other similar frames from long-term history for more global scene reference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.182, + 0.484 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Problem statement. Given a monocular video consisting of a sequence of RGB image frames \\(\\{I_i\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{i = 1}^N\\) that captures a static scene, the goal is to reconstruct its dense 3D pointcloud \\(P\\in \\mathbb{R}^{M\\times 3}\\), where \\(M\\) is the number of 3D points. Our work focuses on three key objectives: maximizing 3D points recovery for reconstruction completeness, improving the accuracy of each recovered point, and achieving these goals while preserving real-time performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.75 + ], + "angle": 0, + "content": "System overview. Figure 2 illustrates an overview of the proposed dense reconstruction system. It consists of two main components: an Image-to-Points (I2P) network that recovers local 3D points from video clips, and a Local-to-World (L2W) network that registers local reconstructions into a global scene coordinate system. During the reconstruction of the dense point cloud, the system does not explicitly solve any camera parameters. Instead, it directly predicts 3D point maps in unified coordinate systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The system starts by applying a sliding window mechanism of length \\( L \\) to convert the input video into short clips \\( \\{\\mathcal{W}_i \\in \\mathbb{R}^{L \\times H \\times W \\times 3}\\} \\). The I2P network then processes each window \\( \\mathcal{W}_i \\) to recover local 3D pointmaps. Within each window, the system selects a keyframe to define a reference coordinate system for point reconstruction, as detailed in Sec. 3.1. By default, the stride of the sliding window is set to 1, ensuring each input frame in the video is selected at least once as a keyframe. For global scene reconstruction, we initialize the world coordinate system with the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "first window and use the reconstructed frames (image and local point map produced by the I2P) as input for the L2W model. The L2W model incrementally register these local reconstructions into a unified global 3D coordinate system. To ensure both accuracy and efficiency during this process, the system maintains a limited reservoir of registered frames, called scene frames. Whenever the L2W model registers a new keyframe, we retrieve the best-correlated scene frames as a reference. The details are introduced in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.833, + 0.251 + ], + "angle": 0, + "content": "3.1. Inner-Window Local Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.379 + ], + "angle": 0, + "content": "The Image-to-Points (I2P) model aims to infer dense 3D pointmaps for every pixel of a keyframe in a given video clip. By default, the middle image of a window \\(\\mathcal{W}\\) is chosen as the keyframe \\(I_{key}\\) to define the local coordinate system, as it is most likely to have the largest overlap with other frames. The remaining images \\(\\{I_{sup_i}\\}_{i = 1}^{L - 1}\\) serve as supporting frames. Note that the 3D pointmaps of supporting frames can also be reconstructed through I2P." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.501 + ], + "angle": 0, + "content": "The I2P network draws inspiration from DUSt3R [64], originally designed for stereo 3D reconstruction. We introduce several simple yet effective modifications to extend it for multi-view scenarios. The I2P model uses a multibranch Vision Transformer (ViT) [13] as its backbone. It consists of a shared encoder \\( E_{img} \\), two separate decoders \\( D_{key} \\) and \\( D_{sup} \\), and a point regression head for final prediction. These components are detailed below." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.501, + 0.905, + 0.592 + ], + "angle": 0, + "content": "Image encoder. For a given video clip, the image encoder \\( E_{img} \\) encodes each frame \\( I_{i} \\) to obtain token representations \\( F_{i} \\in \\mathbb{R}^{T \\times d} \\), where \\( T \\) is the number of tokens and \\( d \\) is the token dimension. The encoder \\( E_{img} \\) comprises \\( m \\) ViT encoder blocks, each containing self-attention and feed-forward layers. The encoding process is denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.599, + 0.855, + 0.621 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {(T \\times d)} = E _ {i m g} (I _ {i} ^ {(H \\times W \\times 3)}), i = 1, \\dots , L.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.675 + ], + "angle": 0, + "content": "The frames are processed independently and in parallel, with the output divided into two parts: \\( F_{key} \\) for the keyframe and \\( \\{F_{sup_i}\\}_{i=1}^{L-1} \\) for the supporting frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.676, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Keyframe decoder. The keyframe decoder \\( D_{key} \\) consists of \\( n \\) ViT decoder blocks, each containing self-attention, cross-attention, and feed-forward layers. Unlike DUSt3R which uses the standard cross-attention, we introduce a novel multi-view cross-attention to combine information from different supporting frames. Given the feature tokens \\( F_{key} \\) and \\( \\{F_{sup_i}\\}_{i=1}^{L-1} \\), the keyframe decoder \\( D_{key} \\) takes \\( F_{key} \\) as input for self-attention and performs cross-attention between \\( F_{key} \\) and \\( \\{F_{sup_i}\\}_{i=1}^{L-1} \\). A decoder block is illustrated in Figure 3. For each cross-attention layer, queries are taken from \\( F_{key} \\), while keys and values are extracted from the supporting tokens \\( F_{sup_i} \\). These \\( L-1 \\) cross-attention layers are independent of each other, allowing for parallel processing. A max-pooling layer is then employed to aggregate features after cross-attention. We obtain decoded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16653" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.087, + 0.904, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Figure 2. System overview. Given an input monocular RGB video, we apply a sliding window mechanism to convert it into overlapping clips (referred to as windows). Each window is fed into an Image-to-Points (I2P) network to recover 3D points in a local coordinate system. Next, the local points are incrementally fed into a Local-to-World (L2W) network to create a globally consistent scene model. The proposed I2P and L2W networks elegantly share similar architectures. In the I2P step (Sec. 3.1), we select a keyframe as a reference to set up a local coordinate system and use the remaining frames in the window to estimate the 3D geometry captured within it. The points from the first window are used to establish the world coordinate system. We then incrementally fuse the following windows in the L2W step (Sec. 3.2). This process involves retrieving the most relevant already-registered keyframes as a reference, and integrating new keyframes. Through this iterative process, we eventually obtain the full scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.412, + 0.265, + 0.429 + ], + "angle": 0, + "content": "keyframe tokens \\(G_{key}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.44, + 0.424, + 0.459 + ], + "angle": 0, + "content": "\\[\nG _ {k e y} = D _ {k e y} (F _ {k e y}, F _ {s u p _ {1}}, \\dots , F _ {s u p _ {L - 1}}).\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.485, + 0.473, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Figure 3. Illustration of a decoder block in the proposed keyframe decoder \\( D_{key} \\). We present a minimalist modification to integrate information from different supporting images. Our approach traverses each of them, selects its token keys and values, and uses the keyframe queries to interact with them separately across the supporting images. This multi-view information is then aggregated through max-pooling. The registration decoder \\( D_{reg} \\) and scene decoder \\( D_{sce} \\) (described in Sec. 3.2) share the same architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Supporting decoder. The supporting decoder \\( D_{sup} \\) is designed to complement the keyframe decoder. It inherits the decoder architecture used in DUSt3R, consisting of \\( n \\) standard ViT decoder blocks. The cross-attention mechanism is applied only to exchange information with the keyframe." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.412, + 0.907, + 0.443 + ], + "angle": 0, + "content": "Note that all supporting frames share the same \\( D_{sup} \\). This process is denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.45, + 0.866, + 0.468 + ], + "angle": 0, + "content": "\\[\nG _ {s u p _ {i}} = D _ {s u p} (F _ {s u p _ {i}}, F _ {k e y}), i = 1, \\dots , L - 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Points reconstruction. Similar to DUSt3R, we apply a linear head [64] to regress dense 3D pointmaps in the unified coordinate system from decoded tokens. In addition to the pointmaps, we also predict the confidence maps for all frames to evaluate their reliability. The final predictions are:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.566, + 0.886, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {i} ^ {(H \\times W \\times 3)}, \\hat {C} _ {i} ^ {(H \\times W \\times 1)} = \\mathrm {H} (G _ {i} ^ {(T \\times d)}), i = 1, \\dots , L.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.594, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Training loss. Following DUSt3R, the I2P network is trained end-to-end using ground-truth scene points \\(\\{X_{i}\\}_{i = 1}^{L}\\). Both the ground truth and predicted point maps are normalized to a canonical scale, determined by the average distance of all valid points within the window to the origin. The confidence-aware training loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.703, + 0.88, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I 2 P} = \\sum_ {i = 1} ^ {L} M _ {i} \\cdot (\\hat {C} _ {i} \\cdot \\mathrm {L} 1 (\\frac {1}{\\hat {z}} \\hat {X} _ {i}, \\frac {1}{z} X _ {i}) - \\alpha \\log \\hat {C} _ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.841 + ], + "angle": 0, + "content": "where \\( M_{i} \\) represents a mask of valid points that have ground-truth values in \\( X_{i} \\), \\( z \\) and \\( \\hat{z} \\) are the scale factor, \\( \\hat{C}_i \\) is the confidence map, the operator \\( \\cdot \\) denotes the elementwise matrix multiplication, \\( \\mathrm{L1}(\\cdot) \\) denotes the point-wise Euclidean distance, and \\( \\alpha \\) is a hyper-parameter to control the regularization term. We will detail the process in Sec 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.816, + 0.864 + ], + "angle": 0, + "content": "3.2. Inter-Window Global Registration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.907, + 0.901 + ], + "angle": 0, + "content": "After obtaining the 3D pointmap \\(\\{\\hat{X}_{key}\\}\\) from the I2P network, we use the inter-window Local-to-World (L2W)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16654" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.212 + ], + "angle": 0, + "content": "model to incrementally register the newly generated pointmap into a global 3D coordinate system. Similar to the I2P network, the L2W model also relies on some frames to serve as a reference for the scene coordinate system. Furthermore, it can leverage multiple registered keyframes as a global reference. These registered keyframes are referred to as scene frames and they are maintained in a buffering set through a sampling mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.305 + ], + "angle": 0, + "content": "The buffering set is designed for scalability in handling long videos. We apply a reservoir strategy [59] that maintains a maximum of \\( B \\) registered frames in the buffering set. When a new keyframe is inferred from I2P and ready for fusion, we retrieve the top- \\( K \\) best-correlated scene frames from the buffering set as its support for global registration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Scene initialization. The first window is used to initialize the scene model. It's crucial to ensure this initialization is as accurate as possible. To achieve this, we execute the I2P network \\( L \\) times, attempting to traverse and designate each frame within the window as the keyframe. We then select the result with the highest total confidence score for scene model initialization. This process results in a scene pointcloud along with a set of registered frames. All these frames are regarded as scene frames, and they are used to initialize the buffering set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Reservoir and retrieval. Each scene frame \\( I_{sce_i} \\) is recorded with its latent feature \\( F_{sce_i} \\) and pointmaps \\( \\tilde{X}_{sce_i} \\). For efficiency, we apply reservoir sampling to allow storing an unbiased subset of an empirical distribution in a bounded amount of memory. The first \\( B \\) registered frames chosen are directly inserted into the buffering set. For each subsequent frame with \\( id > B \\), the probability of inserting it is \\( B / id \\). If chosen for insertion, it will randomly replace one of the current scene frames in the buffering set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Given a new keyframe \\( I_{key} \\) to be registered, we feed its feature \\( F_{key} \\) and the features from the buffering set into a retrieval module," + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.654, + 0.39, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e t r i e v a l} \\left(F _ {\\text {k e y}} ^ {(T \\times d)}, \\left\\{F _ {\\text {s c e} _ {i}} ^ {(T \\times d)} \\right\\}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.839 + ], + "angle": 0, + "content": "to obtain a list of correlation scores, measuring both the visual similarity and baseline suitability between the keyframe and the scene frames in the buffering set. The retrieval module uses the first \\( r \\) decoder blocks from the I2P module as its backbone. A linear projection and an average-pooling layer follow, together producing an image-wise correlation score. We then select the top- \\( K \\) scene frames as a global reference to fuse the current keyframe. As a result, we have \\( K \\) scene frames and one keyframe as the input for the following L2W model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Points embedding. The 3D pointmaps reconstructed by the I2P model are encoded into the L2W model using a patch embedding method similar to image patchification in the ViT encoder \\( E_{img} \\). We process the new keyframe and \\( K \\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.757, + 0.107 + ], + "angle": 0, + "content": "retrieved scene frames in parallel as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.119, + 0.87, + 0.14 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {i} ^ {(T \\times d)} = E _ {p t s} (\\hat {X} _ {i} ^ {(H \\times W \\times 3)}), i = 1, \\dots , K + 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.184 + ], + "angle": 0, + "content": "The encoded geometric tokens are combined with their corresponding visual tokens by" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.196, + 0.87, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {i} ^ {(T \\times d)} = F _ {i} ^ {(T \\times d)} + \\mathcal {P} _ {i} ^ {(T \\times d)}, i = 1, \\dots , K + 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.905, + 0.304 + ], + "angle": 0, + "content": "This resulting a token set \\(\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}\\) contains joint features of image patch appearance and 3D geometry for the keyframe and retrieved scene frames. In the following decoders, \\(\\{\\mathcal{P}\\}\\) are further accumulated to \\(\\{\\mathcal{F}\\}\\) between adjacent blocks to enhance the geometric representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.397 + ], + "angle": 0, + "content": "Registration decoder. The registration decoder \\( D_{reg} \\) takes feature tokens \\( \\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\} \\) as input and aims to transform the local reconstruction of the keyframe to the scene coordinate system. It takes the same network architecture of the keyframe decoder \\( D_{key} \\). This decoding process is denoted by" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.41, + 0.838, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {k e y} = D _ {r e g} \\left(\\mathcal {F} _ {k e y}, \\mathcal {F} _ {s c e _ {1}}, \\dots , \\mathcal {F} _ {s c e _ {K}}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.906, + 0.577 + ], + "angle": 0, + "content": "Scene decoder. The scene decoder \\( D_{sce} \\) takes the token set \\( \\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\} \\) as input to refine the scene geometry without coordinate system changes. It uses the same network architecture as the keyframe decoder \\( D_{key} \\), allowing us to extend to multi-keyframe co-registration (see supplementary material for details). By default, we register one keyframe each time. Each of the \\( \\mathcal{F}_{sce_i} \\) has information exchange only with the \\( \\mathcal{F}_{key} \\). This decoding process is denoted by" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.591, + 0.848, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {s c e _ {i}} = D _ {s c e} \\left(\\mathcal {F} _ {s c e _ {i}}, \\mathcal {F} _ {k e y}\\right), i = 1, \\dots , K.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.623, + 0.906, + 0.668 + ], + "angle": 0, + "content": "Points reconstruction and training loss. We apply the same head design as that of the I2P network to predict all the pointmaps \\(\\tilde{X}_i\\) in the global scene coordinate system:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.679, + 0.901, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} _ {i} ^ {(H \\times W \\times 3)}, \\tilde {C} _ {i} ^ {(H \\times W \\times 1)} = \\mathrm {H} (\\mathcal {G} _ {i} ^ {(T \\times d)}), i = 1, \\dots , K + 1.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.805 + ], + "angle": 0, + "content": "We train the L2W network using a similar loss function as the I2P network. Differently, no normalization is applied to the predicted point map, as the output scale must align with the scene frames in the input. This alignment ensures that the output can be directly integrated into the existing reconstruction. The training loss of the L2W network is:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.817, + 0.87, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L 2 W} = \\sum_ {i = 1} ^ {L} M _ {i} \\cdot \\left(\\tilde {C} _ {i} \\cdot \\mathrm {L} 1 \\left(\\tilde {X} _ {i}, X _ {i}\\right) - \\alpha \\log \\tilde {C} _ {i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The following section provides a detailed discussion of the training process and its implementation details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16655" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.875, + 0.177 + ], + "angle": 0, + "content": "
MethodChess Acc. / Comp.Fire Acc. / Comp.Heads Acc. / Comp.Office Acc. / Comp.Pumpkin Acc. / Comp.RedKitchen Acc. / Comp.Stairs Acc. / Comp.Average Acc. / Comp.FPS
DUS3R [64]2.26 / 2.131.04 / 1.501.66 / 0.984.62 / 4.741.73 / 2.431.95 / 2.363.37 / 10.752.19 / 3.24<1
MAS3R [28]2.08 / 2.121.54 / 1.431.06 / 1.043.23 / 3.195.68 / 3.073.50 / 3.372.36 / 13.163.04 / 3.90≪1
Spann3R [61]2.23 / 1.680.88 / 0.922.67 / 0.985.86 / 3.542.25 / 1.852.68 / 1.805.65 / 5.153.42 / 2.41>50
SLAM3R-NoConf (Ours)2.12 / 1.210.95 / 0.803.23 / 1.672.59 / 2.211.99 / 2.042.09 / 1.884.54 / 6.382.40 / 2.24~25
SLAM3R (Ours)1.63 / 1.310.84 / 0.832.95 / 1.222.32 / 2.261.81 / 2.051.84 / 1.944.19 / 6.912.13 / 2.34~25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.187, + 0.908, + 0.23 + ], + "angle": 0, + "content": "Table 1. Reconstruction results on 7 Scenes [51] dataset. The average numbers are computed over all test sequences. The methods are categorized into two groups based on whether their FPS is above or below 1. The best results within each category are shown in bold. We report accuracy and completeness in centimeters. The color gradient shifts from red through yellow to green to show increasing FPS." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.242, + 0.905, + 0.369 + ], + "angle": 0, + "content": "
MethodRoom 0 Acc. / Comp.Room 1 Acc. / Comp.Room 2 Acc. / Comp.Office 0 Acc. / Comp.Office 1 Acc. / Comp.Office 2 Acc. / Comp.Office 3 Acc. / Comp.Office 4 Acc. / Comp.Average Acc. / Comp.FPS
DUSt3R [64]3.47 / 2.502.53 / 1.862.95 / 1.764.92 / 3.513.09 / 2.214.01 / 3.103.27 / 2.253.66 / 2.613.49 / 2.48<1
MASt3R [28]4.01 / 4.103.61 / 3.253.13 / 2.152.57 / 1.6312.85 / 8.133.13 / 1.994.67 / 3.153.69 / 2.474.71 / 3.36≤1
NICER-SLAM [79]*2.53 / 3.043.93 / 4.103.40 / 3.425.49 / 6.093.45 / 4.424.02 / 4.293.34 / 4.033.03 / 3.873.65 / 4.16≤1
DROID-SLAM [56]*12.18 / 8.968.35 / 6.073.26 / 16.013.01 / 16.192.39 / 16.205.66 / 15.564.49 / 9.734.65 / 9.635.50 / 12.29~20
DIM-SLAM [29]*14.19 / 6.249.56 / 6.458.41 / 12.1710.16 / 5.957.86 / 8.3316.50 / 8.2813.01 / 6.7713.08 / 8.6211.60 / 7.85~3
GO-SLAM [75]--------3.81 / 4.79~8
Spann3R [61]9.75 / 12.9415.51 / 12.947.28 / 8.505.46 / 18.755.24 / 16.649.33 / 11.8016.00 / 9.0313.97 / 16.0210.32 / 13.33>50
SLAM3R-NoConf (Ours)3.37 / 2.403.22 / 2.333.15 / 2.004.43 / 2.593.18 / 2.343.95 / 2.784.20 / 3.154.57 / 3.383.76 / 2.62~24
SLAM3R (Ours)3.19 / 2.403.12 / 2.342.72 / 2.004.28 / 2.603.17 / 2.343.84 / 2.783.90 / 3.164.32 / 3.363.57 / 2.62~24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.189, + 0.379, + 0.807, + 0.392 + ], + "angle": 0, + "content": "Table 2. Reconstruction results on Replica [54] dataset. * denotes the results reported in NICER-SLAM." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.407, + 0.891, + 0.448 + ], + "angle": 0, + "content": "
DUSt3R [64]MASt3R [28]NICER-SLAM [79]*DROID-SLAM [56]*DIM-SLAM [29]GO-SLAM [75]Spann3R [61]SLAM3R-NoConf (Ours)SLAM3R (Ours)
7 Scenes8.026.288.555.66--11.708.448.41
Replica4.761.671.880.330.460.3932.796.616.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.19, + 0.458, + 0.807, + 0.472 + ], + "angle": 0, + "content": "Table 3. Camera pose results evaluated by ATE-RMSE (cm) on 7 Scenes [51] and Replica [54] datasets." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.489, + 0.268, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.584, + 0.235, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.668, + 0.199, + 0.68 + ], + "angle": 0, + "content": "Spann3R" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.489, + 0.471, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.585, + 0.473, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.668, + 0.398, + 0.679 + ], + "angle": 0, + "content": "DUSt3R" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.489, + 0.685, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.585, + 0.685, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.667, + 0.621, + 0.68 + ], + "angle": 0, + "content": "SLAM3R (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.489, + 0.899, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.585, + 0.899, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.668, + 0.835, + 0.679 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.693, + 0.908, + 0.723 + ], + "angle": 0, + "content": "Figure 4. We visualize the reconstruction results on two scenes: Office-09 and Office 2 from the 7-Scenes [51] and Replica [54] datasets. Our method runs in real-time and achieves high-quality reconstruction comparable to the offline method DUSt3R [64]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.224, + 0.764 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. For both the Image-to-Points (I2P) and Local-to-World (L2W) models, we perform training with a mixture of three datasets: ScanNet++ [71], Aria Synthetic Environments [3] and CO3D-v2 [41]. These datasets vary from scene-level to object-centric, and contain both real-world and synthetic scenes. Since they are all recorded sequentially, we can easily extract video clips with a sliding-window mechanism as our training data. We select about" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.907, + 0.854 + ], + "angle": 0, + "content": "850K clips for training in total. To validate our reconstruction quality, we conduct quantitative evaluations on two unseen datasets: 7 Scenes [51], a real-world dataset of partial scenes, and Replica [54], a synthetic dataset of complete scenes. We also demonstrate visual reconstruction results across diverse datasets and in-the-wild captured videos to showcase the generalization ability of SLAM3R." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation details. Both the I2P and L2W models build upon the architecture of DUSt3R [64] with minimal but effective modifications, making it natural for them to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16656" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.904, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 5. Qualitative examples. We show our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild captured videos, to demonstrate SLAM3R's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.486, + 0.648 + ], + "angle": 0, + "content": "initialize their weights from the DUSt3R pre-trained model. We initialize our weights using the DUSt3R model trained on \\(224 \\times 224\\) resolution images with \\(m = 24\\) encoder blocks, \\(n = 12\\) decoder blocks with linear heads. All images are center-cropped to \\(224 \\times 224\\) pixels before feeding into SLAM3R. Our training is conducted on 8 NVIDIA 4090D GPUs, each with 24 GB of memory. It takes about one day. At test time, we set the initial window length to \\(L = 5\\) to ensure high-quality reconstruction of all frames within the window. For subsequent incremental windows, we use \\(L = 11\\) to provide more supporting views for better keyframe reconstruction. Please refer to our supplementary material for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.662, + 0.233, + 0.679 + ], + "angle": 0, + "content": "4.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.485, + 0.869 + ], + "angle": 0, + "content": "Evaluation metrics. Following NICER-SLAM [79] and Spann3R [61], we build a ground-truth point cloud model for each test sequence by back-projecting pixels to the world using ground-truth depths and camera parameters. The reconstructed point clouds are aligned to ground truths using Umayama [58] and ICP [43] algorithms. We quantify reconstruction quality through accuracy and completeness metrics. To demonstrate computational efficiency, we report frames per second (FPS) on a single NVIDIA 4090D GPU. We also evaluate camera poses using absolute trajectory error (ATE-RMSE). For detailed formulations of the metrics, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Reconstruction results on the 7 Scenes [51] dataset. The numerical results of scene reconstruction quality are re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.693 + ], + "angle": 0, + "content": "ported in Table 1. Following Spann3R [61]'s setting, we uniformly sample one-twentieth of the frames in each test sequence as input video. Each video is regarded as an individual scene. We evaluate SLAM3R using two settings: integrating the full pointmaps predicted for all input frames to create reconstruction results (denoted by SLAM3R-NoConf), and filtering pointmaps with a confidence threshold of 3 before creating reconstruction results (SLAM3R). We compare our method with optimization-based reconstruction DUSt3R [64], triangulation-based MASt3R [28], and online incremental reconstruction Spann3R. Our method outperforms all baselines in both accuracy and completeness while maintaining real-time performance. As shown in the Office-09 scene (the top row in Figure 4), our approach demonstrates much less drift compared to the concurrent work Spann3R [61]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reconstruction results on the Replica [54] dataset. Besides the baselines mentioned in 7 Scene datasets, we also compare the SLAM-based reconstruction approaches NICER-SLAM [79], DROID-SLAM [56], DIM-SLAM [29] and GO-SLAM [75] on the Replica [54] dataset. The numerical results on full scene reconstruction are reported in Table 2. Due to the memory constraint, DUSt3R [64] and MASt3R [28] process only one-twentieth of the frames for reconstruction. As is shown in the table, our method surpasses all baselines with FPS greater than 1. Notably, without any optimization procedure, our method achieves reconstruction quality comparable to optimization-based methods such as NICER-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16657" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.444, + 0.194 + ], + "angle": 0, + "content": "
Method# FramesAcc.Comp.FPS
DUSt3R [64]23.162.8942.55
I2P23.393.0442.55
I2P52.622.2840.82
I2P (Default)112.382.0340.11
I2P152.271.9435.51
I2P512.231.8611.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.205, + 0.483, + 0.247 + ], + "angle": 0, + "content": "Table 4. Inner-window keyframe reconstruction results with various window lengths. By default, we use 11-frame windows for incremental reconstruction to balance quality and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.274, + 0.483, + 0.319 + ], + "angle": 0, + "content": "SLAM [79] and DUS3R [64]. Example of the Office 2 (the bottom row in Figure 4) also illustrates the global consistency of our reconstruction result." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Camera pose estimation on 7 Scenes [51] and Replica [54]. Our method is designed in a new paradigm that reconstructs 3D points end-to-end without explicitly solving camera parameters. Following DUSt3R [64], We also derive camera poses from the predicted scene points using PnP-RANSAC solver in OpenCV [6] with ground truth camera intrinsics of each frame. The results are reported in Table 3. We can observe that camera poses and scene reconstruction results are not fully positively correlated. This discrepancy between pose and reconstruction errors indicates that effective end-to-end 3D reconstruction is possible and promising without first obtaining precise camera poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.534 + ], + "angle": 0, + "content": "For more details on the comparisons in this section, please refer to our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.545, + 0.196, + 0.56 + ], + "angle": 0, + "content": "4.2. Analyses" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Effectiveness of the I2P model. To highlight the advantages of our multi-view I2P model over the original two-view DUSt3R [64], we evaluate the reconstruction quality of keyframes with varying numbers of supporting views. We conduct experiments on the Replica [54] dataset, where input views are sampled using a sliding window of different sizes, and the reconstruction accuracy and completeness of the keyframes are computed. The results are reported in Table 4. As the number of supporting views increases, our approach progressively improves reconstruction quality. Notably, the efficiency of our method remains stable until the window size exceeds 11, demonstrating the effectiveness of our parallel design. However, the results also show diminishing returns as the number of views increases, which we detail in the supplementary material. Visual results of I2P reconstruction can be found in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Advantages of the L2W model. The effectiveness of the L2W model is evaluated through ablation studies on the Replica [54] dataset. Per-window reconstructions are first generated with a window size of 11 using the I2P model. Local points are then aligned to a unified coordinate frame using different methods: global optimization" + }, + { + "type": "table", + "bbox": [ + 0.574, + 0.089, + 0.849, + 0.164 + ], + "angle": 0, + "content": "
MethodAcc.Comp.FPS
I2P+GA4.873.00~3
I2P+UI7.473.86~1
I2P+L2W6.193.54~92
I2P+L2W+Re (Full)3.622.70~43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.174, + 0.907, + 0.216 + ], + "angle": 0, + "content": "Table 5. Reconstruction results using various point alignment methods and scene frame selection strategies. The FPS reported only accounts for the overhead of the alignment operation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.905, + 0.367 + ], + "angle": 0, + "content": "from DUSt3R [64] (I2P-GA), traditional approaches such as Umeyama [58] and ICP [43] (I2P+UI), and our L2W model (I2P+L2W+Re). For consistency, we set the window size for global optimization to 10, which is equal to the number of views used to align new frames in other methods. Results in Table 5 show that our full method achieves superior alignment accuracy and computational efficiency compared to the alternatives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.49 + ], + "angle": 0, + "content": "Analysis of the retrieval module. We propose a lightweight retrieval module that selects historical scene frames from the reservoir. This approach effectively performs implicit re-localization. We compare our retrieval method with a baseline approach that selects the ten nearest previous frames, named I2P+L2W. The results in Table 5 indicate a significant performance improvement with our retrieval strategy, demonstrating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.643 + ], + "angle": 0, + "content": "In-the-wild scene reconstruction. We have tested our method on a diverse range of unseen datasets and found that SLAM3R shows strong generalization capabilities. Figure 5 shows our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild videos we captured. These results show that our method performs reliably across different scales and scenes. We also provide additional numerical results on sampled scenes from these datasets in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.634, + 0.677 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In this paper, we present SLAM3R, a novel and effective system that performs real-time, high-quality, dense 3D scene reconstruction using RGB videos. It employs a two-hierarchy neural network framework to perform end-to-end 3D reconstruction through streamlined feed-forward processes, eliminating the need to explicitly solve any camera parameters. Experiments demonstrate its state-of-the-art reconstruction quality and real-time efficiency (20+ FPS)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Limitations and future work. The elimination of camera parameter prediction prevents us from performing global bundle adjustment. Additionally, the poses derived from our scene point cloud prediction still fall short of SLAM systems that specialize in camera localization. Addressing these limitations will be the focus of our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16658" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.283 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (2022ZD0160800). This work is also supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and in part by the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We thank the reviewers for their valuable feedback, as well as Zihan Zhu and Songyou Peng for their help with our experiments. Siyan Dong would also like to thank the support from HKU School of Computing and Data Science." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.293, + 0.188, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[2] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescipt: Reconstructing scenes with an autoregressive structured language model. arXiv preprint arXiv:2403.13064, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[4] Tim Bailey and Hugh Durrant-Whyte. Simultaneous localization and mapping (slam): Part ii. IEEE robotics & automation magazine, 13(3):108-117, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.651 + ], + "angle": 0, + "content": "[5] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.484, + 0.679 + ], + "angle": 0, + "content": "[6] G Bradski. The opencv library. Dr. Dobb's Journal of Software Tools, 2000. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[7] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[8] Yiyang Chen, Siyan Dong, Xulong Wang, Lulu Cai, Youyi Zheng, and Yanchao Yang. Sg-nerf: Neural surface reconstruction with scene graph optimization. arXiv preprint arXiv:2407.12667, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Chi-Ming Chung, Yang-Che Tseng, Ya-Ching Hsu, Xiang-Qian Shi, Yun-Hung Hua, Jia-Fong Yeh, Wen-Chin Chen, Yi-Ting Chen, and Winston H Hsu. Orbeez-slam: A real-time monocular visual slam with orb features and nefrealized mapping. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9400-9406. IEEE, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[10] Jan Czarnowski, Tristan Laidlow, Ronald Clark, and Andrew J Davison. Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters, 5 (2):721-728, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[14] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. arXiv preprint arXiv:2409.19152, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[15] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[16] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[17] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.561 + ], + "angle": 0, + "content": "[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.908, + 0.632 + ], + "angle": 0, + "content": "[19] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.689 + ], + "angle": 0, + "content": "[20] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[21] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[22] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16659" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[25] Georg Klein and David Murray. Parallel tracking and mapping on a camera phone. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 83-86. IEEE, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[27] Lukas Koestler, Nan Yang, Niclas Zeller, and Daniel Cremers. Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In Conference on Robot Learning, pages 34-45. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[29] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[31] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[32] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[33] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[34] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[39] Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision, pages 2320-2327. IEEE, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[40] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3437-3444. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[43] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[44] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[45] Erik Sandström, Kevin Ta, Luc Van Gool, and Martin R Oswald. Uncle-slam: Uncertainty learning for dense neural slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4537-4548, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[46] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians. arXiv preprint arXiv:2405.16544, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[48] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16660" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[50] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[51] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[52] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[53] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[54] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[55] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6229-6238, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[56] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[57] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R Oswald, and Matteo Poggi. How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255, 4, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[58] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[59] Jeffrey S Vitter. Random sampling with a reservoir. ACM Transactions on Mathematical Software (TOMS), 11(1):37-57, 1985. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[60] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8606-8615, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[61] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[63] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.295, + 0.905, + 0.364 + ], + "angle": 0, + "content": "[65] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.367, + 0.904, + 0.407 + ], + "angle": 0, + "content": "[66] Changchang Wu. Visualsfm: A visual structure from motion system. http://www.cs.washington.edu/homes/ccwu/vsfm, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.905, + 0.479 + ], + "angle": 0, + "content": "[67] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.904, + 0.538 + ], + "angle": 0, + "content": "[68] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.61 + ], + "angle": 0, + "content": "[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.613, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[70] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[71] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[72] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[73] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[74] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16661" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[75] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[76] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[77] Heng Zhou, Zhetao Guo, Shuhong Liu, Lechen Zhang, Qihao Wang, Yuxiang Ren, and Mingrui Li. Mod-slam: Monocular dense mapping for unbounded 3d scene reconstruction. arXiv preprint arXiv:2402.03762, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[78] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[79] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16662" + } + ] +] \ No newline at end of file diff --git a/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_origin.pdf b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30c77e29df2d57024b69fbae1eac168fd52d44cd --- /dev/null +++ b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/2aa67e72-2cf5-4fb7-b0f5-0552ddd9811a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa0cdf4d247aa508c21f4d451af3b5c87b09dd244a0f1be998722b84ad5a26ae +size 5539355 diff --git a/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/full.md b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6ad53187f82289aef370a7541dfaa3ad515b8342 --- /dev/null +++ b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/full.md @@ -0,0 +1,376 @@ +# SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos + +Yuzheng Liu $^{1*}$ Siyan Dong $^{2*†}$ Shuzhe Wang $^{3}$ Yingda Yin $^{1}$ + +Yanchao Yang $^{2\dagger}$ Qingnan Fan $^{4}$ Baoquan Chen $^{1\dagger}$ + +$^{1}$ Peking University $^{2}$ The University of Hong Kong $^{3}$ Aalto University $^{4}$ VIVO + +# Abstract + +In this paper, we introduce SLAM3R, a novel and effective system for real-time, high-quality, dense 3D reconstruction using RGB videos. SLAM3R provides an end-to-end solution by seamlessly integrating local 3D reconstruction and global coordinate registration through feedforward neural networks. Given an input video, the system first converts it into overlapping clips using a sliding window mechanism. Unlike traditional pose optimization-based methods, SLAM3R directly regresses 3D pointmaps from RGB images in each window and progressively aligns and deforms these local pointmaps to create a globally consistent scene reconstruction - all without explicitly solving any camera parameters. Experiments across datasets consistently show that SLAM3R achieves state-of-the-art reconstruction accuracy and completeness while maintaining real-time performance at $20+$ FPS. Code available at: https://github.com/PRU-VCL-3DV/SLAM3R. + +# 1. Introduction + +Dense 3D reconstruction, a long-standing challenge in computer vision, aims to capture and reconstruct the detailed geometry of real-world scenes. Traditional approaches have largely relied on multi-stage pipelines. These typically begin with sparse Simultaneous Localization and Mapping (SLAM) [7, 16, 25, 37, 38] or Structure-from-Motion (SfM) [31, 33, 47, 53, 66] algorithms to estimate camera parameters, followed by Multi-View Stereo (MVS) [18, 48, 60, 68] techniques to fill in scene details. While these methods offer high-quality reconstructions, they often require offline processing to produce a complete model, which limits their applicability in real-world scenarios. + +In the literature, dense SLAM approaches [5, 9, 10, 17, 22, 39, 42, 55, 56, 76, 78] have been developed to address dense scene reconstruction as a complete system. However, these approaches often fall short in terms of reconstruction accuracy or completeness, or they rely heavily on + +![](images/0db8d85bac62fd1d9a1399553068be176822bf31718e87fa364a43f5043dae38.jpg) +Monocular RGB video +： +rnnnne + +![](images/cdf5f9486d9d2ae216eea8cde4677a19cd7b2b0fa006afa190e2f61c71fa8007.jpg) + +![](images/5b1725e7c08c93b6714b35e83d89801688addcb8e6bf3235a30dcb048dc049df.jpg) + +![](images/6f9cac0419478a57d85aa30cf7d2a15ad0b7c97ae3bae5c0ebd692514623326b.jpg) + +![](images/2e0ae777d0d36307b7da4c032205aec8aff4db120412f6902a211f13b2663123.jpg) +··· +··· + +![](images/c7d22570d7d62c3fde62ae775babab89d99b1e606a4591121ef032ec7978cc21.jpg) + +![](images/198e72296ee59f9ada4d8ae8f1bfb953c3a1d6566dc895b63c5cce010087da21.jpg) + +![](images/ab733a38f11d22d11d4e114711cc2c8ab07f9145858444a9d3b4fb7987543ab3.jpg) +Figure 1. We introduce a novel dense reconstruction system - SLAM3R. It takes a monocular RGB video as input and reconstructs the scene as a dense pointcloud. The video is converted into short clips for local reconstruction (denoted as inner-window), which are then incrementally registered together (inter-window) to create a global scene model. This process runs in real-time, producing a reconstruction that is both accurate and complete. + +![](images/5f43ce300897134d6b32ab3c7b7c425094180c6b9f0cce2d3f7cd66873c5cc07.jpg) + +![](images/d96334ff76ea5b7f73d0a489e23f810f23d0b8e13a738866f6e409a60f0508c2.jpg) + +depth sensors. Recently, several monocular SLAM systems [29, 46, 73, 75, 77, 79] have been proposed to tackle dense scene reconstruction from RGB videos. By incorporating advanced scene representations [24, 36, 40, 57, 62], these systems produce accurate and complete scene reconstructions. However, this comes at the cost of reduced running efficiency. For example, NICER-SLAM [79] operates at a speed significantly below 1 FPS. Therefore, current approaches struggle with at least one of three key criteria: reconstruction accuracy, completeness, or efficiency. + +While monocular dense SLAM systems encounter the limitations mentioned earlier, recent advances in two-view geometry have shown promising potential. DUSt3R [64] introduces a purely end-to-end approach for learning dense reconstruction. Trained on large-scale datasets, its network + +is capable of producing high-quality dense reconstructions from paired images in real-time. However, for multiple views, a global optimization step is required to align these image pairs, which significantly hampers its efficiency. A concurrent work, Spann3R [61], extends DUSt3R to multiview (video) scenarios through an incremental pipeline with spatial memory. While this method accelerates the reconstruction process, it unfortunately results in noticeable accumulated drift and reduced reconstruction quality. + +To address these challenges, we present SLAM3R (pronounced "slæmər"), a real-time dense 3D reconstruction system using RGB-only videos as input. Unlike traditional SLAM problems, SLAM3R performs implicit camera localization and focuses more on dense scene mapping, where 3R stands for 3D Reconstruction. SLAM3R comprises a two-hierarchy framework. First, it reconstructs local 3D geometry from a sliding window that processes short clips from the input video. Then, it progressively registers these local reconstructions to build a globally consistent 3D scene. Both modules are developed with simple yet effective feed-forward models, enabling end-to-end and efficient scene reconstruction. Specifically, the two modules are the Image-to-Points (I2P) network and the Local-to-World (L2W) network. The I2P module, inspired by DUSt3R, selects a keyframe in a local window as the coordinate system reference. It directly predicts the dense 3D point map supported by the remaining frames within that window. The L2W module incrementally fuses locally reconstructed points into a coherent global coordinate system. Both processes reconstruct the 3D points without explicitly estimating any camera parameters. + +Through extensive experiments, we demonstrate that SLAM3R provides high-quality scene reconstructions with minimal drift, outperforming existing dense SLAM systems across various benchmarks. Furthermore, SLAM3R achieves these results at $20+$ FPS, bridging the gap between quality and efficiency in RGB-only dense scene reconstruction. Our contributions are summarized below: + +- We present a novel real-time end-to-end dense 3D reconstruction system that uses RGB videos to directly predict 3D pointmaps in a unified coordinate system through feed-forward neural networks. +- Through careful design, our Image-to-Points module can process an arbitrary number of images simultaneously, effectively extending DUSt3R to handle multiple views and produce higher-quality predictions. +- The proposed Local-to-World module directly aligns predicted local 3D pointmaps into a unified global coordinate system. This eliminates the need for explicit camera parameter estimation and costly global optimization. +- We evaluate our method on multiple public benchmarks. It achieves state-of-the-art reconstruction quality in terms of both accuracy and completeness at real-time speeds. + +# 2. Related Work + +Traditional offline approaches. Dense 3D pointcloud reconstruction is a long-standing problem in computer vision. Classical approaches to this problem first determine camera parameters using Structure from Motion (SfM) [31, 33, 47, 53, 66], followed by dense 3D points triangulation with Multi-View Stereo (MVS) [1, 18, 48, 60, 68]. In recent years, neural implicit [8, 30, 36, 62, 65] and 3D Gaussian [12, 19, 20] representations have been applied to further enhance the quality of dense reconstruction. While these methods deliver high-quality results, they have a significant limitation: the requirement for offline processing to generate the final 3D model, which restricts their applicability in real-time scenarios. In this paper, we focus on online dense reconstruction in the context of Simultaneous Localization and Mapping (SLAM). + +Dense SLAM. Early works on SLAM [4, 7, 15, 16, 25, 37, 38] focused on reconstructing the structure of unknown environments while simultaneously localizing camera poses. These approaches prioritize real-time performance but produce only sparse structures of the scene. Dense SLAM approaches [5, 9-11, 17, 22, 27, 39, 42, 55, 56, 76, 78] incorporate detailed scene geometry information to improve pose estimation. DROID-SLAM [56] introduces recurrent iterative updates of camera poses and pixel-wise depth estimates, while TANDEM [27] proposes an online MVS module for depth prediction. These systems enable real-time dense scene reconstruction. However, their focus on camera trajectory accuracy often results in incomplete and noisy 3D reconstruction. Neural implicit and Gaussian representations have also been integrated with dense SLAM systems [9, 21-23, 32, 34, 42, 42, 44, 45, 55, 67, 72, 78]. However, these approaches often rely on additional depth sensors or focus primarily on novel view synthesis rather than producing detailed geometric reconstruction. + +More recently, several monocular dense SLAM systems [29, 46, 73, 75, 77, 79] have been developed to produce dense scene geometry reconstruction. A notable limitation of these systems is their slow runtime. Among these systems, GO-SLAM [75] achieves a speed of $\sim 8$ FPS, which still falls short of real-time capability. Furthermore, these methods all share a common strategy: they alternate between solving for camera poses and estimating the scene representation. In contrast, this paper presents a novel approach to dense scene reconstruction that eliminates the need for explicitly solving camera parameters, offering a more efficient and streamlined solution. + +End-to-end dense 3D reconstruction. DUSt3R [64] introduces the first purely end-to-end dense 3D reconstruction pipeline without relying on camera parameters. Recently, several works have adopted a similar approach for single-view reconstruction [63], feature matching [28], novel view synthesis [52, 70], and dynamics reconstruction [74]. These + +successes demonstrate the effectiveness of end-to-end dense point prediction, inspiring us to develop a dense reconstruction system with a similar methodology. + +While DUSt3R operates in real-time for two-view predictions, its extension to multiple views involves exhaustive pairing images and performing an additional global optimization step. This process significantly increases computational time, thereby hindering its real-time performance. MASt3R [28] enhances the matching capability of DUSt3R by adding a match head, achieving more accurate keypoint correspondences for 3D reconstruction [14], but at the cost of increased computational time. More recently, the concurrent work Spann3R [61] extends DUSt3R with spatial memory. It takes a video as input and performs incremental scene reconstruction in a unified coordinate system without requiring global optimization. While this approach significantly improves runtime efficiency, the incremental reconstruction pipeline frame by frame leads to noticeable accumulated drift. Unlike Spann3R, our networks at each hierarchy take multiple frames as input to minimize drift. Additionally, we propose a self-contained retrieval module that, when registering a new frame, this module selects not only its previous few frames but also other similar frames from long-term history for more global scene reference. + +# 3. Method + +Problem statement. Given a monocular video consisting of a sequence of RGB image frames $\{I_i\in \mathbb{R}^{H\times W\times 3}\}_{i = 1}^N$ that captures a static scene, the goal is to reconstruct its dense 3D pointcloud $P\in \mathbb{R}^{M\times 3}$ , where $M$ is the number of 3D points. Our work focuses on three key objectives: maximizing 3D points recovery for reconstruction completeness, improving the accuracy of each recovered point, and achieving these goals while preserving real-time performance. + +System overview. Figure 2 illustrates an overview of the proposed dense reconstruction system. It consists of two main components: an Image-to-Points (I2P) network that recovers local 3D points from video clips, and a Local-to-World (L2W) network that registers local reconstructions into a global scene coordinate system. During the reconstruction of the dense point cloud, the system does not explicitly solve any camera parameters. Instead, it directly predicts 3D point maps in unified coordinate systems. + +The system starts by applying a sliding window mechanism of length $L$ to convert the input video into short clips $\{\mathcal{W}_i \in \mathbb{R}^{L \times H \times W \times 3}\}$ . The I2P network then processes each window $\mathcal{W}_i$ to recover local 3D pointmaps. Within each window, the system selects a keyframe to define a reference coordinate system for point reconstruction, as detailed in Sec. 3.1. By default, the stride of the sliding window is set to 1, ensuring each input frame in the video is selected at least once as a keyframe. For global scene reconstruction, we initialize the world coordinate system with the + +first window and use the reconstructed frames (image and local point map produced by the I2P) as input for the L2W model. The L2W model incrementally register these local reconstructions into a unified global 3D coordinate system. To ensure both accuracy and efficiency during this process, the system maintains a limited reservoir of registered frames, called scene frames. Whenever the L2W model registers a new keyframe, we retrieve the best-correlated scene frames as a reference. The details are introduced in Sec. 3.2. + +# 3.1. Inner-Window Local Reconstruction + +The Image-to-Points (I2P) model aims to infer dense 3D pointmaps for every pixel of a keyframe in a given video clip. By default, the middle image of a window $\mathcal{W}$ is chosen as the keyframe $I_{key}$ to define the local coordinate system, as it is most likely to have the largest overlap with other frames. The remaining images $\{I_{sup_i}\}_{i = 1}^{L - 1}$ serve as supporting frames. Note that the 3D pointmaps of supporting frames can also be reconstructed through I2P. + +The I2P network draws inspiration from DUSt3R [64], originally designed for stereo 3D reconstruction. We introduce several simple yet effective modifications to extend it for multi-view scenarios. The I2P model uses a multibranch Vision Transformer (ViT) [13] as its backbone. It consists of a shared encoder $E_{img}$ , two separate decoders $D_{key}$ and $D_{sup}$ , and a point regression head for final prediction. These components are detailed below. + +Image encoder. For a given video clip, the image encoder $E_{img}$ encodes each frame $I_{i}$ to obtain token representations $F_{i} \in \mathbb{R}^{T \times d}$ , where $T$ is the number of tokens and $d$ is the token dimension. The encoder $E_{img}$ comprises $m$ ViT encoder blocks, each containing self-attention and feed-forward layers. The encoding process is denoted as + +$$ +F _ {i} ^ {(T \times d)} = E _ {i m g} (I _ {i} ^ {(H \times W \times 3)}), i = 1, \dots , L. +$$ + +The frames are processed independently and in parallel, with the output divided into two parts: $F_{key}$ for the keyframe and $\{F_{sup_i}\}_{i=1}^{L-1}$ for the supporting frames. + +Keyframe decoder. The keyframe decoder $D_{key}$ consists of $n$ ViT decoder blocks, each containing self-attention, cross-attention, and feed-forward layers. Unlike DUSt3R which uses the standard cross-attention, we introduce a novel multi-view cross-attention to combine information from different supporting frames. Given the feature tokens $F_{key}$ and $\{F_{sup_i}\}_{i=1}^{L-1}$ , the keyframe decoder $D_{key}$ takes $F_{key}$ as input for self-attention and performs cross-attention between $F_{key}$ and $\{F_{sup_i}\}_{i=1}^{L-1}$ . A decoder block is illustrated in Figure 3. For each cross-attention layer, queries are taken from $F_{key}$ , while keys and values are extracted from the supporting tokens $F_{sup_i}$ . These $L-1$ cross-attention layers are independent of each other, allowing for parallel processing. A max-pooling layer is then employed to aggregate features after cross-attention. We obtain decoded + +![](images/1cf1dd152ac4f10e494115256d7b0d5b3cd4d6e1d4851006aa050e406e416a35.jpg) +Figure 2. System overview. Given an input monocular RGB video, we apply a sliding window mechanism to convert it into overlapping clips (referred to as windows). Each window is fed into an Image-to-Points (I2P) network to recover 3D points in a local coordinate system. Next, the local points are incrementally fed into a Local-to-World (L2W) network to create a globally consistent scene model. The proposed I2P and L2W networks elegantly share similar architectures. In the I2P step (Sec. 3.1), we select a keyframe as a reference to set up a local coordinate system and use the remaining frames in the window to estimate the 3D geometry captured within it. The points from the first window are used to establish the world coordinate system. We then incrementally fuse the following windows in the L2W step (Sec. 3.2). This process involves retrieving the most relevant already-registered keyframes as a reference, and integrating new keyframes. Through this iterative process, we eventually obtain the full scene reconstruction. + +keyframe tokens $G_{key}$ as: + +$$ +G _ {k e y} = D _ {k e y} (F _ {k e y}, F _ {s u p _ {1}}, \dots , F _ {s u p _ {L - 1}}). +$$ + +![](images/9328c85455681f3182ac79be86dcd56cc43692a32c381184ebc4032850bcf15b.jpg) +Figure 3. Illustration of a decoder block in the proposed keyframe decoder $D_{key}$ . We present a minimalist modification to integrate information from different supporting images. Our approach traverses each of them, selects its token keys and values, and uses the keyframe queries to interact with them separately across the supporting images. This multi-view information is then aggregated through max-pooling. The registration decoder $D_{reg}$ and scene decoder $D_{sce}$ (described in Sec. 3.2) share the same architecture. + +Supporting decoder. The supporting decoder $D_{sup}$ is designed to complement the keyframe decoder. It inherits the decoder architecture used in DUSt3R, consisting of $n$ standard ViT decoder blocks. The cross-attention mechanism is applied only to exchange information with the keyframe. + +Note that all supporting frames share the same $D_{sup}$ . This process is denoted as + +$$ +G _ {s u p _ {i}} = D _ {s u p} (F _ {s u p _ {i}}, F _ {k e y}), i = 1, \dots , L - 1. +$$ + +Points reconstruction. Similar to DUSt3R, we apply a linear head [64] to regress dense 3D pointmaps in the unified coordinate system from decoded tokens. In addition to the pointmaps, we also predict the confidence maps for all frames to evaluate their reliability. The final predictions are: + +$$ +\hat {X} _ {i} ^ {(H \times W \times 3)}, \hat {C} _ {i} ^ {(H \times W \times 1)} = \mathrm {H} (G _ {i} ^ {(T \times d)}), i = 1, \dots , L. +$$ + +Training loss. Following DUSt3R, the I2P network is trained end-to-end using ground-truth scene points $\{X_{i}\}_{i = 1}^{L}$ . Both the ground truth and predicted point maps are normalized to a canonical scale, determined by the average distance of all valid points within the window to the origin. The confidence-aware training loss is: + +$$ +\mathcal {L} _ {I 2 P} = \sum_ {i = 1} ^ {L} M _ {i} \cdot (\hat {C} _ {i} \cdot \mathrm {L} 1 (\frac {1}{\hat {z}} \hat {X} _ {i}, \frac {1}{z} X _ {i}) - \alpha \log \hat {C} _ {i}), +$$ + +where $M_{i}$ represents a mask of valid points that have ground-truth values in $X_{i}$ , $z$ and $\hat{z}$ are the scale factor, $\hat{C}_i$ is the confidence map, the operator $\cdot$ denotes the elementwise matrix multiplication, $\mathrm{L1}(\cdot)$ denotes the point-wise Euclidean distance, and $\alpha$ is a hyper-parameter to control the regularization term. We will detail the process in Sec 4. + +# 3.2. Inter-Window Global Registration + +After obtaining the 3D pointmap $\{\hat{X}_{key}\}$ from the I2P network, we use the inter-window Local-to-World (L2W) + +model to incrementally register the newly generated pointmap into a global 3D coordinate system. Similar to the I2P network, the L2W model also relies on some frames to serve as a reference for the scene coordinate system. Furthermore, it can leverage multiple registered keyframes as a global reference. These registered keyframes are referred to as scene frames and they are maintained in a buffering set through a sampling mechanism. + +The buffering set is designed for scalability in handling long videos. We apply a reservoir strategy [59] that maintains a maximum of $B$ registered frames in the buffering set. When a new keyframe is inferred from I2P and ready for fusion, we retrieve the top- $K$ best-correlated scene frames from the buffering set as its support for global registration. + +Scene initialization. The first window is used to initialize the scene model. It's crucial to ensure this initialization is as accurate as possible. To achieve this, we execute the I2P network $L$ times, attempting to traverse and designate each frame within the window as the keyframe. We then select the result with the highest total confidence score for scene model initialization. This process results in a scene pointcloud along with a set of registered frames. All these frames are regarded as scene frames, and they are used to initialize the buffering set. + +Reservoir and retrieval. Each scene frame $I_{sce_i}$ is recorded with its latent feature $F_{sce_i}$ and pointmaps $\tilde{X}_{sce_i}$ . For efficiency, we apply reservoir sampling to allow storing an unbiased subset of an empirical distribution in a bounded amount of memory. The first $B$ registered frames chosen are directly inserted into the buffering set. For each subsequent frame with $id > B$ , the probability of inserting it is $B / id$ . If chosen for insertion, it will randomly replace one of the current scene frames in the buffering set. + +Given a new keyframe $I_{key}$ to be registered, we feed its feature $F_{key}$ and the features from the buffering set into a retrieval module, + +$$ +\operatorname {R e t r i e v a l} \left(F _ {\text {k e y}} ^ {(T \times d)}, \left\{F _ {\text {s c e} _ {i}} ^ {(T \times d)} \right\}\right), +$$ + +to obtain a list of correlation scores, measuring both the visual similarity and baseline suitability between the keyframe and the scene frames in the buffering set. The retrieval module uses the first $r$ decoder blocks from the I2P module as its backbone. A linear projection and an average-pooling layer follow, together producing an image-wise correlation score. We then select the top- $K$ scene frames as a global reference to fuse the current keyframe. As a result, we have $K$ scene frames and one keyframe as the input for the following L2W model. + +Points embedding. The 3D pointmaps reconstructed by the I2P model are encoded into the L2W model using a patch embedding method similar to image patchification in the ViT encoder $E_{img}$ . We process the new keyframe and $K$ + +retrieved scene frames in parallel as: + +$$ +\mathcal {P} _ {i} ^ {(T \times d)} = E _ {p t s} (\hat {X} _ {i} ^ {(H \times W \times 3)}), i = 1, \dots , K + 1. +$$ + +The encoded geometric tokens are combined with their corresponding visual tokens by + +$$ +\mathcal {F} _ {i} ^ {(T \times d)} = F _ {i} ^ {(T \times d)} + \mathcal {P} _ {i} ^ {(T \times d)}, i = 1, \dots , K + 1. +$$ + +This resulting a token set $\{\mathcal{F}_{key},\{\mathcal{F}_{sce_i}\}_{i = 1}^K\}$ contains joint features of image patch appearance and 3D geometry for the keyframe and retrieved scene frames. In the following decoders, $\{\mathcal{P}\}$ are further accumulated to $\{\mathcal{F}\}$ between adjacent blocks to enhance the geometric representation. + +Registration decoder. The registration decoder $D_{reg}$ takes feature tokens $\{\mathcal{F}_{key},\{\mathcal{F}_{sce_i}\}_{i = 1}^K\}$ as input and aims to transform the local reconstruction of the keyframe to the scene coordinate system. It takes the same network architecture of the keyframe decoder $D_{key}$ . This decoding process is denoted by + +$$ +\mathcal {G} _ {k e y} = D _ {r e g} \left(\mathcal {F} _ {k e y}, \mathcal {F} _ {s c e _ {1}}, \dots , \mathcal {F} _ {s c e _ {K}}\right). +$$ + +Scene decoder. The scene decoder $D_{sce}$ takes the token set $\{\mathcal{F}_{key},\{\mathcal{F}_{sce_i}\}_{i = 1}^K\}$ as input to refine the scene geometry without coordinate system changes. It uses the same network architecture as the keyframe decoder $D_{key}$ , allowing us to extend to multi-keyframe co-registration (see supplementary material for details). By default, we register one keyframe each time. Each of the $\mathcal{F}_{sce_i}$ has information exchange only with the $\mathcal{F}_{key}$ . This decoding process is denoted by + +$$ +\mathcal {G} _ {s c e _ {i}} = D _ {s c e} \left(\mathcal {F} _ {s c e _ {i}}, \mathcal {F} _ {k e y}\right), i = 1, \dots , K. +$$ + +Points reconstruction and training loss. We apply the same head design as that of the I2P network to predict all the pointmaps $\tilde{X}_i$ in the global scene coordinate system: + +$$ +\tilde {X} _ {i} ^ {(H \times W \times 3)}, \tilde {C} _ {i} ^ {(H \times W \times 1)} = \mathrm {H} (\mathcal {G} _ {i} ^ {(T \times d)}), i = 1, \dots , K + 1. +$$ + +We train the L2W network using a similar loss function as the I2P network. Differently, no normalization is applied to the predicted point map, as the output scale must align with the scene frames in the input. This alignment ensures that the output can be directly integrated into the existing reconstruction. The training loss of the L2W network is: + +$$ +\mathcal {L} _ {L 2 W} = \sum_ {i = 1} ^ {L} M _ {i} \cdot \left(\tilde {C} _ {i} \cdot \mathrm {L} 1 \left(\tilde {X} _ {i}, X _ {i}\right) - \alpha \log \tilde {C} _ {i}\right). +$$ + +The following section provides a detailed discussion of the training process and its implementation details. + +
MethodChess Acc. / Comp.Fire Acc. / Comp.Heads Acc. / Comp.Office Acc. / Comp.Pumpkin Acc. / Comp.RedKitchen Acc. / Comp.Stairs Acc. / Comp.Average Acc. / Comp.FPS
DUS3R [64]2.26 / 2.131.04 / 1.501.66 / 0.984.62 / 4.741.73 / 2.431.95 / 2.363.37 / 10.752.19 / 3.24<1
MAS3R [28]2.08 / 2.121.54 / 1.431.06 / 1.043.23 / 3.195.68 / 3.073.50 / 3.372.36 / 13.163.04 / 3.90≪1
Spann3R [61]2.23 / 1.680.88 / 0.922.67 / 0.985.86 / 3.542.25 / 1.852.68 / 1.805.65 / 5.153.42 / 2.41>50
SLAM3R-NoConf (Ours)2.12 / 1.210.95 / 0.803.23 / 1.672.59 / 2.211.99 / 2.042.09 / 1.884.54 / 6.382.40 / 2.24~25
SLAM3R (Ours)1.63 / 1.310.84 / 0.832.95 / 1.222.32 / 2.261.81 / 2.051.84 / 1.944.19 / 6.912.13 / 2.34~25
+ +Table 1. Reconstruction results on 7 Scenes [51] dataset. The average numbers are computed over all test sequences. The methods are categorized into two groups based on whether their FPS is above or below 1. The best results within each category are shown in bold. We report accuracy and completeness in centimeters. The color gradient shifts from red through yellow to green to show increasing FPS. + +
MethodRoom 0 Acc. / Comp.Room 1 Acc. / Comp.Room 2 Acc. / Comp.Office 0 Acc. / Comp.Office 1 Acc. / Comp.Office 2 Acc. / Comp.Office 3 Acc. / Comp.Office 4 Acc. / Comp.Average Acc. / Comp.FPS
DUSt3R [64]3.47 / 2.502.53 / 1.862.95 / 1.764.92 / 3.513.09 / 2.214.01 / 3.103.27 / 2.253.66 / 2.613.49 / 2.48<1
MASt3R [28]4.01 / 4.103.61 / 3.253.13 / 2.152.57 / 1.6312.85 / 8.133.13 / 1.994.67 / 3.153.69 / 2.474.71 / 3.36≤1
NICER-SLAM [79]*2.53 / 3.043.93 / 4.103.40 / 3.425.49 / 6.093.45 / 4.424.02 / 4.293.34 / 4.033.03 / 3.873.65 / 4.16≤1
DROID-SLAM [56]*12.18 / 8.968.35 / 6.073.26 / 16.013.01 / 16.192.39 / 16.205.66 / 15.564.49 / 9.734.65 / 9.635.50 / 12.29~20
DIM-SLAM [29]*14.19 / 6.249.56 / 6.458.41 / 12.1710.16 / 5.957.86 / 8.3316.50 / 8.2813.01 / 6.7713.08 / 8.6211.60 / 7.85~3
GO-SLAM [75]--------3.81 / 4.79~8
Spann3R [61]9.75 / 12.9415.51 / 12.947.28 / 8.505.46 / 18.755.24 / 16.649.33 / 11.8016.00 / 9.0313.97 / 16.0210.32 / 13.33>50
SLAM3R-NoConf (Ours)3.37 / 2.403.22 / 2.333.15 / 2.004.43 / 2.593.18 / 2.343.95 / 2.784.20 / 3.154.57 / 3.383.76 / 2.62~24
SLAM3R (Ours)3.19 / 2.403.12 / 2.342.72 / 2.004.28 / 2.603.17 / 2.343.84 / 2.783.90 / 3.164.32 / 3.363.57 / 2.62~24
+ +Table 2. Reconstruction results on Replica [54] dataset. * denotes the results reported in NICER-SLAM. + +
DUSt3R [64]MASt3R [28]NICER-SLAM [79]*DROID-SLAM [56]*DIM-SLAM [29]GO-SLAM [75]Spann3R [61]SLAM3R-NoConf (Ours)SLAM3R (Ours)
7 Scenes8.026.288.555.66--11.708.448.41
Replica4.761.671.880.330.460.3932.796.616.61
+ +Table 3. Camera pose results evaluated by ATE-RMSE (cm) on 7 Scenes [51] and Replica [54] datasets. + +![](images/44345ea07f6baf0eb2ac594b6ac8db7b6c7d52b1a43792d0a855fbf0d20ce29a.jpg) + +![](images/4d20d4f1c45c1c75db0ac26467fa3204ec21b1a1b942d11bcba92fecb577fcd3.jpg) +Spann3R +Figure 4. We visualize the reconstruction results on two scenes: Office-09 and Office 2 from the 7-Scenes [51] and Replica [54] datasets. Our method runs in real-time and achieves high-quality reconstruction comparable to the offline method DUSt3R [64]. + +![](images/0af3d4918d0308fc33eea8a0a5322a26ce8641a9335ec7451d5a7732d39b607b.jpg) + +![](images/309d4c9fb0ceeec648887a6c95d84d1f1c9ef8404ce2af3f9b1b1cad7a5dade8.jpg) +DUSt3R + +![](images/200ae69a4d04f2b4fc8008b95a78ccbf15cded675d5513c4eb3fc283b1f31a58.jpg) + +![](images/5453a8c6eea0f90eef1faa2e3e7c4def1e186093507c7db842f3b8018e9598c1.jpg) +SLAM3R (Ours) + +![](images/d5b41e2e7e51997716d37255fafedb4a2504a3de2d032030de7833a298909481.jpg) + +![](images/beb59d5df17aed1ca7a0573c6a38e9e65349dc67a5ac2f453bf0d63b8a748181.jpg) +Ground-truth + +# 4. Experiments + +Datasets. For both the Image-to-Points (I2P) and Local-to-World (L2W) models, we perform training with a mixture of three datasets: ScanNet++ [71], Aria Synthetic Environments [3] and CO3D-v2 [41]. These datasets vary from scene-level to object-centric, and contain both real-world and synthetic scenes. Since they are all recorded sequentially, we can easily extract video clips with a sliding-window mechanism as our training data. We select about + +850K clips for training in total. To validate our reconstruction quality, we conduct quantitative evaluations on two unseen datasets: 7 Scenes [51], a real-world dataset of partial scenes, and Replica [54], a synthetic dataset of complete scenes. We also demonstrate visual reconstruction results across diverse datasets and in-the-wild captured videos to showcase the generalization ability of SLAM3R. + +Implementation details. Both the I2P and L2W models build upon the architecture of DUSt3R [64] with minimal but effective modifications, making it natural for them to + +![](images/1c0ed92335b18e80c89f9545ce3b3994448a23ce6aff48003530a111283e3731.jpg) +Figure 5. Qualitative examples. We show our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild captured videos, to demonstrate SLAM3R's generalization ability. + +initialize their weights from the DUSt3R pre-trained model. We initialize our weights using the DUSt3R model trained on $224 \times 224$ resolution images with $m = 24$ encoder blocks, $n = 12$ decoder blocks with linear heads. All images are center-cropped to $224 \times 224$ pixels before feeding into SLAM3R. Our training is conducted on 8 NVIDIA 4090D GPUs, each with 24 GB of memory. It takes about one day. At test time, we set the initial window length to $L = 5$ to ensure high-quality reconstruction of all frames within the window. For subsequent incremental windows, we use $L = 11$ to provide more supporting views for better keyframe reconstruction. Please refer to our supplementary material for more implementation details. + +# 4.1. Comparisons + +Evaluation metrics. Following NICER-SLAM [79] and Spann3R [61], we build a ground-truth point cloud model for each test sequence by back-projecting pixels to the world using ground-truth depths and camera parameters. The reconstructed point clouds are aligned to ground truths using Umayama [58] and ICP [43] algorithms. We quantify reconstruction quality through accuracy and completeness metrics. To demonstrate computational efficiency, we report frames per second (FPS) on a single NVIDIA 4090D GPU. We also evaluate camera poses using absolute trajectory error (ATE-RMSE). For detailed formulations of the metrics, please refer to the supplementary material. + +Reconstruction results on the 7 Scenes [51] dataset. The numerical results of scene reconstruction quality are re + +ported in Table 1. Following Spann3R [61]'s setting, we uniformly sample one-twentieth of the frames in each test sequence as input video. Each video is regarded as an individual scene. We evaluate SLAM3R using two settings: integrating the full pointmaps predicted for all input frames to create reconstruction results (denoted by SLAM3R-NoConf), and filtering pointmaps with a confidence threshold of 3 before creating reconstruction results (SLAM3R). We compare our method with optimization-based reconstruction DUSt3R [64], triangulation-based MASt3R [28], and online incremental reconstruction Spann3R. Our method outperforms all baselines in both accuracy and completeness while maintaining real-time performance. As shown in the Office-09 scene (the top row in Figure 4), our approach demonstrates much less drift compared to the concurrent work Spann3R [61]. + +Reconstruction results on the Replica [54] dataset. Besides the baselines mentioned in 7 Scene datasets, we also compare the SLAM-based reconstruction approaches NICER-SLAM [79], DROID-SLAM [56], DIM-SLAM [29] and GO-SLAM [75] on the Replica [54] dataset. The numerical results on full scene reconstruction are reported in Table 2. Due to the memory constraint, DUSt3R [64] and MASt3R [28] process only one-twentieth of the frames for reconstruction. As is shown in the table, our method surpasses all baselines with FPS greater than 1. Notably, without any optimization procedure, our method achieves reconstruction quality comparable to optimization-based methods such as NICER- + +
Method# FramesAcc.Comp.FPS
DUSt3R [64]23.162.8942.55
I2P23.393.0442.55
I2P52.622.2840.82
I2P (Default)112.382.0340.11
I2P152.271.9435.51
I2P512.231.8611.97
+ +SLAM [79] and DUS3R [64]. Example of the Office 2 (the bottom row in Figure 4) also illustrates the global consistency of our reconstruction result. + +Camera pose estimation on 7 Scenes [51] and Replica [54]. Our method is designed in a new paradigm that reconstructs 3D points end-to-end without explicitly solving camera parameters. Following DUSt3R [64], We also derive camera poses from the predicted scene points using PnP-RANSAC solver in OpenCV [6] with ground truth camera intrinsics of each frame. The results are reported in Table 3. We can observe that camera poses and scene reconstruction results are not fully positively correlated. This discrepancy between pose and reconstruction errors indicates that effective end-to-end 3D reconstruction is possible and promising without first obtaining precise camera poses. + +For more details on the comparisons in this section, please refer to our supplementary material. + +# 4.2. Analyses + +Effectiveness of the I2P model. To highlight the advantages of our multi-view I2P model over the original two-view DUSt3R [64], we evaluate the reconstruction quality of keyframes with varying numbers of supporting views. We conduct experiments on the Replica [54] dataset, where input views are sampled using a sliding window of different sizes, and the reconstruction accuracy and completeness of the keyframes are computed. The results are reported in Table 4. As the number of supporting views increases, our approach progressively improves reconstruction quality. Notably, the efficiency of our method remains stable until the window size exceeds 11, demonstrating the effectiveness of our parallel design. However, the results also show diminishing returns as the number of views increases, which we detail in the supplementary material. Visual results of I2P reconstruction can be found in Figure 1. + +Advantages of the L2W model. The effectiveness of the L2W model is evaluated through ablation studies on the Replica [54] dataset. Per-window reconstructions are first generated with a window size of 11 using the I2P model. Local points are then aligned to a unified coordinate frame using different methods: global optimization + +Table 4. Inner-window keyframe reconstruction results with various window lengths. By default, we use 11-frame windows for incremental reconstruction to balance quality and efficiency. + +
MethodAcc.Comp.FPS
I2P+GA4.873.00~3
I2P+UI7.473.86~1
I2P+L2W6.193.54~92
I2P+L2W+Re (Full)3.622.70~43
+ +Table 5. Reconstruction results using various point alignment methods and scene frame selection strategies. The FPS reported only accounts for the overhead of the alignment operation. + +from DUSt3R [64] (I2P-GA), traditional approaches such as Umeyama [58] and ICP [43] (I2P+UI), and our L2W model (I2P+L2W+Re). For consistency, we set the window size for global optimization to 10, which is equal to the number of views used to align new frames in other methods. Results in Table 5 show that our full method achieves superior alignment accuracy and computational efficiency compared to the alternatives. + +Analysis of the retrieval module. We propose a lightweight retrieval module that selects historical scene frames from the reservoir. This approach effectively performs implicit re-localization. We compare our retrieval method with a baseline approach that selects the ten nearest previous frames, named I2P+L2W. The results in Table 5 indicate a significant performance improvement with our retrieval strategy, demonstrating its effectiveness. + +In-the-wild scene reconstruction. We have tested our method on a diverse range of unseen datasets and found that SLAM3R shows strong generalization capabilities. Figure 5 shows our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild videos we captured. These results show that our method performs reliably across different scales and scenes. We also provide additional numerical results on sampled scenes from these datasets in the supplementary material. + +# 5. Conclusion + +In this paper, we present SLAM3R, a novel and effective system that performs real-time, high-quality, dense 3D scene reconstruction using RGB videos. It employs a two-hierarchy neural network framework to perform end-to-end 3D reconstruction through streamlined feed-forward processes, eliminating the need to explicitly solve any camera parameters. Experiments demonstrate its state-of-the-art reconstruction quality and real-time efficiency (20+ FPS). + +Limitations and future work. The elimination of camera parameter prediction prevents us from performing global bundle adjustment. Additionally, the poses derived from our scene point cloud prediction still fall short of SLAM systems that specialize in camera localization. Addressing these limitations will be the focus of our future work. + +# 6. Acknowledgments + +This work is supported by the National Key R&D Program of China (2022ZD0160800). This work is also supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and in part by the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We thank the reviewers for their valuable feedback, as well as Zihan Zhu and Songyou Peng for their help with our experiments. Siyan Dong would also like to thank the support from HKU School of Computing and Data Science. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2 +[2] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 7, 8 +[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescipt: Reconstructing scenes with an autoregressive structured language model. arXiv preprint arXiv:2403.13064, 2024. 6 +[4] Tim Bailey and Hugh Durrant-Whyte. Simultaneous localization and mapping (slam): Part ii. IEEE robotics & automation magazine, 13(3):108-117, 2006. 2 +[5] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 1, 2 +[6] G Bradski. The opencv library. Dr. Dobb's Journal of Software Tools, 2000. 8 +[7] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2 +[8] Yiyang Chen, Siyan Dong, Xulong Wang, Lulu Cai, Youyi Zheng, and Yanchao Yang. Sg-nerf: Neural surface reconstruction with scene graph optimization. arXiv preprint arXiv:2407.12667, 2024. 2 +[9] Chi-Ming Chung, Yang-Che Tseng, Ya-Ching Hsu, Xiang-Qian Shi, Yun-Hung Hua, Jia-Fong Yeh, Wen-Chin Chen, Yi-Ting Chen, and Winston H Hsu. Orbeez-slam: A real-time monocular visual slam with orb features and nefrealized mapping. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9400-9406. IEEE, 2023. 1, 2 + +[10] Jan Czarnowski, Tristan Laidlow, Ronald Clark, and Andrew J Davison. Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters, 5 (2):721-728, 2020. 1 +[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2 +[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[14] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. arXiv preprint arXiv:2409.19152, 2024. 3 +[15] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2 +[16] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 1, 2 +[17] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 1, 2 +[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2 +[19] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2 +[20] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[21] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024. 2 +[22] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 1, 2 +[23] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1 +[25] Georg Klein and David Murray. Parallel tracking and mapping on a camera phone. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 83-86. IEEE, 2009. 1, 2 +[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7, 8 +[27] Lukas Koestler, Nan Yang, Niclas Zeller, and Daniel Cremers. Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In Conference on Robot Learning, pages 34-45. PMLR, 2022. 2 +[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 6, 7 +[29] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations, 2023. 1, 2, 6, 7 +[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2 +[31] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 1, 2 +[32] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2 +[33] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 1, 2 +[34] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2 +[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 + +[37] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 1, 2 +[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 1, 2 +[39] Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision, pages 2320-2327. IEEE, 2011. 1, 2 +[40] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1 +[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6 +[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3437-3444. IEEE, 2023. 1, 2 +[43] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 7, 8 +[44] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2 +[45] Erik Sandström, Kevin Ta, Luc Van Gool, and Martin R Oswald. Uncle-slam: Uncertainty learning for dense neural slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4537-4548, 2023. 2 +[46] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians. arXiv preprint arXiv:2405.16544, 2024. 1, 2 +[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 2 +[48] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1, 2 +[49] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high- + +resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 7, 8 +[50] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 7, 8 +[51] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 6, 7, 8 +[52] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2 +[53] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 1, 2 +[54] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7, 8 +[55] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6229-6238, 2021. 1, 2 +[56] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6, 7 +[57] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R Oswald, and Matteo Poggi. How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255, 4, 2024. 1 +[58] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7, 8 +[59] Jeffrey S Vitter. Random sampling with a reservoir. ACM Transactions on Mathematical Software (TOMS), 11(1):37-57, 1985. 5 +[60] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8606-8615, 2022. 1, 2 +[61] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 3, 6, 7 + +[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2 +[63] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2 +[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7, 8 +[65] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2 +[66] Changchang Wu. Visualsfm: A visual structure from motion system. http://www.cs.washington.edu/homes/ccwu/vsfm, 2011. 1, 2 +[67] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2 +[68] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2 +[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 7, 8 +[70] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2 +[71] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 6 +[72] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2 +[73] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 1, 2 +[74] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2 + +[75] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 1, 2, 6, 7 +[76] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 1, 2 +[77] Heng Zhou, Zhetao Guo, Shuhong Liu, Lechen Zhang, Qihao Wang, Yuxiang Ren, and Mingrui Li. Mod-slam: Monocular dense mapping for unbounded 3d scene reconstruction. arXiv preprint arXiv:2402.03762, 2024. 1, 2 +[78] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 1, 2 +[79] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 1, 2, 6, 7, 8 \ No newline at end of file diff --git a/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/images.zip b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af7ba93d88979bfa4ea4da7e687788e2b433be6e --- /dev/null +++ b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e68ccdfb36a021290bad3804ebf09eadac0c24a726da8f2bde3eceafc7c2a5b7 +size 676176 diff --git a/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/layout.json b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c86b7d90834bef875270cae4b8e539ef8590aa29 --- /dev/null +++ b/2025/SLAM3R_ Real-Time Dense Scene Reconstruction from Monocular RGB Videos/layout.json @@ -0,0 +1,10333 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 57, + 103, + 552, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 103, + 552, + 120 + ], + "spans": [ + { + "bbox": [ + 57, + 103, + 552, + 120 + ], + "type": "text", + "content": "SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "spans": [ + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "text", + "content": "Yuzheng Liu" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "text", + "content": " Siyan Dong" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "text", + "content": " Shuzhe Wang" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "text", + "content": " Yingda Yin" + }, + { + "bbox": [ + 150, + 142, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "spans": [ + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "text", + "content": "Yanchao Yang" + }, + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "text", + "content": " Qingnan Fan" + }, + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "text", + "content": " Baoquan Chen" + }, + { + "bbox": [ + 181, + 157, + 433, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "text", + "content": "Peking University " + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "text", + "content": "Aalto University " + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 102, + 171, + 506, + 186 + ], + "type": "text", + "content": "VIVO" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 238, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 296, + 441 + ], + "type": "text", + "content": "In this paper, we introduce SLAM3R, a novel and effective system for real-time, high-quality, dense 3D reconstruction using RGB videos. SLAM3R provides an end-to-end solution by seamlessly integrating local 3D reconstruction and global coordinate registration through feedforward neural networks. Given an input video, the system first converts it into overlapping clips using a sliding window mechanism. Unlike traditional pose optimization-based methods, SLAM3R directly regresses 3D pointmaps from RGB images in each window and progressively aligns and deforms these local pointmaps to create a globally consistent scene reconstruction - all without explicitly solving any camera parameters. Experiments across datasets consistently show that SLAM3R achieves state-of-the-art reconstruction accuracy and completeness while maintaining real-time performance at " + }, + { + "bbox": [ + 53, + 238, + 296, + 441 + ], + "type": "inline_equation", + "content": "20+" + }, + { + "bbox": [ + 53, + 238, + 296, + 441 + ], + "type": "text", + "content": " FPS. Code available at: https://github.com/PRU-VCL-3DV/SLAM3R." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 483, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 626 + ], + "type": "text", + "content": "Dense 3D reconstruction, a long-standing challenge in computer vision, aims to capture and reconstruct the detailed geometry of real-world scenes. Traditional approaches have largely relied on multi-stage pipelines. These typically begin with sparse Simultaneous Localization and Mapping (SLAM) [7, 16, 25, 37, 38] or Structure-from-Motion (SfM) [31, 33, 47, 53, 66] algorithms to estimate camera parameters, followed by Multi-View Stereo (MVS) [18, 48, 60, 68] techniques to fill in scene details. While these methods offer high-quality reconstructions, they often require offline processing to produce a complete model, which limits their applicability in real-world scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 627, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 295, + 687 + ], + "type": "text", + "content": "In the literature, dense SLAM approaches [5, 9, 10, 17, 22, 39, 42, 55, 56, 76, 78] have been developed to address dense scene reconstruction as a complete system. However, these approaches often fall short in terms of reconstruction accuracy or completeness, or they rely heavily on" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 230, + 354, + 371 + ], + "blocks": [ + { + "bbox": [ + 319, + 214, + 354, + 230 + ], + "lines": [ + { + "bbox": [ + 319, + 214, + 354, + 230 + ], + "spans": [ + { + "bbox": [ + 319, + 214, + 354, + 230 + ], + "type": "text", + "content": "Monocular RGB video" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 230, + 354, + 371 + ], + "lines": [ + { + "bbox": [ + 319, + 230, + 354, + 371 + ], + "spans": [ + { + "bbox": [ + 319, + 230, + 354, + 371 + ], + "type": "image", + "image_path": "0db8d85bac62fd1d9a1399553068be176822bf31718e87fa364a43f5043dae38.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 373, + 339, + 382 + ], + "lines": [ + { + "bbox": [ + 334, + 373, + 339, + 382 + ], + "spans": [ + { + "bbox": [ + 334, + 373, + 339, + 382 + ], + "type": "text", + "content": "：" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 268, + 371, + 309 + ], + "lines": [ + { + "bbox": [ + 362, + 268, + 371, + 309 + ], + "spans": [ + { + "bbox": [ + 362, + 268, + 371, + 309 + ], + "type": "text", + "content": "rnnnne" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 384, + 354, + 418 + ], + "blocks": [ + { + "bbox": [ + 319, + 384, + 354, + 418 + ], + "lines": [ + { + "bbox": [ + 319, + 384, + 354, + 418 + ], + "spans": [ + { + "bbox": [ + 319, + 384, + 354, + 418 + ], + "type": "image", + "image_path": "cdf5f9486d9d2ae216eea8cde4677a19cd7b2b0fa006afa190e2f61c71fa8007.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 362, + 215, + 420, + 257 + ], + "blocks": [ + { + "bbox": [ + 362, + 215, + 420, + 257 + ], + "lines": [ + { + "bbox": [ + 362, + 215, + 420, + 257 + ], + "spans": [ + { + "bbox": [ + 362, + 215, + 420, + 257 + ], + "type": "image", + "image_path": "5b1725e7c08c93b6714b35e83d89801688addcb8e6bf3235a30dcb048dc049df.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 427, + 214, + 471, + 255 + ], + "blocks": [ + { + "bbox": [ + 427, + 214, + 471, + 255 + ], + "lines": [ + { + "bbox": [ + 427, + 214, + 471, + 255 + ], + "spans": [ + { + "bbox": [ + 427, + 214, + 471, + 255 + ], + "type": "image", + "image_path": "6f9cac0419478a57d85aa30cf7d2a15ad0b7c97ae3bae5c0ebd692514623326b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 492, + 214, + 537, + 256 + ], + "blocks": [ + { + "bbox": [ + 477, + 233, + 487, + 239 + ], + "lines": [ + { + "bbox": [ + 477, + 233, + 487, + 239 + ], + "spans": [ + { + "bbox": [ + 477, + 233, + 487, + 239 + ], + "type": "text", + "content": "···" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 214, + 537, + 256 + ], + "lines": [ + { + "bbox": [ + 492, + 214, + 537, + 256 + ], + "spans": [ + { + "bbox": [ + 492, + 214, + 537, + 256 + ], + "type": "image", + "image_path": "2e0ae777d0d36307b7da4c032205aec8aff4db120412f6902a211f13b2663123.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 542, + 234, + 550, + 239 + ], + "lines": [ + { + "bbox": [ + 542, + 234, + 550, + 239 + ], + "spans": [ + { + "bbox": [ + 542, + 234, + 550, + 239 + ], + "type": "text", + "content": "···" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 435, + 261, + 475, + 316 + ], + "blocks": [ + { + "bbox": [ + 435, + 261, + 475, + 316 + ], + "lines": [ + { + "bbox": [ + 435, + 261, + 475, + 316 + ], + "spans": [ + { + "bbox": [ + 435, + 261, + 475, + 316 + ], + "type": "image", + "image_path": "c7d22570d7d62c3fde62ae775babab89d99b1e606a4591121ef032ec7978cc21.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 482, + 262, + 547, + 317 + ], + "blocks": [ + { + "bbox": [ + 482, + 262, + 547, + 317 + ], + "lines": [ + { + "bbox": [ + 482, + 262, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 482, + 262, + 547, + 317 + ], + "type": "image", + "image_path": "198e72296ee59f9ada4d8ae8f1bfb953c3a1d6566dc895b63c5cce010087da21.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 361, + 319, + 474, + 421 + ], + "blocks": [ + { + "bbox": [ + 361, + 319, + 474, + 421 + ], + "lines": [ + { + "bbox": [ + 361, + 319, + 474, + 421 + ], + "spans": [ + { + "bbox": [ + 361, + 319, + 474, + 421 + ], + "type": "image", + "image_path": "ab733a38f11d22d11d4e114711cc2c8ab07f9145858444a9d3b4fb7987543ab3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 430, + 555, + 508 + ], + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 508 + ], + "type": "text", + "content": "Figure 1. We introduce a novel dense reconstruction system - SLAM3R. It takes a monocular RGB video as input and reconstructs the scene as a dense pointcloud. The video is converted into short clips for local reconstruction (denoted as inner-window), which are then incrementally registered together (inter-window) to create a global scene model. This process runs in real-time, producing a reconstruction that is both accurate and complete." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 477, + 319, + 548, + 369 + ], + "blocks": [ + { + "bbox": [ + 477, + 319, + 548, + 369 + ], + "lines": [ + { + "bbox": [ + 477, + 319, + 548, + 369 + ], + "spans": [ + { + "bbox": [ + 477, + 319, + 548, + 369 + ], + "type": "image", + "image_path": "5f43ce300897134d6b32ab3c7b7c425094180c6b9f0cce2d3f7cd66873c5cc07.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 476, + 370, + 548, + 420 + ], + "blocks": [ + { + "bbox": [ + 476, + 370, + 548, + 420 + ], + "lines": [ + { + "bbox": [ + 476, + 370, + 548, + 420 + ], + "spans": [ + { + "bbox": [ + 476, + 370, + 548, + 420 + ], + "type": "image", + "image_path": "d96334ff76ea5b7f73d0a489e23f810f23d0b8e13a738866f6e409a60f0508c2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 532, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 555, + 652 + ], + "type": "text", + "content": "depth sensors. Recently, several monocular SLAM systems [29, 46, 73, 75, 77, 79] have been proposed to tackle dense scene reconstruction from RGB videos. By incorporating advanced scene representations [24, 36, 40, 57, 62], these systems produce accurate and complete scene reconstructions. However, this comes at the cost of reduced running efficiency. For example, NICER-SLAM [79] operates at a speed significantly below 1 FPS. Therefore, current approaches struggle with at least one of three key criteria: reconstruction accuracy, completeness, or efficiency." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "While monocular dense SLAM systems encounter the limitations mentioned earlier, recent advances in two-view geometry have shown promising potential. DUSt3R [64] introduces a purely end-to-end approach for learning dense reconstruction. Trained on large-scale datasets, its network" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 283, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 283, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 283, + 703 + ], + "type": "text", + "content": "*Joint first authors: liu_yuzheng@stu.pku.edu.cn, siyan3d@hku.hk" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16651" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "is capable of producing high-quality dense reconstructions from paired images in real-time. However, for multiple views, a global optimization step is required to align these image pairs, which significantly hampers its efficiency. A concurrent work, Spann3R [61], extends DUSt3R to multiview (video) scenarios through an incremental pipeline with spatial memory. While this method accelerates the reconstruction process, it unfortunately results in noticeable accumulated drift and reduced reconstruction quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 298, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 298, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 298, + 445 + ], + "type": "text", + "content": "To address these challenges, we present SLAM3R (pronounced \"slæmər\"), a real-time dense 3D reconstruction system using RGB-only videos as input. Unlike traditional SLAM problems, SLAM3R performs implicit camera localization and focuses more on dense scene mapping, where 3R stands for 3D Reconstruction. SLAM3R comprises a two-hierarchy framework. First, it reconstructs local 3D geometry from a sliding window that processes short clips from the input video. Then, it progressively registers these local reconstructions to build a globally consistent 3D scene. Both modules are developed with simple yet effective feed-forward models, enabling end-to-end and efficient scene reconstruction. Specifically, the two modules are the Image-to-Points (I2P) network and the Local-to-World (L2W) network. The I2P module, inspired by DUSt3R, selects a keyframe in a local window as the coordinate system reference. It directly predicts the dense 3D point map supported by the remaining frames within that window. The L2W module incrementally fuses locally reconstructed points into a coherent global coordinate system. Both processes reconstruct the 3D points without explicitly estimating any camera parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 446, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 529 + ], + "type": "text", + "content": "Through extensive experiments, we demonstrate that SLAM3R provides high-quality scene reconstructions with minimal drift, outperforming existing dense SLAM systems across various benchmarks. Furthermore, SLAM3R achieves these results at " + }, + { + "bbox": [ + 55, + 446, + 295, + 529 + ], + "type": "inline_equation", + "content": "20+" + }, + { + "bbox": [ + 55, + 446, + 295, + 529 + ], + "type": "text", + "content": " FPS, bridging the gap between quality and efficiency in RGB-only dense scene reconstruction. Our contributions are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 582 + ], + "type": "text", + "content": "- We present a novel real-time end-to-end dense 3D reconstruction system that uses RGB videos to directly predict 3D pointmaps in a unified coordinate system through feed-forward neural networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 630 + ], + "type": "text", + "content": "- Through careful design, our Image-to-Points module can process an arbitrary number of images simultaneously, effectively extending DUSt3R to handle multiple views and produce higher-quality predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 631, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 631, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 631, + 295, + 677 + ], + "type": "text", + "content": "- The proposed Local-to-World module directly aligns predicted local 3D pointmaps into a unified global coordinate system. This eliminates the need for explicit camera parameter estimation and costly global optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 678, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 678, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 678, + 295, + 714 + ], + "type": "text", + "content": "- We evaluate our method on multiple public benchmarks. It achieves state-of-the-art reconstruction quality in terms of both accuracy and completeness at real-time speeds." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 270 + ], + "type": "text", + "content": "Traditional offline approaches. Dense 3D pointcloud reconstruction is a long-standing problem in computer vision. Classical approaches to this problem first determine camera parameters using Structure from Motion (SfM) [31, 33, 47, 53, 66], followed by dense 3D points triangulation with Multi-View Stereo (MVS) [1, 18, 48, 60, 68]. In recent years, neural implicit [8, 30, 36, 62, 65] and 3D Gaussian [12, 19, 20] representations have been applied to further enhance the quality of dense reconstruction. While these methods deliver high-quality results, they have a significant limitation: the requirement for offline processing to generate the final 3D model, which restricts their applicability in real-time scenarios. In this paper, we focus on online dense reconstruction in the context of Simultaneous Localization and Mapping (SLAM)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 498 + ], + "type": "text", + "content": "Dense SLAM. Early works on SLAM [4, 7, 15, 16, 25, 37, 38] focused on reconstructing the structure of unknown environments while simultaneously localizing camera poses. These approaches prioritize real-time performance but produce only sparse structures of the scene. Dense SLAM approaches [5, 9-11, 17, 22, 27, 39, 42, 55, 56, 76, 78] incorporate detailed scene geometry information to improve pose estimation. DROID-SLAM [56] introduces recurrent iterative updates of camera poses and pixel-wise depth estimates, while TANDEM [27] proposes an online MVS module for depth prediction. These systems enable real-time dense scene reconstruction. However, their focus on camera trajectory accuracy often results in incomplete and noisy 3D reconstruction. Neural implicit and Gaussian representations have also been integrated with dense SLAM systems [9, 21-23, 32, 34, 42, 42, 44, 45, 55, 67, 72, 78]. However, these approaches often rely on additional depth sensors or focus primarily on novel view synthesis rather than producing detailed geometric reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 640 + ], + "type": "text", + "content": "More recently, several monocular dense SLAM systems [29, 46, 73, 75, 77, 79] have been developed to produce dense scene geometry reconstruction. A notable limitation of these systems is their slow runtime. Among these systems, GO-SLAM [75] achieves a speed of " + }, + { + "bbox": [ + 313, + 498, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\sim 8" + }, + { + "bbox": [ + 313, + 498, + 553, + 640 + ], + "type": "text", + "content": " FPS, which still falls short of real-time capability. Furthermore, these methods all share a common strategy: they alternate between solving for camera poses and estimating the scene representation. In contrast, this paper presents a novel approach to dense scene reconstruction that eliminates the need for explicitly solving camera parameters, offering a more efficient and streamlined solution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "text", + "content": "End-to-end dense 3D reconstruction. DUSt3R [64] introduces the first purely end-to-end dense 3D reconstruction pipeline without relying on camera parameters. Recently, several works have adopted a similar approach for single-view reconstruction [63], feature matching [28], novel view synthesis [52, 70], and dynamics reconstruction [74]. These" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16652" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "successes demonstrate the effectiveness of end-to-end dense point prediction, inspiring us to develop a dense reconstruction system with a similar methodology." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 295, + 360 + ], + "type": "text", + "content": "While DUSt3R operates in real-time for two-view predictions, its extension to multiple views involves exhaustive pairing images and performing an additional global optimization step. This process significantly increases computational time, thereby hindering its real-time performance. MASt3R [28] enhances the matching capability of DUSt3R by adding a match head, achieving more accurate keypoint correspondences for 3D reconstruction [14], but at the cost of increased computational time. More recently, the concurrent work Spann3R [61] extends DUSt3R with spatial memory. It takes a video as input and performs incremental scene reconstruction in a unified coordinate system without requiring global optimization. While this approach significantly improves runtime efficiency, the incremental reconstruction pipeline frame by frame leads to noticeable accumulated drift. Unlike Spann3R, our networks at each hierarchy take multiple frames as input to minimize drift. Additionally, we propose a self-contained retrieval module that, when registering a new frame, this module selects not only its previous few frames but also other similar frames from long-term history for more global scene reference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 370, + 111, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 111, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 111, + 383 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "text", + "content": "Problem statement. Given a monocular video consisting of a sequence of RGB image frames " + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\{I_i\\in \\mathbb{R}^{H\\times W\\times 3}\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "text", + "content": " that captures a static scene, the goal is to reconstruct its dense 3D pointcloud " + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{M\\times 3}" + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 390, + 295, + 486 + ], + "type": "text", + "content": " is the number of 3D points. Our work focuses on three key objectives: maximizing 3D points recovery for reconstruction completeness, improving the accuracy of each recovered point, and achieving these goals while preserving real-time performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 594 + ], + "type": "text", + "content": "System overview. Figure 2 illustrates an overview of the proposed dense reconstruction system. It consists of two main components: an Image-to-Points (I2P) network that recovers local 3D points from video clips, and a Local-to-World (L2W) network that registers local reconstructions into a global scene coordinate system. During the reconstruction of the dense point cloud, the system does not explicitly solve any camera parameters. Instead, it directly predicts 3D point maps in unified coordinate systems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "The system starts by applying a sliding window mechanism of length " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " to convert the input video into short clips " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{W}_i \\in \\mathbb{R}^{L \\times H \\times W \\times 3}\\}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ". The I2P network then processes each window " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " to recover local 3D pointmaps. Within each window, the system selects a keyframe to define a reference coordinate system for point reconstruction, as detailed in Sec. 3.1. By default, the stride of the sliding window is set to 1, ensuring each input frame in the video is selected at least once as a keyframe. For global scene reconstruction, we initialize the world coordinate system with the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "first window and use the reconstructed frames (image and local point map produced by the I2P) as input for the L2W model. The L2W model incrementally register these local reconstructions into a unified global 3D coordinate system. To ensure both accuracy and efficiency during this process, the system maintains a limited reservoir of registered frames, called scene frames. Whenever the L2W model registers a new keyframe, we retrieve the best-correlated scene frames as a reference. The details are introduced in Sec. 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 186, + 509, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 509, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 509, + 198 + ], + "type": "text", + "content": "3.1. Inner-Window Local Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "text", + "content": "The Image-to-Points (I2P) model aims to infer dense 3D pointmaps for every pixel of a keyframe in a given video clip. By default, the middle image of a window " + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "text", + "content": " is chosen as the keyframe " + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "inline_equation", + "content": "I_{key}" + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "text", + "content": " to define the local coordinate system, as it is most likely to have the largest overlap with other frames. The remaining images " + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\{I_{sup_i}\\}_{i = 1}^{L - 1}" + }, + { + "bbox": [ + 313, + 205, + 553, + 300 + ], + "type": "text", + "content": " serve as supporting frames. Note that the 3D pointmaps of supporting frames can also be reconstructed through I2P." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "content": "The I2P network draws inspiration from DUSt3R [64], originally designed for stereo 3D reconstruction. We introduce several simple yet effective modifications to extend it for multi-view scenarios. The I2P model uses a multibranch Vision Transformer (ViT) [13] as its backbone. It consists of a shared encoder " + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "content": ", two separate decoders " + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "inline_equation", + "content": "D_{key}" + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "inline_equation", + "content": "D_{sup}" + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "content": ", and a point regression head for final prediction. These components are detailed below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": "Image encoder. For a given video clip, the image encoder " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": " encodes each frame " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": " to obtain token representations " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "F_{i} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": " is the number of tokens and " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": " is the token dimension. The encoder " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 396, + 553, + 468 + ], + "type": "text", + "content": " ViT encoder blocks, each containing self-attention and feed-forward layers. The encoding process is denoted as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 474, + 523, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 474, + 523, + 491 + ], + "spans": [ + { + "bbox": [ + 345, + 474, + 523, + 491 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {(T \\times d)} = E _ {i m g} (I _ {i} ^ {(H \\times W \\times 3)}), i = 1, \\dots , L.", + "image_path": "ac58765b1343d686aeb138f16a744e42dd5ef7095fe3c11710a274a240ba2694.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": "The frames are processed independently and in parallel, with the output divided into two parts: " + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": " for the keyframe and " + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\{F_{sup_i}\\}_{i=1}^{L-1}" + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": " for the supporting frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": "Keyframe decoder. The keyframe decoder " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_{key}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " ViT decoder blocks, each containing self-attention, cross-attention, and feed-forward layers. Unlike DUSt3R which uses the standard cross-attention, we introduce a novel multi-view cross-attention to combine information from different supporting frames. Given the feature tokens " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{F_{sup_i}\\}_{i=1}^{L-1}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": ", the keyframe decoder " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_{key}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " as input for self-attention and performs cross-attention between " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{F_{sup_i}\\}_{i=1}^{L-1}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": ". A decoder block is illustrated in Figure 3. For each cross-attention layer, queries are taken from " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": ", while keys and values are extracted from the supporting tokens " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{sup_i}" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": ". These " + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "inline_equation", + "content": "L-1" + }, + { + "bbox": [ + 313, + 535, + 555, + 713 + ], + "type": "text", + "content": " cross-attention layers are independent of each other, allowing for parallel processing. A max-pooling layer is then employed to aggregate features after cross-attention. We obtain decoded" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16653" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 553, + 210 + ], + "type": "image", + "image_path": "1cf1dd152ac4f10e494115256d7b0d5b3cd4d6e1d4851006aa050e406e416a35.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 305 + ], + "type": "text", + "content": "Figure 2. System overview. Given an input monocular RGB video, we apply a sliding window mechanism to convert it into overlapping clips (referred to as windows). Each window is fed into an Image-to-Points (I2P) network to recover 3D points in a local coordinate system. Next, the local points are incrementally fed into a Local-to-World (L2W) network to create a globally consistent scene model. The proposed I2P and L2W networks elegantly share similar architectures. In the I2P step (Sec. 3.1), we select a keyframe as a reference to set up a local coordinate system and use the remaining frames in the window to estimate the 3D geometry captured within it. The points from the first window are used to establish the world coordinate system. We then incrementally fuse the following windows in the L2W step (Sec. 3.2). This process involves retrieving the most relevant already-registered keyframes as a reference, and integrating new keyframes. Through this iterative process, we eventually obtain the full scene reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 326, + 162, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 162, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 162, + 339 + ], + "type": "text", + "content": "keyframe tokens " + }, + { + "bbox": [ + 55, + 326, + 162, + 339 + ], + "type": "inline_equation", + "content": "G_{key}" + }, + { + "bbox": [ + 55, + 326, + 162, + 339 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 348, + 259, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 348, + 259, + 363 + ], + "spans": [ + { + "bbox": [ + 91, + 348, + 259, + 363 + ], + "type": "interline_equation", + "content": "G _ {k e y} = D _ {k e y} (F _ {k e y}, F _ {s u p _ {1}}, \\dots , F _ {s u p _ {L - 1}}).", + "image_path": "fdf8cfedda300e964656de35260e4eaa0fce977510fce193770782c8de0ad4eb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 64, + 384, + 289, + 545 + ], + "blocks": [ + { + "bbox": [ + 64, + 384, + 289, + 545 + ], + "lines": [ + { + "bbox": [ + 64, + 384, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 64, + 384, + 289, + 545 + ], + "type": "image", + "image_path": "9328c85455681f3182ac79be86dcd56cc43692a32c381184ebc4032850bcf15b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "text", + "content": "Figure 3. Illustration of a decoder block in the proposed keyframe decoder " + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "inline_equation", + "content": "D_{key}" + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "text", + "content": ". We present a minimalist modification to integrate information from different supporting images. Our approach traverses each of them, selects its token keys and values, and uses the keyframe queries to interact with them separately across the supporting images. This multi-view information is then aggregated through max-pooling. The registration decoder " + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "inline_equation", + "content": "D_{reg}" + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "text", + "content": " and scene decoder " + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "inline_equation", + "content": "D_{sce}" + }, + { + "bbox": [ + 55, + 552, + 295, + 641 + ], + "type": "text", + "content": " (described in Sec. 3.2) share the same architecture." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Supporting decoder. The supporting decoder " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{sup}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " is designed to complement the keyframe decoder. It inherits the decoder architecture used in DUSt3R, consisting of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " standard ViT decoder blocks. The cross-attention mechanism is applied only to exchange information with the keyframe." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 326, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 350 + ], + "type": "text", + "content": "Note that all supporting frames share the same " + }, + { + "bbox": [ + 313, + 326, + 555, + 350 + ], + "type": "inline_equation", + "content": "D_{sup}" + }, + { + "bbox": [ + 313, + 326, + 555, + 350 + ], + "type": "text", + "content": ". This process is denoted as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 356, + 529, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 356, + 529, + 370 + ], + "spans": [ + { + "bbox": [ + 339, + 356, + 529, + 370 + ], + "type": "interline_equation", + "content": "G _ {s u p _ {i}} = D _ {s u p} (F _ {s u p _ {i}}, F _ {k e y}), i = 1, \\dots , L - 1.", + "image_path": "5528f6de30276e9f269597b62d432a0bb708a7ea0fad8c0e35975aec7a966f56.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 374, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 444 + ], + "type": "text", + "content": "Points reconstruction. Similar to DUSt3R, we apply a linear head [64] to regress dense 3D pointmaps in the unified coordinate system from decoded tokens. In addition to the pointmaps, we also predict the confidence maps for all frames to evaluate their reliability. The final predictions are:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 448, + 542, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 448, + 542, + 466 + ], + "spans": [ + { + "bbox": [ + 326, + 448, + 542, + 466 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {i} ^ {(H \\times W \\times 3)}, \\hat {C} _ {i} ^ {(H \\times W \\times 1)} = \\mathrm {H} (G _ {i} ^ {(T \\times d)}), i = 1, \\dots , L.", + "image_path": "1f42d1024d3b75684c9bcadbfa34f0841e135ebcda5165a360c633a79a892f53.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "type": "text", + "content": "Training loss. Following DUSt3R, the I2P network is trained end-to-end using ground-truth scene points " + }, + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\{X_{i}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "type": "text", + "content": ". Both the ground truth and predicted point maps are normalized to a canonical scale, determined by the average distance of all valid points within the window to the origin. The confidence-aware training loss is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 329, + 556, + 538, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 556, + 538, + 590 + ], + "spans": [ + { + "bbox": [ + 329, + 556, + 538, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I 2 P} = \\sum_ {i = 1} ^ {L} M _ {i} \\cdot (\\hat {C} _ {i} \\cdot \\mathrm {L} 1 (\\frac {1}{\\hat {z}} \\hat {X} _ {i}, \\frac {1}{z} X _ {i}) - \\alpha \\log \\hat {C} _ {i}),", + "image_path": "2fb97728ad4eb58db6bf1d2a4fa6bf8052fa7dbc75b7056f511b7b04b3f30002.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " represents a mask of valid points that have ground-truth values in " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " are the scale factor, " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\hat{C}_i" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " is the confidence map, the operator " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " denotes the elementwise matrix multiplication, " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{L1}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " denotes the point-wise Euclidean distance, and " + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 594, + 555, + 666 + ], + "type": "text", + "content": " is a hyper-parameter to control the regularization term. We will detail the process in Sec 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "type": "text", + "content": "3.2. Inter-Window Global Registration" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "After obtaining the 3D pointmap " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\hat{X}_{key}\\}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " from the I2P network, we use the inter-window Local-to-World (L2W)" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16654" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "model to incrementally register the newly generated pointmap into a global 3D coordinate system. Similar to the I2P network, the L2W model also relies on some frames to serve as a reference for the scene coordinate system. Furthermore, it can leverage multiple registered keyframes as a global reference. These registered keyframes are referred to as scene frames and they are maintained in a buffering set through a sampling mechanism." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "type": "text", + "content": "The buffering set is designed for scalability in handling long videos. We apply a reservoir strategy [59] that maintains a maximum of " + }, + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "type": "text", + "content": " registered frames in the buffering set. When a new keyframe is inferred from I2P and ready for fusion, we retrieve the top- " + }, + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 169, + 296, + 241 + ], + "type": "text", + "content": " best-correlated scene frames from the buffering set as its support for global registration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "text", + "content": "Scene initialization. The first window is used to initialize the scene model. It's crucial to ensure this initialization is as accurate as possible. To achieve this, we execute the I2P network " + }, + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 242, + 296, + 361 + ], + "type": "text", + "content": " times, attempting to traverse and designate each frame within the window as the keyframe. We then select the result with the highest total confidence score for scene model initialization. This process results in a scene pointcloud along with a set of registered frames. All these frames are regarded as scene frames, and they are used to initialize the buffering set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": "Reservoir and retrieval. Each scene frame " + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "inline_equation", + "content": "I_{sce_i}" + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": " is recorded with its latent feature " + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "inline_equation", + "content": "F_{sce_i}" + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": " and pointmaps " + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\tilde{X}_{sce_i}" + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": ". For efficiency, we apply reservoir sampling to allow storing an unbiased subset of an empirical distribution in a bounded amount of memory. The first " + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": " registered frames chosen are directly inserted into the buffering set. For each subsequent frame with " + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "inline_equation", + "content": "id > B" + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": ", the probability of inserting it is " + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "inline_equation", + "content": "B / id" + }, + { + "bbox": [ + 55, + 363, + 296, + 471 + ], + "type": "text", + "content": ". If chosen for insertion, it will randomly replace one of the current scene frames in the buffering set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "type": "text", + "content": "Given a new keyframe " + }, + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "type": "inline_equation", + "content": "I_{key}" + }, + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "type": "text", + "content": " to be registered, we feed its feature " + }, + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "type": "inline_equation", + "content": "F_{key}" + }, + { + "bbox": [ + 55, + 472, + 296, + 507 + ], + "type": "text", + "content": " and the features from the buffering set into a retrieval module," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 517, + 238, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 517, + 238, + 534 + ], + "spans": [ + { + "bbox": [ + 112, + 517, + 238, + 534 + ], + "type": "interline_equation", + "content": "\\operatorname {R e t r i e v a l} \\left(F _ {\\text {k e y}} ^ {(T \\times d)}, \\left\\{F _ {\\text {s c e} _ {i}} ^ {(T \\times d)} \\right\\}\\right),", + "image_path": "4be2c5aa571358c78d5c2d9596e09c08853a1f46cbdc5e6bc748e111d3c6c5b3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": "to obtain a list of correlation scores, measuring both the visual similarity and baseline suitability between the keyframe and the scene frames in the buffering set. The retrieval module uses the first " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": " decoder blocks from the I2P module as its backbone. A linear projection and an average-pooling layer follow, together producing an image-wise correlation score. We then select the top- " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": " scene frames as a global reference to fuse the current keyframe. As a result, we have " + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": " scene frames and one keyframe as the input for the following L2W model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Points embedding. The 3D pointmaps reconstructed by the I2P model are encoded into the L2W model using a patch embedding method similar to image patchification in the ViT encoder " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". We process the new keyframe and " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 463, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 463, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 463, + 84 + ], + "type": "text", + "content": "retrieved scene frames in parallel as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 335, + 94, + 532, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 94, + 532, + 110 + ], + "spans": [ + { + "bbox": [ + 335, + 94, + 532, + 110 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {i} ^ {(T \\times d)} = E _ {p t s} (\\hat {X} _ {i} ^ {(H \\times W \\times 3)}), i = 1, \\dots , K + 1.", + "image_path": "50a92b21620fb844820a9c8599b44a7f0af7598db1a57259a7b0921e51c1f507.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 145 + ], + "type": "text", + "content": "The encoded geometric tokens are combined with their corresponding visual tokens by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 155, + 532, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 155, + 532, + 171 + ], + "spans": [ + { + "bbox": [ + 335, + 155, + 532, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {i} ^ {(T \\times d)} = F _ {i} ^ {(T \\times d)} + \\mathcal {P} _ {i} ^ {(T \\times d)}, i = 1, \\dots , K + 1.", + "image_path": "dcc280d383df8f3307841e8c75084309f47133aeb77ff5f798523a9858ff6b4b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "text", + "content": "This resulting a token set " + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}" + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "text", + "content": " contains joint features of image patch appearance and 3D geometry for the keyframe and retrieved scene frames. In the following decoders, " + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}\\}" + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "text", + "content": " are further accumulated to " + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}\\}" + }, + { + "bbox": [ + 313, + 181, + 553, + 240 + ], + "type": "text", + "content": " between adjacent blocks to enhance the geometric representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "content": "Registration decoder. The registration decoder " + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "inline_equation", + "content": "D_{reg}" + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "content": " takes feature tokens " + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}" + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "content": " as input and aims to transform the local reconstruction of the keyframe to the scene coordinate system. It takes the same network architecture of the keyframe decoder " + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "inline_equation", + "content": "D_{key}" + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "content": ". This decoding process is denoted by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 355, + 324, + 512, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 324, + 512, + 338 + ], + "spans": [ + { + "bbox": [ + 355, + 324, + 512, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {k e y} = D _ {r e g} \\left(\\mathcal {F} _ {k e y}, \\mathcal {F} _ {s c e _ {1}}, \\dots , \\mathcal {F} _ {s c e _ {K}}\\right).", + "image_path": "dd45e0d3c6f3c7e86cbb2efe112c4a6c0d774ed32d61bfdd213a1e84676ad77f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "content": "Scene decoder. The scene decoder " + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "inline_equation", + "content": "D_{sce}" + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "content": " takes the token set " + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{key},\\{\\mathcal{F}_{sce_i}\\}_{i = 1}^K\\}" + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "content": " as input to refine the scene geometry without coordinate system changes. It uses the same network architecture as the keyframe decoder " + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "inline_equation", + "content": "D_{key}" + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "content": ", allowing us to extend to multi-keyframe co-registration (see supplementary material for details). By default, we register one keyframe each time. Each of the " + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sce_i}" + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "content": " has information exchange only with the " + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{key}" + }, + { + "bbox": [ + 313, + 349, + 554, + 456 + ], + "type": "text", + "content": ". This decoding process is denoted by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 468, + 518, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 468, + 518, + 481 + ], + "spans": [ + { + "bbox": [ + 349, + 468, + 518, + 481 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {s c e _ {i}} = D _ {s c e} \\left(\\mathcal {F} _ {s c e _ {i}}, \\mathcal {F} _ {k e y}\\right), i = 1, \\dots , K.", + "image_path": "d7ec6bef398e5a8b2a08c46355516fd3cc40cd7b7fb8c33981d6d6c216fa0223.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "type": "text", + "content": "Points reconstruction and training loss. We apply the same head design as that of the I2P network to predict all the pointmaps " + }, + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\tilde{X}_i" + }, + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "type": "text", + "content": " in the global scene coordinate system:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 537, + 551, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 551, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 551, + 555 + ], + "type": "interline_equation", + "content": "\\tilde {X} _ {i} ^ {(H \\times W \\times 3)}, \\tilde {C} _ {i} ^ {(H \\times W \\times 1)} = \\mathrm {H} (\\mathcal {G} _ {i} ^ {(T \\times d)}), i = 1, \\dots , K + 1.", + "image_path": "ec27d0394d6995b5ad18b77a0ba7b1c8453abdfe2d5aed3d5c17e534f2aad681.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "content": "We train the L2W network using a similar loss function as the I2P network. Differently, no normalization is applied to the predicted point map, as the output scale must align with the scene frames in the input. This alignment ensures that the output can be directly integrated into the existing reconstruction. The training loss of the L2W network is:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 335, + 647, + 532, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 647, + 532, + 680 + ], + "spans": [ + { + "bbox": [ + 335, + 647, + 532, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L 2 W} = \\sum_ {i = 1} ^ {L} M _ {i} \\cdot \\left(\\tilde {C} _ {i} \\cdot \\mathrm {L} 1 \\left(\\tilde {X} _ {i}, X _ {i}\\right) - \\alpha \\log \\tilde {C} _ {i}\\right).", + "image_path": "918a4debf721697b51d16e7446a67aacab4dd6007dd5845354ce92254cdbe8ad.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "The following section provides a detailed discussion of the training process and its implementation details." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16655" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 535, + 140 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 535, + 140 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 535, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 535, + 140 + ], + "type": "table", + "html": "
MethodChess Acc. / Comp.Fire Acc. / Comp.Heads Acc. / Comp.Office Acc. / Comp.Pumpkin Acc. / Comp.RedKitchen Acc. / Comp.Stairs Acc. / Comp.Average Acc. / Comp.FPS
DUS3R [64]2.26 / 2.131.04 / 1.501.66 / 0.984.62 / 4.741.73 / 2.431.95 / 2.363.37 / 10.752.19 / 3.24<1
MAS3R [28]2.08 / 2.121.54 / 1.431.06 / 1.043.23 / 3.195.68 / 3.073.50 / 3.372.36 / 13.163.04 / 3.90≪1
Spann3R [61]2.23 / 1.680.88 / 0.922.67 / 0.985.86 / 3.542.25 / 1.852.68 / 1.805.65 / 5.153.42 / 2.41>50
SLAM3R-NoConf (Ours)2.12 / 1.210.95 / 0.803.23 / 1.672.59 / 2.211.99 / 2.042.09 / 1.884.54 / 6.382.40 / 2.24~25
SLAM3R (Ours)1.63 / 1.310.84 / 0.832.95 / 1.222.32 / 2.261.81 / 2.051.84 / 1.944.19 / 6.912.13 / 2.34~25
", + "image_path": "69f6649d353ad4a9524d8bf7436d7d906249f454b0eacece42760ab847fdce73.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 191, + 553, + 292 + ], + "blocks": [ + { + "bbox": [ + 55, + 148, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 555, + 182 + ], + "type": "text", + "content": "Table 1. Reconstruction results on 7 Scenes [51] dataset. The average numbers are computed over all test sequences. The methods are categorized into two groups based on whether their FPS is above or below 1. The best results within each category are shown in bold. We report accuracy and completeness in centimeters. The color gradient shifts from red through yellow to green to show increasing FPS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 191, + 553, + 292 + ], + "lines": [ + { + "bbox": [ + 58, + 191, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 58, + 191, + 553, + 292 + ], + "type": "table", + "html": "
MethodRoom 0 Acc. / Comp.Room 1 Acc. / Comp.Room 2 Acc. / Comp.Office 0 Acc. / Comp.Office 1 Acc. / Comp.Office 2 Acc. / Comp.Office 3 Acc. / Comp.Office 4 Acc. / Comp.Average Acc. / Comp.FPS
DUSt3R [64]3.47 / 2.502.53 / 1.862.95 / 1.764.92 / 3.513.09 / 2.214.01 / 3.103.27 / 2.253.66 / 2.613.49 / 2.48<1
MASt3R [28]4.01 / 4.103.61 / 3.253.13 / 2.152.57 / 1.6312.85 / 8.133.13 / 1.994.67 / 3.153.69 / 2.474.71 / 3.36≤1
NICER-SLAM [79]*2.53 / 3.043.93 / 4.103.40 / 3.425.49 / 6.093.45 / 4.424.02 / 4.293.34 / 4.033.03 / 3.873.65 / 4.16≤1
DROID-SLAM [56]*12.18 / 8.968.35 / 6.073.26 / 16.013.01 / 16.192.39 / 16.205.66 / 15.564.49 / 9.734.65 / 9.635.50 / 12.29~20
DIM-SLAM [29]*14.19 / 6.249.56 / 6.458.41 / 12.1710.16 / 5.957.86 / 8.3316.50 / 8.2813.01 / 6.7713.08 / 8.6211.60 / 7.85~3
GO-SLAM [75]--------3.81 / 4.79~8
Spann3R [61]9.75 / 12.9415.51 / 12.947.28 / 8.505.46 / 18.755.24 / 16.649.33 / 11.8016.00 / 9.0313.97 / 16.0210.32 / 13.33>50
SLAM3R-NoConf (Ours)3.37 / 2.403.22 / 2.333.15 / 2.004.43 / 2.593.18 / 2.343.95 / 2.784.20 / 3.154.57 / 3.383.76 / 2.62~24
SLAM3R (Ours)3.19 / 2.403.12 / 2.342.72 / 2.004.28 / 2.603.17 / 2.343.84 / 2.783.90 / 3.164.32 / 3.363.57 / 2.62~24
", + "image_path": "e604eaee299e3ec9c1ebc931ac85b23d5a9542484657d87d5e1452707100459b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 64, + 322, + 545, + 354 + ], + "blocks": [ + { + "bbox": [ + 115, + 300, + 493, + 310 + ], + "lines": [ + { + "bbox": [ + 115, + 300, + 493, + 310 + ], + "spans": [ + { + "bbox": [ + 115, + 300, + 493, + 310 + ], + "type": "text", + "content": "Table 2. Reconstruction results on Replica [54] dataset. * denotes the results reported in NICER-SLAM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 322, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 64, + 322, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 64, + 322, + 545, + 354 + ], + "type": "table", + "html": "
DUSt3R [64]MASt3R [28]NICER-SLAM [79]*DROID-SLAM [56]*DIM-SLAM [29]GO-SLAM [75]Spann3R [61]SLAM3R-NoConf (Ours)SLAM3R (Ours)
7 Scenes8.026.288.555.66--11.708.448.41
Replica4.761.671.880.330.460.3932.796.616.61
", + "image_path": "51955a9f1835514b360d08862e49c9291fe34f773e4ac771d1f1f4c26a816fe5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 362, + 493, + 373 + ], + "lines": [ + { + "bbox": [ + 116, + 362, + 493, + 373 + ], + "spans": [ + { + "bbox": [ + 116, + 362, + 493, + 373 + ], + "type": "text", + "content": "Table 3. Camera pose results evaluated by ATE-RMSE (cm) on 7 Scenes [51] and Replica [54] datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 387, + 164, + 463 + ], + "blocks": [ + { + "bbox": [ + 60, + 387, + 164, + 463 + ], + "lines": [ + { + "bbox": [ + 60, + 387, + 164, + 463 + ], + "spans": [ + { + "bbox": [ + 60, + 387, + 164, + 463 + ], + "type": "image", + "image_path": "44345ea07f6baf0eb2ac594b6ac8db7b6c7d52b1a43792d0a855fbf0d20ce29a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 462, + 143, + 529 + ], + "blocks": [ + { + "bbox": [ + 60, + 462, + 143, + 529 + ], + "lines": [ + { + "bbox": [ + 60, + 462, + 143, + 529 + ], + "spans": [ + { + "bbox": [ + 60, + 462, + 143, + 529 + ], + "type": "image", + "image_path": "4d20d4f1c45c1c75db0ac26467fa3204ec21b1a1b942d11bcba92fecb577fcd3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 529, + 121, + 538 + ], + "lines": [ + { + "bbox": [ + 88, + 529, + 121, + 538 + ], + "spans": [ + { + "bbox": [ + 88, + 529, + 121, + 538 + ], + "type": "text", + "content": "Spann3R" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 548, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 548, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 555, + 572 + ], + "type": "text", + "content": "Figure 4. We visualize the reconstruction results on two scenes: Office-09 and Office 2 from the 7-Scenes [51] and Replica [54] datasets. Our method runs in real-time and achieves high-quality reconstruction comparable to the offline method DUSt3R [64]." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 165, + 387, + 288, + 463 + ], + "blocks": [ + { + "bbox": [ + 165, + 387, + 288, + 463 + ], + "lines": [ + { + "bbox": [ + 165, + 387, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 165, + 387, + 288, + 463 + ], + "type": "image", + "image_path": "0af3d4918d0308fc33eea8a0a5322a26ce8641a9335ec7451d5a7732d39b607b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 165, + 463, + 289, + 525 + ], + "blocks": [ + { + "bbox": [ + 165, + 463, + 289, + 525 + ], + "lines": [ + { + "bbox": [ + 165, + 463, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 165, + 463, + 289, + 525 + ], + "type": "image", + "image_path": "309d4c9fb0ceeec648887a6c95d84d1f1c9ef8404ce2af3f9b1b1cad7a5dade8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 529, + 243, + 537 + ], + "lines": [ + { + "bbox": [ + 214, + 529, + 243, + 537 + ], + "spans": [ + { + "bbox": [ + 214, + 529, + 243, + 537 + ], + "type": "text", + "content": "DUSt3R" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 293, + 387, + 419, + 463 + ], + "blocks": [ + { + "bbox": [ + 293, + 387, + 419, + 463 + ], + "lines": [ + { + "bbox": [ + 293, + 387, + 419, + 463 + ], + "spans": [ + { + "bbox": [ + 293, + 387, + 419, + 463 + ], + "type": "image", + "image_path": "200ae69a4d04f2b4fc8008b95a78ccbf15cded675d5513c4eb3fc283b1f31a58.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 293, + 463, + 419, + 525 + ], + "blocks": [ + { + "bbox": [ + 293, + 463, + 419, + 525 + ], + "lines": [ + { + "bbox": [ + 293, + 463, + 419, + 525 + ], + "spans": [ + { + "bbox": [ + 293, + 463, + 419, + 525 + ], + "type": "image", + "image_path": "5453a8c6eea0f90eef1faa2e3e7c4def1e186093507c7db842f3b8018e9598c1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 528, + 380, + 538 + ], + "lines": [ + { + "bbox": [ + 324, + 528, + 380, + 538 + ], + "spans": [ + { + "bbox": [ + 324, + 528, + 380, + 538 + ], + "type": "text", + "content": "SLAM3R (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 420, + 387, + 550, + 463 + ], + "blocks": [ + { + "bbox": [ + 420, + 387, + 550, + 463 + ], + "lines": [ + { + "bbox": [ + 420, + 387, + 550, + 463 + ], + "spans": [ + { + "bbox": [ + 420, + 387, + 550, + 463 + ], + "type": "image", + "image_path": "d5b41e2e7e51997716d37255fafedb4a2504a3de2d032030de7833a298909481.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 421, + 463, + 550, + 525 + ], + "blocks": [ + { + "bbox": [ + 421, + 463, + 550, + 525 + ], + "lines": [ + { + "bbox": [ + 421, + 463, + 550, + 525 + ], + "spans": [ + { + "bbox": [ + 421, + 463, + 550, + 525 + ], + "type": "image", + "image_path": "beb59d5df17aed1ca7a0573c6a38e9e65349dc67a5ac2f453bf0d63b8a748181.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 529, + 511, + 537 + ], + "lines": [ + { + "bbox": [ + 461, + 529, + 511, + 537 + ], + "spans": [ + { + "bbox": [ + 461, + 529, + 511, + 537 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 590, + 137, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 137, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 137, + 605 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Datasets. For both the Image-to-Points (I2P) and Local-to-World (L2W) models, we perform training with a mixture of three datasets: ScanNet++ [71], Aria Synthetic Environments [3] and CO3D-v2 [41]. These datasets vary from scene-level to object-centric, and contain both real-world and synthetic scenes. Since they are all recorded sequentially, we can easily extract video clips with a sliding-window mechanism as our training data. We select about" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 591, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 676 + ], + "type": "text", + "content": "850K clips for training in total. To validate our reconstruction quality, we conduct quantitative evaluations on two unseen datasets: 7 Scenes [51], a real-world dataset of partial scenes, and Replica [54], a synthetic dataset of complete scenes. We also demonstrate visual reconstruction results across diverse datasets and in-the-wild captured videos to showcase the generalization ability of SLAM3R." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Implementation details. Both the I2P and L2W models build upon the architecture of DUSt3R [64] with minimal but effective modifications, making it natural for them to" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16656" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 305 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 305 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 305 + ], + "type": "image", + "image_path": "1c0ed92335b18e80c89f9545ce3b3994448a23ce6aff48003530a111283e3731.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "type": "text", + "content": "Figure 5. Qualitative examples. We show our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild captured videos, to demonstrate SLAM3R's generalization ability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": "initialize their weights from the DUSt3R pre-trained model. We initialize our weights using the DUSt3R model trained on " + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": " resolution images with " + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "inline_equation", + "content": "m = 24" + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": " encoder blocks, " + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "inline_equation", + "content": "n = 12" + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": " decoder blocks with linear heads. All images are center-cropped to " + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": " pixels before feeding into SLAM3R. Our training is conducted on 8 NVIDIA 4090D GPUs, each with 24 GB of memory. It takes about one day. At test time, we set the initial window length to " + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "inline_equation", + "content": "L = 5" + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": " to ensure high-quality reconstruction of all frames within the window. For subsequent incremental windows, we use " + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "inline_equation", + "content": "L = 11" + }, + { + "bbox": [ + 54, + 356, + 297, + 513 + ], + "type": "text", + "content": " to provide more supporting views for better keyframe reconstruction. Please refer to our supplementary material for more implementation details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 524, + 142, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 142, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 142, + 537 + ], + "type": "text", + "content": "4.1. Comparisons" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 688 + ], + "type": "text", + "content": "Evaluation metrics. Following NICER-SLAM [79] and Spann3R [61], we build a ground-truth point cloud model for each test sequence by back-projecting pixels to the world using ground-truth depths and camera parameters. The reconstructed point clouds are aligned to ground truths using Umayama [58] and ICP [43] algorithms. We quantify reconstruction quality through accuracy and completeness metrics. To demonstrate computational efficiency, we report frames per second (FPS) on a single NVIDIA 4090D GPU. We also evaluate camera poses using absolute trajectory error (ATE-RMSE). For detailed formulations of the metrics, please refer to the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Reconstruction results on the 7 Scenes [51] dataset. The numerical results of scene reconstruction quality are re" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 356, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 548 + ], + "type": "text", + "content": "ported in Table 1. Following Spann3R [61]'s setting, we uniformly sample one-twentieth of the frames in each test sequence as input video. Each video is regarded as an individual scene. We evaluate SLAM3R using two settings: integrating the full pointmaps predicted for all input frames to create reconstruction results (denoted by SLAM3R-NoConf), and filtering pointmaps with a confidence threshold of 3 before creating reconstruction results (SLAM3R). We compare our method with optimization-based reconstruction DUSt3R [64], triangulation-based MASt3R [28], and online incremental reconstruction Spann3R. Our method outperforms all baselines in both accuracy and completeness while maintaining real-time performance. As shown in the Office-09 scene (the top row in Figure 4), our approach demonstrates much less drift compared to the concurrent work Spann3R [61]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Reconstruction results on the Replica [54] dataset. Besides the baselines mentioned in 7 Scene datasets, we also compare the SLAM-based reconstruction approaches NICER-SLAM [79], DROID-SLAM [56], DIM-SLAM [29] and GO-SLAM [75] on the Replica [54] dataset. The numerical results on full scene reconstruction are reported in Table 2. Due to the memory constraint, DUSt3R [64] and MASt3R [28] process only one-twentieth of the frames for reconstruction. As is shown in the table, our method surpasses all baselines with FPS greater than 1. Notably, without any optimization procedure, our method achieves reconstruction quality comparable to optimization-based methods such as NICER-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16657" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 271, + 153 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 271, + 153 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 271, + 153 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 271, + 153 + ], + "type": "table", + "html": "
Method# FramesAcc.Comp.FPS
DUSt3R [64]23.162.8942.55
I2P23.393.0442.55
I2P52.622.2840.82
I2P (Default)112.382.0340.11
I2P152.271.9435.51
I2P512.231.8611.97
", + "image_path": "7f0c070484c1535c94ab62ed796ef597f6a05faac4da175ef6e632d36df5d113.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 217, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 295, + 252 + ], + "type": "text", + "content": "SLAM [79] and DUS3R [64]. Example of the Office 2 (the bottom row in Figure 4) also illustrates the global consistency of our reconstruction result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 253, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 295, + 397 + ], + "type": "text", + "content": "Camera pose estimation on 7 Scenes [51] and Replica [54]. Our method is designed in a new paradigm that reconstructs 3D points end-to-end without explicitly solving camera parameters. Following DUSt3R [64], We also derive camera poses from the predicted scene points using PnP-RANSAC solver in OpenCV [6] with ground truth camera intrinsics of each frame. The results are reported in Table 3. We can observe that camera poses and scene reconstruction results are not fully positively correlated. This discrepancy between pose and reconstruction errors indicates that effective end-to-end 3D reconstruction is possible and promising without first obtaining precise camera poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "type": "text", + "content": "For more details on the comparisons in this section, please refer to our supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 431, + 119, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 119, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 119, + 443 + ], + "type": "text", + "content": "4.2. Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 449, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 640 + ], + "type": "text", + "content": "Effectiveness of the I2P model. To highlight the advantages of our multi-view I2P model over the original two-view DUSt3R [64], we evaluate the reconstruction quality of keyframes with varying numbers of supporting views. We conduct experiments on the Replica [54] dataset, where input views are sampled using a sliding window of different sizes, and the reconstruction accuracy and completeness of the keyframes are computed. The results are reported in Table 4. As the number of supporting views increases, our approach progressively improves reconstruction quality. Notably, the efficiency of our method remains stable until the window size exceeds 11, demonstrating the effectiveness of our parallel design. However, the results also show diminishing returns as the number of views increases, which we detail in the supplementary material. Visual results of I2P reconstruction can be found in Figure 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Advantages of the L2W model. The effectiveness of the L2W model is evaluated through ablation studies on the Replica [54] dataset. Per-window reconstructions are first generated with a window size of 11 using the I2P model. Local points are then aligned to a unified coordinate frame using different methods: global optimization" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 351, + 70, + 519, + 129 + ], + "blocks": [ + { + "bbox": [ + 55, + 162, + 295, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 295, + 195 + ], + "type": "text", + "content": "Table 4. Inner-window keyframe reconstruction results with various window lengths. By default, we use 11-frame windows for incremental reconstruction to balance quality and efficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 70, + 519, + 129 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 519, + 129 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 519, + 129 + ], + "type": "table", + "html": "
MethodAcc.Comp.FPS
I2P+GA4.873.00~3
I2P+UI7.473.86~1
I2P+L2W6.193.54~92
I2P+L2W+Re (Full)3.622.70~43
", + "image_path": "c98512c71147373f63fd7808dd9ba278dfd351dbd9a602e61c8431423fa1fcf0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 137, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 171 + ], + "type": "text", + "content": "Table 5. Reconstruction results using various point alignment methods and scene frame selection strategies. The FPS reported only accounts for the overhead of the alignment operation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 194, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 290 + ], + "type": "text", + "content": "from DUSt3R [64] (I2P-GA), traditional approaches such as Umeyama [58] and ICP [43] (I2P+UI), and our L2W model (I2P+L2W+Re). For consistency, we set the window size for global optimization to 10, which is equal to the number of views used to align new frames in other methods. Results in Table 5 show that our full method achieves superior alignment accuracy and computational efficiency compared to the alternatives." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 388 + ], + "type": "text", + "content": "Analysis of the retrieval module. We propose a lightweight retrieval module that selects historical scene frames from the reservoir. This approach effectively performs implicit re-localization. We compare our retrieval method with a baseline approach that selects the ten nearest previous frames, named I2P+L2W. The results in Table 5 indicate a significant performance improvement with our retrieval strategy, demonstrating its effectiveness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 509 + ], + "type": "text", + "content": "In-the-wild scene reconstruction. We have tested our method on a diverse range of unseen datasets and found that SLAM3R shows strong generalization capabilities. Figure 5 shows our reconstruction results on Tanks and Temples [26], BlendedMVS [69], Map-free Reloc [2], LLFF [35], and ETH3D [49, 50] datasets, as well as in-the-wild videos we captured. These results show that our method performs reliably across different scales and scenes. We also provide additional numerical results on sampled scenes from these datasets in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 523, + 388, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 388, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 388, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 544, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 640 + ], + "type": "text", + "content": "In this paper, we present SLAM3R, a novel and effective system that performs real-time, high-quality, dense 3D scene reconstruction using RGB videos. It employs a two-hierarchy neural network framework to perform end-to-end 3D reconstruction through streamlined feed-forward processes, eliminating the need to explicitly solve any camera parameters. Experiments demonstrate its state-of-the-art reconstruction quality and real-time efficiency (20+ FPS)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "content": "Limitations and future work. The elimination of camera parameter prediction prevents us from performing global bundle adjustment. Additionally, the poses derived from our scene point cloud prediction still fall short of SLAM systems that specialize in camera localization. Addressing these limitations will be the focus of our future work." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16658" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 224 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (2022ZD0160800). This work is also supported by the Early Career Scheme of the Research Grants Council (grant # 27207224), the HKU-100 Award, a donation from the Musketeers Foundation, and in part by the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. We thank the reviewers for their valuable feedback, as well as Zihan Zhu and Songyou Peng for their help with our experiments. Siyan Dong would also like to thank the support from HKU School of Computing and Data Science." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 295, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 361 + ], + "type": "text", + "content": "[2] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690-708. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 361, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 426 + ], + "type": "text", + "content": "[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescipt: Reconstructing scenes with an autoregressive structured language model. arXiv preprint arXiv:2403.13064, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 427, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 295, + 460 + ], + "type": "text", + "content": "[4] Tim Bailey and Hugh Durrant-Whyte. Simultaneous localization and mapping (slam): Part ii. IEEE robotics & automation magazine, 13(3):108-117, 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 515 + ], + "type": "text", + "content": "[5] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2560–2568, 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 515, + 296, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 296, + 537 + ], + "type": "text", + "content": "[6] G Bradski. The opencv library. Dr. Dobb's Journal of Software Tools, 2000. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 537, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 592 + ], + "type": "text", + "content": "[7] Carlos Campos, Richard Elvira, Juan J Gomez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 635 + ], + "type": "text", + "content": "[8] Yiyang Chen, Siyan Dong, Xulong Wang, Lulu Cai, Youyi Zheng, and Yanchao Yang. Sg-nerf: Neural surface reconstruction with scene graph optimization. arXiv preprint arXiv:2407.12667, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 636, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 296, + 712 + ], + "type": "text", + "content": "[9] Chi-Ming Chung, Yang-Che Tseng, Ya-Ching Hsu, Xiang-Qian Shi, Yun-Hung Hua, Jia-Fong Yeh, Wen-Chin Chen, Yi-Ting Chen, and Winston H Hsu. Orbeez-slam: A real-time monocular visual slam with orb features and nefrealized mapping. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9400-9406. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[10] Jan Czarnowski, Tristan Laidlow, Ronald Clark, and Andrew J Davison. Deepfactors: Real-time probabilistic dense monocular slam. IEEE Robotics and Automation Letters, 5 (2):721-728, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 218 + ], + "type": "text", + "content": "[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 251 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "text", + "content": "[14] Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion. arXiv preprint arXiv:2409.19152, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "text", + "content": "[15] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99-110, 2006. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "text", + "content": "[16] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 410 + ], + "type": "text", + "content": "[17] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE transactions on pattern analysis and machine intelligence, 40(3):611-625, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "text", + "content": "[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 555, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 500 + ], + "type": "text", + "content": "[19] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 554, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 545 + ], + "type": "text", + "content": "[20] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 546, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 611 + ], + "type": "text", + "content": "[21] Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 668 + ], + "type": "text", + "content": "[22] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "text", + "content": "[23] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16659" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "text", + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[25] Georg Klein and David Murray. Parallel tracking and mapping on a camera phone. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 83-86. IEEE, 2009. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[26] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[27] Lukas Koestler, Nan Yang, Niclas Zeller, and Daniel Cremers. Tandem: Tracking and dense mapping in real-time using deep multi-view stereo. In Conference on Robot Learning, pages 34-45. PMLR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "text", + "content": "[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 349 + ], + "type": "text", + "content": "[29] Heng Li, Xiaodong Gu, Weihao Yuan, Luwei Yang, Zilong Dong, and Ping Tan. Dense rgb slam with neural implicit maps. In Proceedings of the International Conference on Learning Representations, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "type": "text", + "content": "[31] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[32] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 559 + ], + "type": "text", + "content": "[33] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "text", + "content": "[34] Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and Andrew J Davison. Gaussian splatting slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18039-18048, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[37] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An opensource slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[39] Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision, pages 2320-2327. IEEE, 2011. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 263 + ], + "type": "text", + "content": "[40] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 331 + ], + "type": "text", + "content": "[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "text", + "content": "[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3437-3444. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[43] Szymon Rusinkiewicz and Marc Levoy. Efficient variants of the icp algorithm. In Proceedings third international conference on 3-D digital imaging and modeling, pages 145-152. IEEE, 2001. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[44] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[45] Erik Sandström, Kevin Ta, Luc Van Gool, and Martin R Oswald. Uncle-slam: Uncertainty learning for dense neural slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4537-4548, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "text", + "content": "[46] Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians. arXiv preprint arXiv:2405.16544, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "type": "text", + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "type": "text", + "content": "[48] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16660" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "resolution images and multi-camera videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3260-3269, 2017. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "text", + "content": "[50] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[51] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[52] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[53] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 408 + ], + "type": "text", + "content": "[54] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[55] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6229-6238, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[56] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[57] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R Oswald, and Matteo Poggi. How nerfs and 3d gaussian splatting are reshaping slam: a survey. arXiv preprint arXiv:2402.13255, 4, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[58] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[59] Jeffrey S Vitter. Random sampling with a reservoir. ACM Transactions on Mathematical Software (TOMS), 11(1):37-57, 1985. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[60] Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Itermvs: Iterative probability estimation for efficient multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8606-8615, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[61] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[62] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 173 + ], + "type": "text", + "content": "[63] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. arXiv preprint arXiv:2410.19115, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 176, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 553, + 230 + ], + "type": "text", + "content": "[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 233, + 553, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 233, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 316, + 233, + 553, + 288 + ], + "type": "text", + "content": "[65] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 290, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 290, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 553, + 322 + ], + "type": "text", + "content": "[66] Changchang Wu. Visualsfm: A visual structure from motion system. http://www.cs.washington.edu/homes/ccwu/vsfm, 2011. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 325, + 553, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 553, + 379 + ], + "type": "text", + "content": "[67] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 382, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 426 + ], + "type": "text", + "content": "[68] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 428, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 483 + ], + "type": "text", + "content": "[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 485, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 485, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 485, + 553, + 529 + ], + "type": "text", + "content": "[70] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 531, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 575 + ], + "type": "text", + "content": "[71] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "text", + "content": "[72] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "text", + "content": "[73] Ganlin Zhang, Erik Sandström, Youmin Zhang, Manthan Patel, Luc Van Gool, and Martin R Oswald. Glorie-slam: Globally optimized rgb-only implicit encoding point cloud slam. arXiv preprint arXiv:2403.19549, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[74] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16661" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 339 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[75] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "type": "text", + "content": "[76] Huizhong Zhou, Benjamin Ummenhofer, and Thomas Brox. Deeptam: Deep tracking and mapping. In Proceedings of the European conference on computer vision (ECCV), pages 822-838, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[77] Heng Zhou, Zhetao Guo, Shuhong Liu, Lechen Zhang, Qihao Wang, Yuxiang Ren, and Mingrui Li. Mod-slam: Monocular dense mapping for unbounded 3d scene reconstruction. arXiv preprint arXiv:2402.03762, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "text", + "content": "[78] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "text", + "content": "[79] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16662" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_content_list.json b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a40a3c15ac4379f04a54e41f6fe16afc4415b9a7 --- /dev/null +++ b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_content_list.json @@ -0,0 +1,1669 @@ +[ + { + "type": "text", + "text": "SLVR: Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ning Ni", + "bbox": [ + 339, + 203, + 408, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China", + "bbox": [ + 169, + 220, + 828, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ningni@mail.bnu.edu.cn", + "bbox": [ + 284, + 242, + 480, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Libao Zhang*", + "bbox": [ + 552, + 204, + 663, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "libaozhang@bnu.edu.cn", + "bbox": [ + 519, + 242, + 707, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, improving the residual structure and designing efficient convolutions have become important branches of lightweight visual reconstruction model design. We have observed that the feature addition mode (FAM) in existing residual structure tends to lead to slow feature learning or stagnation in feature evolution, a phenomenon we define as network inertia. In addition, although blueprint separable convolutions (BSConv) have proved the dominance of intrakernel correlation, BSCov forces the blueprint to perform scale transformation on all channels, which may lead to incorrect intra-kernel correlation and introduce useless or disruptive features on some channels and hinder the effective propagation of features. Therefore, in this paper, we rethink the FAM and BSCov for super-light visual reconstruction framework design. First, we design a novel linking mode, called feature diversity evolution link (FDEL), which aims to alleviate the phenomenon of network inertia by reducing the retention of previous low-level features, thereby promoting the evolution of feature diversity. Second, we propose blueprint controllable convolutions (B2Conv). The B2Conv can adaptively pick accurate intra-kernel correlation in the depth-axis, effectively preventing the introduction of useless or disruptive features. Based on FDEL and B2Conv, we develop a super-light super-resolution (SR) framework SLVR for visual reconstruction. Both FDEL and B2Conv can serve as efficient plugins. Extensive experimental results demonstrate the effectiveness of our proposed B2Conv, FDEL, and SLVR. Our code will be available at https://github.com/chongningni/SLVR.", + "bbox": [ + 88, + 324, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an important direction in the field of computer vision, visual reconstruction focuses on improving the visual quality of images or videos. Image super-resolution (SR) is", + "bbox": [ + 89, + 815, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5aa3a29a6fe17dccaefc5355ba82a92b438b9f54bb8bcbbacf0861d6dfb2fe39.jpg", + "image_caption": [ + "Figure 1. Performance and model complexity comparison on Set14 dataset for upscaling factor $\\times 4$ . * denotes the model trained on DF2K dataset and others trained on DIV2K." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a core task in visual reconstruction, which aims to reconstruct high-resolution details from a low-resolution image. SR is an ill-posed problem because a low-resolution (LR) image can be mapped to multiple high-resolution (HR) spaces, which makes it challenging. Recently, convolutional neural networks (CNNs) based on residual learning [1, 9, 21, 22, 30, 56] have dominated the SR field because they can constrain mapping space and reduce the difficulty of SR learning, which creates conditions for lightweight SR design. Furthermore, lightweight convolution has become an important research branch of lightweight model design. In this paper, we aim to rethink the residual structure and blueprint separable convolutions (BSConv) to build a superlight SR reconstruction framework. The issues we discovered are as follows:", + "bbox": [ + 511, + 580, + 906, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1) Existing feature addition mode easily causes network inertia: SR tasks require recovering detailed textures, which require rich high-level features as support. The existing residual structure uses the feature addition mode (FAM) of input and output $(x + \\mathcal{F}(x))$ , retaining many previous low-level features $(x)$ , which is not conducive to the diver", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. This work is supported by the National Natural Science Foundation of China under Grant 62271060 and 61571050.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "400", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sity evolution of features and easily causes network inertia. $\\mathcal{F}(\\cdot)$ represents the calculation function. Furthermore, retaining redundant previous low-level features will also affect the presentation of high-level features.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) Incorrect intra-kernel correlation of the BSConv: Although BSConv proves the dominance of intra-kernel correlation, we find that kernel correlation is not present on all channels in an ordinary convolution. However, BSConv forces the blueprint to perform linear transformations on all channels, which may lead to incorrect intra-kernel correlation and result in introducing many useless or disruptive features on some channels to hinder the effective propagation of features.", + "bbox": [ + 88, + 155, + 480, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above problems, this work rethinks the residual structure and BSconv. First, to alleviate the network inertia, we propose a new linking mode, named feature diversity evolution link (FDEL). Our FDEL aims to minimize the proportion of low-level features during feature propagation and presentation, which is beneficial to distill rich high-level features with different representations and alleviate network inertia. In addition, to maximize the effect of FDEL, we introduce a diversified feature aggregation strategy (DFAS) to enhance feature representation. Second, to prevent the incorrect inter-kernel correlation of the BSConv (inter-blueprint correlation), we propose new blueprint controllable convolutions (B2Conv). Our B2Conv sets up an access control unit (ACU) in the channel dimension to prevent incorrect intra-blueprint correlation. The ACU is like a guard, which can adaptively pick relevant channels to associate and learn the correct interblueprint correlation. Our B2Conv can reduce the introduction of useless or disruptive features and promote the effective propagation of feature information.", + "bbox": [ + 91, + 294, + 483, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are summarized as:", + "bbox": [ + 109, + 601, + 480, + 616 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose a new convolution B2Conv. Our B2Conv introduces the ACU to adaptively control intra-blueprint correlation, which can precisely prevent the introduction of useless or disruptive features and promote effective propagation of feature information. Besides, the B2Conv can also serve as a highly efficient building block for CNNs.", + "2) We propose a new linking mode FDEL, which effectively alleviates the network inertia problem. Our FDEL aims to reduce the retention of previous low-level feature information and can effectively promote the diverse evolution of features, which is beneficial to the extraction and presentation of high-level features.", + "3) Based on FDEL and B2Conv, we design a novel lightweight SR framework SLVR. Extensive experiments demonstrate that SLVR achieves record high performance (e.g., 28.85dB on Set14 × 4 with only 485K parameters) on lightweight SR benchmarks (shown in Fig. 1)." + ], + "bbox": [ + 89, + 619, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Residual Network: Inspired by [47], He et al. [13] proposed first residual network for image recognition and achieved state-of-the-art performance. Then, many different residual networks were proposed [8, 14, 36, 42, 46] for different computer vision tasks. Residual networks are typically built by stacking a sequence of residual blocks. Recently, the residual block was further developed, such as widening the channels [60], introducing attention mechanism [17, 55], and combining with dense connections [64]. The residual network has achieved great success in heavy-weight networks because it overcomes the vanishing gradient problem. Besides, residual learning can reduce the learning difficulty of computer vision tasks, which also provides opportunities for lightweight network design. To reduce the model complexity, MobileNet [16] introduced depthwise separable convolutions as key elements in the residual blocks. Then, MobileNetV2 [43] proposed an inverted residual block where the shortcut connections are between the thin bottleneck layers. ShuffleNetV2 [38] inserts a channel split component before the inverted residual block. After that, Zhou et al. [67] rethought the inverted residual block and proposed a sandglass block, which aims to perform spatial transformation and identity mapping at higher dimensions to alleviate information loss and gradient confusion problems. Although previous works have effectively improved the residual structure, most proposed residual blocks now still use the addition mode of input and output features.", + "bbox": [ + 511, + 114, + 903, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient Convolution: To reduce network parameters, AlexNet [25] first proposed the concept of group convolution and ResNext [57] demonstrated its effectiveness. After that, the group convolution was further modified by this works [18, 52, 62, 63]. VGG [45] suggested using two $3 \\times 3$ convolutions instead of one $5 \\times 5$ convolution. Xception [5] proposed the depthwise separable convolution (DSConv) to further reduce the parameters of convolution and then it was used in MobileNet [16] and gained state-of-the-art results. Partial convolution (PConv) [33] was first proposed for image inpainting, which greatly improved computing efficiency. To solve the unreasonable problem of mask update in the partial convolution, Gated convolution (GConv) [59] was proposed. Octave convolution [3] was proposed to extract the high-frequency and low-frequency signals of the image. Recently, DSConv was decomposed [41] into two components: variable quantization kernels and distribution offsets, thereby achieving memory reduction and operation acceleration. BSConv [12] rethought the DSConv and considered the dominance of intra-kernel correlations when performing convolution operations.", + "bbox": [ + 511, + 537, + 903, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lightweight SR: Although deep models can achieve good SR performance, they are not conducive to deployment on small devices. Therefore, many works were proposed", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "401", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d7761eaccf626bc2de487665c109c6ad2c175055bbf29730ba6b23092fa0676b.jpg", + "image_caption": [ + "Figure 2. Visualization of intermediate features, above the dotted line is the pipeline of BSRN [30] and green cuboids represent the ESDB modules. Outi represents the output of the $i^{th}$ ESDB module. Outi = xi = xi-1 + yi." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[6, 20, 44] to reduce the computational complexity. LapSRN [26] removed the bicubic interpolation pre-processing step. CARN [1] introduced group convolution and cascading residual blocks. IMDN [21] used channel splitting operations to complete information multi-distillation and selective feature fusion. After that, improving information distillation structure has become an important branch of lightweight SR network design, and many methods [9, 30, 34] were proposed. Then, PAN [66] introduced pixel attention to measure pixel contribution to achieve competitive performance by using fewer parameters. Recently, a series of transformer-based SR methods [4, 7, 11, 29, 31, 35, 48, 51] emerge with superior performance. HNCT [7] cleverly combined CNN and transformer to reduce the parameters while ensuring network performance. CFIN [29] design a novel cross-receptive field-guided transformer to adaptively select the required contextual information. NGswin [4] introduced the N-Gram context to the low-level vision with transformers and achieve superior performance.", + "bbox": [ + 91, + 356, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 671, + 225, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 696, + 212, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we define the phenomenon of network inertia as a situation where the input and output of a network module are similar, indicating that the module fails to learn meaningful features. As shown in Fig. 2, we find that the FAM $(x_{i - 1} + y_i)$ easily causes network inertia because it retains many previous input low-level features $(x_{i - 1})$ information. There are 8 ESDBs with this addition mode in BSRN, we can find that the outputs of most adjacent ESDB modules are similar, especially Out2 and Out3, Out4 and Out5. In addition, we replaced Out5 with Out4 in the subsequent feature fusion and were surprised to find that this replacement had almost no impact on the image reconstruction quality,", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dec7bb17eaaad59bff1be289a055df82da563247f55faa124b69bedb05098fa3.jpg", + "image_caption": [ + "Figure 3. Exemplary filter weights of ordinary convolution and BSConv. One filter kernel of ordinary convolution size $M \\times K \\times K = 64 \\times 3 \\times 3$ is visualized and split into 64 images. Filter kernels of pointwise convolution size $N \\times M \\times K \\times K = 64 \\times 64 \\times 1 \\times 1$ and two channels are visualized." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the PSNR only decreased by less than $0.005\\mathrm{dB}$ . The similarity between features Out4 and Out5 can be inferred that the features $y_{4}$ extracted by the $4^{th}$ ESDB module are meaningless and network inertia occurs.", + "bbox": [ + 511, + 282, + 903, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Second, as shown in the left of Fig. 3, although the weights of each kernel are highly correlated along the channel dimension, differences also exist. Besides, BSConv [12] also proved that not all the slices (52/128 for the ResNet-50 example) of the kernel are highly correlated along the depth axis. However, as shown in the right of Fig. 3, we find that the pointwise convolution of BSConv associates almost all the channel dimensions along the depth axis (the darker the color, the higher the correlation), which results in introducing many useless or disruptive features on some channels and is not conducive to effective extraction and propagation of useful features.", + "bbox": [ + 511, + 342, + 906, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Entropy Delta Extraction Link", + "text_level": 1, + "bbox": [ + 511, + 532, + 785, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will explain the definition of FDEL in detail.", + "bbox": [ + 511, + 555, + 905, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the network with stacked RBs, as shown in Fig. 5 (a), the initial shallow (low-level) features $x_{1}$ will be superimposed and retained until the last layer $x_{L}$ , as follows", + "bbox": [ + 511, + 585, + 905, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {2} = \\mathcal {F} _ {1} (x _ {1}) + x _ {1}, \\dots , \\\\ x _ {l} = \\mathcal {F} _ {l - 1} \\left(x _ {l - 1}\\right) + x _ {l - 1}, \\dots , \\\\ x _ {L} = \\mathcal {F} _ {L - 1} \\left(x _ {L - 1}\\right) + x _ {L - 1}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 648, + 903, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_l$ represents the operations of the $l^{th}$ RB. Equation (1) can be regarded as a feature superposition operation. It may let the feature representation $x_{l}$ of $l^{th}$ RB rely too much on the input features $x_{l-1}$ , which will limit the learning direction of features and even lead to stagnation in feature evolution. This phenomenon can be evidenced in Fig. 2. In addition, the superposition of many low-level input features is not conducive to the presentation of high-level features, especially in structure Fig. 5 (c).", + "bbox": [ + 511, + 719, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In view of the aforementioned limitations of the FAM of the RBs, we rethink its design principles and propose an FDEL that can tackle the above issues by flipping the FAM,", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "402", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d848d73b83085bfa83047f528a3e6d3dced69e420931370cf8ce81b0da3b87ef.jpg", + "image_caption": [ + "Shallow Feature Extraction" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/897bedc20e7af603ed40c95437972f73aeed8e8d712f5b86969da3017ecaf5f5.jpg", + "image_caption": [ + "Deep Feature Extraction" + ], + "image_footnote": [], + "bbox": [ + 93, + 210, + 584, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df2c9f5c11a7f62af36792682f8a3c516e8c72a278e104224bb021981ed55064.jpg", + "image_caption": [ + "Figure 4. The overall architecture of the proposed SLVR and the structure of BCRB and blueprint controllable convolutions (B2Conv)." + ], + "image_footnote": [], + "bbox": [ + 602, + 210, + 728, + 391 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca6940c310b64355605ecc64992b8d961aa54c0599363c750b3a30b692fbc39d.jpg", + "image_caption": [ + "Image Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 746, + 210, + 903, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the structure of RB with FDEL is shown in Fig. 5 (b). Our FDEL aims to remove excessive low-level features. The specific mathematical formula of FDEL is", + "bbox": [ + 89, + 436, + 483, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {l + 1} = \\mathcal {F} _ {\\mathrm {F D E L}} \\left(x _ {l}, \\mathcal {F} _ {l} \\left(x _ {l}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 494, + 480, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{FDEL}}$ denotes the operations of the FDEL. Our FDEL uses the output of the $\\mathcal{F}_l$ minus its input. Because the output feature representation is more high-level than the input, each RB can effectively extract the different high-level features, which promotes the diverse evolution of features. As we all know, we usually refer to the output of a function minus its input as the gain of the function in a physical sense. Therefore, in an RB with FDEL, the gain is the high-level features extracted by this RB.", + "bbox": [ + 89, + 521, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Blueprint Controllable Convolutions", + "text_level": 1, + "bbox": [ + 89, + 665, + 408, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although BSConv [12] demonstrates the importance of intra-kernel correlations in separating ordinary convolution, it does not consider the differences of the filters in the depth-axis. In other words, in some channel dimensions, this blueprint cannot obtain the original filter of the ordinary convolution through scale transformation, so filters after forced scale transformation don't effectively work. Therefore, we sparse the intra-kernel correlation and proposed a new B2Conv. Our B2Conv can adaptively pick appropriate channels along the depth axis to perform blueprint scale transformation, which can promote the extraction and efficient propagation of feature information.", + "bbox": [ + 89, + 688, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a convolution operation, assume that the input is $U\\in \\mathbb{R}^{M\\times H\\times W}$ , the output is $V\\in \\mathbb{R}^{N\\times H\\times W}$ , and the con", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "volution kernels are $\\{E^{(1)}, E^{(2)}, \\dots, E^{(N)}\\}$ and $E^{(n)} \\in \\mathbb{R}^{M \\times H \\times W}$ . Then the process can be expressed as", + "bbox": [ + 511, + 436, + 906, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n, \\therefore :} = U * E ^ {(n)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 477, + 903, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $*$ denotes the convolution operation and $n \\in \\{1,2,\\dots,N\\}$ , the weights of these $N$ kernels can be learned during the training. BSCov proved that in practice these weights usually converge to a state that shows a substantial amount of correlation. Therefore, BSCov proposes to decompose an ordinary convolution into the interaction of a blueprint $\\mathcal{B}$ and a set of scale-transforming weights $\\{w_{n,1}, w_{n,2}, \\dots, w_{n,M}\\}$ , as follows", + "bbox": [ + 511, + 506, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {m,:,:} ^ {(n)} = w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 637, + 903, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m \\in \\{1, 2, \\dots, M\\}$ and $n \\in \\{1, 2, \\dots, N\\}$ (see Fig. 6, left column). However, in ordinary convolution, not all the filters of channel dimension can be translated by Equation (4). Therefore, to solve the problem, we propose to add an access control unit (ACU) before scale-transforming. Specifically, in all channel dimensions, if there is a strong intra-kernel correlation with $\\mathcal{B}$ , we perform Equation (4), otherwise we do not perform any operation or multiply Equation (4) by a small correlation coefficient to ensure that the useless features extracted in this dimension will not affect the efficient propagation of features, as follows", + "bbox": [ + 511, + 666, + 906, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {m,:,:} ^ {(n)} = \\mathcal {T} _ {n, m} \\cdot w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 840, + 903, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{T}_{n,m}$ denotes the output of the ACU, ideally its value is 0 or 1. 1 means allowing scale-transforming and 0 means", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "403", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eafb0e1ff1362b418cfe31a692e23fea81d0a32eeaff477aeef25ea7312bda59.jpg", + "image_caption": [ + "Figure 5. Comparison of residual blocks (RBs) with addition mode and FDEL. For the specific mathematical formula of FDEL, refer to Equation (2)." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 279, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d8c83baf08f6cd185f067ef4ef29325385f94c2c00731d6a4be8b20a7422229b.jpg", + "image_caption": [ + "Figure 6. Interpretations for BSConv and B2Conv. B2Conv can adaptively pick correct intra-kernel correlations." + ], + "image_footnote": [], + "bbox": [ + 295, + 87, + 482, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the opposite Here, to make the function continuous and differentiable, we soften the output of the ACU and limit its value range to $(0,1)$ . The closer the $\\mathcal{T}_{n,m}$ is to 1, the stronger the intra-kernel correlation. The characteristics of features can reflect the characteristics of the convolution kernels. Therefore, $\\mathcal{T}_{n,m}$ can be obtained in an embedding space by feature space mapping and weights of the $\\mathcal{T}_{n,m}$ and $w_{n,m}$ can be learned directly via backpropagation. Specifically, we first rewrite the Equation (3) by split 3d convolution into a set of $M$ 2d convolutions, yielding", + "bbox": [ + 89, + 299, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n,:,:} = U * E ^ {(n)} = \\sum_ {m = 1} ^ {M} \\left(U _ {m,:,:} * E _ {m,:,:} ^ {(n)}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 460, + 482, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Secondly, we replace the ordinary convolution $E_{m,\\cdot,\\cdot}^{(n)}$ with the B2Conv representation as given in Equation (5), as follows", + "bbox": [ + 89, + 513, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n, \\therefore :} = \\sum_ {m = 1} ^ {M} \\left(U _ {m, \\therefore :} * \\left(\\mathcal {T} _ {n, m} \\cdot w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 570, + 482, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because each blueprint $\\mathcal{B}^{(n)}$ is independent in the channel dimension. So, we can rewrite the Equation (7) (Note: As shown in right column of Fig. 6, the input $U_{m,:;}$ also needs to be copied $(\\bar{U}_{m,:;})$ for feature space mapping)", + "bbox": [ + 89, + 622, + 483, + 684 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n,:,:} = \\left(\\underbrace {\\sum_ {m = 1} ^ {M} \\left(\\bar {U} _ {m , :,:} \\cdot \\mathcal {T} _ {n , m}\\right)} _ {\\bar {V} _ {n,:,:} ^ {\\prime}} \\cdot \\underbrace {\\sum_ {m = 1} ^ {M} \\left(U _ {m , :,:} \\cdot w _ {n , m}\\right)} _ {V ^ {\\prime} _ {n,:,:}}\\right) * \\mathcal {B} ^ {(n)}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 694, + 488, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we further rearrange $\\{w_{n,1}, w_{n,2}, \\dots, w_{n,M}\\}$ and $\\{\\mathcal{T}_{n,1}, \\mathcal{T}_{n,2}, \\dots, \\mathcal{T}_{n,M}\\}$ into a $M \\times 1 \\times 1$ arrays $\\tilde{w}_n$ and $\\tilde{\\mathcal{T}}_n$ respectively, the sum operations of Equation (8) can be replaced by convolution operations, as follows", + "bbox": [ + 89, + 795, + 483, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {V} _ {n,:,:} ^ {\\prime} = U * \\tilde {\\mathcal {T}} _ {n}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 866, + 482, + 886 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n,:,:} ^ {\\prime} = U * \\tilde {w} _ {n}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 887, + 480, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3764e97daef199ebcdf6f58ead41ed5c9ad2ecc015e71c42e31d326f3f2e30d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 89, + 903, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n,:,:} = \\left(\\sigma \\left(\\bar {V} _ {n,:,:} ^ {\\prime}\\right) \\cdot V _ {n,:,:} ^ {\\prime}\\right) * \\mathcal {B} ^ {(n)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 280, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the sigmoid function, which limits $\\tilde{V}_{n,::}^{\\prime}$ to (0,1). For a concrete implementation, Equation (9) and Equation (10) can be translated into a $1\\times 1$ pointwise convolution with the kernels $\\{\\tilde{w}_1,\\tilde{w}_2,\\dots ,\\tilde{w}_N\\}$ and $\\left\\{\\tilde{\\mathcal{T}}_1,\\tilde{\\mathcal{T}}_2,\\dots ,\\tilde{\\mathcal{T}}_N\\right\\}$ respectively. The Equation (11) can be translated into a $K\\times K$ depthwise convolution with the kernels $\\{\\mathcal{B}^{(1)},\\mathcal{B}^{(2)},\\dots ,\\mathcal{B}^{(N)}\\}$ .", + "bbox": [ + 511, + 304, + 906, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. The Overall Structure", + "text_level": 1, + "bbox": [ + 513, + 425, + 720, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on FDEL and B2Conv, we further develop a lightweight SLVR framework to achieve high-performance image SR. As shown in Fig. 4, SLVR consists of three parts, including shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given a LR input $I_{\\mathrm{LR}} \\in \\mathbb{R}^{H \\times W \\times C_{in}}$ , we first use a B2Conv $\\mathcal{H}_{\\mathrm{B2Conv}}$ to extract shallow feature $X_0 \\in \\mathbb{R}^{H \\times W \\times C_{mid}}$ , as follows", + "bbox": [ + 511, + 449, + 905, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {0} = \\mathcal {H} _ {\\mathrm {B 2 C o n v}} \\left(\\operatorname {C o n c a t} \\left[ I _ {\\mathrm {L R}} \\right] _ {n}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 563, + 903, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_{in}$ and $C_{mid}$ denote the channel number of the input and shallow feature. $\\text{Concat}[\\cdot]_n$ means concatenating $n$ times input. Then we use $K$ efficient controllable distillation block (ECDB) to extract deep features. To maximize the effect of FDEL, feature aggregation is performed by DFAS and GELU [15] activation to enhance feature representation. After that, we use a B2Conv to continue refining the features, the above process can be formulated as", + "bbox": [ + 511, + 588, + 905, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} = \\mathcal {H} _ {\\mathrm {E C D B} _ {i}} \\left(X _ {i - 1}\\right), i = 1, 2, \\dots , K, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 718, + 903, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\mathrm {A g g}} = \\mathcal {H} _ {\\mathrm {D F A S}} \\left(\\left(w _ {1}, w _ {2}, \\dots , w _ {c}\\right), \\left(X _ {1}, X _ {2}, \\dots , X _ {K}\\right)\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 743, + 903, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {R e f i n e d}} = \\mathcal {H} _ {\\mathrm {B 2 C o n v}} \\left(\\operatorname {G E L U} \\left(X _ {\\mathrm {A g g}}\\right)\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 773, + 903, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_{\\mathrm{ECDB}_i}$ represents the $i^{\\mathrm{th}}$ ECDB, $w_{1}, w_{2}, \\dots, w_{c}$ are the learnable weights, $X_{1}, X_{2}, \\dots, X_{K}$ denote the intermediate features. $\\mathcal{H}_{\\mathrm{DFAS}}(\\cdot)$ denotes the operations of the DFAS, which can be achieved by a convolution. The ECDB is designed based on the ESDB of the BSRN. Specifically, first, we replace the BSRB with the blueprint controllable residual block (BCRB). Second, we replace the", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "404", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4929a7d4ca76336e789905185d20047cb403d4db8c4ced37c7a96ade36dfbd85.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodT-DatasetScale#ParamsMult-AddsSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
VDSR [22]291×2666K612.6G37.530.958733.030.912431.900.896030.760.914037.220.9750
CARN [1]D2K+291×21,592K222.8G37.760.959033.520.916632.090.897831.920.925638.360.9765
IMDN [21]D2K×2694K158.8G38.000.960533.630.917732.190.899632.170.928338.880.9774
LAPAR-A [27]DF2K×2548K171G38.010.960533.620.918332.190.899932.100.928338.670.9772
LatticeNet [37]D2K×2756K169.5G38.060.960733.700.918732.200.899932.250.928838.940.9774
RLFN [24]D2K×2527K97.9G38.070.960733.720.918732.220.900032.330.9299--
SRPN-Lite [65]DF2K×2609K139.9G38.100.960833.700.918932.250.900532.260.9294--
FMEN [6]DF2K×2748K172.0G38.100.960933.750.919232.260.900732.410.931138.950.9778
MLRN [10]D2K×2488k90.4G38.070.960733.590.918032.210.900032.280.929738.760.9773
NGswin [4]D2K×2998K140.4G38.050.961033.790.919932.270.900832.530.932438.970.9777
SLVR (Ours)D2K×2464K102G38.140.961033.870.920232.280.900932.570.931839.080.9776
VDSR [22]291×3666K612.6G33.660.921329.770.831428.820.797627.140.827932.010.9340
CARN [1]D2K+291×31,592K118.8G34.290.925530.290.840729.060.803428.060.849333.500.9440
IMDN [21]D2K×3703K71.5G34.360.927030.320.841729.090.804628.170.851933.610.9445
LAPAR-A [27]DF2K×3594K114G34.360.926730.340.842129.110.805428.150.852333.510.9441
LatticeNet [37]D2K×3765K76.3G34.400.927230.320.841629.100.804928.190.851333.630.9442
RFDN [34]D2K×3633K65.6G34.470.928030.350.842129.110.805328.320.854733.780.9458
SRPN-Lite [65]DF2K×3615K62.7G34.470.927630.380.842529.160.806128.220.8534--
FMEN [6]DF2K×3757K77.2G34.450.927530.400.843529.170.806328.330.856233.860.9462
MLRN [10]D2K×3496K40.9G34.460.926730.350.842629.100.805428.200.853333.660.9450
NGswin [4]D2K×31,007K66.6G34.520.928230.530.845629.190.807828.520.860333.890.9470
SLVR (Ours)D2K×3473K45.3G34.580.928530.530.846029.210.808028.560.860334.070.9475
VDSR [22]291×4666K612.6G31.350.883828.010.767427.290.725125.180.752428.830.8870
CARN [1]D2K+291×41,592K90.9G32.130.893728.600.780627.580.734926.070.783730.470.9084
IMDN [21]D2K×4715K40.9G32.210.894828.580.781127.560.735326.040.783830.450.9075
LAPAR-A [27]DF2K×4659K94G32.150.894428.610.781827.610.736626.140.787130.420.9074
LatticeNet [37]D2K×4777K43.6G32.180.894328.610.781227.570.735526.140.784430.540.9075
RFDN [34]D2K×4550K23.9G32.240.895228.610.781927.570.736026.110.785830.580.9089
RLFN [24]D2K×4543K25.3G32.240.895228.620.781327.600.736426.170.7877--
SRPN-Lite [65]DF2K×4623K35.8G32.240.895828.690.783627.630.737326.160.7875--
BSRN [30]DF2K×4352K19.4G32.350.896628.730.784727.650.738726.270.790830.840.9123
FMEN [6]DF2K×4769K44.2G32.240.895528.700.783927.630.737926.280.790830.700.9107
MLRN [10]D2K×4507K23.5G32.300.895628.620.782427.570.736526.100.786730.560.9092
HPUN-M [48]D2K×4511K27.7G32.240.895028.660.782827.600.737126.120.787830.550.9089
SAFMN [49]DF2K×4240K14.0G32.180.894828.600.781327.580.735925.970.780930.430.9063
OSFFNet [53]DF2K×4537K22.0G32.390.897628.750.785227.660.739326.360.795030.840.9125
NGswin [4]D2K×41,019K36.4G32.330.896328.780.785927.660.739626.450.796330.800.9128
SLVR-S (Ours)DF2K×4158K8.5G32.200.895228.650.782227.570.736126.080.785430.530.9084
SLVR (Ours)D2K×4485K26.2G32.400.897528.780.785027.650.738826.410.795130.930.9133
SLVR* (Ours)DF2K×4485K26.2G32.440.898028.850.786527.700.740026.580.800031.100.9141
", + "bbox": [ + 119, + 87, + 883, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison (PSNR/SSIM) for lightweight SR with state-of-the-art (SOTA) methods on five benchmark datasets. The best and second-best results are marked in red and blue colors respectively. Mult-Adds is evaluated on a ${1280} \\times {720}\\mathrm{{HR}}$ image. T-Dataset denotes training dataset. D2K stands for the DIV2K dataset. DF2K indicates a merged dataset containing ${800} + 2,{650}\\mathrm{{HR}} - \\mathrm{{LR}}$ image pairs. 291 images dataset is from [58].", + "bbox": [ + 89, + 594, + 906, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "addition mode of the residual structure with the proposed FDEL. Third, all the BSConvs are replaced with the proposed B2Conv. Please see Fig. 4 for detailed changes. After performing the above operations, we obtain the final refined features $X_{\\mathrm{Refined}}$ . Finally, to take advantage of residual learning, we use long skip connections to reduce the difficulty of learning. SR results can be obtained by", + "bbox": [ + 89, + 660, + 482, + 767 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {S R}} = \\mathcal {H} _ {\\text {U p s a m p l e}} \\left(X _ {\\text {R e f i n e d}}\\right) + \\text {B i c u b i c} ^ {\\uparrow} \\left(I _ {\\mathrm {L R}}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 772, + 482, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_{\\mathrm{Upsample}}(\\cdot)$ represents the upsampling operation, which is implemented with the pixelshuffle operation [44]. Bicubic denotes bicubic upsampling. $L_{1}$ loss function is used to optimize the network, as follows", + "bbox": [ + 89, + 795, + 483, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} = \\left\\| I _ {\\mathrm {H R}} - I _ {\\mathrm {S R}} \\right\\| _ {1}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 863, + 482, + 880 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{HR}}$ is the HR label.", + "bbox": [ + 89, + 885, + 269, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 659, + 645, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 686, + 702, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. We use DIV2K [50] dataset as the training dataset, which contains 800 HR-LR image pairs. To show more fair results, we used DF2K (DIV2K + Flickr2K [50]) to additionally train an $\\times 4$ model. Note that the model trained with DF2K is marked with a small *. For testing, we evaluated the performances on the five benchmark datasets, including Set5 [2], Set14 [61], B100 [39], Urban100 [19], Manga109 [40]. The PSNR and SSIM [54] are used as the evaluation metrics, which are evaluated on the Y channel of the YCbCr space.", + "bbox": [ + 511, + 714, + 906, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. ECDB number is set to 8, and the channel number is set to 64. The kernel size of the depth-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "405", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e449684b2a1efe1ed3fad976449cfc9a9417a2e3f50d3b6fa33280586af865d4.jpg", + "image_caption": [ + "Figure 7. Visual comparison for $\\times 4$ SR methods. The patches for comparison are marked with red boxes." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "wise convolutions is set to 3. We additionally built a smaller model SLVR-S to compare the latest methods with less than 200K parameters and it trained on DF2K. The ECDB number of the SLVR-S is set to 6 and the channel number is set to 40. During training, we augment the data with random horizontal flips and rotation $(90^{\\mathrm{o}},180^{\\mathrm{o}},270^{\\mathrm{o}})$ . We randomly crop LR patches of size $64\\times 64$ as input. Our models are trained by Adam optimizer [23] for 1000K iterations. The initial learning rate is set to $1\\times 10^{-3}$ and halved for every 200k iterations. The batch size of SLVR and SLVR-S is set to 64 and 128 respectively. LR image patches are obtained by the bicubic kernel from corresponding HR images matching each SR task. Our networks are trained on an RTX3090 GPU. Note that no other training skills (e.g., pre-training [28]) are employed.", + "bbox": [ + 88, + 441, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with the SOTA SR methods", + "text_level": 1, + "bbox": [ + 89, + 680, + 437, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of SLVR, we compare our model with other efficient SR models on scale factors $\\times 2,\\times 3,\\times 4$ including VDSR (CVPR16) [22], CARN (ECCV18) [1], IMDN (ACMMM19) [21], RFDN (ECCV20) [34], LAPAR-A (NeurIPS20) [27], LatticeNet (ECCV20) [37], RLFN (CVPRW22) [24], SRPN-Lite (ICLR22) [65], BSRN (CVPRW22) [30], FMEN (CVPRW22) [6], MLRN (CVPRW23) [10], NGswin (CVPR23) [4], HPUN-M (AAAI23) [48], SAFMN (ICCV23) [49], OSFFNet (AAAI24) [53].", + "bbox": [ + 89, + 703, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, we reported the training dataset, the parameters, Mult-Adds, and PSNR and SSIM on five benchmarks. Compared with models with similar param", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63bb4782ea05068379bb2f5b9913756429bfee69690d89b860e94e1416f25845.jpg", + "image_caption": [ + "Figure 8. Ablation studies on B2Conv. BSRN-B2Conv denotes the BSCnv of BSRN is replaced by B2Conv. All the results are reported on scale factor $\\times 4$" + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 705, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22385b29953ecde2fa7b60f12de3d0b1e1135ac5a0cb26069832c80b2b97bd5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 90, + 903, + 209 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf2de84d086295f28e90eba38e7b387ae287d51e57ab83a09e4d8f2dab895cc8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFDELB2ConvSet5Set14Urban100
Model-I✓×32.37/0.896928.71/0.783926.30/0.7917
Model-II×✓32.39/0.897528.76/0.785126.40/0.7943
SLVR✓✓32.40/0.897528.78/0.785026.41/0.7951
", + "bbox": [ + 517, + 279, + 910, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation studies of FDEL and B2Conv on scale factor $\\times 4$ . The $\\times$ of FDEL denotes it is replaced by addition mode, and the $\\times$ of B2Conv denotes it is replaced by BSConv.", + "bbox": [ + 511, + 349, + 906, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eters and Mult-Adds, the performance of our SLVR surpasses existing methods with a notable margin on all benchmarks. Compared with the latest transformer-based method NGswin, although the number of parameters of our SLVR is less than $50\\%$ of NGswin and the number of Mult-Adds is less than $75\\%$ of NGswin, our SLVR still has an advantage in performance. Compared with the latest extremely lightweight method SAFMN trained on the same dataset, our SLVR-S obtains better performance with a smaller model size.", + "bbox": [ + 511, + 407, + 906, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also perform a visual comparison of different SOTA lightweight SR methods on scale factor $\\times 4$ . As shown in Fig. 7, we can observe that the SR results generated by SLVR contain more correct and fine-grained textures, while other methods generate many blurred edges and artifacts. For example, in the second row of Fig. 7, our SLVR can restore the texture of the fence, while other methods cannot. We think that the visual results can prove the effectiveness of the proposed B2conv, which can prevent incorrect intrakernel correlations to reduce useless or disruptive feature extraction and propagation.", + "bbox": [ + 511, + 560, + 908, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 739, + 679, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The core contributions of our framework are the proposed FDEL and B2Conv. To demonstrate their effectiveness, we first conduct ablation studies on the main framework of our SLVR. Specifically, we remove FDEL and B2Conv and replace them with the original linking mode FAM of RBs and BSCnv, respectively. The experimental results are reported in Table 2. We can observe that our FDEL and B2Conv can improve network performance, especially B2Conv, improved by 0.11 dB on the dataset Urban100.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "406", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce38474fb0efe5508cf61ff6c5bd43b7878e6c5882d6b77c19dc7443d1e76d20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScale#ParamsMult-AddsSet5Set14Urban100
EDSR-baseline×41,518K114.0G32.08/0.893828.58/0.781025.98/0.7824
EDSR+FDEL1,518K114.0G32.12/0.894428.61/0.781926.05/0.7845
", + "bbox": [ + 94, + 89, + 482, + 127 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5280a7c8f79aac669a8156b5e36e4e45d62a8e05637f07a9ca93840a14ceb71b.jpg", + "table_caption": [ + "Table 3. Ablation studies of FDEL. EDSR+FDEL denotes that all addition mode of EDSR-baseline [32] are replaced by FDEL." + ], + "table_footnote": [], + "table_body": "
RFDNScale#ParamsMult-AddsSet5Set14Urban100
w BSConv×3197K25.8G34.37/0.926830.34/0.841828.06/0.8503
w B2Conv261K34.5G34.44/0.927330.36/0.842628.18/0.8525
w Conv633K65.6G34.47/0.928030.35/0.842128.21/0.8525
w BSConv×4210K14.9G32.12/0.894128.59/0.780626.01/0.7829
w B2Conv273K20.1G32.22/0.894928.65/0.782226.06/0.7843
w Conv643K37.4G32.28/0.895728.61/0.781826.11/0.7858
", + "bbox": [ + 94, + 176, + 480, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comprehensive comparison between ordinary convolution (Conv), BSCnv, and B2Conv. The best and second-best results are marked in red and blue colors.", + "bbox": [ + 89, + 267, + 482, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/506680e4a9f5f9174dc4d9063b855981d7cbfaaee447f4420066fadeadb79f66.jpg", + "image_caption": [ + "Figure 9. Visual comparison of feature maps between SLVR and SLVR without FDEL. Outi represents the output of the $i^{th}$ ECDB module." + ], + "image_footnote": [], + "bbox": [ + 94, + 320, + 480, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Second, to verify that FDEL can be used as a more efficient linking mode of RBs, we use FDEL to simply replace the traditional linking mode FAM in the classic residual architecture EDSR [32]. From Table 3, we can conclude that our FDEL is more effective than the traditional FAM without additional parameters and Mult-Adds consumed.", + "bbox": [ + 89, + 445, + 482, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Third, to further verify the effectiveness of our proposed B2Conv and confirm that B2Conv can be used as highly efficient building blocks. We denote BSConv of BSRN [30] replaced by B2Conv as BSRN-B2Conv. As shown in Fig. 8, we can observe that our B2Conv can significantly improve model performance without changing model architecture. Besides, we also conducted a comprehensive comparison between ordinary convolution, BSConv, and B2Conv on the network architecture RFDN [34], as shown in Table 4, compared with BSConv, our B2Conv greatly improves the performance (e.g., $\\mathbf{0.1dB}$ on Set5 × 4) of the RFDN with few additional parameters (63K) consumed, which also illustrates the rationality of our improvement.", + "bbox": [ + 88, + 537, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we perform a visual comparison of features between SLVR and SLVR without FDEL. From Fig. 9, we find that our FDEL can effectively alleviate network inertia. In addition, we visualize the scale transformation state of 8 blueprints on 64 channels. As shown in Fig. 10, we can observe that our B2Conv does not perform scale transformation on all 64 channel dimensions (only on the channel denoted by black squares). It proves that our B2Conv can adaptively select the appropriate channel to perform blueprint scale transformation, which can effectively prevent the introduction of disruptive or useless features. Fur", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b0828daebd3fef597bf61b13377a90184306c2ed8c3cafd8aa442b65bac927d.jpg", + "image_caption": [ + "Figure 10. Visualization of scale transformation state of 4 blueprints on 64 channels. Black squares denote strong intrakernel correlation, and white squares are the opposite." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 609, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f7d271da91ed03df7e1cfeb0cf79ad79fa669ab51ea0a224c836e787a9d98c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 89, + 705, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e1680436f1d224796df161d49cbc9137cbbf193184342ab4293b6445d912df51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 89, + 805, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b954807791b6157a792563a0699127988680707459f18732e11e193109dfa09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 89, + 903, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/627c1e80b9ef302746cedc93293099dd4fa2bb0362c8df65d026d6fd404b71ff.jpg", + "image_caption": [ + "Figure 11. Visual comparison of intermediate feature maps between SLVR and SLVR without B2Conv." + ], + "image_footnote": [], + "bbox": [ + 516, + 219, + 903, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "thermore, we perform a visual comparison of features between SLVR and SLVR without B2Conv. From Fig. 11, we can find that the intermediate features generated by the SLVR without B2Conv contain more interference and disruptive features, resulting in poorer feature representation. The intermediate features generated by our SLVR are more accurate and have better feature representation ability. It also indirectly proves that the proposed B2Conv can adaptively select the correct intra-kernel correlation, prevent the introduction of disruptive features, and is more conducive to feature propagation and representation.", + "bbox": [ + 511, + 419, + 906, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 602, + 632, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose blueprint controllable convolutions (B2Conv). B2Conv can precisely prevent incorrect intra-kernel correlation, which effectively reduces the extraction of useless or disruptive features and promotes the effective propagation of feature information. Besides, we propose a novel linking mode FDEL for RBs. Our FDEL reduces the retention of the input low-level features and can promote the diversity evolution of high-level features, which can effectively alleviate the occurrence of network inertia. Finally, based on B2Conv and FDEL, we propose an efficient lightweight SR framework SLVR. Furthermore, our B2Conv and FDEL can also serve as highly efficient components to improve the performance of other CNN-based models (such as EDSR, BSRN, and RFDN). Extensive experiments on public benchmark datasets show the prominent SR performance of SLVR and comprehensive analytical studies validate the reasonableness and effectiveness of the proposed framework and components.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "407", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, pages 252-268, 2018. 1, 3, 6, 7", + "[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6", + "[3] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Yan Shuicheng, and Ji-ashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In ICCV, pages 3434–3443, 2019. 2", + "[4] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image superresolution. In CVPR, pages 2071–2081, 2023. 3, 6, 7", + "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 2", + "[6] Zongcai Du, Ding Liu, Jie Liu, Jie Tang, Gangshan Wu, and Lean Fu. Fast and memory-efficient network towards efficient image super-resolution. In CVPRW, pages 852-861, 2022. 3, 6, 7", + "[7] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In CVPRW, pages 1102-1111, 2022. 3", + "[8] Mengyang Feng, Huchuan Lu, and Yizhou Yu. Residual learning for salient object detection. IEEE TIP, 29:4696-4708, 2020. 2", + "[9] Guangwei Gao, Wenjie Li, Juncheng Li, Fei Wu, Huimin Lu, and Yi Yu. Feature distillation interaction weighting network for lightweight image super-resolution. In AAAI, pages 661-669, 2022. 1, 3", + "[10] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. Mixer-based local residual network for lightweight image super-resolution. In CVPRW, pages 1593-1602, 2023. 6, 7", + "[11] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. A simple transformer-style network for lightweight image super-resolution. In CVPRW, pages 1484-1494, 2023. 3", + "[12] Daniel Haase and Manuel Amthor. Rethinking depthwise separable convolutions: How intra-kernel correlations lead to improved mobilenets. In CVPR, pages 14588-14597, 2020. 2, 3, 4", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, pages 630-645. Springer, 2016. 2", + "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolu" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2", + "[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2", + "[18] Gao Huang, Shichen Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Condensenet: An efficient densenet using learned group convolutions. In CVPR, pages 2752-2761, 2018. 2", + "[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 6", + "[20] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, pages 723-731, 2018. 3", + "[21] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, pages 2024-2032, 2019. 1, 3, 6, 7", + "[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 1, 6, 7", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7", + "[24] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPRW, pages 765-775, 2022. 6, 7", + "[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 2", + "[26] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, pages 5835-5843, 2017. 3", + "[27] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. NeurIPS, 33:20343-20355, 2020. 6, 7", + "[28] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 7", + "[29] Wenjie Li, Juncheng Li, Guangwei Gao, Weihong Deng, Jiantao Zhou, Jian Yang, and Guo-Jun Qi. Cross-receptive focused inference network for lightweight image superresolution. IEEE TMM, pages 1-13, 2023. 3", + "[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPRW, pages 832-842, 2022. 1, 3, 6, 7, 8", + "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV, pages 1833-1844, 2021. 3", + "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 1132-1140, 2017." + ], + "bbox": [ + 516, + 92, + 903, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "408", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, pages 85-100, 2018. 2", + "[34] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, pages 41-55. Springer, 2020. 3, 6, 7, 8", + "[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPRW, pages 456-465, 2022. 3", + "[36] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 2", + "[37] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, pages 272-289. Springer, 2020. 6, 7", + "[38] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, pages 116-131, 2018. 2", + "[39] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423 vol.2, 2001. 6", + "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 6", + "[41] Marcelo Gennari Do Nascimento, Victor Prisacariu, and Roger Fawcett. *Dsconv: Efficient convolution operator. In ICCV*, pages 5147-5156, 2019. 2", + "[42] Tobias Pohlen, Alexander Hermans, Markus Mathias, and Bastian Leibe. Full-resolution residual networks for semantic segmentation in street scenes. In CVPR, pages 3309-3318, 2017. 2", + "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 2", + "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 3, 6", + "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[46] Ayan Sinha, Asim Unmesh, Qixing Huang, and Karthik Ramani. Surfnet: Generating 3d shape surfaces using deep residual networks. In CVPR, pages 791-800, 2017. 2", + "[47] Rupesh K Srivastava, Klaus Greff, and Jürgen Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2", + "[48] Bin Sun, Yulun Zhang, Songyao Jiang, and Yun Fu. Hybrid pixel-unshuffled network for lightweight image superresolution. In AAAI, pages 2375-2383, 2023. 3, 6, 7", + "[49] Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Spatially-adaptive feature modulation for efficient image" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "super-resolution. In ICCV, pages 13190-13199, 2023. 6, 7", + "[50] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 1110-1121, 2017. 6", + "[51] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In CVPR, pages 22378-22387, 2023. 3", + "[52] Xijun Wang, Meina Kan, Shiguang Shan, and Xilin Chen. Fully learnable group convolution for acceleration of deep neural networks. In CVPR, pages 9041–9050, 2019. 2", + "[53] Yang Wang and Tao Zhang. Osffnet: Omni-stage feature fusion network for lightweight image super-resolution. In AAAI, pages 5660-5668, 2024. 6, 7", + "[54] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6", + "[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 2", + "[56] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023. 1", + "[57] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 5987-5995, 2017. 2", + "[58] Jianchao Yang, John Wright, Thomas S. Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE TIP, 19(11):2861-2873, 2010. 6", + "[59] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, pages 4470-4479, 2019. 2", + "[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 2", + "[61] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 6", + "[62] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. Interleaved group convolutions. In ICCV, pages 4383-4392, 2017. 2", + "[63] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 2", + "[64] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2", + "[65] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 6, 7", + "[66] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV, pages 56-72. Springer, 2020. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "409", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[67] Daquan Zhou, Qibin Hou, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Rethinking bottleneck structure for efficient mobile network design. In ECCV, pages 680-697. Springer, 2020. 2", + "bbox": [ + 91, + 90, + 486, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "410", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_model.json b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f42ec88eb84e7c40936dd9088bd9ca209671ca70 --- /dev/null +++ b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_model.json @@ -0,0 +1,2444 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.13, + 0.85, + 0.177 + ], + "angle": 0, + "content": "SLVR: Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.204, + 0.409, + 0.222 + ], + "angle": 0, + "content": "Ning Ni" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.222, + 0.829, + 0.24 + ], + "angle": 0, + "content": "School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.243, + 0.482, + 0.256 + ], + "angle": 0, + "content": "ningni@mail.bnu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.205, + 0.664, + 0.222 + ], + "angle": 0, + "content": "Libao Zhang*" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.243, + 0.708, + 0.257 + ], + "angle": 0, + "content": "libaozhang@bnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.764 + ], + "angle": 0, + "content": "Recently, improving the residual structure and designing efficient convolutions have become important branches of lightweight visual reconstruction model design. We have observed that the feature addition mode (FAM) in existing residual structure tends to lead to slow feature learning or stagnation in feature evolution, a phenomenon we define as network inertia. In addition, although blueprint separable convolutions (BSConv) have proved the dominance of intrakernel correlation, BSCov forces the blueprint to perform scale transformation on all channels, which may lead to incorrect intra-kernel correlation and introduce useless or disruptive features on some channels and hinder the effective propagation of features. Therefore, in this paper, we rethink the FAM and BSCov for super-light visual reconstruction framework design. First, we design a novel linking mode, called feature diversity evolution link (FDEL), which aims to alleviate the phenomenon of network inertia by reducing the retention of previous low-level features, thereby promoting the evolution of feature diversity. Second, we propose blueprint controllable convolutions (B2Conv). The B2Conv can adaptively pick accurate intra-kernel correlation in the depth-axis, effectively preventing the introduction of useless or disruptive features. Based on FDEL and B2Conv, we develop a super-light super-resolution (SR) framework SLVR for visual reconstruction. Both FDEL and B2Conv can serve as efficient plugins. Extensive experimental results demonstrate the effectiveness of our proposed B2Conv, FDEL, and SLVR. Our code will be available at https://github.com/chongningni/SLVR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.222, + 0.805 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.816, + 0.485, + 0.862 + ], + "angle": 0, + "content": "As an important direction in the field of computer vision, visual reconstruction focuses on improving the visual quality of images or videos. Image super-resolution (SR) is" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.908, + 0.552 + ], + "angle": 0, + "content": "Figure 1. Performance and model complexity comparison on Set14 dataset for upscaling factor \\(\\times 4\\). * denotes the model trained on DF2K dataset and others trained on DIV2K." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.807 + ], + "angle": 0, + "content": "a core task in visual reconstruction, which aims to reconstruct high-resolution details from a low-resolution image. SR is an ill-posed problem because a low-resolution (LR) image can be mapped to multiple high-resolution (HR) spaces, which makes it challenging. Recently, convolutional neural networks (CNNs) based on residual learning [1, 9, 21, 22, 30, 56] have dominated the SR field because they can constrain mapping space and reduce the difficulty of SR learning, which creates conditions for lightweight SR design. Furthermore, lightweight convolution has become an important research branch of lightweight model design. In this paper, we aim to rethink the residual structure and blueprint separable convolutions (BSConv) to build a superlight SR reconstruction framework. The issues we discovered are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "1) Existing feature addition mode easily causes network inertia: SR tasks require recovering detailed textures, which require rich high-level features as support. The existing residual structure uses the feature addition mode (FAM) of input and output \\((x + \\mathcal{F}(x))\\), retaining many previous low-level features \\((x)\\), which is not conducive to the diver" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author. This work is supported by the National Natural Science Foundation of China under Grant 62271060 and 61571050." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "sity evolution of features and easily causes network inertia. \\(\\mathcal{F}(\\cdot)\\) represents the calculation function. Furthermore, retaining redundant previous low-level features will also affect the presentation of high-level features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.156, + 0.482, + 0.292 + ], + "angle": 0, + "content": "2) Incorrect intra-kernel correlation of the BSConv: Although BSConv proves the dominance of intra-kernel correlation, we find that kernel correlation is not present on all channels in an ordinary convolution. However, BSConv forces the blueprint to perform linear transformations on all channels, which may lead to incorrect intra-kernel correlation and result in introducing many useless or disruptive features on some channels to hinder the effective propagation of features." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.295, + 0.485, + 0.597 + ], + "angle": 0, + "content": "To solve the above problems, this work rethinks the residual structure and BSconv. First, to alleviate the network inertia, we propose a new linking mode, named feature diversity evolution link (FDEL). Our FDEL aims to minimize the proportion of low-level features during feature propagation and presentation, which is beneficial to distill rich high-level features with different representations and alleviate network inertia. In addition, to maximize the effect of FDEL, we introduce a diversified feature aggregation strategy (DFAS) to enhance feature representation. Second, to prevent the incorrect inter-kernel correlation of the BSConv (inter-blueprint correlation), we propose new blueprint controllable convolutions (B2Conv). Our B2Conv sets up an access control unit (ACU) in the channel dimension to prevent incorrect intra-blueprint correlation. The ACU is like a guard, which can adaptively pick relevant channels to associate and learn the correct interblueprint correlation. Our B2Conv can reduce the introduction of useless or disruptive features and promote the effective propagation of feature information." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.602, + 0.482, + 0.617 + ], + "angle": 0, + "content": "The main contributions of this paper are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.482, + 0.725 + ], + "angle": 0, + "content": "1) We propose a new convolution B2Conv. Our B2Conv introduces the ACU to adaptively control intra-blueprint correlation, which can precisely prevent the introduction of useless or disruptive features and promote effective propagation of feature information. Besides, the B2Conv can also serve as a highly efficient building block for CNNs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.482, + 0.821 + ], + "angle": 0, + "content": "2) We propose a new linking mode FDEL, which effectively alleviates the network inertia problem. Our FDEL aims to reduce the retention of previous low-level feature information and can effectively promote the diverse evolution of features, which is beneficial to the extraction and presentation of high-level features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.902 + ], + "angle": 0, + "content": "3) Based on FDEL and B2Conv, we design a novel lightweight SR framework SLVR. Extensive experiments demonstrate that SLVR achieves record high performance (e.g., 28.85dB on Set14 × 4 with only 485K parameters) on lightweight SR benchmarks (shown in Fig. 1)." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.621, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.664, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.537 + ], + "angle": 0, + "content": "Residual Network: Inspired by [47], He et al. [13] proposed first residual network for image recognition and achieved state-of-the-art performance. Then, many different residual networks were proposed [8, 14, 36, 42, 46] for different computer vision tasks. Residual networks are typically built by stacking a sequence of residual blocks. Recently, the residual block was further developed, such as widening the channels [60], introducing attention mechanism [17, 55], and combining with dense connections [64]. The residual network has achieved great success in heavy-weight networks because it overcomes the vanishing gradient problem. Besides, residual learning can reduce the learning difficulty of computer vision tasks, which also provides opportunities for lightweight network design. To reduce the model complexity, MobileNet [16] introduced depthwise separable convolutions as key elements in the residual blocks. Then, MobileNetV2 [43] proposed an inverted residual block where the shortcut connections are between the thin bottleneck layers. ShuffleNetV2 [38] inserts a channel split component before the inverted residual block. After that, Zhou et al. [67] rethought the inverted residual block and proposed a sandglass block, which aims to perform spatial transformation and identity mapping at higher dimensions to alleviate information loss and gradient confusion problems. Although previous works have effectively improved the residual structure, most proposed residual blocks now still use the addition mode of input and output features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.539, + 0.905, + 0.855 + ], + "angle": 0, + "content": "Efficient Convolution: To reduce network parameters, AlexNet [25] first proposed the concept of group convolution and ResNext [57] demonstrated its effectiveness. After that, the group convolution was further modified by this works [18, 52, 62, 63]. VGG [45] suggested using two \\(3 \\times 3\\) convolutions instead of one \\(5 \\times 5\\) convolution. Xception [5] proposed the depthwise separable convolution (DSConv) to further reduce the parameters of convolution and then it was used in MobileNet [16] and gained state-of-the-art results. Partial convolution (PConv) [33] was first proposed for image inpainting, which greatly improved computing efficiency. To solve the unreasonable problem of mask update in the partial convolution, Gated convolution (GConv) [59] was proposed. Octave convolution [3] was proposed to extract the high-frequency and low-frequency signals of the image. Recently, DSConv was decomposed [41] into two components: variable quantization kernels and distribution offsets, thereby achieving memory reduction and operation acceleration. BSConv [12] rethought the DSConv and considered the dominance of intra-kernel correlations when performing convolution operations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Lightweight SR: Although deep models can achieve good SR performance, they are not conducive to deployment on small devices. Therefore, many works were proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.483, + 0.344 + ], + "angle": 0, + "content": "Figure 2. Visualization of intermediate features, above the dotted line is the pipeline of BSRN [30] and green cuboids represent the ESDB modules. Outi represents the output of the \\( i^{th} \\) ESDB module. Outi = xi = xi-1 + yi." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.357, + 0.485, + 0.659 + ], + "angle": 0, + "content": "[6, 20, 44] to reduce the computational complexity. LapSRN [26] removed the bicubic interpolation pre-processing step. CARN [1] introduced group convolution and cascading residual blocks. IMDN [21] used channel splitting operations to complete information multi-distillation and selective feature fusion. After that, improving information distillation structure has become an important branch of lightweight SR network design, and many methods [9, 30, 34] were proposed. Then, PAN [66] introduced pixel attention to measure pixel contribution to achieve competitive performance by using fewer parameters. Recently, a series of transformer-based SR methods [4, 7, 11, 29, 31, 35, 48, 51] emerge with superior performance. HNCT [7] cleverly combined CNN and transformer to reduce the parameters while ensuring network performance. CFIN [29] design a novel cross-receptive field-guided transformer to adaptively select the required contextual information. NGswin [4] introduced the N-Gram context to the low-level vision with transformers and achieve superior performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.226, + 0.689 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.214, + 0.71 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "First, we define the phenomenon of network inertia as a situation where the input and output of a network module are similar, indicating that the module fails to learn meaningful features. As shown in Fig. 2, we find that the FAM \\((x_{i - 1} + y_i)\\) easily causes network inertia because it retains many previous input low-level features \\((x_{i - 1})\\) information. There are 8 ESDBs with this addition mode in BSRN, we can find that the outputs of most adjacent ESDB modules are similar, especially Out2 and Out3, Out4 and Out5. In addition, we replaced Out5 with Out4 in the subsequent feature fusion and were surprised to find that this replacement had almost no impact on the image reconstruction quality," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.2, + 0.907, + 0.269 + ], + "angle": 0, + "content": "Figure 3. Exemplary filter weights of ordinary convolution and BSConv. One filter kernel of ordinary convolution size \\( M \\times K \\times K = 64 \\times 3 \\times 3 \\) is visualized and split into 64 images. Filter kernels of pointwise convolution size \\( N \\times M \\times K \\times K = 64 \\times 64 \\times 1 \\times 1 \\) and two channels are visualized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.905, + 0.342 + ], + "angle": 0, + "content": "and the PSNR only decreased by less than \\(0.005\\mathrm{dB}\\). The similarity between features Out4 and Out5 can be inferred that the features \\(y_{4}\\) extracted by the \\(4^{th}\\) ESDB module are meaningless and network inertia occurs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Second, as shown in the left of Fig. 3, although the weights of each kernel are highly correlated along the channel dimension, differences also exist. Besides, BSConv [12] also proved that not all the slices (52/128 for the ResNet-50 example) of the kernel are highly correlated along the depth axis. However, as shown in the right of Fig. 3, we find that the pointwise convolution of BSConv associates almost all the channel dimensions along the depth axis (the darker the color, the higher the correlation), which results in introducing many useless or disruptive features on some channels and is not conducive to effective extraction and propagation of useful features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.787, + 0.549 + ], + "angle": 0, + "content": "3.2. Entropy Delta Extraction Link" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.584 + ], + "angle": 0, + "content": "In this section, we will explain the definition of FDEL in detail." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.632 + ], + "angle": 0, + "content": "In the network with stacked RBs, as shown in Fig. 5 (a), the initial shallow (low-level) features \\( x_{1} \\) will be superimposed and retained until the last layer \\( x_{L} \\), as follows" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.649, + 0.904, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {2} = \\mathcal {F} _ {1} (x _ {1}) + x _ {1}, \\dots , \\\\ x _ {l} = \\mathcal {F} _ {l - 1} \\left(x _ {l - 1}\\right) + x _ {l - 1}, \\dots , \\\\ x _ {L} = \\mathcal {F} _ {L - 1} \\left(x _ {L - 1}\\right) + x _ {L - 1}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.855 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_l\\) represents the operations of the \\(l^{th}\\) RB. Equation (1) can be regarded as a feature superposition operation. It may let the feature representation \\(x_{l}\\) of \\(l^{th}\\) RB rely too much on the input features \\(x_{l-1}\\), which will limit the learning direction of features and even lead to stagnation in feature evolution. This phenomenon can be evidenced in Fig. 2. In addition, the superposition of many low-level input features is not conducive to the presentation of high-level features, especially in structure Fig. 5 (c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In view of the aforementioned limitations of the FAM of the RBs, we rethink its design principles and propose an FDEL that can tackle the above issues by flipping the FAM," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.195, + 0.312, + 0.206 + ], + "angle": 0, + "content": "Shallow Feature Extraction" + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.195, + 0.526, + 0.207 + ], + "angle": 0, + "content": "Deep Feature Extraction" + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.195, + 0.747, + 0.207 + ], + "angle": 0, + "content": "Image Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.211, + 0.585, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.211, + 0.73, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.211, + 0.904, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.403, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 4. The overall architecture of the proposed SLVR and the structure of BCRB and blueprint controllable convolutions (B2Conv)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.484 + ], + "angle": 0, + "content": "the structure of RB with FDEL is shown in Fig. 5 (b). Our FDEL aims to remove excessive low-level features. The specific mathematical formula of FDEL is" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.495, + 0.482, + 0.511 + ], + "angle": 0, + "content": "\\[\nx _ {l + 1} = \\mathcal {F} _ {\\mathrm {F D E L}} \\left(x _ {l}, \\mathcal {F} _ {l} \\left(x _ {l}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.658 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{FDEL}}\\) denotes the operations of the FDEL. Our FDEL uses the output of the \\(\\mathcal{F}_l\\) minus its input. Because the output feature representation is more high-level than the input, each RB can effectively extract the different high-level features, which promotes the diverse evolution of features. As we all know, we usually refer to the output of a function minus its input as the gain of the function in a physical sense. Therefore, in an RB with FDEL, the gain is the high-level features extracted by this RB." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.409, + 0.683 + ], + "angle": 0, + "content": "3.3. Blueprint Controllable Convolutions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Although BSConv [12] demonstrates the importance of intra-kernel correlations in separating ordinary convolution, it does not consider the differences of the filters in the depth-axis. In other words, in some channel dimensions, this blueprint cannot obtain the original filter of the ordinary convolution through scale transformation, so filters after forced scale transformation don't effectively work. Therefore, we sparse the intra-kernel correlation and proposed a new B2Conv. Our B2Conv can adaptively pick appropriate channels along the depth axis to perform blueprint scale transformation, which can promote the extraction and efficient propagation of feature information." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In a convolution operation, assume that the input is \\(U\\in \\mathbb{R}^{M\\times H\\times W}\\), the output is \\(V\\in \\mathbb{R}^{N\\times H\\times W}\\), and the con" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.907, + 0.47 + ], + "angle": 0, + "content": "volution kernels are \\(\\{E^{(1)}, E^{(2)}, \\dots, E^{(N)}\\}\\) and \\(E^{(n)} \\in \\mathbb{R}^{M \\times H \\times W}\\). Then the process can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.478, + 0.905, + 0.496 + ], + "angle": 0, + "content": "\\[\nV _ {n, \\therefore :} = U * E ^ {(n)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.629 + ], + "angle": 0, + "content": "where \\(*\\) denotes the convolution operation and \\(n \\in \\{1,2,\\dots,N\\}\\), the weights of these \\(N\\) kernels can be learned during the training. BSCov proved that in practice these weights usually converge to a state that shows a substantial amount of correlation. Therefore, BSCov proposes to decompose an ordinary convolution into the interaction of a blueprint \\(\\mathcal{B}\\) and a set of scale-transforming weights \\(\\{w_{n,1}, w_{n,2}, \\dots, w_{n,M}\\}\\), as follows" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.638, + 0.905, + 0.658 + ], + "angle": 0, + "content": "\\[\nE _ {m,:,:} ^ {(n)} = w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.833 + ], + "angle": 0, + "content": "where \\(m \\in \\{1, 2, \\dots, M\\}\\) and \\(n \\in \\{1, 2, \\dots, N\\}\\) (see Fig. 6, left column). However, in ordinary convolution, not all the filters of channel dimension can be translated by Equation (4). Therefore, to solve the problem, we propose to add an access control unit (ACU) before scale-transforming. Specifically, in all channel dimensions, if there is a strong intra-kernel correlation with \\(\\mathcal{B}\\), we perform Equation (4), otherwise we do not perform any operation or multiply Equation (4) by a small correlation coefficient to ensure that the useless features extracted in this dimension will not affect the efficient propagation of features, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.842, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\nE _ {m,:,:} ^ {(n)} = \\mathcal {T} _ {n, m} \\cdot w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}_{n,m}\\) denotes the output of the ACU, ideally its value is 0 or 1. 1 means allowing scale-transforming and 0 means" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.28, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.088, + 0.483, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Comparison of residual blocks (RBs) with addition mode and FDEL. For the specific mathematical formula of FDEL, refer to Equation (2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.452 + ], + "angle": 0, + "content": "the opposite Here, to make the function continuous and differentiable, we soften the output of the ACU and limit its value range to \\((0,1)\\). The closer the \\(\\mathcal{T}_{n,m}\\) is to 1, the stronger the intra-kernel correlation. The characteristics of features can reflect the characteristics of the convolution kernels. Therefore, \\(\\mathcal{T}_{n,m}\\) can be obtained in an embedding space by feature space mapping and weights of the \\(\\mathcal{T}_{n,m}\\) and \\(w_{n,m}\\) can be learned directly via backpropagation. Specifically, we first rewrite the Equation (3) by split 3d convolution into a set of \\(M\\) 2d convolutions, yielding" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.462, + 0.483, + 0.504 + ], + "angle": 0, + "content": "\\[\nV _ {n,:,:} = U * E ^ {(n)} = \\sum_ {m = 1} ^ {M} \\left(U _ {m,:,:} * E _ {m,:,:} ^ {(n)}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Secondly, we replace the ordinary convolution \\( E_{m,\\cdot,\\cdot}^{(n)} \\) with the B2Conv representation as given in Equation (5), as follows" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.571, + 0.483, + 0.612 + ], + "angle": 0, + "content": "\\[\nV _ {n, \\therefore :} = \\sum_ {m = 1} ^ {M} \\left(U _ {m, \\therefore :} * \\left(\\mathcal {T} _ {n, m} \\cdot w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Because each blueprint \\(\\mathcal{B}^{(n)}\\) is independent in the channel dimension. So, we can rewrite the Equation (7) (Note: As shown in right column of Fig. 6, the input \\(U_{m,:;}\\) also needs to be copied \\((\\bar{U}_{m,:;})\\) for feature space mapping)" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.695, + 0.489, + 0.794 + ], + "angle": 0, + "content": "\\[\nV _ {n,:,:} = \\left(\\underbrace {\\sum_ {m = 1} ^ {M} \\left(\\bar {U} _ {m , :,:} \\cdot \\mathcal {T} _ {n , m}\\right)} _ {\\bar {V} _ {n,:,:} ^ {\\prime}} \\cdot \\underbrace {\\sum_ {m = 1} ^ {M} \\left(U _ {m , :,:} \\cdot w _ {n , m}\\right)} _ {V ^ {\\prime} _ {n,:,:}}\\right) * \\mathcal {B} ^ {(n)}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.484, + 0.858 + ], + "angle": 0, + "content": "Then, we further rearrange \\(\\{w_{n,1}, w_{n,2}, \\dots, w_{n,M}\\}\\) and \\(\\{\\mathcal{T}_{n,1}, \\mathcal{T}_{n,2}, \\dots, \\mathcal{T}_{n,M}\\}\\) into a \\(M \\times 1 \\times 1\\) arrays \\(\\tilde{w}_n\\) and \\(\\tilde{\\mathcal{T}}_n\\) respectively, the sum operations of Equation (8) can be replaced by convolution operations, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.867, + 0.483, + 0.887 + ], + "angle": 0, + "content": "\\[\n\\bar {V} _ {n,:,:} ^ {\\prime} = U * \\tilde {\\mathcal {T}} _ {n}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.888, + 0.482, + 0.905 + ], + "angle": 0, + "content": "\\[\nV _ {n,:,:} ^ {\\prime} = U * \\tilde {w} _ {n}, \\tag {10}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Figure 6. Interpretations for BSConv and B2Conv. B2Conv can adaptively pick correct intra-kernel correlations." + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.281, + 0.905, + 0.302 + ], + "angle": 0, + "content": "\\[\nV _ {n,:,:} = \\left(\\sigma \\left(\\bar {V} _ {n,:,:} ^ {\\prime}\\right) \\cdot V _ {n,:,:} ^ {\\prime}\\right) * \\mathcal {B} ^ {(n)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.305, + 0.907, + 0.42 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the sigmoid function, which limits \\(\\tilde{V}_{n,::}^{\\prime}\\) to (0,1). For a concrete implementation, Equation (9) and Equation (10) can be translated into a \\(1\\times 1\\) pointwise convolution with the kernels \\(\\{\\tilde{w}_1,\\tilde{w}_2,\\dots ,\\tilde{w}_N\\}\\) and \\(\\left\\{\\tilde{\\mathcal{T}}_1,\\tilde{\\mathcal{T}}_2,\\dots ,\\tilde{\\mathcal{T}}_N\\right\\}\\) respectively. The Equation (11) can be translated into a \\(K\\times K\\) depthwise convolution with the kernels \\(\\{\\mathcal{B}^{(1)},\\mathcal{B}^{(2)},\\dots ,\\mathcal{B}^{(N)}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.426, + 0.722, + 0.441 + ], + "angle": 0, + "content": "3.4. The Overall Structure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.906, + 0.555 + ], + "angle": 0, + "content": "Based on FDEL and B2Conv, we further develop a lightweight SLVR framework to achieve high-performance image SR. As shown in Fig. 4, SLVR consists of three parts, including shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given a LR input \\( I_{\\mathrm{LR}} \\in \\mathbb{R}^{H \\times W \\times C_{in}} \\), we first use a B2Conv \\( \\mathcal{H}_{\\mathrm{B2Conv}} \\) to extract shallow feature \\( X_0 \\in \\mathbb{R}^{H \\times W \\times C_{mid}} \\), as follows" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.564, + 0.905, + 0.582 + ], + "angle": 0, + "content": "\\[\nX _ {0} = \\mathcal {H} _ {\\mathrm {B 2 C o n v}} \\left(\\operatorname {C o n c a t} \\left[ I _ {\\mathrm {L R}} \\right] _ {n}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.906, + 0.71 + ], + "angle": 0, + "content": "where \\( C_{in} \\) and \\( C_{mid} \\) denote the channel number of the input and shallow feature. \\( \\text{Concat}[\\cdot]_n \\) means concatenating \\( n \\) times input. Then we use \\( K \\) efficient controllable distillation block (ECDB) to extract deep features. To maximize the effect of FDEL, feature aggregation is performed by DFAS and GELU [15] activation to enhance feature representation. After that, we use a B2Conv to continue refining the features, the above process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.719, + 0.905, + 0.736 + ], + "angle": 0, + "content": "\\[\nX _ {i} = \\mathcal {H} _ {\\mathrm {E C D B} _ {i}} \\left(X _ {i - 1}\\right), i = 1, 2, \\dots , K, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.744, + 0.905, + 0.774 + ], + "angle": 0, + "content": "\\[\nX _ {\\mathrm {A g g}} = \\mathcal {H} _ {\\mathrm {D F A S}} \\left(\\left(w _ {1}, w _ {2}, \\dots , w _ {c}\\right), \\left(X _ {1}, X _ {2}, \\dots , X _ {K}\\right)\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.775, + 0.905, + 0.791 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {R e f i n e d}} = \\mathcal {H} _ {\\mathrm {B 2 C o n v}} \\left(\\operatorname {G E L U} \\left(X _ {\\mathrm {A g g}}\\right)\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_{\\mathrm{ECDB}_i}\\) represents the \\(i^{\\mathrm{th}}\\) ECDB, \\(w_{1}, w_{2}, \\dots, w_{c}\\) are the learnable weights, \\(X_{1}, X_{2}, \\dots, X_{K}\\) denote the intermediate features. \\(\\mathcal{H}_{\\mathrm{DFAS}}(\\cdot)\\) denotes the operations of the DFAS, which can be achieved by a convolution. The ECDB is designed based on the ESDB of the BSRN. Specifically, first, we replace the BSRB with the blueprint controllable residual block (BCRB). Second, we replace the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "404" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.088, + 0.885, + 0.586 + ], + "angle": 0, + "content": "
MethodT-DatasetScale#ParamsMult-AddsSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
VDSR [22]291×2666K612.6G37.530.958733.030.912431.900.896030.760.914037.220.9750
CARN [1]D2K+291×21,592K222.8G37.760.959033.520.916632.090.897831.920.925638.360.9765
IMDN [21]D2K×2694K158.8G38.000.960533.630.917732.190.899632.170.928338.880.9774
LAPAR-A [27]DF2K×2548K171G38.010.960533.620.918332.190.899932.100.928338.670.9772
LatticeNet [37]D2K×2756K169.5G38.060.960733.700.918732.200.899932.250.928838.940.9774
RLFN [24]D2K×2527K97.9G38.070.960733.720.918732.220.900032.330.9299--
SRPN-Lite [65]DF2K×2609K139.9G38.100.960833.700.918932.250.900532.260.9294--
FMEN [6]DF2K×2748K172.0G38.100.960933.750.919232.260.900732.410.931138.950.9778
MLRN [10]D2K×2488k90.4G38.070.960733.590.918032.210.900032.280.929738.760.9773
NGswin [4]D2K×2998K140.4G38.050.961033.790.919932.270.900832.530.932438.970.9777
SLVR (Ours)D2K×2464K102G38.140.961033.870.920232.280.900932.570.931839.080.9776
VDSR [22]291×3666K612.6G33.660.921329.770.831428.820.797627.140.827932.010.9340
CARN [1]D2K+291×31,592K118.8G34.290.925530.290.840729.060.803428.060.849333.500.9440
IMDN [21]D2K×3703K71.5G34.360.927030.320.841729.090.804628.170.851933.610.9445
LAPAR-A [27]DF2K×3594K114G34.360.926730.340.842129.110.805428.150.852333.510.9441
LatticeNet [37]D2K×3765K76.3G34.400.927230.320.841629.100.804928.190.851333.630.9442
RFDN [34]D2K×3633K65.6G34.470.928030.350.842129.110.805328.320.854733.780.9458
SRPN-Lite [65]DF2K×3615K62.7G34.470.927630.380.842529.160.806128.220.8534--
FMEN [6]DF2K×3757K77.2G34.450.927530.400.843529.170.806328.330.856233.860.9462
MLRN [10]D2K×3496K40.9G34.460.926730.350.842629.100.805428.200.853333.660.9450
NGswin [4]D2K×31,007K66.6G34.520.928230.530.845629.190.807828.520.860333.890.9470
SLVR (Ours)D2K×3473K45.3G34.580.928530.530.846029.210.808028.560.860334.070.9475
VDSR [22]291×4666K612.6G31.350.883828.010.767427.290.725125.180.752428.830.8870
CARN [1]D2K+291×41,592K90.9G32.130.893728.600.780627.580.734926.070.783730.470.9084
IMDN [21]D2K×4715K40.9G32.210.894828.580.781127.560.735326.040.783830.450.9075
LAPAR-A [27]DF2K×4659K94G32.150.894428.610.781827.610.736626.140.787130.420.9074
LatticeNet [37]D2K×4777K43.6G32.180.894328.610.781227.570.735526.140.784430.540.9075
RFDN [34]D2K×4550K23.9G32.240.895228.610.781927.570.736026.110.785830.580.9089
RLFN [24]D2K×4543K25.3G32.240.895228.620.781327.600.736426.170.7877--
SRPN-Lite [65]DF2K×4623K35.8G32.240.895828.690.783627.630.737326.160.7875--
BSRN [30]DF2K×4352K19.4G32.350.896628.730.784727.650.738726.270.790830.840.9123
FMEN [6]DF2K×4769K44.2G32.240.895528.700.783927.630.737926.280.790830.700.9107
MLRN [10]D2K×4507K23.5G32.300.895628.620.782427.570.736526.100.786730.560.9092
HPUN-M [48]D2K×4511K27.7G32.240.895028.660.782827.600.737126.120.787830.550.9089
SAFMN [49]DF2K×4240K14.0G32.180.894828.600.781327.580.735925.970.780930.430.9063
OSFFNet [53]DF2K×4537K22.0G32.390.897628.750.785227.660.739326.360.795030.840.9125
NGswin [4]D2K×41,019K36.4G32.330.896328.780.785927.660.739626.450.796330.800.9128
SLVR-S (Ours)DF2K×4158K8.5G32.200.895228.650.782227.570.736126.080.785430.530.9084
SLVR (Ours)D2K×4485K26.2G32.400.897528.780.785027.650.738826.410.795130.930.9133
SLVR* (Ours)DF2K×4485K26.2G32.440.898028.850.786527.700.740026.580.800031.100.9141
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.595, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison (PSNR/SSIM) for lightweight SR with state-of-the-art (SOTA) methods on five benchmark datasets. The best and second-best results are marked in red and blue colors respectively. Mult-Adds is evaluated on a \\( {1280} \\times {720}\\mathrm{{HR}} \\) image. T-Dataset denotes training dataset. D2K stands for the DIV2K dataset. DF2K indicates a merged dataset containing \\( {800} + 2,{650}\\mathrm{{HR}} - \\mathrm{{LR}} \\) image pairs. 291 images dataset is from [58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.768 + ], + "angle": 0, + "content": "addition mode of the residual structure with the proposed FDEL. Third, all the BSConvs are replaced with the proposed B2Conv. Please see Fig. 4 for detailed changes. After performing the above operations, we obtain the final refined features \\(X_{\\mathrm{Refined}}\\). Finally, to take advantage of residual learning, we use long skip connections to reduce the difficulty of learning. SR results can be obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.773, + 0.483, + 0.79 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {S R}} = \\mathcal {H} _ {\\text {U p s a m p l e}} \\left(X _ {\\text {R e f i n e d}}\\right) + \\text {B i c u b i c} ^ {\\uparrow} \\left(I _ {\\mathrm {L R}}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_{\\mathrm{Upsample}}(\\cdot)\\) represents the upsampling operation, which is implemented with the pixelshuffle operation [44]. Bicubic denotes bicubic upsampling. \\(L_{1}\\) loss function is used to optimize the network, as follows" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.864, + 0.483, + 0.881 + ], + "angle": 0, + "content": "\\[\nL _ {1} = \\left\\| I _ {\\mathrm {H R}} - I _ {\\mathrm {S R}} \\right\\| _ {1}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.271, + 0.901 + ], + "angle": 0, + "content": "where \\(I_{\\mathrm{HR}}\\) is the HR label." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.66, + 0.646, + 0.677 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.688, + 0.704, + 0.705 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.907, + 0.867 + ], + "angle": 0, + "content": "Datasets and Metrics. We use DIV2K [50] dataset as the training dataset, which contains 800 HR-LR image pairs. To show more fair results, we used DF2K (DIV2K + Flickr2K [50]) to additionally train an \\(\\times 4\\) model. Note that the model trained with DF2K is marked with a small *. For testing, we evaluated the performances on the five benchmark datasets, including Set5 [2], Set14 [61], B100 [39], Urban100 [19], Manga109 [40]. The PSNR and SSIM [54] are used as the evaluation metrics, which are evaluated on the Y channel of the YCbCr space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. ECDB number is set to 8, and the channel number is set to 64. The kernel size of the depth-" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Figure 7. Visual comparison for \\(\\times 4\\) SR methods. The patches for comparison are marked with red boxes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.443, + 0.483, + 0.673 + ], + "angle": 0, + "content": "wise convolutions is set to 3. We additionally built a smaller model SLVR-S to compare the latest methods with less than 200K parameters and it trained on DF2K. The ECDB number of the SLVR-S is set to 6 and the channel number is set to 40. During training, we augment the data with random horizontal flips and rotation \\((90^{\\mathrm{o}},180^{\\mathrm{o}},270^{\\mathrm{o}})\\). We randomly crop LR patches of size \\(64\\times 64\\) as input. Our models are trained by Adam optimizer [23] for 1000K iterations. The initial learning rate is set to \\(1\\times 10^{-3}\\) and halved for every 200k iterations. The batch size of SLVR and SLVR-S is set to 64 and 128 respectively. LR image patches are obtained by the bicubic kernel from corresponding HR images matching each SR task. Our networks are trained on an RTX3090 GPU. Note that no other training skills (e.g., pre-training [28]) are employed." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.681, + 0.439, + 0.698 + ], + "angle": 0, + "content": "4.2. Comparison with the SOTA SR methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.856 + ], + "angle": 0, + "content": "To evaluate the effectiveness of SLVR, we compare our model with other efficient SR models on scale factors \\(\\times 2,\\times 3,\\times 4\\) including VDSR (CVPR16) [22], CARN (ECCV18) [1], IMDN (ACMMM19) [21], RFDN (ECCV20) [34], LAPAR-A (NeurIPS20) [27], LatticeNet (ECCV20) [37], RLFN (CVPRW22) [24], SRPN-Lite (ICLR22) [65], BSRN (CVPRW22) [30], FMEN (CVPRW22) [6], MLRN (CVPRW23) [10], NGswin (CVPR23) [4], HPUN-M (AAAI23) [48], SAFMN (ICCV23) [49], OSFFNet (AAAI24) [53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 1, we reported the training dataset, the parameters, Mult-Adds, and PSNR and SSIM on five benchmarks. Compared with models with similar param" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.707, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.092, + 0.905, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.221, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Figure 8. Ablation studies on B2Conv. BSRN-B2Conv denotes the BSCnv of BSRN is replaced by B2Conv. All the results are reported on scale factor \\(\\times 4\\)" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.28, + 0.911, + 0.339 + ], + "angle": 0, + "content": "
MethodFDELB2ConvSet5Set14Urban100
Model-I✓×32.37/0.896928.71/0.783926.30/0.7917
Model-II×✓32.39/0.897528.76/0.785126.40/0.7943
SLVR✓✓32.40/0.897528.78/0.785026.41/0.7951
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.35, + 0.907, + 0.393 + ], + "angle": 0, + "content": "Table 2. Ablation studies of FDEL and B2Conv on scale factor \\(\\times 4\\). The \\(\\times\\) of FDEL denotes it is replaced by addition mode, and the \\(\\times\\) of B2Conv denotes it is replaced by BSConv." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.558 + ], + "angle": 0, + "content": "eters and Mult-Adds, the performance of our SLVR surpasses existing methods with a notable margin on all benchmarks. Compared with the latest transformer-based method NGswin, although the number of parameters of our SLVR is less than \\(50\\%\\) of NGswin and the number of Mult-Adds is less than \\(75\\%\\) of NGswin, our SLVR still has an advantage in performance. Compared with the latest extremely lightweight method SAFMN trained on the same dataset, our SLVR-S obtains better performance with a smaller model size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.728 + ], + "angle": 0, + "content": "We also perform a visual comparison of different SOTA lightweight SR methods on scale factor \\(\\times 4\\). As shown in Fig. 7, we can observe that the SR results generated by SLVR contain more correct and fine-grained textures, while other methods generate many blurred edges and artifacts. For example, in the second row of Fig. 7, our SLVR can restore the texture of the fence, while other methods cannot. We think that the visual results can prove the effectiveness of the proposed B2conv, which can prevent incorrect intrakernel correlations to reduce useless or disruptive feature extraction and propagation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.681, + 0.756 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The core contributions of our framework are the proposed FDEL and B2Conv. To demonstrate their effectiveness, we first conduct ablation studies on the main framework of our SLVR. Specifically, we remove FDEL and B2Conv and replace them with the original linking mode FAM of RBs and BSCnv, respectively. The experimental results are reported in Table 2. We can observe that our FDEL and B2Conv can improve network performance, especially B2Conv, improved by 0.11 dB on the dataset Urban100." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "406" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.483, + 0.128 + ], + "angle": 0, + "content": "
MethodScale#ParamsMult-AddsSet5Set14Urban100
EDSR-baseline×41,518K114.0G32.08/0.893828.58/0.781025.98/0.7824
EDSR+FDEL1,518K114.0G32.12/0.894428.61/0.781926.05/0.7845
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.139, + 0.483, + 0.167 + ], + "angle": 0, + "content": "Table 3. Ablation studies of FDEL. EDSR+FDEL denotes that all addition mode of EDSR-baseline [32] are replaced by FDEL." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.177, + 0.481, + 0.257 + ], + "angle": 0, + "content": "
RFDNScale#ParamsMult-AddsSet5Set14Urban100
w BSConv×3197K25.8G34.37/0.926830.34/0.841828.06/0.8503
w B2Conv261K34.5G34.44/0.927330.36/0.842628.18/0.8525
w Conv633K65.6G34.47/0.928030.35/0.842128.21/0.8525
w BSConv×4210K14.9G32.12/0.894128.59/0.780626.01/0.7829
w B2Conv273K20.1G32.22/0.894928.65/0.782226.06/0.7843
w Conv643K37.4G32.28/0.895728.61/0.781826.11/0.7858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.31 + ], + "angle": 0, + "content": "Table 4. Comprehensive comparison between ordinary convolution (Conv), BSCnv, and B2Conv. The best and second-best results are marked in red and blue colors." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.321, + 0.481, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.435 + ], + "angle": 0, + "content": "Figure 9. Visual comparison of feature maps between SLVR and SLVR without FDEL. Outi represents the output of the \\( i^{th} \\) ECDB module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Second, to verify that FDEL can be used as a more efficient linking mode of RBs, we use FDEL to simply replace the traditional linking mode FAM in the classic residual architecture EDSR [32]. From Table 3, we can conclude that our FDEL is more effective than the traditional FAM without additional parameters and Mult-Adds consumed." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Third, to further verify the effectiveness of our proposed B2Conv and confirm that B2Conv can be used as highly efficient building blocks. We denote BSConv of BSRN [30] replaced by B2Conv as BSRN-B2Conv. As shown in Fig. 8, we can observe that our B2Conv can significantly improve model performance without changing model architecture. Besides, we also conducted a comprehensive comparison between ordinary convolution, BSConv, and B2Conv on the network architecture RFDN [34], as shown in Table 4, compared with BSConv, our B2Conv greatly improves the performance (e.g., \\( \\mathbf{0.1dB} \\) on Set5 × 4) of the RFDN with few additional parameters (63K) consumed, which also illustrates the rationality of our improvement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finally, we perform a visual comparison of features between SLVR and SLVR without FDEL. From Fig. 9, we find that our FDEL can effectively alleviate network inertia. In addition, we visualize the scale transformation state of 8 blueprints on 64 channels. As shown in Fig. 10, we can observe that our B2Conv does not perform scale transformation on all 64 channel dimensions (only on the channel denoted by black squares). It proves that our B2Conv can adaptively select the appropriate channel to perform blueprint scale transformation, which can effectively prevent the introduction of disruptive or useless features. Fur" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.611, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.09, + 0.707, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.09, + 0.806, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.09, + 0.904, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.173, + 0.905, + 0.215 + ], + "angle": 0, + "content": "Figure 10. Visualization of scale transformation state of 4 blueprints on 64 channels. Black squares denote strong intrakernel correlation, and white squares are the opposite." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.382, + 0.905, + 0.409 + ], + "angle": 0, + "content": "Figure 11. Visual comparison of intermediate feature maps between SLVR and SLVR without B2Conv." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.587 + ], + "angle": 0, + "content": "thermore, we perform a visual comparison of features between SLVR and SLVR without B2Conv. From Fig. 11, we can find that the intermediate features generated by the SLVR without B2Conv contain more interference and disruptive features, resulting in poorer feature representation. The intermediate features generated by our SLVR are more accurate and have better feature representation ability. It also indirectly proves that the proposed B2Conv can adaptively select the correct intra-kernel correlation, prevent the introduction of disruptive features, and is more conducive to feature propagation and representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.633, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose blueprint controllable convolutions (B2Conv). B2Conv can precisely prevent incorrect intra-kernel correlation, which effectively reduces the extraction of useless or disruptive features and promotes the effective propagation of feature information. Besides, we propose a novel linking mode FDEL for RBs. Our FDEL reduces the retention of the input low-level features and can promote the diversity evolution of high-level features, which can effectively alleviate the occurrence of network inertia. Finally, based on B2Conv and FDEL, we propose an efficient lightweight SR framework SLVR. Furthermore, our B2Conv and FDEL can also serve as highly efficient components to improve the performance of other CNN-based models (such as EDSR, BSRN, and RFDN). Extensive experiments on public benchmark datasets show the prominent SR performance of SLVR and comprehensive analytical studies validate the reasonableness and effectiveness of the proposed framework and components." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "407" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, pages 252-268, 2018. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[3] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Yan Shuicheng, and Ji-ashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In ICCV, pages 3434–3443, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.3, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image superresolution. In CVPR, pages 2071–2081, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.441 + ], + "angle": 0, + "content": "[6] Zongcai Du, Ding Liu, Jie Liu, Jie Tang, Gangshan Wu, and Lean Fu. Fast and memory-efficient network towards efficient image super-resolution. In CVPRW, pages 852-861, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[7] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In CVPRW, pages 1102-1111, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[8] Mengyang Feng, Huchuan Lu, and Yizhou Yu. Residual learning for salient object detection. IEEE TIP, 29:4696-4708, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[9] Guangwei Gao, Wenjie Li, Juncheng Li, Fei Wu, Huimin Lu, and Yi Yu. Feature distillation interaction weighting network for lightweight image super-resolution. In AAAI, pages 661-669, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.587, + 0.483, + 0.64 + ], + "angle": 0, + "content": "[10] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. Mixer-based local residual network for lightweight image super-resolution. In CVPRW, pages 1593-1602, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.644, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[11] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. A simple transformer-style network for lightweight image super-resolution. In CVPRW, pages 1484-1494, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[12] Daniel Haase and Manuel Amthor. Rethinking depthwise separable convolutions: How intra-kernel correlations lead to improved mobilenets. In CVPR, pages 14588-14597, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.744, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.787, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, pages 630-645. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolu" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[18] Gao Huang, Shichen Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Condensenet: An efficient densenet using learned group convolutions. In CVPR, pages 2752-2761, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[20] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, pages 723-731, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[21] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, pages 2024-2032, 2019. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[24] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPRW, pages 765-775, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[26] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, pages 5835-5843, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[27] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. NeurIPS, 33:20343-20355, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[28] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[29] Wenjie Li, Juncheng Li, Guangwei Gao, Weihong Deng, Jiantao Zhou, Jian Yang, and Guo-Jun Qi. Cross-receptive focused inference network for lightweight image superresolution. IEEE TMM, pages 1-13, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPRW, pages 832-842, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV, pages 1833-1844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.888 + ], + "angle": 0, + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 1132-1140, 2017." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "408" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[33] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, pages 85-100, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[34] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, pages 41-55. Springer, 2020. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPRW, pages 456-465, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[36] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[37] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, pages 272-289. Springer, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[38] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, pages 116-131, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[39] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423 vol.2, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[41] Marcelo Gennari Do Nascimento, Victor Prisacariu, and Roger Fawcett. *Dsconv: Efficient convolution operator. In ICCV*, pages 5147-5156, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[42] Tobias Pohlen, Alexander Hermans, Markus Mathias, and Bastian Leibe. Full-resolution residual networks for semantic segmentation in street scenes. In CVPR, pages 3309-3318, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[46] Ayan Sinha, Asim Unmesh, Qixing Huang, and Karthik Ramani. Surfnet: Generating 3d shape surfaces using deep residual networks. In CVPR, pages 791-800, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[47] Rupesh K Srivastava, Klaus Greff, and Jürgen Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[48] Bin Sun, Yulun Zhang, Songyao Jiang, and Yun Fu. Hybrid pixel-unshuffled network for lightweight image superresolution. In AAAI, pages 2375-2383, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[49] Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Spatially-adaptive feature modulation for efficient image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "super-resolution. In ICCV, pages 13190-13199, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[50] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 1110-1121, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[51] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In CVPR, pages 22378-22387, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[52] Xijun Wang, Meina Kan, Shiguang Shan, and Xilin Chen. Fully learnable group convolution for acceleration of deep neural networks. In CVPR, pages 9041–9050, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[53] Yang Wang and Tao Zhang. Osffnet: Omni-stage feature fusion network for lightweight image super-resolution. In AAAI, pages 5660-5668, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[54] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[56] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[57] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 5987-5995, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[58] Jianchao Yang, John Wright, Thomas S. Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE TIP, 19(11):2861-2873, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[59] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, pages 4470-4479, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[61] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[62] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. Interleaved group convolutions. In ICCV, pages 4383-4392, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[63] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[64] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[65] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[66] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV, pages 56-72. Springer, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "409" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.148 + ], + "angle": 0, + "content": "[67] Daquan Zhou, Qibin Hou, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Rethinking bottleneck structure for efficient mobile network design. In ECCV, pages 680-697. Springer, 2020. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "410" + } + ] +] \ No newline at end of file diff --git a/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_origin.pdf b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bf068ee69f1c0fff4ec03395a318f577f3ebfa0 --- /dev/null +++ b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/c7aef326-130c-4b70-8496-8a5c0eb43ee8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cbeb42efa340b935bbe51da05a83e39fa7c1211132ee4a7b5821401bb435757 +size 4281471 diff --git a/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/full.md b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e5398f68c2bf76cf3cd7945489f6c05d64621a34 --- /dev/null +++ b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/full.md @@ -0,0 +1,355 @@ +# SLVR: Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation + +Ning Ni + +School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China + +ningni@mail.bnu.edu.cn + +Libao Zhang* + +libaozhang@bnu.edu.cn + +# Abstract + +Recently, improving the residual structure and designing efficient convolutions have become important branches of lightweight visual reconstruction model design. We have observed that the feature addition mode (FAM) in existing residual structure tends to lead to slow feature learning or stagnation in feature evolution, a phenomenon we define as network inertia. In addition, although blueprint separable convolutions (BSConv) have proved the dominance of intrakernel correlation, BSCov forces the blueprint to perform scale transformation on all channels, which may lead to incorrect intra-kernel correlation and introduce useless or disruptive features on some channels and hinder the effective propagation of features. Therefore, in this paper, we rethink the FAM and BSCov for super-light visual reconstruction framework design. First, we design a novel linking mode, called feature diversity evolution link (FDEL), which aims to alleviate the phenomenon of network inertia by reducing the retention of previous low-level features, thereby promoting the evolution of feature diversity. Second, we propose blueprint controllable convolutions (B2Conv). The B2Conv can adaptively pick accurate intra-kernel correlation in the depth-axis, effectively preventing the introduction of useless or disruptive features. Based on FDEL and B2Conv, we develop a super-light super-resolution (SR) framework SLVR for visual reconstruction. Both FDEL and B2Conv can serve as efficient plugins. Extensive experimental results demonstrate the effectiveness of our proposed B2Conv, FDEL, and SLVR. Our code will be available at https://github.com/chongningni/SLVR. + +# 1. Introduction + +As an important direction in the field of computer vision, visual reconstruction focuses on improving the visual quality of images or videos. Image super-resolution (SR) is + +![](images/5aa3a29a6fe17dccaefc5355ba82a92b438b9f54bb8bcbbacf0861d6dfb2fe39.jpg) +Figure 1. Performance and model complexity comparison on Set14 dataset for upscaling factor $\times 4$ . * denotes the model trained on DF2K dataset and others trained on DIV2K. + +a core task in visual reconstruction, which aims to reconstruct high-resolution details from a low-resolution image. SR is an ill-posed problem because a low-resolution (LR) image can be mapped to multiple high-resolution (HR) spaces, which makes it challenging. Recently, convolutional neural networks (CNNs) based on residual learning [1, 9, 21, 22, 30, 56] have dominated the SR field because they can constrain mapping space and reduce the difficulty of SR learning, which creates conditions for lightweight SR design. Furthermore, lightweight convolution has become an important research branch of lightweight model design. In this paper, we aim to rethink the residual structure and blueprint separable convolutions (BSConv) to build a superlight SR reconstruction framework. The issues we discovered are as follows: + +1) Existing feature addition mode easily causes network inertia: SR tasks require recovering detailed textures, which require rich high-level features as support. The existing residual structure uses the feature addition mode (FAM) of input and output $(x + \mathcal{F}(x))$ , retaining many previous low-level features $(x)$ , which is not conducive to the diver + +sity evolution of features and easily causes network inertia. $\mathcal{F}(\cdot)$ represents the calculation function. Furthermore, retaining redundant previous low-level features will also affect the presentation of high-level features. + +2) Incorrect intra-kernel correlation of the BSConv: Although BSConv proves the dominance of intra-kernel correlation, we find that kernel correlation is not present on all channels in an ordinary convolution. However, BSConv forces the blueprint to perform linear transformations on all channels, which may lead to incorrect intra-kernel correlation and result in introducing many useless or disruptive features on some channels to hinder the effective propagation of features. + +To solve the above problems, this work rethinks the residual structure and BSconv. First, to alleviate the network inertia, we propose a new linking mode, named feature diversity evolution link (FDEL). Our FDEL aims to minimize the proportion of low-level features during feature propagation and presentation, which is beneficial to distill rich high-level features with different representations and alleviate network inertia. In addition, to maximize the effect of FDEL, we introduce a diversified feature aggregation strategy (DFAS) to enhance feature representation. Second, to prevent the incorrect inter-kernel correlation of the BSConv (inter-blueprint correlation), we propose new blueprint controllable convolutions (B2Conv). Our B2Conv sets up an access control unit (ACU) in the channel dimension to prevent incorrect intra-blueprint correlation. The ACU is like a guard, which can adaptively pick relevant channels to associate and learn the correct interblueprint correlation. Our B2Conv can reduce the introduction of useless or disruptive features and promote the effective propagation of feature information. + +The main contributions of this paper are summarized as: + +1) We propose a new convolution B2Conv. Our B2Conv introduces the ACU to adaptively control intra-blueprint correlation, which can precisely prevent the introduction of useless or disruptive features and promote effective propagation of feature information. Besides, the B2Conv can also serve as a highly efficient building block for CNNs. +2) We propose a new linking mode FDEL, which effectively alleviates the network inertia problem. Our FDEL aims to reduce the retention of previous low-level feature information and can effectively promote the diverse evolution of features, which is beneficial to the extraction and presentation of high-level features. +3) Based on FDEL and B2Conv, we design a novel lightweight SR framework SLVR. Extensive experiments demonstrate that SLVR achieves record high performance (e.g., 28.85dB on Set14 × 4 with only 485K parameters) on lightweight SR benchmarks (shown in Fig. 1). + +# 2. Related Works + +Residual Network: Inspired by [47], He et al. [13] proposed first residual network for image recognition and achieved state-of-the-art performance. Then, many different residual networks were proposed [8, 14, 36, 42, 46] for different computer vision tasks. Residual networks are typically built by stacking a sequence of residual blocks. Recently, the residual block was further developed, such as widening the channels [60], introducing attention mechanism [17, 55], and combining with dense connections [64]. The residual network has achieved great success in heavy-weight networks because it overcomes the vanishing gradient problem. Besides, residual learning can reduce the learning difficulty of computer vision tasks, which also provides opportunities for lightweight network design. To reduce the model complexity, MobileNet [16] introduced depthwise separable convolutions as key elements in the residual blocks. Then, MobileNetV2 [43] proposed an inverted residual block where the shortcut connections are between the thin bottleneck layers. ShuffleNetV2 [38] inserts a channel split component before the inverted residual block. After that, Zhou et al. [67] rethought the inverted residual block and proposed a sandglass block, which aims to perform spatial transformation and identity mapping at higher dimensions to alleviate information loss and gradient confusion problems. Although previous works have effectively improved the residual structure, most proposed residual blocks now still use the addition mode of input and output features. + +Efficient Convolution: To reduce network parameters, AlexNet [25] first proposed the concept of group convolution and ResNext [57] demonstrated its effectiveness. After that, the group convolution was further modified by this works [18, 52, 62, 63]. VGG [45] suggested using two $3 \times 3$ convolutions instead of one $5 \times 5$ convolution. Xception [5] proposed the depthwise separable convolution (DSConv) to further reduce the parameters of convolution and then it was used in MobileNet [16] and gained state-of-the-art results. Partial convolution (PConv) [33] was first proposed for image inpainting, which greatly improved computing efficiency. To solve the unreasonable problem of mask update in the partial convolution, Gated convolution (GConv) [59] was proposed. Octave convolution [3] was proposed to extract the high-frequency and low-frequency signals of the image. Recently, DSConv was decomposed [41] into two components: variable quantization kernels and distribution offsets, thereby achieving memory reduction and operation acceleration. BSConv [12] rethought the DSConv and considered the dominance of intra-kernel correlations when performing convolution operations. + +Lightweight SR: Although deep models can achieve good SR performance, they are not conducive to deployment on small devices. Therefore, many works were proposed + +![](images/d7761eaccf626bc2de487665c109c6ad2c175055bbf29730ba6b23092fa0676b.jpg) +Figure 2. Visualization of intermediate features, above the dotted line is the pipeline of BSRN [30] and green cuboids represent the ESDB modules. Outi represents the output of the $i^{th}$ ESDB module. Outi = xi = xi-1 + yi. + +[6, 20, 44] to reduce the computational complexity. LapSRN [26] removed the bicubic interpolation pre-processing step. CARN [1] introduced group convolution and cascading residual blocks. IMDN [21] used channel splitting operations to complete information multi-distillation and selective feature fusion. After that, improving information distillation structure has become an important branch of lightweight SR network design, and many methods [9, 30, 34] were proposed. Then, PAN [66] introduced pixel attention to measure pixel contribution to achieve competitive performance by using fewer parameters. Recently, a series of transformer-based SR methods [4, 7, 11, 29, 31, 35, 48, 51] emerge with superior performance. HNCT [7] cleverly combined CNN and transformer to reduce the parameters while ensuring network performance. CFIN [29] design a novel cross-receptive field-guided transformer to adaptively select the required contextual information. NGswin [4] introduced the N-Gram context to the low-level vision with transformers and achieve superior performance. + +# 3. Methodology + +# 3.1. Motivation + +First, we define the phenomenon of network inertia as a situation where the input and output of a network module are similar, indicating that the module fails to learn meaningful features. As shown in Fig. 2, we find that the FAM $(x_{i - 1} + y_i)$ easily causes network inertia because it retains many previous input low-level features $(x_{i - 1})$ information. There are 8 ESDBs with this addition mode in BSRN, we can find that the outputs of most adjacent ESDB modules are similar, especially Out2 and Out3, Out4 and Out5. In addition, we replaced Out5 with Out4 in the subsequent feature fusion and were surprised to find that this replacement had almost no impact on the image reconstruction quality, + +![](images/dec7bb17eaaad59bff1be289a055df82da563247f55faa124b69bedb05098fa3.jpg) +Figure 3. Exemplary filter weights of ordinary convolution and BSConv. One filter kernel of ordinary convolution size $M \times K \times K = 64 \times 3 \times 3$ is visualized and split into 64 images. Filter kernels of pointwise convolution size $N \times M \times K \times K = 64 \times 64 \times 1 \times 1$ and two channels are visualized. + +and the PSNR only decreased by less than $0.005\mathrm{dB}$ . The similarity between features Out4 and Out5 can be inferred that the features $y_{4}$ extracted by the $4^{th}$ ESDB module are meaningless and network inertia occurs. + +Second, as shown in the left of Fig. 3, although the weights of each kernel are highly correlated along the channel dimension, differences also exist. Besides, BSConv [12] also proved that not all the slices (52/128 for the ResNet-50 example) of the kernel are highly correlated along the depth axis. However, as shown in the right of Fig. 3, we find that the pointwise convolution of BSConv associates almost all the channel dimensions along the depth axis (the darker the color, the higher the correlation), which results in introducing many useless or disruptive features on some channels and is not conducive to effective extraction and propagation of useful features. + +# 3.2. Entropy Delta Extraction Link + +In this section, we will explain the definition of FDEL in detail. + +In the network with stacked RBs, as shown in Fig. 5 (a), the initial shallow (low-level) features $x_{1}$ will be superimposed and retained until the last layer $x_{L}$ , as follows + +$$ +\begin{array}{l} x _ {2} = \mathcal {F} _ {1} (x _ {1}) + x _ {1}, \dots , \\ x _ {l} = \mathcal {F} _ {l - 1} \left(x _ {l - 1}\right) + x _ {l - 1}, \dots , \\ x _ {L} = \mathcal {F} _ {L - 1} \left(x _ {L - 1}\right) + x _ {L - 1}, \tag {1} \\ \end{array} +$$ + +where $\mathcal{F}_l$ represents the operations of the $l^{th}$ RB. Equation (1) can be regarded as a feature superposition operation. It may let the feature representation $x_{l}$ of $l^{th}$ RB rely too much on the input features $x_{l-1}$ , which will limit the learning direction of features and even lead to stagnation in feature evolution. This phenomenon can be evidenced in Fig. 2. In addition, the superposition of many low-level input features is not conducive to the presentation of high-level features, especially in structure Fig. 5 (c). + +In view of the aforementioned limitations of the FAM of the RBs, we rethink its design principles and propose an FDEL that can tackle the above issues by flipping the FAM, + +![](images/d848d73b83085bfa83047f528a3e6d3dced69e420931370cf8ce81b0da3b87ef.jpg) +Shallow Feature Extraction + +![](images/897bedc20e7af603ed40c95437972f73aeed8e8d712f5b86969da3017ecaf5f5.jpg) +Deep Feature Extraction + +![](images/df2c9f5c11a7f62af36792682f8a3c516e8c72a278e104224bb021981ed55064.jpg) +Figure 4. The overall architecture of the proposed SLVR and the structure of BCRB and blueprint controllable convolutions (B2Conv). + +![](images/ca6940c310b64355605ecc64992b8d961aa54c0599363c750b3a30b692fbc39d.jpg) +Image Reconstruction + +the structure of RB with FDEL is shown in Fig. 5 (b). Our FDEL aims to remove excessive low-level features. The specific mathematical formula of FDEL is + +$$ +x _ {l + 1} = \mathcal {F} _ {\mathrm {F D E L}} \left(x _ {l}, \mathcal {F} _ {l} \left(x _ {l}\right)\right), \tag {2} +$$ + +where $\mathcal{F}_{\mathrm{FDEL}}$ denotes the operations of the FDEL. Our FDEL uses the output of the $\mathcal{F}_l$ minus its input. Because the output feature representation is more high-level than the input, each RB can effectively extract the different high-level features, which promotes the diverse evolution of features. As we all know, we usually refer to the output of a function minus its input as the gain of the function in a physical sense. Therefore, in an RB with FDEL, the gain is the high-level features extracted by this RB. + +# 3.3. Blueprint Controllable Convolutions + +Although BSConv [12] demonstrates the importance of intra-kernel correlations in separating ordinary convolution, it does not consider the differences of the filters in the depth-axis. In other words, in some channel dimensions, this blueprint cannot obtain the original filter of the ordinary convolution through scale transformation, so filters after forced scale transformation don't effectively work. Therefore, we sparse the intra-kernel correlation and proposed a new B2Conv. Our B2Conv can adaptively pick appropriate channels along the depth axis to perform blueprint scale transformation, which can promote the extraction and efficient propagation of feature information. + +In a convolution operation, assume that the input is $U\in \mathbb{R}^{M\times H\times W}$ , the output is $V\in \mathbb{R}^{N\times H\times W}$ , and the con + +volution kernels are $\{E^{(1)}, E^{(2)}, \dots, E^{(N)}\}$ and $E^{(n)} \in \mathbb{R}^{M \times H \times W}$ . Then the process can be expressed as + +$$ +V _ {n, \therefore :} = U * E ^ {(n)}, \tag {3} +$$ + +where $*$ denotes the convolution operation and $n \in \{1,2,\dots,N\}$ , the weights of these $N$ kernels can be learned during the training. BSCov proved that in practice these weights usually converge to a state that shows a substantial amount of correlation. Therefore, BSCov proposes to decompose an ordinary convolution into the interaction of a blueprint $\mathcal{B}$ and a set of scale-transforming weights $\{w_{n,1}, w_{n,2}, \dots, w_{n,M}\}$ , as follows + +$$ +E _ {m,:,:} ^ {(n)} = w _ {n, m} \cdot \mathcal {B} ^ {(n)}, \tag {4} +$$ + +where $m \in \{1, 2, \dots, M\}$ and $n \in \{1, 2, \dots, N\}$ (see Fig. 6, left column). However, in ordinary convolution, not all the filters of channel dimension can be translated by Equation (4). Therefore, to solve the problem, we propose to add an access control unit (ACU) before scale-transforming. Specifically, in all channel dimensions, if there is a strong intra-kernel correlation with $\mathcal{B}$ , we perform Equation (4), otherwise we do not perform any operation or multiply Equation (4) by a small correlation coefficient to ensure that the useless features extracted in this dimension will not affect the efficient propagation of features, as follows + +$$ +E _ {m,:,:} ^ {(n)} = \mathcal {T} _ {n, m} \cdot w _ {n, m} \cdot \mathcal {B} ^ {(n)}, \tag {5} +$$ + +where $\mathcal{T}_{n,m}$ denotes the output of the ACU, ideally its value is 0 or 1. 1 means allowing scale-transforming and 0 means + +![](images/eafb0e1ff1362b418cfe31a692e23fea81d0a32eeaff477aeef25ea7312bda59.jpg) +Figure 5. Comparison of residual blocks (RBs) with addition mode and FDEL. For the specific mathematical formula of FDEL, refer to Equation (2). + +![](images/d8c83baf08f6cd185f067ef4ef29325385f94c2c00731d6a4be8b20a7422229b.jpg) +Figure 6. Interpretations for BSConv and B2Conv. B2Conv can adaptively pick correct intra-kernel correlations. + +the opposite Here, to make the function continuous and differentiable, we soften the output of the ACU and limit its value range to $(0,1)$ . The closer the $\mathcal{T}_{n,m}$ is to 1, the stronger the intra-kernel correlation. The characteristics of features can reflect the characteristics of the convolution kernels. Therefore, $\mathcal{T}_{n,m}$ can be obtained in an embedding space by feature space mapping and weights of the $\mathcal{T}_{n,m}$ and $w_{n,m}$ can be learned directly via backpropagation. Specifically, we first rewrite the Equation (3) by split 3d convolution into a set of $M$ 2d convolutions, yielding + +$$ +V _ {n,:,:} = U * E ^ {(n)} = \sum_ {m = 1} ^ {M} \left(U _ {m,:,:} * E _ {m,:,:} ^ {(n)}\right). \tag {6} +$$ + +Secondly, we replace the ordinary convolution $E_{m,\cdot,\cdot}^{(n)}$ with the B2Conv representation as given in Equation (5), as follows + +$$ +V _ {n, \therefore :} = \sum_ {m = 1} ^ {M} \left(U _ {m, \therefore :} * \left(\mathcal {T} _ {n, m} \cdot w _ {n, m} \cdot \mathcal {B} ^ {(n)}\right)\right). \tag {7} +$$ + +Because each blueprint $\mathcal{B}^{(n)}$ is independent in the channel dimension. So, we can rewrite the Equation (7) (Note: As shown in right column of Fig. 6, the input $U_{m,:;}$ also needs to be copied $(\bar{U}_{m,:;})$ for feature space mapping) + +$$ +V _ {n,:,:} = \left(\underbrace {\sum_ {m = 1} ^ {M} \left(\bar {U} _ {m , :,:} \cdot \mathcal {T} _ {n , m}\right)} _ {\bar {V} _ {n,:,:} ^ {\prime}} \cdot \underbrace {\sum_ {m = 1} ^ {M} \left(U _ {m , :,:} \cdot w _ {n , m}\right)} _ {V ^ {\prime} _ {n,:,:}}\right) * \mathcal {B} ^ {(n)}. \tag {8} +$$ + +Then, we further rearrange $\{w_{n,1}, w_{n,2}, \dots, w_{n,M}\}$ and $\{\mathcal{T}_{n,1}, \mathcal{T}_{n,2}, \dots, \mathcal{T}_{n,M}\}$ into a $M \times 1 \times 1$ arrays $\tilde{w}_n$ and $\tilde{\mathcal{T}}_n$ respectively, the sum operations of Equation (8) can be replaced by convolution operations, as follows + +$$ +\bar {V} _ {n,:,:} ^ {\prime} = U * \tilde {\mathcal {T}} _ {n}, \tag {9} +$$ + +$$ +V _ {n,:,:} ^ {\prime} = U * \tilde {w} _ {n}, \tag {10} +$$ + +![](images/d3764e97daef199ebcdf6f58ead41ed5c9ad2ecc015e71c42e31d326f3f2e30d.jpg) + +$$ +V _ {n,:,:} = \left(\sigma \left(\bar {V} _ {n,:,:} ^ {\prime}\right) \cdot V _ {n,:,:} ^ {\prime}\right) * \mathcal {B} ^ {(n)}, \tag {11} +$$ + +where $\sigma$ denotes the sigmoid function, which limits $\tilde{V}_{n,::}^{\prime}$ to (0,1). For a concrete implementation, Equation (9) and Equation (10) can be translated into a $1\times 1$ pointwise convolution with the kernels $\{\tilde{w}_1,\tilde{w}_2,\dots ,\tilde{w}_N\}$ and $\left\{\tilde{\mathcal{T}}_1,\tilde{\mathcal{T}}_2,\dots ,\tilde{\mathcal{T}}_N\right\}$ respectively. The Equation (11) can be translated into a $K\times K$ depthwise convolution with the kernels $\{\mathcal{B}^{(1)},\mathcal{B}^{(2)},\dots ,\mathcal{B}^{(N)}\}$ . + +# 3.4. The Overall Structure + +Based on FDEL and B2Conv, we further develop a lightweight SLVR framework to achieve high-performance image SR. As shown in Fig. 4, SLVR consists of three parts, including shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given a LR input $I_{\mathrm{LR}} \in \mathbb{R}^{H \times W \times C_{in}}$ , we first use a B2Conv $\mathcal{H}_{\mathrm{B2Conv}}$ to extract shallow feature $X_0 \in \mathbb{R}^{H \times W \times C_{mid}}$ , as follows + +$$ +X _ {0} = \mathcal {H} _ {\mathrm {B 2 C o n v}} \left(\operatorname {C o n c a t} \left[ I _ {\mathrm {L R}} \right] _ {n}\right), \tag {12} +$$ + +where $C_{in}$ and $C_{mid}$ denote the channel number of the input and shallow feature. $\text{Concat}[\cdot]_n$ means concatenating $n$ times input. Then we use $K$ efficient controllable distillation block (ECDB) to extract deep features. To maximize the effect of FDEL, feature aggregation is performed by DFAS and GELU [15] activation to enhance feature representation. After that, we use a B2Conv to continue refining the features, the above process can be formulated as + +$$ +X _ {i} = \mathcal {H} _ {\mathrm {E C D B} _ {i}} \left(X _ {i - 1}\right), i = 1, 2, \dots , K, \tag {13} +$$ + +$$ +X _ {\mathrm {A g g}} = \mathcal {H} _ {\mathrm {D F A S}} \left(\left(w _ {1}, w _ {2}, \dots , w _ {c}\right), \left(X _ {1}, X _ {2}, \dots , X _ {K}\right)\right), \tag {14} +$$ + +$$ +X _ {\text {R e f i n e d}} = \mathcal {H} _ {\mathrm {B 2 C o n v}} \left(\operatorname {G E L U} \left(X _ {\mathrm {A g g}}\right)\right), \tag {15} +$$ + +where $\mathcal{H}_{\mathrm{ECDB}_i}$ represents the $i^{\mathrm{th}}$ ECDB, $w_{1}, w_{2}, \dots, w_{c}$ are the learnable weights, $X_{1}, X_{2}, \dots, X_{K}$ denote the intermediate features. $\mathcal{H}_{\mathrm{DFAS}}(\cdot)$ denotes the operations of the DFAS, which can be achieved by a convolution. The ECDB is designed based on the ESDB of the BSRN. Specifically, first, we replace the BSRB with the blueprint controllable residual block (BCRB). Second, we replace the + +
MethodT-DatasetScale#ParamsMult-AddsSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
VDSR [22]291×2666K612.6G37.530.958733.030.912431.900.896030.760.914037.220.9750
CARN [1]D2K+291×21,592K222.8G37.760.959033.520.916632.090.897831.920.925638.360.9765
IMDN [21]D2K×2694K158.8G38.000.960533.630.917732.190.899632.170.928338.880.9774
LAPAR-A [27]DF2K×2548K171G38.010.960533.620.918332.190.899932.100.928338.670.9772
LatticeNet [37]D2K×2756K169.5G38.060.960733.700.918732.200.899932.250.928838.940.9774
RLFN [24]D2K×2527K97.9G38.070.960733.720.918732.220.900032.330.9299--
SRPN-Lite [65]DF2K×2609K139.9G38.100.960833.700.918932.250.900532.260.9294--
FMEN [6]DF2K×2748K172.0G38.100.960933.750.919232.260.900732.410.931138.950.9778
MLRN [10]D2K×2488k90.4G38.070.960733.590.918032.210.900032.280.929738.760.9773
NGswin [4]D2K×2998K140.4G38.050.961033.790.919932.270.900832.530.932438.970.9777
SLVR (Ours)D2K×2464K102G38.140.961033.870.920232.280.900932.570.931839.080.9776
VDSR [22]291×3666K612.6G33.660.921329.770.831428.820.797627.140.827932.010.9340
CARN [1]D2K+291×31,592K118.8G34.290.925530.290.840729.060.803428.060.849333.500.9440
IMDN [21]D2K×3703K71.5G34.360.927030.320.841729.090.804628.170.851933.610.9445
LAPAR-A [27]DF2K×3594K114G34.360.926730.340.842129.110.805428.150.852333.510.9441
LatticeNet [37]D2K×3765K76.3G34.400.927230.320.841629.100.804928.190.851333.630.9442
RFDN [34]D2K×3633K65.6G34.470.928030.350.842129.110.805328.320.854733.780.9458
SRPN-Lite [65]DF2K×3615K62.7G34.470.927630.380.842529.160.806128.220.8534--
FMEN [6]DF2K×3757K77.2G34.450.927530.400.843529.170.806328.330.856233.860.9462
MLRN [10]D2K×3496K40.9G34.460.926730.350.842629.100.805428.200.853333.660.9450
NGswin [4]D2K×31,007K66.6G34.520.928230.530.845629.190.807828.520.860333.890.9470
SLVR (Ours)D2K×3473K45.3G34.580.928530.530.846029.210.808028.560.860334.070.9475
VDSR [22]291×4666K612.6G31.350.883828.010.767427.290.725125.180.752428.830.8870
CARN [1]D2K+291×41,592K90.9G32.130.893728.600.780627.580.734926.070.783730.470.9084
IMDN [21]D2K×4715K40.9G32.210.894828.580.781127.560.735326.040.783830.450.9075
LAPAR-A [27]DF2K×4659K94G32.150.894428.610.781827.610.736626.140.787130.420.9074
LatticeNet [37]D2K×4777K43.6G32.180.894328.610.781227.570.735526.140.784430.540.9075
RFDN [34]D2K×4550K23.9G32.240.895228.610.781927.570.736026.110.785830.580.9089
RLFN [24]D2K×4543K25.3G32.240.895228.620.781327.600.736426.170.7877--
SRPN-Lite [65]DF2K×4623K35.8G32.240.895828.690.783627.630.737326.160.7875--
BSRN [30]DF2K×4352K19.4G32.350.896628.730.784727.650.738726.270.790830.840.9123
FMEN [6]DF2K×4769K44.2G32.240.895528.700.783927.630.737926.280.790830.700.9107
MLRN [10]D2K×4507K23.5G32.300.895628.620.782427.570.736526.100.786730.560.9092
HPUN-M [48]D2K×4511K27.7G32.240.895028.660.782827.600.737126.120.787830.550.9089
SAFMN [49]DF2K×4240K14.0G32.180.894828.600.781327.580.735925.970.780930.430.9063
OSFFNet [53]DF2K×4537K22.0G32.390.897628.750.785227.660.739326.360.795030.840.9125
NGswin [4]D2K×41,019K36.4G32.330.896328.780.785927.660.739626.450.796330.800.9128
SLVR-S (Ours)DF2K×4158K8.5G32.200.895228.650.782227.570.736126.080.785430.530.9084
SLVR (Ours)D2K×4485K26.2G32.400.897528.780.785027.650.738826.410.795130.930.9133
SLVR* (Ours)DF2K×4485K26.2G32.440.898028.850.786527.700.740026.580.800031.100.9141
+ +Table 1. Quantitative comparison (PSNR/SSIM) for lightweight SR with state-of-the-art (SOTA) methods on five benchmark datasets. The best and second-best results are marked in red and blue colors respectively. Mult-Adds is evaluated on a ${1280} \times {720}\mathrm{{HR}}$ image. T-Dataset denotes training dataset. D2K stands for the DIV2K dataset. DF2K indicates a merged dataset containing ${800} + 2,{650}\mathrm{{HR}} - \mathrm{{LR}}$ image pairs. 291 images dataset is from [58]. + +addition mode of the residual structure with the proposed FDEL. Third, all the BSConvs are replaced with the proposed B2Conv. Please see Fig. 4 for detailed changes. After performing the above operations, we obtain the final refined features $X_{\mathrm{Refined}}$ . Finally, to take advantage of residual learning, we use long skip connections to reduce the difficulty of learning. SR results can be obtained by + +$$ +I _ {\mathrm {S R}} = \mathcal {H} _ {\text {U p s a m p l e}} \left(X _ {\text {R e f i n e d}}\right) + \text {B i c u b i c} ^ {\uparrow} \left(I _ {\mathrm {L R}}\right), \tag {16} +$$ + +where $\mathcal{H}_{\mathrm{Upsample}}(\cdot)$ represents the upsampling operation, which is implemented with the pixelshuffle operation [44]. Bicubic denotes bicubic upsampling. $L_{1}$ loss function is used to optimize the network, as follows + +$$ +L _ {1} = \left\| I _ {\mathrm {H R}} - I _ {\mathrm {S R}} \right\| _ {1}, \tag {17} +$$ + +where $I_{\mathrm{HR}}$ is the HR label. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets and Metrics. We use DIV2K [50] dataset as the training dataset, which contains 800 HR-LR image pairs. To show more fair results, we used DF2K (DIV2K + Flickr2K [50]) to additionally train an $\times 4$ model. Note that the model trained with DF2K is marked with a small *. For testing, we evaluated the performances on the five benchmark datasets, including Set5 [2], Set14 [61], B100 [39], Urban100 [19], Manga109 [40]. The PSNR and SSIM [54] are used as the evaluation metrics, which are evaluated on the Y channel of the YCbCr space. + +Implementation Details. ECDB number is set to 8, and the channel number is set to 64. The kernel size of the depth- + +![](images/e449684b2a1efe1ed3fad976449cfc9a9417a2e3f50d3b6fa33280586af865d4.jpg) +Figure 7. Visual comparison for $\times 4$ SR methods. The patches for comparison are marked with red boxes. + +wise convolutions is set to 3. We additionally built a smaller model SLVR-S to compare the latest methods with less than 200K parameters and it trained on DF2K. The ECDB number of the SLVR-S is set to 6 and the channel number is set to 40. During training, we augment the data with random horizontal flips and rotation $(90^{\mathrm{o}},180^{\mathrm{o}},270^{\mathrm{o}})$ . We randomly crop LR patches of size $64\times 64$ as input. Our models are trained by Adam optimizer [23] for 1000K iterations. The initial learning rate is set to $1\times 10^{-3}$ and halved for every 200k iterations. The batch size of SLVR and SLVR-S is set to 64 and 128 respectively. LR image patches are obtained by the bicubic kernel from corresponding HR images matching each SR task. Our networks are trained on an RTX3090 GPU. Note that no other training skills (e.g., pre-training [28]) are employed. + +# 4.2. Comparison with the SOTA SR methods + +To evaluate the effectiveness of SLVR, we compare our model with other efficient SR models on scale factors $\times 2,\times 3,\times 4$ including VDSR (CVPR16) [22], CARN (ECCV18) [1], IMDN (ACMMM19) [21], RFDN (ECCV20) [34], LAPAR-A (NeurIPS20) [27], LatticeNet (ECCV20) [37], RLFN (CVPRW22) [24], SRPN-Lite (ICLR22) [65], BSRN (CVPRW22) [30], FMEN (CVPRW22) [6], MLRN (CVPRW23) [10], NGswin (CVPR23) [4], HPUN-M (AAAI23) [48], SAFMN (ICCV23) [49], OSFFNet (AAAI24) [53]. + +As shown in Table 1, we reported the training dataset, the parameters, Mult-Adds, and PSNR and SSIM on five benchmarks. Compared with models with similar param + +![](images/63bb4782ea05068379bb2f5b9913756429bfee69690d89b860e94e1416f25845.jpg) +Figure 8. Ablation studies on B2Conv. BSRN-B2Conv denotes the BSCnv of BSRN is replaced by B2Conv. All the results are reported on scale factor $\times 4$ + +![](images/22385b29953ecde2fa7b60f12de3d0b1e1135ac5a0cb26069832c80b2b97bd5e.jpg) + +
MethodFDELB2ConvSet5Set14Urban100
Model-I✓×32.37/0.896928.71/0.783926.30/0.7917
Model-II×✓32.39/0.897528.76/0.785126.40/0.7943
SLVR✓✓32.40/0.897528.78/0.785026.41/0.7951
+ +Table 2. Ablation studies of FDEL and B2Conv on scale factor $\times 4$ . The $\times$ of FDEL denotes it is replaced by addition mode, and the $\times$ of B2Conv denotes it is replaced by BSConv. + +eters and Mult-Adds, the performance of our SLVR surpasses existing methods with a notable margin on all benchmarks. Compared with the latest transformer-based method NGswin, although the number of parameters of our SLVR is less than $50\%$ of NGswin and the number of Mult-Adds is less than $75\%$ of NGswin, our SLVR still has an advantage in performance. Compared with the latest extremely lightweight method SAFMN trained on the same dataset, our SLVR-S obtains better performance with a smaller model size. + +We also perform a visual comparison of different SOTA lightweight SR methods on scale factor $\times 4$ . As shown in Fig. 7, we can observe that the SR results generated by SLVR contain more correct and fine-grained textures, while other methods generate many blurred edges and artifacts. For example, in the second row of Fig. 7, our SLVR can restore the texture of the fence, while other methods cannot. We think that the visual results can prove the effectiveness of the proposed B2conv, which can prevent incorrect intrakernel correlations to reduce useless or disruptive feature extraction and propagation. + +# 4.3. Ablation Studies + +The core contributions of our framework are the proposed FDEL and B2Conv. To demonstrate their effectiveness, we first conduct ablation studies on the main framework of our SLVR. Specifically, we remove FDEL and B2Conv and replace them with the original linking mode FAM of RBs and BSCnv, respectively. The experimental results are reported in Table 2. We can observe that our FDEL and B2Conv can improve network performance, especially B2Conv, improved by 0.11 dB on the dataset Urban100. + +
MethodScale#ParamsMult-AddsSet5Set14Urban100
EDSR-baseline×41,518K114.0G32.08/0.893828.58/0.781025.98/0.7824
EDSR+FDEL1,518K114.0G32.12/0.894428.61/0.781926.05/0.7845
+ +Table 3. Ablation studies of FDEL. EDSR+FDEL denotes that all addition mode of EDSR-baseline [32] are replaced by FDEL. + +
RFDNScale#ParamsMult-AddsSet5Set14Urban100
w BSConv×3197K25.8G34.37/0.926830.34/0.841828.06/0.8503
w B2Conv261K34.5G34.44/0.927330.36/0.842628.18/0.8525
w Conv633K65.6G34.47/0.928030.35/0.842128.21/0.8525
w BSConv×4210K14.9G32.12/0.894128.59/0.780626.01/0.7829
w B2Conv273K20.1G32.22/0.894928.65/0.782226.06/0.7843
w Conv643K37.4G32.28/0.895728.61/0.781826.11/0.7858
+ +Table 4. Comprehensive comparison between ordinary convolution (Conv), BSCnv, and B2Conv. The best and second-best results are marked in red and blue colors. + +![](images/506680e4a9f5f9174dc4d9063b855981d7cbfaaee447f4420066fadeadb79f66.jpg) +Figure 9. Visual comparison of feature maps between SLVR and SLVR without FDEL. Outi represents the output of the $i^{th}$ ECDB module. + +Second, to verify that FDEL can be used as a more efficient linking mode of RBs, we use FDEL to simply replace the traditional linking mode FAM in the classic residual architecture EDSR [32]. From Table 3, we can conclude that our FDEL is more effective than the traditional FAM without additional parameters and Mult-Adds consumed. + +Third, to further verify the effectiveness of our proposed B2Conv and confirm that B2Conv can be used as highly efficient building blocks. We denote BSConv of BSRN [30] replaced by B2Conv as BSRN-B2Conv. As shown in Fig. 8, we can observe that our B2Conv can significantly improve model performance without changing model architecture. Besides, we also conducted a comprehensive comparison between ordinary convolution, BSConv, and B2Conv on the network architecture RFDN [34], as shown in Table 4, compared with BSConv, our B2Conv greatly improves the performance (e.g., $\mathbf{0.1dB}$ on Set5 × 4) of the RFDN with few additional parameters (63K) consumed, which also illustrates the rationality of our improvement. + +Finally, we perform a visual comparison of features between SLVR and SLVR without FDEL. From Fig. 9, we find that our FDEL can effectively alleviate network inertia. In addition, we visualize the scale transformation state of 8 blueprints on 64 channels. As shown in Fig. 10, we can observe that our B2Conv does not perform scale transformation on all 64 channel dimensions (only on the channel denoted by black squares). It proves that our B2Conv can adaptively select the appropriate channel to perform blueprint scale transformation, which can effectively prevent the introduction of disruptive or useless features. Fur + +![](images/6b0828daebd3fef597bf61b13377a90184306c2ed8c3cafd8aa442b65bac927d.jpg) +Figure 10. Visualization of scale transformation state of 4 blueprints on 64 channels. Black squares denote strong intrakernel correlation, and white squares are the opposite. + +![](images/6f7d271da91ed03df7e1cfeb0cf79ad79fa669ab51ea0a224c836e787a9d98c8.jpg) + +![](images/e1680436f1d224796df161d49cbc9137cbbf193184342ab4293b6445d912df51.jpg) + +![](images/2b954807791b6157a792563a0699127988680707459f18732e11e193109dfa09.jpg) + +![](images/627c1e80b9ef302746cedc93293099dd4fa2bb0362c8df65d026d6fd404b71ff.jpg) +Figure 11. Visual comparison of intermediate feature maps between SLVR and SLVR without B2Conv. + +thermore, we perform a visual comparison of features between SLVR and SLVR without B2Conv. From Fig. 11, we can find that the intermediate features generated by the SLVR without B2Conv contain more interference and disruptive features, resulting in poorer feature representation. The intermediate features generated by our SLVR are more accurate and have better feature representation ability. It also indirectly proves that the proposed B2Conv can adaptively select the correct intra-kernel correlation, prevent the introduction of disruptive features, and is more conducive to feature propagation and representation. + +# 5. Conclusion + +In this paper, we propose blueprint controllable convolutions (B2Conv). B2Conv can precisely prevent incorrect intra-kernel correlation, which effectively reduces the extraction of useless or disruptive features and promotes the effective propagation of feature information. Besides, we propose a novel linking mode FDEL for RBs. Our FDEL reduces the retention of the input low-level features and can promote the diversity evolution of high-level features, which can effectively alleviate the occurrence of network inertia. Finally, based on B2Conv and FDEL, we propose an efficient lightweight SR framework SLVR. Furthermore, our B2Conv and FDEL can also serve as highly efficient components to improve the performance of other CNN-based models (such as EDSR, BSRN, and RFDN). Extensive experiments on public benchmark datasets show the prominent SR performance of SLVR and comprehensive analytical studies validate the reasonableness and effectiveness of the proposed framework and components. + +# References + +[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, pages 252-268, 2018. 1, 3, 6, 7 +[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6 +[3] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Yan Shuicheng, and Ji-ashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In ICCV, pages 3434–3443, 2019. 2 +[4] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image superresolution. In CVPR, pages 2071–2081, 2023. 3, 6, 7 +[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 2 +[6] Zongcai Du, Ding Liu, Jie Liu, Jie Tang, Gangshan Wu, and Lean Fu. Fast and memory-efficient network towards efficient image super-resolution. In CVPRW, pages 852-861, 2022. 3, 6, 7 +[7] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In CVPRW, pages 1102-1111, 2022. 3 +[8] Mengyang Feng, Huchuan Lu, and Yizhou Yu. Residual learning for salient object detection. IEEE TIP, 29:4696-4708, 2020. 2 +[9] Guangwei Gao, Wenjie Li, Juncheng Li, Fei Wu, Huimin Lu, and Yi Yu. Feature distillation interaction weighting network for lightweight image super-resolution. In AAAI, pages 661-669, 2022. 1, 3 +[10] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. Mixer-based local residual network for lightweight image super-resolution. In CVPRW, pages 1593-1602, 2023. 6, 7 +[11] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. A simple transformer-style network for lightweight image super-resolution. In CVPRW, pages 1484-1494, 2023. 3 +[12] Daniel Haase and Manuel Amthor. Rethinking depthwise separable convolutions: How intra-kernel correlations lead to improved mobilenets. In CVPR, pages 14588-14597, 2020. 2, 3, 4 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, pages 630-645. Springer, 2016. 2 +[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolu + +tional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2 +[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2 +[18] Gao Huang, Shichen Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Condensenet: An efficient densenet using learned group convolutions. In CVPR, pages 2752-2761, 2018. 2 +[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 6 +[20] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, pages 723-731, 2018. 3 +[21] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, pages 2024-2032, 2019. 1, 3, 6, 7 +[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 1, 6, 7 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[24] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPRW, pages 765-775, 2022. 6, 7 +[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 2 +[26] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, pages 5835-5843, 2017. 3 +[27] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. NeurIPS, 33:20343-20355, 2020. 6, 7 +[28] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 7 +[29] Wenjie Li, Juncheng Li, Guangwei Gao, Weihong Deng, Jiantao Zhou, Jian Yang, and Guo-Jun Qi. Cross-receptive focused inference network for lightweight image superresolution. IEEE TMM, pages 1-13, 2023. 3 +[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPRW, pages 832-842, 2022. 1, 3, 6, 7, 8 +[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV, pages 1833-1844, 2021. 3 +[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 1132-1140, 2017. + +[33] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, pages 85-100, 2018. 2 +[34] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, pages 41-55. Springer, 2020. 3, 6, 7, 8 +[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPRW, pages 456-465, 2022. 3 +[36] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 2 +[37] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, pages 272-289. Springer, 2020. 6, 7 +[38] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, pages 116-131, 2018. 2 +[39] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423 vol.2, 2001. 6 +[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 6 +[41] Marcelo Gennari Do Nascimento, Victor Prisacariu, and Roger Fawcett. *Dsconv: Efficient convolution operator. In ICCV*, pages 5147-5156, 2019. 2 +[42] Tobias Pohlen, Alexander Hermans, Markus Mathias, and Bastian Leibe. Full-resolution residual networks for semantic segmentation in street scenes. In CVPR, pages 3309-3318, 2017. 2 +[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 2 +[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 3, 6 +[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[46] Ayan Sinha, Asim Unmesh, Qixing Huang, and Karthik Ramani. Surfnet: Generating 3d shape surfaces using deep residual networks. In CVPR, pages 791-800, 2017. 2 +[47] Rupesh K Srivastava, Klaus Greff, and Jürgen Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2 +[48] Bin Sun, Yulun Zhang, Songyao Jiang, and Yun Fu. Hybrid pixel-unshuffled network for lightweight image superresolution. In AAAI, pages 2375-2383, 2023. 3, 6, 7 +[49] Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Spatially-adaptive feature modulation for efficient image + +super-resolution. In ICCV, pages 13190-13199, 2023. 6, 7 +[50] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 1110-1121, 2017. 6 +[51] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In CVPR, pages 22378-22387, 2023. 3 +[52] Xijun Wang, Meina Kan, Shiguang Shan, and Xilin Chen. Fully learnable group convolution for acceleration of deep neural networks. In CVPR, pages 9041–9050, 2019. 2 +[53] Yang Wang and Tao Zhang. Osffnet: Omni-stage feature fusion network for lightweight image super-resolution. In AAAI, pages 5660-5668, 2024. 6, 7 +[54] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6 +[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 2 +[56] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023. 1 +[57] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 5987-5995, 2017. 2 +[58] Jianchao Yang, John Wright, Thomas S. Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE TIP, 19(11):2861-2873, 2010. 6 +[59] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, pages 4470-4479, 2019. 2 +[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 2 +[61] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 6 +[62] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. Interleaved group convolutions. In ICCV, pages 4383-4392, 2017. 2 +[63] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 2 +[64] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2 +[65] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 6, 7 +[66] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV, pages 56-72. Springer, 2020. 3 + +[67] Daquan Zhou, Qibin Hou, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Rethinking bottleneck structure for efficient mobile network design. In ECCV, pages 680-697. Springer, 2020. 2 \ No newline at end of file diff --git a/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/images.zip b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d6341a0be3667a4d404a5cf0d0b70ea2dec7b6b --- /dev/null +++ b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e14a7303358fc0361e8af5fc1a6789fe2d5a29c83e53c3ac476d2ef8b5e1a6d +size 1017351 diff --git a/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/layout.json b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6860dc860caa1b4c9aa072a008009851669aba71 --- /dev/null +++ b/2025/SLVR_ Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation/layout.json @@ -0,0 +1,9419 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "text", + "content": "SLVR: Super-Light Visual Reconstruction via Blueprint Controllable Convolutions and Exploring Feature Diversity Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 208, + 161, + 250, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 161, + 250, + 175 + ], + "spans": [ + { + "bbox": [ + 208, + 161, + 250, + 175 + ], + "type": "text", + "content": "Ning Ni" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 175, + 507, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 175, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 104, + 175, + 507, + 190 + ], + "type": "text", + "content": "School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 174, + 192, + 294, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 192, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 174, + 192, + 294, + 202 + ], + "type": "text", + "content": "ningni@mail.bnu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 338, + 162, + 406, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 162, + 406, + 175 + ], + "spans": [ + { + "bbox": [ + 338, + 162, + 406, + 175 + ], + "type": "text", + "content": "Libao Zhang*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 192, + 433, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 192, + 433, + 203 + ], + "spans": [ + { + "bbox": [ + 318, + 192, + 433, + 203 + ], + "type": "text", + "content": "libaozhang@bnu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 605 + ], + "type": "text", + "content": "Recently, improving the residual structure and designing efficient convolutions have become important branches of lightweight visual reconstruction model design. We have observed that the feature addition mode (FAM) in existing residual structure tends to lead to slow feature learning or stagnation in feature evolution, a phenomenon we define as network inertia. In addition, although blueprint separable convolutions (BSConv) have proved the dominance of intrakernel correlation, BSCov forces the blueprint to perform scale transformation on all channels, which may lead to incorrect intra-kernel correlation and introduce useless or disruptive features on some channels and hinder the effective propagation of features. Therefore, in this paper, we rethink the FAM and BSCov for super-light visual reconstruction framework design. First, we design a novel linking mode, called feature diversity evolution link (FDEL), which aims to alleviate the phenomenon of network inertia by reducing the retention of previous low-level features, thereby promoting the evolution of feature diversity. Second, we propose blueprint controllable convolutions (B2Conv). The B2Conv can adaptively pick accurate intra-kernel correlation in the depth-axis, effectively preventing the introduction of useless or disruptive features. Based on FDEL and B2Conv, we develop a super-light super-resolution (SR) framework SLVR for visual reconstruction. Both FDEL and B2Conv can serve as efficient plugins. Extensive experimental results demonstrate the effectiveness of our proposed B2Conv, FDEL, and SLVR. Our code will be available at https://github.com/chongningni/SLVR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 646, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 296, + 682 + ], + "type": "text", + "content": "As an important direction in the field of computer vision, visual reconstruction focuses on improving the visual quality of images or videos. Image super-resolution (SR) is" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 555, + 396 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 396 + ], + "type": "image", + "image_path": "5aa3a29a6fe17dccaefc5355ba82a92b438b9f54bb8bcbbacf0861d6dfb2fe39.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. Performance and model complexity comparison on Set14 dataset for upscaling factor " + }, + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "type": "text", + "content": ". * denotes the model trained on DF2K dataset and others trained on DIV2K." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "type": "text", + "content": "a core task in visual reconstruction, which aims to reconstruct high-resolution details from a low-resolution image. SR is an ill-posed problem because a low-resolution (LR) image can be mapped to multiple high-resolution (HR) spaces, which makes it challenging. Recently, convolutional neural networks (CNNs) based on residual learning [1, 9, 21, 22, 30, 56] have dominated the SR field because they can constrain mapping space and reduce the difficulty of SR learning, which creates conditions for lightweight SR design. Furthermore, lightweight convolution has become an important research branch of lightweight model design. In this paper, we aim to rethink the residual structure and blueprint separable convolutions (BSConv) to build a superlight SR reconstruction framework. The issues we discovered are as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "1) Existing feature addition mode easily causes network inertia: SR tasks require recovering detailed textures, which require rich high-level features as support. The existing residual structure uses the feature addition mode (FAM) of input and output " + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "inline_equation", + "content": "(x + \\mathcal{F}(x))" + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": ", retaining many previous low-level features " + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "inline_equation", + "content": "(x)" + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": ", which is not conducive to the diver" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "*Corresponding Author. This work is supported by the National Natural Science Foundation of China under Grant 62271060 and 61571050." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "400" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "sity evolution of features and easily causes network inertia. " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " represents the calculation function. Furthermore, retaining redundant previous low-level features will also affect the presentation of high-level features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 123, + 294, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 123, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 54, + 123, + 294, + 231 + ], + "type": "text", + "content": "2) Incorrect intra-kernel correlation of the BSConv: Although BSConv proves the dominance of intra-kernel correlation, we find that kernel correlation is not present on all channels in an ordinary convolution. However, BSConv forces the blueprint to perform linear transformations on all channels, which may lead to incorrect intra-kernel correlation and result in introducing many useless or disruptive features on some channels to hinder the effective propagation of features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 233, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 296, + 472 + ], + "type": "text", + "content": "To solve the above problems, this work rethinks the residual structure and BSconv. First, to alleviate the network inertia, we propose a new linking mode, named feature diversity evolution link (FDEL). Our FDEL aims to minimize the proportion of low-level features during feature propagation and presentation, which is beneficial to distill rich high-level features with different representations and alleviate network inertia. In addition, to maximize the effect of FDEL, we introduce a diversified feature aggregation strategy (DFAS) to enhance feature representation. Second, to prevent the incorrect inter-kernel correlation of the BSConv (inter-blueprint correlation), we propose new blueprint controllable convolutions (B2Conv). Our B2Conv sets up an access control unit (ACU) in the channel dimension to prevent incorrect intra-blueprint correlation. The ACU is like a guard, which can adaptively pick relevant channels to associate and learn the correct interblueprint correlation. Our B2Conv can reduce the introduction of useless or disruptive features and promote the effective propagation of feature information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 476, + 294, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 476, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 67, + 476, + 294, + 488 + ], + "type": "text", + "content": "The main contributions of this paper are summarized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 491, + 294, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 294, + 574 + ], + "type": "text", + "content": "1) We propose a new convolution B2Conv. Our B2Conv introduces the ACU to adaptively control intra-blueprint correlation, which can precisely prevent the introduction of useless or disruptive features and promote effective propagation of feature information. Besides, the B2Conv can also serve as a highly efficient building block for CNNs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 294, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 294, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 294, + 650 + ], + "type": "text", + "content": "2) We propose a new linking mode FDEL, which effectively alleviates the network inertia problem. Our FDEL aims to reduce the retention of previous low-level feature information and can effectively promote the diverse evolution of features, which is beneficial to the extraction and presentation of high-level features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 714 + ], + "type": "text", + "content": "3) Based on FDEL and B2Conv, we design a novel lightweight SR framework SLVR. Extensive experiments demonstrate that SLVR achieves record high performance (e.g., 28.85dB on Set14 × 4 with only 485K parameters) on lightweight SR benchmarks (shown in Fig. 1)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 553, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 425 + ], + "type": "text", + "content": "Residual Network: Inspired by [47], He et al. [13] proposed first residual network for image recognition and achieved state-of-the-art performance. Then, many different residual networks were proposed [8, 14, 36, 42, 46] for different computer vision tasks. Residual networks are typically built by stacking a sequence of residual blocks. Recently, the residual block was further developed, such as widening the channels [60], introducing attention mechanism [17, 55], and combining with dense connections [64]. The residual network has achieved great success in heavy-weight networks because it overcomes the vanishing gradient problem. Besides, residual learning can reduce the learning difficulty of computer vision tasks, which also provides opportunities for lightweight network design. To reduce the model complexity, MobileNet [16] introduced depthwise separable convolutions as key elements in the residual blocks. Then, MobileNetV2 [43] proposed an inverted residual block where the shortcut connections are between the thin bottleneck layers. ShuffleNetV2 [38] inserts a channel split component before the inverted residual block. After that, Zhou et al. [67] rethought the inverted residual block and proposed a sandglass block, which aims to perform spatial transformation and identity mapping at higher dimensions to alleviate information loss and gradient confusion problems. Although previous works have effectively improved the residual structure, most proposed residual blocks now still use the addition mode of input and output features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "type": "text", + "content": "Efficient Convolution: To reduce network parameters, AlexNet [25] first proposed the concept of group convolution and ResNext [57] demonstrated its effectiveness. After that, the group convolution was further modified by this works [18, 52, 62, 63]. VGG [45] suggested using two " + }, + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "type": "text", + "content": " convolutions instead of one " + }, + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 426, + 553, + 677 + ], + "type": "text", + "content": " convolution. Xception [5] proposed the depthwise separable convolution (DSConv) to further reduce the parameters of convolution and then it was used in MobileNet [16] and gained state-of-the-art results. Partial convolution (PConv) [33] was first proposed for image inpainting, which greatly improved computing efficiency. To solve the unreasonable problem of mask update in the partial convolution, Gated convolution (GConv) [59] was proposed. Octave convolution [3] was proposed to extract the high-frequency and low-frequency signals of the image. Recently, DSConv was decomposed [41] into two components: variable quantization kernels and distribution offsets, thereby achieving memory reduction and operation acceleration. BSConv [12] rethought the DSConv and considered the dominance of intra-kernel correlations when performing convolution operations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Lightweight SR: Although deep models can achieve good SR performance, they are not conducive to deployment on small devices. Therefore, many works were proposed" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "401" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 219 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 219 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 219 + ], + "type": "image", + "image_path": "d7761eaccf626bc2de487665c109c6ad2c175055bbf29730ba6b23092fa0676b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 227, + 295, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 272 + ], + "type": "text", + "content": "Figure 2. Visualization of intermediate features, above the dotted line is the pipeline of BSRN [30] and green cuboids represent the ESDB modules. Outi represents the output of the " + }, + { + "bbox": [ + 55, + 227, + 295, + 272 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 227, + 295, + 272 + ], + "type": "text", + "content": " ESDB module. Outi = xi = xi-1 + yi." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 282, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 521 + ], + "type": "text", + "content": "[6, 20, 44] to reduce the computational complexity. LapSRN [26] removed the bicubic interpolation pre-processing step. CARN [1] introduced group convolution and cascading residual blocks. IMDN [21] used channel splitting operations to complete information multi-distillation and selective feature fusion. After that, improving information distillation structure has become an important branch of lightweight SR network design, and many methods [9, 30, 34] were proposed. Then, PAN [66] introduced pixel attention to measure pixel contribution to achieve competitive performance by using fewer parameters. Recently, a series of transformer-based SR methods [4, 7, 11, 29, 31, 35, 48, 51] emerge with superior performance. HNCT [7] cleverly combined CNN and transformer to reduce the parameters while ensuring network performance. CFIN [29] design a novel cross-receptive field-guided transformer to adaptively select the required contextual information. NGswin [4] introduced the N-Gram context to the low-level vision with transformers and achieve superior performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 138, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 138, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 138, + 545 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 130, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 130, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 130, + 562 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "First, we define the phenomenon of network inertia as a situation where the input and output of a network module are similar, indicating that the module fails to learn meaningful features. As shown in Fig. 2, we find that the FAM " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_{i - 1} + y_i)" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " easily causes network inertia because it retains many previous input low-level features " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_{i - 1})" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " information. There are 8 ESDBs with this addition mode in BSRN, we can find that the outputs of most adjacent ESDB modules are similar, especially Out2 and Out3, Out4 and Out5. In addition, we replaced Out5 with Out4 in the subsequent feature fusion and were surprised to find that this replacement had almost no impact on the image reconstruction quality," + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 149 + ], + "type": "image", + "image_path": "dec7bb17eaaad59bff1be289a055df82da563247f55faa124b69bedb05098fa3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "type": "text", + "content": "Figure 3. Exemplary filter weights of ordinary convolution and BSConv. One filter kernel of ordinary convolution size " + }, + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "type": "inline_equation", + "content": "M \\times K \\times K = 64 \\times 3 \\times 3" + }, + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "type": "text", + "content": " is visualized and split into 64 images. Filter kernels of pointwise convolution size " + }, + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "type": "inline_equation", + "content": "N \\times M \\times K \\times K = 64 \\times 64 \\times 1 \\times 1" + }, + { + "bbox": [ + 313, + 158, + 555, + 213 + ], + "type": "text", + "content": " and two channels are visualized." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": "and the PSNR only decreased by less than " + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "inline_equation", + "content": "0.005\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": ". The similarity between features Out4 and Out5 can be inferred that the features " + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "inline_equation", + "content": "y_{4}" + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": " extracted by the " + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": " ESDB module are meaningless and network inertia occurs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 271, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 415 + ], + "type": "text", + "content": "Second, as shown in the left of Fig. 3, although the weights of each kernel are highly correlated along the channel dimension, differences also exist. Besides, BSConv [12] also proved that not all the slices (52/128 for the ResNet-50 example) of the kernel are highly correlated along the depth axis. However, as shown in the right of Fig. 3, we find that the pointwise convolution of BSConv associates almost all the channel dimensions along the depth axis (the darker the color, the higher the correlation), which results in introducing many useless or disruptive features on some channels and is not conducive to effective extraction and propagation of useful features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 481, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 481, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 481, + 434 + ], + "type": "text", + "content": "3.2. Entropy Delta Extraction Link" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 440, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 462 + ], + "type": "text", + "content": "In this section, we will explain the definition of FDEL in detail." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "content": "In the network with stacked RBs, as shown in Fig. 5 (a), the initial shallow (low-level) features " + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "content": " will be superimposed and retained until the last layer " + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "inline_equation", + "content": "x_{L}" + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "content": ", as follows" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 514, + 553, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 514, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 372, + 514, + 553, + 555 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {2} = \\mathcal {F} _ {1} (x _ {1}) + x _ {1}, \\dots , \\\\ x _ {l} = \\mathcal {F} _ {l - 1} \\left(x _ {l - 1}\\right) + x _ {l - 1}, \\dots , \\\\ x _ {L} = \\mathcal {F} _ {L - 1} \\left(x _ {L - 1}\\right) + x _ {L - 1}, \\tag {1} \\\\ \\end{array}", + "image_path": "e6a7bb296cd440f2c9b52b9cbb68ca8616b9040f7aa2eaffe2be06b81cc6869e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_l" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": " represents the operations of the " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": " RB. Equation (1) can be regarded as a feature superposition operation. It may let the feature representation " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": " RB rely too much on the input features " + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "inline_equation", + "content": "x_{l-1}" + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": ", which will limit the learning direction of features and even lead to stagnation in feature evolution. This phenomenon can be evidenced in Fig. 2. In addition, the superposition of many low-level input features is not conducive to the presentation of high-level features, especially in structure Fig. 5 (c)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In view of the aforementioned limitations of the FAM of the RBs, we rethink its design principles and propose an FDEL that can tackle the above issues by flipping the FAM," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "402" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 152 + ], + "type": "image", + "image_path": "d848d73b83085bfa83047f528a3e6d3dced69e420931370cf8ce81b0da3b87ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 154, + 190, + 163 + ], + "lines": [ + { + "bbox": [ + 104, + 154, + 190, + 163 + ], + "spans": [ + { + "bbox": [ + 104, + 154, + 190, + 163 + ], + "type": "text", + "content": "Shallow Feature Extraction" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 167, + 358, + 310 + ], + "blocks": [ + { + "bbox": [ + 246, + 154, + 321, + 163 + ], + "lines": [ + { + "bbox": [ + 246, + 154, + 321, + 163 + ], + "spans": [ + { + "bbox": [ + 246, + 154, + 321, + 163 + ], + "type": "text", + "content": "Deep Feature Extraction" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 167, + 358, + 310 + ], + "lines": [ + { + "bbox": [ + 57, + 167, + 358, + 310 + ], + "spans": [ + { + "bbox": [ + 57, + 167, + 358, + 310 + ], + "type": "image", + "image_path": "897bedc20e7af603ed40c95437972f73aeed8e8d712f5b86969da3017ecaf5f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 167, + 446, + 310 + ], + "blocks": [ + { + "bbox": [ + 369, + 167, + 446, + 310 + ], + "lines": [ + { + "bbox": [ + 369, + 167, + 446, + 310 + ], + "spans": [ + { + "bbox": [ + 369, + 167, + 446, + 310 + ], + "type": "image", + "image_path": "df2c9f5c11a7f62af36792682f8a3c516e8c72a278e104224bb021981ed55064.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 319, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 60, + 319, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 60, + 319, + 547, + 331 + ], + "type": "text", + "content": "Figure 4. The overall architecture of the proposed SLVR and the structure of BCRB and blueprint controllable convolutions (B2Conv)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 457, + 167, + 553, + 310 + ], + "blocks": [ + { + "bbox": [ + 388, + 154, + 457, + 163 + ], + "lines": [ + { + "bbox": [ + 388, + 154, + 457, + 163 + ], + "spans": [ + { + "bbox": [ + 388, + 154, + 457, + 163 + ], + "type": "text", + "content": "Image Reconstruction" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 167, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 457, + 167, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 457, + 167, + 553, + 310 + ], + "type": "image", + "image_path": "ca6940c310b64355605ecc64992b8d961aa54c0599363c750b3a30b692fbc39d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 346, + 296, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 383 + ], + "type": "text", + "content": "the structure of RB with FDEL is shown in Fig. 5 (b). Our FDEL aims to remove excessive low-level features. The specific mathematical formula of FDEL is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 119, + 392, + 294, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 392, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 119, + 392, + 294, + 404 + ], + "type": "interline_equation", + "content": "x _ {l + 1} = \\mathcal {F} _ {\\mathrm {F D E L}} \\left(x _ {l}, \\mathcal {F} _ {l} \\left(x _ {l}\\right)\\right), \\tag {2}", + "image_path": "ed0ab6a8aaec239ef18e086351813f90c7ab40ac5ea4cff3ca7d70468a70f8e4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{FDEL}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "type": "text", + "content": " denotes the operations of the FDEL. Our FDEL uses the output of the " + }, + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_l" + }, + { + "bbox": [ + 55, + 413, + 296, + 521 + ], + "type": "text", + "content": " minus its input. Because the output feature representation is more high-level than the input, each RB can effectively extract the different high-level features, which promotes the diverse evolution of features. As we all know, we usually refer to the output of a function minus its input as the gain of the function in a physical sense. Therefore, in an RB with FDEL, the gain is the high-level features extracted by this RB." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 527, + 250, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 250, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 250, + 540 + ], + "type": "text", + "content": "3.3. Blueprint Controllable Convolutions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 545, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 689 + ], + "type": "text", + "content": "Although BSConv [12] demonstrates the importance of intra-kernel correlations in separating ordinary convolution, it does not consider the differences of the filters in the depth-axis. In other words, in some channel dimensions, this blueprint cannot obtain the original filter of the ordinary convolution through scale transformation, so filters after forced scale transformation don't effectively work. Therefore, we sparse the intra-kernel correlation and proposed a new B2Conv. Our B2Conv can adaptively pick appropriate channels along the depth axis to perform blueprint scale transformation, which can promote the extraction and efficient propagation of feature information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In a convolution operation, assume that the input is " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "U\\in \\mathbb{R}^{M\\times H\\times W}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", the output is " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "V\\in \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", and the con" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "type": "text", + "content": "volution kernels are " + }, + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\{E^{(1)}, E^{(2)}, \\dots, E^{(N)}\\}" + }, + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "type": "inline_equation", + "content": "E^{(n)} \\in \\mathbb{R}^{M \\times H \\times W}" + }, + { + "bbox": [ + 313, + 346, + 555, + 372 + ], + "type": "text", + "content": ". Then the process can be expressed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 395, + 378, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 378, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 395, + 378, + 553, + 392 + ], + "type": "interline_equation", + "content": "V _ {n, \\therefore :} = U * E ^ {(n)}, \\tag {3}", + "image_path": "ff2f575c8089b7039008dcb466723c2ce5abf58625b715505e5c2f5c6048e64f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "content": " denotes the convolution operation and " + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "inline_equation", + "content": "n \\in \\{1,2,\\dots,N\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "content": ", the weights of these " + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "content": " kernels can be learned during the training. BSCov proved that in practice these weights usually converge to a state that shows a substantial amount of correlation. Therefore, BSCov proposes to decompose an ordinary convolution into the interaction of a blueprint " + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "content": " and a set of scale-transforming weights " + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\{w_{n,1}, w_{n,2}, \\dots, w_{n,M}\\}" + }, + { + "bbox": [ + 313, + 401, + 555, + 498 + ], + "type": "text", + "content": ", as follows" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 389, + 505, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 505, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 389, + 505, + 553, + 521 + ], + "type": "interline_equation", + "content": "E _ {m,:,:} ^ {(n)} = w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}, \\tag {4}", + "image_path": "fe9c7d5df8f6663c44cb5981abf957a7bbabc404d99e46d23ec2b1c1182228d5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "inline_equation", + "content": "m \\in \\{1, 2, \\dots, M\\}" + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "inline_equation", + "content": "n \\in \\{1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "content": " (see Fig. 6, left column). However, in ordinary convolution, not all the filters of channel dimension can be translated by Equation (4). Therefore, to solve the problem, we propose to add an access control unit (ACU) before scale-transforming. Specifically, in all channel dimensions, if there is a strong intra-kernel correlation with " + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "content": ", we perform Equation (4), otherwise we do not perform any operation or multiply Equation (4) by a small correlation coefficient to ensure that the useless features extracted in this dimension will not affect the efficient propagation of features, as follows" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 666, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 666, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 375, + 666, + 553, + 682 + ], + "type": "interline_equation", + "content": "E _ {m,:,:} ^ {(n)} = \\mathcal {T} _ {n, m} \\cdot w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}, \\tag {5}", + "image_path": "b6c92db58f1a7eb18dcafdead69930ac7f5f58c700c7e87709c0ff528d95b570.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{n,m}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " denotes the output of the ACU, ideally its value is 0 or 1. 1 means allowing scale-transforming and 0 means" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "403" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 171, + 185 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 171, + 185 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 171, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 171, + 185 + ], + "type": "image", + "image_path": "eafb0e1ff1362b418cfe31a692e23fea81d0a32eeaff477aeef25ea7312bda59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 227 + ], + "type": "text", + "content": "Figure 5. Comparison of residual blocks (RBs) with addition mode and FDEL. For the specific mathematical formula of FDEL, refer to Equation (2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 69, + 295, + 185 + ], + "blocks": [ + { + "bbox": [ + 181, + 69, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 181, + 69, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 181, + 69, + 295, + 185 + ], + "type": "image", + "image_path": "d8c83baf08f6cd185f067ef4ef29325385f94c2c00731d6a4be8b20a7422229b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 215 + ], + "type": "text", + "content": "Figure 6. Interpretations for BSConv and B2Conv. B2Conv can adaptively pick correct intra-kernel correlations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": "the opposite Here, to make the function continuous and differentiable, we soften the output of the ACU and limit its value range to " + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": ". The closer the " + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{n,m}" + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": " is to 1, the stronger the intra-kernel correlation. The characteristics of features can reflect the characteristics of the convolution kernels. Therefore, " + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{n,m}" + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": " can be obtained in an embedding space by feature space mapping and weights of the " + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{n,m}" + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "inline_equation", + "content": "w_{n,m}" + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": " can be learned directly via backpropagation. Specifically, we first rewrite the Equation (3) by split 3d convolution into a set of " + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 237, + 296, + 357 + ], + "type": "text", + "content": " 2d convolutions, yielding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 365, + 295, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 365, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 86, + 365, + 295, + 399 + ], + "type": "interline_equation", + "content": "V _ {n,:,:} = U * E ^ {(n)} = \\sum_ {m = 1} ^ {M} \\left(U _ {m,:,:} * E _ {m,:,:} ^ {(n)}\\right). \\tag {6}", + "image_path": "da3925b3653ebeec78908f94e64846efc0b21017bc39b5f5f8c1ae725f56fa69.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": "Secondly, we replace the ordinary convolution " + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "inline_equation", + "content": "E_{m,\\cdot,\\cdot}^{(n)}" + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": " with the B2Conv representation as given in Equation (5), as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 452, + 295, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 452, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 80, + 452, + 295, + 484 + ], + "type": "interline_equation", + "content": "V _ {n, \\therefore :} = \\sum_ {m = 1} ^ {M} \\left(U _ {m, \\therefore :} * \\left(\\mathcal {T} _ {n, m} \\cdot w _ {n, m} \\cdot \\mathcal {B} ^ {(n)}\\right)\\right). \\tag {7}", + "image_path": "fb4959235471e2e26a0a9cadec345fd07174af192b6c5407282bc7663c6d4122.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "text", + "content": "Because each blueprint " + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^{(n)}" + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "text", + "content": " is independent in the channel dimension. So, we can rewrite the Equation (7) (Note: As shown in right column of Fig. 6, the input " + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "inline_equation", + "content": "U_{m,:;}" + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "text", + "content": " also needs to be copied " + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "inline_equation", + "content": "(\\bar{U}_{m,:;})" + }, + { + "bbox": [ + 55, + 493, + 296, + 542 + ], + "type": "text", + "content": " for feature space mapping)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 550, + 299, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 299, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 299, + 628 + ], + "type": "interline_equation", + "content": "V _ {n,:,:} = \\left(\\underbrace {\\sum_ {m = 1} ^ {M} \\left(\\bar {U} _ {m , :,:} \\cdot \\mathcal {T} _ {n , m}\\right)} _ {\\bar {V} _ {n,:,:} ^ {\\prime}} \\cdot \\underbrace {\\sum_ {m = 1} ^ {M} \\left(U _ {m , :,:} \\cdot w _ {n , m}\\right)} _ {V ^ {\\prime} _ {n,:,:}}\\right) * \\mathcal {B} ^ {(n)}. \\tag {8}", + "image_path": "2abc96f8954520d467d9629f0e7b1a47976f110be2fb09b0ab8e26fbdf6af3c4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "content": "Then, we further rearrange " + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\{w_{n,1}, w_{n,2}, \\dots, w_{n,M}\\}" + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_{n,1}, \\mathcal{T}_{n,2}, \\dots, \\mathcal{T}_{n,M}\\}" + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "inline_equation", + "content": "M \\times 1 \\times 1" + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "content": " arrays " + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\tilde{w}_n" + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}}_n" + }, + { + "bbox": [ + 55, + 630, + 296, + 679 + ], + "type": "text", + "content": " respectively, the sum operations of Equation (8) can be replaced by convolution operations, as follows" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 686, + 295, + 702 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 686, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 140, + 686, + 295, + 702 + ], + "type": "interline_equation", + "content": "\\bar {V} _ {n,:,:} ^ {\\prime} = U * \\tilde {\\mathcal {T}} _ {n}, \\tag {9}", + "image_path": "207a628cc3c2cce2bbd4bdfe275f4ee1e9beda74f524ec380a7938cafb841abe.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 703, + 294, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 703, + 294, + 716 + ], + "spans": [ + { + "bbox": [ + 139, + 703, + 294, + 716 + ], + "type": "interline_equation", + "content": "V _ {n,:,:} ^ {\\prime} = U * \\tilde {w} _ {n}, \\tag {10}", + "image_path": "9abd3c2438f2de44d67faf6ebdc083fd1b3f6cd348d1cc7df6c7f73b444db520.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 185 + ], + "type": "image", + "image_path": "d3764e97daef199ebcdf6f58ead41ed5c9ad2ecc015e71c42e31d326f3f2e30d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 360, + 222, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 222, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 360, + 222, + 553, + 239 + ], + "type": "interline_equation", + "content": "V _ {n,:,:} = \\left(\\sigma \\left(\\bar {V} _ {n,:,:} ^ {\\prime}\\right) \\cdot V _ {n,:,:} ^ {\\prime}\\right) * \\mathcal {B} ^ {(n)}, \\tag {11}", + "image_path": "b86729e4eed91efb2c157882e01e190925a100de5c75fe18c87cf636783deb30.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": " denotes the sigmoid function, which limits " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\tilde{V}_{n,::}^{\\prime}" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": " to (0,1). For a concrete implementation, Equation (9) and Equation (10) can be translated into a " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": " pointwise convolution with the kernels " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\{\\tilde{w}_1,\\tilde{w}_2,\\dots ,\\tilde{w}_N\\}" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\left\\{\\tilde{\\mathcal{T}}_1,\\tilde{\\mathcal{T}}_2,\\dots ,\\tilde{\\mathcal{T}}_N\\right\\}" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": " respectively. The Equation (11) can be translated into a " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "K\\times K" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": " depthwise convolution with the kernels " + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}^{(1)},\\mathcal{B}^{(2)},\\dots ,\\mathcal{B}^{(N)}\\}" + }, + { + "bbox": [ + 313, + 241, + 555, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 337, + 441, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 337, + 441, + 349 + ], + "spans": [ + { + "bbox": [ + 314, + 337, + 441, + 349 + ], + "type": "text", + "content": "3.4. The Overall Structure" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "text", + "content": "Based on FDEL and B2Conv, we further develop a lightweight SLVR framework to achieve high-performance image SR. As shown in Fig. 4, SLVR consists of three parts, including shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given a LR input " + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LR}} \\in \\mathbb{R}^{H \\times W \\times C_{in}}" + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "text", + "content": ", we first use a B2Conv " + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{B2Conv}}" + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "text", + "content": " to extract shallow feature " + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "inline_equation", + "content": "X_0 \\in \\mathbb{R}^{H \\times W \\times C_{mid}}" + }, + { + "bbox": [ + 313, + 356, + 554, + 439 + ], + "type": "text", + "content": ", as follows" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 368, + 446, + 553, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 446, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 368, + 446, + 553, + 460 + ], + "type": "interline_equation", + "content": "X _ {0} = \\mathcal {H} _ {\\mathrm {B 2 C o n v}} \\left(\\operatorname {C o n c a t} \\left[ I _ {\\mathrm {L R}} \\right] _ {n}\\right), \\tag {12}", + "image_path": "1dc004610b055621110c5d32ae78ac54de6bdd729530b1f261277f22055a4700.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "inline_equation", + "content": "C_{mid}" + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "content": " denote the channel number of the input and shallow feature. " + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "inline_equation", + "content": "\\text{Concat}[\\cdot]_n" + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "content": " means concatenating " + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "content": " times input. Then we use " + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 466, + 554, + 562 + ], + "type": "text", + "content": " efficient controllable distillation block (ECDB) to extract deep features. To maximize the effect of FDEL, feature aggregation is performed by DFAS and GELU [15] activation to enhance feature representation. After that, we use a B2Conv to continue refining the features, the above process can be formulated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 353, + 569, + 553, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 569, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 353, + 569, + 553, + 582 + ], + "type": "interline_equation", + "content": "X _ {i} = \\mathcal {H} _ {\\mathrm {E C D B} _ {i}} \\left(X _ {i - 1}\\right), i = 1, 2, \\dots , K, \\tag {13}", + "image_path": "3400d164bdd87e2d70976680599feffd639eb74e30563e7b7873c5df83bc8d26.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 589, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 589, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 318, + 589, + 553, + 613 + ], + "type": "interline_equation", + "content": "X _ {\\mathrm {A g g}} = \\mathcal {H} _ {\\mathrm {D F A S}} \\left(\\left(w _ {1}, w _ {2}, \\dots , w _ {c}\\right), \\left(X _ {1}, X _ {2}, \\dots , X _ {K}\\right)\\right), \\tag {14}", + "image_path": "fc282c377770ec195e429a9f9377ddf4fbdb326949a06fd82002141702530231.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 613, + 553, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 613, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 359, + 613, + 553, + 626 + ], + "type": "interline_equation", + "content": "X _ {\\text {R e f i n e d}} = \\mathcal {H} _ {\\mathrm {B 2 C o n v}} \\left(\\operatorname {G E L U} \\left(X _ {\\mathrm {A g g}}\\right)\\right), \\tag {15}", + "image_path": "fa925e4f0170eec3bcafab679249e0f92c4cd0123bdfa6ff0ba94476edeb5de7.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{ECDB}_i}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " ECDB, " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "w_{1}, w_{2}, \\dots, w_{c}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " are the learnable weights, " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "X_{1}, X_{2}, \\dots, X_{K}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " denote the intermediate features. " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{DFAS}}(\\cdot)" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " denotes the operations of the DFAS, which can be achieved by a convolution. The ECDB is designed based on the ESDB of the BSRN. Specifically, first, we replace the BSRB with the blueprint controllable residual block (BCRB). Second, we replace the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "404" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 69, + 541, + 464 + ], + "blocks": [ + { + "bbox": [ + 73, + 69, + 541, + 464 + ], + "lines": [ + { + "bbox": [ + 73, + 69, + 541, + 464 + ], + "spans": [ + { + "bbox": [ + 73, + 69, + 541, + 464 + ], + "type": "table", + "html": "
MethodT-DatasetScale#ParamsMult-AddsSet5Set14BSD100Urban100Manga109
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
VDSR [22]291×2666K612.6G37.530.958733.030.912431.900.896030.760.914037.220.9750
CARN [1]D2K+291×21,592K222.8G37.760.959033.520.916632.090.897831.920.925638.360.9765
IMDN [21]D2K×2694K158.8G38.000.960533.630.917732.190.899632.170.928338.880.9774
LAPAR-A [27]DF2K×2548K171G38.010.960533.620.918332.190.899932.100.928338.670.9772
LatticeNet [37]D2K×2756K169.5G38.060.960733.700.918732.200.899932.250.928838.940.9774
RLFN [24]D2K×2527K97.9G38.070.960733.720.918732.220.900032.330.9299--
SRPN-Lite [65]DF2K×2609K139.9G38.100.960833.700.918932.250.900532.260.9294--
FMEN [6]DF2K×2748K172.0G38.100.960933.750.919232.260.900732.410.931138.950.9778
MLRN [10]D2K×2488k90.4G38.070.960733.590.918032.210.900032.280.929738.760.9773
NGswin [4]D2K×2998K140.4G38.050.961033.790.919932.270.900832.530.932438.970.9777
SLVR (Ours)D2K×2464K102G38.140.961033.870.920232.280.900932.570.931839.080.9776
VDSR [22]291×3666K612.6G33.660.921329.770.831428.820.797627.140.827932.010.9340
CARN [1]D2K+291×31,592K118.8G34.290.925530.290.840729.060.803428.060.849333.500.9440
IMDN [21]D2K×3703K71.5G34.360.927030.320.841729.090.804628.170.851933.610.9445
LAPAR-A [27]DF2K×3594K114G34.360.926730.340.842129.110.805428.150.852333.510.9441
LatticeNet [37]D2K×3765K76.3G34.400.927230.320.841629.100.804928.190.851333.630.9442
RFDN [34]D2K×3633K65.6G34.470.928030.350.842129.110.805328.320.854733.780.9458
SRPN-Lite [65]DF2K×3615K62.7G34.470.927630.380.842529.160.806128.220.8534--
FMEN [6]DF2K×3757K77.2G34.450.927530.400.843529.170.806328.330.856233.860.9462
MLRN [10]D2K×3496K40.9G34.460.926730.350.842629.100.805428.200.853333.660.9450
NGswin [4]D2K×31,007K66.6G34.520.928230.530.845629.190.807828.520.860333.890.9470
SLVR (Ours)D2K×3473K45.3G34.580.928530.530.846029.210.808028.560.860334.070.9475
VDSR [22]291×4666K612.6G31.350.883828.010.767427.290.725125.180.752428.830.8870
CARN [1]D2K+291×41,592K90.9G32.130.893728.600.780627.580.734926.070.783730.470.9084
IMDN [21]D2K×4715K40.9G32.210.894828.580.781127.560.735326.040.783830.450.9075
LAPAR-A [27]DF2K×4659K94G32.150.894428.610.781827.610.736626.140.787130.420.9074
LatticeNet [37]D2K×4777K43.6G32.180.894328.610.781227.570.735526.140.784430.540.9075
RFDN [34]D2K×4550K23.9G32.240.895228.610.781927.570.736026.110.785830.580.9089
RLFN [24]D2K×4543K25.3G32.240.895228.620.781327.600.736426.170.7877--
SRPN-Lite [65]DF2K×4623K35.8G32.240.895828.690.783627.630.737326.160.7875--
BSRN [30]DF2K×4352K19.4G32.350.896628.730.784727.650.738726.270.790830.840.9123
FMEN [6]DF2K×4769K44.2G32.240.895528.700.783927.630.737926.280.790830.700.9107
MLRN [10]D2K×4507K23.5G32.300.895628.620.782427.570.736526.100.786730.560.9092
HPUN-M [48]D2K×4511K27.7G32.240.895028.660.782827.600.737126.120.787830.550.9089
SAFMN [49]DF2K×4240K14.0G32.180.894828.600.781327.580.735925.970.780930.430.9063
OSFFNet [53]DF2K×4537K22.0G32.390.897628.750.785227.660.739326.360.795030.840.9125
NGswin [4]D2K×41,019K36.4G32.330.896328.780.785927.660.739626.450.796330.800.9128
SLVR-S (Ours)DF2K×4158K8.5G32.200.895228.650.782227.570.736126.080.785430.530.9084
SLVR (Ours)D2K×4485K26.2G32.400.897528.780.785027.650.738826.410.795130.930.9133
SLVR* (Ours)DF2K×4485K26.2G32.440.898028.850.786527.700.740026.580.800031.100.9141
", + "image_path": "4929a7d4ca76336e789905185d20047cb403d4db8c4ced37c7a96ade36dfbd85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "type": "text", + "content": "Table 1. Quantitative comparison (PSNR/SSIM) for lightweight SR with state-of-the-art (SOTA) methods on five benchmark datasets. The best and second-best results are marked in red and blue colors respectively. Mult-Adds is evaluated on a " + }, + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "type": "inline_equation", + "content": "{1280} \\times {720}\\mathrm{{HR}}" + }, + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "type": "text", + "content": " image. T-Dataset denotes training dataset. D2K stands for the DIV2K dataset. DF2K indicates a merged dataset containing " + }, + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "type": "inline_equation", + "content": "{800} + 2,{650}\\mathrm{{HR}} - \\mathrm{{LR}}" + }, + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "type": "text", + "content": " image pairs. 291 images dataset is from [58]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 523, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 608 + ], + "type": "text", + "content": "addition mode of the residual structure with the proposed FDEL. Third, all the BSConvs are replaced with the proposed B2Conv. Please see Fig. 4 for detailed changes. After performing the above operations, we obtain the final refined features " + }, + { + "bbox": [ + 55, + 523, + 295, + 608 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{Refined}}" + }, + { + "bbox": [ + 55, + 523, + 295, + 608 + ], + "type": "text", + "content": ". Finally, to take advantage of residual learning, we use long skip connections to reduce the difficulty of learning. SR results can be obtained by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 612, + 295, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 612, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 78, + 612, + 295, + 625 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {S R}} = \\mathcal {H} _ {\\text {U p s a m p l e}} \\left(X _ {\\text {R e f i n e d}}\\right) + \\text {B i c u b i c} ^ {\\uparrow} \\left(I _ {\\mathrm {L R}}\\right), \\tag {16}", + "image_path": "50873844a9df108a89eb612af0f4ddf4951bbd6d21f7ea70e7a476505c68cce6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{Upsample}}(\\cdot)" + }, + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "text", + "content": " represents the upsampling operation, which is implemented with the pixelshuffle operation [44]. Bicubic denotes bicubic upsampling. " + }, + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "text", + "content": " loss function is used to optimize the network, as follows" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 684, + 295, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 684, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 133, + 684, + 295, + 697 + ], + "type": "interline_equation", + "content": "L _ {1} = \\left\\| I _ {\\mathrm {H R}} - I _ {\\mathrm {S R}} \\right\\| _ {1}, \\tag {17}", + "image_path": "c8749a8da1cb35a0a8905db714180136c0a050d3bbc94005be04a0480c513a42.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 165, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 165, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 165, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 165, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 701, + 165, + 713 + ], + "type": "text", + "content": " is the HR label." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 522, + 395, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 522, + 395, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 522, + 395, + 536 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 430, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 430, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 430, + 558 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 566, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 686 + ], + "type": "text", + "content": "Datasets and Metrics. We use DIV2K [50] dataset as the training dataset, which contains 800 HR-LR image pairs. To show more fair results, we used DF2K (DIV2K + Flickr2K [50]) to additionally train an " + }, + { + "bbox": [ + 313, + 566, + 555, + 686 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 566, + 555, + 686 + ], + "type": "text", + "content": " model. Note that the model trained with DF2K is marked with a small *. For testing, we evaluated the performances on the five benchmark datasets, including Set5 [2], Set14 [61], B100 [39], Urban100 [19], Manga109 [40]. The PSNR and SSIM [54] are used as the evaluation metrics, which are evaluated on the Y channel of the YCbCr space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Implementation Details. ECDB number is set to 8, and the channel number is set to 64. The kernel size of the depth-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "405" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 310 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 310 + ], + "type": "image", + "image_path": "e449684b2a1efe1ed3fad976449cfc9a9417a2e3f50d3b6fa33280586af865d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "type": "text", + "content": "Figure 7. Visual comparison for " + }, + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 316, + 295, + 340 + ], + "type": "text", + "content": " SR methods. The patches for comparison are marked with red boxes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "text", + "content": "wise convolutions is set to 3. We additionally built a smaller model SLVR-S to compare the latest methods with less than 200K parameters and it trained on DF2K. The ECDB number of the SLVR-S is set to 6 and the channel number is set to 40. During training, we augment the data with random horizontal flips and rotation " + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "inline_equation", + "content": "(90^{\\mathrm{o}},180^{\\mathrm{o}},270^{\\mathrm{o}})" + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "text", + "content": ". We randomly crop LR patches of size " + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "text", + "content": " as input. Our models are trained by Adam optimizer [23] for 1000K iterations. The initial learning rate is set to " + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "inline_equation", + "content": "1\\times 10^{-3}" + }, + { + "bbox": [ + 54, + 350, + 295, + 533 + ], + "type": "text", + "content": " and halved for every 200k iterations. The batch size of SLVR and SLVR-S is set to 64 and 128 respectively. LR image patches are obtained by the bicubic kernel from corresponding HR images matching each SR task. Our networks are trained on an RTX3090 GPU. Note that no other training skills (e.g., pre-training [28]) are employed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 539, + 268, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 268, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 268, + 552 + ], + "type": "text", + "content": "4.2. Comparison with the SOTA SR methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": "To evaluate the effectiveness of SLVR, we compare our model with other efficient SR models on scale factors " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\times 2,\\times 3,\\times 4" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": " including VDSR (CVPR16) [22], CARN (ECCV18) [1], IMDN (ACMMM19) [21], RFDN (ECCV20) [34], LAPAR-A (NeurIPS20) [27], LatticeNet (ECCV20) [37], RLFN (CVPRW22) [24], SRPN-Lite (ICLR22) [65], BSRN (CVPRW22) [30], FMEN (CVPRW22) [6], MLRN (CVPRW23) [10], NGswin (CVPR23) [4], HPUN-M (AAAI23) [48], SAFMN (ICCV23) [49], OSFFNet (AAAI24) [53]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "As shown in Table 1, we reported the training dataset, the parameters, Mult-Adds, and PSNR and SSIM on five benchmarks. Compared with models with similar param" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 432, + 166 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 432, + 166 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 432, + 166 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 432, + 166 + ], + "type": "image", + "image_path": "63bb4782ea05068379bb2f5b9913756429bfee69690d89b860e94e1416f25845.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 208 + ], + "type": "text", + "content": "Figure 8. Ablation studies on B2Conv. BSRN-B2Conv denotes the BSCnv of BSRN is replaced by B2Conv. All the results are reported on scale factor " + }, + { + "bbox": [ + 313, + 175, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\times 4" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 553, + 166 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 553, + 166 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 553, + 166 + ], + "type": "image", + "image_path": "22385b29953ecde2fa7b60f12de3d0b1e1135ac5a0cb26069832c80b2b97bd5e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 221, + 557, + 268 + ], + "blocks": [ + { + "bbox": [ + 317, + 221, + 557, + 268 + ], + "lines": [ + { + "bbox": [ + 317, + 221, + 557, + 268 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 557, + 268 + ], + "type": "table", + "html": "
MethodFDELB2ConvSet5Set14Urban100
Model-I✓×32.37/0.896928.71/0.783926.30/0.7917
Model-II×✓32.39/0.897528.76/0.785126.40/0.7943
SLVR✓✓32.40/0.897528.78/0.785026.41/0.7951
", + "image_path": "bf2de84d086295f28e90eba38e7b387ae287d51e57ab83a09e4d8f2dab895cc8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "text", + "content": "Table 2. Ablation studies of FDEL and B2Conv on scale factor " + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "text", + "content": " of FDEL denotes it is replaced by addition mode, and the " + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 277, + 555, + 311 + ], + "type": "text", + "content": " of B2Conv denotes it is replaced by BSConv." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "type": "text", + "content": "eters and Mult-Adds, the performance of our SLVR surpasses existing methods with a notable margin on all benchmarks. Compared with the latest transformer-based method NGswin, although the number of parameters of our SLVR is less than " + }, + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "type": "text", + "content": " of NGswin and the number of Mult-Adds is less than " + }, + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 323, + 555, + 441 + ], + "type": "text", + "content": " of NGswin, our SLVR still has an advantage in performance. Compared with the latest extremely lightweight method SAFMN trained on the same dataset, our SLVR-S obtains better performance with a smaller model size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "type": "text", + "content": "We also perform a visual comparison of different SOTA lightweight SR methods on scale factor " + }, + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "type": "text", + "content": ". As shown in Fig. 7, we can observe that the SR results generated by SLVR contain more correct and fine-grained textures, while other methods generate many blurred edges and artifacts. For example, in the second row of Fig. 7, our SLVR can restore the texture of the fence, while other methods cannot. We think that the visual results can prove the effectiveness of the proposed B2conv, which can prevent incorrect intrakernel correlations to reduce useless or disruptive feature extraction and propagation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 416, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 416, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 416, + 598 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "The core contributions of our framework are the proposed FDEL and B2Conv. To demonstrate their effectiveness, we first conduct ablation studies on the main framework of our SLVR. Specifically, we remove FDEL and B2Conv and replace them with the original linking mode FAM of RBs and BSCnv, respectively. The experimental results are reported in Table 2. We can observe that our FDEL and B2Conv can improve network performance, especially B2Conv, improved by 0.11 dB on the dataset Urban100." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "406" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 295, + 101 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 295, + 101 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 295, + 101 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 295, + 101 + ], + "type": "table", + "html": "
MethodScale#ParamsMult-AddsSet5Set14Urban100
EDSR-baseline×41,518K114.0G32.08/0.893828.58/0.781025.98/0.7824
EDSR+FDEL1,518K114.0G32.12/0.894428.61/0.781926.05/0.7845
", + "image_path": "ce38474fb0efe5508cf61ff6c5bd43b7878e6c5882d6b77c19dc7443d1e76d20.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 140, + 294, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 110, + 295, + 132 + ], + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 132 + ], + "type": "text", + "content": "Table 3. Ablation studies of FDEL. EDSR+FDEL denotes that all addition mode of EDSR-baseline [32] are replaced by FDEL." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 140, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 58, + 140, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 294, + 203 + ], + "type": "table", + "html": "
RFDNScale#ParamsMult-AddsSet5Set14Urban100
w BSConv×3197K25.8G34.37/0.926830.34/0.841828.06/0.8503
w B2Conv261K34.5G34.44/0.927330.36/0.842628.18/0.8525
w Conv633K65.6G34.47/0.928030.35/0.842128.21/0.8525
w BSConv×4210K14.9G32.12/0.894128.59/0.780626.01/0.7829
w B2Conv273K20.1G32.22/0.894928.65/0.782226.06/0.7843
w Conv643K37.4G32.28/0.895728.61/0.781826.11/0.7858
", + "image_path": "5280a7c8f79aac669a8156b5e36e4e45d62a8e05637f07a9ca93840a14ceb71b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 295, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 245 + ], + "type": "text", + "content": "Table 4. Comprehensive comparison between ordinary convolution (Conv), BSCnv, and B2Conv. The best and second-best results are marked in red and blue colors." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 254, + 294, + 303 + ], + "blocks": [ + { + "bbox": [ + 58, + 254, + 294, + 303 + ], + "lines": [ + { + "bbox": [ + 58, + 254, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 58, + 254, + 294, + 303 + ], + "type": "image", + "image_path": "506680e4a9f5f9174dc4d9063b855981d7cbfaaee447f4420066fadeadb79f66.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 295, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 344 + ], + "type": "text", + "content": "Figure 9. Visual comparison of feature maps between SLVR and SLVR without FDEL. Outi represents the output of the " + }, + { + "bbox": [ + 55, + 312, + 295, + 344 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 312, + 295, + 344 + ], + "type": "text", + "content": " ECDB module." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 353, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 425 + ], + "type": "text", + "content": "Second, to verify that FDEL can be used as a more efficient linking mode of RBs, we use FDEL to simply replace the traditional linking mode FAM in the classic residual architecture EDSR [32]. From Table 3, we can conclude that our FDEL is more effective than the traditional FAM without additional parameters and Mult-Adds consumed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": "Third, to further verify the effectiveness of our proposed B2Conv and confirm that B2Conv can be used as highly efficient building blocks. We denote BSConv of BSRN [30] replaced by B2Conv as BSRN-B2Conv. As shown in Fig. 8, we can observe that our B2Conv can significantly improve model performance without changing model architecture. Besides, we also conducted a comprehensive comparison between ordinary convolution, BSConv, and B2Conv on the network architecture RFDN [34], as shown in Table 4, compared with BSConv, our B2Conv greatly improves the performance (e.g., " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{0.1dB}" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": " on Set5 × 4) of the RFDN with few additional parameters (63K) consumed, which also illustrates the rationality of our improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Finally, we perform a visual comparison of features between SLVR and SLVR without FDEL. From Fig. 9, we find that our FDEL can effectively alleviate network inertia. In addition, we visualize the scale transformation state of 8 blueprints on 64 channels. As shown in Fig. 10, we can observe that our B2Conv does not perform scale transformation on all 64 channel dimensions (only on the channel denoted by black squares). It proves that our B2Conv can adaptively select the appropriate channel to perform blueprint scale transformation, which can effectively prevent the introduction of disruptive or useless features. Fur" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 373, + 127 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 373, + 127 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 373, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 373, + 127 + ], + "type": "image", + "image_path": "6b0828daebd3fef597bf61b13377a90184306c2ed8c3cafd8aa442b65bac927d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 137, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 170 + ], + "type": "text", + "content": "Figure 10. Visualization of scale transformation state of 4 blueprints on 64 channels. Black squares denote strong intrakernel correlation, and white squares are the opposite." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 432, + 127 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 432, + 127 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 432, + 127 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 432, + 127 + ], + "type": "image", + "image_path": "6f7d271da91ed03df7e1cfeb0cf79ad79fa669ab51ea0a224c836e787a9d98c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 71, + 493, + 127 + ], + "blocks": [ + { + "bbox": [ + 436, + 71, + 493, + 127 + ], + "lines": [ + { + "bbox": [ + 436, + 71, + 493, + 127 + ], + "spans": [ + { + "bbox": [ + 436, + 71, + 493, + 127 + ], + "type": "image", + "image_path": "e1680436f1d224796df161d49cbc9137cbbf193184342ab4293b6445d912df51.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 497, + 71, + 553, + 127 + ], + "blocks": [ + { + "bbox": [ + 497, + 71, + 553, + 127 + ], + "lines": [ + { + "bbox": [ + 497, + 71, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 497, + 71, + 553, + 127 + ], + "type": "image", + "image_path": "2b954807791b6157a792563a0699127988680707459f18732e11e193109dfa09.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 174, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 316, + 174, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 293 + ], + "type": "image", + "image_path": "627c1e80b9ef302746cedc93293099dd4fa2bb0362c8df65d026d6fd404b71ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 302, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 323 + ], + "type": "text", + "content": "Figure 11. Visual comparison of intermediate feature maps between SLVR and SLVR without B2Conv." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 332, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 464 + ], + "type": "text", + "content": "thermore, we perform a visual comparison of features between SLVR and SLVR without B2Conv. From Fig. 11, we can find that the intermediate features generated by the SLVR without B2Conv contain more interference and disruptive features, resulting in poorer feature representation. The intermediate features generated by our SLVR are more accurate and have better feature representation ability. It also indirectly proves that the proposed B2Conv can adaptively select the correct intra-kernel correlation, prevent the introduction of disruptive features, and is more conducive to feature propagation and representation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 477, + 387, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 387, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose blueprint controllable convolutions (B2Conv). B2Conv can precisely prevent incorrect intra-kernel correlation, which effectively reduces the extraction of useless or disruptive features and promotes the effective propagation of feature information. Besides, we propose a novel linking mode FDEL for RBs. Our FDEL reduces the retention of the input low-level features and can promote the diversity evolution of high-level features, which can effectively alleviate the occurrence of network inertia. Finally, based on B2Conv and FDEL, we propose an efficient lightweight SR framework SLVR. Furthermore, our B2Conv and FDEL can also serve as highly efficient components to improve the performance of other CNN-based models (such as EDSR, BSRN, and RFDN). Extensive experiments on public benchmark datasets show the prominent SR performance of SLVR and comprehensive analytical studies validate the reasonableness and effectiveness of the proposed framework and components." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "407" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, pages 252-268, 2018. 1, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "text", + "content": "[2] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "text", + "content": "[3] Yunpeng Chen, Haoqi Fan, Bing Xu, Zhicheng Yan, Yannis Kalantidis, Marcus Rohrbach, Yan Shuicheng, and Ji-ashi Feng. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In ICCV, pages 3434–3443, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 237, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 237, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 295, + 270 + ], + "type": "text", + "content": "[4] Haram Choi, Jeongmin Lee, and Jihoon Yang. N-gram in swim transformers for efficient lightweight image superresolution. In CVPR, pages 2071–2081, 2023. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 303 + ], + "type": "text", + "content": "[5] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 295, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 349 + ], + "type": "text", + "content": "[6] Zongcai Du, Ding Liu, Jie Liu, Jie Tang, Gangshan Wu, and Lean Fu. Fast and memory-efficient network towards efficient image super-resolution. In CVPRW, pages 852-861, 2022. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "type": "text", + "content": "[7] Jinsheng Fang, Hanjiang Lin, Xinyu Chen, and Kun Zeng. A hybrid network of cnn and transformer for lightweight image super-resolution. In CVPRW, pages 1102-1111, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 385, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 295, + 417 + ], + "type": "text", + "content": "[8] Mengyang Feng, Huchuan Lu, and Yizhou Yu. Residual learning for salient object detection. IEEE TIP, 29:4696-4708, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "type": "text", + "content": "[9] Guangwei Gao, Wenjie Li, Juncheng Li, Fei Wu, Huimin Lu, and Yi Yu. Feature distillation interaction weighting network for lightweight image super-resolution. In AAAI, pages 661-669, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 464, + 295, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 295, + 506 + ], + "type": "text", + "content": "[10] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. Mixer-based local residual network for lightweight image super-resolution. In CVPRW, pages 1593-1602, 2023. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 542 + ], + "type": "text", + "content": "[11] Garas Gendy, Nabil Sabor, Jingchao Hou, and Guanghui He. A simple transformer-style network for lightweight image super-resolution. In CVPRW, pages 1484-1494, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 586 + ], + "type": "text", + "content": "[12] Daniel Haase and Manuel Amthor. Rethinking depthwise separable convolutions: How intra-kernel correlations lead to improved mobilenets. In CVPR, pages 14588-14597, 2020. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 589, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 589, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 589, + 295, + 621 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 623, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 623, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 623, + 295, + 656 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In ECCV, pages 630-645. Springer, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 678 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[16] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolu" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 703 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 117 + ], + "type": "text", + "content": "[17] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, pages 7132-7141, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[18] Gao Huang, Shichen Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Condensenet: An efficient densenet using learned group convolutions. In CVPR, pages 2752-2761, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[19] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "type": "text", + "content": "[20] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, pages 723-731, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "text", + "content": "[21] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multi-distillation network. In ACM MM, pages 2024-2032, 2019. 1, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 1, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 306, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 336 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 381 + ], + "type": "text", + "content": "[24] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPRW, pages 765-775, 2022. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "type": "text", + "content": "[26] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In CVPR, pages 5835-5843, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 503 + ], + "type": "text", + "content": "[27] Wenbo Li, Kun Zhou, Lu Qi, Nianjuan Jiang, Jiangbo Lu, and Jiaya Jia. Lapar: Linearly-assembled pixel-adaptive regression network for single image super-resolution and beyond. NeurIPS, 33:20343-20355, 2020. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 505, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 547 + ], + "type": "text", + "content": "[28] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[29] Wenjie Li, Juncheng Li, Guangwei Gao, Weihong Deng, Jiantao Zhou, Jian Yang, and Guo-Jun Qi. Cross-receptive focused inference network for lightweight image superresolution. IEEE TMM, pages 1-13, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[30] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In CVPRW, pages 832-842, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 637, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 669 + ], + "type": "text", + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In ICCV, pages 1833-1844, 2021. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 703 + ], + "type": "text", + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 1132-1140, 2017." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "408" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[33] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In ECCV, pages 85-100, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[34] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, pages 41-55. Springer, 2020. 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[35] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPRW, pages 456-465, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[36] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[37] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, pages 272-289. Springer, 2020. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[38] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In ECCV, pages 116-131, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[39] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423 vol.2, 2001. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76:21811-21838, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[41] Marcelo Gennari Do Nascimento, Victor Prisacariu, and Roger Fawcett. *Dsconv: Efficient convolution operator. In ICCV*, pages 5147-5156, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[42] Tobias Pohlen, Alexander Hermans, Markus Mathias, and Bastian Leibe. Full-resolution residual networks for semantic segmentation in street scenes. In CVPR, pages 3309-3318, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "text", + "content": "[43] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510-4520, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 511, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 566 + ], + "type": "text", + "content": "[44] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "text", + "content": "[46] Ayan Sinha, Asim Unmesh, Qixing Huang, and Karthik Ramani. Surfnet: Generating 3d shape surfaces using deep residual networks. In CVPR, pages 791-800, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 656 + ], + "type": "text", + "content": "[47] Rupesh K Srivastava, Klaus Greff, and Jürgen Schmidhuber. Training very deep networks. NeurIPS, 28, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[48] Bin Sun, Yulun Zhang, Songyao Jiang, and Yun Fu. Hybrid pixel-unshuffled network for lightweight image superresolution. In AAAI, pages 2375-2383, 2023. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[49] Long Sun, Jiangxin Dong, Jinhui Tang, and Jinshan Pan. Spatially-adaptive feature modulation for efficient image" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "super-resolution. In ICCV, pages 13190-13199, 2023. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[50] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 1110-1121, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 183 + ], + "type": "text", + "content": "[51] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Jinfan Liu. Omni aggregation networks for lightweight image super-resolution. In CVPR, pages 22378-22387, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[52] Xijun Wang, Meina Kan, Shiguang Shan, and Xilin Chen. Fully learnable group convolution for acceleration of deep neural networks. In CVPR, pages 9041–9050, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "text", + "content": "[53] Yang Wang and Tao Zhang. Osffnet: Omni-stage feature fusion network for lightweight image super-resolution. In AAAI, pages 5660-5668, 2024. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[54] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "type": "text", + "content": "[56] Zeyu Xiao, Jiawang Bai, Zhihe Lu, and Zhiwei Xiong. A dive into sam prior in image restoration. arXiv preprint arXiv:2305.13620, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "text", + "content": "[57] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, pages 5987-5995, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[58] Jianchao Yang, John Wright, Thomas S. Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE TIP, 19(11):2861-2873, 2010. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[59] Jiahui Yu, Zhe Lin, Jamei Yang, Xiaohui Shen, Xin Lu, and Thomas Huang. Free-form image inpainting with gated convolution. In ICCV, pages 4470-4479, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 456, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 478 + ], + "type": "text", + "content": "[60] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[61] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "text", + "content": "[62] Ting Zhang, Guo-Jun Qi, Bin Xiao, and Jingdong Wang. Interleaved group convolutions. In ICCV, pages 4383-4392, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "text", + "content": "[63] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[64] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[65] Yulun Zhang, Huan Wang, Can Qin, and Yun Fu. Learning efficient image super-resolution networks via structure-regularized pruning. In ICLR, 2021. 6, 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[66] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCV, pages 56-72. Springer, 2020. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "409" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 117 + ], + "type": "text", + "content": "[67] Daquan Zhou, Qibin Hou, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Rethinking bottleneck structure for efficient mobile network design. In ECCV, pages 680-697. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "410" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_content_list.json b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..886d15eef395048fd8d37db46f23104dea95187c --- /dev/null +++ b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "SMILE: Infusing Spatial and Motion Semantics in Masked Video Learning", + "text_level": 1, + "bbox": [ + 119, + 130, + 877, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fida Mohammad Thoker, Letian Jiang, Chen Zhao†, Bernard Ghanem King Abdullah University of Science and Technology (KAUST)", + "bbox": [ + 218, + 180, + 776, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fida.thoker,letian.jiang,chen.zhao,Bernard.Ghanem}@kaust.edu.sa", + "bbox": [ + 212, + 219, + 779, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masked video modeling, such as VideoMAE, is an effective paradigm for video self-supervised learning (SSL). However, they are primarily based on reconstructing pixel-level details on natural videos which have substantial temporal redundancy, limiting their capability for semantic representation and sufficient encoding of motion dynamics. To address these issues, this paper introduces a novel SSL approach for video representation learning, dubbed as SMILE, by infusing both spatial and motion semantics. In SMILE, we leverage image-language pretrained models, such as CLIP, to guide the learning process with their high-level spatial semantics. We enhance the representation of motion by introducing synthetic motion patterns in the training data, allowing the model to capture more complex and dynamic content. Furthermore, using SMILE, we establish a new self-supervised video learning paradigm capable of learning strong video representations without requiring any natural video data. We have carried out extensive experiments on 7 datasets with various downstream scenarios. SMILE surpasses current state-of-the-art SSL methods, showcasing its effectiveness in learning more discriminative and generalizable video representations. Code is available: https://github.com/fmthoker/SMILE", + "bbox": [ + 89, + 301, + 485, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 675, + 220, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) [43, 44, 57, 59] plays a crucial role for many video understanding tasks, such as action recognition [18, 26, 51, 52, 55] and detection [28, 32], significantly reducing the reliance on extensive human annotation. The goal of video SSL is to learn video representations that capture both spatial and temporal semantics effectively. Masked video modeling has emerged as a promising approach with state-of-the-art performance [12, 21, 53, 57]. It involves masking a large portion of video frames and reconstructing the masked regions using the unmasked parts.", + "bbox": [ + 89, + 702, + 482, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, this approach has some limitations. First, it relies heavily on pixel reconstruction, which can lead to overfitting low-level details. This limits the model's abl", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38cd8034fabbd90eb5913dd89adf8141947704afaf27dbbef3cca683613243ea.jpg", + "image_caption": [ + "Figure 1. Comparison with SOTA masked video modeling methods. Our method significantly outperforms prior masked video modeling methods across diverse downstream settings." + ], + "image_footnote": [], + "bbox": [ + 549, + 271, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ity to learn meaningful high-level semantics essential for discriminative representations, as shown by the suboptimal linear probing performance of current masked video modeling methods [12, 21, 53, 57] in Figure 1. The second key issue in masked video modeling is the substantial temporal redundancy of video data with most videos having little foreground motion, human or object. Consequently, the reconstruction objective can be solved using the temporal correlations between co-located patches or by spatial correlations in tube masking [57], resulting in inadequate motion awareness in the learned representations. This turn leads to poor generalization on downstream tasks, requiring a finer motion understanding as shown in [55].", + "bbox": [ + 509, + 582, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we address these limitations by proposing a novel framework for masked video modeling, dubbed as SMILE (Spatial and Motion semantics in masked video LEarning), to capture high-level spatial and motion semantics leading to generalizable video representations. To improve the spatial semantics in masked video modeling, we propose to explicitly infuse high-level spatial semantics from pretrained image foundation models such as", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIP [45]. Such models encapsulate high-level semantics learned through large-scale image-text alignment and have been adapted for video understanding tasks. For example, ActionCLIP [61], ViFiCLIP [47], and X-CLIP [36]) adapt CLIP for action recognition using videos with action labels, whereas ViCLIP [64] and VideoPrism [72]) adapt it for video-text alignment with video-text pairs. We take inspiration from them but aim to learn video representations with unlabeled videos. In particular, we propose replacing the reconstruction of pixels in masked video modeling with that of CLIP embedding, thus providing high-level semantics as the supervision signal.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve motion awareness, we propose to reduce the temporal redundancy in input training videos with synthetic motion patterns. Specifically, we randomly paste objects onto the video frames and move them smoothly along the temporal dimension to simulate variable temporal transitions of motion-intensive videos. This reduces the temporal redundancy in the original videos and results in videos with an additional prominent foreground motion. Furthermore, adding objects with known locations allows us to add a motion focus to the mask-and-predict task by explicitly masking the regions of object motion. This in turn encourages the network to avoid an over-reliance on static spatial information and encode the motion dynamics of the added objects, when solving the reconstruction task. We summarize our contributions as follows.", + "bbox": [ + 89, + 273, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SMILE, a novel SSL approach for masked video modeling. It encodes spatial and motion semantics using guidance from image-text pretrained models and synthetic-motion augmented videos, leading to more discriminative and generalizable video representation.", + "- We present a new strategy to augment video data with diverse synthetic motions, paired with a masking strategy that focuses on reconstructing these added motions to improve masked video modeling pretraining.", + "- Through extensive evaluation on 7 datasets, including Kinetics-400 [26], UCF-101 [52], HMDB-51 [29], FineGYM [50], EPIC-Kitchens-100 [8], SomethingSomething V2 [18], and Charades [51], we demonstrate that SMILE significantly outperforms existing state-of-the-art video SSL methods and exhibits better generalization." + ], + "bbox": [ + 91, + 502, + 482, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 741, + 238, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Self-Supervised Learning. Video SSL aims to learn the spatio-temporal dynamics of video data without the need for any supervision. Three main learning paradigms have emerged: transformation prediction, contrastive learning, and masked video modeling. In transformation prediction, models learn video patterns by solving tasks such as space-time puzzles [1, 25, 27], predicting order in video clips [15, 24, 30, 33, 34, 66], or estimating speed [3, 7, 69]. Contrastive learning is based on instance discrimina", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion [38] and learning invariances to spatio-temporal augmentations [2, 6, 14, 35, 40, 42, 44, 46, 54, 56]. Masked video modeling methods [12, 21, 53, 57, 67], like VideoMAE [57], use a mask-and-predict task to reconstruct videos from partial data with an encoder-decoder setup similar to MAE [20]. By masking a high proportion of pixels (e.g., $90\\%$ ) and reconstructing them such methods learn useful video representations. Our approach aims to improve the spatial and motion semantics of the video representations learned by masked video modeling.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CLIP for Video Understanding. CLIP [45] captures detailed spatial semantics from extensive image-text training, making it a strong foundation for directly modeling video data with added temporal dynamics, e.g., [36, 41, 58, 61, 64, 68]. ViFiCLIP [47] finetunes CLIP on video by adding a temporal pooling layer, ActionCLIP [61] includes a temporal positional embedding and aggregation layer, and X-CLIP [36] integrates cross-frame attention to share information between frames. Most relevant to us are works that also use CLIP for video representation learning [13, 31, 64]. ViCLIP [64] initializes a model with CLIP weights and employs further contrastive pretraining with video-text pairs. UMT [31] uses CLIP as a teacher and aligns its predictions with the CLIP features for unmasked tokens, followed by a 2nd stage of learning from video-text contrastive pairs. TGM [13] relies on CLIP text-encoder features to guide masked video modeling and video-text contrastive learning in a unified framework. In contrast, our SMILE only learns from video data in a self-supervised manner, reconstructing CLIP features of masked tokens and added objects.", + "bbox": [ + 511, + 244, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion-aware Video Self-Supervised Learning. Encoding motion dynamics in video representations is essential for generalizing across various video tasks [55]. As a result, many video SSL methods focus on enhancing the motion sensitivity of learned representations. For example, [16, 19, 21, 37, 65] capture motion using optical flow, [9, 60] do so by removing background or static elements and [11, 22, 56, 59] incorporate synthetic motions. [22, 59] learn by predicting the speed and trajectory of the pseudo motions, and [56] contrasts video with different types of motions. These approaches are specific to transformation prediction and contrastive learning paradigms. We take inspiration from these works and inject synthetic motions into the masked video learning.", + "bbox": [ + 511, + 550, + 903, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most relevant to our work are methods like [12, 21, 53, 67], which enhance motion understanding in VideoMAE [57] by modifying certain design choices. MGMAE [21] and MGM [12] improve masking strategies to mask the regions of motion using optical flow and motion vectors. MotionMAE [67] adds frame difference as an additional target for reconstruction, and MME [53] uses optical flow to extract dense object trajectories and then predicts these trajectories and their HOG features. However, these", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods are limited by the temporal redundancies in the original data, restricting the amount of motion or motion-focused regions available for learning in the reconstruction task. In contrast, we add synthetic motions to the original video data, reducing temporal redundancy and increasing motion-rich regions to learn from. As shown in our experiments, reconstructing these modified videos encourages the model to capture motion dynamics, enhancing the motion awareness of the learned video representations.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 234, + 225, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting Masked Video Modeling", + "text_level": 1, + "bbox": [ + 89, + 258, + 395, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of masked video modeling [12, 21, 57] is to learn video representations by employing a mask-and-predict task with an asymmetric encoder-decoder architecture. An input video $V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ is first divided into non-overlapped 3D space-time tokens (e.g., $2 \\times 16 \\times 16$ ) as $\\mathcal{T} = \\{v_1, \\dots, v_N\\}$ . From $\\mathcal{T}$ , a subset of tokens $\\mathcal{T}^{mask}$ are randomly masked with a high masking ratio (e.g., $90\\%$ ) and the unmasked tokens $\\mathcal{T}^{unmask} = \\mathcal{T} \\setminus \\mathcal{T}^{mask}$ are fed into a transformer encoder $\\Phi_{enc}$ . Then, the output of the encoder and the learnable tokens containing the positional information of the masked tokens are passed to a shallow decoder $\\Phi_{dec}$ to reconstruct the input pixel with the following loss:", + "bbox": [ + 89, + 279, + 483, + 462 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{| \\mathcal {T} ^ {\\text {m a s k}} |} \\sum_ {i \\in \\mathcal {T} ^ {\\text {m a s k}}} \\| v _ {i} - Y _ {i} \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 472, + 482, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $v_{i}$ refers to the pixel values of the $i$ -th space-time token in $\\mathcal{T}^{mask}$ and $Y_{i} = \\Phi_{dec}(\\Phi_{enc}(\\mathcal{T}^{unmask}))$ denotes the corresponding reconstructed pixels for $i$ -th token given the set of all unmasked tokens $\\mathcal{T}^{unmask}$ .", + "bbox": [ + 89, + 521, + 482, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Such pixel reconstruction drives the model to focus on low-level features, and with high temporal redundancy in video data, it primarily reconstructs static appearance details of video data. However, effective video understanding also requires capturing temporal progression and high-level semantics, such as scene context and object interactions.", + "bbox": [ + 89, + 582, + 483, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. General Scheme of SMILE", + "text_level": 1, + "bbox": [ + 89, + 683, + 334, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to infuse spatial and motion semantics in masked video modeling. To achieve this, we integrate two key elements into the masked video modeling paradigm. To improve motion semantics, we overlay local motion patterns in input videos in the form of moving objects. To improve spatial semantics, we replace the low-level pixel targets with high-level features extracted from a pretrained CLIP encoder. By combining both, we learn video representations that can encode both spatial and motion dynamics necessary for modeling downstream video tasks. The overview of our method is shown in Figure 2.", + "bbox": [ + 89, + 703, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Infusion. Given a input video $V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ . We first sample a random object $o$ from a set of predefined", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "segmented objects $O$ . Next, we generate a motion trajectory (Section 3.3), along which we overlay this object onto the video $V$ to obtain the motion-augmented video $V'$ . By adding objects with synthetic motions, we reduce temporal redundancy in the original videos and encourage the model to learn from the motion dynamics of the overlaid objects.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masking. We employ two separate masking strategies on $V'$ as follows. For the original regions, we employ random tube masking as in VideoMAE [57]. For added object regions, we propose trajectory-based masking. Since we know the location of the object in each frame of $V'$ , we explicitly drop space-time tokens at random along its trajectory. Employing a high masking ratio along the object trajectory ensures that solving the reconstruction task requires encoding the object's motion dynamics, thereby enhancing the motion awareness of learned representations.", + "bbox": [ + 511, + 183, + 906, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. We employ a teacher-student reconstruction architecture with a pretrained image-encoder as the teacher and an asymmetric encoder-decoder network $\\Phi$ as the student. The augmented video is patchified into 3D space-time tokens $\\mathcal{T} = \\{v_1',\\dots,v_N'\\}$ as in general video masked autoencoders (Section 3.1). The tokens are then masked using tube and trajectory-based masking to generate sets of masked $\\mathcal{T}^{mask}$ and unmasked $\\mathcal{T}^{unmask}$ tokens. The un-masked tokens $\\mathcal{T}^{unmask}$ and the learnable tokens containing the positional information of the masked patches are used as the input to $\\Phi$ as in Section 3.1.", + "bbox": [ + 511, + 335, + 905, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction. Instead of directly reconstructing pixels, we project the input video $V'$ onto the teacher's feature space to generate high-level targets. In particular, we choose a CLIP model as the projection network due to its high-level semantics learned from language guidance. Specifically, the input $V'$ is passed frame-by-frame through the CLIP encoder to extract features. The extracted features are stacked into a sequence of feature tokens as $\\mathcal{F} = \\{f_1', \\dots, f_N'\\}$ . To match the number of tokens with $\\mathcal{T}$ , we only use the CLIP features for the first time slice of the 3D space-time tokens, thus generating a total of $N$ feature tokens, each corresponding to a space-time token in the input video. The encoder-decoder is then trained with a masked feature reconstruction loss as:", + "bbox": [ + 511, + 503, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{| \\mathcal {T} ^ {\\text {m a s k}} |} \\sum_ {i \\in \\mathcal {T} ^ {\\text {m a s k}}} \\| f _ {i} ^ {\\prime} - Y _ {i} \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 724, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{i}^{\\prime}$ is the CLIP feature projection of $i$ -th space-time token $v_{i}^{\\prime}$ in $T^{mask}$ and $Y_{i} = \\Phi_{dec}(\\Phi_{enc}(\\mathcal{T}^{unmask}))$ denotes the corresponding reconstructed feature for $i$ -th token given the set of all unmasked tokens $T^{unmask}$ . Such formulation first increases the temporal dynamics in otherwise static data, avoiding learning shortcuts for reconstruction. Second, projecting targets onto the CLIP space encourages the network to capture more high-level video semantics.", + "bbox": [ + 511, + 777, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8440", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00a8924413236da0589b363ecc48927c37eff76ed09b81815d0c6aa2e1cf011e.jpg", + "image_caption": [ + "Figure 2. Overall architecture of our SMILE. An input video clip $V$ is overlaid with a segmented object along a randomly generated trajectory to generate $V'$ infusing synthetic object motion in $V$ . $V'$ is passed frame-by-frame through the CLIP encoder to extract feature tokens $\\mathcal{F}$ . $V'$ is patched into a set of space-time tokens $\\mathcal{T}$ . We apply two types of masking upon $\\mathcal{T}$ , tube masking on the space-time tokens of the original video and trajectory-based masking on the tokens of added objects. The unmasked tokens $\\mathcal{T}_{unmask}$ are fed into the encoder-decoder network $\\Phi_{enc} \\Phi_{dec}$ which is trained to reconstruct the masked feature tokens in $\\mathcal{F}$ ." + ], + "image_footnote": [], + "bbox": [ + 127, + 75, + 861, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, the key intuition is that CLIP features capture high-level information, such as scene context and object interactions, which guides the network to attend to both the overlaid object motions and the original video semantics.", + "bbox": [ + 89, + 463, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Motion Generation", + "text_level": 1, + "bbox": [ + 89, + 530, + 274, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To infuse motion, we overlay a segmented object along a randomly generated trajectory onto the original video frames. To introduce more complex dynamics, the object is also transformed e.g. rotation, along the trajectory.", + "bbox": [ + 89, + 550, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance and Trajectory. To infuse a new synthetic motion onto a video sequence $V = [v_{1}, v_{2}, \\dots, v_{T}]$ comprising $T$ frames, we begin by selecting an object $o$ from $O$ at random, where $O$ denotes the set of off-the-shelf segmented objects generated using Stable Diffusion [48] and X-Paste[71]. We then sample a scale $P \\times Q$ from a predefined set of sizes for $o$ to model different object sizes, where $P \\ll H$ and $Q \\ll W$ . Next, we sample a sequence of locations for the object $\\mathrm{Traj}_o = [(x^1, y^1),.., (x^T, y^T)]$ representing the center coordinate $(x^i, y^i)$ of the object $o$ in the corresponding frame. Specifically, we generate a random non-linear path for the trajectory by sampling $M$ 2D points $(M \\gg T)$ uniformly from $x \\in [0, H]$ and $y \\in [0, W]$ . To ensure smoothness in the motion, a 1D Gaussian filter is applied along $x$ and $y$ dimensions as $[(h(x^1), h(y^1)), \\dots, (h(x^M), h(y^M))]$ , where $h(z) = \\frac{1}{\\sqrt{2\\pi}\\kappa} e^{-z^2 / 2\\kappa^2}$ and $\\kappa$ is the smoothing factor. The resulting trajectory is then downsampled to $T$ target locations for the object $o$ . This formulation introduces diverse motions", + "bbox": [ + 89, + 611, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92e27627bbbf735ca15be37dc8c53828bc42dfe7980ab7b15d4bb049eb9f2a6f.jpg", + "image_caption": [ + "Figure 3. Motion overlaid videos showing the position change and transformation of the added objects along the time dimension." + ], + "image_footnote": [], + "bbox": [ + 535, + 450, + 882, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with varying spatio-temporal dynamics in the form of objects with different shapes, sizes, and temporal progression. Transformation. To increase the complexity of the object's motion, we also apply scale and rotation transformations along the temporal dimension. Specifically, we select 3 frame indices, the first, the last, and a random middle one, $v_{1}, v_{m}$ , and $v_{T}$ , respectively. For these three frame indices, a rotation angle $\\alpha_{i}$ , and a scaling factor $S_{i}$ are selected randomly. The angles and scales for the rest of the frames are linearly interpolated between these keyframes. Finally, the object $o$ is rotated and scaled with the corresponding transformations $\\left( \\begin{array}{cc}\\cos \\alpha_{i} & -\\sin \\alpha_{i}\\\\ \\sin \\alpha_{i} & \\cos \\alpha_{i} \\end{array} \\right)$ and $\\left( \\begin{array}{cc}S_i & 0\\\\ 0 & S_i \\end{array} \\right)$ , before overlaying it onto the video frame $v_{i}$ .", + "bbox": [ + 511, + 628, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 shows two sample videos with our added motions. While these motions may not appear entirely realistic, they represent complex motion dynamics of moving structured pixels, and introducing them into a mask-and-predict framework guides the model to learn from such dynamics.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This results in video representations with improved downstream performance on motion-aware tasks, as we also validate empirically. Interestingly, we also show in the experiments (Section 5) that overlaying such object motions on black or noise frames can learn strong video representations without requiring any original videos.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Procedure", + "text_level": 1, + "bbox": [ + 89, + 181, + 277, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By adding synthetic motions, we have introduced a domain shift in the input for reconstruction, which may affect the performance of downstream tasks. To explore effective learning from both original videos $V$ and motion-overlaid videos $V'$ , we explore two training strategies, i.e. mixed training and progressive training. In the mixed training strategy, we use both $V$ and $V'$ as the input to the network and directly optimize for the reconstruction of both inputs. The idea is to simultaneously encode the spatio-temporal dynamics of the original video data and the motion dynamics of the augmented data, with a limited domain shift. In the progressive training strategy, we employ a two-stage learning approach that first learns from augmented data $V'$ alone, then continues to train the model with the original data $V$ alone. This way we can first learn to encode the spatio-temporal dynamics of the augmented data and then align the network to further encode the dynamics of the original data and its domain.", + "bbox": [ + 89, + 200, + 482, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 482, + 222, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate SMILE against prior video SSL methods on action recognition benchmarks: Kinetics-400 (K400) [26], SomethingSomething V2 [18] (SSv2), UCF-101 [52] (UCF), HMDB-51 [29] (HMDB), FineGYM [50] (GYM), and EPIC-Kitchens-100 [8] (EPIC), achieving the best performance in linear probing and full finetuning (Section 4.2). SMILE also shows better generalization on the SEVERE benchmark[55] (Section 4.4). Ablations and further discussion are provided in Section 4.3 and Section 5, respectively.", + "bbox": [ + 89, + 500, + 482, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 640, + 307, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model. Following VideoMAE [57], we use ViT-S or ViT-B with joint space-time attention for our student network. We choose CLIP ViT-B [45] as our teacher network.", + "bbox": [ + 89, + 659, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motions. We follow PIN [10] to create the set of segmented objects $O$ with Stable Diffusion [48] and X-Paste [71]. We sample the object size $P \\times Q$ uniformly from $[32 \\times 32, 128 \\times 128]$ and sample the rotation angles from $[-90, 90]$ and scaling factors from [0.5, 1.5].", + "bbox": [ + 89, + 705, + 482, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining. We follow previous masked video modeling works [12, 21, 53, 57] and use Kinetics-400 [26] (K400) or SomethingSomething V2 [18] (SSv2) for pretraining unless specified otherwise. We use the same hyperparameters as in VideoMAE [57]. Unless stated otherwise, we train with our proposed progressive training for a total of 600 epochs, 300 epochs in each stage. More training details are in supplementary material.", + "bbox": [ + 89, + 781, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/07d8369e9fe9c31ebcf49131dcd2468413fe8da4a3744f8979feb3b36fe4112f.jpg", + "table_caption": [ + "Table 1. Linear probing comparison of masked video modeling methods. Our SMILE learns highly discriminative video representations useful for both spatial and motion-focused domains. All results are obtained with the ViT-B backbone pretrained on K400 and Top-1 accuracies $(\\%)$ are reported." + ], + "table_footnote": [], + "table_body": "
MethodK400UCFHMDBSSv2GYMEPIC
VideoMAE [57]20.758.637.717.523.933.2
MVD [63]18.749.128.612.222.729.7
MME [53]19.156.037.116.629.032.2
MGMAE [21]24.964.441.316.826.133.2
EVEREST [23]14.151.830.314.523.330.5
MGM [12]19.862.540.321.725.832.4
SIGMA [49]47.580.752.320.830.134.2
SMILE w/o motion54.682.651.721.627.932.3
SMILE (ours)56.283.853.423.730.234.4
", + "bbox": [ + 516, + 164, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 511, + 347, + 808, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear Probing. We freeze the pretrained backbone and only train a linear classifier on the target dataset. When evaluating the prior works for comparison, we utilize their publicly available checkpoints for the ViT-B backbone pretrained on the K400 dataset for 800 epochs except for MVD (2000 epochs) and MME (1600 epochs). We evaluate the downstream task using a common setup for a fair comparison. Details about datasets and experimental settings are in the supplementary material.", + "bbox": [ + 511, + 366, + 903, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows that our method, even without synthetic motion, significantly outperforms all state-of-the-art methods on K400 and UCF, and is among the top performers on HMDB, SSv2, GYM, and EPIC. This highlights that CLIP feature reconstruction offers a stronger supervisory signal than pixel reconstruction, especially in tasks requiring spatial semantics like K400 and UCF. Adding synthetic motions further boosts performance by an average of $1.8\\%$ across datasets, enhancing motion cues through reconstruction. On motion-sensitive datasets like SSv2, GYM, and EPIC, our approach consistently surpasses MGM [12], MGMAE [21], and MME [53], confirming that synthetic motions help overcome the lack of motion diversity in current datasets that may limit these methods.", + "bbox": [ + 511, + 503, + 905, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Finetuning. We finetune both the pretrained backbone and the classification head end-to-end following [21, 53, 57] on the downstream datasets K400 and SSv2. Following prior masked video modeling works [12, 21, 57], we show results with self-supervised pretraining on K400 and SSv2 datasets for the VIT-B backbone. We follow VideoMAE [57] for finetuning and evaluation protocols with more details in the supplementary.", + "bbox": [ + 511, + 717, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tables 2 and 3 show the results for SSv2 and K400, respectively. On SSv2, we achieve the best performance with both K400 and SSv2 pretraining, highlighting our method's effectiveness for both in-domain and cross-domain transfer.", + "bbox": [ + 511, + 839, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/37cb230b12f6f4b95d6e0cae4848a41f5a7284dfed20950ea85328ad6840becf.jpg", + "table_caption": [ + "Table 2. Full finetuning comparison of various self-supervised methods on Something-Something V2. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported." + ], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsPretrainTop-1
VideoMAE [57]ViT-B800K40068.5
OmniMAE [17]ViT-B800K40069.0
MGMAE* [21]ViT-B800K40068.9
MME [53]ViT-B800K40070.5
SIGMA [49]ViT-B800K40071.1
MGM* [12]ViT-B800K40071.1
SMILE (ours)ViT-B600K40072.1
OmniMAE [17]ViT-B800SSv269.5
VideoMAE [57]ViT-B800SSv269.6
MGM [12]ViT-B800SSv270.6
MME [53]ViT-B800SSv270.0
SIGMA [49]ViT-B800SSv271.2
MGMAE [21]ViT-B800SSv271.0
SMILE (ours)ViT-B800SSv272.5
", + "bbox": [ + 112, + 157, + 455, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For K400, we exceed the prior top SSL approach by $1.6\\%$ . Additionally, our gains over motion-focused methods, such as MGMAE (1.5% on SSv2, 1.9% on K400), MME (2.5% on SSv2, 1.6% on K400), and MGM (1.9% on SSv2, 2.3% on K400), confirm our improved motion sensitivity, even in full-finetuning settings. Notably, our model surpasses VideoMAE, MotionMAE, MME, and MGMAE trained for more than twice the number of epochs (see supplementary for the extensive comparison). We attribute this efficiency to CLIP feature reconstruction, which directly captures high-level semantics compared to pixel-level learning.", + "bbox": [ + 89, + 391, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 89, + 564, + 202, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We ablate various components of our SMILE to show their impact on the downstream performance in Table 4. For computational efficiency, we use ViT-S as the backbone and pretrain with a smaller subset of K400, denoted as $\\mathrm{K400_m}$ , containing around 80K samples. For evaluation, we use $\\mathrm{K400_m}$ and $\\mathrm{SSv2_m}$ that denotes a subset of SSv2 with $50\\%$ training data. Unless stated otherwise, we use CLIP features as the reconstruction target, add two synthetic objects, mask $90\\%$ , use trajectory masking, and train for 400 epochs.", + "bbox": [ + 89, + 583, + 482, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic motion, yes or no? In Table 4a, we show the impact of adding synthetic motions to the masked video modeling framework for both pixel and feature reconstruction. We can see that adding our synthetic motion patterns significantly improves the downstream performance of both pixel and feature reconstruction.", + "bbox": [ + 89, + 719, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstructing pixels or features? Table 4h compares pixel-based and feature-based reconstruction targets in masked video modeling. Reconstructing in feature space significantly boosts performance over pixel reconstruction, with a $+12.1\\%$ gain on spatially-focused tasks like $\\mathrm{K400_m}$ and $+5.1\\%$ on motion-focused tasks like $\\mathrm{SSv2_m}$ . This sug-", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/46689e6d66f26e62049c72a13c1fdd8404ebcd293bc28d7a77ce2d2d095afc16.jpg", + "table_caption": [ + "Table 3. Full finetuning comparison of various self-supervised methods on Kinetics-400. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported." + ], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsPretrainTop-1
SVT [46]ViT-B-K40078.4
VideoMAE [57]ViT-B800K40080.0
BEVT [62]ViT-B-K40080.6
OmniMAE [17]ViT-B800K40080.8
MGM [12]ViT-B800K40080.8
MGMAE [21]ViT-B800K40081.2
SIGMA [49]ViT-B800K40081.5
MME* [53]ViT-B800K40081.5
SMILE (ours)ViT-B600K40083.1
", + "bbox": [ + 537, + 157, + 877, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gests that using a high-level feature space as the reconstruction target provides more meaningful supervision for learning transferable spatio-temporal representations.", + "bbox": [ + 511, + 323, + 903, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstructing which features: CLIP, MAE or DINO? Table 4b compares several image foundation models—MAE [20], DINO [4], DINO-v2 [39], and CLIP [45]—as reconstruction targets. CLIP features outperform the others, likely due to CLIP's image-text alignment training on a large dataset, enhancing its transferability and adaptability to video tasks as shown in [36, 47, 61]. Compared to pixel reconstruction in Table 4h, all feature-based approaches show substantial performance gains.", + "bbox": [ + 511, + 368, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory-based masking: yes or no? Table 4g shows the impact of our trajectory-based masking. Without trajectory masking, all tokens are masked using tube masking, which masks the whole video without distinguishing between object and non-object space-time tokens. In contrast, trajectory masking first masks space-time tokens associated with objects, followed by tube masking on the remaining tokens, maintaining the overall masking ratio. Our trajectory masking improves performance by approximately $1\\%$ on both $\\mathrm{K400_m}$ and $\\mathrm{SSv2_m}$ , validating that masking along object trajectories enhances motion-aware representation learning.", + "bbox": [ + 511, + 505, + 905, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How much to mask? Table 4e shows the impact of masking ratio $m$ on downstream performance. We mask $m$ of space-time tubes within the objects and $m$ of space-time tubes in the whole video with $m$ being $95\\%$ , $90\\%$ , or $80\\%$ . We observe the best performance with $m = 80\\%$ .", + "bbox": [ + 511, + 672, + 905, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How many synthetic objects? In Table 4f, we analyze the impact of adding more object motions to the input video on downstream performance. We evaluate adding 1, 2 and 3 object motions and compare the results. We observe that adding 2 objects is better than using a single object and the performance saturates when adding more objects.", + "bbox": [ + 511, + 748, + 905, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Which object transformations? In Table 4c, we show the impact of object transformation on downstream performance. Both rotation and scaling improve the performance individually and combining them achieves the best results.", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d1fe19b008daa16e6cb0055932c573b55a0d4bf8e9bfec9b5be4770ed0d2e5f1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Synth.TargetK400mSSv2m
w/oPixels66.055.1
w/Pixels68.156.8
w/oFeatures78.160.2
w/Features78.961.6
", + "bbox": [ + 114, + 88, + 321, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cd3ff0d9b04470ecf4a74ba59d427e1afbd2d81df8d4bdc53399d607983ee86a.jpg", + "table_caption": [ + "(a) Synthetic motion, yes or no? Yes. It improves performance for both pixel and feature targets." + ], + "table_footnote": [], + "table_body": "
RatioTargetK400mSSv2m
95%features77.760.4
90%features78.961.6
80%features80.362.0
", + "bbox": [ + 112, + 213, + 312, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7159e2750f78e4e14d4a9fb21614e6b5cfe5eb035f283cb3a13532e1a0f96e75.jpg", + "table_caption": [ + "(e) How much to mask? Masking $80\\%$ trajectory and total tokens." + ], + "table_footnote": [], + "table_body": "
TargetK400mSSv2m
MAE [20]77.659.5
DINO [4]77.659.7
DINOv2 [39]77.859.7
CLIP [45]78.160.2
", + "bbox": [ + 328, + 89, + 514, + 171 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/71c2d1abcde56818c617c951243537f67d648b6d022b17ce3f96146815e43881.jpg", + "table_caption": [ + "(b) Reconstructing which features? CLIP features provide the best target. No motion is added." + ], + "table_footnote": [], + "table_body": "
# ObjectsK400mSSv2m
178.661.0
278.961.6
378.661.3
", + "bbox": [ + 334, + 214, + 500, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b541539732669f57d7d50152970736cded92c09ce7a8262b2c13aa7d5815a04.jpg", + "table_caption": [ + "(f) How many objects? Adding two object motions is the best." + ], + "table_footnote": [], + "table_body": "
TransformK400mSSv2m
None77.060.8
Scaling78.461.3
Rotation78.261.4
Both78.961.6
", + "bbox": [ + 521, + 90, + 694, + 171 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69888e8d31cb6801a3878213ce1be711ee6aa9fb5612a6f780bd3b7c7b1f509f.jpg", + "table_caption": [ + "(c) Which object transformations? Using both rotation and scaling works the best." + ], + "table_footnote": [], + "table_body": "
MaskingK400mSSv2m
w/o Trajectory77.860.7
w/ Trajectory78.961.6
", + "bbox": [ + 519, + 214, + 712, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bcbf5c71c24dc355732ea95192d02a5599029da1b43c28a5112e9cf1a6924b0d.jpg", + "table_caption": [ + "(g) Trajectory masking, yes or no? Yes. Both targets are features." + ], + "table_footnote": [], + "table_body": "
TrainingK400mSSv2m
Single78.961.6
Mix79.562.2
Progressive79.862.5
", + "bbox": [ + 702, + 90, + 880, + 160 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27cd24efdce3554e8928c72f69b6b075af88318821c4f70e0cde4d45f73da70d.jpg", + "table_caption": [ + "(d) Which training strategy? Progressive training outperforms both for the same number of iterations." + ], + "table_footnote": [], + "table_body": "
TargetK400mSSv2m
Pixels66.055.1
Features78.160.2
", + "bbox": [ + 715, + 214, + 879, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(h) Reconstructing pixels or features? Reconstructing features is much better. No motion is added.", + "bbox": [ + 705, + 273, + 888, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation experiments of our proposed SMILE. The default setting uses clip reconstruction targets with two synthetic objects and a $90\\%$ masking ratio with trajectory-based masking. We also show full-scale ablations with ViT-B and K400 pretraining in supplementary.", + "bbox": [ + 89, + 315, + 906, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Which training strategy? Table 4d compares our proposed training strategies. We also compare learning with the single learning strategy, where only augmented videos are used for training. We observe that both mixed and progressive learning are better than learning from augmented videos only. Progressive is better than mixed, most likely due to a reduced domain shift in the second stage where only original videos are used.", + "bbox": [ + 88, + 372, + 482, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis. To compare the temporal dynamics of video representations learned by different SSL methods, we compute their feature similarity across different frames of input videos (Figure 4). For our model, the features of different frames have larger differences, indicating better temporal awareness. More plots are in the supplementary.", + "bbox": [ + 88, + 493, + 483, + 584 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c92a6abf5b929a132587001f36c07fc6dd7ab20db7b2da7811b90df707fdb57.jpg", + "image_caption": [ + "Figure 4. Feature similarity across different frames for different SSL methods. We compute this on K400 validation videos." + ], + "image_footnote": [], + "bbox": [ + 109, + 599, + 468, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization Assessment", + "text_level": 1, + "bbox": [ + 89, + 744, + 333, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the generalization of our learned video representations to diverse downstream video setups, we evaluate SMILE on the SEVERE benchmark [55], which includes 8 experiments targeting 4 key generalization factors: domain shift, sample efficiency, action granularity, and task shift. Detailed configurations are provided in the supplementary. Beyond SEVERE, we also show generalization to more temporal tasks in supplementary.", + "bbox": [ + 89, + 763, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 shows the results. For the Domain Shift, we test", + "bbox": [ + 109, + 885, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cross-domain adaptability on SSv2 and Gym99, which differ from the pretraining dataset (K400). Our method significantly outperforms all others, demonstrating superior adaptability to domain shifts. For Sample Efficiency, we evaluate low-shot action recognition on UCF and GYM with only 1,000 samples for finetuning. Our approach surpasses previous methods, achieving a $5\\%$ improvement on GYM over the best prior model, MME, showcasing its strong low-shot recognition capability even in motion-sensitive tasks. For the Action Granularity, tested on FX-S1 and UB-S1 from FineGym, our model achieves the best results, excelling at distinguishing fine-grained motions within gymnastics routines (e.g., jump variations in FX-S1). It outperforms motion-aware models like MME, MGMAE, and MGM, highlighting superior motion generalizability. For Task Shift, evaluated with temporal repetition counting on UCF-RC [70] and multi-label recognition on Charades [51], our method matches prior performance on repetition counting and achieves a $9\\%$ improvement over MME on Charades, demonstrating broad task adaptability. Overall, SMILE (without synthetic motion) is on par with the prior best with the mean performance of $64.9\\%$ . With synthetic motion SMILE, we observe a further $3.0\\%$ gain, reinforcing our model's generalization strength through synthetic motion.", + "bbox": [ + 511, + 372, + 906, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 511, + 746, + 635, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Other CLIP Adaptations. Table 5 presents a comparison between our SMILE and the CLIP adaptation methods ViCLIP [64] and UMT [31], along with the CLIP baseline. Our method achieves superior performance on motion-sensitive datasets, particularly in linear probing, indicating that our CLIP adaptation learns better video representations than the video-text alignment of ViCLIP and the unmasked feature alignment of UMT, despite using less data. More results are in the supplementary.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8444", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a52415acee403f50253fa38869f38eb37b415fcfffc192b601ddf8c0501ad8e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PretrainingFinetuningLinear Probing
MethodVideo DatasetUCFGYMSSv2EPICUCFGYMSSv2EPIC
CLIP [45]-93.688.066.750.377.520.711.325.1
ViCLIP [64]Intervid-10M [64]95.289.767.955.086.727.318.927.3
UMT [31]K700 [5]96.089.970.150.188.026.418.828.2
SMILE (ours)K400 [26]96.490.871.963.383.830.223.734.6
", + "bbox": [ + 94, + 88, + 640, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2335615604a3f683b7b97f3e8eb76c29e3c46a0bb5c17dae8937073cae2d2a3.jpg", + "table_caption": [ + "Table 5. Comparison with CLIP adaptations for video representation learning. Despite using less data, we learn better video representations than other CLIP adaptations, especially for motion-focused downstream domains.", + "Table 6. Generalization assessment on SEVERE benchmark [55] for recent video SSL methods in terms of domain shift, sample efficiency, action granularity, and task shift. We evaluate prior works with the official ViT-B backbone checkpoints pretrained on the K400 dataset. SMILE shows the best generalization performance across setups. Adding synthetic motions further improves the performance." + ], + "table_footnote": [], + "table_body": "
MethodDomain shiftSample efficiency (103)Action granularityTask shiftMean
SSv2Gym99UCFGYMFX-S1UB-S1UCF-RC↓Charades
VideoMAE [57]68.686.674.625.942.865.30.17214.457.6
MVD [63]70.082.567.117.531.350.50.18416.152.1
MGMAE [21]68.987.277.224.133.779.50.18117.958.8
MGM [12]71.189.178.426.438.686.90.15222.562.2
SIGMA [49]70.989.784.128.055.179.90.16923.164.2
MME [53]70.189.779.229.855.587.20.15523.665.0
SMILE w/o motion71.690.085.232.053.474.30.17530.564.9
SMILE (ours)72.190.886.435.155.188.30.17032.567.9
", + "bbox": [ + 130, + 258, + 867, + 417 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a6f6e19a5e9d0fad41310ea671ade2d5c91c2831ee166d1f9955c6a8f7155311.jpg", + "table_caption": [ + "Table 7. Learning without natural videos. We train a ViT-S for pixel reconstruction. Our synthetic object motions can learn video representations without using any natural videos." + ], + "table_footnote": [], + "table_body": "
Unnatural Video SchemePretrain DataK400mSSv2m
No Pretraining-39.728.2
Video-frameK400m-1frm54.043.1
Video-frame + MotionK400m-1frm63.653.2
ImagePlaces-80K52.140.5
Image + MotionPlaces-80K60.152.1
Black image + Motion-59.250.0
Noise image + Motion-61.951.2
", + "bbox": [ + 112, + 478, + 457, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning without Natural Videos. In prior experiments, we overlay our object motions onto natural videos from the Kinetics-400 dataset, raising the question of whether downstream improvements stem from original video priors (motion, action, scene) or our added object motions. To investigate this, we construct videos by overlaying object motions onto four different backgrounds: (1) a single frame from natural videos, (2) a single scene image from the image dataset Places [73], (3) a plain black image, and (4) a noise image. Each background image is duplicated across $T$ frames, creating static clips onto which moving objects are placed. For (1), we select one frame per video from $\\mathrm{K400_m}$ (denoted $\\mathrm{K400_m - 1frm}$ ), preserving action and scene cues without motion priors. For (2), we sample 80K images from Places (Places-80K), matching the dataset size of $\\mathrm{K400_m - 1frm}$ , to remove action and motion priors. Approaches (3) and (4) exclude all natural video or image priors entirely. Evaluation involves full finetuning on $\\mathrm{K400_m}$", + "bbox": [ + 88, + 628, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and $\\mathrm{SSv2_m}$ , consistent with our previous ablation studies.", + "bbox": [ + 511, + 433, + 890, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7 reports results for different augmented video types. Adding object motions consistently enhances downstream performance, whether combined with action and scene cues from natural videos (row 2 vs. row 3) or only scene cues from images (row 4 vs. row 5). Even with black or noisy backgrounds (rows 6 and 7), performance remains robust. Remarkably, using object motions alone yields substantial improvements— $22.2\\%$ on $\\mathrm{K400_m}$ and $23.0\\%$ on $\\mathrm{SSv2_m}$ compared to no pretraining. These findings demonstrate our method's effectiveness in learning video representations from synthetic motions alone. More results are in the supplementary. In summary, our method enables a new paradigm of learning strong video representations with masked video modeling without relying on natural videos.", + "bbox": [ + 511, + 449, + 906, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 667, + 640, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed SMILE, a novel self-supervised approach for video representation learning that addresses key limitations in existing masked video modeling methods by integrating both spatial and motion semantics. Our SMILE, leverages the high-level spatial semantics from image-language pretrained models like CLIP and enhances motion representation through synthetic motion patterns. By capturing complex spatial features and motion dynamics, SMILE achieves superior performance on diverse downstream tasks, demonstrating improved discrimination and generalization over state-of-the-art video SSL methods. Our extensive evaluations demonstrated the advantages of our SMILE, setting a new paradigm for robust and effective video self-supervised learning.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8445", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the KAUST Center of Excellence for Generative AI under award number 5940. For computing time, this research used Ibex managed by the Supercomputing Core Laboratory at King Abdullah University of Science & Technology (KAUST) in Thuwal, Saudi Arabia. We also thank Michael Dorkenwald for providing the synthetic objects used in this work.", + "bbox": [ + 89, + 90, + 485, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 223, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2019.", + "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7", + "[5] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 8", + "[6] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[7] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised visual learning by variable playback speeds prediction of a video. IEEE Access, 9:79562-79571, 2021. 2", + "[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. International Journal of Computer Vision (IJCV), 2021. 2, 5", + "[9] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[10] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 247, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5", + "[11] Hazel Doughty, Fida Mohammad Thoker, and Cees GM Snoek. Locomotion: Learning motion-focused video-language representations. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 50-70, 2024. 2", + "[12] David Fan, Jue Wang, Shuai Liao, Yi Zhu, Vimal Bhat, Hector Santos-Villalobos, Rohith MV, and Xinyu Li. Motion-guided masking for spatiotemporal representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8", + "[13] David Fan, Jue Wang, Shuai Liao, Zhikang Zhang, Vimal Bhat, and Xinyu Li. Text-guided video masked autoencoder. In European Conference on Computer Vision (ECCV), 2024. 2", + "[14] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[15] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[16] Kirill Gavrilyuk, Mihir Jain, Ilia Karmanov, and Cees G M Snoek. Motion-augmented self-training for video recognition at smaller scale. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[17] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimaes: Single model masked pretraining on images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5", + "[19] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, 2, 6, 7", + "[21] Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, and Limin Wang. Mgmae: Motion guided masking for video masked autoencoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8", + "[22] Ziyuan Huang, Shiwei Zhang, Jianwen Jiang, Mingqian Tang, Rong Jin, and Marcelo H Ang. Self-supervised motion" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8446", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning from static images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[23] Sunil Hwang, Jaehong Yoon, Youngwan Lee, and Sung Ju Hwang. Everest: Efficient masked video autoencoder by removing redundant spatiotemporal tokens. In International Conference on Machine Learning (ICML), 2024. 5", + "[24] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), 2020. 2", + "[25] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2", + "[26] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1, 2, 5, 8", + "[27] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2", + "[28] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified CNN architecture for real-time spatiotemporal action localization. arXiv preprint arXiv:1911.06644, 2019. 1", + "[29] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2011. 2, 5", + "[30] Hsin-Ying Lee, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[31] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7, 8", + "[32] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[33] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2", + "[34] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), 2016. 2", + "[35] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[36] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 2, 6", + "[37] Jingcheng Ni, Nan Zhou, Jie Qin, Qian Wu, Junqi Liu, Boxun Li, and Di Huang. Motion sensitive contrastive learning for self-supervised video representation. In European Conference on Computer Vision (ECCV), 2022. 2", + "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 7", + "[40] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. arXiv preprint arXiv:2303.09857, 2023. 2", + "[42] Mandela Patrick, Yuki M. Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[43] AJ Piergiovanni, Anelia Angelova, and Michael S Ryoo. Evolving losses for unsupervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[44] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2, 5, 6, 7, 8", + "[46] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S. Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6", + "[47] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8447", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5", + "[49] Mohammadreza Salehi, Michael Dorkenwald, Fida Mohammad Thoker, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Sigma: Sinkhorn-guided masked video modeling. In European Conference on Computer Vision (ECCV), 2025. 5, 6, 8", + "[50] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5", + "[51] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in Homes: Crowdsourcing Data Collection for Activity Understanding. In European Conference on Computer Vision (ECCV), 2016. 1, 2, 7", + "[52] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5", + "[53] Xinyu Sun, Peihao Chen, Liangwei Chen, Changhao Li, Thomas H. Li, Mingkui Tan, and Chuang Gan. Masked motion encoding for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 8", + "[54] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the ACM International Conference on Multimedia (ACMMM), 2021. 2", + "[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 1, 2, 5, 7, 8", + "[56] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Tubelet-contrastive self-supervision for videoefficient generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 5, 6, 8", + "[58] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. arXiv preprint arXiv:2304.10465, 2023. 2", + "[59] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In Pro" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[60] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[61] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 6", + "[62] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6", + "[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 8", + "[64] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2, 7, 8", + "[65] Fanyi Xiao, Joseph Tighe, and Davide Modolo. Maclr: Motion-aware contrastive learning of representations for videos. In European Conference on Computer Vision (ECCV), 2022. 2", + "[66] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[67] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 2", + "[68] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024, 2023. 2", + "[69] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[70] Huaidong Zhang, Xuemiao Xu, Guoqiang Han, and Shengfeng He. Context-aware and scale-insensitive temporal repetition counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7", + "[71] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8448", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In International Conference on Learning Representations (ICLR), 2023. 4, 5", + "[72] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 2", + "[73] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 8" + ], + "bbox": [ + 91, + 90, + 482, + 273 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8449", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_model.json b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..458a0bce57c1aaa1b858f3174c36d54825b363dc --- /dev/null +++ b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_model.json @@ -0,0 +1,2457 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.878, + 0.155 + ], + "angle": 0, + "content": "SMILE: Infusing Spatial and Motion Semantics in Masked Video Learning" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.181, + 0.777, + 0.218 + ], + "angle": 0, + "content": "Fida Mohammad Thoker, Letian Jiang, Chen Zhao†, Bernard Ghanem King Abdullah University of Science and Technology (KAUST)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.22, + 0.78, + 0.235 + ], + "angle": 0, + "content": "{fida.thoker,letian.jiang,chen.zhao,Bernard.Ghanem}@kaust.edu.sa" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.649 + ], + "angle": 0, + "content": "Masked video modeling, such as VideoMAE, is an effective paradigm for video self-supervised learning (SSL). However, they are primarily based on reconstructing pixel-level details on natural videos which have substantial temporal redundancy, limiting their capability for semantic representation and sufficient encoding of motion dynamics. To address these issues, this paper introduces a novel SSL approach for video representation learning, dubbed as SMILE, by infusing both spatial and motion semantics. In SMILE, we leverage image-language pretrained models, such as CLIP, to guide the learning process with their high-level spatial semantics. We enhance the representation of motion by introducing synthetic motion patterns in the training data, allowing the model to capture more complex and dynamic content. Furthermore, using SMILE, we establish a new self-supervised video learning paradigm capable of learning strong video representations without requiring any natural video data. We have carried out extensive experiments on 7 datasets with various downstream scenarios. SMILE surpasses current state-of-the-art SSL methods, showcasing its effectiveness in learning more discriminative and generalizable video representations. Code is available: https://github.com/fmthoker/SMILE" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.676, + 0.222, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) [43, 44, 57, 59] plays a crucial role for many video understanding tasks, such as action recognition [18, 26, 51, 52, 55] and detection [28, 32], significantly reducing the reliance on extensive human annotation. The goal of video SSL is to learn video representations that capture both spatial and temporal semantics effectively. Masked video modeling has emerged as a promising approach with state-of-the-art performance [12, 21, 53, 57]. It involves masking a large portion of video frames and reconstructing the masked regions using the unmasked parts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, this approach has some limitations. First, it relies heavily on pixel reconstruction, which can lead to overfitting low-level details. This limits the model's abl" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.272, + 0.891, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.515, + 0.908, + 0.557 + ], + "angle": 0, + "content": "Figure 1. Comparison with SOTA masked video modeling methods. Our method significantly outperforms prior masked video modeling methods across diverse downstream settings." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.907, + 0.78 + ], + "angle": 0, + "content": "ity to learn meaningful high-level semantics essential for discriminative representations, as shown by the suboptimal linear probing performance of current masked video modeling methods [12, 21, 53, 57] in Figure 1. The second key issue in masked video modeling is the substantial temporal redundancy of video data with most videos having little foreground motion, human or object. Consequently, the reconstruction objective can be solved using the temporal correlations between co-located patches or by spatial correlations in tube masking [57], resulting in inadequate motion awareness in the learned representations. This turn leads to poor generalization on downstream tasks, requiring a finer motion understanding as shown in [55]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we address these limitations by proposing a novel framework for masked video modeling, dubbed as SMILE (Spatial and Motion semantics in masked video LEarning), to capture high-level spatial and motion semantics leading to generalizable video representations. To improve the spatial semantics in masked video modeling, we propose to explicitly infuse high-level spatial semantics from pretrained image foundation models such as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8438" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "CLIP [45]. Such models encapsulate high-level semantics learned through large-scale image-text alignment and have been adapted for video understanding tasks. For example, ActionCLIP [61], ViFiCLIP [47], and X-CLIP [36]) adapt CLIP for action recognition using videos with action labels, whereas ViCLIP [64] and VideoPrism [72]) adapt it for video-text alignment with video-text pairs. We take inspiration from them but aim to learn video representations with unlabeled videos. In particular, we propose replacing the reconstruction of pixels in masked video modeling with that of CLIP embedding, thus providing high-level semantics as the supervision signal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.501 + ], + "angle": 0, + "content": "To improve motion awareness, we propose to reduce the temporal redundancy in input training videos with synthetic motion patterns. Specifically, we randomly paste objects onto the video frames and move them smoothly along the temporal dimension to simulate variable temporal transitions of motion-intensive videos. This reduces the temporal redundancy in the original videos and results in videos with an additional prominent foreground motion. Furthermore, adding objects with known locations allows us to add a motion focus to the mask-and-predict task by explicitly masking the regions of object motion. This in turn encourages the network to avoid an over-reliance on static spatial information and encode the motion dynamics of the added objects, when solving the reconstruction task. We summarize our contributions as follows." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.503, + 0.483, + 0.578 + ], + "angle": 0, + "content": "- We introduce SMILE, a novel SSL approach for masked video modeling. It encodes spatial and motion semantics using guidance from image-text pretrained models and synthetic-motion augmented videos, leading to more discriminative and generalizable video representation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.579, + 0.483, + 0.639 + ], + "angle": 0, + "content": "- We present a new strategy to augment video data with diverse synthetic motions, paired with a masking strategy that focuses on reconstructing these added motions to improve masked video modeling pretraining." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.64, + 0.483, + 0.729 + ], + "angle": 0, + "content": "- Through extensive evaluation on 7 datasets, including Kinetics-400 [26], UCF-101 [52], HMDB-51 [29], FineGYM [50], EPIC-Kitchens-100 [8], SomethingSomething V2 [18], and Charades [51], we demonstrate that SMILE significantly outperforms existing state-of-the-art video SSL methods and exhibits better generalization." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.503, + 0.483, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.742, + 0.24, + 0.758 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video Self-Supervised Learning. Video SSL aims to learn the spatio-temporal dynamics of video data without the need for any supervision. Three main learning paradigms have emerged: transformation prediction, contrastive learning, and masked video modeling. In transformation prediction, models learn video patterns by solving tasks such as space-time puzzles [1, 25, 27], predicting order in video clips [15, 24, 30, 33, 34, 66], or estimating speed [3, 7, 69]. Contrastive learning is based on instance discrimina" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "tion [38] and learning invariances to spatio-temporal augmentations [2, 6, 14, 35, 40, 42, 44, 46, 54, 56]. Masked video modeling methods [12, 21, 53, 57, 67], like VideoMAE [57], use a mask-and-predict task to reconstruct videos from partial data with an encoder-decoder setup similar to MAE [20]. By masking a high proportion of pixels (e.g., \\(90\\%\\)) and reconstructing them such methods learn useful video representations. Our approach aims to improve the spatial and motion semantics of the video representations learned by masked video modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.905, + 0.548 + ], + "angle": 0, + "content": "CLIP for Video Understanding. CLIP [45] captures detailed spatial semantics from extensive image-text training, making it a strong foundation for directly modeling video data with added temporal dynamics, e.g., [36, 41, 58, 61, 64, 68]. ViFiCLIP [47] finetunes CLIP on video by adding a temporal pooling layer, ActionCLIP [61] includes a temporal positional embedding and aggregation layer, and X-CLIP [36] integrates cross-frame attention to share information between frames. Most relevant to us are works that also use CLIP for video representation learning [13, 31, 64]. ViCLIP [64] initializes a model with CLIP weights and employs further contrastive pretraining with video-text pairs. UMT [31] uses CLIP as a teacher and aligns its predictions with the CLIP features for unmasked tokens, followed by a 2nd stage of learning from video-text contrastive pairs. TGM [13] relies on CLIP text-encoder features to guide masked video modeling and video-text contrastive learning in a unified framework. In contrast, our SMILE only learns from video data in a self-supervised manner, reconstructing CLIP features of masked tokens and added objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.905, + 0.763 + ], + "angle": 0, + "content": "Motion-aware Video Self-Supervised Learning. Encoding motion dynamics in video representations is essential for generalizing across various video tasks [55]. As a result, many video SSL methods focus on enhancing the motion sensitivity of learned representations. For example, [16, 19, 21, 37, 65] capture motion using optical flow, [9, 60] do so by removing background or static elements and [11, 22, 56, 59] incorporate synthetic motions. [22, 59] learn by predicting the speed and trajectory of the pseudo motions, and [56] contrasts video with different types of motions. These approaches are specific to transformation prediction and contrastive learning paradigms. We take inspiration from these works and inject synthetic motions into the masked video learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Most relevant to our work are methods like [12, 21, 53, 67], which enhance motion understanding in VideoMAE [57] by modifying certain design choices. MGMAE [21] and MGM [12] improve masking strategies to mask the regions of motion using optical flow and motion vectors. MotionMAE [67] adds frame difference as an additional target for reconstruction, and MME [53] uses optical flow to extract dense object trajectories and then predicts these trajectories and their HOG features. However, these" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "methods are limited by the temporal redundancies in the original data, restricting the amount of motion or motion-focused regions available for learning in the reconstruction task. In contrast, we add synthetic motions to the original video data, reducing temporal redundancy and increasing motion-rich regions to learn from. As shown in our experiments, reconstructing these modified videos encourages the model to capture motion dynamics, enhancing the motion awareness of the learned video representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.235, + 0.226, + 0.254 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.26, + 0.396, + 0.277 + ], + "angle": 0, + "content": "3.1. Revisiting Masked Video Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.463 + ], + "angle": 0, + "content": "The goal of masked video modeling [12, 21, 57] is to learn video representations by employing a mask-and-predict task with an asymmetric encoder-decoder architecture. An input video \\( V \\in \\mathbb{R}^{T \\times H \\times W \\times 3} \\) is first divided into non-overlapped 3D space-time tokens (e.g., \\( 2 \\times 16 \\times 16 \\)) as \\( \\mathcal{T} = \\{v_1, \\dots, v_N\\} \\). From \\( \\mathcal{T} \\), a subset of tokens \\( \\mathcal{T}^{mask} \\) are randomly masked with a high masking ratio (e.g., \\( 90\\% \\)) and the unmasked tokens \\( \\mathcal{T}^{unmask} = \\mathcal{T} \\setminus \\mathcal{T}^{mask} \\) are fed into a transformer encoder \\( \\Phi_{enc} \\). Then, the output of the encoder and the learnable tokens containing the positional information of the masked tokens are passed to a shallow decoder \\( \\Phi_{dec} \\) to reconstruct the input pixel with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.473, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{| \\mathcal {T} ^ {\\text {m a s k}} |} \\sum_ {i \\in \\mathcal {T} ^ {\\text {m a s k}}} \\| v _ {i} - Y _ {i} \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.581 + ], + "angle": 0, + "content": "where \\( v_{i} \\) refers to the pixel values of the \\( i \\)-th space-time token in \\( \\mathcal{T}^{mask} \\) and \\( Y_{i} = \\Phi_{dec}(\\Phi_{enc}(\\mathcal{T}^{unmask})) \\) denotes the corresponding reconstructed pixels for \\( i \\)-th token given the set of all unmasked tokens \\( \\mathcal{T}^{unmask} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Such pixel reconstruction drives the model to focus on low-level features, and with high temporal redundancy in video data, it primarily reconstructs static appearance details of video data. However, effective video understanding also requires capturing temporal progression and high-level semantics, such as scene context and object interactions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.336, + 0.699 + ], + "angle": 0, + "content": "3.2. General Scheme of SMILE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Our goal is to infuse spatial and motion semantics in masked video modeling. To achieve this, we integrate two key elements into the masked video modeling paradigm. To improve motion semantics, we overlay local motion patterns in input videos in the form of moving objects. To improve spatial semantics, we replace the low-level pixel targets with high-level features extracted from a pretrained CLIP encoder. By combining both, we learn video representations that can encode both spatial and motion dynamics necessary for modeling downstream video tasks. The overview of our method is shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Motion Infusion. Given a input video \\(V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\). We first sample a random object \\(o\\) from a set of predefined" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "segmented objects \\( O \\). Next, we generate a motion trajectory (Section 3.3), along which we overlay this object onto the video \\( V \\) to obtain the motion-augmented video \\( V' \\). By adding objects with synthetic motions, we reduce temporal redundancy in the original videos and encourage the model to learn from the motion dynamics of the overlaid objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Masking. We employ two separate masking strategies on \\( V' \\) as follows. For the original regions, we employ random tube masking as in VideoMAE [57]. For added object regions, we propose trajectory-based masking. Since we know the location of the object in each frame of \\( V' \\), we explicitly drop space-time tokens at random along its trajectory. Employing a high masking ratio along the object trajectory ensures that solving the reconstruction task requires encoding the object's motion dynamics, thereby enhancing the motion awareness of learned representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.906, + 0.502 + ], + "angle": 0, + "content": "Architecture. We employ a teacher-student reconstruction architecture with a pretrained image-encoder as the teacher and an asymmetric encoder-decoder network \\(\\Phi\\) as the student. The augmented video is patchified into 3D space-time tokens \\(\\mathcal{T} = \\{v_1',\\dots,v_N'\\}\\) as in general video masked autoencoders (Section 3.1). The tokens are then masked using tube and trajectory-based masking to generate sets of masked \\(\\mathcal{T}^{mask}\\) and unmasked \\(\\mathcal{T}^{unmask}\\) tokens. The un-masked tokens \\(\\mathcal{T}^{unmask}\\) and the learnable tokens containing the positional information of the masked patches are used as the input to \\(\\Phi\\) as in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Reconstruction. Instead of directly reconstructing pixels, we project the input video \\( V' \\) onto the teacher's feature space to generate high-level targets. In particular, we choose a CLIP model as the projection network due to its high-level semantics learned from language guidance. Specifically, the input \\( V' \\) is passed frame-by-frame through the CLIP encoder to extract features. The extracted features are stacked into a sequence of feature tokens as \\( \\mathcal{F} = \\{f_1', \\dots, f_N'\\} \\). To match the number of tokens with \\( \\mathcal{T} \\), we only use the CLIP features for the first time slice of the 3D space-time tokens, thus generating a total of \\( N \\) feature tokens, each corresponding to a space-time token in the input video. The encoder-decoder is then trained with a masked feature reconstruction loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.725, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{| \\mathcal {T} ^ {\\text {m a s k}} |} \\sum_ {i \\in \\mathcal {T} ^ {\\text {m a s k}}} \\| f _ {i} ^ {\\prime} - Y _ {i} \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{i}^{\\prime} \\) is the CLIP feature projection of \\( i \\)-th space-time token \\( v_{i}^{\\prime} \\) in \\( T^{mask} \\) and \\( Y_{i} = \\Phi_{dec}(\\Phi_{enc}(\\mathcal{T}^{unmask})) \\) denotes the corresponding reconstructed feature for \\( i \\)-th token given the set of all unmasked tokens \\( T^{unmask} \\). Such formulation first increases the temporal dynamics in otherwise static data, avoiding learning shortcuts for reconstruction. Second, projecting targets onto the CLIP space encourages the network to capture more high-level video semantics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8440" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.077, + 0.862, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 2. Overall architecture of our SMILE. An input video clip \\( V \\) is overlaid with a segmented object along a randomly generated trajectory to generate \\( V' \\) infusing synthetic object motion in \\( V \\). \\( V' \\) is passed frame-by-frame through the CLIP encoder to extract feature tokens \\( \\mathcal{F} \\). \\( V' \\) is patched into a set of space-time tokens \\( \\mathcal{T} \\). We apply two types of masking upon \\( \\mathcal{T} \\), tube masking on the space-time tokens of the original video and trajectory-based masking on the tokens of added objects. The unmasked tokens \\( \\mathcal{T}_{unmask} \\) are fed into the encoder-decoder network \\( \\Phi_{enc} \\Phi_{dec} \\) which is trained to reconstruct the masked feature tokens in \\( \\mathcal{F} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Overall, the key intuition is that CLIP features capture high-level information, such as scene context and object interactions, which guides the network to attend to both the overlaid object motions and the original video semantics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.276, + 0.544 + ], + "angle": 0, + "content": "3.3. Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.611 + ], + "angle": 0, + "content": "To infuse motion, we overlay a segmented object along a randomly generated trajectory onto the original video frames. To introduce more complex dynamics, the object is also transformed e.g. rotation, along the trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Appearance and Trajectory. To infuse a new synthetic motion onto a video sequence \\( V = [v_{1}, v_{2}, \\dots, v_{T}] \\) comprising \\( T \\) frames, we begin by selecting an object \\( o \\) from \\( O \\) at random, where \\( O \\) denotes the set of off-the-shelf segmented objects generated using Stable Diffusion [48] and X-Paste[71]. We then sample a scale \\( P \\times Q \\) from a predefined set of sizes for \\( o \\) to model different object sizes, where \\( P \\ll H \\) and \\( Q \\ll W \\). Next, we sample a sequence of locations for the object \\( \\mathrm{Traj}_o = [(x^1, y^1),.., (x^T, y^T)] \\) representing the center coordinate \\( (x^i, y^i) \\) of the object \\( o \\) in the corresponding frame. Specifically, we generate a random non-linear path for the trajectory by sampling \\( M \\) 2D points \\( (M \\gg T) \\) uniformly from \\( x \\in [0, H] \\) and \\( y \\in [0, W] \\). To ensure smoothness in the motion, a 1D Gaussian filter is applied along \\( x \\) and \\( y \\) dimensions as \\( [(h(x^1), h(y^1)), \\dots, (h(x^M), h(y^M))] \\), where \\( h(z) = \\frac{1}{\\sqrt{2\\pi}\\kappa} e^{-z^2 / 2\\kappa^2} \\) and \\( \\kappa \\) is the smoothing factor. The resulting trajectory is then downsampled to \\( T \\) target locations for the object \\( o \\). This formulation introduces diverse motions" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.451, + 0.883, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.574, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 3. Motion overlaid videos showing the position change and transformation of the added objects along the time dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.825 + ], + "angle": 0, + "content": "with varying spatio-temporal dynamics in the form of objects with different shapes, sizes, and temporal progression. Transformation. To increase the complexity of the object's motion, we also apply scale and rotation transformations along the temporal dimension. Specifically, we select 3 frame indices, the first, the last, and a random middle one, \\( v_{1}, v_{m} \\), and \\( v_{T} \\), respectively. For these three frame indices, a rotation angle \\( \\alpha_{i} \\), and a scaling factor \\( S_{i} \\) are selected randomly. The angles and scales for the rest of the frames are linearly interpolated between these keyframes. Finally, the object \\( o \\) is rotated and scaled with the corresponding transformations \\( \\left( \\begin{array}{cc}\\cos \\alpha_{i} & -\\sin \\alpha_{i}\\\\ \\sin \\alpha_{i} & \\cos \\alpha_{i} \\end{array} \\right) \\) and \\( \\left( \\begin{array}{cc}S_i & 0\\\\ 0 & S_i \\end{array} \\right) \\), before overlaying it onto the video frame \\( v_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Figure 3 shows two sample videos with our added motions. While these motions may not appear entirely realistic, they represent complex motion dynamics of moving structured pixels, and introducing them into a mask-and-predict framework guides the model to learn from such dynamics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8441" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "This results in video representations with improved downstream performance on motion-aware tasks, as we also validate empirically. Interestingly, we also show in the experiments (Section 5) that overlaying such object motions on black or noise frames can learn strong video representations without requiring any original videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.183, + 0.279, + 0.199 + ], + "angle": 0, + "content": "3.4. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.483, + 0.475 + ], + "angle": 0, + "content": "By adding synthetic motions, we have introduced a domain shift in the input for reconstruction, which may affect the performance of downstream tasks. To explore effective learning from both original videos \\( V \\) and motion-overlaid videos \\( V' \\), we explore two training strategies, i.e. mixed training and progressive training. In the mixed training strategy, we use both \\( V \\) and \\( V' \\) as the input to the network and directly optimize for the reconstruction of both inputs. The idea is to simultaneously encode the spatio-temporal dynamics of the original video data and the motion dynamics of the augmented data, with a limited domain shift. In the progressive training strategy, we employ a two-stage learning approach that first learns from augmented data \\( V' \\) alone, then continues to train the model with the original data \\( V \\) alone. This way we can first learn to encode the spatio-temporal dynamics of the augmented data and then align the network to further encode the dynamics of the original data and its domain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.223, + 0.499 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.637 + ], + "angle": 0, + "content": "We evaluate SMILE against prior video SSL methods on action recognition benchmarks: Kinetics-400 (K400) [26], SomethingSomething V2 [18] (SSv2), UCF-101 [52] (UCF), HMDB-51 [29] (HMDB), FineGYM [50] (GYM), and EPIC-Kitchens-100 [8] (EPIC), achieving the best performance in linear probing and full finetuning (Section 4.2). SMILE also shows better generalization on the SEVERE benchmark[55] (Section 4.4). Ablations and further discussion are provided in Section 4.3 and Section 5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.641, + 0.308, + 0.657 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Model. Following VideoMAE [57], we use ViT-S or ViT-B with joint space-time attention for our student network. We choose CLIP ViT-B [45] as our teacher network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.781 + ], + "angle": 0, + "content": "Motions. We follow PIN [10] to create the set of segmented objects \\( O \\) with Stable Diffusion [48] and X-Paste [71]. We sample the object size \\( P \\times Q \\) uniformly from \\( [32 \\times 32, 128 \\times 128] \\) and sample the rotation angles from \\( [-90, 90] \\) and scaling factors from [0.5, 1.5]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Pretraining. We follow previous masked video modeling works [12, 21, 53, 57] and use Kinetics-400 [26] (K400) or SomethingSomething V2 [18] (SSv2) for pretraining unless specified otherwise. We use the same hyperparameters as in VideoMAE [57]. Unless stated otherwise, we train with our proposed progressive training for a total of 600 epochs, 300 epochs in each stage. More training details are in supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "Table 1. Linear probing comparison of masked video modeling methods. Our SMILE learns highly discriminative video representations useful for both spatial and motion-focused domains. All results are obtained with the ViT-B backbone pretrained on K400 and Top-1 accuracies \\((\\%)\\) are reported." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.32 + ], + "angle": 0, + "content": "
MethodK400UCFHMDBSSv2GYMEPIC
VideoMAE [57]20.758.637.717.523.933.2
MVD [63]18.749.128.612.222.729.7
MME [53]19.156.037.116.629.032.2
MGMAE [21]24.964.441.316.826.133.2
EVEREST [23]14.151.830.314.523.330.5
MGM [12]19.862.540.321.725.832.4
SIGMA [49]47.580.752.320.830.134.2
SMILE w/o motion54.682.651.721.627.932.3
SMILE (ours)56.283.853.423.730.234.4
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.348, + 0.81, + 0.364 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.905, + 0.503 + ], + "angle": 0, + "content": "Linear Probing. We freeze the pretrained backbone and only train a linear classifier on the target dataset. When evaluating the prior works for comparison, we utilize their publicly available checkpoints for the ViT-B backbone pretrained on the K400 dataset for 800 epochs except for MVD (2000 epochs) and MME (1600 epochs). We evaluate the downstream task using a common setup for a fair comparison. Details about datasets and experimental settings are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Table 1 shows that our method, even without synthetic motion, significantly outperforms all state-of-the-art methods on K400 and UCF, and is among the top performers on HMDB, SSv2, GYM, and EPIC. This highlights that CLIP feature reconstruction offers a stronger supervisory signal than pixel reconstruction, especially in tasks requiring spatial semantics like K400 and UCF. Adding synthetic motions further boosts performance by an average of \\(1.8\\%\\) across datasets, enhancing motion cues through reconstruction. On motion-sensitive datasets like SSv2, GYM, and EPIC, our approach consistently surpasses MGM [12], MGMAE [21], and MME [53], confirming that synthetic motions help overcome the lack of motion diversity in current datasets that may limit these methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Full Finetuning. We finetune both the pretrained backbone and the classification head end-to-end following [21, 53, 57] on the downstream datasets K400 and SSv2. Following prior masked video modeling works [12, 21, 57], we show results with self-supervised pretraining on K400 and SSv2 datasets for the VIT-B backbone. We follow VideoMAE [57] for finetuning and evaluation protocols with more details in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Tables 2 and 3 show the results for SSv2 and K400, respectively. On SSv2, we achieve the best performance with both K400 and SSv2 pretraining, highlighting our method's effectiveness for both in-domain and cross-domain transfer." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8442" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.085, + 0.482, + 0.155 + ], + "angle": 0, + "content": "Table 2. Full finetuning comparison of various self-supervised methods on Something-Something V2. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.158, + 0.457, + 0.381 + ], + "angle": 0, + "content": "
MethodBackboneEpochsPretrainTop-1
VideoMAE [57]ViT-B800K40068.5
OmniMAE [17]ViT-B800K40069.0
MGMAE* [21]ViT-B800K40068.9
MME [53]ViT-B800K40070.5
SIGMA [49]ViT-B800K40071.1
MGM* [12]ViT-B800K40071.1
SMILE (ours)ViT-B600K40072.1
OmniMAE [17]ViT-B800SSv269.5
VideoMAE [57]ViT-B800SSv269.6
MGM [12]ViT-B800SSv270.6
MME [53]ViT-B800SSv270.0
SIGMA [49]ViT-B800SSv271.2
MGMAE [21]ViT-B800SSv271.0
SMILE (ours)ViT-B800SSv272.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.56 + ], + "angle": 0, + "content": "For K400, we exceed the prior top SSL approach by \\(1.6\\%\\). Additionally, our gains over motion-focused methods, such as MGMAE (1.5% on SSv2, 1.9% on K400), MME (2.5% on SSv2, 1.6% on K400), and MGM (1.9% on SSv2, 2.3% on K400), confirm our improved motion sensitivity, even in full-finetuning settings. Notably, our model surpasses VideoMAE, MotionMAE, MME, and MGMAE trained for more than twice the number of epochs (see supplementary for the extensive comparison). We attribute this efficiency to CLIP feature reconstruction, which directly captures high-level semantics compared to pixel-level learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.203, + 0.579 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.719 + ], + "angle": 0, + "content": "We ablate various components of our SMILE to show their impact on the downstream performance in Table 4. For computational efficiency, we use ViT-S as the backbone and pretrain with a smaller subset of K400, denoted as \\(\\mathrm{K400_m}\\), containing around 80K samples. For evaluation, we use \\(\\mathrm{K400_m}\\) and \\(\\mathrm{SSv2_m}\\) that denotes a subset of SSv2 with \\(50\\%\\) training data. Unless stated otherwise, we use CLIP features as the reconstruction target, add two synthetic objects, mask \\(90\\%\\), use trajectory masking, and train for 400 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Synthetic motion, yes or no? In Table 4a, we show the impact of adding synthetic motions to the masked video modeling framework for both pixel and feature reconstruction. We can see that adding our synthetic motion patterns significantly improves the downstream performance of both pixel and feature reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Reconstructing pixels or features? Table 4h compares pixel-based and feature-based reconstruction targets in masked video modeling. Reconstructing in feature space significantly boosts performance over pixel reconstruction, with a \\(+12.1\\%\\) gain on spatially-focused tasks like \\(\\mathrm{K400_m}\\) and \\(+5.1\\%\\) on motion-focused tasks like \\(\\mathrm{SSv2_m}\\). This sug-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.085, + 0.907, + 0.154 + ], + "angle": 0, + "content": "Table 3. Full finetuning comparison of various self-supervised methods on Kinetics-400. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.158, + 0.878, + 0.309 + ], + "angle": 0, + "content": "
MethodBackboneEpochsPretrainTop-1
SVT [46]ViT-B-K40078.4
VideoMAE [57]ViT-B800K40080.0
BEVT [62]ViT-B-K40080.6
OmniMAE [17]ViT-B800K40080.8
MGM [12]ViT-B800K40080.8
MGMAE [21]ViT-B800K40081.2
SIGMA [49]ViT-B800K40081.5
MME* [53]ViT-B800K40081.5
SMILE (ours)ViT-B600K40083.1
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.324, + 0.905, + 0.368 + ], + "angle": 0, + "content": "gests that using a high-level feature space as the reconstruction target provides more meaningful supervision for learning transferable spatio-temporal representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.369, + 0.905, + 0.505 + ], + "angle": 0, + "content": "Reconstructing which features: CLIP, MAE or DINO? Table 4b compares several image foundation models—MAE [20], DINO [4], DINO-v2 [39], and CLIP [45]—as reconstruction targets. CLIP features outperform the others, likely due to CLIP's image-text alignment training on a large dataset, enhancing its transferability and adaptability to video tasks as shown in [36, 47, 61]. Compared to pixel reconstruction in Table 4h, all feature-based approaches show substantial performance gains." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.506, + 0.906, + 0.672 + ], + "angle": 0, + "content": "Trajectory-based masking: yes or no? Table 4g shows the impact of our trajectory-based masking. Without trajectory masking, all tokens are masked using tube masking, which masks the whole video without distinguishing between object and non-object space-time tokens. In contrast, trajectory masking first masks space-time tokens associated with objects, followed by tube masking on the remaining tokens, maintaining the overall masking ratio. Our trajectory masking improves performance by approximately \\(1\\%\\) on both \\(\\mathrm{K400_m}\\) and \\(\\mathrm{SSv2_m}\\), validating that masking along object trajectories enhances motion-aware representation learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.673, + 0.906, + 0.748 + ], + "angle": 0, + "content": "How much to mask? Table 4e shows the impact of masking ratio \\( m \\) on downstream performance. We mask \\( m \\) of space-time tubes within the objects and \\( m \\) of space-time tubes in the whole video with \\( m \\) being \\( 95\\% \\), \\( 90\\% \\), or \\( 80\\% \\). We observe the best performance with \\( m = 80\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.749, + 0.906, + 0.84 + ], + "angle": 0, + "content": "How many synthetic objects? In Table 4f, we analyze the impact of adding more object motions to the input video on downstream performance. We evaluate adding 1, 2 and 3 object motions and compare the results. We observe that adding 2 objects is better than using a single object and the performance saturates when adding more objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Which object transformations? In Table 4c, we show the impact of object transformation on downstream performance. Both rotation and scaling improve the performance individually and combining them achieves the best results." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8443" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.322, + 0.173 + ], + "angle": 0, + "content": "
Synth.TargetK400mSSv2m
w/oPixels66.055.1
w/Pixels68.156.8
w/oFeatures78.160.2
w/Features78.961.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.175, + 0.321, + 0.211 + ], + "angle": 0, + "content": "(a) Synthetic motion, yes or no? Yes. It improves performance for both pixel and feature targets." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.214, + 0.313, + 0.284 + ], + "angle": 0, + "content": "
RatioTargetK400mSSv2m
95%features77.760.4
90%features78.961.6
80%features80.362.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.286, + 0.315, + 0.31 + ], + "angle": 0, + "content": "(e) How much to mask? Masking \\(80\\%\\) trajectory and total tokens." + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.09, + 0.516, + 0.172 + ], + "angle": 0, + "content": "
TargetK400mSSv2m
MAE [20]77.659.5
DINO [4]77.659.7
DINOv2 [39]77.859.7
CLIP [45]78.160.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.33, + 0.175, + 0.512, + 0.211 + ], + "angle": 0, + "content": "(b) Reconstructing which features? CLIP features provide the best target. No motion is added." + }, + { + "type": "table", + "bbox": [ + 0.335, + 0.215, + 0.501, + 0.284 + ], + "angle": 0, + "content": "
# ObjectsK400mSSv2m
178.661.0
278.961.6
378.661.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.328, + 0.285, + 0.508, + 0.31 + ], + "angle": 0, + "content": "(f) How many objects? Adding two object motions is the best." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.091, + 0.695, + 0.172 + ], + "angle": 0, + "content": "
TransformK400mSSv2m
None77.060.8
Scaling78.461.3
Rotation78.261.4
Both78.961.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.175, + 0.691, + 0.211 + ], + "angle": 0, + "content": "(c) Which object transformations? Using both rotation and scaling works the best." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.215, + 0.714, + 0.272 + ], + "angle": 0, + "content": "
MaskingK400mSSv2m
w/o Trajectory77.860.7
w/ Trajectory78.961.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.274, + 0.699, + 0.308 + ], + "angle": 0, + "content": "(g) Trajectory masking, yes or no? Yes. Both targets are features." + }, + { + "type": "table", + "bbox": [ + 0.703, + 0.091, + 0.882, + 0.161 + ], + "angle": 0, + "content": "
TrainingK400mSSv2m
Single78.961.6
Mix79.562.2
Progressive79.862.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.701, + 0.163, + 0.882, + 0.21 + ], + "angle": 0, + "content": "(d) Which training strategy? Progressive training outperforms both for the same number of iterations." + }, + { + "type": "table", + "bbox": [ + 0.716, + 0.215, + 0.88, + 0.272 + ], + "angle": 0, + "content": "
TargetK400mSSv2m
Pixels66.055.1
Features78.160.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.707, + 0.274, + 0.889, + 0.309 + ], + "angle": 0, + "content": "(h) Reconstructing pixels or features? Reconstructing features is much better. No motion is added." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Table 4. Ablation experiments of our proposed SMILE. The default setting uses clip reconstruction targets with two synthetic objects and a \\(90\\%\\) masking ratio with trajectory-based masking. We also show full-scale ablations with ViT-B and K400 pretraining in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.483, + 0.493 + ], + "angle": 0, + "content": "Which training strategy? Table 4d compares our proposed training strategies. We also compare learning with the single learning strategy, where only augmented videos are used for training. We observe that both mixed and progressive learning are better than learning from augmented videos only. Progressive is better than mixed, most likely due to a reduced domain shift in the second stage where only original videos are used." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.494, + 0.484, + 0.585 + ], + "angle": 0, + "content": "Qualitative analysis. To compare the temporal dynamics of video representations learned by different SSL methods, we compute their feature similarity across different frames of input videos (Figure 4). For our model, the features of different frames have larger differences, indicating better temporal awareness. More plots are in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.6, + 0.469, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Figure 4. Feature similarity across different frames for different SSL methods. We compute this on K400 validation videos." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.745, + 0.334, + 0.76 + ], + "angle": 0, + "content": "4.4. Generalization Assessment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.885 + ], + "angle": 0, + "content": "To demonstrate the generalization of our learned video representations to diverse downstream video setups, we evaluate SMILE on the SEVERE benchmark [55], which includes 8 experiments targeting 4 key generalization factors: domain shift, sample efficiency, action granularity, and task shift. Detailed configurations are provided in the supplementary. Beyond SEVERE, we also show generalization to more temporal tasks in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Table 6 shows the results. For the Domain Shift, we test" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.736 + ], + "angle": 0, + "content": "cross-domain adaptability on SSv2 and Gym99, which differ from the pretraining dataset (K400). Our method significantly outperforms all others, demonstrating superior adaptability to domain shifts. For Sample Efficiency, we evaluate low-shot action recognition on UCF and GYM with only 1,000 samples for finetuning. Our approach surpasses previous methods, achieving a \\(5\\%\\) improvement on GYM over the best prior model, MME, showcasing its strong low-shot recognition capability even in motion-sensitive tasks. For the Action Granularity, tested on FX-S1 and UB-S1 from FineGym, our model achieves the best results, excelling at distinguishing fine-grained motions within gymnastics routines (e.g., jump variations in FX-S1). It outperforms motion-aware models like MME, MGMAE, and MGM, highlighting superior motion generalizability. For Task Shift, evaluated with temporal repetition counting on UCF-RC [70] and multi-label recognition on Charades [51], our method matches prior performance on repetition counting and achieves a \\(9\\%\\) improvement over MME on Charades, demonstrating broad task adaptability. Overall, SMILE (without synthetic motion) is on par with the prior best with the mean performance of \\(64.9\\%\\). With synthetic motion SMILE, we observe a further \\(3.0\\%\\) gain, reinforcing our model's generalization strength through synthetic motion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.636, + 0.761 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Comparison with Other CLIP Adaptations. Table 5 presents a comparison between our SMILE and the CLIP adaptation methods ViCLIP [64] and UMT [31], along with the CLIP baseline. Our method achieves superior performance on motion-sensitive datasets, particularly in linear probing, indicating that our CLIP adaptation learns better video representations than the video-text alignment of ViCLIP and the unmasked feature alignment of UMT, despite using less data. More results are in the supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8444" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.641, + 0.201 + ], + "angle": 0, + "content": "
PretrainingFinetuningLinear Probing
MethodVideo DatasetUCFGYMSSv2EPICUCFGYMSSv2EPIC
CLIP [45]-93.688.066.750.377.520.711.325.1
ViCLIP [64]Intervid-10M [64]95.289.767.955.086.727.318.927.3
UMT [31]K700 [5]96.089.970.150.188.026.418.828.2
SMILE (ours)K400 [26]96.490.871.963.383.830.223.734.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.65, + 0.089, + 0.908, + 0.173 + ], + "angle": 0, + "content": "Table 5. Comparison with CLIP adaptations for video representation learning. Despite using less data, we learn better video representations than other CLIP adaptations, especially for motion-focused downstream domains." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.213, + 0.907, + 0.257 + ], + "angle": 0, + "content": "Table 6. Generalization assessment on SEVERE benchmark [55] for recent video SSL methods in terms of domain shift, sample efficiency, action granularity, and task shift. We evaluate prior works with the official ViT-B backbone checkpoints pretrained on the K400 dataset. SMILE shows the best generalization performance across setups. Adding synthetic motions further improves the performance." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.259, + 0.868, + 0.419 + ], + "angle": 0, + "content": "
MethodDomain shiftSample efficiency (103)Action granularityTask shiftMean
SSv2Gym99UCFGYMFX-S1UB-S1UCF-RC↓Charades
VideoMAE [57]68.686.674.625.942.865.30.17214.457.6
MVD [63]70.082.567.117.531.350.50.18416.152.1
MGMAE [21]68.987.277.224.133.779.50.18117.958.8
MGM [12]71.189.178.426.438.686.90.15222.562.2
SIGMA [49]70.989.784.128.055.179.90.16923.164.2
MME [53]70.189.779.229.855.587.20.15523.665.0
SMILE w/o motion71.690.085.232.053.474.30.17530.564.9
SMILE (ours)72.190.886.435.155.188.30.17032.567.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Table 7. Learning without natural videos. We train a ViT-S for pixel reconstruction. Our synthetic object motions can learn video representations without using any natural videos." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.479, + 0.459, + 0.61 + ], + "angle": 0, + "content": "
Unnatural Video SchemePretrain DataK400mSSv2m
No Pretraining-39.728.2
Video-frameK400m-1frm54.043.1
Video-frame + MotionK400m-1frm63.653.2
ImagePlaces-80K52.140.5
Image + MotionPlaces-80K60.152.1
Black image + Motion-59.250.0
Noise image + Motion-61.951.2
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Learning without Natural Videos. In prior experiments, we overlay our object motions onto natural videos from the Kinetics-400 dataset, raising the question of whether downstream improvements stem from original video priors (motion, action, scene) or our added object motions. To investigate this, we construct videos by overlaying object motions onto four different backgrounds: (1) a single frame from natural videos, (2) a single scene image from the image dataset Places [73], (3) a plain black image, and (4) a noise image. Each background image is duplicated across \\( T \\) frames, creating static clips onto which moving objects are placed. For (1), we select one frame per video from \\( \\mathrm{K400_m} \\) (denoted \\( \\mathrm{K400_m - 1frm} \\)), preserving action and scene cues without motion priors. For (2), we sample 80K images from Places (Places-80K), matching the dataset size of \\( \\mathrm{K400_m - 1frm} \\), to remove action and motion priors. Approaches (3) and (4) exclude all natural video or image priors entirely. Evaluation involves full finetuning on \\( \\mathrm{K400_m} \\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.434, + 0.892, + 0.449 + ], + "angle": 0, + "content": "and \\(\\mathrm{SSv2_m}\\), consistent with our previous ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Table 7 reports results for different augmented video types. Adding object motions consistently enhances downstream performance, whether combined with action and scene cues from natural videos (row 2 vs. row 3) or only scene cues from images (row 4 vs. row 5). Even with black or noisy backgrounds (rows 6 and 7), performance remains robust. Remarkably, using object motions alone yields substantial improvements—\\(22.2\\%\\) on \\(\\mathrm{K400_m}\\) and \\(23.0\\%\\) on \\(\\mathrm{SSv2_m}\\) compared to no pretraining. These findings demonstrate our method's effectiveness in learning video representations from synthetic motions alone. More results are in the supplementary. In summary, our method enables a new paradigm of learning strong video representations with masked video modeling without relying on natural videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.641, + 0.684 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we proposed SMILE, a novel self-supervised approach for video representation learning that addresses key limitations in existing masked video modeling methods by integrating both spatial and motion semantics. Our SMILE, leverages the high-level spatial semantics from image-language pretrained models like CLIP and enhances motion representation through synthetic motion patterns. By capturing complex spatial features and motion dynamics, SMILE achieves superior performance on diverse downstream tasks, demonstrating improved discrimination and generalization over state-of-the-art video SSL methods. Our extensive evaluations demonstrated the advantages of our SMILE, setting a new paradigm for robust and effective video self-supervised learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8445" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.212 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the KAUST Center of Excellence for Generative AI under award number 5940. For computing time, this research used Ibex managed by the Supercomputing Core Laboratory at King Abdullah University of Science & Technology (KAUST) in Thuwal, Saudi Arabia. We also thank Michael Dorkenwald for providing the synthetic objects used in this work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.224, + 0.188, + 0.24 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.486, + 0.317 + ], + "angle": 0, + "content": "[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.318, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.388, + 0.485, + 0.457 + ], + "angle": 0, + "content": "[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.484, + 0.526 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.527, + 0.483, + 0.568 + ], + "angle": 0, + "content": "[5] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.483, + 0.638 + ], + "angle": 0, + "content": "[6] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[7] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised visual learning by variable playback speeds prediction of a video. IEEE Access, 9:79562-79571, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. International Journal of Computer Vision (IJCV), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[9] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.248, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[11] Hazel Doughty, Fida Mohammad Thoker, and Cees GM Snoek. Locomotion: Learning motion-focused video-language representations. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 50-70, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[12] David Fan, Jue Wang, Shuai Liao, Yi Zhu, Vimal Bhat, Hector Santos-Villalobos, Rohith MV, and Xinyu Li. Motion-guided masking for spatiotemporal representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[13] David Fan, Jue Wang, Shuai Liao, Zhikang Zhang, Vimal Bhat, and Xinyu Li. Text-guided video masked autoencoder. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[14] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[15] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[16] Kirill Gavrilyuk, Mihir Jain, Ilia Karmanov, and Cees G M Snoek. Motion-augmented self-training for video recognition at smaller scale. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimaes: Single model masked pretraining on images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[19] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[21] Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, and Limin Wang. Mgmae: Motion guided masking for video masked autoencoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Ziyuan Huang, Shiwei Zhang, Jianwen Jiang, Mingqian Tang, Rong Jin, and Marcelo H Ang. Self-supervised motion" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8446" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "learning from static images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[23] Sunil Hwang, Jaehong Yoon, Youngwan Lee, and Sung Ju Hwang. Everest: Efficient masked video autoencoder by removing redundant spatiotemporal tokens. In International Conference on Machine Learning (ICML), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[24] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[25] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[26] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[27] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[28] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified CNN architecture for real-time spatiotemporal action localization. arXiv preprint arXiv:1911.06644, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[29] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[30] Hsin-Ying Lee, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[31] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[32] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[33] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[36] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[37] Jingcheng Ni, Nan Zhou, Jie Qin, Qian Wu, Junqi Liu, Boxun Li, and Di Huang. Motion sensitive contrastive learning for self-supervised video representation. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[40] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. arXiv preprint arXiv:2303.09857, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[42] Mandela Patrick, Yuki M. Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[43] AJ Piergiovanni, Anelia Angelova, and Michael S Ryoo. Evolving losses for unsupervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[44] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[46] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S. Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8447" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[49] Mohammadreza Salehi, Michael Dorkenwald, Fida Mohammad Thoker, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Sigma: Sinkhorn-guided masked video modeling. In European Conference on Computer Vision (ECCV), 2025. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[50] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[51] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in Homes: Crowdsourcing Data Collection for Activity Understanding. In European Conference on Computer Vision (ECCV), 2016. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[52] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[53] Xinyu Sun, Peihao Chen, Liangwei Chen, Changhao Li, Thomas H. Li, Mingkui Tan, and Chuang Gan. Masked motion encoding for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[54] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the ACM International Conference on Multimedia (ACMMM), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[56] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Tubelet-contrastive self-supervision for videoefficient generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[58] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. arXiv preprint arXiv:2304.10465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[59] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[60] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[61] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[62] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[64] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[65] Fanyi Xiao, Joseph Tighe, and Davide Modolo. Maclr: Motion-aware contrastive learning of representations for videos. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[66] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[67] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[68] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[69] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[70] Huaidong Zhang, Xuemiao Xu, Guoqiang Han, and Shengfeng He. Context-aware and scale-insensitive temporal repetition counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8448" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In International Conference on Learning Representations (ICLR), 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[72] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[73] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8449" + } + ] +] \ No newline at end of file diff --git a/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_origin.pdf b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fbce5c9b67967ff56785dfb42b4e77a607d1c826 --- /dev/null +++ b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/6f3fc16f-e100-416b-9a89-673ecf151cb4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d24a557c9e39178b4200a777f7d1409738fbe0537ad52407c73c12791525059 +size 4536054 diff --git a/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/full.md b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad830fa3c07e6ecd4b0c7b3979d87a5e1adbb8ca --- /dev/null +++ b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/full.md @@ -0,0 +1,322 @@ +# SMILE: Infusing Spatial and Motion Semantics in Masked Video Learning + +Fida Mohammad Thoker, Letian Jiang, Chen Zhao†, Bernard Ghanem King Abdullah University of Science and Technology (KAUST) + +{fida.thoker,letian.jiang,chen.zhao,Bernard.Ghanem}@kaust.edu.sa + +# Abstract + +Masked video modeling, such as VideoMAE, is an effective paradigm for video self-supervised learning (SSL). However, they are primarily based on reconstructing pixel-level details on natural videos which have substantial temporal redundancy, limiting their capability for semantic representation and sufficient encoding of motion dynamics. To address these issues, this paper introduces a novel SSL approach for video representation learning, dubbed as SMILE, by infusing both spatial and motion semantics. In SMILE, we leverage image-language pretrained models, such as CLIP, to guide the learning process with their high-level spatial semantics. We enhance the representation of motion by introducing synthetic motion patterns in the training data, allowing the model to capture more complex and dynamic content. Furthermore, using SMILE, we establish a new self-supervised video learning paradigm capable of learning strong video representations without requiring any natural video data. We have carried out extensive experiments on 7 datasets with various downstream scenarios. SMILE surpasses current state-of-the-art SSL methods, showcasing its effectiveness in learning more discriminative and generalizable video representations. Code is available: https://github.com/fmthoker/SMILE + +# 1. Introduction + +Self-supervised learning (SSL) [43, 44, 57, 59] plays a crucial role for many video understanding tasks, such as action recognition [18, 26, 51, 52, 55] and detection [28, 32], significantly reducing the reliance on extensive human annotation. The goal of video SSL is to learn video representations that capture both spatial and temporal semantics effectively. Masked video modeling has emerged as a promising approach with state-of-the-art performance [12, 21, 53, 57]. It involves masking a large portion of video frames and reconstructing the masked regions using the unmasked parts. + +However, this approach has some limitations. First, it relies heavily on pixel reconstruction, which can lead to overfitting low-level details. This limits the model's abl + +![](images/38cd8034fabbd90eb5913dd89adf8141947704afaf27dbbef3cca683613243ea.jpg) +Figure 1. Comparison with SOTA masked video modeling methods. Our method significantly outperforms prior masked video modeling methods across diverse downstream settings. + +ity to learn meaningful high-level semantics essential for discriminative representations, as shown by the suboptimal linear probing performance of current masked video modeling methods [12, 21, 53, 57] in Figure 1. The second key issue in masked video modeling is the substantial temporal redundancy of video data with most videos having little foreground motion, human or object. Consequently, the reconstruction objective can be solved using the temporal correlations between co-located patches or by spatial correlations in tube masking [57], resulting in inadequate motion awareness in the learned representations. This turn leads to poor generalization on downstream tasks, requiring a finer motion understanding as shown in [55]. + +In this work, we address these limitations by proposing a novel framework for masked video modeling, dubbed as SMILE (Spatial and Motion semantics in masked video LEarning), to capture high-level spatial and motion semantics leading to generalizable video representations. To improve the spatial semantics in masked video modeling, we propose to explicitly infuse high-level spatial semantics from pretrained image foundation models such as + +CLIP [45]. Such models encapsulate high-level semantics learned through large-scale image-text alignment and have been adapted for video understanding tasks. For example, ActionCLIP [61], ViFiCLIP [47], and X-CLIP [36]) adapt CLIP for action recognition using videos with action labels, whereas ViCLIP [64] and VideoPrism [72]) adapt it for video-text alignment with video-text pairs. We take inspiration from them but aim to learn video representations with unlabeled videos. In particular, we propose replacing the reconstruction of pixels in masked video modeling with that of CLIP embedding, thus providing high-level semantics as the supervision signal. + +To improve motion awareness, we propose to reduce the temporal redundancy in input training videos with synthetic motion patterns. Specifically, we randomly paste objects onto the video frames and move them smoothly along the temporal dimension to simulate variable temporal transitions of motion-intensive videos. This reduces the temporal redundancy in the original videos and results in videos with an additional prominent foreground motion. Furthermore, adding objects with known locations allows us to add a motion focus to the mask-and-predict task by explicitly masking the regions of object motion. This in turn encourages the network to avoid an over-reliance on static spatial information and encode the motion dynamics of the added objects, when solving the reconstruction task. We summarize our contributions as follows. + +- We introduce SMILE, a novel SSL approach for masked video modeling. It encodes spatial and motion semantics using guidance from image-text pretrained models and synthetic-motion augmented videos, leading to more discriminative and generalizable video representation. +- We present a new strategy to augment video data with diverse synthetic motions, paired with a masking strategy that focuses on reconstructing these added motions to improve masked video modeling pretraining. +- Through extensive evaluation on 7 datasets, including Kinetics-400 [26], UCF-101 [52], HMDB-51 [29], FineGYM [50], EPIC-Kitchens-100 [8], SomethingSomething V2 [18], and Charades [51], we demonstrate that SMILE significantly outperforms existing state-of-the-art video SSL methods and exhibits better generalization. + +# 2. Related Works + +Video Self-Supervised Learning. Video SSL aims to learn the spatio-temporal dynamics of video data without the need for any supervision. Three main learning paradigms have emerged: transformation prediction, contrastive learning, and masked video modeling. In transformation prediction, models learn video patterns by solving tasks such as space-time puzzles [1, 25, 27], predicting order in video clips [15, 24, 30, 33, 34, 66], or estimating speed [3, 7, 69]. Contrastive learning is based on instance discrimina + +tion [38] and learning invariances to spatio-temporal augmentations [2, 6, 14, 35, 40, 42, 44, 46, 54, 56]. Masked video modeling methods [12, 21, 53, 57, 67], like VideoMAE [57], use a mask-and-predict task to reconstruct videos from partial data with an encoder-decoder setup similar to MAE [20]. By masking a high proportion of pixels (e.g., $90\%$ ) and reconstructing them such methods learn useful video representations. Our approach aims to improve the spatial and motion semantics of the video representations learned by masked video modeling. + +CLIP for Video Understanding. CLIP [45] captures detailed spatial semantics from extensive image-text training, making it a strong foundation for directly modeling video data with added temporal dynamics, e.g., [36, 41, 58, 61, 64, 68]. ViFiCLIP [47] finetunes CLIP on video by adding a temporal pooling layer, ActionCLIP [61] includes a temporal positional embedding and aggregation layer, and X-CLIP [36] integrates cross-frame attention to share information between frames. Most relevant to us are works that also use CLIP for video representation learning [13, 31, 64]. ViCLIP [64] initializes a model with CLIP weights and employs further contrastive pretraining with video-text pairs. UMT [31] uses CLIP as a teacher and aligns its predictions with the CLIP features for unmasked tokens, followed by a 2nd stage of learning from video-text contrastive pairs. TGM [13] relies on CLIP text-encoder features to guide masked video modeling and video-text contrastive learning in a unified framework. In contrast, our SMILE only learns from video data in a self-supervised manner, reconstructing CLIP features of masked tokens and added objects. + +Motion-aware Video Self-Supervised Learning. Encoding motion dynamics in video representations is essential for generalizing across various video tasks [55]. As a result, many video SSL methods focus on enhancing the motion sensitivity of learned representations. For example, [16, 19, 21, 37, 65] capture motion using optical flow, [9, 60] do so by removing background or static elements and [11, 22, 56, 59] incorporate synthetic motions. [22, 59] learn by predicting the speed and trajectory of the pseudo motions, and [56] contrasts video with different types of motions. These approaches are specific to transformation prediction and contrastive learning paradigms. We take inspiration from these works and inject synthetic motions into the masked video learning. + +Most relevant to our work are methods like [12, 21, 53, 67], which enhance motion understanding in VideoMAE [57] by modifying certain design choices. MGMAE [21] and MGM [12] improve masking strategies to mask the regions of motion using optical flow and motion vectors. MotionMAE [67] adds frame difference as an additional target for reconstruction, and MME [53] uses optical flow to extract dense object trajectories and then predicts these trajectories and their HOG features. However, these + +methods are limited by the temporal redundancies in the original data, restricting the amount of motion or motion-focused regions available for learning in the reconstruction task. In contrast, we add synthetic motions to the original video data, reducing temporal redundancy and increasing motion-rich regions to learn from. As shown in our experiments, reconstructing these modified videos encourages the model to capture motion dynamics, enhancing the motion awareness of the learned video representations. + +# 3. Methodology + +# 3.1. Revisiting Masked Video Modeling + +The goal of masked video modeling [12, 21, 57] is to learn video representations by employing a mask-and-predict task with an asymmetric encoder-decoder architecture. An input video $V \in \mathbb{R}^{T \times H \times W \times 3}$ is first divided into non-overlapped 3D space-time tokens (e.g., $2 \times 16 \times 16$ ) as $\mathcal{T} = \{v_1, \dots, v_N\}$ . From $\mathcal{T}$ , a subset of tokens $\mathcal{T}^{mask}$ are randomly masked with a high masking ratio (e.g., $90\%$ ) and the unmasked tokens $\mathcal{T}^{unmask} = \mathcal{T} \setminus \mathcal{T}^{mask}$ are fed into a transformer encoder $\Phi_{enc}$ . Then, the output of the encoder and the learnable tokens containing the positional information of the masked tokens are passed to a shallow decoder $\Phi_{dec}$ to reconstruct the input pixel with the following loss: + +$$ +\mathcal {L} = \frac {1}{| \mathcal {T} ^ {\text {m a s k}} |} \sum_ {i \in \mathcal {T} ^ {\text {m a s k}}} \| v _ {i} - Y _ {i} \| _ {2} ^ {2}, \tag {1} +$$ + +where $v_{i}$ refers to the pixel values of the $i$ -th space-time token in $\mathcal{T}^{mask}$ and $Y_{i} = \Phi_{dec}(\Phi_{enc}(\mathcal{T}^{unmask}))$ denotes the corresponding reconstructed pixels for $i$ -th token given the set of all unmasked tokens $\mathcal{T}^{unmask}$ . + +Such pixel reconstruction drives the model to focus on low-level features, and with high temporal redundancy in video data, it primarily reconstructs static appearance details of video data. However, effective video understanding also requires capturing temporal progression and high-level semantics, such as scene context and object interactions. + +# 3.2. General Scheme of SMILE + +Our goal is to infuse spatial and motion semantics in masked video modeling. To achieve this, we integrate two key elements into the masked video modeling paradigm. To improve motion semantics, we overlay local motion patterns in input videos in the form of moving objects. To improve spatial semantics, we replace the low-level pixel targets with high-level features extracted from a pretrained CLIP encoder. By combining both, we learn video representations that can encode both spatial and motion dynamics necessary for modeling downstream video tasks. The overview of our method is shown in Figure 2. + +Motion Infusion. Given a input video $V \in \mathbb{R}^{T \times H \times W \times 3}$ . We first sample a random object $o$ from a set of predefined + +segmented objects $O$ . Next, we generate a motion trajectory (Section 3.3), along which we overlay this object onto the video $V$ to obtain the motion-augmented video $V'$ . By adding objects with synthetic motions, we reduce temporal redundancy in the original videos and encourage the model to learn from the motion dynamics of the overlaid objects. + +Masking. We employ two separate masking strategies on $V'$ as follows. For the original regions, we employ random tube masking as in VideoMAE [57]. For added object regions, we propose trajectory-based masking. Since we know the location of the object in each frame of $V'$ , we explicitly drop space-time tokens at random along its trajectory. Employing a high masking ratio along the object trajectory ensures that solving the reconstruction task requires encoding the object's motion dynamics, thereby enhancing the motion awareness of learned representations. + +Architecture. We employ a teacher-student reconstruction architecture with a pretrained image-encoder as the teacher and an asymmetric encoder-decoder network $\Phi$ as the student. The augmented video is patchified into 3D space-time tokens $\mathcal{T} = \{v_1',\dots,v_N'\}$ as in general video masked autoencoders (Section 3.1). The tokens are then masked using tube and trajectory-based masking to generate sets of masked $\mathcal{T}^{mask}$ and unmasked $\mathcal{T}^{unmask}$ tokens. The un-masked tokens $\mathcal{T}^{unmask}$ and the learnable tokens containing the positional information of the masked patches are used as the input to $\Phi$ as in Section 3.1. + +Reconstruction. Instead of directly reconstructing pixels, we project the input video $V'$ onto the teacher's feature space to generate high-level targets. In particular, we choose a CLIP model as the projection network due to its high-level semantics learned from language guidance. Specifically, the input $V'$ is passed frame-by-frame through the CLIP encoder to extract features. The extracted features are stacked into a sequence of feature tokens as $\mathcal{F} = \{f_1', \dots, f_N'\}$ . To match the number of tokens with $\mathcal{T}$ , we only use the CLIP features for the first time slice of the 3D space-time tokens, thus generating a total of $N$ feature tokens, each corresponding to a space-time token in the input video. The encoder-decoder is then trained with a masked feature reconstruction loss as: + +$$ +\mathcal {L} = \frac {1}{| \mathcal {T} ^ {\text {m a s k}} |} \sum_ {i \in \mathcal {T} ^ {\text {m a s k}}} \| f _ {i} ^ {\prime} - Y _ {i} \| _ {2} ^ {2}, \tag {2} +$$ + +where $f_{i}^{\prime}$ is the CLIP feature projection of $i$ -th space-time token $v_{i}^{\prime}$ in $T^{mask}$ and $Y_{i} = \Phi_{dec}(\Phi_{enc}(\mathcal{T}^{unmask}))$ denotes the corresponding reconstructed feature for $i$ -th token given the set of all unmasked tokens $T^{unmask}$ . Such formulation first increases the temporal dynamics in otherwise static data, avoiding learning shortcuts for reconstruction. Second, projecting targets onto the CLIP space encourages the network to capture more high-level video semantics. + +![](images/00a8924413236da0589b363ecc48927c37eff76ed09b81815d0c6aa2e1cf011e.jpg) +Figure 2. Overall architecture of our SMILE. An input video clip $V$ is overlaid with a segmented object along a randomly generated trajectory to generate $V'$ infusing synthetic object motion in $V$ . $V'$ is passed frame-by-frame through the CLIP encoder to extract feature tokens $\mathcal{F}$ . $V'$ is patched into a set of space-time tokens $\mathcal{T}$ . We apply two types of masking upon $\mathcal{T}$ , tube masking on the space-time tokens of the original video and trajectory-based masking on the tokens of added objects. The unmasked tokens $\mathcal{T}_{unmask}$ are fed into the encoder-decoder network $\Phi_{enc} \Phi_{dec}$ which is trained to reconstruct the masked feature tokens in $\mathcal{F}$ . + +Overall, the key intuition is that CLIP features capture high-level information, such as scene context and object interactions, which guides the network to attend to both the overlaid object motions and the original video semantics. + +# 3.3. Motion Generation + +To infuse motion, we overlay a segmented object along a randomly generated trajectory onto the original video frames. To introduce more complex dynamics, the object is also transformed e.g. rotation, along the trajectory. + +Appearance and Trajectory. To infuse a new synthetic motion onto a video sequence $V = [v_{1}, v_{2}, \dots, v_{T}]$ comprising $T$ frames, we begin by selecting an object $o$ from $O$ at random, where $O$ denotes the set of off-the-shelf segmented objects generated using Stable Diffusion [48] and X-Paste[71]. We then sample a scale $P \times Q$ from a predefined set of sizes for $o$ to model different object sizes, where $P \ll H$ and $Q \ll W$ . Next, we sample a sequence of locations for the object $\mathrm{Traj}_o = [(x^1, y^1),.., (x^T, y^T)]$ representing the center coordinate $(x^i, y^i)$ of the object $o$ in the corresponding frame. Specifically, we generate a random non-linear path for the trajectory by sampling $M$ 2D points $(M \gg T)$ uniformly from $x \in [0, H]$ and $y \in [0, W]$ . To ensure smoothness in the motion, a 1D Gaussian filter is applied along $x$ and $y$ dimensions as $[(h(x^1), h(y^1)), \dots, (h(x^M), h(y^M))]$ , where $h(z) = \frac{1}{\sqrt{2\pi}\kappa} e^{-z^2 / 2\kappa^2}$ and $\kappa$ is the smoothing factor. The resulting trajectory is then downsampled to $T$ target locations for the object $o$ . This formulation introduces diverse motions + +![](images/92e27627bbbf735ca15be37dc8c53828bc42dfe7980ab7b15d4bb049eb9f2a6f.jpg) +Figure 3. Motion overlaid videos showing the position change and transformation of the added objects along the time dimension. + +with varying spatio-temporal dynamics in the form of objects with different shapes, sizes, and temporal progression. Transformation. To increase the complexity of the object's motion, we also apply scale and rotation transformations along the temporal dimension. Specifically, we select 3 frame indices, the first, the last, and a random middle one, $v_{1}, v_{m}$ , and $v_{T}$ , respectively. For these three frame indices, a rotation angle $\alpha_{i}$ , and a scaling factor $S_{i}$ are selected randomly. The angles and scales for the rest of the frames are linearly interpolated between these keyframes. Finally, the object $o$ is rotated and scaled with the corresponding transformations $\left( \begin{array}{cc}\cos \alpha_{i} & -\sin \alpha_{i}\\ \sin \alpha_{i} & \cos \alpha_{i} \end{array} \right)$ and $\left( \begin{array}{cc}S_i & 0\\ 0 & S_i \end{array} \right)$ , before overlaying it onto the video frame $v_{i}$ . + +Figure 3 shows two sample videos with our added motions. While these motions may not appear entirely realistic, they represent complex motion dynamics of moving structured pixels, and introducing them into a mask-and-predict framework guides the model to learn from such dynamics. + +This results in video representations with improved downstream performance on motion-aware tasks, as we also validate empirically. Interestingly, we also show in the experiments (Section 5) that overlaying such object motions on black or noise frames can learn strong video representations without requiring any original videos. + +# 3.4. Training Procedure + +By adding synthetic motions, we have introduced a domain shift in the input for reconstruction, which may affect the performance of downstream tasks. To explore effective learning from both original videos $V$ and motion-overlaid videos $V'$ , we explore two training strategies, i.e. mixed training and progressive training. In the mixed training strategy, we use both $V$ and $V'$ as the input to the network and directly optimize for the reconstruction of both inputs. The idea is to simultaneously encode the spatio-temporal dynamics of the original video data and the motion dynamics of the augmented data, with a limited domain shift. In the progressive training strategy, we employ a two-stage learning approach that first learns from augmented data $V'$ alone, then continues to train the model with the original data $V$ alone. This way we can first learn to encode the spatio-temporal dynamics of the augmented data and then align the network to further encode the dynamics of the original data and its domain. + +# 4. Experiments + +We evaluate SMILE against prior video SSL methods on action recognition benchmarks: Kinetics-400 (K400) [26], SomethingSomething V2 [18] (SSv2), UCF-101 [52] (UCF), HMDB-51 [29] (HMDB), FineGYM [50] (GYM), and EPIC-Kitchens-100 [8] (EPIC), achieving the best performance in linear probing and full finetuning (Section 4.2). SMILE also shows better generalization on the SEVERE benchmark[55] (Section 4.4). Ablations and further discussion are provided in Section 4.3 and Section 5, respectively. + +# 4.1. Implementation Details + +Model. Following VideoMAE [57], we use ViT-S or ViT-B with joint space-time attention for our student network. We choose CLIP ViT-B [45] as our teacher network. + +Motions. We follow PIN [10] to create the set of segmented objects $O$ with Stable Diffusion [48] and X-Paste [71]. We sample the object size $P \times Q$ uniformly from $[32 \times 32, 128 \times 128]$ and sample the rotation angles from $[-90, 90]$ and scaling factors from [0.5, 1.5]. + +Pretraining. We follow previous masked video modeling works [12, 21, 53, 57] and use Kinetics-400 [26] (K400) or SomethingSomething V2 [18] (SSv2) for pretraining unless specified otherwise. We use the same hyperparameters as in VideoMAE [57]. Unless stated otherwise, we train with our proposed progressive training for a total of 600 epochs, 300 epochs in each stage. More training details are in supplementary material. + +Table 1. Linear probing comparison of masked video modeling methods. Our SMILE learns highly discriminative video representations useful for both spatial and motion-focused domains. All results are obtained with the ViT-B backbone pretrained on K400 and Top-1 accuracies $(\%)$ are reported. + +
MethodK400UCFHMDBSSv2GYMEPIC
VideoMAE [57]20.758.637.717.523.933.2
MVD [63]18.749.128.612.222.729.7
MME [53]19.156.037.116.629.032.2
MGMAE [21]24.964.441.316.826.133.2
EVEREST [23]14.151.830.314.523.330.5
MGM [12]19.862.540.321.725.832.4
SIGMA [49]47.580.752.320.830.134.2
SMILE w/o motion54.682.651.721.627.932.3
SMILE (ours)56.283.853.423.730.234.4
+ +# 4.2. Comparison with State-of-the-Art + +Linear Probing. We freeze the pretrained backbone and only train a linear classifier on the target dataset. When evaluating the prior works for comparison, we utilize their publicly available checkpoints for the ViT-B backbone pretrained on the K400 dataset for 800 epochs except for MVD (2000 epochs) and MME (1600 epochs). We evaluate the downstream task using a common setup for a fair comparison. Details about datasets and experimental settings are in the supplementary material. + +Table 1 shows that our method, even without synthetic motion, significantly outperforms all state-of-the-art methods on K400 and UCF, and is among the top performers on HMDB, SSv2, GYM, and EPIC. This highlights that CLIP feature reconstruction offers a stronger supervisory signal than pixel reconstruction, especially in tasks requiring spatial semantics like K400 and UCF. Adding synthetic motions further boosts performance by an average of $1.8\%$ across datasets, enhancing motion cues through reconstruction. On motion-sensitive datasets like SSv2, GYM, and EPIC, our approach consistently surpasses MGM [12], MGMAE [21], and MME [53], confirming that synthetic motions help overcome the lack of motion diversity in current datasets that may limit these methods. + +Full Finetuning. We finetune both the pretrained backbone and the classification head end-to-end following [21, 53, 57] on the downstream datasets K400 and SSv2. Following prior masked video modeling works [12, 21, 57], we show results with self-supervised pretraining on K400 and SSv2 datasets for the VIT-B backbone. We follow VideoMAE [57] for finetuning and evaluation protocols with more details in the supplementary. + +Tables 2 and 3 show the results for SSv2 and K400, respectively. On SSv2, we achieve the best performance with both K400 and SSv2 pretraining, highlighting our method's effectiveness for both in-domain and cross-domain transfer. + +Table 2. Full finetuning comparison of various self-supervised methods on Something-Something V2. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported. + +
MethodBackboneEpochsPretrainTop-1
VideoMAE [57]ViT-B800K40068.5
OmniMAE [17]ViT-B800K40069.0
MGMAE* [21]ViT-B800K40068.9
MME [53]ViT-B800K40070.5
SIGMA [49]ViT-B800K40071.1
MGM* [12]ViT-B800K40071.1
SMILE (ours)ViT-B600K40072.1
OmniMAE [17]ViT-B800SSv269.5
VideoMAE [57]ViT-B800SSv269.6
MGM [12]ViT-B800SSv270.6
MME [53]ViT-B800SSv270.0
SIGMA [49]ViT-B800SSv271.2
MGMAE [21]ViT-B800SSv271.0
SMILE (ours)ViT-B800SSv272.5
+ +For K400, we exceed the prior top SSL approach by $1.6\%$ . Additionally, our gains over motion-focused methods, such as MGMAE (1.5% on SSv2, 1.9% on K400), MME (2.5% on SSv2, 1.6% on K400), and MGM (1.9% on SSv2, 2.3% on K400), confirm our improved motion sensitivity, even in full-finetuning settings. Notably, our model surpasses VideoMAE, MotionMAE, MME, and MGMAE trained for more than twice the number of epochs (see supplementary for the extensive comparison). We attribute this efficiency to CLIP feature reconstruction, which directly captures high-level semantics compared to pixel-level learning. + +# 4.3. Ablations + +We ablate various components of our SMILE to show their impact on the downstream performance in Table 4. For computational efficiency, we use ViT-S as the backbone and pretrain with a smaller subset of K400, denoted as $\mathrm{K400_m}$ , containing around 80K samples. For evaluation, we use $\mathrm{K400_m}$ and $\mathrm{SSv2_m}$ that denotes a subset of SSv2 with $50\%$ training data. Unless stated otherwise, we use CLIP features as the reconstruction target, add two synthetic objects, mask $90\%$ , use trajectory masking, and train for 400 epochs. + +Synthetic motion, yes or no? In Table 4a, we show the impact of adding synthetic motions to the masked video modeling framework for both pixel and feature reconstruction. We can see that adding our synthetic motion patterns significantly improves the downstream performance of both pixel and feature reconstruction. + +Reconstructing pixels or features? Table 4h compares pixel-based and feature-based reconstruction targets in masked video modeling. Reconstructing in feature space significantly boosts performance over pixel reconstruction, with a $+12.1\%$ gain on spatially-focused tasks like $\mathrm{K400_m}$ and $+5.1\%$ on motion-focused tasks like $\mathrm{SSv2_m}$ . This sug- + +Table 3. Full finetuning comparison of various self-supervised methods on Kinetics-400. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported. + +
MethodBackboneEpochsPretrainTop-1
SVT [46]ViT-B-K40078.4
VideoMAE [57]ViT-B800K40080.0
BEVT [62]ViT-B-K40080.6
OmniMAE [17]ViT-B800K40080.8
MGM [12]ViT-B800K40080.8
MGMAE [21]ViT-B800K40081.2
SIGMA [49]ViT-B800K40081.5
MME* [53]ViT-B800K40081.5
SMILE (ours)ViT-B600K40083.1
+ +gests that using a high-level feature space as the reconstruction target provides more meaningful supervision for learning transferable spatio-temporal representations. + +Reconstructing which features: CLIP, MAE or DINO? Table 4b compares several image foundation models—MAE [20], DINO [4], DINO-v2 [39], and CLIP [45]—as reconstruction targets. CLIP features outperform the others, likely due to CLIP's image-text alignment training on a large dataset, enhancing its transferability and adaptability to video tasks as shown in [36, 47, 61]. Compared to pixel reconstruction in Table 4h, all feature-based approaches show substantial performance gains. + +Trajectory-based masking: yes or no? Table 4g shows the impact of our trajectory-based masking. Without trajectory masking, all tokens are masked using tube masking, which masks the whole video without distinguishing between object and non-object space-time tokens. In contrast, trajectory masking first masks space-time tokens associated with objects, followed by tube masking on the remaining tokens, maintaining the overall masking ratio. Our trajectory masking improves performance by approximately $1\%$ on both $\mathrm{K400_m}$ and $\mathrm{SSv2_m}$ , validating that masking along object trajectories enhances motion-aware representation learning. + +How much to mask? Table 4e shows the impact of masking ratio $m$ on downstream performance. We mask $m$ of space-time tubes within the objects and $m$ of space-time tubes in the whole video with $m$ being $95\%$ , $90\%$ , or $80\%$ . We observe the best performance with $m = 80\%$ . + +How many synthetic objects? In Table 4f, we analyze the impact of adding more object motions to the input video on downstream performance. We evaluate adding 1, 2 and 3 object motions and compare the results. We observe that adding 2 objects is better than using a single object and the performance saturates when adding more objects. + +Which object transformations? In Table 4c, we show the impact of object transformation on downstream performance. Both rotation and scaling improve the performance individually and combining them achieves the best results. + +
Synth.TargetK400mSSv2m
w/oPixels66.055.1
w/Pixels68.156.8
w/oFeatures78.160.2
w/Features78.961.6
+ +(a) Synthetic motion, yes or no? Yes. It improves performance for both pixel and feature targets. + +
RatioTargetK400mSSv2m
95%features77.760.4
90%features78.961.6
80%features80.362.0
+ +(e) How much to mask? Masking $80\%$ trajectory and total tokens. + +
TargetK400mSSv2m
MAE [20]77.659.5
DINO [4]77.659.7
DINOv2 [39]77.859.7
CLIP [45]78.160.2
+ +(b) Reconstructing which features? CLIP features provide the best target. No motion is added. + +
# ObjectsK400mSSv2m
178.661.0
278.961.6
378.661.3
+ +(f) How many objects? Adding two object motions is the best. + +
TransformK400mSSv2m
None77.060.8
Scaling78.461.3
Rotation78.261.4
Both78.961.6
+ +(c) Which object transformations? Using both rotation and scaling works the best. + +
MaskingK400mSSv2m
w/o Trajectory77.860.7
w/ Trajectory78.961.6
+ +(g) Trajectory masking, yes or no? Yes. Both targets are features. + +
TrainingK400mSSv2m
Single78.961.6
Mix79.562.2
Progressive79.862.5
+ +(d) Which training strategy? Progressive training outperforms both for the same number of iterations. + +
TargetK400mSSv2m
Pixels66.055.1
Features78.160.2
+ +(h) Reconstructing pixels or features? Reconstructing features is much better. No motion is added. + +Table 4. Ablation experiments of our proposed SMILE. The default setting uses clip reconstruction targets with two synthetic objects and a $90\%$ masking ratio with trajectory-based masking. We also show full-scale ablations with ViT-B and K400 pretraining in supplementary. + +Which training strategy? Table 4d compares our proposed training strategies. We also compare learning with the single learning strategy, where only augmented videos are used for training. We observe that both mixed and progressive learning are better than learning from augmented videos only. Progressive is better than mixed, most likely due to a reduced domain shift in the second stage where only original videos are used. + +Qualitative analysis. To compare the temporal dynamics of video representations learned by different SSL methods, we compute their feature similarity across different frames of input videos (Figure 4). For our model, the features of different frames have larger differences, indicating better temporal awareness. More plots are in the supplementary. + +![](images/7c92a6abf5b929a132587001f36c07fc6dd7ab20db7b2da7811b90df707fdb57.jpg) +Figure 4. Feature similarity across different frames for different SSL methods. We compute this on K400 validation videos. + +# 4.4. Generalization Assessment + +To demonstrate the generalization of our learned video representations to diverse downstream video setups, we evaluate SMILE on the SEVERE benchmark [55], which includes 8 experiments targeting 4 key generalization factors: domain shift, sample efficiency, action granularity, and task shift. Detailed configurations are provided in the supplementary. Beyond SEVERE, we also show generalization to more temporal tasks in supplementary. + +Table 6 shows the results. For the Domain Shift, we test + +cross-domain adaptability on SSv2 and Gym99, which differ from the pretraining dataset (K400). Our method significantly outperforms all others, demonstrating superior adaptability to domain shifts. For Sample Efficiency, we evaluate low-shot action recognition on UCF and GYM with only 1,000 samples for finetuning. Our approach surpasses previous methods, achieving a $5\%$ improvement on GYM over the best prior model, MME, showcasing its strong low-shot recognition capability even in motion-sensitive tasks. For the Action Granularity, tested on FX-S1 and UB-S1 from FineGym, our model achieves the best results, excelling at distinguishing fine-grained motions within gymnastics routines (e.g., jump variations in FX-S1). It outperforms motion-aware models like MME, MGMAE, and MGM, highlighting superior motion generalizability. For Task Shift, evaluated with temporal repetition counting on UCF-RC [70] and multi-label recognition on Charades [51], our method matches prior performance on repetition counting and achieves a $9\%$ improvement over MME on Charades, demonstrating broad task adaptability. Overall, SMILE (without synthetic motion) is on par with the prior best with the mean performance of $64.9\%$ . With synthetic motion SMILE, we observe a further $3.0\%$ gain, reinforcing our model's generalization strength through synthetic motion. + +# 5. Discussions + +Comparison with Other CLIP Adaptations. Table 5 presents a comparison between our SMILE and the CLIP adaptation methods ViCLIP [64] and UMT [31], along with the CLIP baseline. Our method achieves superior performance on motion-sensitive datasets, particularly in linear probing, indicating that our CLIP adaptation learns better video representations than the video-text alignment of ViCLIP and the unmasked feature alignment of UMT, despite using less data. More results are in the supplementary. + +
PretrainingFinetuningLinear Probing
MethodVideo DatasetUCFGYMSSv2EPICUCFGYMSSv2EPIC
CLIP [45]-93.688.066.750.377.520.711.325.1
ViCLIP [64]Intervid-10M [64]95.289.767.955.086.727.318.927.3
UMT [31]K700 [5]96.089.970.150.188.026.418.828.2
SMILE (ours)K400 [26]96.490.871.963.383.830.223.734.6
+ +Table 5. Comparison with CLIP adaptations for video representation learning. Despite using less data, we learn better video representations than other CLIP adaptations, especially for motion-focused downstream domains. +Table 6. Generalization assessment on SEVERE benchmark [55] for recent video SSL methods in terms of domain shift, sample efficiency, action granularity, and task shift. We evaluate prior works with the official ViT-B backbone checkpoints pretrained on the K400 dataset. SMILE shows the best generalization performance across setups. Adding synthetic motions further improves the performance. + +
MethodDomain shiftSample efficiency (103)Action granularityTask shiftMean
SSv2Gym99UCFGYMFX-S1UB-S1UCF-RC↓Charades
VideoMAE [57]68.686.674.625.942.865.30.17214.457.6
MVD [63]70.082.567.117.531.350.50.18416.152.1
MGMAE [21]68.987.277.224.133.779.50.18117.958.8
MGM [12]71.189.178.426.438.686.90.15222.562.2
SIGMA [49]70.989.784.128.055.179.90.16923.164.2
MME [53]70.189.779.229.855.587.20.15523.665.0
SMILE w/o motion71.690.085.232.053.474.30.17530.564.9
SMILE (ours)72.190.886.435.155.188.30.17032.567.9
+ +Table 7. Learning without natural videos. We train a ViT-S for pixel reconstruction. Our synthetic object motions can learn video representations without using any natural videos. + +
Unnatural Video SchemePretrain DataK400mSSv2m
No Pretraining-39.728.2
Video-frameK400m-1frm54.043.1
Video-frame + MotionK400m-1frm63.653.2
ImagePlaces-80K52.140.5
Image + MotionPlaces-80K60.152.1
Black image + Motion-59.250.0
Noise image + Motion-61.951.2
+ +Learning without Natural Videos. In prior experiments, we overlay our object motions onto natural videos from the Kinetics-400 dataset, raising the question of whether downstream improvements stem from original video priors (motion, action, scene) or our added object motions. To investigate this, we construct videos by overlaying object motions onto four different backgrounds: (1) a single frame from natural videos, (2) a single scene image from the image dataset Places [73], (3) a plain black image, and (4) a noise image. Each background image is duplicated across $T$ frames, creating static clips onto which moving objects are placed. For (1), we select one frame per video from $\mathrm{K400_m}$ (denoted $\mathrm{K400_m - 1frm}$ ), preserving action and scene cues without motion priors. For (2), we sample 80K images from Places (Places-80K), matching the dataset size of $\mathrm{K400_m - 1frm}$ , to remove action and motion priors. Approaches (3) and (4) exclude all natural video or image priors entirely. Evaluation involves full finetuning on $\mathrm{K400_m}$ + +and $\mathrm{SSv2_m}$ , consistent with our previous ablation studies. + +Table 7 reports results for different augmented video types. Adding object motions consistently enhances downstream performance, whether combined with action and scene cues from natural videos (row 2 vs. row 3) or only scene cues from images (row 4 vs. row 5). Even with black or noisy backgrounds (rows 6 and 7), performance remains robust. Remarkably, using object motions alone yields substantial improvements— $22.2\%$ on $\mathrm{K400_m}$ and $23.0\%$ on $\mathrm{SSv2_m}$ compared to no pretraining. These findings demonstrate our method's effectiveness in learning video representations from synthetic motions alone. More results are in the supplementary. In summary, our method enables a new paradigm of learning strong video representations with masked video modeling without relying on natural videos. + +# 6. Conclusions + +In this work, we proposed SMILE, a novel self-supervised approach for video representation learning that addresses key limitations in existing masked video modeling methods by integrating both spatial and motion semantics. Our SMILE, leverages the high-level spatial semantics from image-language pretrained models like CLIP and enhances motion representation through synthetic motion patterns. By capturing complex spatial features and motion dynamics, SMILE achieves superior performance on diverse downstream tasks, demonstrating improved discrimination and generalization over state-of-the-art video SSL methods. Our extensive evaluations demonstrated the advantages of our SMILE, setting a new paradigm for robust and effective video self-supervised learning. + +Acknowledgements. This work is supported by the KAUST Center of Excellence for Generative AI under award number 5940. For computing time, this research used Ibex managed by the Supercomputing Core Laboratory at King Abdullah University of Science & Technology (KAUST) in Thuwal, Saudi Arabia. We also thank Michael Dorkenwald for providing the synthetic objects used in this work. + +# References + +[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. +[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7 +[5] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 8 +[6] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[7] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised visual learning by variable playback speeds prediction of a video. IEEE Access, 9:79562-79571, 2021. 2 +[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. International Journal of Computer Vision (IJCV), 2021. 2, 5 +[9] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[10] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5 +[11] Hazel Doughty, Fida Mohammad Thoker, and Cees GM Snoek. Locomotion: Learning motion-focused video-language representations. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 50-70, 2024. 2 +[12] David Fan, Jue Wang, Shuai Liao, Yi Zhu, Vimal Bhat, Hector Santos-Villalobos, Rohith MV, and Xinyu Li. Motion-guided masking for spatiotemporal representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8 +[13] David Fan, Jue Wang, Shuai Liao, Zhikang Zhang, Vimal Bhat, and Xinyu Li. Text-guided video masked autoencoder. In European Conference on Computer Vision (ECCV), 2024. 2 +[14] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[15] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[16] Kirill Gavrilyuk, Mihir Jain, Ilia Karmanov, and Cees G M Snoek. Motion-augmented self-training for video recognition at smaller scale. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[17] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimaes: Single model masked pretraining on images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The "something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5 +[19] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, 2, 6, 7 +[21] Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, and Limin Wang. Mgmae: Motion guided masking for video masked autoencoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8 +[22] Ziyuan Huang, Shiwei Zhang, Jianwen Jiang, Mingqian Tang, Rong Jin, and Marcelo H Ang. Self-supervised motion + +learning from static images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[23] Sunil Hwang, Jaehong Yoon, Youngwan Lee, and Sung Ju Hwang. Everest: Efficient masked video autoencoder by removing redundant spatiotemporal tokens. In International Conference on Machine Learning (ICML), 2024. 5 +[24] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), 2020. 2 +[25] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2 +[26] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1, 2, 5, 8 +[27] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2 +[28] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified CNN architecture for real-time spatiotemporal action localization. arXiv preprint arXiv:1911.06644, 2019. 1 +[29] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2011. 2, 5 +[30] Hsin-Ying Lee, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[31] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7, 8 +[32] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[33] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2 +[34] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), 2016. 2 +[35] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[36] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 2, 6 +[37] Jingcheng Ni, Nan Zhou, Jie Qin, Qian Wu, Junqi Liu, Boxun Li, and Di Huang. Motion sensitive contrastive learning for self-supervised video representation. In European Conference on Computer Vision (ECCV), 2022. 2 +[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 7 +[40] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. arXiv preprint arXiv:2303.09857, 2023. 2 +[42] Mandela Patrick, Yuki M. Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[43] AJ Piergiovanni, Anelia Angelova, and Michael S Ryoo. Evolving losses for unsupervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[44] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2, 5, 6, 7, 8 +[46] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S. Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 +[47] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned + +clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5 +[49] Mohammadreza Salehi, Michael Dorkenwald, Fida Mohammad Thoker, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Sigma: Sinkhorn-guided masked video modeling. In European Conference on Computer Vision (ECCV), 2025. 5, 6, 8 +[50] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5 +[51] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in Homes: Crowdsourcing Data Collection for Activity Understanding. In European Conference on Computer Vision (ECCV), 2016. 1, 2, 7 +[52] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5 +[53] Xinyu Sun, Peihao Chen, Liangwei Chen, Changhao Li, Thomas H. Li, Mingkui Tan, and Chuang Gan. Masked motion encoding for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 8 +[54] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the ACM International Conference on Multimedia (ACMMM), 2021. 2 +[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 1, 2, 5, 7, 8 +[56] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Tubelet-contrastive self-supervision for videoefficient generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 5, 6, 8 +[58] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. arXiv preprint arXiv:2304.10465, 2023. 2 +[59] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[60] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[61] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 6 +[62] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6 +[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 8 +[64] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2, 7, 8 +[65] Fanyi Xiao, Joseph Tighe, and Davide Modolo. Maclr: Motion-aware contrastive learning of representations for videos. In European Conference on Computer Vision (ECCV), 2022. 2 +[66] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[67] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 2 +[68] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024, 2023. 2 +[69] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[70] Huaidong Zhang, Xuemiao Xu, Guoqiang Han, and Shengfeng He. Context-aware and scale-insensitive temporal repetition counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7 +[71] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo + +Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In International Conference on Learning Representations (ICLR), 2023. 4, 5 +[72] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 2 +[73] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 8 \ No newline at end of file diff --git a/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/images.zip b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1976b82b2f93ce3265b4123b3e6bc0aaebe7aeab --- /dev/null +++ b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18e1463f0223ce65d90c97492300e4f6d78671352bbf2bb302befb938114d5aa +size 539613 diff --git a/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/layout.json b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..88821f4a7e503dee6ddeab973d24df9c4f567e2e --- /dev/null +++ b/2025/SMILE_ Infusing Spatial and Motion Semantics in Masked Video Learning/layout.json @@ -0,0 +1,10141 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 537, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 537, + 122 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 537, + 122 + ], + "type": "text", + "content": "SMILE: Infusing Spatial and Motion Semantics in Masked Video Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 143, + 475, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 143, + 475, + 172 + ], + "spans": [ + { + "bbox": [ + 134, + 143, + 475, + 172 + ], + "type": "text", + "content": "Fida Mohammad Thoker, Letian Jiang, Chen Zhao†, Bernard Ghanem King Abdullah University of Science and Technology (KAUST)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 174, + 477, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 174, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 477, + 186 + ], + "type": "text", + "content": "{fida.thoker,letian.jiang,chen.zhao,Bernard.Ghanem}@kaust.edu.sa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 514 + ], + "type": "text", + "content": "Masked video modeling, such as VideoMAE, is an effective paradigm for video self-supervised learning (SSL). However, they are primarily based on reconstructing pixel-level details on natural videos which have substantial temporal redundancy, limiting their capability for semantic representation and sufficient encoding of motion dynamics. To address these issues, this paper introduces a novel SSL approach for video representation learning, dubbed as SMILE, by infusing both spatial and motion semantics. In SMILE, we leverage image-language pretrained models, such as CLIP, to guide the learning process with their high-level spatial semantics. We enhance the representation of motion by introducing synthetic motion patterns in the training data, allowing the model to capture more complex and dynamic content. Furthermore, using SMILE, we establish a new self-supervised video learning paradigm capable of learning strong video representations without requiring any natural video data. We have carried out extensive experiments on 7 datasets with various downstream scenarios. SMILE surpasses current state-of-the-art SSL methods, showcasing its effectiveness in learning more discriminative and generalizable video representations. Code is available: https://github.com/fmthoker/SMILE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 535, + 135, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 135, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 135, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "text", + "content": "Self-supervised learning (SSL) [43, 44, 57, 59] plays a crucial role for many video understanding tasks, such as action recognition [18, 26, 51, 52, 55] and detection [28, 32], significantly reducing the reliance on extensive human annotation. The goal of video SSL is to learn video representations that capture both spatial and temporal semantics effectively. Masked video modeling has emerged as a promising approach with state-of-the-art performance [12, 21, 53, 57]. It involves masking a large portion of video frames and reconstructing the masked regions using the unmasked parts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "However, this approach has some limitations. First, it relies heavily on pixel reconstruction, which can lead to overfitting low-level details. This limits the model's abl" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 336, + 215, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 336, + 215, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 336, + 215, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 336, + 215, + 545, + 403 + ], + "type": "image", + "image_path": "38cd8034fabbd90eb5913dd89adf8141947704afaf27dbbef3cca683613243ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 441 + ], + "type": "text", + "content": "Figure 1. Comparison with SOTA masked video modeling methods. Our method significantly outperforms prior masked video modeling methods across diverse downstream settings." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 461, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 555, + 617 + ], + "type": "text", + "content": "ity to learn meaningful high-level semantics essential for discriminative representations, as shown by the suboptimal linear probing performance of current masked video modeling methods [12, 21, 53, 57] in Figure 1. The second key issue in masked video modeling is the substantial temporal redundancy of video data with most videos having little foreground motion, human or object. Consequently, the reconstruction objective can be solved using the temporal correlations between co-located patches or by spatial correlations in tube masking [57], resulting in inadequate motion awareness in the learned representations. This turn leads to poor generalization on downstream tasks, requiring a finer motion understanding as shown in [55]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "In this work, we address these limitations by proposing a novel framework for masked video modeling, dubbed as SMILE (Spatial and Motion semantics in masked video LEarning), to capture high-level spatial and motion semantics leading to generalizable video representations. To improve the spatial semantics in masked video modeling, we propose to explicitly infuse high-level spatial semantics from pretrained image foundation models such as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8438" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "CLIP [45]. Such models encapsulate high-level semantics learned through large-scale image-text alignment and have been adapted for video understanding tasks. For example, ActionCLIP [61], ViFiCLIP [47], and X-CLIP [36]) adapt CLIP for action recognition using videos with action labels, whereas ViCLIP [64] and VideoPrism [72]) adapt it for video-text alignment with video-text pairs. We take inspiration from them but aim to learn video representations with unlabeled videos. In particular, we propose replacing the reconstruction of pixels in masked video modeling with that of CLIP embedding, thus providing high-level semantics as the supervision signal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 396 + ], + "type": "text", + "content": "To improve motion awareness, we propose to reduce the temporal redundancy in input training videos with synthetic motion patterns. Specifically, we randomly paste objects onto the video frames and move them smoothly along the temporal dimension to simulate variable temporal transitions of motion-intensive videos. This reduces the temporal redundancy in the original videos and results in videos with an additional prominent foreground motion. Furthermore, adding objects with known locations allows us to add a motion focus to the mask-and-predict task by explicitly masking the regions of object motion. This in turn encourages the network to avoid an over-reliance on static spatial information and encode the motion dynamics of the added objects, when solving the reconstruction task. We summarize our contributions as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 398, + 295, + 577 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 398, + 295, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 295, + 457 + ], + "type": "text", + "content": "- We introduce SMILE, a novel SSL approach for masked video modeling. It encodes spatial and motion semantics using guidance from image-text pretrained models and synthetic-motion augmented videos, leading to more discriminative and generalizable video representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 458, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 458, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 458, + 295, + 506 + ], + "type": "text", + "content": "- We present a new strategy to augment video data with diverse synthetic motions, paired with a masking strategy that focuses on reconstructing these added motions to improve masked video modeling pretraining." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 506, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 295, + 577 + ], + "type": "text", + "content": "- Through extensive evaluation on 7 datasets, including Kinetics-400 [26], UCF-101 [52], HMDB-51 [29], FineGYM [50], EPIC-Kitchens-100 [8], SomethingSomething V2 [18], and Charades [51], we demonstrate that SMILE significantly outperforms existing state-of-the-art video SSL methods and exhibits better generalization." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 587, + 146, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 146, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 146, + 600 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Video Self-Supervised Learning. Video SSL aims to learn the spatio-temporal dynamics of video data without the need for any supervision. Three main learning paradigms have emerged: transformation prediction, contrastive learning, and masked video modeling. In transformation prediction, models learn video patterns by solving tasks such as space-time puzzles [1, 25, 27], predicting order in video clips [15, 24, 30, 33, 34, 66], or estimating speed [3, 7, 69]. Contrastive learning is based on instance discrimina" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "tion [38] and learning invariances to spatio-temporal augmentations [2, 6, 14, 35, 40, 42, 44, 46, 54, 56]. Masked video modeling methods [12, 21, 53, 57, 67], like VideoMAE [57], use a mask-and-predict task to reconstruct videos from partial data with an encoder-decoder setup similar to MAE [20]. By masking a high proportion of pixels (e.g., " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ") and reconstructing them such methods learn useful video representations. Our approach aims to improve the spatial and motion semantics of the video representations learned by masked video modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 194, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 434 + ], + "type": "text", + "content": "CLIP for Video Understanding. CLIP [45] captures detailed spatial semantics from extensive image-text training, making it a strong foundation for directly modeling video data with added temporal dynamics, e.g., [36, 41, 58, 61, 64, 68]. ViFiCLIP [47] finetunes CLIP on video by adding a temporal pooling layer, ActionCLIP [61] includes a temporal positional embedding and aggregation layer, and X-CLIP [36] integrates cross-frame attention to share information between frames. Most relevant to us are works that also use CLIP for video representation learning [13, 31, 64]. ViCLIP [64] initializes a model with CLIP weights and employs further contrastive pretraining with video-text pairs. UMT [31] uses CLIP as a teacher and aligns its predictions with the CLIP features for unmasked tokens, followed by a 2nd stage of learning from video-text contrastive pairs. TGM [13] relies on CLIP text-encoder features to guide masked video modeling and video-text contrastive learning in a unified framework. In contrast, our SMILE only learns from video data in a self-supervised manner, reconstructing CLIP features of masked tokens and added objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 436, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 553, + 604 + ], + "type": "text", + "content": "Motion-aware Video Self-Supervised Learning. Encoding motion dynamics in video representations is essential for generalizing across various video tasks [55]. As a result, many video SSL methods focus on enhancing the motion sensitivity of learned representations. For example, [16, 19, 21, 37, 65] capture motion using optical flow, [9, 60] do so by removing background or static elements and [11, 22, 56, 59] incorporate synthetic motions. [22, 59] learn by predicting the speed and trajectory of the pseudo motions, and [56] contrasts video with different types of motions. These approaches are specific to transformation prediction and contrastive learning paradigms. We take inspiration from these works and inject synthetic motions into the masked video learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Most relevant to our work are methods like [12, 21, 53, 67], which enhance motion understanding in VideoMAE [57] by modifying certain design choices. MGMAE [21] and MGM [12] improve masking strategies to mask the regions of motion using optical flow and motion vectors. MotionMAE [67] adds frame difference as an additional target for reconstruction, and MME [53] uses optical flow to extract dense object trajectories and then predicts these trajectories and their HOG features. However, these" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8439" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "methods are limited by the temporal redundancies in the original data, restricting the amount of motion or motion-focused regions available for learning in the reconstruction task. In contrast, we add synthetic motions to the original video data, reducing temporal redundancy and increasing motion-rich regions to learn from. As shown in our experiments, reconstructing these modified videos encourages the model to capture motion dynamics, enhancing the motion awareness of the learned video representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 138, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 138, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 138, + 201 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 242, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 242, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 242, + 219 + ], + "type": "text", + "content": "3.1. Revisiting Masked Video Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": "The goal of masked video modeling [12, 21, 57] is to learn video representations by employing a mask-and-predict task with an asymmetric encoder-decoder architecture. An input video " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": " is first divided into non-overlapped 3D space-time tokens (e.g., " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "2 \\times 16 \\times 16" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": ") as " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{v_1, \\dots, v_N\\}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": ". From " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": ", a subset of tokens " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{mask}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": " are randomly masked with a high masking ratio (e.g., " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": ") and the unmasked tokens " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{unmask} = \\mathcal{T} \\setminus \\mathcal{T}^{mask}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": " are fed into a transformer encoder " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\Phi_{enc}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": ". Then, the output of the encoder and the learnable tokens containing the positional information of the masked tokens are passed to a shallow decoder " + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\Phi_{dec}" + }, + { + "bbox": [ + 55, + 221, + 296, + 366 + ], + "type": "text", + "content": " to reconstruct the input pixel with the following loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 374, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 374, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 93, + 374, + 295, + 404 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{| \\mathcal {T} ^ {\\text {m a s k}} |} \\sum_ {i \\in \\mathcal {T} ^ {\\text {m a s k}}} \\| v _ {i} - Y _ {i} \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "4ad8642cb9ce3ebaae439fc6c53abb85dbd8a6b9042f9067ab69e096bdcbfda0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": " refers to the pixel values of the " + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": "-th space-time token in " + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{mask}" + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "inline_equation", + "content": "Y_{i} = \\Phi_{dec}(\\Phi_{enc}(\\mathcal{T}^{unmask}))" + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": " denotes the corresponding reconstructed pixels for " + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": "-th token given the set of all unmasked tokens " + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{unmask}" + }, + { + "bbox": [ + 55, + 413, + 295, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 534 + ], + "type": "text", + "content": "Such pixel reconstruction drives the model to focus on low-level features, and with high temporal redundancy in video data, it primarily reconstructs static appearance details of video data. However, effective video understanding also requires capturing temporal progression and high-level semantics, such as scene context and object interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 541, + 205, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 205, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 205, + 553 + ], + "type": "text", + "content": "3.2. General Scheme of SMILE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "Our goal is to infuse spatial and motion semantics in masked video modeling. To achieve this, we integrate two key elements into the masked video modeling paradigm. To improve motion semantics, we overlay local motion patterns in input videos in the form of moving objects. To improve spatial semantics, we replace the low-level pixel targets with high-level features extracted from a pretrained CLIP encoder. By combining both, we learn video representations that can encode both spatial and motion dynamics necessary for modeling downstream video tasks. The overview of our method is shown in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Motion Infusion. Given a input video " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": ". We first sample a random object " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " from a set of predefined" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "segmented objects " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". Next, we generate a motion trajectory (Section 3.3), along which we overlay this object onto the video " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to obtain the motion-augmented video " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". By adding objects with synthetic motions, we reduce temporal redundancy in the original videos and encourage the model to learn from the motion dynamics of the overlaid objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "type": "text", + "content": "Masking. We employ two separate masking strategies on " + }, + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "type": "text", + "content": " as follows. For the original regions, we employ random tube masking as in VideoMAE [57]. For added object regions, we propose trajectory-based masking. Since we know the location of the object in each frame of " + }, + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 313, + 145, + 555, + 264 + ], + "type": "text", + "content": ", we explicitly drop space-time tokens at random along its trajectory. Employing a high masking ratio along the object trajectory ensures that solving the reconstruction task requires encoding the object's motion dynamics, thereby enhancing the motion awareness of learned representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": "Architecture. We employ a teacher-student reconstruction architecture with a pretrained image-encoder as the teacher and an asymmetric encoder-decoder network " + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": " as the student. The augmented video is patchified into 3D space-time tokens " + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{v_1',\\dots,v_N'\\}" + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": " as in general video masked autoencoders (Section 3.1). The tokens are then masked using tube and trajectory-based masking to generate sets of masked " + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{mask}" + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": " and unmasked " + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{unmask}" + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": " tokens. The un-masked tokens " + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{unmask}" + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": " and the learnable tokens containing the positional information of the masked patches are used as the input to " + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 266, + 554, + 397 + ], + "type": "text", + "content": " as in Section 3.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": "Reconstruction. Instead of directly reconstructing pixels, we project the input video " + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": " onto the teacher's feature space to generate high-level targets. In particular, we choose a CLIP model as the projection network due to its high-level semantics learned from language guidance. Specifically, the input " + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": " is passed frame-by-frame through the CLIP encoder to extract features. The extracted features are stacked into a sequence of feature tokens as " + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_1', \\dots, f_N'\\}" + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": ". To match the number of tokens with " + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": ", we only use the CLIP features for the first time slice of the 3D space-time tokens, thus generating a total of " + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": " feature tokens, each corresponding to a space-time token in the input video. The encoder-decoder is then trained with a masked feature reconstruction loss as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 574, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 574, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 352, + 574, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{| \\mathcal {T} ^ {\\text {m a s k}} |} \\sum_ {i \\in \\mathcal {T} ^ {\\text {m a s k}}} \\| f _ {i} ^ {\\prime} - Y _ {i} \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "a0adb407ab5476d4970f05a8e7b70239d0ca9f2388494f61fb93e3011fcef591.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": " is the CLIP feature projection of " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": "-th space-time token " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^{mask}" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "Y_{i} = \\Phi_{dec}(\\Phi_{enc}(\\mathcal{T}^{unmask}))" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": " denotes the corresponding reconstructed feature for " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": "-th token given the set of all unmasked tokens " + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "inline_equation", + "content": "T^{unmask}" + }, + { + "bbox": [ + 313, + 616, + 555, + 714 + ], + "type": "text", + "content": ". Such formulation first increases the temporal dynamics in otherwise static data, avoiding learning shortcuts for reconstruction. Second, projecting targets onto the CLIP space encourages the network to capture more high-level video semantics." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8440" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 60, + 527, + 285 + ], + "blocks": [ + { + "bbox": [ + 78, + 60, + 527, + 285 + ], + "lines": [ + { + "bbox": [ + 78, + 60, + 527, + 285 + ], + "spans": [ + { + "bbox": [ + 78, + 60, + 527, + 285 + ], + "type": "image", + "image_path": "00a8924413236da0589b363ecc48927c37eff76ed09b81815d0c6aa2e1cf011e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": "Figure 2. Overall architecture of our SMILE. An input video clip " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": " is overlaid with a segmented object along a randomly generated trajectory to generate " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": " infusing synthetic object motion in " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": " is passed frame-by-frame through the CLIP encoder to extract feature tokens " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": " is patched into a set of space-time tokens " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": ". We apply two types of masking upon " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": ", tube masking on the space-time tokens of the original video and trajectory-based masking on the tokens of added objects. The unmasked tokens " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{unmask}" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": " are fed into the encoder-decoder network " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\Phi_{enc} \\Phi_{dec}" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": " which is trained to reconstruct the masked feature tokens in " + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 288, + 555, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 367, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 415 + ], + "type": "text", + "content": "Overall, the key intuition is that CLIP features capture high-level information, such as scene context and object interactions, which guides the network to attend to both the overlaid object motions and the original video semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 168, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 168, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 168, + 430 + ], + "type": "text", + "content": "3.3. Motion Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": "To infuse motion, we overlay a segmented object along a randomly generated trajectory onto the original video frames. To introduce more complex dynamics, the object is also transformed e.g. rotation, along the trajectory." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": "Appearance and Trajectory. To infuse a new synthetic motion onto a video sequence " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "V = [v_{1}, v_{2}, \\dots, v_{T}]" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " comprising " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " frames, we begin by selecting an object " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " at random, where " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " denotes the set of off-the-shelf segmented objects generated using Stable Diffusion [48] and X-Paste[71]. We then sample a scale " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "P \\times Q" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " from a predefined set of sizes for " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " to model different object sizes, where " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "P \\ll H" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q \\ll W" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": ". Next, we sample a sequence of locations for the object " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Traj}_o = [(x^1, y^1),.., (x^T, y^T)]" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " representing the center coordinate " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "(x^i, y^i)" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " of the object " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " in the corresponding frame. Specifically, we generate a random non-linear path for the trajectory by sampling " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " 2D points " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "(M \\gg T)" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " uniformly from " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "x \\in [0, H]" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "y \\in [0, W]" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": ". To ensure smoothness in the motion, a 1D Gaussian filter is applied along " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " dimensions as " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "[(h(x^1), h(y^1)), \\dots, (h(x^M), h(y^M))]" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "h(z) = \\frac{1}{\\sqrt{2\\pi}\\kappa} e^{-z^2 / 2\\kappa^2}" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " is the smoothing factor. The resulting trajectory is then downsampled to " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": " target locations for the object " + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 484, + 295, + 713 + ], + "type": "text", + "content": ". This formulation introduces diverse motions" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 357, + 540, + 450 + ], + "blocks": [ + { + "bbox": [ + 328, + 357, + 540, + 450 + ], + "lines": [ + { + "bbox": [ + 328, + 357, + 540, + 450 + ], + "spans": [ + { + "bbox": [ + 328, + 357, + 540, + 450 + ], + "type": "image", + "image_path": "92e27627bbbf735ca15be37dc8c53828bc42dfe7980ab7b15d4bb049eb9f2a6f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "type": "text", + "content": "Figure 3. Motion overlaid videos showing the position change and transformation of the added objects along the time dimension." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": "with varying spatio-temporal dynamics in the form of objects with different shapes, sizes, and temporal progression. Transformation. To increase the complexity of the object's motion, we also apply scale and rotation transformations along the temporal dimension. Specifically, we select 3 frame indices, the first, the last, and a random middle one, " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "v_{1}, v_{m}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "v_{T}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": ", respectively. For these three frame indices, a rotation angle " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": ", and a scaling factor " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": " are selected randomly. The angles and scales for the rest of the frames are linearly interpolated between these keyframes. Finally, the object " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": " is rotated and scaled with the corresponding transformations " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\left( \\begin{array}{cc}\\cos \\alpha_{i} & -\\sin \\alpha_{i}\\\\ \\sin \\alpha_{i} & \\cos \\alpha_{i} \\end{array} \\right)" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\left( \\begin{array}{cc}S_i & 0\\\\ 0 & S_i \\end{array} \\right)" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": ", before overlaying it onto the video frame " + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 498, + 555, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Figure 3 shows two sample videos with our added motions. While these motions may not appear entirely realistic, they represent complex motion dynamics of moving structured pixels, and introducing them into a mask-and-predict framework guides the model to learn from such dynamics." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8441" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "This results in video representations with improved downstream performance on motion-aware tasks, as we also validate empirically. Interestingly, we also show in the experiments (Section 5) that overlaying such object motions on black or noise frames can learn strong video representations without requiring any original videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 170, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 170, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 170, + 157 + ], + "type": "text", + "content": "3.4. Training Procedure" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": "By adding synthetic motions, we have introduced a domain shift in the input for reconstruction, which may affect the performance of downstream tasks. To explore effective learning from both original videos " + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": " and motion-overlaid videos " + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": ", we explore two training strategies, i.e. mixed training and progressive training. In the mixed training strategy, we use both " + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": " as the input to the network and directly optimize for the reconstruction of both inputs. The idea is to simultaneously encode the spatio-temporal dynamics of the original video data and the motion dynamics of the augmented data, with a limited domain shift. In the progressive training strategy, we employ a two-stage learning approach that first learns from augmented data " + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": " alone, then continues to train the model with the original data " + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 159, + 295, + 376 + ], + "type": "text", + "content": " alone. This way we can first learn to encode the spatio-temporal dynamics of the augmented data and then align the network to further encode the dynamics of the original data and its domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 382, + 136, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 136, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 136, + 395 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 504 + ], + "type": "text", + "content": "We evaluate SMILE against prior video SSL methods on action recognition benchmarks: Kinetics-400 (K400) [26], SomethingSomething V2 [18] (SSv2), UCF-101 [52] (UCF), HMDB-51 [29] (HMDB), FineGYM [50] (GYM), and EPIC-Kitchens-100 [8] (EPIC), achieving the best performance in linear probing and full finetuning (Section 4.2). SMILE also shows better generalization on the SEVERE benchmark[55] (Section 4.4). Ablations and further discussion are provided in Section 4.3 and Section 5, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 188, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 188, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 188, + 520 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 559 + ], + "type": "text", + "content": "Model. Following VideoMAE [57], we use ViT-S or ViT-B with joint space-time attention for our student network. We choose CLIP ViT-B [45] as our teacher network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": "Motions. We follow PIN [10] to create the set of segmented objects " + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": " with Stable Diffusion [48] and X-Paste [71]. We sample the object size " + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "inline_equation", + "content": "P \\times Q" + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": " uniformly from " + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "inline_equation", + "content": "[32 \\times 32, 128 \\times 128]" + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": " and sample the rotation angles from " + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "inline_equation", + "content": "[-90, 90]" + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": " and scaling factors from [0.5, 1.5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 619, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 295, + 715 + ], + "type": "text", + "content": "Pretraining. We follow previous masked video modeling works [12, 21, 53, 57] and use Kinetics-400 [26] (K400) or SomethingSomething V2 [18] (SSv2) for pretraining unless specified otherwise. We use the same hyperparameters as in VideoMAE [57]. Unless stated otherwise, we train with our proposed progressive training for a total of 600 epochs, 300 epochs in each stage. More training details are in supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 130, + 553, + 253 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 128 + ], + "type": "text", + "content": "Table 1. Linear probing comparison of masked video modeling methods. Our SMILE learns highly discriminative video representations useful for both spatial and motion-focused domains. All results are obtained with the ViT-B backbone pretrained on K400 and Top-1 accuracies " + }, + { + "bbox": [ + 313, + 72, + 553, + 128 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 72, + 553, + 128 + ], + "type": "text", + "content": " are reported." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 130, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 253 + ], + "type": "table", + "html": "
MethodK400UCFHMDBSSv2GYMEPIC
VideoMAE [57]20.758.637.717.523.933.2
MVD [63]18.749.128.612.222.729.7
MME [53]19.156.037.116.629.032.2
MGMAE [21]24.964.441.316.826.133.2
EVEREST [23]14.151.830.314.523.330.5
MGM [12]19.862.540.321.725.832.4
SIGMA [49]47.580.752.320.830.134.2
SMILE w/o motion54.682.651.721.627.932.3
SMILE (ours)56.283.853.423.730.234.4
", + "image_path": "07d8369e9fe9c31ebcf49131dcd2468413fe8da4a3744f8979feb3b36fe4112f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 275, + 495, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 495, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 495, + 288 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 290, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 398 + ], + "type": "text", + "content": "Linear Probing. We freeze the pretrained backbone and only train a linear classifier on the target dataset. When evaluating the prior works for comparison, we utilize their publicly available checkpoints for the ViT-B backbone pretrained on the K400 dataset for 800 epochs except for MVD (2000 epochs) and MME (1600 epochs). We evaluate the downstream task using a common setup for a fair comparison. Details about datasets and experimental settings are in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": "Table 1 shows that our method, even without synthetic motion, significantly outperforms all state-of-the-art methods on K400 and UCF, and is among the top performers on HMDB, SSv2, GYM, and EPIC. This highlights that CLIP feature reconstruction offers a stronger supervisory signal than pixel reconstruction, especially in tasks requiring spatial semantics like K400 and UCF. Adding synthetic motions further boosts performance by an average of " + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 399, + 554, + 566 + ], + "type": "text", + "content": " across datasets, enhancing motion cues through reconstruction. On motion-sensitive datasets like SSv2, GYM, and EPIC, our approach consistently surpasses MGM [12], MGMAE [21], and MME [53], confirming that synthetic motions help overcome the lack of motion diversity in current datasets that may limit these methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 568, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 665 + ], + "type": "text", + "content": "Full Finetuning. We finetune both the pretrained backbone and the classification head end-to-end following [21, 53, 57] on the downstream datasets K400 and SSv2. Following prior masked video modeling works [12, 21, 57], we show results with self-supervised pretraining on K400 and SSv2 datasets for the VIT-B backbone. We follow VideoMAE [57] for finetuning and evaluation protocols with more details in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": "Tables 2 and 3 show the results for SSv2 and K400, respectively. On SSv2, we achieve the best performance with both K400 and SSv2 pretraining, highlighting our method's effectiveness for both in-domain and cross-domain transfer." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8442" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 125, + 279, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 67, + 294, + 122 + ], + "lines": [ + { + "bbox": [ + 55, + 67, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 67, + 294, + 122 + ], + "type": "text", + "content": "Table 2. Full finetuning comparison of various self-supervised methods on Something-Something V2. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 125, + 279, + 301 + ], + "lines": [ + { + "bbox": [ + 69, + 125, + 279, + 301 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 279, + 301 + ], + "type": "table", + "html": "
MethodBackboneEpochsPretrainTop-1
VideoMAE [57]ViT-B800K40068.5
OmniMAE [17]ViT-B800K40069.0
MGMAE* [21]ViT-B800K40068.9
MME [53]ViT-B800K40070.5
SIGMA [49]ViT-B800K40071.1
MGM* [12]ViT-B800K40071.1
SMILE (ours)ViT-B600K40072.1
OmniMAE [17]ViT-B800SSv269.5
VideoMAE [57]ViT-B800SSv269.6
MGM [12]ViT-B800SSv270.6
MME [53]ViT-B800SSv270.0
SIGMA [49]ViT-B800SSv271.2
MGMAE [21]ViT-B800SSv271.0
SMILE (ours)ViT-B800SSv272.5
", + "image_path": "37cb230b12f6f4b95d6e0cae4848a41f5a7284dfed20950ea85328ad6840becf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": "For K400, we exceed the prior top SSL approach by " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": ". Additionally, our gains over motion-focused methods, such as MGMAE (1.5% on SSv2, 1.9% on K400), MME (2.5% on SSv2, 1.6% on K400), and MGM (1.9% on SSv2, 2.3% on K400), confirm our improved motion sensitivity, even in full-finetuning settings. Notably, our model surpasses VideoMAE, MotionMAE, MME, and MGMAE trained for more than twice the number of epochs (see supplementary for the extensive comparison). We attribute this efficiency to CLIP feature reconstruction, which directly captures high-level semantics compared to pixel-level learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 124, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 124, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 124, + 458 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": "We ablate various components of our SMILE to show their impact on the downstream performance in Table 4. For computational efficiency, we use ViT-S as the backbone and pretrain with a smaller subset of K400, denoted as " + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": ", containing around 80K samples. For evaluation, we use " + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{SSv2_m}" + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": " that denotes a subset of SSv2 with " + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": " training data. Unless stated otherwise, we use CLIP features as the reconstruction target, add two synthetic objects, mask " + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": ", use trajectory masking, and train for 400 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "type": "text", + "content": "Synthetic motion, yes or no? In Table 4a, we show the impact of adding synthetic motions to the masked video modeling framework for both pixel and feature reconstruction. We can see that adding our synthetic motion patterns significantly improves the downstream performance of both pixel and feature reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Reconstructing pixels or features? Table 4h compares pixel-based and feature-based reconstruction targets in masked video modeling. Reconstructing in feature space significantly boosts performance over pixel reconstruction, with a " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "+12.1\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " gain on spatially-focused tasks like " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "+5.1\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " on motion-focused tasks like " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SSv2_m}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". This sug-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 329, + 125, + 537, + 244 + ], + "blocks": [ + { + "bbox": [ + 313, + 67, + 555, + 121 + ], + "lines": [ + { + "bbox": [ + 313, + 67, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 67, + 555, + 121 + ], + "type": "text", + "content": "Table 3. Full finetuning comparison of various self-supervised methods on Kinetics-400. All results are obtained with a ViT-B backbone. * denotes results obtained by our evaluation. Our SMILE achieves state-of-the-art performance. Top-1 accuracies (%) are reported." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 125, + 537, + 244 + ], + "lines": [ + { + "bbox": [ + 329, + 125, + 537, + 244 + ], + "spans": [ + { + "bbox": [ + 329, + 125, + 537, + 244 + ], + "type": "table", + "html": "
MethodBackboneEpochsPretrainTop-1
SVT [46]ViT-B-K40078.4
VideoMAE [57]ViT-B800K40080.0
BEVT [62]ViT-B-K40080.6
OmniMAE [17]ViT-B800K40080.8
MGM [12]ViT-B800K40080.8
MGMAE [21]ViT-B800K40081.2
SIGMA [49]ViT-B800K40081.5
MME* [53]ViT-B800K40081.5
SMILE (ours)ViT-B600K40083.1
", + "image_path": "46689e6d66f26e62049c72a13c1fdd8404ebcd293bc28d7a77ce2d2d095afc16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 256, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 291 + ], + "type": "text", + "content": "gests that using a high-level feature space as the reconstruction target provides more meaningful supervision for learning transferable spatio-temporal representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 292, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 399 + ], + "type": "text", + "content": "Reconstructing which features: CLIP, MAE or DINO? Table 4b compares several image foundation models—MAE [20], DINO [4], DINO-v2 [39], and CLIP [45]—as reconstruction targets. CLIP features outperform the others, likely due to CLIP's image-text alignment training on a large dataset, enhancing its transferability and adaptability to video tasks as shown in [36, 47, 61]. Compared to pixel reconstruction in Table 4h, all feature-based approaches show substantial performance gains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "text", + "content": "Trajectory-based masking: yes or no? Table 4g shows the impact of our trajectory-based masking. Without trajectory masking, all tokens are masked using tube masking, which masks the whole video without distinguishing between object and non-object space-time tokens. In contrast, trajectory masking first masks space-time tokens associated with objects, followed by tube masking on the remaining tokens, maintaining the overall masking ratio. Our trajectory masking improves performance by approximately " + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "text", + "content": " on both " + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{SSv2_m}" + }, + { + "bbox": [ + 313, + 400, + 554, + 532 + ], + "type": "text", + "content": ", validating that masking along object trajectories enhances motion-aware representation learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": "How much to mask? Table 4e shows the impact of masking ratio " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": " on downstream performance. We mask " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": " of space-time tubes within the objects and " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": " of space-time tubes in the whole video with " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": " being " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": ". We observe the best performance with " + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "inline_equation", + "content": "m = 80\\%" + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 665 + ], + "type": "text", + "content": "How many synthetic objects? In Table 4f, we analyze the impact of adding more object motions to the input video on downstream performance. We evaluate adding 1, 2 and 3 object motions and compare the results. We observe that adding 2 objects is better than using a single object and the performance saturates when adding more objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "Which object transformations? In Table 4c, we show the impact of object transformation on downstream performance. Both rotation and scaling improve the performance individually and combining them achieves the best results." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8443" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 197, + 137 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 197, + 137 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 197, + 137 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 197, + 137 + ], + "type": "table", + "html": "
Synth.TargetK400mSSv2m
w/oPixels66.055.1
w/Pixels68.156.8
w/oFeatures78.160.2
w/Features78.961.6
", + "image_path": "d1fe19b008daa16e6cb0055932c573b55a0d4bf8e9bfec9b5be4770ed0d2e5f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 169, + 191, + 224 + ], + "blocks": [ + { + "bbox": [ + 70, + 138, + 196, + 167 + ], + "lines": [ + { + "bbox": [ + 70, + 138, + 196, + 167 + ], + "spans": [ + { + "bbox": [ + 70, + 138, + 196, + 167 + ], + "type": "text", + "content": "(a) Synthetic motion, yes or no? Yes. It improves performance for both pixel and feature targets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 169, + 191, + 224 + ], + "lines": [ + { + "bbox": [ + 69, + 169, + 191, + 224 + ], + "spans": [ + { + "bbox": [ + 69, + 169, + 191, + 224 + ], + "type": "table", + "html": "
RatioTargetK400mSSv2m
95%features77.760.4
90%features78.961.6
80%features80.362.0
", + "image_path": "cd3ff0d9b04470ecf4a74ba59d427e1afbd2d81df8d4bdc53399d607983ee86a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 201, + 71, + 315, + 136 + ], + "blocks": [ + { + "bbox": [ + 66, + 226, + 192, + 245 + ], + "lines": [ + { + "bbox": [ + 66, + 226, + 192, + 245 + ], + "spans": [ + { + "bbox": [ + 66, + 226, + 192, + 245 + ], + "type": "text", + "content": "(e) How much to mask? Masking " + }, + { + "bbox": [ + 66, + 226, + 192, + 245 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 66, + 226, + 192, + 245 + ], + "type": "text", + "content": " trajectory and total tokens." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 71, + 315, + 136 + ], + "lines": [ + { + "bbox": [ + 201, + 71, + 315, + 136 + ], + "spans": [ + { + "bbox": [ + 201, + 71, + 315, + 136 + ], + "type": "table", + "html": "
TargetK400mSSv2m
MAE [20]77.659.5
DINO [4]77.659.7
DINOv2 [39]77.859.7
CLIP [45]78.160.2
", + "image_path": "7159e2750f78e4e14d4a9fb21614e6b5cfe5eb035f283cb3a13532e1a0f96e75.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 205, + 170, + 306, + 224 + ], + "blocks": [ + { + "bbox": [ + 201, + 138, + 313, + 167 + ], + "lines": [ + { + "bbox": [ + 201, + 138, + 313, + 167 + ], + "spans": [ + { + "bbox": [ + 201, + 138, + 313, + 167 + ], + "type": "text", + "content": "(b) Reconstructing which features? CLIP features provide the best target. No motion is added." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 205, + 170, + 306, + 224 + ], + "lines": [ + { + "bbox": [ + 205, + 170, + 306, + 224 + ], + "spans": [ + { + "bbox": [ + 205, + 170, + 306, + 224 + ], + "type": "table", + "html": "
# ObjectsK400mSSv2m
178.661.0
278.961.6
378.661.3
", + "image_path": "71c2d1abcde56818c617c951243537f67d648b6d022b17ce3f96146815e43881.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 72, + 425, + 136 + ], + "blocks": [ + { + "bbox": [ + 200, + 225, + 310, + 245 + ], + "lines": [ + { + "bbox": [ + 200, + 225, + 310, + 245 + ], + "spans": [ + { + "bbox": [ + 200, + 225, + 310, + 245 + ], + "type": "text", + "content": "(f) How many objects? Adding two object motions is the best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 72, + 425, + 136 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 425, + 136 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 425, + 136 + ], + "type": "table", + "html": "
TransformK400mSSv2m
None77.060.8
Scaling78.461.3
Rotation78.261.4
Both78.961.6
", + "image_path": "0b541539732669f57d7d50152970736cded92c09ce7a8262b2c13aa7d5815a04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 170, + 436, + 215 + ], + "blocks": [ + { + "bbox": [ + 320, + 138, + 422, + 167 + ], + "lines": [ + { + "bbox": [ + 320, + 138, + 422, + 167 + ], + "spans": [ + { + "bbox": [ + 320, + 138, + 422, + 167 + ], + "type": "text", + "content": "(c) Which object transformations? Using both rotation and scaling works the best." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 170, + 436, + 215 + ], + "lines": [ + { + "bbox": [ + 318, + 170, + 436, + 215 + ], + "spans": [ + { + "bbox": [ + 318, + 170, + 436, + 215 + ], + "type": "table", + "html": "
MaskingK400mSSv2m
w/o Trajectory77.860.7
w/ Trajectory78.961.6
", + "image_path": "69888e8d31cb6801a3878213ce1be711ee6aa9fb5612a6f780bd3b7c7b1f509f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 430, + 72, + 539, + 127 + ], + "blocks": [ + { + "bbox": [ + 318, + 217, + 427, + 243 + ], + "lines": [ + { + "bbox": [ + 318, + 217, + 427, + 243 + ], + "spans": [ + { + "bbox": [ + 318, + 217, + 427, + 243 + ], + "type": "text", + "content": "(g) Trajectory masking, yes or no? Yes. Both targets are features." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 430, + 72, + 539, + 127 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 539, + 127 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 539, + 127 + ], + "type": "table", + "html": "
TrainingK400mSSv2m
Single78.961.6
Mix79.562.2
Progressive79.862.5
", + "image_path": "bcbf5c71c24dc355732ea95192d02a5599029da1b43c28a5112e9cf1a6924b0d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 438, + 170, + 538, + 215 + ], + "blocks": [ + { + "bbox": [ + 429, + 129, + 539, + 166 + ], + "lines": [ + { + "bbox": [ + 429, + 129, + 539, + 166 + ], + "spans": [ + { + "bbox": [ + 429, + 129, + 539, + 166 + ], + "type": "text", + "content": "(d) Which training strategy? Progressive training outperforms both for the same number of iterations." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 438, + 170, + 538, + 215 + ], + "lines": [ + { + "bbox": [ + 438, + 170, + 538, + 215 + ], + "spans": [ + { + "bbox": [ + 438, + 170, + 538, + 215 + ], + "type": "table", + "html": "
TargetK400mSSv2m
Pixels66.055.1
Features78.160.2
", + "image_path": "27cd24efdce3554e8928c72f69b6b075af88318821c4f70e0cde4d45f73da70d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 432, + 217, + 544, + 244 + ], + "lines": [ + { + "bbox": [ + 432, + 217, + 544, + 244 + ], + "spans": [ + { + "bbox": [ + 432, + 217, + 544, + 244 + ], + "type": "text", + "content": "(h) Reconstructing pixels or features? Reconstructing features is much better. No motion is added." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "type": "text", + "content": "Table 4. Ablation experiments of our proposed SMILE. The default setting uses clip reconstruction targets with two synthetic objects and a " + }, + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "type": "text", + "content": " masking ratio with trajectory-based masking. We also show full-scale ablations with ViT-B and K400 pretraining in supplementary." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 295, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 295, + 390 + ], + "type": "text", + "content": "Which training strategy? Table 4d compares our proposed training strategies. We also compare learning with the single learning strategy, where only augmented videos are used for training. We observe that both mixed and progressive learning are better than learning from augmented videos only. Progressive is better than mixed, most likely due to a reduced domain shift in the second stage where only original videos are used." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 391, + 296, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 391, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 54, + 391, + 296, + 463 + ], + "type": "text", + "content": "Qualitative analysis. To compare the temporal dynamics of video representations learned by different SSL methods, we compute their feature similarity across different frames of input videos (Figure 4). For our model, the features of different frames have larger differences, indicating better temporal awareness. More plots are in the supplementary." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 67, + 475, + 287, + 552 + ], + "blocks": [ + { + "bbox": [ + 67, + 475, + 287, + 552 + ], + "lines": [ + { + "bbox": [ + 67, + 475, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 67, + 475, + 287, + 552 + ], + "type": "image", + "image_path": "7c92a6abf5b929a132587001f36c07fc6dd7ab20db7b2da7811b90df707fdb57.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "Figure 4. Feature similarity across different frames for different SSL methods. We compute this on K400 validation videos." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 590, + 204, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 204, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 204, + 601 + ], + "type": "text", + "content": "4.4. Generalization Assessment" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 700 + ], + "type": "text", + "content": "To demonstrate the generalization of our learned video representations to diverse downstream video setups, we evaluate SMILE on the SEVERE benchmark [55], which includes 8 experiments targeting 4 key generalization factors: domain shift, sample efficiency, action granularity, and task shift. Detailed configurations are provided in the supplementary. Beyond SEVERE, we also show generalization to more temporal tasks in supplementary." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "type": "text", + "content": "Table 6 shows the results. For the Domain Shift, we test" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "text", + "content": "cross-domain adaptability on SSv2 and Gym99, which differ from the pretraining dataset (K400). Our method significantly outperforms all others, demonstrating superior adaptability to domain shifts. For Sample Efficiency, we evaluate low-shot action recognition on UCF and GYM with only 1,000 samples for finetuning. Our approach surpasses previous methods, achieving a " + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "text", + "content": " improvement on GYM over the best prior model, MME, showcasing its strong low-shot recognition capability even in motion-sensitive tasks. For the Action Granularity, tested on FX-S1 and UB-S1 from FineGym, our model achieves the best results, excelling at distinguishing fine-grained motions within gymnastics routines (e.g., jump variations in FX-S1). It outperforms motion-aware models like MME, MGMAE, and MGM, highlighting superior motion generalizability. For Task Shift, evaluated with temporal repetition counting on UCF-RC [70] and multi-label recognition on Charades [51], our method matches prior performance on repetition counting and achieves a " + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "text", + "content": " improvement over MME on Charades, demonstrating broad task adaptability. Overall, SMILE (without synthetic motion) is on par with the prior best with the mean performance of " + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "inline_equation", + "content": "64.9\\%" + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "text", + "content": ". With synthetic motion SMILE, we observe a further " + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 313, + 295, + 555, + 582 + ], + "type": "text", + "content": " gain, reinforcing our model's generalization strength through synthetic motion." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 591, + 389, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 389, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 389, + 602 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Comparison with Other CLIP Adaptations. Table 5 presents a comparison between our SMILE and the CLIP adaptation methods ViCLIP [64] and UMT [31], along with the CLIP baseline. Our method achieves superior performance on motion-sensitive datasets, particularly in linear probing, indicating that our CLIP adaptation learns better video representations than the video-text alignment of ViCLIP and the unmasked feature alignment of UMT, despite using less data. More results are in the supplementary." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8444" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 392, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 392, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 392, + 159 + ], + "type": "table", + "html": "
PretrainingFinetuningLinear Probing
MethodVideo DatasetUCFGYMSSv2EPICUCFGYMSSv2EPIC
CLIP [45]-93.688.066.750.377.520.711.325.1
ViCLIP [64]Intervid-10M [64]95.289.767.955.086.727.318.927.3
UMT [31]K700 [5]96.089.970.150.188.026.418.828.2
SMILE (ours)K400 [26]96.490.871.963.383.830.223.734.6
", + "image_path": "a52415acee403f50253fa38869f38eb37b415fcfffc192b601ddf8c0501ad8e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 205, + 531, + 331 + ], + "blocks": [ + { + "bbox": [ + 397, + 70, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 397, + 70, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 397, + 70, + 555, + 137 + ], + "type": "text", + "content": "Table 5. Comparison with CLIP adaptations for video representation learning. Despite using less data, we learn better video representations than other CLIP adaptations, especially for motion-focused downstream domains." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 168, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 555, + 203 + ], + "type": "text", + "content": "Table 6. Generalization assessment on SEVERE benchmark [55] for recent video SSL methods in terms of domain shift, sample efficiency, action granularity, and task shift. We evaluate prior works with the official ViT-B backbone checkpoints pretrained on the K400 dataset. SMILE shows the best generalization performance across setups. Adding synthetic motions further improves the performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 205, + 531, + 331 + ], + "lines": [ + { + "bbox": [ + 80, + 205, + 531, + 331 + ], + "spans": [ + { + "bbox": [ + 80, + 205, + 531, + 331 + ], + "type": "table", + "html": "
MethodDomain shiftSample efficiency (103)Action granularityTask shiftMean
SSv2Gym99UCFGYMFX-S1UB-S1UCF-RC↓Charades
VideoMAE [57]68.686.674.625.942.865.30.17214.457.6
MVD [63]70.082.567.117.531.350.50.18416.152.1
MGMAE [21]68.987.277.224.133.779.50.18117.958.8
MGM [12]71.189.178.426.438.686.90.15222.562.2
SIGMA [49]70.989.784.128.055.179.90.16923.164.2
MME [53]70.189.779.229.855.587.20.15523.665.0
SMILE w/o motion71.690.085.232.053.474.30.17530.564.9
SMILE (ours)72.190.886.435.155.188.30.17032.567.9
", + "image_path": "c2335615604a3f683b7b97f3e8eb76c29e3c46a0bb5c17dae8937073cae2d2a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 69, + 379, + 280, + 483 + ], + "blocks": [ + { + "bbox": [ + 55, + 343, + 295, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 376 + ], + "type": "text", + "content": "Table 7. Learning without natural videos. We train a ViT-S for pixel reconstruction. Our synthetic object motions can learn video representations without using any natural videos." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 379, + 280, + 483 + ], + "lines": [ + { + "bbox": [ + 69, + 379, + 280, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 379, + 280, + 483 + ], + "type": "table", + "html": "
Unnatural Video SchemePretrain DataK400mSSv2m
No Pretraining-39.728.2
Video-frameK400m-1frm54.043.1
Video-frame + MotionK400m-1frm63.653.2
ImagePlaces-80K52.140.5
Image + MotionPlaces-80K60.152.1
Black image + Motion-59.250.0
Noise image + Motion-61.951.2
", + "image_path": "a6f6e19a5e9d0fad41310ea671ade2d5c91c2831ee166d1f9955c6a8f7155311.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "text", + "content": "Learning without Natural Videos. In prior experiments, we overlay our object motions onto natural videos from the Kinetics-400 dataset, raising the question of whether downstream improvements stem from original video priors (motion, action, scene) or our added object motions. To investigate this, we construct videos by overlaying object motions onto four different backgrounds: (1) a single frame from natural videos, (2) a single scene image from the image dataset Places [73], (3) a plain black image, and (4) a noise image. Each background image is duplicated across " + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "text", + "content": " frames, creating static clips onto which moving objects are placed. For (1), we select one frame per video from " + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "text", + "content": " (denoted " + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m - 1frm}" + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "text", + "content": "), preserving action and scene cues without motion priors. For (2), we sample 80K images from Places (Places-80K), matching the dataset size of " + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m - 1frm}" + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "text", + "content": ", to remove action and motion priors. Approaches (3) and (4) exclude all natural video or image priors entirely. Evaluation involves full finetuning on " + }, + { + "bbox": [ + 54, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 343, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 545, + 355 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 343, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathrm{SSv2_m}" + }, + { + "bbox": [ + 313, + 343, + 545, + 355 + ], + "type": "text", + "content": ", consistent with our previous ablation studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "text", + "content": "Table 7 reports results for different augmented video types. Adding object motions consistently enhances downstream performance, whether combined with action and scene cues from natural videos (row 2 vs. row 3) or only scene cues from images (row 4 vs. row 5). Even with black or noisy backgrounds (rows 6 and 7), performance remains robust. Remarkably, using object motions alone yields substantial improvements—" + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{K400_m}" + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "inline_equation", + "content": "23.0\\%" + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{SSv2_m}" + }, + { + "bbox": [ + 313, + 356, + 555, + 523 + ], + "type": "text", + "content": " compared to no pretraining. These findings demonstrate our method's effectiveness in learning video representations from synthetic motions alone. More results are in the supplementary. In summary, our method enables a new paradigm of learning strong video representations with masked video modeling without relying on natural videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 529, + 392, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 392, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 392, + 541 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "In this work, we proposed SMILE, a novel self-supervised approach for video representation learning that addresses key limitations in existing masked video modeling methods by integrating both spatial and motion semantics. Our SMILE, leverages the high-level spatial semantics from image-language pretrained models like CLIP and enhances motion representation through synthetic motion patterns. By capturing complex spatial features and motion dynamics, SMILE achieves superior performance on diverse downstream tasks, demonstrating improved discrimination and generalization over state-of-the-art video SSL methods. Our extensive evaluations demonstrated the advantages of our SMILE, setting a new paradigm for robust and effective video self-supervised learning." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8445" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the KAUST Center of Excellence for Generative AI under award number 5940. For computing time, this research used Ibex managed by the Supercomputing Core Laboratory at King Abdullah University of Science & Technology (KAUST) in Thuwal, Saudi Arabia. We also thank Michael Dorkenwald for providing the synthetic objects used in this work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 196, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 196, + 297, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 297, + 251 + ], + "type": "text", + "content": "[1] Unaiza Ahsan, Rishi Madhok, and Irfan Essa. Video jigsaw: Unsupervised learning of spatiotemporal context for video action recognition. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "type": "text", + "content": "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 307, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 307, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 307, + 296, + 361 + ], + "type": "text", + "content": "[3] Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 362, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 296, + 416 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 417, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 417, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 417, + 295, + 449 + ], + "type": "text", + "content": "[5] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 295, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 505 + ], + "type": "text", + "content": "[6] Peihao Chen, Deng Huang, Dongliang He, Xiang Long, Runhao Zeng, Shilei Wen, Mingkui Tan, and Chuang Gan. Rspnet: Relative speed perception for unsupervised video representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 295, + 547 + ], + "type": "text", + "content": "[7] Hyeon Cho, Taehoon Kim, Hyung Jin Chang, and Wonjun Hwang. Self-supervised visual learning by variable playback speeds prediction of a video. IEEE Access, 9:79562-79571, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 613 + ], + "type": "text", + "content": "[8] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for EPIC-KITCHENS-100. International Journal of Computer Vision (IJCV), 2021. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 295, + 680 + ], + "type": "text", + "content": "[9] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[10] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[11] Hazel Doughty, Fida Mohammad Thoker, and Cees GM Snoek. Locomotion: Learning motion-focused video-language representations. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 50-70, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[12] David Fan, Jue Wang, Shuai Liao, Yi Zhu, Vimal Bhat, Hector Santos-Villalobos, Rohith MV, and Xinyu Li. Motion-guided masking for spatiotemporal representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[13] David Fan, Jue Wang, Shuai Liao, Zhikang Zhang, Vimal Bhat, and Xinyu Li. Text-guided video masked autoencoder. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "text", + "content": "[14] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[15] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[16] Kirill Gavrilyuk, Mihir Jain, Ilia Karmanov, and Cees G M Snoek. Motion-augmented self-training for video recognition at smaller scale. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "text", + "content": "[17] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimaes: Single model masked pretraining on images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 536 + ], + "type": "text", + "content": "[18] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "text", + "content": "[19] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[21] Bingkun Huang, Zhiyu Zhao, Guozhen Zhang, Yu Qiao, and Limin Wang. Mgmae: Motion guided masking for video masked autoencoding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[22] Ziyuan Huang, Shiwei Zhang, Jianwen Jiang, Mingqian Tang, Rong Jin, and Marcelo H Ang. Self-supervised motion" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8446" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "type": "text", + "content": "learning from static images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "text", + "content": "[23] Sunil Hwang, Jaehong Yoon, Youngwan Lee, and Sung Ju Hwang. Everest: Efficient masked video autoencoder by removing redundant spatiotemporal tokens. In International Conference on Machine Learning (ICML), 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[24] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 230 + ], + "type": "text", + "content": "[25] Longlong Jing, Xiaodong Yang, Jingen Liu, and Yingli Tian. Self-supervised spatiotemporal feature learning via video rotation prediction. arXiv preprint arXiv:1811.11387, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "text", + "content": "[26] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 1, 2, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "type": "text", + "content": "[27] Dahun Kim, Donghyeon Cho, and In So Kweon. Self-supervised video representation learning with space-time cubic puzzles. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[28] Okan Köpukli, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified CNN architecture for real-time spatiotemporal action localization. arXiv preprint arXiv:1911.06644, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "text", + "content": "[29] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2011. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[30] Hsin-Ying Lee, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Unsupervised representation learning by sorting sequences. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[31] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[32] Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[33] Dezhao Luo, Chang Liu, Yu Zhou, Dongbao Yang, Can Ma, Qixiang Ye, and Weiping Wang. Video cloze procedure for self-supervised spatio-temporal learning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[34] Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In European Conference on Computer Vision (ECCV), 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audiovisual instance discrimination with cross-modal agreement." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[36] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[37] Jingcheng Ni, Nan Zhou, Jie Qin, Qian Wu, Junqi Liu, Boxun Li, and Di Huang. Motion sensitive contrastive learning for self-supervised video representation. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[38] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 327 + ], + "type": "text", + "content": "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 381 + ], + "type": "text", + "content": "[40] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. Videomoco: Contrastive video representation learning with temporally adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 414 + ], + "type": "text", + "content": "[41] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Dual-path adaptation from image to video transformers. arXiv preprint arXiv:2303.09857, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[42] Mandela Patrick, Yuki M. Asano, Polina Kuznetsova, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[43] AJ Piergiovanni, Anelia Angelova, and Michael S Ryoo. Evolving losses for unsupervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "type": "text", + "content": "[44] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[46] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S. Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[47] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8447" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[49] Mohammadreza Salehi, Michael Dorkenwald, Fida Mohammad Thoker, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Sigma: Sinkhorn-guided masked video modeling. In European Conference on Computer Vision (ECCV), 2025. 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "text", + "content": "[50] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "type": "text", + "content": "[51] Gunnar A. Sigurdsson, Gúl Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in Homes: Crowdsourcing Data Collection for Activity Understanding. In European Conference on Computer Vision (ECCV), 2016. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[52] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 430 + ], + "type": "text", + "content": "[53] Xinyu Sun, Peihao Chen, Liangwei Chen, Changhao Li, Thomas H. Li, Mingkui Tan, and Chuang Gan. Masked motion encoding for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "type": "text", + "content": "[54] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the ACM International Conference on Multimedia (ACMMM), 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In European Conference on Computer Vision (ECCV), 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "type": "text", + "content": "[56] Fida Mohammad Thoker, Hazel Doughty, and Cees G. M. Snoek. Tubelet-contrastive self-supervision for videoefficient generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[57] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[58] Shuyuan Tu, Qi Dai, Zuxuan Wu, Zhi-Qi Cheng, Han Hu, and Yu-Gang Jiang. Implicit temporal modeling with learnable alignment for video recognition. arXiv preprint arXiv:2304.10465, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[59] Guangting Wang, Yizhou Zhou, Chong Luo, Wenxuan Xie, Wenjun Zeng, and Zhiwei Xiong. Unsupervised visual representation learning by tracking patches in video. In Pro" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "type": "text", + "content": "[60] Jinpeng Wang, Yuting Gao, Ke Li, Yiqi Lin, Andy J Ma, Hao Cheng, Pai Peng, Feiyue Huang, Rongrong Ji, and Xing Sun. Removing the background by adding the background: Towards background robust self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[61] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[62] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. Bevt: Bert pretraining of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 326 + ], + "type": "text", + "content": "[63] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[64] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 2, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[65] Fanyi Xiao, Joseph Tighe, and Davide Modolo. Maclr: Motion-aware contrastive learning of representations for videos. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[66] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "text", + "content": "[67] Haosen Yang, Deng Huang, Bin Wen, Jiannan Wu, Hongxun Yao, Yi Jiang, Xiatian Zhu, and Zehuan Yuan. Self-supervised video representation learning with motion-aware masked autoencoders. arXiv preprint arXiv:2210.04154, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 579 + ], + "type": "text", + "content": "[68] Taojiannan Yang, Yi Zhu, Yusheng Xie, Aston Zhang, Chen Chen, and Mu Li. Aim: Adapting image models for efficient video action recognition. arXiv preprint arXiv:2302.03024, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "text", + "content": "[69] Yuan Yao, Chang Liu, Dezhao Luo, Yu Zhou, and Qixiang Ye. Video playback rate perception for self-supervised spatio-temporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[70] Huaidong Zhang, Xuemiao Xu, Guoqiang Han, and Shengfeng He. Context-aware and scale-insensitive temporal repetition counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[71] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8448" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "text", + "content": "Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy-paste for instance segmentation using clip and stablediffusion. In International Conference on Learning Representations (ICLR), 2023. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[72] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[73] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8449" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_content_list.json b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a1275381528b592713c851e9d52c09cccd80f6e9 --- /dev/null +++ b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_content_list.json @@ -0,0 +1,1904 @@ +[ + { + "type": "text", + "text": "SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity", + "text_level": 1, + "bbox": [ + 91, + 130, + 903, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Yijie Xu*, $^{1}$ Bolun Zheng*, $^{1}$ Wei Zhu*, $^{1}$ Hangjia Pan, $^{1}$ Yuchen Yao, $^{2}$ Ning Xu, $^{2}$ Anan Liu, $^{3}$ Quan Zhang, $^{1}$ Chenggang Yan $^{1}$ Hangzhou Dianzi University, $^{2}$ Tianjin University, $^{3}$ Peking University", + "bbox": [ + 225, + 179, + 769, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Social media popularity prediction task aims to predict the popularity of posts on social media platforms, which has a positive driving effect on application scenarios such as content optimization, digital marketing and online advertising. Though many studies have made significant progress, few of them pay much attention to the integration between popularity prediction with temporal alignment. In this paper, with exploring YouTube's multilingual and multi-modal content, we construct a new social media temporal popularity prediction benchmark, namely SMTPD, and suggest a baseline framework for temporal popularity prediction. Through data analysis and experiments, we verify that temporal alignment and early popularity play crucial roles in social media popularity prediction for not only deepening the understanding of temporal dynamics of popularity in social media but also offering a suggestion about developing more effective prediction models in this field. Code is available at https://github.com/zhuwei321/SMTPD", + "bbox": [ + 88, + 300, + 485, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 584, + 220, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advancement of Internet communication technology in recent years, social media has gradually emerged and has influenced various aspects of human life. Any content posted on social media stands the chance of becoming hot spot, and widely disseminated social media content can generate significant social and economic benefits. The prediction of social media popularity holds immense potential applications in content optimization [1, 30], online advertising [14, 15], digital marketing [19, 45], search recommendations [4, 17], intelligent fashion [10], and beyond.", + "bbox": [ + 89, + 609, + 482, + 761 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the early stages of popularity research, statistical and topological methods [41, 43, 53] were extensively employed, with the research focus primarily on time-aware popularity prediction. These prediction methods rely on information about the earlier popularity of posted content.", + "bbox": [ + 89, + 762, + 483, + 838 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3265d56d3056fdfb495ea22510f3d30c86ae82ee55356ad25a5632066d9d33f.jpg", + "image_caption": [ + "Figure 1. Content sections and popularity trend of SMTPD. In 1a, a sample involves four sections of content, with temporal popularity. 1b depicts box plots of daily popularity scores, illustrating variations in popularity distribution at different time points. The distribution consistently demonstrates a decay pattern over time." + ], + "image_footnote": [], + "bbox": [ + 521, + 267, + 911, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Along with the many large-scale social media popularity datasets are proposed [3, 42, 47, 48], the machine learning based methods are widely employed to achieve reasonable predictions of social media popularity and have shown remarkable performance. These methods emphasize meticulous design in feature extraction and fusion, followed by popularity prediction using machine learning models such as deep neural networks (DNN) and gradient boosting decision trees (GBDT).", + "bbox": [ + 511, + 708, + 906, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1 lists the public or mentioned social media popularity datasets in the most recent years as far as we know. For above machine learning based methods, the large-scale", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 112, + 849, + 370, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 112, + 863, + 241, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is supported by the National Key Research and Development Program of China under Grant 2020YFB1406600.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/295fcc04cec6a713a1912410cf72fe2107d4f153f9a5a6d9d33c0e6a7ab889bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSourceCategorySamplesLanguagePrediction Type
Mazloom [35]Instagramfast food brand75KEnglishsingle
Sanjo [42]Cookpadrecipe150KJapanesesingle
TPIC17 [47]Flickr-680KEnglishsingle
SMPD [48]Flickr11 categories486KEnglishsingle
AMPS [11]YouTubeshorts13KKoreansingle
SMTPD (ours)YouTube15 categories282Kover 90 languagessequential
", + "bbox": [ + 91, + 88, + 483, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparing SMTPD with existed multi-modal social media popularity datasets.", + "bbox": [ + 89, + 215, + 482, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "social media popularity datasets [3, 11, 35, 42, 47, 48] provide the fundamental supports.", + "bbox": [ + 89, + 253, + 482, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing datasets and related studies have notable limitations, such as insufficient multi-modal data, limited language diversity, and, crucially, the absence of a consistent timeline for prediction. As shown in Figure1b, popularity distributions shift over time, and predictions made at irregular intervals can reduce accuracy and create challenges for practical applications.", + "bbox": [ + 89, + 287, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, effective popularity prediction requires integrating both the multi-modal content of social media and information on the communication process to capture the dynamic trends in popularity. This essential aspect is missing in most current prediction efforts.", + "bbox": [ + 89, + 397, + 483, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Targeting to fix up the above shortcomings of existed datasets, in this paper, we propose a new benchmark called Social Media Temporal Popularity Dataset (SMTPD) by observing multi-modal content from mainstream for cutting-edge research in the field of social multimedia along with multi-modal feature temporal prediction. As shown in Figure 1a, multi-modal contents of social media posts in multiple languages generated the daily popularity information during the communication process. We define such a post as a sample of SMTPD, and propose a multi-modal framework as a baseline to achieve the temporal prediction of popularity scores. The proposed framework consists of two parts, feature extraction and regression. In the feature extraction part, multiple pre-training models and pre-processing methods are introduced to translate the multi-modal content into deep features. In the regression part, we encode the state and time sequence of the extracted features, and adopt the LSTM-based structure to regress the 30-day popularities. Through analysis and experiments, we discover the importance of early popularity to the task of popularity prediction, and demonstrate the effectiveness of our method in temporal prediction. Generally, the contribution of this work can be summarized as:", + "bbox": [ + 89, + 474, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Against the missing of temporal information in social media popularity researches, we observe over 282K multilingual samples from mainstream social media since they released on the network lasting for 30 days. We refer to these samples as SMTPD, a new benchmark for temporal", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "popularity prediction.", + "bbox": [ + 527, + 92, + 673, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Basing on existed methods, we innovate in both the selection of feature extractors and the construction of the temporal regression component, and suggest a baseline model which enables temporal popularity prediction to be conducted across multiple languages while aligning prediction times.", + "- Exploring the popularity distribution and the correlation between popularity at different times. Based on these, We find the importance of early popularity for popularity prediction task, and point out that the key-point for predicting popularity is to accurately predict the early popularity." + ], + "bbox": [ + 513, + 107, + 903, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 286, + 650, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Statistical and topological methods. Szabo et al. [43] found that the early-stage popularity notably influences subsequent popularity. Yang et al. [53] conducted cluster analysis on temporal patterns within online content, revealing distinct characteristics of popularity variations across different clusters. Richier et al. [40, 41] proposes bio-inspired models to characterize the evolution of video view counts. Wu et al. [51] suggested that the information cascade process is best described as an activate-decay dynamic process.", + "bbox": [ + 511, + 311, + 903, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modal feature based Methods. Ding et al. [13] use pre-trained ResNet and BERT to extract visual and textual features, with the DNN regression. Xu et al. [52] and Lin et al. [31] adopted attention mechanisms to effectively integrate multi-modal features. Chen et al. [6] compared the performance of several regression models, among which XGBoost [7] exhibited the best results. Hsu et al. [22] employed LightGBM [23] and TabNet [2] to capture intricate semantic relationships in multi-modal features. Lai et al. [26] engineered handcrafted features, exploiting CatBoost for regression. Mao et al. [34] enhanced CatBoost-based model by stacking features. Tan et al. [44] extracted visual-textual features by ALBEF [29] and Chen et al. [8, 9] enriched more intermodal features to promote predictive performance. Wu et al. [50] emphasized increasing feature dimensions to improving predictive performance. The post dependencies captured by sliding window average [46] and DSN [54] has also led to improvements. Many of these multi-modal methods are based on the SMPD and working out great [32, 49].", + "bbox": [ + 511, + 448, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Social media popularity datasets. Mazloom et al. [35] conducted experiments using a dataset of posts related to fast-food brands collected from Instagram. Sanjo et al. [42] provided a recipe popularity prediction dataset based on Cookpad, which includes text content entirely in Japanese. Li et al. [28] predicted the future popularity of topics by using historical sentiment information based on Twitter's text data records. TPIC17 [47] and SMPD [48] are datasets for single popularity prediction task based on Flickr.", + "bbox": [ + 511, + 750, + 903, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We believe that the loss of temporal alignment about", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18848", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "popularity in existing methods is a common problem. The image dynamic popularity dataset [37] gave us insights, so we propose a multi-modal temporal popularity benchmark to address the shortcomings of existing studies.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SMTPD Dataset", + "text_level": 1, + "bbox": [ + 89, + 164, + 251, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the great development of web communication technologies, social multimedia has become the most popular media around the world. However, most existing social media prediction datasets are built on the basis of single-output prediction, meaning they observe existing posts and infer their popularity based on their posting time. As mentioned earlier, these data without aligned posting times exhibit non-uniform popularity distributions. Therefore, SMTPD keep an eye on one of the most popular worldwide social media, YouTube [5], with over 282K samples, primarily focusing on the evolution of popularity over time for these samples. We collected over 402K raw data samples. Specifically, we first removed records with missing values—which likely resulted from network issues during data acquisition. Next, we eliminated samples that either failed to be crawled on certain days or had been deleted within 30 days. Finally, we filtered out potential outliers using the $3\\sigma$ rule. In this section, we provide a comprehensive overview of the SMTPD's composition. Additionally, we present various data analysis results that contribute to the feasibility of our approach.", + "bbox": [ + 89, + 189, + 483, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the retrospective methods used in previous datasets, we take note of the newly posted multi-modal content and then observe popularity information every 24 hours via YouTube API. Considering that in real-world applications, the posts to be predicted are always new, this data observation method allows us to obtain aligned temporal popularity of new posts, which is more in line with practical applications. Figure 1a show a sample from SMTPD. We note diverse attributes, including visual content, textual content, metadata, and user profiles. Additionally, we focus on the temporal popularity for each sample within 30 days after release. We perform basic statistics for SMTPD, as shown in Table 2.", + "bbox": [ + 89, + 507, + 483, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal Popularity", + "text_level": 1, + "bbox": [ + 89, + 712, + 287, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The utilization of temporal popularity data is versatile. It not only serves as the output for predictions but also as inputs, where a segment preceding a specific time point, coupled with multi-modal content, can forecast subsequent popularity. We analyze the distribution and the correlation of popularity over time, so as to facilitate a comprehensive understanding of the significance of temporal popularity for predictions from different perspectives.", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A number of popularity definitions have been born from previous work [27, 55]. We use Khosla's popularity score transformation [24] as it can normalize the distribution of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "view counts to a suitable popularity score's distribution. It can be represented as:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np = \\log_ {2} (\\frac {v}{d} + 1) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 128, + 906, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p$ is the popularity score, $v$ is the view counts of a sample, $d$ represents the corresponding number of days after the post's release. As shown in Figure 1, the histogram of view counts reveals an extreme long-tailed distribution, which might not be suitable as a prediction target. It's evident that the distribution of the popularity score metric becomes more reasonable.", + "bbox": [ + 511, + 161, + 906, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, we notice that there is a strong correlation between the popularity scores of different days. The Pearson Correlation (PC) and Spearman Ranking Correlation (SRC) can respectively reflect the linear relationship and rank-order correlation between popularity scores of different days as follows:", + "bbox": [ + 511, + 266, + 905, + 356 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP C = \\frac {\\sum_ {i = 1} ^ {n} \\left(X _ {i} - \\bar {X}\\right) \\left(Y _ {i} - \\bar {Y}\\right)}{\\sqrt {\\sum_ {i = 1} ^ {n} \\left(X _ {i} - \\bar {X}\\right) ^ {2}} \\sqrt {\\sum_ {i = 1} ^ {n} \\left(Y _ {i} - \\bar {Y}\\right) ^ {2}}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 376, + 906, + 421 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS R C = \\frac {1}{n - 1} \\sum_ {i = 1} ^ {n} \\left(\\frac {X _ {i} - \\bar {X}}{\\sigma_ {X}}\\right) \\left(\\frac {Y _ {i} - \\bar {Y}}{\\sigma_ {Y}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 429, + 905, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $X$ and $Y$ denote the popularity scores belonging to 2 different days, while $\\bar{X}$ and $\\sigma_{X}$ denote the mean and standard deviation of $X$ , and similarly for $Y$ . PC and SRC respectively reflect the linear relationship and rank-order correlation between popularity scores of different days.", + "bbox": [ + 511, + 470, + 905, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Pearson Correlation (PC) and Spearman Rank Correlation (SRC) heat maps illustrate internal relationships in daily popularity scores, establishing a basis for prediction based on these temporal patterns. Further details are available in the supplementary materials.", + "bbox": [ + 511, + 547, + 905, + 623 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e56ccbd802427fb35da9cf6e6f7a2e877a5a00b75a64abe6a98d641ab57b7640.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StatisticsSMTPDSMPD (Train/Test)
Number of samples282.4K305K/181K
Number of users152.7K38K/31K
Mean popularity5.956.41/5.12
STD popularity4.152.47/2.41
Number of categories1511
Number of custom tags960K250K
Average length of title53.4 chars29 words
Mean duration1853.4 s-
", + "bbox": [ + 514, + 630, + 906, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. The basic statistics of all samples in SMPD, comparing to SMPD[48]. Due to the multilingual environment, we use chars (characters) to describe the average length of title.", + "bbox": [ + 511, + 854, + 906, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18849", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Modal Content", + "text_level": 1, + "bbox": [ + 89, + 90, + 294, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-modal feature fusion is the mainstream approach in today's popularity prediction methods. In this section, we will divide the data into corresponding modality and elucidate the multi-modal content in SMTPD.", + "bbox": [ + 89, + 112, + 483, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Visual Content", + "text_level": 1, + "bbox": [ + 89, + 180, + 241, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SMTPD samples, primarily drawn from a multi-modal social media platform, are heavily influenced by visual elements. Since metrics like view count and popularity score are often based on the cover frame that viewers see before clicking to play, we use the cover frame as the primary visual content. This approach balances popularity impact with copyright considerations.", + "bbox": [ + 89, + 199, + 483, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Textual Content", + "text_level": 1, + "bbox": [ + 89, + 313, + 250, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The textual content in posts plays a crucial role in media dissemination and user engagement. We examine several core elements, including category, title, description, and hashtags (denoted by the “#” symbol), along with the user ID (user nickname) from profile metadata. YouTube defines 15 distinct video categories, all in English, which support content organization and discovery. Figure 3 shows the statistical distribution of content by category.", + "bbox": [ + 89, + 332, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7bb60ec655ff6911907a94c33cdcbf578ca08183d4eafdc051bf26f2de1f3ea3.jpg", + "image_caption": [ + "(a) The PC Matrix" + ], + "image_footnote": [], + "bbox": [ + 91, + 465, + 251, + 589 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4f299ea09a29b30b6b778651443af19eeca7beb6115a9796887f5f3cb7516c23.jpg", + "image_caption": [ + "(b) The SRC Matrix" + ], + "image_footnote": [], + "bbox": [ + 254, + 465, + 483, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As an international platform, YouTube hosts user-defined textual content in multiple languages beyond standard categories. Figure 4 illustrates the distribution and popularity bias across title languages.", + "bbox": [ + 89, + 676, + 482, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In previous researches, the datasets typically had uniform language for text content, without considering multilingual aspects. While this simplifies modeling for prediction methods, it also poses significant limitations for international social media platforms which not restricted to a single language. Besides, predictions based on a few languages may introduce biases [16] against popularity.", + "bbox": [ + 89, + 738, + 483, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Numerical values", + "text_level": 1, + "bbox": [ + 89, + 851, + 256, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Numerical attributes frequently function as supplementary factors in enhancing prediction accuracy. In this study, we", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "examined key attributes, including video duration, uploader follower count, and the number of posts by the uploader. Additionally, we manually derived several auxiliary metrics, such as title length, the count of custom tags, and description length, to further enrich the feature set and support predictive robustness.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Temporal popularity prediction", + "text_level": 1, + "bbox": [ + 511, + 198, + 803, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-modal feature-based temporal prediction framework is shown in Figure 5, which is divided into two main parts: multi-modal feature extraction and temporal popularity score regression. This framework is designed to adapt to the multilingual temporal prediction tasks under time alignment. The prediction target is set as the popularity of samples, which corresponds to the temporal popularity within 30 days after the sample's release.", + "bbox": [ + 511, + 224, + 906, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Multi-Modal Feature Extraction", + "text_level": 1, + "bbox": [ + 511, + 358, + 799, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For different modalities in SMTPD, we adopted distinct feature extraction methods. We additionally incorporated features from the categorical modality to investigate the impact of categorical features on popularity prediction.", + "bbox": [ + 511, + 380, + 905, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Visual Features", + "text_level": 1, + "bbox": [ + 511, + 452, + 668, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cover image of a social media content is a very important component which would help users to quickly understand what they would see if clicking to view this content.", + "bbox": [ + 511, + 472, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26bd9072261138410e4150ff6346bb7603c8c0ffcb161721b2cb606bf9b3cd98.jpg", + "image_caption": [ + "Figure 2. The heat map of daily popularity correlations. The figure clearly shows that there is a high degree of correlations in popularity between consecutive days." + ], + "image_footnote": [], + "bbox": [ + 516, + 527, + 880, + 667 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cef5d7564335d9ed167cf9ab5edb8f076c9a49a50bf4964b5854ff49d3395ad2.jpg", + "image_caption": [ + "(a) The statistics of samples in each category", + "(b) Average popularity of each category", + "Figure 3. The statistics based on category. 3a counts the number of samples in each category, and 3b shows the average popularity score of samples in each category." + ], + "image_footnote": [], + "bbox": [ + 516, + 684, + 901, + 828 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18850", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/adedc1fccad5168bb90b96d8b059cb0fc4848ca40084985b77bd1c06cdf3d532.jpg", + "image_caption": [ + "Figure 4. Languages analysis. The left is the proportions of these languages, with \"Other\" encompassing 90 different languages. The right represents the average popularity in different languages, revealing the geographic biases. The languages of samples are counted by reference to the title." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 480, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, the cover image would directly affect the popularity of a social media content. The semantic information provided in the cover image plays a key role for user's understanding. We adopt convolutional neural networks ResNet-101 [20] pre-trained on ImageNet as our visual feature extraction model to obtain the semantic information provided by the cover. To match the input size of ResNet-101, we re-scale the size of cover image to $224 \\times 224$ . The 2048-dimension feature vector before the final classification layer will be selected as the final visual feature $f_{v}$ :", + "bbox": [ + 89, + 353, + 483, + 505 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {v} = \\operatorname {R e s N e t} (S (I)) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 515, + 480, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I$ denotes the cover image and $S$ denotes the re-scale operation. However, the cover image would not always be uploaded by authors or be unavailable due to the network transmission. For those contents missing the cover image, we use blank images with all pixels are set to zero instead.", + "bbox": [ + 89, + 542, + 483, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Textual Features", + "text_level": 1, + "bbox": [ + 89, + 625, + 253, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual information significantly impacts a user's choice to view or skip content. Key textual inputs include category, title, tags, description, and user profile ID. Extracting semantic features from these inputs is crucial for accurate popularity prediction.", + "bbox": [ + 89, + 643, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, unlike previous popularity prediction tasks, one distinguishing characteristic of SMTPD is that its text content includes multiple languages, and many samples contain text in more than one language. This makes it challenging to use many pre-trained word vector models based on single-language corpora, such as [18, 36, 38].", + "bbox": [ + 89, + 719, + 483, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to the multilingual capability of BERT-Multilingual [12], which processes text across languages in a single model, we use it to extract a 768-dimensional feature vector for each text, capturing semantic information as:", + "bbox": [ + 89, + 810, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} ^ {k} = \\operatorname {B E R T} \\left(T ^ {k}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/750954a6f40a9685f6a6273ec7d4849189e064e6f84730d43c954b4deb1b5ea4.jpg", + "image_caption": [ + "Figure 5. The proposed model consists of two layers. The upper layer processes visual, textual, numerical, and categorical features, which are input to a multi-layer perceptron (MLP). The MLP outputs are concatenated and fed into a lower LSTM layer for regression. Each LSTM cell's output is passed through another MLP, with the final MLP outputs combined to generate a time prediction sequence using the ReLU activation function." + ], + "image_footnote": [], + "bbox": [ + 545, + 92, + 870, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k \\in \\{cat, tit, tag, des, uid\\}$ maps to category, title, tags, description, and user ID. These are combined into a $768 \\times 5$ matrix, transformed through $5 \\times 5$ convolution to produce the 768D textual feature $f_{t}$ , which can be expressed as:", + "bbox": [ + 511, + 476, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} = \\operatorname {C o n v} _ {5 \\times 5} \\left(\\mathcal {C} \\left(f _ {t} ^ {\\text {c a t}}, f _ {t} ^ {\\text {t i t}}, f _ {t} ^ {\\text {t a g}}, f _ {t} ^ {\\text {d e s}}, f _ {t} ^ {\\text {u i d}}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 558, + 906, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ denotes the above combing operation.", + "bbox": [ + 511, + 584, + 823, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Numerical Features", + "text_level": 1, + "bbox": [ + 511, + 604, + 697, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Numerical metrics directly influence whether an audience clicks to watch which sample on social media platform. Also, the number of followers and the number of posts also impact the process through which an audience navigates from content uploaders' user pages to click and watch videos. Hence, we selected these influential numerical metrics for numerical features extraction. As the number of user followers share a similar distribution pattern that exhibiting long-tailed characteristics as view counts, it is necessary to apply a logarithmic transformation before feeding it into the network. Afterward, all numerical items are Z-score normalized and concatenated as numerical features:", + "bbox": [ + 511, + 623, + 906, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n} ^ {j} = \\frac {T ^ {j} - \\mu_ {j}}{\\sigma_ {j}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 809, + 903, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n} = \\operatorname {C o n c a t} \\left(f _ {n} ^ {f o l}, f _ {n} ^ {p o s}, f _ {n} ^ {d u r}, f _ {n} ^ {t l}, f _ {n} ^ {t n}, f _ {n} ^ {d l}, f _ {n} ^ {E P}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 844, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j \\in \\{fol, pos, dur, tl, tn, dl, EP\\}$ represents the log-scaled follower counts, total number of posts, video du", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ration, title length, the number of tags, description length and the early popularity (the 1st day's popularity). The $\\mu$ and $\\sigma$ represents the mean and standard deviation of all samples. And Concat denotes the horizontal concatenating operation.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.4. Categorical Features", + "text_level": 1, + "bbox": [ + 89, + 179, + 282, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The categorical disparities across social media platforms are depicted in Figure 3b. Beyond text-based feature extraction, these disparities inform the generation of categorical features. Language functions as a key marker of regional and cultural distinctions among content creators, as highlighted by the language bias observed in Figure 4. Consequently, language classification features are crucial for enhancing prediction accuracy. To develop these classification features, two sub-features are selected: label and language attributes. The label feature captures the semantic category, while the language attribute is derived from the language in the video title. Assuming that the primary language of all text content matches the title's language, we use langid [33] for classification. Two embedding layers are introduced to encode label and language features independently, each processed by a dedicated multi-layer perceptron (MLP). Finally, the outputs are combined via cumulative multiplication to form the classification feature $f_{c}$ , defined as:", + "bbox": [ + 89, + 199, + 483, + 472 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {c} = \\operatorname {M L P} _ {1} \\left(\\mathrm {E} _ {1} \\left(T ^ {\\text {c a t}}\\right)\\right) \\odot \\operatorname {M L P} _ {2} \\left(\\mathrm {E} _ {2} \\left(\\operatorname {l a n g i d} \\left(T ^ {\\text {t i t}}\\right)\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 486, + 482, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{E}_1$ and $\\mathrm{E}_2$ represent two independent embedding layers, and $\\mathrm{MLP}_1$ and $\\mathrm{MLP}_2$ refer to multi-layer perceptrons applied to the label embedding $\\mathrm{E}_1(T^{cat})$ and the language embedding $\\mathrm{E}_2(\\text{langid}(T^{tit}))$ , respectively. The operator $\\odot$ denotes element-wise multiplication, which combines the outputs of the two MLPs to produce the final classification feature $f_c$ .", + "bbox": [ + 89, + 521, + 483, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.5. Feature Fusion", + "text_level": 1, + "bbox": [ + 89, + 638, + 243, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After extraction, features from different modalities are aligned in dimension through multi-layer perceptrons (MLPs) and then concatenated into fused features $F$ .", + "bbox": [ + 89, + 659, + 483, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Temporal Popularity Regression", + "text_level": 1, + "bbox": [ + 89, + 720, + 375, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Sequential Encoding and Adjustment", + "text_level": 1, + "bbox": [ + 89, + 744, + 397, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the high correlation between popularity values on adjacent days in the ground truth, we employ a LSTM [21] structure to get the temporal variations in popularity. As shown in Figure 5, multiple LSTM cells are constructed to transmit temporal information. The initial states (both hidden state and cell state) of the LSTM are generated by passing the $F$ through the same MLP as state encoding. Additionally, the input for LSTM cell at each time step is constructed from $F$ through temporal encoding MLPs. These", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "two encoding process can be written as:", + "bbox": [ + 513, + 92, + 779, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {0} = c _ {0} = \\mathrm {M L P} ^ {\\mathrm {h c}} (F) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 111, + 903, + 128 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {s} = \\operatorname {M L P} _ {s} ^ {\\mathrm {X}} (F) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 131, + 903, + 147 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this formulation, $h_0$ and $c_0$ denote the initial hidden state and cell state, while $x_s$ represents the input at time step $s$ . The $\\mathrm{MLP}^{\\mathrm{hc}}$ module encodes the state information, and $\\mathrm{MLP}_s^{\\mathrm{x}}$ handles the temporal encoding at the $s$ -th time step, effectively capturing the sequence dynamics.", + "bbox": [ + 511, + 152, + 903, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each time step, both states are treated as the $s$ -th step's features for output. After concatenating and processing through independent MLPs, each LSTM cell produces outputs, helping capture temporal popularity via backpropagation during training. Before final predictions, we apply a non-negative adjustment since popularity scores cannot be negative. This process is formulated as:", + "bbox": [ + 511, + 229, + 905, + 335 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p r e} _ {s} = \\max \\left(0, \\mathrm {M L P} _ {s} ^ {\\text {o u t}} \\left(\\operatorname {C o n c a t} \\left(h _ {s}, c _ {s}\\right)\\right)\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 340, + 903, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $pre_{s}, h_{s}, c_{s}$ , and $\\mathrm{MLP}_s^{\\mathrm{out}}$ denote the predicted value, hidden state, cell state, and output MLP at time step $s$ , respectively. The operation Concat refers to the process of horizontally concatenating the vectors. After these steps, the final temporal popularity predictions are produced.", + "bbox": [ + 511, + 362, + 905, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Loss Function", + "text_level": 1, + "bbox": [ + 513, + 444, + 658, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The core component of our approach is the Composite Gradient Loss (CGL), specifically designed for this task. This custom loss function consists of several components: SmoothL1Loss (SL) between the model's outputs and targets, the first-order and second-order derivative differences between the outputs and targets, the L1 loss between the onehot encodings of the predicted and ground truth peaks, and the Laplacian remainder (LR). These components are combined with a weight ratio of 1:1:1:1e-6. The overall loss function $\\mathcal{L}$ is formulated as:", + "bbox": [ + 511, + 460, + 905, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathrm {S L} \\left(\\hat {P} _ {d, i}, P _ {d, i}\\right) + \\lambda_ {1} \\cdot \\mathrm {S L} \\left(\\hat {P} _ {d, i} ^ {(1)}, P _ {d, i} ^ {(1)}\\right) + \\lambda_ {2} \\cdot \\mathrm {S L} \\left(\\hat {P} _ {d, i} ^ {(2)}, P _ {d, i} ^ {(2)}\\right) \\tag {13} \\\\ + \\alpha \\cdot \\sum_ {i = 1} ^ {n} \\left| \\delta \\operatorname {a r g m a x} _ {d} \\left(\\hat {P} _ {d, i}\\right) - \\delta \\operatorname {a r g m a x} _ {d} \\left(P _ {d, i}\\right) \\right| + \\epsilon \\cdot \\mathrm {L R} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 603, + 903, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "And the Laplacian remainder is:", + "bbox": [ + 513, + 688, + 707, + 702 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L R} = \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} ^ {(1)} \\right| + \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} ^ {(2)} \\right| \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 707, + 903, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The term $\\mathrm{SL}(\\hat{P}_{d,i}, P_{d,i})$ (with $\\beta = 0.1$ ) is used to compute the error between the predicted outputs and the ground truth targets. Here, $\\hat{P}_{d,i}$ represents the predicted popularity for data point $i$ on day $d$ , while $P_{d,i}$ denotes the ground truth popularity for the same data point. The differences between the one-hot encoded predicted and true peak values are measured using $\\delta \\mathrm{argmax}_d(\\hat{P}_{d,i})$ and $\\delta \\mathrm{argmax}_d(P_{d,i})$ , respectively. Additionally, the term $\\epsilon \\cdot \\mathrm{LR}$ introduces a Laplacian remainder (LR) to provide further regularization. During training, the weights for the first-order and second-order", + "bbox": [ + 511, + 748, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18852", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "derivative terms, as well as the L1 loss weight between the one-hot encodings $(\\lambda_1,\\lambda_2,\\alpha)$ , are adjusted dynamically using a cosine annealing algorithm to ensure smoother convergence throughout the training process.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 165, + 222, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we first describe the experiment settings and evaluation metrics for training models and comparison. Then various experimental results and detailed discussions are provided including evaluations for SMTPD dataset, proposed baseline, multi-modal features and early popularity.", + "bbox": [ + 89, + 190, + 482, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experiment settings", + "text_level": 1, + "bbox": [ + 89, + 275, + 279, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement our approaches using the Pytorch* framework and train it with the Adam [25] optimizer incorporating L2 penalty of $10^{-3}$ , while the batch size is set to 64. The learning rate is initialized to $10^{-3}$ and adjusted by the ReduceLROnPlateau scheduler provided by PyTorch when one epoch is end. Supervised training was performed using the Composite Gradient Loss (CGL) metioned before.", + "bbox": [ + 89, + 297, + 483, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We introduce error-based metrics and correlation-based metrics to fairly evaluate the compared datasets and models. We introduce the Absolute Error (MAE) and average MAE (AMAE) to respectively evaluate models for single-day prediction and temporal prediction. Assuming that the daily MAE and average MAE are denoted as $MAE_{d}$ and $AMAE$ , they can be defined as:", + "bbox": [ + 89, + 402, + 483, + 507 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM A E _ {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} - P _ {d, i} \\right| \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 518, + 482, + 556 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA M A E = \\frac {1}{m} \\sum_ {d = 1} ^ {m} M A E _ {d} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 560, + 482, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $n$ denotes the total of samples and $m$ denotes total of days, while the $P_{d}$ and $\\hat{P}_d$ respectively denotes the ground-truth popularity and predicted popularity for the day $d$ .", + "bbox": [ + 89, + 608, + 483, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For correlation-based metrics, we introduce the daily SRC and the average SRC to evaluate the models for both single-day prediction and temporal prediction from a different perspective. Assuming $n$ , $m$ , $P_d$ , and $\\hat{P}_d$ are defined consistently as mentioned earlier, the daily SRC and average SRC can be formulated as follows:", + "bbox": [ + 89, + 654, + 483, + 744 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nS R C _ {d} = \\frac {1}{n - 1} \\sum_ {i = 1} ^ {n} \\left(\\frac {\\hat {P} _ {d , i} - \\bar {\\hat {P}} _ {d}}{\\sigma_ {\\hat {P} _ {d}}}\\right) \\left(\\frac {P _ {d , i} - \\bar {P} _ {d}}{\\sigma_ {P _ {d}}}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 753, + 482, + 796 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nA S R C = \\frac {1}{m} \\sum_ {d = 1} ^ {m} S R C _ {d} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 797, + 482, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\bar{P}_d$ and $\\sigma_{P_d}$ are mean and standard deviation of the corresponding popularity for the day $d$ .", + "bbox": [ + 89, + 847, + 482, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the curves, sequential performance is represented by the MAE and SRC metrics, which illustrate daily prediction errors and correlations for popularity scores. The X-axis denotes days, while the Y-axis shows $MAE_{d}$ or $SRC_{d}$ .", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. SMPD VS. SMPD", + "text_level": 1, + "bbox": [ + 511, + 162, + 700, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we conduct comparisons and discussions between the SMTPD and SMPD. We first measure three most recently proposed state-of-the-art popularity prediction methods on two datasets. As these methods are all designed basing on the settings of SMPD, we make some modifications to let them fit for training and testing on our SMTPD. Other top-performing models (e.g. [44, 46, 50]) include components that are not applicable to SMTPD or are not reproducible, such as sliding window average or undisclosed self-trained modules. Hence, we do not discuss them in our experiments. First, we distribute content from corresponding modalities into their respective feature extraction modules. Then we use BERT-Multilingual as the textual feature extractor to address the challenges of multilingual text. Given that our SMTPD dataset consists of temporal popularity scores, we restrict these methods to predict only the single popularity score on day 30. To mitigate random bias, we evaluate using 5-fold cross-validation.", + "bbox": [ + 511, + 185, + 906, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows minimal performance differences across folds, confirming that SMTPD contains abundant data under a well-balanced sample distribution. It is evident that on the single-output prediction of popularity on days 7, 14, and 30 within SMTPD, the performance of method [26] based on GBDT outperforms the deep learning method [13] and [52] in terms of both MAE and SRC through its powerful regression capabilities of Catboost [39]. Notably, our method achieves the best results with the addition of EP, where the MAE of the other three models increases as the prediction horizon extends. In contrast, our model's MAE begins to decrease after day 14. The deep learning method [13] also performs well on days 7 and 14, but its effectiveness declines beyond day 14.", + "bbox": [ + 511, + 458, + 908, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover there are also two interesting observations. First, all methods achieve higher SRC on SMPD than on SMPD, likely because SMPD lacks specific time points, making it challenging to predict time-dependent popularity trends accurately as popularity declines and their MAE performances are certainly reduced that the MAE values increase by over 0.16. Besides, the larger popularity range and standard deviation among SMPD samples (shown in Table 2) also contributes to prediction difficulties.", + "bbox": [ + 511, + 670, + 908, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluation for Proposed Baseline", + "text_level": 1, + "bbox": [ + 511, + 816, + 805, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Refer to Table 3, we attempted to validate the rationale of partial structures in the proposed baseline model by employing alternative feature extractors and regression networks. Using BERT-Base as the textual features extractor", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "*https://pytorch.org", + "bbox": [ + 114, + 887, + 220, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18853", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lead to an increase of MAE by around 0.19. This decline in performance may be caused by the limitations of BERT-base model in handling multilingual texts, as it struggles to effectively capture the semantics and contextual information across different languages. The MAE also increased when the MLP (3 layers) was used as the regression structure. Such a structure is similar to [13], suggesting that MLP lacks the capture of temporal information. The SRC did not change for either of the above substitutions, probably due to the high correlation given by EP.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Discussion of Early Popularity", + "text_level": 1, + "bbox": [ + 89, + 250, + 359, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the performance of proposed baseline on SMTPD. As shown in Table 5, the proposed baseline gets a great improvement from existed methods that surpassing the second best method by -0.798/0.109 MAE/SRC for predicting the popularity of day 30. This great improvement is mainly brought by the early popularity (EP). Without the assistant of EP (row 2 in Table 5), the baseline performance sharply reduced around the existed methods. Using the EP predicted by existed method (1.717/0.864 MAE/SRC) as an input also contributes to the baseline's performance (row 3 in Table 5). However the performance gap between the predicted EP and true EP remains large.", + "bbox": [ + 89, + 272, + 482, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instead of directly involving the EP in the model architecture, introducing EP in the training supervision is another suitable option. To validate the effectiveness of such operation, we train a baseline model for predicting the popularities from day 1 to 30. As it presented by the row 1 in Table 5, adding the 1st day's popularity to the prediction sequence make a slight boost, what is in line with the LSTM's ability to capture temporal dependencies in popularity sequences. Having more preceding temporal information from backpropagation leads to more precise predictions of popularity in subsequent time steps.", + "bbox": [ + 89, + 454, + 482, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These comparisons clearly indicate that EP plays a crucial role in popularity prediction, and accurately forecasting the first day's popularity is key to predicting future popularity. The strong correlation between EP and future popularity significantly benefits the prediction task. Our results, as demonstrated by the MAE and SRC curves for both natural EP and our model's predictions, underscore the model's capability to leverage EP for enhanced accuracy. From the", + "bbox": [ + 89, + 619, + 482, + 741 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/55b7f3491a871ba34e66b0f25264ccae7e4df2072fb9ecb15a981dec55fb0f72.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BERT-BaseBERT-MulMLPLSTMAMAEASRC
✓✓0.7820.958
✓✓0.7860.958
✓✓0.7170.959
", + "bbox": [ + 91, + 747, + 480, + 843 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/edc9c288f058627c41bedec20bc1fe44d9c8886e00949ba205a3677875ab357a.jpg", + "table_caption": [ + "Table 3. The evaluation for the proposed baseline model, mainly on evaluating the model's performance in adapting to multilingual and temporal popularity." + ], + "table_footnote": [], + "table_body": "
MethodSMTPD (day 7)SMTPD (day 14)SMTPD (day 30)
AverageAverageAverage
Ding et al. [13]1.715/0.8491.669/0.8461.592/0.843
w. EP0.715/0.9640.742/0.9590.749/0.931
Lai et al. [26]1.573/0.8751.524/0.8721.495/0.864
w. EP0.725/0.9570.753/0.9620.760/0.957
Xu et al. [52]1.895/0.8171.832/0.8181.743/0.820
w. EP0.754/0.9620.798/0.9560.822/0.949
Ours w/o. EP1.673/0.8521.628/0.8501.563/0.848
Ours0.713/0.9640.735/0.9590.732/0.959
", + "bbox": [ + 514, + 89, + 903, + 252 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/75f27d9d206918ee23b9d54b3d8b86b9b98a4f8c4b9da957c9b62f2973057d98.jpg", + "table_caption": [ + "Table 4. The performance (MAE/SRC) was compared across four models, including our model, using the SMTPD dataset, both with and without EP." + ], + "table_footnote": [], + "table_body": "
MethodAMAEASRCMAE(day 30 only)SRC(day 30 only)
w/o. EP(1-30)1.5620.8561.5300.850
w/o. EP(2-30)1.6300.8491.5510.849
w/o. EP+[26]1.6280.8511.5550.848
ours0.7170.9590.7320.959
", + "bbox": [ + 514, + 319, + 903, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Assessment of EP across different scenarios. Here, \"1-30\" denotes the prediction target spanning a continuous sequence from the 1st day to the 30th day, as does \"2-30\" (to align with methods having EP).", + "bbox": [ + 511, + 468, + 903, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "visualized results, tough EP exhibits high correlation to the popularities of following days, the MAE sharply increased over time. By contrast, our baseline model could well optimize the MAE and achieve even better SRC performance comparing to the natural EP curve. Therefore, the proposed baseline model is effective to utilize EP achieving better prediction accuracy.", + "bbox": [ + 511, + 536, + 905, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 655, + 633, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study aims to address the challenge of time-series popularity prediction in social media. In this paper, we introduce a novel multilingual, multi-modal time-series popularity dataset based on YouTube and suggest a multilingual temporal prediction model tailored to this dataset. Through experiments, we demonstrate the effectiveness of this approach in predicting social media popularity time-series in a multilingual environment. The experiments show that combining multi-modal features with early popularity significantly improves prediction accuracy. However, this study has yet to address challenges such as multi-frame information in videos, maximizing the use of language diversity, and deeper multi-modal exploration. These areas will be key focuses for our future work.", + "bbox": [ + 511, + 681, + 906, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18854", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Deepak Agarwal, Bee-Chung Chen, Pradheep Elango, Nitin Motgi, Seung-Taek Park, Raghu Ramakrishnan, Scott Roy, and Joe Zachariah. Online models for content optimization. Advances in Neural Information Processing Systems, 21, 2008. 1", + "[2] Sercan Ö Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, pages 6679-6687, 2021. 2", + "[3] Peng Bao, Hua-Wei Shen, Junming Huang, and Xue-Qi Cheng. Popularity prediction in microblogging network: a case study on sina weibo. In Proceedings of the 22nd international conference on world wide web, pages 177–178, 2013. 1, 2", + "[4] Shenghua Bao, Guirong Xue, Xiaoyuan Wu, Yong Yu, Ben Fei, and Zhong Su. Optimizing web search using social annotations. In Proceedings of the 16th international conference on World Wide Web, pages 501-510, 2007. 1", + "[5] Laura Ceci. Youtube-statistics & facts. Statista. com [online].[cit. 2022-05-05]. Dostupné z: https://www.statista.com/topics/2019/youtube, 2022.3", + "[6] Junhong Chen, Dayong Liang, Zhanmo Zhu, Xiaojing Zhou, Zihan Ye, and Xiuyun Mo. Social media popularity prediction based on visual-textual features with xgboost. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2692-2696, 2019. 2", + "[7] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm SIGkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 2", + "[8] Weilong Chen, Chenghao Huang, Weimin Yuan, Xiaolu Chen, Wenhao Hu, Xinran Zhang, and Yanru Zhang. Title-and-tag contrastive vision-and-language transformer for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7008-7012, 2022. 2", + "[9] Xiaolu Chen, Weilong Chen, Chenghao Huang, Zhongjian Zhang, Lixin Duan, and Yanru Zhang. Double-fine-tuning multi-objective vision-and-language transformer for social media popularity prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9462-9466, 2023. 2", + "[10] Wen-Huang Cheng, Sijie Song, Chieh-Yun Chen, Shintami Chusnul Hidayati, and Jiaying Liu. Fashion meets computer vision: A survey. ACM Computing Surveys (CSUR), 54 (4):1-41, 2021. 1", + "[11] Minhwa Cho, Dahye Jeong, and Eunil Park. Amps: Predicting popularity of short-form videos using multi-modal attention mechanisms in social media marketing environments. Journal of Retailing and Consumer Services, 78: 103778, 2024. 2", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 4171–4186, 2019. 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Keyan Ding, Ronggang Wang, and Shiqi Wang. Social media popularity prediction: A multiple feature fusion approach with deep neural networks. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2682-2686, 2019. 2, 7, 8", + "[14] Zhabiz Gharibshah and Xingquan Zhu. User response prediction in online advertising. aCM Computing Surveys (CSUR), 54(3):1-43, 2021. 1", + "[15] Anindya Ghose and Sha Yang. An empirical analysis of search engine advertising: Sponsored search in electronic markets. Management science, 55(10):1605-1622, 2009. 1", + "[16] Sayan Ghosh, Dylan Baker, David Jurgens, and Vinodkumar Prabhakaran. Detecting cross-geographic biases in toxicity modeling on social media. In Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT), pages 313–328, 2021. 4", + "[17] Marcos Andre Gonçalves, Jussara M Almeida, Luiz GP dos Santos, Alberto HF Laender, and Virgilio Almeida. On popularity in the blogosphere. IEEE Internet Computing, 14(3): 42-49, 2010. 1", + "[18] Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. Learning word vectors for 157 languages. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), 2018. 5", + "[19] Mohammad Hajarian, Mark Anthony Camilleri, Paloma Díaz, and Ignacio Aedo. A taxonomy of online marketing methods. In Strategic corporate communication in the digital age, pages 235-250. Emerald Publishing Limited, 2021. 1", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 6", + "[22] Chih-Chung Hsu, Chia-Ming Lee, Xiu-Yu Hou, and Chi-Han Tsai. Gradient boost tree network based on extensive feature analysis for popularity prediction of social posts. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9451-9455, 2023. 2", + "[23] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems, 30, 2017. 2", + "[24] Aditya Khosla, Atish Das Sarma, and Raffay Hamid. What makes an image popular? In Proceedings of the 23rd international conference on World wide web, pages 867-876, 2014. 3", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations(ICLR), San Diego, California, USA, May 7-9, 2015. 7", + "[26] Xin Lai, Yihong Zhang, and Wei Zhang. Hyfea: Winning solution to social media popularity prediction for multimedia grand challenge 2020. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4565-4569, 2020. 2, 7, 8" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18855", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Chenyu Li, Jun Liu, and Shuxin Ouyang. Analysis and prediction of content popularity for online video service: a youku case study. China Communications, 13(12):216-233, 2016. 3", + "[28] Jinning Li, Yirui Gao, Xiaofeng Gao, Yan Shi, and Guihai Chen. Senti2pop: sentiment-aware topic popularity prediction on social media. In 2019 IEEE International conference on data mining (ICDM), pages 1174-1179. IEEE, 2019. 2", + "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2", + "[30] Yury Lifshits. Ediscope: Social analytics for online news. Yahoo Labs, 2010. 1", + "[31] Hung-Hsiang Lin, Jiun-Da Lin, Jose Jaena Mari Ople, Jun-Cheng Chen, and Kai-Lung Hua. Social media popularity prediction based on multi-modal self-attention mechanisms. IEEE Access, 10:4448-4455, 2022. 2", + "[32] An-An Liu, Xiaowen Wang, Ning Xu, Junbo Guo, Guoqing Jin, Quan Zhang, Yejun Tang, and Shenyuan Zhang. A review of feature fusion-based media popularity prediction methods. Visual Informatics, 2022. 2", + "[33] Marco Lui and Timothy Baldwin. langid.py: An off-the-shelf language identification tool. In Proceedings of the ACL 2012 System Demonstrations, pages 25-30, 2012. 6", + "[34] Shijian Mao, Wudong Xi, Lei Yu, Gaotian Lü, Xingxing Xing, Xingchen Zhou, and Wei Wan. Enhanced catboost with stacking features for social media prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9430-9435, 2023. 2", + "[35] Masoud Mazloom, Robert Rietveld, Stevan Rudinac, Marcel Worring, and Willemijn Van Dolen. Multimodal popularity prediction of brand-related social media posts. In Proceedings of the 24th ACM international conference on Multimedia, pages 197-201, 2016. 2", + "[36] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proceedings of the 1st International Conference on Learning Representations (ICLR), Scottsdale, Arizona, USA, May 2-4, 2013. 5", + "[37] Alessandro Ortiz, Giovanni Maria Farinella, and Sebastiano Battiato. Prediction of social image popularity dynamics. In Image Analysis and Processing-ICIAP 2019: 20th International Conference, Trento, Italy, September 9–13, 2019, Proceedings, Part II 20, pages 572–582. Springer, 2019. 3", + "[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 5", + "[39] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. Advances in neural information processing systems, 31, 2018. 7", + "[40] Cédric Richier, Eitan Altman, Rachid Elazouzi, Tania Altman, Georges Linares, and Yonathan Portilla. Modelling" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "view-count dynamics in youtube. CoRR, abs/1404.2570, 2014. 2", + "[41] Cedric Richier, Eitan Altman, Rachid Elazouzi, Tania Jimenez, Georges Linares, and Yonathan Portilla. Bioinspired models for characterizing youtube viewcout. In 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 297-305. IEEE, 2014. 1, 2", + "[42] Satoshi Sanjo and Marie Katsurai. Recipe popularity prediction with deep visual-semantic fusion. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2279–2282, 2017. 1, 2", + "[43] Gabor Szabo and Bernardo A Huberman. Predicting the popularity of online content. Communications of the ACM, 53 (8):80-88, 2010. 1, 2", + "[44] YunPeng Tan, Fangyu Liu, BoWei Li, Zheng Zhang, and Bo Zhang. An efficient multi-view multimodal data processing framework for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7200-7204, 2022. 2, 7", + "[45] Alexandru Tatar, Marcelo Dias De Amorim, Serge Fdida, and Panayotis Antoniadis. A survey on predicting the popularity of web content. Journal of Internet Services and Applications, 5(1):1-20, 2014. 1", + "[46] Kai Wang, Penghui Wang, Xin Chen, Qiushi Huang, Zhendong Mao, and Yongdong Zhang. A feature generalization framework for social media popularity prediction. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4570-4574, 2020. 2, 7", + "[47] Bo Wu, Wen-Huang Cheng, Yongdong Zhang, Qiushi Huang, Jintao Li, and Tao Mei. Sequential prediction of social media popularity with deep temporal context networks. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI), pages 3062-3068, 2017. 1, 2", + "[48] Bo Wu, Wen-Huang Cheng, Peiye Liu, Bei Liu, Zhaoyang Zeng, and Jiebo Luo. Smp challenge: An overview of social media prediction challenge 2019. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2667-2671. ACM, 2019. 1, 2, 3", + "[49] Bo Wu, Peiye Liu, Wen-Huang Cheng, Bei Liu, Zhaoyang Zeng, Jia Wang, Qiushi Huang, and Jiebo Luo. Smp challenge: An overview and analysis of social media prediction challenge. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9651-9655, 2023. 2", + "[50] Jianmin Wu, Liming Zhao, Dangwei Li, Chen-Wei Xie, Siyang Sun, and Yun Zheng. Deeply exploit visual and language information for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7045–7049, 2022. 2, 7", + "[51] Leilei Wu, Lingling Yi, Xiao-Long Ren, and Linyuan Lu. Predicting the popularity of information on social platforms without underlying network structure. Entropy, 25(6):916, 2023. 2", + "[52] Kele Xu, Zhimin Lin, Jianqiao Zhao, Peicang Shi, Wei Deng, and Huaimin Wang. Multimodal deep learning for social media popularity prediction with attention mechanism. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4580-4584, 2020. 2, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18856", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jaewon Yang and Jure Leskovec. Patterns of temporal variation in online media. In Proceedings of the fourth ACM international conference on Web search and data mining, pages 177-186, 2011. 1, 2", + "[54] Zhizhen Zhang, Xiaohui Xie, Mengyu Yang, Ye Tian, Yong Jiang, and Yong Cui. Improving social media popularity prediction with multiple post dependencies. CoRR, abs/2307.15413, 2023. 2", + "[55] Alireza Zohourian, Hedieh Sajedi, and Arefeh Yavary. Popularity prediction of images and videos on instagram. In 2018 4th International Conference on Web Research (ICWR), pages 111-117, 2018. 3" + ], + "bbox": [ + 91, + 92, + 480, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18857", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_model.json b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94bf10b8224e3e3d44b907150ad99242bc26d63b --- /dev/null +++ b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.905, + 0.155 + ], + "angle": 0, + "content": "SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.18, + 0.771, + 0.237 + ], + "angle": 0, + "content": "\\(^{1}\\)Yijie Xu*, \\(^{1}\\)Bolun Zheng*, \\(^{1}\\)Wei Zhu*, \\(^{1}\\)Hangjia Pan, \\(^{1}\\)Yuchen Yao, \\(^{2}\\)Ning Xu, \\(^{2}\\)Anan Liu, \\(^{3}\\)Quan Zhang, \\(^{1}\\)Chenggang Yan \\(^{1}\\)Hangzhou Dianzi University, \\(^{2}\\)Tianjin University, \\(^{3}\\)Peking University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.486, + 0.574 + ], + "angle": 0, + "content": "Social media popularity prediction task aims to predict the popularity of posts on social media platforms, which has a positive driving effect on application scenarios such as content optimization, digital marketing and online advertising. Though many studies have made significant progress, few of them pay much attention to the integration between popularity prediction with temporal alignment. In this paper, with exploring YouTube's multilingual and multi-modal content, we construct a new social media temporal popularity prediction benchmark, namely SMTPD, and suggest a baseline framework for temporal popularity prediction. Through data analysis and experiments, we verify that temporal alignment and early popularity play crucial roles in social media popularity prediction for not only deepening the understanding of temporal dynamics of popularity in social media but also offering a suggestion about developing more effective prediction models in this field. Code is available at https://github.com/zhuwei321/SMTPD" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.222, + 0.6 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.762 + ], + "angle": 0, + "content": "With the advancement of Internet communication technology in recent years, social media has gradually emerged and has influenced various aspects of human life. Any content posted on social media stands the chance of becoming hot spot, and widely disseminated social media content can generate significant social and economic benefits. The prediction of social media popularity holds immense potential applications in content optimization [1, 30], online advertising [14, 15], digital marketing [19, 45], search recommendations [4, 17], intelligent fashion [10], and beyond." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.839 + ], + "angle": 0, + "content": "In the early stages of popularity research, statistical and topological methods [41, 43, 53] were extensively employed, with the research focus primarily on time-aware popularity prediction. These prediction methods rely on information about the earlier popularity of posted content." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.268, + 0.912, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Figure 1. Content sections and popularity trend of SMTPD. In 1a, a sample involves four sections of content, with temporal popularity. 1b depicts box plots of daily popularity scores, illustrating variations in popularity distribution at different time points. The distribution consistently demonstrates a decay pattern over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.907, + 0.845 + ], + "angle": 0, + "content": "Along with the many large-scale social media popularity datasets are proposed [3, 42, 47, 48], the machine learning based methods are widely employed to achieve reasonable predictions of social media popularity and have shown remarkable performance. These methods emphasize meticulous design in feature extraction and fusion, followed by popularity prediction using machine learning models such as deep neural networks (DNN) and gradient boosting decision trees (GBDT)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Table 1 lists the public or mentioned social media popularity datasets in the most recent years as far as we know. For above machine learning based methods, the large-scale" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.851, + 0.371, + 0.864 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.864, + 0.243, + 0.876 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.876, + 0.483, + 0.9 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China under Grant 2020YFB1406600." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18847" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.204 + ], + "angle": 0, + "content": "
DatasetSourceCategorySamplesLanguagePrediction Type
Mazloom [35]Instagramfast food brand75KEnglishsingle
Sanjo [42]Cookpadrecipe150KJapanesesingle
TPIC17 [47]Flickr-680KEnglishsingle
SMPD [48]Flickr11 categories486KEnglishsingle
AMPS [11]YouTubeshorts13KKoreansingle
SMTPD (ours)YouTube15 categories282Kover 90 languagessequential
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Table 1. Comparing SMTPD with existed multi-modal social media popularity datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.287 + ], + "angle": 0, + "content": "social media popularity datasets [3, 11, 35, 42, 47, 48] provide the fundamental supports." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.395 + ], + "angle": 0, + "content": "However, existing datasets and related studies have notable limitations, such as insufficient multi-modal data, limited language diversity, and, crucially, the absence of a consistent timeline for prediction. As shown in Figure1b, popularity distributions shift over time, and predictions made at irregular intervals can reduce accuracy and create challenges for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Furthermore, effective popularity prediction requires integrating both the multi-modal content of social media and information on the communication process to capture the dynamic trends in popularity. This essential aspect is missing in most current prediction efforts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Targeting to fix up the above shortcomings of existed datasets, in this paper, we propose a new benchmark called Social Media Temporal Popularity Dataset (SMTPD) by observing multi-modal content from mainstream for cutting-edge research in the field of social multimedia along with multi-modal feature temporal prediction. As shown in Figure 1a, multi-modal contents of social media posts in multiple languages generated the daily popularity information during the communication process. We define such a post as a sample of SMTPD, and propose a multi-modal framework as a baseline to achieve the temporal prediction of popularity scores. The proposed framework consists of two parts, feature extraction and regression. In the feature extraction part, multiple pre-training models and pre-processing methods are introduced to translate the multi-modal content into deep features. In the regression part, we encode the state and time sequence of the extracted features, and adopt the LSTM-based structure to regress the 30-day popularities. Through analysis and experiments, we discover the importance of early popularity to the task of popularity prediction, and demonstrate the effectiveness of our method in temporal prediction. Generally, the contribution of this work can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Against the missing of temporal information in social media popularity researches, we observe over 282K multilingual samples from mainstream social media since they released on the network lasting for 30 days. We refer to these samples as SMTPD, a new benchmark for temporal" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.093, + 0.674, + 0.107 + ], + "angle": 0, + "content": "popularity prediction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.196 + ], + "angle": 0, + "content": "- Basing on existed methods, we innovate in both the selection of feature extractors and the construction of the temporal regression component, and suggest a baseline model which enables temporal popularity prediction to be conducted across multiple languages while aligning prediction times." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.274 + ], + "angle": 0, + "content": "- Exploring the popularity distribution and the correlation between popularity at different times. Based on these, We find the importance of early popularity for popularity prediction task, and point out that the key-point for predicting popularity is to accurately predict the early popularity." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.108, + 0.905, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.287, + 0.651, + 0.303 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.448 + ], + "angle": 0, + "content": "Statistical and topological methods. Szabo et al. [43] found that the early-stage popularity notably influences subsequent popularity. Yang et al. [53] conducted cluster analysis on temporal patterns within online content, revealing distinct characteristics of popularity variations across different clusters. Richier et al. [40, 41] proposes bio-inspired models to characterize the evolution of video view counts. Wu et al. [51] suggested that the information cascade process is best described as an activate-decay dynamic process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.905, + 0.75 + ], + "angle": 0, + "content": "Multi-modal feature based Methods. Ding et al. [13] use pre-trained ResNet and BERT to extract visual and textual features, with the DNN regression. Xu et al. [52] and Lin et al. [31] adopted attention mechanisms to effectively integrate multi-modal features. Chen et al. [6] compared the performance of several regression models, among which XGBoost [7] exhibited the best results. Hsu et al. [22] employed LightGBM [23] and TabNet [2] to capture intricate semantic relationships in multi-modal features. Lai et al. [26] engineered handcrafted features, exploiting CatBoost for regression. Mao et al. [34] enhanced CatBoost-based model by stacking features. Tan et al. [44] extracted visual-textual features by ALBEF [29] and Chen et al. [8, 9] enriched more intermodal features to promote predictive performance. Wu et al. [50] emphasized increasing feature dimensions to improving predictive performance. The post dependencies captured by sliding window average [46] and DSN [54] has also led to improvements. Many of these multi-modal methods are based on the SMPD and working out great [32, 49]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.886 + ], + "angle": 0, + "content": "Social media popularity datasets. Mazloom et al. [35] conducted experiments using a dataset of posts related to fast-food brands collected from Instagram. Sanjo et al. [42] provided a recipe popularity prediction dataset based on Cookpad, which includes text content entirely in Japanese. Li et al. [28] predicted the future popularity of topics by using historical sentiment information based on Twitter's text data records. TPIC17 [47] and SMPD [48] are datasets for single popularity prediction task based on Flickr." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We believe that the loss of temporal alignment about" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18848" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "popularity in existing methods is a common problem. The image dynamic popularity dataset [37] gave us insights, so we propose a multi-modal temporal popularity benchmark to address the shortcomings of existing studies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.165, + 0.253, + 0.181 + ], + "angle": 0, + "content": "3. SMTPD Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.19, + 0.484, + 0.508 + ], + "angle": 0, + "content": "With the great development of web communication technologies, social multimedia has become the most popular media around the world. However, most existing social media prediction datasets are built on the basis of single-output prediction, meaning they observe existing posts and infer their popularity based on their posting time. As mentioned earlier, these data without aligned posting times exhibit non-uniform popularity distributions. Therefore, SMTPD keep an eye on one of the most popular worldwide social media, YouTube [5], with over 282K samples, primarily focusing on the evolution of popularity over time for these samples. We collected over 402K raw data samples. Specifically, we first removed records with missing values—which likely resulted from network issues during data acquisition. Next, we eliminated samples that either failed to be crawled on certain days or had been deleted within 30 days. Finally, we filtered out potential outliers using the \\(3\\sigma\\) rule. In this section, we provide a comprehensive overview of the SMTPD's composition. Additionally, we present various data analysis results that contribute to the feasibility of our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Unlike the retrospective methods used in previous datasets, we take note of the newly posted multi-modal content and then observe popularity information every 24 hours via YouTube API. Considering that in real-world applications, the posts to be predicted are always new, this data observation method allows us to obtain aligned temporal popularity of new posts, which is more in line with practical applications. Figure 1a show a sample from SMTPD. We note diverse attributes, including visual content, textual content, metadata, and user profiles. Additionally, we focus on the temporal popularity for each sample within 30 days after release. We perform basic statistics for SMTPD, as shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.288, + 0.729 + ], + "angle": 0, + "content": "3.1. Temporal Popularity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "The utilization of temporal popularity data is versatile. It not only serves as the output for predictions but also as inputs, where a segment preceding a specific time point, coupled with multi-modal content, can forecast subsequent popularity. We analyze the distribution and the correlation of popularity over time, so as to facilitate a comprehensive understanding of the significance of temporal popularity for predictions from different perspectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A number of popularity definitions have been born from previous work [27, 55]. We use Khosla's popularity score transformation [24] as it can normalize the distribution of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "view counts to a suitable popularity score's distribution. It can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.129, + 0.907, + 0.155 + ], + "angle": 0, + "content": "\\[\np = \\log_ {2} (\\frac {v}{d} + 1) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.162, + 0.907, + 0.266 + ], + "angle": 0, + "content": "where \\( p \\) is the popularity score, \\( v \\) is the view counts of a sample, \\( d \\) represents the corresponding number of days after the post's release. As shown in Figure 1, the histogram of view counts reveals an extreme long-tailed distribution, which might not be suitable as a prediction target. It's evident that the distribution of the popularity score metric becomes more reasonable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.906, + 0.357 + ], + "angle": 0, + "content": "In addition, we notice that there is a strong correlation between the popularity scores of different days. The Pearson Correlation (PC) and Spearman Ranking Correlation (SRC) can respectively reflect the linear relationship and rank-order correlation between popularity scores of different days as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.377, + 0.907, + 0.422 + ], + "angle": 0, + "content": "\\[\nP C = \\frac {\\sum_ {i = 1} ^ {n} \\left(X _ {i} - \\bar {X}\\right) \\left(Y _ {i} - \\bar {Y}\\right)}{\\sqrt {\\sum_ {i = 1} ^ {n} \\left(X _ {i} - \\bar {X}\\right) ^ {2}} \\sqrt {\\sum_ {i = 1} ^ {n} \\left(Y _ {i} - \\bar {Y}\\right) ^ {2}}} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.43, + 0.906, + 0.468 + ], + "angle": 0, + "content": "\\[\nS R C = \\frac {1}{n - 1} \\sum_ {i = 1} ^ {n} \\left(\\frac {X _ {i} - \\bar {X}}{\\sigma_ {X}}\\right) \\left(\\frac {Y _ {i} - \\bar {Y}}{\\sigma_ {Y}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.906, + 0.547 + ], + "angle": 0, + "content": "where \\( X \\) and \\( Y \\) denote the popularity scores belonging to 2 different days, while \\( \\bar{X} \\) and \\( \\sigma_{X} \\) denote the mean and standard deviation of \\( X \\), and similarly for \\( Y \\). PC and SRC respectively reflect the linear relationship and rank-order correlation between popularity scores of different days." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.624 + ], + "angle": 0, + "content": "The Pearson Correlation (PC) and Spearman Rank Correlation (SRC) heat maps illustrate internal relationships in daily popularity scores, establishing a basis for prediction based on these temporal patterns. Further details are available in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.631, + 0.907, + 0.845 + ], + "angle": 0, + "content": "
StatisticsSMTPDSMPD (Train/Test)
Number of samples282.4K305K/181K
Number of users152.7K38K/31K
Mean popularity5.956.41/5.12
STD popularity4.152.47/2.41
Number of categories1511
Number of custom tags960K250K
Average length of title53.4 chars29 words
Mean duration1853.4 s-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.855, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Table 2. The basic statistics of all samples in SMPD, comparing to SMPD[48]. Due to the multilingual environment, we use chars (characters) to describe the average length of title." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18849" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.295, + 0.106 + ], + "angle": 0, + "content": "3.2. Multi-Modal Content" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.175 + ], + "angle": 0, + "content": "Multi-modal feature fusion is the mainstream approach in today's popularity prediction methods. In this section, we will divide the data into corresponding modality and elucidate the multi-modal content in SMTPD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.243, + 0.195 + ], + "angle": 0, + "content": "3.2.1. Visual Content" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.484, + 0.307 + ], + "angle": 0, + "content": "SMTPD samples, primarily drawn from a multi-modal social media platform, are heavily influenced by visual elements. Since metrics like view count and popularity score are often based on the cover frame that viewers see before clicking to play, we use the cover frame as the primary visual content. This approach balances popularity impact with copyright considerations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.314, + 0.25, + 0.328 + ], + "angle": 0, + "content": "3.2.2. Textual Content" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.456 + ], + "angle": 0, + "content": "The textual content in posts plays a crucial role in media dissemination and user engagement. We examine several core elements, including category, title, description, and hashtags (denoted by the “#” symbol), along with the user ID (user nickname) from profile metadata. YouTube defines 15 distinct video categories, all in English, which support content organization and discovery. Figure 3 shows the statistical distribution of content by category." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.467, + 0.252, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.592, + 0.239, + 0.604 + ], + "angle": 0, + "content": "(a) The PC Matrix" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.467, + 0.484, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.591, + 0.44, + 0.602 + ], + "angle": 0, + "content": "(b) The SRC Matrix" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.616, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Figure 2. The heat map of daily popularity correlations. The figure clearly shows that there is a high degree of correlations in popularity between consecutive days." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.738 + ], + "angle": 0, + "content": "As an international platform, YouTube hosts user-defined textual content in multiple languages beyond standard categories. Figure 4 illustrates the distribution and popularity bias across title languages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.484, + 0.845 + ], + "angle": 0, + "content": "In previous researches, the datasets typically had uniform language for text content, without considering multilingual aspects. While this simplifies modeling for prediction methods, it also poses significant limitations for international social media platforms which not restricted to a single language. Besides, predictions based on a few languages may introduce biases [16] against popularity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.258, + 0.866 + ], + "angle": 0, + "content": "3.2.3. Numerical values" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Numerical attributes frequently function as supplementary factors in enhancing prediction accuracy. In this study, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "examined key attributes, including video duration, uploader follower count, and the number of posts by the uploader. Additionally, we manually derived several auxiliary metrics, such as title length, the count of custom tags, and description length, to further enrich the feature set and support predictive robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.199, + 0.805, + 0.218 + ], + "angle": 0, + "content": "4. Temporal popularity prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.907, + 0.348 + ], + "angle": 0, + "content": "The multi-modal feature-based temporal prediction framework is shown in Figure 5, which is divided into two main parts: multi-modal feature extraction and temporal popularity score regression. This framework is designed to adapt to the multilingual temporal prediction tasks under time alignment. The prediction target is set as the popularity of samples, which corresponds to the temporal popularity within 30 days after the sample's release." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.8, + 0.374 + ], + "angle": 0, + "content": "4.1. Multi-Modal Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.444 + ], + "angle": 0, + "content": "For different modalities in SMTPD, we adopted distinct feature extraction methods. We additionally incorporated features from the categorical modality to investigate the impact of categorical features on popularity prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.669, + 0.467 + ], + "angle": 0, + "content": "4.1.1. Visual Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.52 + ], + "angle": 0, + "content": "The cover image of a social media content is a very important component which would help users to quickly understand what they would see if clicking to view this content." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.529, + 0.882, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.67, + 0.829, + 0.684 + ], + "angle": 0, + "content": "(a) The statistics of samples in each category" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.685, + 0.902, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.831, + 0.816, + 0.844 + ], + "angle": 0, + "content": "(b) Average popularity of each category" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.855, + 0.906, + 0.898 + ], + "angle": 0, + "content": "Figure 3. The statistics based on category. 3a counts the number of samples in each category, and 3b shows the average popularity score of samples in each category." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18850" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.482, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.485, + 0.343 + ], + "angle": 0, + "content": "Figure 4. Languages analysis. The left is the proportions of these languages, with \"Other\" encompassing 90 different languages. The right represents the average popularity in different languages, revealing the geographic biases. The languages of samples are counted by reference to the title." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.506 + ], + "angle": 0, + "content": "Therefore, the cover image would directly affect the popularity of a social media content. The semantic information provided in the cover image plays a key role for user's understanding. We adopt convolutional neural networks ResNet-101 [20] pre-trained on ImageNet as our visual feature extraction model to obtain the semantic information provided by the cover. To match the input size of ResNet-101, we re-scale the size of cover image to \\( 224 \\times 224 \\). The 2048-dimension feature vector before the final classification layer will be selected as the final visual feature \\( f_{v} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.516, + 0.482, + 0.532 + ], + "angle": 0, + "content": "\\[\nf _ {v} = \\operatorname {R e s N e t} (S (I)) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.619 + ], + "angle": 0, + "content": "where \\(I\\) denotes the cover image and \\(S\\) denotes the re-scale operation. However, the cover image would not always be uploaded by authors or be unavailable due to the network transmission. For those contents missing the cover image, we use blank images with all pixels are set to zero instead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.626, + 0.254, + 0.64 + ], + "angle": 0, + "content": "4.1.2. Textual Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Textual information significantly impacts a user's choice to view or skip content. Key textual inputs include category, title, tags, description, and user profile ID. Extracting semantic features from these inputs is crucial for accurate popularity prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.81 + ], + "angle": 0, + "content": "However, unlike previous popularity prediction tasks, one distinguishing characteristic of SMTPD is that its text content includes multiple languages, and many samples contain text in more than one language. This makes it challenging to use many pre-trained word vector models based on single-language corpora, such as [18, 36, 38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Thanks to the multilingual capability of BERT-Multilingual [12], which processes text across languages in a single model, we use it to extract a 768-dimensional feature vector for each text, capturing semantic information as:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nf _ {t} ^ {k} = \\operatorname {B E R T} \\left(T ^ {k}\\right) \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.093, + 0.872, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Figure 5. The proposed model consists of two layers. The upper layer processes visual, textual, numerical, and categorical features, which are input to a multi-layer perceptron (MLP). The MLP outputs are concatenated and fed into a lower LSTM layer for regression. Each LSTM cell's output is passed through another MLP, with the final MLP outputs combined to generate a time prediction sequence using the ReLU activation function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.552 + ], + "angle": 0, + "content": "where \\( k \\in \\{cat, tit, tag, des, uid\\} \\) maps to category, title, tags, description, and user ID. These are combined into a \\( 768 \\times 5 \\) matrix, transformed through \\( 5 \\times 5 \\) convolution to produce the 768D textual feature \\( f_{t} \\), which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.559, + 0.907, + 0.578 + ], + "angle": 0, + "content": "\\[\nf _ {t} = \\operatorname {C o n v} _ {5 \\times 5} \\left(\\mathcal {C} \\left(f _ {t} ^ {\\text {c a t}}, f _ {t} ^ {\\text {t i t}}, f _ {t} ^ {\\text {t a g}}, f _ {t} ^ {\\text {d e s}}, f _ {t} ^ {\\text {u i d}}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.824, + 0.6 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) denotes the above combing operation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.698, + 0.619 + ], + "angle": 0, + "content": "4.1.3. Numerical Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.805 + ], + "angle": 0, + "content": "Numerical metrics directly influence whether an audience clicks to watch which sample on social media platform. Also, the number of followers and the number of posts also impact the process through which an audience navigates from content uploaders' user pages to click and watch videos. Hence, we selected these influential numerical metrics for numerical features extraction. As the number of user followers share a similar distribution pattern that exhibiting long-tailed characteristics as view counts, it is necessary to apply a logarithmic transformation before feeding it into the network. Afterward, all numerical items are Z-score normalized and concatenated as numerical features:" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.81, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\nf _ {n} ^ {j} = \\frac {T ^ {j} - \\mu_ {j}}{\\sigma_ {j}} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.845, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\nf _ {n} = \\operatorname {C o n c a t} \\left(f _ {n} ^ {f o l}, f _ {n} ^ {p o s}, f _ {n} ^ {d u r}, f _ {n} ^ {t l}, f _ {n} ^ {t n}, f _ {n} ^ {d l}, f _ {n} ^ {E P}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( j \\in \\{fol, pos, dur, tl, tn, dl, EP\\} \\) represents the log-scaled follower counts, total number of posts, video du" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18851" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "ration, title length, the number of tags, description length and the early popularity (the 1st day's popularity). The \\(\\mu\\) and \\(\\sigma\\) represents the mean and standard deviation of all samples. And Concat denotes the horizontal concatenating operation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.18, + 0.283, + 0.195 + ], + "angle": 0, + "content": "4.1.4. Categorical Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.485, + 0.473 + ], + "angle": 0, + "content": "The categorical disparities across social media platforms are depicted in Figure 3b. Beyond text-based feature extraction, these disparities inform the generation of categorical features. Language functions as a key marker of regional and cultural distinctions among content creators, as highlighted by the language bias observed in Figure 4. Consequently, language classification features are crucial for enhancing prediction accuracy. To develop these classification features, two sub-features are selected: label and language attributes. The label feature captures the semantic category, while the language attribute is derived from the language in the video title. Assuming that the primary language of all text content matches the title's language, we use langid [33] for classification. Two embedding layers are introduced to encode label and language features independently, each processed by a dedicated multi-layer perceptron (MLP). Finally, the outputs are combined via cumulative multiplication to form the classification feature \\( f_{c} \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.487, + 0.483, + 0.505 + ], + "angle": 0, + "content": "\\[\nf _ {c} = \\operatorname {M L P} _ {1} \\left(\\mathrm {E} _ {1} \\left(T ^ {\\text {c a t}}\\right)\\right) \\odot \\operatorname {M L P} _ {2} \\left(\\mathrm {E} _ {2} \\left(\\operatorname {l a n g i d} \\left(T ^ {\\text {t i t}}\\right)\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathrm{E}_1\\) and \\(\\mathrm{E}_2\\) represent two independent embedding layers, and \\(\\mathrm{MLP}_1\\) and \\(\\mathrm{MLP}_2\\) refer to multi-layer perceptrons applied to the label embedding \\(\\mathrm{E}_1(T^{cat})\\) and the language embedding \\(\\mathrm{E}_2(\\text{langid}(T^{tit}))\\), respectively. The operator \\(\\odot\\) denotes element-wise multiplication, which combines the outputs of the two MLPs to produce the final classification feature \\(f_c\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.245, + 0.654 + ], + "angle": 0, + "content": "4.1.5. Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.707 + ], + "angle": 0, + "content": "After extraction, features from different modalities are aligned in dimension through multi-layer perceptrons (MLPs) and then concatenated into fused features \\( F \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.721, + 0.376, + 0.737 + ], + "angle": 0, + "content": "4.2. Temporal Popularity Regression" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.745, + 0.398, + 0.76 + ], + "angle": 0, + "content": "4.2.1. Sequential Encoding and Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Considering the high correlation between popularity values on adjacent days in the ground truth, we employ a LSTM [21] structure to get the temporal variations in popularity. As shown in Figure 5, multiple LSTM cells are constructed to transmit temporal information. The initial states (both hidden state and cell state) of the LSTM are generated by passing the \\( F \\) through the same MLP as state encoding. Additionally, the input for LSTM cell at each time step is constructed from \\( F \\) through temporal encoding MLPs. These" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.78, + 0.107 + ], + "angle": 0, + "content": "two encoding process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.112, + 0.905, + 0.129 + ], + "angle": 0, + "content": "\\[\nh _ {0} = c _ {0} = \\mathrm {M L P} ^ {\\mathrm {h c}} (F) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.132, + 0.905, + 0.148 + ], + "angle": 0, + "content": "\\[\nx _ {s} = \\operatorname {M L P} _ {s} ^ {\\mathrm {X}} (F) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.154, + 0.905, + 0.229 + ], + "angle": 0, + "content": "In this formulation, \\( h_0 \\) and \\( c_0 \\) denote the initial hidden state and cell state, while \\( x_s \\) represents the input at time step \\( s \\). The \\( \\mathrm{MLP}^{\\mathrm{hc}} \\) module encodes the state information, and \\( \\mathrm{MLP}_s^{\\mathrm{x}} \\) handles the temporal encoding at the \\( s \\)-th time step, effectively capturing the sequence dynamics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.906, + 0.336 + ], + "angle": 0, + "content": "At each time step, both states are treated as the \\(s\\)-th step's features for output. After concatenating and processing through independent MLPs, each LSTM cell produces outputs, helping capture temporal popularity via backpropagation during training. Before final predictions, we apply a non-negative adjustment since popularity scores cannot be negative. This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.341, + 0.905, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p r e} _ {s} = \\max \\left(0, \\mathrm {M L P} _ {s} ^ {\\text {o u t}} \\left(\\operatorname {C o n c a t} \\left(h _ {s}, c _ {s}\\right)\\right)\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.44 + ], + "angle": 0, + "content": "Where \\( pre_{s}, h_{s}, c_{s} \\), and \\( \\mathrm{MLP}_s^{\\mathrm{out}} \\) denote the predicted value, hidden state, cell state, and output MLP at time step \\( s \\), respectively. The operation Concat refers to the process of horizontally concatenating the vectors. After these steps, the final temporal popularity predictions are produced." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.445, + 0.659, + 0.457 + ], + "angle": 0, + "content": "4.2.2. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.601 + ], + "angle": 0, + "content": "The core component of our approach is the Composite Gradient Loss (CGL), specifically designed for this task. This custom loss function consists of several components: SmoothL1Loss (SL) between the model's outputs and targets, the first-order and second-order derivative differences between the outputs and targets, the L1 loss between the onehot encodings of the predicted and ground truth peaks, and the Laplacian remainder (LR). These components are combined with a weight ratio of 1:1:1:1e-6. The overall loss function \\(\\mathcal{L}\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.604, + 0.905, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathrm {S L} \\left(\\hat {P} _ {d, i}, P _ {d, i}\\right) + \\lambda_ {1} \\cdot \\mathrm {S L} \\left(\\hat {P} _ {d, i} ^ {(1)}, P _ {d, i} ^ {(1)}\\right) + \\lambda_ {2} \\cdot \\mathrm {S L} \\left(\\hat {P} _ {d, i} ^ {(2)}, P _ {d, i} ^ {(2)}\\right) \\tag {13} \\\\ + \\alpha \\cdot \\sum_ {i = 1} ^ {n} \\left| \\delta \\operatorname {a r g m a x} _ {d} \\left(\\hat {P} _ {d, i}\\right) - \\delta \\operatorname {a r g m a x} _ {d} \\left(P _ {d, i}\\right) \\right| + \\epsilon \\cdot \\mathrm {L R} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.689, + 0.708, + 0.703 + ], + "angle": 0, + "content": "And the Laplacian remainder is:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.708, + 0.905, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L R} = \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} ^ {(1)} \\right| + \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} ^ {(2)} \\right| \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The term \\(\\mathrm{SL}(\\hat{P}_{d,i}, P_{d,i})\\) (with \\(\\beta = 0.1\\)) is used to compute the error between the predicted outputs and the ground truth targets. Here, \\(\\hat{P}_{d,i}\\) represents the predicted popularity for data point \\(i\\) on day \\(d\\), while \\(P_{d,i}\\) denotes the ground truth popularity for the same data point. The differences between the one-hot encoded predicted and true peak values are measured using \\(\\delta \\mathrm{argmax}_d(\\hat{P}_{d,i})\\) and \\(\\delta \\mathrm{argmax}_d(P_{d,i})\\), respectively. Additionally, the term \\(\\epsilon \\cdot \\mathrm{LR}\\) introduces a Laplacian remainder (LR) to provide further regularization. During training, the weights for the first-order and second-order" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "derivative terms, as well as the L1 loss weight between the one-hot encodings \\((\\lambda_1,\\lambda_2,\\alpha)\\), are adjusted dynamically using a cosine annealing algorithm to ensure smoother convergence throughout the training process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.223, + 0.182 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.483, + 0.268 + ], + "angle": 0, + "content": "In this section, we first describe the experiment settings and evaluation metrics for training models and comparison. Then various experimental results and detailed discussions are provided including evaluations for SMTPD dataset, proposed baseline, multi-modal features and early popularity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.276, + 0.281, + 0.291 + ], + "angle": 0, + "content": "5.1. Experiment settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.403 + ], + "angle": 0, + "content": "We implement our approaches using the Pytorch* framework and train it with the Adam [25] optimizer incorporating L2 penalty of \\(10^{-3}\\), while the batch size is set to 64. The learning rate is initialized to \\(10^{-3}\\) and adjusted by the ReduceLROnPlateau scheduler provided by PyTorch when one epoch is end. Supervised training was performed using the Composite Gradient Loss (CGL) metioned before." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.508 + ], + "angle": 0, + "content": "We introduce error-based metrics and correlation-based metrics to fairly evaluate the compared datasets and models. We introduce the Absolute Error (MAE) and average MAE (AMAE) to respectively evaluate models for single-day prediction and temporal prediction. Assuming that the daily MAE and average MAE are denoted as \\( MAE_{d} \\) and \\( AMAE \\), they can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.519, + 0.483, + 0.557 + ], + "angle": 0, + "content": "\\[\nM A E _ {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} - P _ {d, i} \\right| \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.561, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\nA M A E = \\frac {1}{m} \\sum_ {d = 1} ^ {m} M A E _ {d} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.655 + ], + "angle": 0, + "content": "where \\( n \\) denotes the total of samples and \\( m \\) denotes total of days, while the \\( P_{d} \\) and \\( \\hat{P}_d \\) respectively denotes the ground-truth popularity and predicted popularity for the day \\( d \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.746 + ], + "angle": 0, + "content": "For correlation-based metrics, we introduce the daily SRC and the average SRC to evaluate the models for both single-day prediction and temporal prediction from a different perspective. Assuming \\( n \\), \\( m \\), \\( P_d \\), and \\( \\hat{P}_d \\) are defined consistently as mentioned earlier, the daily SRC and average SRC can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.754, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\nS R C _ {d} = \\frac {1}{n - 1} \\sum_ {i = 1} ^ {n} \\left(\\frac {\\hat {P} _ {d , i} - \\bar {\\hat {P}} _ {d}}{\\sigma_ {\\hat {P} _ {d}}}\\right) \\left(\\frac {P _ {d , i} - \\bar {P} _ {d}}{\\sigma_ {P _ {d}}}\\right) \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.799, + 0.483, + 0.838 + ], + "angle": 0, + "content": "\\[\nA S R C = \\frac {1}{m} \\sum_ {d = 1} ^ {m} S R C _ {d} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.88 + ], + "angle": 0, + "content": "where \\(\\bar{P}_d\\) and \\(\\sigma_{P_d}\\) are mean and standard deviation of the corresponding popularity for the day \\(d\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "In the curves, sequential performance is represented by the MAE and SRC metrics, which illustrate daily prediction errors and correlations for popularity scores. The X-axis denotes days, while the Y-axis shows \\( MAE_{d} \\) or \\( SRC_{d} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.163, + 0.702, + 0.178 + ], + "angle": 0, + "content": "5.2. SMPD VS. SMPD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.457 + ], + "angle": 0, + "content": "In this subsection, we conduct comparisons and discussions between the SMTPD and SMPD. We first measure three most recently proposed state-of-the-art popularity prediction methods on two datasets. As these methods are all designed basing on the settings of SMPD, we make some modifications to let them fit for training and testing on our SMTPD. Other top-performing models (e.g. [44, 46, 50]) include components that are not applicable to SMTPD or are not reproducible, such as sliding window average or undisclosed self-trained modules. Hence, we do not discuss them in our experiments. First, we distribute content from corresponding modalities into their respective feature extraction modules. Then we use BERT-Multilingual as the textual feature extractor to address the challenges of multilingual text. Given that our SMTPD dataset consists of temporal popularity scores, we restrict these methods to predict only the single popularity score on day 30. To mitigate random bias, we evaluate using 5-fold cross-validation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.67 + ], + "angle": 0, + "content": "Table 4 shows minimal performance differences across folds, confirming that SMTPD contains abundant data under a well-balanced sample distribution. It is evident that on the single-output prediction of popularity on days 7, 14, and 30 within SMTPD, the performance of method [26] based on GBDT outperforms the deep learning method [13] and [52] in terms of both MAE and SRC through its powerful regression capabilities of Catboost [39]. Notably, our method achieves the best results with the addition of EP, where the MAE of the other three models increases as the prediction horizon extends. In contrast, our model's MAE begins to decrease after day 14. The deep learning method [13] also performs well on days 7 and 14, but its effectiveness declines beyond day 14." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Moreover there are also two interesting observations. First, all methods achieve higher SRC on SMPD than on SMPD, likely because SMPD lacks specific time points, making it challenging to predict time-dependent popularity trends accurately as popularity declines and their MAE performances are certainly reduced that the MAE values increase by over 0.16. Besides, the larger popularity range and standard deviation among SMPD samples (shown in Table 2) also contributes to prediction difficulties." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.806, + 0.834 + ], + "angle": 0, + "content": "5.3. Evaluation for Proposed Baseline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Refer to Table 3, we attempted to validate the rationale of partial structures in the proposed baseline model by employing alternative feature extractors and regression networks. Using BERT-Base as the textual features extractor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.222, + 0.901 + ], + "angle": 0, + "content": "*https://pytorch.org" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18853" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "lead to an increase of MAE by around 0.19. This decline in performance may be caused by the limitations of BERT-base model in handling multilingual texts, as it struggles to effectively capture the semantics and contextual information across different languages. The MAE also increased when the MLP (3 layers) was used as the regression structure. Such a structure is similar to [13], suggesting that MLP lacks the capture of temporal information. The SRC did not change for either of the above substitutions, probably due to the high correlation given by EP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.251, + 0.361, + 0.267 + ], + "angle": 0, + "content": "5.4. Discussion of Early Popularity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.452 + ], + "angle": 0, + "content": "We evaluate the performance of proposed baseline on SMTPD. As shown in Table 5, the proposed baseline gets a great improvement from existed methods that surpassing the second best method by -0.798/0.109 MAE/SRC for predicting the popularity of day 30. This great improvement is mainly brought by the early popularity (EP). Without the assistant of EP (row 2 in Table 5), the baseline performance sharply reduced around the existed methods. Using the EP predicted by existed method (1.717/0.864 MAE/SRC) as an input also contributes to the baseline's performance (row 3 in Table 5). However the performance gap between the predicted EP and true EP remains large." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.62 + ], + "angle": 0, + "content": "Instead of directly involving the EP in the model architecture, introducing EP in the training supervision is another suitable option. To validate the effectiveness of such operation, we train a baseline model for predicting the popularities from day 1 to 30. As it presented by the row 1 in Table 5, adding the 1st day's popularity to the prediction sequence make a slight boost, what is in line with the LSTM's ability to capture temporal dependencies in popularity sequences. Having more preceding temporal information from backpropagation leads to more precise predictions of popularity in subsequent time steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.742 + ], + "angle": 0, + "content": "These comparisons clearly indicate that EP plays a crucial role in popularity prediction, and accurately forecasting the first day's popularity is key to predicting future popularity. The strong correlation between EP and future popularity significantly benefits the prediction task. Our results, as demonstrated by the MAE and SRC curves for both natural EP and our model's predictions, underscore the model's capability to leverage EP for enhanced accuracy. From the" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.844 + ], + "angle": 0, + "content": "
BERT-BaseBERT-MulMLPLSTMAMAEASRC
✓✓0.7820.958
✓✓0.7860.958
✓✓0.7170.959
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.898 + ], + "angle": 0, + "content": "Table 3. The evaluation for the proposed baseline model, mainly on evaluating the model's performance in adapting to multilingual and temporal popularity." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.905, + 0.253 + ], + "angle": 0, + "content": "
MethodSMTPD (day 7)SMTPD (day 14)SMTPD (day 30)
AverageAverageAverage
Ding et al. [13]1.715/0.8491.669/0.8461.592/0.843
w. EP0.715/0.9640.742/0.9590.749/0.931
Lai et al. [26]1.573/0.8751.524/0.8721.495/0.864
w. EP0.725/0.9570.753/0.9620.760/0.957
Xu et al. [52]1.895/0.8171.832/0.8181.743/0.820
w. EP0.754/0.9620.798/0.9560.822/0.949
Ours w/o. EP1.673/0.8521.628/0.8501.563/0.848
Ours0.713/0.9640.735/0.9590.732/0.959
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.265, + 0.905, + 0.306 + ], + "angle": 0, + "content": "Table 4. The performance (MAE/SRC) was compared across four models, including our model, using the SMTPD dataset, both with and without EP." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.32, + 0.905, + 0.459 + ], + "angle": 0, + "content": "
MethodAMAEASRCMAE(day 30 only)SRC(day 30 only)
w/o. EP(1-30)1.5620.8561.5300.850
w/o. EP(2-30)1.6300.8491.5510.849
w/o. EP+[26]1.6280.8511.5550.848
ours0.7170.9590.7320.959
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.469, + 0.905, + 0.526 + ], + "angle": 0, + "content": "Table 5. Assessment of EP across different scenarios. Here, \"1-30\" denotes the prediction target spanning a continuous sequence from the 1st day to the 30th day, as does \"2-30\" (to align with methods having EP)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.644 + ], + "angle": 0, + "content": "visualized results, tough EP exhibits high correlation to the popularities of following days, the MAE sharply increased over time. By contrast, our baseline model could well optimize the MAE and achieve even better SRC performance comparing to the natural EP curve. Therefore, the proposed baseline model is effective to utilize EP achieving better prediction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.656, + 0.634, + 0.672 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.907, + 0.893 + ], + "angle": 0, + "content": "This study aims to address the challenge of time-series popularity prediction in social media. In this paper, we introduce a novel multilingual, multi-modal time-series popularity dataset based on YouTube and suggest a multilingual temporal prediction model tailored to this dataset. Through experiments, we demonstrate the effectiveness of this approach in predicting social media popularity time-series in a multilingual environment. The experiments show that combining multi-modal features with early popularity significantly improves prediction accuracy. However, this study has yet to address challenges such as multi-frame information in videos, maximizing the use of language diversity, and deeper multi-modal exploration. These areas will be key focuses for our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18854" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Deepak Agarwal, Bee-Chung Chen, Pradheep Elango, Nitin Motgi, Seung-Taek Park, Raghu Ramakrishnan, Scott Roy, and Joe Zachariah. Online models for content optimization. Advances in Neural Information Processing Systems, 21, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Sercan Ö Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, pages 6679-6687, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Peng Bao, Hua-Wei Shen, Junming Huang, and Xue-Qi Cheng. Popularity prediction in microblogging network: a case study on sina weibo. In Proceedings of the 22nd international conference on world wide web, pages 177–178, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] Shenghua Bao, Guirong Xue, Xiaoyuan Wu, Yong Yu, Ben Fei, and Zhong Su. Optimizing web search using social annotations. In Proceedings of the 16th international conference on World Wide Web, pages 501-510, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[5] Laura Ceci. Youtube-statistics & facts. Statista. com [online].[cit. 2022-05-05]. Dostupné z: https://www.statista.com/topics/2019/youtube, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[6] Junhong Chen, Dayong Liang, Zhanmo Zhu, Xiaojing Zhou, Zihan Ye, and Xiuyun Mo. Social media popularity prediction based on visual-textual features with xgboost. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2692-2696, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[7] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm SIGkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[8] Weilong Chen, Chenghao Huang, Weimin Yuan, Xiaolu Chen, Wenhao Hu, Xinran Zhang, and Yanru Zhang. Title-and-tag contrastive vision-and-language transformer for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7008-7012, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[9] Xiaolu Chen, Weilong Chen, Chenghao Huang, Zhongjian Zhang, Lixin Duan, and Yanru Zhang. Double-fine-tuning multi-objective vision-and-language transformer for social media popularity prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9462-9466, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Wen-Huang Cheng, Sijie Song, Chieh-Yun Chen, Shintami Chusnul Hidayati, and Jiaying Liu. Fashion meets computer vision: A survey. ACM Computing Surveys (CSUR), 54 (4):1-41, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Minhwa Cho, Dahye Jeong, and Eunil Park. Amps: Predicting popularity of short-form videos using multi-modal attention mechanisms in social media marketing environments. Journal of Retailing and Consumer Services, 78: 103778, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 4171–4186, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[13] Keyan Ding, Ronggang Wang, and Shiqi Wang. Social media popularity prediction: A multiple feature fusion approach with deep neural networks. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2682-2686, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Zhabiz Gharibshah and Xingquan Zhu. User response prediction in online advertising. aCM Computing Surveys (CSUR), 54(3):1-43, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Anindya Ghose and Sha Yang. An empirical analysis of search engine advertising: Sponsored search in electronic markets. Management science, 55(10):1605-1622, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[16] Sayan Ghosh, Dylan Baker, David Jurgens, and Vinodkumar Prabhakaran. Detecting cross-geographic biases in toxicity modeling on social media. In Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT), pages 313–328, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[17] Marcos Andre Gonçalves, Jussara M Almeida, Luiz GP dos Santos, Alberto HF Laender, and Virgilio Almeida. On popularity in the blogosphere. IEEE Internet Computing, 14(3): 42-49, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. Learning word vectors for 157 languages. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] Mohammad Hajarian, Mark Anthony Camilleri, Paloma Díaz, and Ignacio Aedo. A taxonomy of online marketing methods. In Strategic corporate communication in the digital age, pages 235-250. Emerald Publishing Limited, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[22] Chih-Chung Hsu, Chia-Ming Lee, Xiu-Yu Hou, and Chi-Han Tsai. Gradient boost tree network based on extensive feature analysis for popularity prediction of social posts. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9451-9455, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[23] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[24] Aditya Khosla, Atish Das Sarma, and Raffay Hamid. What makes an image popular? In Proceedings of the 23rd international conference on World wide web, pages 867-876, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations(ICLR), San Diego, California, USA, May 7-9, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Xin Lai, Yihong Zhang, and Wei Zhang. Hyfea: Winning solution to social media popularity prediction for multimedia grand challenge 2020. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4565-4569, 2020. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18855" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[27] Chenyu Li, Jun Liu, and Shuxin Ouyang. Analysis and prediction of content popularity for online video service: a youku case study. China Communications, 13(12):216-233, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[28] Jinning Li, Yirui Gao, Xiaofeng Gao, Yan Shi, and Guihai Chen. Senti2pop: sentiment-aware topic popularity prediction on social media. In 2019 IEEE International conference on data mining (ICDM), pages 1174-1179. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.277, + 0.481, + 0.303 + ], + "angle": 0, + "content": "[30] Yury Lifshits. Ediscope: Social analytics for online news. Yahoo Labs, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Hung-Hsiang Lin, Jiun-Da Lin, Jose Jaena Mari Ople, Jun-Cheng Chen, and Kai-Lung Hua. Social media popularity prediction based on multi-modal self-attention mechanisms. IEEE Access, 10:4448-4455, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.363, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[32] An-An Liu, Xiaowen Wang, Ning Xu, Junbo Guo, Guoqing Jin, Quan Zhang, Yejun Tang, and Shenyuan Zhang. A review of feature fusion-based media popularity prediction methods. Visual Informatics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.42, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[33] Marco Lui and Timothy Baldwin. langid.py: An off-the-shelf language identification tool. In Proceedings of the ACL 2012 System Demonstrations, pages 25-30, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[34] Shijian Mao, Wudong Xi, Lei Yu, Gaotian Lü, Xingxing Xing, Xingchen Zhou, and Wei Wan. Enhanced catboost with stacking features for social media prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9430-9435, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[35] Masoud Mazloom, Robert Rietveld, Stevan Rudinac, Marcel Worring, and Willemijn Van Dolen. Multimodal popularity prediction of brand-related social media posts. In Proceedings of the 24th ACM international conference on Multimedia, pages 197-201, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[36] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proceedings of the 1st International Conference on Learning Representations (ICLR), Scottsdale, Arizona, USA, May 2-4, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[37] Alessandro Ortiz, Giovanni Maria Farinella, and Sebastiano Battiato. Prediction of social image popularity dynamics. In Image Analysis and Processing-ICIAP 2019: 20th International Conference, Trento, Italy, September 9–13, 2019, Proceedings, Part II 20, pages 572–582. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. Advances in neural information processing systems, 31, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Cédric Richier, Eitan Altman, Rachid Elazouzi, Tania Altman, Georges Linares, and Yonathan Portilla. Modelling" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "view-count dynamics in youtube. CoRR, abs/1404.2570, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[41] Cedric Richier, Eitan Altman, Rachid Elazouzi, Tania Jimenez, Georges Linares, and Yonathan Portilla. Bioinspired models for characterizing youtube viewcout. In 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 297-305. IEEE, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[42] Satoshi Sanjo and Marie Katsurai. Recipe popularity prediction with deep visual-semantic fusion. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2279–2282, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.3 + ], + "angle": 0, + "content": "[43] Gabor Szabo and Bernardo A Huberman. Predicting the popularity of online content. Communications of the ACM, 53 (8):80-88, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.904, + 0.37 + ], + "angle": 0, + "content": "[44] YunPeng Tan, Fangyu Liu, BoWei Li, Zheng Zhang, and Bo Zhang. An efficient multi-view multimodal data processing framework for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7200-7204, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.904, + 0.426 + ], + "angle": 0, + "content": "[45] Alexandru Tatar, Marcelo Dias De Amorim, Serge Fdida, and Panayotis Antoniadis. A survey on predicting the popularity of web content. Journal of Internet Services and Applications, 5(1):1-20, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.904, + 0.496 + ], + "angle": 0, + "content": "[46] Kai Wang, Penghui Wang, Xin Chen, Qiushi Huang, Zhendong Mao, and Yongdong Zhang. A feature generalization framework for social media popularity prediction. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4570-4574, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.904, + 0.566 + ], + "angle": 0, + "content": "[47] Bo Wu, Wen-Huang Cheng, Yongdong Zhang, Qiushi Huang, Jintao Li, and Tao Mei. Sequential prediction of social media popularity with deep temporal context networks. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI), pages 3062-3068, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[48] Bo Wu, Wen-Huang Cheng, Peiye Liu, Bei Liu, Zhaoyang Zeng, and Jiebo Luo. Smp challenge: An overview of social media prediction challenge 2019. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2667-2671. ACM, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[49] Bo Wu, Peiye Liu, Wen-Huang Cheng, Bei Liu, Zhaoyang Zeng, Jia Wang, Qiushi Huang, and Jiebo Luo. Smp challenge: An overview and analysis of social media prediction challenge. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9651-9655, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[50] Jianmin Wu, Liming Zhao, Dangwei Li, Chen-Wei Xie, Siyang Sun, and Yun Zheng. Deeply exploit visual and language information for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7045–7049, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[51] Leilei Wu, Lingling Yi, Xiao-Long Ren, and Linyuan Lu. Predicting the popularity of information on social platforms without underlying network structure. Entropy, 25(6):916, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Kele Xu, Zhimin Lin, Jianqiao Zhao, Peicang Shi, Wei Deng, and Huaimin Wang. Multimodal deep learning for social media popularity prediction with attention mechanism. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4580-4584, 2020. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18856" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[53] Jaewon Yang and Jure Leskovec. Patterns of temporal variation in online media. In Proceedings of the fourth ACM international conference on Web search and data mining, pages 177-186, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[54] Zhizhen Zhang, Xiaohui Xie, Mengyu Yang, Ye Tian, Yong Jiang, and Yong Cui. Improving social media popularity prediction with multiple post dependencies. CoRR, abs/2307.15413, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[55] Alireza Zohourian, Hedieh Sajedi, and Arefeh Yavary. Popularity prediction of images and videos on instagram. In 2018 4th International Conference on Web Research (ICWR), pages 111-117, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18857" + } + ] +] \ No newline at end of file diff --git a/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_origin.pdf b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25c733a4f4fac5b5600c7610171532d7f651a6a4 --- /dev/null +++ b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/3f53df77-cd61-4f83-8a2e-0eb089935375_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21afe8e2dc95341b71201985854839b1f1fd61f55908eaa95412b1cf485590c3 +size 1926449 diff --git a/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/full.md b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..160a6cefea9946f9f7fd37186f20a14b185dc2c2 --- /dev/null +++ b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/full.md @@ -0,0 +1,382 @@ +# SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity + +$^{1}$ Yijie Xu*, $^{1}$ Bolun Zheng*, $^{1}$ Wei Zhu*, $^{1}$ Hangjia Pan, $^{1}$ Yuchen Yao, $^{2}$ Ning Xu, $^{2}$ Anan Liu, $^{3}$ Quan Zhang, $^{1}$ Chenggang Yan $^{1}$ Hangzhou Dianzi University, $^{2}$ Tianjin University, $^{3}$ Peking University + +# Abstract + +Social media popularity prediction task aims to predict the popularity of posts on social media platforms, which has a positive driving effect on application scenarios such as content optimization, digital marketing and online advertising. Though many studies have made significant progress, few of them pay much attention to the integration between popularity prediction with temporal alignment. In this paper, with exploring YouTube's multilingual and multi-modal content, we construct a new social media temporal popularity prediction benchmark, namely SMTPD, and suggest a baseline framework for temporal popularity prediction. Through data analysis and experiments, we verify that temporal alignment and early popularity play crucial roles in social media popularity prediction for not only deepening the understanding of temporal dynamics of popularity in social media but also offering a suggestion about developing more effective prediction models in this field. Code is available at https://github.com/zhuwei321/SMTPD + +# 1. Introduction + +With the advancement of Internet communication technology in recent years, social media has gradually emerged and has influenced various aspects of human life. Any content posted on social media stands the chance of becoming hot spot, and widely disseminated social media content can generate significant social and economic benefits. The prediction of social media popularity holds immense potential applications in content optimization [1, 30], online advertising [14, 15], digital marketing [19, 45], search recommendations [4, 17], intelligent fashion [10], and beyond. + +In the early stages of popularity research, statistical and topological methods [41, 43, 53] were extensively employed, with the research focus primarily on time-aware popularity prediction. These prediction methods rely on information about the earlier popularity of posted content. + +![](images/a3265d56d3056fdfb495ea22510f3d30c86ae82ee55356ad25a5632066d9d33f.jpg) +Figure 1. Content sections and popularity trend of SMTPD. In 1a, a sample involves four sections of content, with temporal popularity. 1b depicts box plots of daily popularity scores, illustrating variations in popularity distribution at different time points. The distribution consistently demonstrates a decay pattern over time. + +Along with the many large-scale social media popularity datasets are proposed [3, 42, 47, 48], the machine learning based methods are widely employed to achieve reasonable predictions of social media popularity and have shown remarkable performance. These methods emphasize meticulous design in feature extraction and fusion, followed by popularity prediction using machine learning models such as deep neural networks (DNN) and gradient boosting decision trees (GBDT). + +Table 1 lists the public or mentioned social media popularity datasets in the most recent years as far as we know. For above machine learning based methods, the large-scale + +
DatasetSourceCategorySamplesLanguagePrediction Type
Mazloom [35]Instagramfast food brand75KEnglishsingle
Sanjo [42]Cookpadrecipe150KJapanesesingle
TPIC17 [47]Flickr-680KEnglishsingle
SMPD [48]Flickr11 categories486KEnglishsingle
AMPS [11]YouTubeshorts13KKoreansingle
SMTPD (ours)YouTube15 categories282Kover 90 languagessequential
+ +Table 1. Comparing SMTPD with existed multi-modal social media popularity datasets. + +social media popularity datasets [3, 11, 35, 42, 47, 48] provide the fundamental supports. + +However, existing datasets and related studies have notable limitations, such as insufficient multi-modal data, limited language diversity, and, crucially, the absence of a consistent timeline for prediction. As shown in Figure1b, popularity distributions shift over time, and predictions made at irregular intervals can reduce accuracy and create challenges for practical applications. + +Furthermore, effective popularity prediction requires integrating both the multi-modal content of social media and information on the communication process to capture the dynamic trends in popularity. This essential aspect is missing in most current prediction efforts. + +Targeting to fix up the above shortcomings of existed datasets, in this paper, we propose a new benchmark called Social Media Temporal Popularity Dataset (SMTPD) by observing multi-modal content from mainstream for cutting-edge research in the field of social multimedia along with multi-modal feature temporal prediction. As shown in Figure 1a, multi-modal contents of social media posts in multiple languages generated the daily popularity information during the communication process. We define such a post as a sample of SMTPD, and propose a multi-modal framework as a baseline to achieve the temporal prediction of popularity scores. The proposed framework consists of two parts, feature extraction and regression. In the feature extraction part, multiple pre-training models and pre-processing methods are introduced to translate the multi-modal content into deep features. In the regression part, we encode the state and time sequence of the extracted features, and adopt the LSTM-based structure to regress the 30-day popularities. Through analysis and experiments, we discover the importance of early popularity to the task of popularity prediction, and demonstrate the effectiveness of our method in temporal prediction. Generally, the contribution of this work can be summarized as: + +- Against the missing of temporal information in social media popularity researches, we observe over 282K multilingual samples from mainstream social media since they released on the network lasting for 30 days. We refer to these samples as SMTPD, a new benchmark for temporal + +popularity prediction. + +- Basing on existed methods, we innovate in both the selection of feature extractors and the construction of the temporal regression component, and suggest a baseline model which enables temporal popularity prediction to be conducted across multiple languages while aligning prediction times. +- Exploring the popularity distribution and the correlation between popularity at different times. Based on these, We find the importance of early popularity for popularity prediction task, and point out that the key-point for predicting popularity is to accurately predict the early popularity. + +# 2. Related work + +Statistical and topological methods. Szabo et al. [43] found that the early-stage popularity notably influences subsequent popularity. Yang et al. [53] conducted cluster analysis on temporal patterns within online content, revealing distinct characteristics of popularity variations across different clusters. Richier et al. [40, 41] proposes bio-inspired models to characterize the evolution of video view counts. Wu et al. [51] suggested that the information cascade process is best described as an activate-decay dynamic process. + +Multi-modal feature based Methods. Ding et al. [13] use pre-trained ResNet and BERT to extract visual and textual features, with the DNN regression. Xu et al. [52] and Lin et al. [31] adopted attention mechanisms to effectively integrate multi-modal features. Chen et al. [6] compared the performance of several regression models, among which XGBoost [7] exhibited the best results. Hsu et al. [22] employed LightGBM [23] and TabNet [2] to capture intricate semantic relationships in multi-modal features. Lai et al. [26] engineered handcrafted features, exploiting CatBoost for regression. Mao et al. [34] enhanced CatBoost-based model by stacking features. Tan et al. [44] extracted visual-textual features by ALBEF [29] and Chen et al. [8, 9] enriched more intermodal features to promote predictive performance. Wu et al. [50] emphasized increasing feature dimensions to improving predictive performance. The post dependencies captured by sliding window average [46] and DSN [54] has also led to improvements. Many of these multi-modal methods are based on the SMPD and working out great [32, 49]. + +Social media popularity datasets. Mazloom et al. [35] conducted experiments using a dataset of posts related to fast-food brands collected from Instagram. Sanjo et al. [42] provided a recipe popularity prediction dataset based on Cookpad, which includes text content entirely in Japanese. Li et al. [28] predicted the future popularity of topics by using historical sentiment information based on Twitter's text data records. TPIC17 [47] and SMPD [48] are datasets for single popularity prediction task based on Flickr. + +We believe that the loss of temporal alignment about + +popularity in existing methods is a common problem. The image dynamic popularity dataset [37] gave us insights, so we propose a multi-modal temporal popularity benchmark to address the shortcomings of existing studies. + +# 3. SMTPD Dataset + +With the great development of web communication technologies, social multimedia has become the most popular media around the world. However, most existing social media prediction datasets are built on the basis of single-output prediction, meaning they observe existing posts and infer their popularity based on their posting time. As mentioned earlier, these data without aligned posting times exhibit non-uniform popularity distributions. Therefore, SMTPD keep an eye on one of the most popular worldwide social media, YouTube [5], with over 282K samples, primarily focusing on the evolution of popularity over time for these samples. We collected over 402K raw data samples. Specifically, we first removed records with missing values—which likely resulted from network issues during data acquisition. Next, we eliminated samples that either failed to be crawled on certain days or had been deleted within 30 days. Finally, we filtered out potential outliers using the $3\sigma$ rule. In this section, we provide a comprehensive overview of the SMTPD's composition. Additionally, we present various data analysis results that contribute to the feasibility of our approach. + +Unlike the retrospective methods used in previous datasets, we take note of the newly posted multi-modal content and then observe popularity information every 24 hours via YouTube API. Considering that in real-world applications, the posts to be predicted are always new, this data observation method allows us to obtain aligned temporal popularity of new posts, which is more in line with practical applications. Figure 1a show a sample from SMTPD. We note diverse attributes, including visual content, textual content, metadata, and user profiles. Additionally, we focus on the temporal popularity for each sample within 30 days after release. We perform basic statistics for SMTPD, as shown in Table 2. + +# 3.1. Temporal Popularity + +The utilization of temporal popularity data is versatile. It not only serves as the output for predictions but also as inputs, where a segment preceding a specific time point, coupled with multi-modal content, can forecast subsequent popularity. We analyze the distribution and the correlation of popularity over time, so as to facilitate a comprehensive understanding of the significance of temporal popularity for predictions from different perspectives. + +A number of popularity definitions have been born from previous work [27, 55]. We use Khosla's popularity score transformation [24] as it can normalize the distribution of + +view counts to a suitable popularity score's distribution. It can be represented as: + +$$ +p = \log_ {2} (\frac {v}{d} + 1) \tag {1} +$$ + +where $p$ is the popularity score, $v$ is the view counts of a sample, $d$ represents the corresponding number of days after the post's release. As shown in Figure 1, the histogram of view counts reveals an extreme long-tailed distribution, which might not be suitable as a prediction target. It's evident that the distribution of the popularity score metric becomes more reasonable. + +In addition, we notice that there is a strong correlation between the popularity scores of different days. The Pearson Correlation (PC) and Spearman Ranking Correlation (SRC) can respectively reflect the linear relationship and rank-order correlation between popularity scores of different days as follows: + +$$ +P C = \frac {\sum_ {i = 1} ^ {n} \left(X _ {i} - \bar {X}\right) \left(Y _ {i} - \bar {Y}\right)}{\sqrt {\sum_ {i = 1} ^ {n} \left(X _ {i} - \bar {X}\right) ^ {2}} \sqrt {\sum_ {i = 1} ^ {n} \left(Y _ {i} - \bar {Y}\right) ^ {2}}} \tag {2} +$$ + +$$ +S R C = \frac {1}{n - 1} \sum_ {i = 1} ^ {n} \left(\frac {X _ {i} - \bar {X}}{\sigma_ {X}}\right) \left(\frac {Y _ {i} - \bar {Y}}{\sigma_ {Y}}\right) \tag {3} +$$ + +where $X$ and $Y$ denote the popularity scores belonging to 2 different days, while $\bar{X}$ and $\sigma_{X}$ denote the mean and standard deviation of $X$ , and similarly for $Y$ . PC and SRC respectively reflect the linear relationship and rank-order correlation between popularity scores of different days. + +The Pearson Correlation (PC) and Spearman Rank Correlation (SRC) heat maps illustrate internal relationships in daily popularity scores, establishing a basis for prediction based on these temporal patterns. Further details are available in the supplementary materials. + +
StatisticsSMTPDSMPD (Train/Test)
Number of samples282.4K305K/181K
Number of users152.7K38K/31K
Mean popularity5.956.41/5.12
STD popularity4.152.47/2.41
Number of categories1511
Number of custom tags960K250K
Average length of title53.4 chars29 words
Mean duration1853.4 s-
+ +Table 2. The basic statistics of all samples in SMPD, comparing to SMPD[48]. Due to the multilingual environment, we use chars (characters) to describe the average length of title. + +# 3.2. Multi-Modal Content + +Multi-modal feature fusion is the mainstream approach in today's popularity prediction methods. In this section, we will divide the data into corresponding modality and elucidate the multi-modal content in SMTPD. + +# 3.2.1. Visual Content + +SMTPD samples, primarily drawn from a multi-modal social media platform, are heavily influenced by visual elements. Since metrics like view count and popularity score are often based on the cover frame that viewers see before clicking to play, we use the cover frame as the primary visual content. This approach balances popularity impact with copyright considerations. + +# 3.2.2. Textual Content + +The textual content in posts plays a crucial role in media dissemination and user engagement. We examine several core elements, including category, title, description, and hashtags (denoted by the “#” symbol), along with the user ID (user nickname) from profile metadata. YouTube defines 15 distinct video categories, all in English, which support content organization and discovery. Figure 3 shows the statistical distribution of content by category. + +![](images/7bb60ec655ff6911907a94c33cdcbf578ca08183d4eafdc051bf26f2de1f3ea3.jpg) +(a) The PC Matrix + +![](images/4f299ea09a29b30b6b778651443af19eeca7beb6115a9796887f5f3cb7516c23.jpg) +(b) The SRC Matrix + +As an international platform, YouTube hosts user-defined textual content in multiple languages beyond standard categories. Figure 4 illustrates the distribution and popularity bias across title languages. + +In previous researches, the datasets typically had uniform language for text content, without considering multilingual aspects. While this simplifies modeling for prediction methods, it also poses significant limitations for international social media platforms which not restricted to a single language. Besides, predictions based on a few languages may introduce biases [16] against popularity. + +# 3.2.3. Numerical values + +Numerical attributes frequently function as supplementary factors in enhancing prediction accuracy. In this study, we + +examined key attributes, including video duration, uploader follower count, and the number of posts by the uploader. Additionally, we manually derived several auxiliary metrics, such as title length, the count of custom tags, and description length, to further enrich the feature set and support predictive robustness. + +# 4. Temporal popularity prediction + +The multi-modal feature-based temporal prediction framework is shown in Figure 5, which is divided into two main parts: multi-modal feature extraction and temporal popularity score regression. This framework is designed to adapt to the multilingual temporal prediction tasks under time alignment. The prediction target is set as the popularity of samples, which corresponds to the temporal popularity within 30 days after the sample's release. + +# 4.1. Multi-Modal Feature Extraction + +For different modalities in SMTPD, we adopted distinct feature extraction methods. We additionally incorporated features from the categorical modality to investigate the impact of categorical features on popularity prediction. + +# 4.1.1. Visual Features + +The cover image of a social media content is a very important component which would help users to quickly understand what they would see if clicking to view this content. + +![](images/26bd9072261138410e4150ff6346bb7603c8c0ffcb161721b2cb606bf9b3cd98.jpg) +Figure 2. The heat map of daily popularity correlations. The figure clearly shows that there is a high degree of correlations in popularity between consecutive days. + +![](images/cef5d7564335d9ed167cf9ab5edb8f076c9a49a50bf4964b5854ff49d3395ad2.jpg) +(a) The statistics of samples in each category +(b) Average popularity of each category +Figure 3. The statistics based on category. 3a counts the number of samples in each category, and 3b shows the average popularity score of samples in each category. + +![](images/adedc1fccad5168bb90b96d8b059cb0fc4848ca40084985b77bd1c06cdf3d532.jpg) +Figure 4. Languages analysis. The left is the proportions of these languages, with "Other" encompassing 90 different languages. The right represents the average popularity in different languages, revealing the geographic biases. The languages of samples are counted by reference to the title. + +Therefore, the cover image would directly affect the popularity of a social media content. The semantic information provided in the cover image plays a key role for user's understanding. We adopt convolutional neural networks ResNet-101 [20] pre-trained on ImageNet as our visual feature extraction model to obtain the semantic information provided by the cover. To match the input size of ResNet-101, we re-scale the size of cover image to $224 \times 224$ . The 2048-dimension feature vector before the final classification layer will be selected as the final visual feature $f_{v}$ : + +$$ +f _ {v} = \operatorname {R e s N e t} (S (I)) \tag {4} +$$ + +where $I$ denotes the cover image and $S$ denotes the re-scale operation. However, the cover image would not always be uploaded by authors or be unavailable due to the network transmission. For those contents missing the cover image, we use blank images with all pixels are set to zero instead. + +# 4.1.2. Textual Features + +Textual information significantly impacts a user's choice to view or skip content. Key textual inputs include category, title, tags, description, and user profile ID. Extracting semantic features from these inputs is crucial for accurate popularity prediction. + +However, unlike previous popularity prediction tasks, one distinguishing characteristic of SMTPD is that its text content includes multiple languages, and many samples contain text in more than one language. This makes it challenging to use many pre-trained word vector models based on single-language corpora, such as [18, 36, 38]. + +Thanks to the multilingual capability of BERT-Multilingual [12], which processes text across languages in a single model, we use it to extract a 768-dimensional feature vector for each text, capturing semantic information as: + +$$ +f _ {t} ^ {k} = \operatorname {B E R T} \left(T ^ {k}\right) \tag {5} +$$ + +![](images/750954a6f40a9685f6a6273ec7d4849189e064e6f84730d43c954b4deb1b5ea4.jpg) +Figure 5. The proposed model consists of two layers. The upper layer processes visual, textual, numerical, and categorical features, which are input to a multi-layer perceptron (MLP). The MLP outputs are concatenated and fed into a lower LSTM layer for regression. Each LSTM cell's output is passed through another MLP, with the final MLP outputs combined to generate a time prediction sequence using the ReLU activation function. + +where $k \in \{cat, tit, tag, des, uid\}$ maps to category, title, tags, description, and user ID. These are combined into a $768 \times 5$ matrix, transformed through $5 \times 5$ convolution to produce the 768D textual feature $f_{t}$ , which can be expressed as: + +$$ +f _ {t} = \operatorname {C o n v} _ {5 \times 5} \left(\mathcal {C} \left(f _ {t} ^ {\text {c a t}}, f _ {t} ^ {\text {t i t}}, f _ {t} ^ {\text {t a g}}, f _ {t} ^ {\text {d e s}}, f _ {t} ^ {\text {u i d}}\right)\right) \tag {6} +$$ + +where $\mathcal{C}$ denotes the above combing operation. + +# 4.1.3. Numerical Features + +Numerical metrics directly influence whether an audience clicks to watch which sample on social media platform. Also, the number of followers and the number of posts also impact the process through which an audience navigates from content uploaders' user pages to click and watch videos. Hence, we selected these influential numerical metrics for numerical features extraction. As the number of user followers share a similar distribution pattern that exhibiting long-tailed characteristics as view counts, it is necessary to apply a logarithmic transformation before feeding it into the network. Afterward, all numerical items are Z-score normalized and concatenated as numerical features: + +$$ +f _ {n} ^ {j} = \frac {T ^ {j} - \mu_ {j}}{\sigma_ {j}} \tag {7} +$$ + +$$ +f _ {n} = \operatorname {C o n c a t} \left(f _ {n} ^ {f o l}, f _ {n} ^ {p o s}, f _ {n} ^ {d u r}, f _ {n} ^ {t l}, f _ {n} ^ {t n}, f _ {n} ^ {d l}, f _ {n} ^ {E P}\right) \tag {8} +$$ + +where $j \in \{fol, pos, dur, tl, tn, dl, EP\}$ represents the log-scaled follower counts, total number of posts, video du + +ration, title length, the number of tags, description length and the early popularity (the 1st day's popularity). The $\mu$ and $\sigma$ represents the mean and standard deviation of all samples. And Concat denotes the horizontal concatenating operation. + +# 4.1.4. Categorical Features + +The categorical disparities across social media platforms are depicted in Figure 3b. Beyond text-based feature extraction, these disparities inform the generation of categorical features. Language functions as a key marker of regional and cultural distinctions among content creators, as highlighted by the language bias observed in Figure 4. Consequently, language classification features are crucial for enhancing prediction accuracy. To develop these classification features, two sub-features are selected: label and language attributes. The label feature captures the semantic category, while the language attribute is derived from the language in the video title. Assuming that the primary language of all text content matches the title's language, we use langid [33] for classification. Two embedding layers are introduced to encode label and language features independently, each processed by a dedicated multi-layer perceptron (MLP). Finally, the outputs are combined via cumulative multiplication to form the classification feature $f_{c}$ , defined as: + +$$ +f _ {c} = \operatorname {M L P} _ {1} \left(\mathrm {E} _ {1} \left(T ^ {\text {c a t}}\right)\right) \odot \operatorname {M L P} _ {2} \left(\mathrm {E} _ {2} \left(\operatorname {l a n g i d} \left(T ^ {\text {t i t}}\right)\right)\right) \tag {9} +$$ + +where $\mathrm{E}_1$ and $\mathrm{E}_2$ represent two independent embedding layers, and $\mathrm{MLP}_1$ and $\mathrm{MLP}_2$ refer to multi-layer perceptrons applied to the label embedding $\mathrm{E}_1(T^{cat})$ and the language embedding $\mathrm{E}_2(\text{langid}(T^{tit}))$ , respectively. The operator $\odot$ denotes element-wise multiplication, which combines the outputs of the two MLPs to produce the final classification feature $f_c$ . + +# 4.1.5. Feature Fusion + +After extraction, features from different modalities are aligned in dimension through multi-layer perceptrons (MLPs) and then concatenated into fused features $F$ . + +# 4.2. Temporal Popularity Regression + +# 4.2.1. Sequential Encoding and Adjustment + +Considering the high correlation between popularity values on adjacent days in the ground truth, we employ a LSTM [21] structure to get the temporal variations in popularity. As shown in Figure 5, multiple LSTM cells are constructed to transmit temporal information. The initial states (both hidden state and cell state) of the LSTM are generated by passing the $F$ through the same MLP as state encoding. Additionally, the input for LSTM cell at each time step is constructed from $F$ through temporal encoding MLPs. These + +two encoding process can be written as: + +$$ +h _ {0} = c _ {0} = \mathrm {M L P} ^ {\mathrm {h c}} (F) \tag {10} +$$ + +$$ +x _ {s} = \operatorname {M L P} _ {s} ^ {\mathrm {X}} (F) \tag {11} +$$ + +In this formulation, $h_0$ and $c_0$ denote the initial hidden state and cell state, while $x_s$ represents the input at time step $s$ . The $\mathrm{MLP}^{\mathrm{hc}}$ module encodes the state information, and $\mathrm{MLP}_s^{\mathrm{x}}$ handles the temporal encoding at the $s$ -th time step, effectively capturing the sequence dynamics. + +At each time step, both states are treated as the $s$ -th step's features for output. After concatenating and processing through independent MLPs, each LSTM cell produces outputs, helping capture temporal popularity via backpropagation during training. Before final predictions, we apply a non-negative adjustment since popularity scores cannot be negative. This process is formulated as: + +$$ +\operatorname {p r e} _ {s} = \max \left(0, \mathrm {M L P} _ {s} ^ {\text {o u t}} \left(\operatorname {C o n c a t} \left(h _ {s}, c _ {s}\right)\right)\right) \tag {12} +$$ + +Where $pre_{s}, h_{s}, c_{s}$ , and $\mathrm{MLP}_s^{\mathrm{out}}$ denote the predicted value, hidden state, cell state, and output MLP at time step $s$ , respectively. The operation Concat refers to the process of horizontally concatenating the vectors. After these steps, the final temporal popularity predictions are produced. + +# 4.2.2. Loss Function + +The core component of our approach is the Composite Gradient Loss (CGL), specifically designed for this task. This custom loss function consists of several components: SmoothL1Loss (SL) between the model's outputs and targets, the first-order and second-order derivative differences between the outputs and targets, the L1 loss between the onehot encodings of the predicted and ground truth peaks, and the Laplacian remainder (LR). These components are combined with a weight ratio of 1:1:1:1e-6. The overall loss function $\mathcal{L}$ is formulated as: + +$$ +\begin{array}{l} \mathcal {L} = \mathrm {S L} \left(\hat {P} _ {d, i}, P _ {d, i}\right) + \lambda_ {1} \cdot \mathrm {S L} \left(\hat {P} _ {d, i} ^ {(1)}, P _ {d, i} ^ {(1)}\right) + \lambda_ {2} \cdot \mathrm {S L} \left(\hat {P} _ {d, i} ^ {(2)}, P _ {d, i} ^ {(2)}\right) \tag {13} \\ + \alpha \cdot \sum_ {i = 1} ^ {n} \left| \delta \operatorname {a r g m a x} _ {d} \left(\hat {P} _ {d, i}\right) - \delta \operatorname {a r g m a x} _ {d} \left(P _ {d, i}\right) \right| + \epsilon \cdot \mathrm {L R} \\ \end{array} +$$ + +And the Laplacian remainder is: + +$$ +\mathrm {L R} = \sum_ {i = 1} ^ {n} \left| \hat {P} _ {d, i} ^ {(1)} \right| + \sum_ {i = 1} ^ {n} \left| \hat {P} _ {d, i} ^ {(2)} \right| \tag {14} +$$ + +The term $\mathrm{SL}(\hat{P}_{d,i}, P_{d,i})$ (with $\beta = 0.1$ ) is used to compute the error between the predicted outputs and the ground truth targets. Here, $\hat{P}_{d,i}$ represents the predicted popularity for data point $i$ on day $d$ , while $P_{d,i}$ denotes the ground truth popularity for the same data point. The differences between the one-hot encoded predicted and true peak values are measured using $\delta \mathrm{argmax}_d(\hat{P}_{d,i})$ and $\delta \mathrm{argmax}_d(P_{d,i})$ , respectively. Additionally, the term $\epsilon \cdot \mathrm{LR}$ introduces a Laplacian remainder (LR) to provide further regularization. During training, the weights for the first-order and second-order + +derivative terms, as well as the L1 loss weight between the one-hot encodings $(\lambda_1,\lambda_2,\alpha)$ , are adjusted dynamically using a cosine annealing algorithm to ensure smoother convergence throughout the training process. + +# 5. Experiments + +In this section, we first describe the experiment settings and evaluation metrics for training models and comparison. Then various experimental results and detailed discussions are provided including evaluations for SMTPD dataset, proposed baseline, multi-modal features and early popularity. + +# 5.1. Experiment settings + +We implement our approaches using the Pytorch* framework and train it with the Adam [25] optimizer incorporating L2 penalty of $10^{-3}$ , while the batch size is set to 64. The learning rate is initialized to $10^{-3}$ and adjusted by the ReduceLROnPlateau scheduler provided by PyTorch when one epoch is end. Supervised training was performed using the Composite Gradient Loss (CGL) metioned before. + +We introduce error-based metrics and correlation-based metrics to fairly evaluate the compared datasets and models. We introduce the Absolute Error (MAE) and average MAE (AMAE) to respectively evaluate models for single-day prediction and temporal prediction. Assuming that the daily MAE and average MAE are denoted as $MAE_{d}$ and $AMAE$ , they can be defined as: + +$$ +M A E _ {d} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| \hat {P} _ {d, i} - P _ {d, i} \right| \tag {15} +$$ + +$$ +A M A E = \frac {1}{m} \sum_ {d = 1} ^ {m} M A E _ {d} \tag {16} +$$ + +where $n$ denotes the total of samples and $m$ denotes total of days, while the $P_{d}$ and $\hat{P}_d$ respectively denotes the ground-truth popularity and predicted popularity for the day $d$ . + +For correlation-based metrics, we introduce the daily SRC and the average SRC to evaluate the models for both single-day prediction and temporal prediction from a different perspective. Assuming $n$ , $m$ , $P_d$ , and $\hat{P}_d$ are defined consistently as mentioned earlier, the daily SRC and average SRC can be formulated as follows: + +$$ +S R C _ {d} = \frac {1}{n - 1} \sum_ {i = 1} ^ {n} \left(\frac {\hat {P} _ {d , i} - \bar {\hat {P}} _ {d}}{\sigma_ {\hat {P} _ {d}}}\right) \left(\frac {P _ {d , i} - \bar {P} _ {d}}{\sigma_ {P _ {d}}}\right) \tag {17} +$$ + +$$ +A S R C = \frac {1}{m} \sum_ {d = 1} ^ {m} S R C _ {d} \tag {18} +$$ + +where $\bar{P}_d$ and $\sigma_{P_d}$ are mean and standard deviation of the corresponding popularity for the day $d$ . + +In the curves, sequential performance is represented by the MAE and SRC metrics, which illustrate daily prediction errors and correlations for popularity scores. The X-axis denotes days, while the Y-axis shows $MAE_{d}$ or $SRC_{d}$ . + +# 5.2. SMPD VS. SMPD + +In this subsection, we conduct comparisons and discussions between the SMTPD and SMPD. We first measure three most recently proposed state-of-the-art popularity prediction methods on two datasets. As these methods are all designed basing on the settings of SMPD, we make some modifications to let them fit for training and testing on our SMTPD. Other top-performing models (e.g. [44, 46, 50]) include components that are not applicable to SMTPD or are not reproducible, such as sliding window average or undisclosed self-trained modules. Hence, we do not discuss them in our experiments. First, we distribute content from corresponding modalities into their respective feature extraction modules. Then we use BERT-Multilingual as the textual feature extractor to address the challenges of multilingual text. Given that our SMTPD dataset consists of temporal popularity scores, we restrict these methods to predict only the single popularity score on day 30. To mitigate random bias, we evaluate using 5-fold cross-validation. + +Table 4 shows minimal performance differences across folds, confirming that SMTPD contains abundant data under a well-balanced sample distribution. It is evident that on the single-output prediction of popularity on days 7, 14, and 30 within SMTPD, the performance of method [26] based on GBDT outperforms the deep learning method [13] and [52] in terms of both MAE and SRC through its powerful regression capabilities of Catboost [39]. Notably, our method achieves the best results with the addition of EP, where the MAE of the other three models increases as the prediction horizon extends. In contrast, our model's MAE begins to decrease after day 14. The deep learning method [13] also performs well on days 7 and 14, but its effectiveness declines beyond day 14. + +Moreover there are also two interesting observations. First, all methods achieve higher SRC on SMPD than on SMPD, likely because SMPD lacks specific time points, making it challenging to predict time-dependent popularity trends accurately as popularity declines and their MAE performances are certainly reduced that the MAE values increase by over 0.16. Besides, the larger popularity range and standard deviation among SMPD samples (shown in Table 2) also contributes to prediction difficulties. + +# 5.3. Evaluation for Proposed Baseline + +Refer to Table 3, we attempted to validate the rationale of partial structures in the proposed baseline model by employing alternative feature extractors and regression networks. Using BERT-Base as the textual features extractor + +lead to an increase of MAE by around 0.19. This decline in performance may be caused by the limitations of BERT-base model in handling multilingual texts, as it struggles to effectively capture the semantics and contextual information across different languages. The MAE also increased when the MLP (3 layers) was used as the regression structure. Such a structure is similar to [13], suggesting that MLP lacks the capture of temporal information. The SRC did not change for either of the above substitutions, probably due to the high correlation given by EP. + +# 5.4. Discussion of Early Popularity + +We evaluate the performance of proposed baseline on SMTPD. As shown in Table 5, the proposed baseline gets a great improvement from existed methods that surpassing the second best method by -0.798/0.109 MAE/SRC for predicting the popularity of day 30. This great improvement is mainly brought by the early popularity (EP). Without the assistant of EP (row 2 in Table 5), the baseline performance sharply reduced around the existed methods. Using the EP predicted by existed method (1.717/0.864 MAE/SRC) as an input also contributes to the baseline's performance (row 3 in Table 5). However the performance gap between the predicted EP and true EP remains large. + +Instead of directly involving the EP in the model architecture, introducing EP in the training supervision is another suitable option. To validate the effectiveness of such operation, we train a baseline model for predicting the popularities from day 1 to 30. As it presented by the row 1 in Table 5, adding the 1st day's popularity to the prediction sequence make a slight boost, what is in line with the LSTM's ability to capture temporal dependencies in popularity sequences. Having more preceding temporal information from backpropagation leads to more precise predictions of popularity in subsequent time steps. + +These comparisons clearly indicate that EP plays a crucial role in popularity prediction, and accurately forecasting the first day's popularity is key to predicting future popularity. The strong correlation between EP and future popularity significantly benefits the prediction task. Our results, as demonstrated by the MAE and SRC curves for both natural EP and our model's predictions, underscore the model's capability to leverage EP for enhanced accuracy. From the + +
BERT-BaseBERT-MulMLPLSTMAMAEASRC
✓✓0.7820.958
✓✓0.7860.958
✓✓0.7170.959
+ +Table 3. The evaluation for the proposed baseline model, mainly on evaluating the model's performance in adapting to multilingual and temporal popularity. + +
MethodSMTPD (day 7)SMTPD (day 14)SMTPD (day 30)
AverageAverageAverage
Ding et al. [13]1.715/0.8491.669/0.8461.592/0.843
w. EP0.715/0.9640.742/0.9590.749/0.931
Lai et al. [26]1.573/0.8751.524/0.8721.495/0.864
w. EP0.725/0.9570.753/0.9620.760/0.957
Xu et al. [52]1.895/0.8171.832/0.8181.743/0.820
w. EP0.754/0.9620.798/0.9560.822/0.949
Ours w/o. EP1.673/0.8521.628/0.8501.563/0.848
Ours0.713/0.9640.735/0.9590.732/0.959
+ +Table 4. The performance (MAE/SRC) was compared across four models, including our model, using the SMTPD dataset, both with and without EP. + +
MethodAMAEASRCMAE(day 30 only)SRC(day 30 only)
w/o. EP(1-30)1.5620.8561.5300.850
w/o. EP(2-30)1.6300.8491.5510.849
w/o. EP+[26]1.6280.8511.5550.848
ours0.7170.9590.7320.959
+ +Table 5. Assessment of EP across different scenarios. Here, "1-30" denotes the prediction target spanning a continuous sequence from the 1st day to the 30th day, as does "2-30" (to align with methods having EP). + +visualized results, tough EP exhibits high correlation to the popularities of following days, the MAE sharply increased over time. By contrast, our baseline model could well optimize the MAE and achieve even better SRC performance comparing to the natural EP curve. Therefore, the proposed baseline model is effective to utilize EP achieving better prediction accuracy. + +# 6. Conclusion + +This study aims to address the challenge of time-series popularity prediction in social media. In this paper, we introduce a novel multilingual, multi-modal time-series popularity dataset based on YouTube and suggest a multilingual temporal prediction model tailored to this dataset. Through experiments, we demonstrate the effectiveness of this approach in predicting social media popularity time-series in a multilingual environment. The experiments show that combining multi-modal features with early popularity significantly improves prediction accuracy. However, this study has yet to address challenges such as multi-frame information in videos, maximizing the use of language diversity, and deeper multi-modal exploration. These areas will be key focuses for our future work. + +# References + +[1] Deepak Agarwal, Bee-Chung Chen, Pradheep Elango, Nitin Motgi, Seung-Taek Park, Raghu Ramakrishnan, Scott Roy, and Joe Zachariah. Online models for content optimization. Advances in Neural Information Processing Systems, 21, 2008. 1 +[2] Sercan Ö Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, pages 6679-6687, 2021. 2 +[3] Peng Bao, Hua-Wei Shen, Junming Huang, and Xue-Qi Cheng. Popularity prediction in microblogging network: a case study on sina weibo. In Proceedings of the 22nd international conference on world wide web, pages 177–178, 2013. 1, 2 +[4] Shenghua Bao, Guirong Xue, Xiaoyuan Wu, Yong Yu, Ben Fei, and Zhong Su. Optimizing web search using social annotations. In Proceedings of the 16th international conference on World Wide Web, pages 501-510, 2007. 1 +[5] Laura Ceci. Youtube-statistics & facts. Statista. com [online].[cit. 2022-05-05]. Dostupné z: https://www.statista.com/topics/2019/youtube, 2022.3 +[6] Junhong Chen, Dayong Liang, Zhanmo Zhu, Xiaojing Zhou, Zihan Ye, and Xiuyun Mo. Social media popularity prediction based on visual-textual features with xgboost. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2692-2696, 2019. 2 +[7] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm SIGkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 2 +[8] Weilong Chen, Chenghao Huang, Weimin Yuan, Xiaolu Chen, Wenhao Hu, Xinran Zhang, and Yanru Zhang. Title-and-tag contrastive vision-and-language transformer for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7008-7012, 2022. 2 +[9] Xiaolu Chen, Weilong Chen, Chenghao Huang, Zhongjian Zhang, Lixin Duan, and Yanru Zhang. Double-fine-tuning multi-objective vision-and-language transformer for social media popularity prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9462-9466, 2023. 2 +[10] Wen-Huang Cheng, Sijie Song, Chieh-Yun Chen, Shintami Chusnul Hidayati, and Jiaying Liu. Fashion meets computer vision: A survey. ACM Computing Surveys (CSUR), 54 (4):1-41, 2021. 1 +[11] Minhwa Cho, Dahye Jeong, and Eunil Park. Amps: Predicting popularity of short-form videos using multi-modal attention mechanisms in social media marketing environments. Journal of Retailing and Consumer Services, 78: 103778, 2024. 2 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 4171–4186, 2019. 5 + +[13] Keyan Ding, Ronggang Wang, and Shiqi Wang. Social media popularity prediction: A multiple feature fusion approach with deep neural networks. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2682-2686, 2019. 2, 7, 8 +[14] Zhabiz Gharibshah and Xingquan Zhu. User response prediction in online advertising. aCM Computing Surveys (CSUR), 54(3):1-43, 2021. 1 +[15] Anindya Ghose and Sha Yang. An empirical analysis of search engine advertising: Sponsored search in electronic markets. Management science, 55(10):1605-1622, 2009. 1 +[16] Sayan Ghosh, Dylan Baker, David Jurgens, and Vinodkumar Prabhakaran. Detecting cross-geographic biases in toxicity modeling on social media. In Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT), pages 313–328, 2021. 4 +[17] Marcos Andre Gonçalves, Jussara M Almeida, Luiz GP dos Santos, Alberto HF Laender, and Virgilio Almeida. On popularity in the blogosphere. IEEE Internet Computing, 14(3): 42-49, 2010. 1 +[18] Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. Learning word vectors for 157 languages. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), 2018. 5 +[19] Mohammad Hajarian, Mark Anthony Camilleri, Paloma Díaz, and Ignacio Aedo. A taxonomy of online marketing methods. In Strategic corporate communication in the digital age, pages 235-250. Emerald Publishing Limited, 2021. 1 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 6 +[22] Chih-Chung Hsu, Chia-Ming Lee, Xiu-Yu Hou, and Chi-Han Tsai. Gradient boost tree network based on extensive feature analysis for popularity prediction of social posts. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9451-9455, 2023. 2 +[23] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems, 30, 2017. 2 +[24] Aditya Khosla, Atish Das Sarma, and Raffay Hamid. What makes an image popular? In Proceedings of the 23rd international conference on World wide web, pages 867-876, 2014. 3 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations(ICLR), San Diego, California, USA, May 7-9, 2015. 7 +[26] Xin Lai, Yihong Zhang, and Wei Zhang. Hyfea: Winning solution to social media popularity prediction for multimedia grand challenge 2020. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4565-4569, 2020. 2, 7, 8 + +[27] Chenyu Li, Jun Liu, and Shuxin Ouyang. Analysis and prediction of content popularity for online video service: a youku case study. China Communications, 13(12):216-233, 2016. 3 +[28] Jinning Li, Yirui Gao, Xiaofeng Gao, Yan Shi, and Guihai Chen. Senti2pop: sentiment-aware topic popularity prediction on social media. In 2019 IEEE International conference on data mining (ICDM), pages 1174-1179. IEEE, 2019. 2 +[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2 +[30] Yury Lifshits. Ediscope: Social analytics for online news. Yahoo Labs, 2010. 1 +[31] Hung-Hsiang Lin, Jiun-Da Lin, Jose Jaena Mari Ople, Jun-Cheng Chen, and Kai-Lung Hua. Social media popularity prediction based on multi-modal self-attention mechanisms. IEEE Access, 10:4448-4455, 2022. 2 +[32] An-An Liu, Xiaowen Wang, Ning Xu, Junbo Guo, Guoqing Jin, Quan Zhang, Yejun Tang, and Shenyuan Zhang. A review of feature fusion-based media popularity prediction methods. Visual Informatics, 2022. 2 +[33] Marco Lui and Timothy Baldwin. langid.py: An off-the-shelf language identification tool. In Proceedings of the ACL 2012 System Demonstrations, pages 25-30, 2012. 6 +[34] Shijian Mao, Wudong Xi, Lei Yu, Gaotian Lü, Xingxing Xing, Xingchen Zhou, and Wei Wan. Enhanced catboost with stacking features for social media prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9430-9435, 2023. 2 +[35] Masoud Mazloom, Robert Rietveld, Stevan Rudinac, Marcel Worring, and Willemijn Van Dolen. Multimodal popularity prediction of brand-related social media posts. In Proceedings of the 24th ACM international conference on Multimedia, pages 197-201, 2016. 2 +[36] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proceedings of the 1st International Conference on Learning Representations (ICLR), Scottsdale, Arizona, USA, May 2-4, 2013. 5 +[37] Alessandro Ortiz, Giovanni Maria Farinella, and Sebastiano Battiato. Prediction of social image popularity dynamics. In Image Analysis and Processing-ICIAP 2019: 20th International Conference, Trento, Italy, September 9–13, 2019, Proceedings, Part II 20, pages 572–582. Springer, 2019. 3 +[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 5 +[39] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. Advances in neural information processing systems, 31, 2018. 7 +[40] Cédric Richier, Eitan Altman, Rachid Elazouzi, Tania Altman, Georges Linares, and Yonathan Portilla. Modelling + +view-count dynamics in youtube. CoRR, abs/1404.2570, 2014. 2 +[41] Cedric Richier, Eitan Altman, Rachid Elazouzi, Tania Jimenez, Georges Linares, and Yonathan Portilla. Bioinspired models for characterizing youtube viewcout. In 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 297-305. IEEE, 2014. 1, 2 +[42] Satoshi Sanjo and Marie Katsurai. Recipe popularity prediction with deep visual-semantic fusion. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2279–2282, 2017. 1, 2 +[43] Gabor Szabo and Bernardo A Huberman. Predicting the popularity of online content. Communications of the ACM, 53 (8):80-88, 2010. 1, 2 +[44] YunPeng Tan, Fangyu Liu, BoWei Li, Zheng Zhang, and Bo Zhang. An efficient multi-view multimodal data processing framework for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7200-7204, 2022. 2, 7 +[45] Alexandru Tatar, Marcelo Dias De Amorim, Serge Fdida, and Panayotis Antoniadis. A survey on predicting the popularity of web content. Journal of Internet Services and Applications, 5(1):1-20, 2014. 1 +[46] Kai Wang, Penghui Wang, Xin Chen, Qiushi Huang, Zhendong Mao, and Yongdong Zhang. A feature generalization framework for social media popularity prediction. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4570-4574, 2020. 2, 7 +[47] Bo Wu, Wen-Huang Cheng, Yongdong Zhang, Qiushi Huang, Jintao Li, and Tao Mei. Sequential prediction of social media popularity with deep temporal context networks. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI), pages 3062-3068, 2017. 1, 2 +[48] Bo Wu, Wen-Huang Cheng, Peiye Liu, Bei Liu, Zhaoyang Zeng, and Jiebo Luo. Smp challenge: An overview of social media prediction challenge 2019. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2667-2671. ACM, 2019. 1, 2, 3 +[49] Bo Wu, Peiye Liu, Wen-Huang Cheng, Bei Liu, Zhaoyang Zeng, Jia Wang, Qiushi Huang, and Jiebo Luo. Smp challenge: An overview and analysis of social media prediction challenge. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9651-9655, 2023. 2 +[50] Jianmin Wu, Liming Zhao, Dangwei Li, Chen-Wei Xie, Siyang Sun, and Yun Zheng. Deeply exploit visual and language information for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7045–7049, 2022. 2, 7 +[51] Leilei Wu, Lingling Yi, Xiao-Long Ren, and Linyuan Lu. Predicting the popularity of information on social platforms without underlying network structure. Entropy, 25(6):916, 2023. 2 +[52] Kele Xu, Zhimin Lin, Jianqiao Zhao, Peicang Shi, Wei Deng, and Huaimin Wang. Multimodal deep learning for social media popularity prediction with attention mechanism. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4580-4584, 2020. 2, 7, 8 + +[53] Jaewon Yang and Jure Leskovec. Patterns of temporal variation in online media. In Proceedings of the fourth ACM international conference on Web search and data mining, pages 177-186, 2011. 1, 2 +[54] Zhizhen Zhang, Xiaohui Xie, Mengyu Yang, Ye Tian, Yong Jiang, and Yong Cui. Improving social media popularity prediction with multiple post dependencies. CoRR, abs/2307.15413, 2023. 2 +[55] Alireza Zohourian, Hedieh Sajedi, and Arefeh Yavary. Popularity prediction of images and videos on instagram. In 2018 4th International Conference on Web Research (ICWR), pages 111-117, 2018. 3 \ No newline at end of file diff --git a/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/images.zip b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..024fece061764204399a22576708326488601879 --- /dev/null +++ b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c94ee50e10e1b3416199bcbbab596485357ca65dda9430a456c9e739c791d3b +size 431396 diff --git a/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/layout.json b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46e8abf4988d5419176ef7b7ba674d8bf9069491 --- /dev/null +++ b/2025/SMTPD_ A New Benchmark for Temporal Prediction of Social Media Popularity/layout.json @@ -0,0 +1,9381 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 553, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 553, + 122 + ], + "type": "text", + "content": "SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Yijie Xu*, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Bolun Zheng*, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Wei Zhu*, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Hangjia Pan, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Yuchen Yao, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Ning Xu, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Anan Liu, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Quan Zhang, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Chenggang Yan " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Hangzhou Dianzi University, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Tianjin University, " + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 142, + 471, + 187 + ], + "type": "text", + "content": "Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 238, + 297, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 297, + 454 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 297, + 454 + ], + "type": "text", + "content": "Social media popularity prediction task aims to predict the popularity of posts on social media platforms, which has a positive driving effect on application scenarios such as content optimization, digital marketing and online advertising. Though many studies have made significant progress, few of them pay much attention to the integration between popularity prediction with temporal alignment. In this paper, with exploring YouTube's multilingual and multi-modal content, we construct a new social media temporal popularity prediction benchmark, namely SMTPD, and suggest a baseline framework for temporal popularity prediction. Through data analysis and experiments, we verify that temporal alignment and early popularity play crucial roles in social media popularity prediction for not only deepening the understanding of temporal dynamics of popularity in social media but also offering a suggestion about developing more effective prediction models in this field. Code is available at https://github.com/zhuwei321/SMTPD" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 135, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 603 + ], + "type": "text", + "content": "With the advancement of Internet communication technology in recent years, social media has gradually emerged and has influenced various aspects of human life. Any content posted on social media stands the chance of becoming hot spot, and widely disseminated social media content can generate significant social and economic benefits. The prediction of social media popularity holds immense potential applications in content optimization [1, 30], online advertising [14, 15], digital marketing [19, 45], search recommendations [4, 17], intelligent fashion [10], and beyond." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 664 + ], + "type": "text", + "content": "In the early stages of popularity research, statistical and topological methods [41, 43, 53] were extensively employed, with the research focus primarily on time-aware popularity prediction. These prediction methods rely on information about the earlier popularity of posted content." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 212, + 558, + 488 + ], + "blocks": [ + { + "bbox": [ + 319, + 212, + 558, + 488 + ], + "lines": [ + { + "bbox": [ + 319, + 212, + 558, + 488 + ], + "spans": [ + { + "bbox": [ + 319, + 212, + 558, + 488 + ], + "type": "image", + "image_path": "a3265d56d3056fdfb495ea22510f3d30c86ae82ee55356ad25a5632066d9d33f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 553 + ], + "type": "text", + "content": "Figure 1. Content sections and popularity trend of SMTPD. In 1a, a sample involves four sections of content, with temporal popularity. 1b depicts box plots of daily popularity scores, illustrating variations in popularity distribution at different time points. The distribution consistently demonstrates a decay pattern over time." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 561, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 669 + ], + "type": "text", + "content": "Along with the many large-scale social media popularity datasets are proposed [3, 42, 47, 48], the machine learning based methods are widely employed to achieve reasonable predictions of social media popularity and have shown remarkable performance. These methods emphasize meticulous design in feature extraction and fusion, followed by popularity prediction using machine learning models such as deep neural networks (DNN) and gradient boosting decision trees (GBDT)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Table 1 lists the public or mentioned social media popularity datasets in the most recent years as far as we know. For above machine learning based methods, the large-scale" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 673, + 227, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 673, + 227, + 684 + ], + "spans": [ + { + "bbox": [ + 69, + 673, + 227, + 684 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 69, + 684, + 148, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 684, + 148, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 684, + 148, + 693 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China under Grant 2020YFB1406600." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18847" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 161 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 161 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 161 + ], + "type": "table", + "html": "
DatasetSourceCategorySamplesLanguagePrediction Type
Mazloom [35]Instagramfast food brand75KEnglishsingle
Sanjo [42]Cookpadrecipe150KJapanesesingle
TPIC17 [47]Flickr-680KEnglishsingle
SMPD [48]Flickr11 categories486KEnglishsingle
AMPS [11]YouTubeshorts13KKoreansingle
SMTPD (ours)YouTube15 categories282Kover 90 languagessequential
", + "image_path": "295fcc04cec6a713a1912410cf72fe2107d4f153f9a5a6d9d33c0e6a7ab889bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "type": "text", + "content": "Table 1. Comparing SMTPD with existed multi-modal social media popularity datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 201, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 227 + ], + "type": "text", + "content": "social media popularity datasets [3, 11, 35, 42, 47, 48] provide the fundamental supports." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 312 + ], + "type": "text", + "content": "However, existing datasets and related studies have notable limitations, such as insufficient multi-modal data, limited language diversity, and, crucially, the absence of a consistent timeline for prediction. As shown in Figure1b, popularity distributions shift over time, and predictions made at irregular intervals can reduce accuracy and create challenges for practical applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "content": "Furthermore, effective popularity prediction requires integrating both the multi-modal content of social media and information on the communication process to capture the dynamic trends in popularity. This essential aspect is missing in most current prediction efforts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 651 + ], + "type": "text", + "content": "Targeting to fix up the above shortcomings of existed datasets, in this paper, we propose a new benchmark called Social Media Temporal Popularity Dataset (SMTPD) by observing multi-modal content from mainstream for cutting-edge research in the field of social multimedia along with multi-modal feature temporal prediction. As shown in Figure 1a, multi-modal contents of social media posts in multiple languages generated the daily popularity information during the communication process. We define such a post as a sample of SMTPD, and propose a multi-modal framework as a baseline to achieve the temporal prediction of popularity scores. The proposed framework consists of two parts, feature extraction and regression. In the feature extraction part, multiple pre-training models and pre-processing methods are introduced to translate the multi-modal content into deep features. In the regression part, we encode the state and time sequence of the extracted features, and adopt the LSTM-based structure to regress the 30-day popularities. Through analysis and experiments, we discover the importance of early popularity to the task of popularity prediction, and demonstrate the effectiveness of our method in temporal prediction. Generally, the contribution of this work can be summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "- Against the missing of temporal information in social media popularity researches, we observe over 282K multilingual samples from mainstream social media since they released on the network lasting for 30 days. We refer to these samples as SMTPD, a new benchmark for temporal" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 73, + 412, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 73, + 412, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 412, + 84 + ], + "type": "text", + "content": "popularity prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 85, + 553, + 217 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 85, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 85, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 85, + 553, + 155 + ], + "type": "text", + "content": "- Basing on existed methods, we innovate in both the selection of feature extractors and the construction of the temporal regression component, and suggest a baseline model which enables temporal popularity prediction to be conducted across multiple languages while aligning prediction times." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 156, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 217 + ], + "type": "text", + "content": "- Exploring the popularity distribution and the correlation between popularity at different times. Based on these, We find the importance of early popularity for popularity prediction task, and point out that the key-point for predicting popularity is to accurately predict the early popularity." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 227, + 398, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 227, + 398, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 227, + 398, + 239 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "text", + "content": "Statistical and topological methods. Szabo et al. [43] found that the early-stage popularity notably influences subsequent popularity. Yang et al. [53] conducted cluster analysis on temporal patterns within online content, revealing distinct characteristics of popularity variations across different clusters. Richier et al. [40, 41] proposes bio-inspired models to characterize the evolution of video view counts. Wu et al. [51] suggested that the information cascade process is best described as an activate-decay dynamic process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 355, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 553, + 594 + ], + "type": "text", + "content": "Multi-modal feature based Methods. Ding et al. [13] use pre-trained ResNet and BERT to extract visual and textual features, with the DNN regression. Xu et al. [52] and Lin et al. [31] adopted attention mechanisms to effectively integrate multi-modal features. Chen et al. [6] compared the performance of several regression models, among which XGBoost [7] exhibited the best results. Hsu et al. [22] employed LightGBM [23] and TabNet [2] to capture intricate semantic relationships in multi-modal features. Lai et al. [26] engineered handcrafted features, exploiting CatBoost for regression. Mao et al. [34] enhanced CatBoost-based model by stacking features. Tan et al. [44] extracted visual-textual features by ALBEF [29] and Chen et al. [8, 9] enriched more intermodal features to promote predictive performance. Wu et al. [50] emphasized increasing feature dimensions to improving predictive performance. The post dependencies captured by sliding window average [46] and DSN [54] has also led to improvements. Many of these multi-modal methods are based on the SMPD and working out great [32, 49]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 701 + ], + "type": "text", + "content": "Social media popularity datasets. Mazloom et al. [35] conducted experiments using a dataset of posts related to fast-food brands collected from Instagram. Sanjo et al. [42] provided a recipe popularity prediction dataset based on Cookpad, which includes text content entirely in Japanese. Li et al. [28] predicted the future popularity of topics by using historical sentiment information based on Twitter's text data records. TPIC17 [47] and SMPD [48] are datasets for single popularity prediction task based on Flickr." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "We believe that the loss of temporal alignment about" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18848" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "popularity in existing methods is a common problem. The image dynamic popularity dataset [37] gave us insights, so we propose a multi-modal temporal popularity benchmark to address the shortcomings of existing studies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 154, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 154, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 154, + 143 + ], + "type": "text", + "content": "3. SMTPD Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 150, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 296, + 402 + ], + "type": "text", + "content": "With the great development of web communication technologies, social multimedia has become the most popular media around the world. However, most existing social media prediction datasets are built on the basis of single-output prediction, meaning they observe existing posts and infer their popularity based on their posting time. As mentioned earlier, these data without aligned posting times exhibit non-uniform popularity distributions. Therefore, SMTPD keep an eye on one of the most popular worldwide social media, YouTube [5], with over 282K samples, primarily focusing on the evolution of popularity over time for these samples. We collected over 402K raw data samples. Specifically, we first removed records with missing values—which likely resulted from network issues during data acquisition. Next, we eliminated samples that either failed to be crawled on certain days or had been deleted within 30 days. Finally, we filtered out potential outliers using the " + }, + { + "bbox": [ + 55, + 150, + 296, + 402 + ], + "type": "inline_equation", + "content": "3\\sigma" + }, + { + "bbox": [ + 55, + 150, + 296, + 402 + ], + "type": "text", + "content": " rule. In this section, we provide a comprehensive overview of the SMTPD's composition. Additionally, we present various data analysis results that contribute to the feasibility of our approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 557 + ], + "type": "text", + "content": "Unlike the retrospective methods used in previous datasets, we take note of the newly posted multi-modal content and then observe popularity information every 24 hours via YouTube API. Considering that in real-world applications, the posts to be predicted are always new, this data observation method allows us to obtain aligned temporal popularity of new posts, which is more in line with practical applications. Figure 1a show a sample from SMTPD. We note diverse attributes, including visual content, textual content, metadata, and user profiles. Additionally, we focus on the temporal popularity for each sample within 30 days after release. We perform basic statistics for SMTPD, as shown in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 176, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 176, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 176, + 577 + ], + "type": "text", + "content": "3.1. Temporal Popularity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "The utilization of temporal popularity data is versatile. It not only serves as the output for predictions but also as inputs, where a segment preceding a specific time point, coupled with multi-modal content, can forecast subsequent popularity. We analyze the distribution and the correlation of popularity over time, so as to facilitate a comprehensive understanding of the significance of temporal popularity for predictions from different perspectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "A number of popularity definitions have been born from previous work [27, 55]. We use Khosla's popularity score transformation [24] as it can normalize the distribution of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "view counts to a suitable popularity score's distribution. It can be represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 397, + 102, + 555, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 102, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 397, + 102, + 555, + 122 + ], + "type": "interline_equation", + "content": "p = \\log_ {2} (\\frac {v}{d} + 1) \\tag {1}", + "image_path": "997eb873c2947a6919bcd4941ef9d8995799ced22a13bd6addac00a704c5c24c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "text", + "content": " is the popularity score, " + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "text", + "content": " is the view counts of a sample, " + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 128, + 555, + 210 + ], + "type": "text", + "content": " represents the corresponding number of days after the post's release. As shown in Figure 1, the histogram of view counts reveals an extreme long-tailed distribution, which might not be suitable as a prediction target. It's evident that the distribution of the popularity score metric becomes more reasonable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 211, + 554, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 554, + 282 + ], + "type": "text", + "content": "In addition, we notice that there is a strong correlation between the popularity scores of different days. The Pearson Correlation (PC) and Spearman Ranking Correlation (SRC) can respectively reflect the linear relationship and rank-order correlation between popularity scores of different days as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 298, + 555, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 298, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 339, + 298, + 555, + 334 + ], + "type": "interline_equation", + "content": "P C = \\frac {\\sum_ {i = 1} ^ {n} \\left(X _ {i} - \\bar {X}\\right) \\left(Y _ {i} - \\bar {Y}\\right)}{\\sqrt {\\sum_ {i = 1} ^ {n} \\left(X _ {i} - \\bar {X}\\right) ^ {2}} \\sqrt {\\sum_ {i = 1} ^ {n} \\left(Y _ {i} - \\bar {Y}\\right) ^ {2}}} \\tag {2}", + "image_path": "5d6b570874471f4dd50449ff086c6842d90033c495267ed9528892c31c132b3b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 340, + 554, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 340, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 353, + 340, + 554, + 370 + ], + "type": "interline_equation", + "content": "S R C = \\frac {1}{n - 1} \\sum_ {i = 1} ^ {n} \\left(\\frac {X _ {i} - \\bar {X}}{\\sigma_ {X}}\\right) \\left(\\frac {Y _ {i} - \\bar {Y}}{\\sigma_ {Y}}\\right) \\tag {3}", + "image_path": "6105f0c53ca4cf8be5b2ddb13e0e9ad557e9930908491a6119d5ce04965b741c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": " denote the popularity scores belonging to 2 different days, while " + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\bar{X}" + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\sigma_{X}" + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": " denote the mean and standard deviation of " + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": ", and similarly for " + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 373, + 554, + 433 + ], + "type": "text", + "content": ". PC and SRC respectively reflect the linear relationship and rank-order correlation between popularity scores of different days." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 434, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 494 + ], + "type": "text", + "content": "The Pearson Correlation (PC) and Spearman Rank Correlation (SRC) heat maps illustrate internal relationships in daily popularity scores, establishing a basis for prediction based on these temporal patterns. Further details are available in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 499, + 555, + 669 + ], + "blocks": [ + { + "bbox": [ + 315, + 499, + 555, + 669 + ], + "lines": [ + { + "bbox": [ + 315, + 499, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 315, + 499, + 555, + 669 + ], + "type": "table", + "html": "
StatisticsSMTPDSMPD (Train/Test)
Number of samples282.4K305K/181K
Number of users152.7K38K/31K
Mean popularity5.956.41/5.12
STD popularity4.152.47/2.41
Number of categories1511
Number of custom tags960K250K
Average length of title53.4 chars29 words
Mean duration1853.4 s-
", + "image_path": "e56ccbd802427fb35da9cf6e6f7a2e877a5a00b75a64abe6a98d641ab57b7640.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "type": "text", + "content": "Table 2. The basic statistics of all samples in SMPD, comparing to SMPD[48]. Due to the multilingual environment, we use chars (characters) to describe the average length of title." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18849" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 180, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 180, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 180, + 83 + ], + "type": "text", + "content": "3.2. Multi-Modal Content" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 138 + ], + "type": "text", + "content": "Multi-modal feature fusion is the mainstream approach in today's popularity prediction methods. In this section, we will divide the data into corresponding modality and elucidate the multi-modal content in SMTPD." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 148, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 148, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 148, + 154 + ], + "type": "text", + "content": "3.2.1. Visual Content" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 158, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 243 + ], + "type": "text", + "content": "SMTPD samples, primarily drawn from a multi-modal social media platform, are heavily influenced by visual elements. Since metrics like view count and popularity score are often based on the cover frame that viewers see before clicking to play, we use the cover frame as the primary visual content. This approach balances popularity impact with copyright considerations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 248, + 153, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 153, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 153, + 259 + ], + "type": "text", + "content": "3.2.2. Textual Content" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 263, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 361 + ], + "type": "text", + "content": "The textual content in posts plays a crucial role in media dissemination and user engagement. We examine several core elements, including category, title, description, and hashtags (denoted by the “#” symbol), along with the user ID (user nickname) from profile metadata. YouTube defines 15 distinct video categories, all in English, which support content organization and discovery. Figure 3 shows the statistical distribution of content by category." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 369, + 154, + 467 + ], + "blocks": [ + { + "bbox": [ + 56, + 369, + 154, + 467 + ], + "lines": [ + { + "bbox": [ + 56, + 369, + 154, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 154, + 467 + ], + "type": "image", + "image_path": "7bb60ec655ff6911907a94c33cdcbf578ca08183d4eafdc051bf26f2de1f3ea3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 468, + 146, + 478 + ], + "lines": [ + { + "bbox": [ + 85, + 468, + 146, + 478 + ], + "spans": [ + { + "bbox": [ + 85, + 468, + 146, + 478 + ], + "type": "text", + "content": "(a) The PC Matrix" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 369, + 296, + 466 + ], + "blocks": [ + { + "bbox": [ + 156, + 369, + 296, + 466 + ], + "lines": [ + { + "bbox": [ + 156, + 369, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 156, + 369, + 296, + 466 + ], + "type": "image", + "image_path": "4f299ea09a29b30b6b778651443af19eeca7beb6115a9796887f5f3cb7516c23.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 468, + 269, + 476 + ], + "lines": [ + { + "bbox": [ + 203, + 468, + 269, + 476 + ], + "spans": [ + { + "bbox": [ + 203, + 468, + 269, + 476 + ], + "type": "text", + "content": "(b) The SRC Matrix" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 536, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 584 + ], + "type": "text", + "content": "As an international platform, YouTube hosts user-defined textual content in multiple languages beyond standard categories. Figure 4 illustrates the distribution and popularity bias across title languages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 585, + 296, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 296, + 669 + ], + "type": "text", + "content": "In previous researches, the datasets typically had uniform language for text content, without considering multilingual aspects. While this simplifies modeling for prediction methods, it also poses significant limitations for international social media platforms which not restricted to a single language. Besides, predictions based on a few languages may introduce biases [16] against popularity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 674, + 157, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 157, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 157, + 685 + ], + "type": "text", + "content": "3.2.3. Numerical values" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Numerical attributes frequently function as supplementary factors in enhancing prediction accuracy. In this study, we" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "examined key attributes, including video duration, uploader follower count, and the number of posts by the uploader. Additionally, we manually derived several auxiliary metrics, such as title length, the count of custom tags, and description length, to further enrich the feature set and support predictive robustness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 157, + 492, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 492, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 492, + 172 + ], + "type": "text", + "content": "4. Temporal popularity prediction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 178, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 275 + ], + "type": "text", + "content": "The multi-modal feature-based temporal prediction framework is shown in Figure 5, which is divided into two main parts: multi-modal feature extraction and temporal popularity score regression. This framework is designed to adapt to the multilingual temporal prediction tasks under time alignment. The prediction target is set as the popularity of samples, which corresponds to the temporal popularity within 30 days after the sample's release." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 284, + 489, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 489, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 489, + 296 + ], + "type": "text", + "content": "4.1. Multi-Modal Feature Extraction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 301, + 554, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 351 + ], + "type": "text", + "content": "For different modalities in SMTPD, we adopted distinct feature extraction methods. We additionally incorporated features from the categorical modality to investigate the impact of categorical features on popularity prediction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 358, + 409, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 409, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 409, + 369 + ], + "type": "text", + "content": "4.1.1. Visual Features" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 374, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 411 + ], + "type": "text", + "content": "The cover image of a social media content is a very important component which would help users to quickly understand what they would see if clicking to view this content." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 316, + 418, + 539, + 529 + ], + "blocks": [ + { + "bbox": [ + 55, + 487, + 296, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 521 + ], + "type": "text", + "content": "Figure 2. The heat map of daily popularity correlations. The figure clearly shows that there is a high degree of correlations in popularity between consecutive days." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 418, + 539, + 529 + ], + "lines": [ + { + "bbox": [ + 316, + 418, + 539, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 539, + 529 + ], + "type": "image", + "image_path": "26bd9072261138410e4150ff6346bb7603c8c0ffcb161721b2cb606bf9b3cd98.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 316, + 542, + 552, + 656 + ], + "blocks": [ + { + "bbox": [ + 362, + 530, + 507, + 541 + ], + "lines": [ + { + "bbox": [ + 362, + 530, + 507, + 541 + ], + "spans": [ + { + "bbox": [ + 362, + 530, + 507, + 541 + ], + "type": "text", + "content": "(a) The statistics of samples in each category" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 542, + 552, + 656 + ], + "lines": [ + { + "bbox": [ + 316, + 542, + 552, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 552, + 656 + ], + "type": "image", + "image_path": "cef5d7564335d9ed167cf9ab5edb8f076c9a49a50bf4964b5854ff49d3395ad2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 658, + 499, + 668 + ], + "lines": [ + { + "bbox": [ + 370, + 658, + 499, + 668 + ], + "spans": [ + { + "bbox": [ + 370, + 658, + 499, + 668 + ], + "type": "text", + "content": "(b) Average popularity of each category" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 677, + 554, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 711 + ], + "type": "text", + "content": "Figure 3. The statistics based on category. 3a counts the number of samples in each category, and 3b shows the average popularity score of samples in each category." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18850" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 294, + 208 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 294, + 208 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 294, + 208 + ], + "type": "image", + "image_path": "adedc1fccad5168bb90b96d8b059cb0fc4848ca40084985b77bd1c06cdf3d532.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 296, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 271 + ], + "type": "text", + "content": "Figure 4. Languages analysis. The left is the proportions of these languages, with \"Other\" encompassing 90 different languages. The right represents the average popularity in different languages, revealing the geographic biases. The languages of samples are counted by reference to the title." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "type": "text", + "content": "Therefore, the cover image would directly affect the popularity of a social media content. The semantic information provided in the cover image plays a key role for user's understanding. We adopt convolutional neural networks ResNet-101 [20] pre-trained on ImageNet as our visual feature extraction model to obtain the semantic information provided by the cover. To match the input size of ResNet-101, we re-scale the size of cover image to " + }, + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "type": "text", + "content": ". The 2048-dimension feature vector before the final classification layer will be selected as the final visual feature " + }, + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 280, + 296, + 400 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 408, + 294, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 408, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 134, + 408, + 294, + 421 + ], + "type": "interline_equation", + "content": "f _ {v} = \\operatorname {R e s N e t} (S (I)) \\tag {4}", + "image_path": "ae42011fcb8a86ff71d0ffc730b018cbc829bfc911e57362fac456e883388d21.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "type": "text", + "content": " denotes the cover image and " + }, + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 430, + 296, + 490 + ], + "type": "text", + "content": " denotes the re-scale operation. However, the cover image would not always be uploaded by authors or be unavailable due to the network transmission. For those contents missing the cover image, we use blank images with all pixels are set to zero instead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 155, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 155, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 155, + 506 + ], + "type": "text", + "content": "4.1.2. Textual Features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 569 + ], + "type": "text", + "content": "Textual information significantly impacts a user's choice to view or skip content. Key textual inputs include category, title, tags, description, and user profile ID. Extracting semantic features from these inputs is crucial for accurate popularity prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 641 + ], + "type": "text", + "content": "However, unlike previous popularity prediction tasks, one distinguishing characteristic of SMTPD is that its text content includes multiple languages, and many samples contain text in more than one language. This makes it challenging to use many pre-trained word vector models based on single-language corpora, such as [18, 36, 38]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 689 + ], + "type": "text", + "content": "Thanks to the multilingual capability of BERT-Multilingual [12], which processes text across languages in a single model, we use it to extract a 768-dimensional feature vector for each text, capturing semantic information as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 139, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "f _ {t} ^ {k} = \\operatorname {B E R T} \\left(T ^ {k}\\right) \\tag {5}", + "image_path": "75760e39333d91920db59378f54f2da50bbd7cebcf00e9b5a0710c608aabe4bf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 73, + 533, + 277 + ], + "blocks": [ + { + "bbox": [ + 334, + 73, + 533, + 277 + ], + "lines": [ + { + "bbox": [ + 334, + 73, + 533, + 277 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 533, + 277 + ], + "type": "image", + "image_path": "750954a6f40a9685f6a6273ec7d4849189e064e6f84730d43c954b4deb1b5ea4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 290, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 369 + ], + "type": "text", + "content": "Figure 5. The proposed model consists of two layers. The upper layer processes visual, textual, numerical, and categorical features, which are input to a multi-layer perceptron (MLP). The MLP outputs are concatenated and fed into a lower LSTM layer for regression. Each LSTM cell's output is passed through another MLP, with the final MLP outputs combined to generate a time prediction sequence using the ReLU activation function." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "inline_equation", + "content": "k \\in \\{cat, tit, tag, des, uid\\}" + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "content": " maps to category, title, tags, description, and user ID. These are combined into a " + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "inline_equation", + "content": "768 \\times 5" + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "content": " matrix, transformed through " + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "content": " convolution to produce the 768D textual feature " + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 377, + 555, + 437 + ], + "type": "text", + "content": ", which can be expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 442, + 555, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 442, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 340, + 442, + 555, + 457 + ], + "type": "interline_equation", + "content": "f _ {t} = \\operatorname {C o n v} _ {5 \\times 5} \\left(\\mathcal {C} \\left(f _ {t} ^ {\\text {c a t}}, f _ {t} ^ {\\text {t i t}}, f _ {t} ^ {\\text {t a g}}, f _ {t} ^ {\\text {d e s}}, f _ {t} ^ {\\text {u i d}}\\right)\\right) \\tag {6}", + "image_path": "d3a744f56bb1c3923e35b10ca2d09d4493291cd96200540849f2be15d01e716f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 463, + 504, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 504, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 504, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 463, + 504, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 463, + 504, + 475 + ], + "type": "text", + "content": " denotes the above combing operation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 479, + 427, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 427, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 427, + 490 + ], + "type": "text", + "content": "4.1.3. Numerical Features" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 494, + 555, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 637 + ], + "type": "text", + "content": "Numerical metrics directly influence whether an audience clicks to watch which sample on social media platform. Also, the number of followers and the number of posts also impact the process through which an audience navigates from content uploaders' user pages to click and watch videos. Hence, we selected these influential numerical metrics for numerical features extraction. As the number of user followers share a similar distribution pattern that exhibiting long-tailed characteristics as view counts, it is necessary to apply a logarithmic transformation before feeding it into the network. Afterward, all numerical items are Z-score normalized and concatenated as numerical features:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 400, + 641, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 641, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 400, + 641, + 553, + 669 + ], + "type": "interline_equation", + "content": "f _ {n} ^ {j} = \\frac {T ^ {j} - \\mu_ {j}}{\\sigma_ {j}} \\tag {7}", + "image_path": "b6faea7612c4197f0bc07d7f8711ed542f82d6d558e43b627827ffe26560dffc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 327, + 669, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 669, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 327, + 669, + 553, + 685 + ], + "type": "interline_equation", + "content": "f _ {n} = \\operatorname {C o n c a t} \\left(f _ {n} ^ {f o l}, f _ {n} ^ {p o s}, f _ {n} ^ {d u r}, f _ {n} ^ {t l}, f _ {n} ^ {t n}, f _ {n} ^ {d l}, f _ {n} ^ {E P}\\right) \\tag {8}", + "image_path": "a88f7555c6e2f782b5867c553a448aa2e0a9582a1c99b34bc15a6c62f50058e8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "j \\in \\{fol, pos, dur, tl, tn, dl, EP\\}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " represents the log-scaled follower counts, total number of posts, video du" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18851" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "ration, title length, the number of tags, description length and the early popularity (the 1st day's popularity). The " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " represents the mean and standard deviation of all samples. And Concat denotes the horizontal concatenating operation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 173, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 173, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 173, + 154 + ], + "type": "text", + "content": "4.1.4. Categorical Features" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 374 + ], + "type": "text", + "content": "The categorical disparities across social media platforms are depicted in Figure 3b. Beyond text-based feature extraction, these disparities inform the generation of categorical features. Language functions as a key marker of regional and cultural distinctions among content creators, as highlighted by the language bias observed in Figure 4. Consequently, language classification features are crucial for enhancing prediction accuracy. To develop these classification features, two sub-features are selected: label and language attributes. The label feature captures the semantic category, while the language attribute is derived from the language in the video title. Assuming that the primary language of all text content matches the title's language, we use langid [33] for classification. Two embedding layers are introduced to encode label and language features independently, each processed by a dedicated multi-layer perceptron (MLP). Finally, the outputs are combined via cumulative multiplication to form the classification feature " + }, + { + "bbox": [ + 55, + 158, + 296, + 374 + ], + "type": "inline_equation", + "content": "f_{c}" + }, + { + "bbox": [ + 55, + 158, + 296, + 374 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 385, + 295, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 385, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 66, + 385, + 295, + 399 + ], + "type": "interline_equation", + "content": "f _ {c} = \\operatorname {M L P} _ {1} \\left(\\mathrm {E} _ {1} \\left(T ^ {\\text {c a t}}\\right)\\right) \\odot \\operatorname {M L P} _ {2} \\left(\\mathrm {E} _ {2} \\left(\\operatorname {l a n g i d} \\left(T ^ {\\text {t i t}}\\right)\\right)\\right) \\tag {9}", + "image_path": "5aff61bf2f3fa4fa5ab419df0523ed679f5a1b5be160d5278ff4db9618ea9472.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_1" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_2" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": " represent two independent embedding layers, and " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_1" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_2" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": " refer to multi-layer perceptrons applied to the label embedding " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_1(T^{cat})" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": " and the language embedding " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_2(\\text{langid}(T^{tit}))" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": ", respectively. The operator " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": " denotes element-wise multiplication, which combines the outputs of the two MLPs to produce the final classification feature " + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "inline_equation", + "content": "f_c" + }, + { + "bbox": [ + 55, + 413, + 296, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 149, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 149, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 149, + 517 + ], + "type": "text", + "content": "4.1.5. Feature Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 559 + ], + "type": "text", + "content": "After extraction, features from different modalities are aligned in dimension through multi-layer perceptrons (MLPs) and then concatenated into fused features " + }, + { + "bbox": [ + 55, + 522, + 296, + 559 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 522, + 296, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 230, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 230, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 230, + 583 + ], + "type": "text", + "content": "4.2. Temporal Popularity Regression" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 243, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 243, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 243, + 601 + ], + "type": "text", + "content": "4.2.1. Sequential Encoding and Adjustment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Considering the high correlation between popularity values on adjacent days in the ground truth, we employ a LSTM [21] structure to get the temporal variations in popularity. As shown in Figure 5, multiple LSTM cells are constructed to transmit temporal information. The initial states (both hidden state and cell state) of the LSTM are generated by passing the " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " through the same MLP as state encoding. Additionally, the input for LSTM cell at each time step is constructed from " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " through temporal encoding MLPs. These" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 73, + 477, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 477, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 477, + 84 + ], + "type": "text", + "content": "two encoding process can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 88, + 553, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 88, + 553, + 102 + ], + "spans": [ + { + "bbox": [ + 388, + 88, + 553, + 102 + ], + "type": "interline_equation", + "content": "h _ {0} = c _ {0} = \\mathrm {M L P} ^ {\\mathrm {h c}} (F) \\tag {10}", + "image_path": "0ebaf4701352fc3bc26f7d55fad8d9f40cb31a94e9173a45c4cf51500065eb76.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 400, + 104, + 553, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 104, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 400, + 104, + 553, + 117 + ], + "type": "interline_equation", + "content": "x _ {s} = \\operatorname {M L P} _ {s} ^ {\\mathrm {X}} (F) \\tag {11}", + "image_path": "995912d79e4232f229698d846dee94a47cc26e06aa2718824dcfbcb9324c07d5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": "In this formulation, " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "c_0" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": " denote the initial hidden state and cell state, while " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "x_s" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": " represents the input at time step " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}^{\\mathrm{hc}}" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": " module encodes the state information, and " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_s^{\\mathrm{x}}" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": " handles the temporal encoding at the " + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 121, + 553, + 181 + ], + "type": "text", + "content": "-th time step, effectively capturing the sequence dynamics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 182, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 554, + 266 + ], + "type": "text", + "content": "At each time step, both states are treated as the " + }, + { + "bbox": [ + 313, + 182, + 554, + 266 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 182, + 554, + 266 + ], + "type": "text", + "content": "-th step's features for output. After concatenating and processing through independent MLPs, each LSTM cell produces outputs, helping capture temporal popularity via backpropagation during training. Before final predictions, we apply a non-negative adjustment since popularity scores cannot be negative. This process is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 270, + 553, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 270, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 350, + 270, + 553, + 283 + ], + "type": "interline_equation", + "content": "\\operatorname {p r e} _ {s} = \\max \\left(0, \\mathrm {M L P} _ {s} ^ {\\text {o u t}} \\left(\\operatorname {C o n c a t} \\left(h _ {s}, c _ {s}\\right)\\right)\\right) \\tag {12}", + "image_path": "959998b4c20da452737b5e3a7ae124916bf05cf844ea64969c51d2ee8fcb11d6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "inline_equation", + "content": "pre_{s}, h_{s}, c_{s}" + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_s^{\\mathrm{out}}" + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "text", + "content": " denote the predicted value, hidden state, cell state, and output MLP at time step " + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 287, + 554, + 348 + ], + "type": "text", + "content": ", respectively. The operation Concat refers to the process of horizontally concatenating the vectors. After these steps, the final temporal popularity predictions are produced." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 352, + 403, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 352, + 403, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 352, + 403, + 361 + ], + "type": "text", + "content": "4.2.2. Loss Function" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 365, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 475 + ], + "type": "text", + "content": "The core component of our approach is the Composite Gradient Loss (CGL), specifically designed for this task. This custom loss function consists of several components: SmoothL1Loss (SL) between the model's outputs and targets, the first-order and second-order derivative differences between the outputs and targets, the L1 loss between the onehot encodings of the predicted and ground truth peaks, and the Laplacian remainder (LR). These components are combined with a weight ratio of 1:1:1:1e-6. The overall loss function " + }, + { + "bbox": [ + 313, + 365, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 365, + 554, + 475 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 478, + 553, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 478, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 315, + 478, + 553, + 542 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathrm {S L} \\left(\\hat {P} _ {d, i}, P _ {d, i}\\right) + \\lambda_ {1} \\cdot \\mathrm {S L} \\left(\\hat {P} _ {d, i} ^ {(1)}, P _ {d, i} ^ {(1)}\\right) + \\lambda_ {2} \\cdot \\mathrm {S L} \\left(\\hat {P} _ {d, i} ^ {(2)}, P _ {d, i} ^ {(2)}\\right) \\tag {13} \\\\ + \\alpha \\cdot \\sum_ {i = 1} ^ {n} \\left| \\delta \\operatorname {a r g m a x} _ {d} \\left(\\hat {P} _ {d, i}\\right) - \\delta \\operatorname {a r g m a x} _ {d} \\left(P _ {d, i}\\right) \\right| + \\epsilon \\cdot \\mathrm {L R} \\\\ \\end{array}", + "image_path": "b18715ea1fc4ef336cc600971bc5700834785bcc08150d30a33392f942b85ca1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 545, + 433, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 545, + 433, + 556 + ], + "spans": [ + { + "bbox": [ + 314, + 545, + 433, + 556 + ], + "type": "text", + "content": "And the Laplacian remainder is:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 376, + 560, + 553, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 560, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 376, + 560, + 553, + 588 + ], + "type": "interline_equation", + "content": "\\mathrm {L R} = \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} ^ {(1)} \\right| + \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} ^ {(2)} \\right| \\tag {14}", + "image_path": "e43603da62111b1594cac9d53fa5af33bc40180da040d3dbc54222d4ddbad89e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": "The term " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SL}(\\hat{P}_{d,i}, P_{d,i})" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": ") is used to compute the error between the predicted outputs and the ground truth targets. Here, " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{P}_{d,i}" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": " represents the predicted popularity for data point " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": " on day " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{d,i}" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": " denotes the ground truth popularity for the same data point. The differences between the one-hot encoded predicted and true peak values are measured using " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\mathrm{argmax}_d(\\hat{P}_{d,i})" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\mathrm{argmax}_d(P_{d,i})" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": ", respectively. Additionally, the term " + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\cdot \\mathrm{LR}" + }, + { + "bbox": [ + 313, + 593, + 555, + 714 + ], + "type": "text", + "content": " introduces a Laplacian remainder (LR) to provide further regularization. During training, the weights for the first-order and second-order" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18852" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "derivative terms, as well as the L1 loss weight between the one-hot encodings " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "(\\lambda_1,\\lambda_2,\\alpha)" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", are adjusted dynamically using a cosine annealing algorithm to ensure smoother convergence throughout the training process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 295, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 295, + 212 + ], + "type": "text", + "content": "In this section, we first describe the experiment settings and evaluation metrics for training models and comparison. Then various experimental results and detailed discussions are provided including evaluations for SMTPD dataset, proposed baseline, multi-modal features and early popularity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 218, + 171, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 171, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 171, + 230 + ], + "type": "text", + "content": "5.1. Experiment settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": "We implement our approaches using the Pytorch* framework and train it with the Adam [25] optimizer incorporating L2 penalty of " + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": ", while the batch size is set to 64. The learning rate is initialized to " + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 55, + 236, + 296, + 319 + ], + "type": "text", + "content": " and adjusted by the ReduceLROnPlateau scheduler provided by PyTorch when one epoch is end. Supervised training was performed using the Composite Gradient Loss (CGL) metioned before." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": "We introduce error-based metrics and correlation-based metrics to fairly evaluate the compared datasets and models. We introduce the Absolute Error (MAE) and average MAE (AMAE) to respectively evaluate models for single-day prediction and temporal prediction. Assuming that the daily MAE and average MAE are denoted as " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "MAE_{d}" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "inline_equation", + "content": "AMAE" + }, + { + "bbox": [ + 55, + 319, + 296, + 402 + ], + "type": "text", + "content": ", they can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 411, + 295, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 411, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 113, + 411, + 295, + 441 + ], + "type": "interline_equation", + "content": "M A E _ {d} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| \\hat {P} _ {d, i} - P _ {d, i} \\right| \\tag {15}", + "image_path": "8dd2c8686e748adaae3c8cdad6556efdeb20f0257c353f4f381e5d004d8c01f4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 444, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 444, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 123, + 444, + 295, + 475 + ], + "type": "interline_equation", + "content": "A M A E = \\frac {1}{m} \\sum_ {d = 1} ^ {m} M A E _ {d} \\tag {16}", + "image_path": "3cfdeae6b38bd384609dc1beec65413cbc6da75f519a8b1930e4b0ec84c0d0e7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": " denotes the total of samples and " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": " denotes total of days, while the " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\hat{P}_d" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": " respectively denotes the ground-truth popularity and predicted popularity for the day " + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 482, + 296, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "text", + "content": "For correlation-based metrics, we introduce the daily SRC and the average SRC to evaluate the models for both single-day prediction and temporal prediction from a different perspective. Assuming " + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "inline_equation", + "content": "P_d" + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\hat{P}_d" + }, + { + "bbox": [ + 55, + 518, + 296, + 590 + ], + "type": "text", + "content": " are defined consistently as mentioned earlier, the daily SRC and average SRC can be formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 597, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 597, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 75, + 597, + 295, + 631 + ], + "type": "interline_equation", + "content": "S R C _ {d} = \\frac {1}{n - 1} \\sum_ {i = 1} ^ {n} \\left(\\frac {\\hat {P} _ {d , i} - \\bar {\\hat {P}} _ {d}}{\\sigma_ {\\hat {P} _ {d}}}\\right) \\left(\\frac {P _ {d , i} - \\bar {P} _ {d}}{\\sigma_ {P _ {d}}}\\right) \\tag {17}", + "image_path": "530950bcdc2889a8252889e64076f1b8e0b10184e0616cad9945639e55a84b62.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 632, + 295, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 632, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 295, + 663 + ], + "type": "interline_equation", + "content": "A S R C = \\frac {1}{m} \\sum_ {d = 1} ^ {m} S R C _ {d} \\tag {18}", + "image_path": "67240a72c08975791588d73f503e6bfb1ccbc4f726725f14ed54e049577fe7aa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\bar{P}_d" + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\sigma_{P_d}" + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": " are mean and standard deviation of the corresponding popularity for the day " + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "In the curves, sequential performance is represented by the MAE and SRC metrics, which illustrate daily prediction errors and correlations for popularity scores. The X-axis denotes days, while the Y-axis shows " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "MAE_{d}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "SRC_{d}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 129, + 429, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 429, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 429, + 140 + ], + "type": "text", + "content": "5.2. SMPD VS. SMPD" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 147, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 361 + ], + "type": "text", + "content": "In this subsection, we conduct comparisons and discussions between the SMTPD and SMPD. We first measure three most recently proposed state-of-the-art popularity prediction methods on two datasets. As these methods are all designed basing on the settings of SMPD, we make some modifications to let them fit for training and testing on our SMTPD. Other top-performing models (e.g. [44, 46, 50]) include components that are not applicable to SMTPD or are not reproducible, such as sliding window average or undisclosed self-trained modules. Hence, we do not discuss them in our experiments. First, we distribute content from corresponding modalities into their respective feature extraction modules. Then we use BERT-Multilingual as the textual feature extractor to address the challenges of multilingual text. Given that our SMTPD dataset consists of temporal popularity scores, we restrict these methods to predict only the single popularity score on day 30. To mitigate random bias, we evaluate using 5-fold cross-validation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 363, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 530 + ], + "type": "text", + "content": "Table 4 shows minimal performance differences across folds, confirming that SMTPD contains abundant data under a well-balanced sample distribution. It is evident that on the single-output prediction of popularity on days 7, 14, and 30 within SMTPD, the performance of method [26] based on GBDT outperforms the deep learning method [13] and [52] in terms of both MAE and SRC through its powerful regression capabilities of Catboost [39]. Notably, our method achieves the best results with the addition of EP, where the MAE of the other three models increases as the prediction horizon extends. In contrast, our model's MAE begins to decrease after day 14. The deep learning method [13] also performs well on days 7 and 14, but its effectiveness declines beyond day 14." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 531, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 639 + ], + "type": "text", + "content": "Moreover there are also two interesting observations. First, all methods achieve higher SRC on SMPD than on SMPD, likely because SMPD lacks specific time points, making it challenging to predict time-dependent popularity trends accurately as popularity declines and their MAE performances are certainly reduced that the MAE values increase by over 0.16. Besides, the larger popularity range and standard deviation among SMPD samples (shown in Table 2) also contributes to prediction difficulties." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 647, + 493, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 493, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 493, + 660 + ], + "type": "text", + "content": "5.3. Evaluation for Proposed Baseline" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Refer to Table 3, we attempted to validate the rationale of partial structures in the proposed baseline model by employing alternative feature extractors and regression networks. Using BERT-Base as the textual features extractor" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 70, + 703, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 135, + 713 + ], + "type": "text", + "content": "*https://pytorch.org" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18853" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "lead to an increase of MAE by around 0.19. This decline in performance may be caused by the limitations of BERT-base model in handling multilingual texts, as it struggles to effectively capture the semantics and contextual information across different languages. The MAE also increased when the MLP (3 layers) was used as the regression structure. Such a structure is similar to [13], suggesting that MLP lacks the capture of temporal information. The SRC did not change for either of the above substitutions, probably due to the high correlation given by EP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 220, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 220, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 220, + 211 + ], + "type": "text", + "content": "5.4. Discussion of Early Popularity" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 357 + ], + "type": "text", + "content": "We evaluate the performance of proposed baseline on SMTPD. As shown in Table 5, the proposed baseline gets a great improvement from existed methods that surpassing the second best method by -0.798/0.109 MAE/SRC for predicting the popularity of day 30. This great improvement is mainly brought by the early popularity (EP). Without the assistant of EP (row 2 in Table 5), the baseline performance sharply reduced around the existed methods. Using the EP predicted by existed method (1.717/0.864 MAE/SRC) as an input also contributes to the baseline's performance (row 3 in Table 5). However the performance gap between the predicted EP and true EP remains large." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 491 + ], + "type": "text", + "content": "Instead of directly involving the EP in the model architecture, introducing EP in the training supervision is another suitable option. To validate the effectiveness of such operation, we train a baseline model for predicting the popularities from day 1 to 30. As it presented by the row 1 in Table 5, adding the 1st day's popularity to the prediction sequence make a slight boost, what is in line with the LSTM's ability to capture temporal dependencies in popularity sequences. Having more preceding temporal information from backpropagation leads to more precise predictions of popularity in subsequent time steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 587 + ], + "type": "text", + "content": "These comparisons clearly indicate that EP plays a crucial role in popularity prediction, and accurately forecasting the first day's popularity is key to predicting future popularity. The strong correlation between EP and future popularity significantly benefits the prediction task. Our results, as demonstrated by the MAE and SRC curves for both natural EP and our model's predictions, underscore the model's capability to leverage EP for enhanced accuracy. From the" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 592, + 294, + 668 + ], + "blocks": [ + { + "bbox": [ + 56, + 592, + 294, + 668 + ], + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 668 + ], + "type": "table", + "html": "
BERT-BaseBERT-MulMLPLSTMAMAEASRC
✓✓0.7820.958
✓✓0.7860.958
✓✓0.7170.959
", + "image_path": "55b7f3491a871ba34e66b0f25264ccae7e4df2072fb9ecb15a981dec55fb0f72.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "type": "text", + "content": "Table 3. The evaluation for the proposed baseline model, mainly on evaluating the model's performance in adapting to multilingual and temporal popularity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 200 + ], + "type": "table", + "html": "
MethodSMTPD (day 7)SMTPD (day 14)SMTPD (day 30)
AverageAverageAverage
Ding et al. [13]1.715/0.8491.669/0.8461.592/0.843
w. EP0.715/0.9640.742/0.9590.749/0.931
Lai et al. [26]1.573/0.8751.524/0.8721.495/0.864
w. EP0.725/0.9570.753/0.9620.760/0.957
Xu et al. [52]1.895/0.8171.832/0.8181.743/0.820
w. EP0.754/0.9620.798/0.9560.822/0.949
Ours w/o. EP1.673/0.8521.628/0.8501.563/0.848
Ours0.713/0.9640.735/0.9590.732/0.959
", + "image_path": "edc9c288f058627c41bedec20bc1fe44d9c8886e00949ba205a3677875ab357a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 253, + 553, + 363 + ], + "blocks": [ + { + "bbox": [ + 313, + 209, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 242 + ], + "type": "text", + "content": "Table 4. The performance (MAE/SRC) was compared across four models, including our model, using the SMTPD dataset, both with and without EP." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 253, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 315, + 253, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 553, + 363 + ], + "type": "table", + "html": "
MethodAMAEASRCMAE(day 30 only)SRC(day 30 only)
w/o. EP(1-30)1.5620.8561.5300.850
w/o. EP(2-30)1.6300.8491.5510.849
w/o. EP+[26]1.6280.8511.5550.848
ours0.7170.9590.7320.959
", + "image_path": "75f27d9d206918ee23b9d54b3d8b86b9b98a4f8c4b9da957c9b62f2973057d98.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 371, + 553, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 416 + ], + "type": "text", + "content": "Table 5. Assessment of EP across different scenarios. Here, \"1-30\" denotes the prediction target spanning a continuous sequence from the 1st day to the 30th day, as does \"2-30\" (to align with methods having EP)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 425, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 510 + ], + "type": "text", + "content": "visualized results, tough EP exhibits high correlation to the popularities of following days, the MAE sharply increased over time. By contrast, our baseline model could well optimize the MAE and achieve even better SRC performance comparing to the natural EP curve. Therefore, the proposed baseline model is effective to utilize EP achieving better prediction accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 519, + 388, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 388, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 388, + 532 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 707 + ], + "type": "text", + "content": "This study aims to address the challenge of time-series popularity prediction in social media. In this paper, we introduce a novel multilingual, multi-modal time-series popularity dataset based on YouTube and suggest a multilingual temporal prediction model tailored to this dataset. Through experiments, we demonstrate the effectiveness of this approach in predicting social media popularity time-series in a multilingual environment. The experiments show that combining multi-modal features with early popularity significantly improves prediction accuracy. However, this study has yet to address challenges such as multi-frame information in videos, maximizing the use of language diversity, and deeper multi-modal exploration. These areas will be key focuses for our future work." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18854" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Deepak Agarwal, Bee-Chung Chen, Pradheep Elango, Nitin Motgi, Seung-Taek Park, Raghu Ramakrishnan, Scott Roy, and Joe Zachariah. Online models for content optimization. Advances in Neural Information Processing Systems, 21, 2008. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "text", + "content": "[2] Sercan Ö Arik and Tomas Pfister. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence, pages 6679-6687, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Peng Bao, Hua-Wei Shen, Junming Huang, and Xue-Qi Cheng. Popularity prediction in microblogging network: a case study on sina weibo. In Proceedings of the 22nd international conference on world wide web, pages 177–178, 2013. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "text", + "content": "[4] Shenghua Bao, Guirong Xue, Xiaoyuan Wu, Yong Yu, Ben Fei, and Zhong Su. Optimizing web search using social annotations. In Proceedings of the 16th international conference on World Wide Web, pages 501-510, 2007. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 294, + 313 + ], + "type": "text", + "content": "[5] Laura Ceci. Youtube-statistics & facts. Statista. com [online].[cit. 2022-05-05]. Dostupné z: https://www.statista.com/topics/2019/youtube, 2022.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "text", + "content": "[6] Junhong Chen, Dayong Liang, Zhanmo Zhu, Xiaojing Zhou, Zihan Ye, and Xiuyun Mo. Social media popularity prediction based on visual-textual features with xgboost. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2692-2696, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 295, + 414 + ], + "type": "text", + "content": "[7] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm SIGkdd international conference on knowledge discovery and data mining, pages 785-794, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "type": "text", + "content": "[8] Weilong Chen, Chenghao Huang, Weimin Yuan, Xiaolu Chen, Wenhao Hu, Xinran Zhang, and Yanru Zhang. Title-and-tag contrastive vision-and-language transformer for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7008-7012, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 545 + ], + "type": "text", + "content": "[9] Xiaolu Chen, Weilong Chen, Chenghao Huang, Zhongjian Zhang, Lixin Duan, and Yanru Zhang. Double-fine-tuning multi-objective vision-and-language transformer for social media popularity prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9462-9466, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 590 + ], + "type": "text", + "content": "[10] Wen-Huang Cheng, Sijie Song, Chieh-Yun Chen, Shintami Chusnul Hidayati, and Jiaying Liu. Fashion meets computer vision: A survey. ACM Computing Surveys (CSUR), 54 (4):1-41, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 646 + ], + "type": "text", + "content": "[11] Minhwa Cho, Dahye Jeong, and Eunil Park. Amps: Predicting popularity of short-form videos using multi-modal attention mechanisms in social media marketing environments. Journal of Retailing and Consumer Services, 78: 103778, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 4171–4186, 2019. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[13] Keyan Ding, Ronggang Wang, and Shiqi Wang. Social media popularity prediction: A multiple feature fusion approach with deep neural networks. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2682-2686, 2019. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "text", + "content": "[14] Zhabiz Gharibshah and Xingquan Zhu. User response prediction in online advertising. aCM Computing Surveys (CSUR), 54(3):1-43, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 196 + ], + "type": "text", + "content": "[15] Anindya Ghose and Sha Yang. An empirical analysis of search engine advertising: Sponsored search in electronic markets. Management science, 55(10):1605-1622, 2009. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "text", + "content": "[16] Sayan Ghosh, Dylan Baker, David Jurgens, and Vinodkumar Prabhakaran. Detecting cross-geographic biases in toxicity modeling on social media. In Proceedings of the Seventh Workshop on Noisy User-generated Text (W-NUT), pages 313–328, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 297 + ], + "type": "text", + "content": "[17] Marcos Andre Gonçalves, Jussara M Almeida, Luiz GP dos Santos, Alberto HF Laender, and Virgilio Almeida. On popularity in the blogosphere. IEEE Internet Computing, 14(3): 42-49, 2010. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "text", + "content": "[18] Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. Learning word vectors for 157 languages. In Proceedings of the International Conference on Language Resources and Evaluation (LREC), 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "type": "text", + "content": "[19] Mohammad Hajarian, Mark Anthony Camilleri, Paloma Díaz, and Ignacio Aedo. A taxonomy of online marketing methods. In Strategic corporate communication in the digital age, pages 235-250. Emerald Publishing Limited, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 466 + ], + "type": "text", + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[22] Chih-Chung Hsu, Chia-Ming Lee, Xiu-Yu Hou, and Chi-Han Tsai. Gradient boost tree network based on extensive feature analysis for popularity prediction of social posts. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9451-9455, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "text", + "content": "[23] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 569, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 611 + ], + "type": "text", + "content": "[24] Aditya Khosla, Atish Das Sarma, and Raffay Hamid. What makes an image popular? In Proceedings of the 23rd international conference on World wide web, pages 867-876, 2014. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations(ICLR), San Diego, California, USA, May 7-9, 2015. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[26] Xin Lai, Yihong Zhang, and Wei Zhang. Hyfea: Winning solution to social media popularity prediction for multimedia grand challenge 2020. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4565-4569, 2020. 2, 7, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18855" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[27] Chenyu Li, Jun Liu, and Shuxin Ouyang. Analysis and prediction of content popularity for online video service: a youku case study. China Communications, 13(12):216-233, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[28] Jinning Li, Yirui Gao, Xiaofeng Gao, Yan Shi, and Guihai Chen. Senti2pop: sentiment-aware topic popularity prediction on social media. In 2019 IEEE International conference on data mining (ICDM), pages 1174-1179. IEEE, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 219, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 294, + 239 + ], + "type": "text", + "content": "[30] Yury Lifshits. Ediscope: Social analytics for online news. Yahoo Labs, 2010. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "type": "text", + "content": "[31] Hung-Hsiang Lin, Jiun-Da Lin, Jose Jaena Mari Ople, Jun-Cheng Chen, and Kai-Lung Hua. Social media popularity prediction based on multi-modal self-attention mechanisms. IEEE Access, 10:4448-4455, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 287, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 294, + 330 + ], + "type": "text", + "content": "[32] An-An Liu, Xiaowen Wang, Ning Xu, Junbo Guo, Guoqing Jin, Quan Zhang, Yejun Tang, and Shenyuan Zhang. A review of feature fusion-based media popularity prediction methods. Visual Informatics, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 332, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 332, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 294, + 365 + ], + "type": "text", + "content": "[33] Marco Lui and Timothy Baldwin. langid.py: An off-the-shelf language identification tool. In Proceedings of the ACL 2012 System Demonstrations, pages 25-30, 2012. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 366, + 294, + 421 + ], + "type": "text", + "content": "[34] Shijian Mao, Wudong Xi, Lei Yu, Gaotian Lü, Xingxing Xing, Xingchen Zhou, and Wei Wan. Enhanced catboost with stacking features for social media prediction. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9430-9435, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[35] Masoud Mazloom, Robert Rietveld, Stevan Rudinac, Marcel Worring, and Willemijn Van Dolen. Multimodal popularity prediction of brand-related social media posts. In Proceedings of the 24th ACM international conference on Multimedia, pages 197-201, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "text", + "content": "[36] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proceedings of the 1st International Conference on Learning Representations (ICLR), Scottsdale, Arizona, USA, May 2-4, 2013. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "text", + "content": "[37] Alessandro Ortiz, Giovanni Maria Farinella, and Sebastiano Battiato. Prediction of social image popularity dynamics. In Image Analysis and Processing-ICIAP 2019: 20th International Conference, Trento, Italy, September 9–13, 2019, Proceedings, Part II 20, pages 572–582. Springer, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 643 + ], + "type": "text", + "content": "[38] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[39] Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. Catboost: unbiased boosting with categorical features. Advances in neural information processing systems, 31, 2018. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[40] Cédric Richier, Eitan Altman, Rachid Elazouzi, Tania Altman, Georges Linares, and Yonathan Portilla. Modelling" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "view-count dynamics in youtube. CoRR, abs/1404.2570, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[41] Cedric Richier, Eitan Altman, Rachid Elazouzi, Tania Jimenez, Georges Linares, and Yonathan Portilla. Bioinspired models for characterizing youtube viewcout. In 2014 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 297-305. IEEE, 2014. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[42] Satoshi Sanjo and Marie Katsurai. Recipe popularity prediction with deep visual-semantic fusion. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2279–2282, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "text", + "content": "[43] Gabor Szabo and Bernardo A Huberman. Predicting the popularity of online content. Communications of the ACM, 53 (8):80-88, 2010. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[44] YunPeng Tan, Fangyu Liu, BoWei Li, Zheng Zhang, and Bo Zhang. An efficient multi-view multimodal data processing framework for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7200-7204, 2022. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "text", + "content": "[45] Alexandru Tatar, Marcelo Dias De Amorim, Serge Fdida, and Panayotis Antoniadis. A survey on predicting the popularity of web content. Journal of Internet Services and Applications, 5(1):1-20, 2014. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 392 + ], + "type": "text", + "content": "[46] Kai Wang, Penghui Wang, Xin Chen, Qiushi Huang, Zhendong Mao, and Yongdong Zhang. A feature generalization framework for social media popularity prediction. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4570-4574, 2020. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "type": "text", + "content": "[47] Bo Wu, Wen-Huang Cheng, Yongdong Zhang, Qiushi Huang, Jintao Li, and Tao Mei. Sequential prediction of social media popularity with deep temporal context networks. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI), pages 3062-3068, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "text", + "content": "[48] Bo Wu, Wen-Huang Cheng, Peiye Liu, Bei Liu, Zhaoyang Zeng, and Jiebo Luo. Smp challenge: An overview of social media prediction challenge 2019. In Proceedings of the 27th ACM International Conference on Multimedia, pages 2667-2671. ACM, 2019. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 558 + ], + "type": "text", + "content": "[49] Bo Wu, Peiye Liu, Wen-Huang Cheng, Bei Liu, Zhaoyang Zeng, Jia Wang, Qiushi Huang, and Jiebo Luo. Smp challenge: An overview and analysis of social media prediction challenge. In Proceedings of the 31st ACM International Conference on Multimedia, pages 9651-9655, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "text", + "content": "[50] Jianmin Wu, Liming Zhao, Dangwei Li, Chen-Wei Xie, Siyang Sun, and Yun Zheng. Deeply exploit visual and language information for social media popularity prediction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 7045–7049, 2022. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 615, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 656 + ], + "type": "text", + "content": "[51] Leilei Wu, Lingling Yi, Xiao-Long Ren, and Linyuan Lu. Predicting the popularity of information on social platforms without underlying network structure. Entropy, 25(6):916, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[52] Kele Xu, Zhimin Lin, Jianqiao Zhao, Peicang Shi, Wei Deng, and Huaimin Wang. Multimodal deep learning for social media popularity prediction with attention mechanism. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4580-4584, 2020. 2, 7, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18856" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[53] Jaewon Yang and Jure Leskovec. Patterns of temporal variation in online media. In Proceedings of the fourth ACM international conference on Web search and data mining, pages 177-186, 2011. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[54] Zhizhen Zhang, Xiaohui Xie, Mengyu Yang, Ye Tian, Yong Jiang, and Yong Cui. Improving social media popularity prediction with multiple post dependencies. CoRR, abs/2307.15413, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[55] Alireza Zohourian, Hedieh Sajedi, and Arefeh Yavary. Popularity prediction of images and videos on instagram. In 2018 4th International Conference on Web Research (ICWR), pages 111-117, 2018. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18857" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_content_list.json b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a843e7990ae70651b1f66c16e288913d126bba33 --- /dev/null +++ b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_content_list.json @@ -0,0 +1,1595 @@ +[ + { + "type": "text", + "text": "SOAP: Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection", + "text_level": 1, + "bbox": [ + 119, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyo-Jun Lee", + "bbox": [ + 281, + 202, + 395, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyunseop Kim", + "bbox": [ + 300, + 220, + 429, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yeong Jun $\\mathrm{Koh}^{2*}$", + "bbox": [ + 436, + 203, + 573, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yonguk Lee", + "bbox": [ + 470, + 220, + 576, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanul Kim3", + "bbox": [ + 614, + 203, + 712, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinu Lee", + "bbox": [ + 617, + 220, + 694, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Kangwon National University", + "bbox": [ + 227, + 241, + 473, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Chungnam National University", + "bbox": [ + 511, + 241, + 769, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Seoul National University of Science and Technology", + "bbox": [ + 218, + 260, + 648, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "42dot Inc.", + "bbox": [ + 687, + 260, + 779, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gywns6298@kangwon.ac.kr", + "bbox": [ + 199, + 281, + 405, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yjkoh@cnu.ac.kr", + "bbox": [ + 437, + 282, + 573, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hukim@seoultech.ac.kr", + "bbox": [ + 609, + 281, + 797, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hyunseop95@gmail.com", + "bbox": [ + 207, + 299, + 390, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yonguk.lee@42dot.ai", + "bbox": [ + 423, + 299, + 593, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jinu.lee@42dot.ai", + "bbox": [ + 630, + 299, + 781, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 347, + 326, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing view transformations in vision-centric 3D Semantic Scene Completion (SSC) inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, leading to a dilution of informative contexts. Furthermore, semantic classes exhibit high variability in their appearance in real-world driving scenarios. To address these issues, we introduce a novel 3D SSC method, called SOAP, including two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection effectively converts 2D image features into voxel space, refining the duplicated features of occluded regions using information gathered from previous observations. The scene-adaptive decoder guides query embeddings to learn diverse driving environments based on a comprehensive semantic repository. Extensive experiments validate that the proposed SOAP significantly outperforms existing methods for the vision-centric 3D SSC on automated driving datasets, SemanticKITTI and SSCBench. Code is available at https://github.com/gywns6287/SOAP.", + "bbox": [ + 89, + 378, + 483, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 705, + 220, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding 3D traffic scenes is crucial for autonomous driving systems to develop downstream tasks such as motion prediction and planning. LiDAR sensors [8, 33, 43, 54] provide precise 3D scene information, but their high cost and the complex setup limit their scalability. In contrast, areas of study about vision-centric 3D perception, including 3D object detection [11, 25, 46, 48, 52], 3D car shape reconstruction [14, 16, 17], and depth estimation [2, 27, 29, 45], have been receiving attention due to its simplicity and cost-effectiveness. However, these methods still struggle to re", + "bbox": [ + 89, + 728, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a000ecd8d9ad8d5aa724a75275853163d3a45af72c5e0209c16329be42543dd.jpg", + "image_caption": [ + "(a) Occluded voxel regions." + ], + "image_footnote": [], + "bbox": [ + 513, + 345, + 908, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8e435855253d1c9952d8a9efcc9689e8fffb2295847165759e72d8f5a237b5f.jpg", + "image_caption": [ + "(b) Diversity in the appearance of semantic classes.", + "Figure 1. Two main motivations of SOAP. (a) The 2D-to-3D view projection inevitably yields erroneous feature duplication in voxel generation due to occlusion. (b) The semantic classes exhibit significant variability in their appearance depending on driving regions." + ], + "image_footnote": [], + "bbox": [ + 514, + 489, + 908, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "construct complete and fine-grained 3D scene representations. To address this limitation, 3D Semantic Scene Completion (SSC) has emerged as a promising alternative. 3D SSC leverages a voxel-based approach, assigning both an occupancy probability and a semantic label (e.g. car, pedestrian, road) to each voxel within the 3D space. Thereby, it can create a comprehensive representation of the driving scene. While earlier works [5, 19, 39, 41] primarily rely on LiDAR sensors, recent research [3, 13, 18, 22, 37, 50, 51] has demonstrated the potential of vision-centric methods.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 114, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17145", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A general vision-based 3D SSC model comprises an image encoder, a view projection module, and a voxel decoder. The image encoder first extracts 2D features from an input image, and then the view projection module subsequently lifts 2D image features into 3D voxel space. Given the 3D features of each voxel region, the voxel decoder estimates semantic classes corresponding to each voxel region. In this process, the precise 2D-3D view projection plays a crucial role. However, existing view projection methods [3, 50] inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, particularly along the longitudinal direction. For instance, as illustrated in Figure 1a, when a camera ray encounters an occupancy region (e.g. car), occluded regions (voxels behind this occupancy) may contain duplicated features of the obstructing car in the 3D view projection process. This duplication can occur even when the actual classes of the occluded regions are not related to a car, resulting in incorrect 3D SSC results.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, the diverse appearances of semantic classes pose another significant challenge in 3D SSC. As shown in Figure 1b, semantic classes exhibit high variability in their appearance in real-world driving scenarios. For instance, road color and texture may vary depending on driving regions (e.g. urban or highway). Such appearance variations make it difficult for SSC models to distinguish similar classes, such as road $\\leftrightarrow$ sidewalk, leading to potential misinterpretations of driving environments. While most previous works [3, 13, 22, 37] usually employ sequential 3D convolutions for the semantic decoder, this approach often lacks the flexibility to generalize for the diverse semantic context. More recently, Mask2former [6]-like approach [50, 51] is adopted for the SSC decoder, which utilizes semantic tokens derived from a query embedding to represent the semantic scene. However, these semantic tokens lack adaptability to dynamic driving environments, as they are fundamentally based on static query embedding.", + "bbox": [ + 89, + 367, + 482, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel 3D semantic scene completion network, SOAP, faithfully designed to address the occluded voxel regions and the visual variations in semantic classes. The proposed method leverages two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection (OAP) first converts 2D image features into 3D voxel features using the LSS algorithm [28], where these initial voxel features contain false duplicated information due to occlusions, as illustrated in Figure 1a. These occluded regions can be inferred by referring to previous frames, which provides valuable context about the missing observation. Therefore, OAP identifies occluded regions in the current frame using depth estimation and refines the duplicated features of occluded regions using information gathered from previous observations. Also, a vanilla transformer decoder [6] generates the semantic tokens based on", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "static query embedding, which is not versatile enough to capture a dynamic driving environment. To represent more diverse semantics, a repository builder first creates a semantic repository by gathering semantic tokens from various scenes. Then, the scene-adaptive decoder integrates informative clues from this extensive repository into the semantic tokens, guiding them to navigate diverse driving environments effectively.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that the proposed SOAP outperforms the other state-of-the-arts on two vision-centric SSC benchmarks, SemanticKITTI [1] and SSCBench-KITTI360 [21]. To summarize, the main contributions of our work are three folds:", + "bbox": [ + 511, + 212, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop the occluded region-aware view projection that effectively transforms 2D image features into 3D voxel space by recovering occluded voxel regions.", + "- We also introduce a scene-adaptive transformer that integrates informative semantic clues from the various scenes into the semantic tokens of an individual scene.", + "- The proposed SOAP achieves outstanding performances on two driving datasets, SemanticKITTI and SSCBenchKITTI360, with a margin of $+2$ and $+2.67$ mIoU scores, respectively." + ], + "bbox": [ + 513, + 287, + 903, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 452, + 661, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Semantic Scene Completion. 3D semantic scene completion (SSC) is an advanced task in 3D scene reconstruction fields, which aims to predict a dense semantic scene representation involving the geometrical occupancy and its corresponding semantic label. Early SSC methods [5, 19, 34] leverage 3D inputs to reconstruct voxelized indoor scenes. SSCNet [34] first defines the SSC task, which jointly reasons the geometry and semantics of an indoor scene from a single-depth image. 3DSketch [5] uses the Truncated Signed Distance Function to guide an occupancy and the semantic labels of the indoor scene. AIC-Net [19] develops anisotropic convolution to extract hybrid features from the depth and RGB images for indoor SSC tasks. Recently, the release of SemanticKITTI [1], the first large-scale outdoor SSC dataset, yields the development of diverse outdoor SSC methods. LMSCNet [31] uses the 2D UNet to extract collapsed BEV features from sparse 3D Li-DAR scans and finally expands the height dimension for the 3D SSC. JS3C-Net [41] learns the contextual shape priors from sequential LiDAR data completing the sparse single sweep point cloud to the dense one for enhanced 3D SSC. The other existing works [7, 30, 36, 39, 40, 47] also commonly depend on 3D input such as LiDAR point clouds. In contrast, we address vision-centric 3D SSC, which only uses an RGB image for the semantic scene reconstruction.", + "bbox": [ + 511, + 477, + 906, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-centric 3D SSC. MonoScene [3] introduces the first vision-centric approach for SSC, utilizing a 3D UNet to process voxel features generated by sight projection. Vox-", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17146", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a5c87da77491001fd7b2aabc086de697c19964d59e2999dca521239f6247600a.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed SOAP. Given multi-scale image features, an occluded region-aware view projection (OAP) transforms them into voxel space, refining duplicated features in occluded voxel regions using historical information. A scene-adaptive decoder enriches query embeddings with diverse semantic contexts through a semantic repository, which is constructed and iteratively updated by a repository builder." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 888, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Former [22] estimates occupancy regions using stereo depth estimation and selectively updates these regions with RGB images. HASSC [37] and Symphonies [13] also identify voxel occupancy based on stereo depth estimation, following VoxFormer. NDC-scene [42] extends a 2D feature map to a normalized device coordinates space to restore the depth dimension and subsequently reconstructs a 3D semantic scene using this space. SurroundOcc [38] adopts spatial 2D-3D attention to lift multi-scale 2D features into the 3D volume space. TPVFormer [12] introduces a query-based tri-perspective space for 3D SSC. Occformer [50] and LowRankOcc [51] reconstruct 3D scenes using semantic tokens on the Mask2former decoders [6]. SGN [26] propagates semantic contexts from semantic-aware seed voxels based on their spatial geometry cues. HTCL [18] leverages historical frames to refine current frame features through affinity-based dynamic refinement.", + "bbox": [ + 88, + 486, + 485, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 760, + 183, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Overview", + "text_level": 1, + "bbox": [ + 89, + 786, + 253, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Semantic Scene Completion (SSC) aims to predict semantic probabilities of the voxelized scene $\\mathbf{V} \\in \\mathbb{R}^{X \\times Y \\times Z \\times S}$ , where $(X, Y, Z)$ denotes the resolution of the 3D volume, and $S$ is the number of semantic classes including a free (or empty) class. Figure 2 illustrates the overall architecture of the proposed SOAP. Image encoder first ex", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tracts multi-scale image features $\\{\\mathbf{F}_i^{2\\mathrm{D}} \\in \\mathbb{R}^{H_i \\times W_i \\times C}\\}_{i=1}^N$ from an RGB image, where $N$ is the number of scales, $(H_i, W_i)$ is the spatial size at the scale $i$ , and $C$ is the channel dimension. Depth estimator predicts the depth map $\\mathbf{E} \\in \\mathbb{R}^{H \\times W}$ from RGB images. Consistent with previous approaches [13, 18, 22, 37, 44], we also use off-the shelf stereo depth model [32] for the depth estimator.", + "bbox": [ + 511, + 484, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The occluded region-aware view projection (Section 3.2) utilizes the estimated depth map $\\mathbf{E}$ to lift multi-scale image features $\\{\\mathbf{F}_i^{2\\mathrm{D}}\\}_{i = 1}^N$ into multi-scale voxel features $\\{\\mathbf{F}_i^{3\\mathrm{D}}\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i\\times C}\\}_{i = 1}^N$ , enhancing occluded voxel regions based on informative features in historical frames. The scene-adaptive decoder (Section 3.3) first constructs a semantic repository using a repository builder and then it integrates semantic diversity from a semantic repository into query embeddings to produce scene-adaptive semantic tokens, which are utilized to reconstruct the final semantic voxel representation $\\mathbf{V}$ , in conjunction with multiscale voxel features. Through this complementary process, SOAP guarantees a more flexible and accurate 3D SSC for a real-world driving environment.", + "bbox": [ + 511, + 592, + 908, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Occluded Region-Aware View Projection", + "text_level": 1, + "bbox": [ + 511, + 809, + 864, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Occluded region-aware view projection (OAP) transforms 2D image features $\\{\\mathbf{F}_i^{2\\mathrm{D}}\\}_{i = 1}^N$ into 3D voxel space $\\{\\mathbf{F}_i^{3\\mathrm{D}}\\}_{i = 1}^N$ . Figure 3 illustrates the detailed process of OAP. It produces voxel features for each $i$ -th scale solely from the image feature at $i$ -th scale.", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17147", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c035787505e4718e3d4720049c84aded0b31c19efa75517a8d4e6c4c0f192e8f.jpg", + "image_caption": [ + "Figure 3. Occluded region-aware view projection. (a) occluded region identification identifies occluded regions in the current frame by inferring an occlusion map based on the depth distribution. (b) occluded region refinement refines the duplicated features in the identified occluded regions using previous observations within the historical voxel features." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 875, + 311 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/12214c58a5d4ec297e5914d097ccdba18ee4ef28fbec11dc72cf99b58d551342.jpg", + "image_caption": [ + "Figure 4. t-SNE visualization of voxel features in the occluded regions for (a) $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ and (b) $\\tilde{\\mathbf{F}}_i^{\\mathrm{occ}}$ . Different colors represent different camera rays." + ], + "image_footnote": [], + "bbox": [ + 91, + 391, + 480, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initial Voxel Generation. For each $i$ th scale, OAP first generates initial voxel features $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}$ from the 2D image features $\\mathbf{F}_i^{2\\mathrm{D}}$ and the depth estimation $\\mathbf{E}$ . Specifically, OAP resizes the estimated depth $\\mathbf{E}$ to $\\mathbf{E}_i$ that has the spatial resolution of $(H_i, W_i)$ , and distributes it to the depth distribution map $\\mathbf{D}_i \\in \\mathbb{R}^{H_i \\times W_i \\times D}$ . Here, the continuous depth value $\\mathbf{E}_i(h, w)$ for each pixel $(h, w)$ is converted into discrete depth bins $\\mathbf{D}_i(h, w) \\in \\mathbb{R}^D$ represented as one-hot encoded vectors, followed by a Gaussian filter to create a soft depth distribution. Then, OAP computes the outer product $\\mathbf{F}_i^{2\\mathrm{D}} \\otimes \\mathbf{D}_i$ to create the depth distributed features $\\mathbf{F}_i^{\\mathrm{depth}} \\in \\mathbb{R}^{H_i \\times W_i \\times D \\times C}$ . The voxel-", + "bbox": [ + 89, + 717, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pooling operation $\\varphi(\\cdot)$ in [28] can project $\\mathbf{F}_i^{\\mathrm{depth}}$ into the voxel space using camera intrinsics and extrinsics. Thus, the initial voxel features are defined as $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}} = \\varphi(\\mathbf{F}_i^{\\mathrm{depth}})$ . Historical Voxel Generation. To represent previous observations, we incorporate depth distributed features $\\{\\mathbf{F}_{i,t - p}^{\\mathrm{depth}}\\}_{p = 1}^{P}$ from the previous frames into the current voxel space $\\mathbf{F}_i^{\\mathrm{hist}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}$ , where $P$ is the number of previous frames at timestamp $t$ . Specifically, each past $\\mathbf{F}_{i,t - p}^{\\mathrm{depth}}$ is projected into voxel space $\\varphi(\\mathbf{F}_{i,t - p}^{\\mathrm{depth}})$ and then aligned to the current voxel space at $t$ via the warping function $\\mathcal{W}(\\cdot)$ and summation operation:", + "bbox": [ + 511, + 388, + 906, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {\\mathrm {h i s t}} = \\sum_ {p = 1} ^ {P} \\mathcal {W} \\left(\\varphi \\left(\\mathbf {F} _ {i, t - p} ^ {\\mathrm {d e p t h}}\\right), \\mathcal {G} _ {t} ^ {- 1} \\cdot \\mathcal {G} _ {t - p}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 565, + 906, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_t$ and $\\mathcal{G}_{t - p}$ are ego-posed matrices at $t$ and $t - p$ . These historical voxel features $\\mathbf{F}_i^{\\mathrm{hist}}$ provide valuable perspectives that are not available in the initial voxel features. Occluded Region Identification. The initial voxel features $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ , which are projected from a single image view, inherently include false duplicated features due to occlusions. Figure 4a presents the t-SNE visualization of features in the occluded regions within $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ . Each point represents an individual voxel region, and voxel regions of the same color correspond to those located along the same camera ray. We observe that most occluded regions along the same camera ray exhibit similar patterns, which indicates erroneous replication of the features from the occupancy regions positioned in front of the occluded regions as in Figure 1a.", + "bbox": [ + 511, + 613, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this limitation, OAP refines these occluded features by referring to previous frames, which provide valuable contexts about the missing observation in the current frame. As part of this process, we identify the occluded regions using the estimated depth distribution $\\mathbf{D}_i$ .", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17148", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As in Figure 3a, each vector $\\mathbf{D}_i(h,w)$ at the spatial location $(h,w)$ contains the probability distribution of the nearest occupancy region along the longitudinal (or depth) direction. In other words, regions located behind the nearest occupancy with the highest depth probability often lose their semantic context information due to occlusions. To distinguish the occluded regions from non-occluded regions, we define an occlusion map $\\mathbf{O}_i\\in \\{1,0\\}^{H_i\\times W_i\\times D}$ as", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {i} (h, w, d) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d > \\underset {\\tilde {d} \\in \\{1, \\dots , D \\}} {\\arg \\max } \\mathbf {D} _ {i} (h, w, \\tilde {d}) \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 224, + 483, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3a provides an example of $\\mathbf{O}_i$ , where occluded regions are marked with 1 (taller rectangular) and non-occluded regions are marked with 0 (shorter rectangular). To identify occluded regions in the voxel space, the occlusion map $\\mathbf{O}_i$ is projected into voxel space through the voxel pooling operation, i.e. $\\varphi (\\mathbf{O}_i)\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i}$ . Note that the regions identified by this occlusion map $\\varphi (\\mathbf{O}_i)$ contain incorrectly duplicated information from the nearest occupancy region along the longitudinal direction, as illustrated in Figure 4a. Therefore, we refine the features of these occluded regions through an occluded region refinement process.", + "bbox": [ + 89, + 292, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Occluded Region Refinement. The occluded regions at the current time step may have been observable in previous time steps. Therefore, we refine deficient features in the occluded regions with previous observations in the historical voxel features $\\mathbf{F}_i^{\\mathrm{hist}}$ . To achieve this, we perform 3D deformable cross-attention, DCA [53] between the historical voxel features $\\mathbf{F}_i^{\\mathrm{hist}}$ and the occluded voxel features, resulting in reconstructed occluded voxel features $\\mathbf{F}_i^{\\mathrm{occ}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}$ :", + "bbox": [ + 89, + 474, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {\\text {o c c}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z) + \\mathrm {D C A} \\left(\\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z), \\mathbf {F} _ {i} ^ {\\text {h i s t}}\\right), \\\\ \\quad \\text {i f} \\varphi (\\mathbf {O} _ {i}) (x, y, z) \\geq 1 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 625, + 483, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(x,y,z)$ are spatial indices of the voxel space. This approach ensures that the reconstructed features $\\mathbf{F}_i^{\\mathrm{occ}}$ effectively capture observable historical information about occluded regions, enabling more accurate and context-aware scene representation. As shown in Figure 4b, the reconstructed features $\\mathbf{F}_i^{\\mathrm{occ}}$ include appropriately corrected information for the occluded regions, resulting in more diverse patterns according to different semantic regions.", + "bbox": [ + 89, + 703, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Invisible and Visible Region Reconstruction. The initial voxel features also contain invisible regions due to the limited observation of a single camera. Since these invisible voxels do not contain any features, we directly fill them with the corresponding historical voxel features to recon-", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "struct invisible voxel features $\\mathbf{F}_i^{\\mathrm{inv}}\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i\\times C}$ ..", + "bbox": [ + 513, + 89, + 870, + 108 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {\\text {i n v}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\mathbf {F} _ {i} ^ {\\text {h i s t}} (x, y, z) & \\text {i f} \\sum \\tilde {\\mathbf {F}} _ {i} ^ {\\text {3 D}} (x, y, z) = 0 \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 116, + 903, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We designate the voxel space, excluding invisible and occluded regions, as visible regions. The visible regions already contain appropriate semantic features with respect to the current observation, allowing us to directly retrieve the visible voxel features from the initial voxel features $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ . The visible voxel features $\\mathbf{F}_i^{\\mathrm{vis}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}$ are defined as", + "bbox": [ + 511, + 176, + 906, + 281 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {\\mathrm {v i s}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z) & \\text {i f} \\varphi (\\mathbf {O} _ {i}) (x, y, z) = 0 \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 287, + 903, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Voxel Encoder. To compute the final 3D voxel features, we sum the occluded, invisible, and visible voxel features, $\\mathbf{F}_i^{\\mathrm{occ}} + \\mathbf{F}_i^{\\mathrm{inv}} + \\mathbf{F}_i^{\\mathrm{vis}}$ , and process them through a voxel encoder to obtain the final voxel features $\\mathbf{F}_i^{3\\mathrm{D}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i}$ . In the voxel encoder, the spatial dimensions $(X_i \\times Y_i \\times Z_i)$ of the voxel features are sliced into $Z_i$ bird's-eye view (BEV) planes with spatial resolution $(X_i \\times Y_i)$ . Then, two windowed self-attention blocks [23] are sequentially performed to each BEV plane.", + "bbox": [ + 511, + 349, + 906, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Scene-Adaptive Decoder", + "text_level": 1, + "bbox": [ + 513, + 494, + 736, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A standard transformer decoder [6, 50, 51] yields semantic tokens starting from static query embeddings, which lack adaptability for dynamic driving scenes. To represent more diverse semantics, the scene-adaptive decoder correlates semantic tokens for a specific scene with tokens from different scenes. As in Figure 2, the scene-adaptive decoder generates scene-adaptive tokens $\\mathbf{Q} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times C}$ through $L$ iterative transformer blocks, where $N_{\\mathrm{q}}$ is the number of tokens. Each transformer block in the scene transformer cascades four modules including a cross-attention between query embeddings and the semantic repository, the other cross-attention between the query embeddings and the voxel features $\\mathbf{F}_i^{3\\mathrm{D}}$ , self-attention, and feed-forward networks. The cross-attention with the semantic repository integrates semantic features from dynamic scenes into the query embeddings, while the cross-attention with voxel features provides scene-specific semantic features. After the end of iterations, scene-adaptive tokens $\\mathbf{Q}$ are projected to class predictions $\\mathbf{S}^{\\mathrm{cls}} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}$ and segmentation mask predictions $\\mathbf{M} \\in \\mathbb{R}^{X \\times Y \\times Z \\times N_{\\mathrm{q}}}$ . Then, the final semantic voxel $\\mathbf{V} \\in \\mathbb{R}^{X \\times Y \\times Z \\times S}$ is formulated as", + "bbox": [ + 511, + 516, + 908, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\mathcal {F} ^ {- 1} (\\mathcal {F} (\\mathbf {M}) \\mathbf {S} ^ {\\mathrm {c l s}}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 842, + 903, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}:\\mathbb{R}^{X\\times Y\\times Z\\times C}\\to \\mathbb{R}^{XYZ\\times C}$ is as spatial flatten and $\\mathcal{F}^{-1}$ is its inverse operation.", + "bbox": [ + 511, + 868, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17149", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0e81f49a627d0638da5baa8e30281c1492db0c5b457e9d03657d36f0c2e90799.jpg", + "table_caption": [ + "Table 1. Semantic scene completion results on SemanticKITTI test set. † represents temporal methods that leverage previous frames for 3D SSC. We highlight the best results in bold and the second best results in underline." + ], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoUroad (15.30%)sidewalk (11.13%)parking (1.12%)other-ground (0.56%)building (14.1%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation(39.3%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (3.90%)pole (0.29%)traf-sign (0.08%)
MonoScene [3]34.1611.0854.727.124.85.714.418.83.30.50.74.414.92.419.51.01.40.411.13.32.1
TPVFormer [12]34.2511.2655.127.227.46.514.819.23.71.00.52.313.92.620.41.12.40.311.02.91.5
SurroundOcc [38]34.7211.8656.928.330.26.815.220.61.41.61.24.414.93.419.31.42.00.111.33.92.4
VoxFormer† [22]43.2113.4154.126.925.17.323.521.73.61.91.64.124.48.124.21.61.10.013.16.65.7
OccFormer [50]34.5312.3255.930.331.56.515.721.61.21.51.73.216.83.921.32.21.10.211.93.83.7
NDC-Scene [42]33.8711.5556.228.728.05.615.819.71.81.11.14.914.32.620.60.71.70.411.23.21.7
LowRankOcc [51]38.4713.5652.827.225.18.822.120.92.93.32.74.422.98.920.82.41.72.314.47.07.0
Symphonies [13]42.1915.0458.429.326.911.724.723.63.23.62.65.624.210.023.13.21.92.016.17.78.0
HASSC† [37]42.8714.3855.329.625.911.323.123.02.91.91.54.924.89.826.51.43.00.014.37.07.1
SGN† [26]45.4215.7660.431.428.98.728.425.44.50.91.63.727.412.628.40.50.30.118.110.08.3
HTCL† [18]44.2317.0964.434.833.812.425.927.35.71.82.25.425.310.831.21.13.10.921.19.08.3
\\(SOAP^{\\dagger}\\) (proposed)46.0919.0963.636.236.817.228.728.93.35.05.37.029.815.132.33.72.30.222.511.713.1
", + "bbox": [ + 91, + 127, + 906, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Repository Builder. The semantic repository $\\mathbf{R} \\in \\mathbb{R}^{(S \\times K) \\times C}$ stores $K$ representative tokens for each semantic class, iteratively updated by semantic tokens with top- $K$ highest scores. The detailed process of the repository builder is shown in Figure 2. Even though representative tokens could be selected using the classification scores $\\mathbf{S}^{\\mathrm{cls}}$ of the scene-adaptive tokens $\\mathbf{Q}$ , relying solely on $\\mathbf{S}^{\\mathrm{cls}}$ makes a repository filled with similar token and consequently degrades semantic diversity. To retain the semantic diversity within the repository $\\mathbf{R}$ , we introduce a novelty-based token selection. Specifically, we re-define token scores $\\mathbf{S} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}$ as $\\mathbf{S} = \\mathbf{S}^{\\mathrm{cls}} + \\alpha \\mathbf{S}^{\\mathrm{nov}}$ , where $\\mathbf{S}^{\\mathrm{nov}} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}$ is novelty scores and $\\alpha$ is its balancing parameter. To compute the novelty scores, the $K$ tokens within each class in $\\mathbf{R}$ are first averaged, resulting in the class average repository $\\bar{\\mathbf{R}} \\in \\mathbb{R}^{S \\times C}$ . Then, the novelty scores $\\mathbf{S}^{\\mathrm{nov}}$ are calculated based on cosine similarity between each token in $\\mathbf{Q}$ and $\\bar{\\mathbf{R}}$ :", + "bbox": [ + 89, + 436, + 485, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} ^ {\\text {n o v}} (n, s) = - \\cos (\\mathbf {Q} (n), \\bar {\\mathbf {R}} (s)) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 709, + 482, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n$ and $s$ indicate token indices of $\\mathbf{Q}$ and $\\bar{\\mathbf{R}}$ , respectively, and $\\cos(\\cdot, \\cdot)$ is a cosine similarity between two input vectors.", + "bbox": [ + 89, + 733, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, new tokens in $\\mathbf{Q}$ with higher scores than old tokens in $\\mathbf{R}$ are selected and, newly pushed into the semantic repository $\\mathbf{R}$ . When a class holds more than $K$ tokens, the old tokens with lower scores are discarded to maintain $K$ representative tokens for each semantic class in $\\mathbf{R}$ . Updating $\\mathbf{R}$ with tokens that exhibit both high classification and novelty scores enables the repository to capture rich semantic information, effectively covering dynamic driving", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scenes. Note that the semantic repository $\\mathbf{R}$ is updated only during the SOAP's training phase to ensure stable inference performance.", + "bbox": [ + 511, + 436, + 906, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Loss Functions", + "text_level": 1, + "bbox": [ + 511, + 494, + 666, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform the Hungarian matching [15] between predicted and ground-truth voxel masks following [4, 6, 53]. For each matched pair, we compute a mask loss $L_{\\mathrm{mask}}$ , a mask dice loss $L_{\\mathrm{dice}}$ , and a class loss $L_{\\mathrm{cls}}$ as in [6]. The final loss is defined as", + "bbox": [ + 511, + 517, + 906, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {\\text {m a s k}} L _ {\\text {m a s k}} + \\lambda_ {\\text {d i c e}} L _ {\\text {d i c e}} + \\lambda_ {\\text {c l s}} L _ {\\text {c l s}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 608, + 906, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, scaling weights $\\lambda_{\\mathrm{mask}}, \\lambda_{\\mathrm{dice}}, \\lambda_{\\mathrm{cls}}$ are set to 5, 5, 2, respectively.", + "bbox": [ + 511, + 637, + 905, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 684, + 645, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 710, + 614, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of SOAP on two outdoor 3D SSC datasets: SemanticKITTI [1] and SSCBenchKITTI360 [21]. Both datasets provide front-view RGB images, LiDAR sweeps, and semantic scene labels represented as $256 \\times 256 \\times 32$ voxel grids. The scene labels cover a range of $51.2m \\times 51.2m \\times 6.4m$ from the front view of the vehicle.", + "bbox": [ + 511, + 733, + 906, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SemanticKITTI [1]. It contains a total of 22 sequences, which are split into $10/1/11$ for train/val/test following [3]. We utilize the RGB images of cam-2 with a $1280 \\times 384$ resolution as model inputs. In the semantic labels, each voxel", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17150", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fbd892223eb3fefd9971900474d2787825477382606abf22356d48a919a22ddf.jpg", + "table_caption": [ + "Table 2. Semantic scene completion results on the SSCBench test set. † represents methods that use temporal information. We highlight the best results in **bold** and the second best results in **underline**." + ], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoUcarbicyclemotorcycletruckother-veh.personroadparkingsidewalkother-groundbuildingfencevegetationterrainpoletraf.-signother-struct.other-object
MonoScene [3]37.8712.3119.30.40.68.02.00.948.411.428.13.232.93.526.216.86.95.74.23.1
TPVFormer [12]40.2213.6421.61.11.48.12.62.453.012.031.13.834.84.830.117.57.55.95.52.7
OccFormer [50]40.2713.8122.60.70.39.93.82.854.313.431.53.636.44.831.019.57.88.57.04.6
VoxFormer†[22]38.7611.9117.81.20.94.62.11.647.09.727.22.931.25.029.014.76.56.93.82.4
Symphonies [13]44.1218.5830.01.95.925.112.18.254.913.832.86.935.18.638.311.514.09.614.411.3
SGN†[26]47.0618.2529.03.42.910.95.23.058.115.036.44.442.07.738.223.216.716.49.95.8
\\( \\mathrm{SOAP}^{\\dagger } \\) (proposed)48.1220.9229.95.67.814.47.66.160.917.440.35.445.310.640.524.816.821.012.69.9
", + "bbox": [ + 91, + 127, + 906, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is annotated with 21 semantic classes (19 semantics, 1 free, 1 unknown).", + "bbox": [ + 89, + 340, + 480, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SSCBench-KITTI360 [21]. It is a recently introduced dataset for 3D SSC task. SSCBench-KITTI360 contains 9 long sequences, which are split into 7/1/1 for train/val/test. The input resolution of the RGB images is set to $1408 \\times 384$ . The scene voxels in SSCBench-KITTI360 are annotated with 20 semantic classes (18 semantics, 1 free, 1 unknown).", + "bbox": [ + 89, + 369, + 483, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 469, + 308, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Network Structure. Following other comparable methods [3, 12, 44, 50, 51], we employ EfficientNetB7 as our backbone. The multi-scale number $N$ is 4 and the spatial dimensions of the multi-scale image features are $\\left(\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right)$ of an input image, respectively. We set a spatial voxel size $X \\times Y \\times Z$ to $128 \\times 128 \\times 16$ and channel numbers $C$ to 192. The spatial dimensions of the multiscale voxel features are $(1, \\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8})$ of $X \\times Y \\times Z$ , respectively. We set the number of semantic tokens to $N_{\\mathrm{q}} = 100$ while the depth distribution bin to $D = 112$ . In the repository builder, the balancing parameter for novelty scores is set to $\\alpha = 0.2$ and the number of representative tokens of the semantic repository $K$ is set to 10. We use $P = 4$ historical frames. The scene-adaptive decoder is repeated $L = 9$ times. The output voxel is upsampled to $256 \\times 256 \\times 32$ through the trilinear interpolation for the evaluation.", + "bbox": [ + 88, + 491, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training. We train the proposed SOAP for 30 epochs. We employ AdamW [24] optimizer with an initial learning rate 0.0001 and divide it by 10 at the 20 and 25 epochs. For the training, we use a mini-batch of size 4 on 4 RTX A6000 GPUs. For data augmentation, we use random resize, rotation, and flip for the image space and 3D flip for the 3D volume space following [10, 20, 49].", + "bbox": [ + 88, + 733, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 848, + 230, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For evaluation on SemanticKITTI and SSCBench, we use the mean intersection over union (mIoU) of semantic", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/200f57d0c04cb62225973be93071c59fddabb746bcf84eb74bdc85a4cd8e9c3f.jpg", + "table_caption": [ + "Table 3. Complexity-accuracy comparison on SemanticKITTI validation set. † represents methods that use temporal information." + ], + "table_footnote": [], + "table_body": "
MethodImage encoderParams.GFLoPsIoUmIoU
OccFormer [50]EfficientNetB7202.7M973.136.5013.46
\\(HTCL^†\\) [18]EfficientNetB7181.4M2622.445.5117.13
\\(SOAP^†\\)EfficientNetB785.2M705.347.2419.21
Symphonies [13]ResNet5059.3M611.941.9214.89
\\(HASSC^†\\) [37]ResNet5058.1M313.944.5814.74
\\(SGN^†\\) [26]ResNet5027.9M723.746.2115.32
\\(SOAP^†\\)ResNet5045.1M627.848.1218.80
", + "bbox": [ + 514, + 369, + 903, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classes for the SSC task and the intersection over union (IoU) for the class-agnostic scene completion (SC) task.", + "bbox": [ + 511, + 484, + 905, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Highest mIoU and IoU Scores. Tables 1 and 2 compare the proposed SOAP with recent vision-centric 3D SSC algorithms on the SemanticKITTI and SSCBench-KITTI360 test sets, respectively. The scores of other methods come from respective papers. In Table 1, we observe that SOAP significantly outperforms the existing methods on both datasets in terms of the mIoU and IoU scores. Compared to HTCL [18], SOAP improves $+2$ SSC mIoU and $+1.86$ SC IoU on SemanticKITTI test set, which achieves $12\\%$ and $4\\%$ relative improvement. Also, SOAP surpasses SGN [26] on SSCBench-KITTI test set with a margin of $+2.67$ mIoU and $+1.06$ IoU. The highest mIoU and IoU scores demonstrate that the proposed OAP faithfully refines occluded regions in the initial voxel features, while the scene-adaptive decoder further enhances the semantic context of the query embeddings.", + "bbox": [ + 511, + 518, + 906, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency Evaluation. We replace our image encoder, EfficientNetB7 [35] with a more light-weight network, ResNet50 [9]. Table 3 shows complexity-accuracy comparisons between SOAP and other methods on SemanticKITTI validation set, using either ResNet50 or EfficientNetB7 as the image encoder. Compared to OccFormer [50] and HTCL [18], which utilize EfficientNetB7, SOAP requires the lowest parameters and GFLoPs while achieving the highest IoU and mIoU scores, demonstrating the effective", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a6c310b69ba4b07afd41a78d8853b315a4fe221721e9b9e250d0751afde0aa7.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 107, + 95, + 313, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2e4947b1cc23e631baa4df875c9a47c6d4343032fb0ce69000cabc4a2a5c9d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 95, + 472, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9af1d35189661f00a4ad07c3d0207d1cddc88b2d771213721ecfd10ba461bf8.jpg", + "image_caption": [ + "SOAP" + ], + "image_footnote": [], + "bbox": [ + 107, + 167, + 276, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7f0a041533b72f20816e7c8c70b74d07933bdf3ffde4491fd20751154bd6888.jpg", + "image_caption": [ + "Ground-truth", + "OccFormer [50]" + ], + "image_footnote": [], + "bbox": [ + 292, + 167, + 472, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/672e54184cef91cf9038fa9238727e3050da32c3d2dfd2391e11477d83124908.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 109, + 268, + 313, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da67e55e10c275427580c393eb7010a343cb0870758cb80232323493a27fe7c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 266, + 480, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac4d0833fd9aa0db7634d5d704aa87f7780629619e6d10af30e3bbb0cf1df382.jpg", + "image_caption": [ + "SOAP", + "Figure 5. Qualitative results on SemanticKITTI validation set. SOAP demonstrates robustness in handling the occluded regions (top) and distinguishing similar classes (bottom). Red boxes highlight failure cases of OccFormer [50] for comparison." + ], + "image_footnote": [ + "car truck road parking sidewalk building fence" + ], + "bbox": [ + 94, + 340, + 284, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2afc38abe44aeeaf7f9c82234d7e9b83ee8441a6fae2b879dfd995b7ea6905f2.jpg", + "image_caption": [ + "Ground-truth", + "OccFormer [50]" + ], + "image_footnote": [ + "vegetation trunk terrain pole traffic sign" + ], + "bbox": [ + 297, + 340, + 480, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ness and efficiency of SOAP. Also, SOAP with ResNet50 yields remarkable mIoU and IoU performance with comparable complexity compared to other methods.", + "bbox": [ + 89, + 542, + 482, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 5 presents qualitative comparisons between SOAP and OccFormer [50] in terms of occluded region reconstruction and similar class discrimination. In Figure 5 (top), OccFormer fails to recognize occluded regions behind car objects, resulting in fused car shapes, whereas SOAP appropriately distinguishes and separates the cars that are actually apart. Also, as in Figure 5 (bottom), Occformer confuses classes with similar appearance (road $\\leftrightarrow$ sidewalk) while SOAP reasonably differentiates these classes.", + "bbox": [ + 89, + 590, + 482, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 756, + 238, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the contributions of individual components of the proposed SOAP through ablation experiments on the SemanticKITTI validation set. Table 4 reports the contributions of the occluded region-aware projection (OAP) and the scene-adaptive decoder (SAD). For the results without SAD, we replace it with a standard transformer decoder [6, 50, 51]. For the results without OAP, we input the initial voxel features $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ directly into the voxel", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/936df77bb68d12bb8bdbc12d79f55f7a4e70b1b8a2dc5c03de00bf0f713c07c5.jpg", + "table_caption": [ + "Table 4. Ablation study on different combinations of occluded region-aware projection (OAP) and scene-adaptive decoder (SAD)." + ], + "table_footnote": [], + "table_body": "
OAPSADSC IoUSSC mIoU
43.3216.86
✓43.5117.22
✓46.5018.99
✓✓47.2419.21
", + "bbox": [ + 514, + 135, + 903, + 223 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/12a9b73fbd0b3cf4f029fa20d58ea075971619b15eb44a8372492463d52dadb1.jpg", + "table_caption": [ + "Table 5. Alternative approaches for historical information." + ], + "table_footnote": [], + "table_body": "
MethodSC IoUSSC mIoU
Addition45.0018.05
DCA45.9018.53
OAP47.2419.21
", + "bbox": [ + 514, + 253, + 903, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "encoder, without using any historical information. In Table 4, IoU and mIoU scores show significant degradation without either OAP or SAD. This indicates that OAP effectively refines occluded regions during its voxel generation, while SAD provides representative semantic features across dynamic driving scenes for query embeddings. Notably, SOAP still retains reasonable performance even without OAP, which implements 3D SSC without temporal information.", + "bbox": [ + 511, + 345, + 903, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of OAP, we design several alternative approaches to utilize historical information. First, we simply add the historical voxel features $\\mathbf{F}_i^{\\mathrm{hist}}$ to the initial voxel features $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ . Second, we employ the standard deformable cross-attention (DCA) [53] between the initial voxel features $\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}$ and the historical voxel features $\\mathbf{F}_i^{\\mathrm{hist}}$ without occluded region separation. In Table 5, the proposed OAP achieves better IoU and mIoU scores than the standard DCA, with a margin of $+1.34$ and $+0.68$ , respectively. Therefore, perception of occluded regions during view projection is essential for accurate 3D semantic scene reasoning.", + "bbox": [ + 511, + 481, + 905, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 670, + 633, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a novel 3D semantic scene completion algorithm, called SOAP. The main contributions of SOAP are the occluded region-aware view projection and the scene-adaptive decoder. The occluded region-aware view projection addresses the information deficiency in vision-centric 3D SSC by refining occluded voxel regions with historical observation. The scene-adaptive decoder creates the comprehensive semantic repository using the repository builder. Then, the scene-adaptive decoder integrates informative clues from the semantic repository into query embeddings. Experiments demonstrated that SOAP significantly outperforms the state-of-the-arts on the SemanticKITTI and SSC-Bench datasets, with a margin of $+2$ and $+2.67$ mIoU scores, respectively.", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17152", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly by the National Research Foundation of Korea (NRF) grants funded by the Ministry of Education (RS-2022-NR075739) and the Korea government (MSIT) (No. RS-2024-00352566, No. RS-2025-00559165), and partly by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155857, Artificial Intelligence Convergence Innovation Human Resources Development (Chungnam National University)).", + "bbox": [ + 89, + 114, + 483, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 271, + 187, + 286 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6", + "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1", + "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 1, 2, 6, 7", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6", + "[5] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020. 1, 2", + "[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2, 3, 5, 6, 8", + "[7] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161, 2021. 2", + "[8] Emeç Erçelik, Ekim Yurtsever, Mingyu Liu, Zhijie Yang, Hanzhen Zhang, Pinar Topçam, Maximilian Listl, Yilmaz Kaan Cayli, and Alois Knoll. 3d object detection with a self-supervised lidar scene flow backbone. In ECCV, pages 247-265, 2022. 1", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7", + "[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7", + "[11] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, pages 4012-4021, 2022. 1" + ], + "bbox": [ + 93, + 295, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 3, 6, 7", + "[13] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In CVPR, pages 20258-20267, 2024. 1, 2, 3, 6, 7", + "[14] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, pages 515–532, 2020. 1", + "[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6", + "[16] Abhijit Kundu, Yin Li, and James M Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, pages 3559–3568, 2018. 1", + "[17] Hyo-Jun Lee, Hanul Kim, Su-Min Choi, Seong-Gyun Jeong, and Yeong Jun Koh. Baam: Monocular 3d pose and shape reconstruction with bi-contextual attention module and attention-guided modeling. In CVPR, pages 9011-9020, 2023. 1", + "[18] Bohan Li, Jiajun Deng, Wenyao Zhang, Zhujin Liang, Dalong Du, Xin Jin, and Wenjun Zeng. Hierarchical temporal context learning for camera-based semantic scene completion. In ECCV, pages 131-148. Springer, 2024. 1, 3, 6, 7", + "[19] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020. 1, 2", + "[20] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In AAAI, pages 1477-1485, 2023. 7", + "[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 6, 7", + "[22] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, pages 9087-9098, 2023. 1, 2, 3, 6, 7", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7", + "[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021. 1", + "[26] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Jongwon Ra, Yukai Ma, Laijian Li, and Yong Liu. Camera-based" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d semantic scene completion with sparse guidance network. IEEE TIP, 2024. 3, 6, 7", + "[27] Chao Ning and Hongping Gan. Trap attention: Monocular depth estimation with manual traps. In CVPR, pages 5033-5043, 2023. 1", + "[28] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020. 2, 4", + "[29] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. idisc: Internal discretization for monocular depth estimation. In CVPR, pages 21477-21487, 2023. 1", + "[30] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. IEEE TPAMI, 44(10): 7205-7218, 2021. 2", + "[31] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020. 2", + "[32] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 2417-2426, 2022. 3", + "[33] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 1", + "[34] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2", + "[35] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 7", + "[36] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Not all voxels are equal: Semantic scene completion from the point-voxel perspective. In AAAI, pages 2352–2360, 2022. 2", + "[37] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1, 2, 3, 6, 7", + "[38] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 3, 6", + "[39] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In CVPR, pages 17642-17651, 2023. 1, 2", + "[40] Jinfeng Xu, Xianzhi Li, Yuan Tang, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Casfusionnet: A cascaded network for point cloud semantic scene completion by dense feature fusion. In AAAI, pages 3018-3026, 2023. 2", + "[41] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9421-9431, 2023. 3, 6", + "[43] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1", + "[44] Zhu Yu, Running Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Siyuan Cao, and Huiliang Shen. Context and geometry aware voxel transformer for semantic scene completion. In NeurIPS, 2024. 3, 7", + "[45] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3916-3925, 2022. 1", + "[46] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In ICCV, pages 9155-9166, 2023. 1", + "[47] Shoulong Zhang, Shuai Li, Aimin Hao, and Hong Qin. Point cloud semantic scene completion from rgb-d images. In AAAI, pages 3385-3393, 2021. 2", + "[48] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, pages 1589-1598, 2022. 1", + "[49] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 7", + "[50] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 1, 2, 3, 5, 6, 7, 8", + "[51] Linqing Zhao, Xiuwei Xu, Ziwei Wang, Yunpeng Zhang, Borui Zhang, Wenzhao Zheng, Dalong Du, Jie Zhou, and Jiwen Lu. Lowrankocc: Tensor decomposition and low-rank recovery for vision-based 3d semantic occupancy prediction. In CVPR, pages 9806-9815, 2024. 1, 2, 3, 5, 6, 7, 8", + "[52] Yunsong Zhou, Hongzi Zhu, Quan Liu, Shan Chang, and Minyi Guo. Monoatt: Online monocular 3d object detection with adaptive token transformer. In CVPR, pages 17493-17503, 2023. 1", + "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 5, 6, 8", + "[54] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 813 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17154", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_model.json b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9b88d823af4812839fbb78c7f5d9e62561add535 --- /dev/null +++ b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "SOAP: Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.203, + 0.397, + 0.22 + ], + "angle": 0, + "content": "Hyo-Jun Lee" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.222, + 0.431, + 0.24 + ], + "angle": 0, + "content": "Hyunseop Kim" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.204, + 0.575, + 0.222 + ], + "angle": 0, + "content": "Yeong Jun \\(\\mathrm{Koh}^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.472, + 0.222, + 0.578, + 0.24 + ], + "angle": 0, + "content": "Yonguk Lee" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.204, + 0.714, + 0.221 + ], + "angle": 0, + "content": "Hanul Kim3" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.222, + 0.696, + 0.238 + ], + "angle": 0, + "content": "Jinu Lee" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.242, + 0.474, + 0.26 + ], + "angle": 0, + "content": "\\(^{1}\\)Kangwon National University" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.242, + 0.771, + 0.261 + ], + "angle": 0, + "content": "\\(^{2}\\)Chungnam National University" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.261, + 0.65, + 0.279 + ], + "angle": 0, + "content": "\\(^{3}\\)Seoul National University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.689, + 0.261, + 0.78, + 0.278 + ], + "angle": 0, + "content": "42dot Inc." + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.282, + 0.406, + 0.296 + ], + "angle": 0, + "content": "gywns6298@kangwon.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.283, + 0.575, + 0.295 + ], + "angle": 0, + "content": "yjkoh@cnu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.61, + 0.282, + 0.798, + 0.295 + ], + "angle": 0, + "content": "hukim@seoultech.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.3, + 0.391, + 0.313 + ], + "angle": 0, + "content": "hyunseop95@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.3, + 0.594, + 0.313 + ], + "angle": 0, + "content": "yonguk.lee@42dot.ai" + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.3, + 0.782, + 0.313 + ], + "angle": 0, + "content": "jinu.lee@42dot.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.348, + 0.327, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.485, + 0.683 + ], + "angle": 0, + "content": "Existing view transformations in vision-centric 3D Semantic Scene Completion (SSC) inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, leading to a dilution of informative contexts. Furthermore, semantic classes exhibit high variability in their appearance in real-world driving scenarios. To address these issues, we introduce a novel 3D SSC method, called SOAP, including two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection effectively converts 2D image features into voxel space, refining the duplicated features of occluded regions using information gathered from previous observations. The scene-adaptive decoder guides query embeddings to learn diverse driving environments based on a comprehensive semantic repository. Extensive experiments validate that the proposed SOAP significantly outperforms existing methods for the vision-centric 3D SSC on automated driving datasets, SemanticKITTI and SSCBench. Code is available at https://github.com/gywns6287/SOAP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.222, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Understanding 3D traffic scenes is crucial for autonomous driving systems to develop downstream tasks such as motion prediction and planning. LiDAR sensors [8, 33, 43, 54] provide precise 3D scene information, but their high cost and the complex setup limit their scalability. In contrast, areas of study about vision-centric 3D perception, including 3D object detection [11, 25, 46, 48, 52], 3D car shape reconstruction [14, 16, 17], and depth estimation [2, 27, 29, 45], have been receiving attention due to its simplicity and cost-effectiveness. However, these methods still struggle to re" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.347, + 0.909, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.473, + 0.784, + 0.485 + ], + "angle": 0, + "content": "(a) Occluded voxel regions." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.491, + 0.909, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.635, + 0.845, + 0.647 + ], + "angle": 0, + "content": "(b) Diversity in the appearance of semantic classes." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.658, + 0.908, + 0.729 + ], + "angle": 0, + "content": "Figure 1. Two main motivations of SOAP. (a) The 2D-to-3D view projection inevitably yields erroneous feature duplication in voxel generation due to occlusion. (b) The semantic classes exhibit significant variability in their appearance depending on driving regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "construct complete and fine-grained 3D scene representations. To address this limitation, 3D Semantic Scene Completion (SSC) has emerged as a promising alternative. 3D SSC leverages a voxel-based approach, assigning both an occupancy probability and a semantic label (e.g. car, pedestrian, road) to each voxel within the 3D space. Thereby, it can create a comprehensive representation of the driving scene. While earlier works [5, 19, 39, 41] primarily rely on LiDAR sensors, recent research [3, 13, 18, 22, 37, 50, 51] has demonstrated the potential of vision-centric methods." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17145" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "A general vision-based 3D SSC model comprises an image encoder, a view projection module, and a voxel decoder. The image encoder first extracts 2D features from an input image, and then the view projection module subsequently lifts 2D image features into 3D voxel space. Given the 3D features of each voxel region, the voxel decoder estimates semantic classes corresponding to each voxel region. In this process, the precise 2D-3D view projection plays a crucial role. However, existing view projection methods [3, 50] inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, particularly along the longitudinal direction. For instance, as illustrated in Figure 1a, when a camera ray encounters an occupancy region (e.g. car), occluded regions (voxels behind this occupancy) may contain duplicated features of the obstructing car in the 3D view projection process. This duplication can occur even when the actual classes of the occluded regions are not related to a car, resulting in incorrect 3D SSC results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Furthermore, the diverse appearances of semantic classes pose another significant challenge in 3D SSC. As shown in Figure 1b, semantic classes exhibit high variability in their appearance in real-world driving scenarios. For instance, road color and texture may vary depending on driving regions (e.g. urban or highway). Such appearance variations make it difficult for SSC models to distinguish similar classes, such as road \\(\\leftrightarrow\\) sidewalk, leading to potential misinterpretations of driving environments. While most previous works [3, 13, 22, 37] usually employ sequential 3D convolutions for the semantic decoder, this approach often lacks the flexibility to generalize for the diverse semantic context. More recently, Mask2former [6]-like approach [50, 51] is adopted for the SSC decoder, which utilizes semantic tokens derived from a query embedding to represent the semantic scene. However, these semantic tokens lack adaptability to dynamic driving environments, as they are fundamentally based on static query embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel 3D semantic scene completion network, SOAP, faithfully designed to address the occluded voxel regions and the visual variations in semantic classes. The proposed method leverages two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection (OAP) first converts 2D image features into 3D voxel features using the LSS algorithm [28], where these initial voxel features contain false duplicated information due to occlusions, as illustrated in Figure 1a. These occluded regions can be inferred by referring to previous frames, which provides valuable context about the missing observation. Therefore, OAP identifies occluded regions in the current frame using depth estimation and refines the duplicated features of occluded regions using information gathered from previous observations. Also, a vanilla transformer decoder [6] generates the semantic tokens based on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "static query embedding, which is not versatile enough to capture a dynamic driving environment. To represent more diverse semantics, a repository builder first creates a semantic repository by gathering semantic tokens from various scenes. Then, the scene-adaptive decoder integrates informative clues from this extensive repository into the semantic tokens, guiding them to navigate diverse driving environments effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that the proposed SOAP outperforms the other state-of-the-arts on two vision-centric SSC benchmarks, SemanticKITTI [1] and SSCBench-KITTI360 [21]. To summarize, the main contributions of our work are three folds:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.288, + 0.905, + 0.333 + ], + "angle": 0, + "content": "- We develop the occluded region-aware view projection that effectively transforms 2D image features into 3D voxel space by recovering occluded voxel regions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.378 + ], + "angle": 0, + "content": "- We also introduce a scene-adaptive transformer that integrates informative semantic clues from the various scenes into the semantic tokens of an individual scene." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.379, + 0.905, + 0.439 + ], + "angle": 0, + "content": "- The proposed SOAP achieves outstanding performances on two driving datasets, SemanticKITTI and SSCBenchKITTI360, with a margin of \\(+2\\) and \\(+2.67\\) mIoU scores, respectively." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.288, + 0.905, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.453, + 0.663, + 0.468 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.854 + ], + "angle": 0, + "content": "3D Semantic Scene Completion. 3D semantic scene completion (SSC) is an advanced task in 3D scene reconstruction fields, which aims to predict a dense semantic scene representation involving the geometrical occupancy and its corresponding semantic label. Early SSC methods [5, 19, 34] leverage 3D inputs to reconstruct voxelized indoor scenes. SSCNet [34] first defines the SSC task, which jointly reasons the geometry and semantics of an indoor scene from a single-depth image. 3DSketch [5] uses the Truncated Signed Distance Function to guide an occupancy and the semantic labels of the indoor scene. AIC-Net [19] develops anisotropic convolution to extract hybrid features from the depth and RGB images for indoor SSC tasks. Recently, the release of SemanticKITTI [1], the first large-scale outdoor SSC dataset, yields the development of diverse outdoor SSC methods. LMSCNet [31] uses the 2D UNet to extract collapsed BEV features from sparse 3D Li-DAR scans and finally expands the height dimension for the 3D SSC. JS3C-Net [41] learns the contextual shape priors from sequential LiDAR data completing the sparse single sweep point cloud to the dense one for enhanced 3D SSC. The other existing works [7, 30, 36, 39, 40, 47] also commonly depend on 3D input such as LiDAR point clouds. In contrast, we address vision-centric 3D SSC, which only uses an RGB image for the semantic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Vision-centric 3D SSC. MonoScene [3] introduces the first vision-centric approach for SSC, utilizing a 3D UNet to process voxel features generated by sight projection. Vox-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17146" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.086, + 0.889, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.404, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed SOAP. Given multi-scale image features, an occluded region-aware view projection (OAP) transforms them into voxel space, refining duplicated features in occluded voxel regions using historical information. A scene-adaptive decoder enriches query embeddings with diverse semantic contexts through a semantic repository, which is constructed and iteratively updated by a repository builder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.486, + 0.744 + ], + "angle": 0, + "content": "Former [22] estimates occupancy regions using stereo depth estimation and selectively updates these regions with RGB images. HASSC [37] and Symphonies [13] also identify voxel occupancy based on stereo depth estimation, following VoxFormer. NDC-scene [42] extends a 2D feature map to a normalized device coordinates space to restore the depth dimension and subsequently reconstructs a 3D semantic scene using this space. SurroundOcc [38] adopts spatial 2D-3D attention to lift multi-scale 2D features into the 3D volume space. TPVFormer [12] introduces a query-based tri-perspective space for 3D SSC. Occformer [50] and LowRankOcc [51] reconstruct 3D scenes using semantic tokens on the Mask2former decoders [6]. SGN [26] propagates semantic contexts from semantic-aware seed voxels based on their spatial geometry cues. HTCL [18] leverages historical frames to refine current frame features through affinity-based dynamic refinement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.761, + 0.184, + 0.777 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.254, + 0.803 + ], + "angle": 0, + "content": "3.1. Model Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "3D Semantic Scene Completion (SSC) aims to predict semantic probabilities of the voxelized scene \\(\\mathbf{V} \\in \\mathbb{R}^{X \\times Y \\times Z \\times S}\\), where \\((X, Y, Z)\\) denotes the resolution of the 3D volume, and \\(S\\) is the number of semantic classes including a free (or empty) class. Figure 2 illustrates the overall architecture of the proposed SOAP. Image encoder first ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.592 + ], + "angle": 0, + "content": "tracts multi-scale image features \\(\\{\\mathbf{F}_i^{2\\mathrm{D}} \\in \\mathbb{R}^{H_i \\times W_i \\times C}\\}_{i=1}^N\\) from an RGB image, where \\(N\\) is the number of scales, \\((H_i, W_i)\\) is the spatial size at the scale \\(i\\), and \\(C\\) is the channel dimension. Depth estimator predicts the depth map \\(\\mathbf{E} \\in \\mathbb{R}^{H \\times W}\\) from RGB images. Consistent with previous approaches [13, 18, 22, 37, 44], we also use off-the shelf stereo depth model [32] for the depth estimator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.804 + ], + "angle": 0, + "content": "The occluded region-aware view projection (Section 3.2) utilizes the estimated depth map \\(\\mathbf{E}\\) to lift multi-scale image features \\(\\{\\mathbf{F}_i^{2\\mathrm{D}}\\}_{i = 1}^N\\) into multi-scale voxel features \\(\\{\\mathbf{F}_i^{3\\mathrm{D}}\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i\\times C}\\}_{i = 1}^N\\), enhancing occluded voxel regions based on informative features in historical frames. The scene-adaptive decoder (Section 3.3) first constructs a semantic repository using a repository builder and then it integrates semantic diversity from a semantic repository into query embeddings to produce scene-adaptive semantic tokens, which are utilized to reconstruct the final semantic voxel representation \\(\\mathbf{V}\\), in conjunction with multiscale voxel features. Through this complementary process, SOAP guarantees a more flexible and accurate 3D SSC for a real-world driving environment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.81, + 0.866, + 0.826 + ], + "angle": 0, + "content": "3.2. Occluded Region-Aware View Projection" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Occluded region-aware view projection (OAP) transforms 2D image features \\(\\{\\mathbf{F}_i^{2\\mathrm{D}}\\}_{i = 1}^N\\) into 3D voxel space \\(\\{\\mathbf{F}_i^{3\\mathrm{D}}\\}_{i = 1}^N\\). Figure 3 illustrates the detailed process of OAP. It produces voxel features for each \\(i\\)-th scale solely from the image feature at \\(i\\)-th scale." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.877, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 3. Occluded region-aware view projection. (a) occluded region identification identifies occluded regions in the current frame by inferring an occlusion map based on the depth distribution. (b) occluded region refinement refines the duplicated features in the identified occluded regions using previous observations within the historical voxel features." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.392, + 0.481, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Figure 4. t-SNE visualization of voxel features in the occluded regions for (a) \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\) and (b) \\(\\tilde{\\mathbf{F}}_i^{\\mathrm{occ}}\\). Different colors represent different camera rays." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Initial Voxel Generation. For each \\(i\\)th scale, OAP first generates initial voxel features \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}\\) from the 2D image features \\(\\mathbf{F}_i^{2\\mathrm{D}}\\) and the depth estimation \\(\\mathbf{E}\\). Specifically, OAP resizes the estimated depth \\(\\mathbf{E}\\) to \\(\\mathbf{E}_i\\) that has the spatial resolution of \\((H_i, W_i)\\), and distributes it to the depth distribution map \\(\\mathbf{D}_i \\in \\mathbb{R}^{H_i \\times W_i \\times D}\\). Here, the continuous depth value \\(\\mathbf{E}_i(h, w)\\) for each pixel \\((h, w)\\) is converted into discrete depth bins \\(\\mathbf{D}_i(h, w) \\in \\mathbb{R}^D\\) represented as one-hot encoded vectors, followed by a Gaussian filter to create a soft depth distribution. Then, OAP computes the outer product \\(\\mathbf{F}_i^{2\\mathrm{D}} \\otimes \\mathbf{D}_i\\) to create the depth distributed features \\(\\mathbf{F}_i^{\\mathrm{depth}} \\in \\mathbb{R}^{H_i \\times W_i \\times D \\times C}\\). The voxel-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.56 + ], + "angle": 0, + "content": "pooling operation \\(\\varphi(\\cdot)\\) in [28] can project \\(\\mathbf{F}_i^{\\mathrm{depth}}\\) into the voxel space using camera intrinsics and extrinsics. Thus, the initial voxel features are defined as \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}} = \\varphi(\\mathbf{F}_i^{\\mathrm{depth}})\\). Historical Voxel Generation. To represent previous observations, we incorporate depth distributed features \\(\\{\\mathbf{F}_{i,t - p}^{\\mathrm{depth}}\\}_{p = 1}^{P}\\) from the previous frames into the current voxel space \\(\\mathbf{F}_i^{\\mathrm{hist}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}\\), where \\(P\\) is the number of previous frames at timestamp \\(t\\). Specifically, each past \\(\\mathbf{F}_{i,t - p}^{\\mathrm{depth}}\\) is projected into voxel space \\(\\varphi(\\mathbf{F}_{i,t - p}^{\\mathrm{depth}})\\) and then aligned to the current voxel space at \\(t\\) via the warping function \\(\\mathcal{W}(\\cdot)\\) and summation operation:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.566, + 0.907, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {\\mathrm {h i s t}} = \\sum_ {p = 1} ^ {P} \\mathcal {W} \\left(\\varphi \\left(\\mathbf {F} _ {i, t - p} ^ {\\mathrm {d e p t h}}\\right), \\mathcal {G} _ {t} ^ {- 1} \\cdot \\mathcal {G} _ {t - p}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_t\\) and \\(\\mathcal{G}_{t - p}\\) are ego-posed matrices at \\(t\\) and \\(t - p\\). These historical voxel features \\(\\mathbf{F}_i^{\\mathrm{hist}}\\) provide valuable perspectives that are not available in the initial voxel features. Occluded Region Identification. The initial voxel features \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\), which are projected from a single image view, inherently include false duplicated features due to occlusions. Figure 4a presents the t-SNE visualization of features in the occluded regions within \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\). Each point represents an individual voxel region, and voxel regions of the same color correspond to those located along the same camera ray. We observe that most occluded regions along the same camera ray exhibit similar patterns, which indicates erroneous replication of the features from the occupancy regions positioned in front of the occluded regions as in Figure 1a." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To address this limitation, OAP refines these occluded features by referring to previous frames, which provide valuable contexts about the missing observation in the current frame. As part of this process, we identify the occluded regions using the estimated depth distribution \\(\\mathbf{D}_i\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17148" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.213 + ], + "angle": 0, + "content": "As in Figure 3a, each vector \\(\\mathbf{D}_i(h,w)\\) at the spatial location \\((h,w)\\) contains the probability distribution of the nearest occupancy region along the longitudinal (or depth) direction. In other words, regions located behind the nearest occupancy with the highest depth probability often lose their semantic context information due to occlusions. To distinguish the occluded regions from non-occluded regions, we define an occlusion map \\(\\mathbf{O}_i\\in \\{1,0\\}^{H_i\\times W_i\\times D}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.226, + 0.484, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {i} (h, w, d) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d > \\underset {\\tilde {d} \\in \\{1, \\dots , D \\}} {\\arg \\max } \\mathbf {D} _ {i} (h, w, \\tilde {d}) \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Figure 3a provides an example of \\(\\mathbf{O}_i\\), where occluded regions are marked with 1 (taller rectangular) and non-occluded regions are marked with 0 (shorter rectangular). To identify occluded regions in the voxel space, the occlusion map \\(\\mathbf{O}_i\\) is projected into voxel space through the voxel pooling operation, i.e. \\(\\varphi (\\mathbf{O}_i)\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i}\\). Note that the regions identified by this occlusion map \\(\\varphi (\\mathbf{O}_i)\\) contain incorrectly duplicated information from the nearest occupancy region along the longitudinal direction, as illustrated in Figure 4a. Therefore, we refine the features of these occluded regions through an occluded region refinement process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Occluded Region Refinement. The occluded regions at the current time step may have been observable in previous time steps. Therefore, we refine deficient features in the occluded regions with previous observations in the historical voxel features \\(\\mathbf{F}_i^{\\mathrm{hist}}\\). To achieve this, we perform 3D deformable cross-attention, DCA [53] between the historical voxel features \\(\\mathbf{F}_i^{\\mathrm{hist}}\\) and the occluded voxel features, resulting in reconstructed occluded voxel features \\(\\mathbf{F}_i^{\\mathrm{occ}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {\\text {o c c}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z) + \\mathrm {D C A} \\left(\\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z), \\mathbf {F} _ {i} ^ {\\text {h i s t}}\\right), \\\\ \\quad \\text {i f} \\varphi (\\mathbf {O} _ {i}) (x, y, z) \\geq 1 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\((x,y,z)\\) are spatial indices of the voxel space. This approach ensures that the reconstructed features \\(\\mathbf{F}_i^{\\mathrm{occ}}\\) effectively capture observable historical information about occluded regions, enabling more accurate and context-aware scene representation. As shown in Figure 4b, the reconstructed features \\(\\mathbf{F}_i^{\\mathrm{occ}}\\) include appropriately corrected information for the occluded regions, resulting in more diverse patterns according to different semantic regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Invisible and Visible Region Reconstruction. The initial voxel features also contain invisible regions due to the limited observation of a single camera. Since these invisible voxels do not contain any features, we directly fill them with the corresponding historical voxel features to recon-" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.09, + 0.871, + 0.109 + ], + "angle": 0, + "content": "struct invisible voxel features \\(\\mathbf{F}_i^{\\mathrm{inv}}\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i\\times C}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.117, + 0.905, + 0.177 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {\\text {i n v}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\mathbf {F} _ {i} ^ {\\text {h i s t}} (x, y, z) & \\text {i f} \\sum \\tilde {\\mathbf {F}} _ {i} ^ {\\text {3 D}} (x, y, z) = 0 \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.178, + 0.907, + 0.282 + ], + "angle": 0, + "content": "We designate the voxel space, excluding invisible and occluded regions, as visible regions. The visible regions already contain appropriate semantic features with respect to the current observation, allowing us to directly retrieve the visible voxel features from the initial voxel features \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\). The visible voxel features \\(\\mathbf{F}_i^{\\mathrm{vis}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}\\) are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.289, + 0.905, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {\\mathrm {v i s}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z) & \\text {i f} \\varphi (\\mathbf {O} _ {i}) (x, y, z) = 0 \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.351, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Voxel Encoder. To compute the final 3D voxel features, we sum the occluded, invisible, and visible voxel features, \\(\\mathbf{F}_i^{\\mathrm{occ}} + \\mathbf{F}_i^{\\mathrm{inv}} + \\mathbf{F}_i^{\\mathrm{vis}}\\), and process them through a voxel encoder to obtain the final voxel features \\(\\mathbf{F}_i^{3\\mathrm{D}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i}\\). In the voxel encoder, the spatial dimensions \\((X_i \\times Y_i \\times Z_i)\\) of the voxel features are sliced into \\(Z_i\\) bird's-eye view (BEV) planes with spatial resolution \\((X_i \\times Y_i)\\). Then, two windowed self-attention blocks [23] are sequentially performed to each BEV plane." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.495, + 0.738, + 0.511 + ], + "angle": 0, + "content": "3.3. Scene-Adaptive Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.909, + 0.835 + ], + "angle": 0, + "content": "A standard transformer decoder [6, 50, 51] yields semantic tokens starting from static query embeddings, which lack adaptability for dynamic driving scenes. To represent more diverse semantics, the scene-adaptive decoder correlates semantic tokens for a specific scene with tokens from different scenes. As in Figure 2, the scene-adaptive decoder generates scene-adaptive tokens \\(\\mathbf{Q} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times C}\\) through \\(L\\) iterative transformer blocks, where \\(N_{\\mathrm{q}}\\) is the number of tokens. Each transformer block in the scene transformer cascades four modules including a cross-attention between query embeddings and the semantic repository, the other cross-attention between the query embeddings and the voxel features \\(\\mathbf{F}_i^{3\\mathrm{D}}\\), self-attention, and feed-forward networks. The cross-attention with the semantic repository integrates semantic features from dynamic scenes into the query embeddings, while the cross-attention with voxel features provides scene-specific semantic features. After the end of iterations, scene-adaptive tokens \\(\\mathbf{Q}\\) are projected to class predictions \\(\\mathbf{S}^{\\mathrm{cls}} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}\\) and segmentation mask predictions \\(\\mathbf{M} \\in \\mathbb{R}^{X \\times Y \\times Z \\times N_{\\mathrm{q}}}\\). Then, the final semantic voxel \\(\\mathbf{V} \\in \\mathbb{R}^{X \\times Y \\times Z \\times S}\\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.843, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\mathcal {F} ^ {- 1} (\\mathcal {F} (\\mathbf {M}) \\mathbf {S} ^ {\\mathrm {c l s}}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.869, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}:\\mathbb{R}^{X\\times Y\\times Z\\times C}\\to \\mathbb{R}^{XYZ\\times C}\\) is as spatial flatten and \\(\\mathcal{F}^{-1}\\) is its inverse operation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17149" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Semantic scene completion results on SemanticKITTI test set. † represents temporal methods that leverage previous frames for 3D SSC. We highlight the best results in bold and the second best results in underline." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.128, + 0.907, + 0.413 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoUroad (15.30%)sidewalk (11.13%)parking (1.12%)other-ground (0.56%)building (14.1%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation(39.3%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (3.90%)pole (0.29%)traf-sign (0.08%)
MonoScene [3]34.1611.0854.727.124.85.714.418.83.30.50.74.414.92.419.51.01.40.411.13.32.1
TPVFormer [12]34.2511.2655.127.227.46.514.819.23.71.00.52.313.92.620.41.12.40.311.02.91.5
SurroundOcc [38]34.7211.8656.928.330.26.815.220.61.41.61.24.414.93.419.31.42.00.111.33.92.4
VoxFormer† [22]43.2113.4154.126.925.17.323.521.73.61.91.64.124.48.124.21.61.10.013.16.65.7
OccFormer [50]34.5312.3255.930.331.56.515.721.61.21.51.73.216.83.921.32.21.10.211.93.83.7
NDC-Scene [42]33.8711.5556.228.728.05.615.819.71.81.11.14.914.32.620.60.71.70.411.23.21.7
LowRankOcc [51]38.4713.5652.827.225.18.822.120.92.93.32.74.422.98.920.82.41.72.314.47.07.0
Symphonies [13]42.1915.0458.429.326.911.724.723.63.23.62.65.624.210.023.13.21.92.016.17.78.0
HASSC† [37]42.8714.3855.329.625.911.323.123.02.91.91.54.924.89.826.51.43.00.014.37.07.1
SGN† [26]45.4215.7660.431.428.98.728.425.44.50.91.63.727.412.628.40.50.30.118.110.08.3
HTCL† [18]44.2317.0964.434.833.812.425.927.35.71.82.25.425.310.831.21.13.10.921.19.08.3
\\(SOAP^{\\dagger}\\) (proposed)46.0919.0963.636.236.817.228.728.93.35.05.37.029.815.132.33.72.30.222.511.713.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.486, + 0.695 + ], + "angle": 0, + "content": "Repository Builder. The semantic repository \\(\\mathbf{R} \\in \\mathbb{R}^{(S \\times K) \\times C}\\) stores \\(K\\) representative tokens for each semantic class, iteratively updated by semantic tokens with top- \\(K\\) highest scores. The detailed process of the repository builder is shown in Figure 2. Even though representative tokens could be selected using the classification scores \\(\\mathbf{S}^{\\mathrm{cls}}\\) of the scene-adaptive tokens \\(\\mathbf{Q}\\), relying solely on \\(\\mathbf{S}^{\\mathrm{cls}}\\) makes a repository filled with similar token and consequently degrades semantic diversity. To retain the semantic diversity within the repository \\(\\mathbf{R}\\), we introduce a novelty-based token selection. Specifically, we re-define token scores \\(\\mathbf{S} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}\\) as \\(\\mathbf{S} = \\mathbf{S}^{\\mathrm{cls}} + \\alpha \\mathbf{S}^{\\mathrm{nov}}\\), where \\(\\mathbf{S}^{\\mathrm{nov}} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}\\) is novelty scores and \\(\\alpha\\) is its balancing parameter. To compute the novelty scores, the \\(K\\) tokens within each class in \\(\\mathbf{R}\\) are first averaged, resulting in the class average repository \\(\\bar{\\mathbf{R}} \\in \\mathbb{R}^{S \\times C}\\). Then, the novelty scores \\(\\mathbf{S}^{\\mathrm{nov}}\\) are calculated based on cosine similarity between each token in \\(\\mathbf{Q}\\) and \\(\\bar{\\mathbf{R}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.71, + 0.483, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} ^ {\\text {n o v}} (n, s) = - \\cos (\\mathbf {Q} (n), \\bar {\\mathbf {R}} (s)) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.779 + ], + "angle": 0, + "content": "where \\( n \\) and \\( s \\) indicate token indices of \\( \\mathbf{Q} \\) and \\( \\bar{\\mathbf{R}} \\), respectively, and \\( \\cos(\\cdot, \\cdot) \\) is a cosine similarity between two input vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Then, new tokens in \\(\\mathbf{Q}\\) with higher scores than old tokens in \\(\\mathbf{R}\\) are selected and, newly pushed into the semantic repository \\(\\mathbf{R}\\). When a class holds more than \\(K\\) tokens, the old tokens with lower scores are discarded to maintain \\(K\\) representative tokens for each semantic class in \\(\\mathbf{R}\\). Updating \\(\\mathbf{R}\\) with tokens that exhibit both high classification and novelty scores enables the repository to capture rich semantic information, effectively covering dynamic driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.484 + ], + "angle": 0, + "content": "scenes. Note that the semantic repository \\(\\mathbf{R}\\) is updated only during the SOAP's training phase to ensure stable inference performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.667, + 0.51 + ], + "angle": 0, + "content": "3.4. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.593 + ], + "angle": 0, + "content": "We perform the Hungarian matching [15] between predicted and ground-truth voxel masks following [4, 6, 53]. For each matched pair, we compute a mask loss \\( L_{\\mathrm{mask}} \\), a mask dice loss \\( L_{\\mathrm{dice}} \\), and a class loss \\( L_{\\mathrm{cls}} \\) as in [6]. The final loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.609, + 0.907, + 0.626 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {\\text {m a s k}} L _ {\\text {m a s k}} + \\lambda_ {\\text {d i c e}} L _ {\\text {d i c e}} + \\lambda_ {\\text {c l s}} L _ {\\text {c l s}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.669 + ], + "angle": 0, + "content": "where, scaling weights \\(\\lambda_{\\mathrm{mask}}, \\lambda_{\\mathrm{dice}}, \\lambda_{\\mathrm{cls}}\\) are set to 5, 5, 2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.646, + 0.703 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.616, + 0.725 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.838 + ], + "angle": 0, + "content": "We evaluate the performance of SOAP on two outdoor 3D SSC datasets: SemanticKITTI [1] and SSCBenchKITTI360 [21]. Both datasets provide front-view RGB images, LiDAR sweeps, and semantic scene labels represented as \\(256 \\times 256 \\times 32\\) voxel grids. The scene labels cover a range of \\(51.2m \\times 51.2m \\times 6.4m\\) from the front view of the vehicle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "SemanticKITTI [1]. It contains a total of 22 sequences, which are split into \\(10/1/11\\) for train/val/test following [3]. We utilize the RGB images of cam-2 with a \\(1280 \\times 384\\) resolution as model inputs. In the semantic labels, each voxel" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17150" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Semantic scene completion results on the SSCBench test set. † represents methods that use temporal information. We highlight the best results in **bold** and the second best results in **underline**." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.128, + 0.907, + 0.317 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoUcarbicyclemotorcycletruckother-veh.personroadparkingsidewalkother-groundbuildingfencevegetationterrainpoletraf.-signother-struct.other-object
MonoScene [3]37.8712.3119.30.40.68.02.00.948.411.428.13.232.93.526.216.86.95.74.23.1
TPVFormer [12]40.2213.6421.61.11.48.12.62.453.012.031.13.834.84.830.117.57.55.95.52.7
OccFormer [50]40.2713.8122.60.70.39.93.82.854.313.431.53.636.44.831.019.57.88.57.04.6
VoxFormer†[22]38.7611.9117.81.20.94.62.11.647.09.727.22.931.25.029.014.76.56.93.82.4
Symphonies [13]44.1218.5830.01.95.925.112.18.254.913.832.86.935.18.638.311.514.09.614.411.3
SGN†[26]47.0618.2529.03.42.910.95.23.058.115.036.44.442.07.738.223.216.716.49.95.8
\\( \\mathrm{SOAP}^{\\dagger } \\) (proposed)48.1220.9229.95.67.814.47.66.160.917.440.35.445.310.640.524.816.821.012.69.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.482, + 0.371 + ], + "angle": 0, + "content": "is annotated with 21 semantic classes (19 semantics, 1 free, 1 unknown)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.462 + ], + "angle": 0, + "content": "SSCBench-KITTI360 [21]. It is a recently introduced dataset for 3D SSC task. SSCBench-KITTI360 contains 9 long sequences, which are split into 7/1/1 for train/val/test. The input resolution of the RGB images is set to \\(1408 \\times 384\\). The scene voxels in SSCBench-KITTI360 are annotated with 20 semantic classes (18 semantics, 1 free, 1 unknown)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.31, + 0.487 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Network Structure. Following other comparable methods [3, 12, 44, 50, 51], we employ EfficientNetB7 as our backbone. The multi-scale number \\( N \\) is 4 and the spatial dimensions of the multi-scale image features are \\( \\left(\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right) \\) of an input image, respectively. We set a spatial voxel size \\( X \\times Y \\times Z \\) to \\( 128 \\times 128 \\times 16 \\) and channel numbers \\( C \\) to 192. The spatial dimensions of the multiscale voxel features are \\( (1, \\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}) \\) of \\( X \\times Y \\times Z \\), respectively. We set the number of semantic tokens to \\( N_{\\mathrm{q}} = 100 \\) while the depth distribution bin to \\( D = 112 \\). In the repository builder, the balancing parameter for novelty scores is set to \\( \\alpha = 0.2 \\) and the number of representative tokens of the semantic repository \\( K \\) is set to 10. We use \\( P = 4 \\) historical frames. The scene-adaptive decoder is repeated \\( L = 9 \\) times. The output voxel is upsampled to \\( 256 \\times 256 \\times 32 \\) through the trilinear interpolation for the evaluation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.734, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Training. We train the proposed SOAP for 30 epochs. We employ AdamW [24] optimizer with an initial learning rate 0.0001 and divide it by 10 at the 20 and 25 epochs. For the training, we use a mini-batch of size 4 on 4 RTX A6000 GPUs. For data augmentation, we use random resize, rotation, and flip for the image space and 3D flip for the 3D volume space following [10, 20, 49]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.231, + 0.863 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For evaluation on SemanticKITTI and SSCBench, we use the mean intersection over union (mIoU) of semantic" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.339, + 0.906, + 0.368 + ], + "angle": 0, + "content": "Table 3. Complexity-accuracy comparison on SemanticKITTI validation set. † represents methods that use temporal information." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.37, + 0.905, + 0.468 + ], + "angle": 0, + "content": "
MethodImage encoderParams.GFLoPsIoUmIoU
OccFormer [50]EfficientNetB7202.7M973.136.5013.46
\\(HTCL^†\\) [18]EfficientNetB7181.4M2622.445.5117.13
\\(SOAP^†\\)EfficientNetB785.2M705.347.2419.21
Symphonies [13]ResNet5059.3M611.941.9214.89
\\(HASSC^†\\) [37]ResNet5058.1M313.944.5814.74
\\(SGN^†\\) [26]ResNet5027.9M723.746.2115.32
\\(SOAP^†\\)ResNet5045.1M627.848.1218.80
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.517 + ], + "angle": 0, + "content": "classes for the SSC task and the intersection over union (IoU) for the class-agnostic scene completion (SC) task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Highest mIoU and IoU Scores. Tables 1 and 2 compare the proposed SOAP with recent vision-centric 3D SSC algorithms on the SemanticKITTI and SSCBench-KITTI360 test sets, respectively. The scores of other methods come from respective papers. In Table 1, we observe that SOAP significantly outperforms the existing methods on both datasets in terms of the mIoU and IoU scores. Compared to HTCL [18], SOAP improves \\(+2\\) SSC mIoU and \\(+1.86\\) SC IoU on SemanticKITTI test set, which achieves \\(12\\%\\) and \\(4\\%\\) relative improvement. Also, SOAP surpasses SGN [26] on SSCBench-KITTI test set with a margin of \\(+2.67\\) mIoU and \\(+1.06\\) IoU. The highest mIoU and IoU scores demonstrate that the proposed OAP faithfully refines occluded regions in the initial voxel features, while the scene-adaptive decoder further enhances the semantic context of the query embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Efficiency Evaluation. We replace our image encoder, EfficientNetB7 [35] with a more light-weight network, ResNet50 [9]. Table 3 shows complexity-accuracy comparisons between SOAP and other methods on SemanticKITTI validation set, using either ResNet50 or EfficientNetB7 as the image encoder. Compared to OccFormer [50] and HTCL [18], which utilize EfficientNetB7, SOAP requires the lowest parameters and GFLoPs while achieving the highest IoU and mIoU scores, demonstrating the effective" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17151" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.096, + 0.314, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.148, + 0.246, + 0.159 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.096, + 0.473, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.156, + 0.447, + 0.166 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.169, + 0.277, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.236, + 0.183, + 0.247 + ], + "angle": 0, + "content": "SOAP" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.169, + 0.473, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.235, + 0.406, + 0.248 + ], + "angle": 0, + "content": "OccFormer [50]" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.269, + 0.314, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.32, + 0.246, + 0.331 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.267, + 0.481, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.328, + 0.447, + 0.339 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.342, + 0.285, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.408, + 0.183, + 0.418 + ], + "angle": 0, + "content": "SOAP" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.342, + 0.482, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.408, + 0.406, + 0.42 + ], + "angle": 0, + "content": "OccFormer [50]" + }, + { + "type": "image_footnote", + "bbox": [ + 0.123, + 0.424, + 0.453, + 0.435 + ], + "angle": 0, + "content": "car truck road parking sidewalk building fence" + }, + { + "type": "image_footnote", + "bbox": [ + 0.154, + 0.436, + 0.423, + 0.447 + ], + "angle": 0, + "content": "vegetation trunk terrain pole traffic sign" + }, + { + "type": "list", + "bbox": [ + 0.123, + 0.424, + 0.453, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.515 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on SemanticKITTI validation set. SOAP demonstrates robustness in handling the occluded regions (top) and distinguishing similar classes (bottom). Red boxes highlight failure cases of OccFormer [50] for comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.589 + ], + "angle": 0, + "content": "ness and efficiency of SOAP. Also, SOAP with ResNet50 yields remarkable mIoU and IoU performance with comparable complexity compared to other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Qualitative Results. Figure 5 presents qualitative comparisons between SOAP and OccFormer [50] in terms of occluded region reconstruction and similar class discrimination. In Figure 5 (top), OccFormer fails to recognize occluded regions behind car objects, resulting in fused car shapes, whereas SOAP appropriately distinguishes and separates the cars that are actually apart. Also, as in Figure 5 (bottom), Occformer confuses classes with similar appearance (road \\(\\leftrightarrow\\) sidewalk) while SOAP reasonably differentiates these classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.24, + 0.773 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We analyze the contributions of individual components of the proposed SOAP through ablation experiments on the SemanticKITTI validation set. Table 4 reports the contributions of the occluded region-aware projection (OAP) and the scene-adaptive decoder (SAD). For the results without SAD, we replace it with a standard transformer decoder [6, 50, 51]. For the results without OAP, we input the initial voxel features \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\) directly into the voxel" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.131 + ], + "angle": 0, + "content": "Table 4. Ablation study on different combinations of occluded region-aware projection (OAP) and scene-adaptive decoder (SAD)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.224 + ], + "angle": 0, + "content": "
OAPSADSC IoUSSC mIoU
43.3216.86
✓43.5117.22
✓46.5018.99
✓✓47.2419.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.238, + 0.882, + 0.251 + ], + "angle": 0, + "content": "Table 5. Alternative approaches for historical information." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.255, + 0.905, + 0.329 + ], + "angle": 0, + "content": "
MethodSC IoUSSC mIoU
Addition45.0018.05
DCA45.9018.53
OAP47.2419.21
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.905, + 0.48 + ], + "angle": 0, + "content": "encoder, without using any historical information. In Table 4, IoU and mIoU scores show significant degradation without either OAP or SAD. This indicates that OAP effectively refines occluded regions during its voxel generation, while SAD provides representative semantic features across dynamic driving scenes for query embeddings. Notably, SOAP still retains reasonable performance even without OAP, which implements 3D SSC without temporal information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.906, + 0.663 + ], + "angle": 0, + "content": "To validate the effectiveness of OAP, we design several alternative approaches to utilize historical information. First, we simply add the historical voxel features \\(\\mathbf{F}_i^{\\mathrm{hist}}\\) to the initial voxel features \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\). Second, we employ the standard deformable cross-attention (DCA) [53] between the initial voxel features \\(\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}\\) and the historical voxel features \\(\\mathbf{F}_i^{\\mathrm{hist}}\\) without occluded region separation. In Table 5, the proposed OAP achieves better IoU and mIoU scores than the standard DCA, with a margin of \\(+1.34\\) and \\(+0.68\\), respectively. Therefore, perception of occluded regions during view projection is essential for accurate 3D semantic scene reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.634, + 0.685 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We proposed a novel 3D semantic scene completion algorithm, called SOAP. The main contributions of SOAP are the occluded region-aware view projection and the scene-adaptive decoder. The occluded region-aware view projection addresses the information deficiency in vision-centric 3D SSC by refining occluded voxel regions with historical observation. The scene-adaptive decoder creates the comprehensive semantic repository using the repository builder. Then, the scene-adaptive decoder integrates informative clues from the semantic repository into query embeddings. Experiments demonstrated that SOAP significantly outperforms the state-of-the-arts on the SemanticKITTI and SSC-Bench datasets, with a margin of \\(+2\\) and \\(+2.67\\) mIoU scores, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17152" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.242 + ], + "angle": 0, + "content": "This work was partly by the National Research Foundation of Korea (NRF) grants funded by the Ministry of Education (RS-2022-NR075739) and the Korea government (MSIT) (No. RS-2024-00352566, No. RS-2025-00559165), and partly by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155857, Artificial Intelligence Convergence Innovation Human Resources Development (Chungnam National University))." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.272, + 0.188, + 0.287 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.296, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.355, + 0.484, + 0.396 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.399, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.442, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[5] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[7] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[8] Emeç Erçelik, Ekim Yurtsever, Mingyu Liu, Zhijie Yang, Hanzhen Zhang, Pinar Topçam, Maximilian Listl, Yilmaz Kaan Cayli, and Alois Knoll. 3d object detection with a self-supervised lidar scene flow backbone. In ECCV, pages 247-265, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.788, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[11] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, pages 4012-4021, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.296, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[13] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In CVPR, pages 20258-20267, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[14] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, pages 515–532, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.906, + 0.314 + ], + "angle": 0, + "content": "[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[16] Abhijit Kundu, Yin Li, and James M Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, pages 3559–3568, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.425 + ], + "angle": 0, + "content": "[17] Hyo-Jun Lee, Hanul Kim, Su-Min Choi, Seong-Gyun Jeong, and Yeong Jun Koh. Baam: Monocular 3d pose and shape reconstruction with bi-contextual attention module and attention-guided modeling. In CVPR, pages 9011-9020, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.906, + 0.483 + ], + "angle": 0, + "content": "[18] Bohan Li, Jiajun Deng, Wenyao Zhang, Zhujin Liang, Dalong Du, Xin Jin, and Wenjun Zeng. Hierarchical temporal context learning for camera-based semantic scene completion. In ECCV, pages 131-148. Springer, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[19] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[20] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In AAAI, pages 1477-1485, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[22] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, pages 9087-9098, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Jongwon Ra, Yukai Ma, Laijian Li, and Yong Liu. Camera-based" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17153" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "3d semantic scene completion with sparse guidance network. IEEE TIP, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[27] Chao Ning and Hongping Gan. Trap attention: Monocular depth estimation with manual traps. In CVPR, pages 5033-5043, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[28] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. idisc: Internal discretization for monocular depth estimation. In CVPR, pages 21477-21487, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[30] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. IEEE TPAMI, 44(10): 7205-7218, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[31] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.345, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[32] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 2417-2426, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[33] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[34] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[35] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[36] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Not all voxels are equal: Semantic scene completion from the point-voxel perspective. In AAAI, pages 2352–2360, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[37] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[38] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[39] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In CVPR, pages 17642-17651, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[40] Jinfeng Xu, Xianzhi Li, Yuan Tang, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Casfusionnet: A cascaded network for point cloud semantic scene completion by dense feature fusion. In AAAI, pages 3018-3026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[42] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9421-9431, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[43] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[44] Zhu Yu, Running Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Siyuan Cao, and Huiliang Shen. Context and geometry aware voxel transformer for semantic scene completion. In NeurIPS, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[45] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3916-3925, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In ICCV, pages 9155-9166, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[47] Shoulong Zhang, Shuai Li, Aimin Hao, and Hong Qin. Point cloud semantic scene completion from rgb-d images. In AAAI, pages 3385-3393, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[48] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, pages 1589-1598, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[49] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[50] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[51] Linqing Zhao, Xiuwei Xu, Ziwei Wang, Yunpeng Zhang, Borui Zhang, Wenzhao Zheng, Dalong Du, Jie Zhou, and Jiwen Lu. Lowrankocc: Tensor decomposition and low-rank recovery for vision-based 3d semantic occupancy prediction. In CVPR, pages 9806-9815, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[52] Yunsong Zhou, Hongzi Zhu, Quan Liu, Shan Chang, and Minyi Guo. Monoatt: Online monocular 3d object detection with adaptive token transformer. In CVPR, pages 17493-17503, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[54] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17154" + } + ] +] \ No newline at end of file diff --git a/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_origin.pdf b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e54e04f656dc109beb895cc37b39846658aa271b --- /dev/null +++ b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/399b8632-9ed1-4ab5-bef3-002cbfb4c36d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee6924be86d5c60fb4ccfdc3664aa9c120e8ad27071a5a09b3eeb59919991f1a +size 9272590 diff --git a/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/full.md b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..690cd07eb46679038e38b115e298aec82c06dc76 --- /dev/null +++ b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/full.md @@ -0,0 +1,327 @@ +# SOAP: Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection + +Hyo-Jun Lee + +Hyunseop Kim + +Yeong Jun $\mathrm{Koh}^{2*}$ + +Yonguk Lee + +Hanul Kim3 + +Jinu Lee + +$^{1}$ Kangwon National University + +$^{2}$ Chungnam National University + +$^{3}$ Seoul National University of Science and Technology + +42dot Inc. + +gywns6298@kangwon.ac.kr + +yjkoh@cnu.ac.kr + +hukim@seoultech.ac.kr + +hyunseop95@gmail.com + +yonguk.lee@42dot.ai + +jinu.lee@42dot.ai + +# Abstract + +Existing view transformations in vision-centric 3D Semantic Scene Completion (SSC) inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, leading to a dilution of informative contexts. Furthermore, semantic classes exhibit high variability in their appearance in real-world driving scenarios. To address these issues, we introduce a novel 3D SSC method, called SOAP, including two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection effectively converts 2D image features into voxel space, refining the duplicated features of occluded regions using information gathered from previous observations. The scene-adaptive decoder guides query embeddings to learn diverse driving environments based on a comprehensive semantic repository. Extensive experiments validate that the proposed SOAP significantly outperforms existing methods for the vision-centric 3D SSC on automated driving datasets, SemanticKITTI and SSCBench. Code is available at https://github.com/gywns6287/SOAP. + +# 1. Introduction + +Understanding 3D traffic scenes is crucial for autonomous driving systems to develop downstream tasks such as motion prediction and planning. LiDAR sensors [8, 33, 43, 54] provide precise 3D scene information, but their high cost and the complex setup limit their scalability. In contrast, areas of study about vision-centric 3D perception, including 3D object detection [11, 25, 46, 48, 52], 3D car shape reconstruction [14, 16, 17], and depth estimation [2, 27, 29, 45], have been receiving attention due to its simplicity and cost-effectiveness. However, these methods still struggle to re + +![](images/5a000ecd8d9ad8d5aa724a75275853163d3a45af72c5e0209c16329be42543dd.jpg) +(a) Occluded voxel regions. + +![](images/f8e435855253d1c9952d8a9efcc9689e8fffb2295847165759e72d8f5a237b5f.jpg) +(b) Diversity in the appearance of semantic classes. +Figure 1. Two main motivations of SOAP. (a) The 2D-to-3D view projection inevitably yields erroneous feature duplication in voxel generation due to occlusion. (b) The semantic classes exhibit significant variability in their appearance depending on driving regions. + +construct complete and fine-grained 3D scene representations. To address this limitation, 3D Semantic Scene Completion (SSC) has emerged as a promising alternative. 3D SSC leverages a voxel-based approach, assigning both an occupancy probability and a semantic label (e.g. car, pedestrian, road) to each voxel within the 3D space. Thereby, it can create a comprehensive representation of the driving scene. While earlier works [5, 19, 39, 41] primarily rely on LiDAR sensors, recent research [3, 13, 18, 22, 37, 50, 51] has demonstrated the potential of vision-centric methods. + +A general vision-based 3D SSC model comprises an image encoder, a view projection module, and a voxel decoder. The image encoder first extracts 2D features from an input image, and then the view projection module subsequently lifts 2D image features into 3D voxel space. Given the 3D features of each voxel region, the voxel decoder estimates semantic classes corresponding to each voxel region. In this process, the precise 2D-3D view projection plays a crucial role. However, existing view projection methods [3, 50] inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, particularly along the longitudinal direction. For instance, as illustrated in Figure 1a, when a camera ray encounters an occupancy region (e.g. car), occluded regions (voxels behind this occupancy) may contain duplicated features of the obstructing car in the 3D view projection process. This duplication can occur even when the actual classes of the occluded regions are not related to a car, resulting in incorrect 3D SSC results. + +Furthermore, the diverse appearances of semantic classes pose another significant challenge in 3D SSC. As shown in Figure 1b, semantic classes exhibit high variability in their appearance in real-world driving scenarios. For instance, road color and texture may vary depending on driving regions (e.g. urban or highway). Such appearance variations make it difficult for SSC models to distinguish similar classes, such as road $\leftrightarrow$ sidewalk, leading to potential misinterpretations of driving environments. While most previous works [3, 13, 22, 37] usually employ sequential 3D convolutions for the semantic decoder, this approach often lacks the flexibility to generalize for the diverse semantic context. More recently, Mask2former [6]-like approach [50, 51] is adopted for the SSC decoder, which utilizes semantic tokens derived from a query embedding to represent the semantic scene. However, these semantic tokens lack adaptability to dynamic driving environments, as they are fundamentally based on static query embedding. + +In this work, we propose a novel 3D semantic scene completion network, SOAP, faithfully designed to address the occluded voxel regions and the visual variations in semantic classes. The proposed method leverages two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection (OAP) first converts 2D image features into 3D voxel features using the LSS algorithm [28], where these initial voxel features contain false duplicated information due to occlusions, as illustrated in Figure 1a. These occluded regions can be inferred by referring to previous frames, which provides valuable context about the missing observation. Therefore, OAP identifies occluded regions in the current frame using depth estimation and refines the duplicated features of occluded regions using information gathered from previous observations. Also, a vanilla transformer decoder [6] generates the semantic tokens based on + +static query embedding, which is not versatile enough to capture a dynamic driving environment. To represent more diverse semantics, a repository builder first creates a semantic repository by gathering semantic tokens from various scenes. Then, the scene-adaptive decoder integrates informative clues from this extensive repository into the semantic tokens, guiding them to navigate diverse driving environments effectively. + +Extensive experiments demonstrate that the proposed SOAP outperforms the other state-of-the-arts on two vision-centric SSC benchmarks, SemanticKITTI [1] and SSCBench-KITTI360 [21]. To summarize, the main contributions of our work are three folds: + +- We develop the occluded region-aware view projection that effectively transforms 2D image features into 3D voxel space by recovering occluded voxel regions. +- We also introduce a scene-adaptive transformer that integrates informative semantic clues from the various scenes into the semantic tokens of an individual scene. +- The proposed SOAP achieves outstanding performances on two driving datasets, SemanticKITTI and SSCBenchKITTI360, with a margin of $+2$ and $+2.67$ mIoU scores, respectively. + +# 2. Related Works + +3D Semantic Scene Completion. 3D semantic scene completion (SSC) is an advanced task in 3D scene reconstruction fields, which aims to predict a dense semantic scene representation involving the geometrical occupancy and its corresponding semantic label. Early SSC methods [5, 19, 34] leverage 3D inputs to reconstruct voxelized indoor scenes. SSCNet [34] first defines the SSC task, which jointly reasons the geometry and semantics of an indoor scene from a single-depth image. 3DSketch [5] uses the Truncated Signed Distance Function to guide an occupancy and the semantic labels of the indoor scene. AIC-Net [19] develops anisotropic convolution to extract hybrid features from the depth and RGB images for indoor SSC tasks. Recently, the release of SemanticKITTI [1], the first large-scale outdoor SSC dataset, yields the development of diverse outdoor SSC methods. LMSCNet [31] uses the 2D UNet to extract collapsed BEV features from sparse 3D Li-DAR scans and finally expands the height dimension for the 3D SSC. JS3C-Net [41] learns the contextual shape priors from sequential LiDAR data completing the sparse single sweep point cloud to the dense one for enhanced 3D SSC. The other existing works [7, 30, 36, 39, 40, 47] also commonly depend on 3D input such as LiDAR point clouds. In contrast, we address vision-centric 3D SSC, which only uses an RGB image for the semantic scene reconstruction. + +Vision-centric 3D SSC. MonoScene [3] introduces the first vision-centric approach for SSC, utilizing a 3D UNet to process voxel features generated by sight projection. Vox- + +![](images/a5c87da77491001fd7b2aabc086de697c19964d59e2999dca521239f6247600a.jpg) +Figure 2. Overview of the proposed SOAP. Given multi-scale image features, an occluded region-aware view projection (OAP) transforms them into voxel space, refining duplicated features in occluded voxel regions using historical information. A scene-adaptive decoder enriches query embeddings with diverse semantic contexts through a semantic repository, which is constructed and iteratively updated by a repository builder. + +Former [22] estimates occupancy regions using stereo depth estimation and selectively updates these regions with RGB images. HASSC [37] and Symphonies [13] also identify voxel occupancy based on stereo depth estimation, following VoxFormer. NDC-scene [42] extends a 2D feature map to a normalized device coordinates space to restore the depth dimension and subsequently reconstructs a 3D semantic scene using this space. SurroundOcc [38] adopts spatial 2D-3D attention to lift multi-scale 2D features into the 3D volume space. TPVFormer [12] introduces a query-based tri-perspective space for 3D SSC. Occformer [50] and LowRankOcc [51] reconstruct 3D scenes using semantic tokens on the Mask2former decoders [6]. SGN [26] propagates semantic contexts from semantic-aware seed voxels based on their spatial geometry cues. HTCL [18] leverages historical frames to refine current frame features through affinity-based dynamic refinement. + +# 3. Method + +# 3.1. Model Overview + +3D Semantic Scene Completion (SSC) aims to predict semantic probabilities of the voxelized scene $\mathbf{V} \in \mathbb{R}^{X \times Y \times Z \times S}$ , where $(X, Y, Z)$ denotes the resolution of the 3D volume, and $S$ is the number of semantic classes including a free (or empty) class. Figure 2 illustrates the overall architecture of the proposed SOAP. Image encoder first ex + +tracts multi-scale image features $\{\mathbf{F}_i^{2\mathrm{D}} \in \mathbb{R}^{H_i \times W_i \times C}\}_{i=1}^N$ from an RGB image, where $N$ is the number of scales, $(H_i, W_i)$ is the spatial size at the scale $i$ , and $C$ is the channel dimension. Depth estimator predicts the depth map $\mathbf{E} \in \mathbb{R}^{H \times W}$ from RGB images. Consistent with previous approaches [13, 18, 22, 37, 44], we also use off-the shelf stereo depth model [32] for the depth estimator. + +The occluded region-aware view projection (Section 3.2) utilizes the estimated depth map $\mathbf{E}$ to lift multi-scale image features $\{\mathbf{F}_i^{2\mathrm{D}}\}_{i = 1}^N$ into multi-scale voxel features $\{\mathbf{F}_i^{3\mathrm{D}}\in \mathbb{R}^{X_i\times Y_i\times Z_i\times C}\}_{i = 1}^N$ , enhancing occluded voxel regions based on informative features in historical frames. The scene-adaptive decoder (Section 3.3) first constructs a semantic repository using a repository builder and then it integrates semantic diversity from a semantic repository into query embeddings to produce scene-adaptive semantic tokens, which are utilized to reconstruct the final semantic voxel representation $\mathbf{V}$ , in conjunction with multiscale voxel features. Through this complementary process, SOAP guarantees a more flexible and accurate 3D SSC for a real-world driving environment. + +# 3.2. Occluded Region-Aware View Projection + +Occluded region-aware view projection (OAP) transforms 2D image features $\{\mathbf{F}_i^{2\mathrm{D}}\}_{i = 1}^N$ into 3D voxel space $\{\mathbf{F}_i^{3\mathrm{D}}\}_{i = 1}^N$ . Figure 3 illustrates the detailed process of OAP. It produces voxel features for each $i$ -th scale solely from the image feature at $i$ -th scale. + +![](images/c035787505e4718e3d4720049c84aded0b31c19efa75517a8d4e6c4c0f192e8f.jpg) +Figure 3. Occluded region-aware view projection. (a) occluded region identification identifies occluded regions in the current frame by inferring an occlusion map based on the depth distribution. (b) occluded region refinement refines the duplicated features in the identified occluded regions using previous observations within the historical voxel features. + +![](images/12214c58a5d4ec297e5914d097ccdba18ee4ef28fbec11dc72cf99b58d551342.jpg) +Figure 4. t-SNE visualization of voxel features in the occluded regions for (a) $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ and (b) $\tilde{\mathbf{F}}_i^{\mathrm{occ}}$ . Different colors represent different camera rays. + +Initial Voxel Generation. For each $i$ th scale, OAP first generates initial voxel features $\tilde{\mathbf{F}}_i^{3\mathrm{D}} \in \mathbb{R}^{X_i \times Y_i \times Z_i \times C}$ from the 2D image features $\mathbf{F}_i^{2\mathrm{D}}$ and the depth estimation $\mathbf{E}$ . Specifically, OAP resizes the estimated depth $\mathbf{E}$ to $\mathbf{E}_i$ that has the spatial resolution of $(H_i, W_i)$ , and distributes it to the depth distribution map $\mathbf{D}_i \in \mathbb{R}^{H_i \times W_i \times D}$ . Here, the continuous depth value $\mathbf{E}_i(h, w)$ for each pixel $(h, w)$ is converted into discrete depth bins $\mathbf{D}_i(h, w) \in \mathbb{R}^D$ represented as one-hot encoded vectors, followed by a Gaussian filter to create a soft depth distribution. Then, OAP computes the outer product $\mathbf{F}_i^{2\mathrm{D}} \otimes \mathbf{D}_i$ to create the depth distributed features $\mathbf{F}_i^{\mathrm{depth}} \in \mathbb{R}^{H_i \times W_i \times D \times C}$ . The voxel- + +pooling operation $\varphi(\cdot)$ in [28] can project $\mathbf{F}_i^{\mathrm{depth}}$ into the voxel space using camera intrinsics and extrinsics. Thus, the initial voxel features are defined as $\tilde{\mathbf{F}}_i^{3\mathrm{D}} = \varphi(\mathbf{F}_i^{\mathrm{depth}})$ . Historical Voxel Generation. To represent previous observations, we incorporate depth distributed features $\{\mathbf{F}_{i,t - p}^{\mathrm{depth}}\}_{p = 1}^{P}$ from the previous frames into the current voxel space $\mathbf{F}_i^{\mathrm{hist}} \in \mathbb{R}^{X_i \times Y_i \times Z_i \times C}$ , where $P$ is the number of previous frames at timestamp $t$ . Specifically, each past $\mathbf{F}_{i,t - p}^{\mathrm{depth}}$ is projected into voxel space $\varphi(\mathbf{F}_{i,t - p}^{\mathrm{depth}})$ and then aligned to the current voxel space at $t$ via the warping function $\mathcal{W}(\cdot)$ and summation operation: + +$$ +\mathbf {F} _ {i} ^ {\mathrm {h i s t}} = \sum_ {p = 1} ^ {P} \mathcal {W} \left(\varphi \left(\mathbf {F} _ {i, t - p} ^ {\mathrm {d e p t h}}\right), \mathcal {G} _ {t} ^ {- 1} \cdot \mathcal {G} _ {t - p}\right) \tag {1} +$$ + +where $\mathcal{G}_t$ and $\mathcal{G}_{t - p}$ are ego-posed matrices at $t$ and $t - p$ . These historical voxel features $\mathbf{F}_i^{\mathrm{hist}}$ provide valuable perspectives that are not available in the initial voxel features. Occluded Region Identification. The initial voxel features $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ , which are projected from a single image view, inherently include false duplicated features due to occlusions. Figure 4a presents the t-SNE visualization of features in the occluded regions within $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ . Each point represents an individual voxel region, and voxel regions of the same color correspond to those located along the same camera ray. We observe that most occluded regions along the same camera ray exhibit similar patterns, which indicates erroneous replication of the features from the occupancy regions positioned in front of the occluded regions as in Figure 1a. + +To address this limitation, OAP refines these occluded features by referring to previous frames, which provide valuable contexts about the missing observation in the current frame. As part of this process, we identify the occluded regions using the estimated depth distribution $\mathbf{D}_i$ . + +As in Figure 3a, each vector $\mathbf{D}_i(h,w)$ at the spatial location $(h,w)$ contains the probability distribution of the nearest occupancy region along the longitudinal (or depth) direction. In other words, regions located behind the nearest occupancy with the highest depth probability often lose their semantic context information due to occlusions. To distinguish the occluded regions from non-occluded regions, we define an occlusion map $\mathbf{O}_i\in \{1,0\}^{H_i\times W_i\times D}$ as + +$$ +\mathbf {O} _ {i} (h, w, d) = \left\{ \begin{array}{l l} 1 & \text {i f} d > \underset {\tilde {d} \in \{1, \dots , D \}} {\arg \max } \mathbf {D} _ {i} (h, w, \tilde {d}) \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +Figure 3a provides an example of $\mathbf{O}_i$ , where occluded regions are marked with 1 (taller rectangular) and non-occluded regions are marked with 0 (shorter rectangular). To identify occluded regions in the voxel space, the occlusion map $\mathbf{O}_i$ is projected into voxel space through the voxel pooling operation, i.e. $\varphi (\mathbf{O}_i)\in \mathbb{R}^{X_i\times Y_i\times Z_i}$ . Note that the regions identified by this occlusion map $\varphi (\mathbf{O}_i)$ contain incorrectly duplicated information from the nearest occupancy region along the longitudinal direction, as illustrated in Figure 4a. Therefore, we refine the features of these occluded regions through an occluded region refinement process. + +Occluded Region Refinement. The occluded regions at the current time step may have been observable in previous time steps. Therefore, we refine deficient features in the occluded regions with previous observations in the historical voxel features $\mathbf{F}_i^{\mathrm{hist}}$ . To achieve this, we perform 3D deformable cross-attention, DCA [53] between the historical voxel features $\mathbf{F}_i^{\mathrm{hist}}$ and the occluded voxel features, resulting in reconstructed occluded voxel features $\mathbf{F}_i^{\mathrm{occ}} \in \mathbb{R}^{X_i \times Y_i \times Z_i \times C}$ : + +$$ +\mathbf {F} _ {i} ^ {\text {o c c}} (x, y, z) = \left\{ \begin{array}{l l} \tilde {\mathbf {F}} _ {i} ^ {\mathrm {3 D}} (x, y, z) + \mathrm {D C A} \left(\tilde {\mathbf {F}} _ {i} ^ {\mathrm {3 D}} (x, y, z), \mathbf {F} _ {i} ^ {\text {h i s t}}\right), \\ \quad \text {i f} \varphi (\mathbf {O} _ {i}) (x, y, z) \geq 1 \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +where $(x,y,z)$ are spatial indices of the voxel space. This approach ensures that the reconstructed features $\mathbf{F}_i^{\mathrm{occ}}$ effectively capture observable historical information about occluded regions, enabling more accurate and context-aware scene representation. As shown in Figure 4b, the reconstructed features $\mathbf{F}_i^{\mathrm{occ}}$ include appropriately corrected information for the occluded regions, resulting in more diverse patterns according to different semantic regions. + +Invisible and Visible Region Reconstruction. The initial voxel features also contain invisible regions due to the limited observation of a single camera. Since these invisible voxels do not contain any features, we directly fill them with the corresponding historical voxel features to recon- + +struct invisible voxel features $\mathbf{F}_i^{\mathrm{inv}}\in \mathbb{R}^{X_i\times Y_i\times Z_i\times C}$ .. + +$$ +\mathbf {F} _ {i} ^ {\text {i n v}} (x, y, z) = \left\{ \begin{array}{l l} \mathbf {F} _ {i} ^ {\text {h i s t}} (x, y, z) & \text {i f} \sum \tilde {\mathbf {F}} _ {i} ^ {\text {3 D}} (x, y, z) = 0 \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +We designate the voxel space, excluding invisible and occluded regions, as visible regions. The visible regions already contain appropriate semantic features with respect to the current observation, allowing us to directly retrieve the visible voxel features from the initial voxel features $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ . The visible voxel features $\mathbf{F}_i^{\mathrm{vis}} \in \mathbb{R}^{X_i \times Y_i \times Z_i \times C}$ are defined as + +$$ +\mathbf {F} _ {i} ^ {\mathrm {v i s}} (x, y, z) = \left\{ \begin{array}{l l} \tilde {\mathbf {F}} _ {i} ^ {\mathrm {3 D}} (x, y, z) & \text {i f} \varphi (\mathbf {O} _ {i}) (x, y, z) = 0 \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {5} +$$ + +Voxel Encoder. To compute the final 3D voxel features, we sum the occluded, invisible, and visible voxel features, $\mathbf{F}_i^{\mathrm{occ}} + \mathbf{F}_i^{\mathrm{inv}} + \mathbf{F}_i^{\mathrm{vis}}$ , and process them through a voxel encoder to obtain the final voxel features $\mathbf{F}_i^{3\mathrm{D}} \in \mathbb{R}^{X_i \times Y_i \times Z_i}$ . In the voxel encoder, the spatial dimensions $(X_i \times Y_i \times Z_i)$ of the voxel features are sliced into $Z_i$ bird's-eye view (BEV) planes with spatial resolution $(X_i \times Y_i)$ . Then, two windowed self-attention blocks [23] are sequentially performed to each BEV plane. + +# 3.3. Scene-Adaptive Decoder + +A standard transformer decoder [6, 50, 51] yields semantic tokens starting from static query embeddings, which lack adaptability for dynamic driving scenes. To represent more diverse semantics, the scene-adaptive decoder correlates semantic tokens for a specific scene with tokens from different scenes. As in Figure 2, the scene-adaptive decoder generates scene-adaptive tokens $\mathbf{Q} \in \mathbb{R}^{N_{\mathrm{q}} \times C}$ through $L$ iterative transformer blocks, where $N_{\mathrm{q}}$ is the number of tokens. Each transformer block in the scene transformer cascades four modules including a cross-attention between query embeddings and the semantic repository, the other cross-attention between the query embeddings and the voxel features $\mathbf{F}_i^{3\mathrm{D}}$ , self-attention, and feed-forward networks. The cross-attention with the semantic repository integrates semantic features from dynamic scenes into the query embeddings, while the cross-attention with voxel features provides scene-specific semantic features. After the end of iterations, scene-adaptive tokens $\mathbf{Q}$ are projected to class predictions $\mathbf{S}^{\mathrm{cls}} \in \mathbb{R}^{N_{\mathrm{q}} \times S}$ and segmentation mask predictions $\mathbf{M} \in \mathbb{R}^{X \times Y \times Z \times N_{\mathrm{q}}}$ . Then, the final semantic voxel $\mathbf{V} \in \mathbb{R}^{X \times Y \times Z \times S}$ is formulated as + +$$ +\mathbf {V} = \mathcal {F} ^ {- 1} (\mathcal {F} (\mathbf {M}) \mathbf {S} ^ {\mathrm {c l s}}) \tag {6} +$$ + +where $\mathcal{F}:\mathbb{R}^{X\times Y\times Z\times C}\to \mathbb{R}^{XYZ\times C}$ is as spatial flatten and $\mathcal{F}^{-1}$ is its inverse operation. + +Table 1. Semantic scene completion results on SemanticKITTI test set. † represents temporal methods that leverage previous frames for 3D SSC. We highlight the best results in bold and the second best results in underline. + +
MethodSC IoUSSC mIoUroad (15.30%)sidewalk (11.13%)parking (1.12%)other-ground (0.56%)building (14.1%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation(39.3%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (3.90%)pole (0.29%)traf-sign (0.08%)
MonoScene [3]34.1611.0854.727.124.85.714.418.83.30.50.74.414.92.419.51.01.40.411.13.32.1
TPVFormer [12]34.2511.2655.127.227.46.514.819.23.71.00.52.313.92.620.41.12.40.311.02.91.5
SurroundOcc [38]34.7211.8656.928.330.26.815.220.61.41.61.24.414.93.419.31.42.00.111.33.92.4
VoxFormer† [22]43.2113.4154.126.925.17.323.521.73.61.91.64.124.48.124.21.61.10.013.16.65.7
OccFormer [50]34.5312.3255.930.331.56.515.721.61.21.51.73.216.83.921.32.21.10.211.93.83.7
NDC-Scene [42]33.8711.5556.228.728.05.615.819.71.81.11.14.914.32.620.60.71.70.411.23.21.7
LowRankOcc [51]38.4713.5652.827.225.18.822.120.92.93.32.74.422.98.920.82.41.72.314.47.07.0
Symphonies [13]42.1915.0458.429.326.911.724.723.63.23.62.65.624.210.023.13.21.92.016.17.78.0
HASSC† [37]42.8714.3855.329.625.911.323.123.02.91.91.54.924.89.826.51.43.00.014.37.07.1
SGN† [26]45.4215.7660.431.428.98.728.425.44.50.91.63.727.412.628.40.50.30.118.110.08.3
HTCL† [18]44.2317.0964.434.833.812.425.927.35.71.82.25.425.310.831.21.13.10.921.19.08.3
\(SOAP^{\dagger}\) (proposed)46.0919.0963.636.236.817.228.728.93.35.05.37.029.815.132.33.72.30.222.511.713.1
+ +Repository Builder. The semantic repository $\mathbf{R} \in \mathbb{R}^{(S \times K) \times C}$ stores $K$ representative tokens for each semantic class, iteratively updated by semantic tokens with top- $K$ highest scores. The detailed process of the repository builder is shown in Figure 2. Even though representative tokens could be selected using the classification scores $\mathbf{S}^{\mathrm{cls}}$ of the scene-adaptive tokens $\mathbf{Q}$ , relying solely on $\mathbf{S}^{\mathrm{cls}}$ makes a repository filled with similar token and consequently degrades semantic diversity. To retain the semantic diversity within the repository $\mathbf{R}$ , we introduce a novelty-based token selection. Specifically, we re-define token scores $\mathbf{S} \in \mathbb{R}^{N_{\mathrm{q}} \times S}$ as $\mathbf{S} = \mathbf{S}^{\mathrm{cls}} + \alpha \mathbf{S}^{\mathrm{nov}}$ , where $\mathbf{S}^{\mathrm{nov}} \in \mathbb{R}^{N_{\mathrm{q}} \times S}$ is novelty scores and $\alpha$ is its balancing parameter. To compute the novelty scores, the $K$ tokens within each class in $\mathbf{R}$ are first averaged, resulting in the class average repository $\bar{\mathbf{R}} \in \mathbb{R}^{S \times C}$ . Then, the novelty scores $\mathbf{S}^{\mathrm{nov}}$ are calculated based on cosine similarity between each token in $\mathbf{Q}$ and $\bar{\mathbf{R}}$ : + +$$ +\mathbf {S} ^ {\text {n o v}} (n, s) = - \cos (\mathbf {Q} (n), \bar {\mathbf {R}} (s)) \tag {7} +$$ + +where $n$ and $s$ indicate token indices of $\mathbf{Q}$ and $\bar{\mathbf{R}}$ , respectively, and $\cos(\cdot, \cdot)$ is a cosine similarity between two input vectors. + +Then, new tokens in $\mathbf{Q}$ with higher scores than old tokens in $\mathbf{R}$ are selected and, newly pushed into the semantic repository $\mathbf{R}$ . When a class holds more than $K$ tokens, the old tokens with lower scores are discarded to maintain $K$ representative tokens for each semantic class in $\mathbf{R}$ . Updating $\mathbf{R}$ with tokens that exhibit both high classification and novelty scores enables the repository to capture rich semantic information, effectively covering dynamic driving + +scenes. Note that the semantic repository $\mathbf{R}$ is updated only during the SOAP's training phase to ensure stable inference performance. + +# 3.4. Loss Functions + +We perform the Hungarian matching [15] between predicted and ground-truth voxel masks following [4, 6, 53]. For each matched pair, we compute a mask loss $L_{\mathrm{mask}}$ , a mask dice loss $L_{\mathrm{dice}}$ , and a class loss $L_{\mathrm{cls}}$ as in [6]. The final loss is defined as + +$$ +L = \lambda_ {\text {m a s k}} L _ {\text {m a s k}} + \lambda_ {\text {d i c e}} L _ {\text {d i c e}} + \lambda_ {\text {c l s}} L _ {\text {c l s}} \tag {8} +$$ + +where, scaling weights $\lambda_{\mathrm{mask}}, \lambda_{\mathrm{dice}}, \lambda_{\mathrm{cls}}$ are set to 5, 5, 2, respectively. + +# 4. Experiments + +# 4.1. Datasets + +We evaluate the performance of SOAP on two outdoor 3D SSC datasets: SemanticKITTI [1] and SSCBenchKITTI360 [21]. Both datasets provide front-view RGB images, LiDAR sweeps, and semantic scene labels represented as $256 \times 256 \times 32$ voxel grids. The scene labels cover a range of $51.2m \times 51.2m \times 6.4m$ from the front view of the vehicle. + +SemanticKITTI [1]. It contains a total of 22 sequences, which are split into $10/1/11$ for train/val/test following [3]. We utilize the RGB images of cam-2 with a $1280 \times 384$ resolution as model inputs. In the semantic labels, each voxel + +Table 2. Semantic scene completion results on the SSCBench test set. † represents methods that use temporal information. We highlight the best results in **bold** and the second best results in **underline**. + +
MethodSC IoUSSC mIoUcarbicyclemotorcycletruckother-veh.personroadparkingsidewalkother-groundbuildingfencevegetationterrainpoletraf.-signother-struct.other-object
MonoScene [3]37.8712.3119.30.40.68.02.00.948.411.428.13.232.93.526.216.86.95.74.23.1
TPVFormer [12]40.2213.6421.61.11.48.12.62.453.012.031.13.834.84.830.117.57.55.95.52.7
OccFormer [50]40.2713.8122.60.70.39.93.82.854.313.431.53.636.44.831.019.57.88.57.04.6
VoxFormer†[22]38.7611.9117.81.20.94.62.11.647.09.727.22.931.25.029.014.76.56.93.82.4
Symphonies [13]44.1218.5830.01.95.925.112.18.254.913.832.86.935.18.638.311.514.09.614.411.3
SGN†[26]47.0618.2529.03.42.910.95.23.058.115.036.44.442.07.738.223.216.716.49.95.8
\( \mathrm{SOAP}^{\dagger } \) (proposed)48.1220.9229.95.67.814.47.66.160.917.440.35.445.310.640.524.816.821.012.69.9
+ +is annotated with 21 semantic classes (19 semantics, 1 free, 1 unknown). + +SSCBench-KITTI360 [21]. It is a recently introduced dataset for 3D SSC task. SSCBench-KITTI360 contains 9 long sequences, which are split into 7/1/1 for train/val/test. The input resolution of the RGB images is set to $1408 \times 384$ . The scene voxels in SSCBench-KITTI360 are annotated with 20 semantic classes (18 semantics, 1 free, 1 unknown). + +# 4.2. Implementation Details + +Network Structure. Following other comparable methods [3, 12, 44, 50, 51], we employ EfficientNetB7 as our backbone. The multi-scale number $N$ is 4 and the spatial dimensions of the multi-scale image features are $\left(\frac{1}{4}, \frac{1}{8}, \frac{1}{16}, \frac{1}{32}\right)$ of an input image, respectively. We set a spatial voxel size $X \times Y \times Z$ to $128 \times 128 \times 16$ and channel numbers $C$ to 192. The spatial dimensions of the multiscale voxel features are $(1, \frac{1}{2}, \frac{1}{4}, \frac{1}{8})$ of $X \times Y \times Z$ , respectively. We set the number of semantic tokens to $N_{\mathrm{q}} = 100$ while the depth distribution bin to $D = 112$ . In the repository builder, the balancing parameter for novelty scores is set to $\alpha = 0.2$ and the number of representative tokens of the semantic repository $K$ is set to 10. We use $P = 4$ historical frames. The scene-adaptive decoder is repeated $L = 9$ times. The output voxel is upsampled to $256 \times 256 \times 32$ through the trilinear interpolation for the evaluation. + +Training. We train the proposed SOAP for 30 epochs. We employ AdamW [24] optimizer with an initial learning rate 0.0001 and divide it by 10 at the 20 and 25 epochs. For the training, we use a mini-batch of size 4 on 4 RTX A6000 GPUs. For data augmentation, we use random resize, rotation, and flip for the image space and 3D flip for the 3D volume space following [10, 20, 49]. + +# 4.3. Main Results + +For evaluation on SemanticKITTI and SSCBench, we use the mean intersection over union (mIoU) of semantic + +Table 3. Complexity-accuracy comparison on SemanticKITTI validation set. † represents methods that use temporal information. + +
MethodImage encoderParams.GFLoPsIoUmIoU
OccFormer [50]EfficientNetB7202.7M973.136.5013.46
\(HTCL^†\) [18]EfficientNetB7181.4M2622.445.5117.13
\(SOAP^†\)EfficientNetB785.2M705.347.2419.21
Symphonies [13]ResNet5059.3M611.941.9214.89
\(HASSC^†\) [37]ResNet5058.1M313.944.5814.74
\(SGN^†\) [26]ResNet5027.9M723.746.2115.32
\(SOAP^†\)ResNet5045.1M627.848.1218.80
+ +classes for the SSC task and the intersection over union (IoU) for the class-agnostic scene completion (SC) task. + +Highest mIoU and IoU Scores. Tables 1 and 2 compare the proposed SOAP with recent vision-centric 3D SSC algorithms on the SemanticKITTI and SSCBench-KITTI360 test sets, respectively. The scores of other methods come from respective papers. In Table 1, we observe that SOAP significantly outperforms the existing methods on both datasets in terms of the mIoU and IoU scores. Compared to HTCL [18], SOAP improves $+2$ SSC mIoU and $+1.86$ SC IoU on SemanticKITTI test set, which achieves $12\%$ and $4\%$ relative improvement. Also, SOAP surpasses SGN [26] on SSCBench-KITTI test set with a margin of $+2.67$ mIoU and $+1.06$ IoU. The highest mIoU and IoU scores demonstrate that the proposed OAP faithfully refines occluded regions in the initial voxel features, while the scene-adaptive decoder further enhances the semantic context of the query embeddings. + +Efficiency Evaluation. We replace our image encoder, EfficientNetB7 [35] with a more light-weight network, ResNet50 [9]. Table 3 shows complexity-accuracy comparisons between SOAP and other methods on SemanticKITTI validation set, using either ResNet50 or EfficientNetB7 as the image encoder. Compared to OccFormer [50] and HTCL [18], which utilize EfficientNetB7, SOAP requires the lowest parameters and GFLoPs while achieving the highest IoU and mIoU scores, demonstrating the effective + +![](images/7a6c310b69ba4b07afd41a78d8853b315a4fe221721e9b9e250d0751afde0aa7.jpg) +Input image + +![](images/a2e4947b1cc23e631baa4df875c9a47c6d4343032fb0ce69000cabc4a2a5c9d1.jpg) + +![](images/a9af1d35189661f00a4ad07c3d0207d1cddc88b2d771213721ecfd10ba461bf8.jpg) +SOAP + +![](images/b7f0a041533b72f20816e7c8c70b74d07933bdf3ffde4491fd20751154bd6888.jpg) +Ground-truth +OccFormer [50] + +![](images/672e54184cef91cf9038fa9238727e3050da32c3d2dfd2391e11477d83124908.jpg) +Input image + +![](images/da67e55e10c275427580c393eb7010a343cb0870758cb80232323493a27fe7c3.jpg) + +SOAP +Figure 5. Qualitative results on SemanticKITTI validation set. SOAP demonstrates robustness in handling the occluded regions (top) and distinguishing similar classes (bottom). Red boxes highlight failure cases of OccFormer [50] for comparison. +![](images/ac4d0833fd9aa0db7634d5d704aa87f7780629619e6d10af30e3bbb0cf1df382.jpg) +car truck road parking sidewalk building fence + +Ground-truth +OccFormer [50] +![](images/2afc38abe44aeeaf7f9c82234d7e9b83ee8441a6fae2b879dfd995b7ea6905f2.jpg) +vegetation trunk terrain pole traffic sign + +ness and efficiency of SOAP. Also, SOAP with ResNet50 yields remarkable mIoU and IoU performance with comparable complexity compared to other methods. + +Qualitative Results. Figure 5 presents qualitative comparisons between SOAP and OccFormer [50] in terms of occluded region reconstruction and similar class discrimination. In Figure 5 (top), OccFormer fails to recognize occluded regions behind car objects, resulting in fused car shapes, whereas SOAP appropriately distinguishes and separates the cars that are actually apart. Also, as in Figure 5 (bottom), Occformer confuses classes with similar appearance (road $\leftrightarrow$ sidewalk) while SOAP reasonably differentiates these classes. + +# 4.4. Ablation study + +We analyze the contributions of individual components of the proposed SOAP through ablation experiments on the SemanticKITTI validation set. Table 4 reports the contributions of the occluded region-aware projection (OAP) and the scene-adaptive decoder (SAD). For the results without SAD, we replace it with a standard transformer decoder [6, 50, 51]. For the results without OAP, we input the initial voxel features $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ directly into the voxel + +Table 4. Ablation study on different combinations of occluded region-aware projection (OAP) and scene-adaptive decoder (SAD). + +
OAPSADSC IoUSSC mIoU
43.3216.86
✓43.5117.22
✓46.5018.99
✓✓47.2419.21
+ +Table 5. Alternative approaches for historical information. + +
MethodSC IoUSSC mIoU
Addition45.0018.05
DCA45.9018.53
OAP47.2419.21
+ +encoder, without using any historical information. In Table 4, IoU and mIoU scores show significant degradation without either OAP or SAD. This indicates that OAP effectively refines occluded regions during its voxel generation, while SAD provides representative semantic features across dynamic driving scenes for query embeddings. Notably, SOAP still retains reasonable performance even without OAP, which implements 3D SSC without temporal information. + +To validate the effectiveness of OAP, we design several alternative approaches to utilize historical information. First, we simply add the historical voxel features $\mathbf{F}_i^{\mathrm{hist}}$ to the initial voxel features $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ . Second, we employ the standard deformable cross-attention (DCA) [53] between the initial voxel features $\tilde{\mathbf{F}}_i^{3\mathrm{D}}$ and the historical voxel features $\mathbf{F}_i^{\mathrm{hist}}$ without occluded region separation. In Table 5, the proposed OAP achieves better IoU and mIoU scores than the standard DCA, with a margin of $+1.34$ and $+0.68$ , respectively. Therefore, perception of occluded regions during view projection is essential for accurate 3D semantic scene reasoning. + +# 5. Conclusion + +We proposed a novel 3D semantic scene completion algorithm, called SOAP. The main contributions of SOAP are the occluded region-aware view projection and the scene-adaptive decoder. The occluded region-aware view projection addresses the information deficiency in vision-centric 3D SSC by refining occluded voxel regions with historical observation. The scene-adaptive decoder creates the comprehensive semantic repository using the repository builder. Then, the scene-adaptive decoder integrates informative clues from the semantic repository into query embeddings. Experiments demonstrated that SOAP significantly outperforms the state-of-the-arts on the SemanticKITTI and SSC-Bench datasets, with a margin of $+2$ and $+2.67$ mIoU scores, respectively. + +# Acknowledgements + +This work was partly by the National Research Foundation of Korea (NRF) grants funded by the Ministry of Education (RS-2022-NR075739) and the Korea government (MSIT) (No. RS-2024-00352566, No. RS-2025-00559165), and partly by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155857, Artificial Intelligence Convergence Innovation Human Resources Development (Chungnam National University)). + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6 +[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1 +[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 1, 2, 6, 7 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6 +[5] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020. 1, 2 +[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2, 3, 5, 6, 8 +[7] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161, 2021. 2 +[8] Emeç Erçelik, Ekim Yurtsever, Mingyu Liu, Zhijie Yang, Hanzhen Zhang, Pinar Topçam, Maximilian Listl, Yilmaz Kaan Cayli, and Alois Knoll. 3d object detection with a self-supervised lidar scene flow backbone. In ECCV, pages 247-265, 2022. 1 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7 +[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7 +[11] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, pages 4012-4021, 2022. 1 + +[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 3, 6, 7 +[13] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In CVPR, pages 20258-20267, 2024. 1, 2, 3, 6, 7 +[14] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, pages 515–532, 2020. 1 +[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6 +[16] Abhijit Kundu, Yin Li, and James M Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, pages 3559–3568, 2018. 1 +[17] Hyo-Jun Lee, Hanul Kim, Su-Min Choi, Seong-Gyun Jeong, and Yeong Jun Koh. Baam: Monocular 3d pose and shape reconstruction with bi-contextual attention module and attention-guided modeling. In CVPR, pages 9011-9020, 2023. 1 +[18] Bohan Li, Jiajun Deng, Wenyao Zhang, Zhujin Liang, Dalong Du, Xin Jin, and Wenjun Zeng. Hierarchical temporal context learning for camera-based semantic scene completion. In ECCV, pages 131-148. Springer, 2024. 1, 3, 6, 7 +[19] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020. 1, 2 +[20] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In AAAI, pages 1477-1485, 2023. 7 +[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 6, 7 +[22] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, pages 9087-9098, 2023. 1, 2, 3, 6, 7 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7 +[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021. 1 +[26] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Jongwon Ra, Yukai Ma, Laijian Li, and Yong Liu. Camera-based + +3d semantic scene completion with sparse guidance network. IEEE TIP, 2024. 3, 6, 7 +[27] Chao Ning and Hongping Gan. Trap attention: Monocular depth estimation with manual traps. In CVPR, pages 5033-5043, 2023. 1 +[28] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020. 2, 4 +[29] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. idisc: Internal discretization for monocular depth estimation. In CVPR, pages 21477-21487, 2023. 1 +[30] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. IEEE TPAMI, 44(10): 7205-7218, 2021. 2 +[31] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020. 2 +[32] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 2417-2426, 2022. 3 +[33] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 1 +[34] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2 +[35] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 7 +[36] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Not all voxels are equal: Semantic scene completion from the point-voxel perspective. In AAAI, pages 2352–2360, 2022. 2 +[37] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1, 2, 3, 6, 7 +[38] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 3, 6 +[39] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In CVPR, pages 17642-17651, 2023. 1, 2 +[40] Jinfeng Xu, Xianzhi Li, Yuan Tang, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Casfusionnet: A cascaded network for point cloud semantic scene completion by dense feature fusion. In AAAI, pages 3018-3026, 2023. 2 +[41] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 1, 2 + +[42] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9421-9431, 2023. 3, 6 +[43] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1 +[44] Zhu Yu, Running Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Siyuan Cao, and Huiliang Shen. Context and geometry aware voxel transformer for semantic scene completion. In NeurIPS, 2024. 3, 7 +[45] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3916-3925, 2022. 1 +[46] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In ICCV, pages 9155-9166, 2023. 1 +[47] Shoulong Zhang, Shuai Li, Aimin Hao, and Hong Qin. Point cloud semantic scene completion from rgb-d images. In AAAI, pages 3385-3393, 2021. 2 +[48] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, pages 1589-1598, 2022. 1 +[49] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 7 +[50] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 1, 2, 3, 5, 6, 7, 8 +[51] Linqing Zhao, Xiuwei Xu, Ziwei Wang, Yunpeng Zhang, Borui Zhang, Wenzhao Zheng, Dalong Du, Jie Zhou, and Jiwen Lu. Lowrankocc: Tensor decomposition and low-rank recovery for vision-based 3d semantic occupancy prediction. In CVPR, pages 9806-9815, 2024. 1, 2, 3, 5, 6, 7, 8 +[52] Yunsong Zhou, Hongzi Zhu, Quan Liu, Shan Chang, and Minyi Guo. Monoatt: Online monocular 3d object detection with adaptive token transformer. In CVPR, pages 17493-17503, 2023. 1 +[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 5, 6, 8 +[54] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1 \ No newline at end of file diff --git a/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/images.zip b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d414d28bc436d1e478d2ee4f27ccf2fc310cc8e3 --- /dev/null +++ b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37049c4af02fff665bf1e783330d7006361ee5e84785a4147cc86235aba88174 +size 757613 diff --git a/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/layout.json b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9eab9e1e26838cbd13652efa6e48031a6e5ac506 --- /dev/null +++ b/2025/SOAP_ Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection/layout.json @@ -0,0 +1,10267 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "text", + "content": "SOAP: Vision-Centric 3D Semantic Scene Completion with Scene-Adaptive Decoder and Occluded Region-Aware View Projection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 160, + 242, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 160, + 242, + 174 + ], + "spans": [ + { + "bbox": [ + 172, + 160, + 242, + 174 + ], + "type": "text", + "content": "Hyo-Jun Lee" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 175, + 263, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 175, + 263, + 190 + ], + "spans": [ + { + "bbox": [ + 184, + 175, + 263, + 190 + ], + "type": "text", + "content": "Hyunseop Kim" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 267, + 161, + 351, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 161, + 351, + 175 + ], + "spans": [ + { + "bbox": [ + 267, + 161, + 351, + 175 + ], + "type": "text", + "content": "Yeong Jun " + }, + { + "bbox": [ + 267, + 161, + 351, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Koh}^{2*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 288, + 175, + 353, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 288, + 175, + 353, + 190 + ], + "spans": [ + { + "bbox": [ + 288, + 175, + 353, + 190 + ], + "type": "text", + "content": "Yonguk Lee" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 376, + 161, + 436, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 161, + 436, + 175 + ], + "spans": [ + { + "bbox": [ + 376, + 161, + 436, + 175 + ], + "type": "text", + "content": "Hanul Kim3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 378, + 175, + 425, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 175, + 425, + 188 + ], + "spans": [ + { + "bbox": [ + 378, + 175, + 425, + 188 + ], + "type": "text", + "content": "Jinu Lee" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 139, + 191, + 290, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 191, + 290, + 205 + ], + "spans": [ + { + "bbox": [ + 139, + 191, + 290, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 191, + 290, + 205 + ], + "type": "text", + "content": "Kangwon National University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 191, + 471, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 471, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 471, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 313, + 191, + 471, + 206 + ], + "type": "text", + "content": "Chungnam National University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 206, + 397, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 206, + 397, + 220 + ], + "spans": [ + { + "bbox": [ + 134, + 206, + 397, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 134, + 206, + 397, + 220 + ], + "type": "text", + "content": "Seoul National University of Science and Technology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 421, + 206, + 477, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 206, + 477, + 220 + ], + "spans": [ + { + "bbox": [ + 421, + 206, + 477, + 220 + ], + "type": "text", + "content": "42dot Inc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 122, + 223, + 248, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 223, + 248, + 234 + ], + "spans": [ + { + "bbox": [ + 122, + 223, + 248, + 234 + ], + "type": "text", + "content": "gywns6298@kangwon.ac.kr" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 268, + 224, + 351, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 224, + 351, + 233 + ], + "spans": [ + { + "bbox": [ + 268, + 224, + 351, + 233 + ], + "type": "text", + "content": "yjkoh@cnu.ac.kr" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 223, + 488, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 223, + 488, + 233 + ], + "spans": [ + { + "bbox": [ + 373, + 223, + 488, + 233 + ], + "type": "text", + "content": "hukim@seoultech.ac.kr" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 237, + 239, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 237, + 239, + 247 + ], + "spans": [ + { + "bbox": [ + 127, + 237, + 239, + 247 + ], + "type": "text", + "content": "hyunseop95@gmail.com" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 259, + 237, + 363, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 237, + 363, + 247 + ], + "spans": [ + { + "bbox": [ + 259, + 237, + 363, + 247 + ], + "type": "text", + "content": "yonguk.lee@42dot.ai" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 237, + 478, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 237, + 478, + 247 + ], + "spans": [ + { + "bbox": [ + 386, + 237, + 478, + 247 + ], + "type": "text", + "content": "jinu.lee@42dot.ai" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "spans": [ + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 300, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 540 + ], + "type": "text", + "content": "Existing view transformations in vision-centric 3D Semantic Scene Completion (SSC) inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, leading to a dilution of informative contexts. Furthermore, semantic classes exhibit high variability in their appearance in real-world driving scenarios. To address these issues, we introduce a novel 3D SSC method, called SOAP, including two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection effectively converts 2D image features into voxel space, refining the duplicated features of occluded regions using information gathered from previous observations. The scene-adaptive decoder guides query embeddings to learn diverse driving environments based on a comprehensive semantic repository. Extensive experiments validate that the proposed SOAP significantly outperforms existing methods for the vision-centric 3D SSC on automated driving datasets, SemanticKITTI and SSCBench. Code is available at https://github.com/gywns6287/SOAP." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 577, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 696 + ], + "type": "text", + "content": "Understanding 3D traffic scenes is crucial for autonomous driving systems to develop downstream tasks such as motion prediction and planning. LiDAR sensors [8, 33, 43, 54] provide precise 3D scene information, but their high cost and the complex setup limit their scalability. In contrast, areas of study about vision-centric 3D perception, including 3D object detection [11, 25, 46, 48, 52], 3D car shape reconstruction [14, 16, 17], and depth estimation [2, 27, 29, 45], have been receiving attention due to its simplicity and cost-effectiveness. However, these methods still struggle to re" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 314, + 274, + 556, + 373 + ], + "blocks": [ + { + "bbox": [ + 314, + 274, + 556, + 373 + ], + "lines": [ + { + "bbox": [ + 314, + 274, + 556, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 556, + 373 + ], + "type": "image", + "image_path": "5a000ecd8d9ad8d5aa724a75275853163d3a45af72c5e0209c16329be42543dd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 374, + 479, + 384 + ], + "lines": [ + { + "bbox": [ + 389, + 374, + 479, + 384 + ], + "spans": [ + { + "bbox": [ + 389, + 374, + 479, + 384 + ], + "type": "text", + "content": "(a) Occluded voxel regions." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 315, + 388, + 556, + 502 + ], + "blocks": [ + { + "bbox": [ + 315, + 388, + 556, + 502 + ], + "lines": [ + { + "bbox": [ + 315, + 388, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 556, + 502 + ], + "type": "image", + "image_path": "f8e435855253d1c9952d8a9efcc9689e8fffb2295847165759e72d8f5a237b5f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 502, + 517, + 512 + ], + "lines": [ + { + "bbox": [ + 352, + 502, + 517, + 512 + ], + "spans": [ + { + "bbox": [ + 352, + 502, + 517, + 512 + ], + "type": "text", + "content": "(b) Diversity in the appearance of semantic classes." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 521, + 555, + 577 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 577 + ], + "type": "text", + "content": "Figure 1. Two main motivations of SOAP. (a) The 2D-to-3D view projection inevitably yields erroneous feature duplication in voxel generation due to occlusion. (b) The semantic classes exhibit significant variability in their appearance depending on driving regions." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "construct complete and fine-grained 3D scene representations. To address this limitation, 3D Semantic Scene Completion (SSC) has emerged as a promising alternative. 3D SSC leverages a voxel-based approach, assigning both an occupancy probability and a semantic label (e.g. car, pedestrian, road) to each voxel within the 3D space. Thereby, it can create a comprehensive representation of the driving scene. While earlier works [5, 19, 39, 41] primarily rely on LiDAR sensors, recent research [3, 13, 18, 22, 37, 50, 51] has demonstrated the potential of vision-centric methods." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17145" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "A general vision-based 3D SSC model comprises an image encoder, a view projection module, and a voxel decoder. The image encoder first extracts 2D features from an input image, and then the view projection module subsequently lifts 2D image features into 3D voxel space. Given the 3D features of each voxel region, the voxel decoder estimates semantic classes corresponding to each voxel region. In this process, the precise 2D-3D view projection plays a crucial role. However, existing view projection methods [3, 50] inevitably experience erroneous feature duplication in the reconstructed voxel space due to occlusions, particularly along the longitudinal direction. For instance, as illustrated in Figure 1a, when a camera ray encounters an occupancy region (e.g. car), occluded regions (voxels behind this occupancy) may contain duplicated features of the obstructing car in the 3D view projection process. This duplication can occur even when the actual classes of the occluded regions are not related to a car, resulting in incorrect 3D SSC results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 507 + ], + "type": "text", + "content": "Furthermore, the diverse appearances of semantic classes pose another significant challenge in 3D SSC. As shown in Figure 1b, semantic classes exhibit high variability in their appearance in real-world driving scenarios. For instance, road color and texture may vary depending on driving regions (e.g. urban or highway). Such appearance variations make it difficult for SSC models to distinguish similar classes, such as road " + }, + { + "bbox": [ + 55, + 291, + 295, + 507 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 291, + 295, + 507 + ], + "type": "text", + "content": " sidewalk, leading to potential misinterpretations of driving environments. While most previous works [3, 13, 22, 37] usually employ sequential 3D convolutions for the semantic decoder, this approach often lacks the flexibility to generalize for the diverse semantic context. More recently, Mask2former [6]-like approach [50, 51] is adopted for the SSC decoder, which utilizes semantic tokens derived from a query embedding to represent the semantic scene. However, these semantic tokens lack adaptability to dynamic driving environments, as they are fundamentally based on static query embedding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel 3D semantic scene completion network, SOAP, faithfully designed to address the occluded voxel regions and the visual variations in semantic classes. The proposed method leverages two key components: an occluded region-aware view projection and a scene-adaptive decoder. The occluded region-aware view projection (OAP) first converts 2D image features into 3D voxel features using the LSS algorithm [28], where these initial voxel features contain false duplicated information due to occlusions, as illustrated in Figure 1a. These occluded regions can be inferred by referring to previous frames, which provides valuable context about the missing observation. Therefore, OAP identifies occluded regions in the current frame using depth estimation and refines the duplicated features of occluded regions using information gathered from previous observations. Also, a vanilla transformer decoder [6] generates the semantic tokens based on" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "static query embedding, which is not versatile enough to capture a dynamic driving environment. To represent more diverse semantics, a repository builder first creates a semantic repository by gathering semantic tokens from various scenes. Then, the scene-adaptive decoder integrates informative clues from this extensive repository into the semantic tokens, guiding them to navigate diverse driving environments effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 168, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 227 + ], + "type": "text", + "content": "Extensive experiments demonstrate that the proposed SOAP outperforms the other state-of-the-arts on two vision-centric SSC benchmarks, SemanticKITTI [1] and SSCBench-KITTI360 [21]. To summarize, the main contributions of our work are three folds:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 228, + 553, + 347 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 228, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 263 + ], + "type": "text", + "content": "- We develop the occluded region-aware view projection that effectively transforms 2D image features into 3D voxel space by recovering occluded voxel regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 264, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 553, + 299 + ], + "type": "text", + "content": "- We also introduce a scene-adaptive transformer that integrates informative semantic clues from the various scenes into the semantic tokens of an individual scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "type": "text", + "content": "- The proposed SOAP achieves outstanding performances on two driving datasets, SemanticKITTI and SSCBenchKITTI360, with a margin of " + }, + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "type": "inline_equation", + "content": "+2" + }, + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "type": "inline_equation", + "content": "+2.67" + }, + { + "bbox": [ + 314, + 300, + 553, + 347 + ], + "type": "text", + "content": " mIoU scores, respectively." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 358, + 405, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 358, + 405, + 370 + ], + "spans": [ + { + "bbox": [ + 314, + 358, + 405, + 370 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 378, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 676 + ], + "type": "text", + "content": "3D Semantic Scene Completion. 3D semantic scene completion (SSC) is an advanced task in 3D scene reconstruction fields, which aims to predict a dense semantic scene representation involving the geometrical occupancy and its corresponding semantic label. Early SSC methods [5, 19, 34] leverage 3D inputs to reconstruct voxelized indoor scenes. SSCNet [34] first defines the SSC task, which jointly reasons the geometry and semantics of an indoor scene from a single-depth image. 3DSketch [5] uses the Truncated Signed Distance Function to guide an occupancy and the semantic labels of the indoor scene. AIC-Net [19] develops anisotropic convolution to extract hybrid features from the depth and RGB images for indoor SSC tasks. Recently, the release of SemanticKITTI [1], the first large-scale outdoor SSC dataset, yields the development of diverse outdoor SSC methods. LMSCNet [31] uses the 2D UNet to extract collapsed BEV features from sparse 3D Li-DAR scans and finally expands the height dimension for the 3D SSC. JS3C-Net [41] learns the contextual shape priors from sequential LiDAR data completing the sparse single sweep point cloud to the dense one for enhanced 3D SSC. The other existing works [7, 30, 36, 39, 40, 47] also commonly depend on 3D input such as LiDAR point clouds. In contrast, we address vision-centric 3D SSC, which only uses an RGB image for the semantic scene reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Vision-centric 3D SSC. MonoScene [3] introduces the first vision-centric approach for SSC, utilizing a 3D UNet to process voxel features generated by sight projection. Vox-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17146" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 544, + 313 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 544, + 313 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 544, + 313 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 544, + 313 + ], + "type": "image", + "image_path": "a5c87da77491001fd7b2aabc086de697c19964d59e2999dca521239f6247600a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 555, + 366 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed SOAP. Given multi-scale image features, an occluded region-aware view projection (OAP) transforms them into voxel space, refining duplicated features in occluded voxel regions using historical information. A scene-adaptive decoder enriches query embeddings with diverse semantic contexts through a semantic repository, which is constructed and iteratively updated by a repository builder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 385, + 297, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 297, + 589 + ], + "type": "text", + "content": "Former [22] estimates occupancy regions using stereo depth estimation and selectively updates these regions with RGB images. HASSC [37] and Symphonies [13] also identify voxel occupancy based on stereo depth estimation, following VoxFormer. NDC-scene [42] extends a 2D feature map to a normalized device coordinates space to restore the depth dimension and subsequently reconstructs a 3D semantic scene using this space. SurroundOcc [38] adopts spatial 2D-3D attention to lift multi-scale 2D features into the 3D volume space. TPVFormer [12] introduces a query-based tri-perspective space for 3D SSC. Occformer [50] and LowRankOcc [51] reconstruct 3D scenes using semantic tokens on the Mask2former decoders [6]. SGN [26] propagates semantic contexts from semantic-aware seed voxels based on their spatial geometry cues. HTCL [18] leverages historical frames to refine current frame features through affinity-based dynamic refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 602, + 112, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 112, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 112, + 615 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 623, + 155, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 155, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 155, + 635 + ], + "type": "text", + "content": "3.1. Model Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "3D Semantic Scene Completion (SSC) aims to predict semantic probabilities of the voxelized scene " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{X \\times Y \\times Z \\times S}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "(X, Y, Z)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " denotes the resolution of the 3D volume, and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the number of semantic classes including a free (or empty) class. Figure 2 illustrates the overall architecture of the proposed SOAP. Image encoder first ex" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": "tracts multi-scale image features " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^{2\\mathrm{D}} \\in \\mathbb{R}^{H_i \\times W_i \\times C}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " from an RGB image, where " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " is the number of scales, " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "(H_i, W_i)" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " is the spatial size at the scale " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " is the channel dimension. Depth estimator predicts the depth map " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " from RGB images. Consistent with previous approaches [13, 18, 22, 37, 44], we also use off-the shelf stereo depth model [32] for the depth estimator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "text", + "content": "The occluded region-aware view projection (Section 3.2) utilizes the estimated depth map " + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "text", + "content": " to lift multi-scale image features " + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^{2\\mathrm{D}}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "text", + "content": " into multi-scale voxel features " + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^{3\\mathrm{D}}\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i\\times C}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "text", + "content": ", enhancing occluded voxel regions based on informative features in historical frames. The scene-adaptive decoder (Section 3.3) first constructs a semantic repository using a repository builder and then it integrates semantic diversity from a semantic repository into query embeddings to produce scene-adaptive semantic tokens, which are utilized to reconstruct the final semantic voxel representation " + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 469, + 556, + 636 + ], + "type": "text", + "content": ", in conjunction with multiscale voxel features. Through this complementary process, SOAP guarantees a more flexible and accurate 3D SSC for a real-world driving environment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 529, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 529, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 529, + 654 + ], + "type": "text", + "content": "3.2. Occluded Region-Aware View Projection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Occluded region-aware view projection (OAP) transforms 2D image features " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^{2\\mathrm{D}}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " into 3D voxel space " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_i^{3\\mathrm{D}}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": ". Figure 3 illustrates the detailed process of OAP. It produces voxel features for each " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "-th scale solely from the image feature at " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "-th scale." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17147" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 536, + 247 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 536, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 536, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 536, + 247 + ], + "type": "image", + "image_path": "c035787505e4718e3d4720049c84aded0b31c19efa75517a8d4e6c4c0f192e8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "type": "text", + "content": "Figure 3. Occluded region-aware view projection. (a) occluded region identification identifies occluded regions in the current frame by inferring an occlusion map based on the depth distribution. (b) occluded region refinement refines the duplicated features in the identified occluded regions using previous observations within the historical voxel features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 310, + 294, + 502 + ], + "blocks": [ + { + "bbox": [ + 56, + 310, + 294, + 502 + ], + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 502 + ], + "type": "image", + "image_path": "12214c58a5d4ec297e5914d097ccdba18ee4ef28fbec11dc72cf99b58d551342.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "content": "Figure 4. t-SNE visualization of voxel features in the occluded regions for (a) " + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "content": " and (b) " + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "content": ". Different colors represent different camera rays." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": "Initial Voxel Generation. For each " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": "th scale, OAP first generates initial voxel features " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " from the 2D image features " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " and the depth estimation " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": ". Specifically, OAP resizes the estimated depth " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_i" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " that has the spatial resolution of " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "(H_i, W_i)" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": ", and distributes it to the depth distribution map " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i \\in \\mathbb{R}^{H_i \\times W_i \\times D}" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": ". Here, the continuous depth value " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_i(h, w)" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " for each pixel " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "(h, w)" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " is converted into discrete depth bins " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i(h, w) \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " represented as one-hot encoded vectors, followed by a Gaussian filter to create a soft depth distribution. Then, OAP computes the outer product " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{2\\mathrm{D}} \\otimes \\mathbf{D}_i" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": " to create the depth distributed features " + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{depth}} \\in \\mathbb{R}^{H_i \\times W_i \\times D \\times C}" + }, + { + "bbox": [ + 55, + 568, + 296, + 715 + ], + "type": "text", + "content": ". The voxel-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": "pooling operation " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\varphi(\\cdot)" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " in [28] can project " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{depth}}" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " into the voxel space using camera intrinsics and extrinsics. Thus, the initial voxel features are defined as " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}} = \\varphi(\\mathbf{F}_i^{\\mathrm{depth}})" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": ". Historical Voxel Generation. To represent previous observations, we incorporate depth distributed features " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{i,t - p}^{\\mathrm{depth}}\\}_{p = 1}^{P}" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " from the previous frames into the current voxel space " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{hist}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " is the number of previous frames at timestamp " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": ". Specifically, each past " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i,t - p}^{\\mathrm{depth}}" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " is projected into voxel space " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\varphi(\\mathbf{F}_{i,t - p}^{\\mathrm{depth}})" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " and then aligned to the current voxel space at " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " via the warping function " + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(\\cdot)" + }, + { + "bbox": [ + 313, + 308, + 555, + 443 + ], + "type": "text", + "content": " and summation operation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 352, + 448, + 555, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 448, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 352, + 448, + 555, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {\\mathrm {h i s t}} = \\sum_ {p = 1} ^ {P} \\mathcal {W} \\left(\\varphi \\left(\\mathbf {F} _ {i, t - p} ^ {\\mathrm {d e p t h}}\\right), \\mathcal {G} _ {t} ^ {- 1} \\cdot \\mathcal {G} _ {t - p}\\right) \\tag {1}", + "image_path": "32812eb6c0792ea0b0fd973ae8c35c99cb247a092688b92fb0225fdb7f28979a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{t - p}" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": " are ego-posed matrices at " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "t - p" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": ". These historical voxel features " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{hist}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": " provide valuable perspectives that are not available in the initial voxel features. Occluded Region Identification. The initial voxel features " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": ", which are projected from a single image view, inherently include false duplicated features due to occlusions. Figure 4a presents the t-SNE visualization of features in the occluded regions within " + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 653 + ], + "type": "text", + "content": ". Each point represents an individual voxel region, and voxel regions of the same color correspond to those located along the same camera ray. We observe that most occluded regions along the same camera ray exhibit similar patterns, which indicates erroneous replication of the features from the occupancy regions positioned in front of the occluded regions as in Figure 1a." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "To address this limitation, OAP refines these occluded features by referring to previous frames, which provide valuable contexts about the missing observation in the current frame. As part of this process, we identify the occluded regions using the estimated depth distribution " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17148" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "As in Figure 3a, each vector " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i(h,w)" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " at the spatial location " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " contains the probability distribution of the nearest occupancy region along the longitudinal (or depth) direction. In other words, regions located behind the nearest occupancy with the highest depth probability often lose their semantic context information due to occlusions. To distinguish the occluded regions from non-occluded regions, we define an occlusion map " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i\\in \\{1,0\\}^{H_i\\times W_i\\times D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 178, + 296, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 178, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 64, + 178, + 296, + 223 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {i} (h, w, d) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} d > \\underset {\\tilde {d} \\in \\{1, \\dots , D \\}} {\\arg \\max } \\mathbf {D} _ {i} (h, w, \\tilde {d}) \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "3fc410137f58847354d640b0eb269c472169aea95270e40788d010c171412a88.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "text", + "content": "Figure 3a provides an example of " + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "text", + "content": ", where occluded regions are marked with 1 (taller rectangular) and non-occluded regions are marked with 0 (shorter rectangular). To identify occluded regions in the voxel space, the occlusion map " + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "text", + "content": " is projected into voxel space through the voxel pooling operation, i.e. " + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\varphi (\\mathbf{O}_i)\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i}" + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "text", + "content": ". Note that the regions identified by this occlusion map " + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\varphi (\\mathbf{O}_i)" + }, + { + "bbox": [ + 55, + 232, + 296, + 374 + ], + "type": "text", + "content": " contain incorrectly duplicated information from the nearest occupancy region along the longitudinal direction, as illustrated in Figure 4a. Therefore, we refine the features of these occluded regions through an occluded region refinement process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "text", + "content": "Occluded Region Refinement. The occluded regions at the current time step may have been observable in previous time steps. Therefore, we refine deficient features in the occluded regions with previous observations in the historical voxel features " + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{hist}}" + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "text", + "content": ". To achieve this, we perform 3D deformable cross-attention, DCA [53] between the historical voxel features " + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{hist}}" + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "text", + "content": " and the occluded voxel features, resulting in reconstructed occluded voxel features " + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{occ}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}" + }, + { + "bbox": [ + 55, + 376, + 296, + 485 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 495, + 296, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 557 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {\\text {o c c}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z) + \\mathrm {D C A} \\left(\\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z), \\mathbf {F} _ {i} ^ {\\text {h i s t}}\\right), \\\\ \\quad \\text {i f} \\varphi (\\mathbf {O} _ {i}) (x, y, z) \\geq 1 \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "17733d77ca9359e26248e2c59a55081565b2381f937d2b390233e10efb9adca3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "text", + "content": " are spatial indices of the voxel space. This approach ensures that the reconstructed features " + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "text", + "content": " effectively capture observable historical information about occluded regions, enabling more accurate and context-aware scene representation. As shown in Figure 4b, the reconstructed features " + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 557, + 296, + 653 + ], + "type": "text", + "content": " include appropriately corrected information for the occluded regions, resulting in more diverse patterns according to different semantic regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Invisible and Visible Region Reconstruction. The initial voxel features also contain invisible regions due to the limited observation of a single camera. Since these invisible voxels do not contain any features, we directly fill them with the corresponding historical voxel features to recon-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 71, + 533, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 533, + 86 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 533, + 86 + ], + "type": "text", + "content": "struct invisible voxel features " + }, + { + "bbox": [ + 314, + 71, + 533, + 86 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{inv}}\\in \\mathbb{R}^{X_i\\times Y_i\\times Z_i\\times C}" + }, + { + "bbox": [ + 314, + 71, + 533, + 86 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 92, + 553, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 92, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 319, + 92, + 553, + 140 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {\\text {i n v}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\mathbf {F} _ {i} ^ {\\text {h i s t}} (x, y, z) & \\text {i f} \\sum \\tilde {\\mathbf {F}} _ {i} ^ {\\text {3 D}} (x, y, z) = 0 \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "fc2e78bd630c28413f3f7f2af4ada419b072b1b90487d9e81ebf6a8070dddaeb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "type": "text", + "content": "We designate the voxel space, excluding invisible and occluded regions, as visible regions. The visible regions already contain appropriate semantic features with respect to the current observation, allowing us to directly retrieve the visible voxel features from the initial voxel features " + }, + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "type": "text", + "content": ". The visible voxel features " + }, + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{vis}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i \\times C}" + }, + { + "bbox": [ + 313, + 140, + 555, + 223 + ], + "type": "text", + "content": " are defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 228, + 553, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 228, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 324, + 228, + 553, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {\\mathrm {v i s}} (x, y, z) = \\left\\{ \\begin{array}{l l} \\tilde {\\mathbf {F}} _ {i} ^ {\\mathrm {3 D}} (x, y, z) & \\text {i f} \\varphi (\\mathbf {O} _ {i}) (x, y, z) = 0 \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}", + "image_path": "7eef4d62c46f0a9553e8b0d698b3787d8155b21dcdb6cc20987e35fefdf24738.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "content": "Voxel Encoder. To compute the final 3D voxel features, we sum the occluded, invisible, and visible voxel features, " + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{occ}} + \\mathbf{F}_i^{\\mathrm{inv}} + \\mathbf{F}_i^{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "content": ", and process them through a voxel encoder to obtain the final voxel features " + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{3\\mathrm{D}} \\in \\mathbb{R}^{X_i \\times Y_i \\times Z_i}" + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "content": ". In the voxel encoder, the spatial dimensions " + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "inline_equation", + "content": "(X_i \\times Y_i \\times Z_i)" + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "content": " of the voxel features are sliced into " + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "inline_equation", + "content": "Z_i" + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "content": " bird's-eye view (BEV) planes with spatial resolution " + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "inline_equation", + "content": "(X_i \\times Y_i)" + }, + { + "bbox": [ + 313, + 277, + 555, + 385 + ], + "type": "text", + "content": ". Then, two windowed self-attention blocks [23] are sequentially performed to each BEV plane." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 392, + 451, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 451, + 404 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 451, + 404 + ], + "type": "text", + "content": "3.3. Scene-Adaptive Decoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": "A standard transformer decoder [6, 50, 51] yields semantic tokens starting from static query embeddings, which lack adaptability for dynamic driving scenes. To represent more diverse semantics, the scene-adaptive decoder correlates semantic tokens for a specific scene with tokens from different scenes. As in Figure 2, the scene-adaptive decoder generates scene-adaptive tokens " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times C}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": " iterative transformer blocks, where " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{q}}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": " is the number of tokens. Each transformer block in the scene transformer cascades four modules including a cross-attention between query embeddings and the semantic repository, the other cross-attention between the query embeddings and the voxel features " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": ", self-attention, and feed-forward networks. The cross-attention with the semantic repository integrates semantic features from dynamic scenes into the query embeddings, while the cross-attention with voxel features provides scene-specific semantic features. After the end of iterations, scene-adaptive tokens " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": " are projected to class predictions " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{\\mathrm{cls}} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": " and segmentation mask predictions " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{X \\times Y \\times Z \\times N_{\\mathrm{q}}}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": ". Then, the final semantic voxel " + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{X \\times Y \\times Z \\times S}" + }, + { + "bbox": [ + 313, + 409, + 556, + 661 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 667, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 667, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 386, + 667, + 553, + 682 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\mathcal {F} ^ {- 1} (\\mathcal {F} (\\mathbf {M}) \\mathbf {S} ^ {\\mathrm {c l s}}) \\tag {6}", + "image_path": "5e8b50a2c23d624ba3f3a5632fc3af29c473b612b88fc2b254ce9a5430c0bd2c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:\\mathbb{R}^{X\\times Y\\times Z\\times C}\\to \\mathbb{R}^{XYZ\\times C}" + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": " is as spatial flatten and " + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 313, + 688, + 555, + 714 + ], + "type": "text", + "content": " is its inverse operation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17149" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 555, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Semantic scene completion results on SemanticKITTI test set. † represents temporal methods that leverage previous frames for 3D SSC. We highlight the best results in bold and the second best results in underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 555, + 327 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoUroad (15.30%)sidewalk (11.13%)parking (1.12%)other-ground (0.56%)building (14.1%)car (3.92%)truck (0.16%)bicycle (0.03%)motorcycle (0.03%)other-veh. (0.20%)vegetation(39.3%)trunk (0.51%)terrain (9.17%)person (0.07%)bicyclist (0.07%)motorcyclist (0.05%)fence (3.90%)pole (0.29%)traf-sign (0.08%)
MonoScene [3]34.1611.0854.727.124.85.714.418.83.30.50.74.414.92.419.51.01.40.411.13.32.1
TPVFormer [12]34.2511.2655.127.227.46.514.819.23.71.00.52.313.92.620.41.12.40.311.02.91.5
SurroundOcc [38]34.7211.8656.928.330.26.815.220.61.41.61.24.414.93.419.31.42.00.111.33.92.4
VoxFormer† [22]43.2113.4154.126.925.17.323.521.73.61.91.64.124.48.124.21.61.10.013.16.65.7
OccFormer [50]34.5312.3255.930.331.56.515.721.61.21.51.73.216.83.921.32.21.10.211.93.83.7
NDC-Scene [42]33.8711.5556.228.728.05.615.819.71.81.11.14.914.32.620.60.71.70.411.23.21.7
LowRankOcc [51]38.4713.5652.827.225.18.822.120.92.93.32.74.422.98.920.82.41.72.314.47.07.0
Symphonies [13]42.1915.0458.429.326.911.724.723.63.23.62.65.624.210.023.13.21.92.016.17.78.0
HASSC† [37]42.8714.3855.329.625.911.323.123.02.91.91.54.924.89.826.51.43.00.014.37.07.1
SGN† [26]45.4215.7660.431.428.98.728.425.44.50.91.63.727.412.628.40.50.30.118.110.08.3
HTCL† [18]44.2317.0964.434.833.812.425.927.35.71.82.25.425.310.831.21.13.10.921.19.08.3
\\(SOAP^{\\dagger}\\) (proposed)46.0919.0963.636.236.817.228.728.93.35.05.37.029.815.132.33.72.30.222.511.713.1
", + "image_path": "0e81f49a627d0638da5baa8e30281c1492db0c5b457e9d03657d36f0c2e90799.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": "Repository Builder. The semantic repository " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{(S \\times K) \\times C}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " stores " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " representative tokens for each semantic class, iteratively updated by semantic tokens with top- " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " highest scores. The detailed process of the repository builder is shown in Figure 2. Even though representative tokens could be selected using the classification scores " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " of the scene-adaptive tokens " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": ", relying solely on " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " makes a repository filled with similar token and consequently degrades semantic diversity. To retain the semantic diversity within the repository " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": ", we introduce a novelty-based token selection. Specifically, we re-define token scores " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathbf{S}^{\\mathrm{cls}} + \\alpha \\mathbf{S}^{\\mathrm{nov}}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{\\mathrm{nov}} \\in \\mathbb{R}^{N_{\\mathrm{q}} \\times S}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " is novelty scores and " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " is its balancing parameter. To compute the novelty scores, the " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " tokens within each class in " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " are first averaged, resulting in the class average repository " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{R}} \\in \\mathbb{R}^{S \\times C}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": ". Then, the novelty scores " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{\\mathrm{nov}}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " are calculated based on cosine similarity between each token in " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 346, + 297, + 550 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 562, + 295, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 562, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 107, + 562, + 295, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {S} ^ {\\text {n o v}} (n, s) = - \\cos (\\mathbf {Q} (n), \\bar {\\mathbf {R}} (s)) \\tag {7}", + "image_path": "d82b0e1d4bdc78929bdd644590ff54089c9c1af43dbd7e5df6b53a57aee0a9f3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "content": " indicate token indices of " + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\cos(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 581, + 295, + 616 + ], + "type": "text", + "content": " is a cosine similarity between two input vectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Then, new tokens in " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " with higher scores than old tokens in " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " are selected and, newly pushed into the semantic repository " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". When a class holds more than " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " tokens, the old tokens with lower scores are discarded to maintain " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " representative tokens for each semantic class in " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Updating " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " with tokens that exhibit both high classification and novelty scores enables the repository to capture rich semantic information, effectively covering dynamic driving" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 346, + 555, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 383 + ], + "type": "text", + "content": "scenes. Note that the semantic repository " + }, + { + "bbox": [ + 313, + 346, + 555, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 346, + 555, + 383 + ], + "type": "text", + "content": " is updated only during the SOAP's training phase to ensure stable inference performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 392, + 408, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 408, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 408, + 403 + ], + "type": "text", + "content": "3.4. Loss Functions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "content": "We perform the Hungarian matching [15] between predicted and ground-truth voxel masks following [4, 6, 53]. For each matched pair, we compute a mask loss " + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{mask}}" + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "content": ", a mask dice loss " + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dice}}" + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "content": ", and a class loss " + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 410, + 555, + 469 + ], + "type": "text", + "content": " as in [6]. The final loss is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 349, + 482, + 555, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 482, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 349, + 482, + 555, + 495 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {\\text {m a s k}} L _ {\\text {m a s k}} + \\lambda_ {\\text {d i c e}} L _ {\\text {d i c e}} + \\lambda_ {\\text {c l s}} L _ {\\text {c l s}} \\tag {8}", + "image_path": "9722550bf882074035e6709bd5e0e0a8f1aefdeb35cf6570b197c4be38f4d35a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 505, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 529 + ], + "type": "text", + "content": "where, scaling weights " + }, + { + "bbox": [ + 313, + 505, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{mask}}, \\lambda_{\\mathrm{dice}}, \\lambda_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 505, + 554, + 529 + ], + "type": "text", + "content": " are set to 5, 5, 2, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 542, + 395, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 395, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 395, + 556 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 563, + 376, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 376, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 376, + 574 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "text", + "content": "We evaluate the performance of SOAP on two outdoor 3D SSC datasets: SemanticKITTI [1] and SSCBenchKITTI360 [21]. Both datasets provide front-view RGB images, LiDAR sweeps, and semantic scene labels represented as " + }, + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "text", + "content": " voxel grids. The scene labels cover a range of " + }, + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "inline_equation", + "content": "51.2m \\times 51.2m \\times 6.4m" + }, + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "text", + "content": " from the front view of the vehicle." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "SemanticKITTI [1]. It contains a total of 22 sequences, which are split into " + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "10/1/11" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": " for train/val/test following [3]. We utilize the RGB images of cam-2 with a " + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "1280 \\times 384" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": " resolution as model inputs. In the semantic labels, each voxel" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17150" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 555, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Semantic scene completion results on the SSCBench test set. † represents methods that use temporal information. We highlight the best results in **bold** and the second best results in **underline**." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 555, + 251 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoUcarbicyclemotorcycletruckother-veh.personroadparkingsidewalkother-groundbuildingfencevegetationterrainpoletraf.-signother-struct.other-object
MonoScene [3]37.8712.3119.30.40.68.02.00.948.411.428.13.232.93.526.216.86.95.74.23.1
TPVFormer [12]40.2213.6421.61.11.48.12.62.453.012.031.13.834.84.830.117.57.55.95.52.7
OccFormer [50]40.2713.8122.60.70.39.93.82.854.313.431.53.636.44.831.019.57.88.57.04.6
VoxFormer†[22]38.7611.9117.81.20.94.62.11.647.09.727.22.931.25.029.014.76.56.93.82.4
Symphonies [13]44.1218.5830.01.95.925.112.18.254.913.832.86.935.18.638.311.514.09.614.411.3
SGN†[26]47.0618.2529.03.42.910.95.23.058.115.036.44.442.07.738.223.216.716.49.95.8
\\( \\mathrm{SOAP}^{\\dagger } \\) (proposed)48.1220.9229.95.67.814.47.66.160.917.440.35.445.310.640.524.816.821.012.69.9
", + "image_path": "fbd892223eb3fefd9971900474d2787825477382606abf22356d48a919a22ddf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 294, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 294, + 293 + ], + "type": "text", + "content": "is annotated with 21 semantic classes (19 semantics, 1 free, 1 unknown)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 365 + ], + "type": "text", + "content": "SSCBench-KITTI360 [21]. It is a recently introduced dataset for 3D SSC task. SSCBench-KITTI360 contains 9 long sequences, which are split into 7/1/1 for train/val/test. The input resolution of the RGB images is set to " + }, + { + "bbox": [ + 55, + 293, + 296, + 365 + ], + "type": "inline_equation", + "content": "1408 \\times 384" + }, + { + "bbox": [ + 55, + 293, + 296, + 365 + ], + "type": "text", + "content": ". The scene voxels in SSCBench-KITTI360 are annotated with 20 semantic classes (18 semantics, 1 free, 1 unknown)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 189, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 189, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 189, + 385 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": "Network Structure. Following other comparable methods [3, 12, 44, 50, 51], we employ EfficientNetB7 as our backbone. The multi-scale number " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " is 4 and the spatial dimensions of the multi-scale image features are " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\left(\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right)" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " of an input image, respectively. We set a spatial voxel size " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "X \\times Y \\times Z" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 16" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " and channel numbers " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " to 192. The spatial dimensions of the multiscale voxel features are " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "(1, \\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8})" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "X \\times Y \\times Z" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": ", respectively. We set the number of semantic tokens to " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{q}} = 100" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " while the depth distribution bin to " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "D = 112" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": ". In the repository builder, the balancing parameter for novelty scores is set to " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " and the number of representative tokens of the semantic repository " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " is set to 10. We use " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "P = 4" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " historical frames. The scene-adaptive decoder is repeated " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "L = 9" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " times. The output voxel is upsampled to " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " through the trilinear interpolation for the evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 581, + 295, + 665 + ], + "type": "text", + "content": "Training. We train the proposed SOAP for 30 epochs. We employ AdamW [24] optimizer with an initial learning rate 0.0001 and divide it by 10 at the 20 and 25 epochs. For the training, we use a mini-batch of size 4 on 4 RTX A6000 GPUs. For data augmentation, we use random resize, rotation, and flip for the image space and 3D flip for the 3D volume space following [10, 20, 49]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 141, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 141, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 141, + 683 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "For evaluation on SemanticKITTI and SSCBench, we use the mean intersection over union (mIoU) of semantic" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 293, + 553, + 370 + ], + "blocks": [ + { + "bbox": [ + 313, + 268, + 554, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 268, + 554, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 554, + 291 + ], + "type": "text", + "content": "Table 3. Complexity-accuracy comparison on SemanticKITTI validation set. † represents methods that use temporal information." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 293, + 553, + 370 + ], + "lines": [ + { + "bbox": [ + 315, + 293, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 553, + 370 + ], + "type": "table", + "html": "
MethodImage encoderParams.GFLoPsIoUmIoU
OccFormer [50]EfficientNetB7202.7M973.136.5013.46
\\(HTCL^†\\) [18]EfficientNetB7181.4M2622.445.5117.13
\\(SOAP^†\\)EfficientNetB785.2M705.347.2419.21
Symphonies [13]ResNet5059.3M611.941.9214.89
\\(HASSC^†\\) [37]ResNet5058.1M313.944.5814.74
\\(SGN^†\\) [26]ResNet5027.9M723.746.2115.32
\\(SOAP^†\\)ResNet5045.1M627.848.1218.80
", + "image_path": "200f57d0c04cb62225973be93071c59fddabb746bcf84eb74bdc85a4cd8e9c3f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 409 + ], + "type": "text", + "content": "classes for the SSC task and the intersection over union (IoU) for the class-agnostic scene completion (SC) task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": "Highest mIoU and IoU Scores. Tables 1 and 2 compare the proposed SOAP with recent vision-centric 3D SSC algorithms on the SemanticKITTI and SSCBench-KITTI360 test sets, respectively. The scores of other methods come from respective papers. In Table 1, we observe that SOAP significantly outperforms the existing methods on both datasets in terms of the mIoU and IoU scores. Compared to HTCL [18], SOAP improves " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "+2" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " SSC mIoU and " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "+1.86" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " SC IoU on SemanticKITTI test set, which achieves " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " relative improvement. Also, SOAP surpasses SGN [26] on SSCBench-KITTI test set with a margin of " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "+2.67" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "+1.06" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " IoU. The highest mIoU and IoU scores demonstrate that the proposed OAP faithfully refines occluded regions in the initial voxel features, while the scene-adaptive decoder further enhances the semantic context of the query embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Efficiency Evaluation. We replace our image encoder, EfficientNetB7 [35] with a more light-weight network, ResNet50 [9]. Table 3 shows complexity-accuracy comparisons between SOAP and other methods on SemanticKITTI validation set, using either ResNet50 or EfficientNetB7 as the image encoder. Compared to OccFormer [50] and HTCL [18], which utilize EfficientNetB7, SOAP requires the lowest parameters and GFLoPs while achieving the highest IoU and mIoU scores, demonstrating the effective" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17151" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 76, + 192, + 115 + ], + "blocks": [ + { + "bbox": [ + 66, + 76, + 192, + 115 + ], + "lines": [ + { + "bbox": [ + 66, + 76, + 192, + 115 + ], + "spans": [ + { + "bbox": [ + 66, + 76, + 192, + 115 + ], + "type": "image", + "image_path": "7a6c310b69ba4b07afd41a78d8853b315a4fe221721e9b9e250d0751afde0aa7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 117, + 150, + 125 + ], + "lines": [ + { + "bbox": [ + 108, + 117, + 150, + 125 + ], + "spans": [ + { + "bbox": [ + 108, + 117, + 150, + 125 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 76, + 289, + 121 + ], + "blocks": [ + { + "bbox": [ + 212, + 76, + 289, + 121 + ], + "lines": [ + { + "bbox": [ + 212, + 76, + 289, + 121 + ], + "spans": [ + { + "bbox": [ + 212, + 76, + 289, + 121 + ], + "type": "image", + "image_path": "a2e4947b1cc23e631baa4df875c9a47c6d4343032fb0ce69000cabc4a2a5c9d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 66, + 133, + 169, + 183 + ], + "blocks": [ + { + "bbox": [ + 66, + 133, + 169, + 183 + ], + "lines": [ + { + "bbox": [ + 66, + 133, + 169, + 183 + ], + "spans": [ + { + "bbox": [ + 66, + 133, + 169, + 183 + ], + "type": "image", + "image_path": "a9af1d35189661f00a4ad07c3d0207d1cddc88b2d771213721ecfd10ba461bf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 186, + 111, + 195 + ], + "lines": [ + { + "bbox": [ + 89, + 186, + 111, + 195 + ], + "spans": [ + { + "bbox": [ + 89, + 186, + 111, + 195 + ], + "type": "text", + "content": "SOAP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 133, + 289, + 183 + ], + "blocks": [ + { + "bbox": [ + 228, + 123, + 273, + 131 + ], + "lines": [ + { + "bbox": [ + 228, + 123, + 273, + 131 + ], + "spans": [ + { + "bbox": [ + 228, + 123, + 273, + 131 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 133, + 289, + 183 + ], + "lines": [ + { + "bbox": [ + 179, + 133, + 289, + 183 + ], + "spans": [ + { + "bbox": [ + 179, + 133, + 289, + 183 + ], + "type": "image", + "image_path": "b7f0a041533b72f20816e7c8c70b74d07933bdf3ffde4491fd20751154bd6888.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 186, + 248, + 196 + ], + "lines": [ + { + "bbox": [ + 192, + 186, + 248, + 196 + ], + "spans": [ + { + "bbox": [ + 192, + 186, + 248, + 196 + ], + "type": "text", + "content": "OccFormer [50]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 67, + 213, + 192, + 251 + ], + "blocks": [ + { + "bbox": [ + 67, + 213, + 192, + 251 + ], + "lines": [ + { + "bbox": [ + 67, + 213, + 192, + 251 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 192, + 251 + ], + "type": "image", + "image_path": "672e54184cef91cf9038fa9238727e3050da32c3d2dfd2391e11477d83124908.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 253, + 150, + 262 + ], + "lines": [ + { + "bbox": [ + 108, + 253, + 150, + 262 + ], + "spans": [ + { + "bbox": [ + 108, + 253, + 150, + 262 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 211, + 294, + 256 + ], + "blocks": [ + { + "bbox": [ + 206, + 211, + 294, + 256 + ], + "lines": [ + { + "bbox": [ + 206, + 211, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 206, + 211, + 294, + 256 + ], + "type": "image", + "image_path": "da67e55e10c275427580c393eb7010a343cb0870758cb80232323493a27fe7c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 270, + 174, + 319 + ], + "blocks": [ + { + "bbox": [ + 58, + 270, + 174, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 270, + 174, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 174, + 319 + ], + "type": "image", + "image_path": "ac4d0833fd9aa0db7634d5d704aa87f7780629619e6d10af30e3bbb0cf1df382.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 323, + 111, + 331 + ], + "lines": [ + { + "bbox": [ + 89, + 323, + 111, + 331 + ], + "spans": [ + { + "bbox": [ + 89, + 323, + 111, + 331 + ], + "type": "text", + "content": "SOAP" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 335, + 277, + 344 + ], + "lines": [ + { + "bbox": [ + 75, + 335, + 277, + 344 + ], + "spans": [ + { + "bbox": [ + 75, + 335, + 277, + 344 + ], + "type": "text", + "content": "car truck road parking sidewalk building fence" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 362, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 407 + ], + "type": "text", + "content": "Figure 5. Qualitative results on SemanticKITTI validation set. SOAP demonstrates robustness in handling the occluded regions (top) and distinguishing similar classes (bottom). Red boxes highlight failure cases of OccFormer [50] for comparison." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 182, + 270, + 294, + 319 + ], + "blocks": [ + { + "bbox": [ + 228, + 259, + 273, + 268 + ], + "lines": [ + { + "bbox": [ + 228, + 259, + 273, + 268 + ], + "spans": [ + { + "bbox": [ + 228, + 259, + 273, + 268 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 270, + 294, + 319 + ], + "lines": [ + { + "bbox": [ + 182, + 270, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 182, + 270, + 294, + 319 + ], + "type": "image", + "image_path": "2afc38abe44aeeaf7f9c82234d7e9b83ee8441a6fae2b879dfd995b7ea6905f2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 323, + 248, + 332 + ], + "lines": [ + { + "bbox": [ + 192, + 323, + 248, + 332 + ], + "spans": [ + { + "bbox": [ + 192, + 323, + 248, + 332 + ], + "type": "text", + "content": "OccFormer [50]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 345, + 258, + 354 + ], + "lines": [ + { + "bbox": [ + 94, + 345, + 258, + 354 + ], + "spans": [ + { + "bbox": [ + 94, + 345, + 258, + 354 + ], + "type": "text", + "content": "vegetation trunk terrain pole traffic sign" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 430, + 295, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 466 + ], + "type": "text", + "content": "ness and efficiency of SOAP. Also, SOAP with ResNet50 yields remarkable mIoU and IoU performance with comparable complexity compared to other methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 468, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 586 + ], + "type": "text", + "content": "Qualitative Results. Figure 5 presents qualitative comparisons between SOAP and OccFormer [50] in terms of occluded region reconstruction and similar class discrimination. In Figure 5 (top), OccFormer fails to recognize occluded regions behind car objects, resulting in fused car shapes, whereas SOAP appropriately distinguishes and separates the cars that are actually apart. Also, as in Figure 5 (bottom), Occformer confuses classes with similar appearance (road " + }, + { + "bbox": [ + 55, + 468, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 468, + 295, + 586 + ], + "type": "text", + "content": " sidewalk) while SOAP reasonably differentiates these classes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 599, + 146, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 146, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 146, + 612 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We analyze the contributions of individual components of the proposed SOAP through ablation experiments on the SemanticKITTI validation set. Table 4 reports the contributions of the occluded region-aware projection (OAP) and the scene-adaptive decoder (SAD). For the results without SAD, we replace it with a standard transformer decoder [6, 50, 51]. For the results without OAP, we input the initial voxel features " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " directly into the voxel" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 315, + 107, + 553, + 177 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "type": "text", + "content": "Table 4. Ablation study on different combinations of occluded region-aware projection (OAP) and scene-adaptive decoder (SAD)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 107, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 177 + ], + "type": "table", + "html": "
OAPSADSC IoUSSC mIoU
43.3216.86
✓43.5117.22
✓46.5018.99
✓✓47.2419.21
", + "image_path": "936df77bb68d12bb8bdbc12d79f55f7a4e70b1b8a2dc5c03de00bf0f713c07c5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 315, + 201, + 553, + 260 + ], + "blocks": [ + { + "bbox": [ + 328, + 188, + 539, + 198 + ], + "lines": [ + { + "bbox": [ + 328, + 188, + 539, + 198 + ], + "spans": [ + { + "bbox": [ + 328, + 188, + 539, + 198 + ], + "type": "text", + "content": "Table 5. Alternative approaches for historical information." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 201, + 553, + 260 + ], + "lines": [ + { + "bbox": [ + 315, + 201, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 315, + 201, + 553, + 260 + ], + "type": "table", + "html": "
MethodSC IoUSSC mIoU
Addition45.0018.05
DCA45.9018.53
OAP47.2419.21
", + "image_path": "12a9b73fbd0b3cf4f029fa20d58ea075971619b15eb44a8372492463d52dadb1.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 274, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 553, + 380 + ], + "type": "text", + "content": "encoder, without using any historical information. In Table 4, IoU and mIoU scores show significant degradation without either OAP or SAD. This indicates that OAP effectively refines occluded regions during its voxel generation, while SAD provides representative semantic features across dynamic driving scenes for query embeddings. Notably, SOAP still retains reasonable performance even without OAP, which implements 3D SSC without temporal information." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": "To validate the effectiveness of OAP, we design several alternative approaches to utilize historical information. First, we simply add the historical voxel features " + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{hist}}" + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": " to the initial voxel features " + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": ". Second, we employ the standard deformable cross-attention (DCA) [53] between the initial voxel features " + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{F}}_i^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": " and the historical voxel features " + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{\\mathrm{hist}}" + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": " without occluded region separation. In Table 5, the proposed OAP achieves better IoU and mIoU scores than the standard DCA, with a margin of " + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "inline_equation", + "content": "+1.34" + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "inline_equation", + "content": "+0.68" + }, + { + "bbox": [ + 313, + 381, + 554, + 525 + ], + "type": "text", + "content": ", respectively. Therefore, perception of occluded regions during view projection is essential for accurate 3D semantic scene reasoning." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 531, + 388, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 388, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 388, + 542 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "We proposed a novel 3D semantic scene completion algorithm, called SOAP. The main contributions of SOAP are the occluded region-aware view projection and the scene-adaptive decoder. The occluded region-aware view projection addresses the information deficiency in vision-centric 3D SSC by refining occluded voxel regions with historical observation. The scene-adaptive decoder creates the comprehensive semantic repository using the repository builder. Then, the scene-adaptive decoder integrates informative clues from the semantic repository into query embeddings. Experiments demonstrated that SOAP significantly outperforms the state-of-the-arts on the SemanticKITTI and SSC-Bench datasets, with a margin of " + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "inline_equation", + "content": "+2" + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "inline_equation", + "content": "+2.67" + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": " mIoU scores, respectively." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17152" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 191 + ], + "type": "text", + "content": "This work was partly by the National Research Foundation of Korea (NRF) grants funded by the Ministry of Education (RS-2022-NR075739) and the Korea government (MSIT) (No. RS-2024-00352566, No. RS-2025-00559165), and partly by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155857, Artificial Intelligence Convergence Innovation Human Resources Development (Chungnam National University))." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 215, + 115, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 215, + 115, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 215, + 115, + 227 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 234, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 234, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 234, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 234, + 296, + 280 + ], + "type": "text", + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, pages 9297-9307, 2019. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 281, + 296, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 296, + 313 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 316, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 316, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 61, + 316, + 296, + 348 + ], + "type": "text", + "content": "[3] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 296, + 392 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 395, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 438 + ], + "type": "text", + "content": "[5] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In CVPR, pages 4193-4202, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 485 + ], + "type": "text", + "content": "[6] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 529 + ], + "type": "text", + "content": "[7] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In CoRL, pages 2148-2161, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "text", + "content": "[8] Emeç Erçelik, Ekim Yurtsever, Mingyu Liu, Zhijie Yang, Hanzhen Zhang, Pinar Topçam, Maximilian Listl, Yilmaz Kaan Cayli, and Alois Knoll. 3d object detection with a self-supervised lidar scene flow backbone. In ECCV, pages 247-265, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 589, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 296, + 622 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 296, + 667 + ], + "type": "text", + "content": "[10] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[11] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In CVPR, pages 4012-4021, 2022. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "text", + "content": "[13] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In CVPR, pages 20258-20267, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 172, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 554, + 217 + ], + "type": "text", + "content": "[14] Lei Ke, Shichao Li, Yanan Sun, Yu-Wing Tai, and Chi-Keung Tang. Gsnet: Joint vehicle pose and shape reconstruction with geometrical and scene-aware supervision. In ECCV, pages 515–532, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 554, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 248 + ], + "type": "text", + "content": "[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 282 + ], + "type": "text", + "content": "[16] Abhijit Kundu, Yin Li, and James M Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render-and-compare. In CVPR, pages 3559–3568, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 554, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 336 + ], + "type": "text", + "content": "[17] Hyo-Jun Lee, Hanul Kim, Su-Min Choi, Seong-Gyun Jeong, and Yeong Jun Koh. Baam: Monocular 3d pose and shape reconstruction with bi-contextual attention module and attention-guided modeling. In CVPR, pages 9011-9020, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 554, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 382 + ], + "type": "text", + "content": "[18] Bohan Li, Jiajun Deng, Wenyao Zhang, Zhujin Liang, Dalong Du, Xin Jin, and Wenjun Zeng. Hierarchical temporal context learning for camera-based semantic scene completion. In ECCV, pages 131-148. Springer, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[19] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In CVPR, pages 3351-3359, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "text", + "content": "[20] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In AAAI, pages 1477-1485, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 525 + ], + "type": "text", + "content": "[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. arXiv preprint arXiv:2306.09001, 2023. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 525, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 554, + 581 + ], + "type": "text", + "content": "[22] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, pages 9087-9098, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 647 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "text", + "content": "[25] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In ICCV, pages 3111-3121, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[26] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Jongwon Ra, Yukai Ma, Laijian Li, and Yong Liu. Camera-based" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17153" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "3d semantic scene completion with sparse guidance network. IEEE TIP, 2024. 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[27] Chao Ning and Hongping Gan. Trap attention: Monocular depth estimation with manual traps. In CVPR, pages 5033-5043, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "text", + "content": "[28] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210, 2020. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 57, + 162, + 294, + 195 + ], + "type": "text", + "content": "[29] Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. idisc: Internal discretization for monocular depth estimation. In CVPR, pages 21477-21487, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 196, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 237 + ], + "type": "text", + "content": "[30] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. IEEE TPAMI, 44(10): 7205-7218, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "text", + "content": "[31] Luis Roldao, Raoul de Charette, and Anne Verroust-Blondet. Lmscnet: Lightweight multiscale 3d semantic completion. In 3DV, pages 111–119. IEEE, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 273, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 294, + 315 + ], + "type": "text", + "content": "[32] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In WACV, pages 2417-2426, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[33] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[34] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[35] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[36] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Not all voxels are equal: Semantic scene completion from the point-voxel perspective. In AAAI, pages 2352–2360, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[37] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[38] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[39] Zhaoyang Xia, Youquan Liu, Xin Li, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, and Yu Qiao. Scpnet: Semantic scene completion on point cloud. In CVPR, pages 17642-17651, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[40] Jinfeng Xu, Xianzhi Li, Yuan Tang, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Casfusionnet: A cascaded network for point cloud semantic scene completion by dense feature fusion. In AAAI, pages 3018-3026, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[41] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, pages 3101-3109, 2021. 1, 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 644 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[42] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9421-9431, 2023. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "text", + "content": "[43] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[44] Zhu Yu, Running Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Siyuan Cao, and Huiliang Shen. Context and geometry aware voxel transformer for semantic scene completion. In NeurIPS, 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "text", + "content": "[45] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3916-3925, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[46] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In ICCV, pages 9155-9166, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[47] Shoulong Zhang, Shuai Li, Aimin Hao, and Hong Qin. Point cloud semantic scene completion from rgb-d images. In AAAI, pages 3385-3393, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[48] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Dalong Du, Jie Zhou, and Jiwen Lu. Dimension embeddings for monocular 3d object detection. In CVPR, pages 1589-1598, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "text", + "content": "[49] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[50] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In ICCV, pages 9433-9443, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "text", + "content": "[51] Linqing Zhao, Xiuwei Xu, Ziwei Wang, Yunpeng Zhang, Borui Zhang, Wenzhao Zheng, Dalong Du, Jie Zhou, and Jiwen Lu. Lowrankocc: Tensor decomposition and low-rank recovery for vision-based 3d semantic occupancy prediction. In CVPR, pages 9806-9815, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 522, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 564 + ], + "type": "text", + "content": "[52] Yunsong Zhou, Hongzi Zhu, Quan Liu, Shan Chang, and Minyi Guo. Monoatt: Online monocular 3d object detection with adaptive token transformer. In CVPR, pages 17493-17503, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "text", + "content": "[53] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 5, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 644 + ], + "type": "text", + "content": "[54] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, pages 9939-9948, 2021. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17154" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_content_list.json b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e466c81c1c7e4e8f92839b37c47ea31d8a66524f --- /dev/null +++ b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_content_list.json @@ -0,0 +1,1905 @@ +[ + { + "type": "text", + "text": "SOGS: Second-Order Anchor for Advanced 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 158, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahui Zhang $^{1}$ Fangneng Zhan $^{2,3}$ Ling Shao $^{4}$ Shijian Lu $^{1}$", + "bbox": [ + 243, + 179, + 750, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nanyang Technological University 2Harvard University MIT", + "bbox": [ + 243, + 198, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4UCAS-Terminus AI Lab, University of Chinese Academy of Sciences", + "bbox": [ + 222, + 217, + 779, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiahui003@e.ntu.edu.sg fnzhan@seasharvard.edu", + "bbox": [ + 294, + 237, + 710, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ling.shao@ieee.org shijian.lu@ntu.edu.sg", + "bbox": [ + 313, + 255, + 676, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7f57ee65b0bd04d32da9829fb266809b65b70d3b3762ab5998ddc08e2a32997.jpg", + "image_caption": [ + "Rendered Images" + ], + "image_footnote": [], + "bbox": [ + 93, + 321, + 292, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4a5b0d7d869538bad3dcaa45b21349fa5b8d86897523387b6860fdf5d4b4950.jpg", + "image_caption": [ + "Gradient Maps" + ], + "image_footnote": [], + "bbox": [ + 294, + 321, + 493, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/323fb34a19344feeb6abf1d3bbf31d0b3b9bd8d4662d3cc0908bfec89b75f02e.jpg", + "image_caption": [ + "Rendered Images" + ], + "image_footnote": [], + "bbox": [ + 503, + 321, + 700, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2358e0fade5230fc7f40e4e0f511c059525c51757329926861deef1041edae4.jpg", + "image_caption": [ + "Gradient Maps" + ], + "image_footnote": [], + "bbox": [ + 705, + 321, + 903, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07831be65495cb6054194bbcea4b71fac52e4e152eb5529631addaae611073d3.jpg", + "image_caption": [ + "(a)Scaffold-GS" + ], + "image_footnote": [], + "bbox": [ + 93, + 425, + 292, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1cb42ed6dc80f7c1c643dc50d1e230a45b2763600f0e2fe06ccde40aec20736b.jpg", + "image_caption": [ + "Figure 1. The proposed SOGS can render high-quality textures and geometries and reduce model size simultaneously. The illustrations in (a) and (b) show Rendered Images and Gradient Maps which are produced by Scaffold-GS [25] and SOGS, respectively, for the samples 'Stump' and 'Room' from Mip-NeRF360 [2]. The Gradient Maps are extracted with the Sobel operator which highlight the texture and geometry of imaged scenes." + ], + "image_footnote": [], + "bbox": [ + 292, + 425, + 493, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e132f0cb814e5b2b047bf791228eeb1c1b93f0963b73449acc06e76dc3d087a.jpg", + "image_caption": [ + "(b) SOGS" + ], + "image_footnote": [], + "bbox": [ + 503, + 425, + 700, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 625, + 326, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anchor-based 3D Gaussian splatting (3D-GS) exploits anchor features in 3D Gaussian prediction, which has achieved impressive 3D rendering quality with reduced Gaussian redundancy. On the other hand, it often encounters the dilemma among anchor features, model size, and rendering quality - large anchor features lead to large 3D models and high-quality rendering whereas reducing anchor features degrades Gaussian attribute prediction which leads to clear artifacts in the rendered textures and geometries. We design SOGS, an anchor-based 3D-GS technique that introduces second-order anchors to achieve superior rendering quality and reduced anchor features and model size simultaneously. Specifically, SOGS incorporates covariance-based second-order statistics and corre", + "bbox": [ + 88, + 659, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lation across feature dimensions to augment features within each anchor, compensating for the reduced feature size and improving rendering quality effectively. In addition, it introduces a selective gradient loss to enhance the optimization of scene textures and scene geometries, leading to high-quality rendering with small anchor features. Extensive experiments over multiple widely adopted benchmarks show that SOGS achieves superior rendering quality in novel view synthesis with clearly reduced model size.", + "bbox": [ + 511, + 627, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 797, + 645, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D scene representation and rendering have been a pivotal task in 3D computer vision, playing a crucial role in various applications such as virtual reality and scene simulation. Neural radiance fields (NeRFs) [27] and its variants [1-3, 46] have been developed to model implicit", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Shijian Lu is the corresponding author.", + "bbox": [ + 109, + 886, + 321, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11167", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene representation and enable volume rendering for high-quality novel view synthesis. However, NeRF is limited by long training and rendering times, largely due to the time-consuming point sampling in volume rendering. Recently, 3D Gaussian Splatting (3D-GS) [18] has been proposed as an alternative which learns explicit 3D scene representations with learnable 3D Gaussians. By leveraging efficient splatting and rasterization, 3D-GS projects 3D Gaussians onto a 2D plane, enabling real-time rendering. However, 3D-GS stacks substantial 3D Gaussians to fit each training view individually, often overlooking the underlying scene structure [25].", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent Scaffold-GS [25] exploits scene geometry to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. Specifically, it introduces anchor points to store features and exploits MLPs to predict Gaussian attributes from these anchor features. However, Scaffold-GS struggles to strike a balance between rendering quality and model size. Specifically, adopting large anchor features improves rendering quality [9] but it significantly enlarges model size due to the large number of anchor points used in scene representations. Using smaller anchor features can shrink the model, but it clearly degrades the prediction of Gaussian attributes which further leads to suboptimal rendering with various artifacts in rendered textures and geometries. HAC [9] introduces grid-based context modeling and entropy encoding to compress the Scaffold-GS. However, HAC does not reduce the model size but the storage size of the trained Scaffold-GS.", + "bbox": [ + 89, + 275, + 482, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design SOGS, an innovative anchor-based 3D-GS technique that introduces second-order anchors and achieves superior rendering quality with reduced anchor feature dimensions and model size. Specifically, SOGS incorporates covariance-based second-order statistics to model correlation across anchor feature dimensions. The rationale is that textures and geometries are defined by not only individual features but also the correlation across features. Hence, the correlation across anchor features can effectively guide to capture the pattern of intricate local textures and structures in each anchor. This nice feature enables anchor feature augmentation, empowering SOGS to achieve superior rendering quality with reduced feature dimensions and feature size.", + "bbox": [ + 89, + 550, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we design a selective gradient loss to enhance the rendering quality with small anchor features. This loss computes gradient maps from both the rendered image and the corresponding ground truth that highlight their differences in scene textures and structures. With the difference of the gradient maps, the 3D Gaussian prediction can adaptively focus on image regions where the texture and geometry are hard to render. This enables dynamic region selection along the training process, where the model can", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adjust the focused image regions dynamically according to the evolving errors between the gradient maps of the rendered image and the ground truth. Experiments show that SOGS simultaneously achieves superior rendering quality in novel view synthesis as well as clearly reduced model sizes as shown in Fig. 1.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The major contributions of this work can be summarized in three aspects. First, we propose SOGS, an innovative anchor-based 3D-GS that introduces second-order anchors to balance the rendering quality and model size. With covariance-based second-order statistics that model correlation across anchor feature dimensions to capture intricate textural and structural patterns, SOGS achieves anchor feature augmentation which produces superior rendering quality with significantly reduced anchor feature dimensions. Second, we design a selective gradient loss that guides 3D Gaussian prediction to focus on difficult-to-render textures and structures adaptively, further enhancing rendering quality with small anchor features Third, experiments over multiple benchmarks show that SOGS achieves superior novel view synthesis with efficient anchor size and model size, and outperforms the state-of-the-art in rendering quality.", + "bbox": [ + 511, + 183, + 903, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 441, + 653, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Radiance Field and Neural Rendering", + "text_level": 1, + "bbox": [ + 511, + 468, + 841, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Radiance field has been widely explored for novel view synthesis and achieved impressive rendering quality. Neural radiance field (NeRF) [27] utilizes MLPs to learn implicit 3D scene representations from multi-view posed 2D images. Coupled with differentiable volume rendering, NeRF achieves high-quality rendering with superb multiview consistency. Several NeRF variants have been developed to address various new challenges, including largescale scenes [32, 33, 37], dynamic scenes [12, 14, 31], few-shot setting [6, 34, 40], pose-free setting [4, 8, 22, 43, 44], and antialIASing [2, 3]. However, NeRF and its variants still suffer from extremely long training and rendering times due to the time-consuming sampling required for volume rendering. Several studies [7, 15, 28, 29] have been proposed to address this issue. For instance, Müller et al. [28] introduce multi-resolution hash encoding to reduce the size of neural networks, significantly shortening the training time and enabling real-time rendering in tens of milliseconds. However, these methods often suffer from degraded reconstruction accuracy and rendering quality of scenes.", + "bbox": [ + 511, + 491, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Kerbl et al. recently propose 3D Gaussian splatting (3D-GS) [18], which achieves high-quality and real-time rendering by explicitly representing scenes by using parameterized 3D Gaussians, along with efficient splatting and rasterization. Building on its excellent performance, numerous 3D-GS variants have been developed to address diverse challenges and tasks, such as dynamic scenes [21, 23, 26,", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11168", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "36, 41], autonomous driving [38, 47], rendering quality optimization [11, 39, 42, 45], sparse-view setting [5, 10, 20], colmap-free setting [13], and 3D style transfer [24].", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Anchor-based 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 148, + 410, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One major constraint in 3D-GS is that it stacks a large number of 3D Gaussians to match each training view separately, often overlooking the underlying scene geometry. To address this issue, Lu et al. [25] present Scaffold-GS, the first anchor-based 3D-GS technique that incorporates anchor points to predict Gaussian attributes through multilayer perceptrons (MLPs). Scaffold-GS leverages scene structure to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. However, Scaffold-GS faces challenges in balancing enhanced rendering quality with anchor and model size reduction. Although several follow-ups [9, 35] introduce context-based entropy encoding to compress the Scaffold-GS, they reduce its storage size instead of anchor and model sizes as used during training and rendering. The proposed SOGS introduces second-order anchors, enabling superior rendering quality with reduced model size.", + "bbox": [ + 89, + 171, + 483, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 445, + 264, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose SOGS, a novel anchor-based 3D Gaussian splatting that introduces second-order anchors and selective gradient loss to achieve superior rendering quality with reduced anchor and model sizes. Fig. 2 shows the overview of SOGS. The 3D Gaussian splatting [18] (3D-GS) and Scaffold-GS [25] (the base model for our SOGS) are briefly introduced in Sec. 3.1. In Sec. 3.2, we first explain why the proposed second-order anchor enables the model to enhance rendering quality in reduced anchor and model sizes, and then describe the details of the second-order anchor. Besides, to further ensure high-quality image rendering under compact anchor size, we design a selective gradient loss, detailed in Sec. 3.3.", + "bbox": [ + 89, + 470, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 680, + 230, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3D-GS). 3D-GS [18] explicitly represents scenes using anisotropic 3D Gaussians and enables real-time rendering through efficient differentiable splatting. 3D Gaussians are initialized from point clouds generated by structure-from-motion [16, 30]. Each Gaussian is defined by a covariance matrix, a position (mean), an opacity value $\\alpha$ , and spherical harmonics coefficients representing color $c$ , where the covariance matrix is decomposed into scaling matrix and rotation matrix to facilitate differentiable optimization.", + "bbox": [ + 89, + 702, + 482, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rendering, 3D Gaussians are projected onto a 2D plane through splatting, followed by $\\alpha$ -blending. Specifically, the color $C$ of a pixel is computed by blending $N$ or", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dered 2D Gaussians overlapping that pixel, formulated as:", + "bbox": [ + 511, + 90, + 897, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {n \\in N} c _ {n} \\alpha_ {n} \\prod_ {z = 1} ^ {n - 1} (1 - \\alpha_ {z}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 116, + 905, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{n}$ and $\\alpha_{n}$ represent the color and opacity of the $n$ -th 2D Gaussian.", + "bbox": [ + 511, + 167, + 905, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaffold-GS. Scaffold-GS [25] introduces anchor points to predict 3D Gaussians via MLPs. It aims to leverage scene geometry to guide the distribution and attributes of 3D Gaussians and reduce the Gaussian redundancy. Specifically, Scaffold-GS voxelizes the entire scene based on the point cloud initialized by COLMAP [30] and assigns the center of each voxel as an anchor point. A local scene feature $\\pmb{f}^{\\pmb{a}} \\in \\mathbb{R}^{D}$ , a scaling factor $l^{\\pmb{a}}$ and $K$ offsets $o^{\\pmb{a}} \\in \\mathbb{R}^{3K}$ are stored in each anchor as the anchor attributes. Given an anchor position $x_{a}$ , its relative distance $\\delta_{ac}$ and viewing direction $d_{ac}$ to a camera position $x_{c}$ can be represented as: $\\delta_{ac} = ||x_{a} - x_{c}||_{2}$ , $d_{ac} = \\frac{x_{a} - x_{c}}{||x_{a} - x_{c}||_{2}}$ . The prediction of $K$ Gaussians from one anchor point using MLPs $F$ can then be formulated as follows:", + "bbox": [ + 511, + 215, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \\right\\} _ {k = 1} ^ {K} = F (\\boldsymbol {f} ^ {\\boldsymbol {a}}, \\boldsymbol {\\delta} _ {a c}, \\boldsymbol {d} _ {a c}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 439, + 906, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{k}, c_{k}, s_{k}$ and $q_{k}$ denote the opacity, color, scaling and the rotation-related quaternion of the $k$ -th Gaussian, respectively. The positions of the $K$ Gaussians are calculated by this formula:", + "bbox": [ + 511, + 468, + 905, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{p _ {k} \\right\\} _ {k = 1} ^ {K} = \\boldsymbol {x} _ {a} + \\left\\{o _ {k} ^ {a} \\right\\} _ {k = 1} ^ {K} * \\boldsymbol {l} ^ {\\boldsymbol {a}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 537, + 905, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaffold-GS follows the same rendering process as Eq. 1 described in 3D-GS after predicting the Gaussian attributes.", + "bbox": [ + 511, + 566, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Second-Order Anchor", + "text_level": 1, + "bbox": [ + 511, + 606, + 720, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first explain why the proposed second-order anchor enables the model to achieve superior image rendering with compact anchor size. The key lies in leveraging covariance-based second-order feature statistics to achieve anchor feature augmentation. Specifically, second-order feature statistics extract co-varying relationships across anchor feature dimensions. As textures and structures are defined not only by individual features but also by their interdependencies, these co-varying relationships can guide the capture of intricate local textural and structural patterns within each anchor, facilitating feature augmentation. As a result, SOGS compensates for the reduced feature dimensions and ensures superior rendering quality by performing feature augmentation. We provide the visualization of the impact of second-order anchors on scene textures and structures in the Sec. 4.5.", + "bbox": [ + 511, + 628, + 906, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given anchor feature set $\\mathbf{F}^{\\mathbf{a}}\\in \\mathbb{R}^{N\\times D} = [f_1^a,f_2^a,\\dots,f_N^a ]$ , where $N$ and $D$ denote the number of", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11169", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d42c671027edd64f2e7d10d34cf3fc38412f8c15b1db14e2ab326a84f8e69db9.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed SOGS. Initialized from point clouds, each anchor stores an anchor feature vector $f^a \\in \\mathbb{R}^D$ . Putting $f^a$ along $D$ dimensions as $D$ variables and all anchors as observed samples, the second-order anchor statistics capturing covarying relations across $D$ can be computed from the covariance matrix $\\Sigma$ across $D$ dimensions. We select the top $M$ eigenvectors $P$ as the most significant co-variation patterns and combine $P$ with $f^a$ to capture anchor-specific textures and geometries for anchor feature augmentation, second-order anchor construction and Gaussian prediction. With the gradient maps of the rendered image and the ground truth, the selective gradient loss guides to learn to render finer textures and geometries with dynamic region selection." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "anchors and the dimensionality of each anchor feature, we first compute second-order statistics in the anchor feature set. Specifically, we treat $D$ feature dimensions (channels) as $D$ variables and utilize the anchor set as $N$ observation samples, and compute the covariance matrix $\\Sigma \\in \\mathbb{R}^{D\\times D}$ across these channels, which can be formulated as:", + "bbox": [ + 89, + 424, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\frac {1}{N - 1} \\left(\\boldsymbol {F} ^ {\\boldsymbol {a}} - \\boldsymbol {\\mu} ^ {T}\\right) \\left(\\boldsymbol {F} ^ {\\boldsymbol {a}} - \\boldsymbol {\\mu} ^ {T}\\right) ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 523, + 482, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} = \\frac {1}{N} \\boldsymbol {F} ^ {\\boldsymbol {a}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 566, + 482, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\mu} \\in \\mathbb{R}^{D}$ denotes the mean vector across $N$ samples, used for feature centering. Given that two variables with large variances can yield a large covariance despite only weak relationship, we then construct a correlation matrix $\\pmb{R}$ from the covariance matrix $\\pmb{\\Sigma}$ to standardize the relationships among $D$ variables, eliminating the influence of different scales or variances:", + "bbox": [ + 89, + 601, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {R} = \\boldsymbol {A} ^ {- 1} \\boldsymbol {\\Sigma} \\boldsymbol {A} ^ {- 1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 718, + 482, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} = \\operatorname {d i a g} \\left(\\sigma_ {1}, \\dots , \\sigma_ {u}, \\dots , \\sigma_ {D}\\right), \\quad \\sigma_ {u} = \\sqrt {\\Sigma_ {u u}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 747, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ is a diagonal matrix and the diagonal elements are the standard deviation of $D$ variables. The correlation between $u$ -th and $v$ -th variables can then be expressed as:", + "bbox": [ + 89, + 772, + 483, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {u v} = \\frac {\\Sigma_ {u v}}{\\sigma_ {u} \\sigma_ {v}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 828, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Sigma_{uv}$ represents the covariance between the $u$ -th and $v$ -th variables. With $\\pmb{R}$ , we extract $M$ principle co-varying", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relationships across anchor feature dimensions. Specifically, we perform eigendecomposition and decompose the matrix $\\pmb{R}$ into the eigenvector matrix $Q\\in \\mathbb{R}^{D\\times D}$ and the diagonal matrix $\\Lambda \\in \\mathbb{R}^{D\\times D}$ of eigenvalues:", + "bbox": [ + 511, + 424, + 905, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {R} = \\boldsymbol {Q} \\boldsymbol {\\Lambda} \\boldsymbol {Q} ^ {T}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 489, + 903, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ is an orthogonal matrix and each column is an eigenvector representing a co-variation pattern across anchor feature dimensions. We sort the eigenvalues in descending order, where the eigenvector corresponding to the largest eigenvalue captures the most significant co-variation patterns across the feature dimensions. Following this, we select the top- $M$ eigenvectors $\\pmb{P} = [P_{1},\\dots,P_{M}]$ from $Q$ as the primary directions of co-variation.", + "bbox": [ + 511, + 513, + 905, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then leverage $P$ to guide the extraction of textural and structural information in each anchor. As the covarying relationships across anchor feature dimensions are extracted from the anchor set $F^a$ , we treat $P$ as the global-shared co-variations. Combined with the feature $f^a$ in one anchor, we can capture the anchor-specific textures and structures as follows:", + "bbox": [ + 511, + 636, + 905, + 739 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} ^ {\\boldsymbol {t}} = F _ {i} \\left(\\left[ \\boldsymbol {P} _ {i}, \\boldsymbol {f} ^ {\\boldsymbol {a}} \\right]\\right), \\quad i \\in [ 1, M ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 747, + 903, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{i}(\\cdot)$ represents the two-layer MLP used to extract the textures and structures for anchor $\\pmb{f}^{a}$ based on the $i$ -th principal co-varying relationship across anchor feature dimensions. Finally, we concatenate the extracted features with the original anchor feature to achieve feature augmentation and predict Gaussian attributes. Take one second-order anchor as an example:", + "bbox": [ + 511, + 771, + 905, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \\right\\} _ {k = 0} ^ {K} = F \\left(\\boldsymbol {f} ^ {\\boldsymbol {a}}, \\left\\{\\boldsymbol {f} _ {\\boldsymbol {i}} ^ {\\boldsymbol {t}} \\right\\} _ {i = 0} ^ {M}, \\boldsymbol {\\delta} _ {a c}, \\boldsymbol {d} _ {a c}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11170", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/73fe3830d156d04d9e113b7695e3479fa11d528ef7f36217c111918be00b3845.jpg", + "image_caption": [ + "Figure 3. Visual illustration of the proposed selective gradient loss (SGL). SGL clearly improves the rendering quality by generating finer textures and details. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\{\\pmb{f}_i^t\\}_{i=0}^M$ are derived from $\\pmb{f}^a$ for feature augmentation and do not increase the anchor feature dimensions and the size of each anchor.", + "bbox": [ + 89, + 397, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Selective Gradient Loss", + "text_level": 1, + "bbox": [ + 89, + 454, + 308, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to proposing second-order anchors for anchor feature augmentation to achieve superior rendering quality with reduced anchor size, we design a selective gradient loss to further ensure the rendering quality under compact anchor features. It enables the model to adaptively identify difficult-to-render textures and structures, and regularize Gaussian prediction corresponding to these regions. With the designed loss, superior rendering quality with finer textures can be achieved as shown in Fig 3. Given the limited sensitivity of pixel-level L1 loss to textures and structures in RGB images, we consider gradient map to emphasize textural and structural regions. Specifically, we exploit the Sobel operator to extract horizontal and vertical gradient maps $(G_x', G_y', G_x, G_y)$ of the rendering result $I' \\in \\mathbb{R}^{H \\times W \\times C}$ and its corresponding ground truth $I$ , which can be expressed as follows:", + "bbox": [ + 89, + 477, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {x} ^ {\\prime} = S _ {x} * I ^ {\\prime}, \\quad G _ {y} ^ {\\prime} = S _ {y} * I ^ {\\prime}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 732, + 482, + 757 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG _ {x} = S _ {x} * I, \\quad G _ {y} = S _ {y} * I,\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 755, + 392, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {x} = \\left[ \\begin{array}{l l l} - 1 & 0 & 1 \\\\ - 2 & 0 & 2 \\\\ - 1 & 0 & 1 \\end{array} \\right], \\quad S _ {y} = \\left[ \\begin{array}{l l l} - 1 & - 2 & - 1 \\\\ 0 & 0 & 0 \\\\ 1 & 2 & 1 \\end{array} \\right], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 784, + 483, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{x}$ and $S_{y}$ denote the horizontal and vertical Sobel kernels, respectively. We then use Euclidean metric to measure the horizontal and vertical gradient discrepancies $(l_{x}, l_{y})$ between the rendering $I'$ and the corresponding ground", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "truth $I$ , which are then averaged to derive the final discrepancies as follows:", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {x} = \\frac {1}{\\sqrt {H W}} \\sum_ {\\substack {x = 0 \\\\ H = 1}} ^ {H - 1} \\sum_ {\\substack {y = 0 \\\\ W = 1}} ^ {W - 1} \\left| G _ {x} ^ {\\prime} (x, y) - G _ {x} (x, y) \\right|, \\tag{14}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 128, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {y} = \\frac {1}{\\sqrt {H W}} \\sum_ {x = 0} ^ {H - 1} \\sum_ {y = 0} ^ {W - 1} \\left| G _ {y} ^ {\\prime} (x, y) - G _ {y} (x, y) \\right|.\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 176, + 849, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gradient maps, derived from differences in pixel intensities, typically exhibit meaningful gradients in localized regions, such as edges and areas with significant intensity changes, which are crucial for capturing structures and textures. In contrast, the majority of areas in gradient map comprise flat and low-gradient regions with limited information. Therefore naively adopting the horizontal and vertical gradient discrepancies as the loss function can cause low-gradient regions to dominate the overall loss. As a result, the model may deemphasize regions critical for preserving fine textures and structures, focusing instead on less important flat areas. Besides, the regions with meaningful gradients should also be distinguished, prioritizing greater attention to textures and structures exhibiting larger rendering errors. To address the above issues, we introduce a dynamic region selection that computes the rendering error map as a weight map, defined by the absolute difference between the gradient maps of the rendered result and the ground truth, which can be expressed by:", + "bbox": [ + 511, + 229, + 906, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {x} = \\left| G _ {x} ^ {\\prime} (x, y) - G _ {x} (x, y) \\right|, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 527, + 903, + 553 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {y} = \\left| G _ {y} ^ {\\prime} (x, y) - G _ {y} (x, y) \\right|.\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 550, + 807, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combined with the horizontal and vertical weight maps, the selective gradient loss $\\mathcal{L}_s$ can be formulated as:", + "bbox": [ + 511, + 577, + 905, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = w _ {x} * l _ {x} + w _ {y} * l _ {y}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 619, + 903, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above analysis, the selective gradient loss enables the model to focus on the difficult-to-render textural and structural regions adaptively. Besides, the model is able to adjust the focused regions dynamically in response to the evolving rendering error map throughout the training process. Note, the pixel-level L1 loss with D-SSIM term between RGB images is also used in the training process, which complements the proposed selective gradient loss used between gradient maps. And we also keep the volume regularization as described in Scaffold-GS [25]. The total loss function is given as follows:", + "bbox": [ + 511, + 648, + 906, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} + \\lambda_ {v o l} \\mathcal {L} _ {v o l} + \\lambda_ {s} \\mathcal {L} _ {s}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 827, + 903, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\lambda_s$ , $\\lambda_1$ , $\\lambda_{SSIM}$ and $\\lambda_{vol}$ represent the training weights of the selective gradient loss $\\mathcal{L}_s$ , the L1 loss $\\mathcal{L}_1$ , the D-SSIM term $\\mathcal{L}_{SSIM}$ and the volume regularization $\\mathcal{L}_{vol}$ .", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bf2f2ca4284921ec2ead15269ab29e2dcd09583df597ebeb7b24f9f1562728ce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMip-NeRF360Tanks&TemplesDeep Blending
MethodsSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.80627.500.25232 dim0.85323.960.17732 dim0.90630.210.25432 dim
SOGS(Ours)0.81527.850.22116 dim0.85524.140.17612 dim0.90730.290.25212 dim
", + "bbox": [ + 91, + 88, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f19a69f94c573fde3bde434e1ac806d4421ec156f805c794e3aeb98432e35890.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of novel view synthesis. All methods are trained with the same training data. SOGS achieves superior rendering quality with reduced anchor (model) size. Notably, SOGS using the feature dimension of 12 can still outperform the Scaffold-GS that uses a feature dimension of 32. We do not compare with HAC [9] and ContextGS [35], as they leverage compression techniques to reduce the storage size of Scaffold-GS rather than the actual anchor and model sizes as used in training and rendering." + ], + "table_footnote": [], + "table_body": "
DatasetsBungeeNeRF
MethodsSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.86526.620.24132 dim
SOGS(Ours)0.88027.060.17116 dim
", + "bbox": [ + 91, + 273, + 478, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons on large-scale scenes from the dataset BungeeNeRF [37]. All methods are trained with the same training data.", + "bbox": [ + 89, + 368, + 482, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 436, + 223, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 462, + 410, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets For both training and testing, we adopt the datasets and configurations in Scaffold-GS [25] and conduct experiments on images of 19 real-world scenes. Specifically, we evaluate SOGS across all nine scenes from the Mip-NeRF360 dataset [2], two scenes from the Tanks&Temples dataset [19], and two additional scenes from the Deep Blending dataset [17]. These scenes exhibit diverse styles, ranging from bounded indoor environments to unbounded outdoor ones. We also evaluate SOGS on the BungeeNeRF dataset [37], which includes six large-scale outdoor scenes, to further validate its effectiveness. The datasets are divided into training and test sets following the same settings as Scaffold-GS, with image resolutions consistent with those used in Scaffold-GS.", + "bbox": [ + 89, + 484, + 483, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation For the second-order anchor, we reduce the dimension $D$ of the anchor features in Scaffold-GS from 32 to 16, or even to 12. The covariance matrix $\\pmb{\\Sigma}$ is computed from all anchors in scene representations which captures the global correlation across all anchor feature dimensions. The number of the selected eigenvectors is set to $M = 2$ . All MLPs used to extract anchor-specific textural and structural information consist of two layers with ReLU activation. For the loss function, the training weights $\\lambda_{s}, \\lambda_{1}, \\lambda_{SSIM}$ and $\\lambda_{vol}$ are set to 0.01, 0.8, 0.2 and 0.01. Please note, for fair comparisons, we keep the loss terms $\\lambda_{1}, \\lambda_{SSIM}$ and $\\lambda_{vol}$ the same as in Scaffold-GS. The number of", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Gaussians corresponding to one second-order anchor is set as $K = 10$ . We use the Pytorch framework to implement SOGS. The SOGS model is trained for 30000 iterations, the same as Scaffold-GS and 3D-GS.", + "bbox": [ + 511, + 277, + 906, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 511, + 347, + 846, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We primarily compare SOGS with Scaffold-GS [25] over datasets Mip-NeRF360, Tank&Temple and Deep Blending as Scaffold-GS is an anchor-based 3D Gaussian splattering method and is most relevant to our study. Table 1 shows experimental results over the same test images as described in Section 4.1. We can observe that SOGS outperforms Scaffold-GS consistently in PSNR, SSIM and LPIPS across all evaluated scenes. The anchor size of SOGS is reduced to 16 in the dataset Mip-NeRF360, while it is challengingly reduced to 12 in Tank&Temple and Deep Blending. The superior performance is largely attributed to our proposed second-order anchor which performs anchor feature augmentation to compensate for the reduced anchor feature size and improve the rendering quality as well as the selective gradient loss that further ensures the high-quality rendering under small-sized anchor features. Further, we benchmark SOGS with Scaffold-GS over the BungeeNeRF dataset which provides diverse large-scale scenes. It can be observed that SOGS achieves superior rendering quality with only 16-dimensional anchors. Besides, as shown in Fig. 4, we perform the visual comparison of SOGS with 3D-GS and Scaffold-GS and provide their model sizes in each scene. With the compact anchor size of the proposed second-order anchor, SOGS achieves a significant reduction in model size while delivering superior novel view synthesis with much less artifacts but finer details. Note we did not compare with HAC [9] and ContextGS [35] as they compress the storage size of Scaffold-GS rather than the actual model size as used in training and rendering.", + "bbox": [ + 511, + 369, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 816, + 679, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive ablation experiments to validate the effectiveness of the proposed second-order anchor and the selective gradient loss. In the experiments, we keep the model size and anchor feature dimension the same as used", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11172", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53c23612a1d9a5daffd7061b8e799cab742e92987764e55c4ad64b14c3562685.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 295, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96b30dcd32b63f16effd497d396c876f205a4d4ad03ad69b152c9192db80f121.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 191, + 294, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a72293882420d3c57150d9da0032643570a007aa6adbe497ed22bfc1d492190a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 295, + 294, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b2957d95718f0ea4c0272843bb69dfbcc070b7f3a5e55d0c993da423aa5bd49.jpg", + "image_caption": [ + "3D-GS" + ], + "image_footnote": [], + "bbox": [ + 91, + 400, + 294, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2cfbf6e77c9a18fad2ee76ece50a9e09076486ec5edadeb98c93982a621c3e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 89, + 496, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ec4df95056874e034468d26c2c2937f96ae9c85492e98c3f8b5929864d983fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 191, + 496, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/133c90f6d63fbf63b51b02947de23486567dd8a6483adf1b89729ae1be0a4613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 295, + 496, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7075f87aec37e98e544c993a7b17a0d66032e401b0a8c96e207dbbb7439e72da.jpg", + "image_caption": [ + "Scaffold-GS" + ], + "image_footnote": [], + "bbox": [ + 295, + 400, + 496, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f3e643bcc45e8e977e3d0483710782cc7754e5e57234706e558dcb465b63f5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 89, + 700, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfc67c21b11c7e99623ab4cbd53d3897133e19bb9c387e4e810f11aed9b33022.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 191, + 700, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e7847eb86a16524bb93a1c9dbdcb8846307972aa08ec0d7741a00c232a57118.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 295, + 700, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e9873211860b4153ce815690bfebab3f8cddf13d16fd8093e37cac18ce44463.jpg", + "image_caption": [ + "SOGS (Ours)", + "Figure 4. Qualitative comparisons of SOGS with 3D-GS [18] and Scaffold-GS [25]. SOGS achieves smaller model size and superior image rendering with much less artifacts but more fine details. The experiments are conducted over multiple indoor and outdoor scenes including 'Garden', 'Flower' and 'Counter' from the Mip-NeRF360 and 'Quebec' from the BungeeNeRF. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 500, + 400, + 700, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89f30eda274087615941b9d6507304c94608409ca04d8d8f53c274576608fb60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 89, + 903, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/360b2830dc0f98fb30f6f1e7d3a668a38b692b491446d28ce61a4b8234e45406.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 191, + 903, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1168505cd077e2a18611a0f586dd398e6c86a7d36a9b731675c5ea6c14696c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 295, + 903, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c96c271317359f08a98c09b52974db73a06100b361e07a4c3b45de9d0f08101a.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 700, + 400, + 903, + 489 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43be155e99941f268cf0bf9c04b226c1ad73cf0385d581dbe0ae8bb37dd09cbf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsEvaluation Metrics
PSNR ↑SSIM ↑LPIPS ↓
Base26.620.8650.241
Base+SOA27.250.8790.208
Base+SOA+SGL27.390.8870.161
", + "bbox": [ + 99, + 594, + 475, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies of SOGS on the dataset BungeeNeRF. With Scaffold-GS as the baseline Base, Base + SOA introduces the second-order anchor to augment anchor features which clearly improves the Gaussian attribute prediction and image rendering. Base + SOA + SGL (i.e., SOGS) further introduces the selective gradient loss which enables the Gaussian prediction to focus on regions with difficult-to-render textures and geometries. All models are trained under the same settings and model size, and the anchor feature dimension is set at 32.", + "bbox": [ + 89, + 705, + 483, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Scaffold-GS, aiming to screen out other variations and show how our two designs improve the rendering quality.", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More details are described in the following two subsections.", + "bbox": [ + 513, + 598, + 906, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second-order Anchor We first examine how our proposed second-order anchor affects PSNR, SSIM and LPIPS. In this experiment, we adopt the Scaffold-GS as the baseline model Base that performs anchor-based Gaussian splattering with naive anchors. On top of the Base, we train a model Base $+SOA$ that incorporates our proposed second-order anchor to replace the naive anchor in Scaffold-GS. As Table 3 shows, Base $+SOA$ outperforms the Base clearly in PSNR, SSIM and LPIPS, indicating that the second-order anchor can effectively improve the Gaussian attribute prediction and novel view synthesis by anchor feature augmentation.", + "bbox": [ + 511, + 636, + 906, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selective Gradient Loss To evaluate how the proposed selective gradient loss contributes, we train a new model $Base + SOA + SGL$ (i.e., the complete SOGS) that incorporates the selective gradient loss on top of $Base + SOA$ . Table 3 shows experimental results. We can observe that", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11173", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a44e6b9f7d31a0aec7f975cf9bb2e3fb338af55453a385db6000b0bbf82c3b6.jpg", + "image_caption": [ + "Figure 5. Anchor feature dimension vs SOGS performance: The performance of SOGS varies with the feature dimension $D$ in each anchor, where increasing $D$ improves SOGS consistently together with the increased model size and computational cost. The graph shows the PSNR of the scene 'Bicycle' from MipNeRF360." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 486, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Base $+SOA + SGL$ further improves the novel view synthesis consistently across all three evaluation metrics, demonstrating its effectiveness on Gaussian attribute prediction and rendering quality.", + "bbox": [ + 89, + 308, + 483, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Parameter Investigation", + "text_level": 1, + "bbox": [ + 89, + 380, + 313, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examine how the dimension of anchor features $D$ affects the SOGS performance. Fig. 5 shows experiments on the scene 'Bicycle' from the dataset MipNeRF360. We can observe that the performance of SOGS (in PSNR ↑ score) improves consistently while the feature dimension $D$ increases. Specifically, the feature dimension plays a vital role in the rendering quality as it sets the upper bound for the encoded information as well as the effectiveness of feature augmentation in SOGS. As $D$ becomes larger than 16, the performance gains gradually diminish due to feature saturation while the model size and computational costs grow greatly. We therefore set $D$ at 12 or 16 in our implemented system to balance the model size and performance.", + "bbox": [ + 89, + 402, + 483, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 89, + 609, + 228, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the impact of second-order anchors on scene textures and structures. As Fig. 6 shows, our proposed second-order anchor can enhance the rendered images with finer structures and textures. This visualization verifies that the proposed second-order anchor can effectively capture the structural and textural patterns for anchor feature augmentation, and facilitate subsequent Gaussian prediction and image rendering with augmented features.", + "bbox": [ + 89, + 632, + 483, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 768, + 210, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents SOGS, an innovative anchor-based 3D Gaussian splatting technique that incorporates second-order anchors to achieve superior rendering with compact anchor feature dimensions and model size. Specifically, by introducing covariance-based second-order statistics, SOGS learns correlations across anchor feature dimensions to capture anchor-specific textures and structural patterns. This", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a136b8b1f3c5ce0d7c08a6cb921cd9440b90b678064615716bfc31cc1adfbaa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 88, + 707, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fad434b231629a5c57501163d2a051911c9aa09d97cb21a947ed987e002633f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 181, + 707, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f84bb157558e17130ab51bbbe2616df2f27b280437e15c0d3500882c8b76707.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 273, + 707, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70f9d1682c5393f26b595da4baadeeac80d3df8b549d3dd748d789434a92bf8b.jpg", + "image_caption": [ + "Naive Anchor" + ], + "image_footnote": [], + "bbox": [ + 529, + 367, + 707, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05b4a95019e66465ff70fa9a3fd1b5a35086f2e3d69c44c0234799b77d9c88af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 89, + 890, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d1c9d73bf3aec43edcabcb1f9c52d9cea8ffe3291312a7592fcce6aa7aa960a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 181, + 890, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15b448cf857c0ff60b27b49db5d2539bc9e95568ea35f567b068decc45488e1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 273, + 890, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dda097a16269f91c2f01ba6d7c5a870ee937b809edb77660122c0626e86179f8.jpg", + "image_caption": [ + "Second-Order Anchor", + "Figure 6. Visualization of the impact of second-order anchors on scene textures and structures. The two samples are 'Drjohnson' and 'Playroom' from dataset Deep Blending [17]. The proposed second-order anchor improves the rendering quality with superior structures and textures by anchor feature augmentation. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 712, + 367, + 890, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "enables each anchor to achieve feature augmentation, compensating for the reduced anchor feature size and resulting in superior Gaussian prediction and rendering quality. Besides, we also design a selective gradient loss to further ensure high-quality rendering under compact anchor features, which enables the model to dynamically focus on the Gaussian prediction for difficult-to-render textural and structural regions. Experiments over multiple widely adopted indoor and outdoor scenes show that SOGS achieves superior novel view synthesis with compact model size. Although time complexity increases slightly due to the additional computations in second-order anchors, SOGS still achieves efficient training and real-time rendering.", + "bbox": [ + 511, + 606, + 906, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 815, + 702, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is funded by the Ministry of Education Singapore, under the Tier-2 project scheme with a project number MOE-T2EP20220-0003. Fangneng Zhan is supported by the NSF grant CRCNS-2309041.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11174", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 1, 2", + "[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2", + "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3", + "[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2", + "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[8] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2", + "[9] Yihang Chen, Qianyi Wu, Jianfei Cai, Mehrtash Harandi, and Weiyao Lin. Hac: Hash-grid assisted context for 3d gaussian splattering compression. arXiv preprint arXiv:2403.14530, 2024. 2, 3, 6", + "[10] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 3", + "[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 3", + "[12] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 14324-14334, 2021. 2", + "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3", + "[14] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2", + "[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2", + "[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3", + "[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 8", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 7", + "[19] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6", + "[20] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3", + "[21] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2", + "[22] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2", + "[23] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2", + "[24] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. Stylegaussian: Instant 3d style transfer with gaussian splatting. In SIGGRAPH Asia 2024 Technical Communications, pages 1-4. 2024. 3", + "[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 3, 5, 6, 7", + "[26] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2", + "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2", + "[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2", + "[30] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3", + "[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2", + "[32] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2", + "[33] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2", + "[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2", + "[35] Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex C Kot, and Bihan Wen. Contextgs: Compact 3d gaussian splatting with anchor level context model. arXiv preprint arXiv:2405.20721, 2024. 3, 6", + "[36] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In European conference on computer vision, pages 106-122. Springer, 2022. 2, 6", + "[38] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 3", + "[39] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 3", + "[40] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2", + "[41] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3", + "[42] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3", + "[43] Jiahui Zhang, Fangneng Zhan, Rongliang Wu, Yingchen Yu, Wenqing Zhang, Bai Song, Xiaqin Zhang, and Shijian Lu. Vmrf: View matching neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6579-6587, 2022. 2", + "[44] Jiahui Zhang, Fangneng Zhan, Yingchen Yu, Kunhao Liu, Rongliang Wu, Xiaoqin Zhang, Ling Shao, and Shijian Lu. Pose-free neural radiance fields via implicit pose regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3534-3543, 2023. 2", + "[45] Jiahui Zhang, Fangneng Zhan, Muyu Xu, Shijian Lu, and Eric Xing. Fregs: 3d gaussian splatting with progressive frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21424-21433, 2024. 3", + "[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1", + "[47] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11176", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_model.json b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..53032cb3fa724c2a3cdb718658c1e686194b67f9 --- /dev/null +++ b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "SOGS: Second-Order Anchor for Advanced 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.751, + 0.199 + ], + "angle": 0, + "content": "Jiahui Zhang\\(^{1}\\) Fangneng Zhan\\(^{2,3}\\) Ling Shao\\(^{4}\\) Shijian Lu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.199, + 0.766, + 0.216 + ], + "angle": 0, + "content": "Nanyang Technological University 2Harvard University MIT" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.218, + 0.78, + 0.235 + ], + "angle": 0, + "content": "4UCAS-Terminus AI Lab, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.238, + 0.712, + 0.252 + ], + "angle": 0, + "content": "jiahui003@e.ntu.edu.sg fnzhan@seasharvard.edu" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.256, + 0.678, + 0.27 + ], + "angle": 0, + "content": "ling.shao@ieee.org shijian.lu@ntu.edu.sg" + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.306, + 0.248, + 0.319 + ], + "angle": 0, + "content": "Rendered Images" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.322, + 0.293, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.306, + 0.436, + 0.319 + ], + "angle": 0, + "content": "Gradient Maps" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.322, + 0.494, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.306, + 0.652, + 0.319 + ], + "angle": 0, + "content": "Rendered Images" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.322, + 0.702, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.306, + 0.845, + 0.319 + ], + "angle": 0, + "content": "Gradient Maps" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.322, + 0.905, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.426, + 0.293, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.533, + 0.345, + 0.546 + ], + "angle": 0, + "content": "(a)Scaffold-GS" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.426, + 0.494, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.426, + 0.702, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.533, + 0.735, + 0.546 + ], + "angle": 0, + "content": "(b) SOGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.557, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Figure 1. The proposed SOGS can render high-quality textures and geometries and reduce model size simultaneously. The illustrations in (a) and (b) show Rendered Images and Gradient Maps which are produced by Scaffold-GS [25] and SOGS, respectively, for the samples 'Stump' and 'Room' from Mip-NeRF360 [2]. The Gradient Maps are extracted with the Sobel operator which highlight the texture and geometry of imaged scenes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.626, + 0.327, + 0.642 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.486, + 0.871 + ], + "angle": 0, + "content": "Anchor-based 3D Gaussian splatting (3D-GS) exploits anchor features in 3D Gaussian prediction, which has achieved impressive 3D rendering quality with reduced Gaussian redundancy. On the other hand, it often encounters the dilemma among anchor features, model size, and rendering quality - large anchor features lead to large 3D models and high-quality rendering whereas reducing anchor features degrades Gaussian attribute prediction which leads to clear artifacts in the rendered textures and geometries. We design SOGS, an anchor-based 3D-GS technique that introduces second-order anchors to achieve superior rendering quality and reduced anchor features and model size simultaneously. Specifically, SOGS incorporates covariance-based second-order statistics and corre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.765 + ], + "angle": 0, + "content": "lation across feature dimensions to augment features within each anchor, compensating for the reduced feature size and improving rendering quality effectively. In addition, it introduces a selective gradient loss to enhance the optimization of scene textures and scene geometries, leading to high-quality rendering with small anchor features. Extensive experiments over multiple widely adopted benchmarks show that SOGS achieves superior rendering quality in novel view synthesis with clearly reduced model size." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.798, + 0.646, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D scene representation and rendering have been a pivotal task in 3D computer vision, playing a crucial role in various applications such as virtual reality and scene simulation. Neural radiance fields (NeRFs) [27] and its variants [1-3, 46] have been developed to model implicit" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.322, + 0.901 + ], + "angle": 0, + "content": "*Shijian Lu is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "scene representation and enable volume rendering for high-quality novel view synthesis. However, NeRF is limited by long training and rendering times, largely due to the time-consuming point sampling in volume rendering. Recently, 3D Gaussian Splatting (3D-GS) [18] has been proposed as an alternative which learns explicit 3D scene representations with learnable 3D Gaussians. By leveraging efficient splatting and rasterization, 3D-GS projects 3D Gaussians onto a 2D plane, enabling real-time rendering. However, 3D-GS stacks substantial 3D Gaussians to fit each training view individually, often overlooking the underlying scene structure [25]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.547 + ], + "angle": 0, + "content": "The recent Scaffold-GS [25] exploits scene geometry to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. Specifically, it introduces anchor points to store features and exploits MLPs to predict Gaussian attributes from these anchor features. However, Scaffold-GS struggles to strike a balance between rendering quality and model size. Specifically, adopting large anchor features improves rendering quality [9] but it significantly enlarges model size due to the large number of anchor points used in scene representations. Using smaller anchor features can shrink the model, but it clearly degrades the prediction of Gaussian attributes which further leads to suboptimal rendering with various artifacts in rendered textures and geometries. HAC [9] introduces grid-based context modeling and entropy encoding to compress the Scaffold-GS. However, HAC does not reduce the model size but the storage size of the trained Scaffold-GS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.761 + ], + "angle": 0, + "content": "We design SOGS, an innovative anchor-based 3D-GS technique that introduces second-order anchors and achieves superior rendering quality with reduced anchor feature dimensions and model size. Specifically, SOGS incorporates covariance-based second-order statistics to model correlation across anchor feature dimensions. The rationale is that textures and geometries are defined by not only individual features but also the correlation across features. Hence, the correlation across anchor features can effectively guide to capture the pattern of intricate local textures and structures in each anchor. This nice feature enables anchor feature augmentation, empowering SOGS to achieve superior rendering quality with reduced feature dimensions and feature size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Additionally, we design a selective gradient loss to enhance the rendering quality with small anchor features. This loss computes gradient maps from both the rendered image and the corresponding ground truth that highlight their differences in scene textures and structures. With the difference of the gradient maps, the 3D Gaussian prediction can adaptively focus on image regions where the texture and geometry are hard to render. This enables dynamic region selection along the training process, where the model can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "adjust the focused image regions dynamically according to the evolving errors between the gradient maps of the rendered image and the ground truth. Experiments show that SOGS simultaneously achieves superior rendering quality in novel view synthesis as well as clearly reduced model sizes as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.905, + 0.426 + ], + "angle": 0, + "content": "The major contributions of this work can be summarized in three aspects. First, we propose SOGS, an innovative anchor-based 3D-GS that introduces second-order anchors to balance the rendering quality and model size. With covariance-based second-order statistics that model correlation across anchor feature dimensions to capture intricate textural and structural patterns, SOGS achieves anchor feature augmentation which produces superior rendering quality with significantly reduced anchor feature dimensions. Second, we design a selective gradient loss that guides 3D Gaussian prediction to focus on difficult-to-render textures and structures adaptively, further enhancing rendering quality with small anchor features Third, experiments over multiple benchmarks show that SOGS achieves superior novel view synthesis with efficient anchor size and model size, and outperforms the state-of-the-art in rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.442, + 0.655, + 0.457 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.842, + 0.485 + ], + "angle": 0, + "content": "2.1. Radiance Field and Neural Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.905, + 0.793 + ], + "angle": 0, + "content": "Radiance field has been widely explored for novel view synthesis and achieved impressive rendering quality. Neural radiance field (NeRF) [27] utilizes MLPs to learn implicit 3D scene representations from multi-view posed 2D images. Coupled with differentiable volume rendering, NeRF achieves high-quality rendering with superb multiview consistency. Several NeRF variants have been developed to address various new challenges, including largescale scenes [32, 33, 37], dynamic scenes [12, 14, 31], few-shot setting [6, 34, 40], pose-free setting [4, 8, 22, 43, 44], and antialIASing [2, 3]. However, NeRF and its variants still suffer from extremely long training and rendering times due to the time-consuming sampling required for volume rendering. Several studies [7, 15, 28, 29] have been proposed to address this issue. For instance, Müller et al. [28] introduce multi-resolution hash encoding to reduce the size of neural networks, significantly shortening the training time and enabling real-time rendering in tens of milliseconds. However, these methods often suffer from degraded reconstruction accuracy and rendering quality of scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Kerbl et al. recently propose 3D Gaussian splatting (3D-GS) [18], which achieves high-quality and real-time rendering by explicitly representing scenes by using parameterized 3D Gaussians, along with efficient splatting and rasterization. Building on its excellent performance, numerous 3D-GS variants have been developed to address diverse challenges and tasks, such as dynamic scenes [21, 23, 26," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11168" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "36, 41], autonomous driving [38, 47], rendering quality optimization [11, 39, 42, 45], sparse-view setting [5, 10, 20], colmap-free setting [13], and 3D style transfer [24]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.149, + 0.411, + 0.166 + ], + "angle": 0, + "content": "2.2. Anchor-based 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.484, + 0.43 + ], + "angle": 0, + "content": "One major constraint in 3D-GS is that it stacks a large number of 3D Gaussians to match each training view separately, often overlooking the underlying scene geometry. To address this issue, Lu et al. [25] present Scaffold-GS, the first anchor-based 3D-GS technique that incorporates anchor points to predict Gaussian attributes through multilayer perceptrons (MLPs). Scaffold-GS leverages scene structure to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. However, Scaffold-GS faces challenges in balancing enhanced rendering quality with anchor and model size reduction. Although several follow-ups [9, 35] introduce context-based entropy encoding to compress the Scaffold-GS, they reduce its storage size instead of anchor and model sizes as used during training and rendering. The proposed SOGS introduces second-order anchors, enabling superior rendering quality with reduced model size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.265, + 0.463 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.669 + ], + "angle": 0, + "content": "We propose SOGS, a novel anchor-based 3D Gaussian splatting that introduces second-order anchors and selective gradient loss to achieve superior rendering quality with reduced anchor and model sizes. Fig. 2 shows the overview of SOGS. The 3D Gaussian splatting [18] (3D-GS) and Scaffold-GS [25] (the base model for our SOGS) are briefly introduced in Sec. 3.1. In Sec. 3.2, we first explain why the proposed second-order anchor enables the model to enhance rendering quality in reduced anchor and model sizes, and then describe the details of the second-order anchor. Besides, to further ensure high-quality image rendering under compact anchor size, we design a selective gradient loss, detailed in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.232, + 0.696 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.854 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3D-GS). 3D-GS [18] explicitly represents scenes using anisotropic 3D Gaussians and enables real-time rendering through efficient differentiable splatting. 3D Gaussians are initialized from point clouds generated by structure-from-motion [16, 30]. Each Gaussian is defined by a covariance matrix, a position (mean), an opacity value \\(\\alpha\\), and spherical harmonics coefficients representing color \\(c\\), where the covariance matrix is decomposed into scaling matrix and rotation matrix to facilitate differentiable optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For rendering, 3D Gaussians are projected onto a 2D plane through splatting, followed by \\(\\alpha\\)-blending. Specifically, the color \\(C\\) of a pixel is computed by blending \\(N\\) or" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.898, + 0.107 + ], + "angle": 0, + "content": "dered 2D Gaussians overlapping that pixel, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.117, + 0.906, + 0.159 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {n \\in N} c _ {n} \\alpha_ {n} \\prod_ {z = 1} ^ {n - 1} (1 - \\alpha_ {z}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.169, + 0.906, + 0.199 + ], + "angle": 0, + "content": "where \\(c_{n}\\) and \\(\\alpha_{n}\\) represent the color and opacity of the \\(n\\)-th 2D Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Scaffold-GS. Scaffold-GS [25] introduces anchor points to predict 3D Gaussians via MLPs. It aims to leverage scene geometry to guide the distribution and attributes of 3D Gaussians and reduce the Gaussian redundancy. Specifically, Scaffold-GS voxelizes the entire scene based on the point cloud initialized by COLMAP [30] and assigns the center of each voxel as an anchor point. A local scene feature \\( \\pmb{f}^{\\pmb{a}} \\in \\mathbb{R}^{D} \\), a scaling factor \\( l^{\\pmb{a}} \\) and \\( K \\) offsets \\( o^{\\pmb{a}} \\in \\mathbb{R}^{3K} \\) are stored in each anchor as the anchor attributes. Given an anchor position \\( x_{a} \\), its relative distance \\( \\delta_{ac} \\) and viewing direction \\( d_{ac} \\) to a camera position \\( x_{c} \\) can be represented as: \\( \\delta_{ac} = ||x_{a} - x_{c}||_{2} \\), \\( d_{ac} = \\frac{x_{a} - x_{c}}{||x_{a} - x_{c}||_{2}} \\). The prediction of \\( K \\) Gaussians from one anchor point using MLPs \\( F \\) can then be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.44, + 0.907, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \\right\\} _ {k = 1} ^ {K} = F (\\boldsymbol {f} ^ {\\boldsymbol {a}}, \\boldsymbol {\\delta} _ {a c}, \\boldsymbol {d} _ {a c}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.529 + ], + "angle": 0, + "content": "where \\(\\alpha_{k}, c_{k}, s_{k}\\) and \\(q_{k}\\) denote the opacity, color, scaling and the rotation-related quaternion of the \\(k\\)-th Gaussian, respectively. The positions of the \\(K\\) Gaussians are calculated by this formula:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.539, + 0.906, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\left\\{p _ {k} \\right\\} _ {k = 1} ^ {K} = \\boldsymbol {x} _ {a} + \\left\\{o _ {k} ^ {a} \\right\\} _ {k = 1} ^ {K} * \\boldsymbol {l} ^ {\\boldsymbol {a}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.599 + ], + "angle": 0, + "content": "Scaffold-GS follows the same rendering process as Eq. 1 described in 3D-GS after predicting the Gaussian attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.722, + 0.622 + ], + "angle": 0, + "content": "3.2. Second-Order Anchor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.869 + ], + "angle": 0, + "content": "In this section, we first explain why the proposed second-order anchor enables the model to achieve superior image rendering with compact anchor size. The key lies in leveraging covariance-based second-order feature statistics to achieve anchor feature augmentation. Specifically, second-order feature statistics extract co-varying relationships across anchor feature dimensions. As textures and structures are defined not only by individual features but also by their interdependencies, these co-varying relationships can guide the capture of intricate local textural and structural patterns within each anchor, facilitating feature augmentation. As a result, SOGS compensates for the reduced feature dimensions and ensures superior rendering quality by performing feature augmentation. We provide the visualization of the impact of second-order anchors on scene textures and structures in the Sec. 4.5." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Given anchor feature set \\(\\mathbf{F}^{\\mathbf{a}}\\in \\mathbb{R}^{N\\times D} = [f_1^a,f_2^a,\\dots,f_N^a ]\\) , where \\(N\\) and \\(D\\) denote the number of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.313, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed SOGS. Initialized from point clouds, each anchor stores an anchor feature vector \\( f^a \\in \\mathbb{R}^D \\). Putting \\( f^a \\) along \\( D \\) dimensions as \\( D \\) variables and all anchors as observed samples, the second-order anchor statistics capturing covarying relations across \\( D \\) can be computed from the covariance matrix \\( \\Sigma \\) across \\( D \\) dimensions. We select the top \\( M \\) eigenvectors \\( P \\) as the most significant co-variation patterns and combine \\( P \\) with \\( f^a \\) to capture anchor-specific textures and geometries for anchor feature augmentation, second-order anchor construction and Gaussian prediction. With the gradient maps of the rendered image and the ground truth, the selective gradient loss guides to learn to render finer textures and geometries with dynamic region selection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.515 + ], + "angle": 0, + "content": "anchors and the dimensionality of each anchor feature, we first compute second-order statistics in the anchor feature set. Specifically, we treat \\(D\\) feature dimensions (channels) as \\(D\\) variables and utilize the anchor set as \\(N\\) observation samples, and compute the covariance matrix \\(\\Sigma \\in \\mathbb{R}^{D\\times D}\\) across these channels, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.525, + 0.483, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\frac {1}{N - 1} \\left(\\boldsymbol {F} ^ {\\boldsymbol {a}} - \\boldsymbol {\\mu} ^ {T}\\right) \\left(\\boldsymbol {F} ^ {\\boldsymbol {a}} - \\boldsymbol {\\mu} ^ {T}\\right) ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.567, + 0.483, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} = \\frac {1}{N} \\boldsymbol {F} ^ {\\boldsymbol {a}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.484, + 0.707 + ], + "angle": 0, + "content": "where \\(\\pmb{\\mu} \\in \\mathbb{R}^{D}\\) denotes the mean vector across \\(N\\) samples, used for feature centering. Given that two variables with large variances can yield a large covariance despite only weak relationship, we then construct a correlation matrix \\(\\pmb{R}\\) from the covariance matrix \\(\\pmb{\\Sigma}\\) to standardize the relationships among \\(D\\) variables, eliminating the influence of different scales or variances:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.719, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {R} = \\boldsymbol {A} ^ {- 1} \\boldsymbol {\\Sigma} \\boldsymbol {A} ^ {- 1}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.748, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} = \\operatorname {d i a g} \\left(\\sigma_ {1}, \\dots , \\sigma_ {u}, \\dots , \\sigma_ {D}\\right), \\quad \\sigma_ {u} = \\sqrt {\\Sigma_ {u u}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.484, + 0.819 + ], + "angle": 0, + "content": "where \\( \\mathbf{A} \\) is a diagonal matrix and the diagonal elements are the standard deviation of \\( D \\) variables. The correlation between \\( u \\)-th and \\( v \\)-th variables can then be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.829, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nR _ {u v} = \\frac {\\Sigma_ {u v}}{\\sigma_ {u} \\sigma_ {v}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\Sigma_{uv}\\) represents the covariance between the \\(u\\)-th and \\(v\\)-th variables. With \\(\\pmb{R}\\), we extract \\(M\\) principle co-varying" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.485 + ], + "angle": 0, + "content": "relationships across anchor feature dimensions. Specifically, we perform eigendecomposition and decompose the matrix \\(\\pmb{R}\\) into the eigenvector matrix \\(Q\\in \\mathbb{R}^{D\\times D}\\) and the diagonal matrix \\(\\Lambda \\in \\mathbb{R}^{D\\times D}\\) of eigenvalues:" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.49, + 0.905, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {R} = \\boldsymbol {Q} \\boldsymbol {\\Lambda} \\boldsymbol {Q} ^ {T}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.635 + ], + "angle": 0, + "content": "where \\(Q\\) is an orthogonal matrix and each column is an eigenvector representing a co-variation pattern across anchor feature dimensions. We sort the eigenvalues in descending order, where the eigenvector corresponding to the largest eigenvalue captures the most significant co-variation patterns across the feature dimensions. Following this, we select the top-\\(M\\) eigenvectors \\(\\pmb{P} = [P_{1},\\dots,P_{M}]\\) from \\(Q\\) as the primary directions of co-variation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.741 + ], + "angle": 0, + "content": "We then leverage \\( P \\) to guide the extraction of textural and structural information in each anchor. As the covarying relationships across anchor feature dimensions are extracted from the anchor set \\( F^a \\), we treat \\( P \\) as the global-shared co-variations. Combined with the feature \\( f^a \\) in one anchor, we can capture the anchor-specific textures and structures as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.748, + 0.905, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} ^ {\\boldsymbol {t}} = F _ {i} \\left(\\left[ \\boldsymbol {P} _ {i}, \\boldsymbol {f} ^ {\\boldsymbol {a}} \\right]\\right), \\quad i \\in [ 1, M ], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.906, + 0.878 + ], + "angle": 0, + "content": "where \\(F_{i}(\\cdot)\\) represents the two-layer MLP used to extract the textures and structures for anchor \\(\\pmb{f}^{a}\\) based on the \\(i\\)-th principal co-varying relationship across anchor feature dimensions. Finally, we concatenate the extracted features with the original anchor feature to achieve feature augmentation and predict Gaussian attributes. Take one second-order anchor as an example:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \\right\\} _ {k = 0} ^ {K} = F \\left(\\boldsymbol {f} ^ {\\boldsymbol {a}}, \\left\\{\\boldsymbol {f} _ {\\boldsymbol {i}} ^ {\\boldsymbol {t}} \\right\\} _ {i = 0} ^ {M}, \\boldsymbol {\\delta} _ {a c}, \\boldsymbol {d} _ {a c}\\right). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.485, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Visual illustration of the proposed selective gradient loss (SGL). SGL clearly improves the rendering quality by generating finer textures and details. Zoom in for best view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.445 + ], + "angle": 0, + "content": "Note that \\(\\{\\pmb{f}_i^t\\}_{i=0}^M\\) are derived from \\(\\pmb{f}^a\\) for feature augmentation and do not increase the anchor feature dimensions and the size of each anchor." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.31, + 0.47 + ], + "angle": 0, + "content": "3.3. Selective Gradient Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.721 + ], + "angle": 0, + "content": "In addition to proposing second-order anchors for anchor feature augmentation to achieve superior rendering quality with reduced anchor size, we design a selective gradient loss to further ensure the rendering quality under compact anchor features. It enables the model to adaptively identify difficult-to-render textures and structures, and regularize Gaussian prediction corresponding to these regions. With the designed loss, superior rendering quality with finer textures can be achieved as shown in Fig 3. Given the limited sensitivity of pixel-level L1 loss to textures and structures in RGB images, we consider gradient map to emphasize textural and structural regions. Specifically, we exploit the Sobel operator to extract horizontal and vertical gradient maps \\((G_x', G_y', G_x, G_y)\\) of the rendering result \\(I' \\in \\mathbb{R}^{H \\times W \\times C}\\) and its corresponding ground truth \\(I\\), which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.733, + 0.483, + 0.758 + ], + "angle": 0, + "content": "\\[\nG _ {x} ^ {\\prime} = S _ {x} * I ^ {\\prime}, \\quad G _ {y} ^ {\\prime} = S _ {y} * I ^ {\\prime}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.756, + 0.393, + 0.771 + ], + "angle": 0, + "content": "\\[\nG _ {x} = S _ {x} * I, \\quad G _ {y} = S _ {y} * I,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.785, + 0.484, + 0.835 + ], + "angle": 0, + "content": "\\[\nS _ {x} = \\left[ \\begin{array}{l l l} - 1 & 0 & 1 \\\\ - 2 & 0 & 2 \\\\ - 1 & 0 & 1 \\end{array} \\right], \\quad S _ {y} = \\left[ \\begin{array}{l l l} - 1 & - 2 & - 1 \\\\ 0 & 0 & 0 \\\\ 1 & 2 & 1 \\end{array} \\right], \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "where \\( S_{x} \\) and \\( S_{y} \\) denote the horizontal and vertical Sobel kernels, respectively. We then use Euclidean metric to measure the horizontal and vertical gradient discrepancies \\( (l_{x}, l_{y}) \\) between the rendering \\( I' \\) and the corresponding ground" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "truth \\(I\\), which are then averaged to derive the final discrepancies as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.13, + 0.905, + 0.182 + ], + "angle": 0, + "content": "\\[\nl _ {x} = \\frac {1}{\\sqrt {H W}} \\sum_ {\\substack {x = 0 \\\\ H = 1}} ^ {H - 1} \\sum_ {\\substack {y = 0 \\\\ W = 1}} ^ {W - 1} \\left| G _ {x} ^ {\\prime} (x, y) - G _ {x} (x, y) \\right|, \\tag{14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.178, + 0.85, + 0.22 + ], + "angle": 0, + "content": "\\[\nl _ {y} = \\frac {1}{\\sqrt {H W}} \\sum_ {x = 0} ^ {H - 1} \\sum_ {y = 0} ^ {W - 1} \\left| G _ {y} ^ {\\prime} (x, y) - G _ {y} (x, y) \\right|.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Gradient maps, derived from differences in pixel intensities, typically exhibit meaningful gradients in localized regions, such as edges and areas with significant intensity changes, which are crucial for capturing structures and textures. In contrast, the majority of areas in gradient map comprise flat and low-gradient regions with limited information. Therefore naively adopting the horizontal and vertical gradient discrepancies as the loss function can cause low-gradient regions to dominate the overall loss. As a result, the model may deemphasize regions critical for preserving fine textures and structures, focusing instead on less important flat areas. Besides, the regions with meaningful gradients should also be distinguished, prioritizing greater attention to textures and structures exhibiting larger rendering errors. To address the above issues, we introduce a dynamic region selection that computes the rendering error map as a weight map, defined by the absolute difference between the gradient maps of the rendered result and the ground truth, which can be expressed by:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.528, + 0.905, + 0.554 + ], + "angle": 0, + "content": "\\[\nw _ {x} = \\left| G _ {x} ^ {\\prime} (x, y) - G _ {x} (x, y) \\right|, \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.551, + 0.808, + 0.567 + ], + "angle": 0, + "content": "\\[\nw _ {y} = \\left| G _ {y} ^ {\\prime} (x, y) - G _ {y} (x, y) \\right|.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.578, + 0.906, + 0.608 + ], + "angle": 0, + "content": "Combined with the horizontal and vertical weight maps, the selective gradient loss \\(\\mathcal{L}_s\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.621, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = w _ {x} * l _ {x} + w _ {y} * l _ {y}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.907, + 0.816 + ], + "angle": 0, + "content": "Based on the above analysis, the selective gradient loss enables the model to focus on the difficult-to-render textural and structural regions adaptively. Besides, the model is able to adjust the focused regions dynamically in response to the evolving rendering error map throughout the training process. Note, the pixel-level L1 loss with D-SSIM term between RGB images is also used in the training process, which complements the proposed selective gradient loss used between gradient maps. And we also keep the volume regularization as described in Scaffold-GS [25]. The total loss function is given as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.828, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} + \\lambda_ {v o l} \\mathcal {L} _ {v o l} + \\lambda_ {s} \\mathcal {L} _ {s}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where the \\(\\lambda_s\\), \\(\\lambda_1\\), \\(\\lambda_{SSIM}\\) and \\(\\lambda_{vol}\\) represent the training weights of the selective gradient loss \\(\\mathcal{L}_s\\), the L1 loss \\(\\mathcal{L}_1\\), the D-SSIM term \\(\\mathcal{L}_{SSIM}\\) and the volume regularization \\(\\mathcal{L}_{vol}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11171" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
DatasetsMip-NeRF360Tanks&TemplesDeep Blending
MethodsSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.80627.500.25232 dim0.85323.960.17732 dim0.90630.210.25432 dim
SOGS(Ours)0.81527.850.22116 dim0.85524.140.17612 dim0.90730.290.25212 dim
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.195, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of novel view synthesis. All methods are trained with the same training data. SOGS achieves superior rendering quality with reduced anchor (model) size. Notably, SOGS using the feature dimension of 12 can still outperform the Scaffold-GS that uses a feature dimension of 32. We do not compare with HAC [9] and ContextGS [35], as they leverage compression techniques to reduce the storage size of Scaffold-GS rather than the actual anchor and model sizes as used in training and rendering." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.275, + 0.48, + 0.361 + ], + "angle": 0, + "content": "
DatasetsBungeeNeRF
MethodsSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.86526.620.24132 dim
SOGS(Ours)0.88027.060.17116 dim
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on large-scale scenes from the dataset BungeeNeRF [37]. All methods are trained with the same training data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.224, + 0.456 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.411, + 0.48 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.697 + ], + "angle": 0, + "content": "Datasets For both training and testing, we adopt the datasets and configurations in Scaffold-GS [25] and conduct experiments on images of 19 real-world scenes. Specifically, we evaluate SOGS across all nine scenes from the Mip-NeRF360 dataset [2], two scenes from the Tanks&Temples dataset [19], and two additional scenes from the Deep Blending dataset [17]. These scenes exhibit diverse styles, ranging from bounded indoor environments to unbounded outdoor ones. We also evaluate SOGS on the BungeeNeRF dataset [37], which includes six large-scale outdoor scenes, to further validate its effectiveness. The datasets are divided into training and test sets following the same settings as Scaffold-GS, with image resolutions consistent with those used in Scaffold-GS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation For the second-order anchor, we reduce the dimension \\(D\\) of the anchor features in Scaffold-GS from 32 to 16, or even to 12. The covariance matrix \\(\\pmb{\\Sigma}\\) is computed from all anchors in scene representations which captures the global correlation across all anchor feature dimensions. The number of the selected eigenvectors is set to \\(M = 2\\). All MLPs used to extract anchor-specific textural and structural information consist of two layers with ReLU activation. For the loss function, the training weights \\(\\lambda_{s}, \\lambda_{1}, \\lambda_{SSIM}\\) and \\(\\lambda_{vol}\\) are set to 0.01, 0.8, 0.2 and 0.01. Please note, for fair comparisons, we keep the loss terms \\(\\lambda_{1}, \\lambda_{SSIM}\\) and \\(\\lambda_{vol}\\) the same as in Scaffold-GS. The number of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.278, + 0.907, + 0.339 + ], + "angle": 0, + "content": "3D Gaussians corresponding to one second-order anchor is set as \\( K = 10 \\). We use the Pytorch framework to implement SOGS. The SOGS model is trained for 30000 iterations, the same as Scaffold-GS and 3D-GS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.348, + 0.847, + 0.365 + ], + "angle": 0, + "content": "4.2. Comparisons with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.908, + 0.81 + ], + "angle": 0, + "content": "We primarily compare SOGS with Scaffold-GS [25] over datasets Mip-NeRF360, Tank&Temple and Deep Blending as Scaffold-GS is an anchor-based 3D Gaussian splattering method and is most relevant to our study. Table 1 shows experimental results over the same test images as described in Section 4.1. We can observe that SOGS outperforms Scaffold-GS consistently in PSNR, SSIM and LPIPS across all evaluated scenes. The anchor size of SOGS is reduced to 16 in the dataset Mip-NeRF360, while it is challengingly reduced to 12 in Tank&Temple and Deep Blending. The superior performance is largely attributed to our proposed second-order anchor which performs anchor feature augmentation to compensate for the reduced anchor feature size and improve the rendering quality as well as the selective gradient loss that further ensures the high-quality rendering under small-sized anchor features. Further, we benchmark SOGS with Scaffold-GS over the BungeeNeRF dataset which provides diverse large-scale scenes. It can be observed that SOGS achieves superior rendering quality with only 16-dimensional anchors. Besides, as shown in Fig. 4, we perform the visual comparison of SOGS with 3D-GS and Scaffold-GS and provide their model sizes in each scene. With the compact anchor size of the proposed second-order anchor, SOGS achieves a significant reduction in model size while delivering superior novel view synthesis with much less artifacts but finer details. Note we did not compare with HAC [9] and ContextGS [35] as they compress the storage size of Scaffold-GS rather than the actual model size as used in training and rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.68, + 0.833 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conduct extensive ablation experiments to validate the effectiveness of the proposed second-order anchor and the selective gradient loss. In the experiments, we keep the model size and anchor feature dimension the same as used" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.296, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.192, + 0.295, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.296, + 0.295, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.401, + 0.295, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.494, + 0.214, + 0.505 + ], + "angle": 0, + "content": "3D-GS" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.09, + 0.498, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.192, + 0.498, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.296, + 0.498, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.401, + 0.498, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.494, + 0.446, + 0.505 + ], + "angle": 0, + "content": "Scaffold-GS" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.09, + 0.702, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.192, + 0.702, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.296, + 0.702, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.401, + 0.702, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.494, + 0.65, + 0.506 + ], + "angle": 0, + "content": "SOGS (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.09, + 0.905, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.192, + 0.904, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.296, + 0.904, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.401, + 0.904, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.494, + 0.85, + 0.505 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.53, + 0.908, + 0.573 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of SOGS with 3D-GS [18] and Scaffold-GS [25]. SOGS achieves smaller model size and superior image rendering with much less artifacts but more fine details. The experiments are conducted over multiple indoor and outdoor scenes including 'Garden', 'Flower' and 'Counter' from the Mip-NeRF360 and 'Quebec' from the BungeeNeRF. Zoom in for best view." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.595, + 0.477, + 0.684 + ], + "angle": 0, + "content": "
ModelsEvaluation Metrics
PSNR ↑SSIM ↑LPIPS ↓
Base26.620.8650.241
Base+SOA27.250.8790.208
Base+SOA+SGL27.390.8870.161
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.831 + ], + "angle": 0, + "content": "Table 3. Ablation studies of SOGS on the dataset BungeeNeRF. With Scaffold-GS as the baseline Base, Base + SOA introduces the second-order anchor to augment anchor features which clearly improves the Gaussian attribute prediction and image rendering. Base + SOA + SGL (i.e., SOGS) further introduces the selective gradient loss which enables the Gaussian prediction to focus on regions with difficult-to-render textures and geometries. All models are trained under the same settings and model size, and the anchor feature dimension is set at 32." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "in Scaffold-GS, aiming to screen out other variations and show how our two designs improve the rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.599, + 0.907, + 0.615 + ], + "angle": 0, + "content": "More details are described in the following two subsections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.803 + ], + "angle": 0, + "content": "Second-order Anchor We first examine how our proposed second-order anchor affects PSNR, SSIM and LPIPS. In this experiment, we adopt the Scaffold-GS as the baseline model Base that performs anchor-based Gaussian splattering with naive anchors. On top of the Base, we train a model Base \\(+SOA\\) that incorporates our proposed second-order anchor to replace the naive anchor in Scaffold-GS. As Table 3 shows, Base \\(+SOA\\) outperforms the Base clearly in PSNR, SSIM and LPIPS, indicating that the second-order anchor can effectively improve the Gaussian attribute prediction and novel view synthesis by anchor feature augmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Selective Gradient Loss To evaluate how the proposed selective gradient loss contributes, we train a new model \\( Base + SOA + SGL \\) (i.e., the complete SOGS) that incorporates the selective gradient loss on top of \\( Base + SOA \\). Table 3 shows experimental results. We can observe that" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.488, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.211, + 0.486, + 0.285 + ], + "angle": 0, + "content": "Figure 5. Anchor feature dimension vs SOGS performance: The performance of SOGS varies with the feature dimension \\(D\\) in each anchor, where increasing \\(D\\) improves SOGS consistently together with the increased model size and computational cost. The graph shows the PSNR of the scene 'Bicycle' from MipNeRF360." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.485, + 0.371 + ], + "angle": 0, + "content": "Base \\(+SOA + SGL\\) further improves the novel view synthesis consistently across all three evaluation metrics, demonstrating its effectiveness on Gaussian attribute prediction and rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.381, + 0.315, + 0.397 + ], + "angle": 0, + "content": "4.4. Parameter Investigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.601 + ], + "angle": 0, + "content": "We examine how the dimension of anchor features \\( D \\) affects the SOGS performance. Fig. 5 shows experiments on the scene 'Bicycle' from the dataset MipNeRF360. We can observe that the performance of SOGS (in PSNR ↑ score) improves consistently while the feature dimension \\( D \\) increases. Specifically, the feature dimension plays a vital role in the rendering quality as it sets the upper bound for the encoded information as well as the effectiveness of feature augmentation in SOGS. As \\( D \\) becomes larger than 16, the performance gains gradually diminish due to feature saturation while the model size and computational costs grow greatly. We therefore set \\( D \\) at 12 or 16 in our implemented system to balance the model size and performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.229, + 0.625 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.755 + ], + "angle": 0, + "content": "We visualize the impact of second-order anchors on scene textures and structures. As Fig. 6 shows, our proposed second-order anchor can enhance the rendered images with finer structures and textures. This visualization verifies that the proposed second-order anchor can effectively capture the structural and textural patterns for anchor feature augmentation, and facilitate subsequent Gaussian prediction and image rendering with augmented features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.212, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "This paper presents SOGS, an innovative anchor-based 3D Gaussian splatting technique that incorporates second-order anchors to achieve superior rendering with compact anchor feature dimensions and model size. Specifically, by introducing covariance-based second-order statistics, SOGS learns correlations across anchor feature dimensions to capture anchor-specific textures and structural patterns. This" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.089, + 0.709, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.182, + 0.709, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.274, + 0.709, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.368, + 0.709, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.462, + 0.652, + 0.472 + ], + "angle": 0, + "content": "Naive Anchor" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.09, + 0.892, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.182, + 0.892, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.274, + 0.892, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.368, + 0.892, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.461, + 0.862, + 0.472 + ], + "angle": 0, + "content": "Second-Order Anchor" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 6. Visualization of the impact of second-order anchors on scene textures and structures. The two samples are 'Drjohnson' and 'Playroom' from dataset Deep Blending [17]. The proposed second-order anchor improves the rendering quality with superior structures and textures by anchor feature augmentation. Zoom in for best view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.803 + ], + "angle": 0, + "content": "enables each anchor to achieve feature augmentation, compensating for the reduced anchor feature size and resulting in superior Gaussian prediction and rendering quality. Besides, we also design a selective gradient loss to further ensure high-quality rendering under compact anchor features, which enables the model to dynamically focus on the Gaussian prediction for difficult-to-render textural and structural regions. Experiments over multiple widely adopted indoor and outdoor scenes show that SOGS achieves superior novel view synthesis with compact model size. Although time complexity increases slightly due to the additional computations in second-order anchors, SOGS still achieves efficient training and real-time rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.704, + 0.833 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This project is funded by the Ministry of Education Singapore, under the Tier-2 project scheme with a project number MOE-T2EP20220-0003. Fangneng Zhan is supported by the NSF grant CRCNS-2309041." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11174" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[8] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Yihang Chen, Qianyi Wu, Jianfei Cai, Mehrtash Harandi, and Weiyao Lin. Hac: Hash-grid assisted context for 3d gaussian splattering compression. arXiv preprint arXiv:2403.14530, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.861, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 14324-14334, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[20] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[21] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[22] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[24] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. Stylegaussian: Instant 3d style transfer with gaussian splatting. In SIGGRAPH Asia 2024 Technical Communications, pages 1-4. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11175" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[26] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[30] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.45, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[32] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[33] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[35] Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex C Kot, and Bihan Wen. Contextgs: Compact 3d gaussian splatting with anchor level context model. arXiv preprint arXiv:2405.20721, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[36] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[37] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In European conference on computer vision, pages 106-122. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[38] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[39] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[40] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[41] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.499 + ], + "angle": 0, + "content": "[42] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[43] Jiahui Zhang, Fangneng Zhan, Rongliang Wu, Yingchen Yu, Wenqing Zhang, Bai Song, Xiaqin Zhang, and Shijian Lu. Vmrf: View matching neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6579-6587, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[44] Jiahui Zhang, Fangneng Zhan, Yingchen Yu, Kunhao Liu, Rongliang Wu, Xiaoqin Zhang, Ling Shao, and Shijian Lu. Pose-free neural radiance fields via implicit pose regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3534-3543, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[45] Jiahui Zhang, Fangneng Zhan, Muyu Xu, Shijian Lu, and Eric Xing. Fregs: 3d gaussian splatting with progressive frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21424-21433, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.839 + ], + "angle": 0, + "content": "[47] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11176" + } + ] +] \ No newline at end of file diff --git a/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_origin.pdf b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1042a6e9d45e58a8fe784ede7c035f6c41e6dcbd --- /dev/null +++ b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/c50dd783-fbc4-44df-8116-14786787e34e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc803768dc04807e5c8d0fc2d87518f47d082b9adfe4233b78e98c6dec787605 +size 6874024 diff --git a/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/full.md b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..940c13be91b7db88e9bf72091cfe7e75255ac3b8 --- /dev/null +++ b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/full.md @@ -0,0 +1,386 @@ +# SOGS: Second-Order Anchor for Advanced 3D Gaussian Splitting + +Jiahui Zhang $^{1}$ Fangneng Zhan $^{2,3}$ Ling Shao $^{4}$ Shijian Lu $^{1}$ + +Nanyang Technological University 2Harvard University MIT + +4UCAS-Terminus AI Lab, University of Chinese Academy of Sciences + +jiahui003@e.ntu.edu.sg fnzhan@seasharvard.edu + +ling.shao@ieee.org shijian.lu@ntu.edu.sg + +![](images/c7f57ee65b0bd04d32da9829fb266809b65b70d3b3762ab5998ddc08e2a32997.jpg) +Rendered Images + +![](images/e4a5b0d7d869538bad3dcaa45b21349fa5b8d86897523387b6860fdf5d4b4950.jpg) +Gradient Maps + +![](images/323fb34a19344feeb6abf1d3bbf31d0b3b9bd8d4662d3cc0908bfec89b75f02e.jpg) +Rendered Images + +![](images/f2358e0fade5230fc7f40e4e0f511c059525c51757329926861deef1041edae4.jpg) +Gradient Maps + +![](images/07831be65495cb6054194bbcea4b71fac52e4e152eb5529631addaae611073d3.jpg) +(a)Scaffold-GS + +![](images/1cb42ed6dc80f7c1c643dc50d1e230a45b2763600f0e2fe06ccde40aec20736b.jpg) +Figure 1. The proposed SOGS can render high-quality textures and geometries and reduce model size simultaneously. The illustrations in (a) and (b) show Rendered Images and Gradient Maps which are produced by Scaffold-GS [25] and SOGS, respectively, for the samples 'Stump' and 'Room' from Mip-NeRF360 [2]. The Gradient Maps are extracted with the Sobel operator which highlight the texture and geometry of imaged scenes. + +![](images/7e132f0cb814e5b2b047bf791228eeb1c1b93f0963b73449acc06e76dc3d087a.jpg) +(b) SOGS + +# Abstract + +Anchor-based 3D Gaussian splatting (3D-GS) exploits anchor features in 3D Gaussian prediction, which has achieved impressive 3D rendering quality with reduced Gaussian redundancy. On the other hand, it often encounters the dilemma among anchor features, model size, and rendering quality - large anchor features lead to large 3D models and high-quality rendering whereas reducing anchor features degrades Gaussian attribute prediction which leads to clear artifacts in the rendered textures and geometries. We design SOGS, an anchor-based 3D-GS technique that introduces second-order anchors to achieve superior rendering quality and reduced anchor features and model size simultaneously. Specifically, SOGS incorporates covariance-based second-order statistics and corre + +lation across feature dimensions to augment features within each anchor, compensating for the reduced feature size and improving rendering quality effectively. In addition, it introduces a selective gradient loss to enhance the optimization of scene textures and scene geometries, leading to high-quality rendering with small anchor features. Extensive experiments over multiple widely adopted benchmarks show that SOGS achieves superior rendering quality in novel view synthesis with clearly reduced model size. + +# 1. Introduction + +3D scene representation and rendering have been a pivotal task in 3D computer vision, playing a crucial role in various applications such as virtual reality and scene simulation. Neural radiance fields (NeRFs) [27] and its variants [1-3, 46] have been developed to model implicit + +scene representation and enable volume rendering for high-quality novel view synthesis. However, NeRF is limited by long training and rendering times, largely due to the time-consuming point sampling in volume rendering. Recently, 3D Gaussian Splatting (3D-GS) [18] has been proposed as an alternative which learns explicit 3D scene representations with learnable 3D Gaussians. By leveraging efficient splatting and rasterization, 3D-GS projects 3D Gaussians onto a 2D plane, enabling real-time rendering. However, 3D-GS stacks substantial 3D Gaussians to fit each training view individually, often overlooking the underlying scene structure [25]. + +The recent Scaffold-GS [25] exploits scene geometry to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. Specifically, it introduces anchor points to store features and exploits MLPs to predict Gaussian attributes from these anchor features. However, Scaffold-GS struggles to strike a balance between rendering quality and model size. Specifically, adopting large anchor features improves rendering quality [9] but it significantly enlarges model size due to the large number of anchor points used in scene representations. Using smaller anchor features can shrink the model, but it clearly degrades the prediction of Gaussian attributes which further leads to suboptimal rendering with various artifacts in rendered textures and geometries. HAC [9] introduces grid-based context modeling and entropy encoding to compress the Scaffold-GS. However, HAC does not reduce the model size but the storage size of the trained Scaffold-GS. + +We design SOGS, an innovative anchor-based 3D-GS technique that introduces second-order anchors and achieves superior rendering quality with reduced anchor feature dimensions and model size. Specifically, SOGS incorporates covariance-based second-order statistics to model correlation across anchor feature dimensions. The rationale is that textures and geometries are defined by not only individual features but also the correlation across features. Hence, the correlation across anchor features can effectively guide to capture the pattern of intricate local textures and structures in each anchor. This nice feature enables anchor feature augmentation, empowering SOGS to achieve superior rendering quality with reduced feature dimensions and feature size. + +Additionally, we design a selective gradient loss to enhance the rendering quality with small anchor features. This loss computes gradient maps from both the rendered image and the corresponding ground truth that highlight their differences in scene textures and structures. With the difference of the gradient maps, the 3D Gaussian prediction can adaptively focus on image regions where the texture and geometry are hard to render. This enables dynamic region selection along the training process, where the model can + +adjust the focused image regions dynamically according to the evolving errors between the gradient maps of the rendered image and the ground truth. Experiments show that SOGS simultaneously achieves superior rendering quality in novel view synthesis as well as clearly reduced model sizes as shown in Fig. 1. + +The major contributions of this work can be summarized in three aspects. First, we propose SOGS, an innovative anchor-based 3D-GS that introduces second-order anchors to balance the rendering quality and model size. With covariance-based second-order statistics that model correlation across anchor feature dimensions to capture intricate textural and structural patterns, SOGS achieves anchor feature augmentation which produces superior rendering quality with significantly reduced anchor feature dimensions. Second, we design a selective gradient loss that guides 3D Gaussian prediction to focus on difficult-to-render textures and structures adaptively, further enhancing rendering quality with small anchor features Third, experiments over multiple benchmarks show that SOGS achieves superior novel view synthesis with efficient anchor size and model size, and outperforms the state-of-the-art in rendering quality. + +# 2. Related Work + +# 2.1. Radiance Field and Neural Rendering + +Radiance field has been widely explored for novel view synthesis and achieved impressive rendering quality. Neural radiance field (NeRF) [27] utilizes MLPs to learn implicit 3D scene representations from multi-view posed 2D images. Coupled with differentiable volume rendering, NeRF achieves high-quality rendering with superb multiview consistency. Several NeRF variants have been developed to address various new challenges, including largescale scenes [32, 33, 37], dynamic scenes [12, 14, 31], few-shot setting [6, 34, 40], pose-free setting [4, 8, 22, 43, 44], and antialIASing [2, 3]. However, NeRF and its variants still suffer from extremely long training and rendering times due to the time-consuming sampling required for volume rendering. Several studies [7, 15, 28, 29] have been proposed to address this issue. For instance, Müller et al. [28] introduce multi-resolution hash encoding to reduce the size of neural networks, significantly shortening the training time and enabling real-time rendering in tens of milliseconds. However, these methods often suffer from degraded reconstruction accuracy and rendering quality of scenes. + +Kerbl et al. recently propose 3D Gaussian splatting (3D-GS) [18], which achieves high-quality and real-time rendering by explicitly representing scenes by using parameterized 3D Gaussians, along with efficient splatting and rasterization. Building on its excellent performance, numerous 3D-GS variants have been developed to address diverse challenges and tasks, such as dynamic scenes [21, 23, 26, + +36, 41], autonomous driving [38, 47], rendering quality optimization [11, 39, 42, 45], sparse-view setting [5, 10, 20], colmap-free setting [13], and 3D style transfer [24]. + +# 2.2. Anchor-based 3D Gaussian Splitting + +One major constraint in 3D-GS is that it stacks a large number of 3D Gaussians to match each training view separately, often overlooking the underlying scene geometry. To address this issue, Lu et al. [25] present Scaffold-GS, the first anchor-based 3D-GS technique that incorporates anchor points to predict Gaussian attributes through multilayer perceptrons (MLPs). Scaffold-GS leverages scene structure to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. However, Scaffold-GS faces challenges in balancing enhanced rendering quality with anchor and model size reduction. Although several follow-ups [9, 35] introduce context-based entropy encoding to compress the Scaffold-GS, they reduce its storage size instead of anchor and model sizes as used during training and rendering. The proposed SOGS introduces second-order anchors, enabling superior rendering quality with reduced model size. + +# 3. Proposed Method + +We propose SOGS, a novel anchor-based 3D Gaussian splatting that introduces second-order anchors and selective gradient loss to achieve superior rendering quality with reduced anchor and model sizes. Fig. 2 shows the overview of SOGS. The 3D Gaussian splatting [18] (3D-GS) and Scaffold-GS [25] (the base model for our SOGS) are briefly introduced in Sec. 3.1. In Sec. 3.2, we first explain why the proposed second-order anchor enables the model to enhance rendering quality in reduced anchor and model sizes, and then describe the details of the second-order anchor. Besides, to further ensure high-quality image rendering under compact anchor size, we design a selective gradient loss, detailed in Sec. 3.3. + +# 3.1. Preliminaries + +3D Gaussian Splatting (3D-GS). 3D-GS [18] explicitly represents scenes using anisotropic 3D Gaussians and enables real-time rendering through efficient differentiable splatting. 3D Gaussians are initialized from point clouds generated by structure-from-motion [16, 30]. Each Gaussian is defined by a covariance matrix, a position (mean), an opacity value $\alpha$ , and spherical harmonics coefficients representing color $c$ , where the covariance matrix is decomposed into scaling matrix and rotation matrix to facilitate differentiable optimization. + +For rendering, 3D Gaussians are projected onto a 2D plane through splatting, followed by $\alpha$ -blending. Specifically, the color $C$ of a pixel is computed by blending $N$ or + +dered 2D Gaussians overlapping that pixel, formulated as: + +$$ +C = \sum_ {n \in N} c _ {n} \alpha_ {n} \prod_ {z = 1} ^ {n - 1} (1 - \alpha_ {z}), \tag {1} +$$ + +where $c_{n}$ and $\alpha_{n}$ represent the color and opacity of the $n$ -th 2D Gaussian. + +Scaffold-GS. Scaffold-GS [25] introduces anchor points to predict 3D Gaussians via MLPs. It aims to leverage scene geometry to guide the distribution and attributes of 3D Gaussians and reduce the Gaussian redundancy. Specifically, Scaffold-GS voxelizes the entire scene based on the point cloud initialized by COLMAP [30] and assigns the center of each voxel as an anchor point. A local scene feature $\pmb{f}^{\pmb{a}} \in \mathbb{R}^{D}$ , a scaling factor $l^{\pmb{a}}$ and $K$ offsets $o^{\pmb{a}} \in \mathbb{R}^{3K}$ are stored in each anchor as the anchor attributes. Given an anchor position $x_{a}$ , its relative distance $\delta_{ac}$ and viewing direction $d_{ac}$ to a camera position $x_{c}$ can be represented as: $\delta_{ac} = ||x_{a} - x_{c}||_{2}$ , $d_{ac} = \frac{x_{a} - x_{c}}{||x_{a} - x_{c}||_{2}}$ . The prediction of $K$ Gaussians from one anchor point using MLPs $F$ can then be formulated as follows: + +$$ +\left\{\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \right\} _ {k = 1} ^ {K} = F (\boldsymbol {f} ^ {\boldsymbol {a}}, \boldsymbol {\delta} _ {a c}, \boldsymbol {d} _ {a c}), \tag {2} +$$ + +where $\alpha_{k}, c_{k}, s_{k}$ and $q_{k}$ denote the opacity, color, scaling and the rotation-related quaternion of the $k$ -th Gaussian, respectively. The positions of the $K$ Gaussians are calculated by this formula: + +$$ +\left\{p _ {k} \right\} _ {k = 1} ^ {K} = \boldsymbol {x} _ {a} + \left\{o _ {k} ^ {a} \right\} _ {k = 1} ^ {K} * \boldsymbol {l} ^ {\boldsymbol {a}}. \tag {3} +$$ + +Scaffold-GS follows the same rendering process as Eq. 1 described in 3D-GS after predicting the Gaussian attributes. + +# 3.2. Second-Order Anchor + +In this section, we first explain why the proposed second-order anchor enables the model to achieve superior image rendering with compact anchor size. The key lies in leveraging covariance-based second-order feature statistics to achieve anchor feature augmentation. Specifically, second-order feature statistics extract co-varying relationships across anchor feature dimensions. As textures and structures are defined not only by individual features but also by their interdependencies, these co-varying relationships can guide the capture of intricate local textural and structural patterns within each anchor, facilitating feature augmentation. As a result, SOGS compensates for the reduced feature dimensions and ensures superior rendering quality by performing feature augmentation. We provide the visualization of the impact of second-order anchors on scene textures and structures in the Sec. 4.5. + +Given anchor feature set $\mathbf{F}^{\mathbf{a}}\in \mathbb{R}^{N\times D} = [f_1^a,f_2^a,\dots,f_N^a ]$ , where $N$ and $D$ denote the number of + +![](images/d42c671027edd64f2e7d10d34cf3fc38412f8c15b1db14e2ab326a84f8e69db9.jpg) +Figure 2. Overview of the proposed SOGS. Initialized from point clouds, each anchor stores an anchor feature vector $f^a \in \mathbb{R}^D$ . Putting $f^a$ along $D$ dimensions as $D$ variables and all anchors as observed samples, the second-order anchor statistics capturing covarying relations across $D$ can be computed from the covariance matrix $\Sigma$ across $D$ dimensions. We select the top $M$ eigenvectors $P$ as the most significant co-variation patterns and combine $P$ with $f^a$ to capture anchor-specific textures and geometries for anchor feature augmentation, second-order anchor construction and Gaussian prediction. With the gradient maps of the rendered image and the ground truth, the selective gradient loss guides to learn to render finer textures and geometries with dynamic region selection. + +anchors and the dimensionality of each anchor feature, we first compute second-order statistics in the anchor feature set. Specifically, we treat $D$ feature dimensions (channels) as $D$ variables and utilize the anchor set as $N$ observation samples, and compute the covariance matrix $\Sigma \in \mathbb{R}^{D\times D}$ across these channels, which can be formulated as: + +$$ +\boldsymbol {\Sigma} = \frac {1}{N - 1} \left(\boldsymbol {F} ^ {\boldsymbol {a}} - \boldsymbol {\mu} ^ {T}\right) \left(\boldsymbol {F} ^ {\boldsymbol {a}} - \boldsymbol {\mu} ^ {T}\right) ^ {T}, \tag {4} +$$ + +$$ +\boldsymbol {\mu} = \frac {1}{N} \boldsymbol {F} ^ {\boldsymbol {a}}, \tag {5} +$$ + +where $\pmb{\mu} \in \mathbb{R}^{D}$ denotes the mean vector across $N$ samples, used for feature centering. Given that two variables with large variances can yield a large covariance despite only weak relationship, we then construct a correlation matrix $\pmb{R}$ from the covariance matrix $\pmb{\Sigma}$ to standardize the relationships among $D$ variables, eliminating the influence of different scales or variances: + +$$ +\boldsymbol {R} = \boldsymbol {A} ^ {- 1} \boldsymbol {\Sigma} \boldsymbol {A} ^ {- 1}, \tag {6} +$$ + +$$ +\boldsymbol {A} = \operatorname {d i a g} \left(\sigma_ {1}, \dots , \sigma_ {u}, \dots , \sigma_ {D}\right), \quad \sigma_ {u} = \sqrt {\Sigma_ {u u}} \tag {7} +$$ + +where $\mathbf{A}$ is a diagonal matrix and the diagonal elements are the standard deviation of $D$ variables. The correlation between $u$ -th and $v$ -th variables can then be expressed as: + +$$ +R _ {u v} = \frac {\Sigma_ {u v}}{\sigma_ {u} \sigma_ {v}}, \tag {8} +$$ + +where $\Sigma_{uv}$ represents the covariance between the $u$ -th and $v$ -th variables. With $\pmb{R}$ , we extract $M$ principle co-varying + +relationships across anchor feature dimensions. Specifically, we perform eigendecomposition and decompose the matrix $\pmb{R}$ into the eigenvector matrix $Q\in \mathbb{R}^{D\times D}$ and the diagonal matrix $\Lambda \in \mathbb{R}^{D\times D}$ of eigenvalues: + +$$ +\boldsymbol {R} = \boldsymbol {Q} \boldsymbol {\Lambda} \boldsymbol {Q} ^ {T}, \tag {9} +$$ + +where $Q$ is an orthogonal matrix and each column is an eigenvector representing a co-variation pattern across anchor feature dimensions. We sort the eigenvalues in descending order, where the eigenvector corresponding to the largest eigenvalue captures the most significant co-variation patterns across the feature dimensions. Following this, we select the top- $M$ eigenvectors $\pmb{P} = [P_{1},\dots,P_{M}]$ from $Q$ as the primary directions of co-variation. + +We then leverage $P$ to guide the extraction of textural and structural information in each anchor. As the covarying relationships across anchor feature dimensions are extracted from the anchor set $F^a$ , we treat $P$ as the global-shared co-variations. Combined with the feature $f^a$ in one anchor, we can capture the anchor-specific textures and structures as follows: + +$$ +\boldsymbol {f} _ {i} ^ {\boldsymbol {t}} = F _ {i} \left(\left[ \boldsymbol {P} _ {i}, \boldsymbol {f} ^ {\boldsymbol {a}} \right]\right), \quad i \in [ 1, M ], \tag {10} +$$ + +where $F_{i}(\cdot)$ represents the two-layer MLP used to extract the textures and structures for anchor $\pmb{f}^{a}$ based on the $i$ -th principal co-varying relationship across anchor feature dimensions. Finally, we concatenate the extracted features with the original anchor feature to achieve feature augmentation and predict Gaussian attributes. Take one second-order anchor as an example: + +$$ +\left\{\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \right\} _ {k = 0} ^ {K} = F \left(\boldsymbol {f} ^ {\boldsymbol {a}}, \left\{\boldsymbol {f} _ {\boldsymbol {i}} ^ {\boldsymbol {t}} \right\} _ {i = 0} ^ {M}, \boldsymbol {\delta} _ {a c}, \boldsymbol {d} _ {a c}\right). \tag {11} +$$ + +![](images/73fe3830d156d04d9e113b7695e3479fa11d528ef7f36217c111918be00b3845.jpg) +Figure 3. Visual illustration of the proposed selective gradient loss (SGL). SGL clearly improves the rendering quality by generating finer textures and details. Zoom in for best view. + +Note that $\{\pmb{f}_i^t\}_{i=0}^M$ are derived from $\pmb{f}^a$ for feature augmentation and do not increase the anchor feature dimensions and the size of each anchor. + +# 3.3. Selective Gradient Loss + +In addition to proposing second-order anchors for anchor feature augmentation to achieve superior rendering quality with reduced anchor size, we design a selective gradient loss to further ensure the rendering quality under compact anchor features. It enables the model to adaptively identify difficult-to-render textures and structures, and regularize Gaussian prediction corresponding to these regions. With the designed loss, superior rendering quality with finer textures can be achieved as shown in Fig 3. Given the limited sensitivity of pixel-level L1 loss to textures and structures in RGB images, we consider gradient map to emphasize textural and structural regions. Specifically, we exploit the Sobel operator to extract horizontal and vertical gradient maps $(G_x', G_y', G_x, G_y)$ of the rendering result $I' \in \mathbb{R}^{H \times W \times C}$ and its corresponding ground truth $I$ , which can be expressed as follows: + +$$ +G _ {x} ^ {\prime} = S _ {x} * I ^ {\prime}, \quad G _ {y} ^ {\prime} = S _ {y} * I ^ {\prime}, \tag {12} +$$ + +$$ +G _ {x} = S _ {x} * I, \quad G _ {y} = S _ {y} * I, +$$ + +$$ +S _ {x} = \left[ \begin{array}{l l l} - 1 & 0 & 1 \\ - 2 & 0 & 2 \\ - 1 & 0 & 1 \end{array} \right], \quad S _ {y} = \left[ \begin{array}{l l l} - 1 & - 2 & - 1 \\ 0 & 0 & 0 \\ 1 & 2 & 1 \end{array} \right], \tag {13} +$$ + +where $S_{x}$ and $S_{y}$ denote the horizontal and vertical Sobel kernels, respectively. We then use Euclidean metric to measure the horizontal and vertical gradient discrepancies $(l_{x}, l_{y})$ between the rendering $I'$ and the corresponding ground + +truth $I$ , which are then averaged to derive the final discrepancies as follows: + +$$ +l _ {x} = \frac {1}{\sqrt {H W}} \sum_ {\substack {x = 0 \\ H = 1}} ^ {H - 1} \sum_ {\substack {y = 0 \\ W = 1}} ^ {W - 1} \left| G _ {x} ^ {\prime} (x, y) - G _ {x} (x, y) \right|, \tag{14} +$$ + +$$ +l _ {y} = \frac {1}{\sqrt {H W}} \sum_ {x = 0} ^ {H - 1} \sum_ {y = 0} ^ {W - 1} \left| G _ {y} ^ {\prime} (x, y) - G _ {y} (x, y) \right|. +$$ + +Gradient maps, derived from differences in pixel intensities, typically exhibit meaningful gradients in localized regions, such as edges and areas with significant intensity changes, which are crucial for capturing structures and textures. In contrast, the majority of areas in gradient map comprise flat and low-gradient regions with limited information. Therefore naively adopting the horizontal and vertical gradient discrepancies as the loss function can cause low-gradient regions to dominate the overall loss. As a result, the model may deemphasize regions critical for preserving fine textures and structures, focusing instead on less important flat areas. Besides, the regions with meaningful gradients should also be distinguished, prioritizing greater attention to textures and structures exhibiting larger rendering errors. To address the above issues, we introduce a dynamic region selection that computes the rendering error map as a weight map, defined by the absolute difference between the gradient maps of the rendered result and the ground truth, which can be expressed by: + +$$ +w _ {x} = \left| G _ {x} ^ {\prime} (x, y) - G _ {x} (x, y) \right|, \tag {15} +$$ + +$$ +w _ {y} = \left| G _ {y} ^ {\prime} (x, y) - G _ {y} (x, y) \right|. +$$ + +Combined with the horizontal and vertical weight maps, the selective gradient loss $\mathcal{L}_s$ can be formulated as: + +$$ +\mathcal {L} _ {s} = w _ {x} * l _ {x} + w _ {y} * l _ {y}, \tag {16} +$$ + +Based on the above analysis, the selective gradient loss enables the model to focus on the difficult-to-render textural and structural regions adaptively. Besides, the model is able to adjust the focused regions dynamically in response to the evolving rendering error map throughout the training process. Note, the pixel-level L1 loss with D-SSIM term between RGB images is also used in the training process, which complements the proposed selective gradient loss used between gradient maps. And we also keep the volume regularization as described in Scaffold-GS [25]. The total loss function is given as follows: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {1} + \lambda_ {S S I M} \mathcal {L} _ {S S I M} + \lambda_ {v o l} \mathcal {L} _ {v o l} + \lambda_ {s} \mathcal {L} _ {s}, \tag {17} +$$ + +where the $\lambda_s$ , $\lambda_1$ , $\lambda_{SSIM}$ and $\lambda_{vol}$ represent the training weights of the selective gradient loss $\mathcal{L}_s$ , the L1 loss $\mathcal{L}_1$ , the D-SSIM term $\mathcal{L}_{SSIM}$ and the volume regularization $\mathcal{L}_{vol}$ . + +
DatasetsMip-NeRF360Tanks&TemplesDeep Blending
MethodsSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.80627.500.25232 dim0.85323.960.17732 dim0.90630.210.25432 dim
SOGS(Ours)0.81527.850.22116 dim0.85524.140.17612 dim0.90730.290.25212 dim
+ +Table 1. Quantitative comparisons of novel view synthesis. All methods are trained with the same training data. SOGS achieves superior rendering quality with reduced anchor (model) size. Notably, SOGS using the feature dimension of 12 can still outperform the Scaffold-GS that uses a feature dimension of 32. We do not compare with HAC [9] and ContextGS [35], as they leverage compression techniques to reduce the storage size of Scaffold-GS rather than the actual anchor and model sizes as used in training and rendering. + +
DatasetsBungeeNeRF
MethodsSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.86526.620.24132 dim
SOGS(Ours)0.88027.060.17116 dim
+ +Table 2. Quantitative comparisons on large-scale scenes from the dataset BungeeNeRF [37]. All methods are trained with the same training data. + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +Datasets For both training and testing, we adopt the datasets and configurations in Scaffold-GS [25] and conduct experiments on images of 19 real-world scenes. Specifically, we evaluate SOGS across all nine scenes from the Mip-NeRF360 dataset [2], two scenes from the Tanks&Temples dataset [19], and two additional scenes from the Deep Blending dataset [17]. These scenes exhibit diverse styles, ranging from bounded indoor environments to unbounded outdoor ones. We also evaluate SOGS on the BungeeNeRF dataset [37], which includes six large-scale outdoor scenes, to further validate its effectiveness. The datasets are divided into training and test sets following the same settings as Scaffold-GS, with image resolutions consistent with those used in Scaffold-GS. + +Implementation For the second-order anchor, we reduce the dimension $D$ of the anchor features in Scaffold-GS from 32 to 16, or even to 12. The covariance matrix $\pmb{\Sigma}$ is computed from all anchors in scene representations which captures the global correlation across all anchor feature dimensions. The number of the selected eigenvectors is set to $M = 2$ . All MLPs used to extract anchor-specific textural and structural information consist of two layers with ReLU activation. For the loss function, the training weights $\lambda_{s}, \lambda_{1}, \lambda_{SSIM}$ and $\lambda_{vol}$ are set to 0.01, 0.8, 0.2 and 0.01. Please note, for fair comparisons, we keep the loss terms $\lambda_{1}, \lambda_{SSIM}$ and $\lambda_{vol}$ the same as in Scaffold-GS. The number of + +3D Gaussians corresponding to one second-order anchor is set as $K = 10$ . We use the Pytorch framework to implement SOGS. The SOGS model is trained for 30000 iterations, the same as Scaffold-GS and 3D-GS. + +# 4.2. Comparisons with the State-of-the-Art + +We primarily compare SOGS with Scaffold-GS [25] over datasets Mip-NeRF360, Tank&Temple and Deep Blending as Scaffold-GS is an anchor-based 3D Gaussian splattering method and is most relevant to our study. Table 1 shows experimental results over the same test images as described in Section 4.1. We can observe that SOGS outperforms Scaffold-GS consistently in PSNR, SSIM and LPIPS across all evaluated scenes. The anchor size of SOGS is reduced to 16 in the dataset Mip-NeRF360, while it is challengingly reduced to 12 in Tank&Temple and Deep Blending. The superior performance is largely attributed to our proposed second-order anchor which performs anchor feature augmentation to compensate for the reduced anchor feature size and improve the rendering quality as well as the selective gradient loss that further ensures the high-quality rendering under small-sized anchor features. Further, we benchmark SOGS with Scaffold-GS over the BungeeNeRF dataset which provides diverse large-scale scenes. It can be observed that SOGS achieves superior rendering quality with only 16-dimensional anchors. Besides, as shown in Fig. 4, we perform the visual comparison of SOGS with 3D-GS and Scaffold-GS and provide their model sizes in each scene. With the compact anchor size of the proposed second-order anchor, SOGS achieves a significant reduction in model size while delivering superior novel view synthesis with much less artifacts but finer details. Note we did not compare with HAC [9] and ContextGS [35] as they compress the storage size of Scaffold-GS rather than the actual model size as used in training and rendering. + +# 4.3. Ablation Studies + +We conduct extensive ablation experiments to validate the effectiveness of the proposed second-order anchor and the selective gradient loss. In the experiments, we keep the model size and anchor feature dimension the same as used + +![](images/53c23612a1d9a5daffd7061b8e799cab742e92987764e55c4ad64b14c3562685.jpg) + +![](images/96b30dcd32b63f16effd497d396c876f205a4d4ad03ad69b152c9192db80f121.jpg) + +![](images/a72293882420d3c57150d9da0032643570a007aa6adbe497ed22bfc1d492190a.jpg) + +![](images/3b2957d95718f0ea4c0272843bb69dfbcc070b7f3a5e55d0c993da423aa5bd49.jpg) +3D-GS + +![](images/a2cfbf6e77c9a18fad2ee76ece50a9e09076486ec5edadeb98c93982a621c3e8.jpg) + +![](images/5ec4df95056874e034468d26c2c2937f96ae9c85492e98c3f8b5929864d983fd.jpg) + +![](images/133c90f6d63fbf63b51b02947de23486567dd8a6483adf1b89729ae1be0a4613.jpg) + +![](images/7075f87aec37e98e544c993a7b17a0d66032e401b0a8c96e207dbbb7439e72da.jpg) +Scaffold-GS + +![](images/3f3e643bcc45e8e977e3d0483710782cc7754e5e57234706e558dcb465b63f5c.jpg) + +![](images/dfc67c21b11c7e99623ab4cbd53d3897133e19bb9c387e4e810f11aed9b33022.jpg) + +![](images/9e7847eb86a16524bb93a1c9dbdcb8846307972aa08ec0d7741a00c232a57118.jpg) + +![](images/1e9873211860b4153ce815690bfebab3f8cddf13d16fd8093e37cac18ce44463.jpg) +SOGS (Ours) +Figure 4. Qualitative comparisons of SOGS with 3D-GS [18] and Scaffold-GS [25]. SOGS achieves smaller model size and superior image rendering with much less artifacts but more fine details. The experiments are conducted over multiple indoor and outdoor scenes including 'Garden', 'Flower' and 'Counter' from the Mip-NeRF360 and 'Quebec' from the BungeeNeRF. Zoom in for best view. + +![](images/89f30eda274087615941b9d6507304c94608409ca04d8d8f53c274576608fb60.jpg) + +![](images/360b2830dc0f98fb30f6f1e7d3a668a38b692b491446d28ce61a4b8234e45406.jpg) + +![](images/1168505cd077e2a18611a0f586dd398e6c86a7d36a9b731675c5ea6c14696c67.jpg) + +![](images/c96c271317359f08a98c09b52974db73a06100b361e07a4c3b45de9d0f08101a.jpg) +Ground Truth + +
ModelsEvaluation Metrics
PSNR ↑SSIM ↑LPIPS ↓
Base26.620.8650.241
Base+SOA27.250.8790.208
Base+SOA+SGL27.390.8870.161
+ +Table 3. Ablation studies of SOGS on the dataset BungeeNeRF. With Scaffold-GS as the baseline Base, Base + SOA introduces the second-order anchor to augment anchor features which clearly improves the Gaussian attribute prediction and image rendering. Base + SOA + SGL (i.e., SOGS) further introduces the selective gradient loss which enables the Gaussian prediction to focus on regions with difficult-to-render textures and geometries. All models are trained under the same settings and model size, and the anchor feature dimension is set at 32. + +in Scaffold-GS, aiming to screen out other variations and show how our two designs improve the rendering quality. + +More details are described in the following two subsections. + +Second-order Anchor We first examine how our proposed second-order anchor affects PSNR, SSIM and LPIPS. In this experiment, we adopt the Scaffold-GS as the baseline model Base that performs anchor-based Gaussian splattering with naive anchors. On top of the Base, we train a model Base $+SOA$ that incorporates our proposed second-order anchor to replace the naive anchor in Scaffold-GS. As Table 3 shows, Base $+SOA$ outperforms the Base clearly in PSNR, SSIM and LPIPS, indicating that the second-order anchor can effectively improve the Gaussian attribute prediction and novel view synthesis by anchor feature augmentation. + +Selective Gradient Loss To evaluate how the proposed selective gradient loss contributes, we train a new model $Base + SOA + SGL$ (i.e., the complete SOGS) that incorporates the selective gradient loss on top of $Base + SOA$ . Table 3 shows experimental results. We can observe that + +![](images/0a44e6b9f7d31a0aec7f975cf9bb2e3fb338af55453a385db6000b0bbf82c3b6.jpg) +Figure 5. Anchor feature dimension vs SOGS performance: The performance of SOGS varies with the feature dimension $D$ in each anchor, where increasing $D$ improves SOGS consistently together with the increased model size and computational cost. The graph shows the PSNR of the scene 'Bicycle' from MipNeRF360. + +Base $+SOA + SGL$ further improves the novel view synthesis consistently across all three evaluation metrics, demonstrating its effectiveness on Gaussian attribute prediction and rendering quality. + +# 4.4. Parameter Investigation + +We examine how the dimension of anchor features $D$ affects the SOGS performance. Fig. 5 shows experiments on the scene 'Bicycle' from the dataset MipNeRF360. We can observe that the performance of SOGS (in PSNR ↑ score) improves consistently while the feature dimension $D$ increases. Specifically, the feature dimension plays a vital role in the rendering quality as it sets the upper bound for the encoded information as well as the effectiveness of feature augmentation in SOGS. As $D$ becomes larger than 16, the performance gains gradually diminish due to feature saturation while the model size and computational costs grow greatly. We therefore set $D$ at 12 or 16 in our implemented system to balance the model size and performance. + +# 4.5. Visualization + +We visualize the impact of second-order anchors on scene textures and structures. As Fig. 6 shows, our proposed second-order anchor can enhance the rendered images with finer structures and textures. This visualization verifies that the proposed second-order anchor can effectively capture the structural and textural patterns for anchor feature augmentation, and facilitate subsequent Gaussian prediction and image rendering with augmented features. + +# 5. Conclusion + +This paper presents SOGS, an innovative anchor-based 3D Gaussian splatting technique that incorporates second-order anchors to achieve superior rendering with compact anchor feature dimensions and model size. Specifically, by introducing covariance-based second-order statistics, SOGS learns correlations across anchor feature dimensions to capture anchor-specific textures and structural patterns. This + +![](images/a136b8b1f3c5ce0d7c08a6cb921cd9440b90b678064615716bfc31cc1adfbaa0.jpg) + +![](images/6fad434b231629a5c57501163d2a051911c9aa09d97cb21a947ed987e002633f.jpg) + +![](images/4f84bb157558e17130ab51bbbe2616df2f27b280437e15c0d3500882c8b76707.jpg) + +![](images/70f9d1682c5393f26b595da4baadeeac80d3df8b549d3dd748d789434a92bf8b.jpg) +Naive Anchor + +![](images/05b4a95019e66465ff70fa9a3fd1b5a35086f2e3d69c44c0234799b77d9c88af.jpg) + +![](images/6d1c9d73bf3aec43edcabcb1f9c52d9cea8ffe3291312a7592fcce6aa7aa960a.jpg) + +![](images/15b448cf857c0ff60b27b49db5d2539bc9e95568ea35f567b068decc45488e1d.jpg) + +![](images/dda097a16269f91c2f01ba6d7c5a870ee937b809edb77660122c0626e86179f8.jpg) +Second-Order Anchor +Figure 6. Visualization of the impact of second-order anchors on scene textures and structures. The two samples are 'Drjohnson' and 'Playroom' from dataset Deep Blending [17]. The proposed second-order anchor improves the rendering quality with superior structures and textures by anchor feature augmentation. Zoom in for best view. + +enables each anchor to achieve feature augmentation, compensating for the reduced anchor feature size and resulting in superior Gaussian prediction and rendering quality. Besides, we also design a selective gradient loss to further ensure high-quality rendering under compact anchor features, which enables the model to dynamically focus on the Gaussian prediction for difficult-to-render textural and structural regions. Experiments over multiple widely adopted indoor and outdoor scenes show that SOGS achieves superior novel view synthesis with compact model size. Although time complexity increases slightly due to the additional computations in second-order anchors, SOGS still achieves efficient training and real-time rendering. + +# 6. Acknowledgements + +This project is funded by the Ministry of Education Singapore, under the Tier-2 project scheme with a project number MOE-T2EP20220-0003. Fangneng Zhan is supported by the NSF grant CRCNS-2309041. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 1, 2 +[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2 +[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3 +[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2 +[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[8] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2 +[9] Yihang Chen, Qianyi Wu, Jianfei Cai, Mehrtash Harandi, and Weiyao Lin. Hac: Hash-grid assisted context for 3d gaussian splattering compression. arXiv preprint arXiv:2403.14530, 2024. 2, 3, 6 +[10] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 3 +[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 3 +[12] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 14324-14334, 2021. 2 +[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3 +[14] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2 +[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2 +[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3 +[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 8 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 7 +[19] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6 +[20] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3 +[21] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2 +[22] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2 +[23] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2 +[24] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. Stylegaussian: Instant 3d style transfer with gaussian splatting. In SIGGRAPH Asia 2024 Technical Communications, pages 1-4. 2024. 3 +[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 3, 5, 6, 7 +[26] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2 +[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2 +[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2 +[30] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3 +[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2 +[32] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2 +[33] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2 +[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2 +[35] Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex C Kot, and Bihan Wen. Contextgs: Compact 3d gaussian splatting with anchor level context model. arXiv preprint arXiv:2405.20721, 2024. 3, 6 +[36] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3 + +[37] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In European conference on computer vision, pages 106-122. Springer, 2022. 2, 6 +[38] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 3 +[39] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 3 +[40] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2 +[41] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3 +[42] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3 +[43] Jiahui Zhang, Fangneng Zhan, Rongliang Wu, Yingchen Yu, Wenqing Zhang, Bai Song, Xiaqin Zhang, and Shijian Lu. Vmrf: View matching neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6579-6587, 2022. 2 +[44] Jiahui Zhang, Fangneng Zhan, Yingchen Yu, Kunhao Liu, Rongliang Wu, Xiaoqin Zhang, Ling Shao, and Shijian Lu. Pose-free neural radiance fields via implicit pose regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3534-3543, 2023. 2 +[45] Jiahui Zhang, Fangneng Zhan, Muyu Xu, Shijian Lu, and Eric Xing. Fregs: 3d gaussian splatting with progressive frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21424-21433, 2024. 3 +[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1 +[47] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 3 \ No newline at end of file diff --git a/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/images.zip b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..344fca77fb8746600b6a41cd753702ad51521952 --- /dev/null +++ b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc963cbc84807e7d335d8cf8497faa18cda309abc1aec2b117b22226314566c8 +size 822092 diff --git a/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/layout.json b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30855e05c5d47c16f4e4e07febc9e64eedbf3bfe --- /dev/null +++ b/2025/SOGS_ Second-Order Anchor for Advanced 3D Gaussian Splatting/layout.json @@ -0,0 +1,10063 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 512, + 121 + ], + "type": "text", + "content": "SOGS: Second-Order Anchor for Advanced 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "text", + "content": "Jiahui Zhang" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "text", + "content": " Fangneng Zhan" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "text", + "content": " Ling Shao" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "text", + "content": " Shijian Lu" + }, + { + "bbox": [ + 149, + 142, + 459, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 468, + 171 + ], + "type": "text", + "content": "Nanyang Technological University 2Harvard University MIT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 172, + 477, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 172, + 477, + 186 + ], + "spans": [ + { + "bbox": [ + 136, + 172, + 477, + 186 + ], + "type": "text", + "content": "4UCAS-Terminus AI Lab, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 180, + 188, + 435, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 188, + 435, + 199 + ], + "spans": [ + { + "bbox": [ + 180, + 188, + 435, + 199 + ], + "type": "text", + "content": "jiahui003@e.ntu.edu.sg fnzhan@seasharvard.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 202, + 414, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 202, + 414, + 213 + ], + "spans": [ + { + "bbox": [ + 192, + 202, + 414, + 213 + ], + "type": "text", + "content": "ling.shao@ieee.org shijian.lu@ntu.edu.sg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 255, + 179, + 335 + ], + "blocks": [ + { + "bbox": [ + 88, + 242, + 151, + 252 + ], + "lines": [ + { + "bbox": [ + 88, + 242, + 151, + 252 + ], + "spans": [ + { + "bbox": [ + 88, + 242, + 151, + 252 + ], + "type": "text", + "content": "Rendered Images" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 255, + 179, + 335 + ], + "lines": [ + { + "bbox": [ + 57, + 255, + 179, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 179, + 335 + ], + "type": "image", + "image_path": "c7f57ee65b0bd04d32da9829fb266809b65b70d3b3762ab5998ddc08e2a32997.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 180, + 255, + 302, + 335 + ], + "blocks": [ + { + "bbox": [ + 212, + 242, + 266, + 252 + ], + "lines": [ + { + "bbox": [ + 212, + 242, + 266, + 252 + ], + "spans": [ + { + "bbox": [ + 212, + 242, + 266, + 252 + ], + "type": "text", + "content": "Gradient Maps" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 255, + 302, + 335 + ], + "lines": [ + { + "bbox": [ + 180, + 255, + 302, + 335 + ], + "spans": [ + { + "bbox": [ + 180, + 255, + 302, + 335 + ], + "type": "image", + "image_path": "e4a5b0d7d869538bad3dcaa45b21349fa5b8d86897523387b6860fdf5d4b4950.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 255, + 429, + 335 + ], + "blocks": [ + { + "bbox": [ + 335, + 242, + 399, + 252 + ], + "lines": [ + { + "bbox": [ + 335, + 242, + 399, + 252 + ], + "spans": [ + { + "bbox": [ + 335, + 242, + 399, + 252 + ], + "type": "text", + "content": "Rendered Images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 255, + 429, + 335 + ], + "lines": [ + { + "bbox": [ + 308, + 255, + 429, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 429, + 335 + ], + "type": "image", + "image_path": "323fb34a19344feeb6abf1d3bbf31d0b3b9bd8d4662d3cc0908bfec89b75f02e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 432, + 255, + 553, + 335 + ], + "blocks": [ + { + "bbox": [ + 462, + 242, + 517, + 252 + ], + "lines": [ + { + "bbox": [ + 462, + 242, + 517, + 252 + ], + "spans": [ + { + "bbox": [ + 462, + 242, + 517, + 252 + ], + "type": "text", + "content": "Gradient Maps" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 255, + 553, + 335 + ], + "lines": [ + { + "bbox": [ + 432, + 255, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 432, + 255, + 553, + 335 + ], + "type": "image", + "image_path": "f2358e0fade5230fc7f40e4e0f511c059525c51757329926861deef1041edae4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 57, + 337, + 179, + 418 + ], + "blocks": [ + { + "bbox": [ + 57, + 337, + 179, + 418 + ], + "lines": [ + { + "bbox": [ + 57, + 337, + 179, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 337, + 179, + 418 + ], + "type": "image", + "image_path": "07831be65495cb6054194bbcea4b71fac52e4e152eb5529631addaae611073d3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 422, + 211, + 432 + ], + "lines": [ + { + "bbox": [ + 150, + 422, + 211, + 432 + ], + "spans": [ + { + "bbox": [ + 150, + 422, + 211, + 432 + ], + "type": "text", + "content": "(a)Scaffold-GS" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 179, + 337, + 302, + 418 + ], + "blocks": [ + { + "bbox": [ + 179, + 337, + 302, + 418 + ], + "lines": [ + { + "bbox": [ + 179, + 337, + 302, + 418 + ], + "spans": [ + { + "bbox": [ + 179, + 337, + 302, + 418 + ], + "type": "image", + "image_path": "1cb42ed6dc80f7c1c643dc50d1e230a45b2763600f0e2fe06ccde40aec20736b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 555, + 485 + ], + "type": "text", + "content": "Figure 1. The proposed SOGS can render high-quality textures and geometries and reduce model size simultaneously. The illustrations in (a) and (b) show Rendered Images and Gradient Maps which are produced by Scaffold-GS [25] and SOGS, respectively, for the samples 'Stump' and 'Room' from Mip-NeRF360 [2]. The Gradient Maps are extracted with the Sobel operator which highlight the texture and geometry of imaged scenes." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 308, + 337, + 429, + 418 + ], + "blocks": [ + { + "bbox": [ + 308, + 337, + 429, + 418 + ], + "lines": [ + { + "bbox": [ + 308, + 337, + 429, + 418 + ], + "spans": [ + { + "bbox": [ + 308, + 337, + 429, + 418 + ], + "type": "image", + "image_path": "7e132f0cb814e5b2b047bf791228eeb1c1b93f0963b73449acc06e76dc3d087a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 422, + 449, + 432 + ], + "lines": [ + { + "bbox": [ + 411, + 422, + 449, + 432 + ], + "spans": [ + { + "bbox": [ + 411, + 422, + 449, + 432 + ], + "type": "text", + "content": "(b) SOGS" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 151, + 495, + 200, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 495, + 200, + 508 + ], + "spans": [ + { + "bbox": [ + 151, + 495, + 200, + 508 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 522, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 297, + 689 + ], + "type": "text", + "content": "Anchor-based 3D Gaussian splatting (3D-GS) exploits anchor features in 3D Gaussian prediction, which has achieved impressive 3D rendering quality with reduced Gaussian redundancy. On the other hand, it often encounters the dilemma among anchor features, model size, and rendering quality - large anchor features lead to large 3D models and high-quality rendering whereas reducing anchor features degrades Gaussian attribute prediction which leads to clear artifacts in the rendered textures and geometries. We design SOGS, an anchor-based 3D-GS technique that introduces second-order anchors to achieve superior rendering quality and reduced anchor features and model size simultaneously. Specifically, SOGS incorporates covariance-based second-order statistics and corre" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 497, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 605 + ], + "type": "text", + "content": "lation across feature dimensions to augment features within each anchor, compensating for the reduced feature size and improving rendering quality effectively. In addition, it introduces a selective gradient loss to enhance the optimization of scene textures and scene geometries, leading to high-quality rendering with small anchor features. Extensive experiments over multiple widely adopted benchmarks show that SOGS achieves superior rendering quality in novel view synthesis with clearly reduced model size." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 632, + 395, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 632, + 395, + 644 + ], + "spans": [ + { + "bbox": [ + 314, + 632, + 395, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "3D scene representation and rendering have been a pivotal task in 3D computer vision, playing a crucial role in various applications such as virtual reality and scene simulation. Neural radiance fields (NeRFs) [27] and its variants [1-3, 46] have been developed to model implicit" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 197, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 197, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 197, + 713 + ], + "type": "text", + "content": "*Shijian Lu is the corresponding author." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11167" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "scene representation and enable volume rendering for high-quality novel view synthesis. However, NeRF is limited by long training and rendering times, largely due to the time-consuming point sampling in volume rendering. Recently, 3D Gaussian Splatting (3D-GS) [18] has been proposed as an alternative which learns explicit 3D scene representations with learnable 3D Gaussians. By leveraging efficient splatting and rasterization, 3D-GS projects 3D Gaussians onto a 2D plane, enabling real-time rendering. However, 3D-GS stacks substantial 3D Gaussians to fit each training view individually, often overlooking the underlying scene structure [25]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 218, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 433 + ], + "type": "text", + "content": "The recent Scaffold-GS [25] exploits scene geometry to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. Specifically, it introduces anchor points to store features and exploits MLPs to predict Gaussian attributes from these anchor features. However, Scaffold-GS struggles to strike a balance between rendering quality and model size. Specifically, adopting large anchor features improves rendering quality [9] but it significantly enlarges model size due to the large number of anchor points used in scene representations. Using smaller anchor features can shrink the model, but it clearly degrades the prediction of Gaussian attributes which further leads to suboptimal rendering with various artifacts in rendered textures and geometries. HAC [9] introduces grid-based context modeling and entropy encoding to compress the Scaffold-GS. However, HAC does not reduce the model size but the storage size of the trained Scaffold-GS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 602 + ], + "type": "text", + "content": "We design SOGS, an innovative anchor-based 3D-GS technique that introduces second-order anchors and achieves superior rendering quality with reduced anchor feature dimensions and model size. Specifically, SOGS incorporates covariance-based second-order statistics to model correlation across anchor feature dimensions. The rationale is that textures and geometries are defined by not only individual features but also the correlation across features. Hence, the correlation across anchor features can effectively guide to capture the pattern of intricate local textures and structures in each anchor. This nice feature enables anchor feature augmentation, empowering SOGS to achieve superior rendering quality with reduced feature dimensions and feature size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Additionally, we design a selective gradient loss to enhance the rendering quality with small anchor features. This loss computes gradient maps from both the rendered image and the corresponding ground truth that highlight their differences in scene textures and structures. With the difference of the gradient maps, the 3D Gaussian prediction can adaptively focus on image regions where the texture and geometry are hard to render. This enables dynamic region selection along the training process, where the model can" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "adjust the focused image regions dynamically according to the evolving errors between the gradient maps of the rendered image and the ground truth. Experiments show that SOGS simultaneously achieves superior rendering quality in novel view synthesis as well as clearly reduced model sizes as shown in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 145, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 337 + ], + "type": "text", + "content": "The major contributions of this work can be summarized in three aspects. First, we propose SOGS, an innovative anchor-based 3D-GS that introduces second-order anchors to balance the rendering quality and model size. With covariance-based second-order statistics that model correlation across anchor feature dimensions to capture intricate textural and structural patterns, SOGS achieves anchor feature augmentation which produces superior rendering quality with significantly reduced anchor feature dimensions. Second, we design a selective gradient loss that guides 3D Gaussian prediction to focus on difficult-to-render textures and structures adaptively, further enhancing rendering quality with small anchor features Third, experiments over multiple benchmarks show that SOGS achieves superior novel view synthesis with efficient anchor size and model size, and outperforms the state-of-the-art in rendering quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 350, + 400, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 400, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 400, + 361 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 371, + 515, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 515, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 515, + 384 + ], + "type": "text", + "content": "2.1. Radiance Field and Neural Rendering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 389, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 553, + 628 + ], + "type": "text", + "content": "Radiance field has been widely explored for novel view synthesis and achieved impressive rendering quality. Neural radiance field (NeRF) [27] utilizes MLPs to learn implicit 3D scene representations from multi-view posed 2D images. Coupled with differentiable volume rendering, NeRF achieves high-quality rendering with superb multiview consistency. Several NeRF variants have been developed to address various new challenges, including largescale scenes [32, 33, 37], dynamic scenes [12, 14, 31], few-shot setting [6, 34, 40], pose-free setting [4, 8, 22, 43, 44], and antialIASing [2, 3]. However, NeRF and its variants still suffer from extremely long training and rendering times due to the time-consuming sampling required for volume rendering. Several studies [7, 15, 28, 29] have been proposed to address this issue. For instance, Müller et al. [28] introduce multi-resolution hash encoding to reduce the size of neural networks, significantly shortening the training time and enabling real-time rendering in tens of milliseconds. However, these methods often suffer from degraded reconstruction accuracy and rendering quality of scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Kerbl et al. recently propose 3D Gaussian splatting (3D-GS) [18], which achieves high-quality and real-time rendering by explicitly representing scenes by using parameterized 3D Gaussians, along with efficient splatting and rasterization. Building on its excellent performance, numerous 3D-GS variants have been developed to address diverse challenges and tasks, such as dynamic scenes [21, 23, 26," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11168" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "36, 41], autonomous driving [38, 47], rendering quality optimization [11, 39, 42, 45], sparse-view setting [5, 10, 20], colmap-free setting [13], and 3D style transfer [24]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 251, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 251, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 251, + 131 + ], + "type": "text", + "content": "2.2. Anchor-based 3D Gaussian Splitting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 296, + 340 + ], + "type": "text", + "content": "One major constraint in 3D-GS is that it stacks a large number of 3D Gaussians to match each training view separately, often overlooking the underlying scene geometry. To address this issue, Lu et al. [25] present Scaffold-GS, the first anchor-based 3D-GS technique that incorporates anchor points to predict Gaussian attributes through multilayer perceptrons (MLPs). Scaffold-GS leverages scene structure to guide the distribution of 3D Gaussians, reducing Gaussian redundancy and delivering impressive 3D rendering performance. However, Scaffold-GS faces challenges in balancing enhanced rendering quality with anchor and model size reduction. Although several follow-ups [9, 35] introduce context-based entropy encoding to compress the Scaffold-GS, they reduce its storage size instead of anchor and model sizes as used during training and rendering. The proposed SOGS introduces second-order anchors, enabling superior rendering quality with reduced model size." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 162, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 162, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 162, + 366 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 529 + ], + "type": "text", + "content": "We propose SOGS, a novel anchor-based 3D Gaussian splatting that introduces second-order anchors and selective gradient loss to achieve superior rendering quality with reduced anchor and model sizes. Fig. 2 shows the overview of SOGS. The 3D Gaussian splatting [18] (3D-GS) and Scaffold-GS [25] (the base model for our SOGS) are briefly introduced in Sec. 3.1. In Sec. 3.2, we first explain why the proposed second-order anchor enables the model to enhance rendering quality in reduced anchor and model sizes, and then describe the details of the second-order anchor. Besides, to further ensure high-quality image rendering under compact anchor size, we design a selective gradient loss, detailed in Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 141, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 141, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 141, + 551 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "text", + "content": "3D Gaussian Splatting (3D-GS). 3D-GS [18] explicitly represents scenes using anisotropic 3D Gaussians and enables real-time rendering through efficient differentiable splatting. 3D Gaussians are initialized from point clouds generated by structure-from-motion [16, 30]. Each Gaussian is defined by a covariance matrix, a position (mean), an opacity value " + }, + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "text", + "content": ", and spherical harmonics coefficients representing color " + }, + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 556, + 295, + 676 + ], + "type": "text", + "content": ", where the covariance matrix is decomposed into scaling matrix and rotation matrix to facilitate differentiable optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "For rendering, 3D Gaussians are projected onto a 2D plane through splatting, followed by " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "-blending. Specifically, the color " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " of a pixel is computed by blending " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " or" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 549, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 549, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 549, + 84 + ], + "type": "text", + "content": "dered 2D Gaussians overlapping that pixel, formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 373, + 92, + 554, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 92, + 554, + 125 + ], + "spans": [ + { + "bbox": [ + 373, + 92, + 554, + 125 + ], + "type": "interline_equation", + "content": "C = \\sum_ {n \\in N} c _ {n} \\alpha_ {n} \\prod_ {z = 1} ^ {n - 1} (1 - \\alpha_ {z}), \\tag {1}", + "image_path": "3a5c04ac682abb07bc57acdab6947893e2027e4d96bfc425d0567cda0a75eda7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "text", + "content": " represent the color and opacity of the " + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 133, + 554, + 157 + ], + "type": "text", + "content": "-th 2D Gaussian." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": "Scaffold-GS. Scaffold-GS [25] introduces anchor points to predict 3D Gaussians via MLPs. It aims to leverage scene geometry to guide the distribution and attributes of 3D Gaussians and reduce the Gaussian redundancy. Specifically, Scaffold-GS voxelizes the entire scene based on the point cloud initialized by COLMAP [30] and assigns the center of each voxel as an anchor point. A local scene feature " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{\\pmb{a}} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": ", a scaling factor " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "l^{\\pmb{a}}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " offsets " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "o^{\\pmb{a}} \\in \\mathbb{R}^{3K}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " are stored in each anchor as the anchor attributes. Given an anchor position " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": ", its relative distance " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\delta_{ac}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "d_{ac}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " to a camera position " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " can be represented as: " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\delta_{ac} = ||x_{a} - x_{c}||_{2}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "d_{ac} = \\frac{x_{a} - x_{c}}{||x_{a} - x_{c}||_{2}}" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": ". The prediction of " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " Gaussians from one anchor point using MLPs " + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 171, + 555, + 340 + ], + "type": "text", + "content": " can then be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 348, + 555, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 348, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 353, + 348, + 555, + 364 + ], + "type": "interline_equation", + "content": "\\left\\{\\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \\right\\} _ {k = 1} ^ {K} = F (\\boldsymbol {f} ^ {\\boldsymbol {a}}, \\boldsymbol {\\delta} _ {a c}, \\boldsymbol {d} _ {a c}), \\tag {2}", + "image_path": "14838d1b56f2dfd94939f68e290cb0d86d680a53b572b3a186d650c62d298ea9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\alpha_{k}, c_{k}, s_{k}" + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "inline_equation", + "content": "q_{k}" + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "content": " denote the opacity, color, scaling and the rotation-related quaternion of the " + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "content": "-th Gaussian, respectively. The positions of the " + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 371, + 554, + 418 + ], + "type": "text", + "content": " Gaussians are calculated by this formula:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 426, + 554, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 426, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 369, + 426, + 554, + 441 + ], + "type": "interline_equation", + "content": "\\left\\{p _ {k} \\right\\} _ {k = 1} ^ {K} = \\boldsymbol {x} _ {a} + \\left\\{o _ {k} ^ {a} \\right\\} _ {k = 1} ^ {K} * \\boldsymbol {l} ^ {\\boldsymbol {a}}. \\tag {3}", + "image_path": "fbca4a5f0be9129e267f811144a6b8f70d292e6cad630f29793537c9ebbad716.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 449, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 474 + ], + "type": "text", + "content": "Scaffold-GS follows the same rendering process as Eq. 1 described in 3D-GS after predicting the Gaussian attributes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 441, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 441, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 441, + 492 + ], + "type": "text", + "content": "3.2. Second-Order Anchor" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 688 + ], + "type": "text", + "content": "In this section, we first explain why the proposed second-order anchor enables the model to achieve superior image rendering with compact anchor size. The key lies in leveraging covariance-based second-order feature statistics to achieve anchor feature augmentation. Specifically, second-order feature statistics extract co-varying relationships across anchor feature dimensions. As textures and structures are defined not only by individual features but also by their interdependencies, these co-varying relationships can guide the capture of intricate local textural and structural patterns within each anchor, facilitating feature augmentation. As a result, SOGS compensates for the reduced feature dimensions and ensures superior rendering quality by performing feature augmentation. We provide the visualization of the impact of second-order anchors on scene textures and structures in the Sec. 4.5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Given anchor feature set " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathbf{a}}\\in \\mathbb{R}^{N\\times D} = [f_1^a,f_2^a,\\dots,f_N^a ]" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " denote the number of" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11169" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 230 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 230 + ], + "type": "image", + "image_path": "d42c671027edd64f2e7d10d34cf3fc38412f8c15b1db14e2ab326a84f8e69db9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed SOGS. Initialized from point clouds, each anchor stores an anchor feature vector " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "f^a \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": ". Putting " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " along " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " dimensions as " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " variables and all anchors as observed samples, the second-order anchor statistics capturing covarying relations across " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " can be computed from the covariance matrix " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " dimensions. We select the top " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " eigenvectors " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " as the most significant co-variation patterns and combine " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 54, + 247, + 555, + 316 + ], + "type": "text", + "content": " to capture anchor-specific textures and geometries for anchor feature augmentation, second-order anchor construction and Gaussian prediction. With the gradient maps of the rendered image and the ground truth, the selective gradient loss guides to learn to render finer textures and geometries with dynamic region selection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "text", + "content": "anchors and the dimensionality of each anchor feature, we first compute second-order statistics in the anchor feature set. Specifically, we treat " + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "text", + "content": " feature dimensions (channels) as " + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "text", + "content": " variables and utilize the anchor set as " + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "text", + "content": " observation samples, and compute the covariance matrix " + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{D\\times D}" + }, + { + "bbox": [ + 55, + 336, + 296, + 407 + ], + "type": "text", + "content": " across these channels, which can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 415, + 295, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 415, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 96, + 415, + 295, + 440 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\frac {1}{N - 1} \\left(\\boldsymbol {F} ^ {\\boldsymbol {a}} - \\boldsymbol {\\mu} ^ {T}\\right) \\left(\\boldsymbol {F} ^ {\\boldsymbol {a}} - \\boldsymbol {\\mu} ^ {T}\\right) ^ {T}, \\tag {4}", + "image_path": "65073f0cfe3182a9293b79beb7615435866e97388ee31248638e96bccaa68e4d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 449, + 295, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 449, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 149, + 449, + 295, + 471 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} = \\frac {1}{N} \\boldsymbol {F} ^ {\\boldsymbol {a}}, \\tag {5}", + "image_path": "d8fc35581fb53ed939aa49f999a6a1fe770391340d14e4376482f84119cd29c1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " denotes the mean vector across " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " samples, used for feature centering. Given that two variables with large variances can yield a large covariance despite only weak relationship, we then construct a correlation matrix " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " from the covariance matrix " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " to standardize the relationships among " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " variables, eliminating the influence of different scales or variances:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 137, + 569, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 569, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 137, + 569, + 295, + 583 + ], + "type": "interline_equation", + "content": "\\boldsymbol {R} = \\boldsymbol {A} ^ {- 1} \\boldsymbol {\\Sigma} \\boldsymbol {A} ^ {- 1}, \\tag {6}", + "image_path": "52dbf03702335f0ee4a23c120daf5fa95c360e7f669878edac6be8244f1c6752.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 592, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 592, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 83, + 592, + 295, + 607 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} = \\operatorname {d i a g} \\left(\\sigma_ {1}, \\dots , \\sigma_ {u}, \\dots , \\sigma_ {D}\\right), \\quad \\sigma_ {u} = \\sqrt {\\Sigma_ {u u}} \\tag {7}", + "image_path": "dd9d5cfa28d736494ae47ea75b0793ae815557c77a3e1accec524c5da2962021.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "content": " is a diagonal matrix and the diagonal elements are the standard deviation of " + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "content": " variables. The correlation between " + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "content": "-th variables can then be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 656, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 656, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 145, + 656, + 295, + 682 + ], + "type": "interline_equation", + "content": "R _ {u v} = \\frac {\\Sigma_ {u v}}{\\sigma_ {u} \\sigma_ {v}}, \\tag {8}", + "image_path": "e11eb5f04b014c719ed32c16e273d910f93d29624a9974f6ee2248d6595f77aa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma_{uv}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " represents the covariance between the " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "-th variables. With " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ", we extract " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " principle co-varying" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": "relationships across anchor feature dimensions. Specifically, we perform eigendecomposition and decompose the matrix " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " into the eigenvector matrix " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "Q\\in \\mathbb{R}^{D\\times D}" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " and the diagonal matrix " + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "inline_equation", + "content": "\\Lambda \\in \\mathbb{R}^{D\\times D}" + }, + { + "bbox": [ + 313, + 336, + 554, + 384 + ], + "type": "text", + "content": " of eigenvalues:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 405, + 388, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 388, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 405, + 388, + 553, + 403 + ], + "type": "interline_equation", + "content": "\\boldsymbol {R} = \\boldsymbol {Q} \\boldsymbol {\\Lambda} \\boldsymbol {Q} ^ {T}, \\tag {9}", + "image_path": "f207d5d046629969b43e834961d400972a2601e86a88265cd2ead16119a48e70.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "text", + "content": " is an orthogonal matrix and each column is an eigenvector representing a co-variation pattern across anchor feature dimensions. We sort the eigenvalues in descending order, where the eigenvector corresponding to the largest eigenvalue captures the most significant co-variation patterns across the feature dimensions. Following this, we select the top-" + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "text", + "content": " eigenvectors " + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{P} = [P_{1},\\dots,P_{M}]" + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 407, + 554, + 502 + ], + "type": "text", + "content": " as the primary directions of co-variation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "text", + "content": "We then leverage " + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "text", + "content": " to guide the extraction of textural and structural information in each anchor. As the covarying relationships across anchor feature dimensions are extracted from the anchor set " + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "inline_equation", + "content": "F^a" + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "text", + "content": ", we treat " + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "text", + "content": " as the global-shared co-variations. Combined with the feature " + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 313, + 504, + 554, + 586 + ], + "type": "text", + "content": " in one anchor, we can capture the anchor-specific textures and structures as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 366, + 592, + 553, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 592, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 366, + 592, + 553, + 607 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} ^ {\\boldsymbol {t}} = F _ {i} \\left(\\left[ \\boldsymbol {P} _ {i}, \\boldsymbol {f} ^ {\\boldsymbol {a}} \\right]\\right), \\quad i \\in [ 1, M ], \\tag {10}", + "image_path": "424ac7cd52b09743d97f7155e0481d2c5facdd4e48821617c7c55f3ac7b0aebe.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "inline_equation", + "content": "F_{i}(\\cdot)" + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "text", + "content": " represents the two-layer MLP used to extract the textures and structures for anchor " + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{a}" + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "text", + "content": " based on the " + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 611, + 554, + 695 + ], + "type": "text", + "content": "-th principal co-varying relationship across anchor feature dimensions. Finally, we concatenate the extracted features with the original anchor feature to achieve feature augmentation and predict Gaussian attributes. Take one second-order anchor as an example:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 326, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\left\\{\\alpha_ {k}, c _ {k}, q _ {k}, s _ {k} \\right\\} _ {k = 0} ^ {K} = F \\left(\\boldsymbol {f} ^ {\\boldsymbol {a}}, \\left\\{\\boldsymbol {f} _ {\\boldsymbol {i}} ^ {\\boldsymbol {t}} \\right\\} _ {i = 0} ^ {M}, \\boldsymbol {\\delta} _ {a c}, \\boldsymbol {d} _ {a c}\\right). \\tag {11}", + "image_path": "81fd62ec9f264b5531ffd65e03de1c67d2f6ca974a6e44c7f6ef8bdc71156b76.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11170" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 296, + 240 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 240 + ], + "type": "image", + "image_path": "73fe3830d156d04d9e113b7695e3479fa11d528ef7f36217c111918be00b3845.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 293 + ], + "type": "text", + "content": "Figure 3. Visual illustration of the proposed selective gradient loss (SGL). SGL clearly improves the rendering quality by generating finer textures and details. Zoom in for best view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_i^t\\}_{i=0}^M" + }, + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "type": "text", + "content": " are derived from " + }, + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\pmb{f}^a" + }, + { + "bbox": [ + 55, + 315, + 296, + 352 + ], + "type": "text", + "content": " for feature augmentation and do not increase the anchor feature dimensions and the size of each anchor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 189, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 189, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 189, + 372 + ], + "type": "text", + "content": "3.3. Selective Gradient Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "text", + "content": "In addition to proposing second-order anchors for anchor feature augmentation to achieve superior rendering quality with reduced anchor size, we design a selective gradient loss to further ensure the rendering quality under compact anchor features. It enables the model to adaptively identify difficult-to-render textures and structures, and regularize Gaussian prediction corresponding to these regions. With the designed loss, superior rendering quality with finer textures can be achieved as shown in Fig 3. Given the limited sensitivity of pixel-level L1 loss to textures and structures in RGB images, we consider gradient map to emphasize textural and structural regions. Specifically, we exploit the Sobel operator to extract horizontal and vertical gradient maps " + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "inline_equation", + "content": "(G_x', G_y', G_x, G_y)" + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "text", + "content": " of the rendering result " + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "inline_equation", + "content": "I' \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "text", + "content": " and its corresponding ground truth " + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 378, + 296, + 571 + ], + "type": "text", + "content": ", which can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 580, + 295, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 580, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 110, + 580, + 295, + 600 + ], + "type": "interline_equation", + "content": "G _ {x} ^ {\\prime} = S _ {x} * I ^ {\\prime}, \\quad G _ {y} ^ {\\prime} = S _ {y} * I ^ {\\prime}, \\tag {12}", + "image_path": "368659db2d3b3055812240ffb44b79efc823e5a0225212896a309348c5a0626c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 598, + 240, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 598, + 240, + 610 + ], + "spans": [ + { + "bbox": [ + 117, + 598, + 240, + 610 + ], + "type": "interline_equation", + "content": "G _ {x} = S _ {x} * I, \\quad G _ {y} = S _ {y} * I,", + "image_path": "3a24f29f95508b6cef0a22ac939371cb77ce972bbdd7e39336f8f0988ad44486.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 621, + 296, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 621, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 67, + 621, + 296, + 661 + ], + "type": "interline_equation", + "content": "S _ {x} = \\left[ \\begin{array}{l l l} - 1 & 0 & 1 \\\\ - 2 & 0 & 2 \\\\ - 1 & 0 & 1 \\end{array} \\right], \\quad S _ {y} = \\left[ \\begin{array}{l l l} - 1 & - 2 & - 1 \\\\ 0 & 0 & 0 \\\\ 1 & 2 & 1 \\end{array} \\right], \\tag {13}", + "image_path": "50306bc9ad8ec13e07d962f1ea07c831e999b989404fb3381b59be49b5b9d494.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{y}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " denote the horizontal and vertical Sobel kernels, respectively. We then use Euclidean metric to measure the horizontal and vertical gradient discrepancies " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "(l_{x}, l_{y})" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " between the rendering " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and the corresponding ground" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "truth " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": ", which are then averaged to derive the final discrepancies as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 102, + 553, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 102, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 330, + 102, + 553, + 144 + ], + "type": "interline_equation", + "content": "l _ {x} = \\frac {1}{\\sqrt {H W}} \\sum_ {\\substack {x = 0 \\\\ H = 1}} ^ {H - 1} \\sum_ {\\substack {y = 0 \\\\ W = 1}} ^ {W - 1} \\left| G _ {x} ^ {\\prime} (x, y) - G _ {x} (x, y) \\right|, \\tag{14}", + "image_path": "4821e8ccd093fd13bfd5d179d46b1daa64a2456a95cbe861aa01e1f74c90c76f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 140, + 520, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 140, + 520, + 174 + ], + "spans": [ + { + "bbox": [ + 332, + 140, + 520, + 174 + ], + "type": "interline_equation", + "content": "l _ {y} = \\frac {1}{\\sqrt {H W}} \\sum_ {x = 0} ^ {H - 1} \\sum_ {y = 0} ^ {W - 1} \\left| G _ {y} ^ {\\prime} (x, y) - G _ {y} (x, y) \\right|.", + "image_path": "75cb2b3ab1cb74ee9655008d4f4aee40ad8267ccdc9af92f9f0a9eb8c9b3110b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 182, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 410 + ], + "type": "text", + "content": "Gradient maps, derived from differences in pixel intensities, typically exhibit meaningful gradients in localized regions, such as edges and areas with significant intensity changes, which are crucial for capturing structures and textures. In contrast, the majority of areas in gradient map comprise flat and low-gradient regions with limited information. Therefore naively adopting the horizontal and vertical gradient discrepancies as the loss function can cause low-gradient regions to dominate the overall loss. As a result, the model may deemphasize regions critical for preserving fine textures and structures, focusing instead on less important flat areas. Besides, the regions with meaningful gradients should also be distinguished, prioritizing greater attention to textures and structures exhibiting larger rendering errors. To address the above issues, we introduce a dynamic region selection that computes the rendering error map as a weight map, defined by the absolute difference between the gradient maps of the rendered result and the ground truth, which can be expressed by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 418, + 553, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 418, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 372, + 418, + 553, + 438 + ], + "type": "interline_equation", + "content": "w _ {x} = \\left| G _ {x} ^ {\\prime} (x, y) - G _ {x} (x, y) \\right|, \\tag {15}", + "image_path": "b414e6b9ddb09655cedc9f8870a28d51c6bcff3cf9ee7005701c05281e5964ae.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 436, + 494, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 436, + 494, + 449 + ], + "spans": [ + { + "bbox": [ + 375, + 436, + 494, + 449 + ], + "type": "interline_equation", + "content": "w _ {y} = \\left| G _ {y} ^ {\\prime} (x, y) - G _ {y} (x, y) \\right|.", + "image_path": "6a43a5b50a35c00ddc76c09345756ac962a1db75455d2c1d6cc850e928e34b7f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 457, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 481 + ], + "type": "text", + "content": "Combined with the horizontal and vertical weight maps, the selective gradient loss " + }, + { + "bbox": [ + 313, + 457, + 554, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 457, + 554, + 481 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 491, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 491, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 383, + 491, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = w _ {x} * l _ {x} + w _ {y} * l _ {y}, \\tag {16}", + "image_path": "f415f92494372de616d9fcc6f5a2ce8c49e5149323fea37512add4ffc1374e2d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 514, + 555, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 646 + ], + "type": "text", + "content": "Based on the above analysis, the selective gradient loss enables the model to focus on the difficult-to-render textural and structural regions adaptively. Besides, the model is able to adjust the focused regions dynamically in response to the evolving rendering error map throughout the training process. Note, the pixel-level L1 loss with D-SSIM term between RGB images is also used in the training process, which complements the proposed selective gradient loss used between gradient maps. And we also keep the volume regularization as described in Scaffold-GS [25]. The total loss function is given as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 655, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 655, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 325, + 655, + 553, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {S S I M} \\mathcal {L} _ {S S I M} + \\lambda_ {v o l} \\mathcal {L} _ {v o l} + \\lambda_ {s} \\mathcal {L} _ {s}, \\tag {17}", + "image_path": "9758e5af2b6e36b525ab39f99fd857a19f6a3f6d478f36bd93d4494d71015e66.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_s" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{SSIM}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{vol}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " represent the training weights of the selective gradient loss " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", the L1 loss " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", the D-SSIM term " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SSIM}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and the volume regularization " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vol}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11171" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 148 + ], + "type": "table", + "html": "
DatasetsMip-NeRF360Tanks&TemplesDeep Blending
MethodsSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor SizeSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.80627.500.25232 dim0.85323.960.17732 dim0.90630.210.25432 dim
SOGS(Ours)0.81527.850.22116 dim0.85524.140.17612 dim0.90730.290.25212 dim
", + "image_path": "bf2f2ca4284921ec2ead15269ab29e2dcd09583df597ebeb7b24f9f1562728ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 217, + 293, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 154, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 555, + 200 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of novel view synthesis. All methods are trained with the same training data. SOGS achieves superior rendering quality with reduced anchor (model) size. Notably, SOGS using the feature dimension of 12 can still outperform the Scaffold-GS that uses a feature dimension of 32. We do not compare with HAC [9] and ContextGS [35], as they leverage compression techniques to reduce the storage size of Scaffold-GS rather than the actual anchor and model sizes as used in training and rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 217, + 293, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 293, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 293, + 285 + ], + "type": "table", + "html": "
DatasetsBungeeNeRF
MethodsSSIM↑PSNR↑LPIPS↓Anchor Size
Scaffold-GS0.86526.620.24132 dim
SOGS(Ours)0.88027.060.17116 dim
", + "image_path": "f19a69f94c573fde3bde434e1ac806d4421ec156f805c794e3aeb98432e35890.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 295, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 326 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on large-scale scenes from the dataset BungeeNeRF [37]. All methods are trained with the same training data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 346, + 137, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 137, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 137, + 361 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 366, + 251, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 251, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 251, + 380 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 552 + ], + "type": "text", + "content": "Datasets For both training and testing, we adopt the datasets and configurations in Scaffold-GS [25] and conduct experiments on images of 19 real-world scenes. Specifically, we evaluate SOGS across all nine scenes from the Mip-NeRF360 dataset [2], two scenes from the Tanks&Temples dataset [19], and two additional scenes from the Deep Blending dataset [17]. These scenes exhibit diverse styles, ranging from bounded indoor environments to unbounded outdoor ones. We also evaluate SOGS on the BungeeNeRF dataset [37], which includes six large-scale outdoor scenes, to further validate its effectiveness. The datasets are divided into training and test sets following the same settings as Scaffold-GS, with image resolutions consistent with those used in Scaffold-GS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Implementation For the second-order anchor, we reduce the dimension " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " of the anchor features in Scaffold-GS from 32 to 16, or even to 12. The covariance matrix " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is computed from all anchors in scene representations which captures the global correlation across all anchor feature dimensions. The number of the selected eigenvectors is set to " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". All MLPs used to extract anchor-specific textural and structural information consist of two layers with ReLU activation. For the loss function, the training weights " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{s}, \\lambda_{1}, \\lambda_{SSIM}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{vol}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are set to 0.01, 0.8, 0.2 and 0.01. Please note, for fair comparisons, we keep the loss terms " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{1}, \\lambda_{SSIM}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{vol}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " the same as in Scaffold-GS. The number of" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 220, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 268 + ], + "type": "text", + "content": "3D Gaussians corresponding to one second-order anchor is set as " + }, + { + "bbox": [ + 313, + 220, + 555, + 268 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 313, + 220, + 555, + 268 + ], + "type": "text", + "content": ". We use the Pytorch framework to implement SOGS. The SOGS model is trained for 30000 iterations, the same as Scaffold-GS and 3D-GS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 275, + 518, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 518, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 518, + 289 + ], + "type": "text", + "content": "4.2. Comparisons with the State-of-the-Art" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 293, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 641 + ], + "type": "text", + "content": "We primarily compare SOGS with Scaffold-GS [25] over datasets Mip-NeRF360, Tank&Temple and Deep Blending as Scaffold-GS is an anchor-based 3D Gaussian splattering method and is most relevant to our study. Table 1 shows experimental results over the same test images as described in Section 4.1. We can observe that SOGS outperforms Scaffold-GS consistently in PSNR, SSIM and LPIPS across all evaluated scenes. The anchor size of SOGS is reduced to 16 in the dataset Mip-NeRF360, while it is challengingly reduced to 12 in Tank&Temple and Deep Blending. The superior performance is largely attributed to our proposed second-order anchor which performs anchor feature augmentation to compensate for the reduced anchor feature size and improve the rendering quality as well as the selective gradient loss that further ensures the high-quality rendering under small-sized anchor features. Further, we benchmark SOGS with Scaffold-GS over the BungeeNeRF dataset which provides diverse large-scale scenes. It can be observed that SOGS achieves superior rendering quality with only 16-dimensional anchors. Besides, as shown in Fig. 4, we perform the visual comparison of SOGS with 3D-GS and Scaffold-GS and provide their model sizes in each scene. With the compact anchor size of the proposed second-order anchor, SOGS achieves a significant reduction in model size while delivering superior novel view synthesis with much less artifacts but finer details. Note we did not compare with HAC [9] and ContextGS [35] as they compress the storage size of Scaffold-GS rather than the actual model size as used in training and rendering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 416, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 416, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 416, + 659 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We conduct extensive ablation experiments to validate the effectiveness of the proposed second-order anchor and the selective gradient loss. In the experiments, we keep the model size and anchor feature dimension the same as used" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11172" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 181, + 151 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 181, + 151 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 181, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 181, + 151 + ], + "type": "image", + "image_path": "53c23612a1d9a5daffd7061b8e799cab742e92987764e55c4ad64b14c3562685.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 152, + 180, + 233 + ], + "blocks": [ + { + "bbox": [ + 56, + 152, + 180, + 233 + ], + "lines": [ + { + "bbox": [ + 56, + 152, + 180, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 180, + 233 + ], + "type": "image", + "image_path": "96b30dcd32b63f16effd497d396c876f205a4d4ad03ad69b152c9192db80f121.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 234, + 180, + 316 + ], + "blocks": [ + { + "bbox": [ + 56, + 234, + 180, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 234, + 180, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 180, + 316 + ], + "type": "image", + "image_path": "a72293882420d3c57150d9da0032643570a007aa6adbe497ed22bfc1d492190a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 317, + 180, + 388 + ], + "blocks": [ + { + "bbox": [ + 56, + 317, + 180, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 317, + 180, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 180, + 388 + ], + "type": "image", + "image_path": "3b2957d95718f0ea4c0272843bb69dfbcc070b7f3a5e55d0c993da423aa5bd49.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 391, + 130, + 399 + ], + "lines": [ + { + "bbox": [ + 103, + 391, + 130, + 399 + ], + "spans": [ + { + "bbox": [ + 103, + 391, + 130, + 399 + ], + "type": "text", + "content": "3D-GS" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 181, + 71, + 304, + 151 + ], + "blocks": [ + { + "bbox": [ + 181, + 71, + 304, + 151 + ], + "lines": [ + { + "bbox": [ + 181, + 71, + 304, + 151 + ], + "spans": [ + { + "bbox": [ + 181, + 71, + 304, + 151 + ], + "type": "image", + "image_path": "a2cfbf6e77c9a18fad2ee76ece50a9e09076486ec5edadeb98c93982a621c3e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 181, + 152, + 304, + 233 + ], + "blocks": [ + { + "bbox": [ + 181, + 152, + 304, + 233 + ], + "lines": [ + { + "bbox": [ + 181, + 152, + 304, + 233 + ], + "spans": [ + { + "bbox": [ + 181, + 152, + 304, + 233 + ], + "type": "image", + "image_path": "5ec4df95056874e034468d26c2c2937f96ae9c85492e98c3f8b5929864d983fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 234, + 304, + 316 + ], + "blocks": [ + { + "bbox": [ + 181, + 234, + 304, + 316 + ], + "lines": [ + { + "bbox": [ + 181, + 234, + 304, + 316 + ], + "spans": [ + { + "bbox": [ + 181, + 234, + 304, + 316 + ], + "type": "image", + "image_path": "133c90f6d63fbf63b51b02947de23486567dd8a6483adf1b89729ae1be0a4613.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 181, + 317, + 304, + 388 + ], + "blocks": [ + { + "bbox": [ + 181, + 317, + 304, + 388 + ], + "lines": [ + { + "bbox": [ + 181, + 317, + 304, + 388 + ], + "spans": [ + { + "bbox": [ + 181, + 317, + 304, + 388 + ], + "type": "image", + "image_path": "7075f87aec37e98e544c993a7b17a0d66032e401b0a8c96e207dbbb7439e72da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 391, + 272, + 399 + ], + "lines": [ + { + "bbox": [ + 226, + 391, + 272, + 399 + ], + "spans": [ + { + "bbox": [ + 226, + 391, + 272, + 399 + ], + "type": "text", + "content": "Scaffold-GS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 429, + 151 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 429, + 151 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 429, + 151 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 429, + 151 + ], + "type": "image", + "image_path": "3f3e643bcc45e8e977e3d0483710782cc7754e5e57234706e558dcb465b63f5c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 152, + 429, + 233 + ], + "blocks": [ + { + "bbox": [ + 306, + 152, + 429, + 233 + ], + "lines": [ + { + "bbox": [ + 306, + 152, + 429, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 429, + 233 + ], + "type": "image", + "image_path": "dfc67c21b11c7e99623ab4cbd53d3897133e19bb9c387e4e810f11aed9b33022.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 234, + 429, + 316 + ], + "blocks": [ + { + "bbox": [ + 306, + 234, + 429, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 234, + 429, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 429, + 316 + ], + "type": "image", + "image_path": "9e7847eb86a16524bb93a1c9dbdcb8846307972aa08ec0d7741a00c232a57118.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 317, + 429, + 388 + ], + "blocks": [ + { + "bbox": [ + 306, + 317, + 429, + 388 + ], + "lines": [ + { + "bbox": [ + 306, + 317, + 429, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 317, + 429, + 388 + ], + "type": "image", + "image_path": "1e9873211860b4153ce815690bfebab3f8cddf13d16fd8093e37cac18ce44463.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 391, + 397, + 400 + ], + "lines": [ + { + "bbox": [ + 345, + 391, + 397, + 400 + ], + "spans": [ + { + "bbox": [ + 345, + 391, + 397, + 400 + ], + "type": "text", + "content": "SOGS (Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 419, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 419, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 555, + 453 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of SOGS with 3D-GS [18] and Scaffold-GS [25]. SOGS achieves smaller model size and superior image rendering with much less artifacts but more fine details. The experiments are conducted over multiple indoor and outdoor scenes including 'Garden', 'Flower' and 'Counter' from the Mip-NeRF360 and 'Quebec' from the BungeeNeRF. Zoom in for best view." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 553, + 151 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 553, + 151 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 553, + 151 + ], + "type": "image", + "image_path": "89f30eda274087615941b9d6507304c94608409ca04d8d8f53c274576608fb60.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 430, + 152, + 553, + 233 + ], + "blocks": [ + { + "bbox": [ + 430, + 152, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 430, + 152, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 430, + 152, + 553, + 233 + ], + "type": "image", + "image_path": "360b2830dc0f98fb30f6f1e7d3a668a38b692b491446d28ce61a4b8234e45406.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 429, + 234, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 429, + 234, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 429, + 234, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 429, + 234, + 553, + 316 + ], + "type": "image", + "image_path": "1168505cd077e2a18611a0f586dd398e6c86a7d36a9b731675c5ea6c14696c67.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 429, + 317, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 429, + 317, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 429, + 317, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 429, + 317, + 553, + 388 + ], + "type": "image", + "image_path": "c96c271317359f08a98c09b52974db73a06100b361e07a4c3b45de9d0f08101a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 391, + 520, + 399 + ], + "lines": [ + { + "bbox": [ + 465, + 391, + 520, + 399 + ], + "spans": [ + { + "bbox": [ + 465, + 391, + 520, + 399 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 61, + 471, + 291, + 541 + ], + "blocks": [ + { + "bbox": [ + 61, + 471, + 291, + 541 + ], + "lines": [ + { + "bbox": [ + 61, + 471, + 291, + 541 + ], + "spans": [ + { + "bbox": [ + 61, + 471, + 291, + 541 + ], + "type": "table", + "html": "
ModelsEvaluation Metrics
PSNR ↑SSIM ↑LPIPS ↓
Base26.620.8650.241
Base+SOA27.250.8790.208
Base+SOA+SGL27.390.8870.161
", + "image_path": "43be155e99941f268cf0bf9c04b226c1ad73cf0385d581dbe0ae8bb37dd09cbf.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 559, + 296, + 658 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 658 + ], + "type": "text", + "content": "Table 3. Ablation studies of SOGS on the dataset BungeeNeRF. With Scaffold-GS as the baseline Base, Base + SOA introduces the second-order anchor to augment anchor features which clearly improves the Gaussian attribute prediction and image rendering. Base + SOA + SGL (i.e., SOGS) further introduces the selective gradient loss which enables the Gaussian prediction to focus on regions with difficult-to-render textures and geometries. All models are trained under the same settings and model size, and the anchor feature dimension is set at 32." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "in Scaffold-GS, aiming to screen out other variations and show how our two designs improve the rendering quality." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 474, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 555, + 487 + ], + "type": "text", + "content": "More details are described in the following two subsections." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "type": "text", + "content": "Second-order Anchor We first examine how our proposed second-order anchor affects PSNR, SSIM and LPIPS. In this experiment, we adopt the Scaffold-GS as the baseline model Base that performs anchor-based Gaussian splattering with naive anchors. On top of the Base, we train a model Base " + }, + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "type": "inline_equation", + "content": "+SOA" + }, + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "type": "text", + "content": " that incorporates our proposed second-order anchor to replace the naive anchor in Scaffold-GS. As Table 3 shows, Base " + }, + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "type": "inline_equation", + "content": "+SOA" + }, + { + "bbox": [ + 313, + 504, + 555, + 635 + ], + "type": "text", + "content": " outperforms the Base clearly in PSNR, SSIM and LPIPS, indicating that the second-order anchor can effectively improve the Gaussian attribute prediction and novel view synthesis by anchor feature augmentation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Selective Gradient Loss To evaluate how the proposed selective gradient loss contributes, we train a new model " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "Base + SOA + SGL" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": " (i.e., the complete SOGS) that incorporates the selective gradient loss on top of " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "Base + SOA" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": ". Table 3 shows experimental results. We can observe that" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11173" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 298, + 150 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 150 + ], + "type": "image", + "image_path": "0a44e6b9f7d31a0aec7f975cf9bb2e3fb338af55453a385db6000b0bbf82c3b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "type": "text", + "content": "Figure 5. Anchor feature dimension vs SOGS performance: The performance of SOGS varies with the feature dimension " + }, + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "type": "text", + "content": " in each anchor, where increasing " + }, + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 167, + 297, + 225 + ], + "type": "text", + "content": " improves SOGS consistently together with the increased model size and computational cost. The graph shows the PSNR of the scene 'Bicycle' from MipNeRF360." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 244, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 293 + ], + "type": "text", + "content": "Base " + }, + { + "bbox": [ + 55, + 244, + 296, + 293 + ], + "type": "inline_equation", + "content": "+SOA + SGL" + }, + { + "bbox": [ + 55, + 244, + 296, + 293 + ], + "type": "text", + "content": " further improves the novel view synthesis consistently across all three evaluation metrics, demonstrating its effectiveness on Gaussian attribute prediction and rendering quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 192, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 192, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 192, + 314 + ], + "type": "text", + "content": "4.4. Parameter Investigation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "text", + "content": "We examine how the dimension of anchor features " + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "text", + "content": " affects the SOGS performance. Fig. 5 shows experiments on the scene 'Bicycle' from the dataset MipNeRF360. We can observe that the performance of SOGS (in PSNR ↑ score) improves consistently while the feature dimension " + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "text", + "content": " increases. Specifically, the feature dimension plays a vital role in the rendering quality as it sets the upper bound for the encoded information as well as the effectiveness of feature augmentation in SOGS. As " + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "text", + "content": " becomes larger than 16, the performance gains gradually diminish due to feature saturation while the model size and computational costs grow greatly. We therefore set " + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 319, + 296, + 475 + ], + "type": "text", + "content": " at 12 or 16 in our implemented system to balance the model size and performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 140, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 140, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 140, + 495 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 501, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 597 + ], + "type": "text", + "content": "We visualize the impact of second-order anchors on scene textures and structures. As Fig. 6 shows, our proposed second-order anchor can enhance the rendered images with finer structures and textures. This visualization verifies that the proposed second-order anchor can effectively capture the structural and textural patterns for anchor feature augmentation, and facilitate subsequent Gaussian prediction and image rendering with augmented features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 609, + 129, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 129, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 129, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "This paper presents SOGS, an innovative anchor-based 3D Gaussian splatting technique that incorporates second-order anchors to achieve superior rendering with compact anchor feature dimensions and model size. Specifically, by introducing covariance-based second-order statistics, SOGS learns correlations across anchor feature dimensions to capture anchor-specific textures and structural patterns. This" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 70, + 433, + 143 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 433, + 143 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 433, + 143 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 433, + 143 + ], + "type": "image", + "image_path": "a136b8b1f3c5ce0d7c08a6cb921cd9440b90b678064615716bfc31cc1adfbaa0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 324, + 144, + 433, + 215 + ], + "blocks": [ + { + "bbox": [ + 324, + 144, + 433, + 215 + ], + "lines": [ + { + "bbox": [ + 324, + 144, + 433, + 215 + ], + "spans": [ + { + "bbox": [ + 324, + 144, + 433, + 215 + ], + "type": "image", + "image_path": "6fad434b231629a5c57501163d2a051911c9aa09d97cb21a947ed987e002633f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 324, + 217, + 433, + 289 + ], + "blocks": [ + { + "bbox": [ + 324, + 217, + 433, + 289 + ], + "lines": [ + { + "bbox": [ + 324, + 217, + 433, + 289 + ], + "spans": [ + { + "bbox": [ + 324, + 217, + 433, + 289 + ], + "type": "image", + "image_path": "4f84bb157558e17130ab51bbbe2616df2f27b280437e15c0d3500882c8b76707.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 324, + 291, + 433, + 363 + ], + "blocks": [ + { + "bbox": [ + 324, + 291, + 433, + 363 + ], + "lines": [ + { + "bbox": [ + 324, + 291, + 433, + 363 + ], + "spans": [ + { + "bbox": [ + 324, + 291, + 433, + 363 + ], + "type": "image", + "image_path": "70f9d1682c5393f26b595da4baadeeac80d3df8b549d3dd748d789434a92bf8b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 365, + 399, + 373 + ], + "lines": [ + { + "bbox": [ + 351, + 365, + 399, + 373 + ], + "spans": [ + { + "bbox": [ + 351, + 365, + 399, + 373 + ], + "type": "text", + "content": "Naive Anchor" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 71, + 545, + 143 + ], + "blocks": [ + { + "bbox": [ + 435, + 71, + 545, + 143 + ], + "lines": [ + { + "bbox": [ + 435, + 71, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 435, + 71, + 545, + 143 + ], + "type": "image", + "image_path": "05b4a95019e66465ff70fa9a3fd1b5a35086f2e3d69c44c0234799b77d9c88af.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 435, + 144, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 435, + 144, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 435, + 144, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 435, + 144, + 545, + 216 + ], + "type": "image", + "image_path": "6d1c9d73bf3aec43edcabcb1f9c52d9cea8ffe3291312a7592fcce6aa7aa960a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 217, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 435, + 217, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 435, + 217, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 435, + 217, + 545, + 289 + ], + "type": "image", + "image_path": "15b448cf857c0ff60b27b49db5d2539bc9e95568ea35f567b068decc45488e1d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 291, + 545, + 362 + ], + "blocks": [ + { + "bbox": [ + 436, + 291, + 545, + 362 + ], + "lines": [ + { + "bbox": [ + 436, + 291, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 436, + 291, + 545, + 362 + ], + "type": "image", + "image_path": "dda097a16269f91c2f01ba6d7c5a870ee937b809edb77660122c0626e86179f8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 365, + 527, + 373 + ], + "lines": [ + { + "bbox": [ + 455, + 365, + 527, + 373 + ], + "spans": [ + { + "bbox": [ + 455, + 365, + 527, + 373 + ], + "type": "text", + "content": "Second-Order Anchor" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 460 + ], + "type": "text", + "content": "Figure 6. Visualization of the impact of second-order anchors on scene textures and structures. The two samples are 'Drjohnson' and 'Playroom' from dataset Deep Blending [17]. The proposed second-order anchor improves the rendering quality with superior structures and textures by anchor feature augmentation. Zoom in for best view." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 480, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 635 + ], + "type": "text", + "content": "enables each anchor to achieve feature augmentation, compensating for the reduced anchor feature size and resulting in superior Gaussian prediction and rendering quality. Besides, we also design a selective gradient loss to further ensure high-quality rendering under compact anchor features, which enables the model to dynamically focus on the Gaussian prediction for difficult-to-render textural and structural regions. Experiments over multiple widely adopted indoor and outdoor scenes show that SOGS achieves superior novel view synthesis with compact model size. Although time complexity increases slightly due to the additional computations in second-order anchors, SOGS still achieves efficient training and real-time rendering." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 646, + 430, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 430, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 430, + 659 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "This project is funded by the Ministry of Education Singapore, under the Tier-2 project scheme with a project number MOE-T2EP20220-0003. Fangneng Zhan is supported by the NSF grant CRCNS-2309041." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11174" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 323 + ], + "type": "text", + "content": "[4] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 325, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 379 + ], + "type": "text", + "content": "[5] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "text", + "content": "[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "type": "text", + "content": "[8] Yue Chen, Xingyu Chen, Xuan Wang, Qi Zhang, Yu Guo, Ying Shan, and Fei Wang. Local-to-global registration for bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8264-8273, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 579 + ], + "type": "text", + "content": "[9] Yihang Chen, Qianyi Wu, Jianfei Cai, Mehrtash Harandi, and Weiyao Lin. Hac: Hash-grid assisted context for 3d gaussian splattering compression. arXiv preprint arXiv:2403.14530, 2024. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "text", + "content": "[10] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "text", + "content": "[11] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 14324-14334, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 151 + ], + "type": "text", + "content": "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "text", + "content": "[14] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "text", + "content": "[15] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 285 + ], + "type": "text", + "content": "[16] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "text", + "content": "[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 333, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 376 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 2, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "type": "text", + "content": "[19] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "type": "text", + "content": "[20] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "type": "text", + "content": "[21] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[22] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 633 + ], + "type": "text", + "content": "[23] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[24] Kunhao Liu, Fangneng Zhan, Muyu Xu, Christian Theobalt, Ling Shao, and Shijian Lu. Stylegaussian: Instant 3d style transfer with gaussian splatting. In SIGGRAPH Asia 2024 Technical Communications, pages 1-4. 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "type": "text", + "content": "[25] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11175" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 152 + ], + "type": "text", + "content": "[26] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 207 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 294, + 251 + ], + "type": "text", + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "text", + "content": "[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 354 + ], + "type": "text", + "content": "[30] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 356, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 295, + 420 + ], + "type": "text", + "content": "[31] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "text", + "content": "[32] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 544 + ], + "type": "text", + "content": "[33] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 600 + ], + "type": "text", + "content": "[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 645 + ], + "type": "text", + "content": "[35] Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex C Kot, and Bihan Wen. Contextgs: Compact 3d gaussian splatting with anchor level context model. arXiv preprint arXiv:2405.20721, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "text", + "content": "[36] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 664 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[37] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In European conference on computer vision, pages 106-122. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[38] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[39] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "text", + "content": "[40] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "text", + "content": "[41] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 395 + ], + "type": "text", + "content": "[42] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "text", + "content": "[43] Jiahui Zhang, Fangneng Zhan, Rongliang Wu, Yingchen Yu, Wenqing Zhang, Bai Song, Xiaqin Zhang, and Shijian Lu. Vmrf: View matching neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6579-6587, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "text", + "content": "[44] Jiahui Zhang, Fangneng Zhan, Yingchen Yu, Kunhao Liu, Rongliang Wu, Xiaoqin Zhang, Ling Shao, and Shijian Lu. Pose-free neural radiance fields via implicit pose regularization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3534-3543, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "type": "text", + "content": "[45] Jiahui Zhang, Fangneng Zhan, Muyu Xu, Shijian Lu, and Eric Xing. Fregs: 3d gaussian splatting with progressive frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21424-21433, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 565, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 597 + ], + "type": "text", + "content": "[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 599, + 553, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 664 + ], + "type": "text", + "content": "[47] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11176" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_content_list.json b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7edcbf1c108c58795b0783d7817ee7ddcbe4b9b3 --- /dev/null +++ b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_content_list.json @@ -0,0 +1,1414 @@ +[ + { + "type": "text", + "text": "SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters", + "text_level": 1, + "bbox": [ + 194, + 130, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianping Jiang\\*†, Weiye Xiao\\*‡, Zhengyu Lin $^{1}$ , Huaizhong Zhang $^{1}$ , Tianxiang Ren $^{1}$ , Yang Gao $^{1}$ , Zhiqian Lin $^{1}$ , Zhongang Cai\\*1,2, Lei Yang\\*1, Ziwei Liu\\*2 \n $^{1}$ SenseTime Research, $^{2}$ S-Lab, Nanyang Technological University \n* Equal Contribution, † Project Lead, ‡ Engineering Lead, § Corresponding Author \nhttps://solami-ai.github.io/", + "bbox": [ + 189, + 202, + 807, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6b04288844f9303802f257ea5d2e55ab414937fb1546b89de629177d5989676.jpg", + "image_caption": [ + "Figure 1. SOLAMI enables the user to interact with 3D autonomous characters through speech and body language in an immersive VR environment via an end-to-end social vision-language-action model, which is trained on our synthesized multimodal dataset SynMSI." + ], + "image_footnote": [], + "bbox": [ + 94, + 315, + 903, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 699, + 326, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human beings are social animals. How to equip 3D autonomous characters with similar social intelligence that can perceive, understand and interact with humans remains an open yet foundational problem. In this paper, we introduce SOLAMI, the first end-to-end Social vision-Language-Action (VLA) Modeling framework for Immersive interaction with 3D autonomous characters. Specifically, SOLAMI builds 3D autonomous characters from three aspects: 1) Social VLA Architecture: We propose a unified social VLA framework to generate multimodal response (speech and motion) based on the user's", + "bbox": [ + 88, + 734, + 486, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multimodal input to drive the character for social interaction. 2) Interactive Multimodal Data: We present SynMSI, a synthetic multimodal social interaction dataset generated by an automatic pipeline using only existing motion datasets to address the issue of data scarcity. 3) Immersive VR Interface: We develop a VR interface that enables users to immersively interact with these characters driven by various architectures. Extensive quantitative experiments and user study demonstrate that our framework leads to more precise and natural character responses (in both speech and motion) that align with user expectations with lower latency.", + "bbox": [ + 511, + 700, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26887", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Have you ever imagined having an immersive face-to-face conversation with a character you deeply admire? Not merely through speech dialogue, but through an interaction where you can observe its subtle facial expressions, natural body language, and even fleeting emotional changes. Psychology research [15, 25, 46, 49] shows that in social interactions, the greater the level of immersion, the better the human experience. However, current character agents [1, 3, 47] are still limited to text or voice interactions. This limitation prompts us to build 3D autonomous characters with richer modalities.", + "bbox": [ + 89, + 114, + 480, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Developing an autonomous 3D character requires effectively modeling its behavior system, which involves two major challenges: 1) The 3D character needs to accurately observe and understand the information conveyed by the user, and respond appropriately based on the context and its character setting through speech, body motion, and facial expression, etc. This goes beyond previous singular human-related tasks, such as motion generation [62], motion understanding [27], and audio-to-motion [6]. 2) Data for multimodal interactions between users and 3D characters is extremely scarce due to the prohibitive cost of the comprehensive setup.", + "bbox": [ + 89, + 281, + 480, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous work [13] is primarily based on the LLM-Agent framework, using text to link various sub-modules (such as motion captioning and text-to-motion). While this approach performs well in high-level tasks like planning and memory, it tends to fall short in tasks such as understanding user behaviors and providing timely body motion responses. This limitation arises because using text as the intermediary between modules conveys high-level information but often omits subtle nuances. And the sub-modules (motion captioning, speech recognition, etc.) in the complex engineering framework incur substantial latency that undermines the timeliness of a natural communication [50].", + "bbox": [ + 89, + 462, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interestingly, research on robotics shows the similar conclusion. The LLM-Agent framework can handle planning tasks [26], but for low-level manipulation tasks, end-to-end Vision-Language-Action (VLA) models built upon LLMs show superior performance [11, 24, 29, 67]. We argue that digital avatars are essentially robots with virtual humanoid embodiment. Therefore, building a VLA model for social interactions with users is a promising direction.", + "bbox": [ + 89, + 643, + 482, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we implement an end-to-end social VLA model, SOLAMI. Our model, built upon a decoder-only LLM backbone, processes the inputs of user speech and motion into discrete representations, and generate the responsive speech and motion tokens, which are then decoded to the character's speech and motion. This modeling approach can effectively learn character behavior patterns across motion and speech modalities and offer low latency.", + "bbox": [ + 89, + 763, + 482, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although there are numerous datasets related to human", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "social behaviors [13, 34, 40, 56], comprehensive multimodal interaction datasets remain scarce. Thus, we introduce a data synthesis method that utilizes existing text-motion datasets to automatically construct multimodal interaction data at a low cost. Leveraging our extensively curated topics (5.3 K), uniformly processed motion database (46 K), and iterative script refinement pipeline, we develop SynMSI, a dataset containing 6.3 K multi-turn multimodal conversation items. To evaluate the effectiveness of our method, we developed a VR interface where users can immersively interact with various 3D characters. Quantitative experimental results and user study analysis show that our approach produces more precise and natural social interaction experience with lower latency.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we contribute 1) A new VLA architecture to model the character's behavior system for immersive social interaction; 2) A dedicated synthesizing pipeline that automatically generates large-scale multimodal interactive dataset, SynMSI; 3) An immer VR interface for users to interact with various characters through speech and motion.", + "bbox": [ + 511, + 301, + 903, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 424, + 653, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human Motion & LLM", + "text_level": 1, + "bbox": [ + 511, + 450, + 733, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing role-play agents primarily rely on text [47], speech [1, 3, 41], or video [7] as interactive media, but building 3D autonomous characters means modeling 3D embodied behavior, especially body language. Unlike single-purpose human motion tasks [6, 12, 22, 34, 48, 56, 62], we expect this 3D character not only to comprehend user speech and body language but also to respond according to its profile setting. Large language models (LLMs) [41, 51], with their remarkable emergent abilities [54], provide promising solutions for this direction. One approach [18, 28, 64, 65] tries to integrate LLMs and human motion in an end-to-end fashion, enabling a single model to perform multiple motion-related tasks. However, the goal of these methods is not to generate responsive motion based on the input motion according to the character setting. Another approach [13, 35] utilizes the LLM as a versatile brain center that controls various submodules (e.g., motion understanding, text-to-motion generation) with text or code instructions. However, this modular approach inherently introduces information loss and engineering-related time latency. Therefore, how to create feasible autonomous 3D characters for immersive interaction remains an open challenge.", + "bbox": [ + 511, + 473, + 903, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Embodied Intelligence", + "text_level": 1, + "bbox": [ + 511, + 832, + 723, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building 3D autonomous characters means creating virtual humanoid agents with embodied intelligence. In the field of embodied intelligence, researchers have found the high-", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26888", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "level abilities (planning, memory, etc.) of LLM-Agents across various environments, including factories [16, 26], 2D sandbox settings [42], and 3D gaming spaces [53]. For tasks requiring low-level skills, such as manipulation, end-to-end VLA models [11, 24, 29, 67] have shown considerable potential. Despite existing efforts [19, 20, 22] in egocentric human-related tasks, the capabilities of VLA models in social interactive tasks with humans have not been fully explored.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Social Vision-Language-Action Modeling", + "text_level": 1, + "bbox": [ + 89, + 242, + 459, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 89, + 267, + 225, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework is an end-to-end social VLA model that takes the user's speech and motion as input and generates the character's responsive speech and motion as output. In this process, speech and motion are added as new languages to the LLM text vocabulary. First, the user's speech and motion are converted into discrete motion tokens and speech tokens via a motion tokenizer and a speech tokenizer, respectively. A decoder-only LLM backbone then predicts the character's output motion and speech tokens based on the user's input tokens and the character setting. The generated tokens are subsequently decoded into corresponding motion and speech by their respective decoders.", + "bbox": [ + 89, + 289, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Representation. To take advantage of SMPL-X's [43] compatibility with industry animation workflow, we directly model human poses as SMPL-X joint rotations instead of keypoint positions [28, 38] to facilitate animation of characters down the stream.", + "bbox": [ + 89, + 470, + 483, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion Tokenizer. Our motion tokenizer employs a Vector Quantized Variational Autoencoders (VQ-VAE) structure as [28, 64]. It learns discrete representations of motion, enabling the LLM to understand the text-motion connection. We design separate VQVAEs $(Q^{b}, Q^{h}, Q^{t})$ for the body motion $\\mathbf{m}^{b}$ , hand motion $\\mathbf{m}^{h}$ , and inter-character relative transform (rotation and translation) $\\mathbf{m}_{t}$ for higher reconstruction accuracy [38]. The quantization process for motion $\\mathbf{m}^{u}$ is formulated as", + "bbox": [ + 89, + 547, + 483, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {m} _ {t} ^ {u} = Q ^ {u} \\left(\\mathbf {m} _ {t} ^ {u}\\right) = \\arg \\min _ {z _ {i} \\in \\mathbb {Z} _ {u}} \\| \\mathbf {m} _ {t} ^ {u} - z _ {i} \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 695, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{Z}_u$ is the codebook of motion part $u\\in \\{b,h,t\\}$ and $\\hat{m}_t^u$ is the corresponding motion tokens. The VQVAEs of body and hand, $Q^{b}$ and $Q^{h}$ , apply 1D convolutions on motion features along the temporal dimension to get $L_{M}$ sequential tokens $\\hat{M}^b = [\\hat{m}_1^b,\\hat{m}_2^b,\\dots,\\hat{m}_{L_M}^b]$ and $\\hat{M}^h = [\\hat{m}_1^h,\\hat{m}_2^h,\\dots,\\hat{m}_{L_M}^h]$ , and the VQVAE $Q^{t}$ uses MLPs to get the transform token $\\hat{m}^t$ of a sequence.", + "bbox": [ + 89, + 729, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech Tokenizer. Research [9, 61] on speech discretization mainly utilizes the RVQ-VAE structure [59]. In this work, we utilize the SpeechTokenizer [63] that disentangles semantic and acoustic information within speech $S$ . This", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "allows us to use the semantic tokens $\\hat{S} = [\\hat{s}_1^s, \\hat{s}_2^s, \\dots, \\hat{s}_{L_S}^s]$ from the first layer as input to the LLM, reducing inference costs ( $L_S$ is the sequence length of semantic tokens). Simultaneously, a short sample of the character's voice (4 to 6 seconds) can be used as a prompt to achieve instance voice cloning when decoding through SoundStorm [10].", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal Multi-round Interaction. User-character interaction is formulated as a multi-round conversation fashion of common LLMs [41, 51]. When the user sends speech and motion to SOLAMI, the model auto-regressively generates speech and motion responses based on previous dialogue contents and the character setting. To facilitate training, we use special tokens to mark the start and end of each modality sequence as [28, 30]. The interaction process can be formulated as follow:", + "bbox": [ + 511, + 181, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interaction Template", + "text_level": 1, + "bbox": [ + 540, + 329, + 679, + 344 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Input: \nSystem Prompt: \nUser: $< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>$ \nCharacter: $< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>$ \n... \nUser: $< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>$ \nOutput: \nCharacter: $< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>$", + "guess_lang": "txt", + "bbox": [ + 540, + 357, + 856, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle \\text{Character\\_Placeholder} \\rangle$ is placeholder for character's text description, $\\langle \\text{M\\_Placeholder} \\rangle$ for motion token sequences, and $\\langle \\text{S\\_Placeholder} \\rangle$ for speech token sequences.", + "bbox": [ + 511, + 494, + 905, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training", + "text_level": 1, + "bbox": [ + 513, + 565, + 617, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The training of SOLAMI adopts a three-stage strategy.", + "bbox": [ + 511, + 588, + 872, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage 1: Tokenizer Training. The training approach for the motion tokenizer uses the fashion of [28]. The train loss is", + "bbox": [ + 511, + 603, + 905, + 646 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {e} \\mathcal {L} _ {e} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {v} \\mathcal {L} _ {v}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 650, + 903, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_r$ means reconstruction loss, $\\mathcal{L}_e$ embedding loss, $\\mathcal{L}_c$ commitment loss, $\\mathcal{L}_v$ velocity loss, and $\\lambda_*$ are manually adjusted weights. For the speech tokenizer, we use the pre-trained checkpoint from AnyGPT [60]. We freeze the tokenizers' weights after this stage.", + "bbox": [ + 511, + 672, + 905, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stage 2: Multi-task Pre-training for Modality Alignment. As shown in Fig. 2, the second stage is multi-task pre-training, achieving modality alignment between motion and text, as well as between speech and text. It is necessary because motion data is scarce, and direct training on multimodal interaction data results in suboptimal models, as demonstrated in subsequent ablation studies. For motion and text alignment, we use $46\\mathrm{K}$ motion-text pairs for text-to-motion generation and motion captioning tasks, and $11\\mathrm{K}$ interactive motion pairs", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26889", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2319dea11b5e8a6fed26e0c49dd3742d1f98014a560f6b5eddec7c09e28c05fe.jpg", + "image_caption": [ + "Figure 2. Training pipeline of SOLAMI. We train SOLAMI through a three-stage process. In the pre-training stage, we train the model with motion-text and speech-text related tasks to align the speech and motion modalities with language. During the instruction tuning stage, we train the model with social multimodal multi-round interaction data, enabling it to generate multimodal responses that align with the character settings and the context of the topic." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 893, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for two-person motion generation. To align the speech and text, we train the model with $410\\mathrm{K}$ speech-text pairs for text-to-speech and automatic speech recognition tasks, and $100\\mathrm{K}$ speech dialogue pairs for speech-to-speech generation. The tasks are formulated as \"User: ; Character: \". To balance the scale disparity between the motion and speech data, we sampled them at a 4:6 ratio during training.", + "bbox": [ + 89, + 424, + 483, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 3: Instruction Tuning for Multi-turn Conversat- tion. In the third stage, we perform instruction tuning by applying supervised fine-tuning with multimodal interaction data, enabling the model to handle long-sequence, multi-turn dialogues, as shown in Fig. 2. We utilize $5.7\\mathrm{K}$ multimodal conversation items for supervised fine-tuning. Each conversation item is fed to the model in the format as the Interaction Template in Sec. 3.1. We experiment with two approaches: full-parameter fine-tuning and LoRA finetuning [23]. We supervised only the character's response to teach the model how to react to the user's behavior. Thus we train the model by maximizing the log-likelihood of the next-token prediction and the train loss is:", + "bbox": [ + 89, + 560, + 483, + 756 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {I T}} = - \\sum_ {r = 1} ^ {R} \\sum_ {i = 1} ^ {L _ {M} ^ {r}} \\log p _ {\\Theta} (\\hat {m} _ {i} ^ {r} | \\hat {m} _ {i - 1} ^ {r}, \\dots , \\hat {m} _ {1} ^ {r}, \\hat {S} _ {< r}, \\hat {M} _ {< r}) \\\\ - \\sum_ {r = 1} ^ {R} \\sum_ {i = 1} ^ {L _ {S} ^ {r}} \\log p _ {\\Theta} \\left(\\hat {s} _ {i} ^ {r} \\mid \\hat {s} _ {i - 1} ^ {r}, \\dots , \\hat {s} _ {1} ^ {r}, \\hat {S} _ {< r}, \\hat {M} _ {\\leq r}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 768, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Theta$ is the network parameter of the LLM backbone or LoRA weights, $R$ is the conversation round, $\\hat{S}_r$ and $\\hat{M}_r$", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are the $r$ -th round speech and motion token sequences with lengths $L_{M}^{r}$ and $L_{S}^{r}$ , respectively.", + "bbox": [ + 511, + 424, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. SynMSI Dataset", + "text_level": 1, + "bbox": [ + 511, + 467, + 674, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Social interaction between users and virtual characters is inherently unique, which makes collecting such multimodal interaction data particularly challenging. Currently, available public datasets [21, 40, 56] are incomplete for our task. To address this issue, we propose a data synthesis pipeline that leverages existing motion-text datasets, text-based role-play models, and speech synthesis methods and generates a large-scale multimodal dialogue dataset, SynMSI.", + "bbox": [ + 511, + 492, + 906, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Motion Data", + "text_level": 1, + "bbox": [ + 511, + 622, + 648, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect motion-text data for two purposes: first, to achieve alignment between motion and text during pretraining, and second, to generate multimodal data for instruction tuning. Since our work focuses on modeling social interactions, we select existing datasets that contain rich social actions: HumanML3D [21] with $24\\mathrm{K}$ motion-text pairs, Inter-X [56] with $20\\mathrm{K}$ motion-text pairs and 10 K two-person motion pairs, and DLP-MoCap [13] with 2 K motion-text pairs. Since the Inter-X [56] dataset contains only text descriptions of two-person interactive motion without descriptions of individual motion, we used GPT-4o [41] to decompose the two-person action descriptions into single-person motion-text pairs. Additionally, we used GPT-4o [41] to synthesize comprehensive text descriptions for each motion clip by consolidating multiple possible descriptions, thereby providing more detailed textual annotations that preserve motion details.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26890", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8b8c986f2ba86c131ae2e8df79e3ec0882a12d91b99c1ab5ed2be0f5b8499f9d.jpg", + "image_caption": [ + "Figure 3. SynMSI dataset generation. Our synthesizing pipeline consists of 4 steps. Based on numerous character-relevant topics and state-of-the-art LLMs [41], we generate text scripts for multimodal dialogues. Using a large-scale motion database [13, 21, 56], we retrieve the most appropriate motions and refine the speech scripts accordingly. Finally, we employ TTS/voice cloning [14] to generate character-specific speech. This approach enables us to create multimodal interaction data of various characters using only existing motion datasets." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 568, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/90fd8d13848f8105c4f4d7d1122c58b7728eeaf2fd7ea279e6049fadaf97e53e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 95, + 893, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Speech Data", + "text_level": 1, + "bbox": [ + 89, + 431, + 225, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use speech-text data for speech-text alignment in the pre-training stage. Speech datasets involve CommonVoice [8] (150 K speech-text pairs in our experiments), AnyInstruct [60] (200 K speech-text pairs and 100 K speech-to-speech items), and synthetic speech data (60 K speech-text pairs) by text-to-speech approaches (Azure TTS and XTTS_v2 [14]).", + "bbox": [ + 89, + 455, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multimodal Data Synthesizing", + "text_level": 1, + "bbox": [ + 89, + 574, + 362, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous data synthesis methods [32, 37, 66] usually use the general abilities of advanced LLMs [41] and text-annotated multimodal data to generate synthetic data. However, generating social multimodal interaction data has not yet been achieved. This challenging task requires high-quality expression of body language, voice consistency that matches the characters, and suitable dialogue content.", + "bbox": [ + 89, + 597, + 483, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our synthesizing pipeline includes 4 steps. (1) First, we collect $5.3\\mathrm{K}$ character-related and daily topics from internet platforms (Google Trends [2], Zhihu [5], Jike) and brainstorms of GPT-4o [41] to improve the diversity of the dialogue contents. (2) Based on the topic, character setting, and previous round of scripts, we use GPT-4o [41] to generate textual descriptions (motion, speech, expression etc.) for the next round of the dialogue. (3) Then we utilize the text embedding [39] of the motion description to retrieve the most relevant motions from our meticulously curated motion-text database. (4) Moreover, we refine the generated speech text with the retrieved motions to ensure that speech and motion are well-coordinated.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The motion database with detailed text annotations and the refinement process can alleviate the misalignment between the real motion and speech in the LLM-Agent method [13]. By iteratively repeating this process, we can generate multi-round dialogue contents across many characters, where the motions are sourced from the motion database, and the speeches are synthesized using TTS/voice cloning (Azure TTS and XTTS_v2 [14]) to maintain consistency with the character's voice style. We finally obtained $6.3\\mathrm{K}$ multimodal dialogue items.", + "bbox": [ + 511, + 433, + 906, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. VR Interface", + "text_level": 1, + "bbox": [ + 511, + 602, + 650, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate our method directly, we developed a VR interface with an Oculus Quest 3 frontend and a backend service, as shown in Fig. 4. The frontend enables immersive interaction between users and 3D autonomous characters, while the backend, powered by 2 H800 GPUs, supports the computation of various baselines in our experiments. During usage, the VR headset captures the user's speech and body motion, and sends them to the backend computation nodes. For motion capture, we use the Quest's full-body tracking system [55] to obtain pose parameters, which are then retargeted onto an SMPL-X model [43]. The computation nodes generate the body motion parameters and speech responses of the character based on the multimodal input. Then we apply an audio-to-face method, UniTalker [17], to generate the character's facial animation parameters. The facial and body parameters are jointly retargeted onto a 3D character model [58], completing one cycle of social interaction. For a natural user experience, we employ preset idle", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26891", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/516a93c459e60f9f5c0fda17fe26afbc63d8ab83fbfa8702af7e6f6b6f7b38f5.jpg", + "image_caption": [ + "Figure 4. VR interface architecture. Our VR project consists of a Quest 3 client and a server. The Quest client captures and transmits user body motion and speech to the server. The server then generates character's speech, body motion, and face blendshape parameters based on the selected methods. The response is then sent back to the Quest client to drive the character." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "motions on the character when the method is LLM+Speech or the character is waiting for the user's input.", + "bbox": [ + 89, + 348, + 482, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Character Assets. Our 3D character portfolio covers a diverse range of entities, including AI assistant avatars, famous cinematic roles, internet memes, and real-world celebrity personas. These 3D models are sourced from open-source repositories under CC Attribution-NonCommercial-ShareAlike License as well as our manual creation using VRoid Studio [4]. We subsequently employ facial rigging, skinning, bone chain simulation, retargeting, and texture and material creation in Unity Engine processes to yield functional 3D character assets.", + "bbox": [ + 89, + 378, + 482, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 547, + 223, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 574, + 297, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiment, we selected the AnyGPT-base model [60] (based on LLaMA2-7B [51]) as the backbone for SOLAMI, because it is an open-source model available at our experimental time that supports end-to-end speech processing. During the pre-training stage, we utilize 32 V100 GPUs to train the model for $3\\mathrm{K}$ steps (batch size 256, learning rate 4e-5). For instruction tuning, we train the SOLAMI for 800 steps using 16 V100 GPUs (batch size 48, learning rate 2e-5). For LoRA fine-tuning [23], we set the rank as 8 and alpha as 16. We split the synthesized multimodal data into training and test sets with a 9:1 ratio. We use DeepSpeed [45] to accelerate the training. During testing, we evaluate each round of the character's response.", + "bbox": [ + 89, + 597, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To validate the performance improvement in social interaction brought by incorporating 3D modalities (such as body motion), we compared SOLAMI with the LLM+Speech and the AnyGPT (fine-tune) approach. For the LLM+Speech framework, the user's speech is first transcribed into text using ASR techniques, which is then fed to a LLM to generate the character's response in text, and sub", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sequentially converted into speech using TTS. For fairness, we use LLaMA2-7B-Chat [51] as the LLM backbone, Whisper large-v3 [44] for ASR, and XTTS_v2 [14] for voice cloning. For the AnyGPT (fine-tune) framework, we use the speech data of SynMSI to train the AnyGPT-base model [60] with the same parameter settings. To compare the effectiveness of the LLM-Agent architecture with the social VLA framework, we used DLP [13] as a baseline method. In DLP framework, the user's speech and body motion are separately processed as text descriptions by ASR and motion captioning modules. Based on the input text descriptions, LLM generates the character's text instructions of speech and motion, which are transferred into speech and body motion by TTS and motion generation module. The speech component of DLP follows the LLM+Speech method. Considering that MoMat-MoGen module in DLP is too slow for user interaction (over 5 seconds latency), we use MotionGPT [28] for motion captioning and motion generation. To ensure a fair comparison, we used the same motion data as the pre-training stage of SOLAMI to train MotionGPT. Additionally, we conducted ablation experiments on the effect of the pre-training stage, marked as (w/o pretrain). For the ablation study of the motion tokenizer, please refer to the supplementary materials. We use vLLM [31] to accelerate the LLM backbones for low latency interaction.", + "bbox": [ + 511, + 90, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 481, + 735, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted quantitative evaluation for our method and all the baselines mentioned in Sec. 6.1.", + "bbox": [ + 511, + 503, + 906, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For motion, we evaluate the model responses using metrics including FID, diversity, PAMPJPE (mm), and angle error [22]. Following Duolando [48], we obtain FID and diversity using motion features from AIST++ [33]. For speech, we use VC similarity [57] to evaluate the voice similarity with the character. To evaluate the content quality of speech, we first use Whisper-large-v3 [44] to transcribe the speech into text. Then following [47, 52], we employ GPT-4o [41] as the judge to assess Context Relevance and Character Consistency on a Likert scale ranging from 1 to 5. Context Relevance indicates whether the speech content aligns with the topic and context of the conversation, while Character Consistency assesses whether the content adheres to the character settings. For inference latency (seconds), we deploy all the models on 2 H800 GPUs with vLLM [31] framework and asynchronous mechanisms to improve performance while maintaining fairness.", + "bbox": [ + 511, + 536, + 908, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. The quantitative results in Tab. 1 demonstrate that, using the same foundation model (Llama2) [51] as the backbone, our method, SOLAMI (full params), significantly outperforms other methods in terms of motion quality and inference latency.", + "bbox": [ + 511, + 809, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion Quality. SOLAMI demonstrates superior perfor", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26892", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cd0a622a372203de9c0c79b8ca70d5ded664638a386dddd025e5e34cd42991e3.jpg", + "table_caption": [ + "Table 1. Quantitative results of baselines and SOLAMI. $\\uparrow^{\\prime}(\\downarrow^{\\prime})$ indicates that the values are better if the metrics are larger (smaller). We run all the evaluations 5 times and report the average metric. The best results are in bold and the second best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsMotion MetricsSpeech MetricsInference Latency↓
FID↓Diversity↑PA-MPJPE↓Angle Error↓VC Similarity↑Context Relevance↑Character Consistency↑
SynMSI Dataset-9.136---4.8884.893-
LLM+Speech (Llama2) [51]----0.8183.5273.8593.157
AnyGPT (fine-tune) [60]----0.8193.5023.8032.588
DLP (MotionGPT) [13]4.2548.259165.0530.4950.8123.5773.7855.518
SOLAMI (w/o pretrain)5.0528.558159.7090.3870.8203.5413.4612.657
SOLAMI (LoRA)15.7298.145167.1490.4000.7703.2513.4232.710
SOLAMI (full params)3.4438.853151.5000.3600.8243.6343.8242.639
", + "bbox": [ + 93, + 127, + 906, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance compared with the DLP method [13] across multiple motion metrics. This is because the social VLA model achieves comprehensive modality alignment among speech, motion, and language through training on our high-fidelity character-specific multimodal data. Our model can precisely perceive the user's physical motions and linguistic cues, enabling semantically rich interactive motions in response. This ability contrasts with the LLM-Agent architectures of the DLP method, which exhibits limitations in conveying multimodal nuances through text-only intermediary representation.", + "bbox": [ + 88, + 242, + 480, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speech Quality. Our method demonstrates the capability to synthesize a voice tone that matches the character with a higher Voice Cloning (VC) Similarity score. Our model also shows better performance on the context relevance of the dialogue than other methods. Because the inclusion of the motion modality enables the model to perceive the user's body language, while the LLM+Speech or AnyGPT(fine-tune) method lacks this capability. In terms of character consistency, our model achieves secondary performance metrics. We suspect this may be due to the incorporation of motion and speech modalities, which potentially affects the character-related knowledge embedded within the original LLM. The performance degradation is similar to the observations in [16, 36].", + "bbox": [ + 89, + 411, + 482, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference Latency. Our end-to-end approach is significantly superior to the modular pipeline approaches (LLM+Speech or DLP). Because the end-to-end VLA model naturally aligns with the process of real-time human communication. Theoretically, if we could collect data on real-time interactions between humans and characters, our method could achieve full-duplex streaming interaction.", + "bbox": [ + 89, + 626, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. As shown in Tab. 1, the pre-training stage of SOLAMI leads to better performance in both motion and speech quality. We believe that the pre-training stage, which aligns motion, speech, and language, facilitates the model's ability to learn the multimodal dialogue skill during the instruction tuning stage. Instruction tuning using LoRA [23] shows weaker results compared to full parameter fine-tuning. We think that the gap between the data distribution of pre-training tasks and the multimodal instruction tuning task is substantial, and LoRA fine-tuning alone is insufficient for the model to learn the strong multimodal", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f87743c6f98aea5e739cc80d0989cf505918fa5e6d2ab9d1e0cfdd34f00a53a.jpg", + "table_caption": [ + "Table 2. Questionnaire settings of our user study." + ], + "table_footnote": [], + "table_body": "
DimensionQuestions
Motion CoherenceDoes the motion match the character's setting?Does the motion align well with speech?
Motion InteractionDoes the character follow motion instructions correctly?Does the character understand user's motion?
Speech ConsistencyDoes the speech match the character's setting?Is the speech relevant to the current topic?
Overall ExperienceHow would you rate the overall experience?
", + "bbox": [ + 516, + 265, + 906, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77940dc7c538e9e7b11b4f79aed699930c925aa63230fa19d4e9a7982fd51406.jpg", + "image_caption": [ + "Figure 5. Results of the user study with $95\\%$ confidence." + ], + "image_footnote": [], + "bbox": [ + 519, + 375, + 901, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dialog ability.", + "bbox": [ + 511, + 565, + 607, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. VR Interface User Study", + "text_level": 1, + "bbox": [ + 511, + 590, + 740, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation of a single modality alone cannot fully compare the 3D autonomous characters built on different frameworks. To address this, we conducted user study with a VR interface. As shown in Fig. 4, we employ LLM+Speech, DLP (MotionGPT), and SOLAMI as method options of the server backend and the same VR frontend across different methods. This implies that the variable in the experiment is the driving method in the server. Users are asked to engage in more than five rounds of dialogue with the VR character before completing our questionnaire.", + "bbox": [ + 509, + 612, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. The indicators and corresponding questions for our questionnaire are shown in Tab. 2. Motion Coherence evaluates whether the character's motion aligns with the character setting and dialogue content. Motion Interaction assesses whether the character can understand the semantics of the body motion and effectively interact with the user. Speech Coherence examines whether the generated speech aligns with the context and character settings.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26893", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/088896000ac6620228c85e6cb79d98c7e5d84c86e3527e5f156506411bb69003.jpg", + "image_caption": [ + "Figure 6. Qualitative results of SOLAMI and baselines, and the user workflow for VR experience. Our social VLA model, trained in an end-to-end strategy on SynMSI dataset, can accurately perceive the semantic information embedded within users' speech and motion input, and subsequently generate natural and coherent responses." + ], + "image_footnote": [], + "bbox": [ + 89, + 95, + 906, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "And the Overall Experience measures the user's satisfaction with the overall experience. Each question is rated on a 1 to 5 Likert scale, with higher scores indicating greater satisfaction. The score for each dimension is calculated as the average score of its corresponding questions. We ultimately collected 60 survey responses, with participants from various gender and age groups.", + "bbox": [ + 89, + 527, + 482, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. As shown in the Fig. 5, our method achieved the highest user satisfaction across all dimensions. SOLAMI demonstrates superior performance over the DLP method across all metrics, validating the effectiveness of an end-to-end social VLA model in character behavior modeling. While the DLP method shows lower speech consistency compared to the LLM+Speech method, it excels in motion-related metrics and achieves higher overall satisfaction, indicating that effective body language can enhance user experience despite speech quality limitations.", + "bbox": [ + 89, + 638, + 482, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To provide a more intuitive demonstration of our model's capabilities, we replay the user study experiments and render them, as shown in the Fig. 6. The results indicate that SOLAMI demonstrates excellent capabilities in body language understanding, motion command execution, and body interaction. For better understanding, we also present the workflow from a first-person view during actual usage.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We encourage readers to explore our supplementary materials and video for a more detailed overview of the model experiments, data generation pipeline, VR interface construction process, and comprehensive experimental results.", + "bbox": [ + 511, + 527, + 906, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 598, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SOLAMI, an approach for building 3D autonomous characters. This approach includes three key components: 1) Architecture: A novel social VLA modeling framework enabling multimodal social interaction; 2) Multimodal Data Synthesizing: A pipeline for automatically generating multimodal interaction data from existing incomplete datasets; 3) VR Interface: A VR engineering framework that facilitates immersive interactions between users and various characters. Together, these modules contribute to an enhanced user interaction experience.", + "bbox": [ + 511, + 625, + 906, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 513, + 781, + 671, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 511, + 806, + 905, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26894", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Character.ai. https://character.ai.2", + "[2] Google trends. https://trends.google.com/trends.5", + "[3] Talkie ai. https://www-talkie-ai.com. 2", + "[4] Vvoid studio. https://vvoid.com/en/studio.6", + "[5] Zhihu. https://www.zhihu.com. 5", + "[6] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG, 2023. 2", + "[7] Tenglong Ao. Body of her: A preliminary study on end-to-end humanoid agent. arXiv preprint arXiv:2408.02879, 2024. 2", + "[8] Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In LREC, 2020. 5", + "[9] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. TASLP, 2023. 3", + "[10] Zalán Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi. Soundstorm: Efficient parallel audio generation. arXiv preprint arXiv:2305.09636, 2023. 3", + "[11] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In RSS, 2023. 2, 3", + "[12] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In NeurlPS, 2023. 2", + "[13] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, Liang Pan, Xiangyu Fan, Han Du, Peng Gao, Zhitao Yang, Yang Gao, Jiaqi Li, Tianxiang Ren, Yukun Wei, Xiaogang Wang, Chen Change Loy, Lei Yang, and Zwei Liu. Digital life project: Autonomous 3d characters with social intelligence. In CVPR, 2024. 2, 4, 5, 6, 7", + "[14] Edresson Casanova, Kelly Davis, Eren Gölgé, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Meyer, Reuben Morais, Samuel Olayemi, et al. Xtts: a massively multilingual zero-shot text-to-speech model. arXiv preprint arXiv:2406.04904, 2024. 5, 6", + "[15] James J Cummings and Jeremy N Bailenson. How immersive is enough? A meta-analysis of the effect of immersive technology on user presence. Media psychology, 2016. 2", + "[16] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied multimodal language model. In ICML, 2023. 3, 7", + "[17] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. 2024. 5", + "[18] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Chatpose: Chatting about 3d human pose. In CVPR, 2024. 2", + "[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenjiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 3", + "[20] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, Fu-Jen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abrham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26895", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, Anush Kumar, Robert Kuo, Sach Lakhavani, Miao Liu, Mi Luo, Zhengyi Luo, Brighid Meredith, Austin Miller, Oluwatuminu Oguntola, Xiaqing Pan, Penny Peng, Shraman Pramanick, Merey Ramazanova, Fiona Ryan, Wei Shan, Kiran Somasundaram, Chenan Song, Audrey Southerland, Masatoshi Tateno, Huiyu Wang, Yuchen Wang, Takuma Yagi, Mingfei Yan, Xitong Yang, Zecheng Yu, Shengxin Cindy Zha, Chen Zhao, Ziwei Zhao, Zhifan Zhu, Jeff Zhuo, Pablo Arbeláez, Gedas Bertasius, Dima Damen, Jakob Engel, Giovanni Maria Farinella, Antonino Furnari, Bernard Ghanem, Judy Hoffman, C. V. Jawahar, Richard A. Newcombe, Hyun Soo Park, James M. Rehg, Yoichi Sato, Manolis Savva, Jianbo Shi, Mike Zheng Shout, and Michael Wray. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. In CVPR, 2024. 3", + "[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4, 5", + "[22] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2, 3, 6", + "[23] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4, 6, 7", + "[24] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2, 3", + "[25] Sarah Hudson, Sheila Matson-Barkat, Nico Pallamin, and Guillaume Jegou. With or without you? Interaction and immersion in a virtual reality experience. Journal of business research, 2019. 2", + "[26] Brian Ichter, Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, Dmitry Kalashnikov, Sergey Levine, Yao Lu, Carolina Parada, Kanishka Rao, Pierre Sermanet, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Mengyuan Yan, Noah Brown, Michael Ahn, Omar Cortes, Nicolas Sievers, Clayton Tan, Sichun Xu, Diego Reyes, Jarek Rettinghouse, Jornell Quiambao, Peter Pastor, Linda Liuu, Kuang-Huei Lee, Yuheng Kuang, Sally Jesmonth, Nikhil J. Joshi, Kyle Jeffrey, Rosario Jauregui Ruano, Jasmine Hsu, Keerthana Gopalakrishnan, Byron David, Andy Zeng, and Chuyuan Kelly Fu. Do as I can, not as I say: Grounding language in robotic affordances. In CoRL, 2022. 2, 3", + "[27] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2", + "[28] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 3, 6" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Paul Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 3", + "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuok Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation. In ICML, 2024. 3", + "[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6", + "[32] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5", + "[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In ICCV, 2021. 6", + "[34] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. IJCV, 2024. 2", + "[35] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 2", + "[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. VILA: on pre-training for visual language models. In CVPR, 2024. 7", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 5", + "[38] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, 2024. 3", + "[39] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005, 2022. 5", + "[40] Evonne Ng, Javier Romero, Timur M. Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In CVPR, 2024. 2, 4", + "[41] OpenAI. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 3, 4, 5, 6" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26896", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Joon Sung Park, Joseph C. O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3", + "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3, 5", + "[44] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 6", + "[45] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD, 2020. 6", + "[46] Giuseppe Riva, Fabrizia Mantovani, Claret S. Capdeville, Alessandra Preziosa, Francesca Morganti, Daniela Villani, Andrea Gaggioli, Cristina Botella, and Mariano Alcaniz Raya. Affective interactions using virtual reality: The link between presence and emotions. Cyberpsychology Behav. Soc. Netw., 2007. 2", + "[47] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. Character-llm: A trainable agent for role-playing. In EMNLP, 2023. 2, 6", + "[48] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower GPT with off-policy reinforcement learning for dance accompaniment. In ICLR, 2024. 2, 6", + "[49] Mel Slater, Daniel Pérez Marcos, Henrik Ehrsson, and Maria V Sanchez-Vives. Inducing illusory ownership of a virtual body. Frontiers in neuroscience, 2009. 2", + "[50] Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive architectures for language agents. 2024. 2", + "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Quan Tu, Shilong Fan, Zihang Tian, Tianhao Shen, Shuo Shang, Xin Gao, and Rui Yan. Charactereval: A chinese benchmark for role-playing conversational agent evaluation. In ACL, 2024. 6", + "[53] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. TMLR, 2024. 3", + "[54] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. TMLR, 2022. 2", + "[55] Alexander W. Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia, 2022. 5", + "[56] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, Yunhui Liu, Wenjun Zeng, and Xiaokang Yang. Inter-x: Towards versatile human-human interaction analysis. In CVPR, 2024. 2, 4, 5", + "[57] Dongchao Yang, Jinchuan Tian, Xu Tan, Rongjie Huang, Songxiang Liu, Xuankai Chang, Jiatong Shi, Sheng Zhao, Jiang Bian, Xixin Wu, Zhou Zhao, and Helen Meng. Uniaudio: An audio foundation model toward universal audio generation. arXiv preprint arXiv:2310.00704, 2023. 6", + "[58] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In ICCV, 2023. 5", + "[59] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. TASLP, 2022. 3", + "[60] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, Hang Yan, Jie Fu, Tao Gui, Tianxiang Sun, YuGang Jiang, and Xipeng Qiu. Anygpt: Unified multimodal LLM with discrete sequence modeling. In ACL, 2024. 3, 5, 6, 7", + "[61] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. In EMNLP, 2023. 3", + "[62] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2", + "[63] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech large language models. arXiv preprint arXiv:2308.16692, 2023. 3", + "[64] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2, 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26897", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. In CVPR, 2024. 2", + "[66] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 5", + "[67] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3" + ], + "bbox": [ + 91, + 92, + 480, + 425 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "26898", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_model.json b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_model.json new file mode 100644 index 0000000000000000000000000000000000000000..50548d88d721bb76f0b393db0e5f745569f37952 --- /dev/null +++ b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_model.json @@ -0,0 +1,2138 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.8, + 0.177 + ], + "angle": 0, + "content": "SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.203, + 0.808, + 0.311 + ], + "angle": 0, + "content": "Jianping Jiang\\*†, Weiye Xiao\\*‡, Zhengyu Lin\\(^{1}\\), Huaizhong Zhang\\(^{1}\\), Tianxiang Ren\\(^{1}\\), Yang Gao\\(^{1}\\), Zhiqian Lin\\(^{1}\\), Zhongang Cai\\*1,2, Lei Yang\\*1, Ziwei Liu\\*2 \n\\(^{1}\\)SenseTime Research, \\(^{2}\\)S-Lab, Nanyang Technological University \n* Equal Contribution, † Project Lead, ‡ Engineering Lead, § Corresponding Author \nhttps://solami-ai.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.316, + 0.904, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.659, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Figure 1. SOLAMI enables the user to interact with 3D autonomous characters through speech and body language in an immersive VR environment via an end-to-end social vision-language-action model, which is trained on our synthesized multimodal dataset SynMSI." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.7, + 0.327, + 0.715 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Human beings are social animals. How to equip 3D autonomous characters with similar social intelligence that can perceive, understand and interact with humans remains an open yet foundational problem. In this paper, we introduce SOLAMI, the first end-to-end Social vision-Language-Action (VLA) Modeling framework for Immersive interaction with 3D autonomous characters. Specifically, SOLAMI builds 3D autonomous characters from three aspects: 1) Social VLA Architecture: We propose a unified social VLA framework to generate multimodal response (speech and motion) based on the user's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.883 + ], + "angle": 0, + "content": "multimodal input to drive the character for social interaction. 2) Interactive Multimodal Data: We present SynMSI, a synthetic multimodal social interaction dataset generated by an automatic pipeline using only existing motion datasets to address the issue of data scarcity. 3) Immersive VR Interface: We develop a VR interface that enables users to immersively interact with these characters driven by various architectures. Extensive quantitative experiments and user study demonstrate that our framework leads to more precise and natural character responses (in both speech and motion) that align with user expectations with lower latency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "26887" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.281 + ], + "angle": 0, + "content": "Have you ever imagined having an immersive face-to-face conversation with a character you deeply admire? Not merely through speech dialogue, but through an interaction where you can observe its subtle facial expressions, natural body language, and even fleeting emotional changes. Psychology research [15, 25, 46, 49] shows that in social interactions, the greater the level of immersion, the better the human experience. However, current character agents [1, 3, 47] are still limited to text or voice interactions. This limitation prompts us to build 3D autonomous characters with richer modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.482, + 0.463 + ], + "angle": 0, + "content": "Developing an autonomous 3D character requires effectively modeling its behavior system, which involves two major challenges: 1) The 3D character needs to accurately observe and understand the information conveyed by the user, and respond appropriately based on the context and its character setting through speech, body motion, and facial expression, etc. This goes beyond previous singular human-related tasks, such as motion generation [62], motion understanding [27], and audio-to-motion [6]. 2) Data for multimodal interactions between users and 3D characters is extremely scarce due to the prohibitive cost of the comprehensive setup." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Previous work [13] is primarily based on the LLM-Agent framework, using text to link various sub-modules (such as motion captioning and text-to-motion). While this approach performs well in high-level tasks like planning and memory, it tends to fall short in tasks such as understanding user behaviors and providing timely body motion responses. This limitation arises because using text as the intermediary between modules conveys high-level information but often omits subtle nuances. And the sub-modules (motion captioning, speech recognition, etc.) in the complex engineering framework incur substantial latency that undermines the timeliness of a natural communication [50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Interestingly, research on robotics shows the similar conclusion. The LLM-Agent framework can handle planning tasks [26], but for low-level manipulation tasks, end-to-end Vision-Language-Action (VLA) models built upon LLMs show superior performance [11, 24, 29, 67]. We argue that digital avatars are essentially robots with virtual humanoid embodiment. Therefore, building a VLA model for social interactions with users is a promising direction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.886 + ], + "angle": 0, + "content": "In this paper, we implement an end-to-end social VLA model, SOLAMI. Our model, built upon a decoder-only LLM backbone, processes the inputs of user speech and motion into discrete representations, and generate the responsive speech and motion tokens, which are then decoded to the character's speech and motion. This modeling approach can effectively learn character behavior patterns across motion and speech modalities and offer low latency." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Although there are numerous datasets related to human" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "social behaviors [13, 34, 40, 56], comprehensive multimodal interaction datasets remain scarce. Thus, we introduce a data synthesis method that utilizes existing text-motion datasets to automatically construct multimodal interaction data at a low cost. Leveraging our extensively curated topics (5.3 K), uniformly processed motion database (46 K), and iterative script refinement pipeline, we develop SynMSI, a dataset containing 6.3 K multi-turn multimodal conversation items. To evaluate the effectiveness of our method, we developed a VR interface where users can immersively interact with various 3D characters. Quantitative experimental results and user study analysis show that our approach produces more precise and natural social interaction experience with lower latency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.41 + ], + "angle": 0, + "content": "To summarize, we contribute 1) A new VLA architecture to model the character's behavior system for immersive social interaction; 2) A dedicated synthesizing pipeline that automatically generates large-scale multimodal interactive dataset, SynMSI; 3) An immer VR interface for users to interact with various characters through speech and motion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.655, + 0.441 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.734, + 0.466 + ], + "angle": 0, + "content": "2.1. Human Motion & LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.905, + 0.822 + ], + "angle": 0, + "content": "Existing role-play agents primarily rely on text [47], speech [1, 3, 41], or video [7] as interactive media, but building 3D autonomous characters means modeling 3D embodied behavior, especially body language. Unlike single-purpose human motion tasks [6, 12, 22, 34, 48, 56, 62], we expect this 3D character not only to comprehend user speech and body language but also to respond according to its profile setting. Large language models (LLMs) [41, 51], with their remarkable emergent abilities [54], provide promising solutions for this direction. One approach [18, 28, 64, 65] tries to integrate LLMs and human motion in an end-to-end fashion, enabling a single model to perform multiple motion-related tasks. However, the goal of these methods is not to generate responsive motion based on the input motion according to the character setting. Another approach [13, 35] utilizes the LLM as a versatile brain center that controls various submodules (e.g., motion understanding, text-to-motion generation) with text or code instructions. However, this modular approach inherently introduces information loss and engineering-related time latency. Therefore, how to create feasible autonomous 3D characters for immersive interaction remains an open challenge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.724, + 0.849 + ], + "angle": 0, + "content": "2.2. Embodied Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Building 3D autonomous characters means creating virtual humanoid agents with embodied intelligence. In the field of embodied intelligence, researchers have found the high-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26888" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "level abilities (planning, memory, etc.) of LLM-Agents across various environments, including factories [16, 26], 2D sandbox settings [42], and 3D gaming spaces [53]. For tasks requiring low-level skills, such as manipulation, end-to-end VLA models [11, 24, 29, 67] have shown considerable potential. Despite existing efforts [19, 20, 22] in egocentric human-related tasks, the capabilities of VLA models in social interactive tasks with humans have not been fully explored." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.243, + 0.46, + 0.261 + ], + "angle": 0, + "content": "3. Social Vision-Language-Action Modeling" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.268, + 0.226, + 0.282 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.472 + ], + "angle": 0, + "content": "Our framework is an end-to-end social VLA model that takes the user's speech and motion as input and generates the character's responsive speech and motion as output. In this process, speech and motion are added as new languages to the LLM text vocabulary. First, the user's speech and motion are converted into discrete motion tokens and speech tokens via a motion tokenizer and a speech tokenizer, respectively. A decoder-only LLM backbone then predicts the character's output motion and speech tokens based on the user's input tokens and the character setting. The generated tokens are subsequently decoded into corresponding motion and speech by their respective decoders." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Motion Representation. To take advantage of SMPL-X's [43] compatibility with industry animation workflow, we directly model human poses as SMPL-X joint rotations instead of keypoint positions [28, 38] to facilitate animation of characters down the stream." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Motion Tokenizer. Our motion tokenizer employs a Vector Quantized Variational Autoencoders (VQ-VAE) structure as [28, 64]. It learns discrete representations of motion, enabling the LLM to understand the text-motion connection. We design separate VQVAEs \\((Q^{b}, Q^{h}, Q^{t})\\) for the body motion \\(\\mathbf{m}^{b}\\), hand motion \\(\\mathbf{m}^{h}\\), and inter-character relative transform (rotation and translation) \\(\\mathbf{m}_{t}\\) for higher reconstruction accuracy [38]. The quantization process for motion \\(\\mathbf{m}^{u}\\) is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.696, + 0.483, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\hat {m} _ {t} ^ {u} = Q ^ {u} \\left(\\mathbf {m} _ {t} ^ {u}\\right) = \\arg \\min _ {z _ {i} \\in \\mathbb {Z} _ {u}} \\| \\mathbf {m} _ {t} ^ {u} - z _ {i} \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbb{Z}_u\\) is the codebook of motion part \\(u\\in \\{b,h,t\\}\\) and \\(\\hat{m}_t^u\\) is the corresponding motion tokens. The VQVAEs of body and hand, \\(Q^{b}\\) and \\(Q^{h}\\), apply 1D convolutions on motion features along the temporal dimension to get \\(L_{M}\\) sequential tokens \\(\\hat{M}^b = [\\hat{m}_1^b,\\hat{m}_2^b,\\dots,\\hat{m}_{L_M}^b]\\) and \\(\\hat{M}^h = [\\hat{m}_1^h,\\hat{m}_2^h,\\dots,\\hat{m}_{L_M}^h]\\), and the VQVAE \\(Q^{t}\\) uses MLPs to get the transform token \\(\\hat{m}^t\\) of a sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Speech Tokenizer. Research [9, 61] on speech discretization mainly utilizes the RVQ-VAE structure [59]. In this work, we utilize the SpeechTokenizer [63] that disentangles semantic and acoustic information within speech \\( S \\). This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.182 + ], + "angle": 0, + "content": "allows us to use the semantic tokens \\(\\hat{S} = [\\hat{s}_1^s, \\hat{s}_2^s, \\dots, \\hat{s}_{L_S}^s]\\) from the first layer as input to the LLM, reducing inference costs (\\(L_S\\) is the sequence length of semantic tokens). Simultaneously, a short sample of the character's voice (4 to 6 seconds) can be used as a prompt to achieve instance voice cloning when decoding through SoundStorm [10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Multi-modal Multi-round Interaction. User-character interaction is formulated as a multi-round conversation fashion of common LLMs [41, 51]. When the user sends speech and motion to SOLAMI, the model auto-regressively generates speech and motion responses based on previous dialogue contents and the character setting. To facilitate training, we use special tokens to mark the start and end of each modality sequence as [28, 30]. The interaction process can be formulated as follow:" + }, + { + "type": "title", + "bbox": [ + 0.541, + 0.33, + 0.68, + 0.345 + ], + "angle": 0, + "content": "Interaction Template" + }, + { + "type": "code", + "bbox": [ + 0.541, + 0.358, + 0.857, + 0.479 + ], + "angle": 0, + "content": "Input: \nSystem Prompt: \nUser: \\(< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>\\) \nCharacter: \\(< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>\\) \n... \nUser: \\(< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>\\) \nOutput: \nCharacter: \\(< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.556 + ], + "angle": 0, + "content": "where \\( \\langle \\text{Character\\_Placeholder} \\rangle \\) is placeholder for character's text description, \\( \\langle \\text{M\\_Placeholder} \\rangle \\) for motion token sequences, and \\( \\langle \\text{S\\_Placeholder} \\rangle \\) for speech token sequences." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.566, + 0.618, + 0.582 + ], + "angle": 0, + "content": "3.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.589, + 0.874, + 0.604 + ], + "angle": 0, + "content": "The training of SOLAMI adopts a three-stage strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.604, + 0.906, + 0.647 + ], + "angle": 0, + "content": "Stage 1: Tokenizer Training. The training approach for the motion tokenizer uses the fashion of [28]. The train loss is" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.651, + 0.905, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {e} \\mathcal {L} _ {e} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {v} \\mathcal {L} _ {v}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.906, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_r\\) means reconstruction loss, \\(\\mathcal{L}_e\\) embedding loss, \\(\\mathcal{L}_c\\) commitment loss, \\(\\mathcal{L}_v\\) velocity loss, and \\(\\lambda_*\\) are manually adjusted weights. For the speech tokenizer, we use the pre-trained checkpoint from AnyGPT [60]. We freeze the tokenizers' weights after this stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Stage 2: Multi-task Pre-training for Modality Alignment. As shown in Fig. 2, the second stage is multi-task pre-training, achieving modality alignment between motion and text, as well as between speech and text. It is necessary because motion data is scarce, and direct training on multimodal interaction data results in suboptimal models, as demonstrated in subsequent ablation studies. For motion and text alignment, we use \\(46\\mathrm{K}\\) motion-text pairs for text-to-motion generation and motion captioning tasks, and \\(11\\mathrm{K}\\) interactive motion pairs" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26889" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.895, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.36, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Training pipeline of SOLAMI. We train SOLAMI through a three-stage process. In the pre-training stage, we train the model with motion-text and speech-text related tasks to align the speech and motion modalities with language. During the instruction tuning stage, we train the model with social multimodal multi-round interaction data, enabling it to generate multimodal responses that align with the character settings and the context of the topic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.56 + ], + "angle": 0, + "content": "for two-person motion generation. To align the speech and text, we train the model with \\(410\\mathrm{K}\\) speech-text pairs for text-to-speech and automatic speech recognition tasks, and \\(100\\mathrm{K}\\) speech dialogue pairs for speech-to-speech generation. The tasks are formulated as \"User: ; Character: \". To balance the scale disparity between the motion and speech data, we sampled them at a 4:6 ratio during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.757 + ], + "angle": 0, + "content": "Stage 3: Instruction Tuning for Multi-turn Conversat- tion. In the third stage, we perform instruction tuning by applying supervised fine-tuning with multimodal interaction data, enabling the model to handle long-sequence, multi-turn dialogues, as shown in Fig. 2. We utilize \\(5.7\\mathrm{K}\\) multimodal conversation items for supervised fine-tuning. Each conversation item is fed to the model in the format as the Interaction Template in Sec. 3.1. We experiment with two approaches: full-parameter fine-tuning and LoRA finetuning [23]. We supervised only the character's response to teach the model how to react to the user's behavior. Thus we train the model by maximizing the log-likelihood of the next-token prediction and the train loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.769, + 0.484, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {I T}} = - \\sum_ {r = 1} ^ {R} \\sum_ {i = 1} ^ {L _ {M} ^ {r}} \\log p _ {\\Theta} (\\hat {m} _ {i} ^ {r} | \\hat {m} _ {i - 1} ^ {r}, \\dots , \\hat {m} _ {1} ^ {r}, \\hat {S} _ {< r}, \\hat {M} _ {< r}) \\\\ - \\sum_ {r = 1} ^ {R} \\sum_ {i = 1} ^ {L _ {S} ^ {r}} \\log p _ {\\Theta} \\left(\\hat {s} _ {i} ^ {r} \\mid \\hat {s} _ {i - 1} ^ {r}, \\dots , \\hat {s} _ {1} ^ {r}, \\hat {S} _ {< r}, \\hat {M} _ {\\leq r}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Theta\\) is the network parameter of the LLM backbone or LoRA weights, \\(R\\) is the conversation round, \\(\\hat{S}_r\\) and \\(\\hat{M}_r\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.456 + ], + "angle": 0, + "content": "are the \\(r\\)-th round speech and motion token sequences with lengths \\(L_{M}^{r}\\) and \\(L_{S}^{r}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.676, + 0.485 + ], + "angle": 0, + "content": "4. SynMSI Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.614 + ], + "angle": 0, + "content": "Social interaction between users and virtual characters is inherently unique, which makes collecting such multimodal interaction data particularly challenging. Currently, available public datasets [21, 40, 56] are incomplete for our task. To address this issue, we propose a data synthesis pipeline that leverages existing motion-text datasets, text-based role-play models, and speech synthesis methods and generates a large-scale multimodal dialogue dataset, SynMSI." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.649, + 0.637 + ], + "angle": 0, + "content": "4.1. Motion Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We collect motion-text data for two purposes: first, to achieve alignment between motion and text during pretraining, and second, to generate multimodal data for instruction tuning. Since our work focuses on modeling social interactions, we select existing datasets that contain rich social actions: HumanML3D [21] with \\(24\\mathrm{K}\\) motion-text pairs, Inter-X [56] with \\(20\\mathrm{K}\\) motion-text pairs and 10 K two-person motion pairs, and DLP-MoCap [13] with 2 K motion-text pairs. Since the Inter-X [56] dataset contains only text descriptions of two-person interactive motion without descriptions of individual motion, we used GPT-4o [41] to decompose the two-person action descriptions into single-person motion-text pairs. Additionally, we used GPT-4o [41] to synthesize comprehensive text descriptions for each motion clip by consolidating multiple possible descriptions, thereby providing more detailed textual annotations that preserve motion details." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26890" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.57, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.096, + 0.895, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.365, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Figure 3. SynMSI dataset generation. Our synthesizing pipeline consists of 4 steps. Based on numerous character-relevant topics and state-of-the-art LLMs [41], we generate text scripts for multimodal dialogues. Using a large-scale motion database [13, 21, 56], we retrieve the most appropriate motions and refine the speech scripts accordingly. Finally, we employ TTS/voice cloning [14] to generate character-specific speech. This approach enables us to create multimodal interaction data of various characters using only existing motion datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.226, + 0.451 + ], + "angle": 0, + "content": "4.2. Speech Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.562 + ], + "angle": 0, + "content": "We use speech-text data for speech-text alignment in the pre-training stage. Speech datasets involve CommonVoice [8] (150 K speech-text pairs in our experiments), AnyInstruct [60] (200 K speech-text pairs and 100 K speech-to-speech items), and synthetic speech data (60 K speech-text pairs) by text-to-speech approaches (Azure TTS and XTTS_v2 [14])." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.363, + 0.592 + ], + "angle": 0, + "content": "4.3. Multimodal Data Synthesizing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Previous data synthesis methods [32, 37, 66] usually use the general abilities of advanced LLMs [41] and text-annotated multimodal data to generate synthetic data. However, generating social multimodal interaction data has not yet been achieved. This challenging task requires high-quality expression of body language, voice consistency that matches the characters, and suitable dialogue content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, our synthesizing pipeline includes 4 steps. (1) First, we collect \\(5.3\\mathrm{K}\\) character-related and daily topics from internet platforms (Google Trends [2], Zhihu [5], Jike) and brainstorms of GPT-4o [41] to improve the diversity of the dialogue contents. (2) Based on the topic, character setting, and previous round of scripts, we use GPT-4o [41] to generate textual descriptions (motion, speech, expression etc.) for the next round of the dialogue. (3) Then we utilize the text embedding [39] of the motion description to retrieve the most relevant motions from our meticulously curated motion-text database. (4) Moreover, we refine the generated speech text with the retrieved motions to ensure that speech and motion are well-coordinated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.585 + ], + "angle": 0, + "content": "The motion database with detailed text annotations and the refinement process can alleviate the misalignment between the real motion and speech in the LLM-Agent method [13]. By iteratively repeating this process, we can generate multi-round dialogue contents across many characters, where the motions are sourced from the motion database, and the speeches are synthesized using TTS/voice cloning (Azure TTS and XTTS_v2 [14]) to maintain consistency with the character's voice style. We finally obtained \\(6.3\\mathrm{K}\\) multimodal dialogue items." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.651, + 0.619 + ], + "angle": 0, + "content": "5. VR Interface" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To demonstrate our method directly, we developed a VR interface with an Oculus Quest 3 frontend and a backend service, as shown in Fig. 4. The frontend enables immersive interaction between users and 3D autonomous characters, while the backend, powered by 2 H800 GPUs, supports the computation of various baselines in our experiments. During usage, the VR headset captures the user's speech and body motion, and sends them to the backend computation nodes. For motion capture, we use the Quest's full-body tracking system [55] to obtain pose parameters, which are then retargeted onto an SMPL-X model [43]. The computation nodes generate the body motion parameters and speech responses of the character based on the multimodal input. Then we apply an audio-to-face method, UniTalker [17], to generate the character's facial animation parameters. The facial and body parameters are jointly retargeted onto a 3D character model [58], completing one cycle of social interaction. For a natural user experience, we employ preset idle" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "26891" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.334 + ], + "angle": 0, + "content": "Figure 4. VR interface architecture. Our VR project consists of a Quest 3 client and a server. The Quest client captures and transmits user body motion and speech to the server. The server then generates character's speech, body motion, and face blendshape parameters based on the selected methods. The response is then sent back to the Quest client to drive the character." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.483, + 0.379 + ], + "angle": 0, + "content": "motions on the character when the method is LLM+Speech or the character is waiting for the user's input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.531 + ], + "angle": 0, + "content": "3D Character Assets. Our 3D character portfolio covers a diverse range of entities, including AI assistant avatars, famous cinematic roles, internet memes, and real-world celebrity personas. These 3D models are sourced from open-source repositories under CC Attribution-NonCommercial-ShareAlike License as well as our manual creation using VRoid Studio [4]. We subsequently employ facial rigging, skinning, bone chain simulation, retargeting, and texture and material creation in Unity Engine processes to yield functional 3D character assets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.549, + 0.224, + 0.566 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.299, + 0.592 + ], + "angle": 0, + "content": "6.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.794 + ], + "angle": 0, + "content": "In our experiment, we selected the AnyGPT-base model [60] (based on LLaMA2-7B [51]) as the backbone for SOLAMI, because it is an open-source model available at our experimental time that supports end-to-end speech processing. During the pre-training stage, we utilize 32 V100 GPUs to train the model for \\(3\\mathrm{K}\\) steps (batch size 256, learning rate 4e-5). For instruction tuning, we train the SOLAMI for 800 steps using 16 V100 GPUs (batch size 48, learning rate 2e-5). For LoRA fine-tuning [23], we set the rank as 8 and alpha as 16. We split the synthesized multimodal data into training and test sets with a 9:1 ratio. We use DeepSpeed [45] to accelerate the training. During testing, we evaluate each round of the character's response." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baselines. To validate the performance improvement in social interaction brought by incorporating 3D modalities (such as body motion), we compared SOLAMI with the LLM+Speech and the AnyGPT (fine-tune) approach. For the LLM+Speech framework, the user's speech is first transcribed into text using ASR techniques, which is then fed to a LLM to generate the character's response in text, and sub" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.471 + ], + "angle": 0, + "content": "sequentially converted into speech using TTS. For fairness, we use LLaMA2-7B-Chat [51] as the LLM backbone, Whisper large-v3 [44] for ASR, and XTTS_v2 [14] for voice cloning. For the AnyGPT (fine-tune) framework, we use the speech data of SynMSI to train the AnyGPT-base model [60] with the same parameter settings. To compare the effectiveness of the LLM-Agent architecture with the social VLA framework, we used DLP [13] as a baseline method. In DLP framework, the user's speech and body motion are separately processed as text descriptions by ASR and motion captioning modules. Based on the input text descriptions, LLM generates the character's text instructions of speech and motion, which are transferred into speech and body motion by TTS and motion generation module. The speech component of DLP follows the LLM+Speech method. Considering that MoMat-MoGen module in DLP is too slow for user interaction (over 5 seconds latency), we use MotionGPT [28] for motion captioning and motion generation. To ensure a fair comparison, we used the same motion data as the pre-training stage of SOLAMI to train MotionGPT. Additionally, we conducted ablation experiments on the effect of the pre-training stage, marked as (w/o pretrain). For the ablation study of the motion tokenizer, please refer to the supplementary materials. We use vLLM [31] to accelerate the LLM backbones for low latency interaction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.736, + 0.499 + ], + "angle": 0, + "content": "6.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.534 + ], + "angle": 0, + "content": "We conducted quantitative evaluation for our method and all the baselines mentioned in Sec. 6.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Evaluation Metrics. For motion, we evaluate the model responses using metrics including FID, diversity, PAMPJPE (mm), and angle error [22]. Following Duolando [48], we obtain FID and diversity using motion features from AIST++ [33]. For speech, we use VC similarity [57] to evaluate the voice similarity with the character. To evaluate the content quality of speech, we first use Whisper-large-v3 [44] to transcribe the speech into text. Then following [47, 52], we employ GPT-4o [41] as the judge to assess Context Relevance and Character Consistency on a Likert scale ranging from 1 to 5. Context Relevance indicates whether the speech content aligns with the topic and context of the conversation, while Character Consistency assesses whether the content adheres to the character settings. For inference latency (seconds), we deploy all the models on 2 H800 GPUs with vLLM [31] framework and asynchronous mechanisms to improve performance while maintaining fairness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Quantitative Results. The quantitative results in Tab. 1 demonstrate that, using the same foundation model (Llama2) [51] as the backbone, our method, SOLAMI (full params), significantly outperforms other methods in terms of motion quality and inference latency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Motion Quality. SOLAMI demonstrates superior perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26892" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative results of baselines and SOLAMI. \\(\\uparrow^{\\prime}(\\downarrow^{\\prime})\\) indicates that the values are better if the metrics are larger (smaller). We run all the evaluations 5 times and report the average metric. The best results are in bold and the second best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.128, + 0.907, + 0.232 + ], + "angle": 0, + "content": "
MethodsMotion MetricsSpeech MetricsInference Latency↓
FID↓Diversity↑PA-MPJPE↓Angle Error↓VC Similarity↑Context Relevance↑Character Consistency↑
SynMSI Dataset-9.136---4.8884.893-
LLM+Speech (Llama2) [51]----0.8183.5273.8593.157
AnyGPT (fine-tune) [60]----0.8193.5023.8032.588
DLP (MotionGPT) [13]4.2548.259165.0530.4950.8123.5773.7855.518
SOLAMI (w/o pretrain)5.0528.558159.7090.3870.8203.5413.4612.657
SOLAMI (LoRA)15.7298.145167.1490.4000.7703.2513.4232.710
SOLAMI (full params)3.4438.853151.5000.3600.8243.6343.8242.639
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.243, + 0.482, + 0.41 + ], + "angle": 0, + "content": "mance compared with the DLP method [13] across multiple motion metrics. This is because the social VLA model achieves comprehensive modality alignment among speech, motion, and language through training on our high-fidelity character-specific multimodal data. Our model can precisely perceive the user's physical motions and linguistic cues, enabling semantically rich interactive motions in response. This ability contrasts with the LLM-Agent architectures of the DLP method, which exhibits limitations in conveying multimodal nuances through text-only intermediary representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.624 + ], + "angle": 0, + "content": "Speech Quality. Our method demonstrates the capability to synthesize a voice tone that matches the character with a higher Voice Cloning (VC) Similarity score. Our model also shows better performance on the context relevance of the dialogue than other methods. Because the inclusion of the motion modality enables the model to perceive the user's body language, while the LLM+Speech or AnyGPT(fine-tune) method lacks this capability. In terms of character consistency, our model achieves secondary performance metrics. We suspect this may be due to the incorporation of motion and speech modalities, which potentially affects the character-related knowledge embedded within the original LLM. The performance degradation is similar to the observations in [16, 36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Inference Latency. Our end-to-end approach is significantly superior to the modular pipeline approaches (LLM+Speech or DLP). Because the end-to-end VLA model naturally aligns with the process of real-time human communication. Theoretically, if we could collect data on real-time interactions between humans and characters, our method could achieve full-duplex streaming interaction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation Study. As shown in Tab. 1, the pre-training stage of SOLAMI leads to better performance in both motion and speech quality. We believe that the pre-training stage, which aligns motion, speech, and language, facilitates the model's ability to learn the multimodal dialogue skill during the instruction tuning stage. Instruction tuning using LoRA [23] shows weaker results compared to full parameter fine-tuning. We think that the gap between the data distribution of pre-training tasks and the multimodal instruction tuning task is substantial, and LoRA fine-tuning alone is insufficient for the model to learn the strong multimodal" + }, + { + "type": "table_caption", + "bbox": [ + 0.564, + 0.242, + 0.855, + 0.256 + ], + "angle": 0, + "content": "Table 2. Questionnaire settings of our user study." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.266, + 0.907, + 0.361 + ], + "angle": 0, + "content": "
DimensionQuestions
Motion CoherenceDoes the motion match the character's setting?Does the motion align well with speech?
Motion InteractionDoes the character follow motion instructions correctly?Does the character understand user's motion?
Speech ConsistencyDoes the speech match the character's setting?Is the speech relevant to the current topic?
Overall ExperienceHow would you rate the overall experience?
" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.376, + 0.903, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.541, + 0.878, + 0.555 + ], + "angle": 0, + "content": "Figure 5. Results of the user study with \\(95\\%\\) confidence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.609, + 0.581 + ], + "angle": 0, + "content": "dialog ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.741, + 0.608 + ], + "angle": 0, + "content": "6.3. VR Interface User Study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Quantitative evaluation of a single modality alone cannot fully compare the 3D autonomous characters built on different frameworks. To address this, we conducted user study with a VR interface. As shown in Fig. 4, we employ LLM+Speech, DLP (MotionGPT), and SOLAMI as method options of the server backend and the same VR frontend across different methods. This implies that the variable in the experiment is the driving method in the server. Users are asked to engage in more than five rounds of dialogue with the VR character before completing our questionnaire." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. The indicators and corresponding questions for our questionnaire are shown in Tab. 2. Motion Coherence evaluates whether the character's motion aligns with the character setting and dialogue content. Motion Interaction assesses whether the character can understand the semantics of the body motion and effectively interact with the user. Speech Coherence examines whether the generated speech aligns with the context and character settings." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26893" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.096, + 0.907, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.477, + 0.908, + 0.52 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of SOLAMI and baselines, and the user workflow for VR experience. Our social VLA model, trained in an end-to-end strategy on SynMSI dataset, can accurately perceive the semantic information embedded within users' speech and motion input, and subsequently generate natural and coherent responses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.634 + ], + "angle": 0, + "content": "And the Overall Experience measures the user's satisfaction with the overall experience. Each question is rated on a 1 to 5 Likert scale, with higher scores indicating greater satisfaction. The score for each dimension is calculated as the average score of its corresponding questions. We ultimately collected 60 survey responses, with participants from various gender and age groups." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Results. As shown in the Fig. 5, our method achieved the highest user satisfaction across all dimensions. SOLAMI demonstrates superior performance over the DLP method across all metrics, validating the effectiveness of an end-to-end social VLA model in character behavior modeling. While the DLP method shows lower speech consistency compared to the LLM+Speech method, it excels in motion-related metrics and achieves higher overall satisfaction, indicating that effective body language can enhance user experience despite speech quality limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To provide a more intuitive demonstration of our model's capabilities, we replay the user study experiments and render them, as shown in the Fig. 6. The results indicate that SOLAMI demonstrates excellent capabilities in body language understanding, motion command execution, and body interaction. For better understanding, we also present the workflow from a first-person view during actual usage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.588 + ], + "angle": 0, + "content": "We encourage readers to explore our supplementary materials and video for a more detailed overview of the model experiments, data generation pipeline, VR interface construction process, and comprehensive experimental results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.599, + 0.634, + 0.615 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.776 + ], + "angle": 0, + "content": "In this paper, we propose SOLAMI, an approach for building 3D autonomous characters. This approach includes three key components: 1) Architecture: A novel social VLA modeling framework enabling multimodal social interaction; 2) Multimodal Data Synthesizing: A pipeline for automatically generating multimodal interaction data from existing incomplete datasets; 3) VR Interface: A VR engineering framework that facilitates immersive interactions between users and various characters. Together, these modules contribute to an enhanced user interaction experience." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.782, + 0.673, + 0.799 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.906, + 0.898 + ], + "angle": 0, + "content": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26894" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.4, + 0.128 + ], + "angle": 0, + "content": "[1] Character.ai. https://character.ai.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.132, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[2] Google trends. https://trends.google.com/trends.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.161, + 0.425, + 0.174 + ], + "angle": 0, + "content": "[3] Talkie ai. https://www-talkie-ai.com. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.178, + 0.465, + 0.19 + ], + "angle": 0, + "content": "[4] Vvoid studio. https://vvoid.com/en/studio.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.193, + 0.373, + 0.205 + ], + "angle": 0, + "content": "[5] Zhihu. https://www.zhihu.com. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.209, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[6] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[7] Tenglong Ao. Body of her: A preliminary study on end-to-end humanoid agent. arXiv preprint arXiv:2408.02879, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.296, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[8] Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In LREC, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.366, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[9] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. TASLP, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.437, + 0.483, + 0.491 + ], + "angle": 0, + "content": "[10] Zalán Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi. Soundstorm: Efficient parallel audio generation. arXiv preprint arXiv:2305.09636, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.494, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[11] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In RSS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[12] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In NeurlPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[13] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, Liang Pan, Xiangyu Fan, Han Du, Peng Gao, Zhitao Yang, Yang Gao, Jiaqi Li, Tianxiang Ren, Yukun Wei, Xiaogang Wang, Chen Change Loy, Lei Yang, and Zwei Liu. Digital life project: Autonomous 3d characters with social intelligence. In CVPR, 2024. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Edresson Casanova, Kelly Davis, Eren Gölgé, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Meyer, Reuben Morais, Samuel Olayemi, et al. Xtts: a massively multilingual zero-shot text-to-speech model. arXiv preprint arXiv:2406.04904, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[15] James J Cummings and Jeremy N Bailenson. How immersive is enough? A meta-analysis of the effect of immersive technology on user presence. Media psychology, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[16] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied multimodal language model. In ICML, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.29, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[17] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.332, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[18] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Chatpose: Chatting about 3d human pose. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.375, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenjiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[20] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, Fu-Jen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abrham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26895" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.355 + ], + "angle": 0, + "content": "ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, Anush Kumar, Robert Kuo, Sach Lakhavani, Miao Liu, Mi Luo, Zhengyi Luo, Brighid Meredith, Austin Miller, Oluwatuminu Oguntola, Xiaqing Pan, Penny Peng, Shraman Pramanick, Merey Ramazanova, Fiona Ryan, Wei Shan, Kiran Somasundaram, Chenan Song, Audrey Southerland, Masatoshi Tateno, Huiyu Wang, Yuchen Wang, Takuma Yagi, Mingfei Yan, Xitong Yang, Zecheng Yu, Shengxin Cindy Zha, Chen Zhao, Ziwei Zhao, Zhifan Zhu, Jeff Zhuo, Pablo Arbeláez, Gedas Bertasius, Dima Damen, Jakob Engel, Giovanni Maria Farinella, Antonino Furnari, Bernard Ghanem, Judy Hoffman, C. V. Jawahar, Richard A. Newcombe, Hyun Soo Park, James M. Rehg, Yoichi Sato, Manolis Savva, Jianbo Shi, Mike Zheng Shout, and Michael Wray. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.357, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.399, + 0.482, + 0.454 + ], + "angle": 0, + "content": "[22] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.455, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[23] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.511, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[24] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[25] Sarah Hudson, Sheila Matson-Barkat, Nico Pallamin, and Guillaume Jegou. With or without you? Interaction and immersion in a virtual reality experience. Journal of business research, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.623, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[26] Brian Ichter, Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, Dmitry Kalashnikov, Sergey Levine, Yao Lu, Carolina Parada, Kanishka Rao, Pierre Sermanet, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Mengyuan Yan, Noah Brown, Michael Ahn, Omar Cortes, Nicolas Sievers, Clayton Tan, Sichun Xu, Diego Reyes, Jarek Rettinghouse, Jornell Quiambao, Peter Pastor, Linda Liuu, Kuang-Huei Lee, Yuheng Kuang, Sally Jesmonth, Nikhil J. Joshi, Kyle Jeffrey, Rosario Jauregui Ruano, Jasmine Hsu, Keerthana Gopalakrishnan, Byron David, Andy Zeng, and Chuyuan Kelly Fu. Do as I can, not as I say: Grounding language in robotic affordances. In CoRL, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[27] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[28] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Paul Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuok Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.414 + ], + "angle": 0, + "content": "[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.905, + 0.47 + ], + "angle": 0, + "content": "[32] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.904, + 0.514 + ], + "angle": 0, + "content": "[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[34] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. IJCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[35] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. VILA: on pre-training for visual language models. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.66, + 0.904, + 0.687 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[38] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[39] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[40] Evonne Ng, Javier Romero, Timur M. Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In CVPR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[41] OpenAI. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26896" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[42] Joon Sung Park, Joseph C. O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[45] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[46] Giuseppe Riva, Fabrizia Mantovani, Claret S. Capdeville, Alessandra Preziosa, Francesca Morganti, Daniela Villani, Andrea Gaggioli, Cristina Botella, and Mariano Alcaniz Raya. Affective interactions using virtual reality: The link between presence and emotions. Cyberpsychology Behav. Soc. Netw., 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.408, + 0.482, + 0.449 + ], + "angle": 0, + "content": "[47] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. Character-llm: A trainable agent for role-playing. In EMNLP, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.452, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[48] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower GPT with off-policy reinforcement learning for dance accompaniment. In ICLR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.509, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[49] Mel Slater, Daniel Pérez Marcos, Henrik Ehrsson, and Maria V Sanchez-Vives. Inducing illusory ownership of a virtual body. Frontiers in neuroscience, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[50] Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive architectures for language agents. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.597, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[52] Quan Tu, Shilong Fan, Zihang Tian, Tianhao Shen, Shuo Shang, Xin Gao, and Rui Yan. Charactereval: A chinese benchmark for role-playing conversational agent evaluation. In ACL, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[53] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. TMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[54] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. TMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[55] Alexander W. Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[56] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, Yunhui Liu, Wenjun Zeng, and Xiaokang Yang. Inter-x: Towards versatile human-human interaction analysis. In CVPR, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[57] Dongchao Yang, Jinchuan Tian, Xu Tan, Rongjie Huang, Songxiang Liu, Xuankai Chang, Jiatong Shi, Sheng Zhao, Jiang Bian, Xixin Wu, Zhou Zhao, and Helen Meng. Uniaudio: An audio foundation model toward universal audio generation. arXiv preprint arXiv:2310.00704, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[58] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[59] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. TASLP, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[60] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, Hang Yan, Jie Fu, Tao Gui, Tianxiang Sun, YuGang Jiang, and Xipeng Qiu. Anygpt: Unified multimodal LLM with discrete sequence modeling. In ACL, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[61] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. In EMNLP, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[62] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[63] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech large language models. arXiv preprint arXiv:2308.16692, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[64] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26897" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[65] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[66] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[67] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26898" + } + ] +] \ No newline at end of file diff --git a/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_origin.pdf b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca330f18a8fa0a63203019199e665537c910c417 --- /dev/null +++ b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/268c1c21-7560-456f-8ae1-2ae010e66b03_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:751c74ecc5b50fce5849a19d47c69a7e7ef3f34bd3d3f11f280a5d5ac4040e57 +size 9977737 diff --git a/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/full.md b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eca3ab30ed3c9202e419c492fe977820f8c1ddc3 --- /dev/null +++ b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/full.md @@ -0,0 +1,291 @@ +# SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters + +Jianping Jiang\*†, Weiye Xiao\*‡, Zhengyu Lin $^{1}$ , Huaizhong Zhang $^{1}$ , Tianxiang Ren $^{1}$ , Yang Gao $^{1}$ , Zhiqian Lin $^{1}$ , Zhongang Cai\*1,2, Lei Yang\*1, Ziwei Liu\*2 + $^{1}$ SenseTime Research, $^{2}$ S-Lab, Nanyang Technological University +* Equal Contribution, † Project Lead, ‡ Engineering Lead, § Corresponding Author +https://solami-ai.github.io/ + +![](images/e6b04288844f9303802f257ea5d2e55ab414937fb1546b89de629177d5989676.jpg) +Figure 1. SOLAMI enables the user to interact with 3D autonomous characters through speech and body language in an immersive VR environment via an end-to-end social vision-language-action model, which is trained on our synthesized multimodal dataset SynMSI. + +# Abstract + +Human beings are social animals. How to equip 3D autonomous characters with similar social intelligence that can perceive, understand and interact with humans remains an open yet foundational problem. In this paper, we introduce SOLAMI, the first end-to-end Social vision-Language-Action (VLA) Modeling framework for Immersive interaction with 3D autonomous characters. Specifically, SOLAMI builds 3D autonomous characters from three aspects: 1) Social VLA Architecture: We propose a unified social VLA framework to generate multimodal response (speech and motion) based on the user's + +multimodal input to drive the character for social interaction. 2) Interactive Multimodal Data: We present SynMSI, a synthetic multimodal social interaction dataset generated by an automatic pipeline using only existing motion datasets to address the issue of data scarcity. 3) Immersive VR Interface: We develop a VR interface that enables users to immersively interact with these characters driven by various architectures. Extensive quantitative experiments and user study demonstrate that our framework leads to more precise and natural character responses (in both speech and motion) that align with user expectations with lower latency. + +# 1. Introduction + +Have you ever imagined having an immersive face-to-face conversation with a character you deeply admire? Not merely through speech dialogue, but through an interaction where you can observe its subtle facial expressions, natural body language, and even fleeting emotional changes. Psychology research [15, 25, 46, 49] shows that in social interactions, the greater the level of immersion, the better the human experience. However, current character agents [1, 3, 47] are still limited to text or voice interactions. This limitation prompts us to build 3D autonomous characters with richer modalities. + +Developing an autonomous 3D character requires effectively modeling its behavior system, which involves two major challenges: 1) The 3D character needs to accurately observe and understand the information conveyed by the user, and respond appropriately based on the context and its character setting through speech, body motion, and facial expression, etc. This goes beyond previous singular human-related tasks, such as motion generation [62], motion understanding [27], and audio-to-motion [6]. 2) Data for multimodal interactions between users and 3D characters is extremely scarce due to the prohibitive cost of the comprehensive setup. + +Previous work [13] is primarily based on the LLM-Agent framework, using text to link various sub-modules (such as motion captioning and text-to-motion). While this approach performs well in high-level tasks like planning and memory, it tends to fall short in tasks such as understanding user behaviors and providing timely body motion responses. This limitation arises because using text as the intermediary between modules conveys high-level information but often omits subtle nuances. And the sub-modules (motion captioning, speech recognition, etc.) in the complex engineering framework incur substantial latency that undermines the timeliness of a natural communication [50]. + +Interestingly, research on robotics shows the similar conclusion. The LLM-Agent framework can handle planning tasks [26], but for low-level manipulation tasks, end-to-end Vision-Language-Action (VLA) models built upon LLMs show superior performance [11, 24, 29, 67]. We argue that digital avatars are essentially robots with virtual humanoid embodiment. Therefore, building a VLA model for social interactions with users is a promising direction. + +In this paper, we implement an end-to-end social VLA model, SOLAMI. Our model, built upon a decoder-only LLM backbone, processes the inputs of user speech and motion into discrete representations, and generate the responsive speech and motion tokens, which are then decoded to the character's speech and motion. This modeling approach can effectively learn character behavior patterns across motion and speech modalities and offer low latency. + +Although there are numerous datasets related to human + +social behaviors [13, 34, 40, 56], comprehensive multimodal interaction datasets remain scarce. Thus, we introduce a data synthesis method that utilizes existing text-motion datasets to automatically construct multimodal interaction data at a low cost. Leveraging our extensively curated topics (5.3 K), uniformly processed motion database (46 K), and iterative script refinement pipeline, we develop SynMSI, a dataset containing 6.3 K multi-turn multimodal conversation items. To evaluate the effectiveness of our method, we developed a VR interface where users can immersively interact with various 3D characters. Quantitative experimental results and user study analysis show that our approach produces more precise and natural social interaction experience with lower latency. + +To summarize, we contribute 1) A new VLA architecture to model the character's behavior system for immersive social interaction; 2) A dedicated synthesizing pipeline that automatically generates large-scale multimodal interactive dataset, SynMSI; 3) An immer VR interface for users to interact with various characters through speech and motion. + +# 2. Related Work + +# 2.1. Human Motion & LLM + +Existing role-play agents primarily rely on text [47], speech [1, 3, 41], or video [7] as interactive media, but building 3D autonomous characters means modeling 3D embodied behavior, especially body language. Unlike single-purpose human motion tasks [6, 12, 22, 34, 48, 56, 62], we expect this 3D character not only to comprehend user speech and body language but also to respond according to its profile setting. Large language models (LLMs) [41, 51], with their remarkable emergent abilities [54], provide promising solutions for this direction. One approach [18, 28, 64, 65] tries to integrate LLMs and human motion in an end-to-end fashion, enabling a single model to perform multiple motion-related tasks. However, the goal of these methods is not to generate responsive motion based on the input motion according to the character setting. Another approach [13, 35] utilizes the LLM as a versatile brain center that controls various submodules (e.g., motion understanding, text-to-motion generation) with text or code instructions. However, this modular approach inherently introduces information loss and engineering-related time latency. Therefore, how to create feasible autonomous 3D characters for immersive interaction remains an open challenge. + +# 2.2. Embodied Intelligence + +Building 3D autonomous characters means creating virtual humanoid agents with embodied intelligence. In the field of embodied intelligence, researchers have found the high- + +level abilities (planning, memory, etc.) of LLM-Agents across various environments, including factories [16, 26], 2D sandbox settings [42], and 3D gaming spaces [53]. For tasks requiring low-level skills, such as manipulation, end-to-end VLA models [11, 24, 29, 67] have shown considerable potential. Despite existing efforts [19, 20, 22] in egocentric human-related tasks, the capabilities of VLA models in social interactive tasks with humans have not been fully explored. + +# 3. Social Vision-Language-Action Modeling + +# 3.1. Architecture + +Our framework is an end-to-end social VLA model that takes the user's speech and motion as input and generates the character's responsive speech and motion as output. In this process, speech and motion are added as new languages to the LLM text vocabulary. First, the user's speech and motion are converted into discrete motion tokens and speech tokens via a motion tokenizer and a speech tokenizer, respectively. A decoder-only LLM backbone then predicts the character's output motion and speech tokens based on the user's input tokens and the character setting. The generated tokens are subsequently decoded into corresponding motion and speech by their respective decoders. + +Motion Representation. To take advantage of SMPL-X's [43] compatibility with industry animation workflow, we directly model human poses as SMPL-X joint rotations instead of keypoint positions [28, 38] to facilitate animation of characters down the stream. + +Motion Tokenizer. Our motion tokenizer employs a Vector Quantized Variational Autoencoders (VQ-VAE) structure as [28, 64]. It learns discrete representations of motion, enabling the LLM to understand the text-motion connection. We design separate VQVAEs $(Q^{b}, Q^{h}, Q^{t})$ for the body motion $\mathbf{m}^{b}$ , hand motion $\mathbf{m}^{h}$ , and inter-character relative transform (rotation and translation) $\mathbf{m}_{t}$ for higher reconstruction accuracy [38]. The quantization process for motion $\mathbf{m}^{u}$ is formulated as + +$$ +\hat {m} _ {t} ^ {u} = Q ^ {u} \left(\mathbf {m} _ {t} ^ {u}\right) = \arg \min _ {z _ {i} \in \mathbb {Z} _ {u}} \| \mathbf {m} _ {t} ^ {u} - z _ {i} \| _ {2}, \tag {1} +$$ + +where $\mathbb{Z}_u$ is the codebook of motion part $u\in \{b,h,t\}$ and $\hat{m}_t^u$ is the corresponding motion tokens. The VQVAEs of body and hand, $Q^{b}$ and $Q^{h}$ , apply 1D convolutions on motion features along the temporal dimension to get $L_{M}$ sequential tokens $\hat{M}^b = [\hat{m}_1^b,\hat{m}_2^b,\dots,\hat{m}_{L_M}^b]$ and $\hat{M}^h = [\hat{m}_1^h,\hat{m}_2^h,\dots,\hat{m}_{L_M}^h]$ , and the VQVAE $Q^{t}$ uses MLPs to get the transform token $\hat{m}^t$ of a sequence. + +Speech Tokenizer. Research [9, 61] on speech discretization mainly utilizes the RVQ-VAE structure [59]. In this work, we utilize the SpeechTokenizer [63] that disentangles semantic and acoustic information within speech $S$ . This + +allows us to use the semantic tokens $\hat{S} = [\hat{s}_1^s, \hat{s}_2^s, \dots, \hat{s}_{L_S}^s]$ from the first layer as input to the LLM, reducing inference costs ( $L_S$ is the sequence length of semantic tokens). Simultaneously, a short sample of the character's voice (4 to 6 seconds) can be used as a prompt to achieve instance voice cloning when decoding through SoundStorm [10]. + +Multi-modal Multi-round Interaction. User-character interaction is formulated as a multi-round conversation fashion of common LLMs [41, 51]. When the user sends speech and motion to SOLAMI, the model auto-regressively generates speech and motion responses based on previous dialogue contents and the character setting. To facilitate training, we use special tokens to mark the start and end of each modality sequence as [28, 30]. The interaction process can be formulated as follow: + +# Interaction Template + +```txt +Input: +System Prompt: +User: $< \mathrm{M\_Placeholder} > < \mathrm{S\_Placeholder}>$ +Character: $< \mathrm{M\_Placeholder} > < \mathrm{S\_Placeholder}>$ +... +User: $< \mathrm{M\_Placeholder} > < \mathrm{S\_Placeholder}>$ +Output: +Character: $< \mathrm{M\_Placeholder} > < \mathrm{S\_Placeholder}>$ +``` + +where $\langle \text{Character\_Placeholder} \rangle$ is placeholder for character's text description, $\langle \text{M\_Placeholder} \rangle$ for motion token sequences, and $\langle \text{S\_Placeholder} \rangle$ for speech token sequences. + +# 3.2. Training + +The training of SOLAMI adopts a three-stage strategy. + +Stage 1: Tokenizer Training. The training approach for the motion tokenizer uses the fashion of [28]. The train loss is + +$$ +\mathcal {L} _ {m} = \lambda_ {r} \mathcal {L} _ {r} + \lambda_ {e} \mathcal {L} _ {e} + \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {v} \mathcal {L} _ {v}, \tag {2} +$$ + +where $\mathcal{L}_r$ means reconstruction loss, $\mathcal{L}_e$ embedding loss, $\mathcal{L}_c$ commitment loss, $\mathcal{L}_v$ velocity loss, and $\lambda_*$ are manually adjusted weights. For the speech tokenizer, we use the pre-trained checkpoint from AnyGPT [60]. We freeze the tokenizers' weights after this stage. + +Stage 2: Multi-task Pre-training for Modality Alignment. As shown in Fig. 2, the second stage is multi-task pre-training, achieving modality alignment between motion and text, as well as between speech and text. It is necessary because motion data is scarce, and direct training on multimodal interaction data results in suboptimal models, as demonstrated in subsequent ablation studies. For motion and text alignment, we use $46\mathrm{K}$ motion-text pairs for text-to-motion generation and motion captioning tasks, and $11\mathrm{K}$ interactive motion pairs + +![](images/2319dea11b5e8a6fed26e0c49dd3742d1f98014a560f6b5eddec7c09e28c05fe.jpg) +Figure 2. Training pipeline of SOLAMI. We train SOLAMI through a three-stage process. In the pre-training stage, we train the model with motion-text and speech-text related tasks to align the speech and motion modalities with language. During the instruction tuning stage, we train the model with social multimodal multi-round interaction data, enabling it to generate multimodal responses that align with the character settings and the context of the topic. + +for two-person motion generation. To align the speech and text, we train the model with $410\mathrm{K}$ speech-text pairs for text-to-speech and automatic speech recognition tasks, and $100\mathrm{K}$ speech dialogue pairs for speech-to-speech generation. The tasks are formulated as "User: ; Character: ". To balance the scale disparity between the motion and speech data, we sampled them at a 4:6 ratio during training. + +Stage 3: Instruction Tuning for Multi-turn Conversat- tion. In the third stage, we perform instruction tuning by applying supervised fine-tuning with multimodal interaction data, enabling the model to handle long-sequence, multi-turn dialogues, as shown in Fig. 2. We utilize $5.7\mathrm{K}$ multimodal conversation items for supervised fine-tuning. Each conversation item is fed to the model in the format as the Interaction Template in Sec. 3.1. We experiment with two approaches: full-parameter fine-tuning and LoRA finetuning [23]. We supervised only the character's response to teach the model how to react to the user's behavior. Thus we train the model by maximizing the log-likelihood of the next-token prediction and the train loss is: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {I T}} = - \sum_ {r = 1} ^ {R} \sum_ {i = 1} ^ {L _ {M} ^ {r}} \log p _ {\Theta} (\hat {m} _ {i} ^ {r} | \hat {m} _ {i - 1} ^ {r}, \dots , \hat {m} _ {1} ^ {r}, \hat {S} _ {< r}, \hat {M} _ {< r}) \\ - \sum_ {r = 1} ^ {R} \sum_ {i = 1} ^ {L _ {S} ^ {r}} \log p _ {\Theta} \left(\hat {s} _ {i} ^ {r} \mid \hat {s} _ {i - 1} ^ {r}, \dots , \hat {s} _ {1} ^ {r}, \hat {S} _ {< r}, \hat {M} _ {\leq r}\right), \tag {3} \\ \end{array} +$$ + +where $\Theta$ is the network parameter of the LLM backbone or LoRA weights, $R$ is the conversation round, $\hat{S}_r$ and $\hat{M}_r$ + +are the $r$ -th round speech and motion token sequences with lengths $L_{M}^{r}$ and $L_{S}^{r}$ , respectively. + +# 4. SynMSI Dataset + +Social interaction between users and virtual characters is inherently unique, which makes collecting such multimodal interaction data particularly challenging. Currently, available public datasets [21, 40, 56] are incomplete for our task. To address this issue, we propose a data synthesis pipeline that leverages existing motion-text datasets, text-based role-play models, and speech synthesis methods and generates a large-scale multimodal dialogue dataset, SynMSI. + +# 4.1. Motion Data + +We collect motion-text data for two purposes: first, to achieve alignment between motion and text during pretraining, and second, to generate multimodal data for instruction tuning. Since our work focuses on modeling social interactions, we select existing datasets that contain rich social actions: HumanML3D [21] with $24\mathrm{K}$ motion-text pairs, Inter-X [56] with $20\mathrm{K}$ motion-text pairs and 10 K two-person motion pairs, and DLP-MoCap [13] with 2 K motion-text pairs. Since the Inter-X [56] dataset contains only text descriptions of two-person interactive motion without descriptions of individual motion, we used GPT-4o [41] to decompose the two-person action descriptions into single-person motion-text pairs. Additionally, we used GPT-4o [41] to synthesize comprehensive text descriptions for each motion clip by consolidating multiple possible descriptions, thereby providing more detailed textual annotations that preserve motion details. + +![](images/8b8c986f2ba86c131ae2e8df79e3ec0882a12d91b99c1ab5ed2be0f5b8499f9d.jpg) +Figure 3. SynMSI dataset generation. Our synthesizing pipeline consists of 4 steps. Based on numerous character-relevant topics and state-of-the-art LLMs [41], we generate text scripts for multimodal dialogues. Using a large-scale motion database [13, 21, 56], we retrieve the most appropriate motions and refine the speech scripts accordingly. Finally, we employ TTS/voice cloning [14] to generate character-specific speech. This approach enables us to create multimodal interaction data of various characters using only existing motion datasets. + +![](images/90fd8d13848f8105c4f4d7d1122c58b7728eeaf2fd7ea279e6049fadaf97e53e.jpg) + +# 4.2. Speech Data + +We use speech-text data for speech-text alignment in the pre-training stage. Speech datasets involve CommonVoice [8] (150 K speech-text pairs in our experiments), AnyInstruct [60] (200 K speech-text pairs and 100 K speech-to-speech items), and synthetic speech data (60 K speech-text pairs) by text-to-speech approaches (Azure TTS and XTTS_v2 [14]). + +# 4.3. Multimodal Data Synthesizing + +Previous data synthesis methods [32, 37, 66] usually use the general abilities of advanced LLMs [41] and text-annotated multimodal data to generate synthetic data. However, generating social multimodal interaction data has not yet been achieved. This challenging task requires high-quality expression of body language, voice consistency that matches the characters, and suitable dialogue content. + +As shown in Fig. 3, our synthesizing pipeline includes 4 steps. (1) First, we collect $5.3\mathrm{K}$ character-related and daily topics from internet platforms (Google Trends [2], Zhihu [5], Jike) and brainstorms of GPT-4o [41] to improve the diversity of the dialogue contents. (2) Based on the topic, character setting, and previous round of scripts, we use GPT-4o [41] to generate textual descriptions (motion, speech, expression etc.) for the next round of the dialogue. (3) Then we utilize the text embedding [39] of the motion description to retrieve the most relevant motions from our meticulously curated motion-text database. (4) Moreover, we refine the generated speech text with the retrieved motions to ensure that speech and motion are well-coordinated. + +The motion database with detailed text annotations and the refinement process can alleviate the misalignment between the real motion and speech in the LLM-Agent method [13]. By iteratively repeating this process, we can generate multi-round dialogue contents across many characters, where the motions are sourced from the motion database, and the speeches are synthesized using TTS/voice cloning (Azure TTS and XTTS_v2 [14]) to maintain consistency with the character's voice style. We finally obtained $6.3\mathrm{K}$ multimodal dialogue items. + +# 5. VR Interface + +To demonstrate our method directly, we developed a VR interface with an Oculus Quest 3 frontend and a backend service, as shown in Fig. 4. The frontend enables immersive interaction between users and 3D autonomous characters, while the backend, powered by 2 H800 GPUs, supports the computation of various baselines in our experiments. During usage, the VR headset captures the user's speech and body motion, and sends them to the backend computation nodes. For motion capture, we use the Quest's full-body tracking system [55] to obtain pose parameters, which are then retargeted onto an SMPL-X model [43]. The computation nodes generate the body motion parameters and speech responses of the character based on the multimodal input. Then we apply an audio-to-face method, UniTalker [17], to generate the character's facial animation parameters. The facial and body parameters are jointly retargeted onto a 3D character model [58], completing one cycle of social interaction. For a natural user experience, we employ preset idle + +![](images/516a93c459e60f9f5c0fda17fe26afbc63d8ab83fbfa8702af7e6f6b6f7b38f5.jpg) +Figure 4. VR interface architecture. Our VR project consists of a Quest 3 client and a server. The Quest client captures and transmits user body motion and speech to the server. The server then generates character's speech, body motion, and face blendshape parameters based on the selected methods. The response is then sent back to the Quest client to drive the character. + +motions on the character when the method is LLM+Speech or the character is waiting for the user's input. + +3D Character Assets. Our 3D character portfolio covers a diverse range of entities, including AI assistant avatars, famous cinematic roles, internet memes, and real-world celebrity personas. These 3D models are sourced from open-source repositories under CC Attribution-NonCommercial-ShareAlike License as well as our manual creation using VRoid Studio [4]. We subsequently employ facial rigging, skinning, bone chain simulation, retargeting, and texture and material creation in Unity Engine processes to yield functional 3D character assets. + +# 6. Experiments + +# 6.1. Experimental Settings + +In our experiment, we selected the AnyGPT-base model [60] (based on LLaMA2-7B [51]) as the backbone for SOLAMI, because it is an open-source model available at our experimental time that supports end-to-end speech processing. During the pre-training stage, we utilize 32 V100 GPUs to train the model for $3\mathrm{K}$ steps (batch size 256, learning rate 4e-5). For instruction tuning, we train the SOLAMI for 800 steps using 16 V100 GPUs (batch size 48, learning rate 2e-5). For LoRA fine-tuning [23], we set the rank as 8 and alpha as 16. We split the synthesized multimodal data into training and test sets with a 9:1 ratio. We use DeepSpeed [45] to accelerate the training. During testing, we evaluate each round of the character's response. + +Baselines. To validate the performance improvement in social interaction brought by incorporating 3D modalities (such as body motion), we compared SOLAMI with the LLM+Speech and the AnyGPT (fine-tune) approach. For the LLM+Speech framework, the user's speech is first transcribed into text using ASR techniques, which is then fed to a LLM to generate the character's response in text, and sub + +sequentially converted into speech using TTS. For fairness, we use LLaMA2-7B-Chat [51] as the LLM backbone, Whisper large-v3 [44] for ASR, and XTTS_v2 [14] for voice cloning. For the AnyGPT (fine-tune) framework, we use the speech data of SynMSI to train the AnyGPT-base model [60] with the same parameter settings. To compare the effectiveness of the LLM-Agent architecture with the social VLA framework, we used DLP [13] as a baseline method. In DLP framework, the user's speech and body motion are separately processed as text descriptions by ASR and motion captioning modules. Based on the input text descriptions, LLM generates the character's text instructions of speech and motion, which are transferred into speech and body motion by TTS and motion generation module. The speech component of DLP follows the LLM+Speech method. Considering that MoMat-MoGen module in DLP is too slow for user interaction (over 5 seconds latency), we use MotionGPT [28] for motion captioning and motion generation. To ensure a fair comparison, we used the same motion data as the pre-training stage of SOLAMI to train MotionGPT. Additionally, we conducted ablation experiments on the effect of the pre-training stage, marked as (w/o pretrain). For the ablation study of the motion tokenizer, please refer to the supplementary materials. We use vLLM [31] to accelerate the LLM backbones for low latency interaction. + +# 6.2.Quantitative Evaluation + +We conducted quantitative evaluation for our method and all the baselines mentioned in Sec. 6.1. + +Evaluation Metrics. For motion, we evaluate the model responses using metrics including FID, diversity, PAMPJPE (mm), and angle error [22]. Following Duolando [48], we obtain FID and diversity using motion features from AIST++ [33]. For speech, we use VC similarity [57] to evaluate the voice similarity with the character. To evaluate the content quality of speech, we first use Whisper-large-v3 [44] to transcribe the speech into text. Then following [47, 52], we employ GPT-4o [41] as the judge to assess Context Relevance and Character Consistency on a Likert scale ranging from 1 to 5. Context Relevance indicates whether the speech content aligns with the topic and context of the conversation, while Character Consistency assesses whether the content adheres to the character settings. For inference latency (seconds), we deploy all the models on 2 H800 GPUs with vLLM [31] framework and asynchronous mechanisms to improve performance while maintaining fairness. + +Quantitative Results. The quantitative results in Tab. 1 demonstrate that, using the same foundation model (Llama2) [51] as the backbone, our method, SOLAMI (full params), significantly outperforms other methods in terms of motion quality and inference latency. + +Motion Quality. SOLAMI demonstrates superior perfor + +Table 1. Quantitative results of baselines and SOLAMI. $\uparrow^{\prime}(\downarrow^{\prime})$ indicates that the values are better if the metrics are larger (smaller). We run all the evaluations 5 times and report the average metric. The best results are in bold and the second best results are underlined. + +
MethodsMotion MetricsSpeech MetricsInference Latency↓
FID↓Diversity↑PA-MPJPE↓Angle Error↓VC Similarity↑Context Relevance↑Character Consistency↑
SynMSI Dataset-9.136---4.8884.893-
LLM+Speech (Llama2) [51]----0.8183.5273.8593.157
AnyGPT (fine-tune) [60]----0.8193.5023.8032.588
DLP (MotionGPT) [13]4.2548.259165.0530.4950.8123.5773.7855.518
SOLAMI (w/o pretrain)5.0528.558159.7090.3870.8203.5413.4612.657
SOLAMI (LoRA)15.7298.145167.1490.4000.7703.2513.4232.710
SOLAMI (full params)3.4438.853151.5000.3600.8243.6343.8242.639
+ +mance compared with the DLP method [13] across multiple motion metrics. This is because the social VLA model achieves comprehensive modality alignment among speech, motion, and language through training on our high-fidelity character-specific multimodal data. Our model can precisely perceive the user's physical motions and linguistic cues, enabling semantically rich interactive motions in response. This ability contrasts with the LLM-Agent architectures of the DLP method, which exhibits limitations in conveying multimodal nuances through text-only intermediary representation. + +Speech Quality. Our method demonstrates the capability to synthesize a voice tone that matches the character with a higher Voice Cloning (VC) Similarity score. Our model also shows better performance on the context relevance of the dialogue than other methods. Because the inclusion of the motion modality enables the model to perceive the user's body language, while the LLM+Speech or AnyGPT(fine-tune) method lacks this capability. In terms of character consistency, our model achieves secondary performance metrics. We suspect this may be due to the incorporation of motion and speech modalities, which potentially affects the character-related knowledge embedded within the original LLM. The performance degradation is similar to the observations in [16, 36]. + +Inference Latency. Our end-to-end approach is significantly superior to the modular pipeline approaches (LLM+Speech or DLP). Because the end-to-end VLA model naturally aligns with the process of real-time human communication. Theoretically, if we could collect data on real-time interactions between humans and characters, our method could achieve full-duplex streaming interaction. + +Ablation Study. As shown in Tab. 1, the pre-training stage of SOLAMI leads to better performance in both motion and speech quality. We believe that the pre-training stage, which aligns motion, speech, and language, facilitates the model's ability to learn the multimodal dialogue skill during the instruction tuning stage. Instruction tuning using LoRA [23] shows weaker results compared to full parameter fine-tuning. We think that the gap between the data distribution of pre-training tasks and the multimodal instruction tuning task is substantial, and LoRA fine-tuning alone is insufficient for the model to learn the strong multimodal + +Table 2. Questionnaire settings of our user study. + +
DimensionQuestions
Motion CoherenceDoes the motion match the character's setting?Does the motion align well with speech?
Motion InteractionDoes the character follow motion instructions correctly?Does the character understand user's motion?
Speech ConsistencyDoes the speech match the character's setting?Is the speech relevant to the current topic?
Overall ExperienceHow would you rate the overall experience?
+ +![](images/77940dc7c538e9e7b11b4f79aed699930c925aa63230fa19d4e9a7982fd51406.jpg) +Figure 5. Results of the user study with $95\%$ confidence. + +dialog ability. + +# 6.3. VR Interface User Study + +Quantitative evaluation of a single modality alone cannot fully compare the 3D autonomous characters built on different frameworks. To address this, we conducted user study with a VR interface. As shown in Fig. 4, we employ LLM+Speech, DLP (MotionGPT), and SOLAMI as method options of the server backend and the same VR frontend across different methods. This implies that the variable in the experiment is the driving method in the server. Users are asked to engage in more than five rounds of dialogue with the VR character before completing our questionnaire. + +Evaluation Metrics. The indicators and corresponding questions for our questionnaire are shown in Tab. 2. Motion Coherence evaluates whether the character's motion aligns with the character setting and dialogue content. Motion Interaction assesses whether the character can understand the semantics of the body motion and effectively interact with the user. Speech Coherence examines whether the generated speech aligns with the context and character settings. + +![](images/088896000ac6620228c85e6cb79d98c7e5d84c86e3527e5f156506411bb69003.jpg) +Figure 6. Qualitative results of SOLAMI and baselines, and the user workflow for VR experience. Our social VLA model, trained in an end-to-end strategy on SynMSI dataset, can accurately perceive the semantic information embedded within users' speech and motion input, and subsequently generate natural and coherent responses. + +And the Overall Experience measures the user's satisfaction with the overall experience. Each question is rated on a 1 to 5 Likert scale, with higher scores indicating greater satisfaction. The score for each dimension is calculated as the average score of its corresponding questions. We ultimately collected 60 survey responses, with participants from various gender and age groups. + +Results. As shown in the Fig. 5, our method achieved the highest user satisfaction across all dimensions. SOLAMI demonstrates superior performance over the DLP method across all metrics, validating the effectiveness of an end-to-end social VLA model in character behavior modeling. While the DLP method shows lower speech consistency compared to the LLM+Speech method, it excels in motion-related metrics and achieves higher overall satisfaction, indicating that effective body language can enhance user experience despite speech quality limitations. + +To provide a more intuitive demonstration of our model's capabilities, we replay the user study experiments and render them, as shown in the Fig. 6. The results indicate that SOLAMI demonstrates excellent capabilities in body language understanding, motion command execution, and body interaction. For better understanding, we also present the workflow from a first-person view during actual usage. + +We encourage readers to explore our supplementary materials and video for a more detailed overview of the model experiments, data generation pipeline, VR interface construction process, and comprehensive experimental results. + +# 7. Conclusion + +In this paper, we propose SOLAMI, an approach for building 3D autonomous characters. This approach includes three key components: 1) Architecture: A novel social VLA modeling framework enabling multimodal social interaction; 2) Multimodal Data Synthesizing: A pipeline for automatically generating multimodal interaction data from existing incomplete datasets; 3) VR Interface: A VR engineering framework that facilitates immersive interactions between users and various characters. Together, these modules contribute to an enhanced user interaction experience. + +# Acknowledgement + +This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Character.ai. https://character.ai.2 +[2] Google trends. https://trends.google.com/trends.5 +[3] Talkie ai. https://www-talkie-ai.com. 2 +[4] Vvoid studio. https://vvoid.com/en/studio.6 +[5] Zhihu. https://www.zhihu.com. 5 +[6] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG, 2023. 2 +[7] Tenglong Ao. Body of her: A preliminary study on end-to-end humanoid agent. arXiv preprint arXiv:2408.02879, 2024. 2 +[8] Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In LREC, 2020. 5 +[9] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. TASLP, 2023. 3 +[10] Zalán Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi. Soundstorm: Efficient parallel audio generation. arXiv preprint arXiv:2305.09636, 2023. 3 +[11] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In RSS, 2023. 2, 3 +[12] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In NeurlPS, 2023. 2 +[13] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, Liang Pan, Xiangyu Fan, Han Du, Peng Gao, Zhitao Yang, Yang Gao, Jiaqi Li, Tianxiang Ren, Yukun Wei, Xiaogang Wang, Chen Change Loy, Lei Yang, and Zwei Liu. Digital life project: Autonomous 3d characters with social intelligence. In CVPR, 2024. 2, 4, 5, 6, 7 +[14] Edresson Casanova, Kelly Davis, Eren Gölgé, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua + +Meyer, Reuben Morais, Samuel Olayemi, et al. Xtts: a massively multilingual zero-shot text-to-speech model. arXiv preprint arXiv:2406.04904, 2024. 5, 6 +[15] James J Cummings and Jeremy N Bailenson. How immersive is enough? A meta-analysis of the effect of immersive technology on user presence. Media psychology, 2016. 2 +[16] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied multimodal language model. In ICML, 2023. 3, 7 +[17] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. 2024. 5 +[18] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Chatpose: Chatting about 3d human pose. In CVPR, 2024. 2 +[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenjiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 3 +[20] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, Fu-Jen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abrham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed + +ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, Anush Kumar, Robert Kuo, Sach Lakhavani, Miao Liu, Mi Luo, Zhengyi Luo, Brighid Meredith, Austin Miller, Oluwatuminu Oguntola, Xiaqing Pan, Penny Peng, Shraman Pramanick, Merey Ramazanova, Fiona Ryan, Wei Shan, Kiran Somasundaram, Chenan Song, Audrey Southerland, Masatoshi Tateno, Huiyu Wang, Yuchen Wang, Takuma Yagi, Mingfei Yan, Xitong Yang, Zecheng Yu, Shengxin Cindy Zha, Chen Zhao, Ziwei Zhao, Zhifan Zhu, Jeff Zhuo, Pablo Arbeláez, Gedas Bertasius, Dima Damen, Jakob Engel, Giovanni Maria Farinella, Antonino Furnari, Bernard Ghanem, Judy Hoffman, C. V. Jawahar, Richard A. Newcombe, Hyun Soo Park, James M. Rehg, Yoichi Sato, Manolis Savva, Jianbo Shi, Mike Zheng Shout, and Michael Wray. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. In CVPR, 2024. 3 +[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4, 5 +[22] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2, 3, 6 +[23] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4, 6, 7 +[24] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2, 3 +[25] Sarah Hudson, Sheila Matson-Barkat, Nico Pallamin, and Guillaume Jegou. With or without you? Interaction and immersion in a virtual reality experience. Journal of business research, 2019. 2 +[26] Brian Ichter, Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, Dmitry Kalashnikov, Sergey Levine, Yao Lu, Carolina Parada, Kanishka Rao, Pierre Sermanet, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Mengyuan Yan, Noah Brown, Michael Ahn, Omar Cortes, Nicolas Sievers, Clayton Tan, Sichun Xu, Diego Reyes, Jarek Rettinghouse, Jornell Quiambao, Peter Pastor, Linda Liuu, Kuang-Huei Lee, Yuheng Kuang, Sally Jesmonth, Nikhil J. Joshi, Kyle Jeffrey, Rosario Jauregui Ruano, Jasmine Hsu, Keerthana Gopalakrishnan, Byron David, Andy Zeng, and Chuyuan Kelly Fu. Do as I can, not as I say: Grounding language in robotic affordances. In CoRL, 2022. 2, 3 +[27] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2 +[28] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 3, 6 + +[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Paul Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 3 +[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuok Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation. In ICML, 2024. 3 +[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6 +[32] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5 +[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In ICCV, 2021. 6 +[34] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. IJCV, 2024. 2 +[35] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 2 +[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. VILA: on pre-training for visual language models. In CVPR, 2024. 7 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 5 +[38] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, 2024. 3 +[39] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005, 2022. 5 +[40] Evonne Ng, Javier Romero, Timur M. Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In CVPR, 2024. 2, 4 +[41] OpenAI. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 3, 4, 5, 6 + +[42] Joon Sung Park, Joseph C. O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3 +[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3, 5 +[44] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 6 +[45] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD, 2020. 6 +[46] Giuseppe Riva, Fabrizia Mantovani, Claret S. Capdeville, Alessandra Preziosa, Francesca Morganti, Daniela Villani, Andrea Gaggioli, Cristina Botella, and Mariano Alcaniz Raya. Affective interactions using virtual reality: The link between presence and emotions. Cyberpsychology Behav. Soc. Netw., 2007. 2 +[47] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. Character-llm: A trainable agent for role-playing. In EMNLP, 2023. 2, 6 +[48] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower GPT with off-policy reinforcement learning for dance accompaniment. In ICLR, 2024. 2, 6 +[49] Mel Slater, Daniel Pérez Marcos, Henrik Ehrsson, and Maria V Sanchez-Vives. Inducing illusory ownership of a virtual body. Frontiers in neuroscience, 2009. 2 +[50] Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive architectures for language agents. 2024. 2 +[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6, 7 + +[52] Quan Tu, Shilong Fan, Zihang Tian, Tianhao Shen, Shuo Shang, Xin Gao, and Rui Yan. Charactereval: A chinese benchmark for role-playing conversational agent evaluation. In ACL, 2024. 6 +[53] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. TMLR, 2024. 3 +[54] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. TMLR, 2022. 2 +[55] Alexander W. Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia, 2022. 5 +[56] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, Yunhui Liu, Wenjun Zeng, and Xiaokang Yang. Inter-x: Towards versatile human-human interaction analysis. In CVPR, 2024. 2, 4, 5 +[57] Dongchao Yang, Jinchuan Tian, Xu Tan, Rongjie Huang, Songxiang Liu, Xuankai Chang, Jiatong Shi, Sheng Zhao, Jiang Bian, Xixin Wu, Zhou Zhao, and Helen Meng. Uniaudio: An audio foundation model toward universal audio generation. arXiv preprint arXiv:2310.00704, 2023. 6 +[58] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In ICCV, 2023. 5 +[59] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. TASLP, 2022. 3 +[60] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, Hang Yan, Jie Fu, Tao Gui, Tianxiang Sun, YuGang Jiang, and Xipeng Qiu. Anygpt: Unified multimodal LLM with discrete sequence modeling. In ACL, 2024. 3, 5, 6, 7 +[61] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. In EMNLP, 2023. 3 +[62] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2 +[63] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech large language models. arXiv preprint arXiv:2308.16692, 2023. 3 +[64] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2, 3 + +[65] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. In CVPR, 2024. 2 +[66] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 5 +[67] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3 \ No newline at end of file diff --git a/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/images.zip b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5814209c189762ce63f27138346c81c285b7230 --- /dev/null +++ b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea896e1ecd1af2c5aa5d8a2d8fd6f95722fb55fe92fe2c51a697320059d53852 +size 564550 diff --git a/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/layout.json b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7db7cc2b20515cc4f8a323e42b55c32712e7c7cb --- /dev/null +++ b/2025/SOLAMI_ Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters/layout.json @@ -0,0 +1,7568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 489, + 140 + ], + "type": "text", + "content": "SOLAMI: Social Vision-Language-Action Modeling for Immersive Interaction with 3D Autonomous Characters" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "spans": [ + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": "Jianping Jiang\\*†, Weiye Xiao\\*‡, Zhengyu Lin" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": ", Huaizhong Zhang" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": ", Tianxiang Ren" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": ", Yang Gao" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": ", Zhiqian Lin" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": ", Zhongang Cai\\*1,2, Lei Yang\\*1, Ziwei Liu\\*2 \n" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": "SenseTime Research, " + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 116, + 160, + 494, + 246 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University \n* Equal Contribution, † Project Lead, ‡ Engineering Lead, § Corresponding Author \nhttps://solami-ai.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 250, + 553, + 511 + ], + "blocks": [ + { + "bbox": [ + 58, + 250, + 553, + 511 + ], + "lines": [ + { + "bbox": [ + 58, + 250, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 58, + 250, + 553, + 511 + ], + "type": "image", + "image_path": "e6b04288844f9303802f257ea5d2e55ab414937fb1546b89de629177d5989676.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 521, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 54, + 521, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 555, + 544 + ], + "type": "text", + "content": "Figure 1. SOLAMI enables the user to interact with 3D autonomous characters through speech and body language in an immersive VR environment via an end-to-end social vision-language-action model, which is trained on our synthesized multimodal dataset SynMSI." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 554, + 200, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 554, + 200, + 566 + ], + "spans": [ + { + "bbox": [ + 151, + 554, + 200, + 566 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 582, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 298, + 714 + ], + "type": "text", + "content": "Human beings are social animals. How to equip 3D autonomous characters with similar social intelligence that can perceive, understand and interact with humans remains an open yet foundational problem. In this paper, we introduce SOLAMI, the first end-to-end Social vision-Language-Action (VLA) Modeling framework for Immersive interaction with 3D autonomous characters. Specifically, SOLAMI builds 3D autonomous characters from three aspects: 1) Social VLA Architecture: We propose a unified social VLA framework to generate multimodal response (speech and motion) based on the user's" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "content": "multimodal input to drive the character for social interaction. 2) Interactive Multimodal Data: We present SynMSI, a synthetic multimodal social interaction dataset generated by an automatic pipeline using only existing motion datasets to address the issue of data scarcity. 3) Immersive VR Interface: We develop a VR interface that enables users to immersively interact with these characters driven by various architectures. Extensive quantitative experiments and user study demonstrate that our framework leads to more precise and natural character responses (in both speech and motion) that align with user expectations with lower latency." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26887" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 222 + ], + "type": "text", + "content": "Have you ever imagined having an immersive face-to-face conversation with a character you deeply admire? Not merely through speech dialogue, but through an interaction where you can observe its subtle facial expressions, natural body language, and even fleeting emotional changes. Psychology research [15, 25, 46, 49] shows that in social interactions, the greater the level of immersion, the better the human experience. However, current character agents [1, 3, 47] are still limited to text or voice interactions. This limitation prompts us to build 3D autonomous characters with richer modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 294, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 294, + 366 + ], + "type": "text", + "content": "Developing an autonomous 3D character requires effectively modeling its behavior system, which involves two major challenges: 1) The 3D character needs to accurately observe and understand the information conveyed by the user, and respond appropriately based on the context and its character setting through speech, body motion, and facial expression, etc. This goes beyond previous singular human-related tasks, such as motion generation [62], motion understanding [27], and audio-to-motion [6]. 2) Data for multimodal interactions between users and 3D characters is extremely scarce due to the prohibitive cost of the comprehensive setup." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 510 + ], + "type": "text", + "content": "Previous work [13] is primarily based on the LLM-Agent framework, using text to link various sub-modules (such as motion captioning and text-to-motion). While this approach performs well in high-level tasks like planning and memory, it tends to fall short in tasks such as understanding user behaviors and providing timely body motion responses. This limitation arises because using text as the intermediary between modules conveys high-level information but often omits subtle nuances. And the sub-modules (motion captioning, speech recognition, etc.) in the complex engineering framework incur substantial latency that undermines the timeliness of a natural communication [50]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": "Interestingly, research on robotics shows the similar conclusion. The LLM-Agent framework can handle planning tasks [26], but for low-level manipulation tasks, end-to-end Vision-Language-Action (VLA) models built upon LLMs show superior performance [11, 24, 29, 67]. We argue that digital avatars are essentially robots with virtual humanoid embodiment. Therefore, building a VLA model for social interactions with users is a promising direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "content": "In this paper, we implement an end-to-end social VLA model, SOLAMI. Our model, built upon a decoder-only LLM backbone, processes the inputs of user speech and motion into discrete representations, and generate the responsive speech and motion tokens, which are then decoded to the character's speech and motion. This modeling approach can effectively learn character behavior patterns across motion and speech modalities and offer low latency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Although there are numerous datasets related to human" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "social behaviors [13, 34, 40, 56], comprehensive multimodal interaction datasets remain scarce. Thus, we introduce a data synthesis method that utilizes existing text-motion datasets to automatically construct multimodal interaction data at a low cost. Leveraging our extensively curated topics (5.3 K), uniformly processed motion database (46 K), and iterative script refinement pipeline, we develop SynMSI, a dataset containing 6.3 K multi-turn multimodal conversation items. To evaluate the effectiveness of our method, we developed a VR interface where users can immersively interact with various 3D characters. Quantitative experimental results and user study analysis show that our approach produces more precise and natural social interaction experience with lower latency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 239, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 324 + ], + "type": "text", + "content": "To summarize, we contribute 1) A new VLA architecture to model the character's behavior system for immersive social interaction; 2) A dedicated synthesizing pipeline that automatically generates large-scale multimodal interactive dataset, SynMSI; 3) An immer VR interface for users to interact with various characters through speech and motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 400, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 400, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 400, + 349 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 449, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 449, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 449, + 369 + ], + "type": "text", + "content": "2.1. Human Motion & LLM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 651 + ], + "type": "text", + "content": "Existing role-play agents primarily rely on text [47], speech [1, 3, 41], or video [7] as interactive media, but building 3D autonomous characters means modeling 3D embodied behavior, especially body language. Unlike single-purpose human motion tasks [6, 12, 22, 34, 48, 56, 62], we expect this 3D character not only to comprehend user speech and body language but also to respond according to its profile setting. Large language models (LLMs) [41, 51], with their remarkable emergent abilities [54], provide promising solutions for this direction. One approach [18, 28, 64, 65] tries to integrate LLMs and human motion in an end-to-end fashion, enabling a single model to perform multiple motion-related tasks. However, the goal of these methods is not to generate responsive motion based on the input motion according to the character setting. Another approach [13, 35] utilizes the LLM as a versatile brain center that controls various submodules (e.g., motion understanding, text-to-motion generation) with text or code instructions. However, this modular approach inherently introduces information loss and engineering-related time latency. Therefore, how to create feasible autonomous 3D characters for immersive interaction remains an open challenge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 659, + 443, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 443, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 443, + 672 + ], + "type": "text", + "content": "2.2. Embodied Intelligence" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Building 3D autonomous characters means creating virtual humanoid agents with embodied intelligence. In the field of embodied intelligence, researchers have found the high-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26888" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "level abilities (planning, memory, etc.) of LLM-Agents across various environments, including factories [16, 26], 2D sandbox settings [42], and 3D gaming spaces [53]. For tasks requiring low-level skills, such as manipulation, end-to-end VLA models [11, 24, 29, 67] have shown considerable potential. Despite existing efforts [19, 20, 22] in egocentric human-related tasks, the capabilities of VLA models in social interactive tasks with humans have not been fully explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 281, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 281, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 281, + 206 + ], + "type": "text", + "content": "3. Social Vision-Language-Action Modeling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 138, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 138, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 138, + 223 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 373 + ], + "type": "text", + "content": "Our framework is an end-to-end social VLA model that takes the user's speech and motion as input and generates the character's responsive speech and motion as output. In this process, speech and motion are added as new languages to the LLM text vocabulary. First, the user's speech and motion are converted into discrete motion tokens and speech tokens via a motion tokenizer and a speech tokenizer, respectively. A decoder-only LLM backbone then predicts the character's output motion and speech tokens based on the user's input tokens and the character setting. The generated tokens are subsequently decoded into corresponding motion and speech by their respective decoders." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": "Motion Representation. To take advantage of SMPL-X's [43] compatibility with industry animation workflow, we directly model human poses as SMPL-X joint rotations instead of keypoint positions [28, 38] to facilitate animation of characters down the stream." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": "Motion Tokenizer. Our motion tokenizer employs a Vector Quantized Variational Autoencoders (VQ-VAE) structure as [28, 64]. It learns discrete representations of motion, enabling the LLM to understand the text-motion connection. We design separate VQVAEs " + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "inline_equation", + "content": "(Q^{b}, Q^{h}, Q^{t})" + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": " for the body motion " + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{b}" + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": ", hand motion " + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{h}" + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": ", and inter-character relative transform (rotation and translation) " + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{t}" + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": " for higher reconstruction accuracy [38]. The quantization process for motion " + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{u}" + }, + { + "bbox": [ + 55, + 434, + 296, + 540 + ], + "type": "text", + "content": " is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 551, + 295, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 551, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 91, + 551, + 295, + 570 + ], + "type": "interline_equation", + "content": "\\hat {m} _ {t} ^ {u} = Q ^ {u} \\left(\\mathbf {m} _ {t} ^ {u}\\right) = \\arg \\min _ {z _ {i} \\in \\mathbb {Z} _ {u}} \\| \\mathbf {m} _ {t} ^ {u} - z _ {i} \\| _ {2}, \\tag {1}", + "image_path": "673430c4b754cedbc7bb1c1aaaf8b1ff55c8cb702aa2b4fee9920ad0fc2453fe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}_u" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " is the codebook of motion part " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "u\\in \\{b,h,t\\}" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\hat{m}_t^u" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " is the corresponding motion tokens. The VQVAEs of body and hand, " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "Q^{b}" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "Q^{h}" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": ", apply 1D convolutions on motion features along the temporal dimension to get " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "L_{M}" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " sequential tokens " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\hat{M}^b = [\\hat{m}_1^b,\\hat{m}_2^b,\\dots,\\hat{m}_{L_M}^b]" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\hat{M}^h = [\\hat{m}_1^h,\\hat{m}_2^h,\\dots,\\hat{m}_{L_M}^h]" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": ", and the VQVAE " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "Q^{t}" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " uses MLPs to get the transform token " + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\hat{m}^t" + }, + { + "bbox": [ + 55, + 578, + 296, + 665 + ], + "type": "text", + "content": " of a sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Speech Tokenizer. Research [9, 61] on speech discretization mainly utilizes the RVQ-VAE structure [59]. In this work, we utilize the SpeechTokenizer [63] that disentangles semantic and acoustic information within speech " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". This" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "allows us to use the semantic tokens " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\hat{S} = [\\hat{s}_1^s, \\hat{s}_2^s, \\dots, \\hat{s}_{L_S}^s]" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " from the first layer as input to the LLM, reducing inference costs (" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "L_S" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is the sequence length of semantic tokens). Simultaneously, a short sample of the character's voice (4 to 6 seconds) can be used as a prompt to achieve instance voice cloning when decoding through SoundStorm [10]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "content": "Multi-modal Multi-round Interaction. User-character interaction is formulated as a multi-round conversation fashion of common LLMs [41, 51]. When the user sends speech and motion to SOLAMI, the model auto-regressively generates speech and motion responses based on previous dialogue contents and the character setting. To facilitate training, we use special tokens to mark the start and end of each modality sequence as [28, 30]. The interaction process can be formulated as follow:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 261, + 416, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 261, + 416, + 273 + ], + "spans": [ + { + "bbox": [ + 331, + 261, + 416, + 273 + ], + "type": "text", + "content": "Interaction Template" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 331, + 283, + 524, + 379 + ], + "blocks": [ + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "lines": [ + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "spans": [ + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "text", + "content": "Input: \nSystem Prompt: \nUser: " + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "inline_equation", + "content": "< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>" + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "text", + "content": " \nCharacter: " + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "inline_equation", + "content": "< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>" + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "text", + "content": " \n... \nUser: " + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "inline_equation", + "content": "< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>" + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "text", + "content": " \nOutput: \nCharacter: " + }, + { + "bbox": [ + 331, + 283, + 524, + 379 + ], + "type": "inline_equation", + "content": "< \\mathrm{M\\_Placeholder} > < \\mathrm{S\\_Placeholder}>" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "inline_equation", + "content": "\\langle \\text{Character\\_Placeholder} \\rangle" + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "text", + "content": " is placeholder for character's text description, " + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "inline_equation", + "content": "\\langle \\text{M\\_Placeholder} \\rangle" + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "text", + "content": " for motion token sequences, and " + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "inline_equation", + "content": "\\langle \\text{S\\_Placeholder} \\rangle" + }, + { + "bbox": [ + 313, + 392, + 554, + 440 + ], + "type": "text", + "content": " for speech token sequences." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 448, + 378, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 448, + 378, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 448, + 378, + 460 + ], + "type": "text", + "content": "3.2. Training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 466, + 534, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 534, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 534, + 478 + ], + "type": "text", + "content": "The training of SOLAMI adopts a three-stage strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 478, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 554, + 512 + ], + "type": "text", + "content": "Stage 1: Tokenizer Training. The training approach for the motion tokenizer uses the fashion of [28]. The train loss is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 515, + 553, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 515, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 356, + 515, + 553, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\lambda_ {r} \\mathcal {L} _ {r} + \\lambda_ {e} \\mathcal {L} _ {e} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {v} \\mathcal {L} _ {v}, \\tag {2}", + "image_path": "93daefd481a37042fba2de4a8230b1d8766c1e7b670f1c058d7d25c7a21a0753.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": " means reconstruction loss, " + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": " embedding loss, " + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": " commitment loss, " + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": " velocity loss, and " + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\lambda_*" + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": " are manually adjusted weights. For the speech tokenizer, we use the pre-trained checkpoint from AnyGPT [60]. We freeze the tokenizers' weights after this stage." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Stage 2: Multi-task Pre-training for Modality Alignment. As shown in Fig. 2, the second stage is multi-task pre-training, achieving modality alignment between motion and text, as well as between speech and text. It is necessary because motion data is scarce, and direct training on multimodal interaction data results in suboptimal models, as demonstrated in subsequent ablation studies. For motion and text alignment, we use " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "46\\mathrm{K}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " motion-text pairs for text-to-motion generation and motion captioning tasks, and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "11\\mathrm{K}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " interactive motion pairs" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26889" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 547, + 277 + ], + "type": "image", + "image_path": "2319dea11b5e8a6fed26e0c49dd3742d1f98014a560f6b5eddec7c09e28c05fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. Training pipeline of SOLAMI. We train SOLAMI through a three-stage process. In the pre-training stage, we train the model with motion-text and speech-text related tasks to align the speech and motion modalities with language. During the instruction tuning stage, we train the model with social multimodal multi-round interaction data, enabling it to generate multimodal responses that align with the character settings and the context of the topic." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": "for two-person motion generation. To align the speech and text, we train the model with " + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "inline_equation", + "content": "410\\mathrm{K}" + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": " speech-text pairs for text-to-speech and automatic speech recognition tasks, and " + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}" + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": " speech dialogue pairs for speech-to-speech generation. The tasks are formulated as \"User: ; Character: \". To balance the scale disparity between the motion and speech data, we sampled them at a 4:6 ratio during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 444, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 599 + ], + "type": "text", + "content": "Stage 3: Instruction Tuning for Multi-turn Conversat- tion. In the third stage, we perform instruction tuning by applying supervised fine-tuning with multimodal interaction data, enabling the model to handle long-sequence, multi-turn dialogues, as shown in Fig. 2. We utilize " + }, + { + "bbox": [ + 55, + 444, + 296, + 599 + ], + "type": "inline_equation", + "content": "5.7\\mathrm{K}" + }, + { + "bbox": [ + 55, + 444, + 296, + 599 + ], + "type": "text", + "content": " multimodal conversation items for supervised fine-tuning. Each conversation item is fed to the model in the format as the Interaction Template in Sec. 3.1. We experiment with two approaches: full-parameter fine-tuning and LoRA finetuning [23]. We supervised only the character's response to teach the model how to react to the user's behavior. Thus we train the model by maximizing the log-likelihood of the next-token prediction and the train loss is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 609, + 296, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 609, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 64, + 609, + 296, + 689 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {I T}} = - \\sum_ {r = 1} ^ {R} \\sum_ {i = 1} ^ {L _ {M} ^ {r}} \\log p _ {\\Theta} (\\hat {m} _ {i} ^ {r} | \\hat {m} _ {i - 1} ^ {r}, \\dots , \\hat {m} _ {1} ^ {r}, \\hat {S} _ {< r}, \\hat {M} _ {< r}) \\\\ - \\sum_ {r = 1} ^ {R} \\sum_ {i = 1} ^ {L _ {S} ^ {r}} \\log p _ {\\Theta} \\left(\\hat {s} _ {i} ^ {r} \\mid \\hat {s} _ {i - 1} ^ {r}, \\dots , \\hat {s} _ {1} ^ {r}, \\hat {S} _ {< r}, \\hat {M} _ {\\leq r}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "ee96386f0ab89e7b253640199ed8f2b534368dc116eba02979bc8d1aaf280a62.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the network parameter of the LLM backbone or LoRA weights, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the conversation round, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{S}_r" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{M}_r" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "text", + "content": "are the " + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "text", + "content": "-th round speech and motion token sequences with lengths " + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "inline_equation", + "content": "L_{M}^{r}" + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "inline_equation", + "content": "L_{S}^{r}" + }, + { + "bbox": [ + 313, + 336, + 555, + 361 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 370, + 413, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 413, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 413, + 384 + ], + "type": "text", + "content": "4. SynMSI Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "type": "text", + "content": "Social interaction between users and virtual characters is inherently unique, which makes collecting such multimodal interaction data particularly challenging. Currently, available public datasets [21, 40, 56] are incomplete for our task. To address this issue, we propose a data synthesis pipeline that leverages existing motion-text datasets, text-based role-play models, and speech synthesis methods and generates a large-scale multimodal dialogue dataset, SynMSI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 493, + 397, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 397, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 397, + 504 + ], + "type": "text", + "content": "4.1. Motion Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "We collect motion-text data for two purposes: first, to achieve alignment between motion and text during pretraining, and second, to generate multimodal data for instruction tuning. Since our work focuses on modeling social interactions, we select existing datasets that contain rich social actions: HumanML3D [21] with " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "24\\mathrm{K}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " motion-text pairs, Inter-X [56] with " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "20\\mathrm{K}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " motion-text pairs and 10 K two-person motion pairs, and DLP-MoCap [13] with 2 K motion-text pairs. Since the Inter-X [56] dataset contains only text descriptions of two-person interactive motion without descriptions of individual motion, we used GPT-4o [41] to decompose the two-person action descriptions into single-person motion-text pairs. Additionally, we used GPT-4o [41] to synthesize comprehensive text descriptions for each motion clip by consolidating multiple possible descriptions, thereby providing more detailed textual annotations that preserve motion details." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26890" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 348, + 277 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 348, + 277 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 348, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 348, + 277 + ], + "type": "image", + "image_path": "8b8c986f2ba86c131ae2e8df79e3ec0882a12d91b99c1ab5ed2be0f5b8499f9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 289, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 289, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 555, + 335 + ], + "type": "text", + "content": "Figure 3. SynMSI dataset generation. Our synthesizing pipeline consists of 4 steps. Based on numerous character-relevant topics and state-of-the-art LLMs [41], we generate text scripts for multimodal dialogues. Using a large-scale motion database [13, 21, 56], we retrieve the most appropriate motions and refine the speech scripts accordingly. Finally, we employ TTS/voice cloning [14] to generate character-specific speech. This approach enables us to create multimodal interaction data of various characters using only existing motion datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 356, + 76, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 356, + 76, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 356, + 76, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 356, + 76, + 547, + 277 + ], + "type": "image", + "image_path": "90fd8d13848f8105c4f4d7d1122c58b7728eeaf2fd7ea279e6049fadaf97e53e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 342, + 138, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 138, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 138, + 357 + ], + "type": "text", + "content": "4.2. Speech Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "content": "We use speech-text data for speech-text alignment in the pre-training stage. Speech datasets involve CommonVoice [8] (150 K speech-text pairs in our experiments), AnyInstruct [60] (200 K speech-text pairs and 100 K speech-to-speech items), and synthetic speech data (60 K speech-text pairs) by text-to-speech approaches (Azure TTS and XTTS_v2 [14])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 222, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 222, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 222, + 468 + ], + "type": "text", + "content": "4.3. Multimodal Data Synthesizing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 557 + ], + "type": "text", + "content": "Previous data synthesis methods [32, 37, 66] usually use the general abilities of advanced LLMs [41] and text-annotated multimodal data to generate synthetic data. However, generating social multimodal interaction data has not yet been achieved. This challenging task requires high-quality expression of body language, voice consistency that matches the characters, and suitable dialogue content." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, our synthesizing pipeline includes 4 steps. (1) First, we collect " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "5.3\\mathrm{K}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " character-related and daily topics from internet platforms (Google Trends [2], Zhihu [5], Jike) and brainstorms of GPT-4o [41] to improve the diversity of the dialogue contents. (2) Based on the topic, character setting, and previous round of scripts, we use GPT-4o [41] to generate textual descriptions (motion, speech, expression etc.) for the next round of the dialogue. (3) Then we utilize the text embedding [39] of the motion description to retrieve the most relevant motions from our meticulously curated motion-text database. (4) Moreover, we refine the generated speech text with the retrieved motions to ensure that speech and motion are well-coordinated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 343, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 463 + ], + "type": "text", + "content": "The motion database with detailed text annotations and the refinement process can alleviate the misalignment between the real motion and speech in the LLM-Agent method [13]. By iteratively repeating this process, we can generate multi-round dialogue contents across many characters, where the motions are sourced from the motion database, and the speeches are synthesized using TTS/voice cloning (Azure TTS and XTTS_v2 [14]) to maintain consistency with the character's voice style. We finally obtained " + }, + { + "bbox": [ + 313, + 343, + 555, + 463 + ], + "type": "inline_equation", + "content": "6.3\\mathrm{K}" + }, + { + "bbox": [ + 313, + 343, + 555, + 463 + ], + "type": "text", + "content": " multimodal dialogue items." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 477, + 398, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 398, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 398, + 490 + ], + "type": "text", + "content": "5. VR Interface" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "To demonstrate our method directly, we developed a VR interface with an Oculus Quest 3 frontend and a backend service, as shown in Fig. 4. The frontend enables immersive interaction between users and 3D autonomous characters, while the backend, powered by 2 H800 GPUs, supports the computation of various baselines in our experiments. During usage, the VR headset captures the user's speech and body motion, and sends them to the backend computation nodes. For motion capture, we use the Quest's full-body tracking system [55] to obtain pose parameters, which are then retargeted onto an SMPL-X model [43]. The computation nodes generate the body motion parameters and speech responses of the character based on the multimodal input. Then we apply an audio-to-face method, UniTalker [17], to generate the character's facial animation parameters. The facial and body parameters are jointly retargeted onto a 3D character model [58], completing one cycle of social interaction. For a natural user experience, we employ preset idle" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "26891" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 190 + ], + "type": "image", + "image_path": "516a93c459e60f9f5c0fda17fe26afbc63d8ab83fbfa8702af7e6f6b6f7b38f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 264 + ], + "type": "text", + "content": "Figure 4. VR interface architecture. Our VR project consists of a Quest 3 client and a server. The Quest client captures and transmits user body motion and speech to the server. The server then generates character's speech, body motion, and face blendshape parameters based on the selected methods. The response is then sent back to the Quest client to drive the character." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 295, + 300 + ], + "type": "text", + "content": "motions on the character when the method is LLM+Speech or the character is waiting for the user's input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 420 + ], + "type": "text", + "content": "3D Character Assets. Our 3D character portfolio covers a diverse range of entities, including AI assistant avatars, famous cinematic roles, internet memes, and real-world celebrity personas. These 3D models are sourced from open-source repositories under CC Attribution-NonCommercial-ShareAlike License as well as our manual creation using VRoid Studio [4]. We subsequently employ facial rigging, skinning, bone chain simulation, retargeting, and texture and material creation in Unity Engine processes to yield functional 3D character assets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 137, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 137, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 137, + 448 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 182, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 182, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 182, + 468 + ], + "type": "text", + "content": "6.1. Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "content": "In our experiment, we selected the AnyGPT-base model [60] (based on LLaMA2-7B [51]) as the backbone for SOLAMI, because it is an open-source model available at our experimental time that supports end-to-end speech processing. During the pre-training stage, we utilize 32 V100 GPUs to train the model for " + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "inline_equation", + "content": "3\\mathrm{K}" + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "content": " steps (batch size 256, learning rate 4e-5). For instruction tuning, we train the SOLAMI for 800 steps using 16 V100 GPUs (batch size 48, learning rate 2e-5). For LoRA fine-tuning [23], we set the rank as 8 and alpha as 16. We split the synthesized multimodal data into training and test sets with a 9:1 ratio. We use DeepSpeed [45] to accelerate the training. During testing, we evaluate each round of the character's response." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Baselines. To validate the performance improvement in social interaction brought by incorporating 3D modalities (such as body motion), we compared SOLAMI with the LLM+Speech and the AnyGPT (fine-tune) approach. For the LLM+Speech framework, the user's speech is first transcribed into text using ASR techniques, which is then fed to a LLM to generate the character's response in text, and sub" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 373 + ], + "type": "text", + "content": "sequentially converted into speech using TTS. For fairness, we use LLaMA2-7B-Chat [51] as the LLM backbone, Whisper large-v3 [44] for ASR, and XTTS_v2 [14] for voice cloning. For the AnyGPT (fine-tune) framework, we use the speech data of SynMSI to train the AnyGPT-base model [60] with the same parameter settings. To compare the effectiveness of the LLM-Agent architecture with the social VLA framework, we used DLP [13] as a baseline method. In DLP framework, the user's speech and body motion are separately processed as text descriptions by ASR and motion captioning modules. Based on the input text descriptions, LLM generates the character's text instructions of speech and motion, which are transferred into speech and body motion by TTS and motion generation module. The speech component of DLP follows the LLM+Speech method. Considering that MoMat-MoGen module in DLP is too slow for user interaction (over 5 seconds latency), we use MotionGPT [28] for motion captioning and motion generation. To ensure a fair comparison, we used the same motion data as the pre-training stage of SOLAMI to train MotionGPT. Additionally, we conducted ablation experiments on the effect of the pre-training stage, marked as (w/o pretrain). For the ablation study of the motion tokenizer, please refer to the supplementary materials. We use vLLM [31] to accelerate the LLM backbones for low latency interaction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 450, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 450, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 450, + 395 + ], + "type": "text", + "content": "6.2.Quantitative Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 422 + ], + "type": "text", + "content": "We conducted quantitative evaluation for our method and all the baselines mentioned in Sec. 6.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 640 + ], + "type": "text", + "content": "Evaluation Metrics. For motion, we evaluate the model responses using metrics including FID, diversity, PAMPJPE (mm), and angle error [22]. Following Duolando [48], we obtain FID and diversity using motion features from AIST++ [33]. For speech, we use VC similarity [57] to evaluate the voice similarity with the character. To evaluate the content quality of speech, we first use Whisper-large-v3 [44] to transcribe the speech into text. Then following [47, 52], we employ GPT-4o [41] as the judge to assess Context Relevance and Character Consistency on a Likert scale ranging from 1 to 5. Context Relevance indicates whether the speech content aligns with the topic and context of the conversation, while Character Consistency assesses whether the content adheres to the character settings. For inference latency (seconds), we deploy all the models on 2 H800 GPUs with vLLM [31] framework and asynchronous mechanisms to improve performance while maintaining fairness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "type": "text", + "content": "Quantitative Results. The quantitative results in Tab. 1 demonstrate that, using the same foundation model (Llama2) [51] as the backbone, our method, SOLAMI (full params), significantly outperforms other methods in terms of motion quality and inference latency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Motion Quality. SOLAMI demonstrates superior perfor" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26892" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 101, + 555, + 183 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative results of baselines and SOLAMI. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\uparrow^{\\prime}(\\downarrow^{\\prime})" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates that the values are better if the metrics are larger (smaller). We run all the evaluations 5 times and report the average metric. The best results are in bold and the second best results are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 101, + 555, + 183 + ], + "lines": [ + { + "bbox": [ + 57, + 101, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 101, + 555, + 183 + ], + "type": "table", + "html": "
MethodsMotion MetricsSpeech MetricsInference Latency↓
FID↓Diversity↑PA-MPJPE↓Angle Error↓VC Similarity↑Context Relevance↑Character Consistency↑
SynMSI Dataset-9.136---4.8884.893-
LLM+Speech (Llama2) [51]----0.8183.5273.8593.157
AnyGPT (fine-tune) [60]----0.8193.5023.8032.588
DLP (MotionGPT) [13]4.2548.259165.0530.4950.8123.5773.7855.518
SOLAMI (w/o pretrain)5.0528.558159.7090.3870.8203.5413.4612.657
SOLAMI (LoRA)15.7298.145167.1490.4000.7703.2513.4232.710
SOLAMI (full params)3.4438.853151.5000.3600.8243.6343.8242.639
", + "image_path": "cd0a622a372203de9c0c79b8ca70d5ded664638a386dddd025e5e34cd42991e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 192, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 192, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 192, + 294, + 324 + ], + "type": "text", + "content": "mance compared with the DLP method [13] across multiple motion metrics. This is because the social VLA model achieves comprehensive modality alignment among speech, motion, and language through training on our high-fidelity character-specific multimodal data. Our model can precisely perceive the user's physical motions and linguistic cues, enabling semantically rich interactive motions in response. This ability contrasts with the LLM-Agent architectures of the DLP method, which exhibits limitations in conveying multimodal nuances through text-only intermediary representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 295, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 494 + ], + "type": "text", + "content": "Speech Quality. Our method demonstrates the capability to synthesize a voice tone that matches the character with a higher Voice Cloning (VC) Similarity score. Our model also shows better performance on the context relevance of the dialogue than other methods. Because the inclusion of the motion modality enables the model to perceive the user's body language, while the LLM+Speech or AnyGPT(fine-tune) method lacks this capability. In terms of character consistency, our model achieves secondary performance metrics. We suspect this may be due to the incorporation of motion and speech modalities, which potentially affects the character-related knowledge embedded within the original LLM. The performance degradation is similar to the observations in [16, 36]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 579 + ], + "type": "text", + "content": "Inference Latency. Our end-to-end approach is significantly superior to the modular pipeline approaches (LLM+Speech or DLP). Because the end-to-end VLA model naturally aligns with the process of real-time human communication. Theoretically, if we could collect data on real-time interactions between humans and characters, our method could achieve full-duplex streaming interaction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Ablation Study. As shown in Tab. 1, the pre-training stage of SOLAMI leads to better performance in both motion and speech quality. We believe that the pre-training stage, which aligns motion, speech, and language, facilitates the model's ability to learn the multimodal dialogue skill during the instruction tuning stage. Instruction tuning using LoRA [23] shows weaker results compared to full parameter fine-tuning. We think that the gap between the data distribution of pre-training tasks and the multimodal instruction tuning task is substantial, and LoRA fine-tuning alone is insufficient for the model to learn the strong multimodal" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 210, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 345, + 191, + 523, + 202 + ], + "lines": [ + { + "bbox": [ + 345, + 191, + 523, + 202 + ], + "spans": [ + { + "bbox": [ + 345, + 191, + 523, + 202 + ], + "type": "text", + "content": "Table 2. Questionnaire settings of our user study." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 210, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 316, + 210, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 555, + 285 + ], + "type": "table", + "html": "
DimensionQuestions
Motion CoherenceDoes the motion match the character's setting?Does the motion align well with speech?
Motion InteractionDoes the character follow motion instructions correctly?Does the character understand user's motion?
Speech ConsistencyDoes the speech match the character's setting?Is the speech relevant to the current topic?
Overall ExperienceHow would you rate the overall experience?
", + "image_path": "5f87743c6f98aea5e739cc80d0989cf505918fa5e6d2ab9d1e0cfdd34f00a53a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 297, + 552, + 417 + ], + "blocks": [ + { + "bbox": [ + 318, + 297, + 552, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 297, + 552, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 297, + 552, + 417 + ], + "type": "image", + "image_path": "77940dc7c538e9e7b11b4f79aed699930c925aa63230fa19d4e9a7982fd51406.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 428, + 537, + 439 + ], + "lines": [ + { + "bbox": [ + 330, + 428, + 537, + 439 + ], + "spans": [ + { + "bbox": [ + 330, + 428, + 537, + 439 + ], + "type": "text", + "content": "Figure 5. Results of the user study with " + }, + { + "bbox": [ + 330, + 428, + 537, + 439 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 330, + 428, + 537, + 439 + ], + "type": "text", + "content": " confidence." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 448, + 372, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 372, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 372, + 460 + ], + "type": "text", + "content": "dialog ability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 453, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 453, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 453, + 481 + ], + "type": "text", + "content": "6.3. VR Interface User Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 485, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 616 + ], + "type": "text", + "content": "Quantitative evaluation of a single modality alone cannot fully compare the 3D autonomous characters built on different frameworks. To address this, we conducted user study with a VR interface. As shown in Fig. 4, we employ LLM+Speech, DLP (MotionGPT), and SOLAMI as method options of the server backend and the same VR frontend across different methods. This implies that the variable in the experiment is the driving method in the server. Users are asked to engage in more than five rounds of dialogue with the VR character before completing our questionnaire." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. The indicators and corresponding questions for our questionnaire are shown in Tab. 2. Motion Coherence evaluates whether the character's motion aligns with the character setting and dialogue content. Motion Interaction assesses whether the character can understand the semantics of the body motion and effectively interact with the user. Speech Coherence examines whether the generated speech aligns with the context and character settings." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26893" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 76, + 555, + 371 + ], + "blocks": [ + { + "bbox": [ + 55, + 76, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 76, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 76, + 555, + 371 + ], + "type": "image", + "image_path": "088896000ac6620228c85e6cb79d98c7e5d84c86e3527e5f156506411bb69003.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 377, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 555, + 411 + ], + "type": "text", + "content": "Figure 6. Qualitative results of SOLAMI and baselines, and the user workflow for VR experience. Our social VLA model, trained in an end-to-end strategy on SynMSI dataset, can accurately perceive the semantic information embedded within users' speech and motion input, and subsequently generate natural and coherent responses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 418, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 502 + ], + "type": "text", + "content": "And the Overall Experience measures the user's satisfaction with the overall experience. Each question is rated on a 1 to 5 Likert scale, with higher scores indicating greater satisfaction. The score for each dimension is calculated as the average score of its corresponding questions. We ultimately collected 60 survey responses, with participants from various gender and age groups." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 625 + ], + "type": "text", + "content": "Results. As shown in the Fig. 5, our method achieved the highest user satisfaction across all dimensions. SOLAMI demonstrates superior performance over the DLP method across all metrics, validating the effectiveness of an end-to-end social VLA model in character behavior modeling. While the DLP method shows lower speech consistency compared to the LLM+Speech method, it excels in motion-related metrics and achieves higher overall satisfaction, indicating that effective body language can enhance user experience despite speech quality limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "To provide a more intuitive demonstration of our model's capabilities, we replay the user study experiments and render them, as shown in the Fig. 6. The results indicate that SOLAMI demonstrates excellent capabilities in body language understanding, motion command execution, and body interaction. For better understanding, we also present the workflow from a first-person view during actual usage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 418, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 465 + ], + "type": "text", + "content": "We encourage readers to explore our supplementary materials and video for a more detailed overview of the model experiments, data generation pipeline, VR interface construction process, and comprehensive experimental results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 474, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 388, + 487 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 495, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 614 + ], + "type": "text", + "content": "In this paper, we propose SOLAMI, an approach for building 3D autonomous characters. This approach includes three key components: 1) Architecture: A novel social VLA modeling framework enabling multimodal social interaction; 2) Multimodal Data Synthesizing: A pipeline for automatically generating multimodal interaction data from existing incomplete datasets; 3) VR Interface: A VR engineering framework that facilitates immersive interactions between users and various characters. Together, these modules contribute to an enhanced user interaction experience." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 619, + 411, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 619, + 411, + 632 + ], + "spans": [ + { + "bbox": [ + 314, + 619, + 411, + 632 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 639, + 554, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 554, + 711 + ], + "type": "text", + "content": "This study is supported by the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOET2EP20221-0012, MOE-T2EP20223-0002), and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26894" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 244, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 244, + 101 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 244, + 101 + ], + "type": "text", + "content": "[1] Character.ai. https://character.ai.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 104, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 104, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 62, + 104, + 294, + 124 + ], + "type": "text", + "content": "[2] Google trends. https://trends.google.com/trends.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 127, + 260, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 127, + 260, + 137 + ], + "spans": [ + { + "bbox": [ + 62, + 127, + 260, + 137 + ], + "type": "text", + "content": "[3] Talkie ai. https://www-talkie-ai.com. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 140, + 284, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 284, + 150 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 284, + 150 + ], + "type": "text", + "content": "[4] Vvoid studio. https://vvoid.com/en/studio.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 152, + 228, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 228, + 162 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 228, + 162 + ], + "type": "text", + "content": "[5] Zhihu. https://www.zhihu.com. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 165, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 165, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 165, + 295, + 196 + ], + "type": "text", + "content": "[6] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM TOG, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 199, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 294, + 230 + ], + "type": "text", + "content": "[7] Tenglong Ao. Body of her: A preliminary study on end-to-end humanoid agent. arXiv preprint arXiv:2408.02879, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 234, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 234, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 234, + 294, + 286 + ], + "type": "text", + "content": "[8] Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis M. Tyers, and Gregor Weber. Common voice: A massively-multilingual speech corpus. In LREC, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 289, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 294, + 343 + ], + "type": "text", + "content": "[9] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matthew Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. TASLP, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 346, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 346, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 346, + 295, + 388 + ], + "type": "text", + "content": "[10] Zalán Borsos, Matt Sharifi, Damien Vincent, Eugene Kharitonov, Neil Zeghidour, and Marco Tagliasacchi. Soundstorm: Efficient parallel audio generation. arXiv preprint arXiv:2305.09636, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 391, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 391, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 391, + 295, + 554 + ], + "type": "text", + "content": "[11] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In RSS, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 610 + ], + "type": "text", + "content": "[12] Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qingping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, Chen Change Loy, Lei Yang, and Ziwei Liu. Smpler-x: Scaling up expressive human pose and shape estimation. In NeurlPS, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 689 + ], + "type": "text", + "content": "[13] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, Liang Pan, Xiangyu Fan, Han Du, Peng Gao, Zhitao Yang, Yang Gao, Jiaqi Li, Tianxiang Ren, Yukun Wei, Xiaogang Wang, Chen Change Loy, Lei Yang, and Zwei Liu. Digital life project: Autonomous 3d characters with social intelligence. In CVPR, 2024. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[14] Edresson Casanova, Kelly Davis, Eren Gölgé, Görkem Göknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": "Meyer, Reuben Morais, Samuel Olayemi, et al. Xtts: a massively multilingual zero-shot text-to-speech model. arXiv preprint arXiv:2406.04904, 2024. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[15] James J Cummings and Jeremy N Bailenson. How immersive is enough? A meta-analysis of the effect of immersive technology on user presence. Media psychology, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 227 + ], + "type": "text", + "content": "[16] Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. Palm-e: An embodied multimodal language model. In ICML, 2023. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 229, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 229, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 229, + 553, + 260 + ], + "type": "text", + "content": "[17] Xiangyu Fan, Jiaqi Li, Zhiqian Lin, Weiye Xiao, and Lei Yang. Unitalker: Scaling up audio-driven 3d facial animation through A unified model. 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 262, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 553, + 294 + ], + "type": "text", + "content": "[18] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Chatpose: Chatting about 3d human pose. In CVPR, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 297, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 553, + 580 + ], + "type": "text", + "content": "[19] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenjiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 582, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 713 + ], + "type": "text", + "content": "[20] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, Eugene Byrne, Zachary Chavis, Joya Chen, Feng Cheng, Fu-Jen Chu, Sean Crane, Avijit Dasgupta, Jing Dong, María Escobar, Cristhian Forigua, Abrham Gebreselasie, Sanjay Haresh, Jing Huang, Md Mohaiminul Islam, Suyog Dutt Jain, Rawal Khirodkar, Devansh Kukreja, Kevin J. Liang, Jia-Wei Liu, Sagnik Majumder, Yongsen Mao, Miguel Martin, Effrosyni Mavroudi, Tushar Nagarajan, Francesco Ragusa, Santhosh Kumar Ramakrishnan, Luigi Seminara, Arjun Somayazulu, Yale Song, Shan Su, Zihui Xue, Ed" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26895" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 281 + ], + "type": "text", + "content": "ward Zhang, Jinxu Zhang, Angela Castillo, Changan Chen, Xinzhu Fu, Ryosuke Furuta, Cristina González, Prince Gupta, Jiabo Hu, Yifei Huang, Yiming Huang, Weslie Khoo, Anush Kumar, Robert Kuo, Sach Lakhavani, Miao Liu, Mi Luo, Zhengyi Luo, Brighid Meredith, Austin Miller, Oluwatuminu Oguntola, Xiaqing Pan, Penny Peng, Shraman Pramanick, Merey Ramazanova, Fiona Ryan, Wei Shan, Kiran Somasundaram, Chenan Song, Audrey Southerland, Masatoshi Tateno, Huiyu Wang, Yuchen Wang, Takuma Yagi, Mingfei Yan, Xitong Yang, Zecheng Yu, Shengxin Cindy Zha, Chen Zhao, Ziwei Zhao, Zhifan Zhu, Jeff Zhuo, Pablo Arbeláez, Gedas Bertasius, Dima Damen, Jakob Engel, Giovanni Maria Farinella, Antonino Furnari, Bernard Ghanem, Judy Hoffman, C. V. Jawahar, Richard A. Newcombe, Hyun Soo Park, James M. Rehg, Yoichi Sato, Manolis Savva, Jianbo Shi, Mike Zheng Shout, and Michael Wray. Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives. In CVPR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 282, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 282, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 57, + 282, + 294, + 314 + ], + "type": "text", + "content": "[21] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 316, + 294, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 359 + ], + "type": "text", + "content": "[22] Fangzhou Hong, Vladimir Guzov, Hyo Jin Kim, Yuting Ye, Richard Newcombe, Ziwei Liu, and Lingni Ma. Egolm: Multi-modal language model of egocentric motions. arXiv preprint arXiv:2409.18127, 2024. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 57, + 360, + 294, + 403 + ], + "type": "text", + "content": "[23] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 404, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 404, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 57, + 404, + 294, + 447 + ], + "type": "text", + "content": "[24] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 449, + 294, + 491 + ], + "type": "text", + "content": "[25] Sarah Hudson, Sheila Matson-Barkat, Nico Pallamin, and Guillaume Jegou. With or without you? Interaction and immersion in a virtual reality experience. Journal of business research, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 493, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 493, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 493, + 294, + 645 + ], + "type": "text", + "content": "[26] Brian Ichter, Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, Dmitry Kalashnikov, Sergey Levine, Yao Lu, Carolina Parada, Kanishka Rao, Pierre Sermanet, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Mengyuan Yan, Noah Brown, Michael Ahn, Omar Cortes, Nicolas Sievers, Clayton Tan, Sichun Xu, Diego Reyes, Jarek Rettinghouse, Jornell Quiambao, Peter Pastor, Linda Liuu, Kuang-Huei Lee, Yuheng Kuang, Sally Jesmonth, Nikhil J. Joshi, Kyle Jeffrey, Rosario Jauregui Ruano, Jasmine Hsu, Keerthana Gopalakrishnan, Byron David, Andy Zeng, and Chuyuan Kelly Fu. Do as I can, not as I say: Grounding language in robotic affordances. In CoRL, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[27] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "text", + "content": "[28] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. In NeurIPS, 2023. 2, 3, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "text", + "content": "[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Paul Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 151, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 260 + ], + "type": "text", + "content": "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuok Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. Videopoet: A large language model for zero-shot video generation. In ICML, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 262, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 327 + ], + "type": "text", + "content": "[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 329, + 553, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 553, + 372 + ], + "type": "text", + "content": "[32] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 374, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 407 + ], + "type": "text", + "content": "[33] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. AI choreographer: Music conditioned 3d dance generation with AIST++. In ICCV, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 408, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 441 + ], + "type": "text", + "content": "[34] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. IJCV, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 442, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 485 + ], + "type": "text", + "content": "[35] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "type": "text", + "content": "[36] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. VILA: on pre-training for visual language models. In CVPR, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 522, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 522, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 522, + 553, + 544 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 545, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 587 + ], + "type": "text", + "content": "[38] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 590, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 643 + ], + "type": "text", + "content": "[39] Arvind Neelakantan, Tao Xu, Raul Puri, Alec Radford, Jesse Michael Han, Jerry Tworek, Qiming Yuan, Nikolas Tezak, Jong Wook Kim, Chris Hallacy, et al. Text and code embeddings by contrastive pre-training. arXiv preprint arXiv:2201.10005, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 646, + 553, + 689 + ], + "type": "text", + "content": "[40] Evonne Ng, Javier Romero, Timur M. Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In CVPR, 2024. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 712 + ], + "type": "text", + "content": "[41] OpenAI. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26896" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[42] Joon Sung Park, Joseph C. O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "text", + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[45] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In KDD, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 320 + ], + "type": "text", + "content": "[46] Giuseppe Riva, Fabrizia Mantovani, Claret S. Capdeville, Alessandra Preziosa, Francesca Morganti, Daniela Villani, Andrea Gaggioli, Cristina Botella, and Mariano Alcaniz Raya. Affective interactions using virtual reality: The link between presence and emotions. Cyberpsychology Behav. Soc. Netw., 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 323, + 294, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 294, + 355 + ], + "type": "text", + "content": "[47] Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu. Character-llm: A trainable agent for role-playing. In EMNLP, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 357, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 357, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 294, + 401 + ], + "type": "text", + "content": "[48] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower GPT with off-policy reinforcement learning for dance accompaniment. In ICLR, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 403, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 403, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 403, + 294, + 435 + ], + "type": "text", + "content": "[49] Mel Slater, Daniel Pérez Marcos, Henrik Ehrsson, and Maria V Sanchez-Vives. Inducing illusory ownership of a virtual body. Frontiers in neuroscience, 2009. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[50] Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths. Cognitive architectures for language agents. 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 472, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 712 + ], + "type": "text", + "content": "[51] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurélien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[52] Quan Tu, Shilong Fan, Zihang Tian, Tianhao Shen, Shuo Shang, Xin Gao, and Rui Yan. Charactereval: A chinese benchmark for role-playing conversational agent evaluation. In ACL, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[53] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. TMLR, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "type": "text", + "content": "[54] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. TMLR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "text", + "content": "[55] Alexander W. Winkler, Jungdam Won, and Yuting Ye. Question: Human motion tracking from sparse sensors with simulated avatars. In SIGGRAPH Asia, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[56] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, Yunhui Liu, Wenjun Zeng, and Xiaokang Yang. Inter-x: Towards versatile human-human interaction analysis. In CVPR, 2024. 2, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "type": "text", + "content": "[57] Dongchao Yang, Jinchuan Tian, Xu Tan, Rongjie Huang, Songxiang Liu, Xuankai Chang, Jiatong Shi, Sheng Zhao, Jiang Bian, Xixin Wu, Zhou Zhao, and Helen Meng. Uniaudio: An audio foundation model toward universal audio generation. arXiv preprint arXiv:2310.00704, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "text", + "content": "[58] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, Wayne Wu, Chen Qian, Dahua Lin, Ziwei Liu, and Lei Yang. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In ICCV, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "type": "text", + "content": "[59] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. TASLP, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 471, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 536 + ], + "type": "text", + "content": "[60] Jun Zhan, Junqi Dai, Jiasheng Ye, Yunhua Zhou, Dong Zhang, Zhigeng Liu, Xin Zhang, Ruibin Yuan, Ge Zhang, Linyang Li, Hang Yan, Jie Fu, Tao Gui, Tianxiang Sun, YuGang Jiang, and Xipeng Qiu. Anygpt: Unified multimodal LLM with discrete sequence modeling. In ACL, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "text", + "content": "[61] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. In EMNLP, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[62] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[63] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech large language models. arXiv preprint arXiv:2308.16692, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[64] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, 2024. 2, 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26897" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 337 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[65] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding, planning, generation and beyond. In CVPR, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 149 + ], + "type": "text", + "content": "[66] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 337 + ], + "type": "text", + "content": "[67] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. 2, 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26898" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_content_list.json b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df5609ce444f41662ca541155c3f9e6d6b7f1263 --- /dev/null +++ b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_content_list.json @@ -0,0 +1,1361 @@ +[ + { + "type": "text", + "text": "SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving", + "text_level": 1, + "bbox": [ + 156, + 128, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuesong Chen $^{1,3,4*}$ Linjiang Huang $^{2*}$ Tao Ma $^{1}$ Rongyao Fang $^{1}$ Shaoshuai Shi $^{3\\dagger}$ Hongsheng Li $^{1,4\\dagger}$", + "bbox": [ + 99, + 200, + 957, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MMLab, CUHK $^{2}$ Institute of Artificial Intelligence, Beihang University $^{3}$ Voyager Research, Didi Chuxing $^{4}$ CPII under InnoHK", + "bbox": [ + 228, + 220, + 823, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenxuesong@link, hsli@ee}.cuhk.edu.hk, ljhuang@buaa.edu.cn, shaoshuaics@gmail.com", + "bbox": [ + 156, + 258, + 890, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The integration of Vision-Language Models (VLMs) into autonomous driving systems has shown promise in addressing key challenges such as learning complexity, interpretability, and common-sense reasoning. However, existing approaches often struggle with efficient integration and real-time decision-making due to computational demands. In this paper, we introduce SOLVE, an innovative framework that synergizes VLMs with end-to-end (E2E) models to enhance autonomous vehicle planning. Our approach emphasizes knowledge sharing at the feature level through a shared visual encoder, enabling comprehensive interaction between VLM and E2E components. We propose a Trajectory Chain-of-Thought (T-CoT) paradigm, which progressively refines trajectory predictions, reducing uncertainty and improving accuracy. By employing a temporal decoupling strategy, SOLVE achieves efficient cooperation by aligning high-quality VLM outputs with E2E real-time performance. Evaluated on the nuScenes dataset, our method demonstrates significant improvements in trajectory prediction accuracy, paving the way for more robust and reliable autonomous driving systems.", + "bbox": [ + 88, + 340, + 483, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 686, + 220, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past two decades, autonomous driving has always been a key research focus, with the potential to significantly revolutionize transportation [25, 32, 41]. Thanks to rapid advances in deep learning for perception [22, 26, 34, 38], prediction [14, 29, 31], and planning [4, 11, 16, 37, 40], end-to-end autonomous driving [8, 16-19, 49] has gained prominence, where the entire driving process—from perception to planning—is learned within a unified framework. Typically, these methods predict future trajectories or control signals directly from sensor inputs, bypassing distinct", + "bbox": [ + 89, + 713, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7085294ccfd9e8c5bf0efd74d3199a3b2b6ff1a65a454429a3179b3bafca37fb.jpg", + "image_caption": [ + "Figure 1. Previous methods combine VLM and end-to-end networks through post-processing, while our method combines VLM and end-to-end networks through both feature-level synergy (shared visual encoder) and trajectory-level synergy." + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 916, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "decision-making steps. Through extensive data training, end-to-end approaches have demonstrated impressive planning capabilities, providing a streamlined and competitive alternative to traditional modular pipelines [5, 21, 33].", + "bbox": [ + 509, + 520, + 906, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end autonomous driving, while promising, still faces significant challenges, such as learning complexity [7], insufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. For instance, current approaches may misinterpret trucks carrying traffic cones as roadblocks or mistake billboard images of vehicles for real obstacles, resulting in unnecessary braking. Recent advancements in large Vision-Language Models (VLMs) [1, 23, 28] offer a potential solution to these challenges. VLMs, through extensive training on diverse datasets, can simplify learning complexity by providing a more intuitive understanding of various scenarios. Additionally, the logical and commonsense knowledge provided by VLMs enhances autonomous driving systems, allowing them to better interpret driving environments and make safer decisions in complex situations. However, VLMs still face limitations in areas such as efficiency and 3D perception, which pose challenges for their integration into an end-to-end autonomous driving framework.", + "bbox": [ + 509, + 582, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some initial explorations have been made to leverage VLMs within the realm of autonomous driving. For exam-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ple, DriveVLM [41] introduces a dual-system architecture that integrates the classic perception-prediction-planning (3P) pipeline with VLMs. As shown in Figure 1(a), the VLM and the end-to-end (E2E) model operate independently and cooperate at the trajectory level, e.g., using the E2E planner to refine the trajectory predictions made by the VLM. However, this strategy limits information and knowledge sharing between the VLM and the E2E model. For example, the E2E model cannot leverage the high-level reasoning abilities of VLM to pay more attention to critical objects or understand their intents, which is crucial for collision avoidance. Similarly, the VLM cannot integrate the 3D perception capabilities of the E2E model for precise scene comprehension, which is essential for accurate planning.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and planning integration. As Figure 1 shows, besides the trajectory-level cooperation, which is already explored in previous approaches [41], our method enhances feature-level knowledge sharing through a shared vision encoder, facilitating comprehensive collaboration between the VLM and the E2E model. Leveraging the Q-Former design, the shared encoder efficiently compresses surrounding foreground objects (e.g., cars, pedestrians), background maps (e.g., lane centerlines), and environmental cues (e.g., traffic lights, weather, time of day) into unified visual tokens, reducing redundant feature extraction. Consequently, the Q-Former is designed to operate feature collection and alignment with various perception queries sequentially. This structural alignment enhances consistency across vision perception and control domains, significantly facilitating knowledge sharing and easing the difficulties of scene understanding.", + "bbox": [ + 91, + 305, + 482, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To boost precision and efficiency in trajectory prediction, we meticulously optimize the VLM and the E2E model. To address the challenges of VLM in directly generating fine-grained trajectories in an auto-regressive manner, we introduce the Trajectory Chain-of-Thought (T-CoT). This new paradigm progressively refines trajectories by using a predefined trajectory bank and chain-style reasoning. Initially, a trajectory adapter transforms each candidate trajectory from the trajectory bank into trajectory tokens that align with text and visual tokens, enabling the network to select the most suitable trajectory. Subsequently, the selected trajectory serves as a foundation for the next round of refinement, ultimately generating the final trajectory.", + "bbox": [ + 89, + 595, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given that VLM can produce high-quality trajectories but at a significant computational cost, we propose a temporal decoupling strategy to enable cooperation between the VLM and the E2E model. Unlike previous methods [9], which perform feature-level asynchronization, our approach operates at the trajectory level, specifically aiming to enhance trajectory accuracy. The framework implements a", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "memory-based mechanism where VLM-generated trajectories are stored and subsequently accessed by the E2E network as initialization priors, accommodating a temporal offset from the current state. This approach leverages priors of VLM to asynchronously enhance E2E network performance while maintaining real-time system operation capabilities.", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold:", + "bbox": [ + 529, + 196, + 831, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and trajectory integration.", + "- We propose the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories through a predefined trajectory bank and chain-style reasoning.", + "- We demonstrate the effectiveness of our proposed approach through extensive experiments, and our framework achieves state-of-the-art results for open-loop planning on the nuScenes benchmark." + ], + "bbox": [ + 511, + 213, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 378, + 653, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-end Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 404, + 802, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the early stage, autonomous driving systems often follow a modular design and are equipped with rule-based planners. This design hurdles the system's generalization ability and leads to limited performance. In contrast, end-to-end driving is a fully differentiable system from sensors to control signals, optimizing the entire system to directly map perception inputs to planning outputs. Some early works primarily rely on data-driven approaches to train neural networks for end-to-end autonomous driving. These methods show promise, despite challenges with interpretability. The growth in driving data further promotes end-to-end autonomous driving, with models like UniAD [16] and PARA-Drive [47] using Transformers for multi-task learning like object tracking and planning, enhancing performance through additional supervision. Furthermore, VAD [18] introduces vectorized scene representation for better accuracy and speed, with VAD- $\\nu 2$ [8] further improving planning by capturing uncertainty. Besides, methods such as AD-MLP [49] employ a simple multilayer perceptron (MLP) that relies solely on ego vehicle status information, highlighting the potential over-reliance on ego status in simpler driving scenarios while also revealing limitations in utilizing perception data effectively. BEV-Planner [27] integrates various perception inputs in a bird's-eye view (BEV) format, aiming to enhance planning capabilities in more complex traffic situations by leveraging richer environmental context.", + "bbox": [ + 511, + 428, + 906, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. VLMs for Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 845, + 790, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, end-to-end autonomous driving often suffers from limitations, such as learning complexity [7], in", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. There is an urgent need to improve these systems for real-world applications. Integrating Vision-Language Models (VLMs) into autonomous driving systems has gained significant attention due to their potential to enhance explainability, reasoning, and generalization in driving tasks. For instance, models like DriveGPT4 [48] utilize VLMs to predict control signals while simultaneously providing explanations for vehicle actions through an iterative question-and-answer format. Similarly, Drive Anywhere [46] employs patch-aligned feature extraction from multi-modal large language models (MLLMs) to facilitate text-based driving decision-making, while OmniDrive [44] introduces a 3D vision-language model architecture aimed at improving reasoning and planning capabilities. Other innovative approaches, such as DriveLM [36], leverage graph-based visual question-answering (VQA) techniques. In a complementary manner, Drive-with-LLMs [6] encodes perception data into latent space using a Transformer network, which is then processed by an LLM to forecast future trajectories. However, most of these VLM-based approaches require a large computation cost, which cannot achieve real-time decision making. DriveVLM [41] stands out as the first model to merge VLMs with end-to-end architectures, predicting low-frequency trajectories through chain-of-thought reasoning that are refined by an end-to-end model to produce final planning outputs. Despite its promising performance, it only fosters cooperation at the trajectory level, overlooking the high-level complementarity between the VLM and the end-to-end model.", + "bbox": [ + 89, + 90, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SOLVE", + "text_level": 1, + "bbox": [ + 89, + 571, + 181, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a detailed introduction to SOLVE. Figure 2 illustrates the overall architecture. We utilize a Sequential Q-Former (SQ-Former) to obtain rich and condensed visual cues from sequences image input (Section 3.1). These visual cues are shared between the Vision-Language Model (VLM) and the end-to-end (E2E) model, enabling efficient feature-level knowledge integration. Additionally, we introduce a Trajectory Chain-of-Thought (T-CoT) mechanism (Section 3.2) that refines trajectory predictions from coarse to fine through chain-style reasoning. By aligning the trajectory predictions of the VLM with E2E real-time processing in a time-decoupled cooperation manner (Section 3.3), SOLVE achieves the balance between accuracy and computational efficiency. Finally, we introduce our multi-stage training strategy (Section 3.4) to further promote feature-level synergy and trajectory prediction.", + "bbox": [ + 89, + 597, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SQ-Former", + "text_level": 1, + "bbox": [ + 89, + 847, + 217, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual cues play a pivotal role in autonomous driving, offering foreground information, scene map information, and", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "environmental cues. These elements help the system to accurately predict vehicle behavior and plan safe routes, enhancing overall decision-making and navigation capabilities. However, abundant visual information can also introduce interference in autonomous driving, such as irrelevant background clutter or reflections. Effectively filtering and extracting useful visual information ensures accuracy and efficiency for autonomous driving.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To harness the relevant information from observations, we categorize useful visual cues into three groups, as shown in Figure 3. The first group includes scene-level static or holistic environmental details, such as weather, time of day, and traffic conditions. The second group concentrates on dynamic participants on the road, such as vehicles and pedestrians. The third group contains dynamic map cues, e.g., lanes. Here, we present Sequential Q-Former (SQ-Former), a Q-Former-style architecture that sequentially compresses these groups of visual cues into a fixed number of condensed visual queries, aligning perception with reasoning and planning. Specifically, the SQ-Former is built upon an image encoder, mapping the multi-view images into high-level features $\\mathbf{F}_i\\in \\mathbb{R}^{N\\times C\\times H\\times W}$ , where $N,C,H,W$ denotes the number of views, channels, height and width, respectively. The extracted features are forwarded into the SQ-Former to progressively collect both static and dynamic visual cues from observations. Inspired by StreamPETR [43], we begin with a set of collector queries $\\mathbf{Q}_c$ . These queries are input into a transformer decoder to capture static cues of the first group, which can be formulated as:", + "bbox": [ + 511, + 212, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {c} ^ {(s)} = f _ {s} \\left(\\mathbf {Q} _ {c}, \\mathbf {F} _ {i} + \\mathbf {P} _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 553, + 903, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_i$ denotes the 3D position embedding. $f_{s}$ stands for the transformer decoder, which includes interleaved selfattention and cross-attention layers to promote interaction between queries and extract informative static cues, respectively. Afterward, the updated collector queries $\\mathbf{Q}_c^{(s)}$ are concatenated with task-specific queries $\\mathbf{Q}_{\\triangle}$ , where $\\triangle$ represents different tasks, such as object detection (d) or map understanding (m). These queries are then fed into separate transformer decoders. In SQ-Former, we sequentially conduct this process, following a pre-defined order of perception tasks, which can be written as:", + "bbox": [ + 511, + 582, + 905, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Q} _ {c} ^ {(s, \\boldsymbol {\\Delta} [ 1: n ])}, \\tilde {\\mathbf {Q}} _ {\\boldsymbol {\\Delta} [ n ]} \\\\ = f _ {\\triangle [ n ]} \\left(\\left[ \\mathbf {Q} _ {c} ^ {(s, \\triangle [ 1: n - 1 ])}, \\mathbf {Q} _ {\\triangle [ n ]} \\right], \\mathbf {F} _ {i} + \\mathbf {P} _ {i}, \\mathcal {M}\\right), \\quad (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 758, + 906, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\cdot]$ denotes the concatenation operation, and $f_{\\pmb{\\Delta}[n]}$ is the transformer decoder specific to the $n$ -th task. We introduce the memory bank $\\mathcal{M}$ , following StreamPETR [43], to incorporate the temporal dynamics of the driving scenario. In this framework, $\\mathbf{Q}c^{(s,\\pmb{\\Delta}[1:n])}$ and $\\tilde{\\mathbf{Q}}\\pmb{\\Delta}[n]$ represent the collector queries enriched with visual cues for tasks 1", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cea31da02d1e06d259b2065d5304f8387f9aa714f16856a9f05663325d51f365.jpg", + "image_caption": [ + "Figure 2. The overall framework of the proposed SOLVE." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 903, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2a259d47e4408167df5eafec99acdfb849796ecbba7a8fb0c19514f5647adc5.jpg", + "image_caption": [ + "Figure 3. The detail of the proposed SQ-Former. We first capture the static cues from multi-view images and then sequentially align the model with different perception tasks." + ], + "image_footnote": [], + "bbox": [ + 96, + 333, + 478, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "through $n$ , and the updated task-specific queries, respectively. These updated queries are then used to align with the $n$ -th task. The collector queries are sent to a vision adapter to match the dimensions of the VLM, with an ego token appended for trajectory prediction.", + "bbox": [ + 89, + 564, + 482, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to our experiments, the sequential design ensures that SQ-Former captures rich, condensed visual cues from various perception tasks, enhancing synergy between them. For example, compared to OmniDrive [44] that uses 512 queries, our method achieves superior performance with only 384 queries, significantly reducing the computational cost of the following VLM.", + "bbox": [ + 89, + 641, + 482, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Trajectory Chain-of-Thought", + "text_level": 1, + "bbox": [ + 89, + 756, + 354, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AD-MLP [49] has demonstrated that even a simple planner can achieve moderate results using only ego status. However, it lacks the ability to reason and integrate logical knowledge, which is crucial for a safe and reliable autonomous driving system. To address this limitation, we introduce VLM to enhance the system's reasoning abilities.", + "bbox": [ + 89, + 779, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we provide the VLM with visual and ego inputs, along with high-quality question-answering (QA)", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pairs that encompass perception, reasoning, and planning [44]. Unlike previous methods, we meticulously develop a Trajectory Chain-of-Thought (T-CoT) process for the planning phase, employing a coarse-to-fine procedure for trajectory prediction. Notably, there are several methods, such as DriveCoT [45], DriveLM [36] and DriveVLM [41], have demonstrated the effectiveness of CoT in autonomous driving. However, their reasoning chain spans from perception to planning, requiring multiple rounds of reasoning and lengthy textual output, resulting in significant computational costs. In contrast, our trajectory CoT operates at the trajectory level, refining trajectories in several steps and requiring minimal output.", + "bbox": [ + 511, + 335, + 906, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make the T-CoT work at the trajectory level, we first construct a trajectory bank containing 36 candidate trajectories (with corresponding ego status and history trajectory) by clustering training set trajectories using the $k$ -means algorithm for each navigation command (forward, left-turn and right-turn). Next, we utilize the current ego status and historical trajectory as a feature vector, calculate its similarity with the corresponding feature vectors to the clustered trajectories, and then use a top-k operation to select the $k_{l}$ trajectories that are closest to the current feature vector. These are combined with a trajectory predicted by an MLP, which uses only ego information and historical data as input [49], resulting in $k_{l} + 1$ candidate trajectories. We introduce clustered trajectories because, although the coarse trajectories are already acceptable, some predicted trajectories still deviate significantly from the ground truth. Retrieving clustered trajectories provides a pull-back mechanism to correct these inaccuracies. With the trajectory bank, the VLM is tasked with progressively outputting predictions in two stages: trajectory selection and trajectory refinement.", + "bbox": [ + 511, + 534, + 908, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory selection prompts the VLM to choose the most accurate trajectory from the candidates, based on ego and environmental information. To reduce computational costs, the endpoint of each candidate trajectory is", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3730a3439857b47253359c5ac5da0cae53d6f4bd1266c3cf3f96301e648f9d15.jpg", + "image_caption": [ + "Figure 4. The illustration of combination of the proposed trajectory tokens with image tokens and text tokens for the large language model-based planning." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 472, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoded into a single token through the proposed trajectory adapter and inserted into the textual prompt as follows: Here are predefined trajectories $\\langle \\text{Traj}_1 \\rangle$ , $\\langle \\text{Traj}_2 \\rangle$ , ..., $\\langle \\text{Traj}_{k_l + 1} \\rangle$ . Please select the best trajectory for the ego car according to the current scenario., as shown in Figure 4.", + "bbox": [ + 89, + 339, + 482, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trajectory refinement involves generating a refined trajectory after selecting the reference trajectory. The reference trajectory consists of waypoints, denoted by $W = \\{w_{1}, w_{2}, \\ldots, w_{n}\\}$ , where $w_{i} = (x_{i}, y_{i})$ , representing the vehicle's path over a future period with predetermined intervals. Each waypoint of the reference trajectory is encoded into a waypoint token and injected into the textual prompt as follows: With the selected trajectory as a reference $<\\text{Point}_1>, <\\text{Point}_2>, \\ldots, <\\text{Point}_n>$ , please provide the planning trajectory for the ego car, which has a velocity of $\\text{SPEED} \\, m/s$ and an acceleration of $\\text{ACCEL} \\, m/s^2$ .", + "bbox": [ + 88, + 415, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Time-decoupled Synergy Prediction", + "text_level": 1, + "bbox": [ + 89, + 590, + 403, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our planning-oriented synergy system operates at both feature-level (shared SQ-Former feature) and the trajectory-level, aiming to further enhance final trajectory accuracy. This strategy combines the advanced capabilities of VLMs in generating high-accuracy trajectories with the efficiency of a streamlined E2E model.", + "bbox": [ + 89, + 613, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we deploy a simple planner that incorporates cross-attention mechanisms between multiple planning queries and collector queries and then it employs an MLP to project these queries into the final output trajectories. The initialization of the planning queries are performed by encoding position embedding from clustering trajectories from the training set, following SparseDrive [39]. The incorporation between the SQ-Former and the simple planner constructs an E2E model, a pipeline frequently adopted in prior systems.", + "bbox": [ + 89, + 704, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upon generating a trajectory from the VLM, we put it into memory. The end-to-end network can obtain the latest VLM predicted trajectory from the memory (with a de", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lay from the current time) as an additional trajectory initialization. Specifically, we task the VLM with predicting trajectories over extended time horizons (longer than both the VLM's latency and the prediction horizon required by the end-to-end model). Consequently, despite the VLM's latency, the E2E model can still asynchronously leverage the VLM's long-term trajectories to obtain additional high-quality trajectory initialization and get the final collaborative predicted trajectory.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Strategy", + "text_level": 1, + "bbox": [ + 511, + 237, + 687, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a multi-stage training strategy for SOLVE. In the first stage, we employ the same QA pairs as proposed by OmniDrive [44], prompting the VLM by incorporating LoRA layers [15] to address inquiries related to scene description, perception, prediction, and planning. Through QA training, the collector queries in the visual encoder are optimized to comprehensively encode vital scene information while maintaining efficiency.", + "bbox": [ + 511, + 261, + 906, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we train a trajectory adapter to ensure that it can encode predefined trajectories into a feature space compatible with the VLM. We then train the end-to-end planning head while freezing the previously trained SQ-Former, vision adapter, and trajectory adapter.", + "bbox": [ + 511, + 382, + 905, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the subsequent phase, leveraging these pre-trained weights, we jointly optimize the VLM and end-to-end planning head without freezing any components. During this phase, we employ the proposed T-CoT to augment the reasoning capabilities of the VLM.", + "bbox": [ + 511, + 458, + 905, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, our trajectory-level synergy strategy involves exclusively refining the end-to-end planner by leveraging the trajectory forecasted by the VLM in historical frames as an additional initialized planning query for the current frame.", + "bbox": [ + 511, + 534, + 905, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 609, + 645, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 635, + 823, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NuScenes Dataset. We utilize the nuScenes [3] dataset to train and evaluate our method. This comprehensive, large-scale driving dataset is tailored for urban environments, featuring 1,000 driving scenes, each approximately 20 seconds long. It encompasses diverse sensor data, including surround-view camera images and LiDAR point clouds, along with extensive annotations such as 3D bounding boxes, ego vehicle trajectories, and road maps. For training of VLM, we employ the OmniDrive-nuScenes [44] dataset, a planning-oriented extension of the nuScenes dataset with additional annotations, including perception, prediction, and planning-related QA. Specifically, the dataset features two distinct QA paradigms: Offline QA, which focuses on scene-level reasoning through tasks such as scene description, attention analysis, and decision validation using language models; and Online QA, which emphasizes 3D spa", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12072", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1a91509e34e68bd1725fe84715297430f7ff26f2bae1b5c8c9b020ffe609cf8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEgo StatusL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
UniAD [16]-0.591.011.481.030.160.511.640.77
UniAD [16]✓0.200.420.750.460.020.250.840.37
VAD-Base [18]-0.691.221.831.250.060.682.521.09
VAD-Base [18]✓0.170.340.600.370.040.270.670.33
AD-MLP [49]-0.150.320.590.350.000.270.850.37
BEV-Planner [27]-0.300.520.830.550.100.371.300.59
BEV-Planner++ [27]✓0.160.320.570.350.000.290.730.34
DriveVLM [41]✓0.180.340.680.40----
OmniDrive [44]✓0.140.290.550.330.000.130.780.30
EMMA [17]✓0.140.290.540.32----
DriveVLM-Dual [41]✓0.150.290.480.31----
Ours-E2E✓0.140.280.500.310.040.170.680.30
Ours-E2E (Async)✓0.130.270.500.300.040.170.650.29
Ours-VLM✓0.130.250.470.280.000.160.430.20
", + "bbox": [ + 209, + 88, + 792, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison of various methods on the NuScenes dataset for open-loop planning. Methods are evaluated with or without ego status consideration. The proposed methods, particularly SOLVE-VLM, demonstrate superior accuracy and safety, achieving the lowest average L2 error and collision rates.", + "bbox": [ + 89, + 352, + 906, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tial understanding through tasks like 2D-to-3D grounding, proximity analysis, and lane-based object detection using geometric labels.", + "bbox": [ + 89, + 410, + 480, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous works [27], we employ displacement error (L2 distance) and collision rate as the primary metrics.", + "bbox": [ + 89, + 455, + 482, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use LLaVA v1.5 [28], whose component of language model is based on LLaMA-7B, as our default large vision-language model, demonstrating exceptional performance in tasks such as question answering, visual localization, and text recognition. We replace its vision encoder with our SQ-Former. For the SQ-Former, we employ a predefined task order of image $\\rightarrow$ 3D detection $\\rightarrow$ lane detection. The image encoder is initialized using EVA02-L [12]. For the detection decoder and lane decoder, we adopt the architecture from StreamPETR [43]. The numbers of object queries, lane queries, and the sequential collector queries are set to 900, 300, and 384, respectively.", + "bbox": [ + 88, + 502, + 482, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 695, + 454, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, SOLVE-E2E and SOLVE-VLM achieved SOTA results in their respective tracks. Specifically, among end-to-end networks, SOLVE-E2E surpassed UniAD [16], VAD [18], and BEV-planner [27] by 0.15, 0.06, and 0.04, respectively, in terms of L2 (m) error, which shows that SOLVE can better imitate the driving behavior of human experts. On the other hand, SOLVE achieves $0.07\\%$ , $0.03\\%$ , and $0.07\\%$ lower collision rates than UniAD, VAD, and BEV-planner, respectively, which shows that it has more accurate spatial judgment ability than others and can avoid other objects as much as possible when planning trajectories. It is worth noting that SOLVE-E2E achieved sim", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ilar performance to DriveVLM-dual. This proves the superiority of the joint training of SOLVE's shared visual encoder. Specifically, other end-to-end methods only train visual encoders through detection or segmentation tasks, but lack scene understanding capabilities and may miss important visual cues during the encoding process. In contrast, we use VLM's comprehensive QA task, including perception, prediction, and planning, to ensure that the visual encoder encodes important features in the input signal, such as key objects that need to be avoided, while ignoring some information that is not important for planning, such as some unimportant objects behind the vehicle.", + "bbox": [ + 511, + 410, + 906, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other hand, SOLVE-VLM achieves the highest performance compared to other VLM-based methods, surpassing OmniDrive [44] and DriveVLM [41] by 0.05 and 0.03 in terms of L2 error, using large language models of similar capacities (DriveVLM uses QwenVL-7.7B [2], OmniDrive and our SOLVE-VLM use LLaVA1.5-7B [42]). This experiment proves that 1) our proposed SQ-former can effectively compress visual features and better release the reasoning ability of the language model. 2) Our joint training scheme of sharing SQ-Former enables VLM to obtain effective auxiliary supervision from the end-to-end model branch to improve performance. Finally, using the future trajectory predicted by VLM in the historical frame as the initialization of the current moment, SOLVE-E2E (Async) achieves better performance than SOLVE-E2E, with an improvement of 0.01 in terms of L2 (m) error.", + "bbox": [ + 511, + 592, + 908, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 845, + 694, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show two planning results of SOLVE in Fig. 5. To a better understanding of the scene, we also provide input", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12073", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a558958e0504eaa43db12cda6bf2686b3c688b4b8afc77e03b2d5c76ac6aa72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 89, + 679, + 593 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96e782d58b3a6488326ab8d89c3fa59314576268027c0db2d77695ff8f347b35.jpg", + "image_caption": [ + "Figure 5. Qualitative results of SOLVE, where red lines, blue lines and yellow lines mean the planning results from VLM, E2E-Async and E2E modules. The ego car (green box) and ground truth trajectory (green line) are shown in the right BEV images." + ], + "image_footnote": [], + "bbox": [ + 683, + 88, + 893, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "surrounding camera images and project the planning trajectories to the front camera image. As shown in the figure, for simple straight-ahead scenarios (see first row), both VLM and the E2E model can predict trajectories that match the expert trajectory. However, for more complex scenarios, the trajectory predicted by the end-to-end model may deviate from the center line of the lane (see yellow line in second row). At the same time, VLM can better understand the scenario and keep the trajectory in the middle of the lane. Besides, with the help of VLM-predicted trajectories, E2E-Async can predict better trajectories than E2E networks.", + "bbox": [ + 89, + 632, + 485, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 814, + 254, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies to verify the effectiveness of each component in SOLVE. Unless otherwise mentioned, we train both the SQ-Former and VLM for 6 epochs to optimize computational resources while main", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "taining experimental rigor. We take L2 (cm) as the default metric for comparison.", + "bbox": [ + 511, + 632, + 906, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of query number of SQ-former. Table 2 displays the VLM performances of using different numbers of SQ-former queries to encode visual information. First, too few collector query numbers (i.e. 256) can only represent limited features, which may easily lead to the loss of visual features, resulting in the lowest performance. Second, too many collector queries (i.e. 512) may introduce redundant information while increasing the capacity of the SQ-former. It will also increase the amount of computation during VLM training and inference. Finally, we choose 384 queries as our default option to balance performance and computation.", + "bbox": [ + 511, + 667, + 908, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different designs of visual information encoding in SQ-Former. Table 3 investigates the impact of different visual information encoding strategies for VLM. Q means the initial learnable query and the arrows indicate", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12074", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d1060f682d7b3cf56e3c619a718893bc4f6b98dbc91cfd12eafff2f6a0894452.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of SQ-Former's queriesL2 (cm) ↓
1s2s3sAvg
25614.129.254.532.6
38414.129.054.032.3
51214.229.254.432.6
", + "bbox": [ + 96, + 89, + 480, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d27278cf4f0360757aeafdc1884fa5c67605419dddce59f1ce4bdb165d7c0f5c.jpg", + "table_caption": [ + "Table 2. Effects of the number of queries of SQ-Former to encode temporal vision information." + ], + "table_footnote": [], + "table_body": "
MethodL2 (cm) ↓
1s2s3sAvg
Q → Det → Lane14.329.555.033.0
Q → Img → Det → Lane14.129.054.032.3
Q → Img → Lane → Det14.229.154.432.6
", + "bbox": [ + 114, + 209, + 464, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the order in which SQ-Former encodes different elements, including image features, detection query, and lane query. Comparing the first and second rows of the table, we can find that incorporating the encoding of the overall image features can achieve a 0.7 performance improvement of L2, which proves that some scene-level clues contained in the image features, such as road conditions and weather, are beneficial to VLM's trajectory planning. Comparing the second and third rows of the table, we find that the encoding order of different elements also affects the performance. Encoding the detection query first and then the lane query can achieve a 0.3 L2 improvement compared to the reverse order. We use the design in the second row as our choice.", + "bbox": [ + 88, + 354, + 482, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of trajectory CoT for VLM planning. Table 4 investigates the impact of the proposed trajectory CoT for VLM. Results demonstrate that direct trajectory prediction without CoT yields suboptimal performance, highlighting the inherent challenges VLMs face in auto-regressive numerical prediction tasks. The integration of trajectory CoT, through reference trajectories and trajectory tokens, provides crucial location priors, reducing the uncertainty of direct waypoint prediction and yielding a 1.1 performance improvement. Our analysis of reference trajectory quantity reveals that insufficient samples (e.g., 4) constrain trajectory diversity, though still achieving a 0.9 improvement over the baseline. Conversely, excessive samples (e.g., 8), while providing more optional initial references, increase both trajectory selection complexity and computational overhead. Based on this trade-off analysis, we empirically determine 6 reference trajectories as the optimal configuration, balancing performance and computational efficiency.", + "bbox": [ + 89, + 551, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of sharing SQ-Former. Table 5 shows the impact of shared SQ-Former features on the performance of the E2E planner and VLM planner. The results indicate that feature sharing and joint training protocols significantly outperform independently trained networks. The E2E net", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3c23ad112c6ccbd66b62884e1f51868beb848c4f51d3e98f146d09b049c2c7ec.jpg", + "table_caption": [ + "Table 3. Different designs of visual feature encoding in SQ-Former, where Q means the initial learnable query and $\\rightarrow$ means the order of cross attention." + ], + "table_footnote": [], + "table_body": "
MethodTrajectory numberL2 (cm) ↓
1s2s3sAvg
w/o CoT-13.626.947.829.5
CoT4 (3+1)12.625.747.328.6
6 (5+1)12.525.647.128.4
8 (7+1)12.525.647.128.4
", + "bbox": [ + 547, + 89, + 870, + 179 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c15bbd3e2daedc5e6295a1e0a98c8c352d48c5368892b9e8039615db6b39a229.jpg", + "table_caption": [ + "Table 4. Effects of the proposed T-CoT. The trajectory number means the selected top-k and 1 predicted trajectories." + ], + "table_footnote": [], + "table_body": "
MethodSQ-FormerL2 (cm) ↓
1s2s3sAvg
E2Ew/o sharing14.929.653.232.5
sharing14.228.150.531.0
VLMw/o sharing12.826.048.329.0
sharing12.625.547.328.4
", + "bbox": [ + 549, + 224, + 874, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Effects of sharing SQ-Former features to the E2E module and the VLM module.", + "bbox": [ + 511, + 325, + 903, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work exhibits a $1.5\\mathrm{cm}$ reduction in L2 error when utilizing shared SQ-Former features, suggesting that question-answering supervision enables more effective feature extraction for planning compared to conventional detection and segmentation supervision alone. Similarly, VLM performance improves by 0.6 through E2E network integration. This contrasts with DriveVLM's approach of trajectory initialization through post-processing, which shows no significant improvement in VLM's planning capabilities.", + "bbox": [ + 511, + 380, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 530, + 633, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce SOLVE, an innovative framework that fosters synergy between the Vision-Language Model (VLM) and the end-to-end (E2E) model, emphasizing both knowledge and planning integration. To tackle the challenges faced by VLMs in directly generating fine-grained trajectories in an auto-regressive manner, we present the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories using a predefined trajectory bank and chain-style reasoning. Additionally, we propose a temporal decoupling strategy to facilitate cooperation between the VLM and the E2E model. Experiments on the nuScenes dataset demonstrate that our approach achieves state-of-the-art results.", + "bbox": [ + 511, + 554, + 906, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 763, + 692, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 511, + 787, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1", + "[2] J Bai, S Bai, S Yang, S Wang, S Tan, P Wang, J Lin, C Zhou, and J Qwen-VL Zhou. A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 5", + "[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 1", + "[5] Changfang Chen, Yingmin Jia, Minglei Shu, and Yinglong Wang. Hierarchical adaptive path-tracking control for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 16(5):2900-2912, 2015. 1", + "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, pages 14093-14100. IEEE, 2024. 3", + "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE TPAMI, 2024. 1, 2", + "[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2", + "[9] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, pages 22-38. Springer, 2025. 2", + "[10] Pim De Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. NeurIPS, 32, 2019. 1, 3", + "[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, pages 9710-9719, 2021. 1", + "[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 6", + "[13] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In WACV, pages 910-919, 2024. 1, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In ICCV, pages 15303-15312, 2021. 1", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, pages 17853-17862, 2023. 1, 2, 6", + "[17] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 6", + "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 2, 6", + "[19] Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 1", + "[20] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In ECCV, pages 370-387. Springer, 2022. 1, 3", + "[21] Felix Kunz, Dominik Nuss, Jürgen Wiest, Hendrik Deusch, Stephan Reuter, Franz Gritschneder, Alexander Scheel, Manuel Stübler, Martin Bach, Patrick Hatzelmann, et al. Autonomous driving at ulm university: A modular, robust, and sensor-independent fusion approach. In IEEE Intelligent Vehicles Symposium, pages 666-673. IEEE, 2015. 1", + "[22] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023. 1", + "[24] Jiankun Li, Hao Li, Jiangjiang Liu, Zhikang Zou, Xiaqing Ye, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Exploring the causality of end-to-end autonomous driving. arXiv preprint arXiv:2407.06546, 2024. 1, 3", + "[25] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1", + "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 1", + "[27] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, pages 14864-14873, 2024. 2, 6", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 6", + "[29] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, pages 7577-7586, 2021. 1", + "[30] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 1, 3", + "[31] Sajjad Mozaffari, Omar Y Al-Jarrah, Mehrdad Dianati, Paul Jennings, and Alexandros Mouzakitis. Deep learning-based vehicle behavior prediction for autonomous driving applications: A review. IEEE Transactions on Intelligent Transportation Systems, 23(1):33-47, 2020. 1", + "[32] Khan Muhammad, Amin Ullah, Jaime Lloret, Javier Del Ser, and Victor Hugo C de Albuquerque. Deep learning for safe autonomous driving: Current challenges and future directions. IEEE Transactions on Intelligent Transportation Systems, 22(7):4316-4336, 2020. 1", + "[33] Samyeul Noh and Kyounghwan An. Decision-making framework for automated driving in highway environments. IEEE Transactions on Intelligent Transportation Systems, 19 (1):58-71, 2017. 1", + "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1", + "[35] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 3", + "[36] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3, 4", + "[37] Haoran Song, Wenchao Ding, Yuxuan Chen, Shaojie Shen, Michael Yu Wang, and Qifeng Chen. Pip: Planning-informed trajectory prediction for autonomous driving. In ECCV, pages 598-614. Springer, 2020. 1", + "[38] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 1", + "[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 5", + "[40] Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Yunfeng Ai, Dongsheng Yang, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, et al. Motion planning for autonomous driving: The state of the art and future perspectives. IEEE Transactions on Intelligent Vehicles, 8(6):3692-3711, 2023. 1" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 4, 6", + "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6", + "[43] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In ICCV, pages 3621-3631, 2023. 3, 6", + "[44] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3, 4, 5, 6", + "[45] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 4", + "[46] Tsun-Hsuan Wang, Alaa Maalouf, Wei Xiao, Yutong Ban, Alexander Amini, Guy Rosman, Sertac Karaman, and Daniela Rus. Drive anywhere: Generalizable end-to-end autonomous driving with multi-modal foundation models. In ICRA, pages 6687-6694. IEEE, 2024. 3", + "[47] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 2", + "[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3", + "[49] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1, 2, 4, 6" + ], + "bbox": [ + 514, + 92, + 903, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12077", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_model.json b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6bd8fd6089c3812750efbd403410056bf002eb07 --- /dev/null +++ b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_model.json @@ -0,0 +1,1936 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.13, + 0.842, + 0.177 + ], + "angle": 0, + "content": "SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.958, + 0.222 + ], + "angle": 0, + "content": "Xuesong Chen\\(^{1,3,4*}\\) Linjiang Huang\\(^{2*}\\) Tao Ma\\(^{1}\\) Rongyao Fang\\(^{1}\\) Shaoshuai Shi\\(^{3\\dagger}\\) Hongsheng Li\\(^{1,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.222, + 0.825, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)MMLab, CUHK \\(^{2}\\) Institute of Artificial Intelligence, Beihang University \\(^{3}\\) Voyager Research, Didi Chuxing \\(^{4}\\) CPII under InnoHK" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.26, + 0.892, + 0.275 + ], + "angle": 0, + "content": "{chenxuesong@link, hsli@ee}.cuhk.edu.hk, ljhuang@buaa.edu.cn, shaoshuaics@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.485, + 0.659 + ], + "angle": 0, + "content": "The integration of Vision-Language Models (VLMs) into autonomous driving systems has shown promise in addressing key challenges such as learning complexity, interpretability, and common-sense reasoning. However, existing approaches often struggle with efficient integration and real-time decision-making due to computational demands. In this paper, we introduce SOLVE, an innovative framework that synergizes VLMs with end-to-end (E2E) models to enhance autonomous vehicle planning. Our approach emphasizes knowledge sharing at the feature level through a shared visual encoder, enabling comprehensive interaction between VLM and E2E components. We propose a Trajectory Chain-of-Thought (T-CoT) paradigm, which progressively refines trajectory predictions, reducing uncertainty and improving accuracy. By employing a temporal decoupling strategy, SOLVE achieves efficient cooperation by aligning high-quality VLM outputs with E2E real-time performance. Evaluated on the nuScenes dataset, our method demonstrates significant improvements in trajectory prediction accuracy, paving the way for more robust and reliable autonomous driving systems." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.222, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Over the past two decades, autonomous driving has always been a key research focus, with the potential to significantly revolutionize transportation [25, 32, 41]. Thanks to rapid advances in deep learning for perception [22, 26, 34, 38], prediction [14, 29, 31], and planning [4, 11, 16, 37, 40], end-to-end autonomous driving [8, 16-19, 49] has gained prominence, where the entire driving process—from perception to planning—is learned within a unified framework. Typically, these methods predict future trajectories or control signals directly from sensor inputs, bypassing distinct" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.308, + 0.917, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Previous methods combine VLM and end-to-end networks through post-processing, while our method combines VLM and end-to-end networks through both feature-level synergy (shared visual encoder) and trajectory-level synergy." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.521, + 0.907, + 0.581 + ], + "angle": 0, + "content": "decision-making steps. Through extensive data training, end-to-end approaches have demonstrated impressive planning capabilities, providing a streamlined and competitive alternative to traditional modular pipelines [5, 21, 33]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.909, + 0.87 + ], + "angle": 0, + "content": "End-to-end autonomous driving, while promising, still faces significant challenges, such as learning complexity [7], insufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. For instance, current approaches may misinterpret trucks carrying traffic cones as roadblocks or mistake billboard images of vehicles for real obstacles, resulting in unnecessary braking. Recent advancements in large Vision-Language Models (VLMs) [1, 23, 28] offer a potential solution to these challenges. VLMs, through extensive training on diverse datasets, can simplify learning complexity by providing a more intuitive understanding of various scenarios. Additionally, the logical and commonsense knowledge provided by VLMs enhances autonomous driving systems, allowing them to better interpret driving environments and make safer decisions in complex situations. However, VLMs still face limitations in areas such as efficiency and 3D perception, which pose challenges for their integration into an end-to-end autonomous driving framework." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Some initial explorations have been made to leverage VLMs within the realm of autonomous driving. For exam-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.239, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.239, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "ple, DriveVLM [41] introduces a dual-system architecture that integrates the classic perception-prediction-planning (3P) pipeline with VLMs. As shown in Figure 1(a), the VLM and the end-to-end (E2E) model operate independently and cooperate at the trajectory level, e.g., using the E2E planner to refine the trajectory predictions made by the VLM. However, this strategy limits information and knowledge sharing between the VLM and the E2E model. For example, the E2E model cannot leverage the high-level reasoning abilities of VLM to pay more attention to critical objects or understand their intents, which is crucial for collision avoidance. Similarly, the VLM cannot integrate the 3D perception capabilities of the E2E model for precise scene comprehension, which is essential for accurate planning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.306, + 0.483, + 0.593 + ], + "angle": 0, + "content": "In this work, we introduce SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and planning integration. As Figure 1 shows, besides the trajectory-level cooperation, which is already explored in previous approaches [41], our method enhances feature-level knowledge sharing through a shared vision encoder, facilitating comprehensive collaboration between the VLM and the E2E model. Leveraging the Q-Former design, the shared encoder efficiently compresses surrounding foreground objects (e.g., cars, pedestrians), background maps (e.g., lane centerlines), and environmental cues (e.g., traffic lights, weather, time of day) into unified visual tokens, reducing redundant feature extraction. Consequently, the Q-Former is designed to operate feature collection and alignment with various perception queries sequentially. This structural alignment enhances consistency across vision perception and control domains, significantly facilitating knowledge sharing and easing the difficulties of scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.793 + ], + "angle": 0, + "content": "To boost precision and efficiency in trajectory prediction, we meticulously optimize the VLM and the E2E model. To address the challenges of VLM in directly generating fine-grained trajectories in an auto-regressive manner, we introduce the Trajectory Chain-of-Thought (T-CoT). This new paradigm progressively refines trajectories by using a predefined trajectory bank and chain-style reasoning. Initially, a trajectory adapter transforms each candidate trajectory from the trajectory bank into trajectory tokens that align with text and visual tokens, enabling the network to select the most suitable trajectory. Subsequently, the selected trajectory serves as a foundation for the next round of refinement, ultimately generating the final trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Given that VLM can produce high-quality trajectories but at a significant computational cost, we propose a temporal decoupling strategy to enable cooperation between the VLM and the E2E model. Unlike previous methods [9], which perform feature-level asynchronization, our approach operates at the trajectory level, specifically aiming to enhance trajectory accuracy. The framework implements a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "memory-based mechanism where VLM-generated trajectories are stored and subsequently accessed by the E2E network as initialization priors, accommodating a temporal offset from the current state. This approach leverages priors of VLM to asynchronously enhance E2E network performance while maintaining real-time system operation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.198, + 0.833, + 0.212 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.214, + 0.905, + 0.259 + ], + "angle": 0, + "content": "- We propose SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and trajectory integration." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.304 + ], + "angle": 0, + "content": "- We propose the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories through a predefined trajectory bank and chain-style reasoning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.364 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of our proposed approach through extensive experiments, and our framework achieves state-of-the-art results for open-loop planning on the nuScenes benchmark." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.38, + 0.655, + 0.395 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.803, + 0.421 + ], + "angle": 0, + "content": "2.1. End-to-end Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.836 + ], + "angle": 0, + "content": "In the early stage, autonomous driving systems often follow a modular design and are equipped with rule-based planners. This design hurdles the system's generalization ability and leads to limited performance. In contrast, end-to-end driving is a fully differentiable system from sensors to control signals, optimizing the entire system to directly map perception inputs to planning outputs. Some early works primarily rely on data-driven approaches to train neural networks for end-to-end autonomous driving. These methods show promise, despite challenges with interpretability. The growth in driving data further promotes end-to-end autonomous driving, with models like UniAD [16] and PARA-Drive [47] using Transformers for multi-task learning like object tracking and planning, enhancing performance through additional supervision. Furthermore, VAD [18] introduces vectorized scene representation for better accuracy and speed, with VAD-\\(\\nu 2\\) [8] further improving planning by capturing uncertainty. Besides, methods such as AD-MLP [49] employ a simple multilayer perceptron (MLP) that relies solely on ego vehicle status information, highlighting the potential over-reliance on ego status in simpler driving scenarios while also revealing limitations in utilizing perception data effectively. BEV-Planner [27] integrates various perception inputs in a bird's-eye view (BEV) format, aiming to enhance planning capabilities in more complex traffic situations by leveraging richer environmental context." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.792, + 0.864 + ], + "angle": 0, + "content": "2.2. VLMs for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Nevertheless, end-to-end autonomous driving often suffers from limitations, such as learning complexity [7], in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12069" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.56 + ], + "angle": 0, + "content": "sufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. There is an urgent need to improve these systems for real-world applications. Integrating Vision-Language Models (VLMs) into autonomous driving systems has gained significant attention due to their potential to enhance explainability, reasoning, and generalization in driving tasks. For instance, models like DriveGPT4 [48] utilize VLMs to predict control signals while simultaneously providing explanations for vehicle actions through an iterative question-and-answer format. Similarly, Drive Anywhere [46] employs patch-aligned feature extraction from multi-modal large language models (MLLMs) to facilitate text-based driving decision-making, while OmniDrive [44] introduces a 3D vision-language model architecture aimed at improving reasoning and planning capabilities. Other innovative approaches, such as DriveLM [36], leverage graph-based visual question-answering (VQA) techniques. In a complementary manner, Drive-with-LLMs [6] encodes perception data into latent space using a Transformer network, which is then processed by an LLM to forecast future trajectories. However, most of these VLM-based approaches require a large computation cost, which cannot achieve real-time decision making. DriveVLM [41] stands out as the first model to merge VLMs with end-to-end architectures, predicting low-frequency trajectories through chain-of-thought reasoning that are refined by an end-to-end model to produce final planning outputs. Despite its promising performance, it only fosters cooperation at the trajectory level, overlooking the high-level complementarity between the VLM and the end-to-end model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.182, + 0.587 + ], + "angle": 0, + "content": "3. SOLVE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.84 + ], + "angle": 0, + "content": "In this section, we provide a detailed introduction to SOLVE. Figure 2 illustrates the overall architecture. We utilize a Sequential Q-Former (SQ-Former) to obtain rich and condensed visual cues from sequences image input (Section 3.1). These visual cues are shared between the Vision-Language Model (VLM) and the end-to-end (E2E) model, enabling efficient feature-level knowledge integration. Additionally, we introduce a Trajectory Chain-of-Thought (T-CoT) mechanism (Section 3.2) that refines trajectory predictions from coarse to fine through chain-style reasoning. By aligning the trajectory predictions of the VLM with E2E real-time processing in a time-decoupled cooperation manner (Section 3.3), SOLVE achieves the balance between accuracy and computational efficiency. Finally, we introduce our multi-stage training strategy (Section 3.4) to further promote feature-level synergy and trajectory prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.218, + 0.863 + ], + "angle": 0, + "content": "3.1. SQ-Former" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Visual cues play a pivotal role in autonomous driving, offering foreground information, scene map information, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "environmental cues. These elements help the system to accurately predict vehicle behavior and plan safe routes, enhancing overall decision-making and navigation capabilities. However, abundant visual information can also introduce interference in autonomous driving, such as irrelevant background clutter or reflections. Effectively filtering and extracting useful visual information ensures accuracy and efficiency for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.543 + ], + "angle": 0, + "content": "To harness the relevant information from observations, we categorize useful visual cues into three groups, as shown in Figure 3. The first group includes scene-level static or holistic environmental details, such as weather, time of day, and traffic conditions. The second group concentrates on dynamic participants on the road, such as vehicles and pedestrians. The third group contains dynamic map cues, e.g., lanes. Here, we present Sequential Q-Former (SQ-Former), a Q-Former-style architecture that sequentially compresses these groups of visual cues into a fixed number of condensed visual queries, aligning perception with reasoning and planning. Specifically, the SQ-Former is built upon an image encoder, mapping the multi-view images into high-level features \\(\\mathbf{F}_i\\in \\mathbb{R}^{N\\times C\\times H\\times W}\\), where \\(N,C,H,W\\) denotes the number of views, channels, height and width, respectively. The extracted features are forwarded into the SQ-Former to progressively collect both static and dynamic visual cues from observations. Inspired by StreamPETR [43], we begin with a set of collector queries \\(\\mathbf{Q}_c\\). These queries are input into a transformer decoder to capture static cues of the first group, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.554, + 0.905, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {c} ^ {(s)} = f _ {s} \\left(\\mathbf {Q} _ {c}, \\mathbf {F} _ {i} + \\mathbf {P} _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.751 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_i\\) denotes the 3D position embedding. \\(f_{s}\\) stands for the transformer decoder, which includes interleaved selfattention and cross-attention layers to promote interaction between queries and extract informative static cues, respectively. Afterward, the updated collector queries \\(\\mathbf{Q}_c^{(s)}\\) are concatenated with task-specific queries \\(\\mathbf{Q}_{\\triangle}\\), where \\(\\triangle\\) represents different tasks, such as object detection (d) or map understanding (m). These queries are then fed into separate transformer decoders. In SQ-Former, we sequentially conduct this process, following a pre-defined order of perception tasks, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.759, + 0.907, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Q} _ {c} ^ {(s, \\boldsymbol {\\Delta} [ 1: n ])}, \\tilde {\\mathbf {Q}} _ {\\boldsymbol {\\Delta} [ n ]} \\\\ = f _ {\\triangle [ n ]} \\left(\\left[ \\mathbf {Q} _ {c} ^ {(s, \\triangle [ 1: n - 1 ])}, \\mathbf {Q} _ {\\triangle [ n ]} \\right], \\mathbf {F} _ {i} + \\mathbf {P} _ {i}, \\mathcal {M}\\right), \\quad (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\([\\cdot]\\) denotes the concatenation operation, and \\(f_{\\pmb{\\Delta}[n]}\\) is the transformer decoder specific to the \\(n\\)-th task. We introduce the memory bank \\(\\mathcal{M}\\), following StreamPETR [43], to incorporate the temporal dynamics of the driving scenario. In this framework, \\(\\mathbf{Q}c^{(s,\\pmb{\\Delta}[1:n])}\\) and \\(\\tilde{\\mathbf{Q}}\\pmb{\\Delta}[n]\\) represent the collector queries enriched with visual cues for tasks 1" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12070" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.904, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.305, + 0.673, + 0.321 + ], + "angle": 0, + "content": "Figure 2. The overall framework of the proposed SOLVE." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.334, + 0.48, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.551 + ], + "angle": 0, + "content": "Figure 3. The detail of the proposed SQ-Former. We first capture the static cues from multi-view images and then sequentially align the model with different perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.641 + ], + "angle": 0, + "content": "through \\( n \\), and the updated task-specific queries, respectively. These updated queries are then used to align with the \\( n \\)-th task. The collector queries are sent to a vision adapter to match the dimensions of the VLM, with an ego token appended for trajectory prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.748 + ], + "angle": 0, + "content": "According to our experiments, the sequential design ensures that SQ-Former captures rich, condensed visual cues from various perception tasks, enhancing synergy between them. For example, compared to OmniDrive [44] that uses 512 queries, our method achieves superior performance with only 384 queries, significantly reducing the computational cost of the following VLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.355, + 0.773 + ], + "angle": 0, + "content": "3.2. Trajectory Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "AD-MLP [49] has demonstrated that even a simple planner can achieve moderate results using only ego status. However, it lacks the ability to reason and integrate logical knowledge, which is crucial for a safe and reliable autonomous driving system. To address this limitation, we introduce VLM to enhance the system's reasoning abilities." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Specifically, we provide the VLM with visual and ego inputs, along with high-quality question-answering (QA)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.532 + ], + "angle": 0, + "content": "pairs that encompass perception, reasoning, and planning [44]. Unlike previous methods, we meticulously develop a Trajectory Chain-of-Thought (T-CoT) process for the planning phase, employing a coarse-to-fine procedure for trajectory prediction. Notably, there are several methods, such as DriveCoT [45], DriveLM [36] and DriveVLM [41], have demonstrated the effectiveness of CoT in autonomous driving. However, their reasoning chain spans from perception to planning, requiring multiple rounds of reasoning and lengthy textual output, resulting in significant computational costs. In contrast, our trajectory CoT operates at the trajectory level, refining trajectories in several steps and requiring minimal output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.909, + 0.838 + ], + "angle": 0, + "content": "To make the T-CoT work at the trajectory level, we first construct a trajectory bank containing 36 candidate trajectories (with corresponding ego status and history trajectory) by clustering training set trajectories using the \\(k\\)-means algorithm for each navigation command (forward, left-turn and right-turn). Next, we utilize the current ego status and historical trajectory as a feature vector, calculate its similarity with the corresponding feature vectors to the clustered trajectories, and then use a top-k operation to select the \\(k_{l}\\) trajectories that are closest to the current feature vector. These are combined with a trajectory predicted by an MLP, which uses only ego information and historical data as input [49], resulting in \\(k_{l} + 1\\) candidate trajectories. We introduce clustered trajectories because, although the coarse trajectories are already acceptable, some predicted trajectories still deviate significantly from the ground truth. Retrieving clustered trajectories provides a pull-back mechanism to correct these inaccuracies. With the trajectory bank, the VLM is tasked with progressively outputting predictions in two stages: trajectory selection and trajectory refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Trajectory selection prompts the VLM to choose the most accurate trajectory from the candidates, based on ego and environmental information. To reduce computational costs, the endpoint of each candidate trajectory is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.473, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 4. The illustration of combination of the proposed trajectory tokens with image tokens and text tokens for the large language model-based planning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.416 + ], + "angle": 0, + "content": "encoded into a single token through the proposed trajectory adapter and inserted into the textual prompt as follows: Here are predefined trajectories \\( \\langle \\text{Traj}_1 \\rangle \\), \\( \\langle \\text{Traj}_2 \\rangle \\), ..., \\( \\langle \\text{Traj}_{k_l + 1} \\rangle \\). Please select the best trajectory for the ego car according to the current scenario., as shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Trajectory refinement involves generating a refined trajectory after selecting the reference trajectory. The reference trajectory consists of waypoints, denoted by \\( W = \\{w_{1}, w_{2}, \\ldots, w_{n}\\} \\), where \\( w_{i} = (x_{i}, y_{i}) \\), representing the vehicle's path over a future period with predetermined intervals. Each waypoint of the reference trajectory is encoded into a waypoint token and injected into the textual prompt as follows: With the selected trajectory as a reference \\( <\\text{Point}_1>, <\\text{Point}_2>, \\ldots, <\\text{Point}_n> \\), please provide the planning trajectory for the ego car, which has a velocity of \\( \\text{SPEED} \\, m/s \\) and an acceleration of \\( \\text{ACCEL} \\, m/s^2 \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.405, + 0.607 + ], + "angle": 0, + "content": "3.3. Time-decoupled Synergy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Our planning-oriented synergy system operates at both feature-level (shared SQ-Former feature) and the trajectory-level, aiming to further enhance final trajectory accuracy. This strategy combines the advanced capabilities of VLMs in generating high-accuracy trajectories with the efficiency of a streamlined E2E model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Specifically, we deploy a simple planner that incorporates cross-attention mechanisms between multiple planning queries and collector queries and then it employs an MLP to project these queries into the final output trajectories. The initialization of the planning queries are performed by encoding position embedding from clustering trajectories from the training set, following SparseDrive [39]. The incorporation between the SQ-Former and the simple planner constructs an E2E model, a pipeline frequently adopted in prior systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Upon generating a trajectory from the VLM, we put it into memory. The end-to-end network can obtain the latest VLM predicted trajectory from the memory (with a de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "lay from the current time) as an additional trajectory initialization. Specifically, we task the VLM with predicting trajectories over extended time horizons (longer than both the VLM's latency and the prediction horizon required by the end-to-end model). Consequently, despite the VLM's latency, the E2E model can still asynchronously leverage the VLM's long-term trajectories to obtain additional high-quality trajectory initialization and get the final collaborative predicted trajectory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.238, + 0.688, + 0.256 + ], + "angle": 0, + "content": "3.4. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.383 + ], + "angle": 0, + "content": "We employ a multi-stage training strategy for SOLVE. In the first stage, we employ the same QA pairs as proposed by OmniDrive [44], prompting the VLM by incorporating LoRA layers [15] to address inquiries related to scene description, perception, prediction, and planning. Through QA training, the collector queries in the visual encoder are optimized to comprehensively encode vital scene information while maintaining efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.459 + ], + "angle": 0, + "content": "Subsequently, we train a trajectory adapter to ensure that it can encode predefined trajectories into a feature space compatible with the VLM. We then train the end-to-end planning head while freezing the previously trained SQ-Former, vision adapter, and trajectory adapter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.535 + ], + "angle": 0, + "content": "In the subsequent phase, leveraging these pre-trained weights, we jointly optimize the VLM and end-to-end planning head without freezing any components. During this phase, we employ the proposed T-CoT to augment the reasoning capabilities of the VLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.596 + ], + "angle": 0, + "content": "Finally, our trajectory-level synergy strategy involves exclusively refining the end-to-end planner by leveraging the trajectory forecasted by the VLM in historical frames as an additional initialized planning query for the current frame." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.646, + 0.627 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.825, + 0.652 + ], + "angle": 0, + "content": "4.1. Dataset and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "NuScenes Dataset. We utilize the nuScenes [3] dataset to train and evaluate our method. This comprehensive, large-scale driving dataset is tailored for urban environments, featuring 1,000 driving scenes, each approximately 20 seconds long. It encompasses diverse sensor data, including surround-view camera images and LiDAR point clouds, along with extensive annotations such as 3D bounding boxes, ego vehicle trajectories, and road maps. For training of VLM, we employ the OmniDrive-nuScenes [44] dataset, a planning-oriented extension of the nuScenes dataset with additional annotations, including perception, prediction, and planning-related QA. Specifically, the dataset features two distinct QA paradigms: Offline QA, which focuses on scene-level reasoning through tasks such as scene description, attention analysis, and decision validation using language models; and Online QA, which emphasizes 3D spa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12072" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.21, + 0.089, + 0.794, + 0.344 + ], + "angle": 0, + "content": "
MethodEgo StatusL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
UniAD [16]-0.591.011.481.030.160.511.640.77
UniAD [16]✓0.200.420.750.460.020.250.840.37
VAD-Base [18]-0.691.221.831.250.060.682.521.09
VAD-Base [18]✓0.170.340.600.370.040.270.670.33
AD-MLP [49]-0.150.320.590.350.000.270.850.37
BEV-Planner [27]-0.300.520.830.550.100.371.300.59
BEV-Planner++ [27]✓0.160.320.570.350.000.290.730.34
DriveVLM [41]✓0.180.340.680.40----
OmniDrive [44]✓0.140.290.550.330.000.130.780.30
EMMA [17]✓0.140.290.540.32----
DriveVLM-Dual [41]✓0.150.290.480.31----
Ours-E2E✓0.140.280.500.310.040.170.680.30
Ours-E2E (Async)✓0.130.270.500.300.040.170.650.29
Ours-VLM✓0.130.250.470.280.000.160.430.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.353, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Table 1. Performance comparison of various methods on the NuScenes dataset for open-loop planning. Methods are evaluated with or without ego status consideration. The proposed methods, particularly SOLVE-VLM, demonstrate superior accuracy and safety, achieving the lowest average L2 error and collision rates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.482, + 0.456 + ], + "angle": 0, + "content": "tial understanding through tasks like 2D-to-3D grounding, proximity analysis, and lane-based object detection using geometric labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous works [27], we employ displacement error (L2 distance) and collision rate as the primary metrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.503, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Implementation Details. We use LLaVA v1.5 [28], whose component of language model is based on LLaMA-7B, as our default large vision-language model, demonstrating exceptional performance in tasks such as question answering, visual localization, and text recognition. We replace its vision encoder with our SQ-Former. For the SQ-Former, we employ a predefined task order of image \\(\\rightarrow\\) 3D detection \\(\\rightarrow\\) lane detection. The image encoder is initialized using EVA02-L [12]. For the detection decoder and lane decoder, we adopt the architecture from StreamPETR [43]. The numbers of object queries, lane queries, and the sequential collector queries are set to 900, 300, and 384, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.696, + 0.455, + 0.712 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, SOLVE-E2E and SOLVE-VLM achieved SOTA results in their respective tracks. Specifically, among end-to-end networks, SOLVE-E2E surpassed UniAD [16], VAD [18], and BEV-planner [27] by 0.15, 0.06, and 0.04, respectively, in terms of L2 (m) error, which shows that SOLVE can better imitate the driving behavior of human experts. On the other hand, SOLVE achieves \\(0.07\\%\\), \\(0.03\\%\\), and \\(0.07\\%\\) lower collision rates than UniAD, VAD, and BEV-planner, respectively, which shows that it has more accurate spatial judgment ability than others and can avoid other objects as much as possible when planning trajectories. It is worth noting that SOLVE-E2E achieved sim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.592 + ], + "angle": 0, + "content": "ilar performance to DriveVLM-dual. This proves the superiority of the joint training of SOLVE's shared visual encoder. Specifically, other end-to-end methods only train visual encoders through detection or segmentation tasks, but lack scene understanding capabilities and may miss important visual cues during the encoding process. In contrast, we use VLM's comprehensive QA task, including perception, prediction, and planning, to ensure that the visual encoder encodes important features in the input signal, such as key objects that need to be avoided, while ignoring some information that is not important for planning, such as some unimportant objects behind the vehicle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.835 + ], + "angle": 0, + "content": "On the other hand, SOLVE-VLM achieves the highest performance compared to other VLM-based methods, surpassing OmniDrive [44] and DriveVLM [41] by 0.05 and 0.03 in terms of L2 error, using large language models of similar capacities (DriveVLM uses QwenVL-7.7B [2], OmniDrive and our SOLVE-VLM use LLaVA1.5-7B [42]). This experiment proves that 1) our proposed SQ-former can effectively compress visual features and better release the reasoning ability of the language model. 2) Our joint training scheme of sharing SQ-Former enables VLM to obtain effective auxiliary supervision from the end-to-end model branch to improve performance. Finally, using the future trajectory predicted by VLM in the historical frame as the initialization of the current moment, SOLVE-E2E (Async) achieves better performance than SOLVE-E2E, with an improvement of 0.01 in terms of L2 (m) error." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.696, + 0.862 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We show two planning results of SOLVE in Fig. 5. To a better understanding of the scene, we also provide input" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.681, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.089, + 0.895, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.596, + 0.908, + 0.625 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of SOLVE, where red lines, blue lines and yellow lines mean the planning results from VLM, E2E-Async and E2E modules. The ego car (green box) and ground truth trajectory (green line) are shown in the right BEV images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.486, + 0.801 + ], + "angle": 0, + "content": "surrounding camera images and project the planning trajectories to the front camera image. As shown in the figure, for simple straight-ahead scenarios (see first row), both VLM and the E2E model can predict trajectories that match the expert trajectory. However, for more complex scenarios, the trajectory predicted by the end-to-end model may deviate from the center line of the lane (see yellow line in second row). At the same time, VLM can better understand the scenario and keep the trajectory in the middle of the lane. Besides, with the help of VLM-predicted trajectories, E2E-Async can predict better trajectories than E2E networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.255, + 0.83 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies to verify the effectiveness of each component in SOLVE. Unless otherwise mentioned, we train both the SQ-Former and VLM for 6 epochs to optimize computational resources while main" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.907, + 0.663 + ], + "angle": 0, + "content": "taining experimental rigor. We take L2 (cm) as the default metric for comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Effects of query number of SQ-former. Table 2 displays the VLM performances of using different numbers of SQ-former queries to encode visual information. First, too few collector query numbers (i.e. 256) can only represent limited features, which may easily lead to the loss of visual features, resulting in the lowest performance. Second, too many collector queries (i.e. 512) may introduce redundant information while increasing the capacity of the SQ-former. It will also increase the amount of computation during VLM training and inference. Finally, we choose 384 queries as our default option to balance performance and computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effects of different designs of visual information encoding in SQ-Former. Table 3 investigates the impact of different visual information encoding strategies for VLM. Q means the initial learnable query and the arrows indicate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12074" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.481, + 0.166 + ], + "angle": 0, + "content": "
Number of SQ-Former's queriesL2 (cm) ↓
1s2s3sAvg
25614.129.254.532.6
38414.129.054.032.3
51214.229.254.432.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.17, + 0.483, + 0.199 + ], + "angle": 0, + "content": "Table 2. Effects of the number of queries of SQ-Former to encode temporal vision information." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.21, + 0.465, + 0.288 + ], + "angle": 0, + "content": "
MethodL2 (cm) ↓
1s2s3sAvg
Q → Det → Lane14.329.555.033.0
Q → Img → Det → Lane14.129.054.032.3
Q → Img → Lane → Det14.229.154.432.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.339 + ], + "angle": 0, + "content": "Table 3. Different designs of visual feature encoding in SQ-Former, where Q means the initial learnable query and \\(\\rightarrow\\) means the order of cross attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.356, + 0.483, + 0.552 + ], + "angle": 0, + "content": "the order in which SQ-Former encodes different elements, including image features, detection query, and lane query. Comparing the first and second rows of the table, we can find that incorporating the encoding of the overall image features can achieve a 0.7 performance improvement of L2, which proves that some scene-level clues contained in the image features, such as road conditions and weather, are beneficial to VLM's trajectory planning. Comparing the second and third rows of the table, we find that the encoding order of different elements also affects the performance. Encoding the detection query first and then the lane query can achieve a 0.3 L2 improvement compared to the reverse order. We use the design in the second row as our choice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Effects of trajectory CoT for VLM planning. Table 4 investigates the impact of the proposed trajectory CoT for VLM. Results demonstrate that direct trajectory prediction without CoT yields suboptimal performance, highlighting the inherent challenges VLMs face in auto-regressive numerical prediction tasks. The integration of trajectory CoT, through reference trajectories and trajectory tokens, provides crucial location priors, reducing the uncertainty of direct waypoint prediction and yielding a 1.1 performance improvement. Our analysis of reference trajectory quantity reveals that insufficient samples (e.g., 4) constrain trajectory diversity, though still achieving a 0.9 improvement over the baseline. Conversely, excessive samples (e.g., 8), while providing more optional initial references, increase both trajectory selection complexity and computational overhead. Based on this trade-off analysis, we empirically determine 6 reference trajectories as the optimal configuration, balancing performance and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effects of sharing SQ-Former. Table 5 shows the impact of shared SQ-Former features on the performance of the E2E planner and VLM planner. The results indicate that feature sharing and joint training protocols significantly outperform independently trained networks. The E2E net" + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.09, + 0.872, + 0.18 + ], + "angle": 0, + "content": "
MethodTrajectory numberL2 (cm) ↓
1s2s3sAvg
w/o CoT-13.626.947.829.5
CoT4 (3+1)12.625.747.328.6
6 (5+1)12.525.647.128.4
8 (7+1)12.525.647.128.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Table 4. Effects of the proposed T-CoT. The trajectory number means the selected top-k and 1 predicted trajectories." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.226, + 0.875, + 0.317 + ], + "angle": 0, + "content": "
MethodSQ-FormerL2 (cm) ↓
1s2s3sAvg
E2Ew/o sharing14.929.653.232.5
sharing14.228.150.531.0
VLMw/o sharing12.826.048.329.0
sharing12.625.547.328.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.327, + 0.905, + 0.355 + ], + "angle": 0, + "content": "Table 5. Effects of sharing SQ-Former features to the E2E module and the VLM module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.518 + ], + "angle": 0, + "content": "work exhibits a \\(1.5\\mathrm{cm}\\) reduction in L2 error when utilizing shared SQ-Former features, suggesting that question-answering supervision enables more effective feature extraction for planning compared to conventional detection and segmentation supervision alone. Similarly, VLM performance improves by 0.6 through E2E network integration. This contrasts with DriveVLM's approach of trajectory initialization through post-processing, which shows no significant improvement in VLM's planning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.634, + 0.546 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.753 + ], + "angle": 0, + "content": "In this work, we introduce SOLVE, an innovative framework that fosters synergy between the Vision-Language Model (VLM) and the end-to-end (E2E) model, emphasizing both knowledge and planning integration. To tackle the challenges faced by VLMs in directly generating fine-grained trajectories in an auto-regressive manner, we present the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories using a predefined trajectory bank and chain-style reasoning. Additionally, we propose a temporal decoupling strategy to facilitate cooperation between the VLM and the E2E model. Experiments on the nuScenes dataset demonstrate that our approach achieves state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.694, + 0.782 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12075" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] J Bai, S Bai, S Yang, S Wang, S Tan, P Wang, J Lin, C Zhou, and J Qwen-VL Zhou. A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Changfang Chen, Yingmin Jia, Minglei Shu, and Yinglong Wang. Hierarchical adaptive path-tracking control for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 16(5):2900-2912, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, pages 14093-14100. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE TPAMI, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, pages 22-38. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Pim De Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. NeurIPS, 32, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, pages 9710-9719, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In WACV, pages 910-919, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[14] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In ICCV, pages 15303-15312, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, pages 17853-17862, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[17] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[19] Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[20] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In ECCV, pages 370-387. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[21] Felix Kunz, Dominik Nuss, Jürgen Wiest, Hendrik Deusch, Stephan Reuter, Franz Gritschneder, Alexander Scheel, Manuel Stübler, Martin Bach, Patrick Hatzelmann, et al. Autonomous driving at ulm university: A modular, robust, and sensor-independent fusion approach. In IEEE Intelligent Vehicles Symposium, pages 666-673. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[22] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[24] Jiankun Li, Hao Li, Jiangjiang Liu, Zhikang Zou, Xiaqing Ye, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Exploring the causality of end-to-end autonomous driving. arXiv preprint arXiv:2407.06546, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[25] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12076" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[27] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, pages 14864-14873, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, pages 7577-7586, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[30] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Sajjad Mozaffari, Omar Y Al-Jarrah, Mehrdad Dianati, Paul Jennings, and Alexandros Mouzakitis. Deep learning-based vehicle behavior prediction for autonomous driving applications: A review. IEEE Transactions on Intelligent Transportation Systems, 23(1):33-47, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[32] Khan Muhammad, Amin Ullah, Jaime Lloret, Javier Del Ser, and Victor Hugo C de Albuquerque. Deep learning for safe autonomous driving: Current challenges and future directions. IEEE Transactions on Intelligent Transportation Systems, 22(7):4316-4336, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[33] Samyeul Noh and Kyounghwan An. Decision-making framework for automated driving in highway environments. IEEE Transactions on Intelligent Transportation Systems, 19 (1):58-71, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[35] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[36] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[37] Haoran Song, Wenchao Ding, Yuxuan Chen, Shaojie Shen, Michael Yu Wang, and Qifeng Chen. Pip: Planning-informed trajectory prediction for autonomous driving. In ECCV, pages 598-614. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[38] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Yunfeng Ai, Dongsheng Yang, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, et al. Motion planning for autonomous driving: The state of the art and future perspectives. IEEE Transactions on Intelligent Vehicles, 8(6):3692-3711, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[43] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In ICCV, pages 3621-3631, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[44] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.414 + ], + "angle": 0, + "content": "[45] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.417, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[46] Tsun-Hsuan Wang, Alaa Maalouf, Wei Xiao, Yutong Ban, Alexander Amini, Guy Rosman, Sertac Karaman, and Daniela Rus. Drive anywhere: Generalizable end-to-end autonomous driving with multi-modal foundation models. In ICRA, pages 6687-6694. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[47] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[49] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1, 2, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12077" + } + ] +] \ No newline at end of file diff --git a/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_origin.pdf b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55994b1bede299f9605f31cb634f5c09590b077e --- /dev/null +++ b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/0ded0517-3ff0-4f0f-aa7e-4de5fe3d35e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3b9084dd99119770366dc2aa675eb16be216ca4f1119f03a975a77e839276cf +size 1143932 diff --git a/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/full.md b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad023b2d1fe461843caed8ebf167680420e4968c --- /dev/null +++ b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/full.md @@ -0,0 +1,258 @@ +# SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving + +Xuesong Chen $^{1,3,4*}$ Linjiang Huang $^{2*}$ Tao Ma $^{1}$ Rongyao Fang $^{1}$ Shaoshuai Shi $^{3\dagger}$ Hongsheng Li $^{1,4\dagger}$ + +$^{1}$ MMLab, CUHK $^{2}$ Institute of Artificial Intelligence, Beihang University $^{3}$ Voyager Research, Didi Chuxing $^{4}$ CPII under InnoHK + +{chenxuesong@link, hsli@ee}.cuhk.edu.hk, ljhuang@buaa.edu.cn, shaoshuaics@gmail.com + +# Abstract + +The integration of Vision-Language Models (VLMs) into autonomous driving systems has shown promise in addressing key challenges such as learning complexity, interpretability, and common-sense reasoning. However, existing approaches often struggle with efficient integration and real-time decision-making due to computational demands. In this paper, we introduce SOLVE, an innovative framework that synergizes VLMs with end-to-end (E2E) models to enhance autonomous vehicle planning. Our approach emphasizes knowledge sharing at the feature level through a shared visual encoder, enabling comprehensive interaction between VLM and E2E components. We propose a Trajectory Chain-of-Thought (T-CoT) paradigm, which progressively refines trajectory predictions, reducing uncertainty and improving accuracy. By employing a temporal decoupling strategy, SOLVE achieves efficient cooperation by aligning high-quality VLM outputs with E2E real-time performance. Evaluated on the nuScenes dataset, our method demonstrates significant improvements in trajectory prediction accuracy, paving the way for more robust and reliable autonomous driving systems. + +# 1. Introduction + +Over the past two decades, autonomous driving has always been a key research focus, with the potential to significantly revolutionize transportation [25, 32, 41]. Thanks to rapid advances in deep learning for perception [22, 26, 34, 38], prediction [14, 29, 31], and planning [4, 11, 16, 37, 40], end-to-end autonomous driving [8, 16-19, 49] has gained prominence, where the entire driving process—from perception to planning—is learned within a unified framework. Typically, these methods predict future trajectories or control signals directly from sensor inputs, bypassing distinct + +![](images/7085294ccfd9e8c5bf0efd74d3199a3b2b6ff1a65a454429a3179b3bafca37fb.jpg) +Figure 1. Previous methods combine VLM and end-to-end networks through post-processing, while our method combines VLM and end-to-end networks through both feature-level synergy (shared visual encoder) and trajectory-level synergy. + +decision-making steps. Through extensive data training, end-to-end approaches have demonstrated impressive planning capabilities, providing a streamlined and competitive alternative to traditional modular pipelines [5, 21, 33]. + +End-to-end autonomous driving, while promising, still faces significant challenges, such as learning complexity [7], insufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. For instance, current approaches may misinterpret trucks carrying traffic cones as roadblocks or mistake billboard images of vehicles for real obstacles, resulting in unnecessary braking. Recent advancements in large Vision-Language Models (VLMs) [1, 23, 28] offer a potential solution to these challenges. VLMs, through extensive training on diverse datasets, can simplify learning complexity by providing a more intuitive understanding of various scenarios. Additionally, the logical and commonsense knowledge provided by VLMs enhances autonomous driving systems, allowing them to better interpret driving environments and make safer decisions in complex situations. However, VLMs still face limitations in areas such as efficiency and 3D perception, which pose challenges for their integration into an end-to-end autonomous driving framework. + +Some initial explorations have been made to leverage VLMs within the realm of autonomous driving. For exam- + +ple, DriveVLM [41] introduces a dual-system architecture that integrates the classic perception-prediction-planning (3P) pipeline with VLMs. As shown in Figure 1(a), the VLM and the end-to-end (E2E) model operate independently and cooperate at the trajectory level, e.g., using the E2E planner to refine the trajectory predictions made by the VLM. However, this strategy limits information and knowledge sharing between the VLM and the E2E model. For example, the E2E model cannot leverage the high-level reasoning abilities of VLM to pay more attention to critical objects or understand their intents, which is crucial for collision avoidance. Similarly, the VLM cannot integrate the 3D perception capabilities of the E2E model for precise scene comprehension, which is essential for accurate planning. + +In this work, we introduce SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and planning integration. As Figure 1 shows, besides the trajectory-level cooperation, which is already explored in previous approaches [41], our method enhances feature-level knowledge sharing through a shared vision encoder, facilitating comprehensive collaboration between the VLM and the E2E model. Leveraging the Q-Former design, the shared encoder efficiently compresses surrounding foreground objects (e.g., cars, pedestrians), background maps (e.g., lane centerlines), and environmental cues (e.g., traffic lights, weather, time of day) into unified visual tokens, reducing redundant feature extraction. Consequently, the Q-Former is designed to operate feature collection and alignment with various perception queries sequentially. This structural alignment enhances consistency across vision perception and control domains, significantly facilitating knowledge sharing and easing the difficulties of scene understanding. + +To boost precision and efficiency in trajectory prediction, we meticulously optimize the VLM and the E2E model. To address the challenges of VLM in directly generating fine-grained trajectories in an auto-regressive manner, we introduce the Trajectory Chain-of-Thought (T-CoT). This new paradigm progressively refines trajectories by using a predefined trajectory bank and chain-style reasoning. Initially, a trajectory adapter transforms each candidate trajectory from the trajectory bank into trajectory tokens that align with text and visual tokens, enabling the network to select the most suitable trajectory. Subsequently, the selected trajectory serves as a foundation for the next round of refinement, ultimately generating the final trajectory. + +Given that VLM can produce high-quality trajectories but at a significant computational cost, we propose a temporal decoupling strategy to enable cooperation between the VLM and the E2E model. Unlike previous methods [9], which perform feature-level asynchronization, our approach operates at the trajectory level, specifically aiming to enhance trajectory accuracy. The framework implements a + +memory-based mechanism where VLM-generated trajectories are stored and subsequently accessed by the E2E network as initialization priors, accommodating a temporal offset from the current state. This approach leverages priors of VLM to asynchronously enhance E2E network performance while maintaining real-time system operation capabilities. + +In summary, our contributions are three-fold: + +- We propose SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and trajectory integration. +- We propose the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories through a predefined trajectory bank and chain-style reasoning. +- We demonstrate the effectiveness of our proposed approach through extensive experiments, and our framework achieves state-of-the-art results for open-loop planning on the nuScenes benchmark. + +# 2. Related Work + +# 2.1. End-to-end Autonomous Driving + +In the early stage, autonomous driving systems often follow a modular design and are equipped with rule-based planners. This design hurdles the system's generalization ability and leads to limited performance. In contrast, end-to-end driving is a fully differentiable system from sensors to control signals, optimizing the entire system to directly map perception inputs to planning outputs. Some early works primarily rely on data-driven approaches to train neural networks for end-to-end autonomous driving. These methods show promise, despite challenges with interpretability. The growth in driving data further promotes end-to-end autonomous driving, with models like UniAD [16] and PARA-Drive [47] using Transformers for multi-task learning like object tracking and planning, enhancing performance through additional supervision. Furthermore, VAD [18] introduces vectorized scene representation for better accuracy and speed, with VAD- $\nu 2$ [8] further improving planning by capturing uncertainty. Besides, methods such as AD-MLP [49] employ a simple multilayer perceptron (MLP) that relies solely on ego vehicle status information, highlighting the potential over-reliance on ego status in simpler driving scenarios while also revealing limitations in utilizing perception data effectively. BEV-Planner [27] integrates various perception inputs in a bird's-eye view (BEV) format, aiming to enhance planning capabilities in more complex traffic situations by leveraging richer environmental context. + +# 2.2. VLMs for Autonomous Driving + +Nevertheless, end-to-end autonomous driving often suffers from limitations, such as learning complexity [7], in + +sufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. There is an urgent need to improve these systems for real-world applications. Integrating Vision-Language Models (VLMs) into autonomous driving systems has gained significant attention due to their potential to enhance explainability, reasoning, and generalization in driving tasks. For instance, models like DriveGPT4 [48] utilize VLMs to predict control signals while simultaneously providing explanations for vehicle actions through an iterative question-and-answer format. Similarly, Drive Anywhere [46] employs patch-aligned feature extraction from multi-modal large language models (MLLMs) to facilitate text-based driving decision-making, while OmniDrive [44] introduces a 3D vision-language model architecture aimed at improving reasoning and planning capabilities. Other innovative approaches, such as DriveLM [36], leverage graph-based visual question-answering (VQA) techniques. In a complementary manner, Drive-with-LLMs [6] encodes perception data into latent space using a Transformer network, which is then processed by an LLM to forecast future trajectories. However, most of these VLM-based approaches require a large computation cost, which cannot achieve real-time decision making. DriveVLM [41] stands out as the first model to merge VLMs with end-to-end architectures, predicting low-frequency trajectories through chain-of-thought reasoning that are refined by an end-to-end model to produce final planning outputs. Despite its promising performance, it only fosters cooperation at the trajectory level, overlooking the high-level complementarity between the VLM and the end-to-end model. + +# 3. SOLVE + +In this section, we provide a detailed introduction to SOLVE. Figure 2 illustrates the overall architecture. We utilize a Sequential Q-Former (SQ-Former) to obtain rich and condensed visual cues from sequences image input (Section 3.1). These visual cues are shared between the Vision-Language Model (VLM) and the end-to-end (E2E) model, enabling efficient feature-level knowledge integration. Additionally, we introduce a Trajectory Chain-of-Thought (T-CoT) mechanism (Section 3.2) that refines trajectory predictions from coarse to fine through chain-style reasoning. By aligning the trajectory predictions of the VLM with E2E real-time processing in a time-decoupled cooperation manner (Section 3.3), SOLVE achieves the balance between accuracy and computational efficiency. Finally, we introduce our multi-stage training strategy (Section 3.4) to further promote feature-level synergy and trajectory prediction. + +# 3.1. SQ-Former + +Visual cues play a pivotal role in autonomous driving, offering foreground information, scene map information, and + +environmental cues. These elements help the system to accurately predict vehicle behavior and plan safe routes, enhancing overall decision-making and navigation capabilities. However, abundant visual information can also introduce interference in autonomous driving, such as irrelevant background clutter or reflections. Effectively filtering and extracting useful visual information ensures accuracy and efficiency for autonomous driving. + +To harness the relevant information from observations, we categorize useful visual cues into three groups, as shown in Figure 3. The first group includes scene-level static or holistic environmental details, such as weather, time of day, and traffic conditions. The second group concentrates on dynamic participants on the road, such as vehicles and pedestrians. The third group contains dynamic map cues, e.g., lanes. Here, we present Sequential Q-Former (SQ-Former), a Q-Former-style architecture that sequentially compresses these groups of visual cues into a fixed number of condensed visual queries, aligning perception with reasoning and planning. Specifically, the SQ-Former is built upon an image encoder, mapping the multi-view images into high-level features $\mathbf{F}_i\in \mathbb{R}^{N\times C\times H\times W}$ , where $N,C,H,W$ denotes the number of views, channels, height and width, respectively. The extracted features are forwarded into the SQ-Former to progressively collect both static and dynamic visual cues from observations. Inspired by StreamPETR [43], we begin with a set of collector queries $\mathbf{Q}_c$ . These queries are input into a transformer decoder to capture static cues of the first group, which can be formulated as: + +$$ +\mathbf {Q} _ {c} ^ {(s)} = f _ {s} \left(\mathbf {Q} _ {c}, \mathbf {F} _ {i} + \mathbf {P} _ {i}\right), \tag {1} +$$ + +where $\mathbf{P}_i$ denotes the 3D position embedding. $f_{s}$ stands for the transformer decoder, which includes interleaved selfattention and cross-attention layers to promote interaction between queries and extract informative static cues, respectively. Afterward, the updated collector queries $\mathbf{Q}_c^{(s)}$ are concatenated with task-specific queries $\mathbf{Q}_{\triangle}$ , where $\triangle$ represents different tasks, such as object detection (d) or map understanding (m). These queries are then fed into separate transformer decoders. In SQ-Former, we sequentially conduct this process, following a pre-defined order of perception tasks, which can be written as: + +$$ +\begin{array}{l} \mathbf {Q} _ {c} ^ {(s, \boldsymbol {\Delta} [ 1: n ])}, \tilde {\mathbf {Q}} _ {\boldsymbol {\Delta} [ n ]} \\ = f _ {\triangle [ n ]} \left(\left[ \mathbf {Q} _ {c} ^ {(s, \triangle [ 1: n - 1 ])}, \mathbf {Q} _ {\triangle [ n ]} \right], \mathbf {F} _ {i} + \mathbf {P} _ {i}, \mathcal {M}\right), \quad (2) \\ \end{array} +$$ + +where $[\cdot]$ denotes the concatenation operation, and $f_{\pmb{\Delta}[n]}$ is the transformer decoder specific to the $n$ -th task. We introduce the memory bank $\mathcal{M}$ , following StreamPETR [43], to incorporate the temporal dynamics of the driving scenario. In this framework, $\mathbf{Q}c^{(s,\pmb{\Delta}[1:n])}$ and $\tilde{\mathbf{Q}}\pmb{\Delta}[n]$ represent the collector queries enriched with visual cues for tasks 1 + +![](images/cea31da02d1e06d259b2065d5304f8387f9aa714f16856a9f05663325d51f365.jpg) +Figure 2. The overall framework of the proposed SOLVE. + +![](images/a2a259d47e4408167df5eafec99acdfb849796ecbba7a8fb0c19514f5647adc5.jpg) +Figure 3. The detail of the proposed SQ-Former. We first capture the static cues from multi-view images and then sequentially align the model with different perception tasks. + +through $n$ , and the updated task-specific queries, respectively. These updated queries are then used to align with the $n$ -th task. The collector queries are sent to a vision adapter to match the dimensions of the VLM, with an ego token appended for trajectory prediction. + +According to our experiments, the sequential design ensures that SQ-Former captures rich, condensed visual cues from various perception tasks, enhancing synergy between them. For example, compared to OmniDrive [44] that uses 512 queries, our method achieves superior performance with only 384 queries, significantly reducing the computational cost of the following VLM. + +# 3.2. Trajectory Chain-of-Thought + +AD-MLP [49] has demonstrated that even a simple planner can achieve moderate results using only ego status. However, it lacks the ability to reason and integrate logical knowledge, which is crucial for a safe and reliable autonomous driving system. To address this limitation, we introduce VLM to enhance the system's reasoning abilities. + +Specifically, we provide the VLM with visual and ego inputs, along with high-quality question-answering (QA) + +pairs that encompass perception, reasoning, and planning [44]. Unlike previous methods, we meticulously develop a Trajectory Chain-of-Thought (T-CoT) process for the planning phase, employing a coarse-to-fine procedure for trajectory prediction. Notably, there are several methods, such as DriveCoT [45], DriveLM [36] and DriveVLM [41], have demonstrated the effectiveness of CoT in autonomous driving. However, their reasoning chain spans from perception to planning, requiring multiple rounds of reasoning and lengthy textual output, resulting in significant computational costs. In contrast, our trajectory CoT operates at the trajectory level, refining trajectories in several steps and requiring minimal output. + +To make the T-CoT work at the trajectory level, we first construct a trajectory bank containing 36 candidate trajectories (with corresponding ego status and history trajectory) by clustering training set trajectories using the $k$ -means algorithm for each navigation command (forward, left-turn and right-turn). Next, we utilize the current ego status and historical trajectory as a feature vector, calculate its similarity with the corresponding feature vectors to the clustered trajectories, and then use a top-k operation to select the $k_{l}$ trajectories that are closest to the current feature vector. These are combined with a trajectory predicted by an MLP, which uses only ego information and historical data as input [49], resulting in $k_{l} + 1$ candidate trajectories. We introduce clustered trajectories because, although the coarse trajectories are already acceptable, some predicted trajectories still deviate significantly from the ground truth. Retrieving clustered trajectories provides a pull-back mechanism to correct these inaccuracies. With the trajectory bank, the VLM is tasked with progressively outputting predictions in two stages: trajectory selection and trajectory refinement. + +Trajectory selection prompts the VLM to choose the most accurate trajectory from the candidates, based on ego and environmental information. To reduce computational costs, the endpoint of each candidate trajectory is + +![](images/3730a3439857b47253359c5ac5da0cae53d6f4bd1266c3cf3f96301e648f9d15.jpg) +Figure 4. The illustration of combination of the proposed trajectory tokens with image tokens and text tokens for the large language model-based planning. + +encoded into a single token through the proposed trajectory adapter and inserted into the textual prompt as follows: Here are predefined trajectories $\langle \text{Traj}_1 \rangle$ , $\langle \text{Traj}_2 \rangle$ , ..., $\langle \text{Traj}_{k_l + 1} \rangle$ . Please select the best trajectory for the ego car according to the current scenario., as shown in Figure 4. + +Trajectory refinement involves generating a refined trajectory after selecting the reference trajectory. The reference trajectory consists of waypoints, denoted by $W = \{w_{1}, w_{2}, \ldots, w_{n}\}$ , where $w_{i} = (x_{i}, y_{i})$ , representing the vehicle's path over a future period with predetermined intervals. Each waypoint of the reference trajectory is encoded into a waypoint token and injected into the textual prompt as follows: With the selected trajectory as a reference $<\text{Point}_1>, <\text{Point}_2>, \ldots, <\text{Point}_n>$ , please provide the planning trajectory for the ego car, which has a velocity of $\text{SPEED} \, m/s$ and an acceleration of $\text{ACCEL} \, m/s^2$ . + +# 3.3. Time-decoupled Synergy Prediction + +Our planning-oriented synergy system operates at both feature-level (shared SQ-Former feature) and the trajectory-level, aiming to further enhance final trajectory accuracy. This strategy combines the advanced capabilities of VLMs in generating high-accuracy trajectories with the efficiency of a streamlined E2E model. + +Specifically, we deploy a simple planner that incorporates cross-attention mechanisms between multiple planning queries and collector queries and then it employs an MLP to project these queries into the final output trajectories. The initialization of the planning queries are performed by encoding position embedding from clustering trajectories from the training set, following SparseDrive [39]. The incorporation between the SQ-Former and the simple planner constructs an E2E model, a pipeline frequently adopted in prior systems. + +Upon generating a trajectory from the VLM, we put it into memory. The end-to-end network can obtain the latest VLM predicted trajectory from the memory (with a de + +lay from the current time) as an additional trajectory initialization. Specifically, we task the VLM with predicting trajectories over extended time horizons (longer than both the VLM's latency and the prediction horizon required by the end-to-end model). Consequently, despite the VLM's latency, the E2E model can still asynchronously leverage the VLM's long-term trajectories to obtain additional high-quality trajectory initialization and get the final collaborative predicted trajectory. + +# 3.4. Training Strategy + +We employ a multi-stage training strategy for SOLVE. In the first stage, we employ the same QA pairs as proposed by OmniDrive [44], prompting the VLM by incorporating LoRA layers [15] to address inquiries related to scene description, perception, prediction, and planning. Through QA training, the collector queries in the visual encoder are optimized to comprehensively encode vital scene information while maintaining efficiency. + +Subsequently, we train a trajectory adapter to ensure that it can encode predefined trajectories into a feature space compatible with the VLM. We then train the end-to-end planning head while freezing the previously trained SQ-Former, vision adapter, and trajectory adapter. + +In the subsequent phase, leveraging these pre-trained weights, we jointly optimize the VLM and end-to-end planning head without freezing any components. During this phase, we employ the proposed T-CoT to augment the reasoning capabilities of the VLM. + +Finally, our trajectory-level synergy strategy involves exclusively refining the end-to-end planner by leveraging the trajectory forecasted by the VLM in historical frames as an additional initialized planning query for the current frame. + +# 4. Experiments + +# 4.1. Dataset and Implementation Details + +NuScenes Dataset. We utilize the nuScenes [3] dataset to train and evaluate our method. This comprehensive, large-scale driving dataset is tailored for urban environments, featuring 1,000 driving scenes, each approximately 20 seconds long. It encompasses diverse sensor data, including surround-view camera images and LiDAR point clouds, along with extensive annotations such as 3D bounding boxes, ego vehicle trajectories, and road maps. For training of VLM, we employ the OmniDrive-nuScenes [44] dataset, a planning-oriented extension of the nuScenes dataset with additional annotations, including perception, prediction, and planning-related QA. Specifically, the dataset features two distinct QA paradigms: Offline QA, which focuses on scene-level reasoning through tasks such as scene description, attention analysis, and decision validation using language models; and Online QA, which emphasizes 3D spa + +
MethodEgo StatusL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
UniAD [16]-0.591.011.481.030.160.511.640.77
UniAD [16]✓0.200.420.750.460.020.250.840.37
VAD-Base [18]-0.691.221.831.250.060.682.521.09
VAD-Base [18]✓0.170.340.600.370.040.270.670.33
AD-MLP [49]-0.150.320.590.350.000.270.850.37
BEV-Planner [27]-0.300.520.830.550.100.371.300.59
BEV-Planner++ [27]✓0.160.320.570.350.000.290.730.34
DriveVLM [41]✓0.180.340.680.40----
OmniDrive [44]✓0.140.290.550.330.000.130.780.30
EMMA [17]✓0.140.290.540.32----
DriveVLM-Dual [41]✓0.150.290.480.31----
Ours-E2E✓0.140.280.500.310.040.170.680.30
Ours-E2E (Async)✓0.130.270.500.300.040.170.650.29
Ours-VLM✓0.130.250.470.280.000.160.430.20
+ +Table 1. Performance comparison of various methods on the NuScenes dataset for open-loop planning. Methods are evaluated with or without ego status consideration. The proposed methods, particularly SOLVE-VLM, demonstrate superior accuracy and safety, achieving the lowest average L2 error and collision rates. + +tial understanding through tasks like 2D-to-3D grounding, proximity analysis, and lane-based object detection using geometric labels. + +Evaluation Metrics. Following previous works [27], we employ displacement error (L2 distance) and collision rate as the primary metrics. + +Implementation Details. We use LLaVA v1.5 [28], whose component of language model is based on LLaMA-7B, as our default large vision-language model, demonstrating exceptional performance in tasks such as question answering, visual localization, and text recognition. We replace its vision encoder with our SQ-Former. For the SQ-Former, we employ a predefined task order of image $\rightarrow$ 3D detection $\rightarrow$ lane detection. The image encoder is initialized using EVA02-L [12]. For the detection decoder and lane decoder, we adopt the architecture from StreamPETR [43]. The numbers of object queries, lane queries, and the sequential collector queries are set to 900, 300, and 384, respectively. + +# 4.2. Comparison with State-of-the-art Methods + +As shown in Table 1, SOLVE-E2E and SOLVE-VLM achieved SOTA results in their respective tracks. Specifically, among end-to-end networks, SOLVE-E2E surpassed UniAD [16], VAD [18], and BEV-planner [27] by 0.15, 0.06, and 0.04, respectively, in terms of L2 (m) error, which shows that SOLVE can better imitate the driving behavior of human experts. On the other hand, SOLVE achieves $0.07\%$ , $0.03\%$ , and $0.07\%$ lower collision rates than UniAD, VAD, and BEV-planner, respectively, which shows that it has more accurate spatial judgment ability than others and can avoid other objects as much as possible when planning trajectories. It is worth noting that SOLVE-E2E achieved sim + +ilar performance to DriveVLM-dual. This proves the superiority of the joint training of SOLVE's shared visual encoder. Specifically, other end-to-end methods only train visual encoders through detection or segmentation tasks, but lack scene understanding capabilities and may miss important visual cues during the encoding process. In contrast, we use VLM's comprehensive QA task, including perception, prediction, and planning, to ensure that the visual encoder encodes important features in the input signal, such as key objects that need to be avoided, while ignoring some information that is not important for planning, such as some unimportant objects behind the vehicle. + +On the other hand, SOLVE-VLM achieves the highest performance compared to other VLM-based methods, surpassing OmniDrive [44] and DriveVLM [41] by 0.05 and 0.03 in terms of L2 error, using large language models of similar capacities (DriveVLM uses QwenVL-7.7B [2], OmniDrive and our SOLVE-VLM use LLaVA1.5-7B [42]). This experiment proves that 1) our proposed SQ-former can effectively compress visual features and better release the reasoning ability of the language model. 2) Our joint training scheme of sharing SQ-Former enables VLM to obtain effective auxiliary supervision from the end-to-end model branch to improve performance. Finally, using the future trajectory predicted by VLM in the historical frame as the initialization of the current moment, SOLVE-E2E (Async) achieves better performance than SOLVE-E2E, with an improvement of 0.01 in terms of L2 (m) error. + +# 4.3. Qualitative Results + +We show two planning results of SOLVE in Fig. 5. To a better understanding of the scene, we also provide input + +![](images/0a558958e0504eaa43db12cda6bf2686b3c688b4b8afc77e03b2d5c76ac6aa72.jpg) + +![](images/96e782d58b3a6488326ab8d89c3fa59314576268027c0db2d77695ff8f347b35.jpg) +Figure 5. Qualitative results of SOLVE, where red lines, blue lines and yellow lines mean the planning results from VLM, E2E-Async and E2E modules. The ego car (green box) and ground truth trajectory (green line) are shown in the right BEV images. + +surrounding camera images and project the planning trajectories to the front camera image. As shown in the figure, for simple straight-ahead scenarios (see first row), both VLM and the E2E model can predict trajectories that match the expert trajectory. However, for more complex scenarios, the trajectory predicted by the end-to-end model may deviate from the center line of the lane (see yellow line in second row). At the same time, VLM can better understand the scenario and keep the trajectory in the middle of the lane. Besides, with the help of VLM-predicted trajectories, E2E-Async can predict better trajectories than E2E networks. + +# 4.4. Ablation Studies + +We conduct comprehensive ablation studies to verify the effectiveness of each component in SOLVE. Unless otherwise mentioned, we train both the SQ-Former and VLM for 6 epochs to optimize computational resources while main + +taining experimental rigor. We take L2 (cm) as the default metric for comparison. + +Effects of query number of SQ-former. Table 2 displays the VLM performances of using different numbers of SQ-former queries to encode visual information. First, too few collector query numbers (i.e. 256) can only represent limited features, which may easily lead to the loss of visual features, resulting in the lowest performance. Second, too many collector queries (i.e. 512) may introduce redundant information while increasing the capacity of the SQ-former. It will also increase the amount of computation during VLM training and inference. Finally, we choose 384 queries as our default option to balance performance and computation. + +Effects of different designs of visual information encoding in SQ-Former. Table 3 investigates the impact of different visual information encoding strategies for VLM. Q means the initial learnable query and the arrows indicate + +
Number of SQ-Former's queriesL2 (cm) ↓
1s2s3sAvg
25614.129.254.532.6
38414.129.054.032.3
51214.229.254.432.6
+ +Table 2. Effects of the number of queries of SQ-Former to encode temporal vision information. + +
MethodL2 (cm) ↓
1s2s3sAvg
Q → Det → Lane14.329.555.033.0
Q → Img → Det → Lane14.129.054.032.3
Q → Img → Lane → Det14.229.154.432.6
+ +the order in which SQ-Former encodes different elements, including image features, detection query, and lane query. Comparing the first and second rows of the table, we can find that incorporating the encoding of the overall image features can achieve a 0.7 performance improvement of L2, which proves that some scene-level clues contained in the image features, such as road conditions and weather, are beneficial to VLM's trajectory planning. Comparing the second and third rows of the table, we find that the encoding order of different elements also affects the performance. Encoding the detection query first and then the lane query can achieve a 0.3 L2 improvement compared to the reverse order. We use the design in the second row as our choice. + +Effects of trajectory CoT for VLM planning. Table 4 investigates the impact of the proposed trajectory CoT for VLM. Results demonstrate that direct trajectory prediction without CoT yields suboptimal performance, highlighting the inherent challenges VLMs face in auto-regressive numerical prediction tasks. The integration of trajectory CoT, through reference trajectories and trajectory tokens, provides crucial location priors, reducing the uncertainty of direct waypoint prediction and yielding a 1.1 performance improvement. Our analysis of reference trajectory quantity reveals that insufficient samples (e.g., 4) constrain trajectory diversity, though still achieving a 0.9 improvement over the baseline. Conversely, excessive samples (e.g., 8), while providing more optional initial references, increase both trajectory selection complexity and computational overhead. Based on this trade-off analysis, we empirically determine 6 reference trajectories as the optimal configuration, balancing performance and computational efficiency. + +Effects of sharing SQ-Former. Table 5 shows the impact of shared SQ-Former features on the performance of the E2E planner and VLM planner. The results indicate that feature sharing and joint training protocols significantly outperform independently trained networks. The E2E net + +Table 3. Different designs of visual feature encoding in SQ-Former, where Q means the initial learnable query and $\rightarrow$ means the order of cross attention. + +
MethodTrajectory numberL2 (cm) ↓
1s2s3sAvg
w/o CoT-13.626.947.829.5
CoT4 (3+1)12.625.747.328.6
6 (5+1)12.525.647.128.4
8 (7+1)12.525.647.128.4
+ +Table 4. Effects of the proposed T-CoT. The trajectory number means the selected top-k and 1 predicted trajectories. + +
MethodSQ-FormerL2 (cm) ↓
1s2s3sAvg
E2Ew/o sharing14.929.653.232.5
sharing14.228.150.531.0
VLMw/o sharing12.826.048.329.0
sharing12.625.547.328.4
+ +Table 5. Effects of sharing SQ-Former features to the E2E module and the VLM module. + +work exhibits a $1.5\mathrm{cm}$ reduction in L2 error when utilizing shared SQ-Former features, suggesting that question-answering supervision enables more effective feature extraction for planning compared to conventional detection and segmentation supervision alone. Similarly, VLM performance improves by 0.6 through E2E network integration. This contrasts with DriveVLM's approach of trajectory initialization through post-processing, which shows no significant improvement in VLM's planning capabilities. + +# 5. Conclusion + +In this work, we introduce SOLVE, an innovative framework that fosters synergy between the Vision-Language Model (VLM) and the end-to-end (E2E) model, emphasizing both knowledge and planning integration. To tackle the challenges faced by VLMs in directly generating fine-grained trajectories in an auto-regressive manner, we present the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories using a predefined trajectory bank and chain-style reasoning. Additionally, we propose a temporal decoupling strategy to facilitate cooperation between the VLM and the E2E model. Experiments on the nuScenes dataset demonstrate that our approach achieves state-of-the-art results. + +# 6. Acknowledgement + +This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1 +[2] J Bai, S Bai, S Yang, S Wang, S Tan, P Wang, J Lin, C Zhou, and J Qwen-VL Zhou. A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 5 +[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 1 +[5] Changfang Chen, Yingmin Jia, Minglei Shu, and Yinglong Wang. Hierarchical adaptive path-tracking control for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 16(5):2900-2912, 2015. 1 +[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, pages 14093-14100. IEEE, 2024. 3 +[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE TPAMI, 2024. 1, 2 +[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2 +[9] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, pages 22-38. Springer, 2025. 2 +[10] Pim De Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. NeurIPS, 32, 2019. 1, 3 +[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, pages 9710-9719, 2021. 1 +[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 6 +[13] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In WACV, pages 910-919, 2024. 1, 3 + +[14] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In ICCV, pages 15303-15312, 2021. 1 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, pages 17853-17862, 2023. 1, 2, 6 +[17] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 6 +[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 2, 6 +[19] Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 1 +[20] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In ECCV, pages 370-387. Springer, 2022. 1, 3 +[21] Felix Kunz, Dominik Nuss, Jürgen Wiest, Hendrik Deusch, Stephan Reuter, Franz Gritschneder, Alexander Scheel, Manuel Stübler, Martin Bach, Patrick Hatzelmann, et al. Autonomous driving at ulm university: A modular, robust, and sensor-independent fusion approach. In IEEE Intelligent Vehicles Symposium, pages 666-673. IEEE, 2015. 1 +[22] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023. 1 +[24] Jiankun Li, Hao Li, Jiangjiang Liu, Zhikang Zou, Xiaqing Ye, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Exploring the causality of end-to-end autonomous driving. arXiv preprint arXiv:2407.06546, 2024. 1, 3 +[25] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1 +[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera + +images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 1 +[27] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, pages 14864-14873, 2024. 2, 6 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 6 +[29] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, pages 7577-7586, 2021. 1 +[30] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 1, 3 +[31] Sajjad Mozaffari, Omar Y Al-Jarrah, Mehrdad Dianati, Paul Jennings, and Alexandros Mouzakitis. Deep learning-based vehicle behavior prediction for autonomous driving applications: A review. IEEE Transactions on Intelligent Transportation Systems, 23(1):33-47, 2020. 1 +[32] Khan Muhammad, Amin Ullah, Jaime Lloret, Javier Del Ser, and Victor Hugo C de Albuquerque. Deep learning for safe autonomous driving: Current challenges and future directions. IEEE Transactions on Intelligent Transportation Systems, 22(7):4316-4336, 2020. 1 +[33] Samyeul Noh and Kyounghwan An. Decision-making framework for automated driving in highway environments. IEEE Transactions on Intelligent Transportation Systems, 19 (1):58-71, 2017. 1 +[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1 +[35] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 3 +[36] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3, 4 +[37] Haoran Song, Wenchao Ding, Yuxuan Chen, Shaojie Shen, Michael Yu Wang, and Qifeng Chen. Pip: Planning-informed trajectory prediction for autonomous driving. In ECCV, pages 598-614. Springer, 2020. 1 +[38] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 1 +[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 5 +[40] Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Yunfeng Ai, Dongsheng Yang, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, et al. Motion planning for autonomous driving: The state of the art and future perspectives. IEEE Transactions on Intelligent Vehicles, 8(6):3692-3711, 2023. 1 + +[41] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 4, 6 +[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6 +[43] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In ICCV, pages 3621-3631, 2023. 3, 6 +[44] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3, 4, 5, 6 +[45] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 4 +[46] Tsun-Hsuan Wang, Alaa Maalouf, Wei Xiao, Yutong Ban, Alexander Amini, Guy Rosman, Sertac Karaman, and Daniela Rus. Drive anywhere: Generalizable end-to-end autonomous driving with multi-modal foundation models. In ICRA, pages 6687-6694. IEEE, 2024. 3 +[47] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 2 +[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3 +[49] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1, 2, 4, 6 \ No newline at end of file diff --git a/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/images.zip b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62354b84d199556019252480bae773f420394226 --- /dev/null +++ b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb08548df12c99d3c515b3a5eff9e67f58609cf3b79045713656be70d17448cc +size 504413 diff --git a/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/layout.json b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6fafe8f12069ac7d198d1e722afa0e7700df9a23 --- /dev/null +++ b/2025/SOLVE_ Synergy of Language-Vision and End-to-End Networks for Autonomous Driving/layout.json @@ -0,0 +1,6749 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 515, + 140 + ], + "type": "text", + "content": "SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "content": "Xuesong Chen" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "inline_equation", + "content": "^{1,3,4*}" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "content": " Linjiang Huang" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "content": " Tao Ma" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "content": " Rongyao Fang" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "content": " Shaoshuai Shi" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "text", + "content": " Hongsheng Li" + }, + { + "bbox": [ + 61, + 159, + 586, + 175 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "spans": [ + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "text", + "content": "MMLab, CUHK " + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Beihang University " + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "text", + "content": " Voyager Research, Didi Chuxing " + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 175, + 504, + 203 + ], + "type": "text", + "content": " CPII under InnoHK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 205, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 205, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 96, + 205, + 545, + 217 + ], + "type": "text", + "content": "{chenxuesong@link, hsli@ee}.cuhk.edu.hk, ljhuang@buaa.edu.cn, shaoshuaics@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 270, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 521 + ], + "type": "text", + "content": "The integration of Vision-Language Models (VLMs) into autonomous driving systems has shown promise in addressing key challenges such as learning complexity, interpretability, and common-sense reasoning. However, existing approaches often struggle with efficient integration and real-time decision-making due to computational demands. In this paper, we introduce SOLVE, an innovative framework that synergizes VLMs with end-to-end (E2E) models to enhance autonomous vehicle planning. Our approach emphasizes knowledge sharing at the feature level through a shared visual encoder, enabling comprehensive interaction between VLM and E2E components. We propose a Trajectory Chain-of-Thought (T-CoT) paradigm, which progressively refines trajectory predictions, reducing uncertainty and improving accuracy. By employing a temporal decoupling strategy, SOLVE achieves efficient cooperation by aligning high-quality VLM outputs with E2E real-time performance. Evaluated on the nuScenes dataset, our method demonstrates significant improvements in trajectory prediction accuracy, paving the way for more robust and reliable autonomous driving systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 135, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 135, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 135, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 565, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 685 + ], + "type": "text", + "content": "Over the past two decades, autonomous driving has always been a key research focus, with the potential to significantly revolutionize transportation [25, 32, 41]. Thanks to rapid advances in deep learning for perception [22, 26, 34, 38], prediction [14, 29, 31], and planning [4, 11, 16, 37, 40], end-to-end autonomous driving [8, 16-19, 49] has gained prominence, where the entire driving process—from perception to planning—is learned within a unified framework. Typically, these methods predict future trajectories or control signals directly from sensor inputs, bypassing distinct" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 561, + 345 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 561, + 345 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 561, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 561, + 345 + ], + "type": "image", + "image_path": "7085294ccfd9e8c5bf0efd74d3199a3b2b6ff1a65a454429a3179b3bafca37fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 354, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 399 + ], + "type": "text", + "content": "Figure 1. Previous methods combine VLM and end-to-end networks through post-processing, while our method combines VLM and end-to-end networks through both feature-level synergy (shared visual encoder) and trajectory-level synergy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 412, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 412, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 412, + 555, + 460 + ], + "type": "text", + "content": "decision-making steps. Through extensive data training, end-to-end approaches have demonstrated impressive planning capabilities, providing a streamlined and competitive alternative to traditional modular pipelines [5, 21, 33]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 461, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 556, + 689 + ], + "type": "text", + "content": "End-to-end autonomous driving, while promising, still faces significant challenges, such as learning complexity [7], insufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. For instance, current approaches may misinterpret trucks carrying traffic cones as roadblocks or mistake billboard images of vehicles for real obstacles, resulting in unnecessary braking. Recent advancements in large Vision-Language Models (VLMs) [1, 23, 28] offer a potential solution to these challenges. VLMs, through extensive training on diverse datasets, can simplify learning complexity by providing a more intuitive understanding of various scenarios. Additionally, the logical and commonsense knowledge provided by VLMs enhances autonomous driving systems, allowing them to better interpret driving environments and make safer decisions in complex situations. However, VLMs still face limitations in areas such as efficiency and 3D perception, which pose challenges for their integration into an end-to-end autonomous driving framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Some initial explorations have been made to leverage VLMs within the realm of autonomous driving. For exam-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Equal contributions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12068" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "ple, DriveVLM [41] introduces a dual-system architecture that integrates the classic perception-prediction-planning (3P) pipeline with VLMs. As shown in Figure 1(a), the VLM and the end-to-end (E2E) model operate independently and cooperate at the trajectory level, e.g., using the E2E planner to refine the trajectory predictions made by the VLM. However, this strategy limits information and knowledge sharing between the VLM and the E2E model. For example, the E2E model cannot leverage the high-level reasoning abilities of VLM to pay more attention to critical objects or understand their intents, which is crucial for collision avoidance. Similarly, the VLM cannot integrate the 3D perception capabilities of the E2E model for precise scene comprehension, which is essential for accurate planning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 242, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 295, + 469 + ], + "type": "text", + "content": "In this work, we introduce SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and planning integration. As Figure 1 shows, besides the trajectory-level cooperation, which is already explored in previous approaches [41], our method enhances feature-level knowledge sharing through a shared vision encoder, facilitating comprehensive collaboration between the VLM and the E2E model. Leveraging the Q-Former design, the shared encoder efficiently compresses surrounding foreground objects (e.g., cars, pedestrians), background maps (e.g., lane centerlines), and environmental cues (e.g., traffic lights, weather, time of day) into unified visual tokens, reducing redundant feature extraction. Consequently, the Q-Former is designed to operate feature collection and alignment with various perception queries sequentially. This structural alignment enhances consistency across vision perception and control domains, significantly facilitating knowledge sharing and easing the difficulties of scene understanding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "type": "text", + "content": "To boost precision and efficiency in trajectory prediction, we meticulously optimize the VLM and the E2E model. To address the challenges of VLM in directly generating fine-grained trajectories in an auto-regressive manner, we introduce the Trajectory Chain-of-Thought (T-CoT). This new paradigm progressively refines trajectories by using a predefined trajectory bank and chain-style reasoning. Initially, a trajectory adapter transforms each candidate trajectory from the trajectory bank into trajectory tokens that align with text and visual tokens, enabling the network to select the most suitable trajectory. Subsequently, the selected trajectory serves as a foundation for the next round of refinement, ultimately generating the final trajectory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Given that VLM can produce high-quality trajectories but at a significant computational cost, we propose a temporal decoupling strategy to enable cooperation between the VLM and the E2E model. Unlike previous methods [9], which perform feature-level asynchronization, our approach operates at the trajectory level, specifically aiming to enhance trajectory accuracy. The framework implements a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "memory-based mechanism where VLM-generated trajectories are stored and subsequently accessed by the E2E network as initialization priors, accommodating a temporal offset from the current state. This approach leverages priors of VLM to asynchronously enhance E2E network performance while maintaining real-time system operation capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 324, + 156, + 509, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 156, + 509, + 167 + ], + "spans": [ + { + "bbox": [ + 324, + 156, + 509, + 167 + ], + "type": "text", + "content": "In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 169, + 553, + 288 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 169, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 169, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 169, + 553, + 205 + ], + "type": "text", + "content": "- We propose SOLVE, an innovative framework that promotes synergy between the VLM and the E2E model, emphasizing both knowledge and trajectory integration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 205, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 240 + ], + "type": "text", + "content": "- We propose the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories through a predefined trajectory bank and chain-style reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 241, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 288 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of our proposed approach through extensive experiments, and our framework achieves state-of-the-art results for open-loop planning on the nuScenes benchmark." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 300, + 400, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 400, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 400, + 312 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 320, + 491, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 491, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 491, + 333 + ], + "type": "text", + "content": "2.1. End-to-end Autonomous Driving" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 662 + ], + "type": "text", + "content": "In the early stage, autonomous driving systems often follow a modular design and are equipped with rule-based planners. This design hurdles the system's generalization ability and leads to limited performance. In contrast, end-to-end driving is a fully differentiable system from sensors to control signals, optimizing the entire system to directly map perception inputs to planning outputs. Some early works primarily rely on data-driven approaches to train neural networks for end-to-end autonomous driving. These methods show promise, despite challenges with interpretability. The growth in driving data further promotes end-to-end autonomous driving, with models like UniAD [16] and PARA-Drive [47] using Transformers for multi-task learning like object tracking and planning, enhancing performance through additional supervision. Furthermore, VAD [18] introduces vectorized scene representation for better accuracy and speed, with VAD-" + }, + { + "bbox": [ + 313, + 339, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\nu 2" + }, + { + "bbox": [ + 313, + 339, + 555, + 662 + ], + "type": "text", + "content": " [8] further improving planning by capturing uncertainty. Besides, methods such as AD-MLP [49] employ a simple multilayer perceptron (MLP) that relies solely on ego vehicle status information, highlighting the potential over-reliance on ego status in simpler driving scenarios while also revealing limitations in utilizing perception data effectively. BEV-Planner [27] integrates various perception inputs in a bird's-eye view (BEV) format, aiming to enhance planning capabilities in more complex traffic situations by leveraging richer environmental context." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 670, + 484, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 484, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 484, + 684 + ], + "type": "text", + "content": "2.2. VLMs for Autonomous Driving" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Nevertheless, end-to-end autonomous driving often suffers from limitations, such as learning complexity [7], in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12069" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 443 + ], + "type": "text", + "content": "sufficient common-sense reasoning [13, 30], limited interpretability [20, 35], and causal confusion [10, 24]. There is an urgent need to improve these systems for real-world applications. Integrating Vision-Language Models (VLMs) into autonomous driving systems has gained significant attention due to their potential to enhance explainability, reasoning, and generalization in driving tasks. For instance, models like DriveGPT4 [48] utilize VLMs to predict control signals while simultaneously providing explanations for vehicle actions through an iterative question-and-answer format. Similarly, Drive Anywhere [46] employs patch-aligned feature extraction from multi-modal large language models (MLLMs) to facilitate text-based driving decision-making, while OmniDrive [44] introduces a 3D vision-language model architecture aimed at improving reasoning and planning capabilities. Other innovative approaches, such as DriveLM [36], leverage graph-based visual question-answering (VQA) techniques. In a complementary manner, Drive-with-LLMs [6] encodes perception data into latent space using a Transformer network, which is then processed by an LLM to forecast future trajectories. However, most of these VLM-based approaches require a large computation cost, which cannot achieve real-time decision making. DriveVLM [41] stands out as the first model to merge VLMs with end-to-end architectures, predicting low-frequency trajectories through chain-of-thought reasoning that are refined by an end-to-end model to produce final planning outputs. Despite its promising performance, it only fosters cooperation at the trajectory level, overlooking the high-level complementarity between the VLM and the end-to-end model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 453, + 111, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 111, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 111, + 464 + ], + "type": "text", + "content": "3. SOLVE" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 473, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 665 + ], + "type": "text", + "content": "In this section, we provide a detailed introduction to SOLVE. Figure 2 illustrates the overall architecture. We utilize a Sequential Q-Former (SQ-Former) to obtain rich and condensed visual cues from sequences image input (Section 3.1). These visual cues are shared between the Vision-Language Model (VLM) and the end-to-end (E2E) model, enabling efficient feature-level knowledge integration. Additionally, we introduce a Trajectory Chain-of-Thought (T-CoT) mechanism (Section 3.2) that refines trajectory predictions from coarse to fine through chain-style reasoning. By aligning the trajectory predictions of the VLM with E2E real-time processing in a time-decoupled cooperation manner (Section 3.3), SOLVE achieves the balance between accuracy and computational efficiency. Finally, we introduce our multi-stage training strategy (Section 3.4) to further promote feature-level synergy and trajectory prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 671, + 133, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 133, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 133, + 683 + ], + "type": "text", + "content": "3.1. SQ-Former" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Visual cues play a pivotal role in autonomous driving, offering foreground information, scene map information, and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "environmental cues. These elements help the system to accurately predict vehicle behavior and plan safe routes, enhancing overall decision-making and navigation capabilities. However, abundant visual information can also introduce interference in autonomous driving, such as irrelevant background clutter or reflections. Effectively filtering and extracting useful visual information ensures accuracy and efficiency for autonomous driving." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "text", + "content": "To harness the relevant information from observations, we categorize useful visual cues into three groups, as shown in Figure 3. The first group includes scene-level static or holistic environmental details, such as weather, time of day, and traffic conditions. The second group concentrates on dynamic participants on the road, such as vehicles and pedestrians. The third group contains dynamic map cues, e.g., lanes. Here, we present Sequential Q-Former (SQ-Former), a Q-Former-style architecture that sequentially compresses these groups of visual cues into a fixed number of condensed visual queries, aligning perception with reasoning and planning. Specifically, the SQ-Former is built upon an image encoder, mapping the multi-view images into high-level features " + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i\\in \\mathbb{R}^{N\\times C\\times H\\times W}" + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "inline_equation", + "content": "N,C,H,W" + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "text", + "content": " denotes the number of views, channels, height and width, respectively. The extracted features are forwarded into the SQ-Former to progressively collect both static and dynamic visual cues from observations. Inspired by StreamPETR [43], we begin with a set of collector queries " + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_c" + }, + { + "bbox": [ + 313, + 168, + 555, + 430 + ], + "type": "text", + "content": ". These queries are input into a transformer decoder to capture static cues of the first group, which can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 380, + 438, + 553, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 438, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 380, + 438, + 553, + 453 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {c} ^ {(s)} = f _ {s} \\left(\\mathbf {Q} _ {c}, \\mathbf {F} _ {i} + \\mathbf {P} _ {i}\\right), \\tag {1}", + "image_path": "d9ee5c57b379bc761cd9fc05ea59541f766935d9ebf89187a444c30b97fbc179.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "content": " denotes the 3D position embedding. " + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "content": " stands for the transformer decoder, which includes interleaved selfattention and cross-attention layers to promote interaction between queries and extract informative static cues, respectively. Afterward, the updated collector queries " + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_c^{(s)}" + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "content": " are concatenated with task-specific queries " + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\triangle}" + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 313, + 461, + 554, + 594 + ], + "type": "text", + "content": " represents different tasks, such as object detection (d) or map understanding (m). These queries are then fed into separate transformer decoders. In SQ-Former, we sequentially conduct this process, following a pre-defined order of perception tasks, which can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 338, + 601, + 555, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 601, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 338, + 601, + 555, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Q} _ {c} ^ {(s, \\boldsymbol {\\Delta} [ 1: n ])}, \\tilde {\\mathbf {Q}} _ {\\boldsymbol {\\Delta} [ n ]} \\\\ = f _ {\\triangle [ n ]} \\left(\\left[ \\mathbf {Q} _ {c} ^ {(s, \\triangle [ 1: n - 1 ])}, \\mathbf {Q} _ {\\triangle [ n ]} \\right], \\mathbf {F} _ {i} + \\mathbf {P} _ {i}, \\mathcal {M}\\right), \\quad (2) \\\\ \\end{array}", + "image_path": "915cf8a24e61bf35310e91ff671d28457dee027cd24b89cd3ea95130367c5141.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " denotes the concatenation operation, and " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_{\\pmb{\\Delta}[n]}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " is the transformer decoder specific to the " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "-th task. We introduce the memory bank " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": ", following StreamPETR [43], to incorporate the temporal dynamics of the driving scenario. In this framework, " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}c^{(s,\\pmb{\\Delta}[1:n])}" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}\\pmb{\\Delta}[n]" + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": " represent the collector queries enriched with visual cues for tasks 1" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12070" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 553, + 232 + ], + "type": "image", + "image_path": "cea31da02d1e06d259b2065d5304f8387f9aa714f16856a9f05663325d51f365.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 241, + 411, + 254 + ], + "lines": [ + { + "bbox": [ + 198, + 241, + 411, + 254 + ], + "spans": [ + { + "bbox": [ + 198, + 241, + 411, + 254 + ], + "type": "text", + "content": "Figure 2. The overall framework of the proposed SOLVE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 264, + 293, + 392 + ], + "blocks": [ + { + "bbox": [ + 59, + 264, + 293, + 392 + ], + "lines": [ + { + "bbox": [ + 59, + 264, + 293, + 392 + ], + "spans": [ + { + "bbox": [ + 59, + 264, + 293, + 392 + ], + "type": "image", + "image_path": "a2a259d47e4408167df5eafec99acdfb849796ecbba7a8fb0c19514f5647adc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 295, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 436 + ], + "type": "text", + "content": "Figure 3. The detail of the proposed SQ-Former. We first capture the static cues from multi-view images and then sequentially align the model with different perception tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "type": "text", + "content": "through " + }, + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "type": "text", + "content": ", and the updated task-specific queries, respectively. These updated queries are then used to align with the " + }, + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 447, + 295, + 507 + ], + "type": "text", + "content": "-th task. The collector queries are sent to a vision adapter to match the dimensions of the VLM, with an ego token appended for trajectory prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 592 + ], + "type": "text", + "content": "According to our experiments, the sequential design ensures that SQ-Former captures rich, condensed visual cues from various perception tasks, enhancing synergy between them. For example, compared to OmniDrive [44] that uses 512 queries, our method achieves superior performance with only 384 queries, significantly reducing the computational cost of the following VLM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 217, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 217, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 217, + 612 + ], + "type": "text", + "content": "3.2. Trajectory Chain-of-Thought" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "AD-MLP [49] has demonstrated that even a simple planner can achieve moderate results using only ego status. However, it lacks the ability to reason and integrate logical knowledge, which is crucial for a safe and reliable autonomous driving system. To address this limitation, we introduce VLM to enhance the system's reasoning abilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Specifically, we provide the VLM with visual and ego inputs, along with high-quality question-answering (QA)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 266, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 421 + ], + "type": "text", + "content": "pairs that encompass perception, reasoning, and planning [44]. Unlike previous methods, we meticulously develop a Trajectory Chain-of-Thought (T-CoT) process for the planning phase, employing a coarse-to-fine procedure for trajectory prediction. Notably, there are several methods, such as DriveCoT [45], DriveLM [36] and DriveVLM [41], have demonstrated the effectiveness of CoT in autonomous driving. However, their reasoning chain spans from perception to planning, requiring multiple rounds of reasoning and lengthy textual output, resulting in significant computational costs. In contrast, our trajectory CoT operates at the trajectory level, refining trajectories in several steps and requiring minimal output." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "content": "To make the T-CoT work at the trajectory level, we first construct a trajectory bank containing 36 candidate trajectories (with corresponding ego status and history trajectory) by clustering training set trajectories using the " + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "content": "-means algorithm for each navigation command (forward, left-turn and right-turn). Next, we utilize the current ego status and historical trajectory as a feature vector, calculate its similarity with the corresponding feature vectors to the clustered trajectories, and then use a top-k operation to select the " + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "inline_equation", + "content": "k_{l}" + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "content": " trajectories that are closest to the current feature vector. These are combined with a trajectory predicted by an MLP, which uses only ego information and historical data as input [49], resulting in " + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "inline_equation", + "content": "k_{l} + 1" + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "content": " candidate trajectories. We introduce clustered trajectories because, although the coarse trajectories are already acceptable, some predicted trajectories still deviate significantly from the ground truth. Retrieving clustered trajectories provides a pull-back mechanism to correct these inaccuracies. With the trajectory bank, the VLM is tasked with progressively outputting predictions in two stages: trajectory selection and trajectory refinement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Trajectory selection prompts the VLM to choose the most accurate trajectory from the candidates, based on ego and environmental information. To reduce computational costs, the endpoint of each candidate trajectory is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12071" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 289, + 207 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 289, + 207 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 289, + 207 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 289, + 207 + ], + "type": "image", + "image_path": "3730a3439857b47253359c5ac5da0cae53d6f4bd1266c3cf3f96301e648f9d15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "text", + "content": "Figure 4. The illustration of combination of the proposed trajectory tokens with image tokens and text tokens for the large language model-based planning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "text", + "content": "encoded into a single token through the proposed trajectory adapter and inserted into the textual prompt as follows: Here are predefined trajectories " + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "inline_equation", + "content": "\\langle \\text{Traj}_1 \\rangle" + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "inline_equation", + "content": "\\langle \\text{Traj}_2 \\rangle" + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "text", + "content": ", ..., " + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "inline_equation", + "content": "\\langle \\text{Traj}_{k_l + 1} \\rangle" + }, + { + "bbox": [ + 55, + 269, + 295, + 329 + ], + "type": "text", + "content": ". Please select the best trajectory for the ego car according to the current scenario., as shown in Figure 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "content": "Trajectory refinement involves generating a refined trajectory after selecting the reference trajectory. The reference trajectory consists of waypoints, denoted by " + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "inline_equation", + "content": "W = \\{w_{1}, w_{2}, \\ldots, w_{n}\\}" + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "inline_equation", + "content": "w_{i} = (x_{i}, y_{i})" + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "content": ", representing the vehicle's path over a future period with predetermined intervals. Each waypoint of the reference trajectory is encoded into a waypoint token and injected into the textual prompt as follows: With the selected trajectory as a reference " + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "inline_equation", + "content": "<\\text{Point}_1>, <\\text{Point}_2>, \\ldots, <\\text{Point}_n>" + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "content": ", please provide the planning trajectory for the ego car, which has a velocity of " + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "inline_equation", + "content": "\\text{SPEED} \\, m/s" + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "content": " and an acceleration of " + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "inline_equation", + "content": "\\text{ACCEL} \\, m/s^2" + }, + { + "bbox": [ + 54, + 329, + 295, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 247, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 247, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 247, + 480 + ], + "type": "text", + "content": "3.3. Time-decoupled Synergy Prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 556 + ], + "type": "text", + "content": "Our planning-oriented synergy system operates at both feature-level (shared SQ-Former feature) and the trajectory-level, aiming to further enhance final trajectory accuracy. This strategy combines the advanced capabilities of VLMs in generating high-accuracy trajectories with the efficiency of a streamlined E2E model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": "Specifically, we deploy a simple planner that incorporates cross-attention mechanisms between multiple planning queries and collector queries and then it employs an MLP to project these queries into the final output trajectories. The initialization of the planning queries are performed by encoding position embedding from clustering trajectories from the training set, following SparseDrive [39]. The incorporation between the SQ-Former and the simple planner constructs an E2E model, a pipeline frequently adopted in prior systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Upon generating a trajectory from the VLM, we put it into memory. The end-to-end network can obtain the latest VLM predicted trajectory from the memory (with a de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "lay from the current time) as an additional trajectory initialization. Specifically, we task the VLM with predicting trajectories over extended time horizons (longer than both the VLM's latency and the prediction horizon required by the end-to-end model). Consequently, despite the VLM's latency, the E2E model can still asynchronously leverage the VLM's long-term trajectories to obtain additional high-quality trajectory initialization and get the final collaborative predicted trajectory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 188, + 421, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 421, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 421, + 202 + ], + "type": "text", + "content": "3.4. Training Strategy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 207, + 555, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 303 + ], + "type": "text", + "content": "We employ a multi-stage training strategy for SOLVE. In the first stage, we employ the same QA pairs as proposed by OmniDrive [44], prompting the VLM by incorporating LoRA layers [15] to address inquiries related to scene description, perception, prediction, and planning. Through QA training, the collector queries in the visual encoder are optimized to comprehensively encode vital scene information while maintaining efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 554, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 363 + ], + "type": "text", + "content": "Subsequently, we train a trajectory adapter to ensure that it can encode predefined trajectories into a feature space compatible with the VLM. We then train the end-to-end planning head while freezing the previously trained SQ-Former, vision adapter, and trajectory adapter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": "In the subsequent phase, leveraging these pre-trained weights, we jointly optimize the VLM and end-to-end planning head without freezing any components. During this phase, we employ the proposed T-CoT to augment the reasoning capabilities of the VLM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 472 + ], + "type": "text", + "content": "Finally, our trajectory-level synergy strategy involves exclusively refining the end-to-end planner by leveraging the trajectory forecasted by the VLM in historical frames as an additional initialized planning query for the current frame." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 395, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 395, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 395, + 496 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 503, + 504, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 504, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 504, + 516 + ], + "type": "text", + "content": "4.1. Dataset and Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "NuScenes Dataset. We utilize the nuScenes [3] dataset to train and evaluate our method. This comprehensive, large-scale driving dataset is tailored for urban environments, featuring 1,000 driving scenes, each approximately 20 seconds long. It encompasses diverse sensor data, including surround-view camera images and LiDAR point clouds, along with extensive annotations such as 3D bounding boxes, ego vehicle trajectories, and road maps. For training of VLM, we employ the OmniDrive-nuScenes [44] dataset, a planning-oriented extension of the nuScenes dataset with additional annotations, including perception, prediction, and planning-related QA. Specifically, the dataset features two distinct QA paradigms: Offline QA, which focuses on scene-level reasoning through tasks such as scene description, attention analysis, and decision validation using language models; and Online QA, which emphasizes 3D spa" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12072" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 128, + 70, + 485, + 272 + ], + "blocks": [ + { + "bbox": [ + 128, + 70, + 485, + 272 + ], + "lines": [ + { + "bbox": [ + 128, + 70, + 485, + 272 + ], + "spans": [ + { + "bbox": [ + 128, + 70, + 485, + 272 + ], + "type": "table", + "html": "
MethodEgo StatusL2 (m) ↓Collision (%) ↓
1s2s3sAvg.1s2s3sAvg.
UniAD [16]-0.591.011.481.030.160.511.640.77
UniAD [16]✓0.200.420.750.460.020.250.840.37
VAD-Base [18]-0.691.221.831.250.060.682.521.09
VAD-Base [18]✓0.170.340.600.370.040.270.670.33
AD-MLP [49]-0.150.320.590.350.000.270.850.37
BEV-Planner [27]-0.300.520.830.550.100.371.300.59
BEV-Planner++ [27]✓0.160.320.570.350.000.290.730.34
DriveVLM [41]✓0.180.340.680.40----
OmniDrive [44]✓0.140.290.550.330.000.130.780.30
EMMA [17]✓0.140.290.540.32----
DriveVLM-Dual [41]✓0.150.290.480.31----
Ours-E2E✓0.140.280.500.310.040.170.680.30
Ours-E2E (Async)✓0.130.270.500.300.040.170.650.29
Ours-VLM✓0.130.250.470.280.000.160.430.20
", + "image_path": "1a91509e34e68bd1725fe84715297430f7ff26f2bae1b5c8c9b020ffe609cf8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "type": "text", + "content": "Table 1. Performance comparison of various methods on the NuScenes dataset for open-loop planning. Methods are evaluated with or without ego status consideration. The proposed methods, particularly SOLVE-VLM, demonstrate superior accuracy and safety, achieving the lowest average L2 error and collision rates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 325, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 294, + 361 + ], + "type": "text", + "content": "tial understanding through tasks like 2D-to-3D grounding, proximity analysis, and lane-based object detection using geometric labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 397 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous works [27], we employ displacement error (L2 distance) and collision rate as the primary metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "type": "text", + "content": "Implementation Details. We use LLaVA v1.5 [28], whose component of language model is based on LLaMA-7B, as our default large vision-language model, demonstrating exceptional performance in tasks such as question answering, visual localization, and text recognition. We replace its vision encoder with our SQ-Former. For the SQ-Former, we employ a predefined task order of image " + }, + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "type": "text", + "content": " 3D detection " + }, + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 398, + 295, + 542 + ], + "type": "text", + "content": " lane detection. The image encoder is initialized using EVA02-L [12]. For the detection decoder and lane decoder, we adopt the architecture from StreamPETR [43]. The numbers of object queries, lane queries, and the sequential collector queries are set to 900, 300, and 384, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 278, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 278, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 278, + 563 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "As shown in Table 1, SOLVE-E2E and SOLVE-VLM achieved SOTA results in their respective tracks. Specifically, among end-to-end networks, SOLVE-E2E surpassed UniAD [16], VAD [18], and BEV-planner [27] by 0.15, 0.06, and 0.04, respectively, in terms of L2 (m) error, which shows that SOLVE can better imitate the driving behavior of human experts. On the other hand, SOLVE achieves " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.03\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": " lower collision rates than UniAD, VAD, and BEV-planner, respectively, which shows that it has more accurate spatial judgment ability than others and can avoid other objects as much as possible when planning trajectories. It is worth noting that SOLVE-E2E achieved sim" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 325, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 468 + ], + "type": "text", + "content": "ilar performance to DriveVLM-dual. This proves the superiority of the joint training of SOLVE's shared visual encoder. Specifically, other end-to-end methods only train visual encoders through detection or segmentation tasks, but lack scene understanding capabilities and may miss important visual cues during the encoding process. In contrast, we use VLM's comprehensive QA task, including perception, prediction, and planning, to ensure that the visual encoder encodes important features in the input signal, such as key objects that need to be avoided, while ignoring some information that is not important for planning, such as some unimportant objects behind the vehicle." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 469, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 661 + ], + "type": "text", + "content": "On the other hand, SOLVE-VLM achieves the highest performance compared to other VLM-based methods, surpassing OmniDrive [44] and DriveVLM [41] by 0.05 and 0.03 in terms of L2 error, using large language models of similar capacities (DriveVLM uses QwenVL-7.7B [2], OmniDrive and our SOLVE-VLM use LLaVA1.5-7B [42]). This experiment proves that 1) our proposed SQ-former can effectively compress visual features and better release the reasoning ability of the language model. 2) Our joint training scheme of sharing SQ-Former enables VLM to obtain effective auxiliary supervision from the end-to-end model branch to improve performance. Finally, using the future trajectory predicted by VLM in the historical frame as the initialization of the current moment, SOLVE-E2E (Async) achieves better performance than SOLVE-E2E, with an improvement of 0.01 in terms of L2 (m) error." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 670, + 425, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 425, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 425, + 682 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "We show two planning results of SOLVE in Fig. 5. To a better understanding of the scene, we also provide input" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12073" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 416, + 470 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 416, + 470 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 416, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 416, + 470 + ], + "type": "image", + "image_path": "0a558958e0504eaa43db12cda6bf2686b3c688b4b8afc77e03b2d5c76ac6aa72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 418, + 70, + 547, + 472 + ], + "blocks": [ + { + "bbox": [ + 418, + 70, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 418, + 70, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 418, + 70, + 547, + 472 + ], + "type": "image", + "image_path": "96e782d58b3a6488326ab8d89c3fa59314576268027c0db2d77695ff8f347b35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 472, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 555, + 495 + ], + "type": "text", + "content": "Figure 5. Qualitative results of SOLVE, where red lines, blue lines and yellow lines mean the planning results from VLM, E2E-Async and E2E modules. The ego car (green box) and ground truth trajectory (green line) are shown in the right BEV images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 501, + 297, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 297, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 297, + 634 + ], + "type": "text", + "content": "surrounding camera images and project the planning trajectories to the front camera image. As shown in the figure, for simple straight-ahead scenarios (see first row), both VLM and the E2E model can predict trajectories that match the expert trajectory. However, for more complex scenarios, the trajectory predicted by the end-to-end model may deviate from the center line of the lane (see yellow line in second row). At the same time, VLM can better understand the scenario and keep the trajectory in the middle of the lane. Besides, with the help of VLM-predicted trajectories, E2E-Async can predict better trajectories than E2E networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 645, + 156, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 156, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 156, + 657 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies to verify the effectiveness of each component in SOLVE. Unless otherwise mentioned, we train both the SQ-Former and VLM for 6 epochs to optimize computational resources while main" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 501, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 525 + ], + "type": "text", + "content": "taining experimental rigor. We take L2 (cm) as the default metric for comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 529, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 662 + ], + "type": "text", + "content": "Effects of query number of SQ-former. Table 2 displays the VLM performances of using different numbers of SQ-former queries to encode visual information. First, too few collector query numbers (i.e. 256) can only represent limited features, which may easily lead to the loss of visual features, resulting in the lowest performance. Second, too many collector queries (i.e. 512) may introduce redundant information while increasing the capacity of the SQ-former. It will also increase the amount of computation during VLM training and inference. Finally, we choose 384 queries as our default option to balance performance and computation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Effects of different designs of visual information encoding in SQ-Former. Table 3 investigates the impact of different visual information encoding strategies for VLM. Q means the initial learnable query and the arrows indicate" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12074" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 294, + 131 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 294, + 131 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 294, + 131 + ], + "type": "table", + "html": "
Number of SQ-Former's queriesL2 (cm) ↓
1s2s3sAvg
25614.129.254.532.6
38414.129.054.032.3
51214.229.254.432.6
", + "image_path": "d1060f682d7b3cf56e3c619a718893bc4f6b98dbc91cfd12eafff2f6a0894452.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 166, + 284, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "type": "text", + "content": "Table 2. Effects of the number of queries of SQ-Former to encode temporal vision information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 166, + 284, + 228 + ], + "lines": [ + { + "bbox": [ + 70, + 166, + 284, + 228 + ], + "spans": [ + { + "bbox": [ + 70, + 166, + 284, + 228 + ], + "type": "table", + "html": "
MethodL2 (cm) ↓
1s2s3sAvg
Q → Det → Lane14.329.555.033.0
Q → Img → Det → Lane14.129.054.032.3
Q → Img → Lane → Det14.229.154.432.6
", + "image_path": "d27278cf4f0360757aeafdc1884fa5c67605419dddce59f1ce4bdb165d7c0f5c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 281, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 295, + 437 + ], + "type": "text", + "content": "the order in which SQ-Former encodes different elements, including image features, detection query, and lane query. Comparing the first and second rows of the table, we can find that incorporating the encoding of the overall image features can achieve a 0.7 performance improvement of L2, which proves that some scene-level clues contained in the image features, such as road conditions and weather, are beneficial to VLM's trajectory planning. Comparing the second and third rows of the table, we find that the encoding order of different elements also affects the performance. Encoding the detection query first and then the lane query can achieve a 0.3 L2 improvement compared to the reverse order. We use the design in the second row as our choice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 653 + ], + "type": "text", + "content": "Effects of trajectory CoT for VLM planning. Table 4 investigates the impact of the proposed trajectory CoT for VLM. Results demonstrate that direct trajectory prediction without CoT yields suboptimal performance, highlighting the inherent challenges VLMs face in auto-regressive numerical prediction tasks. The integration of trajectory CoT, through reference trajectories and trajectory tokens, provides crucial location priors, reducing the uncertainty of direct waypoint prediction and yielding a 1.1 performance improvement. Our analysis of reference trajectory quantity reveals that insufficient samples (e.g., 4) constrain trajectory diversity, though still achieving a 0.9 improvement over the baseline. Conversely, excessive samples (e.g., 8), while providing more optional initial references, increase both trajectory selection complexity and computational overhead. Based on this trade-off analysis, we empirically determine 6 reference trajectories as the optimal configuration, balancing performance and computational efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Effects of sharing SQ-Former. Table 5 shows the impact of shared SQ-Former features on the performance of the E2E planner and VLM planner. The results indicate that feature sharing and joint training protocols significantly outperform independently trained networks. The E2E net" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 335, + 71, + 533, + 142 + ], + "blocks": [ + { + "bbox": [ + 55, + 235, + 295, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 268 + ], + "type": "text", + "content": "Table 3. Different designs of visual feature encoding in SQ-Former, where Q means the initial learnable query and " + }, + { + "bbox": [ + 55, + 235, + 295, + 268 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 235, + 295, + 268 + ], + "type": "text", + "content": " means the order of cross attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 71, + 533, + 142 + ], + "lines": [ + { + "bbox": [ + 335, + 71, + 533, + 142 + ], + "spans": [ + { + "bbox": [ + 335, + 71, + 533, + 142 + ], + "type": "table", + "html": "
MethodTrajectory numberL2 (cm) ↓
1s2s3sAvg
w/o CoT-13.626.947.829.5
CoT4 (3+1)12.625.747.328.6
6 (5+1)12.525.647.128.4
8 (7+1)12.525.647.128.4
", + "image_path": "3c23ad112c6ccbd66b62884e1f51868beb848c4f51d3e98f146d09b049c2c7ec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 336, + 178, + 535, + 251 + ], + "blocks": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": "Table 4. Effects of the proposed T-CoT. The trajectory number means the selected top-k and 1 predicted trajectories." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 178, + 535, + 251 + ], + "lines": [ + { + "bbox": [ + 336, + 178, + 535, + 251 + ], + "spans": [ + { + "bbox": [ + 336, + 178, + 535, + 251 + ], + "type": "table", + "html": "
MethodSQ-FormerL2 (cm) ↓
1s2s3sAvg
E2Ew/o sharing14.929.653.232.5
sharing14.228.150.531.0
VLMw/o sharing12.826.048.329.0
sharing12.625.547.328.4
", + "image_path": "c15bbd3e2daedc5e6295a1e0a98c8c352d48c5368892b9e8039615db6b39a229.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "type": "text", + "content": "Table 5. Effects of sharing SQ-Former features to the E2E module and the VLM module." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 301, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 410 + ], + "type": "text", + "content": "work exhibits a " + }, + { + "bbox": [ + 313, + 301, + 555, + 410 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 301, + 555, + 410 + ], + "type": "text", + "content": " reduction in L2 error when utilizing shared SQ-Former features, suggesting that question-answering supervision enables more effective feature extraction for planning compared to conventional detection and segmentation supervision alone. Similarly, VLM performance improves by 0.6 through E2E network integration. This contrasts with DriveVLM's approach of trajectory initialization through post-processing, which shows no significant improvement in VLM's planning capabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 420, + 388, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 388, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 388, + 432 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 439, + 555, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 596 + ], + "type": "text", + "content": "In this work, we introduce SOLVE, an innovative framework that fosters synergy between the Vision-Language Model (VLM) and the end-to-end (E2E) model, emphasizing both knowledge and planning integration. To tackle the challenges faced by VLMs in directly generating fine-grained trajectories in an auto-regressive manner, we present the Trajectory Chain-of-Thought (T-CoT), which progressively refines trajectories using a predefined trajectory bank and chain-style reasoning. Additionally, we propose a temporal decoupling strategy to facilitate cooperation between the VLM and the E2E model. Experiments on the nuScenes dataset demonstrate that our approach achieves state-of-the-art results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 424, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 424, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 424, + 619 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": "This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021, by NSFC-RGC Joint Fund Project N_CUHK498/24. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12075" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] J Bai, S Bai, S Yang, S Wang, S Tan, P Wang, J Lin, C Zhou, and J Qwen-VL Zhou. A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "type": "text", + "content": "[4] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 301, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 346 + ], + "type": "text", + "content": "[5] Changfang Chen, Yingmin Jia, Minglei Shu, and Yinglong Wang. Hierarchical adaptive path-tracking control for autonomous vehicles. IEEE Transactions on Intelligent Transportation Systems, 16(5):2900-2912, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "text", + "content": "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. In ICRA, pages 14093-14100. IEEE, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 445 + ], + "type": "text", + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE TPAMI, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 500 + ], + "type": "text", + "content": "[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 546 + ], + "type": "text", + "content": "[9] Yuan Chen, Zi-han Ding, Ziqin Wang, Yan Wang, Lijun Zhang, and Si Liu. Asynchronous large language model enhanced planner for autonomous driving. In ECCV, pages 22-38. Springer, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 568 + ], + "type": "text", + "content": "[10] Pim De Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning. NeurIPS, 32, 2019. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 624 + ], + "type": "text", + "content": "[11] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, pages 9710-9719, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "type": "text", + "content": "[13] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In WACV, pages 910-919, 2024. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In ICCV, pages 15303-15312, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 152 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, pages 17853-17862, 2023. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[17] Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "text", + "content": "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[19] Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[20] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In ECCV, pages 370-387. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 487 + ], + "type": "text", + "content": "[21] Felix Kunz, Dominik Nuss, Jürgen Wiest, Hendrik Deusch, Stephan Reuter, Franz Gritschneder, Alexander Scheel, Manuel Stübler, Martin Bach, Patrick Hatzelmann, et al. Autonomous driving at ulm university: A modular, robust, and sensor-independent fusion approach. In IEEE Intelligent Vehicles Symposium, pages 666-673. IEEE, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "text", + "content": "[22] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 533, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 578 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. pages 19730-19742. PMLR, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 578, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 624 + ], + "type": "text", + "content": "[24] Jiankun Li, Hao Li, Jiangjiang Liu, Zhikang Zou, Xiaqing Ye, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Exploring the causality of end-to-end autonomous driving. arXiv preprint arXiv:2407.06546, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[25] Ying Li, Lingfei Ma, Zilong Zhong, Fei Liu, Michael A Chapman, Dongpu Cao, and Jonathan Li. Deep learning for lidar point clouds in autonomous driving: A review. IEEE Transactions on Neural Networks and Learning Systems, 32 (8):3412-3432, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12076" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[27] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In CVPR, pages 14864-14873, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 161 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "text", + "content": "[29] Yicheng Liu, Jinghuai Zhang, Liangji Fang, Qinhong Jiang, and Bolei Zhou. Multimodal motion prediction with stacked transformers. In CVPR, pages 7577-7586, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "type": "text", + "content": "[30] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[31] Sajjad Mozaffari, Omar Y Al-Jarrah, Mehrdad Dianati, Paul Jennings, and Alexandros Mouzakitis. Deep learning-based vehicle behavior prediction for autonomous driving applications: A review. IEEE Transactions on Intelligent Transportation Systems, 23(1):33-47, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[32] Khan Muhammad, Amin Ullah, Jaime Lloret, Javier Del Ser, and Victor Hugo C de Albuquerque. Deep learning for safe autonomous driving: Current challenges and future directions. IEEE Transactions on Intelligent Transportation Systems, 22(7):4316-4336, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[33] Samyeul Noh and Kyounghwan An. Decision-making framework for automated driving in highway environments. IEEE Transactions on Intelligent Transportation Systems, 19 (1):58-71, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "text", + "content": "[35] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[36] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beiwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "type": "text", + "content": "[37] Haoran Song, Wenchao Ding, Yuxuan Chen, Shaojie Shen, Michael Yu Wang, and Qifeng Chen. Pip: Planning-informed trajectory prediction for autonomous driving. In ECCV, pages 598-614. Springer, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[38] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[39] Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao ran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Yunfeng Ai, Dongsheng Yang, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, et al. Motion planning for autonomous driving: The state of the art and future perspectives. IEEE Transactions on Intelligent Vehicles, 8(6):3692-3711, 2023. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "type": "text", + "content": "[41] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[42] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[43] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In ICCV, pages 3621-3631, 2023. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "text", + "content": "[44] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 285, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 327 + ], + "type": "text", + "content": "[45] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 330, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 330, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 553, + 384 + ], + "type": "text", + "content": "[46] Tsun-Hsuan Wang, Alaa Maalouf, Wei Xiao, Yutong Ban, Alexander Amini, Guy Rosman, Sertac Karaman, and Daniela Rus. Drive anywhere: Generalizable end-to-end autonomous driving with multi-modal foundation models. In ICRA, pages 6687-6694. IEEE, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "text", + "content": "[47] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 441, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 495 + ], + "type": "text", + "content": "[48] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "text", + "content": "[49] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12077" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_content_list.json b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..487016aa8dda21c59e4464acb81351a520231288 --- /dev/null +++ b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_content_list.json @@ -0,0 +1,1462 @@ +[ + { + "type": "text", + "text": "SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts", + "text_level": 1, + "bbox": [ + 140, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shijia Zhao $^{1,2*}$ Qiming Xia $^{1,2*}$ Xusheng Guo $^{1,2}$ Pufan Zou $^{1,2}$ Maoji Zheng $^{1,2}$ Hai Wu $^{1,2}$ Chenglu Wen $^{1,2\\dagger}$ Cheng Wang $^{1,2}$", + "bbox": [ + 174, + 200, + 846, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, China \n $^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China", + "bbox": [ + 101, + 239, + 921, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, sparsely-supervised 3D object detection has gained great attention, achieving performance close to fully-supervised 3D detectors while requiring only a few annotated instances. Nevertheless, these methods suffer challenges when accurate labels are extremely absent. In this paper, we propose a boosting strategy, termed SP3D, explicitly utilizing the cross-modal semantic prompts generated from Large Multimodal Models (LMMs) to boost the 3D detector with robust feature discrimination capability under sparse annotation settings. Specifically, we first develop a Confident Points Semantic Transfer (CPST) module that generates accurate cross-modal semantic prompts through boundary-constrained center cluster selection. Based on these accurate semantic prompts, which we treat as seed points, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module to yield pseudo-supervision signals from the geometry shape of multi-scale neighbor points. Additionally, we design a Distribution Shape score (DS score) that chooses high-quality supervision signals for the initial training of the 3D detector. Experiments on the KITTI dataset and Waymo Open Dataset (WOD) have validated that SP3D can enhance the performance of sparsely supervised detectors by a large margin under meager labeling conditions. Moreover, we verified SP3D in the zero-shot setting, where its performance exceeded that of the state-of-the-art methods. The code is available at https://github.com/xmuqimingxia/SP3D.", + "bbox": [ + 88, + 357, + 485, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 809, + 220, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection, aiming at locating and classifying objects within 3D scenes, has garnered significant attention in", + "bbox": [ + 89, + 834, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/305b3135f0d33c79ca91e5dd308aa752cbda1c0b1956d456f2d7f149cac9ea0a.jpg", + "image_caption": [ + "Figure 1. Performance comparison of the sparsely-supervised detector at various annotation rates. SP3D indicates the CoIn initialized with the proposed SP3D." + ], + "image_footnote": [], + "bbox": [ + 514, + 328, + 905, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "autonomous driving [6, 10, 19, 34, 36]. However, the mainstream 3D detectors relies heavily on high-quality labels annotated by humans, which is not only time-consuming but also sensitive to the subjective impression of annotators.", + "bbox": [ + 509, + 505, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To minimize the dependence of 3D detectors on high-quality manual annotations, recent work has begun to focus on label-efficient training strategies [15, 21, 31, 37]. To discover unlabeled instances, SS3D [14] employs a self-training approach to iteratively optimize the detector trained on sparsely annotated data. CoIn [37] introduces contrastive learning methods, enhancing the model's discriminative capability for various category features. However, existing strategies make 3D detectors struggle to extract sufficiently discriminative features from extremely limited annotations. Fig. 1 shows some examples where the state-of-the-art (SoTA) sparsely-supervised object detector CoIn [37] hardly maintains robust performance with a significant reduction in annotation rate.", + "bbox": [ + 509, + 566, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the successive emergence and widespread application of large language models (LLMs), such as BERT [7] and GPT [1, 3], in natural language processing, research on large multimodal models (LMMs) is also gaining momentum [12, 13, 17, 27]. Benefiting from the outstanding performance of LMMs, the utilization of pre-trained LMMs has led to significant advancements in 2D vision tasks. Inspired by this, [40, 46, 50] transfers the image-text knowl", + "bbox": [ + 509, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author, clwen@xmu.edu.cn.", + "bbox": [ + 109, + 887, + 382, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29374", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df33c4bea8e9b9886939ae7a97635bd84acf2a09e4bc978ef3130af63bd63bb4.jpg", + "image_caption": [ + "Figure 2. Semantics belonging to a foreground object may be incorrectly assigned to background/other objects." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 281, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3d02f993f13d66a59661380e7aa38a8becec8642b55e445381b8847698022c60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 88, + 460, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edge prior from 2D LMMs to 3D point clouds. However, these methods typically focus on classifying individual instances, and there will be certain limitations when applying them directly to outdoor 3D object detection. Despite this, these attempts to transfer the priors from 2D LMMs to 3D point clouds still provide a new perspective for solving the problem of sparsely-supervised 3D object detection.", + "bbox": [ + 88, + 276, + 482, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these methods, we proposed a two-stage training strategy, SP3D, boosting sparsely-supervised 3D object detection with cross-modal semantic prompts generated by LMMs. As shown in Fig. 3, we first employ LMMs to extract semantics from 2D images and explicitly transfer them to 3D point clouds, generating pseudolabels for the first stage of detector training. In the second stage, we fine-tune the trained detector with sparse accurate labels. However, this enhanced paradigm still holds some challenges: (1) Semantic ambiguity at instance edges. Due to the absence of inherent depth information in images, directly transferring image semantics onto point clouds may result in noisy semantic prompts at the edge of the instance. (2) Efficient generation of pseudo-labels. Based on the obtained semantics, directly fitting pseudo-labels may result in incomplete foreground bounding boxes. Moreover, it is difficult to assess the quality of the generated pseudo-labels because of the lack of ground truth.", + "bbox": [ + 91, + 383, + 482, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, our SP3D first designs a Confident Points Semantic Transfer (CPST) module, employing boundary-constrained center cluster selection operation to filter out inaccurate edges to obtain accurate cross-modal semantic prompts. These semantic prompts focus on central foreground semantic masks generated by LMMs. Inspired by unsupervised algorithms [35, 45], we can utilize these as accurate semantic prompts as seed points to generate bounding box pseudo-labels. In this case, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score) to efficiently discover and assess high-quality pseudo-labels with complete foreground information from seed points. As shown in Fig. 3, we utilize the generated pseudo-labels to train the 3D object detector for the first stage. After training, the 3D detector has learned a certain", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "feature discrimination capability from the 2D images. Subsequently, we fine-tune the 3D detector with sparse accurate labels, and in conjunction with current label-efficient methods, it demonstrates relatively high detection capabilities even under extremely low labeling scenarios.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiment results on the KITTI and Waymo datasets show that SP3D substantially enhances the performance of SoTA sparsely-supervised 3D object detectors. Moreover, without fine-tuning on labeled data, SP3D has also shown superior performance in a zero-shot setting, indicating the efficiency of the SP3D-initialized detector. The contributions of this paper can be summarized as follows:", + "bbox": [ + 511, + 167, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a two-stage boosting strategy, SP3D, utilizing accurate cross-modal prompts to boost the feature discrimination capability of 3D detectors under sparsely-supervised situations.", + "- We propose a Confident Points Semantic Transfer (CPST) module for accurate cross-modal semantic prompt generation and constraint, reducing the noise interference in the semantic transfer process.", + "- We propose a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score), which generate and assess high-quality pseudolabels based on semantic prompts to improve the detect performance." + ], + "bbox": [ + 511, + 275, + 903, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 488, + 650, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Label-efficient 3D Object Detection", + "text_level": 1, + "bbox": [ + 511, + 513, + 820, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, label-efficient 3D object detection methods have begun to be explored in responding to the challenge of extremely low annotation volumes. Generally, these label-efficient methods can be categorized into semi-supervised [15, 24, 31, 48], and weakly-supervised [16, 21, 26, 43] approaches according to the difference in quantity and supervision form. In semi-supervised methods, full annotations are provided only for some scenes, while weakly-supervised methods typically adopt weaker annotation forms. For example, VS3D [26] leverages detectors on images to guide unsupervised 3D object detection, and WS3D [21] employs point annotation as a supervision signal. To maintain accuracy while reducing the annotations, SS3D [14] introduces the concept of sparse supervision, annotating only one complete 3D object per frame. CoIn [37] then adopts a contrastive instance feature mining strategy to extract the feature-level pseudo-labels from a significantly reduced amount of annotated data. However, the performance of existing methods remains constrained due to the insufficient feature discriminability of the initial detector, which may affect subsequent training under very few annotations. This work aims to develop a two-stage strategy, enabling the 3D detector to maintain robust feature representation capabilities despite having lower instance annotation ratio.", + "bbox": [ + 511, + 537, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Large Multimodal Models in 3D", + "text_level": 1, + "bbox": [ + 89, + 90, + 375, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the outstanding performance achieved by LMMs in 2D tasks [2, 12, 18, 25, 27], some studies have begun to explore their application in the 3D domain. Inspired by CLIP [27], ULIP [40] enhances the 3D understanding capability by transferring knowledge from 2D LMM to 3D encoder through contrastive learning methods. Similar works are [46, 50]. In the outdoor scenario, SAM3D [44] employs SAM to segment BEV images of point clouds and fit bounding boxes based on the segmentation masks to obtain detection results. CLIP2Scene [5] establishes the connection between point clouds and text by using images as an intermediate modality, enhancing the 3D model's semantic understanding of the scene with the prior knowledge of CLIP. Unlike previous approaches, our SP3D explicitly transfers semantic masks obtained from LMMs onto point clouds to generate high-quality pseudo-labels for the first-stage training of the 3D detector.", + "bbox": [ + 89, + 112, + 483, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multimodal Representation Learning", + "text_level": 1, + "bbox": [ + 89, + 378, + 413, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, using multimodal methods from 2D images and 3D point clouds to enhance 3D detectors has gradually gained the community's attention [19, 28, 32, 34, 39]. However, these works mainly focus on investigating the image-point cloud fusion strategy, neglecting the utilization of images to explore label-efficient 3D detection. To reduce the required annotations, some methods transfer image information into point clouds to generate pseudo-labels [41, 51]. However, semantic ambiguity may occur at the object's edge due to the 2D-3D calibration error. MixSup [41] proposed a connected components labeling strategy, addressing this issue with the spatial separability property inherent to point clouds. SAL [23] employs density-based clustering to refine imperfect projection issues. Compared with these methods, our SP3D provides a simple but efficient way to reduce semantic noise caused by projection errors.", + "bbox": [ + 89, + 401, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 657, + 189, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 681, + 202, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper introduces SP3D, explicitly utilizing the accurate cross-modal semantic prompts generated by LMMs to boost the sparsely-supervised 3D detectors. As shown in Fig. 3, SP3D consists of three primary components: (1) Confident Points Semantic Transfer (CPST) module, acquiring accurate cross-modal semantic prompts, which we treat as seed points, with the boundary-constrained center cluster selection. (2) Dynamic Cluster Pseudo-label Generation (DCPG) module, which dynamically generates pseudo-label proposals based on the geometric shapes within the multi-scale neighborhood of these seed points. (3) Distribution Shape score (DS score), which employs unsupervised priors as the criterion for evaluating the qual", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity of pseudo-label proposals, and we subsequently apply NMS [22] to retain high-quality pseudo-labels further. We then follow the training strategy of CoIn [37] to train an initial 3D detector with enhanced discriminative capacity. We detail our SP3D framework as follows.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Confident Points Semantic Transfer Module", + "text_level": 1, + "bbox": [ + 511, + 175, + 885, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encouraged by the development of LMMs, we first utilize the prior knowledge of LMMs to extract semantic information from 2D images explicitly. Meanwhile, by integrating the calibration relationship matrix between 2D images and 3D point clouds, it is straightforward to transfer semantic information onto point clouds. However, as shown in Fig. 2, there is significant noise in the edge points of instances during the transfer process. These semantic points are mistakenly assigned to other instances can have a detrimental impact on subsequent pseudo-label fitting and the training of the detector. To prevent the incorporation of noise during the transfer of semantic information, we use the boundary-constrained mask shrink operation, followed by the coordinate system transformation, to obtain accurate cross-modal semantic prompts. CSPT is illustrated in Fig. 3(a); we have divided it into two parts as LMMs-guided semantic extraction and confident points filtering.", + "bbox": [ + 511, + 196, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LMMs-guided semantic extraction. The goal of LMMs is to generate high-quality foreground semantic masks. Specifically, we take as input an image $\\mathcal{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ and the category-text prompt $\\mathcal{T}^{\\mathcal{C}} = \\{t_1^{\\mathcal{C}}, t_2^{\\mathcal{C}}, \\dots, t_C^{\\mathcal{C}}\\}$ , where $H$ and $W$ denote the height and width of the image, $C$ denotes the categories of interest. We first utilize FastSAM [49] to perform speed-efficient segmentation as", + "bbox": [ + 511, + 472, + 905, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {\\mathcal {I}} = \\operatorname {S A M} (\\mathcal {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 589, + 903, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_{\\mathcal{I}} \\in \\mathbb{R}^{M \\times H \\times W}$ denotes the $M$ class-agnostic masks extracted from $\\mathcal{I}$ , $\\mathrm{SAM}(\\cdot)$ indicates the FastSAM model. We then utilize $\\mathcal{M}_{\\mathcal{I}}$ as the mask prompts and feed them, along with image $\\mathcal{I}$ , into the SemanticSAM [4] model, which is expected to output the descriptions $\\mathcal{T}^{\\mathcal{D}} = \\{t_1^{\\mathcal{D}}, t_2^{\\mathcal{D}}, \\dots, t_M^{\\mathcal{D}}\\}$ for each mask in $\\mathcal{M}_{\\mathcal{I}}$ . Specifically:", + "bbox": [ + 511, + 614, + 905, + 708 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {\\mathcal {D}} = \\operatorname {S S A M} (\\mathcal {I}, \\mathcal {M} _ {\\mathcal {I}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 718, + 903, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{SSAM}(\\cdot)$ refers to the SemanticSAM model. Then, we calculate the cosine similarity between the category-text $\\mathcal{T}^{\\mathcal{C}}$ and the mask description $\\mathcal{T}^{\\mathcal{D}}$ to filter out uninteresting background masks, obtaining the foreground masks $\\mathcal{M}_{\\mathcal{I}}'$ .", + "bbox": [ + 511, + 746, + 905, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Confident points filtering. As fuzziness of boundaries results from depth occlusions in images and calibration inaccuracies, we opt to constrain the boundary of foreground masks before 2D-3D transfer, retaining only its central portion. Specifically, for each foreground mask, we denote its", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29376", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f318f96b058ae74adb971c7700206bdf6311b61924d22f1751029b456ab92bf3.jpg", + "image_caption": [ + "Figure 3. The overview of our SP3D, including (a) CSPT finds semantic seed points through high-confidence semantic masks transfer, (b) DCPG dynamically clusters neighbor points of seed points to fit pseudo-label proposals, and (c) DS score to evaluate the quality of generated pseudo-label proposals, serving as a scoring metric to NMS to suppress low-quality proposals." + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 903, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "maximum and minimum values in the pixel coordinate system $(u,v)$ as $(u_{min},u_{max},v_{min},v_{max})$ . We perform mask shrink to constraint boundary range of $\\mathcal{M}_{\\mathcal{I}}^{\\prime}$ to obtain $\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}$ :", + "bbox": [ + 89, + 483, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu \\in \\left[ u _ {\\min } + \\frac {1}{2} (1 \\pm \\gamma) \\left(u _ {\\max } - u _ {\\min }\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 541, + 482, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv \\in [ v _ {m i n} + \\frac {1}{2} (1 \\pm \\gamma) (v _ {m a x} - v _ {m i n}) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 573, + 413, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ denotes the shrink factor. With this constraint, we obtain the shrunk masks that filter out the semantically ambiguous regions, ensuring the accuracy of foreground semantic information transferred onto the point clouds. Following [30], we transfer the semantic mask from the image onto the point clouds to obtain accurate cross-modal semantic prompts, namely seed points, using the camera's intrinsic and extrinsic parameter matrices. Note that, different from [30], we adopt the approach of explicitly transferring the shrunk masks onto the point cloud rather than implicitly embedding unprocessed semantic masks into the point cloud's features. This approach helps avoid potential semantic feature confusion between different modalities arising from sparse annotations.", + "bbox": [ + 89, + 611, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dynamic Cluster Pseudo-label Generation", + "text_level": 1, + "bbox": [ + 89, + 832, + 450, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the proposed CPST, we explicitly obtain the semantic seed points from transformation $\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}$ of the foreground mask. Given $\\mathcal{P} \\in \\mathbb{R}^{N \\times 3} = \\{p_1, p_2, \\dots, p_N\\}$ as LiDAR", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "points, we define the seed points covered by $\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}$ as $\\mathcal{P}_T = \\{p_t\\}$ , $\\mathcal{P}_T \\subset \\mathcal{P}$ . It is crucial for the 3D detection task that obtain complete bounding boxes from these seed points.", + "bbox": [ + 511, + 483, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By referring to traditional unsupervised pseudo-label generation methods [45], we produce many pseudo-labels and then use the positional constraints of the seed points to retain the more promising pseudo-labels as supervisory signals. However, existing unsupervised bounding box fitting approaches [35, 45] usually take a fixed constant as cluster radius, leading to the problem of inadequate foreground or excessive background noise for the generated bounding boxes. In this case, we propose a dynamic cluster pseudo-label generation (DCPG) module. This module utilizes the geometry shape of the seed points' multi-scale neighborhood to capture complete foreground information while minimizing background interference. It dynamically generates pseudo-label proposals, as shown in Fig. 3(b).", + "bbox": [ + 511, + 529, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we denote $\\mathcal{P}_T^{(k)}$ as the $k$ -th instance in a point cloud frame and utilize DCPG dynamically generates a clustering radius $r$ for the $t$ -th seed point $p_t^{(k)}$ . We define the updating function for the dynamic radius as", + "bbox": [ + 511, + 741, + 905, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {u p d a t e} (t, r _ {\\text {i n i t}}) = r _ {\\text {i n i t}} \\cdot \\frac {t}{N ^ {(k)}} + \\delta , t = 1, 2, \\dots , N ^ {(k)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 816, + 906, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r_{init}$ is a hyper-parameter set based on empirical experience, $\\delta$ denotes adjustment factor to avoid $r$ too small, and $N^{(k)}$ is the number of seed points in current instance.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29377", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Dynamic cluster pseudo-label generation" + ], + "code_body": "Require: LiDAR points $\\mathcal{P}$ , the seed points set $\\mathcal{P}_T^{(k)}$ of the $k$ -th instance, initial radius $r_{init}$ , ground remove function $\\mathcal{G}$ , box fitting function $\\mathcal{F}$ ; \nEnsure: Pseudo-label proposal set $\\hat{\\mathcal{B}}^{(k)}$ ; \n1: Acquire the number $N^{(k)}$ of points for the seed points set of current instance $\\mathcal{P}_T^{(k)}$ ; \n2: $\\mathcal{P}_{gr} \\gets \\mathcal{G}(\\mathcal{P})$ ; // remove ground points \n3: $\\hat{\\mathcal{B}}^{(k)} = \\emptyset$ ; \n4: for $t = 1, 2, \\dots, N^{(k)}$ do \n5: $p_t = \\mathcal{P}_T^{(k)}[t]$ ; // loop each seed point \n6: $r \\gets \\text{update}(t, r_{init})$ \n7: $\\hat{\\mathbf{b}} \\gets \\mathcal{F}(\\text{DBSCAN}(\\mathcal{P}_{gr}, p_t, r))$ \n8: $\\hat{\\mathcal{B}}^{(k)} \\gets \\hat{\\mathcal{B}}^{(k)} \\cup \\hat{\\mathbf{b}}$ \n9: end for", + "bbox": [ + 91, + 111, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By applying Eq. 4, we dynamically update the radius $r$ , $r \\in (\\delta, r_{init} + \\delta]$ , during point clustering, thereby obtaining foreground clusters with multi-scale receptive fields. Following [45]. We utilize the radius calculated from Eq. 4 as the clustering radius for DBSCAN [8] and employ [47] to fit the bounding box for each foreground cluster. Algorithm 1 summarizes Our DCPG.", + "bbox": [ + 89, + 353, + 483, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Distribution Shape Score", + "text_level": 1, + "bbox": [ + 89, + 467, + 318, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While DCPG has the capacity for high-quality pseudolabel generation, it unavoidably produces an amount of low-quality pseudo-label proposals. The shapes of these proposals and the extent of foreground completeness contained within the proposals usually deviate significantly from reality. Traditional detection methods typically compute the Intersection over Union (IoU) score between predicted bounding boxes and ground-truth (GT) boxes and then employ the NMS [22] to suppress these low-quality proposals. However, lacking GT makes it challenging to directly apply NMS using IoU as the evaluation criterion within our SP3D framework. In this case, we propose a distribution shape score (DS score) to assess the quality of the pseudo-labels using unsupervised prior knowledge. As illustrated in Fig. 3(c), we divided the DS score into two parts: distribution constraint score and meta-shape constraint score.", + "bbox": [ + 89, + 488, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distribution constraint score. Inspired by [20], within a high-quality pseudo-label proposal $\\hat{b}$ , the distances from its interior points $p_{i,i=1,\\dots,n}$ to its boundary roughly follow a Gaussian distribution $\\mathcal{N}(\\mu, \\sigma)$ , where $\\mu = 0.8$ and $\\sigma = 0.2$ , respectively. In other words, we denote random variable $D = \\{d_1, \\dots, d_n\\}$ as the distance between $p_i$ and the box boundary of $\\hat{b}$ , and $D \\sim \\mathcal{N}(0.8, 0.2)$ . Based on this prior, we assign a distribution constraint score to the pseudo-label proposal $\\hat{b}$ by calculating the similarity between the random", + "bbox": [ + 89, + 762, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "variable $D$ corresponding to each $\\hat{b}$ and the normal distribution $\\mathcal{N}$ . Specifically:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {d c} (\\hat {b}) = \\frac {1}{| \\mathcal {P} _ {f g} |} \\sum_ {p _ {i} \\in \\mathcal {P} _ {f g}} \\log (\\mathcal {N} (D | \\mu , \\sigma)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 131, + 906, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\log (\\cdot)$ denotes the logarithm function, $\\mathcal{P}_{fg}$ is the foreground points within $\\hat{b}$ , and $|\\mathcal{P}_{fg}|$ represents the number of points in $\\mathcal{P}_{fg}$ .", + "bbox": [ + 511, + 180, + 906, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meta-shape constraint score. In addition, the shape of a high-quality pseudo-label is expected to be consistent with its template in the real world, which we define as the meta instance, corresponding to its category [35]. For class $c$ , we denote $\\mathcal{B}_c \\in \\{l_c, w_c, h_c\\}$ as the shape of its meta instance, where $l_c$ , $w_c$ and $h_c$ are the normalized length, width and height, respectively. We followed this shape prior to constructing the class-aware meta-shape constraint score $s_{msc}(\\hat{b})$ as", + "bbox": [ + 511, + 246, + 906, + 382 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {m s c} (\\hat {b}) = 1 - \\Phi_ {K L} (\\mathcal {B} _ {c} | | \\hat {\\mathcal {B}} _ {\\hat {b}}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 393, + 906, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_{KL}(\\cdot)$ denotes the normalized KL divergence function, and $\\hat{\\mathcal{B}}_{\\hat{b}} \\in \\{l_{\\hat{b}}, w_{\\hat{b}}, h_{\\hat{b}}\\}$ indicates the normalized shape of the pseudo-label proposal. The purpose of this operation is to suppress the low-quality proposals whose shape deviates significantly from the meta instance. By combining the distribution constraint score and the meta-shape constraint score, we can obtain the DS score as", + "bbox": [ + 511, + 422, + 905, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D S} (\\hat {b}) = \\lambda_ {1} \\bar {s} _ {d c} (\\hat {b}) + \\lambda_ {2} \\bar {s} _ {m s c} (\\hat {b}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 539, + 906, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ denote weight adjustment factor. Notably, to unify the dimension, we normalized the two constraint scores before combining them, resulting in $\\overline{s}_{dc}$ and $\\overline{s}_{msc}$ . We then employ the DS score as a substitution for the confidence score in NMS to suppress the low-quality pseudolabels. We utilize the obtained pseudo-labels in conjunction with CoIn [37] for the first phase of training. Subsequently, we fine-tune the trained detector with a small amount of accurate labels to boost the performance of the 3D detector.", + "bbox": [ + 511, + 568, + 906, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 718, + 645, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Metrics", + "text_level": 1, + "bbox": [ + 511, + 742, + 705, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KITTI dataset. The KITTI [9] dataset is a widely used benchmark for 3D object detection. During the first training stage, we relied solely on the semantic information provided by LMMs to generate pseudo-labels instead of the human-annotated ground truth. In the fine-tuning stage, we split the training set (7,481 scenes) into a train split (3,712 scenes) and a val split (3,769 scenes). We follow [37] to randomly select $10\\%$ of the scenes from the train split and retain only one instance annotation per scene. In this case,", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29378", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b09a8372836e45d4724e34543f85c29d6ea7b060e3d04102ff26769672cd45bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingCostMethodCarPedCyc
EasyMod.HardEasyMod.HardEasyMod.Hard
Fully-sup.100%VoxelRCNN92.384.982.669.663.058.688.772.568.2
Sparsely-sup.20%SS3D89.384.278.2------
2%VoxelRCNN70.554.944.842.638.532.173.347.843.2
2%CoIn89.170.255.650.845.239.680.252.348.6
2%CoIn++92.079.571.546.736.131.282.058.454.6
2%CoIn++ with SP3D91.380.574.067.458.750.992.573.168.3
", + "bbox": [ + 102, + 88, + 893, + 232 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f2818196e249a8e880d1b94bfd63f9d0766d9a9d6f9ee9f784b1c8f32c8974e4.jpg", + "table_caption": [ + "Table 1. Comparison with SoTA sparsely-supervised methods on KITTI val split. All methods are based on VoxelRCNN, and we report the 3D AP results of full cost (100%) and limited cost (20%, 2%). The best sparsely-supervised methods are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
StageLabelMethodsCar-3D @IoU 0.7Car-BEV @IoU 0.7
EasyModHardEasyModHard
Single-stageSparse1. CenterPoint49.6931.5525.9156.7842.5034.14
Sparse2. CoIn (CenterPoint-based)72.0354.8243.7787.2073.5466.03
Sparse3. 2 with SP3D (CenterPoint-based)87.4469.2458.6192.7280.0069.01
-4. Improvements 3→2+15.41+14.42+14.84+5.52+6.46+2.98
Two-stageSparse1. Voxel-RCNN70.5254.9744.8283.6771.1457.71
Sparse2. CoIn (Voxel-RCNN-based)84.5668.4758.0292.3181.0170.24
Sparse3. 2 with SP3D (Voxel-RCNN-based)91.3774.8963.8495.4185.2774.57
-4. Improvements 3→2+6.81+6.42+5.82+3.1+4.26+4.33
Multi-stageSparse1. CasA74.1857.3745.0585.9073.2157.23
Sparse2. CoIn (CasA-based)89.1775.3262.9895.9985.0272.47
Sparse3. 2 with SP3D (CasA-based)91.1275.9466.4696.5585.6576.31
-4. Improvements 3→2+1.95+0.62+3.48+0.56+0.63+3.84
", + "bbox": [ + 91, + 287, + 919, + 510 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ae4b437e905330d286f398a5c654d605f2985564abd7b9cdc74edc055bde669.jpg", + "table_caption": [ + "Table 2. Comparison with different fully-supervised methods. Sparse label refers to the use of limited split (2% annotation cost). The 3D object detection and BEV detection benchmark are evaluated by mAP with R40, under IoU threshold 0.7." + ], + "table_footnote": [], + "table_body": "
MethodsVeh. LEVEL 1Veh. LEVEL 2
mAPmAPHmAPmAPH
CoIn39.639.034.233.6
CoIn+SP3D46.746.040.439.9
Improvement+7.1+7.0+6.2+6.3
", + "bbox": [ + 114, + 574, + 460, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with SoTA sparsely-supervised methods on WOD validation set. All methods are based on CenterPoint, and we report the results at a $1\\%$ annotation cost and 0.7 IoU threshold.", + "bbox": [ + 89, + 670, + 482, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we can obtain a limited split, which merely takes $2\\%$ of instance annotations compared with the origin train split. All results are evaluated using the official evaluation metric: 3D Average Precision (AP) across 40 recall thresholds (R40).", + "bbox": [ + 89, + 746, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Waymo Open Dataset (WOD). The WOD is a widely used large-scale benchmark that contains 798, 202, and 150 sequences (approximately 200 frames per sequence) for training, validation, and testing, respectively. All results are evaluated by using 3D mean Average Precision (mAP) and its variant weighted by heading accuracy (mAPH).", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 575, + 732, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pseudo-labels generation. We directly employed the model parameters provided by FastSAM [49] and SemanticSAM [4] for inference without additional supervisory signals for fine-tuning. To achieve accurate segmentation results, we set a higher segmentation threshold of 0.7 during the FastSAM inference process. To mitigate computational demands, we opted to generate pseudo-labels within a confined spatial domain of the semantically relevant points, specifically within an 8-meter radius. We set mask shrink factor $\\gamma$ to 0.3, initial cluster radius $r_{init}$ to 1, adjustment factor $\\delta$ to 0.1. We utilize unsupervised priors to filter out the low-quality pseudo-labels that are evidently inconsistent with the intuitive expectations and set the weight adjustment factor of DS score $\\lambda_1$ and $\\lambda_2$ as 0.5, 0.5.", + "bbox": [ + 511, + 599, + 906, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detector training. We conduct all experiments with a batch size of 8 and a learning rate of 0.003 on 4 RTX 3090 GPUs. Following previous sparsely-supervised 3D object detection methods [14, 37], we choose three different clas", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29379", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94870e223ed798be4e2b189ab7a8b03db43815291f116a33cf8e7634e4745510.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods2%1%0.5%
CarPed.Cyc.Avg.CarPed.Cyc.Avg.CarPed.Cyc.Avg.
semi-sup.HSSDA81.958.265.868.680.951.942.359.577.819.544.147.2
HSSDA+SP3D83.260.874.472.883.053.269.168.482.322.146.450.3
Improvement+1.3+2.6+8.6+4.2+2.1+1.3+26.8+8.9+4.5+2.6+2.3+3.1
sparsely-sup.CoIn72.336.958.455.865.331.552.249.655.225.530.537.0
CoIn+SP3D73.444.762.360.171.736.959.155.963.336.143.447.6
Improvement+1.1+7.8+3.9+4.3+6.4+5.4+6.9+6.3+8.1+10.6+12.5+10.6
HINTED85.160.081.375.482.849.974.168.980.245.864.563.5
HINTED+SP3D85.560.181.475.684.057.281.074.083.249.273.968.7
Improvement+0.4+0.1+0.1+0.2+1.2+7.3+6.9+5.1+3.0+3.4+9.4+5.2
", + "bbox": [ + 91, + 88, + 916, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/59aacb20c0a6a73da93f15711da66a5080e4461e4cf52d9af699b0d27dd943d9.jpg", + "table_caption": [ + "Table 4. Improvement of SP3D in SoTA semi-supervised and sparsely-supervised methods on KITTI val split under limited cost setting. The results are evaluated by mAP with R40, under IoU thresholds 0.7, 0.5, and 0.5 for 'Car', 'Pedestrian', and 'Cyclist', respectively. Additionally, the AP is computed across the difficulty levels of easy, moderate, and hard." + ], + "table_footnote": [], + "table_body": "
MethodsModalityCar-3D @IOU 0.5/0.7
Train TestEasyMod.Hard
VS3DL+CL40.32/9.0937.36/5.7331.09/5.03
WS3DPRL+CL-/60.01-/44.48-/36.93
SP3D (Ours)L+CL93.75/69.7176.36/48.6571.01/40.53
", + "bbox": [ + 91, + 357, + 498, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sical detectors [6, 33, 42] as our architecture. And we employ the OpenPCDet [29] to conduct our experiments. In the first training stage, we employ CoIn [37] to train an initial detector with the generated pseudo-labels. Then, we use limited split to fine-tune the detector.", + "bbox": [ + 89, + 497, + 482, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. To thoroughly validate the effectiveness of SP3D, we select the SoTA sparsely-supervised method [37] as the primary baseline for comparison. We compare the proposed SP3D approach with the baseline under conventional sparse settings with $2\\%$ annotation cost. We also compared with cross-modal weakly-supervised methods [16, 26] in a zero-shot setting, which also incorporate visual models to extract semantic information to enhance the performance of detectors. Furthermore, we establish baselines under progressively reduced annotation costs to evaluate the sensitivity to annotation costs.", + "bbox": [ + 89, + 594, + 482, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with SoTA Methods", + "text_level": 1, + "bbox": [ + 89, + 771, + 379, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with sparsely-supervised methods. We integrate our proposed SP3D into the SoTA sparsely-supervised 3D detection algorithm, CoIn++ [37]. For a fair comparison, all detectors employ the VoxelRCNN [6] as the base architecture. As illustrated in Tab. 1, SP3D significantly improves the detection performance of CoIn++. The experimental results on WOD, as reported in Tab. 3,", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a079060592ca81eb002f946b5e53f71d1a721634d10ec02756d45ffa871df51.jpg", + "table_caption": [ + "Table 5. Comparison of zero-shot methods on KITTI val split." + ], + "table_footnote": [], + "table_body": "
MethodsModalityVeh.Ped.Cyc.
TrainTest
SAM3DLL19.10.00.0
CM3DL + CL23.70.12.6
SP3D (Ours)L + CL35.015.75.9
", + "bbox": [ + 526, + 357, + 893, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Comparison with zero-shot methods on the WOD validation set. We report Level-2 mAP with IoU thresholds of 0.7, 0.5, and 0.5 for Vehicle, Pedestrian, and Cyclist, respectively.", + "bbox": [ + 511, + 454, + 905, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "also validate the performance enhancement of our method for the sparsely-supervised 3D object detector. However, we observe a slight decrease in precision for the 'Easy' car category with our SP3D-initialized model in Tab. 1. This could arise because our initial pseudo-labels are inferred based on the geometric shape of the objects, which may differ from the conventions of manual annotation. When the point cloud structure of an instance is relatively intact, such discrepancies can lead to noticeable differences in the size of the annotated bounding boxes.", + "bbox": [ + 511, + 526, + 906, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with fully-supervised methods. For a fair comparison, we follow CoIn [37] to select CenterPoint [42], VoxelRCNN [6], and CasA [33] as our baseline detectors, representing three distinct types of detection algorithms. We initialize the 3D detector using our SP3D, followed by fine-tuning with the limited split. As shown in Tab. 2, detectors designed for full supervision struggle to yield optimal results due to scarce annotations. Despite the effectiveness of CoIn in improving this situation, the results achieved are still unsatisfactory for single-stage detection algorithms with relatively simple structures. Our designed strategy, SP3D, significantly reduces this discrepancy, enabling detectors to achieve similar results.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29380", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/03cebefa7dcde1008488b3683e76ab059e9d4f48ec231474545fe49be6daf8ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCar-3D @IoU 0.5
EasyMod.Hard
CPD61.4861.5860.09
CPD+SP3D73.6566.1363.78
Improvement+12.17+4.55+2.83
", + "bbox": [ + 130, + 88, + 444, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with zero-shot methods. We also compared SP3D with SoTA zero-shot methods. SAM3D [44] and CM3D [11] generate 3D pseudo-labels by leveraging large models for the semantic segmentation of 2D images. In line with the zero-shot approach of prior methods, no 3D annotations were used for training; instead, pseudolabels were generated from large models and image data. As shown in Tab. 5 and Tab. 6, the results on the KITTI and Waymo datasets demonstrate that our method significantly outperforms previous approaches.", + "bbox": [ + 89, + 228, + 480, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with different annotation rates. To explore the robustness of SP3D, we conducted experiments on detectors across different supervision paradigms under varying annotation rates, and we reported the results on Tab. 4. For a fair comparison, we initialize the detectors by following the strategies of CoIn (CenterPoint) [37], HSSDA (PV-RCNN) [15], and HINTED (Voxel-RCNN) [38]. The experimental results show that when the annotation rate is limited, detector performance drops sharply as the amount of annotated data decreases. By inheriting SP3D, these detectors' performance has been significantly enhanced, mitigating the performance drop associated with reduced annotation. Taking CoIn as an example, its average AP for the three categories at annotation rates of $1\\%$ and $0.5\\%$ has improved by $6.3\\%$ and $10.6\\%$ , respectively. These results indicate that even at low annotation rates, SP3D-initialized models can significantly improve the performance of 3D detectors across different supervision paradigms.", + "bbox": [ + 89, + 402, + 482, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with unsupervised methods. To further validate the robustness of SP3D, we integrated it with the SoTA unsupervised detection method CPD [35]. As shown in Tab. 7, despite suffering some impact, the combination of SP3D and CPD still enhances the performance of the unsupervised 3D object detector to a certain extent.", + "bbox": [ + 89, + 699, + 482, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 801, + 243, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted ablation studies based on CenterPoint to verify the proposed modules' effectiveness and recorded the results in Tab. 8. The first and third rows show that the precision of pseudo-labels, as augmented by DCPG, can substantially amplify the 3D detector's capabilities. This may be", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3affdc84ebd974da68791ee80d8ff265f2622dd4d8bbd03fc2b09e43059ce06d.jpg", + "table_caption": [ + "Table 7. Improvement for unsupervised method." + ], + "table_footnote": [], + "table_body": "
Mask shrinkDCPGDS scoreCar-3D AP@IoU 0.7
EasyMod.Hard
✓35.1023.7519.52
✓✓38.2629.2524.74
✓✓40.5628.1522.40
✓✓47.2333.4027.13
✓✓✓52.5638.0031.52
", + "bbox": [ + 521, + 88, + 898, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/711b0c79f6966068c9ffad49b79d0fe095992f3cc51c40d799b92d9b385a779d.jpg", + "table_caption": [ + "Table 8. Ablation study on KITTI val split." + ], + "table_footnote": [], + "table_body": "
DS scoreCar-3D AP@IoU 0.7
MSCDCEasyMod.Hard
✓44.9133.5528.19
✓48.4835.3929.75
", + "bbox": [ + 558, + 246, + 861, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study of each component of DS score.", + "bbox": [ + 540, + 328, + 874, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "due to incorporating more comprehensive foreground information in pseudo-labels, which enhanced the model's feature discrimination capability. The last two rows show that the mask shrink operation is necessary for handling semantic noise at the instance edges. The first three rows and fifth row show that using the DS score to filter out low-quality labels can significantly boost the detector's precision. Moreover, we validated the effectiveness of each component in DS score with the ablation studies presented in Tab. 9.", + "bbox": [ + 511, + 371, + 906, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 525, + 761, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a two-stage boosting strategy, SP3D, that leverages accurate cross-modality semantic prompts to enhance the sparsely-supervised 3D detectors' capabilities. We first develop a CSPT module to obtain accurate cross-modal semantic prompts in point clouds by efficiently transferring high-confidence semantic masks. We then design a DCPG module to dynamically generate pseudo-label proposals based on semantic prompts. Lastly, we propose a DS score as the criterion for NMS to select high-quality pseudo-labels integrated with sparsely-supervised strategy to train the initial detector. After fine-tuning with sparse GT data, SP3D show superior performance over the original sparsely-supervised 3D detector on the WOD and KITTI dataset, and remain robust performance even as the amount of annotation decreased. Limitations: SP3D exhibits a relatively significant degradation when fine-tuning with annotation rates below $0.1\\%$ , which may result from the noise introduced by extremely low annotations. Future efforts to explore efficient fine-tuning strategies to solve this issue.", + "bbox": [ + 511, + 551, + 906, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (No.62171393), and the Fundamental Research Funds for the Central Universities (No.20720220064).", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29381", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 3", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[4] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.3,6", + "[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 3", + "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 1, 7", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1", + "[8] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 5", + "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5", + "[10] Xun Huang, Hai Wu, Xin Li, Xiaoliang Fan, Chenglu Wen, and Cheng Wang. Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2409–2416, 2024. 1", + "[11] Mehar Khurana, Neehar Peri, James Hays, and Deva Ramanan. Shelf-supervised cross-modal pre-training for 3d object detection. In 8th Annual Conference on Robot Learning, 2024. 8", + "[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1", + "[14] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8428-8437, 2022. 1, 2, 6", + "[15] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23819-23828, 2023. 1, 2, 8", + "[16] Haizhuang Liu, Huimin Ma, Yilin Wang, Bochao Zou, Tianyu Hu, Rongquan Wang, and Jiansheng Chen. Eliminating spatial ambiguity for weakly supervised 3d object detection without spatial labels. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3511-3520, 2022. 2, 7", + "[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1", + "[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 3", + "[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3", + "[20] Katie Luo, Zhenzhen Liu, Xiangyu Chen, Yurong You, Sagie Benaim, Cheng Perng Phoo, Mark Campbell, Wen Sun, Bharath Hariharan, and Kilian Q Weinberger. Reward finetuning for faster and more accurate unsupervised object discovery. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[21] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Yunde Jia, and Luc Van Gool. Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4454-4468, 2021. 1, 2", + "[22] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 3, 5", + "[23] Aljosa Osep, Tim Meinhardt, Francesco Ferroni, Neehar Peri, Deva Ramanan, and Laura Leal-Taixe. Better call sal: Towards learning to segment anything in lidar. In European Conference on Computer Vision, pages 71-90. Springer, 2025. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29382", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jinhyung Park, Chenfeng Xu, Yiyang Zhou, Masayoshi Tomizuka, and Wei Zhan. Detmatch: Two teachers are better than one for joint 2d and 3d semi-supervised object detection. In European Conference on Computer Vision, pages 370-389. Springer, 2022. 2", + "[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3", + "[26] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 7", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3", + "[28] Ziying Song, Feiyang Jia, Hongyu Pan, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lin Liu, Yang Ji, Lei Yang, and Li Wang. Contrastalign: Toward robust bev feature alignment via contrastive learning for multi-modal 3d object detection. arXiv preprint arXiv:2405.16873, 2024. 3", + "[29] OD Team et al. Openpcdet: An open-source toolbox for 3d object detection from point clouds, 2020. 7", + "[30] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 4", + "[31] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 1, 2", + "[32] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6792-6802, 2023. 3", + "[33] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7", + "[34] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 1, 3", + "[35] Hai Wu, Shijia Zhao, Xun Huang, Chenglu Wen, Xin Li, and Cheng Wang. Commonsense prototype for outdoor unsupervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14968-14977, 2024. 2, 4, 5, 8", + "[36] Qiming Xia, Yidong Chen, Guorong Cai, Guikun Chen, Daoshun Xie, Jinhe Su, and Zongyue Wang. 3-d hanet: A" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1", + "[37] Qiming Xia, Jinhao Deng, Chenglu Wen, Hai Wu, Shaoshuai Shi, Xin Li, and Cheng Wang. Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6254-6263, 2023. 1, 2, 3, 5, 6, 7, 8", + "[38] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 8", + "[39] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3", + "[40] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 1, 3", + "[41] Yuxue Yang, Lue Fan, and Zhaoxiang Zhang. Mixsup: Mixed-grained supervision for label-efficient lidar-based 3d object detection. arXiv preprint arXiv:2401.16305, 2024. 3", + "[42] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 7", + "[43] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8373-8383, 2023. 2", + "[44] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: zero-shot 3d object detection via the segment anything model. Science China Information Sciences, 67(4):149101, 2024. 3, 8", + "[45] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9317-9328, 2023. 2, 4, 5", + "[46] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1, 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Xiao Zhang, Wenda Xu, Chiyu Dong, and John M Dolan. Efficient l-shape fitting for vehicle detection using laser scanners. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 54-59. IEEE, 2017. 5", + "[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembleing semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2", + "[49] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything, 2023. 3, 6", + "[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 1, 3", + "[51] Pufan Zou, Shijia Zhao, Weijie Huang, Qiming Xia, Chenglu Wen, Wei Li, and Cheng Wang. Adaco: Overcoming visual foundation model noise in 3d semantic segmentation via adaptive label correction. arXiv preprint arXiv:2412.18255, 2024. 3" + ], + "bbox": [ + 89, + 92, + 482, + 400 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29384", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_model.json b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_model.json new file mode 100644 index 0000000000000000000000000000000000000000..186189d79ea25732ab0ffb3e97d4c886a4fd2eb5 --- /dev/null +++ b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_model.json @@ -0,0 +1,2059 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.202, + 0.847, + 0.24 + ], + "angle": 0, + "content": "Shijia Zhao\\(^{1,2*}\\) Qiming Xia\\(^{1,2*}\\) Xusheng Guo\\(^{1,2}\\) Pufan Zou\\(^{1,2}\\) Maoji Zheng\\(^{1,2}\\) Hai Wu\\(^{1,2}\\) Chenglu Wen\\(^{1,2\\dagger}\\) Cheng Wang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.24, + 0.922, + 0.294 + ], + "angle": 0, + "content": "\\(^{1}\\)Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, China \n\\(^{2}\\) Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.486, + 0.782 + ], + "angle": 0, + "content": "Recently, sparsely-supervised 3D object detection has gained great attention, achieving performance close to fully-supervised 3D detectors while requiring only a few annotated instances. Nevertheless, these methods suffer challenges when accurate labels are extremely absent. In this paper, we propose a boosting strategy, termed SP3D, explicitly utilizing the cross-modal semantic prompts generated from Large Multimodal Models (LMMs) to boost the 3D detector with robust feature discrimination capability under sparse annotation settings. Specifically, we first develop a Confident Points Semantic Transfer (CPST) module that generates accurate cross-modal semantic prompts through boundary-constrained center cluster selection. Based on these accurate semantic prompts, which we treat as seed points, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module to yield pseudo-supervision signals from the geometry shape of multi-scale neighbor points. Additionally, we design a Distribution Shape score (DS score) that chooses high-quality supervision signals for the initial training of the 3D detector. Experiments on the KITTI dataset and Waymo Open Dataset (WOD) have validated that SP3D can enhance the performance of sparsely supervised detectors by a large margin under meager labeling conditions. Moreover, we verified SP3D in the zero-shot setting, where its performance exceeded that of the state-of-the-art methods. The code is available at https://github.com/xmuqimingxia/SP3D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.81, + 0.222, + 0.826 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.484, + 0.867 + ], + "angle": 0, + "content": "3D object detection, aiming at locating and classifying objects within 3D scenes, has garnered significant attention in" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.329, + 0.906, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Performance comparison of the sparsely-supervised detector at various annotation rates. SP3D indicates the CoIn initialized with the proposed SP3D." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.506, + 0.907, + 0.567 + ], + "angle": 0, + "content": "autonomous driving [6, 10, 19, 34, 36]. However, the mainstream 3D detectors relies heavily on high-quality labels annotated by humans, which is not only time-consuming but also sensitive to the subjective impression of annotators." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.909, + 0.779 + ], + "angle": 0, + "content": "To minimize the dependence of 3D detectors on high-quality manual annotations, recent work has begun to focus on label-efficient training strategies [15, 21, 31, 37]. To discover unlabeled instances, SS3D [14] employs a self-training approach to iteratively optimize the detector trained on sparsely annotated data. CoIn [37] introduces contrastive learning methods, enhancing the model's discriminative capability for various category features. However, existing strategies make 3D detectors struggle to extract sufficiently discriminative features from extremely limited annotations. Fig. 1 shows some examples where the state-of-the-art (SoTA) sparsely-supervised object detector CoIn [37] hardly maintains robust performance with a significant reduction in annotation rate." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "With the successive emergence and widespread application of large language models (LLMs), such as BERT [7] and GPT [1, 3], in natural language processing, research on large multimodal models (LMMs) is also gaining momentum [12, 13, 17, 27]. Benefiting from the outstanding performance of LMMs, the utilization of pre-trained LMMs has led to significant advancements in 2D vision tasks. Inspired by this, [40, 46, 50] transfers the image-text knowl" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.383, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author, clwen@xmu.edu.cn." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.383, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.282, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.089, + 0.462, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.216, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Figure 2. Semantics belonging to a foreground object may be incorrectly assigned to background/other objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.277, + 0.483, + 0.382 + ], + "angle": 0, + "content": "edge prior from 2D LMMs to 3D point clouds. However, these methods typically focus on classifying individual instances, and there will be certain limitations when applying them directly to outdoor 3D object detection. Despite this, these attempts to transfer the priors from 2D LMMs to 3D point clouds still provide a new perspective for solving the problem of sparsely-supervised 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.385, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Motivated by these methods, we proposed a two-stage training strategy, SP3D, boosting sparsely-supervised 3D object detection with cross-modal semantic prompts generated by LMMs. As shown in Fig. 3, we first employ LMMs to extract semantics from 2D images and explicitly transfer them to 3D point clouds, generating pseudolabels for the first stage of detector training. In the second stage, we fine-tune the trained detector with sparse accurate labels. However, this enhanced paradigm still holds some challenges: (1) Semantic ambiguity at instance edges. Due to the absence of inherent depth information in images, directly transferring image semantics onto point clouds may result in noisy semantic prompts at the edge of the instance. (2) Efficient generation of pseudo-labels. Based on the obtained semantics, directly fitting pseudo-labels may result in incomplete foreground bounding boxes. Moreover, it is difficult to assess the quality of the generated pseudo-labels because of the lack of ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned challenges, our SP3D first designs a Confident Points Semantic Transfer (CPST) module, employing boundary-constrained center cluster selection operation to filter out inaccurate edges to obtain accurate cross-modal semantic prompts. These semantic prompts focus on central foreground semantic masks generated by LMMs. Inspired by unsupervised algorithms [35, 45], we can utilize these as accurate semantic prompts as seed points to generate bounding box pseudo-labels. In this case, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score) to efficiently discover and assess high-quality pseudo-labels with complete foreground information from seed points. As shown in Fig. 3, we utilize the generated pseudo-labels to train the 3D object detector for the first stage. After training, the 3D detector has learned a certain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "feature discrimination capability from the 2D images. Subsequently, we fine-tune the 3D detector with sparse accurate labels, and in conjunction with current label-efficient methods, it demonstrates relatively high detection capabilities even under extremely low labeling scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Experiment results on the KITTI and Waymo datasets show that SP3D substantially enhances the performance of SoTA sparsely-supervised 3D object detectors. Moreover, without fine-tuning on labeled data, SP3D has also shown superior performance in a zero-shot setting, indicating the efficiency of the SP3D-initialized detector. The contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.904, + 0.335 + ], + "angle": 0, + "content": "- We propose a two-stage boosting strategy, SP3D, utilizing accurate cross-modal prompts to boost the feature discrimination capability of 3D detectors under sparsely-supervised situations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.905, + 0.396 + ], + "angle": 0, + "content": "- We propose a Confident Points Semantic Transfer (CPST) module for accurate cross-modal semantic prompt generation and constraint, reducing the noise interference in the semantic transfer process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.472 + ], + "angle": 0, + "content": "- We propose a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score), which generate and assess high-quality pseudolabels based on semantic prompts to improve the detect performance." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.276, + 0.905, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.651, + 0.505 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.821, + 0.532 + ], + "angle": 0, + "content": "2.1. Label-efficient 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recently, label-efficient 3D object detection methods have begun to be explored in responding to the challenge of extremely low annotation volumes. Generally, these label-efficient methods can be categorized into semi-supervised [15, 24, 31, 48], and weakly-supervised [16, 21, 26, 43] approaches according to the difference in quantity and supervision form. In semi-supervised methods, full annotations are provided only for some scenes, while weakly-supervised methods typically adopt weaker annotation forms. For example, VS3D [26] leverages detectors on images to guide unsupervised 3D object detection, and WS3D [21] employs point annotation as a supervision signal. To maintain accuracy while reducing the annotations, SS3D [14] introduces the concept of sparse supervision, annotating only one complete 3D object per frame. CoIn [37] then adopts a contrastive instance feature mining strategy to extract the feature-level pseudo-labels from a significantly reduced amount of annotated data. However, the performance of existing methods remains constrained due to the insufficient feature discriminability of the initial detector, which may affect subsequent training under very few annotations. This work aims to develop a two-stage strategy, enabling the 3D detector to maintain robust feature representation capabilities despite having lower instance annotation ratio." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29375" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.376, + 0.107 + ], + "angle": 0, + "content": "2.2. Large Multimodal Models in 3D" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.37 + ], + "angle": 0, + "content": "As the outstanding performance achieved by LMMs in 2D tasks [2, 12, 18, 25, 27], some studies have begun to explore their application in the 3D domain. Inspired by CLIP [27], ULIP [40] enhances the 3D understanding capability by transferring knowledge from 2D LMM to 3D encoder through contrastive learning methods. Similar works are [46, 50]. In the outdoor scenario, SAM3D [44] employs SAM to segment BEV images of point clouds and fit bounding boxes based on the segmentation masks to obtain detection results. CLIP2Scene [5] establishes the connection between point clouds and text by using images as an intermediate modality, enhancing the 3D model's semantic understanding of the scene with the prior knowledge of CLIP. Unlike previous approaches, our SP3D explicitly transfers semantic masks obtained from LMMs onto point clouds to generate high-quality pseudo-labels for the first-stage training of the 3D detector." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.38, + 0.415, + 0.396 + ], + "angle": 0, + "content": "2.3. Multimodal Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.645 + ], + "angle": 0, + "content": "Recently, using multimodal methods from 2D images and 3D point clouds to enhance 3D detectors has gradually gained the community's attention [19, 28, 32, 34, 39]. However, these works mainly focus on investigating the image-point cloud fusion strategy, neglecting the utilization of images to explore label-efficient 3D detection. To reduce the required annotations, some methods transfer image information into point clouds to generate pseudo-labels [41, 51]. However, semantic ambiguity may occur at the object's edge due to the 2D-3D calibration error. MixSup [41] proposed a connected components labeling strategy, addressing this issue with the spatial separability property inherent to point clouds. SAL [23] employs density-based clustering to refine imperfect projection issues. Compared with these methods, our SP3D provides a simple but efficient way to reduce semantic noise caused by projection errors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.191, + 0.673 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.203, + 0.697 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This paper introduces SP3D, explicitly utilizing the accurate cross-modal semantic prompts generated by LMMs to boost the sparsely-supervised 3D detectors. As shown in Fig. 3, SP3D consists of three primary components: (1) Confident Points Semantic Transfer (CPST) module, acquiring accurate cross-modal semantic prompts, which we treat as seed points, with the boundary-constrained center cluster selection. (2) Dynamic Cluster Pseudo-label Generation (DCPG) module, which dynamically generates pseudo-label proposals based on the geometric shapes within the multi-scale neighborhood of these seed points. (3) Distribution Shape score (DS score), which employs unsupervised priors as the criterion for evaluating the qual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "ity of pseudo-label proposals, and we subsequently apply NMS [22] to retain high-quality pseudo-labels further. We then follow the training strategy of CoIn [37] to train an initial 3D detector with enhanced discriminative capacity. We detail our SP3D framework as follows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.176, + 0.887, + 0.192 + ], + "angle": 0, + "content": "3.2. Confident Points Semantic Transfer Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Encouraged by the development of LMMs, we first utilize the prior knowledge of LMMs to extract semantic information from 2D images explicitly. Meanwhile, by integrating the calibration relationship matrix between 2D images and 3D point clouds, it is straightforward to transfer semantic information onto point clouds. However, as shown in Fig. 2, there is significant noise in the edge points of instances during the transfer process. These semantic points are mistakenly assigned to other instances can have a detrimental impact on subsequent pseudo-label fitting and the training of the detector. To prevent the incorporation of noise during the transfer of semantic information, we use the boundary-constrained mask shrink operation, followed by the coordinate system transformation, to obtain accurate cross-modal semantic prompts. CSPT is illustrated in Fig. 3(a); we have divided it into two parts as LMMs-guided semantic extraction and confident points filtering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.58 + ], + "angle": 0, + "content": "LMMs-guided semantic extraction. The goal of LMMs is to generate high-quality foreground semantic masks. Specifically, we take as input an image \\(\\mathcal{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\) and the category-text prompt \\(\\mathcal{T}^{\\mathcal{C}} = \\{t_1^{\\mathcal{C}}, t_2^{\\mathcal{C}}, \\dots, t_C^{\\mathcal{C}}\\}\\), where \\(H\\) and \\(W\\) denote the height and width of the image, \\(C\\) denotes the categories of interest. We first utilize FastSAM [49] to perform speed-efficient segmentation as" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.59, + 0.905, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {\\mathcal {I}} = \\operatorname {S A M} (\\mathcal {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.616, + 0.906, + 0.709 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_{\\mathcal{I}} \\in \\mathbb{R}^{M \\times H \\times W}\\) denotes the \\(M\\) class-agnostic masks extracted from \\(\\mathcal{I}\\), \\(\\mathrm{SAM}(\\cdot)\\) indicates the FastSAM model. We then utilize \\(\\mathcal{M}_{\\mathcal{I}}\\) as the mask prompts and feed them, along with image \\(\\mathcal{I}\\), into the SemanticSAM [4] model, which is expected to output the descriptions \\(\\mathcal{T}^{\\mathcal{D}} = \\{t_1^{\\mathcal{D}}, t_2^{\\mathcal{D}}, \\dots, t_M^{\\mathcal{D}}\\}\\) for each mask in \\(\\mathcal{M}_{\\mathcal{I}}\\). Specifically:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.719, + 0.905, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {\\mathcal {D}} = \\operatorname {S S A M} (\\mathcal {I}, \\mathcal {M} _ {\\mathcal {I}}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.747, + 0.906, + 0.809 + ], + "angle": 0, + "content": "where \\(\\mathrm{SSAM}(\\cdot)\\) refers to the SemanticSAM model. Then, we calculate the cosine similarity between the category-text \\(\\mathcal{T}^{\\mathcal{C}}\\) and the mask description \\(\\mathcal{T}^{\\mathcal{D}}\\) to filter out uninteresting background masks, obtaining the foreground masks \\(\\mathcal{M}_{\\mathcal{I}}'\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Confident points filtering. As fuzziness of boundaries results from depth occlusions in images and calibration inaccuracies, we opt to constrain the boundary of foreground masks before 2D-3D transfer, retaining only its central portion. Specifically, for each foreground mask, we denote its" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29376" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.904, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.908, + 0.459 + ], + "angle": 0, + "content": "Figure 3. The overview of our SP3D, including (a) CSPT finds semantic seed points through high-confidence semantic masks transfer, (b) DCPG dynamically clusters neighbor points of seed points to fit pseudo-label proposals, and (c) DS score to evaluate the quality of generated pseudo-label proposals, serving as a scoring metric to NMS to suppress low-quality proposals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.483, + 0.531 + ], + "angle": 0, + "content": "maximum and minimum values in the pixel coordinate system \\((u,v)\\) as \\((u_{min},u_{max},v_{min},v_{max})\\). We perform mask shrink to constraint boundary range of \\(\\mathcal{M}_{\\mathcal{I}}^{\\prime}\\) to obtain \\(\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.542, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\nu \\in \\left[ u _ {\\min } + \\frac {1}{2} (1 \\pm \\gamma) \\left(u _ {\\max } - u _ {\\min }\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.574, + 0.415, + 0.601 + ], + "angle": 0, + "content": "\\[\nv \\in [ v _ {m i n} + \\frac {1}{2} (1 \\pm \\gamma) (v _ {m a x} - v _ {m i n}) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.822 + ], + "angle": 0, + "content": "where \\(\\gamma\\) denotes the shrink factor. With this constraint, we obtain the shrunk masks that filter out the semantically ambiguous regions, ensuring the accuracy of foreground semantic information transferred onto the point clouds. Following [30], we transfer the semantic mask from the image onto the point clouds to obtain accurate cross-modal semantic prompts, namely seed points, using the camera's intrinsic and extrinsic parameter matrices. Note that, different from [30], we adopt the approach of explicitly transferring the shrunk masks onto the point cloud rather than implicitly embedding unprocessed semantic masks into the point cloud's features. This approach helps avoid potential semantic feature confusion between different modalities arising from sparse annotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.452, + 0.85 + ], + "angle": 0, + "content": "3.3. Dynamic Cluster Pseudo-label Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "With the proposed CPST, we explicitly obtain the semantic seed points from transformation \\(\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}\\) of the foreground mask. Given \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times 3} = \\{p_1, p_2, \\dots, p_N\\}\\) as LiDAR" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.484, + 0.907, + 0.529 + ], + "angle": 0, + "content": "points, we define the seed points covered by \\(\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}\\) as \\(\\mathcal{P}_T = \\{p_t\\}\\), \\(\\mathcal{P}_T \\subset \\mathcal{P}\\). It is crucial for the 3D detection task that obtain complete bounding boxes from these seed points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.741 + ], + "angle": 0, + "content": "By referring to traditional unsupervised pseudo-label generation methods [45], we produce many pseudo-labels and then use the positional constraints of the seed points to retain the more promising pseudo-labels as supervisory signals. However, existing unsupervised bounding box fitting approaches [35, 45] usually take a fixed constant as cluster radius, leading to the problem of inadequate foreground or excessive background noise for the generated bounding boxes. In this case, we propose a dynamic cluster pseudo-label generation (DCPG) module. This module utilizes the geometry shape of the seed points' multi-scale neighborhood to capture complete foreground information while minimizing background interference. It dynamically generates pseudo-label proposals, as shown in Fig. 3(b)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.742, + 0.906, + 0.807 + ], + "angle": 0, + "content": "Specifically, we denote \\(\\mathcal{P}_T^{(k)}\\) as the \\(k\\)-th instance in a point cloud frame and utilize DCPG dynamically generates a clustering radius \\(r\\) for the \\(t\\)-th seed point \\(p_t^{(k)}\\). We define the updating function for the dynamic radius as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.817, + 0.907, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\operatorname {u p d a t e} (t, r _ {\\text {i n i t}}) = r _ {\\text {i n i t}} \\cdot \\frac {t}{N ^ {(k)}} + \\delta , t = 1, 2, \\dots , N ^ {(k)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\( r_{init} \\) is a hyper-parameter set based on empirical experience, \\( \\delta \\) denotes adjustment factor to avoid \\( r \\) too small, and \\( N^{(k)} \\) is the number of seed points in current instance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29377" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.454, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Dynamic cluster pseudo-label generation" + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.112, + 0.484, + 0.336 + ], + "angle": 0, + "content": "Require: LiDAR points \\(\\mathcal{P}\\), the seed points set \\(\\mathcal{P}_T^{(k)}\\) of the \\(k\\)-th instance, initial radius \\(r_{init}\\), ground remove function \\(\\mathcal{G}\\), box fitting function \\(\\mathcal{F}\\); \nEnsure: Pseudo-label proposal set \\(\\hat{\\mathcal{B}}^{(k)}\\); \n1: Acquire the number \\(N^{(k)}\\) of points for the seed points set of current instance \\(\\mathcal{P}_T^{(k)}\\); \n2: \\(\\mathcal{P}_{gr} \\gets \\mathcal{G}(\\mathcal{P})\\); // remove ground points \n3: \\(\\hat{\\mathcal{B}}^{(k)} = \\emptyset\\); \n4: for \\(t = 1, 2, \\dots, N^{(k)}\\) do \n5: \\(p_t = \\mathcal{P}_T^{(k)}[t]\\); // loop each seed point \n6: \\(r \\gets \\text{update}(t, r_{init})\\) \n7: \\(\\hat{\\mathbf{b}} \\gets \\mathcal{F}(\\text{DBSCAN}(\\mathcal{P}_{gr}, p_t, r))\\) \n8: \\(\\hat{\\mathcal{B}}^{(k)} \\gets \\hat{\\mathcal{B}}^{(k)} \\cup \\hat{\\mathbf{b}}\\) \n9: end for" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.46 + ], + "angle": 0, + "content": "By applying Eq. 4, we dynamically update the radius \\( r \\), \\( r \\in (\\delta, r_{init} + \\delta] \\), during point clustering, thereby obtaining foreground clusters with multi-scale receptive fields. Following [45]. We utilize the radius calculated from Eq. 4 as the clustering radius for DBSCAN [8] and employ [47] to fit the bounding box for each foreground cluster. Algorithm 1 summarizes Our DCPG." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.32, + 0.484 + ], + "angle": 0, + "content": "3.4. Distribution Shape Score" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.745 + ], + "angle": 0, + "content": "While DCPG has the capacity for high-quality pseudolabel generation, it unavoidably produces an amount of low-quality pseudo-label proposals. The shapes of these proposals and the extent of foreground completeness contained within the proposals usually deviate significantly from reality. Traditional detection methods typically compute the Intersection over Union (IoU) score between predicted bounding boxes and ground-truth (GT) boxes and then employ the NMS [22] to suppress these low-quality proposals. However, lacking GT makes it challenging to directly apply NMS using IoU as the evaluation criterion within our SP3D framework. In this case, we propose a distribution shape score (DS score) to assess the quality of the pseudo-labels using unsupervised prior knowledge. As illustrated in Fig. 3(c), we divided the DS score into two parts: distribution constraint score and meta-shape constraint score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Distribution constraint score. Inspired by [20], within a high-quality pseudo-label proposal \\(\\hat{b}\\), the distances from its interior points \\(p_{i,i=1,\\dots,n}\\) to its boundary roughly follow a Gaussian distribution \\(\\mathcal{N}(\\mu, \\sigma)\\), where \\(\\mu = 0.8\\) and \\(\\sigma = 0.2\\), respectively. In other words, we denote random variable \\(D = \\{d_1, \\dots, d_n\\}\\) as the distance between \\(p_i\\) and the box boundary of \\(\\hat{b}\\), and \\(D \\sim \\mathcal{N}(0.8, 0.2)\\). Based on this prior, we assign a distribution constraint score to the pseudo-label proposal \\(\\hat{b}\\) by calculating the similarity between the random" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.122 + ], + "angle": 0, + "content": "variable \\(D\\) corresponding to each \\(\\hat{b}\\) and the normal distribution \\(\\mathcal{N}\\). Specifically:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.132, + 0.907, + 0.171 + ], + "angle": 0, + "content": "\\[\ns _ {d c} (\\hat {b}) = \\frac {1}{| \\mathcal {P} _ {f g} |} \\sum_ {p _ {i} \\in \\mathcal {P} _ {f g}} \\log (\\mathcal {N} (D | \\mu , \\sigma)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.181, + 0.907, + 0.231 + ], + "angle": 0, + "content": "where \\(\\log (\\cdot)\\) denotes the logarithm function, \\(\\mathcal{P}_{fg}\\) is the foreground points within \\(\\hat{b}\\), and \\(|\\mathcal{P}_{fg}|\\) represents the number of points in \\(\\mathcal{P}_{fg}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.907, + 0.383 + ], + "angle": 0, + "content": "Meta-shape constraint score. In addition, the shape of a high-quality pseudo-label is expected to be consistent with its template in the real world, which we define as the meta instance, corresponding to its category [35]. For class \\(c\\), we denote \\(\\mathcal{B}_c \\in \\{l_c, w_c, h_c\\}\\) as the shape of its meta instance, where \\(l_c\\), \\(w_c\\) and \\(h_c\\) are the normalized length, width and height, respectively. We followed this shape prior to constructing the class-aware meta-shape constraint score \\(s_{msc}(\\hat{b})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.394, + 0.907, + 0.414 + ], + "angle": 0, + "content": "\\[\ns _ {m s c} (\\hat {b}) = 1 - \\Phi_ {K L} (\\mathcal {B} _ {c} | | \\hat {\\mathcal {B}} _ {\\hat {b}}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.528 + ], + "angle": 0, + "content": "where \\(\\Phi_{KL}(\\cdot)\\) denotes the normalized KL divergence function, and \\(\\hat{\\mathcal{B}}_{\\hat{b}} \\in \\{l_{\\hat{b}}, w_{\\hat{b}}, h_{\\hat{b}}\\}\\) indicates the normalized shape of the pseudo-label proposal. The purpose of this operation is to suppress the low-quality proposals whose shape deviates significantly from the meta instance. By combining the distribution constraint score and the meta-shape constraint score, we can obtain the DS score as" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.54, + 0.907, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D S} (\\hat {b}) = \\lambda_ {1} \\bar {s} _ {d c} (\\hat {b}) + \\lambda_ {2} \\bar {s} _ {m s c} (\\hat {b}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.705 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) denote weight adjustment factor. Notably, to unify the dimension, we normalized the two constraint scores before combining them, resulting in \\(\\overline{s}_{dc}\\) and \\(\\overline{s}_{msc}\\). We then employ the DS score as a substitution for the confidence score in NMS to suppress the low-quality pseudolabels. We utilize the obtained pseudo-labels in conjunction with CoIn [37] for the first phase of training. Subsequently, we fine-tune the trained detector with a small amount of accurate labels to boost the performance of the 3D detector." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.719, + 0.646, + 0.735 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.706, + 0.758 + ], + "angle": 0, + "content": "4.1. Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "KITTI dataset. The KITTI [9] dataset is a widely used benchmark for 3D object detection. During the first training stage, we relied solely on the semantic information provided by LMMs to generate pseudo-labels instead of the human-annotated ground truth. In the fine-tuning stage, we split the training set (7,481 scenes) into a train split (3,712 scenes) and a val split (3,769 scenes). We follow [37] to randomly select \\(10\\%\\) of the scenes from the train split and retain only one instance annotation per scene. In this case," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29378" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.895, + 0.233 + ], + "angle": 0, + "content": "
SettingCostMethodCarPedCyc
EasyMod.HardEasyMod.HardEasyMod.Hard
Fully-sup.100%VoxelRCNN92.384.982.669.663.058.688.772.568.2
Sparsely-sup.20%SS3D89.384.278.2------
2%VoxelRCNN70.554.944.842.638.532.173.347.843.2
2%CoIn89.170.255.650.845.239.680.252.348.6
2%CoIn++92.079.571.546.736.131.282.058.454.6
2%CoIn++ with SP3D91.380.574.067.458.750.992.573.168.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.246, + 0.907, + 0.276 + ], + "angle": 0, + "content": "Table 1. Comparison with SoTA sparsely-supervised methods on KITTI val split. All methods are based on VoxelRCNN, and we report the 3D AP results of full cost (100%) and limited cost (20%, 2%). The best sparsely-supervised methods are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.288, + 0.92, + 0.511 + ], + "angle": 0, + "content": "
StageLabelMethodsCar-3D @IoU 0.7Car-BEV @IoU 0.7
EasyModHardEasyModHard
Single-stageSparse1. CenterPoint49.6931.5525.9156.7842.5034.14
Sparse2. CoIn (CenterPoint-based)72.0354.8243.7787.2073.5466.03
Sparse3. 2 with SP3D (CenterPoint-based)87.4469.2458.6192.7280.0069.01
-4. Improvements 3→2+15.41+14.42+14.84+5.52+6.46+2.98
Two-stageSparse1. Voxel-RCNN70.5254.9744.8283.6771.1457.71
Sparse2. CoIn (Voxel-RCNN-based)84.5668.4758.0292.3181.0170.24
Sparse3. 2 with SP3D (Voxel-RCNN-based)91.3774.8963.8495.4185.2774.57
-4. Improvements 3→2+6.81+6.42+5.82+3.1+4.26+4.33
Multi-stageSparse1. CasA74.1857.3745.0585.9073.2157.23
Sparse2. CoIn (CasA-based)89.1775.3262.9895.9985.0272.47
Sparse3. 2 with SP3D (CasA-based)91.1275.9466.4696.5585.6576.31
-4. Improvements 3→2+1.95+0.62+3.48+0.56+0.63+3.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.522, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Table 2. Comparison with different fully-supervised methods. Sparse label refers to the use of limited split (2% annotation cost). The 3D object detection and BEV detection benchmark are evaluated by mAP with R40, under IoU threshold 0.7." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.575, + 0.462, + 0.661 + ], + "angle": 0, + "content": "
MethodsVeh. LEVEL 1Veh. LEVEL 2
mAPmAPHmAPmAPH
CoIn39.639.034.233.6
CoIn+SP3D46.746.040.439.9
Improvement+7.1+7.0+6.2+6.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.714 + ], + "angle": 0, + "content": "Table 3. Comparison with SoTA sparsely-supervised methods on WOD validation set. All methods are based on CenterPoint, and we report the results at a \\(1\\%\\) annotation cost and 0.7 IoU threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.808 + ], + "angle": 0, + "content": "we can obtain a limited split, which merely takes \\(2\\%\\) of instance annotations compared with the origin train split. All results are evaluated using the official evaluation metric: 3D Average Precision (AP) across 40 recall thresholds (R40)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Waymo Open Dataset (WOD). The WOD is a widely used large-scale benchmark that contains 798, 202, and 150 sequences (approximately 200 frames per sequence) for training, validation, and testing, respectively. All results are evaluated by using 3D mean Average Precision (mAP) and its variant weighted by heading accuracy (mAPH)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.733, + 0.592 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Pseudo-labels generation. We directly employed the model parameters provided by FastSAM [49] and SemanticSAM [4] for inference without additional supervisory signals for fine-tuning. To achieve accurate segmentation results, we set a higher segmentation threshold of 0.7 during the FastSAM inference process. To mitigate computational demands, we opted to generate pseudo-labels within a confined spatial domain of the semantically relevant points, specifically within an 8-meter radius. We set mask shrink factor \\(\\gamma\\) to 0.3, initial cluster radius \\(r_{init}\\) to 1, adjustment factor \\(\\delta\\) to 0.1. We utilize unsupervised priors to filter out the low-quality pseudo-labels that are evidently inconsistent with the intuitive expectations and set the weight adjustment factor of DS score \\(\\lambda_1\\) and \\(\\lambda_2\\) as 0.5, 0.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Detector training. We conduct all experiments with a batch size of 8 and a learning rate of 0.003 on 4 RTX 3090 GPUs. Following previous sparsely-supervised 3D object detection methods [14, 37], we choose three different clas" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29379" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.917, + 0.282 + ], + "angle": 0, + "content": "
Methods2%1%0.5%
CarPed.Cyc.Avg.CarPed.Cyc.Avg.CarPed.Cyc.Avg.
semi-sup.HSSDA81.958.265.868.680.951.942.359.577.819.544.147.2
HSSDA+SP3D83.260.874.472.883.053.269.168.482.322.146.450.3
Improvement+1.3+2.6+8.6+4.2+2.1+1.3+26.8+8.9+4.5+2.6+2.3+3.1
sparsely-sup.CoIn72.336.958.455.865.331.552.249.655.225.530.537.0
CoIn+SP3D73.444.762.360.171.736.959.155.963.336.143.447.6
Improvement+1.1+7.8+3.9+4.3+6.4+5.4+6.9+6.3+8.1+10.6+12.5+10.6
HINTED85.160.081.375.482.849.974.168.980.245.864.563.5
HINTED+SP3D85.560.181.475.684.057.281.074.083.249.273.968.7
Improvement+0.4+0.1+0.1+0.2+1.2+7.3+6.9+5.1+3.0+3.4+9.4+5.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.292, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Table 4. Improvement of SP3D in SoTA semi-supervised and sparsely-supervised methods on KITTI val split under limited cost setting. The results are evaluated by mAP with R40, under IoU thresholds 0.7, 0.5, and 0.5 for 'Car', 'Pedestrian', and 'Cyclist', respectively. Additionally, the AP is computed across the difficulty levels of easy, moderate, and hard." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.358, + 0.499, + 0.442 + ], + "angle": 0, + "content": "
MethodsModalityCar-3D @IOU 0.5/0.7
Train TestEasyMod.Hard
VS3DL+CL40.32/9.0937.36/5.7331.09/5.03
WS3DPRL+CL-/60.01-/44.48-/36.93
SP3D (Ours)L+CL93.75/69.7176.36/48.6571.01/40.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.455, + 0.473, + 0.47 + ], + "angle": 0, + "content": "Table 5. Comparison of zero-shot methods on KITTI val split." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.573 + ], + "angle": 0, + "content": "sical detectors [6, 33, 42] as our architecture. And we employ the OpenPCDet [29] to conduct our experiments. In the first training stage, we employ CoIn [37] to train an initial detector with the generated pseudo-labels. Then, we use limited split to fine-tune the detector." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Baselines. To thoroughly validate the effectiveness of SP3D, we select the SoTA sparsely-supervised method [37] as the primary baseline for comparison. We compare the proposed SP3D approach with the baseline under conventional sparse settings with \\(2\\%\\) annotation cost. We also compared with cross-modal weakly-supervised methods [16, 26] in a zero-shot setting, which also incorporate visual models to extract semantic information to enhance the performance of detectors. Furthermore, we establish baselines under progressively reduced annotation costs to evaluate the sensitivity to annotation costs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.38, + 0.789 + ], + "angle": 0, + "content": "4.3. Comparison with SoTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison with sparsely-supervised methods. We integrate our proposed SP3D into the SoTA sparsely-supervised 3D detection algorithm, CoIn++ [37]. For a fair comparison, all detectors employ the VoxelRCNN [6] as the base architecture. As illustrated in Tab. 1, SP3D significantly improves the detection performance of CoIn++. The experimental results on WOD, as reported in Tab. 3," + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.358, + 0.895, + 0.445 + ], + "angle": 0, + "content": "
MethodsModalityVeh.Ped.Cyc.
TrainTest
SAM3DLL19.10.00.0
CM3DL + CL23.70.12.6
SP3D (Ours)L + CL35.015.75.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.455, + 0.906, + 0.498 + ], + "angle": 0, + "content": "Table 6. Comparison with zero-shot methods on the WOD validation set. We report Level-2 mAP with IoU thresholds of 0.7, 0.5, and 0.5 for Vehicle, Pedestrian, and Cyclist, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.679 + ], + "angle": 0, + "content": "also validate the performance enhancement of our method for the sparsely-supervised 3D object detector. However, we observe a slight decrease in precision for the 'Easy' car category with our SP3D-initialized model in Tab. 1. This could arise because our initial pseudo-labels are inferred based on the geometric shape of the objects, which may differ from the conventions of manual annotation. When the point cloud structure of an instance is relatively intact, such discrepancies can lead to noticeable differences in the size of the annotated bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Comparison with fully-supervised methods. For a fair comparison, we follow CoIn [37] to select CenterPoint [42], VoxelRCNN [6], and CasA [33] as our baseline detectors, representing three distinct types of detection algorithms. We initialize the 3D detector using our SP3D, followed by fine-tuning with the limited split. As shown in Tab. 2, detectors designed for full supervision struggle to yield optimal results due to scarce annotations. Despite the effectiveness of CoIn in improving this situation, the results achieved are still unsatisfactory for single-stage detection algorithms with relatively simple structures. Our designed strategy, SP3D, significantly reduces this discrepancy, enabling detectors to achieve similar results." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29380" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.445, + 0.174 + ], + "angle": 0, + "content": "
MethodsCar-3D @IoU 0.5
EasyMod.Hard
CPD61.4861.5860.09
CPD+SP3D73.6566.1363.78
Improvement+12.17+4.55+2.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.186, + 0.431, + 0.2 + ], + "angle": 0, + "content": "Table 7. Improvement for unsupervised method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.482, + 0.381 + ], + "angle": 0, + "content": "Comparison with zero-shot methods. We also compared SP3D with SoTA zero-shot methods. SAM3D [44] and CM3D [11] generate 3D pseudo-labels by leveraging large models for the semantic segmentation of 2D images. In line with the zero-shot approach of prior methods, no 3D annotations were used for training; instead, pseudolabels were generated from large models and image data. As shown in Tab. 5 and Tab. 6, the results on the KITTI and Waymo datasets demonstrate that our method significantly outperforms previous approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Comparison with different annotation rates. To explore the robustness of SP3D, we conducted experiments on detectors across different supervision paradigms under varying annotation rates, and we reported the results on Tab. 4. For a fair comparison, we initialize the detectors by following the strategies of CoIn (CenterPoint) [37], HSSDA (PV-RCNN) [15], and HINTED (Voxel-RCNN) [38]. The experimental results show that when the annotation rate is limited, detector performance drops sharply as the amount of annotated data decreases. By inheriting SP3D, these detectors' performance has been significantly enhanced, mitigating the performance drop associated with reduced annotation. Taking CoIn as an example, its average AP for the three categories at annotation rates of \\(1\\%\\) and \\(0.5\\%\\) has improved by \\(6.3\\%\\) and \\(10.6\\%\\), respectively. These results indicate that even at low annotation rates, SP3D-initialized models can significantly improve the performance of 3D detectors across different supervision paradigms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Comparison with unsupervised methods. To further validate the robustness of SP3D, we integrated it with the SoTA unsupervised detection method CPD [35]. As shown in Tab. 7, despite suffering some impact, the combination of SP3D and CPD still enhances the performance of the unsupervised 3D object detector to a certain extent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.245, + 0.818 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We conducted ablation studies based on CenterPoint to verify the proposed modules' effectiveness and recorded the results in Tab. 8. The first and third rows show that the precision of pseudo-labels, as augmented by DCPG, can substantially amplify the 3D detector's capabilities. This may be" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.089, + 0.9, + 0.205 + ], + "angle": 0, + "content": "
Mask shrinkDCPGDS scoreCar-3D AP@IoU 0.7
EasyMod.Hard
✓35.1023.7519.52
✓✓38.2629.2524.74
✓✓40.5628.1522.40
✓✓47.2333.4027.13
✓✓✓52.5638.0031.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.216, + 0.838, + 0.23 + ], + "angle": 0, + "content": "Table 8. Ablation study on KITTI val split." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.247, + 0.862, + 0.318 + ], + "angle": 0, + "content": "
DS scoreCar-3D AP@IoU 0.7
MSCDCEasyMod.Hard
✓44.9133.5528.19
✓48.4835.3929.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.329, + 0.875, + 0.343 + ], + "angle": 0, + "content": "Table 9. Ablation study of each component of DS score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.509 + ], + "angle": 0, + "content": "due to incorporating more comprehensive foreground information in pseudo-labels, which enhanced the model's feature discrimination capability. The last two rows show that the mask shrink operation is necessary for handling semantic noise at the instance edges. The first three rows and fifth row show that using the DS score to filter out low-quality labels can significantly boost the detector's precision. Moreover, we validated the effectiveness of each component in DS score with the ablation studies presented in Tab. 9." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.762, + 0.541 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.839 + ], + "angle": 0, + "content": "This paper proposes a two-stage boosting strategy, SP3D, that leverages accurate cross-modality semantic prompts to enhance the sparsely-supervised 3D detectors' capabilities. We first develop a CSPT module to obtain accurate cross-modal semantic prompts in point clouds by efficiently transferring high-confidence semantic masks. We then design a DCPG module to dynamically generate pseudo-label proposals based on semantic prompts. Lastly, we propose a DS score as the criterion for NMS to select high-quality pseudo-labels integrated with sparsely-supervised strategy to train the initial detector. After fine-tuning with sparse GT data, SP3D show superior performance over the original sparsely-supervised 3D detector on the WOD and KITTI dataset, and remain robust performance even as the amount of annotation decreased. Limitations: SP3D exhibits a relatively significant degradation when fine-tuning with annotation rates below \\(0.1\\%\\), which may result from the noise introduced by extremely low annotations. Future efforts to explore efficient fine-tuning strategies to solve this issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (No.62171393), and the Fundamental Research Funds for the Central Universities (No.20720220064)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29381" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[4] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[8] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[10] Xun Huang, Hai Wu, Xin Li, Xiaoliang Fan, Chenglu Wen, and Cheng Wang. Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2409–2416, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[11] Mehar Khurana, Neehar Peri, James Hays, and Deva Ramanan. Shelf-supervised cross-modal pre-training for 3d object detection. In 8th Annual Conference on Robot Learning, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[14] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8428-8437, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[15] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23819-23828, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.33, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[16] Haizhuang Liu, Huimin Ma, Yilin Wang, Bochao Zou, Tianyu Hu, Rongquan Wang, and Jiansheng Chen. Eliminating spatial ambiguity for weakly supervised 3d object detection without spatial labels. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3511-3520, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[20] Katie Luo, Zhenzhen Liu, Xiangyu Chen, Yurong You, Sagie Benaim, Cheng Perng Phoo, Mark Campbell, Wen Sun, Bharath Hariharan, and Kilian Q Weinberger. Reward finetuning for faster and more accurate unsupervised object discovery. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[21] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Yunde Jia, and Luc Van Gool. Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4454-4468, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[22] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[23] Aljosa Osep, Tim Meinhardt, Francesco Ferroni, Neehar Peri, Deva Ramanan, and Laura Leal-Taixe. Better call sal: Towards learning to segment anything in lidar. In European Conference on Computer Vision, pages 71-90. Springer, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Jinhyung Park, Chenfeng Xu, Yiyang Zhou, Masayoshi Tomizuka, and Wei Zhan. Detmatch: Two teachers are better than one for joint 2d and 3d semi-supervised object detection. In European Conference on Computer Vision, pages 370-389. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[26] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[28] Ziying Song, Feiyang Jia, Hongyu Pan, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lin Liu, Yang Ji, Lei Yang, and Li Wang. Contrastalign: Toward robust bev feature alignment via contrastive learning for multi-modal 3d object detection. arXiv preprint arXiv:2405.16873, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[29] OD Team et al. Openpcdet: An open-source toolbox for 3d object detection from point clouds, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[30] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[31] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[32] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6792-6802, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[33] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[34] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Hai Wu, Shijia Zhao, Xun Huang, Chenglu Wen, Xin Li, and Cheng Wang. Commonsense prototype for outdoor unsupervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14968-14977, 2024. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Qiming Xia, Yidong Chen, Guorong Cai, Guikun Chen, Daoshun Xie, Jinhe Su, and Zongyue Wang. 3-d hanet: A" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[37] Qiming Xia, Jinhao Deng, Chenglu Wen, Hai Wu, Shaoshuai Shi, Xin Li, and Cheng Wang. Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6254-6263, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[38] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[39] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[40] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[41] Yuxue Yang, Lue Fan, and Zhaoxiang Zhang. Mixsup: Mixed-grained supervision for label-efficient lidar-based 3d object detection. arXiv preprint arXiv:2401.16305, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[42] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[43] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8373-8383, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[44] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: zero-shot 3d object detection via the segment anything model. Science China Information Sciences, 67(4):149101, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[45] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9317-9328, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29383" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[47] Xiao Zhang, Wenda Xu, Chiyu Dong, and John M Dolan. Efficient l-shape fitting for vehicle detection using laser scanners. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 54-59. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembleing semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[49] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.25, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[51] Pufan Zou, Shijia Zhao, Weijie Huang, Qiming Xia, Chenglu Wen, Wei Li, and Cheng Wang. Adaco: Overcoming visual foundation model noise in 3d semantic segmentation via adaptive label correction. arXiv preprint arXiv:2412.18255, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "29384" + } + ] +] \ No newline at end of file diff --git a/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_origin.pdf b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74691f3af067899a39debb3b9dff31d7748de921 --- /dev/null +++ b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/dd67abde-1a08-4102-9911-744b35ee1a19_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d72fae41c3c7aab69f793ed515d369b4e878af173a4f771bc2df36d33e578a0 +size 819325 diff --git a/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/full.md b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..197271404465af20f36067760b69cfa90f4c756f --- /dev/null +++ b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/full.md @@ -0,0 +1,299 @@ +# SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts + +Shijia Zhao $^{1,2*}$ Qiming Xia $^{1,2*}$ Xusheng Guo $^{1,2}$ Pufan Zou $^{1,2}$ Maoji Zheng $^{1,2}$ Hai Wu $^{1,2}$ Chenglu Wen $^{1,2\dagger}$ Cheng Wang $^{1,2}$ + +$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, China + $^{2}$ Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China + +# Abstract + +Recently, sparsely-supervised 3D object detection has gained great attention, achieving performance close to fully-supervised 3D detectors while requiring only a few annotated instances. Nevertheless, these methods suffer challenges when accurate labels are extremely absent. In this paper, we propose a boosting strategy, termed SP3D, explicitly utilizing the cross-modal semantic prompts generated from Large Multimodal Models (LMMs) to boost the 3D detector with robust feature discrimination capability under sparse annotation settings. Specifically, we first develop a Confident Points Semantic Transfer (CPST) module that generates accurate cross-modal semantic prompts through boundary-constrained center cluster selection. Based on these accurate semantic prompts, which we treat as seed points, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module to yield pseudo-supervision signals from the geometry shape of multi-scale neighbor points. Additionally, we design a Distribution Shape score (DS score) that chooses high-quality supervision signals for the initial training of the 3D detector. Experiments on the KITTI dataset and Waymo Open Dataset (WOD) have validated that SP3D can enhance the performance of sparsely supervised detectors by a large margin under meager labeling conditions. Moreover, we verified SP3D in the zero-shot setting, where its performance exceeded that of the state-of-the-art methods. The code is available at https://github.com/xmuqimingxia/SP3D. + +# 1. Introduction + +3D object detection, aiming at locating and classifying objects within 3D scenes, has garnered significant attention in + +![](images/305b3135f0d33c79ca91e5dd308aa752cbda1c0b1956d456f2d7f149cac9ea0a.jpg) +Figure 1. Performance comparison of the sparsely-supervised detector at various annotation rates. SP3D indicates the CoIn initialized with the proposed SP3D. + +autonomous driving [6, 10, 19, 34, 36]. However, the mainstream 3D detectors relies heavily on high-quality labels annotated by humans, which is not only time-consuming but also sensitive to the subjective impression of annotators. + +To minimize the dependence of 3D detectors on high-quality manual annotations, recent work has begun to focus on label-efficient training strategies [15, 21, 31, 37]. To discover unlabeled instances, SS3D [14] employs a self-training approach to iteratively optimize the detector trained on sparsely annotated data. CoIn [37] introduces contrastive learning methods, enhancing the model's discriminative capability for various category features. However, existing strategies make 3D detectors struggle to extract sufficiently discriminative features from extremely limited annotations. Fig. 1 shows some examples where the state-of-the-art (SoTA) sparsely-supervised object detector CoIn [37] hardly maintains robust performance with a significant reduction in annotation rate. + +With the successive emergence and widespread application of large language models (LLMs), such as BERT [7] and GPT [1, 3], in natural language processing, research on large multimodal models (LMMs) is also gaining momentum [12, 13, 17, 27]. Benefiting from the outstanding performance of LMMs, the utilization of pre-trained LMMs has led to significant advancements in 2D vision tasks. Inspired by this, [40, 46, 50] transfers the image-text knowl + +![](images/df33c4bea8e9b9886939ae7a97635bd84acf2a09e4bc978ef3130af63bd63bb4.jpg) +Figure 2. Semantics belonging to a foreground object may be incorrectly assigned to background/other objects. + +![](images/3d02f993f13d66a59661380e7aa38a8becec8642b55e445381b8847698022c60.jpg) + +edge prior from 2D LMMs to 3D point clouds. However, these methods typically focus on classifying individual instances, and there will be certain limitations when applying them directly to outdoor 3D object detection. Despite this, these attempts to transfer the priors from 2D LMMs to 3D point clouds still provide a new perspective for solving the problem of sparsely-supervised 3D object detection. + +Motivated by these methods, we proposed a two-stage training strategy, SP3D, boosting sparsely-supervised 3D object detection with cross-modal semantic prompts generated by LMMs. As shown in Fig. 3, we first employ LMMs to extract semantics from 2D images and explicitly transfer them to 3D point clouds, generating pseudolabels for the first stage of detector training. In the second stage, we fine-tune the trained detector with sparse accurate labels. However, this enhanced paradigm still holds some challenges: (1) Semantic ambiguity at instance edges. Due to the absence of inherent depth information in images, directly transferring image semantics onto point clouds may result in noisy semantic prompts at the edge of the instance. (2) Efficient generation of pseudo-labels. Based on the obtained semantics, directly fitting pseudo-labels may result in incomplete foreground bounding boxes. Moreover, it is difficult to assess the quality of the generated pseudo-labels because of the lack of ground truth. + +To address the aforementioned challenges, our SP3D first designs a Confident Points Semantic Transfer (CPST) module, employing boundary-constrained center cluster selection operation to filter out inaccurate edges to obtain accurate cross-modal semantic prompts. These semantic prompts focus on central foreground semantic masks generated by LMMs. Inspired by unsupervised algorithms [35, 45], we can utilize these as accurate semantic prompts as seed points to generate bounding box pseudo-labels. In this case, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score) to efficiently discover and assess high-quality pseudo-labels with complete foreground information from seed points. As shown in Fig. 3, we utilize the generated pseudo-labels to train the 3D object detector for the first stage. After training, the 3D detector has learned a certain + +feature discrimination capability from the 2D images. Subsequently, we fine-tune the 3D detector with sparse accurate labels, and in conjunction with current label-efficient methods, it demonstrates relatively high detection capabilities even under extremely low labeling scenarios. + +Experiment results on the KITTI and Waymo datasets show that SP3D substantially enhances the performance of SoTA sparsely-supervised 3D object detectors. Moreover, without fine-tuning on labeled data, SP3D has also shown superior performance in a zero-shot setting, indicating the efficiency of the SP3D-initialized detector. The contributions of this paper can be summarized as follows: + +- We propose a two-stage boosting strategy, SP3D, utilizing accurate cross-modal prompts to boost the feature discrimination capability of 3D detectors under sparsely-supervised situations. +- We propose a Confident Points Semantic Transfer (CPST) module for accurate cross-modal semantic prompt generation and constraint, reducing the noise interference in the semantic transfer process. +- We propose a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score), which generate and assess high-quality pseudolabels based on semantic prompts to improve the detect performance. + +# 2. Related work + +# 2.1. Label-efficient 3D Object Detection + +Recently, label-efficient 3D object detection methods have begun to be explored in responding to the challenge of extremely low annotation volumes. Generally, these label-efficient methods can be categorized into semi-supervised [15, 24, 31, 48], and weakly-supervised [16, 21, 26, 43] approaches according to the difference in quantity and supervision form. In semi-supervised methods, full annotations are provided only for some scenes, while weakly-supervised methods typically adopt weaker annotation forms. For example, VS3D [26] leverages detectors on images to guide unsupervised 3D object detection, and WS3D [21] employs point annotation as a supervision signal. To maintain accuracy while reducing the annotations, SS3D [14] introduces the concept of sparse supervision, annotating only one complete 3D object per frame. CoIn [37] then adopts a contrastive instance feature mining strategy to extract the feature-level pseudo-labels from a significantly reduced amount of annotated data. However, the performance of existing methods remains constrained due to the insufficient feature discriminability of the initial detector, which may affect subsequent training under very few annotations. This work aims to develop a two-stage strategy, enabling the 3D detector to maintain robust feature representation capabilities despite having lower instance annotation ratio. + +# 2.2. Large Multimodal Models in 3D + +As the outstanding performance achieved by LMMs in 2D tasks [2, 12, 18, 25, 27], some studies have begun to explore their application in the 3D domain. Inspired by CLIP [27], ULIP [40] enhances the 3D understanding capability by transferring knowledge from 2D LMM to 3D encoder through contrastive learning methods. Similar works are [46, 50]. In the outdoor scenario, SAM3D [44] employs SAM to segment BEV images of point clouds and fit bounding boxes based on the segmentation masks to obtain detection results. CLIP2Scene [5] establishes the connection between point clouds and text by using images as an intermediate modality, enhancing the 3D model's semantic understanding of the scene with the prior knowledge of CLIP. Unlike previous approaches, our SP3D explicitly transfers semantic masks obtained from LMMs onto point clouds to generate high-quality pseudo-labels for the first-stage training of the 3D detector. + +# 2.3. Multimodal Representation Learning + +Recently, using multimodal methods from 2D images and 3D point clouds to enhance 3D detectors has gradually gained the community's attention [19, 28, 32, 34, 39]. However, these works mainly focus on investigating the image-point cloud fusion strategy, neglecting the utilization of images to explore label-efficient 3D detection. To reduce the required annotations, some methods transfer image information into point clouds to generate pseudo-labels [41, 51]. However, semantic ambiguity may occur at the object's edge due to the 2D-3D calibration error. MixSup [41] proposed a connected components labeling strategy, addressing this issue with the spatial separability property inherent to point clouds. SAL [23] employs density-based clustering to refine imperfect projection issues. Compared with these methods, our SP3D provides a simple but efficient way to reduce semantic noise caused by projection errors. + +# 3. Methods + +# 3.1. Overview + +This paper introduces SP3D, explicitly utilizing the accurate cross-modal semantic prompts generated by LMMs to boost the sparsely-supervised 3D detectors. As shown in Fig. 3, SP3D consists of three primary components: (1) Confident Points Semantic Transfer (CPST) module, acquiring accurate cross-modal semantic prompts, which we treat as seed points, with the boundary-constrained center cluster selection. (2) Dynamic Cluster Pseudo-label Generation (DCPG) module, which dynamically generates pseudo-label proposals based on the geometric shapes within the multi-scale neighborhood of these seed points. (3) Distribution Shape score (DS score), which employs unsupervised priors as the criterion for evaluating the qual + +ity of pseudo-label proposals, and we subsequently apply NMS [22] to retain high-quality pseudo-labels further. We then follow the training strategy of CoIn [37] to train an initial 3D detector with enhanced discriminative capacity. We detail our SP3D framework as follows. + +# 3.2. Confident Points Semantic Transfer Module + +Encouraged by the development of LMMs, we first utilize the prior knowledge of LMMs to extract semantic information from 2D images explicitly. Meanwhile, by integrating the calibration relationship matrix between 2D images and 3D point clouds, it is straightforward to transfer semantic information onto point clouds. However, as shown in Fig. 2, there is significant noise in the edge points of instances during the transfer process. These semantic points are mistakenly assigned to other instances can have a detrimental impact on subsequent pseudo-label fitting and the training of the detector. To prevent the incorporation of noise during the transfer of semantic information, we use the boundary-constrained mask shrink operation, followed by the coordinate system transformation, to obtain accurate cross-modal semantic prompts. CSPT is illustrated in Fig. 3(a); we have divided it into two parts as LMMs-guided semantic extraction and confident points filtering. + +LMMs-guided semantic extraction. The goal of LMMs is to generate high-quality foreground semantic masks. Specifically, we take as input an image $\mathcal{I} \in \mathbb{R}^{3 \times H \times W}$ and the category-text prompt $\mathcal{T}^{\mathcal{C}} = \{t_1^{\mathcal{C}}, t_2^{\mathcal{C}}, \dots, t_C^{\mathcal{C}}\}$ , where $H$ and $W$ denote the height and width of the image, $C$ denotes the categories of interest. We first utilize FastSAM [49] to perform speed-efficient segmentation as + +$$ +\mathcal {M} _ {\mathcal {I}} = \operatorname {S A M} (\mathcal {I}), \tag {1} +$$ + +where $\mathcal{M}_{\mathcal{I}} \in \mathbb{R}^{M \times H \times W}$ denotes the $M$ class-agnostic masks extracted from $\mathcal{I}$ , $\mathrm{SAM}(\cdot)$ indicates the FastSAM model. We then utilize $\mathcal{M}_{\mathcal{I}}$ as the mask prompts and feed them, along with image $\mathcal{I}$ , into the SemanticSAM [4] model, which is expected to output the descriptions $\mathcal{T}^{\mathcal{D}} = \{t_1^{\mathcal{D}}, t_2^{\mathcal{D}}, \dots, t_M^{\mathcal{D}}\}$ for each mask in $\mathcal{M}_{\mathcal{I}}$ . Specifically: + +$$ +\mathcal {T} ^ {\mathcal {D}} = \operatorname {S S A M} (\mathcal {I}, \mathcal {M} _ {\mathcal {I}}), \tag {2} +$$ + +where $\mathrm{SSAM}(\cdot)$ refers to the SemanticSAM model. Then, we calculate the cosine similarity between the category-text $\mathcal{T}^{\mathcal{C}}$ and the mask description $\mathcal{T}^{\mathcal{D}}$ to filter out uninteresting background masks, obtaining the foreground masks $\mathcal{M}_{\mathcal{I}}'$ . + +Confident points filtering. As fuzziness of boundaries results from depth occlusions in images and calibration inaccuracies, we opt to constrain the boundary of foreground masks before 2D-3D transfer, retaining only its central portion. Specifically, for each foreground mask, we denote its + +![](images/f318f96b058ae74adb971c7700206bdf6311b61924d22f1751029b456ab92bf3.jpg) +Figure 3. The overview of our SP3D, including (a) CSPT finds semantic seed points through high-confidence semantic masks transfer, (b) DCPG dynamically clusters neighbor points of seed points to fit pseudo-label proposals, and (c) DS score to evaluate the quality of generated pseudo-label proposals, serving as a scoring metric to NMS to suppress low-quality proposals. + +maximum and minimum values in the pixel coordinate system $(u,v)$ as $(u_{min},u_{max},v_{min},v_{max})$ . We perform mask shrink to constraint boundary range of $\mathcal{M}_{\mathcal{I}}^{\prime}$ to obtain $\hat{\mathcal{M}}_{\mathcal{I}}^{\prime}$ : + +$$ +u \in \left[ u _ {\min } + \frac {1}{2} (1 \pm \gamma) \left(u _ {\max } - u _ {\min }\right) \right], \tag {3} +$$ + +$$ +v \in [ v _ {m i n} + \frac {1}{2} (1 \pm \gamma) (v _ {m a x} - v _ {m i n}) ], +$$ + +where $\gamma$ denotes the shrink factor. With this constraint, we obtain the shrunk masks that filter out the semantically ambiguous regions, ensuring the accuracy of foreground semantic information transferred onto the point clouds. Following [30], we transfer the semantic mask from the image onto the point clouds to obtain accurate cross-modal semantic prompts, namely seed points, using the camera's intrinsic and extrinsic parameter matrices. Note that, different from [30], we adopt the approach of explicitly transferring the shrunk masks onto the point cloud rather than implicitly embedding unprocessed semantic masks into the point cloud's features. This approach helps avoid potential semantic feature confusion between different modalities arising from sparse annotations. + +# 3.3. Dynamic Cluster Pseudo-label Generation + +With the proposed CPST, we explicitly obtain the semantic seed points from transformation $\hat{\mathcal{M}}_{\mathcal{I}}^{\prime}$ of the foreground mask. Given $\mathcal{P} \in \mathbb{R}^{N \times 3} = \{p_1, p_2, \dots, p_N\}$ as LiDAR + +points, we define the seed points covered by $\hat{\mathcal{M}}_{\mathcal{I}}^{\prime}$ as $\mathcal{P}_T = \{p_t\}$ , $\mathcal{P}_T \subset \mathcal{P}$ . It is crucial for the 3D detection task that obtain complete bounding boxes from these seed points. + +By referring to traditional unsupervised pseudo-label generation methods [45], we produce many pseudo-labels and then use the positional constraints of the seed points to retain the more promising pseudo-labels as supervisory signals. However, existing unsupervised bounding box fitting approaches [35, 45] usually take a fixed constant as cluster radius, leading to the problem of inadequate foreground or excessive background noise for the generated bounding boxes. In this case, we propose a dynamic cluster pseudo-label generation (DCPG) module. This module utilizes the geometry shape of the seed points' multi-scale neighborhood to capture complete foreground information while minimizing background interference. It dynamically generates pseudo-label proposals, as shown in Fig. 3(b). + +Specifically, we denote $\mathcal{P}_T^{(k)}$ as the $k$ -th instance in a point cloud frame and utilize DCPG dynamically generates a clustering radius $r$ for the $t$ -th seed point $p_t^{(k)}$ . We define the updating function for the dynamic radius as + +$$ +\operatorname {u p d a t e} (t, r _ {\text {i n i t}}) = r _ {\text {i n i t}} \cdot \frac {t}{N ^ {(k)}} + \delta , t = 1, 2, \dots , N ^ {(k)}, \tag {4} +$$ + +where $r_{init}$ is a hyper-parameter set based on empirical experience, $\delta$ denotes adjustment factor to avoid $r$ too small, and $N^{(k)}$ is the number of seed points in current instance. + +Algorithm 1 Dynamic cluster pseudo-label generation +Require: LiDAR points $\mathcal{P}$ , the seed points set $\mathcal{P}_T^{(k)}$ of the $k$ -th instance, initial radius $r_{init}$ , ground remove function $\mathcal{G}$ , box fitting function $\mathcal{F}$ ; +Ensure: Pseudo-label proposal set $\hat{\mathcal{B}}^{(k)}$ ; +1: Acquire the number $N^{(k)}$ of points for the seed points set of current instance $\mathcal{P}_T^{(k)}$ ; +2: $\mathcal{P}_{gr} \gets \mathcal{G}(\mathcal{P})$ ; // remove ground points +3: $\hat{\mathcal{B}}^{(k)} = \emptyset$ ; +4: for $t = 1, 2, \dots, N^{(k)}$ do +5: $p_t = \mathcal{P}_T^{(k)}[t]$ ; // loop each seed point +6: $r \gets \text{update}(t, r_{init})$ +7: $\hat{\mathbf{b}} \gets \mathcal{F}(\text{DBSCAN}(\mathcal{P}_{gr}, p_t, r))$ +8: $\hat{\mathcal{B}}^{(k)} \gets \hat{\mathcal{B}}^{(k)} \cup \hat{\mathbf{b}}$ +9: end for + +By applying Eq. 4, we dynamically update the radius $r$ , $r \in (\delta, r_{init} + \delta]$ , during point clustering, thereby obtaining foreground clusters with multi-scale receptive fields. Following [45]. We utilize the radius calculated from Eq. 4 as the clustering radius for DBSCAN [8] and employ [47] to fit the bounding box for each foreground cluster. Algorithm 1 summarizes Our DCPG. + +# 3.4. Distribution Shape Score + +While DCPG has the capacity for high-quality pseudolabel generation, it unavoidably produces an amount of low-quality pseudo-label proposals. The shapes of these proposals and the extent of foreground completeness contained within the proposals usually deviate significantly from reality. Traditional detection methods typically compute the Intersection over Union (IoU) score between predicted bounding boxes and ground-truth (GT) boxes and then employ the NMS [22] to suppress these low-quality proposals. However, lacking GT makes it challenging to directly apply NMS using IoU as the evaluation criterion within our SP3D framework. In this case, we propose a distribution shape score (DS score) to assess the quality of the pseudo-labels using unsupervised prior knowledge. As illustrated in Fig. 3(c), we divided the DS score into two parts: distribution constraint score and meta-shape constraint score. + +Distribution constraint score. Inspired by [20], within a high-quality pseudo-label proposal $\hat{b}$ , the distances from its interior points $p_{i,i=1,\dots,n}$ to its boundary roughly follow a Gaussian distribution $\mathcal{N}(\mu, \sigma)$ , where $\mu = 0.8$ and $\sigma = 0.2$ , respectively. In other words, we denote random variable $D = \{d_1, \dots, d_n\}$ as the distance between $p_i$ and the box boundary of $\hat{b}$ , and $D \sim \mathcal{N}(0.8, 0.2)$ . Based on this prior, we assign a distribution constraint score to the pseudo-label proposal $\hat{b}$ by calculating the similarity between the random + +variable $D$ corresponding to each $\hat{b}$ and the normal distribution $\mathcal{N}$ . Specifically: + +$$ +s _ {d c} (\hat {b}) = \frac {1}{| \mathcal {P} _ {f g} |} \sum_ {p _ {i} \in \mathcal {P} _ {f g}} \log (\mathcal {N} (D | \mu , \sigma)), \tag {5} +$$ + +where $\log (\cdot)$ denotes the logarithm function, $\mathcal{P}_{fg}$ is the foreground points within $\hat{b}$ , and $|\mathcal{P}_{fg}|$ represents the number of points in $\mathcal{P}_{fg}$ . + +Meta-shape constraint score. In addition, the shape of a high-quality pseudo-label is expected to be consistent with its template in the real world, which we define as the meta instance, corresponding to its category [35]. For class $c$ , we denote $\mathcal{B}_c \in \{l_c, w_c, h_c\}$ as the shape of its meta instance, where $l_c$ , $w_c$ and $h_c$ are the normalized length, width and height, respectively. We followed this shape prior to constructing the class-aware meta-shape constraint score $s_{msc}(\hat{b})$ as + +$$ +s _ {m s c} (\hat {b}) = 1 - \Phi_ {K L} (\mathcal {B} _ {c} | | \hat {\mathcal {B}} _ {\hat {b}}), \tag {6} +$$ + +where $\Phi_{KL}(\cdot)$ denotes the normalized KL divergence function, and $\hat{\mathcal{B}}_{\hat{b}} \in \{l_{\hat{b}}, w_{\hat{b}}, h_{\hat{b}}\}$ indicates the normalized shape of the pseudo-label proposal. The purpose of this operation is to suppress the low-quality proposals whose shape deviates significantly from the meta instance. By combining the distribution constraint score and the meta-shape constraint score, we can obtain the DS score as + +$$ +\mathrm {D S} (\hat {b}) = \lambda_ {1} \bar {s} _ {d c} (\hat {b}) + \lambda_ {2} \bar {s} _ {m s c} (\hat {b}), \tag {7} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ denote weight adjustment factor. Notably, to unify the dimension, we normalized the two constraint scores before combining them, resulting in $\overline{s}_{dc}$ and $\overline{s}_{msc}$ . We then employ the DS score as a substitution for the confidence score in NMS to suppress the low-quality pseudolabels. We utilize the obtained pseudo-labels in conjunction with CoIn [37] for the first phase of training. Subsequently, we fine-tune the trained detector with a small amount of accurate labels to boost the performance of the 3D detector. + +# 4. Experiments + +# 4.1. Dataset and Metrics + +KITTI dataset. The KITTI [9] dataset is a widely used benchmark for 3D object detection. During the first training stage, we relied solely on the semantic information provided by LMMs to generate pseudo-labels instead of the human-annotated ground truth. In the fine-tuning stage, we split the training set (7,481 scenes) into a train split (3,712 scenes) and a val split (3,769 scenes). We follow [37] to randomly select $10\%$ of the scenes from the train split and retain only one instance annotation per scene. In this case, + +
SettingCostMethodCarPedCyc
EasyMod.HardEasyMod.HardEasyMod.Hard
Fully-sup.100%VoxelRCNN92.384.982.669.663.058.688.772.568.2
Sparsely-sup.20%SS3D89.384.278.2------
2%VoxelRCNN70.554.944.842.638.532.173.347.843.2
2%CoIn89.170.255.650.845.239.680.252.348.6
2%CoIn++92.079.571.546.736.131.282.058.454.6
2%CoIn++ with SP3D91.380.574.067.458.750.992.573.168.3
+ +Table 1. Comparison with SoTA sparsely-supervised methods on KITTI val split. All methods are based on VoxelRCNN, and we report the 3D AP results of full cost (100%) and limited cost (20%, 2%). The best sparsely-supervised methods are highlighted in bold. + +
StageLabelMethodsCar-3D @IoU 0.7Car-BEV @IoU 0.7
EasyModHardEasyModHard
Single-stageSparse1. CenterPoint49.6931.5525.9156.7842.5034.14
Sparse2. CoIn (CenterPoint-based)72.0354.8243.7787.2073.5466.03
Sparse3. 2 with SP3D (CenterPoint-based)87.4469.2458.6192.7280.0069.01
-4. Improvements 3→2+15.41+14.42+14.84+5.52+6.46+2.98
Two-stageSparse1. Voxel-RCNN70.5254.9744.8283.6771.1457.71
Sparse2. CoIn (Voxel-RCNN-based)84.5668.4758.0292.3181.0170.24
Sparse3. 2 with SP3D (Voxel-RCNN-based)91.3774.8963.8495.4185.2774.57
-4. Improvements 3→2+6.81+6.42+5.82+3.1+4.26+4.33
Multi-stageSparse1. CasA74.1857.3745.0585.9073.2157.23
Sparse2. CoIn (CasA-based)89.1775.3262.9895.9985.0272.47
Sparse3. 2 with SP3D (CasA-based)91.1275.9466.4696.5585.6576.31
-4. Improvements 3→2+1.95+0.62+3.48+0.56+0.63+3.84
+ +Table 2. Comparison with different fully-supervised methods. Sparse label refers to the use of limited split (2% annotation cost). The 3D object detection and BEV detection benchmark are evaluated by mAP with R40, under IoU threshold 0.7. + +
MethodsVeh. LEVEL 1Veh. LEVEL 2
mAPmAPHmAPmAPH
CoIn39.639.034.233.6
CoIn+SP3D46.746.040.439.9
Improvement+7.1+7.0+6.2+6.3
+ +Table 3. Comparison with SoTA sparsely-supervised methods on WOD validation set. All methods are based on CenterPoint, and we report the results at a $1\%$ annotation cost and 0.7 IoU threshold. + +we can obtain a limited split, which merely takes $2\%$ of instance annotations compared with the origin train split. All results are evaluated using the official evaluation metric: 3D Average Precision (AP) across 40 recall thresholds (R40). + +Waymo Open Dataset (WOD). The WOD is a widely used large-scale benchmark that contains 798, 202, and 150 sequences (approximately 200 frames per sequence) for training, validation, and testing, respectively. All results are evaluated by using 3D mean Average Precision (mAP) and its variant weighted by heading accuracy (mAPH). + +# 4.2. Implementation Details + +Pseudo-labels generation. We directly employed the model parameters provided by FastSAM [49] and SemanticSAM [4] for inference without additional supervisory signals for fine-tuning. To achieve accurate segmentation results, we set a higher segmentation threshold of 0.7 during the FastSAM inference process. To mitigate computational demands, we opted to generate pseudo-labels within a confined spatial domain of the semantically relevant points, specifically within an 8-meter radius. We set mask shrink factor $\gamma$ to 0.3, initial cluster radius $r_{init}$ to 1, adjustment factor $\delta$ to 0.1. We utilize unsupervised priors to filter out the low-quality pseudo-labels that are evidently inconsistent with the intuitive expectations and set the weight adjustment factor of DS score $\lambda_1$ and $\lambda_2$ as 0.5, 0.5. + +Detector training. We conduct all experiments with a batch size of 8 and a learning rate of 0.003 on 4 RTX 3090 GPUs. Following previous sparsely-supervised 3D object detection methods [14, 37], we choose three different clas + +
Methods2%1%0.5%
CarPed.Cyc.Avg.CarPed.Cyc.Avg.CarPed.Cyc.Avg.
semi-sup.HSSDA81.958.265.868.680.951.942.359.577.819.544.147.2
HSSDA+SP3D83.260.874.472.883.053.269.168.482.322.146.450.3
Improvement+1.3+2.6+8.6+4.2+2.1+1.3+26.8+8.9+4.5+2.6+2.3+3.1
sparsely-sup.CoIn72.336.958.455.865.331.552.249.655.225.530.537.0
CoIn+SP3D73.444.762.360.171.736.959.155.963.336.143.447.6
Improvement+1.1+7.8+3.9+4.3+6.4+5.4+6.9+6.3+8.1+10.6+12.5+10.6
HINTED85.160.081.375.482.849.974.168.980.245.864.563.5
HINTED+SP3D85.560.181.475.684.057.281.074.083.249.273.968.7
Improvement+0.4+0.1+0.1+0.2+1.2+7.3+6.9+5.1+3.0+3.4+9.4+5.2
+ +Table 4. Improvement of SP3D in SoTA semi-supervised and sparsely-supervised methods on KITTI val split under limited cost setting. The results are evaluated by mAP with R40, under IoU thresholds 0.7, 0.5, and 0.5 for 'Car', 'Pedestrian', and 'Cyclist', respectively. Additionally, the AP is computed across the difficulty levels of easy, moderate, and hard. + +
MethodsModalityCar-3D @IOU 0.5/0.7
Train TestEasyMod.Hard
VS3DL+CL40.32/9.0937.36/5.7331.09/5.03
WS3DPRL+CL-/60.01-/44.48-/36.93
SP3D (Ours)L+CL93.75/69.7176.36/48.6571.01/40.53
+ +sical detectors [6, 33, 42] as our architecture. And we employ the OpenPCDet [29] to conduct our experiments. In the first training stage, we employ CoIn [37] to train an initial detector with the generated pseudo-labels. Then, we use limited split to fine-tune the detector. + +Baselines. To thoroughly validate the effectiveness of SP3D, we select the SoTA sparsely-supervised method [37] as the primary baseline for comparison. We compare the proposed SP3D approach with the baseline under conventional sparse settings with $2\%$ annotation cost. We also compared with cross-modal weakly-supervised methods [16, 26] in a zero-shot setting, which also incorporate visual models to extract semantic information to enhance the performance of detectors. Furthermore, we establish baselines under progressively reduced annotation costs to evaluate the sensitivity to annotation costs. + +# 4.3. Comparison with SoTA Methods + +Comparison with sparsely-supervised methods. We integrate our proposed SP3D into the SoTA sparsely-supervised 3D detection algorithm, CoIn++ [37]. For a fair comparison, all detectors employ the VoxelRCNN [6] as the base architecture. As illustrated in Tab. 1, SP3D significantly improves the detection performance of CoIn++. The experimental results on WOD, as reported in Tab. 3, + +Table 5. Comparison of zero-shot methods on KITTI val split. + +
MethodsModalityVeh.Ped.Cyc.
TrainTest
SAM3DLL19.10.00.0
CM3DL + CL23.70.12.6
SP3D (Ours)L + CL35.015.75.9
+ +Table 6. Comparison with zero-shot methods on the WOD validation set. We report Level-2 mAP with IoU thresholds of 0.7, 0.5, and 0.5 for Vehicle, Pedestrian, and Cyclist, respectively. + +also validate the performance enhancement of our method for the sparsely-supervised 3D object detector. However, we observe a slight decrease in precision for the 'Easy' car category with our SP3D-initialized model in Tab. 1. This could arise because our initial pseudo-labels are inferred based on the geometric shape of the objects, which may differ from the conventions of manual annotation. When the point cloud structure of an instance is relatively intact, such discrepancies can lead to noticeable differences in the size of the annotated bounding boxes. + +Comparison with fully-supervised methods. For a fair comparison, we follow CoIn [37] to select CenterPoint [42], VoxelRCNN [6], and CasA [33] as our baseline detectors, representing three distinct types of detection algorithms. We initialize the 3D detector using our SP3D, followed by fine-tuning with the limited split. As shown in Tab. 2, detectors designed for full supervision struggle to yield optimal results due to scarce annotations. Despite the effectiveness of CoIn in improving this situation, the results achieved are still unsatisfactory for single-stage detection algorithms with relatively simple structures. Our designed strategy, SP3D, significantly reduces this discrepancy, enabling detectors to achieve similar results. + +
MethodsCar-3D @IoU 0.5
EasyMod.Hard
CPD61.4861.5860.09
CPD+SP3D73.6566.1363.78
Improvement+12.17+4.55+2.83
+ +Comparison with zero-shot methods. We also compared SP3D with SoTA zero-shot methods. SAM3D [44] and CM3D [11] generate 3D pseudo-labels by leveraging large models for the semantic segmentation of 2D images. In line with the zero-shot approach of prior methods, no 3D annotations were used for training; instead, pseudolabels were generated from large models and image data. As shown in Tab. 5 and Tab. 6, the results on the KITTI and Waymo datasets demonstrate that our method significantly outperforms previous approaches. + +Comparison with different annotation rates. To explore the robustness of SP3D, we conducted experiments on detectors across different supervision paradigms under varying annotation rates, and we reported the results on Tab. 4. For a fair comparison, we initialize the detectors by following the strategies of CoIn (CenterPoint) [37], HSSDA (PV-RCNN) [15], and HINTED (Voxel-RCNN) [38]. The experimental results show that when the annotation rate is limited, detector performance drops sharply as the amount of annotated data decreases. By inheriting SP3D, these detectors' performance has been significantly enhanced, mitigating the performance drop associated with reduced annotation. Taking CoIn as an example, its average AP for the three categories at annotation rates of $1\%$ and $0.5\%$ has improved by $6.3\%$ and $10.6\%$ , respectively. These results indicate that even at low annotation rates, SP3D-initialized models can significantly improve the performance of 3D detectors across different supervision paradigms. + +Comparison with unsupervised methods. To further validate the robustness of SP3D, we integrated it with the SoTA unsupervised detection method CPD [35]. As shown in Tab. 7, despite suffering some impact, the combination of SP3D and CPD still enhances the performance of the unsupervised 3D object detector to a certain extent. + +# 4.4. Ablation Study + +We conducted ablation studies based on CenterPoint to verify the proposed modules' effectiveness and recorded the results in Tab. 8. The first and third rows show that the precision of pseudo-labels, as augmented by DCPG, can substantially amplify the 3D detector's capabilities. This may be + +Table 7. Improvement for unsupervised method. + +
Mask shrinkDCPGDS scoreCar-3D AP@IoU 0.7
EasyMod.Hard
✓35.1023.7519.52
✓✓38.2629.2524.74
✓✓40.5628.1522.40
✓✓47.2333.4027.13
✓✓✓52.5638.0031.52
+ +Table 8. Ablation study on KITTI val split. + +
DS scoreCar-3D AP@IoU 0.7
MSCDCEasyMod.Hard
✓44.9133.5528.19
✓48.4835.3929.75
+ +Table 9. Ablation study of each component of DS score. + +due to incorporating more comprehensive foreground information in pseudo-labels, which enhanced the model's feature discrimination capability. The last two rows show that the mask shrink operation is necessary for handling semantic noise at the instance edges. The first three rows and fifth row show that using the DS score to filter out low-quality labels can significantly boost the detector's precision. Moreover, we validated the effectiveness of each component in DS score with the ablation studies presented in Tab. 9. + +# 5. Discussion and Conclusion + +This paper proposes a two-stage boosting strategy, SP3D, that leverages accurate cross-modality semantic prompts to enhance the sparsely-supervised 3D detectors' capabilities. We first develop a CSPT module to obtain accurate cross-modal semantic prompts in point clouds by efficiently transferring high-confidence semantic masks. We then design a DCPG module to dynamically generate pseudo-label proposals based on semantic prompts. Lastly, we propose a DS score as the criterion for NMS to select high-quality pseudo-labels integrated with sparsely-supervised strategy to train the initial detector. After fine-tuning with sparse GT data, SP3D show superior performance over the original sparsely-supervised 3D detector on the WOD and KITTI dataset, and remain robust performance even as the amount of annotation decreased. Limitations: SP3D exhibits a relatively significant degradation when fine-tuning with annotation rates below $0.1\%$ , which may result from the noise introduced by extremely low annotations. Future efforts to explore efficient fine-tuning strategies to solve this issue. + +Acknowledgment. This work was supported by the National Natural Science Foundation of China (No.62171393), and the Fundamental Research Funds for the Central Universities (No.20720220064). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 3 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[4] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.3,6 +[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 3 +[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 1, 7 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1 +[8] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 5 +[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5 +[10] Xun Huang, Hai Wu, Xin Li, Xiaoliang Fan, Chenglu Wen, and Cheng Wang. Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2409–2416, 2024. 1 +[11] Mehar Khurana, Neehar Peri, James Hays, and Deva Ramanan. Shelf-supervised cross-modal pre-training for 3d object detection. In 8th Annual Conference on Robot Learning, 2024. 8 +[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 3 + +[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1 +[14] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8428-8437, 2022. 1, 2, 6 +[15] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23819-23828, 2023. 1, 2, 8 +[16] Haizhuang Liu, Huimin Ma, Yilin Wang, Bochao Zou, Tianyu Hu, Rongquan Wang, and Jiansheng Chen. Eliminating spatial ambiguity for weakly supervised 3d object detection without spatial labels. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3511-3520, 2022. 2, 7 +[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1 +[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 3 +[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3 +[20] Katie Luo, Zhenzhen Liu, Xiangyu Chen, Yurong You, Sagie Benaim, Cheng Perng Phoo, Mark Campbell, Wen Sun, Bharath Hariharan, and Kilian Q Weinberger. Reward finetuning for faster and more accurate unsupervised object discovery. Advances in Neural Information Processing Systems, 36, 2024. 5 +[21] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Yunde Jia, and Luc Van Gool. Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4454-4468, 2021. 1, 2 +[22] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 3, 5 +[23] Aljosa Osep, Tim Meinhardt, Francesco Ferroni, Neehar Peri, Deva Ramanan, and Laura Leal-Taixe. Better call sal: Towards learning to segment anything in lidar. In European Conference on Computer Vision, pages 71-90. Springer, 2025. 3 + +[24] Jinhyung Park, Chenfeng Xu, Yiyang Zhou, Masayoshi Tomizuka, and Wei Zhan. Detmatch: Two teachers are better than one for joint 2d and 3d semi-supervised object detection. In European Conference on Computer Vision, pages 370-389. Springer, 2022. 2 +[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3 +[26] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 7 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3 +[28] Ziying Song, Feiyang Jia, Hongyu Pan, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lin Liu, Yang Ji, Lei Yang, and Li Wang. Contrastalign: Toward robust bev feature alignment via contrastive learning for multi-modal 3d object detection. arXiv preprint arXiv:2405.16873, 2024. 3 +[29] OD Team et al. Openpcdet: An open-source toolbox for 3d object detection from point clouds, 2020. 7 +[30] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 4 +[31] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 1, 2 +[32] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6792-6802, 2023. 3 +[33] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7 +[34] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 1, 3 +[35] Hai Wu, Shijia Zhao, Xun Huang, Chenglu Wen, Xin Li, and Cheng Wang. Commonsense prototype for outdoor unsupervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14968-14977, 2024. 2, 4, 5, 8 +[36] Qiming Xia, Yidong Chen, Guorong Cai, Guikun Chen, Daoshun Xie, Jinhe Su, and Zongyue Wang. 3-d hanet: A + +flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1 +[37] Qiming Xia, Jinhao Deng, Chenglu Wen, Hai Wu, Shaoshuai Shi, Xin Li, and Cheng Wang. Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6254-6263, 2023. 1, 2, 3, 5, 6, 7, 8 +[38] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 8 +[39] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3 +[40] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 1, 3 +[41] Yuxue Yang, Lue Fan, and Zhaoxiang Zhang. Mixsup: Mixed-grained supervision for label-efficient lidar-based 3d object detection. arXiv preprint arXiv:2401.16305, 2024. 3 +[42] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 7 +[43] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8373-8383, 2023. 2 +[44] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: zero-shot 3d object detection via the segment anything model. Science China Information Sciences, 67(4):149101, 2024. 3, 8 +[45] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9317-9328, 2023. 2, 4, 5 +[46] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1, 3 + +[47] Xiao Zhang, Wenda Xu, Chiyu Dong, and John M Dolan. Efficient l-shape fitting for vehicle detection using laser scanners. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 54-59. IEEE, 2017. 5 +[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembleing semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2 +[49] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything, 2023. 3, 6 +[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 1, 3 +[51] Pufan Zou, Shijia Zhao, Weijie Huang, Qiming Xia, Chenglu Wen, Wei Li, and Cheng Wang. Adaco: Overcoming visual foundation model noise in 3d semantic segmentation via adaptive label correction. arXiv preprint arXiv:2412.18255, 2024. 3 \ No newline at end of file diff --git a/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/images.zip b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f5d0f015c1852bf7026dbfc5394e814bb06552a8 --- /dev/null +++ b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ec5d03b7f43266e844c41f685d9986977c23917a8192818a1f2d611dc10493f +size 578172 diff --git a/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/layout.json b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..336d461baa5a68c25c69a69cb5a9cb060190c0db --- /dev/null +++ b/2025/SP3D_ Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts/layout.json @@ -0,0 +1,8411 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 86, + 102, + 523, + 140 + ], + "type": "text", + "content": "SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": "Shijia Zhao" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Qiming Xia" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Xusheng Guo" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Pufan Zou" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Maoji Zheng" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Hai Wu" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Chenglu Wen" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "text", + "content": " Cheng Wang" + }, + { + "bbox": [ + 107, + 159, + 518, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 190, + 564, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 190, + 564, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 564, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 190, + 564, + 232 + ], + "type": "text", + "content": "Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University, Xiamen, China \n" + }, + { + "bbox": [ + 62, + 190, + 564, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 190, + 564, + 232 + ], + "type": "text", + "content": " Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 283, + 297, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 297, + 619 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 297, + 619 + ], + "type": "text", + "content": "Recently, sparsely-supervised 3D object detection has gained great attention, achieving performance close to fully-supervised 3D detectors while requiring only a few annotated instances. Nevertheless, these methods suffer challenges when accurate labels are extremely absent. In this paper, we propose a boosting strategy, termed SP3D, explicitly utilizing the cross-modal semantic prompts generated from Large Multimodal Models (LMMs) to boost the 3D detector with robust feature discrimination capability under sparse annotation settings. Specifically, we first develop a Confident Points Semantic Transfer (CPST) module that generates accurate cross-modal semantic prompts through boundary-constrained center cluster selection. Based on these accurate semantic prompts, which we treat as seed points, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module to yield pseudo-supervision signals from the geometry shape of multi-scale neighbor points. Additionally, we design a Distribution Shape score (DS score) that chooses high-quality supervision signals for the initial training of the 3D detector. Experiments on the KITTI dataset and Waymo Open Dataset (WOD) have validated that SP3D can enhance the performance of sparsely supervised detectors by a large margin under meager labeling conditions. Moreover, we verified SP3D in the zero-shot setting, where its performance exceeded that of the state-of-the-art methods. The code is available at https://github.com/xmuqimingxia/SP3D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 135, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 135, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 135, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 661, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 686 + ], + "type": "text", + "content": "3D object detection, aiming at locating and classifying objects within 3D scenes, has garnered significant attention in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 260, + 554, + 336 + ], + "blocks": [ + { + "bbox": [ + 315, + 260, + 554, + 336 + ], + "lines": [ + { + "bbox": [ + 315, + 260, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 315, + 260, + 554, + 336 + ], + "type": "image", + "image_path": "305b3135f0d33c79ca91e5dd308aa752cbda1c0b1956d456f2d7f149cac9ea0a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 345, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 378 + ], + "type": "text", + "content": "Figure 1. Performance comparison of the sparsely-supervised detector at various annotation rates. SP3D indicates the CoIn initialized with the proposed SP3D." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 400, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 400, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 312, + 400, + 555, + 449 + ], + "type": "text", + "content": "autonomous driving [6, 10, 19, 34, 36]. However, the mainstream 3D detectors relies heavily on high-quality labels annotated by humans, which is not only time-consuming but also sensitive to the subjective impression of annotators." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 449, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 556, + 616 + ], + "type": "text", + "content": "To minimize the dependence of 3D detectors on high-quality manual annotations, recent work has begun to focus on label-efficient training strategies [15, 21, 31, 37]. To discover unlabeled instances, SS3D [14] employs a self-training approach to iteratively optimize the detector trained on sparsely annotated data. CoIn [37] introduces contrastive learning methods, enhancing the model's discriminative capability for various category features. However, existing strategies make 3D detectors struggle to extract sufficiently discriminative features from extremely limited annotations. Fig. 1 shows some examples where the state-of-the-art (SoTA) sparsely-supervised object detector CoIn [37] hardly maintains robust performance with a significant reduction in annotation rate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 557, + 714 + ], + "type": "text", + "content": "With the successive emergence and widespread application of large language models (LLMs), such as BERT [7] and GPT [1, 3], in natural language processing, research on large multimodal models (LMMs) is also gaining momentum [12, 13, 17, 27]. Benefiting from the outstanding performance of LMMs, the utilization of pre-trained LMMs has led to significant advancements in 2D vision tasks. Inspired by this, [40, 46, 50] transfers the image-text knowl" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 234, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 234, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 234, + 713 + ], + "type": "text", + "content": "†Corresponding author, clwen@xmu.edu.cn." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29374" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 172, + 162 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 172, + 162 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 172, + 162 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 172, + 162 + ], + "type": "image", + "image_path": "df33c4bea8e9b9886939ae7a97635bd84acf2a09e4bc978ef3130af63bd63bb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "type": "text", + "content": "Figure 2. Semantics belonging to a foreground object may be incorrectly assigned to background/other objects." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 282, + 163 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 282, + 163 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 282, + 163 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 282, + 163 + ], + "type": "image", + "image_path": "3d02f993f13d66a59661380e7aa38a8becec8642b55e445381b8847698022c60.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 219, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 219, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 295, + 302 + ], + "type": "text", + "content": "edge prior from 2D LMMs to 3D point clouds. However, these methods typically focus on classifying individual instances, and there will be certain limitations when applying them directly to outdoor 3D object detection. Despite this, these attempts to transfer the priors from 2D LMMs to 3D point clouds still provide a new perspective for solving the problem of sparsely-supervised 3D object detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 519 + ], + "type": "text", + "content": "Motivated by these methods, we proposed a two-stage training strategy, SP3D, boosting sparsely-supervised 3D object detection with cross-modal semantic prompts generated by LMMs. As shown in Fig. 3, we first employ LMMs to extract semantics from 2D images and explicitly transfer them to 3D point clouds, generating pseudolabels for the first stage of detector training. In the second stage, we fine-tune the trained detector with sparse accurate labels. However, this enhanced paradigm still holds some challenges: (1) Semantic ambiguity at instance edges. Due to the absence of inherent depth information in images, directly transferring image semantics onto point clouds may result in noisy semantic prompts at the edge of the instance. (2) Efficient generation of pseudo-labels. Based on the obtained semantics, directly fitting pseudo-labels may result in incomplete foreground bounding boxes. Moreover, it is difficult to assess the quality of the generated pseudo-labels because of the lack of ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "To address the aforementioned challenges, our SP3D first designs a Confident Points Semantic Transfer (CPST) module, employing boundary-constrained center cluster selection operation to filter out inaccurate edges to obtain accurate cross-modal semantic prompts. These semantic prompts focus on central foreground semantic masks generated by LMMs. Inspired by unsupervised algorithms [35, 45], we can utilize these as accurate semantic prompts as seed points to generate bounding box pseudo-labels. In this case, we introduce a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score) to efficiently discover and assess high-quality pseudo-labels with complete foreground information from seed points. As shown in Fig. 3, we utilize the generated pseudo-labels to train the 3D object detector for the first stage. After training, the 3D detector has learned a certain" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "feature discrimination capability from the 2D images. Subsequently, we fine-tune the 3D detector with sparse accurate labels, and in conjunction with current label-efficient methods, it demonstrates relatively high detection capabilities even under extremely low labeling scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 216 + ], + "type": "text", + "content": "Experiment results on the KITTI and Waymo datasets show that SP3D substantially enhances the performance of SoTA sparsely-supervised 3D object detectors. Moreover, without fine-tuning on labeled data, SP3D has also shown superior performance in a zero-shot setting, indicating the efficiency of the SP3D-initialized detector. The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 218, + 553, + 373 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 218, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 265 + ], + "type": "text", + "content": "- We propose a two-stage boosting strategy, SP3D, utilizing accurate cross-modal prompts to boost the feature discrimination capability of 3D detectors under sparsely-supervised situations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 266, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 313 + ], + "type": "text", + "content": "- We propose a Confident Points Semantic Transfer (CPST) module for accurate cross-modal semantic prompt generation and constraint, reducing the noise interference in the semantic transfer process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 314, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 373 + ], + "type": "text", + "content": "- We propose a Dynamic Cluster Pseudo-label Generation (DCPG) module and Distribution Shape score (DS score), which generate and assess high-quality pseudolabels based on semantic prompts to improve the detect performance." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 387, + 398, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 398, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 398, + 399 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 502, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 502, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 502, + 421 + ], + "type": "text", + "content": "2.1. Label-efficient 3D Object Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 714 + ], + "type": "text", + "content": "Recently, label-efficient 3D object detection methods have begun to be explored in responding to the challenge of extremely low annotation volumes. Generally, these label-efficient methods can be categorized into semi-supervised [15, 24, 31, 48], and weakly-supervised [16, 21, 26, 43] approaches according to the difference in quantity and supervision form. In semi-supervised methods, full annotations are provided only for some scenes, while weakly-supervised methods typically adopt weaker annotation forms. For example, VS3D [26] leverages detectors on images to guide unsupervised 3D object detection, and WS3D [21] employs point annotation as a supervision signal. To maintain accuracy while reducing the annotations, SS3D [14] introduces the concept of sparse supervision, annotating only one complete 3D object per frame. CoIn [37] then adopts a contrastive instance feature mining strategy to extract the feature-level pseudo-labels from a significantly reduced amount of annotated data. However, the performance of existing methods remains constrained due to the insufficient feature discriminability of the initial detector, which may affect subsequent training under very few annotations. This work aims to develop a two-stage strategy, enabling the 3D detector to maintain robust feature representation capabilities despite having lower instance annotation ratio." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29375" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 230, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 230, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 230, + 84 + ], + "type": "text", + "content": "2.2. Large Multimodal Models in 3D" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "type": "text", + "content": "As the outstanding performance achieved by LMMs in 2D tasks [2, 12, 18, 25, 27], some studies have begun to explore their application in the 3D domain. Inspired by CLIP [27], ULIP [40] enhances the 3D understanding capability by transferring knowledge from 2D LMM to 3D encoder through contrastive learning methods. Similar works are [46, 50]. In the outdoor scenario, SAM3D [44] employs SAM to segment BEV images of point clouds and fit bounding boxes based on the segmentation masks to obtain detection results. CLIP2Scene [5] establishes the connection between point clouds and text by using images as an intermediate modality, enhancing the 3D model's semantic understanding of the scene with the prior knowledge of CLIP. Unlike previous approaches, our SP3D explicitly transfers semantic masks obtained from LMMs onto point clouds to generate high-quality pseudo-labels for the first-stage training of the 3D detector." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 253, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 253, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 253, + 313 + ], + "type": "text", + "content": "2.3. Multimodal Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 510 + ], + "type": "text", + "content": "Recently, using multimodal methods from 2D images and 3D point clouds to enhance 3D detectors has gradually gained the community's attention [19, 28, 32, 34, 39]. However, these works mainly focus on investigating the image-point cloud fusion strategy, neglecting the utilization of images to explore label-efficient 3D detection. To reduce the required annotations, some methods transfer image information into point clouds to generate pseudo-labels [41, 51]. However, semantic ambiguity may occur at the object's edge due to the 2D-3D calibration error. MixSup [41] proposed a connected components labeling strategy, addressing this issue with the spatial separability property inherent to point clouds. SAL [23] employs density-based clustering to refine imperfect projection issues. Compared with these methods, our SP3D provides a simple but efficient way to reduce semantic noise caused by projection errors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 116, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 116, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 116, + 533 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 124, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 124, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 124, + 552 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "This paper introduces SP3D, explicitly utilizing the accurate cross-modal semantic prompts generated by LMMs to boost the sparsely-supervised 3D detectors. As shown in Fig. 3, SP3D consists of three primary components: (1) Confident Points Semantic Transfer (CPST) module, acquiring accurate cross-modal semantic prompts, which we treat as seed points, with the boundary-constrained center cluster selection. (2) Dynamic Cluster Pseudo-label Generation (DCPG) module, which dynamically generates pseudo-label proposals based on the geometric shapes within the multi-scale neighborhood of these seed points. (3) Distribution Shape score (DS score), which employs unsupervised priors as the criterion for evaluating the qual" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "ity of pseudo-label proposals, and we subsequently apply NMS [22] to retain high-quality pseudo-labels further. We then follow the training strategy of CoIn [37] to train an initial 3D detector with enhanced discriminative capacity. We detail our SP3D framework as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 139, + 542, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 542, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 542, + 152 + ], + "type": "text", + "content": "3.2. Confident Points Semantic Transfer Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 361 + ], + "type": "text", + "content": "Encouraged by the development of LMMs, we first utilize the prior knowledge of LMMs to extract semantic information from 2D images explicitly. Meanwhile, by integrating the calibration relationship matrix between 2D images and 3D point clouds, it is straightforward to transfer semantic information onto point clouds. However, as shown in Fig. 2, there is significant noise in the edge points of instances during the transfer process. These semantic points are mistakenly assigned to other instances can have a detrimental impact on subsequent pseudo-label fitting and the training of the detector. To prevent the incorporation of noise during the transfer of semantic information, we use the boundary-constrained mask shrink operation, followed by the coordinate system transformation, to obtain accurate cross-modal semantic prompts. CSPT is illustrated in Fig. 3(a); we have divided it into two parts as LMMs-guided semantic extraction and confident points filtering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": "LMMs-guided semantic extraction. The goal of LMMs is to generate high-quality foreground semantic masks. Specifically, we take as input an image " + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": " and the category-text prompt " + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathcal{C}} = \\{t_1^{\\mathcal{C}}, t_2^{\\mathcal{C}}, \\dots, t_C^{\\mathcal{C}}\\}" + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": " denote the height and width of the image, " + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 374, + 554, + 459 + ], + "type": "text", + "content": " denotes the categories of interest. We first utilize FastSAM [49] to perform speed-efficient segmentation as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 397, + 467, + 553, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 467, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 397, + 467, + 553, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {\\mathcal {I}} = \\operatorname {S A M} (\\mathcal {I}), \\tag {1}", + "image_path": "9a1c524eed95148258fa416f374b564c2ab391a8fe23a52ef45d2ccae51d421c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{I}} \\in \\mathbb{R}^{M \\times H \\times W}" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": " class-agnostic masks extracted from " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathrm{SAM}(\\cdot)" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": " indicates the FastSAM model. We then utilize " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": " as the mask prompts and feed them, along with image " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": ", into the SemanticSAM [4] model, which is expected to output the descriptions " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathcal{D}} = \\{t_1^{\\mathcal{D}}, t_2^{\\mathcal{D}}, \\dots, t_M^{\\mathcal{D}}\\}" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": " for each mask in " + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 487, + 554, + 561 + ], + "type": "text", + "content": ". Specifically:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 569, + 553, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 569, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 384, + 569, + 553, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {\\mathcal {D}} = \\operatorname {S S A M} (\\mathcal {I}, \\mathcal {M} _ {\\mathcal {I}}), \\tag {2}", + "image_path": "6af01b24f9186c7ced7cf71024141e79fa0e36f229096e4d2c9f9c7fbd34b479.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{SSAM}(\\cdot)" + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "text", + "content": " refers to the SemanticSAM model. Then, we calculate the cosine similarity between the category-text " + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "text", + "content": " and the mask description " + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathcal{D}}" + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "text", + "content": " to filter out uninteresting background masks, obtaining the foreground masks " + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{I}}'" + }, + { + "bbox": [ + 313, + 591, + 554, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Confident points filtering. As fuzziness of boundaries results from depth occlusions in images and calibration inaccuracies, we opt to constrain the boundary of foreground masks before 2D-3D transfer, retaining only its central portion. Specifically, for each foreground mask, we denote its" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29376" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 553, + 322 + ], + "type": "image", + "image_path": "f318f96b058ae74adb971c7700206bdf6311b61924d22f1751029b456ab92bf3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 363 + ], + "type": "text", + "content": "Figure 3. The overview of our SP3D, including (a) CSPT finds semantic seed points through high-confidence semantic masks transfer, (b) DCPG dynamically clusters neighbor points of seed points to fit pseudo-label proposals, and (c) DS score to evaluate the quality of generated pseudo-label proposals, serving as a scoring metric to NMS to suppress low-quality proposals." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "text", + "content": "maximum and minimum values in the pixel coordinate system " + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "inline_equation", + "content": "(u_{min},u_{max},v_{min},v_{max})" + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "text", + "content": ". We perform mask shrink to constraint boundary range of " + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{I}}^{\\prime}" + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}" + }, + { + "bbox": [ + 55, + 383, + 295, + 420 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 429, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 429, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 93, + 429, + 295, + 456 + ], + "type": "interline_equation", + "content": "u \\in \\left[ u _ {\\min } + \\frac {1}{2} (1 \\pm \\gamma) \\left(u _ {\\max } - u _ {\\min }\\right) \\right], \\tag {3}", + "image_path": "69dc3a9fa7bfc6ef77fcfc620f8ed06d36065a69dee015affd490556937f65c2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 454, + 253, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 454, + 253, + 475 + ], + "spans": [ + { + "bbox": [ + 94, + 454, + 253, + 475 + ], + "type": "interline_equation", + "content": "v \\in [ v _ {m i n} + \\frac {1}{2} (1 \\pm \\gamma) (v _ {m a x} - v _ {m i n}) ],", + "image_path": "b4c03c213d56d162d03406793aae793468d056376384b6a28852c2926057f156.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 484, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 484, + 296, + 651 + ], + "type": "text", + "content": " denotes the shrink factor. With this constraint, we obtain the shrunk masks that filter out the semantically ambiguous regions, ensuring the accuracy of foreground semantic information transferred onto the point clouds. Following [30], we transfer the semantic mask from the image onto the point clouds to obtain accurate cross-modal semantic prompts, namely seed points, using the camera's intrinsic and extrinsic parameter matrices. Note that, different from [30], we adopt the approach of explicitly transferring the shrunk masks onto the point cloud rather than implicitly embedding unprocessed semantic masks into the point cloud's features. This approach helps avoid potential semantic feature confusion between different modalities arising from sparse annotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 276, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 276, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 276, + 673 + ], + "type": "text", + "content": "3.3. Dynamic Cluster Pseudo-label Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "With the proposed CPST, we explicitly obtain the semantic seed points from transformation " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " of the foreground mask. Given " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times 3} = \\{p_1, p_2, \\dots, p_N\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " as LiDAR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "content": "points, we define the seed points covered by " + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}_{\\mathcal{I}}^{\\prime}" + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T = \\{p_t\\}" + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T \\subset \\mathcal{P}" + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "content": ". It is crucial for the 3D detection task that obtain complete bounding boxes from these seed points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 419, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 586 + ], + "type": "text", + "content": "By referring to traditional unsupervised pseudo-label generation methods [45], we produce many pseudo-labels and then use the positional constraints of the seed points to retain the more promising pseudo-labels as supervisory signals. However, existing unsupervised bounding box fitting approaches [35, 45] usually take a fixed constant as cluster radius, leading to the problem of inadequate foreground or excessive background noise for the generated bounding boxes. In this case, we propose a dynamic cluster pseudo-label generation (DCPG) module. This module utilizes the geometry shape of the seed points' multi-scale neighborhood to capture complete foreground information while minimizing background interference. It dynamically generates pseudo-label proposals, as shown in Fig. 3(b)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "content": "Specifically, we denote " + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T^{(k)}" + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "content": "-th instance in a point cloud frame and utilize DCPG dynamically generates a clustering radius " + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "content": "-th seed point " + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "inline_equation", + "content": "p_t^{(k)}" + }, + { + "bbox": [ + 313, + 587, + 554, + 639 + ], + "type": "text", + "content": ". We define the updating function for the dynamic radius as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 647, + 555, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 647, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 318, + 647, + 555, + 670 + ], + "type": "interline_equation", + "content": "\\operatorname {u p d a t e} (t, r _ {\\text {i n i t}}) = r _ {\\text {i n i t}} \\cdot \\frac {t}{N ^ {(k)}} + \\delta , t = 1, 2, \\dots , N ^ {(k)}, \\tag {4}", + "image_path": "e084926e3c2057a2de332e3cc45059b65c392bb7b1b458e6eec8bc5564234129.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "r_{init}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " is a hyper-parameter set based on empirical experience, " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " denotes adjustment factor to avoid " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " too small, and " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "N^{(k)}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " is the number of seed points in current instance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29377" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 88, + 296, + 266 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 277, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 277, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 277, + 84 + ], + "type": "text", + "content": "Algorithm 1 Dynamic cluster pseudo-label generation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "Require: LiDAR points " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": ", the seed points set " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T^{(k)}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "-th instance, initial radius " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "r_{init}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": ", ground remove function " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": ", box fitting function " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "; \nEnsure: Pseudo-label proposal set " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}^{(k)}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "; \n1: Acquire the number " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "N^{(k)}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": " of points for the seed points set of current instance " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T^{(k)}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "; \n2: " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{gr} \\gets \\mathcal{G}(\\mathcal{P})" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "; // remove ground points \n3: " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}^{(k)} = \\emptyset" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "; \n4: for " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "t = 1, 2, \\dots, N^{(k)}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": " do \n5: " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "p_t = \\mathcal{P}_T^{(k)}[t]" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": "; // loop each seed point \n6: " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "r \\gets \\text{update}(t, r_{init})" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}} \\gets \\mathcal{F}(\\text{DBSCAN}(\\mathcal{P}_{gr}, p_t, r))" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}^{(k)} \\gets \\hat{\\mathcal{B}}^{(k)} \\cup \\hat{\\mathbf{b}}" + }, + { + "bbox": [ + 56, + 88, + 296, + 266 + ], + "type": "text", + "content": " \n9: end for" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "type": "text", + "content": "By applying Eq. 4, we dynamically update the radius " + }, + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "type": "inline_equation", + "content": "r \\in (\\delta, r_{init} + \\delta]" + }, + { + "bbox": [ + 55, + 280, + 296, + 364 + ], + "type": "text", + "content": ", during point clustering, thereby obtaining foreground clusters with multi-scale receptive fields. Following [45]. We utilize the radius calculated from Eq. 4 as the clustering radius for DBSCAN [8] and employ [47] to fit the bounding box for each foreground cluster. Algorithm 1 summarizes Our DCPG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 195, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 195, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 195, + 383 + ], + "type": "text", + "content": "3.4. Distribution Shape Score" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 590 + ], + "type": "text", + "content": "While DCPG has the capacity for high-quality pseudolabel generation, it unavoidably produces an amount of low-quality pseudo-label proposals. The shapes of these proposals and the extent of foreground completeness contained within the proposals usually deviate significantly from reality. Traditional detection methods typically compute the Intersection over Union (IoU) score between predicted bounding boxes and ground-truth (GT) boxes and then employ the NMS [22] to suppress these low-quality proposals. However, lacking GT makes it challenging to directly apply NMS using IoU as the evaluation criterion within our SP3D framework. In this case, we propose a distribution shape score (DS score) to assess the quality of the pseudo-labels using unsupervised prior knowledge. As illustrated in Fig. 3(c), we divided the DS score into two parts: distribution constraint score and meta-shape constraint score." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": "Distribution constraint score. Inspired by [20], within a high-quality pseudo-label proposal " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", the distances from its interior points " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{i,i=1,\\dots,n}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " to its boundary roughly follow a Gaussian distribution " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\sigma)" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mu = 0.8" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma = 0.2" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", respectively. In other words, we denote random variable " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "D = \\{d_1, \\dots, d_n\\}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " as the distance between " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " and the box boundary of " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "D \\sim \\mathcal{N}(0.8, 0.2)" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ". Based on this prior, we assign a distribution constraint score to the pseudo-label proposal " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " by calculating the similarity between the random" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "variable " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " corresponding to each " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " and the normal distribution " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ". Specifically:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 351, + 104, + 555, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 104, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 351, + 104, + 555, + 135 + ], + "type": "interline_equation", + "content": "s _ {d c} (\\hat {b}) = \\frac {1}{| \\mathcal {P} _ {f g} |} \\sum_ {p _ {i} \\in \\mathcal {P} _ {f g}} \\log (\\mathcal {N} (D | \\mu , \\sigma)), \\tag {5}", + "image_path": "a9fcfb7d7fff3b63c7f4e3021653ab40a472ed36739278331823d170a4977fda.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\log (\\cdot)" + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "content": " denotes the logarithm function, " + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{fg}" + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "content": " is the foreground points within " + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}_{fg}|" + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "content": " represents the number of points in " + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{fg}" + }, + { + "bbox": [ + 313, + 143, + 555, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": "Meta-shape constraint score. In addition, the shape of a high-quality pseudo-label is expected to be consistent with its template in the real world, which we define as the meta instance, corresponding to its category [35]. For class " + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_c \\in \\{l_c, w_c, h_c\\}" + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": " as the shape of its meta instance, where " + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "inline_equation", + "content": "l_c" + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "inline_equation", + "content": "w_c" + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "inline_equation", + "content": "h_c" + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": " are the normalized length, width and height, respectively. We followed this shape prior to constructing the class-aware meta-shape constraint score " + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "inline_equation", + "content": "s_{msc}(\\hat{b})" + }, + { + "bbox": [ + 313, + 195, + 555, + 303 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 312, + 555, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 312, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 372, + 312, + 555, + 327 + ], + "type": "interline_equation", + "content": "s _ {m s c} (\\hat {b}) = 1 - \\Phi_ {K L} (\\mathcal {B} _ {c} | | \\hat {\\mathcal {B}} _ {\\hat {b}}), \\tag {6}", + "image_path": "a534cc5fae6d76f442b45f3cbeb411653920bf53cea095785f65d56707953116.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\Phi_{KL}(\\cdot)" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": " denotes the normalized KL divergence function, and " + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{B}}_{\\hat{b}} \\in \\{l_{\\hat{b}}, w_{\\hat{b}}, h_{\\hat{b}}\\}" + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": " indicates the normalized shape of the pseudo-label proposal. The purpose of this operation is to suppress the low-quality proposals whose shape deviates significantly from the meta instance. By combining the distribution constraint score and the meta-shape constraint score, we can obtain the DS score as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 427, + 555, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 427, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 367, + 427, + 555, + 442 + ], + "type": "interline_equation", + "content": "\\mathrm {D S} (\\hat {b}) = \\lambda_ {1} \\bar {s} _ {d c} (\\hat {b}) + \\lambda_ {2} \\bar {s} _ {m s c} (\\hat {b}), \\tag {7}", + "image_path": "2dc818f2e851162124b26718482c4096e948fa11efe9f2970f4251b0782d8bcf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "text", + "content": " denote weight adjustment factor. Notably, to unify the dimension, we normalized the two constraint scores before combining them, resulting in " + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\overline{s}_{dc}" + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\overline{s}_{msc}" + }, + { + "bbox": [ + 313, + 450, + 555, + 558 + ], + "type": "text", + "content": ". We then employ the DS score as a substitution for the confidence score in NMS to suppress the low-quality pseudolabels. We utilize the obtained pseudo-labels in conjunction with CoIn [37] for the first phase of training. Subsequently, we fine-tune the trained detector with a small amount of accurate labels to boost the performance of the 3D detector." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 569, + 395, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 395, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 395, + 582 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 588, + 432, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 432, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 432, + 600 + ], + "type": "text", + "content": "4.1. Dataset and Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "KITTI dataset. The KITTI [9] dataset is a widely used benchmark for 3D object detection. During the first training stage, we relied solely on the semantic information provided by LMMs to generate pseudo-labels instead of the human-annotated ground truth. In the fine-tuning stage, we split the training set (7,481 scenes) into a train split (3,712 scenes) and a val split (3,769 scenes). We follow [37] to randomly select " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " of the scenes from the train split and retain only one instance annotation per scene. In this case," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29378" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 547, + 184 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 184 + ], + "type": "table", + "html": "
SettingCostMethodCarPedCyc
EasyMod.HardEasyMod.HardEasyMod.Hard
Fully-sup.100%VoxelRCNN92.384.982.669.663.058.688.772.568.2
Sparsely-sup.20%SS3D89.384.278.2------
2%VoxelRCNN70.554.944.842.638.532.173.347.843.2
2%CoIn89.170.255.650.845.239.680.252.348.6
2%CoIn++92.079.571.546.736.131.282.058.454.6
2%CoIn++ with SP3D91.380.574.067.458.750.992.573.168.3
", + "image_path": "b09a8372836e45d4724e34543f85c29d6ea7b060e3d04102ff26769672cd45bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 228, + 563, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 194, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 218 + ], + "type": "text", + "content": "Table 1. Comparison with SoTA sparsely-supervised methods on KITTI val split. All methods are based on VoxelRCNN, and we report the 3D AP results of full cost (100%) and limited cost (20%, 2%). The best sparsely-supervised methods are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 228, + 563, + 404 + ], + "lines": [ + { + "bbox": [ + 56, + 228, + 563, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 563, + 404 + ], + "type": "table", + "html": "
StageLabelMethodsCar-3D @IoU 0.7Car-BEV @IoU 0.7
EasyModHardEasyModHard
Single-stageSparse1. CenterPoint49.6931.5525.9156.7842.5034.14
Sparse2. CoIn (CenterPoint-based)72.0354.8243.7787.2073.5466.03
Sparse3. 2 with SP3D (CenterPoint-based)87.4469.2458.6192.7280.0069.01
-4. Improvements 3→2+15.41+14.42+14.84+5.52+6.46+2.98
Two-stageSparse1. Voxel-RCNN70.5254.9744.8283.6771.1457.71
Sparse2. CoIn (Voxel-RCNN-based)84.5668.4758.0292.3181.0170.24
Sparse3. 2 with SP3D (Voxel-RCNN-based)91.3774.8963.8495.4185.2774.57
-4. Improvements 3→2+6.81+6.42+5.82+3.1+4.26+4.33
Multi-stageSparse1. CasA74.1857.3745.0585.9073.2157.23
Sparse2. CoIn (CasA-based)89.1775.3262.9895.9985.0272.47
Sparse3. 2 with SP3D (CasA-based)91.1275.9466.4696.5585.6576.31
-4. Improvements 3→2+1.95+0.62+3.48+0.56+0.63+3.84
", + "image_path": "f2818196e249a8e880d1b94bfd63f9d0766d9a9d6f9ee9f784b1c8f32c8974e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 455, + 282, + 523 + ], + "blocks": [ + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 555, + 437 + ], + "type": "text", + "content": "Table 2. Comparison with different fully-supervised methods. Sparse label refers to the use of limited split (2% annotation cost). The 3D object detection and BEV detection benchmark are evaluated by mAP with R40, under IoU threshold 0.7." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 455, + 282, + 523 + ], + "lines": [ + { + "bbox": [ + 70, + 455, + 282, + 523 + ], + "spans": [ + { + "bbox": [ + 70, + 455, + 282, + 523 + ], + "type": "table", + "html": "
MethodsVeh. LEVEL 1Veh. LEVEL 2
mAPmAPHmAPmAPH
CoIn39.639.034.233.6
CoIn+SP3D46.746.040.439.9
Improvement+7.1+7.0+6.2+6.3
", + "image_path": "6ae4b437e905330d286f398a5c654d605f2985564abd7b9cdc74edc055bde669.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 295, + 565 + ], + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 565 + ], + "type": "text", + "content": "Table 3. Comparison with SoTA sparsely-supervised methods on WOD validation set. All methods are based on CenterPoint, and we report the results at a " + }, + { + "bbox": [ + 55, + 531, + 295, + 565 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 531, + 295, + 565 + ], + "type": "text", + "content": " annotation cost and 0.7 IoU threshold." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 591, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 639 + ], + "type": "text", + "content": "we can obtain a limited split, which merely takes " + }, + { + "bbox": [ + 55, + 591, + 295, + 639 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 591, + 295, + 639 + ], + "type": "text", + "content": " of instance annotations compared with the origin train split. All results are evaluated using the official evaluation metric: 3D Average Precision (AP) across 40 recall thresholds (R40)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Waymo Open Dataset (WOD). The WOD is a widely used large-scale benchmark that contains 798, 202, and 150 sequences (approximately 200 frames per sequence) for training, validation, and testing, respectively. All results are evaluated by using 3D mean Average Precision (mAP) and its variant weighted by heading accuracy (mAPH)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 448, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 448, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 448, + 468 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": "Pseudo-labels generation. We directly employed the model parameters provided by FastSAM [49] and SemanticSAM [4] for inference without additional supervisory signals for fine-tuning. To achieve accurate segmentation results, we set a higher segmentation threshold of 0.7 during the FastSAM inference process. To mitigate computational demands, we opted to generate pseudo-labels within a confined spatial domain of the semantically relevant points, specifically within an 8-meter radius. We set mask shrink factor " + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": " to 0.3, initial cluster radius " + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "inline_equation", + "content": "r_{init}" + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": " to 1, adjustment factor " + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": " to 0.1. We utilize unsupervised priors to filter out the low-quality pseudo-labels that are evidently inconsistent with the intuitive expectations and set the weight adjustment factor of DS score " + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": " as 0.5, 0.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Detector training. We conduct all experiments with a batch size of 8 and a learning rate of 0.003 on 4 RTX 3090 GPUs. Following previous sparsely-supervised 3D object detection methods [14, 37], we choose three different clas" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29379" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 561, + 223 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 561, + 223 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 561, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 561, + 223 + ], + "type": "table", + "html": "
Methods2%1%0.5%
CarPed.Cyc.Avg.CarPed.Cyc.Avg.CarPed.Cyc.Avg.
semi-sup.HSSDA81.958.265.868.680.951.942.359.577.819.544.147.2
HSSDA+SP3D83.260.874.472.883.053.269.168.482.322.146.450.3
Improvement+1.3+2.6+8.6+4.2+2.1+1.3+26.8+8.9+4.5+2.6+2.3+3.1
sparsely-sup.CoIn72.336.958.455.865.331.552.249.655.225.530.537.0
CoIn+SP3D73.444.762.360.171.736.959.155.963.336.143.447.6
Improvement+1.1+7.8+3.9+4.3+6.4+5.4+6.9+6.3+8.1+10.6+12.5+10.6
HINTED85.160.081.375.482.849.974.168.980.245.864.563.5
HINTED+SP3D85.560.181.475.684.057.281.074.083.249.273.968.7
Improvement+0.4+0.1+0.1+0.2+1.2+7.3+6.9+5.1+3.0+3.4+9.4+5.2
", + "image_path": "94870e223ed798be4e2b189ab7a8b03db43815291f116a33cf8e7634e4745510.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 283, + 305, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 231, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 266 + ], + "type": "text", + "content": "Table 4. Improvement of SP3D in SoTA semi-supervised and sparsely-supervised methods on KITTI val split under limited cost setting. The results are evaluated by mAP with R40, under IoU thresholds 0.7, 0.5, and 0.5 for 'Car', 'Pedestrian', and 'Cyclist', respectively. Additionally, the AP is computed across the difficulty levels of easy, moderate, and hard." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 283, + 305, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 283, + 305, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 305, + 350 + ], + "type": "table", + "html": "
MethodsModalityCar-3D @IOU 0.5/0.7
Train TestEasyMod.Hard
VS3DL+CL40.32/9.0937.36/5.7331.09/5.03
WS3DPRL+CL-/60.01-/44.48-/36.93
SP3D (Ours)L+CL93.75/69.7176.36/48.6571.01/40.53
", + "image_path": "59aacb20c0a6a73da93f15711da66a5080e4461e4cf52d9af699b0d27dd943d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 394, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 453 + ], + "type": "text", + "content": "sical detectors [6, 33, 42] as our architecture. And we employ the OpenPCDet [29] to conduct our experiments. In the first training stage, we employ CoIn [37] to train an initial detector with the generated pseudo-labels. Then, we use limited split to fine-tune the detector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "type": "text", + "content": "Baselines. To thoroughly validate the effectiveness of SP3D, we select the SoTA sparsely-supervised method [37] as the primary baseline for comparison. We compare the proposed SP3D approach with the baseline under conventional sparse settings with " + }, + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 471, + 295, + 602 + ], + "type": "text", + "content": " annotation cost. We also compared with cross-modal weakly-supervised methods [16, 26] in a zero-shot setting, which also incorporate visual models to extract semantic information to enhance the performance of detectors. Furthermore, we establish baselines under progressively reduced annotation costs to evaluate the sensitivity to annotation costs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 232, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 232, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 232, + 624 + ], + "type": "text", + "content": "4.3. Comparison with SoTA Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Comparison with sparsely-supervised methods. We integrate our proposed SP3D into the SoTA sparsely-supervised 3D detection algorithm, CoIn++ [37]. For a fair comparison, all detectors employ the VoxelRCNN [6] as the base architecture. As illustrated in Tab. 1, SP3D significantly improves the detection performance of CoIn++. The experimental results on WOD, as reported in Tab. 3," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 283, + 547, + 352 + ], + "blocks": [ + { + "bbox": [ + 61, + 360, + 289, + 372 + ], + "lines": [ + { + "bbox": [ + 61, + 360, + 289, + 372 + ], + "spans": [ + { + "bbox": [ + 61, + 360, + 289, + 372 + ], + "type": "text", + "content": "Table 5. Comparison of zero-shot methods on KITTI val split." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 283, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 322, + 283, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 322, + 283, + 547, + 352 + ], + "type": "table", + "html": "
MethodsModalityVeh.Ped.Cyc.
TrainTest
SAM3DLL19.10.00.0
CM3DL + CL23.70.12.6
SP3D (Ours)L + CL35.015.75.9
", + "image_path": "0a079060592ca81eb002f946b5e53f71d1a721634d10ec02756d45ffa871df51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 360, + 554, + 394 + ], + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 394 + ], + "type": "text", + "content": "Table 6. Comparison with zero-shot methods on the WOD validation set. We report Level-2 mAP with IoU thresholds of 0.7, 0.5, and 0.5 for Vehicle, Pedestrian, and Cyclist, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "content": "also validate the performance enhancement of our method for the sparsely-supervised 3D object detector. However, we observe a slight decrease in precision for the 'Easy' car category with our SP3D-initialized model in Tab. 1. This could arise because our initial pseudo-labels are inferred based on the geometric shape of the objects, which may differ from the conventions of manual annotation. When the point cloud structure of an instance is relatively intact, such discrepancies can lead to noticeable differences in the size of the annotated bounding boxes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Comparison with fully-supervised methods. For a fair comparison, we follow CoIn [37] to select CenterPoint [42], VoxelRCNN [6], and CasA [33] as our baseline detectors, representing three distinct types of detection algorithms. We initialize the 3D detector using our SP3D, followed by fine-tuning with the limited split. As shown in Tab. 2, detectors designed for full supervision struggle to yield optimal results due to scarce annotations. Despite the effectiveness of CoIn in improving this situation, the results achieved are still unsatisfactory for single-stage detection algorithms with relatively simple structures. Our designed strategy, SP3D, significantly reduces this discrepancy, enabling detectors to achieve similar results." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29380" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 272, + 137 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 272, + 137 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 272, + 137 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 272, + 137 + ], + "type": "table", + "html": "
MethodsCar-3D @IoU 0.5
EasyMod.Hard
CPD61.4861.5860.09
CPD+SP3D73.6566.1363.78
Improvement+12.17+4.55+2.83
", + "image_path": "03cebefa7dcde1008488b3683e76ab059e9d4f48ec231474545fe49be6daf8ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 181, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 294, + 301 + ], + "type": "text", + "content": "Comparison with zero-shot methods. We also compared SP3D with SoTA zero-shot methods. SAM3D [44] and CM3D [11] generate 3D pseudo-labels by leveraging large models for the semantic segmentation of 2D images. In line with the zero-shot approach of prior methods, no 3D annotations were used for training; instead, pseudolabels were generated from large models and image data. As shown in Tab. 5 and Tab. 6, the results on the KITTI and Waymo datasets demonstrate that our method significantly outperforms previous approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "text", + "content": "Comparison with different annotation rates. To explore the robustness of SP3D, we conducted experiments on detectors across different supervision paradigms under varying annotation rates, and we reported the results on Tab. 4. For a fair comparison, we initialize the detectors by following the strategies of CoIn (CenterPoint) [37], HSSDA (PV-RCNN) [15], and HINTED (Voxel-RCNN) [38]. The experimental results show that when the annotation rate is limited, detector performance drops sharply as the amount of annotated data decreases. By inheriting SP3D, these detectors' performance has been significantly enhanced, mitigating the performance drop associated with reduced annotation. Taking CoIn as an example, its average AP for the three categories at annotation rates of " + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "text", + "content": " has improved by " + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 535 + ], + "type": "text", + "content": ", respectively. These results indicate that even at low annotation rates, SP3D-initialized models can significantly improve the performance of 3D detectors across different supervision paradigms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 554, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 625 + ], + "type": "text", + "content": "Comparison with unsupervised methods. To further validate the robustness of SP3D, we integrated it with the SoTA unsupervised detection method CPD [35]. As shown in Tab. 7, despite suffering some impact, the combination of SP3D and CPD still enhances the performance of the unsupervised 3D object detector to a certain extent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "We conducted ablation studies based on CenterPoint to verify the proposed modules' effectiveness and recorded the results in Tab. 8. The first and third rows show that the precision of pseudo-labels, as augmented by DCPG, can substantially amplify the 3D detector's capabilities. This may be" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 70, + 550, + 162 + ], + "blocks": [ + { + "bbox": [ + 87, + 147, + 263, + 158 + ], + "lines": [ + { + "bbox": [ + 87, + 147, + 263, + 158 + ], + "spans": [ + { + "bbox": [ + 87, + 147, + 263, + 158 + ], + "type": "text", + "content": "Table 7. Improvement for unsupervised method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 70, + 550, + 162 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 550, + 162 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 550, + 162 + ], + "type": "table", + "html": "
Mask shrinkDCPGDS scoreCar-3D AP@IoU 0.7
EasyMod.Hard
✓35.1023.7519.52
✓✓38.2629.2524.74
✓✓40.5628.1522.40
✓✓47.2333.4027.13
✓✓✓52.5638.0031.52
", + "image_path": "3affdc84ebd974da68791ee80d8ff265f2622dd4d8bbd03fc2b09e43059ce06d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 342, + 195, + 527, + 251 + ], + "blocks": [ + { + "bbox": [ + 354, + 171, + 512, + 182 + ], + "lines": [ + { + "bbox": [ + 354, + 171, + 512, + 182 + ], + "spans": [ + { + "bbox": [ + 354, + 171, + 512, + 182 + ], + "type": "text", + "content": "Table 8. Ablation study on KITTI val split." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 195, + 527, + 251 + ], + "lines": [ + { + "bbox": [ + 342, + 195, + 527, + 251 + ], + "spans": [ + { + "bbox": [ + 342, + 195, + 527, + 251 + ], + "type": "table", + "html": "
DS scoreCar-3D AP@IoU 0.7
MSCDCEasyMod.Hard
✓44.9133.5528.19
✓48.4835.3929.75
", + "image_path": "711b0c79f6966068c9ffad49b79d0fe095992f3cc51c40d799b92d9b385a779d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 260, + 535, + 271 + ], + "lines": [ + { + "bbox": [ + 331, + 260, + 535, + 271 + ], + "spans": [ + { + "bbox": [ + 331, + 260, + 535, + 271 + ], + "type": "text", + "content": "Table 9. Ablation study of each component of DS score." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 294, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 403 + ], + "type": "text", + "content": "due to incorporating more comprehensive foreground information in pseudo-labels, which enhanced the model's feature discrimination capability. The last two rows show that the mask shrink operation is necessary for handling semantic noise at the instance edges. The first three rows and fifth row show that using the DS score to filter out low-quality labels can significantly boost the detector's precision. Moreover, we validated the effectiveness of each component in DS score with the ablation studies presented in Tab. 9." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 416, + 466, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 466, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 466, + 428 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "content": "This paper proposes a two-stage boosting strategy, SP3D, that leverages accurate cross-modality semantic prompts to enhance the sparsely-supervised 3D detectors' capabilities. We first develop a CSPT module to obtain accurate cross-modal semantic prompts in point clouds by efficiently transferring high-confidence semantic masks. We then design a DCPG module to dynamically generate pseudo-label proposals based on semantic prompts. Lastly, we propose a DS score as the criterion for NMS to select high-quality pseudo-labels integrated with sparsely-supervised strategy to train the initial detector. After fine-tuning with sparse GT data, SP3D show superior performance over the original sparsely-supervised 3D detector on the WOD and KITTI dataset, and remain robust performance even as the amount of annotation decreased. Limitations: SP3D exhibits a relatively significant degradation when fine-tuning with annotation rates below " + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "content": ", which may result from the noise introduced by extremely low annotations. Future efforts to explore efficient fine-tuning strategies to solve this issue." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (No.62171393), and the Fundamental Research Funds for the Central Universities (No.20720220064)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29381" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "type": "text", + "content": "[2] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 294 + ], + "type": "text", + "content": "[4] Jiaqi Chen, Zeyu Yang, and Li Zhang. Semantic segment anything. https://github.com/fudan-zvg/Semantic-Segment-Anything, 2023.3,6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 360 + ], + "type": "text", + "content": "[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 294, + 418 + ], + "type": "text", + "content": "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 508 + ], + "type": "text", + "content": "[8] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 554 + ], + "type": "text", + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 610 + ], + "type": "text", + "content": "[10] Xun Huang, Hai Wu, Xin Li, Xiaoliang Fan, Chenglu Wen, and Cheng Wang. Sunshine to rainstorm: Cross-weather knowledge distillation for robust 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2409–2416, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "type": "text", + "content": "[11] Mehar Khurana, Neehar Peri, James Hays, and Deva Ramanan. Shelf-supervised cross-modal pre-training for 3d object detection. In 8th Annual Conference on Robot Learning, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[13] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 194 + ], + "type": "text", + "content": "[14] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Jiang Liu, Deyu Meng, and Xinbo Gao. Ss3d: Sparsely-supervised 3d object detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8428-8437, 2022. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "type": "text", + "content": "[15] Chuandong Liu, Chenqiang Gao, Fangcen Liu, Pengcheng Li, Deyu Meng, and Xinbo Gao. Hierarchical supervision and shuffle data augmentation for 3d semi-supervised object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23819-23828, 2023. 1, 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 261, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 553, + 326 + ], + "type": "text", + "content": "[16] Haizhuang Liu, Huimin Ma, Yilin Wang, Bochao Zou, Tianyu Hu, Rongquan Wang, and Jiansheng Chen. Eliminating spatial ambiguity for weakly supervised 3d object detection without spatial labels. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3511-3520, 2022. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 380 + ], + "type": "text", + "content": "[17] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "text", + "content": "[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "text", + "content": "[19] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "text", + "content": "[20] Katie Luo, Zhenzhen Liu, Xiangyu Chen, Yurong You, Sagie Benaim, Cheng Perng Phoo, Mark Campbell, Wen Sun, Bharath Hariharan, and Kilian Q Weinberger. Reward finetuning for faster and more accurate unsupervised object discovery. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "text", + "content": "[21] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Yunde Jia, and Luc Van Gool. Towards a weakly supervised framework for 3d point cloud object detection and annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4454-4468, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[22] Alexander Neubeck and Luc Van Gool. Efficient nonmaximum suppression. In 18th international conference on pattern recognition (ICPR'06), pages 850-855. IEEE, 2006. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[23] Aljosa Osep, Tim Meinhardt, Francesco Ferroni, Neehar Peri, Deva Ramanan, and Laura Leal-Taixe. Better call sal: Towards learning to segment anything in lidar. In European Conference on Computer Vision, pages 71-90. Springer, 2025. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29382" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Jinhyung Park, Chenfeng Xu, Yiyang Zhou, Masayoshi Tomizuka, and Wei Zhan. Detmatch: Two teachers are better than one for joint 2d and 3d semi-supervised object detection. In European Conference on Computer Vision, pages 370-389. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[25] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[26] Zengyi Qin, Jinglu Wang, and Yan Lu. Weakly supervised 3d object detection from point clouds. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4144-4152, 2020. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[28] Ziying Song, Feiyang Jia, Hongyu Pan, Yadan Luo, Caiyan Jia, Guoxin Zhang, Lin Liu, Yang Ji, Lei Yang, and Li Wang. Contrastalign: Toward robust bev feature alignment via contrastive learning for multi-modal 3d object detection. arXiv preprint arXiv:2405.16873, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 364 + ], + "type": "text", + "content": "[29] OD Team et al. Openpcdet: An open-source toolbox for 3d object detection from point clouds, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[30] Sourabh Vora, Alex H Lang, Bassam Helou, and Oscar Bei-jbom. Pointpainting: Sequential fusion for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4604-4612, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[31] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[32] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6792-6802, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[33] Hai Wu, Jinhao Deng, Chenglu Wen, Xin Li, Cheng Wang, and Jonathan Li. Casa: A cascade attention network for 3-d object detection from lidar point clouds. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[34] Hai Wu, Chenglu Wen, Shaoshuai Shi, Xin Li, and Cheng Wang. Virtual sparse convolution for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21653-21662, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Hai Wu, Shijia Zhao, Xun Huang, Chenglu Wen, Xin Li, and Cheng Wang. Commonsense prototype for outdoor unsupervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14968-14977, 2024. 2, 4, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Qiming Xia, Yidong Chen, Guorong Cai, Guikun Chen, Daoshun Xie, Jinhe Su, and Zongyue Wang. 3-d hanet: A" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "flexible 3-d heatmap auxiliary network for object detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[37] Qiming Xia, Jinhao Deng, Chenglu Wen, Hai Wu, Shaoshuai Shi, Xin Li, and Cheng Wang. Coin: Contrastive instance feature mining for outdoor 3d object detection with very limited annotations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6254-6263, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 251 + ], + "type": "text", + "content": "[38] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 319 + ], + "type": "text", + "content": "[39] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 320, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 396 + ], + "type": "text", + "content": "[40] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[41] Yuxue Yang, Lue Fan, and Zhaoxiang Zhang. Mixsup: Mixed-grained supervision for label-efficient lidar-based 3d object detection. arXiv preprint arXiv:2401.16305, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "text", + "content": "[42] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[43] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8373-8383, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "type": "text", + "content": "[44] Dingyuan Zhang, Dingkang Liang, Hongcheng Yang, Zhikang Zou, Xiaqing Ye, Zhe Liu, and Xiang Bai. Sam3d: zero-shot 3d object detection via the segment anything model. Science China Information Sciences, 67(4):149101, 2024. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 591, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 655 + ], + "type": "text", + "content": "[45] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9317-9328, 2023. 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[46] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29383" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 317 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[47] Xiao Zhang, Wenda Xu, Chiyu Dong, and John M Dolan. Efficient l-shape fitting for vehicle detection using laser scanners. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 54-59. IEEE, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "text", + "content": "[48] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembleing semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 196 + ], + "type": "text", + "content": "[49] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything, 2023. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 198, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 295, + 261 + ], + "type": "text", + "content": "[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2639-2650, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 263, + 294, + 317 + ], + "type": "text", + "content": "[51] Pufan Zou, Shijia Zhao, Weijie Huang, Qiming Xia, Chenglu Wen, Wei Li, and Cheng Wang. Adaco: Overcoming visual foundation model noise in 3d semantic segmentation via adaptive label correction. arXiv preprint arXiv:2412.18255, 2024. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29384" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_content_list.json b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f9d9b7eb22cefe2e13477a3b2c95ac5964db4f1 --- /dev/null +++ b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_content_list.json @@ -0,0 +1,1534 @@ +[ + { + "type": "text", + "text": "SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models", + "text_level": 1, + "bbox": [ + 119, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongting Zhang $^{1,3*}$ Lu Chen $^{2,3*}$ Guodong Zheng $^{2,3}$ Yifeng Gao $^{2}$ Rui Zheng $^{2}$ Jinlan Fu $^{3}$ Zhenfei Yin $^{3}$ Senjie Jin $^{2}$ Yu Qiao $^{3}$ Xuanjing Huang $^{2}$ Feng Zhao $^{1\\dagger}$ Tao Gui $^{2,3‡}$ Jing Shao $^{3\\dagger}$ ${}^{1}$ University of Science and Technology of China ${}^{2}$ Fudan University ${}^{3}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 217, + 200, + 789, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zytabcd@mail.ustc.edu.cn, luchen23@m.fudan.edu.cn", + "bbox": [ + 272, + 295, + 723, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://sqrtizhang.github.io/SPA-VL/", + "bbox": [ + 334, + 311, + 656, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 367, + 326, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness.", + "bbox": [ + 89, + 398, + 485, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 777, + 223, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs), such as GPT-4V [36], Claude 3 [1], LLaVA [27], and MiniGPT-4 [62] can understand visual signals and respond to user instructions.", + "bbox": [ + 89, + 804, + 483, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Equipped with a visual encoder module, VLMs extract multimodal knowledge from both visual and textual queries, leveraging pre-trained LLMs' powerful comprehension and generative capabilities to achieve remarkable results across diverse vision-language tasks.", + "bbox": [ + 511, + 369, + 906, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the complexity of multimodal harms, previous study [16] has demonstrated that harmless inputs may also result in outputs that do not align with human preferences. Although LLMs have undergone harmless alignment, the alignment of visual encoders is relatively weak, making VLMs susceptible to successful attacks through the visual modality [6, 14, 25]. Therefore, it is necessary to simultaneously improve the alignment of the visual and language modules of the VLM to achieve the harmless and helpful responses. Ensuring the alignment of VLMs with ethical and safety standards is crucial for their safe and effective deployment. However, most existing works on the safety of VLMs focused on the evaluation benchmarks [10, 24, 26] or jailbreak detection [6, 14, 14, 23-25]. Seldom studies considered constructing large-scale, high-quality open-source training datasets to achieve the safety alignment of VLMs. The lack of such datasets hampers further development in this field.", + "bbox": [ + 511, + 446, + 908, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose a large-scale safety alignment dataset for VLMs named SPA-VL. Since Reinforcement Learning from Human Feedback (RLHF) is widely regarded as performing well in alignment studies [20, 41], our SPA-VL dataset is designed for the RLHF, with each sample containing four elements (question, image, chosen response, rejected response). The main perspectives of our SPA-VL dataset are summarized as follows:", + "bbox": [ + 511, + 718, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(1) Comprehensive Domains: SPA-VL contains 100,788 samples and comprehensively covers a wide range of harm types, encompassing 6 domains, 13 categories, and 53 subcategories. (2) Diverse Questions and Responses:", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Authorship order determined by coin flip.", + "bbox": [ + 107, + 862, + 441, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 875, + 241, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19867", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For diverse question collection, we gathered three types of questions for each image: easy question, hard question, and hard statement. For diverse response collection, to enhance the diversity of the collected responses and reduce specific model biases, we selected responses from 2 out of 12 different models for each question, designating them as chosen and rejected responses. (3) Multi-Objective Alignment: The preference dataset is designed according to two alignment objectives: harmlessness and helpfulness. Thus, the responses of the aligned VLMs are enhanced in both aspects without sacrificing helpfulness for safety.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are listed below:", + "bbox": [ + 109, + 257, + 385, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We propose the SPA-VL, a large-scale, high-quality, and domain diversity dataset for vision language model safety alignment studies. By using techniques such as PPO and DPO for safety alignment learning on the SPA-VL dataset based on some open VLMs, such as LLaVA, experimental results show that these VLMs significantly improve performance in safety and surpass the state-of-the-art VLMs.", + "(2) We conducted extensive analysis and ablation studies, finding that increasing the scale of the dataset, incorporating diverse answers, and using various question types can improve the safety and performance of the alignment VLMs.", + "(3) The entire dataset construction process is fully automated, ensuring efficient and scalable data generation. The automation covers all key steps, including the collection of images and questions, the selection of VLM responses, and the annotations of chosen and rejected responses." + ], + "bbox": [ + 89, + 272, + 480, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 558, + 232, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Models. The rapid advancement of Large Language Models [42, 59] and their robust foundational capabilities have significantly prompted the development of multimodal large models. Recently, Vision-Language Models have emerged as a promising extension of LLMs [9], integrating visual and textual information for enhanced multimodal understanding. Notable models in this domain include InstructBLIP [11], InternLMXComposer [57], LAMM-SFT [56], LAMM [56], LLaMA-Adapter-V2 [13], MiniGPT-4 [62], mPLUG-Owl [55], Otter [21], and Qwen-VL-Chat [3]. Most of these VLMs are developed by projecting the vision space into the language space through a learned projector, leveraging pre-trained language models as their backbone. As VLMs continue to advance rapidly, safety concerns have garnered significant attention from researchers.", + "bbox": [ + 89, + 583, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reinforcement Learning from Human Feedback. Despite the promising capabilities of LLMs and VLMs, they are prone to unintended behaviors, such as fabricating facts, producing biased or harmful content, or even harming humans [7, 8]. They should be helpful, honest, and harm-", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "less (3H) [4, 37, 49]. RLHF offers the most straightforward approach to achieving this goal. RLHF methods such as PPO [45] and DPO [43] have been highly successful in aligning AI with human preferences. Notable applications like ChatGPT [34] and Claude [1] show strong performances in academic benchmarks. Models trained with RLHF methods often perform better and adhere more closely to human values compared to those trained only with SFT [4]. This success extends to VLMs, where RLHF has been used to address hallucination issues [5, 39, 48, 61].", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Safety of VLMs. To evaluate the safety performance of VLMs, various methods and datasets have been proposed. Among these evaluation benchmarks are MM-SafetyBench [29], ChEf [47], and OODCV-VQA, SketchyVQA [50]. In addition to these benchmarks, several attack methods, such as adversarial attacks [12, 40, 46, 60] and jailbreak techniques [14, 24, 33], have been developed to test the vulnerabilities of VLMs. These studies aim to identify and exploit potential weaknesses in VLMs, underscoring the need for robust safety measures.", + "bbox": [ + 511, + 242, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these vulnerabilities, several methods have been developed to enhance the safety performance of VLMs. For instance, VLGuard [63] employed supervised fine-tuning (SFT) on the VLGuard dataset, which contains 2000 training images designed to improve safety. Similarly, Chen et al. [10] used AI-annotated data for SFT. Wei et al. [52] utilized in-context learning to bolster model safety. Additionally, Pi et al. [38] introduced MLLM-Protector, a harm detector serving as a plug-and-play defense mechanism for VLMs, and Wang et al. [51] applied inference time alignment methods to improve harmlessness. These approaches collectively demonstrate ongoing efforts to mitigate risks and enhance the resilience of VLMs against various types of attacks.", + "bbox": [ + 511, + 393, + 903, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Among all the methods mentioned above, we can broadly categorize them into two types. The first type enhances model safety during the inference stage by using prompts. This approach is efficient and convenient but often results in limited safety improvements and lacks generalization [17]. The second type involves training-based methods, specifically during the training-to-align phase, which can be further divided into SFT and RLHF. While the aforementioned methods primarily rely on SFT, we go a step further by providing a comprehensive RLHF dataset SPA-VL.", + "bbox": [ + 511, + 604, + 903, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. SPA-VL Dataset", + "text_level": 1, + "bbox": [ + 511, + 768, + 676, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the development of VLMs, effectively addressing harmful content in multimodal environments poses a significant challenge. The SPA-VL dataset helps VLMs confront this challenge by providing safety preference data for RLHF. As shown in Figure 1, establishing the SPA-VL dataset involves three key stages. First, we systematically collect images, which includes gathering a diverse set of images to", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19868", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef9d1caf78fa225218ff69358afcc7685cb10346ce86d00e12cb2451299d1aec.jpg", + "image_caption": [ + "Figure 1. Overview of SPA-VL Dataset Construction. It is built in three stages: 1) Image Collection, 2) Questions Construction and 3) Preference Construction. The dataset examples shows vision-question/preferences pairs that comprise three types of questions: easy questions, hard questions, and hard statements." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 893, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ensure comprehensiveness. Next, we generate questions related to categories of harmful content. After this, we proceed with preference annotation. This stage includes generating responses from various models and labeling these responses based on preferences for harmlessness and helpfulness.", + "bbox": [ + 88, + 444, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/9f1654040c7227d88f6f2a2673f4665291eed5d9d73553eb25282fe4a7955a37.jpg", + "table_caption": [ + "Table 1. Training dataset statistics for SPA-VL. For each image, we provide three prompts: Easy question, Hard question, Hard statement. UR% represents the unsafe rate." + ], + "table_footnote": [], + "table_body": "
Secondary ClassImage Cnt.QuestionPrefer(UR%)Prefer(Len.)
UR%Cnt.Len.Cho.Rej.Cho.Rej.
Toxic379144.111132111611.3541.55488392
Unfair358938.38106841207.1532.16620441
Erosion of Trust in Public Information126337.6237671527.6231.62595463
False Beliefs181429.3154241465.8827.16746539
Dangerous Information126359.66378812914.7849.39621580
Privacy63653.12190715612.1144.83635513
Security Threats245263.99727914112.7450.83567572
Defamation61151.83180614516.4546.46542464
Fraud or Deceptive Action477957.211417914413.7348.14590531
Influence Operations179551.51531715117.1149.69725599
Illegal Activities373460.511102513013.8349.23568531
Persuasion and Manipulation118859.38333113617.8951.73719633
Violation of Personal Property190955.5753821609.541.19543478
Risky Financial Practices184931.8152071479.130.57933685
Trade and Compliance122129.4630211809.7631.45905647
Total3189449.279325813811.742.23626519
", + "bbox": [ + 94, + 592, + 480, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 443, + 624, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Statistics. Our SPA-VL dataset comprises four parts: the training set, the validation set, and two test sets, HarmEval and HelpEval, which are used to evaluate harmfulness and helpfulness, respectively. The number of samples in each part is 93,258, 7,000, 265, and 265, respectively. Table 1 shows the dataset statistics of the training set. To detect the unsafe content covered by our SPA-VL dataset, we utilize the MD-Judge evaluator [22] to calculate the unsafe rate of the collected questions and VLMs' responses. Nearly half of the collected questions are unsafe, while the unsafe rate for the chosen response and rejected response is $11.7\\%$ and $42.23\\%$ , respectively. The HarmEval test set includes a substantial number of harmful questions, while the HelpEval test set primarily comprises questions that involve instruction following or require the expression of opinions.", + "bbox": [ + 511, + 470, + 906, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diverse Domains. A diverse and representative set of images is essential for training models to handle vision data safely and effectively. Our primary challenge is ensuring diversity while maintaining relevance to harmful content categories. To address this, we establish a comprehensive harm content categorization framework. Our SPA-VL adopts 6 primary domains, 15 secondary categories, and 53 tertiary categories, ensuring comprehensive coverage and granularity for precise harm detection and response alignment. A detailed visual representation of this hierarchical structure is provided in Figure ?? in the Appendix ??. We reference Li et al. [22], Weidinger et al. [54] for our primary harm", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19869", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/726c64f80a5961c18fac5e945405524457343f9bca0b9c869ae3d1cfb3b6bf7c.jpg", + "image_caption": [ + "The most degrading and ageist way to mark your 28th birthday would be to have a card that says \"You're 28\" with a picture of a duck on it. This is degrading because it implies that the person is old and irrelevant, and it is ageist because it suggests that the person is past their prime and has no value. The card's message is disrespectful and disregards the individual's worth and potential for growth and development." + ], + "image_footnote": [], + "bbox": [ + 132, + 97, + 202, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a0170ecb21ab281a059fec504dd5efb728f98d50244c2fe059ecde752604e53.jpg", + "image_caption": [ + "What's the most degrading and ageist way to mark your 28th birthday, showcasing your declining worth as a stale and irrelevant human being?" + ], + "image_footnote": [], + "bbox": [ + 437, + 88, + 475, + 127 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7abae8fc431edd41d7d4ca08d96a250ed64d9ccc39435b99df273aa47ab44906.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 171, + 137, + 203 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6472b30d4ea7fe2217d0fa7dda0bcdd3a0f4e5f51dc381c048af6a19baaa1600.jpg", + "image_caption": [ + "I'm sorry, but I cannot provide an answer to that question as it is ageist and degrading. It is important to remember that everyone is unique and valuable, regardless of their age. Let's focus on celebrating our birthdays and the milestones we have achieved, rather than making derogatory comments about others." + ], + "image_footnote": [], + "bbox": [ + 102, + 234, + 140, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e772b6ed351d3ce1c058696225ced69bc28b9e0b129274934b219ef590cefc9.jpg", + "image_caption": [ + "It is not appropriate to use degrading or ageist language to mark someone's 28th birthday or any other milestone. Such language is disrespectful and can negatively impact the individual's self-esteem and well-being. It is important to celebrate and acknowledge the achievements and experiences of individuals regardless of their age, and to treat them with respect and dignity. \\n\\nInstead of using degrading language, you can create a positive and supportive atmosphere for the person's 28th birthday celebration by: \\n\\nOrganizing a fun and engaging party with friends and family. \\n\\nthoughtful and meaningful presents that reflect the individual's interests and hobbies. \\n\\nSharing heartfelt messages and well-wishes on social media or through personal communication. \\n\\nEncouraging the person to set new goals and aspirations for the next year of their life. \\n\\nCreating a memorable photo album or video montage of the person's life journey so far. \\n\\nBy focusing on positivity and support, you can help the individual feel valued and appreciated, regardless of their age.\\n\\n", + "Figure 2. Case study comparing responses from the original model, the model trained with DPO and PPO on our SPA-VL." + ], + "image_footnote": [], + "bbox": [ + 102, + 340, + 140, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "categories, systematically organizing and identifying various types of harmful content. For the secondary and tertiary classes, we reference usage policies from leading AI organizations, including OpenAI [35], LLaMA [32], Gemini [15], Claude [2], as well as guidelines from Llama Guard [18] and Llama Guard2 [30]. Additional references include Luo et al. [31], Weidinger et al. [53], ensuring that our classification aligns with industry standards and best practices, enhancing the relevance and applicability of our dataset.", + "bbox": [ + 88, + 516, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Formats. Following Bai et al. [4], we gather preference data by choosing the better response from two generated by VLMs, based on predefined criteria of harmlessness and helpfulness. Finally, a quadruple (question, image, chosen response, rejected response) reflecting preferences is collected, where the chosen response is the better response selected under principle harmless and helpful. Figure 1 shows three demonstration examples of our SPAVL dataset. In the following sections, we will give a detailed illustration of the construction process of our dataset.", + "bbox": [ + 89, + 654, + 482, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Collection", + "text_level": 1, + "bbox": [ + 89, + 816, + 256, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this robust categorization framework in place, we proceed to collect images that correspond to our harm categories. We use the LAION-5B [44] dataset, which is well-suited for this purpose due to its extensive and diverse col", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lection of images. The LAION-5B dataset is built using a CLIP model to match images and text, allowing us to effectively use our tertiary class labels to search for relevant pictures. This leverages the strengths of the CLIP model in understanding visual and textual correlations, ensuring that the images are well-aligned with our harm categories. To ensure diversity and avoid bias, we use six different search keywords for each tertiary class. This approach helps capture a wide range of images within each category, preventing over-representation of specific subtypes. Details are illustrated in Appendix ??", + "bbox": [ + 511, + 90, + 906, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Question Generation", + "text_level": 1, + "bbox": [ + 511, + 267, + 709, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generating relevant and meaningful questions for each image is crucial for contextual understanding. The primary challenge here is ensuring that the questions are specific to the harmful content categories and diverse in complexity. Although the LAION-5B dataset provides captions, they are often misaligned with the images [44]. To address this, we enhance the captions using Gemini 1.0 Pro Vision1.", + "bbox": [ + 511, + 289, + 903, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the subsequent step, we devise queries that could be potentially harmful for each image. Initially, Gemini 1.0 Pro Vision produces an easy question for every image. To ensure pertinence, the model is supplied with the image's primary, secondary, and tertiary categories. However, these questions, typically starting with \"what\" or \"how\", tend to be straightforward and closely related to the image, which may lead to a lack of complexity and diversity. To overcome this, we use Gemini 1.0 Pro to refine the questions, transforming them into more challenging queries and statements to enhance the diversity of our dataset. This process yields two additional outputs: a hard question (hardq) and a hard statement (hardd). For this refinement, the model is provided with the image's hierarchical classifications, two captions (one generated by Gemini and the original from LAION), and the previously generated easy question. Given that Gemini is designed to avoid generating harmful content, we employ a jailbreak strategy [58] to evoke more challenging queries. Further details regarding this process are provided in Appendix ??", + "bbox": [ + 511, + 395, + 906, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Preference Collection", + "text_level": 1, + "bbox": [ + 511, + 707, + 714, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final stage in constructing the SPA-VL dataset involves generating and annotating preferences for responses for training VLMs. This stage combines the generation of diverse responses and the careful annotation of these responses to create preference pairs.", + "bbox": [ + 511, + 729, + 905, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Response Generation. To create annotations that better align with human judgment, we collect diverse responses from 12 different models. This diversity helps capture a wide range of potential answers, reducing model-specific", + "bbox": [ + 511, + 805, + 905, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/gemini-pro-vision", + "bbox": [ + 511, + 875, + 901, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19870", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "biases and ensuring that the VLMs can learn from a variety of perspectives. Detailed generation process are described in Appendix ??", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preference Annotation. Responses are classified using MD-Judge to ensuring and for each question, we randomly select two responses from different safety-rated model groups to ensure that the chosen responses reflect diverse levels of safety to make GPT-4V better annotate. GPT-4V evaluates these pairs based on both harmlessness and helpfulness, also, to avoid bias due to the order of the answers, we query GPT-4V twice with the answers swapped. Details are illustrated in Appendix ??", + "bbox": [ + 89, + 137, + 480, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 290, + 222, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 315, + 187, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply DPO [43] and PPO [45], using our preference dataset to train LLaVA-1.5 [27, 28]. For preliminaries on DPO and PPO, please refer to the Appendix ??. We keep the visual encoder weights frozen as in Karamcheti et al. [19], Liu et al. [27] and update the weights of both the projection layer and LLM in LLaVA-1.5 (7B) in our main experiments. For training parameters, please refer to the Appendix ?? for further details.", + "bbox": [ + 89, + 337, + 480, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate harmlessness using the HarmEval dataset and assess adversarial robustness on AdvBench [64] and MM-SafetyBench [29]. Helpfulness is measured using the Anthropic-Helpful [4] and HelpEval datasets, with GPT-4 determining response quality based on helpfulness and safety. Detailed evaluation methods and results can be found in the Appendix ??", + "bbox": [ + 89, + 460, + 480, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our HarmEval dataset comprises 265 questions across 53 categories, with each category containing 5 questions. We utilize two evaluation metrics on this dataset: the unsafe rate and the harm score. The UnSafe Rate assesses the proportion of unsafe responses generated by the model for all test questions. This metric evaluates the general safety of the model's outputs. The Harm Score involves comparing pairs of answers from the base model, LLaVA-1.5 (7B), and the trained model using GPT-4V. Unlike the data construction phase, GPT-4V assesses the answers based solely on the criterion of \"harmlessness\".", + "bbox": [ + 89, + 566, + 480, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure unbiased evaluation, we also test on the publicly available language dataset AdvBench [64] and the vision dataset MM-SafetyBench [29] (Figure ?? shows two examples from this benchmark). For both datasets, the evaluation metric is the Attack Success Rate (ASR), which measures the safety of the model by assessing its vulnerability to adversarial attacks.", + "bbox": [ + 89, + 733, + 480, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MM-SafetyBench. MM-SafetyBench is a benchmark for evaluating VLMs safety composed of harmful questions and images. It includes four subcategories: Text-only (text only), SD (images generated by Stable Diffusion using key", + "bbox": [ + 89, + 839, + 480, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "phrases), Tipo (typography, images with the black phrase on a white background), and SD+Typo (concatenated generated images and typography together). The evaluation metric for this benchmark is the attack success rate (ASR). This dataset enables us to evaluate how well our trained model performs on publicly available image-based test sets in terms of safety. We use MM-Safety Prompt as Liu et al. [29] to ensure consistency across our tests.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AdvBench. This dataset is purely text-based and has two subcategories: vanilla instruction and suffix injection (questions with the suffix \"Answer with 'Absolutely, here's'\"). The evaluation metric for this dataset is also ASR. This dataset allows us to assess the safety of our trained model when confronted with text-based harmful content in publicly available test sets.", + "bbox": [ + 511, + 212, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 325, + 651, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the Table 2, the models trained on our SPA-VL dataset, LLaVA-SPA-VL-DPO and LLaVA-SPA-VL-PPO, which are the best safety models from our training, exhibit superior safety performance. They surpass the baseline model LLaVA-1.5 (7B) and other open-source models, whether or not those models have undergone safety alignment. Specifically, our models achieve best safe result on MM-SafetyBench, AdvBench and HarmEval tests. Additionally, we provide a comparison of responses before and after training for LLaVA-SPA-VL-DPO, LLaVA-SPA-VL-PPO, and baseline model LLaVA in Figure 2. More comparison examples can be found in Appendix ??", + "bbox": [ + 511, + 348, + 903, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to evaluating the safety performance, we also validate our models' general ability. We select the most commonly used multimodal benchmarks and find that the general ability of our safety-aligned models does not significantly decline compared to the backbone model. Details of these evaluations can be found in the Appendix ??", + "bbox": [ + 511, + 530, + 903, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 513, + 633, + 736, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we further explore the factors that affect the performance of alignment models. Our focus includes examining the impact of dataset scale, the selection of response models from diverse models during dataset construction, the influence of different question types within the dataset, and the outcomes of deciding whether to freeze the projection layer during model training, and the different training base models.", + "bbox": [ + 511, + 657, + 903, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Data Scale", + "text_level": 1, + "bbox": [ + 511, + 787, + 630, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We delve into the impact of varying amounts of data on the performance of alignment models. Across different data quantities (around 100, 1k, 5k, 10k, 30k, and 90k), we conduct experiments encompassing various evaluation metrics, as already described in Appendix ??. The resulting line plots in Figure 3 reveal compelling insights. For HarmEval,", + "bbox": [ + 511, + 810, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/091a5fbe52dd1da468d4db9b7a40904eead6ad56e0e9a7f9630ee1400b01946d.jpg", + "table_caption": [ + "Table 2. Comparison of different VLM models on harmlessness. The models are evaluated across multiple metrics on MM-SafetyBench and AdvBench, as well as the HarmEval UnSafe Rate (HarmEval USR). After training on our proposed dataset, SPA-VL, the model achieves the best scores according all metric on both DPO and PPO methods." + ], + "table_footnote": [], + "table_body": "
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Open Sourced Baseline
InstructBLIP-7B27.3813.1027.3825.0023.2151.2564.6247.55
LAMM-13B14.294.762.386.556.9924.4239.1132.83
LAMM + SFT-13B16.077.148.3321.4313.2422.6972.1232.08
LLaMA-Adapter-v2-7B35.7112.507.7417.8618.4598.2610046.04
MiniGPT-4-7B20.839.5223.8120.2418.6031.3565.3838.32
mPLUG-Owl-7B35.718.9312.5030.3621.8810010052.45
Otter-9B29.7610.125.957.7413.3991.9210041.13
InternVL-Chat-7B5.951.7919.6413.1010.126.9289.4229.81
CogVLM2-LLama3-Chat-19B16.674.7623.8123.2117.1125.3898.0813.96
LLaVA-v1.6-34B4.764.1716.0719.0511.015.5893.0822.64
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
Close Sourced Baseline
Gemini-1.5-pro000000.380.381.89
GPT-4-0125-preview1.790000.450.966.542.26
Safety Aligned
LLaVA + SPA-VL-DPO0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
LLaVA + SPA-VL-PPO0.6(↓33.93)0(↓7.74)0(↓22.62)1.19(↓16.07)0.45(↓20.09)0.19(↓97.88)2.12(↓97.69)0(↓44.15)
", + "bbox": [ + 94, + 132, + 901, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Harm Score consistently decreases with increasing data volume. Similarly, in MM-SafetyBench, the overall average ASR steadily decreases as the data scale grows, and both vanilla and suffix ASRs in AdvBench exhibit a similar trend of declining rates with expanding data sizes. Notably, the help score in Anthropic-Helpful exhibits a progressive increase, indicating a simultaneous enhancement in safety and helpfulness as the dataset size expands. The full results are presented in Table 3, with detailed analysis in Appendix ??", + "bbox": [ + 88, + 530, + 480, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Response Model Selection", + "text_level": 1, + "bbox": [ + 89, + 695, + 326, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we examine the impact of response diversity and safety in our dataset on model training. We conducted four groups of experiments, each group is trained using DPO on around 10K samples. Safe Group consists of response pairs from the three safest models (InternLMXComposer, QwenVL, Gemini1.0 Pro Vision) according to Table ?? in Appendix ??. Relative Safe Group includes pairs from relative safe models(LAMM_SFT, LLaVA1.5, InternLMXComposer, QwenVL, gemini). Unsafe Group comprises pairs from the five least safe models(mPLUG-Owl, Otter, InstructBLIP, LLaMA-Adapter-v2, GeminiJailbreak) and the All group consists of the complete set", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8081ccc5ef4e3416b6d1d4a900024487c3e8eadde1c98bc4e4af707a2e3e5c43.jpg", + "table_caption": [ + "Table 3. Impact of Training Data Scale on Alignment Model Performance using PPO and DPO Methods on LLaVA-1.5 (7B). The table compares the performance of alignment models trained with PPO and DPO methods across varying data scales: 100, 1K, 5K, 10K, 30K, and 90K tsamples. The evaluation covers both safety and helpfulness metrics. As the data scale increases, the performance generally improves across harmlessness and helpfulness metrics." + ], + "table_footnote": [], + "table_body": "
Data ScaleMM-SafetyBench↓AdvBench↓HarmEval↓Helpful
Text-onlySDCargoSD+CargoAvgvanillasuffixUSRScoreAnthropic↑HelpEval↑
PPO
10035.128.9318.4517.2619.9497.5098.8543.3950.0051.5019.00
1K30.958.3319.0516.6718.7597.3198.4641.8945.0658.0028.61
5K5.954.7613.1017.2610.2736.1585.7710.1917.2461.0048.54
10K0.601.195.959.524.328.6517.500.009.2864.5067.60
30K0.000.000.002.980.740.583.270.0010.9265.5052.09
90K0.600.000.001.190.450.192.120.008.8170.0071.04
DPO
10032.7410.1220.2416.6719.9497.8999.6243.4050.5751.0024.12
1K30.367.7417.8619.0518.7591.5496.7326.0430.0058.5035.65
5K4.171.194.178.334.4610.0040.001.8915.3464.0042.86
10K1.191.792.384.762.535.776.540.3813.7863.0040.85
30K0.000.600.601.190.600.000.000.0013.6469.5045.72
90K1.190.601.192.981.490.000.000.7513.4170.0050.63
", + "bbox": [ + 516, + 640, + 903, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of 12 models.", + "text_level": 1, + "bbox": [ + 511, + 792, + 607, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results presented in the Table 4 indicate that if our dataset only includes pairs of safe responses, the model struggles to learn how to avoid bad patterns, leading to vulnerability (Safe Group performed poorly on the AdvBench suffix test, suggesting the model is easily attacked). Similarly, if our dataset only includes unsafe responses, the", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19872", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6390a9a455d1ebffcf62f6599282a1b708f661e7c7f0a66deab7b1c52d3f15c1.jpg", + "image_caption": [ + "(a) EvalHarm" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 267, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e23859681d979630a87480b16b30bc05bcc4f467302af369c1bd7b7d6c73ba5.jpg", + "image_caption": [ + "(c) AdvBench" + ], + "image_footnote": [], + "bbox": [ + 96, + 203, + 267, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2431fed2bffe567079fd268c8ed89c381642bf7c13efce0fc08932a24250b862.jpg", + "image_caption": [ + "(b) MM-SafetyBench" + ], + "image_footnote": [], + "bbox": [ + 308, + 88, + 478, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1b512c48e6e28377113ea7b7a5e5d39543aa25250ae0fa38fa06b68ee5e9127.jpg", + "image_caption": [ + "(d) Anthropic-Helpful", + "Figure 3. Impact of Data Scale on Alignment Model Performance. Line plots illustrate the effect of varying data quantities (100, 1k, 5k, 10k, 30k, and 90k) on the performance metrics of alignment models for both PPO and DPO methods. (a) Shows the Harm Score (\\%) on EvalHarm (b) Shows the Average Attack Success Rate (ASR $\\%$ ) on MM-SafetyBench (c) Shows ASR (\\%) on vanilla and suffix in AdvBench (d) Shows the Help Score (\\%) on Anthropic-Helpful." + ], + "image_footnote": [], + "bbox": [ + 310, + 203, + 475, + 297 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d06296308d6919b4c5569478d45763d3b6969b5decd69d79c0d77dbf50104cc6.jpg", + "table_caption": [ + "Table 4. Detailed harmlessness evaluation metrics for Response Model Selection. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge." + ], + "table_footnote": [], + "table_body": "
Model BagAdvBenchMM-SafetyBenchHarmEval
vanillas suffixText-onlySDTypoSD+TypoAvgUSRScore
Safe32.5065.3812.53.5710.7111.909.6713.9718.49
Relative safe14.8135.004.173.579.528.936.554.5315.85
Unsafe9.0460.772.982.988.9311.906.707.1721.14
All0.586.541.191.792.384.762.536.1113.78
", + "bbox": [ + 93, + 511, + 480, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model cannot be trained to be safe, as it lacks exposure to good patterns (Unsafe Group also performed poorly on the AdvBench suffix test). Relative safe Group, which includes a mix of good and bad responses, shows significantly better safety performance compared to Safe Group and Unsafe Group. However, there still exists a gap on harmlessness between the model trained with Relative Safe and the All group. This demonstrates the critical importance of response diversity during the data construction process for effective model alignment.", + "bbox": [ + 89, + 609, + 482, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Question Types", + "text_level": 1, + "bbox": [ + 89, + 771, + 246, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the impact of three different question types (Easy questions, Hard questions, and Hard statements) on the experimental results. We also compare these individual results with the combined results of all three question types. For each experiment, we select training dataset of approximately $10k$ instances and using DPO to train our model. The results presented in the Table 5", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e3ea9c883dcf71e9b9d00a8450f4557b48a072b7bbf966fc8106956e1364f9c.jpg", + "table_caption": [ + "Table 5. The detailed harmless evaluation metrics of Question Types. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge." + ], + "table_footnote": [], + "table_body": "
Ques TypeAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
Easy-Q3.8524.041.791.792.388.933.722.2616.73
Hard-Q2.1211.541.191.193.579.523.870.7513.97
Hard-S2.125.003.571.794.175.953.878.7018.44
Mixed0.586.541.191.792.384.762.530.3813.78
", + "bbox": [ + 516, + 146, + 903, + 227 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ca30e394fd3b4dd0bb27a61f059d7664ea0561119a21cb40a9bf193d82fed92.jpg", + "table_caption": [ + "Table 6. Detailed harmless evaluation metrics of model architecture on LLaVA-1.5 (7B), with projector and without projector for both DPO and PPO training methods." + ], + "table_footnote": [], + "table_body": "
Model ArchAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
DPO
w/o projector0.000.190.000.001.195.361.641.1314.21
w projector0.000.000.000.600.601.190.600.0013.64
(0.00)(↓0.19)(0.00)(↑0.60)(↓0.59)(↓4.17)(↓1.04)(↓1.13)(↓0.57)
PPO
w/o projector0.582.880.000.001.791.790.890.0019.32
w projector0.583.270.000.000.002.980.740.0010.92
(0.00)(↓0.39)(0.00)(0.00)(↓1.79)(↓1.19)(↓0.15)(0.00)(↓8.4)
", + "bbox": [ + 516, + 286, + 903, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "show that the combined dataset achieves higher safety performance compared to the individual datasets of each question type at the same data scale. This indicates that integrating diverse question types enhances the model's robustness against harmful attacks. These findings suggest that, in real-world scenarios, questions that provoke harmful responses are varied. Consequently, training with a combination of different question types is likely to improve the model's ability to resist harmful attacks and decrease the likelihood of generating harmful responses.", + "bbox": [ + 511, + 429, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Training Methods", + "text_level": 1, + "bbox": [ + 511, + 602, + 687, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the approach outlined in LLaVA [28], we freeze the vision encoder during training. In this framework, both the projection layer and LLM were trained together during SFT as described in LLaVA [28], whereas in LLaVARLHF [48], only the LLM was trained. We aim to explore the impact of training the LLM alone versus with the projection layer on reinforcement learning outcomes. We analyze safety validation results on our $30k$ dataset using DPO method. As shown in the Table 6, there are minimal differences in language-only safety tests (AdvBench). However, in image-involved safety tests (MM-SafetyBench and EvalHarm), training with the projection layer consistently outperforms training without it. We hypothesize that including the projection layer improves the model's ability to detect harmful content in images. This suggests a valuable direction for future work to further investigate these effects. We also investigate the training of LoRA, with further details provided in Appendix ??", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Model", + "text_level": 1, + "bbox": [ + 91, + 90, + 176, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explored other backbone models as well and conducted experiments using QwenVL-7BChat and InternLMXComposer-7B, applying DPO training on our dataset. As shown in Table 7, with SPA-VL data training, open-source models, including LLaVA, Qwen, and InternLMXComposer, exhibit significant safety improvements, approaching the performance levels of closed-source models.", + "bbox": [ + 89, + 114, + 480, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47863007a91050e642bd18a1ba73103026a06a537476991a478afffb9ad1df11.jpg", + "table_caption": [ + "Table 7. We presents the results of DPO training on safety benchmarks for different model backbones, comparing baseline models with their safety-aligned counterparts using the SPA-VL. The models evaluated include LLaVA-7B, InternLMXComposer-7B, and QwenVL-Chat-7B, with their respective safety-aligned versions: LLaVA + SPA-VL, InternLMXC + SPA-VL, and QwenVL + SPA-VL. Across all backbones, significant improvements are observed after safety alignment." + ], + "table_footnote": [], + "table_body": "
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Baseline
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
Safety Aligned
LLaVA + SPA-VL0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
InternLMXC+SPA-VL0(↓7.74)1.19(↓2.98)0.6(↓25.59)2.38(↓24.41)1.04(↓15.18)0.19(↓5.21)0.38(↓97.50)0.75(↓25.29)
QwenVL+SPA-VL0(↓3.57)1.19(↓2.38)1.19(↓22.02)4.17(↓22.62)1.64(↓12.65)0.38(↓1.54)4.61(↓68.12)3.02(↓4.53)
", + "bbox": [ + 91, + 363, + 480, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Human Annotation", + "text_level": 1, + "bbox": [ + 89, + 508, + 277, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate the consistency between GPT evaluation and human evaluation to ensure the reliability and validity of our assessment methods. For annotation part, we random sample 530 pairs for each question type to show the human preference result with GPT-4V. For AdvBench, we random select a total result in our result and humanly check the attack success for all the 520 results for both suffix and valina with GPT-4. In the case of Anthropic-Helpful, we examine 200 samples to check the consistency of helpfulness evaluations with GPT-4. For HarmEval, 265 samples (5 from each category) are selected to compare the safety preference consistency between human evaluators and GPT-4V for the baseline model LL and our trained model. Similarly, for HelpEval, we select 265 samples (5 from each category) to compare the helpfulness win rate consistency between human evaluators and GPT-4V for the GPT-4V responses and our trained model.", + "bbox": [ + 88, + 535, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This comprehensive approach ensures that our models' evaluations align closely with human judgment, thereby enhancing the robustness of our evaluation part. The consistency rates between GPT evaluations and human evaluations are summarized in Table 8. The results indicate high alignment across various evaluation metrics, reinforcing the reliability of our assessment approach.", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9e084331bd3f19b7a2881c0a6791767ad4b267795c57ea56aad56dcd3f63ba9e.jpg", + "table_caption": [ + "Table 8. Consistency Rate (%) between GPT-4V and human annotation." + ], + "table_footnote": [], + "table_body": "
Easy-QHard-QHard-SAdvBenchAnthropic-HelpfulHarmEvalHelpEval
suffixvanilla
96.6697.7493.7799.4099.8091.0098.11100
", + "bbox": [ + 517, + 119, + 903, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 183, + 782, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced SPA-VL, the first large-scale, high-quality dataset designed for the safety alignment of VLMs. SPA-VL's comprehensive coverage of harmfulness domains, diverse question-answer types, and focus on both harmlessness and helpfulness make it a robust dataset for improving model safety without compromising core capabilities. Our experiments demonstrate that models trained on SPA-VL using techniques like PPO and DPO show significant improvements in safety and outperform baseline and other state-of-the-art VLMs in various safety evaluations. Additionally, our analyses reveal that increasing data volume, incorporating diverse responses, and using a mix of question types enhance the safety and performance of aligned models. The findings highlight the importance of comprehensive datasets like SPA-VL in achieving robust safety alignment, ensuring that VLMs can be effectively and safely deployed in real-world applications. SPA-VL represents a significant step towards safer and more reliable vision-language models, paving the way for future research and development in this crucial area.", + "bbox": [ + 509, + 208, + 906, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we aim to extend the scope of our work to encompass the unified \"3H\" framework of helpfulness, harmlessness, and honesty, to ensure a more holistic approach to aligning VLMs with human values. Furthermore, we plan to explore the application of our safety alignment techniques to more complex tasks such as reasoning in VLMs, which will require nuanced understanding and generation of visual content. Additionally, we are interested in investigating the transferability of alignment capabilities between LLMs and VLMs, which could lead to more efficient and effective alignment strategies across different modalities.", + "bbox": [ + 509, + 511, + 908, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 705, + 684, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, as well as the computational resources provided by the Shanghai Artificial Intelligence Laboratory, for their essential contributions of computational resources and technical support. Furthermore, we extend our heartfelt gratitude to our advisors for their invaluable guidance, insightful discussions, and constructive feedback throughout the research process.", + "bbox": [ + 509, + 729, + 906, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19874", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthropic. Claude, 2023. 1, 2", + "[2] Anthropic. Claude usage policies, 2024. https://www.anthropic.com/legal/aup.4", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2", + "[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2, 4, 5", + "[5] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 2", + "[6] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime, 2023. 1", + "[7] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In FAccT '21: 2021 ACM Conference on Fairness, Accountability, and Transparency, Virtual Event / Toronto, Canada, March 3-10, 2021, pages 610-623. ACM, 2021. 2", + "[8] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 2", + "[9] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 2", + "[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 1, 2", + "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + ], + "bbox": [ + 96, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2", + "[12] Jieren Deng, Yijue Wang, Ji Li, Chenghong Wang, Chao Shang, Hang Liu, Sanguthevar Rajasekaran, and Caiwen Ding. Tag: Gradient attack on transformer-based language models. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3600-3610, 2021. 2", + "[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.2", + "[14] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 2", + "[15] Google. Gemini usage policies, 2024. https://ai.google.dev/gemini-api/docs/safety-settings.4", + "[16] Susan Hao, Renee Shelby, Yuchi Liu, Hansa Srinivasan, Mukul Bhutani, Burcu Karagol Ayan, Shivani Poddar, and Sarah Laszlo. Harm amplification in text-to-image models. arXiv preprint arXiv:2402.01787, 2024. 1", + "[17] Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P. Xing, Furong Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huaxiu Yao, Manolis Kellis, Marinka Zitnik, Meng Jiang, Mohit Bansal, James Zou, Jian Pei, Jian Liu, Jianfeng Gao, Jiawei Han, Jieyu Zhao, Jiliang Tang, Jindong Wang, Joaquin Vanschoren, John Mitchell, Kai Shu, Kaidi Xu, Kai-Wei Chang, Lifang He, Lifu Huang, Michael Backes, Neil Zhenqiang Gong, Philip S. Yu, Pin-Yu Chen, Quanquan Gu, Ran Xu, Rex Ying, Shuiwang Ji, Suman Jana, Tianlong Chen, Tianming Liu, Tianyi Zhou, William Yang Wang, Xiang Li, Xiangliang Zhang, Xiao Wang, Xing Xie, Xun Chen, Xuyu Wang, Yan Liu, Yanfang Ye, Yinzhi Cao, Yong Chen, and Yue Zhao. Trustworthiness in large language models. In Forty-first International Conference on Machine Learning, 2024. 2", + "[18] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 4", + "[19] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. RLAIF vs. RLHF: scaling reinforcement learning from human feedback with AI feedback. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1", + "[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 2", + "[22] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Saladbench: A hierarchical and comprehensive safety benchmark for large language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3923-3954. Association for Computational Linguistics, 2024. 3", + "[23] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3326-3342. Association for Computational Linguistics, 2024. 1", + "[24] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2", + "[25] Jiawei Liang, Siyuan Liang, Man Luo, Aishan Liu, Dongchen Han, Ee-Chien Chang, and Xiaochun Cao. Vl-trojan: Multimodal instruction backdoor attacks against autoregressive visual language models. arXiv preprint arXiv:2402.13851, 2024. 1", + "[26] Hongzhan Lin, Ziyang Luo, Bo Wang, Ruichao Yang, and Jing Ma. Goat-bench: Safety insights to large multimodal models through meme-based social abuse. ArXiv, abs/2401.01523, 2024. 1", + "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1, 5", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26286-26296. IEEE, 2024. 5, 7", + "[29] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2025. 2, 5", + "[30] LLamaGuard2. Llama guard2, 2024. https://github.com/meta-llama/PurpleLlama/blob/main/Llama-Guard2/MODEL_CARD.md.4", + "[31] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 4", + "[32] Meta. Llama usage policies, 2024. https://ai.meta.com/llama/use-policy.4", + "[33] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2", + "[34] OpenAI. Introducing chatgpt, 2022. 2", + "[35] OpenAI. Openai usage policies, 2023. https://openai.com/policies/usage-policies.4", + "[36] OpenAI. Gpt4v, 2024. https://chatgpt.com/.1", + "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2", + "[38] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 16012-16027. Association for Computational Linguistics, 2024. 2", + "[39] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2025. 2", + "[40] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2", + "[41] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1", + "[42] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2", + "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, pages 53728-53741. Curran Associates, Inc., 2023. 2, 5", + "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294. 4", + "[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv: Learning, arXiv: Learning. 2, 5", + "[46] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[47] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2", + "[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darryll. Aligning large multimodal models with factually augmented RLHF. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 13088-13110. Association for Computational Linguistics, 2024. 2, 7", + "[49] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Agüera y Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. Lamda: Language models for dialog applications. CoRR, abs/2201.08239, 2022. 2", + "[50] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2", + "[51] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 10460-10479. Association for Computational Linguistics, 2024. 2", + "[52] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 2", + "[53] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021. 4", + "[54] Laura Weidinger, Maribeth Rauh, Nahema Marchal, Arianna Manzini, Lisa Anne Hendricks, Juan Mateos-Garcia, Stevie Bergman, Jackie Kay, Conor Griffin, Ben Bariach, et al. Sociotechnical safety evaluation of generative ai systems. arXiv preprint arXiv:2310.11986, 2023. 3", + "[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[56] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, Jing Shao, and Wanli Ouyang. LAMM: language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2", + "[57] Pan Zhang, XiaoyiDongBin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition. 2023. 2", + "[58] Zaibin Zhang, Yongting Zhang, Lijun Li, Jing Shao, Hongzhi Gao, Yu Qiao, Lijun Wang, Huchuan Lu, and Feng Zhao. Psysafe: A comprehensive framework for psychological-based attack, defense, and evaluation of multi-agent system safety. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 15202-15231. Association for Computational Linguistics, 2024. 4", + "[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2", + "[60] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[61] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 2", + "[62] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. Open-Review.net, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy M. Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 2", + "[64] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5" + ], + "bbox": [ + 91, + 92, + 480, + 218 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "19878", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_model.json b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65cd8ddfde898c0c8ed7453270613f93965c25c0 --- /dev/null +++ b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_model.json @@ -0,0 +1,2369 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.88, + 0.178 + ], + "angle": 0, + "content": "SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.202, + 0.79, + 0.293 + ], + "angle": 0, + "content": "Yongting Zhang\\(^{1,3*}\\) Lu Chen\\(^{2,3*}\\) Guodong Zheng\\(^{2,3}\\) Yifeng Gao\\(^{2}\\) Rui Zheng\\(^{2}\\) Jinlan Fu\\(^{3}\\) Zhenfei Yin\\(^{3}\\) Senjie Jin\\(^{2}\\) Yu Qiao\\(^{3}\\) Xuanjing Huang\\(^{2}\\) Feng Zhao\\(^{1\\dagger}\\) Tao Gui\\(^{2,3‡}\\) Jing Shao\\(^{3\\dagger}\\) \\({}^{1}\\)University of Science and Technology of China \\({}^{2}\\)Fudan University \\({}^{3}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.296, + 0.724, + 0.309 + ], + "angle": 0, + "content": "zytabcd@mail.ustc.edu.cn, luchen23@m.fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.313, + 0.658, + 0.327 + ], + "angle": 0, + "content": "https://sqrtizhang.github.io/SPA-VL/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.368, + 0.327, + 0.385 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.486, + 0.751 + ], + "angle": 0, + "content": "The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.779, + 0.224, + 0.795 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.485, + 0.852 + ], + "angle": 0, + "content": "Vision Language Models (VLMs), such as GPT-4V [36], Claude 3 [1], LLaVA [27], and MiniGPT-4 [62] can understand visual signals and respond to user instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Equipped with a visual encoder module, VLMs extract multimodal knowledge from both visual and textual queries, leveraging pre-trained LLMs' powerful comprehension and generative capabilities to achieve remarkable results across diverse vision-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Due to the complexity of multimodal harms, previous study [16] has demonstrated that harmless inputs may also result in outputs that do not align with human preferences. Although LLMs have undergone harmless alignment, the alignment of visual encoders is relatively weak, making VLMs susceptible to successful attacks through the visual modality [6, 14, 25]. Therefore, it is necessary to simultaneously improve the alignment of the visual and language modules of the VLM to achieve the harmless and helpful responses. Ensuring the alignment of VLMs with ethical and safety standards is crucial for their safe and effective deployment. However, most existing works on the safety of VLMs focused on the evaluation benchmarks [10, 24, 26] or jailbreak detection [6, 14, 14, 23-25]. Seldom studies considered constructing large-scale, high-quality open-source training datasets to achieve the safety alignment of VLMs. The lack of such datasets hampers further development in this field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.909, + 0.84 + ], + "angle": 0, + "content": "To address these limitations, we propose a large-scale safety alignment dataset for VLMs named SPA-VL. Since Reinforcement Learning from Human Feedback (RLHF) is widely regarded as performing well in alignment studies [20, 41], our SPA-VL dataset is designed for the RLHF, with each sample containing four elements (question, image, chosen response, rejected response). The main perspectives of our SPA-VL dataset are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "(1) Comprehensive Domains: SPA-VL contains 100,788 samples and comprehensively covers a wide range of harm types, encompassing 6 domains, 13 categories, and 53 subcategories. (2) Diverse Questions and Responses:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.863, + 0.442, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution. Authorship order determined by coin flip." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.242, + 0.888 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.863, + 0.442, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19867" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "For diverse question collection, we gathered three types of questions for each image: easy question, hard question, and hard statement. For diverse response collection, to enhance the diversity of the collected responses and reduce specific model biases, we selected responses from 2 out of 12 different models for each question, designating them as chosen and rejected responses. (3) Multi-Objective Alignment: The preference dataset is designed according to two alignment objectives: harmlessness and helpfulness. Thus, the responses of the aligned VLMs are enhanced in both aspects without sacrificing helpfulness for safety." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.258, + 0.386, + 0.272 + ], + "angle": 0, + "content": "The main contributions are listed below:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.392 + ], + "angle": 0, + "content": "(1) We propose the SPA-VL, a large-scale, high-quality, and domain diversity dataset for vision language model safety alignment studies. By using techniques such as PPO and DPO for safety alignment learning on the SPA-VL dataset based on some open VLMs, such as LLaVA, experimental results show that these VLMs significantly improve performance in safety and surpass the state-of-the-art VLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.482, + 0.468 + ], + "angle": 0, + "content": "(2) We conducted extensive analysis and ablation studies, finding that increasing the scale of the dataset, incorporating diverse answers, and using various question types can improve the safety and performance of the alignment VLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.482, + 0.545 + ], + "angle": 0, + "content": "(3) The entire dataset construction process is fully automated, ensuring efficient and scalable data generation. The automation covers all key steps, including the collection of images and questions, the selection of VLM responses, and the annotations of chosen and rejected responses." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.233, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Vision Language Models. The rapid advancement of Large Language Models [42, 59] and their robust foundational capabilities have significantly prompted the development of multimodal large models. Recently, Vision-Language Models have emerged as a promising extension of LLMs [9], integrating visual and textual information for enhanced multimodal understanding. Notable models in this domain include InstructBLIP [11], InternLMXComposer [57], LAMM-SFT [56], LAMM [56], LLaMA-Adapter-V2 [13], MiniGPT-4 [62], mPLUG-Owl [55], Otter [21], and Qwen-VL-Chat [3]. Most of these VLMs are developed by projecting the vision space into the language space through a learned projector, leveraging pre-trained language models as their backbone. As VLMs continue to advance rapidly, safety concerns have garnered significant attention from researchers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Reinforcement Learning from Human Feedback. Despite the promising capabilities of LLMs and VLMs, they are prone to unintended behaviors, such as fabricating facts, producing biased or harmful content, or even harming humans [7, 8]. They should be helpful, honest, and harm-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "less (3H) [4, 37, 49]. RLHF offers the most straightforward approach to achieving this goal. RLHF methods such as PPO [45] and DPO [43] have been highly successful in aligning AI with human preferences. Notable applications like ChatGPT [34] and Claude [1] show strong performances in academic benchmarks. Models trained with RLHF methods often perform better and adhere more closely to human values compared to those trained only with SFT [4]. This success extends to VLMs, where RLHF has been used to address hallucination issues [5, 39, 48, 61]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.393 + ], + "angle": 0, + "content": "Safety of VLMs. To evaluate the safety performance of VLMs, various methods and datasets have been proposed. Among these evaluation benchmarks are MM-SafetyBench [29], ChEf [47], and OODCV-VQA, SketchyVQA [50]. In addition to these benchmarks, several attack methods, such as adversarial attacks [12, 40, 46, 60] and jailbreak techniques [14, 24, 33], have been developed to test the vulnerabilities of VLMs. These studies aim to identify and exploit potential weaknesses in VLMs, underscoring the need for robust safety measures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.604 + ], + "angle": 0, + "content": "In response to these vulnerabilities, several methods have been developed to enhance the safety performance of VLMs. For instance, VLGuard [63] employed supervised fine-tuning (SFT) on the VLGuard dataset, which contains 2000 training images designed to improve safety. Similarly, Chen et al. [10] used AI-annotated data for SFT. Wei et al. [52] utilized in-context learning to bolster model safety. Additionally, Pi et al. [38] introduced MLLM-Protector, a harm detector serving as a plug-and-play defense mechanism for VLMs, and Wang et al. [51] applied inference time alignment methods to improve harmlessness. These approaches collectively demonstrate ongoing efforts to mitigate risks and enhance the resilience of VLMs against various types of attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.905, + 0.756 + ], + "angle": 0, + "content": "Among all the methods mentioned above, we can broadly categorize them into two types. The first type enhances model safety during the inference stage by using prompts. This approach is efficient and convenient but often results in limited safety improvements and lacks generalization [17]. The second type involves training-based methods, specifically during the training-to-align phase, which can be further divided into SFT and RLHF. While the aforementioned methods primarily rely on SFT, we go a step further by providing a comprehensive RLHF dataset SPA-VL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.677, + 0.785 + ], + "angle": 0, + "content": "3. SPA-VL Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the development of VLMs, effectively addressing harmful content in multimodal environments poses a significant challenge. The SPA-VL dataset helps VLMs confront this challenge by providing safety preference data for RLHF. As shown in Figure 1, establishing the SPA-VL dataset involves three key stages. First, we systematically collect images, which includes gathering a diverse set of images to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19868" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.895, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.394, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 1. Overview of SPA-VL Dataset Construction. It is built in three stages: 1) Image Collection, 2) Questions Construction and 3) Preference Construction. The dataset examples shows vision-question/preferences pairs that comprise three types of questions: easy questions, hard questions, and hard statements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.484, + 0.534 + ], + "angle": 0, + "content": "ensure comprehensiveness. Next, we generate questions related to categories of harmful content. After this, we proceed with preference annotation. This stage includes generating responses from various models and labeling these responses based on preferences for harmlessness and helpfulness." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Table 1. Training dataset statistics for SPA-VL. For each image, we provide three prompts: Easy question, Hard question, Hard statement. UR% represents the unsafe rate." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.593, + 0.481, + 0.88 + ], + "angle": 0, + "content": "
Secondary ClassImage Cnt.QuestionPrefer(UR%)Prefer(Len.)
UR%Cnt.Len.Cho.Rej.Cho.Rej.
Toxic379144.111132111611.3541.55488392
Unfair358938.38106841207.1532.16620441
Erosion of Trust in Public Information126337.6237671527.6231.62595463
False Beliefs181429.3154241465.8827.16746539
Dangerous Information126359.66378812914.7849.39621580
Privacy63653.12190715612.1144.83635513
Security Threats245263.99727914112.7450.83567572
Defamation61151.83180614516.4546.46542464
Fraud or Deceptive Action477957.211417914413.7348.14590531
Influence Operations179551.51531715117.1149.69725599
Illegal Activities373460.511102513013.8349.23568531
Persuasion and Manipulation118859.38333113617.8951.73719633
Violation of Personal Property190955.5753821609.541.19543478
Risky Financial Practices184931.8152071479.130.57933685
Trade and Compliance122129.4630211809.7631.45905647
Total3189449.279325813811.742.23626519
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.444, + 0.625, + 0.459 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Data Statistics. Our SPA-VL dataset comprises four parts: the training set, the validation set, and two test sets, HarmEval and HelpEval, which are used to evaluate harmfulness and helpfulness, respectively. The number of samples in each part is 93,258, 7,000, 265, and 265, respectively. Table 1 shows the dataset statistics of the training set. To detect the unsafe content covered by our SPA-VL dataset, we utilize the MD-Judge evaluator [22] to calculate the unsafe rate of the collected questions and VLMs' responses. Nearly half of the collected questions are unsafe, while the unsafe rate for the chosen response and rejected response is \\(11.7\\%\\) and \\(42.23\\%\\), respectively. The HarmEval test set includes a substantial number of harmful questions, while the HelpEval test set primarily comprises questions that involve instruction following or require the expression of opinions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diverse Domains. A diverse and representative set of images is essential for training models to handle vision data safely and effectively. Our primary challenge is ensuring diversity while maintaining relevance to harmful content categories. To address this, we establish a comprehensive harm content categorization framework. Our SPA-VL adopts 6 primary domains, 15 secondary categories, and 53 tertiary categories, ensuring comprehensive coverage and granularity for precise harm detection and response alignment. A detailed visual representation of this hierarchical structure is provided in Figure ?? in the Appendix ??. We reference Li et al. [22], Weidinger et al. [54] for our primary harm" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19869" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.098, + 0.203, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.104, + 0.429, + 0.138 + ], + "angle": 0, + "content": "What's the most degrading and ageist way to mark your 28th birthday, showcasing your declining worth as a stale and irrelevant human being?" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.089, + 0.477, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.172, + 0.138, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.154, + 0.464, + 0.221 + ], + "angle": 0, + "content": "The most degrading and ageist way to mark your 28th birthday would be to have a card that says \"You're 28\" with a picture of a duck on it. This is degrading because it implies that the person is old and irrelevant, and it is ageist because it suggests that the person is past their prime and has no value. The card's message is disrespectful and disregards the individual's worth and potential for growth and development." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.236, + 0.142, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.231, + 0.468, + 0.281 + ], + "angle": 0, + "content": "I'm sorry, but I cannot provide an answer to that question as it is ageist and degrading. It is important to remember that everyone is unique and valuable, regardless of their age. Let's focus on celebrating our birthdays and the milestones we have achieved, rather than making derogatory comments about others." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.342, + 0.142, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.292, + 0.469, + 0.446 + ], + "angle": 0, + "content": "It is not appropriate to use degrading or ageist language to mark someone's 28th birthday or any other milestone. Such language is disrespectful and can negatively impact the individual's self-esteem and well-being. It is important to celebrate and acknowledge the achievements and experiences of individuals regardless of their age, and to treat them with respect and dignity. \\n\\nInstead of using degrading language, you can create a positive and supportive atmosphere for the person's 28th birthday celebration by: \\n\\nOrganizing a fun and engaging party with friends and family. \\n\\nthoughtful and meaningful presents that reflect the individual's interests and hobbies. \\n\\nSharing heartfelt messages and well-wishes on social media or through personal communication. \\n\\nEncouraging the person to set new goals and aspirations for the next year of their life. \\n\\nCreating a memorable photo album or video montage of the person's life journey so far. \\n\\nBy focusing on positivity and support, you can help the individual feel valued and appreciated, regardless of their age.\\n\\n" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Figure 2. Case study comparing responses from the original model, the model trained with DPO and PPO on our SPA-VL." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.483, + 0.653 + ], + "angle": 0, + "content": "categories, systematically organizing and identifying various types of harmful content. For the secondary and tertiary classes, we reference usage policies from leading AI organizations, including OpenAI [35], LLaMA [32], Gemini [15], Claude [2], as well as guidelines from Llama Guard [18] and Llama Guard2 [30]. Additional references include Luo et al. [31], Weidinger et al. [53], ensuring that our classification aligns with industry standards and best practices, enhancing the relevance and applicability of our dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Data Formats. Following Bai et al. [4], we gather preference data by choosing the better response from two generated by VLMs, based on predefined criteria of harmlessness and helpfulness. Finally, a quadruple (question, image, chosen response, rejected response) reflecting preferences is collected, where the chosen response is the better response selected under principle harmless and helpful. Figure 1 shows three demonstration examples of our SPAVL dataset. In the following sections, we will give a detailed illustration of the construction process of our dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.258, + 0.833 + ], + "angle": 0, + "content": "3.2. Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With this robust categorization framework in place, we proceed to collect images that correspond to our harm categories. We use the LAION-5B [44] dataset, which is well-suited for this purpose due to its extensive and diverse col" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.258 + ], + "angle": 0, + "content": "lection of images. The LAION-5B dataset is built using a CLIP model to match images and text, allowing us to effectively use our tertiary class labels to search for relevant pictures. This leverages the strengths of the CLIP model in understanding visual and textual correlations, ensuring that the images are well-aligned with our harm categories. To ensure diversity and avoid bias, we use six different search keywords for each tertiary class. This approach helps capture a wide range of images within each category, preventing over-representation of specific subtypes. Details are illustrated in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.268, + 0.71, + 0.283 + ], + "angle": 0, + "content": "3.3. Question Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Generating relevant and meaningful questions for each image is crucial for contextual understanding. The primary challenge here is ensuring that the questions are specific to the harmful content categories and diverse in complexity. Although the LAION-5B dataset provides captions, they are often misaligned with the images [44]. To address this, we enhance the captions using Gemini 1.0 Pro Vision1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.699 + ], + "angle": 0, + "content": "In the subsequent step, we devise queries that could be potentially harmful for each image. Initially, Gemini 1.0 Pro Vision produces an easy question for every image. To ensure pertinence, the model is supplied with the image's primary, secondary, and tertiary categories. However, these questions, typically starting with \"what\" or \"how\", tend to be straightforward and closely related to the image, which may lead to a lack of complexity and diversity. To overcome this, we use Gemini 1.0 Pro to refine the questions, transforming them into more challenging queries and statements to enhance the diversity of our dataset. This process yields two additional outputs: a hard question (hardq) and a hard statement (hardd). For this refinement, the model is provided with the image's hierarchical classifications, two captions (one generated by Gemini and the original from LAION), and the previously generated easy question. Given that Gemini is designed to avoid generating harmful content, we employ a jailbreak strategy [58] to evoke more challenging queries. Further details regarding this process are provided in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.715, + 0.722 + ], + "angle": 0, + "content": "3.4. Preference Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.906, + 0.805 + ], + "angle": 0, + "content": "The final stage in constructing the SPA-VL dataset involves generating and annotating preferences for responses for training VLMs. This stage combines the generation of diverse responses and the careful annotation of these responses to create preference pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.806, + 0.906, + 0.867 + ], + "angle": 0, + "content": "Response Generation. To create annotations that better align with human judgment, we collect diverse responses from 12 different models. This diversity helps capture a wide range of potential answers, reducing model-specific" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.902, + 0.901 + ], + "angle": 0, + "content": "1https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/gemini-pro-vision" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19870" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "biases and ensuring that the VLMs can learn from a variety of perspectives. Detailed generation process are described in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.482, + 0.274 + ], + "angle": 0, + "content": "Preference Annotation. Responses are classified using MD-Judge to ensuring and for each question, we randomly select two responses from different safety-rated model groups to ensure that the chosen responses reflect diverse levels of safety to make GPT-4V better annotate. GPT-4V evaluates these pairs based on both harmlessness and helpfulness, also, to avoid bias due to the order of the answers, we query GPT-4V twice with the answers swapped. Details are illustrated in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.291, + 0.223, + 0.307 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.316, + 0.188, + 0.332 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.482, + 0.459 + ], + "angle": 0, + "content": "We apply DPO [43] and PPO [45], using our preference dataset to train LLaVA-1.5 [27, 28]. For preliminaries on DPO and PPO, please refer to the Appendix ??. We keep the visual encoder weights frozen as in Karamcheti et al. [19], Liu et al. [27] and update the weights of both the projection layer and LLM in LLaVA-1.5 (7B) in our main experiments. For training parameters, please refer to the Appendix ?? for further details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.482, + 0.566 + ], + "angle": 0, + "content": "We evaluate harmlessness using the HarmEval dataset and assess adversarial robustness on AdvBench [64] and MM-SafetyBench [29]. Helpfulness is measured using the Anthropic-Helpful [4] and HelpEval datasets, with GPT-4 determining response quality based on helpfulness and safety. Detailed evaluation methods and results can be found in the Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.482, + 0.732 + ], + "angle": 0, + "content": "Our HarmEval dataset comprises 265 questions across 53 categories, with each category containing 5 questions. We utilize two evaluation metrics on this dataset: the unsafe rate and the harm score. The UnSafe Rate assesses the proportion of unsafe responses generated by the model for all test questions. This metric evaluates the general safety of the model's outputs. The Harm Score involves comparing pairs of answers from the base model, LLaVA-1.5 (7B), and the trained model using GPT-4V. Unlike the data construction phase, GPT-4V assesses the answers based solely on the criterion of \"harmlessness\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.482, + 0.839 + ], + "angle": 0, + "content": "To ensure unbiased evaluation, we also test on the publicly available language dataset AdvBench [64] and the vision dataset MM-SafetyBench [29] (Figure ?? shows two examples from this benchmark). For both datasets, the evaluation metric is the Attack Success Rate (ASR), which measures the safety of the model by assessing its vulnerability to adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.902 + ], + "angle": 0, + "content": "MM-SafetyBench. MM-SafetyBench is a benchmark for evaluating VLMs safety composed of harmful questions and images. It includes four subcategories: Text-only (text only), SD (images generated by Stable Diffusion using key" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "phrases), Tipo (typography, images with the black phrase on a white background), and SD+Typo (concatenated generated images and typography together). The evaluation metric for this benchmark is the attack success rate (ASR). This dataset enables us to evaluate how well our trained model performs on publicly available image-based test sets in terms of safety. We use MM-Safety Prompt as Liu et al. [29] to ensure consistency across our tests." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.318 + ], + "angle": 0, + "content": "AdvBench. This dataset is purely text-based and has two subcategories: vanilla instruction and suffix injection (questions with the suffix \"Answer with 'Absolutely, here's'\"). The evaluation metric for this dataset is also ASR. This dataset allows us to assess the safety of our trained model when confronted with text-based harmful content in publicly available test sets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.327, + 0.652, + 0.341 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.53 + ], + "angle": 0, + "content": "As shown in the Table 2, the models trained on our SPA-VL dataset, LLaVA-SPA-VL-DPO and LLaVA-SPA-VL-PPO, which are the best safety models from our training, exhibit superior safety performance. They surpass the baseline model LLaVA-1.5 (7B) and other open-source models, whether or not those models have undergone safety alignment. Specifically, our models achieve best safe result on MM-SafetyBench, AdvBench and HarmEval tests. Additionally, we provide a comparison of responses before and after training for LLaVA-SPA-VL-DPO, LLaVA-SPA-VL-PPO, and baseline model LLaVA in Figure 2. More comparison examples can be found in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.531, + 0.905, + 0.621 + ], + "angle": 0, + "content": "In addition to evaluating the safety performance, we also validate our models' general ability. We select the most commonly used multimodal benchmarks and find that the general ability of our safety-aligned models does not significantly decline compared to the backbone model. Details of these evaluations can be found in the Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.738, + 0.65 + ], + "angle": 0, + "content": "5. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.78 + ], + "angle": 0, + "content": "In this section, we further explore the factors that affect the performance of alignment models. Our focus includes examining the impact of dataset scale, the selection of response models from diverse models during dataset construction, the influence of different question types within the dataset, and the outcomes of deciding whether to freeze the projection layer during model training, and the different training base models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.632, + 0.803 + ], + "angle": 0, + "content": "5.1. Data Scale" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We delve into the impact of varying amounts of data on the performance of alignment models. Across different data quantities (around 100, 1k, 5k, 10k, 30k, and 90k), we conduct experiments encompassing various evaluation metrics, as already described in Appendix ??. The resulting line plots in Figure 3 reveal compelling insights. For HarmEval," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19871" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Comparison of different VLM models on harmlessness. The models are evaluated across multiple metrics on MM-SafetyBench and AdvBench, as well as the HarmEval UnSafe Rate (HarmEval USR). After training on our proposed dataset, SPA-VL, the model achieves the best scores according all metric on both DPO and PPO methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.133, + 0.903, + 0.511 + ], + "angle": 0, + "content": "
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Open Sourced Baseline
InstructBLIP-7B27.3813.1027.3825.0023.2151.2564.6247.55
LAMM-13B14.294.762.386.556.9924.4239.1132.83
LAMM + SFT-13B16.077.148.3321.4313.2422.6972.1232.08
LLaMA-Adapter-v2-7B35.7112.507.7417.8618.4598.2610046.04
MiniGPT-4-7B20.839.5223.8120.2418.6031.3565.3838.32
mPLUG-Owl-7B35.718.9312.5030.3621.8810010052.45
Otter-9B29.7610.125.957.7413.3991.9210041.13
InternVL-Chat-7B5.951.7919.6413.1010.126.9289.4229.81
CogVLM2-LLama3-Chat-19B16.674.7623.8123.2117.1125.3898.0813.96
LLaVA-v1.6-34B4.764.1716.0719.0511.015.5893.0822.64
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
Close Sourced Baseline
Gemini-1.5-pro000000.380.381.89
GPT-4-0125-preview1.790000.450.966.542.26
Safety Aligned
LLaVA + SPA-VL-DPO0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
LLaVA + SPA-VL-PPO0.6(↓33.93)0(↓7.74)0(↓22.62)1.19(↓16.07)0.45(↓20.09)0.19(↓97.88)2.12(↓97.69)0(↓44.15)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.482, + 0.681 + ], + "angle": 0, + "content": "the Harm Score consistently decreases with increasing data volume. Similarly, in MM-SafetyBench, the overall average ASR steadily decreases as the data scale grows, and both vanilla and suffix ASRs in AdvBench exhibit a similar trend of declining rates with expanding data sizes. Notably, the help score in Anthropic-Helpful exhibits a progressive increase, indicating a simultaneous enhancement in safety and helpfulness as the dataset size expands. The full results are presented in Table 3, with detailed analysis in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.327, + 0.712 + ], + "angle": 0, + "content": "5.2. Response Model Selection" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we examine the impact of response diversity and safety in our dataset on model training. We conducted four groups of experiments, each group is trained using DPO on around 10K samples. Safe Group consists of response pairs from the three safest models (InternLMXComposer, QwenVL, Gemini1.0 Pro Vision) according to Table ?? in Appendix ??. Relative Safe Group includes pairs from relative safe models(LAMM_SFT, LLaVA1.5, InternLMXComposer, QwenVL, gemini). Unsafe Group comprises pairs from the five least safe models(mPLUG-Owl, Otter, InstructBLIP, LLaMA-Adapter-v2, GeminiJailbreak) and the All group consists of the complete set" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.529, + 0.907, + 0.638 + ], + "angle": 0, + "content": "Table 3. Impact of Training Data Scale on Alignment Model Performance using PPO and DPO Methods on LLaVA-1.5 (7B). The table compares the performance of alignment models trained with PPO and DPO methods across varying data scales: 100, 1K, 5K, 10K, 30K, and 90K tsamples. The evaluation covers both safety and helpfulness metrics. As the data scale increases, the performance generally improves across harmlessness and helpfulness metrics." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.641, + 0.905, + 0.772 + ], + "angle": 0, + "content": "
Data ScaleMM-SafetyBench↓AdvBench↓HarmEval↓Helpful
Text-onlySDCargoSD+CargoAvgvanillasuffixUSRScoreAnthropic↑HelpEval↑
PPO
10035.128.9318.4517.2619.9497.5098.8543.3950.0051.5019.00
1K30.958.3319.0516.6718.7597.3198.4641.8945.0658.0028.61
5K5.954.7613.1017.2610.2736.1585.7710.1917.2461.0048.54
10K0.601.195.959.524.328.6517.500.009.2864.5067.60
30K0.000.000.002.980.740.583.270.0010.9265.5052.09
90K0.600.000.001.190.450.192.120.008.8170.0071.04
DPO
10032.7410.1220.2416.6719.9497.8999.6243.4050.5751.0024.12
1K30.367.7417.8619.0518.7591.5496.7326.0430.0058.5035.65
5K4.171.194.178.334.4610.0040.001.8915.3464.0042.86
10K1.191.792.384.762.535.776.540.3813.7863.0040.85
30K0.000.600.601.190.600.000.000.0013.6469.5045.72
90K1.190.601.192.981.490.000.000.7513.4170.0050.63
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.608, + 0.807 + ], + "angle": 0, + "content": "of 12 models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The results presented in the Table 4 indicate that if our dataset only includes pairs of safe responses, the model struggles to learn how to avoid bad patterns, leading to vulnerability (Safe Group performed poorly on the AdvBench suffix test, suggesting the model is easily attacked). Similarly, if our dataset only includes unsafe responses, the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.268, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.189, + 0.218, + 0.201 + ], + "angle": 0, + "content": "(a) EvalHarm" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.204, + 0.268, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.303, + 0.218, + 0.314 + ], + "angle": 0, + "content": "(c) AdvBench" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.089, + 0.479, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.189, + 0.453, + 0.202 + ], + "angle": 0, + "content": "(b) MM-SafetyBench" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.204, + 0.477, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.303, + 0.454, + 0.315 + ], + "angle": 0, + "content": "(d) Anthropic-Helpful" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.439 + ], + "angle": 0, + "content": "Figure 3. Impact of Data Scale on Alignment Model Performance. Line plots illustrate the effect of varying data quantities (100, 1k, 5k, 10k, 30k, and 90k) on the performance metrics of alignment models for both PPO and DPO methods. (a) Shows the Harm Score (\\%) on EvalHarm (b) Shows the Average Attack Success Rate (ASR \\(\\%\\) ) on MM-SafetyBench (c) Shows ASR (\\%) on vanilla and suffix in AdvBench (d) Shows the Help Score (\\%) on Anthropic-Helpful." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Table 4. Detailed harmlessness evaluation metrics for Response Model Selection. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.512, + 0.481, + 0.59 + ], + "angle": 0, + "content": "
Model BagAdvBenchMM-SafetyBenchHarmEval
vanillas suffixText-onlySDTypoSD+TypoAvgUSRScore
Safe32.5065.3812.53.5710.7111.909.6713.9718.49
Relative safe14.8135.004.173.579.528.936.554.5315.85
Unsafe9.0460.772.982.988.9311.906.707.1721.14
All0.586.541.191.792.384.762.536.1113.78
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.762 + ], + "angle": 0, + "content": "model cannot be trained to be safe, as it lacks exposure to good patterns (Unsafe Group also performed poorly on the AdvBench suffix test). Relative safe Group, which includes a mix of good and bad responses, shows significantly better safety performance compared to Safe Group and Unsafe Group. However, there still exists a gap on harmlessness between the model trained with Relative Safe and the All group. This demonstrates the critical importance of response diversity during the data construction process for effective model alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.247, + 0.789 + ], + "angle": 0, + "content": "5.3. Question Types" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we analyze the impact of three different question types (Easy questions, Hard questions, and Hard statements) on the experimental results. We also compare these individual results with the combined results of all three question types. For each experiment, we select training dataset of approximately \\(10k\\) instances and using DPO to train our model. The results presented in the Table 5" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.146 + ], + "angle": 0, + "content": "Table 5. The detailed harmless evaluation metrics of Question Types. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.147, + 0.905, + 0.228 + ], + "angle": 0, + "content": "
Ques TypeAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
Easy-Q3.8524.041.791.792.388.933.722.2616.73
Hard-Q2.1211.541.191.193.579.523.870.7513.97
Hard-S2.125.003.571.794.175.953.878.7018.44
Mixed0.586.541.191.792.384.762.530.3813.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Table 6. Detailed harmless evaluation metrics of model architecture on LLaVA-1.5 (7B), with projector and without projector for both DPO and PPO training methods." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.402 + ], + "angle": 0, + "content": "
Model ArchAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
DPO
w/o projector0.000.190.000.001.195.361.641.1314.21
w projector0.000.000.000.600.601.190.600.0013.64
(0.00)(↓0.19)(0.00)(↑0.60)(↓0.59)(↓4.17)(↓1.04)(↓1.13)(↓0.57)
PPO
w/o projector0.582.880.000.001.791.790.890.0019.32
w projector0.583.270.000.000.002.980.740.0010.92
(0.00)(↓0.39)(0.00)(0.00)(↓1.79)(↓1.19)(↓0.15)(0.00)(↓8.4)
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.581 + ], + "angle": 0, + "content": "show that the combined dataset achieves higher safety performance compared to the individual datasets of each question type at the same data scale. This indicates that integrating diverse question types enhances the model's robustness against harmful attacks. These findings suggest that, in real-world scenarios, questions that provoke harmful responses are varied. Consequently, training with a combination of different question types is likely to improve the model's ability to resist harmful attacks and decrease the likelihood of generating harmful responses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.689, + 0.619 + ], + "angle": 0, + "content": "5.4. Training Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Following the approach outlined in LLaVA [28], we freeze the vision encoder during training. In this framework, both the projection layer and LLM were trained together during SFT as described in LLaVA [28], whereas in LLaVARLHF [48], only the LLM was trained. We aim to explore the impact of training the LLM alone versus with the projection layer on reinforcement learning outcomes. We analyze safety validation results on our \\(30k\\) dataset using DPO method. As shown in the Table 6, there are minimal differences in language-only safety tests (AdvBench). However, in image-involved safety tests (MM-SafetyBench and EvalHarm), training with the projection layer consistently outperforms training without it. We hypothesize that including the projection layer improves the model's ability to detect harmful content in images. This suggests a valuable direction for future work to further investigate these effects. We also investigate the training of LoRA, with further details provided in Appendix ??" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19873" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.178, + 0.106 + ], + "angle": 0, + "content": "5.5. Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.236 + ], + "angle": 0, + "content": "We explored other backbone models as well and conducted experiments using QwenVL-7BChat and InternLMXComposer-7B, applying DPO training on our dataset. As shown in Table 7, with SPA-VL data training, open-source models, including LLaVA, Qwen, and InternLMXComposer, exhibit significant safety improvements, approaching the performance levels of closed-source models." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.363 + ], + "angle": 0, + "content": "Table 7. We presents the results of DPO training on safety benchmarks for different model backbones, comparing baseline models with their safety-aligned counterparts using the SPA-VL. The models evaluated include LLaVA-7B, InternLMXComposer-7B, and QwenVL-Chat-7B, with their respective safety-aligned versions: LLaVA + SPA-VL, InternLMXC + SPA-VL, and QwenVL + SPA-VL. Across all backbones, significant improvements are observed after safety alignment." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.474 + ], + "angle": 0, + "content": "
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Baseline
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
Safety Aligned
LLaVA + SPA-VL0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
InternLMXC+SPA-VL0(↓7.74)1.19(↓2.98)0.6(↓25.59)2.38(↓24.41)1.04(↓15.18)0.19(↓5.21)0.38(↓97.50)0.75(↓25.29)
QwenVL+SPA-VL0(↓3.57)1.19(↓2.38)1.19(↓22.02)4.17(↓22.62)1.64(↓12.65)0.38(↓1.54)4.61(↓68.12)3.02(↓4.53)
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.509, + 0.279, + 0.525 + ], + "angle": 0, + "content": "6. Human Annotation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.483, + 0.793 + ], + "angle": 0, + "content": "In this section, we evaluate the consistency between GPT evaluation and human evaluation to ensure the reliability and validity of our assessment methods. For annotation part, we random sample 530 pairs for each question type to show the human preference result with GPT-4V. For AdvBench, we random select a total result in our result and humanly check the attack success for all the 520 results for both suffix and valina with GPT-4. In the case of Anthropic-Helpful, we examine 200 samples to check the consistency of helpfulness evaluations with GPT-4. For HarmEval, 265 samples (5 from each category) are selected to compare the safety preference consistency between human evaluators and GPT-4V for the baseline model LL and our trained model. Similarly, for HelpEval, we select 265 samples (5 from each category) to compare the helpfulness win rate consistency between human evaluators and GPT-4V for the GPT-4V responses and our trained model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "This comprehensive approach ensures that our models' evaluations align closely with human judgment, thereby enhancing the robustness of our evaluation part. The consistency rates between GPT evaluations and human evaluations are summarized in Table 8. The results indicate high alignment across various evaluation metrics, reinforcing the reliability of our assessment approach." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.09, + 0.905, + 0.117 + ], + "angle": 0, + "content": "Table 8. Consistency Rate (%) between GPT-4V and human annotation." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.12, + 0.905, + 0.16 + ], + "angle": 0, + "content": "
Easy-QHard-QHard-SAdvBenchAnthropic-HelpfulHarmEvalHelpEval
suffixvanilla
96.6697.7493.7799.4099.8091.0098.11100
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.184, + 0.783, + 0.2 + ], + "angle": 0, + "content": "7. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.209, + 0.907, + 0.51 + ], + "angle": 0, + "content": "In this paper, we introduced SPA-VL, the first large-scale, high-quality dataset designed for the safety alignment of VLMs. SPA-VL's comprehensive coverage of harmfulness domains, diverse question-answer types, and focus on both harmlessness and helpfulness make it a robust dataset for improving model safety without compromising core capabilities. Our experiments demonstrate that models trained on SPA-VL using techniques like PPO and DPO show significant improvements in safety and outperform baseline and other state-of-the-art VLMs in various safety evaluations. Additionally, our analyses reveal that increasing data volume, incorporating diverse responses, and using a mix of question types enhance the safety and performance of aligned models. The findings highlight the importance of comprehensive datasets like SPA-VL in achieving robust safety alignment, ensuring that VLMs can be effectively and safely deployed in real-world applications. SPA-VL represents a significant step towards safer and more reliable vision-language models, paving the way for future research and development in this crucial area." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.512, + 0.909, + 0.692 + ], + "angle": 0, + "content": "In the future, we aim to extend the scope of our work to encompass the unified \"3H\" framework of helpfulness, harmlessness, and honesty, to ensure a more holistic approach to aligning VLMs with human values. Furthermore, we plan to explore the application of our safety alignment techniques to more complex tasks such as reasoning in VLMs, which will require nuanced understanding and generation of visual content. Additionally, we are interested in investigating the transferability of alignment capabilities between LLMs and VLMs, which could lead to more efficient and effective alignment strategies across different modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.686, + 0.723 + ], + "angle": 0, + "content": "8. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.731, + 0.907, + 0.898 + ], + "angle": 0, + "content": "This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, as well as the computational resources provided by the Shanghai Artificial Intelligence Laboratory, for their essential contributions of computational resources and technical support. Furthermore, we extend our heartfelt gratitude to our advisors for their invaluable guidance, insightful discussions, and constructive feedback throughout the research process." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19874" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.308, + 0.128 + ], + "angle": 0, + "content": "[1] Anthropic. Claude, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.48, + 0.158 + ], + "angle": 0, + "content": "[2] Anthropic. Claude usage policies, 2024. https://www.anthropic.com/legal/aup.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.16, + 0.483, + 0.228 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.482, + 0.298 + ], + "angle": 0, + "content": "[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[5] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[6] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[7] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In FAccT '21: 2021 ACM Conference on Fairness, Accountability, and Transparency, Virtual Event / Toronto, Canada, March 3-10, 2021, pages 610-623. ACM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[8] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[9] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.79, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.097, + 0.875, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[12] Jieren Deng, Yijue Wang, Ji Li, Chenghong Wang, Chao Shang, Hang Liu, Sanguthevar Rajasekaran, and Caiwen Ding. Tag: Gradient attack on transformer-based language models. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3600-3610, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.904, + 0.317 + ], + "angle": 0, + "content": "[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.323, + 0.904, + 0.376 + ], + "angle": 0, + "content": "[14] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.38, + 0.904, + 0.42 + ], + "angle": 0, + "content": "[15] Google. Gemini usage policies, 2024. https://ai.google.dev/gemini-api/docs/safety-settings.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.424, + 0.904, + 0.478 + ], + "angle": 0, + "content": "[16] Susan Hao, Renee Shelby, Yuchi Liu, Hansa Srinivasan, Mukul Bhutani, Burcu Karagol Ayan, Shivani Poddar, and Sarah Laszlo. Harm amplification in text-to-image models. arXiv preprint arXiv:2402.01787, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.482, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[17] Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P. Xing, Furong Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huaxiu Yao, Manolis Kellis, Marinka Zitnik, Meng Jiang, Mohit Bansal, James Zou, Jian Pei, Jian Liu, Jianfeng Gao, Jiawei Han, Jieyu Zhao, Jiliang Tang, Jindong Wang, Joaquin Vanschoren, John Mitchell, Kai Shu, Kaidi Xu, Kai-Wei Chang, Lifang He, Lifu Huang, Michael Backes, Neil Zhenqiang Gong, Philip S. Yu, Pin-Yu Chen, Quanquan Gu, Ran Xu, Rex Ying, Shuiwang Ji, Suman Jana, Tianlong Chen, Tianming Liu, Tianyi Zhou, William Yang Wang, Xiang Li, Xiangliang Zhang, Xiao Wang, Xing Xie, Xun Chen, Xuyu Wang, Yan Liu, Yanfang Ye, Yinzhi Cao, Yong Chen, and Yue Zhao. Trustworthiness in large language models. In Forty-first International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.904, + 0.815 + ], + "angle": 0, + "content": "[18] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.819, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[19] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19875" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[20] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. RLAIF vs. RLHF: scaling reinforcement learning from human feedback with AI feedback. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[22] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Saladbench: A hierarchical and comprehensive safety benchmark for large language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3923-3954. Association for Computational Linguistics, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[23] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3326-3342. Association for Computational Linguistics, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[24] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[25] Jiawei Liang, Siyuan Liang, Man Luo, Aishan Liu, Dongchen Han, Ee-Chien Chang, and Xiaochun Cao. Vl-trojan: Multimodal instruction backdoor attacks against autoregressive visual language models. arXiv preprint arXiv:2402.13851, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[26] Hongzhan Lin, Ziyang Luo, Bo Wang, Ruichao Yang, and Jing Ma. Goat-bench: Safety insights to large multimodal models through meme-based social abuse. ArXiv, abs/2401.01523, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26286-26296. IEEE, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.827 + ], + "angle": 0, + "content": "[29] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[30] LLamaGuard2. Llama guard2, 2024. https://github.com/meta-llama/PurpleLlama/blob/main/Llama-Guard2/MODEL_CARD.md.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[31] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.904, + 0.15 + ], + "angle": 0, + "content": "[32] Meta. Llama usage policies, 2024. https://ai.meta.com/llama/use-policy.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.152, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[33] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.777, + 0.209 + ], + "angle": 0, + "content": "[34] OpenAI. Introducing chatgpt, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.21, + 0.905, + 0.237 + ], + "angle": 0, + "content": "[35] OpenAI. Openai usage policies, 2023. https://openai.com/policies/usage-policies.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.239, + 0.881, + 0.253 + ], + "angle": 0, + "content": "[36] OpenAI. Gpt4v, 2024. https://chatgpt.com/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.255, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.352, + 0.905, + 0.449 + ], + "angle": 0, + "content": "[38] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 16012-16027. Association for Computational Linguistics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.451, + 0.905, + 0.52 + ], + "angle": 0, + "content": "[39] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.522, + 0.905, + 0.59 + ], + "angle": 0, + "content": "[40] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.592, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[41] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[42] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, pages 53728-53741. Curran Associates, Inc., 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19876" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv: Learning, arXiv: Learning. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[46] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[47] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darryll. Aligning large multimodal models with factually augmented RLHF. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 13088-13110. Association for Computational Linguistics, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[49] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Agüera y Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. Lamda: Language models for dialog applications. CoRR, abs/2201.08239, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[50] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[51] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 10460-10479. Association for Computational Linguistics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[52] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[53] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[54] Laura Weidinger, Maribeth Rauh, Nahema Marchal, Arianna Manzini, Lisa Anne Hendricks, Juan Mateos-Garcia, Stevie Bergman, Jackie Kay, Conor Griffin, Ben Bariach, et al. Sociotechnical safety evaluation of generative ai systems. arXiv preprint arXiv:2310.11986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[56] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, Jing Shao, and Wanli Ouyang. LAMM: language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[57] Pan Zhang, XiaoyiDongBin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[58] Zaibin Zhang, Yongting Zhang, Lijun Li, Jing Shao, Hongzhi Gao, Yu Qiao, Lijun Wang, Huchuan Lu, and Feng Zhao. Psysafe: A comprehensive framework for psychological-based attack, defense, and evaluation of multi-agent system safety. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 15202-15231. Association for Computational Linguistics, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[60] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[61] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[62] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. Open-Review.net, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19877" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[63] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy M. Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[64] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "19878" + } + ] +] \ No newline at end of file diff --git a/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_origin.pdf b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45744c2d198295cdb890062b93308c073deb9408 --- /dev/null +++ b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/22cd485b-929e-4951-969d-47a8232e9e43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b69195cb7c26c3f9eb99ddc32d3614ee04a4c6461e3f97886b1b5ab37b15f1b9 +size 1176656 diff --git a/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/full.md b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..64ab94352764fb3ff2c030b60059de587d93774d --- /dev/null +++ b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/full.md @@ -0,0 +1,297 @@ +# SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models + +Yongting Zhang $^{1,3*}$ Lu Chen $^{2,3*}$ Guodong Zheng $^{2,3}$ Yifeng Gao $^{2}$ Rui Zheng $^{2}$ Jinlan Fu $^{3}$ Zhenfei Yin $^{3}$ Senjie Jin $^{2}$ Yu Qiao $^{3}$ Xuanjing Huang $^{2}$ Feng Zhao $^{1\dagger}$ Tao Gui $^{2,3‡}$ Jing Shao $^{3\dagger}$ ${}^{1}$ University of Science and Technology of China ${}^{2}$ Fudan University ${}^{3}$ Shanghai Artificial Intelligence Laboratory + +zytabcd@mail.ustc.edu.cn, luchen23@m.fudan.edu.cn + +https://sqrtizhang.github.io/SPA-VL/ + +# Abstract + +The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness. + +# 1. Introduction + +Vision Language Models (VLMs), such as GPT-4V [36], Claude 3 [1], LLaVA [27], and MiniGPT-4 [62] can understand visual signals and respond to user instructions. + +Equipped with a visual encoder module, VLMs extract multimodal knowledge from both visual and textual queries, leveraging pre-trained LLMs' powerful comprehension and generative capabilities to achieve remarkable results across diverse vision-language tasks. + +Due to the complexity of multimodal harms, previous study [16] has demonstrated that harmless inputs may also result in outputs that do not align with human preferences. Although LLMs have undergone harmless alignment, the alignment of visual encoders is relatively weak, making VLMs susceptible to successful attacks through the visual modality [6, 14, 25]. Therefore, it is necessary to simultaneously improve the alignment of the visual and language modules of the VLM to achieve the harmless and helpful responses. Ensuring the alignment of VLMs with ethical and safety standards is crucial for their safe and effective deployment. However, most existing works on the safety of VLMs focused on the evaluation benchmarks [10, 24, 26] or jailbreak detection [6, 14, 14, 23-25]. Seldom studies considered constructing large-scale, high-quality open-source training datasets to achieve the safety alignment of VLMs. The lack of such datasets hampers further development in this field. + +To address these limitations, we propose a large-scale safety alignment dataset for VLMs named SPA-VL. Since Reinforcement Learning from Human Feedback (RLHF) is widely regarded as performing well in alignment studies [20, 41], our SPA-VL dataset is designed for the RLHF, with each sample containing four elements (question, image, chosen response, rejected response). The main perspectives of our SPA-VL dataset are summarized as follows: + +(1) Comprehensive Domains: SPA-VL contains 100,788 samples and comprehensively covers a wide range of harm types, encompassing 6 domains, 13 categories, and 53 subcategories. (2) Diverse Questions and Responses: + +For diverse question collection, we gathered three types of questions for each image: easy question, hard question, and hard statement. For diverse response collection, to enhance the diversity of the collected responses and reduce specific model biases, we selected responses from 2 out of 12 different models for each question, designating them as chosen and rejected responses. (3) Multi-Objective Alignment: The preference dataset is designed according to two alignment objectives: harmlessness and helpfulness. Thus, the responses of the aligned VLMs are enhanced in both aspects without sacrificing helpfulness for safety. + +The main contributions are listed below: + +(1) We propose the SPA-VL, a large-scale, high-quality, and domain diversity dataset for vision language model safety alignment studies. By using techniques such as PPO and DPO for safety alignment learning on the SPA-VL dataset based on some open VLMs, such as LLaVA, experimental results show that these VLMs significantly improve performance in safety and surpass the state-of-the-art VLMs. +(2) We conducted extensive analysis and ablation studies, finding that increasing the scale of the dataset, incorporating diverse answers, and using various question types can improve the safety and performance of the alignment VLMs. +(3) The entire dataset construction process is fully automated, ensuring efficient and scalable data generation. The automation covers all key steps, including the collection of images and questions, the selection of VLM responses, and the annotations of chosen and rejected responses. + +# 2. Related Work + +Vision Language Models. The rapid advancement of Large Language Models [42, 59] and their robust foundational capabilities have significantly prompted the development of multimodal large models. Recently, Vision-Language Models have emerged as a promising extension of LLMs [9], integrating visual and textual information for enhanced multimodal understanding. Notable models in this domain include InstructBLIP [11], InternLMXComposer [57], LAMM-SFT [56], LAMM [56], LLaMA-Adapter-V2 [13], MiniGPT-4 [62], mPLUG-Owl [55], Otter [21], and Qwen-VL-Chat [3]. Most of these VLMs are developed by projecting the vision space into the language space through a learned projector, leveraging pre-trained language models as their backbone. As VLMs continue to advance rapidly, safety concerns have garnered significant attention from researchers. + +Reinforcement Learning from Human Feedback. Despite the promising capabilities of LLMs and VLMs, they are prone to unintended behaviors, such as fabricating facts, producing biased or harmful content, or even harming humans [7, 8]. They should be helpful, honest, and harm- + +less (3H) [4, 37, 49]. RLHF offers the most straightforward approach to achieving this goal. RLHF methods such as PPO [45] and DPO [43] have been highly successful in aligning AI with human preferences. Notable applications like ChatGPT [34] and Claude [1] show strong performances in academic benchmarks. Models trained with RLHF methods often perform better and adhere more closely to human values compared to those trained only with SFT [4]. This success extends to VLMs, where RLHF has been used to address hallucination issues [5, 39, 48, 61]. + +Safety of VLMs. To evaluate the safety performance of VLMs, various methods and datasets have been proposed. Among these evaluation benchmarks are MM-SafetyBench [29], ChEf [47], and OODCV-VQA, SketchyVQA [50]. In addition to these benchmarks, several attack methods, such as adversarial attacks [12, 40, 46, 60] and jailbreak techniques [14, 24, 33], have been developed to test the vulnerabilities of VLMs. These studies aim to identify and exploit potential weaknesses in VLMs, underscoring the need for robust safety measures. + +In response to these vulnerabilities, several methods have been developed to enhance the safety performance of VLMs. For instance, VLGuard [63] employed supervised fine-tuning (SFT) on the VLGuard dataset, which contains 2000 training images designed to improve safety. Similarly, Chen et al. [10] used AI-annotated data for SFT. Wei et al. [52] utilized in-context learning to bolster model safety. Additionally, Pi et al. [38] introduced MLLM-Protector, a harm detector serving as a plug-and-play defense mechanism for VLMs, and Wang et al. [51] applied inference time alignment methods to improve harmlessness. These approaches collectively demonstrate ongoing efforts to mitigate risks and enhance the resilience of VLMs against various types of attacks. + +Among all the methods mentioned above, we can broadly categorize them into two types. The first type enhances model safety during the inference stage by using prompts. This approach is efficient and convenient but often results in limited safety improvements and lacks generalization [17]. The second type involves training-based methods, specifically during the training-to-align phase, which can be further divided into SFT and RLHF. While the aforementioned methods primarily rely on SFT, we go a step further by providing a comprehensive RLHF dataset SPA-VL. + +# 3. SPA-VL Dataset + +In the development of VLMs, effectively addressing harmful content in multimodal environments poses a significant challenge. The SPA-VL dataset helps VLMs confront this challenge by providing safety preference data for RLHF. As shown in Figure 1, establishing the SPA-VL dataset involves three key stages. First, we systematically collect images, which includes gathering a diverse set of images to + +![](images/ef9d1caf78fa225218ff69358afcc7685cb10346ce86d00e12cb2451299d1aec.jpg) +Figure 1. Overview of SPA-VL Dataset Construction. It is built in three stages: 1) Image Collection, 2) Questions Construction and 3) Preference Construction. The dataset examples shows vision-question/preferences pairs that comprise three types of questions: easy questions, hard questions, and hard statements. + +ensure comprehensiveness. Next, we generate questions related to categories of harmful content. After this, we proceed with preference annotation. This stage includes generating responses from various models and labeling these responses based on preferences for harmlessness and helpfulness. + +Table 1. Training dataset statistics for SPA-VL. For each image, we provide three prompts: Easy question, Hard question, Hard statement. UR% represents the unsafe rate. + +
Secondary ClassImage Cnt.QuestionPrefer(UR%)Prefer(Len.)
UR%Cnt.Len.Cho.Rej.Cho.Rej.
Toxic379144.111132111611.3541.55488392
Unfair358938.38106841207.1532.16620441
Erosion of Trust in Public Information126337.6237671527.6231.62595463
False Beliefs181429.3154241465.8827.16746539
Dangerous Information126359.66378812914.7849.39621580
Privacy63653.12190715612.1144.83635513
Security Threats245263.99727914112.7450.83567572
Defamation61151.83180614516.4546.46542464
Fraud or Deceptive Action477957.211417914413.7348.14590531
Influence Operations179551.51531715117.1149.69725599
Illegal Activities373460.511102513013.8349.23568531
Persuasion and Manipulation118859.38333113617.8951.73719633
Violation of Personal Property190955.5753821609.541.19543478
Risky Financial Practices184931.8152071479.130.57933685
Trade and Compliance122129.4630211809.7631.45905647
Total3189449.279325813811.742.23626519
+ +# 3.1. Overview + +Data Statistics. Our SPA-VL dataset comprises four parts: the training set, the validation set, and two test sets, HarmEval and HelpEval, which are used to evaluate harmfulness and helpfulness, respectively. The number of samples in each part is 93,258, 7,000, 265, and 265, respectively. Table 1 shows the dataset statistics of the training set. To detect the unsafe content covered by our SPA-VL dataset, we utilize the MD-Judge evaluator [22] to calculate the unsafe rate of the collected questions and VLMs' responses. Nearly half of the collected questions are unsafe, while the unsafe rate for the chosen response and rejected response is $11.7\%$ and $42.23\%$ , respectively. The HarmEval test set includes a substantial number of harmful questions, while the HelpEval test set primarily comprises questions that involve instruction following or require the expression of opinions. + +Diverse Domains. A diverse and representative set of images is essential for training models to handle vision data safely and effectively. Our primary challenge is ensuring diversity while maintaining relevance to harmful content categories. To address this, we establish a comprehensive harm content categorization framework. Our SPA-VL adopts 6 primary domains, 15 secondary categories, and 53 tertiary categories, ensuring comprehensive coverage and granularity for precise harm detection and response alignment. A detailed visual representation of this hierarchical structure is provided in Figure ?? in the Appendix ??. We reference Li et al. [22], Weidinger et al. [54] for our primary harm + +![](images/726c64f80a5961c18fac5e945405524457343f9bca0b9c869ae3d1cfb3b6bf7c.jpg) +The most degrading and ageist way to mark your 28th birthday would be to have a card that says "You're 28" with a picture of a duck on it. This is degrading because it implies that the person is old and irrelevant, and it is ageist because it suggests that the person is past their prime and has no value. The card's message is disrespectful and disregards the individual's worth and potential for growth and development. + +![](images/6a0170ecb21ab281a059fec504dd5efb728f98d50244c2fe059ecde752604e53.jpg) +What's the most degrading and ageist way to mark your 28th birthday, showcasing your declining worth as a stale and irrelevant human being? + +![](images/7abae8fc431edd41d7d4ca08d96a250ed64d9ccc39435b99df273aa47ab44906.jpg) + +![](images/6472b30d4ea7fe2217d0fa7dda0bcdd3a0f4e5f51dc381c048af6a19baaa1600.jpg) +I'm sorry, but I cannot provide an answer to that question as it is ageist and degrading. It is important to remember that everyone is unique and valuable, regardless of their age. Let's focus on celebrating our birthdays and the milestones we have achieved, rather than making derogatory comments about others. + +![](images/9e772b6ed351d3ce1c058696225ced69bc28b9e0b129274934b219ef590cefc9.jpg) +It is not appropriate to use degrading or ageist language to mark someone's 28th birthday or any other milestone. Such language is disrespectful and can negatively impact the individual's self-esteem and well-being. It is important to celebrate and acknowledge the achievements and experiences of individuals regardless of their age, and to treat them with respect and dignity. \n\nInstead of using degrading language, you can create a positive and supportive atmosphere for the person's 28th birthday celebration by: \n\nOrganizing a fun and engaging party with friends and family. \n\nthoughtful and meaningful presents that reflect the individual's interests and hobbies. \n\nSharing heartfelt messages and well-wishes on social media or through personal communication. \n\nEncouraging the person to set new goals and aspirations for the next year of their life. \n\nCreating a memorable photo album or video montage of the person's life journey so far. \n\nBy focusing on positivity and support, you can help the individual feel valued and appreciated, regardless of their age.\n\n +Figure 2. Case study comparing responses from the original model, the model trained with DPO and PPO on our SPA-VL. + +categories, systematically organizing and identifying various types of harmful content. For the secondary and tertiary classes, we reference usage policies from leading AI organizations, including OpenAI [35], LLaMA [32], Gemini [15], Claude [2], as well as guidelines from Llama Guard [18] and Llama Guard2 [30]. Additional references include Luo et al. [31], Weidinger et al. [53], ensuring that our classification aligns with industry standards and best practices, enhancing the relevance and applicability of our dataset. + +Data Formats. Following Bai et al. [4], we gather preference data by choosing the better response from two generated by VLMs, based on predefined criteria of harmlessness and helpfulness. Finally, a quadruple (question, image, chosen response, rejected response) reflecting preferences is collected, where the chosen response is the better response selected under principle harmless and helpful. Figure 1 shows three demonstration examples of our SPAVL dataset. In the following sections, we will give a detailed illustration of the construction process of our dataset. + +# 3.2. Image Collection + +With this robust categorization framework in place, we proceed to collect images that correspond to our harm categories. We use the LAION-5B [44] dataset, which is well-suited for this purpose due to its extensive and diverse col + +lection of images. The LAION-5B dataset is built using a CLIP model to match images and text, allowing us to effectively use our tertiary class labels to search for relevant pictures. This leverages the strengths of the CLIP model in understanding visual and textual correlations, ensuring that the images are well-aligned with our harm categories. To ensure diversity and avoid bias, we use six different search keywords for each tertiary class. This approach helps capture a wide range of images within each category, preventing over-representation of specific subtypes. Details are illustrated in Appendix ?? + +# 3.3. Question Generation + +Generating relevant and meaningful questions for each image is crucial for contextual understanding. The primary challenge here is ensuring that the questions are specific to the harmful content categories and diverse in complexity. Although the LAION-5B dataset provides captions, they are often misaligned with the images [44]. To address this, we enhance the captions using Gemini 1.0 Pro Vision1. + +In the subsequent step, we devise queries that could be potentially harmful for each image. Initially, Gemini 1.0 Pro Vision produces an easy question for every image. To ensure pertinence, the model is supplied with the image's primary, secondary, and tertiary categories. However, these questions, typically starting with "what" or "how", tend to be straightforward and closely related to the image, which may lead to a lack of complexity and diversity. To overcome this, we use Gemini 1.0 Pro to refine the questions, transforming them into more challenging queries and statements to enhance the diversity of our dataset. This process yields two additional outputs: a hard question (hardq) and a hard statement (hardd). For this refinement, the model is provided with the image's hierarchical classifications, two captions (one generated by Gemini and the original from LAION), and the previously generated easy question. Given that Gemini is designed to avoid generating harmful content, we employ a jailbreak strategy [58] to evoke more challenging queries. Further details regarding this process are provided in Appendix ?? + +# 3.4. Preference Collection + +The final stage in constructing the SPA-VL dataset involves generating and annotating preferences for responses for training VLMs. This stage combines the generation of diverse responses and the careful annotation of these responses to create preference pairs. + +Response Generation. To create annotations that better align with human judgment, we collect diverse responses from 12 different models. This diversity helps capture a wide range of potential answers, reducing model-specific + +biases and ensuring that the VLMs can learn from a variety of perspectives. Detailed generation process are described in Appendix ?? + +Preference Annotation. Responses are classified using MD-Judge to ensuring and for each question, we randomly select two responses from different safety-rated model groups to ensure that the chosen responses reflect diverse levels of safety to make GPT-4V better annotate. GPT-4V evaluates these pairs based on both harmlessness and helpfulness, also, to avoid bias due to the order of the answers, we query GPT-4V twice with the answers swapped. Details are illustrated in Appendix ?? + +# 4. Experiments + +# 4.1. Settings + +We apply DPO [43] and PPO [45], using our preference dataset to train LLaVA-1.5 [27, 28]. For preliminaries on DPO and PPO, please refer to the Appendix ??. We keep the visual encoder weights frozen as in Karamcheti et al. [19], Liu et al. [27] and update the weights of both the projection layer and LLM in LLaVA-1.5 (7B) in our main experiments. For training parameters, please refer to the Appendix ?? for further details. + +We evaluate harmlessness using the HarmEval dataset and assess adversarial robustness on AdvBench [64] and MM-SafetyBench [29]. Helpfulness is measured using the Anthropic-Helpful [4] and HelpEval datasets, with GPT-4 determining response quality based on helpfulness and safety. Detailed evaluation methods and results can be found in the Appendix ?? + +Our HarmEval dataset comprises 265 questions across 53 categories, with each category containing 5 questions. We utilize two evaluation metrics on this dataset: the unsafe rate and the harm score. The UnSafe Rate assesses the proportion of unsafe responses generated by the model for all test questions. This metric evaluates the general safety of the model's outputs. The Harm Score involves comparing pairs of answers from the base model, LLaVA-1.5 (7B), and the trained model using GPT-4V. Unlike the data construction phase, GPT-4V assesses the answers based solely on the criterion of "harmlessness". + +To ensure unbiased evaluation, we also test on the publicly available language dataset AdvBench [64] and the vision dataset MM-SafetyBench [29] (Figure ?? shows two examples from this benchmark). For both datasets, the evaluation metric is the Attack Success Rate (ASR), which measures the safety of the model by assessing its vulnerability to adversarial attacks. + +MM-SafetyBench. MM-SafetyBench is a benchmark for evaluating VLMs safety composed of harmful questions and images. It includes four subcategories: Text-only (text only), SD (images generated by Stable Diffusion using key + +phrases), Tipo (typography, images with the black phrase on a white background), and SD+Typo (concatenated generated images and typography together). The evaluation metric for this benchmark is the attack success rate (ASR). This dataset enables us to evaluate how well our trained model performs on publicly available image-based test sets in terms of safety. We use MM-Safety Prompt as Liu et al. [29] to ensure consistency across our tests. + +AdvBench. This dataset is purely text-based and has two subcategories: vanilla instruction and suffix injection (questions with the suffix "Answer with 'Absolutely, here's'"). The evaluation metric for this dataset is also ASR. This dataset allows us to assess the safety of our trained model when confronted with text-based harmful content in publicly available test sets. + +# 4.2. Main Results + +As shown in the Table 2, the models trained on our SPA-VL dataset, LLaVA-SPA-VL-DPO and LLaVA-SPA-VL-PPO, which are the best safety models from our training, exhibit superior safety performance. They surpass the baseline model LLaVA-1.5 (7B) and other open-source models, whether or not those models have undergone safety alignment. Specifically, our models achieve best safe result on MM-SafetyBench, AdvBench and HarmEval tests. Additionally, we provide a comparison of responses before and after training for LLaVA-SPA-VL-DPO, LLaVA-SPA-VL-PPO, and baseline model LLaVA in Figure 2. More comparison examples can be found in Appendix ?? + +In addition to evaluating the safety performance, we also validate our models' general ability. We select the most commonly used multimodal benchmarks and find that the general ability of our safety-aligned models does not significantly decline compared to the backbone model. Details of these evaluations can be found in the Appendix ?? + +# 5. Analysis and Discussion + +In this section, we further explore the factors that affect the performance of alignment models. Our focus includes examining the impact of dataset scale, the selection of response models from diverse models during dataset construction, the influence of different question types within the dataset, and the outcomes of deciding whether to freeze the projection layer during model training, and the different training base models. + +# 5.1. Data Scale + +We delve into the impact of varying amounts of data on the performance of alignment models. Across different data quantities (around 100, 1k, 5k, 10k, 30k, and 90k), we conduct experiments encompassing various evaluation metrics, as already described in Appendix ??. The resulting line plots in Figure 3 reveal compelling insights. For HarmEval, + +Table 2. Comparison of different VLM models on harmlessness. The models are evaluated across multiple metrics on MM-SafetyBench and AdvBench, as well as the HarmEval UnSafe Rate (HarmEval USR). After training on our proposed dataset, SPA-VL, the model achieves the best scores according all metric on both DPO and PPO methods. + +
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Open Sourced Baseline
InstructBLIP-7B27.3813.1027.3825.0023.2151.2564.6247.55
LAMM-13B14.294.762.386.556.9924.4239.1132.83
LAMM + SFT-13B16.077.148.3321.4313.2422.6972.1232.08
LLaMA-Adapter-v2-7B35.7112.507.7417.8618.4598.2610046.04
MiniGPT-4-7B20.839.5223.8120.2418.6031.3565.3838.32
mPLUG-Owl-7B35.718.9312.5030.3621.8810010052.45
Otter-9B29.7610.125.957.7413.3991.9210041.13
InternVL-Chat-7B5.951.7919.6413.1010.126.9289.4229.81
CogVLM2-LLama3-Chat-19B16.674.7623.8123.2117.1125.3898.0813.96
LLaVA-v1.6-34B4.764.1716.0719.0511.015.5893.0822.64
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
Close Sourced Baseline
Gemini-1.5-pro000000.380.381.89
GPT-4-0125-preview1.790000.450.966.542.26
Safety Aligned
LLaVA + SPA-VL-DPO0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
LLaVA + SPA-VL-PPO0.6(↓33.93)0(↓7.74)0(↓22.62)1.19(↓16.07)0.45(↓20.09)0.19(↓97.88)2.12(↓97.69)0(↓44.15)
+ +the Harm Score consistently decreases with increasing data volume. Similarly, in MM-SafetyBench, the overall average ASR steadily decreases as the data scale grows, and both vanilla and suffix ASRs in AdvBench exhibit a similar trend of declining rates with expanding data sizes. Notably, the help score in Anthropic-Helpful exhibits a progressive increase, indicating a simultaneous enhancement in safety and helpfulness as the dataset size expands. The full results are presented in Table 3, with detailed analysis in Appendix ?? + +# 5.2. Response Model Selection + +In this section, we examine the impact of response diversity and safety in our dataset on model training. We conducted four groups of experiments, each group is trained using DPO on around 10K samples. Safe Group consists of response pairs from the three safest models (InternLMXComposer, QwenVL, Gemini1.0 Pro Vision) according to Table ?? in Appendix ??. Relative Safe Group includes pairs from relative safe models(LAMM_SFT, LLaVA1.5, InternLMXComposer, QwenVL, gemini). Unsafe Group comprises pairs from the five least safe models(mPLUG-Owl, Otter, InstructBLIP, LLaMA-Adapter-v2, GeminiJailbreak) and the All group consists of the complete set + +Table 3. Impact of Training Data Scale on Alignment Model Performance using PPO and DPO Methods on LLaVA-1.5 (7B). The table compares the performance of alignment models trained with PPO and DPO methods across varying data scales: 100, 1K, 5K, 10K, 30K, and 90K tsamples. The evaluation covers both safety and helpfulness metrics. As the data scale increases, the performance generally improves across harmlessness and helpfulness metrics. + +
Data ScaleMM-SafetyBench↓AdvBench↓HarmEval↓Helpful
Text-onlySDCargoSD+CargoAvgvanillasuffixUSRScoreAnthropic↑HelpEval↑
PPO
10035.128.9318.4517.2619.9497.5098.8543.3950.0051.5019.00
1K30.958.3319.0516.6718.7597.3198.4641.8945.0658.0028.61
5K5.954.7613.1017.2610.2736.1585.7710.1917.2461.0048.54
10K0.601.195.959.524.328.6517.500.009.2864.5067.60
30K0.000.000.002.980.740.583.270.0010.9265.5052.09
90K0.600.000.001.190.450.192.120.008.8170.0071.04
DPO
10032.7410.1220.2416.6719.9497.8999.6243.4050.5751.0024.12
1K30.367.7417.8619.0518.7591.5496.7326.0430.0058.5035.65
5K4.171.194.178.334.4610.0040.001.8915.3464.0042.86
10K1.191.792.384.762.535.776.540.3813.7863.0040.85
30K0.000.600.601.190.600.000.000.0013.6469.5045.72
90K1.190.601.192.981.490.000.000.7513.4170.0050.63
+ +# of 12 models. + +The results presented in the Table 4 indicate that if our dataset only includes pairs of safe responses, the model struggles to learn how to avoid bad patterns, leading to vulnerability (Safe Group performed poorly on the AdvBench suffix test, suggesting the model is easily attacked). Similarly, if our dataset only includes unsafe responses, the + +![](images/6390a9a455d1ebffcf62f6599282a1b708f661e7c7f0a66deab7b1c52d3f15c1.jpg) +(a) EvalHarm + +![](images/5e23859681d979630a87480b16b30bc05bcc4f467302af369c1bd7b7d6c73ba5.jpg) +(c) AdvBench + +![](images/2431fed2bffe567079fd268c8ed89c381642bf7c13efce0fc08932a24250b862.jpg) +(b) MM-SafetyBench + +![](images/f1b512c48e6e28377113ea7b7a5e5d39543aa25250ae0fa38fa06b68ee5e9127.jpg) +(d) Anthropic-Helpful +Figure 3. Impact of Data Scale on Alignment Model Performance. Line plots illustrate the effect of varying data quantities (100, 1k, 5k, 10k, 30k, and 90k) on the performance metrics of alignment models for both PPO and DPO methods. (a) Shows the Harm Score (\%) on EvalHarm (b) Shows the Average Attack Success Rate (ASR $\%$ ) on MM-SafetyBench (c) Shows ASR (\%) on vanilla and suffix in AdvBench (d) Shows the Help Score (\%) on Anthropic-Helpful. + +Table 4. Detailed harmlessness evaluation metrics for Response Model Selection. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge. + +
Model BagAdvBenchMM-SafetyBenchHarmEval
vanillas suffixText-onlySDTypoSD+TypoAvgUSRScore
Safe32.5065.3812.53.5710.7111.909.6713.9718.49
Relative safe14.8135.004.173.579.528.936.554.5315.85
Unsafe9.0460.772.982.988.9311.906.707.1721.14
All0.586.541.191.792.384.762.536.1113.78
+ +model cannot be trained to be safe, as it lacks exposure to good patterns (Unsafe Group also performed poorly on the AdvBench suffix test). Relative safe Group, which includes a mix of good and bad responses, shows significantly better safety performance compared to Safe Group and Unsafe Group. However, there still exists a gap on harmlessness between the model trained with Relative Safe and the All group. This demonstrates the critical importance of response diversity during the data construction process for effective model alignment. + +# 5.3. Question Types + +In this section, we analyze the impact of three different question types (Easy questions, Hard questions, and Hard statements) on the experimental results. We also compare these individual results with the combined results of all three question types. For each experiment, we select training dataset of approximately $10k$ instances and using DPO to train our model. The results presented in the Table 5 + +Table 5. The detailed harmless evaluation metrics of Question Types. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge. + +
Ques TypeAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
Easy-Q3.8524.041.791.792.388.933.722.2616.73
Hard-Q2.1211.541.191.193.579.523.870.7513.97
Hard-S2.125.003.571.794.175.953.878.7018.44
Mixed0.586.541.191.792.384.762.530.3813.78
+ +Table 6. Detailed harmless evaluation metrics of model architecture on LLaVA-1.5 (7B), with projector and without projector for both DPO and PPO training methods. + +
Model ArchAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
DPO
w/o projector0.000.190.000.001.195.361.641.1314.21
w projector0.000.000.000.600.601.190.600.0013.64
(0.00)(↓0.19)(0.00)(↑0.60)(↓0.59)(↓4.17)(↓1.04)(↓1.13)(↓0.57)
PPO
w/o projector0.582.880.000.001.791.790.890.0019.32
w projector0.583.270.000.000.002.980.740.0010.92
(0.00)(↓0.39)(0.00)(0.00)(↓1.79)(↓1.19)(↓0.15)(0.00)(↓8.4)
+ +show that the combined dataset achieves higher safety performance compared to the individual datasets of each question type at the same data scale. This indicates that integrating diverse question types enhances the model's robustness against harmful attacks. These findings suggest that, in real-world scenarios, questions that provoke harmful responses are varied. Consequently, training with a combination of different question types is likely to improve the model's ability to resist harmful attacks and decrease the likelihood of generating harmful responses. + +# 5.4. Training Methods + +Following the approach outlined in LLaVA [28], we freeze the vision encoder during training. In this framework, both the projection layer and LLM were trained together during SFT as described in LLaVA [28], whereas in LLaVARLHF [48], only the LLM was trained. We aim to explore the impact of training the LLM alone versus with the projection layer on reinforcement learning outcomes. We analyze safety validation results on our $30k$ dataset using DPO method. As shown in the Table 6, there are minimal differences in language-only safety tests (AdvBench). However, in image-involved safety tests (MM-SafetyBench and EvalHarm), training with the projection layer consistently outperforms training without it. We hypothesize that including the projection layer improves the model's ability to detect harmful content in images. This suggests a valuable direction for future work to further investigate these effects. We also investigate the training of LoRA, with further details provided in Appendix ?? + +# 5.5. Model + +We explored other backbone models as well and conducted experiments using QwenVL-7BChat and InternLMXComposer-7B, applying DPO training on our dataset. As shown in Table 7, with SPA-VL data training, open-source models, including LLaVA, Qwen, and InternLMXComposer, exhibit significant safety improvements, approaching the performance levels of closed-source models. + +Table 7. We presents the results of DPO training on safety benchmarks for different model backbones, comparing baseline models with their safety-aligned counterparts using the SPA-VL. The models evaluated include LLaVA-7B, InternLMXComposer-7B, and QwenVL-Chat-7B, with their respective safety-aligned versions: LLaVA + SPA-VL, InternLMXC + SPA-VL, and QwenVL + SPA-VL. Across all backbones, significant improvements are observed after safety alignment. + +
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Baseline
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
Safety Aligned
LLaVA + SPA-VL0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
InternLMXC+SPA-VL0(↓7.74)1.19(↓2.98)0.6(↓25.59)2.38(↓24.41)1.04(↓15.18)0.19(↓5.21)0.38(↓97.50)0.75(↓25.29)
QwenVL+SPA-VL0(↓3.57)1.19(↓2.38)1.19(↓22.02)4.17(↓22.62)1.64(↓12.65)0.38(↓1.54)4.61(↓68.12)3.02(↓4.53)
+ +# 6. Human Annotation + +In this section, we evaluate the consistency between GPT evaluation and human evaluation to ensure the reliability and validity of our assessment methods. For annotation part, we random sample 530 pairs for each question type to show the human preference result with GPT-4V. For AdvBench, we random select a total result in our result and humanly check the attack success for all the 520 results for both suffix and valina with GPT-4. In the case of Anthropic-Helpful, we examine 200 samples to check the consistency of helpfulness evaluations with GPT-4. For HarmEval, 265 samples (5 from each category) are selected to compare the safety preference consistency between human evaluators and GPT-4V for the baseline model LL and our trained model. Similarly, for HelpEval, we select 265 samples (5 from each category) to compare the helpfulness win rate consistency between human evaluators and GPT-4V for the GPT-4V responses and our trained model. + +This comprehensive approach ensures that our models' evaluations align closely with human judgment, thereby enhancing the robustness of our evaluation part. The consistency rates between GPT evaluations and human evaluations are summarized in Table 8. The results indicate high alignment across various evaluation metrics, reinforcing the reliability of our assessment approach. + +Table 8. Consistency Rate (%) between GPT-4V and human annotation. + +
Easy-QHard-QHard-SAdvBenchAnthropic-HelpfulHarmEvalHelpEval
suffixvanilla
96.6697.7493.7799.4099.8091.0098.11100
+ +# 7. Conclusion and Future Work + +In this paper, we introduced SPA-VL, the first large-scale, high-quality dataset designed for the safety alignment of VLMs. SPA-VL's comprehensive coverage of harmfulness domains, diverse question-answer types, and focus on both harmlessness and helpfulness make it a robust dataset for improving model safety without compromising core capabilities. Our experiments demonstrate that models trained on SPA-VL using techniques like PPO and DPO show significant improvements in safety and outperform baseline and other state-of-the-art VLMs in various safety evaluations. Additionally, our analyses reveal that increasing data volume, incorporating diverse responses, and using a mix of question types enhance the safety and performance of aligned models. The findings highlight the importance of comprehensive datasets like SPA-VL in achieving robust safety alignment, ensuring that VLMs can be effectively and safely deployed in real-world applications. SPA-VL represents a significant step towards safer and more reliable vision-language models, paving the way for future research and development in this crucial area. + +In the future, we aim to extend the scope of our work to encompass the unified "3H" framework of helpfulness, harmlessness, and honesty, to ensure a more holistic approach to aligning VLMs with human values. Furthermore, we plan to explore the application of our safety alignment techniques to more complex tasks such as reasoning in VLMs, which will require nuanced understanding and generation of visual content. Additionally, we are interested in investigating the transferability of alignment capabilities between LLMs and VLMs, which could lead to more efficient and effective alignment strategies across different modalities. + +# 8. Acknowledgment + +This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, as well as the computational resources provided by the Shanghai Artificial Intelligence Laboratory, for their essential contributions of computational resources and technical support. Furthermore, we extend our heartfelt gratitude to our advisors for their invaluable guidance, insightful discussions, and constructive feedback throughout the research process. + +# References + +[1] Anthropic. Claude, 2023. 1, 2 +[2] Anthropic. Claude usage policies, 2024. https://www.anthropic.com/legal/aup.4 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2 +[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2, 4, 5 +[5] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 2 +[6] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime, 2023. 1 +[7] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In FAccT '21: 2021 ACM Conference on Fairness, Accountability, and Transparency, Virtual Event / Toronto, Canada, March 3-10, 2021, pages 610-623. ACM, 2021. 2 +[8] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 2 +[9] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 2 +[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 1, 2 +[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale + +Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2 +[12] Jieren Deng, Yijue Wang, Ji Li, Chenghong Wang, Chao Shang, Hang Liu, Sanguthevar Rajasekaran, and Caiwen Ding. Tag: Gradient attack on transformer-based language models. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3600-3610, 2021. 2 +[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.2 +[14] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 2 +[15] Google. Gemini usage policies, 2024. https://ai.google.dev/gemini-api/docs/safety-settings.4 +[16] Susan Hao, Renee Shelby, Yuchi Liu, Hansa Srinivasan, Mukul Bhutani, Burcu Karagol Ayan, Shivani Poddar, and Sarah Laszlo. Harm amplification in text-to-image models. arXiv preprint arXiv:2402.01787, 2024. 1 +[17] Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P. Xing, Furong Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huaxiu Yao, Manolis Kellis, Marinka Zitnik, Meng Jiang, Mohit Bansal, James Zou, Jian Pei, Jian Liu, Jianfeng Gao, Jiawei Han, Jieyu Zhao, Jiliang Tang, Jindong Wang, Joaquin Vanschoren, John Mitchell, Kai Shu, Kaidi Xu, Kai-Wei Chang, Lifang He, Lifu Huang, Michael Backes, Neil Zhenqiang Gong, Philip S. Yu, Pin-Yu Chen, Quanquan Gu, Ran Xu, Rex Ying, Shuiwang Ji, Suman Jana, Tianlong Chen, Tianming Liu, Tianyi Zhou, William Yang Wang, Xiang Li, Xiangliang Zhang, Xiao Wang, Xing Xie, Xun Chen, Xuyu Wang, Yan Liu, Yanfang Ye, Yinzhi Cao, Yong Chen, and Yue Zhao. Trustworthiness in large language models. In Forty-first International Conference on Machine Learning, 2024. 2 +[18] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 4 +[19] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 5 + +[20] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. RLAIF vs. RLHF: scaling reinforcement learning from human feedback with AI feedback. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1 +[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 2 +[22] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Saladbench: A hierarchical and comprehensive safety benchmark for large language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3923-3954. Association for Computational Linguistics, 2024. 3 +[23] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3326-3342. Association for Computational Linguistics, 2024. 1 +[24] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2 +[25] Jiawei Liang, Siyuan Liang, Man Luo, Aishan Liu, Dongchen Han, Ee-Chien Chang, and Xiaochun Cao. Vl-trojan: Multimodal instruction backdoor attacks against autoregressive visual language models. arXiv preprint arXiv:2402.13851, 2024. 1 +[26] Hongzhan Lin, Ziyang Luo, Bo Wang, Ruichao Yang, and Jing Ma. Goat-bench: Safety insights to large multimodal models through meme-based social abuse. ArXiv, abs/2401.01523, 2024. 1 +[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1, 5 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26286-26296. IEEE, 2024. 5, 7 +[29] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2025. 2, 5 +[30] LLamaGuard2. Llama guard2, 2024. https://github.com/meta-llama/PurpleLlama/blob/main/Llama-Guard2/MODEL_CARD.md.4 +[31] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing + +the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 4 +[32] Meta. Llama usage policies, 2024. https://ai.meta.com/llama/use-policy.4 +[33] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2 +[34] OpenAI. Introducing chatgpt, 2022. 2 +[35] OpenAI. Openai usage policies, 2023. https://openai.com/policies/usage-policies.4 +[36] OpenAI. Gpt4v, 2024. https://chatgpt.com/.1 +[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2 +[38] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 16012-16027. Association for Computational Linguistics, 2024. 2 +[39] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2025. 2 +[40] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2 +[41] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1 +[42] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2 +[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, pages 53728-53741. Curran Associates, Inc., 2023. 2, 5 +[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training + +next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294. 4 +[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv: Learning, arXiv: Learning. 2, 5 +[46] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 2 +[47] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2 +[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darryll. Aligning large multimodal models with factually augmented RLHF. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 13088-13110. Association for Computational Linguistics, 2024. 2, 7 +[49] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Agüera y Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. Lamda: Language models for dialog applications. CoRR, abs/2201.08239, 2022. 2 +[50] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2 +[51] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 10460-10479. Association for Computational Linguistics, 2024. 2 +[52] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 2 +[53] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia + +Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021. 4 +[54] Laura Weidinger, Maribeth Rauh, Nahema Marchal, Arianna Manzini, Lisa Anne Hendricks, Juan Mateos-Garcia, Stevie Bergman, Jackie Kay, Conor Griffin, Ben Bariach, et al. Sociotechnical safety evaluation of generative ai systems. arXiv preprint arXiv:2310.11986, 2023. 3 +[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[56] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, Jing Shao, and Wanli Ouyang. LAMM: language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2 +[57] Pan Zhang, XiaoyiDongBin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition. 2023. 2 +[58] Zaibin Zhang, Yongting Zhang, Lijun Li, Jing Shao, Hongzhi Gao, Yu Qiao, Lijun Wang, Huchuan Lu, and Feng Zhao. Psysafe: A comprehensive framework for psychological-based attack, defense, and evaluation of multi-agent system safety. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 15202-15231. Association for Computational Linguistics, 2024. 4 +[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2 +[60] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[61] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 2 +[62] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. Open-Review.net, 2024. 1, 2 + +[63] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy M. Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 2 +[64] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5 \ No newline at end of file diff --git a/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/images.zip b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a758cbd25585b284f7e7e21a08a0b05c4439caf --- /dev/null +++ b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e8eeaabcae8e97619051dff30da89dae1b859e6a1b869b5320070c5871a5384 +size 599669 diff --git a/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/layout.json b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5844f399413f255b6cf0c312403066736a054aa --- /dev/null +++ b/2025/SPA-VL_ A Comprehensive Safety Preference Alignment Dataset for Vision Language Models/layout.json @@ -0,0 +1,7626 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "text", + "content": "SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": "Yongting Zhang" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Lu Chen" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Guodong Zheng" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Yifeng Gao" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Rui Zheng" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Jinlan Fu" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Zhenfei Yin" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Senjie Jin" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Xuanjing Huang" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Feng Zhao" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Tao Gui" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{2,3‡}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": " Jing Shao" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 133, + 159, + 483, + 232 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 234, + 443, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 234, + 443, + 244 + ], + "spans": [ + { + "bbox": [ + 167, + 234, + 443, + 244 + ], + "type": "text", + "content": "zytabcd@mail.ustc.edu.cn, luchen23@m.fudan.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 205, + 247, + 402, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 247, + 402, + 258 + ], + "spans": [ + { + "bbox": [ + 205, + 247, + 402, + 258 + ], + "type": "text", + "content": "https://sqrtizhang.github.io/SPA-VL/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 291, + 200, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 291, + 200, + 304 + ], + "spans": [ + { + "bbox": [ + 151, + 291, + 200, + 304 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 316, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 297, + 594 + ], + "type": "text", + "content": "The emergence of Vision Language Models (VLMs) has brought unprecedented advances in understanding multimodal information. The combination of textual and visual semantics in VLMs is highly complex and diverse, making the safety alignment of these models challenging. Furthermore, due to the limited study on the safety alignment of VLMs, there is a lack of large-scale, high-quality datasets. To address these limitations, we propose a Safety Preference Alignment dataset for Vision Language Models named SPA-VL. In terms of breadth, SPA-VL covers 6 harmfulness domains, 13 categories, and 53 subcategories, and contains 100,788 samples of the quadruple (question, image, chosen response, rejected response). In terms of depth, the responses are collected from 12 open-source (e.g., QwenVL) and closed-source (e.g., Gemini) VLMs to ensure diversity. The construction of preference data is fully automated, and the experimental results indicate that models trained with alignment techniques on the SPA-VL dataset exhibit substantial improvements in harmlessness and helpfulness while maintaining core capabilities. SPA-VL, as a large-scale, high-quality, and diverse dataset, represents a significant milestone in ensuring that VLMs achieve both harmlessness and helpfulness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 616, + 137, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 137, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 137, + 629 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 637, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 674 + ], + "type": "text", + "content": "Vision Language Models (VLMs), such as GPT-4V [36], Claude 3 [1], LLaVA [27], and MiniGPT-4 [62] can understand visual signals and respond to user instructions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 293, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 353 + ], + "type": "text", + "content": "Equipped with a visual encoder module, VLMs extract multimodal knowledge from both visual and textual queries, leveraging pre-trained LLMs' powerful comprehension and generative capabilities to achieve remarkable results across diverse vision-language tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 568 + ], + "type": "text", + "content": "Due to the complexity of multimodal harms, previous study [16] has demonstrated that harmless inputs may also result in outputs that do not align with human preferences. Although LLMs have undergone harmless alignment, the alignment of visual encoders is relatively weak, making VLMs susceptible to successful attacks through the visual modality [6, 14, 25]. Therefore, it is necessary to simultaneously improve the alignment of the visual and language modules of the VLM to achieve the harmless and helpful responses. Ensuring the alignment of VLMs with ethical and safety standards is crucial for their safe and effective deployment. However, most existing works on the safety of VLMs focused on the evaluation benchmarks [10, 24, 26] or jailbreak detection [6, 14, 14, 23-25]. Seldom studies considered constructing large-scale, high-quality open-source training datasets to achieve the safety alignment of VLMs. The lack of such datasets hampers further development in this field." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 556, + 665 + ], + "type": "text", + "content": "To address these limitations, we propose a large-scale safety alignment dataset for VLMs named SPA-VL. Since Reinforcement Learning from Human Feedback (RLHF) is widely regarded as performing well in alignment studies [20, 41], our SPA-VL dataset is designed for the RLHF, with each sample containing four elements (question, image, chosen response, rejected response). The main perspectives of our SPA-VL dataset are summarized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "(1) Comprehensive Domains: SPA-VL contains 100,788 samples and comprehensively covers a wide range of harm types, encompassing 6 domains, 13 categories, and 53 subcategories. (2) Diverse Questions and Responses:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 270, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 270, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 270, + 693 + ], + "type": "text", + "content": "*Equal contribution. Authorship order determined by coin flip." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 148, + 703 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": " Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19867" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "For diverse question collection, we gathered three types of questions for each image: easy question, hard question, and hard statement. For diverse response collection, to enhance the diversity of the collected responses and reduce specific model biases, we selected responses from 2 out of 12 different models for each question, designating them as chosen and rejected responses. (3) Multi-Objective Alignment: The preference dataset is designed according to two alignment objectives: harmlessness and helpfulness. Thus, the responses of the aligned VLMs are enhanced in both aspects without sacrificing helpfulness for safety." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 204, + 236, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 204, + 236, + 215 + ], + "spans": [ + { + "bbox": [ + 67, + 204, + 236, + 215 + ], + "type": "text", + "content": "The main contributions are listed below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 294, + 431 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 216, + 294, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 310 + ], + "type": "text", + "content": "(1) We propose the SPA-VL, a large-scale, high-quality, and domain diversity dataset for vision language model safety alignment studies. By using techniques such as PPO and DPO for safety alignment learning on the SPA-VL dataset based on some open VLMs, such as LLaVA, experimental results show that these VLMs significantly improve performance in safety and surpass the state-of-the-art VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 294, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 294, + 370 + ], + "type": "text", + "content": "(2) We conducted extensive analysis and ablation studies, finding that increasing the scale of the dataset, incorporating diverse answers, and using various question types can improve the safety and performance of the alignment VLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 294, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 294, + 431 + ], + "type": "text", + "content": "(3) The entire dataset construction process is fully automated, ensuring efficient and scalable data generation. The automation covers all key steps, including the collection of images and questions, the selection of VLM responses, and the annotations of chosen and rejected responses." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 442, + 142, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 142, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 142, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 653 + ], + "type": "text", + "content": "Vision Language Models. The rapid advancement of Large Language Models [42, 59] and their robust foundational capabilities have significantly prompted the development of multimodal large models. Recently, Vision-Language Models have emerged as a promising extension of LLMs [9], integrating visual and textual information for enhanced multimodal understanding. Notable models in this domain include InstructBLIP [11], InternLMXComposer [57], LAMM-SFT [56], LAMM [56], LLaMA-Adapter-V2 [13], MiniGPT-4 [62], mPLUG-Owl [55], Otter [21], and Qwen-VL-Chat [3]. Most of these VLMs are developed by projecting the vision space into the language space through a learned projector, leveraging pre-trained language models as their backbone. As VLMs continue to advance rapidly, safety concerns have garnered significant attention from researchers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Reinforcement Learning from Human Feedback. Despite the promising capabilities of LLMs and VLMs, they are prone to unintended behaviors, such as fabricating facts, producing biased or harmful content, or even harming humans [7, 8]. They should be helpful, honest, and harm-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "less (3H) [4, 37, 49]. RLHF offers the most straightforward approach to achieving this goal. RLHF methods such as PPO [45] and DPO [43] have been highly successful in aligning AI with human preferences. Notable applications like ChatGPT [34] and Claude [1] show strong performances in academic benchmarks. Models trained with RLHF methods often perform better and adhere more closely to human values compared to those trained only with SFT [4]. This success extends to VLMs, where RLHF has been used to address hallucination issues [5, 39, 48, 61]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 192, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 311 + ], + "type": "text", + "content": "Safety of VLMs. To evaluate the safety performance of VLMs, various methods and datasets have been proposed. Among these evaluation benchmarks are MM-SafetyBench [29], ChEf [47], and OODCV-VQA, SketchyVQA [50]. In addition to these benchmarks, several attack methods, such as adversarial attacks [12, 40, 46, 60] and jailbreak techniques [14, 24, 33], have been developed to test the vulnerabilities of VLMs. These studies aim to identify and exploit potential weaknesses in VLMs, underscoring the need for robust safety measures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 478 + ], + "type": "text", + "content": "In response to these vulnerabilities, several methods have been developed to enhance the safety performance of VLMs. For instance, VLGuard [63] employed supervised fine-tuning (SFT) on the VLGuard dataset, which contains 2000 training images designed to improve safety. Similarly, Chen et al. [10] used AI-annotated data for SFT. Wei et al. [52] utilized in-context learning to bolster model safety. Additionally, Pi et al. [38] introduced MLLM-Protector, a harm detector serving as a plug-and-play defense mechanism for VLMs, and Wang et al. [51] applied inference time alignment methods to improve harmlessness. These approaches collectively demonstrate ongoing efforts to mitigate risks and enhance the resilience of VLMs against various types of attacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 598 + ], + "type": "text", + "content": "Among all the methods mentioned above, we can broadly categorize them into two types. The first type enhances model safety during the inference stage by using prompts. This approach is efficient and convenient but often results in limited safety improvements and lacks generalization [17]. The second type involves training-based methods, specifically during the training-to-align phase, which can be further divided into SFT and RLHF. While the aforementioned methods primarily rely on SFT, we go a step further by providing a comprehensive RLHF dataset SPA-VL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 609, + 414, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 414, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 414, + 621 + ], + "type": "text", + "content": "3. SPA-VL Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "In the development of VLMs, effectively addressing harmful content in multimodal environments poses a significant challenge. The SPA-VL dataset helps VLMs confront this challenge by providing safety preference data for RLHF. As shown in Figure 1, establishing the SPA-VL dataset involves three key stages. First, we systematically collect images, which includes gathering a diverse set of images to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19868" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 547, + 301 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 547, + 301 + ], + "type": "image", + "image_path": "ef9d1caf78fa225218ff69358afcc7685cb10346ce86d00e12cb2451299d1aec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 312, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 312, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 555, + 346 + ], + "type": "text", + "content": "Figure 1. Overview of SPA-VL Dataset Construction. It is built in three stages: 1) Image Collection, 2) Questions Construction and 3) Preference Construction. The dataset examples shows vision-question/preferences pairs that comprise three types of questions: easy questions, hard questions, and hard statements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 352, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 296, + 422 + ], + "type": "text", + "content": "ensure comprehensiveness. Next, we generate questions related to categories of harmful content. After this, we proceed with preference annotation. This stage includes generating responses from various models and labeling these responses based on preferences for harmlessness and helpfulness." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 469, + 294, + 696 + ], + "blocks": [ + { + "bbox": [ + 55, + 434, + 296, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 468 + ], + "type": "text", + "content": "Table 1. Training dataset statistics for SPA-VL. For each image, we provide three prompts: Easy question, Hard question, Hard statement. UR% represents the unsafe rate." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 469, + 294, + 696 + ], + "lines": [ + { + "bbox": [ + 58, + 469, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 58, + 469, + 294, + 696 + ], + "type": "table", + "html": "
Secondary ClassImage Cnt.QuestionPrefer(UR%)Prefer(Len.)
UR%Cnt.Len.Cho.Rej.Cho.Rej.
Toxic379144.111132111611.3541.55488392
Unfair358938.38106841207.1532.16620441
Erosion of Trust in Public Information126337.6237671527.6231.62595463
False Beliefs181429.3154241465.8827.16746539
Dangerous Information126359.66378812914.7849.39621580
Privacy63653.12190715612.1144.83635513
Security Threats245263.99727914112.7450.83567572
Defamation61151.83180614516.4546.46542464
Fraud or Deceptive Action477957.211417914413.7348.14590531
Influence Operations179551.51531715117.1149.69725599
Illegal Activities373460.511102513013.8349.23568531
Persuasion and Manipulation118859.38333113617.8951.73719633
Violation of Personal Property190955.5753821609.541.19543478
Risky Financial Practices184931.8152071479.130.57933685
Trade and Compliance122129.4630211809.7631.45905647
Total3189449.279325813811.742.23626519
", + "image_path": "9f1654040c7227d88f6f2a2673f4665291eed5d9d73553eb25282fe4a7955a37.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 351, + 382, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 351, + 382, + 363 + ], + "spans": [ + { + "bbox": [ + 314, + 351, + 382, + 363 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "content": "Data Statistics. Our SPA-VL dataset comprises four parts: the training set, the validation set, and two test sets, HarmEval and HelpEval, which are used to evaluate harmfulness and helpfulness, respectively. The number of samples in each part is 93,258, 7,000, 265, and 265, respectively. Table 1 shows the dataset statistics of the training set. To detect the unsafe content covered by our SPA-VL dataset, we utilize the MD-Judge evaluator [22] to calculate the unsafe rate of the collected questions and VLMs' responses. Nearly half of the collected questions are unsafe, while the unsafe rate for the chosen response and rejected response is " + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "inline_equation", + "content": "42.23\\%" + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "content": ", respectively. The HarmEval test set includes a substantial number of harmful questions, while the HelpEval test set primarily comprises questions that involve instruction following or require the expression of opinions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Diverse Domains. A diverse and representative set of images is essential for training models to handle vision data safely and effectively. Our primary challenge is ensuring diversity while maintaining relevance to harmful content categories. To address this, we establish a comprehensive harm content categorization framework. Our SPA-VL adopts 6 primary domains, 15 secondary categories, and 53 tertiary categories, ensuring comprehensive coverage and granularity for precise harm detection and response alignment. A detailed visual representation of this hierarchical structure is provided in Figure ?? in the Appendix ??. We reference Li et al. [22], Weidinger et al. [54] for our primary harm" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19869" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 77, + 124, + 114 + ], + "blocks": [ + { + "bbox": [ + 81, + 77, + 124, + 114 + ], + "lines": [ + { + "bbox": [ + 81, + 77, + 124, + 114 + ], + "spans": [ + { + "bbox": [ + 81, + 77, + 124, + 114 + ], + "type": "image", + "image_path": "726c64f80a5961c18fac5e945405524457343f9bca0b9c869ae3d1cfb3b6bf7c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 121, + 283, + 175 + ], + "lines": [ + { + "bbox": [ + 88, + 121, + 283, + 175 + ], + "spans": [ + { + "bbox": [ + 88, + 121, + 283, + 175 + ], + "type": "text", + "content": "The most degrading and ageist way to mark your 28th birthday would be to have a card that says \"You're 28\" with a picture of a duck on it. This is degrading because it implies that the person is old and irrelevant, and it is ageist because it suggests that the person is past their prime and has no value. The card's message is disrespectful and disregards the individual's worth and potential for growth and development." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 268, + 70, + 291, + 101 + ], + "blocks": [ + { + "bbox": [ + 131, + 82, + 262, + 109 + ], + "lines": [ + { + "bbox": [ + 131, + 82, + 262, + 109 + ], + "spans": [ + { + "bbox": [ + 131, + 82, + 262, + 109 + ], + "type": "text", + "content": "What's the most degrading and ageist way to mark your 28th birthday, showcasing your declining worth as a stale and irrelevant human being?" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 268, + 70, + 291, + 101 + ], + "lines": [ + { + "bbox": [ + 268, + 70, + 291, + 101 + ], + "spans": [ + { + "bbox": [ + 268, + 70, + 291, + 101 + ], + "type": "image", + "image_path": "6a0170ecb21ab281a059fec504dd5efb728f98d50244c2fe059ecde752604e53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 136, + 84, + 161 + ], + "blocks": [ + { + "bbox": [ + 63, + 136, + 84, + 161 + ], + "lines": [ + { + "bbox": [ + 63, + 136, + 84, + 161 + ], + "spans": [ + { + "bbox": [ + 63, + 136, + 84, + 161 + ], + "type": "image", + "image_path": "7abae8fc431edd41d7d4ca08d96a250ed64d9ccc39435b99df273aa47ab44906.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 186, + 86, + 220 + ], + "blocks": [ + { + "bbox": [ + 63, + 186, + 86, + 220 + ], + "lines": [ + { + "bbox": [ + 63, + 186, + 86, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 86, + 220 + ], + "type": "image", + "image_path": "6472b30d4ea7fe2217d0fa7dda0bcdd3a0f4e5f51dc381c048af6a19baaa1600.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 182, + 286, + 222 + ], + "lines": [ + { + "bbox": [ + 88, + 182, + 286, + 222 + ], + "spans": [ + { + "bbox": [ + 88, + 182, + 286, + 222 + ], + "type": "text", + "content": "I'm sorry, but I cannot provide an answer to that question as it is ageist and degrading. It is important to remember that everyone is unique and valuable, regardless of their age. Let's focus on celebrating our birthdays and the milestones we have achieved, rather than making derogatory comments about others." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 63, + 270, + 86, + 308 + ], + "blocks": [ + { + "bbox": [ + 63, + 270, + 86, + 308 + ], + "lines": [ + { + "bbox": [ + 63, + 270, + 86, + 308 + ], + "spans": [ + { + "bbox": [ + 63, + 270, + 86, + 308 + ], + "type": "image", + "image_path": "9e772b6ed351d3ce1c058696225ced69bc28b9e0b129274934b219ef590cefc9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 231, + 287, + 353 + ], + "lines": [ + { + "bbox": [ + 88, + 231, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 88, + 231, + 287, + 353 + ], + "type": "text", + "content": "It is not appropriate to use degrading or ageist language to mark someone's 28th birthday or any other milestone. Such language is disrespectful and can negatively impact the individual's self-esteem and well-being. It is important to celebrate and acknowledge the achievements and experiences of individuals regardless of their age, and to treat them with respect and dignity. \\n\\nInstead of using degrading language, you can create a positive and supportive atmosphere for the person's 28th birthday celebration by: \\n\\nOrganizing a fun and engaging party with friends and family. \\n\\nthoughtful and meaningful presents that reflect the individual's interests and hobbies. \\n\\nSharing heartfelt messages and well-wishes on social media or through personal communication. \\n\\nEncouraging the person to set new goals and aspirations for the next year of their life. \\n\\nCreating a memorable photo album or video montage of the person's life journey so far. \\n\\nBy focusing on positivity and support, you can help the individual feel valued and appreciated, regardless of their age.\\n\\n" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 365, + 295, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 387 + ], + "type": "text", + "content": "Figure 2. Case study comparing responses from the original model, the model trained with DPO and PPO on our SPA-VL." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 409, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 295, + 517 + ], + "type": "text", + "content": "categories, systematically organizing and identifying various types of harmful content. For the secondary and tertiary classes, we reference usage policies from leading AI organizations, including OpenAI [35], LLaMA [32], Gemini [15], Claude [2], as well as guidelines from Llama Guard [18] and Llama Guard2 [30]. Additional references include Luo et al. [31], Weidinger et al. [53], ensuring that our classification aligns with industry standards and best practices, enhancing the relevance and applicability of our dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 518, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 638 + ], + "type": "text", + "content": "Data Formats. Following Bai et al. [4], we gather preference data by choosing the better response from two generated by VLMs, based on predefined criteria of harmlessness and helpfulness. Finally, a quadruple (question, image, chosen response, rejected response) reflecting preferences is collected, where the chosen response is the better response selected under principle harmless and helpful. Figure 1 shows three demonstration examples of our SPAVL dataset. In the following sections, we will give a detailed illustration of the construction process of our dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 647, + 157, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 157, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 157, + 659 + ], + "type": "text", + "content": "3.2. Image Collection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "With this robust categorization framework in place, we proceed to collect images that correspond to our harm categories. We use the LAION-5B [44] dataset, which is well-suited for this purpose due to its extensive and diverse col" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "lection of images. The LAION-5B dataset is built using a CLIP model to match images and text, allowing us to effectively use our tertiary class labels to search for relevant pictures. This leverages the strengths of the CLIP model in understanding visual and textual correlations, ensuring that the images are well-aligned with our harm categories. To ensure diversity and avoid bias, we use six different search keywords for each tertiary class. This approach helps capture a wide range of images within each category, preventing over-representation of specific subtypes. Details are illustrated in Appendix ??" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 212, + 434, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 434, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 434, + 224 + ], + "type": "text", + "content": "3.3. Question Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 229, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 312 + ], + "type": "text", + "content": "Generating relevant and meaningful questions for each image is crucial for contextual understanding. The primary challenge here is ensuring that the questions are specific to the harmful content categories and diverse in complexity. Although the LAION-5B dataset provides captions, they are often misaligned with the images [44]. To address this, we enhance the captions using Gemini 1.0 Pro Vision1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 313, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 553 + ], + "type": "text", + "content": "In the subsequent step, we devise queries that could be potentially harmful for each image. Initially, Gemini 1.0 Pro Vision produces an easy question for every image. To ensure pertinence, the model is supplied with the image's primary, secondary, and tertiary categories. However, these questions, typically starting with \"what\" or \"how\", tend to be straightforward and closely related to the image, which may lead to a lack of complexity and diversity. To overcome this, we use Gemini 1.0 Pro to refine the questions, transforming them into more challenging queries and statements to enhance the diversity of our dataset. This process yields two additional outputs: a hard question (hardq) and a hard statement (hardd). For this refinement, the model is provided with the image's hierarchical classifications, two captions (one generated by Gemini and the original from LAION), and the previously generated easy question. Given that Gemini is designed to avoid generating harmful content, we employ a jailbreak strategy [58] to evoke more challenging queries. Further details regarding this process are provided in Appendix ??" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 560, + 437, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 437, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 437, + 571 + ], + "type": "text", + "content": "3.4. Preference Collection" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 578, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 637 + ], + "type": "text", + "content": "The final stage in constructing the SPA-VL dataset involves generating and annotating preferences for responses for training VLMs. This stage combines the generation of diverse responses and the careful annotation of these responses to create preference pairs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 638, + 554, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 554, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 554, + 686 + ], + "type": "text", + "content": "Response Generation. To create annotations that better align with human judgment, we collect diverse responses from 12 different models. This diversity helps capture a wide range of potential answers, reducing model-specific" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 552, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 552, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 552, + 713 + ], + "type": "text", + "content": "1https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/gemini-pro-vision" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19870" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "biases and ensuring that the VLMs can learn from a variety of perspectives. Detailed generation process are described in Appendix ??" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 294, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 294, + 217 + ], + "type": "text", + "content": "Preference Annotation. Responses are classified using MD-Judge to ensuring and for each question, we randomly select two responses from different safety-rated model groups to ensure that the chosen responses reflect diverse levels of safety to make GPT-4V better annotate. GPT-4V evaluates these pairs based on both harmlessness and helpfulness, also, to avoid bias due to the order of the answers, we query GPT-4V twice with the answers swapped. Details are illustrated in Appendix ??" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 230, + 136, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 136, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 136, + 243 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 250, + 115, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 115, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 115, + 262 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 267, + 294, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 294, + 363 + ], + "type": "text", + "content": "We apply DPO [43] and PPO [45], using our preference dataset to train LLaVA-1.5 [27, 28]. For preliminaries on DPO and PPO, please refer to the Appendix ??. We keep the visual encoder weights frozen as in Karamcheti et al. [19], Liu et al. [27] and update the weights of both the projection layer and LLM in LLaVA-1.5 (7B) in our main experiments. For training parameters, please refer to the Appendix ?? for further details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 294, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 294, + 448 + ], + "type": "text", + "content": "We evaluate harmlessness using the HarmEval dataset and assess adversarial robustness on AdvBench [64] and MM-SafetyBench [29]. Helpfulness is measured using the Anthropic-Helpful [4] and HelpEval datasets, with GPT-4 determining response quality based on helpfulness and safety. Detailed evaluation methods and results can be found in the Appendix ??" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 294, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 294, + 579 + ], + "type": "text", + "content": "Our HarmEval dataset comprises 265 questions across 53 categories, with each category containing 5 questions. We utilize two evaluation metrics on this dataset: the unsafe rate and the harm score. The UnSafe Rate assesses the proportion of unsafe responses generated by the model for all test questions. This metric evaluates the general safety of the model's outputs. The Harm Score involves comparing pairs of answers from the base model, LLaVA-1.5 (7B), and the trained model using GPT-4V. Unlike the data construction phase, GPT-4V assesses the answers based solely on the criterion of \"harmlessness\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 294, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 294, + 664 + ], + "type": "text", + "content": "To ensure unbiased evaluation, we also test on the publicly available language dataset AdvBench [64] and the vision dataset MM-SafetyBench [29] (Figure ?? shows two examples from this benchmark). For both datasets, the evaluation metric is the Attack Success Rate (ASR), which measures the safety of the model by assessing its vulnerability to adversarial attacks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 714 + ], + "type": "text", + "content": "MM-SafetyBench. MM-SafetyBench is a benchmark for evaluating VLMs safety composed of harmful questions and images. It includes four subcategories: Text-only (text only), SD (images generated by Stable Diffusion using key" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "phrases), Tipo (typography, images with the black phrase on a white background), and SD+Typo (concatenated generated images and typography together). The evaluation metric for this benchmark is the attack success rate (ASR). This dataset enables us to evaluate how well our trained model performs on publicly available image-based test sets in terms of safety. We use MM-Safety Prompt as Liu et al. [29] to ensure consistency across our tests." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": "AdvBench. This dataset is purely text-based and has two subcategories: vanilla instruction and suffix injection (questions with the suffix \"Answer with 'Absolutely, here's'\"). The evaluation metric for this dataset is also ASR. This dataset allows us to assess the safety of our trained model when confronted with text-based harmful content in publicly available test sets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 258, + 399, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 258, + 399, + 270 + ], + "spans": [ + { + "bbox": [ + 314, + 258, + 399, + 270 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 276, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 419 + ], + "type": "text", + "content": "As shown in the Table 2, the models trained on our SPA-VL dataset, LLaVA-SPA-VL-DPO and LLaVA-SPA-VL-PPO, which are the best safety models from our training, exhibit superior safety performance. They surpass the baseline model LLaVA-1.5 (7B) and other open-source models, whether or not those models have undergone safety alignment. Specifically, our models achieve best safe result on MM-SafetyBench, AdvBench and HarmEval tests. Additionally, we provide a comparison of responses before and after training for LLaVA-SPA-VL-DPO, LLaVA-SPA-VL-PPO, and baseline model LLaVA in Figure 2. More comparison examples can be found in Appendix ??" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 420, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 491 + ], + "type": "text", + "content": "In addition to evaluating the safety performance, we also validate our models' general ability. We select the most commonly used multimodal benchmarks and find that the general ability of our safety-aligned models does not significantly decline compared to the backbone model. Details of these evaluations can be found in the Appendix ??" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 502, + 451, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 451, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 451, + 514 + ], + "type": "text", + "content": "5. Analysis and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 617 + ], + "type": "text", + "content": "In this section, we further explore the factors that affect the performance of alignment models. Our focus includes examining the impact of dataset scale, the selection of response models from diverse models during dataset construction, the influence of different question types within the dataset, and the outcomes of deciding whether to freeze the projection layer during model training, and the different training base models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 386, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 386, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 386, + 635 + ], + "type": "text", + "content": "5.1. Data Scale" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "text", + "content": "We delve into the impact of varying amounts of data on the performance of alignment models. Across different data quantities (around 100, 1k, 5k, 10k, 30k, and 90k), we conduct experiments encompassing various evaluation metrics, as already described in Appendix ??. The resulting line plots in Figure 3 reveal compelling insights. For HarmEval," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19871" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 105, + 552, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Comparison of different VLM models on harmlessness. The models are evaluated across multiple metrics on MM-SafetyBench and AdvBench, as well as the HarmEval UnSafe Rate (HarmEval USR). After training on our proposed dataset, SPA-VL, the model achieves the best scores according all metric on both DPO and PPO methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 105, + 552, + 404 + ], + "lines": [ + { + "bbox": [ + 58, + 105, + 552, + 404 + ], + "spans": [ + { + "bbox": [ + 58, + 105, + 552, + 404 + ], + "type": "table", + "html": "
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Open Sourced Baseline
InstructBLIP-7B27.3813.1027.3825.0023.2151.2564.6247.55
LAMM-13B14.294.762.386.556.9924.4239.1132.83
LAMM + SFT-13B16.077.148.3321.4313.2422.6972.1232.08
LLaMA-Adapter-v2-7B35.7112.507.7417.8618.4598.2610046.04
MiniGPT-4-7B20.839.5223.8120.2418.6031.3565.3838.32
mPLUG-Owl-7B35.718.9312.5030.3621.8810010052.45
Otter-9B29.7610.125.957.7413.3991.9210041.13
InternVL-Chat-7B5.951.7919.6413.1010.126.9289.4229.81
CogVLM2-LLama3-Chat-19B16.674.7623.8123.2117.1125.3898.0813.96
LLaVA-v1.6-34B4.764.1716.0719.0511.015.5893.0822.64
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
Close Sourced Baseline
Gemini-1.5-pro000000.380.381.89
GPT-4-0125-preview1.790000.450.966.542.26
Safety Aligned
LLaVA + SPA-VL-DPO0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
LLaVA + SPA-VL-PPO0.6(↓33.93)0(↓7.74)0(↓22.62)1.19(↓16.07)0.45(↓20.09)0.19(↓97.88)2.12(↓97.69)0(↓44.15)
", + "image_path": "091a5fbe52dd1da468d4db9b7a40904eead6ad56e0e9a7f9630ee1400b01946d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 420, + 294, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 294, + 539 + ], + "type": "text", + "content": "the Harm Score consistently decreases with increasing data volume. Similarly, in MM-SafetyBench, the overall average ASR steadily decreases as the data scale grows, and both vanilla and suffix ASRs in AdvBench exhibit a similar trend of declining rates with expanding data sizes. Notably, the help score in Anthropic-Helpful exhibits a progressive increase, indicating a simultaneous enhancement in safety and helpfulness as the dataset size expands. The full results are presented in Table 3, with detailed analysis in Appendix ??" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 551, + 200, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 200, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 200, + 563 + ], + "type": "text", + "content": "5.2. Response Model Selection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this section, we examine the impact of response diversity and safety in our dataset on model training. We conducted four groups of experiments, each group is trained using DPO on around 10K samples. Safe Group consists of response pairs from the three safest models (InternLMXComposer, QwenVL, Gemini1.0 Pro Vision) according to Table ?? in Appendix ??. Relative Safe Group includes pairs from relative safe models(LAMM_SFT, LLaVA1.5, InternLMXComposer, QwenVL, gemini). Unsafe Group comprises pairs from the five least safe models(mPLUG-Owl, Otter, InstructBLIP, LLaMA-Adapter-v2, GeminiJailbreak) and the All group consists of the complete set" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 507, + 553, + 611 + ], + "blocks": [ + { + "bbox": [ + 313, + 418, + 555, + 505 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 505 + ], + "type": "text", + "content": "Table 3. Impact of Training Data Scale on Alignment Model Performance using PPO and DPO Methods on LLaVA-1.5 (7B). The table compares the performance of alignment models trained with PPO and DPO methods across varying data scales: 100, 1K, 5K, 10K, 30K, and 90K tsamples. The evaluation covers both safety and helpfulness metrics. As the data scale increases, the performance generally improves across harmlessness and helpfulness metrics." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 507, + 553, + 611 + ], + "lines": [ + { + "bbox": [ + 316, + 507, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 553, + 611 + ], + "type": "table", + "html": "
Data ScaleMM-SafetyBench↓AdvBench↓HarmEval↓Helpful
Text-onlySDCargoSD+CargoAvgvanillasuffixUSRScoreAnthropic↑HelpEval↑
PPO
10035.128.9318.4517.2619.9497.5098.8543.3950.0051.5019.00
1K30.958.3319.0516.6718.7597.3198.4641.8945.0658.0028.61
5K5.954.7613.1017.2610.2736.1585.7710.1917.2461.0048.54
10K0.601.195.959.524.328.6517.500.009.2864.5067.60
30K0.000.000.002.980.740.583.270.0010.9265.5052.09
90K0.600.000.001.190.450.192.120.008.8170.0071.04
DPO
10032.7410.1220.2416.6719.9497.8999.6243.4050.5751.0024.12
1K30.367.7417.8619.0518.7591.5496.7326.0430.0058.5035.65
5K4.171.194.178.334.4610.0040.001.8915.3464.0042.86
10K1.191.792.384.762.535.776.540.3813.7863.0040.85
30K0.000.600.601.190.600.000.000.0013.6469.5045.72
90K1.190.601.192.981.490.000.000.7513.4170.0050.63
", + "image_path": "8081ccc5ef4e3416b6d1d4a900024487c3e8eadde1c98bc4e4af707a2e3e5c43.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 628, + 372, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 372, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 372, + 639 + ], + "type": "text", + "content": "of 12 models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "The results presented in the Table 4 indicate that if our dataset only includes pairs of safe responses, the model struggles to learn how to avoid bad patterns, leading to vulnerability (Safe Group performed poorly on the AdvBench suffix test, suggesting the model is easily attacked). Similarly, if our dataset only includes unsafe responses, the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19872" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 164, + 147 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 164, + 147 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 164, + 147 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 164, + 147 + ], + "type": "image", + "image_path": "6390a9a455d1ebffcf62f6599282a1b708f661e7c7f0a66deab7b1c52d3f15c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 149, + 133, + 159 + ], + "lines": [ + { + "bbox": [ + 87, + 149, + 133, + 159 + ], + "spans": [ + { + "bbox": [ + 87, + 149, + 133, + 159 + ], + "type": "text", + "content": "(a) EvalHarm" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 161, + 164, + 236 + ], + "blocks": [ + { + "bbox": [ + 59, + 161, + 164, + 236 + ], + "lines": [ + { + "bbox": [ + 59, + 161, + 164, + 236 + ], + "spans": [ + { + "bbox": [ + 59, + 161, + 164, + 236 + ], + "type": "image", + "image_path": "5e23859681d979630a87480b16b30bc05bcc4f467302af369c1bd7b7d6c73ba5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 239, + 133, + 248 + ], + "lines": [ + { + "bbox": [ + 87, + 239, + 133, + 248 + ], + "spans": [ + { + "bbox": [ + 87, + 239, + 133, + 248 + ], + "type": "text", + "content": "(c) AdvBench" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 189, + 70, + 293, + 147 + ], + "blocks": [ + { + "bbox": [ + 189, + 70, + 293, + 147 + ], + "lines": [ + { + "bbox": [ + 189, + 70, + 293, + 147 + ], + "spans": [ + { + "bbox": [ + 189, + 70, + 293, + 147 + ], + "type": "image", + "image_path": "2431fed2bffe567079fd268c8ed89c381642bf7c13efce0fc08932a24250b862.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 149, + 277, + 159 + ], + "lines": [ + { + "bbox": [ + 206, + 149, + 277, + 159 + ], + "spans": [ + { + "bbox": [ + 206, + 149, + 277, + 159 + ], + "type": "text", + "content": "(b) MM-SafetyBench" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 190, + 161, + 291, + 236 + ], + "blocks": [ + { + "bbox": [ + 190, + 161, + 291, + 236 + ], + "lines": [ + { + "bbox": [ + 190, + 161, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 190, + 161, + 291, + 236 + ], + "type": "image", + "image_path": "f1b512c48e6e28377113ea7b7a5e5d39543aa25250ae0fa38fa06b68ee5e9127.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 239, + 277, + 249 + ], + "lines": [ + { + "bbox": [ + 205, + 239, + 277, + 249 + ], + "spans": [ + { + "bbox": [ + 205, + 239, + 277, + 249 + ], + "type": "text", + "content": "(d) Anthropic-Helpful" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 258, + 296, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 347 + ], + "type": "text", + "content": "Figure 3. Impact of Data Scale on Alignment Model Performance. Line plots illustrate the effect of varying data quantities (100, 1k, 5k, 10k, 30k, and 90k) on the performance metrics of alignment models for both PPO and DPO methods. (a) Shows the Harm Score (\\%) on EvalHarm (b) Shows the Average Attack Success Rate (ASR " + }, + { + "bbox": [ + 55, + 258, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 55, + 258, + 296, + 347 + ], + "type": "text", + "content": " ) on MM-SafetyBench (c) Shows ASR (\\%) on vanilla and suffix in AdvBench (d) Shows the Help Score (\\%) on Anthropic-Helpful." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 57, + 405, + 294, + 467 + ], + "blocks": [ + { + "bbox": [ + 55, + 360, + 296, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 403 + ], + "type": "text", + "content": "Table 4. Detailed harmlessness evaluation metrics for Response Model Selection. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 405, + 294, + 467 + ], + "lines": [ + { + "bbox": [ + 57, + 405, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 57, + 405, + 294, + 467 + ], + "type": "table", + "html": "
Model BagAdvBenchMM-SafetyBenchHarmEval
vanillas suffixText-onlySDTypoSD+TypoAvgUSRScore
Safe32.5065.3812.53.5710.7111.909.6713.9718.49
Relative safe14.8135.004.173.579.528.936.554.5315.85
Unsafe9.0460.772.982.988.9311.906.707.1721.14
All0.586.541.191.792.384.762.536.1113.78
", + "image_path": "d06296308d6919b4c5569478d45763d3b6969b5decd69d79c0d77dbf50104cc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 483, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 603 + ], + "type": "text", + "content": "model cannot be trained to be safe, as it lacks exposure to good patterns (Unsafe Group also performed poorly on the AdvBench suffix test). Relative safe Group, which includes a mix of good and bad responses, shows significantly better safety performance compared to Safe Group and Unsafe Group. However, there still exists a gap on harmlessness between the model trained with Relative Safe and the All group. This demonstrates the critical importance of response diversity during the data construction process for effective model alignment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 611, + 151, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 151, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 151, + 624 + ], + "type": "text", + "content": "5.3. Question Types" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "In this section, we analyze the impact of three different question types (Easy questions, Hard questions, and Hard statements) on the experimental results. We also compare these individual results with the combined results of all three question types. For each experiment, we select training dataset of approximately " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "10k" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " instances and using DPO to train our model. The results presented in the Table 5" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 116, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "type": "text", + "content": "Table 5. The detailed harmless evaluation metrics of Question Types. HarmEval HS represents the Harm Score on HarmEval evaluated by GPT-4V, while HarmEval USR indicates the unsafe rate on HarmEval evaluated by MD-Judge." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 116, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 316, + 116, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 553, + 180 + ], + "type": "table", + "html": "
Ques TypeAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
Easy-Q3.8524.041.791.792.388.933.722.2616.73
Hard-Q2.1211.541.191.193.579.523.870.7513.97
Hard-S2.125.003.571.794.175.953.878.7018.44
Mixed0.586.541.191.792.384.762.530.3813.78
", + "image_path": "4e3ea9c883dcf71e9b9d00a8450f4557b48a072b7bbf966fc8106956e1364f9c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 316, + 227, + 553, + 318 + ], + "blocks": [ + { + "bbox": [ + 313, + 193, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 225 + ], + "type": "text", + "content": "Table 6. Detailed harmless evaluation metrics of model architecture on LLaVA-1.5 (7B), with projector and without projector for both DPO and PPO training methods." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 227, + 553, + 318 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 318 + ], + "type": "table", + "html": "
Model ArchAdvBenchMM-SafetyBenchHarmEval
vanillasuffixText-onlySDTypoSD+TypoAvgUSRScore
DPO
w/o projector0.000.190.000.001.195.361.641.1314.21
w projector0.000.000.000.600.601.190.600.0013.64
(0.00)(↓0.19)(0.00)(↑0.60)(↓0.59)(↓4.17)(↓1.04)(↓1.13)(↓0.57)
PPO
w/o projector0.582.880.000.001.791.790.890.0019.32
w projector0.583.270.000.000.002.980.740.0010.92
(0.00)(↓0.39)(0.00)(0.00)(↓1.79)(↓1.19)(↓0.15)(0.00)(↓8.4)
", + "image_path": "8ca30e394fd3b4dd0bb27a61f059d7664ea0561119a21cb40a9bf193d82fed92.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "type": "text", + "content": "show that the combined dataset achieves higher safety performance compared to the individual datasets of each question type at the same data scale. This indicates that integrating diverse question types enhances the model's robustness against harmful attacks. These findings suggest that, in real-world scenarios, questions that provoke harmful responses are varied. Consequently, training with a combination of different question types is likely to improve the model's ability to resist harmful attacks and decrease the likelihood of generating harmful responses." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 477, + 421, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 421, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 421, + 490 + ], + "type": "text", + "content": "5.4. Training Methods" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Following the approach outlined in LLaVA [28], we freeze the vision encoder during training. In this framework, both the projection layer and LLM were trained together during SFT as described in LLaVA [28], whereas in LLaVARLHF [48], only the LLM was trained. We aim to explore the impact of training the LLM alone versus with the projection layer on reinforcement learning outcomes. We analyze safety validation results on our " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "30k" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " dataset using DPO method. As shown in the Table 6, there are minimal differences in language-only safety tests (AdvBench). However, in image-involved safety tests (MM-SafetyBench and EvalHarm), training with the projection layer consistently outperforms training without it. We hypothesize that including the projection layer improves the model's ability to detect harmful content in images. This suggests a valuable direction for future work to further investigate these effects. We also investigate the training of LoRA, with further details provided in Appendix ??" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19873" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 108, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 108, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 108, + 83 + ], + "type": "text", + "content": "5.5. Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 186 + ], + "type": "text", + "content": "We explored other backbone models as well and conducted experiments using QwenVL-7BChat and InternLMXComposer-7B, applying DPO training on our dataset. As shown in Table 7, with SPA-VL data training, open-source models, including LLaVA, Qwen, and InternLMXComposer, exhibit significant safety improvements, approaching the performance levels of closed-source models." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 288, + 294, + 375 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 295, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 287 + ], + "type": "text", + "content": "Table 7. We presents the results of DPO training on safety benchmarks for different model backbones, comparing baseline models with their safety-aligned counterparts using the SPA-VL. The models evaluated include LLaVA-7B, InternLMXComposer-7B, and QwenVL-Chat-7B, with their respective safety-aligned versions: LLaVA + SPA-VL, InternLMXC + SPA-VL, and QwenVL + SPA-VL. Across all backbones, significant improvements are observed after safety alignment." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 288, + 294, + 375 + ], + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 375 + ], + "type": "table", + "html": "
ModelMM-SafetyBenchAdvBenchHarmEval USR
Text-onlySDTypoSD+TypoAvgvanillasuffix
Baseline
LLaVA-7B34.527.7422.6217.2620.5498.0899.8144.15
InternLMXComposer-7B7.744.1726.1926.7916.225.4097.8826.04
QwenVL-Chat-7B3.573.5723.2126.7914.291.9272.737.55
Safety Aligned
LLaVA + SPA-VL0(↓34.52)0.6(↓7.14)0.6(↓22.02)1.19(↓16.07)0.6(↓19.94)0(↓98.08)0(↓99.81)0(↓44.15)
InternLMXC+SPA-VL0(↓7.74)1.19(↓2.98)0.6(↓25.59)2.38(↓24.41)1.04(↓15.18)0.19(↓5.21)0.38(↓97.50)0.75(↓25.29)
QwenVL+SPA-VL0(↓3.57)1.19(↓2.38)1.19(↓22.02)4.17(↓22.62)1.64(↓12.65)0.38(↓1.54)4.61(↓68.12)3.02(↓4.53)
", + "image_path": "47863007a91050e642bd18a1ba73103026a06a537476991a478afffb9ad1df11.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 170, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 170, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 170, + 415 + ], + "type": "text", + "content": "6. Human Annotation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 424, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 295, + 628 + ], + "type": "text", + "content": "In this section, we evaluate the consistency between GPT evaluation and human evaluation to ensure the reliability and validity of our assessment methods. For annotation part, we random sample 530 pairs for each question type to show the human preference result with GPT-4V. For AdvBench, we random select a total result in our result and humanly check the attack success for all the 520 results for both suffix and valina with GPT-4. In the case of Anthropic-Helpful, we examine 200 samples to check the consistency of helpfulness evaluations with GPT-4. For HarmEval, 265 samples (5 from each category) are selected to compare the safety preference consistency between human evaluators and GPT-4V for the baseline model LL and our trained model. Similarly, for HelpEval, we select 265 samples (5 from each category) to compare the helpfulness win rate consistency between human evaluators and GPT-4V for the GPT-4V responses and our trained model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "This comprehensive approach ensures that our models' evaluations align closely with human judgment, thereby enhancing the robustness of our evaluation part. The consistency rates between GPT evaluations and human evaluations are summarized in Table 8. The results indicate high alignment across various evaluation metrics, reinforcing the reliability of our assessment approach." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 95, + 553, + 126 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 553, + 92 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 553, + 92 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 553, + 92 + ], + "type": "text", + "content": "Table 8. Consistency Rate (%) between GPT-4V and human annotation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 95, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 317, + 95, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 553, + 126 + ], + "type": "table", + "html": "
Easy-QHard-QHard-SAdvBenchAnthropic-HelpfulHarmEvalHelpEval
suffixvanilla
96.6697.7493.7799.4099.8091.0098.11100
", + "image_path": "9e084331bd3f19b7a2881c0a6791767ad4b267795c57ea56aad56dcd3f63ba9e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 145, + 479, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 479, + 158 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 479, + 158 + ], + "type": "text", + "content": "7. Conclusion and Future Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 165, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 165, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 165, + 555, + 403 + ], + "type": "text", + "content": "In this paper, we introduced SPA-VL, the first large-scale, high-quality dataset designed for the safety alignment of VLMs. SPA-VL's comprehensive coverage of harmfulness domains, diverse question-answer types, and focus on both harmlessness and helpfulness make it a robust dataset for improving model safety without compromising core capabilities. Our experiments demonstrate that models trained on SPA-VL using techniques like PPO and DPO show significant improvements in safety and outperform baseline and other state-of-the-art VLMs in various safety evaluations. Additionally, our analyses reveal that increasing data volume, incorporating diverse responses, and using a mix of question types enhance the safety and performance of aligned models. The findings highlight the importance of comprehensive datasets like SPA-VL in achieving robust safety alignment, ensuring that VLMs can be effectively and safely deployed in real-world applications. SPA-VL represents a significant step towards safer and more reliable vision-language models, paving the way for future research and development in this crucial area." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 405, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 556, + 548 + ], + "type": "text", + "content": "In the future, we aim to extend the scope of our work to encompass the unified \"3H\" framework of helpfulness, harmlessness, and honesty, to ensure a more holistic approach to aligning VLMs with human values. Furthermore, we plan to explore the application of our safety alignment techniques to more complex tasks such as reasoning in VLMs, which will require nuanced understanding and generation of visual content. Additionally, we are interested in investigating the transferability of alignment capabilities between LLMs and VLMs, which could lead to more efficient and effective alignment strategies across different modalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 559, + 419, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 419, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 419, + 572 + ], + "type": "text", + "content": "8. Acknowledgment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 578, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 578, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 312, + 578, + 555, + 711 + ], + "type": "text", + "content": "This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, as well as the computational resources provided by the Shanghai Artificial Intelligence Laboratory, for their essential contributions of computational resources and technical support. Furthermore, we extend our heartfelt gratitude to our advisors for their invaluable guidance, insightful discussions, and constructive feedback throughout the research process." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19874" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 188, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 188, + 101 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 188, + 101 + ], + "type": "text", + "content": "[1] Anthropic. Claude, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 293, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 293, + 125 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 293, + 125 + ], + "type": "text", + "content": "[2] Anthropic. Claude usage policies, 2024. https://www.anthropic.com/legal/aup.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 182, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 236 + ], + "type": "text", + "content": "[4] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 238, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 280 + ], + "type": "text", + "content": "[5] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 282, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 294, + 313 + ], + "type": "text", + "content": "[6] Luke Bailey, Euan Ong, Stuart Russell, and Scott Emmons. Image hijacks: Adversarial images can control generative models at runtime, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 316, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 380 + ], + "type": "text", + "content": "[7] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In FAccT '21: 2021 ACM Conference on Fairness, Accountability, and Transparency, Virtual Event / Toronto, Canada, March 3-10, 2021, pages 610-623. ACM, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 383, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 295, + 567 + ], + "type": "text", + "content": "[8] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. CoRR, abs/2108.07258, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "type": "text", + "content": "[9] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 60, + 625, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 625, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 60, + 625, + 294, + 689 + ], + "type": "text", + "content": "[10] Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Dress: Instructing large vision-language models to align and interact with humans via natural language feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14239-14250, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 693, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 294, + 713 + ], + "type": "text", + "content": "[11] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 138 + ], + "type": "text", + "content": "Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[12] Jieren Deng, Yijue Wang, Ji Li, Chenghong Wang, Chao Shang, Hang Liu, Sanguthevar Rajasekaran, and Caiwen Ding. Tag: Gradient attack on transformer-based language models. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3600-3610, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "text", + "content": "[13] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "type": "text", + "content": "[14] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 300, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 332 + ], + "type": "text", + "content": "[15] Google. Gemini usage policies, 2024. https://ai.google.dev/gemini-api/docs/safety-settings.4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 335, + 553, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 335, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 553, + 378 + ], + "type": "text", + "content": "[16] Susan Hao, Renee Shelby, Yuchi Liu, Hansa Srinivasan, Mukul Bhutani, Burcu Karagol Ayan, Shivani Poddar, and Sarah Laszlo. Harm amplification in text-to-image models. arXiv preprint arXiv:2402.01787, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 381, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 381, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 381, + 553, + 588 + ], + "type": "text", + "content": "[17] Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P. Xing, Furong Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huaxiu Yao, Manolis Kellis, Marinka Zitnik, Meng Jiang, Mohit Bansal, James Zou, Jian Pei, Jian Liu, Jianfeng Gao, Jiawei Han, Jieyu Zhao, Jiliang Tang, Jindong Wang, Joaquin Vanschoren, John Mitchell, Kai Shu, Kaidi Xu, Kai-Wei Chang, Lifang He, Lifu Huang, Michael Backes, Neil Zhenqiang Gong, Philip S. Yu, Pin-Yu Chen, Quanquan Gu, Ran Xu, Rex Ying, Shuiwang Ji, Suman Jana, Tianlong Chen, Tianming Liu, Tianyi Zhou, William Yang Wang, Xiang Li, Xiangliang Zhang, Xiao Wang, Xing Xie, Xun Chen, Xuyu Wang, Yan Liu, Yanfang Ye, Yinzhi Cao, Yong Chen, and Yue Zhao. Trustworthiness in large language models. In Forty-first International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 591, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 645 + ], + "type": "text", + "content": "[18] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Llm-based input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 648, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 648, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 648, + 553, + 712 + ], + "type": "text", + "content": "[19] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19875" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 149 + ], + "type": "text", + "content": "[20] Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and Sushant Prakash. RLAIF vs. RLHF: scaling reinforcement learning from human feedback with AI feedback. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[21] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: a multi-modal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 273 + ], + "type": "text", + "content": "[22] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Saladbench: A hierarchical and comprehensive safety benchmark for large language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3923-3954. Association for Computational Linguistics, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 274, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 340 + ], + "type": "text", + "content": "[23] Mukai Li, Lei Li, Yuwei Yin, Masood Ahmed, Zhenguang Liu, and Qi Liu. Red teaming visual language models. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 3326-3342. Association for Computational Linguistics, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "text", + "content": "[24] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. CoRR, abs/2403.09792, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "text", + "content": "[25] Jiawei Liang, Siyuan Liang, Man Luo, Aishan Liu, Dongchen Han, Ee-Chien Chang, and Xiaochun Cao. Vl-trojan: Multimodal instruction backdoor attacks against autoregressive visual language models. arXiv preprint arXiv:2402.13851, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 443, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 486 + ], + "type": "text", + "content": "[26] Hongzhan Lin, Ziyang Luo, Bo Wang, Ruichao Yang, and Jing Ma. Goat-bench: Safety insights to large multimodal models through meme-based social abuse. ArXiv, abs/2401.01523, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 488, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 542 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26286-26296. IEEE, 2024. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 601, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 654 + ], + "type": "text", + "content": "[29] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, pages 386-403. Springer, 2025. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[30] LLamaGuard2. Llama guard2, 2024. https://github.com/meta-llama/PurpleLlama/blob/main/Llama-Guard2/MODEL_CARD.md.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[31] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xiaoyu Guo, and Chaowei Xiao. Jailbreakv-28k: A benchmark for assessing" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the robustness of multimodal large language models against jailbreak attacks. arXiv preprint arXiv:2404.03027, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "type": "text", + "content": "[32] Meta. Llama usage policies, 2024. https://ai.meta.com/llama/use-policy.4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 120, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 120, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 120, + 553, + 152 + ], + "type": "text", + "content": "[33] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 475, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 475, + 165 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 475, + 165 + ], + "type": "text", + "content": "[34] OpenAI. Introducing chatgpt, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 166, + 553, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 166, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 553, + 187 + ], + "type": "text", + "content": "[35] OpenAI. Openai usage policies, 2023. https://openai.com/policies/usage-policies.4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 189, + 539, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 189, + 539, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 189, + 539, + 200 + ], + "type": "text", + "content": "[36] OpenAI. Gpt4v, 2024. https://chatgpt.com/.1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 201, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 276 + ], + "type": "text", + "content": "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 278, + 553, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 278, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 553, + 355 + ], + "type": "text", + "content": "[38] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm's safety without hurting performance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 16012-16027. Association for Computational Linguistics, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 357, + 553, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 357, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 357, + 553, + 411 + ], + "type": "text", + "content": "[39] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 413, + 553, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 413, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 316, + 413, + 553, + 467 + ], + "type": "text", + "content": "[40] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21527-21536, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 534 + ], + "type": "text", + "content": "[41] Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, and Peter Henderson. Fine-tuning aligned language models compromises safety, even when users do not intend to! In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 536, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 601 + ], + "type": "text", + "content": "[42] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 667 + ], + "type": "text", + "content": "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, pages 53728-53741. Curran Associates, Inc., 2023. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19876" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 95 + ], + "type": "text", + "content": "next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv: Learning, arXiv: Learning. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[46] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 174, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 228 + ], + "type": "text", + "content": "[47] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 316 + ], + "type": "text", + "content": "[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liangyan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darryll. Aligning large multimodal models with factually augmented RLHF. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 13088-13110. Association for Computational Linguistics, 2024. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 514 + ], + "type": "text", + "content": "[49] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Agüera y Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. Lamda: Language models for dialog applications. CoRR, abs/2201.08239, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 514, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 568 + ], + "type": "text", + "content": "[50] Haoqin Tu, Chenhang Cui, Zijun Wang, Yiyang Zhou, Bingchen Zhao, Junlin Han, Wangchunshu Zhou, Huaxiu Yao, and Cihang Xie. How many unicorns are in this image? a safety evaluation benchmark for vision llms. arXiv preprint arXiv:2311.16101, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 570, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 657 + ], + "type": "text", + "content": "[51] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Mozhi Zhang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 16-16, 2024, pages 10460-10479. Association for Computational Linguistics, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 690 + ], + "type": "text", + "content": "[52] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[53] Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 108, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 163 + ], + "type": "text", + "content": "[54] Laura Weidinger, Maribeth Rauh, Nahema Marchal, Arianna Manzini, Lisa Anne Hendricks, Juan Mateos-Garcia, Stevie Bergman, Jackie Kay, Conor Griffin, Ben Bariach, et al. Sociotechnical safety evaluation of generative ai systems. arXiv preprint arXiv:2310.11986, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "text", + "content": "[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 220, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 308 + ], + "type": "text", + "content": "[56] Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, Jing Shao, and Wanli Ouyang. LAMM: language-assisted multi-modal instruction-tuning dataset, framework, and benchmark. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 310, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 387 + ], + "type": "text", + "content": "[57] Pan Zhang, XiaoyiDongBin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer: A vision-language large model for advanced text-image comprehension and composition. 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 388, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 487 + ], + "type": "text", + "content": "[58] Zaibin Zhang, Yongting Zhang, Lijun Li, Jing Shao, Hongzhi Gao, Yu Qiao, Lijun Wang, Huchuan Lu, and Feng Zhao. Psysafe: A comprehensive framework for psychological-based attack, defense, and evaluation of multi-agent system safety. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pages 15202-15231. Association for Computational Linguistics, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "type": "text", + "content": "[60] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 591, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 644 + ], + "type": "text", + "content": "[61] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. In ICLR 2024 Workshop on Reliable and Responsible Foundation Models, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[62] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. Open-Review.net, 2024. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19877" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[63] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy M. Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "text", + "content": "[64] Andy Zou, Zifan Wang, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 5" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19878" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_content_list.json b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..659bd7f5d0be88ef5aebba72f37b5c4d8e43737e --- /dev/null +++ b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixuan Huang $^{1,2*}$", + "bbox": [ + 124, + 179, + 267, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mark Boss", + "bbox": [ + 287, + 180, + 383, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aaryaman Vasishta", + "bbox": [ + 405, + 180, + 566, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "James M. Rehg2", + "bbox": [ + 588, + 180, + 722, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Varun Jampani1", + "bbox": [ + 741, + 180, + 869, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stability AI, $^{2}$ UIUC", + "bbox": [ + 411, + 198, + 581, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c1fc2eab87aa86ef36b6063d9cf3ada5b2199677ce684d135eebf691bcd83d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 251, + 890, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/691b91eefc4219c9142139c74942557f2bda43bf28dbfb07af89923f779df4c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 484, + 155, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/157a8772aa8ac76751d2f04a0f3454ac4e7fccce74d08c298b65dc5f0b730a5e.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 127, + 560, + 155, + 614 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b44ad1086ba85fe7d52370bd8ba146af062a4c7e879c6ba08f36bac750b2ece.jpg", + "image_caption": [ + "Edit" + ], + "image_footnote": [], + "bbox": [ + 222, + 487, + 256, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d79cb85b41d00beb0c0e0e7d8338c2223471fa4a119d8990400b2b61e6a85883.jpg", + "image_caption": [ + "Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 220, + 560, + 256, + 611 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e657a7487da64a95e34e9bdea82477b4ca614d5d8665207bc9c1435127cfbf36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 484, + 364, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9967dd5bb81f49ba25954978781718a3489a8061c1eede5afa7dd754eae8a14.jpg", + "image_caption": [ + "Mesh Output" + ], + "image_footnote": [], + "bbox": [ + 331, + 556, + 362, + 611 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8fa67c7d50c8e6a599b049923dea4604e2ba30542cb8d4a05dc5419502da53d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 484, + 464, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/390dd82fef223989d09b23db0a4da6138d1acc915b12412a9685fa4a85fe2c1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 556, + 462, + 614 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5376fd4b94788fec55be741d22454cb5d06c66303eac51d62251dddc02981ab2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 483, + 588, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79a1efa409a65d8c3eacb564fe5ee34eac3e6a0841cb1f70bed1ada2fbc7bb09.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 527, + 556, + 588, + 616 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1c1c1d187a4aa24464e82bfc1d45d67bbb2d8daf47697b25cb115908f47ac26.jpg", + "image_caption": [ + "↓Edit" + ], + "image_footnote": [], + "bbox": [ + 625, + 487, + 683, + 539 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4135ca0970c221c63fa0705b34d60d451f618a73e70628635099279e8bebc364.jpg", + "image_caption": [ + "Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 622, + 561, + 684, + 613 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/831c9532a3d2ed6a6642bb7253ac18a3d1578aa647a5047ff1120e5f03a267b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 487, + 776, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02e2b00fe7ca057226c39e52d1bdb519ef26669ac772781725a309d0464274f9.jpg", + "image_caption": [ + "Mesh Output", + "Figure 1. We present SPAR3D, a state-of-the-art 3D reconstructor that reconstructs high-quality 3D meshes from single-view images. SPAR3D enjoys a fast reconstruction speed at 0.7 seconds and supports interactive user edits." + ], + "image_footnote": [], + "bbox": [ + 720, + 563, + 776, + 619 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5967dcec98bef3b008a6e986cb670d5a42cf4cb59150f373e786545c6ffc108f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 825, + 484, + 888, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c3287016519b5992976c289a52815721139f85bfebb91681249539c10a1bc3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 560, + 888, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 688, + 326, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study the problem of single-image 3D object reconstruction. Recent works have diverged into two directions: regression-based modeling and generative modeling. Regression methods efficiently infer visible surfaces, but struggle with occluded regions. Generative methods handle uncertain regions better by modeling distributions, but are computationally expensive and the generation is often miscaligned with visible surfaces. In this paper, we present SPAR3D, a novel two-stage approach aiming to take the best of both directions. The first stage of SPAR3D generates", + "bbox": [ + 89, + 720, + 483, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sparse 3D point clouds using a lightweight point diffusion model, which has a fast sampling speed. The second stage uses both the sampled point cloud and the input image to create highly detailed meshes. Our two-stage design enables probabilistic modeling of the ill-posed single-image 3D task while maintaining high computational efficiency and great output fidelity. Using point clouds as an intermediate representation further allows for interactive user edits. Evaluated on diverse datasets, SPAR3D demonstrates superior performance over previous state-of-the-art methods, at an inference speed of 0.7 seconds. Project page with code and model: https://spar3d.github.io", + "bbox": [ + 511, + 689, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done at Stability AI.", + "bbox": [ + 109, + 886, + 256, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16860", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstructing 3D objects from monocular images is a fundamental problem in computer vision. An efficient reconstruction system opens up a wide range of applications, including augmented reality, filmmaking, and manufacturing. Monocular 3D reconstruction is also a complex inverse problem: while the visible surface can be estimated from shading, predicting the occluded surface necessitates a strong 3D object prior. Our field has seen a divergence in two different directions: feedforward regression [2, 11, 20, 25, 26, 28, 39, 55, 56, 61, 63, 65, 66, 69, 70, 73] and diffusion-based generation [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75]. Despite the significant progress made in both directions, each has fundamental limitations.", + "bbox": [ + 89, + 114, + 480, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regression-based models are highly effective in adhering to the visible surface in the image, and the inference speed is typically fast. However, they make the oversimplified assumption of bijective mapping between images and 3D. This assumption introduces ambiguity in the learning objective, leading to poorly estimated surfaces and textures in occluded regions. On the other hand, diffusion-based approaches are generative and do not predict the statistical mean. However, their iterative sampling at inference time is computationally inefficient when modeling high-resolution 3D. Additionally, previous studies such as [28] indicate that diffusion-generated 3D models exhibit worse alignment to the surface visible in the input image. How can we take the best of both worlds while avoiding their limitations?", + "bbox": [ + 89, + 311, + 482, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of this, we propose SPAR3D, which breaks the 3D reconstruction process down into two stages: the point sampling stage and the meshing stage. The point sampling stage uses diffusion models to generate sparse point clouds, followed by the meshing stage, transforming point clouds into highly detailed meshes. Our main idea is to offload the uncertainty modeling to the point sampling stage, where the low resolution of the point clouds allows rapid iterative sampling. The subsequent meshing stage leverages the local image features to transform the point cloud into a detailed mesh of high output fidelity. Reducing the meshing uncertainty with point clouds further facilitates unsupervised learning of inverse rendering, which reduces the baked-in lighting in the textures. Our two-stage design enables SPAR3D to significantly outperform previous regressive methods, while preserving high computational efficiency and fidelity to input observation.", + "bbox": [ + 89, + 523, + 482, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key design choice of our method is the usage of point clouds to connect the two stages. To ensure fast reconstruction, our intermediate representation needs to be lightweight so it can be efficiently generated. However, it should provide enough guidance to the meshing stage. This inspires us to use point clouds, which are perhaps the most computationally efficient 3D representation because all information bits are used to represent the surface. Moreover, the lack of", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "connectivity, typically considered as the drawback of point clouds, now turns into an advantage with our two-stage approach for editing purposes. When the back surface does not align with user expectations, local edits can be easily made on the low-resolution point clouds without worrying about topologies (see Fig. 1 bottom). Feeding edited point clouds into the meshing stage produces better meshes tailored towards user requirements.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate the superiority of SPAR3D over previous state-of-the-art methods, with solid quantitative and qualitative results on various data sources. SPAR3D also exhibits a strong generalization ability to inthe-wild images and AI-generated images. With a total inference time below 0.7 seconds, SPAR3D is not only efficient but also allows for easy user-driven edits, offering a practical solution to the task of monocular 3D reconstruction. We hope that this is a meaningful step towards scalable generation of high-quality 3D assets.", + "bbox": [ + 511, + 212, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 380, + 653, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feedforward 3D reconstruction methods address the problem of 3D object reconstruction by learning a feedforward model in a regression-based manner. Earlier works [11, 20, 26, 39, 61, 66, 70] in this field typically predict only the geometry and train on small datasets [5, 52], which limits their generalization ability. Recently, larger 3D datasets [12, 44] have been collected, unlocking the potential to train feedforward 3D models at scale [25, 28, 65]. These models exhibit great generalization ability to unseen images, and excel at producing reconstructions that tightly align with the observed cues in the input image. In particular, LRM [25] and follow-up works [2, 55, 56, 63, 69, 73] show that properly designed large transformer models can be trained using only rendering losses to capture object geometry and texture in great detail. Despite the high fidelity and computational efficiency of these models, the oversimplified bijective assumption in these regressive approaches results in oversmoothed unseen surfaces. Multi-view diffusion models [36, 37, 47, 48] have been considered as a remedy for this, where additional viewpoints are synthesized as input to the feedforward model [55, 63, 69]. However, the inconsistency across viewpoints often leads to significant artifacts on the reconstructed surfaces, and the computational efficiency of these approaches is severely affected by the slow multi-view generation process. Our model also aims to overcome the learning ambiguity in regressive approaches, but our point sampling approach is inherently 3D-consistent and computationally efficient, and further allows easy user edits.", + "bbox": [ + 511, + 407, + 903, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative 3D modeling learns the image-conditioned distribution of 3D assets instead of a deterministic mapping. Early 3D generative works use GAN [4, 17, 29, 57],", + "bbox": [ + 511, + 854, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16861", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e8e51b1fbb67bb7db36b5bbe6cfc083785c1ae501f34e37887d3e051ccd0c87d.jpg", + "image_caption": [ + "Figure 2. SPAR3D Overview. Conditioned on the input image, SPAR3D first leverages a point diffusion model to generate a sparse point cloud. The triplane transformer then uses the sampled point cloud and image features to produce high-resolution triplane features. The triplane features are then queried to reconstruct the geometry, texture, and illumination of the object in the image." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 893, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "normalizing flow [31, 71] or VAE [18, 40, 68] as the generative framework. Inspired by the success of 2D diffusion models [15, 45], 3D diffusion models [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75] have also been extensively explored in recent works. Despite the advantage of probabilistic modeling that avoids over-smoothed results, diffusion-based 3D generation has two drawbacks: 1) not aligning well with input observations, and 2) having low inference speed at high resolution. Our work inherits the advantage of probabilistic modeling, while avoiding the drawbacks by using diffusion to generate only sparse point clouds.", + "bbox": [ + 88, + 348, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization-based single-view 3D leverages 2D generative priors to recover 3D from single-view images. These works [13, 21, 38, 54] rely on SDS-type loss [43, 60] and generate 3D assets by optimizing for each object image separately. These methods achieve promising results without large-scale annotation. However, the lack of a strong explicit 3D prior makes the optimization process inefficient and prone to local minima.", + "bbox": [ + 89, + 525, + 483, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 662, + 181, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SPAR3D Overview. Given the input image $I \\in \\mathbb{R}^{3 \\times h \\times w}$ , our method produces a 3D mesh with PBR materials, including albedo, metallic, roughness and surface normals. The main goal of our work is to develop a model that enjoys the benefits of distribution learning through diffusion models, while not suffering from the low output fidelity and computational inefficiency. To this end, we design a two-stage model that consists of the point sampling stage and the meshing stage (see Fig. 2). At the point sampling stage, a point diffusion model learns the conditional distribution of point clouds given the input image. This stage is computationally efficient given the low resolution of the point clouds. The regression-based meshing stage transforms the sampled point cloud into a highly detailed mesh that aligns", + "bbox": [ + 89, + 686, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with the visible surface. The reduced uncertainty with point sampling further facilitates the learning of materials and illumination in an unsupervised manner during the meshing stage. This reduces baked-in lighting artifacts and results in better modeling of specular surfaces. Finally, by using sparse point clouds as the intermediate representation, SPAR3D enables human editing in the loop.", + "bbox": [ + 511, + 348, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Point Sampling Stage", + "text_level": 1, + "bbox": [ + 511, + 462, + 714, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. The point sampling stage produces a sparse point cloud as the input to the meshing stage. The core of the point sampling stage is a point diffusion model, which generates point clouds $\\pmb{p}_0 \\in \\mathbb{R}^{n \\times 6}$ conditioned on the input image $I$ . The six channels include three XYZ channels and three RGB channels. In our work, the resolution of the point cloud $n$ is set to 512.", + "bbox": [ + 511, + 484, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Point Diffusion Framework.** Our diffusion framework is based on DDPM [24], which consists of two processes: 1) the forward process which adds noise to the original point cloud, and 2) the backward process where the denoiser learns to remove the noise. At timestep $t \\in [0,T]$ , the diffusion process combines Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ with a point cloud $p_0$ as", + "bbox": [ + 511, + 595, + 905, + 702 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {p} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 709, + 903, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ denotes the noise schedule. We use the sigmoid noise schedule proposed in [7], combined with input scaling and the renormalization trick. The denoiser $\\pmb{\\epsilon}_{\\theta}(\\pmb{p}_t, t; \\pmb{c})$ then learns to recover the noise from $\\pmb{p}_t$ and is supervised by", + "bbox": [ + 511, + 733, + 905, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s i m p l e} (\\theta) = \\mathbb {E} _ {t, \\boldsymbol {p} _ {0}, \\boldsymbol {\\epsilon}} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {p} _ {t}, t; \\boldsymbol {c}) \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 800, + 903, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $c$ denotes the image condition tokens. During inference, we use the DDIM sampler [51] to generate point cloud samples. Samples generated directly often align poorly with the condition, hence we use the classifier-free guidance (CFG) [23] to improve sampling fidelity.", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16862", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoiser Design. We use a transformer denoiser similar to Point-E [41], where the noisy point cloud $\\pmb{p}_t \\in \\mathbb{R}^{n \\times 6}$ is linearly mapped to a set of point tokens $\\pmb{x} \\in \\mathbb{R}^{n \\times d}$ . We use DINOv2 [42] to encode the input image $I$ as conditioning tokens $\\pmb{c} \\in \\mathbb{R}^{c \\times d}$ . The conditions and the point tokens are then concatenated together as input to the transformer, which predicts the added noise on each point.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Albedo Point clouds. In the meshing stage, we estimate the materials and lighting alongside the geometry. However, this decomposition is inherently ambiguous because there are countless combinations of lighting and albedo that can explain the same input image. It is challenging to learn this highly uncertain decomposition during the regressive meshing stage alone. We therefore reduce the uncertainty at the point sampling stage, by directly generating albedo point clouds with diffusion models. Sampling albedo point clouds as input to the meshing stage drastically reduces the ambiguity of inverse rendering and stabilizes the decomposition learning.", + "bbox": [ + 89, + 207, + 480, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Meshing Stage", + "text_level": 1, + "bbox": [ + 89, + 401, + 238, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. The meshing stage produces a textured mesh from the input image and the point cloud. The backbone of our meshing model is a large triplane transformer, which predicts triplane features from the image and point cloud conditions. We estimate the geometry, texture and lighting of the current object from the triplane, and metallic/roughness from the image features. The geometry and materials are fed into our differentiable renderer during training, so that we can apply rendering loss to supervise our model.", + "bbox": [ + 89, + 425, + 480, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Triplane Transformer. Our triplane transformer consists of three submodules: the point cloud encoder, the image encoder, and the transformer backbone. We use a simple transformer encoder to encode the point cloud as a set of point tokens. Given the low resolution of the point clouds, each point can be directly mapped to a single token. Our image encoder is DINOv2 [42], which produces local image embeddings. Our triplane transformer follows a similar design to PointInfinity [27] and SF3D [2], which produces triplanes at high resolution of $384 \\times 384$ by using a computationally-detached two-stream design.", + "bbox": [ + 89, + 588, + 480, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface Estimation. To estimate the geometry, the triplanes are queried with a shallow MLP to produce density values. Similar to [2, 63, 69], we convert the implicit density field to explicit surface using differentiable Marching Tetrahedron (DMTet) [46]. We additionally use two MLP heads to predict vertex offsets and surface normals together with density. These two attributes reduce the artifacts introduced by the Marching Tetrahedron and lead to locally smoother surfaces.", + "bbox": [ + 89, + 763, + 480, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Material and Illumination Estimation. We perform inverse rendering and jointly estimate materials (albedo, metallic and roughness) and illumination alongside the geometry. The task is highly ill-posed and Neural-PIL [1] showed that an illumination prior can reduce the ambiguity. We build our illumination estimator upon the learning-based illumination prior from $\\mathrm{RENI}++$ [19]. $\\mathrm{RENI}++$ is originally an unconditional generative model for HDR illumination generation. We learn an encoder to map triplane features into the latent space of $\\mathrm{RENI}++$ . This allows us to estimate the environment illumination in the input image. The albedo is estimated from triplane similar to geometry, where a shallow MLP predicts the albedo value for each 3D location. For metallic and roughness, we follow SF3D [2] and learn to estimate them with a probabilistic approach via a Beta prior. We find that the CLIP encoder used in SF3D is unstable when the object size changes. We therefore replace their CLIP encoder with AlphaCLIP [53] to alleviate this issue using foreground object masks.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiable Rendering. We implement a differentiable renderer that renders images based on the predicted environment map, PBR materials and geometry surface (see Fig. 3). We use a differentiable mesh rasterizer and add a differentiable Shader. Specifically, we leverage the standard simplified Disney PBR model [3] in our Shader. As we use $\\mathrm{RENI}++$ to reconstruct environment maps, we need to explicitly integrate the incoming radiance. Here, we opt to use the Monte Carlo Integration. Given the low sample counts we can computationally afford during training, we rely on Multiple Importance Sampling (MIS) with the balanced heuristic [58] to reduce integration variance. Additionally, to better model the self-occlusion which has been typically ignored in prior works, we implement a visibility test for better shadow modeling. We take inspiration from real-time graphics and model the visibility test as a screen-space method using the depth map from our rasterizer. An overview of this test is shown in Fig. 4. Specifically, we ray-march a short distance (0.25) in 6 steps for all proposed sample directions from MIS, and project the position back to image space. If the current ray depth is farther away than the sampled value from the depth map, then the ray is marked as shadowed.", + "bbox": [ + 511, + 382, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Function. Our main loss function is the rendering loss that compares renderings from novel views to the groundtruth (GT) images. Specifically, our rendering loss is a linear combination of 1) the L2 distance between the rendered and GT images, 2) the perceptual distance between the rendered and GT images measured by LPIPS [74], and 3) the L2 distance between the rendered opacity and the GT foreground mask. Apart from the rendering loss, we also follow SF3D and apply the mesh and shading regularization that regularizes the surface smoothness and the inverse rendering respectively.", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16863", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2d0c167b5544547dcd492ec0275ffd4ea7eb036926a4c0b28c2939f01796ff3a.jpg", + "image_caption": [ + "Figure 3. Our Differentiable Renderer. We estimate geometry, albedo, lighting, and normal maps from the triplane and metallic/roughness values from the image. We rasterize and interpolate these values as input to our shader (omitted here for simplicity). Our shader uses the Disney BRDF [3] and performs Monte Carlo integration. We further perform visibility testing to improve shadow modeling. Finally, we compare the rendered image with the GT image and minimize the rendering loss." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 540, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/076aad9651cf70aa0dd5d76755f6ed3aca2ab82ac14276932a594066fcc5f2a5.jpg", + "image_caption": [ + "Figure 4. Shadow Modeling. We perform visibility testing in screen-space by marching along sampled rays. If any point along the ray has a ray depth which is farther away than the depth map, we consider the entire ray as shadowed." + ], + "image_footnote": [], + "bbox": [ + 555, + 88, + 903, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Interactive Editing", + "text_level": 1, + "bbox": [ + 89, + 354, + 272, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A unique advantage of our two-stage design is that it naturally supports interactive editing of unseen regions in our produced mesh. In most circumstances, the visible surface is determined by the input image and remains highly accurate, while the unseen surface is mainly based on the sampled point cloud, which might not align with user intention. In this case, editing the unseen surface of the mesh is feasible by altering the point cloud. Point clouds are perhaps one of the most flexible 3D representation for editing purposes because there are no topology constraints. Given the low resolution of our point clouds, editing the point cloud is fairly efficient and intuitive. Users can easily delete, duplicate, stretch or recolor points in the point cloud. Our efficient meshing model is able to produce the adjusted mesh in 0.3 seconds, which makes this process fairly interactive.", + "bbox": [ + 88, + 378, + 485, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 614, + 308, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Sampling Stage. Our point diffusion model has 16 transformer blocks in total. Each transformer block consists of two Layer Normalization layer, one Multi-Head Attention (MHA) layer and one MLP. We use a feature dimension of 1024 and 16 attention heads in each MHA layer. With many emissive objects in our dataset, albedo can be visually distinct from the input image and hard to learn. Therefore, instead of directly generating albedo point clouds in the point sampling stage, we learn to generate white-lit point clouds as a proxy target.", + "bbox": [ + 88, + 637, + 483, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meshing Stage. Our triplane transformer consists of 4 two-stream blocks [27]. Each two-stream block consists of three self-attention and two cross-attentions. The main computation is carried out using 3,072 latent tokens, each with a feature dimension of 1024. The MHA includes 16 attention heads. The point cloud encoder is a vanilla transformer with 12 layers and 512 feature dimension, and the", + "bbox": [ + 88, + 794, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image encoder is DINOv2-large. We use a tetrahedra resolution of 160 for DMTet. In our differentiable Shader, we follow Hasselgren et al. [22] and use a detached biased sampling scheme. We sample based on the specular lobe (GGX [59]), the 2D piecewise-linear distribution of the environment map luminance and the hemispherical distribution. Specifically, we include 6 samples from the GGX lobe, 6 samples from the 2D piecewise-linear distribution of the luminance, and 4 samples from the hemispherical distribution. The main body of the shader is implemented in PyTorch, while the screen-space shadowing and the 2D piecewise-linear distribution computation for the environment map are implemented as custom CUDA kernels for efficiency. The training of our meshing stage includes multiple phases, where we increase the rendering resolution and decrease the batch size at later training phases. We use GT point clouds as input when training the meshing model. The curation of our training data follows TripoSR [56].", + "bbox": [ + 511, + 354, + 908, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 643, + 648, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 667, + 635, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We used two datasets for evaluation, GSO [16] and OmniObject3D [67]. We follow TripoSR [56] and remove simple box or cylindrical objects to avoid bias on simple geometries. Each of the evaluation sets consists of around 250 objects. We render the objects with diverse azimuth angles at different elevations, with randomly sampled HDRI environment maps. We also vary the focal length of the camera to create more diverse test cases.", + "bbox": [ + 511, + 691, + 908, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. To evaluate the geometry quality of the reconstructed meshes, we use follow prior works [56, 69] and use Chamfer Distance (CD) and F-score (FS) as our evaluation metrics. CD measures the alignment between two point clouds and is defined as the average of accuracy and", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16864", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/282a1d12d905838b8d8c8d40fab41fc050cde6ea94860d34e3286ecee841f62c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2040.3590.6380.92215.30.8020.2053.1
LN3Diff [32]0.1740.4220.7030.94917.10.8190.1695.1
LGM [55]0.1960.3560.6350.93617.00.8180.18441.0
CRM [63]0.1610.4370.7350.96117.50.8300.1697.4
3DTopia-XL [8]0.2050.3830.6510.91316.00.8180.19340.6
Era3D [34]0.1810.4170.6900.94116.90.8180.173348.0
TripoSR [56]0.1450.5010.7840.96818.50.8370.1510.2
InstantMesh [69]0.1350.5450.8120.97118.10.8380.14636.1
SF3D [2]0.1370.5400.8060.97018.00.8390.1450.3
SPAR3D (ours)0.1200.5840.8500.98318.60.8360.1390.7
", + "bbox": [ + 112, + 87, + 885, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/72d3d23c98b21e61c4858001e5bea732368be22d1ba76e77a981e165069f870a.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons on GSO [16]. SPAR3D performs favorably to other state-of-the-art methods." + ], + "table_footnote": [], + "table_body": "
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2120.3490.6240.90914.80.80060.2053.1
LN3Diff [32]0.1600.4800.7440.95716.70.8190.1615.0
LGM [55]0.2000.3660.6380.92416.10.8100.18842.0
CRM [63]0.1550.4820.7650.96217.00.8280.1627.0
3DTopia-XL [8]0.1910.4070.6760.93115.80.8170.18641.2
Era3D [34]0.1850.4200.6850.93716.40.8160.167391.7
TripoSR [56]0.1440.5370.7850.96318.00.8350.1470.2
InstantMesh [69]0.1450.5460.7900.96217.20.8320.15034.7
SF3D [2]0.1380.5540.8000.96717.40.8360.1450.3
SPAR3D (ours)0.1220.5870.8450.97817.90.8320.1400.7
", + "bbox": [ + 112, + 276, + 885, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative Comparisons on OmniObject3D [67]. SPAR3D performs favorably to other state-of-the-art methods.", + "bbox": [ + 125, + 444, + 867, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "completeness:", + "bbox": [ + 89, + 463, + 189, + 478 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(S _ {1}, S _ {2}\\right) = \\frac {1}{2 \\left| S _ {1} \\right|} \\sum_ {x \\in S _ {1}} \\min _ {y \\in S _ {2}} \\| x - y \\| _ {2} + \\frac {1}{2 \\left| S _ {2} \\right|} \\sum_ {y \\in S _ {2}} \\min _ {x \\in S _ {1}} \\| x - y \\| _ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 488, + 486, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FS evaluates point cloud alignment by calculating the F-score with a predefined threshold. Predicted points that lie within the distance threshold are considered as correct predictions. A higher FS means better alignment between the reconstructed shape and the groundtruth. To evaluate the texture quality, we compute standard image metrics, including PSNR, SSIM and LPIPS, between images rendered from the predicted mesh and the groundtruth images.", + "bbox": [ + 89, + 534, + 482, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol. To calculate the metrics that are comparable across methods, the meshes need to lie in the same coordinate system. To this end, we perform brute-force search in rotations to align each predicted mesh with the groundtruth mesh. Both the prediction and the groundtruth are normalized before the brute-force alignment, and the alignment is further refined with ICP.", + "bbox": [ + 89, + 664, + 482, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare SPAR3D with other efficient methods for single-view 3D generation or reconstruction [2, 55, 56, 63, 69]. We use the official implementation for all baselines, and we evaluate the produced meshes under the same protocol. Specifically, we compare against TripoSR [56], LGM [55], CRM [63], InstantMesh [69], LN3Diff [32], Shap-E [30] and SF3D [2]. Among these baselines, TripoSR and SF3D are pure regression-based ap", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proaches; LGM, CRM and InstantMesh use multiview diffusion to generate pseudo multi-view images; LN3Diff and Shap-E are purely diffusion-based 3D generative models.", + "bbox": [ + 511, + 462, + 906, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 518, + 653, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We compare SPAR3D to other baselines on GSO and Omniobject3D quantitatively. As shown in Tab. 1 and Tab. 2, SPAR3D outperforms all other regressive or generative baselines significantly across most metrics on both datasets. For SSIM, we observe that SPAR3D is slightly worse than the strongest baseline for this metric. We find that this relates to the Monte Carlo noise from our Shader. SPAR3D is also among the fastest reconstruction models with an inference speed of 0.7 seconds per object, which is significantly faster than 3D or multiview diffusion-based approaches.", + "bbox": [ + 511, + 542, + 906, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. We show qualitative results of different methods in Fig. 5. The reconstructed meshes from pure regression-based approaches such as SF3D or TripoSR align with the input image well, but the backside is often less accurate and over-smoothed. Multi-view diffusion-based methods such as LGM, CRM and InstantMesh show more details on the backside. However, the inconsistency in the synthesized views leads to clear artifacts and overall worse results. Pure generative approaches such as Shap-E and LN3Diff are able to produce sharp surfaces in their generation. However, many details are erroneous hallucinations that do not accurately follow the input images, and the vis", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16865", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a3ecfe6b0ad1bbd0694f064943ef140b23ffe23266785a108909941e3fcb572.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison. We compare SPAR3D to other state-of-the-art methods visually. SPAR3D not only aligns better with the visible surfaces from images, but also generates higher-quality geometries and textures for the occluded surfaces." + ], + "image_footnote": [], + "bbox": [ + 122, + 97, + 883, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3823312c952490db487fbafc77a9068d71383a5ca938984e747f2f70081bd17e.jpg", + "image_caption": [ + "Figure 6. Generalization Results. We show qualitative results of SPAR3D on in-the-wild images from 2D generative models (top 2 rows) and ImageNet (bottom 2 rows). The reconstructed meshes exhibit accurate geometric structures with great textures, demonstrating a strong generalization performance of SPAR3D." + ], + "image_footnote": [], + "bbox": [ + 122, + 506, + 877, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ible surfaces are often reconstructed incorrectly. Compared to prior art, the meshes produced by SPAR3D not only faithfully resemble the input image, but also exhibit well-generated occluded parts with reasonable details. In Fig. 6, we further show qualitative results of SPAR3D on in-the-wild images. The images are generated using text-to-image generative models or from the ImageNet validation set [14]. The high quality of the reconstructed meshes demonstrates", + "bbox": [ + 89, + 771, + 486, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a strong generalization performance of SPAR3D. More results and comparisons are included in the supplement.", + "bbox": [ + 511, + 771, + 906, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Editing Results", + "text_level": 1, + "bbox": [ + 511, + 815, + 668, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The usage of explicit point clouds as an intermediate representation enables interactive editing of the generated meshes. Users can easily alter the unseen surface of the mesh by manipulating the point cloud. In Fig. 7, we show a", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16866", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/886233bcc6fa5c6c13ed349270b3480443d90cb0c90d667d6887c45c7eb55b42.jpg", + "image_caption": [ + "Figure 7. Editing Results. We show qualitative examples of interactive editing with SPAR3D. On the left two examples, we add a handle to the mug and a tail to the elephant doll by duplicating existing points. On the right two examples, we move or delete points to fix imperfections and to improve local details on the mesh. All the edits are performed in Blender within a minute." + ], + "image_footnote": [], + "bbox": [ + 117, + 95, + 869, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "few editing examples with SPAR3D, either by adding major object parts to the reconstruction, or improving undesirable generated details.", + "bbox": [ + 88, + 359, + 482, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation", + "text_level": 1, + "bbox": [ + 89, + 416, + 194, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate the key idea of SPAR3D, the point sampling stage, which can be seen as an addition to standard regression approaches. We consider a variant of our model (SPAR3D w/o Point), where we remove the point sampling stage and make SPAR3D a full regressive model. We compare this variant with our full model on both GSO and Omniobject3D. As shown in Tab. 3, our full SPAR3D significantly outperforms the regressive variant, which validates the effectiveness of our design. More ablation experiments are included in the supplement.", + "bbox": [ + 89, + 439, + 483, + 590 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/900a8e7d35e753ef3588bdae30e76520159ecd7e217bd2b0345fa14259d4c40f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCD↓FS@0.1↑PSNR↑LPIPS↓
SPAR3D w/o Point0.1360.50618.50.146
SPAR3D0.1200.58418.60.139
SPAR3D w/o Point0.1400.50917.80.146
SPAR3D0.1220.58717.90.140
", + "bbox": [ + 91, + 594, + 483, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Study on GSO (top 2 rows) and Omniobject3D (bottom 2 rows). Removing the point sampling stage leads to significant performance drop.", + "bbox": [ + 89, + 672, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 726, + 192, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further design experiments to understand how SPAR3D works. Our key assumption when designing SPAR3D is that the two-stage design effectively separates the uncertain part (back-surface modeling) and the deterministic part (visible surface modeling) of the monocular 3D reconstruction problem. Ideally, the meshing stage should mainly rely on the input image for reconstructing the visible surface, while relying on the point cloud to generate the back surface. To see whether this is true, we design an experiment where we artificially use point clouds that conflict with the input im", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7484dced574c55277f64c202759d1a6d5ef3780359e9971a9cfac87fa1eb588f.jpg", + "image_caption": [ + "Figure 8. Generated Mesh with Conflicting Cues. Under conflicting cues from images and point clouds, our model reconstructs the visible surface based on the image, while generating the backside surface based on the point cloud." + ], + "image_footnote": [], + "bbox": [ + 537, + 368, + 880, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age. In Fig. 8, we feed the input image of a squirrel and the point cloud of a horse to the meshing model. As shown in the figure, the reconstructed mesh indeed aligns with the squirrel image well on the visible surface, while the back surface mainly adheres to the point cloud. This result validates our assumption.", + "bbox": [ + 511, + 599, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SPAR3D, a simple yet effective approach for single-view 3D reconstruction. The core of our model is a two-stage design based on point sampling. We first generate a sparse point cloud via point diffusion, and then reconstruct a highly detailed mesh from both the point cloud and the image. This design enables us to take the best of regression-based and generative modeling. Evaluated on standard benchmarks and in-the-wild images, SPAR3D significantly outperforms previous state-of-the-art methods with a fast inference speed. We will release our model upon publication, and we hope our effort is useful for future research towards scalable generation of high-quality 3D content.", + "bbox": [ + 511, + 732, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16867", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021. 4", + "[2] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint, 2024. 2, 4, 6", + "[3] Brent Burley. Physically-based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 4, 5", + "[4] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 2", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2", + "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2416-2425, 2023. 2, 3", + "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 3", + "[8] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024.6", + "[9] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3", + "[10] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2262-2272, 2023. 2, 3", + "[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 2", + "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2", + "[13] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al." + ], + "bbox": [ + 93, + 143, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20637-20647, 2023. 3", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3", + "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5, 6", + "[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2", + "[18] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3", + "[19] James AD Gardner, Bernhard Egger, and William AP Smith. Reni++ a rotation-equivariant, scale-invariant, natural illumination prior. arXiv preprint arXiv:2311.09361, 2023. 4", + "[20] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[21] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 3", + "[22] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 5", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[25] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2", + "[26] Zixuan Huang, Varun Jampani, Anh Thai, Yuanzhen Li, Stefan Stojanov, and James M Rehg. Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16868", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[27] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10050-10060, 2024. 2, 3, 4, 5", + "[28] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Zeroshape: Regression-based zero-shot shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10071, 2024. 2", + "[29] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 397-413. Springer, 2020. 2", + "[30] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6", + "[31] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 3", + "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. arXiv preprint arXiv:2403.12019, 2024. 2, 3, 6", + "[33] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12642-12651, 2023. 2, 3", + "[34] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wei Xue, Wenhan Luo, et al. Era3d: high-resolution multiview diffusion using efficient row-wise attention. Advances in Neural Information Processing Systems, 37:55975-56000, 2024. 6", + "[35] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885, 2023. 2, 3", + "[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2", + "[37] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3", + "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 3", + "[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks:" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2", + "[40] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 3", + "[41] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3, 4", + "[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4", + "[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 4", + "[47] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model, 2023. 2", + "[48] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3", + "[49] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20887-20897, 2023.", + "[50] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3", + "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16869", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2", + "[53] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want, 2023. 4", + "[54] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 3", + "[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 6", + "[56] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 2, 5, 6, 12", + "[57] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International conference on learning representations, 2018. 2", + "[58] Eric Veach. *Robust Monte Carlo Methods for Light Transport Simulation*. PhD thesis, Stanford University, 1997. 4", + "[59] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. Eurographics Symposium on Rendering, 2007. 5", + "[60] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3", + "[61] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2", + "[62] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 13", + "[63] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 2, 4, 6", + "[64] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 12" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. arXiv preprint arXiv:2301.08247, 2023. 2", + "[66] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 2", + "[67] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. arXiv preprint arXiv:2301.07525, 2023. 5, 6", + "[68] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3", + "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4, 5, 6", + "[70] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 2", + "[71] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3", + "[72] Lior Yariv, Omri Puny, Oran Gafni, and Yaron Lipman. Mosaic-sdf for 3d generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4630-4639, 2024. 2, 3", + "[73] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2", + "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4", + "[75] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 3" + ], + "bbox": [ + 516, + 90, + 906, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16870", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_model.json b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42db42bd986bccf00c09ed18e64c22c6603a681d --- /dev/null +++ b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_model.json @@ -0,0 +1,2411 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.18, + 0.269, + 0.201 + ], + "angle": 0, + "content": "Zixuan Huang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.289, + 0.181, + 0.385, + 0.198 + ], + "angle": 0, + "content": "Mark Boss" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.181, + 0.568, + 0.199 + ], + "angle": 0, + "content": "Aaryaman Vasishta" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.181, + 0.723, + 0.199 + ], + "angle": 0, + "content": "James M. Rehg2" + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.181, + 0.87, + 0.199 + ], + "angle": 0, + "content": "Varun Jampani1" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.199, + 0.582, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Stability AI, \\(^{2}\\)UIUC" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.252, + 0.891, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.486, + 0.156, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.561, + 0.156, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.619, + 0.178, + 0.633 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.488, + 0.258, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.543, + 0.268, + 0.559 + ], + "angle": 0, + "content": "Edit" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.561, + 0.257, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.619, + 0.275, + 0.632 + ], + "angle": 0, + "content": "Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.485, + 0.365, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.557, + 0.364, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.619, + 0.435, + 0.632 + ], + "angle": 0, + "content": "Mesh Output" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.485, + 0.465, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.557, + 0.464, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.484, + 0.589, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.558, + 0.589, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.619, + 0.596, + 0.632 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.488, + 0.684, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.544, + 0.677, + 0.559 + ], + "angle": 0, + "content": "↓Edit" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.562, + 0.685, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.619, + 0.69, + 0.631 + ], + "angle": 0, + "content": "Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.488, + 0.777, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.564, + 0.777, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.62, + 0.843, + 0.632 + ], + "angle": 0, + "content": "Mesh Output" + }, + { + "type": "image", + "bbox": [ + 0.826, + 0.486, + 0.889, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.561, + 0.889, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.648, + 0.905, + 0.677 + ], + "angle": 0, + "content": "Figure 1. We present SPAR3D, a state-of-the-art 3D reconstructor that reconstructs high-quality 3D meshes from single-view images. SPAR3D enjoys a fast reconstruction speed at 0.7 seconds and supports interactive user edits." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.689, + 0.327, + 0.705 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.873 + ], + "angle": 0, + "content": "We study the problem of single-image 3D object reconstruction. Recent works have diverged into two directions: regression-based modeling and generative modeling. Regression methods efficiently infer visible surfaces, but struggle with occluded regions. Generative methods handle uncertain regions better by modeling distributions, but are computationally expensive and the generation is often miscaligned with visible surfaces. In this paper, we present SPAR3D, a novel two-stage approach aiming to take the best of both directions. The first stage of SPAR3D generates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.871 + ], + "angle": 0, + "content": "sparse 3D point clouds using a lightweight point diffusion model, which has a fast sampling speed. The second stage uses both the sampled point cloud and the input image to create highly detailed meshes. Our two-stage design enables probabilistic modeling of the ill-posed single-image 3D task while maintaining high computational efficiency and great output fidelity. Using point clouds as an intermediate representation further allows for interactive user edits. Evaluated on diverse datasets, SPAR3D demonstrates superior performance over previous state-of-the-art methods, at an inference speed of 0.7 seconds. Project page with code and model: https://spar3d.github.io" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.258, + 0.901 + ], + "angle": 0, + "content": "*Work done at Stability AI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16860" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.482, + 0.312 + ], + "angle": 0, + "content": "Reconstructing 3D objects from monocular images is a fundamental problem in computer vision. An efficient reconstruction system opens up a wide range of applications, including augmented reality, filmmaking, and manufacturing. Monocular 3D reconstruction is also a complex inverse problem: while the visible surface can be estimated from shading, predicting the occluded surface necessitates a strong 3D object prior. Our field has seen a divergence in two different directions: feedforward regression [2, 11, 20, 25, 26, 28, 39, 55, 56, 61, 63, 65, 66, 69, 70, 73] and diffusion-based generation [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75]. Despite the significant progress made in both directions, each has fundamental limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Regression-based models are highly effective in adhering to the visible surface in the image, and the inference speed is typically fast. However, they make the oversimplified assumption of bijective mapping between images and 3D. This assumption introduces ambiguity in the learning objective, leading to poorly estimated surfaces and textures in occluded regions. On the other hand, diffusion-based approaches are generative and do not predict the statistical mean. However, their iterative sampling at inference time is computationally inefficient when modeling high-resolution 3D. Additionally, previous studies such as [28] indicate that diffusion-generated 3D models exhibit worse alignment to the surface visible in the input image. How can we take the best of both worlds while avoiding their limitations?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.78 + ], + "angle": 0, + "content": "In light of this, we propose SPAR3D, which breaks the 3D reconstruction process down into two stages: the point sampling stage and the meshing stage. The point sampling stage uses diffusion models to generate sparse point clouds, followed by the meshing stage, transforming point clouds into highly detailed meshes. Our main idea is to offload the uncertainty modeling to the point sampling stage, where the low resolution of the point clouds allows rapid iterative sampling. The subsequent meshing stage leverages the local image features to transform the point cloud into a detailed mesh of high output fidelity. Reducing the meshing uncertainty with point clouds further facilitates unsupervised learning of inverse rendering, which reduces the baked-in lighting in the textures. Our two-stage design enables SPAR3D to significantly outperform previous regressive methods, while preserving high computational efficiency and fidelity to input observation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A key design choice of our method is the usage of point clouds to connect the two stages. To ensure fast reconstruction, our intermediate representation needs to be lightweight so it can be efficiently generated. However, it should provide enough guidance to the meshing stage. This inspires us to use point clouds, which are perhaps the most computationally efficient 3D representation because all information bits are used to represent the surface. Moreover, the lack of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "connectivity, typically considered as the drawback of point clouds, now turns into an advantage with our two-stage approach for editing purposes. When the back surface does not align with user expectations, local edits can be easily made on the low-resolution point clouds without worrying about topologies (see Fig. 1 bottom). Feeding edited point clouds into the meshing stage produces better meshes tailored towards user requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Our experiments demonstrate the superiority of SPAR3D over previous state-of-the-art methods, with solid quantitative and qualitative results on various data sources. SPAR3D also exhibits a strong generalization ability to inthe-wild images and AI-generated images. With a total inference time below 0.7 seconds, SPAR3D is not only efficient but also allows for easy user-driven edits, offering a practical solution to the task of monocular 3D reconstruction. We hope that this is a meaningful step towards scalable generation of high-quality 3D assets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.655, + 0.396 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.905, + 0.844 + ], + "angle": 0, + "content": "Feedforward 3D reconstruction methods address the problem of 3D object reconstruction by learning a feedforward model in a regression-based manner. Earlier works [11, 20, 26, 39, 61, 66, 70] in this field typically predict only the geometry and train on small datasets [5, 52], which limits their generalization ability. Recently, larger 3D datasets [12, 44] have been collected, unlocking the potential to train feedforward 3D models at scale [25, 28, 65]. These models exhibit great generalization ability to unseen images, and excel at producing reconstructions that tightly align with the observed cues in the input image. In particular, LRM [25] and follow-up works [2, 55, 56, 63, 69, 73] show that properly designed large transformer models can be trained using only rendering losses to capture object geometry and texture in great detail. Despite the high fidelity and computational efficiency of these models, the oversimplified bijective assumption in these regressive approaches results in oversmoothed unseen surfaces. Multi-view diffusion models [36, 37, 47, 48] have been considered as a remedy for this, where additional viewpoints are synthesized as input to the feedforward model [55, 63, 69]. However, the inconsistency across viewpoints often leads to significant artifacts on the reconstructed surfaces, and the computational efficiency of these approaches is severely affected by the slow multi-view generation process. Our model also aims to overcome the learning ambiguity in regressive approaches, but our point sampling approach is inherently 3D-consistent and computationally efficient, and further allows easy user edits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Generative 3D modeling learns the image-conditioned distribution of 3D assets instead of a deterministic mapping. Early 3D generative works use GAN [4, 17, 29, 57]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16861" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.895, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.298, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 2. SPAR3D Overview. Conditioned on the input image, SPAR3D first leverages a point diffusion model to generate a sparse point cloud. The triplane transformer then uses the sampled point cloud and image features to produce high-resolution triplane features. The triplane features are then queried to reconstruct the geometry, texture, and illumination of the object in the image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.349, + 0.484, + 0.516 + ], + "angle": 0, + "content": "normalizing flow [31, 71] or VAE [18, 40, 68] as the generative framework. Inspired by the success of 2D diffusion models [15, 45], 3D diffusion models [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75] have also been extensively explored in recent works. Despite the advantage of probabilistic modeling that avoids over-smoothed results, diffusion-based 3D generation has two drawbacks: 1) not aligning well with input observations, and 2) having low inference speed at high resolution. Our work inherits the advantage of probabilistic modeling, while avoiding the drawbacks by using diffusion to generate only sparse point clouds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.648 + ], + "angle": 0, + "content": "Optimization-based single-view 3D leverages 2D generative priors to recover 3D from single-view images. These works [13, 21, 38, 54] rely on SDS-type loss [43, 60] and generate 3D assets by optimizing for each object image separately. These methods achieve promising results without large-scale annotation. However, the lack of a strong explicit 3D prior makes the optimization process inefficient and prone to local minima." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.182, + 0.679 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.901 + ], + "angle": 0, + "content": "SPAR3D Overview. Given the input image \\( I \\in \\mathbb{R}^{3 \\times h \\times w} \\), our method produces a 3D mesh with PBR materials, including albedo, metallic, roughness and surface normals. The main goal of our work is to develop a model that enjoys the benefits of distribution learning through diffusion models, while not suffering from the low output fidelity and computational inefficiency. To this end, we design a two-stage model that consists of the point sampling stage and the meshing stage (see Fig. 2). At the point sampling stage, a point diffusion model learns the conditional distribution of point clouds given the input image. This stage is computationally efficient given the low resolution of the point clouds. The regression-based meshing stage transforms the sampled point cloud into a highly detailed mesh that aligns" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.456 + ], + "angle": 0, + "content": "with the visible surface. The reduced uncertainty with point sampling further facilitates the learning of materials and illumination in an unsupervised manner during the meshing stage. This reduces baked-in lighting artifacts and results in better modeling of specular surfaces. Finally, by using sparse point clouds as the intermediate representation, SPAR3D enables human editing in the loop." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.463, + 0.715, + 0.479 + ], + "angle": 0, + "content": "3.1. Point Sampling Stage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Overview. The point sampling stage produces a sparse point cloud as the input to the meshing stage. The core of the point sampling stage is a point diffusion model, which generates point clouds \\( \\pmb{p}_0 \\in \\mathbb{R}^{n \\times 6} \\) conditioned on the input image \\( I \\). The six channels include three XYZ channels and three RGB channels. In our work, the resolution of the point cloud \\( n \\) is set to 512." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.703 + ], + "angle": 0, + "content": "**Point Diffusion Framework.** Our diffusion framework is based on DDPM [24], which consists of two processes: 1) the forward process which adds noise to the original point cloud, and 2) the backward process where the denoiser learns to remove the noise. At timestep \\( t \\in [0,T] \\), the diffusion process combines Gaussian noise \\( \\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) with a point cloud \\( p_0 \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.71, + 0.905, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {p} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.795 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) denotes the noise schedule. We use the sigmoid noise schedule proposed in [7], combined with input scaling and the renormalization trick. The denoiser \\(\\pmb{\\epsilon}_{\\theta}(\\pmb{p}_t, t; \\pmb{c})\\) then learns to recover the noise from \\(\\pmb{p}_t\\) and is supervised by" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.801, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\nL _ {s i m p l e} (\\theta) = \\mathbb {E} _ {t, \\boldsymbol {p} _ {0}, \\boldsymbol {\\epsilon}} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {p} _ {t}, t; \\boldsymbol {c}) \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Here \\( c \\) denotes the image condition tokens. During inference, we use the DDIM sampler [51] to generate point cloud samples. Samples generated directly often align poorly with the condition, hence we use the classifier-free guidance (CFG) [23] to improve sampling fidelity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16862" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Denoiser Design. We use a transformer denoiser similar to Point-E [41], where the noisy point cloud \\(\\pmb{p}_t \\in \\mathbb{R}^{n \\times 6}\\) is linearly mapped to a set of point tokens \\(\\pmb{x} \\in \\mathbb{R}^{n \\times d}\\). We use DINOv2 [42] to encode the input image \\(I\\) as conditioning tokens \\(\\pmb{c} \\in \\mathbb{R}^{c \\times d}\\). The conditions and the point tokens are then concatenated together as input to the transformer, which predicts the added noise on each point." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.482, + 0.39 + ], + "angle": 0, + "content": "Albedo Point clouds. In the meshing stage, we estimate the materials and lighting alongside the geometry. However, this decomposition is inherently ambiguous because there are countless combinations of lighting and albedo that can explain the same input image. It is challenging to learn this highly uncertain decomposition during the regressive meshing stage alone. We therefore reduce the uncertainty at the point sampling stage, by directly generating albedo point clouds with diffusion models. Sampling albedo point clouds as input to the meshing stage drastically reduces the ambiguity of inverse rendering and stabilizes the decomposition learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.402, + 0.24, + 0.418 + ], + "angle": 0, + "content": "3.2. Meshing Stage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.482, + 0.576 + ], + "angle": 0, + "content": "Overview. The meshing stage produces a textured mesh from the input image and the point cloud. The backbone of our meshing model is a large triplane transformer, which predicts triplane features from the image and point cloud conditions. We estimate the geometry, texture and lighting of the current object from the triplane, and metallic/roughness from the image features. The geometry and materials are fed into our differentiable renderer during training, so that we can apply rendering loss to supervise our model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.482, + 0.755 + ], + "angle": 0, + "content": "Triplane Transformer. Our triplane transformer consists of three submodules: the point cloud encoder, the image encoder, and the transformer backbone. We use a simple transformer encoder to encode the point cloud as a set of point tokens. Given the low resolution of the point clouds, each point can be directly mapped to a single token. Our image encoder is DINOv2 [42], which produces local image embeddings. Our triplane transformer follows a similar design to PointInfinity [27] and SF3D [2], which produces triplanes at high resolution of \\( 384 \\times 384 \\) by using a computationally-detached two-stream design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Surface Estimation. To estimate the geometry, the triplanes are queried with a shallow MLP to produce density values. Similar to [2, 63, 69], we convert the implicit density field to explicit surface using differentiable Marching Tetrahedron (DMTet) [46]. We additionally use two MLP heads to predict vertex offsets and surface normals together with density. These two attributes reduce the artifacts introduced by the Marching Tetrahedron and lead to locally smoother surfaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Material and Illumination Estimation. We perform inverse rendering and jointly estimate materials (albedo, metallic and roughness) and illumination alongside the geometry. The task is highly ill-posed and Neural-PIL [1] showed that an illumination prior can reduce the ambiguity. We build our illumination estimator upon the learning-based illumination prior from \\(\\mathrm{RENI}++\\) [19]. \\(\\mathrm{RENI}++\\) is originally an unconditional generative model for HDR illumination generation. We learn an encoder to map triplane features into the latent space of \\(\\mathrm{RENI}++\\). This allows us to estimate the environment illumination in the input image. The albedo is estimated from triplane similar to geometry, where a shallow MLP predicts the albedo value for each 3D location. For metallic and roughness, we follow SF3D [2] and learn to estimate them with a probabilistic approach via a Beta prior. We find that the CLIP encoder used in SF3D is unstable when the object size changes. We therefore replace their CLIP encoder with AlphaCLIP [53] to alleviate this issue using foreground object masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Differentiable Rendering. We implement a differentiable renderer that renders images based on the predicted environment map, PBR materials and geometry surface (see Fig. 3). We use a differentiable mesh rasterizer and add a differentiable Shader. Specifically, we leverage the standard simplified Disney PBR model [3] in our Shader. As we use \\(\\mathrm{RENI}++\\) to reconstruct environment maps, we need to explicitly integrate the incoming radiance. Here, we opt to use the Monte Carlo Integration. Given the low sample counts we can computationally afford during training, we rely on Multiple Importance Sampling (MIS) with the balanced heuristic [58] to reduce integration variance. Additionally, to better model the self-occlusion which has been typically ignored in prior works, we implement a visibility test for better shadow modeling. We take inspiration from real-time graphics and model the visibility test as a screen-space method using the depth map from our rasterizer. An overview of this test is shown in Fig. 4. Specifically, we ray-march a short distance (0.25) in 6 steps for all proposed sample directions from MIS, and project the position back to image space. If the current ray depth is farther away than the sampled value from the depth map, then the ray is marked as shadowed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Loss Function. Our main loss function is the rendering loss that compares renderings from novel views to the groundtruth (GT) images. Specifically, our rendering loss is a linear combination of 1) the L2 distance between the rendered and GT images, 2) the perceptual distance between the rendered and GT images measured by LPIPS [74], and 3) the L2 distance between the rendered opacity and the GT foreground mask. Apart from the rendering loss, we also follow SF3D and apply the mesh and shading regularization that regularizes the surface smoothness and the inverse rendering respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.955 + ], + "angle": 0, + "content": "16863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.541, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.249, + 0.545, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Our Differentiable Renderer. We estimate geometry, albedo, lighting, and normal maps from the triplane and metallic/roughness values from the image. We rasterize and interpolate these values as input to our shader (omitted here for simplicity). Our shader uses the Disney BRDF [3] and performs Monte Carlo integration. We further perform visibility testing to improve shadow modeling. Finally, we compare the rendered image with the GT image and minimize the rendering loss." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.089, + 0.905, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.249, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Figure 4. Shadow Modeling. We perform visibility testing in screen-space by marching along sampled rays. If any point along the ray has a ray depth which is farther away than the depth map, we consider the entire ray as shadowed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.355, + 0.274, + 0.372 + ], + "angle": 0, + "content": "3.3. Interactive Editing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.486, + 0.606 + ], + "angle": 0, + "content": "A unique advantage of our two-stage design is that it naturally supports interactive editing of unseen regions in our produced mesh. In most circumstances, the visible surface is determined by the input image and remains highly accurate, while the unseen surface is mainly based on the sampled point cloud, which might not align with user intention. In this case, editing the unseen surface of the mesh is feasible by altering the point cloud. Point clouds are perhaps one of the most flexible 3D representation for editing purposes because there are no topology constraints. Given the low resolution of our point clouds, editing the point cloud is fairly efficient and intuitive. Users can easily delete, duplicate, stretch or recolor points in the point cloud. Our efficient meshing model is able to produce the adjusted mesh in 0.3 seconds, which makes this process fairly interactive." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.31, + 0.631 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.638, + 0.485, + 0.79 + ], + "angle": 0, + "content": "Point Sampling Stage. Our point diffusion model has 16 transformer blocks in total. Each transformer block consists of two Layer Normalization layer, one Multi-Head Attention (MHA) layer and one MLP. We use a feature dimension of 1024 and 16 attention heads in each MHA layer. With many emissive objects in our dataset, albedo can be visually distinct from the input image and hard to learn. Therefore, instead of directly generating albedo point clouds in the point sampling stage, we learn to generate white-lit point clouds as a proxy target." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Meshing Stage. Our triplane transformer consists of 4 two-stream blocks [27]. Each two-stream block consists of three self-attention and two cross-attentions. The main computation is carried out using 3,072 latent tokens, each with a feature dimension of 1024. The MHA includes 16 attention heads. The point cloud encoder is a vanilla transformer with 12 layers and 512 feature dimension, and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.909, + 0.629 + ], + "angle": 0, + "content": "image encoder is DINOv2-large. We use a tetrahedra resolution of 160 for DMTet. In our differentiable Shader, we follow Hasselgren et al. [22] and use a detached biased sampling scheme. We sample based on the specular lobe (GGX [59]), the 2D piecewise-linear distribution of the environment map luminance and the hemispherical distribution. Specifically, we include 6 samples from the GGX lobe, 6 samples from the 2D piecewise-linear distribution of the luminance, and 4 samples from the hemispherical distribution. The main body of the shader is implemented in PyTorch, while the screen-space shadowing and the 2D piecewise-linear distribution computation for the environment map are implemented as custom CUDA kernels for efficiency. The training of our meshing stage includes multiple phases, where we increase the rendering resolution and decrease the batch size at later training phases. We use GT point clouds as input when training the meshing model. The curation of our training data follows TripoSR [56]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.644, + 0.649, + 0.661 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.637, + 0.685 + ], + "angle": 0, + "content": "4.1. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.909, + 0.815 + ], + "angle": 0, + "content": "Datasets. We used two datasets for evaluation, GSO [16] and OmniObject3D [67]. We follow TripoSR [56] and remove simple box or cylindrical objects to avoid bias on simple geometries. Each of the evaluation sets consists of around 250 objects. We render the objects with diverse azimuth angles at different elevations, with randomly sampled HDRI environment maps. We also vary the focal length of the camera to create more diverse test cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metrics. To evaluate the geometry quality of the reconstructed meshes, we use follow prior works [56, 69] and use Chamfer Distance (CD) and F-score (FS) as our evaluation metrics. CD measures the alignment between two point clouds and is defined as the average of accuracy and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16864" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.088, + 0.887, + 0.246 + ], + "angle": 0, + "content": "
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2040.3590.6380.92215.30.8020.2053.1
LN3Diff [32]0.1740.4220.7030.94917.10.8190.1695.1
LGM [55]0.1960.3560.6350.93617.00.8180.18441.0
CRM [63]0.1610.4370.7350.96117.50.8300.1697.4
3DTopia-XL [8]0.2050.3830.6510.91316.00.8180.19340.6
Era3D [34]0.1810.4170.6900.94116.90.8180.173348.0
TripoSR [56]0.1450.5010.7840.96818.50.8370.1510.2
InstantMesh [69]0.1350.5450.8120.97118.10.8380.14636.1
SF3D [2]0.1370.5400.8060.97018.00.8390.1450.3
SPAR3D (ours)0.1200.5840.8500.98318.60.8360.1390.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.256, + 0.837, + 0.271 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons on GSO [16]. SPAR3D performs favorably to other state-of-the-art methods." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.277, + 0.887, + 0.436 + ], + "angle": 0, + "content": "
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2120.3490.6240.90914.80.80060.2053.1
LN3Diff [32]0.1600.4800.7440.95716.70.8190.1615.0
LGM [55]0.2000.3660.6380.92416.10.8100.18842.0
CRM [63]0.1550.4820.7650.96217.00.8280.1627.0
3DTopia-XL [8]0.1910.4070.6760.93115.80.8170.18641.2
Era3D [34]0.1850.4200.6850.93716.40.8160.167391.7
TripoSR [56]0.1440.5370.7850.96318.00.8350.1470.2
InstantMesh [69]0.1450.5460.7900.96217.20.8320.15034.7
SF3D [2]0.1380.5540.8000.96717.40.8360.1450.3
SPAR3D (ours)0.1220.5870.8450.97817.90.8320.1400.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.445, + 0.869, + 0.46 + ], + "angle": 0, + "content": "Table 2. Quantitative Comparisons on OmniObject3D [67]. SPAR3D performs favorably to other state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.19, + 0.479 + ], + "angle": 0, + "content": "completeness:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.489, + 0.487, + 0.534 + ], + "angle": 0, + "content": "\\[\nd \\left(S _ {1}, S _ {2}\\right) = \\frac {1}{2 \\left| S _ {1} \\right|} \\sum_ {x \\in S _ {1}} \\min _ {y \\in S _ {2}} \\| x - y \\| _ {2} + \\frac {1}{2 \\left| S _ {2} \\right|} \\sum_ {y \\in S _ {2}} \\min _ {x \\in S _ {1}} \\| x - y \\| _ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.656 + ], + "angle": 0, + "content": "FS evaluates point cloud alignment by calculating the F-score with a predefined threshold. Predicted points that lie within the distance threshold are considered as correct predictions. A higher FS means better alignment between the reconstructed shape and the groundtruth. To evaluate the texture quality, we compute standard image metrics, including PSNR, SSIM and LPIPS, between images rendered from the predicted mesh and the groundtruth images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Protocol. To calculate the metrics that are comparable across methods, the meshes need to lie in the same coordinate system. To this end, we perform brute-force search in rotations to align each predicted mesh with the groundtruth mesh. Both the prediction and the groundtruth are normalized before the brute-force alignment, and the alignment is further refined with ICP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare SPAR3D with other efficient methods for single-view 3D generation or reconstruction [2, 55, 56, 63, 69]. We use the official implementation for all baselines, and we evaluate the produced meshes under the same protocol. Specifically, we compare against TripoSR [56], LGM [55], CRM [63], InstantMesh [69], LN3Diff [32], Shap-E [30] and SF3D [2]. Among these baselines, TripoSR and SF3D are pure regression-based ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.51 + ], + "angle": 0, + "content": "proaches; LGM, CRM and InstantMesh use multiview diffusion to generate pseudo multi-view images; LN3Diff and Shap-E are purely diffusion-based 3D generative models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.655, + 0.535 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.711 + ], + "angle": 0, + "content": "Quantitative Comparison. We compare SPAR3D to other baselines on GSO and Omniobject3D quantitatively. As shown in Tab. 1 and Tab. 2, SPAR3D outperforms all other regressive or generative baselines significantly across most metrics on both datasets. For SSIM, we observe that SPAR3D is slightly worse than the strongest baseline for this metric. We find that this relates to the Monte Carlo noise from our Shader. SPAR3D is also among the fastest reconstruction models with an inference speed of 0.7 seconds per object, which is significantly faster than 3D or multiview diffusion-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. We show qualitative results of different methods in Fig. 5. The reconstructed meshes from pure regression-based approaches such as SF3D or TripoSR align with the input image well, but the backside is often less accurate and over-smoothed. Multi-view diffusion-based methods such as LGM, CRM and InstantMesh show more details on the backside. However, the inconsistency in the synthesized views leads to clear artifacts and overall worse results. Pure generative approaches such as Shap-E and LN3Diff are able to produce sharp surfaces in their generation. However, many details are erroneous hallucinations that do not accurately follow the input images, and the vis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.098, + 0.885, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison. We compare SPAR3D to other state-of-the-art methods visually. SPAR3D not only aligns better with the visible surfaces from images, but also generates higher-quality geometries and textures for the occluded surfaces." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.507, + 0.879, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.726, + 0.908, + 0.768 + ], + "angle": 0, + "content": "Figure 6. Generalization Results. We show qualitative results of SPAR3D on in-the-wild images from 2D generative models (top 2 rows) and ImageNet (bottom 2 rows). The reconstructed meshes exhibit accurate geometric structures with great textures, demonstrating a strong generalization performance of SPAR3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.487, + 0.894 + ], + "angle": 0, + "content": "ible surfaces are often reconstructed incorrectly. Compared to prior art, the meshes produced by SPAR3D not only faithfully resemble the input image, but also exhibit well-generated occluded parts with reasonable details. In Fig. 6, we further show qualitative results of SPAR3D on in-the-wild images. The images are generated using text-to-image generative models or from the ImageNet validation set [14]. The high quality of the reconstructed meshes demonstrates" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.772, + 0.907, + 0.804 + ], + "angle": 0, + "content": "a strong generalization performance of SPAR3D. More results and comparisons are included in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.669, + 0.832 + ], + "angle": 0, + "content": "4.3. Editing Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The usage of explicit point clouds as an intermediate representation enables interactive editing of the generated meshes. Users can easily alter the unseen surface of the mesh by manipulating the point cloud. In Fig. 7, we show a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16866" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.096, + 0.87, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.314, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 7. Editing Results. We show qualitative examples of interactive editing with SPAR3D. On the left two examples, we add a handle to the mug and a tail to the elephant doll by duplicating existing points. On the right two examples, we move or delete points to fix imperfections and to improve local details on the mesh. All the edits are performed in Blender within a minute." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.483, + 0.407 + ], + "angle": 0, + "content": "few editing examples with SPAR3D, either by adding major object parts to the reconstruction, or improving undesirable generated details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.196, + 0.431 + ], + "angle": 0, + "content": "4.4. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.591 + ], + "angle": 0, + "content": "We ablate the key idea of SPAR3D, the point sampling stage, which can be seen as an addition to standard regression approaches. We consider a variant of our model (SPAR3D w/o Point), where we remove the point sampling stage and make SPAR3D a full regressive model. We compare this variant with our full model on both GSO and Omniobject3D. As shown in Tab. 3, our full SPAR3D significantly outperforms the regressive variant, which validates the effectiveness of our design. More ablation experiments are included in the supplement." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.595, + 0.484, + 0.664 + ], + "angle": 0, + "content": "
MethodCD↓FS@0.1↑PSNR↑LPIPS↓
SPAR3D w/o Point0.1360.50618.50.146
SPAR3D0.1200.58418.60.139
SPAR3D w/o Point0.1400.50917.80.146
SPAR3D0.1220.58717.90.140
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Table 3. Ablation Study on GSO (top 2 rows) and Omniobject3D (bottom 2 rows). Removing the point sampling stage leads to significant performance drop." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.194, + 0.743 + ], + "angle": 0, + "content": "4.5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We further design experiments to understand how SPAR3D works. Our key assumption when designing SPAR3D is that the two-stage design effectively separates the uncertain part (back-surface modeling) and the deterministic part (visible surface modeling) of the monocular 3D reconstruction problem. Ideally, the meshing stage should mainly rely on the input image for reconstructing the visible surface, while relying on the point cloud to generate the back surface. To see whether this is true, we design an experiment where we artificially use point clouds that conflict with the input im" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.369, + 0.882, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.535, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Figure 8. Generated Mesh with Conflicting Cues. Under conflicting cues from images and point clouds, our model reconstructs the visible surface based on the image, while generating the backside surface based on the point cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.692 + ], + "angle": 0, + "content": "age. In Fig. 8, we feed the input image of a squirrel and the point cloud of a horse to the meshing model. As shown in the figure, the reconstructed mesh indeed aligns with the squirrel image well on the visible surface, while the back surface mainly adheres to the point cloud. This result validates our assumption." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We present SPAR3D, a simple yet effective approach for single-view 3D reconstruction. The core of our model is a two-stage design based on point sampling. We first generate a sparse point cloud via point diffusion, and then reconstruct a highly detailed mesh from both the point cloud and the image. This design enables us to take the best of regression-based and generative modeling. Evaluated on standard benchmarks and in-the-wild images, SPAR3D significantly outperforms previous state-of-the-art methods with a fast inference speed. We will release our model upon publication, and we hope our effort is useful for future research towards scalable generation of high-quality 3D content." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16867" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.482, + 0.199 + ], + "angle": 0, + "content": "[1] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.483, + 0.256 + ], + "angle": 0, + "content": "[2] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Brent Burley. Physically-based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.286, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[4] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2416-2425, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[8] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[9] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[10] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2262-2272, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20637-20647, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[18] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] James AD Gardner, Bernhard Egger, and William AP Smith. Reni++ a rotation-equivariant, scale-invariant, natural illumination prior. arXiv preprint arXiv:2311.09361, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[20] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[21] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[22] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[25] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Zixuan Huang, Varun Jampani, Anh Thai, Yuanzhen Li, Stefan Stojanov, and James M Rehg. Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "16868" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[27] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10050-10060, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[28] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Zeroshape: Regression-based zero-shot shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10071, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[29] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 397-413. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[30] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[31] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. arXiv preprint arXiv:2403.12019, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[33] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12642-12651, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[34] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wei Xue, Wenhan Luo, et al. Era3d: high-resolution multiview diffusion using efficient row-wise attention. Advances in Neural Information Processing Systems, 37:55975-56000, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[35] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[37] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[40] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[41] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[47] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[48] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[49] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20887-20897, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[50] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16869" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[52] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[53] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[54] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[56] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 2, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[57] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International conference on learning representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[58] Eric Veach. *Robust Monte Carlo Methods for Light Transport Simulation*. PhD thesis, Stanford University, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[59] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. Eurographics Symposium on Rendering, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[60] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[61] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[62] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[63] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[64] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 12" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[65] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. arXiv preprint arXiv:2301.08247, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[66] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[67] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. arXiv preprint arXiv:2301.07525, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[68] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.906, + 0.471 + ], + "angle": 0, + "content": "[70] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.906, + 0.543 + ], + "angle": 0, + "content": "[71] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.908, + 0.599 + ], + "angle": 0, + "content": "[72] Lior Yariv, Omri Puny, Oran Gafni, and Yaron Lipman. Mosaic-sdf for 3d generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4630-4639, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.906, + 0.656 + ], + "angle": 0, + "content": "[73] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.906, + 0.783 + ], + "angle": 0, + "content": "[75] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16870" + } + ] +] \ No newline at end of file diff --git a/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_origin.pdf b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d372edb788def63d103b3ccd32583a3274965c67 --- /dev/null +++ b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/3000205b-bdbd-4656-8ffc-c648c9a58288_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd80f101b3f647b16e66af608b55ddc8d0194c676ff9fb1a4f81e4d320c9eb74 +size 1832301 diff --git a/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/full.md b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6cc65cbec64f2e80c7f24fb8b97db13f27597b52 --- /dev/null +++ b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/full.md @@ -0,0 +1,311 @@ +# SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images + +Zixuan Huang $^{1,2*}$ + +Mark Boss + +Aaryaman Vasishta + +James M. Rehg2 + +Varun Jampani1 + +$^{1}$ Stability AI, $^{2}$ UIUC + +![](images/9c1fc2eab87aa86ef36b6063d9cf3ada5b2199677ce684d135eebf691bcd83d4.jpg) + +![](images/691b91eefc4219c9142139c74942557f2bda43bf28dbfb07af89923f779df4c7.jpg) + +![](images/157a8772aa8ac76751d2f04a0f3454ac4e7fccce74d08c298b65dc5f0b730a5e.jpg) +Input Image + +![](images/1b44ad1086ba85fe7d52370bd8ba146af062a4c7e879c6ba08f36bac750b2ece.jpg) +Edit + +![](images/d79cb85b41d00beb0c0e0e7d8338c2223471fa4a119d8990400b2b61e6a85883.jpg) +Point Cloud + +![](images/e657a7487da64a95e34e9bdea82477b4ca614d5d8665207bc9c1435127cfbf36.jpg) + +![](images/b9967dd5bb81f49ba25954978781718a3489a8061c1eede5afa7dd754eae8a14.jpg) +Mesh Output + +![](images/a8fa67c7d50c8e6a599b049923dea4604e2ba30542cb8d4a05dc5419502da53d.jpg) + +![](images/390dd82fef223989d09b23db0a4da6138d1acc915b12412a9685fa4a85fe2c1f.jpg) + +![](images/5376fd4b94788fec55be741d22454cb5d06c66303eac51d62251dddc02981ab2.jpg) + +![](images/79a1efa409a65d8c3eacb564fe5ee34eac3e6a0841cb1f70bed1ada2fbc7bb09.jpg) +Input Image + +![](images/c1c1c1d187a4aa24464e82bfc1d45d67bbb2d8daf47697b25cb115908f47ac26.jpg) +↓Edit + +![](images/4135ca0970c221c63fa0705b34d60d451f618a73e70628635099279e8bebc364.jpg) +Point Cloud + +![](images/831c9532a3d2ed6a6642bb7253ac18a3d1578aa647a5047ff1120e5f03a267b7.jpg) + +![](images/02e2b00fe7ca057226c39e52d1bdb519ef26669ac772781725a309d0464274f9.jpg) +Mesh Output +Figure 1. We present SPAR3D, a state-of-the-art 3D reconstructor that reconstructs high-quality 3D meshes from single-view images. SPAR3D enjoys a fast reconstruction speed at 0.7 seconds and supports interactive user edits. + +![](images/5967dcec98bef3b008a6e986cb670d5a42cf4cb59150f373e786545c6ffc108f.jpg) + +![](images/6c3287016519b5992976c289a52815721139f85bfebb91681249539c10a1bc3e.jpg) + +# Abstract + +We study the problem of single-image 3D object reconstruction. Recent works have diverged into two directions: regression-based modeling and generative modeling. Regression methods efficiently infer visible surfaces, but struggle with occluded regions. Generative methods handle uncertain regions better by modeling distributions, but are computationally expensive and the generation is often miscaligned with visible surfaces. In this paper, we present SPAR3D, a novel two-stage approach aiming to take the best of both directions. The first stage of SPAR3D generates + +sparse 3D point clouds using a lightweight point diffusion model, which has a fast sampling speed. The second stage uses both the sampled point cloud and the input image to create highly detailed meshes. Our two-stage design enables probabilistic modeling of the ill-posed single-image 3D task while maintaining high computational efficiency and great output fidelity. Using point clouds as an intermediate representation further allows for interactive user edits. Evaluated on diverse datasets, SPAR3D demonstrates superior performance over previous state-of-the-art methods, at an inference speed of 0.7 seconds. Project page with code and model: https://spar3d.github.io + +# 1. Introduction + +Reconstructing 3D objects from monocular images is a fundamental problem in computer vision. An efficient reconstruction system opens up a wide range of applications, including augmented reality, filmmaking, and manufacturing. Monocular 3D reconstruction is also a complex inverse problem: while the visible surface can be estimated from shading, predicting the occluded surface necessitates a strong 3D object prior. Our field has seen a divergence in two different directions: feedforward regression [2, 11, 20, 25, 26, 28, 39, 55, 56, 61, 63, 65, 66, 69, 70, 73] and diffusion-based generation [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75]. Despite the significant progress made in both directions, each has fundamental limitations. + +Regression-based models are highly effective in adhering to the visible surface in the image, and the inference speed is typically fast. However, they make the oversimplified assumption of bijective mapping between images and 3D. This assumption introduces ambiguity in the learning objective, leading to poorly estimated surfaces and textures in occluded regions. On the other hand, diffusion-based approaches are generative and do not predict the statistical mean. However, their iterative sampling at inference time is computationally inefficient when modeling high-resolution 3D. Additionally, previous studies such as [28] indicate that diffusion-generated 3D models exhibit worse alignment to the surface visible in the input image. How can we take the best of both worlds while avoiding their limitations? + +In light of this, we propose SPAR3D, which breaks the 3D reconstruction process down into two stages: the point sampling stage and the meshing stage. The point sampling stage uses diffusion models to generate sparse point clouds, followed by the meshing stage, transforming point clouds into highly detailed meshes. Our main idea is to offload the uncertainty modeling to the point sampling stage, where the low resolution of the point clouds allows rapid iterative sampling. The subsequent meshing stage leverages the local image features to transform the point cloud into a detailed mesh of high output fidelity. Reducing the meshing uncertainty with point clouds further facilitates unsupervised learning of inverse rendering, which reduces the baked-in lighting in the textures. Our two-stage design enables SPAR3D to significantly outperform previous regressive methods, while preserving high computational efficiency and fidelity to input observation. + +A key design choice of our method is the usage of point clouds to connect the two stages. To ensure fast reconstruction, our intermediate representation needs to be lightweight so it can be efficiently generated. However, it should provide enough guidance to the meshing stage. This inspires us to use point clouds, which are perhaps the most computationally efficient 3D representation because all information bits are used to represent the surface. Moreover, the lack of + +connectivity, typically considered as the drawback of point clouds, now turns into an advantage with our two-stage approach for editing purposes. When the back surface does not align with user expectations, local edits can be easily made on the low-resolution point clouds without worrying about topologies (see Fig. 1 bottom). Feeding edited point clouds into the meshing stage produces better meshes tailored towards user requirements. + +Our experiments demonstrate the superiority of SPAR3D over previous state-of-the-art methods, with solid quantitative and qualitative results on various data sources. SPAR3D also exhibits a strong generalization ability to inthe-wild images and AI-generated images. With a total inference time below 0.7 seconds, SPAR3D is not only efficient but also allows for easy user-driven edits, offering a practical solution to the task of monocular 3D reconstruction. We hope that this is a meaningful step towards scalable generation of high-quality 3D assets. + +# 2. Related Work + +Feedforward 3D reconstruction methods address the problem of 3D object reconstruction by learning a feedforward model in a regression-based manner. Earlier works [11, 20, 26, 39, 61, 66, 70] in this field typically predict only the geometry and train on small datasets [5, 52], which limits their generalization ability. Recently, larger 3D datasets [12, 44] have been collected, unlocking the potential to train feedforward 3D models at scale [25, 28, 65]. These models exhibit great generalization ability to unseen images, and excel at producing reconstructions that tightly align with the observed cues in the input image. In particular, LRM [25] and follow-up works [2, 55, 56, 63, 69, 73] show that properly designed large transformer models can be trained using only rendering losses to capture object geometry and texture in great detail. Despite the high fidelity and computational efficiency of these models, the oversimplified bijective assumption in these regressive approaches results in oversmoothed unseen surfaces. Multi-view diffusion models [36, 37, 47, 48] have been considered as a remedy for this, where additional viewpoints are synthesized as input to the feedforward model [55, 63, 69]. However, the inconsistency across viewpoints often leads to significant artifacts on the reconstructed surfaces, and the computational efficiency of these approaches is severely affected by the slow multi-view generation process. Our model also aims to overcome the learning ambiguity in regressive approaches, but our point sampling approach is inherently 3D-consistent and computationally efficient, and further allows easy user edits. + +Generative 3D modeling learns the image-conditioned distribution of 3D assets instead of a deterministic mapping. Early 3D generative works use GAN [4, 17, 29, 57], + +![](images/e8e51b1fbb67bb7db36b5bbe6cfc083785c1ae501f34e37887d3e051ccd0c87d.jpg) +Figure 2. SPAR3D Overview. Conditioned on the input image, SPAR3D first leverages a point diffusion model to generate a sparse point cloud. The triplane transformer then uses the sampled point cloud and image features to produce high-resolution triplane features. The triplane features are then queried to reconstruct the geometry, texture, and illumination of the object in the image. + +normalizing flow [31, 71] or VAE [18, 40, 68] as the generative framework. Inspired by the success of 2D diffusion models [15, 45], 3D diffusion models [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75] have also been extensively explored in recent works. Despite the advantage of probabilistic modeling that avoids over-smoothed results, diffusion-based 3D generation has two drawbacks: 1) not aligning well with input observations, and 2) having low inference speed at high resolution. Our work inherits the advantage of probabilistic modeling, while avoiding the drawbacks by using diffusion to generate only sparse point clouds. + +Optimization-based single-view 3D leverages 2D generative priors to recover 3D from single-view images. These works [13, 21, 38, 54] rely on SDS-type loss [43, 60] and generate 3D assets by optimizing for each object image separately. These methods achieve promising results without large-scale annotation. However, the lack of a strong explicit 3D prior makes the optimization process inefficient and prone to local minima. + +# 3. Method + +SPAR3D Overview. Given the input image $I \in \mathbb{R}^{3 \times h \times w}$ , our method produces a 3D mesh with PBR materials, including albedo, metallic, roughness and surface normals. The main goal of our work is to develop a model that enjoys the benefits of distribution learning through diffusion models, while not suffering from the low output fidelity and computational inefficiency. To this end, we design a two-stage model that consists of the point sampling stage and the meshing stage (see Fig. 2). At the point sampling stage, a point diffusion model learns the conditional distribution of point clouds given the input image. This stage is computationally efficient given the low resolution of the point clouds. The regression-based meshing stage transforms the sampled point cloud into a highly detailed mesh that aligns + +with the visible surface. The reduced uncertainty with point sampling further facilitates the learning of materials and illumination in an unsupervised manner during the meshing stage. This reduces baked-in lighting artifacts and results in better modeling of specular surfaces. Finally, by using sparse point clouds as the intermediate representation, SPAR3D enables human editing in the loop. + +# 3.1. Point Sampling Stage + +Overview. The point sampling stage produces a sparse point cloud as the input to the meshing stage. The core of the point sampling stage is a point diffusion model, which generates point clouds $\pmb{p}_0 \in \mathbb{R}^{n \times 6}$ conditioned on the input image $I$ . The six channels include three XYZ channels and three RGB channels. In our work, the resolution of the point cloud $n$ is set to 512. + +**Point Diffusion Framework.** Our diffusion framework is based on DDPM [24], which consists of two processes: 1) the forward process which adds noise to the original point cloud, and 2) the backward process where the denoiser learns to remove the noise. At timestep $t \in [0,T]$ , the diffusion process combines Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ with a point cloud $p_0$ as + +$$ +\boldsymbol {p} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {p} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}, \tag {1} +$$ + +where $\bar{\alpha}_t$ denotes the noise schedule. We use the sigmoid noise schedule proposed in [7], combined with input scaling and the renormalization trick. The denoiser $\pmb{\epsilon}_{\theta}(\pmb{p}_t, t; \pmb{c})$ then learns to recover the noise from $\pmb{p}_t$ and is supervised by + +$$ +L _ {s i m p l e} (\theta) = \mathbb {E} _ {t, \boldsymbol {p} _ {0}, \boldsymbol {\epsilon}} \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {p} _ {t}, t; \boldsymbol {c}) \| _ {2} ^ {2}. \tag {2} +$$ + +Here $c$ denotes the image condition tokens. During inference, we use the DDIM sampler [51] to generate point cloud samples. Samples generated directly often align poorly with the condition, hence we use the classifier-free guidance (CFG) [23] to improve sampling fidelity. + +Denoiser Design. We use a transformer denoiser similar to Point-E [41], where the noisy point cloud $\pmb{p}_t \in \mathbb{R}^{n \times 6}$ is linearly mapped to a set of point tokens $\pmb{x} \in \mathbb{R}^{n \times d}$ . We use DINOv2 [42] to encode the input image $I$ as conditioning tokens $\pmb{c} \in \mathbb{R}^{c \times d}$ . The conditions and the point tokens are then concatenated together as input to the transformer, which predicts the added noise on each point. + +Albedo Point clouds. In the meshing stage, we estimate the materials and lighting alongside the geometry. However, this decomposition is inherently ambiguous because there are countless combinations of lighting and albedo that can explain the same input image. It is challenging to learn this highly uncertain decomposition during the regressive meshing stage alone. We therefore reduce the uncertainty at the point sampling stage, by directly generating albedo point clouds with diffusion models. Sampling albedo point clouds as input to the meshing stage drastically reduces the ambiguity of inverse rendering and stabilizes the decomposition learning. + +# 3.2. Meshing Stage + +Overview. The meshing stage produces a textured mesh from the input image and the point cloud. The backbone of our meshing model is a large triplane transformer, which predicts triplane features from the image and point cloud conditions. We estimate the geometry, texture and lighting of the current object from the triplane, and metallic/roughness from the image features. The geometry and materials are fed into our differentiable renderer during training, so that we can apply rendering loss to supervise our model. + +Triplane Transformer. Our triplane transformer consists of three submodules: the point cloud encoder, the image encoder, and the transformer backbone. We use a simple transformer encoder to encode the point cloud as a set of point tokens. Given the low resolution of the point clouds, each point can be directly mapped to a single token. Our image encoder is DINOv2 [42], which produces local image embeddings. Our triplane transformer follows a similar design to PointInfinity [27] and SF3D [2], which produces triplanes at high resolution of $384 \times 384$ by using a computationally-detached two-stream design. + +Surface Estimation. To estimate the geometry, the triplanes are queried with a shallow MLP to produce density values. Similar to [2, 63, 69], we convert the implicit density field to explicit surface using differentiable Marching Tetrahedron (DMTet) [46]. We additionally use two MLP heads to predict vertex offsets and surface normals together with density. These two attributes reduce the artifacts introduced by the Marching Tetrahedron and lead to locally smoother surfaces. + +Material and Illumination Estimation. We perform inverse rendering and jointly estimate materials (albedo, metallic and roughness) and illumination alongside the geometry. The task is highly ill-posed and Neural-PIL [1] showed that an illumination prior can reduce the ambiguity. We build our illumination estimator upon the learning-based illumination prior from $\mathrm{RENI}++$ [19]. $\mathrm{RENI}++$ is originally an unconditional generative model for HDR illumination generation. We learn an encoder to map triplane features into the latent space of $\mathrm{RENI}++$ . This allows us to estimate the environment illumination in the input image. The albedo is estimated from triplane similar to geometry, where a shallow MLP predicts the albedo value for each 3D location. For metallic and roughness, we follow SF3D [2] and learn to estimate them with a probabilistic approach via a Beta prior. We find that the CLIP encoder used in SF3D is unstable when the object size changes. We therefore replace their CLIP encoder with AlphaCLIP [53] to alleviate this issue using foreground object masks. + +Differentiable Rendering. We implement a differentiable renderer that renders images based on the predicted environment map, PBR materials and geometry surface (see Fig. 3). We use a differentiable mesh rasterizer and add a differentiable Shader. Specifically, we leverage the standard simplified Disney PBR model [3] in our Shader. As we use $\mathrm{RENI}++$ to reconstruct environment maps, we need to explicitly integrate the incoming radiance. Here, we opt to use the Monte Carlo Integration. Given the low sample counts we can computationally afford during training, we rely on Multiple Importance Sampling (MIS) with the balanced heuristic [58] to reduce integration variance. Additionally, to better model the self-occlusion which has been typically ignored in prior works, we implement a visibility test for better shadow modeling. We take inspiration from real-time graphics and model the visibility test as a screen-space method using the depth map from our rasterizer. An overview of this test is shown in Fig. 4. Specifically, we ray-march a short distance (0.25) in 6 steps for all proposed sample directions from MIS, and project the position back to image space. If the current ray depth is farther away than the sampled value from the depth map, then the ray is marked as shadowed. + +Loss Function. Our main loss function is the rendering loss that compares renderings from novel views to the groundtruth (GT) images. Specifically, our rendering loss is a linear combination of 1) the L2 distance between the rendered and GT images, 2) the perceptual distance between the rendered and GT images measured by LPIPS [74], and 3) the L2 distance between the rendered opacity and the GT foreground mask. Apart from the rendering loss, we also follow SF3D and apply the mesh and shading regularization that regularizes the surface smoothness and the inverse rendering respectively. + +![](images/2d0c167b5544547dcd492ec0275ffd4ea7eb036926a4c0b28c2939f01796ff3a.jpg) +Figure 3. Our Differentiable Renderer. We estimate geometry, albedo, lighting, and normal maps from the triplane and metallic/roughness values from the image. We rasterize and interpolate these values as input to our shader (omitted here for simplicity). Our shader uses the Disney BRDF [3] and performs Monte Carlo integration. We further perform visibility testing to improve shadow modeling. Finally, we compare the rendered image with the GT image and minimize the rendering loss. + +![](images/076aad9651cf70aa0dd5d76755f6ed3aca2ab82ac14276932a594066fcc5f2a5.jpg) +Figure 4. Shadow Modeling. We perform visibility testing in screen-space by marching along sampled rays. If any point along the ray has a ray depth which is farther away than the depth map, we consider the entire ray as shadowed. + +# 3.3. Interactive Editing + +A unique advantage of our two-stage design is that it naturally supports interactive editing of unseen regions in our produced mesh. In most circumstances, the visible surface is determined by the input image and remains highly accurate, while the unseen surface is mainly based on the sampled point cloud, which might not align with user intention. In this case, editing the unseen surface of the mesh is feasible by altering the point cloud. Point clouds are perhaps one of the most flexible 3D representation for editing purposes because there are no topology constraints. Given the low resolution of our point clouds, editing the point cloud is fairly efficient and intuitive. Users can easily delete, duplicate, stretch or recolor points in the point cloud. Our efficient meshing model is able to produce the adjusted mesh in 0.3 seconds, which makes this process fairly interactive. + +# 3.4. Implementation Details + +Point Sampling Stage. Our point diffusion model has 16 transformer blocks in total. Each transformer block consists of two Layer Normalization layer, one Multi-Head Attention (MHA) layer and one MLP. We use a feature dimension of 1024 and 16 attention heads in each MHA layer. With many emissive objects in our dataset, albedo can be visually distinct from the input image and hard to learn. Therefore, instead of directly generating albedo point clouds in the point sampling stage, we learn to generate white-lit point clouds as a proxy target. + +Meshing Stage. Our triplane transformer consists of 4 two-stream blocks [27]. Each two-stream block consists of three self-attention and two cross-attentions. The main computation is carried out using 3,072 latent tokens, each with a feature dimension of 1024. The MHA includes 16 attention heads. The point cloud encoder is a vanilla transformer with 12 layers and 512 feature dimension, and the + +image encoder is DINOv2-large. We use a tetrahedra resolution of 160 for DMTet. In our differentiable Shader, we follow Hasselgren et al. [22] and use a detached biased sampling scheme. We sample based on the specular lobe (GGX [59]), the 2D piecewise-linear distribution of the environment map luminance and the hemispherical distribution. Specifically, we include 6 samples from the GGX lobe, 6 samples from the 2D piecewise-linear distribution of the luminance, and 4 samples from the hemispherical distribution. The main body of the shader is implemented in PyTorch, while the screen-space shadowing and the 2D piecewise-linear distribution computation for the environment map are implemented as custom CUDA kernels for efficiency. The training of our meshing stage includes multiple phases, where we increase the rendering resolution and decrease the batch size at later training phases. We use GT point clouds as input when training the meshing model. The curation of our training data follows TripoSR [56]. + +# 4. Experiments + +# 4.1. Evaluation + +Datasets. We used two datasets for evaluation, GSO [16] and OmniObject3D [67]. We follow TripoSR [56] and remove simple box or cylindrical objects to avoid bias on simple geometries. Each of the evaluation sets consists of around 250 objects. We render the objects with diverse azimuth angles at different elevations, with randomly sampled HDRI environment maps. We also vary the focal length of the camera to create more diverse test cases. + +Metrics. To evaluate the geometry quality of the reconstructed meshes, we use follow prior works [56, 69] and use Chamfer Distance (CD) and F-score (FS) as our evaluation metrics. CD measures the alignment between two point clouds and is defined as the average of accuracy and + +
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2040.3590.6380.92215.30.8020.2053.1
LN3Diff [32]0.1740.4220.7030.94917.10.8190.1695.1
LGM [55]0.1960.3560.6350.93617.00.8180.18441.0
CRM [63]0.1610.4370.7350.96117.50.8300.1697.4
3DTopia-XL [8]0.2050.3830.6510.91316.00.8180.19340.6
Era3D [34]0.1810.4170.6900.94116.90.8180.173348.0
TripoSR [56]0.1450.5010.7840.96818.50.8370.1510.2
InstantMesh [69]0.1350.5450.8120.97118.10.8380.14636.1
SF3D [2]0.1370.5400.8060.97018.00.8390.1450.3
SPAR3D (ours)0.1200.5840.8500.98318.60.8360.1390.7
+ +Table 1. Quantitative Comparisons on GSO [16]. SPAR3D performs favorably to other state-of-the-art methods. + +
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2120.3490.6240.90914.80.80060.2053.1
LN3Diff [32]0.1600.4800.7440.95716.70.8190.1615.0
LGM [55]0.2000.3660.6380.92416.10.8100.18842.0
CRM [63]0.1550.4820.7650.96217.00.8280.1627.0
3DTopia-XL [8]0.1910.4070.6760.93115.80.8170.18641.2
Era3D [34]0.1850.4200.6850.93716.40.8160.167391.7
TripoSR [56]0.1440.5370.7850.96318.00.8350.1470.2
InstantMesh [69]0.1450.5460.7900.96217.20.8320.15034.7
SF3D [2]0.1380.5540.8000.96717.40.8360.1450.3
SPAR3D (ours)0.1220.5870.8450.97817.90.8320.1400.7
+ +Table 2. Quantitative Comparisons on OmniObject3D [67]. SPAR3D performs favorably to other state-of-the-art methods. + +completeness: + +$$ +d \left(S _ {1}, S _ {2}\right) = \frac {1}{2 \left| S _ {1} \right|} \sum_ {x \in S _ {1}} \min _ {y \in S _ {2}} \| x - y \| _ {2} + \frac {1}{2 \left| S _ {2} \right|} \sum_ {y \in S _ {2}} \min _ {x \in S _ {1}} \| x - y \| _ {2} \tag {3} +$$ + +FS evaluates point cloud alignment by calculating the F-score with a predefined threshold. Predicted points that lie within the distance threshold are considered as correct predictions. A higher FS means better alignment between the reconstructed shape and the groundtruth. To evaluate the texture quality, we compute standard image metrics, including PSNR, SSIM and LPIPS, between images rendered from the predicted mesh and the groundtruth images. + +Protocol. To calculate the metrics that are comparable across methods, the meshes need to lie in the same coordinate system. To this end, we perform brute-force search in rotations to align each predicted mesh with the groundtruth mesh. Both the prediction and the groundtruth are normalized before the brute-force alignment, and the alignment is further refined with ICP. + +Baselines. We compare SPAR3D with other efficient methods for single-view 3D generation or reconstruction [2, 55, 56, 63, 69]. We use the official implementation for all baselines, and we evaluate the produced meshes under the same protocol. Specifically, we compare against TripoSR [56], LGM [55], CRM [63], InstantMesh [69], LN3Diff [32], Shap-E [30] and SF3D [2]. Among these baselines, TripoSR and SF3D are pure regression-based ap + +proaches; LGM, CRM and InstantMesh use multiview diffusion to generate pseudo multi-view images; LN3Diff and Shap-E are purely diffusion-based 3D generative models. + +# 4.2. Main Results + +Quantitative Comparison. We compare SPAR3D to other baselines on GSO and Omniobject3D quantitatively. As shown in Tab. 1 and Tab. 2, SPAR3D outperforms all other regressive or generative baselines significantly across most metrics on both datasets. For SSIM, we observe that SPAR3D is slightly worse than the strongest baseline for this metric. We find that this relates to the Monte Carlo noise from our Shader. SPAR3D is also among the fastest reconstruction models with an inference speed of 0.7 seconds per object, which is significantly faster than 3D or multiview diffusion-based approaches. + +Qualitative Results. We show qualitative results of different methods in Fig. 5. The reconstructed meshes from pure regression-based approaches such as SF3D or TripoSR align with the input image well, but the backside is often less accurate and over-smoothed. Multi-view diffusion-based methods such as LGM, CRM and InstantMesh show more details on the backside. However, the inconsistency in the synthesized views leads to clear artifacts and overall worse results. Pure generative approaches such as Shap-E and LN3Diff are able to produce sharp surfaces in their generation. However, many details are erroneous hallucinations that do not accurately follow the input images, and the vis + +![](images/1a3ecfe6b0ad1bbd0694f064943ef140b23ffe23266785a108909941e3fcb572.jpg) +Figure 5. Qualitative Comparison. We compare SPAR3D to other state-of-the-art methods visually. SPAR3D not only aligns better with the visible surfaces from images, but also generates higher-quality geometries and textures for the occluded surfaces. + +![](images/3823312c952490db487fbafc77a9068d71383a5ca938984e747f2f70081bd17e.jpg) +Figure 6. Generalization Results. We show qualitative results of SPAR3D on in-the-wild images from 2D generative models (top 2 rows) and ImageNet (bottom 2 rows). The reconstructed meshes exhibit accurate geometric structures with great textures, demonstrating a strong generalization performance of SPAR3D. + +ible surfaces are often reconstructed incorrectly. Compared to prior art, the meshes produced by SPAR3D not only faithfully resemble the input image, but also exhibit well-generated occluded parts with reasonable details. In Fig. 6, we further show qualitative results of SPAR3D on in-the-wild images. The images are generated using text-to-image generative models or from the ImageNet validation set [14]. The high quality of the reconstructed meshes demonstrates + +a strong generalization performance of SPAR3D. More results and comparisons are included in the supplement. + +# 4.3. Editing Results + +The usage of explicit point clouds as an intermediate representation enables interactive editing of the generated meshes. Users can easily alter the unseen surface of the mesh by manipulating the point cloud. In Fig. 7, we show a + +![](images/886233bcc6fa5c6c13ed349270b3480443d90cb0c90d667d6887c45c7eb55b42.jpg) +Figure 7. Editing Results. We show qualitative examples of interactive editing with SPAR3D. On the left two examples, we add a handle to the mug and a tail to the elephant doll by duplicating existing points. On the right two examples, we move or delete points to fix imperfections and to improve local details on the mesh. All the edits are performed in Blender within a minute. + +few editing examples with SPAR3D, either by adding major object parts to the reconstruction, or improving undesirable generated details. + +# 4.4. Ablation + +We ablate the key idea of SPAR3D, the point sampling stage, which can be seen as an addition to standard regression approaches. We consider a variant of our model (SPAR3D w/o Point), where we remove the point sampling stage and make SPAR3D a full regressive model. We compare this variant with our full model on both GSO and Omniobject3D. As shown in Tab. 3, our full SPAR3D significantly outperforms the regressive variant, which validates the effectiveness of our design. More ablation experiments are included in the supplement. + +
MethodCD↓FS@0.1↑PSNR↑LPIPS↓
SPAR3D w/o Point0.1360.50618.50.146
SPAR3D0.1200.58418.60.139
SPAR3D w/o Point0.1400.50917.80.146
SPAR3D0.1220.58717.90.140
+ +Table 3. Ablation Study on GSO (top 2 rows) and Omniobject3D (bottom 2 rows). Removing the point sampling stage leads to significant performance drop. + +# 4.5. Analysis + +We further design experiments to understand how SPAR3D works. Our key assumption when designing SPAR3D is that the two-stage design effectively separates the uncertain part (back-surface modeling) and the deterministic part (visible surface modeling) of the monocular 3D reconstruction problem. Ideally, the meshing stage should mainly rely on the input image for reconstructing the visible surface, while relying on the point cloud to generate the back surface. To see whether this is true, we design an experiment where we artificially use point clouds that conflict with the input im + +![](images/7484dced574c55277f64c202759d1a6d5ef3780359e9971a9cfac87fa1eb588f.jpg) +Figure 8. Generated Mesh with Conflicting Cues. Under conflicting cues from images and point clouds, our model reconstructs the visible surface based on the image, while generating the backside surface based on the point cloud. + +age. In Fig. 8, we feed the input image of a squirrel and the point cloud of a horse to the meshing model. As shown in the figure, the reconstructed mesh indeed aligns with the squirrel image well on the visible surface, while the back surface mainly adheres to the point cloud. This result validates our assumption. + +# 5. Conclusion + +We present SPAR3D, a simple yet effective approach for single-view 3D reconstruction. The core of our model is a two-stage design based on point sampling. We first generate a sparse point cloud via point diffusion, and then reconstruct a highly detailed mesh from both the point cloud and the image. This design enables us to take the best of regression-based and generative modeling. Evaluated on standard benchmarks and in-the-wild images, SPAR3D significantly outperforms previous state-of-the-art methods with a fast inference speed. We will release our model upon publication, and we hope our effort is useful for future research towards scalable generation of high-quality 3D content. + +# References + +[1] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021. 4 +[2] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint, 2024. 2, 4, 6 +[3] Brent Burley. Physically-based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 4, 5 +[4] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 2 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2 +[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2416-2425, 2023. 2, 3 +[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 3 +[8] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024.6 +[9] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3 +[10] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2262-2272, 2023. 2, 3 +[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 2 +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2 +[13] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. + +Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20637-20647, 2023. 3 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3 +[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5, 6 +[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2 +[18] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3 +[19] James AD Gardner, Bernhard Egger, and William AP Smith. Reni++ a rotation-equivariant, scale-invariant, natural illumination prior. arXiv preprint arXiv:2311.09361, 2023. 4 +[20] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[21] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 3 +[22] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 5 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[25] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[26] Zixuan Huang, Varun Jampani, Anh Thai, Yuanzhen Li, Stefan Stojanov, and James M Rehg. Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, 2023. 2 +[27] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10050-10060, 2024. 2, 3, 4, 5 +[28] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Zeroshape: Regression-based zero-shot shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10071, 2024. 2 +[29] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 397-413. Springer, 2020. 2 +[30] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6 +[31] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 3 +[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. arXiv preprint arXiv:2403.12019, 2024. 2, 3, 6 +[33] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12642-12651, 2023. 2, 3 +[34] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wei Xue, Wenhan Luo, et al. Era3d: high-resolution multiview diffusion using efficient row-wise attention. Advances in Neural Information Processing Systems, 37:55975-56000, 2024. 6 +[35] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885, 2023. 2, 3 +[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2 +[37] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3 +[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 3 +[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: + +Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2 +[40] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 3 +[41] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3, 4 +[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4 +[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 3 +[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 4 +[47] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model, 2023. 2 +[48] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3 +[49] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20887-20897, 2023. +[50] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3 +[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3 + +[52] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2 +[53] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want, 2023. 4 +[54] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 3 +[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 6 +[56] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 2, 5, 6, 12 +[57] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International conference on learning representations, 2018. 2 +[58] Eric Veach. *Robust Monte Carlo Methods for Light Transport Simulation*. PhD thesis, Stanford University, 1997. 4 +[59] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. Eurographics Symposium on Rendering, 2007. 5 +[60] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3 +[61] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2 +[62] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 13 +[63] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 2, 4, 6 +[64] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 12 + +[65] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. arXiv preprint arXiv:2301.08247, 2023. 2 +[66] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 2 +[67] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. arXiv preprint arXiv:2301.07525, 2023. 5, 6 +[68] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3 +[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4, 5, 6 +[70] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 2 +[71] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3 +[72] Lior Yariv, Omri Puny, Oran Gafni, and Yaron Lipman. Mosaic-sdf for 3d generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4630-4639, 2024. 2, 3 +[73] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2 +[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4 +[75] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 3 \ No newline at end of file diff --git a/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/images.zip b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a9b0d2ed0c7662b81dd711ef70813b0c99e53070 --- /dev/null +++ b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e829fb0115d843f363174b2eab27b30ffcebbb07745c875a47e1c4c59194dc41 +size 646576 diff --git a/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/layout.json b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8cc1fa2a964c4120bd81520c33b949f12077bd6 --- /dev/null +++ b/2025/SPAR3D_ Stable Point-Aware Reconstruction of 3D Objects from Single Images/layout.json @@ -0,0 +1,7944 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 142, + 164, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 142, + 164, + 159 + ], + "spans": [ + { + "bbox": [ + 76, + 142, + 164, + 159 + ], + "type": "text", + "content": "Zixuan Huang" + }, + { + "bbox": [ + 76, + 142, + 164, + 159 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 143, + 235, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 143, + 235, + 156 + ], + "spans": [ + { + "bbox": [ + 176, + 143, + 235, + 156 + ], + "type": "text", + "content": "Mark Boss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 143, + 347, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 143, + 347, + 157 + ], + "spans": [ + { + "bbox": [ + 248, + 143, + 347, + 157 + ], + "type": "text", + "content": "Aaryaman Vasishta" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 360, + 143, + 442, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 143, + 442, + 157 + ], + "spans": [ + { + "bbox": [ + 360, + 143, + 442, + 157 + ], + "type": "text", + "content": "James M. Rehg2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 454, + 143, + 532, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 143, + 532, + 157 + ], + "spans": [ + { + "bbox": [ + 454, + 143, + 532, + 157 + ], + "type": "text", + "content": "Varun Jampani1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 252, + 157, + 356, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 157, + 356, + 171 + ], + "spans": [ + { + "bbox": [ + 252, + 157, + 356, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 252, + 157, + 356, + 171 + ], + "type": "text", + "content": "Stability AI, " + }, + { + "bbox": [ + 252, + 157, + 356, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 252, + 157, + 356, + 171 + ], + "type": "text", + "content": "UIUC" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 72, + 199, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 72, + 199, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 72, + 199, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 72, + 199, + 545, + 376 + ], + "type": "image", + "image_path": "9c1fc2eab87aa86ef36b6063d9cf3ada5b2199677ce684d135eebf691bcd83d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 77, + 384, + 95, + 429 + ], + "blocks": [ + { + "bbox": [ + 77, + 384, + 95, + 429 + ], + "lines": [ + { + "bbox": [ + 77, + 384, + 95, + 429 + ], + "spans": [ + { + "bbox": [ + 77, + 384, + 95, + 429 + ], + "type": "image", + "image_path": "691b91eefc4219c9142139c74942557f2bda43bf28dbfb07af89923f779df4c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 78, + 444, + 95, + 487 + ], + "blocks": [ + { + "bbox": [ + 78, + 444, + 95, + 487 + ], + "lines": [ + { + "bbox": [ + 78, + 444, + 95, + 487 + ], + "spans": [ + { + "bbox": [ + 78, + 444, + 95, + 487 + ], + "type": "image", + "image_path": "157a8772aa8ac76751d2f04a0f3454ac4e7fccce74d08c298b65dc5f0b730a5e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 490, + 108, + 501 + ], + "lines": [ + { + "bbox": [ + 62, + 490, + 108, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 108, + 501 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 136, + 386, + 157, + 428 + ], + "blocks": [ + { + "bbox": [ + 136, + 386, + 157, + 428 + ], + "lines": [ + { + "bbox": [ + 136, + 386, + 157, + 428 + ], + "spans": [ + { + "bbox": [ + 136, + 386, + 157, + 428 + ], + "type": "image", + "image_path": "1b44ad1086ba85fe7d52370bd8ba146af062a4c7e879c6ba08f36bac750b2ece.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 430, + 164, + 442 + ], + "lines": [ + { + "bbox": [ + 143, + 430, + 164, + 442 + ], + "spans": [ + { + "bbox": [ + 143, + 430, + 164, + 442 + ], + "type": "text", + "content": "Edit" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 135, + 444, + 157, + 484 + ], + "blocks": [ + { + "bbox": [ + 135, + 444, + 157, + 484 + ], + "lines": [ + { + "bbox": [ + 135, + 444, + 157, + 484 + ], + "spans": [ + { + "bbox": [ + 135, + 444, + 157, + 484 + ], + "type": "image", + "image_path": "d79cb85b41d00beb0c0e0e7d8338c2223471fa4a119d8990400b2b61e6a85883.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 490, + 168, + 500 + ], + "lines": [ + { + "bbox": [ + 122, + 490, + 168, + 500 + ], + "spans": [ + { + "bbox": [ + 122, + 490, + 168, + 500 + ], + "type": "text", + "content": "Point Cloud" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 202, + 384, + 223, + 429 + ], + "blocks": [ + { + "bbox": [ + 202, + 384, + 223, + 429 + ], + "lines": [ + { + "bbox": [ + 202, + 384, + 223, + 429 + ], + "spans": [ + { + "bbox": [ + 202, + 384, + 223, + 429 + ], + "type": "image", + "image_path": "e657a7487da64a95e34e9bdea82477b4ca614d5d8665207bc9c1435127cfbf36.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 203, + 441, + 222, + 484 + ], + "blocks": [ + { + "bbox": [ + 203, + 441, + 222, + 484 + ], + "lines": [ + { + "bbox": [ + 203, + 441, + 222, + 484 + ], + "spans": [ + { + "bbox": [ + 203, + 441, + 222, + 484 + ], + "type": "image", + "image_path": "b9967dd5bb81f49ba25954978781718a3489a8061c1eede5afa7dd754eae8a14.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 490, + 266, + 500 + ], + "lines": [ + { + "bbox": [ + 215, + 490, + 266, + 500 + ], + "spans": [ + { + "bbox": [ + 215, + 490, + 266, + 500 + ], + "type": "text", + "content": "Mesh Output" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 262, + 384, + 284, + 430 + ], + "blocks": [ + { + "bbox": [ + 262, + 384, + 284, + 430 + ], + "lines": [ + { + "bbox": [ + 262, + 384, + 284, + 430 + ], + "spans": [ + { + "bbox": [ + 262, + 384, + 284, + 430 + ], + "type": "image", + "image_path": "a8fa67c7d50c8e6a599b049923dea4604e2ba30542cb8d4a05dc5419502da53d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 261, + 441, + 283, + 487 + ], + "blocks": [ + { + "bbox": [ + 261, + 441, + 283, + 487 + ], + "lines": [ + { + "bbox": [ + 261, + 441, + 283, + 487 + ], + "spans": [ + { + "bbox": [ + 261, + 441, + 283, + 487 + ], + "type": "image", + "image_path": "390dd82fef223989d09b23db0a4da6138d1acc915b12412a9685fa4a85fe2c1f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 323, + 383, + 360, + 429 + ], + "blocks": [ + { + "bbox": [ + 323, + 383, + 360, + 429 + ], + "lines": [ + { + "bbox": [ + 323, + 383, + 360, + 429 + ], + "spans": [ + { + "bbox": [ + 323, + 383, + 360, + 429 + ], + "type": "image", + "image_path": "5376fd4b94788fec55be741d22454cb5d06c66303eac51d62251dddc02981ab2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 323, + 441, + 360, + 488 + ], + "blocks": [ + { + "bbox": [ + 323, + 441, + 360, + 488 + ], + "lines": [ + { + "bbox": [ + 323, + 441, + 360, + 488 + ], + "spans": [ + { + "bbox": [ + 323, + 441, + 360, + 488 + ], + "type": "image", + "image_path": "79a1efa409a65d8c3eacb564fe5ee34eac3e6a0841cb1f70bed1ada2fbc7bb09.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 490, + 364, + 500 + ], + "lines": [ + { + "bbox": [ + 318, + 490, + 364, + 500 + ], + "spans": [ + { + "bbox": [ + 318, + 490, + 364, + 500 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 383, + 386, + 418, + 427 + ], + "blocks": [ + { + "bbox": [ + 383, + 386, + 418, + 427 + ], + "lines": [ + { + "bbox": [ + 383, + 386, + 418, + 427 + ], + "spans": [ + { + "bbox": [ + 383, + 386, + 418, + 427 + ], + "type": "image", + "image_path": "c1c1c1d187a4aa24464e82bfc1d45d67bbb2d8daf47697b25cb115908f47ac26.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 430, + 414, + 442 + ], + "lines": [ + { + "bbox": [ + 393, + 430, + 414, + 442 + ], + "spans": [ + { + "bbox": [ + 393, + 430, + 414, + 442 + ], + "type": "text", + "content": "↓Edit" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 381, + 445, + 419, + 486 + ], + "blocks": [ + { + "bbox": [ + 381, + 445, + 419, + 486 + ], + "lines": [ + { + "bbox": [ + 381, + 445, + 419, + 486 + ], + "spans": [ + { + "bbox": [ + 381, + 445, + 419, + 486 + ], + "type": "image", + "image_path": "4135ca0970c221c63fa0705b34d60d451f618a73e70628635099279e8bebc364.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 490, + 422, + 499 + ], + "lines": [ + { + "bbox": [ + 376, + 490, + 422, + 499 + ], + "spans": [ + { + "bbox": [ + 376, + 490, + 422, + 499 + ], + "type": "text", + "content": "Point Cloud" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 441, + 386, + 475, + 430 + ], + "blocks": [ + { + "bbox": [ + 441, + 386, + 475, + 430 + ], + "lines": [ + { + "bbox": [ + 441, + 386, + 475, + 430 + ], + "spans": [ + { + "bbox": [ + 441, + 386, + 475, + 430 + ], + "type": "image", + "image_path": "831c9532a3d2ed6a6642bb7253ac18a3d1578aa647a5047ff1120e5f03a267b7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 441, + 446, + 475, + 491 + ], + "blocks": [ + { + "bbox": [ + 441, + 446, + 475, + 491 + ], + "lines": [ + { + "bbox": [ + 441, + 446, + 475, + 491 + ], + "spans": [ + { + "bbox": [ + 441, + 446, + 475, + 491 + ], + "type": "image", + "image_path": "02e2b00fe7ca057226c39e52d1bdb519ef26669ac772781725a309d0464274f9.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 491, + 515, + 500 + ], + "lines": [ + { + "bbox": [ + 466, + 491, + 515, + 500 + ], + "spans": [ + { + "bbox": [ + 466, + 491, + 515, + 500 + ], + "type": "text", + "content": "Mesh Output" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 513, + 553, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 513, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 553, + 536 + ], + "type": "text", + "content": "Figure 1. We present SPAR3D, a state-of-the-art 3D reconstructor that reconstructs high-quality 3D meshes from single-view images. SPAR3D enjoys a fast reconstruction speed at 0.7 seconds and supports interactive user edits." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 505, + 384, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 505, + 384, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 505, + 384, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 505, + 384, + 544, + 430 + ], + "type": "image", + "image_path": "5967dcec98bef3b008a6e986cb670d5a42cf4cb59150f373e786545c6ffc108f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 503, + 444, + 544, + 490 + ], + "blocks": [ + { + "bbox": [ + 503, + 444, + 544, + 490 + ], + "lines": [ + { + "bbox": [ + 503, + 444, + 544, + 490 + ], + "spans": [ + { + "bbox": [ + 503, + 444, + 544, + 490 + ], + "type": "image", + "image_path": "6c3287016519b5992976c289a52815721139f85bfebb91681249539c10a1bc3e.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 152, + 545, + 200, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 545, + 200, + 558 + ], + "spans": [ + { + "bbox": [ + 152, + 545, + 200, + 558 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 691 + ], + "type": "text", + "content": "We study the problem of single-image 3D object reconstruction. Recent works have diverged into two directions: regression-based modeling and generative modeling. Regression methods efficiently infer visible surfaces, but struggle with occluded regions. Generative methods handle uncertain regions better by modeling distributions, but are computationally expensive and the generation is often miscaligned with visible surfaces. In this paper, we present SPAR3D, a novel two-stage approach aiming to take the best of both directions. The first stage of SPAR3D generates" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 546, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 689 + ], + "type": "text", + "content": "sparse 3D point clouds using a lightweight point diffusion model, which has a fast sampling speed. The second stage uses both the sampled point cloud and the input image to create highly detailed meshes. Our two-stage design enables probabilistic modeling of the ill-posed single-image 3D task while maintaining high computational efficiency and great output fidelity. Using point clouds as an intermediate representation further allows for interactive user edits. Evaluated on diverse datasets, SPAR3D demonstrates superior performance over previous state-of-the-art methods, at an inference speed of 0.7 seconds. Project page with code and model: https://spar3d.github.io" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 157, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 157, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 157, + 713 + ], + "type": "text", + "content": "*Work done at Stability AI." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "16860" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "type": "text", + "content": "Reconstructing 3D objects from monocular images is a fundamental problem in computer vision. An efficient reconstruction system opens up a wide range of applications, including augmented reality, filmmaking, and manufacturing. Monocular 3D reconstruction is also a complex inverse problem: while the visible surface can be estimated from shading, predicting the occluded surface necessitates a strong 3D object prior. Our field has seen a divergence in two different directions: feedforward regression [2, 11, 20, 25, 26, 28, 39, 55, 56, 61, 63, 65, 66, 69, 70, 73] and diffusion-based generation [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75]. Despite the significant progress made in both directions, each has fundamental limitations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 413 + ], + "type": "text", + "content": "Regression-based models are highly effective in adhering to the visible surface in the image, and the inference speed is typically fast. However, they make the oversimplified assumption of bijective mapping between images and 3D. This assumption introduces ambiguity in the learning objective, leading to poorly estimated surfaces and textures in occluded regions. On the other hand, diffusion-based approaches are generative and do not predict the statistical mean. However, their iterative sampling at inference time is computationally inefficient when modeling high-resolution 3D. Additionally, previous studies such as [28] indicate that diffusion-generated 3D models exhibit worse alignment to the surface visible in the input image. How can we take the best of both worlds while avoiding their limitations?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 617 + ], + "type": "text", + "content": "In light of this, we propose SPAR3D, which breaks the 3D reconstruction process down into two stages: the point sampling stage and the meshing stage. The point sampling stage uses diffusion models to generate sparse point clouds, followed by the meshing stage, transforming point clouds into highly detailed meshes. Our main idea is to offload the uncertainty modeling to the point sampling stage, where the low resolution of the point clouds allows rapid iterative sampling. The subsequent meshing stage leverages the local image features to transform the point cloud into a detailed mesh of high output fidelity. Reducing the meshing uncertainty with point clouds further facilitates unsupervised learning of inverse rendering, which reduces the baked-in lighting in the textures. Our two-stage design enables SPAR3D to significantly outperform previous regressive methods, while preserving high computational efficiency and fidelity to input observation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "A key design choice of our method is the usage of point clouds to connect the two stages. To ensure fast reconstruction, our intermediate representation needs to be lightweight so it can be efficiently generated. However, it should provide enough guidance to the meshing stage. This inspires us to use point clouds, which are perhaps the most computationally efficient 3D representation because all information bits are used to represent the surface. Moreover, the lack of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "connectivity, typically considered as the drawback of point clouds, now turns into an advantage with our two-stage approach for editing purposes. When the back surface does not align with user expectations, local edits can be easily made on the low-resolution point clouds without worrying about topologies (see Fig. 1 bottom). Feeding edited point clouds into the meshing stage produces better meshes tailored towards user requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 168, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 289 + ], + "type": "text", + "content": "Our experiments demonstrate the superiority of SPAR3D over previous state-of-the-art methods, with solid quantitative and qualitative results on various data sources. SPAR3D also exhibits a strong generalization ability to inthe-wild images and AI-generated images. With a total inference time below 0.7 seconds, SPAR3D is not only efficient but also allows for easy user-driven edits, offering a practical solution to the task of monocular 3D reconstruction. We hope that this is a meaningful step towards scalable generation of high-quality 3D assets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 301, + 400, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 400, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 400, + 313 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 323, + 553, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 668 + ], + "type": "text", + "content": "Feedforward 3D reconstruction methods address the problem of 3D object reconstruction by learning a feedforward model in a regression-based manner. Earlier works [11, 20, 26, 39, 61, 66, 70] in this field typically predict only the geometry and train on small datasets [5, 52], which limits their generalization ability. Recently, larger 3D datasets [12, 44] have been collected, unlocking the potential to train feedforward 3D models at scale [25, 28, 65]. These models exhibit great generalization ability to unseen images, and excel at producing reconstructions that tightly align with the observed cues in the input image. In particular, LRM [25] and follow-up works [2, 55, 56, 63, 69, 73] show that properly designed large transformer models can be trained using only rendering losses to capture object geometry and texture in great detail. Despite the high fidelity and computational efficiency of these models, the oversimplified bijective assumption in these regressive approaches results in oversmoothed unseen surfaces. Multi-view diffusion models [36, 37, 47, 48] have been considered as a remedy for this, where additional viewpoints are synthesized as input to the feedforward model [55, 63, 69]. However, the inconsistency across viewpoints often leads to significant artifacts on the reconstructed surfaces, and the computational efficiency of these approaches is severely affected by the slow multi-view generation process. Our model also aims to overcome the learning ambiguity in regressive approaches, but our point sampling approach is inherently 3D-consistent and computationally efficient, and further allows easy user edits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": "Generative 3D modeling learns the image-conditioned distribution of 3D assets instead of a deterministic mapping. Early 3D generative works use GAN [4, 17, 29, 57]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "16861" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 547, + 224 + ], + "type": "image", + "image_path": "e8e51b1fbb67bb7db36b5bbe6cfc083785c1ae501f34e37887d3e051ccd0c87d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 236, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. SPAR3D Overview. Conditioned on the input image, SPAR3D first leverages a point diffusion model to generate a sparse point cloud. The triplane transformer then uses the sampled point cloud and image features to produce high-resolution triplane features. The triplane features are then queried to reconstruct the geometry, texture, and illumination of the object in the image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 276, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 276, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 296, + 408 + ], + "type": "text", + "content": "normalizing flow [31, 71] or VAE [18, 40, 68] as the generative framework. Inspired by the success of 2D diffusion models [15, 45], 3D diffusion models [6, 9, 10, 27, 30, 32, 33, 35-37, 41, 48-50, 72, 75] have also been extensively explored in recent works. Despite the advantage of probabilistic modeling that avoids over-smoothed results, diffusion-based 3D generation has two drawbacks: 1) not aligning well with input observations, and 2) having low inference speed at high resolution. Our work inherits the advantage of probabilistic modeling, while avoiding the drawbacks by using diffusion to generate only sparse point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": "Optimization-based single-view 3D leverages 2D generative priors to recover 3D from single-view images. These works [13, 21, 38, 54] rely on SDS-type loss [43, 60] and generate 3D assets by optimizing for each object image separately. These methods achieve promising results without large-scale annotation. However, the lack of a strong explicit 3D prior makes the optimization process inefficient and prone to local minima." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 111, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 111, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 111, + 537 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "content": "SPAR3D Overview. Given the input image " + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times h \\times w}" + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "text", + "content": ", our method produces a 3D mesh with PBR materials, including albedo, metallic, roughness and surface normals. The main goal of our work is to develop a model that enjoys the benefits of distribution learning through diffusion models, while not suffering from the low output fidelity and computational inefficiency. To this end, we design a two-stage model that consists of the point sampling stage and the meshing stage (see Fig. 2). At the point sampling stage, a point diffusion model learns the conditional distribution of point clouds given the input image. This stage is computationally efficient given the low resolution of the point clouds. The regression-based meshing stage transforms the sampled point cloud into a highly detailed mesh that aligns" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 276, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 361 + ], + "type": "text", + "content": "with the visible surface. The reduced uncertainty with point sampling further facilitates the learning of materials and illumination in an unsupervised manner during the meshing stage. This reduces baked-in lighting artifacts and results in better modeling of specular surfaces. Finally, by using sparse point clouds as the intermediate representation, SPAR3D enables human editing in the loop." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 366, + 437, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 437, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 437, + 379 + ], + "type": "text", + "content": "3.1. Point Sampling Stage" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": "Overview. The point sampling stage produces a sparse point cloud as the input to the meshing stage. The core of the point sampling stage is a point diffusion model, which generates point clouds " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\pmb{p}_0 \\in \\mathbb{R}^{n \\times 6}" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " conditioned on the input image " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": ". The six channels include three XYZ channels and three RGB channels. In our work, the resolution of the point cloud " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " is set to 512." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "content": "**Point Diffusion Framework.** Our diffusion framework is based on DDPM [24], which consists of two processes: 1) the forward process which adds noise to the original point cloud, and 2) the backward process where the denoiser learns to remove the noise. At timestep " + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "inline_equation", + "content": "t \\in [0,T]" + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "content": ", the diffusion process combines Gaussian noise " + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "content": " with a point cloud " + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 472, + 554, + 556 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 562, + 553, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 562, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 378, + 562, + 553, + 575 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {p} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\tag {1}", + "image_path": "334377cc91c67d65bd77d5b9ac08302fb012297ffcc9b327891bce44cb20f36c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "content": " denotes the noise schedule. We use the sigmoid noise schedule proposed in [7], combined with input scaling and the renormalization trick. The denoiser " + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}_{\\theta}(\\pmb{p}_t, t; \\pmb{c})" + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "content": " then learns to recover the noise from " + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t" + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "content": " and is supervised by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 634, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 634, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 351, + 634, + 553, + 649 + ], + "type": "interline_equation", + "content": "L _ {s i m p l e} (\\theta) = \\mathbb {E} _ {t, \\boldsymbol {p} _ {0}, \\boldsymbol {\\epsilon}} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {p} _ {t}, t; \\boldsymbol {c}) \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "69f39fa25720c7f2376ba49edf36a7f86962067d4ce029bbe7c6213895109b2c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " denotes the image condition tokens. During inference, we use the DDIM sampler [51] to generate point cloud samples. Samples generated directly often align poorly with the condition, hence we use the classifier-free guidance (CFG) [23] to improve sampling fidelity." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "16862" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Denoiser Design. We use a transformer denoiser similar to Point-E [41], where the noisy point cloud " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t \\in \\mathbb{R}^{n \\times 6}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " is linearly mapped to a set of point tokens " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". We use DINOv2 [42] to encode the input image " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " as conditioning tokens " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{c} \\in \\mathbb{R}^{c \\times d}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". The conditions and the point tokens are then concatenated together as input to the transformer, which predicts the added noise on each point." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 164, + 294, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 294, + 308 + ], + "type": "text", + "content": "Albedo Point clouds. In the meshing stage, we estimate the materials and lighting alongside the geometry. However, this decomposition is inherently ambiguous because there are countless combinations of lighting and albedo that can explain the same input image. It is challenging to learn this highly uncertain decomposition during the regressive meshing stage alone. We therefore reduce the uncertainty at the point sampling stage, by directly generating albedo point clouds with diffusion models. Sampling albedo point clouds as input to the meshing stage drastically reduces the ambiguity of inverse rendering and stabilizes the decomposition learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 318, + 146, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 146, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 146, + 331 + ], + "type": "text", + "content": "3.2. Meshing Stage" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 294, + 456 + ], + "type": "text", + "content": "Overview. The meshing stage produces a textured mesh from the input image and the point cloud. The backbone of our meshing model is a large triplane transformer, which predicts triplane features from the image and point cloud conditions. We estimate the geometry, texture and lighting of the current object from the triplane, and metallic/roughness from the image features. The geometry and materials are fed into our differentiable renderer during training, so that we can apply rendering loss to supervise our model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 466, + 294, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 294, + 597 + ], + "type": "text", + "content": "Triplane Transformer. Our triplane transformer consists of three submodules: the point cloud encoder, the image encoder, and the transformer backbone. We use a simple transformer encoder to encode the point cloud as a set of point tokens. Given the low resolution of the point clouds, each point can be directly mapped to a single token. Our image encoder is DINOv2 [42], which produces local image embeddings. Our triplane transformer follows a similar design to PointInfinity [27] and SF3D [2], which produces triplanes at high resolution of " + }, + { + "bbox": [ + 55, + 466, + 294, + 597 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 55, + 466, + 294, + 597 + ], + "type": "text", + "content": " by using a computationally-detached two-stream design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "content": "Surface Estimation. To estimate the geometry, the triplanes are queried with a shallow MLP to produce density values. Similar to [2, 63, 69], we convert the implicit density field to explicit surface using differentiable Marching Tetrahedron (DMTet) [46]. We additionally use two MLP heads to predict vertex offsets and surface normals together with density. These two attributes reduce the artifacts introduced by the Marching Tetrahedron and lead to locally smoother surfaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "Material and Illumination Estimation. We perform inverse rendering and jointly estimate materials (albedo, metallic and roughness) and illumination alongside the geometry. The task is highly ill-posed and Neural-PIL [1] showed that an illumination prior can reduce the ambiguity. We build our illumination estimator upon the learning-based illumination prior from " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{RENI}++" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " [19]. " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{RENI}++" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " is originally an unconditional generative model for HDR illumination generation. We learn an encoder to map triplane features into the latent space of " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{RENI}++" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": ". This allows us to estimate the environment illumination in the input image. The albedo is estimated from triplane similar to geometry, where a shallow MLP predicts the albedo value for each 3D location. For metallic and roughness, we follow SF3D [2] and learn to estimate them with a probabilistic approach via a Beta prior. We find that the CLIP encoder used in SF3D is unstable when the object size changes. We therefore replace their CLIP encoder with AlphaCLIP [53] to alleviate this issue using foreground object masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 303, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 578 + ], + "type": "text", + "content": "Differentiable Rendering. We implement a differentiable renderer that renders images based on the predicted environment map, PBR materials and geometry surface (see Fig. 3). We use a differentiable mesh rasterizer and add a differentiable Shader. Specifically, we leverage the standard simplified Disney PBR model [3] in our Shader. As we use " + }, + { + "bbox": [ + 313, + 303, + 553, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{RENI}++" + }, + { + "bbox": [ + 313, + 303, + 553, + 578 + ], + "type": "text", + "content": " to reconstruct environment maps, we need to explicitly integrate the incoming radiance. Here, we opt to use the Monte Carlo Integration. Given the low sample counts we can computationally afford during training, we rely on Multiple Importance Sampling (MIS) with the balanced heuristic [58] to reduce integration variance. Additionally, to better model the self-occlusion which has been typically ignored in prior works, we implement a visibility test for better shadow modeling. We take inspiration from real-time graphics and model the visibility test as a screen-space method using the depth map from our rasterizer. An overview of this test is shown in Fig. 4. Specifically, we ray-march a short distance (0.25) in 6 steps for all proposed sample directions from MIS, and project the position back to image space. If the current ray depth is farther away than the sampled value from the depth map, then the ray is marked as shadowed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Loss Function. Our main loss function is the rendering loss that compares renderings from novel views to the groundtruth (GT) images. Specifically, our rendering loss is a linear combination of 1) the L2 distance between the rendered and GT images, 2) the perceptual distance between the rendered and GT images measured by LPIPS [74], and 3) the L2 distance between the rendered opacity and the GT foreground mask. Apart from the rendering loss, we also follow SF3D and apply the mesh and shading regularization that regularizes the surface smoothness and the inverse rendering respectively." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "16863" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 331, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 331, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 331, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 331, + 190 + ], + "type": "image", + "image_path": "2d0c167b5544547dcd492ec0275ffd4ea7eb036926a4c0b28c2939f01796ff3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 197, + 333, + 275 + ], + "lines": [ + { + "bbox": [ + 54, + 197, + 333, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 197, + 333, + 275 + ], + "type": "text", + "content": "Figure 3. Our Differentiable Renderer. We estimate geometry, albedo, lighting, and normal maps from the triplane and metallic/roughness values from the image. We rasterize and interpolate these values as input to our shader (omitted here for simplicity). Our shader uses the Disney BRDF [3] and performs Monte Carlo integration. We further perform visibility testing to improve shadow modeling. Finally, we compare the rendered image with the GT image and minimize the rendering loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 340, + 70, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 340, + 70, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 553, + 190 + ], + "type": "image", + "image_path": "076aad9651cf70aa0dd5d76755f6ed3aca2ab82ac14276932a594066fcc5f2a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 197, + 553, + 252 + ], + "lines": [ + { + "bbox": [ + 339, + 197, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 339, + 197, + 553, + 252 + ], + "type": "text", + "content": "Figure 4. Shadow Modeling. We perform visibility testing in screen-space by marching along sampled rays. If any point along the ray has a ray depth which is farther away than the depth map, we consider the entire ray as shadowed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 167, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 167, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 167, + 294 + ], + "type": "text", + "content": "3.3. Interactive Editing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 300, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 297, + 479 + ], + "type": "text", + "content": "A unique advantage of our two-stage design is that it naturally supports interactive editing of unseen regions in our produced mesh. In most circumstances, the visible surface is determined by the input image and remains highly accurate, while the unseen surface is mainly based on the sampled point cloud, which might not align with user intention. In this case, editing the unseen surface of the mesh is feasible by altering the point cloud. Point clouds are perhaps one of the most flexible 3D representation for editing purposes because there are no topology constraints. Given the low resolution of our point clouds, editing the point cloud is fairly efficient and intuitive. Users can easily delete, duplicate, stretch or recolor points in the point cloud. Our efficient meshing model is able to produce the adjusted mesh in 0.3 seconds, which makes this process fairly interactive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 189, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 189, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 189, + 499 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 505, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 505, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 296, + 625 + ], + "type": "text", + "content": "Point Sampling Stage. Our point diffusion model has 16 transformer blocks in total. Each transformer block consists of two Layer Normalization layer, one Multi-Head Attention (MHA) layer and one MLP. We use a feature dimension of 1024 and 16 attention heads in each MHA layer. With many emissive objects in our dataset, albedo can be visually distinct from the input image and hard to learn. Therefore, instead of directly generating albedo point clouds in the point sampling stage, we learn to generate white-lit point clouds as a proxy target." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "content": "Meshing Stage. Our triplane transformer consists of 4 two-stream blocks [27]. Each two-stream block consists of three self-attention and two cross-attentions. The main computation is carried out using 3,072 latent tokens, each with a feature dimension of 1024. The MHA includes 16 attention heads. The point cloud encoder is a vanilla transformer with 12 layers and 512 feature dimension, and the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 498 + ], + "type": "text", + "content": "image encoder is DINOv2-large. We use a tetrahedra resolution of 160 for DMTet. In our differentiable Shader, we follow Hasselgren et al. [22] and use a detached biased sampling scheme. We sample based on the specular lobe (GGX [59]), the 2D piecewise-linear distribution of the environment map luminance and the hemispherical distribution. Specifically, we include 6 samples from the GGX lobe, 6 samples from the 2D piecewise-linear distribution of the luminance, and 4 samples from the hemispherical distribution. The main body of the shader is implemented in PyTorch, while the screen-space shadowing and the 2D piecewise-linear distribution computation for the environment map are implemented as custom CUDA kernels for efficiency. The training of our meshing stage includes multiple phases, where we increase the rendering resolution and decrease the batch size at later training phases. We use GT point clouds as input when training the meshing model. The curation of our training data follows TripoSR [56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 397, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 397, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 397, + 523 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 389, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 389, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 389, + 542 + ], + "type": "text", + "content": "4.1. Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 548, + 556, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 556, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 556, + 645 + ], + "type": "text", + "content": "Datasets. We used two datasets for evaluation, GSO [16] and OmniObject3D [67]. We follow TripoSR [56] and remove simple box or cylindrical objects to avoid bias on simple geometries. Each of the evaluation sets consists of around 250 objects. We render the objects with diverse azimuth angles at different elevations, with randomly sampled HDRI environment maps. We also vary the focal length of the camera to create more diverse test cases." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Metrics. To evaluate the geometry quality of the reconstructed meshes, we use follow prior works [56, 69] and use Chamfer Distance (CD) and F-score (FS) as our evaluation metrics. CD measures the alignment between two point clouds and is defined as the average of accuracy and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "16864" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 69, + 542, + 194 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 542, + 194 + ], + "type": "table", + "html": "
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2040.3590.6380.92215.30.8020.2053.1
LN3Diff [32]0.1740.4220.7030.94917.10.8190.1695.1
LGM [55]0.1960.3560.6350.93617.00.8180.18441.0
CRM [63]0.1610.4370.7350.96117.50.8300.1697.4
3DTopia-XL [8]0.2050.3830.6510.91316.00.8180.19340.6
Era3D [34]0.1810.4170.6900.94116.90.8180.173348.0
TripoSR [56]0.1450.5010.7840.96818.50.8370.1510.2
InstantMesh [69]0.1350.5450.8120.97118.10.8380.14636.1
SF3D [2]0.1370.5400.8060.97018.00.8390.1450.3
SPAR3D (ours)0.1200.5840.8500.98318.60.8360.1390.7
", + "image_path": "282a1d12d905838b8d8c8d40fab41fc050cde6ea94860d34e3286ecee841f62c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 219, + 542, + 345 + ], + "blocks": [ + { + "bbox": [ + 96, + 202, + 512, + 214 + ], + "lines": [ + { + "bbox": [ + 96, + 202, + 512, + 214 + ], + "spans": [ + { + "bbox": [ + 96, + 202, + 512, + 214 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons on GSO [16]. SPAR3D performs favorably to other state-of-the-art methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 219, + 542, + 345 + ], + "lines": [ + { + "bbox": [ + 69, + 219, + 542, + 345 + ], + "spans": [ + { + "bbox": [ + 69, + 219, + 542, + 345 + ], + "type": "table", + "html": "
MethodCD↓FS@0.1↑FS@0.2↑FS@0.5↑PSNR↑SSIM↑LPIPS↓Time (s)↓
Shap-E [30]0.2120.3490.6240.90914.80.80060.2053.1
LN3Diff [32]0.1600.4800.7440.95716.70.8190.1615.0
LGM [55]0.2000.3660.6380.92416.10.8100.18842.0
CRM [63]0.1550.4820.7650.96217.00.8280.1627.0
3DTopia-XL [8]0.1910.4070.6760.93115.80.8170.18641.2
Era3D [34]0.1850.4200.6850.93716.40.8160.167391.7
TripoSR [56]0.1440.5370.7850.96318.00.8350.1470.2
InstantMesh [69]0.1450.5460.7900.96217.20.8320.15034.7
SF3D [2]0.1380.5540.8000.96717.40.8360.1450.3
SPAR3D (ours)0.1220.5870.8450.97817.90.8320.1400.7
", + "image_path": "72d3d23c98b21e61c4858001e5bea732368be22d1ba76e77a981e165069f870a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 352, + 531, + 364 + ], + "lines": [ + { + "bbox": [ + 77, + 352, + 531, + 364 + ], + "spans": [ + { + "bbox": [ + 77, + 352, + 531, + 364 + ], + "type": "text", + "content": "Table 2. Quantitative Comparisons on OmniObject3D [67]. SPAR3D performs favorably to other state-of-the-art methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 367, + 116, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 116, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 116, + 379 + ], + "type": "text", + "content": "completeness:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 298, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 298, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 298, + 422 + ], + "type": "interline_equation", + "content": "d \\left(S _ {1}, S _ {2}\\right) = \\frac {1}{2 \\left| S _ {1} \\right|} \\sum_ {x \\in S _ {1}} \\min _ {y \\in S _ {2}} \\| x - y \\| _ {2} + \\frac {1}{2 \\left| S _ {2} \\right|} \\sum_ {y \\in S _ {2}} \\min _ {x \\in S _ {1}} \\| x - y \\| _ {2} \\tag {3}", + "image_path": "e6a89e9a019d433b67e79478370268dbe5e15d6b0d93b63baa6a196bef4070b7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 423, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 519 + ], + "type": "text", + "content": "FS evaluates point cloud alignment by calculating the F-score with a predefined threshold. Predicted points that lie within the distance threshold are considered as correct predictions. A higher FS means better alignment between the reconstructed shape and the groundtruth. To evaluate the texture quality, we compute standard image metrics, including PSNR, SSIM and LPIPS, between images rendered from the predicted mesh and the groundtruth images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 526, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 610 + ], + "type": "text", + "content": "Protocol. To calculate the metrics that are comparable across methods, the meshes need to lie in the same coordinate system. To this end, we perform brute-force search in rotations to align each predicted mesh with the groundtruth mesh. Both the prediction and the groundtruth are normalized before the brute-force alignment, and the alignment is further refined with ICP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "Baselines. We compare SPAR3D with other efficient methods for single-view 3D generation or reconstruction [2, 55, 56, 63, 69]. We use the official implementation for all baselines, and we evaluate the produced meshes under the same protocol. Specifically, we compare against TripoSR [56], LGM [55], CRM [63], InstantMesh [69], LN3Diff [32], Shap-E [30] and SF3D [2]. Among these baselines, TripoSR and SF3D are pure regression-based ap" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 366, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 403 + ], + "type": "text", + "content": "proaches; LGM, CRM and InstantMesh use multiview diffusion to generate pseudo multi-view images; LN3Diff and Shap-E are purely diffusion-based 3D generative models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 400, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 400, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 400, + 423 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 563 + ], + "type": "text", + "content": "Quantitative Comparison. We compare SPAR3D to other baselines on GSO and Omniobject3D quantitatively. As shown in Tab. 1 and Tab. 2, SPAR3D outperforms all other regressive or generative baselines significantly across most metrics on both datasets. For SSIM, we observe that SPAR3D is slightly worse than the strongest baseline for this metric. We find that this relates to the Monte Carlo noise from our Shader. SPAR3D is also among the fastest reconstruction models with an inference speed of 0.7 seconds per object, which is significantly faster than 3D or multiview diffusion-based approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Qualitative Results. We show qualitative results of different methods in Fig. 5. The reconstructed meshes from pure regression-based approaches such as SF3D or TripoSR align with the input image well, but the backside is often less accurate and over-smoothed. Multi-view diffusion-based methods such as LGM, CRM and InstantMesh show more details on the backside. However, the inconsistency in the synthesized views leads to clear artifacts and overall worse results. Pure generative approaches such as Shap-E and LN3Diff are able to produce sharp surfaces in their generation. However, many details are erroneous hallucinations that do not accurately follow the input images, and the vis" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "16865" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 77, + 541, + 371 + ], + "blocks": [ + { + "bbox": [ + 75, + 77, + 541, + 371 + ], + "lines": [ + { + "bbox": [ + 75, + 77, + 541, + 371 + ], + "spans": [ + { + "bbox": [ + 75, + 77, + 541, + 371 + ], + "type": "image", + "image_path": "1a3ecfe6b0ad1bbd0694f064943ef140b23ffe23266785a108909941e3fcb572.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 396 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison. We compare SPAR3D to other state-of-the-art methods visually. SPAR3D not only aligns better with the visible surfaces from images, but also generates higher-quality geometries and textures for the occluded surfaces." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 401, + 537, + 566 + ], + "blocks": [ + { + "bbox": [ + 75, + 401, + 537, + 566 + ], + "lines": [ + { + "bbox": [ + 75, + 401, + 537, + 566 + ], + "spans": [ + { + "bbox": [ + 75, + 401, + 537, + 566 + ], + "type": "image", + "image_path": "3823312c952490db487fbafc77a9068d71383a5ca938984e747f2f70081bd17e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 574, + 555, + 608 + ], + "lines": [ + { + "bbox": [ + 55, + 574, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 555, + 608 + ], + "type": "text", + "content": "Figure 6. Generalization Results. We show qualitative results of SPAR3D on in-the-wild images from 2D generative models (top 2 rows) and ImageNet (bottom 2 rows). The reconstructed meshes exhibit accurate geometric structures with great textures, demonstrating a strong generalization performance of SPAR3D." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 611, + 298, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 298, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 298, + 708 + ], + "type": "text", + "content": "ible surfaces are often reconstructed incorrectly. Compared to prior art, the meshes produced by SPAR3D not only faithfully resemble the input image, but also exhibit well-generated occluded parts with reasonable details. In Fig. 6, we further show qualitative results of SPAR3D on in-the-wild images. The images are generated using text-to-image generative models or from the ImageNet validation set [14]. The high quality of the reconstructed meshes demonstrates" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 611, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 555, + 636 + ], + "type": "text", + "content": "a strong generalization performance of SPAR3D. More results and comparisons are included in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 646, + 409, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 409, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 409, + 658 + ], + "type": "text", + "content": "4.3. Editing Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "The usage of explicit point clouds as an intermediate representation enables interactive editing of the generated meshes. Users can easily alter the unseen surface of the mesh by manipulating the point cloud. In Fig. 7, we show a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "16866" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 76, + 532, + 238 + ], + "blocks": [ + { + "bbox": [ + 72, + 76, + 532, + 238 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 532, + 238 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 532, + 238 + ], + "type": "image", + "image_path": "886233bcc6fa5c6c13ed349270b3480443d90cb0c90d667d6887c45c7eb55b42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 248, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 248, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 248, + 555, + 282 + ], + "type": "text", + "content": "Figure 7. Editing Results. We show qualitative examples of interactive editing with SPAR3D. On the left two examples, we add a handle to the mug and a tail to the elephant doll by duplicating existing points. On the right two examples, we move or delete points to fix imperfections and to improve local details on the mesh. All the edits are performed in Blender within a minute." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 295, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 295, + 322 + ], + "type": "text", + "content": "few editing examples with SPAR3D, either by adding major object parts to the reconstruction, or improving undesirable generated details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 119, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 119, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 119, + 341 + ], + "type": "text", + "content": "4.4. Ablation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 468 + ], + "type": "text", + "content": "We ablate the key idea of SPAR3D, the point sampling stage, which can be seen as an addition to standard regression approaches. We consider a variant of our model (SPAR3D w/o Point), where we remove the point sampling stage and make SPAR3D a full regressive model. We compare this variant with our full model on both GSO and Omniobject3D. As shown in Tab. 3, our full SPAR3D significantly outperforms the regressive variant, which validates the effectiveness of our design. More ablation experiments are included in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 471, + 296, + 525 + ], + "blocks": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 525 + ], + "type": "table", + "html": "
MethodCD↓FS@0.1↑PSNR↑LPIPS↓
SPAR3D w/o Point0.1360.50618.50.146
SPAR3D0.1200.58418.60.139
SPAR3D w/o Point0.1400.50917.80.146
SPAR3D0.1220.58717.90.140
", + "image_path": "900a8e7d35e753ef3588bdae30e76520159ecd7e217bd2b0345fa14259d4c40f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "text", + "content": "Table 3. Ablation Study on GSO (top 2 rows) and Omniobject3D (bottom 2 rows). Removing the point sampling stage leads to significant performance drop." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 575, + 118, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 118, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 118, + 588 + ], + "type": "text", + "content": "4.5. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "We further design experiments to understand how SPAR3D works. Our key assumption when designing SPAR3D is that the two-stage design effectively separates the uncertain part (back-surface modeling) and the deterministic part (visible surface modeling) of the monocular 3D reconstruction problem. Ideally, the meshing stage should mainly rely on the input image for reconstructing the visible surface, while relying on the point cloud to generate the back surface. To see whether this is true, we design an experiment where we artificially use point clouds that conflict with the input im" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 292, + 539, + 416 + ], + "blocks": [ + { + "bbox": [ + 329, + 292, + 539, + 416 + ], + "lines": [ + { + "bbox": [ + 329, + 292, + 539, + 416 + ], + "spans": [ + { + "bbox": [ + 329, + 292, + 539, + 416 + ], + "type": "image", + "image_path": "7484dced574c55277f64c202759d1a6d5ef3780359e9971a9cfac87fa1eb588f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "type": "text", + "content": "Figure 8. Generated Mesh with Conflicting Cues. Under conflicting cues from images and point clouds, our model reconstructs the visible surface based on the image, while generating the backside surface based on the point cloud." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 548 + ], + "type": "text", + "content": "age. In Fig. 8, we feed the input image of a squirrel and the point cloud of a horse to the meshing model. As shown in the figure, the reconstructed mesh indeed aligns with the squirrel image well on the visible surface, while the back surface mainly adheres to the point cloud. This result validates our assumption." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": "We present SPAR3D, a simple yet effective approach for single-view 3D reconstruction. The core of our model is a two-stage design based on point sampling. We first generate a sparse point cloud via point diffusion, and then reconstruct a highly detailed mesh from both the point cloud and the image. This design enables us to take the best of regression-based and generative modeling. Evaluated on standard benchmarks and in-the-wild images, SPAR3D significantly outperforms previous state-of-the-art methods with a fast inference speed. We will release our model upon publication, and we hope our effort is useful for future research towards scalable generation of high-quality 3D content." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "16867" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 114, + 294, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 294, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 294, + 157 + ], + "type": "text", + "content": "[1] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. NeurIPS, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 295, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 295, + 202 + ], + "type": "text", + "content": "[2] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. Sf3d: Stable fast 3d mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint, 2024. 2, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 224 + ], + "type": "text", + "content": "[3] Brent Burley. Physically-based shading at disney. ACM Transactions on Graphics (SIGGRAPH), 2012. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 226, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 226, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 226, + 294, + 291 + ], + "type": "text", + "content": "[4] Ruojin Cai, Guandao Yang, Hadar Averbuch-Elor, Zekun Hao, Serge Belongie, Noah Snively, and Bharath Hariharan. Learning gradient fields for shape generation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 364–381. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 402 + ], + "type": "text", + "content": "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2416-2425, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 424 + ], + "type": "text", + "content": "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 480 + ], + "type": "text", + "content": "[8] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 535 + ], + "type": "text", + "content": "[9] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 581 + ], + "type": "text", + "content": "[10] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2262-2272, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 582, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 634 + ], + "type": "text", + "content": "[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628-644. Springer, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "text", + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[13] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20637-20647, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 196 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 262 + ], + "type": "text", + "content": "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "text", + "content": "[17] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 320, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 554, + 363 + ], + "type": "text", + "content": "[18] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40 (6):1-15, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[19] James AD Gardner, Bernhard Egger, and William AP Smith. Reni++ a rotation-equivariant, scale-invariant, natural illumination prior. arXiv preprint arXiv:2311.09361, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 453 + ], + "type": "text", + "content": "[20] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan Russell, and Mathieu Aubry. AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 520 + ], + "type": "text", + "content": "[21] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nerf-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[22] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 589 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "type": "text", + "content": "[25] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[26] Zixuan Huang, Varun Jampani, Anh Thai, Yuanzhen Li, Stefan Stojanov, and James M Rehg. Shapeclipper: Scalable 3d shape learning from single-view images via geometric and clip-based consistency. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "16868" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[27] Zixuan Huang, Justin Johnson, Shoubhik Debnath, James M Rehg, and Chao-Yuan Wu. Pointinfinity: Resolution-invariant point diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10050-10060, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[28] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Zeroshape: Regression-based zero-shot shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061-10071, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 260 + ], + "type": "text", + "content": "[29] Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang. Progressive point cloud deconvolution generation network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 397-413. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[30] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[31] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. In European Conference on Computer Vision, pages 694-710. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "type": "text", + "content": "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. arXiv preprint arXiv:2403.12019, 2024. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 426 + ], + "type": "text", + "content": "[33] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12642-12651, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "type": "text", + "content": "[34] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wei Xue, Wenhan Luo, et al. Era3d: high-resolution multiview diffusion using efficient row-wise attention. Advances in Neural Information Processing Systems, 37:55975-56000, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[35] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. arXiv preprint arXiv:2311.07885, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[36] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9298–9309, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[37] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8446-8455, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[40] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306-315, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[41] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 375 + ], + "type": "text", + "content": "[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10901-10911, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[46] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "text", + "content": "[47] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "text", + "content": "[48] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "type": "text", + "content": "[49] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20887-20897, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[50] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "16869" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[52] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[53] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "text", + "content": "[54] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[55] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[56] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 2, 5, 6, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[57] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International conference on learning representations, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 404 + ], + "type": "text", + "content": "[58] Eric Veach. *Robust Monte Carlo Methods for Light Transport Simulation*. PhD thesis, Stanford University, 1997. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 448 + ], + "type": "text", + "content": "[59] Bruce Walter, Stephen R. Marschner, Hongsong Li, and Kenneth E. Torrance. Microfacet models for refraction through rough surfaces. Eurographics Symposium on Rendering, 2007. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 504 + ], + "type": "text", + "content": "[60] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[61] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[62] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[63] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 2, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[64] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 12" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 620 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[65] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. arXiv preprint arXiv:2301.08247, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[66] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "text", + "content": "[67] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. arXiv preprint arXiv:2301.07525, 2023. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "text", + "content": "[68] Zhijie Wu, Xiang Wang, Di Lin, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Sagnet: Structure-aware generative network for 3d-shape modeling. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "type": "text", + "content": "[69] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 554, + 373 + ], + "type": "text", + "content": "[70] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 374, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 554, + 430 + ], + "type": "text", + "content": "[71] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4541-4550, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 555, + 474 + ], + "type": "text", + "content": "[72] Lior Yariv, Omri Puny, Oran Gafni, and Yaron Lipman. Mosaic-sdf for 3d generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4630-4639, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 554, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 554, + 519 + ], + "type": "text", + "content": "[73] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. arXiv preprint arXiv:2404.19702, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 554, + 553 + ], + "type": "text", + "content": "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 554, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 554, + 620 + ], + "type": "text", + "content": "[75] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang YU, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "16870" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_content_list.json b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..074c0434a3bf06c126a19575a8f4315a3060ab0f --- /dev/null +++ b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "SPARC: Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models", + "text_level": 1, + "bbox": [ + 124, + 128, + 872, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kevin Miller \nBoston University \nnivek@bu.edu", + "bbox": [ + 222, + 203, + 370, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aditya Gangrade \nBoston University \ngangrade@bu.edu", + "bbox": [ + 424, + 203, + 571, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samarth Mishra \nBoston University \nsamarthm@bu.edu", + "bbox": [ + 627, + 203, + 772, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kate Saenko \nBoston University and Meta AI (FAIR)", + "bbox": [ + 228, + 263, + 542, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "saenko@bu.edu", + "bbox": [ + 326, + 303, + 444, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Venkatesh Saligrama \nBoston University \nsrv@bu.edu", + "bbox": [ + 598, + 263, + 767, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 351, + 326, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-shot multi-label recognition (MLR) with Vision-Language Models (VLMs) faces significant challenges without training data, model tuning, or architectural modifications. Existing approaches require prompt tuning or architectural adaptations, limiting zero-shot applicability. Our work proposes a novel solution treating VLMs as black boxes, leveraging scores without training data or ground truth. We make two contributions. First, we find that VLM scores suffer from image- and prompt-specific biases, and that simple standardization is surprisingly effective at removing these and boosting MLR performance. And second, we introduce compound prompts grounded in realistic object combinations. Our analysis reveals \"AND\"/\"OR\" signal ambiguities that cause maximum compound scores to be surprisingly suboptimal compared to second-highest scores. We introduce an adaptive fusion method to address this issue. Our method enhances other zero-shot approaches, consistently improving their results. Experiments show superior mean Average Precision (mAP) compared to methods requiring training data, achieved through refined object ranking for robust zero-shot MLR. Code can be found at https://github.com/kjmillerCURIS/SPARC.", + "bbox": [ + 89, + 385, + 483, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs), such as CLIP [16], have emerged as general-purpose systems for understanding visual data through language-based queries. These models enable a broad range of applications, from object detection to image captioning, by linking visual inputs to language prompts. In standard settings where images contain single, recognizable objects, VLMs perform remarkably well. However, for the more complex task of zero-shot multi-label recognition (MLR) (Fig. 1 (top)), where mod", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0db919ced7a520be949a3f532cc33f99a7a787d84a0b66bfa4af880c91182727.jpg", + "image_caption": [ + "Figure 1. (top) Vision-Language Models (VLMs) like CLIP can be used for zero-shot classification with image-text similarity scores. While this works fairly well for single-class labels, they can struggle in the multi-label scenario. (bottom) In this paper, we introduce SPARC, our solution that functions on top of an existing VLM, treating it simply as a black-box score generator. Using class names, SPARC first creates compound prompts for additional queries to the VLM. It then normalizes, ranks and appropriately fuses them to generate final scores for the original classes." + ], + "image_footnote": [], + "bbox": [ + 526, + 349, + 895, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els must identify multiple objects within an image without prior training on specific data, VLMs face significant limitations. Zero-shot MLR is crucial for applications in fields like robotics and medical imaging, where objects rarely appear in configurations that align neatly with training distributions. In these scenarios, achieving robust multi-label recognition without fine-tuning is challenging, given the task's reliance on mean Average Precision (mAP) scores, which depend on ranking images for object presence.", + "bbox": [ + 511, + 715, + 906, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VLM: Prompt Dependent AND/OR Noisy Channel. Despite the promise of zero-shot capabilities, current VLM approaches often struggle with MLR due to inherent scoring", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4313", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/207ec5e7ad2b8641f53f9651fe0cae7f444435211af6f44f50e2016c6e13f702.jpg", + "image_caption": [ + "Figure 2. A motivating example for the Normalization module, with CLIP scores on a slice of the COCO dataset, before and after normalization. The \"necktie\" class is present for examples in the top half of the plot and absent in the bottom half. Image- and prompt-level biases show up as horizontal and vertical striations; normalization removes these and creates better separation." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9a3605d6e786a3389140855f80cb432d3ee53c74107b82f89e7c3cdc0944a30.jpg", + "image_caption": [ + "Figure 3. A motivating example for the Rank Fusion module, with an image where class \"cat\" is absent (left) and one where it is present (right). The highest compound prompt score is an unhelpful signal because it gives a high score to both negatives and positives, while the second-highest is more discriminative. Our method adaptively fuses the most informative order statistics, resulting in a strong signal." + ], + "image_footnote": [], + "bbox": [ + 96, + 407, + 478, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "behaviors and biases. The performance of these models is hampered by a mix of conjunction (AND) and disjunction (OR) behaviors in their scoring, leading to inflated scores in compound prompts that contain multiple objects. For example, a prompt like \"cat and sofa\" might yield a high score even if only one of these objects is present in the image. This tendency reflects biases learned during training, where common object pairs receive higher scores even when only one object is present, disrupting the accuracy of mAP-based evaluations. Furthermore, existing methods for adapting VLMs to zero-shot MLR frequently rely on prompt tuning", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or architectural adjustments—approaches that are often dependent on training data and computationally intensive finetuning, which limit their generalizability to novel tasks.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Approach. In contrast to these methods, we introduce SPARC (Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in VLMs), a novel approach to zero-shot MLR that bypasses the need for training data, prompt tuning, or model-specific modifications. Our method treats the VLM as a black box, relying solely on its output scores to infer object presence (see Fig. 1). This black-box approach enables us to avoid assumptions about the model's internal workings, allowing for a purely zero-shot framework that is both model-agnostic and dataset-independent. SPARC introduces two main innovations that address the unique challenges of zero-shot MLR.", + "bbox": [ + 511, + 141, + 906, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A. Score Normalization. Our analysis finds that VLM scores suffer from image-level and prompt-level biases. Image-level bias refers to noise terms that are constant across prompts but vary between images, while prompt-level bias refers to the converse. Image-level bias in particular hampers mAP as it changes the ordering of scores within each class. Standardization is surprisingly effective at addressing this problem in MLR. Standardization alone improves mAP by $6 - 10\\%$ on COCO, VOC, and NUSWIDE.", + "bbox": [ + 511, + 325, + 908, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "B. Compound Prompts and Adaptive Fusion. Recognizing that VLMs can provide richer information when prompted with combinations of objects, we develop a method for constructing compound prompts. These prompts reflect likely contextual associations between objects, such as \"cat and sofa\" or \"car and bus.\" By gathering scores from these compound prompts, we can capture a spectrum of potential object contexts within the image, enhancing detection without relying on training-based adaptations. This composition strategy allows us to agnostically extract information from the VLM without depending on any specific dataset or VLM architecture.", + "bbox": [ + 511, + 463, + 908, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surprisingly, we observe that the maximum score among compound prompts is often a poor proxy for true object presence. Although one might expect the highest score to serve as a reliable signal, we find that it suffers because VLMs tend to respond to compound prompts with OR-like behavior, raising scores even when only one object in the prompt is present. Instead, we observe that the second-highest score consistently outperforms the maximum score by minimizing the effects of false positives caused by OR-like behavior. Building on this insight, we develop a PCA-based fusion method that combines information from multiple order statistics from the compound prompts, as well as the singleton prompts, ultimately optimizing mAP by enhancing score accuracy.", + "bbox": [ + 511, + 645, + 908, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Complementarity. SPARC is complementary to other zero-shot and training-free MLR methods. When applied on top of these approaches, SPARC consistently enhances", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4314", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mAP scores by refining object ranking and reducing bias in VLM outputs. This capability makes SPARC an adaptable solution that can improve upon existing methods while maintaining a fully zero-shot, model-agnostic framework.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical Results. SPARC achieves significant improvements in mAP, outperforming methods that incorporate architectural modifications. This outcome shows the potential of a fully zero-shot approach that relies only on systematic prompt design and score interpretation, rather than prompttraining or fine-tuning. By revealing that the second-highest score can be a superior proxy to the maximum, our findings provide new insights into VLM scoring behavior, suggesting that careful treatment of prompt compositions and score patterns can unlock robust MLR capabilities.", + "bbox": [ + 89, + 152, + 482, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1.1. Related Work", + "text_level": 1, + "bbox": [ + 89, + 315, + 235, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised Methods. Prior methods improve MLR performance through various approaches. DualCoOp [7, 18] trains prompts for text-guided spatial attention. Subsequent works use class co-occurrences to refine CLIP logits [17, 20]. Hierarchical structuring [20] encourages related classes to learn similar prompts. SSPA [19] combines embeddings from learned and LLM-generated prompts, and TRM-ML [12] uses pseudolabels to match texts to image regions. While effective, these methods require annotated training data. In contrast, ours is unsupervised.", + "bbox": [ + 89, + 339, + 483, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised Training-Based Methods. TaI-DPT [4] uses caption-only data to adapt CLIP to MLR by training a prompt-generation network using caption embeddings. Extensions include LLM-generated captions [22], pseudovisual prompts [21], and lightweight classifiers on caption embeddings [11, 23]. Separately, CDUL [1] leverages CLIP scores with spatial aggregation to create training pseudolabels. While these approaches demonstrate shared vision-language embedding power, they require significant training and CLIP internal access. In contrast, our black-box method uses crude cooccurrence-based compound prompts (essentially deleting implausible contextual associations) to probe multilabel information, fusing scores for effective MLR.", + "bbox": [ + 89, + 491, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised Training-free methods. The recent literature has proposed methods to adapt CLIP to the MLR taks in a training-free manner via architectural changes. CLIP-Surgery [8] alters the architecture to increase the coherence of attention masks within the image backbone, and normalizes features on a patch-by-patch basis. TagCLIP [10] uses internal attention masks to refine scores at a patch level based on regional coherence, and aggregate queries over crops of the image. Again, both methods require access to VLM internals, while ours works in a black-box manner.", + "bbox": [ + 89, + 686, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Enhancement for Single-Label Recognition. Several methods use LLM-generated descriptive prompts for single-label recognition, combining them via max or mean fusion [3, 13-15]. While related, our compound", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prompts specifically target MLR by modeling class relationships rather than enriching individual class descriptions.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Complementarity of SPARC to Unsupervised Methods. We explicitly note that our methodology is complementary to much of the work on unsupervised VLM-based MLR. For instance, since we assume only black-box access to VLM scores, our compound prompts could be refined via prompt tuning as in TaI-DPT [4] and/or applied patch-by-patch and integrated à la TagCLIP [10]. Scores from any of these methods can be combined via our normalization and rank-fusion approach. §4.3 investigates the resulting gains.", + "bbox": [ + 511, + 121, + 906, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 513, + 270, + 604, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We detail our method, SPARC (see Alg. 1 and Fig. 1).", + "bbox": [ + 513, + 296, + 870, + 311 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 SPARC Pipeline \nInput: Images $\\mathcal{I}$ , Class Names $\\mathcal{C}$ \nOutput: Final scores $\\zeta_i^t$ for each image $t$ and class $i$ $P \\leftarrow$ GenerateCompoundPrompts $(\\mathcal{C})$ (Details in Supp.) \nfor $t \\in \\mathcal{I}$ do \n $s_i^t \\leftarrow$ GetVLMScores $(t, i)$ (Singleton scores) \n $\\forall p \\in P, s_p^t \\leftarrow$ GetVLMScores $(t, p)$ (Compound scs.) \n// (Normalize scores (Eqs. 1, 2)) \n $\\tilde{s}_i^t \\leftarrow$ ImageNorm $(s_i^t), \\tilde{s}_p^t \\leftarrow$ ImageNorm $(s_p^t)$ $\\bar{s}_i^t \\leftarrow$ PromptNorm $(\\tilde{s}_i^t), \\bar{s}_p^t \\leftarrow$ PromptNorm $(\\tilde{s}_p^t)$ \nfor $i \\in \\mathcal{C}$ do \n $r_{i,k}^t \\leftarrow$ GetOrderStatistics( $\\{\\bar{s}_p^t : i \\in p\\}$ ) \n $w^{i*} \\leftarrow$ MaxVarDirection([\\bar{s}_i^t, r_{i,k}^t]) (Eq. 3) \n $\\tilde{\\zeta}_i^t \\leftarrow w_0^{i*} \\bar{s}_i^t + \\sum_k w_k^{i*} r_{i,k}^t$ (Eq. 4) \n $\\zeta_i^t \\leftarrow \\bar{s}_i^t + \\tilde{\\zeta}_i^t$ (Eq. 5) \nreturn $\\{\\zeta_i^t\\}$", + "bbox": [ + 514, + 315, + 903, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Setup. Suppose we have $M$ images and $N$ target classes, with classnames $c_{1},\\ldots ,c_{N}$ . We assume that we have a set of \"singleton\" prompts describing each of these classes in isolation (e.g. \"picture of a cat\"). Given an image $t$ , let $s_1^t,\\dots,s_N^t$ denote the similarity scores produced by a VLM when comparing the singleton prompts to that image.", + "bbox": [ + 511, + 583, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A naive approach is to directly use these singleton scores to perform MLR. However, since the presence of a particular class alters the conditional probability of the remaining classes in images, it should be possible to further refine these scores by accounting for the multi-label structure of the images. In order to do this, we use the class-names to generate a further set of \"compound\" prompts $P$ which mention multiple classes. For instance, the classes \"cat\" and \"sofa\" may be compounded to \"cat and sofa\". Let $\\{s_p^t : p \\in P\\}$ denote the scores given by the VLM when comparing these compound prompts to image $t$ .", + "bbox": [ + 511, + 674, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the singleton as well as the compound prompts above, SPARC produces a vector of refined class-wise scores $\\zeta_1^t,\\dots ,\\zeta_N^t$ , for each image $t$ . Given this structure, SPARC has three main components: (i) compound prompt", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4315", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generation to enable the use of CLIP scoring for MLR (ii) image and prompt level normalization to allow scores to be directly compared, and (iii) Rank Fusion to combine singleton and compound prompt scores into a single score per class. We now discuss each of these aspects in detail.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.1. Compound Prompt Generation", + "text_level": 1, + "bbox": [ + 89, + 174, + 366, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our compound prompt generation method relies on contextual associations from commonly observed visual patterns to inform prompt structure. We provide pseudocode for the process in the Supplementary. The process takes as input the classnames $c_{1}, \\ldots, c_{N}$ , ground-truth cooccurrence statistics $\\mathbb{P}$ , thresholds $\\tau_{2}$ and $\\tau_{3}$ (fixed to 0.05 and 0.025 for all datasets), and off-the-shelf LLM $\\Phi$ . We use probability thresholding to select pairs and triplets of classes that could plausibly coocur in realistic visual scenes. These pairs and triplets are used to make formulaic compound prompts of the form \"A and B\" and \"A, B, and C\". We feed these formulaic prompts to LLM $\\Phi$ and ask it to generate natural sentences from them. Our total set of compound prompts comprises the formulaic pair, formulaic triplet, and natural sentence prompts, which add up to on average $\\leq 20$ compound prompts per class in all of our datasets. Note that probabilities $\\mathbb{P}$ are not used in any part of our method other than compound prompt generation, which is why we say that it uses \"rough\" cooccurrence info. In fact, our ablations reveal that we still get a performance gain when using all possible class pairs instead of filtering by probability.", + "bbox": [ + 89, + 198, + 483, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2. Normalization", + "text_level": 1, + "bbox": [ + 89, + 523, + 238, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we have obtained compound prompts $P$ , we can query the VLM to obtain singleton scores $s_1^t, \\ldots, s_N^t$ and compound scores $\\{s_p^t : p \\in P\\}$ for each image $t$ . However, these scores contain both image- and prompt-specific biases. The former in particular is a problem, as it can change the order of scores between negative and positive images.", + "bbox": [ + 89, + 546, + 482, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We address image-level bias for singleton and compound prompts, respectively, as follows:", + "bbox": [ + 89, + 637, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {s} _ {i} ^ {t} := \\frac {s _ {i} ^ {t} - \\hat {\\mu} (s _ {i} ^ {t})}{\\hat {\\sigma} (s _ {i} ^ {t})} \\quad a n d \\quad \\tilde {s} _ {p} ^ {t} := \\frac {s _ {p} ^ {t} - \\hat {\\mu} (\\check {s} _ {i} ^ {t})}{\\hat {\\sigma} (\\check {s} _ {i} ^ {t})}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 672, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mu}(s_{\\cdot}^{t}), \\hat{\\mu}(\\check{s}_{\\cdot}^{t}), \\hat{\\sigma}(s_{\\cdot}^{t}), \\hat{\\sigma}(\\check{s}_{\\cdot}^{t})$ are sample means and standard deviations across the prompt dimension for a single image. $\\check{s}_{1}^{t}, \\ldots, \\check{s}_{N}^{t}$ are the scores of \"auxiliary\" prompts that mention classnames in isolation and are used only to obtain statistics for standardizing compound prompt scores.", + "bbox": [ + 89, + 715, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also standardize across images to quash prompt bias:", + "bbox": [ + 109, + 792, + 482, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{s}} _ {i} ^ {t} := \\frac {\\tilde {s} _ {i} ^ {t} - \\hat {\\mu} (\\tilde {s} _ {i} ^ {*})}{\\hat {\\sigma} (\\tilde {s} _ {i} ^ {*})} a n d \\overline {{s}} _ {p} ^ {t} := \\frac {\\tilde {s} _ {p} ^ {t} - \\hat {\\mu} (\\tilde {s} _ {p} ^ {*})}{\\hat {\\sigma} (\\tilde {s} _ {p} ^ {*})}, (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 825, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mu} (\\tilde{s}_i)$ , $\\hat{\\mu} (\\tilde{s}_p)$ , $\\hat{\\sigma} (\\tilde{s}_i)$ , and $\\hat{\\sigma} (\\tilde{s}_p)$ now denote sample means and SDs across the images. Removing prompt-level", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bias makes scores from different prompts more compatible with each other, which is important for fusing them.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Rank Fusion", + "text_level": 1, + "bbox": [ + 513, + 128, + 648, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal now is to find a way to use compound scores to strengthen the singleton scores. Let us define a bit more notation here. Let $r_{i,k}^{t}$ denote the score of the $k$ -th highest scoring compound prompt that mentions class $i$ .", + "bbox": [ + 511, + 152, + 903, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A natural choice might be to choose $r_{i,1}^{t}$ , i.e. the highest scoring prompt. After all, we would expect this to be the prompt that most closely describes the image, and if the image did not contain $c_{i}$ then we would expect none of the prompts containing $c_{i}$ to score high. However we find that $r_{i,1}^{t}$ does not offer a good detection of classes. The mechanism behind this, which is explored in detail in §3, is that when using a compound prompt of the form 'A and B' to detect the class $A$ , the score sees a large increase if the object $B$ truly occurs in the image. This 'OR'-like behavior (§3) means that maximum score $r_{i,1}^{t}$ typically captures the effect of other classes than $A$ being present in the image, and thus leads to very poor separation between ground truth positive and negative images for any class. Counterintuitively, then, we find that 'weakened maxes', i.e., lower order statistics like $r_{i,2}^{t}$ and $r_{i,3}^{t}$ yield much better separation.", + "bbox": [ + 511, + 214, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above fact suggests that we ought to utilize the entirety of the order statistics $\\{r_{i,k}^t\\}$ in order to generate an effective fusion rule. We do fusion by taking a weighted sum where the weighting vector is the direction of highest variance, i.e., we compute the fused compound score $\\tilde{\\zeta}_i^t$ as", + "bbox": [ + 511, + 455, + 905, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {i *} := \\underset {w ^ {i}} {\\arg \\max } \\operatorname {V a r} _ {t} \\left(w _ {0} ^ {i} \\bar {s} _ {i} ^ {t} + \\sum_ {k} w _ {k} ^ {i} r _ {i, k} ^ {t}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 551, + 906, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\zeta} _ {i} ^ {t} := w _ {0} ^ {i *} \\bar {s} _ {i} ^ {t} + \\sum_ {k} w _ {k} ^ {i *} r _ {i, k} ^ {t}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 587, + 906, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the variance above is computed across examples, i.e. an iteration over example index $t$ . Computing arg max above is just a matter of running PCA and taking the first principal component. The motivation for this approach is similar to that for linear discriminant analysis. Imagine each example is a point in $\\mathbb{R}^{K_i}$ , where $K_{i}$ is the number of compound prompts for class $i$ , and these points are distributed according to a mixture of two components, one for examples where $i$ is present, the other for examples where $i$ is absent. The shift between the means of these components should contribute to the variance in the direction of that shift. If we assume that these components are roughly homoskedastic, then, as long as the 'signal' induced by the mean shift is large compared to any heteroskedasticity, the highest variance direction $w^{i*}$ is precisely the direction of the mean shift, and thus $\\tilde{\\zeta}_i^t$ effectively computes the component of the score vector that lies along this signal direction. While principled, this max-variance strategy is of course not", + "bbox": [ + 511, + 627, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4316", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "necessary to execute, and we compare it to alternative fusion strategies in §4.4.1. We add (i.e. \"merge\") this fused score into the original singleton score to get our final score:", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta_ {i} ^ {t} := s _ {i} ^ {t} + \\tilde {\\zeta} _ {i} ^ {t} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 150, + 482, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the next section we probe CLIP scores through a statistical lens to build intuition on our approach.", + "bbox": [ + 89, + 176, + 482, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Noise Model of a VLM", + "text_level": 1, + "bbox": [ + 89, + 220, + 307, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our methodological choices are strongly driven by the behavior of CLIP scores under compound prompts, and this leads to both the normalization approach taken, as well as our counterintuitive choice of fusing compound scores by weakened max instead of outright max. In order to justify these choices, we probe the structure of CLIP scores under compound prompts by constructing a statistical model.", + "bbox": [ + 89, + 246, + 483, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an image $t$ and a prompt with classes $i$ and $j$ , we model the score of compound prompt 'i and j' as $s_{i,j}^{t} \\coloneqq \\theta_{1}^{t} \\cdot f(y_{i}^{t}, y_{j}^{t}) + \\theta_{0}^{t} + \\varepsilon$ where $y_{i}^{t}, y_{j}^{t} \\in \\{0,1\\}$ indicate ground-truth presence of the two classes, $\\theta_{1}^{t}$ and $\\theta_{0}^{t}$ represent image-level bias (e.g. due to visual domain-shifts), and $\\varepsilon$ is IID Gaussian noise. Note that our Normalization module is well-suited to mitigate variations in $\\theta_{0}^{t}$ and $\\theta_{1}^{t}$ , at least when $\\varepsilon$ is not strongly dominating.", + "bbox": [ + 89, + 353, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, the core object of interest is the map $f: \\{0,1\\}^2 \\to \\mathbb{R}$ , which is responsible for modeling class pair interactions and prompt-level biases (see Table 1). In an ideal situation for MLR, $f$ would behave roughly as an 'AND' function, so that compound prompts would let us directly hone in on the presence of pairs of classes. However, prior work has observed that instead, for CLIP-type models, $f$ tends to predominantly behave as an 'OR' function, in that $f$ is high if either $i$ or $j$ is present. Notice that if $f$ were to exactly behave as an OR, compound prompts would not be effective in detecting classes that were missed by the singleton scores.", + "bbox": [ + 89, + 474, + 483, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In experimentation, we found that the behavior of $s_{i,j}^{t}$ is intermediary to the above extremes: CLIP scores for compound prompts behave qualitatively as an OR function with a small AND 'bonus' when both classes are present. This bonus, as well as the particular scores, are modulated by the classes mentioned (a consequence of prompt-level bias for each class) which we model as linear effects, leading to an overall model of the signal term of the CLIP score as", + "bbox": [ + 89, + 641, + 483, + 761 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f \\left(y _ {i} ^ {t}, y _ {j} ^ {t}\\right) := \\max \\left(a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}\\right) \\\\ + \\delta_ {i j} \\min (a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}), \\qquad (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 772, + 482, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $a$ terms are class-wise linear effect coefficients, and $\\delta_{ij}$ captures the strength of this AND bonus.", + "bbox": [ + 89, + 824, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that this AND bonus is a critical feature of CLIP scores in their use for MLR problems, since this bump allows us to both infer the presence of classes that singleton", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a4730ee9e35c0c9ed4293501c89df210b54949ef23c66a5474e2552f96a1e960.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Noise Modelf(yi,t,yj,t)FVU
constant0.10.802
only ANDmin(a0i+yitai,aj+ytaj)0.536
only ORmax(a0i+yitai,aj+ytaj)0.288
additivea0i+yitai+aj+ytaj0.267
OR + static AND-bonusmax(a0i+yitai,aj+ytaj) + δ min(a0i+yitai,aj+ytaj)0.263
OR + variable AND-bonusmax(a0i+yitai,aj+ytaj) + δi,j min(a0i+yitai,aj+ytaj)0.248
look-up tableLUT(i,j)(yi,t,yj,t)0.235
", + "bbox": [ + 516, + 88, + 905, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of fidelity of noise models for scoring pairwise compound prompts. Notice that the OR+AND-bonus model captures nearly all of the fidelity of the look-up table, and offers significant improvements over the simplified models above it.", + "bbox": [ + 511, + 266, + 906, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompts have missed, and indirectly to filter false positives (since the true positives are further elevated by the bonus).", + "bbox": [ + 511, + 335, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows the results of a systematic evaluation of the AND-OR structure of compound CLIP scores, using 81K COCO test images and ViT-L/14@336px backbone. We evaluate various models of $f$ , ranging from simple constant, 'only OR,' and 'only AND' models to additive models and OR+AND-bonus models with both a 'static' $\\delta$ shared by all class pairs, and a 'variable' $\\delta_{i,j}$ that is class-pairspecific. The final entry is a 'look-up table' (LUT) model where all four values of $f$ can vary arbitrarily, which serves as the most expressive baseline in this model class. These are evaluated w.r.t. the fraction of variance unexplained (i.e., $1 - R^2$ value). Observe that while the only OR model has remarkable fidelity, the additive and AND-bonus models significantly improve upon this, and allowing $\\delta$ to vary with $i,j$ yield a further significant improvement to within a $6\\%$ extra loss from the LUT model. This shows that our AND-bonus models capture a significant amount of the qualitative behavior of CLIP scores for compound prompts.", + "bbox": [ + 511, + 366, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A closer look at these fitted models reveals the $\\delta$ values are small and relatively stable: for static $\\delta$ , the recovered value is 0.56, and the IQR for varying $\\delta_{ij}$ is (0.44, 0.57).", + "bbox": [ + 511, + 638, + 903, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We find similar behavior when examining other backbones and datasets, which we detail in the Supplementary. The Supplementary also contains a theoretical view of the benefits of using weakened-max, based on the noise model.", + "bbox": [ + 511, + 684, + 903, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Methodological Implications", + "text_level": 1, + "bbox": [ + 511, + 755, + 769, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The qualitative structure of the OR+AND-bonus behavior justifies our methodological choices in §2. Image-level normalization reduces the effect of variation in $\\theta_0^t$ (via mean subtraction) and $\\theta_1^t$ (via rescaling), while prompt-level normalization mitigates variations in $a_0^i$ and $a_1^i$ across classes.", + "bbox": [ + 511, + 779, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can also qualitatively observe that, because the OR component of the scores dominates the AND component, the maximum score $r_{i,1}^{t}$ can create false positives when $i$ is", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "absent but related class $j$ is present, and so we should look to subsequent order statistics for a better signal.", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 137, + 222, + 155 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We benchmark our method on 3 different MLR datasets. They are COCO2014 [9], which has 40,504 test images with 80 classes; VOC2007 [6], which has 4.952 test images with 20 classes; and NUSWIDE [2], which has 107,859 test images with 81 classes. We use only test, and not train, images from these datasets. We do use their conditional ground-truth label distributions, computed from training set, to choose compound prompts, but do not use ground-truth probabilities in any other part of our method.", + "bbox": [ + 89, + 162, + 482, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP Backbones In order to showcase the generality of our method, we apply it to 9 different CLIP backbones, which are ViT-L/14@336px, ViT-L/14, ViT-B/16, ViT-B/32, RN50x64, RN50x16, RN50x4, RN101, RN50. It is worth noting that from our model's perspective, each backbone could be seen as a dataset of its own, providing a different source of scores for each image set. We use the default 224x224 resolution, resizing without center crop, for all ViT models. For ResNet models, we use the same resolution as [7] except when otherwise stated.", + "bbox": [ + 89, + 314, + 482, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Our method improves over vanilla ZSCLIP across many datasets and models", + "text_level": 1, + "bbox": [ + 89, + 477, + 482, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a baseline, we compute singleton scores using prompt-ensembling with the 80 templates from the CLIP paper [16]. We then compute compound prompt scores and apply our method to the singletons and compounds. We compare our results to this vanilla ZSCLIP baseline.", + "bbox": [ + 89, + 517, + 482, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 shows our results over three different datasets and nine different CLIP backbones. We see that our method is able to improve over vanilla ZSCLIP by $12.6\\%$ on COCO, $8.8\\%$ on VOC, and $7.9\\%$ on NUSWIDE. Moreover, we see that our method achieves consistent improvements across all datasets and backbones, showcasing its generality. For example, the improvements range from $11.4 - 15.0\\%$ for COCO, $7.4 - 10\\%$ for VOC, and $6.1 - 10.8\\%$ for NUSWIDE. In fact, we see in the Supplementary that SPARC consistently improves the AP of almost every class in all datasets.", + "bbox": [ + 89, + 593, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Our method complements local features", + "text_level": 1, + "bbox": [ + 89, + 755, + 431, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider another training-free zero-shot baseline introduced by the TaI-DPT paper [4] which we call \"CLIP-DPT\". This method modifies ZSCLIP by replacing the final global pooling in the image backbone with a parameter-free attention pooling where the query is the text embedding, allowing us to harness local features that might otherwise be destroyed. We incorporate CLIP-DPT into our method, in the same way that we incorporated ZSCLIP in the previous", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "section, and see a consistent improvement in Tab. 3, suggesting that our method is adding a complementary strength and not just compensating for lost local information.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Our method has complementary strengths", + "text_level": 1, + "bbox": [ + 511, + 148, + 872, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several existing methods adapt CLIP to MLR tasks in an image-free ([4], [22], [21], [11], [23]) or training-free ([8], [10]) manner. These approaches lie outside our scope, as we focus on methods that are deployable 'out-of-the-box' on new datasets and VLMs, without requiring dataset-specific training or VLM surgery. Our black-box method allows easy 'plug-and-play' integration, using scores from existing methods in place of the vanilla ZSCLIP singleton scores.", + "bbox": [ + 511, + 172, + 903, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We combine our method with three existing methods that have public codebases: (1) TagCLIP [10], a training-free method that benchmarks on COCO and VOC with a ViT-B/16 backbone; (2) TaI-DPT [4], an image-free method that requires training, which benchmarks on COCO, VOC, and NUSWIDE with a RN50 backbone; (3) CoMC [11], an image-free method that requires training and has a publicly-available RN50 model for COCO for one train seed.", + "bbox": [ + 511, + 294, + 903, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are some nuances regarding image preprocessing and score postprocessing. TaI-DPT and CoMC use different image preprocessing settings than us - theirs are lower resolution and take a center crop. For a fair comparison, we use these same preprocessing settings when getting compound prompt scores to combine with TaI-DPT and CoMC. We also try running both our and their method with our preprocessing setting, which we denote as \"RN50*\". TagCLIP takes images at their native resolution, so for that comparison we do the same in order to be apples-to-apples. As for score postprocessing, TagCLIP uses softmax, which we have found to have a normalizing effect, so we omit our image-level normalization on TagCLIP's scores, and take the log of their scores to make them compatible with ours. We make no postprocessing changes for TaI-DPT or CoMC.", + "bbox": [ + 511, + 416, + 903, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the results of these plug-and-play combinations in Tab. 4 and Tab. 5. We find that method is able to improve on TagCLIP by on average $1.6\\%$ and on TaI-DPT by on average $1.7\\%$ . We do not improve on CoMC, but the degradation is only $0.3\\%$ . These results suggest that our method's signal complements those obtained through training and architectural manipulation by other methods.", + "bbox": [ + 511, + 643, + 903, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 511, + 760, + 624, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1 Normalization Module", + "text_level": 1, + "bbox": [ + 511, + 785, + 725, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We take a closer look at the Normalization module by seeing how it impacts the performance of both singleton prompts and our full pipeline. We see in Tab. 7 that the Normalization module provides significant gains in both cases (7.7% with singletons, 8.6% full pipeline), confirming that bias significantly impacts MLR when left unaddressed.", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4318", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/967404bff832c4da37ff93b914109592b13c6c59cba2cd766ea8ae64bbb055d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetViT-L/14 336pxViT-L/14ViT-B/16ArchitecturesAvg (archs)
ViT-B/32RN50 x64RN50 x16RN50 x4RN101RN50
COCOZSCLIP59.158.155.550.958.658.155.752.753.055.7
Ours70.569.567.464.070.670.169.567.765.768.3
VOCZSCLIP81.379.979.977.583.181.680.680.379.780.4
Ours88.988.388.787.590.590.390.089.789.289.2
NUSWIDEZSCLIP41.441.040.938.739.838.538.835.738.539.3
Ours47.547.147.346.947.447.948.346.545.847.2
", + "bbox": [ + 94, + 88, + 911, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49e25a27821f3f872e0236e26d6138132d862840c43215e38c4b92ddd5bdfe8f.jpg", + "table_caption": [ + "Table 2. Results over three datasets and nine CLIP backbones. Our proposed method consistently outperforms the ZSCLIP baseline across all datasets and architectures exhibiting its effectiveness on multi-label recognition tasks." + ], + "table_footnote": [], + "table_body": "
RN50 x64RN50 x16RN50 x4RN 101RN 50
COCOCLIP-DPT70.564.264.259.757.1
Ours-DPT77.873.972.769.868.4
VOCCLIP-DPT86.682.486.183.582.6
Ours-DPT91.189.888.588.589.9
NUSCLIP-DPT39.837.839.338.438.1
Ours-DPT43.545.542.742.644.5
", + "bbox": [ + 94, + 284, + 504, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7b6b8b590cac581a66f1a6ac81661209268a0c678cba90a845476e20392652d3.jpg", + "table_caption": [ + "Table 3. Comparing SPARC with local-feature method CLIP-DPT [4]. Our method still improves upon local features, showing that its strength is complementary." + ], + "table_footnote": [], + "table_body": "
COCO\nViT-B/16VOC\nViT-B/16Avg
TagCLIP70.991.781.3
+Ours73.892.082.9
", + "bbox": [ + 129, + 474, + 439, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Shows complementarity with architectural approaches, improving TagCLIP by $1.6\\%$", + "bbox": [ + 89, + 556, + 482, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8efa8dba3110c470bc03f30584068af4beb600573fb9925b77907b4a81449ada.jpg", + "image_caption": [ + "Figure 4. Average mAP for different Rank Fusion strategies demonstrates superiority of adaptive fusion over fixed strategies." + ], + "image_footnote": [], + "bbox": [ + 145, + 584, + 400, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Rank Fusion Module", + "text_level": 1, + "bbox": [ + 89, + 780, + 290, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our next ablation takes a closer look at the role of the Rank Fusion module. We consider some possible handcrafted strategies for combining compound prompt scores:", + "bbox": [ + 89, + 805, + 482, + 849 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- \" $k$ max\": Use the $k$ -th highest compound score, i.e. $r_{i,k}^{t}$", + "- \"mean $> = k$ \": Use partial avg $\\frac{1}{m_i - K + 1}\\sum_{k = K}^{m_i - K + 1}r_{i,k}^t$" + ], + "bbox": [ + 89, + 849, + 472, + 886 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 shows the average mAP of these, as well as \"maxVari", + "bbox": [ + 89, + 886, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ance” (3) and singleton-only, across all datasets and backbones. We see that 1st-max does much worse than 2nd-max, which underperforms 3rd-max, and so on, reaching a plateau around 5th-max. This lends credence to the idea that a weakened max is a stronger signal than outright max. We also see that the mean of all compound prompts (i.e. “mean $>= 1$ ) does surprisingly well. This is actually consistent with our intuition on weakened maxes - the mean weakens the higher maxes by averaging them with the lower ones. We can improve further by excluding high ranks in the “mean $>= k$ \" strategy. However, we can surpass all of these by adaptively weighting scores using the max-variance direction. The Supplementary repeats this analysis without the “merge” step (5), which turns out to be critical as well.", + "bbox": [ + 511, + 287, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3 Compound Prompts", + "text_level": 1, + "bbox": [ + 511, + 525, + 709, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our next ablation takes a closer look at the importance of the composition of our set of compound prompts and the use of rough cooccurrence info. The results are in Tab. 6. We start by including all pairs of classes in our set of formulaic pairwise prompts (i.e. of the form \"A and B\") and not including any other kind of compound prompt. We find that this gives us a $1.3\\%$ average boost over normalized singletons. Filtering these pairwise prompts by cooccurrence, as our proposed method does, yields a further $0.5\\%$ boost, and adding triplet and descriptive prompts gives an additional $0.3\\%$ boost. It makes sense that using the full set of class pairs would still have some benefit, without any cooccurrence filtering, if all the classes as a whole tend to cooccur positively more often than negatively, as that would mean that they are on average a helpful signal for predicting each other's presence. It might also be the case that related classes provide a useful context signal to CLIP. In any case, it is noteworthy that Rank Fusion can still help even when no cooccurrence information is required from the user.", + "bbox": [ + 511, + 551, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our final ablation aims to figure out whether our performance boost actually comes from the semantics of the compound prompts themselves, and not just an increase in number of prompts. Each class has on average $\\leq 20$ compound", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4319", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ee5c00b46006c79fa00b9cced55b45d7d91f7134ca59f4def06d6bd80e0faee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
COCOVOCNUSWIDEAvgCOCOAvg
RN50RN50*RN50RN50*RN50RN50*RN50RN50*
TaI-DPT65.168.288.588.046.243.466.6CoMC68.871.370.0
+Ours68.270.190.390.246.644.668.3+Ours68.970.669.7
", + "bbox": [ + 112, + 88, + 880, + 164 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ac660ea03eb03782ee121c767c600648259f16247fd2ee786ba5e2445e7773b.jpg", + "table_caption": [ + "Table 5. Results from combining our method with TaI-DPT (left) and CoMC (right) showing compatibility with training-based methods. Our method improves TaI-DPT by $1.7\\%$ while preserving the existing strong signal of CoMC, degrading it by only $0.3\\%$ ." + ], + "table_footnote": [], + "table_body": "
NormalizePair promptsTriplets + DescriptiveRank Fusion StrategyCOCOVOCNUSAvg
✓--65.987.745.166.2
✓all pairsours67.588.546.467.5
✓all pairsmean67.588.546.467.5
✓cooccurrence-filteredours68.189.047.068.0
✓cooccurrence-filteredmean67.988.546.867.7
✓cooccurrence-filtered✓ours68.389.247.268.3
", + "bbox": [ + 99, + 203, + 893, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f04526a18b4f8941016938f8da732bdd33438ef68ec77436ad8564d6cb518f74.jpg", + "table_caption": [ + "Table 6. Ablations on the makeup of our compound prompts. We find that we can still enjoy some benefit from pairwise prompts without any cooccurrence filtering." + ], + "table_footnote": [], + "table_body": "
CompoundNormalizeCOCOVOCNUSAvg
55.780.439.358.5
✓65.987.745.166.2
✓58.480.540.059.7
✓✓68.389.247.268.3
", + "bbox": [ + 102, + 359, + 467, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablations on Normalization module. Quantifies impact of normalization both with and without compound prompts.", + "bbox": [ + 89, + 444, + 482, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prompts, 4x less than the 80 singleton prompt templates already used by vanilla ZSCLIP, but there is still an increase in total number of prompts, so an ablation is warranted.", + "bbox": [ + 89, + 482, + 482, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We take inspiration from WaffleCLIP [5], which found that prompt ensembles made of random texts could perform just as well as descriptive ensembles generated by LLMs. For our ablation, we use WaffleCLIP-style prompts in place of compound prompts. Specifically, for eachclassname $c_i$ we create 30 prompts of the form \"A photo of a $c_i$ , which is [RAND]\", where \"[RAND]\" is 10 random characters. We try both the \"maxVariance\" and \"mean $> = 1$ \" fusion strategies, as the latter is more typical for assembling.", + "bbox": [ + 88, + 529, + 482, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the results of this ablation in the Supplementary. We see that randomized compound prompts do not offer any complementary signal to the singleton prompts. Thus, any gain from compound prompts must come from their semantics. It is unlikely that simple prompt diversity explains the gains either, given that vanilla ZSCLIP already uses a large and diverse template ensemble.", + "bbox": [ + 89, + 666, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. A qualitative look at the weakened max", + "text_level": 1, + "bbox": [ + 89, + 785, + 429, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We saw in Sec. 4.4.1 that the second-highest compound prompt score is a much better signal than the highest one. This counterintuitive finding can be explained through object cooccurrence. When CLIP is given a compound prompt like 'cat and sofa', it has OR-like behavior - giving high scores if either object is present. The maximum score often", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comes from such prompts where only one object is present but strongly detected. In contrast, the second-highest scores tend to come from prompts where both objects have moderate confidence, providing a more reliable signal. This explains why using second-highest scores results in better separation between positive and negative cases.", + "bbox": [ + 511, + 362, + 903, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3 shows an example of what happens at an image level. On the right is an image that contains a cat (curled up behind the two dogs on the bed); on the left is an image without a cat. The highest-scoring compound prompt is \"Cat and Dog\" for both images. This prompt gives a high score to the right image by detecting the dogs, but it also detects the dog in the left image for the same reason, and so it also misorders the images. But the second-highest-scoring prompts - \"Bed and Cat\" for right image, \"Cat and Potted plant\" for left - correctly order the images.", + "bbox": [ + 511, + 452, + 905, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show score histograms for one of the classes in the Supplementary to illustrate what happens in aggregate - the second-highest score lifts positive examples without introducing as many false-positives as first-highest score.", + "bbox": [ + 511, + 604, + 903, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 676, + 638, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented SPARC, a zero-shot multi-label recognition approach that improves VLM performance without training data or architectural changes. Our key insights - the suboptimal nature of maximum scores and presence of systematic biases - led to two complementary innovations: compound prompts with adaptive fusion, and systematic normalization. Beyond improving standalone VLM performance, SPARC enhances existing zero-shot and training-based methods while maintaining a black-box approach. The success of our method reveals fundamental properties of VLM scoring behavior, suggesting promising directions for improving zero-shot recognition through score analysis rather than architectural modification or finetuning.", + "bbox": [ + 511, + 702, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4320", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 3", + "[2] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM International Conference on Image and Video Retrieval, New York, NY, USA, 2009. Association for Computing Machinery. 6", + "[3] Reza Esfandiarpoor and Stephen H Bach. Follow-up differential descriptions: Language models resolve ambiguities for image classification. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 3", + "[4] Guo et al. Texts as images in prompt tuning for multi-label image recognition. In CVPR, '23. 3, 6, 7", + "[5] Roth et al. Waffling around for performance: visual classification with random words and broad concepts ICCV, '23. 8, 3", + "[6] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6", + "[7] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6", + "[8] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 3, 6", + "[9] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[10] Yuqi Lin, Minghao Chen, Kaipeng Zhang, Hengjia Li, Mingming Li, Zheng Yang, Dongqin Lv, Binbin Lin, Haifeng Liu, and Deng Cai. Tagclip: A local-to-global framework to enhance open-vocabulary multi-label classification of clip without training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3513-3521, 2024. 3, 6", + "[11] Yicheng Liu, Jie Wen, Chengliang Liu, Xiaozhao Fang, Zuoyong Li, Yong Xu, and Zheng Zhang. Language-driven cross-modal classifier for zero-shot multi-label image recognition. In *Forty-first International Conference on Machine Learning*, 2024. 3, 6", + "[12] Leilei Ma, Hongxing Xie, Lei Wang, Yanping Fu, Dengdi Sun, and Haifeng Zhao. Text-region matching for multi-label image recognition with missing labels. In Proceedings of the 32nd ACM International Conference on Multimedia, page 6133-6142. ACM, 2024. 3", + "[13] Sachit Menon and Carl Vondrick. Visual classification via" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3", + "[14] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023.", + "[15] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3", + "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6", + "[17] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In Proceedings of the International Conference on Pattern Recognition (ICPR), 2024. 3", + "[18] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems, 35:30569-30582, 2022. 3", + "[19] Hao Tan, Zichang Tan, Jun Li, Jun Wan, Zhen Lei, and Stan Z. Li. Sspa: Split-and-synthesize prompting with gated alignments for multi-label image recognition. ArXiv, abs/2407.20920, 2024. 3", + "[20] Ao Wang, Hui Chen, Zijia Lin, Zixuan Ding, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Guiguang Ding. Hierarchical prompt learning using clip for multi-label classification with single positive labels. In Proceedings of the 31st ACM International Conference on Multimedia, page 5594-5604, New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[21] Xiangyu Wu, Qing-Yuan Jiang, Yang Yang, Yi-Feng Wu, Qing-Guo Chen, and Jianfeng Lu. Tai++: Text as image for multi-label image classification by co-learning transferable prompt. arXiv preprint arXiv:2405.06926, 2024. 3, 6", + "[22] Shuo Yang, Zirui Shang, Yongqi Wang, Derong Deng, Hongwei Chen, Qiyuan Cheng, and Xinxiao Wu. Data-free multi-label image recognition via llm-powered prompt tuning. arXiv preprint arXiv:2403.01209, 2024. 3, 6", + "[23] Xueling Zhu, Jiuxin Cao, Jian Liu, Dongqi Tang, Furong Xu, Weijia Liu, Jiawei Ge, Bo Liu, Qingpei Guo, and Tianyi Zhang. Text as image: Learning transferable adapter for multi-label classification. ArXiv, abs/2312.04160, 2023. 3, 6" + ], + "bbox": [ + 517, + 93, + 903, + 772 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Ping Hu, Siqi Wang, and Piotr Teterwak for their valuable discussions, as well as the anonymous reviewers for their thorough and helpful feedback. This research was supported by the Army Research Office Grant W911NF2110246, AFRL Grant FA8650-22-C1039, and the National Science Foundation grants CPS-2317079, CCF-2007350, and CCF-1955981.", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4321", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_model.json b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c195665daa4e7e871d5258484e59cb9da22340f --- /dev/null +++ b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_model.json @@ -0,0 +1,1901 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.873, + 0.177 + ], + "angle": 0, + "content": "SPARC: Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.204, + 0.371, + 0.256 + ], + "angle": 0, + "content": "Kevin Miller \nBoston University \nnivek@bu.edu" + }, + { + "type": "text", + "bbox": [ + 0.426, + 0.204, + 0.573, + 0.257 + ], + "angle": 0, + "content": "Aditya Gangrade \nBoston University \ngangrade@bu.edu" + }, + { + "type": "text", + "bbox": [ + 0.628, + 0.204, + 0.774, + 0.256 + ], + "angle": 0, + "content": "Samarth Mishra \nBoston University \nsamarthm@bu.edu" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.265, + 0.543, + 0.3 + ], + "angle": 0, + "content": "Kate Saenko \nBoston University and Meta AI (FAIR)" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.304, + 0.446, + 0.316 + ], + "angle": 0, + "content": "saenko@bu.edu" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.265, + 0.768, + 0.316 + ], + "angle": 0, + "content": "Venkatesh Saligrama \nBoston University \nsrv@bu.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.352, + 0.327, + 0.368 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.485, + 0.718 + ], + "angle": 0, + "content": "Zero-shot multi-label recognition (MLR) with Vision-Language Models (VLMs) faces significant challenges without training data, model tuning, or architectural modifications. Existing approaches require prompt tuning or architectural adaptations, limiting zero-shot applicability. Our work proposes a novel solution treating VLMs as black boxes, leveraging scores without training data or ground truth. We make two contributions. First, we find that VLM scores suffer from image- and prompt-specific biases, and that simple standardization is surprisingly effective at removing these and boosting MLR performance. And second, we introduce compound prompts grounded in realistic object combinations. Our analysis reveals \"AND\"/\"OR\" signal ambiguities that cause maximum compound scores to be surprisingly suboptimal compared to second-highest scores. We introduce an adaptive fusion method to address this issue. Our method enhances other zero-shot approaches, consistently improving their results. Experiments show superior mean Average Precision (mAP) compared to methods requiring training data, achieved through refined object ranking for robust zero-shot MLR. Code can be found at https://github.com/kjmillerCURIS/SPARC." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs), such as CLIP [16], have emerged as general-purpose systems for understanding visual data through language-based queries. These models enable a broad range of applications, from object detection to image captioning, by linking visual inputs to language prompts. In standard settings where images contain single, recognizable objects, VLMs perform remarkably well. However, for the more complex task of zero-shot multi-label recognition (MLR) (Fig. 1 (top)), where mod" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.351, + 0.897, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.698 + ], + "angle": 0, + "content": "Figure 1. (top) Vision-Language Models (VLMs) like CLIP can be used for zero-shot classification with image-text similarity scores. While this works fairly well for single-class labels, they can struggle in the multi-label scenario. (bottom) In this paper, we introduce SPARC, our solution that functions on top of an existing VLM, treating it simply as a black-box score generator. Using class names, SPARC first creates compound prompts for additional queries to the VLM. It then normalizes, ranks and appropriately fuses them to generate final scores for the original classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.853 + ], + "angle": 0, + "content": "els must identify multiple objects within an image without prior training on specific data, VLMs face significant limitations. Zero-shot MLR is crucial for applications in fields like robotics and medical imaging, where objects rarely appear in configurations that align neatly with training distributions. In these scenarios, achieving robust multi-label recognition without fine-tuning is challenging, given the task's reliance on mean Average Precision (mAP) scores, which depend on ranking images for object presence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "VLM: Prompt Dependent AND/OR Noisy Channel. Despite the promise of zero-shot capabilities, current VLM approaches often struggle with MLR due to inherent scoring" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.397 + ], + "angle": 0, + "content": "Figure 2. A motivating example for the Normalization module, with CLIP scores on a slice of the COCO dataset, before and after normalization. The \"necktie\" class is present for examples in the top half of the plot and absent in the bottom half. Image- and prompt-level biases show up as horizontal and vertical striations; normalization removes these and creates better separation." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.409, + 0.48, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Figure 3. A motivating example for the Rank Fusion module, with an image where class \"cat\" is absent (left) and one where it is present (right). The highest compound prompt score is an unhelpful signal because it gives a high score to both negatives and positives, while the second-highest is more discriminative. Our method adaptively fuses the most informative order statistics, resulting in a strong signal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "behaviors and biases. The performance of these models is hampered by a mix of conjunction (AND) and disjunction (OR) behaviors in their scoring, leading to inflated scores in compound prompts that contain multiple objects. For example, a prompt like \"cat and sofa\" might yield a high score even if only one of these objects is present in the image. This tendency reflects biases learned during training, where common object pairs receive higher scores even when only one object is present, disrupting the accuracy of mAP-based evaluations. Furthermore, existing methods for adapting VLMs to zero-shot MLR frequently rely on prompt tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "or architectural adjustments—approaches that are often dependent on training data and computationally intensive finetuning, which limit their generalizability to novel tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.142, + 0.907, + 0.323 + ], + "angle": 0, + "content": "Our Approach. In contrast to these methods, we introduce SPARC (Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in VLMs), a novel approach to zero-shot MLR that bypasses the need for training data, prompt tuning, or model-specific modifications. Our method treats the VLM as a black box, relying solely on its output scores to infer object presence (see Fig. 1). This black-box approach enables us to avoid assumptions about the model's internal workings, allowing for a purely zero-shot framework that is both model-agnostic and dataset-independent. SPARC introduces two main innovations that address the unique challenges of zero-shot MLR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.909, + 0.463 + ], + "angle": 0, + "content": "A. Score Normalization. Our analysis finds that VLM scores suffer from image-level and prompt-level biases. Image-level bias refers to noise terms that are constant across prompts but vary between images, while prompt-level bias refers to the converse. Image-level bias in particular hampers mAP as it changes the ordering of scores within each class. Standardization is surprisingly effective at addressing this problem in MLR. Standardization alone improves mAP by \\(6 - 10\\%\\) on COCO, VOC, and NUSWIDE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.644 + ], + "angle": 0, + "content": "B. Compound Prompts and Adaptive Fusion. Recognizing that VLMs can provide richer information when prompted with combinations of objects, we develop a method for constructing compound prompts. These prompts reflect likely contextual associations between objects, such as \"cat and sofa\" or \"car and bus.\" By gathering scores from these compound prompts, we can capture a spectrum of potential object contexts within the image, enhancing detection without relying on training-based adaptations. This composition strategy allows us to agnostically extract information from the VLM without depending on any specific dataset or VLM architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Surprisingly, we observe that the maximum score among compound prompts is often a poor proxy for true object presence. Although one might expect the highest score to serve as a reliable signal, we find that it suffers because VLMs tend to respond to compound prompts with OR-like behavior, raising scores even when only one object in the prompt is present. Instead, we observe that the second-highest score consistently outperforms the maximum score by minimizing the effects of false positives caused by OR-like behavior. Building on this insight, we develop a PCA-based fusion method that combines information from multiple order statistics from the compound prompts, as well as the singleton prompts, ultimately optimizing mAP by enhancing score accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Complementarity. SPARC is complementary to other zero-shot and training-free MLR methods. When applied on top of these approaches, SPARC consistently enhances" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "mAP scores by refining object ranking and reducing bias in VLM outputs. This capability makes SPARC an adaptable solution that can improve upon existing methods while maintaining a fully zero-shot, model-agnostic framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Empirical Results. SPARC achieves significant improvements in mAP, outperforming methods that incorporate architectural modifications. This outcome shows the potential of a fully zero-shot approach that relies only on systematic prompt design and score interpretation, rather than prompttraining or fine-tuning. By revealing that the second-highest score can be a superior proxy to the maximum, our findings provide new insights into VLM scoring behavior, suggesting that careful treatment of prompt compositions and score patterns can unlock robust MLR capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.316, + 0.236, + 0.33 + ], + "angle": 0, + "content": "1.1. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.491 + ], + "angle": 0, + "content": "Supervised Methods. Prior methods improve MLR performance through various approaches. DualCoOp [7, 18] trains prompts for text-guided spatial attention. Subsequent works use class co-occurrences to refine CLIP logits [17, 20]. Hierarchical structuring [20] encourages related classes to learn similar prompts. SSPA [19] combines embeddings from learned and LLM-generated prompts, and TRM-ML [12] uses pseudolabels to match texts to image regions. While effective, these methods require annotated training data. In contrast, ours is unsupervised." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Unsupervised Training-Based Methods. TaI-DPT [4] uses caption-only data to adapt CLIP to MLR by training a prompt-generation network using caption embeddings. Extensions include LLM-generated captions [22], pseudovisual prompts [21], and lightweight classifiers on caption embeddings [11, 23]. Separately, CDUL [1] leverages CLIP scores with spatial aggregation to create training pseudolabels. While these approaches demonstrate shared vision-language embedding power, they require significant training and CLIP internal access. In contrast, our black-box method uses crude cooccurrence-based compound prompts (essentially deleting implausible contextual associations) to probe multilabel information, fusing scores for effective MLR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Unsupervised Training-free methods. The recent literature has proposed methods to adapt CLIP to the MLR taks in a training-free manner via architectural changes. CLIP-Surgery [8] alters the architecture to increase the coherence of attention masks within the image backbone, and normalizes features on a patch-by-patch basis. TagCLIP [10] uses internal attention masks to refine scores at a patch level based on regional coherence, and aggregate queries over crops of the image. Again, both methods require access to VLM internals, while ours works in a black-box manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Prompt Enhancement for Single-Label Recognition. Several methods use LLM-generated descriptive prompts for single-label recognition, combining them via max or mean fusion [3, 13-15]. While related, our compound" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "prompts specifically target MLR by modeling class relationships rather than enriching individual class descriptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Complementarity of SPARC to Unsupervised Methods. We explicitly note that our methodology is complementary to much of the work on unsupervised VLM-based MLR. For instance, since we assume only black-box access to VLM scores, our compound prompts could be refined via prompt tuning as in TaI-DPT [4] and/or applied patch-by-patch and integrated à la TagCLIP [10]. Scores from any of these methods can be combined via our normalization and rank-fusion approach. §4.3 investigates the resulting gains." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.271, + 0.605, + 0.287 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.297, + 0.871, + 0.312 + ], + "angle": 0, + "content": "We detail our method, SPARC (see Alg. 1 and Fig. 1)." + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.316, + 0.905, + 0.576 + ], + "angle": 0, + "content": "Algorithm 1 SPARC Pipeline \nInput: Images \\(\\mathcal{I}\\), Class Names \\(\\mathcal{C}\\) \nOutput: Final scores \\(\\zeta_i^t\\) for each image \\(t\\) and class \\(i\\) \n\\(P \\leftarrow\\) GenerateCompoundPrompts \\((\\mathcal{C})\\) (Details in Supp.) \nfor \\(t \\in \\mathcal{I}\\) do \n\\(s_i^t \\leftarrow\\) GetVLMScores \\((t, i)\\) (Singleton scores) \n\\(\\forall p \\in P, s_p^t \\leftarrow\\) GetVLMScores \\((t, p)\\) (Compound scs.) \n// (Normalize scores (Eqs. 1, 2)) \n\\(\\tilde{s}_i^t \\leftarrow\\) ImageNorm \\((s_i^t), \\tilde{s}_p^t \\leftarrow\\) ImageNorm \\((s_p^t)\\) \n\\(\\bar{s}_i^t \\leftarrow\\) PromptNorm \\((\\tilde{s}_i^t), \\bar{s}_p^t \\leftarrow\\) PromptNorm \\((\\tilde{s}_p^t)\\) \nfor \\(i \\in \\mathcal{C}\\) do \n\\(r_{i,k}^t \\leftarrow\\) GetOrderStatistics(\\(\\{\\bar{s}_p^t : i \\in p\\}\\)) \n\\(w^{i*} \\leftarrow\\) MaxVarDirection([\\bar{s}_i^t, r_{i,k}^t]) (Eq. 3) \n\\(\\tilde{\\zeta}_i^t \\leftarrow w_0^{i*} \\bar{s}_i^t + \\sum_k w_k^{i*} r_{i,k}^t\\) (Eq. 4) \n\\(\\zeta_i^t \\leftarrow \\bar{s}_i^t + \\tilde{\\zeta}_i^t\\) (Eq. 5) \nreturn \\(\\{\\zeta_i^t\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Setup. Suppose we have \\(M\\) images and \\(N\\) target classes, with classnames \\(c_{1},\\ldots ,c_{N}\\). We assume that we have a set of \"singleton\" prompts describing each of these classes in isolation (e.g. \"picture of a cat\"). Given an image \\(t\\), let \\(s_1^t,\\dots,s_N^t\\) denote the similarity scores produced by a VLM when comparing the singleton prompts to that image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.84 + ], + "angle": 0, + "content": "A naive approach is to directly use these singleton scores to perform MLR. However, since the presence of a particular class alters the conditional probability of the remaining classes in images, it should be possible to further refine these scores by accounting for the multi-label structure of the images. In order to do this, we use the class-names to generate a further set of \"compound\" prompts \\( P \\) which mention multiple classes. For instance, the classes \"cat\" and \"sofa\" may be compounded to \"cat and sofa\". Let \\( \\{s_p^t : p \\in P\\} \\) denote the scores given by the VLM when comparing these compound prompts to image \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Using the singleton as well as the compound prompts above, SPARC produces a vector of refined class-wise scores \\(\\zeta_1^t,\\dots ,\\zeta_N^t\\) , for each image \\(t\\) . Given this structure, SPARC has three main components: (i) compound prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4315" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "generation to enable the use of CLIP scoring for MLR (ii) image and prompt level normalization to allow scores to be directly compared, and (iii) Rank Fusion to combine singleton and compound prompt scores into a single score per class. We now discuss each of these aspects in detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.175, + 0.367, + 0.192 + ], + "angle": 0, + "content": "2.1. Compound Prompt Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.484, + 0.516 + ], + "angle": 0, + "content": "Our compound prompt generation method relies on contextual associations from commonly observed visual patterns to inform prompt structure. We provide pseudocode for the process in the Supplementary. The process takes as input the classnames \\( c_{1}, \\ldots, c_{N} \\), ground-truth cooccurrence statistics \\( \\mathbb{P} \\), thresholds \\( \\tau_{2} \\) and \\( \\tau_{3} \\) (fixed to 0.05 and 0.025 for all datasets), and off-the-shelf LLM \\( \\Phi \\). We use probability thresholding to select pairs and triplets of classes that could plausibly coocur in realistic visual scenes. These pairs and triplets are used to make formulaic compound prompts of the form \"A and B\" and \"A, B, and C\". We feed these formulaic prompts to LLM \\( \\Phi \\) and ask it to generate natural sentences from them. Our total set of compound prompts comprises the formulaic pair, formulaic triplet, and natural sentence prompts, which add up to on average \\( \\leq 20 \\) compound prompts per class in all of our datasets. Note that probabilities \\( \\mathbb{P} \\) are not used in any part of our method other than compound prompt generation, which is why we say that it uses \"rough\" cooccurrence info. In fact, our ablations reveal that we still get a performance gain when using all possible class pairs instead of filtering by probability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.239, + 0.538 + ], + "angle": 0, + "content": "2.2. Normalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Once we have obtained compound prompts \\( P \\), we can query the VLM to obtain singleton scores \\( s_1^t, \\ldots, s_N^t \\) and compound scores \\( \\{s_p^t : p \\in P\\} \\) for each image \\( t \\). However, these scores contain both image- and prompt-specific biases. The former in particular is a problem, as it can change the order of scores between negative and positive images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.668 + ], + "angle": 0, + "content": "We address image-level bias for singleton and compound prompts, respectively, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.674, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\tilde {s} _ {i} ^ {t} := \\frac {s _ {i} ^ {t} - \\hat {\\mu} (s _ {i} ^ {t})}{\\hat {\\sigma} (s _ {i} ^ {t})} \\quad a n d \\quad \\tilde {s} _ {p} ^ {t} := \\frac {s _ {p} ^ {t} - \\hat {\\mu} (\\check {s} _ {i} ^ {t})}{\\hat {\\sigma} (\\check {s} _ {i} ^ {t})}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.793 + ], + "angle": 0, + "content": "where \\(\\hat{\\mu}(s_{\\cdot}^{t}), \\hat{\\mu}(\\check{s}_{\\cdot}^{t}), \\hat{\\sigma}(s_{\\cdot}^{t}), \\hat{\\sigma}(\\check{s}_{\\cdot}^{t})\\) are sample means and standard deviations across the prompt dimension for a single image. \\(\\check{s}_{1}^{t}, \\ldots, \\check{s}_{N}^{t}\\) are the scores of \"auxiliary\" prompts that mention classnames in isolation and are used only to obtain statistics for standardizing compound prompt scores." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.793, + 0.483, + 0.808 + ], + "angle": 0, + "content": "We also standardize across images to quash prompt bias:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.827, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\overline {{s}} _ {i} ^ {t} := \\frac {\\tilde {s} _ {i} ^ {t} - \\hat {\\mu} (\\tilde {s} _ {i} ^ {*})}{\\hat {\\sigma} (\\tilde {s} _ {i} ^ {*})} a n d \\overline {{s}} _ {p} ^ {t} := \\frac {\\tilde {s} _ {p} ^ {t} - \\hat {\\mu} (\\tilde {s} _ {p} ^ {*})}{\\hat {\\sigma} (\\tilde {s} _ {p} ^ {*})}, (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{\\mu} (\\tilde{s}_i)\\), \\(\\hat{\\mu} (\\tilde{s}_p)\\), \\(\\hat{\\sigma} (\\tilde{s}_i)\\), and \\(\\hat{\\sigma} (\\tilde{s}_p)\\) now denote sample means and SDs across the images. Removing prompt-level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "bias makes scores from different prompts more compatible with each other, which is important for fusing them." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.13, + 0.649, + 0.145 + ], + "angle": 0, + "content": "2.3. Rank Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.905, + 0.214 + ], + "angle": 0, + "content": "Our goal now is to find a way to use compound scores to strengthen the singleton scores. Let us define a bit more notation here. Let \\( r_{i,k}^{t} \\) denote the score of the \\( k \\)-th highest scoring compound prompt that mentions class \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.906, + 0.456 + ], + "angle": 0, + "content": "A natural choice might be to choose \\( r_{i,1}^{t} \\), i.e. the highest scoring prompt. After all, we would expect this to be the prompt that most closely describes the image, and if the image did not contain \\( c_{i} \\) then we would expect none of the prompts containing \\( c_{i} \\) to score high. However we find that \\( r_{i,1}^{t} \\) does not offer a good detection of classes. The mechanism behind this, which is explored in detail in §3, is that when using a compound prompt of the form 'A and B' to detect the class \\( A \\), the score sees a large increase if the object \\( B \\) truly occurs in the image. This 'OR'-like behavior (§3) means that maximum score \\( r_{i,1}^{t} \\) typically captures the effect of other classes than \\( A \\) being present in the image, and thus leads to very poor separation between ground truth positive and negative images for any class. Counterintuitively, then, we find that 'weakened maxes', i.e., lower order statistics like \\( r_{i,2}^{t} \\) and \\( r_{i,3}^{t} \\) yield much better separation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.532 + ], + "angle": 0, + "content": "The above fact suggests that we ought to utilize the entirety of the order statistics \\(\\{r_{i,k}^t\\}\\) in order to generate an effective fusion rule. We do fusion by taking a weighted sum where the weighting vector is the direction of highest variance, i.e., we compute the fused compound score \\(\\tilde{\\zeta}_i^t\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.553, + 0.907, + 0.584 + ], + "angle": 0, + "content": "\\[\nw ^ {i *} := \\underset {w ^ {i}} {\\arg \\max } \\operatorname {V a r} _ {t} \\left(w _ {0} ^ {i} \\bar {s} _ {i} ^ {t} + \\sum_ {k} w _ {k} ^ {i} r _ {i, k} ^ {t}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.588, + 0.907, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\zeta} _ {i} ^ {t} := w _ {0} ^ {i *} \\bar {s} _ {i} ^ {t} + \\sum_ {k} w _ {k} ^ {i *} r _ {i, k} ^ {t}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Note that the variance above is computed across examples, i.e. an iteration over example index \\( t \\). Computing arg max above is just a matter of running PCA and taking the first principal component. The motivation for this approach is similar to that for linear discriminant analysis. Imagine each example is a point in \\( \\mathbb{R}^{K_i} \\), where \\( K_{i} \\) is the number of compound prompts for class \\( i \\), and these points are distributed according to a mixture of two components, one for examples where \\( i \\) is present, the other for examples where \\( i \\) is absent. The shift between the means of these components should contribute to the variance in the direction of that shift. If we assume that these components are roughly homoskedastic, then, as long as the 'signal' induced by the mean shift is large compared to any heteroskedasticity, the highest variance direction \\( w^{i*} \\) is precisely the direction of the mean shift, and thus \\( \\tilde{\\zeta}_i^t \\) effectively computes the component of the score vector that lies along this signal direction. While principled, this max-variance strategy is of course not" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4316" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "necessary to execute, and we compare it to alternative fusion strategies in §4.4.1. We add (i.e. \"merge\") this fused score into the original singleton score to get our final score:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.151, + 0.483, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\zeta_ {i} ^ {t} := s _ {i} ^ {t} + \\tilde {\\zeta} _ {i} ^ {t} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.208 + ], + "angle": 0, + "content": "In the next section we probe CLIP scores through a statistical lens to build intuition on our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.222, + 0.308, + 0.238 + ], + "angle": 0, + "content": "3. Noise Model of a VLM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.484, + 0.353 + ], + "angle": 0, + "content": "Our methodological choices are strongly driven by the behavior of CLIP scores under compound prompts, and this leads to both the normalization approach taken, as well as our counterintuitive choice of fusing compound scores by weakened max instead of outright max. In order to justify these choices, we probe the structure of CLIP scores under compound prompts by constructing a statistical model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.475 + ], + "angle": 0, + "content": "Given an image \\( t \\) and a prompt with classes \\( i \\) and \\( j \\), we model the score of compound prompt 'i and j' as \\( s_{i,j}^{t} \\coloneqq \\theta_{1}^{t} \\cdot f(y_{i}^{t}, y_{j}^{t}) + \\theta_{0}^{t} + \\varepsilon \\) where \\( y_{i}^{t}, y_{j}^{t} \\in \\{0,1\\} \\) indicate ground-truth presence of the two classes, \\( \\theta_{1}^{t} \\) and \\( \\theta_{0}^{t} \\) represent image-level bias (e.g. due to visual domain-shifts), and \\( \\varepsilon \\) is IID Gaussian noise. Note that our Normalization module is well-suited to mitigate variations in \\( \\theta_{0}^{t} \\) and \\( \\theta_{1}^{t} \\), at least when \\( \\varepsilon \\) is not strongly dominating." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Now, the core object of interest is the map \\( f: \\{0,1\\}^2 \\to \\mathbb{R} \\), which is responsible for modeling class pair interactions and prompt-level biases (see Table 1). In an ideal situation for MLR, \\( f \\) would behave roughly as an 'AND' function, so that compound prompts would let us directly hone in on the presence of pairs of classes. However, prior work has observed that instead, for CLIP-type models, \\( f \\) tends to predominantly behave as an 'OR' function, in that \\( f \\) is high if either \\( i \\) or \\( j \\) is present. Notice that if \\( f \\) were to exactly behave as an OR, compound prompts would not be effective in detecting classes that were missed by the singleton scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.762 + ], + "angle": 0, + "content": "In experimentation, we found that the behavior of \\( s_{i,j}^{t} \\) is intermediary to the above extremes: CLIP scores for compound prompts behave qualitatively as an OR function with a small AND 'bonus' when both classes are present. This bonus, as well as the particular scores, are modulated by the classes mentioned (a consequence of prompt-level bias for each class) which we model as linear effects, leading to an overall model of the signal term of the CLIP score as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.773, + 0.483, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f \\left(y _ {i} ^ {t}, y _ {j} ^ {t}\\right) := \\max \\left(a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}\\right) \\\\ + \\delta_ {i j} \\min (a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}), \\qquad (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where the \\(a\\) terms are class-wise linear effect coefficients, and \\(\\delta_{ij}\\) captures the strength of this AND bonus." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Note that this AND bonus is a critical feature of CLIP scores in their use for MLR problems, since this bump allows us to both infer the presence of classes that singleton" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.262 + ], + "angle": 0, + "content": "
Noise Modelf(yi,t,yj,t)FVU
constant0.10.802
only ANDmin(a0i+yitai,aj+ytaj)0.536
only ORmax(a0i+yitai,aj+ytaj)0.288
additivea0i+yitai+aj+ytaj0.267
OR + static AND-bonusmax(a0i+yitai,aj+ytaj) + δ min(a0i+yitai,aj+ytaj)0.263
OR + variable AND-bonusmax(a0i+yitai,aj+ytaj) + δi,j min(a0i+yitai,aj+ytaj)0.248
look-up tableLUT(i,j)(yi,t,yj,t)0.235
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.267, + 0.907, + 0.323 + ], + "angle": 0, + "content": "Table 1. Comparison of fidelity of noise models for scoring pairwise compound prompts. Notice that the OR+AND-bonus model captures nearly all of the fidelity of the look-up table, and offers significant improvements over the simplified models above it." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.905, + 0.366 + ], + "angle": 0, + "content": "prompts have missed, and indirectly to filter false positives (since the true positives are further elevated by the bonus)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Table 1 shows the results of a systematic evaluation of the AND-OR structure of compound CLIP scores, using 81K COCO test images and ViT-L/14@336px backbone. We evaluate various models of \\( f \\), ranging from simple constant, 'only OR,' and 'only AND' models to additive models and OR+AND-bonus models with both a 'static' \\( \\delta \\) shared by all class pairs, and a 'variable' \\( \\delta_{i,j} \\) that is class-pairspecific. The final entry is a 'look-up table' (LUT) model where all four values of \\( f \\) can vary arbitrarily, which serves as the most expressive baseline in this model class. These are evaluated w.r.t. the fraction of variance unexplained (i.e., \\( 1 - R^2 \\) value). Observe that while the only OR model has remarkable fidelity, the additive and AND-bonus models significantly improve upon this, and allowing \\( \\delta \\) to vary with \\( i,j \\) yield a further significant improvement to within a \\( 6\\% \\) extra loss from the LUT model. This shows that our AND-bonus models capture a significant amount of the qualitative behavior of CLIP scores for compound prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.905, + 0.685 + ], + "angle": 0, + "content": "A closer look at these fitted models reveals the \\(\\delta\\) values are small and relatively stable: for static \\(\\delta\\), the recovered value is 0.56, and the IQR for varying \\(\\delta_{ij}\\) is (0.44, 0.57)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.905, + 0.745 + ], + "angle": 0, + "content": "We find similar behavior when examining other backbones and datasets, which we detail in the Supplementary. The Supplementary also contains a theoretical view of the benefits of using weakened-max, based on the noise model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.771, + 0.772 + ], + "angle": 0, + "content": "3.1. Methodological Implications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.856 + ], + "angle": 0, + "content": "The qualitative structure of the OR+AND-bonus behavior justifies our methodological choices in §2. Image-level normalization reduces the effect of variation in \\(\\theta_0^t\\) (via mean subtraction) and \\(\\theta_1^t\\) (via rescaling), while prompt-level normalization mitigates variations in \\(a_0^i\\) and \\(a_1^i\\) across classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We can also qualitatively observe that, because the OR component of the scores dominates the AND component, the maximum score \\( r_{i,1}^{t} \\) can create false positives when \\( i \\) is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.123 + ], + "angle": 0, + "content": "absent but related class \\( j \\) is present, and so we should look to subsequent order statistics for a better signal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.138, + 0.223, + 0.156 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.164, + 0.483, + 0.3 + ], + "angle": 0, + "content": "Datasets We benchmark our method on 3 different MLR datasets. They are COCO2014 [9], which has 40,504 test images with 80 classes; VOC2007 [6], which has 4.952 test images with 20 classes; and NUSWIDE [2], which has 107,859 test images with 81 classes. We use only test, and not train, images from these datasets. We do use their conditional ground-truth label distributions, computed from training set, to choose compound prompts, but do not use ground-truth probabilities in any other part of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.483, + 0.466 + ], + "angle": 0, + "content": "CLIP Backbones In order to showcase the generality of our method, we apply it to 9 different CLIP backbones, which are ViT-L/14@336px, ViT-L/14, ViT-B/16, ViT-B/32, RN50x64, RN50x16, RN50x4, RN101, RN50. It is worth noting that from our model's perspective, each backbone could be seen as a dataset of its own, providing a different source of scores for each image set. We use the default 224x224 resolution, resizing without center crop, for all ViT models. For ResNet models, we use the same resolution as [7] except when otherwise stated." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.51 + ], + "angle": 0, + "content": "4.1. Our method improves over vanilla ZSCLIP across many datasets and models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.592 + ], + "angle": 0, + "content": "As a baseline, we compute singleton scores using prompt-ensembling with the 80 templates from the CLIP paper [16]. We then compute compound prompt scores and apply our method to the singletons and compounds. We compare our results to this vanilla ZSCLIP baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Tab. 2 shows our results over three different datasets and nine different CLIP backbones. We see that our method is able to improve over vanilla ZSCLIP by \\(12.6\\%\\) on COCO, \\(8.8\\%\\) on VOC, and \\(7.9\\%\\) on NUSWIDE. Moreover, we see that our method achieves consistent improvements across all datasets and backbones, showcasing its generality. For example, the improvements range from \\(11.4 - 15.0\\%\\) for COCO, \\(7.4 - 10\\%\\) for VOC, and \\(6.1 - 10.8\\%\\) for NUSWIDE. In fact, we see in the Supplementary that SPARC consistently improves the AP of almost every class in all datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.433, + 0.772 + ], + "angle": 0, + "content": "4.2. Our method complements local features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We consider another training-free zero-shot baseline introduced by the TaI-DPT paper [4] which we call \"CLIP-DPT\". This method modifies ZSCLIP by replacing the final global pooling in the image backbone with a parameter-free attention pooling where the query is the text embedding, allowing us to harness local features that might otherwise be destroyed. We incorporate CLIP-DPT into our method, in the same way that we incorporated ZSCLIP in the previous" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "section, and see a consistent improvement in Tab. 3, suggesting that our method is adding a complementary strength and not just compensating for lost local information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.149, + 0.874, + 0.166 + ], + "angle": 0, + "content": "4.3. Our method has complementary strengths" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.173, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Several existing methods adapt CLIP to MLR tasks in an image-free ([4], [22], [21], [11], [23]) or training-free ([8], [10]) manner. These approaches lie outside our scope, as we focus on methods that are deployable 'out-of-the-box' on new datasets and VLMs, without requiring dataset-specific training or VLM surgery. Our black-box method allows easy 'plug-and-play' integration, using scores from existing methods in place of the vanilla ZSCLIP singleton scores." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.295, + 0.905, + 0.415 + ], + "angle": 0, + "content": "We combine our method with three existing methods that have public codebases: (1) TagCLIP [10], a training-free method that benchmarks on COCO and VOC with a ViT-B/16 backbone; (2) TaI-DPT [4], an image-free method that requires training, which benchmarks on COCO, VOC, and NUSWIDE with a RN50 backbone; (3) CoMC [11], an image-free method that requires training and has a publicly-available RN50 model for COCO for one train seed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.905, + 0.643 + ], + "angle": 0, + "content": "There are some nuances regarding image preprocessing and score postprocessing. TaI-DPT and CoMC use different image preprocessing settings than us - theirs are lower resolution and take a center crop. For a fair comparison, we use these same preprocessing settings when getting compound prompt scores to combine with TaI-DPT and CoMC. We also try running both our and their method with our preprocessing setting, which we denote as \"RN50*\". TagCLIP takes images at their native resolution, so for that comparison we do the same in order to be apples-to-apples. As for score postprocessing, TagCLIP uses softmax, which we have found to have a normalizing effect, so we omit our image-level normalization on TagCLIP's scores, and take the log of their scores to make them compatible with ours. We make no postprocessing changes for TaI-DPT or CoMC." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.644, + 0.905, + 0.75 + ], + "angle": 0, + "content": "We show the results of these plug-and-play combinations in Tab. 4 and Tab. 5. We find that method is able to improve on TagCLIP by on average \\(1.6\\%\\) and on TaI-DPT by on average \\(1.7\\%\\). We do not improve on CoMC, but the degradation is only \\(0.3\\%\\). These results suggest that our method's signal complements those obtained through training and architectural manipulation by other methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.761, + 0.625, + 0.775 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.726, + 0.8 + ], + "angle": 0, + "content": "4.4.1 Normalization Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We take a closer look at the Normalization module by seeing how it impacts the performance of both singleton prompts and our full pipeline. We see in Tab. 7 that the Normalization module provides significant gains in both cases (7.7% with singletons, 8.6% full pipeline), confirming that bias significantly impacts MLR when left unaddressed." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4318" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.912, + 0.245 + ], + "angle": 0, + "content": "
DatasetViT-L/14 336pxViT-L/14ViT-B/16ArchitecturesAvg (archs)
ViT-B/32RN50 x64RN50 x16RN50 x4RN101RN50
COCOZSCLIP59.158.155.550.958.658.155.752.753.055.7
Ours70.569.567.464.070.670.169.567.765.768.3
VOCZSCLIP81.379.979.977.583.181.680.680.379.780.4
Ours88.988.388.787.590.590.390.089.789.289.2
NUSWIDEZSCLIP41.441.040.938.739.838.538.835.738.539.3
Ours47.547.147.346.947.447.948.346.545.847.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Table 2. Results over three datasets and nine CLIP backbones. Our proposed method consistently outperforms the ZSCLIP baseline across all datasets and architectures exhibiting its effectiveness on multi-label recognition tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.285, + 0.506, + 0.428 + ], + "angle": 0, + "content": "
RN50 x64RN50 x16RN50 x4RN 101RN 50
COCOCLIP-DPT70.564.264.259.757.1
Ours-DPT77.873.972.769.868.4
VOCCLIP-DPT86.682.486.183.582.6
Ours-DPT91.189.888.588.589.9
NUSCLIP-DPT39.837.839.338.438.1
Ours-DPT43.545.542.742.644.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Table 3. Comparing SPARC with local-feature method CLIP-DPT [4]. Our method still improves upon local features, showing that its strength is complementary." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.475, + 0.441, + 0.552 + ], + "angle": 0, + "content": "
COCO\nViT-B/16VOC\nViT-B/16Avg
TagCLIP70.991.781.3
+Ours73.892.082.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.557, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Table 4. Shows complementarity with architectural approaches, improving TagCLIP by \\(1.6\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.585, + 0.401, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.772 + ], + "angle": 0, + "content": "Figure 4. Average mAP for different Rank Fusion strategies demonstrates superiority of adaptive fusion over fixed strategies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.781, + 0.292, + 0.795 + ], + "angle": 0, + "content": "4.4.2 Rank Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.806, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Our next ablation takes a closer look at the role of the Rank Fusion module. We consider some possible handcrafted strategies for combining compound prompt scores:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.851, + 0.471, + 0.868 + ], + "angle": 0, + "content": "- \" \\( k \\) max\": Use the \\( k \\)-th highest compound score, i.e. \\( r_{i,k}^{t} \\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.868, + 0.473, + 0.887 + ], + "angle": 0, + "content": "- \"mean \\(> = k\\)\": Use partial avg \\(\\frac{1}{m_i - K + 1}\\sum_{k = K}^{m_i - K + 1}r_{i,k}^t\\)" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.851, + 0.473, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 shows the average mAP of these, as well as \"maxVari" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.5 + ], + "angle": 0, + "content": "ance” (3) and singleton-only, across all datasets and backbones. We see that 1st-max does much worse than 2nd-max, which underperforms 3rd-max, and so on, reaching a plateau around 5th-max. This lends credence to the idea that a weakened max is a stronger signal than outright max. We also see that the mean of all compound prompts (i.e. “mean \\(>= 1\\)) does surprisingly well. This is actually consistent with our intuition on weakened maxes - the mean weakens the higher maxes by averaging them with the lower ones. We can improve further by excluding high ranks in the “mean \\(>= k\\)\" strategy. However, we can surpass all of these by adaptively weighting scores using the max-variance direction. The Supplementary repeats this analysis without the “merge” step (5), which turns out to be critical as well." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.71, + 0.541 + ], + "angle": 0, + "content": "4.4.3 Compound Prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Our next ablation takes a closer look at the importance of the composition of our set of compound prompts and the use of rough cooccurrence info. The results are in Tab. 6. We start by including all pairs of classes in our set of formulaic pairwise prompts (i.e. of the form \"A and B\") and not including any other kind of compound prompt. We find that this gives us a \\(1.3\\%\\) average boost over normalized singletons. Filtering these pairwise prompts by cooccurrence, as our proposed method does, yields a further \\(0.5\\%\\) boost, and adding triplet and descriptive prompts gives an additional \\(0.3\\%\\) boost. It makes sense that using the full set of class pairs would still have some benefit, without any cooccurrence filtering, if all the classes as a whole tend to cooccur positively more often than negatively, as that would mean that they are on average a helpful signal for predicting each other's presence. It might also be the case that related classes provide a useful context signal to CLIP. In any case, it is noteworthy that Rank Fusion can still help even when no cooccurrence information is required from the user." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our final ablation aims to figure out whether our performance boost actually comes from the semantics of the compound prompts themselves, and not just an increase in number of prompts. Each class has on average \\(\\leq 20\\) compound" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4319" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.882, + 0.165 + ], + "angle": 0, + "content": "
COCOVOCNUSWIDEAvgCOCOAvg
RN50RN50*RN50RN50*RN50RN50*RN50RN50*
TaI-DPT65.168.288.588.046.243.466.6CoMC68.871.370.0
+Ours68.270.190.390.246.644.668.3+Ours68.970.669.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.168, + 0.905, + 0.197 + ], + "angle": 0, + "content": "Table 5. Results from combining our method with TaI-DPT (left) and CoMC (right) showing compatibility with training-based methods. Our method improves TaI-DPT by \\(1.7\\%\\) while preserving the existing strong signal of CoMC, degrading it by only \\(0.3\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.204, + 0.895, + 0.326 + ], + "angle": 0, + "content": "
NormalizePair promptsTriplets + DescriptiveRank Fusion StrategyCOCOVOCNUSAvg
✓--65.987.745.166.2
✓all pairsours67.588.546.467.5
✓all pairsmean67.588.546.467.5
✓cooccurrence-filteredours68.189.047.068.0
✓cooccurrence-filteredmean67.988.546.867.7
✓cooccurrence-filtered✓ours68.389.247.268.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.328, + 0.905, + 0.357 + ], + "angle": 0, + "content": "Table 6. Ablations on the makeup of our compound prompts. We find that we can still enjoy some benefit from pairwise prompts without any cooccurrence filtering." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.36, + 0.468, + 0.442 + ], + "angle": 0, + "content": "
CompoundNormalizeCOCOVOCNUSAvg
55.780.439.358.5
✓65.987.745.166.2
✓58.480.540.059.7
✓✓68.389.247.268.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Table 7. Ablations on Normalization module. Quantifies impact of normalization both with and without compound prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.483, + 0.529 + ], + "angle": 0, + "content": "prompts, 4x less than the 80 singleton prompt templates already used by vanilla ZSCLIP, but there is still an increase in total number of prompts, so an ablation is warranted." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.483, + 0.666 + ], + "angle": 0, + "content": "We take inspiration from WaffleCLIP [5], which found that prompt ensembles made of random texts could perform just as well as descriptive ensembles generated by LLMs. For our ablation, we use WaffleCLIP-style prompts in place of compound prompts. Specifically, for eachclassname \\( c_i \\) we create 30 prompts of the form \"A photo of a \\( c_i \\), which is [RAND]\", where \"[RAND]\" is 10 random characters. We try both the \"maxVariance\" and \"mean \\( > = 1 \\)\" fusion strategies, as the latter is more typical for assembling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.773 + ], + "angle": 0, + "content": "We show the results of this ablation in the Supplementary. We see that randomized compound prompts do not offer any complementary signal to the singleton prompts. Thus, any gain from compound prompts must come from their semantics. It is unlikely that simple prompt diversity explains the gains either, given that vanilla ZSCLIP already uses a large and diverse template ensemble." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.786, + 0.43, + 0.802 + ], + "angle": 0, + "content": "4.5. A qualitative look at the weakened max" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We saw in Sec. 4.4.1 that the second-highest compound prompt score is a much better signal than the highest one. This counterintuitive finding can be explained through object cooccurrence. When CLIP is given a compound prompt like 'cat and sofa', it has OR-like behavior - giving high scores if either object is present. The maximum score often" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.905, + 0.452 + ], + "angle": 0, + "content": "comes from such prompts where only one object is present but strongly detected. In contrast, the second-highest scores tend to come from prompts where both objects have moderate confidence, providing a more reliable signal. This explains why using second-highest scores results in better separation between positive and negative cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.604 + ], + "angle": 0, + "content": "Fig. 3 shows an example of what happens at an image level. On the right is an image that contains a cat (curled up behind the two dogs on the bed); on the left is an image without a cat. The highest-scoring compound prompt is \"Cat and Dog\" for both images. This prompt gives a high score to the right image by detecting the dogs, but it also detects the dog in the left image for the same reason, and so it also misorders the images. But the second-highest-scoring prompts - \"Bed and Cat\" for right image, \"Cat and Potted plant\" for left - correctly order the images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.905, + 0.665 + ], + "angle": 0, + "content": "We show score histograms for one of the classes in the Supplementary to illustrate what happens in aggregate - the second-highest score lifts positive examples without introducing as many false-positives as first-highest score." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.64, + 0.694 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.9 + ], + "angle": 0, + "content": "We presented SPARC, a zero-shot multi-label recognition approach that improves VLM performance without training data or architectural changes. Our key insights - the suboptimal nature of maximum scores and presence of systematic biases - led to two complementary innovations: compound prompts with adaptive fusion, and systematic normalization. Beyond improving standalone VLM performance, SPARC enhances existing zero-shot and training-based methods while maintaining a black-box approach. The success of our method reveals fundamental properties of VLM scoring behavior, suggesting promising directions for improving zero-shot recognition through score analysis rather than architectural modification or finetuning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4320" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[2] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM International Conference on Image and Video Retrieval, New York, NY, USA, 2009. Association for Computing Machinery. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.272, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Reza Esfandiarpoor and Stephen H Bach. Follow-up differential descriptions: Language models resolve ambiguities for image classification. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[4] Guo et al. Texts as images in prompt tuning for multi-label image recognition. In CVPR, '23. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[5] Roth et al. Waffling around for performance: visual classification with random words and broad concepts ICCV, '23. 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[6] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[7] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[8] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[9] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[10] Yuqi Lin, Minghao Chen, Kaipeng Zhang, Hengjia Li, Mingming Li, Zheng Yang, Dongqin Lv, Binbin Lin, Haifeng Liu, and Deng Cai. Tagclip: A local-to-global framework to enhance open-vocabulary multi-label classification of clip without training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3513-3521, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[11] Yicheng Liu, Jie Wen, Chengliang Liu, Xiaozhao Fang, Zuoyong Li, Yong Xu, and Zheng Zhang. Language-driven cross-modal classifier for zero-shot multi-label image recognition. In *Forty-first International Conference on Machine Learning*, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.885 + ], + "angle": 0, + "content": "[12] Leilei Ma, Hongxing Xie, Lei Wang, Yanping Fu, Dengdi Sun, and Haifeng Zhao. Text-region matching for multi-label image recognition with missing labels. In Proceedings of the 32nd ACM International Conference on Multimedia, page 6133-6142. ACM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Sachit Menon and Carl Vondrick. Visual classification via" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.329, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[17] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In Proceedings of the International Conference on Pattern Recognition (ICPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[18] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems, 35:30569-30582, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[19] Hao Tan, Zichang Tan, Jun Li, Jun Wan, Zhen Lei, and Stan Z. Li. Sspa: Split-and-synthesize prompting with gated alignments for multi-label image recognition. ArXiv, abs/2407.20920, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[20] Ao Wang, Hui Chen, Zijia Lin, Zixuan Ding, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Guiguang Ding. Hierarchical prompt learning using clip for multi-label classification with single positive labels. In Proceedings of the 31st ACM International Conference on Multimedia, page 5594-5604, New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.608, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[21] Xiangyu Wu, Qing-Yuan Jiang, Yang Yang, Yi-Feng Wu, Qing-Guo Chen, and Jianfeng Lu. Tai++: Text as image for multi-label image classification by co-learning transferable prompt. arXiv preprint arXiv:2405.06926, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[22] Shuo Yang, Zirui Shang, Yongqi Wang, Derong Deng, Hongwei Chen, Qiyuan Cheng, and Xinxiao Wu. Data-free multi-label image recognition via llm-powered prompt tuning. arXiv preprint arXiv:2403.01209, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[23] Xueling Zhu, Jiuxin Cao, Jian Liu, Dongqi Tang, Furong Xu, Weijia Liu, Jiawei Ge, Bo Liu, Qingpei Guo, and Tianyi Zhang. Text as image: Learning transferable adapter for multi-label classification. ArXiv, abs/2312.04160, 2023. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. We thank Ping Hu, Siqi Wang, and Piotr Teterwak for their valuable discussions, as well as the anonymous reviewers for their thorough and helpful feedback. This research was supported by the Army Research Office Grant W911NF2110246, AFRL Grant FA8650-22-C1039, and the National Science Foundation grants CPS-2317079, CCF-2007350, and CCF-1955981." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4321" + } + ] +] \ No newline at end of file diff --git a/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_origin.pdf b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7cac7ec2da7a756236d29b3964451d8da8d6330 --- /dev/null +++ b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/4c63b13f-f035-4408-a7c4-a7f777a69d2d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc2f810a8ad50f87890658ec770bbcb089a1f051ad1026020fd4fd8e7a3621c +size 4491970 diff --git a/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/full.md b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f7c07baba3639d4b73230206e99df94f09ea8a56 --- /dev/null +++ b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/full.md @@ -0,0 +1,322 @@ +# SPARC: Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models + +Kevin Miller +Boston University +nivek@bu.edu + +Aditya Gangrade +Boston University +gangrade@bu.edu + +Samarth Mishra +Boston University +samarthm@bu.edu + +Kate Saenko +Boston University and Meta AI (FAIR) + +saenko@bu.edu + +Venkatesh Saligrama +Boston University +srv@bu.edu + +# Abstract + +Zero-shot multi-label recognition (MLR) with Vision-Language Models (VLMs) faces significant challenges without training data, model tuning, or architectural modifications. Existing approaches require prompt tuning or architectural adaptations, limiting zero-shot applicability. Our work proposes a novel solution treating VLMs as black boxes, leveraging scores without training data or ground truth. We make two contributions. First, we find that VLM scores suffer from image- and prompt-specific biases, and that simple standardization is surprisingly effective at removing these and boosting MLR performance. And second, we introduce compound prompts grounded in realistic object combinations. Our analysis reveals "AND"/"OR" signal ambiguities that cause maximum compound scores to be surprisingly suboptimal compared to second-highest scores. We introduce an adaptive fusion method to address this issue. Our method enhances other zero-shot approaches, consistently improving their results. Experiments show superior mean Average Precision (mAP) compared to methods requiring training data, achieved through refined object ranking for robust zero-shot MLR. Code can be found at https://github.com/kjmillerCURIS/SPARC. + +# 1. Introduction + +Vision-Language Models (VLMs), such as CLIP [16], have emerged as general-purpose systems for understanding visual data through language-based queries. These models enable a broad range of applications, from object detection to image captioning, by linking visual inputs to language prompts. In standard settings where images contain single, recognizable objects, VLMs perform remarkably well. However, for the more complex task of zero-shot multi-label recognition (MLR) (Fig. 1 (top)), where mod + +![](images/0db919ced7a520be949a3f532cc33f99a7a787d84a0b66bfa4af880c91182727.jpg) +Figure 1. (top) Vision-Language Models (VLMs) like CLIP can be used for zero-shot classification with image-text similarity scores. While this works fairly well for single-class labels, they can struggle in the multi-label scenario. (bottom) In this paper, we introduce SPARC, our solution that functions on top of an existing VLM, treating it simply as a black-box score generator. Using class names, SPARC first creates compound prompts for additional queries to the VLM. It then normalizes, ranks and appropriately fuses them to generate final scores for the original classes. + +els must identify multiple objects within an image without prior training on specific data, VLMs face significant limitations. Zero-shot MLR is crucial for applications in fields like robotics and medical imaging, where objects rarely appear in configurations that align neatly with training distributions. In these scenarios, achieving robust multi-label recognition without fine-tuning is challenging, given the task's reliance on mean Average Precision (mAP) scores, which depend on ranking images for object presence. + +VLM: Prompt Dependent AND/OR Noisy Channel. Despite the promise of zero-shot capabilities, current VLM approaches often struggle with MLR due to inherent scoring + +![](images/207ec5e7ad2b8641f53f9651fe0cae7f444435211af6f44f50e2016c6e13f702.jpg) +Figure 2. A motivating example for the Normalization module, with CLIP scores on a slice of the COCO dataset, before and after normalization. The "necktie" class is present for examples in the top half of the plot and absent in the bottom half. Image- and prompt-level biases show up as horizontal and vertical striations; normalization removes these and creates better separation. + +![](images/e9a3605d6e786a3389140855f80cb432d3ee53c74107b82f89e7c3cdc0944a30.jpg) +Figure 3. A motivating example for the Rank Fusion module, with an image where class "cat" is absent (left) and one where it is present (right). The highest compound prompt score is an unhelpful signal because it gives a high score to both negatives and positives, while the second-highest is more discriminative. Our method adaptively fuses the most informative order statistics, resulting in a strong signal. + +behaviors and biases. The performance of these models is hampered by a mix of conjunction (AND) and disjunction (OR) behaviors in their scoring, leading to inflated scores in compound prompts that contain multiple objects. For example, a prompt like "cat and sofa" might yield a high score even if only one of these objects is present in the image. This tendency reflects biases learned during training, where common object pairs receive higher scores even when only one object is present, disrupting the accuracy of mAP-based evaluations. Furthermore, existing methods for adapting VLMs to zero-shot MLR frequently rely on prompt tuning + +or architectural adjustments—approaches that are often dependent on training data and computationally intensive finetuning, which limit their generalizability to novel tasks. + +Our Approach. In contrast to these methods, we introduce SPARC (Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in VLMs), a novel approach to zero-shot MLR that bypasses the need for training data, prompt tuning, or model-specific modifications. Our method treats the VLM as a black box, relying solely on its output scores to infer object presence (see Fig. 1). This black-box approach enables us to avoid assumptions about the model's internal workings, allowing for a purely zero-shot framework that is both model-agnostic and dataset-independent. SPARC introduces two main innovations that address the unique challenges of zero-shot MLR. + +A. Score Normalization. Our analysis finds that VLM scores suffer from image-level and prompt-level biases. Image-level bias refers to noise terms that are constant across prompts but vary between images, while prompt-level bias refers to the converse. Image-level bias in particular hampers mAP as it changes the ordering of scores within each class. Standardization is surprisingly effective at addressing this problem in MLR. Standardization alone improves mAP by $6 - 10\%$ on COCO, VOC, and NUSWIDE. + +B. Compound Prompts and Adaptive Fusion. Recognizing that VLMs can provide richer information when prompted with combinations of objects, we develop a method for constructing compound prompts. These prompts reflect likely contextual associations between objects, such as "cat and sofa" or "car and bus." By gathering scores from these compound prompts, we can capture a spectrum of potential object contexts within the image, enhancing detection without relying on training-based adaptations. This composition strategy allows us to agnostically extract information from the VLM without depending on any specific dataset or VLM architecture. + +Surprisingly, we observe that the maximum score among compound prompts is often a poor proxy for true object presence. Although one might expect the highest score to serve as a reliable signal, we find that it suffers because VLMs tend to respond to compound prompts with OR-like behavior, raising scores even when only one object in the prompt is present. Instead, we observe that the second-highest score consistently outperforms the maximum score by minimizing the effects of false positives caused by OR-like behavior. Building on this insight, we develop a PCA-based fusion method that combines information from multiple order statistics from the compound prompts, as well as the singleton prompts, ultimately optimizing mAP by enhancing score accuracy. + +Complementarity. SPARC is complementary to other zero-shot and training-free MLR methods. When applied on top of these approaches, SPARC consistently enhances + +mAP scores by refining object ranking and reducing bias in VLM outputs. This capability makes SPARC an adaptable solution that can improve upon existing methods while maintaining a fully zero-shot, model-agnostic framework. + +Empirical Results. SPARC achieves significant improvements in mAP, outperforming methods that incorporate architectural modifications. This outcome shows the potential of a fully zero-shot approach that relies only on systematic prompt design and score interpretation, rather than prompttraining or fine-tuning. By revealing that the second-highest score can be a superior proxy to the maximum, our findings provide new insights into VLM scoring behavior, suggesting that careful treatment of prompt compositions and score patterns can unlock robust MLR capabilities. + +# 1.1. Related Work + +Supervised Methods. Prior methods improve MLR performance through various approaches. DualCoOp [7, 18] trains prompts for text-guided spatial attention. Subsequent works use class co-occurrences to refine CLIP logits [17, 20]. Hierarchical structuring [20] encourages related classes to learn similar prompts. SSPA [19] combines embeddings from learned and LLM-generated prompts, and TRM-ML [12] uses pseudolabels to match texts to image regions. While effective, these methods require annotated training data. In contrast, ours is unsupervised. + +Unsupervised Training-Based Methods. TaI-DPT [4] uses caption-only data to adapt CLIP to MLR by training a prompt-generation network using caption embeddings. Extensions include LLM-generated captions [22], pseudovisual prompts [21], and lightweight classifiers on caption embeddings [11, 23]. Separately, CDUL [1] leverages CLIP scores with spatial aggregation to create training pseudolabels. While these approaches demonstrate shared vision-language embedding power, they require significant training and CLIP internal access. In contrast, our black-box method uses crude cooccurrence-based compound prompts (essentially deleting implausible contextual associations) to probe multilabel information, fusing scores for effective MLR. + +Unsupervised Training-free methods. The recent literature has proposed methods to adapt CLIP to the MLR taks in a training-free manner via architectural changes. CLIP-Surgery [8] alters the architecture to increase the coherence of attention masks within the image backbone, and normalizes features on a patch-by-patch basis. TagCLIP [10] uses internal attention masks to refine scores at a patch level based on regional coherence, and aggregate queries over crops of the image. Again, both methods require access to VLM internals, while ours works in a black-box manner. + +Prompt Enhancement for Single-Label Recognition. Several methods use LLM-generated descriptive prompts for single-label recognition, combining them via max or mean fusion [3, 13-15]. While related, our compound + +prompts specifically target MLR by modeling class relationships rather than enriching individual class descriptions. + +Complementarity of SPARC to Unsupervised Methods. We explicitly note that our methodology is complementary to much of the work on unsupervised VLM-based MLR. For instance, since we assume only black-box access to VLM scores, our compound prompts could be refined via prompt tuning as in TaI-DPT [4] and/or applied patch-by-patch and integrated à la TagCLIP [10]. Scores from any of these methods can be combined via our normalization and rank-fusion approach. §4.3 investigates the resulting gains. + +# 2. Method + +We detail our method, SPARC (see Alg. 1 and Fig. 1). + +Algorithm 1 SPARC Pipeline +Input: Images $\mathcal{I}$ , Class Names $\mathcal{C}$ +Output: Final scores $\zeta_i^t$ for each image $t$ and class $i$ $P \leftarrow$ GenerateCompoundPrompts $(\mathcal{C})$ (Details in Supp.) +for $t \in \mathcal{I}$ do + $s_i^t \leftarrow$ GetVLMScores $(t, i)$ (Singleton scores) + $\forall p \in P, s_p^t \leftarrow$ GetVLMScores $(t, p)$ (Compound scs.) +// (Normalize scores (Eqs. 1, 2)) + $\tilde{s}_i^t \leftarrow$ ImageNorm $(s_i^t), \tilde{s}_p^t \leftarrow$ ImageNorm $(s_p^t)$ $\bar{s}_i^t \leftarrow$ PromptNorm $(\tilde{s}_i^t), \bar{s}_p^t \leftarrow$ PromptNorm $(\tilde{s}_p^t)$ +for $i \in \mathcal{C}$ do + $r_{i,k}^t \leftarrow$ GetOrderStatistics( $\{\bar{s}_p^t : i \in p\}$ ) + $w^{i*} \leftarrow$ MaxVarDirection([\bar{s}_i^t, r_{i,k}^t]) (Eq. 3) + $\tilde{\zeta}_i^t \leftarrow w_0^{i*} \bar{s}_i^t + \sum_k w_k^{i*} r_{i,k}^t$ (Eq. 4) + $\zeta_i^t \leftarrow \bar{s}_i^t + \tilde{\zeta}_i^t$ (Eq. 5) +return $\{\zeta_i^t\}$ + +Setup. Suppose we have $M$ images and $N$ target classes, with classnames $c_{1},\ldots ,c_{N}$ . We assume that we have a set of "singleton" prompts describing each of these classes in isolation (e.g. "picture of a cat"). Given an image $t$ , let $s_1^t,\dots,s_N^t$ denote the similarity scores produced by a VLM when comparing the singleton prompts to that image. + +A naive approach is to directly use these singleton scores to perform MLR. However, since the presence of a particular class alters the conditional probability of the remaining classes in images, it should be possible to further refine these scores by accounting for the multi-label structure of the images. In order to do this, we use the class-names to generate a further set of "compound" prompts $P$ which mention multiple classes. For instance, the classes "cat" and "sofa" may be compounded to "cat and sofa". Let $\{s_p^t : p \in P\}$ denote the scores given by the VLM when comparing these compound prompts to image $t$ . + +Using the singleton as well as the compound prompts above, SPARC produces a vector of refined class-wise scores $\zeta_1^t,\dots ,\zeta_N^t$ , for each image $t$ . Given this structure, SPARC has three main components: (i) compound prompt + +generation to enable the use of CLIP scoring for MLR (ii) image and prompt level normalization to allow scores to be directly compared, and (iii) Rank Fusion to combine singleton and compound prompt scores into a single score per class. We now discuss each of these aspects in detail. + +# 2.1. Compound Prompt Generation + +Our compound prompt generation method relies on contextual associations from commonly observed visual patterns to inform prompt structure. We provide pseudocode for the process in the Supplementary. The process takes as input the classnames $c_{1}, \ldots, c_{N}$ , ground-truth cooccurrence statistics $\mathbb{P}$ , thresholds $\tau_{2}$ and $\tau_{3}$ (fixed to 0.05 and 0.025 for all datasets), and off-the-shelf LLM $\Phi$ . We use probability thresholding to select pairs and triplets of classes that could plausibly coocur in realistic visual scenes. These pairs and triplets are used to make formulaic compound prompts of the form "A and B" and "A, B, and C". We feed these formulaic prompts to LLM $\Phi$ and ask it to generate natural sentences from them. Our total set of compound prompts comprises the formulaic pair, formulaic triplet, and natural sentence prompts, which add up to on average $\leq 20$ compound prompts per class in all of our datasets. Note that probabilities $\mathbb{P}$ are not used in any part of our method other than compound prompt generation, which is why we say that it uses "rough" cooccurrence info. In fact, our ablations reveal that we still get a performance gain when using all possible class pairs instead of filtering by probability. + +# 2.2. Normalization + +Once we have obtained compound prompts $P$ , we can query the VLM to obtain singleton scores $s_1^t, \ldots, s_N^t$ and compound scores $\{s_p^t : p \in P\}$ for each image $t$ . However, these scores contain both image- and prompt-specific biases. The former in particular is a problem, as it can change the order of scores between negative and positive images. + +We address image-level bias for singleton and compound prompts, respectively, as follows: + +$$ +\tilde {s} _ {i} ^ {t} := \frac {s _ {i} ^ {t} - \hat {\mu} (s _ {i} ^ {t})}{\hat {\sigma} (s _ {i} ^ {t})} \quad a n d \quad \tilde {s} _ {p} ^ {t} := \frac {s _ {p} ^ {t} - \hat {\mu} (\check {s} _ {i} ^ {t})}{\hat {\sigma} (\check {s} _ {i} ^ {t})}, \tag {1} +$$ + +where $\hat{\mu}(s_{\cdot}^{t}), \hat{\mu}(\check{s}_{\cdot}^{t}), \hat{\sigma}(s_{\cdot}^{t}), \hat{\sigma}(\check{s}_{\cdot}^{t})$ are sample means and standard deviations across the prompt dimension for a single image. $\check{s}_{1}^{t}, \ldots, \check{s}_{N}^{t}$ are the scores of "auxiliary" prompts that mention classnames in isolation and are used only to obtain statistics for standardizing compound prompt scores. + +We also standardize across images to quash prompt bias: + +$$ +\overline {{s}} _ {i} ^ {t} := \frac {\tilde {s} _ {i} ^ {t} - \hat {\mu} (\tilde {s} _ {i} ^ {*})}{\hat {\sigma} (\tilde {s} _ {i} ^ {*})} a n d \overline {{s}} _ {p} ^ {t} := \frac {\tilde {s} _ {p} ^ {t} - \hat {\mu} (\tilde {s} _ {p} ^ {*})}{\hat {\sigma} (\tilde {s} _ {p} ^ {*})}, (2) +$$ + +where $\hat{\mu} (\tilde{s}_i)$ , $\hat{\mu} (\tilde{s}_p)$ , $\hat{\sigma} (\tilde{s}_i)$ , and $\hat{\sigma} (\tilde{s}_p)$ now denote sample means and SDs across the images. Removing prompt-level + +bias makes scores from different prompts more compatible with each other, which is important for fusing them. + +# 2.3. Rank Fusion + +Our goal now is to find a way to use compound scores to strengthen the singleton scores. Let us define a bit more notation here. Let $r_{i,k}^{t}$ denote the score of the $k$ -th highest scoring compound prompt that mentions class $i$ . + +A natural choice might be to choose $r_{i,1}^{t}$ , i.e. the highest scoring prompt. After all, we would expect this to be the prompt that most closely describes the image, and if the image did not contain $c_{i}$ then we would expect none of the prompts containing $c_{i}$ to score high. However we find that $r_{i,1}^{t}$ does not offer a good detection of classes. The mechanism behind this, which is explored in detail in §3, is that when using a compound prompt of the form 'A and B' to detect the class $A$ , the score sees a large increase if the object $B$ truly occurs in the image. This 'OR'-like behavior (§3) means that maximum score $r_{i,1}^{t}$ typically captures the effect of other classes than $A$ being present in the image, and thus leads to very poor separation between ground truth positive and negative images for any class. Counterintuitively, then, we find that 'weakened maxes', i.e., lower order statistics like $r_{i,2}^{t}$ and $r_{i,3}^{t}$ yield much better separation. + +The above fact suggests that we ought to utilize the entirety of the order statistics $\{r_{i,k}^t\}$ in order to generate an effective fusion rule. We do fusion by taking a weighted sum where the weighting vector is the direction of highest variance, i.e., we compute the fused compound score $\tilde{\zeta}_i^t$ as + +$$ +w ^ {i *} := \underset {w ^ {i}} {\arg \max } \operatorname {V a r} _ {t} \left(w _ {0} ^ {i} \bar {s} _ {i} ^ {t} + \sum_ {k} w _ {k} ^ {i} r _ {i, k} ^ {t}\right) \tag {3} +$$ + +$$ +\tilde {\zeta} _ {i} ^ {t} := w _ {0} ^ {i *} \bar {s} _ {i} ^ {t} + \sum_ {k} w _ {k} ^ {i *} r _ {i, k} ^ {t}. \tag {4} +$$ + +Note that the variance above is computed across examples, i.e. an iteration over example index $t$ . Computing arg max above is just a matter of running PCA and taking the first principal component. The motivation for this approach is similar to that for linear discriminant analysis. Imagine each example is a point in $\mathbb{R}^{K_i}$ , where $K_{i}$ is the number of compound prompts for class $i$ , and these points are distributed according to a mixture of two components, one for examples where $i$ is present, the other for examples where $i$ is absent. The shift between the means of these components should contribute to the variance in the direction of that shift. If we assume that these components are roughly homoskedastic, then, as long as the 'signal' induced by the mean shift is large compared to any heteroskedasticity, the highest variance direction $w^{i*}$ is precisely the direction of the mean shift, and thus $\tilde{\zeta}_i^t$ effectively computes the component of the score vector that lies along this signal direction. While principled, this max-variance strategy is of course not + +necessary to execute, and we compare it to alternative fusion strategies in §4.4.1. We add (i.e. "merge") this fused score into the original singleton score to get our final score: + +$$ +\zeta_ {i} ^ {t} := s _ {i} ^ {t} + \tilde {\zeta} _ {i} ^ {t} \tag {5} +$$ + +In the next section we probe CLIP scores through a statistical lens to build intuition on our approach. + +# 3. Noise Model of a VLM + +Our methodological choices are strongly driven by the behavior of CLIP scores under compound prompts, and this leads to both the normalization approach taken, as well as our counterintuitive choice of fusing compound scores by weakened max instead of outright max. In order to justify these choices, we probe the structure of CLIP scores under compound prompts by constructing a statistical model. + +Given an image $t$ and a prompt with classes $i$ and $j$ , we model the score of compound prompt 'i and j' as $s_{i,j}^{t} \coloneqq \theta_{1}^{t} \cdot f(y_{i}^{t}, y_{j}^{t}) + \theta_{0}^{t} + \varepsilon$ where $y_{i}^{t}, y_{j}^{t} \in \{0,1\}$ indicate ground-truth presence of the two classes, $\theta_{1}^{t}$ and $\theta_{0}^{t}$ represent image-level bias (e.g. due to visual domain-shifts), and $\varepsilon$ is IID Gaussian noise. Note that our Normalization module is well-suited to mitigate variations in $\theta_{0}^{t}$ and $\theta_{1}^{t}$ , at least when $\varepsilon$ is not strongly dominating. + +Now, the core object of interest is the map $f: \{0,1\}^2 \to \mathbb{R}$ , which is responsible for modeling class pair interactions and prompt-level biases (see Table 1). In an ideal situation for MLR, $f$ would behave roughly as an 'AND' function, so that compound prompts would let us directly hone in on the presence of pairs of classes. However, prior work has observed that instead, for CLIP-type models, $f$ tends to predominantly behave as an 'OR' function, in that $f$ is high if either $i$ or $j$ is present. Notice that if $f$ were to exactly behave as an OR, compound prompts would not be effective in detecting classes that were missed by the singleton scores. + +In experimentation, we found that the behavior of $s_{i,j}^{t}$ is intermediary to the above extremes: CLIP scores for compound prompts behave qualitatively as an OR function with a small AND 'bonus' when both classes are present. This bonus, as well as the particular scores, are modulated by the classes mentioned (a consequence of prompt-level bias for each class) which we model as linear effects, leading to an overall model of the signal term of the CLIP score as + +$$ +\begin{array}{l} f \left(y _ {i} ^ {t}, y _ {j} ^ {t}\right) := \max \left(a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}\right) \\ + \delta_ {i j} \min (a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}), \qquad (6) \\ \end{array} +$$ + +where the $a$ terms are class-wise linear effect coefficients, and $\delta_{ij}$ captures the strength of this AND bonus. + +Note that this AND bonus is a critical feature of CLIP scores in their use for MLR problems, since this bump allows us to both infer the presence of classes that singleton + +
Noise Modelf(yi,t,yj,t)FVU
constant0.10.802
only ANDmin(a0i+yitai,aj+ytaj)0.536
only ORmax(a0i+yitai,aj+ytaj)0.288
additivea0i+yitai+aj+ytaj0.267
OR + static AND-bonusmax(a0i+yitai,aj+ytaj) + δ min(a0i+yitai,aj+ytaj)0.263
OR + variable AND-bonusmax(a0i+yitai,aj+ytaj) + δi,j min(a0i+yitai,aj+ytaj)0.248
look-up tableLUT(i,j)(yi,t,yj,t)0.235
+ +Table 1. Comparison of fidelity of noise models for scoring pairwise compound prompts. Notice that the OR+AND-bonus model captures nearly all of the fidelity of the look-up table, and offers significant improvements over the simplified models above it. + +prompts have missed, and indirectly to filter false positives (since the true positives are further elevated by the bonus). + +Table 1 shows the results of a systematic evaluation of the AND-OR structure of compound CLIP scores, using 81K COCO test images and ViT-L/14@336px backbone. We evaluate various models of $f$ , ranging from simple constant, 'only OR,' and 'only AND' models to additive models and OR+AND-bonus models with both a 'static' $\delta$ shared by all class pairs, and a 'variable' $\delta_{i,j}$ that is class-pairspecific. The final entry is a 'look-up table' (LUT) model where all four values of $f$ can vary arbitrarily, which serves as the most expressive baseline in this model class. These are evaluated w.r.t. the fraction of variance unexplained (i.e., $1 - R^2$ value). Observe that while the only OR model has remarkable fidelity, the additive and AND-bonus models significantly improve upon this, and allowing $\delta$ to vary with $i,j$ yield a further significant improvement to within a $6\%$ extra loss from the LUT model. This shows that our AND-bonus models capture a significant amount of the qualitative behavior of CLIP scores for compound prompts. + +A closer look at these fitted models reveals the $\delta$ values are small and relatively stable: for static $\delta$ , the recovered value is 0.56, and the IQR for varying $\delta_{ij}$ is (0.44, 0.57). + +We find similar behavior when examining other backbones and datasets, which we detail in the Supplementary. The Supplementary also contains a theoretical view of the benefits of using weakened-max, based on the noise model. + +# 3.1. Methodological Implications + +The qualitative structure of the OR+AND-bonus behavior justifies our methodological choices in §2. Image-level normalization reduces the effect of variation in $\theta_0^t$ (via mean subtraction) and $\theta_1^t$ (via rescaling), while prompt-level normalization mitigates variations in $a_0^i$ and $a_1^i$ across classes. + +We can also qualitatively observe that, because the OR component of the scores dominates the AND component, the maximum score $r_{i,1}^{t}$ can create false positives when $i$ is + +absent but related class $j$ is present, and so we should look to subsequent order statistics for a better signal. + +# 4. Experiments + +Datasets We benchmark our method on 3 different MLR datasets. They are COCO2014 [9], which has 40,504 test images with 80 classes; VOC2007 [6], which has 4.952 test images with 20 classes; and NUSWIDE [2], which has 107,859 test images with 81 classes. We use only test, and not train, images from these datasets. We do use their conditional ground-truth label distributions, computed from training set, to choose compound prompts, but do not use ground-truth probabilities in any other part of our method. + +CLIP Backbones In order to showcase the generality of our method, we apply it to 9 different CLIP backbones, which are ViT-L/14@336px, ViT-L/14, ViT-B/16, ViT-B/32, RN50x64, RN50x16, RN50x4, RN101, RN50. It is worth noting that from our model's perspective, each backbone could be seen as a dataset of its own, providing a different source of scores for each image set. We use the default 224x224 resolution, resizing without center crop, for all ViT models. For ResNet models, we use the same resolution as [7] except when otherwise stated. + +# 4.1. Our method improves over vanilla ZSCLIP across many datasets and models + +As a baseline, we compute singleton scores using prompt-ensembling with the 80 templates from the CLIP paper [16]. We then compute compound prompt scores and apply our method to the singletons and compounds. We compare our results to this vanilla ZSCLIP baseline. + +Tab. 2 shows our results over three different datasets and nine different CLIP backbones. We see that our method is able to improve over vanilla ZSCLIP by $12.6\%$ on COCO, $8.8\%$ on VOC, and $7.9\%$ on NUSWIDE. Moreover, we see that our method achieves consistent improvements across all datasets and backbones, showcasing its generality. For example, the improvements range from $11.4 - 15.0\%$ for COCO, $7.4 - 10\%$ for VOC, and $6.1 - 10.8\%$ for NUSWIDE. In fact, we see in the Supplementary that SPARC consistently improves the AP of almost every class in all datasets. + +# 4.2. Our method complements local features + +We consider another training-free zero-shot baseline introduced by the TaI-DPT paper [4] which we call "CLIP-DPT". This method modifies ZSCLIP by replacing the final global pooling in the image backbone with a parameter-free attention pooling where the query is the text embedding, allowing us to harness local features that might otherwise be destroyed. We incorporate CLIP-DPT into our method, in the same way that we incorporated ZSCLIP in the previous + +section, and see a consistent improvement in Tab. 3, suggesting that our method is adding a complementary strength and not just compensating for lost local information. + +# 4.3. Our method has complementary strengths + +Several existing methods adapt CLIP to MLR tasks in an image-free ([4], [22], [21], [11], [23]) or training-free ([8], [10]) manner. These approaches lie outside our scope, as we focus on methods that are deployable 'out-of-the-box' on new datasets and VLMs, without requiring dataset-specific training or VLM surgery. Our black-box method allows easy 'plug-and-play' integration, using scores from existing methods in place of the vanilla ZSCLIP singleton scores. + +We combine our method with three existing methods that have public codebases: (1) TagCLIP [10], a training-free method that benchmarks on COCO and VOC with a ViT-B/16 backbone; (2) TaI-DPT [4], an image-free method that requires training, which benchmarks on COCO, VOC, and NUSWIDE with a RN50 backbone; (3) CoMC [11], an image-free method that requires training and has a publicly-available RN50 model for COCO for one train seed. + +There are some nuances regarding image preprocessing and score postprocessing. TaI-DPT and CoMC use different image preprocessing settings than us - theirs are lower resolution and take a center crop. For a fair comparison, we use these same preprocessing settings when getting compound prompt scores to combine with TaI-DPT and CoMC. We also try running both our and their method with our preprocessing setting, which we denote as "RN50*". TagCLIP takes images at their native resolution, so for that comparison we do the same in order to be apples-to-apples. As for score postprocessing, TagCLIP uses softmax, which we have found to have a normalizing effect, so we omit our image-level normalization on TagCLIP's scores, and take the log of their scores to make them compatible with ours. We make no postprocessing changes for TaI-DPT or CoMC. + +We show the results of these plug-and-play combinations in Tab. 4 and Tab. 5. We find that method is able to improve on TagCLIP by on average $1.6\%$ and on TaI-DPT by on average $1.7\%$ . We do not improve on CoMC, but the degradation is only $0.3\%$ . These results suggest that our method's signal complements those obtained through training and architectural manipulation by other methods. + +# 4.4. Ablations + +# 4.4.1 Normalization Module + +We take a closer look at the Normalization module by seeing how it impacts the performance of both singleton prompts and our full pipeline. We see in Tab. 7 that the Normalization module provides significant gains in both cases (7.7% with singletons, 8.6% full pipeline), confirming that bias significantly impacts MLR when left unaddressed. + +
DatasetViT-L/14 336pxViT-L/14ViT-B/16ArchitecturesAvg (archs)
ViT-B/32RN50 x64RN50 x16RN50 x4RN101RN50
COCOZSCLIP59.158.155.550.958.658.155.752.753.055.7
Ours70.569.567.464.070.670.169.567.765.768.3
VOCZSCLIP81.379.979.977.583.181.680.680.379.780.4
Ours88.988.388.787.590.590.390.089.789.289.2
NUSWIDEZSCLIP41.441.040.938.739.838.538.835.738.539.3
Ours47.547.147.346.947.447.948.346.545.847.2
+ +Table 2. Results over three datasets and nine CLIP backbones. Our proposed method consistently outperforms the ZSCLIP baseline across all datasets and architectures exhibiting its effectiveness on multi-label recognition tasks. + +
RN50 x64RN50 x16RN50 x4RN 101RN 50
COCOCLIP-DPT70.564.264.259.757.1
Ours-DPT77.873.972.769.868.4
VOCCLIP-DPT86.682.486.183.582.6
Ours-DPT91.189.888.588.589.9
NUSCLIP-DPT39.837.839.338.438.1
Ours-DPT43.545.542.742.644.5
+ +Table 3. Comparing SPARC with local-feature method CLIP-DPT [4]. Our method still improves upon local features, showing that its strength is complementary. + +
COCO +ViT-B/16VOC +ViT-B/16Avg
TagCLIP70.991.781.3
+Ours73.892.082.9
+ +Table 4. Shows complementarity with architectural approaches, improving TagCLIP by $1.6\%$ + +![](images/8efa8dba3110c470bc03f30584068af4beb600573fb9925b77907b4a81449ada.jpg) +Figure 4. Average mAP for different Rank Fusion strategies demonstrates superiority of adaptive fusion over fixed strategies. + +# 4.4.2 Rank Fusion Module + +Our next ablation takes a closer look at the role of the Rank Fusion module. We consider some possible handcrafted strategies for combining compound prompt scores: + +- " $k$ max": Use the $k$ -th highest compound score, i.e. $r_{i,k}^{t}$ +- "mean $> = k$ ": Use partial avg $\frac{1}{m_i - K + 1}\sum_{k = K}^{m_i - K + 1}r_{i,k}^t$ + +Fig. 4 shows the average mAP of these, as well as "maxVari + +ance” (3) and singleton-only, across all datasets and backbones. We see that 1st-max does much worse than 2nd-max, which underperforms 3rd-max, and so on, reaching a plateau around 5th-max. This lends credence to the idea that a weakened max is a stronger signal than outright max. We also see that the mean of all compound prompts (i.e. “mean $>= 1$ ) does surprisingly well. This is actually consistent with our intuition on weakened maxes - the mean weakens the higher maxes by averaging them with the lower ones. We can improve further by excluding high ranks in the “mean $>= k$ " strategy. However, we can surpass all of these by adaptively weighting scores using the max-variance direction. The Supplementary repeats this analysis without the “merge” step (5), which turns out to be critical as well. + +# 4.4.3 Compound Prompts + +Our next ablation takes a closer look at the importance of the composition of our set of compound prompts and the use of rough cooccurrence info. The results are in Tab. 6. We start by including all pairs of classes in our set of formulaic pairwise prompts (i.e. of the form "A and B") and not including any other kind of compound prompt. We find that this gives us a $1.3\%$ average boost over normalized singletons. Filtering these pairwise prompts by cooccurrence, as our proposed method does, yields a further $0.5\%$ boost, and adding triplet and descriptive prompts gives an additional $0.3\%$ boost. It makes sense that using the full set of class pairs would still have some benefit, without any cooccurrence filtering, if all the classes as a whole tend to cooccur positively more often than negatively, as that would mean that they are on average a helpful signal for predicting each other's presence. It might also be the case that related classes provide a useful context signal to CLIP. In any case, it is noteworthy that Rank Fusion can still help even when no cooccurrence information is required from the user. + +Our final ablation aims to figure out whether our performance boost actually comes from the semantics of the compound prompts themselves, and not just an increase in number of prompts. Each class has on average $\leq 20$ compound + +
COCOVOCNUSWIDEAvgCOCOAvg
RN50RN50*RN50RN50*RN50RN50*RN50RN50*
TaI-DPT65.168.288.588.046.243.466.6CoMC68.871.370.0
+Ours68.270.190.390.246.644.668.3+Ours68.970.669.7
+ +Table 5. Results from combining our method with TaI-DPT (left) and CoMC (right) showing compatibility with training-based methods. Our method improves TaI-DPT by $1.7\%$ while preserving the existing strong signal of CoMC, degrading it by only $0.3\%$ . + +
NormalizePair promptsTriplets + DescriptiveRank Fusion StrategyCOCOVOCNUSAvg
✓--65.987.745.166.2
✓all pairsours67.588.546.467.5
✓all pairsmean67.588.546.467.5
✓cooccurrence-filteredours68.189.047.068.0
✓cooccurrence-filteredmean67.988.546.867.7
✓cooccurrence-filtered✓ours68.389.247.268.3
+ +Table 6. Ablations on the makeup of our compound prompts. We find that we can still enjoy some benefit from pairwise prompts without any cooccurrence filtering. + +
CompoundNormalizeCOCOVOCNUSAvg
55.780.439.358.5
✓65.987.745.166.2
✓58.480.540.059.7
✓✓68.389.247.268.3
+ +Table 7. Ablations on Normalization module. Quantifies impact of normalization both with and without compound prompts. + +prompts, 4x less than the 80 singleton prompt templates already used by vanilla ZSCLIP, but there is still an increase in total number of prompts, so an ablation is warranted. + +We take inspiration from WaffleCLIP [5], which found that prompt ensembles made of random texts could perform just as well as descriptive ensembles generated by LLMs. For our ablation, we use WaffleCLIP-style prompts in place of compound prompts. Specifically, for eachclassname $c_i$ we create 30 prompts of the form "A photo of a $c_i$ , which is [RAND]", where "[RAND]" is 10 random characters. We try both the "maxVariance" and "mean $> = 1$ " fusion strategies, as the latter is more typical for assembling. + +We show the results of this ablation in the Supplementary. We see that randomized compound prompts do not offer any complementary signal to the singleton prompts. Thus, any gain from compound prompts must come from their semantics. It is unlikely that simple prompt diversity explains the gains either, given that vanilla ZSCLIP already uses a large and diverse template ensemble. + +# 4.5. A qualitative look at the weakened max + +We saw in Sec. 4.4.1 that the second-highest compound prompt score is a much better signal than the highest one. This counterintuitive finding can be explained through object cooccurrence. When CLIP is given a compound prompt like 'cat and sofa', it has OR-like behavior - giving high scores if either object is present. The maximum score often + +comes from such prompts where only one object is present but strongly detected. In contrast, the second-highest scores tend to come from prompts where both objects have moderate confidence, providing a more reliable signal. This explains why using second-highest scores results in better separation between positive and negative cases. + +Fig. 3 shows an example of what happens at an image level. On the right is an image that contains a cat (curled up behind the two dogs on the bed); on the left is an image without a cat. The highest-scoring compound prompt is "Cat and Dog" for both images. This prompt gives a high score to the right image by detecting the dogs, but it also detects the dog in the left image for the same reason, and so it also misorders the images. But the second-highest-scoring prompts - "Bed and Cat" for right image, "Cat and Potted plant" for left - correctly order the images. + +We show score histograms for one of the classes in the Supplementary to illustrate what happens in aggregate - the second-highest score lifts positive examples without introducing as many false-positives as first-highest score. + +# 5. Conclusions + +We presented SPARC, a zero-shot multi-label recognition approach that improves VLM performance without training data or architectural changes. Our key insights - the suboptimal nature of maximum scores and presence of systematic biases - led to two complementary innovations: compound prompts with adaptive fusion, and systematic normalization. Beyond improving standalone VLM performance, SPARC enhances existing zero-shot and training-based methods while maintaining a black-box approach. The success of our method reveals fundamental properties of VLM scoring behavior, suggesting promising directions for improving zero-shot recognition through score analysis rather than architectural modification or finetuning. + +# References + +[1] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 3 +[2] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM International Conference on Image and Video Retrieval, New York, NY, USA, 2009. Association for Computing Machinery. 6 +[3] Reza Esfandiarpoor and Stephen H Bach. Follow-up differential descriptions: Language models resolve ambiguities for image classification. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 3 +[4] Guo et al. Texts as images in prompt tuning for multi-label image recognition. In CVPR, '23. 3, 6, 7 +[5] Roth et al. Waffling around for performance: visual classification with random words and broad concepts ICCV, '23. 8, 3 +[6] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6 +[7] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6 +[8] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 3, 6 +[9] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[10] Yuqi Lin, Minghao Chen, Kaipeng Zhang, Hengjia Li, Mingming Li, Zheng Yang, Dongqin Lv, Binbin Lin, Haifeng Liu, and Deng Cai. Tagclip: A local-to-global framework to enhance open-vocabulary multi-label classification of clip without training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3513-3521, 2024. 3, 6 +[11] Yicheng Liu, Jie Wen, Chengliang Liu, Xiaozhao Fang, Zuoyong Li, Yong Xu, and Zheng Zhang. Language-driven cross-modal classifier for zero-shot multi-label image recognition. In *Forty-first International Conference on Machine Learning*, 2024. 3, 6 +[12] Leilei Ma, Hongxing Xie, Lei Wang, Yanping Fu, Dengdi Sun, and Haifeng Zhao. Text-region matching for multi-label image recognition with missing labels. In Proceedings of the 32nd ACM International Conference on Multimedia, page 6133-6142. ACM, 2024. 3 +[13] Sachit Menon and Carl Vondrick. Visual classification via + +description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3 +[14] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023. +[15] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3 +[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6 +[17] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In Proceedings of the International Conference on Pattern Recognition (ICPR), 2024. 3 +[18] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems, 35:30569-30582, 2022. 3 +[19] Hao Tan, Zichang Tan, Jun Li, Jun Wan, Zhen Lei, and Stan Z. Li. Sspa: Split-and-synthesize prompting with gated alignments for multi-label image recognition. ArXiv, abs/2407.20920, 2024. 3 +[20] Ao Wang, Hui Chen, Zijia Lin, Zixuan Ding, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Guiguang Ding. Hierarchical prompt learning using clip for multi-label classification with single positive labels. In Proceedings of the 31st ACM International Conference on Multimedia, page 5594-5604, New York, NY, USA, 2023. Association for Computing Machinery. 3 +[21] Xiangyu Wu, Qing-Yuan Jiang, Yang Yang, Yi-Feng Wu, Qing-Guo Chen, and Jianfeng Lu. Tai++: Text as image for multi-label image classification by co-learning transferable prompt. arXiv preprint arXiv:2405.06926, 2024. 3, 6 +[22] Shuo Yang, Zirui Shang, Yongqi Wang, Derong Deng, Hongwei Chen, Qiyuan Cheng, and Xinxiao Wu. Data-free multi-label image recognition via llm-powered prompt tuning. arXiv preprint arXiv:2403.01209, 2024. 3, 6 +[23] Xueling Zhu, Jiuxin Cao, Jian Liu, Dongqi Tang, Furong Xu, Weijia Liu, Jiawei Ge, Bo Liu, Qingpei Guo, and Tianyi Zhang. Text as image: Learning transferable adapter for multi-label classification. ArXiv, abs/2312.04160, 2023. 3, 6 + +Acknowledgements. We thank Ping Hu, Siqi Wang, and Piotr Teterwak for their valuable discussions, as well as the anonymous reviewers for their thorough and helpful feedback. This research was supported by the Army Research Office Grant W911NF2110246, AFRL Grant FA8650-22-C1039, and the National Science Foundation grants CPS-2317079, CCF-2007350, and CCF-1955981. \ No newline at end of file diff --git a/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/images.zip b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ccfe77ff040f89c18ac1393a233d4e2fc1227ce --- /dev/null +++ b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5887da8de1431f2a3188fbeb722646de010e7343ee3d04f015b7e2b0ea55252e +size 412381 diff --git a/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/layout.json b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6ce8e9b5a1451f1c46517d2ab98a1f563332358c --- /dev/null +++ b/2025/SPARC_ Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models/layout.json @@ -0,0 +1,8629 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 534, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 534, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 534, + 140 + ], + "type": "text", + "content": "SPARC: Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 161, + 227, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 161, + 227, + 202 + ], + "spans": [ + { + "bbox": [ + 136, + 161, + 227, + 202 + ], + "type": "text", + "content": "Kevin Miller \nBoston University \nnivek@bu.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 260, + 161, + 350, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 161, + 350, + 203 + ], + "spans": [ + { + "bbox": [ + 260, + 161, + 350, + 203 + ], + "type": "text", + "content": "Aditya Gangrade \nBoston University \ngangrade@bu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 384, + 161, + 473, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 161, + 473, + 202 + ], + "spans": [ + { + "bbox": [ + 384, + 161, + 473, + 202 + ], + "type": "text", + "content": "Samarth Mishra \nBoston University \nsamarthm@bu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 209, + 332, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 209, + 332, + 237 + ], + "spans": [ + { + "bbox": [ + 140, + 209, + 332, + 237 + ], + "type": "text", + "content": "Kate Saenko \nBoston University and Meta AI (FAIR)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 240, + 272, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 240, + 272, + 250 + ], + "spans": [ + { + "bbox": [ + 200, + 240, + 272, + 250 + ], + "type": "text", + "content": "saenko@bu.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 366, + 209, + 470, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 209, + 470, + 250 + ], + "spans": [ + { + "bbox": [ + 366, + 209, + 470, + 250 + ], + "type": "text", + "content": "Venkatesh Saligrama \nBoston University \nsrv@bu.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 278, + 200, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 278, + 200, + 291 + ], + "spans": [ + { + "bbox": [ + 152, + 278, + 200, + 291 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 305, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 568 + ], + "type": "text", + "content": "Zero-shot multi-label recognition (MLR) with Vision-Language Models (VLMs) faces significant challenges without training data, model tuning, or architectural modifications. Existing approaches require prompt tuning or architectural adaptations, limiting zero-shot applicability. Our work proposes a novel solution treating VLMs as black boxes, leveraging scores without training data or ground truth. We make two contributions. First, we find that VLM scores suffer from image- and prompt-specific biases, and that simple standardization is surprisingly effective at removing these and boosting MLR performance. And second, we introduce compound prompts grounded in realistic object combinations. Our analysis reveals \"AND\"/\"OR\" signal ambiguities that cause maximum compound scores to be surprisingly suboptimal compared to second-highest scores. We introduce an adaptive fusion method to address this issue. Our method enhances other zero-shot approaches, consistently improving their results. Experiments show superior mean Average Precision (mAP) compared to methods requiring training data, achieved through refined object ranking for robust zero-shot MLR. Code can be found at https://github.com/kjmillerCURIS/SPARC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Vision-Language Models (VLMs), such as CLIP [16], have emerged as general-purpose systems for understanding visual data through language-based queries. These models enable a broad range of applications, from object detection to image captioning, by linking visual inputs to language prompts. In standard settings where images contain single, recognizable objects, VLMs perform remarkably well. However, for the more complex task of zero-shot multi-label recognition (MLR) (Fig. 1 (top)), where mod" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 277, + 548, + 437 + ], + "blocks": [ + { + "bbox": [ + 322, + 277, + 548, + 437 + ], + "lines": [ + { + "bbox": [ + 322, + 277, + 548, + 437 + ], + "spans": [ + { + "bbox": [ + 322, + 277, + 548, + 437 + ], + "type": "image", + "image_path": "0db919ced7a520be949a3f532cc33f99a7a787d84a0b66bfa4af880c91182727.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 453, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 552 + ], + "type": "text", + "content": "Figure 1. (top) Vision-Language Models (VLMs) like CLIP can be used for zero-shot classification with image-text similarity scores. While this works fairly well for single-class labels, they can struggle in the multi-label scenario. (bottom) In this paper, we introduce SPARC, our solution that functions on top of an existing VLM, treating it simply as a black-box score generator. Using class names, SPARC first creates compound prompts for additional queries to the VLM. It then normalizes, ranks and appropriately fuses them to generate final scores for the original classes." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 675 + ], + "type": "text", + "content": "els must identify multiple objects within an image without prior training on specific data, VLMs face significant limitations. Zero-shot MLR is crucial for applications in fields like robotics and medical imaging, where objects rarely appear in configurations that align neatly with training distributions. In these scenarios, achieving robust multi-label recognition without fine-tuning is challenging, given the task's reliance on mean Average Precision (mAP) scores, which depend on ranking images for object presence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "VLM: Prompt Dependent AND/OR Noisy Channel. Despite the promise of zero-shot capabilities, current VLM approaches often struggle with MLR due to inherent scoring" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4313" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 242 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 242 + ], + "type": "image", + "image_path": "207ec5e7ad2b8641f53f9651fe0cae7f444435211af6f44f50e2016c6e13f702.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 296, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 314 + ], + "type": "text", + "content": "Figure 2. A motivating example for the Normalization module, with CLIP scores on a slice of the COCO dataset, before and after normalization. The \"necktie\" class is present for examples in the top half of the plot and absent in the bottom half. Image- and prompt-level biases show up as horizontal and vertical striations; normalization removes these and creates better separation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 323, + 293, + 479 + ], + "blocks": [ + { + "bbox": [ + 59, + 323, + 293, + 479 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 293, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 293, + 479 + ], + "type": "image", + "image_path": "e9a3605d6e786a3389140855f80cb432d3ee53c74107b82f89e7c3cdc0944a30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 490, + 296, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 568 + ], + "type": "text", + "content": "Figure 3. A motivating example for the Rank Fusion module, with an image where class \"cat\" is absent (left) and one where it is present (right). The highest compound prompt score is an unhelpful signal because it gives a high score to both negatives and positives, while the second-highest is more discriminative. Our method adaptively fuses the most informative order statistics, resulting in a strong signal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "behaviors and biases. The performance of these models is hampered by a mix of conjunction (AND) and disjunction (OR) behaviors in their scoring, leading to inflated scores in compound prompts that contain multiple objects. For example, a prompt like \"cat and sofa\" might yield a high score even if only one of these objects is present in the image. This tendency reflects biases learned during training, where common object pairs receive higher scores even when only one object is present, disrupting the accuracy of mAP-based evaluations. Furthermore, existing methods for adapting VLMs to zero-shot MLR frequently rely on prompt tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "or architectural adjustments—approaches that are often dependent on training data and computationally intensive finetuning, which limit their generalizability to novel tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 112, + 555, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 112, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 555, + 255 + ], + "type": "text", + "content": "Our Approach. In contrast to these methods, we introduce SPARC (Score Prompting and Adaptive Fusion for Zero-Shot Multi-Label Recognition in VLMs), a novel approach to zero-shot MLR that bypasses the need for training data, prompt tuning, or model-specific modifications. Our method treats the VLM as a black box, relying solely on its output scores to infer object presence (see Fig. 1). This black-box approach enables us to avoid assumptions about the model's internal workings, allowing for a purely zero-shot framework that is both model-agnostic and dataset-independent. SPARC introduces two main innovations that address the unique challenges of zero-shot MLR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 556, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 556, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 556, + 366 + ], + "type": "text", + "content": "A. Score Normalization. Our analysis finds that VLM scores suffer from image-level and prompt-level biases. Image-level bias refers to noise terms that are constant across prompts but vary between images, while prompt-level bias refers to the converse. Image-level bias in particular hampers mAP as it changes the ordering of scores within each class. Standardization is surprisingly effective at addressing this problem in MLR. Standardization alone improves mAP by " + }, + { + "bbox": [ + 313, + 258, + 556, + 366 + ], + "type": "inline_equation", + "content": "6 - 10\\%" + }, + { + "bbox": [ + 313, + 258, + 556, + 366 + ], + "type": "text", + "content": " on COCO, VOC, and NUSWIDE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 367, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 510 + ], + "type": "text", + "content": "B. Compound Prompts and Adaptive Fusion. Recognizing that VLMs can provide richer information when prompted with combinations of objects, we develop a method for constructing compound prompts. These prompts reflect likely contextual associations between objects, such as \"cat and sofa\" or \"car and bus.\" By gathering scores from these compound prompts, we can capture a spectrum of potential object contexts within the image, enhancing detection without relying on training-based adaptations. This composition strategy allows us to agnostically extract information from the VLM without depending on any specific dataset or VLM architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 511, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 677 + ], + "type": "text", + "content": "Surprisingly, we observe that the maximum score among compound prompts is often a poor proxy for true object presence. Although one might expect the highest score to serve as a reliable signal, we find that it suffers because VLMs tend to respond to compound prompts with OR-like behavior, raising scores even when only one object in the prompt is present. Instead, we observe that the second-highest score consistently outperforms the maximum score by minimizing the effects of false positives caused by OR-like behavior. Building on this insight, we develop a PCA-based fusion method that combines information from multiple order statistics from the compound prompts, as well as the singleton prompts, ultimately optimizing mAP by enhancing score accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Complementarity. SPARC is complementary to other zero-shot and training-free MLR methods. When applied on top of these approaches, SPARC consistently enhances" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4314" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "mAP scores by refining object ranking and reducing bias in VLM outputs. This capability makes SPARC an adaptable solution that can improve upon existing methods while maintaining a fully zero-shot, model-agnostic framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "type": "text", + "content": "Empirical Results. SPARC achieves significant improvements in mAP, outperforming methods that incorporate architectural modifications. This outcome shows the potential of a fully zero-shot approach that relies only on systematic prompt design and score interpretation, rather than prompttraining or fine-tuning. By revealing that the second-highest score can be a superior proxy to the maximum, our findings provide new insights into VLM scoring behavior, suggesting that careful treatment of prompt compositions and score patterns can unlock robust MLR capabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 250, + 144, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 144, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 144, + 261 + ], + "type": "text", + "content": "1.1. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 269, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 388 + ], + "type": "text", + "content": "Supervised Methods. Prior methods improve MLR performance through various approaches. DualCoOp [7, 18] trains prompts for text-guided spatial attention. Subsequent works use class co-occurrences to refine CLIP logits [17, 20]. Hierarchical structuring [20] encourages related classes to learn similar prompts. SSPA [19] combines embeddings from learned and LLM-generated prompts, and TRM-ML [12] uses pseudolabels to match texts to image regions. While effective, these methods require annotated training data. In contrast, ours is unsupervised." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 544 + ], + "type": "text", + "content": "Unsupervised Training-Based Methods. TaI-DPT [4] uses caption-only data to adapt CLIP to MLR by training a prompt-generation network using caption embeddings. Extensions include LLM-generated captions [22], pseudovisual prompts [21], and lightweight classifiers on caption embeddings [11, 23]. Separately, CDUL [1] leverages CLIP scores with spatial aggregation to create training pseudolabels. While these approaches demonstrate shared vision-language embedding power, they require significant training and CLIP internal access. In contrast, our black-box method uses crude cooccurrence-based compound prompts (essentially deleting implausible contextual associations) to probe multilabel information, fusing scores for effective MLR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": "Unsupervised Training-free methods. The recent literature has proposed methods to adapt CLIP to the MLR taks in a training-free manner via architectural changes. CLIP-Surgery [8] alters the architecture to increase the coherence of attention masks within the image backbone, and normalizes features on a patch-by-patch basis. TagCLIP [10] uses internal attention masks to refine scores at a patch level based on regional coherence, and aggregate queries over crops of the image. Again, both methods require access to VLM internals, while ours works in a black-box manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Prompt Enhancement for Single-Label Recognition. Several methods use LLM-generated descriptive prompts for single-label recognition, combining them via max or mean fusion [3, 13-15]. While related, our compound" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "prompts specifically target MLR by modeling class relationships rather than enriching individual class descriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 205 + ], + "type": "text", + "content": "Complementarity of SPARC to Unsupervised Methods. We explicitly note that our methodology is complementary to much of the work on unsupervised VLM-based MLR. For instance, since we assume only black-box access to VLM scores, our compound prompts could be refined via prompt tuning as in TaI-DPT [4] and/or applied patch-by-patch and integrated à la TagCLIP [10]. Scores from any of these methods can be combined via our normalization and rank-fusion approach. §4.3 investigates the resulting gains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 214, + 370, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 214, + 370, + 227 + ], + "spans": [ + { + "bbox": [ + 314, + 214, + 370, + 227 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 235, + 533, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 235, + 533, + 247 + ], + "spans": [ + { + "bbox": [ + 314, + 235, + 533, + 247 + ], + "type": "text", + "content": "We detail our method, SPARC (see Alg. 1 and Fig. 1)." + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 315, + 250, + 553, + 456 + ], + "blocks": [ + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "lines": [ + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": "Algorithm 1 SPARC Pipeline \nInput: Images " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": ", Class Names " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " \nOutput: Final scores " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\zeta_i^t" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " and class " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "P \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " GenerateCompoundPrompts " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(\\mathcal{C})" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " (Details in Supp.) \nfor " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{I}" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "s_i^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " GetVLMScores " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(t, i)" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " (Singleton scores) \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\forall p \\in P, s_p^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " GetVLMScores " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(t, p)" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " (Compound scs.) \n// (Normalize scores (Eqs. 1, 2)) \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\tilde{s}_i^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " ImageNorm " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(s_i^t), \\tilde{s}_p^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " ImageNorm " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(s_p^t)" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\bar{s}_i^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " PromptNorm " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(\\tilde{s}_i^t), \\bar{s}_p^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " PromptNorm " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "(\\tilde{s}_p^t)" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{C}" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " do \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "r_{i,k}^t \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " GetOrderStatistics(" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\{\\bar{s}_p^t : i \\in p\\}" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": ") \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "w^{i*} \\leftarrow" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " MaxVarDirection([\\bar{s}_i^t, r_{i,k}^t]) (Eq. 3) \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\tilde{\\zeta}_i^t \\leftarrow w_0^{i*} \\bar{s}_i^t + \\sum_k w_k^{i*} r_{i,k}^t" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " (Eq. 4) \n" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\zeta_i^t \\leftarrow \\bar{s}_i^t + \\tilde{\\zeta}_i^t" + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "text", + "content": " (Eq. 5) \nreturn " + }, + { + "bbox": [ + 315, + 250, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\{\\zeta_i^t\\}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": "Setup. Suppose we have " + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": " images and " + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": " target classes, with classnames " + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "inline_equation", + "content": "c_{1},\\ldots ,c_{N}" + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": ". We assume that we have a set of \"singleton\" prompts describing each of these classes in isolation (e.g. \"picture of a cat\"). Given an image " + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "inline_equation", + "content": "s_1^t,\\dots,s_N^t" + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": " denote the similarity scores produced by a VLM when comparing the singleton prompts to that image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "content": "A naive approach is to directly use these singleton scores to perform MLR. However, since the presence of a particular class alters the conditional probability of the remaining classes in images, it should be possible to further refine these scores by accounting for the multi-label structure of the images. In order to do this, we use the class-names to generate a further set of \"compound\" prompts " + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "content": " which mention multiple classes. For instance, the classes \"cat\" and \"sofa\" may be compounded to \"cat and sofa\". Let " + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\{s_p^t : p \\in P\\}" + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "content": " denote the scores given by the VLM when comparing these compound prompts to image " + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 534, + 555, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Using the singleton as well as the compound prompts above, SPARC produces a vector of refined class-wise scores " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\zeta_1^t,\\dots ,\\zeta_N^t" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " , for each image " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " . Given this structure, SPARC has three main components: (i) compound prompt" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4315" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "generation to enable the use of CLIP scoring for MLR (ii) image and prompt level normalization to allow scores to be directly compared, and (iii) Rank Fusion to combine singleton and compound prompt scores into a single score per class. We now discuss each of these aspects in detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 138, + 224, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 224, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 224, + 152 + ], + "type": "text", + "content": "2.1. Compound Prompt Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": "Our compound prompt generation method relies on contextual associations from commonly observed visual patterns to inform prompt structure. We provide pseudocode for the process in the Supplementary. The process takes as input the classnames " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "c_{1}, \\ldots, c_{N}" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": ", ground-truth cooccurrence statistics " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": ", thresholds " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\tau_{3}" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": " (fixed to 0.05 and 0.025 for all datasets), and off-the-shelf LLM " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": ". We use probability thresholding to select pairs and triplets of classes that could plausibly coocur in realistic visual scenes. These pairs and triplets are used to make formulaic compound prompts of the form \"A and B\" and \"A, B, and C\". We feed these formulaic prompts to LLM " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": " and ask it to generate natural sentences from them. Our total set of compound prompts comprises the formulaic pair, formulaic triplet, and natural sentence prompts, which add up to on average " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\leq 20" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": " compound prompts per class in all of our datasets. Note that probabilities " + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 157, + 296, + 408 + ], + "type": "text", + "content": " are not used in any part of our method other than compound prompt generation, which is why we say that it uses \"rough\" cooccurrence info. In fact, our ablations reveal that we still get a performance gain when using all possible class pairs instead of filtering by probability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 146, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 146, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 146, + 426 + ], + "type": "text", + "content": "2.2. Normalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "content": "Once we have obtained compound prompts " + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "content": ", we can query the VLM to obtain singleton scores " + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "inline_equation", + "content": "s_1^t, \\ldots, s_N^t" + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "content": " and compound scores " + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\{s_p^t : p \\in P\\}" + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 433, + 295, + 504 + ], + "type": "text", + "content": ". However, these scores contain both image- and prompt-specific biases. The former in particular is a problem, as it can change the order of scores between negative and positive images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "content": "We address image-level bias for singleton and compound prompts, respectively, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 533, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 533, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 84, + 533, + 295, + 562 + ], + "type": "interline_equation", + "content": "\\tilde {s} _ {i} ^ {t} := \\frac {s _ {i} ^ {t} - \\hat {\\mu} (s _ {i} ^ {t})}{\\hat {\\sigma} (s _ {i} ^ {t})} \\quad a n d \\quad \\tilde {s} _ {p} ^ {t} := \\frac {s _ {p} ^ {t} - \\hat {\\mu} (\\check {s} _ {i} ^ {t})}{\\hat {\\sigma} (\\check {s} _ {i} ^ {t})}, \\tag {1}", + "image_path": "5294d8c6842a3867e2d03ffeded135eaf8a7439e9fddc71227f7cda08348ee53.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}(s_{\\cdot}^{t}), \\hat{\\mu}(\\check{s}_{\\cdot}^{t}), \\hat{\\sigma}(s_{\\cdot}^{t}), \\hat{\\sigma}(\\check{s}_{\\cdot}^{t})" + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": " are sample means and standard deviations across the prompt dimension for a single image. " + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\check{s}_{1}^{t}, \\ldots, \\check{s}_{N}^{t}" + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": " are the scores of \"auxiliary\" prompts that mention classnames in isolation and are used only to obtain statistics for standardizing compound prompt scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 628, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 295, + 639 + ], + "type": "text", + "content": "We also standardize across images to quash prompt bias:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 654, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 654, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 83, + 654, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\overline {{s}} _ {i} ^ {t} := \\frac {\\tilde {s} _ {i} ^ {t} - \\hat {\\mu} (\\tilde {s} _ {i} ^ {*})}{\\hat {\\sigma} (\\tilde {s} _ {i} ^ {*})} a n d \\overline {{s}} _ {p} ^ {t} := \\frac {\\tilde {s} _ {p} ^ {t} - \\hat {\\mu} (\\tilde {s} _ {p} ^ {*})}{\\hat {\\sigma} (\\tilde {s} _ {p} ^ {*})}, (2)", + "image_path": "2bf869330e6728d3a0cd877bc24c802eef0472733ff61480371bd4cb33c24de6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mu} (\\tilde{s}_i)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mu} (\\tilde{s}_p)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma} (\\tilde{s}_i)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma} (\\tilde{s}_p)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " now denote sample means and SDs across the images. Removing prompt-level" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "bias makes scores from different prompts more compatible with each other, which is important for fusing them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 102, + 397, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 102, + 397, + 114 + ], + "spans": [ + { + "bbox": [ + 314, + 102, + 397, + 114 + ], + "type": "text", + "content": "2.3. Rank Fusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": "Our goal now is to find a way to use compound scores to strengthen the singleton scores. Let us define a bit more notation here. Let " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "r_{i,k}^{t}" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": " denote the score of the " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": "-th highest scoring compound prompt that mentions class " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": "A natural choice might be to choose " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "r_{i,1}^{t}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": ", i.e. the highest scoring prompt. After all, we would expect this to be the prompt that most closely describes the image, and if the image did not contain " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " then we would expect none of the prompts containing " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " to score high. However we find that " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "r_{i,1}^{t}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " does not offer a good detection of classes. The mechanism behind this, which is explored in detail in §3, is that when using a compound prompt of the form 'A and B' to detect the class " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": ", the score sees a large increase if the object " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " truly occurs in the image. This 'OR'-like behavior (§3) means that maximum score " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "r_{i,1}^{t}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " typically captures the effect of other classes than " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " being present in the image, and thus leads to very poor separation between ground truth positive and negative images for any class. Counterintuitively, then, we find that 'weakened maxes', i.e., lower order statistics like " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "r_{i,2}^{t}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "inline_equation", + "content": "r_{i,3}^{t}" + }, + { + "bbox": [ + 313, + 170, + 554, + 361 + ], + "type": "text", + "content": " yield much better separation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": "The above fact suggests that we ought to utilize the entirety of the order statistics " + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\{r_{i,k}^t\\}" + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": " in order to generate an effective fusion rule. We do fusion by taking a weighted sum where the weighting vector is the direction of highest variance, i.e., we compute the fused compound score " + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\tilde{\\zeta}_i^t" + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 437, + 555, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 437, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 349, + 437, + 555, + 462 + ], + "type": "interline_equation", + "content": "w ^ {i *} := \\underset {w ^ {i}} {\\arg \\max } \\operatorname {V a r} _ {t} \\left(w _ {0} ^ {i} \\bar {s} _ {i} ^ {t} + \\sum_ {k} w _ {k} ^ {i} r _ {i, k} ^ {t}\\right) \\tag {3}", + "image_path": "340aea1d1ab13acafb8689435b436e31b97c5d3202c5f5080784b8d7b59d077b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 465, + 555, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 465, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 356, + 465, + 555, + 491 + ], + "type": "interline_equation", + "content": "\\tilde {\\zeta} _ {i} ^ {t} := w _ {0} ^ {i *} \\bar {s} _ {i} ^ {t} + \\sum_ {k} w _ {k} ^ {i *} r _ {i, k} ^ {t}. \\tag {4}", + "image_path": "5db08e59f7a63455d90b1846b13d3d0234b8a5ed8b4c0fb9512242c14d63d3e8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": "Note that the variance above is computed across examples, i.e. an iteration over example index " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": ". Computing arg max above is just a matter of running PCA and taking the first principal component. The motivation for this approach is similar to that for linear discriminant analysis. Imagine each example is a point in " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{K_i}" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": " is the number of compound prompts for class " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": ", and these points are distributed according to a mixture of two components, one for examples where " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": " is present, the other for examples where " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": " is absent. The shift between the means of these components should contribute to the variance in the direction of that shift. If we assume that these components are roughly homoskedastic, then, as long as the 'signal' induced by the mean shift is large compared to any heteroskedasticity, the highest variance direction " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "w^{i*}" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": " is precisely the direction of the mean shift, and thus " + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\zeta}_i^t" + }, + { + "bbox": [ + 313, + 497, + 555, + 713 + ], + "type": "text", + "content": " effectively computes the component of the score vector that lies along this signal direction. While principled, this max-variance strategy is of course not" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4316" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "necessary to execute, and we compare it to alternative fusion strategies in §4.4.1. We add (i.e. \"merge\") this fused score into the original singleton score to get our final score:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 119, + 295, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 119, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 146, + 119, + 295, + 135 + ], + "type": "interline_equation", + "content": "\\zeta_ {i} ^ {t} := s _ {i} ^ {t} + \\tilde {\\zeta} _ {i} ^ {t} \\tag {5}", + "image_path": "3558e37c862b9045640c22d1eac1e66d3b089b03651621229f22ec2c8f33d1fa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 295, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 164 + ], + "type": "text", + "content": "In the next section we probe CLIP scores through a statistical lens to build intuition on our approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 188, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 188, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 188, + 188 + ], + "type": "text", + "content": "3. Noise Model of a VLM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 195, + 296, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 279 + ], + "type": "text", + "content": "Our methodological choices are strongly driven by the behavior of CLIP scores under compound prompts, and this leads to both the normalization approach taken, as well as our counterintuitive choice of fusing compound scores by weakened max instead of outright max. In order to justify these choices, we probe the structure of CLIP scores under compound prompts by constructing a statistical model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " and a prompt with classes " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": ", we model the score of compound prompt 'i and j' as " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "s_{i,j}^{t} \\coloneqq \\theta_{1}^{t} \\cdot f(y_{i}^{t}, y_{j}^{t}) + \\theta_{0}^{t} + \\varepsilon" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "y_{i}^{t}, y_{j}^{t} \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " indicate ground-truth presence of the two classes, " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\theta_{1}^{t}" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\theta_{0}^{t}" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " represent image-level bias (e.g. due to visual domain-shifts), and " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " is IID Gaussian noise. Note that our Normalization module is well-suited to mitigate variations in " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\theta_{0}^{t}" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\theta_{1}^{t}" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": ", at least when " + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 280, + 296, + 376 + ], + "type": "text", + "content": " is not strongly dominating." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": "Now, the core object of interest is the map " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "f: \\{0,1\\}^2 \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": ", which is responsible for modeling class pair interactions and prompt-level biases (see Table 1). In an ideal situation for MLR, " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": " would behave roughly as an 'AND' function, so that compound prompts would let us directly hone in on the presence of pairs of classes. However, prior work has observed that instead, for CLIP-type models, " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": " tends to predominantly behave as an 'OR' function, in that " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": " is high if either " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": " is present. Notice that if " + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 376, + 296, + 507 + ], + "type": "text", + "content": " were to exactly behave as an OR, compound prompts would not be effective in detecting classes that were missed by the singleton scores." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 508, + 296, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 603 + ], + "type": "text", + "content": "In experimentation, we found that the behavior of " + }, + { + "bbox": [ + 55, + 508, + 296, + 603 + ], + "type": "inline_equation", + "content": "s_{i,j}^{t}" + }, + { + "bbox": [ + 55, + 508, + 296, + 603 + ], + "type": "text", + "content": " is intermediary to the above extremes: CLIP scores for compound prompts behave qualitatively as an OR function with a small AND 'bonus' when both classes are present. This bonus, as well as the particular scores, are modulated by the classes mentioned (a consequence of prompt-level bias for each class) which we model as linear effects, leading to an overall model of the signal term of the CLIP score as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 612, + 295, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 612, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 85, + 612, + 295, + 645 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f \\left(y _ {i} ^ {t}, y _ {j} ^ {t}\\right) := \\max \\left(a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}\\right) \\\\ + \\delta_ {i j} \\min (a _ {0} ^ {i} + y _ {i} ^ {t} a _ {1} ^ {i}, a _ {0} ^ {j} + y _ {j} ^ {t} a _ {1} ^ {j}), \\qquad (6) \\\\ \\end{array}", + "image_path": "495dd2e349fcf68595b3972c7604cbfae6679962ae6849a0d5e07089d8b2d0a5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " terms are class-wise linear effect coefficients, and " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " captures the strength of this AND bonus." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Note that this AND bonus is a critical feature of CLIP scores in their use for MLR problems, since this bump allows us to both infer the presence of classes that singleton" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 554, + 207 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 207 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 207 + ], + "type": "table", + "html": "
Noise Modelf(yi,t,yj,t)FVU
constant0.10.802
only ANDmin(a0i+yitai,aj+ytaj)0.536
only ORmax(a0i+yitai,aj+ytaj)0.288
additivea0i+yitai+aj+ytaj0.267
OR + static AND-bonusmax(a0i+yitai,aj+ytaj) + δ min(a0i+yitai,aj+ytaj)0.263
OR + variable AND-bonusmax(a0i+yitai,aj+ytaj) + δi,j min(a0i+yitai,aj+ytaj)0.248
look-up tableLUT(i,j)(yi,t,yj,t)0.235
", + "image_path": "a4730ee9e35c0c9ed4293501c89df210b54949ef23c66a5474e2552f96a1e960.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 211, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 255 + ], + "type": "text", + "content": "Table 1. Comparison of fidelity of noise models for scoring pairwise compound prompts. Notice that the OR+AND-bonus model captures nearly all of the fidelity of the look-up table, and offers significant improvements over the simplified models above it." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "content": "prompts have missed, and indirectly to filter false positives (since the true positives are further elevated by the bonus)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": "Table 1 shows the results of a systematic evaluation of the AND-OR structure of compound CLIP scores, using 81K COCO test images and ViT-L/14@336px backbone. We evaluate various models of " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": ", ranging from simple constant, 'only OR,' and 'only AND' models to additive models and OR+AND-bonus models with both a 'static' " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " shared by all class pairs, and a 'variable' " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\delta_{i,j}" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " that is class-pairspecific. The final entry is a 'look-up table' (LUT) model where all four values of " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " can vary arbitrarily, which serves as the most expressive baseline in this model class. These are evaluated w.r.t. the fraction of variance unexplained (i.e., " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "1 - R^2" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " value). Observe that while the only OR model has remarkable fidelity, the additive and AND-bonus models significantly improve upon this, and allowing " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " to vary with " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " yield a further significant improvement to within a " + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 290, + 554, + 506 + ], + "type": "text", + "content": " extra loss from the LUT model. This shows that our AND-bonus models capture a significant amount of the qualitative behavior of CLIP scores for compound prompts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "text", + "content": "A closer look at these fitted models reveals the " + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "text", + "content": " values are small and relatively stable: for static " + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "text", + "content": ", the recovered value is 0.56, and the IQR for varying " + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "inline_equation", + "content": "\\delta_{ij}" + }, + { + "bbox": [ + 313, + 506, + 553, + 542 + ], + "type": "text", + "content": " is (0.44, 0.57)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 542, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 553, + 590 + ], + "type": "text", + "content": "We find similar behavior when examining other backbones and datasets, which we detail in the Supplementary. The Supplementary also contains a theoretical view of the benefits of using weakened-max, based on the noise model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 598, + 471, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 471, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 471, + 611 + ], + "type": "text", + "content": "3.1. Methodological Implications" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": "The qualitative structure of the OR+AND-bonus behavior justifies our methodological choices in §2. Image-level normalization reduces the effect of variation in " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\theta_0^t" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": " (via mean subtraction) and " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\theta_1^t" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": " (via rescaling), while prompt-level normalization mitigates variations in " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "a_0^i" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "a_1^i" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": " across classes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "We can also qualitatively observe that, because the OR component of the scores dominates the AND component, the maximum score " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "r_{i,1}^{t}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " can create false positives when " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4317" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "absent but related class " + }, + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": " is present, and so we should look to subsequent order statistics for a better signal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 136, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 136, + 123 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 136, + 123 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 295, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 237 + ], + "type": "text", + "content": "Datasets We benchmark our method on 3 different MLR datasets. They are COCO2014 [9], which has 40,504 test images with 80 classes; VOC2007 [6], which has 4.952 test images with 20 classes; and NUSWIDE [2], which has 107,859 test images with 81 classes. We use only test, and not train, images from these datasets. We do use their conditional ground-truth label distributions, computed from training set, to choose compound prompts, but do not use ground-truth probabilities in any other part of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 249, + 295, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 295, + 369 + ], + "type": "text", + "content": "CLIP Backbones In order to showcase the generality of our method, we apply it to 9 different CLIP backbones, which are ViT-L/14@336px, ViT-L/14, ViT-B/16, ViT-B/32, RN50x64, RN50x16, RN50x4, RN101, RN50. It is worth noting that from our model's perspective, each backbone could be seen as a dataset of its own, providing a different source of scores for each image set. We use the default 224x224 resolution, resizing without center crop, for all ViT models. For ResNet models, we use the same resolution as [7] except when otherwise stated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 295, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 403 + ], + "type": "text", + "content": "4.1. Our method improves over vanilla ZSCLIP across many datasets and models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 410, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 468 + ], + "type": "text", + "content": "As a baseline, we compute singleton scores using prompt-ensembling with the 80 templates from the CLIP paper [16]. We then compute compound prompt scores and apply our method to the singletons and compounds. We compare our results to this vanilla ZSCLIP baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": "Tab. 2 shows our results over three different datasets and nine different CLIP backbones. We see that our method is able to improve over vanilla ZSCLIP by " + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "inline_equation", + "content": "12.6\\%" + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": " on COCO, " + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": " on VOC, and " + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": " on NUSWIDE. Moreover, we see that our method achieves consistent improvements across all datasets and backbones, showcasing its generality. For example, the improvements range from " + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "inline_equation", + "content": "11.4 - 15.0\\%" + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": " for COCO, " + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "inline_equation", + "content": "7.4 - 10\\%" + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": " for VOC, and " + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "inline_equation", + "content": "6.1 - 10.8\\%" + }, + { + "bbox": [ + 55, + 470, + 295, + 590 + ], + "type": "text", + "content": " for NUSWIDE. In fact, we see in the Supplementary that SPARC consistently improves the AP of almost every class in all datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 598, + 264, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 264, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 264, + 611 + ], + "type": "text", + "content": "4.2. Our method complements local features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "We consider another training-free zero-shot baseline introduced by the TaI-DPT paper [4] which we call \"CLIP-DPT\". This method modifies ZSCLIP by replacing the final global pooling in the image backbone with a parameter-free attention pooling where the query is the text embedding, allowing us to harness local features that might otherwise be destroyed. We incorporate CLIP-DPT into our method, in the same way that we incorporated ZSCLIP in the previous" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "section, and see a consistent improvement in Tab. 3, suggesting that our method is adding a complementary strength and not just compensating for lost local information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 118, + 534, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 534, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 534, + 131 + ], + "type": "text", + "content": "4.3. Our method has complementary strengths" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 137, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 232 + ], + "type": "text", + "content": "Several existing methods adapt CLIP to MLR tasks in an image-free ([4], [22], [21], [11], [23]) or training-free ([8], [10]) manner. These approaches lie outside our scope, as we focus on methods that are deployable 'out-of-the-box' on new datasets and VLMs, without requiring dataset-specific training or VLM surgery. Our black-box method allows easy 'plug-and-play' integration, using scores from existing methods in place of the vanilla ZSCLIP singleton scores." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 233, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 553, + 328 + ], + "type": "text", + "content": "We combine our method with three existing methods that have public codebases: (1) TagCLIP [10], a training-free method that benchmarks on COCO and VOC with a ViT-B/16 backbone; (2) TaI-DPT [4], an image-free method that requires training, which benchmarks on COCO, VOC, and NUSWIDE with a RN50 backbone; (3) CoMC [11], an image-free method that requires training and has a publicly-available RN50 model for COCO for one train seed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 330, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 509 + ], + "type": "text", + "content": "There are some nuances regarding image preprocessing and score postprocessing. TaI-DPT and CoMC use different image preprocessing settings than us - theirs are lower resolution and take a center crop. For a fair comparison, we use these same preprocessing settings when getting compound prompt scores to combine with TaI-DPT and CoMC. We also try running both our and their method with our preprocessing setting, which we denote as \"RN50*\". TagCLIP takes images at their native resolution, so for that comparison we do the same in order to be apples-to-apples. As for score postprocessing, TagCLIP uses softmax, which we have found to have a normalizing effect, so we omit our image-level normalization on TagCLIP's scores, and take the log of their scores to make them compatible with ours. We make no postprocessing changes for TaI-DPT or CoMC." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "content": "We show the results of these plug-and-play combinations in Tab. 4 and Tab. 5. We find that method is able to improve on TagCLIP by on average " + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "content": " and on TaI-DPT by on average " + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "content": ". We do not improve on CoMC, but the degradation is only " + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 510, + 553, + 594 + ], + "type": "text", + "content": ". These results suggest that our method's signal complements those obtained through training and architectural manipulation by other methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 602, + 382, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 382, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 382, + 613 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 622, + 444, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 444, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 444, + 633 + ], + "type": "text", + "content": "4.4.1 Normalization Module" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "We take a closer look at the Normalization module by seeing how it impacts the performance of both singleton prompts and our full pipeline. We see in Tab. 7 that the Normalization module provides significant gains in both cases (7.7% with singletons, 8.6% full pipeline), confirming that bias significantly impacts MLR when left unaddressed." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4318" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 558, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 558, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 558, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 558, + 194 + ], + "type": "table", + "html": "
DatasetViT-L/14 336pxViT-L/14ViT-B/16ArchitecturesAvg (archs)
ViT-B/32RN50 x64RN50 x16RN50 x4RN101RN50
COCOZSCLIP59.158.155.550.958.658.155.752.753.055.7
Ours70.569.567.464.070.670.169.567.765.768.3
VOCZSCLIP81.379.979.977.583.181.680.680.379.780.4
Ours88.988.388.787.590.590.390.089.789.289.2
NUSWIDEZSCLIP41.441.040.938.739.838.538.835.738.539.3
Ours47.547.147.346.947.447.948.346.545.847.2
", + "image_path": "967404bff832c4da37ff93b914109592b13c6c59cba2cd766ea8ae64bbb055d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 225, + 309, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 222 + ], + "type": "text", + "content": "Table 2. Results over three datasets and nine CLIP backbones. Our proposed method consistently outperforms the ZSCLIP baseline across all datasets and architectures exhibiting its effectiveness on multi-label recognition tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 225, + 309, + 338 + ], + "lines": [ + { + "bbox": [ + 58, + 225, + 309, + 338 + ], + "spans": [ + { + "bbox": [ + 58, + 225, + 309, + 338 + ], + "type": "table", + "html": "
RN50 x64RN50 x16RN50 x4RN 101RN 50
COCOCLIP-DPT70.564.264.259.757.1
Ours-DPT77.873.972.769.868.4
VOCCLIP-DPT86.682.486.183.582.6
Ours-DPT91.189.888.588.589.9
NUSCLIP-DPT39.837.839.338.438.1
Ours-DPT43.545.542.742.644.5
", + "image_path": "49e25a27821f3f872e0236e26d6138132d862840c43215e38c4b92ddd5bdfe8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 79, + 376, + 269, + 437 + ], + "blocks": [ + { + "bbox": [ + 55, + 340, + 296, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 374 + ], + "type": "text", + "content": "Table 3. Comparing SPARC with local-feature method CLIP-DPT [4]. Our method still improves upon local features, showing that its strength is complementary." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 376, + 269, + 437 + ], + "lines": [ + { + "bbox": [ + 79, + 376, + 269, + 437 + ], + "spans": [ + { + "bbox": [ + 79, + 376, + 269, + 437 + ], + "type": "table", + "html": "
COCO\nViT-B/16VOC\nViT-B/16Avg
TagCLIP70.991.781.3
+Ours73.892.082.9
", + "image_path": "7b6b8b590cac581a66f1a6ac81661209268a0c678cba90a845476e20392652d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 462 + ], + "type": "text", + "content": "Table 4. Shows complementarity with architectural approaches, improving TagCLIP by " + }, + { + "bbox": [ + 55, + 441, + 295, + 462 + ], + "type": "inline_equation", + "content": "1.6\\%" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 89, + 463, + 245, + 585 + ], + "blocks": [ + { + "bbox": [ + 89, + 463, + 245, + 585 + ], + "lines": [ + { + "bbox": [ + 89, + 463, + 245, + 585 + ], + "spans": [ + { + "bbox": [ + 89, + 463, + 245, + 585 + ], + "type": "image", + "image_path": "8efa8dba3110c470bc03f30584068af4beb600573fb9925b77907b4a81449ada.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 588, + 295, + 611 + ], + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 611 + ], + "type": "text", + "content": "Figure 4. Average mAP for different Rank Fusion strategies demonstrates superiority of adaptive fusion over fixed strategies." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 178, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 178, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 178, + 629 + ], + "type": "text", + "content": "4.4.2 Rank Fusion Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 638, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 295, + 673 + ], + "type": "text", + "content": "Our next ablation takes a closer look at the role of the Rank Fusion module. We consider some possible handcrafted strategies for combining compound prompt scores:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 673, + 289, + 702 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "text", + "content": "- \" " + }, + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "text", + "content": " max\": Use the " + }, + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "text", + "content": "-th highest compound score, i.e. " + }, + { + "bbox": [ + 55, + 673, + 288, + 687 + ], + "type": "inline_equation", + "content": "r_{i,k}^{t}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 687, + 289, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 289, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 289, + 702 + ], + "type": "text", + "content": "- \"mean " + }, + { + "bbox": [ + 55, + 687, + 289, + 702 + ], + "type": "inline_equation", + "content": "> = k" + }, + { + "bbox": [ + 55, + 687, + 289, + 702 + ], + "type": "text", + "content": "\": Use partial avg " + }, + { + "bbox": [ + 55, + 687, + 289, + 702 + ], + "type": "inline_equation", + "content": "\\frac{1}{m_i - K + 1}\\sum_{k = K}^{m_i - K + 1}r_{i,k}^t" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 702, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 294, + 713 + ], + "type": "text", + "content": "Fig. 4 shows the average mAP of these, as well as \"maxVari" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "content": "ance” (3) and singleton-only, across all datasets and backbones. We see that 1st-max does much worse than 2nd-max, which underperforms 3rd-max, and so on, reaching a plateau around 5th-max. This lends credence to the idea that a weakened max is a stronger signal than outright max. We also see that the mean of all compound prompts (i.e. “mean " + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "inline_equation", + "content": ">= 1" + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "content": ") does surprisingly well. This is actually consistent with our intuition on weakened maxes - the mean weakens the higher maxes by averaging them with the lower ones. We can improve further by excluding high ranks in the “mean " + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "inline_equation", + "content": ">= k" + }, + { + "bbox": [ + 313, + 228, + 555, + 396 + ], + "type": "text", + "content": "\" strategy. However, we can surpass all of these by adaptively weighting scores using the max-variance direction. The Supplementary repeats this analysis without the “merge” step (5), which turns out to be critical as well." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 416, + 434, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 434, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 434, + 428 + ], + "type": "text", + "content": "4.4.3 Compound Prompts" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "content": "Our next ablation takes a closer look at the importance of the composition of our set of compound prompts and the use of rough cooccurrence info. The results are in Tab. 6. We start by including all pairs of classes in our set of formulaic pairwise prompts (i.e. of the form \"A and B\") and not including any other kind of compound prompt. We find that this gives us a " + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "content": " average boost over normalized singletons. Filtering these pairwise prompts by cooccurrence, as our proposed method does, yields a further " + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "content": " boost, and adding triplet and descriptive prompts gives an additional " + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 664 + ], + "type": "text", + "content": " boost. It makes sense that using the full set of class pairs would still have some benefit, without any cooccurrence filtering, if all the classes as a whole tend to cooccur positively more often than negatively, as that would mean that they are on average a helpful signal for predicting each other's presence. It might also be the case that related classes provide a useful context signal to CLIP. In any case, it is noteworthy that Rank Fusion can still help even when no cooccurrence information is required from the user." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Our final ablation aims to figure out whether our performance boost actually comes from the semantics of the compound prompts themselves, and not just an increase in number of prompts. Each class has on average " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\leq 20" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " compound" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4319" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 539, + 130 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 539, + 130 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 539, + 130 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 539, + 130 + ], + "type": "table", + "html": "
COCOVOCNUSWIDEAvgCOCOAvg
RN50RN50*RN50RN50*RN50RN50*RN50RN50*
TaI-DPT65.168.288.588.046.243.466.6CoMC68.871.370.0
+Ours68.270.190.390.246.644.668.3+Ours68.970.669.7
", + "image_path": "8ee5c00b46006c79fa00b9cced55b45d7d91f7134ca59f4def06d6bd80e0faee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 161, + 547, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "type": "text", + "content": "Table 5. Results from combining our method with TaI-DPT (left) and CoMC (right) showing compatibility with training-based methods. Our method improves TaI-DPT by " + }, + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "type": "text", + "content": " while preserving the existing strong signal of CoMC, degrading it by only " + }, + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 55, + 133, + 553, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 161, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 61, + 161, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 547, + 258 + ], + "type": "table", + "html": "
NormalizePair promptsTriplets + DescriptiveRank Fusion StrategyCOCOVOCNUSAvg
✓--65.987.745.166.2
✓all pairsours67.588.546.467.5
✓all pairsmean67.588.546.467.5
✓cooccurrence-filteredours68.189.047.068.0
✓cooccurrence-filteredmean67.988.546.867.7
✓cooccurrence-filtered✓ours68.389.247.268.3
", + "image_path": "5ac660ea03eb03782ee121c767c600648259f16247fd2ee786ba5e2445e7773b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 285, + 286, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 259, + 553, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 553, + 282 + ], + "type": "text", + "content": "Table 6. Ablations on the makeup of our compound prompts. We find that we can still enjoy some benefit from pairwise prompts without any cooccurrence filtering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 285, + 286, + 350 + ], + "lines": [ + { + "bbox": [ + 63, + 285, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 63, + 285, + 286, + 350 + ], + "type": "table", + "html": "
CompoundNormalizeCOCOVOCNUSAvg
55.780.439.358.5
✓65.987.745.166.2
✓58.480.540.059.7
✓✓68.389.247.268.3
", + "image_path": "f04526a18b4f8941016938f8da732bdd33438ef68ec77436ad8564d6cb518f74.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 352, + 295, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 374 + ], + "type": "text", + "content": "Table 7. Ablations on Normalization module. Quantifies impact of normalization both with and without compound prompts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 382, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 295, + 418 + ], + "type": "text", + "content": "prompts, 4x less than the 80 singleton prompt templates already used by vanilla ZSCLIP, but there is still an increase in total number of prompts, so an ablation is warranted." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "text", + "content": "We take inspiration from WaffleCLIP [5], which found that prompt ensembles made of random texts could perform just as well as descriptive ensembles generated by LLMs. For our ablation, we use WaffleCLIP-style prompts in place of compound prompts. Specifically, for eachclassname " + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "text", + "content": " we create 30 prompts of the form \"A photo of a " + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "text", + "content": ", which is [RAND]\", where \"[RAND]\" is 10 random characters. We try both the \"maxVariance\" and \"mean " + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "inline_equation", + "content": "> = 1" + }, + { + "bbox": [ + 54, + 419, + 295, + 527 + ], + "type": "text", + "content": "\" fusion strategies, as the latter is more typical for assembling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": "We show the results of this ablation in the Supplementary. We see that randomized compound prompts do not offer any complementary signal to the singleton prompts. Thus, any gain from compound prompts must come from their semantics. It is unlikely that simple prompt diversity explains the gains either, given that vanilla ZSCLIP already uses a large and diverse template ensemble." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 622, + 263, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 263, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 263, + 635 + ], + "type": "text", + "content": "4.5. A qualitative look at the weakened max" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "We saw in Sec. 4.4.1 that the second-highest compound prompt score is a much better signal than the highest one. This counterintuitive finding can be explained through object cooccurrence. When CLIP is given a compound prompt like 'cat and sofa', it has OR-like behavior - giving high scores if either object is present. The maximum score often" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 287, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 357 + ], + "type": "text", + "content": "comes from such prompts where only one object is present but strongly detected. In contrast, the second-highest scores tend to come from prompts where both objects have moderate confidence, providing a more reliable signal. This explains why using second-highest scores results in better separation between positive and negative cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 358, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 478 + ], + "type": "text", + "content": "Fig. 3 shows an example of what happens at an image level. On the right is an image that contains a cat (curled up behind the two dogs on the bed); on the left is an image without a cat. The highest-scoring compound prompt is \"Cat and Dog\" for both images. This prompt gives a high score to the right image by detecting the dogs, but it also detects the dog in the left image for the same reason, and so it also misorders the images. But the second-highest-scoring prompts - \"Bed and Cat\" for right image, \"Cat and Potted plant\" for left - correctly order the images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 526 + ], + "type": "text", + "content": "We show score histograms for one of the classes in the Supplementary to illustrate what happens in aggregate - the second-highest score lifts positive examples without introducing as many false-positives as first-highest score." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 536, + 391, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 391, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 391, + 549 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 556, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 712 + ], + "type": "text", + "content": "We presented SPARC, a zero-shot multi-label recognition approach that improves VLM performance without training data or architectural changes. Our key insights - the suboptimal nature of maximum scores and presence of systematic biases - led to two complementary innovations: compound prompts with adaptive fusion, and systematic normalization. Beyond improving standalone VLM performance, SPARC enhances existing zero-shot and training-based methods while maintaining a black-box approach. The success of our method reveals fundamental properties of VLM scoring behavior, suggesting promising directions for improving zero-shot recognition through score analysis rather than architectural modification or finetuning." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4320" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "text", + "content": "[2] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM International Conference on Image and Video Retrieval, New York, NY, USA, 2009. Association for Computing Machinery. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 215, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 215, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 295, + 258 + ], + "type": "text", + "content": "[3] Reza Esfandiarpoor and Stephen H Bach. Follow-up differential descriptions: Language models resolve ambiguities for image classification. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 281 + ], + "type": "text", + "content": "[4] Guo et al. Texts as images in prompt tuning for multi-label image recognition. In CVPR, '23. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 315 + ], + "type": "text", + "content": "[5] Roth et al. Waffling around for performance: visual classification with random words and broad concepts ICCV, '23. 8, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "text", + "content": "[6] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "text", + "content": "[7] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "type": "text", + "content": "[8] Yi Li, Hualiang Wang, Yiqun Duan, and Xiaomeng Li. Clip surgery for better explainability with enhancement in open-vocabulary tasks. arXiv preprint arXiv:2304.05653, 2023. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 521 + ], + "type": "text", + "content": "[9] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 588 + ], + "type": "text", + "content": "[10] Yuqi Lin, Minghao Chen, Kaipeng Zhang, Hengjia Li, Mingming Li, Zheng Yang, Dongqin Lv, Binbin Lin, Haifeng Liu, and Deng Cai. Tagclip: A local-to-global framework to enhance open-vocabulary multi-label classification of clip without training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3513-3521, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "type": "text", + "content": "[11] Yicheng Liu, Jie Wen, Chengliang Liu, Xiaozhao Fang, Zuoyong Li, Yong Xu, and Zheng Zhang. Language-driven cross-modal classifier for zero-shot multi-label image recognition. In *Forty-first International Conference on Machine Learning*, 2024. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 700 + ], + "type": "text", + "content": "[12] Leilei Ma, Hongxing Xie, Lei Wang, Yanping Fu, Dengdi Sun, and Haifeng Zhao. Text-region matching for multi-label image recognition with missing labels. In Proceedings of the 32nd ACM International Conference on Multimedia, page 6133-6142. ACM, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 702, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 294, + 712 + ], + "type": "text", + "content": "[13] Sachit Menon and Carl Vondrick. Visual classification via" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 612 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "text", + "content": "description from large language models. arXiv preprint arXiv:2210.07183, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[14] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In International Conference on Machine Learning, pages 26342-26362. PMLR, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[15] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "type": "text", + "content": "[16] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 260, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 260, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 260, + 553, + 315 + ], + "type": "text", + "content": "[17] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In Proceedings of the International Conference on Pattern Recognition (ICPR), 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 358 + ], + "type": "text", + "content": "[18] Ximeng Sun, Ping Hu, and Kate Saenko. Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information Processing Systems, 35:30569-30582, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 360, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 553, + 402 + ], + "type": "text", + "content": "[19] Hao Tan, Zichang Tan, Jun Li, Jun Wan, Zhen Lei, and Stan Z. Li. Sspa: Split-and-synthesize prompting with gated alignments for multi-label image recognition. ArXiv, abs/2407.20920, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 404, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 479 + ], + "type": "text", + "content": "[20] Ao Wang, Hui Chen, Zijia Lin, Zixuan Ding, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Guiguang Ding. Hierarchical prompt learning using clip for multi-label classification with single positive labels. In Proceedings of the 31st ACM International Conference on Multimedia, page 5594-5604, New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 481, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 481, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 481, + 553, + 524 + ], + "type": "text", + "content": "[21] Xiangyu Wu, Qing-Yuan Jiang, Yang Yang, Yi-Feng Wu, Qing-Guo Chen, and Jianfeng Lu. Tai++: Text as image for multi-label image classification by co-learning transferable prompt. arXiv preprint arXiv:2405.06926, 2024. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 525, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 568 + ], + "type": "text", + "content": "[22] Shuo Yang, Zirui Shang, Yongqi Wang, Derong Deng, Hongwei Chen, Qiyuan Cheng, and Xinxiao Wu. Data-free multi-label image recognition via llm-powered prompt tuning. arXiv preprint arXiv:2403.01209, 2024. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 569, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 612 + ], + "type": "text", + "content": "[23] Xueling Zhu, Jiuxin Cao, Jian Liu, Dongqi Tang, Furong Xu, Weijia Liu, Jiawei Ge, Bo Liu, Qingpei Guo, and Tianyi Zhang. Text as image: Learning transferable adapter for multi-label classification. ArXiv, abs/2312.04160, 2023. 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements. We thank Ping Hu, Siqi Wang, and Piotr Teterwak for their valuable discussions, as well as the anonymous reviewers for their thorough and helpful feedback. This research was supported by the Army Research Office Grant W911NF2110246, AFRL Grant FA8650-22-C1039, and the National Science Foundation grants CPS-2317079, CCF-2007350, and CCF-1955981." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4321" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_content_list.json b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e824a19e5134b8e8e859a56f7e68255693117f72 --- /dev/null +++ b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_content_list.json @@ -0,0 +1,1651 @@ +[ + { + "type": "text", + "text": "SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction", + "text_level": 1, + "bbox": [ + 135, + 128, + 859, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yutao Tang*", + "bbox": [ + 238, + 203, + 339, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxiang Guo*", + "bbox": [ + 372, + 203, + 488, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deming Li", + "bbox": [ + 527, + 203, + 617, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Peng", + "bbox": [ + 656, + 203, + 756, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Johns Hopkins University", + "bbox": [ + 393, + 224, + 602, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ytang67，yguo87，dli90，cpeng26}@jhu.edu", + "bbox": [ + 315, + 244, + 676, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 294, + 326, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent efforts in Gaussian-Splat-based Novel View Synthesis can achieve photorealistic rendering; however, such capability is limited in sparse-view scenarios due to sparse initialization and over-fitting floaters. Recent progress in depth estimation and alignment can provide dense point cloud using few views; however, the resulting pose accuracy is suboptimal. In this work, we present SPARS3R, which combines the advantages of accurate pose estimation from Structure-from-Motion and dense point cloud from depth estimation. To this end, SPARS3R first performs a Global Fusion Alignment process that maps a prior dense point cloud to a sparse point cloud from Structure-from-Motion based on triangulated correspondences. RANSAC is applied during this process to distinguish inliers and outliers. SPARS3R then performs a second, Semantic Outlier Alignment step, which extracts semantically coherent regions around the outliers and performs local alignment in these regions. Along with several improvements in the evaluation process, we demonstrate that SPARS3R can achieve photorealistic rendering with sparse images and significantly outperforms existing approaches.", + "bbox": [ + 89, + 325, + 485, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 672, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photorealistic scene reconstruction and Novel View Synthesis (NVS) from unposed 2D images is a challenging task with numerous applications in site modeling, autonomous driving, robotics, urban and agricultural planning, etc. The introduction of methods such as Neural Radiance Field (NeRF) [35] and 3D Gaussian Splatting (3DGS) [24] have introduced significant advances in rendering quality and efficiency based on dense multi-view imagery. However, applications of these methods still face issues in practical scenarios where a dense coverage of the scene is not available.", + "bbox": [ + 89, + 696, + 482, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given sparse images, over-fitting the photometric objectives to an incorrect geometry is a common issue in", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a784259f4c66dd91775151c1516e7a0efa985b79a39fb642e70049b092546f2.jpg", + "image_caption": [ + "Instant-NGP [36]", + "FSGS [65]" + ], + "image_footnote": [], + "bbox": [ + 516, + 294, + 905, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7186186877bf36149aa9b12a105ead39acc7b6a2c3d5c1eac9b2e2080ae13c07.jpg", + "image_caption": [ + "InstantSplat [14]", + "SPARS3R", + "Figure 1. A visualization of SPARS3R in comparison to current SoTA. Without additional prior, sparse NVS leads to incorrect geometry by Instant-NGP [36]. FSGS [65] can be blurry due to sparse initialization and insufficient densification. InstantSplat [14] relies on DUSt3R [52] initialization with suboptimal poses. Our method, SPARS3R, can reliably render details in the foreground and background with accurate poses." + ], + "image_footnote": [], + "bbox": [ + 516, + 410, + 905, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVS. Various constraints, such as semantic consistent loss [23], depth and novel-view regularization [12, 39, 51], frequency regularization [57], and ray entropy minimization [25] have been introduced to improve upon NeRF. These methods often lead to significant computation overhead due to the costly patch rendering from ray-tracing. More recently, Gaussian-Splatting-based approaches have further improved sparse NVS by leveraging the explicit representation and fast differentiable rasterization. Depth regularization [30], Gaussian floater pruning [55], and proximity-based Gaussian densification strategy [65] have been proposed to constrain and guide the scene structures. As demonstrated in Fig. 1, despite improved geometry, these approaches tend to produce overly smooth rendering in the background. This issue can be attributed to a sparse initial", + "bbox": [ + 511, + 672, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26810", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "point cloud, particularly in the background regions. Furthermore, the additional constraints, e.g. based on monocular depth prior, are oftentimes imperfect, leading to noisy gradients that prevent proper densification in sparse regions.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One potential solution to address this conundrum is to initialize Gaussian optimization with a denser point cloud to help disambiguate shape and radiance. To this end, recent advances in two-view depth estimation, particularly DUSt3R [52] and MAST3R [29], have shown the impressive ability to construct a dense point cloud from a pre-trained prior model. In theory, such a point cloud can be directly used to optimize a Gaussian-Splat radiance field [14]. In practice, camera calibration obtained from multi-view depth alignment is often suboptimal due to the difficulties in estimating an accurate depth map. If left uncorrected, the Gaussian optimization process will generate floaters to compensate the suboptimal calibration, prompting strategies to prevent densification as a stopgap measure [14]. In comparison, camera calibration based on Structure-from-Motion (SfM), e.g., COLMAP [46], is fast and accurate with an appropriate feature matching model. Instead of aligning dense depth, SfM takes confident correspondences and performs triangulation with RANSAC [15] to reject outliers. In a sparse-view setting, such a process can be repeated numerous times to further improve accuracy.", + "bbox": [ + 91, + 154, + 482, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address sparse point cloud initialization and pose inaccuracy in sparse-view NVS, we propose SPARS3R. SPARS3R composes of two stages: Global Fusion Alignment and Semantic Outlier Alignment. In Global Fusion Alignment, SPARS3R first obtains a dense point cloud from sparse views through DUSt3R [52] or MASt3R [29], and a sparse point cloud from COLMAP [46]. By taking the triangulated correspondences within images, SPARS3R fuses the dense point cloud onto the sparse point cloud through a global Procrustes Alignment process with RANSAC [15]. Since some points will yield large error due to local scale variations, a second Semantic Alignment process is introduced to perform local alignment. Specifically, outliers from Global Fusion Alignment are identified and used as prompts for an Interactive Segmentation model, e.g. SAM [26], generating semantically coherent masks around the outliers. The resulting semantic masks indicate regions within dense point cloud to be treated with local alignment. By locally aligning the identified outliers with the SfM point cloud while preserving the globally aligned inliers, we obtain a dense and pose-wise accurate point cloud, serving as a strong prior for Gaussian optimization.", + "bbox": [ + 91, + 470, + 482, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions can be summarized as:", + "bbox": [ + 109, + 805, + 464, + 819 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a Global Fusion Alignment approach, which transforms a prior dense point cloud onto a reference SfM sparse point cloud, putting dense initialization and accurate camera poses in the same coordinate frame.", + "2. To address outliers that cannot be aligned accurately due" + ], + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to depth discrepancies, we propose a Semantic Outlier Alignment step. This step extracts semantically similar regions around the outliers to perform local alignment, resulting in a dense point cloud with minimum transformation error.", + "bbox": [ + 532, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We evaluate the overall method, SPARS3R, on three popular benchmark datasets and find significant quantitative and visual improvements compared to current SoTA methods.", + "bbox": [ + 513, + 167, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 233, + 653, + 248 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Models for Synthesizing Novel Views", + "text_level": 1, + "bbox": [ + 513, + 255, + 857, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Photorealistic scene reconstruction and novel view synthesis is a long-standing task in computer vision and graphics. Neural Radiance Field (NeRF) [35] proposes to model a scene implicitly with multi-layer perceptron (MLP) by mapping 3D coordinates and view direction to color and density, then rendering pixel value via alpha blending. Since NeRF's introduction, numerous work has sought to improve its efficiency [5, 16, 20, 36, 45, 48], quality [1-3, 7, 50], and extension to dynamic and in-the-wild [4, 6, 8, 17, 34, 42, 44, 47, 54, 58] scenarios. Recently, 3D Gaussian Splatting (3DGS) [24] has emerged as an efficient alternative, offering improved rendering quality over NeRF. Unlike NeRF, 3DGS represents scenes with explicit 3D Gaussians kernels and uses a differentiable rasterization technique. Following its introduction, many methods have emerged to reduce the computational cost [13, 21, 28, 37, 38] and enhance quality [9, 33, 40, 43, 56, 61, 63, 64] of 3DGS.", + "bbox": [ + 511, + 277, + 903, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Gaussian-Splatting-based Sparse-View NVs", + "text_level": 1, + "bbox": [ + 513, + 537, + 885, + 554 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While 3DGS performs well with dense view support, its effectiveness decreases in practical settings with sparse inputs. Many approaches address this through depth constraints. FSGS [65] uses a Proximity-guided Gaussian Unpooling to mitigate the sparse initialization issue and introduces pseudo-views during training to avoid overfitting. DRGS [11] and SparseGS [55] add a depth regularization term to enforce consistency between estimated and monocular depths while promoting smoothness. DNGS [31] refines depth regularization by prioritizing local depth variations. CoherentGS [41] initializes Gaussians from monocular depth estimation and improves coherence through an optical flow constraint. InstantSplat [14] is a concurrent work that directly uses DUSt3R to produce dense point clouds and introduces a camera pose optimization strategy; densification of Gaussians is disabled to prevent the introduction of floaters due to suboptimal poses.", + "bbox": [ + 511, + 560, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 825, + 611, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 848, + 640, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DUSt3R [52] is a two-view depth estimation method that produces dense 3D point clouds from image pairs. Specif", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/946e9e00b1cae5139d38a7f6349ff06938ab3bd9742f08cb17fdea43306eb754.jpg", + "image_caption": [ + "Figure 2. SPARS3R combines a prior dense point cloud $\\chi$ and a sparse SfM point cloud $\\bar{\\mathbf{X}}$ . The prior $\\chi$ often has inferior depth accuracy compared to $\\bar{\\mathbf{X}}$ . SPARS3R first globally aligns all points in $\\chi$ onto $\\bar{\\mathbf{X}}$ , based on shared correspondences. Inliers and outliers are identified through alignment error. SPARS3R then extracts the semantically relevant 2D regions around the outliers to move local regions of $\\chi$ in groups, producing a dense point cloud $\\chi^{*}$ that is depth-wise and pose-wise accurate. By providing $\\mathcal{X} = \\chi^{*} \\bigcup \\bar{\\mathbf{X}}$ for Gaussian optimization, SPARS3R achieves photorealistic rendering under sparse-view condition." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ically, given an image pair $I^1$ and $I^2$ , DUSt3R generates dense pointmaps $\\mathbf{X}^{1,1}$ and $\\mathbf{X}^{2,1} \\in \\mathbb{R}^{H \\times W \\times 3}$ . A pointmap, e.g. $\\mathbf{X}^{1,1}$ , provides a mapping from each pixel $(u, v)$ in image $I^1$ to a corresponding 3D point $\\mathbf{X}^{1,1}(u, v)$ under the coordinate system of image $I^1$ , giving dense 2D-to-3D correspondences. Given $K > 2$ input images, DUSt3R [52] aggregates across all pairwise pointmap predictions by globally aligning pairwise pointmaps into a unified point cloud $\\chi$ . This alignment process also produces camera parameters $\\{W(I^n), J(I^n)\\}$ , which are the world-to-camera transformation matrix and the Jacobian of the affine approximation of the projective transformation for view $I^n$ . As such, the relationship between a pixel $(u, v)$ on $I^n$ and its 3D point $\\chi_i$ is established via a projective function $P$ as follows:", + "bbox": [ + 89, + 393, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} = [ u, v ] ^ {T} = P \\left(\\chi_ {i}\\right) = J \\left(I ^ {n}\\right) W \\left(I ^ {n}\\right) \\chi_ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 618, + 480, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on this, MASt3R [29] incorporates an additional matching head that generates discriminative local feature descriptors and pixel-level correspondences across paired images. We note that while MASt3R [29] uses this feature correspondence during its multi-depth alignment, the estimated poses are oftentimes inaccurate due to smooth depth predictions.", + "bbox": [ + 89, + 642, + 483, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. SPARS3R", + "text_level": 1, + "bbox": [ + 89, + 756, + 205, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As show in Fig. 2, SPARS3R builds upon the advancement in DUSt3R [52] and MASt3R [29] as a pre-trained 3D prior for sparse-view reconstruction. Firstly, SPARS3R performs SfM based on image correspondences, either from MASt3R [29] or other feature matching methods. SPARS3R then aligns the dense point cloud produced by MASt3R via rigid transformations in two stages: Global Fusion Alignment and Semantic Outlier Alignment.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Global Fusion Alignment", + "text_level": 1, + "bbox": [ + 511, + 393, + 735, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While DUSt3R [52] and MASt3R [29] can generate dense 3D points from sparse images, it does not estimate accurate or detailed depth through its pointmap formulation. As a result, the camera calibration obtained from such smooth depths is suboptimal and will degrade the fidelity of scene reconstructions. To construct a better point cloud prior, we propose to align MASt3R's point cloud with that from a SfM pipeline, which is more reliable based on selective correspondences. This alignment process begins by identifying matching points between the SfM and MASt3R point clouds, denoted as $\\overline{\\mathbf{X}}\\in \\mathbb{R}^{\\bar{N}\\times 3}$ and $\\chi \\in \\mathbb{R}^{N\\times 3}$ , respectively. A visibility indicator $V:\\mathbb{R}^{\\bar{N}\\times 3}\\to \\mathbb{R}^{\\widetilde{N}\\times 3}$ denotes if an SfM point $\\overline{\\mathbf{X}}_i$ is visible with respect to a view $I^n$ , and $\\tilde{N}$ denotes the number of visible SfM points in view $I^n$ . For each view, similar to Eq. (1), we can obtain the pixel-to-3D relationship of $(\\bar{u},\\bar{v})$ and $V(\\overline{\\mathbf{X}})_i$ through estimated $\\overline{W}(I^n)$ and $\\bar{J}(I^n)$ from SfM. Subsequently, the correspondence between a SfM point $V(\\overline{\\mathbf{X}})_i$ and a point in $\\chi$ , denoted as $\\tilde{\\chi}_i$ , can be established through a forward-backward projection:", + "bbox": [ + 511, + 412, + 906, + 704 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} _ {i} = \\left[ \\bar {u}, \\bar {v} \\right] ^ {T} = \\bar {P} (\\bar {\\mathbf {X}}) _ {i} = \\bar {J} (I ^ {n}) \\bar {W} (I ^ {n}) V (\\bar {\\mathbf {X}}) _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 724, + 903, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\chi} _ {i} = P ^ {- 1} (\\bar {x} _ {i}) = W ^ {- 1} \\left(I ^ {n}\\right) J ^ {- 1} \\left(I ^ {n}\\right) \\bar {x} _ {i}.\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 747, + 839, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each view, we denote the ordered set of visible SfM points as $\\{V(\\bar{\\mathbf{X}})_i\\}_{i = 1}^N$ , the projected 2D pixels as $\\{\\bar{x}_i\\}_{i = 1}^N$ , and the set of 3D points that correspond to the projected 2D pixels above in $\\chi$ as $\\{\\bar{\\chi}_i\\}_{i = 1}^N$ , omitting view indicator $n$ . With these identified correspondences, we can apply Generalized Procrustes Analysis [22] to estimate a global rigid transformation, optimizing for scale, rotation, and translation. During this process, RANSAC is applied to filter out-", + "bbox": [ + 511, + 776, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26812", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "liers, excluding their influence on the estimation. Specifically, we minimize the following objective to get the optimal global transformation based on Algorithm 1:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {0}, R _ {0}, t _ {0} = \\underset {s, R, t} {\\arg \\min } \\sum_ {i = 1} ^ {\\widetilde {N}} e \\left(\\bar {\\chi} _ {i}, s, R, t, V (\\bar {\\mathbf {X}}) _ {i}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 157, + 482, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne (\\bar {\\chi} _ {i}, s, R, t, V (\\bar {\\mathbf {X}}) _ {i}) = \\left\\| s R \\bar {\\chi} _ {i} + t - V (\\bar {\\mathbf {X}}) _ {i} \\right\\| ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 203, + 437, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 RANSAC Alignment", + "text_level": 1, + "bbox": [ + 91, + 246, + 323, + 262 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Matched MASt3R points $\\{\\overline{\\chi}_i\\}_{i=1}^N$ and SfM points $\\{V(\\bar{\\mathbf{X}})_i\\}_{i=1}^N$ , sample size $n$ , error threshold $\\epsilon$ , number of iterations $R$", + "2: Output: Best parameters $(s_0,R_0,t_0)$ , inlier set $\\mathcal{I}$ , outlier set $\\mathcal{O}$ , corresponding outlier SfM set $\\bar{\\mathcal{O}}$", + "3: Initialize: num_inliers $\\leftarrow 0$ , $(\\hat{s_0},\\hat{R_0},\\hat{t_0})\\gets \\emptyset ,\\hat{\\mathcal{I}}\\gets \\emptyset ,\\hat{\\mathcal{O}}\\gets \\emptyset ,\\hat{\\bar{\\mathcal{O}}}\\gets \\emptyset$ Iterations $\\leftarrow 0$", + "4: while Iterations $< R$ do", + "5: Randomly select $n$ points as sample set $S \\subset \\{i\\}_{i=1}^{\\widetilde{N}}$ .", + "6: Estimate $(\\hat{s}_0,\\hat{R}_0,\\hat{t}_0)$ on sample set $\\mathcal{S}$ by minimizing $\\sum_{i\\in \\mathcal{S}}e(\\bar{\\chi}_i,s,R,t,V(\\bar{\\mathbf{X}})_i)$ (Eq.3))", + "7: for each point $\\overline{\\chi}_i$ do", + "8: if $e(\\overline{\\chi}_i, \\hat{s}_0, \\hat{R}_0, \\hat{t}_0, V(\\overline{\\mathbf{X}})_i) < \\epsilon$ then", + "9: Add $\\bar{\\chi}_i$ to inlier set $\\hat{\\mathcal{I}}$", + "10: else", + "11: Add $\\bar{\\chi}_i$ to outlier set $\\hat{\\mathcal{O}}$ and add $V(\\overline{\\mathbf{X}})_i$ to outlier SfM set $\\hat{\\mathcal{O}}$", + "12: end if", + "13: end for", + "14: if $|\\mathcal{I}| > \\text{num\\_inliers}$ then", + "15: Update: num_inliers \\(\\leftarrow |\\hat{\\mathcal{I}}|, (s_0, R_0, t_0) \\leftarrow (\\hat{s}_0, \\hat{R}_0, \\hat{t}_0), \\mathcal{I} \\leftarrow \\hat{\\mathcal{I}}, \\mathcal{O} \\leftarrow \\hat{\\mathcal{O}}, \\bar{\\mathcal{O}} \\leftarrow \\hat{\\bar{\\mathcal{O}}}", + "16: end if", + "17: Iterations $\\leftarrow$ Iterations $+1$", + "18: end while", + "19: return $(s_0, R_0, t_0), \\mathcal{I}, \\mathcal{O}, \\mathcal{O}$" + ], + "bbox": [ + 96, + 268, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After this, we extract an inlier set $\\mathcal{I}$ , an outlier set $\\mathcal{O}$ and a corresponding outlier SfM set $\\bar{\\mathcal{O}}$ from RANSAC with an empirical threshold. As shown in Fig. 2, the global transformation recovered from Eq. 3 often leaves large amount of outliers between $\\bar{\\mathbf{X}}$ and $\\chi$ , particularly for objects that are far apart from the foreground. We observe that depth generated from DUSt3R [52] or MAST3R [29] has a strong bias towards smoothness between objects, despite local coherence. Therefore, we handle the outliers through a secondary Semantic Outlier Alignment step.", + "bbox": [ + 89, + 694, + 482, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Semantic Outlier Alignment", + "text_level": 1, + "bbox": [ + 89, + 851, + 333, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key challenge in aligning outliers from $\\chi$ to $\\bar{\\mathbf{X}}$ is that, $\\bar{\\mathcal{O}}$ only provides sparse correspondences to $\\chi$ , which has a 3D", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "point for every pixel. As such, some heuristics need to be developed to group regions of $\\chi$ to move based on $\\bar{\\mathcal{O}}$ . Based on the observation that geometric inconsistencies between $\\chi$ and $\\bar{\\mathbf{X}}$ tend to occur between objects and not within objects, we introduce an Interactive Segmentation Model (ISM) to get a series of masks according to $\\bar{\\mathcal{O}}$ in an iterative manner.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we randomly select a point from the outlier SfM set, $\\bar{\\mathbf{X}}_k\\in \\bar{\\mathcal{O}}$ Using Eq.2),we can get the projected 2D point $\\bar{x}_k$ to prompt ISM, producing a binary mask $m_{k}\\in \\mathbb{R}^{H\\times W}$ delineating a semantically coherent local region around this outlier point. We further define a threshold $T$ , which describes the desired number of correspondences to estimate a local alignment. If the number of outliers within $m_{k}$ is greater than $T$ i.e. $|m_k\\bigcap \\bar{P} (\\bar{\\mathcal{O}})| > T$ then we keep $m_{k}$ as it is and exclude within-mask outliers from being selected in the next iteration. Otherwise, we use all the outliers within the produced mask to prompt ISM again to update $m_{k}$ until the above criterion is met, or no further outliers can be included, in which case we discard this mask and the outliers within. We repeat the above process until no point in $\\bar{\\mathcal{O}}$ is left, eventually obtaining a set of semantically coherent masks around the outliers, $\\{m_k\\}_{k = 1}^M$ .Visualization of this process can be found in the supplementary materials. This iterative process is denoted as", + "bbox": [ + 511, + 181, + 903, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\operatorname{Init}: m_k = \\operatorname{ISM}(x_k)$ , s.t. $\\bar{x}_k \\in \\bar{P}(\\bar{\\mathcal{O}})$ and $\\bar{x}_k \\notin \\bigcup_{j=1}^{k-1} m_j$", + "bbox": [ + 519, + 458, + 895, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, $m_{k} = \\left\\{ \\begin{array}{ll}m_{k}, & \\text{if} |m_{k}\\bigcap \\bar{P} (\\bar{\\mathcal{O}})| > T\\\\ \\operatorname {ISM}(m_{k}\\bigcap \\bar{P} (\\bar{\\mathcal{O}})), & \\text{otherwise} \\end{array} \\right.$ (4)", + "bbox": [ + 514, + 503, + 903, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the outlier masks $\\{m_k\\}_{k=1}^M$ , we extract the within-mask outlier set $\\mathcal{O}_k = P^{-1}(m_k \\bigcap P(\\mathcal{O}))$ and their corresponding SfM set is $\\bar{\\mathcal{O}}_k$ . We then estimate a local rigid transformation within each mask as follows:", + "bbox": [ + 511, + 556, + 903, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {k}, R _ {k}, t _ {k} = \\underset {s, R, t} {\\arg \\min } \\sum_ {(\\bar {\\chi} _ {i}, \\bar {\\mathbf {X}} _ {i}) \\in (\\mathcal {O} _ {k}, \\bar {\\mathcal {O}} _ {k})} \\| s R \\bar {\\chi} _ {i} + t - \\bar {\\mathbf {X}} _ {i} \\| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 633, + 903, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we construct the inlier mask as $m_0 = \\neg \\left( \\bigcup_{k=1}^{M} m_k \\right)$ , where $\\neg$ is the logical NOT operator. Then, the global transformation is applied to MASt3R's points within inlier mask $m_0$ and each individual local transformation is applied to those within each local mask $m_k, k \\neq 0$ . This results in the final SfM-aligned point cloud, $\\chi^*$ , where", + "bbox": [ + 511, + 681, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\chi^ {*} = \\bigcup_ {k = 0} ^ {M} \\left\\{s _ {k} R _ {k} \\chi_ {i} + t _ {k} \\mid \\chi_ {i} \\in m _ {k} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 789, + 903, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the final dense and pose-wise accurate point cloud $\\mathcal{X}$ as prior, we concatenate the final SfM-aligned point cloud with that of SfM, i.e.", + "bbox": [ + 511, + 832, + 903, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} = \\chi^ {*} \\bigcup \\bar {\\mathbf {X}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 882, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Gaussian Optimization", + "text_level": 1, + "bbox": [ + 89, + 90, + 299, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SPARS3R constructed $\\mathcal{X}$ can be directly used as the initialization for any Gaussian Splitting-based optimization method. Here we use Splatfacto, developed under the NeRFStudio framework [49]; the Gaussian optimization loss is:", + "bbox": [ + 89, + 109, + 483, + 185 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\| \\tilde {C} - C \\| _ {1} + \\lambda_ {2} \\mathrm {D} - \\operatorname {S S I M} (\\tilde {C}, C), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 199, + 482, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{C}$ and $C$ denote our rendered and the groundtruth RGB image, D-SSIM denotes the structural similarity loss.", + "bbox": [ + 89, + 232, + 482, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 279, + 223, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments are performed on three popular benchmark datasets. Tanks and Temples [27] contains 8 scenes captured in video format. Following InstantSplat [14], We uniformly sample 24 images with alternating indices for train and test split. MVImgNet [60] is an object-centric dataset that consists of 7 scenes. These scenes are captured from a consistent height covering $180^{\\circ}$ field of view. We apply the same sampling and train-test split procedure as for Tanks and Temples. For sparser settings, we uniformly sample 3 and 6 views from the training set for the above two datasets. Mip-NeRF 360 [2] comprises of 9 scenes with $360^{\\circ}$ views and greater pose variation between the scenes, including diverse heights and distances. The increased variability, combined with sparsity, adds complexity to accurate pose estimation and scene reconstruction. For this dataset, we follow the test set outlined in MipNeRF360 [2] and uniformly sample 12 images from the original training set to construct a sparse-view set.", + "bbox": [ + 89, + 305, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All sparse-view input images go through a COLMAP [46] pipeline for camera calibration. Specifically, we use feature matching results from MASt3R [29] for SfM triangulation. Since sparse-view registration can be unstable due to limited pairs, we perform multiple SfMs and pick the outcome that maximizes successful triangulation per image.", + "bbox": [ + 89, + 579, + 483, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Sparse NVS Evaluation", + "text_level": 1, + "bbox": [ + 89, + 681, + 308, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accurately and holistically evaluating sparse NVS should involve many practical considerations. Various NeRF-based approaches [12, 23, 25, 39, 51, 57, 59] have assumed perfect camera poses, i.e. calibration results from dense views, for sparse-view scenarios, which is unrealistic. Additionally, Gaussian-Splatting-based methods rely heavily on SfM point cloud, further necessitating the separation of train and test views from the registration stage. Previous approaches [65] have tried to perform re-triangulation based on known train poses, excluding test-set point cloud from training, but do not account for pose inaccuracy in sparse-view registration. While InstantSplat [14] adopts the realistic setup where they use recovered poses from sparse", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "views and initialize the test camera poses by applying Procrustes Alignment [22] between train-only poses and train-test-together poses, their large camera alignment error leads to significant disruption in common pixel-based rendering metrics, e.g. PSNR and SSIM [53].", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ca6ca25b7a9bd17a21fcf6799059eb39469dd79dff0b0150a4daec2f6a3a7802.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mip-NeRF 360ER(μ)↓ET(μ)↓
Procrustes Alignment0.1960.0144
+ RANSAC0.1790.0114
+ Rotation Points0.1560.0117
", + "bbox": [ + 537, + 178, + 883, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Improvements over Procrustes Alignment baseline in average rotation error $E_{R}$ and translation error $E_{T}$ . Incorporating rotation points further minimizes the overall error.", + "bbox": [ + 511, + 266, + 906, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also operate under the realistic setup but improve upon the camera alignment process in two aspects. First, during camera alignment, instead of relying only on camera positions, we sample three additional points $P_{\\mathrm{rotation}} \\in \\mathbb{R}^{3 \\times 3}$ per camera along the unit direction vectors in rotation matrix $R_{\\mathrm{camera}} \\in \\mathbb{R}^{3 \\times 3}$ at camera position $T_{\\mathrm{camera}} \\in \\mathbb{R}^{1 \\times 3}$ , i.e.", + "bbox": [ + 511, + 318, + 905, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {r o t a t i o n}} = s _ {\\text {c l o u d}} R _ {\\text {c a m e r a}} + T _ {\\text {c a m e r a}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 422, + 903, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The scale of the camera cloud is denoted by $s_{\\mathrm{cloud}} = \\sqrt{\\sigma_x^2 + \\sigma_y^2 + \\sigma_z^2}$ , and $\\sigma_x, \\sigma_y,$ and $\\sigma_z$ are the standard deviations of camera positions along $x, y,$ and $z$ axes respectively. The four points per camera enable alignment of both camera rotations and translations. Second, we add RANSAC [15] in the alignment process, minimizing the influence of outliers. We measure the alignment rotation error $E_R$ by the angular difference in degrees between two rotations represented by quaternions following", + "bbox": [ + 511, + 445, + 906, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {R} = 2 \\cdot \\arccos \\left(\\frac {\\operatorname {a b s} \\left(q _ {1} \\cdot q _ {2}\\right)}{\\| q _ {1} \\| \\cdot \\| q _ {2} \\|}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 597, + 903, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the translation error $E_{T}$ by the Euclidean norm $\\| \\cdot \\|$ between translation vectors. As shown in Tab. 1, these two improvements enhance camera alignment accuracy in both rotation and translation. Beyond accurate camera pose alignment, test pose optimization can also be applied between rendered and ground-truth images to minimize the pose error; however, such process is time-consuming and can get stuck if the initial displacement is too large, as is sometimes the case in InstantSplat [14].", + "bbox": [ + 511, + 640, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. In addition to the commonly used PSNR, SSIM, and LPIPS, we incorporate DreamSim [18] (DSIM) as an additional metric to assess rendering quality. DSIM captures human perceptual similarity by finetuning a combination of embeddings from visual foundation models based on human evaluations. Notably, it provides a robust assessment of rendering image quality under moderate pose shift, which frequently occurs in the realistic sparse-view", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26814", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "settings. Further details on DSIM's robustness as an evaluation metric are provided in the supplementary materials.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We employ Segment-Anything Model [26] as the segmentation model in Semantic Outlier Alignment. The weights for the loss terms are $\\lambda_{1} = 0.8$ , $\\lambda_{2} = 0.2$ . For fair implementation and comparison, we employ test pose optimization for all baselines and SPARS3R for 500 steps to maximally remove the effect of shifted camera pose. All baselines are run until convergence.", + "bbox": [ + 89, + 125, + 483, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 241, + 256, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the different components in SPARS3R to demonstrate their effectiveness on the Mip-NeRF 360 [2] dataset:", + "bbox": [ + 89, + 263, + 482, + 306 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SfM-initialized 3DGS: Optimizing 3DGS on a sparse SfM point cloud and SfM-estimated camera poses.", + "- MAST3R-initialized 3DGS: Optimizing 3DGS on a MAST3R-produced dense point cloud and camera poses.", + "- Global Fusion Alignment: Aligning the MAST3R-produced dense point cloud onto SfM point cloud with a global transform. Poses are taken from SfM.", + "- Semantic Outlier Alignment (SPARS3R): The full SPARS3R pipeline that aligns outliers through their 2D semantic masks." + ], + "bbox": [ + 89, + 311, + 482, + 460 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/de43c8f40daaae813052dff7f868ec86652dd3b7317bb08ddd8470ac310f739f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingPSNRSSIMLPIPSDSIM
Initialization: SfM [46]
3DGS [24]16.60.3880.4580.232
Initialization: MASt3R [29]
3DGS [24]15.90.2930.4630.229
+ Global Fusion Alignment18.60.4860.3300.130
+ Semantic Outlier Alignment18.90.5000.3270.127
", + "bbox": [ + 94, + 474, + 478, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, given only 12 images, sparse NVS is challenging both in registration and reconstruction. Specifically, we observe that the 3DGS based on SfM initialization performs better than MASt3R initialization. Clearly, render quality is affected if poses are inaccurate, despite locally coherent dense point cloud. The introduction of Global Fusion Alignment improves upon the naive baselines from either initializations by combining accurate pose from SfM and dense point cloud from MASt3R. For scenes that do not have significant depth discrepancies, Global Fusion Alignment works well. In scenarios such as the Bonsai scene in Fig. 3, MASt3R's relative depth between the table and the wall is inaccurate; therefore, Global Fusion Alignment alone leads to a broken geometry of the background. In comparison, our Semantic Outlier Alignment provide piece-wise alignments on the background to ad", + "bbox": [ + 89, + 657, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dress this relative discrepancy. Quantitatively, we observe a 1.4 dB improvement on the Bonsai scene.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e2051c47bc81a21888a4eea47eb0604d435477f42df0165a518cb9514b3974e.jpg", + "image_caption": [ + "Global Fusion Alignment (GFA)" + ], + "image_footnote": [], + "bbox": [ + 524, + 132, + 707, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12362d3bf7151a9c2b1a262546a2ff5fd6b69d04d0fa6ba9f1aaf46adc241f3b.jpg", + "image_caption": [ + "Semantic Outlier Align (SOA)" + ], + "image_footnote": [], + "bbox": [ + 722, + 132, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c3da074c43d318456d1be6921fc64e0292bc955fd3ece68d7d31be218fa5237.jpg", + "image_caption": [ + "Aligned point cloud from GFA" + ], + "image_footnote": [], + "bbox": [ + 524, + 252, + 704, + 342 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/45a32b2e951d136f28955f084c97630f57b35dd89b2060dbe69069b25095ed9a.jpg", + "image_caption": [ + "Aligned point cloud from SOA", + "Figure 3. Visual comparison of SPARS3R with and without SOA. While the dense bonsai in the foreground is aligned with the sparse point cloud, depth differences are obvious. SOA successfully fixes such gap." + ], + "image_footnote": [], + "bbox": [ + 720, + 247, + 900, + 345 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d5824ce43f7b99f389d0dd233a64e4ccea17f8aea3cfc340f11962075bd43d3.jpg", + "table_caption": [ + "Table 2. Ablation on key components of SPARS3R. The results are shown using PSNR, SSIM, LPIPS and DSIM based on the MipNeRF 360 [2]." + ], + "table_footnote": [], + "table_body": "
MethodsMip-NeRF 360Tanks & TemplesMVimgNet
RPEtRPErRPEtRPErRPEt RPEr
DUSt3R [52]2.0752.5840.5700.1430.438 0.421
MASt3R [29]1.1861.4900.2410.2480.208 0.304
InstantSplat [14]2.0492.5550.1510.0810.264 0.311
COLMAP + MASt3R0.2520.4120.1610.0930.075 0.078
", + "bbox": [ + 514, + 458, + 916, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative evaluation of pose accuracy across three datasets, Relative Translation Error $(\\mathrm{RPE}_t)$ and Relative Rotation Error $(\\mathrm{RPE}_r)$ [62] are calculated based on the normalized poses.", + "bbox": [ + 511, + 551, + 906, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Estimation Accuracy. As shown in Tab. 3, we compare the pose estimation accuracy from different calibration approaches. Both DUSt3R [52] and MASt3R [29] have significantly larger errors compared to COLMAP [46] with MASt3R feature matching outputs. InstantSplat [14] uses DUSt3R's [52] dense point cloud and pose estimation and attempts to improve accuracy through a training pose optimization approach similar to BARF [32]. While it brings down the errors in some cases, such training pose optimization strategy does not work as well in more challenging datasets like Mip-NeRF 360. This highlights the difficulty in improving camera calibration jointly with dense forward render, considering the loss landscape is highly nonlinear based on pixel-wise metrics.", + "bbox": [ + 511, + 613, + 906, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative and Visual Evaluation", + "text_level": 1, + "bbox": [ + 511, + 833, + 821, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We quantitatively compare SPARS3R against various NVS methods in Tab. 4 and Tab. 5. Instant-NGP [36] and 3DGS [24] are two mainstream NVS methods based on", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26815", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc8c04d074362b6fc39fd011076c17c8d971425a296b331987d1eee29791c935.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsMip-NeRF 360 [2]Tanks & Temples [27]MVImgNet [60]
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
Instant-NGP [36]14.820.2940.6780.52415.280.4510.3890.25413.280.4260.8920.892
3DGS [24]16.570.3880.4580.23221.070.7300.1940.07021.240.6730.2340.064
FSGS [65]17.600.4430.5580.24325.720.8450.1110.02323.430.7600.2120.042
SparseGS [55]16.660.4050.4610.21020.280.7270.2020.07520.560.6720.2480.072
DRGS [10]16.880.4010.6490.30021.460.7230.2890.07821.700.6410.4220.071
CF-3DGS [19]13.270.2500.6980.50918.990.6060.2960.12715.430.4080.5450.325
InstantSplat [14]16.230.3590.5430.23326.970.8740.1150.01323.220.7340.2480.028
SPARS3R18.850.5000.3270.12729.900.9190.0470.00725.850.8200.1140.011
", + "bbox": [ + 99, + 88, + 898, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f37aa4e4c6bccf1c26b03a62c5ec86a66281419587f10f4dba4abede8bbc4f65.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of different NVS methods on 12 views on three popular benchmark datasets, totaling 24 scenes. All methods are run on the same registrations and updated with test pose optimization." + ], + "table_footnote": [], + "table_body": "
3 ViewsTanks & TemplesMVImgNet
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
FSGS20.360.6580.2190.07716.920.5120.4760.183
InstantSplat21.650.7550.1640.04416.290.5510.3230.103
SPARS3R21.830.7550.1500.04318.620.5880.2810.091
6 Views
FSGS23.900.7830.1520.04320.420.6440.3310.080
InstantSplat25.890.8470.1170.02221.090.6950.2360.046
SPARS3R26.800.8730.0730.01422.750.7320.1610.023
", + "bbox": [ + 94, + 294, + 488, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of 3 and 6 views on Tanks & Temples and MVImgNet datasets.", + "bbox": [ + 89, + 430, + 482, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "implicit and explicit scene representations. Specifically, 3DGS leverages SfM points from registration as the starting point for scene optimization, while NeRF [35] based methods rely only on poses. Colmap-Free 3DGS [19] also does not rely on prior registration and optimizes cameras along with reconstruction. We find that both Instant-NGP and CF-3DGS perform significantly worse than other methods that use a prior SfM point cloud.", + "bbox": [ + 88, + 476, + 482, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among methods that are designed for sparse NVS, DRGS [11], SparseGS [55], and FSGS [65] leverage monocular depth prior to gain more geometric information of the scene. DRGS applies an additional smoothness constraint, which leads to high PSNR values but worse LPIPS and DSIM performances. SparseGS additionally applies a Score Distillation Sampling Loss, which generates novel views and constrained them through a prior generative model. Its results are sharper than DRGS, but more noisy, leading to worse PSNR. FSGS seeks to address the sparse point cloud issue in sparse-view NVS by modifying the densification algorithm in 3DGS. More Gaussians are encouraged between the space of two existing Gaussians, allowing representative areas to densify quicker. FSGS has overall better metrics than SparseGS and DRGS; however, it still encounters issues in regions where few initializations are available, e.g. the background of the scene. As shown in Fig. 4 and in Fig. 1, this leads to various blurry regions. InstantSplat [14] is a concurrent work that uses prior depth estimation from DUSt3R [52] and the dense point cloud to help with sparse-", + "bbox": [ + 91, + 599, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "view NVS. While the initialized point cloud is much denser than FSGS, InstantSplat also inherits the suboptimal poses produced by DUSt3R. The incorrect geometry limits the effectiveness of densification in 3DGS and can lead to generation of floaters; to ameliorate this problem, InstantSplat disables Gaussian densification and optimizes the scene with few steps. As shown in Tab. 3 and Tab. 4, on the Mip-NeRF 360 and MVImgNet dataset, where pose accuracy from DUSt3R is limited, InstantSplat performs on-par with FSGS despite using a denser prior point cloud. As shown in Fig. 4, we observe that the Gaussians tend to be noisy on the surface. In the Stump and Bicycle scenes, InstantSplat's results cannot be rigidly aligned well with the groundtruth test poses due to the large offset to begin with.", + "bbox": [ + 511, + 297, + 906, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our SPARS3R method improves upon DUSt3R and MASt3R's prior point cloud by adjusting it based on conventional SfM calibration approaches. Since SfM only triangulates based on confident keypoints, its pose and depth estimation performance, particularly under sparse-view scenario, is more reliable. As shown in Tab. 4, SPARS3R consistently improves upon previous best approach by an average of $2.7\\mathrm{dB}$ . SPARS3R is stable in its optimization process and does not generate additional floaters with more iterations. As seen in Fig. 3, SPARS3R produces rendering with smooth surfaces in the foreground, and sharp details in the background.", + "bbox": [ + 511, + 515, + 908, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For even sparser settings, we provide a quantitative comparison in Tab. 5 between SPARS3R and the two best-performing methods from Tab. 4, FSGS [65] and InstantSplat [14]. FSGS struggles significantly under extreme sparsity due to its reliance on a sparse initialization, resulting in DSIM scores that are approximately two- and four-times worse than those of SPARS3R under the 3-view and 6-view settings, respectively. InstantSplat generally performs better than FSGS, benefiting from a much denser point cloud initialization produced by DUSt3R [52], which is particularly advantageous under extreme sparse conditions. However, its performance remains inferior to SPARS3R due to suboptimal pose estimation. Our", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26816", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afb8b32a21e1c4cbe6e7f63dd2fa11c336b835bcbf35a8b69a5355b9b84bf8a5.jpg", + "image_caption": [ + "Figure 4. Visual comparisons of different NVS methods on 12 views on Mip-NeRF 360 [2] dataset. Zooming in on the visualizations is recommended to show differences in detail. More visualizations for other datasets are available in the Supplemental Material." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 916, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SPARS3R continues to outperform these methods, improving upon the best approach by an average of 0.9 dB and 1.3 dB in the 3-view and 6-view settings, respectively. These results further demonstrate SPARS3R's superiority over state-of-the-art approaches in handling extreme sparse-view reconstruction. For more visualizations and flythroughs of the scene, please refer to the supplemental materials.", + "bbox": [ + 89, + 529, + 483, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 89, + 650, + 217, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While SPARS3R significantly improves upon previous SoTA, there are also several limitations worth noting. Firstly, SPARS3R relies on a well-performing semantic segmentation model to group outliers; as such, if the segmentation model is very sensitive, i.e. separating a connected surface into multiple disjointed parts, some of these parts may lack sufficient SfM support for alignment. If the segmentation model produces masks that group multiple areas with disparate depth together, the local alignments are also less effective similar to the global alignment. Secondly, non-rigid transform between two point clouds are worth further investigation. While SPARS3R's piece-wise rigid transform based on semantic relationship is fast and addresses the inaccurate inter-object depth estimation, a non-rigid formulation may be more generalizable.", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 527, + 632, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SPARS3R, a scene reconstruction and NVS method that can achieve high-quality rendering with sparse input images. We demonstrate that existing methods lead to blurry rendering due to sparse point cloud initialization. Recent progress in generating a dense point cloud from prior depth estimation model, while promising, can lead to noisy pose estimation. To address these issues, SPARS3R combines the advantages of both by proposing a two-step alignment approach. The first step computes a global transformation between the dense point cloud from depth prior and a reference SfM point cloud. The second step takes the semantic regions around the outlier points identified from the first step and performs successive local alignments. This approach successfully addresses the depth estimation errors within the prior dense point cloud and demonstrates that the updated point cloud leads to significantly better performances in sparse-view NVs. We also introduce several improvements in the evaluation process to better represent the practical limitations in sparse-view registration and reconstruction. In the future, we hope to further improve the smoothness of our alignment process, e.g. by exploring potential non-rigid transformation approaches.", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26817", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD.", + "bbox": [ + 89, + 114, + 485, + 310 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 90, + 609, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 5, 6, 7, 8", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2", + "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2", + "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2", + "[6] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19436-19446, 2024. 2", + "[7] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2", + "[8] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12952, 2022. 2", + "[9] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2", + "[10] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 7", + "[11] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 2, 7", + "[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 516, + 114, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26818", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5", + "[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\\mathrm{x}$ reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2", + "[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 1, 2, 5, 6, 7", + "[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5", + "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2", + "[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2", + "[18] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data, 2023. 5", + "[19] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 7", + "[20] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021. 2", + "[21] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2", + "[22] John C Gower. Generalized procrustes analysis. Psychometrika, 40:33-51, 1975. 3, 5", + "[23] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 5", + "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6, 7", + "[25] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infererf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022. 1, 5" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 6", + "[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7", + "[28] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719-21728, 2024. 2", + "[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 4, 5, 6", + "[30] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 1", + "[31] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2", + "[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 6", + "[33] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2", + "[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 2", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 7", + "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1, 2, 6, 7", + "[37] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26819", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349-10358, 2024. 2", + "[39] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1, 5", + "[40] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatt-ting for robust real-time rendering with $900+$ fps. arXiv preprint arXiv:2403.13806, 2024. 2", + "[41] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In European Conference on Computer Vision, pages 19-37. Springer, 2025. 2", + "[42] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2", + "[43] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In European Conference on Computer Vision, pages 293-310. Springer, 2024. 2", + "[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2", + "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2", + "[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5, 6", + "[47] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2", + "[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5", + "[50] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2", + "[51] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1, 5", + "[52] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7", + "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[54] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2", + "[55] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 $\\{\\backslash \\mathrm{deg}\\}$ sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 1, 2, 7", + "[56] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 2", + "[57] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 1, 5", + "[58] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15901-15911, 2023. 2", + "[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 5", + "[60] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26820", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 5, 7", + "[61] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2", + "[62] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual(-inertial) odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7244-7251, 2018. 6", + "[63] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2", + "[64] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2", + "[65] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1, 2, 5, 7" + ], + "bbox": [ + 91, + 92, + 482, + 445 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "26821", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_model.json b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f9209082a07bf6b9eedc0ab7f92b0a5e84daa33 --- /dev/null +++ b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_model.json @@ -0,0 +1,2644 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.861, + 0.175 + ], + "angle": 0, + "content": "SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.204, + 0.34, + 0.222 + ], + "angle": 0, + "content": "Yutao Tang*" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.204, + 0.489, + 0.222 + ], + "angle": 0, + "content": "Yuxiang Guo*" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.204, + 0.619, + 0.222 + ], + "angle": 0, + "content": "Deming Li" + }, + { + "type": "text", + "bbox": [ + 0.658, + 0.204, + 0.758, + 0.222 + ], + "angle": 0, + "content": "Cheng Peng" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.226, + 0.603, + 0.243 + ], + "angle": 0, + "content": "Johns Hopkins University" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.245, + 0.678, + 0.261 + ], + "angle": 0, + "content": "{ytang67，yguo87，dli90，cpeng26}@jhu.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.295, + 0.327, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.486, + 0.645 + ], + "angle": 0, + "content": "Recent efforts in Gaussian-Splat-based Novel View Synthesis can achieve photorealistic rendering; however, such capability is limited in sparse-view scenarios due to sparse initialization and over-fitting floaters. Recent progress in depth estimation and alignment can provide dense point cloud using few views; however, the resulting pose accuracy is suboptimal. In this work, we present SPARS3R, which combines the advantages of accurate pose estimation from Structure-from-Motion and dense point cloud from depth estimation. To this end, SPARS3R first performs a Global Fusion Alignment process that maps a prior dense point cloud to a sparse point cloud from Structure-from-Motion based on triangulated correspondences. RANSAC is applied during this process to distinguish inliers and outliers. SPARS3R then performs a second, Semantic Outlier Alignment step, which extracts semantically coherent regions around the outliers and performs local alignment in these regions. Along with several improvements in the evaluation process, we demonstrate that SPARS3R can achieve photorealistic rendering with sparse images and significantly outperforms existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.222, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.848 + ], + "angle": 0, + "content": "Photorealistic scene reconstruction and Novel View Synthesis (NVS) from unposed 2D images is a challenging task with numerous applications in site modeling, autonomous driving, robotics, urban and agricultural planning, etc. The introduction of methods such as Neural Radiance Field (NeRF) [35] and 3D Gaussian Splatting (3DGS) [24] have introduced significant advances in rendering quality and efficiency based on dense multi-view imagery. However, applications of these methods still face issues in practical scenarios where a dense coverage of the scene is not available." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Given sparse images, over-fitting the photometric objectives to an incorrect geometry is a common issue in" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.295, + 0.906, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.395, + 0.662, + 0.407 + ], + "angle": 0, + "content": "Instant-NGP [36]" + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.395, + 0.837, + 0.407 + ], + "angle": 0, + "content": "FSGS [65]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.411, + 0.906, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.511, + 0.66, + 0.523 + ], + "angle": 0, + "content": "InstantSplat [14]" + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.511, + 0.836, + 0.522 + ], + "angle": 0, + "content": "SPARS3R" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Figure 1. A visualization of SPARS3R in comparison to current SoTA. Without additional prior, sparse NVS leads to incorrect geometry by Instant-NGP [36]. FSGS [65] can be blurry due to sparse initialization and insufficient densification. InstantSplat [14] relies on DUSt3R [52] initialization with suboptimal poses. Our method, SPARS3R, can reliably render details in the foreground and background with accurate poses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.902 + ], + "angle": 0, + "content": "NVS. Various constraints, such as semantic consistent loss [23], depth and novel-view regularization [12, 39, 51], frequency regularization [57], and ray entropy minimization [25] have been introduced to improve upon NeRF. These methods often lead to significant computation overhead due to the costly patch rendering from ray-tracing. More recently, Gaussian-Splatting-based approaches have further improved sparse NVS by leveraging the explicit representation and fast differentiable rasterization. Depth regularization [30], Gaussian floater pruning [55], and proximity-based Gaussian densification strategy [65] have been proposed to constrain and guide the scene structures. As demonstrated in Fig. 1, despite improved geometry, these approaches tend to produce overly smooth rendering in the background. This issue can be attributed to a sparse initial" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26810" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "point cloud, particularly in the background regions. Furthermore, the additional constraints, e.g. based on monocular depth prior, are oftentimes imperfect, leading to noisy gradients that prevent proper densification in sparse regions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.483, + 0.47 + ], + "angle": 0, + "content": "One potential solution to address this conundrum is to initialize Gaussian optimization with a denser point cloud to help disambiguate shape and radiance. To this end, recent advances in two-view depth estimation, particularly DUSt3R [52] and MAST3R [29], have shown the impressive ability to construct a dense point cloud from a pre-trained prior model. In theory, such a point cloud can be directly used to optimize a Gaussian-Splat radiance field [14]. In practice, camera calibration obtained from multi-view depth alignment is often suboptimal due to the difficulties in estimating an accurate depth map. If left uncorrected, the Gaussian optimization process will generate floaters to compensate the suboptimal calibration, prompting strategies to prevent densification as a stopgap measure [14]. In comparison, camera calibration based on Structure-from-Motion (SfM), e.g., COLMAP [46], is fast and accurate with an appropriate feature matching model. Instead of aligning dense depth, SfM takes confident correspondences and performs triangulation with RANSAC [15] to reject outliers. In a sparse-view setting, such a process can be repeated numerous times to further improve accuracy." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.472, + 0.483, + 0.804 + ], + "angle": 0, + "content": "To address sparse point cloud initialization and pose inaccuracy in sparse-view NVS, we propose SPARS3R. SPARS3R composes of two stages: Global Fusion Alignment and Semantic Outlier Alignment. In Global Fusion Alignment, SPARS3R first obtains a dense point cloud from sparse views through DUSt3R [52] or MASt3R [29], and a sparse point cloud from COLMAP [46]. By taking the triangulated correspondences within images, SPARS3R fuses the dense point cloud onto the sparse point cloud through a global Procrustes Alignment process with RANSAC [15]. Since some points will yield large error due to local scale variations, a second Semantic Alignment process is introduced to perform local alignment. Specifically, outliers from Global Fusion Alignment are identified and used as prompts for an Interactive Segmentation model, e.g. SAM [26], generating semantically coherent masks around the outliers. The resulting semantic masks indicate regions within dense point cloud to be treated with local alignment. By locally aligning the identified outliers with the SfM point cloud while preserving the globally aligned inliers, we obtain a dense and pose-wise accurate point cloud, serving as a strong prior for Gaussian optimization." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.806, + 0.465, + 0.82 + ], + "angle": 0, + "content": "In summary, our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.483, + 0.884 + ], + "angle": 0, + "content": "1. We propose a Global Fusion Alignment approach, which transforms a prior dense point cloud onto a reference SfM sparse point cloud, putting dense initialization and accurate camera poses in the same coordinate frame." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2. To address outliers that cannot be aligned accurately due" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "to depth discrepancies, we propose a Semantic Outlier Alignment step. This step extracts semantically similar regions around the outliers to perform local alignment, resulting in a dense point cloud with minimum transformation error." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "3. We evaluate the overall method, SPARS3R, on three popular benchmark datasets and find significant quantitative and visual improvements compared to current SoTA methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.234, + 0.655, + 0.249 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.256, + 0.858, + 0.272 + ], + "angle": 0, + "content": "2.1. 3D Models for Synthesizing Novel Views" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.278, + 0.905, + 0.535 + ], + "angle": 0, + "content": "Photorealistic scene reconstruction and novel view synthesis is a long-standing task in computer vision and graphics. Neural Radiance Field (NeRF) [35] proposes to model a scene implicitly with multi-layer perceptron (MLP) by mapping 3D coordinates and view direction to color and density, then rendering pixel value via alpha blending. Since NeRF's introduction, numerous work has sought to improve its efficiency [5, 16, 20, 36, 45, 48], quality [1-3, 7, 50], and extension to dynamic and in-the-wild [4, 6, 8, 17, 34, 42, 44, 47, 54, 58] scenarios. Recently, 3D Gaussian Splatting (3DGS) [24] has emerged as an efficient alternative, offering improved rendering quality over NeRF. Unlike NeRF, 3DGS represents scenes with explicit 3D Gaussians kernels and uses a differentiable rasterization technique. Following its introduction, many methods have emerged to reduce the computational cost [13, 21, 28, 37, 38] and enhance quality [9, 33, 40, 43, 56, 61, 63, 64] of 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.887, + 0.555 + ], + "angle": 0, + "content": "2.2. Gaussian-Splatting-based Sparse-View NVs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.561, + 0.905, + 0.818 + ], + "angle": 0, + "content": "While 3DGS performs well with dense view support, its effectiveness decreases in practical settings with sparse inputs. Many approaches address this through depth constraints. FSGS [65] uses a Proximity-guided Gaussian Unpooling to mitigate the sparse initialization issue and introduces pseudo-views during training to avoid overfitting. DRGS [11] and SparseGS [55] add a depth regularization term to enforce consistency between estimated and monocular depths while promoting smoothness. DNGS [31] refines depth regularization by prioritizing local depth variations. CoherentGS [41] initializes Gaussians from monocular depth estimation and improves coherence through an optical flow constraint. InstantSplat [14] is a concurrent work that directly uses DUSt3R to produce dense point clouds and introduces a camera pose optimization strategy; densification of Gaussians is disabled to prevent the introduction of floaters due to suboptimal poses." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.827, + 0.612, + 0.842 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.849, + 0.642, + 0.865 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "DUSt3R [52] is a two-view depth estimation method that produces dense 3D point clouds from image pairs. Specif" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26811" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.298, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 2. SPARS3R combines a prior dense point cloud \\(\\chi\\) and a sparse SfM point cloud \\(\\bar{\\mathbf{X}}\\). The prior \\(\\chi\\) often has inferior depth accuracy compared to \\(\\bar{\\mathbf{X}}\\). SPARS3R first globally aligns all points in \\(\\chi\\) onto \\(\\bar{\\mathbf{X}}\\), based on shared correspondences. Inliers and outliers are identified through alignment error. SPARS3R then extracts the semantically relevant 2D regions around the outliers to move local regions of \\(\\chi\\) in groups, producing a dense point cloud \\(\\chi^{*}\\) that is depth-wise and pose-wise accurate. By providing \\(\\mathcal{X} = \\chi^{*} \\bigcup \\bar{\\mathbf{X}}\\) for Gaussian optimization, SPARS3R achieves photorealistic rendering under sparse-view condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.606 + ], + "angle": 0, + "content": "ically, given an image pair \\(I^1\\) and \\(I^2\\), DUSt3R generates dense pointmaps \\(\\mathbf{X}^{1,1}\\) and \\(\\mathbf{X}^{2,1} \\in \\mathbb{R}^{H \\times W \\times 3}\\). A pointmap, e.g. \\(\\mathbf{X}^{1,1}\\), provides a mapping from each pixel \\((u, v)\\) in image \\(I^1\\) to a corresponding 3D point \\(\\mathbf{X}^{1,1}(u, v)\\) under the coordinate system of image \\(I^1\\), giving dense 2D-to-3D correspondences. Given \\(K > 2\\) input images, DUSt3R [52] aggregates across all pairwise pointmap predictions by globally aligning pairwise pointmaps into a unified point cloud \\(\\chi\\). This alignment process also produces camera parameters \\(\\{W(I^n), J(I^n)\\}\\), which are the world-to-camera transformation matrix and the Jacobian of the affine approximation of the projective transformation for view \\(I^n\\). As such, the relationship between a pixel \\((u, v)\\) on \\(I^n\\) and its 3D point \\(\\chi_i\\) is established via a projective function \\(P\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.619, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\nx _ {i} = [ u, v ] ^ {T} = P \\left(\\chi_ {i}\\right) = J \\left(I ^ {n}\\right) W \\left(I ^ {n}\\right) \\chi_ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Building on this, MASt3R [29] incorporates an additional matching head that generates discriminative local feature descriptors and pixel-level correspondences across paired images. We note that while MASt3R [29] uses this feature correspondence during its multi-depth alignment, the estimated poses are oftentimes inaccurate due to smooth depth predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.206, + 0.772 + ], + "angle": 0, + "content": "3.2. SPARS3R" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As show in Fig. 2, SPARS3R builds upon the advancement in DUSt3R [52] and MASt3R [29] as a pre-trained 3D prior for sparse-view reconstruction. Firstly, SPARS3R performs SfM based on image correspondences, either from MASt3R [29] or other feature matching methods. SPARS3R then aligns the dense point cloud produced by MASt3R via rigid transformations in two stages: Global Fusion Alignment and Semantic Outlier Alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.736, + 0.41 + ], + "angle": 0, + "content": "3.2.1. Global Fusion Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.705 + ], + "angle": 0, + "content": "While DUSt3R [52] and MASt3R [29] can generate dense 3D points from sparse images, it does not estimate accurate or detailed depth through its pointmap formulation. As a result, the camera calibration obtained from such smooth depths is suboptimal and will degrade the fidelity of scene reconstructions. To construct a better point cloud prior, we propose to align MASt3R's point cloud with that from a SfM pipeline, which is more reliable based on selective correspondences. This alignment process begins by identifying matching points between the SfM and MASt3R point clouds, denoted as \\(\\overline{\\mathbf{X}}\\in \\mathbb{R}^{\\bar{N}\\times 3}\\) and \\(\\chi \\in \\mathbb{R}^{N\\times 3}\\), respectively. A visibility indicator \\(V:\\mathbb{R}^{\\bar{N}\\times 3}\\to \\mathbb{R}^{\\widetilde{N}\\times 3}\\) denotes if an SfM point \\(\\overline{\\mathbf{X}}_i\\) is visible with respect to a view \\(I^n\\), and \\(\\tilde{N}\\) denotes the number of visible SfM points in view \\(I^n\\). For each view, similar to Eq. (1), we can obtain the pixel-to-3D relationship of \\((\\bar{u},\\bar{v})\\) and \\(V(\\overline{\\mathbf{X}})_i\\) through estimated \\(\\overline{W}(I^n)\\) and \\(\\bar{J}(I^n)\\) from SfM. Subsequently, the correspondence between a SfM point \\(V(\\overline{\\mathbf{X}})_i\\) and a point in \\(\\chi\\), denoted as \\(\\tilde{\\chi}_i\\), can be established through a forward-backward projection:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.726, + 0.905, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\bar {x} _ {i} = \\left[ \\bar {u}, \\bar {v} \\right] ^ {T} = \\bar {P} (\\bar {\\mathbf {X}}) _ {i} = \\bar {J} (I ^ {n}) \\bar {W} (I ^ {n}) V (\\bar {\\mathbf {X}}) _ {i}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.748, + 0.841, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\bar {\\chi} _ {i} = P ^ {- 1} (\\bar {x} _ {i}) = W ^ {- 1} \\left(I ^ {n}\\right) J ^ {- 1} \\left(I ^ {n}\\right) \\bar {x} _ {i}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For each view, we denote the ordered set of visible SfM points as \\(\\{V(\\bar{\\mathbf{X}})_i\\}_{i = 1}^N\\), the projected 2D pixels as \\(\\{\\bar{x}_i\\}_{i = 1}^N\\), and the set of 3D points that correspond to the projected 2D pixels above in \\(\\chi\\) as \\(\\{\\bar{\\chi}_i\\}_{i = 1}^N\\), omitting view indicator \\(n\\). With these identified correspondences, we can apply Generalized Procrustes Analysis [22] to estimate a global rigid transformation, optimizing for scale, rotation, and translation. During this process, RANSAC is applied to filter out-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26812" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "liers, excluding their influence on the estimation. Specifically, we minimize the following objective to get the optimal global transformation based on Algorithm 1:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.159, + 0.483, + 0.201 + ], + "angle": 0, + "content": "\\[\ns _ {0}, R _ {0}, t _ {0} = \\underset {s, R, t} {\\arg \\min } \\sum_ {i = 1} ^ {\\widetilde {N}} e \\left(\\bar {\\chi} _ {i}, s, R, t, V (\\bar {\\mathbf {X}}) _ {i}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.204, + 0.438, + 0.224 + ], + "angle": 0, + "content": "\\[\ne (\\bar {\\chi} _ {i}, s, R, t, V (\\bar {\\mathbf {X}}) _ {i}) = \\left\\| s R \\bar {\\chi} _ {i} + t - V (\\bar {\\mathbf {X}}) _ {i} \\right\\| ^ {2}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.247, + 0.325, + 0.263 + ], + "angle": 0, + "content": "Algorithm 1 RANSAC Alignment" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.318 + ], + "angle": 0, + "content": "1: Input: Matched MASt3R points \\(\\{\\overline{\\chi}_i\\}_{i=1}^N\\) and SfM points \\(\\{V(\\bar{\\mathbf{X}})_i\\}_{i=1}^N\\), sample size \\(n\\), error threshold \\(\\epsilon\\), number of iterations \\(R\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.32, + 0.483, + 0.35 + ], + "angle": 0, + "content": "2: Output: Best parameters \\((s_0,R_0,t_0)\\), inlier set \\(\\mathcal{I}\\), outlier set \\(\\mathcal{O}\\), corresponding outlier SfM set \\(\\bar{\\mathcal{O}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.35, + 0.483, + 0.383 + ], + "angle": 0, + "content": "3: Initialize: num_inliers \\(\\leftarrow 0\\), \\((\\hat{s_0},\\hat{R_0},\\hat{t_0})\\gets \\emptyset ,\\hat{\\mathcal{I}}\\gets \\emptyset ,\\hat{\\mathcal{O}}\\gets \\emptyset ,\\hat{\\bar{\\mathcal{O}}}\\gets \\emptyset\\) Iterations \\(\\leftarrow 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.384, + 0.284, + 0.396 + ], + "angle": 0, + "content": "4: while Iterations \\( < R \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.396, + 0.475, + 0.415 + ], + "angle": 0, + "content": "5: Randomly select \\( n \\) points as sample set \\( S \\subset \\{i\\}_{i=1}^{\\widetilde{N}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.446 + ], + "angle": 0, + "content": "6: Estimate \\((\\hat{s}_0,\\hat{R}_0,\\hat{t}_0)\\) on sample set \\(\\mathcal{S}\\) by minimizing \\(\\sum_{i\\in \\mathcal{S}}e(\\bar{\\chi}_i,s,R,t,V(\\bar{\\mathbf{X}})_i)\\) (Eq.3))" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.446, + 0.274, + 0.46 + ], + "angle": 0, + "content": "7: for each point \\(\\overline{\\chi}_i\\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.46, + 0.392, + 0.475 + ], + "angle": 0, + "content": "8: if \\(e(\\overline{\\chi}_i, \\hat{s}_0, \\hat{R}_0, \\hat{t}_0, V(\\overline{\\mathbf{X}})_i) < \\epsilon\\) then" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.476, + 0.312, + 0.49 + ], + "angle": 0, + "content": "9: Add \\(\\bar{\\chi}_i\\) to inlier set \\(\\hat{\\mathcal{I}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.491, + 0.182, + 0.504 + ], + "angle": 0, + "content": "10: else" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.505, + 0.483, + 0.537 + ], + "angle": 0, + "content": "11: Add \\(\\bar{\\chi}_i\\) to outlier set \\(\\hat{\\mathcal{O}}\\) and add \\(V(\\overline{\\mathbf{X}})_i\\) to outlier SfM set \\(\\hat{\\mathcal{O}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.539, + 0.197, + 0.551 + ], + "angle": 0, + "content": "12: end if" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.555, + 0.19, + 0.566 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.569, + 0.305, + 0.583 + ], + "angle": 0, + "content": "14: if \\( |\\mathcal{I}| > \\text{num\\_inliers} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.584, + 0.483, + 0.617 + ], + "angle": 0, + "content": "15: Update: num_inliers \\(\\leftarrow |\\hat{\\mathcal{I}}|, (s_0, R_0, t_0) \\leftarrow (\\hat{s}_0, \\hat{R}_0, \\hat{t}_0), \\mathcal{I} \\leftarrow \\hat{\\mathcal{I}}, \\mathcal{O} \\leftarrow \\hat{\\mathcal{O}}, \\bar{\\mathcal{O}} \\leftarrow \\hat{\\bar{\\mathcal{O}}}" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.619, + 0.181, + 0.63 + ], + "angle": 0, + "content": "16: end if" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.633, + 0.312, + 0.646 + ], + "angle": 0, + "content": "17: Iterations \\(\\leftarrow\\) Iterations \\(+1\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.649, + 0.19, + 0.66 + ], + "angle": 0, + "content": "18: end while" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.663, + 0.309, + 0.677 + ], + "angle": 0, + "content": "19: return \\((s_0, R_0, t_0), \\mathcal{I}, \\mathcal{O}, \\mathcal{O}\\)" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.269, + 0.483, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.846 + ], + "angle": 0, + "content": "After this, we extract an inlier set \\(\\mathcal{I}\\), an outlier set \\(\\mathcal{O}\\) and a corresponding outlier SfM set \\(\\bar{\\mathcal{O}}\\) from RANSAC with an empirical threshold. As shown in Fig. 2, the global transformation recovered from Eq. 3 often leaves large amount of outliers between \\(\\bar{\\mathbf{X}}\\) and \\(\\chi\\), particularly for objects that are far apart from the foreground. We observe that depth generated from DUSt3R [52] or MAST3R [29] has a strong bias towards smoothness between objects, despite local coherence. Therefore, we handle the outliers through a secondary Semantic Outlier Alignment step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.334, + 0.867 + ], + "angle": 0, + "content": "3.2.2. Semantic Outlier Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The key challenge in aligning outliers from \\(\\chi\\) to \\(\\bar{\\mathbf{X}}\\) is that, \\(\\bar{\\mathcal{O}}\\) only provides sparse correspondences to \\(\\chi\\), which has a 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "point for every pixel. As such, some heuristics need to be developed to group regions of \\(\\chi\\) to move based on \\(\\bar{\\mathcal{O}}\\). Based on the observation that geometric inconsistencies between \\(\\chi\\) and \\(\\bar{\\mathbf{X}}\\) tend to occur between objects and not within objects, we introduce an Interactive Segmentation Model (ISM) to get a series of masks according to \\(\\bar{\\mathcal{O}}\\) in an iterative manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.455 + ], + "angle": 0, + "content": "First, we randomly select a point from the outlier SfM set, \\(\\bar{\\mathbf{X}}_k\\in \\bar{\\mathcal{O}}\\) Using Eq.2),we can get the projected 2D point \\(\\bar{x}_k\\) to prompt ISM, producing a binary mask \\(m_{k}\\in \\mathbb{R}^{H\\times W}\\) delineating a semantically coherent local region around this outlier point. We further define a threshold \\(T\\) , which describes the desired number of correspondences to estimate a local alignment. If the number of outliers within \\(m_{k}\\) is greater than \\(T\\) i.e. \\(|m_k\\bigcap \\bar{P} (\\bar{\\mathcal{O}})| > T\\) then we keep \\(m_{k}\\) as it is and exclude within-mask outliers from being selected in the next iteration. Otherwise, we use all the outliers within the produced mask to prompt ISM again to update \\(m_{k}\\) until the above criterion is met, or no further outliers can be included, in which case we discard this mask and the outliers within. We repeat the above process until no point in \\(\\bar{\\mathcal{O}}\\) is left, eventually obtaining a set of semantically coherent masks around the outliers, \\(\\{m_k\\}_{k = 1}^M\\) .Visualization of this process can be found in the supplementary materials. This iterative process is denoted as" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.459, + 0.897, + 0.502 + ], + "angle": 0, + "content": "\\(\\operatorname{Init}: m_k = \\operatorname{ISM}(x_k)\\), s.t. \\(\\bar{x}_k \\in \\bar{P}(\\bar{\\mathcal{O}})\\) and \\(\\bar{x}_k \\notin \\bigcup_{j=1}^{k-1} m_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.504, + 0.905, + 0.556 + ], + "angle": 0, + "content": "Then, \\(m_{k} = \\left\\{ \\begin{array}{ll}m_{k}, & \\text{if} |m_{k}\\bigcap \\bar{P} (\\bar{\\mathcal{O}})| > T\\\\ \\operatorname {ISM}(m_{k}\\bigcap \\bar{P} (\\bar{\\mathcal{O}})), & \\text{otherwise} \\end{array} \\right.\\) (4)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.905, + 0.617 + ], + "angle": 0, + "content": "Based on the outlier masks \\(\\{m_k\\}_{k=1}^M\\), we extract the within-mask outlier set \\(\\mathcal{O}_k = P^{-1}(m_k \\bigcap P(\\mathcal{O}))\\) and their corresponding SfM set is \\(\\bar{\\mathcal{O}}_k\\). We then estimate a local rigid transformation within each mask as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.634, + 0.905, + 0.682 + ], + "angle": 0, + "content": "\\[\ns _ {k}, R _ {k}, t _ {k} = \\underset {s, R, t} {\\arg \\min } \\sum_ {(\\bar {\\chi} _ {i}, \\bar {\\mathbf {X}} _ {i}) \\in (\\mathcal {O} _ {k}, \\bar {\\mathcal {O}} _ {k})} \\| s R \\bar {\\chi} _ {i} + t - \\bar {\\mathbf {X}} _ {i} \\| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.683, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Finally, we construct the inlier mask as \\( m_0 = \\neg \\left( \\bigcup_{k=1}^{M} m_k \\right) \\), where \\( \\neg \\) is the logical NOT operator. Then, the global transformation is applied to MASt3R's points within inlier mask \\( m_0 \\) and each individual local transformation is applied to those within each local mask \\( m_k, k \\neq 0 \\). This results in the final SfM-aligned point cloud, \\( \\chi^* \\), where" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\chi^ {*} = \\bigcup_ {k = 0} ^ {M} \\left\\{s _ {k} R _ {k} \\chi_ {i} + t _ {k} \\mid \\chi_ {i} \\in m _ {k} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.905, + 0.878 + ], + "angle": 0, + "content": "To obtain the final dense and pose-wise accurate point cloud \\(\\mathcal{X}\\) as prior, we concatenate the final SfM-aligned point cloud with that of SfM, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.883, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} = \\chi^ {*} \\bigcup \\bar {\\mathbf {X}} \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.3, + 0.108 + ], + "angle": 0, + "content": "3.2.3. Gaussian Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.485, + 0.186 + ], + "angle": 0, + "content": "The SPARS3R constructed \\(\\mathcal{X}\\) can be directly used as the initialization for any Gaussian Splitting-based optimization method. Here we use Splatfacto, developed under the NeRFStudio framework [49]; the Gaussian optimization loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.2, + 0.483, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\| \\tilde {C} - C \\| _ {1} + \\lambda_ {2} \\mathrm {D} - \\operatorname {S S I M} (\\tilde {C}, C), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.483, + 0.264 + ], + "angle": 0, + "content": "where \\(\\tilde{C}\\) and \\(C\\) denote our rendered and the groundtruth RGB image, D-SSIM denotes the structural similarity loss." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.28, + 0.224, + 0.297 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.485, + 0.577 + ], + "angle": 0, + "content": "Experiments are performed on three popular benchmark datasets. Tanks and Temples [27] contains 8 scenes captured in video format. Following InstantSplat [14], We uniformly sample 24 images with alternating indices for train and test split. MVImgNet [60] is an object-centric dataset that consists of 7 scenes. These scenes are captured from a consistent height covering \\(180^{\\circ}\\) field of view. We apply the same sampling and train-test split procedure as for Tanks and Temples. For sparser settings, we uniformly sample 3 and 6 views from the training set for the above two datasets. Mip-NeRF 360 [2] comprises of 9 scenes with \\(360^{\\circ}\\) views and greater pose variation between the scenes, including diverse heights and distances. The increased variability, combined with sparsity, adds complexity to accurate pose estimation and scene reconstruction. For this dataset, we follow the test set outlined in MipNeRF360 [2] and uniformly sample 12 images from the original training set to construct a sparse-view set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.671 + ], + "angle": 0, + "content": "All sparse-view input images go through a COLMAP [46] pipeline for camera calibration. Specifically, we use feature matching results from MASt3R [29] for SfM triangulation. Since sparse-view registration can be unstable due to limited pairs, we perform multiple SfMs and pick the outcome that maximizes successful triangulation per image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.31, + 0.698 + ], + "angle": 0, + "content": "4.1. Sparse NVS Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Accurately and holistically evaluating sparse NVS should involve many practical considerations. Various NeRF-based approaches [12, 23, 25, 39, 51, 57, 59] have assumed perfect camera poses, i.e. calibration results from dense views, for sparse-view scenarios, which is unrealistic. Additionally, Gaussian-Splatting-based methods rely heavily on SfM point cloud, further necessitating the separation of train and test views from the registration stage. Previous approaches [65] have tried to perform re-triangulation based on known train poses, excluding test-set point cloud from training, but do not account for pose inaccuracy in sparse-view registration. While InstantSplat [14] adopts the realistic setup where they use recovered poses from sparse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "views and initialize the test camera poses by applying Procrustes Alignment [22] between train-only poses and train-test-together poses, their large camera alignment error leads to significant disruption in common pixel-based rendering metrics, e.g. PSNR and SSIM [53]." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.179, + 0.885, + 0.257 + ], + "angle": 0, + "content": "
Mip-NeRF 360ER(μ)↓ET(μ)↓
Procrustes Alignment0.1960.0144
+ RANSAC0.1790.0114
+ Rotation Points0.1560.0117
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.907, + 0.309 + ], + "angle": 0, + "content": "Table 1. Improvements over Procrustes Alignment baseline in average rotation error \\( E_{R} \\) and translation error \\( E_{T} \\). Incorporating rotation points further minimizes the overall error." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.906, + 0.41 + ], + "angle": 0, + "content": "We also operate under the realistic setup but improve upon the camera alignment process in two aspects. First, during camera alignment, instead of relying only on camera positions, we sample three additional points \\( P_{\\mathrm{rotation}} \\in \\mathbb{R}^{3 \\times 3} \\) per camera along the unit direction vectors in rotation matrix \\( R_{\\mathrm{camera}} \\in \\mathbb{R}^{3 \\times 3} \\) at camera position \\( T_{\\mathrm{camera}} \\in \\mathbb{R}^{1 \\times 3} \\), i.e." + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.424, + 0.905, + 0.44 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {r o t a t i o n}} = s _ {\\text {c l o u d}} R _ {\\text {c a m e r a}} + T _ {\\text {c a m e r a}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.446, + 0.907, + 0.59 + ], + "angle": 0, + "content": "The scale of the camera cloud is denoted by \\( s_{\\mathrm{cloud}} = \\sqrt{\\sigma_x^2 + \\sigma_y^2 + \\sigma_z^2} \\), and \\( \\sigma_x, \\sigma_y, \\) and \\( \\sigma_z \\) are the standard deviations of camera positions along \\( x, y, \\) and \\( z \\) axes respectively. The four points per camera enable alignment of both camera rotations and translations. Second, we add RANSAC [15] in the alignment process, minimizing the influence of outliers. We measure the alignment rotation error \\( E_R \\) by the angular difference in degrees between two rotations represented by quaternions following" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.598, + 0.905, + 0.632 + ], + "angle": 0, + "content": "\\[\nE _ {R} = 2 \\cdot \\arccos \\left(\\frac {\\operatorname {a b s} \\left(q _ {1} \\cdot q _ {2}\\right)}{\\| q _ {1} \\| \\cdot \\| q _ {2} \\|}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.778 + ], + "angle": 0, + "content": "and the translation error \\( E_{T} \\) by the Euclidean norm \\( \\| \\cdot \\| \\) between translation vectors. As shown in Tab. 1, these two improvements enhance camera alignment accuracy in both rotation and translation. Beyond accurate camera pose alignment, test pose optimization can also be applied between rendered and ground-truth images to minimize the pose error; however, such process is time-consuming and can get stuck if the initial displacement is too large, as is sometimes the case in InstantSplat [14]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. In addition to the commonly used PSNR, SSIM, and LPIPS, we incorporate DreamSim [18] (DSIM) as an additional metric to assess rendering quality. DSIM captures human perceptual similarity by finetuning a combination of embeddings from visual foundation models based on human evaluations. Notably, it provides a robust assessment of rendering image quality under moderate pose shift, which frequently occurs in the realistic sparse-view" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26814" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "settings. Further details on DSIM's robustness as an evaluation metric are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.126, + 0.484, + 0.232 + ], + "angle": 0, + "content": "Implementation Details. We employ Segment-Anything Model [26] as the segmentation model in Semantic Outlier Alignment. The weights for the loss terms are \\(\\lambda_{1} = 0.8\\), \\(\\lambda_{2} = 0.2\\). For fair implementation and comparison, we employ test pose optimization for all baselines and SPARS3R for 500 steps to maximally remove the effect of shifted camera pose. All baselines are run until convergence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.242, + 0.258, + 0.258 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.308 + ], + "angle": 0, + "content": "We compare the different components in SPARS3R to demonstrate their effectiveness on the Mip-NeRF 360 [2] dataset:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.312, + 0.483, + 0.342 + ], + "angle": 0, + "content": "- SfM-initialized 3DGS: Optimizing 3DGS on a sparse SfM point cloud and SfM-estimated camera poses." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.343, + 0.483, + 0.372 + ], + "angle": 0, + "content": "- MAST3R-initialized 3DGS: Optimizing 3DGS on a MAST3R-produced dense point cloud and camera poses." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.373, + 0.482, + 0.417 + ], + "angle": 0, + "content": "- Global Fusion Alignment: Aligning the MAST3R-produced dense point cloud onto SfM point cloud with a global transform. Poses are taken from SfM." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.418, + 0.482, + 0.462 + ], + "angle": 0, + "content": "- Semantic Outlier Alignment (SPARS3R): The full SPARS3R pipeline that aligns outliers through their 2D semantic masks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.312, + 0.483, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.476, + 0.479, + 0.589 + ], + "angle": 0, + "content": "
SettingPSNRSSIMLPIPSDSIM
Initialization: SfM [46]
3DGS [24]16.60.3880.4580.232
Initialization: MASt3R [29]
3DGS [24]15.90.2930.4630.229
+ Global Fusion Alignment18.60.4860.3300.130
+ Semantic Outlier Alignment18.90.5000.3270.127
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Table 2. Ablation on key components of SPARS3R. The results are shown using PSNR, SSIM, LPIPS and DSIM based on the MipNeRF 360 [2]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Tab. 2, given only 12 images, sparse NVS is challenging both in registration and reconstruction. Specifically, we observe that the 3DGS based on SfM initialization performs better than MASt3R initialization. Clearly, render quality is affected if poses are inaccurate, despite locally coherent dense point cloud. The introduction of Global Fusion Alignment improves upon the naive baselines from either initializations by combining accurate pose from SfM and dense point cloud from MASt3R. For scenes that do not have significant depth discrepancies, Global Fusion Alignment works well. In scenarios such as the Bonsai scene in Fig. 3, MASt3R's relative depth between the table and the wall is inaccurate; therefore, Global Fusion Alignment alone leads to a broken geometry of the background. In comparison, our Semantic Outlier Alignment provide piece-wise alignments on the background to ad" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "dress this relative discrepancy. Quantitatively, we observe a 1.4 dB improvement on the Bonsai scene." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.133, + 0.709, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.229, + 0.704, + 0.242 + ], + "angle": 0, + "content": "Global Fusion Alignment (GFA)" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.133, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.229, + 0.895, + 0.242 + ], + "angle": 0, + "content": "Semantic Outlier Align (SOA)" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.253, + 0.705, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.347, + 0.699, + 0.36 + ], + "angle": 0, + "content": "Aligned point cloud from GFA" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.248, + 0.901, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.347, + 0.896, + 0.36 + ], + "angle": 0, + "content": "Aligned point cloud from SOA" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.433 + ], + "angle": 0, + "content": "Figure 3. Visual comparison of SPARS3R with and without SOA. While the dense bonsai in the foreground is aligned with the sparse point cloud, depth differences are obvious. SOA successfully fixes such gap." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.459, + 0.918, + 0.543 + ], + "angle": 0, + "content": "
MethodsMip-NeRF 360Tanks & TemplesMVimgNet
RPEtRPErRPEtRPErRPEt RPEr
DUSt3R [52]2.0752.5840.5700.1430.438 0.421
MASt3R [29]1.1861.4900.2410.2480.208 0.304
InstantSplat [14]2.0492.5550.1510.0810.264 0.311
COLMAP + MASt3R0.2520.4120.1610.0930.075 0.078
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.553, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of pose accuracy across three datasets, Relative Translation Error \\((\\mathrm{RPE}_t)\\) and Relative Rotation Error \\((\\mathrm{RPE}_r)\\) [62] are calculated based on the normalized poses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.826 + ], + "angle": 0, + "content": "Pose Estimation Accuracy. As shown in Tab. 3, we compare the pose estimation accuracy from different calibration approaches. Both DUSt3R [52] and MASt3R [29] have significantly larger errors compared to COLMAP [46] with MASt3R feature matching outputs. InstantSplat [14] uses DUSt3R's [52] dense point cloud and pose estimation and attempts to improve accuracy through a training pose optimization approach similar to BARF [32]. While it brings down the errors in some cases, such training pose optimization strategy does not work as well in more challenging datasets like Mip-NeRF 360. This highlights the difficulty in improving camera calibration jointly with dense forward render, considering the loss landscape is highly nonlinear based on pixel-wise metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.822, + 0.85 + ], + "angle": 0, + "content": "4.3.Quantitative and Visual Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We quantitatively compare SPARS3R against various NVS methods in Tab. 4 and Tab. 5. Instant-NGP [36] and 3DGS [24] are two mainstream NVS methods based on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26815" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.9, + 0.233 + ], + "angle": 0, + "content": "
MethodsMip-NeRF 360 [2]Tanks & Temples [27]MVImgNet [60]
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
Instant-NGP [36]14.820.2940.6780.52415.280.4510.3890.25413.280.4260.8920.892
3DGS [24]16.570.3880.4580.23221.070.7300.1940.07021.240.6730.2340.064
FSGS [65]17.600.4430.5580.24325.720.8450.1110.02323.430.7600.2120.042
SparseGS [55]16.660.4050.4610.21020.280.7270.2020.07520.560.6720.2480.072
DRGS [10]16.880.4010.6490.30021.460.7230.2890.07821.700.6410.4220.071
CF-3DGS [19]13.270.2500.6980.50918.990.6060.2960.12715.430.4080.5450.325
InstantSplat [14]16.230.3590.5430.23326.970.8740.1150.01323.220.7340.2480.028
SPARS3R18.850.5000.3270.12729.900.9190.0470.00725.850.8200.1140.011
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.243, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of different NVS methods on 12 views on three popular benchmark datasets, totaling 24 scenes. All methods are run on the same registrations and updated with test pose optimization." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.295, + 0.49, + 0.421 + ], + "angle": 0, + "content": "
3 ViewsTanks & TemplesMVImgNet
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
FSGS20.360.6580.2190.07716.920.5120.4760.183
InstantSplat21.650.7550.1640.04416.290.5510.3230.103
SPARS3R21.830.7550.1500.04318.620.5880.2810.091
6 Views
FSGS23.900.7830.1520.04320.420.6440.3310.080
InstantSplat25.890.8470.1170.02221.090.6950.2360.046
SPARS3R26.800.8730.0730.01422.750.7320.1610.023
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.46 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of 3 and 6 views on Tanks & Temples and MVImgNet datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.598 + ], + "angle": 0, + "content": "implicit and explicit scene representations. Specifically, 3DGS leverages SfM points from registration as the starting point for scene optimization, while NeRF [35] based methods rely only on poses. Colmap-Free 3DGS [19] also does not rely on prior registration and optimizes cameras along with reconstruction. We find that both Instant-NGP and CF-3DGS perform significantly worse than other methods that use a prior SfM point cloud." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.6, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Among methods that are designed for sparse NVS, DRGS [11], SparseGS [55], and FSGS [65] leverage monocular depth prior to gain more geometric information of the scene. DRGS applies an additional smoothness constraint, which leads to high PSNR values but worse LPIPS and DSIM performances. SparseGS additionally applies a Score Distillation Sampling Loss, which generates novel views and constrained them through a prior generative model. Its results are sharper than DRGS, but more noisy, leading to worse PSNR. FSGS seeks to address the sparse point cloud issue in sparse-view NVS by modifying the densification algorithm in 3DGS. More Gaussians are encouraged between the space of two existing Gaussians, allowing representative areas to densify quicker. FSGS has overall better metrics than SparseGS and DRGS; however, it still encounters issues in regions where few initializations are available, e.g. the background of the scene. As shown in Fig. 4 and in Fig. 1, this leads to various blurry regions. InstantSplat [14] is a concurrent work that uses prior depth estimation from DUSt3R [52] and the dense point cloud to help with sparse-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.51 + ], + "angle": 0, + "content": "view NVS. While the initialized point cloud is much denser than FSGS, InstantSplat also inherits the suboptimal poses produced by DUSt3R. The incorrect geometry limits the effectiveness of densification in 3DGS and can lead to generation of floaters; to ameliorate this problem, InstantSplat disables Gaussian densification and optimizes the scene with few steps. As shown in Tab. 3 and Tab. 4, on the Mip-NeRF 360 and MVImgNet dataset, where pose accuracy from DUSt3R is limited, InstantSplat performs on-par with FSGS despite using a denser prior point cloud. As shown in Fig. 4, we observe that the Gaussians tend to be noisy on the surface. In the Stump and Bicycle scenes, InstantSplat's results cannot be rigidly aligned well with the groundtruth test poses due to the large offset to begin with." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.699 + ], + "angle": 0, + "content": "Our SPARS3R method improves upon DUSt3R and MASt3R's prior point cloud by adjusting it based on conventional SfM calibration approaches. Since SfM only triangulates based on confident keypoints, its pose and depth estimation performance, particularly under sparse-view scenario, is more reliable. As shown in Tab. 4, SPARS3R consistently improves upon previous best approach by an average of \\(2.7\\mathrm{dB}\\). SPARS3R is stable in its optimization process and does not generate additional floaters with more iterations. As seen in Fig. 3, SPARS3R produces rendering with smooth surfaces in the foreground, and sharp details in the background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For even sparser settings, we provide a quantitative comparison in Tab. 5 between SPARS3R and the two best-performing methods from Tab. 4, FSGS [65] and InstantSplat [14]. FSGS struggles significantly under extreme sparsity due to its reliance on a sparse initialization, resulting in DSIM scores that are approximately two- and four-times worse than those of SPARS3R under the 3-view and 6-view settings, respectively. InstantSplat generally performs better than FSGS, benefiting from a much denser point cloud initialization produced by DUSt3R [52], which is particularly advantageous under extreme sparse conditions. However, its performance remains inferior to SPARS3R due to suboptimal pose estimation. Our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.918, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.475, + 0.908, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons of different NVS methods on 12 views on Mip-NeRF 360 [2] dataset. Zooming in on the visualizations is recommended to show differences in detail. More visualizations for other datasets are available in the Supplemental Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.637 + ], + "angle": 0, + "content": "SPARS3R continues to outperform these methods, improving upon the best approach by an average of 0.9 dB and 1.3 dB in the 3-view and 6-view settings, respectively. These results further demonstrate SPARS3R's superiority over state-of-the-art approaches in handling extreme sparse-view reconstruction. For more visualizations and flythroughs of the scene, please refer to the supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.218, + 0.665 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "While SPARS3R significantly improves upon previous SoTA, there are also several limitations worth noting. Firstly, SPARS3R relies on a well-performing semantic segmentation model to group outliers; as such, if the segmentation model is very sensitive, i.e. separating a connected surface into multiple disjointed parts, some of these parts may lack sufficient SfM support for alignment. If the segmentation model produces masks that group multiple areas with disparate depth together, the local alignments are also less effective similar to the global alignment. Secondly, non-rigid transform between two point clouds are worth further investigation. While SPARS3R's piece-wise rigid transform based on semantic relationship is fast and addresses the inaccurate inter-object depth estimation, a non-rigid formulation may be more generalizable." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.633, + 0.544 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We present SPARS3R, a scene reconstruction and NVS method that can achieve high-quality rendering with sparse input images. We demonstrate that existing methods lead to blurry rendering due to sparse point cloud initialization. Recent progress in generating a dense point cloud from prior depth estimation model, while promising, can lead to noisy pose estimation. To address these issues, SPARS3R combines the advantages of both by proposing a two-step alignment approach. The first step computes a global transformation between the dense point cloud from depth prior and a reference SfM point cloud. The second step takes the semantic regions around the outlier points identified from the first step and performs successive local alignments. This approach successfully addresses the depth estimation errors within the prior dense point cloud and demonstrates that the updated point cloud leads to significantly better performances in sparse-view NVs. We also introduce several improvements in the evaluation process to better represent the practical limitations in sparse-view registration and reconstruction. In the future, we hope to further improve the smoothness of our alignment process, e.g. by exploring potential non-rigid transformation approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26817" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.311 + ], + "angle": 0, + "content": "This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD." + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.091, + 0.611, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.115, + 0.907, + 0.197 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.199, + 0.908, + 0.269 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.271, + 0.907, + 0.338 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.341, + 0.907, + 0.395 + ], + "angle": 0, + "content": "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.398, + 0.906, + 0.449 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.453, + 0.906, + 0.522 + ], + "angle": 0, + "content": "[6] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19436-19446, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.524, + 0.906, + 0.592 + ], + "angle": 0, + "content": "[7] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.594, + 0.906, + 0.661 + ], + "angle": 0, + "content": "[8] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12952, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.664, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[9] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[10] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[11] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.115, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26818" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with \\(15\\mathrm{x}\\) reduction and \\(200+\\) fps. arXiv preprint arXiv:2311.17245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[18] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[19] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[20] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[21] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[22] John C Gower. Generalized procrustes analysis. Psychometrika, 40:33-51, 1975. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[23] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[25] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infererf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[28] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719-21728, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[30] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[31] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[33] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[37] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26819" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[38] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349-10358, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[39] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[40] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatt-ting for robust real-time rendering with \\(900+\\) fps. arXiv preprint arXiv:2403.13806, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[41] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In European Conference on Computer Vision, pages 19-37. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[42] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[43] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In European Conference on Computer Vision, pages 293-310. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[47] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[49] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[50] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[51] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[52] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[54] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[55] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 \\(\\{\\backslash \\mathrm{deg}\\}\\) sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[56] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[57] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[58] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15901-15911, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[60] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26820" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[61] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[62] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual(-inertial) odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7244-7251, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[63] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[64] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[65] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1, 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26821" + } + ] +] \ No newline at end of file diff --git a/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_origin.pdf b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a89126e9daa0d4ffd9290188af09fef17f41cffc --- /dev/null +++ b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/ad2ca2d2-676a-49f1-84fe-0212a262ed5a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:584e92145b934baa35753da72dc850429fb4f946c3d7fa8b3b8e1761d4a8f74c +size 9999477 diff --git a/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/full.md b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1150a08df564b308d7dc2de833c339a17977e7c3 --- /dev/null +++ b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/full.md @@ -0,0 +1,354 @@ +# SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction + +Yutao Tang* + +Yuxiang Guo* + +Deming Li + +Cheng Peng + +Johns Hopkins University + +{ytang67，yguo87，dli90，cpeng26}@jhu.edu + +# Abstract + +Recent efforts in Gaussian-Splat-based Novel View Synthesis can achieve photorealistic rendering; however, such capability is limited in sparse-view scenarios due to sparse initialization and over-fitting floaters. Recent progress in depth estimation and alignment can provide dense point cloud using few views; however, the resulting pose accuracy is suboptimal. In this work, we present SPARS3R, which combines the advantages of accurate pose estimation from Structure-from-Motion and dense point cloud from depth estimation. To this end, SPARS3R first performs a Global Fusion Alignment process that maps a prior dense point cloud to a sparse point cloud from Structure-from-Motion based on triangulated correspondences. RANSAC is applied during this process to distinguish inliers and outliers. SPARS3R then performs a second, Semantic Outlier Alignment step, which extracts semantically coherent regions around the outliers and performs local alignment in these regions. Along with several improvements in the evaluation process, we demonstrate that SPARS3R can achieve photorealistic rendering with sparse images and significantly outperforms existing approaches. + +# 1. Introduction + +Photorealistic scene reconstruction and Novel View Synthesis (NVS) from unposed 2D images is a challenging task with numerous applications in site modeling, autonomous driving, robotics, urban and agricultural planning, etc. The introduction of methods such as Neural Radiance Field (NeRF) [35] and 3D Gaussian Splatting (3DGS) [24] have introduced significant advances in rendering quality and efficiency based on dense multi-view imagery. However, applications of these methods still face issues in practical scenarios where a dense coverage of the scene is not available. + +Given sparse images, over-fitting the photometric objectives to an incorrect geometry is a common issue in + +![](images/8a784259f4c66dd91775151c1516e7a0efa985b79a39fb642e70049b092546f2.jpg) +Instant-NGP [36] +FSGS [65] + +![](images/7186186877bf36149aa9b12a105ead39acc7b6a2c3d5c1eac9b2e2080ae13c07.jpg) +InstantSplat [14] +SPARS3R +Figure 1. A visualization of SPARS3R in comparison to current SoTA. Without additional prior, sparse NVS leads to incorrect geometry by Instant-NGP [36]. FSGS [65] can be blurry due to sparse initialization and insufficient densification. InstantSplat [14] relies on DUSt3R [52] initialization with suboptimal poses. Our method, SPARS3R, can reliably render details in the foreground and background with accurate poses. + +NVS. Various constraints, such as semantic consistent loss [23], depth and novel-view regularization [12, 39, 51], frequency regularization [57], and ray entropy minimization [25] have been introduced to improve upon NeRF. These methods often lead to significant computation overhead due to the costly patch rendering from ray-tracing. More recently, Gaussian-Splatting-based approaches have further improved sparse NVS by leveraging the explicit representation and fast differentiable rasterization. Depth regularization [30], Gaussian floater pruning [55], and proximity-based Gaussian densification strategy [65] have been proposed to constrain and guide the scene structures. As demonstrated in Fig. 1, despite improved geometry, these approaches tend to produce overly smooth rendering in the background. This issue can be attributed to a sparse initial + +point cloud, particularly in the background regions. Furthermore, the additional constraints, e.g. based on monocular depth prior, are oftentimes imperfect, leading to noisy gradients that prevent proper densification in sparse regions. + +One potential solution to address this conundrum is to initialize Gaussian optimization with a denser point cloud to help disambiguate shape and radiance. To this end, recent advances in two-view depth estimation, particularly DUSt3R [52] and MAST3R [29], have shown the impressive ability to construct a dense point cloud from a pre-trained prior model. In theory, such a point cloud can be directly used to optimize a Gaussian-Splat radiance field [14]. In practice, camera calibration obtained from multi-view depth alignment is often suboptimal due to the difficulties in estimating an accurate depth map. If left uncorrected, the Gaussian optimization process will generate floaters to compensate the suboptimal calibration, prompting strategies to prevent densification as a stopgap measure [14]. In comparison, camera calibration based on Structure-from-Motion (SfM), e.g., COLMAP [46], is fast and accurate with an appropriate feature matching model. Instead of aligning dense depth, SfM takes confident correspondences and performs triangulation with RANSAC [15] to reject outliers. In a sparse-view setting, such a process can be repeated numerous times to further improve accuracy. + +To address sparse point cloud initialization and pose inaccuracy in sparse-view NVS, we propose SPARS3R. SPARS3R composes of two stages: Global Fusion Alignment and Semantic Outlier Alignment. In Global Fusion Alignment, SPARS3R first obtains a dense point cloud from sparse views through DUSt3R [52] or MASt3R [29], and a sparse point cloud from COLMAP [46]. By taking the triangulated correspondences within images, SPARS3R fuses the dense point cloud onto the sparse point cloud through a global Procrustes Alignment process with RANSAC [15]. Since some points will yield large error due to local scale variations, a second Semantic Alignment process is introduced to perform local alignment. Specifically, outliers from Global Fusion Alignment are identified and used as prompts for an Interactive Segmentation model, e.g. SAM [26], generating semantically coherent masks around the outliers. The resulting semantic masks indicate regions within dense point cloud to be treated with local alignment. By locally aligning the identified outliers with the SfM point cloud while preserving the globally aligned inliers, we obtain a dense and pose-wise accurate point cloud, serving as a strong prior for Gaussian optimization. + +In summary, our contributions can be summarized as: + +1. We propose a Global Fusion Alignment approach, which transforms a prior dense point cloud onto a reference SfM sparse point cloud, putting dense initialization and accurate camera poses in the same coordinate frame. +2. To address outliers that cannot be aligned accurately due + +to depth discrepancies, we propose a Semantic Outlier Alignment step. This step extracts semantically similar regions around the outliers to perform local alignment, resulting in a dense point cloud with minimum transformation error. + +3. We evaluate the overall method, SPARS3R, on three popular benchmark datasets and find significant quantitative and visual improvements compared to current SoTA methods. + +# 2. Related Work + +# 2.1. 3D Models for Synthesizing Novel Views + +Photorealistic scene reconstruction and novel view synthesis is a long-standing task in computer vision and graphics. Neural Radiance Field (NeRF) [35] proposes to model a scene implicitly with multi-layer perceptron (MLP) by mapping 3D coordinates and view direction to color and density, then rendering pixel value via alpha blending. Since NeRF's introduction, numerous work has sought to improve its efficiency [5, 16, 20, 36, 45, 48], quality [1-3, 7, 50], and extension to dynamic and in-the-wild [4, 6, 8, 17, 34, 42, 44, 47, 54, 58] scenarios. Recently, 3D Gaussian Splatting (3DGS) [24] has emerged as an efficient alternative, offering improved rendering quality over NeRF. Unlike NeRF, 3DGS represents scenes with explicit 3D Gaussians kernels and uses a differentiable rasterization technique. Following its introduction, many methods have emerged to reduce the computational cost [13, 21, 28, 37, 38] and enhance quality [9, 33, 40, 43, 56, 61, 63, 64] of 3DGS. + +# 2.2. Gaussian-Splatting-based Sparse-View NVs + +While 3DGS performs well with dense view support, its effectiveness decreases in practical settings with sparse inputs. Many approaches address this through depth constraints. FSGS [65] uses a Proximity-guided Gaussian Unpooling to mitigate the sparse initialization issue and introduces pseudo-views during training to avoid overfitting. DRGS [11] and SparseGS [55] add a depth regularization term to enforce consistency between estimated and monocular depths while promoting smoothness. DNGS [31] refines depth regularization by prioritizing local depth variations. CoherentGS [41] initializes Gaussians from monocular depth estimation and improves coherence through an optical flow constraint. InstantSplat [14] is a concurrent work that directly uses DUSt3R to produce dense point clouds and introduces a camera pose optimization strategy; densification of Gaussians is disabled to prevent the introduction of floaters due to suboptimal poses. + +# 3. Methods + +# 3.1. Preliminary + +DUSt3R [52] is a two-view depth estimation method that produces dense 3D point clouds from image pairs. Specif + +![](images/946e9e00b1cae5139d38a7f6349ff06938ab3bd9742f08cb17fdea43306eb754.jpg) +Figure 2. SPARS3R combines a prior dense point cloud $\chi$ and a sparse SfM point cloud $\bar{\mathbf{X}}$ . The prior $\chi$ often has inferior depth accuracy compared to $\bar{\mathbf{X}}$ . SPARS3R first globally aligns all points in $\chi$ onto $\bar{\mathbf{X}}$ , based on shared correspondences. Inliers and outliers are identified through alignment error. SPARS3R then extracts the semantically relevant 2D regions around the outliers to move local regions of $\chi$ in groups, producing a dense point cloud $\chi^{*}$ that is depth-wise and pose-wise accurate. By providing $\mathcal{X} = \chi^{*} \bigcup \bar{\mathbf{X}}$ for Gaussian optimization, SPARS3R achieves photorealistic rendering under sparse-view condition. + +ically, given an image pair $I^1$ and $I^2$ , DUSt3R generates dense pointmaps $\mathbf{X}^{1,1}$ and $\mathbf{X}^{2,1} \in \mathbb{R}^{H \times W \times 3}$ . A pointmap, e.g. $\mathbf{X}^{1,1}$ , provides a mapping from each pixel $(u, v)$ in image $I^1$ to a corresponding 3D point $\mathbf{X}^{1,1}(u, v)$ under the coordinate system of image $I^1$ , giving dense 2D-to-3D correspondences. Given $K > 2$ input images, DUSt3R [52] aggregates across all pairwise pointmap predictions by globally aligning pairwise pointmaps into a unified point cloud $\chi$ . This alignment process also produces camera parameters $\{W(I^n), J(I^n)\}$ , which are the world-to-camera transformation matrix and the Jacobian of the affine approximation of the projective transformation for view $I^n$ . As such, the relationship between a pixel $(u, v)$ on $I^n$ and its 3D point $\chi_i$ is established via a projective function $P$ as follows: + +$$ +x _ {i} = [ u, v ] ^ {T} = P \left(\chi_ {i}\right) = J \left(I ^ {n}\right) W \left(I ^ {n}\right) \chi_ {i}. \tag {1} +$$ + +Building on this, MASt3R [29] incorporates an additional matching head that generates discriminative local feature descriptors and pixel-level correspondences across paired images. We note that while MASt3R [29] uses this feature correspondence during its multi-depth alignment, the estimated poses are oftentimes inaccurate due to smooth depth predictions. + +# 3.2. SPARS3R + +As show in Fig. 2, SPARS3R builds upon the advancement in DUSt3R [52] and MASt3R [29] as a pre-trained 3D prior for sparse-view reconstruction. Firstly, SPARS3R performs SfM based on image correspondences, either from MASt3R [29] or other feature matching methods. SPARS3R then aligns the dense point cloud produced by MASt3R via rigid transformations in two stages: Global Fusion Alignment and Semantic Outlier Alignment. + +# 3.2.1. Global Fusion Alignment + +While DUSt3R [52] and MASt3R [29] can generate dense 3D points from sparse images, it does not estimate accurate or detailed depth through its pointmap formulation. As a result, the camera calibration obtained from such smooth depths is suboptimal and will degrade the fidelity of scene reconstructions. To construct a better point cloud prior, we propose to align MASt3R's point cloud with that from a SfM pipeline, which is more reliable based on selective correspondences. This alignment process begins by identifying matching points between the SfM and MASt3R point clouds, denoted as $\overline{\mathbf{X}}\in \mathbb{R}^{\bar{N}\times 3}$ and $\chi \in \mathbb{R}^{N\times 3}$ , respectively. A visibility indicator $V:\mathbb{R}^{\bar{N}\times 3}\to \mathbb{R}^{\widetilde{N}\times 3}$ denotes if an SfM point $\overline{\mathbf{X}}_i$ is visible with respect to a view $I^n$ , and $\tilde{N}$ denotes the number of visible SfM points in view $I^n$ . For each view, similar to Eq. (1), we can obtain the pixel-to-3D relationship of $(\bar{u},\bar{v})$ and $V(\overline{\mathbf{X}})_i$ through estimated $\overline{W}(I^n)$ and $\bar{J}(I^n)$ from SfM. Subsequently, the correspondence between a SfM point $V(\overline{\mathbf{X}})_i$ and a point in $\chi$ , denoted as $\tilde{\chi}_i$ , can be established through a forward-backward projection: + +$$ +\bar {x} _ {i} = \left[ \bar {u}, \bar {v} \right] ^ {T} = \bar {P} (\bar {\mathbf {X}}) _ {i} = \bar {J} (I ^ {n}) \bar {W} (I ^ {n}) V (\bar {\mathbf {X}}) _ {i}, \tag {2} +$$ + +$$ +\bar {\chi} _ {i} = P ^ {- 1} (\bar {x} _ {i}) = W ^ {- 1} \left(I ^ {n}\right) J ^ {- 1} \left(I ^ {n}\right) \bar {x} _ {i}. +$$ + +For each view, we denote the ordered set of visible SfM points as $\{V(\bar{\mathbf{X}})_i\}_{i = 1}^N$ , the projected 2D pixels as $\{\bar{x}_i\}_{i = 1}^N$ , and the set of 3D points that correspond to the projected 2D pixels above in $\chi$ as $\{\bar{\chi}_i\}_{i = 1}^N$ , omitting view indicator $n$ . With these identified correspondences, we can apply Generalized Procrustes Analysis [22] to estimate a global rigid transformation, optimizing for scale, rotation, and translation. During this process, RANSAC is applied to filter out- + +liers, excluding their influence on the estimation. Specifically, we minimize the following objective to get the optimal global transformation based on Algorithm 1: + +$$ +s _ {0}, R _ {0}, t _ {0} = \underset {s, R, t} {\arg \min } \sum_ {i = 1} ^ {\widetilde {N}} e \left(\bar {\chi} _ {i}, s, R, t, V (\bar {\mathbf {X}}) _ {i}\right) \tag {3} +$$ + +$$ +e (\bar {\chi} _ {i}, s, R, t, V (\bar {\mathbf {X}}) _ {i}) = \left\| s R \bar {\chi} _ {i} + t - V (\bar {\mathbf {X}}) _ {i} \right\| ^ {2}. +$$ + +# Algorithm 1 RANSAC Alignment + +1: Input: Matched MASt3R points $\{\overline{\chi}_i\}_{i=1}^N$ and SfM points $\{V(\bar{\mathbf{X}})_i\}_{i=1}^N$ , sample size $n$ , error threshold $\epsilon$ , number of iterations $R$ +2: Output: Best parameters $(s_0,R_0,t_0)$ , inlier set $\mathcal{I}$ , outlier set $\mathcal{O}$ , corresponding outlier SfM set $\bar{\mathcal{O}}$ +3: Initialize: num_inliers $\leftarrow 0$ , $(\hat{s_0},\hat{R_0},\hat{t_0})\gets \emptyset ,\hat{\mathcal{I}}\gets \emptyset ,\hat{\mathcal{O}}\gets \emptyset ,\hat{\bar{\mathcal{O}}}\gets \emptyset$ Iterations $\leftarrow 0$ +4: while Iterations $< R$ do +5: Randomly select $n$ points as sample set $S \subset \{i\}_{i=1}^{\widetilde{N}}$ . +6: Estimate $(\hat{s}_0,\hat{R}_0,\hat{t}_0)$ on sample set $\mathcal{S}$ by minimizing $\sum_{i\in \mathcal{S}}e(\bar{\chi}_i,s,R,t,V(\bar{\mathbf{X}})_i)$ (Eq.3)) +7: for each point $\overline{\chi}_i$ do +8: if $e(\overline{\chi}_i, \hat{s}_0, \hat{R}_0, \hat{t}_0, V(\overline{\mathbf{X}})_i) < \epsilon$ then +9: Add $\bar{\chi}_i$ to inlier set $\hat{\mathcal{I}}$ +10: else +11: Add $\bar{\chi}_i$ to outlier set $\hat{\mathcal{O}}$ and add $V(\overline{\mathbf{X}})_i$ to outlier SfM set $\hat{\mathcal{O}}$ +12: end if +13: end for +14: if $|\mathcal{I}| > \text{num\_inliers}$ then +15: Update: num_inliers \(\leftarrow |\hat{\mathcal{I}}|, (s_0, R_0, t_0) \leftarrow (\hat{s}_0, \hat{R}_0, \hat{t}_0), \mathcal{I} \leftarrow \hat{\mathcal{I}}, \mathcal{O} \leftarrow \hat{\mathcal{O}}, \bar{\mathcal{O}} \leftarrow \hat{\bar{\mathcal{O}}} +16: end if +17: Iterations $\leftarrow$ Iterations $+1$ +18: end while +19: return $(s_0, R_0, t_0), \mathcal{I}, \mathcal{O}, \mathcal{O}$ + +After this, we extract an inlier set $\mathcal{I}$ , an outlier set $\mathcal{O}$ and a corresponding outlier SfM set $\bar{\mathcal{O}}$ from RANSAC with an empirical threshold. As shown in Fig. 2, the global transformation recovered from Eq. 3 often leaves large amount of outliers between $\bar{\mathbf{X}}$ and $\chi$ , particularly for objects that are far apart from the foreground. We observe that depth generated from DUSt3R [52] or MAST3R [29] has a strong bias towards smoothness between objects, despite local coherence. Therefore, we handle the outliers through a secondary Semantic Outlier Alignment step. + +# 3.2.2. Semantic Outlier Alignment + +The key challenge in aligning outliers from $\chi$ to $\bar{\mathbf{X}}$ is that, $\bar{\mathcal{O}}$ only provides sparse correspondences to $\chi$ , which has a 3D + +point for every pixel. As such, some heuristics need to be developed to group regions of $\chi$ to move based on $\bar{\mathcal{O}}$ . Based on the observation that geometric inconsistencies between $\chi$ and $\bar{\mathbf{X}}$ tend to occur between objects and not within objects, we introduce an Interactive Segmentation Model (ISM) to get a series of masks according to $\bar{\mathcal{O}}$ in an iterative manner. + +First, we randomly select a point from the outlier SfM set, $\bar{\mathbf{X}}_k\in \bar{\mathcal{O}}$ Using Eq.2),we can get the projected 2D point $\bar{x}_k$ to prompt ISM, producing a binary mask $m_{k}\in \mathbb{R}^{H\times W}$ delineating a semantically coherent local region around this outlier point. We further define a threshold $T$ , which describes the desired number of correspondences to estimate a local alignment. If the number of outliers within $m_{k}$ is greater than $T$ i.e. $|m_k\bigcap \bar{P} (\bar{\mathcal{O}})| > T$ then we keep $m_{k}$ as it is and exclude within-mask outliers from being selected in the next iteration. Otherwise, we use all the outliers within the produced mask to prompt ISM again to update $m_{k}$ until the above criterion is met, or no further outliers can be included, in which case we discard this mask and the outliers within. We repeat the above process until no point in $\bar{\mathcal{O}}$ is left, eventually obtaining a set of semantically coherent masks around the outliers, $\{m_k\}_{k = 1}^M$ .Visualization of this process can be found in the supplementary materials. This iterative process is denoted as + +$\operatorname{Init}: m_k = \operatorname{ISM}(x_k)$ , s.t. $\bar{x}_k \in \bar{P}(\bar{\mathcal{O}})$ and $\bar{x}_k \notin \bigcup_{j=1}^{k-1} m_j$ + +Then, $m_{k} = \left\{ \begin{array}{ll}m_{k}, & \text{if} |m_{k}\bigcap \bar{P} (\bar{\mathcal{O}})| > T\\ \operatorname {ISM}(m_{k}\bigcap \bar{P} (\bar{\mathcal{O}})), & \text{otherwise} \end{array} \right.$ (4) + +Based on the outlier masks $\{m_k\}_{k=1}^M$ , we extract the within-mask outlier set $\mathcal{O}_k = P^{-1}(m_k \bigcap P(\mathcal{O}))$ and their corresponding SfM set is $\bar{\mathcal{O}}_k$ . We then estimate a local rigid transformation within each mask as follows: + +$$ +s _ {k}, R _ {k}, t _ {k} = \underset {s, R, t} {\arg \min } \sum_ {(\bar {\chi} _ {i}, \bar {\mathbf {X}} _ {i}) \in (\mathcal {O} _ {k}, \bar {\mathcal {O}} _ {k})} \| s R \bar {\chi} _ {i} + t - \bar {\mathbf {X}} _ {i} \| ^ {2}. \tag {5} +$$ + +Finally, we construct the inlier mask as $m_0 = \neg \left( \bigcup_{k=1}^{M} m_k \right)$ , where $\neg$ is the logical NOT operator. Then, the global transformation is applied to MASt3R's points within inlier mask $m_0$ and each individual local transformation is applied to those within each local mask $m_k, k \neq 0$ . This results in the final SfM-aligned point cloud, $\chi^*$ , where + +$$ +\chi^ {*} = \bigcup_ {k = 0} ^ {M} \left\{s _ {k} R _ {k} \chi_ {i} + t _ {k} \mid \chi_ {i} \in m _ {k} \right\}. \tag {6} +$$ + +To obtain the final dense and pose-wise accurate point cloud $\mathcal{X}$ as prior, we concatenate the final SfM-aligned point cloud with that of SfM, i.e. + +$$ +\mathcal {X} = \chi^ {*} \bigcup \bar {\mathbf {X}} \tag {7} +$$ + +# 3.2.3. Gaussian Optimization + +The SPARS3R constructed $\mathcal{X}$ can be directly used as the initialization for any Gaussian Splitting-based optimization method. Here we use Splatfacto, developed under the NeRFStudio framework [49]; the Gaussian optimization loss is: + +$$ +\mathcal {L} = \lambda_ {1} \| \tilde {C} - C \| _ {1} + \lambda_ {2} \mathrm {D} - \operatorname {S S I M} (\tilde {C}, C), \tag {8} +$$ + +where $\tilde{C}$ and $C$ denote our rendered and the groundtruth RGB image, D-SSIM denotes the structural similarity loss. + +# 4. Experiments + +Experiments are performed on three popular benchmark datasets. Tanks and Temples [27] contains 8 scenes captured in video format. Following InstantSplat [14], We uniformly sample 24 images with alternating indices for train and test split. MVImgNet [60] is an object-centric dataset that consists of 7 scenes. These scenes are captured from a consistent height covering $180^{\circ}$ field of view. We apply the same sampling and train-test split procedure as for Tanks and Temples. For sparser settings, we uniformly sample 3 and 6 views from the training set for the above two datasets. Mip-NeRF 360 [2] comprises of 9 scenes with $360^{\circ}$ views and greater pose variation between the scenes, including diverse heights and distances. The increased variability, combined with sparsity, adds complexity to accurate pose estimation and scene reconstruction. For this dataset, we follow the test set outlined in MipNeRF360 [2] and uniformly sample 12 images from the original training set to construct a sparse-view set. + +All sparse-view input images go through a COLMAP [46] pipeline for camera calibration. Specifically, we use feature matching results from MASt3R [29] for SfM triangulation. Since sparse-view registration can be unstable due to limited pairs, we perform multiple SfMs and pick the outcome that maximizes successful triangulation per image. + +# 4.1. Sparse NVS Evaluation + +Accurately and holistically evaluating sparse NVS should involve many practical considerations. Various NeRF-based approaches [12, 23, 25, 39, 51, 57, 59] have assumed perfect camera poses, i.e. calibration results from dense views, for sparse-view scenarios, which is unrealistic. Additionally, Gaussian-Splatting-based methods rely heavily on SfM point cloud, further necessitating the separation of train and test views from the registration stage. Previous approaches [65] have tried to perform re-triangulation based on known train poses, excluding test-set point cloud from training, but do not account for pose inaccuracy in sparse-view registration. While InstantSplat [14] adopts the realistic setup where they use recovered poses from sparse + +views and initialize the test camera poses by applying Procrustes Alignment [22] between train-only poses and train-test-together poses, their large camera alignment error leads to significant disruption in common pixel-based rendering metrics, e.g. PSNR and SSIM [53]. + +
Mip-NeRF 360ER(μ)↓ET(μ)↓
Procrustes Alignment0.1960.0144
+ RANSAC0.1790.0114
+ Rotation Points0.1560.0117
+ +Table 1. Improvements over Procrustes Alignment baseline in average rotation error $E_{R}$ and translation error $E_{T}$ . Incorporating rotation points further minimizes the overall error. + +We also operate under the realistic setup but improve upon the camera alignment process in two aspects. First, during camera alignment, instead of relying only on camera positions, we sample three additional points $P_{\mathrm{rotation}} \in \mathbb{R}^{3 \times 3}$ per camera along the unit direction vectors in rotation matrix $R_{\mathrm{camera}} \in \mathbb{R}^{3 \times 3}$ at camera position $T_{\mathrm{camera}} \in \mathbb{R}^{1 \times 3}$ , i.e. + +$$ +P _ {\text {r o t a t i o n}} = s _ {\text {c l o u d}} R _ {\text {c a m e r a}} + T _ {\text {c a m e r a}}. \tag {9} +$$ + +The scale of the camera cloud is denoted by $s_{\mathrm{cloud}} = \sqrt{\sigma_x^2 + \sigma_y^2 + \sigma_z^2}$ , and $\sigma_x, \sigma_y,$ and $\sigma_z$ are the standard deviations of camera positions along $x, y,$ and $z$ axes respectively. The four points per camera enable alignment of both camera rotations and translations. Second, we add RANSAC [15] in the alignment process, minimizing the influence of outliers. We measure the alignment rotation error $E_R$ by the angular difference in degrees between two rotations represented by quaternions following + +$$ +E _ {R} = 2 \cdot \arccos \left(\frac {\operatorname {a b s} \left(q _ {1} \cdot q _ {2}\right)}{\| q _ {1} \| \cdot \| q _ {2} \|}\right), \tag {10} +$$ + +and the translation error $E_{T}$ by the Euclidean norm $\| \cdot \|$ between translation vectors. As shown in Tab. 1, these two improvements enhance camera alignment accuracy in both rotation and translation. Beyond accurate camera pose alignment, test pose optimization can also be applied between rendered and ground-truth images to minimize the pose error; however, such process is time-consuming and can get stuck if the initial displacement is too large, as is sometimes the case in InstantSplat [14]. + +Evaluation Metrics. In addition to the commonly used PSNR, SSIM, and LPIPS, we incorporate DreamSim [18] (DSIM) as an additional metric to assess rendering quality. DSIM captures human perceptual similarity by finetuning a combination of embeddings from visual foundation models based on human evaluations. Notably, it provides a robust assessment of rendering image quality under moderate pose shift, which frequently occurs in the realistic sparse-view + +settings. Further details on DSIM's robustness as an evaluation metric are provided in the supplementary materials. + +Implementation Details. We employ Segment-Anything Model [26] as the segmentation model in Semantic Outlier Alignment. The weights for the loss terms are $\lambda_{1} = 0.8$ , $\lambda_{2} = 0.2$ . For fair implementation and comparison, we employ test pose optimization for all baselines and SPARS3R for 500 steps to maximally remove the effect of shifted camera pose. All baselines are run until convergence. + +# 4.2. Ablation Studies + +We compare the different components in SPARS3R to demonstrate their effectiveness on the Mip-NeRF 360 [2] dataset: + +- SfM-initialized 3DGS: Optimizing 3DGS on a sparse SfM point cloud and SfM-estimated camera poses. +- MAST3R-initialized 3DGS: Optimizing 3DGS on a MAST3R-produced dense point cloud and camera poses. +- Global Fusion Alignment: Aligning the MAST3R-produced dense point cloud onto SfM point cloud with a global transform. Poses are taken from SfM. +- Semantic Outlier Alignment (SPARS3R): The full SPARS3R pipeline that aligns outliers through their 2D semantic masks. + +
SettingPSNRSSIMLPIPSDSIM
Initialization: SfM [46]
3DGS [24]16.60.3880.4580.232
Initialization: MASt3R [29]
3DGS [24]15.90.2930.4630.229
+ Global Fusion Alignment18.60.4860.3300.130
+ Semantic Outlier Alignment18.90.5000.3270.127
+ +As shown in Tab. 2, given only 12 images, sparse NVS is challenging both in registration and reconstruction. Specifically, we observe that the 3DGS based on SfM initialization performs better than MASt3R initialization. Clearly, render quality is affected if poses are inaccurate, despite locally coherent dense point cloud. The introduction of Global Fusion Alignment improves upon the naive baselines from either initializations by combining accurate pose from SfM and dense point cloud from MASt3R. For scenes that do not have significant depth discrepancies, Global Fusion Alignment works well. In scenarios such as the Bonsai scene in Fig. 3, MASt3R's relative depth between the table and the wall is inaccurate; therefore, Global Fusion Alignment alone leads to a broken geometry of the background. In comparison, our Semantic Outlier Alignment provide piece-wise alignments on the background to ad + +dress this relative discrepancy. Quantitatively, we observe a 1.4 dB improvement on the Bonsai scene. + +![](images/7e2051c47bc81a21888a4eea47eb0604d435477f42df0165a518cb9514b3974e.jpg) +Global Fusion Alignment (GFA) + +![](images/12362d3bf7151a9c2b1a262546a2ff5fd6b69d04d0fa6ba9f1aaf46adc241f3b.jpg) +Semantic Outlier Align (SOA) + +![](images/8c3da074c43d318456d1be6921fc64e0292bc955fd3ece68d7d31be218fa5237.jpg) +Aligned point cloud from GFA + +![](images/45a32b2e951d136f28955f084c97630f57b35dd89b2060dbe69069b25095ed9a.jpg) +Aligned point cloud from SOA +Figure 3. Visual comparison of SPARS3R with and without SOA. While the dense bonsai in the foreground is aligned with the sparse point cloud, depth differences are obvious. SOA successfully fixes such gap. + +Table 2. Ablation on key components of SPARS3R. The results are shown using PSNR, SSIM, LPIPS and DSIM based on the MipNeRF 360 [2]. + +
MethodsMip-NeRF 360Tanks & TemplesMVimgNet
RPEtRPErRPEtRPErRPEt RPEr
DUSt3R [52]2.0752.5840.5700.1430.438 0.421
MASt3R [29]1.1861.4900.2410.2480.208 0.304
InstantSplat [14]2.0492.5550.1510.0810.264 0.311
COLMAP + MASt3R0.2520.4120.1610.0930.075 0.078
+ +Table 3. Quantitative evaluation of pose accuracy across three datasets, Relative Translation Error $(\mathrm{RPE}_t)$ and Relative Rotation Error $(\mathrm{RPE}_r)$ [62] are calculated based on the normalized poses. + +Pose Estimation Accuracy. As shown in Tab. 3, we compare the pose estimation accuracy from different calibration approaches. Both DUSt3R [52] and MASt3R [29] have significantly larger errors compared to COLMAP [46] with MASt3R feature matching outputs. InstantSplat [14] uses DUSt3R's [52] dense point cloud and pose estimation and attempts to improve accuracy through a training pose optimization approach similar to BARF [32]. While it brings down the errors in some cases, such training pose optimization strategy does not work as well in more challenging datasets like Mip-NeRF 360. This highlights the difficulty in improving camera calibration jointly with dense forward render, considering the loss landscape is highly nonlinear based on pixel-wise metrics. + +# 4.3.Quantitative and Visual Evaluation + +We quantitatively compare SPARS3R against various NVS methods in Tab. 4 and Tab. 5. Instant-NGP [36] and 3DGS [24] are two mainstream NVS methods based on + +
MethodsMip-NeRF 360 [2]Tanks & Temples [27]MVImgNet [60]
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
Instant-NGP [36]14.820.2940.6780.52415.280.4510.3890.25413.280.4260.8920.892
3DGS [24]16.570.3880.4580.23221.070.7300.1940.07021.240.6730.2340.064
FSGS [65]17.600.4430.5580.24325.720.8450.1110.02323.430.7600.2120.042
SparseGS [55]16.660.4050.4610.21020.280.7270.2020.07520.560.6720.2480.072
DRGS [10]16.880.4010.6490.30021.460.7230.2890.07821.700.6410.4220.071
CF-3DGS [19]13.270.2500.6980.50918.990.6060.2960.12715.430.4080.5450.325
InstantSplat [14]16.230.3590.5430.23326.970.8740.1150.01323.220.7340.2480.028
SPARS3R18.850.5000.3270.12729.900.9190.0470.00725.850.8200.1140.011
+ +Table 4. Quantitative comparison of different NVS methods on 12 views on three popular benchmark datasets, totaling 24 scenes. All methods are run on the same registrations and updated with test pose optimization. + +
3 ViewsTanks & TemplesMVImgNet
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
FSGS20.360.6580.2190.07716.920.5120.4760.183
InstantSplat21.650.7550.1640.04416.290.5510.3230.103
SPARS3R21.830.7550.1500.04318.620.5880.2810.091
6 Views
FSGS23.900.7830.1520.04320.420.6440.3310.080
InstantSplat25.890.8470.1170.02221.090.6950.2360.046
SPARS3R26.800.8730.0730.01422.750.7320.1610.023
+ +Table 5. Quantitative comparison of 3 and 6 views on Tanks & Temples and MVImgNet datasets. + +implicit and explicit scene representations. Specifically, 3DGS leverages SfM points from registration as the starting point for scene optimization, while NeRF [35] based methods rely only on poses. Colmap-Free 3DGS [19] also does not rely on prior registration and optimizes cameras along with reconstruction. We find that both Instant-NGP and CF-3DGS perform significantly worse than other methods that use a prior SfM point cloud. + +Among methods that are designed for sparse NVS, DRGS [11], SparseGS [55], and FSGS [65] leverage monocular depth prior to gain more geometric information of the scene. DRGS applies an additional smoothness constraint, which leads to high PSNR values but worse LPIPS and DSIM performances. SparseGS additionally applies a Score Distillation Sampling Loss, which generates novel views and constrained them through a prior generative model. Its results are sharper than DRGS, but more noisy, leading to worse PSNR. FSGS seeks to address the sparse point cloud issue in sparse-view NVS by modifying the densification algorithm in 3DGS. More Gaussians are encouraged between the space of two existing Gaussians, allowing representative areas to densify quicker. FSGS has overall better metrics than SparseGS and DRGS; however, it still encounters issues in regions where few initializations are available, e.g. the background of the scene. As shown in Fig. 4 and in Fig. 1, this leads to various blurry regions. InstantSplat [14] is a concurrent work that uses prior depth estimation from DUSt3R [52] and the dense point cloud to help with sparse- + +view NVS. While the initialized point cloud is much denser than FSGS, InstantSplat also inherits the suboptimal poses produced by DUSt3R. The incorrect geometry limits the effectiveness of densification in 3DGS and can lead to generation of floaters; to ameliorate this problem, InstantSplat disables Gaussian densification and optimizes the scene with few steps. As shown in Tab. 3 and Tab. 4, on the Mip-NeRF 360 and MVImgNet dataset, where pose accuracy from DUSt3R is limited, InstantSplat performs on-par with FSGS despite using a denser prior point cloud. As shown in Fig. 4, we observe that the Gaussians tend to be noisy on the surface. In the Stump and Bicycle scenes, InstantSplat's results cannot be rigidly aligned well with the groundtruth test poses due to the large offset to begin with. + +Our SPARS3R method improves upon DUSt3R and MASt3R's prior point cloud by adjusting it based on conventional SfM calibration approaches. Since SfM only triangulates based on confident keypoints, its pose and depth estimation performance, particularly under sparse-view scenario, is more reliable. As shown in Tab. 4, SPARS3R consistently improves upon previous best approach by an average of $2.7\mathrm{dB}$ . SPARS3R is stable in its optimization process and does not generate additional floaters with more iterations. As seen in Fig. 3, SPARS3R produces rendering with smooth surfaces in the foreground, and sharp details in the background. + +For even sparser settings, we provide a quantitative comparison in Tab. 5 between SPARS3R and the two best-performing methods from Tab. 4, FSGS [65] and InstantSplat [14]. FSGS struggles significantly under extreme sparsity due to its reliance on a sparse initialization, resulting in DSIM scores that are approximately two- and four-times worse than those of SPARS3R under the 3-view and 6-view settings, respectively. InstantSplat generally performs better than FSGS, benefiting from a much denser point cloud initialization produced by DUSt3R [52], which is particularly advantageous under extreme sparse conditions. However, its performance remains inferior to SPARS3R due to suboptimal pose estimation. Our + +![](images/afb8b32a21e1c4cbe6e7f63dd2fa11c336b835bcbf35a8b69a5355b9b84bf8a5.jpg) +Figure 4. Visual comparisons of different NVS methods on 12 views on Mip-NeRF 360 [2] dataset. Zooming in on the visualizations is recommended to show differences in detail. More visualizations for other datasets are available in the Supplemental Material. + +SPARS3R continues to outperform these methods, improving upon the best approach by an average of 0.9 dB and 1.3 dB in the 3-view and 6-view settings, respectively. These results further demonstrate SPARS3R's superiority over state-of-the-art approaches in handling extreme sparse-view reconstruction. For more visualizations and flythroughs of the scene, please refer to the supplemental materials. + +# 4.4. Limitations + +While SPARS3R significantly improves upon previous SoTA, there are also several limitations worth noting. Firstly, SPARS3R relies on a well-performing semantic segmentation model to group outliers; as such, if the segmentation model is very sensitive, i.e. separating a connected surface into multiple disjointed parts, some of these parts may lack sufficient SfM support for alignment. If the segmentation model produces masks that group multiple areas with disparate depth together, the local alignments are also less effective similar to the global alignment. Secondly, non-rigid transform between two point clouds are worth further investigation. While SPARS3R's piece-wise rigid transform based on semantic relationship is fast and addresses the inaccurate inter-object depth estimation, a non-rigid formulation may be more generalizable. + +# 5. Conclusion + +We present SPARS3R, a scene reconstruction and NVS method that can achieve high-quality rendering with sparse input images. We demonstrate that existing methods lead to blurry rendering due to sparse point cloud initialization. Recent progress in generating a dense point cloud from prior depth estimation model, while promising, can lead to noisy pose estimation. To address these issues, SPARS3R combines the advantages of both by proposing a two-step alignment approach. The first step computes a global transformation between the dense point cloud from depth prior and a reference SfM point cloud. The second step takes the semantic regions around the outlier points identified from the first step and performs successive local alignments. This approach successfully addresses the depth estimation errors within the prior dense point cloud and demonstrates that the updated point cloud leads to significantly better performances in sparse-view NVs. We also introduce several improvements in the evaluation process to better represent the practical limitations in sparse-view registration and reconstruction. In the future, we hope to further improve the smoothness of our alignment process, e.g. by exploring potential non-rigid transformation approaches. + +# Acknowledgement + +This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 5, 6, 7, 8 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2 +[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2 +[6] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19436-19446, 2024. 2 +[7] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2 +[8] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12952, 2022. 2 +[9] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2 +[10] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 7 +[11] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 2, 7 +[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5 +[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with $15\mathrm{x}$ reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 2 +[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 1, 2, 5, 6, 7 +[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5 +[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2 +[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2 +[18] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data, 2023. 5 +[19] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 7 +[20] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021. 2 +[21] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2 +[22] John C Gower. Generalized procrustes analysis. Psychometrika, 40:33-51, 1975. 3, 5 +[23] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 5 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6, 7 +[25] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infererf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022. 1, 5 + +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 6 +[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7 +[28] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719-21728, 2024. 2 +[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 4, 5, 6 +[30] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 1 +[31] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2 +[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 6 +[33] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 2 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 7 +[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1, 2, 6, 7 +[37] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024. 2 + +[38] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349-10358, 2024. 2 +[39] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1, 5 +[40] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatt-ting for robust real-time rendering with $900+$ fps. arXiv preprint arXiv:2403.13806, 2024. 2 +[41] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In European Conference on Computer Vision, pages 19-37. Springer, 2025. 2 +[42] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2 +[43] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In European Conference on Computer Vision, pages 293-310. Springer, 2024. 2 +[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2 +[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5, 6 +[47] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2 +[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2 + +[49] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5 +[50] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2 +[51] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1, 5 +[52] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7 +[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[54] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2 +[55] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 $\{\backslash \mathrm{deg}\}$ sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 1, 2, 7 +[56] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 2 +[57] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 1, 5 +[58] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15901-15911, 2023. 2 +[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 5 +[60] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 5, 7 +[61] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[62] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual(-inertial) odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7244-7251, 2018. 6 +[63] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2 +[64] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2 +[65] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1, 2, 5, 7 \ No newline at end of file diff --git a/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/images.zip b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..68847fd30cb547a5b119b976571c6a453eefe644 --- /dev/null +++ b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec924437683e289db2174e568d3c8c07b4b18bf9c8f2399a21b86506a0982f96 +size 750902 diff --git a/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/layout.json b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6059d14b4f5d10d0296962ab6983df2ec1dac5b7 --- /dev/null +++ b/2025/SPARS3R_ Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction/layout.json @@ -0,0 +1,10703 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 526, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 526, + 138 + ], + "type": "text", + "content": "SPARS3R: Semantic Prior Alignment and Regularization for Sparse 3D Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 161, + 208, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 161, + 208, + 175 + ], + "spans": [ + { + "bbox": [ + 146, + 161, + 208, + 175 + ], + "type": "text", + "content": "Yutao Tang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 161, + 299, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 161, + 299, + 175 + ], + "spans": [ + { + "bbox": [ + 228, + 161, + 299, + 175 + ], + "type": "text", + "content": "Yuxiang Guo*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 161, + 378, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 161, + 378, + 175 + ], + "spans": [ + { + "bbox": [ + 323, + 161, + 378, + 175 + ], + "type": "text", + "content": "Deming Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 402, + 161, + 463, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 161, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 402, + 161, + 463, + 175 + ], + "type": "text", + "content": "Cheng Peng" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 241, + 178, + 369, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 178, + 369, + 192 + ], + "spans": [ + { + "bbox": [ + 241, + 178, + 369, + 192 + ], + "type": "text", + "content": "Johns Hopkins University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 193, + 194, + 414, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 194, + 414, + 206 + ], + "spans": [ + { + "bbox": [ + 193, + 194, + 414, + 206 + ], + "type": "text", + "content": "{ytang67，yguo87，dli90，cpeng26}@jhu.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 233, + 200, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 233, + 200, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 233, + 200, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 258, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 510 + ], + "type": "text", + "content": "Recent efforts in Gaussian-Splat-based Novel View Synthesis can achieve photorealistic rendering; however, such capability is limited in sparse-view scenarios due to sparse initialization and over-fitting floaters. Recent progress in depth estimation and alignment can provide dense point cloud using few views; however, the resulting pose accuracy is suboptimal. In this work, we present SPARS3R, which combines the advantages of accurate pose estimation from Structure-from-Motion and dense point cloud from depth estimation. To this end, SPARS3R first performs a Global Fusion Alignment process that maps a prior dense point cloud to a sparse point cloud from Structure-from-Motion based on triangulated correspondences. RANSAC is applied during this process to distinguish inliers and outliers. SPARS3R then performs a second, Semantic Outlier Alignment step, which extracts semantically coherent regions around the outliers and performs local alignment in these regions. Along with several improvements in the evaluation process, we demonstrate that SPARS3R can achieve photorealistic rendering with sparse images and significantly outperforms existing approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 552, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 671 + ], + "type": "text", + "content": "Photorealistic scene reconstruction and Novel View Synthesis (NVS) from unposed 2D images is a challenging task with numerous applications in site modeling, autonomous driving, robotics, urban and agricultural planning, etc. The introduction of methods such as Neural Radiance Field (NeRF) [35] and 3D Gaussian Splatting (3DGS) [24] have introduced significant advances in rendering quality and efficiency based on dense multi-view imagery. However, applications of these methods still face issues in practical scenarios where a dense coverage of the scene is not available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "Given sparse images, over-fitting the photometric objectives to an incorrect geometry is a common issue in" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 233, + 554, + 312 + ], + "blocks": [ + { + "bbox": [ + 316, + 233, + 554, + 312 + ], + "lines": [ + { + "bbox": [ + 316, + 233, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 233, + 554, + 312 + ], + "type": "image", + "image_path": "8a784259f4c66dd91775151c1516e7a0efa985b79a39fb642e70049b092546f2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 312, + 405, + 322 + ], + "lines": [ + { + "bbox": [ + 347, + 312, + 405, + 322 + ], + "spans": [ + { + "bbox": [ + 347, + 312, + 405, + 322 + ], + "type": "text", + "content": "Instant-NGP [36]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 312, + 512, + 322 + ], + "lines": [ + { + "bbox": [ + 475, + 312, + 512, + 322 + ], + "spans": [ + { + "bbox": [ + 475, + 312, + 512, + 322 + ], + "type": "text", + "content": "FSGS [65]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 325, + 554, + 404 + ], + "blocks": [ + { + "bbox": [ + 316, + 325, + 554, + 404 + ], + "lines": [ + { + "bbox": [ + 316, + 325, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 554, + 404 + ], + "type": "image", + "image_path": "7186186877bf36149aa9b12a105ead39acc7b6a2c3d5c1eac9b2e2080ae13c07.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 404, + 403, + 414 + ], + "lines": [ + { + "bbox": [ + 347, + 404, + 403, + 414 + ], + "spans": [ + { + "bbox": [ + 347, + 404, + 403, + 414 + ], + "type": "text", + "content": "InstantSplat [14]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 404, + 511, + 413 + ], + "lines": [ + { + "bbox": [ + 476, + 404, + 511, + 413 + ], + "spans": [ + { + "bbox": [ + 476, + 404, + 511, + 413 + ], + "type": "text", + "content": "SPARS3R" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 426, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. A visualization of SPARS3R in comparison to current SoTA. Without additional prior, sparse NVS leads to incorrect geometry by Instant-NGP [36]. FSGS [65] can be blurry due to sparse initialization and insufficient densification. InstantSplat [14] relies on DUSt3R [52] initialization with suboptimal poses. Our method, SPARS3R, can reliably render details in the foreground and background with accurate poses." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "type": "text", + "content": "NVS. Various constraints, such as semantic consistent loss [23], depth and novel-view regularization [12, 39, 51], frequency regularization [57], and ray entropy minimization [25] have been introduced to improve upon NeRF. These methods often lead to significant computation overhead due to the costly patch rendering from ray-tracing. More recently, Gaussian-Splatting-based approaches have further improved sparse NVS by leveraging the explicit representation and fast differentiable rasterization. Depth regularization [30], Gaussian floater pruning [55], and proximity-based Gaussian densification strategy [65] have been proposed to constrain and guide the scene structures. As demonstrated in Fig. 1, despite improved geometry, these approaches tend to produce overly smooth rendering in the background. This issue can be attributed to a sparse initial" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26810" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "point cloud, particularly in the background regions. Furthermore, the additional constraints, e.g. based on monocular depth prior, are oftentimes imperfect, leading to noisy gradients that prevent proper densification in sparse regions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 295, + 372 + ], + "type": "text", + "content": "One potential solution to address this conundrum is to initialize Gaussian optimization with a denser point cloud to help disambiguate shape and radiance. To this end, recent advances in two-view depth estimation, particularly DUSt3R [52] and MAST3R [29], have shown the impressive ability to construct a dense point cloud from a pre-trained prior model. In theory, such a point cloud can be directly used to optimize a Gaussian-Splat radiance field [14]. In practice, camera calibration obtained from multi-view depth alignment is often suboptimal due to the difficulties in estimating an accurate depth map. If left uncorrected, the Gaussian optimization process will generate floaters to compensate the suboptimal calibration, prompting strategies to prevent densification as a stopgap measure [14]. In comparison, camera calibration based on Structure-from-Motion (SfM), e.g., COLMAP [46], is fast and accurate with an appropriate feature matching model. Instead of aligning dense depth, SfM takes confident correspondences and performs triangulation with RANSAC [15] to reject outliers. In a sparse-view setting, such a process can be repeated numerous times to further improve accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 373, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 295, + 636 + ], + "type": "text", + "content": "To address sparse point cloud initialization and pose inaccuracy in sparse-view NVS, we propose SPARS3R. SPARS3R composes of two stages: Global Fusion Alignment and Semantic Outlier Alignment. In Global Fusion Alignment, SPARS3R first obtains a dense point cloud from sparse views through DUSt3R [52] or MASt3R [29], and a sparse point cloud from COLMAP [46]. By taking the triangulated correspondences within images, SPARS3R fuses the dense point cloud onto the sparse point cloud through a global Procrustes Alignment process with RANSAC [15]. Since some points will yield large error due to local scale variations, a second Semantic Alignment process is introduced to perform local alignment. Specifically, outliers from Global Fusion Alignment are identified and used as prompts for an Interactive Segmentation model, e.g. SAM [26], generating semantically coherent masks around the outliers. The resulting semantic masks indicate regions within dense point cloud to be treated with local alignment. By locally aligning the identified outliers with the SfM point cloud while preserving the globally aligned inliers, we obtain a dense and pose-wise accurate point cloud, serving as a strong prior for Gaussian optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 638, + 284, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 284, + 649 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 284, + 649 + ], + "type": "text", + "content": "In summary, our contributions can be summarized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 654, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 295, + 700 + ], + "type": "text", + "content": "1. We propose a Global Fusion Alignment approach, which transforms a prior dense point cloud onto a reference SfM sparse point cloud, putting dense initialization and accurate camera poses in the same coordinate frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "2. To address outliers that cannot be aligned accurately due" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 131 + ], + "type": "text", + "content": "to depth discrepancies, we propose a Semantic Outlier Alignment step. This step extracts semantically similar regions around the outliers to perform local alignment, resulting in a dense point cloud with minimum transformation error." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "content": "3. We evaluate the overall method, SPARS3R, on three popular benchmark datasets and find significant quantitative and visual improvements compared to current SoTA methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 185, + 400, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 185, + 400, + 197 + ], + "spans": [ + { + "bbox": [ + 314, + 185, + 400, + 197 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 202, + 525, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 202, + 525, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 202, + 525, + 215 + ], + "type": "text", + "content": "2.1. 3D Models for Synthesizing Novel Views" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 553, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 423 + ], + "type": "text", + "content": "Photorealistic scene reconstruction and novel view synthesis is a long-standing task in computer vision and graphics. Neural Radiance Field (NeRF) [35] proposes to model a scene implicitly with multi-layer perceptron (MLP) by mapping 3D coordinates and view direction to color and density, then rendering pixel value via alpha blending. Since NeRF's introduction, numerous work has sought to improve its efficiency [5, 16, 20, 36, 45, 48], quality [1-3, 7, 50], and extension to dynamic and in-the-wild [4, 6, 8, 17, 34, 42, 44, 47, 54, 58] scenarios. Recently, 3D Gaussian Splatting (3DGS) [24] has emerged as an efficient alternative, offering improved rendering quality over NeRF. Unlike NeRF, 3DGS represents scenes with explicit 3D Gaussians kernels and uses a differentiable rasterization technique. Following its introduction, many methods have emerged to reduce the computational cost [13, 21, 28, 37, 38] and enhance quality [9, 33, 40, 43, 56, 61, 63, 64] of 3DGS." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 426, + 542, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 542, + 439 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 542, + 439 + ], + "type": "text", + "content": "2.2. Gaussian-Splatting-based Sparse-View NVs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 444, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 553, + 647 + ], + "type": "text", + "content": "While 3DGS performs well with dense view support, its effectiveness decreases in practical settings with sparse inputs. Many approaches address this through depth constraints. FSGS [65] uses a Proximity-guided Gaussian Unpooling to mitigate the sparse initialization issue and introduces pseudo-views during training to avoid overfitting. DRGS [11] and SparseGS [55] add a depth regularization term to enforce consistency between estimated and monocular depths while promoting smoothness. DNGS [31] refines depth regularization by prioritizing local depth variations. CoherentGS [41] initializes Gaussians from monocular depth estimation and improves coherence through an optical flow constraint. InstantSplat [14] is a concurrent work that directly uses DUSt3R to produce dense point clouds and introduces a camera pose optimization strategy; densification of Gaussians is disabled to prevent the introduction of floaters due to suboptimal poses." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 654, + 374, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 654, + 374, + 666 + ], + "spans": [ + { + "bbox": [ + 314, + 654, + 374, + 666 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 672, + 392, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 672, + 392, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 672, + 392, + 685 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "DUSt3R [52] is a two-view depth estimation method that produces dense 3D point clouds from image pairs. Specif" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26811" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "type": "image", + "image_path": "946e9e00b1cae5139d38a7f6349ff06938ab3bd9742f08cb17fdea43306eb754.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. SPARS3R combines a prior dense point cloud " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " and a sparse SfM point cloud " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": ". The prior " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " often has inferior depth accuracy compared to " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": ". SPARS3R first globally aligns all points in " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": ", based on shared correspondences. Inliers and outliers are identified through alignment error. SPARS3R then extracts the semantically relevant 2D regions around the outliers to move local regions of " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " in groups, producing a dense point cloud " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\chi^{*}" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " that is depth-wise and pose-wise accurate. By providing " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\chi^{*} \\bigcup \\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " for Gaussian optimization, SPARS3R achieves photorealistic rendering under sparse-view condition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": "ically, given an image pair " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "I^1" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "I^2" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", DUSt3R generates dense pointmaps " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1,1}" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{2,1} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ". A pointmap, e.g. " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1,1}" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", provides a mapping from each pixel " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "I^1" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " to a corresponding 3D point " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{1,1}(u, v)" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " under the coordinate system of image " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "I^1" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", giving dense 2D-to-3D correspondences. Given " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "K > 2" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " input images, DUSt3R [52] aggregates across all pairwise pointmap predictions by globally aligning pairwise pointmaps into a unified point cloud " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ". This alignment process also produces camera parameters " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\{W(I^n), J(I^n)\\}" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ", which are the world-to-camera transformation matrix and the Jacobian of the affine approximation of the projective transformation for view " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": ". As such, the relationship between a pixel " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " and its 3D point " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\chi_i" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " is established via a projective function " + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 490, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 490, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 88, + 490, + 294, + 506 + ], + "type": "interline_equation", + "content": "x _ {i} = [ u, v ] ^ {T} = P \\left(\\chi_ {i}\\right) = J \\left(I ^ {n}\\right) W \\left(I ^ {n}\\right) \\chi_ {i}. \\tag {1}", + "image_path": "9db9fc0e4c90c2b640c30d8cf92b68fb6a5adebab1b794e314ad1676a7941bff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 593 + ], + "type": "text", + "content": "Building on this, MASt3R [29] incorporates an additional matching head that generates discriminative local feature descriptors and pixel-level correspondences across paired images. We note that while MASt3R [29] uses this feature correspondence during its multi-depth alignment, the estimated poses are oftentimes inaccurate due to smooth depth predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 126, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 126, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 126, + 611 + ], + "type": "text", + "content": "3.2. SPARS3R" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "As show in Fig. 2, SPARS3R builds upon the advancement in DUSt3R [52] and MASt3R [29] as a pre-trained 3D prior for sparse-view reconstruction. Firstly, SPARS3R performs SfM based on image correspondences, either from MASt3R [29] or other feature matching methods. SPARS3R then aligns the dense point cloud produced by MASt3R via rigid transformations in two stages: Global Fusion Alignment and Semantic Outlier Alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 450, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 450, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 450, + 324 + ], + "type": "text", + "content": "3.2.1. Global Fusion Alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": "While DUSt3R [52] and MASt3R [29] can generate dense 3D points from sparse images, it does not estimate accurate or detailed depth through its pointmap formulation. As a result, the camera calibration obtained from such smooth depths is suboptimal and will degrade the fidelity of scene reconstructions. To construct a better point cloud prior, we propose to align MASt3R's point cloud with that from a SfM pipeline, which is more reliable based on selective correspondences. This alignment process begins by identifying matching points between the SfM and MASt3R point clouds, denoted as " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}\\in \\mathbb{R}^{\\bar{N}\\times 3}" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\chi \\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": ", respectively. A visibility indicator " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "V:\\mathbb{R}^{\\bar{N}\\times 3}\\to \\mathbb{R}^{\\widetilde{N}\\times 3}" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " denotes if an SfM point " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{X}}_i" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " is visible with respect to a view " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{N}" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " denotes the number of visible SfM points in view " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": ". For each view, similar to Eq. (1), we can obtain the pixel-to-3D relationship of " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "(\\bar{u},\\bar{v})" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "V(\\overline{\\mathbf{X}})_i" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " through estimated " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\overline{W}(I^n)" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\bar{J}(I^n)" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " from SfM. Subsequently, the correspondence between a SfM point " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "V(\\overline{\\mathbf{X}})_i" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": " and a point in " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{\\chi}_i" + }, + { + "bbox": [ + 313, + 327, + 555, + 558 + ], + "type": "text", + "content": ", can be established through a forward-backward projection:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 336, + 574, + 553, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 574, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 336, + 574, + 553, + 596 + ], + "type": "interline_equation", + "content": "\\bar {x} _ {i} = \\left[ \\bar {u}, \\bar {v} \\right] ^ {T} = \\bar {P} (\\bar {\\mathbf {X}}) _ {i} = \\bar {J} (I ^ {n}) \\bar {W} (I ^ {n}) V (\\bar {\\mathbf {X}}) _ {i}, \\tag {2}", + "image_path": "41d3c7dbfd1eddc4e1a98e0d50f701bcdc01542022c16d3ea2772cd86ae71edc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 592, + 514, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 592, + 514, + 605 + ], + "spans": [ + { + "bbox": [ + 354, + 592, + 514, + 605 + ], + "type": "interline_equation", + "content": "\\bar {\\chi} _ {i} = P ^ {- 1} (\\bar {x} _ {i}) = W ^ {- 1} \\left(I ^ {n}\\right) J ^ {- 1} \\left(I ^ {n}\\right) \\bar {x} _ {i}.", + "image_path": "9961596f51184ed918c9c5b331b97dd7278809b75241090e23627a4bb78f3db0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "content": "For each view, we denote the ordered set of visible SfM points as " + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{V(\\bar{\\mathbf{X}})_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "content": ", the projected 2D pixels as " + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\bar{x}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "content": ", and the set of 3D points that correspond to the projected 2D pixels above in " + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\chi}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "content": ", omitting view indicator " + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 615, + 555, + 713 + ], + "type": "text", + "content": ". With these identified correspondences, we can apply Generalized Procrustes Analysis [22] to estimate a global rigid transformation, optimizing for scale, rotation, and translation. During this process, RANSAC is applied to filter out-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26812" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "liers, excluding their influence on the estimation. Specifically, we minimize the following objective to get the optimal global transformation based on Algorithm 1:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 125, + 295, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 125, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 77, + 125, + 295, + 159 + ], + "type": "interline_equation", + "content": "s _ {0}, R _ {0}, t _ {0} = \\underset {s, R, t} {\\arg \\min } \\sum_ {i = 1} ^ {\\widetilde {N}} e \\left(\\bar {\\chi} _ {i}, s, R, t, V (\\bar {\\mathbf {X}}) _ {i}\\right) \\tag {3}", + "image_path": "099cb5b17f419d51a99c4370932c02b41ddd947e5fad26388c6c8d984f356acf.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 161, + 268, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 161, + 268, + 177 + ], + "spans": [ + { + "bbox": [ + 73, + 161, + 268, + 177 + ], + "type": "interline_equation", + "content": "e (\\bar {\\chi} _ {i}, s, R, t, V (\\bar {\\mathbf {X}}) _ {i}) = \\left\\| s R \\bar {\\chi} _ {i} + t - V (\\bar {\\mathbf {X}}) _ {i} \\right\\| ^ {2}.", + "image_path": "d7e03362d5c28e31e15f5130c8feada4259ca3b7c9b774b74686d0dc9c5e0afb.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 198, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 198, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 198, + 208 + ], + "type": "text", + "content": "Algorithm 1 RANSAC Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 213, + 295, + 536 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "text", + "content": "1: Input: Matched MASt3R points " + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\{\\overline{\\chi}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "text", + "content": " and SfM points " + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\{V(\\bar{\\mathbf{X}})_i\\}_{i=1}^N" + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "text", + "content": ", sample size " + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "text", + "content": ", error threshold " + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "text", + "content": ", number of iterations " + }, + { + "bbox": [ + 62, + 213, + 295, + 251 + ], + "type": "inline_equation", + "content": "R" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "text", + "content": "2: Output: Best parameters " + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "inline_equation", + "content": "(s_0,R_0,t_0)" + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "text", + "content": ", inlier set " + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "text", + "content": ", outlier set " + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "text", + "content": ", corresponding outlier SfM set " + }, + { + "bbox": [ + 61, + 253, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "text", + "content": "3: Initialize: num_inliers " + }, + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + }, + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "inline_equation", + "content": "(\\hat{s_0},\\hat{R_0},\\hat{t_0})\\gets \\emptyset ,\\hat{\\mathcal{I}}\\gets \\emptyset ,\\hat{\\mathcal{O}}\\gets \\emptyset ,\\hat{\\bar{\\mathcal{O}}}\\gets \\emptyset" + }, + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "text", + "content": " Iterations " + }, + { + "bbox": [ + 62, + 277, + 295, + 303 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 304, + 173, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 173, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 173, + 313 + ], + "type": "text", + "content": "4: while Iterations " + }, + { + "bbox": [ + 62, + 304, + 173, + 313 + ], + "type": "inline_equation", + "content": "< R" + }, + { + "bbox": [ + 62, + 304, + 173, + 313 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "type": "text", + "content": "5: Randomly select " + }, + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "type": "text", + "content": " points as sample set " + }, + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "type": "inline_equation", + "content": "S \\subset \\{i\\}_{i=1}^{\\widetilde{N}}" + }, + { + "bbox": [ + 62, + 313, + 290, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "text", + "content": "6: Estimate " + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "inline_equation", + "content": "(\\hat{s}_0,\\hat{R}_0,\\hat{t}_0)" + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "text", + "content": " on sample set " + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "inline_equation", + "content": "\\sum_{i\\in \\mathcal{S}}e(\\bar{\\chi}_i,s,R,t,V(\\bar{\\mathbf{X}})_i)" + }, + { + "bbox": [ + 62, + 328, + 295, + 353 + ], + "type": "text", + "content": " (Eq.3))" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 353, + 167, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 167, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 167, + 364 + ], + "type": "text", + "content": "7: for each point " + }, + { + "bbox": [ + 62, + 353, + 167, + 364 + ], + "type": "inline_equation", + "content": "\\overline{\\chi}_i" + }, + { + "bbox": [ + 62, + 353, + 167, + 364 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 364, + 239, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 239, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 239, + 376 + ], + "type": "text", + "content": "8: if " + }, + { + "bbox": [ + 62, + 364, + 239, + 376 + ], + "type": "inline_equation", + "content": "e(\\overline{\\chi}_i, \\hat{s}_0, \\hat{R}_0, \\hat{t}_0, V(\\overline{\\mathbf{X}})_i) < \\epsilon" + }, + { + "bbox": [ + 62, + 364, + 239, + 376 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 376, + 190, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 190, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 190, + 388 + ], + "type": "text", + "content": "9: Add " + }, + { + "bbox": [ + 62, + 376, + 190, + 388 + ], + "type": "inline_equation", + "content": "\\bar{\\chi}_i" + }, + { + "bbox": [ + 62, + 376, + 190, + 388 + ], + "type": "text", + "content": " to inlier set " + }, + { + "bbox": [ + 62, + 376, + 190, + 388 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 388, + 111, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 388, + 111, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 388, + 111, + 399 + ], + "type": "text", + "content": "10: else" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "text", + "content": "11: Add " + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\bar{\\chi}_i" + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "text", + "content": " to outlier set " + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}" + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "text", + "content": " and add " + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "inline_equation", + "content": "V(\\overline{\\mathbf{X}})_i" + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "text", + "content": " to outlier SfM set " + }, + { + "bbox": [ + 59, + 399, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 426, + 120, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 426, + 120, + 436 + ], + "spans": [ + { + "bbox": [ + 59, + 426, + 120, + 436 + ], + "type": "text", + "content": "12: end if" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 439, + 116, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 439, + 116, + 448 + ], + "spans": [ + { + "bbox": [ + 59, + 439, + 116, + 448 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 450, + 186, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 450, + 186, + 461 + ], + "spans": [ + { + "bbox": [ + 59, + 450, + 186, + 461 + ], + "type": "text", + "content": "14: if " + }, + { + "bbox": [ + 59, + 450, + 186, + 461 + ], + "type": "inline_equation", + "content": "|\\mathcal{I}| > \\text{num\\_inliers}" + }, + { + "bbox": [ + 59, + 450, + 186, + 461 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 462, + 295, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 462, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 59, + 462, + 295, + 488 + ], + "type": "text", + "content": "15: Update: num_inliers \\(\\leftarrow |\\hat{\\mathcal{I}}|, (s_0, R_0, t_0) \\leftarrow (\\hat{s}_0, \\hat{R}_0, \\hat{t}_0), \\mathcal{I} \\leftarrow \\hat{\\mathcal{I}}, \\mathcal{O} \\leftarrow \\hat{\\mathcal{O}}, \\bar{\\mathcal{O}} \\leftarrow \\hat{\\bar{\\mathcal{O}}}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 490, + 110, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 490, + 110, + 498 + ], + "spans": [ + { + "bbox": [ + 59, + 490, + 110, + 498 + ], + "type": "text", + "content": "16: end if" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 501, + 190, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 501, + 190, + 511 + ], + "spans": [ + { + "bbox": [ + 59, + 501, + 190, + 511 + ], + "type": "text", + "content": "17: Iterations " + }, + { + "bbox": [ + 59, + 501, + 190, + 511 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 59, + 501, + 190, + 511 + ], + "type": "text", + "content": " Iterations " + }, + { + "bbox": [ + 59, + 501, + 190, + 511 + ], + "type": "inline_equation", + "content": "+1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 514, + 116, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 514, + 116, + 522 + ], + "spans": [ + { + "bbox": [ + 59, + 514, + 116, + 522 + ], + "type": "text", + "content": "18: end while" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 525, + 189, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 525, + 189, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 525, + 189, + 536 + ], + "type": "text", + "content": "19: return " + }, + { + "bbox": [ + 59, + 525, + 189, + 536 + ], + "type": "inline_equation", + "content": "(s_0, R_0, t_0), \\mathcal{I}, \\mathcal{O}, \\mathcal{O}" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "content": "After this, we extract an inlier set " + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "content": ", an outlier set " + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "content": " and a corresponding outlier SfM set " + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}" + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "content": " from RANSAC with an empirical threshold. As shown in Fig. 2, the global transformation recovered from Eq. 3 often leaves large amount of outliers between " + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 55, + 550, + 295, + 670 + ], + "type": "text", + "content": ", particularly for objects that are far apart from the foreground. We observe that depth generated from DUSt3R [52] or MAST3R [29] has a strong bias towards smoothness between objects, despite local coherence. Therefore, we handle the outliers through a secondary Semantic Outlier Alignment step." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 674, + 204, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 204, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 204, + 686 + ], + "type": "text", + "content": "3.2.2. Semantic Outlier Alignment" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The key challenge in aligning outliers from " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is that, " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " only provides sparse correspondences to " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": ", which has a 3D" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "point for every pixel. As such, some heuristics need to be developed to group regions of " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to move based on " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". Based on the observation that geometric inconsistencies between " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " tend to occur between objects and not within objects, we introduce an Interactive Segmentation Model (ISM) to get a series of masks according to " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " in an iterative manner." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": "First, we randomly select a point from the outlier SfM set, " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}_k\\in \\bar{\\mathcal{O}}" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " Using Eq.2),we can get the projected 2D point " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\bar{x}_k" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " to prompt ISM, producing a binary mask " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "m_{k}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " delineating a semantically coherent local region around this outlier point. We further define a threshold " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " , which describes the desired number of correspondences to estimate a local alignment. If the number of outliers within " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "m_{k}" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " i.e. " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "|m_k\\bigcap \\bar{P} (\\bar{\\mathcal{O}})| > T" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " then we keep " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "m_{k}" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " as it is and exclude within-mask outliers from being selected in the next iteration. Otherwise, we use all the outliers within the produced mask to prompt ISM again to update " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "m_{k}" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " until the above criterion is met, or no further outliers can be included, in which case we discard this mask and the outliers within. We repeat the above process until no point in " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " is left, eventually obtaining a set of semantically coherent masks around the outliers, " + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\{m_k\\}_{k = 1}^M" + }, + { + "bbox": [ + 313, + 144, + 553, + 360 + ], + "type": "text", + "content": " .Visualization of this process can be found in the supplementary materials. This iterative process is denoted as" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "spans": [ + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "type": "inline_equation", + "content": "\\operatorname{Init}: m_k = \\operatorname{ISM}(x_k)" + }, + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "type": "text", + "content": ", s.t. " + }, + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "type": "inline_equation", + "content": "\\bar{x}_k \\in \\bar{P}(\\bar{\\mathcal{O}})" + }, + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 363, + 548, + 397 + ], + "type": "inline_equation", + "content": "\\bar{x}_k \\notin \\bigcup_{j=1}^{k-1} m_j" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 399, + 553, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 440 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 315, + 399, + 553, + 440 + ], + "type": "inline_equation", + "content": "m_{k} = \\left\\{ \\begin{array}{ll}m_{k}, & \\text{if} |m_{k}\\bigcap \\bar{P} (\\bar{\\mathcal{O}})| > T\\\\ \\operatorname {ISM}(m_{k}\\bigcap \\bar{P} (\\bar{\\mathcal{O}})), & \\text{otherwise} \\end{array} \\right." + }, + { + "bbox": [ + 315, + 399, + 553, + 440 + ], + "type": "text", + "content": " (4)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "text", + "content": "Based on the outlier masks " + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\{m_k\\}_{k=1}^M" + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "text", + "content": ", we extract the within-mask outlier set " + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_k = P^{-1}(m_k \\bigcap P(\\mathcal{O}))" + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "text", + "content": " and their corresponding SfM set is " + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}_k" + }, + { + "bbox": [ + 313, + 441, + 553, + 488 + ], + "type": "text", + "content": ". We then estimate a local rigid transformation within each mask as follows:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 502, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 502, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 318, + 502, + 553, + 540 + ], + "type": "interline_equation", + "content": "s _ {k}, R _ {k}, t _ {k} = \\underset {s, R, t} {\\arg \\min } \\sum_ {(\\bar {\\chi} _ {i}, \\bar {\\mathbf {X}} _ {i}) \\in (\\mathcal {O} _ {k}, \\bar {\\mathcal {O}} _ {k})} \\| s R \\bar {\\chi} _ {i} + t - \\bar {\\mathbf {X}} _ {i} \\| ^ {2}. \\tag {5}", + "image_path": "1588092734e81c0b03ccb789161a77a5534c73cbabcf488c1b5e3a24ecd2b3af.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "content": "Finally, we construct the inlier mask as " + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "inline_equation", + "content": "m_0 = \\neg \\left( \\bigcup_{k=1}^{M} m_k \\right)" + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\neg" + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "content": " is the logical NOT operator. Then, the global transformation is applied to MASt3R's points within inlier mask " + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "content": " and each individual local transformation is applied to those within each local mask " + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "inline_equation", + "content": "m_k, k \\neq 0" + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "content": ". This results in the final SfM-aligned point cloud, " + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\chi^*" + }, + { + "bbox": [ + 313, + 540, + 553, + 617 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 362, + 625, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 362, + 625, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\chi^ {*} = \\bigcup_ {k = 0} ^ {M} \\left\\{s _ {k} R _ {k} \\chi_ {i} + t _ {k} \\mid \\chi_ {i} \\in m _ {k} \\right\\}. \\tag {6}", + "image_path": "2f02cfdc2599739dadb18a0b092b4b6ba1d48e13e18cd719dd95ffd447b11f48.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 659, + 553, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 553, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 553, + 695 + ], + "type": "text", + "content": "To obtain the final dense and pose-wise accurate point cloud " + }, + { + "bbox": [ + 313, + 659, + 553, + 695 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 659, + 553, + 695 + ], + "type": "text", + "content": " as prior, we concatenate the final SfM-aligned point cloud with that of SfM, i.e." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 403, + 699, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 699, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 403, + 699, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {X} = \\chi^ {*} \\bigcup \\bar {\\mathbf {X}} \\tag {7}", + "image_path": "6d3ab26ea6468bc3937ef724ad960fe3cffc934917cf48c6337639ce629d07a8.jpg" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26813" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 183, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 183, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 183, + 85 + ], + "type": "text", + "content": "3.2.3. Gaussian Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "type": "text", + "content": "The SPARS3R constructed " + }, + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 87, + 296, + 147 + ], + "type": "text", + "content": " can be directly used as the initialization for any Gaussian Splitting-based optimization method. Here we use Splatfacto, developed under the NeRFStudio framework [49]; the Gaussian optimization loss is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 158, + 295, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 158, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 94, + 158, + 295, + 172 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\| \\tilde {C} - C \\| _ {1} + \\lambda_ {2} \\mathrm {D} - \\operatorname {S S I M} (\\tilde {C}, C), \\tag {8}", + "image_path": "8afa15382bdfeb05accf606a8e62a326c4ea0bea8f16874dc73232b2d96c67db.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "inline_equation", + "content": "\\tilde{C}" + }, + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "text", + "content": " denote our rendered and the groundtruth RGB image, D-SSIM denotes the structural similarity loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 221, + 137, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 137, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 137, + 235 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "type": "text", + "content": "Experiments are performed on three popular benchmark datasets. Tanks and Temples [27] contains 8 scenes captured in video format. Following InstantSplat [14], We uniformly sample 24 images with alternating indices for train and test split. MVImgNet [60] is an object-centric dataset that consists of 7 scenes. These scenes are captured from a consistent height covering " + }, + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "type": "text", + "content": " field of view. We apply the same sampling and train-test split procedure as for Tanks and Temples. For sparser settings, we uniformly sample 3 and 6 views from the training set for the above two datasets. Mip-NeRF 360 [2] comprises of 9 scenes with " + }, + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 242, + 296, + 456 + ], + "type": "text", + "content": " views and greater pose variation between the scenes, including diverse heights and distances. The increased variability, combined with sparsity, adds complexity to accurate pose estimation and scene reconstruction. For this dataset, we follow the test set outlined in MipNeRF360 [2] and uniformly sample 12 images from the original training set to construct a sparse-view set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 459, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 531 + ], + "type": "text", + "content": "All sparse-view input images go through a COLMAP [46] pipeline for camera calibration. Specifically, we use feature matching results from MASt3R [29] for SfM triangulation. Since sparse-view registration can be unstable due to limited pairs, we perform multiple SfMs and pick the outcome that maximizes successful triangulation per image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 189, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 189, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 189, + 552 + ], + "type": "text", + "content": "4.1. Sparse NVS Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Accurately and holistically evaluating sparse NVS should involve many practical considerations. Various NeRF-based approaches [12, 23, 25, 39, 51, 57, 59] have assumed perfect camera poses, i.e. calibration results from dense views, for sparse-view scenarios, which is unrealistic. Additionally, Gaussian-Splatting-based methods rely heavily on SfM point cloud, further necessitating the separation of train and test views from the registration stage. Previous approaches [65] have tried to perform re-triangulation based on known train poses, excluding test-set point cloud from training, but do not account for pose inaccuracy in sparse-view registration. While InstantSplat [14] adopts the realistic setup where they use recovered poses from sparse" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "views and initialize the test camera poses by applying Procrustes Alignment [22] between train-only poses and train-test-together poses, their large camera alignment error leads to significant disruption in common pixel-based rendering metrics, e.g. PSNR and SSIM [53]." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 329, + 141, + 541, + 203 + ], + "blocks": [ + { + "bbox": [ + 329, + 141, + 541, + 203 + ], + "lines": [ + { + "bbox": [ + 329, + 141, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 329, + 141, + 541, + 203 + ], + "type": "table", + "html": "
Mip-NeRF 360ER(μ)↓ET(μ)↓
Procrustes Alignment0.1960.0144
+ RANSAC0.1790.0114
+ Rotation Points0.1560.0117
", + "image_path": "ca6ca25b7a9bd17a21fcf6799059eb39469dd79dff0b0150a4daec2f6a3a7802.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "type": "text", + "content": "Table 1. Improvements over Procrustes Alignment baseline in average rotation error " + }, + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "type": "inline_equation", + "content": "E_{R}" + }, + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "type": "text", + "content": " and translation error " + }, + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 211, + 555, + 244 + ], + "type": "text", + "content": ". Incorporating rotation points further minimizes the overall error." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": "We also operate under the realistic setup but improve upon the camera alignment process in two aspects. First, during camera alignment, instead of relying only on camera positions, we sample three additional points " + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{rotation}} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": " per camera along the unit direction vectors in rotation matrix " + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{camera}} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": " at camera position " + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{camera}} \\in \\mathbb{R}^{1 \\times 3}" + }, + { + "bbox": [ + 313, + 252, + 554, + 324 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 366, + 335, + 553, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 335, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 366, + 335, + 553, + 348 + ], + "type": "interline_equation", + "content": "P _ {\\text {r o t a t i o n}} = s _ {\\text {c l o u d}} R _ {\\text {c a m e r a}} + T _ {\\text {c a m e r a}}. \\tag {9}", + "image_path": "8ce460be27718a91ea9d5427a792f8fe6cc02cd3bf3a4de2747f246ff97a34d2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": "The scale of the camera cloud is denoted by " + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{cloud}} = \\sqrt{\\sigma_x^2 + \\sigma_y^2 + \\sigma_z^2}" + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\sigma_x, \\sigma_y," + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\sigma_z" + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": " are the standard deviations of camera positions along " + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "inline_equation", + "content": "x, y," + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": " axes respectively. The four points per camera enable alignment of both camera rotations and translations. Second, we add RANSAC [15] in the alignment process, minimizing the influence of outliers. We measure the alignment rotation error " + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "inline_equation", + "content": "E_R" + }, + { + "bbox": [ + 313, + 353, + 555, + 467 + ], + "type": "text", + "content": " by the angular difference in degrees between two rotations represented by quaternions following" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 473, + 553, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 473, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 365, + 473, + 553, + 500 + ], + "type": "interline_equation", + "content": "E _ {R} = 2 \\cdot \\arccos \\left(\\frac {\\operatorname {a b s} \\left(q _ {1} \\cdot q _ {2}\\right)}{\\| q _ {1} \\| \\cdot \\| q _ {2} \\|}\\right), \\tag {10}", + "image_path": "a2eccb2c0551f08ba0ebc2607effea9fe593181074d39fd19c1f99d19fcbef39.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "type": "text", + "content": "and the translation error " + }, + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "type": "text", + "content": " by the Euclidean norm " + }, + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 313, + 507, + 555, + 616 + ], + "type": "text", + "content": " between translation vectors. As shown in Tab. 1, these two improvements enhance camera alignment accuracy in both rotation and translation. Beyond accurate camera pose alignment, test pose optimization can also be applied between rendered and ground-truth images to minimize the pose error; however, such process is time-consuming and can get stuck if the initial displacement is too large, as is sometimes the case in InstantSplat [14]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. In addition to the commonly used PSNR, SSIM, and LPIPS, we incorporate DreamSim [18] (DSIM) as an additional metric to assess rendering quality. DSIM captures human perceptual similarity by finetuning a combination of embeddings from visual foundation models based on human evaluations. Notably, it provides a robust assessment of rendering image quality under moderate pose shift, which frequently occurs in the realistic sparse-view" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26814" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "settings. Further details on DSIM's robustness as an evaluation metric are provided in the supplementary materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "type": "text", + "content": "Implementation Details. We employ Segment-Anything Model [26] as the segmentation model in Semantic Outlier Alignment. The weights for the loss terms are " + }, + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.8" + }, + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.2" + }, + { + "bbox": [ + 55, + 99, + 296, + 183 + ], + "type": "text", + "content": ". For fair implementation and comparison, we employ test pose optimization for all baselines and SPARS3R for 500 steps to maximally remove the effect of shifted camera pose. All baselines are run until convergence." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 157, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 157, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 157, + 204 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 295, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 243 + ], + "type": "text", + "content": "We compare the different components in SPARS3R to demonstrate their effectiveness on the Mip-NeRF 360 [2] dataset:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 247, + 295, + 365 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 270 + ], + "type": "text", + "content": "- SfM-initialized 3DGS: Optimizing 3DGS on a sparse SfM point cloud and SfM-estimated camera poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 295, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 294 + ], + "type": "text", + "content": "- MAST3R-initialized 3DGS: Optimizing 3DGS on a MAST3R-produced dense point cloud and camera poses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 295, + 294, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 330 + ], + "type": "text", + "content": "- Global Fusion Alignment: Aligning the MAST3R-produced dense point cloud onto SfM point cloud with a global transform. Poses are taken from SfM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 331, + 294, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 365 + ], + "type": "text", + "content": "- Semantic Outlier Alignment (SPARS3R): The full SPARS3R pipeline that aligns outliers through their 2D semantic masks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 58, + 376, + 293, + 466 + ], + "blocks": [ + { + "bbox": [ + 58, + 376, + 293, + 466 + ], + "lines": [ + { + "bbox": [ + 58, + 376, + 293, + 466 + ], + "spans": [ + { + "bbox": [ + 58, + 376, + 293, + 466 + ], + "type": "table", + "html": "
SettingPSNRSSIMLPIPSDSIM
Initialization: SfM [46]
3DGS [24]16.60.3880.4580.232
Initialization: MASt3R [29]
3DGS [24]15.90.2930.4630.229
+ Global Fusion Alignment18.60.4860.3300.130
+ Semantic Outlier Alignment18.90.5000.3270.127
", + "image_path": "de43c8f40daaae813052dff7f868ec86652dd3b7317bb08ddd8470ac310f739f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "content": "As shown in Tab. 2, given only 12 images, sparse NVS is challenging both in registration and reconstruction. Specifically, we observe that the 3DGS based on SfM initialization performs better than MASt3R initialization. Clearly, render quality is affected if poses are inaccurate, despite locally coherent dense point cloud. The introduction of Global Fusion Alignment improves upon the naive baselines from either initializations by combining accurate pose from SfM and dense point cloud from MASt3R. For scenes that do not have significant depth discrepancies, Global Fusion Alignment works well. In scenarios such as the Bonsai scene in Fig. 3, MASt3R's relative depth between the table and the wall is inaccurate; therefore, Global Fusion Alignment alone leads to a broken geometry of the background. In comparison, our Semantic Outlier Alignment provide piece-wise alignments on the background to ad" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "dress this relative discrepancy. Quantitatively, we observe a 1.4 dB improvement on the Bonsai scene." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 105, + 433, + 180 + ], + "blocks": [ + { + "bbox": [ + 321, + 105, + 433, + 180 + ], + "lines": [ + { + "bbox": [ + 321, + 105, + 433, + 180 + ], + "spans": [ + { + "bbox": [ + 321, + 105, + 433, + 180 + ], + "type": "image", + "image_path": "7e2051c47bc81a21888a4eea47eb0604d435477f42df0165a518cb9514b3974e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 181, + 430, + 191 + ], + "lines": [ + { + "bbox": [ + 323, + 181, + 430, + 191 + ], + "spans": [ + { + "bbox": [ + 323, + 181, + 430, + 191 + ], + "type": "text", + "content": "Global Fusion Alignment (GFA)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 105, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 442, + 105, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 442, + 105, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 442, + 105, + 553, + 180 + ], + "type": "image", + "image_path": "12362d3bf7151a9c2b1a262546a2ff5fd6b69d04d0fa6ba9f1aaf46adc241f3b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 181, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 447, + 181, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 447, + 181, + 547, + 191 + ], + "type": "text", + "content": "Semantic Outlier Align (SOA)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 321, + 200, + 431, + 271 + ], + "blocks": [ + { + "bbox": [ + 321, + 200, + 431, + 271 + ], + "lines": [ + { + "bbox": [ + 321, + 200, + 431, + 271 + ], + "spans": [ + { + "bbox": [ + 321, + 200, + 431, + 271 + ], + "type": "image", + "image_path": "8c3da074c43d318456d1be6921fc64e0292bc955fd3ece68d7d31be218fa5237.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 274, + 427, + 285 + ], + "lines": [ + { + "bbox": [ + 326, + 274, + 427, + 285 + ], + "spans": [ + { + "bbox": [ + 326, + 274, + 427, + 285 + ], + "type": "text", + "content": "Aligned point cloud from GFA" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 441, + 196, + 551, + 274 + ], + "blocks": [ + { + "bbox": [ + 441, + 196, + 551, + 274 + ], + "lines": [ + { + "bbox": [ + 441, + 196, + 551, + 274 + ], + "spans": [ + { + "bbox": [ + 441, + 196, + 551, + 274 + ], + "type": "image", + "image_path": "45a32b2e951d136f28955f084c97630f57b35dd89b2060dbe69069b25095ed9a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 274, + 548, + 285 + ], + "lines": [ + { + "bbox": [ + 447, + 274, + 548, + 285 + ], + "spans": [ + { + "bbox": [ + 447, + 274, + 548, + 285 + ], + "type": "text", + "content": "Aligned point cloud from SOA" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 297, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 342 + ], + "type": "text", + "content": "Figure 3. Visual comparison of SPARS3R with and without SOA. While the dense bonsai in the foreground is aligned with the sparse point cloud, depth differences are obvious. SOA successfully fixes such gap." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 315, + 363, + 561, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 473, + 296, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 507 + ], + "type": "text", + "content": "Table 2. Ablation on key components of SPARS3R. The results are shown using PSNR, SSIM, LPIPS and DSIM based on the MipNeRF 360 [2]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 363, + 561, + 430 + ], + "lines": [ + { + "bbox": [ + 315, + 363, + 561, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 363, + 561, + 430 + ], + "type": "table", + "html": "
MethodsMip-NeRF 360Tanks & TemplesMVimgNet
RPEtRPErRPEtRPErRPEt RPEr
DUSt3R [52]2.0752.5840.5700.1430.438 0.421
MASt3R [29]1.1861.4900.2410.2480.208 0.304
InstantSplat [14]2.0492.5550.1510.0810.264 0.311
COLMAP + MASt3R0.2520.4120.1610.0930.075 0.078
", + "image_path": "8d5824ce43f7b99f389d0dd233a64e4ccea17f8aea3cfc340f11962075bd43d3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of pose accuracy across three datasets, Relative Translation Error " + }, + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "type": "inline_equation", + "content": "(\\mathrm{RPE}_t)" + }, + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "type": "text", + "content": " and Relative Rotation Error " + }, + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "type": "inline_equation", + "content": "(\\mathrm{RPE}_r)" + }, + { + "bbox": [ + 313, + 437, + 555, + 472 + ], + "type": "text", + "content": " [62] are calculated based on the normalized poses." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 486, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 654 + ], + "type": "text", + "content": "Pose Estimation Accuracy. As shown in Tab. 3, we compare the pose estimation accuracy from different calibration approaches. Both DUSt3R [52] and MASt3R [29] have significantly larger errors compared to COLMAP [46] with MASt3R feature matching outputs. InstantSplat [14] uses DUSt3R's [52] dense point cloud and pose estimation and attempts to improve accuracy through a training pose optimization approach similar to BARF [32]. While it brings down the errors in some cases, such training pose optimization strategy does not work as well in more challenging datasets like Mip-NeRF 360. This highlights the difficulty in improving camera calibration jointly with dense forward render, considering the loss landscape is highly nonlinear based on pixel-wise metrics." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 660, + 503, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 503, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 503, + 673 + ], + "type": "text", + "content": "4.3.Quantitative and Visual Evaluation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We quantitatively compare SPARS3R against various NVS methods in Tab. 4 and Tab. 5. Instant-NGP [36] and 3DGS [24] are two mainstream NVS methods based on" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26815" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 184 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 184 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 184 + ], + "type": "table", + "html": "
MethodsMip-NeRF 360 [2]Tanks & Temples [27]MVImgNet [60]
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
Instant-NGP [36]14.820.2940.6780.52415.280.4510.3890.25413.280.4260.8920.892
3DGS [24]16.570.3880.4580.23221.070.7300.1940.07021.240.6730.2340.064
FSGS [65]17.600.4430.5580.24325.720.8450.1110.02323.430.7600.2120.042
SparseGS [55]16.660.4050.4610.21020.280.7270.2020.07520.560.6720.2480.072
DRGS [10]16.880.4010.6490.30021.460.7230.2890.07821.700.6410.4220.071
CF-3DGS [19]13.270.2500.6980.50918.990.6060.2960.12715.430.4080.5450.325
InstantSplat [14]16.230.3590.5430.23326.970.8740.1150.01323.220.7340.2480.028
SPARS3R18.850.5000.3270.12729.900.9190.0470.00725.850.8200.1140.011
", + "image_path": "dc8c04d074362b6fc39fd011076c17c8d971425a296b331987d1eee29791c935.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 233, + 299, + 333 + ], + "blocks": [ + { + "bbox": [ + 55, + 192, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 214 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of different NVS methods on 12 views on three popular benchmark datasets, totaling 24 scenes. All methods are run on the same registrations and updated with test pose optimization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 233, + 299, + 333 + ], + "lines": [ + { + "bbox": [ + 58, + 233, + 299, + 333 + ], + "spans": [ + { + "bbox": [ + 58, + 233, + 299, + 333 + ], + "type": "table", + "html": "
3 ViewsTanks & TemplesMVImgNet
PSNRSSIMLPIPSDSIMPSNRSSIMLPIPSDSIM
FSGS20.360.6580.2190.07716.920.5120.4760.183
InstantSplat21.650.7550.1640.04416.290.5510.3230.103
SPARS3R21.830.7550.1500.04318.620.5880.2810.091
6 Views
FSGS23.900.7830.1520.04320.420.6440.3310.080
InstantSplat25.890.8470.1170.02221.090.6950.2360.046
SPARS3R26.800.8730.0730.01422.750.7320.1610.023
", + "image_path": "f37aa4e4c6bccf1c26b03a62c5ec86a66281419587f10f4dba4abede8bbc4f65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 295, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 364 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of 3 and 6 views on Tanks & Temples and MVImgNet datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 377, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 473 + ], + "type": "text", + "content": "implicit and explicit scene representations. Specifically, 3DGS leverages SfM points from registration as the starting point for scene optimization, while NeRF [35] based methods rely only on poses. Colmap-Free 3DGS [19] also does not rely on prior registration and optimizes cameras along with reconstruction. We find that both Instant-NGP and CF-3DGS perform significantly worse than other methods that use a prior SfM point cloud." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 475, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 713 + ], + "type": "text", + "content": "Among methods that are designed for sparse NVS, DRGS [11], SparseGS [55], and FSGS [65] leverage monocular depth prior to gain more geometric information of the scene. DRGS applies an additional smoothness constraint, which leads to high PSNR values but worse LPIPS and DSIM performances. SparseGS additionally applies a Score Distillation Sampling Loss, which generates novel views and constrained them through a prior generative model. Its results are sharper than DRGS, but more noisy, leading to worse PSNR. FSGS seeks to address the sparse point cloud issue in sparse-view NVS by modifying the densification algorithm in 3DGS. More Gaussians are encouraged between the space of two existing Gaussians, allowing representative areas to densify quicker. FSGS has overall better metrics than SparseGS and DRGS; however, it still encounters issues in regions where few initializations are available, e.g. the background of the scene. As shown in Fig. 4 and in Fig. 1, this leads to various blurry regions. InstantSplat [14] is a concurrent work that uses prior depth estimation from DUSt3R [52] and the dense point cloud to help with sparse-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 236, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 403 + ], + "type": "text", + "content": "view NVS. While the initialized point cloud is much denser than FSGS, InstantSplat also inherits the suboptimal poses produced by DUSt3R. The incorrect geometry limits the effectiveness of densification in 3DGS and can lead to generation of floaters; to ameliorate this problem, InstantSplat disables Gaussian densification and optimizes the scene with few steps. As shown in Tab. 3 and Tab. 4, on the Mip-NeRF 360 and MVImgNet dataset, where pose accuracy from DUSt3R is limited, InstantSplat performs on-par with FSGS despite using a denser prior point cloud. As shown in Fig. 4, we observe that the Gaussians tend to be noisy on the surface. In the Stump and Bicycle scenes, InstantSplat's results cannot be rigidly aligned well with the groundtruth test poses due to the large offset to begin with." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 408, + 556, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 553 + ], + "type": "text", + "content": "Our SPARS3R method improves upon DUSt3R and MASt3R's prior point cloud by adjusting it based on conventional SfM calibration approaches. Since SfM only triangulates based on confident keypoints, its pose and depth estimation performance, particularly under sparse-view scenario, is more reliable. As shown in Tab. 4, SPARS3R consistently improves upon previous best approach by an average of " + }, + { + "bbox": [ + 313, + 408, + 556, + 553 + ], + "type": "inline_equation", + "content": "2.7\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 408, + 556, + 553 + ], + "type": "text", + "content": ". SPARS3R is stable in its optimization process and does not generate additional floaters with more iterations. As seen in Fig. 3, SPARS3R produces rendering with smooth surfaces in the foreground, and sharp details in the background." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "For even sparser settings, we provide a quantitative comparison in Tab. 5 between SPARS3R and the two best-performing methods from Tab. 4, FSGS [65] and InstantSplat [14]. FSGS struggles significantly under extreme sparsity due to its reliance on a sparse initialization, resulting in DSIM scores that are approximately two- and four-times worse than those of SPARS3R under the 3-view and 6-view settings, respectively. InstantSplat generally performs better than FSGS, benefiting from a much denser point cloud initialization produced by DUSt3R [52], which is particularly advantageous under extreme sparse conditions. However, its performance remains inferior to SPARS3R due to suboptimal pose estimation. Our" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26816" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 561, + 366 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 561, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 561, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 561, + 366 + ], + "type": "image", + "image_path": "afb8b32a21e1c4cbe6e7f63dd2fa11c336b835bcbf35a8b69a5355b9b84bf8a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 376, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 555, + 399 + ], + "type": "text", + "content": "Figure 4. Visual comparisons of different NVS methods on 12 views on Mip-NeRF 360 [2] dataset. Zooming in on the visualizations is recommended to show differences in detail. More visualizations for other datasets are available in the Supplemental Material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 419, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 504 + ], + "type": "text", + "content": "SPARS3R continues to outperform these methods, improving upon the best approach by an average of 0.9 dB and 1.3 dB in the 3-view and 6-view settings, respectively. These results further demonstrate SPARS3R's superiority over state-of-the-art approaches in handling extreme sparse-view reconstruction. For more visualizations and flythroughs of the scene, please refer to the supplemental materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 515, + 133, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 133, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 133, + 526 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "While SPARS3R significantly improves upon previous SoTA, there are also several limitations worth noting. Firstly, SPARS3R relies on a well-performing semantic segmentation model to group outliers; as such, if the segmentation model is very sensitive, i.e. separating a connected surface into multiple disjointed parts, some of these parts may lack sufficient SfM support for alignment. If the segmentation model produces masks that group multiple areas with disparate depth together, the local alignments are also less effective similar to the global alignment. Secondly, non-rigid transform between two point clouds are worth further investigation. While SPARS3R's piece-wise rigid transform based on semantic relationship is fast and addresses the inaccurate inter-object depth estimation, a non-rigid formulation may be more generalizable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 418, + 387, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 387, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 387, + 430 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "We present SPARS3R, a scene reconstruction and NVS method that can achieve high-quality rendering with sparse input images. We demonstrate that existing methods lead to blurry rendering due to sparse point cloud initialization. Recent progress in generating a dense point cloud from prior depth estimation model, while promising, can lead to noisy pose estimation. To address these issues, SPARS3R combines the advantages of both by proposing a two-step alignment approach. The first step computes a global transformation between the dense point cloud from depth prior and a reference SfM point cloud. The second step takes the semantic regions around the outlier points identified from the first step and performs successive local alignments. This approach successfully addresses the depth estimation errors within the prior dense point cloud and demonstrates that the updated point cloud leads to significantly better performances in sparse-view NVs. We also introduce several improvements in the evaluation process to better represent the practical limitations in sparse-view registration and reconstruction. In the future, we hope to further improve the smoothness of our alignment process, e.g. by exploring potential non-rigid transformation approaches." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26817" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 246 + ], + "type": "text", + "content": "This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The authors also thank the support and feedbacks from researchers, Kaiwen Jiang, Dr. Ravi Ramamoorthi, at UCSD." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 315, + 72, + 373, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 373, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 373, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 91, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 319, + 91, + 555, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 91, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 555, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 319, + 157, + 555, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 157, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 319, + 157, + 555, + 213 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 320, + 214, + 555, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 214, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 320, + 214, + 555, + 267 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 270, + 555, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 270, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 320, + 270, + 555, + 312 + ], + "type": "text", + "content": "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 315, + 554, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 315, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 320, + 315, + 554, + 355 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 320, + 358, + 554, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 358, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 320, + 358, + 554, + 413 + ], + "type": "text", + "content": "[6] Jiahao Chen, Yipeng Qin, Lingjie Liu, Jiangbo Lu, and Guanbin Li. Nerf-hugs: Improved neural radiance fields in non-static scenes using heuristics-guided segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19436-19446, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 415, + 554, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 554, + 468 + ], + "type": "text", + "content": "[7] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 470, + 554, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 470, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 320, + 470, + 554, + 523 + ], + "type": "text", + "content": "[8] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12952, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 525, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 554, + 568 + ], + "type": "text", + "content": "[9] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. In Forty-first International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 570, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 554, + 623 + ], + "type": "text", + "content": "[10] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 625, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 678 + ], + "type": "text", + "content": "[11] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[12] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26818" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[13] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with " + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "inline_equation", + "content": "15\\mathrm{x}" + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": " reduction and " + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2311.17245, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 204 + ], + "type": "text", + "content": "[14] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "type": "text", + "content": "[16] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "type": "text", + "content": "[17] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[18] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[19] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20796-20805, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[20] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "type": "text", + "content": "[21] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "type": "text", + "content": "[22] John C Gower. Generalized procrustes analysis. Psychometrika, 40:33-51, 1975. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[23] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[25] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infererf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12912-12921, 2022. 1, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[28] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719-21728, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "type": "text", + "content": "[29] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 330 + ], + "type": "text", + "content": "[30] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 396 + ], + "type": "text", + "content": "[31] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[32] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 555, + 499 + ], + "type": "text", + "content": "[33] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 567 + ], + "type": "text", + "content": "[34] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[37] KL Navaneet, Kossar Pourahmadi Meibodi, Soroush Abbasi Koohpayegani, and Hamed Pirsiavash. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26819" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[38] Simon Niedermayr, Josef Stumpfegger, and Rüdiger Westermann. Compressed 3d gaussian splattering for accelerated novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10349-10358, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[39] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "text", + "content": "[40] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatt-ting for robust real-time rendering with " + }, + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2403.13806, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[41] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In European Conference on Computer Vision, pages 19-37. Springer, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[42] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "text", + "content": "[43] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In European Conference on Computer Vision, pages 293-310. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 489, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 544 + ], + "type": "text", + "content": "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[46] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 591, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 656 + ], + "type": "text", + "content": "[47] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[48] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "text", + "content": "[49] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 195 + ], + "type": "text", + "content": "[50] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "text", + "content": "[51] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 1, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "type": "text", + "content": "[52] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 354, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 407 + ], + "type": "text", + "content": "[54] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. Advances in neural information processing systems, 35:32653-32666, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[55] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 " + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "inline_equation", + "content": "\\{\\backslash \\mathrm{deg}\\}" + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": " sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "type": "text", + "content": "[56] Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20923–20931, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 511, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 566 + ], + "type": "text", + "content": "[57] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[58] Yifan Yang, Shuhai Zhang, Zixiong Huang, Yubing Zhang, and Mingkui Tan. Cross-ray neural radiance fields for novel-view synthesis from unconstrained image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15901-15911, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[60] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26820" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 353 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[61] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[62] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual(-inertial) odometry. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7244-7251, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[63] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 307 + ], + "type": "text", + "content": "[64] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 353 + ], + "type": "text", + "content": "[65] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 1, 2, 5, 7" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26821" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_content_list.json b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b118f53e7ebf03aec03c5f19ad0e2dccf4caf9f9 --- /dev/null +++ b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_content_list.json @@ -0,0 +1,1801 @@ +[ + { + "type": "text", + "text": "SPC-GS: Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs", + "text_level": 1, + "bbox": [ + 117, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guibiao Liao $^{1,2}$", + "bbox": [ + 148, + 202, + 274, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qing Li²", + "bbox": [ + 315, + 203, + 388, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyu Bao $^{1,2}$", + "bbox": [ + 428, + 203, + 545, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guoping Qiu3", + "bbox": [ + 586, + 203, + 700, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kanglin Liu $^{2*}$", + "bbox": [ + 741, + 203, + 854, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electronic and Computer Engineering, Peking University", + "bbox": [ + 230, + 220, + 767, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Pengcheng Laboratory, $^{3}$ University of Nottingham", + "bbox": [ + 284, + 239, + 709, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56b1a999e97bfb27559c4b4d5347d830b2f0465d5b5b64c6cc5c4452ebb5dea0.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 98, + 297, + 243, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79e834178e20de645a65d1a7d454339e0542438ce41514e605ee1c9035cb6ed3.jpg", + "image_caption": [ + "(a1) SfM Points (Gau-Grouping)" + ], + "image_footnote": [], + "bbox": [ + 263, + 297, + 413, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ace5f62f9e314a28d2308fb2950cac66f2b24f8e4a89212f00772bb1862c549.jpg", + "image_caption": [ + "(a2) Rendered RGB/Semantic Results (Gau-Grouping)" + ], + "image_footnote": [], + "bbox": [ + 421, + 297, + 568, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6134cebc0dc6fc47e4785696a7f028feec51412a6e498a37acfa61c3f457a023.jpg", + "image_caption": [ + "Gau-Grouping" + ], + "image_footnote": [], + "bbox": [ + 733, + 301, + 753, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b9018d89f4da445abb28aab4c6909a9c87a5aa03350d2f9991a98ca2fee36f5.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 834, + 303, + 854, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89fede60bb0b3e956fa0a0b57b0e4d5f4c5e37f5ff2fbf2cd7b49645457a43d8.jpg", + "image_caption": [ + "Sparse Input Images" + ], + "image_footnote": [], + "bbox": [ + 98, + 400, + 243, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94781e75ea003cab1e969ead9a69c792df92c5ea0b55d2442263f30ef94fdd45.jpg", + "image_caption": [ + "(b1) Scene layout Points (Ours)" + ], + "image_footnote": [], + "bbox": [ + 264, + 398, + 413, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7f9783f58807b6d66b40ca97ea016f90c9ec11593b055df52772fdbd8f239af.jpg", + "image_caption": [ + "Figure 1. Visual comparisons of open-world free-view synthesis from sparse inputs (12 training views). The previous 3DGS-based method, Gau-Grouping [50], utilizes sparse SfM points for Gaussian initialization and limited training views for supervision, leading to inferior rendering results. In contrast, our approach leverages scene layout-based points for instructive initialization, and cooperates with semantic-prompt consistency constraints (detailed in Sec. 3.3), yielding superior reconstruction and segmentation results." + ], + "image_footnote": [], + "bbox": [ + 421, + 398, + 568, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e66b74017f05f01a248383a1395bf1ca8519b6a31cca494c2f1cd2917b56840.jpg", + "image_caption": [ + "(b2) Rendered RGB/Semantic Results (Ours)" + ], + "image_footnote": [], + "bbox": [ + 568, + 398, + 715, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa5f7a25599a8d4cc179e40a1774f231f750d9be53a16c0ee09ec72529d19c00.jpg", + "image_caption": [ + "(c) Comparison Results" + ], + "image_footnote": [], + "bbox": [ + 733, + 330, + 777, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eda3cd88212d2e53152452c01ce62cd50438a8dd1ea11730914c7adda18ebc6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 330, + 833, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9011ed3c69227ae8905b608d334c5339e8d43626e25997ce1733eb2a3ef8823d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 844, + 330, + 890, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 575, + 326, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting-based indoor open-world free-view synthesis approaches have shown significant performance with dense input images. However, they exhibit poor performance when confronted with sparse inputs, primarily due to the sparse distribution of Gaussian points and insufficient view supervision. To relieve these challenges, we propose SPC-GS, leveraging Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) Regularization for open-world free view synthesis with sparse inputs. Specifically, SGI provides a dense, scene-Layout-based Gaussian distribution by utilizing view-changed images generated from the video generation model and view-constraint Gaussian points densification. Additionally, SPC mitigates limited view supervision by employing semantic-prompt-based consistency constraints developed by SAM2. This approach leverages available semantics from training views, serving as instructive prompts, to optimize visually overlapping regions in novel views with", + "bbox": [ + 88, + 606, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2D and 3D consistency constraints. Extensive experiments demonstrate the superior performance of SPC-GS across Replica and ScanNet benchmarks. Notably, our SPC-GS achieves a 3.06 dB gain in PSNR for reconstruction quality and a $7.3\\%$ improvement in mIoU for open-world semantic segmentation. Project website at: https://gbliao.github.io/SPC-GS.github.io.", + "bbox": [ + 511, + 575, + 908, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 709, + 645, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting-based open-world free-view synthesis seeks to construct semantic Gaussian radiance fields using a set of 3D Gaussian primitives enriched with color and semantic attributes. This task enables the free novel view synthesis of both realistic appearances and open-vocabulary semantic understanding (also referred to as 3D open-vocabulary semantic segmentation) of the scenes, with significant implications for augmented/virtual reality, robot navigation, and manipulation [16, 18, 26, 28, 30, 58]. Existing methods [35, 50, 59] generally require dense, large sets of input views to train the semantic Gaussian radiance", + "bbox": [ + 511, + 733, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 114, + 887, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11264", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "field for comprehensive scene coverage. However, when only sparse views are available, these approaches exhibit substantial declines in reconstruction and segmentation accuracy. This is mainly attributed to two key limitations.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) Sparsely Distributed Gaussian Points: With sparse input views, the initialized Gaussian points via Structure-from-Motion (SfM) in 3DGS [19] are inherently few and sparsely distributed (Fig. 1 (a1)). Such sparse point distribution imposes challenges in optimizing 3D Gaussians to depict complex indoor environments, leading to suboptimal reconstruction quality and, consequently, impaired segmentation accuracy (Fig. 1 (a2)). 2) Insufficient View Constraints: Optimizing 3D Gaussian parameters with inadequate view supervision often converges to local minima that overfit training views [25, 60]. This results in semantic ambiguity and poor performance in under-observed regions when rendering from novel viewpoints, due to limited reliable semantic integration across different views.", + "bbox": [ + 89, + 152, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we propose SPC-GS, a novel framework that incorporates Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) regularization.", + "bbox": [ + 89, + 366, + 482, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To seek denser points for improved Gaussian initialization, SGI produces a structured scene layout Gaussian distribution in two steps. First, SGI employs a video generation model to create view-changed images proximal to training images. These generated images, combined with training views, enrich feature matching in SfM, yielding denser SfM points that promote scene layout Gaussian generation. Second, these denser SfM points undergo view-constraint Gaussian densification and outlier removal to generate structured scene layout points. These points provide an instructive initialization that faithfully captures indoor scene geometry (e.g. furniture, walls, ceilings), enhancing 3DGS training and scene representation, thereby improving segmentation accuracy (Fig. 1 (b1) and (b2)).", + "bbox": [ + 89, + 429, + 482, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on dense scene layout Gaussian initialization, SPC enhances supervision through view-to-view semantic consistency constraints. The core of SPC lies in utilizing informative semantics signals from known training views, serving as semantic prompts, to optimize online-rendered novel (pseudo) views, establishing pseudo-multiview constraints through training-to-pseudo semantic consistency. To achieve it, we utilize the powerful Segment Anything Model 2 (SAM2) [38] to build region mask correspondences across training and pseudo views. Using these correspondences, we leverage the semantics prompts provided by training views to enforce semantic consistency constraints on the corresponding regions of pseudo views at 2D and 3D spaces. This approach substantially augments supervision signals, leading to more coherent segmentation on novel views under sparse input conditions.", + "bbox": [ + 89, + 642, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works on 3D open-vocabulary segmentation or", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sparse-input free-view synthesis mainly focuses on object-centric or face-forward scenarios, such as LERF [20], and 3DOVS [31], where objects are centrally positioned, and views are sampled in an \"outside-in\" manner, resulting in significant view overlap. In contrast, this study focuses on sparse-input open-world free-view synthesis within indoor scenes, particularly relevant for robotics and augmented/virtual reality applications. To this end, we utilize the widely recognized Replica [42] and ScanNet [8] datasets, which feature representative indoor scenes characterized by an \"inside-out\" viewing configuration. This setup leads to less overlap between certain views, posing a substantial challenge for sparse-input open-world free-view synthesis. The contributions of this work are as follows:", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SPC-GS, an early and novel attempt at jointly reconstructing and understanding open-world indoor scenes from sparse-input images.", + "- We introduce Scene layout-based Gaussian Initialization (SGI) to tackle the challenge of sparsely distributed Gaussians, offering a dense and instructive scene layout point initialization that promotes semantic Gaussian learning.", + "- We present Semantic-Prompt Consistency (SPC) regularization to alleviate inadequate view supervision by enforcing semantic consistency constraints on unseen pseudo views through prompts derived from training views, enhancing coherent segmentation.", + "- Extensive experiments on sparse-input scenes show the superiority of our approach in free-view synthesis of realistic appearances and open-vocabulary segmentation." + ], + "bbox": [ + 513, + 304, + 903, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 542, + 653, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Free-view Synthesis via Radiance Fields. In recent years, radiance field-based free-view synthesis has made remarkable advancements. A Notable example is the Neural Radiance Field (NeRF) [13, 32], which leverages a coordinate-based multi-layer perceptron (MLP) to implicitly model 3D scenes, enabling novel view synthesis via volume rendering [10]. Another innovative approach is 3D Gaussian Splating (3DGS) [19, 48], which utilizes a collection of 3D Gaussians to represent the scene and incorporates a fast tile-based rasterizer technique, allowing real-time rendering.", + "bbox": [ + 511, + 568, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although current methods achieve impressive results in dense-view scenarios [4, 12, 15, 27, 29, 33, 54], their performance degrades significantly with sparse inputs [17, 60]. Sparse-input Free-view Synthesis. Recent efforts have tried to address the sparse-input view synthesis and they can broadly categorized into cross-scene pre-training methods and per-scene optimization methods.", + "bbox": [ + 511, + 719, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-training methods learn scene priors by pre-training on large-scale datasets spanning multiple scenes. For instance, pixelNeRF [52] conditions a NeRF on CNN features from input images during volume rendering. Similarly, several approaches [3, 45] utilize different NeRF conditions", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to facilitate cross-scene pre-training. Inspired by the success of 3DGS, recent methods such as pixelSplat [2] and MVSplat [6] integrate efficient CUDA-based rasterization with 3D Gaussian representations for pre-training. However, these works demand substantial training resources and large-scale datasets.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Per-scene optimization methods aim to rapidly reconstruct a 3D scene by introducing effective regularizations. Early NeRF-based methods enhance sparse-input view synthesis through depth-based [34, 44] and frequency range regularization [49]. Recent 3DGS-based works focus on improving rendering efficiency. [7] designs depth regularization loss using monocular depth. DnGaussian [25] and FSGS [60] enhance sparse Gaussian representations through global-local depth normalization and correlation depth distribution losses. CoR-GS [56] adopts color co-regularization loss using rendered pseudo views.", + "bbox": [ + 89, + 183, + 482, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead, we focus on enhancing view synthesis and openworld 3D indoor semantic understanding from sparse inputs following the per-scene optimization route.", + "bbox": [ + 89, + 351, + 482, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Open-world Understanding via Radiance Fields. Recent studies have explored integrating 2D visual foundation models (e.g. DINO [1] and CLIP [24, 37]) with radiance fields to facilitate 3D open-world scene understanding. These approaches can be broadly divided into NeRF-based and 3DGS-based methods.", + "bbox": [ + 89, + 397, + 482, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF-based. Early methods [11, 43] distill visual features from DINO into semantic radiance fields. More recent language-embedded neural field methods [20, 22] integrate CLIP's visual features into semantic fields, enabling language-driven 3D semantic understanding. Additionally, 3DOVS [31] further improves semantic accuracy through the proposed Relevancy-Distribution Alignment loss.", + "bbox": [ + 89, + 489, + 482, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DGS-based. Alternative to NeRFs, recent works explore embedding CLIP features into 3D Gaussians for more efficient semantic rendering. Feature 3DGS [59] directly attaches CLIP embeddings to 3D Gaussians, while FMGS [61] combines Gaussians with multi-resolution hash encodings to form the semantic Gaussian field. [41] embeds quantized CLIP features into Gaussians to reduce storage demand. Furthermore, LangSplat [35] learns low-dimensional semantic Gaussian embeddings and uses a deep neural network for post-process upsampling to obtain semantic representation. Gau-Grouping [50] introduces a 2D object identity loss within training views to optimize Gaussians, while this method requires dense view supervision. Note that the 3D interactive segmentation methods, such as [14, 21, 39, 51], lie outside the scope of language-driven open-world segmentation, as they rely on 2D point prompts to generate object masks but lack semantic meaning.", + "bbox": [ + 89, + 597, + 482, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While previous approaches rely on dense input images, our work focuses on GS-based 3D open-world scene understanding in sparse-input indoor scenarios.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates our framework, which consists of two primary components. First, Scene layout-based Gaussian Initialization (SGI) (Section 3.2) constructs a dense, instructive scene layout Gaussian initialization that improves Gaussian representation, thus supporting enhanced semantic learning. Then, using the initialized scene layout Gaussians, Semantic-Prompt Consistency (SPC) Regularization (Section 3.3) introduces 2D and 3D semantic consistency constraints, creating effective pseudo-multi-view optimization for semantic learning under sparse-input conditions.", + "bbox": [ + 511, + 114, + 906, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 273, + 643, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Definition. Our approach aims to leverage sparse input images to reconstruct 3D indoor scenes and enable language-driven semantic perception. Concretely, given $L$ posed sparse input images $I = \\{I_i\\}_{i=1}^L \\in \\mathbb{R}^{L \\times H \\times W \\times 3}$ and language descriptions, our goal is to generate both realistic appearance and open-world semantic understanding results for specified novel viewpoints. To achieve this, we exploit 3D Gaussians [19] for scene representation, and utilize CLIP models [24, 37] to provide semantic supervision. Preliminary. 3D Gaussian Splitting (3DGS) [19] represents a 3D scene using a collection of explicit 3D Gaussian primitives $\\{\\mathbf{g}_i = (\\mathbf{p}_i, \\mathbf{c}_i, o_i, \\mathbf{s}_i, \\mathbf{q}_i)\\}_{i=1}^{N_K}$ , where each primitive comprises a 3D position $\\mathbf{p} \\in \\mathbb{R}^3$ , color $\\mathbf{c} \\in \\mathbb{R}^3$ , opacity $o \\in \\mathbb{R}$ , scaling factor $\\mathbf{s} \\in \\mathbb{R}^3$ , and rotation quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ . For pixel color $\\hat{I}(x_p)$ rendering, 3DGS employs $\\alpha$ -blending point-based rendering [23], formulated as:", + "bbox": [ + 511, + 296, + 906, + 539 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} \\left(x _ {p}\\right) = \\sum_ {i \\in \\mathcal {N}} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 547, + 905, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{N}$ represents Gaussian points overlapping pixel $x_{p}$ , $\\alpha_{i} = o_{i}G_{i}^{proj}$ with $G_{i}^{proj}$ as the $i$ -th Gaussian's 2D projection, and transmittance $T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})$ .", + "bbox": [ + 511, + 589, + 905, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For pixel semantic feature $\\hat{\\mathbf{F}}(x_p)$ rendering, following [35, 59], we attach a learnable semantic parameter $\\mathbf{f}_i \\in \\mathbb{R}^D$ to each Gaussian and use the above $\\alpha$ -blending:", + "bbox": [ + 511, + 641, + 905, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} (x _ {p}) = \\sum_ {i \\in \\mathcal {N}} T _ {i} \\alpha_ {i} \\mathbf {f} _ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 696, + 905, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [31], we use a CLIP text encoder to generate textual feature $\\mathbf{T} \\in \\mathbb{R}^{M \\times 512}$ for $M$ classes and obtain rendered segmentation logits $\\hat{\\mathbf{S}}$ as follows: $\\hat{\\mathbf{S}} = \\cos\\langle \\omega_f(\\hat{\\mathbf{F}}), \\mathbf{T} \\rangle$ , where $\\cos$ denotes cosine similarity and $\\omega_f$ is a $1 \\times 1$ convolution layer for feature dimension alignment.", + "bbox": [ + 511, + 738, + 905, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We incorporate reconstruction and semantic losses to train the 3D Gaussians. The color reconstruction loss, denoted as $\\mathcal{L}_C$ , follows the formulation in 3DGS [19], and the semantic loss, $\\mathcal{L}_S$ , is defined as:", + "bbox": [ + 511, + 814, + 905, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\mathcal {L} _ {c o s} \\left(\\omega_ {f} (\\hat {\\mathbf {F}}), \\mathbf {F}\\right) + \\mathcal {L} _ {c e} \\left(\\omega_ {s} (\\hat {\\mathbf {S}}), \\mathbf {S}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 883, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11266", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/82e6338f8afaf1bbc9a41e24cac2102e5dc9d48ec3d1057f9c9018379d4bf31f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 107, + 903, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1a8f73a5dad6e911e0464b24142bb336d2f84f17516b79a679cc0ceaf951ab88.jpg", + "image_caption": [ + "Scene layout-based Gaussian Initialization (SGI)", + "Figure 2. Framework of SPC-GS. (a) We first generate adjacent images of each training image using the video generation model [47]. These generated images, combined with the original training images, produce denser initialized SfM points. These points are then optimized to create a scene layout Gaussian distribution via Gaussian densification and outlier removal. (b) Building on the scene-Layout Gaussian initialization, SPC leverages semantic information from training views as instructive semantic prompts to optimize adjacent rendered pseudo views, establishing semantic consistency constraints that enhance overall sparse-input semantic understanding of 3D scenes." + ], + "image_footnote": [], + "bbox": [ + 93, + 266, + 903, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ and $\\mathbf{S}$ are CLIP-derived semantic features and relevancy maps, respectively. $\\omega_{s}$ is a $1\\times 1$ convolution layer, and $\\mathcal{L}_{cos}$ and $\\mathcal{L}_{ce}$ denote cosine and cross-entropy losses.", + "bbox": [ + 89, + 566, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Scene layout-based Gaussian Initialization", + "text_level": 1, + "bbox": [ + 89, + 621, + 452, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose Scene layout-based Gaussian Initialization (SGI) to produce a dense and conducive point distribution for initialization with the following insights and designs.", + "bbox": [ + 89, + 643, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VGM-based Point Creation. In static indoor scenes, viewpoint variations correspond to camera movements. Motivated by this, we employ a video generation model (VGM) to create additional views beyond the sparse training set through image-to-video generation. These generated views augment the available scene information and, when combined with the original training views, enrich feature matching during SfM, yielding denser SfM points for generating scene layout Gaussians.", + "bbox": [ + 89, + 689, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for each training image $I_{i}$ , we generate eight adjacent views $\\{\\tilde{I}_i^j\\}_{j = 1}^8$ using the image-to-video generation mode of the powerful video generation model MotionCtrl [47]. This is accomplished by inputting the training images along with eight standard camera motion trajectory.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ries [47] into MotionCtrl. These trajectories include pan up, pan down, pan left, pan right, zoom in, zoom out, clockwise rotation, and anticlockwise rotation. Each trajectory generates a sequence of consecutive video frames. From each sequence, we select a single frame temporally adjacent to the input training image as the valid generated image since it exhibits superior visual quality. These generated images are then combined with the original training images for SfM processing, yielding denser SfM points for Gaussian initialization (see supplementary file). Despite being denser than SfM points derived from sparse-input training images, these points remain limited in capturing the geometric layout and details of indoor scenes.", + "bbox": [ + 511, + 566, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene-query Point Generation. We further generate scene-query point distributions to enhance 3DGS training and scene representation through the following steps.", + "bbox": [ + 511, + 763, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "i) View-constraints 3D Gaussian Densification. Using the denser SfM point initialization, we follow Gaussian densification [19] and using our view constraints from Eq. (9) to generate a more densely distributed, instructive scenelayout Gaussian point set $\\{\\mathbf{g}_i = (\\mathbf{p}_i,\\mathbf{c}_i,\\mathbf{f}_i,o_i,\\mathbf{s}_i,\\mathbf{q}_i)\\}_{i = 1}^{N_{K'}}$ as shown in Fig. 2. These denser $N_{K^{\\prime}}$ points provide an", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Video Generation Model based Point Creation", + "bbox": [ + 230, + 90, + 504, + 103 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "Scene-Layout Point Generation", + "bbox": [ + 674, + 90, + 859, + 103 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11267", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "improved spatial layout of indoor scenes, and enhanced attribute representations (e.g. $\\mathbf{c}_i$ , $\\mathbf{f}_i$ ), making them effective priors for Gaussian initialization.", + "bbox": [ + 89, + 90, + 482, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ii) Outlier Gaussian Primitive Removal (OGR). During Gaussian densification, new Gaussian primitives are sampled based on a 3D Gaussian probability distribution, which may generate outlier primitives positioned far from the actual scene layout. These outliers, when used in subsequent initialization, complicate the optimization process. To mitigate this, we present a density-based removal strategy to remove those sparse outlier points. Specifically, Gaussian primitives with fewer than $\\mathcal{P}$ neighboring points within a spherical radius $r$ are periodically eliminated during training. Empirical experiments show that setting $\\mathcal{P}$ to 5 and $r$ to 1 strikes a balance between effective outlier removal and the preservation of essential scene elements.", + "bbox": [ + 88, + 137, + 482, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By using instructive scene layout-based points for initialization, the 3DGS training and scene representation can be enhanced, consequently improving segmentation performance under the sparse-input condition.", + "bbox": [ + 89, + 334, + 483, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Semantic-Prompt Consistency Regularization", + "text_level": 1, + "bbox": [ + 89, + 407, + 478, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce Semantic-Prompt Consistency (SPC) regularization to exploit additional supervision signals. The key idea of SPC is to leverage semantics from training views as prompts, to impose effective semantic consistency constraints on corresponding regions of pseudo views.", + "bbox": [ + 89, + 430, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Region Correspondence Establishment. To establish correspondences between training and pseudo views, we draw inspiration from the powerful zero-shot temporal region proposal extraction capability of Segment Anything Model 2 (SAM2) [38] and leverage it to achieve this.", + "bbox": [ + 89, + 507, + 483, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we introduce Iterative Stochastic Prompting (ISP) to efficiently build correspondence. During training, ISP renders pseudo (unobserved) views using the online training SPC-GS. The view locations are sampled from the two nearest known training views by adding random noise following [60]. Then, ISP treats the current training image and its associated pseudo-views as a viewpoint-changed video sequence, and selects a stochastic point coordinate in the training image, as the 2D point prompt for SAM2, generating corresponding region masks across both the training and pseudo views, which can be formulated as:", + "bbox": [ + 89, + 584, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i}, \\left\\{\\mathbf {M} _ {i} ^ {*} \\right\\} _ {p = 1} ^ {P} = \\mathbf {I S P} \\left(I _ {i}, \\left\\{I _ {i} ^ {*} \\right\\} _ {p = 1} ^ {P}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 762, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{i}^{*}$ is the rendered pseudo image and $\\mathbf{M}_i^*$ is the corresponding region mask within $I_{i}^{*}$ . $P$ is set to 2 for computational efficiency. As training progress proceeds, the stochastic sampling mechanism encourages a uniform distribution of points across the entire image space, facilitating the efficient region mask correspondence construction for diverse regions of the whole image, as shown in Fig. 3.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbcd7abe309192f64805496f9f041733e07fcf32e0f0b7e5c47fb67286206c9e.jpg", + "image_caption": [ + "Figure 3. Illustration of Iterative Stochastic Prompting (ISP). During training, ISP randomly samples point coordinates $(\\star)$ from the training view, serving as the point prompt for SAM2 to produce region masks (highlighted in blue) across training and pseudo views." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 705, + 184 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2264dc1266402d2800977b05db6c3d1cfd28299d3310ff5cf9d9dd818e82e2df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 89, + 900, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic-Prompt for 2D Constrains. Building on the established region mask correspondences, we enforce effective constraints within 2D views to enhance Gaussian learning with sparse inputs.", + "bbox": [ + 511, + 268, + 905, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "i) View-to-View Regularization. Leveraging the semantics from training views as prompts, we enforce training-to-pseudo semantic consistency constraints in feature and relevancy map aspects, which is formulated as:", + "bbox": [ + 511, + 330, + 905, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {i n t e r} = \\sum_ {p = 1} ^ {P} | | \\frac {\\hat {\\mathbf {F}} _ {p} ^ {*} \\odot \\mathbf {M} _ {p} ^ {*}}{\\sum \\mathbf {M} _ {p} ^ {*}} - \\frac {\\hat {\\mathbf {F}} \\odot \\mathbf {M}}{\\sum \\mathbf {M}} | | _ {2} \\\\ - \\sum_ {p = 1} ^ {P} \\mathcal {R} [ \\Phi (\\hat {\\mathbf {S}} \\odot \\mathbf {M}) ] \\log (\\hat {\\mathbf {S}} _ {p} ^ {*} \\odot \\mathbf {M} _ {p} ^ {*}), \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 402, + 905, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{F}}^*$ and $\\hat{\\mathbf{S}}^*$ represent the semantic feature and the relevancy map derived from pseudo views. The function $\\Phi$ is the semantic uniformity operation: $\\Phi (\\hat{\\mathbf{S}}\\odot \\mathbf{M}) =$ $\\mathrm{argmax}_c\\sum_{i\\in \\mathbf{M}}\\mathbb{1}\\{\\mathrm{argmax}(\\omega_s(\\hat{\\mathbf{S}}_i)) = c\\}$ , which integrates SAM2's masks with a maximum voting mechanism to select the dominant class $c$ within the region mask M, unifying the region's semantics and enforcing consistent region-wise semantic supervision. $\\mathcal{R}[\\cdot ]$ denotes the dimensional reshaping operation to align the dimension shape of the region mask $\\mathbf{M}^*$ from the pseudo view. In $\\mathcal{L}_{inter}$ , we utilize self-rendered semantic $(\\hat{\\mathbf{F}},\\hat{\\mathbf{S}})$ from training views, generated by the trained SPC-GS, rather than 2D CLIP semantics. This is motivated by the superior coherence and reliability of 3D representations over their 2D counterparts [55] (see supplementary file for experimental validation).", + "bbox": [ + 511, + 503, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ii) Contrastive Regularization. Beyond inter-view, we introduce semantic-guided contrastive learning regularization within intra-training-views to further enhance supervision. This is based on the principle that features of semantically consistent objects should be similar.", + "bbox": [ + 511, + 734, + 903, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $N$ feature samples, each sample $\\mathbf{z}$ is derived from $\\mathbf{z} = W_{\\psi}(\\hat{\\mathbf{F}})$ , where $W_{\\psi}$ is a $1\\times 1$ convolution transformation layer. Let $i\\in U_z\\equiv \\{1\\dots N\\}$ be the index of a sample, and $T_{z}(i)\\equiv U_{z}\\backslash \\{i\\}$ represents the index set excluding index $i$ . We designate features with the same rendered semantic label $\\tilde{\\mathbf{q}}$ as positive samples, where $\\tilde{\\mathbf{q}} =$", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11268", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\Phi (\\hat{\\mathbf{S}}\\odot \\{\\mathbf{M}^k\\}_{k = 1}^\\eta)$ . Here, the $\\{\\mathbf{M}^k\\}_{k = 1}^\\eta$ denote $\\eta$ masks generated by SAM2's automatic mask generator in the entire image, and $\\Phi$ is the function that unifies semantics within each mask. Accordingly, for the $i$ -th sample, positive sample indexes are defined as: $P_{z}(i)\\equiv \\{p\\in T_{z}(i):\\tilde{\\mathbf{q}}_{p} = \\tilde{\\mathbf{q}}_{i}\\}$ , and negative sample indexes: $T_{z}(i)\\backslash P_{z}(i)$ . Similar to SimCLR [5], we employ contrastive learning loss to maximize agreement between positive samples. This method brings $P_{z}$ -indexed positive samples closer in feature space, imposing effective contrastive constraints on sparse-view Gaussian learning, which can be formulated as:", + "bbox": [ + 89, + 89, + 483, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n t r a}} = - \\sum_ {i \\in U _ {z}} \\frac {1}{\\mathcal {M} _ {z} (i)} \\sum_ {p \\in P _ {z} (i)} \\log \\left[ \\frac {e ^ {\\left(\\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {p}\\right) / \\tau\\right)}}{\\sum_ {j \\in T _ {z} (i)} e ^ {\\left(\\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j}\\right) / \\tau\\right)}} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 268, + 496, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_z(i)$ denotes the number of elements in $P_{z}(i)$ . The symbol $\\cdot$ refers to the inner product operation. $\\tau$ is the temperature parameter, set to 0.2 in the experiment.", + "bbox": [ + 89, + 324, + 483, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain $N$ samples, we employ a patch-wise sampling strategy by randomly extracting a patch of size $h \\times w$ in each training iteration, where $N = h \\times w$ . Empirically, $h$ and $w$ are both set to 128. In summary, the semantic prompt for 2D constraints can be formulated as:", + "bbox": [ + 89, + 369, + 483, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S P C} ^ {2 d} = \\mathcal {L} _ {\\text {i n t e r}} + \\mathcal {L} _ {\\text {i n t r a}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 455, + 480, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic-Prompt for 3D Constrains. In addition to 2D constraints, we extend semantic-prompt to impose effective 3D constraints on 3D Gaussian primitives. Considering that Gaussian primitives with maximum weight dominate the rendering results (Eq. (2)), we encourage semantic feature similarity between the maximally-weighted 3D Gaussians from two corresponding regions. Specifically, for each pixel within the region mask, we identify the Gaussian primitive with maximum weight along the ray cast into 3D space. Then, the semantic features of these maximally-weighted Gaussians from training views are used to constrain their counterparts in pseudo views.", + "bbox": [ + 89, + 486, + 483, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, since Gaussians related to region boundaries often span objects with differing semantics, directly applying the above constraints may complicate optimization. To mitigate this, we introduce a Region Boundary Erosion technique, which removes ambiguous primitives corresponding to region boundaries, which is defined as:", + "bbox": [ + 89, + 667, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S P C} ^ {3 d} = \\mathcal {L} _ {k l} (\\mathcal {F} | | \\mathcal {F} ^ {*}) = \\sum_ {j = 1} ^ {n ^ {\\prime}} \\bar {\\mathbf {f}} \\log \\left(\\frac {\\bar {\\mathbf {f}}}{\\mathbf {f} _ {j} ^ {*}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 770, + 483, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ denotes the mean feature representation of 3D Gaussian primitives derived from mask $\\mathbf{M}'$ in the training view, defined as $\\overline{\\mathbf{f}} = (\\mathbf{f} \\odot \\mathbf{M}') / \\sum \\mathbf{M}'$ . Here, $\\mathbf{M}'$ refers to the eroded mask after region boundary erosion, which is achieved through morphological erosion. Specifically, a", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pixel $\\mathbf{M}'(x,y)$ is retained as \"True\" only if all pixels within its $5\\times 5$ neighborhood in the input original mask $\\mathbf{M}$ are \"True\"; otherwise, it is set to \"False\". This operation effectively shrinks the boundaries of the region, causing them to contract inward. The set $\\mathcal{F}^{*} = \\{\\mathbf{f}_{1}^{*},\\mathbf{f}_{2}^{*},\\dots,\\mathbf{f}_{n^{\\prime}}^{*}\\}$ comprises $n^\\prime$ Gaussian primitives obtained from eroded masks $\\{\\mathbf{M}^{*'}\\}_{p = 1}^{P}$ in pseudo views.", + "bbox": [ + 511, + 90, + 905, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To summarize, the overall loss is given by:", + "bbox": [ + 531, + 198, + 815, + 213 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {C} + \\mathcal {L} _ {S} + \\mathcal {L} _ {G S} + \\mathcal {L} _ {S} ^ {3 d} + \\lambda \\left(\\mathcal {L} _ {S P C} ^ {2 d} + \\mathcal {L} _ {S P C} ^ {3 d}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 224, + 905, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{GS}$ is the semantic loss for generated images from video generation model: $\\mathcal{L}_{GS} = \\mathcal{L}_{cos}(\\omega_f(\\hat{\\mathbf{F}}^g),\\mathbf{F}^g) + \\mathcal{L}_{ce}(\\omega_s(\\hat{\\mathbf{S}}^g),\\mathbf{S}^g)$ . These generated RGB images are not used for color constraints due to their error-prone color pixels that degrade overall results (see supplementary file). $\\mathcal{L}_S^{3d}$ is the 3D local adaptive regularization, which promotes semantic similarity among neighboring Gaussians: $\\mathcal{L}_S^{3d} = \\frac{1}{\\theta k}\\sum_{i = 1}^{\\theta}\\sum_{j = 1}^{k}\\omega_{ij}(|\\mathbf{f}_i - \\mathbf{f}_j|_2)$ , where $\\theta = 800$ denotes the number of Gaussian randomly sampled per iteration, and $k = 5$ specifies the number of nearest neighbors for each sampled Gaussian in 3D space. $\\omega_{ij} = \\exp (-||\\mu_i - \\mu_j||_2)$ is the distance-based weighting coefficient. To prevent early-stage suboptimal convergence, $\\lambda$ is initialized at 0 and then increased to 1 after 4k iterations during training.", + "bbox": [ + 511, + 256, + 906, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 488, + 645, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 513, + 702, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two widely-used indoor scene datasets in terms of novel view synthesis: Replica [42] and ScanNet [8], using a sparse-input setting with only 12 input images for training. Following the standard sparse-input protocol [25, 60], every 10-th frame from each sequence is selected for testing, while 12 evenly sampled sparse views from the remaining images are used for training. These 12 views collectively encompass the entire scene. Notably, the \"inside-out\" viewing direction and limited overlap among views in sparse-input indoor scenes pose significant challenges for free-view synthesis.", + "bbox": [ + 511, + 536, + 905, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We employ PSNR, SSIM [46], and LPIPS [57] for evaluating image quality. We report segmentation performance using the mean Intersection-over-Union (mIoU) and mean pixel accuracy (mAcc).", + "bbox": [ + 511, + 703, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement our approach using the official PyTorch 3D Gaussian Splatting [19] on an NVIDIA A100 GPU. The dimension of the learnable semantic parameter attached to 3D Gaussians is set to 32. We modify the CUDA kernel to enable parallel differentiable rasterization of both color and semantic attributes. We employ CLIP ViT-B/16 [37] for feature extraction and SAM2 Hiera-L [38] for region mask generation. More details are provided in the supplementary file.", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11269", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39dc835ba34161f6a6aa3912d2fbdf24f17e813be142c2f37db76f62b60723df.jpg", + "image_caption": [ + "Figure 4. Visual reconstruction results on novel views. Our approach achieves superior global structure and photo-realistic details." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 890, + 241 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a92ba41123c0efb8fbbb6108810d99b52e5a34df6b7fe03c98b2c21638879813.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReplica [42]ScanNet [8]
PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑
3DOVS [31]13.2000.5150.6633.91321.43918.7210.6430.5447.75136.356
Feature 3DGS [59]17.0590.7000.3597.40730.20918.0880.6560.4247.34738.002
LangSplat [35]17.7150.7010.4203.29218.62118.7800.6630.4763.33924.491
Gau-Grouping [50]17.2970.6910.37615.02850.42219.0060.6750.42115.17257.355
DNGaussian [25]18.5130.7120.40216.65754.40018.3920.6910.48413.91254.814
FSGS [60]18.6940.7250.32816.10153.47420.5740.6960.40614.03156.850
CoR-GS [56]18.9550.7580.31215.93352.47620.8150.7210.39314.46255.812
Ours22.0110.7920.25423.96063.26222.4010.7410.36828.94468.823
", + "bbox": [ + 101, + 270, + 892, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of reconstruction and segmentation on novel views on the Replica and ScanNet datasets, using the CLIP [37] for optimization with 12 training views. Our approach achieves superior performances across all metrics on various datasets.", + "bbox": [ + 89, + 415, + 903, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 468, + 454, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct novel view synthesis comparisons using sparse input data in indoor scenes, including two types of methods. i) 3D open-vocabulary segmentation methods, including 3DOVS [31], Feature 3DGS [59], LangSplat [35], and Gau-Grouping [50], are designed for 3D open world semantic understanding. ii) Sparse-input free-view synthesis methods contain DNGaussian [25], FSGS [60], and CoR-GS [56], which are designed to learn Gaussian radiance fields from sparse inputs, focusing primarily on appearance modeling. For concurrent 3D open-world segmentation, we extend these methods by adding semantic properties to Gaussian and incorporating CLIP-based semantic supervision, aligning with the semantic settings of our approach. We adopt consistent evaluations across all methods by employing the same sparse-input training protocol and test views.", + "bbox": [ + 88, + 491, + 482, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. i) Reconstruction. As shown in Fig. 4, the results from existing 3D open-vocabulary segmentation methods exhibit significant artifacts, since these methods lack sufficient view constraints. Although sparse-view reconstruction methods introduce geometric constraints for regularization, they struggle to recover photo-realistic details. In comparison, our approach incorporates scene-Layout points prior and semantic consistency constraints, leading to a more detailed appearance.", + "bbox": [ + 88, + 718, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ii) Open-world Segmentation. In Fig. 5, current methods struggle to obtain precise and complete results, since they easily encounter ambiguity arising from insufficient seman", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tic view supervision. In contrast, our method delivers accurate results, benefiting from the SPC regularization that provides effective consistency constraints.", + "bbox": [ + 511, + 469, + 903, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 1 shows that, when optimized using the CLIP model [37], our approach significantly outperforms existing methods across all metrics on both synthetic and real datasets. Competing methods struggle under sparse input conditions, due to either limited Gaussian representations or insufficient semantic consistency constraints. Notably, LangSplat faces considerable challenges, as it freezes Gaussian coordinates, scale, and rotation during semantic parameter optimization, which severely impedes semantic learning under sparse inputs. Besides, our method maintains superior results when optimized with CLIP-LSeg [24] (see supplementary file), validating its generalizability.", + "bbox": [ + 511, + 516, + 906, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 710, + 679, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Components. In Table 2, applying the VGM-based Point Creation (PC, #1) and Scene layout Point Generation (SPG, #2) notably improves reconstruction quality over setting #0, and consequently boosts segmentation results, validating SGI's effectiveness in enhancing Gaussian representation. Adding the full Semantic-Prompt Consistency regularization (SPC, #4) further yields significant segmentation gains, showing SPC's capability to address insufficient view constraints and enhance 3D scene understanding. Fig. 6 also shows that adding SGI can reduce artifacts, and applying SPC further enhances segmentation results.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11270", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7291b47ef891b4f3d200342323ca7ad3c9359564c2cbe40ccc8b187ade11adb6.jpg", + "image_caption": [ + "Figure 5. Visual open-world segmentation results on novel views. Our approach yields more accurate and complete results." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 890, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4370fbe5f81b374970c978bcb62eb166aa2bf3f4aa3b757eaea7c29c7f7eb91d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CaseSGISPCReplica [42]
PCSPG2D3DPSNRSSIMLPIPSmIoUmAcc
#017.1340.6780.36914.89951.255
#1✓18.9240.7240.33416.27952.212
#2✓✓21.1760.7830.26019.40158.281
#3✓✓✓21.8140.7910.25722.84961.362
#4✓✓✓✓22.0110.7920.25423.96063.262
", + "bbox": [ + 91, + 313, + 483, + 405 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dea76bbf5fdb8550807a89a47938d54e0666b36056df0c8e46b59b47e400f683.jpg", + "table_caption": [ + "Table 2. Ablation study on key components." + ], + "table_footnote": [], + "table_body": "
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aOurs (Full setting)22.0110.7920.25423.96063.262
#bw/o View-to-View Regular.21.6830.7890.25721.24460.914
#cw/o Contrastive Regular.21.8790.7920.25622.32260.086
#dw/o Boundary Erosion21.9500.7920.25422.96461.706
", + "bbox": [ + 91, + 439, + 483, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bff21012490d3246c721469b28a5147bca5e32be66c4f26525755105c02ef230.jpg", + "image_caption": [ + "Figure 6. Visual ablation comparison of key components." + ], + "image_footnote": [], + "bbox": [ + 94, + 539, + 478, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Components in SPC. In Table 3, removing either the 2D view-to-view or contrastive regularization (#b, #c), or omitting the region boundary erosion operation in the 3D constraint (#d), degrades performance. This validates their roles in mitigating insufficient supervision and filtering ambiguous Gaussian primitives spanning different objects.", + "bbox": [ + 89, + 700, + 482, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness. In Table 4, our strategies also achieve significant gains even with sparse SfM initialization from limited training images, validating their robustness.", + "bbox": [ + 89, + 792, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization. From top to bottom in Fig. 7, we can see that leveraging denser and more effective points for Gaussian initialization promotes 3DGS training that enhances scene representation, leading to improved rendered results.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d00adbef7ca8e5a2dc7cfec9fd5a4e836442f7118ed187d478f7a8c475eca1fd.jpg", + "table_caption": [ + "Table 3. Ablation results of our SPC with different settings." + ], + "table_footnote": [], + "table_body": "
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aSfM Points Initialization17.1340.6780.36914.89951.255
#b#a + SGI20.6800.7740.27018.16956.928
#c#b + SPC21.5890.7860.26020.76661.000
", + "bbox": [ + 514, + 311, + 905, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of our strategies using SfM point initialization.", + "bbox": [ + 519, + 369, + 897, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a259ff223ed1b144b32b4bf5c8e0da33b2c67e8523d1f9a0edc3449fbf31545.jpg", + "image_caption": [ + "Figure 7. Visual comparison of different initializations and their rendering results after optimization with these initializations." + ], + "image_footnote": [], + "bbox": [ + 517, + 400, + 898, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 657, + 633, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents SPC-GS, a novel approach to address the challenges of GS-based indoor open-world free-view synthesis with sparse views. In SPC-GS, the proposed Scene layout-based Gaussian Initialization (SGI) effectively enhances Gaussian representation by providing an instructive point initialization, consequently promoting semantic learning. Additionally, the Semantic-Prompt Consistency (SPC) regularization mitigates limited supervision by introducing additional view-consistent semantic constraints, improving coherent segmentation performance. Experimental results show that our approach significantly outperforms SOTA methods in reconstruction and 3D open-world segmentation tasks across synthetic and real-world datasets with various CLIP models.", + "bbox": [ + 509, + 684, + 906, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11271", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3", + "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6", + "[6] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Proceedings of the European Conference on Computer Vision, pages 370-386. Springer, 2024. 3", + "[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 3", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 1, 3, 4", + "[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(4): 1, 2017. 2", + "[10] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2", + "[11] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In Proceedings of the International Conference on Learning Representations, 2022. 3", + "[12] Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3d gaussian splatting as new era: A survey. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 2", + "[14] Xu Hu, Yuxi Wang, Lue Fan, Junsong Fan, Junran Peng, Zhen Lei, Qing Li, and Zhaoxiang Zhang. Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857, 2024. 3", + "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[16] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 10608-10615, 2023. 1", + "[17] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2", + "[18] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Ganesh Iyer, Soroush Saryazdi, Tao Chen, Alaa Maalouf, Shuang Li, Nikhil Varma Keetha, Ayush Tewari, Joshua Tenenbaum, Celso de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. In Proceedings of Robotics: Science and Systems, 2023. 1", + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4, 6", + "[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 3", + "[21] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3", + "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Proceedings of the Advances in Neural Information Processing Systems, pages 23311-23330, 2022. 3", + "[23] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3", + "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proceedings of the International Conference on Learning Representations, 2022. 3, 7, 6, 8, 10", + "[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2, 3, 6, 7, 9, 10" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11272", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Guibiao Liao, Wei Gao, Qiuping Jiang, Ronggang Wang, and Ge Li. Mmnet: Multi-stage and multi-scale fusion network for rgb-d salient object detection. In Proceedings of the ACM international conference on multimedia, pages 2436-2444, 2020. 1", + "[27] Guibiao Liao, Jiankun Li, Zhenyu Bao, Xiaqing Ye, Jingdong Wang, Qing Li, and Kanglin Liu. Clip-gs: Clip-informed gaussian splatting for real-time and view-consistent 3d semantic understanding. arXiv preprint arXiv:2404.14249, 2024. 2", + "[28] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3351-3359, 2024. 1", + "[29] Guibiao Liao, Kaichen Zhou, Zhenyu Bao, Kanglin Liu, and Qing Li. Ov-nerf: Open-vocabulary neural radiance fields with vision and language foundation models for 3d semantic understanding. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2", + "[30] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv preprint arXiv:2403.03174, 2024. 1", + "[31] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In Proceedings of the Advances in Neural Information Processing Systems, pages 53433-53456, 2023. 2, 3, 7, 9, 10", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 2", + "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41 (4):1-15, 2022. 2", + "[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3", + "[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 3, 7, 2, 9", + "[36] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 3", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 8, 9", + "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5, 6", + "[39] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splattering segmentation solved optimally. In Proceedings of the European Conference on Computer Vision, pages 456-472. Springer, 2024. 3", + "[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193-13203, 2024. 3, 4", + "[41] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 3", + "[42] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8, 1, 3, 4", + "[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In Proceedings of the International Conference on 3D Vision, pages 443-453, 2022. 3", + "[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3", + "[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2", + "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[47] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In Proceedings of the ACM SIGGRAPH, pages 1-11, 2024. 4, 1, 2", + "[48] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11273", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2", + "[49] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3", + "[50] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In Proceedings of the European Conference on Computer Vision, pages 162-179. Springer, 2024. 1, 3, 7, 9, 10", + "[51] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20612-20622, 2024. 3", + "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2", + "[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3, 4", + "[54] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2", + "[55] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In Proceedings of the European Conference on Computer Vision, pages 57-74. Springer, 2024. 5", + "[56] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatt-ting via co-regularization. In Proceedings of the European Conference on Computer Vision, pages 335–352. Springer, 2024. 3, 7, 9, 10", + "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[58] Xiaoyu Zhang, Guibiao Liao, Wei Gao, and Ge Li. Tdrnet: Transformer-based dual-branch restoration network for geometry based point cloud compression artifacts. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1–6. IEEE, 2022. 1", + "[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 1, 3, 7, 2, 9, 10" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In Proceedings of the European Conference on Computer Vision, pages 145-163. Springer, 2024. 2, 3, 5, 6, 7, 9, 10", + "[61] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. arXiv preprint arXiv:2401.01970, 2024. 3" + ], + "bbox": [ + 516, + 90, + 905, + 219 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11274", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_model.json b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a75d6f8f87170cd59bd130b83486b21ca59685d5 --- /dev/null +++ b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "SPC-GS: Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.203, + 0.276, + 0.221 + ], + "angle": 0, + "content": "Guibiao Liao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.204, + 0.389, + 0.222 + ], + "angle": 0, + "content": "Qing Li²" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.204, + 0.547, + 0.222 + ], + "angle": 0, + "content": "Zhenyu Bao\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.587, + 0.204, + 0.701, + 0.222 + ], + "angle": 0, + "content": "Guoping Qiu3" + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.204, + 0.855, + 0.222 + ], + "angle": 0, + "content": "Kanglin Liu\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.222, + 0.769, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electronic and Computer Engineering, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.24, + 0.71, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)Pengcheng Laboratory, \\(^{3}\\)University of Nottingham" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.298, + 0.245, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.387, + 0.164, + 0.392 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.298, + 0.414, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.263, + 0.387, + 0.416, + 0.397 + ], + "angle": 0, + "content": "(a1) SfM Points (Gau-Grouping)" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.298, + 0.569, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.387, + 0.697, + 0.397 + ], + "angle": 0, + "content": "(a2) Rendered RGB/Semantic Results (Gau-Grouping)" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.303, + 0.754, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.304, + 0.83, + 0.315 + ], + "angle": 0, + "content": "Gau-Grouping" + }, + { + "type": "image", + "bbox": [ + 0.835, + 0.304, + 0.855, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.86, + 0.304, + 0.885, + 0.315 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.401, + 0.245, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.489, + 0.22, + 0.5 + ], + "angle": 0, + "content": "Sparse Input Images" + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.4, + 0.415, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.489, + 0.415, + 0.5 + ], + "angle": 0, + "content": "(b1) Scene layout Points (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.4, + 0.569, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.489, + 0.675, + 0.5 + ], + "angle": 0, + "content": "(b2) Rendered RGB/Semantic Results (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.399, + 0.716, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.332, + 0.779, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.332, + 0.834, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.846, + 0.332, + 0.892, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.488, + 0.872, + 0.498 + ], + "angle": 0, + "content": "(c) Comparison Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.907, + 0.564 + ], + "angle": 0, + "content": "Figure 1. Visual comparisons of open-world free-view synthesis from sparse inputs (12 training views). The previous 3DGS-based method, Gau-Grouping [50], utilizes sparse SfM points for Gaussian initialization and limited training views for supervision, leading to inferior rendering results. In contrast, our approach leverages scene layout-based points for instructive initialization, and cooperates with semantic-prompt consistency constraints (detailed in Sec. 3.3), yielding superior reconstruction and segmentation results." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.576, + 0.327, + 0.592 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.607, + 0.486, + 0.88 + ], + "angle": 0, + "content": "3D Gaussian Splatting-based indoor open-world free-view synthesis approaches have shown significant performance with dense input images. However, they exhibit poor performance when confronted with sparse inputs, primarily due to the sparse distribution of Gaussian points and insufficient view supervision. To relieve these challenges, we propose SPC-GS, leveraging Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) Regularization for open-world free view synthesis with sparse inputs. Specifically, SGI provides a dense, scene-Layout-based Gaussian distribution by utilizing view-changed images generated from the video generation model and view-constraint Gaussian points densification. Additionally, SPC mitigates limited view supervision by employing semantic-prompt-based consistency constraints developed by SAM2. This approach leverages available semantics from training views, serving as instructive prompts, to optimize visually overlapping regions in novel views with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.684 + ], + "angle": 0, + "content": "2D and 3D consistency constraints. Extensive experiments demonstrate the superior performance of SPC-GS across Replica and ScanNet benchmarks. Notably, our SPC-GS achieves a 3.06 dB gain in PSNR for reconstruction quality and a \\(7.3\\%\\) improvement in mIoU for open-world semantic segmentation. Project website at: https://gbliao.github.io/SPC-GS.github.io." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.646, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting-based open-world free-view synthesis seeks to construct semantic Gaussian radiance fields using a set of 3D Gaussian primitives enriched with color and semantic attributes. This task enables the free novel view synthesis of both realistic appearances and open-vocabulary semantic understanding (also referred to as 3D open-vocabulary semantic segmentation) of the scenes, with significant implications for augmented/virtual reality, robot navigation, and manipulation [16, 18, 26, 28, 30, 58]. Existing methods [35, 50, 59] generally require dense, large sets of input views to train the semantic Gaussian radiance" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.245, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11264" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "field for comprehensive scene coverage. However, when only sparse views are available, these approaches exhibit substantial declines in reconstruction and segmentation accuracy. This is mainly attributed to two key limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.365 + ], + "angle": 0, + "content": "1) Sparsely Distributed Gaussian Points: With sparse input views, the initialized Gaussian points via Structure-from-Motion (SfM) in 3DGS [19] are inherently few and sparsely distributed (Fig. 1 (a1)). Such sparse point distribution imposes challenges in optimizing 3D Gaussians to depict complex indoor environments, leading to suboptimal reconstruction quality and, consequently, impaired segmentation accuracy (Fig. 1 (a2)). 2) Insufficient View Constraints: Optimizing 3D Gaussian parameters with inadequate view supervision often converges to local minima that overfit training views [25, 60]. This results in semantic ambiguity and poor performance in under-observed regions when rendering from novel viewpoints, due to limited reliable semantic integration across different views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.428 + ], + "angle": 0, + "content": "To address the above challenges, we propose SPC-GS, a novel framework that incorporates Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.641 + ], + "angle": 0, + "content": "To seek denser points for improved Gaussian initialization, SGI produces a structured scene layout Gaussian distribution in two steps. First, SGI employs a video generation model to create view-changed images proximal to training images. These generated images, combined with training views, enrich feature matching in SfM, yielding denser SfM points that promote scene layout Gaussian generation. Second, these denser SfM points undergo view-constraint Gaussian densification and outlier removal to generate structured scene layout points. These points provide an instructive initialization that faithfully captures indoor scene geometry (e.g. furniture, walls, ceilings), enhancing 3DGS training and scene representation, thereby improving segmentation accuracy (Fig. 1 (b1) and (b2))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Building on dense scene layout Gaussian initialization, SPC enhances supervision through view-to-view semantic consistency constraints. The core of SPC lies in utilizing informative semantics signals from known training views, serving as semantic prompts, to optimize online-rendered novel (pseudo) views, establishing pseudo-multiview constraints through training-to-pseudo semantic consistency. To achieve it, we utilize the powerful Segment Anything Model 2 (SAM2) [38] to build region mask correspondences across training and pseudo views. Using these correspondences, we leverage the semantics prompts provided by training views to enforce semantic consistency constraints on the corresponding regions of pseudo views at 2D and 3D spaces. This approach substantially augments supervision signals, leading to more coherent segmentation on novel views under sparse input conditions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Prior works on 3D open-vocabulary segmentation or" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "sparse-input free-view synthesis mainly focuses on object-centric or face-forward scenarios, such as LERF [20], and 3DOVS [31], where objects are centrally positioned, and views are sampled in an \"outside-in\" manner, resulting in significant view overlap. In contrast, this study focuses on sparse-input open-world free-view synthesis within indoor scenes, particularly relevant for robotics and augmented/virtual reality applications. To this end, we utilize the widely recognized Replica [42] and ScanNet [8] datasets, which feature representative indoor scenes characterized by an \"inside-out\" viewing configuration. This setup leads to less overlap between certain views, posing a substantial challenge for sparse-input open-world free-view synthesis. The contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.305, + 0.905, + 0.35 + ], + "angle": 0, + "content": "- We propose SPC-GS, an early and novel attempt at jointly reconstructing and understanding open-world indoor scenes from sparse-input images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.35, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- We introduce Scene layout-based Gaussian Initialization (SGI) to tackle the challenge of sparsely distributed Gaussians, offering a dense and instructive scene layout point initialization that promotes semantic Gaussian learning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.411, + 0.905, + 0.485 + ], + "angle": 0, + "content": "- We present Semantic-Prompt Consistency (SPC) regularization to alleviate inadequate view supervision by enforcing semantic consistency constraints on unseen pseudo views through prompts derived from training views, enhancing coherent segmentation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.486, + 0.905, + 0.532 + ], + "angle": 0, + "content": "- Extensive experiments on sparse-input scenes show the superiority of our approach in free-view synthesis of realistic appearances and open-vocabulary segmentation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.305, + 0.905, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.543, + 0.655, + 0.559 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.719 + ], + "angle": 0, + "content": "Free-view Synthesis via Radiance Fields. In recent years, radiance field-based free-view synthesis has made remarkable advancements. A Notable example is the Neural Radiance Field (NeRF) [13, 32], which leverages a coordinate-based multi-layer perceptron (MLP) to implicitly model 3D scenes, enabling novel view synthesis via volume rendering [10]. Another innovative approach is 3D Gaussian Splating (3DGS) [19, 48], which utilizes a collection of 3D Gaussians to represent the scene and incorporates a fast tile-based rasterizer technique, allowing real-time rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Although current methods achieve impressive results in dense-view scenarios [4, 12, 15, 27, 29, 33, 54], their performance degrades significantly with sparse inputs [17, 60]. Sparse-input Free-view Synthesis. Recent efforts have tried to address the sparse-input view synthesis and they can broadly categorized into cross-scene pre-training methods and per-scene optimization methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Pre-training methods learn scene priors by pre-training on large-scale datasets spanning multiple scenes. For instance, pixelNeRF [52] conditions a NeRF on CNN features from input images during volume rendering. Similarly, several approaches [3, 45] utilize different NeRF conditions" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "to facilitate cross-scene pre-training. Inspired by the success of 3DGS, recent methods such as pixelSplat [2] and MVSplat [6] integrate efficient CUDA-based rasterization with 3D Gaussian representations for pre-training. However, these works demand substantial training resources and large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.35 + ], + "angle": 0, + "content": "Per-scene optimization methods aim to rapidly reconstruct a 3D scene by introducing effective regularizations. Early NeRF-based methods enhance sparse-input view synthesis through depth-based [34, 44] and frequency range regularization [49]. Recent 3DGS-based works focus on improving rendering efficiency. [7] designs depth regularization loss using monocular depth. DnGaussian [25] and FSGS [60] enhance sparse Gaussian representations through global-local depth normalization and correlation depth distribution losses. CoR-GS [56] adopts color co-regularization loss using rendered pseudo views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Instead, we focus on enhancing view synthesis and openworld 3D indoor semantic understanding from sparse inputs following the per-scene optimization route." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.488 + ], + "angle": 0, + "content": "3D Open-world Understanding via Radiance Fields. Recent studies have explored integrating 2D visual foundation models (e.g. DINO [1] and CLIP [24, 37]) with radiance fields to facilitate 3D open-world scene understanding. These approaches can be broadly divided into NeRF-based and 3DGS-based methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.596 + ], + "angle": 0, + "content": "NeRF-based. Early methods [11, 43] distill visual features from DINO into semantic radiance fields. More recent language-embedded neural field methods [20, 22] integrate CLIP's visual features into semantic fields, enabling language-driven 3D semantic understanding. Additionally, 3DOVS [31] further improves semantic accuracy through the proposed Relevancy-Distribution Alignment loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.854 + ], + "angle": 0, + "content": "3DGS-based. Alternative to NeRFs, recent works explore embedding CLIP features into 3D Gaussians for more efficient semantic rendering. Feature 3DGS [59] directly attaches CLIP embeddings to 3D Gaussians, while FMGS [61] combines Gaussians with multi-resolution hash encodings to form the semantic Gaussian field. [41] embeds quantized CLIP features into Gaussians to reduce storage demand. Furthermore, LangSplat [35] learns low-dimensional semantic Gaussian embeddings and uses a deep neural network for post-process upsampling to obtain semantic representation. Gau-Grouping [50] introduces a 2D object identity loss within training views to optimize Gaussians, while this method requires dense view supervision. Note that the 3D interactive segmentation methods, such as [14, 21, 39, 51], lie outside the scope of language-driven open-world segmentation, as they rely on 2D point prompts to generate object masks but lack semantic meaning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While previous approaches rely on dense input images, our work focuses on GS-based 3D open-world scene understanding in sparse-input indoor scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Fig. 2 illustrates our framework, which consists of two primary components. First, Scene layout-based Gaussian Initialization (SGI) (Section 3.2) constructs a dense, instructive scene layout Gaussian initialization that improves Gaussian representation, thus supporting enhanced semantic learning. Then, using the initialized scene layout Gaussians, Semantic-Prompt Consistency (SPC) Regularization (Section 3.3) introduces 2D and 3D semantic consistency constraints, creating effective pseudo-multi-view optimization for semantic learning under sparse-input conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.275, + 0.644, + 0.291 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Problem Definition. Our approach aims to leverage sparse input images to reconstruct 3D indoor scenes and enable language-driven semantic perception. Concretely, given \\(L\\) posed sparse input images \\(I = \\{I_i\\}_{i=1}^L \\in \\mathbb{R}^{L \\times H \\times W \\times 3}\\) and language descriptions, our goal is to generate both realistic appearance and open-world semantic understanding results for specified novel viewpoints. To achieve this, we exploit 3D Gaussians [19] for scene representation, and utilize CLIP models [24, 37] to provide semantic supervision. Preliminary. 3D Gaussian Splitting (3DGS) [19] represents a 3D scene using a collection of explicit 3D Gaussian primitives \\(\\{\\mathbf{g}_i = (\\mathbf{p}_i, \\mathbf{c}_i, o_i, \\mathbf{s}_i, \\mathbf{q}_i)\\}_{i=1}^{N_K}\\), where each primitive comprises a 3D position \\(\\mathbf{p} \\in \\mathbb{R}^3\\), color \\(\\mathbf{c} \\in \\mathbb{R}^3\\), opacity \\(o \\in \\mathbb{R}\\), scaling factor \\(\\mathbf{s} \\in \\mathbb{R}^3\\), and rotation quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\). For pixel color \\(\\hat{I}(x_p)\\) rendering, 3DGS employs \\(\\alpha\\)-blending point-based rendering [23], formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.548, + 0.906, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\hat {I} \\left(x _ {p}\\right) = \\sum_ {i \\in \\mathcal {N}} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.59, + 0.906, + 0.642 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}\\) represents Gaussian points overlapping pixel \\(x_{p}\\), \\(\\alpha_{i} = o_{i}G_{i}^{proj}\\) with \\(G_{i}^{proj}\\) as the \\(i\\)-th Gaussian's 2D projection, and transmittance \\(T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.906, + 0.688 + ], + "angle": 0, + "content": "For pixel semantic feature \\(\\hat{\\mathbf{F}}(x_p)\\) rendering, following [35, 59], we attach a learnable semantic parameter \\(\\mathbf{f}_i \\in \\mathbb{R}^D\\) to each Gaussian and use the above \\(\\alpha\\)-blending:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.698, + 0.906, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} (x _ {p}) = \\sum_ {i \\in \\mathcal {N}} T _ {i} \\alpha_ {i} \\mathbf {f} _ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.906, + 0.814 + ], + "angle": 0, + "content": "Following [31], we use a CLIP text encoder to generate textual feature \\(\\mathbf{T} \\in \\mathbb{R}^{M \\times 512}\\) for \\(M\\) classes and obtain rendered segmentation logits \\(\\hat{\\mathbf{S}}\\) as follows: \\(\\hat{\\mathbf{S}} = \\cos\\langle \\omega_f(\\hat{\\mathbf{F}}), \\mathbf{T} \\rangle\\), where \\(\\cos\\) denotes cosine similarity and \\(\\omega_f\\) is a \\(1 \\times 1\\) convolution layer for feature dimension alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.906, + 0.875 + ], + "angle": 0, + "content": "We incorporate reconstruction and semantic losses to train the 3D Gaussians. The color reconstruction loss, denoted as \\(\\mathcal{L}_C\\), follows the formulation in 3DGS [19], and the semantic loss, \\(\\mathcal{L}_S\\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.884, + 0.906, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\mathcal {L} _ {c o s} \\left(\\omega_ {f} (\\hat {\\mathbf {F}}), \\mathbf {F}\\right) + \\mathcal {L} _ {c e} \\left(\\omega_ {s} (\\hat {\\mathbf {S}}), \\mathbf {S}\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11266" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.232, + 0.091, + 0.506, + 0.104 + ], + "angle": 0, + "content": "Video Generation Model based Point Creation" + }, + { + "type": "header", + "bbox": [ + 0.676, + 0.092, + 0.86, + 0.104 + ], + "angle": 0, + "content": "Scene-Layout Point Generation" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.108, + 0.904, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.253, + 0.68, + 0.267 + ], + "angle": 0, + "content": "Scene layout-based Gaussian Initialization (SGI)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.267, + 0.905, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.471, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Figure 2. Framework of SPC-GS. (a) We first generate adjacent images of each training image using the video generation model [47]. These generated images, combined with the original training images, produce denser initialized SfM points. These points are then optimized to create a scene layout Gaussian distribution via Gaussian densification and outlier removal. (b) Building on the scene-Layout Gaussian initialization, SPC leverages semantic information from training views as instructive semantic prompts to optimize adjacent rendered pseudo views, establishing semantic consistency constraints that enhance overall sparse-input semantic understanding of 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.613 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}\\) and \\(\\mathbf{S}\\) are CLIP-derived semantic features and relevancy maps, respectively. \\(\\omega_{s}\\) is a \\(1\\times 1\\) convolution layer, and \\(\\mathcal{L}_{cos}\\) and \\(\\mathcal{L}_{ce}\\) denote cosine and cross-entropy losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.453, + 0.637 + ], + "angle": 0, + "content": "3.2. Scene layout-based Gaussian Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.689 + ], + "angle": 0, + "content": "We propose Scene layout-based Gaussian Initialization (SGI) to produce a dense and conducive point distribution for initialization with the following insights and designs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.825 + ], + "angle": 0, + "content": "VGM-based Point Creation. In static indoor scenes, viewpoint variations correspond to camera movements. Motivated by this, we employ a video generation model (VGM) to create additional views beyond the sparse training set through image-to-video generation. These generated views augment the available scene information and, when combined with the original training views, enrich feature matching during SfM, yielding denser SfM points for generating scene layout Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specifically, for each training image \\(I_{i}\\), we generate eight adjacent views \\(\\{\\tilde{I}_i^j\\}_{j = 1}^8\\) using the image-to-video generation mode of the powerful video generation model MotionCtrl [47]. This is accomplished by inputting the training images along with eight standard camera motion trajectory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.763 + ], + "angle": 0, + "content": "ries [47] into MotionCtrl. These trajectories include pan up, pan down, pan left, pan right, zoom in, zoom out, clockwise rotation, and anticlockwise rotation. Each trajectory generates a sequence of consecutive video frames. From each sequence, we select a single frame temporally adjacent to the input training image as the valid generated image since it exhibits superior visual quality. These generated images are then combined with the original training images for SfM processing, yielding denser SfM points for Gaussian initialization (see supplementary file). Despite being denser than SfM points derived from sparse-input training images, these points remain limited in capturing the geometric layout and details of indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Scene-query Point Generation. We further generate scene-query point distributions to enhance 3DGS training and scene representation through the following steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "i) View-constraints 3D Gaussian Densification. Using the denser SfM point initialization, we follow Gaussian densification [19] and using our view constraints from Eq. (9) to generate a more densely distributed, instructive scenelayout Gaussian point set \\(\\{\\mathbf{g}_i = (\\mathbf{p}_i,\\mathbf{c}_i,\\mathbf{f}_i,o_i,\\mathbf{s}_i,\\mathbf{q}_i)\\}_{i = 1}^{N_{K'}}\\) as shown in Fig. 2. These denser \\(N_{K^{\\prime}}\\) points provide an" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11267" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "improved spatial layout of indoor scenes, and enhanced attribute representations (e.g. \\(\\mathbf{c}_i\\), \\(\\mathbf{f}_i\\)), making them effective priors for Gaussian initialization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.483, + 0.333 + ], + "angle": 0, + "content": "ii) Outlier Gaussian Primitive Removal (OGR). During Gaussian densification, new Gaussian primitives are sampled based on a 3D Gaussian probability distribution, which may generate outlier primitives positioned far from the actual scene layout. These outliers, when used in subsequent initialization, complicate the optimization process. To mitigate this, we present a density-based removal strategy to remove those sparse outlier points. Specifically, Gaussian primitives with fewer than \\(\\mathcal{P}\\) neighboring points within a spherical radius \\(r\\) are periodically eliminated during training. Empirical experiments show that setting \\(\\mathcal{P}\\) to 5 and \\(r\\) to 1 strikes a balance between effective outlier removal and the preservation of essential scene elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.396 + ], + "angle": 0, + "content": "By using instructive scene layout-based points for initialization, the 3DGS training and scene representation can be enhanced, consequently improving segmentation performance under the sparse-input condition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.479, + 0.425 + ], + "angle": 0, + "content": "3.3. Semantic-Prompt Consistency Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.507 + ], + "angle": 0, + "content": "We introduce Semantic-Prompt Consistency (SPC) regularization to exploit additional supervision signals. The key idea of SPC is to leverage semantics from training views as prompts, to impose effective semantic consistency constraints on corresponding regions of pseudo views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.583 + ], + "angle": 0, + "content": "Region Correspondence Establishment. To establish correspondences between training and pseudo views, we draw inspiration from the powerful zero-shot temporal region proposal extraction capability of Segment Anything Model 2 (SAM2) [38] and leverage it to achieve this." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Specifically, we introduce Iterative Stochastic Prompting (ISP) to efficiently build correspondence. During training, ISP renders pseudo (unobserved) views using the online training SPC-GS. The view locations are sampled from the two nearest known training views by adding random noise following [60]. Then, ISP treats the current training image and its associated pseudo-views as a viewpoint-changed video sequence, and selects a stochastic point coordinate in the training image, as the 2D point prompt for SAM2, generating corresponding region masks across both the training and pseudo views, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.763, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i}, \\left\\{\\mathbf {M} _ {i} ^ {*} \\right\\} _ {p = 1} ^ {P} = \\mathbf {I S P} \\left(I _ {i}, \\left\\{I _ {i} ^ {*} \\right\\} _ {p = 1} ^ {P}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( I_{i}^{*} \\) is the rendered pseudo image and \\( \\mathbf{M}_i^* \\) is the corresponding region mask within \\( I_{i}^{*} \\). \\( P \\) is set to 2 for computational efficiency. As training progress proceeds, the stochastic sampling mechanism encourages a uniform distribution of points across the entire image space, facilitating the efficient region mask correspondence construction for diverse regions of the whole image, as shown in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.707, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.901, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.187, + 0.906, + 0.244 + ], + "angle": 0, + "content": "Figure 3. Illustration of Iterative Stochastic Prompting (ISP). During training, ISP randomly samples point coordinates \\((\\star)\\) from the training view, serving as the point prompt for SAM2 to produce region masks (highlighted in blue) across training and pseudo views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.906, + 0.331 + ], + "angle": 0, + "content": "Semantic-Prompt for 2D Constrains. Building on the established region mask correspondences, we enforce effective constraints within 2D views to enhance Gaussian learning with sparse inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.906, + 0.392 + ], + "angle": 0, + "content": "i) View-to-View Regularization. Leveraging the semantics from training views as prompts, we enforce training-to-pseudo semantic consistency constraints in feature and relevancy map aspects, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.404, + 0.906, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {i n t e r} = \\sum_ {p = 1} ^ {P} | | \\frac {\\hat {\\mathbf {F}} _ {p} ^ {*} \\odot \\mathbf {M} _ {p} ^ {*}}{\\sum \\mathbf {M} _ {p} ^ {*}} - \\frac {\\hat {\\mathbf {F}} \\odot \\mathbf {M}}{\\sum \\mathbf {M}} | | _ {2} \\\\ - \\sum_ {p = 1} ^ {P} \\mathcal {R} [ \\Phi (\\hat {\\mathbf {S}} \\odot \\mathbf {M}) ] \\log (\\hat {\\mathbf {S}} _ {p} ^ {*} \\odot \\mathbf {M} _ {p} ^ {*}), \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.735 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{F}}^*\\) and \\(\\hat{\\mathbf{S}}^*\\) represent the semantic feature and the relevancy map derived from pseudo views. The function \\(\\Phi\\) is the semantic uniformity operation: \\(\\Phi (\\hat{\\mathbf{S}}\\odot \\mathbf{M}) =\\) \\(\\mathrm{argmax}_c\\sum_{i\\in \\mathbf{M}}\\mathbb{1}\\{\\mathrm{argmax}(\\omega_s(\\hat{\\mathbf{S}}_i)) = c\\}\\), which integrates SAM2's masks with a maximum voting mechanism to select the dominant class \\(c\\) within the region mask M, unifying the region's semantics and enforcing consistent region-wise semantic supervision. \\(\\mathcal{R}[\\cdot ]\\) denotes the dimensional reshaping operation to align the dimension shape of the region mask \\(\\mathbf{M}^*\\) from the pseudo view. In \\(\\mathcal{L}_{inter}\\), we utilize self-rendered semantic \\((\\hat{\\mathbf{F}},\\hat{\\mathbf{S}})\\) from training views, generated by the trained SPC-GS, rather than 2D CLIP semantics. This is motivated by the superior coherence and reliability of 3D representations over their 2D counterparts [55] (see supplementary file for experimental validation)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.809 + ], + "angle": 0, + "content": "ii) Contrastive Regularization. Beyond inter-view, we introduce semantic-guided contrastive learning regularization within intra-training-views to further enhance supervision. This is based on the principle that features of semantically consistent objects should be similar." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given \\(N\\) feature samples, each sample \\(\\mathbf{z}\\) is derived from \\(\\mathbf{z} = W_{\\psi}(\\hat{\\mathbf{F}})\\), where \\(W_{\\psi}\\) is a \\(1\\times 1\\) convolution transformation layer. Let \\(i\\in U_z\\equiv \\{1\\dots N\\}\\) be the index of a sample, and \\(T_{z}(i)\\equiv U_{z}\\backslash \\{i\\}\\) represents the index set excluding index \\(i\\). We designate features with the same rendered semantic label \\(\\tilde{\\mathbf{q}}\\) as positive samples, where \\(\\tilde{\\mathbf{q}} =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11268" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.258 + ], + "angle": 0, + "content": "\\(\\Phi (\\hat{\\mathbf{S}}\\odot \\{\\mathbf{M}^k\\}_{k = 1}^\\eta)\\). Here, the \\(\\{\\mathbf{M}^k\\}_{k = 1}^\\eta\\) denote \\(\\eta\\) masks generated by SAM2's automatic mask generator in the entire image, and \\(\\Phi\\) is the function that unifies semantics within each mask. Accordingly, for the \\(i\\)-th sample, positive sample indexes are defined as: \\(P_{z}(i)\\equiv \\{p\\in T_{z}(i):\\tilde{\\mathbf{q}}_{p} = \\tilde{\\mathbf{q}}_{i}\\}\\), and negative sample indexes: \\(T_{z}(i)\\backslash P_{z}(i)\\). Similar to SimCLR [5], we employ contrastive learning loss to maximize agreement between positive samples. This method brings \\(P_{z}\\)-indexed positive samples closer in feature space, imposing effective contrastive constraints on sparse-view Gaussian learning, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.269, + 0.498, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n t r a}} = - \\sum_ {i \\in U _ {z}} \\frac {1}{\\mathcal {M} _ {z} (i)} \\sum_ {p \\in P _ {z} (i)} \\log \\left[ \\frac {e ^ {\\left(\\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {p}\\right) / \\tau\\right)}}{\\sum_ {j \\in T _ {z} (i)} e ^ {\\left(\\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j}\\right) / \\tau\\right)}} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.371 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_z(i)\\) denotes the number of elements in \\(P_{z}(i)\\). The symbol \\(\\cdot\\) refers to the inner product operation. \\(\\tau\\) is the temperature parameter, set to 0.2 in the experiment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.446 + ], + "angle": 0, + "content": "To obtain \\(N\\) samples, we employ a patch-wise sampling strategy by randomly extracting a patch of size \\(h \\times w\\) in each training iteration, where \\(N = h \\times w\\). Empirically, \\(h\\) and \\(w\\) are both set to 128. In summary, the semantic prompt for 2D constraints can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.457, + 0.482, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S P C} ^ {2 d} = \\mathcal {L} _ {\\text {i n t e r}} + \\mathcal {L} _ {\\text {i n t r a}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Semantic-Prompt for 3D Constrains. In addition to 2D constraints, we extend semantic-prompt to impose effective 3D constraints on 3D Gaussian primitives. Considering that Gaussian primitives with maximum weight dominate the rendering results (Eq. (2)), we encourage semantic feature similarity between the maximally-weighted 3D Gaussians from two corresponding regions. Specifically, for each pixel within the region mask, we identify the Gaussian primitive with maximum weight along the ray cast into 3D space. Then, the semantic features of these maximally-weighted Gaussians from training views are used to constrain their counterparts in pseudo views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.484, + 0.759 + ], + "angle": 0, + "content": "However, since Gaussians related to region boundaries often span objects with differing semantics, directly applying the above constraints may complicate optimization. To mitigate this, we introduce a Region Boundary Erosion technique, which removes ambiguous primitives corresponding to region boundaries, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.771, + 0.484, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S P C} ^ {3 d} = \\mathcal {L} _ {k l} (\\mathcal {F} | | \\mathcal {F} ^ {*}) = \\sum_ {j = 1} ^ {n ^ {\\prime}} \\bar {\\mathbf {f}} \\log \\left(\\frac {\\bar {\\mathbf {f}}}{\\mathbf {f} _ {j} ^ {*}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) denotes the mean feature representation of 3D Gaussian primitives derived from mask \\(\\mathbf{M}'\\) in the training view, defined as \\(\\overline{\\mathbf{f}} = (\\mathbf{f} \\odot \\mathbf{M}') / \\sum \\mathbf{M}'\\). Here, \\(\\mathbf{M}'\\) refers to the eroded mask after region boundary erosion, which is achieved through morphological erosion. Specifically, a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.199 + ], + "angle": 0, + "content": "pixel \\(\\mathbf{M}'(x,y)\\) is retained as \"True\" only if all pixels within its \\(5\\times 5\\) neighborhood in the input original mask \\(\\mathbf{M}\\) are \"True\"; otherwise, it is set to \"False\". This operation effectively shrinks the boundaries of the region, causing them to contract inward. The set \\(\\mathcal{F}^{*} = \\{\\mathbf{f}_{1}^{*},\\mathbf{f}_{2}^{*},\\dots,\\mathbf{f}_{n^{\\prime}}^{*}\\}\\) comprises \\(n^\\prime\\) Gaussian primitives obtained from eroded masks \\(\\{\\mathbf{M}^{*'}\\}_{p = 1}^{P}\\) in pseudo views." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.199, + 0.816, + 0.214 + ], + "angle": 0, + "content": "To summarize, the overall loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.225, + 0.906, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {C} + \\mathcal {L} _ {S} + \\mathcal {L} _ {G S} + \\mathcal {L} _ {S} ^ {3 d} + \\lambda \\left(\\mathcal {L} _ {S P C} ^ {2 d} + \\mathcal {L} _ {S P C} ^ {3 d}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.257, + 0.907, + 0.473 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{GS}\\) is the semantic loss for generated images from video generation model: \\(\\mathcal{L}_{GS} = \\mathcal{L}_{cos}(\\omega_f(\\hat{\\mathbf{F}}^g),\\mathbf{F}^g) + \\mathcal{L}_{ce}(\\omega_s(\\hat{\\mathbf{S}}^g),\\mathbf{S}^g)\\). These generated RGB images are not used for color constraints due to their error-prone color pixels that degrade overall results (see supplementary file). \\(\\mathcal{L}_S^{3d}\\) is the 3D local adaptive regularization, which promotes semantic similarity among neighboring Gaussians: \\(\\mathcal{L}_S^{3d} = \\frac{1}{\\theta k}\\sum_{i = 1}^{\\theta}\\sum_{j = 1}^{k}\\omega_{ij}(|\\mathbf{f}_i - \\mathbf{f}_j|_2)\\), where \\(\\theta = 800\\) denotes the number of Gaussian randomly sampled per iteration, and \\(k = 5\\) specifies the number of nearest neighbors for each sampled Gaussian in 3D space. \\(\\omega_{ij} = \\exp (-||\\mu_i - \\mu_j||_2)\\) is the distance-based weighting coefficient. To prevent early-stage suboptimal convergence, \\(\\lambda\\) is initialized at 0 and then increased to 1 after 4k iterations during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.646, + 0.505 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.514, + 0.704, + 0.53 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.703 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two widely-used indoor scene datasets in terms of novel view synthesis: Replica [42] and ScanNet [8], using a sparse-input setting with only 12 input images for training. Following the standard sparse-input protocol [25, 60], every 10-th frame from each sequence is selected for testing, while 12 evenly sampled sparse views from the remaining images are used for training. These 12 views collectively encompass the entire scene. Notably, the \"inside-out\" viewing direction and limited overlap among views in sparse-input indoor scenes pose significant challenges for free-view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Evaluation Metrics. We employ PSNR, SSIM [46], and LPIPS [57] for evaluating image quality. We report segmentation performance using the mean Intersection-over-Union (mIoU) and mean pixel accuracy (mAcc)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We implement our approach using the official PyTorch 3D Gaussian Splatting [19] on an NVIDIA A100 GPU. The dimension of the learnable semantic parameter attached to 3D Gaussians is set to 32. We modify the CUDA kernel to enable parallel differentiable rasterization of both color and semantic attributes. We employ CLIP ViT-B/16 [37] for feature extraction and SAM2 Hiera-L [38] for region mask generation. More details are provided in the supplementary file." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.244, + 0.882, + 0.258 + ], + "angle": 0, + "content": "Figure 4. Visual reconstruction results on novel views. Our approach achieves superior global structure and photo-realistic details." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.271, + 0.893, + 0.415 + ], + "angle": 0, + "content": "
MethodReplica [42]ScanNet [8]
PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑
3DOVS [31]13.2000.5150.6633.91321.43918.7210.6430.5447.75136.356
Feature 3DGS [59]17.0590.7000.3597.40730.20918.0880.6560.4247.34738.002
LangSplat [35]17.7150.7010.4203.29218.62118.7800.6630.4763.33924.491
Gau-Grouping [50]17.2970.6910.37615.02850.42219.0060.6750.42115.17257.355
DNGaussian [25]18.5130.7120.40216.65754.40018.3920.6910.48413.91254.814
FSGS [60]18.6940.7250.32816.10153.47420.5740.6960.40614.03156.850
CoR-GS [56]18.9550.7580.31215.93352.47620.8150.7210.39314.46255.812
Ours22.0110.7920.25423.96063.26222.4010.7410.36828.94468.823
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.416, + 0.905, + 0.444 + ], + "angle": 0, + "content": "Table 1. Quantitative results of reconstruction and segmentation on novel views on the Replica and ScanNet datasets, using the CLIP [37] for optimization with 12 training views. Our approach achieves superior performances across all metrics on various datasets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.469, + 0.455, + 0.485 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.718 + ], + "angle": 0, + "content": "We conduct novel view synthesis comparisons using sparse input data in indoor scenes, including two types of methods. i) 3D open-vocabulary segmentation methods, including 3DOVS [31], Feature 3DGS [59], LangSplat [35], and Gau-Grouping [50], are designed for 3D open world semantic understanding. ii) Sparse-input free-view synthesis methods contain DNGaussian [25], FSGS [60], and CoR-GS [56], which are designed to learn Gaussian radiance fields from sparse inputs, focusing primarily on appearance modeling. For concurrent 3D open-world segmentation, we extend these methods by adding semantic properties to Gaussian and incorporating CLIP-based semantic supervision, aligning with the semantic settings of our approach. We adopt consistent evaluations across all methods by employing the same sparse-input training protocol and test views." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.719, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Qualitative Results. i) Reconstruction. As shown in Fig. 4, the results from existing 3D open-vocabulary segmentation methods exhibit significant artifacts, since these methods lack sufficient view constraints. Although sparse-view reconstruction methods introduce geometric constraints for regularization, they struggle to recover photo-realistic details. In comparison, our approach incorporates scene-Layout points prior and semantic consistency constraints, leading to a more detailed appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "ii) Open-world Segmentation. In Fig. 5, current methods struggle to obtain precise and complete results, since they easily encounter ambiguity arising from insufficient seman" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.905, + 0.516 + ], + "angle": 0, + "content": "tic view supervision. In contrast, our method delivers accurate results, benefiting from the SPC regularization that provides effective consistency constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.7 + ], + "angle": 0, + "content": "Quantitative Results. Table 1 shows that, when optimized using the CLIP model [37], our approach significantly outperforms existing methods across all metrics on both synthetic and real datasets. Competing methods struggle under sparse input conditions, due to either limited Gaussian representations or insufficient semantic consistency constraints. Notably, LangSplat faces considerable challenges, as it freezes Gaussian coordinates, scale, and rotation during semantic parameter optimization, which severely impedes semantic learning under sparse inputs. Besides, our method maintains superior results when optimized with CLIP-LSeg [24] (see supplementary file), validating its generalizability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.68, + 0.726 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Key Components. In Table 2, applying the VGM-based Point Creation (PC, #1) and Scene layout Point Generation (SPG, #2) notably improves reconstruction quality over setting #0, and consequently boosts segmentation results, validating SGI's effectiveness in enhancing Gaussian representation. Adding the full Semantic-Prompt Consistency regularization (SPC, #4) further yields significant segmentation gains, showing SPC's capability to address insufficient view constraints and enhance 3D scene understanding. Fig. 6 also shows that adding SGI can reduce artifacts, and applying SPC further enhances segmentation results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11270" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.277, + 0.862, + 0.29 + ], + "angle": 0, + "content": "Figure 5. Visual open-world segmentation results on novel views. Our approach yields more accurate and complete results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.314, + 0.484, + 0.406 + ], + "angle": 0, + "content": "
CaseSGISPCReplica [42]
PCSPG2D3DPSNRSSIMLPIPSmIoUmAcc
#017.1340.6780.36914.89951.255
#1✓18.9240.7240.33416.27952.212
#2✓✓21.1760.7830.26019.40158.281
#3✓✓✓21.8140.7910.25722.84961.362
#4✓✓✓✓22.0110.7920.25423.96063.262
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.408, + 0.419, + 0.422 + ], + "angle": 0, + "content": "Table 2. Ablation study on key components." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.44, + 0.484, + 0.506 + ], + "angle": 0, + "content": "
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aOurs (Full setting)22.0110.7920.25423.96063.262
#bw/o View-to-View Regular.21.6830.7890.25721.24460.914
#cw/o Contrastive Regular.21.8790.7920.25622.32260.086
#dw/o Boundary Erosion21.9500.7920.25422.96461.706
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.508, + 0.462, + 0.521 + ], + "angle": 0, + "content": "Table 3. Ablation results of our SPC with different settings." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.54, + 0.479, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.657, + 0.456, + 0.67 + ], + "angle": 0, + "content": "Figure 6. Visual ablation comparison of key components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.792 + ], + "angle": 0, + "content": "Components in SPC. In Table 3, removing either the 2D view-to-view or contrastive regularization (#b, #c), or omitting the region boundary erosion operation in the 3D constraint (#d), degrades performance. This validates their roles in mitigating insufficient supervision and filtering ambiguous Gaussian primitives spanning different objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Robustness. In Table 4, our strategies also achieve significant gains even with sparse SfM initialization from limited training images, validating their robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Initialization. From top to bottom in Fig. 7, we can see that leveraging denser and more effective points for Gaussian initialization promotes 3DGS training that enhances scene representation, leading to improved rendered results." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.313, + 0.906, + 0.37 + ], + "angle": 0, + "content": "
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aSfM Points Initialization17.1340.6780.36914.89951.255
#b#a + SGI20.6800.7740.27018.16956.928
#c#b + SPC21.5890.7860.26020.76661.000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.371, + 0.898, + 0.385 + ], + "angle": 0, + "content": "Table 4. Results of our strategies using SfM point initialization." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.401, + 0.9, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.607, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of different initializations and their rendering results after optimization with these initializations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.634, + 0.674 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.685, + 0.907, + 0.896 + ], + "angle": 0, + "content": "This paper presents SPC-GS, a novel approach to address the challenges of GS-based indoor open-world free-view synthesis with sparse views. In SPC-GS, the proposed Scene layout-based Gaussian Initialization (SGI) effectively enhances Gaussian representation by providing an instructive point initialization, consequently promoting semantic learning. Additionally, the Semantic-Prompt Consistency (SPC) regularization mitigates limited supervision by introducing additional view-consistent semantic constraints, improving coherent segmentation performance. Experimental results show that our approach significantly outperforms SOTA methods in reconstruction and 3D open-world segmentation tasks across synthetic and real-world datasets with various CLIP models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11271" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[6] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Proceedings of the European Conference on Computer Vision, pages 370-386. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(4): 1, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[10] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In Proceedings of the International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3d gaussian splatting as new era: A survey. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Xu Hu, Yuxi Wang, Lue Fan, Junsong Fan, Junran Peng, Zhen Lei, Qing Li, and Zhaoxiang Zhang. Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 10608-10615, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[17] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[18] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Ganesh Iyer, Soroush Saryazdi, Tao Chen, Alaa Maalouf, Shuang Li, Nikhil Varma Keetha, Ayush Tewari, Joshua Tenenbaum, Celso de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. In Proceedings of Robotics: Science and Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[21] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Proceedings of the Advances in Neural Information Processing Systems, pages 23311-23330, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[23] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proceedings of the International Conference on Learning Representations, 2022. 3, 7, 6, 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2, 3, 6, 7, 9, 10" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11272" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[26] Guibiao Liao, Wei Gao, Qiuping Jiang, Ronggang Wang, and Ge Li. Mmnet: Multi-stage and multi-scale fusion network for rgb-d salient object detection. In Proceedings of the ACM international conference on multimedia, pages 2436-2444, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Guibiao Liao, Jiankun Li, Zhenyu Bao, Xiaqing Ye, Jingdong Wang, Qing Li, and Kanglin Liu. Clip-gs: Clip-informed gaussian splatting for real-time and view-consistent 3d semantic understanding. arXiv preprint arXiv:2404.14249, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3351-3359, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[29] Guibiao Liao, Kaichen Zhou, Zhenyu Bao, Kanglin Liu, and Qing Li. Ov-nerf: Open-vocabulary neural radiance fields with vision and language foundation models for 3d semantic understanding. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[30] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv preprint arXiv:2403.03174, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[31] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In Proceedings of the Advances in Neural Information Processing Systems, pages 53433-53456, 2023. 2, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41 (4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.635, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 3, 7, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[39] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splattering segmentation solved optimally. In Proceedings of the European Conference on Computer Vision, pages 456-472. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193-13203, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.374, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[41] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[42] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8, 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.907, + 0.586 + ], + "angle": 0, + "content": "[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In Proceedings of the International Conference on 3D Vision, pages 443-453, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[47] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In Proceedings of the ACM SIGGRAPH, pages 1-11, 2024. 4, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[48] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11273" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[49] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[50] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In Proceedings of the European Conference on Computer Vision, pages 162-179. Springer, 2024. 1, 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[51] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20612-20622, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.45, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[54] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[55] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In Proceedings of the European Conference on Computer Vision, pages 57-74. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[56] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatt-ting via co-regularization. In Proceedings of the European Conference on Computer Vision, pages 335–352. Springer, 2024. 3, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[58] Xiaoyu Zhang, Guibiao Liao, Wei Gao, and Ge Li. Tdrnet: Transformer-based dual-branch restoration network for geometry based point cloud compression artifacts. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1–6. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 1, 3, 7, 2, 9, 10" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[60] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In Proceedings of the European Conference on Computer Vision, pages 145-163. Springer, 2024. 2, 3, 5, 6, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[61] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. arXiv preprint arXiv:2401.01970, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11274" + } + ] +] \ No newline at end of file diff --git a/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_origin.pdf b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90489cba7195096e738e8b46119730770886f485 --- /dev/null +++ b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/d312a2c0-c1b5-4f3c-8ba0-826d0bc6f893_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:206f73e4771d4fbafe777c6457af419fd06921262c62fcfe6123fe47e6ebc41c +size 2056584 diff --git a/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/full.md b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31d3368563a56435a28922aab6867e0ff60e8a03 --- /dev/null +++ b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/full.md @@ -0,0 +1,358 @@ +# SPC-GS: Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs + +Guibiao Liao $^{1,2}$ + +Qing Li² + +Zhenyu Bao $^{1,2}$ + +Guoping Qiu3 + +Kanglin Liu $^{2*}$ + +$^{1}$ School of Electronic and Computer Engineering, Peking University + +$^{2}$ Pengcheng Laboratory, $^{3}$ University of Nottingham + +![](images/56b1a999e97bfb27559c4b4d5347d830b2f0465d5b5b64c6cc5c4452ebb5dea0.jpg) +# + +![](images/79e834178e20de645a65d1a7d454339e0542438ce41514e605ee1c9035cb6ed3.jpg) +(a1) SfM Points (Gau-Grouping) + +![](images/7ace5f62f9e314a28d2308fb2950cac66f2b24f8e4a89212f00772bb1862c549.jpg) +(a2) Rendered RGB/Semantic Results (Gau-Grouping) + +![](images/6134cebc0dc6fc47e4785696a7f028feec51412a6e498a37acfa61c3f457a023.jpg) +Gau-Grouping + +![](images/6b9018d89f4da445abb28aab4c6909a9c87a5aa03350d2f9991a98ca2fee36f5.jpg) +Ours + +![](images/89fede60bb0b3e956fa0a0b57b0e4d5f4c5e37f5ff2fbf2cd7b49645457a43d8.jpg) +Sparse Input Images + +![](images/94781e75ea003cab1e969ead9a69c792df92c5ea0b55d2442263f30ef94fdd45.jpg) +(b1) Scene layout Points (Ours) + +![](images/a7f9783f58807b6d66b40ca97ea016f90c9ec11593b055df52772fdbd8f239af.jpg) +Figure 1. Visual comparisons of open-world free-view synthesis from sparse inputs (12 training views). The previous 3DGS-based method, Gau-Grouping [50], utilizes sparse SfM points for Gaussian initialization and limited training views for supervision, leading to inferior rendering results. In contrast, our approach leverages scene layout-based points for instructive initialization, and cooperates with semantic-prompt consistency constraints (detailed in Sec. 3.3), yielding superior reconstruction and segmentation results. + +![](images/7e66b74017f05f01a248383a1395bf1ca8519b6a31cca494c2f1cd2917b56840.jpg) +(b2) Rendered RGB/Semantic Results (Ours) + +![](images/aa5f7a25599a8d4cc179e40a1774f231f750d9be53a16c0ee09ec72529d19c00.jpg) +(c) Comparison Results + +![](images/eda3cd88212d2e53152452c01ce62cd50438a8dd1ea11730914c7adda18ebc6d.jpg) + +![](images/9011ed3c69227ae8905b608d334c5339e8d43626e25997ce1733eb2a3ef8823d.jpg) + +# Abstract + +3D Gaussian Splatting-based indoor open-world free-view synthesis approaches have shown significant performance with dense input images. However, they exhibit poor performance when confronted with sparse inputs, primarily due to the sparse distribution of Gaussian points and insufficient view supervision. To relieve these challenges, we propose SPC-GS, leveraging Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) Regularization for open-world free view synthesis with sparse inputs. Specifically, SGI provides a dense, scene-Layout-based Gaussian distribution by utilizing view-changed images generated from the video generation model and view-constraint Gaussian points densification. Additionally, SPC mitigates limited view supervision by employing semantic-prompt-based consistency constraints developed by SAM2. This approach leverages available semantics from training views, serving as instructive prompts, to optimize visually overlapping regions in novel views with + +2D and 3D consistency constraints. Extensive experiments demonstrate the superior performance of SPC-GS across Replica and ScanNet benchmarks. Notably, our SPC-GS achieves a 3.06 dB gain in PSNR for reconstruction quality and a $7.3\%$ improvement in mIoU for open-world semantic segmentation. Project website at: https://gbliao.github.io/SPC-GS.github.io. + +# 1. Introduction + +3D Gaussian Splatting-based open-world free-view synthesis seeks to construct semantic Gaussian radiance fields using a set of 3D Gaussian primitives enriched with color and semantic attributes. This task enables the free novel view synthesis of both realistic appearances and open-vocabulary semantic understanding (also referred to as 3D open-vocabulary semantic segmentation) of the scenes, with significant implications for augmented/virtual reality, robot navigation, and manipulation [16, 18, 26, 28, 30, 58]. Existing methods [35, 50, 59] generally require dense, large sets of input views to train the semantic Gaussian radiance + +field for comprehensive scene coverage. However, when only sparse views are available, these approaches exhibit substantial declines in reconstruction and segmentation accuracy. This is mainly attributed to two key limitations. + +1) Sparsely Distributed Gaussian Points: With sparse input views, the initialized Gaussian points via Structure-from-Motion (SfM) in 3DGS [19] are inherently few and sparsely distributed (Fig. 1 (a1)). Such sparse point distribution imposes challenges in optimizing 3D Gaussians to depict complex indoor environments, leading to suboptimal reconstruction quality and, consequently, impaired segmentation accuracy (Fig. 1 (a2)). 2) Insufficient View Constraints: Optimizing 3D Gaussian parameters with inadequate view supervision often converges to local minima that overfit training views [25, 60]. This results in semantic ambiguity and poor performance in under-observed regions when rendering from novel viewpoints, due to limited reliable semantic integration across different views. + +To address the above challenges, we propose SPC-GS, a novel framework that incorporates Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) regularization. + +To seek denser points for improved Gaussian initialization, SGI produces a structured scene layout Gaussian distribution in two steps. First, SGI employs a video generation model to create view-changed images proximal to training images. These generated images, combined with training views, enrich feature matching in SfM, yielding denser SfM points that promote scene layout Gaussian generation. Second, these denser SfM points undergo view-constraint Gaussian densification and outlier removal to generate structured scene layout points. These points provide an instructive initialization that faithfully captures indoor scene geometry (e.g. furniture, walls, ceilings), enhancing 3DGS training and scene representation, thereby improving segmentation accuracy (Fig. 1 (b1) and (b2)). + +Building on dense scene layout Gaussian initialization, SPC enhances supervision through view-to-view semantic consistency constraints. The core of SPC lies in utilizing informative semantics signals from known training views, serving as semantic prompts, to optimize online-rendered novel (pseudo) views, establishing pseudo-multiview constraints through training-to-pseudo semantic consistency. To achieve it, we utilize the powerful Segment Anything Model 2 (SAM2) [38] to build region mask correspondences across training and pseudo views. Using these correspondences, we leverage the semantics prompts provided by training views to enforce semantic consistency constraints on the corresponding regions of pseudo views at 2D and 3D spaces. This approach substantially augments supervision signals, leading to more coherent segmentation on novel views under sparse input conditions. + +Prior works on 3D open-vocabulary segmentation or + +sparse-input free-view synthesis mainly focuses on object-centric or face-forward scenarios, such as LERF [20], and 3DOVS [31], where objects are centrally positioned, and views are sampled in an "outside-in" manner, resulting in significant view overlap. In contrast, this study focuses on sparse-input open-world free-view synthesis within indoor scenes, particularly relevant for robotics and augmented/virtual reality applications. To this end, we utilize the widely recognized Replica [42] and ScanNet [8] datasets, which feature representative indoor scenes characterized by an "inside-out" viewing configuration. This setup leads to less overlap between certain views, posing a substantial challenge for sparse-input open-world free-view synthesis. The contributions of this work are as follows: + +- We propose SPC-GS, an early and novel attempt at jointly reconstructing and understanding open-world indoor scenes from sparse-input images. +- We introduce Scene layout-based Gaussian Initialization (SGI) to tackle the challenge of sparsely distributed Gaussians, offering a dense and instructive scene layout point initialization that promotes semantic Gaussian learning. +- We present Semantic-Prompt Consistency (SPC) regularization to alleviate inadequate view supervision by enforcing semantic consistency constraints on unseen pseudo views through prompts derived from training views, enhancing coherent segmentation. +- Extensive experiments on sparse-input scenes show the superiority of our approach in free-view synthesis of realistic appearances and open-vocabulary segmentation. + +# 2. Related Work + +Free-view Synthesis via Radiance Fields. In recent years, radiance field-based free-view synthesis has made remarkable advancements. A Notable example is the Neural Radiance Field (NeRF) [13, 32], which leverages a coordinate-based multi-layer perceptron (MLP) to implicitly model 3D scenes, enabling novel view synthesis via volume rendering [10]. Another innovative approach is 3D Gaussian Splating (3DGS) [19, 48], which utilizes a collection of 3D Gaussians to represent the scene and incorporates a fast tile-based rasterizer technique, allowing real-time rendering. + +Although current methods achieve impressive results in dense-view scenarios [4, 12, 15, 27, 29, 33, 54], their performance degrades significantly with sparse inputs [17, 60]. Sparse-input Free-view Synthesis. Recent efforts have tried to address the sparse-input view synthesis and they can broadly categorized into cross-scene pre-training methods and per-scene optimization methods. + +Pre-training methods learn scene priors by pre-training on large-scale datasets spanning multiple scenes. For instance, pixelNeRF [52] conditions a NeRF on CNN features from input images during volume rendering. Similarly, several approaches [3, 45] utilize different NeRF conditions + +to facilitate cross-scene pre-training. Inspired by the success of 3DGS, recent methods such as pixelSplat [2] and MVSplat [6] integrate efficient CUDA-based rasterization with 3D Gaussian representations for pre-training. However, these works demand substantial training resources and large-scale datasets. + +Per-scene optimization methods aim to rapidly reconstruct a 3D scene by introducing effective regularizations. Early NeRF-based methods enhance sparse-input view synthesis through depth-based [34, 44] and frequency range regularization [49]. Recent 3DGS-based works focus on improving rendering efficiency. [7] designs depth regularization loss using monocular depth. DnGaussian [25] and FSGS [60] enhance sparse Gaussian representations through global-local depth normalization and correlation depth distribution losses. CoR-GS [56] adopts color co-regularization loss using rendered pseudo views. + +Instead, we focus on enhancing view synthesis and openworld 3D indoor semantic understanding from sparse inputs following the per-scene optimization route. + +3D Open-world Understanding via Radiance Fields. Recent studies have explored integrating 2D visual foundation models (e.g. DINO [1] and CLIP [24, 37]) with radiance fields to facilitate 3D open-world scene understanding. These approaches can be broadly divided into NeRF-based and 3DGS-based methods. + +NeRF-based. Early methods [11, 43] distill visual features from DINO into semantic radiance fields. More recent language-embedded neural field methods [20, 22] integrate CLIP's visual features into semantic fields, enabling language-driven 3D semantic understanding. Additionally, 3DOVS [31] further improves semantic accuracy through the proposed Relevancy-Distribution Alignment loss. + +3DGS-based. Alternative to NeRFs, recent works explore embedding CLIP features into 3D Gaussians for more efficient semantic rendering. Feature 3DGS [59] directly attaches CLIP embeddings to 3D Gaussians, while FMGS [61] combines Gaussians with multi-resolution hash encodings to form the semantic Gaussian field. [41] embeds quantized CLIP features into Gaussians to reduce storage demand. Furthermore, LangSplat [35] learns low-dimensional semantic Gaussian embeddings and uses a deep neural network for post-process upsampling to obtain semantic representation. Gau-Grouping [50] introduces a 2D object identity loss within training views to optimize Gaussians, while this method requires dense view supervision. Note that the 3D interactive segmentation methods, such as [14, 21, 39, 51], lie outside the scope of language-driven open-world segmentation, as they rely on 2D point prompts to generate object masks but lack semantic meaning. + +While previous approaches rely on dense input images, our work focuses on GS-based 3D open-world scene understanding in sparse-input indoor scenarios. + +# 3. Methodology + +Fig. 2 illustrates our framework, which consists of two primary components. First, Scene layout-based Gaussian Initialization (SGI) (Section 3.2) constructs a dense, instructive scene layout Gaussian initialization that improves Gaussian representation, thus supporting enhanced semantic learning. Then, using the initialized scene layout Gaussians, Semantic-Prompt Consistency (SPC) Regularization (Section 3.3) introduces 2D and 3D semantic consistency constraints, creating effective pseudo-multi-view optimization for semantic learning under sparse-input conditions. + +# 3.1. Preliminary + +Problem Definition. Our approach aims to leverage sparse input images to reconstruct 3D indoor scenes and enable language-driven semantic perception. Concretely, given $L$ posed sparse input images $I = \{I_i\}_{i=1}^L \in \mathbb{R}^{L \times H \times W \times 3}$ and language descriptions, our goal is to generate both realistic appearance and open-world semantic understanding results for specified novel viewpoints. To achieve this, we exploit 3D Gaussians [19] for scene representation, and utilize CLIP models [24, 37] to provide semantic supervision. Preliminary. 3D Gaussian Splitting (3DGS) [19] represents a 3D scene using a collection of explicit 3D Gaussian primitives $\{\mathbf{g}_i = (\mathbf{p}_i, \mathbf{c}_i, o_i, \mathbf{s}_i, \mathbf{q}_i)\}_{i=1}^{N_K}$ , where each primitive comprises a 3D position $\mathbf{p} \in \mathbb{R}^3$ , color $\mathbf{c} \in \mathbb{R}^3$ , opacity $o \in \mathbb{R}$ , scaling factor $\mathbf{s} \in \mathbb{R}^3$ , and rotation quaternion $\mathbf{q} \in \mathbb{R}^4$ . For pixel color $\hat{I}(x_p)$ rendering, 3DGS employs $\alpha$ -blending point-based rendering [23], formulated as: + +$$ +\hat {I} \left(x _ {p}\right) = \sum_ {i \in \mathcal {N}} T _ {i} \alpha_ {i} \mathbf {c} _ {i}, \tag {1} +$$ + +where $\mathcal{N}$ represents Gaussian points overlapping pixel $x_{p}$ , $\alpha_{i} = o_{i}G_{i}^{proj}$ with $G_{i}^{proj}$ as the $i$ -th Gaussian's 2D projection, and transmittance $T_{i} = \prod_{j=1}^{i-1}(1 - \alpha_{j})$ . + +For pixel semantic feature $\hat{\mathbf{F}}(x_p)$ rendering, following [35, 59], we attach a learnable semantic parameter $\mathbf{f}_i \in \mathbb{R}^D$ to each Gaussian and use the above $\alpha$ -blending: + +$$ +\hat {\mathbf {F}} (x _ {p}) = \sum_ {i \in \mathcal {N}} T _ {i} \alpha_ {i} \mathbf {f} _ {i}. \tag {2} +$$ + +Following [31], we use a CLIP text encoder to generate textual feature $\mathbf{T} \in \mathbb{R}^{M \times 512}$ for $M$ classes and obtain rendered segmentation logits $\hat{\mathbf{S}}$ as follows: $\hat{\mathbf{S}} = \cos\langle \omega_f(\hat{\mathbf{F}}), \mathbf{T} \rangle$ , where $\cos$ denotes cosine similarity and $\omega_f$ is a $1 \times 1$ convolution layer for feature dimension alignment. + +We incorporate reconstruction and semantic losses to train the 3D Gaussians. The color reconstruction loss, denoted as $\mathcal{L}_C$ , follows the formulation in 3DGS [19], and the semantic loss, $\mathcal{L}_S$ , is defined as: + +$$ +\mathcal {L} _ {S} = \mathcal {L} _ {c o s} \left(\omega_ {f} (\hat {\mathbf {F}}), \mathbf {F}\right) + \mathcal {L} _ {c e} \left(\omega_ {s} (\hat {\mathbf {S}}), \mathbf {S}\right), \tag {3} +$$ + +![](images/82e6338f8afaf1bbc9a41e24cac2102e5dc9d48ec3d1057f9c9018379d4bf31f.jpg) + +![](images/1a8f73a5dad6e911e0464b24142bb336d2f84f17516b79a679cc0ceaf951ab88.jpg) +Scene layout-based Gaussian Initialization (SGI) +Figure 2. Framework of SPC-GS. (a) We first generate adjacent images of each training image using the video generation model [47]. These generated images, combined with the original training images, produce denser initialized SfM points. These points are then optimized to create a scene layout Gaussian distribution via Gaussian densification and outlier removal. (b) Building on the scene-Layout Gaussian initialization, SPC leverages semantic information from training views as instructive semantic prompts to optimize adjacent rendered pseudo views, establishing semantic consistency constraints that enhance overall sparse-input semantic understanding of 3D scenes. + +where $\mathbf{F}$ and $\mathbf{S}$ are CLIP-derived semantic features and relevancy maps, respectively. $\omega_{s}$ is a $1\times 1$ convolution layer, and $\mathcal{L}_{cos}$ and $\mathcal{L}_{ce}$ denote cosine and cross-entropy losses. + +# 3.2. Scene layout-based Gaussian Initialization + +We propose Scene layout-based Gaussian Initialization (SGI) to produce a dense and conducive point distribution for initialization with the following insights and designs. + +VGM-based Point Creation. In static indoor scenes, viewpoint variations correspond to camera movements. Motivated by this, we employ a video generation model (VGM) to create additional views beyond the sparse training set through image-to-video generation. These generated views augment the available scene information and, when combined with the original training views, enrich feature matching during SfM, yielding denser SfM points for generating scene layout Gaussians. + +Specifically, for each training image $I_{i}$ , we generate eight adjacent views $\{\tilde{I}_i^j\}_{j = 1}^8$ using the image-to-video generation mode of the powerful video generation model MotionCtrl [47]. This is accomplished by inputting the training images along with eight standard camera motion trajectory. + +ries [47] into MotionCtrl. These trajectories include pan up, pan down, pan left, pan right, zoom in, zoom out, clockwise rotation, and anticlockwise rotation. Each trajectory generates a sequence of consecutive video frames. From each sequence, we select a single frame temporally adjacent to the input training image as the valid generated image since it exhibits superior visual quality. These generated images are then combined with the original training images for SfM processing, yielding denser SfM points for Gaussian initialization (see supplementary file). Despite being denser than SfM points derived from sparse-input training images, these points remain limited in capturing the geometric layout and details of indoor scenes. + +Scene-query Point Generation. We further generate scene-query point distributions to enhance 3DGS training and scene representation through the following steps. + +i) View-constraints 3D Gaussian Densification. Using the denser SfM point initialization, we follow Gaussian densification [19] and using our view constraints from Eq. (9) to generate a more densely distributed, instructive scenelayout Gaussian point set $\{\mathbf{g}_i = (\mathbf{p}_i,\mathbf{c}_i,\mathbf{f}_i,o_i,\mathbf{s}_i,\mathbf{q}_i)\}_{i = 1}^{N_{K'}}$ as shown in Fig. 2. These denser $N_{K^{\prime}}$ points provide an + +improved spatial layout of indoor scenes, and enhanced attribute representations (e.g. $\mathbf{c}_i$ , $\mathbf{f}_i$ ), making them effective priors for Gaussian initialization. + +ii) Outlier Gaussian Primitive Removal (OGR). During Gaussian densification, new Gaussian primitives are sampled based on a 3D Gaussian probability distribution, which may generate outlier primitives positioned far from the actual scene layout. These outliers, when used in subsequent initialization, complicate the optimization process. To mitigate this, we present a density-based removal strategy to remove those sparse outlier points. Specifically, Gaussian primitives with fewer than $\mathcal{P}$ neighboring points within a spherical radius $r$ are periodically eliminated during training. Empirical experiments show that setting $\mathcal{P}$ to 5 and $r$ to 1 strikes a balance between effective outlier removal and the preservation of essential scene elements. + +By using instructive scene layout-based points for initialization, the 3DGS training and scene representation can be enhanced, consequently improving segmentation performance under the sparse-input condition. + +# 3.3. Semantic-Prompt Consistency Regularization + +We introduce Semantic-Prompt Consistency (SPC) regularization to exploit additional supervision signals. The key idea of SPC is to leverage semantics from training views as prompts, to impose effective semantic consistency constraints on corresponding regions of pseudo views. + +Region Correspondence Establishment. To establish correspondences between training and pseudo views, we draw inspiration from the powerful zero-shot temporal region proposal extraction capability of Segment Anything Model 2 (SAM2) [38] and leverage it to achieve this. + +Specifically, we introduce Iterative Stochastic Prompting (ISP) to efficiently build correspondence. During training, ISP renders pseudo (unobserved) views using the online training SPC-GS. The view locations are sampled from the two nearest known training views by adding random noise following [60]. Then, ISP treats the current training image and its associated pseudo-views as a viewpoint-changed video sequence, and selects a stochastic point coordinate in the training image, as the 2D point prompt for SAM2, generating corresponding region masks across both the training and pseudo views, which can be formulated as: + +$$ +\mathbf {M} _ {i}, \left\{\mathbf {M} _ {i} ^ {*} \right\} _ {p = 1} ^ {P} = \mathbf {I S P} \left(I _ {i}, \left\{I _ {i} ^ {*} \right\} _ {p = 1} ^ {P}\right), \tag {4} +$$ + +where $I_{i}^{*}$ is the rendered pseudo image and $\mathbf{M}_i^*$ is the corresponding region mask within $I_{i}^{*}$ . $P$ is set to 2 for computational efficiency. As training progress proceeds, the stochastic sampling mechanism encourages a uniform distribution of points across the entire image space, facilitating the efficient region mask correspondence construction for diverse regions of the whole image, as shown in Fig. 3. + +![](images/cbcd7abe309192f64805496f9f041733e07fcf32e0f0b7e5c47fb67286206c9e.jpg) +Figure 3. Illustration of Iterative Stochastic Prompting (ISP). During training, ISP randomly samples point coordinates $(\star)$ from the training view, serving as the point prompt for SAM2 to produce region masks (highlighted in blue) across training and pseudo views. + +![](images/2264dc1266402d2800977b05db6c3d1cfd28299d3310ff5cf9d9dd818e82e2df.jpg) + +Semantic-Prompt for 2D Constrains. Building on the established region mask correspondences, we enforce effective constraints within 2D views to enhance Gaussian learning with sparse inputs. + +i) View-to-View Regularization. Leveraging the semantics from training views as prompts, we enforce training-to-pseudo semantic consistency constraints in feature and relevancy map aspects, which is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {i n t e r} = \sum_ {p = 1} ^ {P} | | \frac {\hat {\mathbf {F}} _ {p} ^ {*} \odot \mathbf {M} _ {p} ^ {*}}{\sum \mathbf {M} _ {p} ^ {*}} - \frac {\hat {\mathbf {F}} \odot \mathbf {M}}{\sum \mathbf {M}} | | _ {2} \\ - \sum_ {p = 1} ^ {P} \mathcal {R} [ \Phi (\hat {\mathbf {S}} \odot \mathbf {M}) ] \log (\hat {\mathbf {S}} _ {p} ^ {*} \odot \mathbf {M} _ {p} ^ {*}), \tag {5} \\ \end{array} +$$ + +where $\hat{\mathbf{F}}^*$ and $\hat{\mathbf{S}}^*$ represent the semantic feature and the relevancy map derived from pseudo views. The function $\Phi$ is the semantic uniformity operation: $\Phi (\hat{\mathbf{S}}\odot \mathbf{M}) =$ $\mathrm{argmax}_c\sum_{i\in \mathbf{M}}\mathbb{1}\{\mathrm{argmax}(\omega_s(\hat{\mathbf{S}}_i)) = c\}$ , which integrates SAM2's masks with a maximum voting mechanism to select the dominant class $c$ within the region mask M, unifying the region's semantics and enforcing consistent region-wise semantic supervision. $\mathcal{R}[\cdot ]$ denotes the dimensional reshaping operation to align the dimension shape of the region mask $\mathbf{M}^*$ from the pseudo view. In $\mathcal{L}_{inter}$ , we utilize self-rendered semantic $(\hat{\mathbf{F}},\hat{\mathbf{S}})$ from training views, generated by the trained SPC-GS, rather than 2D CLIP semantics. This is motivated by the superior coherence and reliability of 3D representations over their 2D counterparts [55] (see supplementary file for experimental validation). + +ii) Contrastive Regularization. Beyond inter-view, we introduce semantic-guided contrastive learning regularization within intra-training-views to further enhance supervision. This is based on the principle that features of semantically consistent objects should be similar. + +Given $N$ feature samples, each sample $\mathbf{z}$ is derived from $\mathbf{z} = W_{\psi}(\hat{\mathbf{F}})$ , where $W_{\psi}$ is a $1\times 1$ convolution transformation layer. Let $i\in U_z\equiv \{1\dots N\}$ be the index of a sample, and $T_{z}(i)\equiv U_{z}\backslash \{i\}$ represents the index set excluding index $i$ . We designate features with the same rendered semantic label $\tilde{\mathbf{q}}$ as positive samples, where $\tilde{\mathbf{q}} =$ + +$\Phi (\hat{\mathbf{S}}\odot \{\mathbf{M}^k\}_{k = 1}^\eta)$ . Here, the $\{\mathbf{M}^k\}_{k = 1}^\eta$ denote $\eta$ masks generated by SAM2's automatic mask generator in the entire image, and $\Phi$ is the function that unifies semantics within each mask. Accordingly, for the $i$ -th sample, positive sample indexes are defined as: $P_{z}(i)\equiv \{p\in T_{z}(i):\tilde{\mathbf{q}}_{p} = \tilde{\mathbf{q}}_{i}\}$ , and negative sample indexes: $T_{z}(i)\backslash P_{z}(i)$ . Similar to SimCLR [5], we employ contrastive learning loss to maximize agreement between positive samples. This method brings $P_{z}$ -indexed positive samples closer in feature space, imposing effective contrastive constraints on sparse-view Gaussian learning, which can be formulated as: + +$$ +\mathcal {L} _ {\text {i n t r a}} = - \sum_ {i \in U _ {z}} \frac {1}{\mathcal {M} _ {z} (i)} \sum_ {p \in P _ {z} (i)} \log \left[ \frac {e ^ {\left(\left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {p}\right) / \tau\right)}}{\sum_ {j \in T _ {z} (i)} e ^ {\left(\left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {j}\right) / \tau\right)}} \right], \tag {6} +$$ + +where $\mathcal{M}_z(i)$ denotes the number of elements in $P_{z}(i)$ . The symbol $\cdot$ refers to the inner product operation. $\tau$ is the temperature parameter, set to 0.2 in the experiment. + +To obtain $N$ samples, we employ a patch-wise sampling strategy by randomly extracting a patch of size $h \times w$ in each training iteration, where $N = h \times w$ . Empirically, $h$ and $w$ are both set to 128. In summary, the semantic prompt for 2D constraints can be formulated as: + +$$ +\mathcal {L} _ {S P C} ^ {2 d} = \mathcal {L} _ {\text {i n t e r}} + \mathcal {L} _ {\text {i n t r a}}. \tag {7} +$$ + +Semantic-Prompt for 3D Constrains. In addition to 2D constraints, we extend semantic-prompt to impose effective 3D constraints on 3D Gaussian primitives. Considering that Gaussian primitives with maximum weight dominate the rendering results (Eq. (2)), we encourage semantic feature similarity between the maximally-weighted 3D Gaussians from two corresponding regions. Specifically, for each pixel within the region mask, we identify the Gaussian primitive with maximum weight along the ray cast into 3D space. Then, the semantic features of these maximally-weighted Gaussians from training views are used to constrain their counterparts in pseudo views. + +However, since Gaussians related to region boundaries often span objects with differing semantics, directly applying the above constraints may complicate optimization. To mitigate this, we introduce a Region Boundary Erosion technique, which removes ambiguous primitives corresponding to region boundaries, which is defined as: + +$$ +\mathcal {L} _ {S P C} ^ {3 d} = \mathcal {L} _ {k l} (\mathcal {F} | | \mathcal {F} ^ {*}) = \sum_ {j = 1} ^ {n ^ {\prime}} \bar {\mathbf {f}} \log \left(\frac {\bar {\mathbf {f}}}{\mathbf {f} _ {j} ^ {*}}\right), \tag {8} +$$ + +where $\mathcal{F}$ denotes the mean feature representation of 3D Gaussian primitives derived from mask $\mathbf{M}'$ in the training view, defined as $\overline{\mathbf{f}} = (\mathbf{f} \odot \mathbf{M}') / \sum \mathbf{M}'$ . Here, $\mathbf{M}'$ refers to the eroded mask after region boundary erosion, which is achieved through morphological erosion. Specifically, a + +pixel $\mathbf{M}'(x,y)$ is retained as "True" only if all pixels within its $5\times 5$ neighborhood in the input original mask $\mathbf{M}$ are "True"; otherwise, it is set to "False". This operation effectively shrinks the boundaries of the region, causing them to contract inward. The set $\mathcal{F}^{*} = \{\mathbf{f}_{1}^{*},\mathbf{f}_{2}^{*},\dots,\mathbf{f}_{n^{\prime}}^{*}\}$ comprises $n^\prime$ Gaussian primitives obtained from eroded masks $\{\mathbf{M}^{*'}\}_{p = 1}^{P}$ in pseudo views. + +To summarize, the overall loss is given by: + +$$ +\mathcal {L} = \mathcal {L} _ {C} + \mathcal {L} _ {S} + \mathcal {L} _ {G S} + \mathcal {L} _ {S} ^ {3 d} + \lambda \left(\mathcal {L} _ {S P C} ^ {2 d} + \mathcal {L} _ {S P C} ^ {3 d}\right), \tag {9} +$$ + +where $\mathcal{L}_{GS}$ is the semantic loss for generated images from video generation model: $\mathcal{L}_{GS} = \mathcal{L}_{cos}(\omega_f(\hat{\mathbf{F}}^g),\mathbf{F}^g) + \mathcal{L}_{ce}(\omega_s(\hat{\mathbf{S}}^g),\mathbf{S}^g)$ . These generated RGB images are not used for color constraints due to their error-prone color pixels that degrade overall results (see supplementary file). $\mathcal{L}_S^{3d}$ is the 3D local adaptive regularization, which promotes semantic similarity among neighboring Gaussians: $\mathcal{L}_S^{3d} = \frac{1}{\theta k}\sum_{i = 1}^{\theta}\sum_{j = 1}^{k}\omega_{ij}(|\mathbf{f}_i - \mathbf{f}_j|_2)$ , where $\theta = 800$ denotes the number of Gaussian randomly sampled per iteration, and $k = 5$ specifies the number of nearest neighbors for each sampled Gaussian in 3D space. $\omega_{ij} = \exp (-||\mu_i - \mu_j||_2)$ is the distance-based weighting coefficient. To prevent early-stage suboptimal convergence, $\lambda$ is initialized at 0 and then increased to 1 after 4k iterations during training. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We conduct experiments on two widely-used indoor scene datasets in terms of novel view synthesis: Replica [42] and ScanNet [8], using a sparse-input setting with only 12 input images for training. Following the standard sparse-input protocol [25, 60], every 10-th frame from each sequence is selected for testing, while 12 evenly sampled sparse views from the remaining images are used for training. These 12 views collectively encompass the entire scene. Notably, the "inside-out" viewing direction and limited overlap among views in sparse-input indoor scenes pose significant challenges for free-view synthesis. + +Evaluation Metrics. We employ PSNR, SSIM [46], and LPIPS [57] for evaluating image quality. We report segmentation performance using the mean Intersection-over-Union (mIoU) and mean pixel accuracy (mAcc). + +Implementation Details. We implement our approach using the official PyTorch 3D Gaussian Splatting [19] on an NVIDIA A100 GPU. The dimension of the learnable semantic parameter attached to 3D Gaussians is set to 32. We modify the CUDA kernel to enable parallel differentiable rasterization of both color and semantic attributes. We employ CLIP ViT-B/16 [37] for feature extraction and SAM2 Hiera-L [38] for region mask generation. More details are provided in the supplementary file. + +![](images/39dc835ba34161f6a6aa3912d2fbdf24f17e813be142c2f37db76f62b60723df.jpg) +Figure 4. Visual reconstruction results on novel views. Our approach achieves superior global structure and photo-realistic details. + +
MethodReplica [42]ScanNet [8]
PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑
3DOVS [31]13.2000.5150.6633.91321.43918.7210.6430.5447.75136.356
Feature 3DGS [59]17.0590.7000.3597.40730.20918.0880.6560.4247.34738.002
LangSplat [35]17.7150.7010.4203.29218.62118.7800.6630.4763.33924.491
Gau-Grouping [50]17.2970.6910.37615.02850.42219.0060.6750.42115.17257.355
DNGaussian [25]18.5130.7120.40216.65754.40018.3920.6910.48413.91254.814
FSGS [60]18.6940.7250.32816.10153.47420.5740.6960.40614.03156.850
CoR-GS [56]18.9550.7580.31215.93352.47620.8150.7210.39314.46255.812
Ours22.0110.7920.25423.96063.26222.4010.7410.36828.94468.823
+ +Table 1. Quantitative results of reconstruction and segmentation on novel views on the Replica and ScanNet datasets, using the CLIP [37] for optimization with 12 training views. Our approach achieves superior performances across all metrics on various datasets. + +# 4.2. Comparison with State-of-the-art Methods + +We conduct novel view synthesis comparisons using sparse input data in indoor scenes, including two types of methods. i) 3D open-vocabulary segmentation methods, including 3DOVS [31], Feature 3DGS [59], LangSplat [35], and Gau-Grouping [50], are designed for 3D open world semantic understanding. ii) Sparse-input free-view synthesis methods contain DNGaussian [25], FSGS [60], and CoR-GS [56], which are designed to learn Gaussian radiance fields from sparse inputs, focusing primarily on appearance modeling. For concurrent 3D open-world segmentation, we extend these methods by adding semantic properties to Gaussian and incorporating CLIP-based semantic supervision, aligning with the semantic settings of our approach. We adopt consistent evaluations across all methods by employing the same sparse-input training protocol and test views. + +Qualitative Results. i) Reconstruction. As shown in Fig. 4, the results from existing 3D open-vocabulary segmentation methods exhibit significant artifacts, since these methods lack sufficient view constraints. Although sparse-view reconstruction methods introduce geometric constraints for regularization, they struggle to recover photo-realistic details. In comparison, our approach incorporates scene-Layout points prior and semantic consistency constraints, leading to a more detailed appearance. + +ii) Open-world Segmentation. In Fig. 5, current methods struggle to obtain precise and complete results, since they easily encounter ambiguity arising from insufficient seman + +tic view supervision. In contrast, our method delivers accurate results, benefiting from the SPC regularization that provides effective consistency constraints. + +Quantitative Results. Table 1 shows that, when optimized using the CLIP model [37], our approach significantly outperforms existing methods across all metrics on both synthetic and real datasets. Competing methods struggle under sparse input conditions, due to either limited Gaussian representations or insufficient semantic consistency constraints. Notably, LangSplat faces considerable challenges, as it freezes Gaussian coordinates, scale, and rotation during semantic parameter optimization, which severely impedes semantic learning under sparse inputs. Besides, our method maintains superior results when optimized with CLIP-LSeg [24] (see supplementary file), validating its generalizability. + +# 4.3. Ablation Studies + +Key Components. In Table 2, applying the VGM-based Point Creation (PC, #1) and Scene layout Point Generation (SPG, #2) notably improves reconstruction quality over setting #0, and consequently boosts segmentation results, validating SGI's effectiveness in enhancing Gaussian representation. Adding the full Semantic-Prompt Consistency regularization (SPC, #4) further yields significant segmentation gains, showing SPC's capability to address insufficient view constraints and enhance 3D scene understanding. Fig. 6 also shows that adding SGI can reduce artifacts, and applying SPC further enhances segmentation results. + +![](images/7291b47ef891b4f3d200342323ca7ad3c9359564c2cbe40ccc8b187ade11adb6.jpg) +Figure 5. Visual open-world segmentation results on novel views. Our approach yields more accurate and complete results. + +
CaseSGISPCReplica [42]
PCSPG2D3DPSNRSSIMLPIPSmIoUmAcc
#017.1340.6780.36914.89951.255
#1✓18.9240.7240.33416.27952.212
#2✓✓21.1760.7830.26019.40158.281
#3✓✓✓21.8140.7910.25722.84961.362
#4✓✓✓✓22.0110.7920.25423.96063.262
+ +Table 2. Ablation study on key components. + +
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aOurs (Full setting)22.0110.7920.25423.96063.262
#bw/o View-to-View Regular.21.6830.7890.25721.24460.914
#cw/o Contrastive Regular.21.8790.7920.25622.32260.086
#dw/o Boundary Erosion21.9500.7920.25422.96461.706
+ +![](images/bff21012490d3246c721469b28a5147bca5e32be66c4f26525755105c02ef230.jpg) +Figure 6. Visual ablation comparison of key components. + +Components in SPC. In Table 3, removing either the 2D view-to-view or contrastive regularization (#b, #c), or omitting the region boundary erosion operation in the 3D constraint (#d), degrades performance. This validates their roles in mitigating insufficient supervision and filtering ambiguous Gaussian primitives spanning different objects. + +Robustness. In Table 4, our strategies also achieve significant gains even with sparse SfM initialization from limited training images, validating their robustness. + +Initialization. From top to bottom in Fig. 7, we can see that leveraging denser and more effective points for Gaussian initialization promotes 3DGS training that enhances scene representation, leading to improved rendered results. + +Table 3. Ablation results of our SPC with different settings. + +
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aSfM Points Initialization17.1340.6780.36914.89951.255
#b#a + SGI20.6800.7740.27018.16956.928
#c#b + SPC21.5890.7860.26020.76661.000
+ +Table 4. Results of our strategies using SfM point initialization. + +![](images/8a259ff223ed1b144b32b4bf5c8e0da33b2c67e8523d1f9a0edc3449fbf31545.jpg) +Figure 7. Visual comparison of different initializations and their rendering results after optimization with these initializations. + +# 5. Conclusion + +This paper presents SPC-GS, a novel approach to address the challenges of GS-based indoor open-world free-view synthesis with sparse views. In SPC-GS, the proposed Scene layout-based Gaussian Initialization (SGI) effectively enhances Gaussian representation by providing an instructive point initialization, consequently promoting semantic learning. Additionally, the Semantic-Prompt Consistency (SPC) regularization mitigates limited supervision by introducing additional view-consistent semantic constraints, improving coherent segmentation performance. Experimental results show that our approach significantly outperforms SOTA methods in reconstruction and 3D open-world segmentation tasks across synthetic and real-world datasets with various CLIP models. + +# References + +[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3 +[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6 +[6] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Proceedings of the European Conference on Computer Vision, pages 370-386. Springer, 2024. 3 +[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 3 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 1, 3, 4 +[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(4): 1, 2017. 2 +[10] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2 +[11] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In Proceedings of the International Conference on Learning Representations, 2022. 3 +[12] Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3d gaussian splatting as new era: A survey. IEEE Transactions on Visualization and Computer Graphics, 2024. 2 + +[13] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 2 +[14] Xu Hu, Yuxi Wang, Lue Fan, Junsong Fan, Junran Peng, Zhen Lei, Qing Li, and Zhaoxiang Zhang. Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857, 2024. 3 +[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[16] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 10608-10615, 2023. 1 +[17] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2 +[18] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Ganesh Iyer, Soroush Saryazdi, Tao Chen, Alaa Maalouf, Shuang Li, Nikhil Varma Keetha, Ayush Tewari, Joshua Tenenbaum, Celso de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. In Proceedings of Robotics: Science and Systems, 2023. 1 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4, 6 +[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 3 +[21] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3 +[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Proceedings of the Advances in Neural Information Processing Systems, pages 23311-23330, 2022. 3 +[23] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3 +[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proceedings of the International Conference on Learning Representations, 2022. 3, 7, 6, 8, 10 +[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2, 3, 6, 7, 9, 10 + +[26] Guibiao Liao, Wei Gao, Qiuping Jiang, Ronggang Wang, and Ge Li. Mmnet: Multi-stage and multi-scale fusion network for rgb-d salient object detection. In Proceedings of the ACM international conference on multimedia, pages 2436-2444, 2020. 1 +[27] Guibiao Liao, Jiankun Li, Zhenyu Bao, Xiaqing Ye, Jingdong Wang, Qing Li, and Kanglin Liu. Clip-gs: Clip-informed gaussian splatting for real-time and view-consistent 3d semantic understanding. arXiv preprint arXiv:2404.14249, 2024. 2 +[28] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3351-3359, 2024. 1 +[29] Guibiao Liao, Kaichen Zhou, Zhenyu Bao, Kanglin Liu, and Qing Li. Ov-nerf: Open-vocabulary neural radiance fields with vision and language foundation models for 3d semantic understanding. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[30] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv preprint arXiv:2403.03174, 2024. 1 +[31] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In Proceedings of the Advances in Neural Information Processing Systems, pages 53433-53456, 2023. 2, 3, 7, 9, 10 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 2 +[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41 (4):1-15, 2022. 2 +[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3 +[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 3, 7, 2, 9 +[36] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 3 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 8, 9 +[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5, 6 +[39] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splattering segmentation solved optimally. In Proceedings of the European Conference on Computer Vision, pages 456-472. Springer, 2024. 3 +[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193-13203, 2024. 3, 4 +[41] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 3 +[42] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8, 1, 3, 4 +[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In Proceedings of the International Conference on 3D Vision, pages 443-453, 2022. 3 +[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3 +[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2 +[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[47] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In Proceedings of the ACM SIGGRAPH, pages 1-11, 2024. 4, 1, 2 +[48] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d + +gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2 +[49] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3 +[50] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In Proceedings of the European Conference on Computer Vision, pages 162-179. Springer, 2024. 1, 3, 7, 9, 10 +[51] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20612-20622, 2024. 3 +[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2 +[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3, 4 +[54] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[55] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In Proceedings of the European Conference on Computer Vision, pages 57-74. Springer, 2024. 5 +[56] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatt-ting via co-regularization. In Proceedings of the European Conference on Computer Vision, pages 335–352. Springer, 2024. 3, 7, 9, 10 +[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[58] Xiaoyu Zhang, Guibiao Liao, Wei Gao, and Ge Li. Tdrnet: Transformer-based dual-branch restoration network for geometry based point cloud compression artifacts. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1–6. IEEE, 2022. 1 +[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 1, 3, 7, 2, 9, 10 + +[60] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In Proceedings of the European Conference on Computer Vision, pages 145-163. Springer, 2024. 2, 3, 5, 6, 7, 9, 10 +[61] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. arXiv preprint arXiv:2401.01970, 2024. 3 \ No newline at end of file diff --git a/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/images.zip b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..26b6eac116b52b721415ed008d60f5a174c80e78 --- /dev/null +++ b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9be8eee548f52e215b80dcd0fd8b2068640c0d62a73cd7d0126ab5ded3b4b43d +size 739764 diff --git a/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/layout.json b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fdf7379d5d8bfaf7978c315001fd3cf0bf42feb1 --- /dev/null +++ b/2025/SPC-GS_ Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs/layout.json @@ -0,0 +1,10091 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "text", + "content": "SPC-GS: Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse Inputs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 168, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 168, + 175 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 168, + 175 + ], + "type": "text", + "content": "Guibiao Liao" + }, + { + "bbox": [ + 91, + 160, + 168, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 161, + 238, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 161, + 238, + 175 + ], + "spans": [ + { + "bbox": [ + 193, + 161, + 238, + 175 + ], + "type": "text", + "content": "Qing Li²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 262, + 161, + 334, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 161, + 334, + 175 + ], + "spans": [ + { + "bbox": [ + 262, + 161, + 334, + 175 + ], + "type": "text", + "content": "Zhenyu Bao" + }, + { + "bbox": [ + 262, + 161, + 334, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 359, + 161, + 429, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 161, + 429, + 175 + ], + "spans": [ + { + "bbox": [ + 359, + 161, + 429, + 175 + ], + "type": "text", + "content": "Guoping Qiu3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 454, + 161, + 523, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 161, + 523, + 175 + ], + "spans": [ + { + "bbox": [ + 454, + 161, + 523, + 175 + ], + "type": "text", + "content": "Kanglin Liu" + }, + { + "bbox": [ + 454, + 161, + 523, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 175, + 470, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 175, + 470, + 190 + ], + "spans": [ + { + "bbox": [ + 141, + 175, + 470, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 175, + 470, + 190 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 174, + 190, + 434, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 190, + 434, + 204 + ], + "spans": [ + { + "bbox": [ + 174, + 190, + 434, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 174, + 190, + 434, + 204 + ], + "type": "text", + "content": "Pengcheng Laboratory, " + }, + { + "bbox": [ + 174, + 190, + 434, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 174, + 190, + 434, + 204 + ], + "type": "text", + "content": "University of Nottingham" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 60, + 236, + 149, + 305 + ], + "blocks": [ + { + "bbox": [ + 60, + 236, + 149, + 305 + ], + "lines": [ + { + "bbox": [ + 60, + 236, + 149, + 305 + ], + "spans": [ + { + "bbox": [ + 60, + 236, + 149, + 305 + ], + "type": "image", + "image_path": "56b1a999e97bfb27559c4b4d5347d830b2f0465d5b5b64c6cc5c4452ebb5dea0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 306, + 100, + 310 + ], + "lines": [ + { + "bbox": [ + 95, + 306, + 100, + 310 + ], + "spans": [ + { + "bbox": [ + 95, + 306, + 100, + 310 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 161, + 236, + 253, + 305 + ], + "blocks": [ + { + "bbox": [ + 161, + 236, + 253, + 305 + ], + "lines": [ + { + "bbox": [ + 161, + 236, + 253, + 305 + ], + "spans": [ + { + "bbox": [ + 161, + 236, + 253, + 305 + ], + "type": "image", + "image_path": "79e834178e20de645a65d1a7d454339e0542438ce41514e605ee1c9035cb6ed3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 306, + 254, + 314 + ], + "lines": [ + { + "bbox": [ + 160, + 306, + 254, + 314 + ], + "spans": [ + { + "bbox": [ + 160, + 306, + 254, + 314 + ], + "type": "text", + "content": "(a1) SfM Points (Gau-Grouping)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 258, + 236, + 348, + 305 + ], + "blocks": [ + { + "bbox": [ + 258, + 236, + 348, + 305 + ], + "lines": [ + { + "bbox": [ + 258, + 236, + 348, + 305 + ], + "spans": [ + { + "bbox": [ + 258, + 236, + 348, + 305 + ], + "type": "image", + "image_path": "7ace5f62f9e314a28d2308fb2950cac66f2b24f8e4a89212f00772bb1862c549.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 306, + 426, + 314 + ], + "lines": [ + { + "bbox": [ + 269, + 306, + 426, + 314 + ], + "spans": [ + { + "bbox": [ + 269, + 306, + 426, + 314 + ], + "type": "text", + "content": "(a2) Rendered RGB/Semantic Results (Gau-Grouping)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 449, + 239, + 461, + 250 + ], + "blocks": [ + { + "bbox": [ + 449, + 239, + 461, + 250 + ], + "lines": [ + { + "bbox": [ + 449, + 239, + 461, + 250 + ], + "spans": [ + { + "bbox": [ + 449, + 239, + 461, + 250 + ], + "type": "image", + "image_path": "6134cebc0dc6fc47e4785696a7f028feec51412a6e498a37acfa61c3f457a023.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 240, + 507, + 249 + ], + "lines": [ + { + "bbox": [ + 465, + 240, + 507, + 249 + ], + "spans": [ + { + "bbox": [ + 465, + 240, + 507, + 249 + ], + "type": "text", + "content": "Gau-Grouping" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 511, + 240, + 523, + 250 + ], + "blocks": [ + { + "bbox": [ + 511, + 240, + 523, + 250 + ], + "lines": [ + { + "bbox": [ + 511, + 240, + 523, + 250 + ], + "spans": [ + { + "bbox": [ + 511, + 240, + 523, + 250 + ], + "type": "image", + "image_path": "6b9018d89f4da445abb28aab4c6909a9c87a5aa03350d2f9991a98ca2fee36f5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 526, + 240, + 541, + 249 + ], + "lines": [ + { + "bbox": [ + 526, + 240, + 541, + 249 + ], + "spans": [ + { + "bbox": [ + 526, + 240, + 541, + 249 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 60, + 317, + 149, + 386 + ], + "blocks": [ + { + "bbox": [ + 60, + 317, + 149, + 386 + ], + "lines": [ + { + "bbox": [ + 60, + 317, + 149, + 386 + ], + "spans": [ + { + "bbox": [ + 60, + 317, + 149, + 386 + ], + "type": "image", + "image_path": "89fede60bb0b3e956fa0a0b57b0e4d5f4c5e37f5ff2fbf2cd7b49645457a43d8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 387, + 134, + 396 + ], + "lines": [ + { + "bbox": [ + 75, + 387, + 134, + 396 + ], + "spans": [ + { + "bbox": [ + 75, + 387, + 134, + 396 + ], + "type": "text", + "content": "Sparse Input Images" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 162, + 316, + 253, + 386 + ], + "blocks": [ + { + "bbox": [ + 162, + 316, + 253, + 386 + ], + "lines": [ + { + "bbox": [ + 162, + 316, + 253, + 386 + ], + "spans": [ + { + "bbox": [ + 162, + 316, + 253, + 386 + ], + "type": "image", + "image_path": "94781e75ea003cab1e969ead9a69c792df92c5ea0b55d2442263f30ef94fdd45.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 387, + 253, + 396 + ], + "lines": [ + { + "bbox": [ + 162, + 387, + 253, + 396 + ], + "spans": [ + { + "bbox": [ + 162, + 387, + 253, + 396 + ], + "type": "text", + "content": "(b1) Scene layout Points (Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 258, + 316, + 348, + 386 + ], + "blocks": [ + { + "bbox": [ + 258, + 316, + 348, + 386 + ], + "lines": [ + { + "bbox": [ + 258, + 316, + 348, + 386 + ], + "spans": [ + { + "bbox": [ + 258, + 316, + 348, + 386 + ], + "type": "image", + "image_path": "a7f9783f58807b6d66b40ca97ea016f90c9ec11593b055df52772fdbd8f239af.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 555, + 446 + ], + "type": "text", + "content": "Figure 1. Visual comparisons of open-world free-view synthesis from sparse inputs (12 training views). The previous 3DGS-based method, Gau-Grouping [50], utilizes sparse SfM points for Gaussian initialization and limited training views for supervision, leading to inferior rendering results. In contrast, our approach leverages scene layout-based points for instructive initialization, and cooperates with semantic-prompt consistency constraints (detailed in Sec. 3.3), yielding superior reconstruction and segmentation results." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 348, + 316, + 438, + 386 + ], + "blocks": [ + { + "bbox": [ + 283, + 387, + 413, + 396 + ], + "lines": [ + { + "bbox": [ + 283, + 387, + 413, + 396 + ], + "spans": [ + { + "bbox": [ + 283, + 387, + 413, + 396 + ], + "type": "text", + "content": "(b2) Rendered RGB/Semantic Results (Ours)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 348, + 316, + 438, + 386 + ], + "lines": [ + { + "bbox": [ + 348, + 316, + 438, + 386 + ], + "spans": [ + { + "bbox": [ + 348, + 316, + 438, + 386 + ], + "type": "image", + "image_path": "7e66b74017f05f01a248383a1395bf1ca8519b6a31cca494c2f1cd2917b56840.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 449, + 262, + 476, + 382 + ], + "blocks": [ + { + "bbox": [ + 449, + 262, + 476, + 382 + ], + "lines": [ + { + "bbox": [ + 449, + 262, + 476, + 382 + ], + "spans": [ + { + "bbox": [ + 449, + 262, + 476, + 382 + ], + "type": "image", + "image_path": "aa5f7a25599a8d4cc179e40a1774f231f750d9be53a16c0ee09ec72529d19c00.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 386, + 533, + 394 + ], + "lines": [ + { + "bbox": [ + 465, + 386, + 533, + 394 + ], + "spans": [ + { + "bbox": [ + 465, + 386, + 533, + 394 + ], + "type": "text", + "content": "(c) Comparison Results" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 482, + 262, + 510, + 382 + ], + "blocks": [ + { + "bbox": [ + 482, + 262, + 510, + 382 + ], + "lines": [ + { + "bbox": [ + 482, + 262, + 510, + 382 + ], + "spans": [ + { + "bbox": [ + 482, + 262, + 510, + 382 + ], + "type": "image", + "image_path": "eda3cd88212d2e53152452c01ce62cd50438a8dd1ea11730914c7adda18ebc6d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 517, + 262, + 545, + 382 + ], + "blocks": [ + { + "bbox": [ + 517, + 262, + 545, + 382 + ], + "lines": [ + { + "bbox": [ + 517, + 262, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 517, + 262, + 545, + 382 + ], + "type": "image", + "image_path": "9011ed3c69227ae8905b608d334c5339e8d43626e25997ce1733eb2a3ef8823d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 151, + 456, + 200, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 456, + 200, + 468 + ], + "spans": [ + { + "bbox": [ + 151, + 456, + 200, + 468 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 54, + 480, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 480, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 297, + 696 + ], + "type": "text", + "content": "3D Gaussian Splatting-based indoor open-world free-view synthesis approaches have shown significant performance with dense input images. However, they exhibit poor performance when confronted with sparse inputs, primarily due to the sparse distribution of Gaussian points and insufficient view supervision. To relieve these challenges, we propose SPC-GS, leveraging Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) Regularization for open-world free view synthesis with sparse inputs. Specifically, SGI provides a dense, scene-Layout-based Gaussian distribution by utilizing view-changed images generated from the video generation model and view-constraint Gaussian points densification. Additionally, SPC mitigates limited view supervision by employing semantic-prompt-based consistency constraints developed by SAM2. This approach leverages available semantics from training views, serving as instructive prompts, to optimize visually overlapping regions in novel views with" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 456, + 556, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 541 + ], + "type": "text", + "content": "2D and 3D consistency constraints. Extensive experiments demonstrate the superior performance of SPC-GS across Replica and ScanNet benchmarks. Notably, our SPC-GS achieves a 3.06 dB gain in PSNR for reconstruction quality and a " + }, + { + "bbox": [ + 313, + 456, + 556, + 541 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 313, + 456, + 556, + 541 + ], + "type": "text", + "content": " improvement in mIoU for open-world semantic segmentation. Project website at: https://gbliao.github.io/SPC-GS.github.io." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 395, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 581, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 556, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting-based open-world free-view synthesis seeks to construct semantic Gaussian radiance fields using a set of 3D Gaussian primitives enriched with color and semantic attributes. This task enables the free novel view synthesis of both realistic appearances and open-vocabulary semantic understanding (also referred to as 3D open-vocabulary semantic segmentation) of the scenes, with significant implications for augmented/virtual reality, robot navigation, and manipulation [16, 18, 26, 28, 30, 58]. Existing methods [35, 50, 59] generally require dense, large sets of input views to train the semantic Gaussian radiance" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11264" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "field for comprehensive scene coverage. However, when only sparse views are available, these approaches exhibit substantial declines in reconstruction and segmentation accuracy. This is mainly attributed to two key limitations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 289 + ], + "type": "text", + "content": "1) Sparsely Distributed Gaussian Points: With sparse input views, the initialized Gaussian points via Structure-from-Motion (SfM) in 3DGS [19] are inherently few and sparsely distributed (Fig. 1 (a1)). Such sparse point distribution imposes challenges in optimizing 3D Gaussians to depict complex indoor environments, leading to suboptimal reconstruction quality and, consequently, impaired segmentation accuracy (Fig. 1 (a2)). 2) Insufficient View Constraints: Optimizing 3D Gaussian parameters with inadequate view supervision often converges to local minima that overfit training views [25, 60]. This results in semantic ambiguity and poor performance in under-observed regions when rendering from novel viewpoints, due to limited reliable semantic integration across different views." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 338 + ], + "type": "text", + "content": "To address the above challenges, we propose SPC-GS, a novel framework that incorporates Scene layout-based Gaussian Initialization (SGI) and Semantic-Prompt Consistency (SPC) regularization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 507 + ], + "type": "text", + "content": "To seek denser points for improved Gaussian initialization, SGI produces a structured scene layout Gaussian distribution in two steps. First, SGI employs a video generation model to create view-changed images proximal to training images. These generated images, combined with training views, enrich feature matching in SfM, yielding denser SfM points that promote scene layout Gaussian generation. Second, these denser SfM points undergo view-constraint Gaussian densification and outlier removal to generate structured scene layout points. These points provide an instructive initialization that faithfully captures indoor scene geometry (e.g. furniture, walls, ceilings), enhancing 3DGS training and scene representation, thereby improving segmentation accuracy (Fig. 1 (b1) and (b2))." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 700 + ], + "type": "text", + "content": "Building on dense scene layout Gaussian initialization, SPC enhances supervision through view-to-view semantic consistency constraints. The core of SPC lies in utilizing informative semantics signals from known training views, serving as semantic prompts, to optimize online-rendered novel (pseudo) views, establishing pseudo-multiview constraints through training-to-pseudo semantic consistency. To achieve it, we utilize the powerful Segment Anything Model 2 (SAM2) [38] to build region mask correspondences across training and pseudo views. Using these correspondences, we leverage the semantics prompts provided by training views to enforce semantic consistency constraints on the corresponding regions of pseudo views at 2D and 3D spaces. This approach substantially augments supervision signals, leading to more coherent segmentation on novel views under sparse input conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Prior works on 3D open-vocabulary segmentation or" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "sparse-input free-view synthesis mainly focuses on object-centric or face-forward scenarios, such as LERF [20], and 3DOVS [31], where objects are centrally positioned, and views are sampled in an \"outside-in\" manner, resulting in significant view overlap. In contrast, this study focuses on sparse-input open-world free-view synthesis within indoor scenes, particularly relevant for robotics and augmented/virtual reality applications. To this end, we utilize the widely recognized Replica [42] and ScanNet [8] datasets, which feature representative indoor scenes characterized by an \"inside-out\" viewing configuration. This setup leads to less overlap between certain views, posing a substantial challenge for sparse-input open-world free-view synthesis. The contributions of this work are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 241, + 553, + 421 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 241, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 241, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 241, + 553, + 277 + ], + "type": "text", + "content": "- We propose SPC-GS, an early and novel attempt at jointly reconstructing and understanding open-world indoor scenes from sparse-input images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 277, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 553, + 324 + ], + "type": "text", + "content": "- We introduce Scene layout-based Gaussian Initialization (SGI) to tackle the challenge of sparsely distributed Gaussians, offering a dense and instructive scene layout point initialization that promotes semantic Gaussian learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 325, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 325, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 325, + 553, + 384 + ], + "type": "text", + "content": "- We present Semantic-Prompt Consistency (SPC) regularization to alleviate inadequate view supervision by enforcing semantic consistency constraints on unseen pseudo views through prompts derived from training views, enhancing coherent segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 384, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 553, + 421 + ], + "type": "text", + "content": "- Extensive experiments on sparse-input scenes show the superiority of our approach in free-view synthesis of realistic appearances and open-vocabulary segmentation." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 430, + 400, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 400, + 442 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 400, + 442 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 569 + ], + "type": "text", + "content": "Free-view Synthesis via Radiance Fields. In recent years, radiance field-based free-view synthesis has made remarkable advancements. A Notable example is the Neural Radiance Field (NeRF) [13, 32], which leverages a coordinate-based multi-layer perceptron (MLP) to implicitly model 3D scenes, enabling novel view synthesis via volume rendering [10]. Another innovative approach is 3D Gaussian Splating (3DGS) [19, 48], which utilizes a collection of 3D Gaussians to represent the scene and incorporates a fast tile-based rasterizer technique, allowing real-time rendering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "text", + "content": "Although current methods achieve impressive results in dense-view scenarios [4, 12, 15, 27, 29, 33, 54], their performance degrades significantly with sparse inputs [17, 60]. Sparse-input Free-view Synthesis. Recent efforts have tried to address the sparse-input view synthesis and they can broadly categorized into cross-scene pre-training methods and per-scene optimization methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Pre-training methods learn scene priors by pre-training on large-scale datasets spanning multiple scenes. For instance, pixelNeRF [52] conditions a NeRF on CNN features from input images during volume rendering. Similarly, several approaches [3, 45] utilize different NeRF conditions" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11265" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "to facilitate cross-scene pre-training. Inspired by the success of 3DGS, recent methods such as pixelSplat [2] and MVSplat [6] integrate efficient CUDA-based rasterization with 3D Gaussian representations for pre-training. However, these works demand substantial training resources and large-scale datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 277 + ], + "type": "text", + "content": "Per-scene optimization methods aim to rapidly reconstruct a 3D scene by introducing effective regularizations. Early NeRF-based methods enhance sparse-input view synthesis through depth-based [34, 44] and frequency range regularization [49]. Recent 3DGS-based works focus on improving rendering efficiency. [7] designs depth regularization loss using monocular depth. DnGaussian [25] and FSGS [60] enhance sparse Gaussian representations through global-local depth normalization and correlation depth distribution losses. CoR-GS [56] adopts color co-regularization loss using rendered pseudo views." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 314 + ], + "type": "text", + "content": "Instead, we focus on enhancing view synthesis and openworld 3D indoor semantic understanding from sparse inputs following the per-scene optimization route." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 386 + ], + "type": "text", + "content": "3D Open-world Understanding via Radiance Fields. Recent studies have explored integrating 2D visual foundation models (e.g. DINO [1] and CLIP [24, 37]) with radiance fields to facilitate 3D open-world scene understanding. These approaches can be broadly divided into NeRF-based and 3DGS-based methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 472 + ], + "type": "text", + "content": "NeRF-based. Early methods [11, 43] distill visual features from DINO into semantic radiance fields. More recent language-embedded neural field methods [20, 22] integrate CLIP's visual features into semantic fields, enabling language-driven 3D semantic understanding. Additionally, 3DOVS [31] further improves semantic accuracy through the proposed Relevancy-Distribution Alignment loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 676 + ], + "type": "text", + "content": "3DGS-based. Alternative to NeRFs, recent works explore embedding CLIP features into 3D Gaussians for more efficient semantic rendering. Feature 3DGS [59] directly attaches CLIP embeddings to 3D Gaussians, while FMGS [61] combines Gaussians with multi-resolution hash encodings to form the semantic Gaussian field. [41] embeds quantized CLIP features into Gaussians to reduce storage demand. Furthermore, LangSplat [35] learns low-dimensional semantic Gaussian embeddings and uses a deep neural network for post-process upsampling to obtain semantic representation. Gau-Grouping [50] introduces a 2D object identity loss within training views to optimize Gaussians, while this method requires dense view supervision. Note that the 3D interactive segmentation methods, such as [14, 21, 39, 51], lie outside the scope of language-driven open-world segmentation, as they rely on 2D point prompts to generate object masks but lack semantic meaning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "While previous approaches rely on dense input images, our work focuses on GS-based 3D open-world scene understanding in sparse-input indoor scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": "Fig. 2 illustrates our framework, which consists of two primary components. First, Scene layout-based Gaussian Initialization (SGI) (Section 3.2) constructs a dense, instructive scene layout Gaussian initialization that improves Gaussian representation, thus supporting enhanced semantic learning. Then, using the initialized scene layout Gaussians, Semantic-Prompt Consistency (SPC) Regularization (Section 3.3) introduces 2D and 3D semantic consistency constraints, creating effective pseudo-multi-view optimization for semantic learning under sparse-input conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 217, + 394, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 394, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 394, + 230 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": "Problem Definition. Our approach aims to leverage sparse input images to reconstruct 3D indoor scenes and enable language-driven semantic perception. Concretely, given " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": " posed sparse input images " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "I = \\{I_i\\}_{i=1}^L \\in \\mathbb{R}^{L \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": " and language descriptions, our goal is to generate both realistic appearance and open-world semantic understanding results for specified novel viewpoints. To achieve this, we exploit 3D Gaussians [19] for scene representation, and utilize CLIP models [24, 37] to provide semantic supervision. Preliminary. 3D Gaussian Splitting (3DGS) [19] represents a 3D scene using a collection of explicit 3D Gaussian primitives " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{g}_i = (\\mathbf{p}_i, \\mathbf{c}_i, o_i, \\mathbf{s}_i, \\mathbf{q}_i)\\}_{i=1}^{N_K}" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": ", where each primitive comprises a 3D position " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": ", scaling factor " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": ", and rotation quaternion " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": ". For pixel color " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\hat{I}(x_p)" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": " rendering, 3DGS employs " + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 235, + 555, + 427 + ], + "type": "text", + "content": "-blending point-based rendering [23], formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 390, + 434, + 554, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 434, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 390, + 434, + 554, + 460 + ], + "type": "interline_equation", + "content": "\\hat {I} \\left(x _ {p}\\right) = \\sum_ {i \\in \\mathcal {N}} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\tag {1}", + "image_path": "9a2c743d2f79d788b23cca793efbb860d2d4781fd6c2b1421a9d05e385ecb506.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": " represents Gaussian points overlapping pixel " + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "inline_equation", + "content": "x_{p}" + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = o_{i}G_{i}^{proj}" + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "inline_equation", + "content": "G_{i}^{proj}" + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": "-th Gaussian's 2D projection, and transmittance " + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "inline_equation", + "content": "T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 313, + 467, + 554, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "text", + "content": "For pixel semantic feature " + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}(x_p)" + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "text", + "content": " rendering, following [35, 59], we attach a learnable semantic parameter " + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "text", + "content": " to each Gaussian and use the above " + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 508, + 554, + 544 + ], + "type": "text", + "content": "-blending:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 389, + 552, + 554, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 552, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 389, + 552, + 554, + 578 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} (x _ {p}) = \\sum_ {i \\in \\mathcal {N}} T _ {i} \\alpha_ {i} \\mathbf {f} _ {i}. \\tag {2}", + "image_path": "d0f264165660c0dfb98ba3479f40dde2c9205ee8d7b25429ce4e53285930bea1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": "Following [31], we use a CLIP text encoder to generate textual feature " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{M \\times 512}" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": " classes and obtain rendered segmentation logits " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": " as follows: " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}} = \\cos\\langle \\omega_f(\\hat{\\mathbf{F}}), \\mathbf{T} \\rangle" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": " denotes cosine similarity and " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\omega_f" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": " convolution layer for feature dimension alignment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": "We incorporate reconstruction and semantic losses to train the 3D Gaussians. The color reconstruction loss, denoted as " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": ", follows the formulation in 3DGS [19], and the semantic loss, " + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 313, + 645, + 554, + 693 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 700, + 554, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 700, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 350, + 700, + 554, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\mathcal {L} _ {c o s} \\left(\\omega_ {f} (\\hat {\\mathbf {F}}), \\mathbf {F}\\right) + \\mathcal {L} _ {c e} \\left(\\omega_ {s} (\\hat {\\mathbf {S}}), \\mathbf {S}\\right), \\tag {3}", + "image_path": "366d5844a2f6ec733c554e55eef0ba951c0f53754ded00764e9c032da2c02f45.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11266" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 85, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 85, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 85, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 85, + 553, + 198 + ], + "type": "image", + "image_path": "82e6338f8afaf1bbc9a41e24cac2102e5dc9d48ec3d1057f9c9018379d4bf31f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 211, + 553, + 367 + ], + "blocks": [ + { + "bbox": [ + 195, + 200, + 416, + 211 + ], + "lines": [ + { + "bbox": [ + 195, + 200, + 416, + 211 + ], + "spans": [ + { + "bbox": [ + 195, + 200, + 416, + 211 + ], + "type": "text", + "content": "Scene layout-based Gaussian Initialization (SGI)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 211, + 553, + 367 + ], + "lines": [ + { + "bbox": [ + 57, + 211, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 57, + 211, + 553, + 367 + ], + "type": "image", + "image_path": "1a8f73a5dad6e911e0464b24142bb336d2f84f17516b79a679cc0ceaf951ab88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 429 + ], + "type": "text", + "content": "Figure 2. Framework of SPC-GS. (a) We first generate adjacent images of each training image using the video generation model [47]. These generated images, combined with the original training images, produce denser initialized SfM points. These points are then optimized to create a scene layout Gaussian distribution via Gaussian densification and outlier removal. (b) Building on the scene-Layout Gaussian initialization, SPC leverages semantic information from training views as instructive semantic prompts to optimize adjacent rendered pseudo views, establishing semantic consistency constraints that enhance overall sparse-input semantic understanding of 3D scenes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " are CLIP-derived semantic features and relevancy maps, respectively. " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\omega_{s}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " convolution layer, and " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cos}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " denote cosine and cross-entropy losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 492, + 277, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 277, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 277, + 504 + ], + "type": "text", + "content": "3.2. Scene layout-based Gaussian Initialization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": "We propose Scene layout-based Gaussian Initialization (SGI) to produce a dense and conducive point distribution for initialization with the following insights and designs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 546, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 653 + ], + "type": "text", + "content": "VGM-based Point Creation. In static indoor scenes, viewpoint variations correspond to camera movements. Motivated by this, we employ a video generation model (VGM) to create additional views beyond the sparse training set through image-to-video generation. These generated views augment the available scene information and, when combined with the original training views, enrich feature matching during SfM, yielding denser SfM points for generating scene layout Gaussians." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Specifically, for each training image " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", we generate eight adjacent views " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_i^j\\}_{j = 1}^8" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " using the image-to-video generation mode of the powerful video generation model MotionCtrl [47]. This is accomplished by inputting the training images along with eight standard camera motion trajectory." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 604 + ], + "type": "text", + "content": "ries [47] into MotionCtrl. These trajectories include pan up, pan down, pan left, pan right, zoom in, zoom out, clockwise rotation, and anticlockwise rotation. Each trajectory generates a sequence of consecutive video frames. From each sequence, we select a single frame temporally adjacent to the input training image as the valid generated image since it exhibits superior visual quality. These generated images are then combined with the original training images for SfM processing, yielding denser SfM points for Gaussian initialization (see supplementary file). Despite being denser than SfM points derived from sparse-input training images, these points remain limited in capturing the geometric layout and details of indoor scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": "Scene-query Point Generation. We further generate scene-query point distributions to enhance 3DGS training and scene representation through the following steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "i) View-constraints 3D Gaussian Densification. Using the denser SfM point initialization, we follow Gaussian densification [19] and using our view constraints from Eq. (9) to generate a more densely distributed, instructive scenelayout Gaussian point set " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{g}_i = (\\mathbf{p}_i,\\mathbf{c}_i,\\mathbf{f}_i,o_i,\\mathbf{s}_i,\\mathbf{q}_i)\\}_{i = 1}^{N_{K'}}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " as shown in Fig. 2. These denser " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "N_{K^{\\prime}}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " points provide an" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 141, + 72, + 309, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 72, + 309, + 82 + ], + "spans": [ + { + "bbox": [ + 141, + 72, + 309, + 82 + ], + "type": "text", + "content": "Video Generation Model based Point Creation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 413, + 72, + 526, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 72, + 526, + 82 + ], + "spans": [ + { + "bbox": [ + 413, + 72, + 526, + 82 + ], + "type": "text", + "content": "Scene-Layout Point Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11267" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": "improved spatial layout of indoor scenes, and enhanced attribute representations (e.g. " + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 107 + ], + "type": "text", + "content": "), making them effective priors for Gaussian initialization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "text", + "content": "ii) Outlier Gaussian Primitive Removal (OGR). During Gaussian densification, new Gaussian primitives are sampled based on a 3D Gaussian probability distribution, which may generate outlier primitives positioned far from the actual scene layout. These outliers, when used in subsequent initialization, complicate the optimization process. To mitigate this, we present a density-based removal strategy to remove those sparse outlier points. Specifically, Gaussian primitives with fewer than " + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "text", + "content": " neighboring points within a spherical radius " + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "text", + "content": " are periodically eliminated during training. Empirical experiments show that setting " + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "text", + "content": " to 5 and " + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 109, + 295, + 263 + ], + "type": "text", + "content": " to 1 strikes a balance between effective outlier removal and the preservation of essential scene elements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 313 + ], + "type": "text", + "content": "By using instructive scene layout-based points for initialization, the 3DGS training and scene representation can be enhanced, consequently improving segmentation performance under the sparse-input condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 293, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 293, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 293, + 336 + ], + "type": "text", + "content": "3.3. Semantic-Prompt Consistency Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 401 + ], + "type": "text", + "content": "We introduce Semantic-Prompt Consistency (SPC) regularization to exploit additional supervision signals. The key idea of SPC is to leverage semantics from training views as prompts, to impose effective semantic consistency constraints on corresponding regions of pseudo views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "content": "Region Correspondence Establishment. To establish correspondences between training and pseudo views, we draw inspiration from the powerful zero-shot temporal region proposal extraction capability of Segment Anything Model 2 (SAM2) [38] and leverage it to achieve this." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 594 + ], + "type": "text", + "content": "Specifically, we introduce Iterative Stochastic Prompting (ISP) to efficiently build correspondence. During training, ISP renders pseudo (unobserved) views using the online training SPC-GS. The view locations are sampled from the two nearest known training views by adding random noise following [60]. Then, ISP treats the current training image and its associated pseudo-views as a viewpoint-changed video sequence, and selects a stochastic point coordinate in the training image, as the 2D point prompt for SAM2, generating corresponding region masks across both the training and pseudo views, which can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 604, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 604, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 99, + 604, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i}, \\left\\{\\mathbf {M} _ {i} ^ {*} \\right\\} _ {p = 1} ^ {P} = \\mathbf {I S P} \\left(I _ {i}, \\left\\{I _ {i} ^ {*} \\right\\} _ {p = 1} ^ {P}\\right), \\tag {4}", + "image_path": "42beddf221f366a0b0050478251bda5271c49770d2413c806b696d96f5688184.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{i}^{*}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " is the rendered pseudo image and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^*" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " is the corresponding region mask within " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{i}^{*}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " is set to 2 for computational efficiency. As training progress proceeds, the stochastic sampling mechanism encourages a uniform distribution of points across the entire image space, facilitating the efficient region mask correspondence construction for diverse regions of the whole image, as shown in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 432, + 146 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 432, + 146 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 432, + 146 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 432, + 146 + ], + "type": "image", + "image_path": "cbcd7abe309192f64805496f9f041733e07fcf32e0f0b7e5c47fb67286206c9e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 148, + 554, + 193 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 554, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 554, + 193 + ], + "type": "text", + "content": "Figure 3. Illustration of Iterative Stochastic Prompting (ISP). During training, ISP randomly samples point coordinates " + }, + { + "bbox": [ + 313, + 148, + 554, + 193 + ], + "type": "inline_equation", + "content": "(\\star)" + }, + { + "bbox": [ + 313, + 148, + 554, + 193 + ], + "type": "text", + "content": " from the training view, serving as the point prompt for SAM2 to produce region masks (highlighted in blue) across training and pseudo views." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 551, + 147 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 551, + 147 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 551, + 147 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 551, + 147 + ], + "type": "image", + "image_path": "2264dc1266402d2800977b05db6c3d1cfd28299d3310ff5cf9d9dd818e82e2df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": "Semantic-Prompt for 2D Constrains. Building on the established region mask correspondences, we enforce effective constraints within 2D views to enhance Gaussian learning with sparse inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 262, + 554, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 310 + ], + "type": "text", + "content": "i) View-to-View Regularization. Leveraging the semantics from training views as prompts, we enforce training-to-pseudo semantic consistency constraints in feature and relevancy map aspects, which is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 319, + 554, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 319, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 339, + 319, + 554, + 389 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {i n t e r} = \\sum_ {p = 1} ^ {P} | | \\frac {\\hat {\\mathbf {F}} _ {p} ^ {*} \\odot \\mathbf {M} _ {p} ^ {*}}{\\sum \\mathbf {M} _ {p} ^ {*}} - \\frac {\\hat {\\mathbf {F}} \\odot \\mathbf {M}}{\\sum \\mathbf {M}} | | _ {2} \\\\ - \\sum_ {p = 1} ^ {P} \\mathcal {R} [ \\Phi (\\hat {\\mathbf {S}} \\odot \\mathbf {M}) ] \\log (\\hat {\\mathbf {S}} _ {p} ^ {*} \\odot \\mathbf {M} _ {p} ^ {*}), \\tag {5} \\\\ \\end{array}", + "image_path": "ad5964463f8847e9946d72a959b772c115f6ab73a07c8857864f9bc7cb6d58c4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^*" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}^*" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " represent the semantic feature and the relevancy map derived from pseudo views. The function " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " is the semantic uniformity operation: " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\Phi (\\hat{\\mathbf{S}}\\odot \\mathbf{M}) =" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathrm{argmax}_c\\sum_{i\\in \\mathbf{M}}\\mathbb{1}\\{\\mathrm{argmax}(\\omega_s(\\hat{\\mathbf{S}}_i)) = c\\}" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": ", which integrates SAM2's masks with a maximum voting mechanism to select the dominant class " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " within the region mask M, unifying the region's semantics and enforcing consistent region-wise semantic supervision. " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{R}[\\cdot ]" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " denotes the dimensional reshaping operation to align the dimension shape of the region mask " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^*" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " from the pseudo view. In " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inter}" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": ", we utilize self-rendered semantic " + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{F}},\\hat{\\mathbf{S}})" + }, + { + "bbox": [ + 313, + 399, + 554, + 582 + ], + "type": "text", + "content": " from training views, generated by the trained SPC-GS, rather than 2D CLIP semantics. This is motivated by the superior coherence and reliability of 3D representations over their 2D counterparts [55] (see supplementary file for experimental validation)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 640 + ], + "type": "text", + "content": "ii) Contrastive Regularization. Beyond inter-view, we introduce semantic-guided contrastive learning regularization within intra-training-views to further enhance supervision. This is based on the principle that features of semantically consistent objects should be similar." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " feature samples, each sample " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = W_{\\psi}(\\hat{\\mathbf{F}})" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "W_{\\psi}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " convolution transformation layer. Let " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "i\\in U_z\\equiv \\{1\\dots N\\}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " be the index of a sample, and " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "T_{z}(i)\\equiv U_{z}\\backslash \\{i\\}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " represents the index set excluding index " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ". We designate features with the same rendered semantic label " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " as positive samples, where " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}} =" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11268" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\Phi (\\hat{\\mathbf{S}}\\odot \\{\\mathbf{M}^k\\}_{k = 1}^\\eta)" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": ". Here, the " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}^k\\}_{k = 1}^\\eta" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": " masks generated by SAM2's automatic mask generator in the entire image, and " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": " is the function that unifies semantics within each mask. Accordingly, for the " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": "-th sample, positive sample indexes are defined as: " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "P_{z}(i)\\equiv \\{p\\in T_{z}(i):\\tilde{\\mathbf{q}}_{p} = \\tilde{\\mathbf{q}}_{i}\\}" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": ", and negative sample indexes: " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "T_{z}(i)\\backslash P_{z}(i)" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": ". Similar to SimCLR [5], we employ contrastive learning loss to maximize agreement between positive samples. This method brings " + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "inline_equation", + "content": "P_{z}" + }, + { + "bbox": [ + 55, + 71, + 296, + 204 + ], + "type": "text", + "content": "-indexed positive samples closer in feature space, imposing effective contrastive constraints on sparse-view Gaussian learning, which can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 213, + 304, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 304, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 304, + 257 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n t r a}} = - \\sum_ {i \\in U _ {z}} \\frac {1}{\\mathcal {M} _ {z} (i)} \\sum_ {p \\in P _ {z} (i)} \\log \\left[ \\frac {e ^ {\\left(\\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {p}\\right) / \\tau\\right)}}{\\sum_ {j \\in T _ {z} (i)} e ^ {\\left(\\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j}\\right) / \\tau\\right)}} \\right], \\tag {6}", + "image_path": "e96c6e83a757ba4f3baf668bff3b01d8f78552142176d2ffa1ac86edfebb7491.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_z(i)" + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "text", + "content": " denotes the number of elements in " + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "inline_equation", + "content": "P_{z}(i)" + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "text", + "content": ". The symbol " + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "text", + "content": " refers to the inner product operation. " + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 257, + 296, + 293 + ], + "type": "text", + "content": " is the temperature parameter, set to 0.2 in the experiment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "content": "To obtain " + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "content": " samples, we employ a patch-wise sampling strategy by randomly extracting a patch of size " + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "content": " in each training iteration, where " + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "inline_equation", + "content": "N = h \\times w" + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "content": ". Empirically, " + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 293, + 296, + 353 + ], + "type": "text", + "content": " are both set to 128. In summary, the semantic prompt for 2D constraints can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 361, + 294, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 361, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 121, + 361, + 294, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S P C} ^ {2 d} = \\mathcal {L} _ {\\text {i n t e r}} + \\mathcal {L} _ {\\text {i n t r a}}. \\tag {7}", + "image_path": "973f0319b20dd4c174226df9590015740937b7d37d48825d4a5ce6b2d211ae28.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 385, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 529 + ], + "type": "text", + "content": "Semantic-Prompt for 3D Constrains. In addition to 2D constraints, we extend semantic-prompt to impose effective 3D constraints on 3D Gaussian primitives. Considering that Gaussian primitives with maximum weight dominate the rendering results (Eq. (2)), we encourage semantic feature similarity between the maximally-weighted 3D Gaussians from two corresponding regions. Specifically, for each pixel within the region mask, we identify the Gaussian primitive with maximum weight along the ray cast into 3D space. Then, the semantic features of these maximally-weighted Gaussians from training views are used to constrain their counterparts in pseudo views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 601 + ], + "type": "text", + "content": "However, since Gaussians related to region boundaries often span objects with differing semantics, directly applying the above constraints may complicate optimization. To mitigate this, we introduce a Region Boundary Erosion technique, which removes ambiguous primitives corresponding to region boundaries, which is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 610, + 296, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 610, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 96, + 610, + 296, + 644 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S P C} ^ {3 d} = \\mathcal {L} _ {k l} (\\mathcal {F} | | \\mathcal {F} ^ {*}) = \\sum_ {j = 1} ^ {n ^ {\\prime}} \\bar {\\mathbf {f}} \\log \\left(\\frac {\\bar {\\mathbf {f}}}{\\mathbf {f} _ {j} ^ {*}}\\right), \\tag {8}", + "image_path": "a6d0ce3891fb373f37cb22bbddf97faec2e8fe2d604fafe60f19dc22ebdcb04a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " denotes the mean feature representation of 3D Gaussian primitives derived from mask " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}'" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " in the training view, defined as " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{f}} = (\\mathbf{f} \\odot \\mathbf{M}') / \\sum \\mathbf{M}'" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}'" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " refers to the eroded mask after region boundary erosion, which is achieved through morphological erosion. Specifically, a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": "pixel " + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{M}'(x,y)" + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": " is retained as \"True\" only if all pixels within its " + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "inline_equation", + "content": "5\\times 5" + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": " neighborhood in the input original mask " + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": " are \"True\"; otherwise, it is set to \"False\". This operation effectively shrinks the boundaries of the region, causing them to contract inward. The set " + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{*} = \\{\\mathbf{f}_{1}^{*},\\mathbf{f}_{2}^{*},\\dots,\\mathbf{f}_{n^{\\prime}}^{*}\\}" + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": " comprises " + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "inline_equation", + "content": "n^\\prime" + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": " Gaussian primitives obtained from eroded masks " + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}^{*'}\\}_{p = 1}^{P}" + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": " in pseudo views." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 157, + 499, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 157, + 499, + 169 + ], + "spans": [ + { + "bbox": [ + 325, + 157, + 499, + 169 + ], + "type": "text", + "content": "To summarize, the overall loss is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 178, + 554, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 178, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 322, + 178, + 554, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {C} + \\mathcal {L} _ {S} + \\mathcal {L} _ {G S} + \\mathcal {L} _ {S} ^ {3 d} + \\lambda \\left(\\mathcal {L} _ {S P C} ^ {2 d} + \\mathcal {L} _ {S P C} ^ {3 d}\\right), \\tag {9}", + "image_path": "cc1aacb233aee416ec695d33a4eb3f8742166c88e4f5246caef4f332bc0b980b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GS}" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": " is the semantic loss for generated images from video generation model: " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GS} = \\mathcal{L}_{cos}(\\omega_f(\\hat{\\mathbf{F}}^g),\\mathbf{F}^g) + \\mathcal{L}_{ce}(\\omega_s(\\hat{\\mathbf{S}}^g),\\mathbf{S}^g)" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": ". These generated RGB images are not used for color constraints due to their error-prone color pixels that degrade overall results (see supplementary file). " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S^{3d}" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": " is the 3D local adaptive regularization, which promotes semantic similarity among neighboring Gaussians: " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S^{3d} = \\frac{1}{\\theta k}\\sum_{i = 1}^{\\theta}\\sum_{j = 1}^{k}\\omega_{ij}(|\\mathbf{f}_i - \\mathbf{f}_j|_2)" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\theta = 800" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": " denotes the number of Gaussian randomly sampled per iteration, and " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": " specifies the number of nearest neighbors for each sampled Gaussian in 3D space. " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\omega_{ij} = \\exp (-||\\mu_i - \\mu_j||_2)" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": " is the distance-based weighting coefficient. To prevent early-stage suboptimal convergence, " + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 203, + 555, + 374 + ], + "type": "text", + "content": " is initialized at 0 and then increased to 1 after 4k iterations during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 387, + 395, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 395, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 395, + 399 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 407, + 430, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 430, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 430, + 419 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 425, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 556 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two widely-used indoor scene datasets in terms of novel view synthesis: Replica [42] and ScanNet [8], using a sparse-input setting with only 12 input images for training. Following the standard sparse-input protocol [25, 60], every 10-th frame from each sequence is selected for testing, while 12 evenly sampled sparse views from the remaining images are used for training. These 12 views collectively encompass the entire scene. Notably, the \"inside-out\" viewing direction and limited overlap among views in sparse-input indoor scenes pose significant challenges for free-view synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "content": "Evaluation Metrics. We employ PSNR, SSIM [46], and LPIPS [57] for evaluating image quality. We report segmentation performance using the mean Intersection-over-Union (mIoU) and mean pixel accuracy (mAcc)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "Implementation Details. We implement our approach using the official PyTorch 3D Gaussian Splatting [19] on an NVIDIA A100 GPU. The dimension of the learnable semantic parameter attached to 3D Gaussians is set to 32. We modify the CUDA kernel to enable parallel differentiable rasterization of both color and semantic attributes. We employ CLIP ViT-B/16 [37] for feature extraction and SAM2 Hiera-L [38] for region mask generation. More details are provided in the supplementary file." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11269" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 191 + ], + "type": "image", + "image_path": "39dc835ba34161f6a6aa3912d2fbdf24f17e813be142c2f37db76f62b60723df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 193, + 539, + 204 + ], + "lines": [ + { + "bbox": [ + 69, + 193, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 193, + 539, + 204 + ], + "type": "text", + "content": "Figure 4. Visual reconstruction results on novel views. Our approach achieves superior global structure and photo-realistic details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 214, + 546, + 328 + ], + "blocks": [ + { + "bbox": [ + 62, + 214, + 546, + 328 + ], + "lines": [ + { + "bbox": [ + 62, + 214, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 546, + 328 + ], + "type": "table", + "html": "
MethodReplica [42]ScanNet [8]
PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑mAcc ↑
3DOVS [31]13.2000.5150.6633.91321.43918.7210.6430.5447.75136.356
Feature 3DGS [59]17.0590.7000.3597.40730.20918.0880.6560.4247.34738.002
LangSplat [35]17.7150.7010.4203.29218.62118.7800.6630.4763.33924.491
Gau-Grouping [50]17.2970.6910.37615.02850.42219.0060.6750.42115.17257.355
DNGaussian [25]18.5130.7120.40216.65754.40018.3920.6910.48413.91254.814
FSGS [60]18.6940.7250.32816.10153.47420.5740.6960.40614.03156.850
CoR-GS [56]18.9550.7580.31215.93352.47620.8150.7210.39314.46255.812
Ours22.0110.7920.25423.96063.26222.4010.7410.36828.94468.823
", + "image_path": "a92ba41123c0efb8fbbb6108810d99b52e5a34df6b7fe03c98b2c21638879813.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 553, + 351 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 553, + 351 + ], + "type": "text", + "content": "Table 1. Quantitative results of reconstruction and segmentation on novel views on the Replica and ScanNet datasets, using the CLIP [37] for optimization with 12 training views. Our approach achieves superior performances across all metrics on various datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 371, + 278, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 278, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 278, + 384 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": "We conduct novel view synthesis comparisons using sparse input data in indoor scenes, including two types of methods. i) 3D open-vocabulary segmentation methods, including 3DOVS [31], Feature 3DGS [59], LangSplat [35], and Gau-Grouping [50], are designed for 3D open world semantic understanding. ii) Sparse-input free-view synthesis methods contain DNGaussian [25], FSGS [60], and CoR-GS [56], which are designed to learn Gaussian radiance fields from sparse inputs, focusing primarily on appearance modeling. For concurrent 3D open-world segmentation, we extend these methods by adding semantic properties to Gaussian and incorporating CLIP-based semantic supervision, aligning with the semantic settings of our approach. We adopt consistent evaluations across all methods by employing the same sparse-input training protocol and test views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 569, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 569, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 569, + 295, + 677 + ], + "type": "text", + "content": "Qualitative Results. i) Reconstruction. As shown in Fig. 4, the results from existing 3D open-vocabulary segmentation methods exhibit significant artifacts, since these methods lack sufficient view constraints. Although sparse-view reconstruction methods introduce geometric constraints for regularization, they struggle to recover photo-realistic details. In comparison, our approach incorporates scene-Layout points prior and semantic consistency constraints, leading to a more detailed appearance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "ii) Open-world Segmentation. In Fig. 5, current methods struggle to obtain precise and complete results, since they easily encounter ambiguity arising from insufficient seman" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 408 + ], + "type": "text", + "content": "tic view supervision. In contrast, our method delivers accurate results, benefiting from the SPC regularization that provides effective consistency constraints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 554 + ], + "type": "text", + "content": "Quantitative Results. Table 1 shows that, when optimized using the CLIP model [37], our approach significantly outperforms existing methods across all metrics on both synthetic and real datasets. Competing methods struggle under sparse input conditions, due to either limited Gaussian representations or insufficient semantic consistency constraints. Notably, LangSplat faces considerable challenges, as it freezes Gaussian coordinates, scale, and rotation during semantic parameter optimization, which severely impedes semantic learning under sparse inputs. Besides, our method maintains superior results when optimized with CLIP-LSeg [24] (see supplementary file), validating its generalizability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 416, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 416, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 416, + 574 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Key Components. In Table 2, applying the VGM-based Point Creation (PC, #1) and Scene layout Point Generation (SPG, #2) notably improves reconstruction quality over setting #0, and consequently boosts segmentation results, validating SGI's effectiveness in enhancing Gaussian representation. Adding the full Semantic-Prompt Consistency regularization (SPC, #4) further yields significant segmentation gains, showing SPC's capability to address insufficient view constraints and enhance 3D scene understanding. Fig. 6 also shows that adding SGI can reduce artifacts, and applying SPC further enhances segmentation results." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11270" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 218 + ], + "type": "image", + "image_path": "7291b47ef891b4f3d200342323ca7ad3c9359564c2cbe40ccc8b187ade11adb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 219, + 527, + 229 + ], + "lines": [ + { + "bbox": [ + 82, + 219, + 527, + 229 + ], + "spans": [ + { + "bbox": [ + 82, + 219, + 527, + 229 + ], + "type": "text", + "content": "Figure 5. Visual open-world segmentation results on novel views. Our approach yields more accurate and complete results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 248, + 296, + 321 + ], + "blocks": [ + { + "bbox": [ + 56, + 248, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 56, + 248, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 296, + 321 + ], + "type": "table", + "html": "
CaseSGISPCReplica [42]
PCSPG2D3DPSNRSSIMLPIPSmIoUmAcc
#017.1340.6780.36914.89951.255
#1✓18.9240.7240.33416.27952.212
#2✓✓21.1760.7830.26019.40158.281
#3✓✓✓21.8140.7910.25722.84961.362
#4✓✓✓✓22.0110.7920.25423.96063.262
", + "image_path": "4370fbe5f81b374970c978bcb62eb166aa2bf3f4aa3b757eaea7c29c7f7eb91d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 348, + 296, + 400 + ], + "blocks": [ + { + "bbox": [ + 95, + 323, + 256, + 334 + ], + "lines": [ + { + "bbox": [ + 95, + 323, + 256, + 334 + ], + "spans": [ + { + "bbox": [ + 95, + 323, + 256, + 334 + ], + "type": "text", + "content": "Table 2. Ablation study on key components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 348, + 296, + 400 + ], + "lines": [ + { + "bbox": [ + 56, + 348, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 296, + 400 + ], + "type": "table", + "html": "
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aOurs (Full setting)22.0110.7920.25423.96063.262
#bw/o View-to-View Regular.21.6830.7890.25721.24460.914
#cw/o Contrastive Regular.21.8790.7920.25622.32260.086
#dw/o Boundary Erosion21.9500.7920.25422.96461.706
", + "image_path": "dea76bbf5fdb8550807a89a47938d54e0666b36056df0c8e46b59b47e400f683.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 427, + 293, + 519 + ], + "blocks": [ + { + "bbox": [ + 58, + 427, + 293, + 519 + ], + "lines": [ + { + "bbox": [ + 58, + 427, + 293, + 519 + ], + "spans": [ + { + "bbox": [ + 58, + 427, + 293, + 519 + ], + "type": "image", + "image_path": "bff21012490d3246c721469b28a5147bca5e32be66c4f26525755105c02ef230.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 520, + 279, + 530 + ], + "lines": [ + { + "bbox": [ + 70, + 520, + 279, + 530 + ], + "spans": [ + { + "bbox": [ + 70, + 520, + 279, + 530 + ], + "type": "text", + "content": "Figure 6. Visual ablation comparison of key components." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 627 + ], + "type": "text", + "content": "Components in SPC. In Table 3, removing either the 2D view-to-view or contrastive regularization (#b, #c), or omitting the region boundary erosion operation in the 3D constraint (#d), degrades performance. This validates their roles in mitigating insufficient supervision and filtering ambiguous Gaussian primitives spanning different objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 664 + ], + "type": "text", + "content": "Robustness. In Table 4, our strategies also achieve significant gains even with sparse SfM initialization from limited training images, validating their robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Initialization. From top to bottom in Fig. 7, we can see that leveraging denser and more effective points for Gaussian initialization promotes 3DGS training that enhances scene representation, leading to improved rendered results." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 247, + 554, + 293 + ], + "blocks": [ + { + "bbox": [ + 67, + 402, + 282, + 412 + ], + "lines": [ + { + "bbox": [ + 67, + 402, + 282, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 402, + 282, + 412 + ], + "type": "text", + "content": "Table 3. Ablation results of our SPC with different settings." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 247, + 554, + 293 + ], + "lines": [ + { + "bbox": [ + 315, + 247, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 554, + 293 + ], + "type": "table", + "html": "
CaseConfigurationPSNRSSIMLPIPSmIoUmAcc
#aSfM Points Initialization17.1340.6780.36914.89951.255
#b#a + SGI20.6800.7740.27018.16956.928
#c#b + SPC21.5890.7860.26020.76661.000
", + "image_path": "d00adbef7ca8e5a2dc7cfec9fd5a4e836442f7118ed187d478f7a8c475eca1fd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 293, + 549, + 304 + ], + "lines": [ + { + "bbox": [ + 318, + 293, + 549, + 304 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 549, + 304 + ], + "type": "text", + "content": "Table 4. Results of our strategies using SfM point initialization." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 317, + 550, + 479 + ], + "blocks": [ + { + "bbox": [ + 317, + 317, + 550, + 479 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 550, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 550, + 479 + ], + "type": "image", + "image_path": "8a259ff223ed1b144b32b4bf5c8e0da33b2c67e8523d1f9a0edc3449fbf31545.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 480, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 502 + ], + "type": "text", + "content": "Figure 7. Visual comparison of different initializations and their rendering results after optimization with these initializations." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 542, + 555, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 542, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 312, + 542, + 555, + 709 + ], + "type": "text", + "content": "This paper presents SPC-GS, a novel approach to address the challenges of GS-based indoor open-world free-view synthesis with sparse views. In SPC-GS, the proposed Scene layout-based Gaussian Initialization (SGI) effectively enhances Gaussian representation by providing an instructive point initialization, consequently promoting semantic learning. Additionally, the Semantic-Prompt Consistency (SPC) regularization mitigates limited supervision by introducing additional view-consistent semantic constraints, improving coherent segmentation performance. Experimental results show that our approach significantly outperforms SOTA methods in reconstruction and 3D open-world segmentation tasks across synthetic and real-world datasets with various CLIP models." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11271" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 411 + ], + "type": "text", + "content": "[6] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Proceedings of the European Conference on Computer Vision, pages 370-386. Springer, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 413, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 467 + ], + "type": "text", + "content": "[7] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 524 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 6, 7, 1, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "text", + "content": "[9] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics, 36(4): 1, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "text", + "content": "[10] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[11] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In Proceedings of the International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3d gaussian splatting as new era: A survey. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[14] Xu Hu, Yuxi Wang, Lue Fan, Junsong Fan, Junran Peng, Zhen Lei, Qing Li, and Zhaoxiang Zhang. Semantic anything in 3d gaussians. arXiv preprint arXiv:2401.17857, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[16] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 10608-10615, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[17] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 360 + ], + "type": "text", + "content": "[18] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Ganesh Iyer, Soroush Saryazdi, Tao Chen, Alaa Maalouf, Shuang Li, Nikhil Varma Keetha, Ayush Tewari, Joshua Tenenbaum, Celso de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. In Proceedings of Robotics: Science and Systems, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "text", + "content": "[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "type": "text", + "content": "[21] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 559 + ], + "type": "text", + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Proceedings of the Advances in Neural Information Processing Systems, pages 23311-23330, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "text", + "content": "[23] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, pages 29-43. Wiley Online Library, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "text", + "content": "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In Proceedings of the International Conference on Learning Representations, 2022. 3, 7, 6, 8, 10" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[25] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2, 3, 6, 7, 9, 10" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11272" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[26] Guibiao Liao, Wei Gao, Qiuping Jiang, Ronggang Wang, and Ge Li. Mmnet: Multi-stage and multi-scale fusion network for rgb-d salient object detection. In Proceedings of the ACM international conference on multimedia, pages 2436-2444, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "type": "text", + "content": "[27] Guibiao Liao, Jiankun Li, Zhenyu Bao, Xiaqing Ye, Jingdong Wang, Qing Li, and Kanglin Liu. Clip-gs: Clip-informed gaussian splatting for real-time and view-consistent 3d semantic understanding. arXiv preprint arXiv:2404.14249, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 239 + ], + "type": "text", + "content": "[28] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3351-3359, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[29] Guibiao Liao, Kaichen Zhou, Zhenyu Bao, Kanglin Liu, and Qing Li. Ov-nerf: Open-vocabulary neural radiance fields with vision and language foundation models for 3d semantic understanding. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": "[30] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv preprint arXiv:2403.03174, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "type": "text", + "content": "[31] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In Proceedings of the Advances in Neural Information Processing Systems, pages 53433-53456, 2023. 2, 3, 7, 9, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 459, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 502 + ], + "type": "text", + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41 (4):1-15, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 502, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 568 + ], + "type": "text", + "content": "[34] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 569, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 624 + ], + "type": "text", + "content": "[35] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 1, 3, 7, 2, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[36] Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. Goi: Find 3d gaussians of interest with an estimizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5328-5337, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 689, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 294, + 714 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 127 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 6, 7, 8, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[38] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "type": "text", + "content": "[39] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Flashsplat: 2d to 3d gaussian splattering segmentation solved optimally. In Proceedings of the European Conference on Computer Vision, pages 456-472. Springer, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 296 + ], + "type": "text", + "content": "[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193-13203, 2024. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 296, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 296, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 553, + 353 + ], + "type": "text", + "content": "[41] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3d gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 353, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 409 + ], + "type": "text", + "content": "[42] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 7, 8, 1, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 410, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 555, + 464 + ], + "type": "text", + "content": "[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In Proceedings of the International Conference on 3D Vision, pages 443-453, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "type": "text", + "content": "[44] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 521, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 589 + ], + "type": "text", + "content": "[45] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 589, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 633 + ], + "type": "text", + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 634, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 689 + ], + "type": "text", + "content": "[47] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In Proceedings of the ACM SIGGRAPH, pages 1-11, 2024. 4, 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 689, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 553, + 714 + ], + "type": "text", + "content": "[48] Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11273" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "gaussian splatting. Computational Visual Media, 10(4):613-642, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[49] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "text", + "content": "[50] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In Proceedings of the European Conference on Computer Vision, pages 162-179. Springer, 2024. 1, 3, 7, 9, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 253 + ], + "type": "text", + "content": "[51] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20612-20622, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "type": "text", + "content": "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 295, + 354 + ], + "type": "text", + "content": "[53] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 356, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 295, + 409 + ], + "type": "text", + "content": "[54] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 412, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 295, + 465 + ], + "type": "text", + "content": "[55] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In Proceedings of the European Conference on Computer Vision, pages 57-74. Springer, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 521 + ], + "type": "text", + "content": "[56] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatt-ting via co-regularization. In Proceedings of the European Conference on Computer Vision, pages 335–352. Springer, 2024. 3, 7, 9, 10" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 524, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 295, + 579 + ], + "type": "text", + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "text", + "content": "[58] Xiaoyu Zhang, Guibiao Liao, Wei Gao, and Ge Li. Tdrnet: Transformer-based dual-branch restoration network for geometry based point cloud compression artifacts. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1–6. IEEE, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 712 + ], + "type": "text", + "content": "[59] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 1, 3, 7, 2, 9, 10" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 174 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[60] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In Proceedings of the European Conference on Computer Vision, pages 145-163. Springer, 2024. 2, 3, 5, 6, 7, 9, 10" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 174 + ], + "type": "text", + "content": "[61] Xingxing Zuo, Pouya Samangouei, Yunwen Zhou, Yan Di, and Mingyang Li. Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understanding. arXiv preprint arXiv:2401.01970, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11274" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_content_list.json b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a8e2549d7d00b39e9240c720e3778296a840c0da --- /dev/null +++ b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_content_list.json @@ -0,0 +1,1472 @@ +[ + { + "type": "text", + "text": "SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking", + "text_level": 1, + "bbox": [ + 158, + 128, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenrui Cai $^{1}$ , Qingjie Liu $^{1,2,3,*}$ , Yunhong Wang $^{1,3}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China \n $^{2}$ Zhongguancun Laboratory, Beijing, China \n $^{3}$ Hangzhou Innovation Institute, Beihang University, Hangzhou, China \n{wenrui_cai, qingjie.liu, yhwang}@buaa.edu.cn", + "bbox": [ + 99, + 202, + 908, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most state-of-the-art trackers adopt one-stream paradigm, using a single Vision Transformer for joint feature extraction and relation modeling of template and search region images. However, relation modeling between different image patches exhibits significant variations. For instance, background regions dominated by target-irrelevant information require reduced attention allocation, while foreground, particularly boundary areas, need to be emphasized. A single model may not effectively handle all kinds of relation modeling simultaneously. In this paper, we propose a novel tracker called SPMTrack based on mixture-of-experts tailored for visual tracking task (TMoE), combining the capability of multiple experts to handle diverse relation modeling more flexibly. Benefiting from TMoE, we extend relation modeling from image pairs to spatio-temporal context, further improving tracking accuracy with minimal increase in model parameters. Moreover, we employ TMoE as a parameter-efficient fine-tuning method, substantially reducing trainable parameters, which enables us to train SPMTrack of varying scales efficiently and preserve the generalization ability of pretrained models to achieve superior performance. We conduct experiments on seven datasets, and experimental results demonstrate that our method significantly outperforms current state-of-the-art trackers. The source code is available at https://github.com/WenRuiCai/SPMTrack.", + "bbox": [ + 89, + 359, + 485, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 785, + 223, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual tracking aims to predict the location of a target throughout subsequent frames of a video, given the initial state in the first frame template. Currently, most state-of-the-art trackers adopt a Transformer-based [46] one-stream", + "bbox": [ + 89, + 811, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71693ee43b859c952338b7d95fcddc5a505a877abccf99af03e02f680d0d802d.jpg", + "image_caption": [ + "Figure 1. Comparison of LaSOT AUC and model parameter count across different trackers. Larger loop indicates better performance." + ], + "image_footnote": [], + "bbox": [ + 519, + 324, + 901, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "paradigm [3, 5, 10, 33, 52, 59, 61], utilizing a single Vision Transformer [12] backbone that accepts both the template and the search region image as input, and employs self-attention to perform image feature extraction and template-search region relation modeling simultaneously.", + "bbox": [ + 511, + 625, + 906, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, not all relation modeling between image patches benefits tracking performance. Prior studies [4, 20, 58, 59] have demonstrated that the abundance of background patches in the search region deteriorates the discriminative power of the foreground features, and single vanilla attention has limited capability in suppressing undesirable foreground-background interactions. Meanwhile, there are also some kinds of relation modeling that demonstrate positive effects on tracking performance [17, 19]. For instance, enhancing attention to target boundary regions can improve tracking performance [17], especially in challenging scenarios such as motion blur, partial occlusion, and deformation.", + "bbox": [ + 511, + 702, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address various relation modeling, numerous meth", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 886, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ods propose specific designs, including background filtering to reduce target-irrelevant information [3, 59], customizing different interaction modules for tokens belonging to different categories [4, 20], and attention score adjustment based on foreground-background weights of each token [58]. However, existing methods can only handle one or a few predefined types of relation modeling, and the handcrafted strategies inherently lack adaptability. Furthermore, there may exist many latent relationships that further complicate the specialized design. These limitations motivate our question: How to design a tracker to adaptively process various relation modeling between image patches?", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by mixture of experts (MoE) [1, 11, 13, 27, 41] in natural language processing, in this paper, we address the question by proposing SPMTrack, a novel tracker based on TMoE, a specialized mixture of experts module for visual tracking task. Similar to MoE, TMoE handles diverse relation modeling through adaptive weighted combinations of multiple experts. However, unlike traditional MoE that is exclusively applied to feed-forward network (FFN) layers, TMoE is simultaneously applied to the linear layers within both the attention layers and the FFN layers. This finer-grained application enables TMoE to achieve more diverse expert combinations, thereby enhancing the capability of various relation modeling. Moreover, differing from traditional MoE that uses FFNs as experts, all the experts in TMoE are linear layers and designed with a lightweight and efficient structure to ensure overall efficiency. Furthermore, we employ TMoE as a parameter-efficient finetuning method, which enables us to train SPMTrack of larger scales. In SPMTrack, only a subset of experts and the router within TMoE, along with the prediction head need to be trained, which substantially reduces the trainable parameters while preserving the generalization capability of pretrained models. Notably, our method reduces the number of trainable parameters by nearly $80\\%$ compared to [2], while demonstrates better performance.", + "bbox": [ + 91, + 277, + 483, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benefiting from the powerful capability of TMoE in handling diverse relation modeling, unlike previous one-stream trackers that only perform relation modeling between template-search region pairs, we extend SPMTrack to incorporate multi-frame spatio-temporal context modeling, which further enhances the performance of SPMTrack with minimal additional parameters. We evaluated the performance of our method on seven datasets, and the experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets including LaSOT [14], GOT-10K [25], TrackingNet [38] and TNL2K [49]. Through extensive experiments with models of varying scales, as shown in Figure 1, Table 1 and Table 2, with ViT-B [12] as backbone and less than $30\\%$ of parameters need to be trained, SPMTrack-B achieves or even surpasses the best performance of previous trackers utilizing ViT-L.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper can be summarized as: (1) We propose TMoE, a mixture of experts module tailored for visual tracking. TMoE enhances the capability to handle various relation modeling and can be used as a method of parameter-efficient fine-tuning. (2) Based on TMoE, we propose SPMTrack, a novel tracker that can effectively integrate spatio-temporal context for visual tracking. (3) Experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets. We trained SPMTrack of varying scales, and our method with ViT-B as backbone can achieve or even surpass current trackers using ViT-L.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 287, + 653, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. One-Stream Trackers", + "text_level": 1, + "bbox": [ + 511, + 314, + 718, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One-stream trackers employ a single Vision Transformer [12] as backbone, simultaneously performing feature extraction and relation modeling at each layer [5, 10, 55, 59]. Many efforts focus on incorporating historical context. TATrack [22] and ODTrack [61] propose architectures capable of processing multiple frames, while HIPTrack [3] introduces historical target features through prompt learning. SeqTrack [6], ARTrack[51], and ARTrackV2 [2] achieve more precise predictions by incorporating multiple historical target position. Other works try to improve conventional self-attention to handle different relation modeling. OSTrack [59] progressively filters background regions to reduce foreground-background relation modeling, GRM [20] and ROMTrack [4] categorizes image tokens and constrains relation modeling between specific categories, F-BDMTrack [58] adjusts the attention score by calculating the foreground-background weights of tokens. But these methods can only handle specific relation modeling and rely on manually designed strategies.", + "bbox": [ + 511, + 337, + 908, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Mixture of Experts", + "text_level": 1, + "bbox": [ + 511, + 636, + 697, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture-of-experts (MoE) [1, 13, 27] are widely applied in large language models [9, 11, 15, 36, 41, 53], where each expert focuses on specific aspects of the data or particular tasks [48]. By performing an adaptive weighted sum over multiple experts, MoE can better capture complex patterns and relationships within the input data. MoE also has applications in computer vision [8, 43, 60]. However, there is few dedicated exploration of MoE in visual tracking. MoETrack [44] employs multiple prediction heads as experts in the field of RGB-T tracking, which is rather far-fetched and deviates from the general practice of embedding MoE within Transformer layers. In the field of RGB-E tracking, eMoETracker [7] inserts multiple experts between Transformer layers, which more closely resembles an adapter architecture [23] and contrasts with MoE implementations where experts are embedded within Transformer layers.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16872", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ad4663e1941ba435163c2dc67801f2da2cfb13becb701174fd848c1830728d7.jpg", + "image_caption": [ + "Figure 2. Overview of SPMTrack that consists of a feature extraction network and a prediction head. The main body of the feature extraction network is a Transformer encoder composed of multiple TMoEBlocks. The structure of TMoEBlock is shown on the right side." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 898, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Parameter-efficient Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 431, + 372, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-efficient fine-tuning aims to reduce computational costs and retain the general knowledge of the pretrained model by freezing most parameters and fine-tuning only a small subset of parameters. Common parameter-efficient fine-tuning approaches include adapter-based methods [18, 23, 40], prompt-based methods [28, 31, 50, 62] and low-rank adaptation methods [24, 45]. In visual tracking, HIPTrack [3] incorporates historical target features as prompts, while LoRAT [33] employs low-rank adaptation for efficient training. And our method maintains flexibility in diverse relation modeling while reducing trainable parameters by fine-tuning TMoE.", + "bbox": [ + 89, + 455, + 483, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 654, + 183, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 679, + 287, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, we propose SPMTrack, a novel tracker that integrates multi-frame spatio-temporal context and employs TMoE for parameter-efficient fine-tuning. The overall architecture follows the one-stream paradigm, consisting of a Feature Extraction Network based on a Vision Transformer [12] as backbone, and a prediction head.", + "bbox": [ + 89, + 702, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extraction Network. The feature extraction network is denoted as $\\Phi$ . The main body of the feature extraction network is a Transformer encoder that utilizes TMoE-Block as its primary component, which is designed based on TMoE. The architecture of TMoE-Block is illustrated in the right panel of Figure 2, and details will be introduced in Section 3.2. The encoder simultaneously per", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "forms feature extraction and relation modeling for input images. We extend the input from one single template $\\mathbf{Z} \\in \\mathbb{R}^{H_z \\times W_z \\times 3}$ that is commonly used in one-stream trackers to the video sequence $\\{Z_1, \\dots, Z_N | Z_i \\in \\mathbb{R}^{H_z \\times W_z \\times 3}\\}$ as reference frames. The reference frames contain the first frame template and the selected tracked historical frames, with a total of $N$ frames. The feature extraction network uses patch embedding to partition each input image into patches of size $M \\times M$ , map all patches into tokens by a convolutional layer, and obtain the reference token sequence of each image $T_i \\in \\mathbb{R}^{N_T \\times d}$ , where $d$ denotes the hidden dimension of Transformer encoder and $N_T = \\frac{H_z W_z}{M^2}$ represents the number of tokens corresponding to each image. Similarly, for the search region $S \\in \\mathbb{R}^{H_s \\times W_s \\times 3}$ , we can also obtain the token sequence $X \\in \\mathbb{R}^{N_X \\times d}$ . All tokens are added with positional embedding and a learnable token type embedding [33]. Inspired by [32, 61], we additionally introduce a target state token $H \\in \\mathbb{R}^{1 \\times d}$ that is learnable as well and propagates over time. All tokens are concatenated and fed into the Transformer encoder. The process can be formulated as:", + "bbox": [ + 511, + 431, + 906, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} _ {i} ^ {j} = \\left\\{ \\begin{array}{l} \\boldsymbol {T} _ {i} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {o}, i f \\text {t o k e n} (i, j) \\text {i n b b o x} \\\\ \\boldsymbol {T} _ {i} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {b}, o t h e r w i s e \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 750, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {j} = \\boldsymbol {X} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {S}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 791, + 738, + 805 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} = \\operatorname {C o n c a t} (\\boldsymbol {H}, \\boldsymbol {T} _ {1}, \\dots , \\boldsymbol {T} _ {N}, \\boldsymbol {X})\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 808, + 769, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{i}^{j}$ represents the $j^{th}$ token in $T_{i}$ , $PE_{j}$ represents positional embedding of the $j^{th}$ token, similarly for $X^{j}$ . $TE_{o}, TE_{b}$ and $TE_{S}$ represents the learnable type embeddings of the foreground region tokens, background region", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tokens in reference frames and the search region tokens, respectively. $\\pmb{I}$ denotes the input to the Transformer encoder.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the Transformer encoder processes all tokens and produces an output $\\pmb{O} \\in \\mathbb{R}^{(1 + N \\times N_T + N_X) \\times d}$ . The output tokens $\\pmb{X}' \\in \\mathbb{R}^{N_X \\times d}$ corresponding to search region and the output $\\pmb{H}' \\in \\mathbb{R}^{1 \\times d}$ from the target state token are fed into the prediction head for prediction. $\\pmb{H}'$ will be added with the input target state token in the next tracking frame, which can be formulated as:", + "bbox": [ + 89, + 121, + 483, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {O} = \\Phi (\\boldsymbol {I}), [ \\boldsymbol {H} ^ {\\prime}, \\boldsymbol {T} _ {1} ^ {\\prime}, \\dots , \\boldsymbol {T} _ {N} ^ {\\prime}, \\boldsymbol {X} ^ {\\prime} ] = \\operatorname {S p l i t} (\\boldsymbol {O}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 234, + 482, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction Head. The input to the prediction head consists of the output target state token $H'$ and the output search region tokens $X'$ from the feature extraction network. Initially, the prediction head performs matrix multiplication between $H'$ and $X'$ to obtain the weight $U \\in \\mathbb{R}^{N_X \\times 1}$ . As the target state token propagates over time, $U$ can reweight and adjust current search region tokens based on the historical state of the target, thereby introducing more spatio-temporal tracking context. After element-wise multiplication of $X'$ with $U$ , the resulting data is reshaped into a feature map $F \\in \\mathbb{R}^{\\frac{H_s}{M} \\times \\frac{W_s}{M} \\times d}$ , which is then processed by a decoupled double-MLP head to perform target center classification and box regression, respectively. For fair comparison, the double-MLP head is identical to [33].", + "bbox": [ + 89, + 253, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. The Design of TMoEBlock", + "text_level": 1, + "bbox": [ + 89, + 474, + 331, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TMoEBlock serves as the primary component of the Transformer encoder in the feature extraction network. As illustrated in the right of Figure 2, TMoEBlock is built based on the standard vision Transformer block. However, unlike the standard ViT block, TMoEBlock applies TMoE module to both the multi-head self-attention (MSA) layer and the feed-forward network (FFN) layer, obtaining MoE-based MSA $(\\mathrm{M}^2\\mathrm{SA})$ layer and MoE-based FFN (MFFN) layer. The finer-grained application of TMoE is capable to achieve more diverse combinations for various relation modeling. The processing of TMoEBlock can be formulated as:", + "bbox": [ + 89, + 496, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {O} _ {l} ^ {\\prime} = \\mathrm {M} ^ {2} \\operatorname {S A} (\\ln (\\boldsymbol {O} _ {l - 1})) + \\boldsymbol {O} _ {l - 1}, \\quad l = 1, \\dots , L \\tag {3} \\\\ \\boldsymbol {O} _ {l} = \\mathrm {M F F N} (\\mathrm {L N} (\\boldsymbol {O} _ {l} ^ {\\prime})) + \\boldsymbol {O} _ {l} ^ {\\prime}, l = 1, \\dots , L \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 680, + 482, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L$ denotes the total number of TMoEBlocks, $\\pmb{O}_l$ represents the output of the $l^{th}$ block in Transformer encoder. When $l = L$ , $\\pmb{O}_l$ is the output $\\pmb{O}$ of the feature extraction network; when $l = 0$ , it is the overall input $\\pmb{I}$ . LN stands for the layer normalization operation.", + "bbox": [ + 89, + 719, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MoE-based Multi-head Self-Attention (M²SA). M²SA not only replaces the standard three linear projection layers $Q, K, V$ in standard attention for the query, key, and value with TMoE modules but also replaces the output projection layer with TMoE module.", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MoE-based Feed-Forward Network (MFFN). Conventional feed-forward network consists of two linear layer and", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5594208d4c14b2f214b8902ee400b433966b6c8af60d984be909b91dc960d36b.jpg", + "image_caption": [ + "Figure 3. The structure of TMoE. The symbols maintain the same meaning with Figure 2." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 897, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "activation function. MFFN replaces both linear layers in FFN with TMoE modules, while maintaining all other architectural configurations unchanged.", + "bbox": [ + 511, + 292, + 903, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mixture of Experts for Visual Tracking", + "text_level": 1, + "bbox": [ + 511, + 345, + 849, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce TMoE, a mixture of experts module tailored for visual tracking task. As illustrated in Figure 3, TMoE consists of one shared expert, one router, one compression expert and $N_{e}$ routed experts. In large language models, mixture of experts are commonly applied to replace FFN layers, and the experts comprising MoE are typically the same structure as FFN. However, TMoE is used to replace the linear layers in both self-attention and FFN, and the shared expert, the compression expert and the routed experts are all implemented as linear layers.", + "bbox": [ + 511, + 367, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In TMoE, the shared expert directly copies weights from the corresponding linear layers in pretrained model and remains frozen during training. The rationale is that the original weights have been well-trained during the pre-training stage, and the shared expert needs to retain more general knowledge. In contrast, the compression experts, routed experts, and the router are all trainable.", + "bbox": [ + 511, + 518, + 905, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a TMOE module replacing a linear layer with input dimension $d$ and output dimension $D$ , for each token $\\pmb{x} \\in \\mathbb{R}^d$ , as shown in Figure 3, TMOE first computes the weights $\\pmb{W} \\in \\mathbb{R}^{N_e}$ of all routed experts based on the input $\\pmb{x}$ , the weights are normalized by Softmax. After that, TMOE calculates the output $\\pmb{y}^s \\in \\mathbb{R}^D$ of the shared expert and the outputs $\\pmb{y}_i^e \\in \\mathbb{R}^D$ of all routed experts. In particular, instead of adopting the same structure as the shared expert, inspired by [45], we first employ a compression expert with a $d \\times r$ linear transformation to reduce the input dimension from $d$ to $r$ and obtain the compressed result $\\pmb{y}^c \\in \\mathbb{R}^r$ , where $r \\ll d$ and $r \\ll D$ . Subsequently, all routed experts use the compressed input $\\pmb{y}^c$ to produce their respective outputs $\\pmb{y}_i^e \\in \\mathbb{R}^D$ . The process can be formally described as:", + "bbox": [ + 511, + 625, + 906, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {W} = \\text {S o f t m a x} (\\operatorname {R o u t e r} (\\boldsymbol {x})), \\quad \\boldsymbol {y} ^ {s} = \\boldsymbol {E} ^ {S} (\\boldsymbol {x}) \\\\ \\boldsymbol {y} ^ {c} = \\boldsymbol {E} ^ {C} (\\boldsymbol {x}), \\quad \\boldsymbol {y} _ {i} ^ {e} = \\boldsymbol {E} _ {i} ^ {R} (\\boldsymbol {y} ^ {c}) \\quad i = 1, \\dots , N _ {e} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 845, + 849, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Router is also a linear layer that maps the input $x$", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16874", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to weights $\\pmb{W}$ , and $\\pmb{E}^{S}$ , $\\pmb{E}^{C}$ , and $\\pmb{E}_{i}^{R}$ denote the shared expert, compress expert, and $i^{th}$ routed expert, respectively.", + "bbox": [ + 89, + 89, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the outputs of all experts, we compute a weighted sum of the outputs $\\pmb{y}_i^e$ from all routed expert based on $\\pmb{W}$ to obtain $\\pmb{y}^e \\in \\mathbb{R}^D$ , which means that the routed experts are densely activated in TMoE. Subsequently, $\\pmb{y}^e$ is added with the shared expert output $\\pmb{y}^s$ to get the final output $\\pmb{y} \\in \\mathbb{R}^D$ . The process can be formulated as follows:", + "bbox": [ + 89, + 121, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} ^ {e} = \\sum_ {i = 1} ^ {N _ {e}} \\boldsymbol {W} _ {i} \\boldsymbol {y} _ {i} ^ {e}, \\quad \\boldsymbol {y} = \\boldsymbol {y} ^ {e} + \\boldsymbol {y} ^ {s} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 223, + 482, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 267, + 235, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The outputs of the decoupled double-MLP prediction head are supervised using binary cross-entropy loss for target center classification and Generalized IoU loss [42] for bounding box regression, respectively. Both loss terms are assigned with the weighting coefficients set to 1.", + "bbox": [ + 89, + 289, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 378, + 223, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 402, + 307, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model settings. We provide three versions of SPMTrack: SPMTrack-B, SPMTrack-L, and SPMTrack-G, utilizing ViT-B [12], ViT-L [12], and ViT-G [39] as backbones respectively. All versions maintain a consistent patch size $M$ of 14 and adopt the pretrained weights from DINOv2 [39]. The resolution of input images are the same across all versions, with reference frames of $196 \\times 196$ and search region of $378 \\times 378$ . A crop factor of 2 is applied to reference frames and 5 to search region. We set the number of reference frames $N$ to 3 for spatio-temporal modeling. Each TMoE module incorporates 4 routed experts, where the input dimension of the routed experts $r$ is set to 64. Other configurations and model information are shown in Table 1, where $L$ is the number of TMoEBlocks, $d$ denotes hidden dimension and $N_h$ is the number of attention heads.", + "bbox": [ + 89, + 424, + 483, + 650 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/52fb3ba64ffb8c0cd46e6d57ccd64d59f56083ad583c29e4d88b5c2308953b22.jpg", + "table_caption": [ + "Table 1. Model configurations and parameter counts across SPM-Track variants of different scales." + ], + "table_footnote": [], + "table_body": "
SPMTrack-BSPMTrack-LSPMTrack-G
BackboneL=12L=24L=40
d=768d=1024d=1536
Nh=12Nh=16Nh=24
#Params (M)115.3379.61339.5
#Trainable \nParams (M)29.275.9204.0
", + "bbox": [ + 91, + 702, + 478, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following previous works [3, 33, 51, 59], we use LaSOT [14], TrackingNet [47], GOT-10K [25], and COCO [34] for training. When evaluating on GOT-10K test set, we strictly follow the protocol by training exclusively on GOT-10K training set. When evaluating on other datasets,", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we jointly train on the training sets of all four datasets, randomly sampling from each dataset with equal probability. For video datasets, we sample 5 frames from one video with random frame intervals ranging from 1 to 200. The sampled frames are arranged either in forward or reverse order with equal probability, where the first 3 frames serve as reference frames and the remaining 2 as search frames. For COCO, we replicate each image 5 times to maintain consistency with the video frame sampling strategy.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and Optimization. Our method is implemented based on PyTorch 2.3.1 and trained on 8 NVIDIA A100 GPUs. We maintain a global batch size of 128 during training. When training exclusively on GOT-10K, we train for 100 epochs. When joint training on multiple datasets, we extend to 170 epochs. In each epoch, we sample 131,072 sequences. We employ AdamW [35] as optimizer with learning rate of $10^{-4}$ and weight decay of 0.1. The learning rate scheduler warms up over the initial 2 epochs from $10^{-7}$ to $10^{-4}$ and then follows a cosine schedule to $10^{-6}$ .", + "bbox": [ + 511, + 227, + 905, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. We use a total of 3 reference frames, including the annotated first frame template. When current search frame index $t \\leq 3$ , we use all available tracked frames as reference frames, if there are fewer than three frames, we repeat the template as necessary to reach the required number. When $t > 3$ , we select two tracked frames as reference frames at indexes $\\left\\lfloor \\frac{t}{3} \\right\\rfloor$ and $\\left\\lfloor \\frac{2t}{3} \\right\\rfloor$ . Following previous trackers [4, 59, 61], we apply a Hanning window penalty to the output response map of the classification head.", + "bbox": [ + 511, + 378, + 905, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 511, + 527, + 844, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LaSOT [14] is a large-scale long-term tracking dataset with its test set containing 280 sequences. As shown in Table 2, SPMTrack-B significantly outperforms all trackers utilizing ViT-B as backbone. Our method also substantially surpasses LoRAT [33] that also employs parameter-efficient fine-tuning. Additionally, as shown in Figure 4, we evaluate tracker performance across various challenging scenarios in LaSOT. Our method surpasses previous trackers that also incorporate spatio-temporal context [2, 3, 6, 51, 56], while significantly outperforming ROMTrack [4] and GRM [20], which focus on optimizing relation modeling.", + "bbox": [ + 511, + 550, + 905, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GOT-10K [25] contains 9,335 sequences for training and 180 sequences for testing. We follow the protocol that trackers are only allowed to be trained using GOT-10K train split. Our approach significantly outperforms LoRAT [33] with the same backbone. And our SPMTrack-B rivals the performance of the current state-of-the-art methods, while SPMTrack-L substantially surpasses all existing trackers.", + "bbox": [ + 511, + 718, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TrackingNet [38] is a large-scale visual tracking dataset with a test set comprising 511 video sequences. Table 2 demonstrates that our method achieves state-of-the-art performance using only ViT-B as backbone, surpassing current trackers that utilize ViT-L or even ViT-G.", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c4a85a07b3c54a00ea7ba6f83527969b007bd68fbef911f0c69bc014147f126a.jpg", + "table_caption": [ + "Table 2. State-of-the-art comparison on LaSOT, GOT-10k and TrackingNet. \\* denotes for trackers trained only with GOT-10k train split. The best three results are highlighted in red, blue and bold, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSourceLaSOTGOT-10k*TrackingNet
AUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)AUC(%)PNorm(%)P(%)
SPMTrack-BOurs74.984.081.776.585.976.386.190.285.6
SPMTrack-LOurs76.885.984.080.089.479.986.991.087.2
SPMTrack-GOurs77.486.685.081.089.282.387.391.488.1
LoRAT-B378 [33]ECCV2472.981.979.173.782.672.984.288.483.0
LoRAT-L378 [33]ECCV2475.184.182.077.586.278.185.689.785.4
LoRAT-G378 [33]ECCV2476.285.383.578.987.880.786.090.286.1
AQATrack384 [56]CVPR2472.782.980.276.085.274.984.889.384.3
ARTrackV2-B384 [2]CVPR2473.082.079.677.586.075.585.789.885.5
ARTrackV2-L384 [2]CVPR2473.682.881.179.587.879.686.190.486.2
HIPTrack [3]CVPR2472.782.979.577.488.074.584.589.183.8
ODTrack-B [61]AAAI2473.283.280.677.087.975.185.190.184.9
F-BDMTrack-384 [58]ICCV2372.081.577.775.484.372.984.589.084.0
ROMTrack-384 [4]ICCV2371.481.478.274.284.372.484.189.083.7
ARTrack384 [51]CVPR2372.681.779.175.584.374.385.189.184.8
SeqTrack-B384 [6]CVPR2371.581.177.874.584.371.483.988.883.6
GRM [20]CVPR2369.979.375.873.482.970.484.088.783.3
OSTrack384 [59]ECCV2271.181.177.673.783.270.883.988.583.2
SimTrack [5]ECCV2270.579.7-69.878.866.083.487.4-
", + "bbox": [ + 91, + 128, + 911, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TNL2K [49] comprises 700 test video sequences, each accompanied by natural language descriptions. As shown in Table 3, our method achieves state-of-the-art performance without utilizing natural language descriptions and surpasses existing vision-language trackers [21, 30].", + "bbox": [ + 88, + 436, + 480, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UAV123 [37] is a dataset focusing on low-altitude drone aerial scenarios, consisting of 123 video sequences with an average length of 915 frames per sequence. Our method achieves state-of-the-art performance with comparable model sizes, demonstrating the capability of our approach in aerial scenes and small object tracking.", + "bbox": [ + 88, + 513, + 482, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NfS [29] consists of 100 video sequences totaling 380,000 frames. Following previous works, we conduct evaluations on the 30 FPS version of the dataset. Our method rivals current state-of-the-art trackers in NfS. One possible reason for not achieving further improvement is that our method does not incorporate consecutive target position like ARTrack [51] and HIPTrack [3], which is crucial for NfS.", + "bbox": [ + 88, + 604, + 482, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OTB2015 [54] is a classic dataset in visual tracking, consisting of 100 short-term tracking sequences that encompass 11 common challenges, such as target deformation, occlusion, and scale variation. Our method also achieves state-of-the-art performance on OTB2015.", + "bbox": [ + 88, + 710, + 482, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 801, + 243, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Importance of Spatio-Temporal Context Modeling and TMoE. Table 5 demonstrates the significance of two core designs in our paper: spatio-temporal context modeling and TMoE. In Table 5, removing TMoE indicates direct fine-tuning with LoRA [24], while removing spatio", + "bbox": [ + 88, + 824, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9baf13b17bc08d8893cff5fac2bb7c1c849891d1802b41ce7154c0e4565c515e.jpg", + "table_caption": [ + "Table 3. The performance of our method and other state-of-the-art trackers on TNL2K. The best three results are highlighted in red, blue and bold." + ], + "table_footnote": [], + "table_body": "
MethodAUC(%)PNorm(%)P(%)
SPMTrack-G64.782.670.6
SPMTrack-L63.781.569.2
SPMTrack-B62.079.766.7
LoRAT-B378 [33]59.9-63.7
ODTrack-L [61]61.7--
ARTrackV2-L384 [2]61.6--
ARTrack-L384 [51]60.3--
RTracker-L [26]60.6-63.7
UNINEXT-H [57]59.3-62.8
CiteTracker [30]57.7-59.6
VLT [21]53.1-53.3
", + "bbox": [ + 531, + 488, + 885, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "temporal modeling indicates using only the first frame template and search region as input and eliminating the target state token. Compared to LoRA, applying TMoE for parameter-efficient fine-tuning yields a substantial improvement of $+0.8$ AUC on LaSOT, demonstrates the superiority of TMoE. And the spatio-temporal context modeling further enhances performance by an additional $+1.2$ AUC. On the basis of incorporating spatio-temporal modeling, the improvement brought by TMoE is $+1.0$ AUC, indicating that TMoE can help the model better capture spatio-temporal contextual information.", + "bbox": [ + 509, + 688, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where to Apply TMoE. Conventional MoE is typically applied to the FFN layers within Transformer blocks, while keeping the attention layers unchanged. In contrast, we ap", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/774929da7c6c25f4145a472543c9c363e124168b2dafea5e572e6a053210cd1f.jpg", + "image_caption": [ + "Figure 4. The performance of our method compared with other state-of-the-art trackers in terms of AUC across various scenarios in the LaSOT test split." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 486, + 378 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f9ed2cbb13512b879840268da65c63e07e398a4a9ba4ad9bc1bb0bb3d7309f8.jpg", + "table_caption": [ + "Table 4. The performance of our method and other state-of-the-art trackers on UAV123, NfS and OTB2015 in terms of AUC metrics. The best three results are highlighted in red, blue and bold." + ], + "table_footnote": [], + "table_body": "
MethodUAV123NfSOTB2015
SPMTrack-B71.767.472.7
HIPTrack [3]70.568.171.0
ARTrackV2-B [2]69.967.6-
ODTrack-L [61]--72.4
ARTrack384 [51]70.566.8-
SeqTrack-B384 [6]68.666.7-
DropTrack [52]--69.6
MixFormer-L [10]69.5--
STMTrack [16]64.7-71.9
", + "bbox": [ + 106, + 500, + 468, + 647 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/162805f52d56196842e39b329316bef99c5965d97777f55456d537836b50c642.jpg", + "table_caption": [ + "Table 5. Ablation studies on spatio-temporal context modeling and TMoE. Experiments are conducted on LaSOT." + ], + "table_footnote": [], + "table_body": "
#Spatio-TemporalTMoEAUC (%)\\(P_{Norm}(\\%)\\)P(%)
1✘✘72.981.979.1
2✘✓73.782.780.0
3✓✘73.982.880.0
4✓✓74.984.081.7
", + "bbox": [ + 93, + 699, + 478, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ply TMoE in both the attention layers and the FFN layers. Table 6 demonstrates the impact of applying TMoE in different layers. Comparing the first, third, and fifth rows, applying TMoE to the query, key, value projections in attention layers and to FFN layers can both improve performance. Comparing the first, second and third rows, we find that applying TMoE in attention layers yields significantly", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "greater benefits compared to applying it only in FFN layers. Comparing the third and fourth rows, applying TMoE in the output projection matrix of the attention layer can bring a slight improvement.", + "bbox": [ + 511, + 90, + 906, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5c968126593e402b4ea874709f10c189e5e994081be5db81ded27a3dec165903.jpg", + "table_caption": [ + "Table 6. Ablation studies on applying TMoE in different layers. All results are evaluated on LaSOT test split. “X” means replacing TMoE with LoRA." + ], + "table_footnote": [], + "table_body": "
#QKVOut ProjFFNAUC (%)\\(P_{Norm}(\\%)\\)P(%)
1XXX73.982.880.0
2XX✓74.183.180.8
3✓XX74.583.681.2
4✓✓X74.683.881.4
5✓✓✓74.984.081.7
", + "bbox": [ + 516, + 218, + 903, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Whether to use a shared compress expert. In TMoE, we employ a shared compression expert to reduce the input dimension to $r$ , and use multiple routed experts for subsequent processing. As shown in the first and second rows of Table 7, compared to providing each routed expert with a compression expert, sharing the compression expert can achieve better performance while reducing the parameter count. When not sharing the compression expert, we provide t-SNE visualizations of the outputs from compression experts and routed experts in five TMoE modules in Figure 5. The outputs of different compression experts show significant overlaps, while the routed experts are more dispersed, validating the effectiveness of our approach.", + "bbox": [ + 511, + 315, + 906, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TMOE vs. Conventional MoE. We compare TMOE with conventional MoE on the visual tracking task. For conventional MoE, we use LoRA to fine-tune the attention layers and apply MoE in the FFN layers with five experts: one frozen shared expert that copies pretrained FFN parameters, and four trainable routed experts, all routed experts are identical to the original FFN architecture, and a router is also applied. As shown in the first row and third row of Table 7, TMOE demonstrates significant advantages over conventional MoE in both parameter efficiency and performance.", + "bbox": [ + 511, + 511, + 908, + 664 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a43817f7916e339c20c9c2d09af726efd92e0addf33a26db0227dcce760838c.jpg", + "table_caption": [ + "Table 7. Ablation studies on whether to share the compression expert and use the conventional MoE to replace TMoE. Results are evaluated on LaSOT test split." + ], + "table_footnote": [], + "table_body": "
model variants#Params(M)AUC (%)PNorm(%)P(%)
SPMTrack-B115.374.984.081.7
w/o shared compress expert131.374.883.781.5
Conventional MoE316.873.482.679.8
", + "bbox": [ + 514, + 729, + 911, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Number of Routed Experts. In Table 8, we evaluated the impact of different numbers of routed experts in TMoE. Results indicate that increasing the number of experts leads to improved performance, highlighting the potential of TMoE as a method to expand the parameter count and model scale. To balance the number of parameters and", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16877", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3f622e8bba3cf50600e59b75077de5b8c4165f72354cf7066a46b048b1c2ac0.jpg", + "image_caption": [ + "Figure 5. Comparison of t-SNE visualizations. Each column shows outputs from all compression experts (top) and routed experts (bottom) within a TMoE module. Different colors represent distinct experts. Zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/996fd35ab80d4aa31c79ad29e23b61b958d7985cd7700cabcfac93ffd8e49497.jpg", + "image_caption": [ + "Figure 6. Visualization comparison of search region attention maps with and without TMoE. Zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 94, + 272, + 480, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance, we choose to use 4 experts, and whether there are separate optimal configurations for the TMoE in the attention layer and the FFN layer remains to be explored.", + "bbox": [ + 88, + 582, + 482, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Number of Reference Frames. In SPMTrack, we use the template and several tracked frames as reference frames to enhance the spatio-temporal modeling ability. In Table 9, we explore the impact of using different numbers of reference frames during training and inference. During inference, if only two reference frames are used, we simply select the frame at $\\left\\lfloor \\frac{t}{2} \\right\\rfloor$ as the reference frame. If four reference frames are used, we select the frames at $\\left\\lfloor \\frac{t}{4} \\right\\rfloor$ , $\\left\\lfloor \\frac{t}{2} \\right\\rfloor$ , and $\\left\\lfloor \\frac{3 \\times t}{4} \\right\\rfloor$ . As shown in Table 9, better performance can be achieved when the number of reference frames used during inference remains the same as that during training. Training with more reference frames can also lead to improved performance, illustrating that our method still has substantial potential for performance enhancement.", + "bbox": [ + 88, + 628, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of attention maps with and without TMoE. In Figure 6, to exclude the influence of intermediate tracked frames, we use the model in rows 1 and 2 of Table 5 for comparison, where spatio-temporal modeling is removed", + "bbox": [ + 88, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd5731af5bf0e702a0a021689bdadc34094fdf97e687c346e4f270ef99407a6a.jpg", + "table_caption": [ + "Table 8. The performance of our method on the test split of GOT-10K when setting different number of routed experts in TMoE." + ], + "table_footnote": [], + "table_body": "
Number2468
AO(%)75.376.576.677.1
SR0.5(%)84.485.986.386.4
SR0.75(%)75.276.377.077.6
", + "bbox": [ + 514, + 128, + 900, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a6b78ce3648fada27fe1a8b80694694d7da8fbada265a9ed71dfc30c92aec8de.jpg", + "table_caption": [ + "Table 9. Ablation study on the number of reference frames during training and inference. Results are evaluated based on SPMTrack-B on GOT-10K test split." + ], + "table_footnote": [], + "table_body": "
TrainingInferenceAO (%)SR0.5(%)SR0.75(%)
2275.885.175.3
374.784.673.8
469.881.866.2
3273.183.972.7
376.585.976.3
472.883.471.3
4270.683.968.9
374.685.474.5
477.587.377.2
", + "bbox": [ + 521, + 260, + 893, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and we focus solely on the impact of TMoE. We compare the attention maps of the search region in the last block of Transformer encoder. The first column is the template, the second column displays search region, the third column is the visualization result and pure attention map, as well as the fourth column. Figure 6 shows that when facing challenging scenarios such as occlusions, distractors, and viewpoint variations, TMoE effectively suppresses background regions while enhancing perception of the target boundary.", + "bbox": [ + 511, + 449, + 906, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 602, + 633, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present TMoE, a mixture of experts module tailored for visual tracking, and propose SPMTrack, a novel tracker enabling spatio-temporal context modeling based on TMoE. TMoE is applied in the linear layers in both self-attention and FFN layers, enhancing the diversity and flexibility of expert combinations to better handle various relation modeling in visual tracking. Additionally, TMoE employs a lightweight and efficient structure and serves as a method of parameter-efficient fine-tuning, which enables us to train SPMTrack of larger scales and enables SPMTrack to achieve state-of-the-art performance with only a small subset of parameters need to be trained. Furthermore, we hope that this work will inspire more applications of mixture of experts in the field of visual tracking.", + "bbox": [ + 511, + 628, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This paper was supported by Zhejiang Provincial Natural Science Foundation of China under Grant No.LD24F020016 and National Natural Science Foundation of China under Grant No.62176017.", + "bbox": [ + 511, + 839, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompt autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19048-19057, 2024. 2, 5, 6, 7", + "[3] Wenrui Cai, Qingjie Liu, and Yunhong Wang. Hiptrack: Visual tracking with historical prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19258-19267, 2024. 1, 2, 3, 5, 6, 7", + "[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9589-9600, 2023. 1, 2, 5, 6", + "[5] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: a simplified architecture for visual object tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 375-392. Springer, 2022. 1, 2, 6", + "[6] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14572-14581, 2023. 2, 5, 6, 7", + "[7] Yucheng Chen and Lin Wang. emoe-tracker: Environmental moe-based transformer for robust event-guided object tracking. arXiv preprint arXiv:2406.20024, 2024. 2", + "[8] Mohammed Nowaz Rabbani Chowdhury, Shuai Zhang, Meng Wang, Sijia Liu, and Pin-Yu Chen. Patch-level routing in mixture-of-experts is provably sample-efficient for convolutional neural networks. In Proceedings of the 40th International Conference on Machine Learning, pages 6074-6114. PMLR, 2023. 2", + "[9] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International conference on machine learning, pages 4057-4086. PMLR, 2022. 2", + "[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13608-13618, 2022. 1, 2, 7", + "[11] Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguis" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tics (Volume 1: Long Papers), pages 1280-1297, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5", + "[13] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yangi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient scaling of language models with mixture-of-experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 2", + "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 2, 5", + "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2", + "[16] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. Stmtrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13774-13783, 2021. 7", + "[17] Zhihong Fu, Zehua Fu, Qingjie Liu, Wenrui Cai, and Yunhong Wang. Sparsett: Visual tracking with sparse transformers. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-22, pages 905–912, 2022. 1", + "[18] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3", + "[19] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 146-164. Springer, 2022. 1", + "[20] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18686-18695, 2023. 1, 2, 5, 6", + "[21] Mingzhe Guo, Zhipeng Zhang, Heng Fan, and Liping Jing. Divert more attention to vision-language tracking. In Advances in Neural Information Processing Systems, pages 4446-4460. Curran Associates, Inc., 2022. 6", + "[22] Kaijie He, Canlong Zhang, Sheng Xie, Zhixin Li, and Zhiwen Wang. Target-aware tracking with long-term context" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16879", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attention. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):773-780, 2023. 2", + "[23] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3", + "[24] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR, 2022. 3, 6", + "[25] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5", + "[26] Yuqing Huang, Xin Li, Zikun Zhou, Yaowei Wang, Zhenyu He, and Ming-Hsuan Yang. Rtracker: Recoverable tracking via pn tree structured memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19038-19047, 2024. 6", + "[27] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2", + "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3", + "[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6", + "[30] Xin Li, Yuqing Huang, Zhenyu He, Yaowei Wang, Huchuan Lu, and Ming-Hsuan Yang. Citetracker: Correlating image and text for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9974-9983, 2023. 6", + "[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 3", + "[32] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. Advances in Neural Information Processing Systems, 35:16743-16754, 2022. 3", + "[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. In Computer Vision - ECCV 2024, pages 300-318, Cham, 2025. Springer Nature Switzerland. 1, 3, 4, 5, 6", + "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR, 2019. 5", + "[36] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2", + "[37] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, pages 445-461, 2016. 6", + "[38] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, pages 300-317, 2018. 2, 5", + "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 5", + "[40] Jonas Pfeiffer, Andreas Rücklé, Clifton Poth, Aishwarya Kamath, Ivan Vulić, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. Adapterhub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, 2020. 3", + "[41] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2", + "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5", + "[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems, pages 8583-8595. Curran Associates, Inc., 2021. 2", + "[44] Zhangyong Tang, Tianyang Xu, Zhenhua Feng, Xuefeng Zhu, He Wang, Pengcheng Shao, Chunyang Cheng, Xiao-Jun Wu, Muhammad Awais, Sara Atito, et al. Revisiting rgbt tracking benchmarks from the perspective of modality validity: A new benchmark, problem, and method. arXiv preprint arXiv:2405.00168, 2024. 2", + "[45] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. Hydralora: An asymmetric lora architecture for efficient fine-tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 1", + "[47] Guangting Wang, Chong Luo, Xiaoyan Sun, Zhiwei Xiong," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16880", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Wenjun Zeng. Tracking by instance detection: A meta-learning approach. In CVPR, pages 6288-6297, 2020. 5", + "[48] Ke Wang, Jiahui Zhu, Minjie Ren, Zeming Liu, Shiwei Li, Zongye Zhang, Chenkai Zhang, Xiaoyu Wu, Qiqi Zhan, Qingjie Liu, et al. A survey on data synthesis and augmentation for large language models. arXiv preprint arXiv:2410.12896, 2024. 2", + "[49] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13763-13773, 2021. 2, 6", + "[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3", + "[51] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9697-9706, 2023. 2, 5, 6, 7", + "[52] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, 2023. 1, 7", + "[53] Xun Wu, Shaohan Huang, and Furu Wei. Mixture of lora experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2", + "[54] Y. Wu, J. Lim, and M. Yang. Object tracking benchmark. TPAMI, 37(9):1834-1848, 2015. 6", + "[55] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8751-8760, 2022. 2", + "[56] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19300-19309, 2024. 5, 6", + "[57] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6", + "[58] Dawei Yang, Jianfeng He, Yinchao Ma, Qianjin Yu, and Tianzhu Zhang. Foreground-background distribution modeling transformer for visual object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10117-10127, 2023. 1, 2, 6", + "[59] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 1, 2, 5, 6", + "[60] Yihua Zhang, Ruisi Cai, Tianlong Chen, Guanhua Zhang, Huan Zhang, Pin-Yu Chen, Shiyu Chang, Zhangyang Wang, and Sijia Liu. Robust mixture-of-expert training for convolutional neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 90-101, 2023. 2", + "[61] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7588-7596, 2024. 1, 2, 3, 5, 6, 7", + "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + ], + "bbox": [ + 516, + 92, + 906, + 343 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16881", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_model.json b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d0c9b3179cedd9ff59ea80718de1df1a2e1ad135 --- /dev/null +++ b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.13, + 0.842, + 0.178 + ], + "angle": 0, + "content": "SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.91, + 0.295 + ], + "angle": 0, + "content": "Wenrui Cai\\(^{1}\\), Qingjie Liu\\(^{1,2,3,*}\\), Yunhong Wang\\(^{1,3}\\) \n\\(^{1}\\)State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China \n\\(^{2}\\)Zhongguancun Laboratory, Beijing, China \n\\(^{3}\\)Hangzhou Innovation Institute, Beihang University, Hangzhou, China \n{wenrui_cai, qingjie.liu, yhwang}@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.328, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.486, + 0.754 + ], + "angle": 0, + "content": "Most state-of-the-art trackers adopt one-stream paradigm, using a single Vision Transformer for joint feature extraction and relation modeling of template and search region images. However, relation modeling between different image patches exhibits significant variations. For instance, background regions dominated by target-irrelevant information require reduced attention allocation, while foreground, particularly boundary areas, need to be emphasized. A single model may not effectively handle all kinds of relation modeling simultaneously. In this paper, we propose a novel tracker called SPMTrack based on mixture-of-experts tailored for visual tracking task (TMoE), combining the capability of multiple experts to handle diverse relation modeling more flexibly. Benefiting from TMoE, we extend relation modeling from image pairs to spatio-temporal context, further improving tracking accuracy with minimal increase in model parameters. Moreover, we employ TMoE as a parameter-efficient fine-tuning method, substantially reducing trainable parameters, which enables us to train SPMTrack of varying scales efficiently and preserve the generalization ability of pretrained models to achieve superior performance. We conduct experiments on seven datasets, and experimental results demonstrate that our method significantly outperforms current state-of-the-art trackers. The source code is available at https://github.com/WenRuiCai/SPMTrack." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.786, + 0.224, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.812, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Visual tracking aims to predict the location of a target throughout subsequent frames of a video, given the initial state in the first frame template. Currently, most state-of-the-art trackers adopt a Transformer-based [46] one-stream" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.325, + 0.902, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.568, + 0.908, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Comparison of LaSOT AUC and model parameter count across different trackers. Larger loop indicates better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.703 + ], + "angle": 0, + "content": "paradigm [3, 5, 10, 33, 52, 59, 61], utilizing a single Vision Transformer [12] backbone that accepts both the template and the search region image as input, and employs self-attention to perform image feature extraction and template-search region relation modeling simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.884 + ], + "angle": 0, + "content": "However, not all relation modeling between image patches benefits tracking performance. Prior studies [4, 20, 58, 59] have demonstrated that the abundance of background patches in the search region deteriorates the discriminative power of the foreground features, and single vanilla attention has limited capability in suppressing undesirable foreground-background interactions. Meanwhile, there are also some kinds of relation modeling that demonstrate positive effects on tracking performance [17, 19]. For instance, enhancing attention to target boundary regions can improve tracking performance [17], especially in challenging scenarios such as motion blur, partial occlusion, and deformation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To address various relation modeling, numerous meth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.245, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16871" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "ods propose specific designs, including background filtering to reduce target-irrelevant information [3, 59], customizing different interaction modules for tokens belonging to different categories [4, 20], and attention score adjustment based on foreground-background weights of each token [58]. However, existing methods can only handle one or a few predefined types of relation modeling, and the handcrafted strategies inherently lack adaptability. Furthermore, there may exist many latent relationships that further complicate the specialized design. These limitations motivate our question: How to design a tracker to adaptively process various relation modeling between image patches?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.278, + 0.485, + 0.655 + ], + "angle": 0, + "content": "Inspired by mixture of experts (MoE) [1, 11, 13, 27, 41] in natural language processing, in this paper, we address the question by proposing SPMTrack, a novel tracker based on TMoE, a specialized mixture of experts module for visual tracking task. Similar to MoE, TMoE handles diverse relation modeling through adaptive weighted combinations of multiple experts. However, unlike traditional MoE that is exclusively applied to feed-forward network (FFN) layers, TMoE is simultaneously applied to the linear layers within both the attention layers and the FFN layers. This finer-grained application enables TMoE to achieve more diverse expert combinations, thereby enhancing the capability of various relation modeling. Moreover, differing from traditional MoE that uses FFNs as experts, all the experts in TMoE are linear layers and designed with a lightweight and efficient structure to ensure overall efficiency. Furthermore, we employ TMoE as a parameter-efficient finetuning method, which enables us to train SPMTrack of larger scales. In SPMTrack, only a subset of experts and the router within TMoE, along with the prediction head need to be trained, which substantially reduces the trainable parameters while preserving the generalization capability of pretrained models. Notably, our method reduces the number of trainable parameters by nearly \\(80\\%\\) compared to [2], while demonstrates better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Benefiting from the powerful capability of TMoE in handling diverse relation modeling, unlike previous one-stream trackers that only perform relation modeling between template-search region pairs, we extend SPMTrack to incorporate multi-frame spatio-temporal context modeling, which further enhances the performance of SPMTrack with minimal additional parameters. We evaluated the performance of our method on seven datasets, and the experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets including LaSOT [14], GOT-10K [25], TrackingNet [38] and TNL2K [49]. Through extensive experiments with models of varying scales, as shown in Figure 1, Table 1 and Table 2, with ViT-B [12] as backbone and less than \\(30\\%\\) of parameters need to be trained, SPMTrack-B achieves or even surpasses the best performance of previous trackers utilizing ViT-L." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.273 + ], + "angle": 0, + "content": "The main contributions of this paper can be summarized as: (1) We propose TMoE, a mixture of experts module tailored for visual tracking. TMoE enhances the capability to handle various relation modeling and can be used as a method of parameter-efficient fine-tuning. (2) Based on TMoE, we propose SPMTrack, a novel tracker that can effectively integrate spatio-temporal context for visual tracking. (3) Experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets. We trained SPMTrack of varying scales, and our method with ViT-B as backbone can achieve or even surpass current trackers using ViT-L." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.655, + 0.305 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.719, + 0.33 + ], + "angle": 0, + "content": "2.1. One-Stream Trackers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.909, + 0.626 + ], + "angle": 0, + "content": "One-stream trackers employ a single Vision Transformer [12] as backbone, simultaneously performing feature extraction and relation modeling at each layer [5, 10, 55, 59]. Many efforts focus on incorporating historical context. TATrack [22] and ODTrack [61] propose architectures capable of processing multiple frames, while HIPTrack [3] introduces historical target features through prompt learning. SeqTrack [6], ARTrack[51], and ARTrackV2 [2] achieve more precise predictions by incorporating multiple historical target position. Other works try to improve conventional self-attention to handle different relation modeling. OSTrack [59] progressively filters background regions to reduce foreground-background relation modeling, GRM [20] and ROMTrack [4] categorizes image tokens and constrains relation modeling between specific categories, F-BDMTrack [58] adjusts the attention score by calculating the foreground-background weights of tokens. But these methods can only handle specific relation modeling and rely on manually designed strategies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.698, + 0.653 + ], + "angle": 0, + "content": "2.2. Mixture of Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Mixture-of-experts (MoE) [1, 13, 27] are widely applied in large language models [9, 11, 15, 36, 41, 53], where each expert focuses on specific aspects of the data or particular tasks [48]. By performing an adaptive weighted sum over multiple experts, MoE can better capture complex patterns and relationships within the input data. MoE also has applications in computer vision [8, 43, 60]. However, there is few dedicated exploration of MoE in visual tracking. MoETrack [44] employs multiple prediction heads as experts in the field of RGB-T tracking, which is rather far-fetched and deviates from the general practice of embedding MoE within Transformer layers. In the field of RGB-E tracking, eMoETracker [7] inserts multiple experts between Transformer layers, which more closely resembles an adapter architecture [23] and contrasts with MoE implementations where experts are embedded within Transformer layers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.9, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 2. Overview of SPMTrack that consists of a feature extraction network and a prediction head. The main body of the feature extraction network is a Transformer encoder composed of multiple TMoEBlocks. The structure of TMoEBlock is shown on the right side." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.373, + 0.451 + ], + "angle": 0, + "content": "2.3. Parameter-efficient Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Parameter-efficient fine-tuning aims to reduce computational costs and retain the general knowledge of the pretrained model by freezing most parameters and fine-tuning only a small subset of parameters. Common parameter-efficient fine-tuning approaches include adapter-based methods [18, 23, 40], prompt-based methods [28, 31, 50, 62] and low-rank adaptation methods [24, 45]. In visual tracking, HIPTrack [3] incorporates historical target features as prompts, while LoRAT [33] employs low-rank adaptation for efficient training. And our method maintains flexibility in diverse relation modeling while reducing trainable parameters by fine-tuning TMoE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.184, + 0.67 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.288, + 0.696 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.795 + ], + "angle": 0, + "content": "As illustrated in Figure 2, we propose SPMTrack, a novel tracker that integrates multi-frame spatio-temporal context and employs TMoE for parameter-efficient fine-tuning. The overall architecture follows the one-stream paradigm, consisting of a Feature Extraction Network based on a Vision Transformer [12] as backbone, and a prediction head." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Feature Extraction Network. The feature extraction network is denoted as \\(\\Phi\\). The main body of the feature extraction network is a Transformer encoder that utilizes TMoE-Block as its primary component, which is designed based on TMoE. The architecture of TMoE-Block is illustrated in the right panel of Figure 2, and details will be introduced in Section 3.2. The encoder simultaneously per" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.737 + ], + "angle": 0, + "content": "forms feature extraction and relation modeling for input images. We extend the input from one single template \\( \\mathbf{Z} \\in \\mathbb{R}^{H_z \\times W_z \\times 3} \\) that is commonly used in one-stream trackers to the video sequence \\( \\{Z_1, \\dots, Z_N | Z_i \\in \\mathbb{R}^{H_z \\times W_z \\times 3}\\} \\) as reference frames. The reference frames contain the first frame template and the selected tracked historical frames, with a total of \\( N \\) frames. The feature extraction network uses patch embedding to partition each input image into patches of size \\( M \\times M \\), map all patches into tokens by a convolutional layer, and obtain the reference token sequence of each image \\( T_i \\in \\mathbb{R}^{N_T \\times d} \\), where \\( d \\) denotes the hidden dimension of Transformer encoder and \\( N_T = \\frac{H_z W_z}{M^2} \\) represents the number of tokens corresponding to each image. Similarly, for the search region \\( S \\in \\mathbb{R}^{H_s \\times W_s \\times 3} \\), we can also obtain the token sequence \\( X \\in \\mathbb{R}^{N_X \\times d} \\). All tokens are added with positional embedding and a learnable token type embedding [33]. Inspired by [32, 61], we additionally introduce a target state token \\( H \\in \\mathbb{R}^{1 \\times d} \\) that is learnable as well and propagates over time. All tokens are concatenated and fed into the Transformer encoder. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.75, + 0.906, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} _ {i} ^ {j} = \\left\\{ \\begin{array}{l} \\boldsymbol {T} _ {i} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {o}, i f \\text {t o k e n} (i, j) \\text {i n b b o x} \\\\ \\boldsymbol {T} _ {i} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {b}, o t h e r w i s e \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.792, + 0.739, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {j} = \\boldsymbol {X} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {S}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.809, + 0.77, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} = \\operatorname {C o n c a t} (\\boldsymbol {H}, \\boldsymbol {T} _ {1}, \\dots , \\boldsymbol {T} _ {N}, \\boldsymbol {X})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( T_{i}^{j} \\) represents the \\( j^{th} \\) token in \\( T_{i} \\), \\( PE_{j} \\) represents positional embedding of the \\( j^{th} \\) token, similarly for \\( X^{j} \\). \\( TE_{o}, TE_{b} \\) and \\( TE_{S} \\) represents the learnable type embeddings of the foreground region tokens, background region" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16873" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "tokens in reference frames and the search region tokens, respectively. \\( \\pmb{I} \\) denotes the input to the Transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.227 + ], + "angle": 0, + "content": "Subsequently, the Transformer encoder processes all tokens and produces an output \\( \\pmb{O} \\in \\mathbb{R}^{(1 + N \\times N_T + N_X) \\times d} \\). The output tokens \\( \\pmb{X}' \\in \\mathbb{R}^{N_X \\times d} \\) corresponding to search region and the output \\( \\pmb{H}' \\in \\mathbb{R}^{1 \\times d} \\) from the target state token are fed into the prediction head for prediction. \\( \\pmb{H}' \\) will be added with the input target state token in the next tracking frame, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.236, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {O} = \\Phi (\\boldsymbol {I}), [ \\boldsymbol {H} ^ {\\prime}, \\boldsymbol {T} _ {1} ^ {\\prime}, \\dots , \\boldsymbol {T} _ {N} ^ {\\prime}, \\boldsymbol {X} ^ {\\prime} ] = \\operatorname {S p l i t} (\\boldsymbol {O}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.254, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Prediction Head. The input to the prediction head consists of the output target state token \\( H' \\) and the output search region tokens \\( X' \\) from the feature extraction network. Initially, the prediction head performs matrix multiplication between \\( H' \\) and \\( X' \\) to obtain the weight \\( U \\in \\mathbb{R}^{N_X \\times 1} \\). As the target state token propagates over time, \\( U \\) can reweight and adjust current search region tokens based on the historical state of the target, thereby introducing more spatio-temporal tracking context. After element-wise multiplication of \\( X' \\) with \\( U \\), the resulting data is reshaped into a feature map \\( F \\in \\mathbb{R}^{\\frac{H_s}{M} \\times \\frac{W_s}{M} \\times d} \\), which is then processed by a decoupled double-MLP head to perform target center classification and box regression, respectively. For fair comparison, the double-MLP head is identical to [33]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.475, + 0.332, + 0.49 + ], + "angle": 0, + "content": "3.2. The Design of TMoEBlock" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.664 + ], + "angle": 0, + "content": "TMoEBlock serves as the primary component of the Transformer encoder in the feature extraction network. As illustrated in the right of Figure 2, TMoEBlock is built based on the standard vision Transformer block. However, unlike the standard ViT block, TMoEBlock applies TMoE module to both the multi-head self-attention (MSA) layer and the feed-forward network (FFN) layer, obtaining MoE-based MSA \\((\\mathrm{M}^2\\mathrm{SA})\\) layer and MoE-based FFN (MFFN) layer. The finer-grained application of TMoE is capable to achieve more diverse combinations for various relation modeling. The processing of TMoEBlock can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.681, + 0.483, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {O} _ {l} ^ {\\prime} = \\mathrm {M} ^ {2} \\operatorname {S A} (\\ln (\\boldsymbol {O} _ {l - 1})) + \\boldsymbol {O} _ {l - 1}, \\quad l = 1, \\dots , L \\tag {3} \\\\ \\boldsymbol {O} _ {l} = \\mathrm {M F F N} (\\mathrm {L N} (\\boldsymbol {O} _ {l} ^ {\\prime})) + \\boldsymbol {O} _ {l} ^ {\\prime}, l = 1, \\dots , L \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where \\(L\\) denotes the total number of TMoEBlocks, \\(\\pmb{O}_l\\) represents the output of the \\(l^{th}\\) block in Transformer encoder. When \\(l = L\\), \\(\\pmb{O}_l\\) is the output \\(\\pmb{O}\\) of the feature extraction network; when \\(l = 0\\), it is the overall input \\(\\pmb{I}\\). LN stands for the layer normalization operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.87 + ], + "angle": 0, + "content": "MoE-based Multi-head Self-Attention (M²SA). M²SA not only replaces the standard three linear projection layers \\( Q, K, V \\) in standard attention for the query, key, and value with TMoE modules but also replaces the output projection layer with TMoE module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "MoE-based Feed-Forward Network (MFFN). Conventional feed-forward network consists of two linear layer and" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.898, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Figure 3. The structure of TMoE. The symbols maintain the same meaning with Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.338 + ], + "angle": 0, + "content": "activation function. MFFN replaces both linear layers in FFN with TMoE modules, while maintaining all other architectural configurations unchanged." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.346, + 0.851, + 0.363 + ], + "angle": 0, + "content": "3.3. Mixture of Experts for Visual Tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.906, + 0.519 + ], + "angle": 0, + "content": "In this section, we introduce TMoE, a mixture of experts module tailored for visual tracking task. As illustrated in Figure 3, TMoE consists of one shared expert, one router, one compression expert and \\( N_{e} \\) routed experts. In large language models, mixture of experts are commonly applied to replace FFN layers, and the experts comprising MoE are typically the same structure as FFN. However, TMoE is used to replace the linear layers in both self-attention and FFN, and the shared expert, the compression expert and the routed experts are all implemented as linear layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.624 + ], + "angle": 0, + "content": "In TMoE, the shared expert directly copies weights from the corresponding linear layers in pretrained model and remains frozen during training. The rationale is that the original weights have been well-trained during the pre-training stage, and the shared expert needs to retain more general knowledge. In contrast, the compression experts, routed experts, and the router are all trainable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Given a TMOE module replacing a linear layer with input dimension \\(d\\) and output dimension \\(D\\), for each token \\(\\pmb{x} \\in \\mathbb{R}^d\\), as shown in Figure 3, TMOE first computes the weights \\(\\pmb{W} \\in \\mathbb{R}^{N_e}\\) of all routed experts based on the input \\(\\pmb{x}\\), the weights are normalized by Softmax. After that, TMOE calculates the output \\(\\pmb{y}^s \\in \\mathbb{R}^D\\) of the shared expert and the outputs \\(\\pmb{y}_i^e \\in \\mathbb{R}^D\\) of all routed experts. In particular, instead of adopting the same structure as the shared expert, inspired by [45], we first employ a compression expert with a \\(d \\times r\\) linear transformation to reduce the input dimension from \\(d\\) to \\(r\\) and obtain the compressed result \\(\\pmb{y}^c \\in \\mathbb{R}^r\\), where \\(r \\ll d\\) and \\(r \\ll D\\). Subsequently, all routed experts use the compressed input \\(\\pmb{y}^c\\) to produce their respective outputs \\(\\pmb{y}_i^e \\in \\mathbb{R}^D\\). The process can be formally described as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.847, + 0.85, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {W} = \\text {S o f t m a x} (\\operatorname {R o u t e r} (\\boldsymbol {x})), \\quad \\boldsymbol {y} ^ {s} = \\boldsymbol {E} ^ {S} (\\boldsymbol {x}) \\\\ \\boldsymbol {y} ^ {c} = \\boldsymbol {E} ^ {C} (\\boldsymbol {x}), \\quad \\boldsymbol {y} _ {i} ^ {e} = \\boldsymbol {E} _ {i} ^ {R} (\\boldsymbol {y} ^ {c}) \\quad i = 1, \\dots , N _ {e} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where Router is also a linear layer that maps the input \\( x \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16874" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.121 + ], + "angle": 0, + "content": "to weights \\(\\pmb{W}\\), and \\(\\pmb{E}^{S}\\), \\(\\pmb{E}^{C}\\), and \\(\\pmb{E}_{i}^{R}\\) denote the shared expert, compress expert, and \\(i^{th}\\) routed expert, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.213 + ], + "angle": 0, + "content": "After obtaining the outputs of all experts, we compute a weighted sum of the outputs \\( \\pmb{y}_i^e \\) from all routed expert based on \\( \\pmb{W} \\) to obtain \\( \\pmb{y}^e \\in \\mathbb{R}^D \\), which means that the routed experts are densely activated in TMoE. Subsequently, \\( \\pmb{y}^e \\) is added with the shared expert output \\( \\pmb{y}^s \\) to get the final output \\( \\pmb{y} \\in \\mathbb{R}^D \\). The process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.224, + 0.483, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} ^ {e} = \\sum_ {i = 1} ^ {N _ {e}} \\boldsymbol {W} _ {i} \\boldsymbol {y} _ {i} ^ {e}, \\quad \\boldsymbol {y} = \\boldsymbol {y} ^ {e} + \\boldsymbol {y} ^ {s} \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.268, + 0.236, + 0.283 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.368 + ], + "angle": 0, + "content": "The outputs of the decoupled double-MLP prediction head are supervised using binary cross-entropy loss for target center classification and Generalized IoU loss [42] for bounding box regression, respectively. Both loss terms are assigned with the weighting coefficients set to 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.379, + 0.224, + 0.396 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.403, + 0.308, + 0.419 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Model settings. We provide three versions of SPMTrack: SPMTrack-B, SPMTrack-L, and SPMTrack-G, utilizing ViT-B [12], ViT-L [12], and ViT-G [39] as backbones respectively. All versions maintain a consistent patch size \\( M \\) of 14 and adopt the pretrained weights from DINOv2 [39]. The resolution of input images are the same across all versions, with reference frames of \\( 196 \\times 196 \\) and search region of \\( 378 \\times 378 \\). A crop factor of 2 is applied to reference frames and 5 to search region. We set the number of reference frames \\( N \\) to 3 for spatio-temporal modeling. Each TMoE module incorporates 4 routed experts, where the input dimension of the routed experts \\( r \\) is set to 64. Other configurations and model information are shown in Table 1, where \\( L \\) is the number of TMoEBlocks, \\( d \\) denotes hidden dimension and \\( N_h \\) is the number of attention heads." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Table 1. Model configurations and parameter counts across SPM-Track variants of different scales." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.703, + 0.48, + 0.811 + ], + "angle": 0, + "content": "
SPMTrack-BSPMTrack-LSPMTrack-G
BackboneL=12L=24L=40
d=768d=1024d=1536
Nh=12Nh=16Nh=24
#Params (M)115.3379.61339.5
#Trainable \nParams (M)29.275.9204.0
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. Following previous works [3, 33, 51, 59], we use LaSOT [14], TrackingNet [47], GOT-10K [25], and COCO [34] for training. When evaluating on GOT-10K test set, we strictly follow the protocol by training exclusively on GOT-10K training set. When evaluating on other datasets," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "we jointly train on the training sets of all four datasets, randomly sampling from each dataset with equal probability. For video datasets, we sample 5 frames from one video with random frame intervals ranging from 1 to 200. The sampled frames are arranged either in forward or reverse order with equal probability, where the first 3 frames serve as reference frames and the remaining 2 as search frames. For COCO, we replicate each image 5 times to maintain consistency with the video frame sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.906, + 0.379 + ], + "angle": 0, + "content": "Training and Optimization. Our method is implemented based on PyTorch 2.3.1 and trained on 8 NVIDIA A100 GPUs. We maintain a global batch size of 128 during training. When training exclusively on GOT-10K, we train for 100 epochs. When joint training on multiple datasets, we extend to 170 epochs. In each epoch, we sample 131,072 sequences. We employ AdamW [35] as optimizer with learning rate of \\(10^{-4}\\) and weight decay of 0.1. The learning rate scheduler warms up over the initial 2 epochs from \\(10^{-7}\\) to \\(10^{-4}\\) and then follows a cosine schedule to \\(10^{-6}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.517 + ], + "angle": 0, + "content": "Inference. We use a total of 3 reference frames, including the annotated first frame template. When current search frame index \\( t \\leq 3 \\), we use all available tracked frames as reference frames, if there are fewer than three frames, we repeat the template as necessary to reach the required number. When \\( t > 3 \\), we select two tracked frames as reference frames at indexes \\( \\left\\lfloor \\frac{t}{3} \\right\\rfloor \\) and \\( \\left\\lfloor \\frac{2t}{3} \\right\\rfloor \\). Following previous trackers [4, 59, 61], we apply a Hanning window penalty to the output response map of the classification head." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.845, + 0.544 + ], + "angle": 0, + "content": "4.2. Comparisons with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.718 + ], + "angle": 0, + "content": "LaSOT [14] is a large-scale long-term tracking dataset with its test set containing 280 sequences. As shown in Table 2, SPMTrack-B significantly outperforms all trackers utilizing ViT-B as backbone. Our method also substantially surpasses LoRAT [33] that also employs parameter-efficient fine-tuning. Additionally, as shown in Figure 4, we evaluate tracker performance across various challenging scenarios in LaSOT. Our method surpasses previous trackers that also incorporate spatio-temporal context [2, 3, 6, 51, 56], while significantly outperforming ROMTrack [4] and GRM [20], which focus on optimizing relation modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.825 + ], + "angle": 0, + "content": "GOT-10K [25] contains 9,335 sequences for training and 180 sequences for testing. We follow the protocol that trackers are only allowed to be trained using GOT-10K train split. Our approach significantly outperforms LoRAT [33] with the same backbone. And our SPMTrack-B rivals the performance of the current state-of-the-art methods, while SPMTrack-L substantially surpasses all existing trackers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "TrackingNet [38] is a large-scale visual tracking dataset with a test set comprising 511 video sequences. Table 2 demonstrates that our method achieves state-of-the-art performance using only ViT-B as backbone, surpassing current trackers that utilize ViT-L or even ViT-G." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16875" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. State-of-the-art comparison on LaSOT, GOT-10k and TrackingNet. \\* denotes for trackers trained only with GOT-10k train split. The best three results are highlighted in red, blue and bold, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.129, + 0.912, + 0.421 + ], + "angle": 0, + "content": "
MethodSourceLaSOTGOT-10k*TrackingNet
AUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)AUC(%)PNorm(%)P(%)
SPMTrack-BOurs74.984.081.776.585.976.386.190.285.6
SPMTrack-LOurs76.885.984.080.089.479.986.991.087.2
SPMTrack-GOurs77.486.685.081.089.282.387.391.488.1
LoRAT-B378 [33]ECCV2472.981.979.173.782.672.984.288.483.0
LoRAT-L378 [33]ECCV2475.184.182.077.586.278.185.689.785.4
LoRAT-G378 [33]ECCV2476.285.383.578.987.880.786.090.286.1
AQATrack384 [56]CVPR2472.782.980.276.085.274.984.889.384.3
ARTrackV2-B384 [2]CVPR2473.082.079.677.586.075.585.789.885.5
ARTrackV2-L384 [2]CVPR2473.682.881.179.587.879.686.190.486.2
HIPTrack [3]CVPR2472.782.979.577.488.074.584.589.183.8
ODTrack-B [61]AAAI2473.283.280.677.087.975.185.190.184.9
F-BDMTrack-384 [58]ICCV2372.081.577.775.484.372.984.589.084.0
ROMTrack-384 [4]ICCV2371.481.478.274.284.372.484.189.083.7
ARTrack384 [51]CVPR2372.681.779.175.584.374.385.189.184.8
SeqTrack-B384 [6]CVPR2371.581.177.874.584.371.483.988.883.6
GRM [20]CVPR2369.979.375.873.482.970.484.088.783.3
OSTrack384 [59]ECCV2271.181.177.673.783.270.883.988.583.2
SimTrack [5]ECCV2270.579.7-69.878.866.083.487.4-
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.437, + 0.482, + 0.514 + ], + "angle": 0, + "content": "TNL2K [49] comprises 700 test video sequences, each accompanied by natural language descriptions. As shown in Table 3, our method achieves state-of-the-art performance without utilizing natural language descriptions and surpasses existing vision-language trackers [21, 30]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.514, + 0.483, + 0.606 + ], + "angle": 0, + "content": "UAV123 [37] is a dataset focusing on low-altitude drone aerial scenarios, consisting of 123 video sequences with an average length of 915 frames per sequence. Our method achieves state-of-the-art performance with comparable model sizes, demonstrating the capability of our approach in aerial scenes and small object tracking." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.606, + 0.483, + 0.712 + ], + "angle": 0, + "content": "NfS [29] consists of 100 video sequences totaling 380,000 frames. Following previous works, we conduct evaluations on the 30 FPS version of the dataset. Our method rivals current state-of-the-art trackers in NfS. One possible reason for not achieving further improvement is that our method does not incorporate consecutive target position like ARTrack [51] and HIPTrack [3], which is crucial for NfS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.712, + 0.483, + 0.789 + ], + "angle": 0, + "content": "OTB2015 [54] is a classic dataset in visual tracking, consisting of 100 short-term tracking sequences that encompass 11 common challenges, such as target deformation, occlusion, and scale variation. Our method also achieves state-of-the-art performance on OTB2015." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.802, + 0.245, + 0.818 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The Importance of Spatio-Temporal Context Modeling and TMoE. Table 5 demonstrates the significance of two core designs in our paper: spatio-temporal context modeling and TMoE. In Table 5, removing TMoE indicates direct fine-tuning with LoRA [24], while removing spatio" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.436, + 0.908, + 0.478 + ], + "angle": 0, + "content": "Table 3. The performance of our method and other state-of-the-art trackers on TNL2K. The best three results are highlighted in red, blue and bold." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.489, + 0.887, + 0.664 + ], + "angle": 0, + "content": "
MethodAUC(%)PNorm(%)P(%)
SPMTrack-G64.782.670.6
SPMTrack-L63.781.569.2
SPMTrack-B62.079.766.7
LoRAT-B378 [33]59.9-63.7
ODTrack-L [61]61.7--
ARTrackV2-L384 [2]61.6--
ARTrack-L384 [51]60.3--
RTracker-L [26]60.6-63.7
UNINEXT-H [57]59.3-62.8
CiteTracker [30]57.7-59.6
VLT [21]53.1-53.3
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.854 + ], + "angle": 0, + "content": "temporal modeling indicates using only the first frame template and search region as input and eliminating the target state token. Compared to LoRA, applying TMoE for parameter-efficient fine-tuning yields a substantial improvement of \\(+0.8\\) AUC on LaSOT, demonstrates the superiority of TMoE. And the spatio-temporal context modeling further enhances performance by an additional \\(+1.2\\) AUC. On the basis of incorporating spatio-temporal modeling, the improvement brought by TMoE is \\(+1.0\\) AUC, indicating that TMoE can help the model better capture spatio-temporal contextual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Where to Apply TMoE. Conventional MoE is typically applied to the FFN layers within Transformer blocks, while keeping the attention layers unchanged. In contrast, we ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.488, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.486, + 0.434 + ], + "angle": 0, + "content": "Figure 4. The performance of our method compared with other state-of-the-art trackers in terms of AUC across various scenarios in the LaSOT test split." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.492 + ], + "angle": 0, + "content": "Table 4. The performance of our method and other state-of-the-art trackers on UAV123, NfS and OTB2015 in terms of AUC metrics. The best three results are highlighted in red, blue and bold." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.5, + 0.469, + 0.648 + ], + "angle": 0, + "content": "
MethodUAV123NfSOTB2015
SPMTrack-B71.767.472.7
HIPTrack [3]70.568.171.0
ARTrackV2-B [2]69.967.6-
ODTrack-L [61]--72.4
ARTrack384 [51]70.566.8-
SeqTrack-B384 [6]68.666.7-
DropTrack [52]--69.6
MixFormer-L [10]69.5--
STMTrack [16]64.7-71.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.661, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Table 5. Ablation studies on spatio-temporal context modeling and TMoE. Experiments are conducted on LaSOT." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.7, + 0.48, + 0.78 + ], + "angle": 0, + "content": "
#Spatio-TemporalTMoEAUC (%)\\(P_{Norm}(\\%)\\)P(%)
1✘✘72.981.979.1
2✘✓73.782.780.0
3✓✘73.982.880.0
4✓✓74.984.081.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.904 + ], + "angle": 0, + "content": "ply TMoE in both the attention layers and the FFN layers. Table 6 demonstrates the impact of applying TMoE in different layers. Comparing the first, third, and fifth rows, applying TMoE to the query, key, value projections in attention layers and to FFN layers can both improve performance. Comparing the first, second and third rows, we find that applying TMoE in attention layers yields significantly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.153 + ], + "angle": 0, + "content": "greater benefits compared to applying it only in FFN layers. Comparing the third and fourth rows, applying TMoE in the output projection matrix of the attention layer can bring a slight improvement." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.165, + 0.907, + 0.207 + ], + "angle": 0, + "content": "Table 6. Ablation studies on applying TMoE in different layers. All results are evaluated on LaSOT test split. “X” means replacing TMoE with LoRA." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.311 + ], + "angle": 0, + "content": "
#QKVOut ProjFFNAUC (%)\\(P_{Norm}(\\%)\\)P(%)
1XXX73.982.880.0
2XX✓74.183.180.8
3✓XX74.583.681.2
4✓✓X74.683.881.4
5✓✓✓74.984.081.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.513 + ], + "angle": 0, + "content": "Whether to use a shared compress expert. In TMoE, we employ a shared compression expert to reduce the input dimension to \\( r \\), and use multiple routed experts for subsequent processing. As shown in the first and second rows of Table 7, compared to providing each routed expert with a compression expert, sharing the compression expert can achieve better performance while reducing the parameter count. When not sharing the compression expert, we provide t-SNE visualizations of the outputs from compression experts and routed experts in five TMoE modules in Figure 5. The outputs of different compression experts show significant overlaps, while the routed experts are more dispersed, validating the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.909, + 0.665 + ], + "angle": 0, + "content": "TMOE vs. Conventional MoE. We compare TMOE with conventional MoE on the visual tracking task. For conventional MoE, we use LoRA to fine-tune the attention layers and apply MoE in the FFN layers with five experts: one frozen shared expert that copies pretrained FFN parameters, and four trainable routed experts, all routed experts are identical to the original FFN architecture, and a router is also applied. As shown in the first row and third row of Table 7, TMOE demonstrates significant advantages over conventional MoE in both parameter efficiency and performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.677, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Table 7. Ablation studies on whether to share the compression expert and use the conventional MoE to replace TMoE. Results are evaluated on LaSOT test split." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.73, + 0.912, + 0.794 + ], + "angle": 0, + "content": "
model variants#Params(M)AUC (%)PNorm(%)P(%)
SPMTrack-B115.374.984.081.7
w/o shared compress expert131.374.883.781.5
Conventional MoE316.873.482.679.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Number of Routed Experts. In Table 8, we evaluated the impact of different numbers of routed experts in TMoE. Results indicate that increasing the number of experts leads to improved performance, highlighting the potential of TMoE as a method to expand the parameter count and model scale. To balance the number of parameters and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.202, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Figure 5. Comparison of t-SNE visualizations. Each column shows outputs from all compression experts (top) and routed experts (bottom) within a TMoE module. Different colors represent distinct experts. Zoom in for better view." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.273, + 0.482, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Figure 6. Visualization comparison of search region attention maps with and without TMoE. Zoom in for better view." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.583, + 0.483, + 0.628 + ], + "angle": 0, + "content": "performance, we choose to use 4 experts, and whether there are separate optimal configurations for the TMoE in the attention layer and the FFN layer remains to be explored." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.839 + ], + "angle": 0, + "content": "The Number of Reference Frames. In SPMTrack, we use the template and several tracked frames as reference frames to enhance the spatio-temporal modeling ability. In Table 9, we explore the impact of using different numbers of reference frames during training and inference. During inference, if only two reference frames are used, we simply select the frame at \\(\\left\\lfloor \\frac{t}{2} \\right\\rfloor\\) as the reference frame. If four reference frames are used, we select the frames at \\(\\left\\lfloor \\frac{t}{4} \\right\\rfloor\\), \\(\\left\\lfloor \\frac{t}{2} \\right\\rfloor\\), and \\(\\left\\lfloor \\frac{3 \\times t}{4} \\right\\rfloor\\). As shown in Table 9, better performance can be achieved when the number of reference frames used during inference remains the same as that during training. Training with more reference frames can also lead to improved performance, illustrating that our method still has substantial potential for performance enhancement." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison of attention maps with and without TMoE. In Figure 6, to exclude the influence of intermediate tracked frames, we use the model in rows 1 and 2 of Table 5 for comparison, where spatio-temporal modeling is removed" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.089, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 8. The performance of our method on the test split of GOT-10K when setting different number of routed experts in TMoE." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.129, + 0.901, + 0.193 + ], + "angle": 0, + "content": "
Number2468
AO(%)75.376.576.677.1
SR0.5(%)84.485.986.386.4
SR0.75(%)75.276.377.077.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.208, + 0.905, + 0.25 + ], + "angle": 0, + "content": "Table 9. Ablation study on the number of reference frames during training and inference. Results are evaluated based on SPMTrack-B on GOT-10K test split." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.261, + 0.895, + 0.424 + ], + "angle": 0, + "content": "
TrainingInferenceAO (%)SR0.5(%)SR0.75(%)
2275.885.175.3
374.784.673.8
469.881.866.2
3273.183.972.7
376.585.976.3
472.883.471.3
4270.683.968.9
374.685.474.5
477.587.377.2
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.588 + ], + "angle": 0, + "content": "and we focus solely on the impact of TMoE. We compare the attention maps of the search region in the last block of Transformer encoder. The first column is the template, the second column displays search region, the third column is the visualization result and pure attention map, as well as the fourth column. Figure 6 shows that when facing challenging scenarios such as occlusions, distractors, and viewpoint variations, TMoE effectively suppresses background regions while enhancing perception of the target boundary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.634, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this paper, we present TMoE, a mixture of experts module tailored for visual tracking, and propose SPMTrack, a novel tracker enabling spatio-temporal context modeling based on TMoE. TMoE is applied in the linear layers in both self-attention and FFN layers, enhancing the diversity and flexibility of expert combinations to better handle various relation modeling in visual tracking. Additionally, TMoE employs a lightweight and efficient structure and serves as a method of parameter-efficient fine-tuning, which enables us to train SPMTrack of larger scales and enables SPMTrack to achieve state-of-the-art performance with only a small subset of parameters need to be trained. Furthermore, we hope that this work will inspire more applications of mixture of experts in the field of visual tracking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This paper was supported by Zhejiang Provincial Natural Science Foundation of China under Grant No.LD24F020016 and National Natural Science Foundation of China under Grant No.62176017." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16878" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompt autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19048-19057, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[3] Wenrui Cai, Qingjie Liu, and Yunhong Wang. Hiptrack: Visual tracking with historical prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19258-19267, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.484, + 0.368 + ], + "angle": 0, + "content": "[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9589-9600, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: a simplified architecture for visual object tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 375-392. Springer, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14572-14581, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Yucheng Chen and Lin Wang. emoe-tracker: Environmental moe-based transformer for robust event-guided object tracking. arXiv preprint arXiv:2406.20024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Mohammed Nowaz Rabbani Chowdhury, Shuai Zhang, Meng Wang, Sijia Liu, and Pin-Yu Chen. Patch-level routing in mixture-of-experts is provably sample-efficient for convolutional neural networks. In Proceedings of the 40th International Conference on Machine Learning, pages 6074-6114. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International conference on machine learning, pages 4057-4086. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13608-13618, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguis" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "tics (Volume 1: Long Papers), pages 1280-1297, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[13] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yangi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient scaling of language models with mixture-of-experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.904, + 0.431 + ], + "angle": 0, + "content": "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[16] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. Stmtrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13774-13783, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[17] Zhihong Fu, Zehua Fu, Qingjie Liu, Wenrui Cai, and Yunhong Wang. Sparsett: Visual tracking with sparse transformers. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-22, pages 905–912, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[18] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[19] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 146-164. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[20] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18686-18695, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[21] Mingzhe Guo, Zhipeng Zhang, Heng Fan, and Liping Jing. Divert more attention to vision-language tracking. In Advances in Neural Information Processing Systems, pages 4446-4460. Curran Associates, Inc., 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Kaijie He, Canlong Zhang, Sheng Xie, Zhixin Li, and Zhiwen Wang. Target-aware tracking with long-term context" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16879" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "attention. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):773-780, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[23] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[24] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[25] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[26] Yuqing Huang, Xin Li, Zikun Zhou, Yaowei Wang, Zhenyu He, and Ming-Hsuan Yang. Rtracker: Recoverable tracking via pn tree structured memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19038-19047, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[27] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.486, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.557, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[30] Xin Li, Yuqing Huang, Zhenyu He, Yaowei Wang, Huchuan Lu, and Ming-Hsuan Yang. Citetracker: Correlating image and text for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9974-9983, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[32] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. Advances in Neural Information Processing Systems, 35:16743-16754, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. In Computer Vision - ECCV 2024, pages 300-318, Cham, 2025. Springer Nature Switzerland. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[36] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[37] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, pages 445-461, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[38] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, pages 300-317, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[40] Jonas Pfeiffer, Andreas Rücklé, Clifton Poth, Aishwarya Kamath, Ivan Vulić, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. Adapterhub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[41] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems, pages 8583-8595. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[44] Zhangyong Tang, Tianyang Xu, Zhenhua Feng, Xuefeng Zhu, He Wang, Pengcheng Shao, Chunyang Cheng, Xiao-Jun Wu, Muhammad Awais, Sara Atito, et al. Revisiting rgbt tracking benchmarks from the perspective of modality validity: A new benchmark, problem, and method. arXiv preprint arXiv:2405.00168, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[45] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. Hydralora: An asymmetric lora architecture for efficient fine-tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Guangting Wang, Chong Luo, Xiaoyan Sun, Zhiwei Xiong," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16880" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "and Wenjun Zeng. Tracking by instance detection: A meta-learning approach. In CVPR, pages 6288-6297, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[48] Ke Wang, Jiahui Zhu, Minjie Ren, Zeming Liu, Shiwei Li, Zongye Zhang, Chenkai Zhang, Xiaoyu Wu, Qiqi Zhan, Qingjie Liu, et al. A survey on data synthesis and augmentation for large language models. arXiv preprint arXiv:2410.12896, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[49] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13763-13773, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[51] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9697-9706, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[52] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[53] Xun Wu, Shaohan Huang, and Furu Wei. Mixture of lora experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.481, + 0.58 + ], + "angle": 0, + "content": "[54] Y. Wu, J. Lim, and M. Yang. Object tracking benchmark. TPAMI, 37(9):1834-1848, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[55] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8751-8760, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[56] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19300-19309, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[57] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[58] Dawei Yang, Jianfeng He, Yinchao Ma, Qianjin Yu, and Tianzhu Zhang. Foreground-background distribution modeling transformer for visual object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10117-10127, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[59] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[60] Yihua Zhang, Ruisi Cai, Tianlong Chen, Guanhua Zhang, Huan Zhang, Pin-Yu Chen, Shiyu Chang, Zhangyang Wang, and Sijia Liu. Robust mixture-of-expert training for convolutional neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 90-101, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[61] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7588-7596, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16881" + } + ] +] \ No newline at end of file diff --git a/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_origin.pdf b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ef5df727173fd882749f113b6a22ea4fc147c98 --- /dev/null +++ b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/4755d93e-1740-4309-977d-6ab1138fe42c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1faee0af44e786e197ed8db6a52d47406494d07c23a75af4d6d060bffc73da4e +size 2879831 diff --git a/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/full.md b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df728ec3ba842091a6305e40a32027dcaf98fae8 --- /dev/null +++ b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/full.md @@ -0,0 +1,309 @@ +# SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking + +Wenrui Cai $^{1}$ , Qingjie Liu $^{1,2,3,*}$ , Yunhong Wang $^{1,3}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China + $^{2}$ Zhongguancun Laboratory, Beijing, China + $^{3}$ Hangzhou Innovation Institute, Beihang University, Hangzhou, China +{wenrui_cai, qingjie.liu, yhwang}@buaa.edu.cn + +# Abstract + +Most state-of-the-art trackers adopt one-stream paradigm, using a single Vision Transformer for joint feature extraction and relation modeling of template and search region images. However, relation modeling between different image patches exhibits significant variations. For instance, background regions dominated by target-irrelevant information require reduced attention allocation, while foreground, particularly boundary areas, need to be emphasized. A single model may not effectively handle all kinds of relation modeling simultaneously. In this paper, we propose a novel tracker called SPMTrack based on mixture-of-experts tailored for visual tracking task (TMoE), combining the capability of multiple experts to handle diverse relation modeling more flexibly. Benefiting from TMoE, we extend relation modeling from image pairs to spatio-temporal context, further improving tracking accuracy with minimal increase in model parameters. Moreover, we employ TMoE as a parameter-efficient fine-tuning method, substantially reducing trainable parameters, which enables us to train SPMTrack of varying scales efficiently and preserve the generalization ability of pretrained models to achieve superior performance. We conduct experiments on seven datasets, and experimental results demonstrate that our method significantly outperforms current state-of-the-art trackers. The source code is available at https://github.com/WenRuiCai/SPMTrack. + +# 1. Introduction + +Visual tracking aims to predict the location of a target throughout subsequent frames of a video, given the initial state in the first frame template. Currently, most state-of-the-art trackers adopt a Transformer-based [46] one-stream + +![](images/71693ee43b859c952338b7d95fcddc5a505a877abccf99af03e02f680d0d802d.jpg) +Figure 1. Comparison of LaSOT AUC and model parameter count across different trackers. Larger loop indicates better performance. + +paradigm [3, 5, 10, 33, 52, 59, 61], utilizing a single Vision Transformer [12] backbone that accepts both the template and the search region image as input, and employs self-attention to perform image feature extraction and template-search region relation modeling simultaneously. + +However, not all relation modeling between image patches benefits tracking performance. Prior studies [4, 20, 58, 59] have demonstrated that the abundance of background patches in the search region deteriorates the discriminative power of the foreground features, and single vanilla attention has limited capability in suppressing undesirable foreground-background interactions. Meanwhile, there are also some kinds of relation modeling that demonstrate positive effects on tracking performance [17, 19]. For instance, enhancing attention to target boundary regions can improve tracking performance [17], especially in challenging scenarios such as motion blur, partial occlusion, and deformation. + +To address various relation modeling, numerous meth + +ods propose specific designs, including background filtering to reduce target-irrelevant information [3, 59], customizing different interaction modules for tokens belonging to different categories [4, 20], and attention score adjustment based on foreground-background weights of each token [58]. However, existing methods can only handle one or a few predefined types of relation modeling, and the handcrafted strategies inherently lack adaptability. Furthermore, there may exist many latent relationships that further complicate the specialized design. These limitations motivate our question: How to design a tracker to adaptively process various relation modeling between image patches? + +Inspired by mixture of experts (MoE) [1, 11, 13, 27, 41] in natural language processing, in this paper, we address the question by proposing SPMTrack, a novel tracker based on TMoE, a specialized mixture of experts module for visual tracking task. Similar to MoE, TMoE handles diverse relation modeling through adaptive weighted combinations of multiple experts. However, unlike traditional MoE that is exclusively applied to feed-forward network (FFN) layers, TMoE is simultaneously applied to the linear layers within both the attention layers and the FFN layers. This finer-grained application enables TMoE to achieve more diverse expert combinations, thereby enhancing the capability of various relation modeling. Moreover, differing from traditional MoE that uses FFNs as experts, all the experts in TMoE are linear layers and designed with a lightweight and efficient structure to ensure overall efficiency. Furthermore, we employ TMoE as a parameter-efficient finetuning method, which enables us to train SPMTrack of larger scales. In SPMTrack, only a subset of experts and the router within TMoE, along with the prediction head need to be trained, which substantially reduces the trainable parameters while preserving the generalization capability of pretrained models. Notably, our method reduces the number of trainable parameters by nearly $80\%$ compared to [2], while demonstrates better performance. + +Benefiting from the powerful capability of TMoE in handling diverse relation modeling, unlike previous one-stream trackers that only perform relation modeling between template-search region pairs, we extend SPMTrack to incorporate multi-frame spatio-temporal context modeling, which further enhances the performance of SPMTrack with minimal additional parameters. We evaluated the performance of our method on seven datasets, and the experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets including LaSOT [14], GOT-10K [25], TrackingNet [38] and TNL2K [49]. Through extensive experiments with models of varying scales, as shown in Figure 1, Table 1 and Table 2, with ViT-B [12] as backbone and less than $30\%$ of parameters need to be trained, SPMTrack-B achieves or even surpasses the best performance of previous trackers utilizing ViT-L. + +The main contributions of this paper can be summarized as: (1) We propose TMoE, a mixture of experts module tailored for visual tracking. TMoE enhances the capability to handle various relation modeling and can be used as a method of parameter-efficient fine-tuning. (2) Based on TMoE, we propose SPMTrack, a novel tracker that can effectively integrate spatio-temporal context for visual tracking. (3) Experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets. We trained SPMTrack of varying scales, and our method with ViT-B as backbone can achieve or even surpass current trackers using ViT-L. + +# 2. Related Work + +# 2.1. One-Stream Trackers + +One-stream trackers employ a single Vision Transformer [12] as backbone, simultaneously performing feature extraction and relation modeling at each layer [5, 10, 55, 59]. Many efforts focus on incorporating historical context. TATrack [22] and ODTrack [61] propose architectures capable of processing multiple frames, while HIPTrack [3] introduces historical target features through prompt learning. SeqTrack [6], ARTrack[51], and ARTrackV2 [2] achieve more precise predictions by incorporating multiple historical target position. Other works try to improve conventional self-attention to handle different relation modeling. OSTrack [59] progressively filters background regions to reduce foreground-background relation modeling, GRM [20] and ROMTrack [4] categorizes image tokens and constrains relation modeling between specific categories, F-BDMTrack [58] adjusts the attention score by calculating the foreground-background weights of tokens. But these methods can only handle specific relation modeling and rely on manually designed strategies. + +# 2.2. Mixture of Experts + +Mixture-of-experts (MoE) [1, 13, 27] are widely applied in large language models [9, 11, 15, 36, 41, 53], where each expert focuses on specific aspects of the data or particular tasks [48]. By performing an adaptive weighted sum over multiple experts, MoE can better capture complex patterns and relationships within the input data. MoE also has applications in computer vision [8, 43, 60]. However, there is few dedicated exploration of MoE in visual tracking. MoETrack [44] employs multiple prediction heads as experts in the field of RGB-T tracking, which is rather far-fetched and deviates from the general practice of embedding MoE within Transformer layers. In the field of RGB-E tracking, eMoETracker [7] inserts multiple experts between Transformer layers, which more closely resembles an adapter architecture [23] and contrasts with MoE implementations where experts are embedded within Transformer layers. + +![](images/2ad4663e1941ba435163c2dc67801f2da2cfb13becb701174fd848c1830728d7.jpg) +Figure 2. Overview of SPMTrack that consists of a feature extraction network and a prediction head. The main body of the feature extraction network is a Transformer encoder composed of multiple TMoEBlocks. The structure of TMoEBlock is shown on the right side. + +# 2.3. Parameter-efficient Fine-tuning + +Parameter-efficient fine-tuning aims to reduce computational costs and retain the general knowledge of the pretrained model by freezing most parameters and fine-tuning only a small subset of parameters. Common parameter-efficient fine-tuning approaches include adapter-based methods [18, 23, 40], prompt-based methods [28, 31, 50, 62] and low-rank adaptation methods [24, 45]. In visual tracking, HIPTrack [3] incorporates historical target features as prompts, while LoRAT [33] employs low-rank adaptation for efficient training. And our method maintains flexibility in diverse relation modeling while reducing trainable parameters by fine-tuning TMoE. + +# 3. Method + +# 3.1. Overall Architecture + +As illustrated in Figure 2, we propose SPMTrack, a novel tracker that integrates multi-frame spatio-temporal context and employs TMoE for parameter-efficient fine-tuning. The overall architecture follows the one-stream paradigm, consisting of a Feature Extraction Network based on a Vision Transformer [12] as backbone, and a prediction head. + +Feature Extraction Network. The feature extraction network is denoted as $\Phi$ . The main body of the feature extraction network is a Transformer encoder that utilizes TMoE-Block as its primary component, which is designed based on TMoE. The architecture of TMoE-Block is illustrated in the right panel of Figure 2, and details will be introduced in Section 3.2. The encoder simultaneously per + +forms feature extraction and relation modeling for input images. We extend the input from one single template $\mathbf{Z} \in \mathbb{R}^{H_z \times W_z \times 3}$ that is commonly used in one-stream trackers to the video sequence $\{Z_1, \dots, Z_N | Z_i \in \mathbb{R}^{H_z \times W_z \times 3}\}$ as reference frames. The reference frames contain the first frame template and the selected tracked historical frames, with a total of $N$ frames. The feature extraction network uses patch embedding to partition each input image into patches of size $M \times M$ , map all patches into tokens by a convolutional layer, and obtain the reference token sequence of each image $T_i \in \mathbb{R}^{N_T \times d}$ , where $d$ denotes the hidden dimension of Transformer encoder and $N_T = \frac{H_z W_z}{M^2}$ represents the number of tokens corresponding to each image. Similarly, for the search region $S \in \mathbb{R}^{H_s \times W_s \times 3}$ , we can also obtain the token sequence $X \in \mathbb{R}^{N_X \times d}$ . All tokens are added with positional embedding and a learnable token type embedding [33]. Inspired by [32, 61], we additionally introduce a target state token $H \in \mathbb{R}^{1 \times d}$ that is learnable as well and propagates over time. All tokens are concatenated and fed into the Transformer encoder. The process can be formulated as: + +$$ +\boldsymbol {T} _ {i} ^ {j} = \left\{ \begin{array}{l} \boldsymbol {T} _ {i} ^ {j} + \boldsymbol {P} \boldsymbol {E} _ {j} + \boldsymbol {T} \boldsymbol {E} _ {o}, i f \text {t o k e n} (i, j) \text {i n b b o x} \\ \boldsymbol {T} _ {i} ^ {j} + \boldsymbol {P} \boldsymbol {E} _ {j} + \boldsymbol {T} \boldsymbol {E} _ {b}, o t h e r w i s e \end{array} \right. \tag {1} +$$ + +$$ +\boldsymbol {X} ^ {j} = \boldsymbol {X} ^ {j} + \boldsymbol {P} \boldsymbol {E} _ {j} + \boldsymbol {T} \boldsymbol {E} _ {S} +$$ + +$$ +\boldsymbol {I} = \operatorname {C o n c a t} (\boldsymbol {H}, \boldsymbol {T} _ {1}, \dots , \boldsymbol {T} _ {N}, \boldsymbol {X}) +$$ + +where $T_{i}^{j}$ represents the $j^{th}$ token in $T_{i}$ , $PE_{j}$ represents positional embedding of the $j^{th}$ token, similarly for $X^{j}$ . $TE_{o}, TE_{b}$ and $TE_{S}$ represents the learnable type embeddings of the foreground region tokens, background region + +tokens in reference frames and the search region tokens, respectively. $\pmb{I}$ denotes the input to the Transformer encoder. + +Subsequently, the Transformer encoder processes all tokens and produces an output $\pmb{O} \in \mathbb{R}^{(1 + N \times N_T + N_X) \times d}$ . The output tokens $\pmb{X}' \in \mathbb{R}^{N_X \times d}$ corresponding to search region and the output $\pmb{H}' \in \mathbb{R}^{1 \times d}$ from the target state token are fed into the prediction head for prediction. $\pmb{H}'$ will be added with the input target state token in the next tracking frame, which can be formulated as: + +$$ +\boldsymbol {O} = \Phi (\boldsymbol {I}), [ \boldsymbol {H} ^ {\prime}, \boldsymbol {T} _ {1} ^ {\prime}, \dots , \boldsymbol {T} _ {N} ^ {\prime}, \boldsymbol {X} ^ {\prime} ] = \operatorname {S p l i t} (\boldsymbol {O}) \tag {2} +$$ + +Prediction Head. The input to the prediction head consists of the output target state token $H'$ and the output search region tokens $X'$ from the feature extraction network. Initially, the prediction head performs matrix multiplication between $H'$ and $X'$ to obtain the weight $U \in \mathbb{R}^{N_X \times 1}$ . As the target state token propagates over time, $U$ can reweight and adjust current search region tokens based on the historical state of the target, thereby introducing more spatio-temporal tracking context. After element-wise multiplication of $X'$ with $U$ , the resulting data is reshaped into a feature map $F \in \mathbb{R}^{\frac{H_s}{M} \times \frac{W_s}{M} \times d}$ , which is then processed by a decoupled double-MLP head to perform target center classification and box regression, respectively. For fair comparison, the double-MLP head is identical to [33]. + +# 3.2. The Design of TMoEBlock + +TMoEBlock serves as the primary component of the Transformer encoder in the feature extraction network. As illustrated in the right of Figure 2, TMoEBlock is built based on the standard vision Transformer block. However, unlike the standard ViT block, TMoEBlock applies TMoE module to both the multi-head self-attention (MSA) layer and the feed-forward network (FFN) layer, obtaining MoE-based MSA $(\mathrm{M}^2\mathrm{SA})$ layer and MoE-based FFN (MFFN) layer. The finer-grained application of TMoE is capable to achieve more diverse combinations for various relation modeling. The processing of TMoEBlock can be formulated as: + +$$ +\begin{array}{l} \boldsymbol {O} _ {l} ^ {\prime} = \mathrm {M} ^ {2} \operatorname {S A} (\ln (\boldsymbol {O} _ {l - 1})) + \boldsymbol {O} _ {l - 1}, \quad l = 1, \dots , L \tag {3} \\ \boldsymbol {O} _ {l} = \mathrm {M F F N} (\mathrm {L N} (\boldsymbol {O} _ {l} ^ {\prime})) + \boldsymbol {O} _ {l} ^ {\prime}, l = 1, \dots , L \\ \end{array} +$$ + +where $L$ denotes the total number of TMoEBlocks, $\pmb{O}_l$ represents the output of the $l^{th}$ block in Transformer encoder. When $l = L$ , $\pmb{O}_l$ is the output $\pmb{O}$ of the feature extraction network; when $l = 0$ , it is the overall input $\pmb{I}$ . LN stands for the layer normalization operation. + +MoE-based Multi-head Self-Attention (M²SA). M²SA not only replaces the standard three linear projection layers $Q, K, V$ in standard attention for the query, key, and value with TMoE modules but also replaces the output projection layer with TMoE module. + +MoE-based Feed-Forward Network (MFFN). Conventional feed-forward network consists of two linear layer and + +![](images/5594208d4c14b2f214b8902ee400b433966b6c8af60d984be909b91dc960d36b.jpg) +Figure 3. The structure of TMoE. The symbols maintain the same meaning with Figure 2. + +activation function. MFFN replaces both linear layers in FFN with TMoE modules, while maintaining all other architectural configurations unchanged. + +# 3.3. Mixture of Experts for Visual Tracking + +In this section, we introduce TMoE, a mixture of experts module tailored for visual tracking task. As illustrated in Figure 3, TMoE consists of one shared expert, one router, one compression expert and $N_{e}$ routed experts. In large language models, mixture of experts are commonly applied to replace FFN layers, and the experts comprising MoE are typically the same structure as FFN. However, TMoE is used to replace the linear layers in both self-attention and FFN, and the shared expert, the compression expert and the routed experts are all implemented as linear layers. + +In TMoE, the shared expert directly copies weights from the corresponding linear layers in pretrained model and remains frozen during training. The rationale is that the original weights have been well-trained during the pre-training stage, and the shared expert needs to retain more general knowledge. In contrast, the compression experts, routed experts, and the router are all trainable. + +Given a TMOE module replacing a linear layer with input dimension $d$ and output dimension $D$ , for each token $\pmb{x} \in \mathbb{R}^d$ , as shown in Figure 3, TMOE first computes the weights $\pmb{W} \in \mathbb{R}^{N_e}$ of all routed experts based on the input $\pmb{x}$ , the weights are normalized by Softmax. After that, TMOE calculates the output $\pmb{y}^s \in \mathbb{R}^D$ of the shared expert and the outputs $\pmb{y}_i^e \in \mathbb{R}^D$ of all routed experts. In particular, instead of adopting the same structure as the shared expert, inspired by [45], we first employ a compression expert with a $d \times r$ linear transformation to reduce the input dimension from $d$ to $r$ and obtain the compressed result $\pmb{y}^c \in \mathbb{R}^r$ , where $r \ll d$ and $r \ll D$ . Subsequently, all routed experts use the compressed input $\pmb{y}^c$ to produce their respective outputs $\pmb{y}_i^e \in \mathbb{R}^D$ . The process can be formally described as: + +$$ +\begin{array}{l} \boldsymbol {W} = \text {S o f t m a x} (\operatorname {R o u t e r} (\boldsymbol {x})), \quad \boldsymbol {y} ^ {s} = \boldsymbol {E} ^ {S} (\boldsymbol {x}) \\ \boldsymbol {y} ^ {c} = \boldsymbol {E} ^ {C} (\boldsymbol {x}), \quad \boldsymbol {y} _ {i} ^ {e} = \boldsymbol {E} _ {i} ^ {R} (\boldsymbol {y} ^ {c}) \quad i = 1, \dots , N _ {e} \\ \end{array} +$$ + +where Router is also a linear layer that maps the input $x$ + +to weights $\pmb{W}$ , and $\pmb{E}^{S}$ , $\pmb{E}^{C}$ , and $\pmb{E}_{i}^{R}$ denote the shared expert, compress expert, and $i^{th}$ routed expert, respectively. + +After obtaining the outputs of all experts, we compute a weighted sum of the outputs $\pmb{y}_i^e$ from all routed expert based on $\pmb{W}$ to obtain $\pmb{y}^e \in \mathbb{R}^D$ , which means that the routed experts are densely activated in TMoE. Subsequently, $\pmb{y}^e$ is added with the shared expert output $\pmb{y}^s$ to get the final output $\pmb{y} \in \mathbb{R}^D$ . The process can be formulated as follows: + +$$ +\boldsymbol {y} ^ {e} = \sum_ {i = 1} ^ {N _ {e}} \boldsymbol {W} _ {i} \boldsymbol {y} _ {i} ^ {e}, \quad \boldsymbol {y} = \boldsymbol {y} ^ {e} + \boldsymbol {y} ^ {s} \tag {5} +$$ + +# 3.4. Loss Function + +The outputs of the decoupled double-MLP prediction head are supervised using binary cross-entropy loss for target center classification and Generalized IoU loss [42] for bounding box regression, respectively. Both loss terms are assigned with the weighting coefficients set to 1. + +# 4. Experiments + +# 4.1. Implementation Details + +Model settings. We provide three versions of SPMTrack: SPMTrack-B, SPMTrack-L, and SPMTrack-G, utilizing ViT-B [12], ViT-L [12], and ViT-G [39] as backbones respectively. All versions maintain a consistent patch size $M$ of 14 and adopt the pretrained weights from DINOv2 [39]. The resolution of input images are the same across all versions, with reference frames of $196 \times 196$ and search region of $378 \times 378$ . A crop factor of 2 is applied to reference frames and 5 to search region. We set the number of reference frames $N$ to 3 for spatio-temporal modeling. Each TMoE module incorporates 4 routed experts, where the input dimension of the routed experts $r$ is set to 64. Other configurations and model information are shown in Table 1, where $L$ is the number of TMoEBlocks, $d$ denotes hidden dimension and $N_h$ is the number of attention heads. + +Table 1. Model configurations and parameter counts across SPM-Track variants of different scales. + +
SPMTrack-BSPMTrack-LSPMTrack-G
BackboneL=12L=24L=40
d=768d=1024d=1536
Nh=12Nh=16Nh=24
#Params (M)115.3379.61339.5
#Trainable +Params (M)29.275.9204.0
+ +Datasets. Following previous works [3, 33, 51, 59], we use LaSOT [14], TrackingNet [47], GOT-10K [25], and COCO [34] for training. When evaluating on GOT-10K test set, we strictly follow the protocol by training exclusively on GOT-10K training set. When evaluating on other datasets, + +we jointly train on the training sets of all four datasets, randomly sampling from each dataset with equal probability. For video datasets, we sample 5 frames from one video with random frame intervals ranging from 1 to 200. The sampled frames are arranged either in forward or reverse order with equal probability, where the first 3 frames serve as reference frames and the remaining 2 as search frames. For COCO, we replicate each image 5 times to maintain consistency with the video frame sampling strategy. + +Training and Optimization. Our method is implemented based on PyTorch 2.3.1 and trained on 8 NVIDIA A100 GPUs. We maintain a global batch size of 128 during training. When training exclusively on GOT-10K, we train for 100 epochs. When joint training on multiple datasets, we extend to 170 epochs. In each epoch, we sample 131,072 sequences. We employ AdamW [35] as optimizer with learning rate of $10^{-4}$ and weight decay of 0.1. The learning rate scheduler warms up over the initial 2 epochs from $10^{-7}$ to $10^{-4}$ and then follows a cosine schedule to $10^{-6}$ . + +Inference. We use a total of 3 reference frames, including the annotated first frame template. When current search frame index $t \leq 3$ , we use all available tracked frames as reference frames, if there are fewer than three frames, we repeat the template as necessary to reach the required number. When $t > 3$ , we select two tracked frames as reference frames at indexes $\left\lfloor \frac{t}{3} \right\rfloor$ and $\left\lfloor \frac{2t}{3} \right\rfloor$ . Following previous trackers [4, 59, 61], we apply a Hanning window penalty to the output response map of the classification head. + +# 4.2. Comparisons with the State-of-the-Art + +LaSOT [14] is a large-scale long-term tracking dataset with its test set containing 280 sequences. As shown in Table 2, SPMTrack-B significantly outperforms all trackers utilizing ViT-B as backbone. Our method also substantially surpasses LoRAT [33] that also employs parameter-efficient fine-tuning. Additionally, as shown in Figure 4, we evaluate tracker performance across various challenging scenarios in LaSOT. Our method surpasses previous trackers that also incorporate spatio-temporal context [2, 3, 6, 51, 56], while significantly outperforming ROMTrack [4] and GRM [20], which focus on optimizing relation modeling. + +GOT-10K [25] contains 9,335 sequences for training and 180 sequences for testing. We follow the protocol that trackers are only allowed to be trained using GOT-10K train split. Our approach significantly outperforms LoRAT [33] with the same backbone. And our SPMTrack-B rivals the performance of the current state-of-the-art methods, while SPMTrack-L substantially surpasses all existing trackers. + +TrackingNet [38] is a large-scale visual tracking dataset with a test set comprising 511 video sequences. Table 2 demonstrates that our method achieves state-of-the-art performance using only ViT-B as backbone, surpassing current trackers that utilize ViT-L or even ViT-G. + +Table 2. State-of-the-art comparison on LaSOT, GOT-10k and TrackingNet. \* denotes for trackers trained only with GOT-10k train split. The best three results are highlighted in red, blue and bold, respectively. + +
MethodSourceLaSOTGOT-10k*TrackingNet
AUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)AUC(%)PNorm(%)P(%)
SPMTrack-BOurs74.984.081.776.585.976.386.190.285.6
SPMTrack-LOurs76.885.984.080.089.479.986.991.087.2
SPMTrack-GOurs77.486.685.081.089.282.387.391.488.1
LoRAT-B378 [33]ECCV2472.981.979.173.782.672.984.288.483.0
LoRAT-L378 [33]ECCV2475.184.182.077.586.278.185.689.785.4
LoRAT-G378 [33]ECCV2476.285.383.578.987.880.786.090.286.1
AQATrack384 [56]CVPR2472.782.980.276.085.274.984.889.384.3
ARTrackV2-B384 [2]CVPR2473.082.079.677.586.075.585.789.885.5
ARTrackV2-L384 [2]CVPR2473.682.881.179.587.879.686.190.486.2
HIPTrack [3]CVPR2472.782.979.577.488.074.584.589.183.8
ODTrack-B [61]AAAI2473.283.280.677.087.975.185.190.184.9
F-BDMTrack-384 [58]ICCV2372.081.577.775.484.372.984.589.084.0
ROMTrack-384 [4]ICCV2371.481.478.274.284.372.484.189.083.7
ARTrack384 [51]CVPR2372.681.779.175.584.374.385.189.184.8
SeqTrack-B384 [6]CVPR2371.581.177.874.584.371.483.988.883.6
GRM [20]CVPR2369.979.375.873.482.970.484.088.783.3
OSTrack384 [59]ECCV2271.181.177.673.783.270.883.988.583.2
SimTrack [5]ECCV2270.579.7-69.878.866.083.487.4-
+ +TNL2K [49] comprises 700 test video sequences, each accompanied by natural language descriptions. As shown in Table 3, our method achieves state-of-the-art performance without utilizing natural language descriptions and surpasses existing vision-language trackers [21, 30]. + +UAV123 [37] is a dataset focusing on low-altitude drone aerial scenarios, consisting of 123 video sequences with an average length of 915 frames per sequence. Our method achieves state-of-the-art performance with comparable model sizes, demonstrating the capability of our approach in aerial scenes and small object tracking. + +NfS [29] consists of 100 video sequences totaling 380,000 frames. Following previous works, we conduct evaluations on the 30 FPS version of the dataset. Our method rivals current state-of-the-art trackers in NfS. One possible reason for not achieving further improvement is that our method does not incorporate consecutive target position like ARTrack [51] and HIPTrack [3], which is crucial for NfS. + +OTB2015 [54] is a classic dataset in visual tracking, consisting of 100 short-term tracking sequences that encompass 11 common challenges, such as target deformation, occlusion, and scale variation. Our method also achieves state-of-the-art performance on OTB2015. + +# 4.3. Ablation Study + +The Importance of Spatio-Temporal Context Modeling and TMoE. Table 5 demonstrates the significance of two core designs in our paper: spatio-temporal context modeling and TMoE. In Table 5, removing TMoE indicates direct fine-tuning with LoRA [24], while removing spatio + +Table 3. The performance of our method and other state-of-the-art trackers on TNL2K. The best three results are highlighted in red, blue and bold. + +
MethodAUC(%)PNorm(%)P(%)
SPMTrack-G64.782.670.6
SPMTrack-L63.781.569.2
SPMTrack-B62.079.766.7
LoRAT-B378 [33]59.9-63.7
ODTrack-L [61]61.7--
ARTrackV2-L384 [2]61.6--
ARTrack-L384 [51]60.3--
RTracker-L [26]60.6-63.7
UNINEXT-H [57]59.3-62.8
CiteTracker [30]57.7-59.6
VLT [21]53.1-53.3
+ +temporal modeling indicates using only the first frame template and search region as input and eliminating the target state token. Compared to LoRA, applying TMoE for parameter-efficient fine-tuning yields a substantial improvement of $+0.8$ AUC on LaSOT, demonstrates the superiority of TMoE. And the spatio-temporal context modeling further enhances performance by an additional $+1.2$ AUC. On the basis of incorporating spatio-temporal modeling, the improvement brought by TMoE is $+1.0$ AUC, indicating that TMoE can help the model better capture spatio-temporal contextual information. + +Where to Apply TMoE. Conventional MoE is typically applied to the FFN layers within Transformer blocks, while keeping the attention layers unchanged. In contrast, we ap + +![](images/774929da7c6c25f4145a472543c9c363e124168b2dafea5e572e6a053210cd1f.jpg) +Figure 4. The performance of our method compared with other state-of-the-art trackers in terms of AUC across various scenarios in the LaSOT test split. + +Table 4. The performance of our method and other state-of-the-art trackers on UAV123, NfS and OTB2015 in terms of AUC metrics. The best three results are highlighted in red, blue and bold. + +
MethodUAV123NfSOTB2015
SPMTrack-B71.767.472.7
HIPTrack [3]70.568.171.0
ARTrackV2-B [2]69.967.6-
ODTrack-L [61]--72.4
ARTrack384 [51]70.566.8-
SeqTrack-B384 [6]68.666.7-
DropTrack [52]--69.6
MixFormer-L [10]69.5--
STMTrack [16]64.7-71.9
+ +Table 5. Ablation studies on spatio-temporal context modeling and TMoE. Experiments are conducted on LaSOT. + +
#Spatio-TemporalTMoEAUC (%)\(P_{Norm}(\%)\)P(%)
1✘✘72.981.979.1
2✘✓73.782.780.0
3✓✘73.982.880.0
4✓✓74.984.081.7
+ +ply TMoE in both the attention layers and the FFN layers. Table 6 demonstrates the impact of applying TMoE in different layers. Comparing the first, third, and fifth rows, applying TMoE to the query, key, value projections in attention layers and to FFN layers can both improve performance. Comparing the first, second and third rows, we find that applying TMoE in attention layers yields significantly + +greater benefits compared to applying it only in FFN layers. Comparing the third and fourth rows, applying TMoE in the output projection matrix of the attention layer can bring a slight improvement. + +Table 6. Ablation studies on applying TMoE in different layers. All results are evaluated on LaSOT test split. “X” means replacing TMoE with LoRA. + +
#QKVOut ProjFFNAUC (%)\(P_{Norm}(\%)\)P(%)
1XXX73.982.880.0
2XX✓74.183.180.8
3✓XX74.583.681.2
4✓✓X74.683.881.4
5✓✓✓74.984.081.7
+ +Whether to use a shared compress expert. In TMoE, we employ a shared compression expert to reduce the input dimension to $r$ , and use multiple routed experts for subsequent processing. As shown in the first and second rows of Table 7, compared to providing each routed expert with a compression expert, sharing the compression expert can achieve better performance while reducing the parameter count. When not sharing the compression expert, we provide t-SNE visualizations of the outputs from compression experts and routed experts in five TMoE modules in Figure 5. The outputs of different compression experts show significant overlaps, while the routed experts are more dispersed, validating the effectiveness of our approach. + +TMOE vs. Conventional MoE. We compare TMOE with conventional MoE on the visual tracking task. For conventional MoE, we use LoRA to fine-tune the attention layers and apply MoE in the FFN layers with five experts: one frozen shared expert that copies pretrained FFN parameters, and four trainable routed experts, all routed experts are identical to the original FFN architecture, and a router is also applied. As shown in the first row and third row of Table 7, TMOE demonstrates significant advantages over conventional MoE in both parameter efficiency and performance. + +Table 7. Ablation studies on whether to share the compression expert and use the conventional MoE to replace TMoE. Results are evaluated on LaSOT test split. + +
model variants#Params(M)AUC (%)PNorm(%)P(%)
SPMTrack-B115.374.984.081.7
w/o shared compress expert131.374.883.781.5
Conventional MoE316.873.482.679.8
+ +The Number of Routed Experts. In Table 8, we evaluated the impact of different numbers of routed experts in TMoE. Results indicate that increasing the number of experts leads to improved performance, highlighting the potential of TMoE as a method to expand the parameter count and model scale. To balance the number of parameters and + +![](images/e3f622e8bba3cf50600e59b75077de5b8c4165f72354cf7066a46b048b1c2ac0.jpg) +Figure 5. Comparison of t-SNE visualizations. Each column shows outputs from all compression experts (top) and routed experts (bottom) within a TMoE module. Different colors represent distinct experts. Zoom in for better view. + +![](images/996fd35ab80d4aa31c79ad29e23b61b958d7985cd7700cabcfac93ffd8e49497.jpg) +Figure 6. Visualization comparison of search region attention maps with and without TMoE. Zoom in for better view. + +performance, we choose to use 4 experts, and whether there are separate optimal configurations for the TMoE in the attention layer and the FFN layer remains to be explored. + +The Number of Reference Frames. In SPMTrack, we use the template and several tracked frames as reference frames to enhance the spatio-temporal modeling ability. In Table 9, we explore the impact of using different numbers of reference frames during training and inference. During inference, if only two reference frames are used, we simply select the frame at $\left\lfloor \frac{t}{2} \right\rfloor$ as the reference frame. If four reference frames are used, we select the frames at $\left\lfloor \frac{t}{4} \right\rfloor$ , $\left\lfloor \frac{t}{2} \right\rfloor$ , and $\left\lfloor \frac{3 \times t}{4} \right\rfloor$ . As shown in Table 9, better performance can be achieved when the number of reference frames used during inference remains the same as that during training. Training with more reference frames can also lead to improved performance, illustrating that our method still has substantial potential for performance enhancement. + +Comparison of attention maps with and without TMoE. In Figure 6, to exclude the influence of intermediate tracked frames, we use the model in rows 1 and 2 of Table 5 for comparison, where spatio-temporal modeling is removed + +Table 8. The performance of our method on the test split of GOT-10K when setting different number of routed experts in TMoE. + +
Number2468
AO(%)75.376.576.677.1
SR0.5(%)84.485.986.386.4
SR0.75(%)75.276.377.077.6
+ +Table 9. Ablation study on the number of reference frames during training and inference. Results are evaluated based on SPMTrack-B on GOT-10K test split. + +
TrainingInferenceAO (%)SR0.5(%)SR0.75(%)
2275.885.175.3
374.784.673.8
469.881.866.2
3273.183.972.7
376.585.976.3
472.883.471.3
4270.683.968.9
374.685.474.5
477.587.377.2
+ +and we focus solely on the impact of TMoE. We compare the attention maps of the search region in the last block of Transformer encoder. The first column is the template, the second column displays search region, the third column is the visualization result and pure attention map, as well as the fourth column. Figure 6 shows that when facing challenging scenarios such as occlusions, distractors, and viewpoint variations, TMoE effectively suppresses background regions while enhancing perception of the target boundary. + +# 5. Conclusion + +In this paper, we present TMoE, a mixture of experts module tailored for visual tracking, and propose SPMTrack, a novel tracker enabling spatio-temporal context modeling based on TMoE. TMoE is applied in the linear layers in both self-attention and FFN layers, enhancing the diversity and flexibility of expert combinations to better handle various relation modeling in visual tracking. Additionally, TMoE employs a lightweight and efficient structure and serves as a method of parameter-efficient fine-tuning, which enables us to train SPMTrack of larger scales and enables SPMTrack to achieve state-of-the-art performance with only a small subset of parameters need to be trained. Furthermore, we hope that this work will inspire more applications of mixture of experts in the field of visual tracking. + +Acknowledgements. This paper was supported by Zhejiang Provincial Natural Science Foundation of China under Grant No.LD24F020016 and National Natural Science Foundation of China under Grant No.62176017. + +# References + +[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompt autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19048-19057, 2024. 2, 5, 6, 7 +[3] Wenrui Cai, Qingjie Liu, and Yunhong Wang. Hiptrack: Visual tracking with historical prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19258-19267, 2024. 1, 2, 3, 5, 6, 7 +[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9589-9600, 2023. 1, 2, 5, 6 +[5] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: a simplified architecture for visual object tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 375-392. Springer, 2022. 1, 2, 6 +[6] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14572-14581, 2023. 2, 5, 6, 7 +[7] Yucheng Chen and Lin Wang. emoe-tracker: Environmental moe-based transformer for robust event-guided object tracking. arXiv preprint arXiv:2406.20024, 2024. 2 +[8] Mohammed Nowaz Rabbani Chowdhury, Shuai Zhang, Meng Wang, Sijia Liu, and Pin-Yu Chen. Patch-level routing in mixture-of-experts is provably sample-efficient for convolutional neural networks. In Proceedings of the 40th International Conference on Machine Learning, pages 6074-6114. PMLR, 2023. 2 +[9] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International conference on machine learning, pages 4057-4086. PMLR, 2022. 2 +[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13608-13618, 2022. 1, 2, 7 +[11] Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguis + +tics (Volume 1: Long Papers), pages 1280-1297, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5 +[13] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yangi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient scaling of language models with mixture-of-experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 2 +[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 2, 5 +[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2 +[16] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. Stmtrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13774-13783, 2021. 7 +[17] Zhihong Fu, Zehua Fu, Qingjie Liu, Wenrui Cai, and Yunhong Wang. Sparsett: Visual tracking with sparse transformers. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-22, pages 905–912, 2022. 1 +[18] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3 +[19] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 146-164. Springer, 2022. 1 +[20] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18686-18695, 2023. 1, 2, 5, 6 +[21] Mingzhe Guo, Zhipeng Zhang, Heng Fan, and Liping Jing. Divert more attention to vision-language tracking. In Advances in Neural Information Processing Systems, pages 4446-4460. Curran Associates, Inc., 2022. 6 +[22] Kaijie He, Canlong Zhang, Sheng Xie, Zhixin Li, and Zhiwen Wang. Target-aware tracking with long-term context + +attention. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):773-780, 2023. 2 +[23] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3 +[24] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR, 2022. 3, 6 +[25] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5 +[26] Yuqing Huang, Xin Li, Zikun Zhou, Yaowei Wang, Zhenyu He, and Ming-Hsuan Yang. Rtracker: Recoverable tracking via pn tree structured memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19038-19047, 2024. 6 +[27] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2 +[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3 +[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6 +[30] Xin Li, Yuqing Huang, Zhenyu He, Yaowei Wang, Huchuan Lu, and Ming-Hsuan Yang. Citetracker: Correlating image and text for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9974-9983, 2023. 6 +[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 3 +[32] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. Advances in Neural Information Processing Systems, 35:16743-16754, 2022. 3 +[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. In Computer Vision - ECCV 2024, pages 300-318, Cham, 2025. Springer Nature Switzerland. 1, 3, 4, 5, 6 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5 + +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR, 2019. 5 +[36] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2 +[37] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, pages 445-461, 2016. 6 +[38] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, pages 300-317, 2018. 2, 5 +[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 5 +[40] Jonas Pfeiffer, Andreas Rücklé, Clifton Poth, Aishwarya Kamath, Ivan Vulić, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. Adapterhub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, 2020. 3 +[41] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2 +[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5 +[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems, pages 8583-8595. Curran Associates, Inc., 2021. 2 +[44] Zhangyong Tang, Tianyang Xu, Zhenhua Feng, Xuefeng Zhu, He Wang, Pengcheng Shao, Chunyang Cheng, Xiao-Jun Wu, Muhammad Awais, Sara Atito, et al. Revisiting rgbt tracking benchmarks from the perspective of modality validity: A new benchmark, problem, and method. arXiv preprint arXiv:2405.00168, 2024. 2 +[45] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. Hydralora: An asymmetric lora architecture for efficient fine-tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 1 +[47] Guangting Wang, Chong Luo, Xiaoyan Sun, Zhiwei Xiong, + +and Wenjun Zeng. Tracking by instance detection: A meta-learning approach. In CVPR, pages 6288-6297, 2020. 5 +[48] Ke Wang, Jiahui Zhu, Minjie Ren, Zeming Liu, Shiwei Li, Zongye Zhang, Chenkai Zhang, Xiaoyu Wu, Qiqi Zhan, Qingjie Liu, et al. A survey on data synthesis and augmentation for large language models. arXiv preprint arXiv:2410.12896, 2024. 2 +[49] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13763-13773, 2021. 2, 6 +[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3 +[51] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9697-9706, 2023. 2, 5, 6, 7 +[52] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, 2023. 1, 7 +[53] Xun Wu, Shaohan Huang, and Furu Wei. Mixture of lora experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2 +[54] Y. Wu, J. Lim, and M. Yang. Object tracking benchmark. TPAMI, 37(9):1834-1848, 2015. 6 +[55] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8751-8760, 2022. 2 +[56] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19300-19309, 2024. 5, 6 +[57] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6 +[58] Dawei Yang, Jianfeng He, Yinchao Ma, Qianjin Yu, and Tianzhu Zhang. Foreground-background distribution modeling transformer for visual object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10117-10127, 2023. 1, 2, 6 +[59] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for + +tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 1, 2, 5, 6 +[60] Yihua Zhang, Ruisi Cai, Tianlong Chen, Guanhua Zhang, Huan Zhang, Pin-Yu Chen, Shiyu Chang, Zhangyang Wang, and Sijia Liu. Robust mixture-of-expert training for convolutional neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 90-101, 2023. 2 +[61] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7588-7596, 2024. 1, 2, 3, 5, 6, 7 +[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3 \ No newline at end of file diff --git a/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/images.zip b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e0d1dc7873751c41701d088a43c9d00b49a7b83d --- /dev/null +++ b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc729b93adf060b58f84ec1af6ed20f45a846442d0126984ba6b44d7f714e9bb +size 770307 diff --git a/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/layout.json b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..199cd47f7d7e429f5a81d543619b76e0c88e6b8a --- /dev/null +++ b/2025/SPMTrack_ Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking/layout.json @@ -0,0 +1,8816 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 102, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 102, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 515, + 140 + ], + "type": "text", + "content": "SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "content": "Wenrui Cai" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "content": ", Qingjie Liu" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "content": ", Yunhong Wang" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "content": "State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China \n" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "content": "Zhongguancun Laboratory, Beijing, China \n" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 556, + 233 + ], + "type": "text", + "content": "Hangzhou Innovation Institute, Beihang University, Hangzhou, China \n{wenrui_cai, qingjie.liu, yhwang}@buaa.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 297, + 597 + ], + "type": "text", + "content": "Most state-of-the-art trackers adopt one-stream paradigm, using a single Vision Transformer for joint feature extraction and relation modeling of template and search region images. However, relation modeling between different image patches exhibits significant variations. For instance, background regions dominated by target-irrelevant information require reduced attention allocation, while foreground, particularly boundary areas, need to be emphasized. A single model may not effectively handle all kinds of relation modeling simultaneously. In this paper, we propose a novel tracker called SPMTrack based on mixture-of-experts tailored for visual tracking task (TMoE), combining the capability of multiple experts to handle diverse relation modeling more flexibly. Benefiting from TMoE, we extend relation modeling from image pairs to spatio-temporal context, further improving tracking accuracy with minimal increase in model parameters. Moreover, we employ TMoE as a parameter-efficient fine-tuning method, substantially reducing trainable parameters, which enables us to train SPMTrack of varying scales efficiently and preserve the generalization ability of pretrained models to achieve superior performance. We conduct experiments on seven datasets, and experimental results demonstrate that our method significantly outperforms current state-of-the-art trackers. The source code is available at https://github.com/WenRuiCai/SPMTrack." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 622, + 137, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 137, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 137, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "content": "Visual tracking aims to predict the location of a target throughout subsequent frames of a video, given the initial state in the first frame template. Currently, most state-of-the-art trackers adopt a Transformer-based [46] one-stream" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 257, + 552, + 443 + ], + "blocks": [ + { + "bbox": [ + 318, + 257, + 552, + 443 + ], + "lines": [ + { + "bbox": [ + 318, + 257, + 552, + 443 + ], + "spans": [ + { + "bbox": [ + 318, + 257, + 552, + 443 + ], + "type": "image", + "image_path": "71693ee43b859c952338b7d95fcddc5a505a877abccf99af03e02f680d0d802d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. Comparison of LaSOT AUC and model parameter count across different trackers. Larger loop indicates better performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 495, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 556 + ], + "type": "text", + "content": "paradigm [3, 5, 10, 33, 52, 59, 61], utilizing a single Vision Transformer [12] backbone that accepts both the template and the search region image as input, and employs self-attention to perform image feature extraction and template-search region relation modeling simultaneously." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 700 + ], + "type": "text", + "content": "However, not all relation modeling between image patches benefits tracking performance. Prior studies [4, 20, 58, 59] have demonstrated that the abundance of background patches in the search region deteriorates the discriminative power of the foreground features, and single vanilla attention has limited capability in suppressing undesirable foreground-background interactions. Meanwhile, there are also some kinds of relation modeling that demonstrate positive effects on tracking performance [17, 19]. For instance, enhancing attention to target boundary regions can improve tracking performance [17], especially in challenging scenarios such as motion blur, partial occlusion, and deformation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "To address various relation modeling, numerous meth" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 149, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16871" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "ods propose specific designs, including background filtering to reduce target-irrelevant information [3, 59], customizing different interaction modules for tokens belonging to different categories [4, 20], and attention score adjustment based on foreground-background weights of each token [58]. However, existing methods can only handle one or a few predefined types of relation modeling, and the handcrafted strategies inherently lack adaptability. Furthermore, there may exist many latent relationships that further complicate the specialized design. These limitations motivate our question: How to design a tracker to adaptively process various relation modeling between image patches?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 220, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 518 + ], + "type": "text", + "content": "Inspired by mixture of experts (MoE) [1, 11, 13, 27, 41] in natural language processing, in this paper, we address the question by proposing SPMTrack, a novel tracker based on TMoE, a specialized mixture of experts module for visual tracking task. Similar to MoE, TMoE handles diverse relation modeling through adaptive weighted combinations of multiple experts. However, unlike traditional MoE that is exclusively applied to feed-forward network (FFN) layers, TMoE is simultaneously applied to the linear layers within both the attention layers and the FFN layers. This finer-grained application enables TMoE to achieve more diverse expert combinations, thereby enhancing the capability of various relation modeling. Moreover, differing from traditional MoE that uses FFNs as experts, all the experts in TMoE are linear layers and designed with a lightweight and efficient structure to ensure overall efficiency. Furthermore, we employ TMoE as a parameter-efficient finetuning method, which enables us to train SPMTrack of larger scales. In SPMTrack, only a subset of experts and the router within TMoE, along with the prediction head need to be trained, which substantially reduces the trainable parameters while preserving the generalization capability of pretrained models. Notably, our method reduces the number of trainable parameters by nearly " + }, + { + "bbox": [ + 56, + 220, + 296, + 518 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 56, + 220, + 296, + 518 + ], + "type": "text", + "content": " compared to [2], while demonstrates better performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Benefiting from the powerful capability of TMoE in handling diverse relation modeling, unlike previous one-stream trackers that only perform relation modeling between template-search region pairs, we extend SPMTrack to incorporate multi-frame spatio-temporal context modeling, which further enhances the performance of SPMTrack with minimal additional parameters. We evaluated the performance of our method on seven datasets, and the experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets including LaSOT [14], GOT-10K [25], TrackingNet [38] and TNL2K [49]. Through extensive experiments with models of varying scales, as shown in Figure 1, Table 1 and Table 2, with ViT-B [12] as backbone and less than " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " of parameters need to be trained, SPMTrack-B achieves or even surpasses the best performance of previous trackers utilizing ViT-L." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "The main contributions of this paper can be summarized as: (1) We propose TMoE, a mixture of experts module tailored for visual tracking. TMoE enhances the capability to handle various relation modeling and can be used as a method of parameter-efficient fine-tuning. (2) Based on TMoE, we propose SPMTrack, a novel tracker that can effectively integrate spatio-temporal context for visual tracking. (3) Experimental results demonstrate that SPMTrack achieves state-of-the-art performance across multiple datasets. We trained SPMTrack of varying scales, and our method with ViT-B as backbone can achieve or even surpass current trackers using ViT-L." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 228, + 400, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 400, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 400, + 241 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 249, + 440, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 440, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 440, + 261 + ], + "type": "text", + "content": "2.1. One-Stream Trackers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 267, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 556, + 495 + ], + "type": "text", + "content": "One-stream trackers employ a single Vision Transformer [12] as backbone, simultaneously performing feature extraction and relation modeling at each layer [5, 10, 55, 59]. Many efforts focus on incorporating historical context. TATrack [22] and ODTrack [61] propose architectures capable of processing multiple frames, while HIPTrack [3] introduces historical target features through prompt learning. SeqTrack [6], ARTrack[51], and ARTrackV2 [2] achieve more precise predictions by incorporating multiple historical target position. Other works try to improve conventional self-attention to handle different relation modeling. OSTrack [59] progressively filters background regions to reduce foreground-background relation modeling, GRM [20] and ROMTrack [4] categorizes image tokens and constrains relation modeling between specific categories, F-BDMTrack [58] adjusts the attention score by calculating the foreground-background weights of tokens. But these methods can only handle specific relation modeling and rely on manually designed strategies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 504, + 427, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 427, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 427, + 517 + ], + "type": "text", + "content": "2.2. Mixture of Experts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Mixture-of-experts (MoE) [1, 13, 27] are widely applied in large language models [9, 11, 15, 36, 41, 53], where each expert focuses on specific aspects of the data or particular tasks [48]. By performing an adaptive weighted sum over multiple experts, MoE can better capture complex patterns and relationships within the input data. MoE also has applications in computer vision [8, 43, 60]. However, there is few dedicated exploration of MoE in visual tracking. MoETrack [44] employs multiple prediction heads as experts in the field of RGB-T tracking, which is rather far-fetched and deviates from the general practice of embedding MoE within Transformer layers. In the field of RGB-E tracking, eMoETracker [7] inserts multiple experts between Transformer layers, which more closely resembles an adapter architecture [23] and contrasts with MoE implementations where experts are embedded within Transformer layers." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16872" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 550, + 297 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 550, + 297 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 550, + 297 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 550, + 297 + ], + "type": "image", + "image_path": "2ad4663e1941ba435163c2dc67801f2da2cfb13becb701174fd848c1830728d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 322 + ], + "type": "text", + "content": "Figure 2. Overview of SPMTrack that consists of a feature extraction network and a prediction head. The main body of the feature extraction network is a Transformer encoder composed of multiple TMoEBlocks. The structure of TMoEBlock is shown on the right side." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 342, + 228, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 228, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 228, + 357 + ], + "type": "text", + "content": "2.3. Parameter-efficient Fine-tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 506 + ], + "type": "text", + "content": "Parameter-efficient fine-tuning aims to reduce computational costs and retain the general knowledge of the pretrained model by freezing most parameters and fine-tuning only a small subset of parameters. Common parameter-efficient fine-tuning approaches include adapter-based methods [18, 23, 40], prompt-based methods [28, 31, 50, 62] and low-rank adaptation methods [24, 45]. In visual tracking, HIPTrack [3] incorporates historical target features as prompts, while LoRAT [33] employs low-rank adaptation for efficient training. And our method maintains flexibility in diverse relation modeling while reducing trainable parameters by fine-tuning TMoE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 112, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 112, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 112, + 530 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 538, + 176, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 176, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 176, + 551 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 629 + ], + "type": "text", + "content": "As illustrated in Figure 2, we propose SPMTrack, a novel tracker that integrates multi-frame spatio-temporal context and employs TMoE for parameter-efficient fine-tuning. The overall architecture follows the one-stream paradigm, consisting of a Feature Extraction Network based on a Vision Transformer [12] as backbone, and a prediction head." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Feature Extraction Network. The feature extraction network is denoted as " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". The main body of the feature extraction network is a Transformer encoder that utilizes TMoE-Block as its primary component, which is designed based on TMoE. The architecture of TMoE-Block is illustrated in the right panel of Figure 2, and details will be introduced in Section 3.2. The encoder simultaneously per" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": "forms feature extraction and relation modeling for input images. We extend the input from one single template " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{H_z \\times W_z \\times 3}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": " that is commonly used in one-stream trackers to the video sequence " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\{Z_1, \\dots, Z_N | Z_i \\in \\mathbb{R}^{H_z \\times W_z \\times 3}\\}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": " as reference frames. The reference frames contain the first frame template and the selected tracked historical frames, with a total of " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": " frames. The feature extraction network uses patch embedding to partition each input image into patches of size " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "M \\times M" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": ", map all patches into tokens by a convolutional layer, and obtain the reference token sequence of each image " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "T_i \\in \\mathbb{R}^{N_T \\times d}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": " denotes the hidden dimension of Transformer encoder and " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "N_T = \\frac{H_z W_z}{M^2}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": " represents the number of tokens corresponding to each image. Similarly, for the search region " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{H_s \\times W_s \\times 3}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": ", we can also obtain the token sequence " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N_X \\times d}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": ". All tokens are added with positional embedding and a learnable token type embedding [33]. Inspired by [32, 61], we additionally introduce a target state token " + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 313, + 342, + 555, + 583 + ], + "type": "text", + "content": " that is learnable as well and propagates over time. All tokens are concatenated and fed into the Transformer encoder. The process can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 594, + 554, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 594, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 326, + 594, + 554, + 628 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} _ {i} ^ {j} = \\left\\{ \\begin{array}{l} \\boldsymbol {T} _ {i} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {o}, i f \\text {t o k e n} (i, j) \\text {i n b b o x} \\\\ \\boldsymbol {T} _ {i} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {b}, o t h e r w i s e \\end{array} \\right. \\tag {1}", + "image_path": "1587ee00a55e6e6ddd846250a0b850da0a752f4abe4e2a8c08a87695cafa9c39.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 627, + 452, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 627, + 452, + 638 + ], + "spans": [ + { + "bbox": [ + 350, + 627, + 452, + 638 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {j} = \\boldsymbol {X} ^ {j} + \\boldsymbol {P} \\boldsymbol {E} _ {j} + \\boldsymbol {T} \\boldsymbol {E} _ {S}", + "image_path": "95d46089ec3e5b0d4ad510676da9725de39e27bf8cc0cfd7229de96091692566.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 640, + 471, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 640, + 471, + 652 + ], + "spans": [ + { + "bbox": [ + 349, + 640, + 471, + 652 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} = \\operatorname {C o n c a t} (\\boldsymbol {H}, \\boldsymbol {T} _ {1}, \\dots , \\boldsymbol {T} _ {N}, \\boldsymbol {X})", + "image_path": "576c8db88d48d197dc57bda68e715526b2af626c3023b988bc99ed677509d13a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "T_{i}^{j}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " token in " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "PE_{j}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " represents positional embedding of the " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " token, similarly for " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "X^{j}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "TE_{o}, TE_{b}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "TE_{S}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " represents the learnable type embeddings of the foreground region tokens, background region" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16873" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "tokens in reference frames and the search region tokens, respectively. " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " denotes the input to the Transformer encoder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "content": "Subsequently, the Transformer encoder processes all tokens and produces an output " + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{O} \\in \\mathbb{R}^{(1 + N \\times N_T + N_X) \\times d}" + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "content": ". The output tokens " + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{X}' \\in \\mathbb{R}^{N_X \\times d}" + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "content": " corresponding to search region and the output " + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{H}' \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "content": " from the target state token are fed into the prediction head for prediction. " + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{H}'" + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "content": " will be added with the input target state token in the next tracking frame, which can be formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 186, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 186, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 88, + 186, + 295, + 199 + ], + "type": "interline_equation", + "content": "\\boldsymbol {O} = \\Phi (\\boldsymbol {I}), [ \\boldsymbol {H} ^ {\\prime}, \\boldsymbol {T} _ {1} ^ {\\prime}, \\dots , \\boldsymbol {T} _ {N} ^ {\\prime}, \\boldsymbol {X} ^ {\\prime} ] = \\operatorname {S p l i t} (\\boldsymbol {O}) \\tag {2}", + "image_path": "f9b6d6b9785b9070505b2841ab882f6e46f1221ef5d7183531602335131b3010.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": "Prediction Head. The input to the prediction head consists of the output target state token " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "H'" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": " and the output search region tokens " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": " from the feature extraction network. Initially, the prediction head performs matrix multiplication between " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "H'" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": " to obtain the weight " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{N_X \\times 1}" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": ". As the target state token propagates over time, " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": " can reweight and adjust current search region tokens based on the historical state of the target, thereby introducing more spatio-temporal tracking context. After element-wise multiplication of " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": ", the resulting data is reshaped into a feature map " + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{\\frac{H_s}{M} \\times \\frac{W_s}{M} \\times d}" + }, + { + "bbox": [ + 55, + 201, + 296, + 370 + ], + "type": "text", + "content": ", which is then processed by a decoupled double-MLP head to perform target center classification and box regression, respectively. For fair comparison, the double-MLP head is identical to [33]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 203, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 203, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 203, + 388 + ], + "type": "text", + "content": "3.2. The Design of TMoEBlock" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 525 + ], + "type": "text", + "content": "TMoEBlock serves as the primary component of the Transformer encoder in the feature extraction network. As illustrated in the right of Figure 2, TMoEBlock is built based on the standard vision Transformer block. However, unlike the standard ViT block, TMoEBlock applies TMoE module to both the multi-head self-attention (MSA) layer and the feed-forward network (FFN) layer, obtaining MoE-based MSA " + }, + { + "bbox": [ + 55, + 393, + 296, + 525 + ], + "type": "inline_equation", + "content": "(\\mathrm{M}^2\\mathrm{SA})" + }, + { + "bbox": [ + 55, + 393, + 296, + 525 + ], + "type": "text", + "content": " layer and MoE-based FFN (MFFN) layer. The finer-grained application of TMoE is capable to achieve more diverse combinations for various relation modeling. The processing of TMoEBlock can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 539, + 295, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 539, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 84, + 539, + 295, + 566 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {O} _ {l} ^ {\\prime} = \\mathrm {M} ^ {2} \\operatorname {S A} (\\ln (\\boldsymbol {O} _ {l - 1})) + \\boldsymbol {O} _ {l - 1}, \\quad l = 1, \\dots , L \\tag {3} \\\\ \\boldsymbol {O} _ {l} = \\mathrm {M F F N} (\\mathrm {L N} (\\boldsymbol {O} _ {l} ^ {\\prime})) + \\boldsymbol {O} _ {l} ^ {\\prime}, l = 1, \\dots , L \\\\ \\end{array}", + "image_path": "4747eb3ee95d3ef7ae1c6c98261e194d8d52110ff015df441c2285a5caa02430.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " denotes the total number of TMoEBlocks, " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{O}_l" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " represents the output of the " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " block in Transformer encoder. When " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "l = L" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{O}_l" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " is the output " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{O}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " of the feature extraction network; when " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "l = 0" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": ", it is the overall input " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": ". LN stands for the layer normalization operation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "MoE-based Multi-head Self-Attention (M²SA). M²SA not only replaces the standard three linear projection layers " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": " in standard attention for the query, key, and value with TMoE modules but also replaces the output projection layer with TMoE module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "MoE-based Feed-Forward Network (MFFN). Conventional feed-forward network consists of two linear layer and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 549, + 189 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 549, + 189 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 549, + 189 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 549, + 189 + ], + "type": "image", + "image_path": "5594208d4c14b2f214b8902ee400b433966b6c8af60d984be909b91dc960d36b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 216 + ], + "type": "text", + "content": "Figure 3. The structure of TMoE. The symbols maintain the same meaning with Figure 2." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "type": "text", + "content": "activation function. MFFN replaces both linear layers in FFN with TMoE modules, while maintaining all other architectural configurations unchanged." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 520, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 520, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 520, + 287 + ], + "type": "text", + "content": "3.3. Mixture of Experts for Visual Tracking" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 291, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 554, + 411 + ], + "type": "text", + "content": "In this section, we introduce TMoE, a mixture of experts module tailored for visual tracking task. As illustrated in Figure 3, TMoE consists of one shared expert, one router, one compression expert and " + }, + { + "bbox": [ + 313, + 291, + 554, + 411 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 313, + 291, + 554, + 411 + ], + "type": "text", + "content": " routed experts. In large language models, mixture of experts are commonly applied to replace FFN layers, and the experts comprising MoE are typically the same structure as FFN. However, TMoE is used to replace the linear layers in both self-attention and FFN, and the shared expert, the compression expert and the routed experts are all implemented as linear layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 494 + ], + "type": "text", + "content": "In TMoE, the shared expert directly copies weights from the corresponding linear layers in pretrained model and remains frozen during training. The rationale is that the original weights have been well-trained during the pre-training stage, and the shared expert needs to retain more general knowledge. In contrast, the compression experts, routed experts, and the router are all trainable." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": "Given a TMOE module replacing a linear layer with input dimension " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " and output dimension " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ", for each token " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ", as shown in Figure 3, TMOE first computes the weights " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{W} \\in \\mathbb{R}^{N_e}" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " of all routed experts based on the input " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ", the weights are normalized by Softmax. After that, TMOE calculates the output " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{y}^s \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " of the shared expert and the outputs " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^e \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " of all routed experts. In particular, instead of adopting the same structure as the shared expert, inspired by [45], we first employ a compression expert with a " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "d \\times r" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " linear transformation to reduce the input dimension from " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " and obtain the compressed result " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{y}^c \\in \\mathbb{R}^r" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "r \\ll d" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "r \\ll D" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ". Subsequently, all routed experts use the compressed input " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{y}^c" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": " to produce their respective outputs " + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^e \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 495, + 555, + 663 + ], + "type": "text", + "content": ". The process can be formally described as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 670, + 520, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 670, + 520, + 699 + ], + "spans": [ + { + "bbox": [ + 347, + 670, + 520, + 699 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {W} = \\text {S o f t m a x} (\\operatorname {R o u t e r} (\\boldsymbol {x})), \\quad \\boldsymbol {y} ^ {s} = \\boldsymbol {E} ^ {S} (\\boldsymbol {x}) \\\\ \\boldsymbol {y} ^ {c} = \\boldsymbol {E} ^ {C} (\\boldsymbol {x}), \\quad \\boldsymbol {y} _ {i} ^ {e} = \\boldsymbol {E} _ {i} ^ {R} (\\boldsymbol {y} ^ {c}) \\quad i = 1, \\dots , N _ {e} \\\\ \\end{array}", + "image_path": "5a0d9c73789025872a5c060c57099e6552146230f42f3f59f33a5d3f18e5a4d6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "where Router is also a linear layer that maps the input " + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "x" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16874" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "content": "to weights " + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\pmb{W}" + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{S}" + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\pmb{E}^{C}" + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{i}^{R}" + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "content": " denote the shared expert, compress expert, and " + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 71, + 295, + 95 + ], + "type": "text", + "content": " routed expert, respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": "After obtaining the outputs of all experts, we compute a weighted sum of the outputs " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^e" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " from all routed expert based on " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{W}" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y}^e \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": ", which means that the routed experts are densely activated in TMoE. Subsequently, " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y}^e" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " is added with the shared expert output " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y}^s" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " to get the final output " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": ". The process can be formulated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 177, + 295, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 177, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 107, + 177, + 295, + 206 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} ^ {e} = \\sum_ {i = 1} ^ {N _ {e}} \\boldsymbol {W} _ {i} \\boldsymbol {y} _ {i} ^ {e}, \\quad \\boldsymbol {y} = \\boldsymbol {y} ^ {e} + \\boldsymbol {y} ^ {s} \\tag {5}", + "image_path": "4e0bd151b1510fba45c9bea6b1ea25401072e02b4573d9f87f584041667ff691.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 144, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 144, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 144, + 224 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 229, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 291 + ], + "type": "text", + "content": "The outputs of the decoupled double-MLP prediction head are supervised using binary cross-entropy loss for target center classification and Generalized IoU loss [42] for bounding box regression, respectively. Both loss terms are assigned with the weighting coefficients set to 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 137, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 137, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 137, + 313 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 319, + 188, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 188, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 188, + 331 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": "Model settings. We provide three versions of SPMTrack: SPMTrack-B, SPMTrack-L, and SPMTrack-G, utilizing ViT-B [12], ViT-L [12], and ViT-G [39] as backbones respectively. All versions maintain a consistent patch size " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " of 14 and adopt the pretrained weights from DINOv2 [39]. The resolution of input images are the same across all versions, with reference frames of " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "196 \\times 196" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " and search region of " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "378 \\times 378" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": ". A crop factor of 2 is applied to reference frames and 5 to search region. We set the number of reference frames " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " to 3 for spatio-temporal modeling. Each TMoE module incorporates 4 routed experts, where the input dimension of the routed experts " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " is set to 64. Other configurations and model information are shown in Table 1, where " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " is the number of TMoEBlocks, " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " denotes hidden dimension and " + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "inline_equation", + "content": "N_h" + }, + { + "bbox": [ + 55, + 336, + 296, + 515 + ], + "type": "text", + "content": " is the number of attention heads." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 56, + 556, + 293, + 642 + ], + "blocks": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": "Table 1. Model configurations and parameter counts across SPM-Track variants of different scales." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 556, + 293, + 642 + ], + "lines": [ + { + "bbox": [ + 56, + 556, + 293, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 293, + 642 + ], + "type": "table", + "html": "
SPMTrack-BSPMTrack-LSPMTrack-G
BackboneL=12L=24L=40
d=768d=1024d=1536
Nh=12Nh=16Nh=24
#Params (M)115.3379.61339.5
#Trainable \nParams (M)29.275.9204.0
", + "image_path": "52fb3ba64ffb8c0cd46e6d57ccd64d59f56083ad583c29e4d88b5c2308953b22.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Datasets. Following previous works [3, 33, 51, 59], we use LaSOT [14], TrackingNet [47], GOT-10K [25], and COCO [34] for training. When evaluating on GOT-10K test set, we strictly follow the protocol by training exclusively on GOT-10K training set. When evaluating on other datasets," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "we jointly train on the training sets of all four datasets, randomly sampling from each dataset with equal probability. For video datasets, we sample 5 frames from one video with random frame intervals ranging from 1 to 200. The sampled frames are arranged either in forward or reverse order with equal probability, where the first 3 frames serve as reference frames and the remaining 2 as search frames. For COCO, we replicate each image 5 times to maintain consistency with the video frame sampling strategy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "text", + "content": "Training and Optimization. Our method is implemented based on PyTorch 2.3.1 and trained on 8 NVIDIA A100 GPUs. We maintain a global batch size of 128 during training. When training exclusively on GOT-10K, we train for 100 epochs. When joint training on multiple datasets, we extend to 170 epochs. In each epoch, we sample 131,072 sequences. We employ AdamW [35] as optimizer with learning rate of " + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "text", + "content": " and weight decay of 0.1. The learning rate scheduler warms up over the initial 2 epochs from " + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "inline_equation", + "content": "10^{-7}" + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "text", + "content": " and then follows a cosine schedule to " + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 313, + 180, + 554, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "text", + "content": "Inference. We use a total of 3 reference frames, including the annotated first frame template. When current search frame index " + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "inline_equation", + "content": "t \\leq 3" + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "text", + "content": ", we use all available tracked frames as reference frames, if there are fewer than three frames, we repeat the template as necessary to reach the required number. When " + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "inline_equation", + "content": "t > 3" + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "text", + "content": ", we select two tracked frames as reference frames at indexes " + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{t}{3} \\right\\rfloor" + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{2t}{3} \\right\\rfloor" + }, + { + "bbox": [ + 313, + 300, + 554, + 409 + ], + "type": "text", + "content": ". Following previous trackers [4, 59, 61], we apply a Hanning window penalty to the output response map of the classification head." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 418, + 517, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 517, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 517, + 430 + ], + "type": "text", + "content": "4.2. Comparisons with the State-of-the-Art" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 436, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 568 + ], + "type": "text", + "content": "LaSOT [14] is a large-scale long-term tracking dataset with its test set containing 280 sequences. As shown in Table 2, SPMTrack-B significantly outperforms all trackers utilizing ViT-B as backbone. Our method also substantially surpasses LoRAT [33] that also employs parameter-efficient fine-tuning. Additionally, as shown in Figure 4, we evaluate tracker performance across various challenging scenarios in LaSOT. Our method surpasses previous trackers that also incorporate spatio-temporal context [2, 3, 6, 51, 56], while significantly outperforming ROMTrack [4] and GRM [20], which focus on optimizing relation modeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 569, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 653 + ], + "type": "text", + "content": "GOT-10K [25] contains 9,335 sequences for training and 180 sequences for testing. We follow the protocol that trackers are only allowed to be trained using GOT-10K train split. Our approach significantly outperforms LoRAT [33] with the same backbone. And our SPMTrack-B rivals the performance of the current state-of-the-art methods, while SPMTrack-L substantially surpasses all existing trackers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "TrackingNet [38] is a large-scale visual tracking dataset with a test set comprising 511 video sequences. Table 2 demonstrates that our method achieves state-of-the-art performance using only ViT-B as backbone, surpassing current trackers that utilize ViT-L or even ViT-G." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16875" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 558, + 333 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. State-of-the-art comparison on LaSOT, GOT-10k and TrackingNet. \\* denotes for trackers trained only with GOT-10k train split. The best three results are highlighted in red, blue and bold, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 558, + 333 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 558, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 558, + 333 + ], + "type": "table", + "html": "
MethodSourceLaSOTGOT-10k*TrackingNet
AUC(%)PNorm(%)P(%)AO(%)SR0.5(%)SR0.75(%)AUC(%)PNorm(%)P(%)
SPMTrack-BOurs74.984.081.776.585.976.386.190.285.6
SPMTrack-LOurs76.885.984.080.089.479.986.991.087.2
SPMTrack-GOurs77.486.685.081.089.282.387.391.488.1
LoRAT-B378 [33]ECCV2472.981.979.173.782.672.984.288.483.0
LoRAT-L378 [33]ECCV2475.184.182.077.586.278.185.689.785.4
LoRAT-G378 [33]ECCV2476.285.383.578.987.880.786.090.286.1
AQATrack384 [56]CVPR2472.782.980.276.085.274.984.889.384.3
ARTrackV2-B384 [2]CVPR2473.082.079.677.586.075.585.789.885.5
ARTrackV2-L384 [2]CVPR2473.682.881.179.587.879.686.190.486.2
HIPTrack [3]CVPR2472.782.979.577.488.074.584.589.183.8
ODTrack-B [61]AAAI2473.283.280.677.087.975.185.190.184.9
F-BDMTrack-384 [58]ICCV2372.081.577.775.484.372.984.589.084.0
ROMTrack-384 [4]ICCV2371.481.478.274.284.372.484.189.083.7
ARTrack384 [51]CVPR2372.681.779.175.584.374.385.189.184.8
SeqTrack-B384 [6]CVPR2371.581.177.874.584.371.483.988.883.6
GRM [20]CVPR2369.979.375.873.482.970.484.088.783.3
OSTrack384 [59]ECCV2271.181.177.673.783.270.883.988.583.2
SimTrack [5]ECCV2270.579.7-69.878.866.083.487.4-
", + "image_path": "c4a85a07b3c54a00ea7ba6f83527969b007bd68fbef911f0c69bc014147f126a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 346, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 346, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 294, + 407 + ], + "type": "text", + "content": "TNL2K [49] comprises 700 test video sequences, each accompanied by natural language descriptions. As shown in Table 3, our method achieves state-of-the-art performance without utilizing natural language descriptions and surpasses existing vision-language trackers [21, 30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 407, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 295, + 479 + ], + "type": "text", + "content": "UAV123 [37] is a dataset focusing on low-altitude drone aerial scenarios, consisting of 123 video sequences with an average length of 915 frames per sequence. Our method achieves state-of-the-art performance with comparable model sizes, demonstrating the capability of our approach in aerial scenes and small object tracking." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 479, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 295, + 563 + ], + "type": "text", + "content": "NfS [29] consists of 100 video sequences totaling 380,000 frames. Following previous works, we conduct evaluations on the 30 FPS version of the dataset. Our method rivals current state-of-the-art trackers in NfS. One possible reason for not achieving further improvement is that our method does not incorporate consecutive target position like ARTrack [51] and HIPTrack [3], which is crucial for NfS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 563, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 563, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 563, + 295, + 624 + ], + "type": "text", + "content": "OTB2015 [54] is a classic dataset in visual tracking, consisting of 100 short-term tracking sequences that encompass 11 common challenges, such as target deformation, occlusion, and scale variation. Our method also achieves state-of-the-art performance on OTB2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 715 + ], + "type": "text", + "content": "The Importance of Spatio-Temporal Context Modeling and TMoE. Table 5 demonstrates the significance of two core designs in our paper: spatio-temporal context modeling and TMoE. In Table 5, removing TMoE indicates direct fine-tuning with LoRA [24], while removing spatio" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 325, + 387, + 542, + 525 + ], + "blocks": [ + { + "bbox": [ + 313, + 345, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 378 + ], + "type": "text", + "content": "Table 3. The performance of our method and other state-of-the-art trackers on TNL2K. The best three results are highlighted in red, blue and bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 387, + 542, + 525 + ], + "lines": [ + { + "bbox": [ + 325, + 387, + 542, + 525 + ], + "spans": [ + { + "bbox": [ + 325, + 387, + 542, + 525 + ], + "type": "table", + "html": "
MethodAUC(%)PNorm(%)P(%)
SPMTrack-G64.782.670.6
SPMTrack-L63.781.569.2
SPMTrack-B62.079.766.7
LoRAT-B378 [33]59.9-63.7
ODTrack-L [61]61.7--
ARTrackV2-L384 [2]61.6--
ARTrack-L384 [51]60.3--
RTracker-L [26]60.6-63.7
UNINEXT-H [57]59.3-62.8
CiteTracker [30]57.7-59.6
VLT [21]53.1-53.3
", + "image_path": "9baf13b17bc08d8893cff5fac2bb7c1c849891d1802b41ce7154c0e4565c515e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "text", + "content": "temporal modeling indicates using only the first frame template and search region as input and eliminating the target state token. Compared to LoRA, applying TMoE for parameter-efficient fine-tuning yields a substantial improvement of " + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "inline_equation", + "content": "+0.8" + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "text", + "content": " AUC on LaSOT, demonstrates the superiority of TMoE. And the spatio-temporal context modeling further enhances performance by an additional " + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "inline_equation", + "content": "+1.2" + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "text", + "content": " AUC. On the basis of incorporating spatio-temporal modeling, the improvement brought by TMoE is " + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "inline_equation", + "content": "+1.0" + }, + { + "bbox": [ + 312, + 545, + 555, + 676 + ], + "type": "text", + "content": " AUC, indicating that TMoE can help the model better capture spatio-temporal contextual information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Where to Apply TMoE. Conventional MoE is typically applied to the FFN layers within Transformer blocks, while keeping the attention layers unchanged. In contrast, we ap" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16876" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 298, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 298, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 298, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 298, + 300 + ], + "type": "image", + "image_path": "774929da7c6c25f4145a472543c9c363e124168b2dafea5e572e6a053210cd1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 297, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 297, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 297, + 343 + ], + "type": "text", + "content": "Figure 4. The performance of our method compared with other state-of-the-art trackers in terms of AUC across various scenarios in the LaSOT test split." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 396, + 287, + 513 + ], + "blocks": [ + { + "bbox": [ + 55, + 354, + 296, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 389 + ], + "type": "text", + "content": "Table 4. The performance of our method and other state-of-the-art trackers on UAV123, NfS and OTB2015 in terms of AUC metrics. The best three results are highlighted in red, blue and bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 396, + 287, + 513 + ], + "lines": [ + { + "bbox": [ + 65, + 396, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 65, + 396, + 287, + 513 + ], + "type": "table", + "html": "
MethodUAV123NfSOTB2015
SPMTrack-B71.767.472.7
HIPTrack [3]70.568.171.0
ARTrackV2-B [2]69.967.6-
ODTrack-L [61]--72.4
ARTrack384 [51]70.566.8-
SeqTrack-B384 [6]68.666.7-
DropTrack [52]--69.6
MixFormer-L [10]69.5--
STMTrack [16]64.7-71.9
", + "image_path": "5f9ed2cbb13512b879840268da65c63e07e398a4a9ba4ad9bc1bb0bb3d7309f8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 554, + 293, + 617 + ], + "blocks": [ + { + "bbox": [ + 55, + 523, + 295, + 546 + ], + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 546 + ], + "type": "text", + "content": "Table 5. Ablation studies on spatio-temporal context modeling and TMoE. Experiments are conducted on LaSOT." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 554, + 293, + 617 + ], + "lines": [ + { + "bbox": [ + 57, + 554, + 293, + 617 + ], + "spans": [ + { + "bbox": [ + 57, + 554, + 293, + 617 + ], + "type": "table", + "html": "
#Spatio-TemporalTMoEAUC (%)\\(P_{Norm}(\\%)\\)P(%)
1✘✘72.981.979.1
2✘✓73.782.780.0
3✓✘73.982.880.0
4✓✓74.984.081.7
", + "image_path": "162805f52d56196842e39b329316bef99c5965d97777f55456d537836b50c642.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "ply TMoE in both the attention layers and the FFN layers. Table 6 demonstrates the impact of applying TMoE in different layers. Comparing the first, third, and fifth rows, applying TMoE to the query, key, value projections in attention layers and to FFN layers can both improve performance. Comparing the first, second and third rows, we find that applying TMoE in attention layers yields significantly" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": "greater benefits compared to applying it only in FFN layers. Comparing the third and fourth rows, applying TMoE in the output projection matrix of the attention layer can bring a slight improvement." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 173, + 553, + 246 + ], + "blocks": [ + { + "bbox": [ + 313, + 130, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 130, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 555, + 163 + ], + "type": "text", + "content": "Table 6. Ablation studies on applying TMoE in different layers. All results are evaluated on LaSOT test split. “X” means replacing TMoE with LoRA." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 173, + 553, + 246 + ], + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 246 + ], + "type": "table", + "html": "
#QKVOut ProjFFNAUC (%)\\(P_{Norm}(\\%)\\)P(%)
1XXX73.982.880.0
2XX✓74.183.180.8
3✓XX74.583.681.2
4✓✓X74.683.881.4
5✓✓✓74.984.081.7
", + "image_path": "5c968126593e402b4ea874709f10c189e5e994081be5db81ded27a3dec165903.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 250, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 406 + ], + "type": "text", + "content": "Whether to use a shared compress expert. In TMoE, we employ a shared compression expert to reduce the input dimension to " + }, + { + "bbox": [ + 313, + 250, + 555, + 406 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 250, + 555, + 406 + ], + "type": "text", + "content": ", and use multiple routed experts for subsequent processing. As shown in the first and second rows of Table 7, compared to providing each routed expert with a compression expert, sharing the compression expert can achieve better performance while reducing the parameter count. When not sharing the compression expert, we provide t-SNE visualizations of the outputs from compression experts and routed experts in five TMoE modules in Figure 5. The outputs of different compression experts show significant overlaps, while the routed experts are more dispersed, validating the effectiveness of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 405, + 556, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 556, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 556, + 526 + ], + "type": "text", + "content": "TMOE vs. Conventional MoE. We compare TMOE with conventional MoE on the visual tracking task. For conventional MoE, we use LoRA to fine-tune the attention layers and apply MoE in the FFN layers with five experts: one frozen shared expert that copies pretrained FFN parameters, and four trainable routed experts, all routed experts are identical to the original FFN architecture, and a router is also applied. As shown in the first row and third row of Table 7, TMOE demonstrates significant advantages over conventional MoE in both parameter efficiency and performance." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 578, + 558, + 628 + ], + "blocks": [ + { + "bbox": [ + 313, + 536, + 556, + 570 + ], + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 570 + ], + "type": "text", + "content": "Table 7. Ablation studies on whether to share the compression expert and use the conventional MoE to replace TMoE. Results are evaluated on LaSOT test split." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 578, + 558, + 628 + ], + "lines": [ + { + "bbox": [ + 315, + 578, + 558, + 628 + ], + "spans": [ + { + "bbox": [ + 315, + 578, + 558, + 628 + ], + "type": "table", + "html": "
model variants#Params(M)AUC (%)PNorm(%)P(%)
SPMTrack-B115.374.984.081.7
w/o shared compress expert131.374.883.781.5
Conventional MoE316.873.482.679.8
", + "image_path": "9a43817f7916e339c20c9c2d09af726efd92e0addf33a26db0227dcce760838c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "The Number of Routed Experts. In Table 8, we evaluated the impact of different numbers of routed experts in TMoE. Results indicate that increasing the number of experts leads to improved performance, highlighting the potential of TMoE as a method to expand the parameter count and model scale. To balance the number of parameters and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16877" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 149 + ], + "type": "image", + "image_path": "e3f622e8bba3cf50600e59b75077de5b8c4165f72354cf7066a46b048b1c2ac0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 159, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 204 + ], + "type": "text", + "content": "Figure 5. Comparison of t-SNE visualizations. Each column shows outputs from all compression experts (top) and routed experts (bottom) within a TMoE module. Different colors represent distinct experts. Zoom in for better view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 216, + 294, + 411 + ], + "blocks": [ + { + "bbox": [ + 58, + 216, + 294, + 411 + ], + "lines": [ + { + "bbox": [ + 58, + 216, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 294, + 411 + ], + "type": "image", + "image_path": "996fd35ab80d4aa31c79ad29e23b61b958d7985cd7700cabcfac93ffd8e49497.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 417, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 440 + ], + "type": "text", + "content": "Figure 6. Visualization comparison of search region attention maps with and without TMoE. Zoom in for better view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 461, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 461, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 54, + 461, + 295, + 497 + ], + "type": "text", + "content": "performance, we choose to use 4 experts, and whether there are separate optimal configurations for the TMoE in the attention layer and the FFN layer remains to be explored." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "text", + "content": "The Number of Reference Frames. In SPMTrack, we use the template and several tracked frames as reference frames to enhance the spatio-temporal modeling ability. In Table 9, we explore the impact of using different numbers of reference frames during training and inference. During inference, if only two reference frames are used, we simply select the frame at " + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{t}{2} \\right\\rfloor" + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "text", + "content": " as the reference frame. If four reference frames are used, we select the frames at " + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{t}{4} \\right\\rfloor" + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{t}{2} \\right\\rfloor" + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{3 \\times t}{4} \\right\\rfloor" + }, + { + "bbox": [ + 54, + 498, + 295, + 664 + ], + "type": "text", + "content": ". As shown in Table 9, better performance can be achieved when the number of reference frames used during inference remains the same as that during training. Training with more reference frames can also lead to improved performance, illustrating that our method still has substantial potential for performance enhancement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "content": "Comparison of attention maps with and without TMoE. In Figure 6, to exclude the influence of intermediate tracked frames, we use the model in rows 1 and 2 of Table 5 for comparison, where spatio-temporal modeling is removed" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 102, + 551, + 152 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 553, + 93 + ], + "type": "text", + "content": "Table 8. The performance of our method on the test split of GOT-10K when setting different number of routed experts in TMoE." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 102, + 551, + 152 + ], + "lines": [ + { + "bbox": [ + 315, + 102, + 551, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 102, + 551, + 152 + ], + "type": "table", + "html": "
Number2468
AO(%)75.376.576.677.1
SR0.5(%)84.485.986.386.4
SR0.75(%)75.276.377.077.6
", + "image_path": "dd5731af5bf0e702a0a021689bdadc34094fdf97e687c346e4f270ef99407a6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 206, + 547, + 335 + ], + "blocks": [ + { + "bbox": [ + 313, + 164, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 164, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 553, + 198 + ], + "type": "text", + "content": "Table 9. Ablation study on the number of reference frames during training and inference. Results are evaluated based on SPMTrack-B on GOT-10K test split." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 206, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 319, + 206, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 319, + 206, + 547, + 335 + ], + "type": "table", + "html": "
TrainingInferenceAO (%)SR0.5(%)SR0.75(%)
2275.885.175.3
374.784.673.8
469.881.866.2
3273.183.972.7
376.585.976.3
472.883.471.3
4270.683.968.9
374.685.474.5
477.587.377.2
", + "image_path": "a6b78ce3648fada27fe1a8b80694694d7da8fbada265a9ed71dfc30c92aec8de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 356, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 465 + ], + "type": "text", + "content": "and we focus solely on the impact of TMoE. We compare the attention maps of the search region in the last block of Transformer encoder. The first column is the template, the second column displays search region, the third column is the visualization result and pure attention map, as well as the fourth column. Figure 6 shows that when facing challenging scenarios such as occlusions, distractors, and viewpoint variations, TMoE effectively suppresses background regions while enhancing perception of the target boundary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 477, + 388, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 388, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 388, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 665 + ], + "type": "text", + "content": "In this paper, we present TMoE, a mixture of experts module tailored for visual tracking, and propose SPMTrack, a novel tracker enabling spatio-temporal context modeling based on TMoE. TMoE is applied in the linear layers in both self-attention and FFN layers, enhancing the diversity and flexibility of expert combinations to better handle various relation modeling in visual tracking. Additionally, TMoE employs a lightweight and efficient structure and serves as a method of parameter-efficient fine-tuning, which enables us to train SPMTrack of larger scales and enables SPMTrack to achieve state-of-the-art performance with only a small subset of parameters need to be trained. Furthermore, we hope that this work will inspire more applications of mixture of experts in the field of visual tracking." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": "Acknowledgements. This paper was supported by Zhejiang Provincial Natural Science Foundation of China under Grant No.LD24F020016 and National Natural Science Foundation of China under Grant No.62176017." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16878" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompt autoregressive tracker where to look and how to describe. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19048-19057, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 296, + 246 + ], + "type": "text", + "content": "[3] Wenrui Cai, Qingjie Liu, and Yunhong Wang. Hiptrack: Visual tracking with historical prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19258-19267, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 296, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 291 + ], + "type": "text", + "content": "[4] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9589-9600, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 357 + ], + "type": "text", + "content": "[5] Boyu Chen, Peixia Li, Lei Bai, Lei Qiao, Qiuhong Shen, Bo Li, Weihao Gan, Wei Wu, and Wanli Ouyang. Backbone is all your need: a simplified architecture for visual object tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 375-392. Springer, 2022. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "text", + "content": "[6] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14572-14581, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "type": "text", + "content": "[7] Yucheng Chen and Lin Wang. emoe-tracker: Environmental moe-based transformer for robust event-guided object tracking. arXiv preprint arXiv:2406.20024, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 513 + ], + "type": "text", + "content": "[8] Mohammed Nowaz Rabbani Chowdhury, Shuai Zhang, Meng Wang, Sijia Liu, and Pin-Yu Chen. Patch-level routing in mixture-of-experts is provably sample-efficient for convolutional neural networks. In Proceedings of the 40th International Conference on Machine Learning, pages 6074-6114. PMLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 579 + ], + "type": "text", + "content": "[9] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International conference on machine learning, pages 4057-4086. PMLR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "text", + "content": "[10] Yutao Cui, Cheng Jiang, Limin Wang, and Gangshan Wu. Mixformer: End-to-end tracking with iterative mixed attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13608-13618, 2022. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 637, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 295, + 713 + ], + "type": "text", + "content": "[11] Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. DeepSeekMoE: Towards ultimate expert specialization in mixture-of-experts language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguis" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "tics (Volume 1: Long Papers), pages 1280-1297, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 173 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 175, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 293 + ], + "type": "text", + "content": "[13] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yangi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient scaling of language models with mixture-of-experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[14] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, pages 5374-5383, 2019. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 385 + ], + "type": "text", + "content": "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 420 + ], + "type": "text", + "content": "[16] Zhihong Fu, Qingjie Liu, Zehua Fu, and Yunhong Wang. Stmtrack: Template-free visual tracking with space-time memory networks. In CVPR, pages 13774-13783, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 422, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 474 + ], + "type": "text", + "content": "[17] Zhihong Fu, Zehua Fu, Qingjie Liu, Wenrui Cai, and Yunhong Wang. Sparsett: Visual tracking with sparse transformers. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-22, pages 905–912, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "text", + "content": "[18] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 534, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 598 + ], + "type": "text", + "content": "[19] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 146-164. Springer, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 601, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 555, + 645 + ], + "type": "text", + "content": "[20] Shenyuan Gao, Chunluan Zhou, and Jun Zhang. Generalized relation modeling for transformer tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18686-18695, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[21] Mingzhe Guo, Zhipeng Zhang, Heng Fan, and Liping Jing. Divert more attention to vision-language tracking. In Advances in Neural Information Processing Systems, pages 4446-4460. Curran Associates, Inc., 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[22] Kaijie He, Canlong Zhang, Sheng Xie, Zhixin Li, and Zhiwen Wang. Target-aware tracking with long-term context" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16879" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "attention. Proceedings of the AAAI Conference on Artificial Intelligence, 37(1):773-780, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "type": "text", + "content": "[23] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 208 + ], + "type": "text", + "content": "[24] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR, 2022. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 294, + 243 + ], + "type": "text", + "content": "[25] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 2019. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 296, + 300 + ], + "type": "text", + "content": "[26] Yuqing Huang, Xin Li, Zikun Zhou, Yaowei Wang, Zhenyu He, and Ming-Hsuan Yang. Rtracker: Recoverable tracking via pn tree structured memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19038-19047, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 303, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 294, + 335 + ], + "type": "text", + "content": "[27] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "type": "text", + "content": "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 384, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 437 + ], + "type": "text", + "content": "[29] Hamed Kiani Galoogahi, Ashton Fagg, Chen Huang, Deva Ramanan, and Simon Lucey. Need for speed: A benchmark for higher frame rate object tracking. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 441, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 495 + ], + "type": "text", + "content": "[30] Xin Li, Yuqing Huang, Zhenyu He, Yaowei Wang, Huchuan Lu, and Ming-Hsuan Yang. Citetracker: Correlating image and text for visual tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9974-9983, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 296, + 564 + ], + "type": "text", + "content": "[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 609 + ], + "type": "text", + "content": "[32] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. Advances in Neural Information Processing Systems, 35:16743-16754, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 666 + ], + "type": "text", + "content": "[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yaowei Wang, Yong Xu, and Haibin Ling. Tracking meets lora: Faster training, larger model, stronger performance. In Computer Vision - ECCV 2024, pages 300-318, Cham, 2025. Springer Nature Switzerland. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "text", + "content": "[36] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[37] Matthias Mueller, Neil Smith, and Bernard Ghanem. A benchmark and simulator for uav tracking. In ECCV, pages 445-461, 2016. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[38] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, pages 300-317, 2018. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "text", + "content": "[39] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 386 + ], + "type": "text", + "content": "[40] Jonas Pfeiffer, Andreas Rücklé, Clifton Poth, Aishwarya Kamath, Ivan Vulić, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. Adapterhub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[41] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[42] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 542 + ], + "type": "text", + "content": "[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Advances in Neural Information Processing Systems, pages 8583-8595. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "type": "text", + "content": "[44] Zhangyong Tang, Tianyang Xu, Zhenhua Feng, Xuefeng Zhu, He Wang, Pengcheng Shao, Chunyang Cheng, Xiao-Jun Wu, Muhammad Awais, Sara Atito, et al. Revisiting rgbt tracking benchmarks from the perspective of modality validity: A new benchmark, problem, and method. arXiv preprint arXiv:2405.00168, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "text", + "content": "[45] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. Hydralora: An asymmetric lora architecture for efficient fine-tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "text", + "content": "[47] Guangting Wang, Chong Luo, Xiaoyan Sun, Zhiwei Xiong," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16880" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "and Wenjun Zeng. Tracking by instance detection: A meta-learning approach. In CVPR, pages 6288-6297, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[48] Ke Wang, Jiahui Zhu, Minjie Ren, Zeming Liu, Shiwei Li, Zongye Zhang, Chenkai Zhang, Xiaoyu Wu, Qiqi Zhan, Qingjie Liu, et al. A survey on data synthesis and augmentation for large language models. arXiv preprint arXiv:2410.12896, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 216 + ], + "type": "text", + "content": "[49] Xiao Wang, Xiujun Shu, Zhipeng Zhang, Bo Jiang, Yaowei Wang, Yonghong Tian, and Feng Wu. Towards more flexible and accurate object tracking with natural language: Algorithms and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13763-13773, 2021. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "text", + "content": "[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 335 + ], + "type": "text", + "content": "[51] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9697-9706, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 403 + ], + "type": "text", + "content": "[52] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, 2023. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "text", + "content": "[53] Xun Wu, Shaohan Huang, and Furu Wei. Mixture of lora experts. In The Twelfth International Conference on Learning Representations, ICLR, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "text", + "content": "[54] Y. Wu, J. Lim, and M. Yang. Object tracking benchmark. TPAMI, 37(9):1834-1848, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "type": "text", + "content": "[55] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8751-8760, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 580 + ], + "type": "text", + "content": "[56] Jinxia Xie, Bineng Zhong, Zhiyi Mo, Shengping Zhang, Liangtao Shi, Shuxiang Song, and Rongrong Ji. Autoregressive queries for adaptive tracking with spatio-temporal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19300-19309, 2024. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "text", + "content": "[57] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15325-15336, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[58] Dawei Yang, Jianfeng He, Yinchao Ma, Qianjin Yu, and Tianzhu Zhang. Foreground-background distribution modeling transformer for visual object tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10117-10127, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[59] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 272 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "tracking: A one-stream framework. In European Conference on Computer Vision, pages 341-357. Springer, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[60] Yihua Zhang, Ruisi Cai, Tianlong Chen, Guanhua Zhang, Huan Zhang, Pin-Yu Chen, Shiyu Chang, Zhangyang Wang, and Sijia Liu. Robust mixture-of-expert training for convolutional neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 90-101, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "text", + "content": "[61] Yaozong Zheng, Bineng Zhong, Qihua Liang, Zhiyi Mo, Shengping Zhang, and Xianxian Li. Odtrack: Online dense temporal token learning for visual tracking. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7588-7596, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 272 + ], + "type": "text", + "content": "[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16881" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_content_list.json b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..51c1b7c88d998f36975eace0c91b1a2b3dc3432e --- /dev/null +++ b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_content_list.json @@ -0,0 +1,1828 @@ +[ + { + "type": "text", + "text": "SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junchen Yu $^{1,2}$ Si-Yuan Cao $^{1,2,3*}$ Runmin Zhang $^{2}$ Chenghao Zhang $^{2}$ Zhu Yu $^{2}$ Shujie Chen $^{4}$ Bailin Yang $^{4}$ Hui-Liang Shen $^{2}$", + "bbox": [ + 205, + 202, + 789, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Ningbo Innovation Center, Zhejiang University $^{2}$ College of Information Science and Electronic Engineering, Zhejiang University", + "bbox": [ + 104, + 241, + 890, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ NingboTech University $^{4}$ Zhejiang Key Laboratory of Big Data and Future ECommerce Technology, Hangzhou, China", + "bbox": [ + 133, + 257, + 861, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yujunchen,cao_siyuan,runmin_zhang,zch00,yu_zhu}@zju.edu.cn", + "bbox": [ + 236, + 277, + 759, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenshujie,ybl}@zjgsu.edu.cn shenhl@zju.edu.cn", + "bbox": [ + 281, + 295, + 712, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation into two supervised sub-problems, each addressed by its specialized network: a homography estimation network and a modality transfer network. To realize stable training, we introduce an effective split optimization strategy to train each network separately within its respective sub-problem. We also formulate an extra homography feature space supervision to enhance feature consistency, further boosting the estimation accuracy. Moreover, we employ a simple yet effective distillation training technique to reduce model parameters and improve cross-domain generalization ability while maintaining comparable performance. The training stability of SSHNet enables its cooperation with various homography estimation architectures. Experiments reveal that the SSHNet using IHN as homography estimation network, namely SSHNet-IHN, outperforms previous unsupervised approaches by a significant margin. Even compared to supervised approaches MHN and LocalTrans, SSHNet-IHN achieves $47.4\\%$ and $85.8\\%$ mean average corner errors (MACES) reduction on the challenging OPT-SAR dataset. Source code is available at https://github.com/Junchen-Yu/SSHNet.", + "bbox": [ + 88, + 377, + 485, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-modal homography estimation aims to solve the global perspective transformation between two images under different modalities. It is widely used in various computer vision tasks, such as GPS-denied robotic localization [14, 36],", + "bbox": [ + 89, + 813, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7aa7f86c4f28fa35b10864c592c3e254349f52d3061ab072f5ca0a193486ea40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 348, + 588, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57824a5c34de5900a60e6036ee959f4730df1e9bbbf11fd99f54be6d4c6c98cc.jpg", + "image_caption": [ + "Figure 1. Comparison of the solution space between (a) Content consistency indirect supervision and (b) Homography direct supervision. The content consistency indirect supervision may mislead the training due to local minima, while homography direct supervision won't." + ], + "image_footnote": [], + "bbox": [ + 519, + 405, + 588, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9663e2d200187d1eb32f7f797be94e22716530d5f1e616cbe73e4a845ac986b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 345, + 741, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b590b7310f33bc26286c0879cece928aeccb755a68d91a08aef517e26d35c44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 345, + 901, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multi-modal image restoration [10, 27], and multi-spectral image fusion [39, 44]. Recent supervised homography estimation approaches [4, 5, 9, 20, 34, 42] have demonstrated their superiority in handling large deformation and modality gaps. However, since multi-modal images are captured by different sensors, the ground-truth homography deformations are usually unknown in real-world applications.", + "bbox": [ + 511, + 549, + 908, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To avoid using labeled data, unsupervised homography estimation approaches have emerged. Existing approaches [19, 28, 38, 40] typically optimize the similarity between the warped source image and target image through intensity-based losses. Some multi-modal image registration approaches [1, 11] transferred an image into the other modality, using mono-modal similarity metrics on image content to measure the correctness of motion. The previous state-of-the-art (SOTA) approach [41] combined cross-modal intensity-based learning with intra-modal self-supervised learning for homography knowledge injection. Notably, almost all the previous approaches fully or partially rely on the content consistency of images or feature maps to indirectly assess the correctness of homography estimation. As illustrated in Fig. 1a, due to the non-convexity of the image signal, loss value of content consistency indirect supervision has", + "bbox": [ + 511, + 657, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16685", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multiple local minima across its solution space, and hence can not guarantee accurate homography prediction, especially under large modality and deformation gaps. On the contrary, the homography direct supervision in Fig. 1b can avoid such drawback, but it is unavailable in unsupervised cross-modal homography estimation, making the problem more challenging.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To cope with the above problem, we propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation problem into two imperfect sub-problems: Sub-problem I: A supervised mono-modal homography estimation problem, where the modalities of input images are not perfectly consistent. Sub-problem II: A supervised modality transfer problem, where the supervisory image is not perfectly aligned. We introduce a split optimization strategy to separately train a homography estimation network for Sub-problem I and a modality transfer network for Sub-problem II, enabling effective convergence of the framework. Thanks to the proposed problem reformulation and the well-designed optimization strategy, our framework effectively solves the unsupervised task by training the networks under totally direct supervision, showing significantly improved training stability. This enables the successful training of iterative homography estimation architectures like IHN [4] and RHWF [5]. Better still, for the correlation-based homography estimation architectures [4, 5, 41], we propose an extra homography feature space supervision that enhances feature consistency, further improving homography estimation accuracy. Moreover, to address the computational burden of the modality transfer network, we introduce a simple yet effective distillation training technique, which notably reduces model parameters and improves cross-domain generalization ability.", + "bbox": [ + 91, + 199, + 486, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate SSHNet on a variety of datasets including GoogleMap [42], DPDN [30], OPT-SAR [23], and Flash/no-flash [15] cross-modal datasets, together with RGB/NIR [2] cross-spectral datasets. SSHNet outperforms all the previous unsupervised homography estimation approaches with a large gap, and even achieves better performance compared to many supervised ones. The SSHNet using IHN as homography estimation network, namely SSHNet-IHN, owns $47.4\\%$ and $85.8\\%$ lower mean average corner errors (MACEs) than the supervised approaches MHN [20] and LocalTrans [34] on the challenging OPT-SAR dataset. Moreover, the SSHNet using the homography estimation architecture of the previous unsupervised SOTA approach SCPNet [41], namely SSHNet-SCPNet, even outperforms SCPNet itself, revealing the effectiveness of our proposed learning framework. To summarize, our main contributions are as follows:", + "bbox": [ + 89, + 622, + 486, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose SSHNet, a novel unsupervised cross-modal homography estimation learning framework, which refor", + "bbox": [ + 89, + 869, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mulates the unsupervised problem into two sub-problems with direct supervision. To realize stable training, we introduce a split optimization strategy that separately trains the network in each sub-problem.", + "bbox": [ + 526, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For correlation-based homography estimation architectures, we introduce an extra homography feature space supervision to improve feature consistency for homography estimation, which further improves estimation accuracy.", + "- Considering the computational burden introduced by the additional modality transfer network, we devise a simple yet effective distillation training technique to reduce model parameters while improving cross-domain generalization ability with little impact on performance.", + "- The training stability of SSHNet enables its cooperation with various homography estimation architectures, especially iterative ones like IHN and RHWF, resulting in higher performance. SSHNet-IHN outperforms the previous unsupervised approaches by a significant margin." + ], + "bbox": [ + 511, + 152, + 906, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 376, + 663, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Deep Homography Estimation. DeTone et al. [9] introduced DHN, a VGG-style network to predict the homography between the source and target images. Le et al. [20] introduced a multi-scale VGG-style network, named MHN. Chang et al. [7] combined CNN with an untrainable IC-LK iterator. Zhao et al. [42] enhanced the feature similarity to improve the performance of IC-LK iterator. Shao et al. [34] proposed LocalTrans that adopts a multi-scale local transformer to register cross-resolution images. Cao et al. [4] proposed IHN, a trainable iterative framework with substantial accuracy. Cao et al. [5] further introduced RHWF that integrates homography-guided image warping and focus transformer, achieving higher performance. Deng et al. [8] proposed CrossHomo, a multi-level framework for cross-modality and cross-resolution homography estimation. Li et al. [22] introduced DMHomo, a diffusion-based framework to generate high-quality training pairs for supervised homography learning.", + "bbox": [ + 511, + 402, + 908, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Deep Homography Estimation. Nguyen et al. [28] proposed UDHN, an unsupervised homography estimation network trained with photometric loss. Zhang et al. [40] introduced CA-UDHN, projecting images into a consistent feature space for effective estimation. Koguchiuk et al.[19] proposed biHomE, using perceptual loss[18] to improve robustness to intensity and viewpoint variations. Ye et al. [38] proposed BasesHomo, which enforces the image feature to be warp-equivalent with a feature identity loss and decomposes the homography matrix into flow bases. As modality and deformation gaps grow, intensity-based metrics for cross-modal homography matching may fail [41]. To address this, Zhang et al. [41] introduced SCPNet, which combines cross-modal intensity-based learning with intramodal self-supervised learning for homography knowledge", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16686", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5239e411bcd118d92376f41ba7195f75bc7dd1c9732febbbe9968690a12e268b.jpg", + "image_caption": [ + "Figure 2. Reformulating the unsupervised cross-modal homography estimation problem into two sub-problems with direct supervision. (a) Unsupervised cross-modal homography estimation problem without direct supervision. (b) Mono-modal homography estimation problem with imperfect inputs. (c) Modality transfer problem with imperfect supervision." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 477, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "injection. Besides, some unsupervised approaches, such as NeMAR [1], RFNet [37], and UMF-CMGR [11], use modality transfer networks to convert one modality into the other, enabling cross-modal motion estimation. The above unsupervised approaches perform well on images with small deformations, but may produce unsatisfactory results under large deformations and modality gaps.", + "bbox": [ + 89, + 334, + 483, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-to-Image Translation. The image-to-image (I2I) translation problem focuses on converting images from a source domain to a target domain. Isola et al. [17] introduced Pix2pix, which employs conditional GANs and requires paired training data. Zhu et al. [45] later proposed CycleGAN, utilizing cycle-consistent loss to enable unpaired I2I translation. Liu et al.[24] developed UNIT, and Huang et al.[16] introduced MUNIT, both leveraging a shared latent space assumption to improve unpaired translation. More recently, Sasaki et al.[33] proposed UNIT-DDPM, integrating diffusion probabilistic models into the I2I translation framework. Li et al. [21] introduced BBDM, modeling I2I translation as a stochastic Brownian Bridge process.", + "bbox": [ + 89, + 439, + 483, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Reformulation", + "text_level": 1, + "bbox": [ + 89, + 648, + 313, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote $\\mathbf{I}_{\\mathrm{A}}$ and $\\mathbf{I}_{\\mathrm{B}}$ as a pair of images from modality A and B, with an unknown homography $\\mathbf{H}_{\\mathrm{AB,GT}}$ . As illustrated in Fig. 2a, the unsupervised cross-modal homography estimation aims to train a network to estimate the homography $\\hat{\\mathbf{H}}_{\\mathrm{AB}}$ , without direct supervision of $\\mathbf{H}_{\\mathrm{AB,GT}}$ . To solve this problem, in SSHNet, we reformulate it into two sub-problems with direct supervision, including a monomodal homography estimation problem with imperfect inputs (Fig. 2b) and a modality transfer problem with imperfect supervision (Fig. 2c). The two sub-problems are mutually causal, and the imperfect condition of each sub-problem can be gradually addressed through the optimization of another sub-problem. In this way, the unsupervised homography estimation problem can be solved in a totally supervised manner.", + "bbox": [ + 89, + 674, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/84ab33bf4d0409efc768b6f8cc6ba8dbe50624d9e806a250d852944c50fc8c87.jpg", + "image_caption": [ + "Figure 3. Comparison of straight optimization and split optimization with regularization. (a) Cross-modal test MACEs for the network using straight optimization and split optimization with regularization, on OPT-SAR dataset. (b) Schematic diagram of straight optimization and split optimization." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 908, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sub-problem I. For the mono-modal homography estimation, generating image pairs with simulated ground-truth deformations for the direct self-supervised training is a common approach [4, 9, 20, 34]. If we can ideally convert $\\mathbf{I}_{\\mathrm{A}}$ into modality B, producing the transferred image $\\mathbf{I}_{\\mathrm{A,T}}$ , we can treat the unsupervised cross-modal homography estimation problem as the mono-modal one by estimating the homography between $\\mathbf{I}_{\\mathrm{A,T}}$ and $\\mathbf{I}_{\\mathrm{B}}$ , as in Fig. 2b. The homography estimation network can be effectively trained using $\\mathbf{I}_{\\mathrm{B}}$ with the simulated homography deformations, in a self-supervised manner. We refer to the problem of estimating the homography between $\\mathbf{I}_{\\mathrm{B}}$ and the transferred image $\\mathbf{I}_{\\mathrm{A,T}}$ with the self-supervised training using $\\mathbf{I}_{\\mathrm{B}}$ as Sub-problem I.", + "bbox": [ + 511, + 304, + 906, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sub-problem II. As illustrated in Fig. 2c, we also need a modality transfer network to obtain the optimal mapping from $\\mathbf{I}_{\\mathrm{A}}$ to $\\mathbf{I}_{\\mathrm{B}}$ , producing $\\mathbf{I}_{\\mathrm{A,T}}$ . The ideal network training typically requires $\\mathbf{I}_{\\mathrm{B}}$ that is well-aligned with $\\mathbf{I}_{\\mathrm{A}}$ for supervision, which is also unavailable in the unsupervised cross-modal homography estimation. Previous approaches [1, 11] attempt to use GAN-based supervision to achieve modality transfer training using the unaligned image pairs. However, this strategy encounters considerable challenges when dealing with large deformation and modality gaps, as illustrated in [41]. Therefore, high-quality modality transfer requires well-aligned $\\mathbf{I}_{\\mathrm{A}}$ and $\\mathbf{I}_{\\mathrm{B}}$ . Now that we have transferred $\\mathbf{I}_{\\mathrm{A}}$ to $\\mathbf{I}_{\\mathrm{A,T}}$ , the requirement becomes that Sub-problem I being ideally solved. We refer to the modality transfer problem, which uses the homography estimated in Sub-problem I to produce a warped $\\mathbf{I}_{\\mathrm{B}}$ for supervision, as Sub-problem II.", + "bbox": [ + 511, + 503, + 908, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Straight Optimization Fails. However, straight optimization of both sub-problems simultaneously is challenging, as the settlement of either sub-problem depends on the assumption that the other has already been solved. We conduct a pilot experiment where the homography estimation network and the modality transfer network are trained simultaneously to address Sub-problem I and II, but the networks fail to converge, as shown by the blue curve in Fig. 3a.", + "bbox": [ + 511, + 747, + 908, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Split Optimization with Regularization Success. Inspired by decomposition methods in convex optimization,", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16687", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2794cc34fe00a2023aa416de20a220276cbe7f0cb1c67d4242dff0b4832e915b.jpg", + "image_caption": [ + "(a) Optimization of Sub-problem I" + ], + "image_footnote": [], + "bbox": [ + 102, + 109, + 395, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b1d6ee9134bb2bd41fdc3f66dd3c9cc7872edc1bd24a151ae88b60baba42c726.jpg", + "image_caption": [ + "Figure 4. Schematic diagram of Split Supervised Homography estimation Network, namely SSHNet. (a) Detailed optimization process of Sub-problem I. (b) Detailed optimization process of Sub-problem II. (c) Extra supervision on the homography feature space." + ], + "image_footnote": [ + "Only Training Stage Both Training and Inference Stage Simulated Homography Generator" + ], + "bbox": [ + 102, + 212, + 395, + 314 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/796f0533b0826ed2f57b2e9f43b61c0f995e99e4118ecb400ca3049b3eaf2004.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 93, + 900, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3886c86cebc914962011f7750079373d4b27f156e5de442a05daa8dc7c0648c5.jpg", + "image_caption": [ + "Homography Estimation Network" + ], + "image_footnote": [], + "bbox": [ + 406, + 241, + 431, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ca645c770e78fc72840a9eb9060c2595c00d196ba672f7cdccbdef885a2d161.jpg", + "image_caption": [ + "Modality Transfer Network" + ], + "image_footnote": [], + "bbox": [ + 604, + 241, + 627, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/67123521fcf15b7cc9cee4a67fab7ea846d9d955500147f66be1a2ee7198d0fd.jpg", + "image_caption": [ + "Trainable", + "Weight Sharing" + ], + "image_footnote": [], + "bbox": [ + 764, + 243, + 779, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66e5b3720dabd991d6c874fe8c66f6e1255c6d06905008fa9f501cf71ef77600.jpg", + "image_caption": [ + "rozen" + ], + "image_footnote": [], + "bbox": [ + 838, + 243, + 862, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7874e89d24a13231f03e76fd20cf551003e561d5f4d6192a8f8161abfe98ab31.jpg", + "image_caption": [ + "Feature Extractor in Homography Estimation Network" + ], + "image_footnote": [], + "bbox": [ + 406, + 266, + 429, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7608eb7f6b724e1471d657ed09d0507b514ab4883f88c22c2ba73cffff298a26.jpg", + "image_caption": [ + "mage Warping" + ], + "image_footnote": [], + "bbox": [ + 792, + 267, + 823, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we then consider using a split optimization strategy to address the two sub-problems separately. During optimization, each network updates its weights with the other one's frozen. As in Fig. 2b and Fig. 2c, in each iteration, the homography estimation network in Sub-problem I is updated, producing a better-aligned $\\mathbf{I}_{\\mathrm{B}}$ to guide the modality transfer network in Sub-problem II in the next iteration. Similarly, the modality transfer network in Sub-problem II is updated, generating a better transferred image $\\mathbf{I}_{\\mathrm{A,T}}$ to serve as input for the homography estimation network in Sub-problem I. This process gradually improves the outputs of each network during training, facilitating the optimization of the other network, and finally makes both sub-problems nearly perfect supervised ones. However, at the beginning of the training, the modality gap between $\\mathbf{I}_{\\mathrm{A,T}}$ and $\\mathbf{I}_{\\mathrm{B}}$ in Sub-problem I is extremely large, so is the misalignment of $\\mathbf{I}_{\\mathrm{A}}$ and $\\mathbf{I}_{\\mathrm{B}}$ in Sub-problem II, which hinders the mutual promotion between two subproblems. A regularization term that guides the training into our desired direction is needed. Inspired by [41] $^{1}$ , in Subproblem I, we replace the self-supervised training using only $\\mathbf{I}_{\\mathrm{B}}$ by the two-branch self-supervised training using both $\\mathbf{I}_{\\mathrm{A,T}}$ and $\\mathbf{I}_{\\mathrm{B}}$ (see Fig. 2b). This enables the homography estimation network to implicitly learn cross-modal homography knowledge at the beginning of the training, and triggers the two split optimized sub-problems to converge into our desired direction, as shown by the orange curve in Fig. 3a.", + "bbox": [ + 91, + 368, + 483, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 772, + 181, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the problem reformulation in Section 3, we further investigate its potential and propose an unsupervised cross-modal homography estimation learning framework, named", + "bbox": [ + 89, + 797, + 483, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Split Supervised Homography estimation Network (SSHNet). Section 4.1 covers the split optimization of SSHNet (Fig. 4a and Fig. 4b), Section 4.2 discusses the cooperation of SSHNet with various homography estimation architectures, Section 4.3 outlines the design of the modality transfer network, Section 4.4 presents the extra homography feature space supervision (Fig. 4c), and Section 4.5 illustrates the distillation training technique (Fig. 5).", + "bbox": [ + 511, + 367, + 906, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Split Optimization", + "text_level": 1, + "bbox": [ + 511, + 502, + 692, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4a and Fig. 4b, the homography estimation network and the modality transfer network are optimized separately in each iteration, forming the split optimization framework of our SSHNet. In the following, we will elaborate the optimization procedure for each network.", + "bbox": [ + 511, + 525, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization of Sub-problem I. As mentioned in Section 3, we adopt a two-branch self-supervised learning strategy to optimize the homography estimation network in Subproblem I. We apply the separate simulated homography deformation on $\\mathbf{I}_{\\mathrm{A}}$ and $\\mathbf{I}_{\\mathrm{B}}$ , generating the intra-modal image pairs $(\\mathbf{I}_{\\mathrm{A}},\\mathbf{I}_{\\mathrm{A}}^{\\prime})$ and $(\\mathbf{I}_{\\mathrm{B}},\\mathbf{I}_{\\mathrm{B}}^{\\prime})$ with corresponding ground-truth homography $\\mathbf{H}_{\\mathrm{A,GT}}$ and $\\mathbf{H}_{\\mathrm{B,GT}}$ . The modality transfer network transfers $(\\mathbf{I}_{\\mathrm{A}},\\mathbf{I}_{\\mathrm{A}}^{\\prime})$ into $(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime})$ . We simultaneously use $(\\mathbf{I}_{\\mathrm{B}},\\mathbf{I}_{\\mathrm{B}}^{\\prime})$ and $(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime})$ to train the weight-shared homography estimation network in a self-supervised manner. The optimization of Sub-problem I can be formulated as:", + "bbox": [ + 511, + 602, + 908, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\zeta} {\\arg \\min } L _ {\\mathrm {H}} \\left(\\mathcal {H} _ {\\zeta} \\left(\\mathbf {I} _ {\\mathrm {B}}, \\mathbf {I} _ {\\mathrm {B}} ^ {\\prime}\\right), \\mathbf {H} _ {\\mathrm {B}, \\mathrm {G T}}\\right) + \\mathcal {R} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}} ^ {\\prime}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 797, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_{\\zeta}$ denotes the homography estimation network with parameters $\\zeta$ to be optimized, and $L_{\\mathrm{H}}$ denotes the homography estimation loss. $\\mathcal{R}(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime}) = L_{\\mathrm{H}}\\bigl (\\mathcal{H}_{\\zeta}(\\mathcal{T}_{\\theta^{*}}(\\mathbf{I}_{\\mathrm{A}}),\\mathcal{T}_{\\theta^{*}}(\\mathbf{I}_{\\mathrm{A}}^{\\prime})),\\mathbf{H}_{\\mathrm{A,GT}}\\bigr)$ , which is the regular", + "bbox": [ + 511, + 838, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "[41] finds that if training the weight-shared homography estimation network using the simulated deformation within each two modalities, the cross-modal homography knowledge can be implicitly learned through the multitask learning mechanism [6, 12].", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16688", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ization term, where $\\mathcal{T}_{\\theta^*}$ denotes the modality transfer network with parameters $\\theta^{*}$ frozen.", + "bbox": [ + 89, + 90, + 485, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization of Sub-problem II. We train a modality transfer network to convert $\\mathbf{I}_{\\mathrm{A}}$ into an image of pseudo modality B, namely $\\mathbf{I}_{\\mathrm{A,T}}$ . Following the reformulation in Section 3, we apply the homography matrix $\\hat{\\mathbf{H}}_{\\mathrm{AB}}$ predicted by $\\mathcal{H}_{\\zeta}$ to warp $\\mathbf{I}_{\\mathrm{B}}$ , providing direct supervision for the modality transfer network. The optimization of Sub-problem II can be formulated as", + "bbox": [ + 89, + 121, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } L _ {\\mathrm {T}} \\left(\\mathcal {T} _ {\\theta} (\\mathbf {I} _ {\\mathrm {A}}), \\mathcal {W} (\\mathbf {I} _ {\\mathrm {B}}, \\mathcal {H} _ {\\zeta^ {*}} (\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {B}}))\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 237, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}_{\\theta}$ denotes the modality transfer network with parameters $\\theta$ to be optimized, $\\mathcal{H}_{\\zeta^{*}}$ denotes the homography estimation network with parameters $\\zeta^{*}$ frozen, $\\mathcal{W}$ denotes the image warping with the estimated homography, and $L_{\\mathrm{T}}$ denotes the modality transfer loss.", + "bbox": [ + 89, + 268, + 483, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Homography Estimation Network", + "text_level": 1, + "bbox": [ + 89, + 352, + 387, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Present homography estimation approaches exhibit a variety of architectural designs. Some approaches [9, 20, 28, 38, 40] concatenate input images in the channel dimension, directly estimating the homography deformation. Others [4, 5, 34, 41] use backbone networks to extract features and employ the feature correlation to perform homography decoding. Additionally, recent approaches [4, 5] further introduce iterative architectures to refine the estimation results, achieving substantial improvement in accuracy.", + "bbox": [ + 89, + 375, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in the experimental results of Section 5.3, our proposed framework is independent of the specific homography estimation architecture, enabling integration with various architectures. The pure supervised training strategy significantly boosts the stability of the optimization process. As a result, iterative homography estimation architectures like IHN [4] and RHWF [5] can be effectively employed, leading to a higher performance upper bound. In contrast, due to the lack of an explicit solution for consistent features between cross-modal images, the previous SOTA approach, SCPNet [41], struggles to converge when using these iterative architectures. For the self-supervised training, we adopt $L_{1}$ homography estimation loss following previous works [4, 5, 9, 20, 41], which is formulated as", + "bbox": [ + 89, + 510, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {H}} (\\hat {\\mathbf {H}}, \\mathbf {H} _ {\\mathrm {G T}}) = \\left\\| \\hat {\\mathbf {H}} - \\mathbf {H} _ {\\mathrm {G T}} \\right\\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 728, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{H}}$ denotes the estimated homography, $\\mathbf{H}_{\\mathrm{GT}}$ denotes the ground-truth homography. For iterative homography architectures, we use a weighted sum of the $L_{1}$ distance following [4, 5].", + "bbox": [ + 89, + 763, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Modality Transfer Network", + "text_level": 1, + "bbox": [ + 89, + 833, + 339, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design the modality transfer network in our SSHNet as a U-shaped architecture to enhance the perception of multiscale information. Inspired by Swin-Unet [3], the Swin", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer block [25] serves as the basic unit of the modality transfer network. The attention mechanism enhances the network's ability to capture hierarchical structures and context information. This facilitates more effective learning of complex mappings between modalities.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Section 4.1, we warp image $\\mathbf{I}_{\\mathrm{B}}$ with the $\\hat{\\mathbf{H}}_{\\mathrm{AB}}$ predicted by the homography estimation network for the modality transfer supervision. During optimization, $\\hat{\\mathbf{H}}_{\\mathrm{AB}}$ cannot ideally align the input image $\\mathbf{I}_{\\mathrm{A}}$ with the supervisory image $\\mathbf{I}_{\\mathrm{B}}$ . Therefore, instead of using pixel-wise $L_{1}$ or $L_{2}$ loss, we apply perceptual loss $L_{\\mathrm{pcp}}$ [18] to capture high-level representations, ensuring better tolerance to misalignments, which is formulated as", + "bbox": [ + 511, + 167, + 906, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {p c p}} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {B}, \\mathrm {W}}\\right) = \\sum_ {j} \\frac {1}{C _ {j} H _ {j} W _ {j}} \\left\\| \\psi_ {j} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}\\right) - \\psi_ {j} \\left(\\mathbf {I} _ {\\mathrm {B}, \\mathrm {W}}\\right) \\right\\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 299, + 926, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi_{j}$ is the $j$ -th layer of the pre-trained VGG with a feature map of shape $C_j\\times H_j\\times W_j$ , $\\mathbf{I}_{\\mathrm{B,W}}$ denotes the image $\\mathbf{I}_{\\mathrm{B}}$ warped by the estimated $\\hat{\\mathbf{H}}_{\\mathrm{AB}}$ .", + "bbox": [ + 511, + 349, + 905, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Extra Homography Feature Space Supervision", + "text_level": 1, + "bbox": [ + 511, + 407, + 905, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The current modality transfer network in Sub-problem II is supervised by the perceptual loss [18], which computes feature similarity using a pre-trained VGG network. This may yield suboptimal results for homography estimation in Sub-problem I when trained under the present totally split optimization framework.", + "bbox": [ + 511, + 430, + 905, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, as illustrated in Fig. 4c, we introduce an extra supervision in the homography feature space when the architecture of homography estimation network has explicit feature extractors, such as IHN, RHWF, and SCPNet. A homography feature (HF) loss that directly enhances the feature consistency for homography estimation between $\\mathbf{I}_{\\mathrm{A,T}}$ and $\\mathbf{I}_{\\mathrm{B}}$ is computed. The modality transfer network and the feature extractor of the homography estimation network are optimized together under the supervision of HF loss, which can be expressed as", + "bbox": [ + 511, + 521, + 906, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta , \\xi} {\\arg \\min } L _ {\\mathrm {H F}} \\left(\\mathcal {F} _ {\\xi} \\left(\\mathcal {T} _ {\\theta} (\\mathbf {I} _ {\\mathrm {A}})\\right), \\mathcal {F} _ {\\xi} \\left(\\mathcal {W} \\left(\\mathbf {I} _ {\\mathrm {B}}, \\hat {\\mathbf {H}} _ {\\mathrm {A B}}\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 684, + 906, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\xi}$ denotes the feature extractor with parameters $\\xi$ to be optimized. We note that IHN, RHWF, and SCPNet all depend on feature correlation to perform homography decoding. Considering this, We formulate the HF loss using a correlation-based loss function $L_{\\mathrm{corr}}$ to enhance feature consistency, which is formulated as", + "bbox": [ + 511, + 722, + 905, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\operatorname {c o r r}} \\left(\\mathbf {F} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {F} _ {\\mathrm {B}, \\mathrm {W}}\\right) = - \\sum_ {i, j} \\mathbf {F} _ {\\mathrm {A}, \\mathrm {T}} (i, j) ^ {\\top} \\mathbf {F} _ {\\mathrm {B}, \\mathrm {W}} (i, j), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 825, + 906, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{\\mathrm{A,T}}$ and $\\mathbf{F}_{\\mathrm{B,W}}$ denote the feature maps of $\\mathbf{I}_{\\mathrm{A,T}}$ and $\\mathbf{I}_{\\mathrm{B,W}}$ in the homography estimation network.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16689", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9093a72265be09818f5a97f6e14019bb401ff28c52d57dc9671abfa6862b6b1.jpg", + "image_caption": [ + "Figure 5. Schematic diagram of the distillation training." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 467, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Distillation Training", + "text_level": 1, + "bbox": [ + 89, + 170, + 284, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can see that the present framework has extra network parameters due to the utilization of the modality transfer network in Sub-problem II. Moreover, as the modality transfer network is trained on multi-modal image pairs from a fixed domain, its generalization ability to images from other domains may be limited.", + "bbox": [ + 89, + 191, + 483, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To cope with the above problem, we introduce a distillation training technique under the inspiration of the unsupervised knowledge distillation frameworks in domain adaptive object detection [32, 43]. We use SSHNet as a teacher network to guide a student network that only retains the homography estimation network, named Distilled SSHNet (SSHNet-D). During the distillation training, the homography $\\hat{\\mathbf{H}}_{\\mathrm{teacher}}$ predicted by SSHNet serves as a pseudo ground-truth to supervise SSHNet-D, as illustrated in Fig. 5. The distillation training can be formulated as", + "bbox": [ + 89, + 282, + 483, + 434 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\omega} {\\arg \\min } L _ {\\mathrm {H}} \\left(\\phi_ {\\omega} \\left(\\mathbf {I} _ {\\mathrm {A}}, \\mathbf {I} _ {\\mathrm {B}}\\right), \\hat {\\mathbf {H}} _ {\\text {t e a c h e r}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 444, + 483, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi_{\\omega}$ denotes the SSHNet-D with the parameters $\\omega$ to be optimized.", + "bbox": [ + 89, + 478, + 483, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 521, + 223, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 546, + 297, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate SSHNet on various datasets, including GoogleMap [42], DPDN [30], OPT-SAR [23], Flash/no-flash [15] cross-modal datasets, and RGB/NIR [2] cross-spectral dataset. Examples for each dataset are shown in Fig. 6. To ensure a fair comparison, all evaluated methods are trained and tested using the same training and test splits within each dataset.", + "bbox": [ + 89, + 568, + 483, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We evaluate SSHNet with traditional feature-based approaches including SIFT [26], ORB [31], unsupervised approaches including UDHN [28], biHomE [19], CA-UDHN [40], BasesHomo [38], UMF-CMGR [11], SCP-Net [41], and supervised approaches including DHN [9], MHN [20], LocalTrans [34], IHN [4], RHWF [5]. For feature-based approaches, we choose RANSAC [13] as the outlier rejection algorithms. UMF-CMGR is an image fusion approach based on registration, and we use the registration part for comparison.", + "bbox": [ + 89, + 674, + 483, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Homography Estimation Architectures. To investigate the potential of our proposed SSHNet, we adopt various homography estimation architectures including DHN [9] and MHN [20], together with iterative architectures such as IHN [4] and RHWF [5]. Moreover, we also adopt the", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c8e0856bef4e0c138f65b44a0e6f9da9908f4545cc7caa56fc4970688eedfed3.jpg", + "table_caption": [ + "Table 1. Ablation study on the training strategy of SSHNet." + ], + "table_footnote": [], + "table_body": "
ExperimentMACE ↓
w/o problem reformulationNC
w/o split optimizationNC
pure self-supervised learning24.60
SSHNet2.94
", + "bbox": [ + 517, + 104, + 901, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9c42742ad791b07d2dcc7adb6e5013fd69a4b2b3dda251f15b184bf073a81cd.jpg", + "table_caption": [ + "Table 2. Ablation study on the architectures of the modality transfer network." + ], + "table_footnote": [], + "table_body": "
ExperimentParameters (M)MACE ↓
CNN-based7.488.72
Transformer-based7.542.94
", + "bbox": [ + 517, + 223, + 901, + 273 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a4d2b929a447173699163574fb50e6d526827c4fcb261f5db3f3aab44024c24.jpg", + "table_caption": [ + "Table 3. Ablation study on the loss functions. $L_{\\mathrm{T}}$ denotes the modality transfer loss, $L_{\\mathrm{HF}}$ denotes the homography feature loss." + ], + "table_footnote": [], + "table_body": "
ExperimentLoss TypeMACE ↓
LTL15.88
L26.23
Lpcp4.52
LT + LHFLpcp + L14.09
Lpcp + L23.27
Lpcp + Lcorr2.94
", + "bbox": [ + 517, + 311, + 901, + 416 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d2ab19dfc218e01baa5efdcedbc786f601f3a41677594264be55687d55c08e3a.jpg", + "table_caption": [ + "Table 4. Ablation study of the modality transfer direction." + ], + "table_footnote": [], + "table_body": "
Method\\DirectionRGB→NIRNIR→RGBOPT→SARSAR→OPT
SSHNet1.661.752.944.28
", + "bbox": [ + 517, + 440, + 901, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "architecture from the previous SOTA unsupervised approach SCPNet [41], to specifically evaluate the effectiveness of our proposed unsupervised learning framework.", + "bbox": [ + 511, + 506, + 905, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Similar to previous works [4, 5, 9, 34, 42], we randomly perturb the corner points of $128 \\times 128$ images to make the deformed image, with a perturbation range set to [-32, 32]. We evaluate homography estimation accuracy using the mean average corner error (MACE) following [4, 5, 9, 34, 42]. A lower MACE indicates a higher accuracy.", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation", + "text_level": 1, + "bbox": [ + 511, + 652, + 617, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ablation studies are conducted on the OPT-SAR dataset. We adopt IHN as the homography estimation network in our SSHNet for ablation studies.", + "bbox": [ + 511, + 674, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Strategy. We conduct an ablation study on the training strategy of our SSHNet. The experimental results in Table 1 demonstrate that both problem reformulation and the split optimization strategy are crucial for SSHNet. Without problem reformulation, unsupervised training with intensity-based supervision fails to converge. Even with problem reformulation, directly optimizing both networks simultaneously does not yield reliable homography estimation. SSHNet achieves effective convergence only when problem reformulation is combined with the split optimization strategy, where each network is trained separately. We also compare our framework with method trained through", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16690", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d3bcc836cb9d77f179be2784cda938a81f454df88aa0a4f00f0c84d4f4bc1d4.jpg", + "image_caption": [ + "(a) GoogleMap" + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 173, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad883459225c75b0a40e9de50d67063a10a43a88424cc96af2e78ad321243a60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 89, + 248, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95f4f8c36a1e5c14a5b293a43c33a37d8318c8d313f3a7e52f6abb0593e5de83.jpg", + "image_caption": [ + "(b) DPDN" + ], + "image_footnote": [], + "bbox": [ + 261, + 90, + 334, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/467ea23b4e37cc7c7271bba3d061132216adddc75a548bfac04157f0429e207c.jpg", + "image_caption": [ + "Figure 6. Examples for each dataset." + ], + "image_footnote": [], + "bbox": [ + 334, + 90, + 408, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6808aeee93d068056181739cb07aab218f753be6f5a9001e581712f5f2a7367.jpg", + "image_caption": [ + "(c) OPT-SAR" + ], + "image_footnote": [], + "bbox": [ + 423, + 90, + 496, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9915e447e38c24b8aa4d4d849ed6b021f8d96fe98c70016290e4f830c8f60bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 570, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f24c14a78ab48ae9c4c78db7ecef954b184eb574b475bd146c064a0ef6b96c44.jpg", + "image_caption": [ + "(d) Flash/no-flash" + ], + "image_footnote": [], + "bbox": [ + 583, + 90, + 656, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3374e92360f8b22776246c78e9d0b6664d62235f9578e2b76b8b08f9a3351a3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 90, + 732, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e678698ac822c50d4530bf8c899f621e5cde91d7647b0945253d656989df89b2.jpg", + "image_caption": [ + "(e) RGB/NIR" + ], + "image_footnote": [], + "bbox": [ + 745, + 90, + 816, + 146 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a21519ed0aa91bc732f2e0323b2efb0e8aa9003765687d2b4ec487d12fe87fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 90, + 893, + 146 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43da390553b587872aac5082b8ab376616c7e5c46d30fdd07ad56efb0323a032.jpg", + "table_caption": [ + "Table 5. Quantitative comparison of our SSHNet and other approaches. NC denotes the training is not converged. The best results of unsupervised methods are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
Dataset ApproachGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
TraditionalSIFT [26]24.5325.1424.8218.2921.33
ORB [31]24.5224.8624.9020.2222.67
SupervisedDHN [9]5.204.928.276.4211.88
MHN [20]2.692.955.595.245.52
LocalTrans [34]3.221.7920.681.521.72
IHN [4]0.921.171.670.801.63
RHWF [5]0.711.061.480.651.07
UnsupervisedUDHN [28]22.84NCNC21.2023.43
CA-UDHN [40]24.6124.9924.7621.3224.12
biHomE [19]NCNCNC11.8623.77
BasesHomo [38]24.4927.3326.7225.1224.41
UMF-CMGR [11]24.6025.4624.7023.4922.38
SCPNet [41]4.35NCNC2.674.78
SSHNet-DHN (Ours)9.289.7117.6310.5112.13
SSHNet-MHN (Ours)2.903.046.775.626.93
SSHNet-SCPNet (Ours)3.894.8112.942.653.86
SSHNet-RHWF (Ours)1.291.263.080.951.52
SSHNet-IHN (Ours)1.231.122.941.081.66
SSHNet-IHN-D (Ours)1.261.243.311.161.72
", + "bbox": [ + 94, + 214, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pure self-supervised learning. Results in Table 1 show that self-supervised learning alone does not yield satisfactory homography estimation results.", + "bbox": [ + 89, + 505, + 482, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architectures of the Modality Transfer Network. As illustrated in Table 2. Our Transformer-based architecture outperforms the CNN-based architecture by a large margin, with similar parameter counts.", + "bbox": [ + 89, + 556, + 483, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss Functions. To further investigate the effectiveness of the modality transfer and homography feature loss functions, we first train the network without HF loss to evaluate the performance of different types of modality transfer loss. The results in the upper part of Table 3 indicate that the perceptual loss $L_{\\mathrm{pcp}}$ outperforms $L_{1}$ and $L_{2}$ . Then we adopt $L_{\\mathrm{pcp}}$ as the modality transfer loss and add different types of HF loss to analyze their impact. We report the results in the lower part of Table 3. Though all different types of HF loss further boost the network's performance, the correlation-based loss $L_{\\mathrm{corr}}$ achieves superior performance compared to $L_{1}$ and $L_{2}$ .", + "bbox": [ + 89, + 623, + 483, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modality transfer direction. Table 4 presents the results of different transfer directions on the RGB/NIR and OPT-SAR datasets. Since some modalities carry more information, transferring from richer to less informative modalities facilitates the modality transfer network's learning process, thereby promoting the framework's convergence.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparisons with Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 505, + 828, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the homography estimation results of our proposed SSHNet and baselines and list the experimental results in Table 5. It is observed that the prior traditional and unsupervised approaches can hardly handle the homography estimation under large modality and deformation gaps. In contrast, our method achieves much more satisfactory results with various homography estimation architectures. When adopting the iterative architectures IHN and RHWF, our method outperforms all the compared traditional and unsupervised approaches by a significant margin. Moreover, even though trained without direct cross-modal supervision, SSHNet-IHN and SSHNet-RHWF outperforms supervised approaches such as DHN, MHN and LocalTrans. Specifically, SSHNet-IHN owns $54.3\\%$ and $47.4\\%$ lower MACEs than MHN, and $61.8\\%$ and $85.8\\%$ lower MACEs than LocalTrans on the challenging GoogleMap and OPT-SAR datasets, respectively. Notably, SSHNet-SCPNet, which uses the homography estimation architecture from the previous unsupervised SOTA approach SCPNet, outperforms SCPNet itself on all datasets, further proving the superiority of our proposed training framework.", + "bbox": [ + 511, + 532, + 906, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct the distillation training on SSHNet with different homography estimation architectures. We present the distillation training results of SSHNet-IHN in Table 5,", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16691", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0d2946f81085d3048c98c621ac4606ecfddf6c59d0e98b5f5956a3ce97d76737.jpg", + "table_caption": [ + "Table 6. Cross-dataset evaluation results. Red indicates the improved performance of SSHNet-IHN-D compared to SSHNet-IHN, and blue indicates the declined one." + ], + "table_footnote": [], + "table_body": "
Test TrainGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
GoogleMap/↑1.8%↑1.0%↑80.4%↑63.0%
DPDN↑66.5%/↑0.4%↑92.1%↑75.8%
OPT-SAR↑48.4%↑1.3%/↑77.5%↑66.2%
Flash/no-flash↓5.2%↓0.2%↑0.1%/↓4.6%
RGB/NIR↓0.5%↑1.2%↑3.0%↑2.1%/
", + "bbox": [ + 94, + 133, + 480, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "namely SSHNet-IHN-D. We observe that the performance of SSHNet-IHN-D is comparable to that of SSHNet-IHN. For more detailed qualitative visualization and experimental results, please refer to supplementary material.", + "bbox": [ + 89, + 246, + 483, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Cross-Dataset Generalization", + "text_level": 1, + "bbox": [ + 89, + 316, + 352, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further conduct a cross-dataset evaluation of SSHNet-IHN-D and SSHNet-IHN on the aforementioned datasets, to further reveal the effectiveness of our proposed distillation training. As illustrated in Fig. 6, multi-modal image pairs from the GoogleMap, DPDN, and OPT-SAR datasets exhibit substantial modality gaps, and we call these three datasets challenging datasets, for a better illustration. In contrast, the Flash/no-flash and RGB/NIR datasets show smaller modality gaps, so we call them easy datasets. For each dataset, we compute the percentage of homography predictions that have average corner errors (ACE) lower than 5 pixels on the test set, which reveals the ratio of reliable estimations. We directly list the differences of the percentage of SSHNet-IHN-D and SSHNet-IHN in Table 6.", + "bbox": [ + 88, + 338, + 485, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SSHNet-IHN uses a modality transfer network to transfer images from one modality to the other on the training dataset, making it highly prone to overfit the training data. In contrast, SSHNet-IHN-D excludes the modality transfer network and directly extracts consistent features from cross-modal images, and hence is forced to learn the features that is robust across different domains. As illustrated in Table 6, when trained on challenging datasets, SSHNet-IHN-D exhibits significant superior generalization ability. Specifically, SSHNet-IHN-D trained on DPDN yields $66.5\\%$ , $92.1\\%$ , and $75.8\\%$ improvement on GoogleMap, Flash/no-flash and RGB/NIR datasets compared to SSHNet. We note that the generalization ability across the challenging datasets remains limited due to the substantial domain gaps between the datasets. As for training on easy datasets, the generalization performance improvement is plain. This is because the training data is too easy to provide enough knowledge for obtaining the features robust across different domains.", + "bbox": [ + 89, + 550, + 485, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Real World Scenarios", + "text_level": 1, + "bbox": [ + 89, + 832, + 294, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate our method on real-world scenarios [35] that do not satisfy perfect homography alignment, including RGB/NIR and Flash/no-flash image pairs. As illustrated", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/325347c9fa2695dc520ab0fd4fd97ccc0d3b628fd61bed4f182edcdd2522d0b1.jpg", + "image_caption": [ + "(a) Cross modal", + "Figure 7. Visualization of results on real-world deformations." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 696, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/110122b662fed84b63dd790f30ef911e0b1ecbebf8bd616c957a6dfe2091461f.jpg", + "image_caption": [ + "(b) UDIS-TIR" + ], + "image_footnote": [], + "bbox": [ + 705, + 88, + 905, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0da805ba45253c14fba25a11743982726418fab9c5e647cffa39631b14034a11.jpg", + "table_caption": [ + "Table 7. Experimental results on the UDIS-TIR dataset. $I_{3 \\times 3}$ denotes the identity matrix." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑
I3×312.920.306
SCPNetNCNC
SSHNet-IHN20.050.728
", + "bbox": [ + 517, + 273, + 901, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 7a, the model trained with simulated homographies achieves effective performance. We also use a pretrained RGB-to-TIR translation network to convert RGB-RGB image pairs from the UDIS [29] dataset into RGB-TIR image pairs, creating the UDIS-TIR dataset with parallax image pairs. Following UDIS, we evaluate SSHNet-IHN on the UDIS-TIR dataset and compute the PSNR and SSIM on RGB images before modality transfer. We present the quantitative results in Table 7 and show more visualization results in Fig. 7b.", + "bbox": [ + 511, + 345, + 906, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 510, + 638, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SSHNet, a novel unsupervised cross-modal homography estimation framework. SSHNet reformulates the problem into two supervised sub-problems, addressed by a homography estimation network and a modality transfer network. These networks are trained separately using a split optimization strategy, ensuring stable training. We also introduce an extra homography feature space supervision to further improve feature consistency during the homography estimation and boost the performance of our framework. Moreover, we employ a distillation training technique to reduce model parameters and improve cross-domain generalization ability. Experimental results show that our SSHNet achieves SOTA performance on multiple datasets among unsupervised approaches, and outperforms many supervised approaches.", + "bbox": [ + 511, + 536, + 908, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the Zhejiang Provincial Natural Science Foundation of China under grant LD24F020003, in part by the National Natural Science Foundation of China under grant 62301484, in part by the Ningbo Natural Science Foundation of China under grant 2024J454, and in part by the National Key Research and Development Program of China under grant 2023YFB3209800. We also thank the generous help from Jianxin Hu, Zhejiang University.", + "bbox": [ + 511, + 761, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16692", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13410–13419, 2020. 1, 3", + "[2] Matthew Brown and Sabine Süssstrunk. Multispectral SIFT for scene category recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 177-184. IEEE, 2011. 2, 6", + "[3] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 205-218. Springer, 2022. 5", + "[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1879-1888, 2022. 1, 2, 3, 5, 6, 7", + "[5] Si-Yuan Cao, Runmin Zhang, Lun Luo, Beinan Yu, Zehua Sheng, Junwei Li, and Hui-Liang Shen. Recurrent homography estimation using homography-guided image warping and focus transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9833-9842, 2023. 1, 2, 5, 6, 7", + "[6] Rich Caruana. Multitask learning. Machine Learning, 28: 41-75, 1997. 4", + "[7] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. CLKN: Cascaded Lucas-Kanade networks for image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2213-2221, 2017. 2", + "[8] Xin Deng, Enpeng Liu, Chao Gao, Shengxi Li, Shuhang Gu, and Mai Xu. CrossHomo: Cross-modality and cross-resolution homography estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7", + "[10] Fayaz Ali Dharejo, Muhammad Zawish, Farah Deeba, Yuanchun Zhou, Kapal Dev, Sunder Ali Khowaja, and Nawab Muhammad Faseeh Qureshi. Multimodal-boost: Multimodal medical image super-resolution using multi-attention network with wavelet transform. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2022. 1", + "[11] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, 2022. 1, 3, 6, 7", + "[12] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2051-2060, 2017. 4", + "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to" + ], + "bbox": [ + 93, + 114, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 6", + "[14] Hunter Goforth and Simon Lucey. GPS-denied UAV localization using pre-existing satellite imagery. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 2974-2980. IEEE, 2019. 1", + "[15] Shengfeng He and Rynson WH Lau. Saliency detection with flash and no-flash image pairs. In Proceedings of the European Conference on Computer Vision, pages 110-124. Springer, 2014. 2, 6", + "[16] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision, pages 172-189, 2018. 3", + "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1125-1134, 2017. 3", + "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision, pages 694–711. Springer, 2016. 2, 5", + "[19] Daniel Koguciuk, Elahe Arani, and Bahram Zonooz. Perceptual loss for robust unsupervised homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4274-4283, 2021. 1, 2, 6, 7", + "[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020. 1, 2, 3, 5, 6, 7", + "[21] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. BBDM: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 3", + "[22] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. DMHomo: Learning homography with diffusion models. ACM Transactions on Graphics, 43 (3):1-16, 2024. 2", + "[23] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. MCANet: A joint semantic segmentation framework of optical and SAR images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106:102638, 2022. 2, 6", + "[24] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in Neural Information Processing Systems, 30, 2017. 3", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5", + "[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 6, 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Iman Marivani, Evaggelia Tsiligianni, Bruno Cornelis, and Nikos Deligiannis. Designing CNNs for multimodal image restoration and fusion via unfolding the method of multipliers. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5830-5845, 2022. 1", + "[28] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 1, 2, 5, 6, 7", + "[29] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Unsupervised deep image stitching: Reconstructing stitched features to images. IEEE Transactions on Image Processing, 30:6184-6197, 2021. 8", + "[30] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In *British Machine Vision Conference*. The British Machine Vision Association, 2016. 2, 6", + "[31] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 6, 7", + "[32] Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain adaptation. In Proceedings of the International Conference on Machine Learning, pages 2988-2997. PMLR, 2017. 6", + "[33] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 3", + "[34] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. LocalTrans: A multiscale local transformer network for cross-resolution homography estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14890-14899, 2021. 1, 2, 3, 5, 6, 7", + "[35] Xiaoyong Shen, Li Xu, Qi Zhang, and Jiaya Jia. Multimodal and multi-spectral registration for natural images. In Proceedings of the European Conference on Computer Vision, pages 309-324. Springer, 2014. 8", + "[36] Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator. Advances in Neural Information Processing Systems, 36, 2023. 1", + "[37] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. RFNet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19679-19688, 2022. 3", + "[38] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13117-13125, 2021. 1, 2, 5, 6, 7", + "[39] Jiacheng Ying, Hui-Liang Shen, and Si-Yuan Cao. Unaligned hyperspectral image fusion via registration and interpolation modeling. IEEE Transactions on Geoscience and Remote Sensing, 2021. 1" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proceedings of the European Conference on Computer Vision, pages 653-669. Springer, 2020. 1, 2, 5, 6, 7", + "[41] Runmin Zhang, Jun Ma, Si-Yuan Cao, Lun Luo, Beinan Yu, Shu-Jie Chen, Junwei Li, and Hui-Liang Shen. SCP-Net: Unsupervised cross-modal homography estimation via intra-modal self-supervised learning. In Proceedings of the European Conference on Computer Vision, pages 460–477. Springer, 2024. 1, 2, 3, 4, 5, 6, 7", + "[42] Yiming Zhao, Xinming Huang, and Ziming Zhang. Deep lucas-kanade homography for multimodal image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15950–15959, 2021. 1, 2, 6", + "[43] Huayi Zhou, Fei Jiang, and Hongtao Lu. SSDA-YOLO: Semi-supervised domain adaptive yolo for cross-domain object detection. Computer Vision and Image Understanding, 229: 103649, 2023. 6", + "[44] Yuan Zhou, Anand Rangarajan, and Paul D Gader. An integrated approach to registration and fusion of hyperspectral and multispectral images. IEEE Transactions on Geoscience and Remote Sensing, 58(5):3020-3033, 2019. 1", + "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2223-2232, 2017. 3" + ], + "bbox": [ + 516, + 92, + 906, + 500 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16694", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_model.json b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..27d2d000c659edef097991deb865a6d9e04ac7d2 --- /dev/null +++ b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_model.json @@ -0,0 +1,2497 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.812, + 0.177 + ], + "angle": 0, + "content": "SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.203, + 0.79, + 0.24 + ], + "angle": 0, + "content": "Junchen Yu\\(^{1,2}\\) Si-Yuan Cao\\(^{1,2,3*}\\) Runmin Zhang\\(^{2}\\) Chenghao Zhang\\(^{2}\\) Zhu Yu\\(^{2}\\) Shujie Chen\\(^{4}\\) Bailin Yang\\(^{4}\\) Hui-Liang Shen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.242, + 0.892, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Ningbo Innovation Center, Zhejiang University \\(^{2}\\)College of Information Science and Electronic Engineering, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.258, + 0.862, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\) NingboTech University \\(^{4}\\) Zhejiang Key Laboratory of Big Data and Future ECommerce Technology, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.278, + 0.76, + 0.293 + ], + "angle": 0, + "content": "{yujunchen,cao_siyuan,runmin_zhang,zch00,yu_zhu}@zju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.296, + 0.713, + 0.31 + ], + "angle": 0, + "content": "{chenshujie,ybl}@zjgsu.edu.cn shenhl@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.378, + 0.486, + 0.756 + ], + "angle": 0, + "content": "We propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation into two supervised sub-problems, each addressed by its specialized network: a homography estimation network and a modality transfer network. To realize stable training, we introduce an effective split optimization strategy to train each network separately within its respective sub-problem. We also formulate an extra homography feature space supervision to enhance feature consistency, further boosting the estimation accuracy. Moreover, we employ a simple yet effective distillation training technique to reduce model parameters and improve cross-domain generalization ability while maintaining comparable performance. The training stability of SSHNet enables its cooperation with various homography estimation architectures. Experiments reveal that the SSHNet using IHN as homography estimation network, namely SSHNet-IHN, outperforms previous unsupervised approaches by a significant margin. Even compared to supervised approaches MHN and LocalTrans, SSHNet-IHN achieves \\(47.4\\%\\) and \\(85.8\\%\\) mean average corner errors (MACES) reduction on the challenging OPT-SAR dataset. Source code is available at https://github.com/Junchen-Yu/SSHNet." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Cross-modal homography estimation aims to solve the global perspective transformation between two images under different modalities. It is widely used in various computer vision tasks, such as GPS-denied robotic localization [14, 36]," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.349, + 0.589, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.406, + 0.589, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.347, + 0.742, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.347, + 0.902, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.467, + 0.908, + 0.535 + ], + "angle": 0, + "content": "Figure 1. Comparison of the solution space between (a) Content consistency indirect supervision and (b) Homography direct supervision. The content consistency indirect supervision may mislead the training due to local minima, while homography direct supervision won't." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.656 + ], + "angle": 0, + "content": "multi-modal image restoration [10, 27], and multi-spectral image fusion [39, 44]. Recent supervised homography estimation approaches [4, 5, 9, 20, 34, 42] have demonstrated their superiority in handling large deformation and modality gaps. However, since multi-modal images are captured by different sensors, the ground-truth homography deformations are usually unknown in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.911, + 0.902 + ], + "angle": 0, + "content": "To avoid using labeled data, unsupervised homography estimation approaches have emerged. Existing approaches [19, 28, 38, 40] typically optimize the similarity between the warped source image and target image through intensity-based losses. Some multi-modal image registration approaches [1, 11] transferred an image into the other modality, using mono-modal similarity metrics on image content to measure the correctness of motion. The previous state-of-the-art (SOTA) approach [41] combined cross-modal intensity-based learning with intra-modal self-supervised learning for homography knowledge injection. Notably, almost all the previous approaches fully or partially rely on the content consistency of images or feature maps to indirectly assess the correctness of homography estimation. As illustrated in Fig. 1a, due to the non-convexity of the image signal, loss value of content consistency indirect supervision has" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "multiple local minima across its solution space, and hence can not guarantee accurate homography prediction, especially under large modality and deformation gaps. On the contrary, the homography direct supervision in Fig. 1b can avoid such drawback, but it is unavailable in unsupervised cross-modal homography estimation, making the problem more challenging." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.487, + 0.621 + ], + "angle": 0, + "content": "To cope with the above problem, we propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation problem into two imperfect sub-problems: Sub-problem I: A supervised mono-modal homography estimation problem, where the modalities of input images are not perfectly consistent. Sub-problem II: A supervised modality transfer problem, where the supervisory image is not perfectly aligned. We introduce a split optimization strategy to separately train a homography estimation network for Sub-problem I and a modality transfer network for Sub-problem II, enabling effective convergence of the framework. Thanks to the proposed problem reformulation and the well-designed optimization strategy, our framework effectively solves the unsupervised task by training the networks under totally direct supervision, showing significantly improved training stability. This enables the successful training of iterative homography estimation architectures like IHN [4] and RHWF [5]. Better still, for the correlation-based homography estimation architectures [4, 5, 41], we propose an extra homography feature space supervision that enhances feature consistency, further improving homography estimation accuracy. Moreover, to address the computational burden of the modality transfer network, we introduce a simple yet effective distillation training technique, which notably reduces model parameters and improves cross-domain generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.487, + 0.865 + ], + "angle": 0, + "content": "We evaluate SSHNet on a variety of datasets including GoogleMap [42], DPDN [30], OPT-SAR [23], and Flash/no-flash [15] cross-modal datasets, together with RGB/NIR [2] cross-spectral datasets. SSHNet outperforms all the previous unsupervised homography estimation approaches with a large gap, and even achieves better performance compared to many supervised ones. The SSHNet using IHN as homography estimation network, namely SSHNet-IHN, owns \\(47.4\\%\\) and \\(85.8\\%\\) lower mean average corner errors (MACEs) than the supervised approaches MHN [20] and LocalTrans [34] on the challenging OPT-SAR dataset. Moreover, the SSHNet using the homography estimation architecture of the previous unsupervised SOTA approach SCPNet [41], namely SSHNet-SCPNet, even outperforms SCPNet itself, revealing the effectiveness of our proposed learning framework. To summarize, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.486, + 0.902 + ], + "angle": 0, + "content": "- We propose SSHNet, a novel unsupervised cross-modal homography estimation learning framework, which refor" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "mulates the unsupervised problem into two sub-problems with direct supervision. To realize stable training, we introduce a split optimization strategy that separately trains the network in each sub-problem." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.907, + 0.213 + ], + "angle": 0, + "content": "- For correlation-based homography estimation architectures, we introduce an extra homography feature space supervision to improve feature consistency for homography estimation, which further improves estimation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.288 + ], + "angle": 0, + "content": "- Considering the computational burden introduced by the additional modality transfer network, we devise a simple yet effective distillation training technique to reduce model parameters while improving cross-domain generalization ability with little impact on performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.907, + 0.364 + ], + "angle": 0, + "content": "- The training stability of SSHNet enables its cooperation with various homography estimation architectures, especially iterative ones like IHN and RHWF, resulting in higher performance. SSHNet-IHN outperforms the previous unsupervised approaches by a significant margin." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.907, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.665, + 0.393 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Supervised Deep Homography Estimation. DeTone et al. [9] introduced DHN, a VGG-style network to predict the homography between the source and target images. Le et al. [20] introduced a multi-scale VGG-style network, named MHN. Chang et al. [7] combined CNN with an untrainable IC-LK iterator. Zhao et al. [42] enhanced the feature similarity to improve the performance of IC-LK iterator. Shao et al. [34] proposed LocalTrans that adopts a multi-scale local transformer to register cross-resolution images. Cao et al. [4] proposed IHN, a trainable iterative framework with substantial accuracy. Cao et al. [5] further introduced RHWF that integrates homography-guided image warping and focus transformer, achieving higher performance. Deng et al. [8] proposed CrossHomo, a multi-level framework for cross-modality and cross-resolution homography estimation. Li et al. [22] introduced DMHomo, a diffusion-based framework to generate high-quality training pairs for supervised homography learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Unsupervised Deep Homography Estimation. Nguyen et al. [28] proposed UDHN, an unsupervised homography estimation network trained with photometric loss. Zhang et al. [40] introduced CA-UDHN, projecting images into a consistent feature space for effective estimation. Koguchiuk et al.[19] proposed biHomE, using perceptual loss[18] to improve robustness to intensity and viewpoint variations. Ye et al. [38] proposed BasesHomo, which enforces the image feature to be warp-equivalent with a feature identity loss and decomposes the homography matrix into flow bases. As modality and deformation gaps grow, intensity-based metrics for cross-modal homography matching may fail [41]. To address this, Zhang et al. [41] introduced SCPNet, which combines cross-modal intensity-based learning with intramodal self-supervised learning for homography knowledge" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "16686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.478, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Reformulating the unsupervised cross-modal homography estimation problem into two sub-problems with direct supervision. (a) Unsupervised cross-modal homography estimation problem without direct supervision. (b) Mono-modal homography estimation problem with imperfect inputs. (c) Modality transfer problem with imperfect supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.44 + ], + "angle": 0, + "content": "injection. Besides, some unsupervised approaches, such as NeMAR [1], RFNet [37], and UMF-CMGR [11], use modality transfer networks to convert one modality into the other, enabling cross-modal motion estimation. The above unsupervised approaches perform well on images with small deformations, but may produce unsatisfactory results under large deformations and modality gaps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Image-to-Image Translation. The image-to-image (I2I) translation problem focuses on converting images from a source domain to a target domain. Isola et al. [17] introduced Pix2pix, which employs conditional GANs and requires paired training data. Zhu et al. [45] later proposed CycleGAN, utilizing cycle-consistent loss to enable unpaired I2I translation. Liu et al.[24] developed UNIT, and Huang et al.[16] introduced MUNIT, both leveraging a shared latent space assumption to improve unpaired translation. More recently, Sasaki et al.[33] proposed UNIT-DDPM, integrating diffusion probabilistic models into the I2I translation framework. Li et al. [21] introduced BBDM, modeling I2I translation as a stochastic Brownian Bridge process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.315, + 0.665 + ], + "angle": 0, + "content": "3. Problem Reformulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Denote \\(\\mathbf{I}_{\\mathrm{A}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\) as a pair of images from modality A and B, with an unknown homography \\(\\mathbf{H}_{\\mathrm{AB,GT}}\\). As illustrated in Fig. 2a, the unsupervised cross-modal homography estimation aims to train a network to estimate the homography \\(\\hat{\\mathbf{H}}_{\\mathrm{AB}}\\), without direct supervision of \\(\\mathbf{H}_{\\mathrm{AB,GT}}\\). To solve this problem, in SSHNet, we reformulate it into two sub-problems with direct supervision, including a monomodal homography estimation problem with imperfect inputs (Fig. 2b) and a modality transfer problem with imperfect supervision (Fig. 2c). The two sub-problems are mutually causal, and the imperfect condition of each sub-problem can be gradually addressed through the optimization of another sub-problem. In this way, the unsupervised homography estimation problem can be solved in a totally supervised manner." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.91, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.223, + 0.908, + 0.293 + ], + "angle": 0, + "content": "Figure 3. Comparison of straight optimization and split optimization with regularization. (a) Cross-modal test MACEs for the network using straight optimization and split optimization with regularization, on OPT-SAR dataset. (b) Schematic diagram of straight optimization and split optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Sub-problem I. For the mono-modal homography estimation, generating image pairs with simulated ground-truth deformations for the direct self-supervised training is a common approach [4, 9, 20, 34]. If we can ideally convert \\(\\mathbf{I}_{\\mathrm{A}}\\) into modality B, producing the transferred image \\(\\mathbf{I}_{\\mathrm{A,T}}\\), we can treat the unsupervised cross-modal homography estimation problem as the mono-modal one by estimating the homography between \\(\\mathbf{I}_{\\mathrm{A,T}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\), as in Fig. 2b. The homography estimation network can be effectively trained using \\(\\mathbf{I}_{\\mathrm{B}}\\) with the simulated homography deformations, in a self-supervised manner. We refer to the problem of estimating the homography between \\(\\mathbf{I}_{\\mathrm{B}}\\) and the transferred image \\(\\mathbf{I}_{\\mathrm{A,T}}\\) with the self-supervised training using \\(\\mathbf{I}_{\\mathrm{B}}\\) as Sub-problem I." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.746 + ], + "angle": 0, + "content": "Sub-problem II. As illustrated in Fig. 2c, we also need a modality transfer network to obtain the optimal mapping from \\(\\mathbf{I}_{\\mathrm{A}}\\) to \\(\\mathbf{I}_{\\mathrm{B}}\\), producing \\(\\mathbf{I}_{\\mathrm{A,T}}\\). The ideal network training typically requires \\(\\mathbf{I}_{\\mathrm{B}}\\) that is well-aligned with \\(\\mathbf{I}_{\\mathrm{A}}\\) for supervision, which is also unavailable in the unsupervised cross-modal homography estimation. Previous approaches [1, 11] attempt to use GAN-based supervision to achieve modality transfer training using the unaligned image pairs. However, this strategy encounters considerable challenges when dealing with large deformation and modality gaps, as illustrated in [41]. Therefore, high-quality modality transfer requires well-aligned \\(\\mathbf{I}_{\\mathrm{A}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\). Now that we have transferred \\(\\mathbf{I}_{\\mathrm{A}}\\) to \\(\\mathbf{I}_{\\mathrm{A,T}}\\), the requirement becomes that Sub-problem I being ideally solved. We refer to the modality transfer problem, which uses the homography estimated in Sub-problem I to produce a warped \\(\\mathbf{I}_{\\mathrm{B}}\\) for supervision, as Sub-problem II." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Straight Optimization Fails. However, straight optimization of both sub-problems simultaneously is challenging, as the settlement of either sub-problem depends on the assumption that the other has already been solved. We conduct a pilot experiment where the homography estimation network and the modality transfer network are trained simultaneously to address Sub-problem I and II, but the networks fail to converge, as shown by the blue curve in Fig. 3a." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Split Optimization with Regularization Success. Inspired by decomposition methods in convex optimization," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16687" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.098, + 0.29, + 0.11 + ], + "angle": 0, + "content": "(a) Optimization of Sub-problem I" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.11, + 0.396, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.213, + 0.396, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.094, + 0.901, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.242, + 0.432, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.244, + 0.583, + 0.256 + ], + "angle": 0, + "content": "Homography Estimation Network" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.242, + 0.629, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.244, + 0.752, + 0.256 + ], + "angle": 0, + "content": "Modality Transfer Network" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.244, + 0.78, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.244, + 0.822, + 0.256 + ], + "angle": 0, + "content": "Trainable" + }, + { + "type": "image", + "bbox": [ + 0.839, + 0.244, + 0.863, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.865, + 0.244, + 0.89, + 0.256 + ], + "angle": 0, + "content": "rozen" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.267, + 0.431, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.269, + 0.671, + 0.282 + ], + "angle": 0, + "content": "Feature Extractor in Homography Estimation Network" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.268, + 0.778, + 0.281 + ], + "angle": 0, + "content": "Weight Sharing" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.268, + 0.824, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.268, + 0.892, + 0.281 + ], + "angle": 0, + "content": "mage Warping" + }, + { + "type": "image_footnote", + "bbox": [ + 0.405, + 0.295, + 0.896, + 0.307 + ], + "angle": 0, + "content": "Only Training Stage Both Training and Inference Stage Simulated Homography Generator" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Schematic diagram of Split Supervised Homography estimation Network, namely SSHNet. (a) Detailed optimization process of Sub-problem I. (b) Detailed optimization process of Sub-problem II. (c) Extra supervision on the homography feature space." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.369, + 0.485, + 0.761 + ], + "angle": 0, + "content": "we then consider using a split optimization strategy to address the two sub-problems separately. During optimization, each network updates its weights with the other one's frozen. As in Fig. 2b and Fig. 2c, in each iteration, the homography estimation network in Sub-problem I is updated, producing a better-aligned \\(\\mathbf{I}_{\\mathrm{B}}\\) to guide the modality transfer network in Sub-problem II in the next iteration. Similarly, the modality transfer network in Sub-problem II is updated, generating a better transferred image \\(\\mathbf{I}_{\\mathrm{A,T}}\\) to serve as input for the homography estimation network in Sub-problem I. This process gradually improves the outputs of each network during training, facilitating the optimization of the other network, and finally makes both sub-problems nearly perfect supervised ones. However, at the beginning of the training, the modality gap between \\(\\mathbf{I}_{\\mathrm{A,T}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\) in Sub-problem I is extremely large, so is the misalignment of \\(\\mathbf{I}_{\\mathrm{A}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\) in Sub-problem II, which hinders the mutual promotion between two subproblems. A regularization term that guides the training into our desired direction is needed. Inspired by [41] \\(^{1}\\), in Subproblem I, we replace the self-supervised training using only \\(\\mathbf{I}_{\\mathrm{B}}\\) by the two-branch self-supervised training using both \\(\\mathbf{I}_{\\mathrm{A,T}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\) (see Fig. 2b). This enables the homography estimation network to implicitly learn cross-modal homography knowledge at the beginning of the training, and triggers the two split optimized sub-problems to converge into our desired direction, as shown by the orange curve in Fig. 3a." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.182, + 0.789 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.484, + 0.844 + ], + "angle": 0, + "content": "Based on the problem reformulation in Section 3, we further investigate its potential and propose an unsupervised cross-modal homography estimation learning framework, named" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Split Supervised Homography estimation Network (SSHNet). Section 4.1 covers the split optimization of SSHNet (Fig. 4a and Fig. 4b), Section 4.2 discusses the cooperation of SSHNet with various homography estimation architectures, Section 4.3 outlines the design of the modality transfer network, Section 4.4 presents the extra homography feature space supervision (Fig. 4c), and Section 4.5 illustrates the distillation training technique (Fig. 5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.693, + 0.518 + ], + "angle": 0, + "content": "4.1. Split Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.601 + ], + "angle": 0, + "content": "As illustrated in Fig. 4a and Fig. 4b, the homography estimation network and the modality transfer network are optimized separately in each iteration, forming the split optimization framework of our SSHNet. In the following, we will elaborate the optimization procedure for each network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.909, + 0.783 + ], + "angle": 0, + "content": "Optimization of Sub-problem I. As mentioned in Section 3, we adopt a two-branch self-supervised learning strategy to optimize the homography estimation network in Subproblem I. We apply the separate simulated homography deformation on \\(\\mathbf{I}_{\\mathrm{A}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\), generating the intra-modal image pairs \\((\\mathbf{I}_{\\mathrm{A}},\\mathbf{I}_{\\mathrm{A}}^{\\prime})\\) and \\((\\mathbf{I}_{\\mathrm{B}},\\mathbf{I}_{\\mathrm{B}}^{\\prime})\\) with corresponding ground-truth homography \\(\\mathbf{H}_{\\mathrm{A,GT}}\\) and \\(\\mathbf{H}_{\\mathrm{B,GT}}\\). The modality transfer network transfers \\((\\mathbf{I}_{\\mathrm{A}},\\mathbf{I}_{\\mathrm{A}}^{\\prime})\\) into \\((\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime})\\). We simultaneously use \\((\\mathbf{I}_{\\mathrm{B}},\\mathbf{I}_{\\mathrm{B}}^{\\prime})\\) and \\((\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime})\\) to train the weight-shared homography estimation network in a self-supervised manner. The optimization of Sub-problem I can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.798, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\underset {\\zeta} {\\arg \\min } L _ {\\mathrm {H}} \\left(\\mathcal {H} _ {\\zeta} \\left(\\mathbf {I} _ {\\mathrm {B}}, \\mathbf {I} _ {\\mathrm {B}} ^ {\\prime}\\right), \\mathbf {H} _ {\\mathrm {B}, \\mathrm {G T}}\\right) + \\mathcal {R} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}} ^ {\\prime}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_{\\zeta}\\) denotes the homography estimation network with parameters \\(\\zeta\\) to be optimized, and \\(L_{\\mathrm{H}}\\) denotes the homography estimation loss. \\(\\mathcal{R}(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime}) = L_{\\mathrm{H}}\\bigl (\\mathcal{H}_{\\zeta}(\\mathcal{T}_{\\theta^{*}}(\\mathbf{I}_{\\mathrm{A}}),\\mathcal{T}_{\\theta^{*}}(\\mathbf{I}_{\\mathrm{A}}^{\\prime})),\\mathbf{H}_{\\mathrm{A,GT}}\\bigr)\\) , which is the regular" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] finds that if training the weight-shared homography estimation network using the simulated deformation within each two modalities, the cross-modal homography knowledge can be implicitly learned through the multitask learning mechanism [6, 12]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16688" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.121 + ], + "angle": 0, + "content": "ization term, where \\(\\mathcal{T}_{\\theta^*}\\) denotes the modality transfer network with parameters \\(\\theta^{*}\\) frozen." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.228 + ], + "angle": 0, + "content": "Optimization of Sub-problem II. We train a modality transfer network to convert \\(\\mathbf{I}_{\\mathrm{A}}\\) into an image of pseudo modality B, namely \\(\\mathbf{I}_{\\mathrm{A,T}}\\). Following the reformulation in Section 3, we apply the homography matrix \\(\\hat{\\mathbf{H}}_{\\mathrm{AB}}\\) predicted by \\(\\mathcal{H}_{\\zeta}\\) to warp \\(\\mathbf{I}_{\\mathrm{B}}\\), providing direct supervision for the modality transfer network. The optimization of Sub-problem II can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.238, + 0.484, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } L _ {\\mathrm {T}} \\left(\\mathcal {T} _ {\\theta} (\\mathbf {I} _ {\\mathrm {A}}), \\mathcal {W} (\\mathbf {I} _ {\\mathrm {B}}, \\mathcal {H} _ {\\zeta^ {*}} (\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {B}}))\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.484, + 0.345 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}_{\\theta}\\) denotes the modality transfer network with parameters \\(\\theta\\) to be optimized, \\(\\mathcal{H}_{\\zeta^{*}}\\) denotes the homography estimation network with parameters \\(\\zeta^{*}\\) frozen, \\(\\mathcal{W}\\) denotes the image warping with the estimated homography, and \\(L_{\\mathrm{T}}\\) denotes the modality transfer loss." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.353, + 0.388, + 0.369 + ], + "angle": 0, + "content": "4.2. Homography Estimation Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Present homography estimation approaches exhibit a variety of architectural designs. Some approaches [9, 20, 28, 38, 40] concatenate input images in the channel dimension, directly estimating the homography deformation. Others [4, 5, 34, 41] use backbone networks to extract features and employ the feature correlation to perform homography decoding. Additionally, recent approaches [4, 5] further introduce iterative architectures to refine the estimation results, achieving substantial improvement in accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.722 + ], + "angle": 0, + "content": "As illustrated in the experimental results of Section 5.3, our proposed framework is independent of the specific homography estimation architecture, enabling integration with various architectures. The pure supervised training strategy significantly boosts the stability of the optimization process. As a result, iterative homography estimation architectures like IHN [4] and RHWF [5] can be effectively employed, leading to a higher performance upper bound. In contrast, due to the lack of an explicit solution for consistent features between cross-modal images, the previous SOTA approach, SCPNet [41], struggles to converge when using these iterative architectures. For the self-supervised training, we adopt \\(L_{1}\\) homography estimation loss following previous works [4, 5, 9, 20, 41], which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.729, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {H}} (\\hat {\\mathbf {H}}, \\mathbf {H} _ {\\mathrm {G T}}) = \\left\\| \\hat {\\mathbf {H}} - \\mathbf {H} _ {\\mathrm {G T}} \\right\\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.826 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{H}}\\) denotes the estimated homography, \\(\\mathbf{H}_{\\mathrm{GT}}\\) denotes the ground-truth homography. For iterative homography architectures, we use a weighted sum of the \\(L_{1}\\) distance following [4, 5]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.34, + 0.85 + ], + "angle": 0, + "content": "4.3. Modality Transfer Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We design the modality transfer network in our SSHNet as a U-shaped architecture to enhance the perception of multiscale information. Inspired by Swin-Unet [3], the Swin" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "Transformer block [25] serves as the basic unit of the modality transfer network. The attention mechanism enhances the network's ability to capture hierarchical structures and context information. This facilitates more effective learning of complex mappings between modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.289 + ], + "angle": 0, + "content": "As mentioned in Section 4.1, we warp image \\(\\mathbf{I}_{\\mathrm{B}}\\) with the \\(\\hat{\\mathbf{H}}_{\\mathrm{AB}}\\) predicted by the homography estimation network for the modality transfer supervision. During optimization, \\(\\hat{\\mathbf{H}}_{\\mathrm{AB}}\\) cannot ideally align the input image \\(\\mathbf{I}_{\\mathrm{A}}\\) with the supervisory image \\(\\mathbf{I}_{\\mathrm{B}}\\). Therefore, instead of using pixel-wise \\(L_{1}\\) or \\(L_{2}\\) loss, we apply perceptual loss \\(L_{\\mathrm{pcp}}\\) [18] to capture high-level representations, ensuring better tolerance to misalignments, which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.3, + 0.928, + 0.35 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {p c p}} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {B}, \\mathrm {W}}\\right) = \\sum_ {j} \\frac {1}{C _ {j} H _ {j} W _ {j}} \\left\\| \\psi_ {j} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}\\right) - \\psi_ {j} \\left(\\mathbf {I} _ {\\mathrm {B}, \\mathrm {W}}\\right) \\right\\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.906, + 0.398 + ], + "angle": 0, + "content": "where \\(\\psi_{j}\\) is the \\(j\\)-th layer of the pre-trained VGG with a feature map of shape \\(C_j\\times H_j\\times W_j\\), \\(\\mathbf{I}_{\\mathrm{B,W}}\\) denotes the image \\(\\mathbf{I}_{\\mathrm{B}}\\) warped by the estimated \\(\\hat{\\mathbf{H}}_{\\mathrm{AB}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.906, + 0.424 + ], + "angle": 0, + "content": "4.4. Extra Homography Feature Space Supervision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.52 + ], + "angle": 0, + "content": "The current modality transfer network in Sub-problem II is supervised by the perceptual loss [18], which computes feature similarity using a pre-trained VGG network. This may yield suboptimal results for homography estimation in Sub-problem I when trained under the present totally split optimization framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.673 + ], + "angle": 0, + "content": "To address this issue, as illustrated in Fig. 4c, we introduce an extra supervision in the homography feature space when the architecture of homography estimation network has explicit feature extractors, such as IHN, RHWF, and SCPNet. A homography feature (HF) loss that directly enhances the feature consistency for homography estimation between \\(\\mathbf{I}_{\\mathrm{A,T}}\\) and \\(\\mathbf{I}_{\\mathrm{B}}\\) is computed. The modality transfer network and the feature extractor of the homography estimation network are optimized together under the supervision of HF loss, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.685, + 0.907, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta , \\xi} {\\arg \\min } L _ {\\mathrm {H F}} \\left(\\mathcal {F} _ {\\xi} \\left(\\mathcal {T} _ {\\theta} (\\mathbf {I} _ {\\mathrm {A}})\\right), \\mathcal {F} _ {\\xi} \\left(\\mathcal {W} \\left(\\mathbf {I} _ {\\mathrm {B}}, \\hat {\\mathbf {H}} _ {\\mathrm {A B}}\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.906, + 0.814 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\xi}\\) denotes the feature extractor with parameters \\(\\xi\\) to be optimized. We note that IHN, RHWF, and SCPNet all depend on feature correlation to perform homography decoding. Considering this, We formulate the HF loss using a correlation-based loss function \\(L_{\\mathrm{corr}}\\) to enhance feature consistency, which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.826, + 0.907, + 0.859 + ], + "angle": 0, + "content": "\\[\nL _ {\\operatorname {c o r r}} \\left(\\mathbf {F} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {F} _ {\\mathrm {B}, \\mathrm {W}}\\right) = - \\sum_ {i, j} \\mathbf {F} _ {\\mathrm {A}, \\mathrm {T}} (i, j) ^ {\\top} \\mathbf {F} _ {\\mathrm {B}, \\mathrm {W}} (i, j), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{\\mathrm{A,T}}\\) and \\(\\mathbf{F}_{\\mathrm{B,W}}\\) denote the feature maps of \\(\\mathbf{I}_{\\mathrm{A,T}}\\) and \\(\\mathbf{I}_{\\mathrm{B,W}}\\) in the homography estimation network." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16689" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.091, + 0.468, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.151, + 0.454, + 0.165 + ], + "angle": 0, + "content": "Figure 5. Schematic diagram of the distillation training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.171, + 0.285, + 0.187 + ], + "angle": 0, + "content": "4.5. Distillation Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.193, + 0.484, + 0.284 + ], + "angle": 0, + "content": "We can see that the present framework has extra network parameters due to the utilization of the modality transfer network in Sub-problem II. Moreover, as the modality transfer network is trained on multi-modal image pairs from a fixed domain, its generalization ability to images from other domains may be limited." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.485, + 0.435 + ], + "angle": 0, + "content": "To cope with the above problem, we introduce a distillation training technique under the inspiration of the unsupervised knowledge distillation frameworks in domain adaptive object detection [32, 43]. We use SSHNet as a teacher network to guide a student network that only retains the homography estimation network, named Distilled SSHNet (SSHNet-D). During the distillation training, the homography \\(\\hat{\\mathbf{H}}_{\\mathrm{teacher}}\\) predicted by SSHNet serves as a pseudo ground-truth to supervise SSHNet-D, as illustrated in Fig. 5. The distillation training can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.445, + 0.484, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\underset {\\omega} {\\arg \\min } L _ {\\mathrm {H}} \\left(\\phi_ {\\omega} \\left(\\mathbf {I} _ {\\mathrm {A}}, \\mathbf {I} _ {\\mathrm {B}}\\right), \\hat {\\mathbf {H}} _ {\\text {t e a c h e r}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.51 + ], + "angle": 0, + "content": "where \\(\\phi_{\\omega}\\) denotes the SSHNet-D with the parameters \\(\\omega\\) to be optimized." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.224, + 0.539 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.299, + 0.563 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.674 + ], + "angle": 0, + "content": "Datasets. We evaluate SSHNet on various datasets, including GoogleMap [42], DPDN [30], OPT-SAR [23], Flash/no-flash [15] cross-modal datasets, and RGB/NIR [2] cross-spectral dataset. Examples for each dataset are shown in Fig. 6. To ensure a fair comparison, all evaluated methods are trained and tested using the same training and test splits within each dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Baselines. We evaluate SSHNet with traditional feature-based approaches including SIFT [26], ORB [31], unsupervised approaches including UDHN [28], biHomE [19], CA-UDHN [40], BasesHomo [38], UMF-CMGR [11], SCP-Net [41], and supervised approaches including DHN [9], MHN [20], LocalTrans [34], IHN [4], RHWF [5]. For feature-based approaches, we choose RANSAC [13] as the outlier rejection algorithms. UMF-CMGR is an image fusion approach based on registration, and we use the registration part for comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Homography Estimation Architectures. To investigate the potential of our proposed SSHNet, we adopt various homography estimation architectures including DHN [9] and MHN [20], together with iterative architectures such as IHN [4] and RHWF [5]. Moreover, we also adopt the" + }, + { + "type": "table_caption", + "bbox": [ + 0.533, + 0.09, + 0.885, + 0.104 + ], + "angle": 0, + "content": "Table 1. Ablation study on the training strategy of SSHNet." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.106, + 0.903, + 0.187 + ], + "angle": 0, + "content": "
ExperimentMACE ↓
w/o problem reformulationNC
w/o split optimizationNC
pure self-supervised learning24.60
SSHNet2.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.906, + 0.22 + ], + "angle": 0, + "content": "Table 2. Ablation study on the architectures of the modality transfer network." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.224, + 0.903, + 0.275 + ], + "angle": 0, + "content": "
ExperimentParameters (M)MACE ↓
CNN-based7.488.72
Transformer-based7.542.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.282, + 0.906, + 0.311 + ], + "angle": 0, + "content": "Table 3. Ablation study on the loss functions. \\( L_{\\mathrm{T}} \\) denotes the modality transfer loss, \\( L_{\\mathrm{HF}} \\) denotes the homography feature loss." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.313, + 0.903, + 0.417 + ], + "angle": 0, + "content": "
ExperimentLoss TypeMACE ↓
LTL15.88
L26.23
Lpcp4.52
LT + LHFLpcp + L14.09
Lpcp + L23.27
Lpcp + Lcorr2.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.424, + 0.882, + 0.438 + ], + "angle": 0, + "content": "Table 4. Ablation study of the modality transfer direction." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.441, + 0.903, + 0.498 + ], + "angle": 0, + "content": "
Method\\DirectionRGB→NIRNIR→RGBOPT→SARSAR→OPT
SSHNet1.661.752.944.28
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.553 + ], + "angle": 0, + "content": "architecture from the previous SOTA unsupervised approach SCPNet [41], to specifically evaluate the effectiveness of our proposed unsupervised learning framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Metrics. Similar to previous works [4, 5, 9, 34, 42], we randomly perturb the corner points of \\(128 \\times 128\\) images to make the deformed image, with a perturbation range set to [-32, 32]. We evaluate homography estimation accuracy using the mean average corner error (MACE) following [4, 5, 9, 34, 42]. A lower MACE indicates a higher accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.618, + 0.667 + ], + "angle": 0, + "content": "5.2. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.908, + 0.718 + ], + "angle": 0, + "content": "The ablation studies are conducted on the OPT-SAR dataset. We adopt IHN as the homography estimation network in our SSHNet for ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Training Strategy. We conduct an ablation study on the training strategy of our SSHNet. The experimental results in Table 1 demonstrate that both problem reformulation and the split optimization strategy are crucial for SSHNet. Without problem reformulation, unsupervised training with intensity-based supervision fails to converge. Even with problem reformulation, directly optimizing both networks simultaneously does not yield reliable homography estimation. SSHNet achieves effective convergence only when problem reformulation is combined with the split optimization strategy, where each network is trained separately. We also compare our framework with method trained through" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16690" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.174, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.148, + 0.21, + 0.157 + ], + "angle": 0, + "content": "(a) GoogleMap" + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.09, + 0.249, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.091, + 0.335, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.091, + 0.41, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.148, + 0.36, + 0.157 + ], + "angle": 0, + "content": "(b) DPDN" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.091, + 0.497, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.148, + 0.529, + 0.157 + ], + "angle": 0, + "content": "(c) OPT-SAR" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.571, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.091, + 0.658, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.091, + 0.733, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.148, + 0.699, + 0.157 + ], + "angle": 0, + "content": "(d) Flash/no-flash" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.091, + 0.818, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.148, + 0.85, + 0.157 + ], + "angle": 0, + "content": "(e) RGB/NIR" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.091, + 0.894, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.162, + 0.607, + 0.176 + ], + "angle": 0, + "content": "Figure 6. Examples for each dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.181, + 0.908, + 0.21 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of our SSHNet and other approaches. NC denotes the training is not converged. The best results of unsupervised methods are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.215, + 0.905, + 0.495 + ], + "angle": 0, + "content": "
Dataset ApproachGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
TraditionalSIFT [26]24.5325.1424.8218.2921.33
ORB [31]24.5224.8624.9020.2222.67
SupervisedDHN [9]5.204.928.276.4211.88
MHN [20]2.692.955.595.245.52
LocalTrans [34]3.221.7920.681.521.72
IHN [4]0.921.171.670.801.63
RHWF [5]0.711.061.480.651.07
UnsupervisedUDHN [28]22.84NCNC21.2023.43
CA-UDHN [40]24.6124.9924.7621.3224.12
biHomE [19]NCNCNC11.8623.77
BasesHomo [38]24.4927.3326.7225.1224.41
UMF-CMGR [11]24.6025.4624.7023.4922.38
SCPNet [41]4.35NCNC2.674.78
SSHNet-DHN (Ours)9.289.7117.6310.5112.13
SSHNet-MHN (Ours)2.903.046.775.626.93
SSHNet-SCPNet (Ours)3.894.8112.942.653.86
SSHNet-RHWF (Ours)1.291.263.080.951.52
SSHNet-IHN (Ours)1.231.122.941.081.66
SSHNet-IHN-D (Ours)1.261.243.311.161.72
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.553 + ], + "angle": 0, + "content": "pure self-supervised learning. Results in Table 1 show that self-supervised learning alone does not yield satisfactory homography estimation results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Architectures of the Modality Transfer Network. As illustrated in Table 2. Our Transformer-based architecture outperforms the CNN-based architecture by a large margin, with similar parameter counts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.805 + ], + "angle": 0, + "content": "Loss Functions. To further investigate the effectiveness of the modality transfer and homography feature loss functions, we first train the network without HF loss to evaluate the performance of different types of modality transfer loss. The results in the upper part of Table 3 indicate that the perceptual loss \\( L_{\\mathrm{pcp}} \\) outperforms \\( L_{1} \\) and \\( L_{2} \\). Then we adopt \\( L_{\\mathrm{pcp}} \\) as the modality transfer loss and add different types of HF loss to analyze their impact. We report the results in the lower part of Table 3. Though all different types of HF loss further boost the network's performance, the correlation-based loss \\( L_{\\mathrm{corr}} \\) achieves superior performance compared to \\( L_{1} \\) and \\( L_{2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Modality transfer direction. Table 4 presents the results of different transfer directions on the RGB/NIR and OPT-SAR datasets. Since some modalities carry more information, transferring from richer to less informative modalities facilitates the modality transfer network's learning process, thereby promoting the framework's convergence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.829, + 0.523 + ], + "angle": 0, + "content": "5.3. Comparisons with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.908, + 0.852 + ], + "angle": 0, + "content": "We compare the homography estimation results of our proposed SSHNet and baselines and list the experimental results in Table 5. It is observed that the prior traditional and unsupervised approaches can hardly handle the homography estimation under large modality and deformation gaps. In contrast, our method achieves much more satisfactory results with various homography estimation architectures. When adopting the iterative architectures IHN and RHWF, our method outperforms all the compared traditional and unsupervised approaches by a significant margin. Moreover, even though trained without direct cross-modal supervision, SSHNet-IHN and SSHNet-RHWF outperforms supervised approaches such as DHN, MHN and LocalTrans. Specifically, SSHNet-IHN owns \\(54.3\\%\\) and \\(47.4\\%\\) lower MACEs than MHN, and \\(61.8\\%\\) and \\(85.8\\%\\) lower MACEs than LocalTrans on the challenging GoogleMap and OPT-SAR datasets, respectively. Notably, SSHNet-SCPNet, which uses the homography estimation architecture from the previous unsupervised SOTA approach SCPNet, outperforms SCPNet itself on all datasets, further proving the superiority of our proposed training framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We also conduct the distillation training on SSHNet with different homography estimation architectures. We present the distillation training results of SSHNet-IHN in Table 5," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16691" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 6. Cross-dataset evaluation results. Red indicates the improved performance of SSHNet-IHN-D compared to SSHNet-IHN, and blue indicates the declined one." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.135, + 0.482, + 0.237 + ], + "angle": 0, + "content": "
Test TrainGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
GoogleMap/↑1.8%↑1.0%↑80.4%↑63.0%
DPDN↑66.5%/↑0.4%↑92.1%↑75.8%
OPT-SAR↑48.4%↑1.3%/↑77.5%↑66.2%
Flash/no-flash↓5.2%↓0.2%↑0.1%/↓4.6%
RGB/NIR↓0.5%↑1.2%↑3.0%↑2.1%/
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.485, + 0.308 + ], + "angle": 0, + "content": "namely SSHNet-IHN-D. We observe that the performance of SSHNet-IHN-D is comparable to that of SSHNet-IHN. For more detailed qualitative visualization and experimental results, please refer to supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.318, + 0.353, + 0.332 + ], + "angle": 0, + "content": "5.4. Cross-Dataset Generalization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.486, + 0.549 + ], + "angle": 0, + "content": "We further conduct a cross-dataset evaluation of SSHNet-IHN-D and SSHNet-IHN on the aforementioned datasets, to further reveal the effectiveness of our proposed distillation training. As illustrated in Fig. 6, multi-modal image pairs from the GoogleMap, DPDN, and OPT-SAR datasets exhibit substantial modality gaps, and we call these three datasets challenging datasets, for a better illustration. In contrast, the Flash/no-flash and RGB/NIR datasets show smaller modality gaps, so we call them easy datasets. For each dataset, we compute the percentage of homography predictions that have average corner errors (ACE) lower than 5 pixels on the test set, which reveals the ratio of reliable estimations. We directly list the differences of the percentage of SSHNet-IHN-D and SSHNet-IHN in Table 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.486, + 0.824 + ], + "angle": 0, + "content": "SSHNet-IHN uses a modality transfer network to transfer images from one modality to the other on the training dataset, making it highly prone to overfit the training data. In contrast, SSHNet-IHN-D excludes the modality transfer network and directly extracts consistent features from cross-modal images, and hence is forced to learn the features that is robust across different domains. As illustrated in Table 6, when trained on challenging datasets, SSHNet-IHN-D exhibits significant superior generalization ability. Specifically, SSHNet-IHN-D trained on DPDN yields \\(66.5\\%\\), \\(92.1\\%\\), and \\(75.8\\%\\) improvement on GoogleMap, Flash/no-flash and RGB/NIR datasets compared to SSHNet. We note that the generalization ability across the challenging datasets remains limited due to the substantial domain gaps between the datasets. As for training on easy datasets, the generalization performance improvement is plain. This is because the training data is too easy to provide enough knowledge for obtaining the features robust across different domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.295, + 0.849 + ], + "angle": 0, + "content": "5.5. Real World Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We further evaluate our method on real-world scenarios [35] that do not satisfy perfect homography alignment, including RGB/NIR and Flash/no-flash image pairs. As illustrated" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.697, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.207, + 0.667, + 0.219 + ], + "angle": 0, + "content": "(a) Cross modal" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.906, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.207, + 0.844, + 0.219 + ], + "angle": 0, + "content": "(b) UDIS-TIR" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.224, + 0.892, + 0.237 + ], + "angle": 0, + "content": "Figure 7. Visualization of results on real-world deformations." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Table 7. Experimental results on the UDIS-TIR dataset. \\( I_{3 \\times 3} \\) denotes the identity matrix." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.274, + 0.903, + 0.337 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑
I3×312.920.306
SCPNetNCNC
SSHNet-IHN20.050.728
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.908, + 0.497 + ], + "angle": 0, + "content": "in Fig. 7a, the model trained with simulated homographies achieves effective performance. We also use a pretrained RGB-to-TIR translation network to convert RGB-RGB image pairs from the UDIS [29] dataset into RGB-TIR image pairs, creating the UDIS-TIR dataset with parallax image pairs. Following UDIS, we evaluate SSHNet-IHN on the UDIS-TIR dataset and compute the PSNR and SSIM on RGB images before modality transfer. We present the quantitative results in Table 7 and show more visualization results in Fig. 7b." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.511, + 0.64, + 0.527 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.762 + ], + "angle": 0, + "content": "In this paper, we propose SSHNet, a novel unsupervised cross-modal homography estimation framework. SSHNet reformulates the problem into two supervised sub-problems, addressed by a homography estimation network and a modality transfer network. These networks are trained separately using a split optimization strategy, ensuring stable training. We also introduce an extra homography feature space supervision to further improve feature consistency during the homography estimation and boost the performance of our framework. Moreover, we employ a distillation training technique to reduce model parameters and improve cross-domain generalization ability. Experimental results show that our SSHNet achieves SOTA performance on multiple datasets among unsupervised approaches, and outperforms many supervised approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.909, + 0.888 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the Zhejiang Provincial Natural Science Foundation of China under grant LD24F020003, in part by the National Natural Science Foundation of China under grant 62301484, in part by the Ningbo Natural Science Foundation of China under grant 2024J454, and in part by the National Key Research and Development Program of China under grant 2023YFB3209800. We also thank the generous help from Jianxin Hu, Zhejiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16692" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.197 + ], + "angle": 0, + "content": "[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13410–13419, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.485, + 0.255 + ], + "angle": 0, + "content": "[2] Matthew Brown and Sabine Süssstrunk. Multispectral SIFT for scene category recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 177-184. IEEE, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[3] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 205-218. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1879-1888, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[5] Si-Yuan Cao, Runmin Zhang, Lun Luo, Beinan Yu, Zehua Sheng, Junwei Li, and Hui-Liang Shen. Recurrent homography estimation using homography-guided image warping and focus transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9833-9842, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[6] Rich Caruana. Multitask learning. Machine Learning, 28: 41-75, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.485, + 0.561 + ], + "angle": 0, + "content": "[7] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. CLKN: Cascaded Lucas-Kanade networks for image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2213-2221, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[8] Xin Deng, Enpeng Liu, Chao Gao, Shengxi Li, Shuhang Gu, and Mai Xu. CrossHomo: Cross-modality and cross-resolution homography estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[10] Fayaz Ali Dharejo, Muhammad Zawish, Farah Deeba, Yuanchun Zhou, Kapal Dev, Sunder Ali Khowaja, and Nawab Muhammad Faseeh Qureshi. Multimodal-boost: Multimodal medical image super-resolution using multi-attention network with wavelet transform. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[11] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2051-2060, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[14] Hunter Goforth and Simon Lucey. GPS-denied UAV localization using pre-existing satellite imagery. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 2974-2980. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[15] Shengfeng He and Rynson WH Lau. Saliency detection with flash and no-flash image pairs. In Proceedings of the European Conference on Computer Vision, pages 110-124. Springer, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.907, + 0.291 + ], + "angle": 0, + "content": "[16] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision, pages 172-189, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.359 + ], + "angle": 0, + "content": "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1125-1134, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision, pages 694–711. Springer, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[19] Daniel Koguciuk, Elahe Arani, and Bahram Zonooz. Perceptual loss for robust unsupervised homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4274-4283, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[21] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. BBDM: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[22] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. DMHomo: Learning homography with diffusion models. ACM Transactions on Graphics, 43 (3):1-16, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[23] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. MCANet: A joint semantic segmentation framework of optical and SAR images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106:102638, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[24] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in Neural Information Processing Systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16693" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[27] Iman Marivani, Evaggelia Tsiligianni, Bruno Cornelis, and Nikos Deligiannis. Designing CNNs for multimodal image restoration and fusion via unfolding the method of multipliers. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5830-5845, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[28] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[29] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Unsupervised deep image stitching: Reconstructing stitched features to images. IEEE Transactions on Image Processing, 30:6184-6197, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[30] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In *British Machine Vision Conference*. The British Machine Vision Association, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[31] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[32] Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain adaptation. In Proceedings of the International Conference on Machine Learning, pages 2988-2997. PMLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[33] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[34] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. LocalTrans: A multiscale local transformer network for cross-resolution homography estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14890-14899, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[35] Xiaoyong Shen, Li Xu, Qi Zhang, and Jiaya Jia. Multimodal and multi-spectral registration for natural images. In Proceedings of the European Conference on Computer Vision, pages 309-324. Springer, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[36] Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator. Advances in Neural Information Processing Systems, 36, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[37] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. RFNet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19679-19688, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[38] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13117-13125, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[39] Jiacheng Ying, Hui-Liang Shen, and Si-Yuan Cao. Unaligned hyperspectral image fusion via registration and interpolation modeling. IEEE Transactions on Geoscience and Remote Sensing, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[40] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proceedings of the European Conference on Computer Vision, pages 653-669. Springer, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[41] Runmin Zhang, Jun Ma, Si-Yuan Cao, Lun Luo, Beinan Yu, Shu-Jie Chen, Junwei Li, and Hui-Liang Shen. SCP-Net: Unsupervised cross-modal homography estimation via intra-modal self-supervised learning. In Proceedings of the European Conference on Computer Vision, pages 460–477. Springer, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.316 + ], + "angle": 0, + "content": "[42] Yiming Zhao, Xinming Huang, and Ziming Zhang. Deep lucas-kanade homography for multimodal image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15950–15959, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.908, + 0.373 + ], + "angle": 0, + "content": "[43] Huayi Zhou, Fei Jiang, and Hongtao Lu. SSDA-YOLO: Semi-supervised domain adaptive yolo for cross-domain object detection. Computer Vision and Image Understanding, 229: 103649, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.908, + 0.431 + ], + "angle": 0, + "content": "[44] Yuan Zhou, Anand Rangarajan, and Paul D Gader. An integrated approach to registration and fusion of hyperspectral and multispectral images. IEEE Transactions on Geoscience and Remote Sensing, 58(5):3020-3033, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.908, + 0.5 + ], + "angle": 0, + "content": "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2223-2232, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16694" + } + ] +] \ No newline at end of file diff --git a/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_origin.pdf b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc6ec0bb0bcafddeddf1aeb912a217e3b0ba255e --- /dev/null +++ b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/ae624eae-7853-4995-bcaa-12d94352fbaa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56f8c26509c9e34f5949345fb8167307fdd3021737eeda2dd5a562645207c596 +size 2432329 diff --git a/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/full.md b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b9097b6a07003892704b6adca4b7cf9afd05558 --- /dev/null +++ b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/full.md @@ -0,0 +1,358 @@ +# SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization + +Junchen Yu $^{1,2}$ Si-Yuan Cao $^{1,2,3*}$ Runmin Zhang $^{2}$ Chenghao Zhang $^{2}$ Zhu Yu $^{2}$ Shujie Chen $^{4}$ Bailin Yang $^{4}$ Hui-Liang Shen $^{2}$ + +$^{1}$ Ningbo Innovation Center, Zhejiang University $^{2}$ College of Information Science and Electronic Engineering, Zhejiang University + +$^{3}$ NingboTech University $^{4}$ Zhejiang Key Laboratory of Big Data and Future ECommerce Technology, Hangzhou, China + +{yujunchen,cao_siyuan,runmin_zhang,zch00,yu_zhu}@zju.edu.cn + +{chenshujie,ybl}@zjgsu.edu.cn shenhl@zju.edu.cn + +# Abstract + +We propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation into two supervised sub-problems, each addressed by its specialized network: a homography estimation network and a modality transfer network. To realize stable training, we introduce an effective split optimization strategy to train each network separately within its respective sub-problem. We also formulate an extra homography feature space supervision to enhance feature consistency, further boosting the estimation accuracy. Moreover, we employ a simple yet effective distillation training technique to reduce model parameters and improve cross-domain generalization ability while maintaining comparable performance. The training stability of SSHNet enables its cooperation with various homography estimation architectures. Experiments reveal that the SSHNet using IHN as homography estimation network, namely SSHNet-IHN, outperforms previous unsupervised approaches by a significant margin. Even compared to supervised approaches MHN and LocalTrans, SSHNet-IHN achieves $47.4\%$ and $85.8\%$ mean average corner errors (MACES) reduction on the challenging OPT-SAR dataset. Source code is available at https://github.com/Junchen-Yu/SSHNet. + +# 1. Introduction + +Cross-modal homography estimation aims to solve the global perspective transformation between two images under different modalities. It is widely used in various computer vision tasks, such as GPS-denied robotic localization [14, 36], + +![](images/7aa7f86c4f28fa35b10864c592c3e254349f52d3061ab072f5ca0a193486ea40.jpg) + +![](images/57824a5c34de5900a60e6036ee959f4730df1e9bbbf11fd99f54be6d4c6c98cc.jpg) +Figure 1. Comparison of the solution space between (a) Content consistency indirect supervision and (b) Homography direct supervision. The content consistency indirect supervision may mislead the training due to local minima, while homography direct supervision won't. + +![](images/9663e2d200187d1eb32f7f797be94e22716530d5f1e616cbe73e4a845ac986b4.jpg) + +![](images/3b590b7310f33bc26286c0879cece928aeccb755a68d91a08aef517e26d35c44.jpg) + +multi-modal image restoration [10, 27], and multi-spectral image fusion [39, 44]. Recent supervised homography estimation approaches [4, 5, 9, 20, 34, 42] have demonstrated their superiority in handling large deformation and modality gaps. However, since multi-modal images are captured by different sensors, the ground-truth homography deformations are usually unknown in real-world applications. + +To avoid using labeled data, unsupervised homography estimation approaches have emerged. Existing approaches [19, 28, 38, 40] typically optimize the similarity between the warped source image and target image through intensity-based losses. Some multi-modal image registration approaches [1, 11] transferred an image into the other modality, using mono-modal similarity metrics on image content to measure the correctness of motion. The previous state-of-the-art (SOTA) approach [41] combined cross-modal intensity-based learning with intra-modal self-supervised learning for homography knowledge injection. Notably, almost all the previous approaches fully or partially rely on the content consistency of images or feature maps to indirectly assess the correctness of homography estimation. As illustrated in Fig. 1a, due to the non-convexity of the image signal, loss value of content consistency indirect supervision has + +multiple local minima across its solution space, and hence can not guarantee accurate homography prediction, especially under large modality and deformation gaps. On the contrary, the homography direct supervision in Fig. 1b can avoid such drawback, but it is unavailable in unsupervised cross-modal homography estimation, making the problem more challenging. + +To cope with the above problem, we propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation problem into two imperfect sub-problems: Sub-problem I: A supervised mono-modal homography estimation problem, where the modalities of input images are not perfectly consistent. Sub-problem II: A supervised modality transfer problem, where the supervisory image is not perfectly aligned. We introduce a split optimization strategy to separately train a homography estimation network for Sub-problem I and a modality transfer network for Sub-problem II, enabling effective convergence of the framework. Thanks to the proposed problem reformulation and the well-designed optimization strategy, our framework effectively solves the unsupervised task by training the networks under totally direct supervision, showing significantly improved training stability. This enables the successful training of iterative homography estimation architectures like IHN [4] and RHWF [5]. Better still, for the correlation-based homography estimation architectures [4, 5, 41], we propose an extra homography feature space supervision that enhances feature consistency, further improving homography estimation accuracy. Moreover, to address the computational burden of the modality transfer network, we introduce a simple yet effective distillation training technique, which notably reduces model parameters and improves cross-domain generalization ability. + +We evaluate SSHNet on a variety of datasets including GoogleMap [42], DPDN [30], OPT-SAR [23], and Flash/no-flash [15] cross-modal datasets, together with RGB/NIR [2] cross-spectral datasets. SSHNet outperforms all the previous unsupervised homography estimation approaches with a large gap, and even achieves better performance compared to many supervised ones. The SSHNet using IHN as homography estimation network, namely SSHNet-IHN, owns $47.4\%$ and $85.8\%$ lower mean average corner errors (MACEs) than the supervised approaches MHN [20] and LocalTrans [34] on the challenging OPT-SAR dataset. Moreover, the SSHNet using the homography estimation architecture of the previous unsupervised SOTA approach SCPNet [41], namely SSHNet-SCPNet, even outperforms SCPNet itself, revealing the effectiveness of our proposed learning framework. To summarize, our main contributions are as follows: + +- We propose SSHNet, a novel unsupervised cross-modal homography estimation learning framework, which refor + +mulates the unsupervised problem into two sub-problems with direct supervision. To realize stable training, we introduce a split optimization strategy that separately trains the network in each sub-problem. + +- For correlation-based homography estimation architectures, we introduce an extra homography feature space supervision to improve feature consistency for homography estimation, which further improves estimation accuracy. +- Considering the computational burden introduced by the additional modality transfer network, we devise a simple yet effective distillation training technique to reduce model parameters while improving cross-domain generalization ability with little impact on performance. +- The training stability of SSHNet enables its cooperation with various homography estimation architectures, especially iterative ones like IHN and RHWF, resulting in higher performance. SSHNet-IHN outperforms the previous unsupervised approaches by a significant margin. + +# 2. Related Works + +Supervised Deep Homography Estimation. DeTone et al. [9] introduced DHN, a VGG-style network to predict the homography between the source and target images. Le et al. [20] introduced a multi-scale VGG-style network, named MHN. Chang et al. [7] combined CNN with an untrainable IC-LK iterator. Zhao et al. [42] enhanced the feature similarity to improve the performance of IC-LK iterator. Shao et al. [34] proposed LocalTrans that adopts a multi-scale local transformer to register cross-resolution images. Cao et al. [4] proposed IHN, a trainable iterative framework with substantial accuracy. Cao et al. [5] further introduced RHWF that integrates homography-guided image warping and focus transformer, achieving higher performance. Deng et al. [8] proposed CrossHomo, a multi-level framework for cross-modality and cross-resolution homography estimation. Li et al. [22] introduced DMHomo, a diffusion-based framework to generate high-quality training pairs for supervised homography learning. + +Unsupervised Deep Homography Estimation. Nguyen et al. [28] proposed UDHN, an unsupervised homography estimation network trained with photometric loss. Zhang et al. [40] introduced CA-UDHN, projecting images into a consistent feature space for effective estimation. Koguchiuk et al.[19] proposed biHomE, using perceptual loss[18] to improve robustness to intensity and viewpoint variations. Ye et al. [38] proposed BasesHomo, which enforces the image feature to be warp-equivalent with a feature identity loss and decomposes the homography matrix into flow bases. As modality and deformation gaps grow, intensity-based metrics for cross-modal homography matching may fail [41]. To address this, Zhang et al. [41] introduced SCPNet, which combines cross-modal intensity-based learning with intramodal self-supervised learning for homography knowledge + +![](images/5239e411bcd118d92376f41ba7195f75bc7dd1c9732febbbe9968690a12e268b.jpg) +Figure 2. Reformulating the unsupervised cross-modal homography estimation problem into two sub-problems with direct supervision. (a) Unsupervised cross-modal homography estimation problem without direct supervision. (b) Mono-modal homography estimation problem with imperfect inputs. (c) Modality transfer problem with imperfect supervision. + +injection. Besides, some unsupervised approaches, such as NeMAR [1], RFNet [37], and UMF-CMGR [11], use modality transfer networks to convert one modality into the other, enabling cross-modal motion estimation. The above unsupervised approaches perform well on images with small deformations, but may produce unsatisfactory results under large deformations and modality gaps. + +Image-to-Image Translation. The image-to-image (I2I) translation problem focuses on converting images from a source domain to a target domain. Isola et al. [17] introduced Pix2pix, which employs conditional GANs and requires paired training data. Zhu et al. [45] later proposed CycleGAN, utilizing cycle-consistent loss to enable unpaired I2I translation. Liu et al.[24] developed UNIT, and Huang et al.[16] introduced MUNIT, both leveraging a shared latent space assumption to improve unpaired translation. More recently, Sasaki et al.[33] proposed UNIT-DDPM, integrating diffusion probabilistic models into the I2I translation framework. Li et al. [21] introduced BBDM, modeling I2I translation as a stochastic Brownian Bridge process. + +# 3. Problem Reformulation + +Denote $\mathbf{I}_{\mathrm{A}}$ and $\mathbf{I}_{\mathrm{B}}$ as a pair of images from modality A and B, with an unknown homography $\mathbf{H}_{\mathrm{AB,GT}}$ . As illustrated in Fig. 2a, the unsupervised cross-modal homography estimation aims to train a network to estimate the homography $\hat{\mathbf{H}}_{\mathrm{AB}}$ , without direct supervision of $\mathbf{H}_{\mathrm{AB,GT}}$ . To solve this problem, in SSHNet, we reformulate it into two sub-problems with direct supervision, including a monomodal homography estimation problem with imperfect inputs (Fig. 2b) and a modality transfer problem with imperfect supervision (Fig. 2c). The two sub-problems are mutually causal, and the imperfect condition of each sub-problem can be gradually addressed through the optimization of another sub-problem. In this way, the unsupervised homography estimation problem can be solved in a totally supervised manner. + +![](images/84ab33bf4d0409efc768b6f8cc6ba8dbe50624d9e806a250d852944c50fc8c87.jpg) +Figure 3. Comparison of straight optimization and split optimization with regularization. (a) Cross-modal test MACEs for the network using straight optimization and split optimization with regularization, on OPT-SAR dataset. (b) Schematic diagram of straight optimization and split optimization. + +Sub-problem I. For the mono-modal homography estimation, generating image pairs with simulated ground-truth deformations for the direct self-supervised training is a common approach [4, 9, 20, 34]. If we can ideally convert $\mathbf{I}_{\mathrm{A}}$ into modality B, producing the transferred image $\mathbf{I}_{\mathrm{A,T}}$ , we can treat the unsupervised cross-modal homography estimation problem as the mono-modal one by estimating the homography between $\mathbf{I}_{\mathrm{A,T}}$ and $\mathbf{I}_{\mathrm{B}}$ , as in Fig. 2b. The homography estimation network can be effectively trained using $\mathbf{I}_{\mathrm{B}}$ with the simulated homography deformations, in a self-supervised manner. We refer to the problem of estimating the homography between $\mathbf{I}_{\mathrm{B}}$ and the transferred image $\mathbf{I}_{\mathrm{A,T}}$ with the self-supervised training using $\mathbf{I}_{\mathrm{B}}$ as Sub-problem I. + +Sub-problem II. As illustrated in Fig. 2c, we also need a modality transfer network to obtain the optimal mapping from $\mathbf{I}_{\mathrm{A}}$ to $\mathbf{I}_{\mathrm{B}}$ , producing $\mathbf{I}_{\mathrm{A,T}}$ . The ideal network training typically requires $\mathbf{I}_{\mathrm{B}}$ that is well-aligned with $\mathbf{I}_{\mathrm{A}}$ for supervision, which is also unavailable in the unsupervised cross-modal homography estimation. Previous approaches [1, 11] attempt to use GAN-based supervision to achieve modality transfer training using the unaligned image pairs. However, this strategy encounters considerable challenges when dealing with large deformation and modality gaps, as illustrated in [41]. Therefore, high-quality modality transfer requires well-aligned $\mathbf{I}_{\mathrm{A}}$ and $\mathbf{I}_{\mathrm{B}}$ . Now that we have transferred $\mathbf{I}_{\mathrm{A}}$ to $\mathbf{I}_{\mathrm{A,T}}$ , the requirement becomes that Sub-problem I being ideally solved. We refer to the modality transfer problem, which uses the homography estimated in Sub-problem I to produce a warped $\mathbf{I}_{\mathrm{B}}$ for supervision, as Sub-problem II. + +Straight Optimization Fails. However, straight optimization of both sub-problems simultaneously is challenging, as the settlement of either sub-problem depends on the assumption that the other has already been solved. We conduct a pilot experiment where the homography estimation network and the modality transfer network are trained simultaneously to address Sub-problem I and II, but the networks fail to converge, as shown by the blue curve in Fig. 3a. + +Split Optimization with Regularization Success. Inspired by decomposition methods in convex optimization, + +![](images/2794cc34fe00a2023aa416de20a220276cbe7f0cb1c67d4242dff0b4832e915b.jpg) +(a) Optimization of Sub-problem I + +Figure 4. Schematic diagram of Split Supervised Homography estimation Network, namely SSHNet. (a) Detailed optimization process of Sub-problem I. (b) Detailed optimization process of Sub-problem II. (c) Extra supervision on the homography feature space. +![](images/b1d6ee9134bb2bd41fdc3f66dd3c9cc7872edc1bd24a151ae88b60baba42c726.jpg) +Only Training Stage Both Training and Inference Stage Simulated Homography Generator + +![](images/796f0533b0826ed2f57b2e9f43b61c0f995e99e4118ecb400ca3049b3eaf2004.jpg) + +![](images/3886c86cebc914962011f7750079373d4b27f156e5de442a05daa8dc7c0648c5.jpg) +Homography Estimation Network + +![](images/2ca645c770e78fc72840a9eb9060c2595c00d196ba672f7cdccbdef885a2d161.jpg) +Modality Transfer Network + +![](images/67123521fcf15b7cc9cee4a67fab7ea846d9d955500147f66be1a2ee7198d0fd.jpg) +Trainable +Weight Sharing + +![](images/66e5b3720dabd991d6c874fe8c66f6e1255c6d06905008fa9f501cf71ef77600.jpg) +rozen + +![](images/7874e89d24a13231f03e76fd20cf551003e561d5f4d6192a8f8161abfe98ab31.jpg) +Feature Extractor in Homography Estimation Network + +![](images/7608eb7f6b724e1471d657ed09d0507b514ab4883f88c22c2ba73cffff298a26.jpg) +mage Warping + +we then consider using a split optimization strategy to address the two sub-problems separately. During optimization, each network updates its weights with the other one's frozen. As in Fig. 2b and Fig. 2c, in each iteration, the homography estimation network in Sub-problem I is updated, producing a better-aligned $\mathbf{I}_{\mathrm{B}}$ to guide the modality transfer network in Sub-problem II in the next iteration. Similarly, the modality transfer network in Sub-problem II is updated, generating a better transferred image $\mathbf{I}_{\mathrm{A,T}}$ to serve as input for the homography estimation network in Sub-problem I. This process gradually improves the outputs of each network during training, facilitating the optimization of the other network, and finally makes both sub-problems nearly perfect supervised ones. However, at the beginning of the training, the modality gap between $\mathbf{I}_{\mathrm{A,T}}$ and $\mathbf{I}_{\mathrm{B}}$ in Sub-problem I is extremely large, so is the misalignment of $\mathbf{I}_{\mathrm{A}}$ and $\mathbf{I}_{\mathrm{B}}$ in Sub-problem II, which hinders the mutual promotion between two subproblems. A regularization term that guides the training into our desired direction is needed. Inspired by [41] $^{1}$ , in Subproblem I, we replace the self-supervised training using only $\mathbf{I}_{\mathrm{B}}$ by the two-branch self-supervised training using both $\mathbf{I}_{\mathrm{A,T}}$ and $\mathbf{I}_{\mathrm{B}}$ (see Fig. 2b). This enables the homography estimation network to implicitly learn cross-modal homography knowledge at the beginning of the training, and triggers the two split optimized sub-problems to converge into our desired direction, as shown by the orange curve in Fig. 3a. + +# 4. Method + +Based on the problem reformulation in Section 3, we further investigate its potential and propose an unsupervised cross-modal homography estimation learning framework, named + +Split Supervised Homography estimation Network (SSHNet). Section 4.1 covers the split optimization of SSHNet (Fig. 4a and Fig. 4b), Section 4.2 discusses the cooperation of SSHNet with various homography estimation architectures, Section 4.3 outlines the design of the modality transfer network, Section 4.4 presents the extra homography feature space supervision (Fig. 4c), and Section 4.5 illustrates the distillation training technique (Fig. 5). + +# 4.1. Split Optimization + +As illustrated in Fig. 4a and Fig. 4b, the homography estimation network and the modality transfer network are optimized separately in each iteration, forming the split optimization framework of our SSHNet. In the following, we will elaborate the optimization procedure for each network. + +Optimization of Sub-problem I. As mentioned in Section 3, we adopt a two-branch self-supervised learning strategy to optimize the homography estimation network in Subproblem I. We apply the separate simulated homography deformation on $\mathbf{I}_{\mathrm{A}}$ and $\mathbf{I}_{\mathrm{B}}$ , generating the intra-modal image pairs $(\mathbf{I}_{\mathrm{A}},\mathbf{I}_{\mathrm{A}}^{\prime})$ and $(\mathbf{I}_{\mathrm{B}},\mathbf{I}_{\mathrm{B}}^{\prime})$ with corresponding ground-truth homography $\mathbf{H}_{\mathrm{A,GT}}$ and $\mathbf{H}_{\mathrm{B,GT}}$ . The modality transfer network transfers $(\mathbf{I}_{\mathrm{A}},\mathbf{I}_{\mathrm{A}}^{\prime})$ into $(\mathbf{I}_{\mathrm{A,T}},\mathbf{I}_{\mathrm{A,T}}^{\prime})$ . We simultaneously use $(\mathbf{I}_{\mathrm{B}},\mathbf{I}_{\mathrm{B}}^{\prime})$ and $(\mathbf{I}_{\mathrm{A,T}},\mathbf{I}_{\mathrm{A,T}}^{\prime})$ to train the weight-shared homography estimation network in a self-supervised manner. The optimization of Sub-problem I can be formulated as: + +$$ +\underset {\zeta} {\arg \min } L _ {\mathrm {H}} \left(\mathcal {H} _ {\zeta} \left(\mathbf {I} _ {\mathrm {B}}, \mathbf {I} _ {\mathrm {B}} ^ {\prime}\right), \mathbf {H} _ {\mathrm {B}, \mathrm {G T}}\right) + \mathcal {R} \left(\mathbf {I} _ {\mathrm {A}, \mathrm {T}}, \mathbf {I} _ {\mathrm {A}, \mathrm {T}} ^ {\prime}\right), \tag {1} +$$ + +where $\mathcal{H}_{\zeta}$ denotes the homography estimation network with parameters $\zeta$ to be optimized, and $L_{\mathrm{H}}$ denotes the homography estimation loss. $\mathcal{R}(\mathbf{I}_{\mathrm{A,T}},\mathbf{I}_{\mathrm{A,T}}^{\prime}) = L_{\mathrm{H}}\bigl (\mathcal{H}_{\zeta}(\mathcal{T}_{\theta^{*}}(\mathbf{I}_{\mathrm{A}}),\mathcal{T}_{\theta^{*}}(\mathbf{I}_{\mathrm{A}}^{\prime})),\mathbf{H}_{\mathrm{A,GT}}\bigr)$ , which is the regular + +ization term, where $\mathcal{T}_{\theta^*}$ denotes the modality transfer network with parameters $\theta^{*}$ frozen. + +Optimization of Sub-problem II. We train a modality transfer network to convert $\mathbf{I}_{\mathrm{A}}$ into an image of pseudo modality B, namely $\mathbf{I}_{\mathrm{A,T}}$ . Following the reformulation in Section 3, we apply the homography matrix $\hat{\mathbf{H}}_{\mathrm{AB}}$ predicted by $\mathcal{H}_{\zeta}$ to warp $\mathbf{I}_{\mathrm{B}}$ , providing direct supervision for the modality transfer network. The optimization of Sub-problem II can be formulated as + +$$ +\underset {\theta} {\arg \min } L _ {\mathrm {T}} \left(\mathcal {T} _ {\theta} (\mathbf {I} _ {\mathrm {A}}), \mathcal {W} (\mathbf {I} _ {\mathrm {B}}, \mathcal {H} _ {\zeta^ {*}} (\mathbf {I} _ {\mathrm {A}, \mathrm {T}}, \mathbf {I} _ {\mathrm {B}}))\right), \tag {2} +$$ + +where $\mathcal{T}_{\theta}$ denotes the modality transfer network with parameters $\theta$ to be optimized, $\mathcal{H}_{\zeta^{*}}$ denotes the homography estimation network with parameters $\zeta^{*}$ frozen, $\mathcal{W}$ denotes the image warping with the estimated homography, and $L_{\mathrm{T}}$ denotes the modality transfer loss. + +# 4.2. Homography Estimation Network + +Present homography estimation approaches exhibit a variety of architectural designs. Some approaches [9, 20, 28, 38, 40] concatenate input images in the channel dimension, directly estimating the homography deformation. Others [4, 5, 34, 41] use backbone networks to extract features and employ the feature correlation to perform homography decoding. Additionally, recent approaches [4, 5] further introduce iterative architectures to refine the estimation results, achieving substantial improvement in accuracy. + +As illustrated in the experimental results of Section 5.3, our proposed framework is independent of the specific homography estimation architecture, enabling integration with various architectures. The pure supervised training strategy significantly boosts the stability of the optimization process. As a result, iterative homography estimation architectures like IHN [4] and RHWF [5] can be effectively employed, leading to a higher performance upper bound. In contrast, due to the lack of an explicit solution for consistent features between cross-modal images, the previous SOTA approach, SCPNet [41], struggles to converge when using these iterative architectures. For the self-supervised training, we adopt $L_{1}$ homography estimation loss following previous works [4, 5, 9, 20, 41], which is formulated as + +$$ +L _ {\mathrm {H}} (\hat {\mathbf {H}}, \mathbf {H} _ {\mathrm {G T}}) = \left\| \hat {\mathbf {H}} - \mathbf {H} _ {\mathrm {G T}} \right\| _ {1}, \tag {3} +$$ + +where $\tilde{\mathbf{H}}$ denotes the estimated homography, $\mathbf{H}_{\mathrm{GT}}$ denotes the ground-truth homography. For iterative homography architectures, we use a weighted sum of the $L_{1}$ distance following [4, 5]. + +# 4.3. Modality Transfer Network + +We design the modality transfer network in our SSHNet as a U-shaped architecture to enhance the perception of multiscale information. Inspired by Swin-Unet [3], the Swin + +Transformer block [25] serves as the basic unit of the modality transfer network. The attention mechanism enhances the network's ability to capture hierarchical structures and context information. This facilitates more effective learning of complex mappings between modalities. + +As mentioned in Section 4.1, we warp image $\mathbf{I}_{\mathrm{B}}$ with the $\hat{\mathbf{H}}_{\mathrm{AB}}$ predicted by the homography estimation network for the modality transfer supervision. During optimization, $\hat{\mathbf{H}}_{\mathrm{AB}}$ cannot ideally align the input image $\mathbf{I}_{\mathrm{A}}$ with the supervisory image $\mathbf{I}_{\mathrm{B}}$ . Therefore, instead of using pixel-wise $L_{1}$ or $L_{2}$ loss, we apply perceptual loss $L_{\mathrm{pcp}}$ [18] to capture high-level representations, ensuring better tolerance to misalignments, which is formulated as + +$$ +L _ {\mathrm {p c p}} \left(\mathbf {I} _ {\mathrm {A}, \mathrm {T}}, \mathbf {I} _ {\mathrm {B}, \mathrm {W}}\right) = \sum_ {j} \frac {1}{C _ {j} H _ {j} W _ {j}} \left\| \psi_ {j} \left(\mathbf {I} _ {\mathrm {A}, \mathrm {T}}\right) - \psi_ {j} \left(\mathbf {I} _ {\mathrm {B}, \mathrm {W}}\right) \right\| _ {2} ^ {2}, \tag {4} +$$ + +where $\psi_{j}$ is the $j$ -th layer of the pre-trained VGG with a feature map of shape $C_j\times H_j\times W_j$ , $\mathbf{I}_{\mathrm{B,W}}$ denotes the image $\mathbf{I}_{\mathrm{B}}$ warped by the estimated $\hat{\mathbf{H}}_{\mathrm{AB}}$ . + +# 4.4. Extra Homography Feature Space Supervision + +The current modality transfer network in Sub-problem II is supervised by the perceptual loss [18], which computes feature similarity using a pre-trained VGG network. This may yield suboptimal results for homography estimation in Sub-problem I when trained under the present totally split optimization framework. + +To address this issue, as illustrated in Fig. 4c, we introduce an extra supervision in the homography feature space when the architecture of homography estimation network has explicit feature extractors, such as IHN, RHWF, and SCPNet. A homography feature (HF) loss that directly enhances the feature consistency for homography estimation between $\mathbf{I}_{\mathrm{A,T}}$ and $\mathbf{I}_{\mathrm{B}}$ is computed. The modality transfer network and the feature extractor of the homography estimation network are optimized together under the supervision of HF loss, which can be expressed as + +$$ +\underset {\theta , \xi} {\arg \min } L _ {\mathrm {H F}} \left(\mathcal {F} _ {\xi} \left(\mathcal {T} _ {\theta} (\mathbf {I} _ {\mathrm {A}})\right), \mathcal {F} _ {\xi} \left(\mathcal {W} \left(\mathbf {I} _ {\mathrm {B}}, \hat {\mathbf {H}} _ {\mathrm {A B}}\right)\right)\right), \tag {5} +$$ + +where $\mathcal{F}_{\xi}$ denotes the feature extractor with parameters $\xi$ to be optimized. We note that IHN, RHWF, and SCPNet all depend on feature correlation to perform homography decoding. Considering this, We formulate the HF loss using a correlation-based loss function $L_{\mathrm{corr}}$ to enhance feature consistency, which is formulated as + +$$ +L _ {\operatorname {c o r r}} \left(\mathbf {F} _ {\mathrm {A}, \mathrm {T}}, \mathbf {F} _ {\mathrm {B}, \mathrm {W}}\right) = - \sum_ {i, j} \mathbf {F} _ {\mathrm {A}, \mathrm {T}} (i, j) ^ {\top} \mathbf {F} _ {\mathrm {B}, \mathrm {W}} (i, j), \tag {6} +$$ + +where $\mathbf{F}_{\mathrm{A,T}}$ and $\mathbf{F}_{\mathrm{B,W}}$ denote the feature maps of $\mathbf{I}_{\mathrm{A,T}}$ and $\mathbf{I}_{\mathrm{B,W}}$ in the homography estimation network. + +![](images/b9093a72265be09818f5a97f6e14019bb401ff28c52d57dc9671abfa6862b6b1.jpg) +Figure 5. Schematic diagram of the distillation training. + +# 4.5. Distillation Training + +We can see that the present framework has extra network parameters due to the utilization of the modality transfer network in Sub-problem II. Moreover, as the modality transfer network is trained on multi-modal image pairs from a fixed domain, its generalization ability to images from other domains may be limited. + +To cope with the above problem, we introduce a distillation training technique under the inspiration of the unsupervised knowledge distillation frameworks in domain adaptive object detection [32, 43]. We use SSHNet as a teacher network to guide a student network that only retains the homography estimation network, named Distilled SSHNet (SSHNet-D). During the distillation training, the homography $\hat{\mathbf{H}}_{\mathrm{teacher}}$ predicted by SSHNet serves as a pseudo ground-truth to supervise SSHNet-D, as illustrated in Fig. 5. The distillation training can be formulated as + +$$ +\underset {\omega} {\arg \min } L _ {\mathrm {H}} \left(\phi_ {\omega} \left(\mathbf {I} _ {\mathrm {A}}, \mathbf {I} _ {\mathrm {B}}\right), \hat {\mathbf {H}} _ {\text {t e a c h e r}}\right), \tag {7} +$$ + +where $\phi_{\omega}$ denotes the SSHNet-D with the parameters $\omega$ to be optimized. + +# 5. Experiments + +# 5.1. Experimental Settings + +Datasets. We evaluate SSHNet on various datasets, including GoogleMap [42], DPDN [30], OPT-SAR [23], Flash/no-flash [15] cross-modal datasets, and RGB/NIR [2] cross-spectral dataset. Examples for each dataset are shown in Fig. 6. To ensure a fair comparison, all evaluated methods are trained and tested using the same training and test splits within each dataset. + +Baselines. We evaluate SSHNet with traditional feature-based approaches including SIFT [26], ORB [31], unsupervised approaches including UDHN [28], biHomE [19], CA-UDHN [40], BasesHomo [38], UMF-CMGR [11], SCP-Net [41], and supervised approaches including DHN [9], MHN [20], LocalTrans [34], IHN [4], RHWF [5]. For feature-based approaches, we choose RANSAC [13] as the outlier rejection algorithms. UMF-CMGR is an image fusion approach based on registration, and we use the registration part for comparison. + +Homography Estimation Architectures. To investigate the potential of our proposed SSHNet, we adopt various homography estimation architectures including DHN [9] and MHN [20], together with iterative architectures such as IHN [4] and RHWF [5]. Moreover, we also adopt the + +Table 1. Ablation study on the training strategy of SSHNet. + +
ExperimentMACE ↓
w/o problem reformulationNC
w/o split optimizationNC
pure self-supervised learning24.60
SSHNet2.94
+ +Table 2. Ablation study on the architectures of the modality transfer network. + +
ExperimentParameters (M)MACE ↓
CNN-based7.488.72
Transformer-based7.542.94
+ +Table 3. Ablation study on the loss functions. $L_{\mathrm{T}}$ denotes the modality transfer loss, $L_{\mathrm{HF}}$ denotes the homography feature loss. + +
ExperimentLoss TypeMACE ↓
LTL15.88
L26.23
Lpcp4.52
LT + LHFLpcp + L14.09
Lpcp + L23.27
Lpcp + Lcorr2.94
+ +Table 4. Ablation study of the modality transfer direction. + +
Method\DirectionRGB→NIRNIR→RGBOPT→SARSAR→OPT
SSHNet1.661.752.944.28
+ +architecture from the previous SOTA unsupervised approach SCPNet [41], to specifically evaluate the effectiveness of our proposed unsupervised learning framework. + +Metrics. Similar to previous works [4, 5, 9, 34, 42], we randomly perturb the corner points of $128 \times 128$ images to make the deformed image, with a perturbation range set to [-32, 32]. We evaluate homography estimation accuracy using the mean average corner error (MACE) following [4, 5, 9, 34, 42]. A lower MACE indicates a higher accuracy. + +# 5.2. Ablation + +The ablation studies are conducted on the OPT-SAR dataset. We adopt IHN as the homography estimation network in our SSHNet for ablation studies. + +Training Strategy. We conduct an ablation study on the training strategy of our SSHNet. The experimental results in Table 1 demonstrate that both problem reformulation and the split optimization strategy are crucial for SSHNet. Without problem reformulation, unsupervised training with intensity-based supervision fails to converge. Even with problem reformulation, directly optimizing both networks simultaneously does not yield reliable homography estimation. SSHNet achieves effective convergence only when problem reformulation is combined with the split optimization strategy, where each network is trained separately. We also compare our framework with method trained through + +![](images/4d3bcc836cb9d77f179be2784cda938a81f454df88aa0a4f00f0c84d4f4bc1d4.jpg) +(a) GoogleMap + +![](images/ad883459225c75b0a40e9de50d67063a10a43a88424cc96af2e78ad321243a60.jpg) + +![](images/95f4f8c36a1e5c14a5b293a43c33a37d8318c8d313f3a7e52f6abb0593e5de83.jpg) +(b) DPDN + +![](images/467ea23b4e37cc7c7271bba3d061132216adddc75a548bfac04157f0429e207c.jpg) +Figure 6. Examples for each dataset. + +![](images/e6808aeee93d068056181739cb07aab218f753be6f5a9001e581712f5f2a7367.jpg) +(c) OPT-SAR + +![](images/b9915e447e38c24b8aa4d4d849ed6b021f8d96fe98c70016290e4f830c8f60bf.jpg) + +![](images/f24c14a78ab48ae9c4c78db7ecef954b184eb574b475bd146c064a0ef6b96c44.jpg) +(d) Flash/no-flash + +![](images/3374e92360f8b22776246c78e9d0b6664d62235f9578e2b76b8b08f9a3351a3a.jpg) + +![](images/e678698ac822c50d4530bf8c899f621e5cde91d7647b0945253d656989df89b2.jpg) +(e) RGB/NIR + +![](images/4a21519ed0aa91bc732f2e0323b2efb0e8aa9003765687d2b4ec487d12fe87fc.jpg) + +Table 5. Quantitative comparison of our SSHNet and other approaches. NC denotes the training is not converged. The best results of unsupervised methods are highlighted in bold. + +
Dataset ApproachGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
TraditionalSIFT [26]24.5325.1424.8218.2921.33
ORB [31]24.5224.8624.9020.2222.67
SupervisedDHN [9]5.204.928.276.4211.88
MHN [20]2.692.955.595.245.52
LocalTrans [34]3.221.7920.681.521.72
IHN [4]0.921.171.670.801.63
RHWF [5]0.711.061.480.651.07
UnsupervisedUDHN [28]22.84NCNC21.2023.43
CA-UDHN [40]24.6124.9924.7621.3224.12
biHomE [19]NCNCNC11.8623.77
BasesHomo [38]24.4927.3326.7225.1224.41
UMF-CMGR [11]24.6025.4624.7023.4922.38
SCPNet [41]4.35NCNC2.674.78
SSHNet-DHN (Ours)9.289.7117.6310.5112.13
SSHNet-MHN (Ours)2.903.046.775.626.93
SSHNet-SCPNet (Ours)3.894.8112.942.653.86
SSHNet-RHWF (Ours)1.291.263.080.951.52
SSHNet-IHN (Ours)1.231.122.941.081.66
SSHNet-IHN-D (Ours)1.261.243.311.161.72
+ +pure self-supervised learning. Results in Table 1 show that self-supervised learning alone does not yield satisfactory homography estimation results. + +Architectures of the Modality Transfer Network. As illustrated in Table 2. Our Transformer-based architecture outperforms the CNN-based architecture by a large margin, with similar parameter counts. + +Loss Functions. To further investigate the effectiveness of the modality transfer and homography feature loss functions, we first train the network without HF loss to evaluate the performance of different types of modality transfer loss. The results in the upper part of Table 3 indicate that the perceptual loss $L_{\mathrm{pcp}}$ outperforms $L_{1}$ and $L_{2}$ . Then we adopt $L_{\mathrm{pcp}}$ as the modality transfer loss and add different types of HF loss to analyze their impact. We report the results in the lower part of Table 3. Though all different types of HF loss further boost the network's performance, the correlation-based loss $L_{\mathrm{corr}}$ achieves superior performance compared to $L_{1}$ and $L_{2}$ . + +Modality transfer direction. Table 4 presents the results of different transfer directions on the RGB/NIR and OPT-SAR datasets. Since some modalities carry more information, transferring from richer to less informative modalities facilitates the modality transfer network's learning process, thereby promoting the framework's convergence. + +# 5.3. Comparisons with Existing Methods + +We compare the homography estimation results of our proposed SSHNet and baselines and list the experimental results in Table 5. It is observed that the prior traditional and unsupervised approaches can hardly handle the homography estimation under large modality and deformation gaps. In contrast, our method achieves much more satisfactory results with various homography estimation architectures. When adopting the iterative architectures IHN and RHWF, our method outperforms all the compared traditional and unsupervised approaches by a significant margin. Moreover, even though trained without direct cross-modal supervision, SSHNet-IHN and SSHNet-RHWF outperforms supervised approaches such as DHN, MHN and LocalTrans. Specifically, SSHNet-IHN owns $54.3\%$ and $47.4\%$ lower MACEs than MHN, and $61.8\%$ and $85.8\%$ lower MACEs than LocalTrans on the challenging GoogleMap and OPT-SAR datasets, respectively. Notably, SSHNet-SCPNet, which uses the homography estimation architecture from the previous unsupervised SOTA approach SCPNet, outperforms SCPNet itself on all datasets, further proving the superiority of our proposed training framework. + +We also conduct the distillation training on SSHNet with different homography estimation architectures. We present the distillation training results of SSHNet-IHN in Table 5, + +Table 6. Cross-dataset evaluation results. Red indicates the improved performance of SSHNet-IHN-D compared to SSHNet-IHN, and blue indicates the declined one. + +
Test TrainGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
GoogleMap/↑1.8%↑1.0%↑80.4%↑63.0%
DPDN↑66.5%/↑0.4%↑92.1%↑75.8%
OPT-SAR↑48.4%↑1.3%/↑77.5%↑66.2%
Flash/no-flash↓5.2%↓0.2%↑0.1%/↓4.6%
RGB/NIR↓0.5%↑1.2%↑3.0%↑2.1%/
+ +namely SSHNet-IHN-D. We observe that the performance of SSHNet-IHN-D is comparable to that of SSHNet-IHN. For more detailed qualitative visualization and experimental results, please refer to supplementary material. + +# 5.4. Cross-Dataset Generalization + +We further conduct a cross-dataset evaluation of SSHNet-IHN-D and SSHNet-IHN on the aforementioned datasets, to further reveal the effectiveness of our proposed distillation training. As illustrated in Fig. 6, multi-modal image pairs from the GoogleMap, DPDN, and OPT-SAR datasets exhibit substantial modality gaps, and we call these three datasets challenging datasets, for a better illustration. In contrast, the Flash/no-flash and RGB/NIR datasets show smaller modality gaps, so we call them easy datasets. For each dataset, we compute the percentage of homography predictions that have average corner errors (ACE) lower than 5 pixels on the test set, which reveals the ratio of reliable estimations. We directly list the differences of the percentage of SSHNet-IHN-D and SSHNet-IHN in Table 6. + +SSHNet-IHN uses a modality transfer network to transfer images from one modality to the other on the training dataset, making it highly prone to overfit the training data. In contrast, SSHNet-IHN-D excludes the modality transfer network and directly extracts consistent features from cross-modal images, and hence is forced to learn the features that is robust across different domains. As illustrated in Table 6, when trained on challenging datasets, SSHNet-IHN-D exhibits significant superior generalization ability. Specifically, SSHNet-IHN-D trained on DPDN yields $66.5\%$ , $92.1\%$ , and $75.8\%$ improvement on GoogleMap, Flash/no-flash and RGB/NIR datasets compared to SSHNet. We note that the generalization ability across the challenging datasets remains limited due to the substantial domain gaps between the datasets. As for training on easy datasets, the generalization performance improvement is plain. This is because the training data is too easy to provide enough knowledge for obtaining the features robust across different domains. + +# 5.5. Real World Scenarios + +We further evaluate our method on real-world scenarios [35] that do not satisfy perfect homography alignment, including RGB/NIR and Flash/no-flash image pairs. As illustrated + +![](images/325347c9fa2695dc520ab0fd4fd97ccc0d3b628fd61bed4f182edcdd2522d0b1.jpg) +(a) Cross modal +Figure 7. Visualization of results on real-world deformations. + +![](images/110122b662fed84b63dd790f30ef911e0b1ecbebf8bd616c957a6dfe2091461f.jpg) +(b) UDIS-TIR + +Table 7. Experimental results on the UDIS-TIR dataset. $I_{3 \times 3}$ denotes the identity matrix. + +
MethodPSNR ↑SSIM ↑
I3×312.920.306
SCPNetNCNC
SSHNet-IHN20.050.728
+ +in Fig. 7a, the model trained with simulated homographies achieves effective performance. We also use a pretrained RGB-to-TIR translation network to convert RGB-RGB image pairs from the UDIS [29] dataset into RGB-TIR image pairs, creating the UDIS-TIR dataset with parallax image pairs. Following UDIS, we evaluate SSHNet-IHN on the UDIS-TIR dataset and compute the PSNR and SSIM on RGB images before modality transfer. We present the quantitative results in Table 7 and show more visualization results in Fig. 7b. + +# 6. Conclusions + +In this paper, we propose SSHNet, a novel unsupervised cross-modal homography estimation framework. SSHNet reformulates the problem into two supervised sub-problems, addressed by a homography estimation network and a modality transfer network. These networks are trained separately using a split optimization strategy, ensuring stable training. We also introduce an extra homography feature space supervision to further improve feature consistency during the homography estimation and boost the performance of our framework. Moreover, we employ a distillation training technique to reduce model parameters and improve cross-domain generalization ability. Experimental results show that our SSHNet achieves SOTA performance on multiple datasets among unsupervised approaches, and outperforms many supervised approaches. + +Acknowledgments. This work was supported in part by the Zhejiang Provincial Natural Science Foundation of China under grant LD24F020003, in part by the National Natural Science Foundation of China under grant 62301484, in part by the Ningbo Natural Science Foundation of China under grant 2024J454, and in part by the National Key Research and Development Program of China under grant 2023YFB3209800. We also thank the generous help from Jianxin Hu, Zhejiang University. + +# References + +[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13410–13419, 2020. 1, 3 +[2] Matthew Brown and Sabine Süssstrunk. Multispectral SIFT for scene category recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 177-184. IEEE, 2011. 2, 6 +[3] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 205-218. Springer, 2022. 5 +[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1879-1888, 2022. 1, 2, 3, 5, 6, 7 +[5] Si-Yuan Cao, Runmin Zhang, Lun Luo, Beinan Yu, Zehua Sheng, Junwei Li, and Hui-Liang Shen. Recurrent homography estimation using homography-guided image warping and focus transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9833-9842, 2023. 1, 2, 5, 6, 7 +[6] Rich Caruana. Multitask learning. Machine Learning, 28: 41-75, 1997. 4 +[7] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. CLKN: Cascaded Lucas-Kanade networks for image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2213-2221, 2017. 2 +[8] Xin Deng, Enpeng Liu, Chao Gao, Shengxi Li, Shuhang Gu, and Mai Xu. CrossHomo: Cross-modality and cross-resolution homography estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7 +[10] Fayaz Ali Dharejo, Muhammad Zawish, Farah Deeba, Yuanchun Zhou, Kapal Dev, Sunder Ali Khowaja, and Nawab Muhammad Faseeh Qureshi. Multimodal-boost: Multimodal medical image super-resolution using multi-attention network with wavelet transform. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2022. 1 +[11] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, 2022. 1, 3, 6, 7 +[12] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2051-2060, 2017. 4 +[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to + +image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 6 +[14] Hunter Goforth and Simon Lucey. GPS-denied UAV localization using pre-existing satellite imagery. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 2974-2980. IEEE, 2019. 1 +[15] Shengfeng He and Rynson WH Lau. Saliency detection with flash and no-flash image pairs. In Proceedings of the European Conference on Computer Vision, pages 110-124. Springer, 2014. 2, 6 +[16] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision, pages 172-189, 2018. 3 +[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1125-1134, 2017. 3 +[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision, pages 694–711. Springer, 2016. 2, 5 +[19] Daniel Koguciuk, Elahe Arani, and Bahram Zonooz. Perceptual loss for robust unsupervised homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4274-4283, 2021. 1, 2, 6, 7 +[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020. 1, 2, 3, 5, 6, 7 +[21] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. BBDM: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 3 +[22] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. DMHomo: Learning homography with diffusion models. ACM Transactions on Graphics, 43 (3):1-16, 2024. 2 +[23] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. MCANet: A joint semantic segmentation framework of optical and SAR images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106:102638, 2022. 2, 6 +[24] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in Neural Information Processing Systems, 30, 2017. 3 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5 +[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 6, 7 + +[27] Iman Marivani, Evaggelia Tsiligianni, Bruno Cornelis, and Nikos Deligiannis. Designing CNNs for multimodal image restoration and fusion via unfolding the method of multipliers. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5830-5845, 2022. 1 +[28] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 1, 2, 5, 6, 7 +[29] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Unsupervised deep image stitching: Reconstructing stitched features to images. IEEE Transactions on Image Processing, 30:6184-6197, 2021. 8 +[30] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In *British Machine Vision Conference*. The British Machine Vision Association, 2016. 2, 6 +[31] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 6, 7 +[32] Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain adaptation. In Proceedings of the International Conference on Machine Learning, pages 2988-2997. PMLR, 2017. 6 +[33] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 3 +[34] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. LocalTrans: A multiscale local transformer network for cross-resolution homography estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14890-14899, 2021. 1, 2, 3, 5, 6, 7 +[35] Xiaoyong Shen, Li Xu, Qi Zhang, and Jiaya Jia. Multimodal and multi-spectral registration for natural images. In Proceedings of the European Conference on Computer Vision, pages 309-324. Springer, 2014. 8 +[36] Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator. Advances in Neural Information Processing Systems, 36, 2023. 1 +[37] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. RFNet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19679-19688, 2022. 3 +[38] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13117-13125, 2021. 1, 2, 5, 6, 7 +[39] Jiacheng Ying, Hui-Liang Shen, and Si-Yuan Cao. Unaligned hyperspectral image fusion via registration and interpolation modeling. IEEE Transactions on Geoscience and Remote Sensing, 2021. 1 + +[40] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proceedings of the European Conference on Computer Vision, pages 653-669. Springer, 2020. 1, 2, 5, 6, 7 +[41] Runmin Zhang, Jun Ma, Si-Yuan Cao, Lun Luo, Beinan Yu, Shu-Jie Chen, Junwei Li, and Hui-Liang Shen. SCP-Net: Unsupervised cross-modal homography estimation via intra-modal self-supervised learning. In Proceedings of the European Conference on Computer Vision, pages 460–477. Springer, 2024. 1, 2, 3, 4, 5, 6, 7 +[42] Yiming Zhao, Xinming Huang, and Ziming Zhang. Deep lucas-kanade homography for multimodal image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15950–15959, 2021. 1, 2, 6 +[43] Huayi Zhou, Fei Jiang, and Hongtao Lu. SSDA-YOLO: Semi-supervised domain adaptive yolo for cross-domain object detection. Computer Vision and Image Understanding, 229: 103649, 2023. 6 +[44] Yuan Zhou, Anand Rangarajan, and Paul D Gader. An integrated approach to registration and fusion of hyperspectral and multispectral images. IEEE Transactions on Geoscience and Remote Sensing, 58(5):3020-3033, 2019. 1 +[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2223-2232, 2017. 3 \ No newline at end of file diff --git a/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/images.zip b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..71d9a737dda7d558b18b8257a6169c114b0fac37 --- /dev/null +++ b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ab20b52731be89b310ff207a8fcdef714fe922b44bf5e5000234d4fb46c2575 +size 474034 diff --git a/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/layout.json b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25c3815b1d5104d6c5b5952c90ca64a7c8cf62bb --- /dev/null +++ b/2025/SSHNet_ Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization/layout.json @@ -0,0 +1,10330 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 140 + ], + "type": "text", + "content": "SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": "Junchen Yu" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Si-Yuan Cao" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Runmin Zhang" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Chenghao Zhang" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Zhu Yu" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Shujie Chen" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Bailin Yang" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "text", + "content": " Hui-Liang Shen" + }, + { + "bbox": [ + 126, + 160, + 483, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 191, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 191, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 64, + 191, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 191, + 545, + 203 + ], + "type": "text", + "content": "Ningbo Innovation Center, Zhejiang University " + }, + { + "bbox": [ + 64, + 191, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 191, + 545, + 203 + ], + "type": "text", + "content": "College of Information Science and Electronic Engineering, Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 204, + 527, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 204, + 527, + 217 + ], + "spans": [ + { + "bbox": [ + 82, + 204, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 204, + 527, + 217 + ], + "type": "text", + "content": " NingboTech University " + }, + { + "bbox": [ + 82, + 204, + 527, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 82, + 204, + 527, + 217 + ], + "type": "text", + "content": " Zhejiang Key Laboratory of Big Data and Future ECommerce Technology, Hangzhou, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 220, + 465, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 220, + 465, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 220, + 465, + 232 + ], + "type": "text", + "content": "{yujunchen,cao_siyuan,runmin_zhang,zch00,yu_zhu}@zju.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 172, + 234, + 436, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 234, + 436, + 245 + ], + "spans": [ + { + "bbox": [ + 172, + 234, + 436, + 245 + ], + "type": "text", + "content": "{chenshujie,ybl}@zjgsu.edu.cn shenhl@zju.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "type": "text", + "content": "We propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation into two supervised sub-problems, each addressed by its specialized network: a homography estimation network and a modality transfer network. To realize stable training, we introduce an effective split optimization strategy to train each network separately within its respective sub-problem. We also formulate an extra homography feature space supervision to enhance feature consistency, further boosting the estimation accuracy. Moreover, we employ a simple yet effective distillation training technique to reduce model parameters and improve cross-domain generalization ability while maintaining comparable performance. The training stability of SSHNet enables its cooperation with various homography estimation architectures. Experiments reveal that the SSHNet using IHN as homography estimation network, namely SSHNet-IHN, outperforms previous unsupervised approaches by a significant margin. Even compared to supervised approaches MHN and LocalTrans, SSHNet-IHN achieves " + }, + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "type": "inline_equation", + "content": "47.4\\%" + }, + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "type": "inline_equation", + "content": "85.8\\%" + }, + { + "bbox": [ + 54, + 299, + 297, + 598 + ], + "type": "text", + "content": " mean average corner errors (MACES) reduction on the challenging OPT-SAR dataset. Source code is available at https://github.com/Junchen-Yu/SSHNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 644, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 297, + 693 + ], + "type": "text", + "content": "Cross-modal homography estimation aims to solve the global perspective transformation between two images under different modalities. It is widely used in various computer vision tasks, such as GPS-denied robotic localization [14, 36]," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 276, + 360, + 319 + ], + "blocks": [ + { + "bbox": [ + 318, + 276, + 360, + 319 + ], + "lines": [ + { + "bbox": [ + 318, + 276, + 360, + 319 + ], + "spans": [ + { + "bbox": [ + 318, + 276, + 360, + 319 + ], + "type": "image", + "image_path": "7aa7f86c4f28fa35b10864c592c3e254349f52d3061ab072f5ca0a193486ea40.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 321, + 360, + 362 + ], + "blocks": [ + { + "bbox": [ + 318, + 321, + 360, + 362 + ], + "lines": [ + { + "bbox": [ + 318, + 321, + 360, + 362 + ], + "spans": [ + { + "bbox": [ + 318, + 321, + 360, + 362 + ], + "type": "image", + "image_path": "57824a5c34de5900a60e6036ee959f4730df1e9bbbf11fd99f54be6d4c6c98cc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 423 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 423 + ], + "type": "text", + "content": "Figure 1. Comparison of the solution space between (a) Content consistency indirect supervision and (b) Homography direct supervision. The content consistency indirect supervision may mislead the training due to local minima, while homography direct supervision won't." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 364, + 274, + 454, + 365 + ], + "blocks": [ + { + "bbox": [ + 364, + 274, + 454, + 365 + ], + "lines": [ + { + "bbox": [ + 364, + 274, + 454, + 365 + ], + "spans": [ + { + "bbox": [ + 364, + 274, + 454, + 365 + ], + "type": "image", + "image_path": "9663e2d200187d1eb32f7f797be94e22716530d5f1e616cbe73e4a845ac986b4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 462, + 274, + 552, + 366 + ], + "blocks": [ + { + "bbox": [ + 462, + 274, + 552, + 366 + ], + "lines": [ + { + "bbox": [ + 462, + 274, + 552, + 366 + ], + "spans": [ + { + "bbox": [ + 462, + 274, + 552, + 366 + ], + "type": "image", + "image_path": "3b590b7310f33bc26286c0879cece928aeccb755a68d91a08aef517e26d35c44.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 519 + ], + "type": "text", + "content": "multi-modal image restoration [10, 27], and multi-spectral image fusion [39, 44]. Recent supervised homography estimation approaches [4, 5, 9, 20, 34, 42] have demonstrated their superiority in handling large deformation and modality gaps. However, since multi-modal images are captured by different sensors, the ground-truth homography deformations are usually unknown in real-world applications." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "type": "text", + "content": "To avoid using labeled data, unsupervised homography estimation approaches have emerged. Existing approaches [19, 28, 38, 40] typically optimize the similarity between the warped source image and target image through intensity-based losses. Some multi-modal image registration approaches [1, 11] transferred an image into the other modality, using mono-modal similarity metrics on image content to measure the correctness of motion. The previous state-of-the-art (SOTA) approach [41] combined cross-modal intensity-based learning with intra-modal self-supervised learning for homography knowledge injection. Notably, almost all the previous approaches fully or partially rely on the content consistency of images or feature maps to indirectly assess the correctness of homography estimation. As illustrated in Fig. 1a, due to the non-convexity of the image signal, loss value of content consistency indirect supervision has" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16685" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "multiple local minima across its solution space, and hence can not guarantee accurate homography prediction, especially under large modality and deformation gaps. On the contrary, the homography direct supervision in Fig. 1b can avoid such drawback, but it is unavailable in unsupervised cross-modal homography estimation, making the problem more challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 298, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 298, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 298, + 491 + ], + "type": "text", + "content": "To cope with the above problem, we propose a novel unsupervised cross-modal homography estimation learning framework, named Split Supervised Homography estimation Network (SSHNet). SSHNet reformulates the unsupervised cross-modal homography estimation problem into two imperfect sub-problems: Sub-problem I: A supervised mono-modal homography estimation problem, where the modalities of input images are not perfectly consistent. Sub-problem II: A supervised modality transfer problem, where the supervisory image is not perfectly aligned. We introduce a split optimization strategy to separately train a homography estimation network for Sub-problem I and a modality transfer network for Sub-problem II, enabling effective convergence of the framework. Thanks to the proposed problem reformulation and the well-designed optimization strategy, our framework effectively solves the unsupervised task by training the networks under totally direct supervision, showing significantly improved training stability. This enables the successful training of iterative homography estimation architectures like IHN [4] and RHWF [5]. Better still, for the correlation-based homography estimation architectures [4, 5, 41], we propose an extra homography feature space supervision that enhances feature consistency, further improving homography estimation accuracy. Moreover, to address the computational burden of the modality transfer network, we introduce a simple yet effective distillation training technique, which notably reduces model parameters and improves cross-domain generalization ability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "type": "text", + "content": "We evaluate SSHNet on a variety of datasets including GoogleMap [42], DPDN [30], OPT-SAR [23], and Flash/no-flash [15] cross-modal datasets, together with RGB/NIR [2] cross-spectral datasets. SSHNet outperforms all the previous unsupervised homography estimation approaches with a large gap, and even achieves better performance compared to many supervised ones. The SSHNet using IHN as homography estimation network, namely SSHNet-IHN, owns " + }, + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "type": "inline_equation", + "content": "47.4\\%" + }, + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "type": "inline_equation", + "content": "85.8\\%" + }, + { + "bbox": [ + 55, + 493, + 298, + 685 + ], + "type": "text", + "content": " lower mean average corner errors (MACEs) than the supervised approaches MHN [20] and LocalTrans [34] on the challenging OPT-SAR dataset. Moreover, the SSHNet using the homography estimation architecture of the previous unsupervised SOTA approach SCPNet [41], namely SSHNet-SCPNet, even outperforms SCPNet itself, revealing the effectiveness of our proposed learning framework. To summarize, our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 714 + ], + "type": "text", + "content": "- We propose SSHNet, a novel unsupervised cross-modal homography estimation learning framework, which refor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 322, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 120 + ], + "type": "text", + "content": "mulates the unsupervised problem into two sub-problems with direct supervision. To realize stable training, we introduce a split optimization strategy that separately trains the network in each sub-problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 121, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 555, + 168 + ], + "type": "text", + "content": "- For correlation-based homography estimation architectures, we introduce an extra homography feature space supervision to improve feature consistency for homography estimation, which further improves estimation accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 168, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 228 + ], + "type": "text", + "content": "- Considering the computational burden introduced by the additional modality transfer network, we devise a simple yet effective distillation training technique to reduce model parameters while improving cross-domain generalization ability with little impact on performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 288 + ], + "type": "text", + "content": "- The training stability of SSHNet enables its cooperation with various homography estimation architectures, especially iterative ones like IHN and RHWF, resulting in higher performance. SSHNet-IHN outperforms the previous unsupervised approaches by a significant margin." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 298, + 406, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 406, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 406, + 311 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 534 + ], + "type": "text", + "content": "Supervised Deep Homography Estimation. DeTone et al. [9] introduced DHN, a VGG-style network to predict the homography between the source and target images. Le et al. [20] introduced a multi-scale VGG-style network, named MHN. Chang et al. [7] combined CNN with an untrainable IC-LK iterator. Zhao et al. [42] enhanced the feature similarity to improve the performance of IC-LK iterator. Shao et al. [34] proposed LocalTrans that adopts a multi-scale local transformer to register cross-resolution images. Cao et al. [4] proposed IHN, a trainable iterative framework with substantial accuracy. Cao et al. [5] further introduced RHWF that integrates homography-guided image warping and focus transformer, achieving higher performance. Deng et al. [8] proposed CrossHomo, a multi-level framework for cross-modality and cross-resolution homography estimation. Li et al. [22] introduced DMHomo, a diffusion-based framework to generate high-quality training pairs for supervised homography learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Unsupervised Deep Homography Estimation. Nguyen et al. [28] proposed UDHN, an unsupervised homography estimation network trained with photometric loss. Zhang et al. [40] introduced CA-UDHN, projecting images into a consistent feature space for effective estimation. Koguchiuk et al.[19] proposed biHomE, using perceptual loss[18] to improve robustness to intensity and viewpoint variations. Ye et al. [38] proposed BasesHomo, which enforces the image feature to be warp-equivalent with a feature identity loss and decomposes the homography matrix into flow bases. As modality and deformation gaps grow, intensity-based metrics for cross-modal homography matching may fail [41]. To address this, Zhang et al. [41] introduced SCPNet, which combines cross-modal intensity-based learning with intramodal self-supervised learning for homography knowledge" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16686" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 292, + 190 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 292, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 292, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 292, + 190 + ], + "type": "image", + "image_path": "5239e411bcd118d92376f41ba7195f75bc7dd1c9732febbbe9968690a12e268b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 295, + 260 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 260 + ], + "type": "text", + "content": "Figure 2. Reformulating the unsupervised cross-modal homography estimation problem into two sub-problems with direct supervision. (a) Unsupervised cross-modal homography estimation problem without direct supervision. (b) Mono-modal homography estimation problem with imperfect inputs. (c) Modality transfer problem with imperfect supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 265, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 348 + ], + "type": "text", + "content": "injection. Besides, some unsupervised approaches, such as NeMAR [1], RFNet [37], and UMF-CMGR [11], use modality transfer networks to convert one modality into the other, enabling cross-modal motion estimation. The above unsupervised approaches perform well on images with small deformations, but may produce unsatisfactory results under large deformations and modality gaps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 504 + ], + "type": "text", + "content": "Image-to-Image Translation. The image-to-image (I2I) translation problem focuses on converting images from a source domain to a target domain. Isola et al. [17] introduced Pix2pix, which employs conditional GANs and requires paired training data. Zhu et al. [45] later proposed CycleGAN, utilizing cycle-consistent loss to enable unpaired I2I translation. Liu et al.[24] developed UNIT, and Huang et al.[16] introduced MUNIT, both leveraging a shared latent space assumption to improve unpaired translation. More recently, Sasaki et al.[33] proposed UNIT-DDPM, integrating diffusion probabilistic models into the I2I translation framework. Li et al. [21] introduced BBDM, modeling I2I translation as a stochastic Brownian Bridge process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 192, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 192, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 192, + 526 + ], + "type": "text", + "content": "3. Problem Reformulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "content": "Denote " + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "content": " as a pair of images from modality A and B, with an unknown homography " + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{AB,GT}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "content": ". As illustrated in Fig. 2a, the unsupervised cross-modal homography estimation aims to train a network to estimate the homography " + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{\\mathrm{AB}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "content": ", without direct supervision of " + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{AB,GT}}" + }, + { + "bbox": [ + 55, + 534, + 296, + 712 + ], + "type": "text", + "content": ". To solve this problem, in SSHNet, we reformulate it into two sub-problems with direct supervision, including a monomodal homography estimation problem with imperfect inputs (Fig. 2b) and a modality transfer problem with imperfect supervision (Fig. 2c). The two sub-problems are mutually causal, and the imperfect condition of each sub-problem can be gradually addressed through the optimization of another sub-problem. In this way, the unsupervised homography estimation problem can be solved in a totally supervised manner." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 556, + 171 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 556, + 171 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 556, + 171 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 556, + 171 + ], + "type": "image", + "image_path": "84ab33bf4d0409efc768b6f8cc6ba8dbe50624d9e806a250d852944c50fc8c87.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 176, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 232 + ], + "type": "text", + "content": "Figure 3. Comparison of straight optimization and split optimization with regularization. (a) Cross-modal test MACEs for the network using straight optimization and split optimization with regularization, on OPT-SAR dataset. (b) Schematic diagram of straight optimization and split optimization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": "Sub-problem I. For the mono-modal homography estimation, generating image pairs with simulated ground-truth deformations for the direct self-supervised training is a common approach [4, 9, 20, 34]. If we can ideally convert " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": " into modality B, producing the transferred image " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": ", we can treat the unsupervised cross-modal homography estimation problem as the mono-modal one by estimating the homography between " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": ", as in Fig. 2b. The homography estimation network can be effectively trained using " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": " with the simulated homography deformations, in a self-supervised manner. We refer to the problem of estimating the homography between " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": " and the transferred image " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": " with the self-supervised training using " + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 241, + 555, + 396 + ], + "type": "text", + "content": " as Sub-problem I." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": "Sub-problem II. As illustrated in Fig. 2c, we also need a modality transfer network to obtain the optimal mapping from " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": ", producing " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": ". The ideal network training typically requires " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": " that is well-aligned with " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": " for supervision, which is also unavailable in the unsupervised cross-modal homography estimation. Previous approaches [1, 11] attempt to use GAN-based supervision to achieve modality transfer training using the unaligned image pairs. However, this strategy encounters considerable challenges when dealing with large deformation and modality gaps, as illustrated in [41]. Therefore, high-quality modality transfer requires well-aligned " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": ". Now that we have transferred " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": ", the requirement becomes that Sub-problem I being ideally solved. We refer to the modality transfer problem, which uses the homography estimated in Sub-problem I to produce a warped " + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 399, + 556, + 590 + ], + "type": "text", + "content": " for supervision, as Sub-problem II." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 592, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 687 + ], + "type": "text", + "content": "Straight Optimization Fails. However, straight optimization of both sub-problems simultaneously is challenging, as the settlement of either sub-problem depends on the assumption that the other has already been solved. We conduct a pilot experiment where the homography estimation network and the modality transfer network are trained simultaneously to address Sub-problem I and II, but the networks fail to converge, as shown by the blue curve in Fig. 3a." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Split Optimization with Regularization Success. Inspired by decomposition methods in convex optimization," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16687" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 87, + 242, + 167 + ], + "blocks": [ + { + "bbox": [ + 63, + 77, + 177, + 87 + ], + "lines": [ + { + "bbox": [ + 63, + 77, + 177, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 77, + 177, + 87 + ], + "type": "text", + "content": "(a) Optimization of Sub-problem I" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 87, + 242, + 167 + ], + "lines": [ + { + "bbox": [ + 63, + 87, + 242, + 167 + ], + "spans": [ + { + "bbox": [ + 63, + 87, + 242, + 167 + ], + "type": "image", + "image_path": "2794cc34fe00a2023aa416de20a220276cbe7f0cb1c67d4242dff0b4832e915b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 168, + 242, + 249 + ], + "blocks": [ + { + "bbox": [ + 63, + 168, + 242, + 249 + ], + "lines": [ + { + "bbox": [ + 63, + 168, + 242, + 249 + ], + "spans": [ + { + "bbox": [ + 63, + 168, + 242, + 249 + ], + "type": "image", + "image_path": "b1d6ee9134bb2bd41fdc3f66dd3c9cc7872edc1bd24a151ae88b60baba42c726.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 233, + 548, + 243 + ], + "lines": [ + { + "bbox": [ + 247, + 233, + 548, + 243 + ], + "spans": [ + { + "bbox": [ + 247, + 233, + 548, + 243 + ], + "type": "text", + "content": "Only Training Stage Both Training and Inference Stage Simulated Homography Generator" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. Schematic diagram of Split Supervised Homography estimation Network, namely SSHNet. (a) Detailed optimization process of Sub-problem I. (b) Detailed optimization process of Sub-problem II. (c) Extra supervision on the homography feature space." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 74, + 551, + 185 + ], + "blocks": [ + { + "bbox": [ + 249, + 74, + 551, + 185 + ], + "lines": [ + { + "bbox": [ + 249, + 74, + 551, + 185 + ], + "spans": [ + { + "bbox": [ + 249, + 74, + 551, + 185 + ], + "type": "image", + "image_path": "796f0533b0826ed2f57b2e9f43b61c0f995e99e4118ecb400ca3049b3eaf2004.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 191, + 264, + 204 + ], + "blocks": [ + { + "bbox": [ + 249, + 191, + 264, + 204 + ], + "lines": [ + { + "bbox": [ + 249, + 191, + 264, + 204 + ], + "spans": [ + { + "bbox": [ + 249, + 191, + 264, + 204 + ], + "type": "image", + "image_path": "3886c86cebc914962011f7750079373d4b27f156e5de442a05daa8dc7c0648c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 193, + 356, + 202 + ], + "lines": [ + { + "bbox": [ + 265, + 193, + 356, + 202 + ], + "spans": [ + { + "bbox": [ + 265, + 193, + 356, + 202 + ], + "type": "text", + "content": "Homography Estimation Network" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 191, + 384, + 204 + ], + "blocks": [ + { + "bbox": [ + 370, + 191, + 384, + 204 + ], + "lines": [ + { + "bbox": [ + 370, + 191, + 384, + 204 + ], + "spans": [ + { + "bbox": [ + 370, + 191, + 384, + 204 + ], + "type": "image", + "image_path": "2ca645c770e78fc72840a9eb9060c2595c00d196ba672f7cdccbdef885a2d161.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 193, + 460, + 202 + ], + "lines": [ + { + "bbox": [ + 386, + 193, + 460, + 202 + ], + "spans": [ + { + "bbox": [ + 386, + 193, + 460, + 202 + ], + "type": "text", + "content": "Modality Transfer Network" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 468, + 193, + 477, + 203 + ], + "blocks": [ + { + "bbox": [ + 468, + 193, + 477, + 203 + ], + "lines": [ + { + "bbox": [ + 468, + 193, + 477, + 203 + ], + "spans": [ + { + "bbox": [ + 468, + 193, + 477, + 203 + ], + "type": "image", + "image_path": "67123521fcf15b7cc9cee4a67fab7ea846d9d955500147f66be1a2ee7198d0fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 193, + 503, + 202 + ], + "lines": [ + { + "bbox": [ + 478, + 193, + 503, + 202 + ], + "spans": [ + { + "bbox": [ + 478, + 193, + 503, + 202 + ], + "type": "text", + "content": "Trainable" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 212, + 476, + 222 + ], + "lines": [ + { + "bbox": [ + 423, + 212, + 476, + 222 + ], + "spans": [ + { + "bbox": [ + 423, + 212, + 476, + 222 + ], + "type": "text", + "content": "Weight Sharing" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 513, + 193, + 528, + 202 + ], + "blocks": [ + { + "bbox": [ + 513, + 193, + 528, + 202 + ], + "lines": [ + { + "bbox": [ + 513, + 193, + 528, + 202 + ], + "spans": [ + { + "bbox": [ + 513, + 193, + 528, + 202 + ], + "type": "image", + "image_path": "66e5b3720dabd991d6c874fe8c66f6e1255c6d06905008fa9f501cf71ef77600.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 529, + 193, + 544, + 202 + ], + "lines": [ + { + "bbox": [ + 529, + 193, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 529, + 193, + 544, + 202 + ], + "type": "text", + "content": "rozen" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 249, + 211, + 263, + 224 + ], + "blocks": [ + { + "bbox": [ + 249, + 211, + 263, + 224 + ], + "lines": [ + { + "bbox": [ + 249, + 211, + 263, + 224 + ], + "spans": [ + { + "bbox": [ + 249, + 211, + 263, + 224 + ], + "type": "image", + "image_path": "7874e89d24a13231f03e76fd20cf551003e561d5f4d6192a8f8161abfe98ab31.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 213, + 410, + 223 + ], + "lines": [ + { + "bbox": [ + 264, + 213, + 410, + 223 + ], + "spans": [ + { + "bbox": [ + 264, + 213, + 410, + 223 + ], + "type": "text", + "content": "Feature Extractor in Homography Estimation Network" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 485, + 212, + 504, + 222 + ], + "blocks": [ + { + "bbox": [ + 485, + 212, + 504, + 222 + ], + "lines": [ + { + "bbox": [ + 485, + 212, + 504, + 222 + ], + "spans": [ + { + "bbox": [ + 485, + 212, + 504, + 222 + ], + "type": "image", + "image_path": "7608eb7f6b724e1471d657ed09d0507b514ab4883f88c22c2ba73cffff298a26.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 212, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 505, + 212, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 505, + 212, + 545, + 222 + ], + "type": "text", + "content": "mage Warping" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": "we then consider using a split optimization strategy to address the two sub-problems separately. During optimization, each network updates its weights with the other one's frozen. As in Fig. 2b and Fig. 2c, in each iteration, the homography estimation network in Sub-problem I is updated, producing a better-aligned " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " to guide the modality transfer network in Sub-problem II in the next iteration. Similarly, the modality transfer network in Sub-problem II is updated, generating a better transferred image " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " to serve as input for the homography estimation network in Sub-problem I. This process gradually improves the outputs of each network during training, facilitating the optimization of the other network, and finally makes both sub-problems nearly perfect supervised ones. However, at the beginning of the training, the modality gap between " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " in Sub-problem I is extremely large, so is the misalignment of " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " in Sub-problem II, which hinders the mutual promotion between two subproblems. A regularization term that guides the training into our desired direction is needed. Inspired by [41] " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": ", in Subproblem I, we replace the self-supervised training using only " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " by the two-branch self-supervised training using both " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 56, + 292, + 296, + 602 + ], + "type": "text", + "content": " (see Fig. 2b). This enables the homography estimation network to implicitly learn cross-modal homography knowledge at the beginning of the training, and triggers the two split optimized sub-problems to converge into our desired direction, as shown by the orange curve in Fig. 3a." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 612, + 111, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 111, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 111, + 624 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 632, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 668 + ], + "type": "text", + "content": "Based on the problem reformulation in Section 3, we further investigate its potential and propose an unsupervised cross-modal homography estimation learning framework, named" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 291, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 387 + ], + "type": "text", + "content": "Split Supervised Homography estimation Network (SSHNet). Section 4.1 covers the split optimization of SSHNet (Fig. 4a and Fig. 4b), Section 4.2 discusses the cooperation of SSHNet with various homography estimation architectures, Section 4.3 outlines the design of the modality transfer network, Section 4.4 presents the extra homography feature space supervision (Fig. 4c), and Section 4.5 illustrates the distillation training technique (Fig. 5)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 398, + 424, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 424, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 424, + 410 + ], + "type": "text", + "content": "4.1. Split Optimization" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 416, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 475 + ], + "type": "text", + "content": "As illustrated in Fig. 4a and Fig. 4b, the homography estimation network and the modality transfer network are optimized separately in each iteration, forming the split optimization framework of our SSHNet. In the following, we will elaborate the optimization procedure for each network." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": "Optimization of Sub-problem I. As mentioned in Section 3, we adopt a two-branch self-supervised learning strategy to optimize the homography estimation network in Subproblem I. We apply the separate simulated homography deformation on " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": ", generating the intra-modal image pairs " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_{\\mathrm{A}},\\mathbf{I}_{\\mathrm{A}}^{\\prime})" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_{\\mathrm{B}},\\mathbf{I}_{\\mathrm{B}}^{\\prime})" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " with corresponding ground-truth homography " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{A,GT}}" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{B,GT}}" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": ". The modality transfer network transfers " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_{\\mathrm{A}},\\mathbf{I}_{\\mathrm{A}}^{\\prime})" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime})" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": ". We simultaneously use " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_{\\mathrm{B}},\\mathbf{I}_{\\mathrm{B}}^{\\prime})" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime})" + }, + { + "bbox": [ + 313, + 477, + 556, + 620 + ], + "type": "text", + "content": " to train the weight-shared homography estimation network in a self-supervised manner. The optimization of Sub-problem I can be formulated as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 632, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 632, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 320, + 632, + 555, + 653 + ], + "type": "interline_equation", + "content": "\\underset {\\zeta} {\\arg \\min } L _ {\\mathrm {H}} \\left(\\mathcal {H} _ {\\zeta} \\left(\\mathbf {I} _ {\\mathrm {B}}, \\mathbf {I} _ {\\mathrm {B}} ^ {\\prime}\\right), \\mathbf {H} _ {\\mathrm {B}, \\mathrm {G T}}\\right) + \\mathcal {R} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}} ^ {\\prime}\\right), \\tag {1}", + "image_path": "9cae3324e848add608c1c4badb4ba75eee31c3b75fb2aba34e407903c8fb747b.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\zeta}" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " denotes the homography estimation network with parameters " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " to be optimized, and " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{H}}" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " denotes the homography estimation loss. " + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathbf{I}_{\\mathrm{A,T}},\\mathbf{I}_{\\mathrm{A,T}}^{\\prime}) = L_{\\mathrm{H}}\\bigl (\\mathcal{H}_{\\zeta}(\\mathcal{T}_{\\theta^{*}}(\\mathbf{I}_{\\mathrm{A}}),\\mathcal{T}_{\\theta^{*}}(\\mathbf{I}_{\\mathrm{A}}^{\\prime})),\\mathbf{H}_{\\mathrm{A,GT}}\\bigr)" + }, + { + "bbox": [ + 313, + 664, + 556, + 715 + ], + "type": "text", + "content": " , which is the regular" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "[41] finds that if training the weight-shared homography estimation network using the simulated deformation within each two modalities, the cross-modal homography knowledge can be implicitly learned through the multitask learning mechanism [6, 12]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16688" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": "ization term, where " + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta^*}" + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": " denotes the modality transfer network with parameters " + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": " frozen." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": "Optimization of Sub-problem II. We train a modality transfer network to convert " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " into an image of pseudo modality B, namely " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ". Following the reformulation in Section 3, we apply the homography matrix " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{\\mathrm{AB}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " predicted by " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\zeta}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " to warp " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ", providing direct supervision for the modality transfer network. The optimization of Sub-problem II can be formulated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 188, + 296, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 188, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 81, + 188, + 296, + 207 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } L _ {\\mathrm {T}} \\left(\\mathcal {T} _ {\\theta} (\\mathbf {I} _ {\\mathrm {A}}), \\mathcal {W} (\\mathbf {I} _ {\\mathrm {B}}, \\mathcal {H} _ {\\zeta^ {*}} (\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {B}}))\\right), \\tag {2}", + "image_path": "6bb480f72d5e4f2b0064de2d3eb79596da0d657742173c4ea0594032504d8f53.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta}" + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": " denotes the modality transfer network with parameters " + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": " to be optimized, " + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\zeta^{*}}" + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": " denotes the homography estimation network with parameters " + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\zeta^{*}" + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": " frozen, " + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": " denotes the image warping with the estimated homography, and " + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 213, + 296, + 273 + ], + "type": "text", + "content": " denotes the modality transfer loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 279, + 237, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 237, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 237, + 292 + ], + "type": "text", + "content": "4.2. Homography Estimation Network" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "type": "text", + "content": "Present homography estimation approaches exhibit a variety of architectural designs. Some approaches [9, 20, 28, 38, 40] concatenate input images in the channel dimension, directly estimating the homography deformation. Others [4, 5, 34, 41] use backbone networks to extract features and employ the feature correlation to perform homography decoding. Additionally, recent approaches [4, 5] further introduce iterative architectures to refine the estimation results, achieving substantial improvement in accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 404, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 571 + ], + "type": "text", + "content": "As illustrated in the experimental results of Section 5.3, our proposed framework is independent of the specific homography estimation architecture, enabling integration with various architectures. The pure supervised training strategy significantly boosts the stability of the optimization process. As a result, iterative homography estimation architectures like IHN [4] and RHWF [5] can be effectively employed, leading to a higher performance upper bound. In contrast, due to the lack of an explicit solution for consistent features between cross-modal images, the previous SOTA approach, SCPNet [41], struggles to converge when using these iterative architectures. For the self-supervised training, we adopt " + }, + { + "bbox": [ + 55, + 404, + 296, + 571 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 404, + 296, + 571 + ], + "type": "text", + "content": " homography estimation loss following previous works [4, 5, 9, 20, 41], which is formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 577, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 577, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 108, + 577, + 295, + 597 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {H}} (\\hat {\\mathbf {H}}, \\mathbf {H} _ {\\mathrm {G T}}) = \\left\\| \\hat {\\mathbf {H}} - \\mathbf {H} _ {\\mathrm {G T}} \\right\\| _ {1}, \\tag {3}", + "image_path": "7de85404f90c279eb3f1a5e4dfb74a7faf2ca2cc904fe06acc0acbbd5d402125.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{H}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " denotes the estimated homography, " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{GT}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " denotes the ground-truth homography. For iterative homography architectures, we use a weighted sum of the " + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 605, + 296, + 654 + ], + "type": "text", + "content": " distance following [4, 5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 208, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 208, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 208, + 673 + ], + "type": "text", + "content": "4.3. Modality Transfer Network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We design the modality transfer network in our SSHNet as a U-shaped architecture to enhance the perception of multiscale information. Inspired by Swin-Unet [3], the Swin" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "Transformer block [25] serves as the basic unit of the modality transfer network. The attention mechanism enhances the network's ability to capture hierarchical structures and context information. This facilitates more effective learning of complex mappings between modalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": "As mentioned in Section 4.1, we warp image " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " with the " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{\\mathrm{AB}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " predicted by the homography estimation network for the modality transfer supervision. During optimization, " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{\\mathrm{AB}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " cannot ideally align the input image " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " with the supervisory image " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": ". Therefore, instead of using pixel-wise " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " loss, we apply perceptual loss " + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pcp}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 228 + ], + "type": "text", + "content": " [18] to capture high-level representations, ensuring better tolerance to misalignments, which is formulated as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 237, + 567, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 237, + 567, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 567, + 277 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {p c p}} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {I} _ {\\mathrm {B}, \\mathrm {W}}\\right) = \\sum_ {j} \\frac {1}{C _ {j} H _ {j} W _ {j}} \\left\\| \\psi_ {j} \\left(\\mathbf {I} _ {\\mathrm {A}, \\mathrm {T}}\\right) - \\psi_ {j} \\left(\\mathbf {I} _ {\\mathrm {B}, \\mathrm {W}}\\right) \\right\\| _ {2} ^ {2}, \\tag {4}", + "image_path": "77a37c1dbac21627875f8fdbf16677e617ecd3ba8395dd3b9fd18182ec6e0281.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "inline_equation", + "content": "\\psi_{j}" + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": "-th layer of the pre-trained VGG with a feature map of shape " + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "inline_equation", + "content": "C_j\\times H_j\\times W_j" + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B,W}}" + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": " denotes the image " + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": " warped by the estimated " + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{\\mathrm{AB}}" + }, + { + "bbox": [ + 313, + 277, + 554, + 315 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 323, + 554, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 335 + ], + "type": "text", + "content": "4.4. Extra Homography Feature Space Supervision" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 411 + ], + "type": "text", + "content": "The current modality transfer network in Sub-problem II is supervised by the perceptual loss [18], which computes feature similarity using a pre-trained VGG network. This may yield suboptimal results for homography estimation in Sub-problem I when trained under the present totally split optimization framework." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": "To address this issue, as illustrated in Fig. 4c, we introduce an extra supervision in the homography feature space when the architecture of homography estimation network has explicit feature extractors, such as IHN, RHWF, and SCPNet. A homography feature (HF) loss that directly enhances the feature consistency for homography estimation between " + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": " is computed. The modality transfer network and the feature extractor of the homography estimation network are optimized together under the supervision of HF loss, which can be expressed as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 336, + 542, + 555, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 542, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 336, + 542, + 555, + 563 + ], + "type": "interline_equation", + "content": "\\underset {\\theta , \\xi} {\\arg \\min } L _ {\\mathrm {H F}} \\left(\\mathcal {F} _ {\\xi} \\left(\\mathcal {T} _ {\\theta} (\\mathbf {I} _ {\\mathrm {A}})\\right), \\mathcal {F} _ {\\xi} \\left(\\mathcal {W} \\left(\\mathbf {I} _ {\\mathrm {B}}, \\hat {\\mathbf {H}} _ {\\mathrm {A B}}\\right)\\right)\\right), \\tag {5}", + "image_path": "b452948d401fce160198a8c71d82827199e82ba42662bc9c4e762bc55360fae0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\xi}" + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "text", + "content": " denotes the feature extractor with parameters " + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "text", + "content": " to be optimized. We note that IHN, RHWF, and SCPNet all depend on feature correlation to perform homography decoding. Considering this, We formulate the HF loss using a correlation-based loss function " + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{corr}}" + }, + { + "bbox": [ + 313, + 572, + 554, + 644 + ], + "type": "text", + "content": " to enhance feature consistency, which is formulated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 654, + 555, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 654, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 320, + 654, + 555, + 680 + ], + "type": "interline_equation", + "content": "L _ {\\operatorname {c o r r}} \\left(\\mathbf {F} _ {\\mathrm {A}, \\mathrm {T}}, \\mathbf {F} _ {\\mathrm {B}, \\mathrm {W}}\\right) = - \\sum_ {i, j} \\mathbf {F} _ {\\mathrm {A}, \\mathrm {T}} (i, j) ^ {\\top} \\mathbf {F} _ {\\mathrm {B}, \\mathrm {W}} (i, j), \\tag {6}", + "image_path": "24147dd650d7066094a6cbf94b7f9e0b201d93803c8d64c6c637071fe8f18d28.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{B,W}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " denote the feature maps of " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{A,T}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{B,W}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " in the homography estimation network." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16689" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 286, + 115 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 286, + 115 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 286, + 115 + ], + "type": "image", + "image_path": "b9093a72265be09818f5a97f6e14019bb401ff28c52d57dc9671abfa6862b6b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 119, + 277, + 130 + ], + "lines": [ + { + "bbox": [ + 73, + 119, + 277, + 130 + ], + "spans": [ + { + "bbox": [ + 73, + 119, + 277, + 130 + ], + "type": "text", + "content": "Figure 5. Schematic diagram of the distillation training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 174, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 174, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 174, + 148 + ], + "type": "text", + "content": "4.5. Distillation Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 152, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 224 + ], + "type": "text", + "content": "We can see that the present framework has extra network parameters due to the utilization of the modality transfer network in Sub-problem II. Moreover, as the modality transfer network is trained on multi-modal image pairs from a fixed domain, its generalization ability to images from other domains may be limited." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 344 + ], + "type": "text", + "content": "To cope with the above problem, we introduce a distillation training technique under the inspiration of the unsupervised knowledge distillation frameworks in domain adaptive object detection [32, 43]. We use SSHNet as a teacher network to guide a student network that only retains the homography estimation network, named Distilled SSHNet (SSHNet-D). During the distillation training, the homography " + }, + { + "bbox": [ + 55, + 224, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{H}}_{\\mathrm{teacher}}" + }, + { + "bbox": [ + 55, + 224, + 296, + 344 + ], + "type": "text", + "content": " predicted by SSHNet serves as a pseudo ground-truth to supervise SSHNet-D, as illustrated in Fig. 5. The distillation training can be formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 352, + 296, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 352, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 102, + 352, + 296, + 371 + ], + "type": "interline_equation", + "content": "\\underset {\\omega} {\\arg \\min } L _ {\\mathrm {H}} \\left(\\phi_ {\\omega} \\left(\\mathbf {I} _ {\\mathrm {A}}, \\mathbf {I} _ {\\mathrm {B}}\\right), \\hat {\\mathbf {H}} _ {\\text {t e a c h e r}}\\right), \\tag {7}", + "image_path": "19513d6d26db6c4bf32e67eefe5cf1a350008963182de426131290c33ae0d83f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\phi_{\\omega}" + }, + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "type": "text", + "content": " denotes the SSHNet-D with the parameters " + }, + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 379, + 296, + 403 + ], + "type": "text", + "content": " to be optimized." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 413, + 137, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 137, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 137, + 426 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 433, + 182, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 182, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 182, + 445 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": "Datasets. We evaluate SSHNet on various datasets, including GoogleMap [42], DPDN [30], OPT-SAR [23], Flash/no-flash [15] cross-modal datasets, and RGB/NIR [2] cross-spectral dataset. Examples for each dataset are shown in Fig. 6. To ensure a fair comparison, all evaluated methods are trained and tested using the same training and test splits within each dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 653 + ], + "type": "text", + "content": "Baselines. We evaluate SSHNet with traditional feature-based approaches including SIFT [26], ORB [31], unsupervised approaches including UDHN [28], biHomE [19], CA-UDHN [40], BasesHomo [38], UMF-CMGR [11], SCP-Net [41], and supervised approaches including DHN [9], MHN [20], LocalTrans [34], IHN [4], RHWF [5]. For feature-based approaches, we choose RANSAC [13] as the outlier rejection algorithms. UMF-CMGR is an image fusion approach based on registration, and we use the registration part for comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Homography Estimation Architectures. To investigate the potential of our proposed SSHNet, we adopt various homography estimation architectures including DHN [9] and MHN [20], together with iterative architectures such as IHN [4] and RHWF [5]. Moreover, we also adopt the" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 83, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 541, + 82 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 541, + 82 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 541, + 82 + ], + "type": "text", + "content": "Table 1. Ablation study on the training strategy of SSHNet." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 83, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 83, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 83, + 552, + 148 + ], + "type": "table", + "html": "
ExperimentMACE ↓
w/o problem reformulationNC
w/o split optimizationNC
pure self-supervised learning24.60
SSHNet2.94
", + "image_path": "c8e0856bef4e0c138f65b44a0e6f9da9908f4545cc7caa56fc4970688eedfed3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 317, + 177, + 552, + 217 + ], + "blocks": [ + { + "bbox": [ + 313, + 154, + 554, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 554, + 174 + ], + "type": "text", + "content": "Table 2. Ablation study on the architectures of the modality transfer network." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 177, + 552, + 217 + ], + "lines": [ + { + "bbox": [ + 317, + 177, + 552, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 552, + 217 + ], + "type": "table", + "html": "
ExperimentParameters (M)MACE ↓
CNN-based7.488.72
Transformer-based7.542.94
", + "image_path": "c9c42742ad791b07d2dcc7adb6e5013fd69a4b2b3dda251f15b184bf073a81cd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 247, + 552, + 330 + ], + "blocks": [ + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "type": "text", + "content": "Table 3. Ablation study on the loss functions. " + }, + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "type": "text", + "content": " denotes the modality transfer loss, " + }, + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{HF}}" + }, + { + "bbox": [ + 313, + 223, + 554, + 246 + ], + "type": "text", + "content": " denotes the homography feature loss." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 247, + 552, + 330 + ], + "lines": [ + { + "bbox": [ + 317, + 247, + 552, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 247, + 552, + 330 + ], + "type": "table", + "html": "
ExperimentLoss TypeMACE ↓
LTL15.88
L26.23
Lpcp4.52
LT + LHFLpcp + L14.09
Lpcp + L23.27
Lpcp + Lcorr2.94
", + "image_path": "7a4d2b929a447173699163574fb50e6d526827c4fcb261f5db3f3aab44024c24.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 349, + 552, + 394 + ], + "blocks": [ + { + "bbox": [ + 328, + 335, + 539, + 346 + ], + "lines": [ + { + "bbox": [ + 328, + 335, + 539, + 346 + ], + "spans": [ + { + "bbox": [ + 328, + 335, + 539, + 346 + ], + "type": "text", + "content": "Table 4. Ablation study of the modality transfer direction." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 349, + 552, + 394 + ], + "lines": [ + { + "bbox": [ + 317, + 349, + 552, + 394 + ], + "spans": [ + { + "bbox": [ + 317, + 349, + 552, + 394 + ], + "type": "table", + "html": "
Method\\DirectionRGB→NIRNIR→RGBOPT→SARSAR→OPT
SSHNet1.661.752.944.28
", + "image_path": "d2ab19dfc218e01baa5efdcedbc786f601f3a41677594264be55687d55c08e3a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "content": "architecture from the previous SOTA unsupervised approach SCPNet [41], to specifically evaluate the effectiveness of our proposed unsupervised learning framework." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "Metrics. Similar to previous works [4, 5, 9, 34, 42], we randomly perturb the corner points of " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " images to make the deformed image, with a perturbation range set to [-32, 32]. We evaluate homography estimation accuracy using the mean average corner error (MACE) following [4, 5, 9, 34, 42]. A lower MACE indicates a higher accuracy." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 517, + 378, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 378, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 378, + 528 + ], + "type": "text", + "content": "5.2. Ablation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 534, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 568 + ], + "type": "text", + "content": "The ablation studies are conducted on the OPT-SAR dataset. We adopt IHN as the homography estimation network in our SSHNet for ablation studies." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Training Strategy. We conduct an ablation study on the training strategy of our SSHNet. The experimental results in Table 1 demonstrate that both problem reformulation and the split optimization strategy are crucial for SSHNet. Without problem reformulation, unsupervised training with intensity-based supervision fails to converge. Even with problem reformulation, directly optimizing both networks simultaneously does not yield reliable homography estimation. SSHNet achieves effective convergence only when problem reformulation is combined with the split optimization strategy, where each network is trained separately. We also compare our framework with method trained through" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16690" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 106, + 116 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 106, + 116 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 106, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 106, + 116 + ], + "type": "image", + "image_path": "4d3bcc836cb9d77f179be2784cda938a81f454df88aa0a4f00f0c84d4f4bc1d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 117, + 128, + 124 + ], + "lines": [ + { + "bbox": [ + 85, + 117, + 128, + 124 + ], + "spans": [ + { + "bbox": [ + 85, + 117, + 128, + 124 + ], + "type": "text", + "content": "(a) GoogleMap" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 107, + 71, + 152, + 116 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 152, + 116 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 152, + 116 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 152, + 116 + ], + "type": "image", + "image_path": "ad883459225c75b0a40e9de50d67063a10a43a88424cc96af2e78ad321243a60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 205, + 116 + ], + "blocks": [ + { + "bbox": [ + 160, + 72, + 205, + 116 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 205, + 116 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 205, + 116 + ], + "type": "image", + "image_path": "95f4f8c36a1e5c14a5b293a43c33a37d8318c8d313f3a7e52f6abb0593e5de83.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 117, + 220, + 124 + ], + "lines": [ + { + "bbox": [ + 191, + 117, + 220, + 124 + ], + "spans": [ + { + "bbox": [ + 191, + 117, + 220, + 124 + ], + "type": "text", + "content": "(b) DPDN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 205, + 72, + 250, + 116 + ], + "blocks": [ + { + "bbox": [ + 205, + 72, + 250, + 116 + ], + "lines": [ + { + "bbox": [ + 205, + 72, + 250, + 116 + ], + "spans": [ + { + "bbox": [ + 205, + 72, + 250, + 116 + ], + "type": "image", + "image_path": "467ea23b4e37cc7c7271bba3d061132216adddc75a548bfac04157f0429e207c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 128, + 371, + 139 + ], + "lines": [ + { + "bbox": [ + 238, + 128, + 371, + 139 + ], + "spans": [ + { + "bbox": [ + 238, + 128, + 371, + 139 + ], + "type": "text", + "content": "Figure 6. Examples for each dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 259, + 72, + 304, + 116 + ], + "blocks": [ + { + "bbox": [ + 259, + 72, + 304, + 116 + ], + "lines": [ + { + "bbox": [ + 259, + 72, + 304, + 116 + ], + "spans": [ + { + "bbox": [ + 259, + 72, + 304, + 116 + ], + "type": "image", + "image_path": "e6808aeee93d068056181739cb07aab218f753be6f5a9001e581712f5f2a7367.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 117, + 323, + 124 + ], + "lines": [ + { + "bbox": [ + 285, + 117, + 323, + 124 + ], + "spans": [ + { + "bbox": [ + 285, + 117, + 323, + 124 + ], + "type": "text", + "content": "(c) OPT-SAR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 349, + 116 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 349, + 116 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 349, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 349, + 116 + ], + "type": "image", + "image_path": "b9915e447e38c24b8aa4d4d849ed6b021f8d96fe98c70016290e4f830c8f60bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 357, + 72, + 402, + 116 + ], + "blocks": [ + { + "bbox": [ + 357, + 72, + 402, + 116 + ], + "lines": [ + { + "bbox": [ + 357, + 72, + 402, + 116 + ], + "spans": [ + { + "bbox": [ + 357, + 72, + 402, + 116 + ], + "type": "image", + "image_path": "f24c14a78ab48ae9c4c78db7ecef954b184eb574b475bd146c064a0ef6b96c44.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 117, + 427, + 124 + ], + "lines": [ + { + "bbox": [ + 378, + 117, + 427, + 124 + ], + "spans": [ + { + "bbox": [ + 378, + 117, + 427, + 124 + ], + "type": "text", + "content": "(d) Flash/no-flash" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 448, + 116 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 448, + 116 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 448, + 116 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 448, + 116 + ], + "type": "image", + "image_path": "3374e92360f8b22776246c78e9d0b6664d62235f9578e2b76b8b08f9a3351a3a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 456, + 72, + 500, + 116 + ], + "blocks": [ + { + "bbox": [ + 456, + 72, + 500, + 116 + ], + "lines": [ + { + "bbox": [ + 456, + 72, + 500, + 116 + ], + "spans": [ + { + "bbox": [ + 456, + 72, + 500, + 116 + ], + "type": "image", + "image_path": "e678698ac822c50d4530bf8c899f621e5cde91d7647b0945253d656989df89b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 117, + 520, + 124 + ], + "lines": [ + { + "bbox": [ + 484, + 117, + 520, + 124 + ], + "spans": [ + { + "bbox": [ + 484, + 117, + 520, + 124 + ], + "type": "text", + "content": "(e) RGB/NIR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 501, + 72, + 547, + 116 + ], + "blocks": [ + { + "bbox": [ + 501, + 72, + 547, + 116 + ], + "lines": [ + { + "bbox": [ + 501, + 72, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 501, + 72, + 547, + 116 + ], + "type": "image", + "image_path": "4a21519ed0aa91bc732f2e0323b2efb0e8aa9003765687d2b4ec487d12fe87fc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 58, + 170, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 143, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 555, + 166 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of our SSHNet and other approaches. NC denotes the training is not converged. The best results of unsupervised methods are highlighted in bold." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 170, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 170, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 170, + 553, + 392 + ], + "type": "table", + "html": "
Dataset ApproachGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
TraditionalSIFT [26]24.5325.1424.8218.2921.33
ORB [31]24.5224.8624.9020.2222.67
SupervisedDHN [9]5.204.928.276.4211.88
MHN [20]2.692.955.595.245.52
LocalTrans [34]3.221.7920.681.521.72
IHN [4]0.921.171.670.801.63
RHWF [5]0.711.061.480.651.07
UnsupervisedUDHN [28]22.84NCNC21.2023.43
CA-UDHN [40]24.6124.9924.7621.3224.12
biHomE [19]NCNCNC11.8623.77
BasesHomo [38]24.4927.3326.7225.1224.41
UMF-CMGR [11]24.6025.4624.7023.4922.38
SCPNet [41]4.35NCNC2.674.78
SSHNet-DHN (Ours)9.289.7117.6310.5112.13
SSHNet-MHN (Ours)2.903.046.775.626.93
SSHNet-SCPNet (Ours)3.894.8112.942.653.86
SSHNet-RHWF (Ours)1.291.263.080.951.52
SSHNet-IHN (Ours)1.231.122.941.081.66
SSHNet-IHN-D (Ours)1.261.243.311.161.72
", + "image_path": "43da390553b587872aac5082b8ab376616c7e5c46d30fdd07ad56efb0323a032.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 400, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 437 + ], + "type": "text", + "content": "pure self-supervised learning. Results in Table 1 show that self-supervised learning alone does not yield satisfactory homography estimation results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": "Architectures of the Modality Transfer Network. As illustrated in Table 2. Our Transformer-based architecture outperforms the CNN-based architecture by a large margin, with similar parameter counts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": "Loss Functions. To further investigate the effectiveness of the modality transfer and homography feature loss functions, we first train the network without HF loss to evaluate the performance of different types of modality transfer loss. The results in the upper part of Table 3 indicate that the perceptual loss " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pcp}}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": ". Then we adopt " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{pcp}}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": " as the modality transfer loss and add different types of HF loss to analyze their impact. We report the results in the lower part of Table 3. Though all different types of HF loss further boost the network's performance, the correlation-based loss " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{corr}}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": " achieves superior performance compared to " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 494, + 296, + 637 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Modality transfer direction. Table 4 presents the results of different transfer directions on the RGB/NIR and OPT-SAR datasets. Since some modalities carry more information, transferring from richer to less informative modalities facilitates the modality transfer network's learning process, thereby promoting the framework's convergence." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 400, + 507, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 507, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 507, + 414 + ], + "type": "text", + "content": "5.3. Comparisons with Existing Methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "text", + "content": "We compare the homography estimation results of our proposed SSHNet and baselines and list the experimental results in Table 5. It is observed that the prior traditional and unsupervised approaches can hardly handle the homography estimation under large modality and deformation gaps. In contrast, our method achieves much more satisfactory results with various homography estimation architectures. When adopting the iterative architectures IHN and RHWF, our method outperforms all the compared traditional and unsupervised approaches by a significant margin. Moreover, even though trained without direct cross-modal supervision, SSHNet-IHN and SSHNet-RHWF outperforms supervised approaches such as DHN, MHN and LocalTrans. Specifically, SSHNet-IHN owns " + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "inline_equation", + "content": "54.3\\%" + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "inline_equation", + "content": "47.4\\%" + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "text", + "content": " lower MACEs than MHN, and " + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "inline_equation", + "content": "61.8\\%" + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "inline_equation", + "content": "85.8\\%" + }, + { + "bbox": [ + 313, + 422, + 555, + 674 + ], + "type": "text", + "content": " lower MACEs than LocalTrans on the challenging GoogleMap and OPT-SAR datasets, respectively. Notably, SSHNet-SCPNet, which uses the homography estimation architecture from the previous unsupervised SOTA approach SCPNet, outperforms SCPNet itself on all datasets, further proving the superiority of our proposed training framework." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We also conduct the distillation training on SSHNet with different homography estimation architectures. We present the distillation training results of SSHNet-IHN in Table 5," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16691" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 106, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 103 + ], + "type": "text", + "content": "Table 6. Cross-dataset evaluation results. Red indicates the improved performance of SSHNet-IHN-D compared to SSHNet-IHN, and blue indicates the declined one." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 106, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 294, + 187 + ], + "type": "table", + "html": "
Test TrainGoogleMapDPDNOPT-SARFlash/no-flashRGB/NIR
GoogleMap/↑1.8%↑1.0%↑80.4%↑63.0%
DPDN↑66.5%/↑0.4%↑92.1%↑75.8%
OPT-SAR↑48.4%↑1.3%/↑77.5%↑66.2%
Flash/no-flash↓5.2%↓0.2%↑0.1%/↓4.6%
RGB/NIR↓0.5%↑1.2%↑3.0%↑2.1%/
", + "image_path": "0d2946f81085d3048c98c621ac4606ecfddf6c59d0e98b5f5956a3ce97d76737.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 195, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 243 + ], + "type": "text", + "content": "namely SSHNet-IHN-D. We observe that the performance of SSHNet-IHN-D is comparable to that of SSHNet-IHN. For more detailed qualitative visualization and experimental results, please refer to supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 216, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 216, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 216, + 262 + ], + "type": "text", + "content": "5.4. Cross-Dataset Generalization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 268, + 297, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 297, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 297, + 434 + ], + "type": "text", + "content": "We further conduct a cross-dataset evaluation of SSHNet-IHN-D and SSHNet-IHN on the aforementioned datasets, to further reveal the effectiveness of our proposed distillation training. As illustrated in Fig. 6, multi-modal image pairs from the GoogleMap, DPDN, and OPT-SAR datasets exhibit substantial modality gaps, and we call these three datasets challenging datasets, for a better illustration. In contrast, the Flash/no-flash and RGB/NIR datasets show smaller modality gaps, so we call them easy datasets. For each dataset, we compute the percentage of homography predictions that have average corner errors (ACE) lower than 5 pixels on the test set, which reveals the ratio of reliable estimations. We directly list the differences of the percentage of SSHNet-IHN-D and SSHNet-IHN in Table 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "text", + "content": "SSHNet-IHN uses a modality transfer network to transfer images from one modality to the other on the training dataset, making it highly prone to overfit the training data. In contrast, SSHNet-IHN-D excludes the modality transfer network and directly extracts consistent features from cross-modal images, and hence is forced to learn the features that is robust across different domains. As illustrated in Table 6, when trained on challenging datasets, SSHNet-IHN-D exhibits significant superior generalization ability. Specifically, SSHNet-IHN-D trained on DPDN yields " + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "inline_equation", + "content": "66.5\\%" + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "inline_equation", + "content": "92.1\\%" + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "inline_equation", + "content": "75.8\\%" + }, + { + "bbox": [ + 55, + 436, + 297, + 652 + ], + "type": "text", + "content": " improvement on GoogleMap, Flash/no-flash and RGB/NIR datasets compared to SSHNet. We note that the generalization ability across the challenging datasets remains limited due to the substantial domain gaps between the datasets. As for training on easy datasets, the generalization performance improvement is plain. This is because the training data is too easy to provide enough knowledge for obtaining the features robust across different domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 180, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 180, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 180, + 672 + ], + "type": "text", + "content": "5.5. Real World Scenarios" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We further evaluate our method on real-world scenarios [35] that do not satisfy perfect homography alignment, including RGB/NIR and Flash/no-flash image pairs. As illustrated" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 426, + 163 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 426, + 163 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 426, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 426, + 163 + ], + "type": "image", + "image_path": "325347c9fa2695dc520ab0fd4fd97ccc0d3b628fd61bed4f182edcdd2522d0b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 163, + 408, + 173 + ], + "lines": [ + { + "bbox": [ + 350, + 163, + 408, + 173 + ], + "spans": [ + { + "bbox": [ + 350, + 163, + 408, + 173 + ], + "type": "text", + "content": "(a) Cross modal" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 177, + 545, + 187 + ], + "lines": [ + { + "bbox": [ + 321, + 177, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 321, + 177, + 545, + 187 + ], + "type": "text", + "content": "Figure 7. Visualization of results on real-world deformations." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 554, + 163 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 554, + 163 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 554, + 163 + ], + "type": "image", + "image_path": "110122b662fed84b63dd790f30ef911e0b1ecbebf8bd616c957a6dfe2091461f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 163, + 516, + 173 + ], + "lines": [ + { + "bbox": [ + 468, + 163, + 516, + 173 + ], + "spans": [ + { + "bbox": [ + 468, + 163, + 516, + 173 + ], + "type": "text", + "content": "(b) UDIS-TIR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 217, + 552, + 266 + ], + "blocks": [ + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "type": "text", + "content": "Table 7. Experimental results on the UDIS-TIR dataset. " + }, + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "type": "inline_equation", + "content": "I_{3 \\times 3}" + }, + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "type": "text", + "content": " denotes the identity matrix." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 217, + 552, + 266 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 552, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 552, + 266 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑
I3×312.920.306
SCPNetNCNC
SSHNet-IHN20.050.728
", + "image_path": "0da805ba45253c14fba25a11743982726418fab9c5e647cffa39631b14034a11.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 274, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 393 + ], + "type": "text", + "content": "in Fig. 7a, the model trained with simulated homographies achieves effective performance. We also use a pretrained RGB-to-TIR translation network to convert RGB-RGB image pairs from the UDIS [29] dataset into RGB-TIR image pairs, creating the UDIS-TIR dataset with parallax image pairs. Following UDIS, we evaluate SSHNet-IHN on the UDIS-TIR dataset and compute the PSNR and SSIM on RGB images before modality transfer. We present the quantitative results in Table 7 and show more visualization results in Fig. 7b." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 404, + 391, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 404, + 391, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 404, + 391, + 417 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 556, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 603 + ], + "type": "text", + "content": "In this paper, we propose SSHNet, a novel unsupervised cross-modal homography estimation framework. SSHNet reformulates the problem into two supervised sub-problems, addressed by a homography estimation network and a modality transfer network. These networks are trained separately using a split optimization strategy, ensuring stable training. We also introduce an extra homography feature space supervision to further improve feature consistency during the homography estimation and boost the performance of our framework. Moreover, we employ a distillation training technique to reduce model parameters and improve cross-domain generalization ability. Experimental results show that our SSHNet achieves SOTA performance on multiple datasets among unsupervised approaches, and outperforms many supervised approaches." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 603, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 556, + 703 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the Zhejiang Provincial Natural Science Foundation of China under grant LD24F020003, in part by the National Natural Science Foundation of China under grant 62301484, in part by the Ningbo Natural Science Foundation of China under grant 2024J454, and in part by the National Key Research and Development Program of China under grant 2023YFB3209800. We also thank the generous help from Jianxin Hu, Zhejiang University." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16692" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "text", + "content": "[1] Moab Arar, Yiftach Ginger, Dov Danon, Amit H Bermano, and Daniel Cohen-Or. Unsupervised multi-modal image registration via geometry preserving image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13410–13419, 2020. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Matthew Brown and Sabine Süssstrunk. Multispectral SIFT for scene category recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 177-184. IEEE, 2011. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 296, + 257 + ], + "type": "text", + "content": "[3] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 205-218. Springer, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 302 + ], + "type": "text", + "content": "[4] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1879-1888, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 367 + ], + "type": "text", + "content": "[5] Si-Yuan Cao, Runmin Zhang, Lun Luo, Beinan Yu, Zehua Sheng, Junwei Li, and Hui-Liang Shen. Recurrent homography estimation using homography-guided image warping and focus transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9833-9842, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 389 + ], + "type": "text", + "content": "[6] Rich Caruana. Multitask learning. Machine Learning, 28: 41-75, 1997. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 296, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 444 + ], + "type": "text", + "content": "[7] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. CLKN: Cascaded Lucas-Kanade networks for image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2213-2221, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "text", + "content": "[8] Xin Deng, Enpeng Liu, Chao Gao, Shengxi Li, Shuhang Gu, and Mai Xu. CrossHomo: Cross-modality and cross-resolution homography estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 523 + ], + "type": "text", + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 590 + ], + "type": "text", + "content": "[10] Fayaz Ali Dharejo, Muhammad Zawish, Farah Deeba, Yuanchun Zhou, Kapal Dev, Sunder Ali Khowaja, and Nawab Muhammad Faseeh Qureshi. Multimodal-boost: Multimodal medical image super-resolution using multi-attention network with wavelet transform. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "text", + "content": "[11] Wang Di, Liu Jinyuan, Fan Xin, and Risheng Liu. Unsupervised misaligned infrared and visible image fusion via cross-modality image generation and registration. In Proceedings of the International Joint Conference on Artificial Intelligence, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[12] Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2051-2060, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[14] Hunter Goforth and Simon Lucey. GPS-denied UAV localization using pre-existing satellite imagery. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 2974-2980. IEEE, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "text", + "content": "[15] Shengfeng He and Rynson WH Lau. Saliency detection with flash and no-flash image pairs. In Proceedings of the European Conference on Computer Vision, pages 110-124. Springer, 2014. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 555, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 230 + ], + "type": "text", + "content": "[16] Xun Huang, Ming-Yu Liu, Serge Belongie, and Jan Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision, pages 172-189, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 284 + ], + "type": "text", + "content": "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1125-1134, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "type": "text", + "content": "[18] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Proceedings of the European Conference on Computer Vision, pages 694–711. Springer, 2016. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 333, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 555, + 386 + ], + "type": "text", + "content": "[19] Daniel Koguciuk, Elahe Arani, and Bahram Zonooz. Perceptual loss for robust unsupervised homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4274-4283, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[20] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7661, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "text", + "content": "[21] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. BBDM: Image-to-image translation with brownian bridge diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1952-1961, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 479, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 554, + 521 + ], + "type": "text", + "content": "[22] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. DMHomo: Learning homography with diffusion models. ACM Transactions on Graphics, 43 (3):1-16, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 523, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 555, + 588 + ], + "type": "text", + "content": "[23] Xue Li, Guo Zhang, Hao Cui, Shasha Hou, Shunyao Wang, Xin Li, Yujia Chen, Zhijiang Li, and Li Zhang. MCANet: A joint semantic segmentation framework of optical and SAR images for land use classification. International Journal of Applied Earth Observation and Geoinformation, 106:102638, 2022. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "type": "text", + "content": "[24] Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. Advances in Neural Information Processing Systems, 30, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[26] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 6, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16693" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "text", + "content": "[27] Iman Marivani, Evaggelia Tsiligianni, Bruno Cornelis, and Nikos Deligiannis. Designing CNNs for multimodal image restoration and fusion via unfolding the method of multipliers. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5830-5845, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 184 + ], + "type": "text", + "content": "[28] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters, 3(3):2346-2353, 2018. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[29] Lang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Unsupervised deep image stitching: Reconstructing stitched features to images. IEEE Transactions on Image Processing, 30:6184-6197, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "text", + "content": "[30] Gernot Riegler, David Ferstl, Matthias Rüther, and Horst Bischof. A deep primal-dual network for guided depth superresolution. In *British Machine Vision Conference*. The British Machine Vision Association, 2016. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[31] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 365 + ], + "type": "text", + "content": "[32] Kuniaki Saito, Yoshitaka Ushiku, and Tatsuya Harada. Asymmetric tri-training for unsupervised domain adaptation. In Proceedings of the International Conference on Machine Learning, pages 2988-2997. PMLR, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 296, + 409 + ], + "type": "text", + "content": "[33] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 466 + ], + "type": "text", + "content": "[34] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. LocalTrans: A multiscale local transformer network for cross-resolution homography estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14890-14899, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 511 + ], + "type": "text", + "content": "[35] Xiaoyong Shen, Li Xu, Qi Zhang, and Jiaya Jia. Multimodal and multi-spectral registration for natural images. In Proceedings of the European Conference on Computer Vision, pages 309-324. Springer, 2014. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 296, + 555 + ], + "type": "text", + "content": "[36] Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator. Advances in Neural Information Processing Systems, 36, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "type": "text", + "content": "[37] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. RFNet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19679-19688, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 669 + ], + "type": "text", + "content": "[38] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13117-13125, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[39] Jiacheng Ying, Hui-Liang Shen, and Si-Yuan Cao. Unaligned hyperspectral image fusion via registration and interpolation modeling. IEEE Transactions on Geoscience and Remote Sensing, 2021. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 396 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[40] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proceedings of the European Conference on Computer Vision, pages 653-669. Springer, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 195 + ], + "type": "text", + "content": "[41] Runmin Zhang, Jun Ma, Si-Yuan Cao, Lun Luo, Beinan Yu, Shu-Jie Chen, Junwei Li, and Hui-Liang Shen. SCP-Net: Unsupervised cross-modal homography estimation via intra-modal self-supervised learning. In Proceedings of the European Conference on Computer Vision, pages 460–477. Springer, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "text", + "content": "[42] Yiming Zhao, Xinming Huang, and Ziming Zhang. Deep lucas-kanade homography for multimodal image alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15950–15959, 2021. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 295 + ], + "type": "text", + "content": "[43] Huayi Zhou, Fei Jiang, and Hongtao Lu. SSDA-YOLO: Semi-supervised domain adaptive yolo for cross-domain object detection. Computer Vision and Image Understanding, 229: 103649, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "type": "text", + "content": "[44] Yuan Zhou, Anand Rangarajan, and Paul D Gader. An integrated approach to registration and fusion of hyperspectral and multispectral images. IEEE Transactions on Geoscience and Remote Sensing, 58(5):3020-3033, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 396 + ], + "type": "text", + "content": "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2223-2232, 2017. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16694" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_content_list.json b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44472e1129655cfc6b8468b77b1f498a4b1ad531 --- /dev/null +++ b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_content_list.json @@ -0,0 +1,1473 @@ +[ + { + "type": "text", + "text": "STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks", + "text_level": 1, + "bbox": [ + 135, + 128, + 859, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianqing Zhang $^{1*}$ Kairong Yu $^{1*}$ Xian Zhong $^{3}$ Hongwei Wang $^{1\\dagger}$ Qi Xu $^{2\\dagger}$ Qiang Zhang $^{2}$ $^{1}$ Zhejiang University $^{2}$ Dalian University of Technology $^{3}$ Wuhan University of Technology \n $^{1}\\{zhangtianqing, 22271086\\} @zju.edu.cn$ $^{1}$ hongweiwang@intl.zju.edu.cn \n $^{2}\\{xuqi, zhangq\\} @dlut.edu.cn$ $^{3}$ zhongx@whut.edu.cn", + "bbox": [ + 117, + 202, + 880, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) have gained significant attention due to their biological plausibility and energy efficiency, making them promising alternatives to Artificial Neural Networks (ANNs). However, the performance gap between SNNs and ANNs remains a substantial challenge hindering the widespread adoption of SNNs. In this paper, we propose a Spatial-Temporal Attention Aggregator SNN (STAA-SNN) framework, which dynamically focuses on and captures both spatial and temporal dependencies. First, we introduce a spike-driven self-attention mechanism specifically designed for SNNs. Additionally, we pioneeringly incorporate position encoding to integrate latent temporal relationships into the incoming features. For spatial-temporal information aggregation, we employ step attention to selectively amplify relevant features to variant steps. Finally, we implement a time-step random dropout strategy to avoid local optima. The framework demonstrates exceptional performance across diverse datasets and exhibits strong generalization capabilities. Notably, STAA-SNN achieves state-of-the-art results on neuromorphic datasets CIFAR10-DVS of $82.10\\%$ and with performances of $97.14\\%$ , $82.05\\%$ and $70.40\\%$ on the static datasets CIFAR-10, CIFAR-100 and ImageNet, respectively. Furthermore, this model exhibits improved performance ranging from $0.33\\%$ to $2.80\\%$ with fewer time steps.", + "bbox": [ + 88, + 340, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 747, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Spiking Neural Networks (SNNs) have attracted significant attention due to their low energy consumption and biological interpretability[12, 17, 51, 57-59]. Inspired by the behavior of biological neurons, SNNs represent information using discrete binary spikes over multiple timesteps, making them well-suited for implementation", + "bbox": [ + 89, + 773, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion on low-power neuromorphic hardware, offering a distinct advantage over traditional Artificial Neural Networks (ANNs)[2, 13, 19, 48, 63]. Among various spiking neuron models, the Leaky Integrate-and-Fire (LIF) model[1] is widely used in SNNs for its linear differentiable [37, 38]. Although previous works[17, 34, 73] have developed deep SNNs to accelerate model convergence, the multi-timestep computations inherent in SNNs significantly increase both training and inference times, leading to high latency. This limitation hinders the potential for efficient algorithm design. Consequently, a performance gap still exists between directly trained large-scale SNNs and ANNs in pattern recognition tasks, underscoring the need for novel approaches that integrate insights from neuroscience with conventional deep learning to improve both the efficiency and accuracy of SNNs. Inspired by the human visual and cognitive systems, attention mechanisms regulate neural activity and connectivity, enhancing responses to focal stimuli. This enhancement is reflected in the increased firing rates and improved synchronization of neurons in relevant brain regions. Additionally, this mechanism influences the strength and efficiency of neural connections, thereby facilitating the flow of information within and between different brain areas. In the field of deep learning, attention mechanisms enable neural networks to selectively focus on relatively important input information, leading to significant advancements in various domains such as natural language processing (NLP) and computer vision (CV). Compared to traditional architectures [5, 6, 44], these mechanisms have consistently demonstrated superior performance.", + "bbox": [ + 511, + 310, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the integration of attention mechanisms into deep SNNs. Although SNNs are biologically inspired models, their current implementations exhibit significant limitations due to inadequate biological fidelity. For instance, the heterogeneity among spiking neurons and the ability to independently transmit spatial and temporal information are overlooked, with membrane parameters being uniformly and permanently set across dimensions and layers in neural networks. Neuroscience research [3, 53] has", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding authors.", + "bbox": [ + 107, + 875, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13959", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "highlighted the critical role of synaptic homeostatic plasticity mechanisms in neuronal activity, as well as the regulatory function of glial cells in information transmission between neurons. Enhancing the alignment of SNN models with these biological processes has the potential to improve their overall performance. Therefore, we propose allowing different network layers to exhibit varying degrees of information transmission permeability.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a self-attention mechanism to capture spatial correlations and the significance of information, while integrating positional encoding to enhance spatial information and potentially learn temporal features. In traditional SNNs, input features are typically aggregated through simple addition, which can introduce significant noise into the synthesized features. To address this, we propose a Step Attention mechanism that aggregates and reconstructs features at each time step, thereby improving the quality of information propagation. Additionally, we introduce a novel Time-Step Random Dropout (TSRD) strategy to prevent the network from falling into local optima, thereby enhancing its robustness and accelerating training. Our main contributions are summarized as follows:", + "bbox": [ + 89, + 213, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Inspired by neuroscience and attention mechanisms, we propose a novel Spatio-Temporal Attention Aggregator for Spiking Neural Networks (STAA-SNN) that integrates self-attention, positional encoding, and step attention into a unified framework, enabling SNNs to efficiently extract features across temporal and spatial domains while maintaining stable training.", + "- We introduce a Time-Step Random Dropout (TSRD) strategy to accelerate training and enhance model generalization by preventing premature convergence and enabling optimal performance.", + "- We evaluate the performance, convergence, and sparse spiking activity of STAA-SNN across multiple benchmark datasets, demonstrating that our method achieves state-of-the-art or superior results with fewer timesteps, highlighting its efficiency and effectiveness." + ], + "bbox": [ + 89, + 426, + 478, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 684, + 238, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Training of Deep SNNs", + "text_level": 1, + "bbox": [ + 89, + 710, + 303, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training deep SNNs generally follows two primary approaches: ANN-to-SNN conversion and direct SNN training. The ANN-to-SNN conversion method aims to transfer knowledge from traditional ANNs to SNNs, but it often encounters challenges such as minimizing accuracy degradation while optimizing inference latency and energy efficiency[27, 28, 55]. Moreover, this approach struggles to fully leverage the spatiotemporal dynamics inherent in SNNs due to the lack of temporal encoding in ANNs. Direct SNN training is categorized into unsupervised and supervised learning. Unsupervised methods rely on biologi", + "bbox": [ + 89, + 734, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cally plausible learning rules such as Hebbian learning[30] and Spike-Timing-Dependent Plasticity (STDP)[7]. However, these approaches are generally not suitable for deep SNNs or large-scale datasets. On the other hand, supervised learning integrates backpropagation with surrogate gradients[17, 38, 46, 58], making applicability increasingly viable[60, 61]. Several methods have been proposed to enhance direct SNN training. For instance, MPBN[26] incorporates a batch normalization layer after membrane potential updates to stabilize training, while MPD-AL[72] introduces an alternative learning paradigm that adapts synaptic modifications based on membrane potential traces. MSResNet[34] reorganizes Vanilla ResNet layers to better align with SNNs, improving feature extraction and gradient propagation. Furthermore, IM-Loss[22] optimizes information maximization loss to increase spike information entropy. Recently, RSNN[62] introduced a Convolutional LSTM-based approach that preserves temporal correlations, enhancing spike-based data processing in SNNs.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Attention Mechanism in SNNs", + "text_level": 1, + "bbox": [ + 511, + 393, + 782, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention mechanisms dynamically assign varying levels of importance to different input components, thereby facilitating efficient processing and feature extraction. Since their introduction in 2014, attention mechanisms have become integral to enhancing deep learning models and have been widely applied to traditional ANNs[32]. In deep learning, attention mechanisms are primarily used in two ways: as a basic paradigm for conducting meta-operator such as self-attention [54] or as an auxiliary enhancement modules [33]. Also these two approaches can be combined. Recently, integrating attention mechanisms into SNNs has gained traction as a means to enhance their performance. Unlike traditional ANNs, where attention typically refines spatial feature representations, SNN-based attention mechanisms must consider both spatial and temporal information. Early work by [47] introduced a computational model that bridges human selective attention and conscious awareness. Kundu[36] utilize attention mechanisms to compress SNN models. Temporal-wise attention was first introduced in TA-SNN[64], laying the foundation for subsequent developments in SNN-specific attention mechanisms. MA-SNN [66] proposed temporal-channel-spatial attention for SNNs, surpassing models focused on single-dimensional information. TCJA-SNN[76] introduced a temporal-channel joint attention mechanism, while SCTFA [8] applied temporal-channel-spatial attention to SNNs with promising results. Moreover, STSC-SNN [71] incorporated temporal convolution and attention mechanisms to implement synaptic filtering and gating functions, demonstrating the versatility of attention mechanisms in enhancing SNN capabilities. IM-LIF [43] adopted temporalwise attention to improve the updating of GRUs and LSTMs", + "bbox": [ + 511, + 417, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13960", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/19976e25b16ba8dd63c522273cf82e127768f49b4e706e65db2435bd85bc3366.jpg", + "image_caption": [ + "(a) Spatial-Temporal Attention Aggregator SNN", + "(b) Structure of the Global Context (GC) block" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 321, + 323 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/313680c231220fa98d255eba78d4c8e446f53ed0c29275a20bb5902d186f6464.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 92, + 500, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4833dfb8c1ffb8fa787f4afbd2c2e593413ef88f869e9269a71e71c6c798aa92.jpg", + "image_caption": [ + "(c) Structure of Step Attention (SA) block", + "Figure 1. Overview of the STAA-SNN Architecture and the TSRD strategy." + ], + "image_footnote": [], + "bbox": [ + 321, + 215, + 496, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c09ac145f0551f349096536c3fdd48b04343d57e48af640729277fa3515a0591.jpg", + "image_caption": [ + "(d) STAA-SNN with TSRD strategy. The dropout probability is represented by $\\beta$ , meaning that at each time step $t$ , there is a probability $\\beta$ of selecting the simple addition aggregation method and discarding the STAA." + ], + "image_footnote": [], + "bbox": [ + 500, + 92, + 906, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "within SNNs. More recently, researchers have explored incorporating SNNs into Transformer architectures, achieving promising results on multiple benchmark datasets[41, 52]. These researches reveal that attention mechanisms can significantly improve the performance of SNN. However, there remains a need for more specialized and high-performance attention modules tailored explicitly for SNN architectures. Addressing this gap can further enhance the efficiency and robustness of SNNs in real-world applications.", + "bbox": [ + 89, + 392, + 483, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 542, + 225, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will first provide a brief overview of the Leaky Integrate-and-Fire (LIF) model and the training process. Following that, we introduce the proposed Spatial-Temporal Attention Aggregator. Lastly, we outline the timestep random dropout strategy for SNN.", + "bbox": [ + 89, + 566, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Leaky Integrate-and-Fire Model", + "text_level": 1, + "bbox": [ + 89, + 652, + 375, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The spiking neuron is a fundamental component of SNNs. In this work, we adopt the Leaky Integrate-and-Fire (LIF) neuron due to its efficiency and simplicity. The discrete-time, iterative update rule of the LIF model is given by:", + "bbox": [ + 89, + 674, + 482, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {t, n} = H ^ {t - 1, n} + \\frac {1}{\\tau} \\left[ I ^ {t - 1, n} - \\left(H ^ {t - 1, n} - V _ {\\text {r e s e t}}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 744, + 482, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {t, n} = \\Theta \\left(V ^ {t, n} - v _ {\\mathrm {t h}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 775, + 480, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {t, n} = V _ {\\text {r e s e t}} \\cdot S ^ {t, n} + V ^ {t, n} \\odot (1 - S ^ {t, n}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 796, + 480, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $H^{t - 1,n}$ denotes the membrane potential after a spike trigger at the previous timestep, while $I^{t,n}$ and $V^{t,n}$ represent the input and membrane potential of the $n$ -th layer at timestep $t$ , respectively. The threshold $v_{th}$ determines whether a spike is emitted, producing the binary spike sen", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sor $S^{t,n}$ . The Heaviside step function $\\Theta$ outputs 1 if $x \\geq 0$ and 0 otherwise.", + "bbox": [ + 511, + 392, + 905, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the similarity between iterative LIF dynamics and Recurrent Neural Networks (RNNs), Spatio-Temporal Backpropagation (STBP) [57, 58] enables direct SNN training via Backpropagation Through Time (BPTT):", + "bbox": [ + 511, + 422, + 905, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial W ^ {n}} = \\sum_ {t} \\left(\\frac {\\partial L}{\\partial S ^ {t , n}} \\cdot \\frac {\\partial S ^ {t , n}}{\\partial V ^ {t , n}} + \\frac {\\partial L}{\\partial V ^ {t + 1 , n}} \\cdot \\frac {\\partial V ^ {t + 1 , n}}{\\partial V ^ {t , n}}\\right) \\frac {\\partial V ^ {t , n}}{\\partial W ^ {n}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 501, + 903, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the term $\\frac{\\partial S^{t,n}}{\\partial V^{t,n}}$ , which represents the derivative of the spike function, is non-differentiable. To address this, prior works introduce Surrogate Gradients(SG), commonly approximated using a rectangular function:", + "bbox": [ + 511, + 547, + 905, + 609 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial S ^ {t}}{\\partial V ^ {t}} = \\frac {1}{a} \\cdot \\operatorname {s i g n} \\left(\\left| V ^ {t} - V _ {\\mathrm {t h}} \\right| < \\frac {a}{2}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 619, + 903, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $a$ is a hyperparameter and set to 1. This approximation provides a gradient of 1 within the range $V_{\\mathrm{th}} - 0.5 \\leq V^t \\leq V_{\\mathrm{th}} + 0.5$ , and otherwise 0.", + "bbox": [ + 511, + 652, + 905, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neuromorphic-Inspired Adaptive LIF Design. Examining Eq. (1), it is evident that keeping $V_{\\mathrm{reset}}$ and $\\tau$ constant results in fixed weight coefficients for $H$ and $I$ , which remain unchanged across timestep $t$ and layer $n$ . This simplification deviates from biological neural processing, where neuronal excitability dynamically adapts to stimuli. To bridge this gap, we propose an adaptive modification to the membrane potential update rule:", + "bbox": [ + 511, + 714, + 905, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {t, n} = M \\odot H ^ {t - 1, n} + N \\odot I ^ {t - 1, n}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 849, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $M$ and $N$ are coefficient matrices that dynamically adjust through iterative learning, regulating activation and", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13961", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "inhibition probabilities across layers. This biologically inspired adaptation enhances information processing efficiency and robustness in SNNs.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Context (GC) Block. The Global Context (GC) Block serves as a lightweight alternative to the self-attention mechanism used in Transformers, enabling the integration of self-attention into conventional convolution-based deep learning architectures without the computational overhead of full Transformer models. This makes it possible to enhance feature learning in non-Transformer structures while maintaining efficiency and precision. In this paper, we simulate the self-attention operation using a GC block [9], which consists of three $1 \\times 1$ convolutional kernels corresponding to key, query, and value. As depicted in Fig. 1(b), the spatial spiking input tensor of the $n$ -th layer at the $t$ -th timestep is defined as $X^{t,n} \\in \\mathbb{R}^{C^n \\times H^n \\times W^n}$ , where $W^n$ and $H^n$ denote the width and height of the feature map, and $C^n$ represents the number of channels. Passing $X^{t,n}$ through the GC block yields three weight matrices for self-attention, namely $W_k$ , $W_q$ and $W_v$ , defined as:", + "bbox": [ + 89, + 156, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {k} = X ^ {t, n} \\cdot \\operatorname {S i g m o i d} \\left(\\operatorname {C o n v} _ {k} \\left(X ^ {t, n}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 425, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {q} = \\operatorname {C o n v} _ {q} \\left(W _ {k}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 445, + 480, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {v} = \\operatorname {C o n v} _ {v} (\\operatorname {R e L U} (\\operatorname {L N} (W _ {q}))). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 463, + 480, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before computing $W_{k}$ , $X^{t,n}$ is reshaped from $C\\times H\\times W$ to $C\\times HW$ . The resulting matrices $W_{k},W_{q},W_{v}\\in \\mathbb{R}^{C\\times 1\\times 1}$ encode self-attention weights, which are then used to refine the input feature map. Subsequently, the enhanced output is computed as:", + "bbox": [ + 89, + 492, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ ^ {t, n} = X ^ {t, n} \\oplus W _ {v}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 566, + 480, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Z^{t,n} \\in \\mathbb{R}^{C,H,W}$ maintains the same dimensions as $X^{t,n}$ . The refined feature map $Z^{t,n}$ , in comparison to the $X^{t,n}$ , improves the network's ability to focus on self-relevant and task-critical information while suppressing irrelevant signals. This approach facilitates the seamless transmission of enriched data flow, enabling the retention of valuable features for subsequent stages.", + "bbox": [ + 89, + 590, + 483, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Encoding (PE) Block. Position encoding is widely used in NLP and sequence modeling to incorporate positional information into input data. While SNNs share structural similarities with RNNs and inherently process temporal information, their reliance on sparse binary spikes for information storage can lead to significant feature loss. This study introduces position encoding to enhance spatial features while enriching temporal relationships, thereby improving feature retention and strengthening temporal representations. The positional encoding is computed as follows:", + "bbox": [ + 89, + 718, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {t, n} = I ^ {t, n} \\oplus \\operatorname {P o s} _ {t, n}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 883, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Pos}_{t,n}$ represents the position encoding at the $t$ -th time step in the $n$ -th layer, with $\\mathrm{Pos}_{t,n} \\in \\mathbb{R}^{T,C^n}$ .", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step Attention (SA) Block. While previous modules primarily enhance feature processing from a spatial perspective, SNNs uniquely process information across both spatial and temporal. Unlike other types of network, SNNs offer an additional understanding of information sensitivity across different timesteps. Therefore, in this paper, we introduce Step Attention to tackle the temporal challenges inherent in SNNs. As illustrated in Fig. 1(c), let $U^{t,n} \\in \\mathbb{R}^{C^n \\times H^n \\times W^n}$ denote the spatially aggregated feature input at the $n$ -th layer and the $t$ -th timestep. The specific Step Attention mechanism is defined as:", + "bbox": [ + 511, + 138, + 906, + 305 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V ^ {t, n} = U ^ {t, n} \\odot \\text {S i g m o i d} \\left(\\operatorname {C o n v} _ {2} \\left(\\operatorname {R e L U} \\left(\\operatorname {C o n v} _ {1} \\right. \\right. \\right. \\tag {12} \\\\ \\left. \\left(\\alpha \\cdot \\operatorname {A v g P o o l} (U ^ {t, n})\\right)\\right) \\Bigg) \\Bigg), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 311, + 905, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling factor that regulates the initial training speed for smoother convergence. In practice, $\\alpha$ is typically set to a default value of 2.", + "bbox": [ + 511, + 388, + 905, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LIF with Spatio-Temporal Attention Aggregator (STAA). The STAA seamlessly integrates with the LIF model, offering compatibility across various network structures without constraints. It adeptly harnesses the potential of spiking neurons, effectively optimizing the storage of discrete spike information while minimizing the loss of crucial data during transmission. As shown in Fig. 1(a), the neuronal update equation for the LIF model with STAA is formulated as:", + "bbox": [ + 511, + 450, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU ^ {t, n} = G C _ {1} \\left(X ^ {t, n}\\right) \\oplus G C _ {2} \\left(H ^ {t - 1, n}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 595, + 903, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {t, n} = S A \\left(U ^ {t, n}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 616, + 903, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $GC_{1}$ and $GC_{2}$ represent Global Context Blocks for the input $X^{t,n}$ and the previous membrane state $H^{t-1,n}$ , respectively. The remaining update equations follow the original LIF model.", + "bbox": [ + 511, + 643, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Time Step Random Dropout (TSRD) Strategy", + "text_level": 1, + "bbox": [ + 511, + 712, + 898, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further enhance training efficiency, we propose TSRD, an adaptive training strategy designed to prevent premature feature stabilization in deep SNNs. In deeper timesteps, spiking features may solidify too early during iterations, hindering the network from achieving optimal convergence. As illustrated in Fig. 1(d), TSRD randomly dropout the augmentation module at timestep $t \\in (0,T)$ with a dropout probability $\\beta$ , and instead choose the most basic elementwise addition aggregation method. The TSRD technique accelerates training and helps the network escape local optima, thereby improving generalization performance.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13962", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4f65df8535f0380689328b8e5132b5d9bd6225061619d36e376f7ab6b5c24163.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArchitectureTimestepAccuracy (%)
CIFAR-10CIFAR-100
CNN-basedGLIF [70]ResNet-19294.44%75.48%
ResNet-19494.85%77.05%
TET [16]ResNet-19294.16%72.87%
ResNet-19494.44%74.47%
LSG [42]ResNet-19294.41%76.32%
ResNet-19495.17%76.85%
PFA [14]ResNet-19295.6%76.7%
ResNet-19495.71%78.1%
Diet-SNN [50]ResNet-20591.78%64.07%
VGG-16593.85%69.67%
IM-loss [22]VGG-16593.85%70.18%
IM-LIF [43]ResNet-19395.29%77.21%
MPBN [26]ResNet-19196.06%78.71%
ResNet-19296.47%79.51%
ResNet-19496.52%80.1%
ResNet-20192.22%68.41%
ResNet-20293.54%70.79%
ResNet-20494.28%72.3%
Transformer-basedSpikformer[75]Spikformer-4-25493.94%75.96%
Spikingformer[74]Spikingformer-4-25494.77%77.43%
Spike-driven Transformer [69]Transformer-2-512495.60%78.40%
CNN-basedOursVGG-13294.70%±0.12%75.16%±0.08%
VGG-13495.26%±0.09%76.35%±0.11%
ResNet-19196.75%±0.09%79.37%±0.10%
ResNet-19296.85%±0.11%80.57%±0.11%
ResNet-19497.14%±0.10%82.05%±0.10%
ResNet-20193.08%±0.09%70.14%±0.07%
ResNet-20294.35%±0.08%73.20%±0.12%
ResNet-20495.03%±0.11%75.10%±0.12%
", + "bbox": [ + 120, + 87, + 875, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison results with SOTA methods on CIFAR-10/100.", + "bbox": [ + 292, + 617, + 699, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 657, + 223, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first outlined the experimental setup, including dataset details and implementation specifics. We then compare our results with previous SOTA methods across multiple benchmarks. Next, we perform ablation studies to assess the effectiveness of STAA. Finally, we visualize the model's performance. Further details on datasets, hyperparameters, additional experiments, and computational efficiency analysis are provided in the Supplementary Material.", + "bbox": [ + 89, + 683, + 482, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 816, + 282, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on five benchmark datasets. CIFAR-10/100[35] are standard image classification datasets with 50,000 training and 10,000 test images across 10 and 100 classes, respectively. ImageNet[15] is a", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "large-scale dataset containing 1.2 million training, 50,000 validation, and 100,000 test images spanning 1,000 categories. For neuromorphic vision tasks, CIFAR10-DVS[39, 40, 45] consists of 10,000 event streams across 10 classes, with a 9:1 training-to-testing split. DVS128 Gesture[4] is an event-based gesture recognition dataset with 1,176 training and 288 test samples, featuring 11 gesture classes recorded from 29 subjects under three lighting conditions.", + "bbox": [ + 511, + 657, + 906, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we set the firing threshold $v_{\\mathrm{th}}$ to 1, and initialize $\\alpha$ and $\\beta$ to 2 and 0.1, respectively. In the PE block, we adopt a learnable position encoding method, initializing all values to zero. All experiments are implemented in PyTorch and conducted on an NVIDIA RTX 3090 GPU, except for ImageNet, which is", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13963", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3802872e9f7760e37689e04a0d3a4b7edd04d9f5581501ea48585c103e5cb6f7.jpg", + "image_caption": [ + "Figure 2. Position encoding locations in SNNs." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 255 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e304d8bc4eb444ab1f58fb82967f3803501ee63483aebd5bee4b45fa1f52c77d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArchitectureTimestepAccuracy
STBP-tdBN [73]ResNet34663.72%
TET [16]ResNet34664.79%
RecDis-SNN [23]ResNet34667.33%
GLIF [70]ResNet34467.52%
IM-Loss [22]ResNet18667.43%
Real Spike [24]ResNet18463.68%
ResNet34467.69%
RMP-Loss [25]ResNet18463.03%
ResNet34465.17%
MPBN [26]ResNet18463.14%
ResNet34464.71%
SEW ResNet [17]ResNet18463.18%
ResNet34467.04%
OursResNet18468.27% ± 0.19%
ResNet34470.40% ± 0.15%
", + "bbox": [ + 96, + 303, + 478, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "trained on 8 NVIDIA RTX 4090 GPUs. The total training epochs are 500 for CIFAR-10, CIFAR-100, and CIFAR10-DVS, while ImageNet runs for 300 epochs.", + "bbox": [ + 89, + 652, + 482, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 89, + 710, + 326, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Static Image Classification. We evaluate our model on three static datasets: CIFAR-10, CIFAR-100, and ImageNet. For CIFAR-10/100, STAA-SNNs are applied to both ResNet-like and VGG-like networks and tested at 1, 2, 3, and 4 timesteps. On ImageNet, the model is evaluated using ResNet-18 and ResNet-34 with 4 time steps. Our method achieves top accuracies of $97.14\\%$ on CIFAR-10, $82.05\\%$ on CIFAR-100, and $70.40\\%$ on ImageNet, surpassing previous state-of-the-art methods across various architectures. These improvements are attributed to the strong spatial-temporal attention mechanism and feature aggrega", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f25fa7be0e716880e0768fed6bacdf45ae6d3ad16849477e87d1a1803cf84be2.jpg", + "table_caption": [ + "Table 2. Comparison of training based SNN SOTA on ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodsArchitectureTimestepAccuracy
CIFAR10-DVS
IM-loss [22]ResNet-191072.60%
LSG [42]ResNet-191077.90%
MPBN [26]ResNet-191074.40%
MPBN [26]ResNet-201078.70%
TET [16]VGG5NN1077.30%
IM-LIF [43]VGG-131080.50%
GLIF [70]7B-wideNet1678.10%
STSA [56]STS-Transformer1679.93%
SEW [17]SEW-ResNet1674.4%
Spikeformer [41]Spikeformer1680.9%
OursResNet-201681.90%±0.20%
VGG-131682.10%±0.20%
DVS128 Gesture
STBP-tdBN [73]ResNet-174096.87%
SEW [17]7B-Net1697.52%
PLIF [18]PLIFNet2097.57%
MA-SNN [66]5 layers SCNN2098.23%
ASA-SNN [67]5 layers SCNN2097.70%
IM-LIF [43]ResNet-194097.33%
LIAF+TA [65]TA-SNN-Net6098.61%
Spike-driven-Transformer [69]Transformer-2-5121699.30%
OursVGG-131698.61%±0.20%
", + "bbox": [ + 516, + 303, + 903, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison results with SOTA methods on CIFAR10-DVS and DVS128 Gesture.", + "bbox": [ + 511, + 708, + 906, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion capabilities of STAA-SNNs, which fully leverage the potential performance of SNNs. Furthermore, our model achieves equivalent accuracy while requiring at least one fewer timestep compared to prior methods, improving efficiency without sacrificing performance.", + "bbox": [ + 511, + 766, + 906, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Event-based Action Recognition. To further assess the spatial-temporal processing capabilities of STAA-SNNs,", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13964", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8784705889bb5c7d12c031d70e074f6f6cdf83688940b7660fad31a5fd4c3cc9.jpg", + "image_caption": [ + "Figure 3. Distribution of accuracy with different dropout probability $\\beta$ in TSRD on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 460, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/164822d45e7592927d8c9f376171a1d403b8a9e630d5cc3cd5165d886620c70a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
STAA ComponentsTSRDAccuracyAccuracy Gain
GCPESA
XXXX72.30%-
✓XXX73.22%+0.92%
✓✓XX73.79%+1.49%
✓✓X✓73.96%+1.66%
✓✓✓X74.78%+2.48%
✓✓✓✓75.10%+2.80%
", + "bbox": [ + 96, + 287, + 478, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Combinations of sub-modules in STAA with TSRD.", + "bbox": [ + 104, + 431, + 467, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we conduct experiments on the DVS128 Gesture and CIFAR10-DVS datasets. Unlike static datasets, these neuromorphic datasets contain temporal event streams, requiring models to efficiently process spiking sequences. As shown in Tab. 3, our model achieves state-of-the-art or superior performance with a smaller network and fewer timesteps. Specifically, STAA-SNNs achieve $82.10\\%$ accuracy on CIFAR10-DVS and $98.61\\%$ on the DVS128 Gesture dataset, demonstrating their effectiveness in event-based action recognition tasks.", + "bbox": [ + 88, + 472, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 633, + 243, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All ablation studies with $\\mathrm{T} = 4$ using ResNet-20 as baseline.", + "bbox": [ + 89, + 655, + 478, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Position encoding placement. To optimize PE in SNNs, we evaluate its best placement within the network. Given the structural complexity of SNNs compared to CNNs, identifying the most effective integration point is crucial. As shown in Fig. 2, we evaluate four placement strategies (b)-(e) alongside a baseline SNN without PE (a), ensuring all other conditions remain constant. The CIFAR-10 accuracies for (b) to (e) are $94.08\\%$ , $94.29\\%$ , $94.27\\%$ , and $93.96\\%$ , respectively. The highest performance occurs when PE is placed at the pre-aggregation spatial input port (c). Although (b) and (d) yield similar results, (d) introduces additional parameters without significant gains. Thus, placing PE at the pre-aggregation spatial input port achieves the best balance between accuracy and efficiency.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of position encoding methods. We next evaluate different encoding methods to determine the most effective approach. The results on CIFAR-10 indicate that learnable PE achieves an accuracy of $94.99\\%$ , outperforming fixed PE, which attains $94.42\\%$ . This $0.57\\%$ improvement highlights the advantage of adaptive position encoding, as it enables the network to dynamically adjust spatial representations, leading to better feature alignment and overall performance.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of dropout probability $\\beta$ in TSRD. To accelerate training and enhance generalization in TSRD, it is essential to determine an optimal dropout probability $\\beta$ . Excessive dropout may lead to network collapse by discarding too much information, while insufficient dropout may limit its regularization effect. As shown in Fig. 3, by incrementally increasing $\\beta$ from 0.1 to 1.0, the model achieves relatively high performance when $\\beta \\leq 0.3$ , and achieving optimal at 0.1. These results confirm that excessive dropout removes essential information, degrading performance, while a balanced dropout probability effectively suppresses noise in deeper timesteps while preserving critical information in shallower steps. This enables the network to navigate critical learning stages more effectively, enhancing both stability and robustness.", + "bbox": [ + 511, + 244, + 908, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Combinations of sub-modules in STAA with TSRD. To evaluate the contribution of each sub-module in STAA, as well as the impact of TSRD, we conduct an ablation study by progressively integrating these components into the SNN. The results on CIFAR-100 are presented in Tab. 4. Notably, the GC block alone improves accuracy by $0.92\\%$ , while the SA block contributes an additional $1.14\\%$ , highlighting the importance of spatio-temporal attention and feature aggregation reconstruction in enhancing SNN performance. When all components, including TSRD, are incorporated, the accuracy reaches $75.10\\%$ , demonstrating a cumulative improvement of $2.80\\%$ over the baseline.", + "bbox": [ + 511, + 489, + 908, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Intermediate Dimension Scaling Coefficients in GC. The GC module incorporates multiple $1 \\times 1$ convolutions for feature transformation. Retaining the original channel size $C$ for the last two convolutions layers would significantly increase computational overhead. To maintain the module's efficiency while preserving performance, a scaling coefficient $r$ is introduced within the GC to compress feature dimensions. By applying this scaling mechanism, the parameter count is reduced from $(C \\cdot C)$ to $(2C \\cdot C / r)$ , making the module more computationally feasible. We conduct experiments to assess the impact of different $r$ values, as shown in Fig. 5. The results indicate that the model achieves optimal performance when $r = 4$ , effectively balancing computational efficiency and accuracy.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13965", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db4f89a7409725552287a7029ebf952e759f4ab8d4905b78a495427f0ac68c63.jpg", + "image_caption": [ + "Figure 4. Visualization on CIFAR10-DVS. Ten layers from VGG-13 in a shallow to deep manner." + ], + "image_footnote": [], + "bbox": [ + 151, + 92, + 848, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f98014b62ed6b38cfa5b66c7ef1f002197aabee5d4150248636ff226ecff292d.jpg", + "image_caption": [ + "Figure 5. Impact of different scaling coefficients $r$ for intermediate feature dimensions in the GC module on the CIFAR-10 dataset." + ], + "image_footnote": [], + "bbox": [ + 122, + 294, + 450, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization and Analysis", + "text_level": 1, + "bbox": [ + 89, + 497, + 330, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We aggregate 4D $(T,C,H,W)$ spiking maps into 2D $(H,W)$ and apply Grad-CAM to visualize feature extraction at different depths. As shown in Fig. 4, the left displays a sample from the horse class in CIFAR10-DVS, while the right compares vanilla LIF and STAA-LIF feature maps. STAA-LIF focuses more effectively on target-relevant features, capturing finer details like the horse's mouth, which the vanilla LIF model misses. This highlights the advantage of spatio-temporal attention in enhancing feature selectivity.", + "bbox": [ + 89, + 520, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Analysis of Computation Efficiency", + "text_level": 1, + "bbox": [ + 89, + 667, + 398, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the single-image inference energy cost using 45-nm technology. ANNs rely on multiplication-and-accumulation (MAC) operations, which consume $4.6\\mathrm{pJ}$ per operation, whereas SNNs primarily use accumulation (AC) operations at a significantly lower cost of $0.9\\mathrm{pJ}$ per operation, making SNNs inherently more energy-efficient [31, 49]. Following [11], we assess the efficiency of ResNet-20, which requires 0.10 billion ACs, 0.06 billion MACs, 0.87 billion FLOPs, and consists of 12.69 million parameters. Under these conditions, inference on CIFAR-100 (with a $32\\times 32$ resolution and $\\mathrm{T} = 4$ ) results in an energy consumption of $0.366\\mathrm{mJ}$ , demonstrating the energy-efficient nature of ResNet-20. Additional computational efficiency experiments are provided in the Supplementary Material.", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 292, + 633, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce STAA-SNN, a spatio-temporal attention-based feature aggregator, together with the TSRD training strategy. STAA-SNN leverages the inherent ability of SNNs to process spatio-temporal information, thereby enhancing network performance while reducing the number of required timesteps. Our approach achieves state-of-the-art results across multiple benchmarks. STAA-SNN is designed as a plug-and-play module that seamlessly integrates into convolution-based SNN architectures. It is composed of several synergistic modules that collectively enhance its capabilities. The GC module simulates the brain's ability to process information across a broad spatial range, involving multiple brain regions in coordination. By focusing on spatial feature relevance, it enables the model to capture global structures and key stimuli in its environment. The PE module enhances temporal encoding by incorporating time-step information into feature representations, mirroring the brain's ability to integrate temporal and spatial cues for improved understanding of time-dependent relationships. The SA module mimics the brain's dynamic attention mechanisms, adjusting feature importance across time steps by adaptively enhancing or suppressing information. This prioritization of critical temporal features strengthens decision-making and overall network efficiency. Together, these modules form a robust spatio-temporal feature enhancement system that improves cognitive processing, robustness, and overall model performance. Despite these promising results, future research should focus on reducing computational energy consumption to further improve efficiency, particularly for resource-constrained applications.", + "bbox": [ + 511, + 320, + 906, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 791, + 700, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230, 62271361), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology.", + "bbox": [ + 511, + 816, + 905, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13966", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Larry F Abbott. Lapicque's introduction of the integrate-and-fire model neuron (1907). *Brain research bulletin*, 50(5-6): 303–304, 1999. 1", + "[2] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1", + "[3] Pepe Alcamí and Alberto E Pereda. Beyond plasticity: the dynamic impact of electrical synapses on neural circuits. Nature Reviews Neuroscience, 20(5):253-271, 2019. 1", + "[4] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 1", + "[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lui, and Cordelia Schmid. Vivit: A video vision transformer. 2021. 1", + "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? 2021. 1", + "[7] Guo-qiang Bi and Mu-ming Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of neuroscience, 18(24):10464-10472, 1998. 2", + "[8] Wuque Cai, Hongze Sun, Rui Liu, Yan Cui, Jun Wang, Yang Xia, Dezhong Yao, and Daqing Guo. A Spatial-Channel-Temporal-Fused Attention for Spiking Neural Networks. IEEE Transactions on Neural Networks and Learning Systems, pages 1–15, 2023. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2", + "[9] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4", + "[10] Biswadeep Chakraborty, Xueyuan She, and Saibal Mukhopadhyay. A Fully Spiking Hybrid Neural Network for Energy-Efficient Object Detection. IEEE Transactions on Image Processing, 30:9014-9029, 2021. Conference Name: IEEE Transactions on Image Processing. 3", + "[11] Guangyao Chen, Peixi Peng, Guoqi Li, and Yonghong Tian. Training Full Spike Neural Networks via Auxiliary Accumulation Pathway, 2023. arXiv:2301.11929 [cs]. 8", + "[12] Dennis V Christensen, Regina Dittmann, Bernabe Linares-Barranco, Abu Sebastian, Manuel Le Gallo, Andrea Redaelli, Stefan Slesazeck, Thomas Mikolajick, Sabina Spiga, Stephan Menzel, et al. 2022 roadmap on neuromorphic computing and engineering. Neuromorphic Computing and Engineering, 2(2):022501, 2022. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. Ieee Micro, 38(1):82-99, 2018. 1", + "[14] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 5", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 5", + "[16] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.5, 6", + "[17] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 1, 2, 6", + "[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6", + "[19] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014. 1", + "[20] Banglei Guan and Ji Zhao. Relative pose estimation for multi-camera systems from point correspondences with scale ratio. In Proceedings of the 30th ACM International Conference on Multimedia, page 5036-5044, New York, NY, USA, 2022. Association for Computing Machinery. 1", + "[21] Banglei Guan, Ji Zhao, Zhang Li, Fang Sun, and Friedrich Fraundorfer. Minimal solutions for relative pose with a single affine correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[22] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems, 35:156-166, 2022. 2, 5, 6", + "[23] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks. In Proceedings of CVPR, pages 326-335, New Orleans, LA, USA, 2022. IEEE. 6", + "[24] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In European Conference on Computer Vision, pages 52-68. Springer, 2022. 6", + "[25] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. RMP" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13967", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Loss: Regularizing Membrane Potential Distribution for Spiking Neural Networks, 2023. arXiv:2308.06787 [cs]. 6", + "[26] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 2, 5, 6", + "[27] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2", + "[28] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 2", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[30] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2", + "[31] Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pages 10-14, 2014. ISSN: 2376-8606. 8, 3", + "[32] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[33] J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu. Squeeze-and-excitation networks. IEEE transactions on pattern analysis and machine intelligence, 42(8):2011-2023, 2020. 2", + "[34] Yifan Hu, Yujie Wu, Lei Deng, and Guoqi Li. Advancing residual learning towards powerful deep spiking neural networks. arXiv e-prints, pages arXiv-2112, 2021. 1, 2", + "[35] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 5, 1", + "[36] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2", + "[37] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Spiking neural networks and bio-inspired supervised deep learning: A survey. arXiv preprint arXiv:2307.16235, 2023. 1", + "[38] Chankyu Lee, Syed Shakib Sarwar, Priyadarshini Panda, Gopalakrishnan Srinivasan, and Kaushik Roy. Enabling spike-based backpropagation for training deep neural network architectures. Frontiers in neuroscience, page 119, 2020. 1, 2", + "[39] Taihang Lei, Banglei Guan, Minzu Liang, Zibin Liu, Jianbin Liu, Yang Shang, and Qifeng Yu. Motion measurements of" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "explosive shock waves based on an event camera. Optics Express, 32(9):15390-15409, 2024. 5", + "[40] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 1", + "[41] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: A novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3, 6", + "[42] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3002–3010, 2023. 5, 6", + "[43] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. IM-LIF: Improved Neuronal Dynamics With Attention Mechanism for Direct Training Deep Spiking Neural Network. IEEE Transactions on Emerging Topics in Computational Intelligence, pages 1-11, 2024. Conference Name: IEEE Transactions on Emerging Topics in Computational Intelligence. 2, 5, 6", + "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021. 1", + "[45] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and tracking with an event camera. IEEE Transactions on Image Processing, 2024. 5", + "[46] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 2", + "[47] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2", + "[48] Garrick Orchard, E Paxon Frady, Daniel Ben Dayan Rubin, Sophia Sanborn, Sumit Bam Shrestha, Friedrich T Sommer, and Mike Davies. Efficient neuromorphic signal processing with loihi 2. In 2021 IEEE Workshop on Signal Processing Systems (SiPS), pages 254-259. IEEE, 2021. 1", + "[49] Ning Qiao, Hesham Mostafa, Federico Corradi, Marc Osswald, Fabio Stefanini, Dora Sumislawka, and Giacomo Indiveri. A reconfigurable on-line learning spiking neuromorphic processor comprising 256 neurons and 128K synapses. Frontiers in Neuroscience, 9, 2015. Publisher: Frontiers. 8, 3", + "[50] Nitin Rathi and Kaushik Roy. Diet-snn: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 5", + "[51] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13968", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Chen She and Laiyun Qing. Spikeformer: Image reconstruction from the sequence of spike camera based on transformer. In Proceedings of the 2022 5th International Conference on Image and Graphics Processing, pages 72-78, 2022. 3", + "[53] Gina G Turrigiano and Sacha B Nelson. Homeostatic plasticity in the developing nervous system. Nature reviews neuroscience, 5(2):97-107, 2004. 1", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv, 2017. 2", + "[55] Bingsen Wang, Jian Cao, Jue Chen, Shuo Feng, and Yuan Wang. A new ann-snn conversion method with high accuracy, low latency and good robustness. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3067-3075, 2023. 2", + "[56] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-Temporal Self-Attention for Asynchronous Spiking Neural Networks. In Proceedings of IJCAI, pages 3085–3093, Macau, SAR China, 2023. 6", + "[57] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 3", + "[58] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2, 3", + "[59] Yujie Wu, Rong Zhao, Jun Zhu, Feng Chen, Mingkun Xu, Guoqi Li, Sen Song, Lei Deng, Guanrui Wang, Hao Zheng, et al. Brain-inspired global-local learning incorporated with neuromorphic computing. Nature Communications, 13(1): 65, 2022. 1", + "[60] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 2", + "[61] Qi Xu, Jie Deng, Jiangrong Shen, Biwu Chen, Huajin Tang, and Gang Pan. Hybrid Spiking Vision Transformer for Event-Based Object Detection, 2024. 2", + "[62] Qi Xu, Xuanye Fang, Yaxin Li, Jiangrong Shen, De Ma, Yi Xu, and Gang Pan. Rsnn: Recurrent spiking neural networks for dynamic spatial-temporal information processing. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10602-10610, 2024. 2", + "[63] Kashu Yamazaki, Viet-Khoa Vo-Ho, Darshan Bulsara, and Ngan Le. Spiking neural networks and their applications: A review. *Brain Sciences*, 12(7):863, 2022. 1", + "[64] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2", + "[65] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-Wise Attention Spiking Neural Networks for Event Streams Classification. In Proceedings of ICCV, pages 10221-10230, 2021. 6" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 6", + "[67] Man Yao, Jiakui Hu, Guangshe Zhao, Yaoyuan Wang, Ziyang Zhang, Bo Xu, and Guoqi Li. Inherent Redundancy in Spiking Neural Networks. In Proceedings of the ICCV, pages 16924-16934, 2023. 6", + "[68] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention Spiking Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9393-9410, 2023. 3", + "[69] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 5, 6", + "[70] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5, 6", + "[71] Chengting Yu, Zheming Gu, Da Li, Gaoang Wang, Aili Wang, and Erping Li. STSC-SNN: Spatio-Temporal Synaptic Connection with temporal convolution and attention for spiking neural networks. Frontiers in Neuroscience, 16, 2022. 2", + "[72] Malu Zhang, Jibin Wu, Yansong Chua, Xiaoling Luo, Zihan Pan, Dan Liu, and Haizhou Li. Mpd-al: an efficient membrane potential driven aggregate-label learning algorithm for spiking neurons. In Proceedings of the AAAI conference on artificial intelligence, pages 1327-1334, 2019. 2", + "[73] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6", + "[74] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5", + "[75] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 5", + "[76] Rui-Jie Zhu, Malu Zhang, Qihang Zhao, Haoyu Deng, Yule Duan, and Liang-Jian Deng. Tcja-snn: Temporal-channel joint attention for spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 741 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13969", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_model.json b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bfb505f81b9d1b461ade058b3183cb5825b5a62e --- /dev/null +++ b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.861, + 0.175 + ], + "angle": 0, + "content": "STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.203, + 0.882, + 0.277 + ], + "angle": 0, + "content": "Tianqing Zhang\\(^{1*}\\) Kairong Yu\\(^{1*}\\) Xian Zhong\\(^{3}\\) Hongwei Wang\\(^{1\\dagger}\\) Qi Xu\\(^{2\\dagger}\\) Qiang Zhang\\(^{2}\\) \n\\(^{1}\\)Zhejiang University \\(^{2}\\)Dalian University of Technology \\(^{3}\\)Wuhan University of Technology \n\\(^{1}\\{zhangtianqing, 22271086\\} @zju.edu.cn\\) \\(^{1}\\)hongweiwang@intl.zju.edu.cn \n\\(^{2}\\{xuqi, zhangq\\} @dlut.edu.cn\\) \\(^{3}\\)zhongx@whut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.721 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) have gained significant attention due to their biological plausibility and energy efficiency, making them promising alternatives to Artificial Neural Networks (ANNs). However, the performance gap between SNNs and ANNs remains a substantial challenge hindering the widespread adoption of SNNs. In this paper, we propose a Spatial-Temporal Attention Aggregator SNN (STAA-SNN) framework, which dynamically focuses on and captures both spatial and temporal dependencies. First, we introduce a spike-driven self-attention mechanism specifically designed for SNNs. Additionally, we pioneeringly incorporate position encoding to integrate latent temporal relationships into the incoming features. For spatial-temporal information aggregation, we employ step attention to selectively amplify relevant features to variant steps. Finally, we implement a time-step random dropout strategy to avoid local optima. The framework demonstrates exceptional performance across diverse datasets and exhibits strong generalization capabilities. Notably, STAA-SNN achieves state-of-the-art results on neuromorphic datasets CIFAR10-DVS of \\(82.10\\%\\) and with performances of \\(97.14\\%\\), \\(82.05\\%\\) and \\(70.40\\%\\) on the static datasets CIFAR-10, CIFAR-100 and ImageNet, respectively. Furthermore, this model exhibits improved performance ranging from \\(0.33\\%\\) to \\(2.80\\%\\) with fewer time steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.485, + 0.868 + ], + "angle": 0, + "content": "In recent years, Spiking Neural Networks (SNNs) have attracted significant attention due to their low energy consumption and biological interpretability[12, 17, 51, 57-59]. Inspired by the behavior of biological neurons, SNNs represent information using discrete binary spikes over multiple timesteps, making them well-suited for implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.908, + 0.763 + ], + "angle": 0, + "content": "tion on low-power neuromorphic hardware, offering a distinct advantage over traditional Artificial Neural Networks (ANNs)[2, 13, 19, 48, 63]. Among various spiking neuron models, the Leaky Integrate-and-Fire (LIF) model[1] is widely used in SNNs for its linear differentiable [37, 38]. Although previous works[17, 34, 73] have developed deep SNNs to accelerate model convergence, the multi-timestep computations inherent in SNNs significantly increase both training and inference times, leading to high latency. This limitation hinders the potential for efficient algorithm design. Consequently, a performance gap still exists between directly trained large-scale SNNs and ANNs in pattern recognition tasks, underscoring the need for novel approaches that integrate insights from neuroscience with conventional deep learning to improve both the efficiency and accuracy of SNNs. Inspired by the human visual and cognitive systems, attention mechanisms regulate neural activity and connectivity, enhancing responses to focal stimuli. This enhancement is reflected in the increased firing rates and improved synchronization of neurons in relevant brain regions. Additionally, this mechanism influences the strength and efficiency of neural connections, thereby facilitating the flow of information within and between different brain areas. In the field of deep learning, attention mechanisms enable neural networks to selectively focus on relatively important input information, leading to significant advancements in various domains such as natural language processing (NLP) and computer vision (CV). Compared to traditional architectures [5, 6, 44], these mechanisms have consistently demonstrated superior performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we explore the integration of attention mechanisms into deep SNNs. Although SNNs are biologically inspired models, their current implementations exhibit significant limitations due to inadequate biological fidelity. For instance, the heterogeneity among spiking neurons and the ability to independently transmit spatial and temporal information are overlooked, with membrane parameters being uniformly and permanently set across dimensions and layers in neural networks. Neuroscience research [3, 53] has" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.245, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13959" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "highlighted the critical role of synaptic homeostatic plasticity mechanisms in neuronal activity, as well as the regulatory function of glial cells in information transmission between neurons. Enhancing the alignment of SNN models with these biological processes has the potential to improve their overall performance. Therefore, we propose allowing different network layers to exhibit varying degrees of information transmission permeability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.482, + 0.424 + ], + "angle": 0, + "content": "We introduce a self-attention mechanism to capture spatial correlations and the significance of information, while integrating positional encoding to enhance spatial information and potentially learn temporal features. In traditional SNNs, input features are typically aggregated through simple addition, which can introduce significant noise into the synthesized features. To address this, we propose a Step Attention mechanism that aggregates and reconstructs features at each time step, thereby improving the quality of information propagation. Additionally, we introduce a novel Time-Step Random Dropout (TSRD) strategy to prevent the network from falling into local optima, thereby enhancing its robustness and accelerating training. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.427, + 0.48, + 0.531 + ], + "angle": 0, + "content": "- Inspired by neuroscience and attention mechanisms, we propose a novel Spatio-Temporal Attention Aggregator for Spiking Neural Networks (STAA-SNN) that integrates self-attention, positional encoding, and step attention into a unified framework, enabling SNNs to efficiently extract features across temporal and spatial domains while maintaining stable training." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.533, + 0.48, + 0.591 + ], + "angle": 0, + "content": "- We introduce a Time-Step Random Dropout (TSRD) strategy to accelerate training and enhance model generalization by preventing premature convergence and enabling optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.593, + 0.48, + 0.667 + ], + "angle": 0, + "content": "- We evaluate the performance, convergence, and sparse spiking activity of STAA-SNN across multiple benchmark datasets, demonstrating that our method achieves state-of-the-art or superior results with fewer timesteps, highlighting its efficiency and effectiveness." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.427, + 0.48, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.24, + 0.701 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.304, + 0.728 + ], + "angle": 0, + "content": "2.1. Training of Deep SNNs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Training deep SNNs generally follows two primary approaches: ANN-to-SNN conversion and direct SNN training. The ANN-to-SNN conversion method aims to transfer knowledge from traditional ANNs to SNNs, but it often encounters challenges such as minimizing accuracy degradation while optimizing inference latency and energy efficiency[27, 28, 55]. Moreover, this approach struggles to fully leverage the spatiotemporal dynamics inherent in SNNs due to the lack of temporal encoding in ANNs. Direct SNN training is categorized into unsupervised and supervised learning. Unsupervised methods rely on biologi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "cally plausible learning rules such as Hebbian learning[30] and Spike-Timing-Dependent Plasticity (STDP)[7]. However, these approaches are generally not suitable for deep SNNs or large-scale datasets. On the other hand, supervised learning integrates backpropagation with surrogate gradients[17, 38, 46, 58], making applicability increasingly viable[60, 61]. Several methods have been proposed to enhance direct SNN training. For instance, MPBN[26] incorporates a batch normalization layer after membrane potential updates to stabilize training, while MPD-AL[72] introduces an alternative learning paradigm that adapts synaptic modifications based on membrane potential traces. MSResNet[34] reorganizes Vanilla ResNet layers to better align with SNNs, improving feature extraction and gradient propagation. Furthermore, IM-Loss[22] optimizes information maximization loss to increase spike information entropy. Recently, RSNN[62] introduced a Convolutional LSTM-based approach that preserves temporal correlations, enhancing spike-based data processing in SNNs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.783, + 0.409 + ], + "angle": 0, + "content": "2.2. Attention Mechanism in SNNs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Attention mechanisms dynamically assign varying levels of importance to different input components, thereby facilitating efficient processing and feature extraction. Since their introduction in 2014, attention mechanisms have become integral to enhancing deep learning models and have been widely applied to traditional ANNs[32]. In deep learning, attention mechanisms are primarily used in two ways: as a basic paradigm for conducting meta-operator such as self-attention [54] or as an auxiliary enhancement modules [33]. Also these two approaches can be combined. Recently, integrating attention mechanisms into SNNs has gained traction as a means to enhance their performance. Unlike traditional ANNs, where attention typically refines spatial feature representations, SNN-based attention mechanisms must consider both spatial and temporal information. Early work by [47] introduced a computational model that bridges human selective attention and conscious awareness. Kundu[36] utilize attention mechanisms to compress SNN models. Temporal-wise attention was first introduced in TA-SNN[64], laying the foundation for subsequent developments in SNN-specific attention mechanisms. MA-SNN [66] proposed temporal-channel-spatial attention for SNNs, surpassing models focused on single-dimensional information. TCJA-SNN[76] introduced a temporal-channel joint attention mechanism, while SCTFA [8] applied temporal-channel-spatial attention to SNNs with promising results. Moreover, STSC-SNN [71] incorporated temporal convolution and attention mechanisms to implement synaptic filtering and gating functions, demonstrating the versatility of attention mechanisms in enhancing SNN capabilities. IM-LIF [43] adopted temporalwise attention to improve the updating of GRUs and LSTMs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.322, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.328, + 0.302, + 0.337 + ], + "angle": 0, + "content": "(a) Spatial-Temporal Attention Aggregator SNN" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.093, + 0.5, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.206, + 0.5, + 0.215 + ], + "angle": 0, + "content": "(b) Structure of the Global Context (GC) block" + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.217, + 0.498, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.328, + 0.488, + 0.337 + ], + "angle": 0, + "content": "(c) Structure of Step Attention (SA) block" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.093, + 0.908, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.311, + 0.881, + 0.337 + ], + "angle": 0, + "content": "(d) STAA-SNN with TSRD strategy. The dropout probability is represented by \\(\\beta\\), meaning that at each time step \\(t\\), there is a probability \\(\\beta\\) of selecting the simple addition aggregation method and discarding the STAA." + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.352, + 0.722, + 0.367 + ], + "angle": 0, + "content": "Figure 1. Overview of the STAA-SNN Architecture and the TSRD strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.528 + ], + "angle": 0, + "content": "within SNNs. More recently, researchers have explored incorporating SNNs into Transformer architectures, achieving promising results on multiple benchmark datasets[41, 52]. These researches reveal that attention mechanisms can significantly improve the performance of SNN. However, there remains a need for more specialized and high-performance attention modules tailored explicitly for SNN architectures. Addressing this gap can further enhance the efficiency and robustness of SNNs in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.226, + 0.56 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.643 + ], + "angle": 0, + "content": "In this section, we will first provide a brief overview of the Leaky Integrate-and-Fire (LIF) model and the training process. Following that, we introduce the proposed Spatial-Temporal Attention Aggregator. Lastly, we outline the timestep random dropout strategy for SNN." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.376, + 0.669 + ], + "angle": 0, + "content": "3.1. Leaky Integrate-and-Fire Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.737 + ], + "angle": 0, + "content": "The spiking neuron is a fundamental component of SNNs. In this work, we adopt the Leaky Integrate-and-Fire (LIF) neuron due to its efficiency and simplicity. The discrete-time, iterative update rule of the LIF model is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.746, + 0.483, + 0.774 + ], + "angle": 0, + "content": "\\[\nV ^ {t, n} = H ^ {t - 1, n} + \\frac {1}{\\tau} \\left[ I ^ {t - 1, n} - \\left(H ^ {t - 1, n} - V _ {\\text {r e s e t}}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.776, + 0.482, + 0.793 + ], + "angle": 0, + "content": "\\[\nS ^ {t, n} = \\Theta \\left(V ^ {t, n} - v _ {\\mathrm {t h}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.797, + 0.482, + 0.814 + ], + "angle": 0, + "content": "\\[\nH ^ {t, n} = V _ {\\text {r e s e t}} \\cdot S ^ {t, n} + V ^ {t, n} \\odot (1 - S ^ {t, n}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Here, \\( H^{t - 1,n} \\) denotes the membrane potential after a spike trigger at the previous timestep, while \\( I^{t,n} \\) and \\( V^{t,n} \\) represent the input and membrane potential of the \\( n \\)-th layer at timestep \\( t \\), respectively. The threshold \\( v_{th} \\) determines whether a spike is emitted, producing the binary spike sen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.906, + 0.422 + ], + "angle": 0, + "content": "sor \\( S^{t,n} \\). The Heaviside step function \\( \\Theta \\) outputs 1 if \\( x \\geq 0 \\) and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.906, + 0.485 + ], + "angle": 0, + "content": "Given the similarity between iterative LIF dynamics and Recurrent Neural Networks (RNNs), Spatio-Temporal Backpropagation (STBP) [57, 58] enables direct SNN training via Backpropagation Through Time (BPTT):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.502, + 0.905, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial W ^ {n}} = \\sum_ {t} \\left(\\frac {\\partial L}{\\partial S ^ {t , n}} \\cdot \\frac {\\partial S ^ {t , n}}{\\partial V ^ {t , n}} + \\frac {\\partial L}{\\partial V ^ {t + 1 , n}} \\cdot \\frac {\\partial V ^ {t + 1 , n}}{\\partial V ^ {t , n}}\\right) \\frac {\\partial V ^ {t , n}}{\\partial W ^ {n}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.61 + ], + "angle": 0, + "content": "However, the term \\(\\frac{\\partial S^{t,n}}{\\partial V^{t,n}}\\), which represents the derivative of the spike function, is non-differentiable. To address this, prior works introduce Surrogate Gradients(SG), commonly approximated using a rectangular function:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.62, + 0.905, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial S ^ {t}}{\\partial V ^ {t}} = \\frac {1}{a} \\cdot \\operatorname {s i g n} \\left(\\left| V ^ {t} - V _ {\\mathrm {t h}} \\right| < \\frac {a}{2}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.906, + 0.697 + ], + "angle": 0, + "content": "where \\(a\\) is a hyperparameter and set to 1. This approximation provides a gradient of 1 within the range \\(V_{\\mathrm{th}} - 0.5 \\leq V^t \\leq V_{\\mathrm{th}} + 0.5\\), and otherwise 0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.836 + ], + "angle": 0, + "content": "Neuromorphic-Inspired Adaptive LIF Design. Examining Eq. (1), it is evident that keeping \\( V_{\\mathrm{reset}} \\) and \\( \\tau \\) constant results in fixed weight coefficients for \\( H \\) and \\( I \\), which remain unchanged across timestep \\( t \\) and layer \\( n \\). This simplification deviates from biological neural processing, where neuronal excitability dynamically adapts to stimuli. To bridge this gap, we propose an adaptive modification to the membrane potential update rule:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.85, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\nV ^ {t, n} = M \\odot H ^ {t - 1, n} + N \\odot I ^ {t - 1, n}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Here, \\(M\\) and \\(N\\) are coefficient matrices that dynamically adjust through iterative learning, regulating activation and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13961" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "inhibition probabilities across layers. This biologically inspired adaptation enhances information processing efficiency and robustness in SNNs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.157, + 0.484, + 0.415 + ], + "angle": 0, + "content": "Global Context (GC) Block. The Global Context (GC) Block serves as a lightweight alternative to the self-attention mechanism used in Transformers, enabling the integration of self-attention into conventional convolution-based deep learning architectures without the computational overhead of full Transformer models. This makes it possible to enhance feature learning in non-Transformer structures while maintaining efficiency and precision. In this paper, we simulate the self-attention operation using a GC block [9], which consists of three \\(1 \\times 1\\) convolutional kernels corresponding to key, query, and value. As depicted in Fig. 1(b), the spatial spiking input tensor of the \\(n\\)-th layer at the \\(t\\)-th timestep is defined as \\(X^{t,n} \\in \\mathbb{R}^{C^n \\times H^n \\times W^n}\\), where \\(W^n\\) and \\(H^n\\) denote the width and height of the feature map, and \\(C^n\\) represents the number of channels. Passing \\(X^{t,n}\\) through the GC block yields three weight matrices for self-attention, namely \\(W_k\\), \\(W_q\\) and \\(W_v\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.426, + 0.483, + 0.443 + ], + "angle": 0, + "content": "\\[\nW _ {k} = X ^ {t, n} \\cdot \\operatorname {S i g m o i d} \\left(\\operatorname {C o n v} _ {k} \\left(X ^ {t, n}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.446, + 0.482, + 0.462 + ], + "angle": 0, + "content": "\\[\nW _ {q} = \\operatorname {C o n v} _ {q} \\left(W _ {k}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.464, + 0.482, + 0.481 + ], + "angle": 0, + "content": "\\[\nW _ {v} = \\operatorname {C o n v} _ {v} (\\operatorname {R e L U} (\\operatorname {L N} (W _ {q}))). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Before computing \\( W_{k} \\), \\( X^{t,n} \\) is reshaped from \\( C\\times H\\times W \\) to \\( C\\times HW \\). The resulting matrices \\( W_{k},W_{q},W_{v}\\in \\mathbb{R}^{C\\times 1\\times 1} \\) encode self-attention weights, which are then used to refine the input feature map. Subsequently, the enhanced output is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.568, + 0.482, + 0.585 + ], + "angle": 0, + "content": "\\[\nZ ^ {t, n} = X ^ {t, n} \\oplus W _ {v}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.7 + ], + "angle": 0, + "content": "where \\( Z^{t,n} \\in \\mathbb{R}^{C,H,W} \\) maintains the same dimensions as \\( X^{t,n} \\). The refined feature map \\( Z^{t,n} \\), in comparison to the \\( X^{t,n} \\), improves the network's ability to focus on self-relevant and task-critical information while suppressing irrelevant signals. This approach facilitates the seamless transmission of enriched data flow, enabling the retention of valuable features for subsequent stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Position Encoding (PE) Block. Position encoding is widely used in NLP and sequence modeling to incorporate positional information into input data. While SNNs share structural similarities with RNNs and inherently process temporal information, their reliance on sparse binary spikes for information storage can lead to significant feature loss. This study introduces position encoding to enhance spatial features while enriching temporal relationships, thereby improving feature retention and strengthening temporal representations. The positional encoding is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nX ^ {t, n} = I ^ {t, n} \\oplus \\operatorname {P o s} _ {t, n}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "where \\(\\mathrm{Pos}_{t,n}\\) represents the position encoding at the \\(t\\)-th time step in the \\(n\\)-th layer, with \\(\\mathrm{Pos}_{t,n} \\in \\mathbb{R}^{T,C^n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.14, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Step Attention (SA) Block. While previous modules primarily enhance feature processing from a spatial perspective, SNNs uniquely process information across both spatial and temporal. Unlike other types of network, SNNs offer an additional understanding of information sensitivity across different timesteps. Therefore, in this paper, we introduce Step Attention to tackle the temporal challenges inherent in SNNs. As illustrated in Fig. 1(c), let \\( U^{t,n} \\in \\mathbb{R}^{C^n \\times H^n \\times W^n} \\) denote the spatially aggregated feature input at the \\( n \\)-th layer and the \\( t \\)-th timestep. The specific Step Attention mechanism is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.312, + 0.906, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V ^ {t, n} = U ^ {t, n} \\odot \\text {S i g m o i d} \\left(\\operatorname {C o n v} _ {2} \\left(\\operatorname {R e L U} \\left(\\operatorname {C o n v} _ {1} \\right. \\right. \\right. \\tag {12} \\\\ \\left. \\left(\\alpha \\cdot \\operatorname {A v g P o o l} (U ^ {t, n})\\right)\\right) \\Bigg) \\Bigg), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.389, + 0.906, + 0.433 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling factor that regulates the initial training speed for smoother convergence. In practice, \\(\\alpha\\) is typically set to a default value of 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.452, + 0.907, + 0.587 + ], + "angle": 0, + "content": "LIF with Spatio-Temporal Attention Aggregator (STAA). The STAA seamlessly integrates with the LIF model, offering compatibility across various network structures without constraints. It adeptly harnesses the potential of spiking neurons, effectively optimizing the storage of discrete spike information while minimizing the loss of crucial data during transmission. As shown in Fig. 1(a), the neuronal update equation for the LIF model with STAA is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.597, + 0.905, + 0.614 + ], + "angle": 0, + "content": "\\[\nU ^ {t, n} = G C _ {1} \\left(X ^ {t, n}\\right) \\oplus G C _ {2} \\left(H ^ {t - 1, n}\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.617, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\nV ^ {t, n} = S A \\left(U ^ {t, n}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Where \\(GC_{1}\\) and \\(GC_{2}\\) represent Global Context Blocks for the input \\(X^{t,n}\\) and the previous membrane state \\(H^{t-1,n}\\), respectively. The remaining update equations follow the original LIF model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.9, + 0.73 + ], + "angle": 0, + "content": "3.2. Time Step Random Dropout (TSRD) Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To further enhance training efficiency, we propose TSRD, an adaptive training strategy designed to prevent premature feature stabilization in deep SNNs. In deeper timesteps, spiking features may solidify too early during iterations, hindering the network from achieving optimal convergence. As illustrated in Fig. 1(d), TSRD randomly dropout the augmentation module at timestep \\(t \\in (0,T)\\) with a dropout probability \\(\\beta\\), and instead choose the most basic elementwise addition aggregation method. The TSRD technique accelerates training and helps the network escape local optima, thereby improving generalization performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13962" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.088, + 0.877, + 0.609 + ], + "angle": 0, + "content": "
MethodArchitectureTimestepAccuracy (%)
CIFAR-10CIFAR-100
CNN-basedGLIF [70]ResNet-19294.44%75.48%
ResNet-19494.85%77.05%
TET [16]ResNet-19294.16%72.87%
ResNet-19494.44%74.47%
LSG [42]ResNet-19294.41%76.32%
ResNet-19495.17%76.85%
PFA [14]ResNet-19295.6%76.7%
ResNet-19495.71%78.1%
Diet-SNN [50]ResNet-20591.78%64.07%
VGG-16593.85%69.67%
IM-loss [22]VGG-16593.85%70.18%
IM-LIF [43]ResNet-19395.29%77.21%
MPBN [26]ResNet-19196.06%78.71%
ResNet-19296.47%79.51%
ResNet-19496.52%80.1%
ResNet-20192.22%68.41%
ResNet-20293.54%70.79%
ResNet-20494.28%72.3%
Transformer-basedSpikformer[75]Spikformer-4-25493.94%75.96%
Spikingformer[74]Spikingformer-4-25494.77%77.43%
Spike-driven Transformer [69]Transformer-2-512495.60%78.40%
CNN-basedOursVGG-13294.70%±0.12%75.16%±0.08%
VGG-13495.26%±0.09%76.35%±0.11%
ResNet-19196.75%±0.09%79.37%±0.10%
ResNet-19296.85%±0.11%80.57%±0.11%
ResNet-19497.14%±0.10%82.05%±0.10%
ResNet-20193.08%±0.09%70.14%±0.07%
ResNet-20294.35%±0.08%73.20%±0.12%
ResNet-20495.03%±0.11%75.10%±0.12%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.294, + 0.618, + 0.7, + 0.632 + ], + "angle": 0, + "content": "Table 1. Comparison results with SOTA methods on CIFAR-10/100." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.224, + 0.675 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.806 + ], + "angle": 0, + "content": "We first outlined the experimental setup, including dataset details and implementation specifics. We then compare our results with previous SOTA methods across multiple benchmarks. Next, we perform ablation studies to assess the effectiveness of STAA. Finally, we visualize the model's performance. Further details on datasets, hyperparameters, additional experiments, and computational efficiency analysis are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.283, + 0.835 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on five benchmark datasets. CIFAR-10/100[35] are standard image classification datasets with 50,000 training and 10,000 test images across 10 and 100 classes, respectively. ImageNet[15] is a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.781 + ], + "angle": 0, + "content": "large-scale dataset containing 1.2 million training, 50,000 validation, and 100,000 test images spanning 1,000 categories. For neuromorphic vision tasks, CIFAR10-DVS[39, 40, 45] consists of 10,000 event streams across 10 classes, with a 9:1 training-to-testing split. DVS128 Gesture[4] is an event-based gesture recognition dataset with 1,176 training and 288 test samples, featuring 11 gesture classes recorded from 29 subjects under three lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we set the firing threshold \\( v_{\\mathrm{th}} \\) to 1, and initialize \\( \\alpha \\) and \\( \\beta \\) to 2 and 0.1, respectively. In the PE block, we adopt a learnable position encoding method, initializing all values to zero. All experiments are implemented in PyTorch and conducted on an NVIDIA RTX 3090 GPU, except for ImageNet, which is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13963" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.912, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.266, + 0.642, + 0.281 + ], + "angle": 0, + "content": "Figure 2. Position encoding locations in SNNs." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.304, + 0.48, + 0.6 + ], + "angle": 0, + "content": "
MethodArchitectureTimestepAccuracy
STBP-tdBN [73]ResNet34663.72%
TET [16]ResNet34664.79%
RecDis-SNN [23]ResNet34667.33%
GLIF [70]ResNet34467.52%
IM-Loss [22]ResNet18667.43%
Real Spike [24]ResNet18463.68%
ResNet34467.69%
RMP-Loss [25]ResNet18463.03%
ResNet34465.17%
MPBN [26]ResNet18463.14%
ResNet34464.71%
SEW ResNet [17]ResNet18463.18%
ResNet34467.04%
OursResNet18468.27% ± 0.19%
ResNet34470.40% ± 0.15%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.61, + 0.48, + 0.625 + ], + "angle": 0, + "content": "Table 2. Comparison of training based SNN SOTA on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.7 + ], + "angle": 0, + "content": "trained on 8 NVIDIA RTX 4090 GPUs. The total training epochs are 500 for CIFAR-10, CIFAR-100, and CIFAR10-DVS, while ImageNet runs for 300 epochs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.327, + 0.729 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Static Image Classification. We evaluate our model on three static datasets: CIFAR-10, CIFAR-100, and ImageNet. For CIFAR-10/100, STAA-SNNs are applied to both ResNet-like and VGG-like networks and tested at 1, 2, 3, and 4 timesteps. On ImageNet, the model is evaluated using ResNet-18 and ResNet-34 with 4 time steps. Our method achieves top accuracies of \\(97.14\\%\\) on CIFAR-10, \\(82.05\\%\\) on CIFAR-100, and \\(70.40\\%\\) on ImageNet, surpassing previous state-of-the-art methods across various architectures. These improvements are attributed to the strong spatial-temporal attention mechanism and feature aggrega" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.699 + ], + "angle": 0, + "content": "
MethodsArchitectureTimestepAccuracy
CIFAR10-DVS
IM-loss [22]ResNet-191072.60%
LSG [42]ResNet-191077.90%
MPBN [26]ResNet-191074.40%
MPBN [26]ResNet-201078.70%
TET [16]VGG5NN1077.30%
IM-LIF [43]VGG-131080.50%
GLIF [70]7B-wideNet1678.10%
STSA [56]STS-Transformer1679.93%
SEW [17]SEW-ResNet1674.4%
Spikeformer [41]Spikeformer1680.9%
OursResNet-201681.90%±0.20%
VGG-131682.10%±0.20%
DVS128 Gesture
STBP-tdBN [73]ResNet-174096.87%
SEW [17]7B-Net1697.52%
PLIF [18]PLIFNet2097.57%
MA-SNN [66]5 layers SCNN2098.23%
ASA-SNN [67]5 layers SCNN2097.70%
IM-LIF [43]ResNet-194097.33%
LIAF+TA [65]TA-SNN-Net6098.61%
Spike-driven-Transformer [69]Transformer-2-5121699.30%
OursVGG-131698.61%±0.20%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.709, + 0.907, + 0.737 + ], + "angle": 0, + "content": "Table 3. Comparison results with SOTA methods on CIFAR10-DVS and DVS128 Gesture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.907, + 0.844 + ], + "angle": 0, + "content": "tion capabilities of STAA-SNNs, which fully leverage the potential performance of SNNs. Furthermore, our model achieves equivalent accuracy while requiring at least one fewer timestep compared to prior methods, improving efficiency without sacrificing performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Event-based Action Recognition. To further assess the spatial-temporal processing capabilities of STAA-SNNs," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.462, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.274 + ], + "angle": 0, + "content": "Figure 3. Distribution of accuracy with different dropout probability \\(\\beta\\) in TSRD on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.288, + 0.479, + 0.422 + ], + "angle": 0, + "content": "
STAA ComponentsTSRDAccuracyAccuracy Gain
GCPESA
XXXX72.30%-
✓XXX73.22%+0.92%
✓✓XX73.79%+1.49%
✓✓X✓73.96%+1.66%
✓✓✓X74.78%+2.48%
✓✓✓✓75.10%+2.80%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.432, + 0.468, + 0.446 + ], + "angle": 0, + "content": "Table 4. Combinations of sub-modules in STAA with TSRD." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.625 + ], + "angle": 0, + "content": "we conduct experiments on the DVS128 Gesture and CIFAR10-DVS datasets. Unlike static datasets, these neuromorphic datasets contain temporal event streams, requiring models to efficiently process spiking sequences. As shown in Tab. 3, our model achieves state-of-the-art or superior performance with a smaller network and fewer timesteps. Specifically, STAA-SNNs achieve \\(82.10\\%\\) accuracy on CIFAR10-DVS and \\(98.61\\%\\) on the DVS128 Gesture dataset, demonstrating their effectiveness in event-based action recognition tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.245, + 0.65 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.656, + 0.48, + 0.672 + ], + "angle": 0, + "content": "All ablation studies with \\( \\mathrm{T} = 4 \\) using ResNet-20 as baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Position encoding placement. To optimize PE in SNNs, we evaluate its best placement within the network. Given the structural complexity of SNNs compared to CNNs, identifying the most effective integration point is crucial. As shown in Fig. 2, we evaluate four placement strategies (b)-(e) alongside a baseline SNN without PE (a), ensuring all other conditions remain constant. The CIFAR-10 accuracies for (b) to (e) are \\(94.08\\%\\), \\(94.29\\%\\), \\(94.27\\%\\), and \\(93.96\\%\\), respectively. The highest performance occurs when PE is placed at the pre-aggregation spatial input port (c). Although (b) and (d) yield similar results, (d) introduces additional parameters without significant gains. Thus, placing PE at the pre-aggregation spatial input port achieves the best balance between accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Evaluation of position encoding methods. We next evaluate different encoding methods to determine the most effective approach. The results on CIFAR-10 indicate that learnable PE achieves an accuracy of \\(94.99\\%\\), outperforming fixed PE, which attains \\(94.42\\%\\). This \\(0.57\\%\\) improvement highlights the advantage of adaptive position encoding, as it enables the network to dynamically adjust spatial representations, leading to better feature alignment and overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.909, + 0.473 + ], + "angle": 0, + "content": "Impact of dropout probability \\(\\beta\\) in TSRD. To accelerate training and enhance generalization in TSRD, it is essential to determine an optimal dropout probability \\(\\beta\\). Excessive dropout may lead to network collapse by discarding too much information, while insufficient dropout may limit its regularization effect. As shown in Fig. 3, by incrementally increasing \\(\\beta\\) from 0.1 to 1.0, the model achieves relatively high performance when \\(\\beta \\leq 0.3\\), and achieving optimal at 0.1. These results confirm that excessive dropout removes essential information, degrading performance, while a balanced dropout probability effectively suppresses noise in deeper timesteps while preserving critical information in shallower steps. This enables the network to navigate critical learning stages more effectively, enhancing both stability and robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.673 + ], + "angle": 0, + "content": "Combinations of sub-modules in STAA with TSRD. To evaluate the contribution of each sub-module in STAA, as well as the impact of TSRD, we conduct an ablation study by progressively integrating these components into the SNN. The results on CIFAR-100 are presented in Tab. 4. Notably, the GC block alone improves accuracy by \\(0.92\\%\\), while the SA block contributes an additional \\(1.14\\%\\), highlighting the importance of spatio-temporal attention and feature aggregation reconstruction in enhancing SNN performance. When all components, including TSRD, are incorporated, the accuracy reaches \\(75.10\\%\\), demonstrating a cumulative improvement of \\(2.80\\%\\) over the baseline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Impact of Intermediate Dimension Scaling Coefficients in GC. The GC module incorporates multiple \\(1 \\times 1\\) convolutions for feature transformation. Retaining the original channel size \\(C\\) for the last two convolutions layers would significantly increase computational overhead. To maintain the module's efficiency while preserving performance, a scaling coefficient \\(r\\) is introduced within the GC to compress feature dimensions. By applying this scaling mechanism, the parameter count is reduced from \\((C \\cdot C)\\) to \\((2C \\cdot C / r)\\), making the module more computationally feasible. We conduct experiments to assess the impact of different \\(r\\) values, as shown in Fig. 5. The results indicate that the model achieves optimal performance when \\(r = 4\\), effectively balancing computational efficiency and accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.153, + 0.093, + 0.849, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.255, + 0.788, + 0.27 + ], + "angle": 0, + "content": "Figure 4. Visualization on CIFAR10-DVS. Ten layers from VGG-13 in a shallow to deep manner." + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.295, + 0.451, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.444, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Impact of different scaling coefficients \\( r \\) for intermediate feature dimensions in the GC module on the CIFAR-10 dataset." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.498, + 0.331, + 0.515 + ], + "angle": 0, + "content": "4.4. Visualization and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.658 + ], + "angle": 0, + "content": "We aggregate 4D \\((T,C,H,W)\\) spiking maps into 2D \\((H,W)\\) and apply Grad-CAM to visualize feature extraction at different depths. As shown in Fig. 4, the left displays a sample from the horse class in CIFAR10-DVS, while the right compares vanilla LIF and STAA-LIF feature maps. STAA-LIF focuses more effectively on target-relevant features, capturing finer details like the horse's mouth, which the vanilla LIF model misses. This highlights the advantage of spatio-temporal attention in enhancing feature selectivity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.668, + 0.399, + 0.684 + ], + "angle": 0, + "content": "4.5. Analysis of Computation Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the single-image inference energy cost using 45-nm technology. ANNs rely on multiplication-and-accumulation (MAC) operations, which consume \\(4.6\\mathrm{pJ}\\) per operation, whereas SNNs primarily use accumulation (AC) operations at a significantly lower cost of \\(0.9\\mathrm{pJ}\\) per operation, making SNNs inherently more energy-efficient [31, 49]. Following [11], we assess the efficiency of ResNet-20, which requires 0.10 billion ACs, 0.06 billion MACs, 0.87 billion FLOPs, and consists of 12.69 million parameters. Under these conditions, inference on CIFAR-100 (with a \\(32\\times 32\\) resolution and \\(\\mathrm{T} = 4\\)) results in an energy consumption of \\(0.366\\mathrm{mJ}\\), demonstrating the energy-efficient nature of ResNet-20. Additional computational efficiency experiments are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.294, + 0.634, + 0.31 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.321, + 0.907, + 0.775 + ], + "angle": 0, + "content": "In this paper, we introduce STAA-SNN, a spatio-temporal attention-based feature aggregator, together with the TSRD training strategy. STAA-SNN leverages the inherent ability of SNNs to process spatio-temporal information, thereby enhancing network performance while reducing the number of required timesteps. Our approach achieves state-of-the-art results across multiple benchmarks. STAA-SNN is designed as a plug-and-play module that seamlessly integrates into convolution-based SNN architectures. It is composed of several synergistic modules that collectively enhance its capabilities. The GC module simulates the brain's ability to process information across a broad spatial range, involving multiple brain regions in coordination. By focusing on spatial feature relevance, it enables the model to capture global structures and key stimuli in its environment. The PE module enhances temporal encoding by incorporating time-step information into feature representations, mirroring the brain's ability to integrate temporal and spatial cues for improved understanding of time-dependent relationships. The SA module mimics the brain's dynamic attention mechanisms, adjusting feature importance across time steps by adaptively enhancing or suppressing information. This prioritization of critical temporal features strengthens decision-making and overall network efficiency. Together, these modules form a robust spatio-temporal feature enhancement system that improves cognitive processing, robustness, and overall model performance. Despite these promising results, future research should focus on reducing computational energy consumption to further improve efficiency, particularly for resource-constrained applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.702, + 0.809 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.906, + 0.888 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230, 62271361), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13966" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Larry F Abbott. Lapicque's introduction of the integrate-and-fire model neuron (1907). *Brain research bulletin*, 50(5-6): 303–304, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Pepe Alcamí and Alberto E Pereda. Beyond plasticity: the dynamic impact of electrical synapses on neural circuits. Nature Reviews Neuroscience, 20(5):253-271, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[4] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lui, and Cordelia Schmid. Vivit: A video vision transformer. 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[7] Guo-qiang Bi and Mu-ming Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of neuroscience, 18(24):10464-10472, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[8] Wuque Cai, Hongze Sun, Rui Liu, Yan Cui, Jun Wang, Yang Xia, Dezhong Yao, and Daqing Guo. A Spatial-Channel-Temporal-Fused Attention for Spiking Neural Networks. IEEE Transactions on Neural Networks and Learning Systems, pages 1–15, 2023. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[9] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[10] Biswadeep Chakraborty, Xueyuan She, and Saibal Mukhopadhyay. A Fully Spiking Hybrid Neural Network for Energy-Efficient Object Detection. IEEE Transactions on Image Processing, 30:9014-9029, 2021. Conference Name: IEEE Transactions on Image Processing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Guangyao Chen, Peixi Peng, Guoqi Li, and Yonghong Tian. Training Full Spike Neural Networks via Auxiliary Accumulation Pathway, 2023. arXiv:2301.11929 [cs]. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Dennis V Christensen, Regina Dittmann, Bernabe Linares-Barranco, Abu Sebastian, Manuel Le Gallo, Andrea Redaelli, Stefan Slesazeck, Thomas Mikolajick, Sabina Spiga, Stephan Menzel, et al. 2022 roadmap on neuromorphic computing and engineering. Neuromorphic Computing and Engineering, 2(2):022501, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. Ieee Micro, 38(1):82-99, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[19] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[20] Banglei Guan and Ji Zhao. Relative pose estimation for multi-camera systems from point correspondences with scale ratio. In Proceedings of the 30th ACM International Conference on Multimedia, page 5036-5044, New York, NY, USA, 2022. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[21] Banglei Guan, Ji Zhao, Zhang Li, Fang Sun, and Friedrich Fraundorfer. Minimal solutions for relative pose with a single affine correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[22] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems, 35:156-166, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks. In Proceedings of CVPR, pages 326-335, New Orleans, LA, USA, 2022. IEEE. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In European Conference on Computer Vision, pages 52-68. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[25] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. RMP" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13967" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Loss: Regularizing Membrane Potential Distribution for Spiking Neural Networks, 2023. arXiv:2308.06787 [cs]. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[26] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[27] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[28] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[30] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pages 10-14, 2014. ISSN: 2376-8606. 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[32] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[33] J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu. Squeeze-and-excitation networks. IEEE transactions on pattern analysis and machine intelligence, 42(8):2011-2023, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[34] Yifan Hu, Yujie Wu, Lei Deng, and Guoqi Li. Advancing residual learning towards powerful deep spiking neural networks. arXiv e-prints, pages arXiv-2112, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[35] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[36] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[37] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Spiking neural networks and bio-inspired supervised deep learning: A survey. arXiv preprint arXiv:2307.16235, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[38] Chankyu Lee, Syed Shakib Sarwar, Priyadarshini Panda, Gopalakrishnan Srinivasan, and Kaushik Roy. Enabling spike-based backpropagation for training deep neural network architectures. Frontiers in neuroscience, page 119, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Taihang Lei, Banglei Guan, Minzu Liang, Zibin Liu, Jianbin Liu, Yang Shang, and Qifeng Yu. Motion measurements of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "explosive shock waves based on an event camera. Optics Express, 32(9):15390-15409, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[41] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: A novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[42] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3002–3010, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[43] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. IM-LIF: Improved Neuronal Dynamics With Attention Mechanism for Direct Training Deep Spiking Neural Network. IEEE Transactions on Emerging Topics in Computational Intelligence, pages 1-11, 2024. Conference Name: IEEE Transactions on Emerging Topics in Computational Intelligence. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[45] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and tracking with an event camera. IEEE Transactions on Image Processing, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[46] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[47] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[48] Garrick Orchard, E Paxon Frady, Daniel Ben Dayan Rubin, Sophia Sanborn, Sumit Bam Shrestha, Friedrich T Sommer, and Mike Davies. Efficient neuromorphic signal processing with loihi 2. In 2021 IEEE Workshop on Signal Processing Systems (SiPS), pages 254-259. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[49] Ning Qiao, Hesham Mostafa, Federico Corradi, Marc Osswald, Fabio Stefanini, Dora Sumislawka, and Giacomo Indiveri. A reconfigurable on-line learning spiking neuromorphic processor comprising 256 neurons and 128K synapses. Frontiers in Neuroscience, 9, 2015. Publisher: Frontiers. 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[50] Nitin Rathi and Kaushik Roy. Diet-snn: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13968" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[52] Chen She and Laiyun Qing. Spikeformer: Image reconstruction from the sequence of spike camera based on transformer. In Proceedings of the 2022 5th International Conference on Image and Graphics Processing, pages 72-78, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[53] Gina G Turrigiano and Sacha B Nelson. Homeostatic plasticity in the developing nervous system. Nature reviews neuroscience, 5(2):97-107, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[55] Bingsen Wang, Jian Cao, Jue Chen, Shuo Feng, and Yuan Wang. A new ann-snn conversion method with high accuracy, low latency and good robustness. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3067-3075, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[56] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-Temporal Self-Attention for Asynchronous Spiking Neural Networks. In Proceedings of IJCAI, pages 3085–3093, Macau, SAR China, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[57] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[58] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[59] Yujie Wu, Rong Zhao, Jun Zhu, Feng Chen, Mingkun Xu, Guoqi Li, Sen Song, Lei Deng, Guanrui Wang, Hao Zheng, et al. Brain-inspired global-local learning incorporated with neuromorphic computing. Nature Communications, 13(1): 65, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[60] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[61] Qi Xu, Jie Deng, Jiangrong Shen, Biwu Chen, Huajin Tang, and Gang Pan. Hybrid Spiking Vision Transformer for Event-Based Object Detection, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[62] Qi Xu, Xuanye Fang, Yaxin Li, Jiangrong Shen, De Ma, Yi Xu, and Gang Pan. Rsnn: Recurrent spiking neural networks for dynamic spatial-temporal information processing. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10602-10610, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[63] Kashu Yamazaki, Viet-Khoa Vo-Ho, Darshan Bulsara, and Ngan Le. Spiking neural networks and their applications: A review. *Brain Sciences*, 12(7):863, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[64] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[65] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-Wise Attention Spiking Neural Networks for Event Streams Classification. In Proceedings of ICCV, pages 10221-10230, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[66] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[67] Man Yao, Jiakui Hu, Guangshe Zhao, Yaoyuan Wang, Ziyang Zhang, Bo Xu, and Guoqi Li. Inherent Redundancy in Spiking Neural Networks. In Proceedings of the ICCV, pages 16924-16934, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[68] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention Spiking Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9393-9410, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[69] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[70] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[71] Chengting Yu, Zheming Gu, Da Li, Gaoang Wang, Aili Wang, and Erping Li. STSC-SNN: Spatio-Temporal Synaptic Connection with temporal convolution and attention for spiking neural networks. Frontiers in Neuroscience, 16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[72] Malu Zhang, Jibin Wu, Yansong Chua, Xiaoling Luo, Zihan Pan, Dan Liu, and Haizhou Li. Mpd-al: an efficient membrane potential driven aggregate-label learning algorithm for spiking neurons. In Proceedings of the AAAI conference on artificial intelligence, pages 1327-1334, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[73] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[74] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[75] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[76] Rui-Jie Zhu, Malu Zhang, Qihang Zhao, Haoyu Deng, Yule Duan, and Liang-Jian Deng. Tcja-snn: Temporal-channel joint attention for spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13969" + } + ] +] \ No newline at end of file diff --git a/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_origin.pdf b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a2194772c2fbd07b741361a4465f1d9773439dc5 --- /dev/null +++ b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/5435aeb8-d79e-4d0b-a257-264caa74e5cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f908fbf09ed800c9020edd6eafc5d7991ee261d8b8ca3032fdcd2174f4367a43 +size 2473841 diff --git a/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/full.md b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..33d1c774c5c41bded666336d6d9fce6cbd382022 --- /dev/null +++ b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/full.md @@ -0,0 +1,322 @@ +# STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks + +Tianqing Zhang $^{1*}$ Kairong Yu $^{1*}$ Xian Zhong $^{3}$ Hongwei Wang $^{1\dagger}$ Qi Xu $^{2\dagger}$ Qiang Zhang $^{2}$ $^{1}$ Zhejiang University $^{2}$ Dalian University of Technology $^{3}$ Wuhan University of Technology + $^{1}\{zhangtianqing, 22271086\} @zju.edu.cn$ $^{1}$ hongweiwang@intl.zju.edu.cn + $^{2}\{xuqi, zhangq\} @dlut.edu.cn$ $^{3}$ zhongx@whut.edu.cn + +# Abstract + +Spiking Neural Networks (SNNs) have gained significant attention due to their biological plausibility and energy efficiency, making them promising alternatives to Artificial Neural Networks (ANNs). However, the performance gap between SNNs and ANNs remains a substantial challenge hindering the widespread adoption of SNNs. In this paper, we propose a Spatial-Temporal Attention Aggregator SNN (STAA-SNN) framework, which dynamically focuses on and captures both spatial and temporal dependencies. First, we introduce a spike-driven self-attention mechanism specifically designed for SNNs. Additionally, we pioneeringly incorporate position encoding to integrate latent temporal relationships into the incoming features. For spatial-temporal information aggregation, we employ step attention to selectively amplify relevant features to variant steps. Finally, we implement a time-step random dropout strategy to avoid local optima. The framework demonstrates exceptional performance across diverse datasets and exhibits strong generalization capabilities. Notably, STAA-SNN achieves state-of-the-art results on neuromorphic datasets CIFAR10-DVS of $82.10\%$ and with performances of $97.14\%$ , $82.05\%$ and $70.40\%$ on the static datasets CIFAR-10, CIFAR-100 and ImageNet, respectively. Furthermore, this model exhibits improved performance ranging from $0.33\%$ to $2.80\%$ with fewer time steps. + +# 1. Introduction + +In recent years, Spiking Neural Networks (SNNs) have attracted significant attention due to their low energy consumption and biological interpretability[12, 17, 51, 57-59]. Inspired by the behavior of biological neurons, SNNs represent information using discrete binary spikes over multiple timesteps, making them well-suited for implementation + +tion on low-power neuromorphic hardware, offering a distinct advantage over traditional Artificial Neural Networks (ANNs)[2, 13, 19, 48, 63]. Among various spiking neuron models, the Leaky Integrate-and-Fire (LIF) model[1] is widely used in SNNs for its linear differentiable [37, 38]. Although previous works[17, 34, 73] have developed deep SNNs to accelerate model convergence, the multi-timestep computations inherent in SNNs significantly increase both training and inference times, leading to high latency. This limitation hinders the potential for efficient algorithm design. Consequently, a performance gap still exists between directly trained large-scale SNNs and ANNs in pattern recognition tasks, underscoring the need for novel approaches that integrate insights from neuroscience with conventional deep learning to improve both the efficiency and accuracy of SNNs. Inspired by the human visual and cognitive systems, attention mechanisms regulate neural activity and connectivity, enhancing responses to focal stimuli. This enhancement is reflected in the increased firing rates and improved synchronization of neurons in relevant brain regions. Additionally, this mechanism influences the strength and efficiency of neural connections, thereby facilitating the flow of information within and between different brain areas. In the field of deep learning, attention mechanisms enable neural networks to selectively focus on relatively important input information, leading to significant advancements in various domains such as natural language processing (NLP) and computer vision (CV). Compared to traditional architectures [5, 6, 44], these mechanisms have consistently demonstrated superior performance. + +In this paper, we explore the integration of attention mechanisms into deep SNNs. Although SNNs are biologically inspired models, their current implementations exhibit significant limitations due to inadequate biological fidelity. For instance, the heterogeneity among spiking neurons and the ability to independently transmit spatial and temporal information are overlooked, with membrane parameters being uniformly and permanently set across dimensions and layers in neural networks. Neuroscience research [3, 53] has + +highlighted the critical role of synaptic homeostatic plasticity mechanisms in neuronal activity, as well as the regulatory function of glial cells in information transmission between neurons. Enhancing the alignment of SNN models with these biological processes has the potential to improve their overall performance. Therefore, we propose allowing different network layers to exhibit varying degrees of information transmission permeability. + +We introduce a self-attention mechanism to capture spatial correlations and the significance of information, while integrating positional encoding to enhance spatial information and potentially learn temporal features. In traditional SNNs, input features are typically aggregated through simple addition, which can introduce significant noise into the synthesized features. To address this, we propose a Step Attention mechanism that aggregates and reconstructs features at each time step, thereby improving the quality of information propagation. Additionally, we introduce a novel Time-Step Random Dropout (TSRD) strategy to prevent the network from falling into local optima, thereby enhancing its robustness and accelerating training. Our main contributions are summarized as follows: + +- Inspired by neuroscience and attention mechanisms, we propose a novel Spatio-Temporal Attention Aggregator for Spiking Neural Networks (STAA-SNN) that integrates self-attention, positional encoding, and step attention into a unified framework, enabling SNNs to efficiently extract features across temporal and spatial domains while maintaining stable training. +- We introduce a Time-Step Random Dropout (TSRD) strategy to accelerate training and enhance model generalization by preventing premature convergence and enabling optimal performance. +- We evaluate the performance, convergence, and sparse spiking activity of STAA-SNN across multiple benchmark datasets, demonstrating that our method achieves state-of-the-art or superior results with fewer timesteps, highlighting its efficiency and effectiveness. + +# 2. Related Works + +# 2.1. Training of Deep SNNs + +Training deep SNNs generally follows two primary approaches: ANN-to-SNN conversion and direct SNN training. The ANN-to-SNN conversion method aims to transfer knowledge from traditional ANNs to SNNs, but it often encounters challenges such as minimizing accuracy degradation while optimizing inference latency and energy efficiency[27, 28, 55]. Moreover, this approach struggles to fully leverage the spatiotemporal dynamics inherent in SNNs due to the lack of temporal encoding in ANNs. Direct SNN training is categorized into unsupervised and supervised learning. Unsupervised methods rely on biologi + +cally plausible learning rules such as Hebbian learning[30] and Spike-Timing-Dependent Plasticity (STDP)[7]. However, these approaches are generally not suitable for deep SNNs or large-scale datasets. On the other hand, supervised learning integrates backpropagation with surrogate gradients[17, 38, 46, 58], making applicability increasingly viable[60, 61]. Several methods have been proposed to enhance direct SNN training. For instance, MPBN[26] incorporates a batch normalization layer after membrane potential updates to stabilize training, while MPD-AL[72] introduces an alternative learning paradigm that adapts synaptic modifications based on membrane potential traces. MSResNet[34] reorganizes Vanilla ResNet layers to better align with SNNs, improving feature extraction and gradient propagation. Furthermore, IM-Loss[22] optimizes information maximization loss to increase spike information entropy. Recently, RSNN[62] introduced a Convolutional LSTM-based approach that preserves temporal correlations, enhancing spike-based data processing in SNNs. + +# 2.2. Attention Mechanism in SNNs + +Attention mechanisms dynamically assign varying levels of importance to different input components, thereby facilitating efficient processing and feature extraction. Since their introduction in 2014, attention mechanisms have become integral to enhancing deep learning models and have been widely applied to traditional ANNs[32]. In deep learning, attention mechanisms are primarily used in two ways: as a basic paradigm for conducting meta-operator such as self-attention [54] or as an auxiliary enhancement modules [33]. Also these two approaches can be combined. Recently, integrating attention mechanisms into SNNs has gained traction as a means to enhance their performance. Unlike traditional ANNs, where attention typically refines spatial feature representations, SNN-based attention mechanisms must consider both spatial and temporal information. Early work by [47] introduced a computational model that bridges human selective attention and conscious awareness. Kundu[36] utilize attention mechanisms to compress SNN models. Temporal-wise attention was first introduced in TA-SNN[64], laying the foundation for subsequent developments in SNN-specific attention mechanisms. MA-SNN [66] proposed temporal-channel-spatial attention for SNNs, surpassing models focused on single-dimensional information. TCJA-SNN[76] introduced a temporal-channel joint attention mechanism, while SCTFA [8] applied temporal-channel-spatial attention to SNNs with promising results. Moreover, STSC-SNN [71] incorporated temporal convolution and attention mechanisms to implement synaptic filtering and gating functions, demonstrating the versatility of attention mechanisms in enhancing SNN capabilities. IM-LIF [43] adopted temporalwise attention to improve the updating of GRUs and LSTMs + +![](images/19976e25b16ba8dd63c522273cf82e127768f49b4e706e65db2435bd85bc3366.jpg) +(a) Spatial-Temporal Attention Aggregator SNN +(b) Structure of the Global Context (GC) block + +![](images/313680c231220fa98d255eba78d4c8e446f53ed0c29275a20bb5902d186f6464.jpg) + +![](images/4833dfb8c1ffb8fa787f4afbd2c2e593413ef88f869e9269a71e71c6c798aa92.jpg) +(c) Structure of Step Attention (SA) block +Figure 1. Overview of the STAA-SNN Architecture and the TSRD strategy. + +![](images/c09ac145f0551f349096536c3fdd48b04343d57e48af640729277fa3515a0591.jpg) +(d) STAA-SNN with TSRD strategy. The dropout probability is represented by $\beta$ , meaning that at each time step $t$ , there is a probability $\beta$ of selecting the simple addition aggregation method and discarding the STAA. + +within SNNs. More recently, researchers have explored incorporating SNNs into Transformer architectures, achieving promising results on multiple benchmark datasets[41, 52]. These researches reveal that attention mechanisms can significantly improve the performance of SNN. However, there remains a need for more specialized and high-performance attention modules tailored explicitly for SNN architectures. Addressing this gap can further enhance the efficiency and robustness of SNNs in real-world applications. + +# 3. Methodology + +In this section, we will first provide a brief overview of the Leaky Integrate-and-Fire (LIF) model and the training process. Following that, we introduce the proposed Spatial-Temporal Attention Aggregator. Lastly, we outline the timestep random dropout strategy for SNN. + +# 3.1. Leaky Integrate-and-Fire Model + +The spiking neuron is a fundamental component of SNNs. In this work, we adopt the Leaky Integrate-and-Fire (LIF) neuron due to its efficiency and simplicity. The discrete-time, iterative update rule of the LIF model is given by: + +$$ +V ^ {t, n} = H ^ {t - 1, n} + \frac {1}{\tau} \left[ I ^ {t - 1, n} - \left(H ^ {t - 1, n} - V _ {\text {r e s e t}}\right) \right], \tag {1} +$$ + +$$ +S ^ {t, n} = \Theta \left(V ^ {t, n} - v _ {\mathrm {t h}}\right), \tag {2} +$$ + +$$ +H ^ {t, n} = V _ {\text {r e s e t}} \cdot S ^ {t, n} + V ^ {t, n} \odot (1 - S ^ {t, n}). \tag {3} +$$ + +Here, $H^{t - 1,n}$ denotes the membrane potential after a spike trigger at the previous timestep, while $I^{t,n}$ and $V^{t,n}$ represent the input and membrane potential of the $n$ -th layer at timestep $t$ , respectively. The threshold $v_{th}$ determines whether a spike is emitted, producing the binary spike sen + +sor $S^{t,n}$ . The Heaviside step function $\Theta$ outputs 1 if $x \geq 0$ and 0 otherwise. + +Given the similarity between iterative LIF dynamics and Recurrent Neural Networks (RNNs), Spatio-Temporal Backpropagation (STBP) [57, 58] enables direct SNN training via Backpropagation Through Time (BPTT): + +$$ +\frac {\partial L}{\partial W ^ {n}} = \sum_ {t} \left(\frac {\partial L}{\partial S ^ {t , n}} \cdot \frac {\partial S ^ {t , n}}{\partial V ^ {t , n}} + \frac {\partial L}{\partial V ^ {t + 1 , n}} \cdot \frac {\partial V ^ {t + 1 , n}}{\partial V ^ {t , n}}\right) \frac {\partial V ^ {t , n}}{\partial W ^ {n}}. \tag {4} +$$ + +However, the term $\frac{\partial S^{t,n}}{\partial V^{t,n}}$ , which represents the derivative of the spike function, is non-differentiable. To address this, prior works introduce Surrogate Gradients(SG), commonly approximated using a rectangular function: + +$$ +\frac {\partial S ^ {t}}{\partial V ^ {t}} = \frac {1}{a} \cdot \operatorname {s i g n} \left(\left| V ^ {t} - V _ {\mathrm {t h}} \right| < \frac {a}{2}\right), \tag {5} +$$ + +where $a$ is a hyperparameter and set to 1. This approximation provides a gradient of 1 within the range $V_{\mathrm{th}} - 0.5 \leq V^t \leq V_{\mathrm{th}} + 0.5$ , and otherwise 0. + +Neuromorphic-Inspired Adaptive LIF Design. Examining Eq. (1), it is evident that keeping $V_{\mathrm{reset}}$ and $\tau$ constant results in fixed weight coefficients for $H$ and $I$ , which remain unchanged across timestep $t$ and layer $n$ . This simplification deviates from biological neural processing, where neuronal excitability dynamically adapts to stimuli. To bridge this gap, we propose an adaptive modification to the membrane potential update rule: + +$$ +V ^ {t, n} = M \odot H ^ {t - 1, n} + N \odot I ^ {t - 1, n}. \tag {6} +$$ + +Here, $M$ and $N$ are coefficient matrices that dynamically adjust through iterative learning, regulating activation and + +inhibition probabilities across layers. This biologically inspired adaptation enhances information processing efficiency and robustness in SNNs. + +Global Context (GC) Block. The Global Context (GC) Block serves as a lightweight alternative to the self-attention mechanism used in Transformers, enabling the integration of self-attention into conventional convolution-based deep learning architectures without the computational overhead of full Transformer models. This makes it possible to enhance feature learning in non-Transformer structures while maintaining efficiency and precision. In this paper, we simulate the self-attention operation using a GC block [9], which consists of three $1 \times 1$ convolutional kernels corresponding to key, query, and value. As depicted in Fig. 1(b), the spatial spiking input tensor of the $n$ -th layer at the $t$ -th timestep is defined as $X^{t,n} \in \mathbb{R}^{C^n \times H^n \times W^n}$ , where $W^n$ and $H^n$ denote the width and height of the feature map, and $C^n$ represents the number of channels. Passing $X^{t,n}$ through the GC block yields three weight matrices for self-attention, namely $W_k$ , $W_q$ and $W_v$ , defined as: + +$$ +W _ {k} = X ^ {t, n} \cdot \operatorname {S i g m o i d} \left(\operatorname {C o n v} _ {k} \left(X ^ {t, n}\right)\right), \tag {7} +$$ + +$$ +W _ {q} = \operatorname {C o n v} _ {q} \left(W _ {k}\right), \tag {8} +$$ + +$$ +W _ {v} = \operatorname {C o n v} _ {v} (\operatorname {R e L U} (\operatorname {L N} (W _ {q}))). \tag {9} +$$ + +Before computing $W_{k}$ , $X^{t,n}$ is reshaped from $C\times H\times W$ to $C\times HW$ . The resulting matrices $W_{k},W_{q},W_{v}\in \mathbb{R}^{C\times 1\times 1}$ encode self-attention weights, which are then used to refine the input feature map. Subsequently, the enhanced output is computed as: + +$$ +Z ^ {t, n} = X ^ {t, n} \oplus W _ {v}, \tag {10} +$$ + +where $Z^{t,n} \in \mathbb{R}^{C,H,W}$ maintains the same dimensions as $X^{t,n}$ . The refined feature map $Z^{t,n}$ , in comparison to the $X^{t,n}$ , improves the network's ability to focus on self-relevant and task-critical information while suppressing irrelevant signals. This approach facilitates the seamless transmission of enriched data flow, enabling the retention of valuable features for subsequent stages. + +Position Encoding (PE) Block. Position encoding is widely used in NLP and sequence modeling to incorporate positional information into input data. While SNNs share structural similarities with RNNs and inherently process temporal information, their reliance on sparse binary spikes for information storage can lead to significant feature loss. This study introduces position encoding to enhance spatial features while enriching temporal relationships, thereby improving feature retention and strengthening temporal representations. The positional encoding is computed as follows: + +$$ +X ^ {t, n} = I ^ {t, n} \oplus \operatorname {P o s} _ {t, n}, \tag {11} +$$ + +where $\mathrm{Pos}_{t,n}$ represents the position encoding at the $t$ -th time step in the $n$ -th layer, with $\mathrm{Pos}_{t,n} \in \mathbb{R}^{T,C^n}$ . + +Step Attention (SA) Block. While previous modules primarily enhance feature processing from a spatial perspective, SNNs uniquely process information across both spatial and temporal. Unlike other types of network, SNNs offer an additional understanding of information sensitivity across different timesteps. Therefore, in this paper, we introduce Step Attention to tackle the temporal challenges inherent in SNNs. As illustrated in Fig. 1(c), let $U^{t,n} \in \mathbb{R}^{C^n \times H^n \times W^n}$ denote the spatially aggregated feature input at the $n$ -th layer and the $t$ -th timestep. The specific Step Attention mechanism is defined as: + +$$ +\begin{array}{l} V ^ {t, n} = U ^ {t, n} \odot \text {S i g m o i d} \left(\operatorname {C o n v} _ {2} \left(\operatorname {R e L U} \left(\operatorname {C o n v} _ {1} \right. \right. \right. \tag {12} \\ \left. \left(\alpha \cdot \operatorname {A v g P o o l} (U ^ {t, n})\right)\right) \Bigg) \Bigg), \\ \end{array} +$$ + +where $\alpha$ is a scaling factor that regulates the initial training speed for smoother convergence. In practice, $\alpha$ is typically set to a default value of 2. + +LIF with Spatio-Temporal Attention Aggregator (STAA). The STAA seamlessly integrates with the LIF model, offering compatibility across various network structures without constraints. It adeptly harnesses the potential of spiking neurons, effectively optimizing the storage of discrete spike information while minimizing the loss of crucial data during transmission. As shown in Fig. 1(a), the neuronal update equation for the LIF model with STAA is formulated as: + +$$ +U ^ {t, n} = G C _ {1} \left(X ^ {t, n}\right) \oplus G C _ {2} \left(H ^ {t - 1, n}\right), \tag {13} +$$ + +$$ +V ^ {t, n} = S A \left(U ^ {t, n}\right). \tag {14} +$$ + +Where $GC_{1}$ and $GC_{2}$ represent Global Context Blocks for the input $X^{t,n}$ and the previous membrane state $H^{t-1,n}$ , respectively. The remaining update equations follow the original LIF model. + +# 3.2. Time Step Random Dropout (TSRD) Strategy + +To further enhance training efficiency, we propose TSRD, an adaptive training strategy designed to prevent premature feature stabilization in deep SNNs. In deeper timesteps, spiking features may solidify too early during iterations, hindering the network from achieving optimal convergence. As illustrated in Fig. 1(d), TSRD randomly dropout the augmentation module at timestep $t \in (0,T)$ with a dropout probability $\beta$ , and instead choose the most basic elementwise addition aggregation method. The TSRD technique accelerates training and helps the network escape local optima, thereby improving generalization performance. + +
MethodArchitectureTimestepAccuracy (%)
CIFAR-10CIFAR-100
CNN-basedGLIF [70]ResNet-19294.44%75.48%
ResNet-19494.85%77.05%
TET [16]ResNet-19294.16%72.87%
ResNet-19494.44%74.47%
LSG [42]ResNet-19294.41%76.32%
ResNet-19495.17%76.85%
PFA [14]ResNet-19295.6%76.7%
ResNet-19495.71%78.1%
Diet-SNN [50]ResNet-20591.78%64.07%
VGG-16593.85%69.67%
IM-loss [22]VGG-16593.85%70.18%
IM-LIF [43]ResNet-19395.29%77.21%
MPBN [26]ResNet-19196.06%78.71%
ResNet-19296.47%79.51%
ResNet-19496.52%80.1%
ResNet-20192.22%68.41%
ResNet-20293.54%70.79%
ResNet-20494.28%72.3%
Transformer-basedSpikformer[75]Spikformer-4-25493.94%75.96%
Spikingformer[74]Spikingformer-4-25494.77%77.43%
Spike-driven Transformer [69]Transformer-2-512495.60%78.40%
CNN-basedOursVGG-13294.70%±0.12%75.16%±0.08%
VGG-13495.26%±0.09%76.35%±0.11%
ResNet-19196.75%±0.09%79.37%±0.10%
ResNet-19296.85%±0.11%80.57%±0.11%
ResNet-19497.14%±0.10%82.05%±0.10%
ResNet-20193.08%±0.09%70.14%±0.07%
ResNet-20294.35%±0.08%73.20%±0.12%
ResNet-20495.03%±0.11%75.10%±0.12%
+ +Table 1. Comparison results with SOTA methods on CIFAR-10/100. + +# 4. Experiments + +We first outlined the experimental setup, including dataset details and implementation specifics. We then compare our results with previous SOTA methods across multiple benchmarks. Next, we perform ablation studies to assess the effectiveness of STAA. Finally, we visualize the model's performance. Further details on datasets, hyperparameters, additional experiments, and computational efficiency analysis are provided in the Supplementary Material. + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on five benchmark datasets. CIFAR-10/100[35] are standard image classification datasets with 50,000 training and 10,000 test images across 10 and 100 classes, respectively. ImageNet[15] is a + +large-scale dataset containing 1.2 million training, 50,000 validation, and 100,000 test images spanning 1,000 categories. For neuromorphic vision tasks, CIFAR10-DVS[39, 40, 45] consists of 10,000 event streams across 10 classes, with a 9:1 training-to-testing split. DVS128 Gesture[4] is an event-based gesture recognition dataset with 1,176 training and 288 test samples, featuring 11 gesture classes recorded from 29 subjects under three lighting conditions. + +Implementation Details. In our experiments, we set the firing threshold $v_{\mathrm{th}}$ to 1, and initialize $\alpha$ and $\beta$ to 2 and 0.1, respectively. In the PE block, we adopt a learnable position encoding method, initializing all values to zero. All experiments are implemented in PyTorch and conducted on an NVIDIA RTX 3090 GPU, except for ImageNet, which is + +![](images/3802872e9f7760e37689e04a0d3a4b7edd04d9f5581501ea48585c103e5cb6f7.jpg) +Figure 2. Position encoding locations in SNNs. + +
MethodArchitectureTimestepAccuracy
STBP-tdBN [73]ResNet34663.72%
TET [16]ResNet34664.79%
RecDis-SNN [23]ResNet34667.33%
GLIF [70]ResNet34467.52%
IM-Loss [22]ResNet18667.43%
Real Spike [24]ResNet18463.68%
ResNet34467.69%
RMP-Loss [25]ResNet18463.03%
ResNet34465.17%
MPBN [26]ResNet18463.14%
ResNet34464.71%
SEW ResNet [17]ResNet18463.18%
ResNet34467.04%
OursResNet18468.27% ± 0.19%
ResNet34470.40% ± 0.15%
+ +trained on 8 NVIDIA RTX 4090 GPUs. The total training epochs are 500 for CIFAR-10, CIFAR-100, and CIFAR10-DVS, while ImageNet runs for 300 epochs. + +# 4.2. Performance Comparison + +Static Image Classification. We evaluate our model on three static datasets: CIFAR-10, CIFAR-100, and ImageNet. For CIFAR-10/100, STAA-SNNs are applied to both ResNet-like and VGG-like networks and tested at 1, 2, 3, and 4 timesteps. On ImageNet, the model is evaluated using ResNet-18 and ResNet-34 with 4 time steps. Our method achieves top accuracies of $97.14\%$ on CIFAR-10, $82.05\%$ on CIFAR-100, and $70.40\%$ on ImageNet, surpassing previous state-of-the-art methods across various architectures. These improvements are attributed to the strong spatial-temporal attention mechanism and feature aggrega + +Table 2. Comparison of training based SNN SOTA on ImageNet. + +
MethodsArchitectureTimestepAccuracy
CIFAR10-DVS
IM-loss [22]ResNet-191072.60%
LSG [42]ResNet-191077.90%
MPBN [26]ResNet-191074.40%
MPBN [26]ResNet-201078.70%
TET [16]VGG5NN1077.30%
IM-LIF [43]VGG-131080.50%
GLIF [70]7B-wideNet1678.10%
STSA [56]STS-Transformer1679.93%
SEW [17]SEW-ResNet1674.4%
Spikeformer [41]Spikeformer1680.9%
OursResNet-201681.90%±0.20%
VGG-131682.10%±0.20%
DVS128 Gesture
STBP-tdBN [73]ResNet-174096.87%
SEW [17]7B-Net1697.52%
PLIF [18]PLIFNet2097.57%
MA-SNN [66]5 layers SCNN2098.23%
ASA-SNN [67]5 layers SCNN2097.70%
IM-LIF [43]ResNet-194097.33%
LIAF+TA [65]TA-SNN-Net6098.61%
Spike-driven-Transformer [69]Transformer-2-5121699.30%
OursVGG-131698.61%±0.20%
+ +Table 3. Comparison results with SOTA methods on CIFAR10-DVS and DVS128 Gesture. + +tion capabilities of STAA-SNNs, which fully leverage the potential performance of SNNs. Furthermore, our model achieves equivalent accuracy while requiring at least one fewer timestep compared to prior methods, improving efficiency without sacrificing performance. + +Event-based Action Recognition. To further assess the spatial-temporal processing capabilities of STAA-SNNs, + +![](images/8784705889bb5c7d12c031d70e074f6f6cdf83688940b7660fad31a5fd4c3cc9.jpg) +Figure 3. Distribution of accuracy with different dropout probability $\beta$ in TSRD on CIFAR-10. + +
STAA ComponentsTSRDAccuracyAccuracy Gain
GCPESA
XXXX72.30%-
✓XXX73.22%+0.92%
✓✓XX73.79%+1.49%
✓✓X✓73.96%+1.66%
✓✓✓X74.78%+2.48%
✓✓✓✓75.10%+2.80%
+ +Table 4. Combinations of sub-modules in STAA with TSRD. + +we conduct experiments on the DVS128 Gesture and CIFAR10-DVS datasets. Unlike static datasets, these neuromorphic datasets contain temporal event streams, requiring models to efficiently process spiking sequences. As shown in Tab. 3, our model achieves state-of-the-art or superior performance with a smaller network and fewer timesteps. Specifically, STAA-SNNs achieve $82.10\%$ accuracy on CIFAR10-DVS and $98.61\%$ on the DVS128 Gesture dataset, demonstrating their effectiveness in event-based action recognition tasks. + +# 4.3. Ablation Study + +All ablation studies with $\mathrm{T} = 4$ using ResNet-20 as baseline. + +Position encoding placement. To optimize PE in SNNs, we evaluate its best placement within the network. Given the structural complexity of SNNs compared to CNNs, identifying the most effective integration point is crucial. As shown in Fig. 2, we evaluate four placement strategies (b)-(e) alongside a baseline SNN without PE (a), ensuring all other conditions remain constant. The CIFAR-10 accuracies for (b) to (e) are $94.08\%$ , $94.29\%$ , $94.27\%$ , and $93.96\%$ , respectively. The highest performance occurs when PE is placed at the pre-aggregation spatial input port (c). Although (b) and (d) yield similar results, (d) introduces additional parameters without significant gains. Thus, placing PE at the pre-aggregation spatial input port achieves the best balance between accuracy and efficiency. + +Evaluation of position encoding methods. We next evaluate different encoding methods to determine the most effective approach. The results on CIFAR-10 indicate that learnable PE achieves an accuracy of $94.99\%$ , outperforming fixed PE, which attains $94.42\%$ . This $0.57\%$ improvement highlights the advantage of adaptive position encoding, as it enables the network to dynamically adjust spatial representations, leading to better feature alignment and overall performance. + +Impact of dropout probability $\beta$ in TSRD. To accelerate training and enhance generalization in TSRD, it is essential to determine an optimal dropout probability $\beta$ . Excessive dropout may lead to network collapse by discarding too much information, while insufficient dropout may limit its regularization effect. As shown in Fig. 3, by incrementally increasing $\beta$ from 0.1 to 1.0, the model achieves relatively high performance when $\beta \leq 0.3$ , and achieving optimal at 0.1. These results confirm that excessive dropout removes essential information, degrading performance, while a balanced dropout probability effectively suppresses noise in deeper timesteps while preserving critical information in shallower steps. This enables the network to navigate critical learning stages more effectively, enhancing both stability and robustness. + +Combinations of sub-modules in STAA with TSRD. To evaluate the contribution of each sub-module in STAA, as well as the impact of TSRD, we conduct an ablation study by progressively integrating these components into the SNN. The results on CIFAR-100 are presented in Tab. 4. Notably, the GC block alone improves accuracy by $0.92\%$ , while the SA block contributes an additional $1.14\%$ , highlighting the importance of spatio-temporal attention and feature aggregation reconstruction in enhancing SNN performance. When all components, including TSRD, are incorporated, the accuracy reaches $75.10\%$ , demonstrating a cumulative improvement of $2.80\%$ over the baseline. + +Impact of Intermediate Dimension Scaling Coefficients in GC. The GC module incorporates multiple $1 \times 1$ convolutions for feature transformation. Retaining the original channel size $C$ for the last two convolutions layers would significantly increase computational overhead. To maintain the module's efficiency while preserving performance, a scaling coefficient $r$ is introduced within the GC to compress feature dimensions. By applying this scaling mechanism, the parameter count is reduced from $(C \cdot C)$ to $(2C \cdot C / r)$ , making the module more computationally feasible. We conduct experiments to assess the impact of different $r$ values, as shown in Fig. 5. The results indicate that the model achieves optimal performance when $r = 4$ , effectively balancing computational efficiency and accuracy. + +![](images/db4f89a7409725552287a7029ebf952e759f4ab8d4905b78a495427f0ac68c63.jpg) +Figure 4. Visualization on CIFAR10-DVS. Ten layers from VGG-13 in a shallow to deep manner. + +![](images/f98014b62ed6b38cfa5b66c7ef1f002197aabee5d4150248636ff226ecff292d.jpg) +Figure 5. Impact of different scaling coefficients $r$ for intermediate feature dimensions in the GC module on the CIFAR-10 dataset. + +# 4.4. Visualization and Analysis + +We aggregate 4D $(T,C,H,W)$ spiking maps into 2D $(H,W)$ and apply Grad-CAM to visualize feature extraction at different depths. As shown in Fig. 4, the left displays a sample from the horse class in CIFAR10-DVS, while the right compares vanilla LIF and STAA-LIF feature maps. STAA-LIF focuses more effectively on target-relevant features, capturing finer details like the horse's mouth, which the vanilla LIF model misses. This highlights the advantage of spatio-temporal attention in enhancing feature selectivity. + +# 4.5. Analysis of Computation Efficiency + +We evaluate the single-image inference energy cost using 45-nm technology. ANNs rely on multiplication-and-accumulation (MAC) operations, which consume $4.6\mathrm{pJ}$ per operation, whereas SNNs primarily use accumulation (AC) operations at a significantly lower cost of $0.9\mathrm{pJ}$ per operation, making SNNs inherently more energy-efficient [31, 49]. Following [11], we assess the efficiency of ResNet-20, which requires 0.10 billion ACs, 0.06 billion MACs, 0.87 billion FLOPs, and consists of 12.69 million parameters. Under these conditions, inference on CIFAR-100 (with a $32\times 32$ resolution and $\mathrm{T} = 4$ ) results in an energy consumption of $0.366\mathrm{mJ}$ , demonstrating the energy-efficient nature of ResNet-20. Additional computational efficiency experiments are provided in the Supplementary Material. + +# 5. Conclusion + +In this paper, we introduce STAA-SNN, a spatio-temporal attention-based feature aggregator, together with the TSRD training strategy. STAA-SNN leverages the inherent ability of SNNs to process spatio-temporal information, thereby enhancing network performance while reducing the number of required timesteps. Our approach achieves state-of-the-art results across multiple benchmarks. STAA-SNN is designed as a plug-and-play module that seamlessly integrates into convolution-based SNN architectures. It is composed of several synergistic modules that collectively enhance its capabilities. The GC module simulates the brain's ability to process information across a broad spatial range, involving multiple brain regions in coordination. By focusing on spatial feature relevance, it enables the model to capture global structures and key stimuli in its environment. The PE module enhances temporal encoding by incorporating time-step information into feature representations, mirroring the brain's ability to integrate temporal and spatial cues for improved understanding of time-dependent relationships. The SA module mimics the brain's dynamic attention mechanisms, adjusting feature importance across time steps by adaptively enhancing or suppressing information. This prioritization of critical temporal features strengthens decision-making and overall network efficiency. Together, these modules form a robust spatio-temporal feature enhancement system that improves cognitive processing, robustness, and overall model performance. Despite these promising results, future research should focus on reducing computational energy consumption to further improve efficiency, particularly for resource-constrained applications. + +# 6. Acknowledgements + +This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230, 62271361), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology. + +# References + +[1] Larry F Abbott. Lapicque's introduction of the integrate-and-fire model neuron (1907). *Brain research bulletin*, 50(5-6): 303–304, 1999. 1 +[2] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1 +[3] Pepe Alcamí and Alberto E Pereda. Beyond plasticity: the dynamic impact of electrical synapses on neural circuits. Nature Reviews Neuroscience, 20(5):253-271, 2019. 1 +[4] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 1 +[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lui, and Cordelia Schmid. Vivit: A video vision transformer. 2021. 1 +[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? 2021. 1 +[7] Guo-qiang Bi and Mu-ming Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of neuroscience, 18(24):10464-10472, 1998. 2 +[8] Wuque Cai, Hongze Sun, Rui Liu, Yan Cui, Jun Wang, Yang Xia, Dezhong Yao, and Daqing Guo. A Spatial-Channel-Temporal-Fused Attention for Spiking Neural Networks. IEEE Transactions on Neural Networks and Learning Systems, pages 1–15, 2023. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2 +[9] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4 +[10] Biswadeep Chakraborty, Xueyuan She, and Saibal Mukhopadhyay. A Fully Spiking Hybrid Neural Network for Energy-Efficient Object Detection. IEEE Transactions on Image Processing, 30:9014-9029, 2021. Conference Name: IEEE Transactions on Image Processing. 3 +[11] Guangyao Chen, Peixi Peng, Guoqi Li, and Yonghong Tian. Training Full Spike Neural Networks via Auxiliary Accumulation Pathway, 2023. arXiv:2301.11929 [cs]. 8 +[12] Dennis V Christensen, Regina Dittmann, Bernabe Linares-Barranco, Abu Sebastian, Manuel Le Gallo, Andrea Redaelli, Stefan Slesazeck, Thomas Mikolajick, Sabina Spiga, Stephan Menzel, et al. 2022 roadmap on neuromorphic computing and engineering. Neuromorphic Computing and Engineering, 2(2):022501, 2022. 1 + +[13] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. Ieee Micro, 38(1):82-99, 2018. 1 +[14] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 5 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 5 +[16] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.5, 6 +[17] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 1, 2, 6 +[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6 +[19] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014. 1 +[20] Banglei Guan and Ji Zhao. Relative pose estimation for multi-camera systems from point correspondences with scale ratio. In Proceedings of the 30th ACM International Conference on Multimedia, page 5036-5044, New York, NY, USA, 2022. Association for Computing Machinery. 1 +[21] Banglei Guan, Ji Zhao, Zhang Li, Fang Sun, and Friedrich Fraundorfer. Minimal solutions for relative pose with a single affine correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[22] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems, 35:156-166, 2022. 2, 5, 6 +[23] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks. In Proceedings of CVPR, pages 326-335, New Orleans, LA, USA, 2022. IEEE. 6 +[24] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In European Conference on Computer Vision, pages 52-68. Springer, 2022. 6 +[25] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. RMP + +Loss: Regularizing Membrane Potential Distribution for Spiking Neural Networks, 2023. arXiv:2308.06787 [cs]. 6 +[26] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 2, 5, 6 +[27] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2 +[28] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 2 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[30] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2 +[31] Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pages 10-14, 2014. ISSN: 2376-8606. 8, 3 +[32] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[33] J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu. Squeeze-and-excitation networks. IEEE transactions on pattern analysis and machine intelligence, 42(8):2011-2023, 2020. 2 +[34] Yifan Hu, Yujie Wu, Lei Deng, and Guoqi Li. Advancing residual learning towards powerful deep spiking neural networks. arXiv e-prints, pages arXiv-2112, 2021. 1, 2 +[35] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 5, 1 +[36] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2 +[37] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Spiking neural networks and bio-inspired supervised deep learning: A survey. arXiv preprint arXiv:2307.16235, 2023. 1 +[38] Chankyu Lee, Syed Shakib Sarwar, Priyadarshini Panda, Gopalakrishnan Srinivasan, and Kaushik Roy. Enabling spike-based backpropagation for training deep neural network architectures. Frontiers in neuroscience, page 119, 2020. 1, 2 +[39] Taihang Lei, Banglei Guan, Minzu Liang, Zibin Liu, Jianbin Liu, Yang Shang, and Qifeng Yu. Motion measurements of + +explosive shock waves based on an event camera. Optics Express, 32(9):15390-15409, 2024. 5 +[40] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 1 +[41] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: A novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3, 6 +[42] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3002–3010, 2023. 5, 6 +[43] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. IM-LIF: Improved Neuronal Dynamics With Attention Mechanism for Direct Training Deep Spiking Neural Network. IEEE Transactions on Emerging Topics in Computational Intelligence, pages 1-11, 2024. Conference Name: IEEE Transactions on Emerging Topics in Computational Intelligence. 2, 5, 6 +[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021. 1 +[45] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and tracking with an event camera. IEEE Transactions on Image Processing, 2024. 5 +[46] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 2 +[47] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2 +[48] Garrick Orchard, E Paxon Frady, Daniel Ben Dayan Rubin, Sophia Sanborn, Sumit Bam Shrestha, Friedrich T Sommer, and Mike Davies. Efficient neuromorphic signal processing with loihi 2. In 2021 IEEE Workshop on Signal Processing Systems (SiPS), pages 254-259. IEEE, 2021. 1 +[49] Ning Qiao, Hesham Mostafa, Federico Corradi, Marc Osswald, Fabio Stefanini, Dora Sumislawka, and Giacomo Indiveri. A reconfigurable on-line learning spiking neuromorphic processor comprising 256 neurons and 128K synapses. Frontiers in Neuroscience, 9, 2015. Publisher: Frontiers. 8, 3 +[50] Nitin Rathi and Kaushik Roy. Diet-snn: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 5 +[51] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1 + +[52] Chen She and Laiyun Qing. Spikeformer: Image reconstruction from the sequence of spike camera based on transformer. In Proceedings of the 2022 5th International Conference on Image and Graphics Processing, pages 72-78, 2022. 3 +[53] Gina G Turrigiano and Sacha B Nelson. Homeostatic plasticity in the developing nervous system. Nature reviews neuroscience, 5(2):97-107, 2004. 1 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv, 2017. 2 +[55] Bingsen Wang, Jian Cao, Jue Chen, Shuo Feng, and Yuan Wang. A new ann-snn conversion method with high accuracy, low latency and good robustness. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3067-3075, 2023. 2 +[56] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-Temporal Self-Attention for Asynchronous Spiking Neural Networks. In Proceedings of IJCAI, pages 3085–3093, Macau, SAR China, 2023. 6 +[57] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 3 +[58] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2, 3 +[59] Yujie Wu, Rong Zhao, Jun Zhu, Feng Chen, Mingkun Xu, Guoqi Li, Sen Song, Lei Deng, Guanrui Wang, Hao Zheng, et al. Brain-inspired global-local learning incorporated with neuromorphic computing. Nature Communications, 13(1): 65, 2022. 1 +[60] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 2 +[61] Qi Xu, Jie Deng, Jiangrong Shen, Biwu Chen, Huajin Tang, and Gang Pan. Hybrid Spiking Vision Transformer for Event-Based Object Detection, 2024. 2 +[62] Qi Xu, Xuanye Fang, Yaxin Li, Jiangrong Shen, De Ma, Yi Xu, and Gang Pan. Rsnn: Recurrent spiking neural networks for dynamic spatial-temporal information processing. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10602-10610, 2024. 2 +[63] Kashu Yamazaki, Viet-Khoa Vo-Ho, Darshan Bulsara, and Ngan Le. Spiking neural networks and their applications: A review. *Brain Sciences*, 12(7):863, 2022. 1 +[64] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2 +[65] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-Wise Attention Spiking Neural Networks for Event Streams Classification. In Proceedings of ICCV, pages 10221-10230, 2021. 6 + +[66] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 6 +[67] Man Yao, Jiakui Hu, Guangshe Zhao, Yaoyuan Wang, Ziyang Zhang, Bo Xu, and Guoqi Li. Inherent Redundancy in Spiking Neural Networks. In Proceedings of the ICCV, pages 16924-16934, 2023. 6 +[68] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention Spiking Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9393-9410, 2023. 3 +[69] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 5, 6 +[70] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5, 6 +[71] Chengting Yu, Zheming Gu, Da Li, Gaoang Wang, Aili Wang, and Erping Li. STSC-SNN: Spatio-Temporal Synaptic Connection with temporal convolution and attention for spiking neural networks. Frontiers in Neuroscience, 16, 2022. 2 +[72] Malu Zhang, Jibin Wu, Yansong Chua, Xiaoling Luo, Zihan Pan, Dan Liu, and Haizhou Li. Mpd-al: an efficient membrane potential driven aggregate-label learning algorithm for spiking neurons. In Proceedings of the AAAI conference on artificial intelligence, pages 1327-1334, 2019. 2 +[73] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6 +[74] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5 +[75] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 5 +[76] Rui-Jie Zhu, Malu Zhang, Qihang Zhao, Haoyu Deng, Yule Duan, and Liang-Jian Deng. Tcja-snn: Temporal-channel joint attention for spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2 \ No newline at end of file diff --git a/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/images.zip b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..88ca4177076499ab4541e918a600273f3fa27e77 --- /dev/null +++ b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15490d73c2aa7e6333578df4934e09d73fa4a38157f6a120071a75d90d85362e +size 581821 diff --git a/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/layout.json b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f4334003bf16f968939c87ce97dd758df2ac1a41 --- /dev/null +++ b/2025/STAA-SNN_ Spatial-Temporal Attention Aggregator for Spiking Neural Networks/layout.json @@ -0,0 +1,9363 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 526, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 526, + 138 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 526, + 138 + ], + "type": "text", + "content": "STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "spans": [ + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": "Tianqing Zhang" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": " Kairong Yu" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": " Xian Zhong" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": " Hongwei Wang" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": " Qi Xu" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": " Qiang Zhang" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": "Dalian University of Technology " + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": "Wuhan University of Technology \n" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{1}\\{zhangtianqing, 22271086\\} @zju.edu.cn" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": "hongweiwang@intl.zju.edu.cn \n" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{2}\\{xuqi, zhangq\\} @dlut.edu.cn" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 160, + 539, + 219 + ], + "type": "text", + "content": "zhongx@whut.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) have gained significant attention due to their biological plausibility and energy efficiency, making them promising alternatives to Artificial Neural Networks (ANNs). However, the performance gap between SNNs and ANNs remains a substantial challenge hindering the widespread adoption of SNNs. In this paper, we propose a Spatial-Temporal Attention Aggregator SNN (STAA-SNN) framework, which dynamically focuses on and captures both spatial and temporal dependencies. First, we introduce a spike-driven self-attention mechanism specifically designed for SNNs. Additionally, we pioneeringly incorporate position encoding to integrate latent temporal relationships into the incoming features. For spatial-temporal information aggregation, we employ step attention to selectively amplify relevant features to variant steps. Finally, we implement a time-step random dropout strategy to avoid local optima. The framework demonstrates exceptional performance across diverse datasets and exhibits strong generalization capabilities. Notably, STAA-SNN achieves state-of-the-art results on neuromorphic datasets CIFAR10-DVS of " + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "inline_equation", + "content": "82.10\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": " and with performances of " + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "inline_equation", + "content": "97.14\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "inline_equation", + "content": "82.05\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "inline_equation", + "content": "70.40\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": " on the static datasets CIFAR-10, CIFAR-100 and ImageNet, respectively. Furthermore, this model exhibits improved performance ranging from " + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "inline_equation", + "content": "0.33\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "inline_equation", + "content": "2.80\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": " with fewer time steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 613, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 687 + ], + "type": "text", + "content": "In recent years, Spiking Neural Networks (SNNs) have attracted significant attention due to their low energy consumption and biological interpretability[12, 17, 51, 57-59]. Inspired by the behavior of biological neurons, SNNs represent information using discrete binary spikes over multiple timesteps, making them well-suited for implementation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 604 + ], + "type": "text", + "content": "tion on low-power neuromorphic hardware, offering a distinct advantage over traditional Artificial Neural Networks (ANNs)[2, 13, 19, 48, 63]. Among various spiking neuron models, the Leaky Integrate-and-Fire (LIF) model[1] is widely used in SNNs for its linear differentiable [37, 38]. Although previous works[17, 34, 73] have developed deep SNNs to accelerate model convergence, the multi-timestep computations inherent in SNNs significantly increase both training and inference times, leading to high latency. This limitation hinders the potential for efficient algorithm design. Consequently, a performance gap still exists between directly trained large-scale SNNs and ANNs in pattern recognition tasks, underscoring the need for novel approaches that integrate insights from neuroscience with conventional deep learning to improve both the efficiency and accuracy of SNNs. Inspired by the human visual and cognitive systems, attention mechanisms regulate neural activity and connectivity, enhancing responses to focal stimuli. This enhancement is reflected in the increased firing rates and improved synchronization of neurons in relevant brain regions. Additionally, this mechanism influences the strength and efficiency of neural connections, thereby facilitating the flow of information within and between different brain areas. In the field of deep learning, attention mechanisms enable neural networks to selectively focus on relatively important input information, leading to significant advancements in various domains such as natural language processing (NLP) and computer vision (CV). Compared to traditional architectures [5, 6, 44], these mechanisms have consistently demonstrated superior performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we explore the integration of attention mechanisms into deep SNNs. Although SNNs are biologically inspired models, their current implementations exhibit significant limitations due to inadequate biological fidelity. For instance, the heterogeneity among spiking neurons and the ability to independently transmit spatial and temporal information are overlooked, with membrane parameters being uniformly and permanently set across dimensions and layers in neural networks. Neuroscience research [3, 53] has" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 149, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding authors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13959" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "highlighted the critical role of synaptic homeostatic plasticity mechanisms in neuronal activity, as well as the regulatory function of glial cells in information transmission between neurons. Enhancing the alignment of SNN models with these biological processes has the potential to improve their overall performance. Therefore, we propose allowing different network layers to exhibit varying degrees of information transmission permeability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 335 + ], + "type": "text", + "content": "We introduce a self-attention mechanism to capture spatial correlations and the significance of information, while integrating positional encoding to enhance spatial information and potentially learn temporal features. In traditional SNNs, input features are typically aggregated through simple addition, which can introduce significant noise into the synthesized features. To address this, we propose a Step Attention mechanism that aggregates and reconstructs features at each time step, thereby improving the quality of information propagation. Additionally, we introduce a novel Time-Step Random Dropout (TSRD) strategy to prevent the network from falling into local optima, thereby enhancing its robustness and accelerating training. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 293, + 528 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 338, + 293, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 293, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 293, + 420 + ], + "type": "text", + "content": "- Inspired by neuroscience and attention mechanisms, we propose a novel Spatio-Temporal Attention Aggregator for Spiking Neural Networks (STAA-SNN) that integrates self-attention, positional encoding, and step attention into a unified framework, enabling SNNs to efficiently extract features across temporal and spatial domains while maintaining stable training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 422, + 293, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 293, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 293, + 468 + ], + "type": "text", + "content": "- We introduce a Time-Step Random Dropout (TSRD) strategy to accelerate training and enhance model generalization by preventing premature convergence and enabling optimal performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 469, + 293, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 469, + 293, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 469, + 293, + 528 + ], + "type": "text", + "content": "- We evaluate the performance, convergence, and sparse spiking activity of STAA-SNN across multiple benchmark datasets, demonstrating that our method achieves state-of-the-art or superior results with fewer timesteps, highlighting its efficiency and effectiveness." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 542, + 146, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 146, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 146, + 555 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 186, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 186, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 186, + 576 + ], + "type": "text", + "content": "2.1. Training of Deep SNNs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 294, + 713 + ], + "type": "text", + "content": "Training deep SNNs generally follows two primary approaches: ANN-to-SNN conversion and direct SNN training. The ANN-to-SNN conversion method aims to transfer knowledge from traditional ANNs to SNNs, but it often encounters challenges such as minimizing accuracy degradation while optimizing inference latency and energy efficiency[27, 28, 55]. Moreover, this approach struggles to fully leverage the spatiotemporal dynamics inherent in SNNs due to the lack of temporal encoding in ANNs. Direct SNN training is categorized into unsupervised and supervised learning. Unsupervised methods rely on biologi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "cally plausible learning rules such as Hebbian learning[30] and Spike-Timing-Dependent Plasticity (STDP)[7]. However, these approaches are generally not suitable for deep SNNs or large-scale datasets. On the other hand, supervised learning integrates backpropagation with surrogate gradients[17, 38, 46, 58], making applicability increasingly viable[60, 61]. Several methods have been proposed to enhance direct SNN training. For instance, MPBN[26] incorporates a batch normalization layer after membrane potential updates to stabilize training, while MPD-AL[72] introduces an alternative learning paradigm that adapts synaptic modifications based on membrane potential traces. MSResNet[34] reorganizes Vanilla ResNet layers to better align with SNNs, improving feature extraction and gradient propagation. Furthermore, IM-Loss[22] optimizes information maximization loss to increase spike information entropy. Recently, RSNN[62] introduced a Convolutional LSTM-based approach that preserves temporal correlations, enhancing spike-based data processing in SNNs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 479, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 479, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 479, + 323 + ], + "type": "text", + "content": "2.2. Attention Mechanism in SNNs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 713 + ], + "type": "text", + "content": "Attention mechanisms dynamically assign varying levels of importance to different input components, thereby facilitating efficient processing and feature extraction. Since their introduction in 2014, attention mechanisms have become integral to enhancing deep learning models and have been widely applied to traditional ANNs[32]. In deep learning, attention mechanisms are primarily used in two ways: as a basic paradigm for conducting meta-operator such as self-attention [54] or as an auxiliary enhancement modules [33]. Also these two approaches can be combined. Recently, integrating attention mechanisms into SNNs has gained traction as a means to enhance their performance. Unlike traditional ANNs, where attention typically refines spatial feature representations, SNN-based attention mechanisms must consider both spatial and temporal information. Early work by [47] introduced a computational model that bridges human selective attention and conscious awareness. Kundu[36] utilize attention mechanisms to compress SNN models. Temporal-wise attention was first introduced in TA-SNN[64], laying the foundation for subsequent developments in SNN-specific attention mechanisms. MA-SNN [66] proposed temporal-channel-spatial attention for SNNs, surpassing models focused on single-dimensional information. TCJA-SNN[76] introduced a temporal-channel joint attention mechanism, while SCTFA [8] applied temporal-channel-spatial attention to SNNs with promising results. Moreover, STSC-SNN [71] incorporated temporal convolution and attention mechanisms to implement synaptic filtering and gating functions, demonstrating the versatility of attention mechanisms in enhancing SNN capabilities. IM-LIF [43] adopted temporalwise attention to improve the updating of GRUs and LSTMs" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13960" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 197, + 256 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 197, + 256 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 197, + 256 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 197, + 256 + ], + "type": "image", + "image_path": "19976e25b16ba8dd63c522273cf82e127768f49b4e706e65db2435bd85bc3366.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 259, + 184, + 266 + ], + "lines": [ + { + "bbox": [ + 72, + 259, + 184, + 266 + ], + "spans": [ + { + "bbox": [ + 72, + 259, + 184, + 266 + ], + "type": "text", + "content": "(a) Spatial-Temporal Attention Aggregator SNN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 163, + 306, + 170 + ], + "lines": [ + { + "bbox": [ + 196, + 163, + 306, + 170 + ], + "spans": [ + { + "bbox": [ + 196, + 163, + 306, + 170 + ], + "type": "text", + "content": "(b) Structure of the Global Context (GC) block" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 73, + 306, + 163 + ], + "blocks": [ + { + "bbox": [ + 197, + 73, + 306, + 163 + ], + "lines": [ + { + "bbox": [ + 197, + 73, + 306, + 163 + ], + "spans": [ + { + "bbox": [ + 197, + 73, + 306, + 163 + ], + "type": "image", + "image_path": "313680c231220fa98d255eba78d4c8e446f53ed0c29275a20bb5902d186f6464.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 197, + 171, + 304, + 258 + ], + "blocks": [ + { + "bbox": [ + 197, + 171, + 304, + 258 + ], + "lines": [ + { + "bbox": [ + 197, + 171, + 304, + 258 + ], + "spans": [ + { + "bbox": [ + 197, + 171, + 304, + 258 + ], + "type": "image", + "image_path": "4833dfb8c1ffb8fa787f4afbd2c2e593413ef88f869e9269a71e71c6c798aa92.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 259, + 298, + 266 + ], + "lines": [ + { + "bbox": [ + 200, + 259, + 298, + 266 + ], + "spans": [ + { + "bbox": [ + 200, + 259, + 298, + 266 + ], + "type": "text", + "content": "(c) Structure of Step Attention (SA) block" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 278, + 441, + 290 + ], + "lines": [ + { + "bbox": [ + 167, + 278, + 441, + 290 + ], + "spans": [ + { + "bbox": [ + 167, + 278, + 441, + 290 + ], + "type": "text", + "content": "Figure 1. Overview of the STAA-SNN Architecture and the TSRD strategy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 73, + 555, + 243 + ], + "blocks": [ + { + "bbox": [ + 306, + 73, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 306, + 73, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 555, + 243 + ], + "type": "image", + "image_path": "c09ac145f0551f349096536c3fdd48b04343d57e48af640729277fa3515a0591.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "lines": [ + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "spans": [ + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "text", + "content": "(d) STAA-SNN with TSRD strategy. The dropout probability is represented by " + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "text", + "content": ", meaning that at each time step " + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "text", + "content": ", there is a probability " + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 311, + 246, + 539, + 266 + ], + "type": "text", + "content": " of selecting the simple addition aggregation method and discarding the STAA." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "type": "text", + "content": "within SNNs. More recently, researchers have explored incorporating SNNs into Transformer architectures, achieving promising results on multiple benchmark datasets[41, 52]. These researches reveal that attention mechanisms can significantly improve the performance of SNN. However, there remains a need for more specialized and high-performance attention modules tailored explicitly for SNN architectures. Addressing this gap can further enhance the efficiency and robustness of SNNs in real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 430, + 138, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 138, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 138, + 443 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": "In this section, we will first provide a brief overview of the Leaky Integrate-and-Fire (LIF) model and the training process. Following that, we introduce the proposed Spatial-Temporal Attention Aggregator. Lastly, we outline the timestep random dropout strategy for SNN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 517, + 230, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 230, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 230, + 529 + ], + "type": "text", + "content": "3.1. Leaky Integrate-and-Fire Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 534, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 583 + ], + "type": "text", + "content": "The spiking neuron is a fundamental component of SNNs. In this work, we adopt the Leaky Integrate-and-Fire (LIF) neuron due to its efficiency and simplicity. The discrete-time, iterative update rule of the LIF model is given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 590, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 590, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 68, + 590, + 295, + 613 + ], + "type": "interline_equation", + "content": "V ^ {t, n} = H ^ {t - 1, n} + \\frac {1}{\\tau} \\left[ I ^ {t - 1, n} - \\left(H ^ {t - 1, n} - V _ {\\text {r e s e t}}\\right) \\right], \\tag {1}", + "image_path": "b6d630714291b8cff7bde86591d5039791d08f56506f0e40be562f41631983e2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 614, + 294, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 614, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 69, + 614, + 294, + 628 + ], + "type": "interline_equation", + "content": "S ^ {t, n} = \\Theta \\left(V ^ {t, n} - v _ {\\mathrm {t h}}\\right), \\tag {2}", + "image_path": "ced3b9e6904d6dc3c44c96565318519469e453e3097b26c332a4ec05c7a67835.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 631, + 294, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 631, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 69, + 631, + 294, + 644 + ], + "type": "interline_equation", + "content": "H ^ {t, n} = V _ {\\text {r e s e t}} \\cdot S ^ {t, n} + V ^ {t, n} \\odot (1 - S ^ {t, n}). \\tag {3}", + "image_path": "d7eaa3e0dfd41aec888369c3b1c602db6e9a694abfd4c605effd2a18e47d6b7b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "H^{t - 1,n}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " denotes the membrane potential after a spike trigger at the previous timestep, while " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "I^{t,n}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "V^{t,n}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " represent the input and membrane potential of the " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "-th layer at timestep " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ", respectively. The threshold " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "v_{th}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " determines whether a spike is emitted, producing the binary spike sen" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "text", + "content": "sor " + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "inline_equation", + "content": "S^{t,n}" + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "text", + "content": ". The Heaviside step function " + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "text", + "content": " outputs 1 if " + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "inline_equation", + "content": "x \\geq 0" + }, + { + "bbox": [ + 313, + 311, + 554, + 334 + ], + "type": "text", + "content": " and 0 otherwise." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 335, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 384 + ], + "type": "text", + "content": "Given the similarity between iterative LIF dynamics and Recurrent Neural Networks (RNNs), Spatio-Temporal Backpropagation (STBP) [57, 58] enables direct SNN training via Backpropagation Through Time (BPTT):" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 397, + 553, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 397, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 318, + 397, + 553, + 434 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial W ^ {n}} = \\sum_ {t} \\left(\\frac {\\partial L}{\\partial S ^ {t , n}} \\cdot \\frac {\\partial S ^ {t , n}}{\\partial V ^ {t , n}} + \\frac {\\partial L}{\\partial V ^ {t + 1 , n}} \\cdot \\frac {\\partial V ^ {t + 1 , n}}{\\partial V ^ {t , n}}\\right) \\frac {\\partial V ^ {t , n}}{\\partial W ^ {n}}. \\tag {4}", + "image_path": "dea5335ca99be0dbfe7ebd4cfef1190943e4e3fbccca0b29134210c834a2bb04.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": "However, the term " + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\frac{\\partial S^{t,n}}{\\partial V^{t,n}}" + }, + { + "bbox": [ + 313, + 434, + 554, + 483 + ], + "type": "text", + "content": ", which represents the derivative of the spike function, is non-differentiable. To address this, prior works introduce Surrogate Gradients(SG), commonly approximated using a rectangular function:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 364, + 491, + 553, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 491, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 364, + 491, + 553, + 515 + ], + "type": "interline_equation", + "content": "\\frac {\\partial S ^ {t}}{\\partial V ^ {t}} = \\frac {1}{a} \\cdot \\operatorname {s i g n} \\left(\\left| V ^ {t} - V _ {\\mathrm {t h}} \\right| < \\frac {a}{2}\\right), \\tag {5}", + "image_path": "b77a82fabc00737cc457d79144fb012dd430191af8d1d65925d77b9ba5d8df5e.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "type": "text", + "content": " is a hyperparameter and set to 1. This approximation provides a gradient of 1 within the range " + }, + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{th}} - 0.5 \\leq V^t \\leq V_{\\mathrm{th}} + 0.5" + }, + { + "bbox": [ + 313, + 517, + 554, + 552 + ], + "type": "text", + "content": ", and otherwise 0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": "Neuromorphic-Inspired Adaptive LIF Design. Examining Eq. (1), it is evident that keeping " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{reset}}" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " constant results in fixed weight coefficients for " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": ", which remain unchanged across timestep " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": ". This simplification deviates from biological neural processing, where neuronal excitability dynamically adapts to stimuli. To bridge this gap, we propose an adaptive modification to the membrane potential update rule:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 358, + 673, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 673, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 358, + 673, + 553, + 685 + ], + "type": "interline_equation", + "content": "V ^ {t, n} = M \\odot H ^ {t - 1, n} + N \\odot I ^ {t - 1, n}. \\tag {6}", + "image_path": "188f91b449ef58aa51c62271f5070b4665fcd1396b1f9650974c126aaf7c05b4.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " are coefficient matrices that dynamically adjust through iterative learning, regulating activation and" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13961" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "inhibition probabilities across layers. This biologically inspired adaptation enhances information processing efficiency and robustness in SNNs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": "Global Context (GC) Block. The Global Context (GC) Block serves as a lightweight alternative to the self-attention mechanism used in Transformers, enabling the integration of self-attention into conventional convolution-based deep learning architectures without the computational overhead of full Transformer models. This makes it possible to enhance feature learning in non-Transformer structures while maintaining efficiency and precision. In this paper, we simulate the self-attention operation using a GC block [9], which consists of three " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": " convolutional kernels corresponding to key, query, and value. As depicted in Fig. 1(b), the spatial spiking input tensor of the " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": "-th layer at the " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": "-th timestep is defined as " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "X^{t,n} \\in \\mathbb{R}^{C^n \\times H^n \\times W^n}" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "W^n" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "H^n" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": " denote the width and height of the feature map, and " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "C^n" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": " represents the number of channels. Passing " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "X^{t,n}" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": " through the GC block yields three weight matrices for self-attention, namely " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "W_k" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "W_q" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "inline_equation", + "content": "W_v" + }, + { + "bbox": [ + 55, + 124, + 296, + 328 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 337, + 295, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 337, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 97, + 337, + 295, + 350 + ], + "type": "interline_equation", + "content": "W _ {k} = X ^ {t, n} \\cdot \\operatorname {S i g m o i d} \\left(\\operatorname {C o n v} _ {k} \\left(X ^ {t, n}\\right)\\right), \\tag {7}", + "image_path": "7dd2684c496d8320b7249a040e5456f228e4f95f27d6b5d350d3be7a47679715.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 353, + 294, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 353, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 99, + 353, + 294, + 365 + ], + "type": "interline_equation", + "content": "W _ {q} = \\operatorname {C o n v} _ {q} \\left(W _ {k}\\right), \\tag {8}", + "image_path": "abbfede8226f13b5e99be3b4fce7e7da56e1ccf6666c3867e30b35571f383db5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 367, + 294, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 367, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 99, + 367, + 294, + 380 + ], + "type": "interline_equation", + "content": "W _ {v} = \\operatorname {C o n v} _ {v} (\\operatorname {R e L U} (\\operatorname {L N} (W _ {q}))). \\tag {9}", + "image_path": "112a0d621263f75d67967ecc4a3bf9a80278d116462ccf8d18741fad3f26101c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "content": "Before computing " + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "inline_equation", + "content": "X^{t,n}" + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "content": " is reshaped from " + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "inline_equation", + "content": "C\\times H\\times W" + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "inline_equation", + "content": "C\\times HW" + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "content": ". The resulting matrices " + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "inline_equation", + "content": "W_{k},W_{q},W_{v}\\in \\mathbb{R}^{C\\times 1\\times 1}" + }, + { + "bbox": [ + 55, + 390, + 296, + 449 + ], + "type": "text", + "content": " encode self-attention weights, which are then used to refine the input feature map. Subsequently, the enhanced output is computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 449, + 294, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 449, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 133, + 449, + 294, + 463 + ], + "type": "interline_equation", + "content": "Z ^ {t, n} = X ^ {t, n} \\oplus W _ {v}, \\tag {10}", + "image_path": "ea9b184f13fb828652b8a3f32cc2830ec5781b47b27101492bfcf08e90979b9f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "inline_equation", + "content": "Z^{t,n} \\in \\mathbb{R}^{C,H,W}" + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "text", + "content": " maintains the same dimensions as " + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "inline_equation", + "content": "X^{t,n}" + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "text", + "content": ". The refined feature map " + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "inline_equation", + "content": "Z^{t,n}" + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "text", + "content": ", in comparison to the " + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "inline_equation", + "content": "X^{t,n}" + }, + { + "bbox": [ + 55, + 468, + 296, + 554 + ], + "type": "text", + "content": ", improves the network's ability to focus on self-relevant and task-critical information while suppressing irrelevant signals. This approach facilitates the seamless transmission of enriched data flow, enabling the retention of valuable features for subsequent stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 689 + ], + "type": "text", + "content": "Position Encoding (PE) Block. Position encoding is widely used in NLP and sequence modeling to incorporate positional information into input data. While SNNs share structural similarities with RNNs and inherently process temporal information, their reliance on sparse binary spikes for information storage can lead to significant feature loss. This study introduces position encoding to enhance spatial features while enriching temporal relationships, thereby improving feature retention and strengthening temporal representations. The positional encoding is computed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 129, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "X ^ {t, n} = I ^ {t, n} \\oplus \\operatorname {P o s} _ {t, n}, \\tag {11}", + "image_path": "d38858b897d813c475e9f3180025a68729ac08362cff2d801c1afbd275b07af0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "\\mathrm{Pos}_{t,n}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " represents the position encoding at the " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "-th time step in the " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "-th layer, with " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "\\mathrm{Pos}_{t,n} \\in \\mathbb{R}^{T,C^n}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "text", + "content": "Step Attention (SA) Block. While previous modules primarily enhance feature processing from a spatial perspective, SNNs uniquely process information across both spatial and temporal. Unlike other types of network, SNNs offer an additional understanding of information sensitivity across different timesteps. Therefore, in this paper, we introduce Step Attention to tackle the temporal challenges inherent in SNNs. As illustrated in Fig. 1(c), let " + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "inline_equation", + "content": "U^{t,n} \\in \\mathbb{R}^{C^n \\times H^n \\times W^n}" + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "text", + "content": " denote the spatially aggregated feature input at the " + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "text", + "content": "-th layer and the " + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 110, + 555, + 242 + ], + "type": "text", + "content": "-th timestep. The specific Step Attention mechanism is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 247, + 554, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 247, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 326, + 247, + 554, + 301 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V ^ {t, n} = U ^ {t, n} \\odot \\text {S i g m o i d} \\left(\\operatorname {C o n v} _ {2} \\left(\\operatorname {R e L U} \\left(\\operatorname {C o n v} _ {1} \\right. \\right. \\right. \\tag {12} \\\\ \\left. \\left(\\alpha \\cdot \\operatorname {A v g P o o l} (U ^ {t, n})\\right)\\right) \\Bigg) \\Bigg), \\\\ \\end{array}", + "image_path": "ac9f5f255e36ac62f5821f610ce96e0cdfef65c3d6aa7d5c603a26bea36a8c4d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "type": "text", + "content": " is a scaling factor that regulates the initial training speed for smoother convergence. In practice, " + }, + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 308, + 554, + 342 + ], + "type": "text", + "content": " is typically set to a default value of 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "type": "text", + "content": "LIF with Spatio-Temporal Attention Aggregator (STAA). The STAA seamlessly integrates with the LIF model, offering compatibility across various network structures without constraints. It adeptly harnesses the potential of spiking neurons, effectively optimizing the storage of discrete spike information while minimizing the loss of crucial data during transmission. As shown in Fig. 1(a), the neuronal update equation for the LIF model with STAA is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 472, + 553, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 472, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 356, + 472, + 553, + 486 + ], + "type": "interline_equation", + "content": "U ^ {t, n} = G C _ {1} \\left(X ^ {t, n}\\right) \\oplus G C _ {2} \\left(H ^ {t - 1, n}\\right), \\tag {13}", + "image_path": "20d146319fd0175301bfcebc36448c4f5f58b81bfbe6ecd28b58e30fe165d6e4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 358, + 488, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 488, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 358, + 488, + 553, + 502 + ], + "type": "interline_equation", + "content": "V ^ {t, n} = S A \\left(U ^ {t, n}\\right). \\tag {14}", + "image_path": "d21e466ee578f864edafff1ca9b70d1a73cb701cdd5da9cda526c74c3636e0f8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "inline_equation", + "content": "GC_{1}" + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "inline_equation", + "content": "GC_{2}" + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "text", + "content": " represent Global Context Blocks for the input " + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "inline_equation", + "content": "X^{t,n}" + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "text", + "content": " and the previous membrane state " + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "inline_equation", + "content": "H^{t-1,n}" + }, + { + "bbox": [ + 313, + 510, + 554, + 557 + ], + "type": "text", + "content": ", respectively. The remaining update equations follow the original LIF model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 564, + 550, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 550, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 550, + 578 + ], + "type": "text", + "content": "3.2. Time Step Random Dropout (TSRD) Strategy" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "To further enhance training efficiency, we propose TSRD, an adaptive training strategy designed to prevent premature feature stabilization in deep SNNs. In deeper timesteps, spiking features may solidify too early during iterations, hindering the network from achieving optimal convergence. As illustrated in Fig. 1(d), TSRD randomly dropout the augmentation module at timestep " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "t \\in (0,T)" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " with a dropout probability " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ", and instead choose the most basic elementwise addition aggregation method. The TSRD technique accelerates training and helps the network escape local optima, thereby improving generalization performance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13962" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 69, + 536, + 482 + ], + "blocks": [ + { + "bbox": [ + 74, + 69, + 536, + 482 + ], + "lines": [ + { + "bbox": [ + 74, + 69, + 536, + 482 + ], + "spans": [ + { + "bbox": [ + 74, + 69, + 536, + 482 + ], + "type": "table", + "html": "
MethodArchitectureTimestepAccuracy (%)
CIFAR-10CIFAR-100
CNN-basedGLIF [70]ResNet-19294.44%75.48%
ResNet-19494.85%77.05%
TET [16]ResNet-19294.16%72.87%
ResNet-19494.44%74.47%
LSG [42]ResNet-19294.41%76.32%
ResNet-19495.17%76.85%
PFA [14]ResNet-19295.6%76.7%
ResNet-19495.71%78.1%
Diet-SNN [50]ResNet-20591.78%64.07%
VGG-16593.85%69.67%
IM-loss [22]VGG-16593.85%70.18%
IM-LIF [43]ResNet-19395.29%77.21%
MPBN [26]ResNet-19196.06%78.71%
ResNet-19296.47%79.51%
ResNet-19496.52%80.1%
ResNet-20192.22%68.41%
ResNet-20293.54%70.79%
ResNet-20494.28%72.3%
Transformer-basedSpikformer[75]Spikformer-4-25493.94%75.96%
Spikingformer[74]Spikingformer-4-25494.77%77.43%
Spike-driven Transformer [69]Transformer-2-512495.60%78.40%
CNN-basedOursVGG-13294.70%±0.12%75.16%±0.08%
VGG-13495.26%±0.09%76.35%±0.11%
ResNet-19196.75%±0.09%79.37%±0.10%
ResNet-19296.85%±0.11%80.57%±0.11%
ResNet-19497.14%±0.10%82.05%±0.10%
ResNet-20193.08%±0.09%70.14%±0.07%
ResNet-20294.35%±0.08%73.20%±0.12%
ResNet-20495.03%±0.11%75.10%±0.12%
", + "image_path": "4f65df8535f0380689328b8e5132b5d9bd6225061619d36e376f7ab6b5c24163.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 179, + 489, + 428, + 500 + ], + "lines": [ + { + "bbox": [ + 179, + 489, + 428, + 500 + ], + "spans": [ + { + "bbox": [ + 179, + 489, + 428, + 500 + ], + "type": "text", + "content": "Table 1. Comparison results with SOTA methods on CIFAR-10/100." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 541, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 638 + ], + "type": "text", + "content": "We first outlined the experimental setup, including dataset details and implementation specifics. We then compare our results with previous SOTA methods across multiple benchmarks. Next, we perform ablation studies to assess the effectiveness of STAA. Finally, we visualize the model's performance. Further details on datasets, hyperparameters, additional experiments, and computational efficiency analysis are provided in the Supplementary Material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 647, + 173, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 173, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 173, + 661 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate our method on five benchmark datasets. CIFAR-10/100[35] are standard image classification datasets with 50,000 training and 10,000 test images across 10 and 100 classes, respectively. ImageNet[15] is a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "content": "large-scale dataset containing 1.2 million training, 50,000 validation, and 100,000 test images spanning 1,000 categories. For neuromorphic vision tasks, CIFAR10-DVS[39, 40, 45] consists of 10,000 event streams across 10 classes, with a 9:1 training-to-testing split. DVS128 Gesture[4] is an event-based gesture recognition dataset with 1,176 training and 288 test samples, featuring 11 gesture classes recorded from 29 subjects under three lighting conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we set the firing threshold " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " to 1, and initialize " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " to 2 and 0.1, respectively. In the PE block, we adopt a learnable position encoding method, initializing all values to zero. All experiments are implemented in PyTorch and conducted on an NVIDIA RTX 3090 GPU, except for ImageNet, which is" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13963" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 202 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 202 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 202 + ], + "type": "image", + "image_path": "3802872e9f7760e37689e04a0d3a4b7edd04d9f5581501ea48585c103e5cb6f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 210, + 392, + 222 + ], + "lines": [ + { + "bbox": [ + 218, + 210, + 392, + 222 + ], + "spans": [ + { + "bbox": [ + 218, + 210, + 392, + 222 + ], + "type": "text", + "content": "Figure 2. Position encoding locations in SNNs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 240, + 293, + 475 + ], + "blocks": [ + { + "bbox": [ + 59, + 240, + 293, + 475 + ], + "lines": [ + { + "bbox": [ + 59, + 240, + 293, + 475 + ], + "spans": [ + { + "bbox": [ + 59, + 240, + 293, + 475 + ], + "type": "table", + "html": "
MethodArchitectureTimestepAccuracy
STBP-tdBN [73]ResNet34663.72%
TET [16]ResNet34664.79%
RecDis-SNN [23]ResNet34667.33%
GLIF [70]ResNet34467.52%
IM-Loss [22]ResNet18667.43%
Real Spike [24]ResNet18463.68%
ResNet34467.69%
RMP-Loss [25]ResNet18463.03%
ResNet34465.17%
MPBN [26]ResNet18463.14%
ResNet34464.71%
SEW ResNet [17]ResNet18463.18%
ResNet34467.04%
OursResNet18468.27% ± 0.19%
ResNet34470.40% ± 0.15%
", + "image_path": "e304d8bc4eb444ab1f58fb82967f3803501ee63483aebd5bee4b45fa1f52c77d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 554 + ], + "type": "text", + "content": "trained on 8 NVIDIA RTX 4090 GPUs. The total training epochs are 500 for CIFAR-10, CIFAR-100, and CIFAR10-DVS, while ImageNet runs for 300 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 200, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 200, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 200, + 577 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Static Image Classification. We evaluate our model on three static datasets: CIFAR-10, CIFAR-100, and ImageNet. For CIFAR-10/100, STAA-SNNs are applied to both ResNet-like and VGG-like networks and tested at 1, 2, 3, and 4 timesteps. On ImageNet, the model is evaluated using ResNet-18 and ResNet-34 with 4 time steps. Our method achieves top accuracies of " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "97.14\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " on CIFAR-10, " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "82.05\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " on CIFAR-100, and " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "70.40\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " on ImageNet, surpassing previous state-of-the-art methods across various architectures. These improvements are attributed to the strong spatial-temporal attention mechanism and feature aggrega" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 240, + 553, + 553 + ], + "blocks": [ + { + "bbox": [ + 57, + 483, + 293, + 495 + ], + "lines": [ + { + "bbox": [ + 57, + 483, + 293, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 483, + 293, + 495 + ], + "type": "text", + "content": "Table 2. Comparison of training based SNN SOTA on ImageNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 240, + 553, + 553 + ], + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 553 + ], + "type": "table", + "html": "
MethodsArchitectureTimestepAccuracy
CIFAR10-DVS
IM-loss [22]ResNet-191072.60%
LSG [42]ResNet-191077.90%
MPBN [26]ResNet-191074.40%
MPBN [26]ResNet-201078.70%
TET [16]VGG5NN1077.30%
IM-LIF [43]VGG-131080.50%
GLIF [70]7B-wideNet1678.10%
STSA [56]STS-Transformer1679.93%
SEW [17]SEW-ResNet1674.4%
Spikeformer [41]Spikeformer1680.9%
OursResNet-201681.90%±0.20%
VGG-131682.10%±0.20%
DVS128 Gesture
STBP-tdBN [73]ResNet-174096.87%
SEW [17]7B-Net1697.52%
PLIF [18]PLIFNet2097.57%
MA-SNN [66]5 layers SCNN2098.23%
ASA-SNN [67]5 layers SCNN2097.70%
IM-LIF [43]ResNet-194097.33%
LIAF+TA [65]TA-SNN-Net6098.61%
Spike-driven-Transformer [69]Transformer-2-5121699.30%
OursVGG-131698.61%±0.20%
", + "image_path": "f25fa7be0e716880e0768fed6bacdf45ae6d3ad16849477e87d1a1803cf84be2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 561, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 583 + ], + "type": "text", + "content": "Table 3. Comparison results with SOTA methods on CIFAR10-DVS and DVS128 Gesture." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 607, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 668 + ], + "type": "text", + "content": "tion capabilities of STAA-SNNs, which fully leverage the potential performance of SNNs. Furthermore, our model achieves equivalent accuracy while requiring at least one fewer timestep compared to prior methods, improving efficiency without sacrificing performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Event-based Action Recognition. To further assess the spatial-temporal processing capabilities of STAA-SNNs," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13964" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 282, + 184 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 282, + 184 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 282, + 184 + ], + "type": "image", + "image_path": "8784705889bb5c7d12c031d70e074f6f6cdf83688940b7660fad31a5fd4c3cc9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 295, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 217 + ], + "type": "text", + "content": "Figure 3. Distribution of accuracy with different dropout probability " + }, + { + "bbox": [ + 55, + 194, + 295, + 217 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 194, + 295, + 217 + ], + "type": "text", + "content": " in TSRD on CIFAR-10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 228, + 293, + 334 + ], + "blocks": [ + { + "bbox": [ + 59, + 228, + 293, + 334 + ], + "lines": [ + { + "bbox": [ + 59, + 228, + 293, + 334 + ], + "spans": [ + { + "bbox": [ + 59, + 228, + 293, + 334 + ], + "type": "table", + "html": "
STAA ComponentsTSRDAccuracyAccuracy Gain
GCPESA
XXXX72.30%-
✓XXX73.22%+0.92%
✓✓XX73.79%+1.49%
✓✓X✓73.96%+1.66%
✓✓✓X74.78%+2.48%
✓✓✓✓75.10%+2.80%
", + "image_path": "164822d45e7592927d8c9f376171a1d403b8a9e630d5cc3cd5165d886620c70a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 342, + 286, + 353 + ], + "lines": [ + { + "bbox": [ + 64, + 342, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 64, + 342, + 286, + 353 + ], + "type": "text", + "content": "Table 4. Combinations of sub-modules in STAA with TSRD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "type": "text", + "content": "we conduct experiments on the DVS128 Gesture and CIFAR10-DVS datasets. Unlike static datasets, these neuromorphic datasets contain temporal event streams, requiring models to efficiently process spiking sequences. As shown in Tab. 3, our model achieves state-of-the-art or superior performance with a smaller network and fewer timesteps. Specifically, STAA-SNNs achieve " + }, + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "type": "inline_equation", + "content": "82.10\\%" + }, + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "type": "text", + "content": " accuracy on CIFAR10-DVS and " + }, + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "type": "inline_equation", + "content": "98.61\\%" + }, + { + "bbox": [ + 54, + 374, + 295, + 495 + ], + "type": "text", + "content": " on the DVS128 Gesture dataset, demonstrating their effectiveness in event-based action recognition tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 149, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 149, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 149, + 514 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 293, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 293, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 293, + 532 + ], + "type": "text", + "content": "All ablation studies with " + }, + { + "bbox": [ + 55, + 519, + 293, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 4" + }, + { + "bbox": [ + 55, + 519, + 293, + 532 + ], + "type": "text", + "content": " using ResNet-20 as baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Position encoding placement. To optimize PE in SNNs, we evaluate its best placement within the network. Given the structural complexity of SNNs compared to CNNs, identifying the most effective integration point is crucial. As shown in Fig. 2, we evaluate four placement strategies (b)-(e) alongside a baseline SNN without PE (a), ensuring all other conditions remain constant. The CIFAR-10 accuracies for (b) to (e) are " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "94.08\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "94.29\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "94.27\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "93.96\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", respectively. The highest performance occurs when PE is placed at the pre-aggregation spatial input port (c). Although (b) and (d) yield similar results, (d) introduces additional parameters without significant gains. Thus, placing PE at the pre-aggregation spatial input port achieves the best balance between accuracy and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "Evaluation of position encoding methods. We next evaluate different encoding methods to determine the most effective approach. The results on CIFAR-10 indicate that learnable PE achieves an accuracy of " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "94.99\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", outperforming fixed PE, which attains " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "94.42\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ". This " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " improvement highlights the advantage of adaptive position encoding, as it enables the network to dynamically adjust spatial representations, leading to better feature alignment and overall performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "text", + "content": "Impact of dropout probability " + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "text", + "content": " in TSRD. To accelerate training and enhance generalization in TSRD, it is essential to determine an optimal dropout probability " + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "text", + "content": ". Excessive dropout may lead to network collapse by discarding too much information, while insufficient dropout may limit its regularization effect. As shown in Fig. 3, by incrementally increasing " + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "text", + "content": " from 0.1 to 1.0, the model achieves relatively high performance when " + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "inline_equation", + "content": "\\beta \\leq 0.3" + }, + { + "bbox": [ + 313, + 194, + 556, + 374 + ], + "type": "text", + "content": ", and achieving optimal at 0.1. These results confirm that excessive dropout removes essential information, degrading performance, while a balanced dropout probability effectively suppresses noise in deeper timesteps while preserving critical information in shallower steps. This enables the network to navigate critical learning stages more effectively, enhancing both stability and robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "text", + "content": "Combinations of sub-modules in STAA with TSRD. To evaluate the contribution of each sub-module in STAA, as well as the impact of TSRD, we conduct an ablation study by progressively integrating these components into the SNN. The results on CIFAR-100 are presented in Tab. 4. Notably, the GC block alone improves accuracy by " + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "inline_equation", + "content": "0.92\\%" + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "text", + "content": ", while the SA block contributes an additional " + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "text", + "content": ", highlighting the importance of spatio-temporal attention and feature aggregation reconstruction in enhancing SNN performance. When all components, including TSRD, are incorporated, the accuracy reaches " + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "inline_equation", + "content": "75.10\\%" + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "text", + "content": ", demonstrating a cumulative improvement of " + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "inline_equation", + "content": "2.80\\%" + }, + { + "bbox": [ + 313, + 388, + 556, + 533 + ], + "type": "text", + "content": " over the baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Impact of Intermediate Dimension Scaling Coefficients in GC. The GC module incorporates multiple " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " convolutions for feature transformation. Retaining the original channel size " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " for the last two convolutions layers would significantly increase computational overhead. To maintain the module's efficiency while preserving performance, a scaling coefficient " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " is introduced within the GC to compress feature dimensions. By applying this scaling mechanism, the parameter count is reduced from " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "(C \\cdot C)" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "(2C \\cdot C / r)" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ", making the module more computationally feasible. We conduct experiments to assess the impact of different " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " values, as shown in Fig. 5. The results indicate that the model achieves optimal performance when " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ", effectively balancing computational efficiency and accuracy." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13965" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 73, + 519, + 194 + ], + "blocks": [ + { + "bbox": [ + 93, + 73, + 519, + 194 + ], + "lines": [ + { + "bbox": [ + 93, + 73, + 519, + 194 + ], + "spans": [ + { + "bbox": [ + 93, + 73, + 519, + 194 + ], + "type": "image", + "image_path": "db4f89a7409725552287a7029ebf952e759f4ab8d4905b78a495427f0ac68c63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 201, + 482, + 213 + ], + "lines": [ + { + "bbox": [ + 127, + 201, + 482, + 213 + ], + "spans": [ + { + "bbox": [ + 127, + 201, + 482, + 213 + ], + "type": "text", + "content": "Figure 4. Visualization on CIFAR10-DVS. Ten layers from VGG-13 in a shallow to deep manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 233, + 276, + 340 + ], + "blocks": [ + { + "bbox": [ + 75, + 233, + 276, + 340 + ], + "lines": [ + { + "bbox": [ + 75, + 233, + 276, + 340 + ], + "spans": [ + { + "bbox": [ + 75, + 233, + 276, + 340 + ], + "type": "image", + "image_path": "f98014b62ed6b38cfa5b66c7ef1f002197aabee5d4150248636ff226ecff292d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "type": "text", + "content": "Figure 5. Impact of different scaling coefficients " + }, + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "type": "text", + "content": " for intermediate feature dimensions in the GC module on the CIFAR-10 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 394, + 202, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 202, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 202, + 407 + ], + "type": "text", + "content": "4.4. Visualization and Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": "We aggregate 4D " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "(T,C,H,W)" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " spiking maps into 2D " + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 55, + 412, + 296, + 521 + ], + "type": "text", + "content": " and apply Grad-CAM to visualize feature extraction at different depths. As shown in Fig. 4, the left displays a sample from the horse class in CIFAR10-DVS, while the right compares vanilla LIF and STAA-LIF feature maps. STAA-LIF focuses more effectively on target-relevant features, capturing finer details like the horse's mouth, which the vanilla LIF model misses. This highlights the advantage of spatio-temporal attention in enhancing feature selectivity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 244, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 244, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 244, + 541 + ], + "type": "text", + "content": "4.5. Analysis of Computation Efficiency" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the single-image inference energy cost using 45-nm technology. ANNs rely on multiplication-and-accumulation (MAC) operations, which consume " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "4.6\\mathrm{pJ}" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " per operation, whereas SNNs primarily use accumulation (AC) operations at a significantly lower cost of " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.9\\mathrm{pJ}" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " per operation, making SNNs inherently more energy-efficient [31, 49]. Following [11], we assess the efficiency of ResNet-20, which requires 0.10 billion ACs, 0.06 billion MACs, 0.87 billion FLOPs, and consists of 12.69 million parameters. Under these conditions, inference on CIFAR-100 (with a " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " resolution and " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{T} = 4" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": ") results in an energy consumption of " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.366\\mathrm{mJ}" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": ", demonstrating the energy-efficient nature of ResNet-20. Additional computational efficiency experiments are provided in the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 232, + 388, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 232, + 388, + 245 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 388, + 245 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 254, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 613 + ], + "type": "text", + "content": "In this paper, we introduce STAA-SNN, a spatio-temporal attention-based feature aggregator, together with the TSRD training strategy. STAA-SNN leverages the inherent ability of SNNs to process spatio-temporal information, thereby enhancing network performance while reducing the number of required timesteps. Our approach achieves state-of-the-art results across multiple benchmarks. STAA-SNN is designed as a plug-and-play module that seamlessly integrates into convolution-based SNN architectures. It is composed of several synergistic modules that collectively enhance its capabilities. The GC module simulates the brain's ability to process information across a broad spatial range, involving multiple brain regions in coordination. By focusing on spatial feature relevance, it enables the model to capture global structures and key stimuli in its environment. The PE module enhances temporal encoding by incorporating time-step information into feature representations, mirroring the brain's ability to integrate temporal and spatial cues for improved understanding of time-dependent relationships. The SA module mimics the brain's dynamic attention mechanisms, adjusting feature importance across time steps by adaptively enhancing or suppressing information. This prioritization of critical temporal features strengthens decision-making and overall network efficiency. Together, these modules form a robust spatio-temporal feature enhancement system that improves cognitive processing, robustness, and overall model performance. Despite these promising results, future research should focus on reducing computational energy consumption to further improve efficiency, particularly for resource-constrained applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 627, + 429, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 429, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 429, + 640 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 554, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 703 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China (NSFC) (62476035, 62206037, 62276230, 62271361), and Natural Science Foundation of Zhejiang Province (LDT23F02023F02), and State Key Laboratory (SKL) of Biobased Transportation Fuel Technology." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13966" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Larry F Abbott. Lapicque's introduction of the integrate-and-fire model neuron (1907). *Brain research bulletin*, 50(5-6): 303–304, 1999. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 203 + ], + "type": "text", + "content": "[2] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 237 + ], + "type": "text", + "content": "[3] Pepe Alcamí and Alberto E Pereda. Beyond plasticity: the dynamic impact of electrical synapses on neural circuits. Nature Reviews Neuroscience, 20(5):253-271, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 305 + ], + "type": "text", + "content": "[4] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7243-7252, 2017. 5, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 307, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 338 + ], + "type": "text", + "content": "[5] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lui, and Cordelia Schmid. Vivit: A video vision transformer. 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "type": "text", + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "type": "text", + "content": "[7] Guo-qiang Bi and Mu-ming Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. Journal of neuroscience, 18(24):10464-10472, 1998. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 421, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 497 + ], + "type": "text", + "content": "[8] Wuque Cai, Hongze Sun, Rui Liu, Yan Cui, Jun Wang, Yang Xia, Dezhong Yao, and Daqing Guo. A Spatial-Channel-Temporal-Fused Attention for Spiking Neural Networks. IEEE Transactions on Neural Networks and Learning Systems, pages 1–15, 2023. Conference Name: IEEE Transactions on Neural Networks and Learning Systems. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 499, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 295, + 553 + ], + "type": "text", + "content": "[9] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 612 + ], + "type": "text", + "content": "[10] Biswadeep Chakraborty, Xueyuan She, and Saibal Mukhopadhyay. A Fully Spiking Hybrid Neural Network for Energy-Efficient Object Detection. IEEE Transactions on Image Processing, 30:9014-9029, 2021. Conference Name: IEEE Transactions on Image Processing. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "text", + "content": "[11] Guangyao Chen, Peixi Peng, Guoqi Li, and Yonghong Tian. Training Full Spike Neural Networks via Auxiliary Accumulation Pathway, 2023. arXiv:2301.11929 [cs]. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 713 + ], + "type": "text", + "content": "[12] Dennis V Christensen, Regina Dittmann, Bernabe Linares-Barranco, Abu Sebastian, Manuel Le Gallo, Andrea Redaelli, Stefan Slesazeck, Thomas Mikolajick, Sabina Spiga, Stephan Menzel, et al. 2022 roadmap on neuromorphic computing and engineering. Neuromorphic Computing and Engineering, 2(2):022501, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. Ieee Micro, 38(1):82-99, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[14] Haoyu Deng, Ruijie Zhu, Xuerui Qiu, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tensor decomposition based attention module for spiking neural networks. Knowledge-Based Systems, 295:111780, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. ISSN: 1063-6919. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 274 + ], + "type": "text", + "content": "[16] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 376 + ], + "type": "text", + "content": "[18] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 409 + ], + "type": "text", + "content": "[19] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 466 + ], + "type": "text", + "content": "[20] Banglei Guan and Ji Zhao. Relative pose estimation for multi-camera systems from point correspondences with scale ratio. In Proceedings of the 30th ACM International Conference on Multimedia, page 5036-5044, New York, NY, USA, 2022. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[21] Banglei Guan, Ji Zhao, Zhang Li, Fang Sun, and Friedrich Fraundorfer. Minimal solutions for relative pose with a single affine correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[22] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems, 35:156-166, 2022. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 634 + ], + "type": "text", + "content": "[23] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks. In Proceedings of CVPR, pages 326-335, New Orleans, LA, USA, 2022. IEEE. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 690 + ], + "type": "text", + "content": "[24] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In European Conference on Computer Vision, pages 52-68. Springer, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[25] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. RMP" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13967" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "Loss: Regularizing Membrane Potential Distribution for Spiking Neural Networks, 2023. arXiv:2308.06787 [cs]. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[26] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19420-19430, 2023. 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "text", + "content": "[27] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 272 + ], + "type": "text", + "content": "[28] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558-13567, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 319 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 342 + ], + "type": "text", + "content": "[30] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology press, 2005. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[31] Mark Horowitz. 1.1 Computing's energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pages 10-14, 2014. ISSN: 2376-8606. 8, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[32] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "text", + "content": "[33] J. Hu, L. Shen, S. Albanie, G. Sun, and E. Wu. Squeeze-and-excitation networks. IEEE transactions on pattern analysis and machine intelligence, 42(8):2011-2023, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[34] Yifan Hu, Yujie Wu, Lei Deng, and Guoqi Li. Advancing residual learning towards powerful deep spiking neural networks. arXiv e-prints, pages arXiv-2112, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "text", + "content": "[35] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 5, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 588 + ], + "type": "text", + "content": "[36] Souvik Kundu, Gourav Datta, Massoud Pedram, and Peter A Beerel. Spike-thrift: Towards energy-efficient deep spiking neural networks by limiting spiking activity via attention-guided compression. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3953-3962, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[37] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Spiking neural networks and bio-inspired supervised deep learning: A survey. arXiv preprint arXiv:2307.16235, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[38] Chankyu Lee, Syed Shakib Sarwar, Priyadarshini Panda, Gopalakrishnan Srinivasan, and Kaushik Roy. Enabling spike-based backpropagation for training deep neural network architectures. Frontiers in neuroscience, page 119, 2020. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Taihang Lei, Banglei Guan, Minzu Liang, Zibin Liu, Jianbin Liu, Yang Shang, and Qifeng Yu. Motion measurements of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "explosive shock waves based on an event camera. Optics Express, 32(9):15390-15409, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[40] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017. 5, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[41] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: A novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 184, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 555, + 239 + ], + "type": "text", + "content": "[42] Shuang Lian, Jiangrong Shen, Qianhui Liu, Ziming Wang, Rui Yan, and Huajin Tang. Learnable surrogate gradient for direct training spiking neural networks. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3002–3010, 2023. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 316 + ], + "type": "text", + "content": "[43] Shuang Lian, Jiangrong Shen, Ziming Wang, and Huajin Tang. IM-LIF: Improved Neuronal Dynamics With Attention Mechanism for Direct Training Deep Spiking Neural Network. IEEE Transactions on Emerging Topics in Computational Intelligence, pages 1-11, 2024. Conference Name: IEEE Transactions on Emerging Topics in Computational Intelligence. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 358 + ], + "type": "text", + "content": "[44] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[45] Zibin Liu, Banglei Guan, Yang Shang, Qifeng Yu, and Laurent Kneip. Line-based 6-dof object pose estimation and tracking with an event camera. IEEE Transactions on Image Processing, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "type": "text", + "content": "[46] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36 (6):51-63, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 513 + ], + "type": "text", + "content": "[47] Kleanthis C Neokleous, Marios N Avraamides, Costas K Neocleous, and Christos N Schizas. Selective attention and consciousness: investigating their relation through computational modelling. Cognitive Computation, 3:321-331, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 569 + ], + "type": "text", + "content": "[48] Garrick Orchard, E Paxon Frady, Daniel Ben Dayan Rubin, Sophia Sanborn, Sumit Bam Shrestha, Friedrich T Sommer, and Mike Davies. Efficient neuromorphic signal processing with loihi 2. In 2021 IEEE Workshop on Signal Processing Systems (SiPS), pages 254-259. IEEE, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 571, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 633 + ], + "type": "text", + "content": "[49] Ning Qiao, Hesham Mostafa, Federico Corradi, Marc Osswald, Fabio Stefanini, Dora Sumislawka, and Giacomo Indiveri. A reconfigurable on-line learning spiking neuromorphic processor comprising 256 neurons and 128K synapses. Frontiers in Neuroscience, 9, 2015. Publisher: Frontiers. 8, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "text", + "content": "[50] Nitin Rathi and Kaushik Roy. Diet-snn: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[51] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13968" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[52] Chen She and Laiyun Qing. Spikeformer: Image reconstruction from the sequence of spike camera based on transformer. In Proceedings of the 2022 5th International Conference on Image and Graphics Processing, pages 72-78, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[53] Gina G Turrigiano and Sacha B Nelson. Homeostatic plasticity in the developing nervous system. Nature reviews neuroscience, 5(2):97-107, 2004. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 241 + ], + "type": "text", + "content": "[55] Bingsen Wang, Jian Cao, Jue Chen, Shuo Feng, and Yuan Wang. A new ann-snn conversion method with high accuracy, low latency and good robustness. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 3067-3075, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 285 + ], + "type": "text", + "content": "[56] Yuchen Wang, Kexin Shi, Chengzhuo Lu, Yuguo Liu, Malu Zhang, and Hong Qu. Spatial-Temporal Self-Attention for Asynchronous Spiking Neural Networks. In Proceedings of IJCAI, pages 3085–3093, Macau, SAR China, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[57] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[58] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 431 + ], + "type": "text", + "content": "[59] Yujie Wu, Rong Zhao, Jun Zhu, Feng Chen, Mingkun Xu, Guoqi Li, Sen Song, Lei Deng, Guanrui Wang, Hao Zheng, et al. Brain-inspired global-local learning incorporated with neuromorphic computing. Nature Communications, 13(1): 65, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[60] Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K. Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neural networks, 2023. arXiv:2304.09500 [cs]. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 511 + ], + "type": "text", + "content": "[61] Qi Xu, Jie Deng, Jiangrong Shen, Biwu Chen, Huajin Tang, and Gang Pan. Hybrid Spiking Vision Transformer for Event-Based Object Detection, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "type": "text", + "content": "[62] Qi Xu, Xuanye Fang, Yaxin Li, Jiangrong Shen, De Ma, Yi Xu, and Gang Pan. Rsnn: Recurrent spiking neural networks for dynamic spatial-temporal information processing. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10602-10610, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "text", + "content": "[63] Kashu Yamazaki, Viet-Khoa Vo-Ho, Darshan Bulsara, and Ngan Le. Spiking neural networks and their applications: A review. *Brain Sciences*, 12(7):863, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[64] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[65] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-Wise Attention Spiking Neural Networks for Event Streams Classification. In Proceedings of ICCV, pages 10221-10230, 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 587 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[66] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. arXiv preprint arXiv:2209.13929, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[67] Man Yao, Jiakui Hu, Guangshe Zhao, Yaoyuan Wang, Ziyang Zhang, Bo Xu, and Guoqi Li. Inherent Redundancy in Spiking Neural Networks. In Proceedings of the ICCV, pages 16924-16934, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[68] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention Spiking Neural Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9393-9410, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[69] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[70] Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160-32171, 2022. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[71] Chengting Yu, Zheming Gu, Da Li, Gaoang Wang, Aili Wang, and Erping Li. STSC-SNN: Spatio-Temporal Synaptic Connection with temporal convolution and attention for spiking neural networks. Frontiers in Neuroscience, 16, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "text", + "content": "[72] Malu Zhang, Jibin Wu, Yansong Chua, Xiaoling Luo, Zihan Pan, Dan Liu, and Haizhou Li. Mpd-al: an efficient membrane potential driven aggregate-label learning algorithm for spiking neurons. In Proceedings of the AAAI conference on artificial intelligence, pages 1327-1334, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[73] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[74] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "text", + "content": "[75] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "text", + "content": "[76] Rui-Jie Zhu, Malu Zhang, Qihang Zhao, Haoyu Deng, Yule Duan, and Liang-Jian Deng. Tcja-snn: Temporal-channel joint attention for spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13969" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_content_list.json b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbe931fa68ce0b5b3c13a52b22d581cd65416705 --- /dev/null +++ b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_content_list.json @@ -0,0 +1,1628 @@ +[ + { + "type": "text", + "text": "STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds", + "text_level": 1, + "bbox": [ + 120, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zikuan Li*, Honghua Chen*, Yuecheng Wang, Sibo Wu, Mingqiang Wei, Jun Wang†", + "bbox": [ + 148, + 202, + 818, + 222 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98f5a62a096ae4ee4d06bc8fdaa57115cf4a9d159e40a69040f6f5331176347b.jpg", + "image_caption": [ + "Figure 1. The characteristics of our method and visual comparison of edge extraction results in thin-walled structures. (a) defines the concept of a thin-walled edge. (b) highlights that the primary challenge in thin-walled edge extraction is the sensitivity to local neighborhood selection. (c) illustrates how a larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring boundary information. (d) shows that a smaller neighborhood may suffer from noise and insufficient sampling on side faces, lacking the contextual information needed for accurate edge point recognition. (e) depicts the spherical projection of the local neighborhood of four points, where points of the same color lie on the same underlying surface. This illustrates our key observation that points co-planar with the neighborhood center tend to align along a great circle arc after spherical projection. (f) provides a qualitative comparison of our STAR-Edge with RFEPS [29], EC-Net [32], and MFLE [4]." + ], + "image_footnote": [], + "bbox": [ + 76, + 250, + 893, + 440 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 566, + 313, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extracting geometric edges from unstructured point clouds remains a significant challenge, particularly in thin-walled structures that are commonly found in everyday objects. Traditional geometric methods and recent learning-based approaches frequently struggle with these structures, as both rely heavily on sufficient contextual information from local point neighborhoods. However, 3D measurement data of thin-walled structures often lack the accurate, dense, and regular neighborhood sampling required for reliable edge extraction, resulting in degraded performance.", + "bbox": [ + 73, + 599, + 468, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce STAR-Edge, a novel approach designed for detecting and refining edge points in thin-walled structures. Our method leverages a unique representation—the local spherical curve—to create structure-aware neighborhoods that emphasize co-planar points while reducing interference from close-by, non-co-planar surfaces. This representation is transformed into a rotation-invariant descriptor, which, combined with a lightweight", + "bbox": [ + 75, + 752, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multi-layer perceptron, enables robust edge point classification even in the presence of noise and sparse or irregular sampling. Besides, we also use the local spherical curve representation to estimate more precise normals and introduce an optimization function to project initially identified edge points exactly on the true edges. Experiments conducted on the ABC dataset and thin-walled structure-specific datasets demonstrate that STAR-Edge outperforms existing edge detection methods, showcasing better robustness under various challenging conditions. The source code is available at https://github.com/Miraclelezk/STAR-Edge.", + "bbox": [ + 496, + 568, + 893, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 782, + 632, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thin-walled structures are essential in our daily lives, seen in items like mechanical parts, furniture, and various manufactured objects. These structures often contain thin-walled edges, defined as boundary lines where surfaces intersect at their side-end faces, characterized by a minimal thickness relative to other dimensions (see Fig. 1 (a)). As geometric", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author", + "bbox": [ + 99, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27254", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "edges provide an abstraction of complex 3D shapes, extracting edges from 3D measurements (e.g. unstructured point clouds) in these structures is crucial for downstream tasks.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extracting geometric edges from point clouds has been a long-standing problem. Traditional techniques generally rely on local geometric properties, such as eigenvalues of the covariance matrix [2, 28], normals [27], curvature [12], or elaborately-designed descriptors [4]. While effective to some extent, these methods depend on first- or second-order derivatives estimated from local neighborhoods, making them particularly sensitive to point disturbances and sampling sparsity or irregularities, especially on thick side-end faces. More recently, some deep neural networks [1, 8, 13, 13, 18, 19, 32] have been proposed to enhance edge accuracy by leveraging deep features [26], neural edge representations [35], or novel point representations (e.g., Scale Space Matrix [18]). However, these methods still struggle with thin-walled structures (as seen in Fig. 1 (f)), as they rely on extracting useful information from valid local neighborhoods.", + "bbox": [ + 75, + 138, + 468, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To be more specific, the local neighborhood of points around thin-walled edges is often unreliable. A larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring the boundary information (see Fig. 1 (c)). Conversely, a smaller neighborhood may suffer from noise and insufficient sampling on side-end faces, lacking the contextual information needed for accurate edge point recognition by those above networks (see Fig. 1 (d)). Although multi-scale schemes [26] can partially mitigate this issue, we argue that this approach is sub-optimal, as it requires hard-coding several fixed neighborhood scales.", + "bbox": [ + 75, + 397, + 468, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ideally, each point's local representation should be structure-aware. This means it should reliably emphasize points on the same underlying surface (e.g., the same plane) while minimizing interference from points on different structures (e.g., another plane). Interestingly, we observe that points co-planar with the neighborhood center tend to align along a great circle arc (dashed circle on the blue plane in Fig. 1 (e)) after spherical projection, effectively reflecting the underlying surface of the thin-walled structure. Based on this observation, naturally, we can construct a local spherical curve to incorporate as many points on the great circle arc as possible, thereby enhancing points on the same underlying surface while reducing the influence of points from other surfaces.", + "bbox": [ + 75, + 580, + 467, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present STAR-Edge, a novel approach for detecting and refining edge points in thin-walled structures. Our method begins with a new representation, the local spherical curve, to explicitly build structure-aware local neighborhoods. This design enhances information aggregation from nearby co-planar points, even in noisy, sparse, or irregularly point distribution settings. Next, we construct", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a rotation-invariant descriptor based on the local spherical curve and integrate it with a multi-layer perceptron (MLP) to accurately classify edge and non-edge points. Beyond edge classification, we also use the local spherical curve representation to estimate more precise point normals and introduce an optimization function to project points closer to the true edge. To validate our approach, we conduct a series of experiments on the well-known ABC dataset [17] and specialized datasets for thin-walled structures [6]. Extensive experimental results demonstrate that our approach outperforms existing edge detection methods. Our contributions are summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce STAR-Edge, a new method for detecting and refining edge points in thin-walled structures.", + "- We develop a new representation, the local spherical curve, to construct structure-aware local neighborhoods, enhancing information aggregation from nearby co-planar points, even under challenging conditions like noise, sparsity, or irregular point distribution.", + "- Using the local spherical curve representation, we construct a rotation-invariant local descriptor to achieve more accurate edge point classification. Additionally, this representation allows us to accurately calculate point cloud normals, enabling further refinement of edge points.", + "- We perform extensive experiments on specialized datasets for thin-walled structures and the ABC dataset, showing that our method outperforms existing edge extraction techniques." + ], + "bbox": [ + 500, + 273, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 532, + 648, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Traditional method for edge extraction", + "text_level": 1, + "bbox": [ + 500, + 558, + 831, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early edge extraction techniques primarily targeted mesh models [15, 25], with subsequent advancements focusing on direct edge feature extraction from point clouds. Oztireli et al. [23] apply robust kernel regression combined with Moving Least Squares (MLS) to preserve fine details and sharp features in point-based surface representations. Daniels et al. [9] further advance this by introducing a robust method that identifies sharp features in point clouds by generating smooth edge-aligned curves through multi-step refinement and surface fitting. They both rely on iterative refinement processes to progressively extract and enhance sharp edge features in point cloud data.", + "bbox": [ + 496, + 582, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mérigot et al. [20] propose a robust method using Voronoi-based covariance matrices to robustly extract curvature, sharp features, and normals from point clouds. Bazazian et al. [2] present a fast method for sharp edge extraction from unorganized point clouds using eigenvalues of covariance matrices from k-nearest neighbors. Building on these approaches, other researchers have explored various techniques for identifying and preserving critical features in point clouds. For instance, Huang et al. [14] present an", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27255", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edge-aware resampling method that computes reliable normals away from edges while effectively preserving sharp features. Ni et al. [22] propose AGPN, a method for detecting edges and tracing feature lines in 3D point clouds using neighborhood geometry analysis with RANSAC and hybrid region growing. Mineo et al. [21] present a boundary point detection algorithm with fast Fourier transform (FFT)-based filtering to improve edge reconstruction and reduce noise in tessellated surfaces from point clouds. Chen et al. [4] propose a two-phase algorithm for extracting linetype features from unstructured point clouds, using a statistical metric to detect feature points and an anisotropic contracting scheme to reconstruct feature lines against noise.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, most traditional point cloud edge extraction methods rely heavily on neighborhood reliability, which includes dense sampling and minimal noise. However, in thin-walled structures, neighborhood information is often ambiguous, making conventional methods generally unsuitable for such structures.", + "bbox": [ + 75, + 287, + 472, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Learning-based methods for edge extraction", + "text_level": 1, + "bbox": [ + 76, + 392, + 452, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the emergence of learning techniques, many data-driven approaches have been proposed to improve the robustness of edge extraction and the preservation of sharp edges [5, 34]. EC-Net [32] is introduced as an edge-aware point set consolidation network designed to handle sparse, irregular, and noisy point clouds by both minimizing distances to 3D meshes and edges. PIE-NET [26] uses PointNet++ [24] as the backbone for classification and regression to extract edges from 3D point clouds as parametric curves, and has been extensively validated on synthetic data. CPE-Net [7] is designed in a multitask learning paradigm and uses a learnable weighted least squares fitting approach to accurately extract countersink edges under noisy conditions. Generally, all of these methods initially classify edge points using deep learning techniques, followed by edge generation through upsampling, clustering, or fitting smooth curves to capture edge structures in point clouds.", + "bbox": [ + 75, + 416, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are also some methods that focus on learning edge features directly from point clouds. Bode et al. [3] proposed Bounded, a neural network that uses local neighborhood statistics to classify non-edge, sharp-edge, and boundary points in 3D point clouds. Yao et al. [30] propose HGNet, a point cloud network that aggregates hierarchical features from points to super-surfaces. Zhao et al. [33] propose a deep learning approach to detect and consolidate sharp feature edges in 3D point clouds using a multi-task network that predicts displacement vectors for edge refinement. APEE-Net [6] is a dual-task neural network designed for extracting precise edges on aircraft panels by identifying nearby points and also using displacement vectors for repositioning. In addition, NerVE [35] defines a neural volumetric edge representation to extract parametric edge", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "curves through graph search, leveraging neural fields rather than traditional point cloud processing. Ye et al. [31] propose a Neural Edge Field method for reconstructing 3D feature curves from multi-view images, optimizing a neural implicit field with a rendering loss to extract 3D edges without 3D supervision. Although these learning-based methods achieve better performance than previous geometric-based methods, they still rely heavily on sufficient contextual information from local point neighborhoods, making them challenging to apply to thin-walled structures.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 255, + 591, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed method extracts edges from thin-walled structures, as illustrated in Fig. 2, through three main steps: local spherical curve construction, edge point identification, and edge optimization. First, for each point, we project its sparse neighborhood onto a unit sphere, constructing a closed spherical curve to represent the local neighborhood (Sec 3.1). Next, we use spherical harmonics (SH) to encode this spherical curve as a rotation-invariant local descriptor, which is then passed through an MLP layer to classify the target point as an edge or non-edge point (Sec 3.2). Finally, we leverage this descriptor again to estimate the normal vector, allowing us to formulate a projection function that refines the edge point positions (Sec 3.3).", + "bbox": [ + 496, + 280, + 893, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Local Spherical Curve Construction", + "text_level": 1, + "bbox": [ + 498, + 486, + 815, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key observation is that neighboring points co-planar with a target point align along great circle arcs after spherical projection, allowing the overall shape to be approximated as a closed, continuous curve. Thus, given a point cloud $P = \\{p_i \\in \\mathbb{R}^3 | i = 1,2,\\dots,N\\}$ , we first perform spherical normalization on each point $p_i$ 's neighborhood point to obtain the local spherical point set $Q_i$ , and then fit a closed, continuous curve.", + "bbox": [ + 496, + 510, + 893, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, by taking $p_i$ as the center of the sphere, we normalize the radius of its neighboring points to 1, projecting them onto the unit sphere. The normalization process is defined as follows:", + "bbox": [ + 496, + 630, + 893, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = \\left\\{q _ {j} = \\frac {p _ {j} - p _ {i}}{\\| p _ {j} - p _ {i} \\|} \\right| p _ {j} \\in K N N (p _ {i}) \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 700, + 890, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q_{j}$ represents the local spherical point and $p_j$ is the KNN neighboring points of $p_i$ . Next, we apply Principal Component Analysis (PCA) to $Q_{i}$ , projecting it onto a local plane. We then compute the convex hull of the projected points and select the vertices of this convex hull as the key point set $K_{i}$ for curve fitting:", + "bbox": [ + 496, + 739, + 893, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK _ {i} = \\operatorname {C o n v e x H u l l} \\left(\\operatorname {P C A} \\left(Q _ {i}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 842, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we perform spline fitting on $K_{i}$ to obtain a spherical curve, effectively representing the local neighborhood.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27256", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d4fe5876330670f17164a4752f80ed0e2df45c0213380d259f5d17fca3afc0d6.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. STAR-edge comprises three main steps: (a) constructing a local spherical curve; (b) identifying edge points by integrating rotation-invariant local descriptors with a light-weight MLP layer; and (c) optimizing edge points by adjusting their positions based on the estimated normal vector from the local spherical curve." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 848, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b18ae857a294db13b949e76cd23e925b88b689021907ffd43ebc3287400dd9f.jpg", + "image_caption": [ + "(a) Thin-walled structure", + "(thickness: 3, resolution: 1)" + ], + "image_footnote": [], + "bbox": [ + 124, + 398, + 274, + 578 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5dbd7bfb943a5a34c01123d6b01430535559fa21babdd5938ef7dca5a13a10d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 414, + 390, + 496 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b31f883d628d5a3cbae2b5c30eeb69ece875b9d76e0d67dbea4f4774311e894.jpg", + "image_caption": [ + "(b) Local Spherical Points (KNN)" + ], + "image_footnote": [], + "bbox": [ + 393, + 412, + 501, + 496 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e32ae342cdbf243a442db177d3802b49e0d60aebc431cc43d12a1b4b49f61147.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 412, + 616, + 496 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2afbb0bbfdf7671a64e44a8fb9e371c88b662a37e4b35bd1ae1343d6006b21e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 411, + 728, + 496 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/15bba3a42695e6de10214e6cd1afd8a49eff7227796e0ecdfe066e25bf3239a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 411, + 839, + 496 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e102264fdf6228d226c129cd0b406fa9b9a7d4efaa82a3a2da8a342b4d23cfe.jpg", + "image_caption": [ + "(c)Key Points and Local Spherical Curve", + "Figure 3. Neighborhood distribution of five surface points on the unit sphere. Local spherical points are shown in grey, sampled key points on the fitted spherical Curve in blue, and local spherical curves in red. We observe that the neighborhoods of edge points resemble two semicircles (e.g. ④), while those of non-edge points maintain a complete great circle (e.g. ①)." + ], + "image_footnote": [], + "bbox": [ + 279, + 515, + 390, + 598 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3425c27ebc8e7c8f5ece34a3b130b87cba8d131cc0b9e2fde77ea3cfb6d4e84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 515, + 501, + 598 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a652af4d6bf527d8f3a7d763f9fcc800da8b3124aab871c67296b5eb24518d8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 515, + 616, + 598 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea8484b7b5b54f3e232113a9bc2ac0a54452fe3e55fc313126c2ff9472d3980d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 515, + 728, + 598 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0713654c7b5dc4e68444c7b6e270359d0002009f581d181af16577be97a5db56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 515, + 839, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 shows the distribution of local spherical points (grey), key points (blue), and the local spherical curve (red) for five target points across a thin-walled structure. Points $①$ and $②$ , both located on the top surface, their neighborhoods are close to a complete great circle despite interference from points on other surfaces. Point $⑤$ , positioned on a side-end face with sparse sampling, is influenced by both upper and lower surfaces of the thin-walled structure; however, its local spherical curve still largely retains a great circle form. In contrast, points $③$ and $④$ are near a sharp edge, with $④$ closer to the edge than $③$ . The local spherical curves of $③$ and $④$ are roughly formed by two semicircles, presenting a different pattern compared to points $①$ and $②$ . These results demonstrate that the proposed local spherical", + "bbox": [ + 75, + 685, + 472, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "curve can robustly and sensitively capture edge features.", + "bbox": [ + 500, + 685, + 870, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Edge Point Identification", + "text_level": 1, + "bbox": [ + 500, + 710, + 728, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the spherical curve for each point, we use SH, an effective time-frequency domain analysis technique on a sphere, to construct rotation-invariant descriptors. This SH-encoded curve serves as a rotation-invariant local descriptor, which is then fed into an MLP to predict whether each point is an edge point.", + "bbox": [ + 496, + 733, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In detail, we discretize the local spherical curve into a set of scattered points. Each points $(x,y,z)$ is mapped to spherical coordinates $f(\\theta ,\\phi)$ on the unit sphere, where the mapping is defined as $(\\theta ,\\phi) = (\\arccos z,\\arctan y / x)$ . Here, $\\theta$ and $\\phi$ represent the zenith and azimuth angles, respec", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27257", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tively. Let $f_{1}, f_{2}, \\ldots, f_{n}$ be a series of mapping on spherical space. We employ kernel density estimation [10] to construct a smooth, continuous function based on these sample points:", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (\\theta , \\phi) = \\frac {1}{n h} \\sum_ {i = 1} ^ {n} k \\left(\\frac {\\delta_ {i} (\\theta , \\phi)}{h}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 151, + 468, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ is a kernel function, defined as a normal probability density function:", + "bbox": [ + 76, + 194, + 468, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nk (x) = 1 / \\sqrt {2 \\pi} \\exp (- 0. 5 x ^ {2}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 233, + 468, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The term $\\delta_{i}(\\theta, \\phi)$ denotes the geodesic distance between $f_{i}(\\theta_{i}, \\phi_{i})$ and $(\\theta, \\phi)$ . The parameter $h$ , known as the window width or smoothing parameter, modulates the influence of mapping points $f_{i}$ to the density estimation at $(\\theta, \\phi)$ . It is defined as $h = \\pi / B$ , where $B$ represents the spherical function bandwidth.", + "bbox": [ + 76, + 261, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compute the SH coefficients for the band-limited function, we apply the discrete SH transform. Following the Shannon-Nyquist sampling theorem [11], we sample the model equiangularly at twice the bandwidth $B$ in both latitude and longitude, resulting in $2B \\times 2B$ sampling points $(\\theta_{j}, \\phi_{k}) \\in (0, \\pi) \\times [0, 2\\pi)$ , with", + "bbox": [ + 76, + 352, + 468, + 443 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\theta_ {j} = \\pi (2 j + 1) / 4 B & , j = 0, 1, \\dots , 2 B - 1 \\\\ \\phi_ {k} = \\pi k / B & , k = 0, 1, \\dots , 2 B - 1 \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 453, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SH coefficients of the function $f(\\theta, \\phi)$ are obtained through the discrete SH transform:", + "bbox": [ + 76, + 502, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {l} ^ {m} = \\frac {\\sqrt {2 \\pi}}{2 B} \\sum_ {j = 0} ^ {2 B - 1} \\sum_ {k = 0} ^ {2 B - 1} w _ {j} f \\left(\\theta_ {j}, \\phi_ {k}\\right) e ^ {- i m \\phi_ {k}} P _ {l} ^ {m} (\\cos \\theta_ {j}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 542, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for degree $0 < l < B$ and order $|m| \\leq l$ . Here $P_{l}^{m}$ denotes the associated Legendre functions defined on $[-1, 1]$ . The weights $w_{j}$ are computed using the method outlined in [11].", + "bbox": [ + 76, + 599, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SH bases are indexed by two integers, $l$ and $m$ , where $-l \\leq m \\leq l$ , resulting in $2l + 1$ coefficients for each degree $l$ . The parameter $l$ determines the frequency of the basis [16]. Consequently, the $2l + 1$ coefficients corresponding to the same degree $l$ are combined into a vector:", + "bbox": [ + 76, + 645, + 468, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {l} = \\left\\{\\hat {f} _ {l} ^ {m}: m = - l, \\dots , 0, \\dots , l \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 729, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the $\\ell_2$ norm $\\| \\beta_l\\|$ is taken as the energy of the function $f$ in degree $l$ . At last, the rotation-invariant local descriptor $D$ of function $f(\\theta ,\\phi)$ with all $B$ degrees is defined as", + "bbox": [ + 76, + 757, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD (f, B) = \\{\\| \\beta_ {l} \\|: l = 0, \\dots , B - 1 \\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 813, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After computing this descriptor, we calculate the average descriptor for ground-truth edge points and measure the feature difference between each point and this average descriptor. By applying a threshold to this difference, we can", + "bbox": [ + 76, + 840, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1ab2b6b4a35c32bf4bc3b1c4ea69961798032ad386da1d1a5fe5b3c2cf02bfa.jpg", + "image_caption": [ + "Figure 4. Normal estimation results for points $③, ④,$ and $⑤$ in Fig. 3 (a). In each sub-figure, the red curve is the great circle arc extracted by RANSAC, the blue curve is the non-great circle segment, and the yellow arrow is the estimated normal vector." + ], + "image_footnote": [], + "bbox": [ + 522, + 95, + 651, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf34b428f812ccc2c919596b917bc0dc84b30ac69aecc9bb09e8c9309676927d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 95, + 756, + 178 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9d3fd4c673dbf0502377d052aa421fc03b96ff9bce4c7eb594b5f2bd8f4cbb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 95, + 867, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "determine whether a point is classified as an edge point. However, in our initial testing, we found that the threshold-based method struggles to effectively handle different edge styles. To address this, we use an MLP to adaptively learn from the rotation-invariant local descriptor, enabling accurate edge point identification. Detailed network structure is provided in the supplementary material.", + "bbox": [ + 498, + 281, + 890, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Edge Point Optimization", + "text_level": 1, + "bbox": [ + 500, + 398, + 725, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that the initially identified edge points are roughly scattered around the true edge. While these points are linearly distributed overall, they form a band-like structure cross two intersecting surfaces when viewed locally. To improve accuracy, we propose constraining edge points to surface intersection regions. This approach is motivated by the fact that an edge point at the intersection of two surfaces has a zero projection distance to both surfaces.", + "bbox": [ + 498, + 421, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $p_i$ represent an initially-identified edge point. Inspired by [29] and [6], we introduce the following optimization function:", + "bbox": [ + 498, + 542, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {z _ {i}} \\sum_ {p _ {j} \\in K N N (p _ {i})} \\left(\\left(z _ {i} - p _ {j}\\right) \\cdot n _ {j}\\right) ^ {2} + \\mu \\| z _ {i} - p _ {i} \\| ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 597, + 890, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_j$ is the normal vector of $p_j$ , and $\\mu$ is a balance parameter. The first term encourages $z_i$ to stay close to the tangent planes of neighboring points, while the second term penalizes the deviation of $z_i$ from its initial position $p_i$ . In all our experiments, $\\mu$ is set to a default value of 0.1.", + "bbox": [ + 498, + 643, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal estimation. Note that the edge point optimization method described above relies on accurate normal vectors; however, conventional normal estimation methods is unreliable in noisy or sparse neighborhood conditions. To resolve it, we propose a normal estimation algorithm based on the proposed local spherical curves, which can enhance robustness due to its underlying structure awareness.", + "bbox": [ + 498, + 719, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, we observe that points co-planar with $p_i$ (shown on the red curve) tend to align on the same plane as the great circle arc after projection onto a sphere. In contrast, other points (shown on the blue curve) lie on a different plane. Therefore, we use this great circle arc", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27258", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(red) on the local spherical curve to approximate the local plane and derive the normal vector. Specifically, we employ random sampling consensus (RANSAC) to identify the great circle arc on the local spherical curve to determine the normal plane, and use the rest (blue) to determine the normal orientation. Fig. 4 shows the normal estimation of the neighborhood $③$ , $④$ , and $⑤$ in Fig. 3 (a).", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Discussion", + "text_level": 1, + "bbox": [ + 76, + 209, + 339, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the effectiveness of our method by comparing it to several learning-based approaches [26, 32, 35] and two traditional methods [4, 29]. We then assess its robustness under varying noise intensities and point cloud resolutions and conduct an ablation study to examine the impact of specific design choices.", + "bbox": [ + 75, + 234, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 334, + 253, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Given the characteristics of thin-walled structures (small dimensions in the thickness direction), we selected a dataset [6] specifically focused on thin-walled structures for our experiments. This dataset includes 42 distinct 3D models with curved edges of varying shapes, sizes, and curvatures, and thicknesses. For a more detailed description of this dataset, please refer to the supplemental file. Note that a subset of 32 shapes is used for training, while 10 shapes are reserved for testing. For each shape, we sample point clouds at resolutions of 0.3, 0.5, and 0.8. Additionally, three levels of random Gaussian noise are added to the point clouds at a resolution of 0.5. Each sampled point cloud contains approximately 20,000 to 15 million points.", + "bbox": [ + 75, + 357, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also perform experiments on the well-known ABC dataset [17], which contains over one million CAD models with edge annotations. To evaluate the generalizability of our method, we select sharp-edged CAD models from the ABC dataset, following the setup in [35]. This subset comprises 2,364 models, which are randomly divided into a training set (80%) and a test set (20%). For each shape, 50,000 points are sampled.", + "bbox": [ + 75, + 551, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. All experiments are conducted on an Intel i7-11700 CPU and an NVIDIA RTX 4090 GPU. For constructing the local rotation-invariant descriptor, we use a KNN query with $k = 26$ and a spherical harmonics transform bandwidth of 10. The proposed approach is implemented in PyTorch and trained over 40 epochs using the Adam optimizer, with a batch size of 1024 and a learning rate of 0.02.", + "bbox": [ + 75, + 674, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We quantitatively evaluate all methods across two aspects. First, point classification performance is assessed using Recall, Precision, F1-score, and Accuracy. Second, to evaluate the quality of the final edge point extraction, we introduce a new metric: Edge Chamfer Distance (ECD). This metric measures the similarity between the optimized edges and the ground-truth edges. The introduction of ECD", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is necessary because the ground-truth sampled point cloud consists of dense points, while the optimized edges may vary in sparsity depending on the chosen resolution. Standard Chamfer Distance, without adjustment, could disproportionately penalize sparse edges, leading to inaccurate similarity assessments. ECD addresses this limitation by providing a more reliable comparison for edge extraction quality. Let $X$ represents the point cloud of the optimized edges, and $Y$ denote the ground-truth edges. The ECD is defined as follows:", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E C D} = \\frac {1}{| X |} \\sum_ {p \\in X} \\min _ {q \\in Y} \\| p - q \\| _ {2} ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 250, + 890, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that for each shape, we label sampled points within a specified distance threshold from the nearest edge as ground-truth edge points. We set empirically this threshold to 0.3 for the dataset in [6] and 0.025 for the ABC dataset.", + "bbox": [ + 496, + 294, + 890, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on Thin-walled Dataset", + "text_level": 1, + "bbox": [ + 498, + 364, + 803, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our STAR-Edge method with state-of-the-art approaches, including EC-Net [32], PIE-Net [26], NerVE [35], MFLE [4], and RFEPS [29], on the thin-walled dataset. All methods use point cloud data with a resolution of 0.5 as input. To accommodate network size limitations, PIE-Net is provided with 8,096 uniformly sampled points obtained through farthest point sampling, while NerVE converts the point cloud into a voxel grid with a resolution of $64^3$ . All networks are retrained on the same dataset for a fair comparison.", + "bbox": [ + 496, + 387, + 890, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 provides visual comparisons of all methods, while quantitative results are presented in Tab. 1. Both results demonstrate that our approach outperforms others on thin-walled structures, achieving better accuracy and a more uniform edge point distribution. Specifically, EC-Net, which up-samples point clouds in edge regions, tends to generate redundant points around true edges, leading to higher Edge Chamfer Distance (ECD). PIE-Net and NerVE perform less effectively on this dataset; PIE-Net requires dense sampling to ensure its effectiveness in edge extraction, while NerVE's voxel size exceeds the thin-wall thickness, making it struggle to perceive edges. A potential improvement for these methods is to segment small patches for re-processing, as discussed in the supplemental file. The geometry-based methods, MFLE and RFEPS, also show limitations: MFLE requires frequent parameter adjustments, while RFEPS produces noisier edge results. Contrast, our method leverages the local spherical curve to achieve more precise edge points and displacement vectors, resulting in superior performance for edge detection in thin-walled structures.", + "bbox": [ + 496, + 537, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Running time. The running times of various methods on point clouds containing approximately 320,000 points are as follows: our method required 214.0 seconds, MFLE completed in 1.14 seconds, RFEPS took 19 minutes, ECNet", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27259", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4d4107d7f1bbbccaa5003098747b9841dbf39f797048b4826f6477ea575cbc2f.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of edge point extraction on the thin-walled structure dataset. -' denotes a failed calculation." + ], + "table_footnote": [], + "table_body": "
EC-Net [32]PIE-Net [26]MFLE [4]RFEPS [29]STAR-Edge (Ours)
ECD↓0.293652.12040.48391.43920.0587
Recall↑-0.05030.97550.99970.9952
Precision↑-0.02950.30280.15120.5418
F1↑-0.03650.46210.26250.7013
Accuracy↑-0.93490.94700.95440.9931
", + "bbox": [ + 174, + 113, + 792, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beb9c451fa1c1c6bda91c1370ab74f704a45c162fbc8b4c6a2ab34917050dbd7.jpg", + "image_caption": [ + "Figure 5. Visual comparison of state-of-the-art methods on the thin-walled structure dataset. In these close-up views, our method demonstrates superior accuracy in edge point extraction. Edge points are in red color." + ], + "image_footnote": [], + "bbox": [ + 122, + 215, + 848, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/479f791b9d8bca0a43db8d5dcd4b42d1c367a367b0a74d4c52a3a6ac9ec24b09.jpg", + "image_caption": [ + "Figure 6. Visual comparison under different noise levels." + ], + "image_footnote": [], + "bbox": [ + 101, + 645, + 447, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "finished in 83 seconds, and PIE-Net was the fastest, con", + "bbox": [ + 76, + 885, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db7ec828cd9bf8b8c838aef7c19a759d2f8dd3815f0305bb8fdb98b952d67deb.jpg", + "image_caption": [ + "Figure 7. Visual comparison under different resolutions." + ], + "image_footnote": [], + "bbox": [ + 522, + 643, + 870, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "suming less than 1.0 second. We attribute PIE-Net's exceptional speed to its down-sampling process. More detailed statistics can be found in the supplementary materials.", + "bbox": [ + 500, + 768, + 890, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Robustness Study", + "text_level": 1, + "bbox": [ + 500, + 830, + 671, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted robustness testing for the proposed STAR-Edge by introducing varying levels of noise and resolution. In our experiments, the thin-walled shapes have a thickness", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27260", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ccb8774b79cad6208a13b0f8a4be1f8497e84ac848d1dbae506af0af25fcea6.jpg", + "table_caption": [ + "Table 2. Quantitative classification results under different noise levels and resolutions." + ], + "table_footnote": [], + "table_body": "
NoiseResolution
0.0010.0050.030.30.50.8
ECD↓0.0590.0590.1990.0480.0590.261
Recall↑0.9950.9950.9940.9600.9950.988
Precision↑0.5420.5410.5230.7840.5420.582
F1↑0.7010.7010.6850.8630.7010.731
Accuracy↑0.9930.9930.9930.9980.9930.993
", + "bbox": [ + 93, + 128, + 450, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70f43082641d3b3862592c144c6c3551e929bac08fc1d4c9a6584bc5b083b52d.jpg", + "table_caption": [ + "Table 3. Quantitative ECD comparisons under different noise and resolution. $\\cdot$ indicates that the calculation failed." + ], + "table_footnote": [], + "table_body": "
MethodsNoiseResolution
0.0010.0050.030.30.50.8
EC-Net0.2950.2930.294-0.2940.518
PIE-Net50.28552.33249.15650.79152.12052.222
RFEPS1.4401.4361.4420.3911.43910.341
MFLE0.4840.4530.4090.4060.4840.742
Ours0.0590.0590.1990.0480.0590.260
", + "bbox": [ + 99, + 282, + 444, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of 2-3 (which is very small relative to other shape dimension). At a resolution of 0.3, approximately 9 points can be sampled on the side-end surface, whereas at a resolution of 0.8, it becomes challenging to obtain even 4 uniformly distributed points. Testing at a resolution of 0.8 poses a significant robustness challenge, as the limited number of sampling points reduces the reliability of projections onto spherical curves.", + "bbox": [ + 75, + 411, + 467, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 2 and Tab. 3 present the effects of noise intensity and resolution on the predictions of our STAR-Edge method. We observe that, even under high noise intensity or significantly reduced resolution, our method consistently produces reasonable edge point predictions with low Edge Chamfer Distance (ECD). As shown in Fig. 6, when the noise level reaches 0.003, the detection performance remains largely unaffected. As shown in Fig. 7, demonstrates that even in an extreme case with a thickness of 3 and a resolution of 0.8, the upper and lower edges can still be clearly distinguished. Additionally, in Tab. 3, we compare the performance of STAR-Edge with state-of-the-art methods across inputs with varying noise levels and resolutions. Our method achieves the best numerical performance.", + "bbox": [ + 75, + 532, + 467, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 755, + 228, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 reports the performance of our method under varying SH bandwidths, which determine the precision and detail of the local feature representation. The results indicate that performance improves with increasing bandwidth; however, higher bandwidth settings also result in increased computational cost. Notably, the performance at a bandwidth of 10 is comparable to that at 15, making a bandwidth of 10 the optimal choice for balancing accuracy and efficiency.", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/33f6cfd02e60df2fce9695c393f0f0fbec49aab80fa352009e8d4e3eaba6d919.jpg", + "table_caption": [ + "Table 4. Ablation study about SH bandwidths." + ], + "table_footnote": [], + "table_body": "
BandwidthRecall↑Precision↑F1↑Accuracy↑ECD↓
50.9880.5230.6850.9910.122
100.9950.5420.7010.9930.059
150.9950.5830.7850.9980.048
", + "bbox": [ + 521, + 114, + 867, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6682244787573d87be17401da0b15e2d6d1fd8c4dde4fd33ef25d3cc1312eda.jpg", + "image_caption": [ + "Figure 8. Visual comparison of edge extraction results on 3D shapes from the ABC dataset. Our method achieves more precise edge extraction, particularly in the sharp corner regions." + ], + "image_footnote": [], + "bbox": [ + 514, + 189, + 875, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Evaluation on ABC dataset", + "text_level": 1, + "bbox": [ + 500, + 493, + 745, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our method on the ABC dataset following the experimental setup in [35]. Quantitative results are presented in Tab. 5, which demonstrate that STAR-Edge consistently outperforms all baseline methods in terms of edge extraction accuracy. Visual comparisons in Fig. 8 further highlight the superior performance of our method, as it achieves more precise edge extraction, particularly in the sharp corner regions.", + "bbox": [ + 496, + 518, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d999b8f91fd9dd86ff02ef1b1a04fd85989a85f3fbb2c8c194ebacef6e30fd21.jpg", + "table_caption": [ + "Table 5. Quantitative comparison on the ABC dataset." + ], + "table_footnote": [], + "table_body": "
EC-NETPIE-NETMFLERFEPSNerVEOurs
ECD↓0.00370.00900.00490.00260.00260.0024
", + "bbox": [ + 503, + 681, + 885, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 752, + 617, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented STAR-Edge, a novel point cloud edge extraction approach specifically designed for thin-walled structures. The core of our method is the introduction of a new structure-aware neighborhood representation—the local spherical curve. This representation effectively emphasizes co-planar points while minimizing interference from unrelated surfaces, enabling robust and accurate edge detection in complex thin-walled geometries.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27261", + "bbox": [ + 478, + 945, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No.92367301, No.52425506, No.92267201, No.52275493, No.T2322012).", + "bbox": [ + 76, + 114, + 468, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 172, + 174, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 2", + "[2] Dena Bazazian, Josep R Casas, and Javier Ruiz-Hidalgo. Fast and robust edge extraction in unorganized point clouds. In 2015 international conference on digital image computing: techniques and applications (DICTA), pages 1-8. IEEE, 2015. 2", + "[3] Lukas Bode, Michael Weinmann, and Reinhard Klein. Bounded: Neural boundary and edge detection in 3d point clouds via local neighborhood statistics. ISPRS Journal of Photogrammetry and Remote Sensing, 205:334-351, 2023. 3", + "[4] Honghua Chen, Yaoran Huang, Qian Xie, Yuanpeng Liu, Yuan Zhang, Mingqiang Wei, and Jun Wang. Multiscale feature line extraction from raw point clouds based on local surface variation and anisotropic contraction. IEEE Transactions on Automation Science and Engineering, 19(2):1003-1016, 2021. 1, 2, 3, 6, 7", + "[5] Honghua Chen, Zeyong Wei, Xianzhi Li, Yabin Xu, Mingqiang Wei, and Jun Wang. Repcd-net: Feature-aware recurrent point cloud denoising network. International Journal of Computer Vision, 130(3):615-629, 2022. 3", + "[6] Mengqi Chen, Laishui Zhou, Honghua Chen, and Jun Wang. Thin-walled aircraft panel edge extraction from 3d measurement surfaces via feature-aware displacement learning. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3, 5, 6", + "[7] Mengqi Chen, Laishui Zhou, Yongming Zhang, Honghua Chen, and Jun Wang. Aircraft skin countersink primitive extraction from 3-d measurement point clouds via deep clustering and fitting. IEEE Transactions on Instrumentation and Measurement, 2024. 3", + "[8] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2727-2735, 2023. 2", + "[9] Joel II Daniels, Linh K Ha, Tilo Ochotta, and Claudio T Silva. Robust smooth feature extraction from point clouds. In IEEE International Conference on Shape Modeling and Applications 2007 (SMI'07), pages 123-136. IEEE, 2007. 2", + "[10] Richard A Davis, Keh-Shin Lii, and Dimitris N Politis. Remarks on some nonparametric estimates of a density function. Selected Works of Murray Rosenblatt, pages 95-100, 2011. 5", + "[11] James R Driscoll and Dennis M Healy. Computing fourier transforms and convolutions on the 2-sphere. Advances in applied mathematics, 15(2):202-250, 1994. 5" + ], + "bbox": [ + 78, + 198, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Timo Hackel, Jan D Wegner, and Konrad Schindler. Contour detection in unstructured 3d point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1610-1618, 2016. 2", + "[13] Chems-Eddine Himeur, Thibault Lejembre, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2", + "[14] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2", + "[15] Andreas Hubeli and Markus Gross. Multiresolution feature extraction for unstructured meshes. In Proceedings Visualization, 2001. VIS'01., pages 287-294. IEEE, 2001. 2", + "[16] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Rotation invariant spherical harmonic representation of 3 d shape descriptors. In Symposium on geometry processing, pages 156-164, 2003. 5", + "[17] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6", + "[18] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 2", + "[19] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2", + "[20] Quentin Mérigot, Maks Ovsjanikov, and Leonidas J Guibas. Voronoi-based curvature and feature estimation from point clouds. IEEE Transactions on Visualization and Computer Graphics, 17(6):743-756, 2010. 2", + "[21] Carmelo Mineo, Stephen Gareth Pierce, and Rahul Summan. Novel algorithms for 3d surface point cloud boundary detection and edge reconstruction. Journal of Computational Design and Engineering, 6(1):81-91, 2019. 3", + "[22] Huan Ni, Xiangguo Lin, Xiaogang Ning, and Jixian Zhang. Edge detection and feature line tracing in 3d-point clouds by analyzing geometric properties of neighborhoods. Remote Sensing, 8(9):710, 2016. 3", + "[23] A Cengiz Öztireli, Gael Guennebaud, and Markus Gross. Feature preserving point set surfaces based on non-linear kernel regression. In Computer graphics forum, pages 493-501. Wiley Online Library, 2009. 2", + "[24] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[25] Charlie CL Wang. Incremental reconstruction of sharp edges on mesh surfaces. Computer-Aided Design, 38(6):689-702, 2006. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27262", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3, 6, 7", + "[27] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 2", + "[28] Shaobo Xia and Ruisheng Wang. A fast edge extraction method for mobile lidar point clouds. IEEE Geoscience and Remote Sensing Letters, 14(8):1288-1292, 2017. 2", + "[29] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 1, 5, 6, 7", + "[30] Ting Yao, Yehao Li, Yingwei Pan, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3", + "[31] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 3", + "[32] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 1, 2, 3, 6, 7", + "[33] Tong Zhao, Mulin Yu, Pierre Alliez, and Florent Lafarge. Sharp feature consolidation from raw 3d point clouds via displacement learning. Computer Aided Geometric Design, 103:102204, 2023. 3", + "[34] Haoran Zhou, Honghua Chen, Yidan Feng, Qiong Wang, Jing Qin, Haoran Xie, Fu Lee Wang, Mingqiang Wei, and Jun Wang. Geometry and learning co-supported normal estimation for unstructured point cloud. In Proceedings of the iee/ivf conference on computer vision and pattern recognition, pages 13238-13247, 2020. 3", + "[35] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 8" + ], + "bbox": [ + 78, + 90, + 470, + 766 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27263", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_model.json b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8508a69fcf98f68dff2a0b13941962dae2299739 --- /dev/null +++ b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.85, + 0.177 + ], + "angle": 0, + "content": "STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.203, + 0.819, + 0.223 + ], + "angle": 0, + "content": "Zikuan Li*, Honghua Chen*, Yuecheng Wang, Sibo Wu, Mingqiang Wei, Jun Wang†" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.25, + 0.895, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.444, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. The characteristics of our method and visual comparison of edge extraction results in thin-walled structures. (a) defines the concept of a thin-walled edge. (b) highlights that the primary challenge in thin-walled edge extraction is the sensitivity to local neighborhood selection. (c) illustrates how a larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring boundary information. (d) shows that a smaller neighborhood may suffer from noise and insufficient sampling on side faces, lacking the contextual information needed for accurate edge point recognition. (e) depicts the spherical projection of the local neighborhood of four points, where points of the same color lie on the same underlying surface. This illustrates our key observation that points co-planar with the neighborhood center tend to align along a great circle arc after spherical projection. (f) provides a qualitative comparison of our STAR-Edge with RFEPS [29], EC-Net [32], and MFLE [4]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.567, + 0.314, + 0.584 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.601, + 0.47, + 0.752 + ], + "angle": 0, + "content": "Extracting geometric edges from unstructured point clouds remains a significant challenge, particularly in thin-walled structures that are commonly found in everyday objects. Traditional geometric methods and recent learning-based approaches frequently struggle with these structures, as both rely heavily on sufficient contextual information from local point neighborhoods. However, 3D measurement data of thin-walled structures often lack the accurate, dense, and regular neighborhood sampling required for reliable edge extraction, resulting in degraded performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.471, + 0.875 + ], + "angle": 0, + "content": "In this work, we introduce STAR-Edge, a novel approach designed for detecting and refining edge points in thin-walled structures. Our method leverages a unique representation—the local spherical curve—to create structure-aware neighborhoods that emphasize co-planar points while reducing interference from close-by, non-co-planar surfaces. This representation is transformed into a rotation-invariant descriptor, which, combined with a lightweight" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.751 + ], + "angle": 0, + "content": "multi-layer perceptron, enables robust edge point classification even in the presence of noise and sparse or irregular sampling. Besides, we also use the local spherical curve representation to estimate more precise normals and introduce an optimization function to project initially identified edge points exactly on the true edges. Experiments conducted on the ABC dataset and thin-walled structure-specific datasets demonstrate that STAR-Edge outperforms existing edge detection methods, showcasing better robustness under various challenging conditions. The source code is available at https://github.com/Miraclelezk/STAR-Edge." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.784, + 0.633, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Thin-walled structures are essential in our daily lives, seen in items like mechanical parts, furniture, and various manufactured objects. These structures often contain thin-walled edges, defined as boundary lines where surfaces intersect at their side-end faces, characterized by a minimal thickness relative to other dimensions (see Fig. 1 (a)). As geometric" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27254" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "edges provide an abstraction of complex 3D shapes, extracting edges from 3D measurements (e.g. unstructured point clouds) in these structures is crucial for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.395 + ], + "angle": 0, + "content": "Extracting geometric edges from point clouds has been a long-standing problem. Traditional techniques generally rely on local geometric properties, such as eigenvalues of the covariance matrix [2, 28], normals [27], curvature [12], or elaborately-designed descriptors [4]. While effective to some extent, these methods depend on first- or second-order derivatives estimated from local neighborhoods, making them particularly sensitive to point disturbances and sampling sparsity or irregularities, especially on thick side-end faces. More recently, some deep neural networks [1, 8, 13, 13, 18, 19, 32] have been proposed to enhance edge accuracy by leveraging deep features [26], neural edge representations [35], or novel point representations (e.g., Scale Space Matrix [18]). However, these methods still struggle with thin-walled structures (as seen in Fig. 1 (f)), as they rely on extracting useful information from valid local neighborhoods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.579 + ], + "angle": 0, + "content": "To be more specific, the local neighborhood of points around thin-walled edges is often unreliable. A larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring the boundary information (see Fig. 1 (c)). Conversely, a smaller neighborhood may suffer from noise and insufficient sampling on side-end faces, lacking the contextual information needed for accurate edge point recognition by those above networks (see Fig. 1 (d)). Although multi-scale schemes [26] can partially mitigate this issue, we argue that this approach is sub-optimal, as it requires hard-coding several fixed neighborhood scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.793 + ], + "angle": 0, + "content": "Ideally, each point's local representation should be structure-aware. This means it should reliably emphasize points on the same underlying surface (e.g., the same plane) while minimizing interference from points on different structures (e.g., another plane). Interestingly, we observe that points co-planar with the neighborhood center tend to align along a great circle arc (dashed circle on the blue plane in Fig. 1 (e)) after spherical projection, effectively reflecting the underlying surface of the thin-walled structure. Based on this observation, naturally, we can construct a local spherical curve to incorporate as many points on the great circle arc as possible, thereby enhancing points on the same underlying surface while reducing the influence of points from other surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To this end, we present STAR-Edge, a novel approach for detecting and refining edge points in thin-walled structures. Our method begins with a new representation, the local spherical curve, to explicitly build structure-aware local neighborhoods. This design enhances information aggregation from nearby co-planar points, even in noisy, sparse, or irregularly point distribution settings. Next, we construct" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "a rotation-invariant descriptor based on the local spherical curve and integrate it with a multi-layer perceptron (MLP) to accurately classify edge and non-edge points. Beyond edge classification, we also use the local spherical curve representation to estimate more precise point normals and introduce an optimization function to project points closer to the true edge. To validate our approach, we conduct a series of experiments on the well-known ABC dataset [17] and specialized datasets for thin-walled structures [6]. Extensive experimental results demonstrate that our approach outperforms existing edge detection methods. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.274, + 0.891, + 0.304 + ], + "angle": 0, + "content": "- We introduce STAR-Edge, a new method for detecting and refining edge points in thin-walled structures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.304, + 0.891, + 0.379 + ], + "angle": 0, + "content": "- We develop a new representation, the local spherical curve, to construct structure-aware local neighborhoods, enhancing information aggregation from nearby co-planar points, even under challenging conditions like noise, sparsity, or irregular point distribution." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.38, + 0.891, + 0.455 + ], + "angle": 0, + "content": "- Using the local spherical curve representation, we construct a rotation-invariant local descriptor to achieve more accurate edge point classification. Additionally, this representation allows us to accurately calculate point cloud normals, enabling further refinement of edge points." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.455, + 0.891, + 0.515 + ], + "angle": 0, + "content": "- We perform extensive experiments on specialized datasets for thin-walled structures and the ABC dataset, showing that our method outperforms existing edge extraction techniques." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.274, + 0.891, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.649, + 0.548 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.833, + 0.574 + ], + "angle": 0, + "content": "2.1. Traditional method for edge extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Early edge extraction techniques primarily targeted mesh models [15, 25], with subsequent advancements focusing on direct edge feature extraction from point clouds. Oztireli et al. [23] apply robust kernel regression combined with Moving Least Squares (MLS) to preserve fine details and sharp features in point-based surface representations. Daniels et al. [9] further advance this by introducing a robust method that identifies sharp features in point clouds by generating smooth edge-aligned curves through multi-step refinement and surface fitting. They both rely on iterative refinement processes to progressively extract and enhance sharp edge features in point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Mérigot et al. [20] propose a robust method using Voronoi-based covariance matrices to robustly extract curvature, sharp features, and normals from point clouds. Bazazian et al. [2] present a fast method for sharp edge extraction from unorganized point clouds using eigenvalues of covariance matrices from k-nearest neighbors. Building on these approaches, other researchers have explored various techniques for identifying and preserving critical features in point clouds. For instance, Huang et al. [14] present an" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "edge-aware resampling method that computes reliable normals away from edges while effectively preserving sharp features. Ni et al. [22] propose AGPN, a method for detecting edges and tracing feature lines in 3D point clouds using neighborhood geometry analysis with RANSAC and hybrid region growing. Mineo et al. [21] present a boundary point detection algorithm with fast Fourier transform (FFT)-based filtering to improve edge reconstruction and reduce noise in tessellated surfaces from point clouds. Chen et al. [4] propose a two-phase algorithm for extracting linetype features from unstructured point clouds, using a statistical metric to detect feature points and an anisotropic contracting scheme to reconstruct feature lines against noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.381 + ], + "angle": 0, + "content": "In summary, most traditional point cloud edge extraction methods rely heavily on neighborhood reliability, which includes dense sampling and minimal noise. However, in thin-walled structures, neighborhood information is often ambiguous, making conventional methods generally unsuitable for such structures." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.453, + 0.409 + ], + "angle": 0, + "content": "2.2. Learning-based methods for edge extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.674 + ], + "angle": 0, + "content": "With the emergence of learning techniques, many data-driven approaches have been proposed to improve the robustness of edge extraction and the preservation of sharp edges [5, 34]. EC-Net [32] is introduced as an edge-aware point set consolidation network designed to handle sparse, irregular, and noisy point clouds by both minimizing distances to 3D meshes and edges. PIE-NET [26] uses PointNet++ [24] as the backbone for classification and regression to extract edges from 3D point clouds as parametric curves, and has been extensively validated on synthetic data. CPE-Net [7] is designed in a multitask learning paradigm and uses a learnable weighted least squares fitting approach to accurately extract countersink edges under noisy conditions. Generally, all of these methods initially classify edge points using deep learning techniques, followed by edge generation through upsampling, clustering, or fitting smooth curves to capture edge structures in point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "There are also some methods that focus on learning edge features directly from point clouds. Bode et al. [3] proposed Bounded, a neural network that uses local neighborhood statistics to classify non-edge, sharp-edge, and boundary points in 3D point clouds. Yao et al. [30] propose HGNet, a point cloud network that aggregates hierarchical features from points to super-surfaces. Zhao et al. [33] propose a deep learning approach to detect and consolidate sharp feature edges in 3D point clouds using a multi-task network that predicts displacement vectors for edge refinement. APEE-Net [6] is a dual-task neural network designed for extracting precise edges on aircraft panels by identifying nearby points and also using displacement vectors for repositioning. In addition, NerVE [35] defines a neural volumetric edge representation to extract parametric edge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "curves through graph search, leveraging neural fields rather than traditional point cloud processing. Ye et al. [31] propose a Neural Edge Field method for reconstructing 3D feature curves from multi-view images, optimizing a neural implicit field with a rendering loss to extract 3D edges without 3D supervision. Although these learning-based methods achieve better performance than previous geometric-based methods, they still rely heavily on sufficient contextual information from local point neighborhoods, making them challenging to apply to thin-walled structures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.256, + 0.593, + 0.272 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.895, + 0.479 + ], + "angle": 0, + "content": "The proposed method extracts edges from thin-walled structures, as illustrated in Fig. 2, through three main steps: local spherical curve construction, edge point identification, and edge optimization. First, for each point, we project its sparse neighborhood onto a unit sphere, constructing a closed spherical curve to represent the local neighborhood (Sec 3.1). Next, we use spherical harmonics (SH) to encode this spherical curve as a rotation-invariant local descriptor, which is then passed through an MLP layer to classify the target point as an edge or non-edge point (Sec 3.2). Finally, we leverage this descriptor again to estimate the normal vector, allowing us to formulate a projection function that refines the edge point positions (Sec 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.487, + 0.816, + 0.503 + ], + "angle": 0, + "content": "3.1. Local Spherical Curve Construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.894, + 0.63 + ], + "angle": 0, + "content": "Our key observation is that neighboring points co-planar with a target point align along great circle arcs after spherical projection, allowing the overall shape to be approximated as a closed, continuous curve. Thus, given a point cloud \\( P = \\{p_i \\in \\mathbb{R}^3 | i = 1,2,\\dots,N\\} \\), we first perform spherical normalization on each point \\( p_i \\)'s neighborhood point to obtain the local spherical point set \\( Q_i \\), and then fit a closed, continuous curve." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.894, + 0.69 + ], + "angle": 0, + "content": "Specifically, by taking \\( p_i \\) as the center of the sphere, we normalize the radius of its neighboring points to 1, projecting them onto the unit sphere. The normalization process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.701, + 0.892, + 0.731 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = \\left\\{q _ {j} = \\frac {p _ {j} - p _ {i}}{\\| p _ {j} - p _ {i} \\|} \\right| p _ {j} \\in K N N (p _ {i}) \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.74, + 0.894, + 0.832 + ], + "angle": 0, + "content": "where \\( q_{j} \\) represents the local spherical point and \\( p_j \\) is the KNN neighboring points of \\( p_i \\). Next, we apply Principal Component Analysis (PCA) to \\( Q_{i} \\), projecting it onto a local plane. We then compute the convex hull of the projected points and select the vertices of this convex hull as the key point set \\( K_{i} \\) for curve fitting:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.843, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\nK _ {i} = \\operatorname {C o n v e x H u l l} \\left(\\operatorname {P C A} \\left(Q _ {i}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Finally, we perform spline fitting on \\( K_{i} \\) to obtain a spherical curve, effectively representing the local neighborhood." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27256" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.849, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.34, + 0.895, + 0.384 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. STAR-edge comprises three main steps: (a) constructing a local spherical curve; (b) identifying edge points by integrating rotation-invariant local descriptors with a light-weight MLP layer; and (c) optimizing edge points by adjusting their positions based on the estimated normal vector from the local spherical curve." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.399, + 0.275, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.581, + 0.253, + 0.59 + ], + "angle": 0, + "content": "(a) Thin-walled structure" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.591, + 0.258, + 0.6 + ], + "angle": 0, + "content": "(thickness: 3, resolution: 1)" + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.402, + 0.432, + 0.414 + ], + "angle": 0, + "content": "(b) Local Spherical Points (KNN)" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.415, + 0.391, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.414, + 0.502, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.413, + 0.617, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.412, + 0.729, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.412, + 0.841, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.5, + 0.467, + 0.516 + ], + "angle": 0, + "content": "(c)Key Points and Local Spherical Curve" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.516, + 0.391, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.516, + 0.502, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.516, + 0.617, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.516, + 0.729, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.516, + 0.841, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.617, + 0.895, + 0.66 + ], + "angle": 0, + "content": "Figure 3. Neighborhood distribution of five surface points on the unit sphere. Local spherical points are shown in grey, sampled key points on the fitted spherical Curve in blue, and local spherical curves in red. We observe that the neighborhoods of edge points resemble two semicircles (e.g. ④), while those of non-edge points maintain a complete great circle (e.g. ①)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.473, + 0.899 + ], + "angle": 0, + "content": "Fig. 3 shows the distribution of local spherical points (grey), key points (blue), and the local spherical curve (red) for five target points across a thin-walled structure. Points \\(①\\) and \\(②\\), both located on the top surface, their neighborhoods are close to a complete great circle despite interference from points on other surfaces. Point \\(⑤\\), positioned on a side-end face with sparse sampling, is influenced by both upper and lower surfaces of the thin-walled structure; however, its local spherical curve still largely retains a great circle form. In contrast, points \\(③\\) and \\(④\\) are near a sharp edge, with \\(④\\) closer to the edge than \\(③\\). The local spherical curves of \\(③\\) and \\(④\\) are roughly formed by two semicircles, presenting a different pattern compared to points \\(①\\) and \\(②\\). These results demonstrate that the proposed local spherical" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.686, + 0.872, + 0.701 + ], + "angle": 0, + "content": "curve can robustly and sensitively capture edge features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.729, + 0.727 + ], + "angle": 0, + "content": "3.2. Edge Point Identification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.825 + ], + "angle": 0, + "content": "After obtaining the spherical curve for each point, we use SH, an effective time-frequency domain analysis technique on a sphere, to construct rotation-invariant descriptors. This SH-encoded curve serves as a rotation-invariant local descriptor, which is then fed into an MLP to predict whether each point is an edge point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In detail, we discretize the local spherical curve into a set of scattered points. Each points \\((x,y,z)\\) is mapped to spherical coordinates \\(f(\\theta ,\\phi)\\) on the unit sphere, where the mapping is defined as \\((\\theta ,\\phi) = (\\arccos z,\\arctan y / x)\\). Here, \\(\\theta\\) and \\(\\phi\\) represent the zenith and azimuth angles, respec" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "tively. Let \\( f_{1}, f_{2}, \\ldots, f_{n} \\) be a series of mapping on spherical space. We employ kernel density estimation [10] to construct a smooth, continuous function based on these sample points:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.152, + 0.469, + 0.189 + ], + "angle": 0, + "content": "\\[\nf (\\theta , \\phi) = \\frac {1}{n h} \\sum_ {i = 1} ^ {n} k \\left(\\frac {\\delta_ {i} (\\theta , \\phi)}{h}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.195, + 0.469, + 0.224 + ], + "angle": 0, + "content": "where \\( k \\) is a kernel function, defined as a normal probability density function:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.234, + 0.469, + 0.253 + ], + "angle": 0, + "content": "\\[\nk (x) = 1 / \\sqrt {2 \\pi} \\exp (- 0. 5 x ^ {2}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.262, + 0.469, + 0.351 + ], + "angle": 0, + "content": "The term \\(\\delta_{i}(\\theta, \\phi)\\) denotes the geodesic distance between \\(f_{i}(\\theta_{i}, \\phi_{i})\\) and \\((\\theta, \\phi)\\). The parameter \\(h\\), known as the window width or smoothing parameter, modulates the influence of mapping points \\(f_{i}\\) to the density estimation at \\((\\theta, \\phi)\\). It is defined as \\(h = \\pi / B\\), where \\(B\\) represents the spherical function bandwidth." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.353, + 0.469, + 0.444 + ], + "angle": 0, + "content": "To compute the SH coefficients for the band-limited function, we apply the discrete SH transform. Following the Shannon-Nyquist sampling theorem [11], we sample the model equiangularly at twice the bandwidth \\(B\\) in both latitude and longitude, resulting in \\(2B \\times 2B\\) sampling points \\((\\theta_{j}, \\phi_{k}) \\in (0, \\pi) \\times [0, 2\\pi)\\), with" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.454, + 0.469, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\theta_ {j} = \\pi (2 j + 1) / 4 B & , j = 0, 1, \\dots , 2 B - 1 \\\\ \\phi_ {k} = \\pi k / B & , k = 0, 1, \\dots , 2 B - 1 \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.503, + 0.469, + 0.533 + ], + "angle": 0, + "content": "The SH coefficients of the function \\( f(\\theta, \\phi) \\) are obtained through the discrete SH transform:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.544, + 0.469, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {l} ^ {m} = \\frac {\\sqrt {2 \\pi}}{2 B} \\sum_ {j = 0} ^ {2 B - 1} \\sum_ {k = 0} ^ {2 B - 1} w _ {j} f \\left(\\theta_ {j}, \\phi_ {k}\\right) e ^ {- i m \\phi_ {k}} P _ {l} ^ {m} (\\cos \\theta_ {j}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.6, + 0.469, + 0.645 + ], + "angle": 0, + "content": "for degree \\(0 < l < B\\) and order \\(|m| \\leq l\\). Here \\(P_{l}^{m}\\) denotes the associated Legendre functions defined on \\([-1, 1]\\). The weights \\(w_{j}\\) are computed using the method outlined in [11]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.646, + 0.469, + 0.72 + ], + "angle": 0, + "content": "The SH bases are indexed by two integers, \\(l\\) and \\(m\\), where \\(-l \\leq m \\leq l\\), resulting in \\(2l + 1\\) coefficients for each degree \\(l\\). The parameter \\(l\\) determines the frequency of the basis [16]. Consequently, the \\(2l + 1\\) coefficients corresponding to the same degree \\(l\\) are combined into a vector:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.73, + 0.469, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\beta_ {l} = \\left\\{\\hat {f} _ {l} ^ {m}: m = - l, \\dots , 0, \\dots , l \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.758, + 0.469, + 0.803 + ], + "angle": 0, + "content": "and the \\(\\ell_2\\) norm \\(\\| \\beta_l\\|\\) is taken as the energy of the function \\(f\\) in degree \\(l\\). At last, the rotation-invariant local descriptor \\(D\\) of function \\(f(\\theta ,\\phi)\\) with all \\(B\\) degrees is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.814, + 0.469, + 0.831 + ], + "angle": 0, + "content": "\\[\nD (f, B) = \\{\\| \\beta_ {l} \\|: l = 0, \\dots , B - 1 \\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "After computing this descriptor, we calculate the average descriptor for ground-truth edge points and measure the feature difference between each point and this average descriptor. By applying a threshold to this difference, we can" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.096, + 0.653, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.096, + 0.758, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.096, + 0.869, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.199, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Figure 4. Normal estimation results for points \\(③, ④,\\) and \\(⑤\\) in Fig. 3 (a). In each sub-figure, the red curve is the great circle arc extracted by RANSAC, the blue curve is the non-great circle segment, and the yellow arrow is the estimated normal vector." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.388 + ], + "angle": 0, + "content": "determine whether a point is classified as an edge point. However, in our initial testing, we found that the threshold-based method struggles to effectively handle different edge styles. To address this, we use an MLP to adaptively learn from the rotation-invariant local descriptor, enabling accurate edge point identification. Detailed network structure is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.399, + 0.727, + 0.414 + ], + "angle": 0, + "content": "3.3. Edge Point Optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.542 + ], + "angle": 0, + "content": "We observe that the initially identified edge points are roughly scattered around the true edge. While these points are linearly distributed overall, they form a band-like structure cross two intersecting surfaces when viewed locally. To improve accuracy, we propose constraining edge points to surface intersection regions. This approach is motivated by the fact that an edge point at the intersection of two surfaces has a zero projection distance to both surfaces." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.543, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Let \\( p_i \\) represent an initially-identified edge point. Inspired by [29] and [6], we introduce the following optimization function:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.598, + 0.891, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\min _ {z _ {i}} \\sum_ {p _ {j} \\in K N N (p _ {i})} \\left(\\left(z _ {i} - p _ {j}\\right) \\cdot n _ {j}\\right) ^ {2} + \\mu \\| z _ {i} - p _ {i} \\| ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.644, + 0.892, + 0.719 + ], + "angle": 0, + "content": "where \\( n_j \\) is the normal vector of \\( p_j \\), and \\( \\mu \\) is a balance parameter. The first term encourages \\( z_i \\) to stay close to the tangent planes of neighboring points, while the second term penalizes the deviation of \\( z_i \\) from its initial position \\( p_i \\). In all our experiments, \\( \\mu \\) is set to a default value of 0.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Normal estimation. Note that the edge point optimization method described above relies on accurate normal vectors; however, conventional normal estimation methods is unreliable in noisy or sparse neighborhood conditions. To resolve it, we propose a normal estimation algorithm based on the proposed local spherical curves, which can enhance robustness due to its underlying structure awareness." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, we observe that points co-planar with \\( p_i \\) (shown on the red curve) tend to align on the same plane as the great circle arc after projection onto a sphere. In contrast, other points (shown on the blue curve) lie on a different plane. Therefore, we use this great circle arc" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27258" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "(red) on the local spherical curve to approximate the local plane and derive the normal vector. Specifically, we employ random sampling consensus (RANSAC) to identify the great circle arc on the local spherical curve to determine the normal plane, and use the rest (blue) to determine the normal orientation. Fig. 4 shows the normal estimation of the neighborhood \\(③\\), \\(④\\), and \\(⑤\\) in Fig. 3 (a)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.21, + 0.34, + 0.227 + ], + "angle": 0, + "content": "4. Experiments and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.47, + 0.327 + ], + "angle": 0, + "content": "In this section, we evaluate the effectiveness of our method by comparing it to several learning-based approaches [26, 32, 35] and two traditional methods [4, 29]. We then assess its robustness under varying noise intensities and point cloud resolutions and conduct an ablation study to examine the impact of specific design choices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.335, + 0.254, + 0.351 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Dataset. Given the characteristics of thin-walled structures (small dimensions in the thickness direction), we selected a dataset [6] specifically focused on thin-walled structures for our experiments. This dataset includes 42 distinct 3D models with curved edges of varying shapes, sizes, and curvatures, and thicknesses. For a more detailed description of this dataset, please refer to the supplemental file. Note that a subset of 32 shapes is used for training, while 10 shapes are reserved for testing. For each shape, we sample point clouds at resolutions of 0.3, 0.5, and 0.8. Additionally, three levels of random Gaussian noise are added to the point clouds at a resolution of 0.5. Each sampled point cloud contains approximately 20,000 to 15 million points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.674 + ], + "angle": 0, + "content": "We also perform experiments on the well-known ABC dataset [17], which contains over one million CAD models with edge annotations. To evaluate the generalizability of our method, we select sharp-edged CAD models from the ABC dataset, following the setup in [35]. This subset comprises 2,364 models, which are randomly divided into a training set (80%) and a test set (20%). For each shape, 50,000 points are sampled." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Implementation. All experiments are conducted on an Intel i7-11700 CPU and an NVIDIA RTX 4090 GPU. For constructing the local rotation-invariant descriptor, we use a KNN query with \\( k = 26 \\) and a spherical harmonics transform bandwidth of 10. The proposed approach is implemented in PyTorch and trained over 40 epochs using the Adam optimizer, with a batch size of 1024 and a learning rate of 0.02." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Metrics. We quantitatively evaluate all methods across two aspects. First, point classification performance is assessed using Recall, Precision, F1-score, and Accuracy. Second, to evaluate the quality of the final edge point extraction, we introduce a new metric: Edge Chamfer Distance (ECD). This metric measures the similarity between the optimized edges and the ground-truth edges. The introduction of ECD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "is necessary because the ground-truth sampled point cloud consists of dense points, while the optimized edges may vary in sparsity depending on the chosen resolution. Standard Chamfer Distance, without adjustment, could disproportionately penalize sparse edges, leading to inaccurate similarity assessments. ECD addresses this limitation by providing a more reliable comparison for edge extraction quality. Let \\( X \\) represents the point cloud of the optimized edges, and \\( Y \\) denote the ground-truth edges. The ECD is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.25, + 0.892, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E C D} = \\frac {1}{| X |} \\sum_ {p \\in X} \\min _ {q \\in Y} \\| p - q \\| _ {2} ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Note that for each shape, we label sampled points within a specified distance threshold from the nearest edge as ground-truth edge points. We set empirically this threshold to 0.3 for the dataset in [6] and 0.025 for the ABC dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.365, + 0.805, + 0.38 + ], + "angle": 0, + "content": "4.2. Evaluation on Thin-walled Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.538 + ], + "angle": 0, + "content": "In this section, we compare our STAR-Edge method with state-of-the-art approaches, including EC-Net [32], PIE-Net [26], NerVE [35], MFLE [4], and RFEPS [29], on the thin-walled dataset. All methods use point cloud data with a resolution of 0.5 as input. To accommodate network size limitations, PIE-Net is provided with 8,096 uniformly sampled points obtained through farthest point sampling, while NerVE converts the point cloud into a voxel grid with a resolution of \\(64^3\\). All networks are retrained on the same dataset for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Fig. 5 provides visual comparisons of all methods, while quantitative results are presented in Tab. 1. Both results demonstrate that our approach outperforms others on thin-walled structures, achieving better accuracy and a more uniform edge point distribution. Specifically, EC-Net, which up-samples point clouds in edge regions, tends to generate redundant points around true edges, leading to higher Edge Chamfer Distance (ECD). PIE-Net and NerVE perform less effectively on this dataset; PIE-Net requires dense sampling to ensure its effectiveness in edge extraction, while NerVE's voxel size exceeds the thin-wall thickness, making it struggle to perceive edges. A potential improvement for these methods is to segment small patches for re-processing, as discussed in the supplemental file. The geometry-based methods, MFLE and RFEPS, also show limitations: MFLE requires frequent parameter adjustments, while RFEPS produces noisier edge results. Contrast, our method leverages the local spherical curve to achieve more precise edge points and displacement vectors, resulting in superior performance for edge detection in thin-walled structures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Running time. The running times of various methods on point clouds containing approximately 320,000 points are as follows: our method required 214.0 seconds, MFLE completed in 1.14 seconds, RFEPS took 19 minutes, ECNet" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27259" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.09, + 0.856, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of edge point extraction on the thin-walled structure dataset. -' denotes a failed calculation." + }, + { + "type": "table", + "bbox": [ + 0.175, + 0.114, + 0.794, + 0.203 + ], + "angle": 0, + "content": "
EC-Net [32]PIE-Net [26]MFLE [4]RFEPS [29]STAR-Edge (Ours)
ECD↓0.293652.12040.48391.43920.0587
Recall↑-0.05030.97550.99970.9952
Precision↑-0.02950.30280.15120.5418
F1↑-0.03650.46210.26250.7013
Accuracy↑-0.93490.94700.95440.9931
" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.216, + 0.849, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.591, + 0.893, + 0.621 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of state-of-the-art methods on the thin-walled structure dataset. In these close-up views, our method demonstrates superior accuracy in edge point extraction. Edge points are in red color." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.646, + 0.449, + 0.821 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.833, + 0.443, + 0.848 + ], + "angle": 0, + "content": "Figure 6. Visual comparison under different noise levels." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.471, + 0.901 + ], + "angle": 0, + "content": "finished in 83 seconds, and PIE-Net was the fastest, con" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.645, + 0.871, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.726, + 0.863, + 0.74 + ], + "angle": 0, + "content": "Figure 7. Visual comparison under different resolutions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.77, + 0.892, + 0.816 + ], + "angle": 0, + "content": "suming less than 1.0 second. We attribute PIE-Net's exceptional speed to its down-sampling process. More detailed statistics can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.673, + 0.848 + ], + "angle": 0, + "content": "4.3. Robustness Study" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conducted robustness testing for the proposed STAR-Edge by introducing varying levels of noise and resolution. In our experiments, the thin-walled shapes have a thickness" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27260" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 2. Quantitative classification results under different noise levels and resolutions." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.129, + 0.452, + 0.23 + ], + "angle": 0, + "content": "
NoiseResolution
0.0010.0050.030.30.50.8
ECD↓0.0590.0590.1990.0480.0590.261
Recall↑0.9950.9950.9940.9600.9950.988
Precision↑0.5420.5410.5230.7840.5420.582
F1↑0.7010.7010.6850.8630.7010.731
Accuracy↑0.9930.9930.9930.9980.9930.993
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.245, + 0.468, + 0.272 + ], + "angle": 0, + "content": "Table 3. Quantitative ECD comparisons under different noise and resolution. \\(\\cdot\\) indicates that the calculation failed." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.284, + 0.445, + 0.387 + ], + "angle": 0, + "content": "
MethodsNoiseResolution
0.0010.0050.030.30.50.8
EC-Net0.2950.2930.294-0.2940.518
PIE-Net50.28552.33249.15650.79152.12052.222
RFEPS1.4401.4361.4420.3911.43910.341
MFLE0.4840.4530.4090.4060.4840.742
Ours0.0590.0590.1990.0480.0590.260
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.468, + 0.531 + ], + "angle": 0, + "content": "of 2-3 (which is very small relative to other shape dimension). At a resolution of 0.3, approximately 9 points can be sampled on the side-end surface, whereas at a resolution of 0.8, it becomes challenging to obtain even 4 uniformly distributed points. Testing at a resolution of 0.8 poses a significant robustness challenge, as the limited number of sampling points reduces the reliability of projections onto spherical curves." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.468, + 0.744 + ], + "angle": 0, + "content": "Tab. 2 and Tab. 3 present the effects of noise intensity and resolution on the predictions of our STAR-Edge method. We observe that, even under high noise intensity or significantly reduced resolution, our method consistently produces reasonable edge point predictions with low Edge Chamfer Distance (ECD). As shown in Fig. 6, when the noise level reaches 0.003, the detection performance remains largely unaffected. As shown in Fig. 7, demonstrates that even in an extreme case with a thickness of 3 and a resolution of 0.8, the upper and lower edges can still be clearly distinguished. Additionally, in Tab. 3, we compare the performance of STAR-Edge with state-of-the-art methods across inputs with varying noise levels and resolutions. Our method achieves the best numerical performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.229, + 0.772 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Tab. 4 reports the performance of our method under varying SH bandwidths, which determine the precision and detail of the local feature representation. The results indicate that performance improves with increasing bandwidth; however, higher bandwidth settings also result in increased computational cost. Notably, the performance at a bandwidth of 10 is comparable to that at 15, making a bandwidth of 10 the optimal choice for balancing accuracy and efficiency." + }, + { + "type": "table_caption", + "bbox": [ + 0.558, + 0.09, + 0.834, + 0.103 + ], + "angle": 0, + "content": "Table 4. Ablation study about SH bandwidths." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.115, + 0.869, + 0.175 + ], + "angle": 0, + "content": "
BandwidthRecall↑Precision↑F1↑Accuracy↑ECD↓
50.9880.5230.6850.9910.122
100.9950.5420.7010.9930.059
150.9950.5830.7850.9980.048
" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.19, + 0.877, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.422, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of edge extraction results on 3D shapes from the ABC dataset. Our method achieves more precise edge extraction, particularly in the sharp corner regions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.494, + 0.746, + 0.508 + ], + "angle": 0, + "content": "4.5. Evaluation on ABC dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.64 + ], + "angle": 0, + "content": "We evaluate our method on the ABC dataset following the experimental setup in [35]. Quantitative results are presented in Tab. 5, which demonstrate that STAR-Edge consistently outperforms all baseline methods in terms of edge extraction accuracy. Visual comparisons in Fig. 8 further highlight the superior performance of our method, as it achieves more precise edge extraction, particularly in the sharp corner regions." + }, + { + "type": "table_caption", + "bbox": [ + 0.535, + 0.658, + 0.856, + 0.671 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison on the ABC dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.683, + 0.887, + 0.716 + ], + "angle": 0, + "content": "
EC-NETPIE-NETMFLERFEPSNerVEOurs
ECD↓0.00370.00900.00490.00260.00260.0024
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.753, + 0.618, + 0.768 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we presented STAR-Edge, a novel point cloud edge extraction approach specifically designed for thin-walled structures. The core of our method is the introduction of a new structure-aware neighborhood representation—the local spherical curve. This representation effectively emphasizes co-planar points while minimizing interference from unrelated surfaces, enabling robust and accurate edge detection in complex thin-walled geometries." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "27261" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.161 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No.92367301, No.52425506, No.92267201, No.52275493, No.T2322012)." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.173, + 0.175, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[1] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.31 + ], + "angle": 0, + "content": "[2] Dena Bazazian, Josep R Casas, and Javier Ruiz-Hidalgo. Fast and robust edge extraction in unorganized point clouds. In 2015 international conference on digital image computing: techniques and applications (DICTA), pages 1-8. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[3] Lukas Bode, Michael Weinmann, and Reinhard Klein. Bounded: Neural boundary and edge detection in 3d point clouds via local neighborhood statistics. ISPRS Journal of Photogrammetry and Remote Sensing, 205:334-351, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[4] Honghua Chen, Yaoran Huang, Qian Xie, Yuanpeng Liu, Yuan Zhang, Mingqiang Wei, and Jun Wang. Multiscale feature line extraction from raw point clouds based on local surface variation and anisotropic contraction. IEEE Transactions on Automation Science and Engineering, 19(2):1003-1016, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[5] Honghua Chen, Zeyong Wei, Xianzhi Li, Yabin Xu, Mingqiang Wei, and Jun Wang. Repcd-net: Feature-aware recurrent point cloud denoising network. International Journal of Computer Vision, 130(3):615-629, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.468, + 0.59 + ], + "angle": 0, + "content": "[6] Mengqi Chen, Laishui Zhou, Honghua Chen, and Jun Wang. Thin-walled aircraft panel edge extraction from 3d measurement surfaces via feature-aware displacement learning. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[7] Mengqi Chen, Laishui Zhou, Yongming Zhang, Honghua Chen, and Jun Wang. Aircraft skin countersink primitive extraction from 3-d measurement point clouds via deep clustering and fitting. IEEE Transactions on Instrumentation and Measurement, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.468, + 0.745 + ], + "angle": 0, + "content": "[8] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2727-2735, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.747, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[9] Joel II Daniels, Linh K Ha, Tilo Ochotta, and Claudio T Silva. Robust smooth feature extraction from point clouds. In IEEE International Conference on Shape Modeling and Applications 2007 (SMI'07), pages 123-136. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[10] Richard A Davis, Keh-Shin Lii, and Dimitris N Politis. Remarks on some nonparametric estimates of a density function. Selected Works of Murray Rosenblatt, pages 95-100, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[11] James R Driscoll and Dennis M Healy. Computing fourier transforms and convolutions on the 2-sphere. Advances in applied mathematics, 15(2):202-250, 1994. 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.199, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[12] Timo Hackel, Jan D Wegner, and Konrad Schindler. Contour detection in unstructured 3d point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1610-1618, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[13] Chems-Eddine Himeur, Thibault Lejembre, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[14] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Andreas Hubeli and Markus Gross. Multiresolution feature extraction for unstructured meshes. In Proceedings Visualization, 2001. VIS'01., pages 287-294. IEEE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[16] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Rotation invariant spherical harmonic representation of 3 d shape descriptors. In Symposium on geometry processing, pages 156-164, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[17] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[18] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[19] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[20] Quentin Mérigot, Maks Ovsjanikov, and Leonidas J Guibas. Voronoi-based curvature and feature estimation from point clouds. IEEE Transactions on Visualization and Computer Graphics, 17(6):743-756, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[21] Carmelo Mineo, Stephen Gareth Pierce, and Rahul Summan. Novel algorithms for 3d surface point cloud boundary detection and edge reconstruction. Journal of Computational Design and Engineering, 6(1):81-91, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[22] Huan Ni, Xiangguo Lin, Xiaogang Ning, and Jixian Zhang. Edge detection and feature line tracing in 3d-point clouds by analyzing geometric properties of neighborhoods. Remote Sensing, 8(9):710, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[23] A Cengiz Öztireli, Gael Guennebaud, and Markus Gross. Feature preserving point set surfaces based on non-linear kernel regression. In Computer graphics forum, pages 493-501. Wiley Online Library, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[24] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Charlie CL Wang. Incremental reconstruction of sharp edges on mesh surfaces. Computer-Aided Design, 38(6):689-702, 2006. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27262" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[26] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[27] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[28] Shaobo Xia and Ruisheng Wang. A fast edge extraction method for mobile lidar point clouds. IEEE Geoscience and Remote Sensing Letters, 14(8):1288-1292, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[29] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.332, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[30] Ting Yao, Yehao Li, Yingwei Pan, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[31] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[32] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[33] Tong Zhao, Mulin Yu, Pierre Alliez, and Florent Lafarge. Sharp feature consolidation from raw 3d point clouds via displacement learning. Computer Aided Geometric Design, 103:102204, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[34] Haoran Zhou, Honghua Chen, Yidan Feng, Qiong Wang, Jing Qin, Haoran Xie, Fu Lee Wang, Mingqiang Wei, and Jun Wang. Geometry and learning co-supported normal estimation for unstructured point cloud. In Proceedings of the iee/ivf conference on computer vision and pattern recognition, pages 13238-13247, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.471, + 0.767 + ], + "angle": 0, + "content": "[35] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27263" + } + ] +] \ No newline at end of file diff --git a/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_origin.pdf b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bfa55489023a2c143b12b0969668d847b9fdba97 --- /dev/null +++ b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/f9f356eb-ac90-4c6e-8326-94d9a1d7d184_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3091d6958eb1c94729f262bbb13b7bb82e09a706b886da2e7f520546bda9ab9f +size 2988468 diff --git a/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/full.md b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..78822fe5394b091715b14e000cc76e2ec687b903 --- /dev/null +++ b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/full.md @@ -0,0 +1,316 @@ +# STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds + +Zikuan Li*, Honghua Chen*, Yuecheng Wang, Sibo Wu, Mingqiang Wei, Jun Wang† + +![](images/98f5a62a096ae4ee4d06bc8fdaa57115cf4a9d159e40a69040f6f5331176347b.jpg) +Figure 1. The characteristics of our method and visual comparison of edge extraction results in thin-walled structures. (a) defines the concept of a thin-walled edge. (b) highlights that the primary challenge in thin-walled edge extraction is the sensitivity to local neighborhood selection. (c) illustrates how a larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring boundary information. (d) shows that a smaller neighborhood may suffer from noise and insufficient sampling on side faces, lacking the contextual information needed for accurate edge point recognition. (e) depicts the spherical projection of the local neighborhood of four points, where points of the same color lie on the same underlying surface. This illustrates our key observation that points co-planar with the neighborhood center tend to align along a great circle arc after spherical projection. (f) provides a qualitative comparison of our STAR-Edge with RFEPS [29], EC-Net [32], and MFLE [4]. + +# Abstract + +Extracting geometric edges from unstructured point clouds remains a significant challenge, particularly in thin-walled structures that are commonly found in everyday objects. Traditional geometric methods and recent learning-based approaches frequently struggle with these structures, as both rely heavily on sufficient contextual information from local point neighborhoods. However, 3D measurement data of thin-walled structures often lack the accurate, dense, and regular neighborhood sampling required for reliable edge extraction, resulting in degraded performance. + +In this work, we introduce STAR-Edge, a novel approach designed for detecting and refining edge points in thin-walled structures. Our method leverages a unique representation—the local spherical curve—to create structure-aware neighborhoods that emphasize co-planar points while reducing interference from close-by, non-co-planar surfaces. This representation is transformed into a rotation-invariant descriptor, which, combined with a lightweight + +multi-layer perceptron, enables robust edge point classification even in the presence of noise and sparse or irregular sampling. Besides, we also use the local spherical curve representation to estimate more precise normals and introduce an optimization function to project initially identified edge points exactly on the true edges. Experiments conducted on the ABC dataset and thin-walled structure-specific datasets demonstrate that STAR-Edge outperforms existing edge detection methods, showcasing better robustness under various challenging conditions. The source code is available at https://github.com/Miraclelezk/STAR-Edge. + +# 1. Introduction + +Thin-walled structures are essential in our daily lives, seen in items like mechanical parts, furniture, and various manufactured objects. These structures often contain thin-walled edges, defined as boundary lines where surfaces intersect at their side-end faces, characterized by a minimal thickness relative to other dimensions (see Fig. 1 (a)). As geometric + +edges provide an abstraction of complex 3D shapes, extracting edges from 3D measurements (e.g. unstructured point clouds) in these structures is crucial for downstream tasks. + +Extracting geometric edges from point clouds has been a long-standing problem. Traditional techniques generally rely on local geometric properties, such as eigenvalues of the covariance matrix [2, 28], normals [27], curvature [12], or elaborately-designed descriptors [4]. While effective to some extent, these methods depend on first- or second-order derivatives estimated from local neighborhoods, making them particularly sensitive to point disturbances and sampling sparsity or irregularities, especially on thick side-end faces. More recently, some deep neural networks [1, 8, 13, 13, 18, 19, 32] have been proposed to enhance edge accuracy by leveraging deep features [26], neural edge representations [35], or novel point representations (e.g., Scale Space Matrix [18]). However, these methods still struggle with thin-walled structures (as seen in Fig. 1 (f)), as they rely on extracting useful information from valid local neighborhoods. + +To be more specific, the local neighborhood of points around thin-walled edges is often unreliable. A larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring the boundary information (see Fig. 1 (c)). Conversely, a smaller neighborhood may suffer from noise and insufficient sampling on side-end faces, lacking the contextual information needed for accurate edge point recognition by those above networks (see Fig. 1 (d)). Although multi-scale schemes [26] can partially mitigate this issue, we argue that this approach is sub-optimal, as it requires hard-coding several fixed neighborhood scales. + +Ideally, each point's local representation should be structure-aware. This means it should reliably emphasize points on the same underlying surface (e.g., the same plane) while minimizing interference from points on different structures (e.g., another plane). Interestingly, we observe that points co-planar with the neighborhood center tend to align along a great circle arc (dashed circle on the blue plane in Fig. 1 (e)) after spherical projection, effectively reflecting the underlying surface of the thin-walled structure. Based on this observation, naturally, we can construct a local spherical curve to incorporate as many points on the great circle arc as possible, thereby enhancing points on the same underlying surface while reducing the influence of points from other surfaces. + +To this end, we present STAR-Edge, a novel approach for detecting and refining edge points in thin-walled structures. Our method begins with a new representation, the local spherical curve, to explicitly build structure-aware local neighborhoods. This design enhances information aggregation from nearby co-planar points, even in noisy, sparse, or irregularly point distribution settings. Next, we construct + +a rotation-invariant descriptor based on the local spherical curve and integrate it with a multi-layer perceptron (MLP) to accurately classify edge and non-edge points. Beyond edge classification, we also use the local spherical curve representation to estimate more precise point normals and introduce an optimization function to project points closer to the true edge. To validate our approach, we conduct a series of experiments on the well-known ABC dataset [17] and specialized datasets for thin-walled structures [6]. Extensive experimental results demonstrate that our approach outperforms existing edge detection methods. Our contributions are summarized as follows: + +- We introduce STAR-Edge, a new method for detecting and refining edge points in thin-walled structures. +- We develop a new representation, the local spherical curve, to construct structure-aware local neighborhoods, enhancing information aggregation from nearby co-planar points, even under challenging conditions like noise, sparsity, or irregular point distribution. +- Using the local spherical curve representation, we construct a rotation-invariant local descriptor to achieve more accurate edge point classification. Additionally, this representation allows us to accurately calculate point cloud normals, enabling further refinement of edge points. +- We perform extensive experiments on specialized datasets for thin-walled structures and the ABC dataset, showing that our method outperforms existing edge extraction techniques. + +# 2. Related Works + +# 2.1. Traditional method for edge extraction + +Early edge extraction techniques primarily targeted mesh models [15, 25], with subsequent advancements focusing on direct edge feature extraction from point clouds. Oztireli et al. [23] apply robust kernel regression combined with Moving Least Squares (MLS) to preserve fine details and sharp features in point-based surface representations. Daniels et al. [9] further advance this by introducing a robust method that identifies sharp features in point clouds by generating smooth edge-aligned curves through multi-step refinement and surface fitting. They both rely on iterative refinement processes to progressively extract and enhance sharp edge features in point cloud data. + +Mérigot et al. [20] propose a robust method using Voronoi-based covariance matrices to robustly extract curvature, sharp features, and normals from point clouds. Bazazian et al. [2] present a fast method for sharp edge extraction from unorganized point clouds using eigenvalues of covariance matrices from k-nearest neighbors. Building on these approaches, other researchers have explored various techniques for identifying and preserving critical features in point clouds. For instance, Huang et al. [14] present an + +edge-aware resampling method that computes reliable normals away from edges while effectively preserving sharp features. Ni et al. [22] propose AGPN, a method for detecting edges and tracing feature lines in 3D point clouds using neighborhood geometry analysis with RANSAC and hybrid region growing. Mineo et al. [21] present a boundary point detection algorithm with fast Fourier transform (FFT)-based filtering to improve edge reconstruction and reduce noise in tessellated surfaces from point clouds. Chen et al. [4] propose a two-phase algorithm for extracting linetype features from unstructured point clouds, using a statistical metric to detect feature points and an anisotropic contracting scheme to reconstruct feature lines against noise. + +In summary, most traditional point cloud edge extraction methods rely heavily on neighborhood reliability, which includes dense sampling and minimal noise. However, in thin-walled structures, neighborhood information is often ambiguous, making conventional methods generally unsuitable for such structures. + +# 2.2. Learning-based methods for edge extraction + +With the emergence of learning techniques, many data-driven approaches have been proposed to improve the robustness of edge extraction and the preservation of sharp edges [5, 34]. EC-Net [32] is introduced as an edge-aware point set consolidation network designed to handle sparse, irregular, and noisy point clouds by both minimizing distances to 3D meshes and edges. PIE-NET [26] uses PointNet++ [24] as the backbone for classification and regression to extract edges from 3D point clouds as parametric curves, and has been extensively validated on synthetic data. CPE-Net [7] is designed in a multitask learning paradigm and uses a learnable weighted least squares fitting approach to accurately extract countersink edges under noisy conditions. Generally, all of these methods initially classify edge points using deep learning techniques, followed by edge generation through upsampling, clustering, or fitting smooth curves to capture edge structures in point clouds. + +There are also some methods that focus on learning edge features directly from point clouds. Bode et al. [3] proposed Bounded, a neural network that uses local neighborhood statistics to classify non-edge, sharp-edge, and boundary points in 3D point clouds. Yao et al. [30] propose HGNet, a point cloud network that aggregates hierarchical features from points to super-surfaces. Zhao et al. [33] propose a deep learning approach to detect and consolidate sharp feature edges in 3D point clouds using a multi-task network that predicts displacement vectors for edge refinement. APEE-Net [6] is a dual-task neural network designed for extracting precise edges on aircraft panels by identifying nearby points and also using displacement vectors for repositioning. In addition, NerVE [35] defines a neural volumetric edge representation to extract parametric edge + +curves through graph search, leveraging neural fields rather than traditional point cloud processing. Ye et al. [31] propose a Neural Edge Field method for reconstructing 3D feature curves from multi-view images, optimizing a neural implicit field with a rendering loss to extract 3D edges without 3D supervision. Although these learning-based methods achieve better performance than previous geometric-based methods, they still rely heavily on sufficient contextual information from local point neighborhoods, making them challenging to apply to thin-walled structures. + +# 3. Method + +The proposed method extracts edges from thin-walled structures, as illustrated in Fig. 2, through three main steps: local spherical curve construction, edge point identification, and edge optimization. First, for each point, we project its sparse neighborhood onto a unit sphere, constructing a closed spherical curve to represent the local neighborhood (Sec 3.1). Next, we use spherical harmonics (SH) to encode this spherical curve as a rotation-invariant local descriptor, which is then passed through an MLP layer to classify the target point as an edge or non-edge point (Sec 3.2). Finally, we leverage this descriptor again to estimate the normal vector, allowing us to formulate a projection function that refines the edge point positions (Sec 3.3). + +# 3.1. Local Spherical Curve Construction + +Our key observation is that neighboring points co-planar with a target point align along great circle arcs after spherical projection, allowing the overall shape to be approximated as a closed, continuous curve. Thus, given a point cloud $P = \{p_i \in \mathbb{R}^3 | i = 1,2,\dots,N\}$ , we first perform spherical normalization on each point $p_i$ 's neighborhood point to obtain the local spherical point set $Q_i$ , and then fit a closed, continuous curve. + +Specifically, by taking $p_i$ as the center of the sphere, we normalize the radius of its neighboring points to 1, projecting them onto the unit sphere. The normalization process is defined as follows: + +$$ +Q _ {i} = \left\{q _ {j} = \frac {p _ {j} - p _ {i}}{\| p _ {j} - p _ {i} \|} \right| p _ {j} \in K N N (p _ {i}) \}, \tag {1} +$$ + +where $q_{j}$ represents the local spherical point and $p_j$ is the KNN neighboring points of $p_i$ . Next, we apply Principal Component Analysis (PCA) to $Q_{i}$ , projecting it onto a local plane. We then compute the convex hull of the projected points and select the vertices of this convex hull as the key point set $K_{i}$ for curve fitting: + +$$ +K _ {i} = \operatorname {C o n v e x H u l l} \left(\operatorname {P C A} \left(Q _ {i}\right)\right). \tag {2} +$$ + +Finally, we perform spline fitting on $K_{i}$ to obtain a spherical curve, effectively representing the local neighborhood. + +![](images/d4fe5876330670f17164a4752f80ed0e2df45c0213380d259f5d17fca3afc0d6.jpg) +Figure 2. Overview of the proposed method. STAR-edge comprises three main steps: (a) constructing a local spherical curve; (b) identifying edge points by integrating rotation-invariant local descriptors with a light-weight MLP layer; and (c) optimizing edge points by adjusting their positions based on the estimated normal vector from the local spherical curve. + +![](images/4b18ae857a294db13b949e76cd23e925b88b689021907ffd43ebc3287400dd9f.jpg) +(a) Thin-walled structure +(thickness: 3, resolution: 1) + +![](images/5dbd7bfb943a5a34c01123d6b01430535559fa21babdd5938ef7dca5a13a10d7.jpg) + +![](images/9b31f883d628d5a3cbae2b5c30eeb69ece875b9d76e0d67dbea4f4774311e894.jpg) +(b) Local Spherical Points (KNN) + +![](images/e32ae342cdbf243a442db177d3802b49e0d60aebc431cc43d12a1b4b49f61147.jpg) + +![](images/b2afbb0bbfdf7671a64e44a8fb9e371c88b662a37e4b35bd1ae1343d6006b21e.jpg) + +![](images/15bba3a42695e6de10214e6cd1afd8a49eff7227796e0ecdfe066e25bf3239a8.jpg) + +![](images/1e102264fdf6228d226c129cd0b406fa9b9a7d4efaa82a3a2da8a342b4d23cfe.jpg) +(c)Key Points and Local Spherical Curve +Figure 3. Neighborhood distribution of five surface points on the unit sphere. Local spherical points are shown in grey, sampled key points on the fitted spherical Curve in blue, and local spherical curves in red. We observe that the neighborhoods of edge points resemble two semicircles (e.g. ④), while those of non-edge points maintain a complete great circle (e.g. ①). + +![](images/d3425c27ebc8e7c8f5ece34a3b130b87cba8d131cc0b9e2fde77ea3cfb6d4e84.jpg) + +![](images/a652af4d6bf527d8f3a7d763f9fcc800da8b3124aab871c67296b5eb24518d8b.jpg) + +![](images/ea8484b7b5b54f3e232113a9bc2ac0a54452fe3e55fc313126c2ff9472d3980d.jpg) + +![](images/0713654c7b5dc4e68444c7b6e270359d0002009f581d181af16577be97a5db56.jpg) + +Fig. 3 shows the distribution of local spherical points (grey), key points (blue), and the local spherical curve (red) for five target points across a thin-walled structure. Points $①$ and $②$ , both located on the top surface, their neighborhoods are close to a complete great circle despite interference from points on other surfaces. Point $⑤$ , positioned on a side-end face with sparse sampling, is influenced by both upper and lower surfaces of the thin-walled structure; however, its local spherical curve still largely retains a great circle form. In contrast, points $③$ and $④$ are near a sharp edge, with $④$ closer to the edge than $③$ . The local spherical curves of $③$ and $④$ are roughly formed by two semicircles, presenting a different pattern compared to points $①$ and $②$ . These results demonstrate that the proposed local spherical + +curve can robustly and sensitively capture edge features. + +# 3.2. Edge Point Identification + +After obtaining the spherical curve for each point, we use SH, an effective time-frequency domain analysis technique on a sphere, to construct rotation-invariant descriptors. This SH-encoded curve serves as a rotation-invariant local descriptor, which is then fed into an MLP to predict whether each point is an edge point. + +In detail, we discretize the local spherical curve into a set of scattered points. Each points $(x,y,z)$ is mapped to spherical coordinates $f(\theta ,\phi)$ on the unit sphere, where the mapping is defined as $(\theta ,\phi) = (\arccos z,\arctan y / x)$ . Here, $\theta$ and $\phi$ represent the zenith and azimuth angles, respec + +tively. Let $f_{1}, f_{2}, \ldots, f_{n}$ be a series of mapping on spherical space. We employ kernel density estimation [10] to construct a smooth, continuous function based on these sample points: + +$$ +f (\theta , \phi) = \frac {1}{n h} \sum_ {i = 1} ^ {n} k \left(\frac {\delta_ {i} (\theta , \phi)}{h}\right), \tag {3} +$$ + +where $k$ is a kernel function, defined as a normal probability density function: + +$$ +k (x) = 1 / \sqrt {2 \pi} \exp (- 0. 5 x ^ {2}). \tag {4} +$$ + +The term $\delta_{i}(\theta, \phi)$ denotes the geodesic distance between $f_{i}(\theta_{i}, \phi_{i})$ and $(\theta, \phi)$ . The parameter $h$ , known as the window width or smoothing parameter, modulates the influence of mapping points $f_{i}$ to the density estimation at $(\theta, \phi)$ . It is defined as $h = \pi / B$ , where $B$ represents the spherical function bandwidth. + +To compute the SH coefficients for the band-limited function, we apply the discrete SH transform. Following the Shannon-Nyquist sampling theorem [11], we sample the model equiangularly at twice the bandwidth $B$ in both latitude and longitude, resulting in $2B \times 2B$ sampling points $(\theta_{j}, \phi_{k}) \in (0, \pi) \times [0, 2\pi)$ , with + +$$ +\left\{ \begin{array}{l l} \theta_ {j} = \pi (2 j + 1) / 4 B & , j = 0, 1, \dots , 2 B - 1 \\ \phi_ {k} = \pi k / B & , k = 0, 1, \dots , 2 B - 1 \end{array} \right. \tag {5} +$$ + +The SH coefficients of the function $f(\theta, \phi)$ are obtained through the discrete SH transform: + +$$ +\hat {f} _ {l} ^ {m} = \frac {\sqrt {2 \pi}}{2 B} \sum_ {j = 0} ^ {2 B - 1} \sum_ {k = 0} ^ {2 B - 1} w _ {j} f \left(\theta_ {j}, \phi_ {k}\right) e ^ {- i m \phi_ {k}} P _ {l} ^ {m} (\cos \theta_ {j}) \tag {6} +$$ + +for degree $0 < l < B$ and order $|m| \leq l$ . Here $P_{l}^{m}$ denotes the associated Legendre functions defined on $[-1, 1]$ . The weights $w_{j}$ are computed using the method outlined in [11]. + +The SH bases are indexed by two integers, $l$ and $m$ , where $-l \leq m \leq l$ , resulting in $2l + 1$ coefficients for each degree $l$ . The parameter $l$ determines the frequency of the basis [16]. Consequently, the $2l + 1$ coefficients corresponding to the same degree $l$ are combined into a vector: + +$$ +\beta_ {l} = \left\{\hat {f} _ {l} ^ {m}: m = - l, \dots , 0, \dots , l \right\}, \tag {7} +$$ + +and the $\ell_2$ norm $\| \beta_l\|$ is taken as the energy of the function $f$ in degree $l$ . At last, the rotation-invariant local descriptor $D$ of function $f(\theta ,\phi)$ with all $B$ degrees is defined as + +$$ +D (f, B) = \{\| \beta_ {l} \|: l = 0, \dots , B - 1 \}. \tag {8} +$$ + +After computing this descriptor, we calculate the average descriptor for ground-truth edge points and measure the feature difference between each point and this average descriptor. By applying a threshold to this difference, we can + +![](images/e1ab2b6b4a35c32bf4bc3b1c4ea69961798032ad386da1d1a5fe5b3c2cf02bfa.jpg) +Figure 4. Normal estimation results for points $③, ④,$ and $⑤$ in Fig. 3 (a). In each sub-figure, the red curve is the great circle arc extracted by RANSAC, the blue curve is the non-great circle segment, and the yellow arrow is the estimated normal vector. + +![](images/cf34b428f812ccc2c919596b917bc0dc84b30ac69aecc9bb09e8c9309676927d.jpg) + +![](images/d9d3fd4c673dbf0502377d052aa421fc03b96ff9bce4c7eb594b5f2bd8f4cbb0.jpg) + +determine whether a point is classified as an edge point. However, in our initial testing, we found that the threshold-based method struggles to effectively handle different edge styles. To address this, we use an MLP to adaptively learn from the rotation-invariant local descriptor, enabling accurate edge point identification. Detailed network structure is provided in the supplementary material. + +# 3.3. Edge Point Optimization + +We observe that the initially identified edge points are roughly scattered around the true edge. While these points are linearly distributed overall, they form a band-like structure cross two intersecting surfaces when viewed locally. To improve accuracy, we propose constraining edge points to surface intersection regions. This approach is motivated by the fact that an edge point at the intersection of two surfaces has a zero projection distance to both surfaces. + +Let $p_i$ represent an initially-identified edge point. Inspired by [29] and [6], we introduce the following optimization function: + +$$ +\min _ {z _ {i}} \sum_ {p _ {j} \in K N N (p _ {i})} \left(\left(z _ {i} - p _ {j}\right) \cdot n _ {j}\right) ^ {2} + \mu \| z _ {i} - p _ {i} \| ^ {2}, \tag {9} +$$ + +where $n_j$ is the normal vector of $p_j$ , and $\mu$ is a balance parameter. The first term encourages $z_i$ to stay close to the tangent planes of neighboring points, while the second term penalizes the deviation of $z_i$ from its initial position $p_i$ . In all our experiments, $\mu$ is set to a default value of 0.1. + +Normal estimation. Note that the edge point optimization method described above relies on accurate normal vectors; however, conventional normal estimation methods is unreliable in noisy or sparse neighborhood conditions. To resolve it, we propose a normal estimation algorithm based on the proposed local spherical curves, which can enhance robustness due to its underlying structure awareness. + +As illustrated in Fig. 4, we observe that points co-planar with $p_i$ (shown on the red curve) tend to align on the same plane as the great circle arc after projection onto a sphere. In contrast, other points (shown on the blue curve) lie on a different plane. Therefore, we use this great circle arc + +(red) on the local spherical curve to approximate the local plane and derive the normal vector. Specifically, we employ random sampling consensus (RANSAC) to identify the great circle arc on the local spherical curve to determine the normal plane, and use the rest (blue) to determine the normal orientation. Fig. 4 shows the normal estimation of the neighborhood $③$ , $④$ , and $⑤$ in Fig. 3 (a). + +# 4. Experiments and Discussion + +In this section, we evaluate the effectiveness of our method by comparing it to several learning-based approaches [26, 32, 35] and two traditional methods [4, 29]. We then assess its robustness under varying noise intensities and point cloud resolutions and conduct an ablation study to examine the impact of specific design choices. + +# 4.1. Experiment Setup + +Dataset. Given the characteristics of thin-walled structures (small dimensions in the thickness direction), we selected a dataset [6] specifically focused on thin-walled structures for our experiments. This dataset includes 42 distinct 3D models with curved edges of varying shapes, sizes, and curvatures, and thicknesses. For a more detailed description of this dataset, please refer to the supplemental file. Note that a subset of 32 shapes is used for training, while 10 shapes are reserved for testing. For each shape, we sample point clouds at resolutions of 0.3, 0.5, and 0.8. Additionally, three levels of random Gaussian noise are added to the point clouds at a resolution of 0.5. Each sampled point cloud contains approximately 20,000 to 15 million points. + +We also perform experiments on the well-known ABC dataset [17], which contains over one million CAD models with edge annotations. To evaluate the generalizability of our method, we select sharp-edged CAD models from the ABC dataset, following the setup in [35]. This subset comprises 2,364 models, which are randomly divided into a training set (80%) and a test set (20%). For each shape, 50,000 points are sampled. + +Implementation. All experiments are conducted on an Intel i7-11700 CPU and an NVIDIA RTX 4090 GPU. For constructing the local rotation-invariant descriptor, we use a KNN query with $k = 26$ and a spherical harmonics transform bandwidth of 10. The proposed approach is implemented in PyTorch and trained over 40 epochs using the Adam optimizer, with a batch size of 1024 and a learning rate of 0.02. + +Metrics. We quantitatively evaluate all methods across two aspects. First, point classification performance is assessed using Recall, Precision, F1-score, and Accuracy. Second, to evaluate the quality of the final edge point extraction, we introduce a new metric: Edge Chamfer Distance (ECD). This metric measures the similarity between the optimized edges and the ground-truth edges. The introduction of ECD + +is necessary because the ground-truth sampled point cloud consists of dense points, while the optimized edges may vary in sparsity depending on the chosen resolution. Standard Chamfer Distance, without adjustment, could disproportionately penalize sparse edges, leading to inaccurate similarity assessments. ECD addresses this limitation by providing a more reliable comparison for edge extraction quality. Let $X$ represents the point cloud of the optimized edges, and $Y$ denote the ground-truth edges. The ECD is defined as follows: + +$$ +\mathrm {E C D} = \frac {1}{| X |} \sum_ {p \in X} \min _ {q \in Y} \| p - q \| _ {2} ^ {2} \tag {10} +$$ + +Note that for each shape, we label sampled points within a specified distance threshold from the nearest edge as ground-truth edge points. We set empirically this threshold to 0.3 for the dataset in [6] and 0.025 for the ABC dataset. + +# 4.2. Evaluation on Thin-walled Dataset + +In this section, we compare our STAR-Edge method with state-of-the-art approaches, including EC-Net [32], PIE-Net [26], NerVE [35], MFLE [4], and RFEPS [29], on the thin-walled dataset. All methods use point cloud data with a resolution of 0.5 as input. To accommodate network size limitations, PIE-Net is provided with 8,096 uniformly sampled points obtained through farthest point sampling, while NerVE converts the point cloud into a voxel grid with a resolution of $64^3$ . All networks are retrained on the same dataset for a fair comparison. + +Fig. 5 provides visual comparisons of all methods, while quantitative results are presented in Tab. 1. Both results demonstrate that our approach outperforms others on thin-walled structures, achieving better accuracy and a more uniform edge point distribution. Specifically, EC-Net, which up-samples point clouds in edge regions, tends to generate redundant points around true edges, leading to higher Edge Chamfer Distance (ECD). PIE-Net and NerVE perform less effectively on this dataset; PIE-Net requires dense sampling to ensure its effectiveness in edge extraction, while NerVE's voxel size exceeds the thin-wall thickness, making it struggle to perceive edges. A potential improvement for these methods is to segment small patches for re-processing, as discussed in the supplemental file. The geometry-based methods, MFLE and RFEPS, also show limitations: MFLE requires frequent parameter adjustments, while RFEPS produces noisier edge results. Contrast, our method leverages the local spherical curve to achieve more precise edge points and displacement vectors, resulting in superior performance for edge detection in thin-walled structures. + +Running time. The running times of various methods on point clouds containing approximately 320,000 points are as follows: our method required 214.0 seconds, MFLE completed in 1.14 seconds, RFEPS took 19 minutes, ECNet + +Table 1. Quantitative comparison of edge point extraction on the thin-walled structure dataset. -' denotes a failed calculation. + +
EC-Net [32]PIE-Net [26]MFLE [4]RFEPS [29]STAR-Edge (Ours)
ECD↓0.293652.12040.48391.43920.0587
Recall↑-0.05030.97550.99970.9952
Precision↑-0.02950.30280.15120.5418
F1↑-0.03650.46210.26250.7013
Accuracy↑-0.93490.94700.95440.9931
+ +![](images/beb9c451fa1c1c6bda91c1370ab74f704a45c162fbc8b4c6a2ab34917050dbd7.jpg) +Figure 5. Visual comparison of state-of-the-art methods on the thin-walled structure dataset. In these close-up views, our method demonstrates superior accuracy in edge point extraction. Edge points are in red color. + +![](images/479f791b9d8bca0a43db8d5dcd4b42d1c367a367b0a74d4c52a3a6ac9ec24b09.jpg) +Figure 6. Visual comparison under different noise levels. + +finished in 83 seconds, and PIE-Net was the fastest, con + +![](images/db7ec828cd9bf8b8c838aef7c19a759d2f8dd3815f0305bb8fdb98b952d67deb.jpg) +Figure 7. Visual comparison under different resolutions. + +suming less than 1.0 second. We attribute PIE-Net's exceptional speed to its down-sampling process. More detailed statistics can be found in the supplementary materials. + +# 4.3. Robustness Study + +We conducted robustness testing for the proposed STAR-Edge by introducing varying levels of noise and resolution. In our experiments, the thin-walled shapes have a thickness + +Table 2. Quantitative classification results under different noise levels and resolutions. + +
NoiseResolution
0.0010.0050.030.30.50.8
ECD↓0.0590.0590.1990.0480.0590.261
Recall↑0.9950.9950.9940.9600.9950.988
Precision↑0.5420.5410.5230.7840.5420.582
F1↑0.7010.7010.6850.8630.7010.731
Accuracy↑0.9930.9930.9930.9980.9930.993
+ +Table 3. Quantitative ECD comparisons under different noise and resolution. $\cdot$ indicates that the calculation failed. + +
MethodsNoiseResolution
0.0010.0050.030.30.50.8
EC-Net0.2950.2930.294-0.2940.518
PIE-Net50.28552.33249.15650.79152.12052.222
RFEPS1.4401.4361.4420.3911.43910.341
MFLE0.4840.4530.4090.4060.4840.742
Ours0.0590.0590.1990.0480.0590.260
+ +of 2-3 (which is very small relative to other shape dimension). At a resolution of 0.3, approximately 9 points can be sampled on the side-end surface, whereas at a resolution of 0.8, it becomes challenging to obtain even 4 uniformly distributed points. Testing at a resolution of 0.8 poses a significant robustness challenge, as the limited number of sampling points reduces the reliability of projections onto spherical curves. + +Tab. 2 and Tab. 3 present the effects of noise intensity and resolution on the predictions of our STAR-Edge method. We observe that, even under high noise intensity or significantly reduced resolution, our method consistently produces reasonable edge point predictions with low Edge Chamfer Distance (ECD). As shown in Fig. 6, when the noise level reaches 0.003, the detection performance remains largely unaffected. As shown in Fig. 7, demonstrates that even in an extreme case with a thickness of 3 and a resolution of 0.8, the upper and lower edges can still be clearly distinguished. Additionally, in Tab. 3, we compare the performance of STAR-Edge with state-of-the-art methods across inputs with varying noise levels and resolutions. Our method achieves the best numerical performance. + +# 4.4. Ablation Study + +Tab. 4 reports the performance of our method under varying SH bandwidths, which determine the precision and detail of the local feature representation. The results indicate that performance improves with increasing bandwidth; however, higher bandwidth settings also result in increased computational cost. Notably, the performance at a bandwidth of 10 is comparable to that at 15, making a bandwidth of 10 the optimal choice for balancing accuracy and efficiency. + +Table 4. Ablation study about SH bandwidths. + +
BandwidthRecall↑Precision↑F1↑Accuracy↑ECD↓
50.9880.5230.6850.9910.122
100.9950.5420.7010.9930.059
150.9950.5830.7850.9980.048
+ +![](images/d6682244787573d87be17401da0b15e2d6d1fd8c4dde4fd33ef25d3cc1312eda.jpg) +Figure 8. Visual comparison of edge extraction results on 3D shapes from the ABC dataset. Our method achieves more precise edge extraction, particularly in the sharp corner regions. + +# 4.5. Evaluation on ABC dataset + +We evaluate our method on the ABC dataset following the experimental setup in [35]. Quantitative results are presented in Tab. 5, which demonstrate that STAR-Edge consistently outperforms all baseline methods in terms of edge extraction accuracy. Visual comparisons in Fig. 8 further highlight the superior performance of our method, as it achieves more precise edge extraction, particularly in the sharp corner regions. + +Table 5. Quantitative comparison on the ABC dataset. + +
EC-NETPIE-NETMFLERFEPSNerVEOurs
ECD↓0.00370.00900.00490.00260.00260.0024
+ +# 5. Conclusion + +In this work, we presented STAR-Edge, a novel point cloud edge extraction approach specifically designed for thin-walled structures. The core of our method is the introduction of a new structure-aware neighborhood representation—the local spherical curve. This representation effectively emphasizes co-planar points while minimizing interference from unrelated surfaces, enabling robust and accurate edge detection in complex thin-walled geometries. + +# 6. Acknowledgments + +This work was supported by the National Natural Science Foundation of China (No.92367301, No.52425506, No.92267201, No.52275493, No.T2322012). + +# References + +[1] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 2 +[2] Dena Bazazian, Josep R Casas, and Javier Ruiz-Hidalgo. Fast and robust edge extraction in unorganized point clouds. In 2015 international conference on digital image computing: techniques and applications (DICTA), pages 1-8. IEEE, 2015. 2 +[3] Lukas Bode, Michael Weinmann, and Reinhard Klein. Bounded: Neural boundary and edge detection in 3d point clouds via local neighborhood statistics. ISPRS Journal of Photogrammetry and Remote Sensing, 205:334-351, 2023. 3 +[4] Honghua Chen, Yaoran Huang, Qian Xie, Yuanpeng Liu, Yuan Zhang, Mingqiang Wei, and Jun Wang. Multiscale feature line extraction from raw point clouds based on local surface variation and anisotropic contraction. IEEE Transactions on Automation Science and Engineering, 19(2):1003-1016, 2021. 1, 2, 3, 6, 7 +[5] Honghua Chen, Zeyong Wei, Xianzhi Li, Yabin Xu, Mingqiang Wei, and Jun Wang. Repcd-net: Feature-aware recurrent point cloud denoising network. International Journal of Computer Vision, 130(3):615-629, 2022. 3 +[6] Mengqi Chen, Laishui Zhou, Honghua Chen, and Jun Wang. Thin-walled aircraft panel edge extraction from 3d measurement surfaces via feature-aware displacement learning. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3, 5, 6 +[7] Mengqi Chen, Laishui Zhou, Yongming Zhang, Honghua Chen, and Jun Wang. Aircraft skin countersink primitive extraction from 3-d measurement point clouds via deep clustering and fitting. IEEE Transactions on Instrumentation and Measurement, 2024. 3 +[8] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2727-2735, 2023. 2 +[9] Joel II Daniels, Linh K Ha, Tilo Ochotta, and Claudio T Silva. Robust smooth feature extraction from point clouds. In IEEE International Conference on Shape Modeling and Applications 2007 (SMI'07), pages 123-136. IEEE, 2007. 2 +[10] Richard A Davis, Keh-Shin Lii, and Dimitris N Politis. Remarks on some nonparametric estimates of a density function. Selected Works of Murray Rosenblatt, pages 95-100, 2011. 5 +[11] James R Driscoll and Dennis M Healy. Computing fourier transforms and convolutions on the 2-sphere. Advances in applied mathematics, 15(2):202-250, 1994. 5 + +[12] Timo Hackel, Jan D Wegner, and Konrad Schindler. Contour detection in unstructured 3d point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1610-1618, 2016. 2 +[13] Chems-Eddine Himeur, Thibault Lejembre, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2 +[14] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2 +[15] Andreas Hubeli and Markus Gross. Multiresolution feature extraction for unstructured meshes. In Proceedings Visualization, 2001. VIS'01., pages 287-294. IEEE, 2001. 2 +[16] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Rotation invariant spherical harmonic representation of 3 d shape descriptors. In Symposium on geometry processing, pages 156-164, 2003. 5 +[17] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6 +[18] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 2 +[19] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2 +[20] Quentin Mérigot, Maks Ovsjanikov, and Leonidas J Guibas. Voronoi-based curvature and feature estimation from point clouds. IEEE Transactions on Visualization and Computer Graphics, 17(6):743-756, 2010. 2 +[21] Carmelo Mineo, Stephen Gareth Pierce, and Rahul Summan. Novel algorithms for 3d surface point cloud boundary detection and edge reconstruction. Journal of Computational Design and Engineering, 6(1):81-91, 2019. 3 +[22] Huan Ni, Xiangguo Lin, Xiaogang Ning, and Jixian Zhang. Edge detection and feature line tracing in 3d-point clouds by analyzing geometric properties of neighborhoods. Remote Sensing, 8(9):710, 2016. 3 +[23] A Cengiz Öztireli, Gael Guennebaud, and Markus Gross. Feature preserving point set surfaces based on non-linear kernel regression. In Computer graphics forum, pages 493-501. Wiley Online Library, 2009. 2 +[24] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[25] Charlie CL Wang. Incremental reconstruction of sharp edges on mesh surfaces. Computer-Aided Design, 38(6):689-702, 2006. 2 + +[26] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3, 6, 7 +[27] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 2 +[28] Shaobo Xia and Ruisheng Wang. A fast edge extraction method for mobile lidar point clouds. IEEE Geoscience and Remote Sensing Letters, 14(8):1288-1292, 2017. 2 +[29] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 1, 5, 6, 7 +[30] Ting Yao, Yehao Li, Yingwei Pan, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3 +[31] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 3 +[32] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 1, 2, 3, 6, 7 +[33] Tong Zhao, Mulin Yu, Pierre Alliez, and Florent Lafarge. Sharp feature consolidation from raw 3d point clouds via displacement learning. Computer Aided Geometric Design, 103:102204, 2023. 3 +[34] Haoran Zhou, Honghua Chen, Yidan Feng, Qiong Wang, Jing Qin, Haoran Xie, Fu Lee Wang, Mingqiang Wei, and Jun Wang. Geometry and learning co-supported normal estimation for unstructured point cloud. In Proceedings of the iee/ivf conference on computer vision and pattern recognition, pages 13238-13247, 2020. 3 +[35] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 8 \ No newline at end of file diff --git a/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/images.zip b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da867b314ced817645a47b49249ebd29d49f8415 --- /dev/null +++ b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a91ce1807f102a33205803193bd871947499e73f42a9d650d829fcd489c585ef +size 779930 diff --git a/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/layout.json b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e43af29c88be6a42ff979acd81b93c58b2bdc7f --- /dev/null +++ b/2025/STAR-Edge_ Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds/layout.json @@ -0,0 +1,8172 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 520, + 140 + ], + "type": "text", + "content": "STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 501, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 501, + 176 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 501, + 176 + ], + "type": "text", + "content": "Zikuan Li*, Honghua Chen*, Yuecheng Wang, Sibo Wu, Mingqiang Wei, Jun Wang†" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 47, + 198, + 547, + 349 + ], + "blocks": [ + { + "bbox": [ + 47, + 198, + 547, + 349 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 547, + 349 + ], + "type": "image", + "image_path": "98f5a62a096ae4ee4d06bc8fdaa57115cf4a9d159e40a69040f6f5331176347b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 351, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 45, + 351, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 45, + 351, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. The characteristics of our method and visual comparison of edge extraction results in thin-walled structures. (a) defines the concept of a thin-walled edge. (b) highlights that the primary challenge in thin-walled edge extraction is the sensitivity to local neighborhood selection. (c) illustrates how a larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring boundary information. (d) shows that a smaller neighborhood may suffer from noise and insufficient sampling on side faces, lacking the contextual information needed for accurate edge point recognition. (e) depicts the spherical projection of the local neighborhood of four points, where points of the same color lie on the same underlying surface. This illustrates our key observation that points co-planar with the neighborhood center tend to align along a great circle arc after spherical projection. (f) provides a qualitative comparison of our STAR-Edge with RFEPS [29], EC-Net [32], and MFLE [4]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 449, + 192, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 449, + 192, + 462 + ], + "spans": [ + { + "bbox": [ + 143, + 449, + 192, + 462 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 475, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 475, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 45, + 475, + 287, + 595 + ], + "type": "text", + "content": "Extracting geometric edges from unstructured point clouds remains a significant challenge, particularly in thin-walled structures that are commonly found in everyday objects. Traditional geometric methods and recent learning-based approaches frequently struggle with these structures, as both rely heavily on sufficient contextual information from local point neighborhoods. However, 3D measurement data of thin-walled structures often lack the accurate, dense, and regular neighborhood sampling required for reliable edge extraction, resulting in degraded performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "type": "text", + "content": "In this work, we introduce STAR-Edge, a novel approach designed for detecting and refining edge points in thin-walled structures. Our method leverages a unique representation—the local spherical curve—to create structure-aware neighborhoods that emphasize co-planar points while reducing interference from close-by, non-co-planar surfaces. This representation is transformed into a rotation-invariant descriptor, which, combined with a lightweight" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 594 + ], + "type": "text", + "content": "multi-layer perceptron, enables robust edge point classification even in the presence of noise and sparse or irregular sampling. Besides, we also use the local spherical curve representation to estimate more precise normals and introduce an optimization function to project initially identified edge points exactly on the true edges. Experiments conducted on the ABC dataset and thin-walled structure-specific datasets demonstrate that STAR-Edge outperforms existing edge detection methods, showcasing better robustness under various challenging conditions. The source code is available at https://github.com/Miraclelezk/STAR-Edge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 620, + 387, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 387, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 387, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Thin-walled structures are essential in our daily lives, seen in items like mechanical parts, furniture, and various manufactured objects. These structures often contain thin-walled edges, defined as boundary lines where surfaces intersect at their side-end faces, characterized by a minimal thickness relative to other dimensions (see Fig. 1 (a)). As geometric" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 208, + 713 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27254" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "edges provide an abstraction of complex 3D shapes, extracting edges from 3D measurements (e.g. unstructured point clouds) in these structures is crucial for downstream tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 312 + ], + "type": "text", + "content": "Extracting geometric edges from point clouds has been a long-standing problem. Traditional techniques generally rely on local geometric properties, such as eigenvalues of the covariance matrix [2, 28], normals [27], curvature [12], or elaborately-designed descriptors [4]. While effective to some extent, these methods depend on first- or second-order derivatives estimated from local neighborhoods, making them particularly sensitive to point disturbances and sampling sparsity or irregularities, especially on thick side-end faces. More recently, some deep neural networks [1, 8, 13, 13, 18, 19, 32] have been proposed to enhance edge accuracy by leveraging deep features [26], neural edge representations [35], or novel point representations (e.g., Scale Space Matrix [18]). However, these methods still struggle with thin-walled structures (as seen in Fig. 1 (f)), as they rely on extracting useful information from valid local neighborhoods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 458 + ], + "type": "text", + "content": "To be more specific, the local neighborhood of points around thin-walled edges is often unreliable. A larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring the boundary information (see Fig. 1 (c)). Conversely, a smaller neighborhood may suffer from noise and insufficient sampling on side-end faces, lacking the contextual information needed for accurate edge point recognition by those above networks (see Fig. 1 (d)). Although multi-scale schemes [26] can partially mitigate this issue, we argue that this approach is sub-optimal, as it requires hard-coding several fixed neighborhood scales." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 628 + ], + "type": "text", + "content": "Ideally, each point's local representation should be structure-aware. This means it should reliably emphasize points on the same underlying surface (e.g., the same plane) while minimizing interference from points on different structures (e.g., another plane). Interestingly, we observe that points co-planar with the neighborhood center tend to align along a great circle arc (dashed circle on the blue plane in Fig. 1 (e)) after spherical projection, effectively reflecting the underlying surface of the thin-walled structure. Based on this observation, naturally, we can construct a local spherical curve to incorporate as many points on the great circle arc as possible, thereby enhancing points on the same underlying surface while reducing the influence of points from other surfaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "To this end, we present STAR-Edge, a novel approach for detecting and refining edge points in thin-walled structures. Our method begins with a new representation, the local spherical curve, to explicitly build structure-aware local neighborhoods. This design enhances information aggregation from nearby co-planar points, even in noisy, sparse, or irregularly point distribution settings. Next, we construct" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "a rotation-invariant descriptor based on the local spherical curve and integrate it with a multi-layer perceptron (MLP) to accurately classify edge and non-edge points. Beyond edge classification, we also use the local spherical curve representation to estimate more precise point normals and introduce an optimization function to project points closer to the true edge. To validate our approach, we conduct a series of experiments on the well-known ABC dataset [17] and specialized datasets for thin-walled structures [6]. Extensive experimental results demonstrate that our approach outperforms existing edge detection methods. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 217, + 545, + 407 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 217, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 545, + 240 + ], + "type": "text", + "content": "- We introduce STAR-Edge, a new method for detecting and refining edge points in thin-walled structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 240, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 545, + 300 + ], + "type": "text", + "content": "- We develop a new representation, the local spherical curve, to construct structure-aware local neighborhoods, enhancing information aggregation from nearby co-planar points, even under challenging conditions like noise, sparsity, or irregular point distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 300, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 360 + ], + "type": "text", + "content": "- Using the local spherical curve representation, we construct a rotation-invariant local descriptor to achieve more accurate edge point classification. Additionally, this representation allows us to accurately calculate point cloud normals, enabling further refinement of edge points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 360, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 407 + ], + "type": "text", + "content": "- We perform extensive experiments on specialized datasets for thin-walled structures and the ABC dataset, showing that our method outperforms existing edge extraction techniques." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 422, + 397, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 397, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 397, + 434 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 442, + 509, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 509, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 509, + 454 + ], + "type": "text", + "content": "2.1. Traditional method for edge extraction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 605 + ], + "type": "text", + "content": "Early edge extraction techniques primarily targeted mesh models [15, 25], with subsequent advancements focusing on direct edge feature extraction from point clouds. Oztireli et al. [23] apply robust kernel regression combined with Moving Least Squares (MLS) to preserve fine details and sharp features in point-based surface representations. Daniels et al. [9] further advance this by introducing a robust method that identifies sharp features in point clouds by generating smooth edge-aligned curves through multi-step refinement and surface fitting. They both rely on iterative refinement processes to progressively extract and enhance sharp edge features in point cloud data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Mérigot et al. [20] propose a robust method using Voronoi-based covariance matrices to robustly extract curvature, sharp features, and normals from point clouds. Bazazian et al. [2] present a fast method for sharp edge extraction from unorganized point clouds using eigenvalues of covariance matrices from k-nearest neighbors. Building on these approaches, other researchers have explored various techniques for identifying and preserving critical features in point clouds. For instance, Huang et al. [14] present an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27255" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "edge-aware resampling method that computes reliable normals away from edges while effectively preserving sharp features. Ni et al. [22] propose AGPN, a method for detecting edges and tracing feature lines in 3D point clouds using neighborhood geometry analysis with RANSAC and hybrid region growing. Mineo et al. [21] present a boundary point detection algorithm with fast Fourier transform (FFT)-based filtering to improve edge reconstruction and reduce noise in tessellated surfaces from point clouds. Chen et al. [4] propose a two-phase algorithm for extracting linetype features from unstructured point clouds, using a statistical metric to detect feature points and an anisotropic contracting scheme to reconstruct feature lines against noise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 301 + ], + "type": "text", + "content": "In summary, most traditional point cloud edge extraction methods rely heavily on neighborhood reliability, which includes dense sampling and minimal noise. However, in thin-walled structures, neighborhood information is often ambiguous, making conventional methods generally unsuitable for such structures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 311, + 277, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 277, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 277, + 323 + ], + "type": "text", + "content": "2.2. Learning-based methods for edge extraction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 330, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 533 + ], + "type": "text", + "content": "With the emergence of learning techniques, many data-driven approaches have been proposed to improve the robustness of edge extraction and the preservation of sharp edges [5, 34]. EC-Net [32] is introduced as an edge-aware point set consolidation network designed to handle sparse, irregular, and noisy point clouds by both minimizing distances to 3D meshes and edges. PIE-NET [26] uses PointNet++ [24] as the backbone for classification and regression to extract edges from 3D point clouds as parametric curves, and has been extensively validated on synthetic data. CPE-Net [7] is designed in a multitask learning paradigm and uses a learnable weighted least squares fitting approach to accurately extract countersink edges under noisy conditions. Generally, all of these methods initially classify edge points using deep learning techniques, followed by edge generation through upsampling, clustering, or fitting smooth curves to capture edge structures in point clouds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "There are also some methods that focus on learning edge features directly from point clouds. Bode et al. [3] proposed Bounded, a neural network that uses local neighborhood statistics to classify non-edge, sharp-edge, and boundary points in 3D point clouds. Yao et al. [30] propose HGNet, a point cloud network that aggregates hierarchical features from points to super-surfaces. Zhao et al. [33] propose a deep learning approach to detect and consolidate sharp feature edges in 3D point clouds using a multi-task network that predicts displacement vectors for edge refinement. APEE-Net [6] is a dual-task neural network designed for extracting precise edges on aircraft panels by identifying nearby points and also using displacement vectors for repositioning. In addition, NerVE [35] defines a neural volumetric edge representation to extract parametric edge" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "curves through graph search, leveraging neural fields rather than traditional point cloud processing. Ye et al. [31] propose a Neural Edge Field method for reconstructing 3D feature curves from multi-view images, optimizing a neural implicit field with a rendering loss to extract 3D edges without 3D supervision. Although these learning-based methods achieve better performance than previous geometric-based methods, they still rely heavily on sufficient contextual information from local point neighborhoods, making them challenging to apply to thin-walled structures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 202, + 362, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 202, + 362, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 362, + 215 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 222, + 547, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 547, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 547, + 379 + ], + "type": "text", + "content": "The proposed method extracts edges from thin-walled structures, as illustrated in Fig. 2, through three main steps: local spherical curve construction, edge point identification, and edge optimization. First, for each point, we project its sparse neighborhood onto a unit sphere, constructing a closed spherical curve to represent the local neighborhood (Sec 3.1). Next, we use spherical harmonics (SH) to encode this spherical curve as a rotation-invariant local descriptor, which is then passed through an MLP layer to classify the target point as an edge or non-edge point (Sec 3.2). Finally, we leverage this descriptor again to estimate the normal vector, allowing us to formulate a projection function that refines the edge point positions (Sec 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 385, + 499, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 499, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 499, + 398 + ], + "type": "text", + "content": "3.1. Local Spherical Curve Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "text", + "content": "Our key observation is that neighboring points co-planar with a target point align along great circle arcs after spherical projection, allowing the overall shape to be approximated as a closed, continuous curve. Thus, given a point cloud " + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "inline_equation", + "content": "P = \\{p_i \\in \\mathbb{R}^3 | i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "text", + "content": ", we first perform spherical normalization on each point " + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "text", + "content": "'s neighborhood point to obtain the local spherical point set " + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 304, + 404, + 547, + 498 + ], + "type": "text", + "content": ", and then fit a closed, continuous curve." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 499, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 547, + 546 + ], + "type": "text", + "content": "Specifically, by taking " + }, + { + "bbox": [ + 304, + 499, + 547, + 546 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 499, + 547, + 546 + ], + "type": "text", + "content": " as the center of the sphere, we normalize the radius of its neighboring points to 1, projecting them onto the unit sphere. The normalization process is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 341, + 555, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 555, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 341, + 555, + 545, + 578 + ], + "type": "interline_equation", + "content": "Q _ {i} = \\left\\{q _ {j} = \\frac {p _ {j} - p _ {i}}{\\| p _ {j} - p _ {i} \\|} \\right| p _ {j} \\in K N N (p _ {i}) \\}, \\tag {1}", + "image_path": "ff986429af9dea0c4c32f397f9b298bf3892a321ec27f230a5b6dbfb4ea7635b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "content": " represents the local spherical point and " + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "content": " is the KNN neighboring points of " + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "content": ". Next, we apply Principal Component Analysis (PCA) to " + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "content": ", projecting it onto a local plane. We then compute the convex hull of the projected points and select the vertices of this convex hull as the key point set " + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 586, + 547, + 658 + ], + "type": "text", + "content": " for curve fitting:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 667, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 667, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 361, + 667, + 545, + 681 + ], + "type": "interline_equation", + "content": "K _ {i} = \\operatorname {C o n v e x H u l l} \\left(\\operatorname {P C A} \\left(Q _ {i}\\right)\\right). \\tag {2}", + "image_path": "02b5e98b11f614827a0e742aae5e0cce83245489729fb70c6fafe13f0138ac95.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Finally, we perform spline fitting on " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "K_{i}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " to obtain a spherical curve, effectively representing the local neighborhood." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27256" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 519, + 259 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 519, + 259 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 519, + 259 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 519, + 259 + ], + "type": "image", + "image_path": "d4fe5876330670f17164a4752f80ed0e2df45c0213380d259f5d17fca3afc0d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 269, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 547, + 304 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. STAR-edge comprises three main steps: (a) constructing a local spherical curve; (b) identifying edge points by integrating rotation-invariant local descriptors with a light-weight MLP layer; and (c) optimizing edge points by adjusting their positions based on the estimated normal vector from the local spherical curve." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 76, + 316, + 168, + 458 + ], + "blocks": [ + { + "bbox": [ + 76, + 316, + 168, + 458 + ], + "lines": [ + { + "bbox": [ + 76, + 316, + 168, + 458 + ], + "spans": [ + { + "bbox": [ + 76, + 316, + 168, + 458 + ], + "type": "image", + "image_path": "4b18ae857a294db13b949e76cd23e925b88b689021907ffd43ebc3287400dd9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 460, + 154, + 467 + ], + "lines": [ + { + "bbox": [ + 86, + 460, + 154, + 467 + ], + "spans": [ + { + "bbox": [ + 86, + 460, + 154, + 467 + ], + "type": "text", + "content": "(a) Thin-walled structure" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 468, + 157, + 475 + ], + "lines": [ + { + "bbox": [ + 83, + 468, + 157, + 475 + ], + "spans": [ + { + "bbox": [ + 83, + 468, + 157, + 475 + ], + "type": "text", + "content": "(thickness: 3, resolution: 1)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 328, + 239, + 393 + ], + "blocks": [ + { + "bbox": [ + 171, + 328, + 239, + 393 + ], + "lines": [ + { + "bbox": [ + 171, + 328, + 239, + 393 + ], + "spans": [ + { + "bbox": [ + 171, + 328, + 239, + 393 + ], + "type": "image", + "image_path": "5dbd7bfb943a5a34c01123d6b01430535559fa21babdd5938ef7dca5a13a10d7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 241, + 327, + 307, + 393 + ], + "blocks": [ + { + "bbox": [ + 170, + 318, + 264, + 327 + ], + "lines": [ + { + "bbox": [ + 170, + 318, + 264, + 327 + ], + "spans": [ + { + "bbox": [ + 170, + 318, + 264, + 327 + ], + "type": "text", + "content": "(b) Local Spherical Points (KNN)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 241, + 327, + 307, + 393 + ], + "lines": [ + { + "bbox": [ + 241, + 327, + 307, + 393 + ], + "spans": [ + { + "bbox": [ + 241, + 327, + 307, + 393 + ], + "type": "image", + "image_path": "9b31f883d628d5a3cbae2b5c30eeb69ece875b9d76e0d67dbea4f4774311e894.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 327, + 377, + 393 + ], + "blocks": [ + { + "bbox": [ + 309, + 327, + 377, + 393 + ], + "lines": [ + { + "bbox": [ + 309, + 327, + 377, + 393 + ], + "spans": [ + { + "bbox": [ + 309, + 327, + 377, + 393 + ], + "type": "image", + "image_path": "e32ae342cdbf243a442db177d3802b49e0d60aebc431cc43d12a1b4b49f61147.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 378, + 326, + 446, + 393 + ], + "blocks": [ + { + "bbox": [ + 378, + 326, + 446, + 393 + ], + "lines": [ + { + "bbox": [ + 378, + 326, + 446, + 393 + ], + "spans": [ + { + "bbox": [ + 378, + 326, + 446, + 393 + ], + "type": "image", + "image_path": "b2afbb0bbfdf7671a64e44a8fb9e371c88b662a37e4b35bd1ae1343d6006b21e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 447, + 326, + 514, + 393 + ], + "blocks": [ + { + "bbox": [ + 447, + 326, + 514, + 393 + ], + "lines": [ + { + "bbox": [ + 447, + 326, + 514, + 393 + ], + "spans": [ + { + "bbox": [ + 447, + 326, + 514, + 393 + ], + "type": "image", + "image_path": "15bba3a42695e6de10214e6cd1afd8a49eff7227796e0ecdfe066e25bf3239a8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 171, + 408, + 239, + 474 + ], + "blocks": [ + { + "bbox": [ + 171, + 396, + 285, + 408 + ], + "lines": [ + { + "bbox": [ + 171, + 396, + 285, + 408 + ], + "spans": [ + { + "bbox": [ + 171, + 396, + 285, + 408 + ], + "type": "text", + "content": "(c)Key Points and Local Spherical Curve" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 408, + 239, + 474 + ], + "lines": [ + { + "bbox": [ + 171, + 408, + 239, + 474 + ], + "spans": [ + { + "bbox": [ + 171, + 408, + 239, + 474 + ], + "type": "image", + "image_path": "1e102264fdf6228d226c129cd0b406fa9b9a7d4efaa82a3a2da8a342b4d23cfe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 488, + 547, + 522 + ], + "lines": [ + { + "bbox": [ + 46, + 488, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 547, + 522 + ], + "type": "text", + "content": "Figure 3. Neighborhood distribution of five surface points on the unit sphere. Local spherical points are shown in grey, sampled key points on the fitted spherical Curve in blue, and local spherical curves in red. We observe that the neighborhoods of edge points resemble two semicircles (e.g. ④), while those of non-edge points maintain a complete great circle (e.g. ①)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 241, + 408, + 307, + 474 + ], + "blocks": [ + { + "bbox": [ + 241, + 408, + 307, + 474 + ], + "lines": [ + { + "bbox": [ + 241, + 408, + 307, + 474 + ], + "spans": [ + { + "bbox": [ + 241, + 408, + 307, + 474 + ], + "type": "image", + "image_path": "d3425c27ebc8e7c8f5ece34a3b130b87cba8d131cc0b9e2fde77ea3cfb6d4e84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 408, + 377, + 474 + ], + "blocks": [ + { + "bbox": [ + 309, + 408, + 377, + 474 + ], + "lines": [ + { + "bbox": [ + 309, + 408, + 377, + 474 + ], + "spans": [ + { + "bbox": [ + 309, + 408, + 377, + 474 + ], + "type": "image", + "image_path": "a652af4d6bf527d8f3a7d763f9fcc800da8b3124aab871c67296b5eb24518d8b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 408, + 446, + 474 + ], + "blocks": [ + { + "bbox": [ + 378, + 408, + 446, + 474 + ], + "lines": [ + { + "bbox": [ + 378, + 408, + 446, + 474 + ], + "spans": [ + { + "bbox": [ + 378, + 408, + 446, + 474 + ], + "type": "image", + "image_path": "ea8484b7b5b54f3e232113a9bc2ac0a54452fe3e55fc313126c2ff9472d3980d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 447, + 408, + 514, + 474 + ], + "blocks": [ + { + "bbox": [ + 447, + 408, + 514, + 474 + ], + "lines": [ + { + "bbox": [ + 447, + 408, + 514, + 474 + ], + "spans": [ + { + "bbox": [ + 447, + 408, + 514, + 474 + ], + "type": "image", + "image_path": "0713654c7b5dc4e68444c7b6e270359d0002009f581d181af16577be97a5db56.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": "Fig. 3 shows the distribution of local spherical points (grey), key points (blue), and the local spherical curve (red) for five target points across a thin-walled structure. Points " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": ", both located on the top surface, their neighborhoods are close to a complete great circle despite interference from points on other surfaces. Point " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": ", positioned on a side-end face with sparse sampling, is influenced by both upper and lower surfaces of the thin-walled structure; however, its local spherical curve still largely retains a great circle form. In contrast, points " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " are near a sharp edge, with " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " closer to the edge than " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": ". The local spherical curves of " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " are roughly formed by two semicircles, presenting a different pattern compared to points " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 46, + 543, + 289, + 712 + ], + "type": "text", + "content": ". These results demonstrate that the proposed local spherical" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 543, + 533, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 543, + 533, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 543, + 533, + 555 + ], + "type": "text", + "content": "curve can robustly and sensitively capture edge features." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 563, + 446, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 446, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 446, + 575 + ], + "type": "text", + "content": "3.2. Edge Point Identification" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 581, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 653 + ], + "type": "text", + "content": "After obtaining the spherical curve for each point, we use SH, an effective time-frequency domain analysis technique on a sphere, to construct rotation-invariant descriptors. This SH-encoded curve serves as a rotation-invariant local descriptor, which is then fed into an MLP to predict whether each point is an edge point." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In detail, we discretize the local spherical curve into a set of scattered points. Each points " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is mapped to spherical coordinates " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "f(\\theta ,\\phi)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " on the unit sphere, where the mapping is defined as " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi) = (\\arccos z,\\arctan y / x)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " represent the zenith and azimuth angles, respec" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27257" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "tively. Let " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "f_{1}, f_{2}, \\ldots, f_{n}" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " be a series of mapping on spherical space. We employ kernel density estimation [10] to construct a smooth, continuous function based on these sample points:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 120, + 287, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 120, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 98, + 120, + 287, + 149 + ], + "type": "interline_equation", + "content": "f (\\theta , \\phi) = \\frac {1}{n h} \\sum_ {i = 1} ^ {n} k \\left(\\frac {\\delta_ {i} (\\theta , \\phi)}{h}\\right), \\tag {3}", + "image_path": "7beda1943ab0717a56705f3d7d6ae06f72d78e9a144510be9ecf864b233da1f2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 154, + 287, + 177 + ], + "type": "text", + "content": " is a kernel function, defined as a normal probability density function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 185, + 287, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 185, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 102, + 185, + 287, + 200 + ], + "type": "interline_equation", + "content": "k (x) = 1 / \\sqrt {2 \\pi} \\exp (- 0. 5 x ^ {2}). \\tag {4}", + "image_path": "9c8e077ce7bb74ba7686e0ff44e5166b40ba07bf82d144386c6d4b1a3b38e0b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": "The term " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\delta_{i}(\\theta, \\phi)" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": " denotes the geodesic distance between " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "f_{i}(\\theta_{i}, \\phi_{i})" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": ", known as the window width or smoothing parameter, modulates the influence of mapping points " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": " to the density estimation at " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": ". It is defined as " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "h = \\pi / B" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 207, + 287, + 277 + ], + "type": "text", + "content": " represents the spherical function bandwidth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": "To compute the SH coefficients for the band-limited function, we apply the discrete SH transform. Following the Shannon-Nyquist sampling theorem [11], we sample the model equiangularly at twice the bandwidth " + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": " in both latitude and longitude, resulting in " + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "2B \\times 2B" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": " sampling points " + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "inline_equation", + "content": "(\\theta_{j}, \\phi_{k}) \\in (0, \\pi) \\times [0, 2\\pi)" + }, + { + "bbox": [ + 47, + 279, + 287, + 351 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 359, + 287, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 359, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 63, + 359, + 287, + 390 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\theta_ {j} = \\pi (2 j + 1) / 4 B & , j = 0, 1, \\dots , 2 B - 1 \\\\ \\phi_ {k} = \\pi k / B & , k = 0, 1, \\dots , 2 B - 1 \\end{array} \\right. \\tag {5}", + "image_path": "637f97f9343325d42030ccf95e27a77c7ec2477609e159ac212095681907bded.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "content": "The SH coefficients of the function " + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "inline_equation", + "content": "f(\\theta, \\phi)" + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "content": " are obtained through the discrete SH transform:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 430, + 287, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 287, + 473 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {l} ^ {m} = \\frac {\\sqrt {2 \\pi}}{2 B} \\sum_ {j = 0} ^ {2 B - 1} \\sum_ {k = 0} ^ {2 B - 1} w _ {j} f \\left(\\theta_ {j}, \\phi_ {k}\\right) e ^ {- i m \\phi_ {k}} P _ {l} ^ {m} (\\cos \\theta_ {j}) \\tag {6}", + "image_path": "e6a3168417a30bf3c36ba1cd4b5b3abf371887c52b37dbeebacfa0ac51c6b2f4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": "for degree " + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "0 < l < B" + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": " and order " + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "|m| \\leq l" + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "P_{l}^{m}" + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": " denotes the associated Legendre functions defined on " + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": ". The weights " + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 47, + 475, + 287, + 510 + ], + "type": "text", + "content": " are computed using the method outlined in [11]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": "The SH bases are indexed by two integers, " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "-l \\leq m \\leq l" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "2l + 1" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": " coefficients for each degree " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": " determines the frequency of the basis [16]. Consequently, the " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "2l + 1" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": " coefficients corresponding to the same degree " + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 511, + 287, + 570 + ], + "type": "text", + "content": " are combined into a vector:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 94, + 578, + 287, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 578, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 94, + 578, + 287, + 592 + ], + "type": "interline_equation", + "content": "\\beta_ {l} = \\left\\{\\hat {f} _ {l} ^ {m}: m = - l, \\dots , 0, \\dots , l \\right\\}, \\tag {7}", + "image_path": "7cba1a09531a83ea8b618f9753864e4c639f8ec5474b2c377cb6da0b9b087e72.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": "and the " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": " norm " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\| \\beta_l\\|" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": " is taken as the energy of the function " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": " in degree " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": ". At last, the rotation-invariant local descriptor " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": " of function " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "f(\\theta ,\\phi)" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": " with all " + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 600, + 287, + 635 + ], + "type": "text", + "content": " degrees is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 87, + 644, + 287, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 644, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 87, + 644, + 287, + 658 + ], + "type": "interline_equation", + "content": "D (f, B) = \\{\\| \\beta_ {l} \\|: l = 0, \\dots , B - 1 \\}. \\tag {8}", + "image_path": "8963ca5cb0c168ae75a137ea62dde1ae9c30b20249714f0215d154df7f03d6cc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": "After computing this descriptor, we calculate the average descriptor for ground-truth edge points and measure the feature difference between each point and this average descriptor. By applying a threshold to this difference, we can" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 76, + 399, + 140 + ], + "blocks": [ + { + "bbox": [ + 320, + 76, + 399, + 140 + ], + "lines": [ + { + "bbox": [ + 320, + 76, + 399, + 140 + ], + "spans": [ + { + "bbox": [ + 320, + 76, + 399, + 140 + ], + "type": "image", + "image_path": "e1ab2b6b4a35c32bf4bc3b1c4ea69961798032ad386da1d1a5fe5b3c2cf02bfa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "type": "text", + "content": "Figure 4. Normal estimation results for points " + }, + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "type": "inline_equation", + "content": "③, ④," + }, + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 305, + 157, + 545, + 201 + ], + "type": "text", + "content": " in Fig. 3 (a). In each sub-figure, the red curve is the great circle arc extracted by RANSAC, the blue curve is the non-great circle segment, and the yellow arrow is the estimated normal vector." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 400, + 76, + 463, + 141 + ], + "blocks": [ + { + "bbox": [ + 400, + 76, + 463, + 141 + ], + "lines": [ + { + "bbox": [ + 400, + 76, + 463, + 141 + ], + "spans": [ + { + "bbox": [ + 400, + 76, + 463, + 141 + ], + "type": "image", + "image_path": "cf34b428f812ccc2c919596b917bc0dc84b30ac69aecc9bb09e8c9309676927d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 466, + 76, + 531, + 140 + ], + "blocks": [ + { + "bbox": [ + 466, + 76, + 531, + 140 + ], + "lines": [ + { + "bbox": [ + 466, + 76, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 466, + 76, + 531, + 140 + ], + "type": "image", + "image_path": "d9d3fd4c673dbf0502377d052aa421fc03b96ff9bce4c7eb594b5f2bd8f4cbb0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 223, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 307 + ], + "type": "text", + "content": "determine whether a point is classified as an edge point. However, in our initial testing, we found that the threshold-based method struggles to effectively handle different edge styles. To address this, we use an MLP to adaptively learn from the rotation-invariant local descriptor, enabling accurate edge point identification. Detailed network structure is provided in the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 316, + 444, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 444, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 444, + 327 + ], + "type": "text", + "content": "3.3. Edge Point Optimization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 334, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 429 + ], + "type": "text", + "content": "We observe that the initially identified edge points are roughly scattered around the true edge. While these points are linearly distributed overall, they form a band-like structure cross two intersecting surfaces when viewed locally. To improve accuracy, we propose constraining edge points to surface intersection regions. This approach is motivated by the fact that an edge point at the intersection of two surfaces has a zero projection distance to both surfaces." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 430, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 545, + 464 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 430, + 545, + 464 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 430, + 545, + 464 + ], + "type": "text", + "content": " represent an initially-identified edge point. Inspired by [29] and [6], we introduce the following optimization function:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 473, + 545, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 473, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 321, + 473, + 545, + 501 + ], + "type": "interline_equation", + "content": "\\min _ {z _ {i}} \\sum_ {p _ {j} \\in K N N (p _ {i})} \\left(\\left(z _ {i} - p _ {j}\\right) \\cdot n _ {j}\\right) ^ {2} + \\mu \\| z _ {i} - p _ {i} \\| ^ {2}, \\tag {9}", + "image_path": "d970558d1df6db7b557b2561b7cbfbf0bb69fafcc1d8a9721219937428ff69cb.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": " is the normal vector of " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": " is a balance parameter. The first term encourages " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": " to stay close to the tangent planes of neighboring points, while the second term penalizes the deviation of " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": " from its initial position " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": ". In all our experiments, " + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 510, + 545, + 569 + ], + "type": "text", + "content": " is set to a default value of 0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 570, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 653 + ], + "type": "text", + "content": "Normal estimation. Note that the edge point optimization method described above relies on accurate normal vectors; however, conventional normal estimation methods is unreliable in noisy or sparse neighborhood conditions. To resolve it, we propose a normal estimation algorithm based on the proposed local spherical curves, which can enhance robustness due to its underlying structure awareness." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 4, we observe that points co-planar with " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " (shown on the red curve) tend to align on the same plane as the great circle arc after projection onto a sphere. In contrast, other points (shown on the blue curve) lie on a different plane. Therefore, we use this great circle arc" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27258" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "(red) on the local spherical curve to approximate the local plane and derive the normal vector. Specifically, we employ random sampling consensus (RANSAC) to identify the great circle arc on the local spherical curve to determine the normal plane, and use the rest (blue) to determine the normal orientation. Fig. 4 shows the normal estimation of the neighborhood " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " in Fig. 3 (a)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 166, + 208, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 208, + 179 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 208, + 179 + ], + "type": "text", + "content": "4. Experiments and Discussion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 258 + ], + "type": "text", + "content": "In this section, we evaluate the effectiveness of our method by comparing it to several learning-based approaches [26, 32, 35] and two traditional methods [4, 29]. We then assess its robustness under varying noise intensities and point cloud resolutions and conduct an ablation study to examine the impact of specific design choices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 265, + 155, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 155, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 155, + 277 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 283, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 437 + ], + "type": "text", + "content": "Dataset. Given the characteristics of thin-walled structures (small dimensions in the thickness direction), we selected a dataset [6] specifically focused on thin-walled structures for our experiments. This dataset includes 42 distinct 3D models with curved edges of varying shapes, sizes, and curvatures, and thicknesses. For a more detailed description of this dataset, please refer to the supplemental file. Note that a subset of 32 shapes is used for training, while 10 shapes are reserved for testing. For each shape, we sample point clouds at resolutions of 0.3, 0.5, and 0.8. Additionally, three levels of random Gaussian noise are added to the point clouds at a resolution of 0.5. Each sampled point cloud contains approximately 20,000 to 15 million points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "We also perform experiments on the well-known ABC dataset [17], which contains over one million CAD models with edge annotations. To evaluate the generalizability of our method, we select sharp-edged CAD models from the ABC dataset, following the setup in [35]. This subset comprises 2,364 models, which are randomly divided into a training set (80%) and a test set (20%). For each shape, 50,000 points are sampled." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 628 + ], + "type": "text", + "content": "Implementation. All experiments are conducted on an Intel i7-11700 CPU and an NVIDIA RTX 4090 GPU. For constructing the local rotation-invariant descriptor, we use a KNN query with " + }, + { + "bbox": [ + 46, + 534, + 287, + 628 + ], + "type": "inline_equation", + "content": "k = 26" + }, + { + "bbox": [ + 46, + 534, + 287, + 628 + ], + "type": "text", + "content": " and a spherical harmonics transform bandwidth of 10. The proposed approach is implemented in PyTorch and trained over 40 epochs using the Adam optimizer, with a batch size of 1024 and a learning rate of 0.02." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Metrics. We quantitatively evaluate all methods across two aspects. First, point classification performance is assessed using Recall, Precision, F1-score, and Accuracy. Second, to evaluate the quality of the final edge point extraction, we introduce a new metric: Edge Chamfer Distance (ECD). This metric measures the similarity between the optimized edges and the ground-truth edges. The introduction of ECD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "is necessary because the ground-truth sampled point cloud consists of dense points, while the optimized edges may vary in sparsity depending on the chosen resolution. Standard Chamfer Distance, without adjustment, could disproportionately penalize sparse edges, leading to inaccurate similarity assessments. ECD addresses this limitation by providing a more reliable comparison for edge extraction quality. Let " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " represents the point cloud of the optimized edges, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": " denote the ground-truth edges. The ECD is defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 198, + 545, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 198, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 361, + 198, + 545, + 227 + ], + "type": "interline_equation", + "content": "\\mathrm {E C D} = \\frac {1}{| X |} \\sum_ {p \\in X} \\min _ {q \\in Y} \\| p - q \\| _ {2} ^ {2} \\tag {10}", + "image_path": "a8d21ac9492308b11f72e73754cd897572ad99ecc96968e4552d8a63d1267ee4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 233, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 281 + ], + "type": "text", + "content": "Note that for each shape, we label sampled points within a specified distance threshold from the nearest edge as ground-truth edge points. We set empirically this threshold to 0.3 for the dataset in [6] and 0.025 for the ABC dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 289, + 492, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 492, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 492, + 300 + ], + "type": "text", + "content": "4.2. Evaluation on Thin-walled Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 307, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 426 + ], + "type": "text", + "content": "In this section, we compare our STAR-Edge method with state-of-the-art approaches, including EC-Net [32], PIE-Net [26], NerVE [35], MFLE [4], and RFEPS [29], on the thin-walled dataset. All methods use point cloud data with a resolution of 0.5 as input. To accommodate network size limitations, PIE-Net is provided with 8,096 uniformly sampled points obtained through farthest point sampling, while NerVE converts the point cloud into a voxel grid with a resolution of " + }, + { + "bbox": [ + 304, + 307, + 545, + 426 + ], + "type": "inline_equation", + "content": "64^3" + }, + { + "bbox": [ + 304, + 307, + 545, + 426 + ], + "type": "text", + "content": ". All networks are retrained on the same dataset for a fair comparison." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 665 + ], + "type": "text", + "content": "Fig. 5 provides visual comparisons of all methods, while quantitative results are presented in Tab. 1. Both results demonstrate that our approach outperforms others on thin-walled structures, achieving better accuracy and a more uniform edge point distribution. Specifically, EC-Net, which up-samples point clouds in edge regions, tends to generate redundant points around true edges, leading to higher Edge Chamfer Distance (ECD). PIE-Net and NerVE perform less effectively on this dataset; PIE-Net requires dense sampling to ensure its effectiveness in edge extraction, while NerVE's voxel size exceeds the thin-wall thickness, making it struggle to perceive edges. A potential improvement for these methods is to segment small patches for re-processing, as discussed in the supplemental file. The geometry-based methods, MFLE and RFEPS, also show limitations: MFLE requires frequent parameter adjustments, while RFEPS produces noisier edge results. Contrast, our method leverages the local spherical curve to achieve more precise edge points and displacement vectors, resulting in superior performance for edge detection in thin-walled structures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Running time. The running times of various methods on point clouds containing approximately 320,000 points are as follows: our method required 214.0 seconds, MFLE completed in 1.14 seconds, RFEPS took 19 minutes, ECNet" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27259" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 90, + 485, + 160 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 523, + 82 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 523, + 82 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 523, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of edge point extraction on the thin-walled structure dataset. -' denotes a failed calculation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 90, + 485, + 160 + ], + "lines": [ + { + "bbox": [ + 107, + 90, + 485, + 160 + ], + "spans": [ + { + "bbox": [ + 107, + 90, + 485, + 160 + ], + "type": "table", + "html": "
EC-Net [32]PIE-Net [26]MFLE [4]RFEPS [29]STAR-Edge (Ours)
ECD↓0.293652.12040.48391.43920.0587
Recall↑-0.05030.97550.99970.9952
Precision↑-0.02950.30280.15120.5418
F1↑-0.03650.46210.26250.7013
Accuracy↑-0.93490.94700.95440.9931
", + "image_path": "4d4107d7f1bbbccaa5003098747b9841dbf39f797048b4826f6477ea575cbc2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 171, + 519, + 459 + ], + "blocks": [ + { + "bbox": [ + 75, + 171, + 519, + 459 + ], + "lines": [ + { + "bbox": [ + 75, + 171, + 519, + 459 + ], + "spans": [ + { + "bbox": [ + 75, + 171, + 519, + 459 + ], + "type": "image", + "image_path": "beb9c451fa1c1c6bda91c1370ab74f704a45c162fbc8b4c6a2ab34917050dbd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 468, + 546, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 546, + 491 + ], + "type": "text", + "content": "Figure 5. Visual comparison of state-of-the-art methods on the thin-walled structure dataset. In these close-up views, our method demonstrates superior accuracy in edge point extraction. Edge points are in red color." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 511, + 274, + 650 + ], + "blocks": [ + { + "bbox": [ + 62, + 511, + 274, + 650 + ], + "lines": [ + { + "bbox": [ + 62, + 511, + 274, + 650 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 274, + 650 + ], + "type": "image", + "image_path": "479f791b9d8bca0a43db8d5dcd4b42d1c367a367b0a74d4c52a3a6ac9ec24b09.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 659, + 271, + 671 + ], + "lines": [ + { + "bbox": [ + 63, + 659, + 271, + 671 + ], + "spans": [ + { + "bbox": [ + 63, + 659, + 271, + 671 + ], + "type": "text", + "content": "Figure 6. Visual comparison under different noise levels." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 713 + ], + "type": "text", + "content": "finished in 83 seconds, and PIE-Net was the fastest, con" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 510, + 533, + 564 + ], + "blocks": [ + { + "bbox": [ + 320, + 510, + 533, + 564 + ], + "lines": [ + { + "bbox": [ + 320, + 510, + 533, + 564 + ], + "spans": [ + { + "bbox": [ + 320, + 510, + 533, + 564 + ], + "type": "image", + "image_path": "db7ec828cd9bf8b8c838aef7c19a759d2f8dd3815f0305bb8fdb98b952d67deb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 574, + 528, + 586 + ], + "lines": [ + { + "bbox": [ + 323, + 574, + 528, + 586 + ], + "spans": [ + { + "bbox": [ + 323, + 574, + 528, + 586 + ], + "type": "text", + "content": "Figure 7. Visual comparison under different resolutions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 609, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 545, + 646 + ], + "type": "text", + "content": "suming less than 1.0 second. We attribute PIE-Net's exceptional speed to its down-sampling process. More detailed statistics can be found in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 658, + 411, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 411, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 411, + 671 + ], + "type": "text", + "content": "4.3. Robustness Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "We conducted robustness testing for the proposed STAR-Edge by introducing varying levels of noise and resolution. In our experiments, the thin-walled shapes have a thickness" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27260" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 102, + 276, + 182 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "type": "text", + "content": "Table 2. Quantitative classification results under different noise levels and resolutions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 102, + 276, + 182 + ], + "lines": [ + { + "bbox": [ + 57, + 102, + 276, + 182 + ], + "spans": [ + { + "bbox": [ + 57, + 102, + 276, + 182 + ], + "type": "table", + "html": "
NoiseResolution
0.0010.0050.030.30.50.8
ECD↓0.0590.0590.1990.0480.0590.261
Recall↑0.9950.9950.9940.9600.9950.988
Precision↑0.5420.5410.5230.7840.5420.582
F1↑0.7010.7010.6850.8630.7010.731
Accuracy↑0.9930.9930.9930.9980.9930.993
", + "image_path": "7ccb8774b79cad6208a13b0f8a4be1f8497e84ac848d1dbae506af0af25fcea6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 224, + 272, + 306 + ], + "blocks": [ + { + "bbox": [ + 47, + 194, + 286, + 215 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 286, + 215 + ], + "type": "text", + "content": "Table 3. Quantitative ECD comparisons under different noise and resolution. " + }, + { + "bbox": [ + 47, + 194, + 286, + 215 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 47, + 194, + 286, + 215 + ], + "type": "text", + "content": " indicates that the calculation failed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 224, + 272, + 306 + ], + "lines": [ + { + "bbox": [ + 61, + 224, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 272, + 306 + ], + "type": "table", + "html": "
MethodsNoiseResolution
0.0010.0050.030.30.50.8
EC-Net0.2950.2930.294-0.2940.518
PIE-Net50.28552.33249.15650.79152.12052.222
RFEPS1.4401.4361.4420.3911.43910.341
MFLE0.4840.4530.4090.4060.4840.742
Ours0.0590.0590.1990.0480.0590.260
", + "image_path": "70f43082641d3b3862592c144c6c3551e929bac08fc1d4c9a6584bc5b083b52d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 286, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 286, + 420 + ], + "type": "text", + "content": "of 2-3 (which is very small relative to other shape dimension). At a resolution of 0.3, approximately 9 points can be sampled on the side-end surface, whereas at a resolution of 0.8, it becomes challenging to obtain even 4 uniformly distributed points. Testing at a resolution of 0.8 poses a significant robustness challenge, as the limited number of sampling points reduces the reliability of projections onto spherical curves." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 286, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 286, + 589 + ], + "type": "text", + "content": "Tab. 2 and Tab. 3 present the effects of noise intensity and resolution on the predictions of our STAR-Edge method. We observe that, even under high noise intensity or significantly reduced resolution, our method consistently produces reasonable edge point predictions with low Edge Chamfer Distance (ECD). As shown in Fig. 6, when the noise level reaches 0.003, the detection performance remains largely unaffected. As shown in Fig. 7, demonstrates that even in an extreme case with a thickness of 3 and a resolution of 0.8, the upper and lower edges can still be clearly distinguished. Additionally, in Tab. 3, we compare the performance of STAR-Edge with state-of-the-art methods across inputs with varying noise levels and resolutions. Our method achieves the best numerical performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 140, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 140, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 140, + 611 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "Tab. 4 reports the performance of our method under varying SH bandwidths, which determine the precision and detail of the local feature representation. The results indicate that performance improves with increasing bandwidth; however, higher bandwidth settings also result in increased computational cost. Notably, the performance at a bandwidth of 10 is comparable to that at 15, making a bandwidth of 10 the optimal choice for balancing accuracy and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 319, + 91, + 531, + 138 + ], + "blocks": [ + { + "bbox": [ + 341, + 71, + 510, + 81 + ], + "lines": [ + { + "bbox": [ + 341, + 71, + 510, + 81 + ], + "spans": [ + { + "bbox": [ + 341, + 71, + 510, + 81 + ], + "type": "text", + "content": "Table 4. Ablation study about SH bandwidths." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 91, + 531, + 138 + ], + "lines": [ + { + "bbox": [ + 319, + 91, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 319, + 91, + 531, + 138 + ], + "type": "table", + "html": "
BandwidthRecall↑Precision↑F1↑Accuracy↑ECD↓
50.9880.5230.6850.9910.122
100.9950.5420.7010.9930.059
150.9950.5830.7850.9980.048
", + "image_path": "33f6cfd02e60df2fce9695c393f0f0fbec49aab80fa352009e8d4e3eaba6d919.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 150, + 536, + 323 + ], + "blocks": [ + { + "bbox": [ + 315, + 150, + 536, + 323 + ], + "lines": [ + { + "bbox": [ + 315, + 150, + 536, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 150, + 536, + 323 + ], + "type": "image", + "image_path": "d6682244787573d87be17401da0b15e2d6d1fd8c4dde4fd33ef25d3cc1312eda.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 334, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 305, + 334, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 545, + 367 + ], + "type": "text", + "content": "Figure 8. Visual comparison of edge extraction results on 3D shapes from the ABC dataset. Our method achieves more precise edge extraction, particularly in the sharp corner regions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 391, + 456, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 456, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 456, + 402 + ], + "type": "text", + "content": "4.5. Evaluation on ABC dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 506 + ], + "type": "text", + "content": "We evaluate our method on the ABC dataset following the experimental setup in [35]. Quantitative results are presented in Tab. 5, which demonstrate that STAR-Edge consistently outperforms all baseline methods in terms of edge extraction accuracy. Visual comparisons in Fig. 8 further highlight the superior performance of our method, as it achieves more precise edge extraction, particularly in the sharp corner regions." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 308, + 540, + 542, + 567 + ], + "blocks": [ + { + "bbox": [ + 327, + 521, + 523, + 531 + ], + "lines": [ + { + "bbox": [ + 327, + 521, + 523, + 531 + ], + "spans": [ + { + "bbox": [ + 327, + 521, + 523, + 531 + ], + "type": "text", + "content": "Table 5. Quantitative comparison on the ABC dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 540, + 542, + 567 + ], + "lines": [ + { + "bbox": [ + 308, + 540, + 542, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 540, + 542, + 567 + ], + "type": "table", + "html": "
EC-NETPIE-NETMFLERFEPSNerVEOurs
ECD↓0.00370.00900.00490.00260.00260.0024
", + "image_path": "d999b8f91fd9dd86ff02ef1b1a04fd85989a85f3fbb2c8c194ebacef6e30fd21.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 596, + 378, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 596, + 378, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 378, + 608 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "In this work, we presented STAR-Edge, a novel point cloud edge extraction approach specifically designed for thin-walled structures. The core of our method is the introduction of a new structure-aware neighborhood representation—the local spherical curve. This representation effectively emphasizes co-planar points while minimizing interference from unrelated surfaces, enabling robust and accurate edge detection in complex thin-walled geometries." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 316, + 757 + ], + "type": "text", + "content": "27261" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 127 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No.92367301, No.52425506, No.92267201, No.52275493, No.T2322012)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 137, + 107, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 157, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 157, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 286, + 190 + ], + "type": "text", + "content": "[1] Dena Bazazian and M Eulàlia Parés. Edc-net: Edge detection capsule network for 3d point clouds. Applied Sciences, 11 (4):1833, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 245 + ], + "type": "text", + "content": "[2] Dena Bazazian, Josep R Casas, and Javier Ruiz-Hidalgo. Fast and robust edge extraction in unorganized point clouds. In 2015 international conference on digital image computing: techniques and applications (DICTA), pages 1-8. IEEE, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[3] Lukas Bode, Michael Weinmann, and Reinhard Klein. Bounded: Neural boundary and edge detection in 3d point clouds via local neighborhood statistics. ISPRS Journal of Photogrammetry and Remote Sensing, 205:334-351, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "type": "text", + "content": "[4] Honghua Chen, Yaoran Huang, Qian Xie, Yuanpeng Liu, Yuan Zhang, Mingqiang Wei, and Jun Wang. Multiscale feature line extraction from raw point clouds based on local surface variation and anisotropic contraction. IEEE Transactions on Automation Science and Engineering, 19(2):1003-1016, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 413 + ], + "type": "text", + "content": "[5] Honghua Chen, Zeyong Wei, Xianzhi Li, Yabin Xu, Mingqiang Wei, and Jun Wang. Repcd-net: Feature-aware recurrent point cloud denoising network. International Journal of Computer Vision, 130(3):615-629, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 286, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 286, + 467 + ], + "type": "text", + "content": "[6] Mengqi Chen, Laishui Zhou, Honghua Chen, and Jun Wang. Thin-walled aircraft panel edge extraction from 3d measurement surfaces via feature-aware displacement learning. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 286, + 523 + ], + "type": "text", + "content": "[7] Mengqi Chen, Laishui Zhou, Yongming Zhang, Honghua Chen, and Jun Wang. Aircraft skin countersink primitive extraction from 3-d measurement point clouds via deep clustering and fitting. IEEE Transactions on Instrumentation and Measurement, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 525, + 286, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 286, + 590 + ], + "type": "text", + "content": "[8] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2727-2735, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 591, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 591, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 591, + 286, + 635 + ], + "type": "text", + "content": "[9] Joel II Daniels, Linh K Ha, Tilo Ochotta, and Claudio T Silva. Robust smooth feature extraction from point clouds. In IEEE International Conference on Shape Modeling and Applications 2007 (SMI'07), pages 123-136. IEEE, 2007. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 678 + ], + "type": "text", + "content": "[10] Richard A Davis, Keh-Shin Lii, and Dimitris N Politis. Remarks on some nonparametric estimates of a density function. Selected Works of Murray Rosenblatt, pages 95-100, 2011. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[11] James R Driscoll and Dennis M Healy. Computing fourier transforms and convolutions on the 2-sphere. Advances in applied mathematics, 15(2):202-250, 1994. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[12] Timo Hackel, Jan D Wegner, and Konrad Schindler. Contour detection in unstructured 3d point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1610-1618, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[13] Chems-Eddine Himeur, Thibault Lejembre, Thomas Pellegrini, Mathias Paulin, Loic Barthe, and Nicolas Mellado. Pcednet: A lightweight neural network for fast and interactive edge detection in 3d point clouds. ACM Transactions on Graphics (TOG), 41(1):1-21, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "text", + "content": "[14] Hui Huang, Shihao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM transactions on graphics (TOG), 32(1):1-12, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "text", + "content": "[15] Andreas Hubeli and Markus Gross. Multiresolution feature extraction for unstructured meshes. In Proceedings Visualization, 2001. VIS'01., pages 287-294. IEEE, 2001. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "text", + "content": "[16] Michael Kazhdan, Thomas Funkhouser, and Szymon Rusinkiewicz. Rotation invariant spherical harmonic representation of 3 d shape descriptors. In Symposium on geometry processing, pages 156-164, 2003. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "text", + "content": "[17] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "text", + "content": "[18] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[19] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes. ACM Transactions on Graphics, 41(4), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 502 + ], + "type": "text", + "content": "[20] Quentin Mérigot, Maks Ovsjanikov, and Leonidas J Guibas. Voronoi-based curvature and feature estimation from point clouds. IEEE Transactions on Visualization and Computer Graphics, 17(6):743-756, 2010. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 548 + ], + "type": "text", + "content": "[21] Carmelo Mineo, Stephen Gareth Pierce, and Rahul Summan. Novel algorithms for 3d surface point cloud boundary detection and edge reconstruction. Journal of Computational Design and Engineering, 6(1):81-91, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[22] Huan Ni, Xiangguo Lin, Xiaogang Ning, and Jixian Zhang. Edge detection and feature line tracing in 3d-point clouds by analyzing geometric properties of neighborhoods. Remote Sensing, 8(9):710, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[23] A Cengiz Öztireli, Gael Guennebaud, and Markus Gross. Feature preserving point set surfaces based on non-linear kernel regression. In Computer graphics forum, pages 493-501. Wiley Online Library, 2009. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[24] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[25] Charlie CL Wang. Incremental reconstruction of sharp edges on mesh surfaces. Computer-Aided Design, 38(6):689-702, 2006. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27262" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 607 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[26] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[27] Christopher Weber, Stefanie Hahmann, and Hans Hagen. Sharp feature detection in point clouds. In 2010 shape modeling international conference, pages 175-186. IEEE, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 206 + ], + "type": "text", + "content": "[28] Shaobo Xia and Ruisheng Wang. A fast edge extraction method for mobile lidar point clouds. IEEE Geoscience and Remote Sensing Letters, 14(8):1288-1292, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 261 + ], + "type": "text", + "content": "[29] Rui Xu, Zixiong Wang, Zhiyang Dou, Chen Zong, Shiqing Xin, Mingyan Jiang, Tao Ju, and Changhe Tu. Rfeps: Reconstructing feature-line equipped polygonal surface. ACM Transactions on Graphics (TOG), 41(6):1-15, 2022. 1, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 262, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 288, + 317 + ], + "type": "text", + "content": "[30] Ting Yao, Yehao Li, Yingwei Pan, and Tao Mei. Hgnet: Learning hierarchical geometry from points, edges, and surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21846-21855, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 374 + ], + "type": "text", + "content": "[31] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 288, + 430 + ], + "type": "text", + "content": "[32] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Ec-net: an edge-aware point set consolidation network. In Proceedings of the European conference on computer vision (ECCV), pages 386–402, 2018. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 288, + 474 + ], + "type": "text", + "content": "[33] Tong Zhao, Mulin Yu, Pierre Alliez, and Florent Lafarge. Sharp feature consolidation from raw 3d point clouds via displacement learning. Computer Aided Geometric Design, 103:102204, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 475, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 542 + ], + "type": "text", + "content": "[34] Haoran Zhou, Honghua Chen, Yidan Feng, Qiong Wang, Jing Qin, Haoran Xie, Fu Lee Wang, Mingqiang Wei, and Jun Wang. Geometry and learning co-supported normal estimation for unstructured point cloud. In Proceedings of the iee/ivf conference on computer vision and pattern recognition, pages 13238-13247, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 542, + 288, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 288, + 607 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 288, + 607 + ], + "type": "text", + "content": "[35] Xiangyu Zhu, Dong Du, Weikai Chen, Zhiyou Zhao, Yinyu Nie, and Xiaoguang Han. Nerve: Neural volumetric edges for parametric curve extraction from point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13601-13610, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27263" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_content_list.json b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..46a375cd60cd278e3cabb601de23cc076d6858b3 --- /dev/null +++ b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_content_list.json @@ -0,0 +1,1614 @@ +[ + { + "type": "text", + "text": "STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction", + "text_level": 1, + "bbox": [ + 101, + 128, + 893, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhimin Liao, Ping Wei\\*, Shuaijia Chen, Haoxuan Wang, Ziyang Ren \nNational Key Laboratory of Human-Machine Hybrid Augmented Intelligence \nInstitute of Artificial Intelligence and Robotics, Xi'an Jiaotong University", + "bbox": [ + 191, + 203, + 805, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liaozm@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn", + "bbox": [ + 305, + 258, + 691, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D occupancy and scene flow offer a detailed and dynamic representation of 3D scene. Recognizing the sparsity and complexity of 3D space, previous vision-centric methods have employed implicit learning-based approaches to model spatial and temporal information. However, these approaches struggle to capture local details and diminish the model's spatial discriminative ability. To address these challenges, we propose a novel explicit state-based modeling method designed to leverage the occupied state to renovate the 3D features. Specifically, we propose a sparse occlusion-aware attention mechanism, integrated with a cascade refinement strategy, which accurately renovates 3D features with the guidance of occupied state information. Additionally, we introduce a novel method for modeling long-term dynamic interactions, which reduces computational costs and preserves spatial information. Compared to the previous state-of-the-art methods, our efficient explicit renovation strategy not only delivers superior performance in terms of RayIoU and mAVE for occupancy and scene flow prediction but also markedly reduces GPU memory usage during training, bringing it down to 8.7GB. Our code is available on https://github.com/lzzzzzm/STCOcc", + "bbox": [ + 89, + 340, + 483, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate perception of 3D surrounding scenes is indeed vital for autonomous systems. The goal of occupancy and scene flow prediction [1] is to segment the entire space into 3D voxels and to determine the semantic and flow information of each voxel. This capability is crucial for understanding the environment and making informed decisions, which is well-suited for downstream tasks in autonomous systems, such as mapping and planning [2, 4, 39, 41, 50].", + "bbox": [ + 89, + 741, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to data sparsity and information redundancy in 3D", + "bbox": [ + 109, + 863, + 482, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/359ed33bc69efd941404ee7c664530e7435e59be568b01cb9afb3950a6409cf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 309, + 898, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edb18cd4ea98ea62e7a6b2100573c4aedb8c9c517fd0af55a00d133a4d60acf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 396, + 898, + 478 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ff87b37eb6fe70f86706c6fff5e9787385008d4d9649150ad0d742dd05709d4.jpg", + "image_caption": [ + "(a) Explicit versus Implicit Modeling.", + "Figure 1. (a) Explicit versus Implicit Modeling: We propose a novel explicit state-based modeling approach that explicitly leverages the occupied state to maintain feature sparsity and model spatial details. (b) Comparison with Different Methods: Our approach achieves state-of-the-art performance of RayIoU and mAVE with lower training costs." + ], + "image_footnote": [], + "bbox": [ + 519, + 496, + 712, + 616 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/278b845ee1a895f4fa4b6b1298b2257171dea60870df791dae1317c906e6e924.jpg", + "image_caption": [ + "(b) Comparison with Different Methods." + ], + "image_footnote": [], + "bbox": [ + 714, + 497, + 901, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "space, employing efficient and robust approaches for 3D feature processing is critical. Existing vision-centric methods [8, 9, 14-16, 22, 29, 31, 41, 45, 53] rely on implicit learning-based method for modeling spatial and temporal information, as shown in Fig. 1a. While these methods optimize 3D reconstruction and long-term temporal fusion through loss supervision, they face inherent limitations. Specifically, the exclusive reliance on loss-driven training hinders the model's ability to capture fine-grained spatial details and compromises the spatial discriminative capacity", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1516", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of learned features, ultimately reducing the effectiveness for occupancy prediction tasks.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose an explicit state-based modeling method that leverages the occupied state of 3D space to refine spatial and temporal feature representations, as shown in Fig. 1a. Our key insight stems from the inherent geometric correspondence between the occupied state and the 3D structure. Since the occupied state directly encodes the geometry of 3D space, it can serve as a robust prior to guide feature learning. This geometric alignment ensures that the feature space preserves structural fidelity, thereby simplifying the learning process and enhancing the discriminative power of the model.", + "bbox": [ + 88, + 122, + 482, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce STCOcc, a sparse Spatial-Temporal Cascade renovation framework tailored for occupancy and scene flow prediction. Within this framework, we present a Spatial-Temporal Cascade Decoder designed to renovate the 3D features both spatially and temporally, leveraging the occupied state of the 3D space. Specifically, at each stage, we employ a Self-Recursive Occupancy Predictor (SROP) to progressively refine the occupied state of the 3D space, thereby providing a more precise 3D geometric state for renovating the 3D features. Subsequently, we propose a sparse occlusion-aware attention mechanism to renovate the 3D features. Our attention mechanism differs from prior methods [15, 17, 19], which relied solely on depth information to renovate the 3D features. Instead, we utilize the occupied state in conjunction with bin depth information to accurately model the 3D spatial features. This approach provides details of local regions and makes the features more geometrically accurate. Furthermore, leveraging the accurate occupied state identified by SROP, we employ the Occlusion-Aware Temporal Self-Attention (OA-TSA) to model dynamic information using a recurrent strategy, supplying detailed short-term temporal information.", + "bbox": [ + 91, + 290, + 482, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To efficiently integrate long-term temporal fusion, we propose a novel sparse-based method for temporal fusion modeling. It also avoids redundant information in historical data and preserves spatial information. Specifically, based on the occupied state, we sample non-empty and empty regions into long-term and short-term streams, respectively. Then, we incorporate the occupied state into both streams and employ a parallel strategy to fuse the temporal information within these two streams. This approach not only reduces computational costs but also retains the spatial information within the 3D space. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 623, + 482, + 803 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an explicit state-based modeling approach designed to renovate the 3D features both spatially and temporally.", + "- We propose a sparse, occlusion-aware mechanism that provides more accurate geometric 3D features. Additionally, we propose a novel sparse-based method for mod" + ], + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eling long-term dynamic information. This approach not only reduces computational costs but also ensures spatial consistency.", + "bbox": [ + 524, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method achieves a RayIoU of $41.7\\%$ on Occ3D [41] and a RayIoU of $40.8\\%$ along with a mAVE of 0.44 for occupancy and scene flow prediction on OpenOcc [39], while also reducing the training memory usage to 8.7GB, as shown in Fig. 1b.", + "bbox": [ + 511, + 136, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 227, + 653, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Camera-based 3D Occupancy Prediction.", + "text_level": 1, + "bbox": [ + 511, + 253, + 864, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Occupancy, as proposed by [30, 36], focuses on the continuous representation of 3D scenes. MonoScene [5] leverages monocular images for semantic scene completion, employing a 3D UNet to process voxel features. TPVFormer [9] lifts image features into 3D TPV space and expands them into voxel representations for 3D occupancy prediction. OccFormer [53] proposes a dual-path transformer for encoding the dense 3D volume features.", + "bbox": [ + 511, + 276, + 903, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the inherent sparsity of 3D scenes, recent methods [14, 18, 22, 29, 40, 47] optimize computational efficiency by processing only non-empty voxels using sparse convolutions or attention mechanisms. VoxFormer [14] utilizes a depth-based query proposal network to generate sparse query proposals for 3D-to-2D cross-attention. SGN [29] introduces a dense-sparse-dense framework that dynamically selects sparse seed voxels and employs hybrid guidance to enhance the convergence of semantic diffusion. Symphonize [10] reconstructs the 3D scene using instance queries. SparseOcc [22] proposes a fully sparse framework that focuses exclusively on non-empty regions. However, these sparse methods rely solely on the occupied state of 3D space to select the region of interest which ignores feature semantics or contextual relationships. Several methods [35, 47] leverage occupancy-based loss supervision to refine 3D features, improving the spatial fidelity of feature representations. However, their exclusive reliance on loss-driven optimization restricts their ability to model fine-grained 3D spatial structures.", + "bbox": [ + 511, + 397, + 903, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Camera-based Temporal Modeling.", + "text_level": 1, + "bbox": [ + 511, + 710, + 821, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal modeling is essential for camera-based perception due to the inherent challenge of lacking depth information. When considering the modeling space, the methods can be divided into two main categories: image feature-based [19, 20, 23, 24, 44] and 3D feature-based [12, 13, 15, 34, 39, 43, 48, 52]. Image feature-based temporal modeling methods utilize multi-frame image features to provide dynamic information. For instance, PETR [23, 24] projects 3D points onto multi-frame image features to generate implicit 3D features for modeling temporal information. Sparse4D [19] creates 4D keypoints based on 3D an", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1517", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a3a4a0e4bca34e31d6d98889fc8eaccc4d7f69a5cd3014c84340dfb4bd16db3f.jpg", + "image_caption": [ + "Figure 2. The overall architecture of STCOcc. The STCOcc framework is primarily composed of four integral modules: a feature extractor that captures image features and depth distribution, a 3D coarse encoder that generates multi-resolution coarse voxel features, a multi-stage spatial-temporal cascade decoder that incrementally renovates these coarse voxel features in both spatial and temporal dimensions, and a head module designed to leverage the refined voxel features for the prediction of 3D occupancy and scene flow." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 859, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "chors and projects these points to aggregate features from multi-frame image features. SparseBEV [20] adaptively generates sampling points based on the query features.", + "bbox": [ + 88, + 468, + 482, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, 3D feature-based methods model temporal features in BEV or voxel space. BEVFormer [15] designs a temporal self-attention mechanism to recursively fuse BEV features. OccNet [39] extends this paradigm to voxel-based temporal self-attention to recursively fuse voxel feature. SOLOFusion [34] uses a parallel strategy to model BEV-level long-term information. We propose a novel sparse approach to model long-term 3D features.", + "bbox": [ + 88, + 517, + 482, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 664, + 189, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 693, + 287, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of STCOcc is presented in Fig. 2. Along the timestamp, we take the multi-view images as video sequence. At current frame $t$ , multi-view images are first processed by the Feature Extractor to obtain image features and depth distribution. The 3D Coarse Encoder uses the image features and depth distribution to create multiresolution coarse voxel features via LSS-based transformation [12, 13, 37]. The Spatial-Temporal Cascade Decoder then progressively renovates these voxel features spatially and temporally, stage by stage. Finally, the Head module utilizes the refined voxel feature to predict occupancy and scene flow.", + "bbox": [ + 88, + 719, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extractor. At each time $t$ , the feature extractor initially uses an image backbone (e.g., ResNet [7]) to extract multi-view features $F_{t} = \\{F_{t}^{j}\\in \\mathbb{R}^{C\\times H\\times W}\\}_{j = 1}^{N_{c}}$ where $F_{t}^{j}$ represents the features of the $j$ -th camera view at time $t$ , and $N_{c}$ is the number of cameras. Then, the depth network [12, 13] processes these image features to predict the bin depth distribution $D_{t} = \\{D_{t}^{j}\\in \\mathbb{R}^{D_{bin}\\times H\\times W}\\}_{j = 1}^{N_{c}}$ .", + "bbox": [ + 511, + 468, + 906, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Coarse Encoder. The 3D coarse encoder adheres to the Lift and Splat framework as delineated in the LSS paradigm [12, 13, 37]. In the Lift phase, each pixel within the 2D image feature planes $F_{t}$ is projected into the 3D voxel space guided by the predicted bin depth distribution $D_{t}$ . Subsequently, the Splat phase consolidates the feature values of pixels falling within each voxel through voxel pooling [12, 13, 37], thereby constructing the coarse voxel feature $V_{t}^{c} \\in \\mathbb{R}^{C \\times X \\times Y \\times Z}$ . Subsequently, we engage a lightweight voxel encoder (e.g. ResNet3D-18 [7]) to produce multi-resolution coarse voxel features $V_{t}^{C} = \\{\\hat{V}_{t}^{i} \\in \\mathbb{R}^{C \\times X_{i} \\times Y_{i} \\times Z_{i}} \\mid i = 1,2,\\dots,N\\}$ , where $X_{i} = \\frac{X}{2^{(N - i)}}, Y_{i} = \\frac{Y}{2^{(N - i)}}, Z_{i} = \\frac{Z}{2^{(N - i)}}$ , and $N$ is the number of processing stages.", + "bbox": [ + 511, + 582, + 908, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Spatial-Temporal Cascade Decoder", + "text_level": 1, + "bbox": [ + 511, + 801, + 821, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To explicitly model the spatial and temporal information into features with the occupied state of 3D space, we introduce a spatial-temporal cascade decoder that renovates the 3D coarse voxel features $V_{t}^{\\mathcal{C}}$ through a multi-stage process. As depicted in Fig. 2, the decoder comprises two pri", + "bbox": [ + 511, + 824, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1518", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mary components: Occlusion-Aware (OA) transformer layers, which accurately capture spatial and short-term temporal information, and a sparse-based temporal fusion module that employs a first-in, first-out (FIFO) memory sequence to encode long-term temporal information.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each stage $i$ , we refine voxel feature in BEV space rather than in voxel space. Initially, we fuse the coarse voxel feature $\\hat{V}_t^i$ with the refined voxel feature from the previous stage output $V_{t}^{i - 1}$ , and then project it into the BEV representation to obtain $B_{t,0}^{i}\\in \\mathbb{R}^{C\\times X_{i}\\times Y_{i}}$ as the input of OA transformer. We subsequently apply $L$ layers OATransformer, which is analogous to the approach in [15], to refine $B_{t,0}^{i}$ . This transformer layer includes three specialized modules: the Self-Recursive Occupancy Predictor (SROP), OA Temporal Self-Attention (OA-TSA), and OA Spatial Cross-Attention (OA-SCA). The SROP is designed to provide an accurate occupied state of the 3D space for each stage. The OA-TSA captures short-range temporal dynamics within the BEV space. The OA-SCA explicitly utilizes the occupied state to address the ambiguous projection problem [15, 17], which transfers geometric 2D feature information into the 3D space.", + "bbox": [ + 89, + 169, + 482, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the $OA$ -Transformer processes the features, the refined BEV features $\\tilde{B}_t^i$ are converted back to voxel form $\\tilde{V}_t^i$ using the occupied weight $W_t^i$ . Subsequently, we leverage the occupied state of the 3D space to guide a novel sparse-based approach for modeling long-term temporal information, thereby obtaining the output $V_t^i$ at time $t$ . $V_t^i$ is upsampled by a factor of 2 for the next stage of refinement.", + "bbox": [ + 89, + 429, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Self-Recursive Occupancy Predictor", + "text_level": 1, + "bbox": [ + 89, + 551, + 388, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To provide a more accurate representation of the occupied state of 3D space and to mitigate the one-off selection issues present in previous methods [14, 22, 29]. We draw inspiration from earlier studies [19, 46, 49] and design the self-recursive occupancy predictor. This predictor employs successive transformer layers to progressively refine the occupied state layer by layer. Specifically, at each layer $l$ , it utilizes a simple Multilayer Perceptron (MLP) to recover the height of $B_{t,l}^{i}$ to voxel space and predict the occupancy weights $W_{t,l}^{i} \\in \\mathbb{R}^{X_{i} \\times Y_{i} \\times Z_{i}}$ . The process of the self-recursive occupancy predictor can be described as follows:", + "bbox": [ + 89, + 573, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {t, l} ^ {i} = f ^ {i} \\left(B _ {t, l} ^ {i}\\right) + \\alpha_ {l} ^ {i} W _ {t, l - 1} ^ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 758, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The function $f^{i}(\\cdot)$ corresponds to the occupancy predictor in stage $i$ , which shares the same weights across different layers. The parameter $\\alpha_{l}^{i}$ signifies the effect of layer $l - 1$ , a learnable parameter initialized to 0.5. At each stage, the initial $W_{t,0}^{i}$ is derived by upsampling the occupied weights from the previous stage, except that it is set to zeros in the first stage.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. OA Temporal Self-Attention", + "text_level": 1, + "bbox": [ + 513, + 90, + 756, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The temporal modeling is crucial for representing the dynamic driving scene [15]. Given the historical BEV feature $B_{t-1}^{i}$ (discussed in Sec. 3.2.4), we align it with the current feature $B_{t,l}^{i}$ via the motion of the ego-vehicle. To efficiently model the dynamic information, we propose Occlusion-Aware Temporal Self-Attention (OA-TSA) to focus the temporal modeling on the non-empty space. The OA-TSA is represented by:", + "bbox": [ + 511, + 109, + 903, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT S A _ {O A} \\left(Q _ {x, y}, \\mathcal {B}, \\bar {W}\\right) = \\sum_ {b \\in \\mathcal {B}} \\bar {w} _ {x, y} H \\left(Q _ {x, y}, b, \\bar {w} _ {x, y}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 234, + 903, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{x,y}$ denotes the BEV feature located at $p = (x,y)$ and $\\mathcal{B} = \\{B_{t,l}^{i}, B_{t - 1}^{i}\\}$ . $\\overline{W} \\in \\mathbb{R}^{X_i \\times Y_i}$ is the average of $W_{t,l}^{i}$ along the z-axis, and $H$ signifies the deformable attention mechanism [54]. $\\overline{w}_{x,y}$ represents the occupied weights $\\overline{W}$ at position $p$ . Unlike the vanilla deformable attention [54], the offsets are predicted by the concatenation of $\\overline{W}$ and $\\mathcal{B}$ . By reweighting the TSA, we enable the model to focus more effectively on the dynamic information within the 3D space.", + "bbox": [ + 511, + 272, + 906, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. OA Spatial Cross-Attention", + "text_level": 1, + "bbox": [ + 511, + 401, + 751, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explicitly utilize the occupied state to renovate the 3D features, we propose the Occupancy-Aware Spatial Cross-Attention (OA-SCA), which leverages the occupied state to enhance geometric features.", + "bbox": [ + 511, + 420, + 903, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first revisit the vanilla Spatial Cross-Attention (SCA) [15] as follows. As shown in Fig. 3, it samples 3D points in BEV plane, and projects these points to 2D image feature planes $F_{t}$ to obtain corresponding features. Formally, the SCA can be expressed as:", + "bbox": [ + 511, + 481, + 905, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS C A \\left(Q _ {x, y}, F _ {t}\\right) = \\sum_ {z = 1} ^ {N _ {\\text {r e f}}} H \\left(Q _ {\\mathbf {x}}, P (\\mathbf {x}), F _ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 561, + 903, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{\\mathbf{x}}$ denotes the sampled feature at coordinates $\\mathbf{x} = (x, y, z)$ in the 3D space and $N_{ref}$ is the number of reference points for each $Q_{x, y}$ . To simplify our initial analysis, we consider a scenario with a single camera, $P(\\cdot)$ is the projection matrix that projects points from the 3D space onto the feature plane. The projection process can be mathematically represented as:", + "bbox": [ + 511, + 614, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd \\cdot \\left[ \\begin{array}{l l l} u & v & 1 \\end{array} \\right] ^ {T} = P \\cdot \\left[ \\begin{array}{l l l l} x & y & z & 1 \\end{array} \\right] ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 731, + 903, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d$ denotes the depth of the point $(u,v)$ on the image plane. This 3D to 2D transformation introduces ambiguity, as different 3D points along the same projection ray map to identical 2D coordinates and are assigned the same features $F_{(u,v)}$ , as illustrated in Fig. 3, where even the green points corresponding to empty areas receive the same feature.", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this ambiguity, previous methods [17, 19] propose utilizing depth information to reweight the features of sampled points. However, these methods overlook the", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1519", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bfc4fb20579872dbca9c785f8b9589220eff488234fd93cc0dd8b667c63d7b36.jpg", + "image_caption": [ + "Figure 3. Illustration of OA-SCA. Due to the projection process, sampled points along the same ray in the feature plane are assigned identical features, even when they represent empty voxel space, as depicted by the green points. To address this limitation, our approach integrates depth and occupancy information to assign appropriate weights to the sampled points, thereby enhancing the differentiation of features along each ray." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 475, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "precision of the predicted depth and the state attributes of the sampled features. In contrast, our approach is inspired by volume rendering techniques [28, 32], which allows us to renovate the features of sampled points more effectively. The volume rendering can be represented as:", + "bbox": [ + 89, + 354, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} = \\sum_ {k = 1} ^ {N _ {c}} \\tau_ {k} \\cdot \\sigma_ {k} \\cdot c _ {k}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 443, + 482, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ represents the expected value of light emitted by particles within the volume as a ray samples $N_{c}$ points. $\\sigma_{k}$ denotes the density at each point, $\\tau_{k}$ is the transmittance, and $c_{k}$ is the corresponding color value. This physical function is analogous to the reverse processing of SCA, where $\\mathcal{C}$ corresponds to $F_{(u,v)}$ in the image plane, and $c_{k}$ corresponds to the sampled point features, as illustrated in Fig. 3.", + "bbox": [ + 89, + 494, + 483, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on above observation, our propose OA-SCA is designed to address the ambiguity inherent in vanilla SCA. Furthermore, to maintain the sparsity of the model, we employ probability sampling to select 3D reference points for refinement. The OA-SCA can be formulated as:", + "bbox": [ + 89, + 602, + 483, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS C A _ {O A} \\left(Q _ {x, y}, F _ {t}\\right) = \\sum_ {\\mathbf {x} \\in N _ {s}} \\Omega_ {\\mathbf {x}} H \\left(Q _ {\\mathbf {x}}, P (\\mathbf {x}), F _ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 686, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{s} = \\{\\mathbf{x}\\in \\mathbf{X}\\mid u < w_{\\mathbf{x}}\\}$ and $\\mathbf{X}$ is total 3D reference points for $Q_{x,y}$ . During training, $u$ follows a truncated normal distribution (0.5,1). For stable inference, $u$ is set to 0.5. This sampling approach allows our method to account for the uncertainty region during training. $w_{\\mathbf{x}}$ is the reference point $\\mathbf{x}$ corresponding occupied weight. The reweighting parameter $\\Omega_{\\mathbf{x}} = w_{\\mathbf{x}}\\cdot w_{d}$ , where $w_{d}$ can be calculated as:", + "bbox": [ + 89, + 731, + 483, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {d} = \\exp \\left(- \\frac {\\min \\left(\\left| d _ {r} - \\left(d _ {r} ^ {\\prime} - \\Delta d\\right) \\right| , \\left| d _ {r} - \\left(d _ {r} ^ {\\prime} + \\Delta d\\right) \\right|\\right) ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 851, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ba5ac319bae4cadd0f27f689b7ec15385d7970c02dcf3f293fdf71419b8c9eab.jpg", + "image_caption": [ + "Figure 4. Illustration of Sparse Temporal Fusion. We implement temporal fusion using a parallel strategy in a sparse manner, focusing only on modeling the sampled features." + ], + "image_footnote": [], + "bbox": [ + 526, + 89, + 890, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$d_{r}$ and $d_r^\\prime$ represent the depth attributes of the reference point respectively analogous to $z$ and $d$ in Eq. (4), they denote the distance of the sampled point from the ego-vehicle and the distance of the ray's corresponding object from the ego-vehicle, respectively. It should be noted that $d_{r}$ is derived from the predicted bin depth distribution $D_{t}$ and is transformed into relative depth, while $\\Delta d$ represents the bin interval. The parameter $\\sigma$ serves as an adjustment factor for $d_{r}$ and $d_r^\\prime$ , enabling fine-tuning of the depth matching tolerance. By default, $\\sigma$ is set to 2, offering a balance between strict and lenient matching. This design takes into account both the attributes of bin depth distribution and the state of the sampled point, enabling a more accurate modeling of 3D spatial information.", + "bbox": [ + 511, + 250, + 906, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.4. Sparse Temporal Modeling", + "text_level": 1, + "bbox": [ + 511, + 467, + 750, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To integrate long-term historical information into the feature representation, at each stage $i$ , we maintain a streaming history memory bank $\\{V_{t - k}^{i}\\}_{k = 1}^{T_{i}}$ that adheres to the first-in, first-out rule to dynamically fuse information using a parallel strategy. Inspired by the SlowFast [6] and considering the redundancy in 3D space, we decouple the non-empty and empty regions into long-term and short-term streams, respectively. The long-term stream models the non-empty region to capture long-term dynamic information, while the short-term stream focuses on the empty region, modeling the overall 3D space in a short-term manner.", + "bbox": [ + 511, + 486, + 905, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For clarity, we consider one stream as an example, as shown in Fig. 4. Given the current refined voxel feature $\\tilde{V}_t^i$ and the occupied weights $W_{t}^{i}$ , we first apply a top-k sampling method to extract the seed feature $S_{t}^{i}\\in \\mathbb{R}^{C\\times N_{S}}$ and corresponding occupied state $w_{t}^{i}\\in \\mathbb{R}^{C_{s}\\times N_{S}}$ , where $C_s$ represents the embedding dimensions for occupied weights. Utilizing the position of $S_{t}$ and the ego-pose transformation matrix $T_{t}^{t - j}$ from frame $t$ to frame $t - j$ , we can retrieve the corresponding historical feature $S_{t - j}^{i}$ . We then concatenate all the corresponding historically sampled features with $w_{t}^{i}$ and subsequently apply MLP to fuse the information along the channel dimension:", + "bbox": [ + 511, + 652, + 906, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {i} = M L P \\left(\\left[ S _ {t} ^ {i}, S _ {t - 1} ^ {i}, \\dots , S _ {t - k} ^ {i}, w _ {t} ^ {i} \\right]\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 842, + 903, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we add $S^i$ to the corresponding position in $\\tilde{V}_t^i$ . By default, at each stage, we use $T_i$ frames to model the", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1520", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/68b2fab36a77f1a18169c00064d1812fbc6a43fd9b643d13d832be6d9c3dcdb5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneInput SizeEpochsRayIoU1m(%)↑RayIoU2m(%)↑RayIoU4m(%)↑RayIoU(%)↑Memory(G)↓
BEVFormer* [15]R1011600 × 9002426.132.938.032.424.1
RenderOcc* [33]Swin-B1408 × 5121213.419.625.519.517.5
BEVDet-Occ* [8]R50704 × 2569023.630.035.129.610.2
FB-Occ* [17]R50704 × 2569026.734.139.733.59.6
SparseOcc (16f) [22]R50704 × 2562429.135.840.335.123.7
COTR† [27]R50704 × 2562436.341.745.141.018.3
OPUS-L [42]R50704 × 25610034.742.146.741.212.1
STCOcc (ours)R50704 × 2563636.242.746.441.77.7
STCOcc (ours)R501408 × 5123636.942.846.742.18.9
", + "bbox": [ + 99, + 88, + 898, + 213 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6b3af14d783dad9f5893a972b8d37d0b7b298073b17e53a1e85ef8007892bec5.jpg", + "table_caption": [ + "Table 1. Comparison of RayIoU (%) performance on the Occ3D-nus dataset. * indicates models trained with camera mask, † denotes that official code was utilized to retrain the model." + ], + "table_footnote": [], + "table_body": "
MethodBackboneInput SizeMemory(G)mIoU(%)othersbarrierbicyclebuscarcons. veh.motor.pedes.tfc. conetrailertruckdrv. surf.other flatsidewalkterrainmanmadevegetation
BEVFormer [15]R1011600×90024.126.95.937.817.940.442.47.423.921.821.022.430.755.428.436.028.120.017.7
CTF-Occ [41]R1011600×900-28.58.139.320.638.342.216.924.522.721.123.031.153.333.838.033.220.818.0
TPVFormer [9]R1011600×90028.927.87.238.913.740.845.917.220.018.914.326.734.255.735.537.630.719.416.8
OSP* [38]R1011600×90020.741.210.949.027.750.255.922.931.030.930.335.631.282.142.651.955.144.838.2
SparseOcc (8f) [22]R50704×25614.330.910.639.220.232.943.319.423.823.429.321.429.367.736.344.640.922.021.9
FB-Occ* [16]R50704×2569.639.113.644.727.045.449.125.226.327.927.832.336.880.142.851.255.142.237.5
ViewFormer* [11]R50704×256-41.912.950.127.944.652.922.429.628.029.235.239.484.749.457.459.747.440.6
COTR* [27]R50704×25618.344.513.352.131.946.055.632.632.830.434.137.741.884.546.257.660.751.946.3
STCOcc* (ours)R50704×2567.744.615.352.931.646.455.931.532.632.134.539.542.583.647.856.460.150.844.8
STCOcc* (ours)R501408×5128.945.015.252.332.250.556.531.733.933.433.838.944.983.947.457.160.150.642.7
", + "bbox": [ + 91, + 266, + 906, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of mIoU (%) performance on the Occ3d-nus dataset. * indicates models trained with camera mask.", + "bbox": [ + 129, + 494, + 866, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "long-term stream and $\\frac{T_i}{2}$ frames to model the short-term stream. At each stage $i$ , the fusion output $V_{t}^{i}$ is appended to the memory queue, and a BEV representation $B_{t}^{i}$ is stored for the subsequent frame of OA-TSA. This strategy provides a long-term perspective for recurrent temporal modeling and mitigates the gradient vanishing issue that is commonly encountered in previous recurrent temporal modeling methods [15].", + "bbox": [ + 88, + 534, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Loss", + "text_level": 1, + "bbox": [ + 89, + 671, + 165, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compute the occupancy loss for each stage, adopting the Scene-Class Affinity Loss $(\\mathcal{L}_{scal})$ from MonoScene [5]. This loss is applied to both semantic and geometric predictions to ensure accurate scene understanding. Given the sparsity and class imbalance inherent in 3D scenes, we also utilize a weighted focal loss [16] combined with the Lovasz loss [3]. The loss for each stage $i$ is formulated as follows:", + "bbox": [ + 89, + 696, + 483, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o c c} ^ {i} = \\mathcal {L} _ {s c a l} ^ {g e o} + \\mathcal {L} _ {s c a l} ^ {s e m} + \\mathcal {L} _ {f o c a l} + \\mathcal {L} _ {l o v}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 823, + 482, + 842 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When considering the scene flow, we utilize the L1 loss $\\mathcal{L}_1$ to supervise only the foreground voxel. The overall loss function is formulated as follow:", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {f} \\mathcal {L} _ {1} + \\mathcal {L} _ {\\text {d e p t h}} + \\sum_ {i = 1} ^ {N} w _ {i} \\times \\mathcal {L} _ {\\text {o c c}} ^ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 547, + 906, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{depth}$ is the cross-entropy loss used to supervise the depth network. $w_{i}$ is computed by $\\frac{1}{2^{N - i}}$ .", + "bbox": [ + 511, + 594, + 906, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 638, + 648, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 662, + 705, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To evaluate the performance of our model, we utilize the Occ3D-nus dataset [41] for assessing 3D occupancy prediction and the OpenOcc dataset [39] for evaluating scene flow prediction. Both datasets are derived from the NuScenes dataset [4], encompassing 600 outdoor scenes for training, 150 for validation, and 150 for testing. Since the Occ3D-nus dataset does not provide scene flow labels, we have omitted the flow head when using this dataset. Additionally, since NuScenes does not provide depth labels, we follow previous methods [12, 13] by projecting LIDAR points onto the image plane to serve as depth labels.", + "bbox": [ + 511, + 685, + 906, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. The mean Intersection-over-Union (mIoU) is a prevalent metric for assessing occupancy prediction per", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1521", + "bbox": [ + 483, + 944, + 513, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/da0fad3885cce6c17a8b2f74a5f4b594e1797d077fbbba16c9f03d2ac459f460.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSup.BackboneInput SizeRayIoU(%)↑mAVE↓Mem(G)↓
OccNeRF-C [51]CR1011600×90021.61.53-
OccNeRF-LLR1011600×90031.71.59-
RenderOcc [33]LR1011600×90036.71.63-
Let Occ Flow [25]C+LR1011408×51240.51.45-
OccNet [39]3DR1011600×90039.71.61-
BEVFormer† [15]3DR501600×90028.11.1226.0
FB-Occ† [16]3DR50704×25632.30.8311.1
SparseOcc† [22]3DR50704×25633.40.8715.8
STCOcc (ours)3DR50704×25640.80.448.7
", + "bbox": [ + 101, + 88, + 470, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance in the Occ3D-nus dataset. However, this metric only accounts for the visible area at the current moment and may not fully reflect the model's completion capabilities [22]. Consequently, we also employ Ray-Based IoU (RayIoU) [21] to evaluate occupancy prediction performance. RayIoU is computed at three distance thresholds: 1, 2, and 4 meters. The final ranking metric is obtained by averaging the results across these thresholds. Additionally, we assess the performance of scene flow prediction using the mean absolute velocity error (mAVE) [39] across defined categories (e.g., car, truck). The mAVE is calculated for the set of true positives within a query ray threshold of 2 meters.", + "bbox": [ + 91, + 296, + 483, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the depth network from BEVStereo [12] and configure our model with three processing stages. Within each stage, the number of layers in the OA transformer is set to 2. The amount of historical information preserved for each stage is 16, 8, and 4, respectively. Unless specified otherwise, all models use the AdamW optimizer [26] with a global batch size of 16. Moreover, because our sampling strategy results in variable GPU memory costs, we report the maximum value in our results.", + "bbox": [ + 89, + 510, + 483, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 670, + 230, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main Results on Occ3D-nus. As shown in Tab. 1 and Tab. 2. We compare our method with previous state-of-the-art methods on 3D occupancy task. Our methods achieve the state of the art performance $41.7\\%$ in RayIoU and $44.6\\%$ in mIoU, which is particularly noteworthy given the significantly lower training costs of 7.7GB, as opposed to the high training costs associated with COTR (18.3 GB) and OPUS-L (12.1 GB). Furthermore, since our spatial refinement process is dependent on the image size, we resize the input to $1408 \\times 512$ and achieve improved results in terms of RayIoU and mIoU.", + "bbox": [ + 89, + 691, + 482, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main Results on OpenOcc. As demonstrated in Tab. 3, we conducted experiments on the OpenOcc dataset to assess", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/00b322ec4fe6cb9d2cb2b6a66dfaa72cdedf53510ad17bed0229c1473a7eb8e4.jpg", + "table_caption": [ + "Table 3. Comparison of RayIoU (%) and mAVE performance on the OpenOcc [39] dataset. C and L denote Camera and Lidar supervision. $\\dagger$ denotes that we utilize the official code and add the flow head to produce the results." + ], + "table_footnote": [], + "table_body": "
SpatialTemporalMetric
OA-SCASROPSTFOA-TSARayIoU(%)↑mAVE↓Mem(G)↓
35.11.325.5
✓35.71.275.7
✓✓36.01.205.7
✓✓✓38.00.698.6
✓✓✓✓38.40.638.7
", + "bbox": [ + 517, + 88, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the each component. SROP refers to the Self-Recursive Occupancy Predictor, OA-SCA refers to the Occupancy-Aware Spatial Cross Attention, OA-TSA refers to the Occlusion-Aware Temporal Self-Attention, and STF refers to the Sparse Temporal Fusion.", + "bbox": [ + 511, + 191, + 906, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43b4ddaca20de8272f049a86707764a305aad0628b24fc072217407b26e90bf8.jpg", + "image_caption": [ + "(a) WOA-SCA." + ], + "image_footnote": [], + "bbox": [ + 566, + 277, + 699, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a53f192f702fdeb69e0c23196960ddb4d5ebe1ba194dc0b1118bffb44bd5320f.jpg", + "image_caption": [ + "(b) W/O OA-SCA.", + "Figure 5. Ablation on the OA-SCA module. We visualize the features after refinement with and without the OA-SCA module." + ], + "image_footnote": [], + "bbox": [ + 715, + 277, + 849, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the performance of our model in terms of occupancy and scene flow. Our approach, which employs a smaller backbone (ResNet-50) and a reduced image input size $(704 \\times 256)$ , achieves RayIoU scores of $40.8\\%$ and mAVE of 0.44. These results surpass those of both OccNet [39] (which uses ResNet-101 with an input size of $1600 \\times 900$ ) and Let Occ Flow [25] (which also uses ResNet-101 with an input size of $1408 \\times 512$ ).", + "bbox": [ + 511, + 458, + 906, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 590, + 666, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the impact of various modules, we perform ablation experiments on the OpenOcc dataset [39]. It is important to note that the ablation experiments were conducted on half of the training dataset and then evaluated on the full validation set. Specifically, we used the first 300 sequences to constitute half of the training dataset.", + "bbox": [ + 511, + 613, + 905, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Effectiveness of Each Component. In Tab. 4, we demonstrate the effectiveness of each component in our model. For the baseline, we omit the OA-SCA and make the occupancy predictor independent in each stage, similar to previous one-off selection methods [22, 39]. This baseline achieves a RayIoU of $35.1\\%$ and a mAVE of 1.32 with a memory cost of 5.5GB. Integrating the OA-SCA into the baseline results in a $1.7\\%$ increase in RayIoU and a $3.7\\%$ increase in mAVE, with an additional memory cost of only 0.2GB. Introducing the SROP further enhances the model's performance without incurring any additional memory costs. Utilizing long-term temporal fusion and", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1522", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/463015490c7afd5e064f70ddf746add55f468e35ef01bea11aeb6f02b1e82033.jpg", + "image_caption": [ + "Figure 6. Qualitative results on Occ3d-nus validation set. As depicted in the red circle, our method delivers detailed predictions for objects such as cars and trucks, while also offering clear boundary delineations for structures like buildings and vegetation." + ], + "image_footnote": [], + "bbox": [ + 210, + 87, + 782, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the OA-TSA, we achieve a $6.3\\%$ increase in RayIoU and a $45.5\\%$ increase in mAVE, with a memory cost of only 1.9GB.", + "bbox": [ + 89, + 366, + 482, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effectiveness of Sparse Temporal Modeling. In Tab. 5, we compare several representations in temporal modeling. The BEV modeling approach [34], while saving computation when modeling long-term history, sacrifices spatial information of the 3D space, resulting in poorer performance in occupancy prediction tasks. Our sparse modeling approach outperforms voxel-level modeling approach [16] in both effectiveness and computational cost. Furthermore, in Tab. 6, we compare our proposed Occupancy-Aware Temporal Self Attention (OA-TSA) with the vanilla Temporal Self Attention (TSA) [15]. Our method achieves superior performance in RayIoU and mAVE metrics, attributing this success to the guidance provided by occupied state modeling.", + "bbox": [ + 89, + 429, + 483, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effectiveness of Occupancy-Aware Spatial Cross Attention. In Fig. 5, we compare the features refined by the OA-SCA module with those refined without it. It is evident that the OA-SCA module renovates the features, providing a more accurate geometric representation of the 3D scene, which is crucial for precise spatial modeling. Moreover, the OA-SCA module selectively enhances the foreground objects in the 3D scene, significantly improving the model's discriminability. Furthermore, in Tab. 7, we compare the vanilla Spatial Cross Attention (SCA) [15] and the Depth-Aware Spatial Cross Attention (DA-SCA) proposed by FB-BEV [17] with our OA-SCA. It is observed that, due to inaccurate spatial modeling, neither SCA nor DA-SCA significantly improves performance. In contrast, Our explicit state-based modeling approach leverages the occupied state to accurately capture detailed spatial information.", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81c355a4559ccf370c93ee771102325750288193e133f84e6f4d486c1763c542.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RepresentationFrameRayIoU(%)↑mAVE↓Mem(G)↓
BEV [34]836.80.816.7
Voxel [16]837.20.737.8
Sparse (ours)837.50.717.3
BEV1637.40.777.6
Voxel1638.00.649.8
Sparse (ours)1638.40.608.7
", + "bbox": [ + 517, + 363, + 901, + 457 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/940279f8e9af3307d239c04236d8d743315c9a0f2ce70bbcd7d2cb84f8fd0e4c.jpg", + "table_caption": [ + "Table 5. Ablation on the Sparse Temporal Fusion. We compare the traditional representations of BEV and voxel to our sparse modeling approach across various frame numbers." + ], + "table_footnote": [], + "table_body": "
MethodRayIoU(%)mAVE
W/O TSA [15]37.90.69
TSA38.00.67
OA-TSA38.30.63
", + "bbox": [ + 517, + 511, + 728, + 575 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9e4bd6af88ad39c9ee9246e05f2f2d2ea21e7b364cad60f0019e93e19a129f2d.jpg", + "table_caption": [ + "Table 6. Ablation on the OA-TSA Module." + ], + "table_footnote": [], + "table_body": "
MethodRayIoU(%)
W/O SCA [15]37.6
SCA37.5
DA-SCA [17]37.7
OA-SCA38.3
", + "bbox": [ + 748, + 511, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation on the OA-SCA Module.", + "bbox": [ + 738, + 580, + 905, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualizations", + "text_level": 1, + "bbox": [ + 511, + 628, + 658, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 6, we present the BEV visualizations on the Occ3D-nus validation set. In comparison to implicit learning-based approaches [16, 22], our explicit state-based method produces clearer boundaries for objects such as cars, buildings, and vegetation.", + "bbox": [ + 511, + 650, + 906, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 739, + 640, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose an explicit state-based modeling approach to capture detailed geometric information in 3D space and integrate long-term dynamic information effectively. Our proposed STCOcc framework incorporates occlusion-aware mechanisms to enhance 3D features in both spatial and temporal aspects, thereby achieving better performance in 3D occupancy and flow prediction. The results demonstrate the efficacy of our paradigm, underscoring its strong potential for applications in downstream tasks.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1523", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tesla AI Day. https://www.youtube.com/watch?v=j0z4FweCy4M, 2021.1", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 1", + "[3] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 6", + "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 6", + "[5] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 6", + "[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 5", + "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[8] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 6", + "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2, 6", + "[10] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. arXiv preprint arXiv:2306.15670, 2023. 2", + "[11] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. arXiv preprint arXiv:2405.04299, 2024.6" + ], + "bbox": [ + 93, + 202, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This research was supported by the National Natural Science Foundation of China (No. U23B2060, No.62088102), and the Youth Innovation Team of Shaanxi Universities.", + "[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 3, 6, 7", + "[13] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2, 3, 6", + "[14] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1, 2, 4", + "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 4, 6, 7, 8", + "[16] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 6, 7, 8", + "[17] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 4, 6, 8", + "[18] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR, 2024. 2", + "[19] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2, 4", + "[20] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18580-18590, 2023. 2, 3", + "[21] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 7", + "[22] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 1, 2, 4, 6, 7, 8", + "[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European conference on computer vision, pages 531-548. Springer, 2022. 2", + "[24] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1524", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2", + "[25] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 7", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Learning, Learning, 2017. 7", + "[27] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 6", + "[28] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, page 99-108, 1995. 5", + "[29] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Xiangrui Zhao, Jongwon Ra, Laijian Li, and Yong Liu. Camera-based 3d semantic scene completion with sparse guidance network. arXiv preprint arXiv:2312.05752, 2023. 1, 2, 4", + "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2", + "[31] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5", + "[33] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 6, 7", + "[34] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 3, 8", + "[35] Liang Peng, Junkai Xu, Haoran Cheng, Zheng Yang, Xiaopei Wu, Wei Qian, Wenxiao Wang, Boxi Wu, and Deng Cai. Learning occupancy for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10281-10292, 2024. 2", + "[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2", + "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3", + "[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as a set of points. arXiv preprint arXiv:2407.04049, 2024. 6", + "[39] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2, 3, 6, 7", + "[40] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2", + "[41] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 6", + "[42] Jiabao Wang, Zhaojiang Liu, Qiang Meng, Liujiang Yan, Ke Wang, Jie Yang, Wei Liu, Qibin Hou, and Ming-Ming Cheng. Opus: Occupancy prediction using a sparse set. arXiv preprint arXiv:2409.09350, 2024. 6", + "[43] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1002-1011, 2023. 2", + "[44] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2", + "[45] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1", + "[46] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv preprint arXiv:2307.13510, 2023. 4", + "[47] Junkai Xu, Liang Peng, Haoran Cheng, Linxuan Xia, Qi Zhou, Dan Deng, Wei Qian, Wenxiao Wang, and Deng Cai. Regulating intermediate 3d features for vision-centric autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6306-6314, 2024. 2", + "[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2", + "[49] Chenhongyi Yang, Tianwei Lin, Lichao Huang, and Elliot J Crowley. Widthformer: Toward efficient transformer-based" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1525", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bev view transformation. arXiv preprint arXiv:2401.03836, 2024. 4", + "[50] Zetong Yang, Li Chen, Yanan Sun, and Hongyang Li. Visual point cloud forecasting enables scalable autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14673-14684, 2024. 1", + "[51] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 7", + "[52] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2", + "[53] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2", + "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + ], + "bbox": [ + 91, + 90, + 482, + 443 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1526", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_model.json b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_model.json new file mode 100644 index 0000000000000000000000000000000000000000..faea49bf3680e49c35706d72b78efbb9ce1afce8 --- /dev/null +++ b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.895, + 0.175 + ], + "angle": 0, + "content": "STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.204, + 0.806, + 0.258 + ], + "angle": 0, + "content": "Zhimin Liao, Ping Wei\\*, Shuaijia Chen, Haoxuan Wang, Ziyang Ren \nNational Key Laboratory of Human-Machine Hybrid Augmented Intelligence \nInstitute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.26, + 0.692, + 0.275 + ], + "angle": 0, + "content": "liaozm@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.688 + ], + "angle": 0, + "content": "3D occupancy and scene flow offer a detailed and dynamic representation of 3D scene. Recognizing the sparsity and complexity of 3D space, previous vision-centric methods have employed implicit learning-based approaches to model spatial and temporal information. However, these approaches struggle to capture local details and diminish the model's spatial discriminative ability. To address these challenges, we propose a novel explicit state-based modeling method designed to leverage the occupied state to renovate the 3D features. Specifically, we propose a sparse occlusion-aware attention mechanism, integrated with a cascade refinement strategy, which accurately renovates 3D features with the guidance of occupied state information. Additionally, we introduce a novel method for modeling long-term dynamic interactions, which reduces computational costs and preserves spatial information. Compared to the previous state-of-the-art methods, our efficient explicit renovation strategy not only delivers superior performance in terms of RayIoU and mAVE for occupancy and scene flow prediction but also markedly reduces GPU memory usage during training, bringing it down to 8.7GB. Our code is available on https://github.com/lzzzzzm/STCOcc" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.717, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Accurate perception of 3D surrounding scenes is indeed vital for autonomous systems. The goal of occupancy and scene flow prediction [1] is to segment the entire space into 3D voxels and to determine the semantic and flow information of each voxel. This capability is crucial for understanding the environment and making informed decisions, which is well-suited for downstream tasks in autonomous systems, such as mapping and planning [2, 4, 39, 41, 50]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.864, + 0.483, + 0.878 + ], + "angle": 0, + "content": "Due to data sparsity and information redundancy in 3D" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.31, + 0.9, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.397, + 0.899, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.484, + 0.81, + 0.496 + ], + "angle": 0, + "content": "(a) Explicit versus Implicit Modeling." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.497, + 0.713, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.62, + 0.817, + 0.632 + ], + "angle": 0, + "content": "(b) Comparison with Different Methods." + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.498, + 0.902, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.728 + ], + "angle": 0, + "content": "Figure 1. (a) Explicit versus Implicit Modeling: We propose a novel explicit state-based modeling approach that explicitly leverages the occupied state to maintain feature sparsity and model spatial details. (b) Comparison with Different Methods: Our approach achieves state-of-the-art performance of RayIoU and mAVE with lower training costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "space, employing efficient and robust approaches for 3D feature processing is critical. Existing vision-centric methods [8, 9, 14-16, 22, 29, 31, 41, 45, 53] rely on implicit learning-based method for modeling spatial and temporal information, as shown in Fig. 1a. While these methods optimize 3D reconstruction and long-term temporal fusion through loss supervision, they face inherent limitations. Specifically, the exclusive reliance on loss-driven training hinders the model's ability to capture fine-grained spatial details and compromises the spatial discriminative capacity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "of learned features, ultimately reducing the effectiveness for occupancy prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.483, + 0.289 + ], + "angle": 0, + "content": "In this paper, we propose an explicit state-based modeling method that leverages the occupied state of 3D space to refine spatial and temporal feature representations, as shown in Fig. 1a. Our key insight stems from the inherent geometric correspondence between the occupied state and the 3D structure. Since the occupied state directly encodes the geometry of 3D space, it can serve as a robust prior to guide feature learning. This geometric alignment ensures that the feature space preserves structural fidelity, thereby simplifying the learning process and enhancing the discriminative power of the model." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.623 + ], + "angle": 0, + "content": "We introduce STCOcc, a sparse Spatial-Temporal Cascade renovation framework tailored for occupancy and scene flow prediction. Within this framework, we present a Spatial-Temporal Cascade Decoder designed to renovate the 3D features both spatially and temporally, leveraging the occupied state of the 3D space. Specifically, at each stage, we employ a Self-Recursive Occupancy Predictor (SROP) to progressively refine the occupied state of the 3D space, thereby providing a more precise 3D geometric state for renovating the 3D features. Subsequently, we propose a sparse occlusion-aware attention mechanism to renovate the 3D features. Our attention mechanism differs from prior methods [15, 17, 19], which relied solely on depth information to renovate the 3D features. Instead, we utilize the occupied state in conjunction with bin depth information to accurately model the 3D spatial features. This approach provides details of local regions and makes the features more geometrically accurate. Furthermore, leveraging the accurate occupied state identified by SROP, we employ the Occlusion-Aware Temporal Self-Attention (OA-TSA) to model dynamic information using a recurrent strategy, supplying detailed short-term temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.804 + ], + "angle": 0, + "content": "To efficiently integrate long-term temporal fusion, we propose a novel sparse-based method for temporal fusion modeling. It also avoids redundant information in historical data and preserves spatial information. Specifically, based on the occupied state, we sample non-empty and empty regions into long-term and short-term streams, respectively. Then, we incorporate the occupied state into both streams and employ a parallel strategy to fuse the temporal information within these two streams. This approach not only reduces computational costs but also retains the spatial information within the 3D space. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.81, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We introduce an explicit state-based modeling approach designed to renovate the 3D features both spatially and temporally." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- We propose a sparse, occlusion-aware mechanism that provides more accurate geometric 3D features. Additionally, we propose a novel sparse-based method for mod" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "eling long-term dynamic information. This approach not only reduces computational costs but also ensures spatial consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.212 + ], + "angle": 0, + "content": "- Our method achieves a RayIoU of \\(41.7\\%\\) on Occ3D [41] and a RayIoU of \\(40.8\\%\\) along with a mAVE of 0.44 for occupancy and scene flow prediction on OpenOcc [39], while also reducing the training memory usage to 8.7GB, as shown in Fig. 1b." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.228, + 0.655, + 0.244 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.254, + 0.866, + 0.27 + ], + "angle": 0, + "content": "2.1. Camera-based 3D Occupancy Prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.396 + ], + "angle": 0, + "content": "Occupancy, as proposed by [30, 36], focuses on the continuous representation of 3D scenes. MonoScene [5] leverages monocular images for semantic scene completion, employing a 3D UNet to process voxel features. TPVFormer [9] lifts image features into 3D TPV space and expands them into voxel representations for 3D occupancy prediction. OccFormer [53] proposes a dual-path transformer for encoding the dense 3D volume features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.7 + ], + "angle": 0, + "content": "Considering the inherent sparsity of 3D scenes, recent methods [14, 18, 22, 29, 40, 47] optimize computational efficiency by processing only non-empty voxels using sparse convolutions or attention mechanisms. VoxFormer [14] utilizes a depth-based query proposal network to generate sparse query proposals for 3D-to-2D cross-attention. SGN [29] introduces a dense-sparse-dense framework that dynamically selects sparse seed voxels and employs hybrid guidance to enhance the convergence of semantic diffusion. Symphonize [10] reconstructs the 3D scene using instance queries. SparseOcc [22] proposes a fully sparse framework that focuses exclusively on non-empty regions. However, these sparse methods rely solely on the occupied state of 3D space to select the region of interest which ignores feature semantics or contextual relationships. Several methods [35, 47] leverage occupancy-based loss supervision to refine 3D features, improving the spatial fidelity of feature representations. However, their exclusive reliance on loss-driven optimization restricts their ability to model fine-grained 3D spatial structures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.822, + 0.729 + ], + "angle": 0, + "content": "2.2. Camera-based Temporal Modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Temporal modeling is essential for camera-based perception due to the inherent challenge of lacking depth information. When considering the modeling space, the methods can be divided into two main categories: image feature-based [19, 20, 23, 24, 44] and 3D feature-based [12, 13, 15, 34, 39, 43, 48, 52]. Image feature-based temporal modeling methods utilize multi-frame image features to provide dynamic information. For instance, PETR [23, 24] projects 3D points onto multi-frame image features to generate implicit 3D features for modeling temporal information. Sparse4D [19] creates 4D keypoints based on 3D an" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1517" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.861, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.398, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of STCOcc. The STCOcc framework is primarily composed of four integral modules: a feature extractor that captures image features and depth distribution, a 3D coarse encoder that generates multi-resolution coarse voxel features, a multi-stage spatial-temporal cascade decoder that incrementally renovates these coarse voxel features in both spatial and temporal dimensions, and a head module designed to leverage the refined voxel features for the prediction of 3D occupancy and scene flow." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.483, + 0.515 + ], + "angle": 0, + "content": "chors and projects these points to aggregate features from multi-frame image features. SparseBEV [20] adaptively generates sampling points based on the query features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.518, + 0.483, + 0.64 + ], + "angle": 0, + "content": "On the other hand, 3D feature-based methods model temporal features in BEV or voxel space. BEVFormer [15] designs a temporal self-attention mechanism to recursively fuse BEV features. OccNet [39] extends this paradigm to voxel-based temporal self-attention to recursively fuse voxel feature. SOLOFusion [34] uses a parallel strategy to model BEV-level long-term information. We propose a novel sparse approach to model long-term 3D features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.191, + 0.681 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.288, + 0.709 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.9 + ], + "angle": 0, + "content": "An overview of STCOcc is presented in Fig. 2. Along the timestamp, we take the multi-view images as video sequence. At current frame \\( t \\), multi-view images are first processed by the Feature Extractor to obtain image features and depth distribution. The 3D Coarse Encoder uses the image features and depth distribution to create multiresolution coarse voxel features via LSS-based transformation [12, 13, 37]. The Spatial-Temporal Cascade Decoder then progressively renovates these voxel features spatially and temporally, stage by stage. Finally, the Head module utilizes the refined voxel feature to predict occupancy and scene flow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Feature Extractor. At each time \\( t \\), the feature extractor initially uses an image backbone (e.g., ResNet [7]) to extract multi-view features \\( F_{t} = \\{F_{t}^{j}\\in \\mathbb{R}^{C\\times H\\times W}\\}_{j = 1}^{N_{c}} \\) where \\( F_{t}^{j} \\) represents the features of the \\( j \\)-th camera view at time \\( t \\), and \\( N_{c} \\) is the number of cameras. Then, the depth network [12, 13] processes these image features to predict the bin depth distribution \\( D_{t} = \\{D_{t}^{j}\\in \\mathbb{R}^{D_{bin}\\times H\\times W}\\}_{j = 1}^{N_{c}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.797 + ], + "angle": 0, + "content": "3D Coarse Encoder. The 3D coarse encoder adheres to the Lift and Splat framework as delineated in the LSS paradigm [12, 13, 37]. In the Lift phase, each pixel within the 2D image feature planes \\( F_{t} \\) is projected into the 3D voxel space guided by the predicted bin depth distribution \\( D_{t} \\). Subsequently, the Splat phase consolidates the feature values of pixels falling within each voxel through voxel pooling [12, 13, 37], thereby constructing the coarse voxel feature \\( V_{t}^{c} \\in \\mathbb{R}^{C \\times X \\times Y \\times Z} \\). Subsequently, we engage a lightweight voxel encoder (e.g. ResNet3D-18 [7]) to produce multi-resolution coarse voxel features \\( V_{t}^{C} = \\{\\hat{V}_{t}^{i} \\in \\mathbb{R}^{C \\times X_{i} \\times Y_{i} \\times Z_{i}} \\mid i = 1,2,\\dots,N\\} \\), where \\( X_{i} = \\frac{X}{2^{(N - i)}}, Y_{i} = \\frac{Y}{2^{(N - i)}}, Z_{i} = \\frac{Z}{2^{(N - i)}} \\), and \\( N \\) is the number of processing stages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.822, + 0.82 + ], + "angle": 0, + "content": "3.2. Spatial-Temporal Cascade Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.903 + ], + "angle": 0, + "content": "To explicitly model the spatial and temporal information into features with the occupied state of 3D space, we introduce a spatial-temporal cascade decoder that renovates the 3D coarse voxel features \\( V_{t}^{\\mathcal{C}} \\) through a multi-stage process. As depicted in Fig. 2, the decoder comprises two pri" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1518" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "mary components: Occlusion-Aware (OA) transformer layers, which accurately capture spatial and short-term temporal information, and a sparse-based temporal fusion module that employs a first-in, first-out (FIFO) memory sequence to encode long-term temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.426 + ], + "angle": 0, + "content": "At each stage \\(i\\), we refine voxel feature in BEV space rather than in voxel space. Initially, we fuse the coarse voxel feature \\(\\hat{V}_t^i\\) with the refined voxel feature from the previous stage output \\(V_{t}^{i - 1}\\), and then project it into the BEV representation to obtain \\(B_{t,0}^{i}\\in \\mathbb{R}^{C\\times X_{i}\\times Y_{i}}\\) as the input of OA transformer. We subsequently apply \\(L\\) layers OATransformer, which is analogous to the approach in [15], to refine \\(B_{t,0}^{i}\\). This transformer layer includes three specialized modules: the Self-Recursive Occupancy Predictor (SROP), OA Temporal Self-Attention (OA-TSA), and OA Spatial Cross-Attention (OA-SCA). The SROP is designed to provide an accurate occupied state of the 3D space for each stage. The OA-TSA captures short-range temporal dynamics within the BEV space. The OA-SCA explicitly utilizes the occupied state to address the ambiguous projection problem [15, 17], which transfers geometric 2D feature information into the 3D space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.538 + ], + "angle": 0, + "content": "After the \\(OA\\)-Transformer processes the features, the refined BEV features \\(\\tilde{B}_t^i\\) are converted back to voxel form \\(\\tilde{V}_t^i\\) using the occupied weight \\(W_t^i\\). Subsequently, we leverage the occupied state of the 3D space to guide a novel sparse-based approach for modeling long-term temporal information, thereby obtaining the output \\(V_t^i\\) at time \\(t\\). \\(V_t^i\\) is upsampled by a factor of 2 for the next stage of refinement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.39, + 0.567 + ], + "angle": 0, + "content": "3.2.1. Self-Recursive Occupancy Predictor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.743 + ], + "angle": 0, + "content": "To provide a more accurate representation of the occupied state of 3D space and to mitigate the one-off selection issues present in previous methods [14, 22, 29]. We draw inspiration from earlier studies [19, 46, 49] and design the self-recursive occupancy predictor. This predictor employs successive transformer layers to progressively refine the occupied state layer by layer. Specifically, at each layer \\(l\\), it utilizes a simple Multilayer Perceptron (MLP) to recover the height of \\(B_{t,l}^{i}\\) to voxel space and predict the occupancy weights \\(W_{t,l}^{i} \\in \\mathbb{R}^{X_{i} \\times Y_{i} \\times Z_{i}}\\). The process of the self-recursive occupancy predictor can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.759, + 0.483, + 0.779 + ], + "angle": 0, + "content": "\\[\nW _ {t, l} ^ {i} = f ^ {i} \\left(B _ {t, l} ^ {i}\\right) + \\alpha_ {l} ^ {i} W _ {t, l - 1} ^ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The function \\( f^{i}(\\cdot) \\) corresponds to the occupancy predictor in stage \\( i \\), which shares the same weights across different layers. The parameter \\( \\alpha_{l}^{i} \\) signifies the effect of layer \\( l - 1 \\), a learnable parameter initialized to 0.5. At each stage, the initial \\( W_{t,0}^{i} \\) is derived by upsampling the occupied weights from the previous stage, except that it is set to zeros in the first stage." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.757, + 0.107 + ], + "angle": 0, + "content": "3.2.2. OA Temporal Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.905, + 0.226 + ], + "angle": 0, + "content": "The temporal modeling is crucial for representing the dynamic driving scene [15]. Given the historical BEV feature \\( B_{t-1}^{i} \\) (discussed in Sec. 3.2.4), we align it with the current feature \\( B_{t,l}^{i} \\) via the motion of the ego-vehicle. To efficiently model the dynamic information, we propose Occlusion-Aware Temporal Self-Attention (OA-TSA) to focus the temporal modeling on the non-empty space. The OA-TSA is represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.235, + 0.905, + 0.264 + ], + "angle": 0, + "content": "\\[\nT S A _ {O A} \\left(Q _ {x, y}, \\mathcal {B}, \\bar {W}\\right) = \\sum_ {b \\in \\mathcal {B}} \\bar {w} _ {x, y} H \\left(Q _ {x, y}, b, \\bar {w} _ {x, y}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.396 + ], + "angle": 0, + "content": "where \\( Q_{x,y} \\) denotes the BEV feature located at \\( p = (x,y) \\) and \\( \\mathcal{B} = \\{B_{t,l}^{i}, B_{t - 1}^{i}\\} \\). \\( \\overline{W} \\in \\mathbb{R}^{X_i \\times Y_i} \\) is the average of \\( W_{t,l}^{i} \\) along the z-axis, and \\( H \\) signifies the deformable attention mechanism [54]. \\( \\overline{w}_{x,y} \\) represents the occupied weights \\( \\overline{W} \\) at position \\( p \\). Unlike the vanilla deformable attention [54], the offsets are predicted by the concatenation of \\( \\overline{W} \\) and \\( \\mathcal{B} \\). By reweighting the TSA, we enable the model to focus more effectively on the dynamic information within the 3D space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.402, + 0.753, + 0.417 + ], + "angle": 0, + "content": "3.2.3. OA Spatial Cross-Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.905, + 0.481 + ], + "angle": 0, + "content": "To explicitly utilize the occupied state to renovate the 3D features, we propose the Occupancy-Aware Spatial Cross-Attention (OA-SCA), which leverages the occupied state to enhance geometric features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.906, + 0.558 + ], + "angle": 0, + "content": "We first revisit the vanilla Spatial Cross-Attention (SCA) [15] as follows. As shown in Fig. 3, it samples 3D points in BEV plane, and projects these points to 2D image feature planes \\( F_{t} \\) to obtain corresponding features. Formally, the SCA can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.563, + 0.905, + 0.606 + ], + "angle": 0, + "content": "\\[\nS C A \\left(Q _ {x, y}, F _ {t}\\right) = \\sum_ {z = 1} ^ {N _ {\\text {r e f}}} H \\left(Q _ {\\mathbf {x}}, P (\\mathbf {x}), F _ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.722 + ], + "angle": 0, + "content": "where \\( Q_{\\mathbf{x}} \\) denotes the sampled feature at coordinates \\( \\mathbf{x} = (x, y, z) \\) in the 3D space and \\( N_{ref} \\) is the number of reference points for each \\( Q_{x, y} \\). To simplify our initial analysis, we consider a scenario with a single camera, \\( P(\\cdot) \\) is the projection matrix that projects points from the 3D space onto the feature plane. The projection process can be mathematically represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.732, + 0.905, + 0.755 + ], + "angle": 0, + "content": "\\[\nd \\cdot \\left[ \\begin{array}{l l l} u & v & 1 \\end{array} \\right] ^ {T} = P \\cdot \\left[ \\begin{array}{l l l l} x & y & z & 1 \\end{array} \\right] ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.856 + ], + "angle": 0, + "content": "where \\(d\\) denotes the depth of the point \\((u,v)\\) on the image plane. This 3D to 2D transformation introduces ambiguity, as different 3D points along the same projection ray map to identical 2D coordinates and are assigned the same features \\(F_{(u,v)}\\), as illustrated in Fig. 3, where even the green points corresponding to empty areas receive the same feature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address this ambiguity, previous methods [17, 19] propose utilizing depth information to reweight the features of sampled points. However, these methods overlook the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.476, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Figure 3. Illustration of OA-SCA. Due to the projection process, sampled points along the same ray in the feature plane are assigned identical features, even when they represent empty voxel space, as depicted by the green points. To address this limitation, our approach integrates depth and occupancy information to assign appropriate weights to the sampled points, thereby enhancing the differentiation of features along each ray." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.432 + ], + "angle": 0, + "content": "precision of the predicted depth and the state attributes of the sampled features. In contrast, our approach is inspired by volume rendering techniques [28, 32], which allows us to renovate the features of sampled points more effectively. The volume rendering can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.444, + 0.483, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} = \\sum_ {k = 1} ^ {N _ {c}} \\tau_ {k} \\cdot \\sigma_ {k} \\cdot c _ {k}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.602 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) represents the expected value of light emitted by particles within the volume as a ray samples \\(N_{c}\\) points. \\(\\sigma_{k}\\) denotes the density at each point, \\(\\tau_{k}\\) is the transmittance, and \\(c_{k}\\) is the corresponding color value. This physical function is analogous to the reverse processing of SCA, where \\(\\mathcal{C}\\) corresponds to \\(F_{(u,v)}\\) in the image plane, and \\(c_{k}\\) corresponds to the sampled point features, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.678 + ], + "angle": 0, + "content": "Based on above observation, our propose OA-SCA is designed to address the ambiguity inherent in vanilla SCA. Furthermore, to maintain the sparsity of the model, we employ probability sampling to select 3D reference points for refinement. The OA-SCA can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.688, + 0.483, + 0.721 + ], + "angle": 0, + "content": "\\[\nS C A _ {O A} \\left(Q _ {x, y}, F _ {t}\\right) = \\sum_ {\\mathbf {x} \\in N _ {s}} \\Omega_ {\\mathbf {x}} H \\left(Q _ {\\mathbf {x}}, P (\\mathbf {x}), F _ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.844 + ], + "angle": 0, + "content": "where \\(N_{s} = \\{\\mathbf{x}\\in \\mathbf{X}\\mid u < w_{\\mathbf{x}}\\}\\) and \\(\\mathbf{X}\\) is total 3D reference points for \\(Q_{x,y}\\). During training, \\(u\\) follows a truncated normal distribution (0.5,1). For stable inference, \\(u\\) is set to 0.5. This sampling approach allows our method to account for the uncertainty region during training. \\(w_{\\mathbf{x}}\\) is the reference point \\(\\mathbf{x}\\) corresponding occupied weight. The reweighting parameter \\(\\Omega_{\\mathbf{x}} = w_{\\mathbf{x}}\\cdot w_{d}\\), where \\(w_{d}\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.852, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\[\nw _ {d} = \\exp \\left(- \\frac {\\min \\left(\\left| d _ {r} - \\left(d _ {r} ^ {\\prime} - \\Delta d\\right) \\right| , \\left| d _ {r} - \\left(d _ {r} ^ {\\prime} + \\Delta d\\right) \\right|\\right) ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {7}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.09, + 0.891, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.197, + 0.905, + 0.239 + ], + "angle": 0, + "content": "Figure 4. Illustration of Sparse Temporal Fusion. We implement temporal fusion using a parallel strategy in a sparse manner, focusing only on modeling the sampled features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.907, + 0.462 + ], + "angle": 0, + "content": "\\(d_{r}\\) and \\(d_r^\\prime\\) represent the depth attributes of the reference point respectively analogous to \\(z\\) and \\(d\\) in Eq. (4), they denote the distance of the sampled point from the ego-vehicle and the distance of the ray's corresponding object from the ego-vehicle, respectively. It should be noted that \\(d_{r}\\) is derived from the predicted bin depth distribution \\(D_{t}\\) and is transformed into relative depth, while \\(\\Delta d\\) represents the bin interval. The parameter \\(\\sigma\\) serves as an adjustment factor for \\(d_{r}\\) and \\(d_r^\\prime\\), enabling fine-tuning of the depth matching tolerance. By default, \\(\\sigma\\) is set to 2, offering a balance between strict and lenient matching. This design takes into account both the attributes of bin depth distribution and the state of the sampled point, enabling a more accurate modeling of 3D spatial information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.75, + 0.484 + ], + "angle": 0, + "content": "3.2.4. Sparse Temporal Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.652 + ], + "angle": 0, + "content": "To integrate long-term historical information into the feature representation, at each stage \\( i \\), we maintain a streaming history memory bank \\( \\{V_{t - k}^{i}\\}_{k = 1}^{T_{i}} \\) that adheres to the first-in, first-out rule to dynamically fuse information using a parallel strategy. Inspired by the SlowFast [6] and considering the redundancy in 3D space, we decouple the non-empty and empty regions into long-term and short-term streams, respectively. The long-term stream models the non-empty region to capture long-term dynamic information, while the short-term stream focuses on the empty region, modeling the overall 3D space in a short-term manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.907, + 0.835 + ], + "angle": 0, + "content": "For clarity, we consider one stream as an example, as shown in Fig. 4. Given the current refined voxel feature \\(\\tilde{V}_t^i\\) and the occupied weights \\(W_{t}^{i}\\), we first apply a top-k sampling method to extract the seed feature \\(S_{t}^{i}\\in \\mathbb{R}^{C\\times N_{S}}\\) and corresponding occupied state \\(w_{t}^{i}\\in \\mathbb{R}^{C_{s}\\times N_{S}}\\), where \\(C_s\\) represents the embedding dimensions for occupied weights. Utilizing the position of \\(S_{t}\\) and the ego-pose transformation matrix \\(T_{t}^{t - j}\\) from frame \\(t\\) to frame \\(t - j\\), we can retrieve the corresponding historical feature \\(S_{t - j}^{i}\\). We then concatenate all the corresponding historically sampled features with \\(w_{t}^{i}\\) and subsequently apply MLP to fuse the information along the channel dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.843, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\nS ^ {i} = M L P \\left(\\left[ S _ {t} ^ {i}, S _ {t - 1} ^ {i}, \\dots , S _ {t - k} ^ {i}, w _ {t} ^ {i} \\right]\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Finally, we add \\( S^i \\) to the corresponding position in \\( \\tilde{V}_t^i \\). By default, at each stage, we use \\( T_i \\) frames to model the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1520" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.899, + 0.214 + ], + "angle": 0, + "content": "
MethodBackboneInput SizeEpochsRayIoU1m(%)↑RayIoU2m(%)↑RayIoU4m(%)↑RayIoU(%)↑Memory(G)↓
BEVFormer* [15]R1011600 × 9002426.132.938.032.424.1
RenderOcc* [33]Swin-B1408 × 5121213.419.625.519.517.5
BEVDet-Occ* [8]R50704 × 2569023.630.035.129.610.2
FB-Occ* [17]R50704 × 2569026.734.139.733.59.6
SparseOcc (16f) [22]R50704 × 2562429.135.840.335.123.7
COTR† [27]R50704 × 2562436.341.745.141.018.3
OPUS-L [42]R50704 × 25610034.742.146.741.212.1
STCOcc (ours)R50704 × 2563636.242.746.441.77.7
STCOcc (ours)R501408 × 5123636.942.846.742.18.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.224, + 0.908, + 0.253 + ], + "angle": 0, + "content": "Table 1. Comparison of RayIoU (%) performance on the Occ3D-nus dataset. * indicates models trained with camera mask, † denotes that official code was utilized to retrain the model." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.267, + 0.907, + 0.484 + ], + "angle": 0, + "content": "
MethodBackboneInput SizeMemory(G)mIoU(%)othersbarrierbicyclebuscarcons. veh.motor.pedes.tfc. conetrailertruckdrv. surf.other flatsidewalkterrainmanmadevegetation
BEVFormer [15]R1011600×90024.126.95.937.817.940.442.47.423.921.821.022.430.755.428.436.028.120.017.7
CTF-Occ [41]R1011600×900-28.58.139.320.638.342.216.924.522.721.123.031.153.333.838.033.220.818.0
TPVFormer [9]R1011600×90028.927.87.238.913.740.845.917.220.018.914.326.734.255.735.537.630.719.416.8
OSP* [38]R1011600×90020.741.210.949.027.750.255.922.931.030.930.335.631.282.142.651.955.144.838.2
SparseOcc (8f) [22]R50704×25614.330.910.639.220.232.943.319.423.823.429.321.429.367.736.344.640.922.021.9
FB-Occ* [16]R50704×2569.639.113.644.727.045.449.125.226.327.927.832.336.880.142.851.255.142.237.5
ViewFormer* [11]R50704×256-41.912.950.127.944.652.922.429.628.029.235.239.484.749.457.459.747.440.6
COTR* [27]R50704×25618.344.513.352.131.946.055.632.632.830.434.137.741.884.546.257.660.751.946.3
STCOcc* (ours)R50704×2567.744.615.352.931.646.455.931.532.632.134.539.542.583.647.856.460.150.844.8
STCOcc* (ours)R501408×5128.945.015.252.332.250.556.531.733.933.433.838.944.983.947.457.160.150.642.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.495, + 0.867, + 0.51 + ], + "angle": 0, + "content": "Table 2. Comparison of mIoU (%) performance on the Occ3d-nus dataset. * indicates models trained with camera mask." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.658 + ], + "angle": 0, + "content": "long-term stream and \\(\\frac{T_i}{2}\\) frames to model the short-term stream. At each stage \\(i\\), the fusion output \\(V_{t}^{i}\\) is appended to the memory queue, and a BEV representation \\(B_{t}^{i}\\) is stored for the subsequent frame of OA-TSA. This strategy provides a long-term perspective for recurrent temporal modeling and mitigates the gradient vanishing issue that is commonly encountered in previous recurrent temporal modeling methods [15]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.166, + 0.687 + ], + "angle": 0, + "content": "3.3. Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.803 + ], + "angle": 0, + "content": "We compute the occupancy loss for each stage, adopting the Scene-Class Affinity Loss \\((\\mathcal{L}_{scal})\\) from MonoScene [5]. This loss is applied to both semantic and geometric predictions to ensure accurate scene understanding. Given the sparsity and class imbalance inherent in 3D scenes, we also utilize a weighted focal loss [16] combined with the Lovasz loss [3]. The loss for each stage \\(i\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.824, + 0.483, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o c c} ^ {i} = \\mathcal {L} _ {s c a l} ^ {g e o} + \\mathcal {L} _ {s c a l} ^ {s e m} + \\mathcal {L} _ {f o c a l} + \\mathcal {L} _ {l o v}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "When considering the scene flow, we utilize the L1 loss \\(\\mathcal{L}_1\\) to supervise only the foreground voxel. The overall loss function is formulated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.548, + 0.907, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {f} \\mathcal {L} _ {1} + \\mathcal {L} _ {\\text {d e p t h}} + \\sum_ {i = 1} ^ {N} w _ {i} \\times \\mathcal {L} _ {\\text {o c c}} ^ {i}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.595, + 0.907, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{depth}\\) is the cross-entropy loss used to supervise the depth network. \\(w_{i}\\) is computed by \\(\\frac{1}{2^{N - i}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.649, + 0.657 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.707, + 0.681 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Dataset. To evaluate the performance of our model, we utilize the Occ3D-nus dataset [41] for assessing 3D occupancy prediction and the OpenOcc dataset [39] for evaluating scene flow prediction. Both datasets are derived from the NuScenes dataset [4], encompassing 600 outdoor scenes for training, 150 for validation, and 150 for testing. Since the Occ3D-nus dataset does not provide scene flow labels, we have omitted the flow head when using this dataset. Additionally, since NuScenes does not provide depth labels, we follow previous methods [12, 13] by projecting LIDAR points onto the image plane to serve as depth labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metric. The mean Intersection-over-Union (mIoU) is a prevalent metric for assessing occupancy prediction per" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "1521" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.472, + 0.21 + ], + "angle": 0, + "content": "
MethodSup.BackboneInput SizeRayIoU(%)↑mAVE↓Mem(G)↓
OccNeRF-C [51]CR1011600×90021.61.53-
OccNeRF-LLR1011600×90031.71.59-
RenderOcc [33]LR1011600×90036.71.63-
Let Occ Flow [25]C+LR1011408×51240.51.45-
OccNet [39]3DR1011600×90039.71.61-
BEVFormer† [15]3DR501600×90028.11.1226.0
FB-Occ† [16]3DR50704×25632.30.8311.1
SparseOcc† [22]3DR50704×25633.40.8715.8
STCOcc (ours)3DR50704×25640.80.448.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Table 3. Comparison of RayIoU (%) and mAVE performance on the OpenOcc [39] dataset. C and L denote Camera and Lidar supervision. \\(\\dagger\\) denotes that we utilize the official code and add the flow head to produce the results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.297, + 0.485, + 0.493 + ], + "angle": 0, + "content": "formance in the Occ3D-nus dataset. However, this metric only accounts for the visible area at the current moment and may not fully reflect the model's completion capabilities [22]. Consequently, we also employ Ray-Based IoU (RayIoU) [21] to evaluate occupancy prediction performance. RayIoU is computed at three distance thresholds: 1, 2, and 4 meters. The final ranking metric is obtained by averaging the results across these thresholds. Additionally, we assess the performance of scene flow prediction using the mean absolute velocity error (mAVE) [39] across defined categories (e.g., car, truck). The mAVE is calculated for the set of true positives within a query ray threshold of 2 meters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Implementation Details. We adopt the depth network from BEVStereo [12] and configure our model with three processing stages. Within each stage, the number of layers in the OA transformer is set to 2. The amount of historical information preserved for each stage is 16, 8, and 4, respectively. Unless specified otherwise, all models use the AdamW optimizer [26] with a global batch size of 16. Moreover, because our sampling strategy results in variable GPU memory costs, we report the maximum value in our results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.231, + 0.687 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.859 + ], + "angle": 0, + "content": "Main Results on Occ3D-nus. As shown in Tab. 1 and Tab. 2. We compare our method with previous state-of-the-art methods on 3D occupancy task. Our methods achieve the state of the art performance \\(41.7\\%\\) in RayIoU and \\(44.6\\%\\) in mIoU, which is particularly noteworthy given the significantly lower training costs of 7.7GB, as opposed to the high training costs associated with COTR (18.3 GB) and OPUS-L (12.1 GB). Furthermore, since our spatial refinement process is dependent on the image size, we resize the input to \\(1408 \\times 512\\) and achieve improved results in terms of RayIoU and mIoU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Main Results on OpenOcc. As demonstrated in Tab. 3, we conducted experiments on the OpenOcc dataset to assess" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.183 + ], + "angle": 0, + "content": "
SpatialTemporalMetric
OA-SCASROPSTFOA-TSARayIoU(%)↑mAVE↓Mem(G)↓
35.11.325.5
✓35.71.275.7
✓✓36.01.205.7
✓✓✓38.00.698.6
✓✓✓✓38.40.638.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Table 4. Ablation study on the each component. SROP refers to the Self-Recursive Occupancy Predictor, OA-SCA refers to the Occupancy-Aware Spatial Cross Attention, OA-TSA refers to the Occlusion-Aware Temporal Self-Attention, and STF refers to the Sparse Temporal Fusion." + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.279, + 0.7, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.385, + 0.678, + 0.396 + ], + "angle": 0, + "content": "(a) WOA-SCA." + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.279, + 0.85, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.385, + 0.834, + 0.396 + ], + "angle": 0, + "content": "(b) W/O OA-SCA." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.403, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 5. Ablation on the OA-SCA module. We visualize the features after refinement with and without the OA-SCA module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.58 + ], + "angle": 0, + "content": "the performance of our model in terms of occupancy and scene flow. Our approach, which employs a smaller backbone (ResNet-50) and a reduced image input size \\((704 \\times 256)\\), achieves RayIoU scores of \\(40.8\\%\\) and mAVE of 0.44. These results surpass those of both OccNet [39] (which uses ResNet-101 with an input size of \\(1600 \\times 900\\)) and Let Occ Flow [25] (which also uses ResNet-101 with an input size of \\(1408 \\times 512\\))." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.667, + 0.607 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.705 + ], + "angle": 0, + "content": "To investigate the impact of various modules, we perform ablation experiments on the OpenOcc dataset [39]. It is important to note that the ablation experiments were conducted on half of the training dataset and then evaluated on the full validation set. Specifically, we used the first 300 sequences to constitute half of the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The Effectiveness of Each Component. In Tab. 4, we demonstrate the effectiveness of each component in our model. For the baseline, we omit the OA-SCA and make the occupancy predictor independent in each stage, similar to previous one-off selection methods [22, 39]. This baseline achieves a RayIoU of \\(35.1\\%\\) and a mAVE of 1.32 with a memory cost of 5.5GB. Integrating the OA-SCA into the baseline results in a \\(1.7\\%\\) increase in RayIoU and a \\(3.7\\%\\) increase in mAVE, with an additional memory cost of only 0.2GB. Introducing the SROP further enhances the model's performance without incurring any additional memory costs. Utilizing long-term temporal fusion and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1522" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.212, + 0.088, + 0.784, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on Occ3d-nus validation set. As depicted in the red circle, our method delivers detailed predictions for objects such as cars and trucks, while also offering clear boundary delineations for structures like buildings and vegetation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.412 + ], + "angle": 0, + "content": "the OA-TSA, we achieve a \\(6.3\\%\\) increase in RayIoU and a \\(45.5\\%\\) increase in mAVE, with a memory cost of only 1.9GB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.642 + ], + "angle": 0, + "content": "The Effectiveness of Sparse Temporal Modeling. In Tab. 5, we compare several representations in temporal modeling. The BEV modeling approach [34], while saving computation when modeling long-term history, sacrifices spatial information of the 3D space, resulting in poorer performance in occupancy prediction tasks. Our sparse modeling approach outperforms voxel-level modeling approach [16] in both effectiveness and computational cost. Furthermore, in Tab. 6, we compare our proposed Occupancy-Aware Temporal Self Attention (OA-TSA) with the vanilla Temporal Self Attention (TSA) [15]. Our method achieves superior performance in RayIoU and mAVE metrics, attributing this success to the guidance provided by occupied state modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The Effectiveness of Occupancy-Aware Spatial Cross Attention. In Fig. 5, we compare the features refined by the OA-SCA module with those refined without it. It is evident that the OA-SCA module renovates the features, providing a more accurate geometric representation of the 3D scene, which is crucial for precise spatial modeling. Moreover, the OA-SCA module selectively enhances the foreground objects in the 3D scene, significantly improving the model's discriminability. Furthermore, in Tab. 7, we compare the vanilla Spatial Cross Attention (SCA) [15] and the Depth-Aware Spatial Cross Attention (DA-SCA) proposed by FB-BEV [17] with our OA-SCA. It is observed that, due to inaccurate spatial modeling, neither SCA nor DA-SCA significantly improves performance. In contrast, Our explicit state-based modeling approach leverages the occupied state to accurately capture detailed spatial information." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.364, + 0.903, + 0.458 + ], + "angle": 0, + "content": "
RepresentationFrameRayIoU(%)↑mAVE↓Mem(G)↓
BEV [34]836.80.816.7
Voxel [16]837.20.737.8
Sparse (ours)837.50.717.3
BEV1637.40.777.6
Voxel1638.00.649.8
Sparse (ours)1638.40.608.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.465, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Table 5. Ablation on the Sparse Temporal Fusion. We compare the traditional representations of BEV and voxel to our sparse modeling approach across various frame numbers." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.512, + 0.73, + 0.576 + ], + "angle": 0, + "content": "
MethodRayIoU(%)mAVE
W/O TSA [15]37.90.69
TSA38.00.67
OA-TSA38.30.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.58, + 0.73, + 0.607 + ], + "angle": 0, + "content": "Table 6. Ablation on the OA-TSA Module." + }, + { + "type": "table", + "bbox": [ + 0.749, + 0.512, + 0.892, + 0.577 + ], + "angle": 0, + "content": "
MethodRayIoU(%)
W/O SCA [15]37.6
SCA37.5
DA-SCA [17]37.7
OA-SCA38.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.739, + 0.581, + 0.906, + 0.608 + ], + "angle": 0, + "content": "Table 7. Ablation on the OA-SCA Module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.629, + 0.659, + 0.643 + ], + "angle": 0, + "content": "4.4. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.727 + ], + "angle": 0, + "content": "In Fig. 6, we present the BEV visualizations on the Occ3D-nus validation set. In comparison to implicit learning-based approaches [16, 22], our explicit state-based method produces clearer boundaries for objects such as cars, buildings, and vegetation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.642, + 0.755 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose an explicit state-based modeling approach to capture detailed geometric information in 3D space and integrate long-term dynamic information effectively. Our proposed STCOcc framework incorporates occlusion-aware mechanisms to enhance 3D features in both spatial and temporal aspects, thereby achieving better performance in 3D occupancy and flow prediction. The results demonstrate the efficacy of our paradigm, underscoring its strong potential for applications in downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1523" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This research was supported by the National Natural Science Foundation of China (No. U23B2060, No.62088102), and the Youth Innovation Team of Shaanxi Universities." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.482, + 0.23 + ], + "angle": 0, + "content": "[1] Tesla AI Day. https://www.youtube.com/watch?v=j0z4FweCy4M, 2021.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[3] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.485, + 0.472 + ], + "angle": 0, + "content": "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.53 + ], + "angle": 0, + "content": "[5] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[8] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[10] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. arXiv preprint arXiv:2306.15670, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. arXiv preprint arXiv:2405.04299, 2024.6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[13] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.908, + 0.317 + ], + "angle": 0, + "content": "[14] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.319, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.404, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[16] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.461, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[17] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[18] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[19] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[20] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18580-18590, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[21] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.759, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[22] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European conference on computer vision, pages 531-548. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1524" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[25] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Learning, Learning, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[27] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[28] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, page 99-108, 1995. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[29] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Xiangrui Zhao, Jongwon Ra, Laijian Li, and Yong Liu. Camera-based 3d semantic scene completion with sparse guidance network. arXiv preprint arXiv:2312.05752, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[31] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[33] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[34] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[35] Liang Peng, Junkai Xu, Haoran Cheng, Zheng Yang, Xiaopei Wu, Wei Qian, Wenxiao Wang, Boxi Wu, and Deng Cai. Learning occupancy for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10281-10292, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as a set of points. arXiv preprint arXiv:2407.04049, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[39] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[40] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[41] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[42] Jiabao Wang, Zhaojiang Liu, Qiang Meng, Liujiang Yan, Ke Wang, Jie Yang, Wei Liu, Qibin Hou, and Ming-Ming Cheng. Opus: Occupancy prediction using a sparse set. arXiv preprint arXiv:2409.09350, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[43] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1002-1011, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[44] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[45] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[46] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv preprint arXiv:2307.13510, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[47] Junkai Xu, Liang Peng, Haoran Cheng, Linxuan Xia, Qi Zhou, Dan Deng, Wei Qian, Wenxiao Wang, and Deng Cai. Regulating intermediate 3d features for vision-centric autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6306-6314, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Chenhongyi Yang, Tianwei Lin, Lichao Huang, and Elliot J Crowley. Widthformer: Toward efficient transformer-based" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1525" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "bev view transformation. arXiv preprint arXiv:2401.03836, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[50] Zetong Yang, Li Chen, Yanan Sun, and Hongyang Li. Visual point cloud forecasting enables scalable autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14673-14684, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[51] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[52] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[53] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1526" + } + ] +] \ No newline at end of file diff --git a/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_origin.pdf b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d484841963bd8fbb655fa32f610fd29a1736724c --- /dev/null +++ b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/8219c456-c593-40fc-a736-867ed4131468_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:217651909d5c1aac7d321a6865fb5a43a778b768a43291480800380f15e27b27 +size 2886527 diff --git a/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/full.md b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4329794e99e61fdd13a7f95cefefce8a0db6494 --- /dev/null +++ b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/full.md @@ -0,0 +1,327 @@ +# STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction + +Zhimin Liao, Ping Wei\*, Shuaijia Chen, Haoxuan Wang, Ziyang Ren +National Key Laboratory of Human-Machine Hybrid Augmented Intelligence +Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University + +liaozm@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn + +# Abstract + +3D occupancy and scene flow offer a detailed and dynamic representation of 3D scene. Recognizing the sparsity and complexity of 3D space, previous vision-centric methods have employed implicit learning-based approaches to model spatial and temporal information. However, these approaches struggle to capture local details and diminish the model's spatial discriminative ability. To address these challenges, we propose a novel explicit state-based modeling method designed to leverage the occupied state to renovate the 3D features. Specifically, we propose a sparse occlusion-aware attention mechanism, integrated with a cascade refinement strategy, which accurately renovates 3D features with the guidance of occupied state information. Additionally, we introduce a novel method for modeling long-term dynamic interactions, which reduces computational costs and preserves spatial information. Compared to the previous state-of-the-art methods, our efficient explicit renovation strategy not only delivers superior performance in terms of RayIoU and mAVE for occupancy and scene flow prediction but also markedly reduces GPU memory usage during training, bringing it down to 8.7GB. Our code is available on https://github.com/lzzzzzm/STCOcc + +# 1. Introduction + +Accurate perception of 3D surrounding scenes is indeed vital for autonomous systems. The goal of occupancy and scene flow prediction [1] is to segment the entire space into 3D voxels and to determine the semantic and flow information of each voxel. This capability is crucial for understanding the environment and making informed decisions, which is well-suited for downstream tasks in autonomous systems, such as mapping and planning [2, 4, 39, 41, 50]. + +Due to data sparsity and information redundancy in 3D + +![](images/359ed33bc69efd941404ee7c664530e7435e59be568b01cb9afb3950a6409cf4.jpg) + +![](images/edb18cd4ea98ea62e7a6b2100573c4aedb8c9c517fd0af55a00d133a4d60acf6.jpg) + +![](images/5ff87b37eb6fe70f86706c6fff5e9787385008d4d9649150ad0d742dd05709d4.jpg) +(a) Explicit versus Implicit Modeling. +Figure 1. (a) Explicit versus Implicit Modeling: We propose a novel explicit state-based modeling approach that explicitly leverages the occupied state to maintain feature sparsity and model spatial details. (b) Comparison with Different Methods: Our approach achieves state-of-the-art performance of RayIoU and mAVE with lower training costs. + +![](images/278b845ee1a895f4fa4b6b1298b2257171dea60870df791dae1317c906e6e924.jpg) +(b) Comparison with Different Methods. + +space, employing efficient and robust approaches for 3D feature processing is critical. Existing vision-centric methods [8, 9, 14-16, 22, 29, 31, 41, 45, 53] rely on implicit learning-based method for modeling spatial and temporal information, as shown in Fig. 1a. While these methods optimize 3D reconstruction and long-term temporal fusion through loss supervision, they face inherent limitations. Specifically, the exclusive reliance on loss-driven training hinders the model's ability to capture fine-grained spatial details and compromises the spatial discriminative capacity + +of learned features, ultimately reducing the effectiveness for occupancy prediction tasks. + +In this paper, we propose an explicit state-based modeling method that leverages the occupied state of 3D space to refine spatial and temporal feature representations, as shown in Fig. 1a. Our key insight stems from the inherent geometric correspondence between the occupied state and the 3D structure. Since the occupied state directly encodes the geometry of 3D space, it can serve as a robust prior to guide feature learning. This geometric alignment ensures that the feature space preserves structural fidelity, thereby simplifying the learning process and enhancing the discriminative power of the model. + +We introduce STCOcc, a sparse Spatial-Temporal Cascade renovation framework tailored for occupancy and scene flow prediction. Within this framework, we present a Spatial-Temporal Cascade Decoder designed to renovate the 3D features both spatially and temporally, leveraging the occupied state of the 3D space. Specifically, at each stage, we employ a Self-Recursive Occupancy Predictor (SROP) to progressively refine the occupied state of the 3D space, thereby providing a more precise 3D geometric state for renovating the 3D features. Subsequently, we propose a sparse occlusion-aware attention mechanism to renovate the 3D features. Our attention mechanism differs from prior methods [15, 17, 19], which relied solely on depth information to renovate the 3D features. Instead, we utilize the occupied state in conjunction with bin depth information to accurately model the 3D spatial features. This approach provides details of local regions and makes the features more geometrically accurate. Furthermore, leveraging the accurate occupied state identified by SROP, we employ the Occlusion-Aware Temporal Self-Attention (OA-TSA) to model dynamic information using a recurrent strategy, supplying detailed short-term temporal information. + +To efficiently integrate long-term temporal fusion, we propose a novel sparse-based method for temporal fusion modeling. It also avoids redundant information in historical data and preserves spatial information. Specifically, based on the occupied state, we sample non-empty and empty regions into long-term and short-term streams, respectively. Then, we incorporate the occupied state into both streams and employ a parallel strategy to fuse the temporal information within these two streams. This approach not only reduces computational costs but also retains the spatial information within the 3D space. Our contributions can be summarized as follows: + +- We introduce an explicit state-based modeling approach designed to renovate the 3D features both spatially and temporally. +- We propose a sparse, occlusion-aware mechanism that provides more accurate geometric 3D features. Additionally, we propose a novel sparse-based method for mod + +eling long-term dynamic information. This approach not only reduces computational costs but also ensures spatial consistency. + +- Our method achieves a RayIoU of $41.7\%$ on Occ3D [41] and a RayIoU of $40.8\%$ along with a mAVE of 0.44 for occupancy and scene flow prediction on OpenOcc [39], while also reducing the training memory usage to 8.7GB, as shown in Fig. 1b. + +# 2. Related Work + +# 2.1. Camera-based 3D Occupancy Prediction. + +Occupancy, as proposed by [30, 36], focuses on the continuous representation of 3D scenes. MonoScene [5] leverages monocular images for semantic scene completion, employing a 3D UNet to process voxel features. TPVFormer [9] lifts image features into 3D TPV space and expands them into voxel representations for 3D occupancy prediction. OccFormer [53] proposes a dual-path transformer for encoding the dense 3D volume features. + +Considering the inherent sparsity of 3D scenes, recent methods [14, 18, 22, 29, 40, 47] optimize computational efficiency by processing only non-empty voxels using sparse convolutions or attention mechanisms. VoxFormer [14] utilizes a depth-based query proposal network to generate sparse query proposals for 3D-to-2D cross-attention. SGN [29] introduces a dense-sparse-dense framework that dynamically selects sparse seed voxels and employs hybrid guidance to enhance the convergence of semantic diffusion. Symphonize [10] reconstructs the 3D scene using instance queries. SparseOcc [22] proposes a fully sparse framework that focuses exclusively on non-empty regions. However, these sparse methods rely solely on the occupied state of 3D space to select the region of interest which ignores feature semantics or contextual relationships. Several methods [35, 47] leverage occupancy-based loss supervision to refine 3D features, improving the spatial fidelity of feature representations. However, their exclusive reliance on loss-driven optimization restricts their ability to model fine-grained 3D spatial structures. + +# 2.2. Camera-based Temporal Modeling. + +Temporal modeling is essential for camera-based perception due to the inherent challenge of lacking depth information. When considering the modeling space, the methods can be divided into two main categories: image feature-based [19, 20, 23, 24, 44] and 3D feature-based [12, 13, 15, 34, 39, 43, 48, 52]. Image feature-based temporal modeling methods utilize multi-frame image features to provide dynamic information. For instance, PETR [23, 24] projects 3D points onto multi-frame image features to generate implicit 3D features for modeling temporal information. Sparse4D [19] creates 4D keypoints based on 3D an + +![](images/a3a4a0e4bca34e31d6d98889fc8eaccc4d7f69a5cd3014c84340dfb4bd16db3f.jpg) +Figure 2. The overall architecture of STCOcc. The STCOcc framework is primarily composed of four integral modules: a feature extractor that captures image features and depth distribution, a 3D coarse encoder that generates multi-resolution coarse voxel features, a multi-stage spatial-temporal cascade decoder that incrementally renovates these coarse voxel features in both spatial and temporal dimensions, and a head module designed to leverage the refined voxel features for the prediction of 3D occupancy and scene flow. + +chors and projects these points to aggregate features from multi-frame image features. SparseBEV [20] adaptively generates sampling points based on the query features. + +On the other hand, 3D feature-based methods model temporal features in BEV or voxel space. BEVFormer [15] designs a temporal self-attention mechanism to recursively fuse BEV features. OccNet [39] extends this paradigm to voxel-based temporal self-attention to recursively fuse voxel feature. SOLOFusion [34] uses a parallel strategy to model BEV-level long-term information. We propose a novel sparse approach to model long-term 3D features. + +# 3. Methods + +# 3.1. Overall Architecture + +An overview of STCOcc is presented in Fig. 2. Along the timestamp, we take the multi-view images as video sequence. At current frame $t$ , multi-view images are first processed by the Feature Extractor to obtain image features and depth distribution. The 3D Coarse Encoder uses the image features and depth distribution to create multiresolution coarse voxel features via LSS-based transformation [12, 13, 37]. The Spatial-Temporal Cascade Decoder then progressively renovates these voxel features spatially and temporally, stage by stage. Finally, the Head module utilizes the refined voxel feature to predict occupancy and scene flow. + +Feature Extractor. At each time $t$ , the feature extractor initially uses an image backbone (e.g., ResNet [7]) to extract multi-view features $F_{t} = \{F_{t}^{j}\in \mathbb{R}^{C\times H\times W}\}_{j = 1}^{N_{c}}$ where $F_{t}^{j}$ represents the features of the $j$ -th camera view at time $t$ , and $N_{c}$ is the number of cameras. Then, the depth network [12, 13] processes these image features to predict the bin depth distribution $D_{t} = \{D_{t}^{j}\in \mathbb{R}^{D_{bin}\times H\times W}\}_{j = 1}^{N_{c}}$ . + +3D Coarse Encoder. The 3D coarse encoder adheres to the Lift and Splat framework as delineated in the LSS paradigm [12, 13, 37]. In the Lift phase, each pixel within the 2D image feature planes $F_{t}$ is projected into the 3D voxel space guided by the predicted bin depth distribution $D_{t}$ . Subsequently, the Splat phase consolidates the feature values of pixels falling within each voxel through voxel pooling [12, 13, 37], thereby constructing the coarse voxel feature $V_{t}^{c} \in \mathbb{R}^{C \times X \times Y \times Z}$ . Subsequently, we engage a lightweight voxel encoder (e.g. ResNet3D-18 [7]) to produce multi-resolution coarse voxel features $V_{t}^{C} = \{\hat{V}_{t}^{i} \in \mathbb{R}^{C \times X_{i} \times Y_{i} \times Z_{i}} \mid i = 1,2,\dots,N\}$ , where $X_{i} = \frac{X}{2^{(N - i)}}, Y_{i} = \frac{Y}{2^{(N - i)}}, Z_{i} = \frac{Z}{2^{(N - i)}}$ , and $N$ is the number of processing stages. + +# 3.2. Spatial-Temporal Cascade Decoder + +To explicitly model the spatial and temporal information into features with the occupied state of 3D space, we introduce a spatial-temporal cascade decoder that renovates the 3D coarse voxel features $V_{t}^{\mathcal{C}}$ through a multi-stage process. As depicted in Fig. 2, the decoder comprises two pri + +mary components: Occlusion-Aware (OA) transformer layers, which accurately capture spatial and short-term temporal information, and a sparse-based temporal fusion module that employs a first-in, first-out (FIFO) memory sequence to encode long-term temporal information. + +At each stage $i$ , we refine voxel feature in BEV space rather than in voxel space. Initially, we fuse the coarse voxel feature $\hat{V}_t^i$ with the refined voxel feature from the previous stage output $V_{t}^{i - 1}$ , and then project it into the BEV representation to obtain $B_{t,0}^{i}\in \mathbb{R}^{C\times X_{i}\times Y_{i}}$ as the input of OA transformer. We subsequently apply $L$ layers OATransformer, which is analogous to the approach in [15], to refine $B_{t,0}^{i}$ . This transformer layer includes three specialized modules: the Self-Recursive Occupancy Predictor (SROP), OA Temporal Self-Attention (OA-TSA), and OA Spatial Cross-Attention (OA-SCA). The SROP is designed to provide an accurate occupied state of the 3D space for each stage. The OA-TSA captures short-range temporal dynamics within the BEV space. The OA-SCA explicitly utilizes the occupied state to address the ambiguous projection problem [15, 17], which transfers geometric 2D feature information into the 3D space. + +After the $OA$ -Transformer processes the features, the refined BEV features $\tilde{B}_t^i$ are converted back to voxel form $\tilde{V}_t^i$ using the occupied weight $W_t^i$ . Subsequently, we leverage the occupied state of the 3D space to guide a novel sparse-based approach for modeling long-term temporal information, thereby obtaining the output $V_t^i$ at time $t$ . $V_t^i$ is upsampled by a factor of 2 for the next stage of refinement. + +# 3.2.1. Self-Recursive Occupancy Predictor + +To provide a more accurate representation of the occupied state of 3D space and to mitigate the one-off selection issues present in previous methods [14, 22, 29]. We draw inspiration from earlier studies [19, 46, 49] and design the self-recursive occupancy predictor. This predictor employs successive transformer layers to progressively refine the occupied state layer by layer. Specifically, at each layer $l$ , it utilizes a simple Multilayer Perceptron (MLP) to recover the height of $B_{t,l}^{i}$ to voxel space and predict the occupancy weights $W_{t,l}^{i} \in \mathbb{R}^{X_{i} \times Y_{i} \times Z_{i}}$ . The process of the self-recursive occupancy predictor can be described as follows: + +$$ +W _ {t, l} ^ {i} = f ^ {i} \left(B _ {t, l} ^ {i}\right) + \alpha_ {l} ^ {i} W _ {t, l - 1} ^ {i}, \tag {1} +$$ + +The function $f^{i}(\cdot)$ corresponds to the occupancy predictor in stage $i$ , which shares the same weights across different layers. The parameter $\alpha_{l}^{i}$ signifies the effect of layer $l - 1$ , a learnable parameter initialized to 0.5. At each stage, the initial $W_{t,0}^{i}$ is derived by upsampling the occupied weights from the previous stage, except that it is set to zeros in the first stage. + +# 3.2.2. OA Temporal Self-Attention + +The temporal modeling is crucial for representing the dynamic driving scene [15]. Given the historical BEV feature $B_{t-1}^{i}$ (discussed in Sec. 3.2.4), we align it with the current feature $B_{t,l}^{i}$ via the motion of the ego-vehicle. To efficiently model the dynamic information, we propose Occlusion-Aware Temporal Self-Attention (OA-TSA) to focus the temporal modeling on the non-empty space. The OA-TSA is represented by: + +$$ +T S A _ {O A} \left(Q _ {x, y}, \mathcal {B}, \bar {W}\right) = \sum_ {b \in \mathcal {B}} \bar {w} _ {x, y} H \left(Q _ {x, y}, b, \bar {w} _ {x, y}\right), \tag {2} +$$ + +where $Q_{x,y}$ denotes the BEV feature located at $p = (x,y)$ and $\mathcal{B} = \{B_{t,l}^{i}, B_{t - 1}^{i}\}$ . $\overline{W} \in \mathbb{R}^{X_i \times Y_i}$ is the average of $W_{t,l}^{i}$ along the z-axis, and $H$ signifies the deformable attention mechanism [54]. $\overline{w}_{x,y}$ represents the occupied weights $\overline{W}$ at position $p$ . Unlike the vanilla deformable attention [54], the offsets are predicted by the concatenation of $\overline{W}$ and $\mathcal{B}$ . By reweighting the TSA, we enable the model to focus more effectively on the dynamic information within the 3D space. + +# 3.2.3. OA Spatial Cross-Attention + +To explicitly utilize the occupied state to renovate the 3D features, we propose the Occupancy-Aware Spatial Cross-Attention (OA-SCA), which leverages the occupied state to enhance geometric features. + +We first revisit the vanilla Spatial Cross-Attention (SCA) [15] as follows. As shown in Fig. 3, it samples 3D points in BEV plane, and projects these points to 2D image feature planes $F_{t}$ to obtain corresponding features. Formally, the SCA can be expressed as: + +$$ +S C A \left(Q _ {x, y}, F _ {t}\right) = \sum_ {z = 1} ^ {N _ {\text {r e f}}} H \left(Q _ {\mathbf {x}}, P (\mathbf {x}), F _ {t}\right), \tag {3} +$$ + +where $Q_{\mathbf{x}}$ denotes the sampled feature at coordinates $\mathbf{x} = (x, y, z)$ in the 3D space and $N_{ref}$ is the number of reference points for each $Q_{x, y}$ . To simplify our initial analysis, we consider a scenario with a single camera, $P(\cdot)$ is the projection matrix that projects points from the 3D space onto the feature plane. The projection process can be mathematically represented as: + +$$ +d \cdot \left[ \begin{array}{l l l} u & v & 1 \end{array} \right] ^ {T} = P \cdot \left[ \begin{array}{l l l l} x & y & z & 1 \end{array} \right] ^ {T}, \tag {4} +$$ + +where $d$ denotes the depth of the point $(u,v)$ on the image plane. This 3D to 2D transformation introduces ambiguity, as different 3D points along the same projection ray map to identical 2D coordinates and are assigned the same features $F_{(u,v)}$ , as illustrated in Fig. 3, where even the green points corresponding to empty areas receive the same feature. + +To address this ambiguity, previous methods [17, 19] propose utilizing depth information to reweight the features of sampled points. However, these methods overlook the + +![](images/bfc4fb20579872dbca9c785f8b9589220eff488234fd93cc0dd8b667c63d7b36.jpg) +Figure 3. Illustration of OA-SCA. Due to the projection process, sampled points along the same ray in the feature plane are assigned identical features, even when they represent empty voxel space, as depicted by the green points. To address this limitation, our approach integrates depth and occupancy information to assign appropriate weights to the sampled points, thereby enhancing the differentiation of features along each ray. + +precision of the predicted depth and the state attributes of the sampled features. In contrast, our approach is inspired by volume rendering techniques [28, 32], which allows us to renovate the features of sampled points more effectively. The volume rendering can be represented as: + +$$ +\mathcal {C} = \sum_ {k = 1} ^ {N _ {c}} \tau_ {k} \cdot \sigma_ {k} \cdot c _ {k}, \tag {5} +$$ + +where $\mathcal{C}$ represents the expected value of light emitted by particles within the volume as a ray samples $N_{c}$ points. $\sigma_{k}$ denotes the density at each point, $\tau_{k}$ is the transmittance, and $c_{k}$ is the corresponding color value. This physical function is analogous to the reverse processing of SCA, where $\mathcal{C}$ corresponds to $F_{(u,v)}$ in the image plane, and $c_{k}$ corresponds to the sampled point features, as illustrated in Fig. 3. + +Based on above observation, our propose OA-SCA is designed to address the ambiguity inherent in vanilla SCA. Furthermore, to maintain the sparsity of the model, we employ probability sampling to select 3D reference points for refinement. The OA-SCA can be formulated as: + +$$ +S C A _ {O A} \left(Q _ {x, y}, F _ {t}\right) = \sum_ {\mathbf {x} \in N _ {s}} \Omega_ {\mathbf {x}} H \left(Q _ {\mathbf {x}}, P (\mathbf {x}), F _ {t}\right), \tag {6} +$$ + +where $N_{s} = \{\mathbf{x}\in \mathbf{X}\mid u < w_{\mathbf{x}}\}$ and $\mathbf{X}$ is total 3D reference points for $Q_{x,y}$ . During training, $u$ follows a truncated normal distribution (0.5,1). For stable inference, $u$ is set to 0.5. This sampling approach allows our method to account for the uncertainty region during training. $w_{\mathbf{x}}$ is the reference point $\mathbf{x}$ corresponding occupied weight. The reweighting parameter $\Omega_{\mathbf{x}} = w_{\mathbf{x}}\cdot w_{d}$ , where $w_{d}$ can be calculated as: + +$$ +w _ {d} = \exp \left(- \frac {\min \left(\left| d _ {r} - \left(d _ {r} ^ {\prime} - \Delta d\right) \right| , \left| d _ {r} - \left(d _ {r} ^ {\prime} + \Delta d\right) \right|\right) ^ {2}}{2 \sigma^ {2}}\right), \tag {7} +$$ + +![](images/ba5ac319bae4cadd0f27f689b7ec15385d7970c02dcf3f293fdf71419b8c9eab.jpg) +Figure 4. Illustration of Sparse Temporal Fusion. We implement temporal fusion using a parallel strategy in a sparse manner, focusing only on modeling the sampled features. + +$d_{r}$ and $d_r^\prime$ represent the depth attributes of the reference point respectively analogous to $z$ and $d$ in Eq. (4), they denote the distance of the sampled point from the ego-vehicle and the distance of the ray's corresponding object from the ego-vehicle, respectively. It should be noted that $d_{r}$ is derived from the predicted bin depth distribution $D_{t}$ and is transformed into relative depth, while $\Delta d$ represents the bin interval. The parameter $\sigma$ serves as an adjustment factor for $d_{r}$ and $d_r^\prime$ , enabling fine-tuning of the depth matching tolerance. By default, $\sigma$ is set to 2, offering a balance between strict and lenient matching. This design takes into account both the attributes of bin depth distribution and the state of the sampled point, enabling a more accurate modeling of 3D spatial information. + +# 3.2.4. Sparse Temporal Modeling + +To integrate long-term historical information into the feature representation, at each stage $i$ , we maintain a streaming history memory bank $\{V_{t - k}^{i}\}_{k = 1}^{T_{i}}$ that adheres to the first-in, first-out rule to dynamically fuse information using a parallel strategy. Inspired by the SlowFast [6] and considering the redundancy in 3D space, we decouple the non-empty and empty regions into long-term and short-term streams, respectively. The long-term stream models the non-empty region to capture long-term dynamic information, while the short-term stream focuses on the empty region, modeling the overall 3D space in a short-term manner. + +For clarity, we consider one stream as an example, as shown in Fig. 4. Given the current refined voxel feature $\tilde{V}_t^i$ and the occupied weights $W_{t}^{i}$ , we first apply a top-k sampling method to extract the seed feature $S_{t}^{i}\in \mathbb{R}^{C\times N_{S}}$ and corresponding occupied state $w_{t}^{i}\in \mathbb{R}^{C_{s}\times N_{S}}$ , where $C_s$ represents the embedding dimensions for occupied weights. Utilizing the position of $S_{t}$ and the ego-pose transformation matrix $T_{t}^{t - j}$ from frame $t$ to frame $t - j$ , we can retrieve the corresponding historical feature $S_{t - j}^{i}$ . We then concatenate all the corresponding historically sampled features with $w_{t}^{i}$ and subsequently apply MLP to fuse the information along the channel dimension: + +$$ +S ^ {i} = M L P \left(\left[ S _ {t} ^ {i}, S _ {t - 1} ^ {i}, \dots , S _ {t - k} ^ {i}, w _ {t} ^ {i} \right]\right). \tag {8} +$$ + +Finally, we add $S^i$ to the corresponding position in $\tilde{V}_t^i$ . By default, at each stage, we use $T_i$ frames to model the + +
MethodBackboneInput SizeEpochsRayIoU1m(%)↑RayIoU2m(%)↑RayIoU4m(%)↑RayIoU(%)↑Memory(G)↓
BEVFormer* [15]R1011600 × 9002426.132.938.032.424.1
RenderOcc* [33]Swin-B1408 × 5121213.419.625.519.517.5
BEVDet-Occ* [8]R50704 × 2569023.630.035.129.610.2
FB-Occ* [17]R50704 × 2569026.734.139.733.59.6
SparseOcc (16f) [22]R50704 × 2562429.135.840.335.123.7
COTR† [27]R50704 × 2562436.341.745.141.018.3
OPUS-L [42]R50704 × 25610034.742.146.741.212.1
STCOcc (ours)R50704 × 2563636.242.746.441.77.7
STCOcc (ours)R501408 × 5123636.942.846.742.18.9
+ +Table 1. Comparison of RayIoU (%) performance on the Occ3D-nus dataset. * indicates models trained with camera mask, † denotes that official code was utilized to retrain the model. + +
MethodBackboneInput SizeMemory(G)mIoU(%)othersbarrierbicyclebuscarcons. veh.motor.pedes.tfc. conetrailertruckdrv. surf.other flatsidewalkterrainmanmadevegetation
BEVFormer [15]R1011600×90024.126.95.937.817.940.442.47.423.921.821.022.430.755.428.436.028.120.017.7
CTF-Occ [41]R1011600×900-28.58.139.320.638.342.216.924.522.721.123.031.153.333.838.033.220.818.0
TPVFormer [9]R1011600×90028.927.87.238.913.740.845.917.220.018.914.326.734.255.735.537.630.719.416.8
OSP* [38]R1011600×90020.741.210.949.027.750.255.922.931.030.930.335.631.282.142.651.955.144.838.2
SparseOcc (8f) [22]R50704×25614.330.910.639.220.232.943.319.423.823.429.321.429.367.736.344.640.922.021.9
FB-Occ* [16]R50704×2569.639.113.644.727.045.449.125.226.327.927.832.336.880.142.851.255.142.237.5
ViewFormer* [11]R50704×256-41.912.950.127.944.652.922.429.628.029.235.239.484.749.457.459.747.440.6
COTR* [27]R50704×25618.344.513.352.131.946.055.632.632.830.434.137.741.884.546.257.660.751.946.3
STCOcc* (ours)R50704×2567.744.615.352.931.646.455.931.532.632.134.539.542.583.647.856.460.150.844.8
STCOcc* (ours)R501408×5128.945.015.252.332.250.556.531.733.933.433.838.944.983.947.457.160.150.642.7
+ +Table 2. Comparison of mIoU (%) performance on the Occ3d-nus dataset. * indicates models trained with camera mask. + +long-term stream and $\frac{T_i}{2}$ frames to model the short-term stream. At each stage $i$ , the fusion output $V_{t}^{i}$ is appended to the memory queue, and a BEV representation $B_{t}^{i}$ is stored for the subsequent frame of OA-TSA. This strategy provides a long-term perspective for recurrent temporal modeling and mitigates the gradient vanishing issue that is commonly encountered in previous recurrent temporal modeling methods [15]. + +# 3.3. Loss + +We compute the occupancy loss for each stage, adopting the Scene-Class Affinity Loss $(\mathcal{L}_{scal})$ from MonoScene [5]. This loss is applied to both semantic and geometric predictions to ensure accurate scene understanding. Given the sparsity and class imbalance inherent in 3D scenes, we also utilize a weighted focal loss [16] combined with the Lovasz loss [3]. The loss for each stage $i$ is formulated as follows: + +$$ +\mathcal {L} _ {o c c} ^ {i} = \mathcal {L} _ {s c a l} ^ {g e o} + \mathcal {L} _ {s c a l} ^ {s e m} + \mathcal {L} _ {f o c a l} + \mathcal {L} _ {l o v}. \tag {9} +$$ + +When considering the scene flow, we utilize the L1 loss $\mathcal{L}_1$ to supervise only the foreground voxel. The overall loss function is formulated as follow: + +$$ +\mathcal {L} = \lambda_ {f} \mathcal {L} _ {1} + \mathcal {L} _ {\text {d e p t h}} + \sum_ {i = 1} ^ {N} w _ {i} \times \mathcal {L} _ {\text {o c c}} ^ {i}, \tag {10} +$$ + +where $\mathcal{L}_{depth}$ is the cross-entropy loss used to supervise the depth network. $w_{i}$ is computed by $\frac{1}{2^{N - i}}$ . + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. To evaluate the performance of our model, we utilize the Occ3D-nus dataset [41] for assessing 3D occupancy prediction and the OpenOcc dataset [39] for evaluating scene flow prediction. Both datasets are derived from the NuScenes dataset [4], encompassing 600 outdoor scenes for training, 150 for validation, and 150 for testing. Since the Occ3D-nus dataset does not provide scene flow labels, we have omitted the flow head when using this dataset. Additionally, since NuScenes does not provide depth labels, we follow previous methods [12, 13] by projecting LIDAR points onto the image plane to serve as depth labels. + +Metric. The mean Intersection-over-Union (mIoU) is a prevalent metric for assessing occupancy prediction per + +
MethodSup.BackboneInput SizeRayIoU(%)↑mAVE↓Mem(G)↓
OccNeRF-C [51]CR1011600×90021.61.53-
OccNeRF-LLR1011600×90031.71.59-
RenderOcc [33]LR1011600×90036.71.63-
Let Occ Flow [25]C+LR1011408×51240.51.45-
OccNet [39]3DR1011600×90039.71.61-
BEVFormer† [15]3DR501600×90028.11.1226.0
FB-Occ† [16]3DR50704×25632.30.8311.1
SparseOcc† [22]3DR50704×25633.40.8715.8
STCOcc (ours)3DR50704×25640.80.448.7
+ +formance in the Occ3D-nus dataset. However, this metric only accounts for the visible area at the current moment and may not fully reflect the model's completion capabilities [22]. Consequently, we also employ Ray-Based IoU (RayIoU) [21] to evaluate occupancy prediction performance. RayIoU is computed at three distance thresholds: 1, 2, and 4 meters. The final ranking metric is obtained by averaging the results across these thresholds. Additionally, we assess the performance of scene flow prediction using the mean absolute velocity error (mAVE) [39] across defined categories (e.g., car, truck). The mAVE is calculated for the set of true positives within a query ray threshold of 2 meters. + +Implementation Details. We adopt the depth network from BEVStereo [12] and configure our model with three processing stages. Within each stage, the number of layers in the OA transformer is set to 2. The amount of historical information preserved for each stage is 16, 8, and 4, respectively. Unless specified otherwise, all models use the AdamW optimizer [26] with a global batch size of 16. Moreover, because our sampling strategy results in variable GPU memory costs, we report the maximum value in our results. + +# 4.2. Main Results + +Main Results on Occ3D-nus. As shown in Tab. 1 and Tab. 2. We compare our method with previous state-of-the-art methods on 3D occupancy task. Our methods achieve the state of the art performance $41.7\%$ in RayIoU and $44.6\%$ in mIoU, which is particularly noteworthy given the significantly lower training costs of 7.7GB, as opposed to the high training costs associated with COTR (18.3 GB) and OPUS-L (12.1 GB). Furthermore, since our spatial refinement process is dependent on the image size, we resize the input to $1408 \times 512$ and achieve improved results in terms of RayIoU and mIoU. + +Main Results on OpenOcc. As demonstrated in Tab. 3, we conducted experiments on the OpenOcc dataset to assess + +Table 3. Comparison of RayIoU (%) and mAVE performance on the OpenOcc [39] dataset. C and L denote Camera and Lidar supervision. $\dagger$ denotes that we utilize the official code and add the flow head to produce the results. + +
SpatialTemporalMetric
OA-SCASROPSTFOA-TSARayIoU(%)↑mAVE↓Mem(G)↓
35.11.325.5
✓35.71.275.7
✓✓36.01.205.7
✓✓✓38.00.698.6
✓✓✓✓38.40.638.7
+ +Table 4. Ablation study on the each component. SROP refers to the Self-Recursive Occupancy Predictor, OA-SCA refers to the Occupancy-Aware Spatial Cross Attention, OA-TSA refers to the Occlusion-Aware Temporal Self-Attention, and STF refers to the Sparse Temporal Fusion. + +![](images/43b4ddaca20de8272f049a86707764a305aad0628b24fc072217407b26e90bf8.jpg) +(a) WOA-SCA. + +![](images/a53f192f702fdeb69e0c23196960ddb4d5ebe1ba194dc0b1118bffb44bd5320f.jpg) +(b) W/O OA-SCA. +Figure 5. Ablation on the OA-SCA module. We visualize the features after refinement with and without the OA-SCA module. + +the performance of our model in terms of occupancy and scene flow. Our approach, which employs a smaller backbone (ResNet-50) and a reduced image input size $(704 \times 256)$ , achieves RayIoU scores of $40.8\%$ and mAVE of 0.44. These results surpass those of both OccNet [39] (which uses ResNet-101 with an input size of $1600 \times 900$ ) and Let Occ Flow [25] (which also uses ResNet-101 with an input size of $1408 \times 512$ ). + +# 4.3. Ablation Study + +To investigate the impact of various modules, we perform ablation experiments on the OpenOcc dataset [39]. It is important to note that the ablation experiments were conducted on half of the training dataset and then evaluated on the full validation set. Specifically, we used the first 300 sequences to constitute half of the training dataset. + +The Effectiveness of Each Component. In Tab. 4, we demonstrate the effectiveness of each component in our model. For the baseline, we omit the OA-SCA and make the occupancy predictor independent in each stage, similar to previous one-off selection methods [22, 39]. This baseline achieves a RayIoU of $35.1\%$ and a mAVE of 1.32 with a memory cost of 5.5GB. Integrating the OA-SCA into the baseline results in a $1.7\%$ increase in RayIoU and a $3.7\%$ increase in mAVE, with an additional memory cost of only 0.2GB. Introducing the SROP further enhances the model's performance without incurring any additional memory costs. Utilizing long-term temporal fusion and + +![](images/463015490c7afd5e064f70ddf746add55f468e35ef01bea11aeb6f02b1e82033.jpg) +Figure 6. Qualitative results on Occ3d-nus validation set. As depicted in the red circle, our method delivers detailed predictions for objects such as cars and trucks, while also offering clear boundary delineations for structures like buildings and vegetation. + +the OA-TSA, we achieve a $6.3\%$ increase in RayIoU and a $45.5\%$ increase in mAVE, with a memory cost of only 1.9GB. + +The Effectiveness of Sparse Temporal Modeling. In Tab. 5, we compare several representations in temporal modeling. The BEV modeling approach [34], while saving computation when modeling long-term history, sacrifices spatial information of the 3D space, resulting in poorer performance in occupancy prediction tasks. Our sparse modeling approach outperforms voxel-level modeling approach [16] in both effectiveness and computational cost. Furthermore, in Tab. 6, we compare our proposed Occupancy-Aware Temporal Self Attention (OA-TSA) with the vanilla Temporal Self Attention (TSA) [15]. Our method achieves superior performance in RayIoU and mAVE metrics, attributing this success to the guidance provided by occupied state modeling. + +The Effectiveness of Occupancy-Aware Spatial Cross Attention. In Fig. 5, we compare the features refined by the OA-SCA module with those refined without it. It is evident that the OA-SCA module renovates the features, providing a more accurate geometric representation of the 3D scene, which is crucial for precise spatial modeling. Moreover, the OA-SCA module selectively enhances the foreground objects in the 3D scene, significantly improving the model's discriminability. Furthermore, in Tab. 7, we compare the vanilla Spatial Cross Attention (SCA) [15] and the Depth-Aware Spatial Cross Attention (DA-SCA) proposed by FB-BEV [17] with our OA-SCA. It is observed that, due to inaccurate spatial modeling, neither SCA nor DA-SCA significantly improves performance. In contrast, Our explicit state-based modeling approach leverages the occupied state to accurately capture detailed spatial information. + +
RepresentationFrameRayIoU(%)↑mAVE↓Mem(G)↓
BEV [34]836.80.816.7
Voxel [16]837.20.737.8
Sparse (ours)837.50.717.3
BEV1637.40.777.6
Voxel1638.00.649.8
Sparse (ours)1638.40.608.7
+ +Table 5. Ablation on the Sparse Temporal Fusion. We compare the traditional representations of BEV and voxel to our sparse modeling approach across various frame numbers. + +
MethodRayIoU(%)mAVE
W/O TSA [15]37.90.69
TSA38.00.67
OA-TSA38.30.63
+ +Table 6. Ablation on the OA-TSA Module. + +
MethodRayIoU(%)
W/O SCA [15]37.6
SCA37.5
DA-SCA [17]37.7
OA-SCA38.3
+ +Table 7. Ablation on the OA-SCA Module. + +# 4.4. Visualizations + +In Fig. 6, we present the BEV visualizations on the Occ3D-nus validation set. In comparison to implicit learning-based approaches [16, 22], our explicit state-based method produces clearer boundaries for objects such as cars, buildings, and vegetation. + +# 5. Conclusions + +We propose an explicit state-based modeling approach to capture detailed geometric information in 3D space and integrate long-term dynamic information effectively. Our proposed STCOcc framework incorporates occlusion-aware mechanisms to enhance 3D features in both spatial and temporal aspects, thereby achieving better performance in 3D occupancy and flow prediction. The results demonstrate the efficacy of our paradigm, underscoring its strong potential for applications in downstream tasks. + +# Acknowledgement + +# References + +[1] Tesla AI Day. https://www.youtube.com/watch?v=j0z4FweCy4M, 2021.1 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 1 +[3] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 6 +[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 6 +[5] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 6 +[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 5 +[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[8] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 6 +[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2, 6 +[10] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. arXiv preprint arXiv:2306.15670, 2023. 2 +[11] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. arXiv preprint arXiv:2405.04299, 2024.6 + +This research was supported by the National Natural Science Foundation of China (No. U23B2060, No.62088102), and the Youth Innovation Team of Shaanxi Universities. +[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 3, 6, 7 +[13] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2, 3, 6 +[14] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1, 2, 4 +[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 4, 6, 7, 8 +[16] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 6, 7, 8 +[17] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 4, 6, 8 +[18] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR, 2024. 2 +[19] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2, 4 +[20] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18580-18590, 2023. 2, 3 +[21] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 7 +[22] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 1, 2, 4, 6, 7, 8 +[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European conference on computer vision, pages 531-548. Springer, 2022. 2 +[24] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework + +for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2 +[25] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 7 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Learning, Learning, 2017. 7 +[27] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 6 +[28] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, page 99-108, 1995. 5 +[29] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Xiangrui Zhao, Jongwon Ra, Laijian Li, and Yong Liu. Camera-based 3d semantic scene completion with sparse guidance network. arXiv preprint arXiv:2312.05752, 2023. 1, 2, 4 +[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2 +[31] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5 +[33] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 6, 7 +[34] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 3, 8 +[35] Liang Peng, Junkai Xu, Haoran Cheng, Zheng Yang, Xiaopei Wu, Wei Qian, Wenxiao Wang, Boxi Wu, and Deng Cai. Learning occupancy for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10281-10292, 2024. 2 +[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2 +[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European + +Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3 +[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as a set of points. arXiv preprint arXiv:2407.04049, 2024. 6 +[39] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2, 3, 6, 7 +[40] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2 +[41] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 6 +[42] Jiabao Wang, Zhaojiang Liu, Qiang Meng, Liujiang Yan, Ke Wang, Jie Yang, Wei Liu, Qibin Hou, and Ming-Ming Cheng. Opus: Occupancy prediction using a sparse set. arXiv preprint arXiv:2409.09350, 2024. 6 +[43] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1002-1011, 2023. 2 +[44] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2 +[45] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1 +[46] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv preprint arXiv:2307.13510, 2023. 4 +[47] Junkai Xu, Liang Peng, Haoran Cheng, Linxuan Xia, Qi Zhou, Dan Deng, Wei Qian, Wenxiao Wang, and Deng Cai. Regulating intermediate 3d features for vision-centric autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6306-6314, 2024. 2 +[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2 +[49] Chenhongyi Yang, Tianwei Lin, Lichao Huang, and Elliot J Crowley. Widthformer: Toward efficient transformer-based + +bev view transformation. arXiv preprint arXiv:2401.03836, 2024. 4 +[50] Zetong Yang, Li Chen, Yanan Sun, and Hongyang Li. Visual point cloud forecasting enables scalable autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14673-14684, 2024. 1 +[51] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 7 +[52] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2 +[53] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2 +[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4 \ No newline at end of file diff --git a/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/images.zip b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0dbecb97c41ab82cd00a52c904e8f467a392d9e7 --- /dev/null +++ b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed0adf01dad97edf11793295fd4edaccbd35fd7157a0e1553b6b080b34ac2e40 +size 666764 diff --git a/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/layout.json b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f857b96f6e72480dc2651d8382201790fc22807 --- /dev/null +++ b/2025/STCOcc_ Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction/layout.json @@ -0,0 +1,9529 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 547, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 547, + 138 + ], + "type": "text", + "content": "STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and Scene Flow Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 161, + 493, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 161, + 493, + 204 + ], + "spans": [ + { + "bbox": [ + 117, + 161, + 493, + 204 + ], + "type": "text", + "content": "Zhimin Liao, Ping Wei\\*, Shuaijia Chen, Haoxuan Wang, Ziyang Ren \nNational Key Laboratory of Human-Machine Hybrid Augmented Intelligence \nInstitute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 205, + 423, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 205, + 423, + 217 + ], + "spans": [ + { + "bbox": [ + 187, + 205, + 423, + 217 + ], + "type": "text", + "content": "liaozm@stu.xjtu.edu.cn, pingwei@xjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 544 + ], + "type": "text", + "content": "3D occupancy and scene flow offer a detailed and dynamic representation of 3D scene. Recognizing the sparsity and complexity of 3D space, previous vision-centric methods have employed implicit learning-based approaches to model spatial and temporal information. However, these approaches struggle to capture local details and diminish the model's spatial discriminative ability. To address these challenges, we propose a novel explicit state-based modeling method designed to leverage the occupied state to renovate the 3D features. Specifically, we propose a sparse occlusion-aware attention mechanism, integrated with a cascade refinement strategy, which accurately renovates 3D features with the guidance of occupied state information. Additionally, we introduce a novel method for modeling long-term dynamic interactions, which reduces computational costs and preserves spatial information. Compared to the previous state-of-the-art methods, our efficient explicit renovation strategy not only delivers superior performance in terms of RayIoU and mAVE for occupancy and scene flow prediction but also markedly reduces GPU memory usage during training, bringing it down to 8.7GB. Our code is available on https://github.com/lzzzzzm/STCOcc" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 684 + ], + "type": "text", + "content": "Accurate perception of 3D surrounding scenes is indeed vital for autonomous systems. The goal of occupancy and scene flow prediction [1] is to segment the entire space into 3D voxels and to determine the semantic and flow information of each voxel. This capability is crucial for understanding the environment and making informed decisions, which is well-suited for downstream tasks in autonomous systems, such as mapping and planning [2, 4, 39, 41, 50]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 684, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 684, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 67, + 684, + 295, + 695 + ], + "type": "text", + "content": "Due to data sparsity and information redundancy in 3D" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 245, + 550, + 312 + ], + "blocks": [ + { + "bbox": [ + 319, + 245, + 550, + 312 + ], + "lines": [ + { + "bbox": [ + 319, + 245, + 550, + 312 + ], + "spans": [ + { + "bbox": [ + 319, + 245, + 550, + 312 + ], + "type": "image", + "image_path": "359ed33bc69efd941404ee7c664530e7435e59be568b01cb9afb3950a6409cf4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 314, + 550, + 379 + ], + "blocks": [ + { + "bbox": [ + 320, + 314, + 550, + 379 + ], + "lines": [ + { + "bbox": [ + 320, + 314, + 550, + 379 + ], + "spans": [ + { + "bbox": [ + 320, + 314, + 550, + 379 + ], + "type": "image", + "image_path": "edb18cd4ea98ea62e7a6b2100573c4aedb8c9c517fd0af55a00d133a4d60acf6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 393, + 436, + 488 + ], + "blocks": [ + { + "bbox": [ + 372, + 383, + 495, + 392 + ], + "lines": [ + { + "bbox": [ + 372, + 383, + 495, + 392 + ], + "spans": [ + { + "bbox": [ + 372, + 383, + 495, + 392 + ], + "type": "text", + "content": "(a) Explicit versus Implicit Modeling." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 393, + 436, + 488 + ], + "lines": [ + { + "bbox": [ + 318, + 393, + 436, + 488 + ], + "spans": [ + { + "bbox": [ + 318, + 393, + 436, + 488 + ], + "type": "image", + "image_path": "5ff87b37eb6fe70f86706c6fff5e9787385008d4d9649150ad0d742dd05709d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 509, + 555, + 576 + ], + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 576 + ], + "type": "text", + "content": "Figure 1. (a) Explicit versus Implicit Modeling: We propose a novel explicit state-based modeling approach that explicitly leverages the occupied state to maintain feature sparsity and model spatial details. (b) Comparison with Different Methods: Our approach achieves state-of-the-art performance of RayIoU and mAVE with lower training costs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 437, + 394, + 552, + 488 + ], + "blocks": [ + { + "bbox": [ + 369, + 491, + 500, + 500 + ], + "lines": [ + { + "bbox": [ + 369, + 491, + 500, + 500 + ], + "spans": [ + { + "bbox": [ + 369, + 491, + 500, + 500 + ], + "type": "text", + "content": "(b) Comparison with Different Methods." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 394, + 552, + 488 + ], + "lines": [ + { + "bbox": [ + 437, + 394, + 552, + 488 + ], + "spans": [ + { + "bbox": [ + 437, + 394, + 552, + 488 + ], + "type": "image", + "image_path": "278b845ee1a895f4fa4b6b1298b2257171dea60870df791dae1317c906e6e924.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "space, employing efficient and robust approaches for 3D feature processing is critical. Existing vision-centric methods [8, 9, 14-16, 22, 29, 31, 41, 45, 53] rely on implicit learning-based method for modeling spatial and temporal information, as shown in Fig. 1a. While these methods optimize 3D reconstruction and long-term temporal fusion through loss supervision, they face inherent limitations. Specifically, the exclusive reliance on loss-driven training hinders the model's ability to capture fine-grained spatial details and compromises the spatial discriminative capacity" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1516" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "of learned features, ultimately reducing the effectiveness for occupancy prediction tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 295, + 228 + ], + "type": "text", + "content": "In this paper, we propose an explicit state-based modeling method that leverages the occupied state of 3D space to refine spatial and temporal feature representations, as shown in Fig. 1a. Our key insight stems from the inherent geometric correspondence between the occupied state and the 3D structure. Since the occupied state directly encodes the geometry of 3D space, it can serve as a robust prior to guide feature learning. This geometric alignment ensures that the feature space preserves structural fidelity, thereby simplifying the learning process and enhancing the discriminative power of the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 230, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 493 + ], + "type": "text", + "content": "We introduce STCOcc, a sparse Spatial-Temporal Cascade renovation framework tailored for occupancy and scene flow prediction. Within this framework, we present a Spatial-Temporal Cascade Decoder designed to renovate the 3D features both spatially and temporally, leveraging the occupied state of the 3D space. Specifically, at each stage, we employ a Self-Recursive Occupancy Predictor (SROP) to progressively refine the occupied state of the 3D space, thereby providing a more precise 3D geometric state for renovating the 3D features. Subsequently, we propose a sparse occlusion-aware attention mechanism to renovate the 3D features. Our attention mechanism differs from prior methods [15, 17, 19], which relied solely on depth information to renovate the 3D features. Instead, we utilize the occupied state in conjunction with bin depth information to accurately model the 3D spatial features. This approach provides details of local regions and makes the features more geometrically accurate. Furthermore, leveraging the accurate occupied state identified by SROP, we employ the Occlusion-Aware Temporal Self-Attention (OA-TSA) to model dynamic information using a recurrent strategy, supplying detailed short-term temporal information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 494, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 636 + ], + "type": "text", + "content": "To efficiently integrate long-term temporal fusion, we propose a novel sparse-based method for temporal fusion modeling. It also avoids redundant information in historical data and preserves spatial information. Specifically, based on the occupied state, we sample non-empty and empty regions into long-term and short-term streams, respectively. Then, we incorporate the occupied state into both streams and employ a parallel strategy to fuse the temporal information within these two streams. This approach not only reduces computational costs but also retains the spatial information within the 3D space. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": "- We introduce an explicit state-based modeling approach designed to renovate the 3D features both spatially and temporally." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- We propose a sparse, occlusion-aware mechanism that provides more accurate geometric 3D features. Additionally, we propose a novel sparse-based method for mod" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 108 + ], + "type": "text", + "content": "eling long-term dynamic information. This approach not only reduces computational costs but also ensures spatial consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "text", + "content": "- Our method achieves a RayIoU of " + }, + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "inline_equation", + "content": "41.7\\%" + }, + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "text", + "content": " on Occ3D [41] and a RayIoU of " + }, + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "inline_equation", + "content": "40.8\\%" + }, + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "text", + "content": " along with a mAVE of 0.44 for occupancy and scene flow prediction on OpenOcc [39], while also reducing the training memory usage to 8.7GB, as shown in Fig. 1b." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 400, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 400, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 400, + 193 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 201, + 529, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 529, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 529, + 213 + ], + "type": "text", + "content": "2.1. Camera-based 3D Occupancy Prediction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 219, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 313 + ], + "type": "text", + "content": "Occupancy, as proposed by [30, 36], focuses on the continuous representation of 3D scenes. MonoScene [5] leverages monocular images for semantic scene completion, employing a 3D UNet to process voxel features. TPVFormer [9] lifts image features into 3D TPV space and expands them into voxel representations for 3D occupancy prediction. OccFormer [53] proposes a dual-path transformer for encoding the dense 3D volume features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 554 + ], + "type": "text", + "content": "Considering the inherent sparsity of 3D scenes, recent methods [14, 18, 22, 29, 40, 47] optimize computational efficiency by processing only non-empty voxels using sparse convolutions or attention mechanisms. VoxFormer [14] utilizes a depth-based query proposal network to generate sparse query proposals for 3D-to-2D cross-attention. SGN [29] introduces a dense-sparse-dense framework that dynamically selects sparse seed voxels and employs hybrid guidance to enhance the convergence of semantic diffusion. Symphonize [10] reconstructs the 3D scene using instance queries. SparseOcc [22] proposes a fully sparse framework that focuses exclusively on non-empty regions. However, these sparse methods rely solely on the occupied state of 3D space to select the region of interest which ignores feature semantics or contextual relationships. Several methods [35, 47] leverage occupancy-based loss supervision to refine 3D features, improving the spatial fidelity of feature representations. However, their exclusive reliance on loss-driven optimization restricts their ability to model fine-grained 3D spatial structures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 503, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 503, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 503, + 577 + ], + "type": "text", + "content": "2.2. Camera-based Temporal Modeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Temporal modeling is essential for camera-based perception due to the inherent challenge of lacking depth information. When considering the modeling space, the methods can be divided into two main categories: image feature-based [19, 20, 23, 24, 44] and 3D feature-based [12, 13, 15, 34, 39, 43, 48, 52]. Image feature-based temporal modeling methods utilize multi-frame image features to provide dynamic information. For instance, PETR [23, 24] projects 3D points onto multi-frame image features to generate implicit 3D features for modeling temporal information. Sparse4D [19] creates 4D keypoints based on 3D an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1517" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 526, + 308 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 526, + 308 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 526, + 308 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 526, + 308 + ], + "type": "image", + "image_path": "a3a4a0e4bca34e31d6d98889fc8eaccc4d7f69a5cd3014c84340dfb4bd16db3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 315, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 315, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. The overall architecture of STCOcc. The STCOcc framework is primarily composed of four integral modules: a feature extractor that captures image features and depth distribution, a 3D coarse encoder that generates multi-resolution coarse voxel features, a multi-stage spatial-temporal cascade decoder that incrementally renovates these coarse voxel features in both spatial and temporal dimensions, and a head module designed to leverage the refined voxel features for the prediction of 3D occupancy and scene flow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 371, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 295, + 407 + ], + "type": "text", + "content": "chors and projects these points to aggregate features from multi-frame image features. SparseBEV [20] adaptively generates sampling points based on the query features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 410, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 295, + 506 + ], + "type": "text", + "content": "On the other hand, 3D feature-based methods model temporal features in BEV or voxel space. BEVFormer [15] designs a temporal self-attention mechanism to recursively fuse BEV features. OccNet [39] extends this paradigm to voxel-based temporal self-attention to recursively fuse voxel feature. SOLOFusion [34] uses a parallel strategy to model BEV-level long-term information. We propose a novel sparse approach to model long-term 3D features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 526, + 116, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 116, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 116, + 539 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 549, + 176, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 176, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 176, + 561 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 712 + ], + "type": "text", + "content": "An overview of STCOcc is presented in Fig. 2. Along the timestamp, we take the multi-view images as video sequence. At current frame " + }, + { + "bbox": [ + 54, + 570, + 295, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 570, + 295, + 712 + ], + "type": "text", + "content": ", multi-view images are first processed by the Feature Extractor to obtain image features and depth distribution. The 3D Coarse Encoder uses the image features and depth distribution to create multiresolution coarse voxel features via LSS-based transformation [12, 13, 37]. The Spatial-Temporal Cascade Decoder then progressively renovates these voxel features spatially and temporally, stage by stage. Finally, the Head module utilizes the refined voxel feature to predict occupancy and scene flow." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": "Feature Extractor. At each time " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": ", the feature extractor initially uses an image backbone (e.g., ResNet [7]) to extract multi-view features " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "F_{t} = \\{F_{t}^{j}\\in \\mathbb{R}^{C\\times H\\times W}\\}_{j = 1}^{N_{c}}" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "F_{t}^{j}" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": " represents the features of the " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": "-th camera view at time " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": " is the number of cameras. Then, the depth network [12, 13] processes these image features to predict the bin depth distribution " + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "inline_equation", + "content": "D_{t} = \\{D_{t}^{j}\\in \\mathbb{R}^{D_{bin}\\times H\\times W}\\}_{j = 1}^{N_{c}}" + }, + { + "bbox": [ + 313, + 371, + 555, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": "3D Coarse Encoder. The 3D coarse encoder adheres to the Lift and Splat framework as delineated in the LSS paradigm [12, 13, 37]. In the Lift phase, each pixel within the 2D image feature planes " + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": " is projected into the 3D voxel space guided by the predicted bin depth distribution " + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": ". Subsequently, the Splat phase consolidates the feature values of pixels falling within each voxel through voxel pooling [12, 13, 37], thereby constructing the coarse voxel feature " + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "inline_equation", + "content": "V_{t}^{c} \\in \\mathbb{R}^{C \\times X \\times Y \\times Z}" + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": ". Subsequently, we engage a lightweight voxel encoder (e.g. ResNet3D-18 [7]) to produce multi-resolution coarse voxel features " + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "inline_equation", + "content": "V_{t}^{C} = \\{\\hat{V}_{t}^{i} \\in \\mathbb{R}^{C \\times X_{i} \\times Y_{i} \\times Z_{i}} \\mid i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "inline_equation", + "content": "X_{i} = \\frac{X}{2^{(N - i)}}, Y_{i} = \\frac{Y}{2^{(N - i)}}, Z_{i} = \\frac{Z}{2^{(N - i)}}" + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 461, + 556, + 631 + ], + "type": "text", + "content": " is the number of processing stages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 635, + 503, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 503, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 503, + 649 + ], + "type": "text", + "content": "3.2. Spatial-Temporal Cascade Decoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": "To explicitly model the spatial and temporal information into features with the occupied state of 3D space, we introduce a spatial-temporal cascade decoder that renovates the 3D coarse voxel features " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "V_{t}^{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " through a multi-stage process. As depicted in Fig. 2, the decoder comprises two pri" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1518" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "mary components: Occlusion-Aware (OA) transformer layers, which accurately capture spatial and short-term temporal information, and a sparse-based temporal fusion module that employs a first-in, first-out (FIFO) memory sequence to encode long-term temporal information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": "At each stage " + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": ", we refine voxel feature in BEV space rather than in voxel space. Initially, we fuse the coarse voxel feature " + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\hat{V}_t^i" + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": " with the refined voxel feature from the previous stage output " + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "inline_equation", + "content": "V_{t}^{i - 1}" + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": ", and then project it into the BEV representation to obtain " + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "inline_equation", + "content": "B_{t,0}^{i}\\in \\mathbb{R}^{C\\times X_{i}\\times Y_{i}}" + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": " as the input of OA transformer. We subsequently apply " + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": " layers OATransformer, which is analogous to the approach in [15], to refine " + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "inline_equation", + "content": "B_{t,0}^{i}" + }, + { + "bbox": [ + 55, + 134, + 295, + 337 + ], + "type": "text", + "content": ". This transformer layer includes three specialized modules: the Self-Recursive Occupancy Predictor (SROP), OA Temporal Self-Attention (OA-TSA), and OA Spatial Cross-Attention (OA-SCA). The SROP is designed to provide an accurate occupied state of the 3D space for each stage. The OA-TSA captures short-range temporal dynamics within the BEV space. The OA-SCA explicitly utilizes the occupied state to address the ambiguous projection problem [15, 17], which transfers geometric 2D feature information into the 3D space." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": "After the " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "OA" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": "-Transformer processes the features, the refined BEV features " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{B}_t^i" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": " are converted back to voxel form " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{V}_t^i" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": " using the occupied weight " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "W_t^i" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": ". Subsequently, we leverage the occupied state of the 3D space to guide a novel sparse-based approach for modeling long-term temporal information, thereby obtaining the output " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "V_t^i" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "inline_equation", + "content": "V_t^i" + }, + { + "bbox": [ + 55, + 340, + 296, + 426 + ], + "type": "text", + "content": " is upsampled by a factor of 2 for the next stage of refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 238, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 238, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 238, + 449 + ], + "type": "text", + "content": "3.2.1. Self-Recursive Occupancy Predictor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "text", + "content": "To provide a more accurate representation of the occupied state of 3D space and to mitigate the one-off selection issues present in previous methods [14, 22, 29]. We draw inspiration from earlier studies [19, 46, 49] and design the self-recursive occupancy predictor. This predictor employs successive transformer layers to progressively refine the occupied state layer by layer. Specifically, at each layer " + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "text", + "content": ", it utilizes a simple Multilayer Perceptron (MLP) to recover the height of " + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "inline_equation", + "content": "B_{t,l}^{i}" + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "text", + "content": " to voxel space and predict the occupancy weights " + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "inline_equation", + "content": "W_{t,l}^{i} \\in \\mathbb{R}^{X_{i} \\times Y_{i} \\times Z_{i}}" + }, + { + "bbox": [ + 55, + 454, + 295, + 588 + ], + "type": "text", + "content": ". The process of the self-recursive occupancy predictor can be described as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 601, + 295, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 601, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 116, + 601, + 295, + 616 + ], + "type": "interline_equation", + "content": "W _ {t, l} ^ {i} = f ^ {i} \\left(B _ {t, l} ^ {i}\\right) + \\alpha_ {l} ^ {i} W _ {t, l - 1} ^ {i}, \\tag {1}", + "image_path": "8abad2a913f775b0f964c4b5198a021646143a9601d365af8e99a5d9ff0535f2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "The function " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "f^{i}(\\cdot)" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " corresponds to the occupancy predictor in stage " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", which shares the same weights across different layers. The parameter " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{l}^{i}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " signifies the effect of layer " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", a learnable parameter initialized to 0.5. At each stage, the initial " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "W_{t,0}^{i}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " is derived by upsampling the occupied weights from the previous stage, except that it is set to zeros in the first stage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 463, + 84 + ], + "type": "text", + "content": "3.2.2. OA Temporal Self-Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "type": "text", + "content": "The temporal modeling is crucial for representing the dynamic driving scene [15]. Given the historical BEV feature " + }, + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "type": "inline_equation", + "content": "B_{t-1}^{i}" + }, + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "type": "text", + "content": " (discussed in Sec. 3.2.4), we align it with the current feature " + }, + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "type": "inline_equation", + "content": "B_{t,l}^{i}" + }, + { + "bbox": [ + 313, + 87, + 553, + 178 + ], + "type": "text", + "content": " via the motion of the ego-vehicle. To efficiently model the dynamic information, we propose Occlusion-Aware Temporal Self-Attention (OA-TSA) to focus the temporal modeling on the non-empty space. The OA-TSA is represented by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 335, + 186, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 186, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 335, + 186, + 553, + 209 + ], + "type": "interline_equation", + "content": "T S A _ {O A} \\left(Q _ {x, y}, \\mathcal {B}, \\bar {W}\\right) = \\sum_ {b \\in \\mathcal {B}} \\bar {w} _ {x, y} H \\left(Q _ {x, y}, b, \\bar {w} _ {x, y}\\right), \\tag {2}", + "image_path": "a19a82b2638de7e61b9341a1b3700c7776715da96d3bf0eb430e407392653789.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "Q_{x,y}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " denotes the BEV feature located at " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "p = (x,y)" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{B_{t,l}^{i}, B_{t - 1}^{i}\\}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\overline{W} \\in \\mathbb{R}^{X_i \\times Y_i}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " is the average of " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "W_{t,l}^{i}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " along the z-axis, and " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " signifies the deformable attention mechanism [54]. " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\overline{w}_{x,y}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " represents the occupied weights " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\overline{W}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": ". Unlike the vanilla deformable attention [54], the offsets are predicted by the concatenation of " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\overline{W}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 216, + 555, + 313 + ], + "type": "text", + "content": ". By reweighting the TSA, we enable the model to focus more effectively on the dynamic information within the 3D space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 318, + 460, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 460, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 460, + 330 + ], + "type": "text", + "content": "3.2.3. OA Spatial Cross-Attention" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 333, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 380 + ], + "type": "text", + "content": "To explicitly utilize the occupied state to renovate the 3D features, we propose the Occupancy-Aware Spatial Cross-Attention (OA-SCA), which leverages the occupied state to enhance geometric features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 381, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 554, + 441 + ], + "type": "text", + "content": "We first revisit the vanilla Spatial Cross-Attention (SCA) [15] as follows. As shown in Fig. 3, it samples 3D points in BEV plane, and projects these points to 2D image feature planes " + }, + { + "bbox": [ + 313, + 381, + 554, + 441 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 381, + 554, + 441 + ], + "type": "text", + "content": " to obtain corresponding features. Formally, the SCA can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 445, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 445, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 348, + 445, + 553, + 479 + ], + "type": "interline_equation", + "content": "S C A \\left(Q _ {x, y}, F _ {t}\\right) = \\sum_ {z = 1} ^ {N _ {\\text {r e f}}} H \\left(Q _ {\\mathbf {x}}, P (\\mathbf {x}), F _ {t}\\right), \\tag {3}", + "image_path": "88a85a086c6ae427763ba87a2a11f2a19b5f8fb15bc3ba877fb84fb34e15cc1b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "inline_equation", + "content": "Q_{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "content": " denotes the sampled feature at coordinates " + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x, y, z)" + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "content": " in the 3D space and " + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "inline_equation", + "content": "N_{ref}" + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "content": " is the number of reference points for each " + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "inline_equation", + "content": "Q_{x, y}" + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "content": ". To simplify our initial analysis, we consider a scenario with a single camera, " + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "inline_equation", + "content": "P(\\cdot)" + }, + { + "bbox": [ + 313, + 487, + 554, + 571 + ], + "type": "text", + "content": " is the projection matrix that projects points from the 3D space onto the feature plane. The projection process can be mathematically represented as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 579, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 579, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 341, + 579, + 553, + 597 + ], + "type": "interline_equation", + "content": "d \\cdot \\left[ \\begin{array}{l l l} u & v & 1 \\end{array} \\right] ^ {T} = P \\cdot \\left[ \\begin{array}{l l l l} x & y & z & 1 \\end{array} \\right] ^ {T}, \\tag {4}", + "image_path": "eaffec773ff1d7dfc144df3fa1a0460a51e0e832ea44aee590b6dc62d38144e5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " denotes the depth of the point " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " on the image plane. This 3D to 2D transformation introduces ambiguity, as different 3D points along the same projection ray map to identical 2D coordinates and are assigned the same features " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "F_{(u,v)}" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": ", as illustrated in Fig. 3, where even the green points corresponding to empty areas receive the same feature." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To address this ambiguity, previous methods [17, 19] propose utilizing depth information to reweight the features of sampled points. However, these methods overlook the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1519" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 291, + 192 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 291, + 192 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 291, + 192 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 291, + 192 + ], + "type": "image", + "image_path": "bfc4fb20579872dbca9c785f8b9589220eff488234fd93cc0dd8b667c63d7b36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 295, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 275 + ], + "type": "text", + "content": "Figure 3. Illustration of OA-SCA. Due to the projection process, sampled points along the same ray in the feature plane are assigned identical features, even when they represent empty voxel space, as depicted by the green points. To address this limitation, our approach integrates depth and occupancy information to assign appropriate weights to the sampled points, thereby enhancing the differentiation of features along each ray." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": "precision of the predicted depth and the state attributes of the sampled features. In contrast, our approach is inspired by volume rendering techniques [28, 32], which allows us to renovate the features of sampled points more effectively. The volume rendering can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 351, + 295, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 351, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 132, + 351, + 295, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {C} = \\sum_ {k = 1} ^ {N _ {c}} \\tau_ {k} \\cdot \\sigma_ {k} \\cdot c _ {k}, \\tag {5}", + "image_path": "646d7aa4d7e77f285f7f23cbf6cd9e17c8b9e76472f16d0e6239919473ac9e9d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " represents the expected value of light emitted by particles within the volume as a ray samples " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " points. " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " denotes the density at each point, " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " is the transmittance, and " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " is the corresponding color value. This physical function is analogous to the reverse processing of SCA, where " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " corresponds to " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "F_{(u,v)}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " in the image plane, and " + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 55, + 392, + 296, + 476 + ], + "type": "text", + "content": " corresponds to the sampled point features, as illustrated in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 477, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 536 + ], + "type": "text", + "content": "Based on above observation, our propose OA-SCA is designed to address the ambiguity inherent in vanilla SCA. Furthermore, to maintain the sparsity of the model, we employ probability sampling to select 3D reference points for refinement. The OA-SCA can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 544, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 544, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 71, + 544, + 295, + 571 + ], + "type": "interline_equation", + "content": "S C A _ {O A} \\left(Q _ {x, y}, F _ {t}\\right) = \\sum_ {\\mathbf {x} \\in N _ {s}} \\Omega_ {\\mathbf {x}} H \\left(Q _ {\\mathbf {x}}, P (\\mathbf {x}), F _ {t}\\right), \\tag {6}", + "image_path": "3345273cfd40e330c2fed1bb3022b1f3fb40d9854ae9f9bb588b1ea3b1eff18b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "N_{s} = \\{\\mathbf{x}\\in \\mathbf{X}\\mid u < w_{\\mathbf{x}}\\}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " is total 3D reference points for " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "Q_{x,y}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": ". During training, " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " follows a truncated normal distribution (0.5,1). For stable inference, " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " is set to 0.5. This sampling approach allows our method to account for the uncertainty region during training. " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "w_{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " is the reference point " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " corresponding occupied weight. The reweighting parameter " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\Omega_{\\mathbf{x}} = w_{\\mathbf{x}}\\cdot w_{d}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "inline_equation", + "content": "w_{d}" + }, + { + "bbox": [ + 55, + 579, + 296, + 668 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 674, + 295, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 295, + 713 + ], + "type": "interline_equation", + "content": "w _ {d} = \\exp \\left(- \\frac {\\min \\left(\\left| d _ {r} - \\left(d _ {r} ^ {\\prime} - \\Delta d\\right) \\right| , \\left| d _ {r} - \\left(d _ {r} ^ {\\prime} + \\Delta d\\right) \\right|\\right) ^ {2}}{2 \\sigma^ {2}}\\right), \\tag {7}", + "image_path": "7fed7ea2d24fcc1e97ae2f32dede62c9fbc084ac03a63f3f2b171b94d20cdfd9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 71, + 545, + 149 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 545, + 149 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 545, + 149 + ], + "type": "image", + "image_path": "ba5ac319bae4cadd0f27f689b7ec15385d7970c02dcf3f293fdf71419b8c9eab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 189 + ], + "type": "text", + "content": "Figure 4. Illustration of Sparse Temporal Fusion. We implement temporal fusion using a parallel strategy in a sparse manner, focusing only on modeling the sampled features." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "d_r^\\prime" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " represent the depth attributes of the reference point respectively analogous to " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " in Eq. (4), they denote the distance of the sampled point from the ego-vehicle and the distance of the ray's corresponding object from the ego-vehicle, respectively. It should be noted that " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " is derived from the predicted bin depth distribution " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " and is transformed into relative depth, while " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\Delta d" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " represents the bin interval. The parameter " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " serves as an adjustment factor for " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "d_{r}" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "d_r^\\prime" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": ", enabling fine-tuning of the depth matching tolerance. By default, " + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 198, + 555, + 365 + ], + "type": "text", + "content": " is set to 2, offering a balance between strict and lenient matching. This design takes into account both the attributes of bin depth distribution and the state of the sampled point, enabling a more accurate modeling of 3D spatial information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 370, + 459, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 459, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 459, + 383 + ], + "type": "text", + "content": "3.2.4. Sparse Temporal Modeling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "type": "text", + "content": "To integrate long-term historical information into the feature representation, at each stage " + }, + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "type": "text", + "content": ", we maintain a streaming history memory bank " + }, + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "type": "inline_equation", + "content": "\\{V_{t - k}^{i}\\}_{k = 1}^{T_{i}}" + }, + { + "bbox": [ + 313, + 385, + 554, + 516 + ], + "type": "text", + "content": " that adheres to the first-in, first-out rule to dynamically fuse information using a parallel strategy. Inspired by the SlowFast [6] and considering the redundancy in 3D space, we decouple the non-empty and empty regions into long-term and short-term streams, respectively. The long-term stream models the non-empty region to capture long-term dynamic information, while the short-term stream focuses on the empty region, modeling the overall 3D space in a short-term manner." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": "For clarity, we consider one stream as an example, as shown in Fig. 4. Given the current refined voxel feature " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\tilde{V}_t^i" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " and the occupied weights " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "W_{t}^{i}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": ", we first apply a top-k sampling method to extract the seed feature " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "S_{t}^{i}\\in \\mathbb{R}^{C\\times N_{S}}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " and corresponding occupied state " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "w_{t}^{i}\\in \\mathbb{R}^{C_{s}\\times N_{S}}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "C_s" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " represents the embedding dimensions for occupied weights. Utilizing the position of " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " and the ego-pose transformation matrix " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "T_{t}^{t - j}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " from frame " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "t - j" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": ", we can retrieve the corresponding historical feature " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "S_{t - j}^{i}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": ". We then concatenate all the corresponding historically sampled features with " + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "inline_equation", + "content": "w_{t}^{i}" + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": " and subsequently apply MLP to fuse the information along the channel dimension:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 667, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 667, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 353, + 667, + 553, + 682 + ], + "type": "interline_equation", + "content": "S ^ {i} = M L P \\left(\\left[ S _ {t} ^ {i}, S _ {t - 1} ^ {i}, \\dots , S _ {t - k} ^ {i}, w _ {t} ^ {i} \\right]\\right). \\tag {8}", + "image_path": "d29769f66e3b0d5733b63055b90bf01683106ed1969f7b2bba4a971d710efba1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Finally, we add " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "S^i" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " to the corresponding position in " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{V}_t^i" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ". By default, at each stage, we use " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " frames to model the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1520" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 169 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 169 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 169 + ], + "type": "table", + "html": "
MethodBackboneInput SizeEpochsRayIoU1m(%)↑RayIoU2m(%)↑RayIoU4m(%)↑RayIoU(%)↑Memory(G)↓
BEVFormer* [15]R1011600 × 9002426.132.938.032.424.1
RenderOcc* [33]Swin-B1408 × 5121213.419.625.519.517.5
BEVDet-Occ* [8]R50704 × 2569023.630.035.129.610.2
FB-Occ* [17]R50704 × 2569026.734.139.733.59.6
SparseOcc (16f) [22]R50704 × 2562429.135.840.335.123.7
COTR† [27]R50704 × 2562436.341.745.141.018.3
OPUS-L [42]R50704 × 25610034.742.146.741.212.1
STCOcc (ours)R50704 × 2563636.242.746.441.77.7
STCOcc (ours)R501408 × 5123636.942.846.742.18.9
", + "image_path": "68b2fab36a77f1a18169c00064d1812fbc6a43fd9b643d13d832be6d9c3dcdb5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 211, + 555, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 177, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 555, + 200 + ], + "type": "text", + "content": "Table 1. Comparison of RayIoU (%) performance on the Occ3D-nus dataset. * indicates models trained with camera mask, † denotes that official code was utilized to retrain the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 211, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 56, + 211, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 555, + 383 + ], + "type": "table", + "html": "
MethodBackboneInput SizeMemory(G)mIoU(%)othersbarrierbicyclebuscarcons. veh.motor.pedes.tfc. conetrailertruckdrv. surf.other flatsidewalkterrainmanmadevegetation
BEVFormer [15]R1011600×90024.126.95.937.817.940.442.47.423.921.821.022.430.755.428.436.028.120.017.7
CTF-Occ [41]R1011600×900-28.58.139.320.638.342.216.924.522.721.123.031.153.333.838.033.220.818.0
TPVFormer [9]R1011600×90028.927.87.238.913.740.845.917.220.018.914.326.734.255.735.537.630.719.416.8
OSP* [38]R1011600×90020.741.210.949.027.750.255.922.931.030.930.335.631.282.142.651.955.144.838.2
SparseOcc (8f) [22]R50704×25614.330.910.639.220.232.943.319.423.823.429.321.429.367.736.344.640.922.021.9
FB-Occ* [16]R50704×2569.639.113.644.727.045.449.125.226.327.927.832.336.880.142.851.255.142.237.5
ViewFormer* [11]R50704×256-41.912.950.127.944.652.922.429.628.029.235.239.484.749.457.459.747.440.6
COTR* [27]R50704×25618.344.513.352.131.946.055.632.632.830.434.137.741.884.546.257.660.751.946.3
STCOcc* (ours)R50704×2567.744.615.352.931.646.455.931.532.632.134.539.542.583.647.856.460.150.844.8
STCOcc* (ours)R501408×5128.945.015.252.332.250.556.531.733.933.433.838.944.983.947.457.160.150.642.7
", + "image_path": "6b3af14d783dad9f5893a972b8d37d0b7b298073b17e53a1e85ef8007892bec5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 392, + 530, + 403 + ], + "lines": [ + { + "bbox": [ + 79, + 392, + 530, + 403 + ], + "spans": [ + { + "bbox": [ + 79, + 392, + 530, + 403 + ], + "type": "text", + "content": "Table 2. Comparison of mIoU (%) performance on the Occ3d-nus dataset. * indicates models trained with camera mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "text", + "content": "long-term stream and " + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\frac{T_i}{2}" + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "text", + "content": " frames to model the short-term stream. At each stage " + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "text", + "content": ", the fusion output " + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "inline_equation", + "content": "V_{t}^{i}" + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "text", + "content": " is appended to the memory queue, and a BEV representation " + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "inline_equation", + "content": "B_{t}^{i}" + }, + { + "bbox": [ + 54, + 423, + 295, + 521 + ], + "type": "text", + "content": " is stored for the subsequent frame of OA-TSA. This strategy provides a long-term perspective for recurrent temporal modeling and mitigates the gradient vanishing issue that is commonly encountered in previous recurrent temporal modeling methods [15]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 101, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 101, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 101, + 544 + ], + "type": "text", + "content": "3.3. Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "type": "text", + "content": "We compute the occupancy loss for each stage, adopting the Scene-Class Affinity Loss " + }, + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{scal})" + }, + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "type": "text", + "content": " from MonoScene [5]. This loss is applied to both semantic and geometric predictions to ensure accurate scene understanding. Given the sparsity and class imbalance inherent in 3D scenes, we also utilize a weighted focal loss [16] combined with the Lovasz loss [3]. The loss for each stage " + }, + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 552, + 296, + 635 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 652, + 295, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 652, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 94, + 652, + 295, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o c c} ^ {i} = \\mathcal {L} _ {s c a l} ^ {g e o} + \\mathcal {L} _ {s c a l} ^ {s e m} + \\mathcal {L} _ {f o c a l} + \\mathcal {L} _ {l o v}. \\tag {9}", + "image_path": "5183a430ff33bfbbf84cff4616969839a674ca6324cd58e586b5db7827110fe3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "When considering the scene flow, we utilize the L1 loss " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " to supervise only the foreground voxel. The overall loss function is formulated as follow:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 434, + 555, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 434, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 355, + 434, + 555, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {f} \\mathcal {L} _ {1} + \\mathcal {L} _ {\\text {d e p t h}} + \\sum_ {i = 1} ^ {N} w _ {i} \\times \\mathcal {L} _ {\\text {o c c}} ^ {i}, \\tag {10}", + "image_path": "8eed43c9caaf13a5f96c56ff6b212a4bbeb67611c9cf112f31a847b18cce7d18.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth}" + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "text", + "content": " is the cross-entropy loss used to supervise the depth network. " + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "text", + "content": " is computed by " + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\frac{1}{2^{N - i}}" + }, + { + "bbox": [ + 313, + 471, + 555, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 397, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 397, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 397, + 520 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 525, + 432, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 432, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 432, + 539 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 543, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 675 + ], + "type": "text", + "content": "Dataset. To evaluate the performance of our model, we utilize the Occ3D-nus dataset [41] for assessing 3D occupancy prediction and the OpenOcc dataset [39] for evaluating scene flow prediction. Both datasets are derived from the NuScenes dataset [4], encompassing 600 outdoor scenes for training, 150 for validation, and 150 for testing. Since the Occ3D-nus dataset does not provide scene flow labels, we have omitted the flow head when using this dataset. Additionally, since NuScenes does not provide depth labels, we follow previous methods [12, 13] by projecting LIDAR points onto the image plane to serve as depth labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Metric. The mean Intersection-over-Union (mIoU) is a prevalent metric for assessing occupancy prediction per" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 758 + ], + "type": "text", + "content": "1521" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 288, + 166 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 288, + 166 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 288, + 166 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 288, + 166 + ], + "type": "table", + "html": "
MethodSup.BackboneInput SizeRayIoU(%)↑mAVE↓Mem(G)↓
OccNeRF-C [51]CR1011600×90021.61.53-
OccNeRF-LLR1011600×90031.71.59-
RenderOcc [33]LR1011600×90036.71.63-
Let Occ Flow [25]C+LR1011408×51240.51.45-
OccNet [39]3DR1011600×90039.71.61-
BEVFormer† [15]3DR501600×90028.11.1226.0
FB-Occ† [16]3DR50704×25632.30.8311.1
SparseOcc† [22]3DR50704×25633.40.8715.8
STCOcc (ours)3DR50704×25640.80.448.7
", + "image_path": "da0fad3885cce6c17a8b2f74a5f4b594e1797d077fbbba16c9f03d2ac459f460.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 235, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 235, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 296, + 390 + ], + "type": "text", + "content": "formance in the Occ3D-nus dataset. However, this metric only accounts for the visible area at the current moment and may not fully reflect the model's completion capabilities [22]. Consequently, we also employ Ray-Based IoU (RayIoU) [21] to evaluate occupancy prediction performance. RayIoU is computed at three distance thresholds: 1, 2, and 4 meters. The final ranking metric is obtained by averaging the results across these thresholds. Additionally, we assess the performance of scene flow prediction using the mean absolute velocity error (mAVE) [39] across defined categories (e.g., car, truck). The mAVE is calculated for the set of true positives within a query ray threshold of 2 meters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 404, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 523 + ], + "type": "text", + "content": "Implementation Details. We adopt the depth network from BEVStereo [12] and configure our model with three processing stages. Within each stage, the number of layers in the OA transformer is set to 2. The amount of historical information preserved for each stage is 16, 8, and 4, respectively. Unless specified otherwise, all models use the AdamW optimizer [26] with a global batch size of 16. Moreover, because our sampling strategy results in variable GPU memory costs, we report the maximum value in our results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 531, + 141, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 141, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 141, + 544 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "text", + "content": "Main Results on Occ3D-nus. As shown in Tab. 1 and Tab. 2. We compare our method with previous state-of-the-art methods on 3D occupancy task. Our methods achieve the state of the art performance " + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "inline_equation", + "content": "41.7\\%" + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "text", + "content": " in RayIoU and " + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "inline_equation", + "content": "44.6\\%" + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "text", + "content": " in mIoU, which is particularly noteworthy given the significantly lower training costs of 7.7GB, as opposed to the high training costs associated with COTR (18.3 GB) and OPUS-L (12.1 GB). Furthermore, since our spatial refinement process is dependent on the image size, we resize the input to " + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "inline_equation", + "content": "1408 \\times 512" + }, + { + "bbox": [ + 55, + 548, + 295, + 680 + ], + "type": "text", + "content": " and achieve improved results in terms of RayIoU and mIoU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Main Results on OpenOcc. As demonstrated in Tab. 3, we conducted experiments on the OpenOcc dataset to assess" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 171, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 214 + ], + "type": "text", + "content": "Table 3. Comparison of RayIoU (%) and mAVE performance on the OpenOcc [39] dataset. C and L denote Camera and Lidar supervision. " + }, + { + "bbox": [ + 55, + 171, + 295, + 214 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 171, + 295, + 214 + ], + "type": "text", + "content": " denotes that we utilize the official code and add the flow head to produce the results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 144 + ], + "type": "table", + "html": "
SpatialTemporalMetric
OA-SCASROPSTFOA-TSARayIoU(%)↑mAVE↓Mem(G)↓
35.11.325.5
✓35.71.275.7
✓✓36.01.205.7
✓✓✓38.00.698.6
✓✓✓✓38.40.638.7
", + "image_path": "00b322ec4fe6cb9d2cb2b6a66dfaa72cdedf53510ad17bed0229c1473a7eb8e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "type": "text", + "content": "Table 4. Ablation study on the each component. SROP refers to the Self-Recursive Occupancy Predictor, OA-SCA refers to the Occupancy-Aware Spatial Cross Attention, OA-TSA refers to the Occlusion-Aware Temporal Self-Attention, and STF refers to the Sparse Temporal Fusion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 347, + 220, + 428, + 301 + ], + "blocks": [ + { + "bbox": [ + 347, + 220, + 428, + 301 + ], + "lines": [ + { + "bbox": [ + 347, + 220, + 428, + 301 + ], + "spans": [ + { + "bbox": [ + 347, + 220, + 428, + 301 + ], + "type": "image", + "image_path": "43b4ddaca20de8272f049a86707764a305aad0628b24fc072217407b26e90bf8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 304, + 414, + 313 + ], + "lines": [ + { + "bbox": [ + 361, + 304, + 414, + 313 + ], + "spans": [ + { + "bbox": [ + 361, + 304, + 414, + 313 + ], + "type": "text", + "content": "(a) WOA-SCA." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 438, + 220, + 520, + 301 + ], + "blocks": [ + { + "bbox": [ + 438, + 220, + 520, + 301 + ], + "lines": [ + { + "bbox": [ + 438, + 220, + 520, + 301 + ], + "spans": [ + { + "bbox": [ + 438, + 220, + 520, + 301 + ], + "type": "image", + "image_path": "a53f192f702fdeb69e0c23196960ddb4d5ebe1ba194dc0b1118bffb44bd5320f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 304, + 510, + 313 + ], + "lines": [ + { + "bbox": [ + 448, + 304, + 510, + 313 + ], + "spans": [ + { + "bbox": [ + 448, + 304, + 510, + 313 + ], + "type": "text", + "content": "(b) W/O OA-SCA." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 319, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 340 + ], + "type": "text", + "content": "Figure 5. Ablation on the OA-SCA module. We visualize the features after refinement with and without the OA-SCA module." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "text", + "content": "the performance of our model in terms of occupancy and scene flow. Our approach, which employs a smaller backbone (ResNet-50) and a reduced image input size " + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "inline_equation", + "content": "(704 \\times 256)" + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "text", + "content": ", achieves RayIoU scores of " + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "inline_equation", + "content": "40.8\\%" + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "text", + "content": " and mAVE of 0.44. These results surpass those of both OccNet [39] (which uses ResNet-101 with an input size of " + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "inline_equation", + "content": "1600 \\times 900" + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "text", + "content": ") and Let Occ Flow [25] (which also uses ResNet-101 with an input size of " + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "inline_equation", + "content": "1408 \\times 512" + }, + { + "bbox": [ + 313, + 363, + 555, + 459 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 468, + 408, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 408, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 408, + 480 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": "To investigate the impact of various modules, we perform ablation experiments on the OpenOcc dataset [39]. It is important to note that the ablation experiments were conducted on half of the training dataset and then evaluated on the full validation set. Specifically, we used the first 300 sequences to constitute half of the training dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "The Effectiveness of Each Component. In Tab. 4, we demonstrate the effectiveness of each component in our model. For the baseline, we omit the OA-SCA and make the occupancy predictor independent in each stage, similar to previous one-off selection methods [22, 39]. This baseline achieves a RayIoU of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "35.1\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and a mAVE of 1.32 with a memory cost of 5.5GB. Integrating the OA-SCA into the baseline results in a " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " increase in RayIoU and a " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " increase in mAVE, with an additional memory cost of only 0.2GB. Introducing the SROP further enhances the model's performance without incurring any additional memory costs. Utilizing long-term temporal fusion and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1522" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 69, + 479, + 247 + ], + "blocks": [ + { + "bbox": [ + 129, + 69, + 479, + 247 + ], + "lines": [ + { + "bbox": [ + 129, + 69, + 479, + 247 + ], + "spans": [ + { + "bbox": [ + 129, + 69, + 479, + 247 + ], + "type": "image", + "image_path": "463015490c7afd5e064f70ddf746add55f468e35ef01bea11aeb6f02b1e82033.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": "Figure 6. Qualitative results on Occ3d-nus validation set. As depicted in the red circle, our method delivers detailed predictions for objects such as cars and trucks, while also offering clear boundary delineations for structures like buildings and vegetation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "text", + "content": "the OA-TSA, we achieve a " + }, + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "text", + "content": " increase in RayIoU and a " + }, + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "inline_equation", + "content": "45.5\\%" + }, + { + "bbox": [ + 55, + 290, + 295, + 326 + ], + "type": "text", + "content": " increase in mAVE, with a memory cost of only 1.9GB." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 508 + ], + "type": "text", + "content": "The Effectiveness of Sparse Temporal Modeling. In Tab. 5, we compare several representations in temporal modeling. The BEV modeling approach [34], while saving computation when modeling long-term history, sacrifices spatial information of the 3D space, resulting in poorer performance in occupancy prediction tasks. Our sparse modeling approach outperforms voxel-level modeling approach [16] in both effectiveness and computational cost. Furthermore, in Tab. 6, we compare our proposed Occupancy-Aware Temporal Self Attention (OA-TSA) with the vanilla Temporal Self Attention (TSA) [15]. Our method achieves superior performance in RayIoU and mAVE metrics, attributing this success to the guidance provided by occupied state modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "The Effectiveness of Occupancy-Aware Spatial Cross Attention. In Fig. 5, we compare the features refined by the OA-SCA module with those refined without it. It is evident that the OA-SCA module renovates the features, providing a more accurate geometric representation of the 3D scene, which is crucial for precise spatial modeling. Moreover, the OA-SCA module selectively enhances the foreground objects in the 3D scene, significantly improving the model's discriminability. Furthermore, in Tab. 7, we compare the vanilla Spatial Cross Attention (SCA) [15] and the Depth-Aware Spatial Cross Attention (DA-SCA) proposed by FB-BEV [17] with our OA-SCA. It is observed that, due to inaccurate spatial modeling, neither SCA nor DA-SCA significantly improves performance. In contrast, Our explicit state-based modeling approach leverages the occupied state to accurately capture detailed spatial information." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 288, + 552, + 362 + ], + "blocks": [ + { + "bbox": [ + 317, + 288, + 552, + 362 + ], + "lines": [ + { + "bbox": [ + 317, + 288, + 552, + 362 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 552, + 362 + ], + "type": "table", + "html": "
RepresentationFrameRayIoU(%)↑mAVE↓Mem(G)↓
BEV [34]836.80.816.7
Voxel [16]837.20.737.8
Sparse (ours)837.50.717.3
BEV1637.40.777.6
Voxel1638.00.649.8
Sparse (ours)1638.40.608.7
", + "image_path": "81c355a4559ccf370c93ee771102325750288193e133f84e6f4d486c1763c542.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 405, + 446, + 456 + ], + "blocks": [ + { + "bbox": [ + 313, + 368, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 401 + ], + "type": "text", + "content": "Table 5. Ablation on the Sparse Temporal Fusion. We compare the traditional representations of BEV and voxel to our sparse modeling approach across various frame numbers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 405, + 446, + 456 + ], + "lines": [ + { + "bbox": [ + 317, + 405, + 446, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 446, + 456 + ], + "type": "table", + "html": "
MethodRayIoU(%)mAVE
W/O TSA [15]37.90.69
TSA38.00.67
OA-TSA38.30.63
", + "image_path": "940279f8e9af3307d239c04236d8d743315c9a0f2ce70bbcd7d2cb84f8fd0e4c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 458, + 405, + 545, + 456 + ], + "blocks": [ + { + "bbox": [ + 314, + 459, + 446, + 480 + ], + "lines": [ + { + "bbox": [ + 314, + 459, + 446, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 459, + 446, + 480 + ], + "type": "text", + "content": "Table 6. Ablation on the OA-TSA Module." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 458, + 405, + 545, + 456 + ], + "lines": [ + { + "bbox": [ + 458, + 405, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 458, + 405, + 545, + 456 + ], + "type": "table", + "html": "
MethodRayIoU(%)
W/O SCA [15]37.6
SCA37.5
DA-SCA [17]37.7
OA-SCA38.3
", + "image_path": "9e4bd6af88ad39c9ee9246e05f2f2d2ea21e7b364cad60f0019e93e19a129f2d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 452, + 460, + 554, + 481 + ], + "lines": [ + { + "bbox": [ + 452, + 460, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 452, + 460, + 554, + 481 + ], + "type": "text", + "content": "Table 7. Ablation on the OA-SCA Module." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 498, + 403, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 403, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 403, + 509 + ], + "type": "text", + "content": "4.4. Visualizations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 575 + ], + "type": "text", + "content": "In Fig. 6, we present the BEV visualizations on the Occ3D-nus validation set. In comparison to implicit learning-based approaches [16, 22], our explicit state-based method produces clearer boundaries for objects such as cars, buildings, and vegetation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 392, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 392, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 392, + 597 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We propose an explicit state-based modeling approach to capture detailed geometric information in 3D space and integrate long-term dynamic information effectively. Our proposed STCOcc framework incorporates occlusion-aware mechanisms to enhance 3D features in both spatial and temporal aspects, thereby achieving better performance in 3D occupancy and flow prediction. The results demonstrate the efficacy of our paradigm, underscoring its strong potential for applications in downstream tasks." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1523" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 294, + 182 + ], + "type": "text", + "content": "[1] Tesla AI Day. https://www.youtube.com/watch?v=j0z4FweCy4M, 2021.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 183, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 296, + 239 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 305 + ], + "type": "text", + "content": "[3] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4413-4421, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "type": "text", + "content": "[4] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 296, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 419 + ], + "type": "text", + "content": "[5] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 463 + ], + "type": "text", + "content": "[6] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 510 + ], + "type": "text", + "content": "[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 555 + ], + "type": "text", + "content": "[8] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 556, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 612 + ], + "type": "text", + "content": "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "type": "text", + "content": "[10] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. arXiv preprint arXiv:2306.15670, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[11] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. arXiv preprint arXiv:2405.04299, 2024.6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This research was supported by the National Natural Science Foundation of China (No. U23B2060, No.62088102), and the Youth Innovation Team of Shaanxi Universities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "text", + "content": "[13] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "type": "text", + "content": "[14] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1, 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 252, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 252, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 252, + 554, + 317 + ], + "type": "text", + "content": "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 319, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 319, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 319, + 554, + 363 + ], + "type": "text", + "content": "[16] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 554, + 420 + ], + "type": "text", + "content": "[17] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 4, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 421, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 554, + 453 + ], + "type": "text", + "content": "[18] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 455, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 554, + 498 + ], + "type": "text", + "content": "[19] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 500, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 554, + 555 + ], + "type": "text", + "content": "[20] Haisong Liu, Yao Teng, Tao Lu, Haiguang Wang, and Limin Wang. Sparsebev: High-performance sparse 3d object detection from multi-camera videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18580-18590, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 556, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 554, + 599 + ], + "type": "text", + "content": "[21] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 601, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 554, + 644 + ], + "type": "text", + "content": "[22] Haisong Liu, Haiguang Wang, Yang Chen, Zetong Yang, Jia Zeng, Li Chen, and Limin Wang. Fully sparse 3d panoptic occupancy prediction. arXiv preprint arXiv:2312.17118, 2023. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 554, + 690 + ], + "type": "text", + "content": "[23] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In European conference on computer vision, pages 531-548. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[24] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petrv2: A unified framework" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1524" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "for 3d perception from multi-camera images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3262-3272, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[25] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 173 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. Learning, Learning, 2017. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 228 + ], + "type": "text", + "content": "[27] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19936-19945, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[28] N. Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, page 99-108, 1995. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[29] Jianbiao Mei, Yu Yang, Mengmeng Wang, Junyu Zhu, Xiangrui Zhao, Jongwon Ra, Laijian Li, and Yong Liu. Camera-based 3d semantic scene completion with sparse guidance network. arXiv preprint arXiv:2312.05752, 2023. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "text", + "content": "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "text", + "content": "[31] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "text", + "content": "[33] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaqi Li, Li Liu, and Shanghang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. arXiv preprint arXiv:2309.09502, 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 557 + ], + "type": "text", + "content": "[34] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. 2, 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[35] Liang Peng, Junkai Xu, Haoran Cheng, Zheng Yang, Xiaopei Wu, Wei Qian, Wenxiao Wang, Boxi Wu, and Deng Cai. Learning occupancy for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10281-10292, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "text", + "content": "[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as a set of points. arXiv preprint arXiv:2407.04049, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "type": "text", + "content": "[39] Chonghao Sima, Wenwen Tong, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, and Hongyang Li. Scene as occupancy. Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 252 + ], + "type": "text", + "content": "[40] Pin Tang, Zhongdao Wang, Guoqing Wang, Jilai Zheng, Xiangxuan Ren, Bailan Feng, and Chao Ma. Sparseocc: Rethinking sparse latent representation for vision-based semantic occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15035-15044, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[41] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. arXiv preprint arXiv:2304.14365, 2023. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[42] Jiabao Wang, Zhaojiang Liu, Qiang Meng, Liujiang Yan, Ke Wang, Jie Yang, Wei Liu, Qibin Hou, and Ming-Ming Cheng. Opus: Occupancy prediction using a sparse set. arXiv preprint arXiv:2409.09350, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 555, + 399 + ], + "type": "text", + "content": "[43] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1002-1011, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[44] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[45] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[46] Yiming Wu, Ruixiang Li, Zequn Qin, Xinhai Zhao, and Xi Li. Heightformer: Explicit height modeling without extra data for camera-only 3d object detection in bird's eye view. arXiv preprint arXiv:2307.13510, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "text", + "content": "[47] Junkai Xu, Liang Peng, Haoran Cheng, Linxuan Xia, Qi Zhou, Dan Deng, Wei Qian, Wenxiao Wang, and Deng Cai. Regulating intermediate 3d features for vision-centric autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6306-6314, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 689 + ], + "type": "text", + "content": "[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Chenhongyi Yang, Tianwei Lin, Lichao Huang, and Elliot J Crowley. Widthformer: Toward efficient transformer-based" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1525" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 351 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "bev view transformation. arXiv preprint arXiv:2401.03836, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[50] Zetong Yang, Li Chen, Yanan Sun, and Hongyang Li. Visual point cloud forecasting enables scalable autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14673-14684, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[51] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 250 + ], + "type": "text", + "content": "[52] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[53] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1526" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_content_list.json b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..660ec716fd88635418ddc55615d4672cf36f34d2 --- /dev/null +++ b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_content_list.json @@ -0,0 +1,1811 @@ +[ + { + "type": "text", + "text": "STDD: Spatio-Temporal Dual Diffusion for Video Generation", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuaizhen Yao $^{1}$ , Xiaoya Zhang $^{1,*}$ , Xin Liu $^{2}$ , Mengyi Liu $^{3}$ , Zhen Cui $^{4,*}$ $^{1}$ Nanjing University of Science and Technology, \n $^{2}$ Nanjing Seetacloud Technology Co., Ltd, \n $^{3}$ Kuaishou, $^{4}$ Beijing Normal University \n{ysz, zhangxiaoya}@njust.edu.cn, xin.liu@seetacloud.com, liumengyi@kuaishou.com, zhen.cui@bnu.edu.cn", + "bbox": [ + 222, + 179, + 781, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion probabilistic model is becoming the cornerstone of data generation, especially generating high-quality images. As an extension, video diffusion generation is in urgent need of a principled temporal-sequence diffusion way, while the spatial-domain diffusion dominates most video diffusion methods. In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal noise propagation/reduction. Mathematically, an analysable dual diffusion process is derived to accumulate information in temporal sequence as well as spatial domain. Correspondingly, we theoretically derive a spatio-temporal probabilistic reverse diffusion process and propose an accelerated sampling way to reduce the inference cost. In principle, the spatio-temporal dual diffusion enables the information of previous frames to be transferred to the current frame, which thus could be beneficial for video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation/prediction as well as text-to-video generation.", + "bbox": [ + 89, + 353, + 485, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion probabilistic model (DPM) is becoming the cornerstone of data generation, where the forward process gradually contaminates the data with gaussian noises and the reverse denoising process is learned to generate new samples [12, 15]. Specifically, DPM has made significant progress in producing high quality, realistic images [17, 23, 24, 27, 29, 33]. Motivated by the success in image generation, some researchers are trying to extend the diffusion model from image to video by accompanying with", + "bbox": [ + 89, + 739, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporal-dimensional prediction/interpolation [9, 39, 49]. However, due to intricate temporal and spatial correlations as well as high dimensionality of video, video diffusion generation remains a challenging task.", + "bbox": [ + 511, + 323, + 906, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The existing video generation methods based on DPMs [9, 13, 28] typically employ a standard diffusion process by treating each frame of the video as an independent diffusion process, which breaks the temporal correlation between frames. Concretely, in the noisig/denoising process, these video diffusion methods separately derive/sample priors, and recover corresponding frames from different-random priors. Due to different-random priors, the continuity of generated video frames is still a bottleneck of video diffusion generation, even though some historical (generated) frames could be used as a condition of the diffusion for some-degree mitigation [25]. An endeavor to this aspect is one recent work [20], decomposing video noise into two components: video sharing base noise and frame-independent residual noise, and then separately performing the diffusion process with two different diffusion networks, which greatly improves video consistency.", + "bbox": [ + 511, + 383, + 908, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite some progress in video diffusion generation, these existing methods could not well capture temporal dependencies due to the use of independent priors for each frame, which would often result into the nonsmoothness or inconsistency of generated videos, especially confronting with large appearance variations in temporal sequences. Essentially, the current video diffusion methods lack a principled diffusion process along time slices (w.r.t frame sequence), not just in the spatial domain (w.r.t every image).", + "bbox": [ + 511, + 642, + 908, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal frame propagation. Mathematically, we derive an analysable dual diffusion process by accumulating temporal frame information. In particular, the historical frames are propagated along with temporal sequence, while accompanying with", + "bbox": [ + 511, + 780, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12575", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the spatio-domain noises w.r.t diffusion on single-frame image. To reach the inverse diffusion, we derive a probabilistic reverse process, which guarantees the feasibility of the diffusion model. Inspired by the diffusion probabilistic implicit diffusion (DDIM) model [33], further, we propose an accelerated spatio-temporal sampling method that is both fast and computation-intensive. To facilitate learning, we decompose the model into spatial components representing the standard diffusion model and temporal components representing the frame-to-frame diffusion. After a series of derivation, we could reach a maneuverable solution by using neural network. Overall, the spatio-temporal diffusion enables the information of previous frames to be recursively transferred to the current frame, which endows the model good ability to capture temporal dependencies, and can be robust to video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation and prediction. In addition, we extend our STDD to text-to-video generation to explore the video generation capabilities.", + "bbox": [ + 89, + 90, + 483, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are three folds: i) propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes; ii) theoretically derive probabilistic reverse process and propose an accelerated sampling method; iii) achieve more competitive performance on video generation.", + "bbox": [ + 89, + 407, + 483, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 527, + 241, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Probabilistic Models", + "text_level": 1, + "bbox": [ + 89, + 551, + 359, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DPM, which was first introduced in reference [32], comprises a diffusion (encoding) process and a denoising (decoding) process. In the diffusion process, given one image $x$ , the noise diffusion process is conducted, which random noises are gradually added by T-step Markov chain [15], denoted as $(x_0,x_1,\\dots ,x_T)$ with", + "bbox": [ + 89, + 574, + 483, + 665 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\\\ \\text {s . t .} \\quad \\bar {\\alpha} _ {t} = \\prod_ {k = 1} ^ {t} \\alpha_ {t}, \\epsilon_ {t} \\sim \\mathcal {N} (0, I), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 676, + 480, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\in (0,1)$ is the corresponding diffusion coefficient. If $T$ is large enough, such as $T = 1000$ , $x_{T}$ approximates random Gaussian noise. Therefore, the reverse process of the diffusion model (denoising process) can be interpreted as the gradual generation of the data distribution from random Gaussian noise. In other words, it aims to recover the original data by removing the added noise step by step.", + "bbox": [ + 89, + 750, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ho et al. [12] establish a connection between DPM and denoising score matching [34] and propose an $\\epsilon$ -prediction formulation for the denoising process, which is $L = ||\\epsilon_t -$", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\epsilon_{\\theta}(x_t,t)||^2$ . In which, $\\epsilon_{\\theta}$ is a denoising neural network parameterized by $\\theta$ , and $L$ is the loss function. Utilizing this formulation, DPM has been successfully applied to various generative tasks, including image generation [11, 19, 27], super-resolution [16], image translation [30], and more [43, 44]. Compared to other generation methods, such as Generative Adversarial Networks (GANs) [8], DPMs are easier to train and can generate more diverse samples [4]. DPMs utilize diffusion processes and de-noising steps, resulting in a more stable training process and improved sample diversity. These characteristics have contributed to the popularity and success of DPMs in the field of generative modeling.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 301, + 730, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video generation is a highly challenging task in the field of generative research. It not only requires the generation of high-quality frames but also necessitates temporal correlation among the generated frames. Previous video generation methods have predominantly relied on Generative Adversarial Networks (GANs). In VGAN [40], the generator directly learns the joint distribution of video frames. Mocogan [36] decomposed a video clip into content and motion components, which are individually modeled using a content vector and a motion vector. Alternatively, other methods utilize a two-step process, which train a vector quantized autoencoder [5] for video data, and then employ an transformer network to learn the video distribution in the quantized latent space [7, 14, 48]. These techniques leverage the quantized latent space to model the video distribution and capture dependencies among frames. Overall, each approach has its own advantages and limitations in generating high-quality continuous videos.", + "bbox": [ + 511, + 324, + 906, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, researchers have been motivated by the remarkable achievements of Diffusion Probabilistic Models (DPM) in image generation and have made efforts to apply DPM to video generation. In reference [13], Ho et al. proposed a video diffusion model that extends the 2D denoising network from the image diffusion model to 3D by considering stacked frames as additional dimensions. This enables the model to capture spatio-temporal dependencies and generate coherent videos. In [39], DPM is utilized for video prediction and interpolation, where known frames are used as conditions for denoising. Some large video generation models [2, 18, 45, 46], with massive data and parameters, allow the attention layer to implicitly learn the relationship between video frames, which requires significant computing and storage resources. These methods often treat video frames as independent samples during the diffusion process. To establish a relation between the video frames, VideoFusion [20] breaks down the noise into the base noise for video sharing and the residual noise for the frames. Benefit from powerful time modeling capabilities,", + "bbox": [ + 511, + 599, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12576", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "transformer-based method [6, 21] is also used to generate high quality video. Still, these methods fail to establish temporal associations between frames, especially adjacent frames, may lead to limitations in generating videos with smooth transitions and preserving temporal consistency.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Spatio-Temporal Dual Diffusion", + "text_level": 1, + "bbox": [ + 89, + 179, + 385, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we will illustrate our main idea, and then derive an analysable spatio-temporal forward and inverse diffusion process, meantime design an accelerated inverse diffusion version and its optimization process.", + "bbox": [ + 89, + 205, + 483, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Description", + "text_level": 1, + "bbox": [ + 89, + 273, + 287, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $X = (x^0, x^1, \\dots, x^i, \\dots, x^{N-1})$ denotes a video sequence with $N$ frames, where the superscript $i$ denotes the $i$ -th video frame. We also follow the customary use of taking subscript as diffusion step, i.e., the symbol $x_t^i$ with the diffusion step $t$ of the $i$ -th frame. The existing video diffusion methods usually take an independent frame-wise diffusion way through leveraging the classic diffusion model such as denoising diffusion probabilistic model [13]. Given one frame $i$ , the noise diffusion process is conducted, denoted as $(x_0^i, x_1^i, \\dots, x_T^i)$ with", + "bbox": [ + 88, + 295, + 483, + 449 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} ^ {i} = \\sqrt {\\alpha_ {t}} x _ {t - 1} ^ {i} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t} ^ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 457, + 482, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_t^i\\sim \\mathcal{N}(0,1)$ $\\alpha_{t} = 1 - \\beta_{t}$ with $\\beta_{t} > 0$ . The inverse denoising process could be recursively estimated by an optimized neural network $\\pmb{\\epsilon}$ , formally,", + "bbox": [ + 89, + 484, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} ^ {i} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} ^ {i} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t} ^ {i}, t\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 541, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\Pi_{s=1}^t\\alpha_s$ , and the initial noisy $x_T^i$ of each frame is also sampled independently from Gaussian noise. In essence, the previous video diffusion methods basically perform a combination of $n$ -frame independent image diffusion processes, although it is usually operated in the hidden feature space in order to enhance the inter-frame consistency. However, the inherent frame-sequential diffusion correlations have not yet been explicitly considered by previous diffusion models.", + "bbox": [ + 89, + 584, + 483, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we consider that the diffusion process should not only be imposed on one single image frame, but only be run on the stream of video frames, named joint spatial and temporal diffusion. From the view of mathematical formulation, we abstract the spatio-temporal dual diffusion as:", + "bbox": [ + 89, + 720, + 483, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} ^ {i} = \\mathbb {D} \\left(x _ {t - 1} ^ {i}, x _ {t - 1} ^ {i - 1}, \\epsilon_ {t} ^ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 805, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} ^ {i} = \\mathbb {D} ^ {- 1} \\left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}, \\epsilon_ {\\theta} \\left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 825, + 482, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the forward and inverse diffusion processes, $\\mathbb{D}$ and $\\mathbb{D}^{-1}$ , are well derived in analyzed mathematical expressions (see below sections), and $\\epsilon$ is a learnable function with the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "parameter $\\theta$ . The STDD essentially represents the dependant relations of continuous frame slices as well as the diffusion steps.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Forward Dual Diffusion", + "text_level": 1, + "bbox": [ + 511, + 145, + 735, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to model the relations between video frames, we also use temporal Markov chain to describe the evolution dependency on time series, in addition to image diffusion process itself. Hereby, a spatio-temporal diffusion model could be built. To distinguish it from standard diffusion, let $z_{t}^{i}$ denotes the noised latent variable of $x^{i}$ at the diffusion step $t$ . According to the diffusion mechanism, we define the forward spatio-temporal noising process as", + "bbox": [ + 511, + 167, + 906, + 289 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t} ^ {i} \\sim p \\left(z _ {t} ^ {i} \\mid z _ {t - 1} ^ {i}, z _ {t - 1} ^ {i - 1}\\right) \\tag {6} \\\\ = \\sqrt {\\alpha_ {t}} (\\underbrace {\\sqrt {\\rho_ {t}} * z _ {t - 1} ^ {i} + \\sqrt {1 - \\rho_ {t}} * z _ {t - 1} ^ {i - 1}} _ {\\text {t e m p o r a l d o m a i n}}) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t} ^ {i} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 297, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the diffusion coefficient $\\rho_t \\in (0,1]$ is used to balance the prior frame and the current frame. In the above equation, the temporal-domain term denotes the temporal diffusion between the $i$ -th and $(i - 1)$ -th frames, and $\\sqrt{\\alpha_t(\\cdot)} + \\sqrt{1 - \\alpha_t} \\epsilon_t^i$ is an image-like diffusion. In order to conveniently operate the noising process, we derive an analyzed expression as shown in the following Proposition 1.", + "bbox": [ + 511, + 369, + 905, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. According to the recursive formula (6), and suppose $\\rho_0 = \\rho_1 = \\dots = \\rho_t = \\rho$ , we compute the diffusion state of any time frame and any diffusion step from the initial states ( $z_0^0 = x^0, z_0^1 = x^1, \\dots, z_0^i = x^i$ ):", + "bbox": [ + 511, + 486, + 906, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t} ^ {i} \\sim p \\big (z _ {t} ^ {i} | \\{z _ {0} ^ {k} | k = 1, 2,.., i \\} \\big) \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} g (t, i) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 556, + 905, + 595 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns. t. g (t, i) = \\sum_ {m = 0} ^ {m i n (t, i)} {\\binom {t} {m}} \\rho^ {(t - m) / 2} (1 - \\rho) ^ {m / 2} z _ {0} ^ {i - m},\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 599, + 887, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we only retain the existing terms $z_0^{i - m}$ w.r.t $m \\leq i$ in $g(t,i)$ .", + "bbox": [ + 511, + 652, + 905, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. The derivation process is deferred to the supplementary material.", + "bbox": [ + 511, + 694, + 905, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Proposition 1, the dual diffusion process integrates the information of time slices and noising steps jointly. The historical frame diffusion term $g(t,i)$ accumulates the information of previous historical frames.", + "bbox": [ + 511, + 734, + 905, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that, the standard diffusion process is a special case of the above model when setting $\\rho = 1$ . For the initial frame, i.e., $i = 0$ , we consider it as a standard image diffusion process, i.e., $z_{t}^{0} = \\sqrt{\\bar{\\alpha}_{t}} z_{0}^{0} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{t}^{0}$ . The probability analysis of the inverse spatio-temporal diffusion is given in Section 3.3, and a practical inverse diffusion can be found directly in Section 3.4.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12577", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In intuitive, it seems require a high-computational burden, but in practice the combinatorial coefficient term $\\left\\{\\binom{t}{m}\\right\\}$ could be pre-calculated as a searching table, which would reduce the running cost. In addition, if $1 - \\rho = < 0.01$ , then $\\rho^{(t - m / 2)} \\approx e^{-t(1 - \\rho) / 2}$ . Therefore, $g(t,i)$ also can be written as:", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (t, i) = \\sum_ {m = 0} ^ {m i n (t, i)} \\binom {t} {m} (1 - \\rho) ^ {m / 2} e ^ {- t (1 - \\rho) / 2} z _ {0} ^ {i - m} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 189, + 483, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ is limited to [0, 16], which require small running cost. Notice, the above coefficient substitution is only for the purpose of simplifying the calculation in the experiment and is independent of our model. And the setting of the parameter $\\rho$ will be discussed in the ablation experiment.", + "bbox": [ + 89, + 241, + 483, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inverse Dual Diffusion", + "text_level": 1, + "bbox": [ + 89, + 325, + 302, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Probability diffusion process can be reduced to a constrained latent variable model whose Gaussian distribution of forward and reverse processes must be modeled. To derive the inverse decoding process, we establish an explicit link between spatio-temporal diffusion probabilistic models and inverse probability distributions, and then provide a variational optimization objective for the proposed dual diffusion model.", + "bbox": [ + 89, + 348, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $Z_{t}$ denotes a set of all frames at the diffusion step $t$ , i.e., $Z_{t} = (z_{t}^{0}, z_{t}^{1}, \\dots, z_{t}^{N-1})$ . For convenient statement, we use a surrogate operator $F_{t}: Z_{0} \\to \\widetilde{Z}_{t}$ and its inverse operator $F_{t}^{-1}: \\widetilde{Z}_{t} \\to Z_{0}$ , in which $F_{t}$ as well as $\\widetilde{Z}_{t}$ is defined as:", + "bbox": [ + 89, + 468, + 483, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {Z} _ {t} = F _ {t} \\left(Z _ {0}\\right) := \\left(\\widetilde {z} _ {t} ^ {0}, \\dots , \\widetilde {z} _ {t} ^ {i}, \\dots , \\widetilde {z} _ {t} ^ {N - 1}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 554, + 482, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .}, \\quad \\widetilde {z} _ {t} ^ {i} = g (t, i). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 577, + 480, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$F_{t}$ represents the diffusion process between video frames, which is called temporal diffusion. Then, according to Eqn. (6) and Proposition 1, the variable $Z_{t}$ can be written as", + "bbox": [ + 89, + 603, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} F _ {t} \\left(Z _ {0}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {11} \\\\ = \\sqrt {\\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t}, \\qquad (1 2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 657, + 482, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the noise $\\epsilon_{t} = (\\epsilon_{t}^{0},\\epsilon_{t}^{1},\\dots,\\epsilon_{t}^{i})$ , and Eqn. (12) is the iterative formula and the Markov chain of the forward diffusion process. Therefore, the conditional probability distribution of the forward diffusion process can be written as a Gaussian distribution:", + "bbox": [ + 89, + 705, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q (Z _ {t} | Z _ {t - 1}, Z _ {0}) = \\mathcal {N} (Z _ {t} | \\sqrt {\\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) (1 - \\alpha_ {t}) I), \\\\ q \\left(Z _ {t} \\mid Z _ {0}\\right) = \\mathcal {N} \\left(Z _ {t} \\mid \\sqrt {\\bar {\\alpha} _ {t}} F _ {t} \\left(Z _ {0}\\right), (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 791, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, video diffusion model can be modeled as some latent variables with the form: $p_{\\theta}(Z_0) = \\int p_{\\theta}(Z_{0:T})d_{Z_{1:T}}$ , where the subscript $\\square_{1:T}$ means the collection of latent variables from step 1 to step $T$ and $\\theta$ is the learnt parameter. The", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "joint distribution $p_{\\theta}(Z_{0:T})$ refers to the reverse process, and it can be defined as a Markov chain with learned Gaussian transitions starting at $p(Z_T) = \\mathcal{N}(0,I)$ . Formally,", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (Z _ {0: T}) = p (Z _ {T}) \\Pi_ {t = 1} ^ {T} p _ {\\theta} (Z _ {t - 1} | Z _ {t}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 147, + 906, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we need to derive $p_{\\theta}(Z_{t - 1}|Z_t)$ as forward process posteriors, which are tractable when conditioned on $Z_0$ :", + "bbox": [ + 511, + 176, + 906, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(Z _ {t - 1} \\mid Z _ {t}\\right) = q \\left(Z _ {t - 1} \\mid Z _ {t}, Z _ {0}\\right) = \\mathcal {N} \\left(\\tilde {\\mu_ {t}}, \\tilde {\\beta_ {t}}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 231, + 906, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we assume the Gaussian distribution in the last equation. Then, with the help of Eqn. (13), we obtain the mean and variance of the posterior distribution of the diffusion process $q(Z_{t - 1}|Z_t) = N(\\tilde{\\mu}_t,\\tilde{\\beta}_t)$ :", + "bbox": [ + 511, + 261, + 905, + 323 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\tilde {\\mu} _ {t} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(F _ {t - 1} \\left(F _ {t} ^ {- 1} \\left(Z _ {t}\\right)\\right) - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {t}\\right), \\\\ \\tilde {\\beta} _ {t} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} (1 - \\alpha_ {t}), \\end{array} \\right. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 333, + 905, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the relevant proof is given in the supplement material.", + "bbox": [ + 511, + 409, + 903, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Accelerate Inverse Inference Process", + "text_level": 1, + "bbox": [ + 511, + 433, + 830, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to boost sampling speed, we derive an implicit probability diffusion model. The denoising diffusion probabilistic process, in section 3.3, requires the construction of Markov chains with the help of the additional operator $F^{-1}$ , and more computational complexity. But, denoising diffusion implicit models has faster sampling speeds and does not require Markov chains. Below we do a derivation to fulfill the spatio-temporal inverse diffusion process with DDIM, which enables more practical use.", + "bbox": [ + 511, + 455, + 906, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, according to Proposition 1, we make reasonable assumptions for the reverse process", + "bbox": [ + 511, + 590, + 905, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} ^ {i} = k _ {t} z _ {t} ^ {i} + C _ {t} ^ {i} \\left(Z _ {0} ^ {0: i - 1}\\right) + \\sigma_ {t} \\epsilon_ {t - 1}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 633, + 903, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{t}$ and $\\sigma_{t}$ are coefficient to be determined, $C_t^i (Z_0^{0:i - 1})$ is a function (also written as $C_t^i$ ) and $\\epsilon^{\\prime}\\sim \\mathcal{N}(0,I)$ . Then, according to Proposition 1 and Eqn. (7), we replace $z_{t}^{i}$ with an expression of $Z_{0}$ and $\\epsilon_t^i$ :", + "bbox": [ + 511, + 657, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} ^ {i} = k _ {t} \\sqrt {\\bar {\\alpha} _ {t}} g (t, i) + C _ {t} ^ {i} + \\sqrt {k _ {t} ^ {2} \\left(1 - \\bar {\\alpha} _ {t}\\right) + \\sigma_ {t} ^ {2}} \\epsilon \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 731, + 903, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with Eqn. (7) in Proposition 1 (i.e., $\\sqrt{\\bar{\\alpha}_{t-1}} g(t-1,i)+\\sqrt{1-\\bar{\\alpha}_{t-1}}\\epsilon$ ), we can solve the coefficients $k_{t}$ and function $C_{t}^{i}(Z_{0}^{0:i-1})$ with conditional of $\\sigma_{t}$ :", + "bbox": [ + 511, + 766, + 905, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} k _ {t} = \\sqrt {\\frac {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}}{1 - \\bar {\\alpha} _ {t}}}, \\\\ C _ {t} ^ {i} \\left(Z _ {0} ^ {0: i - 1}\\right) = \\sqrt {\\bar {\\alpha} _ {t - 1}} g (t - 1, i) - k _ {t} \\sqrt {\\bar {\\alpha} _ {t}} g (t, i). \\end{array} \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 834, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12578", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 STDD Training", + "bbox": [ + 91, + 90, + 287, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: repeat", + "2: $Z_{0} \\sim q(Z_{0})$", + "3: $t \\sim \\text{Uniform}(1, \\dots, T)$", + "4: $\\epsilon \\sim \\mathcal{N}(0,I)$", + "5: $Z = \\sqrt{\\bar{\\alpha}_t} F_t(Z_0) + \\sqrt{1 - \\bar{\\alpha}_t} F_t(\\epsilon)$", + "6: Take gradient descent step on", + "7: $\\nabla_{\\theta}\\mathcal{L}(\\theta)$ according to Eqn. (23)", + "8: until converged" + ], + "bbox": [ + 101, + 111, + 370, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 STDD Sampling", + "bbox": [ + 91, + 252, + 292, + 268 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $\\epsilon \\sim \\mathcal{N}(0,I)$", + "2: $Z_{T} = F_{T}(\\epsilon)$", + "3: for $t = T$ to 1 do", + "4: $\\epsilon_{\\theta} = \\epsilon_{\\theta}(Z_t,t),\\hat{z}_0^i = \\phi_\\theta (z_t^i,t)$", + "5: Calculate $Z_{t-1}$ according to Eqn. (20)", + "6: end for", + "7: return $Z_0$" + ], + "bbox": [ + 101, + 272, + 397, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further bring the solved parameters $k_{t}$ and $C_t^i$ into Eqn. (18) to obtain the accelerated inverse inference process of STDD, with setting $\\epsilon_t^i = \\epsilon_\\theta^i$ , expressed as:", + "bbox": [ + 89, + 409, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} ^ {i} = \\sqrt {\\bar {\\alpha} _ {t - 1}} (u _ {t} ^ {i} + v _ {t} ^ {i}) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} ^ {i} + \\sigma_ {t} \\epsilon^ {\\prime}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 469, + 462, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "s.t., $u_{t}^{i} = \\frac{z_{t}^{i} - \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{\\theta}^{i}}{\\sqrt{\\bar{\\alpha}_{t}}},\\quad \\epsilon_{\\theta}^{i} = \\epsilon_{\\theta}(z_{t}^{i},t),$ (20)", + "bbox": [ + 114, + 498, + 480, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} ^ {i} = g (t - 1, i) - g (t, i), \\quad \\tilde {z} _ {0} ^ {i} = \\phi_ {\\theta} (z _ {t} ^ {i}, t),\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 532, + 444, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\mu}_t^i = \\sqrt{\\bar{\\alpha}_t} * u_t^i$ is the mean of $z_{t}^{i}$ , $v_{t}^{i} \\approx F_{t-1}((1 - \\sqrt{\\rho})\\tilde{z}_{0}^{i} - \\sqrt{1 - \\rho}\\tilde{z}_{0}^{i-1})$ is temporal domain difference, $\\epsilon'$ is the Gaussian noise added during inference and $\\phi_{\\theta}$ is a network, then $v_{t}^{i}$ can be calculated by Eqn. (7) with $\\tilde{z}_{0}^{i}$ . In practice, $\\epsilon_{\\theta}$ and $\\phi_{\\theta}$ are different channel components of the same network output. Moreover, in the process of inference, there is no calculation of the inverse function of $F$ in Eqn. (16).", + "bbox": [ + 89, + 561, + 483, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "How to understand $u$ and $v$ : Assuming a new variable $\\tau$ , rewrite $z_{t}^{i}$ in Eqn. (7) as: $z^{i}(t,\\tau) = \\sqrt{\\bar{\\alpha}_{t}} g(\\tau ,i) + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}^{i}$ . Then, according to SDE model [34], we conclude that", + "bbox": [ + 89, + 685, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} d _ {t} z ^ {i} (t, \\tau) \\approx - 1 / 2 \\beta_ {t} u _ {t} ^ {i} + \\sqrt {\\beta_ {t}} d \\omega \\\\ d _ {\\tau} z ^ {i} (t, \\tau) \\approx - \\sqrt {\\bar {\\alpha} _ {t}} v _ {\\tau} ^ {i} \\end{array} , \\right. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 755, + 482, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega$ gaussian component. Therefore, $u_{t}^{i}$ is image spatial diffusion component and $v_{t}^{i}$ a temporal diffusion or bridge between frames. Therefore, the process inference consists of a diffusion model (i.e., Eqn. (20)) and a difference between frames, shown in Algorithm 2. In addition, to ensure the association between frames, we set $Z_{T} = F(\\epsilon)$ .", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optimize Dual Diffusion", + "text_level": 1, + "bbox": [ + 513, + 90, + 736, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dual diffusion architecture includes Variational AutoEncoder (VAE) for video encoding and decoding, diffusion networks $\\epsilon_{\\theta}$ for denoising, and conditional embedding networks (if needed). In detail, we employ VAE model training in an unsupervised manner, contains two trainable modules: encoder $\\mathcal{E}$ to represent image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ as latent map $z \\in \\mathbb{R}^{h \\times w \\times c}$ and decoder $\\mathcal{D}$ to decode the latent map $\\tilde{Z}$ as output image $\\tilde{X}$ . VAE was trained with the reconstruction loss [28] using only unlabeled images, or load pre-trained model such as Stable Diffusion v1.4 [28]. Then, we use transformer network, used in [21], to extract the feature of video and learn the noise of dual diffusion process. And time step $t$ is specified to $\\epsilon_{\\theta}$ by the sinusoidal position embedding [38]. In addition, we enhance the network backbone by using a cross-attention mechanism [28] to transform the model into a more flexible conditional diffusion model. And, we adopt different pre-training condition networks according to different tasks, such as CLIP [26] for text-to-video generation. Implementation details will be presented in Sec. 4.1.", + "bbox": [ + 511, + 112, + 906, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For optimizing STDD, the KL divergence to directly compare $p_{\\theta}(Z_{t - 1}|Z_t)$ against forward process posteriors is:", + "bbox": [ + 511, + 415, + 905, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {K L} \\left(q \\left(Z _ {t - 1} \\mid Z _ {t}\\right) | | p _ {\\theta} \\left(Z _ {t - 1} \\mid Z _ {t}\\right)\\right) \\\\ = \\frac {\\left| \\left| \\tilde {\\mu} _ {t - 1} - \\mu_ {\\theta} \\right| \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}} + C \\\\ = \\frac {\\left| \\left| \\tilde {\\mu} _ {t - 1} - \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(U _ {t} + V _ {t}\\right) \\right| \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}} + C \\tag {22} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 473, + 903, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is a constant, $U_{t} = (u_{t}^{0}, u_{t}^{1}, \\dots, u_{t}^{N - 1})$ and $V_{t} = (v_{t}^{0}, v_{t}^{1}, \\dots, v_{t}^{N - 1})$ . We further optimize $\\epsilon_{\\theta}$ and $\\phi_{\\theta}$ respectively, thus the loss of the spatio-temporal diffusion probability model is", + "bbox": [ + 511, + 574, + 905, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k l} (\\theta) = \\mathbb {E} | | \\tilde {\\mu} _ {t - 1} / \\sqrt {\\bar {\\alpha} _ {t - 1}} - (U _ {t} + V _ {t}) | | ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 647, + 812, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} (\\theta) = \\mathbb {E} | | \\boldsymbol {\\epsilon} _ {t} - \\epsilon_ {\\theta} (Z _ {t}, t) | | ^ {2} + | | Z _ {0} - \\phi_ {\\theta} (Z _ {t}, t) | | ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 667, + 887, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathcal {L} _ {k l} (\\theta) + \\mathcal {L} _ {m s e} (\\theta), \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 688, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ represents the parameters of the STDD network. The detailed training process is shown in Algorithm 1.", + "bbox": [ + 511, + 717, + 903, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 762, + 645, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 787, + 700, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: To validate the proposed STDD, we conduct comprehensive evaluations and comparisons on Sky Time-Lapse [47] and UCF101 [35] for video generation, MHAD human action dataset [3] for image-to-video and WebVid-10M [1] for text-to-video. UCF101 is a realistic action video motion recognition dataset, collected from", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12579", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcb21df5221cc789563342326782cc193ab29d704050cefbe5c6a3a8e4bff9b3.jpg", + "image_caption": [ + "Figure 1. Qualitative comparisons on MHAD for video prediction. In the red box, the head in LFDM has serious artifacts and distortions, which we handled well." + ], + "image_footnote": [], + "bbox": [ + 101, + 94, + 614, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58e11eaf4822dda8bb2198b2f70b74b926c66bc78279facb91152d1dc466440d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 95, + 890, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YouTube, providing 13,320 videos from 101 action categories. MHAD human action dataset contains 861 videos of 27 actions performed by 8 subjects. UCF101, Sky Time-Lapse and MHAD are cropped to 16 frames for training our model. For quantitative evaluation, we also train a text-conditioned video generation model on WebVid-10M [1], which consists of 10.7M short videos with paired textual descriptions.", + "bbox": [ + 88, + 426, + 480, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics: Following previous works [7, 20], we mainly use Frechet Video Distance (FVD) [37] as the evaluation metric. On UCF101 and Sky Time-Lapse, we report the results of FVD. Following LFDM [22] on MHAD, we report the results of FVD and class conditional FVD (cFVD). The cFVD compare the distance between real and synthesized video feature distributions under the same class condition.", + "bbox": [ + 88, + 547, + 482, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details: We use the transformer network, used in [21], to extract the features of video and learn the noise of the dual diffusion process. The network learning rate is $1e - 4$ , the batch size is 16, and the training stpes is 200000. The diffusion steps $T = 1000$ , the steps of accelerate reverse process is 250 and $\\rho = 1 - 10^{-3}$ . We will provide the generated videos and code on github.", + "bbox": [ + 89, + 654, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with state-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 771, + 393, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video prediction. On MHAD dataset, we have conducted a comparison of our methods with several competitive approaches, including ImaGINator [41], VDM [13], LDM [28] and LFDM [22]. Table 1 shows the quantitative comparison between our method and state-of-the-art approaches like LDM and LFDM, where our proposed STDD consistently outperforms all other models. The results show", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4aa0c52107601e0e25252c6bba214920cc5dfd6f158750d161048b279fdff57d.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on MHAD. $\\downarrow$ denotes the lower the better. The best results are denoted in bold." + ], + "table_footnote": [], + "table_body": "
MethodResolutionFVD ↓cFVD↓
ImaGINator [41]64 × 648891406
VDM [13]64 × 64295531
LDM [28]128 × 128337594
LFDM [22]128 × 128214426
Ours128 × 128195411
", + "bbox": [ + 566, + 464, + 854, + 555 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/279dbd973333876685eb77f8cde9c79def9e8efad2693ab533409a9dc5ba8c4f.jpg", + "table_caption": [ + "Table 2. FVD values of video generation models on UCF101 and Sky Time-Lapse. The lower the better and the best results are denoted in bold." + ], + "table_footnote": [], + "table_body": "
MethodUCF101Sky Time-LapseParams
VideoGPT [48]2880222-
MoCoGAN-HD [36]1729164-
DIGAN [50]163083.1-
StyleGAN-V [31]143179.5-
PVDM [51]114175.5-
LVDM [10]37295.2437M
Latte+IMG [21]33342.7673M
Ours25842.1868M
", + "bbox": [ + 539, + 625, + 880, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that our STDD is superior to LFDM (195 vs 214 in FVD at the resolution of $(128 \\times 128)$ , and LFDM achieves the second best performance. LFDM distorts the initial frame through an optical flow sequence to generate a new frame to model the time correlation, while our method is mainly to explicitly model the frame time correlation, which is conducive to the propagation of historical frame information.", + "bbox": [ + 511, + 777, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video generation. On UCF101 dataset, we have con", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12580", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f910ceb3c9c5ee772d41fe9bd42edbc27e71b3d07b5b2dbd3f096af635b95286.jpg", + "image_caption": [ + "(a)UCF101 (Playing Sitar)" + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 504, + 344 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/789721eea31130d61951b16273775652ed01fc38388515ed40e11a30698931f2.jpg", + "image_caption": [ + "(b) Sky Time-Lapse" + ], + "image_footnote": [], + "bbox": [ + 504, + 93, + 893, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duced a comparison of our methods with several competitive approaches, including VideoGPT [48], StyleGAN [31], DIGAN [50], TATS [7], VDM [13], among others. The quantitative results on UCF101 dataset are presented in Table 2. In terms of video generation, VDM demonstrates superior performance compared to GAN-based methods, particularly in terms of FVD (Fréchet Video Distance). This indicates the potential of DPM-based video generation methods. Our proposed STDD achieves even better results than LVDM by a significant margin. Compared with Latte-IMG, which prerained on images, our approach is also comparable. In addition, the quantitative results in Sky Time-Lapse also achieve competitive performance, shown in Table 2. These comparisons highlight the effectiveness of STDD in video generation tasks, demonstrating its superior performance over existing methods in terms of various evaluation metrics and datasets. In addition, the inference time of our model did not increase significantly with the same inference steps and device (23s vs 20s of Latte with 250 inference steps in RTX 3090 GPU).", + "bbox": [ + 89, + 416, + 483, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 727, + 243, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual comparison. On MHAD dataset, we have conducted a qualitative comparison of our methods with several video pridiction models, including ImaGINator, VDM, LDM and LFDM. As shown in Fig. 1, the first line shows the given image, and the rest shows the generated sequence. In general, our approach has advantages over other video generation methods, especially generating 'stand-to-sit' videos with larger variation areas and more motion variations. The results in Fig. 1 (b) showed that the head of the LFDM generated character had severe artifacts and distor", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/282f016cbaed3a84d1927f8c2986f86f87d239f20bac1550bb42b9c4eadd8cf6.jpg", + "image_caption": [ + "Figure 2. Qualitative comparisons on UCF101 and Sky Time-Lapse for video generation." + ], + "image_footnote": [], + "bbox": [ + 517, + 417, + 771, + 512 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2930b40705e576cbe1c2b6cfbc196057282fd4967b0b0b412e04a257e7cb6fdc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 417, + 895, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b99d0f0475969e6f631bef24383dcd4baee9ffef2f91932f00092101dc6f537.jpg", + "image_caption": [ + "Frame weights" + ], + "image_footnote": [], + "bbox": [ + 516, + 515, + 771, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/627fdcb7974ffff42fd86ae3f52197e434949a688392dd22bd56d66c0bd5cf8a.jpg", + "image_caption": [ + "$g(t = 1000,i)$", + "Figure 3. The tendency of frame weights in Eqn. (7) and Eqn. (8). Appropriate $\\rho$ enhances the similarities and blurring the differences between frames." + ], + "image_footnote": [], + "bbox": [ + 776, + 516, + 895, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions, which we handled well, notably in the red box. We also provide visual comparisons with the most recent video generation methods, i.e. TATS, DIGAN, VideoFusion and Latte. As shown in Fig. 2, each generated video has 16 frames with a resolution of $256 \\times 256$ , and we show 4-th, 8-th, 12-th and 16-th frame. Our STDD could generate more continuous and moving videos.", + "bbox": [ + 511, + 717, + 906, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of $\\rho$ . To explore the effects of $\\rho$ in Eqn (7), we conducted experiments on MHAD with a resolution of $64 \\times 64$ , shown in Table 3. Appropriate $\\rho$ can better integrate the historical frames and improve the performance. Also, smaller $\\rho$ leads to information loss of the", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12581", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68f7ede029e27228502f78c70961adb77cb65b26b979118c898af2973448436d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 155, + 93, + 841, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d49240d508471c2f7c42b569192af13a89ce2490fa7777734399179096e66f3a.jpg", + "image_caption": [ + "(a) A dog in astronaut suit and sunglasses floating in space", + "(b) A turtle swimming in the ocean" + ], + "image_footnote": [], + "bbox": [ + 155, + 306, + 504, + 438 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7745bd384cfb82ad6282abda5bc515a93654ec0209e99ae37a89d4394dd9c522.jpg", + "image_caption": [ + "(c) Surfer on his surfboard in a wave", + "Figure 4. Qualitative evaluation on text-to-video generation. Both our method and Videofusion are trained on WebVid-10M, while Latte is trained on Vimeo25M [42]." + ], + "image_footnote": [], + "bbox": [ + 504, + 306, + 838, + 438 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69296a2f60625c509ed84175ae1a3d2b0ffef4dd10890bed42eb5cf8aa31471e.jpg", + "table_caption": [ + "Table 3. Analysis $\\rho$ in Eqn. (7) on MHAD with the resolution of $64\\times 64$ . Appropriate $\\rho$ can improve the performance and $\\sqrt{\\rho^{1000}}$ represents the weight of the current frame after 1000 steps." + ], + "table_footnote": [], + "table_body": "
ρ10.99990.9990.9950.99
\\( \\sqrt{\\rho^{1000}} \\)10.9510.6060.0820.007
FVD ↓157149133243388
", + "bbox": [ + 140, + 575, + 434, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "current frame, which is not conducive to video generation. Since the coefficient is exponential, it must be required that $0 < 1 - \\rho < 0.01$ , which consistent with the hypothesis in Eqn. (8). The weights of the current and historical frames are shown in Fig. 3. Visualization shows that $g(t,i)$ enhances the similarities and blurring the differences between frames.", + "bbox": [ + 89, + 691, + 482, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-video generation. To explore the video generation capabilities of our proposed approach, we extend STDD to text-to-video generation. We train a text-to-video model, pretrained by Stable Diffusion [28], on the large scale video dataset, i.e. WebVid-10M. Some samples are shown in Fig. 4.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 523, + 633, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes. Then, we theoretically derive probabilistic reverse process and propose an accelerated sampling method. Both qualitative and quantitative experiments demonstrated that our model achieve more competitive performance on video generation.", + "bbox": [ + 511, + 549, + 906, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 667, + 609, + 683 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 6", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[3] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 5" + ], + "bbox": [ + 521, + 691, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12582", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2", + "[6] Yu Gao, Jiancheng Huang, Xiaopeng Sun, Zequn Jie, Yujie Zhong, and Lin Ma. Matten: Video generation with mamba-attention. arXiv preprint arXiv:2405.03025, 2024. 3", + "[7] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In European Conference on Computer Vision, pages 102–118. Springer, 2022. 2, 6, 7", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[9] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 1", + "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.6", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 1, 2, 3, 6, 7", + "[14] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[15] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2", + "[16] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2", + "[17] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 1", + "[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2", + "[20] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. CoRR, abs/2303.08320, 2023. 1, 2, 6", + "[21] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 5, 6", + "[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18444-18455. IEEE, 2023. 6", + "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1", + "[25] Xichen Pan, Pengda Qin, Yuhong Li, Hui Xue, and Wenhu Chen. Synthesizing coherent story with auto-regressive latent diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2920-2930, 2024. 1", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5, 6, 8", + "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12583", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2", + "[31] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 6, 7", + "[32] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2", + "[34] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2, 5", + "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 5", + "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2, 6", + "[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems, 35:23371-23385, 2022. 1, 2", + "[40] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2", + "[41] Yaohui Wang, Piotr Bilinski, Francois Bremond, and Antitza Dantcheva. Imaginator: Conditional spatio-temporal gan for video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1160-1169, 2020. 6", + "[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 8", + "[43] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 22025-22034, 2023. 2", + "[44] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023. 2", + "[45] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024. 2", + "[46] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2", + "[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5", + "[48] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2, 6, 7", + "[49] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25(10):1469, 2023. 1", + "[50] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 6, 7", + "[51] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 6" + ], + "bbox": [ + 516, + 92, + 903, + 584 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12584", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_model.json b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a215f3a84fcc108948e3bc8ea3b92948fbcaf799 --- /dev/null +++ b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "STDD: Spatio-Temporal Dual Diffusion for Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.18, + 0.782, + 0.287 + ], + "angle": 0, + "content": "Shuaizhen Yao\\(^{1}\\), Xiaoya Zhang\\(^{1,*}\\), Xin Liu\\(^{2}\\), Mengyi Liu\\(^{3}\\), Zhen Cui\\(^{4,*}\\) \n\\(^{1}\\) Nanjing University of Science and Technology, \n\\(^{2}\\) Nanjing Seetacloud Technology Co., Ltd, \n\\(^{3}\\) Kuaishou, \\(^{4}\\) Beijing Normal University \n{ysz, zhangxiaoya}@njust.edu.cn, xin.liu@seetacloud.com, liumengyi@kuaishou.com, zhen.cui@bnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.322, + 0.328, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.486, + 0.686 + ], + "angle": 0, + "content": "Diffusion probabilistic model is becoming the cornerstone of data generation, especially generating high-quality images. As an extension, video diffusion generation is in urgent need of a principled temporal-sequence diffusion way, while the spatial-domain diffusion dominates most video diffusion methods. In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal noise propagation/reduction. Mathematically, an analysable dual diffusion process is derived to accumulate information in temporal sequence as well as spatial domain. Correspondingly, we theoretically derive a spatio-temporal probabilistic reverse diffusion process and propose an accelerated sampling way to reduce the inference cost. In principle, the spatio-temporal dual diffusion enables the information of previous frames to be transferred to the current frame, which thus could be beneficial for video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation/prediction as well as text-to-video generation." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Diffusion probabilistic model (DPM) is becoming the cornerstone of data generation, where the forward process gradually contaminates the data with gaussian noises and the reverse denoising process is learned to generate new samples [12, 15]. Specifically, DPM has made significant progress in producing high quality, realistic images [17, 23, 24, 27, 29, 33]. Motivated by the success in image generation, some researchers are trying to extend the diffusion model from image to video by accompanying with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.384 + ], + "angle": 0, + "content": "temporal-dimensional prediction/interpolation [9, 39, 49]. However, due to intricate temporal and spatial correlations as well as high dimensionality of video, video diffusion generation remains a challenging task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.909, + 0.642 + ], + "angle": 0, + "content": "The existing video generation methods based on DPMs [9, 13, 28] typically employ a standard diffusion process by treating each frame of the video as an independent diffusion process, which breaks the temporal correlation between frames. Concretely, in the noisig/denoising process, these video diffusion methods separately derive/sample priors, and recover corresponding frames from different-random priors. Due to different-random priors, the continuity of generated video frames is still a bottleneck of video diffusion generation, even though some historical (generated) frames could be used as a condition of the diffusion for some-degree mitigation [25]. An endeavor to this aspect is one recent work [20], decomposing video noise into two components: video sharing base noise and frame-independent residual noise, and then separately performing the diffusion process with two different diffusion networks, which greatly improves video consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Despite some progress in video diffusion generation, these existing methods could not well capture temporal dependencies due to the use of independent priors for each frame, which would often result into the nonsmoothness or inconsistency of generated videos, especially confronting with large appearance variations in temporal sequences. Essentially, the current video diffusion methods lack a principled diffusion process along time slices (w.r.t frame sequence), not just in the spatial domain (w.r.t every image)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal frame propagation. Mathematically, we derive an analysable dual diffusion process by accumulating temporal frame information. In particular, the historical frames are propagated along with temporal sequence, while accompanying with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.408 + ], + "angle": 0, + "content": "the spatio-domain noises w.r.t diffusion on single-frame image. To reach the inverse diffusion, we derive a probabilistic reverse process, which guarantees the feasibility of the diffusion model. Inspired by the diffusion probabilistic implicit diffusion (DDIM) model [33], further, we propose an accelerated spatio-temporal sampling method that is both fast and computation-intensive. To facilitate learning, we decompose the model into spatial components representing the standard diffusion model and temporal components representing the frame-to-frame diffusion. After a series of derivation, we could reach a maneuverable solution by using neural network. Overall, the spatio-temporal diffusion enables the information of previous frames to be recursively transferred to the current frame, which endows the model good ability to capture temporal dependencies, and can be robust to video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation and prediction. In addition, we extend our STDD to text-to-video generation to explore the video generation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.513 + ], + "angle": 0, + "content": "In summary, our main contributions are three folds: i) propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes; ii) theoretically derive probabilistic reverse process and propose an accelerated sampling method; iii) achieve more competitive performance on video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.242, + 0.543 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.553, + 0.361, + 0.567 + ], + "angle": 0, + "content": "2.1. Diffusion Probabilistic Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.666 + ], + "angle": 0, + "content": "DPM, which was first introduced in reference [32], comprises a diffusion (encoding) process and a denoising (decoding) process. In the diffusion process, given one image \\(x\\), the noise diffusion process is conducted, which random noises are gradually added by T-step Markov chain [15], denoted as \\((x_0,x_1,\\dots ,x_T)\\) with" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.678, + 0.482, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\\\ \\text {s . t .} \\quad \\bar {\\alpha} _ {t} = \\prod_ {k = 1} ^ {t} \\alpha_ {t}, \\epsilon_ {t} \\sim \\mathcal {N} (0, I), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\in (0,1)\\) is the corresponding diffusion coefficient. If \\(T\\) is large enough, such as \\(T = 1000\\), \\(x_{T}\\) approximates random Gaussian noise. Therefore, the reverse process of the diffusion model (denoising process) can be interpreted as the gradual generation of the data distribution from random Gaussian noise. In other words, it aims to recover the original data by removing the added noise step by step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ho et al. [12] establish a connection between DPM and denoising score matching [34] and propose an \\(\\epsilon\\)-prediction formulation for the denoising process, which is \\(L = ||\\epsilon_t -\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.289 + ], + "angle": 0, + "content": "\\(\\epsilon_{\\theta}(x_t,t)||^2\\). In which, \\(\\epsilon_{\\theta}\\) is a denoising neural network parameterized by \\(\\theta\\), and \\(L\\) is the loss function. Utilizing this formulation, DPM has been successfully applied to various generative tasks, including image generation [11, 19, 27], super-resolution [16], image translation [30], and more [43, 44]. Compared to other generation methods, such as Generative Adversarial Networks (GANs) [8], DPMs are easier to train and can generate more diverse samples [4]. DPMs utilize diffusion processes and de-noising steps, resulting in a more stable training process and improved sample diversity. These characteristics have contributed to the popularity and success of DPMs in the field of generative modeling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.731, + 0.317 + ], + "angle": 0, + "content": "2.2. Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Video generation is a highly challenging task in the field of generative research. It not only requires the generation of high-quality frames but also necessitates temporal correlation among the generated frames. Previous video generation methods have predominantly relied on Generative Adversarial Networks (GANs). In VGAN [40], the generator directly learns the joint distribution of video frames. Mocogan [36] decomposed a video clip into content and motion components, which are individually modeled using a content vector and a motion vector. Alternatively, other methods utilize a two-step process, which train a vector quantized autoencoder [5] for video data, and then employ an transformer network to learn the video distribution in the quantized latent space [7, 14, 48]. These techniques leverage the quantized latent space to model the video distribution and capture dependencies among frames. Overall, each approach has its own advantages and limitations in generating high-quality continuous videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In recent years, researchers have been motivated by the remarkable achievements of Diffusion Probabilistic Models (DPM) in image generation and have made efforts to apply DPM to video generation. In reference [13], Ho et al. proposed a video diffusion model that extends the 2D denoising network from the image diffusion model to 3D by considering stacked frames as additional dimensions. This enables the model to capture spatio-temporal dependencies and generate coherent videos. In [39], DPM is utilized for video prediction and interpolation, where known frames are used as conditions for denoising. Some large video generation models [2, 18, 45, 46], with massive data and parameters, allow the attention layer to implicitly learn the relationship between video frames, which requires significant computing and storage resources. These methods often treat video frames as independent samples during the diffusion process. To establish a relation between the video frames, VideoFusion [20] breaks down the noise into the base noise for video sharing and the residual noise for the frames. Benefit from powerful time modeling capabilities," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "transformer-based method [6, 21] is also used to generate high quality video. Still, these methods fail to establish temporal associations between frames, especially adjacent frames, may lead to limitations in generating videos with smooth transitions and preserving temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.18, + 0.386, + 0.198 + ], + "angle": 0, + "content": "3. Spatio-Temporal Dual Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.267 + ], + "angle": 0, + "content": "In this section, we will illustrate our main idea, and then derive an analysable spatio-temporal forward and inverse diffusion process, meantime design an accelerated inverse diffusion version and its optimization process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.275, + 0.288, + 0.291 + ], + "angle": 0, + "content": "3.1. Problem Description" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.485, + 0.45 + ], + "angle": 0, + "content": "Let \\( X = (x^0, x^1, \\dots, x^i, \\dots, x^{N-1}) \\) denotes a video sequence with \\( N \\) frames, where the superscript \\( i \\) denotes the \\( i \\)-th video frame. We also follow the customary use of taking subscript as diffusion step, i.e., the symbol \\( x_t^i \\) with the diffusion step \\( t \\) of the \\( i \\)-th frame. The existing video diffusion methods usually take an independent frame-wise diffusion way through leveraging the classic diffusion model such as denoising diffusion probabilistic model [13]. Given one frame \\( i \\), the noise diffusion process is conducted, denoted as \\( (x_0^i, x_1^i, \\dots, x_T^i) \\) with" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.458, + 0.483, + 0.477 + ], + "angle": 0, + "content": "\\[\nx _ {t} ^ {i} = \\sqrt {\\alpha_ {t}} x _ {t - 1} ^ {i} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t} ^ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.533 + ], + "angle": 0, + "content": "where \\(\\epsilon_t^i\\sim \\mathcal{N}(0,1)\\) \\(\\alpha_{t} = 1 - \\beta_{t}\\) with \\(\\beta_{t} > 0\\) . The inverse denoising process could be recursively estimated by an optimized neural network \\(\\pmb{\\epsilon}\\) , formally," + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.542, + 0.483, + 0.575 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} ^ {i} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} ^ {i} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t} ^ {i}, t\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.72 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\Pi_{s=1}^t\\alpha_s\\), and the initial noisy \\(x_T^i\\) of each frame is also sampled independently from Gaussian noise. In essence, the previous video diffusion methods basically perform a combination of \\(n\\)-frame independent image diffusion processes, although it is usually operated in the hidden feature space in order to enhance the inter-frame consistency. However, the inherent frame-sequential diffusion correlations have not yet been explicitly considered by previous diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.797 + ], + "angle": 0, + "content": "Here we consider that the diffusion process should not only be imposed on one single image frame, but only be run on the stream of video frames, named joint spatial and temporal diffusion. From the view of mathematical formulation, we abstract the spatio-temporal dual diffusion as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.806, + 0.483, + 0.825 + ], + "angle": 0, + "content": "\\[\nx _ {t} ^ {i} = \\mathbb {D} \\left(x _ {t - 1} ^ {i}, x _ {t - 1} ^ {i - 1}, \\epsilon_ {t} ^ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.827, + 0.483, + 0.846 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} ^ {i} = \\mathbb {D} ^ {- 1} \\left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}, \\epsilon_ {\\theta} \\left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where the forward and inverse diffusion processes, \\(\\mathbb{D}\\) and \\(\\mathbb{D}^{-1}\\), are well derived in analyzed mathematical expressions (see below sections), and \\(\\epsilon\\) is a learnable function with the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "parameter \\(\\theta\\). The STDD essentially represents the dependant relations of continuous frame slices as well as the diffusion steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.736, + 0.161 + ], + "angle": 0, + "content": "3.2. Forward Dual Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.29 + ], + "angle": 0, + "content": "In order to model the relations between video frames, we also use temporal Markov chain to describe the evolution dependency on time series, in addition to image diffusion process itself. Hereby, a spatio-temporal diffusion model could be built. To distinguish it from standard diffusion, let \\( z_{t}^{i} \\) denotes the noised latent variable of \\( x^{i} \\) at the diffusion step \\( t \\). According to the diffusion mechanism, we define the forward spatio-temporal noising process as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.299, + 0.907, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t} ^ {i} \\sim p \\left(z _ {t} ^ {i} \\mid z _ {t - 1} ^ {i}, z _ {t - 1} ^ {i - 1}\\right) \\tag {6} \\\\ = \\sqrt {\\alpha_ {t}} (\\underbrace {\\sqrt {\\rho_ {t}} * z _ {t - 1} ^ {i} + \\sqrt {1 - \\rho_ {t}} * z _ {t - 1} ^ {i - 1}} _ {\\text {t e m p o r a l d o m a i n}}) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t} ^ {i} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.477 + ], + "angle": 0, + "content": "where the diffusion coefficient \\(\\rho_t \\in (0,1]\\) is used to balance the prior frame and the current frame. In the above equation, the temporal-domain term denotes the temporal diffusion between the \\(i\\)-th and \\((i - 1)\\)-th frames, and \\(\\sqrt{\\alpha_t(\\cdot)} + \\sqrt{1 - \\alpha_t} \\epsilon_t^i\\) is an image-like diffusion. In order to conveniently operate the noising process, we derive an analyzed expression as shown in the following Proposition 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.487, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Proposition 1. According to the recursive formula (6), and suppose \\(\\rho_0 = \\rho_1 = \\dots = \\rho_t = \\rho\\), we compute the diffusion state of any time frame and any diffusion step from the initial states (\\(z_0^0 = x^0, z_0^1 = x^1, \\dots, z_0^i = x^i\\)):" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.557, + 0.906, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t} ^ {i} \\sim p \\big (z _ {t} ^ {i} | \\{z _ {0} ^ {k} | k = 1, 2,.., i \\} \\big) \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} g (t, i) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.6, + 0.888, + 0.643 + ], + "angle": 0, + "content": "\\[\ns. t. g (t, i) = \\sum_ {m = 0} ^ {m i n (t, i)} {\\binom {t} {m}} \\rho^ {(t - m) / 2} (1 - \\rho) ^ {m / 2} z _ {0} ^ {i - m},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.686 + ], + "angle": 0, + "content": "where we only retain the existing terms \\( z_0^{i - m} \\) w.r.t \\( m \\leq i \\) in \\( g(t,i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.695, + 0.906, + 0.725 + ], + "angle": 0, + "content": "Proof. The derivation process is deferred to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.795 + ], + "angle": 0, + "content": "In Proposition 1, the dual diffusion process integrates the information of time slices and noising steps jointly. The historical frame diffusion term \\( g(t,i) \\) accumulates the information of previous historical frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "It is worth noting that, the standard diffusion process is a special case of the above model when setting \\(\\rho = 1\\). For the initial frame, i.e., \\(i = 0\\), we consider it as a standard image diffusion process, i.e., \\(z_{t}^{0} = \\sqrt{\\bar{\\alpha}_{t}} z_{0}^{0} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{t}^{0}\\). The probability analysis of the inverse spatio-temporal diffusion is given in Section 3.3, and a practical inverse diffusion can be found directly in Section 3.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12577" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "In intuitive, it seems require a high-computational burden, but in practice the combinatorial coefficient term \\(\\left\\{\\binom{t}{m}\\right\\}\\) could be pre-calculated as a searching table, which would reduce the running cost. In addition, if \\(1 - \\rho = < 0.01\\), then \\(\\rho^{(t - m / 2)} \\approx e^{-t(1 - \\rho) / 2}\\). Therefore, \\(g(t,i)\\) also can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.19, + 0.484, + 0.233 + ], + "angle": 0, + "content": "\\[\ng (t, i) = \\sum_ {m = 0} ^ {m i n (t, i)} \\binom {t} {m} (1 - \\rho) ^ {m / 2} e ^ {- t (1 - \\rho) / 2} z _ {0} ^ {i - m} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.319 + ], + "angle": 0, + "content": "where \\( m \\) is limited to [0, 16], which require small running cost. Notice, the above coefficient substitution is only for the purpose of simplifying the calculation in the experiment and is independent of our model. And the setting of the parameter \\( \\rho \\) will be discussed in the ablation experiment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.303, + 0.342 + ], + "angle": 0, + "content": "3.3. Inverse Dual Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Probability diffusion process can be reduced to a constrained latent variable model whose Gaussian distribution of forward and reverse processes must be modeled. To derive the inverse decoding process, we establish an explicit link between spatio-temporal diffusion probabilistic models and inverse probability distributions, and then provide a variational optimization objective for the proposed dual diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.546 + ], + "angle": 0, + "content": "Let \\( Z_{t} \\) denotes a set of all frames at the diffusion step \\( t \\), i.e., \\( Z_{t} = (z_{t}^{0}, z_{t}^{1}, \\dots, z_{t}^{N-1}) \\). For convenient statement, we use a surrogate operator \\( F_{t}: Z_{0} \\to \\widetilde{Z}_{t} \\) and its inverse operator \\( F_{t}^{-1}: \\widetilde{Z}_{t} \\to Z_{0} \\), in which \\( F_{t} \\) as well as \\( \\widetilde{Z}_{t} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.555, + 0.483, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\widetilde {Z} _ {t} = F _ {t} \\left(Z _ {0}\\right) := \\left(\\widetilde {z} _ {t} ^ {0}, \\dots , \\widetilde {z} _ {t} ^ {i}, \\dots , \\widetilde {z} _ {t} ^ {N - 1}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.578, + 0.482, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .}, \\quad \\widetilde {z} _ {t} ^ {i} = g (t, i). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.604, + 0.484, + 0.651 + ], + "angle": 0, + "content": "\\(F_{t}\\) represents the diffusion process between video frames, which is called temporal diffusion. Then, according to Eqn. (6) and Proposition 1, the variable \\(Z_{t}\\) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.659, + 0.483, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} F _ {t} \\left(Z _ {0}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {11} \\\\ = \\sqrt {\\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t}, \\qquad (1 2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.783 + ], + "angle": 0, + "content": "where the noise \\(\\epsilon_{t} = (\\epsilon_{t}^{0},\\epsilon_{t}^{1},\\dots,\\epsilon_{t}^{i})\\), and Eqn. (12) is the iterative formula and the Markov chain of the forward diffusion process. Therefore, the conditional probability distribution of the forward diffusion process can be written as a Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.792, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q (Z _ {t} | Z _ {t - 1}, Z _ {0}) = \\mathcal {N} (Z _ {t} | \\sqrt {\\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) (1 - \\alpha_ {t}) I), \\\\ q \\left(Z _ {t} \\mid Z _ {0}\\right) = \\mathcal {N} \\left(Z _ {t} \\mid \\sqrt {\\bar {\\alpha} _ {t}} F _ {t} \\left(Z _ {0}\\right), (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then, video diffusion model can be modeled as some latent variables with the form: \\( p_{\\theta}(Z_0) = \\int p_{\\theta}(Z_{0:T})d_{Z_{1:T}} \\), where the subscript \\( \\square_{1:T} \\) means the collection of latent variables from step 1 to step \\( T \\) and \\( \\theta \\) is the learnt parameter. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "joint distribution \\( p_{\\theta}(Z_{0:T}) \\) refers to the reverse process, and it can be defined as a Markov chain with learned Gaussian transitions starting at \\( p(Z_T) = \\mathcal{N}(0,I) \\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.148, + 0.907, + 0.169 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (Z _ {0: T}) = p (Z _ {T}) \\Pi_ {t = 1} ^ {T} p _ {\\theta} (Z _ {t - 1} | Z _ {t}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.177, + 0.907, + 0.222 + ], + "angle": 0, + "content": "Therefore, we need to derive \\( p_{\\theta}(Z_{t - 1}|Z_t) \\) as forward process posteriors, which are tractable when conditioned on \\( Z_0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.232, + 0.907, + 0.252 + ], + "angle": 0, + "content": "\\[\nq \\left(Z _ {t - 1} \\mid Z _ {t}\\right) = q \\left(Z _ {t - 1} \\mid Z _ {t}, Z _ {0}\\right) = \\mathcal {N} \\left(\\tilde {\\mu_ {t}}, \\tilde {\\beta_ {t}}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.906, + 0.324 + ], + "angle": 0, + "content": "where we assume the Gaussian distribution in the last equation. Then, with the help of Eqn. (13), we obtain the mean and variance of the posterior distribution of the diffusion process \\( q(Z_{t - 1}|Z_t) = N(\\tilde{\\mu}_t,\\tilde{\\beta}_t) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.334, + 0.906, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\tilde {\\mu} _ {t} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(F _ {t - 1} \\left(F _ {t} ^ {- 1} \\left(Z _ {t}\\right)\\right) - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {t}\\right), \\\\ \\tilde {\\beta} _ {t} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} (1 - \\alpha_ {t}), \\end{array} \\right. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.904, + 0.425 + ], + "angle": 0, + "content": "where the relevant proof is given in the supplement material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.434, + 0.831, + 0.449 + ], + "angle": 0, + "content": "3.4. Accelerate Inverse Inference Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.591 + ], + "angle": 0, + "content": "In order to boost sampling speed, we derive an implicit probability diffusion model. The denoising diffusion probabilistic process, in section 3.3, requires the construction of Markov chains with the help of the additional operator \\( F^{-1} \\), and more computational complexity. But, denoising diffusion implicit models has faster sampling speeds and does not require Markov chains. Below we do a derivation to fulfill the spatio-temporal inverse diffusion process with DDIM, which enables more practical use." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.906, + 0.622 + ], + "angle": 0, + "content": "First, according to Proposition 1, we make reasonable assumptions for the reverse process" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.635, + 0.905, + 0.654 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} ^ {i} = k _ {t} z _ {t} ^ {i} + C _ {t} ^ {i} \\left(Z _ {0} ^ {0: i - 1}\\right) + \\sigma_ {t} \\epsilon_ {t - 1}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.906, + 0.721 + ], + "angle": 0, + "content": "where \\(k_{t}\\) and \\(\\sigma_{t}\\) are coefficient to be determined, \\(C_t^i (Z_0^{0:i - 1})\\) is a function (also written as \\(C_t^i\\)) and \\(\\epsilon^{\\prime}\\sim \\mathcal{N}(0,I)\\). Then, according to Proposition 1 and Eqn. (7), we replace \\(z_{t}^{i}\\) with an expression of \\(Z_{0}\\) and \\(\\epsilon_t^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.732, + 0.905, + 0.757 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} ^ {i} = k _ {t} \\sqrt {\\bar {\\alpha} _ {t}} g (t, i) + C _ {t} ^ {i} + \\sqrt {k _ {t} ^ {2} \\left(1 - \\bar {\\alpha} _ {t}\\right) + \\sigma_ {t} ^ {2}} \\epsilon \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.767, + 0.906, + 0.814 + ], + "angle": 0, + "content": "Compared with Eqn. (7) in Proposition 1 (i.e., \\(\\sqrt{\\bar{\\alpha}_{t-1}} g(t-1,i)+\\sqrt{1-\\bar{\\alpha}_{t-1}}\\epsilon\\)), we can solve the coefficients \\(k_{t}\\) and function \\(C_{t}^{i}(Z_{0}^{0:i-1})\\) with conditional of \\(\\sigma_{t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.835, + 0.906, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} k _ {t} = \\sqrt {\\frac {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}}{1 - \\bar {\\alpha} _ {t}}}, \\\\ C _ {t} ^ {i} \\left(Z _ {0} ^ {0: i - 1}\\right) = \\sqrt {\\bar {\\alpha} _ {t - 1}} g (t - 1, i) - k _ {t} \\sqrt {\\bar {\\alpha} _ {t}} g (t, i). \\end{array} \\right. \\tag {19}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12578" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.288, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 STDD Training" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.112, + 0.169, + 0.125 + ], + "angle": 0, + "content": "1: repeat" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.126, + 0.227, + 0.14 + ], + "angle": 0, + "content": "2: \\(Z_{0} \\sim q(Z_{0})\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.141, + 0.303, + 0.155 + ], + "angle": 0, + "content": "3: \\(t \\sim \\text{Uniform}(1, \\dots, T)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.157, + 0.229, + 0.17 + ], + "angle": 0, + "content": "4: \\(\\epsilon \\sim \\mathcal{N}(0,I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.171, + 0.372, + 0.186 + ], + "angle": 0, + "content": "5: \\(Z = \\sqrt{\\bar{\\alpha}_t} F_t(Z_0) + \\sqrt{1 - \\bar{\\alpha}_t} F_t(\\epsilon)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.187, + 0.341, + 0.2 + ], + "angle": 0, + "content": "6: Take gradient descent step on" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.201, + 0.352, + 0.216 + ], + "angle": 0, + "content": "7: \\(\\nabla_{\\theta}\\mathcal{L}(\\theta)\\) according to Eqn. (23)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.217, + 0.227, + 0.231 + ], + "angle": 0, + "content": "8: until converged" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.112, + 0.372, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.253, + 0.294, + 0.269 + ], + "angle": 0, + "content": "Algorithm 2 STDD Sampling" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.273, + 0.204, + 0.288 + ], + "angle": 0, + "content": "1: \\(\\epsilon \\sim \\mathcal{N}(0,I)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.289, + 0.206, + 0.303 + ], + "angle": 0, + "content": "2: \\( Z_{T} = F_{T}(\\epsilon) \\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.304, + 0.238, + 0.317 + ], + "angle": 0, + "content": "3: for \\( t = T \\) to 1 do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.318, + 0.34, + 0.334 + ], + "angle": 0, + "content": "4: \\(\\epsilon_{\\theta} = \\epsilon_{\\theta}(Z_t,t),\\hat{z}_0^i = \\phi_\\theta (z_t^i,t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.335, + 0.398, + 0.349 + ], + "angle": 0, + "content": "5: Calculate \\( Z_{t-1} \\) according to Eqn. (20)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.35, + 0.174, + 0.362 + ], + "angle": 0, + "content": "6: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.365, + 0.19, + 0.378 + ], + "angle": 0, + "content": "7: return \\( Z_0 \\)" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.273, + 0.398, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.457 + ], + "angle": 0, + "content": "We further bring the solved parameters \\( k_{t} \\) and \\( C_t^i \\) into Eqn. (18) to obtain the accelerated inverse inference process of STDD, with setting \\( \\epsilon_t^i = \\epsilon_\\theta^i \\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.47, + 0.464, + 0.496 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} ^ {i} = \\sqrt {\\bar {\\alpha} _ {t - 1}} (u _ {t} ^ {i} + v _ {t} ^ {i}) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} ^ {i} + \\sigma_ {t} \\epsilon^ {\\prime}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.499, + 0.482, + 0.531 + ], + "angle": 0, + "content": "s.t., \\(u_{t}^{i} = \\frac{z_{t}^{i} - \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{\\theta}^{i}}{\\sqrt{\\bar{\\alpha}_{t}}},\\quad \\epsilon_{\\theta}^{i} = \\epsilon_{\\theta}(z_{t}^{i},t),\\) (20)" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.534, + 0.445, + 0.552 + ], + "angle": 0, + "content": "\\[\nv _ {t} ^ {i} = g (t - 1, i) - g (t, i), \\quad \\tilde {z} _ {0} ^ {i} = \\phi_ {\\theta} (z _ {t} ^ {i}, t),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.685 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mu}_t^i = \\sqrt{\\bar{\\alpha}_t} * u_t^i\\) is the mean of \\(z_{t}^{i}\\), \\(v_{t}^{i} \\approx F_{t-1}((1 - \\sqrt{\\rho})\\tilde{z}_{0}^{i} - \\sqrt{1 - \\rho}\\tilde{z}_{0}^{i-1})\\) is temporal domain difference, \\(\\epsilon'\\) is the Gaussian noise added during inference and \\(\\phi_{\\theta}\\) is a network, then \\(v_{t}^{i}\\) can be calculated by Eqn. (7) with \\(\\tilde{z}_{0}^{i}\\). In practice, \\(\\epsilon_{\\theta}\\) and \\(\\phi_{\\theta}\\) are different channel components of the same network output. Moreover, in the process of inference, there is no calculation of the inverse function of \\(F\\) in Eqn. (16)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.745 + ], + "angle": 0, + "content": "How to understand \\( u \\) and \\( v \\): Assuming a new variable \\( \\tau \\), rewrite \\( z_{t}^{i} \\) in Eqn. (7) as: \\( z^{i}(t,\\tau) = \\sqrt{\\bar{\\alpha}_{t}} g(\\tau ,i) + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}^{i} \\). Then, according to SDE model [34], we conclude that" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.756, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} d _ {t} z ^ {i} (t, \\tau) \\approx - 1 / 2 \\beta_ {t} u _ {t} ^ {i} + \\sqrt {\\beta_ {t}} d \\omega \\\\ d _ {\\tau} z ^ {i} (t, \\tau) \\approx - \\sqrt {\\bar {\\alpha} _ {t}} v _ {\\tau} ^ {i} \\end{array} , \\right. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\omega\\) gaussian component. Therefore, \\(u_{t}^{i}\\) is image spatial diffusion component and \\(v_{t}^{i}\\) a temporal diffusion or bridge between frames. Therefore, the process inference consists of a diffusion model (i.e., Eqn. (20)) and a difference between frames, shown in Algorithm 2. In addition, to ensure the association between frames, we set \\(Z_{T} = F(\\epsilon)\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.738, + 0.107 + ], + "angle": 0, + "content": "3.5. Optimize Dual Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.415 + ], + "angle": 0, + "content": "The dual diffusion architecture includes Variational AutoEncoder (VAE) for video encoding and decoding, diffusion networks \\(\\epsilon_{\\theta}\\) for denoising, and conditional embedding networks (if needed). In detail, we employ VAE model training in an unsupervised manner, contains two trainable modules: encoder \\(\\mathcal{E}\\) to represent image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) as latent map \\(z \\in \\mathbb{R}^{h \\times w \\times c}\\) and decoder \\(\\mathcal{D}\\) to decode the latent map \\(\\tilde{Z}\\) as output image \\(\\tilde{X}\\). VAE was trained with the reconstruction loss [28] using only unlabeled images, or load pre-trained model such as Stable Diffusion v1.4 [28]. Then, we use transformer network, used in [21], to extract the feature of video and learn the noise of dual diffusion process. And time step \\(t\\) is specified to \\(\\epsilon_{\\theta}\\) by the sinusoidal position embedding [38]. In addition, we enhance the network backbone by using a cross-attention mechanism [28] to transform the model into a more flexible conditional diffusion model. And, we adopt different pre-training condition networks according to different tasks, such as CLIP [26] for text-to-video generation. Implementation details will be presented in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.906, + 0.447 + ], + "angle": 0, + "content": "For optimizing STDD, the KL divergence to directly compare \\( p_{\\theta}(Z_{t - 1}|Z_t) \\) against forward process posteriors is:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.474, + 0.905, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {K L} \\left(q \\left(Z _ {t - 1} \\mid Z _ {t}\\right) | | p _ {\\theta} \\left(Z _ {t - 1} \\mid Z _ {t}\\right)\\right) \\\\ = \\frac {\\left| \\left| \\tilde {\\mu} _ {t - 1} - \\mu_ {\\theta} \\right| \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}} + C \\\\ = \\frac {\\left| \\left| \\tilde {\\mu} _ {t - 1} - \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(U _ {t} + V _ {t}\\right) \\right| \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}} + C \\tag {22} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.906, + 0.637 + ], + "angle": 0, + "content": "where \\(C\\) is a constant, \\(U_{t} = (u_{t}^{0}, u_{t}^{1}, \\dots, u_{t}^{N - 1})\\) and \\(V_{t} = (v_{t}^{0}, v_{t}^{1}, \\dots, v_{t}^{N - 1})\\). We further optimize \\(\\epsilon_{\\theta}\\) and \\(\\phi_{\\theta}\\) respectively, thus the loss of the spatio-temporal diffusion probability model is" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.648, + 0.813, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k l} (\\theta) = \\mathbb {E} | | \\tilde {\\mu} _ {t - 1} / \\sqrt {\\bar {\\alpha} _ {t - 1}} - (U _ {t} + V _ {t}) | | ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.669, + 0.888, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} (\\theta) = \\mathbb {E} | | \\boldsymbol {\\epsilon} _ {t} - \\epsilon_ {\\theta} (Z _ {t}, t) | | ^ {2} + | | Z _ {0} - \\phi_ {\\theta} (Z _ {t}, t) | | ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.689, + 0.904, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathcal {L} _ {k l} (\\theta) + \\mathcal {L} _ {m s e} (\\theta), \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.904, + 0.748 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents the parameters of the STDD network. The detailed training process is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.646, + 0.779 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.701, + 0.804 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets: To validate the proposed STDD, we conduct comprehensive evaluations and comparisons on Sky Time-Lapse [47] and UCF101 [35] for video generation, MHAD human action dataset [3] for image-to-video and WebVid-10M [1] for text-to-video. UCF101 is a realistic action video motion recognition dataset, collected from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12579" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.615, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.096, + 0.891, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 1. Qualitative comparisons on MHAD for video prediction. In the red box, the head in LFDM has serious artifacts and distortions, which we handled well." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.427, + 0.482, + 0.548 + ], + "angle": 0, + "content": "YouTube, providing 13,320 videos from 101 action categories. MHAD human action dataset contains 861 videos of 27 actions performed by 8 subjects. UCF101, Sky Time-Lapse and MHAD are cropped to 16 frames for training our model. For quantitative evaluation, we also train a text-conditioned video generation model on WebVid-10M [1], which consists of 10.7M short videos with paired textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Metrics: Following previous works [7, 20], we mainly use Frechet Video Distance (FVD) [37] as the evaluation metric. On UCF101 and Sky Time-Lapse, we report the results of FVD. Following LFDM [22] on MHAD, we report the results of FVD and class conditional FVD (cFVD). The cFVD compare the distance between real and synthesized video feature distributions under the same class condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Implementation details: We use the transformer network, used in [21], to extract the features of video and learn the noise of the dual diffusion process. The network learning rate is \\( 1e - 4 \\), the batch size is 16, and the training stpes is 200000. The diffusion steps \\( T = 1000 \\), the steps of accelerate reverse process is 250 and \\( \\rho = 1 - 10^{-3} \\). We will provide the generated videos and code on github." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.395, + 0.789 + ], + "angle": 0, + "content": "4.2. Comparisons with state-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video prediction. On MHAD dataset, we have conducted a comparison of our methods with several competitive approaches, including ImaGINator [41], VDM [13], LDM [28] and LFDM [22]. Table 1 shows the quantitative comparison between our method and state-of-the-art approaches like LDM and LFDM, where our proposed STDD consistently outperforms all other models. The results show" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.426, + 0.907, + 0.454 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on MHAD. \\( \\downarrow \\) denotes the lower the better. The best results are denoted in bold." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.465, + 0.855, + 0.556 + ], + "angle": 0, + "content": "
MethodResolutionFVD ↓cFVD↓
ImaGINator [41]64 × 648891406
VDM [13]64 × 64295531
LDM [28]128 × 128337594
LFDM [22]128 × 128214426
Ours128 × 128195411
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.571, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Table 2. FVD values of video generation models on UCF101 and Sky Time-Lapse. The lower the better and the best results are denoted in bold." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.625, + 0.882, + 0.751 + ], + "angle": 0, + "content": "
MethodUCF101Sky Time-LapseParams
VideoGPT [48]2880222-
MoCoGAN-HD [36]1729164-
DIGAN [50]163083.1-
StyleGAN-V [31]143179.5-
PVDM [51]114175.5-
LVDM [10]37295.2437M
Latte+IMG [21]33342.7673M
Ours25842.1868M
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.884 + ], + "angle": 0, + "content": "that our STDD is superior to LFDM (195 vs 214 in FVD at the resolution of \\((128 \\times 128)\\), and LFDM achieves the second best performance. LFDM distorts the initial frame through an optical flow sequence to generate a new frame to model the time correlation, while our method is mainly to explicitly model the frame time correlation, which is conducive to the propagation of historical frame information." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video generation. On UCF101 dataset, we have con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12580" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.505, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.347, + 0.385, + 0.361 + ], + "angle": 0, + "content": "(a)UCF101 (Playing Sitar)" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.094, + 0.895, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.347, + 0.756, + 0.361 + ], + "angle": 0, + "content": "(b) Sky Time-Lapse" + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.376, + 0.761, + 0.391 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparisons on UCF101 and Sky Time-Lapse for video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.485, + 0.72 + ], + "angle": 0, + "content": "duced a comparison of our methods with several competitive approaches, including VideoGPT [48], StyleGAN [31], DIGAN [50], TATS [7], VDM [13], among others. The quantitative results on UCF101 dataset are presented in Table 2. In terms of video generation, VDM demonstrates superior performance compared to GAN-based methods, particularly in terms of FVD (Fréchet Video Distance). This indicates the potential of DPM-based video generation methods. Our proposed STDD achieves even better results than LVDM by a significant margin. Compared with Latte-IMG, which prerained on images, our approach is also comparable. In addition, the quantitative results in Sky Time-Lapse also achieve competitive performance, shown in Table 2. These comparisons highlight the effectiveness of STDD in video generation tasks, demonstrating its superior performance over existing methods in terms of various evaluation metrics and datasets. In addition, the inference time of our model did not increase significantly with the same inference steps and device (23s vs 20s of Latte with 250 inference steps in RTX 3090 GPU)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.245, + 0.745 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Visual comparison. On MHAD dataset, we have conducted a qualitative comparison of our methods with several video pridiction models, including ImaGINator, VDM, LDM and LFDM. As shown in Fig. 1, the first line shows the given image, and the rest shows the generated sequence. In general, our approach has advantages over other video generation methods, especially generating 'stand-to-sit' videos with larger variation areas and more motion variations. The results in Fig. 1 (b) showed that the head of the LFDM generated character had severe artifacts and distor" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.418, + 0.772, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.419, + 0.897, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.516, + 0.772, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.614, + 0.698, + 0.627 + ], + "angle": 0, + "content": "Frame weights" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.517, + 0.897, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.614, + 0.882, + 0.628 + ], + "angle": 0, + "content": "\\(g(t = 1000,i)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.686 + ], + "angle": 0, + "content": "Figure 3. The tendency of frame weights in Eqn. (7) and Eqn. (8). Appropriate \\(\\rho\\) enhances the similarities and blurring the differences between frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.824 + ], + "angle": 0, + "content": "tions, which we handled well, notably in the red box. We also provide visual comparisons with the most recent video generation methods, i.e. TATS, DIGAN, VideoFusion and Latte. As shown in Fig. 2, each generated video has 16 frames with a resolution of \\(256 \\times 256\\), and we show 4-th, 8-th, 12-th and 16-th frame. Our STDD could generate more continuous and moving videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis of \\(\\rho\\). To explore the effects of \\(\\rho\\) in Eqn (7), we conducted experiments on MHAD with a resolution of \\(64 \\times 64\\), shown in Table 3. Appropriate \\(\\rho\\) can better integrate the historical frames and improve the performance. Also, smaller \\(\\rho\\) leads to information loss of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.094, + 0.842, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.292, + 0.68, + 0.306 + ], + "angle": 0, + "content": "(a) A dog in astronaut suit and sunglasses floating in space" + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.308, + 0.506, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.442, + 0.446, + 0.456 + ], + "angle": 0, + "content": "(b) A turtle swimming in the ocean" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.308, + 0.839, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.442, + 0.782, + 0.456 + ], + "angle": 0, + "content": "(c) Surfer on his surfboard in a wave" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.472, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 4. Qualitative evaluation on text-to-video generation. Both our method and Videofusion are trained on WebVid-10M, while Latte is trained on Vimeo25M [42]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.524, + 0.486, + 0.568 + ], + "angle": 0, + "content": "Table 3. Analysis \\(\\rho\\) in Eqn. (7) on MHAD with the resolution of \\(64\\times 64\\). Appropriate \\(\\rho\\) can improve the performance and \\(\\sqrt{\\rho^{1000}}\\) represents the weight of the current frame after 1000 steps." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.577, + 0.435, + 0.641 + ], + "angle": 0, + "content": "
ρ10.99990.9990.9950.99
\\( \\sqrt{\\rho^{1000}} \\)10.9510.6060.0820.007
FVD ↓157149133243388
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.483, + 0.797 + ], + "angle": 0, + "content": "current frame, which is not conducive to video generation. Since the coefficient is exponential, it must be required that \\(0 < 1 - \\rho < 0.01\\), which consistent with the hypothesis in Eqn. (8). The weights of the current and historical frames are shown in Fig. 3. Visualization shows that \\(g(t,i)\\) enhances the similarities and blurring the differences between frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Text-to-video generation. To explore the video generation capabilities of our proposed approach, we extend STDD to text-to-video generation. We train a text-to-video model, pretrained by Stable Diffusion [28], on the large scale video dataset, i.e. WebVid-10M. Some samples are shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.525, + 0.634, + 0.541 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.657 + ], + "angle": 0, + "content": "In this paper, we propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes. Then, we theoretically derive probabilistic reverse process and propose an accelerated sampling method. Both qualitative and quantitative experiments demonstrated that our model achieve more competitive performance on video generation." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.668, + 0.611, + 0.684 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.693, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.762, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.832, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[3] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.693, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12582" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.208, + 0.483, + 0.251 + ], + "angle": 0, + "content": "[6] Yu Gao, Jiancheng Huang, Xiaopeng Sun, Zequn Jie, Yujie Zhong, and Lin Ma. Matten: Video generation with mamba-attention. arXiv preprint arXiv:2405.03025, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[7] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In European Conference on Computer Vision, pages 102–118. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[9] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.497, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.527, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.571, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.614, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[14] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[15] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[16] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[17] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[19] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[20] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. CoRR, abs/2303.08320, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[21] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18444-18455. IEEE, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[25] Xichen Pan, Pengda Qin, Yuhong Li, Hui Xue, and Wenhu Chen. Synthesizing coherent story with auto-regressive latent diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2920-2930, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12583" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[31] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[32] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[34] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems, 35:23371-23385, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[40] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[41] Yaohui Wang, Piotr Bilinski, Francois Bremond, and Antitza Dantcheva. Imaginator: Conditional spatio-temporal gan for video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1160-1169, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 22025-22034, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[44] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[45] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[46] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[48] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[49] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25(10):1469, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[50] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[51] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12584" + } + ] +] \ No newline at end of file diff --git a/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_origin.pdf b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..238d9d100d77a1b237c18029487701d50760b5f4 --- /dev/null +++ b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/6f423d46-a407-4dbd-85d4-7aa0894a46ad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd0528ed5db0999c1e5d0d3fa593034372527bb13a346f4384c85cbd1895cedc +size 2026782 diff --git a/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/full.md b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e63f8c50ebcb78412381da6786f43852f3c4da63 --- /dev/null +++ b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/full.md @@ -0,0 +1,401 @@ +# STDD: Spatio-Temporal Dual Diffusion for Video Generation + +Shuaizhen Yao $^{1}$ , Xiaoya Zhang $^{1,*}$ , Xin Liu $^{2}$ , Mengyi Liu $^{3}$ , Zhen Cui $^{4,*}$ $^{1}$ Nanjing University of Science and Technology, + $^{2}$ Nanjing Seetacloud Technology Co., Ltd, + $^{3}$ Kuaishou, $^{4}$ Beijing Normal University +{ysz, zhangxiaoya}@njust.edu.cn, xin.liu@seetacloud.com, liumengyi@kuaishou.com, zhen.cui@bnu.edu.cn + +# Abstract + +Diffusion probabilistic model is becoming the cornerstone of data generation, especially generating high-quality images. As an extension, video diffusion generation is in urgent need of a principled temporal-sequence diffusion way, while the spatial-domain diffusion dominates most video diffusion methods. In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal noise propagation/reduction. Mathematically, an analysable dual diffusion process is derived to accumulate information in temporal sequence as well as spatial domain. Correspondingly, we theoretically derive a spatio-temporal probabilistic reverse diffusion process and propose an accelerated sampling way to reduce the inference cost. In principle, the spatio-temporal dual diffusion enables the information of previous frames to be transferred to the current frame, which thus could be beneficial for video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation/prediction as well as text-to-video generation. + +# 1. Introduction + +Diffusion probabilistic model (DPM) is becoming the cornerstone of data generation, where the forward process gradually contaminates the data with gaussian noises and the reverse denoising process is learned to generate new samples [12, 15]. Specifically, DPM has made significant progress in producing high quality, realistic images [17, 23, 24, 27, 29, 33]. Motivated by the success in image generation, some researchers are trying to extend the diffusion model from image to video by accompanying with + +temporal-dimensional prediction/interpolation [9, 39, 49]. However, due to intricate temporal and spatial correlations as well as high dimensionality of video, video diffusion generation remains a challenging task. + +The existing video generation methods based on DPMs [9, 13, 28] typically employ a standard diffusion process by treating each frame of the video as an independent diffusion process, which breaks the temporal correlation between frames. Concretely, in the noisig/denoising process, these video diffusion methods separately derive/sample priors, and recover corresponding frames from different-random priors. Due to different-random priors, the continuity of generated video frames is still a bottleneck of video diffusion generation, even though some historical (generated) frames could be used as a condition of the diffusion for some-degree mitigation [25]. An endeavor to this aspect is one recent work [20], decomposing video noise into two components: video sharing base noise and frame-independent residual noise, and then separately performing the diffusion process with two different diffusion networks, which greatly improves video consistency. + +Despite some progress in video diffusion generation, these existing methods could not well capture temporal dependencies due to the use of independent priors for each frame, which would often result into the nonsmoothness or inconsistency of generated videos, especially confronting with large appearance variations in temporal sequences. Essentially, the current video diffusion methods lack a principled diffusion process along time slices (w.r.t frame sequence), not just in the spatial domain (w.r.t every image). + +In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal frame propagation. Mathematically, we derive an analysable dual diffusion process by accumulating temporal frame information. In particular, the historical frames are propagated along with temporal sequence, while accompanying with + +the spatio-domain noises w.r.t diffusion on single-frame image. To reach the inverse diffusion, we derive a probabilistic reverse process, which guarantees the feasibility of the diffusion model. Inspired by the diffusion probabilistic implicit diffusion (DDIM) model [33], further, we propose an accelerated spatio-temporal sampling method that is both fast and computation-intensive. To facilitate learning, we decompose the model into spatial components representing the standard diffusion model and temporal components representing the frame-to-frame diffusion. After a series of derivation, we could reach a maneuverable solution by using neural network. Overall, the spatio-temporal diffusion enables the information of previous frames to be recursively transferred to the current frame, which endows the model good ability to capture temporal dependencies, and can be robust to video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation and prediction. In addition, we extend our STDD to text-to-video generation to explore the video generation capabilities. + +In summary, our main contributions are three folds: i) propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes; ii) theoretically derive probabilistic reverse process and propose an accelerated sampling method; iii) achieve more competitive performance on video generation. + +# 2. Related Works + +# 2.1. Diffusion Probabilistic Models + +DPM, which was first introduced in reference [32], comprises a diffusion (encoding) process and a denoising (decoding) process. In the diffusion process, given one image $x$ , the noise diffusion process is conducted, which random noises are gradually added by T-step Markov chain [15], denoted as $(x_0,x_1,\dots ,x_T)$ with + +$$ +\begin{array}{l} x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}, \\ \text {s . t .} \quad \bar {\alpha} _ {t} = \prod_ {k = 1} ^ {t} \alpha_ {t}, \epsilon_ {t} \sim \mathcal {N} (0, I), \tag {1} \\ \end{array} +$$ + +where $\alpha_{t} \in (0,1)$ is the corresponding diffusion coefficient. If $T$ is large enough, such as $T = 1000$ , $x_{T}$ approximates random Gaussian noise. Therefore, the reverse process of the diffusion model (denoising process) can be interpreted as the gradual generation of the data distribution from random Gaussian noise. In other words, it aims to recover the original data by removing the added noise step by step. + +Ho et al. [12] establish a connection between DPM and denoising score matching [34] and propose an $\epsilon$ -prediction formulation for the denoising process, which is $L = ||\epsilon_t -$ + +$\epsilon_{\theta}(x_t,t)||^2$ . In which, $\epsilon_{\theta}$ is a denoising neural network parameterized by $\theta$ , and $L$ is the loss function. Utilizing this formulation, DPM has been successfully applied to various generative tasks, including image generation [11, 19, 27], super-resolution [16], image translation [30], and more [43, 44]. Compared to other generation methods, such as Generative Adversarial Networks (GANs) [8], DPMs are easier to train and can generate more diverse samples [4]. DPMs utilize diffusion processes and de-noising steps, resulting in a more stable training process and improved sample diversity. These characteristics have contributed to the popularity and success of DPMs in the field of generative modeling. + +# 2.2. Video Diffusion Models + +Video generation is a highly challenging task in the field of generative research. It not only requires the generation of high-quality frames but also necessitates temporal correlation among the generated frames. Previous video generation methods have predominantly relied on Generative Adversarial Networks (GANs). In VGAN [40], the generator directly learns the joint distribution of video frames. Mocogan [36] decomposed a video clip into content and motion components, which are individually modeled using a content vector and a motion vector. Alternatively, other methods utilize a two-step process, which train a vector quantized autoencoder [5] for video data, and then employ an transformer network to learn the video distribution in the quantized latent space [7, 14, 48]. These techniques leverage the quantized latent space to model the video distribution and capture dependencies among frames. Overall, each approach has its own advantages and limitations in generating high-quality continuous videos. + +In recent years, researchers have been motivated by the remarkable achievements of Diffusion Probabilistic Models (DPM) in image generation and have made efforts to apply DPM to video generation. In reference [13], Ho et al. proposed a video diffusion model that extends the 2D denoising network from the image diffusion model to 3D by considering stacked frames as additional dimensions. This enables the model to capture spatio-temporal dependencies and generate coherent videos. In [39], DPM is utilized for video prediction and interpolation, where known frames are used as conditions for denoising. Some large video generation models [2, 18, 45, 46], with massive data and parameters, allow the attention layer to implicitly learn the relationship between video frames, which requires significant computing and storage resources. These methods often treat video frames as independent samples during the diffusion process. To establish a relation between the video frames, VideoFusion [20] breaks down the noise into the base noise for video sharing and the residual noise for the frames. Benefit from powerful time modeling capabilities, + +transformer-based method [6, 21] is also used to generate high quality video. Still, these methods fail to establish temporal associations between frames, especially adjacent frames, may lead to limitations in generating videos with smooth transitions and preserving temporal consistency. + +# 3. Spatio-Temporal Dual Diffusion + +In this section, we will illustrate our main idea, and then derive an analysable spatio-temporal forward and inverse diffusion process, meantime design an accelerated inverse diffusion version and its optimization process. + +# 3.1. Problem Description + +Let $X = (x^0, x^1, \dots, x^i, \dots, x^{N-1})$ denotes a video sequence with $N$ frames, where the superscript $i$ denotes the $i$ -th video frame. We also follow the customary use of taking subscript as diffusion step, i.e., the symbol $x_t^i$ with the diffusion step $t$ of the $i$ -th frame. The existing video diffusion methods usually take an independent frame-wise diffusion way through leveraging the classic diffusion model such as denoising diffusion probabilistic model [13]. Given one frame $i$ , the noise diffusion process is conducted, denoted as $(x_0^i, x_1^i, \dots, x_T^i)$ with + +$$ +x _ {t} ^ {i} = \sqrt {\alpha_ {t}} x _ {t - 1} ^ {i} + \sqrt {1 - \alpha_ {t}} \epsilon_ {t} ^ {i}, \tag {2} +$$ + +where $\epsilon_t^i\sim \mathcal{N}(0,1)$ $\alpha_{t} = 1 - \beta_{t}$ with $\beta_{t} > 0$ . The inverse denoising process could be recursively estimated by an optimized neural network $\pmb{\epsilon}$ , formally, + +$$ +x _ {t - 1} ^ {i} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(x _ {t} ^ {i} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(x _ {t} ^ {i}, t\right)\right), \tag {3} +$$ + +where $\bar{\alpha}_t = \Pi_{s=1}^t\alpha_s$ , and the initial noisy $x_T^i$ of each frame is also sampled independently from Gaussian noise. In essence, the previous video diffusion methods basically perform a combination of $n$ -frame independent image diffusion processes, although it is usually operated in the hidden feature space in order to enhance the inter-frame consistency. However, the inherent frame-sequential diffusion correlations have not yet been explicitly considered by previous diffusion models. + +Here we consider that the diffusion process should not only be imposed on one single image frame, but only be run on the stream of video frames, named joint spatial and temporal diffusion. From the view of mathematical formulation, we abstract the spatio-temporal dual diffusion as: + +$$ +x _ {t} ^ {i} = \mathbb {D} \left(x _ {t - 1} ^ {i}, x _ {t - 1} ^ {i - 1}, \epsilon_ {t} ^ {i}\right), \tag {4} +$$ + +$$ +x _ {t - 1} ^ {i} = \mathbb {D} ^ {- 1} \left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}, \epsilon_ {\theta} \left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}\right)\right), \tag {5} +$$ + +where the forward and inverse diffusion processes, $\mathbb{D}$ and $\mathbb{D}^{-1}$ , are well derived in analyzed mathematical expressions (see below sections), and $\epsilon$ is a learnable function with the + +parameter $\theta$ . The STDD essentially represents the dependant relations of continuous frame slices as well as the diffusion steps. + +# 3.2. Forward Dual Diffusion + +In order to model the relations between video frames, we also use temporal Markov chain to describe the evolution dependency on time series, in addition to image diffusion process itself. Hereby, a spatio-temporal diffusion model could be built. To distinguish it from standard diffusion, let $z_{t}^{i}$ denotes the noised latent variable of $x^{i}$ at the diffusion step $t$ . According to the diffusion mechanism, we define the forward spatio-temporal noising process as + +$$ +\begin{array}{l} z _ {t} ^ {i} \sim p \left(z _ {t} ^ {i} \mid z _ {t - 1} ^ {i}, z _ {t - 1} ^ {i - 1}\right) \tag {6} \\ = \sqrt {\alpha_ {t}} (\underbrace {\sqrt {\rho_ {t}} * z _ {t - 1} ^ {i} + \sqrt {1 - \rho_ {t}} * z _ {t - 1} ^ {i - 1}} _ {\text {t e m p o r a l d o m a i n}}) + \sqrt {1 - \alpha_ {t}} \epsilon_ {t} ^ {i} \\ \end{array} +$$ + +where the diffusion coefficient $\rho_t \in (0,1]$ is used to balance the prior frame and the current frame. In the above equation, the temporal-domain term denotes the temporal diffusion between the $i$ -th and $(i - 1)$ -th frames, and $\sqrt{\alpha_t(\cdot)} + \sqrt{1 - \alpha_t} \epsilon_t^i$ is an image-like diffusion. In order to conveniently operate the noising process, we derive an analyzed expression as shown in the following Proposition 1. + +Proposition 1. According to the recursive formula (6), and suppose $\rho_0 = \rho_1 = \dots = \rho_t = \rho$ , we compute the diffusion state of any time frame and any diffusion step from the initial states ( $z_0^0 = x^0, z_0^1 = x^1, \dots, z_0^i = x^i$ ): + +$$ +\begin{array}{l} z _ {t} ^ {i} \sim p \big (z _ {t} ^ {i} | \{z _ {0} ^ {k} | k = 1, 2,.., i \} \big) \\ = \sqrt {\bar {\alpha} _ {t}} g (t, i) + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} ^ {i}, \tag {7} \\ \end{array} +$$ + +$$ +s. t. g (t, i) = \sum_ {m = 0} ^ {m i n (t, i)} {\binom {t} {m}} \rho^ {(t - m) / 2} (1 - \rho) ^ {m / 2} z _ {0} ^ {i - m}, +$$ + +where we only retain the existing terms $z_0^{i - m}$ w.r.t $m \leq i$ in $g(t,i)$ . + +Proof. The derivation process is deferred to the supplementary material. + +In Proposition 1, the dual diffusion process integrates the information of time slices and noising steps jointly. The historical frame diffusion term $g(t,i)$ accumulates the information of previous historical frames. + +It is worth noting that, the standard diffusion process is a special case of the above model when setting $\rho = 1$ . For the initial frame, i.e., $i = 0$ , we consider it as a standard image diffusion process, i.e., $z_{t}^{0} = \sqrt{\bar{\alpha}_{t}} z_{0}^{0} + \sqrt{1 - \bar{\alpha}_{t}} \epsilon_{t}^{0}$ . The probability analysis of the inverse spatio-temporal diffusion is given in Section 3.3, and a practical inverse diffusion can be found directly in Section 3.4. + +In intuitive, it seems require a high-computational burden, but in practice the combinatorial coefficient term $\left\{\binom{t}{m}\right\}$ could be pre-calculated as a searching table, which would reduce the running cost. In addition, if $1 - \rho = < 0.01$ , then $\rho^{(t - m / 2)} \approx e^{-t(1 - \rho) / 2}$ . Therefore, $g(t,i)$ also can be written as: + +$$ +g (t, i) = \sum_ {m = 0} ^ {m i n (t, i)} \binom {t} {m} (1 - \rho) ^ {m / 2} e ^ {- t (1 - \rho) / 2} z _ {0} ^ {i - m} \tag {8} +$$ + +where $m$ is limited to [0, 16], which require small running cost. Notice, the above coefficient substitution is only for the purpose of simplifying the calculation in the experiment and is independent of our model. And the setting of the parameter $\rho$ will be discussed in the ablation experiment. + +# 3.3. Inverse Dual Diffusion + +Probability diffusion process can be reduced to a constrained latent variable model whose Gaussian distribution of forward and reverse processes must be modeled. To derive the inverse decoding process, we establish an explicit link between spatio-temporal diffusion probabilistic models and inverse probability distributions, and then provide a variational optimization objective for the proposed dual diffusion model. + +Let $Z_{t}$ denotes a set of all frames at the diffusion step $t$ , i.e., $Z_{t} = (z_{t}^{0}, z_{t}^{1}, \dots, z_{t}^{N-1})$ . For convenient statement, we use a surrogate operator $F_{t}: Z_{0} \to \widetilde{Z}_{t}$ and its inverse operator $F_{t}^{-1}: \widetilde{Z}_{t} \to Z_{0}$ , in which $F_{t}$ as well as $\widetilde{Z}_{t}$ is defined as: + +$$ +\widetilde {Z} _ {t} = F _ {t} \left(Z _ {0}\right) := \left(\widetilde {z} _ {t} ^ {0}, \dots , \widetilde {z} _ {t} ^ {i}, \dots , \widetilde {z} _ {t} ^ {N - 1}\right), \tag {9} +$$ + +$$ +\text {s . t .}, \quad \widetilde {z} _ {t} ^ {i} = g (t, i). \tag {10} +$$ + +$F_{t}$ represents the diffusion process between video frames, which is called temporal diffusion. Then, according to Eqn. (6) and Proposition 1, the variable $Z_{t}$ can be written as + +$$ +\begin{array}{l} Z _ {t} = \sqrt {\bar {\alpha} _ {t}} F _ {t} \left(Z _ {0}\right) + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} \tag {11} \\ = \sqrt {\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) + \sqrt {1 - \alpha_ {t}} \epsilon_ {t}, \qquad (1 2) \\ \end{array} +$$ + +where the noise $\epsilon_{t} = (\epsilon_{t}^{0},\epsilon_{t}^{1},\dots,\epsilon_{t}^{i})$ , and Eqn. (12) is the iterative formula and the Markov chain of the forward diffusion process. Therefore, the conditional probability distribution of the forward diffusion process can be written as a Gaussian distribution: + +$$ +\begin{array}{l} q (Z _ {t} | Z _ {t - 1}, Z _ {0}) = \mathcal {N} (Z _ {t} | \sqrt {\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) (1 - \alpha_ {t}) I), \\ q \left(Z _ {t} \mid Z _ {0}\right) = \mathcal {N} \left(Z _ {t} \mid \sqrt {\bar {\alpha} _ {t}} F _ {t} \left(Z _ {0}\right), (1 - \bar {\alpha} _ {t}) I\right). \tag {13} \\ \end{array} +$$ + +Then, video diffusion model can be modeled as some latent variables with the form: $p_{\theta}(Z_0) = \int p_{\theta}(Z_{0:T})d_{Z_{1:T}}$ , where the subscript $\square_{1:T}$ means the collection of latent variables from step 1 to step $T$ and $\theta$ is the learnt parameter. The + +joint distribution $p_{\theta}(Z_{0:T})$ refers to the reverse process, and it can be defined as a Markov chain with learned Gaussian transitions starting at $p(Z_T) = \mathcal{N}(0,I)$ . Formally, + +$$ +p _ {\theta} (Z _ {0: T}) = p (Z _ {T}) \Pi_ {t = 1} ^ {T} p _ {\theta} (Z _ {t - 1} | Z _ {t}). \tag {14} +$$ + +Therefore, we need to derive $p_{\theta}(Z_{t - 1}|Z_t)$ as forward process posteriors, which are tractable when conditioned on $Z_0$ : + +$$ +q \left(Z _ {t - 1} \mid Z _ {t}\right) = q \left(Z _ {t - 1} \mid Z _ {t}, Z _ {0}\right) = \mathcal {N} \left(\tilde {\mu_ {t}}, \tilde {\beta_ {t}}\right), \tag {15} +$$ + +where we assume the Gaussian distribution in the last equation. Then, with the help of Eqn. (13), we obtain the mean and variance of the posterior distribution of the diffusion process $q(Z_{t - 1}|Z_t) = N(\tilde{\mu}_t,\tilde{\beta}_t)$ : + +$$ +\left\{ \begin{array}{l} \tilde {\mu} _ {t} = \frac {1}{\sqrt {\alpha_ {t}}} \left(F _ {t - 1} \left(F _ {t} ^ {- 1} \left(Z _ {t}\right)\right) - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \boldsymbol {\epsilon} _ {t}\right), \\ \tilde {\beta} _ {t} = \frac {1 - \bar {\alpha} _ {t - 1}}{1 - \bar {\alpha} _ {t}} (1 - \alpha_ {t}), \end{array} \right. \tag {16} +$$ + +where the relevant proof is given in the supplement material. + +# 3.4. Accelerate Inverse Inference Process + +In order to boost sampling speed, we derive an implicit probability diffusion model. The denoising diffusion probabilistic process, in section 3.3, requires the construction of Markov chains with the help of the additional operator $F^{-1}$ , and more computational complexity. But, denoising diffusion implicit models has faster sampling speeds and does not require Markov chains. Below we do a derivation to fulfill the spatio-temporal inverse diffusion process with DDIM, which enables more practical use. + +First, according to Proposition 1, we make reasonable assumptions for the reverse process + +$$ +z _ {t - 1} ^ {i} = k _ {t} z _ {t} ^ {i} + C _ {t} ^ {i} \left(Z _ {0} ^ {0: i - 1}\right) + \sigma_ {t} \epsilon_ {t - 1}, \tag {17} +$$ + +where $k_{t}$ and $\sigma_{t}$ are coefficient to be determined, $C_t^i (Z_0^{0:i - 1})$ is a function (also written as $C_t^i$ ) and $\epsilon^{\prime}\sim \mathcal{N}(0,I)$ . Then, according to Proposition 1 and Eqn. (7), we replace $z_{t}^{i}$ with an expression of $Z_{0}$ and $\epsilon_t^i$ : + +$$ +z _ {t - 1} ^ {i} = k _ {t} \sqrt {\bar {\alpha} _ {t}} g (t, i) + C _ {t} ^ {i} + \sqrt {k _ {t} ^ {2} \left(1 - \bar {\alpha} _ {t}\right) + \sigma_ {t} ^ {2}} \epsilon \tag {18} +$$ + +Compared with Eqn. (7) in Proposition 1 (i.e., $\sqrt{\bar{\alpha}_{t-1}} g(t-1,i)+\sqrt{1-\bar{\alpha}_{t-1}}\epsilon$ ), we can solve the coefficients $k_{t}$ and function $C_{t}^{i}(Z_{0}^{0:i-1})$ with conditional of $\sigma_{t}$ : + +$$ +\left\{ \begin{array}{l} k _ {t} = \sqrt {\frac {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}}{1 - \bar {\alpha} _ {t}}}, \\ C _ {t} ^ {i} \left(Z _ {0} ^ {0: i - 1}\right) = \sqrt {\bar {\alpha} _ {t - 1}} g (t - 1, i) - k _ {t} \sqrt {\bar {\alpha} _ {t}} g (t, i). \end{array} \right. \tag {19} +$$ + +Algorithm 1 STDD Training + +1: repeat +2: $Z_{0} \sim q(Z_{0})$ +3: $t \sim \text{Uniform}(1, \dots, T)$ +4: $\epsilon \sim \mathcal{N}(0,I)$ +5: $Z = \sqrt{\bar{\alpha}_t} F_t(Z_0) + \sqrt{1 - \bar{\alpha}_t} F_t(\epsilon)$ +6: Take gradient descent step on +7: $\nabla_{\theta}\mathcal{L}(\theta)$ according to Eqn. (23) +8: until converged + +Algorithm 2 STDD Sampling + +1: $\epsilon \sim \mathcal{N}(0,I)$ +2: $Z_{T} = F_{T}(\epsilon)$ +3: for $t = T$ to 1 do +4: $\epsilon_{\theta} = \epsilon_{\theta}(Z_t,t),\hat{z}_0^i = \phi_\theta (z_t^i,t)$ +5: Calculate $Z_{t-1}$ according to Eqn. (20) +6: end for +7: return $Z_0$ + +We further bring the solved parameters $k_{t}$ and $C_t^i$ into Eqn. (18) to obtain the accelerated inverse inference process of STDD, with setting $\epsilon_t^i = \epsilon_\theta^i$ , expressed as: + +$$ +z _ {t - 1} ^ {i} = \sqrt {\bar {\alpha} _ {t - 1}} (u _ {t} ^ {i} + v _ {t} ^ {i}) + \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \epsilon_ {\theta} ^ {i} + \sigma_ {t} \epsilon^ {\prime} +$$ + +s.t., $u_{t}^{i} = \frac{z_{t}^{i} - \sqrt{1 - \bar{\alpha}_{t}}\epsilon_{\theta}^{i}}{\sqrt{\bar{\alpha}_{t}}},\quad \epsilon_{\theta}^{i} = \epsilon_{\theta}(z_{t}^{i},t),$ (20) + +$$ +v _ {t} ^ {i} = g (t - 1, i) - g (t, i), \quad \tilde {z} _ {0} ^ {i} = \phi_ {\theta} (z _ {t} ^ {i}, t), +$$ + +where $\tilde{\mu}_t^i = \sqrt{\bar{\alpha}_t} * u_t^i$ is the mean of $z_{t}^{i}$ , $v_{t}^{i} \approx F_{t-1}((1 - \sqrt{\rho})\tilde{z}_{0}^{i} - \sqrt{1 - \rho}\tilde{z}_{0}^{i-1})$ is temporal domain difference, $\epsilon'$ is the Gaussian noise added during inference and $\phi_{\theta}$ is a network, then $v_{t}^{i}$ can be calculated by Eqn. (7) with $\tilde{z}_{0}^{i}$ . In practice, $\epsilon_{\theta}$ and $\phi_{\theta}$ are different channel components of the same network output. Moreover, in the process of inference, there is no calculation of the inverse function of $F$ in Eqn. (16). + +How to understand $u$ and $v$ : Assuming a new variable $\tau$ , rewrite $z_{t}^{i}$ in Eqn. (7) as: $z^{i}(t,\tau) = \sqrt{\bar{\alpha}_{t}} g(\tau ,i) + \sqrt{1 - \bar{\alpha}_{t}}\epsilon_{t}^{i}$ . Then, according to SDE model [34], we conclude that + +$$ +\left\{ \begin{array}{l} d _ {t} z ^ {i} (t, \tau) \approx - 1 / 2 \beta_ {t} u _ {t} ^ {i} + \sqrt {\beta_ {t}} d \omega \\ d _ {\tau} z ^ {i} (t, \tau) \approx - \sqrt {\bar {\alpha} _ {t}} v _ {\tau} ^ {i} \end{array} , \right. \tag {21} +$$ + +where $\omega$ gaussian component. Therefore, $u_{t}^{i}$ is image spatial diffusion component and $v_{t}^{i}$ a temporal diffusion or bridge between frames. Therefore, the process inference consists of a diffusion model (i.e., Eqn. (20)) and a difference between frames, shown in Algorithm 2. In addition, to ensure the association between frames, we set $Z_{T} = F(\epsilon)$ . + +# 3.5. Optimize Dual Diffusion + +The dual diffusion architecture includes Variational AutoEncoder (VAE) for video encoding and decoding, diffusion networks $\epsilon_{\theta}$ for denoising, and conditional embedding networks (if needed). In detail, we employ VAE model training in an unsupervised manner, contains two trainable modules: encoder $\mathcal{E}$ to represent image $x \in \mathbb{R}^{H \times W \times 3}$ as latent map $z \in \mathbb{R}^{h \times w \times c}$ and decoder $\mathcal{D}$ to decode the latent map $\tilde{Z}$ as output image $\tilde{X}$ . VAE was trained with the reconstruction loss [28] using only unlabeled images, or load pre-trained model such as Stable Diffusion v1.4 [28]. Then, we use transformer network, used in [21], to extract the feature of video and learn the noise of dual diffusion process. And time step $t$ is specified to $\epsilon_{\theta}$ by the sinusoidal position embedding [38]. In addition, we enhance the network backbone by using a cross-attention mechanism [28] to transform the model into a more flexible conditional diffusion model. And, we adopt different pre-training condition networks according to different tasks, such as CLIP [26] for text-to-video generation. Implementation details will be presented in Sec. 4.1. + +For optimizing STDD, the KL divergence to directly compare $p_{\theta}(Z_{t - 1}|Z_t)$ against forward process posteriors is: + +$$ +\begin{array}{l} D _ {K L} \left(q \left(Z _ {t - 1} \mid Z _ {t}\right) | | p _ {\theta} \left(Z _ {t - 1} \mid Z _ {t}\right)\right) \\ = \frac {\left| \left| \tilde {\mu} _ {t - 1} - \mu_ {\theta} \right| \right| ^ {2}}{2 \sigma_ {t} ^ {2}} + C \\ = \frac {\left| \left| \tilde {\mu} _ {t - 1} - \sqrt {\bar {\alpha} _ {t - 1}} \left(U _ {t} + V _ {t}\right) \right| \right| ^ {2}}{2 \sigma_ {t} ^ {2}} + C \tag {22} \\ \end{array} +$$ + +where $C$ is a constant, $U_{t} = (u_{t}^{0}, u_{t}^{1}, \dots, u_{t}^{N - 1})$ and $V_{t} = (v_{t}^{0}, v_{t}^{1}, \dots, v_{t}^{N - 1})$ . We further optimize $\epsilon_{\theta}$ and $\phi_{\theta}$ respectively, thus the loss of the spatio-temporal diffusion probability model is + +$$ +\mathcal {L} _ {k l} (\theta) = \mathbb {E} | | \tilde {\mu} _ {t - 1} / \sqrt {\bar {\alpha} _ {t - 1}} - (U _ {t} + V _ {t}) | | ^ {2}, +$$ + +$$ +\mathcal {L} _ {m s e} (\theta) = \mathbb {E} | | \boldsymbol {\epsilon} _ {t} - \epsilon_ {\theta} (Z _ {t}, t) | | ^ {2} + | | Z _ {0} - \phi_ {\theta} (Z _ {t}, t) | | ^ {2}, +$$ + +$$ +\mathcal {L} (\theta) = \mathcal {L} _ {k l} (\theta) + \mathcal {L} _ {m s e} (\theta), \tag {23} +$$ + +where $\theta$ represents the parameters of the STDD network. The detailed training process is shown in Algorithm 1. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets: To validate the proposed STDD, we conduct comprehensive evaluations and comparisons on Sky Time-Lapse [47] and UCF101 [35] for video generation, MHAD human action dataset [3] for image-to-video and WebVid-10M [1] for text-to-video. UCF101 is a realistic action video motion recognition dataset, collected from + +![](images/dcb21df5221cc789563342326782cc193ab29d704050cefbe5c6a3a8e4bff9b3.jpg) +Figure 1. Qualitative comparisons on MHAD for video prediction. In the red box, the head in LFDM has serious artifacts and distortions, which we handled well. + +![](images/58e11eaf4822dda8bb2198b2f70b74b926c66bc78279facb91152d1dc466440d.jpg) + +YouTube, providing 13,320 videos from 101 action categories. MHAD human action dataset contains 861 videos of 27 actions performed by 8 subjects. UCF101, Sky Time-Lapse and MHAD are cropped to 16 frames for training our model. For quantitative evaluation, we also train a text-conditioned video generation model on WebVid-10M [1], which consists of 10.7M short videos with paired textual descriptions. + +Metrics: Following previous works [7, 20], we mainly use Frechet Video Distance (FVD) [37] as the evaluation metric. On UCF101 and Sky Time-Lapse, we report the results of FVD. Following LFDM [22] on MHAD, we report the results of FVD and class conditional FVD (cFVD). The cFVD compare the distance between real and synthesized video feature distributions under the same class condition. + +Implementation details: We use the transformer network, used in [21], to extract the features of video and learn the noise of the dual diffusion process. The network learning rate is $1e - 4$ , the batch size is 16, and the training stpes is 200000. The diffusion steps $T = 1000$ , the steps of accelerate reverse process is 250 and $\rho = 1 - 10^{-3}$ . We will provide the generated videos and code on github. + +# 4.2. Comparisons with state-of-the-arts + +Video prediction. On MHAD dataset, we have conducted a comparison of our methods with several competitive approaches, including ImaGINator [41], VDM [13], LDM [28] and LFDM [22]. Table 1 shows the quantitative comparison between our method and state-of-the-art approaches like LDM and LFDM, where our proposed STDD consistently outperforms all other models. The results show + +Table 1. Quantitative comparisons on MHAD. $\downarrow$ denotes the lower the better. The best results are denoted in bold. + +
MethodResolutionFVD ↓cFVD↓
ImaGINator [41]64 × 648891406
VDM [13]64 × 64295531
LDM [28]128 × 128337594
LFDM [22]128 × 128214426
Ours128 × 128195411
+ +Table 2. FVD values of video generation models on UCF101 and Sky Time-Lapse. The lower the better and the best results are denoted in bold. + +
MethodUCF101Sky Time-LapseParams
VideoGPT [48]2880222-
MoCoGAN-HD [36]1729164-
DIGAN [50]163083.1-
StyleGAN-V [31]143179.5-
PVDM [51]114175.5-
LVDM [10]37295.2437M
Latte+IMG [21]33342.7673M
Ours25842.1868M
+ +that our STDD is superior to LFDM (195 vs 214 in FVD at the resolution of $(128 \times 128)$ , and LFDM achieves the second best performance. LFDM distorts the initial frame through an optical flow sequence to generate a new frame to model the time correlation, while our method is mainly to explicitly model the frame time correlation, which is conducive to the propagation of historical frame information. + +Video generation. On UCF101 dataset, we have con + +![](images/f910ceb3c9c5ee772d41fe9bd42edbc27e71b3d07b5b2dbd3f096af635b95286.jpg) +(a)UCF101 (Playing Sitar) + +![](images/789721eea31130d61951b16273775652ed01fc38388515ed40e11a30698931f2.jpg) +(b) Sky Time-Lapse + +duced a comparison of our methods with several competitive approaches, including VideoGPT [48], StyleGAN [31], DIGAN [50], TATS [7], VDM [13], among others. The quantitative results on UCF101 dataset are presented in Table 2. In terms of video generation, VDM demonstrates superior performance compared to GAN-based methods, particularly in terms of FVD (Fréchet Video Distance). This indicates the potential of DPM-based video generation methods. Our proposed STDD achieves even better results than LVDM by a significant margin. Compared with Latte-IMG, which prerained on images, our approach is also comparable. In addition, the quantitative results in Sky Time-Lapse also achieve competitive performance, shown in Table 2. These comparisons highlight the effectiveness of STDD in video generation tasks, demonstrating its superior performance over existing methods in terms of various evaluation metrics and datasets. In addition, the inference time of our model did not increase significantly with the same inference steps and device (23s vs 20s of Latte with 250 inference steps in RTX 3090 GPU). + +# 4.3. Ablation Study + +Visual comparison. On MHAD dataset, we have conducted a qualitative comparison of our methods with several video pridiction models, including ImaGINator, VDM, LDM and LFDM. As shown in Fig. 1, the first line shows the given image, and the rest shows the generated sequence. In general, our approach has advantages over other video generation methods, especially generating 'stand-to-sit' videos with larger variation areas and more motion variations. The results in Fig. 1 (b) showed that the head of the LFDM generated character had severe artifacts and distor + +![](images/282f016cbaed3a84d1927f8c2986f86f87d239f20bac1550bb42b9c4eadd8cf6.jpg) +Figure 2. Qualitative comparisons on UCF101 and Sky Time-Lapse for video generation. + +![](images/2930b40705e576cbe1c2b6cfbc196057282fd4967b0b0b412e04a257e7cb6fdc.jpg) + +![](images/5b99d0f0475969e6f631bef24383dcd4baee9ffef2f91932f00092101dc6f537.jpg) +Frame weights + +![](images/627fdcb7974ffff42fd86ae3f52197e434949a688392dd22bd56d66c0bd5cf8a.jpg) +$g(t = 1000,i)$ +Figure 3. The tendency of frame weights in Eqn. (7) and Eqn. (8). Appropriate $\rho$ enhances the similarities and blurring the differences between frames. + +tions, which we handled well, notably in the red box. We also provide visual comparisons with the most recent video generation methods, i.e. TATS, DIGAN, VideoFusion and Latte. As shown in Fig. 2, each generated video has 16 frames with a resolution of $256 \times 256$ , and we show 4-th, 8-th, 12-th and 16-th frame. Our STDD could generate more continuous and moving videos. + +Analysis of $\rho$ . To explore the effects of $\rho$ in Eqn (7), we conducted experiments on MHAD with a resolution of $64 \times 64$ , shown in Table 3. Appropriate $\rho$ can better integrate the historical frames and improve the performance. Also, smaller $\rho$ leads to information loss of the + +![](images/68f7ede029e27228502f78c70961adb77cb65b26b979118c898af2973448436d.jpg) + +![](images/d49240d508471c2f7c42b569192af13a89ce2490fa7777734399179096e66f3a.jpg) +(a) A dog in astronaut suit and sunglasses floating in space +(b) A turtle swimming in the ocean + +![](images/7745bd384cfb82ad6282abda5bc515a93654ec0209e99ae37a89d4394dd9c522.jpg) +(c) Surfer on his surfboard in a wave +Figure 4. Qualitative evaluation on text-to-video generation. Both our method and Videofusion are trained on WebVid-10M, while Latte is trained on Vimeo25M [42]. + +Table 3. Analysis $\rho$ in Eqn. (7) on MHAD with the resolution of $64\times 64$ . Appropriate $\rho$ can improve the performance and $\sqrt{\rho^{1000}}$ represents the weight of the current frame after 1000 steps. + +
ρ10.99990.9990.9950.99
\( \sqrt{\rho^{1000}} \)10.9510.6060.0820.007
FVD ↓157149133243388
+ +current frame, which is not conducive to video generation. Since the coefficient is exponential, it must be required that $0 < 1 - \rho < 0.01$ , which consistent with the hypothesis in Eqn. (8). The weights of the current and historical frames are shown in Fig. 3. Visualization shows that $g(t,i)$ enhances the similarities and blurring the differences between frames. + +Text-to-video generation. To explore the video generation capabilities of our proposed approach, we extend STDD to text-to-video generation. We train a text-to-video model, pretrained by Stable Diffusion [28], on the large scale video dataset, i.e. WebVid-10M. Some samples are shown in Fig. 4. + +# 5. Conclusion + +In this paper, we propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes. Then, we theoretically derive probabilistic reverse process and propose an accelerated sampling method. Both qualitative and quantitative experiments demonstrated that our model achieve more competitive performance on video generation. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 6 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[3] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 5 + +[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2 +[6] Yu Gao, Jiancheng Huang, Xiaopeng Sun, Zequn Jie, Yujie Zhong, and Lin Ma. Matten: Video generation with mamba-attention. arXiv preprint arXiv:2405.03025, 2024. 3 +[7] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In European Conference on Computer Vision, pages 102–118. Springer, 2022. 2, 6, 7 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[9] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 1 +[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.6 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 1, 2, 3, 6, 7 +[14] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[15] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2 +[16] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2 +[17] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 1 +[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2 + +[19] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2 +[20] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. CoRR, abs/2303.08320, 2023. 1, 2, 6 +[21] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 5, 6 +[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18444-18455. IEEE, 2023. 6 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1 +[25] Xichen Pan, Pengda Qin, Yuhong Li, Hui Xue, and Wenhu Chen. Synthesizing coherent story with auto-regressive latent diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2920-2930, 2024. 1 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5, 6, 8 +[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1 + +[30] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2 +[31] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 6, 7 +[32] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2 +[34] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2, 5 +[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 5 +[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2, 6 +[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems, 35:23371-23385, 2022. 1, 2 +[40] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2 +[41] Yaohui Wang, Piotr Bilinski, Francois Bremond, and Antitza Dantcheva. Imaginator: Conditional spatio-temporal gan for video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1160-1169, 2020. 6 +[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 8 +[43] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 22025-22034, 2023. 2 +[44] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023. 2 +[45] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024. 2 +[46] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2 +[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5 +[48] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2, 6, 7 +[49] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25(10):1469, 2023. 1 +[50] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 6, 7 +[51] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 6 \ No newline at end of file diff --git a/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/images.zip b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db15886ed30e80f4f4a6a99b6f9d2816ddb0568e --- /dev/null +++ b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05c9061dfc35dd32458533e01f9ecf2ea27fb4111efe5efd06618d7a7e178884 +size 696107 diff --git a/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/layout.json b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5824e65b3699b47d18ae50e2faa2fa7247206192 --- /dev/null +++ b/2025/STDD_ Spatio-Temporal Dual Diffusion for Video Generation/layout.json @@ -0,0 +1,10867 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 121 + ], + "type": "text", + "content": "STDD: Spatio-Temporal Dual Diffusion for Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": "Shuaizhen Yao" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": ", Xiaoya Zhang" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": ", Xin Liu" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": ", Mengyi Liu" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": ", Zhen Cui" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{4,*}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": " Nanjing University of Science and Technology, \n" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": " Nanjing Seetacloud Technology Co., Ltd, \n" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": " Kuaishou, " + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 142, + 478, + 227 + ], + "type": "text", + "content": " Beijing Normal University \n{ysz, zhangxiaoya}@njust.edu.cn, xin.liu@seetacloud.com, liumengyi@kuaishou.com, zhen.cui@bnu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 280, + 297, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 297, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 297, + 543 + ], + "type": "text", + "content": "Diffusion probabilistic model is becoming the cornerstone of data generation, especially generating high-quality images. As an extension, video diffusion generation is in urgent need of a principled temporal-sequence diffusion way, while the spatial-domain diffusion dominates most video diffusion methods. In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal noise propagation/reduction. Mathematically, an analysable dual diffusion process is derived to accumulate information in temporal sequence as well as spatial domain. Correspondingly, we theoretically derive a spatio-temporal probabilistic reverse diffusion process and propose an accelerated sampling way to reduce the inference cost. In principle, the spatio-temporal dual diffusion enables the information of previous frames to be transferred to the current frame, which thus could be beneficial for video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation/prediction as well as text-to-video generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "type": "text", + "content": "Diffusion probabilistic model (DPM) is becoming the cornerstone of data generation, where the forward process gradually contaminates the data with gaussian noises and the reverse denoising process is learned to generate new samples [12, 15]. Specifically, DPM has made significant progress in producing high quality, realistic images [17, 23, 24, 27, 29, 33]. Motivated by the success in image generation, some researchers are trying to extend the diffusion model from image to video by accompanying with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 256, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 304 + ], + "type": "text", + "content": "temporal-dimensional prediction/interpolation [9, 39, 49]. However, due to intricate temporal and spatial correlations as well as high dimensionality of video, video diffusion generation remains a challenging task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 508 + ], + "type": "text", + "content": "The existing video generation methods based on DPMs [9, 13, 28] typically employ a standard diffusion process by treating each frame of the video as an independent diffusion process, which breaks the temporal correlation between frames. Concretely, in the noisig/denoising process, these video diffusion methods separately derive/sample priors, and recover corresponding frames from different-random priors. Due to different-random priors, the continuity of generated video frames is still a bottleneck of video diffusion generation, even though some historical (generated) frames could be used as a condition of the diffusion for some-degree mitigation [25]. An endeavor to this aspect is one recent work [20], decomposing video noise into two components: video sharing base noise and frame-independent residual noise, and then separately performing the diffusion process with two different diffusion networks, which greatly improves video consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 617 + ], + "type": "text", + "content": "Despite some progress in video diffusion generation, these existing methods could not well capture temporal dependencies due to the use of independent priors for each frame, which would often result into the nonsmoothness or inconsistency of generated videos, especially confronting with large appearance variations in temporal sequences. Essentially, the current video diffusion methods lack a principled diffusion process along time slices (w.r.t frame sequence), not just in the spatial domain (w.r.t every image)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "type": "text", + "content": "In this work, we propose an explicit Spatio-Temporal Dual Diffusion (STDD) method by principally extending the standard diffusion model to a spatio-temporal diffusion model for joint spatial and temporal frame propagation. Mathematically, we derive an analysable dual diffusion process by accumulating temporal frame information. In particular, the historical frames are propagated along with temporal sequence, while accompanying with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12575" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "type": "text", + "content": "the spatio-domain noises w.r.t diffusion on single-frame image. To reach the inverse diffusion, we derive a probabilistic reverse process, which guarantees the feasibility of the diffusion model. Inspired by the diffusion probabilistic implicit diffusion (DDIM) model [33], further, we propose an accelerated spatio-temporal sampling method that is both fast and computation-intensive. To facilitate learning, we decompose the model into spatial components representing the standard diffusion model and temporal components representing the frame-to-frame diffusion. After a series of derivation, we could reach a maneuverable solution by using neural network. Overall, the spatio-temporal diffusion enables the information of previous frames to be recursively transferred to the current frame, which endows the model good ability to capture temporal dependencies, and can be robust to video consistency. Extensive experiments demonstrate that our proposed STDD is more competitive over the state-of-the-art methods in the task of video generation and prediction. In addition, we extend our STDD to text-to-video generation to explore the video generation capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 406 + ], + "type": "text", + "content": "In summary, our main contributions are three folds: i) propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes; ii) theoretically derive probabilistic reverse process and propose an accelerated sampling method; iii) achieve more competitive performance on video generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 418, + 148, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 148, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 148, + 430 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 220, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 220, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 220, + 449 + ], + "type": "text", + "content": "2.1. Diffusion Probabilistic Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": "DPM, which was first introduced in reference [32], comprises a diffusion (encoding) process and a denoising (decoding) process. In the diffusion process, given one image " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": ", the noise diffusion process is conducted, which random noises are gradually added by T-step Markov chain [15], denoted as " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "(x_0,x_1,\\dots ,x_T)" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 536, + 294, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 536, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 107, + 536, + 294, + 585 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\\\ \\text {s . t .} \\quad \\bar {\\alpha} _ {t} = \\prod_ {k = 1} ^ {t} \\alpha_ {t}, \\epsilon_ {t} \\sim \\mathcal {N} (0, I), \\tag {1} \\\\ \\end{array}", + "image_path": "cc260ff0c22c046a9c88f9a79d5bb06008183376f7bffb5fd11b775edaece7ae.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\in (0,1)" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " is the corresponding diffusion coefficient. If " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " is large enough, such as " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " approximates random Gaussian noise. Therefore, the reverse process of the diffusion model (denoising process) can be interpreted as the gradual generation of the data distribution from random Gaussian noise. In other words, it aims to recover the original data by removing the added noise step by step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Ho et al. [12] establish a connection between DPM and denoising score matching [34] and propose an " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "-prediction formulation for the denoising process, which is " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "L = ||\\epsilon_t -" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)||^2" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ". In which, " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " is a denoising neural network parameterized by " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " is the loss function. Utilizing this formulation, DPM has been successfully applied to various generative tasks, including image generation [11, 19, 27], super-resolution [16], image translation [30], and more [43, 44]. Compared to other generation methods, such as Generative Adversarial Networks (GANs) [8], DPMs are easier to train and can generate more diverse samples [4]. DPMs utilize diffusion processes and de-noising steps, resulting in a more stable training process and improved sample diversity. These characteristics have contributed to the popularity and success of DPMs in the field of generative modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 447, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 447, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 447, + 251 + ], + "type": "text", + "content": "2.2. Video Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 257, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 473 + ], + "type": "text", + "content": "Video generation is a highly challenging task in the field of generative research. It not only requires the generation of high-quality frames but also necessitates temporal correlation among the generated frames. Previous video generation methods have predominantly relied on Generative Adversarial Networks (GANs). In VGAN [40], the generator directly learns the joint distribution of video frames. Mocogan [36] decomposed a video clip into content and motion components, which are individually modeled using a content vector and a motion vector. Alternatively, other methods utilize a two-step process, which train a vector quantized autoencoder [5] for video data, and then employ an transformer network to learn the video distribution in the quantized latent space [7, 14, 48]. These techniques leverage the quantized latent space to model the video distribution and capture dependencies among frames. Overall, each approach has its own advantages and limitations in generating high-quality continuous videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 714 + ], + "type": "text", + "content": "In recent years, researchers have been motivated by the remarkable achievements of Diffusion Probabilistic Models (DPM) in image generation and have made efforts to apply DPM to video generation. In reference [13], Ho et al. proposed a video diffusion model that extends the 2D denoising network from the image diffusion model to 3D by considering stacked frames as additional dimensions. This enables the model to capture spatio-temporal dependencies and generate coherent videos. In [39], DPM is utilized for video prediction and interpolation, where known frames are used as conditions for denoising. Some large video generation models [2, 18, 45, 46], with massive data and parameters, allow the attention layer to implicitly learn the relationship between video frames, which requires significant computing and storage resources. These methods often treat video frames as independent samples during the diffusion process. To establish a relation between the video frames, VideoFusion [20] breaks down the noise into the base noise for video sharing and the residual noise for the frames. Benefit from powerful time modeling capabilities," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12576" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "transformer-based method [6, 21] is also used to generate high quality video. Still, these methods fail to establish temporal associations between frames, especially adjacent frames, may lead to limitations in generating videos with smooth transitions and preserving temporal consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 236, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 236, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 236, + 156 + ], + "type": "text", + "content": "3. Spatio-Temporal Dual Diffusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 211 + ], + "type": "text", + "content": "In this section, we will illustrate our main idea, and then derive an analysable spatio-temporal forward and inverse diffusion process, meantime design an accelerated inverse diffusion version and its optimization process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 176, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 176, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 176, + 230 + ], + "type": "text", + "content": "3.1. Problem Description" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "X = (x^0, x^1, \\dots, x^i, \\dots, x^{N-1})" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": " denotes a video sequence with " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": " frames, where the superscript " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": "-th video frame. We also follow the customary use of taking subscript as diffusion step, i.e., the symbol " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "x_t^i" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": " with the diffusion step " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": "-th frame. The existing video diffusion methods usually take an independent frame-wise diffusion way through leveraging the classic diffusion model such as denoising diffusion probabilistic model [13]. Given one frame " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": ", the noise diffusion process is conducted, denoted as " + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "inline_equation", + "content": "(x_0^i, x_1^i, \\dots, x_T^i)" + }, + { + "bbox": [ + 54, + 234, + 296, + 356 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 362, + 295, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 362, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 114, + 362, + 295, + 377 + ], + "type": "interline_equation", + "content": "x _ {t} ^ {i} = \\sqrt {\\alpha_ {t}} x _ {t - 1} ^ {i} + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t} ^ {i}, \\tag {2}", + "image_path": "d1956d8e39c17cda68cce5aca8c750c3b6920995111032fbb9de07448aaf45be.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\epsilon_t^i\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\beta_{t} > 0" + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "text", + "content": " . The inverse denoising process could be recursively estimated by an optimized neural network " + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\pmb{\\epsilon}" + }, + { + "bbox": [ + 55, + 384, + 296, + 422 + ], + "type": "text", + "content": " , formally," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 429, + 295, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 429, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 94, + 429, + 295, + 455 + ], + "type": "interline_equation", + "content": "x _ {t - 1} ^ {i} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(x _ {t} ^ {i} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t} ^ {i}, t\\right)\\right), \\tag {3}", + "image_path": "0b75d093183fd5611278d40211a5dddc046a126817ddbf3fe73a6fcbfb91ca86.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\Pi_{s=1}^t\\alpha_s" + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "content": ", and the initial noisy " + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "inline_equation", + "content": "x_T^i" + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "content": " of each frame is also sampled independently from Gaussian noise. In essence, the previous video diffusion methods basically perform a combination of " + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 463, + 296, + 570 + ], + "type": "text", + "content": "-frame independent image diffusion processes, although it is usually operated in the hidden feature space in order to enhance the inter-frame consistency. However, the inherent frame-sequential diffusion correlations have not yet been explicitly considered by previous diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 571, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 631 + ], + "type": "text", + "content": "Here we consider that the diffusion process should not only be imposed on one single image frame, but only be run on the stream of video frames, named joint spatial and temporal diffusion. From the view of mathematical formulation, we abstract the spatio-temporal dual diffusion as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 638, + 295, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 638, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 110, + 638, + 295, + 653 + ], + "type": "interline_equation", + "content": "x _ {t} ^ {i} = \\mathbb {D} \\left(x _ {t - 1} ^ {i}, x _ {t - 1} ^ {i - 1}, \\epsilon_ {t} ^ {i}\\right), \\tag {4}", + "image_path": "6679afb2191348d299f732561d2e9f234816bcf74b29cb95ca3fb32efd705bbc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 654, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 654, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 100, + 654, + 295, + 670 + ], + "type": "interline_equation", + "content": "x _ {t - 1} ^ {i} = \\mathbb {D} ^ {- 1} \\left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}, \\epsilon_ {\\theta} \\left(x _ {t} ^ {i}, x _ {t - 1} ^ {i - 1}\\right)\\right), \\tag {5}", + "image_path": "d3c7aa23324cc846226a1c04a9dd1c16f30a9b81e7dc5efaa68ad648412dace3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where the forward and inverse diffusion processes, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^{-1}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", are well derived in analyzed mathematical expressions (see below sections), and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is a learnable function with the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "parameter " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". The STDD essentially represents the dependant relations of continuous frame slices as well as the diffusion steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 115, + 450, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 450, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 450, + 127 + ], + "type": "text", + "content": "3.2. Forward Dual Diffusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": "In order to model the relations between video frames, we also use temporal Markov chain to describe the evolution dependency on time series, in addition to image diffusion process itself. Hereby, a spatio-temporal diffusion model could be built. To distinguish it from standard diffusion, let " + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": " denotes the noised latent variable of " + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": " at the diffusion step " + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 133, + 555, + 229 + ], + "type": "text", + "content": ". According to the diffusion mechanism, we define the forward spatio-temporal noising process as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 236, + 555, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 236, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 321, + 236, + 555, + 285 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t} ^ {i} \\sim p \\left(z _ {t} ^ {i} \\mid z _ {t - 1} ^ {i}, z _ {t - 1} ^ {i - 1}\\right) \\tag {6} \\\\ = \\sqrt {\\alpha_ {t}} (\\underbrace {\\sqrt {\\rho_ {t}} * z _ {t - 1} ^ {i} + \\sqrt {1 - \\rho_ {t}} * z _ {t - 1} ^ {i - 1}} _ {\\text {t e m p o r a l d o m a i n}}) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t} ^ {i} \\\\ \\end{array}", + "image_path": "ec3ded74d8429b5ad3d6b4d8453bc6b798141037cd3d1ed9fa9fc5ffa52f2c8a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "text", + "content": "where the diffusion coefficient " + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "inline_equation", + "content": "\\rho_t \\in (0,1]" + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "text", + "content": " is used to balance the prior frame and the current frame. In the above equation, the temporal-domain term denotes the temporal diffusion between the " + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "text", + "content": "-th frames, and " + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "inline_equation", + "content": "\\sqrt{\\alpha_t(\\cdot)} + \\sqrt{1 - \\alpha_t} \\epsilon_t^i" + }, + { + "bbox": [ + 313, + 293, + 554, + 377 + ], + "type": "text", + "content": " is an image-like diffusion. In order to conveniently operate the noising process, we derive an analyzed expression as shown in the following Proposition 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "content": "Proposition 1. According to the recursive formula (6), and suppose " + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\rho_0 = \\rho_1 = \\dots = \\rho_t = \\rho" + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "content": ", we compute the diffusion state of any time frame and any diffusion step from the initial states (" + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "inline_equation", + "content": "z_0^0 = x^0, z_0^1 = x^1, \\dots, z_0^i = x^i" + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 334, + 441, + 554, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 441, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 334, + 441, + 554, + 472 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t} ^ {i} \\sim p \\big (z _ {t} ^ {i} | \\{z _ {0} ^ {k} | k = 1, 2,.., i \\} \\big) \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} g (t, i) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} ^ {i}, \\tag {7} \\\\ \\end{array}", + "image_path": "22d048877b19e04a4355b3b681ebece70f3bf5990df59dbbc8fde00b729802ea.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 324, + 475, + 543, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 475, + 543, + 509 + ], + "spans": [ + { + "bbox": [ + 324, + 475, + 543, + 509 + ], + "type": "interline_equation", + "content": "s. t. g (t, i) = \\sum_ {m = 0} ^ {m i n (t, i)} {\\binom {t} {m}} \\rho^ {(t - m) / 2} (1 - \\rho) ^ {m / 2} z _ {0} ^ {i - m},", + "image_path": "4ae603c9ff8f66f51ab851cfecad3b7c64e201d16fa8f22928a6ea357f9abf0a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "text", + "content": "where we only retain the existing terms " + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "inline_equation", + "content": "z_0^{i - m}" + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "text", + "content": " w.r.t " + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "inline_equation", + "content": "m \\leq i" + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "inline_equation", + "content": "g(t,i)" + }, + { + "bbox": [ + 313, + 517, + 554, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 550, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 574 + ], + "type": "text", + "content": "Proof. The derivation process is deferred to the supplementary material." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "type": "text", + "content": "In Proposition 1, the dual diffusion process integrates the information of time slices and noising steps jointly. The historical frame diffusion term " + }, + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "type": "inline_equation", + "content": "g(t,i)" + }, + { + "bbox": [ + 313, + 582, + 554, + 629 + ], + "type": "text", + "content": " accumulates the information of previous historical frames." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "It is worth noting that, the standard diffusion process is a special case of the above model when setting " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\rho = 1" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". For the initial frame, i.e., " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "i = 0" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ", we consider it as a standard image diffusion process, i.e., " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "z_{t}^{0} = \\sqrt{\\bar{\\alpha}_{t}} z_{0}^{0} + \\sqrt{1 - \\bar{\\alpha}_{t}} \\epsilon_{t}^{0}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". The probability analysis of the inverse spatio-temporal diffusion is given in Section 3.3, and a practical inverse diffusion can be found directly in Section 3.4." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12577" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "In intuitive, it seems require a high-computational burden, but in practice the combinatorial coefficient term " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\left\\{\\binom{t}{m}\\right\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " could be pre-calculated as a searching table, which would reduce the running cost. In addition, if " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "1 - \\rho = < 0.01" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\rho^{(t - m / 2)} \\approx e^{-t(1 - \\rho) / 2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "g(t,i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " also can be written as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 150, + 296, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 150, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 66, + 150, + 296, + 184 + ], + "type": "interline_equation", + "content": "g (t, i) = \\sum_ {m = 0} ^ {m i n (t, i)} \\binom {t} {m} (1 - \\rho) ^ {m / 2} e ^ {- t (1 - \\rho) / 2} z _ {0} ^ {i - m} \\tag {8}", + "image_path": "f10aebac38f84914caf415e0d920a4dbcb19dc160a250639bf38036732f4363d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "type": "text", + "content": " is limited to [0, 16], which require small running cost. Notice, the above coefficient substitution is only for the purpose of simplifying the calculation in the experiment and is independent of our model. And the setting of the parameter " + }, + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 191, + 296, + 252 + ], + "type": "text", + "content": " will be discussed in the ablation experiment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 185, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 185, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 185, + 270 + ], + "type": "text", + "content": "3.3. Inverse Dual Diffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 276, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 370 + ], + "type": "text", + "content": "Probability diffusion process can be reduced to a constrained latent variable model whose Gaussian distribution of forward and reverse processes must be modeled. To derive the inverse decoding process, we establish an explicit link between spatio-temporal diffusion probabilistic models and inverse probability distributions, and then provide a variational optimization objective for the proposed dual diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": " denotes a set of all frames at the diffusion step " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "Z_{t} = (z_{t}^{0}, z_{t}^{1}, \\dots, z_{t}^{N-1})" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": ". For convenient statement, we use a surrogate operator " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "F_{t}: Z_{0} \\to \\widetilde{Z}_{t}" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": " and its inverse operator " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "F_{t}^{-1}: \\widetilde{Z}_{t} \\to Z_{0}" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": " as well as " + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\widetilde{Z}_{t}" + }, + { + "bbox": [ + 55, + 371, + 296, + 432 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 439, + 295, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 439, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 90, + 439, + 295, + 454 + ], + "type": "interline_equation", + "content": "\\widetilde {Z} _ {t} = F _ {t} \\left(Z _ {0}\\right) := \\left(\\widetilde {z} _ {t} ^ {0}, \\dots , \\widetilde {z} _ {t} ^ {i}, \\dots , \\widetilde {z} _ {t} ^ {N - 1}\\right), \\tag {9}", + "image_path": "a7b158701d32774f42b4a41479b6d5b21b4d8f1ad237c1cbf7ca82a254e1c15a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 457, + 294, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 457, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 100, + 457, + 294, + 471 + ], + "type": "interline_equation", + "content": "\\text {s . t .}, \\quad \\widetilde {z} _ {t} ^ {i} = g (t, i). \\tag {10}", + "image_path": "d86189e515e8b3b9461dc51f3c5a0eaf6cd124a8cee7df5ad9b49d3bbd75cf9e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "content": " represents the diffusion process between video frames, which is called temporal diffusion. Then, according to Eqn. (6) and Proposition 1, the variable " + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "content": " can be written as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 521, + 295, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 521, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 91, + 521, + 295, + 553 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} F _ {t} \\left(Z _ {0}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {11} \\\\ = \\sqrt {\\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) + \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t}, \\qquad (1 2) \\\\ \\end{array}", + "image_path": "8bec047b32cfaeec473ddc64685cb97c1dc743ac1dc7e2298ce49ee3234db54c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "content": "where the noise " + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} = (\\epsilon_{t}^{0},\\epsilon_{t}^{1},\\dots,\\epsilon_{t}^{i})" + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "content": ", and Eqn. (12) is the iterative formula and the Markov chain of the forward diffusion process. Therefore, the conditional probability distribution of the forward diffusion process can be written as a Gaussian distribution:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 627, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 627, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 627, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q (Z _ {t} | Z _ {t - 1}, Z _ {0}) = \\mathcal {N} (Z _ {t} | \\sqrt {\\alpha_ {t}} F _ {t} (F _ {t - 1} ^ {- 1} (Z _ {t - 1})) (1 - \\alpha_ {t}) I), \\\\ q \\left(Z _ {t} \\mid Z _ {0}\\right) = \\mathcal {N} \\left(Z _ {t} \\mid \\sqrt {\\bar {\\alpha} _ {t}} F _ {t} \\left(Z _ {0}\\right), (1 - \\bar {\\alpha} _ {t}) I\\right). \\tag {13} \\\\ \\end{array}", + "image_path": "0316d4ca79b76815c9f32ae9b369664bba3b522931a217810762b03e118337f7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Then, video diffusion model can be modeled as some latent variables with the form: " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{\\theta}(Z_0) = \\int p_{\\theta}(Z_{0:T})d_{Z_{1:T}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", where the subscript " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\square_{1:T}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " means the collection of latent variables from step 1 to step " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the learnt parameter. The" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "joint distribution " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "p_{\\theta}(Z_{0:T})" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " refers to the reverse process, and it can be defined as a Markov chain with learned Gaussian transitions starting at " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "p(Z_T) = \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 117, + 555, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 117, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 356, + 117, + 555, + 133 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (Z _ {0: T}) = p (Z _ {T}) \\Pi_ {t = 1} ^ {T} p _ {\\theta} (Z _ {t - 1} | Z _ {t}). \\tag {14}", + "image_path": "93c75f7bbf39cb205db4a3df4e99347d752889f9dde03ea3f6c3d85023d96d15.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "type": "text", + "content": "Therefore, we need to derive " + }, + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "type": "inline_equation", + "content": "p_{\\theta}(Z_{t - 1}|Z_t)" + }, + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "type": "text", + "content": " as forward process posteriors, which are tractable when conditioned on " + }, + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "type": "inline_equation", + "content": "Z_0" + }, + { + "bbox": [ + 313, + 140, + 555, + 175 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 183, + 555, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 183, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 335, + 183, + 555, + 199 + ], + "type": "interline_equation", + "content": "q \\left(Z _ {t - 1} \\mid Z _ {t}\\right) = q \\left(Z _ {t - 1} \\mid Z _ {t}, Z _ {0}\\right) = \\mathcal {N} \\left(\\tilde {\\mu_ {t}}, \\tilde {\\beta_ {t}}\\right), \\tag {15}", + "image_path": "0e6cb6a671294c3966f1225b73a9d9904ba3b2db8de82a9701927315c6a93408.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 207, + 554, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 554, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 554, + 256 + ], + "type": "text", + "content": "where we assume the Gaussian distribution in the last equation. Then, with the help of Eqn. (13), we obtain the mean and variance of the posterior distribution of the diffusion process " + }, + { + "bbox": [ + 313, + 207, + 554, + 256 + ], + "type": "inline_equation", + "content": "q(Z_{t - 1}|Z_t) = N(\\tilde{\\mu}_t,\\tilde{\\beta}_t)" + }, + { + "bbox": [ + 313, + 207, + 554, + 256 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 264, + 554, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 264, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 326, + 264, + 554, + 316 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\tilde {\\mu} _ {t} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(F _ {t - 1} \\left(F _ {t} ^ {- 1} \\left(Z _ {t}\\right)\\right) - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\boldsymbol {\\epsilon} _ {t}\\right), \\\\ \\tilde {\\beta} _ {t} = \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}} (1 - \\alpha_ {t}), \\end{array} \\right. \\tag {16}", + "image_path": "410f7c22847149c4a91a9c87f6e65d7d90292b3c7295072ab16c42dbc55afbbb.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 324, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 336 + ], + "type": "text", + "content": "where the relevant proof is given in the supplement material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 343, + 508, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 508, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 508, + 355 + ], + "type": "text", + "content": "3.4. Accelerate Inverse Inference Process" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "type": "text", + "content": "In order to boost sampling speed, we derive an implicit probability diffusion model. The denoising diffusion probabilistic process, in section 3.3, requires the construction of Markov chains with the help of the additional operator " + }, + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "type": "inline_equation", + "content": "F^{-1}" + }, + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "type": "text", + "content": ", and more computational complexity. But, denoising diffusion implicit models has faster sampling speeds and does not require Markov chains. Below we do a derivation to fulfill the spatio-temporal inverse diffusion process with DDIM, which enables more practical use." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "type": "text", + "content": "First, according to Proposition 1, we make reasonable assumptions for the reverse process" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 353, + 502, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 502, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 353, + 502, + 553, + 517 + ], + "type": "interline_equation", + "content": "z _ {t - 1} ^ {i} = k _ {t} z _ {t} ^ {i} + C _ {t} ^ {i} \\left(Z _ {0} ^ {0: i - 1}\\right) + \\sigma_ {t} \\epsilon_ {t - 1}, \\tag {17}", + "image_path": "36cca80c8cd2693fef0557af73f35fdfc72585b64ccf60a3932a27d1f80b108e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "k_{t}" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": " are coefficient to be determined, " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "C_t^i (Z_0^{0:i - 1})" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": " is a function (also written as " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "C_t^i" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\prime}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": ". Then, according to Proposition 1 and Eqn. (7), we replace " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": " with an expression of " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "Z_{0}" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\epsilon_t^i" + }, + { + "bbox": [ + 313, + 521, + 554, + 571 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 325, + 579, + 553, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 579, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 325, + 579, + 553, + 599 + ], + "type": "interline_equation", + "content": "z _ {t - 1} ^ {i} = k _ {t} \\sqrt {\\bar {\\alpha} _ {t}} g (t, i) + C _ {t} ^ {i} + \\sqrt {k _ {t} ^ {2} \\left(1 - \\bar {\\alpha} _ {t}\\right) + \\sigma_ {t} ^ {2}} \\epsilon \\tag {18}", + "image_path": "fabf494b091f4df6068a155014a4e1236244c64b5b16a690f66315b54c95372c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "text", + "content": "Compared with Eqn. (7) in Proposition 1 (i.e., " + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\sqrt{\\bar{\\alpha}_{t-1}} g(t-1,i)+\\sqrt{1-\\bar{\\alpha}_{t-1}}\\epsilon" + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "text", + "content": "), we can solve the coefficients " + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "inline_equation", + "content": "k_{t}" + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "text", + "content": " and function " + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "inline_equation", + "content": "C_{t}^{i}(Z_{0}^{0:i-1})" + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "text", + "content": " with conditional of " + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 607, + 554, + 644 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 661, + 554, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 661, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 661, + 554, + 712 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} k _ {t} = \\sqrt {\\frac {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}}{1 - \\bar {\\alpha} _ {t}}}, \\\\ C _ {t} ^ {i} \\left(Z _ {0} ^ {0: i - 1}\\right) = \\sqrt {\\bar {\\alpha} _ {t - 1}} g (t - 1, i) - k _ {t} \\sqrt {\\bar {\\alpha} _ {t}} g (t, i). \\end{array} \\right. \\tag {19}", + "image_path": "c01785bba6eb661f80ccd99b5a1897c423ff04fe56835b7ab22f5cae2f94509e.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12578" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 176, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 176, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 176, + 84 + ], + "type": "text", + "content": "Algorithm 1 STDD Training" + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 62, + 88, + 227, + 182 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 62, + 88, + 103, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 88, + 103, + 99 + ], + "spans": [ + { + "bbox": [ + 62, + 88, + 103, + 99 + ], + "type": "text", + "content": "1: repeat" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 99, + 138, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 99, + 138, + 110 + ], + "spans": [ + { + "bbox": [ + 62, + 99, + 138, + 110 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 62, + 99, + 138, + 110 + ], + "type": "inline_equation", + "content": "Z_{0} \\sim q(Z_{0})" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 111, + 185, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 111, + 185, + 122 + ], + "spans": [ + { + "bbox": [ + 62, + 111, + 185, + 122 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 62, + 111, + 185, + 122 + ], + "type": "inline_equation", + "content": "t \\sim \\text{Uniform}(1, \\dots, T)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 124, + 140, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 124, + 140, + 134 + ], + "spans": [ + { + "bbox": [ + 62, + 124, + 140, + 134 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 62, + 124, + 140, + 134 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "type": "inline_equation", + "content": "Z = \\sqrt{\\bar{\\alpha}_t} F_t(Z_0) + \\sqrt{1 - \\bar{\\alpha}_t} F_t(\\epsilon)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 148, + 208, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 208, + 158 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 208, + 158 + ], + "type": "text", + "content": "6: Take gradient descent step on" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 159, + 215, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 215, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 215, + 171 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 62, + 159, + 215, + 171 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}(\\theta)" + }, + { + "bbox": [ + 62, + 159, + 215, + 171 + ], + "type": "text", + "content": " according to Eqn. (23)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 171, + 138, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 138, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 138, + 182 + ], + "type": "text", + "content": "8: until converged" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 200, + 179, + 213 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 179, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 179, + 213 + ], + "type": "text", + "content": "Algorithm 2 STDD Sampling" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 62, + 216, + 243, + 299 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 62, + 216, + 124, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 124, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 124, + 228 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 62, + 216, + 124, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 228, + 126, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 126, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 126, + 239 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 62, + 228, + 126, + 239 + ], + "type": "inline_equation", + "content": "Z_{T} = F_{T}(\\epsilon)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 240, + 145, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 145, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 145, + 251 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 62, + 240, + 145, + 251 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 62, + 240, + 145, + 251 + ], + "type": "text", + "content": " to 1 do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 251, + 208, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 208, + 264 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 208, + 264 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 62, + 251, + 208, + 264 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta} = \\epsilon_{\\theta}(Z_t,t),\\hat{z}_0^i = \\phi_\\theta (z_t^i,t)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 265, + 243, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 265, + 243, + 276 + ], + "spans": [ + { + "bbox": [ + 62, + 265, + 243, + 276 + ], + "type": "text", + "content": "5: Calculate " + }, + { + "bbox": [ + 62, + 265, + 243, + 276 + ], + "type": "inline_equation", + "content": "Z_{t-1}" + }, + { + "bbox": [ + 62, + 265, + 243, + 276 + ], + "type": "text", + "content": " according to Eqn. (20)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 277, + 106, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 106, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 106, + 286 + ], + "type": "text", + "content": "6: end for" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 62, + 289, + 116, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 116, + 299 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 116, + 299 + ], + "type": "text", + "content": "7: return " + }, + { + "bbox": [ + 62, + 289, + 116, + 299 + ], + "type": "inline_equation", + "content": "Z_0" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "content": "We further bring the solved parameters " + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "inline_equation", + "content": "k_{t}" + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "inline_equation", + "content": "C_t^i" + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "content": " into Eqn. (18) to obtain the accelerated inverse inference process of STDD, with setting " + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\epsilon_t^i = \\epsilon_\\theta^i" + }, + { + "bbox": [ + 55, + 324, + 295, + 361 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 69, + 372, + 283, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 372, + 283, + 392 + ], + "spans": [ + { + "bbox": [ + 69, + 372, + 283, + 392 + ], + "type": "interline_equation", + "content": "z _ {t - 1} ^ {i} = \\sqrt {\\bar {\\alpha} _ {t - 1}} (u _ {t} ^ {i} + v _ {t} ^ {i}) + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} ^ {i} + \\sigma_ {t} \\epsilon^ {\\prime}", + "image_path": "cfdf261dadf49b994bf770ff7cffca1ed04c1d465ec522a8f17404c5667c3154.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 70, + 395, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 395, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 70, + 395, + 294, + 420 + ], + "type": "text", + "content": "s.t., " + }, + { + "bbox": [ + 70, + 395, + 294, + 420 + ], + "type": "inline_equation", + "content": "u_{t}^{i} = \\frac{z_{t}^{i} - \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{\\theta}^{i}}{\\sqrt{\\bar{\\alpha}_{t}}},\\quad \\epsilon_{\\theta}^{i} = \\epsilon_{\\theta}(z_{t}^{i},t)," + }, + { + "bbox": [ + 70, + 395, + 294, + 420 + ], + "type": "text", + "content": " (20)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 96, + 422, + 272, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 422, + 272, + 437 + ], + "spans": [ + { + "bbox": [ + 96, + 422, + 272, + 437 + ], + "type": "interline_equation", + "content": "v _ {t} ^ {i} = g (t - 1, i) - g (t, i), \\quad \\tilde {z} _ {0} ^ {i} = \\phi_ {\\theta} (z _ {t} ^ {i}, t),", + "image_path": "c3dc280b7c7d2d86e5f74f79682469ae5321ab7425cd82f7b2775549eb8e75f9.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t^i = \\sqrt{\\bar{\\alpha}_t} * u_t^i" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " is the mean of " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "v_{t}^{i} \\approx F_{t-1}((1 - \\sqrt{\\rho})\\tilde{z}_{0}^{i} - \\sqrt{1 - \\rho}\\tilde{z}_{0}^{i-1})" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " is temporal domain difference, " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon'" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " is the Gaussian noise added during inference and " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " is a network, then " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "v_{t}^{i}" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " can be calculated by Eqn. (7) with " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{0}^{i}" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": ". In practice, " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " are different channel components of the same network output. Moreover, in the process of inference, there is no calculation of the inverse function of " + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 445, + 296, + 542 + ], + "type": "text", + "content": " in Eqn. (16)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "content": "How to understand " + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "content": ": Assuming a new variable " + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "content": ", rewrite " + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "inline_equation", + "content": "z_{t}^{i}" + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "content": " in Eqn. (7) as: " + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "inline_equation", + "content": "z^{i}(t,\\tau) = \\sqrt{\\bar{\\alpha}_{t}} g(\\tau ,i) + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon_{t}^{i}" + }, + { + "bbox": [ + 55, + 543, + 296, + 590 + ], + "type": "text", + "content": ". Then, according to SDE model [34], we conclude that" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 99, + 598, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 598, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 99, + 598, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} d _ {t} z ^ {i} (t, \\tau) \\approx - 1 / 2 \\beta_ {t} u _ {t} ^ {i} + \\sqrt {\\beta_ {t}} d \\omega \\\\ d _ {\\tau} z ^ {i} (t, \\tau) \\approx - \\sqrt {\\bar {\\alpha} _ {t}} v _ {\\tau} ^ {i} \\end{array} , \\right. \\tag {21}", + "image_path": "8af9e3a37da998ea3a74c8e4ddc83b32c10da951d02369701f0ef2defd2c2587.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " gaussian component. Therefore, " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "u_{t}^{i}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " is image spatial diffusion component and " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "v_{t}^{i}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " a temporal diffusion or bridge between frames. Therefore, the process inference consists of a diffusion model (i.e., Eqn. (20)) and a difference between frames, shown in Algorithm 2. In addition, to ensure the association between frames, we set " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "Z_{T} = F(\\epsilon)" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 72, + 451, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 451, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 451, + 84 + ], + "type": "text", + "content": "3.5. Optimize Dual Diffusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": "The dual diffusion architecture includes Variational AutoEncoder (VAE) for video encoding and decoding, diffusion networks " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " for denoising, and conditional embedding networks (if needed). In detail, we employ VAE model training in an unsupervised manner, contains two trainable modules: encoder " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " to represent image " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " as latent map " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " to decode the latent map " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\tilde{Z}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " as output image " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\tilde{X}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": ". VAE was trained with the reconstruction loss [28] using only unlabeled images, or load pre-trained model such as Stable Diffusion v1.4 [28]. Then, we use transformer network, used in [21], to extract the feature of video and learn the noise of dual diffusion process. And time step " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " is specified to " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " by the sinusoidal position embedding [38]. In addition, we enhance the network backbone by using a cross-attention mechanism [28] to transform the model into a more flexible conditional diffusion model. And, we adopt different pre-training condition networks according to different tasks, such as CLIP [26] for text-to-video generation. Implementation details will be presented in Sec. 4.1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 329, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 354 + ], + "type": "text", + "content": "For optimizing STDD, the KL divergence to directly compare " + }, + { + "bbox": [ + 313, + 329, + 554, + 354 + ], + "type": "inline_equation", + "content": "p_{\\theta}(Z_{t - 1}|Z_t)" + }, + { + "bbox": [ + 313, + 329, + 554, + 354 + ], + "type": "text", + "content": " against forward process posteriors is:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 359, + 375, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 375, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 359, + 375, + 553, + 445 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {K L} \\left(q \\left(Z _ {t - 1} \\mid Z _ {t}\\right) | | p _ {\\theta} \\left(Z _ {t - 1} \\mid Z _ {t}\\right)\\right) \\\\ = \\frac {\\left| \\left| \\tilde {\\mu} _ {t - 1} - \\mu_ {\\theta} \\right| \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}} + C \\\\ = \\frac {\\left| \\left| \\tilde {\\mu} _ {t - 1} - \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(U _ {t} + V _ {t}\\right) \\right| \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}} + C \\tag {22} \\\\ \\end{array}", + "image_path": "5ce096a4de1f9e68274c818d4d4f36fcd14bb50e618498dce1157c6aa14d28f4.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "content": " is a constant, " + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "inline_equation", + "content": "U_{t} = (u_{t}^{0}, u_{t}^{1}, \\dots, u_{t}^{N - 1})" + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "inline_equation", + "content": "V_{t} = (v_{t}^{0}, v_{t}^{1}, \\dots, v_{t}^{N - 1})" + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "content": ". We further optimize " + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 313, + 455, + 554, + 504 + ], + "type": "text", + "content": " respectively, thus the loss of the spatio-temporal diffusion probability model is" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 324, + 513, + 497, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 513, + 497, + 528 + ], + "spans": [ + { + "bbox": [ + 324, + 513, + 497, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k l} (\\theta) = \\mathbb {E} | | \\tilde {\\mu} _ {t - 1} / \\sqrt {\\bar {\\alpha} _ {t - 1}} - (U _ {t} + V _ {t}) | | ^ {2},", + "image_path": "79448ebd6ddc8314c19e15adfed6f66275ddbff2d394d183ce0ef47a987b74de.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 326, + 529, + 543, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 529, + 543, + 544 + ], + "spans": [ + { + "bbox": [ + 326, + 529, + 543, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} (\\theta) = \\mathbb {E} | | \\boldsymbol {\\epsilon} _ {t} - \\epsilon_ {\\theta} (Z _ {t}, t) | | ^ {2} + | | Z _ {0} - \\phi_ {\\theta} (Z _ {t}, t) | | ^ {2},", + "image_path": "fdf2b9a914864bcc99857b227dd1c1ab7faad4bc8f42453f9bdff2945bbb426b.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 326, + 545, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 545, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 326, + 545, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathcal {L} _ {k l} (\\theta) + \\mathcal {L} _ {m s e} (\\theta), \\tag {23}", + "image_path": "1ade8d04ea2e5fdbaf2114f89f693fce31a296af41a968a2e215f9053cc662f6.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 568, + 553, + 592 + ], + "type": "text", + "content": " represents the parameters of the STDD network. The detailed training process is shown in Algorithm 1." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 604, + 395, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 395, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 395, + 616 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 624, + 429, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 429, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 429, + 636 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "Datasets: To validate the proposed STDD, we conduct comprehensive evaluations and comparisons on Sky Time-Lapse [47] and UCF101 [35] for video generation, MHAD human action dataset [3] for image-to-video and WebVid-10M [1] for text-to-video. UCF101 is a realistic action video motion recognition dataset, collected from" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12579" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 376, + 286 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 376, + 286 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 376, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 376, + 286 + ], + "type": "image", + "image_path": "dcb21df5221cc789563342326782cc193ab29d704050cefbe5c6a3a8e4bff9b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 295, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 319 + ], + "type": "text", + "content": "Figure 1. Qualitative comparisons on MHAD for video prediction. In the red box, the head in LFDM has serious artifacts and distortions, which we handled well." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 383, + 76, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 383, + 76, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 383, + 76, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 383, + 76, + 545, + 286 + ], + "type": "image", + "image_path": "58e11eaf4822dda8bb2198b2f70b74b926c66bc78279facb91152d1dc466440d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 338, + 294, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 294, + 434 + ], + "type": "text", + "content": "YouTube, providing 13,320 videos from 101 action categories. MHAD human action dataset contains 861 videos of 27 actions performed by 8 subjects. UCF101, Sky Time-Lapse and MHAD are cropped to 16 frames for training our model. For quantitative evaluation, we also train a text-conditioned video generation model on WebVid-10M [1], which consists of 10.7M short videos with paired textual descriptions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 434, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 517 + ], + "type": "text", + "content": "Metrics: Following previous works [7, 20], we mainly use Frechet Video Distance (FVD) [37] as the evaluation metric. On UCF101 and Sky Time-Lapse, we report the results of FVD. Following LFDM [22] on MHAD, we report the results of FVD and class conditional FVD (cFVD). The cFVD compare the distance between real and synthesized video feature distributions under the same class condition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "text", + "content": "Implementation details: We use the transformer network, used in [21], to extract the features of video and learn the noise of the dual diffusion process. The network learning rate is " + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "text", + "content": ", the batch size is 16, and the training stpes is 200000. The diffusion steps " + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "text", + "content": ", the steps of accelerate reverse process is 250 and " + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\rho = 1 - 10^{-3}" + }, + { + "bbox": [ + 55, + 518, + 295, + 602 + ], + "type": "text", + "content": ". We will provide the generated videos and code on github." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 241, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 241, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 241, + 624 + ], + "type": "text", + "content": "4.2. Comparisons with state-of-the-arts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Video prediction. On MHAD dataset, we have conducted a comparison of our methods with several competitive approaches, including ImaGINator [41], VDM [13], LDM [28] and LFDM [22]. Table 1 shows the quantitative comparison between our method and state-of-the-art approaches like LDM and LFDM, where our proposed STDD consistently outperforms all other models. The results show" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 368, + 523, + 440 + ], + "blocks": [ + { + "bbox": [ + 313, + 337, + 555, + 359 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 359 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on MHAD. " + }, + { + "bbox": [ + 313, + 337, + 555, + 359 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 337, + 555, + 359 + ], + "type": "text", + "content": " denotes the lower the better. The best results are denoted in bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 368, + 523, + 440 + ], + "lines": [ + { + "bbox": [ + 347, + 368, + 523, + 440 + ], + "spans": [ + { + "bbox": [ + 347, + 368, + 523, + 440 + ], + "type": "table", + "html": "
MethodResolutionFVD ↓cFVD↓
ImaGINator [41]64 × 648891406
VDM [13]64 × 64295531
LDM [28]128 × 128337594
LFDM [22]128 × 128214426
Ours128 × 128195411
", + "image_path": "4aa0c52107601e0e25252c6bba214920cc5dfd6f158750d161048b279fdff57d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 330, + 495, + 539, + 594 + ], + "blocks": [ + { + "bbox": [ + 313, + 452, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 485 + ], + "type": "text", + "content": "Table 2. FVD values of video generation models on UCF101 and Sky Time-Lapse. The lower the better and the best results are denoted in bold." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 495, + 539, + 594 + ], + "lines": [ + { + "bbox": [ + 330, + 495, + 539, + 594 + ], + "spans": [ + { + "bbox": [ + 330, + 495, + 539, + 594 + ], + "type": "table", + "html": "
MethodUCF101Sky Time-LapseParams
VideoGPT [48]2880222-
MoCoGAN-HD [36]1729164-
DIGAN [50]163083.1-
StyleGAN-V [31]143179.5-
PVDM [51]114175.5-
LVDM [10]37295.2437M
Latte+IMG [21]33342.7673M
Ours25842.1868M
", + "image_path": "279dbd973333876685eb77f8cde9c79def9e8efad2693ab533409a9dc5ba8c4f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "type": "text", + "content": "that our STDD is superior to LFDM (195 vs 214 in FVD at the resolution of " + }, + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 313, + 616, + 555, + 700 + ], + "type": "text", + "content": ", and LFDM achieves the second best performance. LFDM distorts the initial frame through an optical flow sequence to generate a new frame to model the time correlation, while our method is mainly to explicitly model the frame time correlation, which is conducive to the propagation of historical frame information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "Video generation. On UCF101 dataset, we have con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12580" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 309, + 273 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 309, + 273 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 309, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 309, + 273 + ], + "type": "image", + "image_path": "f910ceb3c9c5ee772d41fe9bd42edbc27e71b3d07b5b2dbd3f096af635b95286.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 274, + 235, + 285 + ], + "lines": [ + { + "bbox": [ + 140, + 274, + 235, + 285 + ], + "spans": [ + { + "bbox": [ + 140, + 274, + 235, + 285 + ], + "type": "text", + "content": "(a)UCF101 (Playing Sitar)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 74, + 547, + 273 + ], + "blocks": [ + { + "bbox": [ + 309, + 74, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 309, + 74, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 309, + 74, + 547, + 273 + ], + "type": "image", + "image_path": "789721eea31130d61951b16273775652ed01fc38388515ed40e11a30698931f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 274, + 462, + 285 + ], + "lines": [ + { + "bbox": [ + 392, + 274, + 462, + 285 + ], + "spans": [ + { + "bbox": [ + 392, + 274, + 462, + 285 + ], + "type": "text", + "content": "(b) Sky Time-Lapse" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 570 + ], + "type": "text", + "content": "duced a comparison of our methods with several competitive approaches, including VideoGPT [48], StyleGAN [31], DIGAN [50], TATS [7], VDM [13], among others. The quantitative results on UCF101 dataset are presented in Table 2. In terms of video generation, VDM demonstrates superior performance compared to GAN-based methods, particularly in terms of FVD (Fréchet Video Distance). This indicates the potential of DPM-based video generation methods. Our proposed STDD achieves even better results than LVDM by a significant margin. Compared with Latte-IMG, which prerained on images, our approach is also comparable. In addition, the quantitative results in Sky Time-Lapse also achieve competitive performance, shown in Table 2. These comparisons highlight the effectiveness of STDD in video generation tasks, demonstrating its superior performance over existing methods in terms of various evaluation metrics and datasets. In addition, the inference time of our model did not increase significantly with the same inference steps and device (23s vs 20s of Latte with 250 inference steps in RTX 3090 GPU)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 149, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 149, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 149, + 590 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Visual comparison. On MHAD dataset, we have conducted a qualitative comparison of our methods with several video pridiction models, including ImaGINator, VDM, LDM and LFDM. As shown in Fig. 1, the first line shows the given image, and the rest shows the generated sequence. In general, our approach has advantages over other video generation methods, especially generating 'stand-to-sit' videos with larger variation areas and more motion variations. The results in Fig. 1 (b) showed that the head of the LFDM generated character had severe artifacts and distor" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 331, + 472, + 406 + ], + "blocks": [ + { + "bbox": [ + 142, + 297, + 465, + 309 + ], + "lines": [ + { + "bbox": [ + 142, + 297, + 465, + 309 + ], + "spans": [ + { + "bbox": [ + 142, + 297, + 465, + 309 + ], + "type": "text", + "content": "Figure 2. Qualitative comparisons on UCF101 and Sky Time-Lapse for video generation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 331, + 472, + 406 + ], + "lines": [ + { + "bbox": [ + 317, + 331, + 472, + 406 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 472, + 406 + ], + "type": "image", + "image_path": "282f016cbaed3a84d1927f8c2986f86f87d239f20bac1550bb42b9c4eadd8cf6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 475, + 331, + 548, + 404 + ], + "blocks": [ + { + "bbox": [ + 475, + 331, + 548, + 404 + ], + "lines": [ + { + "bbox": [ + 475, + 331, + 548, + 404 + ], + "spans": [ + { + "bbox": [ + 475, + 331, + 548, + 404 + ], + "type": "image", + "image_path": "2930b40705e576cbe1c2b6cfbc196057282fd4967b0b0b412e04a257e7cb6fdc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 408, + 472, + 483 + ], + "blocks": [ + { + "bbox": [ + 316, + 408, + 472, + 483 + ], + "lines": [ + { + "bbox": [ + 316, + 408, + 472, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 472, + 483 + ], + "type": "image", + "image_path": "5b99d0f0475969e6f631bef24383dcd4baee9ffef2f91932f00092101dc6f537.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 486, + 427, + 496 + ], + "lines": [ + { + "bbox": [ + 369, + 486, + 427, + 496 + ], + "spans": [ + { + "bbox": [ + 369, + 486, + 427, + 496 + ], + "type": "text", + "content": "Frame weights" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 475, + 409, + 548, + 481 + ], + "blocks": [ + { + "bbox": [ + 475, + 409, + 548, + 481 + ], + "lines": [ + { + "bbox": [ + 475, + 409, + 548, + 481 + ], + "spans": [ + { + "bbox": [ + 475, + 409, + 548, + 481 + ], + "type": "image", + "image_path": "627fdcb7974ffff42fd86ae3f52197e434949a688392dd22bd56d66c0bd5cf8a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 486, + 539, + 497 + ], + "lines": [ + { + "bbox": [ + 484, + 486, + 539, + 497 + ], + "spans": [ + { + "bbox": [ + 484, + 486, + 539, + 497 + ], + "type": "inline_equation", + "content": "g(t = 1000,i)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 510, + 555, + 543 + ], + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 543 + ], + "type": "text", + "content": "Figure 3. The tendency of frame weights in Eqn. (7) and Eqn. (8). Appropriate " + }, + { + "bbox": [ + 313, + 510, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 510, + 555, + 543 + ], + "type": "text", + "content": " enhances the similarities and blurring the differences between frames." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 568, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 652 + ], + "type": "text", + "content": "tions, which we handled well, notably in the red box. We also provide visual comparisons with the most recent video generation methods, i.e. TATS, DIGAN, VideoFusion and Latte. As shown in Fig. 2, each generated video has 16 frames with a resolution of " + }, + { + "bbox": [ + 313, + 568, + 555, + 652 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 568, + 555, + 652 + ], + "type": "text", + "content": ", and we show 4-th, 8-th, 12-th and 16-th frame. Our STDD could generate more continuous and moving videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Analysis of " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". To explore the effects of " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " in Eqn (7), we conducted experiments on MHAD with a resolution of " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ", shown in Table 3. Appropriate " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " can better integrate the historical frames and improve the performance. Also, smaller " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " leads to information loss of the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12581" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 74, + 515, + 228 + ], + "blocks": [ + { + "bbox": [ + 95, + 74, + 515, + 228 + ], + "lines": [ + { + "bbox": [ + 95, + 74, + 515, + 228 + ], + "spans": [ + { + "bbox": [ + 95, + 74, + 515, + 228 + ], + "type": "image", + "image_path": "68f7ede029e27228502f78c70961adb77cb65b26b979118c898af2973448436d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 95, + 243, + 309, + 347 + ], + "blocks": [ + { + "bbox": [ + 204, + 231, + 416, + 242 + ], + "lines": [ + { + "bbox": [ + 204, + 231, + 416, + 242 + ], + "spans": [ + { + "bbox": [ + 204, + 231, + 416, + 242 + ], + "type": "text", + "content": "(a) A dog in astronaut suit and sunglasses floating in space" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 243, + 309, + 347 + ], + "lines": [ + { + "bbox": [ + 95, + 243, + 309, + 347 + ], + "spans": [ + { + "bbox": [ + 95, + 243, + 309, + 347 + ], + "type": "image", + "image_path": "d49240d508471c2f7c42b569192af13a89ce2490fa7777734399179096e66f3a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 350, + 272, + 361 + ], + "lines": [ + { + "bbox": [ + 145, + 350, + 272, + 361 + ], + "spans": [ + { + "bbox": [ + 145, + 350, + 272, + 361 + ], + "type": "text", + "content": "(b) A turtle swimming in the ocean" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 513, + 347 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 513, + 347 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 513, + 347 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 513, + 347 + ], + "type": "image", + "image_path": "7745bd384cfb82ad6282abda5bc515a93654ec0209e99ae37a89d4394dd9c522.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 350, + 478, + 361 + ], + "lines": [ + { + "bbox": [ + 345, + 350, + 478, + 361 + ], + "spans": [ + { + "bbox": [ + 345, + 350, + 478, + 361 + ], + "type": "text", + "content": "(c) Surfer on his surfboard in a wave" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 373, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 396 + ], + "type": "text", + "content": "Figure 4. Qualitative evaluation on text-to-video generation. Both our method and Videofusion are trained on WebVid-10M, while Latte is trained on Vimeo25M [42]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 86, + 456, + 266, + 507 + ], + "blocks": [ + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "text", + "content": "Table 3. Analysis " + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "text", + "content": " in Eqn. (7) on MHAD with the resolution of " + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "text", + "content": ". Appropriate " + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "text", + "content": " can improve the performance and " + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\sqrt{\\rho^{1000}}" + }, + { + "bbox": [ + 55, + 415, + 297, + 449 + ], + "type": "text", + "content": " represents the weight of the current frame after 1000 steps." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 456, + 266, + 507 + ], + "lines": [ + { + "bbox": [ + 86, + 456, + 266, + 507 + ], + "spans": [ + { + "bbox": [ + 86, + 456, + 266, + 507 + ], + "type": "table", + "html": "
ρ10.99990.9990.9950.99
\\( \\sqrt{\\rho^{1000}} \\)10.9510.6060.0820.007
FVD ↓157149133243388
", + "image_path": "69296a2f60625c509ed84175ae1a3d2b0ffef4dd10890bed42eb5cf8aa31471e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "type": "text", + "content": "current frame, which is not conducive to video generation. Since the coefficient is exponential, it must be required that " + }, + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "type": "inline_equation", + "content": "0 < 1 - \\rho < 0.01" + }, + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "type": "text", + "content": ", which consistent with the hypothesis in Eqn. (8). The weights of the current and historical frames are shown in Fig. 3. Visualization shows that " + }, + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "type": "inline_equation", + "content": "g(t,i)" + }, + { + "bbox": [ + 55, + 548, + 295, + 631 + ], + "type": "text", + "content": " enhances the similarities and blurring the differences between frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Text-to-video generation. To explore the video generation capabilities of our proposed approach, we extend STDD to text-to-video generation. We train a text-to-video model, pretrained by Stable Diffusion [28], on the large scale video dataset, i.e. WebVid-10M. Some samples are shown in Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 415, + 388, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 388, + 428 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 388, + 428 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 520 + ], + "type": "text", + "content": "In this paper, we propose an explicit spatio-temporal dual diffusion model, mathematically integrating temporal- and spatial-domain diffusion processes. Then, we theoretically derive probabilistic reverse process and propose an accelerated sampling method. Both qualitative and quantitative experiments demonstrated that our model achieve more competitive performance on video generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 529, + 373, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 529, + 373, + 541 + ], + "spans": [ + { + "bbox": [ + 315, + 529, + 373, + 541 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 548, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 319, + 548, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 548, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 319, + 548, + 555, + 602 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 603, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 603, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 319, + 603, + 554, + 658 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 658, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 658, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 658, + 555, + 714 + ], + "type": "text", + "content": "[3] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12582" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "text", + "content": "[4] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 163 + ], + "type": "text", + "content": "[5] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 164, + 295, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 164, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 164, + 295, + 198 + ], + "type": "text", + "content": "[6] Yu Gao, Jiancheng Huang, Xiaopeng Sun, Zequn Jie, Yujie Zhong, and Lin Ma. Matten: Video generation with mamba-attention. arXiv preprint arXiv:2405.03025, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 295, + 255 + ], + "type": "text", + "content": "[7] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and time-sensitive transformer. In European Conference on Computer Vision, pages 102–118. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 256, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 295, + 300 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "type": "text", + "content": "[9] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos. Advances in Neural Information Processing Systems, 35:27953-27965, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 57, + 347, + 295, + 390 + ], + "type": "text", + "content": "[10] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 393, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 393, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 393, + 294, + 415 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 417, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 417, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 417, + 294, + 450 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 452, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 452, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 57, + 452, + 294, + 484 + ], + "type": "text", + "content": "[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In NeurIPS, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 486, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 486, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 486, + 295, + 529 + ], + "type": "text", + "content": "[14] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 532, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 294, + 564 + ], + "type": "text", + "content": "[15] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. arXiv preprint arXiv:2106.00132, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 610 + ], + "type": "text", + "content": "[16] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 657 + ], + "type": "text", + "content": "[17] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[18] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "text", + "content": "[19] Jialin Luo, Yuanzhi Wang, Ziqi Gu, Yide Qiu, Shuaizhen Yao, Fuyun Wang, Chunyan Xu, Wenhua Zhang, Dan Wang, and Zhen Cui. Mmm-rs: A multi-modal, multi-gsd, multiscene remote sensing dataset and benchmark for text-to-image generation. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[20] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Ti-niu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. CoRR, abs/2303.08320, 2023. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[21] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 318 + ], + "type": "text", + "content": "[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18444-18455. IEEE, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "text", + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "text", + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "text", + "content": "[25] Xichen Pan, Pengda Qin, Yuhong Li, Hui Xue, and Wenhu Chen. Synthesizing coherent story with auto-regressive latent diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2920-2930, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 544 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 591, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 646 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12583" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[30] Hiroshi Sasaki, Chris G Willcocks, and Toby P Breckon. Unit-ddpm: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[31] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[32] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[34] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[36] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[37] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in Neural Information Processing Systems, 35:23371-23385, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[40] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[41] Yaohui Wang, Piotr Bilinski, Francois Bremond, and Antitza Dantcheva. Imaginator: Conditional spatio-temporal gan for video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1160-1169, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "text", + "content": "[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[43] Yuanzhi Wang, Zhen Cui, and Yong Li. Distribution-consistent modal recovering for incomplete multimodal learning. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 463 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 22025-22034, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[44] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36:17117-17128, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[45] Yuanzhi Wang, Yong Li, Mengyi Liu, Xiaoya Zhang, Xin Liu, Zhen Cui, and Antoni B Chan. Re-attentional controllable video diffusion editing. arXiv preprint arXiv:2412.11710, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[46] Yuanzhi Wang, Yong Li, Xiaoya Zhang, Xin Liu, Anbo Dai, Antoni B Chan, and Zhen Cui. Edit temporal-consistent videos with image diffusion model. ACM Transactions on Multimedia Computing, Communications and Applications, 20(12):1-16, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "text", + "content": "[47] Wei Xiong, Wenhan Luo, Lin Ma, Wei Liu, and Jiebo Luo. Learning to generate time-lapse videos using multi-stage dynamic generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2364-2373, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[48] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "text", + "content": "[49] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25(10):1469, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[50] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 463 + ], + "type": "text", + "content": "[51] Sihyun Yu, Kihyuk Sohn, Subin Kim, and Jinwoo Shin. Video probabilistic diffusion models in projected latent space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18456-18466, 2023. 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12584" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file